diff --git a/.gitattributes b/.gitattributes index fa1bdcdda66a059fd38f6e4b8e2aca3b61aac8f6..9b9ca7395b6f670eafe210434074bcfb54fba88a 100644 --- a/.gitattributes +++ b/.gitattributes @@ -4772,3 +4772,60 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Segment[[:space:]]Every[[:space:]]Out-of-Distribution[[:space:]]Object/abe08b56-ae22-4da3-bb6d-5f88d5a262ae_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Segment[[:space:]]and[[:space:]]Caption[[:space:]]Anything/959b817d-225c-463d-a58a-99ae6f11b04f_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Selective[[:space:]]Hourglass[[:space:]]Mapping[[:space:]]for[[:space:]]Universal[[:space:]]Image[[:space:]]Restoration[[:space:]]Based[[:space:]]on[[:space:]]Diffusion[[:space:]]Model/c78838dc-b69c-4fcf-b130-147beadbf273_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Selective[[:space:]]Interpretable[[:space:]]and[[:space:]]Motion[[:space:]]Consistent[[:space:]]Privacy[[:space:]]Attribute[[:space:]]Obfuscation[[:space:]]for[[:space:]]Action[[:space:]]Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Selective[[:space:]]Nonlinearities[[:space:]]Removal[[:space:]]from[[:space:]]Digital[[:space:]]Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Selective-Stereo_[[:space:]]Adaptive[[:space:]]Frequency[[:space:]]Information[[:space:]]Selection[[:space:]]for[[:space:]]Stereo[[:space:]]Matching/584de1de-5f52-4125-8106-326f8d8f0441_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Selectively[[:space:]]Informative[[:space:]]Description[[:space:]]can[[:space:]]Reduce[[:space:]]Undesired[[:space:]]Embedding[[:space:]]Entanglements[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Adaptive[[:space:]]Reality-Guided[[:space:]]Diffusion[[:space:]]for[[:space:]]Artifact-Free[[:space:]]Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Calibrating[[:space:]]Vicinal[[:space:]]Risk[[:space:]]Minimisation[[:space:]]for[[:space:]]Model[[:space:]]Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Discovering[[:space:]]Interpretable[[:space:]]Diffusion[[:space:]]Latent[[:space:]]Directions[[:space:]]for[[:space:]]Responsible[[:space:]]Text-to-Image[[:space:]]Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Distilled[[:space:]]Masked[[:space:]]Auto-Encoders[[:space:]]are[[:space:]]Efficient[[:space:]]Video[[:space:]]Anomaly[[:space:]]Detectors/b79b2889-9304-49a2-9631-466d8e496991_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Supervised[[:space:]]Class-Agnostic[[:space:]]Motion[[:space:]]Prediction[[:space:]]with[[:space:]]Spatial[[:space:]]and[[:space:]]Temporal[[:space:]]Consistency[[:space:]]Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Supervised[[:space:]]Dual[[:space:]]Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Supervised[[:space:]]Facial[[:space:]]Representation[[:space:]]Learning[[:space:]]with[[:space:]]Facial[[:space:]]Region[[:space:]]Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Supervised[[:space:]]Multi-Object[[:space:]]Tracking[[:space:]]with[[:space:]]Path[[:space:]]Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Supervised[[:space:]]Representation[[:space:]]Learning[[:space:]]from[[:space:]]Arbitrary[[:space:]]Scenarios/90fe2be8-b444-4296-b490-1791873650b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-Training[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Improved[[:space:]]Visual[[:space:]]Program[[:space:]]Synthesis[[:space:]]With[[:space:]]Visual[[:space:]]Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-correcting[[:space:]]LLM-controlled[[:space:]]Diffusion[[:space:]]Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Self-supervised[[:space:]]Debiasing[[:space:]]Using[[:space:]]Low[[:space:]]Rank[[:space:]]Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SelfOcc_[[:space:]]Self-Supervised[[:space:]]Vision-Based[[:space:]]3D[[:space:]]Occupancy[[:space:]]Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SelfPose3d_[[:space:]]Self-Supervised[[:space:]]Multi-Person[[:space:]]Multi-View[[:space:]]3d[[:space:]]Pose[[:space:]]Estimation/4659f454-3934-4652-8e58-699358f95d03_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SemCity_[[:space:]]Semantic[[:space:]]Scene[[:space:]]Generation[[:space:]]with[[:space:]]Triplane[[:space:]]Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Semantic[[:space:]]Human[[:space:]]Mesh[[:space:]]Reconstruction[[:space:]]with[[:space:]]Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Semantic[[:space:]]Line[[:space:]]Combination[[:space:]]Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Separating[[:space:]]the[[:space:]]'Chirp'[[:space:]]from[[:space:]]the[[:space:]]'Chat'_[[:space:]]Self-supervised[[:space:]]Visual[[:space:]]Grounding[[:space:]]of[[:space:]]Sound[[:space:]]and[[:space:]]Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Sequential[[:space:]]Modeling[[:space:]]Enables[[:space:]]Scalable[[:space:]]Learning[[:space:]]for[[:space:]]Large[[:space:]]Vision[[:space:]]Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SfmCAD_[[:space:]]Unsupervised[[:space:]]CAD[[:space:]]Reconstruction[[:space:]]by[[:space:]]Learning[[:space:]]Sketch-based[[:space:]]Feature[[:space:]]Modeling[[:space:]]Operations/a4171579-c596-406f-93c6-62c45be3925d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Shadow[[:space:]]Generation[[:space:]]for[[:space:]]Composite[[:space:]]Image[[:space:]]Using[[:space:]]Diffusion[[:space:]]Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Shadow-Enlightened[[:space:]]Image[[:space:]]Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Shadows[[:space:]]Don't[[:space:]]Lie[[:space:]]and[[:space:]]Lines[[:space:]]Can't[[:space:]]Bend![[:space:]]Generative[[:space:]]Models[[:space:]]don't[[:space:]]know[[:space:]]Projective[[:space:]]Geometry...for[[:space:]]now/6c6001bd-500a-4e2e-8037-41da45d3b01b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Shallow-Deep[[:space:]]Collaborative[[:space:]]Learning[[:space:]]for[[:space:]]Unsupervised[[:space:]]Visible-Infrared[[:space:]]Person[[:space:]]Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ShapeMatcher_[[:space:]]Self-Supervised[[:space:]]Joint[[:space:]]Shape[[:space:]]Canonicalization[[:space:]]Segmentation[[:space:]]Retrieval[[:space:]]and[[:space:]]Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ShapeWalk_[[:space:]]Compositional[[:space:]]Shape[[:space:]]Editing[[:space:]]Through[[:space:]]Language-Guided[[:space:]]Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Sharingan_[[:space:]]A[[:space:]]Transformer[[:space:]]Architecture[[:space:]]for[[:space:]]Multi-Person[[:space:]]Gaze[[:space:]]Following/df6417d4-023a-4e78-a931-75b2f738abf7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Sheared[[:space:]]Backpropagation[[:space:]]for[[:space:]]Fine-tuning[[:space:]]Foundation[[:space:]]Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Sherpa3D_[[:space:]]Boosting[[:space:]]High-Fidelity[[:space:]]Text-to-3D[[:space:]]Generation[[:space:]]via[[:space:]]Coarse[[:space:]]3D[[:space:]]Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SiTH_[[:space:]]Single-view[[:space:]]Textured[[:space:]]Human[[:space:]]Reconstruction[[:space:]]with[[:space:]]Image-Conditioned[[:space:]]Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Siamese[[:space:]]Learning[[:space:]]with[[:space:]]Joint[[:space:]]Alignment[[:space:]]and[[:space:]]Regression[[:space:]]for[[:space:]]Weakly-Supervised[[:space:]]Video[[:space:]]Paragraph[[:space:]]Grounding/b2a21229-501b-4026-b12f-315fb210d141_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Sieve_[[:space:]]Multimodal[[:space:]]Dataset[[:space:]]Pruning[[:space:]]using[[:space:]]Image[[:space:]]Captioning[[:space:]]Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SignGraph_[[:space:]]A[[:space:]]Sign[[:space:]]Sequence[[:space:]]is[[:space:]]Worth[[:space:]]Graphs[[:space:]]of[[:space:]]Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SimAC_[[:space:]]A[[:space:]]Simple[[:space:]]Anti-Customization[[:space:]]Method[[:space:]]for[[:space:]]Protecting[[:space:]]Face[[:space:]]Privacy[[:space:]]against[[:space:]]Text-to-Image[[:space:]]Synthesis[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/95390555-fbca-4ec8-a42e-5840969340e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SimDA_[[:space:]]Simple[[:space:]]Diffusion[[:space:]]Adapter[[:space:]]for[[:space:]]Efficient[[:space:]]Video[[:space:]]Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Simple[[:space:]]Semantic-Aided[[:space:]]Few-Shot[[:space:]]Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SinSR_[[:space:]]Diffusion-Based[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]in[[:space:]]a[[:space:]]Single[[:space:]]Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Single[[:space:]]Domain[[:space:]]Generalization[[:space:]]for[[:space:]]Crowd[[:space:]]Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Single[[:space:]]Mesh[[:space:]]Diffusion[[:space:]]Models[[:space:]]with[[:space:]]Field[[:space:]]Latents[[:space:]]for[[:space:]]Texture[[:space:]]Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Single[[:space:]]View[[:space:]]Refractive[[:space:]]Index[[:space:]]Tomography[[:space:]]with[[:space:]]Neural[[:space:]]Fields/887b4199-788b-44cf-9fe5-91d50d30037c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Single-Model[[:space:]]and[[:space:]]Any-Modality[[:space:]]for[[:space:]]Video[[:space:]]Object[[:space:]]Tracking/06dd6563-2018-4142-8391-b2e6da00243a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Single-View[[:space:]]Scene[[:space:]]Point[[:space:]]Cloud[[:space:]]Human[[:space:]]Grasp[[:space:]]Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Single-to-Dual-View[[:space:]]Adaptation[[:space:]]for[[:space:]]Egocentric[[:space:]]3D[[:space:]]Hand[[:space:]]Pose[[:space:]]Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SingularTrajectory_[[:space:]]Universal[[:space:]]Trajectory[[:space:]]Predictor[[:space:]]Using[[:space:]]Diffusion[[:space:]]Model/2a853686-8806-4e80-a078-f0cdecbd3caf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Situational[[:space:]]Awareness[[:space:]]Matters[[:space:]]in[[:space:]]3D[[:space:]]Vision[[:space:]]Language[[:space:]]Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Skeleton-in-Context_[[:space:]]Unified[[:space:]]Skeleton[[:space:]]Sequence[[:space:]]Modeling[[:space:]]with[[:space:]]In-Context[[:space:]]Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SketchINR_[[:space:]]A[[:space:]]First[[:space:]]Look[[:space:]]into[[:space:]]Sketches[[:space:]]as[[:space:]]Implicit[[:space:]]Neural[[:space:]]Representations/ab26a978-e483-4292-89b7-9fe0134cf389_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SkillDiffuser_[[:space:]]Interpretable[[:space:]]Hierarchical[[:space:]]Planning[[:space:]]via[[:space:]]Skill[[:space:]]Abstractions[[:space:]]in[[:space:]]Diffusion-Based[[:space:]]Task[[:space:]]Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SkySense_[[:space:]]A[[:space:]]Multi-Modal[[:space:]]Remote[[:space:]]Sensing[[:space:]]Foundation[[:space:]]Model[[:space:]]Towards[[:space:]]Universal[[:space:]]Interpretation[[:space:]]for[[:space:]]Earth[[:space:]]Observation[[:space:]]Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SleepVST_[[:space:]]Sleep[[:space:]]Staging[[:space:]]from[[:space:]]Near-Infrared[[:space:]]Video[[:space:]]Signals[[:space:]]using[[:space:]]Pre-Trained[[:space:]]Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Slice3D_[[:space:]]Multi-Slice[[:space:]]Occlusion-Revealing[[:space:]]Single[[:space:]]View[[:space:]]3D[[:space:]]Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SlowFormer_[[:space:]]Adversarial[[:space:]]Attack[[:space:]]on[[:space:]]Compute[[:space:]]and[[:space:]]Energy[[:space:]]Consumption[[:space:]]of[[:space:]]Efficient[[:space:]]Vision[[:space:]]Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Small[[:space:]]Scale[[:space:]]Data-Free[[:space:]]Knowledge[[:space:]]Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_content_list.json b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5bc446622ff6d5de099fb5c370af54bb5fbc88c1 --- /dev/null +++ b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "Selective, Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition", + "text_level": 1, + "bbox": [ + 212, + 130, + 754, + 176 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20a30826185149aac9aaa823e6ef0fd01965888a5509f7faf4339f067a10a313.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 203, + 285, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e85fb2a714e6b94a3065b4c7169f115b57fed2cfa9d75a5ce179716998ca2446.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 203, + 493, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3aa1d670ae2438f54a386a119664bcd81c22e48eba197fa1b1ba9df0146d9c07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 203, + 658, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9419fbb9a64b9dd1e6a203d4f61065c9aedee8d92065285a64e280b44bde4c12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 203, + 859, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6b16d27e3b4ab0312ed64869d755ee383bcc60a8243d1809e6f07d85fd17fad.jpg", + "image_caption": [ + "Figure 1. Our goal is to hide privacy attributes without action recognition performance dropping. Left: Arbitrary images can be used to specify an interpretable template library defined by privacy attributes. Middle: A salience map is generated from privacy templates; example illustrates use of templates for personal identification. Right: The source video is masked with noise as guided by salience and animated by source video optical flow. Salience makes masking selective to privacy sensitive regions, while preserving scene context; optical flow preserves motion – both of which are critical for action recognition. The obfuscated video can be input directly to arbitrary privacy and action recognition systems without retraining. Zoomed circles highlight details only for illustration." + ], + "image_footnote": [], + "bbox": [ + 78, + 282, + 302, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da0e3093f1ba3ca6d02de67525ec1e1a61a791809e036dd5497d1969b9082244.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 277, + 609, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cccfc1f37d98162a2a7cdc77b62521acbed9ef08f9b8f4198d105ce83cd6f84f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 273, + 890, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 575, + 313, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concerns for the privacy of individuals captured in public imagery have led to privacy-preserving action recognition. Existing approaches often suffer from issues arising through obfuscation being applied globally and a lack of interpretability. Global obfuscation hides privacy sensitive regions, but also contextual regions important for action recognition. Lack of interpretability erodes trust in these new technologies. We highlight the limitations of current paradigms and propose a solution: Human selected privacy templates that yield interpretability by design, an obfuscation scheme that selectively hides attributes and also induces temporal consistency, which is important in action recognition. Our approach is architecture agnostic and directly modifies input imagery, while existing approaches generally require architecture training. Our approach offers more flexibility, as no training is required, and outperforms alternatives on three widely used datasets.", + "bbox": [ + 73, + 595, + 470, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 574, + 630, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advances in state-of-the-art computer vision and machine learning enable deployment of such systems in the public sphere. Accompanying these initiatives, concerns arise for the privacy of individuals that are captured in acquired imagery [11, 19, 35, 46]. In particular, considerations arise regarding attributes that individuals want to keep confidential, yet that are revealed through visual information, even though they are not critical for the functioning of the deployed system, e.g. identity, age, gender and race. Video-based action recognition is an area of consideration as it has potential for widespread applications in surveillance and monitoring. These concerns have sparked interest in privacy preserving action recognition [13, 24, 34, 58]. These approaches process input imagery to obscure privacy attributes while maintaining action recognition performance. Contemporary approaches typically apply their obfuscation across entire input video frames and improvements have been made within this paradigm. Notably, however, there are downsides to this paradigm, as follows.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code available f-iliic.github.io/SelectivePrivacyPreservation", + "bbox": [ + 99, + 887, + 415, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18730", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Collateral damage. Global masking strategies indiscriminately obscure the entire image, impacting regions within the scene that may exhibit high correlations with actions, albeit lack relevance to privacy. For example, it is known that masking objects and scene context can impair action recognition performance [61], yet these are lost in global masking. Furthermore, global masking strategies do not allow for selective attribute obfuscation and generally hide all attributes at once, even when not all are of concern.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Loss of dynamic information. The large change in input modalities from global masking necessitates the retraining of the action recognition module or the design of custom modules (adversarial training), which adds to the challenge of practical deployment. Indeed, even if applied more locally, the obfuscation can compromise the motion of the actors, which also can be important in action recognition [50].", + "bbox": [ + 75, + 229, + 472, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lack of interpretability. Finally, given that state-of-the-art approaches are end-to-end trained with limited concern for interpretability, the exact nature of what is being masked and how it is achieved can remain unclear. Lack of interpretability is an important concern in privacy preservation, because its lack can compromise user trust [29, 38].", + "bbox": [ + 75, + 339, + 472, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Contributions", + "text_level": 1, + "bbox": [ + 76, + 441, + 222, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present an approach to privacy preserving action recognition that responds directly to current limitations in four ways, as illustrated in Fig. 1. (i) The approach is based on local detection and selective obfuscation of privacy sensitive regions. This selectivity maintains global context information that is crucial to action recognition, yet unimportant for privacy. (ii) The local processing avoids large modality shifts in the imagery and is independent of the action recognition module itself; therefore, it does not require algorithm retraining, which is sometimes infeasible. (iii) The masking preserves interframe motion; so, that information is available for action recognition. (iv) The privacy sensitive masking is interpretable by design, and allows inspection through the explicitly generated saliency maps.", + "bbox": [ + 75, + 465, + 472, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 693, + 215, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy in machine learning. The need to protect privacy has garnered increased attention in the vision research community. Current models commonly consume large amounts of web data to learn generalizable representations [20, 33, 53], which inevitably invade personal information, such as identity and location. Moreover, in model deployment it also may be desirable to preserve privacy information. The concern for privacy is not limited to vision research, but extends across artificial intelligence, including natural language processing [5] and more general machine learning [37]. As these technologies find their way into broader society, privacy concerns must be considered.", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy preserving action recognition. We focus on video-based action recognition. Recent developments in this area have yielded systems capable of strong performance on challenging datasets; for review see, e.g. [48]. Similarly, applications, including those in privacy-sensitive scenarios (e.g. surveillance [10] and monitoring [51]), are being developed. To provide useful spatiotemporal signals for action recognition, video clips are mostly collected to capture actors with a great level of clarity throughout the actions, thus increasing the chance of privacy leakage. Moreover, such datasets are expanding rapidly, e.g. [4, 30].", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these concerns, research on recognizing actions while preserving private information has emerged in recent years. Early work concentrated on devising models that can work on low-resolution videos [8, 43, 44]; these methods often operate at the cost of sacrificing action recognition performance. Other work along these lines developed face-anonymization techniques to prevent models from yielding high accuracy on face recognition [28, 42]. Still, those approaches cannot easily extend to other attributes and are correspondingly limited. More recent work has focused on implicitly learning transformation functions to anonymize videos in a data-driven fashion [13, 24, 36, 59], and also extending such approaches to anomaly detection [18]. To reduce model complexity, a competing approach follows a simple, yet effective procedure [34]: Frame subtraction, followed by broadband-filtering to yield motion descriptors for action recognition, while suppressing privacy attributes. A notable limitation of that approach is the relative weakness of frame differences as motion descriptors compared to common alternatives, e.g. optical flow. Another common limitation to all existing work is the lack of flexibility to select arbitrary private information to hide, i.e. preserving privacy of only critical attributes, while maintaining the visibility of others, to avoid obfuscation of visual cues that are essential to action classification.", + "bbox": [ + 496, + 277, + 893, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Template matching. Building correspondences between templates and target images via matching is a foundation for modern vision research, e.g. SfM [21], image correspondences [45], detection [12] and tracking [49]. Features used in matching have advanced rapidly from primarily handcrafted (e.g. [12, 39]) to learning-based convolutional [23] or transformers [14]. Our solution is inspired by techniques seen in feature-based template matching. We apply their insights on matching templates to localize and selectively obfuscate privacy-sensitive regions. More specifically, we adopt DINO-ViT features [6] and compute similarities using the keys of the last attention layer from that architecture. Our use of DINO-ViT keys (rather than queries or values) is based on previous work finding them to perform best when applied to visual correspondence [1, 40].", + "bbox": [ + 496, + 674, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18731", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cd80802ccf9e38d4ba8b524ce926a7f393454301b3166b2d902f36c9109cbf2a.jpg", + "image_caption": [ + "(a) Overview" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 315, + 196 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0001021df2ac512b719a1903e53a751cdc84e614b311782cdc4f2f3e49898646.jpg", + "image_caption": [ + "(b) Matcher", + "Figure 2. Overview of Method. 2a We present a privacy module that builds atop three components: (i) a semantic template library that contains attributes to be hidden, (ii) a descriptor matcher to localize template features in videos to be obscured and, (iii) an obfuscation method that is sensitive with respect to motion present in the scene. 2b A semantic descriptor matcher based on DINO [6]-ViT [14] keys is used to determine privacy salient regions in a video based on the template library. In our case, regions of interest correspond to those that can identify a person; however, this component can be adapted for other privacy attributes through specification of different templates. The result is a saliency map. 2c The saliency map is used as a weight to apply noise to the regions. The noise, however, is not static, but is warped with optical flow with an initial noise pattern image, $N(\\mathbf{p},1)$ , for the purposes of preserving motion information in the source video. The similarity maps of all aggregated relevant privacy attributes are used to weigh the noise and apply it to the input image, obfuscating privacy sensitive information while not destroying the underlying temporal signal." + ], + "image_footnote": [], + "bbox": [ + 351, + 88, + 578, + 198 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb360d16467fe008ad852374cb7588d352e9783feadd1efb0009344f4f7e91ed.jpg", + "image_caption": [ + "(c) Temporally Consistent Selective Obfuscation" + ], + "image_footnote": [], + "bbox": [ + 614, + 88, + 887, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Technical approach", + "text_level": 1, + "bbox": [ + 76, + 375, + 264, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach to privacy preserving action recognition is a stand alone module that operates by preprocessing video that subsequently is input to arbitrary action and privacy recognition algorithms, i.e. it is independent of the recognition algorithms and does not entail any retraining of those algorithms. Our method consists of three key components; see Fig. 2a: (i) A template library that covers privacy attributes to be preserved, (ii) a matcher that produces saliency maps between selected templates and images where privacy is to be preserved and (iii) an obfuscator that uses the saliency maps to hide privacy attributes of concern in a temporally consistent fashion to preserve motion in the video. The remainder of this section details each of these components.", + "bbox": [ + 75, + 400, + 470, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Template library", + "text_level": 1, + "bbox": [ + 76, + 619, + 243, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obfuscate privacy sensitive regions in images in an interpretable fashion, we need an explicit set of “attributes” to be hidden from the target video. Such a concept template library, T, can be built manually choosing image patches corresponding to features one wishes to obfuscate. In this paper, we concentrate on privacy attributes related to personal identification. Therefore, we use landmark anatomical features, corresponding to detailed facial landmarks and the hand (Fig. 3 left: forehead, hair, eye, cheek, lips, hand) as well as larger body parts (Fig. 3 right: arm, torso, legs). The choice to focus on preservation of human identity is motivated by the fact that all three of our evaluation datasets define their privacy attributes on attributes relating to person identity. Notably, however, such a template library easily can be extended depending on the particular task at hand. For example, attributes pertaining to location (e.g. street signs, distinctive scene objects) could guide development", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of a complimentary set of templates. In any case, given a template library, a user can selectively combine templates to obscure attributes of concern in a particular application.", + "bbox": [ + 496, + 377, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we define each template $\\tau_{i}$ as an element in the set $\\mathsf{T} = \\{\\tau_1,\\tau_2,\\dots,\\tau_n\\}$ . Further, let $\\tilde{\\mathsf{T}}\\subseteq \\mathsf{T}$ be the subset of templates that the user selects from the library, $\\mathsf{T}$ , for a particular subset of privacy attributes to be preserved for a given dataset or application. Our manual approach to template selection leads to concepts that are interpretable by design. In particular, we use two source images, shown in Fig. 3, one for small-scale features on the face, and one for large regions, taken from the IPN [3] and SBU [60] datasets. We choose these two datasets because of their complimentarity in template selection: IPN focuses on small scale features (e.g. facial and hand), while SBU focuses on larger scale features (e.g. larger body parts), as detailed later in Sec. 4.1. The choice of the particular template images is not critical to the functioning of our approach, because we extract semantic features from the templates that are known to generalize well across images instances [6], as described next.", + "bbox": [ + 496, + 424, + 892, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Matching: Local patch descriptor templates", + "text_level": 1, + "bbox": [ + 500, + 709, + 874, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic features. Our requirements for good features to match between privacy templates in our library and frames in an action recognition video are straightforward: We require (i) semantically rich features that generalize across (in our case) different individuals and (ii) high spatial resolution to perform privacy obfuscation in a localized manner without destroying regions that are required for action recognition. We find that DINO-ViT features fit our needs very well [6, 56]: (i) They have been trained to yield high similarity for semantically related concepts (e.g. objects and their parts), while suppressing the similarity of unrelated", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18732", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87f98b68be415c66c55c4db4b8056043811dc2e38745915d0d25612367d5c7dc.jpg", + "image_caption": [ + "Figure 3. Template Library consisting of Patches Chosen from Anatomical Landmark Regions. These specific images are passed through a DINO-ViT feature extractor. The keys, corresponding to spatial locations of the highlighted patches, are chosen as the templates for matching to input images to obtain semantically similar regions for obfuscation." + ], + "image_footnote": [], + "bbox": [ + 80, + 92, + 305, + 195 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab1be94bbfe9748f517c48530b3a0601a0b534c117fedcf324a51e92309b390b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 90, + 464, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "matters (e.g. background). (ii) They support local patch feature extraction over high resolution images without losing global context. Elsewhere, DINO-ViT keys have proven especially useful in matching between templates and target images [1, 40]. Following these advances, we use last attention layer DINO-ViT keys computed from privacy templates (e.g. Fig. 3) to match against input images to compute privacy saliency maps.", + "bbox": [ + 75, + 320, + 468, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Privacy saliency matching. To find privacy salient regions in an input video, a similarity map, $S$ , is computed between each frame, $I$ , in the video and the selected set of privacy templates, $\\tilde{\\mathsf{T}}$ . The image is tiled with $m$ patches, $I_{j}, j \\in \\{1,\\dots,m\\}$ , and for each patch DINO-ViT keys, $\\mathsf{K}(I_j)$ , are extracted. These features are matched with DINO-ViT keys, $\\mathsf{K}(\\tau_i)$ , for all selected templates, $\\tau_{i} \\in \\tilde{\\mathsf{T}}$ . Clipped cosine similarity is used to establish the saliency of each image patch according to", + "bbox": [ + 76, + 448, + 470, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {j} = \\frac {1}{| \\tilde {\\mathbf {T}} |} \\sum_ {i = 1} ^ {| \\tilde {\\mathbf {T}} |} \\max \\left(0, \\frac {\\langle \\mathrm {K} (\\tau_ {i}) , \\mathrm {K} (I _ {j}) \\rangle}{\\| \\mathrm {K} (\\tau_ {i}) \\| \\| \\mathrm {K} (I _ {j}) \\|}\\right), \\forall j \\in [ [ 1, m ] ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 597, + 468, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle \\cdot ,\\cdot \\rangle$ is inner product and $|\\tilde{\\mathsf{T}} |$ is the cardinality of $\\tilde{\\mathsf{T}}$ the set of selected privacy templates. Clipping is used because only positive values imply salience. This calculation is performed for every patch in every image of the video. Subsequently, the $m$ saliency patches are reassembled in the shape of the original image to produce the final saliency map,", + "bbox": [ + 75, + 648, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\mathcal {R} \\left(s _ {1}, \\dots , s _ {m}; h, w\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 756, + 468, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\mathcal{R}$ a function that accepts image tiles, $s_j$ , and reshapes them into their original image format of height $h$ and width $w$ . Note that a separate salience map, $S$ , is calculated for each frame in a video of interest, $I$ . The entire process of feature extraction, matching and the resulting saliency maps are summarized in Fig. 2b. Example saliency maps in Fig. 4 illustrate the ability of our approach to capture all of our templates in a variety of scenarios.", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporally consistent obfuscation", + "text_level": 1, + "bbox": [ + 500, + 90, + 797, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to mask privacy sensitive regions in images to obfuscate them. If we were to apply masks to every frame independently, then temporal information important to action recognition, e.g. motion of actors, would be destroyed as well. We empirically document the issue in Sec. 4.2. In response to this challenge, we follow previous work that presented a method for producing temporally consistent spatial noise across videos that supported action recognition, while obscuring appearance information in single frames [25]. While the previous work applied noise patterns uniformly across entire frames, we instead weight them by our privacy salience maps, $S$ , to preserve as much context information as possible. The remainder of this subsection details our approach, with an outline shown in Fig. 2c.", + "bbox": [ + 496, + 113, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Noise pattern initialization. Let $\\mathbf{p} = (x, y)$ be image coordinates and $t$ time. We initialize a noise image, $N(\\mathbf{p}, 1)$ , with the same dimensions as a frame from the input video (i.e. $h \\times w \\times 3$ ), with $h$ height, $w$ width and 3 the number of colour channels. The dataset mean, $\\mu$ , and standard deviation, $\\sigma$ , are used to define a uniform distribution from which individual pixel intensities are drawn according to", + "bbox": [ + 496, + 330, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN (\\mathbf {p}, t = 1) \\sim \\mathcal {U} [ \\mu - \\sigma , \\mu + \\sigma ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 445, + 890, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion consistent noise. To create motion consistent noise the initial random frame, $N(\\mathbf{p}, 1)$ , is warped forward with flow fields derived from the original video, $I(\\mathbf{p}, t)$ , as extracted by an optical flow algorithm. Let $\\mathbf{v}(\\mathbf{p}, t) = (u(\\mathbf{p}, t), v(\\mathbf{p}, t))$ be the flow field that maps points, $\\mathbf{p}$ , in frame $t$ to those in frame $t - 1$ , with $u$ and $v$ the horizontal and vertical components of the flow. Then, a motion consistent noise sequence is generated as", + "bbox": [ + 496, + 476, + 892, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN (\\mathbf {p}, t) = N \\left(\\mathbf {p} + \\mathbf {v} (\\mathbf {p}, t), 1\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 604, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $t \\in \\{2, \\dots, T\\}$ and $T$ the number of frames in the original input video.", + "bbox": [ + 496, + 628, + 890, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selective privacy obfuscation. The computed video sequence, $N$ , shows no single frame appearance related to the original video as it is random noise; however, when viewed as a video it reveals the motion present in the original, cf. [25]. Direct use of this synthesized video obscures privacy attributes; however, it also obscures other context information that could be of use in action recognition. So, instead we selectively apply $N$ to every frame $I$ in the video by using the privacy salience maps $S$ , according to", + "bbox": [ + 496, + 665, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO (\\mathbf {p}, t) = I (\\mathbf {p}, t) + \\left(S (\\mathbf {p}, t) \\times \\left(N (\\mathbf {p}, t) - I (\\mathbf {p}, t)\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 809, + 890, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The resulting video, $O$ , contains selectively obfuscated regions, built with interpretable templates by design and contains motion information that (in principle) does not differ from the original input video.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18733", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/393cb808ada3384756f3cfcb6fb3848694468154b8f7f136eef732f817a541c3.jpg", + "image_caption": [ + "Figure 4. Saliency Maps for Descriptors in the Template Library. The manual selection of these templates allows for interpretability of the obfuscated parts of the image by design. The matched DINO-ViT features capture rich semantic information and allow for detailed spatial localization due to the nature of vision transformers. These saliency maps can then be combined for obfuscating any combination of templates depending on the task at hand." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 158, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/385e2b1c70afa594723da8bf34675e4199f1c012ba5660100d32790efc41a6a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 90, + 238, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/96aa1c9561d7f95c812a34214e2bd196b1070a08eb06e63d7d9cbb1760b8196e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 90, + 320, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eca47d4f1a9f4c1a1a330221b5ea3742f3053d9ab9f74854c2a5c5c3e9cc4c5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 90, + 401, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/67661313effcfe1e50450dac036f4e13da91324ffd9a8d422d10b4b59d3707a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 90, + 482, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0138c17a11f2f31975363bcfbec31f46a84aafd6abdf2012011f8be55ffd016e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 90, + 563, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/48454e4984540219b0e5cbf8c31af116f1e1bb4717811c005447c82d6f201803.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 90, + 645, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfc3bb592be2fd5d95212638237843e46953e8c3604e8ce16bd5825ff95dc22a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 90, + 727, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/408507cccd7eaef775fa8935ba63295d1e3aa4ad5bf3a3fbf16819f28bc0f5bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 90, + 807, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ca77c0593b1eee06660f5b8134f754cb13fa0e2c128961af5722c09a1de0b412.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cfecd684f96f8771134147364d00b9ceb3a58d319a9aa12deee01e600fbfdd87.jpg", + "table_caption": [ + "ACTION RECOGNITION ON SOURCE DATASETS" + ], + "table_footnote": [ + "Table 1. Top-1 Accuracy for all Privacy and Action models on the original unmodified (source) videos, i.e. without privacy obfuscation. Bold and underline indicate first and second best, resp. Number of frames and temporal sampling rate indicated as $f \\times r$ ." + ], + "table_body": "
Networkf × rIPNKTHSBU
C2D [57]8 × 874.5683.0087.50
CSN [55]32 × 291.7388.3390.91
E2S X3D L [25]16 × 595.1595.3391.11
E2S X3D M [25]16 × 589.3892.3386.36
E2S X3D S [25]13 × 685.1692.0082.98
I3D [7]8 × 883.1589.6782.95
MVIT[15]16 × 488.0090.0092.55
R2+1D [54]16 × 489.8087.3381.91
Slow [17]8 × 885.7689.0088.64
SlowFast [17]32 × 288.0687.3390.00
X3D L [16]16 × 594.4194.0090.22
X3D M [16]16 × 591.6793.0081.91
X3D S [16]13 × 687.8190.6775.00
Average88.0590.1586.31
PRIVACY PRESERVATION FOR SOURCE DATASETS
NetworkIPNKTHSBU
ResNet18 [22]88.4687.3390.43
ResNet50 [22]92.3190.0096.81
ResNet101 [22]94.2394.0084.04
ViTb/16 [14]94.2394.0079.79
ViTb/32 [14]90.3894.0078.72
Average91.9291.8785.96
", + "bbox": [ + 78, + 309, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Empirical evaluation", + "text_level": 1, + "bbox": [ + 76, + 722, + 277, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The privacy obfuscated video, (5), serves directly as input to action and privacy recognition. No specialized development, training or other modification of the recognition algorithms nor adaptation of the privacy preservation system is necessary. Indeed, a major difference compared to competing state-of-the-art approaches (e.g. [34, 43, 59]) is that we do not retrain networks with our obfuscated data, while they do retrain. We exploit the fact that privacy attributes are generally independent from action recognition cues, which is enabled by our unique selective obfuscation approach.", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Protocol", + "text_level": 1, + "bbox": [ + 500, + 300, + 601, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use three datasets commonly used for investigating action recognition and privacy, IPN [3], SBU [60] and KTH [47], which pose different challenges concerning both action and privacy.", + "bbox": [ + 496, + 323, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "IPN is a large-scale video-based hand gesture recognition dataset that consists of 50 actors performing 13 static or dynamic gestures, against three different backgrounds [3]. The performed gestures are used as the action labels and actor genders are used as the privacy labels.", + "bbox": [ + 496, + 383, + 890, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SBU is a video dataset depicting eight human interactions. Each video is a video of actor-pairs, in the same laboratory environment. Action labels are derived from the actor interactions and the privacy attributes are the unique pairings of seven different actors resulting in 13 privacy labels.", + "bbox": [ + 496, + 459, + 890, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KTH has 25 actors doing one of six actions [47]. Each action is performed four times in different environments. The six action classes are used for action recognition and 25 actor identities serve as privacy labels. In the originally proposed splits, videos of any one actor are fully contained in one set, since the intent of the KTH dataset was not concerned with actor identity recognition. For privacy identity, each actor must appear in both sets; so, the data is split such that each action is performed twice by each person in the training set, and once in the validation and test sets.", + "bbox": [ + 496, + 535, + 890, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Privacy recognition results are obtained following standard practice [34], averaging outputs over multiple frames (32 for IPN and KTH, and 16 for SBU) from the same video. Action recognition results are obtained by reporting the top-1 accuracy, as per convention [24, 58, 59].", + "bbox": [ + 496, + 686, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It also is interesting to gauge the trade-off between action and privacy recognition. Let $0 \\leq a \\leq 1$ be normalized action recognition performance derived from dividing accuracy percentages by 100, and let $p$ be defined analogously for privacy recognition. To quantify their trade-off we use a linear combination of $a$ and $1 - p$ and define", + "bbox": [ + 496, + 762, + 890, + 852 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\lambda} (a, p) = (1 - \\lambda) a + \\lambda (1 - p), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 861, + 890, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $0 \\leq \\lambda \\leq 1$ weighing the relative importance of action", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18734", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c32ab3289d0c3ea9673aedc497605898b3b92c621699edd203f27d69fcc4f83a.jpg", + "image_caption": [ + "Figure 5. Obfuscation with a Single Attribute and the Impact on Performance. Attribute importance is dataset dependent. For example, notice how the 'Hand' template contributes to a large decrease in action recognition performance on IPN, as the action is determined solely by the hand, whereas on SBU it does not. Optimally blue is high and red is low. Corresponding qualitative examples of saliency maps for each individual template are shown in Fig. 4. Bold text along the abscissa of each plot indicates the templates used for the final results." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 349, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/48b07ac760a5cea5141dff92858a6fcba3687aceb3a0e4aa908559c068c88661.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 614, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6e6b41fce5d496f4afc9d92b0f3fbf2c2b503e078f288bf163871a499117817.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 90, + 880, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f86584d0f0a14883dc713d8206978360ea165299a85c2039fa0bef242006685b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 292, + 176, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1888d6682e9aad01a779860a77a1508c62aad6d73f6632d7b9f4717bb894e604.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 292, + 269, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8b8eb8fdee01620a612a400d88df91adaf82b28df088ef324b01e83f3dcfbfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 292, + 364, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/388c0d6c7f41982a1f29deb40f1da619ba0c6fb13cdd949c6543bcb893ab4abd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 292, + 459, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da0848338641762a6c711f60f856772bbee510eb4619b6928a565b5bb9e71722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 367, + 176, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3c98780a73bd5540276e4e877f578689a983daae423c64252319eacdaf0f723.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 367, + 271, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f6802e627f6b6aacca0c7b0fe71f31281d73b93fd33a28e28870e21bb41eeef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 367, + 367, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/99bff6933cd376107dff4c5d4515f11b049fc0dee9194a377feb64c7054fd2b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 367, + 464, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0ffec1a07cc18dfc1d3bc23e7fcab63829511580fbcde41c9db5c037cf18442.jpg", + "image_caption": [ + "ELR [44]" + ], + "image_footnote": [], + "bbox": [ + 93, + 436, + 181, + 506 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3f411b3776886002ea48644423e1c1c1614f19d08325bb176802ef0ed9b0778.jpg", + "image_caption": [ + "BDQ [34]" + ], + "image_footnote": [], + "bbox": [ + 187, + 436, + 274, + 506 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3f598f0afd372a17bbdb3c97a431af731c749233b381dfe23c3c6ebb6fba902.jpg", + "image_caption": [ + "ALF [59]", + "Figure 6. Optical flow from two consecutive frames from competing approaches. Only our approach retains dynamic information that supports high-quality optical flow recovery. Top row: Single frames as processed by comparison algorithms. Middle row: Full frame optical flow recovery. Bottom row: Zoomed optical flow details. Optical flow shown in Middlebury colour coding [2]." + ], + "image_footnote": [], + "bbox": [ + 279, + 436, + 369, + 507 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29c4ded6542c9c0f4a50bc61e275dbfe845beef23912b749877a2bfdc4ee4ce6.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 375, + 436, + 464, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "recognition vs. privacy. This metric is suitable to create a linear ranking as $a$ , and $(1 - p)$ are optimally 1 ensuring $f_{\\lambda} \\in [0,1]$ . If privacy or action recognition are not equally important, then they can be weighted accordingly; e.g. if privacy is critical, then $\\lambda$ can be increased. In our main experiments we use $\\lambda = 0.5$ , abbreviated as $f_{0.5}$ in Tab. 2 and show an ablation over different values of $\\lambda$ in Fig. 7.", + "bbox": [ + 75, + 633, + 468, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competing obfuscation approaches. We briefly describe NAIVE BASELINE approaches that solely rely on full-frame pixelation and blurring as well as actor masking that have been studied across the literature [13, 58]. We also highlight other State-of-The-Art approaches (SOTA) [34, 43, 59] that we compare against. Visual examples of the SoTA approaches are shown in Fig. 6; also shown are optical flows of consecutive frames for qualitative comparison.", + "bbox": [ + 75, + 746, + 468, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mask Obfuscation. All our datasets involve people performing actions and the privacy attributes are related to geo-", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ple. A naive way to preserve privacy in such videos is to completely mask out the actors. To do so, we use the YOLOv8 implementation [27] based on the original YOLO [41]. We report results based on the masked region filled with the mean intensity of the image covered by the mask.", + "bbox": [ + 496, + 292, + 890, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pixelation. Pixelation applies average pooling over regions of dimensions $x \\times x$ on the input image sequence. We choose two scales of $x \\in \\{4,16\\}$ for the patch-size to pool. In the result tables these methods are abbreviated as $\\operatorname{Pix}_{x \\times x}$ .", + "bbox": [ + 496, + 371, + 890, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Blur. Blur applies a Gaussian blur to the images that have been rescaled to $224 \\times 224$ pixels. The parameters of the Gaussian are the kernel size, $\\kappa$ , and the standard deviation of the kernel, $\\sigma$ . We choose values for weak and strong blurs, $\\kappa = 13$ and $\\kappa = 21$ respectively, as consistent with other work on obfuscation methods [59].", + "bbox": [ + 496, + 435, + 890, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ELR initially reduces frames to Extreme Low Resolution and subsequently applies a set of learned inverse superresolution transforms to support action recognition [43].", + "bbox": [ + 496, + 529, + 890, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ALF is an Adversarial Learning Framework for action recognition that takes into account a privacy budget [59].", + "bbox": [ + 496, + 578, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BDQ is a privacy-preserving encoder that sequentially Blurs, Differences and Quantizes frames. The blur and quantization parameters are learned to maximize action recognition while minimizing privacy recognition [34].", + "bbox": [ + 496, + 612, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, all the compared SoTA methods operate across entire frames (i.e. without selectivity) and have limited interpretability due to their learning-based obfuscation.", + "bbox": [ + 496, + 680, + 890, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details.", + "text_level": 1, + "bbox": [ + 500, + 733, + 669, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. No large scale dataset exists that allows for joint training of action and privacy classification on the same videos. Therefore, we pretrain our action and privacy networks on Kinetics400 [7] and Imagenet1k [33], respectively, and then finetune for specific datasets. We use the AdamW [31] optimizer with a learning rate of $3e^{-4}$ . The networks are trained with a patience scheme of 100 epochs that monitors the loss on the validation set. For action recognition, videos are temporally uniformly subsampled according to the architecture (Tab. 1, $f \\times r$ column).", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18735", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/85ce39d98ec7dd4b2e3fcc83c564d2e5f07d73f1e5eed4ca217001d86fe550a7.jpg", + "table_caption": [ + "RESULTS ACROSS NAIVE BASELINE OBFUSCATION METHODS" + ], + "table_footnote": [], + "table_body": "
MethodIPNKTHSBU
↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ
Original88.0591.920.4890.1591.870.4986.3185.960.50
Masking36.4564.870.3637.5435.220.5152.6230.10.61
Pixelate4×485.5973.650.5683.6459.390.6273.2446.270.63
Pixelate16×1649.8165.760.4238.3338.430.5025.5329.840.48
Blur κ=13σ=10 (weak)76.2467.400.5444.5937.730.5372.7535.690.69
Blur κ=21σ=10 (strong)58.8065.920.4630.3831.840.4957.7734.210.62
Ours87.1151.930.68+0.1188.675.460.92+0.2986.7413.190.87+0.18
RESULTS ACROSS SOTA OBFUSCATION METHODS
BDQ [34]81.0059.000.6191.117.150.9284.0434.180.75
ALF [59]76.0065.000.5685.8919.270.8382.0048.000.67
ELR[44] s=1670.8264.320.5391.2288.860.5196.2782.970.57
ELR[44] s=3252.9663.290.4585.5782.560.5292.4264.890.64
ELR[44] s=6431.6362.700.3456.2158.350.4980.0543.610.68
Ours †85.2551.670.67+0.0689.444.310.93+0.0184.0411.700.86+0.11
Ours87.1151.930.68+0.0788.675.460.92±0.0086.7413.190.87+0.12
", + "bbox": [ + 76, + 102, + 893, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison between Ours vs. Naive Baseline (top) and SOTA (bottom) Approaches as Top-1 Accuracy for both Action and Privacy labels, as well as $f_{0.5}$ Introduced in Sec. 4.1. We show Ours and Naive Baseline results averaged across all recognition algorithms presented in Tab. 1. Competing SOTA approaches only present results on one specific algorithm for action recognition (I3D [7]) and privacy attribute detection (ResNet50 [22]) as those approaches are network specific. To showcase a fair comparison we also present results with the same single recognition algorithms indicated as \"Ours †\". First and second best results indicated by bold and underline, respectively. Relative performance delta ( $\\Delta$ ) is indicated in green (improvement) and orange (tie) between best and second best method.", + "bbox": [ + 73, + 377, + 893, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that we never perform any training on videos obfuscated by our approach. To implement warping, (4), we use pretrained RAFT to extract optical flow [52].", + "bbox": [ + 75, + 470, + 470, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Privacy template selection. Our approach affords selective combination of a predefined set of privacy templates for a given dataset or application. For the experiments, we choose a subset of templates, $\\tilde{\\mathsf{T}}$ , from our complete identity preserving template library, $\\mathsf{T}$ , shown in Fig. 3, to optimize performance on a given dataset. Figure 5 shows template-wise performance based on both action recognition and privacy for each dataset. For SBU and KTH, action recognition is stable with respect to templates; however, the same four templates notably reduce privacy recognition. For consistency on IPN, we also select the four templates that most reduce privacy recognition, even while preserving action recognition; although, the distinction is less striking. This selection process results in templates $\\tilde{\\mathsf{T}} = \\{\\text{torso, arm, leg, hair}\\}$ for SBU and KTH vs. $\\tilde{\\mathsf{T}} = \\{\\text{cheek, eyes, forehead, hair}\\}$ for IPN. Our code yields more results on the overlap of individual saliency maps. Runtime. The constant overhead to produce our obfuscated videos is $\\approx 100ms/frame$ on a NVIDIA RTX4080 GPU.", + "bbox": [ + 75, + 518, + 472, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 816, + 171, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our approach to privacy preserving action recognition on 13 different action recognition models and five different privacy attribute recognition models. Competing approaches evaluate on at most five action models", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[34, 43, 59]. Table 1 shows results on the original videos from IPN [3], SBU [60] and KTH [47], i.e. without any privacy preserving processing. Table 2 shows results comparing our approach vs. NaIVE BASELINES and SOTA. Comparison is given as the average across recognition algorithms presented in Table 1. Notably, while our approach compromises at most $1.48\\%$ action recognition accuracy compared to performance on the original video (88.67% vs. $90.15\\%$ on KTH), it always greatly improves privacy.", + "bbox": [ + 496, + 470, + 890, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Naive Baselines. Compared to the baselines, our approach scores highest with respect to the $f_{0.5}$ metric across all three datasets: IPN (+0.11), KTH (+0.29), and SBU (+0.18). Naive BASELINES show that action recognition correlates highly with privacy performance, as $f_{0.5}$ hovers around 0.5. The notable exception is $\\mathrm{Pix}_{4\\times 4}$ that performs well on IPN and KTH, and is the second best Naive BASELINE on SBU. A notable benefit of all these approaches is that they do not require retraining of the recognition algorithms. Still, none have a selective ability to obfuscate only certain parts. State of the art. A noteworthy difference between our vs. the alternative SoTA approaches is that they only work with the recognition algorithms on which they were trained. While this fact disadvantages our approach, as it lacks such retraining, it still outperforms the competing methods on all datasets: IPN (+0.06), KTH (+0.01), and SBU (+0.11).", + "bbox": [ + 496, + 609, + 892, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is valuable to consider obfuscation approaches in terms of the relative importance of action recognition vs. privacy preservation. Figure 7 compares results for the SoTA ap", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18736", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c12f8d35a275018076c280720822421dc99213da6e5dccd32577da5169ea4e3.jpg", + "image_caption": [ + "Figure 7. Performance across all datasets with varying $\\lambda$ ; see (6). Values of $\\lambda$ closer to 0 weigh action recognition higher, whereas values closer to 1 increase the importance of privacy preservation." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 464, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proaches as that trade-off is varied in terms of $f_{\\lambda}$ , (6). The sweep of $\\lambda$ shows that our obfuscation approach performs better than any competing approach across the entire range. Especially with increasing $\\lambda$ , the gap to other approaches widens as more emphasis is put on privacy preservation.", + "bbox": [ + 75, + 253, + 467, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of selective obfuscation. Our approach is unique in its ability to selectively obfuscate particular regions within a video based on specific privacy attributes. This ability yields two benefits:", + "bbox": [ + 75, + 332, + 468, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(i) Selectivity aids in interpretability. Each selected template results in a saliency map, (2), which allows for visual inspection of what information is being obscured. Figure 4 highlights this benefit as the heat maps reveal the degree of obfuscation to be applied on a template-by-template basis.", + "bbox": [ + 75, + 395, + 468, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(ii) Individual privacy templates can be chosen and combined for best performance; see Fig. 5. Depending on the dataset, different privacy templates differently impact the performance of privacy and action recognition. Action recognition on SBU and KTH is relatively robust to privacy template selection; however, privacy preservation is best when a subset of templates (torso, leg, arm, hair) is selected and the rest remain unobfuscated. This fact derives directly from the saliency calculation, (1): If multiple individual saliency maps have small values, then other strong responses are scaled down in the combined final saliency map. Subsequently, this effect leads to less noise being applied in our selective obfuscation, (5), which can deteriorate privacy preservation. In contrast, action recognition on IPN is compromised if the hand template is selected (as expected with a gesture centric dataset), which documents that simply obfuscating the entire person leads to inferior action performance compared to selectively applied obfuscation.", + "bbox": [ + 75, + 472, + 468, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of motion consistent noise. Action recognition is complicated as different datasets, and even individual actions within a dataset might require different recognition capabilities and can rely to different degrees on the modeling of motion [9]. This uncertainty on the role of motion in action recognition is exacerbated by the fact that different deep learning-based architectures are successful in capturing motion to varying degrees [32].", + "bbox": [ + 75, + 748, + 468, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To see the importance of our proposed motion consistent noise for privacy preservation while maintaining good", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/081822a89c02b238f22cbf300ef27ae3793c55608a9d1f71827e13dd308fd9ec.jpg", + "image_caption": [ + "Figure 8. Comparing Temporally Consistent vs. $iid$ Obfuscation for Action Recognition and Privacy Preservation. Action recognition performance decreases if noise is not temporally consistent." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 697, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0475d0648df86f4dd5d69147a4c7e55028cefd9a8e00c3ef7324b00ae78f784d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 90, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "action recognition, we compare to another version of our pipeline that obfuscates video frames using independent, identically distributed (iid) noise; see Fig. 8 and project page for video results. For all datasets, action recognition performance decreases if the noise is $iid$ . This result is sensible, because the $iid$ noise does not capture the temporal dynamics of the source video. In contrast, privacy preservation is robust to noise in either case. There is less impact on IPN action recognition, as the critical hand motion is never obfuscated, which underlines the importance of selective masking. Further insight on why motion consistent noise supports better action can be had through consideration of Fig. 6, where the ability of such noise to support optical flow estimation is illustrated.", + "bbox": [ + 496, + 243, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Inevitably, there will be a trade-off between action recognition and privacy, as relevant information may be shared. In our approach, that trade-off could happen because our temporally consistent noise maintains dynamic information important for action recognition; however, it also might support motion based identification, e.g. gait recognition. Current protocols in privacy preserving action recognition do not consider motion-based identification. However, gait and related motion-based measurements are weak biometrics [26]; so, favouring action recognition may be apt. Moreover, if it becomes a concern, then it may be possible to apply motion perturbations that impede personal identification while maintaining action recognition.", + "bbox": [ + 496, + 457, + 890, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 665, + 617, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work highlights that it is not necessary to train action recognition and privacy networks in an adversarial fashion for effective obfuscation of privacy attributes while maintaining strong action recognition performance. We show that a system based on local privacy templates, deep features that capture template semantics and selective noise obfuscation that is animated with source video motion can uphold privacy without hindering action recognition. Our approach is unique compared to alternative recent approaches in terms of interpretability and independence from particular action and privacy recognition algorithms. Our nine manually chosen templates, in combination with our proposed obfuscation technique outperforms other state-of-the-art approaches across three different datasets.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18737", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. ECCVW, 2022. 2, 4", + "[2] Simon Baker, Daniel Scharstein, JP Lewis, Stefan Roth, Michael J Black, and Richard Szeliski. A database and evaluation methodology for optical flow. IJCV, 92:1-31, 2011. 6", + "[3] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. IPN Hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In ICPR, 2021. 3, 5, 7", + "[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2", + "[5] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium, 2021. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3", + "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 5, 6, 7", + "[8] Jiawei Chen, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Semi-coupled two-stream fusion ConvNets for action recognition at extremely low resolutions. In WACV, 2017. 2", + "[9] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't I dance in the mall? Learning to mitigate scene bias in action recognition. NeurIPS, 32, 2019. 8", + "[10] Robert T. Collins, Alan J Lipton, and Takeo Kanade. Introduction to the special section on video surveillance. IEEE TPAMI, 22(8):745-746, 2000. 2", + "[11] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (SMID): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 1", + "[12] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005. 2", + "[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. SPAct: Self-supervised privacy preservation for action recognition. In CVPR, 2022. 1, 2, 6", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3, 5", + "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Christoph Feichtenhofer. X3D: Expanding architectures for efficient video recognition. In CVPR, 2020. 5", + "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 5", + "[18] Joseph Fioresi, Ishan Rajendrakumar Dave, and Mubarak Shah. Ted-SPAD: Temporal distinctiveness for self-supervised privacy-preservation for video anomaly detection. In ICCV, 2023. 2", + "[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, and Omer Ovenc. Scalable detection of offensive and non-compliant content/logo in product images. In WACV, 2020. 1", + "[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2", + "[21] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2", + "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2", + "[24] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. PrivHAR: Recognizing human actions from privacy-preserving lens. In ECCV, 2022. 1, 2, 5", + "[25] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In ECCV, 2022. 4, 5", + "[26] Anil K Jain, Patrick Flynn, and Arun A Ross. Handbook of Biometrics. Springer Science & Business Media, 2007. 8", + "[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLO by Ultralytics, 2023. 6", + "[28] Amin Jourabloo, Xi Yin, and Xiaoming Liu. Attribute preserved face de-identification. In International Conference on Biometrics, 2015. 2", + "[29] Alexander John Karran, Théophile Demazure, Antoine Hudon, Sylvain Senecal, and Pierre-Majorique Léger. Designing for confidence: The impact of visualizing artificial intelligence decisions. Frontiers in Neuroscience, 16: 883385, 2022. 2", + "[30] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The Kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2", + "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 6", + "[32] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal net" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18738", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works encode: Quantifying static vs. dynamic information. In CVPR, 2022. 8", + "[33] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2, 6", + "[34] Sudhakar Kumawat and Hajime Nagahara. Privacy-preserving action recognition via motion difference quantization. In ECCV, 2022. 1, 2, 5, 6, 7, 8", + "[35] Agostina J Larrazabal, Nicolas Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020. 1", + "[36] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. STPrivacy: Spatio-temporal privacy-preserving action recognition. In ICCV, 2023. 2", + "[37] Bo Liu, Ming Ding, Sina Shaham, Wenny Rahayu, Farhad Farokhi, and Zihuai Lin. When machine learning meets privacy: A survey and outlook. ACM Computing Surveys, 54 (2):1-36, 2021. 2", + "[38] Haochen Liu, Yiqi Wang, Wenqi Fan, Xiaorui Liu, Yaxin Li, Shaili Jain, Yunhao Liu, Anil Jain, and Jiliang Tang. Trustworthy AI: A computational perspective. ACM Transactions on Intelligent Systems and Technology, 14(1):1-59, 2022. 2", + "[39] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91–110, 2004. 2", + "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4", + "[41] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 6", + "[42] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action detection. In ECCV, 2018. 2", + "[43] Michael Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In AAAI, 2017. 2, 5, 6, 7", + "[44] Michael Ryoo, Kiyoon Kim, and Hyun Yang. Extreme low resolution activity recognition with multi-siamese embedding learning. In AAAI, 2018. 2, 6, 7, 8", + "[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 2", + "[46] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 1", + "[47] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a local SVM approach. In ICPR, 2004. 5, 7", + "[48] Vijeta Sharma, Manjari Gupta, Anil Pandey, Deepti Mishra, and Anil Kumar. A review of deep-learning-based human activity recognition on benchmark video datasets. Applied Artificial Intelligence, 36(1):2093705, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Jianbo Shi and Carlo Tomasi. Good features to track. In CVPR, 1994. 2", + "[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 2", + "[51] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 2", + "[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 7", + "[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2", + "[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 5", + "[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5", + "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3", + "[57] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local Neural Networks. In CVPR, 2018. 5", + "[58] Zhenyu Wu, Zhangyang Wang, Zhaowen Wang, and Hailin Jin. Towards privacy-preserving visual recognition via adversarial training: A pilot study. In ECCV, 2018. 1, 5, 6", + "[59] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE TPAMI, 44(4):2126-2139, 2020. 2, 5, 6, 7, 8", + "[60] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In CVPR, 2012. 3, 5, 7", + "[61] Yingying Zhu, Nandita M Nayak, and Amit K Roy-Chowdhury. Context-aware modeling and recognition of activities in video. In CVPR, 2013. 2" + ], + "bbox": [ + 501, + 92, + 890, + 632 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18739", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_model.json b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b6454887bfb34d72acd6834dff360ac733f40d1c --- /dev/null +++ b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.756, + 0.177 + ], + "angle": 0, + "content": "Selective, Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.204, + 0.286, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.204, + 0.495, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.204, + 0.659, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.204, + 0.86, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.283, + 0.303, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.279, + 0.61, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.275, + 0.891, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.479, + 0.893, + 0.563 + ], + "angle": 0, + "content": "Figure 1. Our goal is to hide privacy attributes without action recognition performance dropping. Left: Arbitrary images can be used to specify an interpretable template library defined by privacy attributes. Middle: A salience map is generated from privacy templates; example illustrates use of templates for personal identification. Right: The source video is masked with noise as guided by salience and animated by source video optical flow. Salience makes masking selective to privacy sensitive regions, while preserving scene context; optical flow preserves motion – both of which are critical for action recognition. The obfuscated video can be input directly to arbitrary privacy and action recognition systems without retraining. Zoomed circles highlight details only for illustration." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.576, + 0.314, + 0.59 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.597, + 0.471, + 0.854 + ], + "angle": 0, + "content": "Concerns for the privacy of individuals captured in public imagery have led to privacy-preserving action recognition. Existing approaches often suffer from issues arising through obfuscation being applied globally and a lack of interpretability. Global obfuscation hides privacy sensitive regions, but also contextual regions important for action recognition. Lack of interpretability erodes trust in these new technologies. We highlight the limitations of current paradigms and propose a solution: Human selected privacy templates that yield interpretability by design, an obfuscation scheme that selectively hides attributes and also induces temporal consistency, which is important in action recognition. Our approach is architecture agnostic and directly modifies input imagery, while existing approaches generally require architecture training. Our approach offers more flexibility, as no training is required, and outperforms alternatives on three widely used datasets." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.575, + 0.631, + 0.59 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Advances in state-of-the-art computer vision and machine learning enable deployment of such systems in the public sphere. Accompanying these initiatives, concerns arise for the privacy of individuals that are captured in acquired imagery [11, 19, 35, 46]. In particular, considerations arise regarding attributes that individuals want to keep confidential, yet that are revealed through visual information, even though they are not critical for the functioning of the deployed system, e.g. identity, age, gender and race. Video-based action recognition is an area of consideration as it has potential for widespread applications in surveillance and monitoring. These concerns have sparked interest in privacy preserving action recognition [13, 24, 34, 58]. These approaches process input imagery to obscure privacy attributes while maintaining action recognition performance. Contemporary approaches typically apply their obfuscation across entire input video frames and improvements have been made within this paradigm. Notably, however, there are downsides to this paradigm, as follows." + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.416, + 0.9 + ], + "angle": 0, + "content": "Code available f-iliic.github.io/SelectivePrivacyPreservation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18730" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "Collateral damage. Global masking strategies indiscriminately obscure the entire image, impacting regions within the scene that may exhibit high correlations with actions, albeit lack relevance to privacy. For example, it is known that masking objects and scene context can impair action recognition performance [61], yet these are lost in global masking. Furthermore, global masking strategies do not allow for selective attribute obfuscation and generally hide all attributes at once, even when not all are of concern." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.473, + 0.339 + ], + "angle": 0, + "content": "Loss of dynamic information. The large change in input modalities from global masking necessitates the retraining of the action recognition module or the design of custom modules (adversarial training), which adds to the challenge of practical deployment. Indeed, even if applied more locally, the obfuscation can compromise the motion of the actors, which also can be important in action recognition [50]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.432 + ], + "angle": 0, + "content": "Lack of interpretability. Finally, given that state-of-the-art approaches are end-to-end trained with limited concern for interpretability, the exact nature of what is being masked and how it is achieved can remain unclear. Lack of interpretability is an important concern in privacy preservation, because its lack can compromise user trust [29, 38]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.442, + 0.223, + 0.456 + ], + "angle": 0, + "content": "1.1. Contributions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.473, + 0.68 + ], + "angle": 0, + "content": "We present an approach to privacy preserving action recognition that responds directly to current limitations in four ways, as illustrated in Fig. 1. (i) The approach is based on local detection and selective obfuscation of privacy sensitive regions. This selectivity maintains global context information that is crucial to action recognition, yet unimportant for privacy. (ii) The local processing avoids large modality shifts in the imagery and is independent of the action recognition module itself; therefore, it does not require algorithm retraining, which is sometimes infeasible. (iii) The masking preserves interframe motion; so, that information is available for action recognition. (iv) The privacy sensitive masking is interpretable by design, and allows inspection through the explicitly generated saliency maps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.216, + 0.71 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Privacy in machine learning. The need to protect privacy has garnered increased attention in the vision research community. Current models commonly consume large amounts of web data to learn generalizable representations [20, 33, 53], which inevitably invade personal information, such as identity and location. Moreover, in model deployment it also may be desirable to preserve privacy information. The concern for privacy is not limited to vision research, but extends across artificial intelligence, including natural language processing [5] and more general machine learning [37]. As these technologies find their way into broader society, privacy concerns must be considered." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.26 + ], + "angle": 0, + "content": "Privacy preserving action recognition. We focus on video-based action recognition. Recent developments in this area have yielded systems capable of strong performance on challenging datasets; for review see, e.g. [48]. Similarly, applications, including those in privacy-sensitive scenarios (e.g. surveillance [10] and monitoring [51]), are being developed. To provide useful spatiotemporal signals for action recognition, video clips are mostly collected to capture actors with a great level of clarity throughout the actions, thus increasing the chance of privacy leakage. Moreover, such datasets are expanding rapidly, e.g. [4, 30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.895, + 0.655 + ], + "angle": 0, + "content": "In response to these concerns, research on recognizing actions while preserving private information has emerged in recent years. Early work concentrated on devising models that can work on low-resolution videos [8, 43, 44]; these methods often operate at the cost of sacrificing action recognition performance. Other work along these lines developed face-anonymization techniques to prevent models from yielding high accuracy on face recognition [28, 42]. Still, those approaches cannot easily extend to other attributes and are correspondingly limited. More recent work has focused on implicitly learning transformation functions to anonymize videos in a data-driven fashion [13, 24, 36, 59], and also extending such approaches to anomaly detection [18]. To reduce model complexity, a competing approach follows a simple, yet effective procedure [34]: Frame subtraction, followed by broadband-filtering to yield motion descriptors for action recognition, while suppressing privacy attributes. A notable limitation of that approach is the relative weakness of frame differences as motion descriptors compared to common alternatives, e.g. optical flow. Another common limitation to all existing work is the lack of flexibility to select arbitrary private information to hide, i.e. preserving privacy of only critical attributes, while maintaining the visibility of others, to avoid obfuscation of visual cues that are essential to action classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Template matching. Building correspondences between templates and target images via matching is a foundation for modern vision research, e.g. SfM [21], image correspondences [45], detection [12] and tracking [49]. Features used in matching have advanced rapidly from primarily handcrafted (e.g. [12, 39]) to learning-based convolutional [23] or transformers [14]. Our solution is inspired by techniques seen in feature-based template matching. We apply their insights on matching templates to localize and selectively obfuscate privacy-sensitive regions. More specifically, we adopt DINO-ViT features [6] and compute similarities using the keys of the last attention layer from that architecture. Our use of DINO-ViT keys (rather than queries or values) is based on previous work finding them to perform best when applied to visual correspondence [1, 40]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18731" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.316, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.202, + 0.234, + 0.213 + ], + "angle": 0, + "content": "(a) Overview" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.58, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.201, + 0.498, + 0.213 + ], + "angle": 0, + "content": "(b) Matcher" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.089, + 0.888, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.201, + 0.88, + 0.213 + ], + "angle": 0, + "content": "(c) Temporally Consistent Selective Obfuscation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Figure 2. Overview of Method. 2a We present a privacy module that builds atop three components: (i) a semantic template library that contains attributes to be hidden, (ii) a descriptor matcher to localize template features in videos to be obscured and, (iii) an obfuscation method that is sensitive with respect to motion present in the scene. 2b A semantic descriptor matcher based on DINO [6]-ViT [14] keys is used to determine privacy salient regions in a video based on the template library. In our case, regions of interest correspond to those that can identify a person; however, this component can be adapted for other privacy attributes through specification of different templates. The result is a saliency map. 2c The saliency map is used as a weight to apply noise to the regions. The noise, however, is not static, but is warped with optical flow with an initial noise pattern image, \\( N(\\mathbf{p},1) \\), for the purposes of preserving motion information in the source video. The similarity maps of all aggregated relevant privacy attributes are used to weigh the noise and apply it to the input image, obfuscating privacy sensitive information while not destroying the underlying temporal signal." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.266, + 0.394 + ], + "angle": 0, + "content": "3. Technical approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.471, + 0.613 + ], + "angle": 0, + "content": "Our approach to privacy preserving action recognition is a stand alone module that operates by preprocessing video that subsequently is input to arbitrary action and privacy recognition algorithms, i.e. it is independent of the recognition algorithms and does not entail any retraining of those algorithms. Our method consists of three key components; see Fig. 2a: (i) A template library that covers privacy attributes to be preserved, (ii) a matcher that produces saliency maps between selected templates and images where privacy is to be preserved and (iii) an obfuscator that uses the saliency maps to hide privacy attributes of concern in a temporally consistent fashion to preserve motion in the video. The remainder of this section details each of these components." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.245, + 0.638 + ], + "angle": 0, + "content": "3.1. Template library" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To obfuscate privacy sensitive regions in images in an interpretable fashion, we need an explicit set of “attributes” to be hidden from the target video. Such a concept template library, T, can be built manually choosing image patches corresponding to features one wishes to obfuscate. In this paper, we concentrate on privacy attributes related to personal identification. Therefore, we use landmark anatomical features, corresponding to detailed facial landmarks and the hand (Fig. 3 left: forehead, hair, eye, cheek, lips, hand) as well as larger body parts (Fig. 3 right: arm, torso, legs). The choice to focus on preservation of human identity is motivated by the fact that all three of our evaluation datasets define their privacy attributes on attributes relating to person identity. Notably, however, such a template library easily can be extended depending on the particular task at hand. For example, attributes pertaining to location (e.g. street signs, distinctive scene objects) could guide development" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.423 + ], + "angle": 0, + "content": "of a complimentary set of templates. In any case, given a template library, a user can selectively combine templates to obscure attributes of concern in a particular application." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.893, + 0.695 + ], + "angle": 0, + "content": "Formally, we define each template \\(\\tau_{i}\\) as an element in the set \\(\\mathsf{T} = \\{\\tau_1,\\tau_2,\\dots,\\tau_n\\}\\). Further, let \\(\\tilde{\\mathsf{T}}\\subseteq \\mathsf{T}\\) be the subset of templates that the user selects from the library, \\(\\mathsf{T}\\), for a particular subset of privacy attributes to be preserved for a given dataset or application. Our manual approach to template selection leads to concepts that are interpretable by design. In particular, we use two source images, shown in Fig. 3, one for small-scale features on the face, and one for large regions, taken from the IPN [3] and SBU [60] datasets. We choose these two datasets because of their complimentarity in template selection: IPN focuses on small scale features (e.g. facial and hand), while SBU focuses on larger scale features (e.g. larger body parts), as detailed later in Sec. 4.1. The choice of the particular template images is not critical to the functioning of our approach, because we extract semantic features from the templates that are known to generalize well across images instances [6], as described next." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.875, + 0.727 + ], + "angle": 0, + "content": "3.2. Matching: Local patch descriptor templates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Semantic features. Our requirements for good features to match between privacy templates in our library and frames in an action recognition video are straightforward: We require (i) semantically rich features that generalize across (in our case) different individuals and (ii) high spatial resolution to perform privacy obfuscation in a localized manner without destroying regions that are required for action recognition. We find that DINO-ViT features fit our needs very well [6, 56]: (i) They have been trained to yield high similarity for semantically related concepts (e.g. objects and their parts), while suppressing the similarity of unrelated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18732" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.093, + 0.307, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.092, + 0.465, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.295 + ], + "angle": 0, + "content": "Figure 3. Template Library consisting of Patches Chosen from Anatomical Landmark Regions. These specific images are passed through a DINO-ViT feature extractor. The keys, corresponding to spatial locations of the highlighted patches, are chosen as the templates for matching to input images to obtain semantically similar regions for obfuscation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.47, + 0.444 + ], + "angle": 0, + "content": "matters (e.g. background). (ii) They support local patch feature extraction over high resolution images without losing global context. Elsewhere, DINO-ViT keys have proven especially useful in matching between templates and target images [1, 40]. Following these advances, we use last attention layer DINO-ViT keys computed from privacy templates (e.g. Fig. 3) to match against input images to compute privacy saliency maps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.449, + 0.471, + 0.587 + ], + "angle": 0, + "content": "Privacy saliency matching. To find privacy salient regions in an input video, a similarity map, \\( S \\), is computed between each frame, \\( I \\), in the video and the selected set of privacy templates, \\( \\tilde{\\mathsf{T}} \\). The image is tiled with \\( m \\) patches, \\( I_{j}, j \\in \\{1,\\dots,m\\} \\), and for each patch DINO-ViT keys, \\( \\mathsf{K}(I_j) \\), are extracted. These features are matched with DINO-ViT keys, \\( \\mathsf{K}(\\tau_i) \\), for all selected templates, \\( \\tau_{i} \\in \\tilde{\\mathsf{T}} \\). Clipped cosine similarity is used to establish the saliency of each image patch according to" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.598, + 0.47, + 0.638 + ], + "angle": 0, + "content": "\\[\ns _ {j} = \\frac {1}{| \\tilde {\\mathbf {T}} |} \\sum_ {i = 1} ^ {| \\tilde {\\mathbf {T}} |} \\max \\left(0, \\frac {\\langle \\mathrm {K} (\\tau_ {i}) , \\mathrm {K} (I _ {j}) \\rangle}{\\| \\mathrm {K} (\\tau_ {i}) \\| \\| \\mathrm {K} (I _ {j}) \\|}\\right), \\forall j \\in [ [ 1, m ] ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.469, + 0.756 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot ,\\cdot \\rangle\\) is inner product and \\(|\\tilde{\\mathsf{T}} |\\) is the cardinality of \\(\\tilde{\\mathsf{T}}\\) the set of selected privacy templates. Clipping is used because only positive values imply salience. This calculation is performed for every patch in every image of the video. Subsequently, the \\(m\\) saliency patches are reassembled in the shape of the original image to produce the final saliency map," + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.757, + 0.469, + 0.773 + ], + "angle": 0, + "content": "\\[\nS = \\mathcal {R} \\left(s _ {1}, \\dots , s _ {m}; h, w\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "with \\(\\mathcal{R}\\) a function that accepts image tiles, \\(s_j\\), and reshapes them into their original image format of height \\(h\\) and width \\(w\\). Note that a separate salience map, \\(S\\), is calculated for each frame in a video of interest, \\(I\\). The entire process of feature extraction, matching and the resulting saliency maps are summarized in Fig. 2b. Example saliency maps in Fig. 4 illustrate the ability of our approach to capture all of our templates in a variety of scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.799, + 0.107 + ], + "angle": 0, + "content": "3.3. Temporally consistent obfuscation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Our goal is to mask privacy sensitive regions in images to obfuscate them. If we were to apply masks to every frame independently, then temporal information important to action recognition, e.g. motion of actors, would be destroyed as well. We empirically document the issue in Sec. 4.2. In response to this challenge, we follow previous work that presented a method for producing temporally consistent spatial noise across videos that supported action recognition, while obscuring appearance information in single frames [25]. While the previous work applied noise patterns uniformly across entire frames, we instead weight them by our privacy salience maps, \\( S \\), to preserve as much context information as possible. The remainder of this subsection details our approach, with an outline shown in Fig. 2c." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Noise pattern initialization. Let \\(\\mathbf{p} = (x, y)\\) be image coordinates and \\(t\\) time. We initialize a noise image, \\(N(\\mathbf{p}, 1)\\), with the same dimensions as a frame from the input video (i.e. \\(h \\times w \\times 3\\)), with \\(h\\) height, \\(w\\) width and 3 the number of colour channels. The dataset mean, \\(\\mu\\), and standard deviation, \\(\\sigma\\), are used to define a uniform distribution from which individual pixel intensities are drawn according to" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.446, + 0.892, + 0.463 + ], + "angle": 0, + "content": "\\[\nN (\\mathbf {p}, t = 1) \\sim \\mathcal {U} [ \\mu - \\sigma , \\mu + \\sigma ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.893, + 0.597 + ], + "angle": 0, + "content": "Motion consistent noise. To create motion consistent noise the initial random frame, \\( N(\\mathbf{p}, 1) \\), is warped forward with flow fields derived from the original video, \\( I(\\mathbf{p}, t) \\), as extracted by an optical flow algorithm. Let \\( \\mathbf{v}(\\mathbf{p}, t) = (u(\\mathbf{p}, t), v(\\mathbf{p}, t)) \\) be the flow field that maps points, \\( \\mathbf{p} \\), in frame \\( t \\) to those in frame \\( t - 1 \\), with \\( u \\) and \\( v \\) the horizontal and vertical components of the flow. Then, a motion consistent noise sequence is generated as" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.605, + 0.892, + 0.623 + ], + "angle": 0, + "content": "\\[\nN (\\mathbf {p}, t) = N \\left(\\mathbf {p} + \\mathbf {v} (\\mathbf {p}, t), 1\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.891, + 0.66 + ], + "angle": 0, + "content": "with \\( t \\in \\{2, \\dots, T\\} \\) and \\( T \\) the number of frames in the original input video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Selective privacy obfuscation. The computed video sequence, \\(N\\), shows no single frame appearance related to the original video as it is random noise; however, when viewed as a video it reveals the motion present in the original, cf. [25]. Direct use of this synthesized video obscures privacy attributes; however, it also obscures other context information that could be of use in action recognition. So, instead we selectively apply \\(N\\) to every frame \\(I\\) in the video by using the privacy salience maps \\(S\\), according to" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.81, + 0.892, + 0.835 + ], + "angle": 0, + "content": "\\[\nO (\\mathbf {p}, t) = I (\\mathbf {p}, t) + \\left(S (\\mathbf {p}, t) \\times \\left(N (\\mathbf {p}, t) - I (\\mathbf {p}, t)\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The resulting video, \\(O\\), contains selectively obfuscated regions, built with interpretable templates by design and contains motion information that (in principle) does not differ from the original input video." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18733" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.16, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.091, + 0.24, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.091, + 0.321, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.091, + 0.402, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.091, + 0.483, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.565, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.091, + 0.646, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.091, + 0.728, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.091, + 0.808, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.091, + 0.891, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.219, + 0.894, + 0.275 + ], + "angle": 0, + "content": "Figure 4. Saliency Maps for Descriptors in the Template Library. The manual selection of these templates allows for interpretability of the obfuscated parts of the image by design. The matched DINO-ViT features capture rich semantic information and allow for detailed spatial localization due to the nature of vision transformers. These saliency maps can then be combined for obfuscating any combination of templates depending on the task at hand." + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.298, + 0.404, + 0.309 + ], + "angle": 0, + "content": "ACTION RECOGNITION ON SOURCE DATASETS" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.31, + 0.47, + 0.634 + ], + "angle": 0, + "content": "
Networkf × rIPNKTHSBU
C2D [57]8 × 874.5683.0087.50
CSN [55]32 × 291.7388.3390.91
E2S X3D L [25]16 × 595.1595.3391.11
E2S X3D M [25]16 × 589.3892.3386.36
E2S X3D S [25]13 × 685.1692.0082.98
I3D [7]8 × 883.1589.6782.95
MVIT[15]16 × 488.0090.0092.55
R2+1D [54]16 × 489.8087.3381.91
Slow [17]8 × 885.7689.0088.64
SlowFast [17]32 × 288.0687.3390.00
X3D L [16]16 × 594.4194.0090.22
X3D M [16]16 × 591.6793.0081.91
X3D S [16]13 × 687.8190.6775.00
Average88.0590.1586.31
PRIVACY PRESERVATION FOR SOURCE DATASETS
NetworkIPNKTHSBU
ResNet18 [22]88.4687.3390.43
ResNet50 [22]92.3190.0096.81
ResNet101 [22]94.2394.0084.04
ViTb/16 [14]94.2394.0079.79
ViTb/32 [14]90.3894.0078.72
Average91.9291.8785.96
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Table 1. Top-1 Accuracy for all Privacy and Action models on the original unmodified (source) videos, i.e. without privacy obfuscation. Bold and underline indicate first and second best, resp. Number of frames and temporal sampling rate indicated as \\( f \\times r \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.279, + 0.74 + ], + "angle": 0, + "content": "4. Empirical evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The privacy obfuscated video, (5), serves directly as input to action and privacy recognition. No specialized development, training or other modification of the recognition algorithms nor adaptation of the privacy preservation system is necessary. Indeed, a major difference compared to competing state-of-the-art approaches (e.g. [34, 43, 59]) is that we do not retrain networks with our obfuscated data, while they do retrain. We exploit the fact that privacy attributes are generally independent from action recognition cues, which is enabled by our unique selective obfuscation approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.602, + 0.315 + ], + "angle": 0, + "content": "4.1. Protocol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Datasets. We use three datasets commonly used for investigating action recognition and privacy, IPN [3], SBU [60] and KTH [47], which pose different challenges concerning both action and privacy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.46 + ], + "angle": 0, + "content": "IPN is a large-scale video-based hand gesture recognition dataset that consists of 50 actors performing 13 static or dynamic gestures, against three different backgrounds [3]. The performed gestures are used as the action labels and actor genders are used as the privacy labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.535 + ], + "angle": 0, + "content": "SBU is a video dataset depicting eight human interactions. Each video is a video of actor-pairs, in the same laboratory environment. Action labels are derived from the actor interactions and the privacy attributes are the unique pairings of seven different actors resulting in 13 privacy labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.686 + ], + "angle": 0, + "content": "KTH has 25 actors doing one of six actions [47]. Each action is performed four times in different environments. The six action classes are used for action recognition and 25 actor identities serve as privacy labels. In the originally proposed splits, videos of any one actor are fully contained in one set, since the intent of the KTH dataset was not concerned with actor identity recognition. For privacy identity, each actor must appear in both sets; so, the data is split such that each action is performed twice by each person in the training set, and once in the validation and test sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Metrics. Privacy recognition results are obtained following standard practice [34], averaging outputs over multiple frames (32 for IPN and KTH, and 16 for SBU) from the same video. Action recognition results are obtained by reporting the top-1 accuracy, as per convention [24, 58, 59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.853 + ], + "angle": 0, + "content": "It also is interesting to gauge the trade-off between action and privacy recognition. Let \\( 0 \\leq a \\leq 1 \\) be normalized action recognition performance derived from dividing accuracy percentages by 100, and let \\( p \\) be defined analogously for privacy recognition. To quantify their trade-off we use a linear combination of \\( a \\) and \\( 1 - p \\) and define" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.862, + 0.891, + 0.878 + ], + "angle": 0, + "content": "\\[\nf _ {\\lambda} (a, p) = (1 - \\lambda) a + \\lambda (1 - p), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "with \\(0 \\leq \\lambda \\leq 1\\) weighing the relative importance of action" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18734" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.091, + 0.35, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.092, + 0.615, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.092, + 0.882, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.21, + 0.894, + 0.268 + ], + "angle": 0, + "content": "Figure 5. Obfuscation with a Single Attribute and the Impact on Performance. Attribute importance is dataset dependent. For example, notice how the 'Hand' template contributes to a large decrease in action recognition performance on IPN, as the action is determined solely by the hand, whereas on SBU it does not. Optimally blue is high and red is low. Corresponding qualitative examples of saliency maps for each individual template are shown in Fig. 4. Bold text along the abscissa of each plot indicates the templates used for the final results." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.293, + 0.177, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.294, + 0.271, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.294, + 0.365, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.294, + 0.46, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.368, + 0.178, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.368, + 0.272, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.368, + 0.369, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.368, + 0.465, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.438, + 0.182, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.511, + 0.159, + 0.524 + ], + "angle": 0, + "content": "ELR [44]" + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.438, + 0.276, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.511, + 0.257, + 0.524 + ], + "angle": 0, + "content": "BDQ [34]" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.438, + 0.37, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.511, + 0.352, + 0.524 + ], + "angle": 0, + "content": "ALF [59]" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.438, + 0.465, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.511, + 0.434, + 0.523 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Figure 6. Optical flow from two consecutive frames from competing approaches. Only our approach retains dynamic information that supports high-quality optical flow recovery. Top row: Single frames as processed by comparison algorithms. Middle row: Full frame optical flow recovery. Bottom row: Zoomed optical flow details. Optical flow shown in Middlebury colour coding [2]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.742 + ], + "angle": 0, + "content": "recognition vs. privacy. This metric is suitable to create a linear ranking as \\(a\\), and \\((1 - p)\\) are optimally 1 ensuring \\(f_{\\lambda} \\in [0,1]\\). If privacy or action recognition are not equally important, then they can be weighted accordingly; e.g. if privacy is critical, then \\(\\lambda\\) can be increased. In our main experiments we use \\(\\lambda = 0.5\\), abbreviated as \\(f_{0.5}\\) in Tab. 2 and show an ablation over different values of \\(\\lambda\\) in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.469, + 0.868 + ], + "angle": 0, + "content": "Competing obfuscation approaches. We briefly describe NAIVE BASELINE approaches that solely rely on full-frame pixelation and blurring as well as actor masking that have been studied across the literature [13, 58]. We also highlight other State-of-The-Art approaches (SOTA) [34, 43, 59] that we compare against. Visual examples of the SoTA approaches are shown in Fig. 6; also shown are optical flows of consecutive frames for qualitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Mask Obfuscation. All our datasets involve people performing actions and the privacy attributes are related to geo-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.369 + ], + "angle": 0, + "content": "ple. A naive way to preserve privacy in such videos is to completely mask out the actors. To do so, we use the YOLOv8 implementation [27] based on the original YOLO [41]. We report results based on the masked region filled with the mean intensity of the image covered by the mask." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.434 + ], + "angle": 0, + "content": "Pixelation. Pixelation applies average pooling over regions of dimensions \\( x \\times x \\) on the input image sequence. We choose two scales of \\( x \\in \\{4,16\\} \\) for the patch-size to pool. In the result tables these methods are abbreviated as \\( \\operatorname{Pix}_{x \\times x} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Blur. Blur applies a Gaussian blur to the images that have been rescaled to \\(224 \\times 224\\) pixels. The parameters of the Gaussian are the kernel size, \\(\\kappa\\), and the standard deviation of the kernel, \\(\\sigma\\). We choose values for weak and strong blurs, \\(\\kappa = 13\\) and \\(\\kappa = 21\\) respectively, as consistent with other work on obfuscation methods [59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.576 + ], + "angle": 0, + "content": "ELR initially reduces frames to Extreme Low Resolution and subsequently applies a set of learned inverse superresolution transforms to support action recognition [43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.61 + ], + "angle": 0, + "content": "ALF is an Adversarial Learning Framework for action recognition that takes into account a privacy budget [59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.674 + ], + "angle": 0, + "content": "BDQ is a privacy-preserving encoder that sequentially Blurs, Differences and Quantizes frames. The blur and quantization parameters are learned to maximize action recognition while minimizing privacy recognition [34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Notably, all the compared SoTA methods operate across entire frames (i.e. without selectivity) and have limited interpretability due to their learning-based obfuscation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.734, + 0.671, + 0.748 + ], + "angle": 0, + "content": "Implementation details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training. No large scale dataset exists that allows for joint training of action and privacy classification on the same videos. Therefore, we pretrain our action and privacy networks on Kinetics400 [7] and Imagenet1k [33], respectively, and then finetune for specific datasets. We use the AdamW [31] optimizer with a learning rate of \\( 3e^{-4} \\). The networks are trained with a patience scheme of 100 epochs that monitors the loss on the validation set. For action recognition, videos are temporally uniformly subsampled according to the architecture (Tab. 1, \\( f \\times r \\) column)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18735" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.293, + 0.088, + 0.679, + 0.101 + ], + "angle": 0, + "content": "RESULTS ACROSS NAIVE BASELINE OBFUSCATION METHODS" + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.103, + 0.895, + 0.372 + ], + "angle": 0, + "content": "
MethodIPNKTHSBU
↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ
Original88.0591.920.4890.1591.870.4986.3185.960.50
Masking36.4564.870.3637.5435.220.5152.6230.10.61
Pixelate4×485.5973.650.5683.6459.390.6273.2446.270.63
Pixelate16×1649.8165.760.4238.3338.430.5025.5329.840.48
Blur κ=13σ=10 (weak)76.2467.400.5444.5937.730.5372.7535.690.69
Blur κ=21σ=10 (strong)58.8065.920.4630.3831.840.4957.7734.210.62
Ours87.1151.930.68+0.1188.675.460.92+0.2986.7413.190.87+0.18
RESULTS ACROSS SOTA OBFUSCATION METHODS
BDQ [34]81.0059.000.6191.117.150.9284.0434.180.75
ALF [59]76.0065.000.5685.8919.270.8382.0048.000.67
ELR[44] s=1670.8264.320.5391.2288.860.5196.2782.970.57
ELR[44] s=3252.9663.290.4585.5782.560.5292.4264.890.64
ELR[44] s=6431.6362.700.3456.2158.350.4980.0543.610.68
Ours †85.2551.670.67+0.0689.444.310.93+0.0184.0411.700.86+0.11
Ours87.1151.930.68+0.0788.675.460.92±0.0086.7413.190.87+0.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.378, + 0.895, + 0.465 + ], + "angle": 0, + "content": "Table 2. Comparison between Ours vs. Naive Baseline (top) and SOTA (bottom) Approaches as Top-1 Accuracy for both Action and Privacy labels, as well as \\( f_{0.5} \\) Introduced in Sec. 4.1. We show Ours and Naive Baseline results averaged across all recognition algorithms presented in Tab. 1. Competing SOTA approaches only present results on one specific algorithm for action recognition (I3D [7]) and privacy attribute detection (ResNet50 [22]) as those approaches are network specific. To showcase a fair comparison we also present results with the same single recognition algorithms indicated as \"Ours †\". First and second best results indicated by bold and underline, respectively. Relative performance delta (\\( \\Delta \\)) is indicated in green (improvement) and orange (tie) between best and second best method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.471, + 0.517 + ], + "angle": 0, + "content": "Note that we never perform any training on videos obfuscated by our approach. To implement warping, (4), we use pretrained RAFT to extract optical flow [52]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.807 + ], + "angle": 0, + "content": "Privacy template selection. Our approach affords selective combination of a predefined set of privacy templates for a given dataset or application. For the experiments, we choose a subset of templates, \\(\\tilde{\\mathsf{T}}\\), from our complete identity preserving template library, \\(\\mathsf{T}\\), shown in Fig. 3, to optimize performance on a given dataset. Figure 5 shows template-wise performance based on both action recognition and privacy for each dataset. For SBU and KTH, action recognition is stable with respect to templates; however, the same four templates notably reduce privacy recognition. For consistency on IPN, we also select the four templates that most reduce privacy recognition, even while preserving action recognition; although, the distinction is less striking. This selection process results in templates \\(\\tilde{\\mathsf{T}} = \\{\\text{torso, arm, leg, hair}\\}\\) for SBU and KTH vs. \\(\\tilde{\\mathsf{T}} = \\{\\text{cheek, eyes, forehead, hair}\\}\\) for IPN. Our code yields more results on the overlap of individual saliency maps. Runtime. The constant overhead to produce our obfuscated videos is \\(\\approx 100ms/frame\\) on a NVIDIA RTX4080 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.172, + 0.832 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We evaluate our approach to privacy preserving action recognition on 13 different action recognition models and five different privacy attribute recognition models. Competing approaches evaluate on at most five action models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[34, 43, 59]. Table 1 shows results on the original videos from IPN [3], SBU [60] and KTH [47], i.e. without any privacy preserving processing. Table 2 shows results comparing our approach vs. NaIVE BASELINES and SOTA. Comparison is given as the average across recognition algorithms presented in Table 1. Notably, while our approach compromises at most \\(1.48\\%\\) action recognition accuracy compared to performance on the original video (88.67% vs. \\(90.15\\%\\) on KTH), it always greatly improves privacy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.893, + 0.853 + ], + "angle": 0, + "content": "Naive Baselines. Compared to the baselines, our approach scores highest with respect to the \\( f_{0.5} \\) metric across all three datasets: IPN (+0.11), KTH (+0.29), and SBU (+0.18). Naive BASELINES show that action recognition correlates highly with privacy performance, as \\( f_{0.5} \\) hovers around 0.5. The notable exception is \\( \\mathrm{Pix}_{4\\times 4} \\) that performs well on IPN and KTH, and is the second best Naive BASELINE on SBU. A notable benefit of all these approaches is that they do not require retraining of the recognition algorithms. Still, none have a selective ability to obfuscate only certain parts. State of the art. A noteworthy difference between our vs. the alternative SoTA approaches is that they only work with the recognition algorithms on which they were trained. While this fact disadvantages our approach, as it lacks such retraining, it still outperforms the competing methods on all datasets: IPN (+0.06), KTH (+0.01), and SBU (+0.11)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "It is valuable to consider obfuscation approaches in terms of the relative importance of action recognition vs. privacy preservation. Figure 7 compares results for the SoTA ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18736" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.465, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Figure 7. Performance across all datasets with varying \\(\\lambda\\); see (6). Values of \\(\\lambda\\) closer to 0 weigh action recognition higher, whereas values closer to 1 increase the importance of privacy preservation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.468, + 0.33 + ], + "angle": 0, + "content": "proaches as that trade-off is varied in terms of \\( f_{\\lambda} \\), (6). The sweep of \\( \\lambda \\) shows that our obfuscation approach performs better than any competing approach across the entire range. Especially with increasing \\( \\lambda \\), the gap to other approaches widens as more emphasis is put on privacy preservation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.393 + ], + "angle": 0, + "content": "Importance of selective obfuscation. Our approach is unique in its ability to selectively obfuscate particular regions within a video based on specific privacy attributes. This ability yields two benefits:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.469, + 0.471 + ], + "angle": 0, + "content": "(i) Selectivity aids in interpretability. Each selected template results in a saliency map, (2), which allows for visual inspection of what information is being obscured. Figure 4 highlights this benefit as the heat maps reveal the degree of obfuscation to be applied on a template-by-template basis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.744 + ], + "angle": 0, + "content": "(ii) Individual privacy templates can be chosen and combined for best performance; see Fig. 5. Depending on the dataset, different privacy templates differently impact the performance of privacy and action recognition. Action recognition on SBU and KTH is relatively robust to privacy template selection; however, privacy preservation is best when a subset of templates (torso, leg, arm, hair) is selected and the rest remain unobfuscated. This fact derives directly from the saliency calculation, (1): If multiple individual saliency maps have small values, then other strong responses are scaled down in the combined final saliency map. Subsequently, this effect leads to less noise being applied in our selective obfuscation, (5), which can deteriorate privacy preservation. In contrast, action recognition on IPN is compromised if the hand template is selected (as expected with a gesture centric dataset), which documents that simply obfuscating the entire person leads to inferior action performance compared to selectively applied obfuscation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Importance of motion consistent noise. Action recognition is complicated as different datasets, and even individual actions within a dataset might require different recognition capabilities and can rely to different degrees on the modeling of motion [9]. This uncertainty on the role of motion in action recognition is exacerbated by the fact that different deep learning-based architectures are successful in capturing motion to varying degrees [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To see the importance of our proposed motion consistent noise for privacy preservation while maintaining good" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.092, + 0.699, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.091, + 0.891, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.188, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Figure 8. Comparing Temporally Consistent vs. \\( iid \\) Obfuscation for Action Recognition and Privacy Preservation. Action recognition performance decreases if noise is not temporally consistent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.456 + ], + "angle": 0, + "content": "action recognition, we compare to another version of our pipeline that obfuscates video frames using independent, identically distributed (iid) noise; see Fig. 8 and project page for video results. For all datasets, action recognition performance decreases if the noise is \\( iid \\). This result is sensible, because the \\( iid \\) noise does not capture the temporal dynamics of the source video. In contrast, privacy preservation is robust to noise in either case. There is less impact on IPN action recognition, as the critical hand motion is never obfuscated, which underlines the importance of selective masking. Further insight on why motion consistent noise supports better action can be had through consideration of Fig. 6, where the ability of such noise to support optical flow estimation is illustrated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Limitations. Inevitably, there will be a trade-off between action recognition and privacy, as relevant information may be shared. In our approach, that trade-off could happen because our temporally consistent noise maintains dynamic information important for action recognition; however, it also might support motion based identification, e.g. gait recognition. Current protocols in privacy preserving action recognition do not consider motion-based identification. However, gait and related motion-based measurements are weak biometrics [26]; so, favouring action recognition may be apt. Moreover, if it becomes a concern, then it may be possible to apply motion perturbations that impede personal identification while maintaining action recognition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.619, + 0.681 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our work highlights that it is not necessary to train action recognition and privacy networks in an adversarial fashion for effective obfuscation of privacy attributes while maintaining strong action recognition performance. We show that a system based on local privacy templates, deep features that capture template semantics and selective noise obfuscation that is animated with source video motion can uphold privacy without hindering action recognition. Our approach is unique compared to alternative recent approaches in terms of interpretability and independence from particular action and privacy recognition algorithms. Our nine manually chosen templates, in combination with our proposed obfuscation technique outperforms other state-of-the-art approaches across three different datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18737" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. ECCVW, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Simon Baker, Daniel Scharstein, JP Lewis, Stefan Roth, Michael J Black, and Richard Szeliski. A database and evaluation methodology for optical flow. IJCV, 92:1-31, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. IPN Hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In ICPR, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[5] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.47, + 0.499 + ], + "angle": 0, + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[8] Jiawei Chen, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Semi-coupled two-stream fusion ConvNets for action recognition at extremely low resolutions. In WACV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[9] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't I dance in the mall? Learning to mitigate scene bias in action recognition. NeurIPS, 32, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[10] Robert T. Collins, Alan J Lipton, and Takeo Kanade. Introduction to the special section on video surveillance. IEEE TPAMI, 22(8):745-746, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[11] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (SMID): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[12] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. SPAct: Self-supervised privacy preservation for action recognition. In CVPR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer. X3D: Expanding architectures for efficient video recognition. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Joseph Fioresi, Ishan Rajendrakumar Dave, and Mubarak Shah. Ted-SPAD: Temporal distinctiveness for self-supervised privacy-preservation for video anomaly detection. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, and Omer Ovenc. Scalable detection of offensive and non-compliant content/logo in product images. In WACV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[21] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[24] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. PrivHAR: Recognizing human actions from privacy-preserving lens. In ECCV, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[25] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In ECCV, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[26] Anil K Jain, Patrick Flynn, and Arun A Ross. Handbook of Biometrics. Springer Science & Business Media, 2007. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLO by Ultralytics, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[28] Amin Jourabloo, Xi Yin, and Xiaoming Liu. Attribute preserved face de-identification. In International Conference on Biometrics, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[29] Alexander John Karran, Théophile Demazure, Antoine Hudon, Sylvain Senecal, and Pierre-Majorique Léger. Designing for confidence: The impact of visualizing artificial intelligence decisions. Frontiers in Neuroscience, 16: 883385, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[30] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The Kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal net" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18738" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "works encode: Quantifying static vs. dynamic information. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.162 + ], + "angle": 0, + "content": "[33] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[34] Sudhakar Kumawat and Hajime Nagahara. Privacy-preserving action recognition via motion difference quantization. In ECCV, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[35] Agostina J Larrazabal, Nicolas Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[36] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. STPrivacy: Spatio-temporal privacy-preserving action recognition. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[37] Bo Liu, Ming Ding, Sina Shaham, Wenny Rahayu, Farhad Farokhi, and Zihuai Lin. When machine learning meets privacy: A survey and outlook. ACM Computing Surveys, 54 (2):1-36, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[38] Haochen Liu, Yiqi Wang, Wenqi Fan, Xiaorui Liu, Yaxin Li, Shaili Jain, Yunhao Liu, Anil Jain, and Jiliang Tang. Trustworthy AI: A computational perspective. ACM Transactions on Intelligent Systems and Technology, 14(1):1-59, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[39] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91–110, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[41] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[42] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action detection. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[43] Michael Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In AAAI, 2017. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[44] Michael Ryoo, Kiyoon Kim, and Hyun Yang. Extreme low resolution activity recognition with multi-siamese embedding learning. In AAAI, 2018. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[46] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[47] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a local SVM approach. In ICPR, 2004. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[48] Vijeta Sharma, Manjari Gupta, Anil Pandey, Deepti Mishra, and Anil Kumar. A review of deep-learning-based human activity recognition on benchmark video datasets. Applied Artificial Intelligence, 36(1):2093705, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[49] Jianbo Shi and Carlo Tomasi. Good features to track. In CVPR, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[51] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.435 + ], + "angle": 0, + "content": "[57] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local Neural Networks. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.436, + 0.892, + 0.478 + ], + "angle": 0, + "content": "[58] Zhenyu Wu, Zhangyang Wang, Zhaowen Wang, and Hailin Jin. Towards privacy-preserving visual recognition via adversarial training: A pilot study. In ECCV, 2018. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.479, + 0.892, + 0.534 + ], + "angle": 0, + "content": "[59] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE TPAMI, 44(4):2126-2139, 2020. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.59 + ], + "angle": 0, + "content": "[60] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In CVPR, 2012. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[61] Yingying Zhu, Nandita M Nayak, and Amit K Roy-Chowdhury. Context-aware modeling and recognition of activities in video. In CVPR, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18739" + } + ] +] \ No newline at end of file diff --git a/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_origin.pdf b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..44b2e4db8b38f11959bfa73cbd5996e7db98664a --- /dev/null +++ b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/a1bdfbcc-dfb6-44e6-9a88-6216e9fd5ee6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87e0cb0a0bc7d031ba39de0198c5fd5fe643e1d616257ac35088b1a95240a886 +size 7440712 diff --git a/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/full.md b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..641bba8174cedb19bb3c2d23796810ee004e7ee1 --- /dev/null +++ b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/full.md @@ -0,0 +1,344 @@ +# Selective, Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition + +![](images/20a30826185149aac9aaa823e6ef0fd01965888a5509f7faf4339f067a10a313.jpg) + +![](images/e85fb2a714e6b94a3065b4c7169f115b57fed2cfa9d75a5ce179716998ca2446.jpg) + +![](images/3aa1d670ae2438f54a386a119664bcd81c22e48eba197fa1b1ba9df0146d9c07.jpg) + +![](images/9419fbb9a64b9dd1e6a203d4f61065c9aedee8d92065285a64e280b44bde4c12.jpg) + +![](images/f6b16d27e3b4ab0312ed64869d755ee383bcc60a8243d1809e6f07d85fd17fad.jpg) +Figure 1. Our goal is to hide privacy attributes without action recognition performance dropping. Left: Arbitrary images can be used to specify an interpretable template library defined by privacy attributes. Middle: A salience map is generated from privacy templates; example illustrates use of templates for personal identification. Right: The source video is masked with noise as guided by salience and animated by source video optical flow. Salience makes masking selective to privacy sensitive regions, while preserving scene context; optical flow preserves motion – both of which are critical for action recognition. The obfuscated video can be input directly to arbitrary privacy and action recognition systems without retraining. Zoomed circles highlight details only for illustration. + +![](images/da0e3093f1ba3ca6d02de67525ec1e1a61a791809e036dd5497d1969b9082244.jpg) + +![](images/cccfc1f37d98162a2a7cdc77b62521acbed9ef08f9b8f4198d105ce83cd6f84f.jpg) + +# Abstract + +Concerns for the privacy of individuals captured in public imagery have led to privacy-preserving action recognition. Existing approaches often suffer from issues arising through obfuscation being applied globally and a lack of interpretability. Global obfuscation hides privacy sensitive regions, but also contextual regions important for action recognition. Lack of interpretability erodes trust in these new technologies. We highlight the limitations of current paradigms and propose a solution: Human selected privacy templates that yield interpretability by design, an obfuscation scheme that selectively hides attributes and also induces temporal consistency, which is important in action recognition. Our approach is architecture agnostic and directly modifies input imagery, while existing approaches generally require architecture training. Our approach offers more flexibility, as no training is required, and outperforms alternatives on three widely used datasets. + +# 1. Introduction + +Advances in state-of-the-art computer vision and machine learning enable deployment of such systems in the public sphere. Accompanying these initiatives, concerns arise for the privacy of individuals that are captured in acquired imagery [11, 19, 35, 46]. In particular, considerations arise regarding attributes that individuals want to keep confidential, yet that are revealed through visual information, even though they are not critical for the functioning of the deployed system, e.g. identity, age, gender and race. Video-based action recognition is an area of consideration as it has potential for widespread applications in surveillance and monitoring. These concerns have sparked interest in privacy preserving action recognition [13, 24, 34, 58]. These approaches process input imagery to obscure privacy attributes while maintaining action recognition performance. Contemporary approaches typically apply their obfuscation across entire input video frames and improvements have been made within this paradigm. Notably, however, there are downsides to this paradigm, as follows. + +Collateral damage. Global masking strategies indiscriminately obscure the entire image, impacting regions within the scene that may exhibit high correlations with actions, albeit lack relevance to privacy. For example, it is known that masking objects and scene context can impair action recognition performance [61], yet these are lost in global masking. Furthermore, global masking strategies do not allow for selective attribute obfuscation and generally hide all attributes at once, even when not all are of concern. + +Loss of dynamic information. The large change in input modalities from global masking necessitates the retraining of the action recognition module or the design of custom modules (adversarial training), which adds to the challenge of practical deployment. Indeed, even if applied more locally, the obfuscation can compromise the motion of the actors, which also can be important in action recognition [50]. + +Lack of interpretability. Finally, given that state-of-the-art approaches are end-to-end trained with limited concern for interpretability, the exact nature of what is being masked and how it is achieved can remain unclear. Lack of interpretability is an important concern in privacy preservation, because its lack can compromise user trust [29, 38]. + +# 1.1. Contributions + +We present an approach to privacy preserving action recognition that responds directly to current limitations in four ways, as illustrated in Fig. 1. (i) The approach is based on local detection and selective obfuscation of privacy sensitive regions. This selectivity maintains global context information that is crucial to action recognition, yet unimportant for privacy. (ii) The local processing avoids large modality shifts in the imagery and is independent of the action recognition module itself; therefore, it does not require algorithm retraining, which is sometimes infeasible. (iii) The masking preserves interframe motion; so, that information is available for action recognition. (iv) The privacy sensitive masking is interpretable by design, and allows inspection through the explicitly generated saliency maps. + +# 2. Related work + +Privacy in machine learning. The need to protect privacy has garnered increased attention in the vision research community. Current models commonly consume large amounts of web data to learn generalizable representations [20, 33, 53], which inevitably invade personal information, such as identity and location. Moreover, in model deployment it also may be desirable to preserve privacy information. The concern for privacy is not limited to vision research, but extends across artificial intelligence, including natural language processing [5] and more general machine learning [37]. As these technologies find their way into broader society, privacy concerns must be considered. + +Privacy preserving action recognition. We focus on video-based action recognition. Recent developments in this area have yielded systems capable of strong performance on challenging datasets; for review see, e.g. [48]. Similarly, applications, including those in privacy-sensitive scenarios (e.g. surveillance [10] and monitoring [51]), are being developed. To provide useful spatiotemporal signals for action recognition, video clips are mostly collected to capture actors with a great level of clarity throughout the actions, thus increasing the chance of privacy leakage. Moreover, such datasets are expanding rapidly, e.g. [4, 30]. + +In response to these concerns, research on recognizing actions while preserving private information has emerged in recent years. Early work concentrated on devising models that can work on low-resolution videos [8, 43, 44]; these methods often operate at the cost of sacrificing action recognition performance. Other work along these lines developed face-anonymization techniques to prevent models from yielding high accuracy on face recognition [28, 42]. Still, those approaches cannot easily extend to other attributes and are correspondingly limited. More recent work has focused on implicitly learning transformation functions to anonymize videos in a data-driven fashion [13, 24, 36, 59], and also extending such approaches to anomaly detection [18]. To reduce model complexity, a competing approach follows a simple, yet effective procedure [34]: Frame subtraction, followed by broadband-filtering to yield motion descriptors for action recognition, while suppressing privacy attributes. A notable limitation of that approach is the relative weakness of frame differences as motion descriptors compared to common alternatives, e.g. optical flow. Another common limitation to all existing work is the lack of flexibility to select arbitrary private information to hide, i.e. preserving privacy of only critical attributes, while maintaining the visibility of others, to avoid obfuscation of visual cues that are essential to action classification. + +Template matching. Building correspondences between templates and target images via matching is a foundation for modern vision research, e.g. SfM [21], image correspondences [45], detection [12] and tracking [49]. Features used in matching have advanced rapidly from primarily handcrafted (e.g. [12, 39]) to learning-based convolutional [23] or transformers [14]. Our solution is inspired by techniques seen in feature-based template matching. We apply their insights on matching templates to localize and selectively obfuscate privacy-sensitive regions. More specifically, we adopt DINO-ViT features [6] and compute similarities using the keys of the last attention layer from that architecture. Our use of DINO-ViT keys (rather than queries or values) is based on previous work finding them to perform best when applied to visual correspondence [1, 40]. + +![](images/cd80802ccf9e38d4ba8b524ce926a7f393454301b3166b2d902f36c9109cbf2a.jpg) +(a) Overview + +![](images/0001021df2ac512b719a1903e53a751cdc84e614b311782cdc4f2f3e49898646.jpg) +(b) Matcher +Figure 2. Overview of Method. 2a We present a privacy module that builds atop three components: (i) a semantic template library that contains attributes to be hidden, (ii) a descriptor matcher to localize template features in videos to be obscured and, (iii) an obfuscation method that is sensitive with respect to motion present in the scene. 2b A semantic descriptor matcher based on DINO [6]-ViT [14] keys is used to determine privacy salient regions in a video based on the template library. In our case, regions of interest correspond to those that can identify a person; however, this component can be adapted for other privacy attributes through specification of different templates. The result is a saliency map. 2c The saliency map is used as a weight to apply noise to the regions. The noise, however, is not static, but is warped with optical flow with an initial noise pattern image, $N(\mathbf{p},1)$ , for the purposes of preserving motion information in the source video. The similarity maps of all aggregated relevant privacy attributes are used to weigh the noise and apply it to the input image, obfuscating privacy sensitive information while not destroying the underlying temporal signal. + +![](images/bb360d16467fe008ad852374cb7588d352e9783feadd1efb0009344f4f7e91ed.jpg) +(c) Temporally Consistent Selective Obfuscation + +# 3. Technical approach + +Our approach to privacy preserving action recognition is a stand alone module that operates by preprocessing video that subsequently is input to arbitrary action and privacy recognition algorithms, i.e. it is independent of the recognition algorithms and does not entail any retraining of those algorithms. Our method consists of three key components; see Fig. 2a: (i) A template library that covers privacy attributes to be preserved, (ii) a matcher that produces saliency maps between selected templates and images where privacy is to be preserved and (iii) an obfuscator that uses the saliency maps to hide privacy attributes of concern in a temporally consistent fashion to preserve motion in the video. The remainder of this section details each of these components. + +# 3.1. Template library + +To obfuscate privacy sensitive regions in images in an interpretable fashion, we need an explicit set of “attributes” to be hidden from the target video. Such a concept template library, T, can be built manually choosing image patches corresponding to features one wishes to obfuscate. In this paper, we concentrate on privacy attributes related to personal identification. Therefore, we use landmark anatomical features, corresponding to detailed facial landmarks and the hand (Fig. 3 left: forehead, hair, eye, cheek, lips, hand) as well as larger body parts (Fig. 3 right: arm, torso, legs). The choice to focus on preservation of human identity is motivated by the fact that all three of our evaluation datasets define their privacy attributes on attributes relating to person identity. Notably, however, such a template library easily can be extended depending on the particular task at hand. For example, attributes pertaining to location (e.g. street signs, distinctive scene objects) could guide development + +of a complimentary set of templates. In any case, given a template library, a user can selectively combine templates to obscure attributes of concern in a particular application. + +Formally, we define each template $\tau_{i}$ as an element in the set $\mathsf{T} = \{\tau_1,\tau_2,\dots,\tau_n\}$ . Further, let $\tilde{\mathsf{T}}\subseteq \mathsf{T}$ be the subset of templates that the user selects from the library, $\mathsf{T}$ , for a particular subset of privacy attributes to be preserved for a given dataset or application. Our manual approach to template selection leads to concepts that are interpretable by design. In particular, we use two source images, shown in Fig. 3, one for small-scale features on the face, and one for large regions, taken from the IPN [3] and SBU [60] datasets. We choose these two datasets because of their complimentarity in template selection: IPN focuses on small scale features (e.g. facial and hand), while SBU focuses on larger scale features (e.g. larger body parts), as detailed later in Sec. 4.1. The choice of the particular template images is not critical to the functioning of our approach, because we extract semantic features from the templates that are known to generalize well across images instances [6], as described next. + +# 3.2. Matching: Local patch descriptor templates + +Semantic features. Our requirements for good features to match between privacy templates in our library and frames in an action recognition video are straightforward: We require (i) semantically rich features that generalize across (in our case) different individuals and (ii) high spatial resolution to perform privacy obfuscation in a localized manner without destroying regions that are required for action recognition. We find that DINO-ViT features fit our needs very well [6, 56]: (i) They have been trained to yield high similarity for semantically related concepts (e.g. objects and their parts), while suppressing the similarity of unrelated + +![](images/87f98b68be415c66c55c4db4b8056043811dc2e38745915d0d25612367d5c7dc.jpg) +Figure 3. Template Library consisting of Patches Chosen from Anatomical Landmark Regions. These specific images are passed through a DINO-ViT feature extractor. The keys, corresponding to spatial locations of the highlighted patches, are chosen as the templates for matching to input images to obtain semantically similar regions for obfuscation. + +![](images/ab1be94bbfe9748f517c48530b3a0601a0b534c117fedcf324a51e92309b390b.jpg) + +matters (e.g. background). (ii) They support local patch feature extraction over high resolution images without losing global context. Elsewhere, DINO-ViT keys have proven especially useful in matching between templates and target images [1, 40]. Following these advances, we use last attention layer DINO-ViT keys computed from privacy templates (e.g. Fig. 3) to match against input images to compute privacy saliency maps. + +Privacy saliency matching. To find privacy salient regions in an input video, a similarity map, $S$ , is computed between each frame, $I$ , in the video and the selected set of privacy templates, $\tilde{\mathsf{T}}$ . The image is tiled with $m$ patches, $I_{j}, j \in \{1,\dots,m\}$ , and for each patch DINO-ViT keys, $\mathsf{K}(I_j)$ , are extracted. These features are matched with DINO-ViT keys, $\mathsf{K}(\tau_i)$ , for all selected templates, $\tau_{i} \in \tilde{\mathsf{T}}$ . Clipped cosine similarity is used to establish the saliency of each image patch according to + +$$ +s _ {j} = \frac {1}{| \tilde {\mathbf {T}} |} \sum_ {i = 1} ^ {| \tilde {\mathbf {T}} |} \max \left(0, \frac {\langle \mathrm {K} (\tau_ {i}) , \mathrm {K} (I _ {j}) \rangle}{\| \mathrm {K} (\tau_ {i}) \| \| \mathrm {K} (I _ {j}) \|}\right), \forall j \in [ [ 1, m ] ], \tag {1} +$$ + +where $\langle \cdot ,\cdot \rangle$ is inner product and $|\tilde{\mathsf{T}} |$ is the cardinality of $\tilde{\mathsf{T}}$ the set of selected privacy templates. Clipping is used because only positive values imply salience. This calculation is performed for every patch in every image of the video. Subsequently, the $m$ saliency patches are reassembled in the shape of the original image to produce the final saliency map, + +$$ +S = \mathcal {R} \left(s _ {1}, \dots , s _ {m}; h, w\right), \tag {2} +$$ + +with $\mathcal{R}$ a function that accepts image tiles, $s_j$ , and reshapes them into their original image format of height $h$ and width $w$ . Note that a separate salience map, $S$ , is calculated for each frame in a video of interest, $I$ . The entire process of feature extraction, matching and the resulting saliency maps are summarized in Fig. 2b. Example saliency maps in Fig. 4 illustrate the ability of our approach to capture all of our templates in a variety of scenarios. + +# 3.3. Temporally consistent obfuscation + +Our goal is to mask privacy sensitive regions in images to obfuscate them. If we were to apply masks to every frame independently, then temporal information important to action recognition, e.g. motion of actors, would be destroyed as well. We empirically document the issue in Sec. 4.2. In response to this challenge, we follow previous work that presented a method for producing temporally consistent spatial noise across videos that supported action recognition, while obscuring appearance information in single frames [25]. While the previous work applied noise patterns uniformly across entire frames, we instead weight them by our privacy salience maps, $S$ , to preserve as much context information as possible. The remainder of this subsection details our approach, with an outline shown in Fig. 2c. + +Noise pattern initialization. Let $\mathbf{p} = (x, y)$ be image coordinates and $t$ time. We initialize a noise image, $N(\mathbf{p}, 1)$ , with the same dimensions as a frame from the input video (i.e. $h \times w \times 3$ ), with $h$ height, $w$ width and 3 the number of colour channels. The dataset mean, $\mu$ , and standard deviation, $\sigma$ , are used to define a uniform distribution from which individual pixel intensities are drawn according to + +$$ +N (\mathbf {p}, t = 1) \sim \mathcal {U} [ \mu - \sigma , \mu + \sigma ]. \tag {3} +$$ + +Motion consistent noise. To create motion consistent noise the initial random frame, $N(\mathbf{p}, 1)$ , is warped forward with flow fields derived from the original video, $I(\mathbf{p}, t)$ , as extracted by an optical flow algorithm. Let $\mathbf{v}(\mathbf{p}, t) = (u(\mathbf{p}, t), v(\mathbf{p}, t))$ be the flow field that maps points, $\mathbf{p}$ , in frame $t$ to those in frame $t - 1$ , with $u$ and $v$ the horizontal and vertical components of the flow. Then, a motion consistent noise sequence is generated as + +$$ +N (\mathbf {p}, t) = N \left(\mathbf {p} + \mathbf {v} (\mathbf {p}, t), 1\right) \tag {4} +$$ + +with $t \in \{2, \dots, T\}$ and $T$ the number of frames in the original input video. + +Selective privacy obfuscation. The computed video sequence, $N$ , shows no single frame appearance related to the original video as it is random noise; however, when viewed as a video it reveals the motion present in the original, cf. [25]. Direct use of this synthesized video obscures privacy attributes; however, it also obscures other context information that could be of use in action recognition. So, instead we selectively apply $N$ to every frame $I$ in the video by using the privacy salience maps $S$ , according to + +$$ +O (\mathbf {p}, t) = I (\mathbf {p}, t) + \left(S (\mathbf {p}, t) \times \left(N (\mathbf {p}, t) - I (\mathbf {p}, t)\right)\right). \tag {5} +$$ + +The resulting video, $O$ , contains selectively obfuscated regions, built with interpretable templates by design and contains motion information that (in principle) does not differ from the original input video. + +![](images/393cb808ada3384756f3cfcb6fb3848694468154b8f7f136eef732f817a541c3.jpg) +Figure 4. Saliency Maps for Descriptors in the Template Library. The manual selection of these templates allows for interpretability of the obfuscated parts of the image by design. The matched DINO-ViT features capture rich semantic information and allow for detailed spatial localization due to the nature of vision transformers. These saliency maps can then be combined for obfuscating any combination of templates depending on the task at hand. + +![](images/385e2b1c70afa594723da8bf34675e4199f1c012ba5660100d32790efc41a6a7.jpg) + +![](images/96aa1c9561d7f95c812a34214e2bd196b1070a08eb06e63d7d9cbb1760b8196e.jpg) + +![](images/eca47d4f1a9f4c1a1a330221b5ea3742f3053d9ab9f74854c2a5c5c3e9cc4c5b.jpg) + +![](images/67661313effcfe1e50450dac036f4e13da91324ffd9a8d422d10b4b59d3707a7.jpg) + +![](images/0138c17a11f2f31975363bcfbec31f46a84aafd6abdf2012011f8be55ffd016e.jpg) + +![](images/48454e4984540219b0e5cbf8c31af116f1e1bb4717811c005447c82d6f201803.jpg) + +![](images/bfc3bb592be2fd5d95212638237843e46953e8c3604e8ce16bd5825ff95dc22a.jpg) + +![](images/408507cccd7eaef775fa8935ba63295d1e3aa4ad5bf3a3fbf16819f28bc0f5bb.jpg) + +![](images/ca77c0593b1eee06660f5b8134f754cb13fa0e2c128961af5722c09a1de0b412.jpg) + +ACTION RECOGNITION ON SOURCE DATASETS + +
Networkf × rIPNKTHSBU
C2D [57]8 × 874.5683.0087.50
CSN [55]32 × 291.7388.3390.91
E2S X3D L [25]16 × 595.1595.3391.11
E2S X3D M [25]16 × 589.3892.3386.36
E2S X3D S [25]13 × 685.1692.0082.98
I3D [7]8 × 883.1589.6782.95
MVIT[15]16 × 488.0090.0092.55
R2+1D [54]16 × 489.8087.3381.91
Slow [17]8 × 885.7689.0088.64
SlowFast [17]32 × 288.0687.3390.00
X3D L [16]16 × 594.4194.0090.22
X3D M [16]16 × 591.6793.0081.91
X3D S [16]13 × 687.8190.6775.00
Average88.0590.1586.31
PRIVACY PRESERVATION FOR SOURCE DATASETS
NetworkIPNKTHSBU
ResNet18 [22]88.4687.3390.43
ResNet50 [22]92.3190.0096.81
ResNet101 [22]94.2394.0084.04
ViTb/16 [14]94.2394.0079.79
ViTb/32 [14]90.3894.0078.72
Average91.9291.8785.96
+ +Table 1. Top-1 Accuracy for all Privacy and Action models on the original unmodified (source) videos, i.e. without privacy obfuscation. Bold and underline indicate first and second best, resp. Number of frames and temporal sampling rate indicated as $f \times r$ . + +# 4. Empirical evaluation + +The privacy obfuscated video, (5), serves directly as input to action and privacy recognition. No specialized development, training or other modification of the recognition algorithms nor adaptation of the privacy preservation system is necessary. Indeed, a major difference compared to competing state-of-the-art approaches (e.g. [34, 43, 59]) is that we do not retrain networks with our obfuscated data, while they do retrain. We exploit the fact that privacy attributes are generally independent from action recognition cues, which is enabled by our unique selective obfuscation approach. + +# 4.1. Protocol + +Datasets. We use three datasets commonly used for investigating action recognition and privacy, IPN [3], SBU [60] and KTH [47], which pose different challenges concerning both action and privacy. + +IPN is a large-scale video-based hand gesture recognition dataset that consists of 50 actors performing 13 static or dynamic gestures, against three different backgrounds [3]. The performed gestures are used as the action labels and actor genders are used as the privacy labels. + +SBU is a video dataset depicting eight human interactions. Each video is a video of actor-pairs, in the same laboratory environment. Action labels are derived from the actor interactions and the privacy attributes are the unique pairings of seven different actors resulting in 13 privacy labels. + +KTH has 25 actors doing one of six actions [47]. Each action is performed four times in different environments. The six action classes are used for action recognition and 25 actor identities serve as privacy labels. In the originally proposed splits, videos of any one actor are fully contained in one set, since the intent of the KTH dataset was not concerned with actor identity recognition. For privacy identity, each actor must appear in both sets; so, the data is split such that each action is performed twice by each person in the training set, and once in the validation and test sets. + +Metrics. Privacy recognition results are obtained following standard practice [34], averaging outputs over multiple frames (32 for IPN and KTH, and 16 for SBU) from the same video. Action recognition results are obtained by reporting the top-1 accuracy, as per convention [24, 58, 59]. + +It also is interesting to gauge the trade-off between action and privacy recognition. Let $0 \leq a \leq 1$ be normalized action recognition performance derived from dividing accuracy percentages by 100, and let $p$ be defined analogously for privacy recognition. To quantify their trade-off we use a linear combination of $a$ and $1 - p$ and define + +$$ +f _ {\lambda} (a, p) = (1 - \lambda) a + \lambda (1 - p), \tag {6} +$$ + +with $0 \leq \lambda \leq 1$ weighing the relative importance of action + +![](images/c32ab3289d0c3ea9673aedc497605898b3b92c621699edd203f27d69fcc4f83a.jpg) +Figure 5. Obfuscation with a Single Attribute and the Impact on Performance. Attribute importance is dataset dependent. For example, notice how the 'Hand' template contributes to a large decrease in action recognition performance on IPN, as the action is determined solely by the hand, whereas on SBU it does not. Optimally blue is high and red is low. Corresponding qualitative examples of saliency maps for each individual template are shown in Fig. 4. Bold text along the abscissa of each plot indicates the templates used for the final results. + +![](images/48b07ac760a5cea5141dff92858a6fcba3687aceb3a0e4aa908559c068c88661.jpg) + +![](images/d6e6b41fce5d496f4afc9d92b0f3fbf2c2b503e078f288bf163871a499117817.jpg) + +![](images/f86584d0f0a14883dc713d8206978360ea165299a85c2039fa0bef242006685b.jpg) + +![](images/1888d6682e9aad01a779860a77a1508c62aad6d73f6632d7b9f4717bb894e604.jpg) + +![](images/e8b8eb8fdee01620a612a400d88df91adaf82b28df088ef324b01e83f3dcfbfc.jpg) + +![](images/388c0d6c7f41982a1f29deb40f1da619ba0c6fb13cdd949c6543bcb893ab4abd.jpg) + +![](images/da0848338641762a6c711f60f856772bbee510eb4619b6928a565b5bb9e71722.jpg) + +![](images/d3c98780a73bd5540276e4e877f578689a983daae423c64252319eacdaf0f723.jpg) + +![](images/9f6802e627f6b6aacca0c7b0fe71f31281d73b93fd33a28e28870e21bb41eeef.jpg) + +![](images/99bff6933cd376107dff4c5d4515f11b049fc0dee9194a377feb64c7054fd2b1.jpg) + +![](images/e0ffec1a07cc18dfc1d3bc23e7fcab63829511580fbcde41c9db5c037cf18442.jpg) +ELR [44] + +![](images/d3f411b3776886002ea48644423e1c1c1614f19d08325bb176802ef0ed9b0778.jpg) +BDQ [34] + +![](images/a3f598f0afd372a17bbdb3c97a431af731c749233b381dfe23c3c6ebb6fba902.jpg) +ALF [59] +Figure 6. Optical flow from two consecutive frames from competing approaches. Only our approach retains dynamic information that supports high-quality optical flow recovery. Top row: Single frames as processed by comparison algorithms. Middle row: Full frame optical flow recovery. Bottom row: Zoomed optical flow details. Optical flow shown in Middlebury colour coding [2]. + +![](images/29c4ded6542c9c0f4a50bc61e275dbfe845beef23912b749877a2bfdc4ee4ce6.jpg) +Ours + +recognition vs. privacy. This metric is suitable to create a linear ranking as $a$ , and $(1 - p)$ are optimally 1 ensuring $f_{\lambda} \in [0,1]$ . If privacy or action recognition are not equally important, then they can be weighted accordingly; e.g. if privacy is critical, then $\lambda$ can be increased. In our main experiments we use $\lambda = 0.5$ , abbreviated as $f_{0.5}$ in Tab. 2 and show an ablation over different values of $\lambda$ in Fig. 7. + +Competing obfuscation approaches. We briefly describe NAIVE BASELINE approaches that solely rely on full-frame pixelation and blurring as well as actor masking that have been studied across the literature [13, 58]. We also highlight other State-of-The-Art approaches (SOTA) [34, 43, 59] that we compare against. Visual examples of the SoTA approaches are shown in Fig. 6; also shown are optical flows of consecutive frames for qualitative comparison. + +Mask Obfuscation. All our datasets involve people performing actions and the privacy attributes are related to geo- + +ple. A naive way to preserve privacy in such videos is to completely mask out the actors. To do so, we use the YOLOv8 implementation [27] based on the original YOLO [41]. We report results based on the masked region filled with the mean intensity of the image covered by the mask. + +Pixelation. Pixelation applies average pooling over regions of dimensions $x \times x$ on the input image sequence. We choose two scales of $x \in \{4,16\}$ for the patch-size to pool. In the result tables these methods are abbreviated as $\operatorname{Pix}_{x \times x}$ . + +Blur. Blur applies a Gaussian blur to the images that have been rescaled to $224 \times 224$ pixels. The parameters of the Gaussian are the kernel size, $\kappa$ , and the standard deviation of the kernel, $\sigma$ . We choose values for weak and strong blurs, $\kappa = 13$ and $\kappa = 21$ respectively, as consistent with other work on obfuscation methods [59]. + +ELR initially reduces frames to Extreme Low Resolution and subsequently applies a set of learned inverse superresolution transforms to support action recognition [43]. + +ALF is an Adversarial Learning Framework for action recognition that takes into account a privacy budget [59]. + +BDQ is a privacy-preserving encoder that sequentially Blurs, Differences and Quantizes frames. The blur and quantization parameters are learned to maximize action recognition while minimizing privacy recognition [34]. + +Notably, all the compared SoTA methods operate across entire frames (i.e. without selectivity) and have limited interpretability due to their learning-based obfuscation. + +# Implementation details. + +Training. No large scale dataset exists that allows for joint training of action and privacy classification on the same videos. Therefore, we pretrain our action and privacy networks on Kinetics400 [7] and Imagenet1k [33], respectively, and then finetune for specific datasets. We use the AdamW [31] optimizer with a learning rate of $3e^{-4}$ . The networks are trained with a patience scheme of 100 epochs that monitors the loss on the validation set. For action recognition, videos are temporally uniformly subsampled according to the architecture (Tab. 1, $f \times r$ column). + +RESULTS ACROSS NAIVE BASELINE OBFUSCATION METHODS + +
MethodIPNKTHSBU
↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ
Original88.0591.920.4890.1591.870.4986.3185.960.50
Masking36.4564.870.3637.5435.220.5152.6230.10.61
Pixelate4×485.5973.650.5683.6459.390.6273.2446.270.63
Pixelate16×1649.8165.760.4238.3338.430.5025.5329.840.48
Blur κ=13σ=10 (weak)76.2467.400.5444.5937.730.5372.7535.690.69
Blur κ=21σ=10 (strong)58.8065.920.4630.3831.840.4957.7734.210.62
Ours87.1151.930.68+0.1188.675.460.92+0.2986.7413.190.87+0.18
RESULTS ACROSS SOTA OBFUSCATION METHODS
BDQ [34]81.0059.000.6191.117.150.9284.0434.180.75
ALF [59]76.0065.000.5685.8919.270.8382.0048.000.67
ELR[44] s=1670.8264.320.5391.2288.860.5196.2782.970.57
ELR[44] s=3252.9663.290.4585.5782.560.5292.4264.890.64
ELR[44] s=6431.6362.700.3456.2158.350.4980.0543.610.68
Ours †85.2551.670.67+0.0689.444.310.93+0.0184.0411.700.86+0.11
Ours87.1151.930.68+0.0788.675.460.92±0.0086.7413.190.87+0.12
+ +Table 2. Comparison between Ours vs. Naive Baseline (top) and SOTA (bottom) Approaches as Top-1 Accuracy for both Action and Privacy labels, as well as $f_{0.5}$ Introduced in Sec. 4.1. We show Ours and Naive Baseline results averaged across all recognition algorithms presented in Tab. 1. Competing SOTA approaches only present results on one specific algorithm for action recognition (I3D [7]) and privacy attribute detection (ResNet50 [22]) as those approaches are network specific. To showcase a fair comparison we also present results with the same single recognition algorithms indicated as "Ours †". First and second best results indicated by bold and underline, respectively. Relative performance delta ( $\Delta$ ) is indicated in green (improvement) and orange (tie) between best and second best method. + +Note that we never perform any training on videos obfuscated by our approach. To implement warping, (4), we use pretrained RAFT to extract optical flow [52]. + +Privacy template selection. Our approach affords selective combination of a predefined set of privacy templates for a given dataset or application. For the experiments, we choose a subset of templates, $\tilde{\mathsf{T}}$ , from our complete identity preserving template library, $\mathsf{T}$ , shown in Fig. 3, to optimize performance on a given dataset. Figure 5 shows template-wise performance based on both action recognition and privacy for each dataset. For SBU and KTH, action recognition is stable with respect to templates; however, the same four templates notably reduce privacy recognition. For consistency on IPN, we also select the four templates that most reduce privacy recognition, even while preserving action recognition; although, the distinction is less striking. This selection process results in templates $\tilde{\mathsf{T}} = \{\text{torso, arm, leg, hair}\}$ for SBU and KTH vs. $\tilde{\mathsf{T}} = \{\text{cheek, eyes, forehead, hair}\}$ for IPN. Our code yields more results on the overlap of individual saliency maps. Runtime. The constant overhead to produce our obfuscated videos is $\approx 100ms/frame$ on a NVIDIA RTX4080 GPU. + +# 4.2. Results + +We evaluate our approach to privacy preserving action recognition on 13 different action recognition models and five different privacy attribute recognition models. Competing approaches evaluate on at most five action models + +[34, 43, 59]. Table 1 shows results on the original videos from IPN [3], SBU [60] and KTH [47], i.e. without any privacy preserving processing. Table 2 shows results comparing our approach vs. NaIVE BASELINES and SOTA. Comparison is given as the average across recognition algorithms presented in Table 1. Notably, while our approach compromises at most $1.48\%$ action recognition accuracy compared to performance on the original video (88.67% vs. $90.15\%$ on KTH), it always greatly improves privacy. + +Naive Baselines. Compared to the baselines, our approach scores highest with respect to the $f_{0.5}$ metric across all three datasets: IPN (+0.11), KTH (+0.29), and SBU (+0.18). Naive BASELINES show that action recognition correlates highly with privacy performance, as $f_{0.5}$ hovers around 0.5. The notable exception is $\mathrm{Pix}_{4\times 4}$ that performs well on IPN and KTH, and is the second best Naive BASELINE on SBU. A notable benefit of all these approaches is that they do not require retraining of the recognition algorithms. Still, none have a selective ability to obfuscate only certain parts. State of the art. A noteworthy difference between our vs. the alternative SoTA approaches is that they only work with the recognition algorithms on which they were trained. While this fact disadvantages our approach, as it lacks such retraining, it still outperforms the competing methods on all datasets: IPN (+0.06), KTH (+0.01), and SBU (+0.11). + +It is valuable to consider obfuscation approaches in terms of the relative importance of action recognition vs. privacy preservation. Figure 7 compares results for the SoTA ap + +![](images/9c12f8d35a275018076c280720822421dc99213da6e5dccd32577da5169ea4e3.jpg) +Figure 7. Performance across all datasets with varying $\lambda$ ; see (6). Values of $\lambda$ closer to 0 weigh action recognition higher, whereas values closer to 1 increase the importance of privacy preservation. + +proaches as that trade-off is varied in terms of $f_{\lambda}$ , (6). The sweep of $\lambda$ shows that our obfuscation approach performs better than any competing approach across the entire range. Especially with increasing $\lambda$ , the gap to other approaches widens as more emphasis is put on privacy preservation. + +Importance of selective obfuscation. Our approach is unique in its ability to selectively obfuscate particular regions within a video based on specific privacy attributes. This ability yields two benefits: + +(i) Selectivity aids in interpretability. Each selected template results in a saliency map, (2), which allows for visual inspection of what information is being obscured. Figure 4 highlights this benefit as the heat maps reveal the degree of obfuscation to be applied on a template-by-template basis. + +(ii) Individual privacy templates can be chosen and combined for best performance; see Fig. 5. Depending on the dataset, different privacy templates differently impact the performance of privacy and action recognition. Action recognition on SBU and KTH is relatively robust to privacy template selection; however, privacy preservation is best when a subset of templates (torso, leg, arm, hair) is selected and the rest remain unobfuscated. This fact derives directly from the saliency calculation, (1): If multiple individual saliency maps have small values, then other strong responses are scaled down in the combined final saliency map. Subsequently, this effect leads to less noise being applied in our selective obfuscation, (5), which can deteriorate privacy preservation. In contrast, action recognition on IPN is compromised if the hand template is selected (as expected with a gesture centric dataset), which documents that simply obfuscating the entire person leads to inferior action performance compared to selectively applied obfuscation. + +Importance of motion consistent noise. Action recognition is complicated as different datasets, and even individual actions within a dataset might require different recognition capabilities and can rely to different degrees on the modeling of motion [9]. This uncertainty on the role of motion in action recognition is exacerbated by the fact that different deep learning-based architectures are successful in capturing motion to varying degrees [32]. + +To see the importance of our proposed motion consistent noise for privacy preservation while maintaining good + +![](images/081822a89c02b238f22cbf300ef27ae3793c55608a9d1f71827e13dd308fd9ec.jpg) +Figure 8. Comparing Temporally Consistent vs. $iid$ Obfuscation for Action Recognition and Privacy Preservation. Action recognition performance decreases if noise is not temporally consistent. + +![](images/0475d0648df86f4dd5d69147a4c7e55028cefd9a8e00c3ef7324b00ae78f784d.jpg) + +action recognition, we compare to another version of our pipeline that obfuscates video frames using independent, identically distributed (iid) noise; see Fig. 8 and project page for video results. For all datasets, action recognition performance decreases if the noise is $iid$ . This result is sensible, because the $iid$ noise does not capture the temporal dynamics of the source video. In contrast, privacy preservation is robust to noise in either case. There is less impact on IPN action recognition, as the critical hand motion is never obfuscated, which underlines the importance of selective masking. Further insight on why motion consistent noise supports better action can be had through consideration of Fig. 6, where the ability of such noise to support optical flow estimation is illustrated. + +Limitations. Inevitably, there will be a trade-off between action recognition and privacy, as relevant information may be shared. In our approach, that trade-off could happen because our temporally consistent noise maintains dynamic information important for action recognition; however, it also might support motion based identification, e.g. gait recognition. Current protocols in privacy preserving action recognition do not consider motion-based identification. However, gait and related motion-based measurements are weak biometrics [26]; so, favouring action recognition may be apt. Moreover, if it becomes a concern, then it may be possible to apply motion perturbations that impede personal identification while maintaining action recognition. + +# 5. Conclusion + +Our work highlights that it is not necessary to train action recognition and privacy networks in an adversarial fashion for effective obfuscation of privacy attributes while maintaining strong action recognition performance. We show that a system based on local privacy templates, deep features that capture template semantics and selective noise obfuscation that is animated with source video motion can uphold privacy without hindering action recognition. Our approach is unique compared to alternative recent approaches in terms of interpretability and independence from particular action and privacy recognition algorithms. Our nine manually chosen templates, in combination with our proposed obfuscation technique outperforms other state-of-the-art approaches across three different datasets. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. ECCVW, 2022. 2, 4 +[2] Simon Baker, Daniel Scharstein, JP Lewis, Stefan Roth, Michael J Black, and Richard Szeliski. A database and evaluation methodology for optical flow. IJCV, 92:1-31, 2011. 6 +[3] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. IPN Hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In ICPR, 2021. 3, 5, 7 +[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2 +[5] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium, 2021. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3 +[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 5, 6, 7 +[8] Jiawei Chen, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Semi-coupled two-stream fusion ConvNets for action recognition at extremely low resolutions. In WACV, 2017. 2 +[9] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't I dance in the mall? Learning to mitigate scene bias in action recognition. NeurIPS, 32, 2019. 8 +[10] Robert T. Collins, Alan J Lipton, and Takeo Kanade. Introduction to the special section on video surveillance. IEEE TPAMI, 22(8):745-746, 2000. 2 +[11] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (SMID): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 1 +[12] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005. 2 +[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. SPAct: Self-supervised privacy preservation for action recognition. In CVPR, 2022. 1, 2, 6 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3, 5 +[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 5 + +[16] Christoph Feichtenhofer. X3D: Expanding architectures for efficient video recognition. In CVPR, 2020. 5 +[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 5 +[18] Joseph Fioresi, Ishan Rajendrakumar Dave, and Mubarak Shah. Ted-SPAD: Temporal distinctiveness for self-supervised privacy-preservation for video anomaly detection. In ICCV, 2023. 2 +[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, and Omer Ovenc. Scalable detection of offensive and non-compliant content/logo in product images. In WACV, 2020. 1 +[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2 +[21] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2 +[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2 +[24] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. PrivHAR: Recognizing human actions from privacy-preserving lens. In ECCV, 2022. 1, 2, 5 +[25] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In ECCV, 2022. 4, 5 +[26] Anil K Jain, Patrick Flynn, and Arun A Ross. Handbook of Biometrics. Springer Science & Business Media, 2007. 8 +[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLO by Ultralytics, 2023. 6 +[28] Amin Jourabloo, Xi Yin, and Xiaoming Liu. Attribute preserved face de-identification. In International Conference on Biometrics, 2015. 2 +[29] Alexander John Karran, Théophile Demazure, Antoine Hudon, Sylvain Senecal, and Pierre-Majorique Léger. Designing for confidence: The impact of visualizing artificial intelligence decisions. Frontiers in Neuroscience, 16: 883385, 2022. 2 +[30] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The Kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2 +[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 6 +[32] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal net + +works encode: Quantifying static vs. dynamic information. In CVPR, 2022. 8 +[33] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2, 6 +[34] Sudhakar Kumawat and Hajime Nagahara. Privacy-preserving action recognition via motion difference quantization. In ECCV, 2022. 1, 2, 5, 6, 7, 8 +[35] Agostina J Larrazabal, Nicolas Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020. 1 +[36] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. STPrivacy: Spatio-temporal privacy-preserving action recognition. In ICCV, 2023. 2 +[37] Bo Liu, Ming Ding, Sina Shaham, Wenny Rahayu, Farhad Farokhi, and Zihuai Lin. When machine learning meets privacy: A survey and outlook. ACM Computing Surveys, 54 (2):1-36, 2021. 2 +[38] Haochen Liu, Yiqi Wang, Wenqi Fan, Xiaorui Liu, Yaxin Li, Shaili Jain, Yunhao Liu, Anil Jain, and Jiliang Tang. Trustworthy AI: A computational perspective. ACM Transactions on Intelligent Systems and Technology, 14(1):1-59, 2022. 2 +[39] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91–110, 2004. 2 +[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4 +[41] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 6 +[42] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action detection. In ECCV, 2018. 2 +[43] Michael Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In AAAI, 2017. 2, 5, 6, 7 +[44] Michael Ryoo, Kiyoon Kim, and Hyun Yang. Extreme low resolution activity recognition with multi-siamese embedding learning. In AAAI, 2018. 2, 6, 7, 8 +[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 2 +[46] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 1 +[47] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a local SVM approach. In ICPR, 2004. 5, 7 +[48] Vijeta Sharma, Manjari Gupta, Anil Pandey, Deepti Mishra, and Anil Kumar. A review of deep-learning-based human activity recognition on benchmark video datasets. Applied Artificial Intelligence, 36(1):2093705, 2022. 2 + +[49] Jianbo Shi and Carlo Tomasi. Good features to track. In CVPR, 1994. 2 +[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 2 +[51] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 2 +[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 7 +[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2 +[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 5 +[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3 +[57] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local Neural Networks. In CVPR, 2018. 5 +[58] Zhenyu Wu, Zhangyang Wang, Zhaowen Wang, and Hailin Jin. Towards privacy-preserving visual recognition via adversarial training: A pilot study. In ECCV, 2018. 1, 5, 6 +[59] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE TPAMI, 44(4):2126-2139, 2020. 2, 5, 6, 7, 8 +[60] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In CVPR, 2012. 3, 5, 7 +[61] Yingying Zhu, Nandita M Nayak, and Amit K Roy-Chowdhury. Context-aware modeling and recognition of activities in video. In CVPR, 2013. 2 \ No newline at end of file diff --git a/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/images.zip b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8891e05a2a6dbee0dd54ebfcad9840d3122ce488 --- /dev/null +++ b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0261d201ada5a3630975d4afa9b737fc59c8cda1f0f26716f11b211df218589 +size 605381 diff --git a/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/layout.json b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e0c883b29606cf180732cfbd625c4af0ede26eba --- /dev/null +++ b/2024/Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition/layout.json @@ -0,0 +1,9612 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 462, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 462, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 462, + 140 + ], + "type": "text", + "content": "Selective, Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 64, + 161, + 175, + 204 + ], + "blocks": [ + { + "bbox": [ + 64, + 161, + 175, + 204 + ], + "lines": [ + { + "bbox": [ + 64, + 161, + 175, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 161, + 175, + 204 + ], + "type": "image", + "image_path": "20a30826185149aac9aaa823e6ef0fd01965888a5509f7faf4339f067a10a313.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 198, + 161, + 302, + 203 + ], + "blocks": [ + { + "bbox": [ + 198, + 161, + 302, + 203 + ], + "lines": [ + { + "bbox": [ + 198, + 161, + 302, + 203 + ], + "spans": [ + { + "bbox": [ + 198, + 161, + 302, + 203 + ], + "type": "image", + "image_path": "e85fb2a714e6b94a3065b4c7169f115b57fed2cfa9d75a5ce179716998ca2446.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 161, + 403, + 203 + ], + "blocks": [ + { + "bbox": [ + 326, + 161, + 403, + 203 + ], + "lines": [ + { + "bbox": [ + 326, + 161, + 403, + 203 + ], + "spans": [ + { + "bbox": [ + 326, + 161, + 403, + 203 + ], + "type": "image", + "image_path": "3aa1d670ae2438f54a386a119664bcd81c22e48eba197fa1b1ba9df0146d9c07.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 426, + 161, + 526, + 203 + ], + "blocks": [ + { + "bbox": [ + 426, + 161, + 526, + 203 + ], + "lines": [ + { + "bbox": [ + 426, + 161, + 526, + 203 + ], + "spans": [ + { + "bbox": [ + 426, + 161, + 526, + 203 + ], + "type": "image", + "image_path": "9419fbb9a64b9dd1e6a203d4f61065c9aedee8d92065285a64e280b44bde4c12.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 224, + 185, + 373 + ], + "blocks": [ + { + "bbox": [ + 48, + 224, + 185, + 373 + ], + "lines": [ + { + "bbox": [ + 48, + 224, + 185, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 185, + 373 + ], + "type": "image", + "image_path": "f6b16d27e3b4ab0312ed64869d755ee383bcc60a8243d1809e6f07d85fd17fad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 379, + 546, + 445 + ], + "lines": [ + { + "bbox": [ + 45, + 379, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 45, + 379, + 546, + 445 + ], + "type": "text", + "content": "Figure 1. Our goal is to hide privacy attributes without action recognition performance dropping. Left: Arbitrary images can be used to specify an interpretable template library defined by privacy attributes. Middle: A salience map is generated from privacy templates; example illustrates use of templates for personal identification. Right: The source video is masked with noise as guided by salience and animated by source video optical flow. Salience makes masking selective to privacy sensitive regions, while preserving scene context; optical flow preserves motion – both of which are critical for action recognition. The obfuscated video can be input directly to arbitrary privacy and action recognition systems without retraining. Zoomed circles highlight details only for illustration." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 216, + 220, + 373, + 365 + ], + "blocks": [ + { + "bbox": [ + 216, + 220, + 373, + 365 + ], + "lines": [ + { + "bbox": [ + 216, + 220, + 373, + 365 + ], + "spans": [ + { + "bbox": [ + 216, + 220, + 373, + 365 + ], + "type": "image", + "image_path": "da0e3093f1ba3ca6d02de67525ec1e1a61a791809e036dd5497d1969b9082244.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 217, + 545, + 365 + ], + "blocks": [ + { + "bbox": [ + 389, + 217, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 389, + 217, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 389, + 217, + 545, + 365 + ], + "type": "image", + "image_path": "cccfc1f37d98162a2a7cdc77b62521acbed9ef08f9b8f4198d105ce83cd6f84f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 456, + 192, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 456, + 192, + 467 + ], + "spans": [ + { + "bbox": [ + 143, + 456, + 192, + 467 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 472, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 472, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 45, + 472, + 288, + 676 + ], + "type": "text", + "content": "Concerns for the privacy of individuals captured in public imagery have led to privacy-preserving action recognition. Existing approaches often suffer from issues arising through obfuscation being applied globally and a lack of interpretability. Global obfuscation hides privacy sensitive regions, but also contextual regions important for action recognition. Lack of interpretability erodes trust in these new technologies. We highlight the limitations of current paradigms and propose a solution: Human selected privacy templates that yield interpretability by design, an obfuscation scheme that selectively hides attributes and also induces temporal consistency, which is important in action recognition. Our approach is architecture agnostic and directly modifies input imagery, while existing approaches generally require architecture training. Our approach offers more flexibility, as no training is required, and outperforms alternatives on three widely used datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 455, + 386, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 386, + 467 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 386, + 467 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Advances in state-of-the-art computer vision and machine learning enable deployment of such systems in the public sphere. Accompanying these initiatives, concerns arise for the privacy of individuals that are captured in acquired imagery [11, 19, 35, 46]. In particular, considerations arise regarding attributes that individuals want to keep confidential, yet that are revealed through visual information, even though they are not critical for the functioning of the deployed system, e.g. identity, age, gender and race. Video-based action recognition is an area of consideration as it has potential for widespread applications in surveillance and monitoring. These concerns have sparked interest in privacy preserving action recognition [13, 24, 34, 58]. These approaches process input imagery to obscure privacy attributes while maintaining action recognition performance. Contemporary approaches typically apply their obfuscation across entire input video frames and improvements have been made within this paradigm. Notably, however, there are downsides to this paradigm, as follows." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 254, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 254, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 254, + 712 + ], + "type": "text", + "content": "Code available f-iliic.github.io/SelectivePrivacyPreservation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18730" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "Collateral damage. Global masking strategies indiscriminately obscure the entire image, impacting regions within the scene that may exhibit high correlations with actions, albeit lack relevance to privacy. For example, it is known that masking objects and scene context can impair action recognition performance [61], yet these are lost in global masking. Furthermore, global masking strategies do not allow for selective attribute obfuscation and generally hide all attributes at once, even when not all are of concern." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 268 + ], + "type": "text", + "content": "Loss of dynamic information. The large change in input modalities from global masking necessitates the retraining of the action recognition module or the design of custom modules (adversarial training), which adds to the challenge of practical deployment. Indeed, even if applied more locally, the obfuscation can compromise the motion of the actors, which also can be important in action recognition [50]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 269, + 289, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 342 + ], + "type": "text", + "content": "Lack of interpretability. Finally, given that state-of-the-art approaches are end-to-end trained with limited concern for interpretability, the exact nature of what is being masked and how it is achieved can remain unclear. Lack of interpretability is an important concern in privacy preservation, because its lack can compromise user trust [29, 38]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 350, + 136, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 136, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 136, + 361 + ], + "type": "text", + "content": "1.1. Contributions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 289, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 538 + ], + "type": "text", + "content": "We present an approach to privacy preserving action recognition that responds directly to current limitations in four ways, as illustrated in Fig. 1. (i) The approach is based on local detection and selective obfuscation of privacy sensitive regions. This selectivity maintains global context information that is crucial to action recognition, yet unimportant for privacy. (ii) The local processing avoids large modality shifts in the imagery and is independent of the action recognition module itself; therefore, it does not require algorithm retraining, which is sometimes infeasible. (iii) The masking preserves interframe motion; so, that information is available for action recognition. (iv) The privacy sensitive masking is interpretable by design, and allows inspection through the explicitly generated saliency maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 549, + 132, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 132, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 132, + 562 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Privacy in machine learning. The need to protect privacy has garnered increased attention in the vision research community. Current models commonly consume large amounts of web data to learn generalizable representations [20, 33, 53], which inevitably invade personal information, such as identity and location. Moreover, in model deployment it also may be desirable to preserve privacy information. The concern for privacy is not limited to vision research, but extends across artificial intelligence, including natural language processing [5] and more general machine learning [37]. As these technologies find their way into broader society, privacy concerns must be considered." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "Privacy preserving action recognition. We focus on video-based action recognition. Recent developments in this area have yielded systems capable of strong performance on challenging datasets; for review see, e.g. [48]. Similarly, applications, including those in privacy-sensitive scenarios (e.g. surveillance [10] and monitoring [51]), are being developed. To provide useful spatiotemporal signals for action recognition, video clips are mostly collected to capture actors with a great level of clarity throughout the actions, thus increasing the chance of privacy leakage. Moreover, such datasets are expanding rapidly, e.g. [4, 30]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 220, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 518 + ], + "type": "text", + "content": "In response to these concerns, research on recognizing actions while preserving private information has emerged in recent years. Early work concentrated on devising models that can work on low-resolution videos [8, 43, 44]; these methods often operate at the cost of sacrificing action recognition performance. Other work along these lines developed face-anonymization techniques to prevent models from yielding high accuracy on face recognition [28, 42]. Still, those approaches cannot easily extend to other attributes and are correspondingly limited. More recent work has focused on implicitly learning transformation functions to anonymize videos in a data-driven fashion [13, 24, 36, 59], and also extending such approaches to anomaly detection [18]. To reduce model complexity, a competing approach follows a simple, yet effective procedure [34]: Frame subtraction, followed by broadband-filtering to yield motion descriptors for action recognition, while suppressing privacy attributes. A notable limitation of that approach is the relative weakness of frame differences as motion descriptors compared to common alternatives, e.g. optical flow. Another common limitation to all existing work is the lack of flexibility to select arbitrary private information to hide, i.e. preserving privacy of only critical attributes, while maintaining the visibility of others, to avoid obfuscation of visual cues that are essential to action classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "content": "Template matching. Building correspondences between templates and target images via matching is a foundation for modern vision research, e.g. SfM [21], image correspondences [45], detection [12] and tracking [49]. Features used in matching have advanced rapidly from primarily handcrafted (e.g. [12, 39]) to learning-based convolutional [23] or transformers [14]. Our solution is inspired by techniques seen in feature-based template matching. We apply their insights on matching templates to localize and selectively obfuscate privacy-sensitive regions. More specifically, we adopt DINO-ViT features [6] and compute similarities using the keys of the last attention layer from that architecture. Our use of DINO-ViT keys (rather than queries or values) is based on previous work finding them to perform best when applied to visual correspondence [1, 40]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18731" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 193, + 156 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 193, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 193, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 193, + 156 + ], + "type": "image", + "image_path": "cd80802ccf9e38d4ba8b524ce926a7f393454301b3166b2d902f36c9109cbf2a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 159, + 143, + 168 + ], + "lines": [ + { + "bbox": [ + 99, + 159, + 143, + 168 + ], + "spans": [ + { + "bbox": [ + 99, + 159, + 143, + 168 + ], + "type": "text", + "content": "(a) Overview" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 354, + 157 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 354, + 157 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 354, + 157 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 354, + 157 + ], + "type": "image", + "image_path": "0001021df2ac512b719a1903e53a751cdc84e614b311782cdc4f2f3e49898646.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 159, + 304, + 168 + ], + "lines": [ + { + "bbox": [ + 265, + 159, + 304, + 168 + ], + "spans": [ + { + "bbox": [ + 265, + 159, + 304, + 168 + ], + "type": "text", + "content": "(b) Matcher" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 178, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 545, + 278 + ], + "type": "text", + "content": "Figure 2. Overview of Method. 2a We present a privacy module that builds atop three components: (i) a semantic template library that contains attributes to be hidden, (ii) a descriptor matcher to localize template features in videos to be obscured and, (iii) an obfuscation method that is sensitive with respect to motion present in the scene. 2b A semantic descriptor matcher based on DINO [6]-ViT [14] keys is used to determine privacy salient regions in a video based on the template library. In our case, regions of interest correspond to those that can identify a person; however, this component can be adapted for other privacy attributes through specification of different templates. The result is a saliency map. 2c The saliency map is used as a weight to apply noise to the regions. The noise, however, is not static, but is warped with optical flow with an initial noise pattern image, " + }, + { + "bbox": [ + 46, + 178, + 545, + 278 + ], + "type": "inline_equation", + "content": "N(\\mathbf{p},1)" + }, + { + "bbox": [ + 46, + 178, + 545, + 278 + ], + "type": "text", + "content": ", for the purposes of preserving motion information in the source video. The similarity maps of all aggregated relevant privacy attributes are used to weigh the noise and apply it to the input image, obfuscating privacy sensitive information while not destroying the underlying temporal signal." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 543, + 156 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 543, + 156 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 543, + 156 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 543, + 156 + ], + "type": "image", + "image_path": "bb360d16467fe008ad852374cb7588d352e9783feadd1efb0009344f4f7e91ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 159, + 538, + 168 + ], + "lines": [ + { + "bbox": [ + 382, + 159, + 538, + 168 + ], + "spans": [ + { + "bbox": [ + 382, + 159, + 538, + 168 + ], + "type": "text", + "content": "(c) Temporally Consistent Selective Obfuscation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 297, + 162, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 162, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 162, + 312 + ], + "type": "text", + "content": "3. Technical approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 317, + 288, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 288, + 485 + ], + "type": "text", + "content": "Our approach to privacy preserving action recognition is a stand alone module that operates by preprocessing video that subsequently is input to arbitrary action and privacy recognition algorithms, i.e. it is independent of the recognition algorithms and does not entail any retraining of those algorithms. Our method consists of three key components; see Fig. 2a: (i) A template library that covers privacy attributes to be preserved, (ii) a matcher that produces saliency maps between selected templates and images where privacy is to be preserved and (iii) an obfuscator that uses the saliency maps to hide privacy attributes of concern in a temporally consistent fashion to preserve motion in the video. The remainder of this section details each of these components." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 491, + 149, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 149, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 149, + 505 + ], + "type": "text", + "content": "3.1. Template library" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "To obfuscate privacy sensitive regions in images in an interpretable fashion, we need an explicit set of “attributes” to be hidden from the target video. Such a concept template library, T, can be built manually choosing image patches corresponding to features one wishes to obfuscate. In this paper, we concentrate on privacy attributes related to personal identification. Therefore, we use landmark anatomical features, corresponding to detailed facial landmarks and the hand (Fig. 3 left: forehead, hair, eye, cheek, lips, hand) as well as larger body parts (Fig. 3 right: arm, torso, legs). The choice to focus on preservation of human identity is motivated by the fact that all three of our evaluation datasets define their privacy attributes on attributes relating to person identity. Notably, however, such a template library easily can be extended depending on the particular task at hand. For example, attributes pertaining to location (e.g. street signs, distinctive scene objects) could guide development" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 299, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 335 + ], + "type": "text", + "content": "of a complimentary set of templates. In any case, given a template library, a user can selectively combine templates to obscure attributes of concern in a particular application." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "text", + "content": "Formally, we define each template " + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "text", + "content": " as an element in the set " + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "inline_equation", + "content": "\\mathsf{T} = \\{\\tau_1,\\tau_2,\\dots,\\tau_n\\}" + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "text", + "content": ". Further, let " + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{T}}\\subseteq \\mathsf{T}" + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "text", + "content": " be the subset of templates that the user selects from the library, " + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "inline_equation", + "content": "\\mathsf{T}" + }, + { + "bbox": [ + 304, + 336, + 546, + 550 + ], + "type": "text", + "content": ", for a particular subset of privacy attributes to be preserved for a given dataset or application. Our manual approach to template selection leads to concepts that are interpretable by design. In particular, we use two source images, shown in Fig. 3, one for small-scale features on the face, and one for large regions, taken from the IPN [3] and SBU [60] datasets. We choose these two datasets because of their complimentarity in template selection: IPN focuses on small scale features (e.g. facial and hand), while SBU focuses on larger scale features (e.g. larger body parts), as detailed later in Sec. 4.1. The choice of the particular template images is not critical to the functioning of our approach, because we extract semantic features from the templates that are known to generalize well across images instances [6], as described next." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 562, + 535, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 535, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 535, + 575 + ], + "type": "text", + "content": "3.2. Matching: Local patch descriptor templates" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Semantic features. Our requirements for good features to match between privacy templates in our library and frames in an action recognition video are straightforward: We require (i) semantically rich features that generalize across (in our case) different individuals and (ii) high spatial resolution to perform privacy obfuscation in a localized manner without destroying regions that are required for action recognition. We find that DINO-ViT features fit our needs very well [6, 56]: (i) They have been trained to yield high similarity for semantically related concepts (e.g. objects and their parts), while suppressing the similarity of unrelated" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18732" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 73, + 187, + 155 + ], + "blocks": [ + { + "bbox": [ + 49, + 73, + 187, + 155 + ], + "lines": [ + { + "bbox": [ + 49, + 73, + 187, + 155 + ], + "spans": [ + { + "bbox": [ + 49, + 73, + 187, + 155 + ], + "type": "image", + "image_path": "87f98b68be415c66c55c4db4b8056043811dc2e38745915d0d25612367d5c7dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 288, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 233 + ], + "type": "text", + "content": "Figure 3. Template Library consisting of Patches Chosen from Anatomical Landmark Regions. These specific images are passed through a DINO-ViT feature extractor. The keys, corresponding to spatial locations of the highlighted patches, are chosen as the templates for matching to input images to obtain semantically similar regions for obfuscation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 72, + 284, + 156 + ], + "blocks": [ + { + "bbox": [ + 211, + 72, + 284, + 156 + ], + "lines": [ + { + "bbox": [ + 211, + 72, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 211, + 72, + 284, + 156 + ], + "type": "image", + "image_path": "ab1be94bbfe9748f517c48530b3a0601a0b534c117fedcf324a51e92309b390b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "text", + "content": "matters (e.g. background). (ii) They support local patch feature extraction over high resolution images without losing global context. Elsewhere, DINO-ViT keys have proven especially useful in matching between templates and target images [1, 40]. Following these advances, we use last attention layer DINO-ViT keys computed from privacy templates (e.g. Fig. 3) to match against input images to compute privacy saliency maps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": "Privacy saliency matching. To find privacy salient regions in an input video, a similarity map, " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": ", is computed between each frame, " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": ", in the video and the selected set of privacy templates, " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{T}}" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": ". The image is tiled with " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": " patches, " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "I_{j}, j \\in \\{1,\\dots,m\\}" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": ", and for each patch DINO-ViT keys, " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\mathsf{K}(I_j)" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": ", are extracted. These features are matched with DINO-ViT keys, " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\mathsf{K}(\\tau_i)" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": ", for all selected templates, " + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\tau_{i} \\in \\tilde{\\mathsf{T}}" + }, + { + "bbox": [ + 47, + 355, + 288, + 464 + ], + "type": "text", + "content": ". Clipped cosine similarity is used to establish the saliency of each image patch according to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 287, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 287, + 505 + ], + "type": "interline_equation", + "content": "s _ {j} = \\frac {1}{| \\tilde {\\mathbf {T}} |} \\sum_ {i = 1} ^ {| \\tilde {\\mathbf {T}} |} \\max \\left(0, \\frac {\\langle \\mathrm {K} (\\tau_ {i}) , \\mathrm {K} (I _ {j}) \\rangle}{\\| \\mathrm {K} (\\tau_ {i}) \\| \\| \\mathrm {K} (I _ {j}) \\|}\\right), \\forall j \\in [ [ 1, m ] ], \\tag {1}", + "image_path": "ae6b2403e7bd643bca144ce7f774083ba75902a668fd8ee6c943f3a5529c82d1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " is inner product and " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "|\\tilde{\\mathsf{T}} |" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " is the cardinality of " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{T}}" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " the set of selected privacy templates. Clipping is used because only positive values imply salience. This calculation is performed for every patch in every image of the video. Subsequently, the " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " saliency patches are reassembled in the shape of the original image to produce the final saliency map," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 599, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 599, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 287, + 612 + ], + "type": "interline_equation", + "content": "S = \\mathcal {R} \\left(s _ {1}, \\dots , s _ {m}; h, w\\right), \\tag {2}", + "image_path": "9867da10826050cbc7fa0e4738f45c9251ac0f5f34a88cebb4f717f233f55a4d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " a function that accepts image tiles, " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", and reshapes them into their original image format of height " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". Note that a separate salience map, " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", is calculated for each frame in a video of interest, " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". The entire process of feature extraction, matching and the resulting saliency maps are summarized in Fig. 2b. Example saliency maps in Fig. 4 illustrate the ability of our approach to capture all of our templates in a variety of scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 488, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 488, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 488, + 84 + ], + "type": "text", + "content": "3.3. Temporally consistent obfuscation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 90, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 258 + ], + "type": "text", + "content": "Our goal is to mask privacy sensitive regions in images to obfuscate them. If we were to apply masks to every frame independently, then temporal information important to action recognition, e.g. motion of actors, would be destroyed as well. We empirically document the issue in Sec. 4.2. In response to this challenge, we follow previous work that presented a method for producing temporally consistent spatial noise across videos that supported action recognition, while obscuring appearance information in single frames [25]. While the previous work applied noise patterns uniformly across entire frames, we instead weight them by our privacy salience maps, " + }, + { + "bbox": [ + 304, + 90, + 545, + 258 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 90, + 545, + 258 + ], + "type": "text", + "content": ", to preserve as much context information as possible. The remainder of this subsection details our approach, with an outline shown in Fig. 2c." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": "Noise pattern initialization. Let " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (x, y)" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": " be image coordinates and " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": " time. We initialize a noise image, " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "N(\\mathbf{p}, 1)" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": ", with the same dimensions as a frame from the input video (i.e. " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "h \\times w \\times 3" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": "), with " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": " height, " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": " width and 3 the number of colour channels. The dataset mean, " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": ", and standard deviation, " + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": ", are used to define a uniform distribution from which individual pixel intensities are drawn according to" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 353, + 545, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 353, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 359, + 353, + 545, + 366 + ], + "type": "interline_equation", + "content": "N (\\mathbf {p}, t = 1) \\sim \\mathcal {U} [ \\mu - \\sigma , \\mu + \\sigma ]. \\tag {3}", + "image_path": "a6bff20d951095391f2230c80e095456007e15ed8e96fab5fec81ab0ca80cecc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": "Motion consistent noise. To create motion consistent noise the initial random frame, " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "N(\\mathbf{p}, 1)" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": ", is warped forward with flow fields derived from the original video, " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "I(\\mathbf{p}, t)" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": ", as extracted by an optical flow algorithm. Let " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{v}(\\mathbf{p}, t) = (u(\\mathbf{p}, t), v(\\mathbf{p}, t))" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": " be the flow field that maps points, " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": ", in frame " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": " to those in frame " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 377, + 546, + 472 + ], + "type": "text", + "content": " the horizontal and vertical components of the flow. Then, a motion consistent noise sequence is generated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 479, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 479, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 364, + 479, + 545, + 493 + ], + "type": "interline_equation", + "content": "N (\\mathbf {p}, t) = N \\left(\\mathbf {p} + \\mathbf {v} (\\mathbf {p}, t), 1\\right) \\tag {4}", + "image_path": "e0891f5964f323a18b0edf24dde6ac9e0aa594947b283e5e8ecf210b03420c90.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 522 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 498, + 545, + 522 + ], + "type": "inline_equation", + "content": "t \\in \\{2, \\dots, T\\}" + }, + { + "bbox": [ + 304, + 498, + 545, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 545, + 522 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 498, + 545, + 522 + ], + "type": "text", + "content": " the number of frames in the original input video." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "text", + "content": "Selective privacy obfuscation. The computed video sequence, " + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "text", + "content": ", shows no single frame appearance related to the original video as it is random noise; however, when viewed as a video it reveals the motion present in the original, cf. [25]. Direct use of this synthesized video obscures privacy attributes; however, it also obscures other context information that could be of use in action recognition. So, instead we selectively apply " + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "text", + "content": " to every frame " + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "text", + "content": " in the video by using the privacy salience maps " + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 527, + 545, + 635 + ], + "type": "text", + "content": ", according to" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 641, + 545, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 641, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 310, + 641, + 545, + 661 + ], + "type": "interline_equation", + "content": "O (\\mathbf {p}, t) = I (\\mathbf {p}, t) + \\left(S (\\mathbf {p}, t) \\times \\left(N (\\mathbf {p}, t) - I (\\mathbf {p}, t)\\right)\\right). \\tag {5}", + "image_path": "cfab593256fc41f6cbb533cbfebd4bd5f573eb66456ababa9f1bac4b91dce179.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The resulting video, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", contains selectively obfuscated regions, built with interpretable templates by design and contains motion information that (in principle) does not differ from the original input video." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18733" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 97, + 168 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 97, + 168 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 97, + 168 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 97, + 168 + ], + "type": "image", + "image_path": "393cb808ada3384756f3cfcb6fb3848694468154b8f7f136eef732f817a541c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 173, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 173, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 547, + 217 + ], + "type": "text", + "content": "Figure 4. Saliency Maps for Descriptors in the Template Library. The manual selection of these templates allows for interpretability of the obfuscated parts of the image by design. The matched DINO-ViT features capture rich semantic information and allow for detailed spatial localization due to the nature of vision transformers. These saliency maps can then be combined for obfuscating any combination of templates depending on the task at hand." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 99, + 72, + 146, + 168 + ], + "blocks": [ + { + "bbox": [ + 99, + 72, + 146, + 168 + ], + "lines": [ + { + "bbox": [ + 99, + 72, + 146, + 168 + ], + "spans": [ + { + "bbox": [ + 99, + 72, + 146, + 168 + ], + "type": "image", + "image_path": "385e2b1c70afa594723da8bf34675e4199f1c012ba5660100d32790efc41a6a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 149, + 72, + 196, + 168 + ], + "blocks": [ + { + "bbox": [ + 149, + 72, + 196, + 168 + ], + "lines": [ + { + "bbox": [ + 149, + 72, + 196, + 168 + ], + "spans": [ + { + "bbox": [ + 149, + 72, + 196, + 168 + ], + "type": "image", + "image_path": "96aa1c9561d7f95c812a34214e2bd196b1070a08eb06e63d7d9cbb1760b8196e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 198, + 72, + 246, + 168 + ], + "blocks": [ + { + "bbox": [ + 198, + 72, + 246, + 168 + ], + "lines": [ + { + "bbox": [ + 198, + 72, + 246, + 168 + ], + "spans": [ + { + "bbox": [ + 198, + 72, + 246, + 168 + ], + "type": "image", + "image_path": "eca47d4f1a9f4c1a1a330221b5ea3742f3053d9ab9f74854c2a5c5c3e9cc4c5b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 72, + 295, + 168 + ], + "blocks": [ + { + "bbox": [ + 249, + 72, + 295, + 168 + ], + "lines": [ + { + "bbox": [ + 249, + 72, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 249, + 72, + 295, + 168 + ], + "type": "image", + "image_path": "67661313effcfe1e50450dac036f4e13da91324ffd9a8d422d10b4b59d3707a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 345, + 168 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 345, + 168 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 345, + 168 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 345, + 168 + ], + "type": "image", + "image_path": "0138c17a11f2f31975363bcfbec31f46a84aafd6abdf2012011f8be55ffd016e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 347, + 72, + 395, + 168 + ], + "blocks": [ + { + "bbox": [ + 347, + 72, + 395, + 168 + ], + "lines": [ + { + "bbox": [ + 347, + 72, + 395, + 168 + ], + "spans": [ + { + "bbox": [ + 347, + 72, + 395, + 168 + ], + "type": "image", + "image_path": "48454e4984540219b0e5cbf8c31af116f1e1bb4717811c005447c82d6f201803.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 397, + 72, + 445, + 168 + ], + "blocks": [ + { + "bbox": [ + 397, + 72, + 445, + 168 + ], + "lines": [ + { + "bbox": [ + 397, + 72, + 445, + 168 + ], + "spans": [ + { + "bbox": [ + 397, + 72, + 445, + 168 + ], + "type": "image", + "image_path": "bfc3bb592be2fd5d95212638237843e46953e8c3604e8ce16bd5825ff95dc22a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 447, + 72, + 494, + 168 + ], + "blocks": [ + { + "bbox": [ + 447, + 72, + 494, + 168 + ], + "lines": [ + { + "bbox": [ + 447, + 72, + 494, + 168 + ], + "spans": [ + { + "bbox": [ + 447, + 72, + 494, + 168 + ], + "type": "image", + "image_path": "408507cccd7eaef775fa8935ba63295d1e3aa4ad5bf3a3fbf16819f28bc0f5bb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 497, + 72, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 497, + 72, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 497, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 497, + 72, + 545, + 168 + ], + "type": "image", + "image_path": "ca77c0593b1eee06660f5b8134f754cb13fa0e2c128961af5722c09a1de0b412.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 48, + 245, + 287, + 502 + ], + "blocks": [ + { + "bbox": [ + 87, + 236, + 247, + 244 + ], + "lines": [ + { + "bbox": [ + 87, + 236, + 247, + 244 + ], + "spans": [ + { + "bbox": [ + 87, + 236, + 247, + 244 + ], + "type": "text", + "content": "ACTION RECOGNITION ON SOURCE DATASETS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 245, + 287, + 502 + ], + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 502 + ], + "type": "table", + "html": "
Networkf × rIPNKTHSBU
C2D [57]8 × 874.5683.0087.50
CSN [55]32 × 291.7388.3390.91
E2S X3D L [25]16 × 595.1595.3391.11
E2S X3D M [25]16 × 589.3892.3386.36
E2S X3D S [25]13 × 685.1692.0082.98
I3D [7]8 × 883.1589.6782.95
MVIT[15]16 × 488.0090.0092.55
R2+1D [54]16 × 489.8087.3381.91
Slow [17]8 × 885.7689.0088.64
SlowFast [17]32 × 288.0687.3390.00
X3D L [16]16 × 594.4194.0090.22
X3D M [16]16 × 591.6793.0081.91
X3D S [16]13 × 687.8190.6775.00
Average88.0590.1586.31
PRIVACY PRESERVATION FOR SOURCE DATASETS
NetworkIPNKTHSBU
ResNet18 [22]88.4687.3390.43
ResNet50 [22]92.3190.0096.81
ResNet101 [22]94.2394.0084.04
ViTb/16 [14]94.2394.0079.79
ViTb/32 [14]90.3894.0078.72
Average91.9291.8785.96
", + "image_path": "cfecd684f96f8771134147364d00b9ceb3a58d319a9aa12deee01e600fbfdd87.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "type": "text", + "content": "Table 1. Top-1 Accuracy for all Privacy and Action models on the original unmodified (source) videos, i.e. without privacy obfuscation. Bold and underline indicate first and second best, resp. Number of frames and temporal sampling rate indicated as " + }, + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "type": "inline_equation", + "content": "f \\times r" + }, + { + "bbox": [ + 46, + 506, + 287, + 551 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 572, + 170, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 170, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 170, + 586 + ], + "type": "text", + "content": "4. Empirical evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "The privacy obfuscated video, (5), serves directly as input to action and privacy recognition. No specialized development, training or other modification of the recognition algorithms nor adaptation of the privacy preservation system is necessary. Indeed, a major difference compared to competing state-of-the-art approaches (e.g. [34, 43, 59]) is that we do not retrain networks with our obfuscated data, while they do retrain. We exploit the fact that privacy attributes are generally independent from action recognition cues, which is enabled by our unique selective obfuscation approach." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 238, + 368, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 368, + 249 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 368, + 249 + ], + "type": "text", + "content": "4.1. Protocol" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": "Datasets. We use three datasets commonly used for investigating action recognition and privacy, IPN [3], SBU [60] and KTH [47], which pose different challenges concerning both action and privacy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "text", + "content": "IPN is a large-scale video-based hand gesture recognition dataset that consists of 50 actors performing 13 static or dynamic gestures, against three different backgrounds [3]. The performed gestures are used as the action labels and actor genders are used as the privacy labels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 364, + 545, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 423 + ], + "type": "text", + "content": "SBU is a video dataset depicting eight human interactions. Each video is a video of actor-pairs, in the same laboratory environment. Action labels are derived from the actor interactions and the privacy attributes are the unique pairings of seven different actors resulting in 13 privacy labels." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 424, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 543 + ], + "type": "text", + "content": "KTH has 25 actors doing one of six actions [47]. Each action is performed four times in different environments. The six action classes are used for action recognition and 25 actor identities serve as privacy labels. In the originally proposed splits, videos of any one actor are fully contained in one set, since the intent of the KTH dataset was not concerned with actor identity recognition. For privacy identity, each actor must appear in both sets; so, the data is split such that each action is performed twice by each person in the training set, and once in the validation and test sets." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": "Metrics. Privacy recognition results are obtained following standard practice [34], averaging outputs over multiple frames (32 for IPN and KTH, and 16 for SBU) from the same video. Action recognition results are obtained by reporting the top-1 accuracy, as per convention [24, 58, 59]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "content": "It also is interesting to gauge the trade-off between action and privacy recognition. Let " + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "inline_equation", + "content": "0 \\leq a \\leq 1" + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "content": " be normalized action recognition performance derived from dividing accuracy percentages by 100, and let " + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "content": " be defined analogously for privacy recognition. To quantify their trade-off we use a linear combination of " + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "inline_equation", + "content": "1 - p" + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "content": " and define" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 358, + 682, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 682, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 358, + 682, + 545, + 695 + ], + "type": "interline_equation", + "content": "f _ {\\lambda} (a, p) = (1 - \\lambda) a + \\lambda (1 - p), \\tag {6}", + "image_path": "dde6bd98733a7eb1d90f5a6ebb11f5d3e40d30f04d5248a0305cbf4f52a18c0b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "0 \\leq \\lambda \\leq 1" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " weighing the relative importance of action" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18734" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 214, + 159 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 214, + 159 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 214, + 159 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 214, + 159 + ], + "type": "image", + "image_path": "c32ab3289d0c3ea9673aedc497605898b3b92c621699edd203f27d69fcc4f83a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 166, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 212 + ], + "type": "text", + "content": "Figure 5. Obfuscation with a Single Attribute and the Impact on Performance. Attribute importance is dataset dependent. For example, notice how the 'Hand' template contributes to a large decrease in action recognition performance on IPN, as the action is determined solely by the hand, whereas on SBU it does not. Optimally blue is high and red is low. Corresponding qualitative examples of saliency maps for each individual template are shown in Fig. 4. Bold text along the abscissa of each plot indicates the templates used for the final results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 376, + 159 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 376, + 159 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 376, + 159 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 376, + 159 + ], + "type": "image", + "image_path": "48b07ac760a5cea5141dff92858a6fcba3687aceb3a0e4aa908559c068c88661.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 539, + 159 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 539, + 159 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 539, + 159 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 539, + 159 + ], + "type": "image", + "image_path": "d6e6b41fce5d496f4afc9d92b0f3fbf2c2b503e078f288bf163871a499117817.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 54, + 232, + 108, + 286 + ], + "blocks": [ + { + "bbox": [ + 54, + 232, + 108, + 286 + ], + "lines": [ + { + "bbox": [ + 54, + 232, + 108, + 286 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 108, + 286 + ], + "type": "image", + "image_path": "f86584d0f0a14883dc713d8206978360ea165299a85c2039fa0bef242006685b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 112, + 232, + 165, + 286 + ], + "blocks": [ + { + "bbox": [ + 112, + 232, + 165, + 286 + ], + "lines": [ + { + "bbox": [ + 112, + 232, + 165, + 286 + ], + "spans": [ + { + "bbox": [ + 112, + 232, + 165, + 286 + ], + "type": "image", + "image_path": "1888d6682e9aad01a779860a77a1508c62aad6d73f6632d7b9f4717bb894e604.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 170, + 232, + 223, + 286 + ], + "blocks": [ + { + "bbox": [ + 170, + 232, + 223, + 286 + ], + "lines": [ + { + "bbox": [ + 170, + 232, + 223, + 286 + ], + "spans": [ + { + "bbox": [ + 170, + 232, + 223, + 286 + ], + "type": "image", + "image_path": "e8b8eb8fdee01620a612a400d88df91adaf82b28df088ef324b01e83f3dcfbfc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 227, + 232, + 281, + 286 + ], + "blocks": [ + { + "bbox": [ + 227, + 232, + 281, + 286 + ], + "lines": [ + { + "bbox": [ + 227, + 232, + 281, + 286 + ], + "spans": [ + { + "bbox": [ + 227, + 232, + 281, + 286 + ], + "type": "image", + "image_path": "388c0d6c7f41982a1f29deb40f1da619ba0c6fb13cdd949c6543bcb893ab4abd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 54, + 291, + 108, + 345 + ], + "blocks": [ + { + "bbox": [ + 54, + 291, + 108, + 345 + ], + "lines": [ + { + "bbox": [ + 54, + 291, + 108, + 345 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 108, + 345 + ], + "type": "image", + "image_path": "da0848338641762a6c711f60f856772bbee510eb4619b6928a565b5bb9e71722.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 112, + 291, + 166, + 345 + ], + "blocks": [ + { + "bbox": [ + 112, + 291, + 166, + 345 + ], + "lines": [ + { + "bbox": [ + 112, + 291, + 166, + 345 + ], + "spans": [ + { + "bbox": [ + 112, + 291, + 166, + 345 + ], + "type": "image", + "image_path": "d3c98780a73bd5540276e4e877f578689a983daae423c64252319eacdaf0f723.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 170, + 291, + 225, + 345 + ], + "blocks": [ + { + "bbox": [ + 170, + 291, + 225, + 345 + ], + "lines": [ + { + "bbox": [ + 170, + 291, + 225, + 345 + ], + "spans": [ + { + "bbox": [ + 170, + 291, + 225, + 345 + ], + "type": "image", + "image_path": "9f6802e627f6b6aacca0c7b0fe71f31281d73b93fd33a28e28870e21bb41eeef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 229, + 291, + 284, + 345 + ], + "blocks": [ + { + "bbox": [ + 229, + 291, + 284, + 345 + ], + "lines": [ + { + "bbox": [ + 229, + 291, + 284, + 345 + ], + "spans": [ + { + "bbox": [ + 229, + 291, + 284, + 345 + ], + "type": "image", + "image_path": "99bff6933cd376107dff4c5d4515f11b049fc0dee9194a377feb64c7054fd2b1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 57, + 346, + 111, + 401 + ], + "blocks": [ + { + "bbox": [ + 57, + 346, + 111, + 401 + ], + "lines": [ + { + "bbox": [ + 57, + 346, + 111, + 401 + ], + "spans": [ + { + "bbox": [ + 57, + 346, + 111, + 401 + ], + "type": "image", + "image_path": "e0ffec1a07cc18dfc1d3bc23e7fcab63829511580fbcde41c9db5c037cf18442.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 404, + 97, + 415 + ], + "lines": [ + { + "bbox": [ + 61, + 404, + 97, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 404, + 97, + 415 + ], + "type": "text", + "content": "ELR [44]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 115, + 346, + 168, + 401 + ], + "blocks": [ + { + "bbox": [ + 115, + 346, + 168, + 401 + ], + "lines": [ + { + "bbox": [ + 115, + 346, + 168, + 401 + ], + "spans": [ + { + "bbox": [ + 115, + 346, + 168, + 401 + ], + "type": "image", + "image_path": "d3f411b3776886002ea48644423e1c1c1614f19d08325bb176802ef0ed9b0778.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 404, + 157, + 415 + ], + "lines": [ + { + "bbox": [ + 119, + 404, + 157, + 415 + ], + "spans": [ + { + "bbox": [ + 119, + 404, + 157, + 415 + ], + "type": "text", + "content": "BDQ [34]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 171, + 346, + 226, + 402 + ], + "blocks": [ + { + "bbox": [ + 171, + 346, + 226, + 402 + ], + "lines": [ + { + "bbox": [ + 171, + 346, + 226, + 402 + ], + "spans": [ + { + "bbox": [ + 171, + 346, + 226, + 402 + ], + "type": "image", + "image_path": "a3f598f0afd372a17bbdb3c97a431af731c749233b381dfe23c3c6ebb6fba902.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 404, + 215, + 415 + ], + "lines": [ + { + "bbox": [ + 179, + 404, + 215, + 415 + ], + "spans": [ + { + "bbox": [ + 179, + 404, + 215, + 415 + ], + "type": "text", + "content": "ALF [59]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 422, + 287, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 489 + ], + "type": "text", + "content": "Figure 6. Optical flow from two consecutive frames from competing approaches. Only our approach retains dynamic information that supports high-quality optical flow recovery. Top row: Single frames as processed by comparison algorithms. Middle row: Full frame optical flow recovery. Bottom row: Zoomed optical flow details. Optical flow shown in Middlebury colour coding [2]." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 230, + 346, + 284, + 402 + ], + "blocks": [ + { + "bbox": [ + 230, + 346, + 284, + 402 + ], + "lines": [ + { + "bbox": [ + 230, + 346, + 284, + 402 + ], + "spans": [ + { + "bbox": [ + 230, + 346, + 284, + 402 + ], + "type": "image", + "image_path": "29c4ded6542c9c0f4a50bc61e275dbfe845beef23912b749877a2bfdc4ee4ce6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 404, + 265, + 414 + ], + "lines": [ + { + "bbox": [ + 247, + 404, + 265, + 414 + ], + "spans": [ + { + "bbox": [ + 247, + 404, + 265, + 414 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": "recognition vs. privacy. This metric is suitable to create a linear ranking as " + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "inline_equation", + "content": "(1 - p)" + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": " are optimally 1 ensuring " + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "inline_equation", + "content": "f_{\\lambda} \\in [0,1]" + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": ". If privacy or action recognition are not equally important, then they can be weighted accordingly; e.g. if privacy is critical, then " + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": " can be increased. In our main experiments we use " + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": ", abbreviated as " + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "inline_equation", + "content": "f_{0.5}" + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": " in Tab. 2 and show an ablation over different values of " + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 502, + 287, + 587 + ], + "type": "text", + "content": " in Fig. 7." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 591, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 687 + ], + "type": "text", + "content": "Competing obfuscation approaches. We briefly describe NAIVE BASELINE approaches that solely rely on full-frame pixelation and blurring as well as actor masking that have been studied across the literature [13, 58]. We also highlight other State-of-The-Art approaches (SOTA) [34, 43, 59] that we compare against. Visual examples of the SoTA approaches are shown in Fig. 6; also shown are optical flows of consecutive frames for qualitative comparison." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Mask Obfuscation. All our datasets involve people performing actions and the privacy attributes are related to geo-" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 232, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 292 + ], + "type": "text", + "content": "ple. A naive way to preserve privacy in such videos is to completely mask out the actors. To do so, we use the YOLOv8 implementation [27] based on the original YOLO [41]. We report results based on the masked region filled with the mean intensity of the image covered by the mask." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "text", + "content": "Pixelation. Pixelation applies average pooling over regions of dimensions " + }, + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "inline_equation", + "content": "x \\times x" + }, + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "text", + "content": " on the input image sequence. We choose two scales of " + }, + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "inline_equation", + "content": "x \\in \\{4,16\\}" + }, + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "text", + "content": " for the patch-size to pool. In the result tables these methods are abbreviated as " + }, + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "inline_equation", + "content": "\\operatorname{Pix}_{x \\times x}" + }, + { + "bbox": [ + 304, + 294, + 545, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "content": "Blur. Blur applies a Gaussian blur to the images that have been rescaled to " + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "content": " pixels. The parameters of the Gaussian are the kernel size, " + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "content": ", and the standard deviation of the kernel, " + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "content": ". We choose values for weak and strong blurs, " + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\kappa = 13" + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\kappa = 21" + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "content": " respectively, as consistent with other work on obfuscation methods [59]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 419, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 456 + ], + "type": "text", + "content": "ELR initially reduces frames to Extreme Low Resolution and subsequently applies a set of learned inverse superresolution transforms to support action recognition [43]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 458, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 483 + ], + "type": "text", + "content": "ALF is an Adversarial Learning Framework for action recognition that takes into account a privacy budget [59]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 485, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 533 + ], + "type": "text", + "content": "BDQ is a privacy-preserving encoder that sequentially Blurs, Differences and Quantizes frames. The blur and quantization parameters are learned to maximize action recognition while minimizing privacy recognition [34]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "content": "Notably, all the compared SoTA methods operate across entire frames (i.e. without selectivity) and have limited interpretability due to their learning-based obfuscation." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 581, + 410, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 581, + 410, + 592 + ], + "spans": [ + { + "bbox": [ + 306, + 581, + 410, + 592 + ], + "type": "text", + "content": "Implementation details." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Training. No large scale dataset exists that allows for joint training of action and privacy classification on the same videos. Therefore, we pretrain our action and privacy networks on Kinetics400 [7] and Imagenet1k [33], respectively, and then finetune for specific datasets. We use the AdamW [31] optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "3e^{-4}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". The networks are trained with a patience scheme of 100 epochs that monitors the loss on the validation set. For action recognition, videos are temporally uniformly subsampled according to the architecture (Tab. 1, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "f \\times r" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " column)." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18735" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 81, + 547, + 294 + ], + "blocks": [ + { + "bbox": [ + 179, + 69, + 415, + 79 + ], + "lines": [ + { + "bbox": [ + 179, + 69, + 415, + 79 + ], + "spans": [ + { + "bbox": [ + 179, + 69, + 415, + 79 + ], + "type": "text", + "content": "RESULTS ACROSS NAIVE BASELINE OBFUSCATION METHODS" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 81, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 47, + 81, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 81, + 547, + 294 + ], + "type": "table", + "html": "
MethodIPNKTHSBU
↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ↑Action↓Privacy↑f0.5Δ
Original88.0591.920.4890.1591.870.4986.3185.960.50
Masking36.4564.870.3637.5435.220.5152.6230.10.61
Pixelate4×485.5973.650.5683.6459.390.6273.2446.270.63
Pixelate16×1649.8165.760.4238.3338.430.5025.5329.840.48
Blur κ=13σ=10 (weak)76.2467.400.5444.5937.730.5372.7535.690.69
Blur κ=21σ=10 (strong)58.8065.920.4630.3831.840.4957.7734.210.62
Ours87.1151.930.68+0.1188.675.460.92+0.2986.7413.190.87+0.18
RESULTS ACROSS SOTA OBFUSCATION METHODS
BDQ [34]81.0059.000.6191.117.150.9284.0434.180.75
ALF [59]76.0065.000.5685.8919.270.8382.0048.000.67
ELR[44] s=1670.8264.320.5391.2288.860.5196.2782.970.57
ELR[44] s=3252.9663.290.4585.5782.560.5292.4264.890.64
ELR[44] s=6431.6362.700.3456.2158.350.4980.0543.610.68
Ours †85.2551.670.67+0.0689.444.310.93+0.0184.0411.700.86+0.11
Ours87.1151.930.68+0.0788.675.460.92±0.0086.7413.190.87+0.12
", + "image_path": "85ce39d98ec7dd4b2e3fcc83c564d2e5f07d73f1e5eed4ca217001d86fe550a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 299, + 547, + 368 + ], + "lines": [ + { + "bbox": [ + 45, + 299, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 45, + 299, + 547, + 368 + ], + "type": "text", + "content": "Table 2. Comparison between Ours vs. Naive Baseline (top) and SOTA (bottom) Approaches as Top-1 Accuracy for both Action and Privacy labels, as well as " + }, + { + "bbox": [ + 45, + 299, + 547, + 368 + ], + "type": "inline_equation", + "content": "f_{0.5}" + }, + { + "bbox": [ + 45, + 299, + 547, + 368 + ], + "type": "text", + "content": " Introduced in Sec. 4.1. We show Ours and Naive Baseline results averaged across all recognition algorithms presented in Tab. 1. Competing SOTA approaches only present results on one specific algorithm for action recognition (I3D [7]) and privacy attribute detection (ResNet50 [22]) as those approaches are network specific. To showcase a fair comparison we also present results with the same single recognition algorithms indicated as \"Ours †\". First and second best results indicated by bold and underline, respectively. Relative performance delta (" + }, + { + "bbox": [ + 45, + 299, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 45, + 299, + 547, + 368 + ], + "type": "text", + "content": ") is indicated in green (improvement) and orange (tie) between best and second best method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "type": "text", + "content": "Note that we never perform any training on videos obfuscated by our approach. To implement warping, (4), we use pretrained RAFT to extract optical flow [52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": "Privacy template selection. Our approach affords selective combination of a predefined set of privacy templates for a given dataset or application. For the experiments, we choose a subset of templates, " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{T}}" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": ", from our complete identity preserving template library, " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\mathsf{T}" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": ", shown in Fig. 3, to optimize performance on a given dataset. Figure 5 shows template-wise performance based on both action recognition and privacy for each dataset. For SBU and KTH, action recognition is stable with respect to templates; however, the same four templates notably reduce privacy recognition. For consistency on IPN, we also select the four templates that most reduce privacy recognition, even while preserving action recognition; although, the distinction is less striking. This selection process results in templates " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{T}} = \\{\\text{torso, arm, leg, hair}\\}" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": " for SBU and KTH vs. " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathsf{T}} = \\{\\text{cheek, eyes, forehead, hair}\\}" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": " for IPN. Our code yields more results on the overlap of individual saliency maps. Runtime. The constant overhead to produce our obfuscated videos is " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\approx 100ms/frame" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": " on a NVIDIA RTX4080 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 105, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 105, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 105, + 658 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "We evaluate our approach to privacy preserving action recognition on 13 different action recognition models and five different privacy attribute recognition models. Competing approaches evaluate on at most five action models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": "[34, 43, 59]. Table 1 shows results on the original videos from IPN [3], SBU [60] and KTH [47], i.e. without any privacy preserving processing. Table 2 shows results comparing our approach vs. NaIVE BASELINES and SOTA. Comparison is given as the average across recognition algorithms presented in Table 1. Notably, while our approach compromises at most " + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "inline_equation", + "content": "1.48\\%" + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": " action recognition accuracy compared to performance on the original video (88.67% vs. " + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "inline_equation", + "content": "90.15\\%" + }, + { + "bbox": [ + 304, + 373, + 545, + 481 + ], + "type": "text", + "content": " on KTH), it always greatly improves privacy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "text", + "content": "Naive Baselines. Compared to the baselines, our approach scores highest with respect to the " + }, + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "inline_equation", + "content": "f_{0.5}" + }, + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "text", + "content": " metric across all three datasets: IPN (+0.11), KTH (+0.29), and SBU (+0.18). Naive BASELINES show that action recognition correlates highly with privacy performance, as " + }, + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "inline_equation", + "content": "f_{0.5}" + }, + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "text", + "content": " hovers around 0.5. The notable exception is " + }, + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{Pix}_{4\\times 4}" + }, + { + "bbox": [ + 304, + 483, + 546, + 675 + ], + "type": "text", + "content": " that performs well on IPN and KTH, and is the second best Naive BASELINE on SBU. A notable benefit of all these approaches is that they do not require retraining of the recognition algorithms. Still, none have a selective ability to obfuscate only certain parts. State of the art. A noteworthy difference between our vs. the alternative SoTA approaches is that they only work with the recognition algorithms on which they were trained. While this fact disadvantages our approach, as it lacks such retraining, it still outperforms the competing methods on all datasets: IPN (+0.06), KTH (+0.01), and SBU (+0.11)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "It is valuable to consider obfuscation approaches in terms of the relative importance of action recognition vs. privacy preservation. Figure 7 compares results for the SoTA ap" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18736" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 284, + 157 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 284, + 157 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 284, + 157 + ], + "type": "image", + "image_path": "9c12f8d35a275018076c280720822421dc99213da6e5dccd32577da5169ea4e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 158, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 191 + ], + "type": "text", + "content": "Figure 7. Performance across all datasets with varying " + }, + { + "bbox": [ + 46, + 158, + 287, + 191 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 158, + 287, + 191 + ], + "type": "text", + "content": "; see (6). Values of " + }, + { + "bbox": [ + 46, + 158, + 287, + 191 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 158, + 287, + 191 + ], + "type": "text", + "content": " closer to 0 weigh action recognition higher, whereas values closer to 1 increase the importance of privacy preservation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "text", + "content": "proaches as that trade-off is varied in terms of " + }, + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "inline_equation", + "content": "f_{\\lambda}" + }, + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "text", + "content": ", (6). The sweep of " + }, + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "text", + "content": " shows that our obfuscation approach performs better than any competing approach across the entire range. Especially with increasing " + }, + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 201, + 286, + 261 + ], + "type": "text", + "content": ", the gap to other approaches widens as more emphasis is put on privacy preservation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 263, + 287, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 311 + ], + "type": "text", + "content": "Importance of selective obfuscation. Our approach is unique in its ability to selectively obfuscate particular regions within a video based on specific privacy attributes. This ability yields two benefits:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 313, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 373 + ], + "type": "text", + "content": "(i) Selectivity aids in interpretability. Each selected template results in a saliency map, (2), which allows for visual inspection of what information is being obscured. Figure 4 highlights this benefit as the heat maps reveal the degree of obfuscation to be applied on a template-by-template basis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "type": "text", + "content": "(ii) Individual privacy templates can be chosen and combined for best performance; see Fig. 5. Depending on the dataset, different privacy templates differently impact the performance of privacy and action recognition. Action recognition on SBU and KTH is relatively robust to privacy template selection; however, privacy preservation is best when a subset of templates (torso, leg, arm, hair) is selected and the rest remain unobfuscated. This fact derives directly from the saliency calculation, (1): If multiple individual saliency maps have small values, then other strong responses are scaled down in the combined final saliency map. Subsequently, this effect leads to less noise being applied in our selective obfuscation, (5), which can deteriorate privacy preservation. In contrast, action recognition on IPN is compromised if the hand template is selected (as expected with a gesture centric dataset), which documents that simply obfuscating the entire person leads to inferior action performance compared to selectively applied obfuscation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 688 + ], + "type": "text", + "content": "Importance of motion consistent noise. Action recognition is complicated as different datasets, and even individual actions within a dataset might require different recognition capabilities and can rely to different degrees on the modeling of motion [9]. This uncertainty on the role of motion in action recognition is exacerbated by the fact that different deep learning-based architectures are successful in capturing motion to varying degrees [32]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "To see the importance of our proposed motion consistent noise for privacy preservation while maintaining good" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 427, + 144 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 427, + 144 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 427, + 144 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 427, + 144 + ], + "type": "image", + "image_path": "081822a89c02b238f22cbf300ef27ae3793c55608a9d1f71827e13dd308fd9ec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "type": "text", + "content": "Figure 8. Comparing Temporally Consistent vs. " + }, + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "type": "inline_equation", + "content": "iid" + }, + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "type": "text", + "content": " Obfuscation for Action Recognition and Privacy Preservation. Action recognition performance decreases if noise is not temporally consistent." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 437, + 72, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 437, + 72, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 437, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 437, + 72, + 545, + 144 + ], + "type": "image", + "image_path": "0475d0648df86f4dd5d69147a4c7e55028cefd9a8e00c3ef7324b00ae78f784d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": "action recognition, we compare to another version of our pipeline that obfuscates video frames using independent, identically distributed (iid) noise; see Fig. 8 and project page for video results. For all datasets, action recognition performance decreases if the noise is " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "iid" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": ". This result is sensible, because the " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "iid" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " noise does not capture the temporal dynamics of the source video. In contrast, privacy preservation is robust to noise in either case. There is less impact on IPN action recognition, as the critical hand motion is never obfuscated, which underlines the importance of selective masking. Further insight on why motion consistent noise supports better action can be had through consideration of Fig. 6, where the ability of such noise to support optical flow estimation is illustrated." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 362, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 518 + ], + "type": "text", + "content": "Limitations. Inevitably, there will be a trade-off between action recognition and privacy, as relevant information may be shared. In our approach, that trade-off could happen because our temporally consistent noise maintains dynamic information important for action recognition; however, it also might support motion based identification, e.g. gait recognition. Current protocols in privacy preserving action recognition do not consider motion-based identification. However, gait and related motion-based measurements are weak biometrics [26]; so, favouring action recognition may be apt. Moreover, if it becomes a concern, then it may be possible to apply motion perturbations that impede personal identification while maintaining action recognition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 527, + 378, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 378, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 378, + 539 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "Our work highlights that it is not necessary to train action recognition and privacy networks in an adversarial fashion for effective obfuscation of privacy attributes while maintaining strong action recognition performance. We show that a system based on local privacy templates, deep features that capture template semantics and selective noise obfuscation that is animated with source video motion can uphold privacy without hindering action recognition. Our approach is unique compared to alternative recent approaches in terms of interpretability and independence from particular action and privacy recognition algorithms. Our nine manually chosen templates, in combination with our proposed obfuscation technique outperforms other state-of-the-art approaches across three different datasets." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18737" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. ECCVW, 2022. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Simon Baker, Daniel Scharstein, JP Lewis, Stefan Roth, Michael J Black, and Richard Szeliski. A database and evaluation methodology for optical flow. IJCV, 92:1-31, 2011. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "text", + "content": "[3] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. IPN Hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In ICPR, 2021. 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 258 + ], + "type": "text", + "content": "[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 316 + ], + "type": "text", + "content": "[5] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 362, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 287, + 395 + ], + "type": "text", + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 396, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 430 + ], + "type": "text", + "content": "[8] Jiawei Chen, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Semi-coupled two-stream fusion ConvNets for action recognition at extremely low resolutions. In WACV, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 474 + ], + "type": "text", + "content": "[9] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't I dance in the mall? Learning to mitigate scene bias in action recognition. NeurIPS, 32, 2019. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 476, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 508 + ], + "type": "text", + "content": "[10] Robert T. Collins, Alan J Lipton, and Takeo Kanade. Introduction to the special section on video surveillance. IEEE TPAMI, 22(8):745-746, 2000. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "text", + "content": "[11] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (SMID): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 577 + ], + "type": "text", + "content": "[12] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 612 + ], + "type": "text", + "content": "[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. SPAct: Self-supervised privacy preservation for action recognition. In CVPR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer. X3D: Expanding architectures for efficient video recognition. In CVPR, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "text", + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast networks for video recognition. In ICCV, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "text", + "content": "[18] Joseph Fioresi, Ishan Rajendrakumar Dave, and Mubarak Shah. Ted-SPAD: Temporal distinctiveness for self-supervised privacy-preservation for video anomaly detection. In ICCV, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[19] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, and Omer Ovenc. Scalable detection of offensive and non-compliant content/logo in product images. In WACV, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 284 + ], + "type": "text", + "content": "[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 318 + ], + "type": "text", + "content": "[21] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "text", + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[24] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. PrivHAR: Recognizing human actions from privacy-preserving lens. In ECCV, 2022. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "text", + "content": "[25] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In ECCV, 2022. 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 466, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 547, + 488 + ], + "type": "text", + "content": "[26] Anil K Jain, Patrick Flynn, and Arun A Ross. Handbook of Biometrics. Springer Science & Business Media, 2007. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 490, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 511 + ], + "type": "text", + "content": "[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLO by Ultralytics, 2023. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 513, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 544 + ], + "type": "text", + "content": "[28] Amin Jourabloo, Xi Yin, and Xiaoming Liu. Attribute preserved face de-identification. In International Conference on Biometrics, 2015. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "type": "text", + "content": "[29] Alexander John Karran, Théophile Demazure, Antoine Hudon, Sylvain Senecal, and Pierre-Majorique Léger. Designing for confidence: The impact of visualizing artificial intelligence decisions. Frontiers in Neuroscience, 16: 883385, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 602, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 655 + ], + "type": "text", + "content": "[30] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The Kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 658, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 679 + ], + "type": "text", + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal net" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18738" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "works encode: Quantifying static vs. dynamic information. In CVPR, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[33] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "text", + "content": "[34] Sudhakar Kumawat and Hajime Nagahara. Privacy-preserving action recognition via motion difference quantization. In ECCV, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 161, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 216 + ], + "type": "text", + "content": "[35] Agostina J Larrazabal, Nicolas Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592–12594, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[36] Ming Li, Xiangyu Xu, Hehe Fan, Pan Zhou, Jun Liu, Jia-Wei Liu, Jiahe Li, Jussi Keppo, Mike Zheng Shou, and Shuicheng Yan. STPrivacy: Spatio-temporal privacy-preserving action recognition. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[37] Bo Liu, Ming Ding, Sina Shaham, Wenny Rahayu, Farhad Farokhi, and Zihuai Lin. When machine learning meets privacy: A survey and outlook. ACM Computing Surveys, 54 (2):1-36, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": "[38] Haochen Liu, Yiqi Wang, Wenqi Fan, Xiaorui Liu, Yaxin Li, Shaili Jain, Yunhao Liu, Anil Jain, and Jiliang Tang. Trustworthy AI: A computational perspective. ACM Transactions on Intelligent Systems and Technology, 14(1):1-59, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 370 + ], + "type": "text", + "content": "[39] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91–110, 2004. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "text", + "content": "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 458 + ], + "type": "text", + "content": "[41] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "type": "text", + "content": "[42] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action detection. In ECCV, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "type": "text", + "content": "[43] Michael Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In AAAI, 2017. 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[44] Michael Ryoo, Kiyoon Kim, and Hyun Yang. Extreme low resolution activity recognition with multi-siamese embedding learning. In AAAI, 2018. 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[46] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "text", + "content": "[47] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a local SVM approach. In ICPR, 2004. 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[48] Vijeta Sharma, Manjari Gupta, Anil Pandey, Deepti Mishra, and Anil Kumar. A review of deep-learning-based human activity recognition on benchmark video datasets. Applied Artificial Intelligence, 36(1):2093705, 2022. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 501 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[49] Jianbo Shi and Carlo Tomasi. Good features to track. In CVPR, 1994. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "text", + "content": "[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 27, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 129, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 152 + ], + "type": "text", + "content": "[51] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "text", + "content": "[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "text", + "content": "[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "text", + "content": "[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 254, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 286 + ], + "type": "text", + "content": "[55] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 287, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 320 + ], + "type": "text", + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 322, + 545, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 344 + ], + "type": "text", + "content": "[57] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local Neural Networks. In CVPR, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 345, + 545, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 345, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 307, + 345, + 545, + 378 + ], + "type": "text", + "content": "[58] Zhenyu Wu, Zhangyang Wang, Zhaowen Wang, and Hailin Jin. Towards privacy-preserving visual recognition via adversarial training: A pilot study. In ECCV, 2018. 1, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 379, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 379, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 379, + 545, + 422 + ], + "type": "text", + "content": "[59] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE TPAMI, 44(4):2126-2139, 2020. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 423, + 545, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 467 + ], + "type": "text", + "content": "[60] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In CVPR, 2012. 3, 5, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 468, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 501 + ], + "type": "text", + "content": "[61] Yingying Zhu, Nandita M Nayak, and Amit K Roy-Chowdhury. Context-aware modeling and recognition of activities in video. In CVPR, 2013. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18739" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_content_list.json b/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..68e8098167eb0bd999a195dc47327a156bb748bb --- /dev/null +++ b/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_content_list.json @@ -0,0 +1,1494 @@ +[ + { + "type": "text", + "text": "Selective nonlinearities removal from digital signals", + "text_level": 1, + "bbox": [ + 223, + 130, + 746, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Krzysztof A. Maliszewski \nSchool of Mathematics and Statistics, University of Canterbury \nChristchurch, New Zealand \nkrzysztof.maliszewski@pg.canterbury.ac.nz", + "bbox": [ + 233, + 181, + 733, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Magdalena A. Urbańska \nSchool of Food and Advanced Technology, Massey University Palmerston North, New Zealand", + "bbox": [ + 238, + 258, + 730, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "murbanska@massey.ac.nz", + "bbox": [ + 383, + 316, + 583, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Varvara Vetrova \nSchool of Mathematics and Statistics, University of Canterbury \nChristchurch, New Zealand", + "bbox": [ + 233, + 337, + 733, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "varvara.vetrova@canterbury.ac.nz", + "bbox": [ + 341, + 393, + 625, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sylwia M. Kolenderska \nInstitute of Physics, Faculty of Physics, Astronomy and Informatics \nNicolaus Copernicus University in Torun \nul. Grudziadzka 5, 87-100 Torun, Poland", + "bbox": [ + 215, + 415, + 753, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Physical and Chemical Sciences, University of Canterbury", + "bbox": [ + 212, + 486, + 754, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christchurch, New Zealand", + "bbox": [ + 372, + 503, + 593, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "skol745@aucklanduni.ac.nz", + "bbox": [ + 370, + 523, + 594, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 573, + 313, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many instruments performing optical and non-optical imaging and sensing, such as Optical Coherence Tomography (OCT), Magnetic Resonance Imaging or Fourier-transform spectrometry, produce digital signals containing modulations, sine-like components, which only after Fourier transformation give information about the structure or characteristics of the investigated object. Due to the fundamental physics-related limitations of such methods, the distribution of these signal components is often nonlinear and, when not properly compensated, leads to the resolution, precision or quality drop in the final image. Here, we propose an innovative approach that has the potential to allow cleaning of the signal from the nonlinearities but most of all, it now allows to switch the given order off, leaving all others intact. The latter provides a tool for more in-depth analysis of the nonlinearity-inducing properties of the investigated object, which can lead to applications in early disease detection or more sensitive sensing of chemical compounds. We consider OCT signals and nonlinear", + "bbox": [ + 76, + 613, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ities up to the third order. In our approach, we propose two neural networks: one to remove solely the second-order nonlinearity and the other for removing solely the third-order nonlinearity. The input of the networks is a novel two-dimensional data structure with all the information needed for the network to infer a nonlinearity-free signal. We describe the developed networks and present the results for second-order and third-order nonlinearity removal in OCT data representing the images of various objects: a mirror, glass, and fruits.", + "bbox": [ + 496, + 574, + 893, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 753, + 632, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Optical Coherence Tomography (OCT), Magnetic Resonance Imaging (MRI) or Fourier-transform Spectrometry (FTS) belong to a class of methods where useful, interpretable signal is obtained in an indirect way, more specifically by Fourier transforming the raw signal acquired by the instrument and taking its absolute value. On the one hand, such an indirect approach makes OCT-based imaging and FTS-based sensing faster and enables MRI, but on", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25028", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/048039b52eca5292a5bd2bb86f3c0b5b328a9d8ae45e104c6bbef271fe54bd27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 88, + 297, + 165 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/be826ae52a302df15bba60a416d8760e8e10906ce0eb7c235ab1aa530d22e642.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 88, + 436, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/994652ee43083019d4be515be09da064c3aa67f81ac026118a4e08fbf465a239.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 167, + 297, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e654a35bc7c766314fad7c929a8f446db8507bac22b1e21bff53076cc67193de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 167, + 434, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c5146151ad4384e5a1760f5c5636843a608be79cfd3c5da55fd065bb8c8129f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 252, + 297, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/86493d112cadd76d5ffdf22b70498c898eac0adc23229436d8ea0281a32b32c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 252, + 434, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2e5f6d1da725eacad35c8bddf6c4afb89b654619651969ead4300c7356acfa2.jpg", + "image_caption": [ + "Figure 1. Effects of signal nonlinearity on the signal's Fourier transform. (a) A signal with a linear modulation distribution Fourier transform to (b) a narrow peak. (c) If the nonlinearity of the modulation is of the second order, (d) the peak gets broadened. (e) If the nonlinearity is of the third order, (f) the peak gets distorted. In OCT, this will result in deteriorated image quality. (g) In practice, both nonlinearities appear at once, resulting in (h) the peak being elongated in the direction of the third-order distortion." + ], + "image_footnote": [], + "bbox": [ + 107, + 335, + 297, + 412 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9b34c1c2f3bda77e01e581f952567a4d157e3a5c92bf230b8d329d074952c464.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 335, + 434, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the other hand, it makes room for errors related to the limitations of Fourier transformation. One such limitation is the Fourier transformation's sensitivity to nonlinearities in the input signal. In the simplest case of a signal being a single-frequency modulation, a second-order nonlinear distribution will Fourier transform to a broader peak (compare Fig. 1a,b with c,d). If this distribution is of a third-order nonlinear character, the peak in the Fourier transform will get distorted (Fig. 1e,f). In practice, both nonlinearities appear at once, the second-order one bigger than the third-order one, leading to a broad peak elongated in the direction of the third-order nonlinearity distortion (Fig. 1g,h).", + "bbox": [ + 75, + 551, + 467, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The influence of nonlinearities is best visible in OCT, where the Fourier transform of the raw signal constitutes one line of the image. Since the width of the peak represents the depth resolution (i.e. how small a detail can be observed), and the peak's shape correlates with the image quality, nonlinearities are generally thought of as detrimental and various techniques are devised to remove them [1, 2]. These techniques do not achieve absolute removal of nonlinearity, preventing single-shot imaging of bulk objects, such as the entire eye [3, 4] and consequently necessitating elaborate workaround solutions [5-8]. A Machine", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning solution for nonlinearity removal was proposed as well [9] – although robust and reliable, it is biased towards specific kinds of objects with specific nonlinear properties.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, the inherent optical properties that produce nonlinearities in the imaged objects have been shown to enable the characterisation of the objects [10-12] and correlate the change of their nonlinear character with the progression of diseases [13]. Unfortunately, the current methods for extracting the second-order nonlinearity are very object-specific as well as error-prone preventing them from being applied in real-world scenarios.", + "bbox": [ + 496, + 137, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present an approach that enables flexible and comfortable management of the signal nonlinearities, letting the user study one order in the absence of the others, a capability previously impossible to achieve using current standard or Machine Learning methods. In our approach, for each nonlinearity order, a neural network is trained whose input is a special two-dimensional data structure, a stack. Each row of the stack represents an amplitude of the Fourier transform of an input signal multiplied by a different nonlinearity-correcting factor. Consequently, it contains nonlinearity-free regions which the network identifies and uses to build a nonlinearity-free Fourier transform amplitude. We present two networks, one to remove the second-order nonlinearity while keeping the third-order one intact, and the second network to remove the third-order nonlinearity while keeping the second-order one intact. We test our approach on OCT signals that naturally contain substantial amounts of second-order nonlinearity and small amounts of third-order nonlinearity. The tests on both computer-generated and experimental signals indicate the robustness of our approach.", + "bbox": [ + 496, + 258, + 892, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start by describing our approach in Section 2, including the creation of the stacks in Subsection 2.1, their application in the training processes in Subsection 2.2, presentation of the used architecture and methodology in Subsection 2.3, and the description of the datasets (Subsection 2.4) used for training of our models. In Section 3, we report the performance of the networks in removing the respective nonlinearity, complete with the analysis of the training process and the subsequent tests on computer-generated data (Subsection 3.1) and experimental data (Subsections 3.2, 3.3 and 3.4). In the final sections, we summarise the obtained results (Section 4), pointing to the strong points of our approach, identifying the observed and potential shortcomings, and discussing potential solutions and the scope of future work (Section 5).", + "bbox": [ + 496, + 577, + 890, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Concept", + "text_level": 1, + "bbox": [ + 500, + 820, + 594, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Amplitude stacks as inputs", + "text_level": 1, + "bbox": [ + 500, + 845, + 741, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our networks rely on a two-dimensional data structure as inputs: a stack (Fig. 2). The rows of a stack are created", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25029", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a16dbcc707930011ceab6b8624b3f09ee051c45d8fa50269d9bbf99f0758b4ab.jpg", + "image_caption": [ + "(a)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 127, + 89, + 431, + 155 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c1c30731d0abafa704dbb5a36e90f55a1ff49d6b1a8911bb320c5c3b121e37f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 83, + 178, + 233, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/97def600b40e5ba6e9df3cd0024dc57242b310ca75f8f7bf4948020d928c7f75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 161, + 460, + 295 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac0c70b4f65a46e82a7e214d19f98664a95e42d451f10d7b5444245b8ed4d0f9.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 81, + 313, + 233, + 429 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab7a219ead3eadb8a1081c5f6d5f309fb6fea3665d9b981175de4c8830e8fad7.jpg", + "image_caption": [ + "(e)", + "Figure 2. Calculation of the input stacks. (a) The raw signal is (b) multiplied by 32 different complex exponents representing 32 different levels of the second-order nonlinearity: $C_0, C_1, C_2, \\ldots, C_{31}$ . Each row is then Fourier transformed. (c) The 2nd-order stack is created from the absolute value of the complex Fourier transforms. Similarly, (d) the raw signal is multiplied by 32 different complex exponents representing 32 different levels of the third-order nonlinearity: $D_0, D_1, D_2, \\ldots, D_{31}$ and then Fourier transformed to build (e) a 3rd-order stack." + ], + "image_footnote": [], + "bbox": [ + 264, + 303, + 460, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the following way. First, the raw signal is multiplied (Fig. 2a) by a complex exponent with a different quadratic argument in the case of the stack used for removing the second-order nonlinearity (Fig. 2b) or with a different cubic argument in the case of the stack for the third-order nonlinearity removal (Fig. 2d). Coefficients in the quadratic arguments, $(C_0,\\dots,C_{31}$ in Fig. 2b), and coefficients in the cubic arguments $(D_0,\\dots,D_{31}$ in Fig. 2d), are chosen to cover a wide range of the second- and third-order nonlinearity, ensuring that the nonlinearities present in the signal are fully compensated. Then, each row is Fourier transformed, and an absolute value is calculated to obtain a 2nd-order stack (Fig. 2c) or a 3rd-order stack (Fig. 2e). The range of the nonlinear coefficients is symmetrical around 0 and broad enough to ensure that each modulation's nonlinearity in the raw signal is compensated. In the 2nd-order stack, these will be the places with the smallest width (marked 1, 2, 3 and 4 in Fig. 3b) and in the 3rd-order stack, the places containing no distortions (marked 1, 2, 3 and 4 in Fig. 3c).", + "bbox": [ + 75, + 613, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f782900b0485a1b983e5c1474a319e8adffe2eef9e06b102fec9a952f84f0124.jpg", + "image_caption": [ + "Figure 3. (a) The Fourier transform amplitude of a raw signal, presented in Fig. 2a, shows four peaks affected by a different amount and type of nonlinearity. The second-order nonlinearity is removed with Network #1 trained on (b) the 2nd-order stack as input and (d) Fourier transform amplitude free of the second-order nonlinearity as output. Similarly, the third-order nonlinearity is removed with Network #2 trained on (b) the 3rd-order stack as input and (d) Fourier transform amplitude free of the third-order nonlinearity as output. As a result, (f) the second-order nonlinearity is removed leaving the third-order one (see peak 4) using Network #1, (g) the third-order nonlinearity is removed leaving the second-order one intact (see peak 4) using Network #2." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Nonlinearity-removing networks", + "text_level": 1, + "bbox": [ + 500, + 676, + 785, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The approach for removing the second-order nonlinearity and the approach for removing the third-order nonlinearity are similar, presented schematically in Fig. 3. The input of the network for the second-order nonlinearity removal, Network #1, is the 2nd-order stack (Fig. 3b) and the output is a Fourier transform amplitude free of second-order nonlinearity (Fig. 3d). Because the input 2nd-order stack contains regions corresponding to perfectly removed second-order nonlinearity, the training process can be viewed as learning how to identify these regions and then stitching them all up. Similarly, the network for removing third-order nonlinearity, Network #2, takes the 3rd-order stack as input (Fig. 3c) and outputs a Fourier transform amplitude free", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25030", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c182229d79c02093451922524225c06d0d70644057fd459307819d87e3aecfa5.jpg", + "image_caption": [ + "Figure 4. Architecture of our nonlinearity-removing networks." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 460, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the 3rd-order nonlinearity (Fig. 3e). Input for both networks has a shape of 32 by 1024 pixels with a single channel whereas output is a vector of size 1024. When compared to the Fourier transform amplitude of the raw signal (Fig. 3a), the signal returned by Network #1 (Fig. 3f) incorporates only the third-order nonlinearity, and the signal returned by Network #2 (Fig. 3g) contains only the second-order nonlinearity.", + "bbox": [ + 75, + 335, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 467, + 215, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our nonlinearity-removing networks are inspired by the U-Net architecture [14]. U-Net stands out for its use of a contracting path to capture contextual information, an expanding path for precise feature localisation, and the incorporation of skip connections that preserve fine-grained details, particularly beneficial for precise object delineation. We leverage these capabilities to identify and extract nonlinearity-related features from the stacks.", + "bbox": [ + 75, + 489, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the classical U-Net deals with 2D data structures for both input and output, in our case, we perform dimension reduction to output 1D signals by incorporating a max-pooling layer. We further refined the architecture by optimising hyperparameters with the validation dataset using Optuna [15]. As a result of the tuning, we adopted the Mean Absolute Error (MAE) loss function and the Adam optimiser [16] with a set learning rate of 0.0002. Our final network architecture is presented in Fig. 4.", + "bbox": [ + 75, + 611, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Datasets", + "text_level": 1, + "bbox": [ + 76, + 756, + 181, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our datasets consist of noiseless computer-generated signals that are Gaussian-shaped, and diverse and randomised in terms of their content. If put in the OCT-imaging-related terms, they represent a broad range of possible imaged object configurations. These objects are simulated to have a random number of interfaces (from 2 to 12) positioned at arbitrary locations, with each object layer exhibiting varying random reflectivity and second- and third-order nonlin", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ba29dd818ac660451fabe13244685a6b03cf21effed3cba4580072f9d54318b.jpg", + "image_caption": [ + "Figure 5. Network #1 and Network #2 performance measured with the Mean Absolute Error (MAE) loss function across 30 epochs of training. Blue lines depict Network #1, while orange lines showcase the Network #2 performance. Dotted vertical lines highlight epochs exhibiting the best performance." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "earity values. These object parameters are used to synthesise modulatory signals equivalent to those acquired in an OCT measurement (see example signals in Fig. 1a, c, e, g or Fig. 2a). Next, the input stacks and outputs, the Fourier transform amplitudes of signals free of the respective nonlinearity, are calculated.", + "bbox": [ + 496, + 356, + 893, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The risk of overfitting was minimised by using the training datasets with an optimised size [17]. In our research, this was found to be between 160,000 and 240,000 objects. The validation and test datasets include 14,993 objects. The distribution is uniform across the objects with a specific number of interfaces, resulting in an equal allocation of 1,363 data samples for each distinct object type.", + "bbox": [ + 496, + 446, + 893, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before starting the training, as part of the preprocessing stage, both the input and output data were normalised through Min-Max scaling, ensuring that values fell within the standardised range of 0 to 1.", + "bbox": [ + 496, + 553, + 893, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Performance", + "text_level": 1, + "bbox": [ + 500, + 626, + 633, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. In silico data tests", + "text_level": 1, + "bbox": [ + 500, + 650, + 673, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each nonlinearity-removing neural network underwent training using a training dataset consisting of 200,000 objects. Within each mini-batch, 8 input stacks were processed. The proposed algorithm requires 39 ms to process a stack on a Quadro RTX 6000 GPU. Network #1 (Fig. 5, blue lines) and Network #2 (Fig. 5, orange lines) both displayed convergence within just a few epochs (Fig. 5) and were trained until the thirtieth epoch.", + "bbox": [ + 496, + 674, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better evaluate the networks and results, we employed the Goodness-of-Fit Approximity (GoF) metric introduced in [12]. The GoF metric assesses the accuracy of predictions by calculating the percentage of values that lie within a predefined distance from the corresponding ground truth values. For this GoF metric, we define an error threshold as a ratio between the predefined distance and the value range", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25031", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b01f51ee0a3c226f49bad1848ea6588ba9e68bd73cf0c597b6880e5ba54f49d.jpg", + "image_caption": [ + "Figure 6. Networks performance assessed through the metric of GoF, measuring the degree to which predictions align with the ground truth within $0.1\\%$ error in the test dataset comprising 14,993 data samples for (a) Network #1 and (b) Network #2. The graphs at the top (a1,a2) are zoomed-in areas of the top area of (b1,b2) the whole graphs presented." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 468, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "span. In our case, where the values range from 0 to 1, the distance of 0.01 means $1\\%$ error threshold.", + "bbox": [ + 75, + 440, + 468, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the post-training evaluation using a $1\\%$ error threshold and the test dataset, Network #1 started to overfit at epoch 25 (Fig. 5, indicated by the blue dotted vertical line). In contrast, Network #2 demonstrated resilience against overfitting and achieved its best results in the 28th epoch (Fig. 5, marked by the red dotted vertical line).", + "bbox": [ + 75, + 472, + 468, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the test dataset and GoF with a $1\\%$ error threshold, both networks achieved a remarkable mean GoF value of $99.98\\%$ , which means that $99.98\\%$ of the predictions differed from the ground truth by less than $1\\%$ , with only one example below $95\\%$ GoF. In light of this, to be able to carry out a meaningful analysis, the error threshold was further reduced to $0.1\\%$ and a boxplot graph was plotted for each model in Fig. 6.", + "bbox": [ + 75, + 565, + 470, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both networks, Network #1 (Fig. 6a) and Network #2 (Fig. 6b), demonstrated very good performance even for error threshold $0.1\\%$ . Only approximately $9.52\\%$ and $3.07\\%$ of the predictions from Network #1 and Network #2, respectively, fell below the $95\\%$ GoF for such a small error, as showed in Tab. 1 that also includes GoF values for objects with distinct number of layers. Fig. 7 offers an in-depth analysis of the worst-predicted examples for the Network #1 (Fig. 7a) and the Network #2 (Fig. 7b) showcasing GoF values of $65.5\\%$ and $45.5\\%$ , respectively. Predictions of peak amplitudes in the nonlinearity-free Fourier transform amplitude (Fig. 7(3), orange line) closely align the expected outcomes (blue line) deviating only by the error threshold level. Also, the networks predict values close to", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4bb2ede470abb4d7063722526af6ed5e95e5fd76a7ad78f7b5a87bbcc63ddd95.jpg", + "image_caption": [ + "Figure 7. Predictions with very low GoF for $0.1\\%$ error threshold for (a) Network #1 and (b) Network #2. Based on the input stack (1), models predict (2, orange line, PR) amplitudes free of the respective nonlinearity that exhibit a strong correlation with the ground truth (2, blue line, GT). (3) Nonlinearity-free peaks are recovered with high fidelity, but (4) the models sometimes struggle with zero-valued places." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 883, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "zero rather than precisely zero, contributing to worse GoF metric results. One should keep in mind here that these discrepancies are extremely low: at the level of around 2e-3 as presented in Fig. 7a4 and b4.", + "bbox": [ + 496, + 595, + 893, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Experimental signals: mirror at different depths", + "text_level": 1, + "bbox": [ + 498, + 665, + 893, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Networks were tested on the signals obtained through an OCT imaging of a mirror placed at 11 different depth positions. Each acquired signal Fourier transforms to a peak whose amount of second-order-nonlinearity-related broadening and third-order-nonlinearity-related distortion increases with depth (Fig. 8a). Due to their simultaneous appearance, the third-order nonlinearity manifests itself in the form of peak elongation. The depth-dependent deterioration is due to the fundamental limitations of OCT's spectrometer-based light detection. Additionally, the second- and third-order nonlinearities originating from the OCT's interferometer add up - depth-independently - to that effect, further worsening the outcome. Typically,", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25032", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cd91d46d5b548845b89f48297f4cb25440891e888f5ff2eb1a99a27148859cd9.jpg", + "table_caption": [ + "Table 1. The number of cases where GoF with error threshold ${0.1}\\%$ falls below ${95}\\%$ for objects with a specific number of reflective layers." + ], + "table_footnote": [], + "table_body": "
Net.Interfaces23456789101112Totals
#1# below 95%3842601931381058779554544371381
# in dataset1363136313631363136313631363136313631363136314993
% below28.1719.0814.1610.127.706.385.804.043.303.322.719.52
#2# below 95%10782575438323122131311460
# in dataset1363136313631363136313631363136313631363136314993
% below7.856.024.183.962.792.352.271.610.950.950.813.07
", + "bbox": [ + 158, + 114, + 805, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f1e9e381f89c8ff7b700fe887e4032e6521a8952802db1eda91af5e979aae43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 217, + 395, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d0ee500d14f3736bdb0d46ea3ec54d1ba04ec0b079b13b39a99eb207f80b209.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 299, + 398, + 380 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2620ec16a20ec7e8d3b0ad08f3edc044a095f704114d0f43f055873e201fabcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 382, + 398, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e928aad4622d3ffb33a2b0c9c743c1e900eeab944af0122a2bb452dd9a9b981e.jpg", + "image_caption": [ + "Figure 8. Mirror at 11 different depths. (a) Fourier transforming raw signals results in peaks that are both broadened by the second-order nonlinearity and distorted by the third-order nonlinearity. The deeper the peak, the bigger the nonlinearity effect, a behaviour characteristic of OCT signals acquired with a spectrometer. (b) Network #1 removes the second-order nonlinearity, leaving the third-order one intact. (c) Network #2 removes the third-order nonlinearity, leaving the third-order one intact. (d, e) The traditional linearisation (grey line) is not as precise as our approach (green lines)." + ], + "image_footnote": [], + "bbox": [ + 109, + 484, + 276, + 580 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ab486d8fc1f593adf72fef4b6ad47350d57db4cb81d0209bbc379694d983211.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 484, + 437, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "spectrometer- and interferometer-related nonlinearities can be satisfactorily removed in a calibration step before the measurement.", + "bbox": [ + 75, + 744, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The application of Network #1 removes the second-order nonlinearity in its entirety and consequently, provides a dramatic reduction of peaks' widths (Fig. 8b). Because the third-order nonlinearity remains intact, the peaks' shapes incorporate the characteristic \"tail\". On the other hand, Network #2 removes the third-order nonlinearity, leading to a complete reduction of the elongation (Fig. 8c).", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One observes the superiority of our approach when the peaks obtained with Network #1 are compared to the peaks obtained as a result of traditional linearisation methods which remove all nonlinearity orders. Here, we used the method in [1] where two raw signals corresponding to the mirror placed at two different locations are used. First, their phases are subtracted to remove the constant interferometer-related contribution, enabling the calculation of the correction vector for spectrometer-related nonlinearities. Once the spectrometer's nonlinearity is removed, one of the initial spectra is used to generate an interferometer-nonlinearities-compensating vector. The peaks located close to 0 depth index obtained using our approach and the standard linearisation method (Fig. 8d, green and grey lines) are similar in shape: the network-outputted peak is higher which is very favourable - it means a higher signal-to-noise ratio, which is associated with better quality imaging. One observes a small \"tail\" at the side, most probably due to a small amount of third-order nonlinearity present at such shallow depths, whereas the peak obtained using the traditional methods is a smooth Gaussian. More interestingly, for a peak at a substantial depth (Fig. 8e), where the third-order nonlinearity is naturally much bigger (let us remind that this depth-increasing nonlinearity is tied to OCT spectrometer), the \"tail\" is much more pronounced in case of our network but not present in case of the traditional method. In the latter case, the peak is elongated which suggests the presence of both second- and third-order nonlinearities. This leads to the conclusion that traditional linearisation, although very satisfactory in practical applications, still has room for improvement. In fact, despite 30 years of OCT in science and industry, new methods for more precise OCT data linearisation are still being developed.", + "bbox": [ + 496, + 218, + 893, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Glass", + "text_level": 1, + "bbox": [ + 500, + 726, + 578, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another test was performed on data corresponding to a $300\\mu \\mathrm{m}$ thick sapphire glass put at an angle obtained with the same instrument. This data consisted of 256 raw OCT signals which represent a different transversal location on the glass. When Fourier transformed one after another, they form a 2D depth image of the object (Fig. 9a, 1 and 2 denote the front and back surface of the glass, the third line is an artefact placed at the depth equal to the thickness of the glass, inherent to OCT). Without linearisation, the image elements are broadened the deeper they are. When the", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25033", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db05d6db814cd63f778520fd8cbf79e2753d3aca766d02d665146c796554e892.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 90, + 367, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29b29f391d403a08ee4c1fb4296adb274529a948df5f08f0247909f1669380e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 195, + 272, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c3bc75914cc1ad76b547fa132501ec56d85bbcd9eab0f85bdfb679050fd43b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 194, + 444, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b06b3c978a03e9edc2508a382adca684a7507d4598e03555fd20d3dc58c5f8d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 296, + 271, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5832df6ccea4e3ea07b2dd0e74e603d312d1e31cd18c06eb847279c016401075.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 296, + 444, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d199ff707b3a1e13fa742a2a2e0da1df1dd4a3f1bd2c4dba222464fb32f3630b.jpg", + "image_caption": [ + "Figure 9. Sapphire glass. (a) Fourier transformation of raw signals gives an OCT image where the quality deteriorates with depth. (b) Network #1 removes the second-order nonlinearity which accounts for the most of quality deterioration, (c) Due to a small amount of the third-order nonlinearity, the application of Network #2 does not improve the image quality. (d), Image line #152 from Network #1 (green line) is compared with its raw image equivalent (grey line) to show dramatic peaks' width reduction, (e) The image line #152 from Network #2 output (green line) and its raw image equivalent (grey line) show no significant difference except for a more favourable shape for the former. (f) Network #1 (green line) performs quite well when compared to the traditional method output (black line). 1 and 2 mark the front and back surface of the imaged glass, the middle line is an artefact." + ], + "image_footnote": [], + "bbox": [ + 150, + 397, + 390, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "same data is processed with Network #1, the broadening is removed from all the image elements (see the image in Fig. 9b and the 152nd image line in Fig. 9d). Due to very small amounts of the third-order nonlinearity, Network #2 does not bring perceptible improvements (Fig. 9c), but under a closer inspection of a single image line (Fig. 9e), one can observe some improvement in the peak height and a possible side elongation reduction.", + "bbox": [ + 75, + 703, + 467, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 9f compares the output of Network #1 (green line) and the traditional linearisation method (black line), indicating that this model could be a satisfactory alternative for the traditional image calculation algorithm when the third-order nonlinearities are negligible. It can be used as such an alter", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "native for the OCT laboratory instrument which was used here to obtain the data and which is nevertheless a fairly good representative of a standard OCT instrument. We note that Network #1 improved the shape of the artefact (the middle peak in Fig. 9f), which - although generally irrelevant for OCT imaging - is still something noteworthy as such a feat is beyond the capabilities of traditional linearisation methods, showing that our approach, unlike the traditional ones, removes the nonlinearities nondiscriminatively.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Grape and cucumber", + "text_level": 1, + "bbox": [ + 500, + 234, + 699, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The final tests were performed on the data corresponding to a grape (Fig. 10) and a cucumber (Fig. 11). The former represents a high signal-to-noise ratio (highly visible modulation in the signal as depicted in Fig. 10a) and the latter a low signal-to-noise ratio (poorly visible modulation in the signal as depicted in Fig. 11a).", + "bbox": [ + 496, + 258, + 890, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fourier transformation of raw signals gives images of a low quality that further deteriorates with depth (Fig. 10b, Fig. 11b). Again, the improvement obtained when applying Network #1 (Fig. 10c, Fig. 11c) is comparable to that obtained with traditional methods (Fig. 10d, Fig. 11d). Under closer inspection (Fig. 10e, Fig. 11e), one notices that Network #1 output is less detailed for the higher signal-to-noise grape data - most probably due to the remaining third-order nonlinearities - and loses some details for the lower signal-to-noise ratio cucumber data. The latter loss is the direct consequence of training data representing predominantly high signal-to-noise ratio signals.", + "bbox": [ + 496, + 349, + 890, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Just as in the case of the glass data, due to small amounts of third-order nonlinearity, Network #2 did not provide visually significant improvement and therefore, was omitted.", + "bbox": [ + 496, + 530, + 890, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Summary and discussion", + "text_level": 1, + "bbox": [ + 500, + 588, + 732, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two models were demonstrated: one for removing second-order nonlinearity from digital signals and leaving the third-order one intact, and another one for removing the third-order nonlinearity but leaving the second-order one intact. Their nonlinearity-removing capabilities were extensively analysed using computer-generated signals as well as experimental signals acquired with a laboratory Optical Coherence Tomography (OCT) device. Tests on the OCT signals, which corresponded to objects with varying levels of complication: mirror, sapphire, grape and cucumber, allowed us to evaluate the networks in terms of their applicability in a world-relevant technology. Whereas the tests and analysis done on the computer-generated signals revealed extraordinary performance in terms of similarity of the ground truth and prediction signals, the tests on OCT-acquired data indicated the networks' advantageous operation when compared to approaches traditionally used in OCT.", + "bbox": [ + 496, + 613, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, both networks converged exceptionally fast and reached remarkable loss function values. Nevertheless", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25034", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e91691d4c6ee5a8513471026196bc126a65485cb245696d7a3e87b0a4c6aa30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 90, + 433, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11cef7ae94bbb341424bf00dabf5797ce05db10ee810bf72a9a19dd6b73afea2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 167, + 444, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff57d1cd76565e64994e4471ce3345eff28d89eb9e775d37f069acc6a2163995.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 253, + 444, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfb73ce259b04d136b209571ff4afd7e95e70ca39b85bd5278c4fcc5ea31b04b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 340, + 444, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fc46eb2c470196332d40602656d8a40560b948ae526f2e6ee8b773941028d93.jpg", + "image_caption": [ + "Figure 10. Grape. (a) One of the raw signals representing a single line in the image of a grape. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image when compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees very well with its equivalent from the image obtained using traditional methods." + ], + "image_footnote": [], + "bbox": [ + 101, + 449, + 434, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "less, we have identified possible improvements: the closer inspection of the predictions revealed non-zero values that are zero-valued areas in the ground-truth signals. Another current shortcoming of both networks is their time of processing an entire OCT image; consequently, for now, the processing speed is insufficient for the real-time processing of data coming straight from OCT devices.", + "bbox": [ + 75, + 646, + 468, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Future work", + "text_level": 1, + "bbox": [ + 76, + 768, + 207, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The future work will mainly consist in developing an approach that enables one to use both networks in a sequence so that both the second- and third-order nonlinearity removal is achieved. Since, as was shown here, the networks remove the nonlinearities completely, their sequential use will provide a very valuable tool for extremely precise data linearisation, especially attractive for OCT where any un", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62313f48081c7d2fb57f479ede3d9b1be037e21aee669caf260aace6251bb469.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 93, + 864, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f91b7b15067482bb19a0dc332343e88654fea8d32a1418a4276c70135d4a55c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 169, + 870, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/108a614d87b771b1ecca5c1e048a8402bddcf6819fbad750134be4c6db4038a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 255, + 870, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a122f416a3da4332676b271f026a756ad0aae55a4e44d28b77efac3cdaeacc00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 342, + 867, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37755de439c2620b9482844e69b68ef5b17823c6d4c1ec5ddc693267d4de479d.jpg", + "image_caption": [ + "Figure 11. Cucumber. (a) One of the raw signals representing a single line in the image of a cucumber. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees well with its equivalent from the image obtained using traditional methods." + ], + "image_footnote": [], + "bbox": [ + 524, + 450, + 861, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compensated nonlinearities lead to the quality drop.", + "bbox": [ + 498, + 645, + 841, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also, in terms of the network performance, we will try to remove the discrepancy for zero-valued elements in the ground truth and prediction signals by adding extra convolutional layers before the output to filter out the unwanted values as well as performing additional hyperparameter tuning. Additionally, we plan to speed up the processing of the stacks by lowering the amount of calculations and the size of the networks.", + "bbox": [ + 496, + 660, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Supported by the New Zealand Ministry of Business, Innovation and Employment (MBIE), Smart Ideas grant \"Extending the Boundaries of Digital Signal Processing: AI-powered Fourier Transformation Alternative (E7943)\". We also thank Oliver Batchelor, François Bissey, and Piotr Kolenderski for sharing their computational resources.", + "bbox": [ + 496, + 780, + 890, + 885 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25035", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kai Wang and Zhihua Ding. Spectral calibration in spectral domain optical coherence tomography. Chinese Optics Letters, 6(12):902-904, 2008. 2, 6", + "[2] Gongpu Lan and Guoqiang Li. Design of ak-space spectrometer for ultra-broad waveband spectral domain optical coherence tomography. Scientific reports, 7(1):42353, 2017. 2", + "[3] Ireneusz Grulkowski, Silvestre Manzanera, Lukasz Cwiklinski, Franciszek Sobczuk, Karol Karnowski, and Pablo Artal. Swept source optical coherence tomography and tunable lens technology for comprehensive imaging and biometry of the whole eye. Optica, 5(1):52-59, 2018. 2", + "[4] Timothy R Hillman and David D Sampson. The effect of water dispersion and absorption on axial resolution in ultrahigh-resolution optical coherence tomography. Optics express, 13 (6):1860-1874, 2005. 2", + "[5] Siyu Chen, Benjamin Potsaid, Yan Li, Junhong Lin, Yun- chan Hwang, Eric M Moult, Jason Zhang, David Huang, and James G Fujimoto. High speed, long range, deep penetration swept source oct for structural and angiographic imaging of the anterior eye. Scientific reports, 12(1):992, 2022. 2", + "[6] Jiangjie Huang, Jinyu Fan, Yi He, and Guohua Shi. Physical compensation method for dispersion of multiple materials in swept source optical coherence tomography. Journal of Biophotonics, 16(10):e202300167, 2023.", + "[7] Daniel L Marks, Amy L Oldenburg, J Joshua Reynolds, and Stephen A Boppart. Digital algorithm for dispersion correction in optical coherence tomography for homogeneous and stratified media. Applied optics, 42(2):204-217, 2003.", + "[8] Liuhua Pan, Xiangzhao Wang, Zhongliang Li, Xiangyang Zhang, Yang Bu, Nan Nan, Yan Chen, Xuan Wang, and Fengzhao Dai. Depth-dependent dispersion compensation for full-depth oct image. Optics express, 25(9):10345-10354, 2017. 2", + "[9] Shaiban Ahmed, David Le, Taeyoon Son, Tobiloba Adejumo, Guangying Ma, and Xincheng Yao. Asc-net: an opensource deep learning network for automated dispersion compensation in oct. Frontiers in Medicine, 9:864879, 2022. 2", + "[10] Sylwia M Kolenderska, Bastian Brauer, and Frédérique Vanholsbeeck. Dispersion mapping as a simple postprocessing step for fourier domain optical coherence tomography data. Scientific reports, 8(1):9244, 2018. 2", + "[11] Christos Photiou and Costas Pitris. Comparison of tissue dispersion measurement techniques based on optical coherence tomography. Journal of biomedical optics, 24(4):046003-046003, 2019.", + "[12] Krzysztof A Maliszewski, Magdalena A Urbańska, Piotr Kolenderski, Varvara Vetrova, and Sylwia M Kolenderska. Extracting group velocity dispersion values using quantum-mimic optical coherence tomography and machine learning. Scientific Reports, 13(1):6596, 2023. 2, 4", + "[13] Christos Photiou, Evgenia Bousi, Ioanna Zouvani, and Costas Pitris. Using speckle to measure tissue dispersion in optical coherence tomography. Biomedical optics express, 8 (5):2528-2535, 2017. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4", + "[15] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623-2631, 2019. 4", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4", + "[17] Jayme Garcia Arnal Barbedo. Impact of dataset size and variety on the effectiveness of deep learning and transfer learning for plant disease classification. Computers and electronics in agriculture, 153:46-53, 2018. 4" + ], + "bbox": [ + 501, + 92, + 890, + 344 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25036", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_model.json b/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..866074f00a27a2eb75ca21c08d422075724de93f --- /dev/null +++ b/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_model.json @@ -0,0 +1,1736 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.131, + 0.748, + 0.154 + ], + "angle": 0, + "content": "Selective nonlinearities removal from digital signals" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.182, + 0.734, + 0.252 + ], + "angle": 0, + "content": "Krzysztof A. Maliszewski \nSchool of Mathematics and Statistics, University of Canterbury \nChristchurch, New Zealand \nkrzysztof.maliszewski@pg.canterbury.ac.nz" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.26, + 0.731, + 0.311 + ], + "angle": 0, + "content": "Magdalena A. Urbańska \nSchool of Food and Advanced Technology, Massey University Palmerston North, New Zealand" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.317, + 0.584, + 0.329 + ], + "angle": 0, + "content": "murbanska@massey.ac.nz" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.338, + 0.734, + 0.39 + ], + "angle": 0, + "content": "Varvara Vetrova \nSchool of Mathematics and Statistics, University of Canterbury \nChristchurch, New Zealand" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.394, + 0.627, + 0.407 + ], + "angle": 0, + "content": "varvara.vetrova@canterbury.ac.nz" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.416, + 0.754, + 0.485 + ], + "angle": 0, + "content": "Sylwia M. Kolenderska \nInstitute of Physics, Faculty of Physics, Astronomy and Informatics \nNicolaus Copernicus University in Torun \nul. Grudziadzka 5, 87-100 Torun, Poland" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.756, + 0.504 + ], + "angle": 0, + "content": "School of Physical and Chemical Sciences, University of Canterbury" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.505, + 0.594, + 0.519 + ], + "angle": 0, + "content": "Christchurch, New Zealand" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.525, + 0.595, + 0.537 + ], + "angle": 0, + "content": "skol745@aucklanduni.ac.nz" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.574, + 0.314, + 0.59 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Many instruments performing optical and non-optical imaging and sensing, such as Optical Coherence Tomography (OCT), Magnetic Resonance Imaging or Fourier-transform spectrometry, produce digital signals containing modulations, sine-like components, which only after Fourier transformation give information about the structure or characteristics of the investigated object. Due to the fundamental physics-related limitations of such methods, the distribution of these signal components is often nonlinear and, when not properly compensated, leads to the resolution, precision or quality drop in the final image. Here, we propose an innovative approach that has the potential to allow cleaning of the signal from the nonlinearities but most of all, it now allows to switch the given order off, leaving all others intact. The latter provides a tool for more in-depth analysis of the nonlinearity-inducing properties of the investigated object, which can lead to applications in early disease detection or more sensitive sensing of chemical compounds. We consider OCT signals and nonlinear" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.895, + 0.728 + ], + "angle": 0, + "content": "ities up to the third order. In our approach, we propose two neural networks: one to remove solely the second-order nonlinearity and the other for removing solely the third-order nonlinearity. The input of the networks is a novel two-dimensional data structure with all the information needed for the network to infer a nonlinearity-free signal. We describe the developed networks and present the results for second-order and third-order nonlinearity removal in OCT data representing the images of various objects: a mirror, glass, and fruits." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.755, + 0.633, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Optical Coherence Tomography (OCT), Magnetic Resonance Imaging (MRI) or Fourier-transform Spectrometry (FTS) belong to a class of methods where useful, interpretable signal is obtained in an indirect way, more specifically by Fourier transforming the raw signal acquired by the instrument and taking its absolute value. On the one hand, such an indirect approach makes OCT-based imaging and FTS-based sensing faster and enables MRI, but on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25028" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.299, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.089, + 0.437, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.168, + 0.299, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.169, + 0.436, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.253, + 0.299, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.253, + 0.436, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.337, + 0.299, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.337, + 0.435, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Figure 1. Effects of signal nonlinearity on the signal's Fourier transform. (a) A signal with a linear modulation distribution Fourier transform to (b) a narrow peak. (c) If the nonlinearity of the modulation is of the second order, (d) the peak gets broadened. (e) If the nonlinearity is of the third order, (f) the peak gets distorted. In OCT, this will result in deteriorated image quality. (g) In practice, both nonlinearities appear at once, resulting in (h) the peak being elongated in the direction of the third-order distortion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.468, + 0.734 + ], + "angle": 0, + "content": "the other hand, it makes room for errors related to the limitations of Fourier transformation. One such limitation is the Fourier transformation's sensitivity to nonlinearities in the input signal. In the simplest case of a signal being a single-frequency modulation, a second-order nonlinear distribution will Fourier transform to a broader peak (compare Fig. 1a,b with c,d). If this distribution is of a third-order nonlinear character, the peak in the Fourier transform will get distorted (Fig. 1e,f). In practice, both nonlinearities appear at once, the second-order one bigger than the third-order one, leading to a broad peak elongated in the direction of the third-order nonlinearity distortion (Fig. 1g,h)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The influence of nonlinearities is best visible in OCT, where the Fourier transform of the raw signal constitutes one line of the image. Since the width of the peak represents the depth resolution (i.e. how small a detail can be observed), and the peak's shape correlates with the image quality, nonlinearities are generally thought of as detrimental and various techniques are devised to remove them [1, 2]. These techniques do not achieve absolute removal of nonlinearity, preventing single-shot imaging of bulk objects, such as the entire eye [3, 4] and consequently necessitating elaborate workaround solutions [5-8]. A Machine" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "Learning solution for nonlinearity removal was proposed as well [9] – although robust and reliable, it is biased towards specific kinds of objects with specific nonlinear properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.258 + ], + "angle": 0, + "content": "On the other hand, the inherent optical properties that produce nonlinearities in the imaged objects have been shown to enable the characterisation of the objects [10-12] and correlate the change of their nonlinear character with the progression of diseases [13]. Unfortunately, the current methods for extracting the second-order nonlinearity are very object-specific as well as error-prone preventing them from being applied in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.893, + 0.577 + ], + "angle": 0, + "content": "We present an approach that enables flexible and comfortable management of the signal nonlinearities, letting the user study one order in the absence of the others, a capability previously impossible to achieve using current standard or Machine Learning methods. In our approach, for each nonlinearity order, a neural network is trained whose input is a special two-dimensional data structure, a stack. Each row of the stack represents an amplitude of the Fourier transform of an input signal multiplied by a different nonlinearity-correcting factor. Consequently, it contains nonlinearity-free regions which the network identifies and uses to build a nonlinearity-free Fourier transform amplitude. We present two networks, one to remove the second-order nonlinearity while keeping the third-order one intact, and the second network to remove the third-order nonlinearity while keeping the second-order one intact. We test our approach on OCT signals that naturally contain substantial amounts of second-order nonlinearity and small amounts of third-order nonlinearity. The tests on both computer-generated and experimental signals indicate the robustness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.804 + ], + "angle": 0, + "content": "We start by describing our approach in Section 2, including the creation of the stacks in Subsection 2.1, their application in the training processes in Subsection 2.2, presentation of the used architecture and methodology in Subsection 2.3, and the description of the datasets (Subsection 2.4) used for training of our models. In Section 3, we report the performance of the networks in removing the respective nonlinearity, complete with the analysis of the training process and the subsequent tests on computer-generated data (Subsection 3.1) and experimental data (Subsections 3.2, 3.3 and 3.4). In the final sections, we summarise the obtained results (Section 4), pointing to the strong points of our approach, identifying the observed and potential shortcomings, and discussing potential solutions and the scope of future work (Section 5)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.821, + 0.596, + 0.838 + ], + "angle": 0, + "content": "2. Concept" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.742, + 0.863 + ], + "angle": 0, + "content": "2.1. Amplitude stacks as inputs" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our networks rely on a two-dimensional data structure as inputs: a stack (Fig. 2). The rows of a stack are created" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25029" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.09, + 0.148, + 0.102 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.09, + 0.432, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.162, + 0.101, + 0.173 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.179, + 0.235, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.161, + 0.264, + 0.173 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.162, + 0.462, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.3, + 0.1, + 0.312 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.314, + 0.235, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.3, + 0.264, + 0.312 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.304, + 0.462, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Figure 2. Calculation of the input stacks. (a) The raw signal is (b) multiplied by 32 different complex exponents representing 32 different levels of the second-order nonlinearity: \\( C_0, C_1, C_2, \\ldots, C_{31} \\). Each row is then Fourier transformed. (c) The 2nd-order stack is created from the absolute value of the complex Fourier transforms. Similarly, (d) the raw signal is multiplied by 32 different complex exponents representing 32 different levels of the third-order nonlinearity: \\( D_0, D_1, D_2, \\ldots, D_{31} \\) and then Fourier transformed to build (e) a 3rd-order stack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.901 + ], + "angle": 0, + "content": "in the following way. First, the raw signal is multiplied (Fig. 2a) by a complex exponent with a different quadratic argument in the case of the stack used for removing the second-order nonlinearity (Fig. 2b) or with a different cubic argument in the case of the stack for the third-order nonlinearity removal (Fig. 2d). Coefficients in the quadratic arguments, \\((C_0,\\dots,C_{31}\\) in Fig. 2b), and coefficients in the cubic arguments \\((D_0,\\dots,D_{31}\\) in Fig. 2d), are chosen to cover a wide range of the second- and third-order nonlinearity, ensuring that the nonlinearities present in the signal are fully compensated. Then, each row is Fourier transformed, and an absolute value is calculated to obtain a 2nd-order stack (Fig. 2c) or a 3rd-order stack (Fig. 2e). The range of the nonlinear coefficients is symmetrical around 0 and broad enough to ensure that each modulation's nonlinearity in the raw signal is compensated. In the 2nd-order stack, these will be the places with the smallest width (marked 1, 2, 3 and 4 in Fig. 3b) and in the 3rd-order stack, the places containing no distortions (marked 1, 2, 3 and 4 in Fig. 3c)." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Figure 3. (a) The Fourier transform amplitude of a raw signal, presented in Fig. 2a, shows four peaks affected by a different amount and type of nonlinearity. The second-order nonlinearity is removed with Network #1 trained on (b) the 2nd-order stack as input and (d) Fourier transform amplitude free of the second-order nonlinearity as output. Similarly, the third-order nonlinearity is removed with Network #2 trained on (b) the 3rd-order stack as input and (d) Fourier transform amplitude free of the third-order nonlinearity as output. As a result, (f) the second-order nonlinearity is removed leaving the third-order one (see peak 4) using Network #1, (g) the third-order nonlinearity is removed leaving the second-order one intact (see peak 4) using Network #2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.787, + 0.694 + ], + "angle": 0, + "content": "2.2. Nonlinearity-removing networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The approach for removing the second-order nonlinearity and the approach for removing the third-order nonlinearity are similar, presented schematically in Fig. 3. The input of the network for the second-order nonlinearity removal, Network #1, is the 2nd-order stack (Fig. 3b) and the output is a Fourier transform amplitude free of second-order nonlinearity (Fig. 3d). Because the input 2nd-order stack contains regions corresponding to perfectly removed second-order nonlinearity, the training process can be viewed as learning how to identify these regions and then stitching them all up. Similarly, the network for removing third-order nonlinearity, Network #2, takes the 3rd-order stack as input (Fig. 3c) and outputs a Fourier transform amplitude free" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25030" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.462, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.307, + 0.462, + 0.323 + ], + "angle": 0, + "content": "Figure 4. Architecture of our nonlinearity-removing networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.457 + ], + "angle": 0, + "content": "from the 3rd-order nonlinearity (Fig. 3e). Input for both networks has a shape of 32 by 1024 pixels with a single channel whereas output is a vector of size 1024. When compared to the Fourier transform amplitude of the raw signal (Fig. 3a), the signal returned by Network #1 (Fig. 3f) incorporates only the third-order nonlinearity, and the signal returned by Network #2 (Fig. 3g) contains only the second-order nonlinearity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.216, + 0.484 + ], + "angle": 0, + "content": "2.3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Our nonlinearity-removing networks are inspired by the U-Net architecture [14]. U-Net stands out for its use of a contracting path to capture contextual information, an expanding path for precise feature localisation, and the incorporation of skip connections that preserve fine-grained details, particularly beneficial for precise object delineation. We leverage these capabilities to identify and extract nonlinearity-related features from the stacks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.749 + ], + "angle": 0, + "content": "While the classical U-Net deals with 2D data structures for both input and output, in our case, we perform dimension reduction to output 1D signals by incorporating a max-pooling layer. We further refined the architecture by optimising hyperparameters with the validation dataset using Optuna [15]. As a result of the tuning, we adopted the Mean Absolute Error (MAE) loss function and the Adam optimiser [16] with a set learning rate of 0.0002. Our final network architecture is presented in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.182, + 0.772 + ], + "angle": 0, + "content": "2.4. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our datasets consist of noiseless computer-generated signals that are Gaussian-shaped, and diverse and randomised in terms of their content. If put in the OCT-imaging-related terms, they represent a broad range of possible imaged object configurations. These objects are simulated to have a random number of interfaces (from 2 to 12) positioned at arbitrary locations, with each object layer exhibiting varying random reflectivity and second- and third-order nonlin" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.891, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.272, + 0.895, + 0.344 + ], + "angle": 0, + "content": "Figure 5. Network #1 and Network #2 performance measured with the Mean Absolute Error (MAE) loss function across 30 epochs of training. Blue lines depict Network #1, while orange lines showcase the Network #2 performance. Dotted vertical lines highlight epochs exhibiting the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.894, + 0.447 + ], + "angle": 0, + "content": "earity values. These object parameters are used to synthesise modulatory signals equivalent to those acquired in an OCT measurement (see example signals in Fig. 1a, c, e, g or Fig. 2a). Next, the input stacks and outputs, the Fourier transform amplitudes of signals free of the respective nonlinearity, are calculated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.894, + 0.553 + ], + "angle": 0, + "content": "The risk of overfitting was minimised by using the training datasets with an optimised size [17]. In our research, this was found to be between 160,000 and 240,000 objects. The validation and test datasets include 14,993 objects. The distribution is uniform across the objects with a specific number of interfaces, resulting in an equal allocation of 1,363 data samples for each distinct object type." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.894, + 0.614 + ], + "angle": 0, + "content": "Before starting the training, as part of the preprocessing stage, both the input and output data were normalised through Min-Max scaling, ensuring that values fell within the standardised range of 0 to 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.634, + 0.643 + ], + "angle": 0, + "content": "3. Performance" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.674, + 0.667 + ], + "angle": 0, + "content": "3.1. In silico data tests" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Each nonlinearity-removing neural network underwent training using a training dataset consisting of 200,000 objects. Within each mini-batch, 8 input stacks were processed. The proposed algorithm requires 39 ms to process a stack on a Quadro RTX 6000 GPU. Network #1 (Fig. 5, blue lines) and Network #2 (Fig. 5, orange lines) both displayed convergence within just a few epochs (Fig. 5) and were trained until the thirtieth epoch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To better evaluate the networks and results, we employed the Goodness-of-Fit Approximity (GoF) metric introduced in [12]. The GoF metric assesses the accuracy of predictions by calculating the percentage of values that lie within a predefined distance from the corresponding ground truth values. For this GoF metric, we define an error threshold as a ratio between the predefined distance and the value range" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "25031" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.09, + 0.47, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.421 + ], + "angle": 0, + "content": "Figure 6. Networks performance assessed through the metric of GoF, measuring the degree to which predictions align with the ground truth within \\(0.1\\%\\) error in the test dataset comprising 14,993 data samples for (a) Network #1 and (b) Network #2. The graphs at the top (a1,a2) are zoomed-in areas of the top area of (b1,b2) the whole graphs presented." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.47 + ], + "angle": 0, + "content": "span. In our case, where the values range from 0 to 1, the distance of 0.01 means \\(1\\%\\) error threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.565 + ], + "angle": 0, + "content": "Based on the post-training evaluation using a \\(1\\%\\) error threshold and the test dataset, Network #1 started to overfit at epoch 25 (Fig. 5, indicated by the blue dotted vertical line). In contrast, Network #2 demonstrated resilience against overfitting and achieved its best results in the 28th epoch (Fig. 5, marked by the red dotted vertical line)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.471, + 0.688 + ], + "angle": 0, + "content": "Using the test dataset and GoF with a \\(1\\%\\) error threshold, both networks achieved a remarkable mean GoF value of \\(99.98\\%\\), which means that \\(99.98\\%\\) of the predictions differed from the ground truth by less than \\(1\\%\\), with only one example below \\(95\\%\\) GoF. In light of this, to be able to carry out a meaningful analysis, the error threshold was further reduced to \\(0.1\\%\\) and a boxplot graph was plotted for each model in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Both networks, Network #1 (Fig. 6a) and Network #2 (Fig. 6b), demonstrated very good performance even for error threshold \\(0.1\\%\\). Only approximately \\(9.52\\%\\) and \\(3.07\\%\\) of the predictions from Network #1 and Network #2, respectively, fell below the \\(95\\%\\) GoF for such a small error, as showed in Tab. 1 that also includes GoF values for objects with distinct number of layers. Fig. 7 offers an in-depth analysis of the worst-predicted examples for the Network #1 (Fig. 7a) and the Network #2 (Fig. 7b) showcasing GoF values of \\(65.5\\%\\) and \\(45.5\\%\\), respectively. Predictions of peak amplitudes in the nonlinearity-free Fourier transform amplitude (Fig. 7(3), orange line) closely align the expected outcomes (blue line) deviating only by the error threshold level. Also, the networks predict values close to" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.885, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.47, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Figure 7. Predictions with very low GoF for \\(0.1\\%\\) error threshold for (a) Network #1 and (b) Network #2. Based on the input stack (1), models predict (2, orange line, PR) amplitudes free of the respective nonlinearity that exhibit a strong correlation with the ground truth (2, blue line, GT). (3) Nonlinearity-free peaks are recovered with high fidelity, but (4) the models sometimes struggle with zero-valued places." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.658 + ], + "angle": 0, + "content": "zero rather than precisely zero, contributing to worse GoF metric results. One should keep in mind here that these discrepancies are extremely low: at the level of around 2e-3 as presented in Fig. 7a4 and b4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.895, + 0.698 + ], + "angle": 0, + "content": "3.2. Experimental signals: mirror at different depths" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Networks were tested on the signals obtained through an OCT imaging of a mirror placed at 11 different depth positions. Each acquired signal Fourier transforms to a peak whose amount of second-order-nonlinearity-related broadening and third-order-nonlinearity-related distortion increases with depth (Fig. 8a). Due to their simultaneous appearance, the third-order nonlinearity manifests itself in the form of peak elongation. The depth-dependent deterioration is due to the fundamental limitations of OCT's spectrometer-based light detection. Additionally, the second- and third-order nonlinearities originating from the OCT's interferometer add up - depth-independently - to that effect, further worsening the outcome. Typically," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25032" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.104 + ], + "angle": 0, + "content": "Table 1. The number of cases where GoF with error threshold \\( {0.1}\\% \\) falls below \\( {95}\\% \\) for objects with a specific number of reflective layers." + }, + { + "type": "table", + "bbox": [ + 0.16, + 0.115, + 0.807, + 0.209 + ], + "angle": 0, + "content": "
Net.Interfaces23456789101112Totals
#1# below 95%3842601931381058779554544371381
# in dataset1363136313631363136313631363136313631363136314993
% below28.1719.0814.1610.127.706.385.804.043.303.322.719.52
#2# below 95%10782575438323122131311460
# in dataset1363136313631363136313631363136313631363136314993
% below7.856.024.183.962.792.352.271.610.950.950.813.07
" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.218, + 0.397, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.3, + 0.4, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.383, + 0.4, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.485, + 0.277, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.485, + 0.438, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.733 + ], + "angle": 0, + "content": "Figure 8. Mirror at 11 different depths. (a) Fourier transforming raw signals results in peaks that are both broadened by the second-order nonlinearity and distorted by the third-order nonlinearity. The deeper the peak, the bigger the nonlinearity effect, a behaviour characteristic of OCT signals acquired with a spectrometer. (b) Network #1 removes the second-order nonlinearity, leaving the third-order one intact. (c) Network #2 removes the third-order nonlinearity, leaving the third-order one intact. (d, e) The traditional linearisation (grey line) is not as precise as our approach (green lines)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.79 + ], + "angle": 0, + "content": "spectrometer- and interferometer-related nonlinearities can be satisfactorily removed in a calibration step before the measurement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The application of Network #1 removes the second-order nonlinearity in its entirety and consequently, provides a dramatic reduction of peaks' widths (Fig. 8b). Because the third-order nonlinearity remains intact, the peaks' shapes incorporate the characteristic \"tail\". On the other hand, Network #2 removes the third-order nonlinearity, leading to a complete reduction of the elongation (Fig. 8c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.718 + ], + "angle": 0, + "content": "One observes the superiority of our approach when the peaks obtained with Network #1 are compared to the peaks obtained as a result of traditional linearisation methods which remove all nonlinearity orders. Here, we used the method in [1] where two raw signals corresponding to the mirror placed at two different locations are used. First, their phases are subtracted to remove the constant interferometer-related contribution, enabling the calculation of the correction vector for spectrometer-related nonlinearities. Once the spectrometer's nonlinearity is removed, one of the initial spectra is used to generate an interferometer-nonlinearities-compensating vector. The peaks located close to 0 depth index obtained using our approach and the standard linearisation method (Fig. 8d, green and grey lines) are similar in shape: the network-outputted peak is higher which is very favourable - it means a higher signal-to-noise ratio, which is associated with better quality imaging. One observes a small \"tail\" at the side, most probably due to a small amount of third-order nonlinearity present at such shallow depths, whereas the peak obtained using the traditional methods is a smooth Gaussian. More interestingly, for a peak at a substantial depth (Fig. 8e), where the third-order nonlinearity is naturally much bigger (let us remind that this depth-increasing nonlinearity is tied to OCT spectrometer), the \"tail\" is much more pronounced in case of our network but not present in case of the traditional method. In the latter case, the peak is elongated which suggests the presence of both second- and third-order nonlinearities. This leads to the conclusion that traditional linearisation, although very satisfactory in practical applications, still has room for improvement. In fact, despite 30 years of OCT in science and industry, new methods for more precise OCT data linearisation are still being developed." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.727, + 0.58, + 0.741 + ], + "angle": 0, + "content": "3.3. Glass" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Another test was performed on data corresponding to a \\(300\\mu \\mathrm{m}\\) thick sapphire glass put at an angle obtained with the same instrument. This data consisted of 256 raw OCT signals which represent a different transversal location on the glass. When Fourier transformed one after another, they form a 2D depth image of the object (Fig. 9a, 1 and 2 denote the front and back surface of the glass, the third line is an artefact placed at the depth equal to the thickness of the glass, inherent to OCT). Without linearisation, the image elements are broadened the deeper they are. When the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.193, + 0.091, + 0.368, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.196, + 0.273, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.195, + 0.446, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.297, + 0.272, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.297, + 0.445, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.151, + 0.398, + 0.391, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Figure 9. Sapphire glass. (a) Fourier transformation of raw signals gives an OCT image where the quality deteriorates with depth. (b) Network #1 removes the second-order nonlinearity which accounts for the most of quality deterioration, (c) Due to a small amount of the third-order nonlinearity, the application of Network #2 does not improve the image quality. (d), Image line #152 from Network #1 (green line) is compared with its raw image equivalent (grey line) to show dramatic peaks' width reduction, (e) The image line #152 from Network #2 output (green line) and its raw image equivalent (grey line) show no significant difference except for a more favourable shape for the former. (f) Network #1 (green line) performs quite well when compared to the traditional method output (black line). 1 and 2 mark the front and back surface of the imaged glass, the middle line is an artefact." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.825 + ], + "angle": 0, + "content": "same data is processed with Network #1, the broadening is removed from all the image elements (see the image in Fig. 9b and the 152nd image line in Fig. 9d). Due to very small amounts of the third-order nonlinearity, Network #2 does not bring perceptible improvements (Fig. 9c), but under a closer inspection of a single image line (Fig. 9e), one can observe some improvement in the peak height and a possible side elongation reduction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Fig. 9f compares the output of Network #1 (green line) and the traditional linearisation method (black line), indicating that this model could be a satisfactory alternative for the traditional image calculation algorithm when the third-order nonlinearities are negligible. It can be used as such an alter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "native for the OCT laboratory instrument which was used here to obtain the data and which is nevertheless a fairly good representative of a standard OCT instrument. We note that Network #1 improved the shape of the artefact (the middle peak in Fig. 9f), which - although generally irrelevant for OCT imaging - is still something noteworthy as such a feat is beyond the capabilities of traditional linearisation methods, showing that our approach, unlike the traditional ones, removes the nonlinearities nondiscriminatively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.236, + 0.7, + 0.251 + ], + "angle": 0, + "content": "3.4. Grape and cucumber" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.349 + ], + "angle": 0, + "content": "The final tests were performed on the data corresponding to a grape (Fig. 10) and a cucumber (Fig. 11). The former represents a high signal-to-noise ratio (highly visible modulation in the signal as depicted in Fig. 10a) and the latter a low signal-to-noise ratio (poorly visible modulation in the signal as depicted in Fig. 11a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Fourier transformation of raw signals gives images of a low quality that further deteriorates with depth (Fig. 10b, Fig. 11b). Again, the improvement obtained when applying Network #1 (Fig. 10c, Fig. 11c) is comparable to that obtained with traditional methods (Fig. 10d, Fig. 11d). Under closer inspection (Fig. 10e, Fig. 11e), one notices that Network #1 output is less detailed for the higher signal-to-noise grape data - most probably due to the remaining third-order nonlinearities - and loses some details for the lower signal-to-noise ratio cucumber data. The latter loss is the direct consequence of training data representing predominantly high signal-to-noise ratio signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.891, + 0.576 + ], + "angle": 0, + "content": "Just as in the case of the glass data, due to small amounts of third-order nonlinearity, Network #2 did not provide visually significant improvement and therefore, was omitted." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.733, + 0.605 + ], + "angle": 0, + "content": "4. Summary and discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Two models were demonstrated: one for removing second-order nonlinearity from digital signals and leaving the third-order one intact, and another one for removing the third-order nonlinearity but leaving the second-order one intact. Their nonlinearity-removing capabilities were extensively analysed using computer-generated signals as well as experimental signals acquired with a laboratory Optical Coherence Tomography (OCT) device. Tests on the OCT signals, which corresponded to objects with varying levels of complication: mirror, sapphire, grape and cucumber, allowed us to evaluate the networks in terms of their applicability in a world-relevant technology. Whereas the tests and analysis done on the computer-generated signals revealed extraordinary performance in terms of similarity of the ground truth and prediction signals, the tests on OCT-acquired data indicated the networks' advantageous operation when compared to approaches traditionally used in OCT." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.899 + ], + "angle": 0, + "content": "During training, both networks converged exceptionally fast and reached remarkable loss function values. Nevertheless" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25034" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.434, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.168, + 0.446, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.254, + 0.446, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.342, + 0.446, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.45, + 0.436, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Figure 10. Grape. (a) One of the raw signals representing a single line in the image of a grape. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image when compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees very well with its equivalent from the image obtained using traditional methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.753 + ], + "angle": 0, + "content": "less, we have identified possible improvements: the closer inspection of the predictions revealed non-zero values that are zero-valued areas in the ground-truth signals. Another current shortcoming of both networks is their time of processing an entire OCT image; consequently, for now, the processing speed is insufficient for the real-time processing of data coming straight from OCT devices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.208, + 0.785 + ], + "angle": 0, + "content": "5. Future work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The future work will mainly consist in developing an approach that enables one to use both networks in a sequence so that both the second- and third-order nonlinearity removal is achieved. Since, as was shown here, the networks remove the nonlinearities completely, their sequential use will provide a very valuable tool for extremely precise data linearisation, especially attractive for OCT where any un" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.094, + 0.865, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.17, + 0.872, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.256, + 0.872, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.343, + 0.869, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.452, + 0.862, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Figure 11. Cucumber. (a) One of the raw signals representing a single line in the image of a cucumber. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees well with its equivalent from the image obtained using traditional methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.646, + 0.842, + 0.661 + ], + "angle": 0, + "content": "compensated nonlinearities lead to the quality drop." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Also, in terms of the network performance, we will try to remove the discrepancy for zero-valued elements in the ground truth and prediction signals by adding extra convolutional layers before the output to filter out the unwanted values as well as performing additional hyperparameter tuning. Additionally, we plan to speed up the processing of the stacks by lowering the amount of calculations and the size of the networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Acknowledgements. Supported by the New Zealand Ministry of Business, Innovation and Employment (MBIE), Smart Ideas grant \"Extending the Boundaries of Digital Signal Processing: AI-powered Fourier Transformation Alternative (E7943)\". We also thank Oliver Batchelor, François Bissey, and Piotr Kolenderski for sharing their computational resources." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25035" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Kai Wang and Zhihua Ding. Spectral calibration in spectral domain optical coherence tomography. Chinese Optics Letters, 6(12):902-904, 2008. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.212 + ], + "angle": 0, + "content": "[2] Gongpu Lan and Guoqiang Li. Design of ak-space spectrometer for ultra-broad waveband spectral domain optical coherence tomography. Scientific reports, 7(1):42353, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.468, + 0.285 + ], + "angle": 0, + "content": "[3] Ireneusz Grulkowski, Silvestre Manzanera, Lukasz Cwiklinski, Franciszek Sobczuk, Karol Karnowski, and Pablo Artal. Swept source optical coherence tomography and tunable lens technology for comprehensive imaging and biometry of the whole eye. Optica, 5(1):52-59, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[4] Timothy R Hillman and David D Sampson. The effect of water dispersion and absorption on axial resolution in ultrahigh-resolution optical coherence tomography. Optics express, 13 (6):1860-1874, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.468, + 0.413 + ], + "angle": 0, + "content": "[5] Siyu Chen, Benjamin Potsaid, Yan Li, Junhong Lin, Yun- chan Hwang, Eric M Moult, Jason Zhang, David Huang, and James G Fujimoto. High speed, long range, deep penetration swept source oct for structural and angiographic imaging of the anterior eye. Scientific reports, 12(1):992, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.468, + 0.47 + ], + "angle": 0, + "content": "[6] Jiangjie Huang, Jinyu Fan, Yi He, and Guohua Shi. Physical compensation method for dispersion of multiple materials in swept source optical coherence tomography. Journal of Biophotonics, 16(10):e202300167, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.468, + 0.529 + ], + "angle": 0, + "content": "[7] Daniel L Marks, Amy L Oldenburg, J Joshua Reynolds, and Stephen A Boppart. Digital algorithm for dispersion correction in optical coherence tomography for homogeneous and stratified media. Applied optics, 42(2):204-217, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.468, + 0.599 + ], + "angle": 0, + "content": "[8] Liuhua Pan, Xiangzhao Wang, Zhongliang Li, Xiangyang Zhang, Yang Bu, Nan Nan, Yan Chen, Xuan Wang, and Fengzhao Dai. Depth-dependent dispersion compensation for full-depth oct image. Optics express, 25(9):10345-10354, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[9] Shaiban Ahmed, David Le, Taeyoon Son, Tobiloba Adejumo, Guangying Ma, and Xincheng Yao. Asc-net: an opensource deep learning network for automated dispersion compensation in oct. Frontiers in Medicine, 9:864879, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[10] Sylwia M Kolenderska, Bastian Brauer, and Frédérique Vanholsbeeck. Dispersion mapping as a simple postprocessing step for fourier domain optical coherence tomography data. Scientific reports, 8(1):9244, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[11] Christos Photiou and Costas Pitris. Comparison of tissue dispersion measurement techniques based on optical coherence tomography. Journal of biomedical optics, 24(4):046003-046003, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[12] Krzysztof A Maliszewski, Magdalena A Urbańska, Piotr Kolenderski, Varvara Vetrova, and Sylwia M Kolenderska. Extracting group velocity dispersion values using quantum-mimic optical coherence tomography and machine learning. Scientific Reports, 13(1):6596, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[13] Christos Photiou, Evgenia Bousi, Ioanna Zouvani, and Costas Pitris. Using speckle to measure tissue dispersion in optical coherence tomography. Biomedical optics express, 8 (5):2528-2535, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[15] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623-2631, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[17] Jayme Garcia Arnal Barbedo. Impact of dataset size and variety on the effectiveness of deep learning and transfer learning for plant disease classification. Computers and electronics in agriculture, 153:46-53, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25036" + } + ] +] \ No newline at end of file diff --git a/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_origin.pdf b/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7f3b3da839039e31e1c3fb2ffcacc80c5a2e67a0 --- /dev/null +++ b/2024/Selective Nonlinearities Removal from Digital Signals/ce4a4f68-a75b-4b4e-affa-326923d0b5b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8c4915483e6395acdd8992599e124f86215f92e8bcec338155ffc064dc4dc63 +size 5239854 diff --git a/2024/Selective Nonlinearities Removal from Digital Signals/full.md b/2024/Selective Nonlinearities Removal from Digital Signals/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f218430d90b6ba6a0df623d7574adeb39553046b --- /dev/null +++ b/2024/Selective Nonlinearities Removal from Digital Signals/full.md @@ -0,0 +1,263 @@ +# Selective nonlinearities removal from digital signals + +Krzysztof A. Maliszewski +School of Mathematics and Statistics, University of Canterbury +Christchurch, New Zealand +krzysztof.maliszewski@pg.canterbury.ac.nz + +Magdalena A. Urbańska +School of Food and Advanced Technology, Massey University Palmerston North, New Zealand + +murbanska@massey.ac.nz + +Varvara Vetrova +School of Mathematics and Statistics, University of Canterbury +Christchurch, New Zealand + +varvara.vetrova@canterbury.ac.nz + +Sylwia M. Kolenderska +Institute of Physics, Faculty of Physics, Astronomy and Informatics +Nicolaus Copernicus University in Torun +ul. Grudziadzka 5, 87-100 Torun, Poland + +School of Physical and Chemical Sciences, University of Canterbury + +Christchurch, New Zealand + +skol745@aucklanduni.ac.nz + +# Abstract + +Many instruments performing optical and non-optical imaging and sensing, such as Optical Coherence Tomography (OCT), Magnetic Resonance Imaging or Fourier-transform spectrometry, produce digital signals containing modulations, sine-like components, which only after Fourier transformation give information about the structure or characteristics of the investigated object. Due to the fundamental physics-related limitations of such methods, the distribution of these signal components is often nonlinear and, when not properly compensated, leads to the resolution, precision or quality drop in the final image. Here, we propose an innovative approach that has the potential to allow cleaning of the signal from the nonlinearities but most of all, it now allows to switch the given order off, leaving all others intact. The latter provides a tool for more in-depth analysis of the nonlinearity-inducing properties of the investigated object, which can lead to applications in early disease detection or more sensitive sensing of chemical compounds. We consider OCT signals and nonlinear + +ities up to the third order. In our approach, we propose two neural networks: one to remove solely the second-order nonlinearity and the other for removing solely the third-order nonlinearity. The input of the networks is a novel two-dimensional data structure with all the information needed for the network to infer a nonlinearity-free signal. We describe the developed networks and present the results for second-order and third-order nonlinearity removal in OCT data representing the images of various objects: a mirror, glass, and fruits. + +# 1. Introduction + +Optical Coherence Tomography (OCT), Magnetic Resonance Imaging (MRI) or Fourier-transform Spectrometry (FTS) belong to a class of methods where useful, interpretable signal is obtained in an indirect way, more specifically by Fourier transforming the raw signal acquired by the instrument and taking its absolute value. On the one hand, such an indirect approach makes OCT-based imaging and FTS-based sensing faster and enables MRI, but on + +![](images/048039b52eca5292a5bd2bb86f3c0b5b328a9d8ae45e104c6bbef271fe54bd27.jpg) + +![](images/be826ae52a302df15bba60a416d8760e8e10906ce0eb7c235ab1aa530d22e642.jpg) + +![](images/994652ee43083019d4be515be09da064c3aa67f81ac026118a4e08fbf465a239.jpg) + +![](images/e654a35bc7c766314fad7c929a8f446db8507bac22b1e21bff53076cc67193de.jpg) + +![](images/8c5146151ad4384e5a1760f5c5636843a608be79cfd3c5da55fd065bb8c8129f.jpg) + +![](images/86493d112cadd76d5ffdf22b70498c898eac0adc23229436d8ea0281a32b32c1.jpg) + +![](images/a2e5f6d1da725eacad35c8bddf6c4afb89b654619651969ead4300c7356acfa2.jpg) +Figure 1. Effects of signal nonlinearity on the signal's Fourier transform. (a) A signal with a linear modulation distribution Fourier transform to (b) a narrow peak. (c) If the nonlinearity of the modulation is of the second order, (d) the peak gets broadened. (e) If the nonlinearity is of the third order, (f) the peak gets distorted. In OCT, this will result in deteriorated image quality. (g) In practice, both nonlinearities appear at once, resulting in (h) the peak being elongated in the direction of the third-order distortion. + +![](images/9b34c1c2f3bda77e01e581f952567a4d157e3a5c92bf230b8d329d074952c464.jpg) + +the other hand, it makes room for errors related to the limitations of Fourier transformation. One such limitation is the Fourier transformation's sensitivity to nonlinearities in the input signal. In the simplest case of a signal being a single-frequency modulation, a second-order nonlinear distribution will Fourier transform to a broader peak (compare Fig. 1a,b with c,d). If this distribution is of a third-order nonlinear character, the peak in the Fourier transform will get distorted (Fig. 1e,f). In practice, both nonlinearities appear at once, the second-order one bigger than the third-order one, leading to a broad peak elongated in the direction of the third-order nonlinearity distortion (Fig. 1g,h). + +The influence of nonlinearities is best visible in OCT, where the Fourier transform of the raw signal constitutes one line of the image. Since the width of the peak represents the depth resolution (i.e. how small a detail can be observed), and the peak's shape correlates with the image quality, nonlinearities are generally thought of as detrimental and various techniques are devised to remove them [1, 2]. These techniques do not achieve absolute removal of nonlinearity, preventing single-shot imaging of bulk objects, such as the entire eye [3, 4] and consequently necessitating elaborate workaround solutions [5-8]. A Machine + +Learning solution for nonlinearity removal was proposed as well [9] – although robust and reliable, it is biased towards specific kinds of objects with specific nonlinear properties. + +On the other hand, the inherent optical properties that produce nonlinearities in the imaged objects have been shown to enable the characterisation of the objects [10-12] and correlate the change of their nonlinear character with the progression of diseases [13]. Unfortunately, the current methods for extracting the second-order nonlinearity are very object-specific as well as error-prone preventing them from being applied in real-world scenarios. + +We present an approach that enables flexible and comfortable management of the signal nonlinearities, letting the user study one order in the absence of the others, a capability previously impossible to achieve using current standard or Machine Learning methods. In our approach, for each nonlinearity order, a neural network is trained whose input is a special two-dimensional data structure, a stack. Each row of the stack represents an amplitude of the Fourier transform of an input signal multiplied by a different nonlinearity-correcting factor. Consequently, it contains nonlinearity-free regions which the network identifies and uses to build a nonlinearity-free Fourier transform amplitude. We present two networks, one to remove the second-order nonlinearity while keeping the third-order one intact, and the second network to remove the third-order nonlinearity while keeping the second-order one intact. We test our approach on OCT signals that naturally contain substantial amounts of second-order nonlinearity and small amounts of third-order nonlinearity. The tests on both computer-generated and experimental signals indicate the robustness of our approach. + +We start by describing our approach in Section 2, including the creation of the stacks in Subsection 2.1, their application in the training processes in Subsection 2.2, presentation of the used architecture and methodology in Subsection 2.3, and the description of the datasets (Subsection 2.4) used for training of our models. In Section 3, we report the performance of the networks in removing the respective nonlinearity, complete with the analysis of the training process and the subsequent tests on computer-generated data (Subsection 3.1) and experimental data (Subsections 3.2, 3.3 and 3.4). In the final sections, we summarise the obtained results (Section 4), pointing to the strong points of our approach, identifying the observed and potential shortcomings, and discussing potential solutions and the scope of future work (Section 5). + +# 2. Concept + +# 2.1. Amplitude stacks as inputs + +Our networks rely on a two-dimensional data structure as inputs: a stack (Fig. 2). The rows of a stack are created + +![](images/a16dbcc707930011ceab6b8624b3f09ee051c45d8fa50269d9bbf99f0758b4ab.jpg) +(a) +(c) + +![](images/2c1c30731d0abafa704dbb5a36e90f55a1ff49d6b1a8911bb320c5c3b121e37f.jpg) +(b) + +![](images/97def600b40e5ba6e9df3cd0024dc57242b310ca75f8f7bf4948020d928c7f75.jpg) + +![](images/ac0c70b4f65a46e82a7e214d19f98664a95e42d451f10d7b5444245b8ed4d0f9.jpg) +(d) + +![](images/ab7a219ead3eadb8a1081c5f6d5f309fb6fea3665d9b981175de4c8830e8fad7.jpg) +(e) +Figure 2. Calculation of the input stacks. (a) The raw signal is (b) multiplied by 32 different complex exponents representing 32 different levels of the second-order nonlinearity: $C_0, C_1, C_2, \ldots, C_{31}$ . Each row is then Fourier transformed. (c) The 2nd-order stack is created from the absolute value of the complex Fourier transforms. Similarly, (d) the raw signal is multiplied by 32 different complex exponents representing 32 different levels of the third-order nonlinearity: $D_0, D_1, D_2, \ldots, D_{31}$ and then Fourier transformed to build (e) a 3rd-order stack. + +in the following way. First, the raw signal is multiplied (Fig. 2a) by a complex exponent with a different quadratic argument in the case of the stack used for removing the second-order nonlinearity (Fig. 2b) or with a different cubic argument in the case of the stack for the third-order nonlinearity removal (Fig. 2d). Coefficients in the quadratic arguments, $(C_0,\dots,C_{31}$ in Fig. 2b), and coefficients in the cubic arguments $(D_0,\dots,D_{31}$ in Fig. 2d), are chosen to cover a wide range of the second- and third-order nonlinearity, ensuring that the nonlinearities present in the signal are fully compensated. Then, each row is Fourier transformed, and an absolute value is calculated to obtain a 2nd-order stack (Fig. 2c) or a 3rd-order stack (Fig. 2e). The range of the nonlinear coefficients is symmetrical around 0 and broad enough to ensure that each modulation's nonlinearity in the raw signal is compensated. In the 2nd-order stack, these will be the places with the smallest width (marked 1, 2, 3 and 4 in Fig. 3b) and in the 3rd-order stack, the places containing no distortions (marked 1, 2, 3 and 4 in Fig. 3c). + +![](images/f782900b0485a1b983e5c1474a319e8adffe2eef9e06b102fec9a952f84f0124.jpg) +Figure 3. (a) The Fourier transform amplitude of a raw signal, presented in Fig. 2a, shows four peaks affected by a different amount and type of nonlinearity. The second-order nonlinearity is removed with Network #1 trained on (b) the 2nd-order stack as input and (d) Fourier transform amplitude free of the second-order nonlinearity as output. Similarly, the third-order nonlinearity is removed with Network #2 trained on (b) the 3rd-order stack as input and (d) Fourier transform amplitude free of the third-order nonlinearity as output. As a result, (f) the second-order nonlinearity is removed leaving the third-order one (see peak 4) using Network #1, (g) the third-order nonlinearity is removed leaving the second-order one intact (see peak 4) using Network #2. + +# 2.2. Nonlinearity-removing networks + +The approach for removing the second-order nonlinearity and the approach for removing the third-order nonlinearity are similar, presented schematically in Fig. 3. The input of the network for the second-order nonlinearity removal, Network #1, is the 2nd-order stack (Fig. 3b) and the output is a Fourier transform amplitude free of second-order nonlinearity (Fig. 3d). Because the input 2nd-order stack contains regions corresponding to perfectly removed second-order nonlinearity, the training process can be viewed as learning how to identify these regions and then stitching them all up. Similarly, the network for removing third-order nonlinearity, Network #2, takes the 3rd-order stack as input (Fig. 3c) and outputs a Fourier transform amplitude free + +![](images/c182229d79c02093451922524225c06d0d70644057fd459307819d87e3aecfa5.jpg) +Figure 4. Architecture of our nonlinearity-removing networks. + +from the 3rd-order nonlinearity (Fig. 3e). Input for both networks has a shape of 32 by 1024 pixels with a single channel whereas output is a vector of size 1024. When compared to the Fourier transform amplitude of the raw signal (Fig. 3a), the signal returned by Network #1 (Fig. 3f) incorporates only the third-order nonlinearity, and the signal returned by Network #2 (Fig. 3g) contains only the second-order nonlinearity. + +# 2.3. Methodology + +Our nonlinearity-removing networks are inspired by the U-Net architecture [14]. U-Net stands out for its use of a contracting path to capture contextual information, an expanding path for precise feature localisation, and the incorporation of skip connections that preserve fine-grained details, particularly beneficial for precise object delineation. We leverage these capabilities to identify and extract nonlinearity-related features from the stacks. + +While the classical U-Net deals with 2D data structures for both input and output, in our case, we perform dimension reduction to output 1D signals by incorporating a max-pooling layer. We further refined the architecture by optimising hyperparameters with the validation dataset using Optuna [15]. As a result of the tuning, we adopted the Mean Absolute Error (MAE) loss function and the Adam optimiser [16] with a set learning rate of 0.0002. Our final network architecture is presented in Fig. 4. + +# 2.4. Datasets + +Our datasets consist of noiseless computer-generated signals that are Gaussian-shaped, and diverse and randomised in terms of their content. If put in the OCT-imaging-related terms, they represent a broad range of possible imaged object configurations. These objects are simulated to have a random number of interfaces (from 2 to 12) positioned at arbitrary locations, with each object layer exhibiting varying random reflectivity and second- and third-order nonlin + +![](images/6ba29dd818ac660451fabe13244685a6b03cf21effed3cba4580072f9d54318b.jpg) +Figure 5. Network #1 and Network #2 performance measured with the Mean Absolute Error (MAE) loss function across 30 epochs of training. Blue lines depict Network #1, while orange lines showcase the Network #2 performance. Dotted vertical lines highlight epochs exhibiting the best performance. + +earity values. These object parameters are used to synthesise modulatory signals equivalent to those acquired in an OCT measurement (see example signals in Fig. 1a, c, e, g or Fig. 2a). Next, the input stacks and outputs, the Fourier transform amplitudes of signals free of the respective nonlinearity, are calculated. + +The risk of overfitting was minimised by using the training datasets with an optimised size [17]. In our research, this was found to be between 160,000 and 240,000 objects. The validation and test datasets include 14,993 objects. The distribution is uniform across the objects with a specific number of interfaces, resulting in an equal allocation of 1,363 data samples for each distinct object type. + +Before starting the training, as part of the preprocessing stage, both the input and output data were normalised through Min-Max scaling, ensuring that values fell within the standardised range of 0 to 1. + +# 3. Performance + +# 3.1. In silico data tests + +Each nonlinearity-removing neural network underwent training using a training dataset consisting of 200,000 objects. Within each mini-batch, 8 input stacks were processed. The proposed algorithm requires 39 ms to process a stack on a Quadro RTX 6000 GPU. Network #1 (Fig. 5, blue lines) and Network #2 (Fig. 5, orange lines) both displayed convergence within just a few epochs (Fig. 5) and were trained until the thirtieth epoch. + +To better evaluate the networks and results, we employed the Goodness-of-Fit Approximity (GoF) metric introduced in [12]. The GoF metric assesses the accuracy of predictions by calculating the percentage of values that lie within a predefined distance from the corresponding ground truth values. For this GoF metric, we define an error threshold as a ratio between the predefined distance and the value range + +![](images/2b01f51ee0a3c226f49bad1848ea6588ba9e68bd73cf0c597b6880e5ba54f49d.jpg) +Figure 6. Networks performance assessed through the metric of GoF, measuring the degree to which predictions align with the ground truth within $0.1\%$ error in the test dataset comprising 14,993 data samples for (a) Network #1 and (b) Network #2. The graphs at the top (a1,a2) are zoomed-in areas of the top area of (b1,b2) the whole graphs presented. + +span. In our case, where the values range from 0 to 1, the distance of 0.01 means $1\%$ error threshold. + +Based on the post-training evaluation using a $1\%$ error threshold and the test dataset, Network #1 started to overfit at epoch 25 (Fig. 5, indicated by the blue dotted vertical line). In contrast, Network #2 demonstrated resilience against overfitting and achieved its best results in the 28th epoch (Fig. 5, marked by the red dotted vertical line). + +Using the test dataset and GoF with a $1\%$ error threshold, both networks achieved a remarkable mean GoF value of $99.98\%$ , which means that $99.98\%$ of the predictions differed from the ground truth by less than $1\%$ , with only one example below $95\%$ GoF. In light of this, to be able to carry out a meaningful analysis, the error threshold was further reduced to $0.1\%$ and a boxplot graph was plotted for each model in Fig. 6. + +Both networks, Network #1 (Fig. 6a) and Network #2 (Fig. 6b), demonstrated very good performance even for error threshold $0.1\%$ . Only approximately $9.52\%$ and $3.07\%$ of the predictions from Network #1 and Network #2, respectively, fell below the $95\%$ GoF for such a small error, as showed in Tab. 1 that also includes GoF values for objects with distinct number of layers. Fig. 7 offers an in-depth analysis of the worst-predicted examples for the Network #1 (Fig. 7a) and the Network #2 (Fig. 7b) showcasing GoF values of $65.5\%$ and $45.5\%$ , respectively. Predictions of peak amplitudes in the nonlinearity-free Fourier transform amplitude (Fig. 7(3), orange line) closely align the expected outcomes (blue line) deviating only by the error threshold level. Also, the networks predict values close to + +![](images/4bb2ede470abb4d7063722526af6ed5e95e5fd76a7ad78f7b5a87bbcc63ddd95.jpg) +Figure 7. Predictions with very low GoF for $0.1\%$ error threshold for (a) Network #1 and (b) Network #2. Based on the input stack (1), models predict (2, orange line, PR) amplitudes free of the respective nonlinearity that exhibit a strong correlation with the ground truth (2, blue line, GT). (3) Nonlinearity-free peaks are recovered with high fidelity, but (4) the models sometimes struggle with zero-valued places. + +zero rather than precisely zero, contributing to worse GoF metric results. One should keep in mind here that these discrepancies are extremely low: at the level of around 2e-3 as presented in Fig. 7a4 and b4. + +# 3.2. Experimental signals: mirror at different depths + +Networks were tested on the signals obtained through an OCT imaging of a mirror placed at 11 different depth positions. Each acquired signal Fourier transforms to a peak whose amount of second-order-nonlinearity-related broadening and third-order-nonlinearity-related distortion increases with depth (Fig. 8a). Due to their simultaneous appearance, the third-order nonlinearity manifests itself in the form of peak elongation. The depth-dependent deterioration is due to the fundamental limitations of OCT's spectrometer-based light detection. Additionally, the second- and third-order nonlinearities originating from the OCT's interferometer add up - depth-independently - to that effect, further worsening the outcome. Typically, + +Table 1. The number of cases where GoF with error threshold ${0.1}\%$ falls below ${95}\%$ for objects with a specific number of reflective layers. + +
Net.Interfaces23456789101112Totals
#1# below 95%3842601931381058779554544371381
# in dataset1363136313631363136313631363136313631363136314993
% below28.1719.0814.1610.127.706.385.804.043.303.322.719.52
#2# below 95%10782575438323122131311460
# in dataset1363136313631363136313631363136313631363136314993
% below7.856.024.183.962.792.352.271.610.950.950.813.07
+ +![](images/5f1e9e381f89c8ff7b700fe887e4032e6521a8952802db1eda91af5e979aae43.jpg) + +![](images/5d0ee500d14f3736bdb0d46ea3ec54d1ba04ec0b079b13b39a99eb207f80b209.jpg) + +![](images/2620ec16a20ec7e8d3b0ad08f3edc044a095f704114d0f43f055873e201fabcb.jpg) + +![](images/e928aad4622d3ffb33a2b0c9c743c1e900eeab944af0122a2bb452dd9a9b981e.jpg) +Figure 8. Mirror at 11 different depths. (a) Fourier transforming raw signals results in peaks that are both broadened by the second-order nonlinearity and distorted by the third-order nonlinearity. The deeper the peak, the bigger the nonlinearity effect, a behaviour characteristic of OCT signals acquired with a spectrometer. (b) Network #1 removes the second-order nonlinearity, leaving the third-order one intact. (c) Network #2 removes the third-order nonlinearity, leaving the third-order one intact. (d, e) The traditional linearisation (grey line) is not as precise as our approach (green lines). + +![](images/3ab486d8fc1f593adf72fef4b6ad47350d57db4cb81d0209bbc379694d983211.jpg) + +spectrometer- and interferometer-related nonlinearities can be satisfactorily removed in a calibration step before the measurement. + +The application of Network #1 removes the second-order nonlinearity in its entirety and consequently, provides a dramatic reduction of peaks' widths (Fig. 8b). Because the third-order nonlinearity remains intact, the peaks' shapes incorporate the characteristic "tail". On the other hand, Network #2 removes the third-order nonlinearity, leading to a complete reduction of the elongation (Fig. 8c). + +One observes the superiority of our approach when the peaks obtained with Network #1 are compared to the peaks obtained as a result of traditional linearisation methods which remove all nonlinearity orders. Here, we used the method in [1] where two raw signals corresponding to the mirror placed at two different locations are used. First, their phases are subtracted to remove the constant interferometer-related contribution, enabling the calculation of the correction vector for spectrometer-related nonlinearities. Once the spectrometer's nonlinearity is removed, one of the initial spectra is used to generate an interferometer-nonlinearities-compensating vector. The peaks located close to 0 depth index obtained using our approach and the standard linearisation method (Fig. 8d, green and grey lines) are similar in shape: the network-outputted peak is higher which is very favourable - it means a higher signal-to-noise ratio, which is associated with better quality imaging. One observes a small "tail" at the side, most probably due to a small amount of third-order nonlinearity present at such shallow depths, whereas the peak obtained using the traditional methods is a smooth Gaussian. More interestingly, for a peak at a substantial depth (Fig. 8e), where the third-order nonlinearity is naturally much bigger (let us remind that this depth-increasing nonlinearity is tied to OCT spectrometer), the "tail" is much more pronounced in case of our network but not present in case of the traditional method. In the latter case, the peak is elongated which suggests the presence of both second- and third-order nonlinearities. This leads to the conclusion that traditional linearisation, although very satisfactory in practical applications, still has room for improvement. In fact, despite 30 years of OCT in science and industry, new methods for more precise OCT data linearisation are still being developed. + +# 3.3. Glass + +Another test was performed on data corresponding to a $300\mu \mathrm{m}$ thick sapphire glass put at an angle obtained with the same instrument. This data consisted of 256 raw OCT signals which represent a different transversal location on the glass. When Fourier transformed one after another, they form a 2D depth image of the object (Fig. 9a, 1 and 2 denote the front and back surface of the glass, the third line is an artefact placed at the depth equal to the thickness of the glass, inherent to OCT). Without linearisation, the image elements are broadened the deeper they are. When the + +![](images/db05d6db814cd63f778520fd8cbf79e2753d3aca766d02d665146c796554e892.jpg) + +![](images/29b29f391d403a08ee4c1fb4296adb274529a948df5f08f0247909f1669380e9.jpg) + +![](images/3c3bc75914cc1ad76b547fa132501ec56d85bbcd9eab0f85bdfb679050fd43b7.jpg) + +![](images/b06b3c978a03e9edc2508a382adca684a7507d4598e03555fd20d3dc58c5f8d7.jpg) + +![](images/5832df6ccea4e3ea07b2dd0e74e603d312d1e31cd18c06eb847279c016401075.jpg) + +![](images/d199ff707b3a1e13fa742a2a2e0da1df1dd4a3f1bd2c4dba222464fb32f3630b.jpg) +Figure 9. Sapphire glass. (a) Fourier transformation of raw signals gives an OCT image where the quality deteriorates with depth. (b) Network #1 removes the second-order nonlinearity which accounts for the most of quality deterioration, (c) Due to a small amount of the third-order nonlinearity, the application of Network #2 does not improve the image quality. (d), Image line #152 from Network #1 (green line) is compared with its raw image equivalent (grey line) to show dramatic peaks' width reduction, (e) The image line #152 from Network #2 output (green line) and its raw image equivalent (grey line) show no significant difference except for a more favourable shape for the former. (f) Network #1 (green line) performs quite well when compared to the traditional method output (black line). 1 and 2 mark the front and back surface of the imaged glass, the middle line is an artefact. + +same data is processed with Network #1, the broadening is removed from all the image elements (see the image in Fig. 9b and the 152nd image line in Fig. 9d). Due to very small amounts of the third-order nonlinearity, Network #2 does not bring perceptible improvements (Fig. 9c), but under a closer inspection of a single image line (Fig. 9e), one can observe some improvement in the peak height and a possible side elongation reduction. + +Fig. 9f compares the output of Network #1 (green line) and the traditional linearisation method (black line), indicating that this model could be a satisfactory alternative for the traditional image calculation algorithm when the third-order nonlinearities are negligible. It can be used as such an alter + +native for the OCT laboratory instrument which was used here to obtain the data and which is nevertheless a fairly good representative of a standard OCT instrument. We note that Network #1 improved the shape of the artefact (the middle peak in Fig. 9f), which - although generally irrelevant for OCT imaging - is still something noteworthy as such a feat is beyond the capabilities of traditional linearisation methods, showing that our approach, unlike the traditional ones, removes the nonlinearities nondiscriminatively. + +# 3.4. Grape and cucumber + +The final tests were performed on the data corresponding to a grape (Fig. 10) and a cucumber (Fig. 11). The former represents a high signal-to-noise ratio (highly visible modulation in the signal as depicted in Fig. 10a) and the latter a low signal-to-noise ratio (poorly visible modulation in the signal as depicted in Fig. 11a). + +Fourier transformation of raw signals gives images of a low quality that further deteriorates with depth (Fig. 10b, Fig. 11b). Again, the improvement obtained when applying Network #1 (Fig. 10c, Fig. 11c) is comparable to that obtained with traditional methods (Fig. 10d, Fig. 11d). Under closer inspection (Fig. 10e, Fig. 11e), one notices that Network #1 output is less detailed for the higher signal-to-noise grape data - most probably due to the remaining third-order nonlinearities - and loses some details for the lower signal-to-noise ratio cucumber data. The latter loss is the direct consequence of training data representing predominantly high signal-to-noise ratio signals. + +Just as in the case of the glass data, due to small amounts of third-order nonlinearity, Network #2 did not provide visually significant improvement and therefore, was omitted. + +# 4. Summary and discussion + +Two models were demonstrated: one for removing second-order nonlinearity from digital signals and leaving the third-order one intact, and another one for removing the third-order nonlinearity but leaving the second-order one intact. Their nonlinearity-removing capabilities were extensively analysed using computer-generated signals as well as experimental signals acquired with a laboratory Optical Coherence Tomography (OCT) device. Tests on the OCT signals, which corresponded to objects with varying levels of complication: mirror, sapphire, grape and cucumber, allowed us to evaluate the networks in terms of their applicability in a world-relevant technology. Whereas the tests and analysis done on the computer-generated signals revealed extraordinary performance in terms of similarity of the ground truth and prediction signals, the tests on OCT-acquired data indicated the networks' advantageous operation when compared to approaches traditionally used in OCT. + +During training, both networks converged exceptionally fast and reached remarkable loss function values. Nevertheless + +![](images/3e91691d4c6ee5a8513471026196bc126a65485cb245696d7a3e87b0a4c6aa30.jpg) + +![](images/11cef7ae94bbb341424bf00dabf5797ce05db10ee810bf72a9a19dd6b73afea2.jpg) + +![](images/ff57d1cd76565e64994e4471ce3345eff28d89eb9e775d37f069acc6a2163995.jpg) + +![](images/cfb73ce259b04d136b209571ff4afd7e95e70ca39b85bd5278c4fcc5ea31b04b.jpg) + +![](images/5fc46eb2c470196332d40602656d8a40560b948ae526f2e6ee8b773941028d93.jpg) +Figure 10. Grape. (a) One of the raw signals representing a single line in the image of a grape. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image when compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees very well with its equivalent from the image obtained using traditional methods. + +less, we have identified possible improvements: the closer inspection of the predictions revealed non-zero values that are zero-valued areas in the ground-truth signals. Another current shortcoming of both networks is their time of processing an entire OCT image; consequently, for now, the processing speed is insufficient for the real-time processing of data coming straight from OCT devices. + +# 5. Future work + +The future work will mainly consist in developing an approach that enables one to use both networks in a sequence so that both the second- and third-order nonlinearity removal is achieved. Since, as was shown here, the networks remove the nonlinearities completely, their sequential use will provide a very valuable tool for extremely precise data linearisation, especially attractive for OCT where any un + +![](images/62313f48081c7d2fb57f479ede3d9b1be037e21aee669caf260aace6251bb469.jpg) + +![](images/0f91b7b15067482bb19a0dc332343e88654fea8d32a1418a4276c70135d4a55c.jpg) + +![](images/108a614d87b771b1ecca5c1e048a8402bddcf6819fbad750134be4c6db4038a4.jpg) + +![](images/a122f416a3da4332676b271f026a756ad0aae55a4e44d28b77efac3cdaeacc00.jpg) + +![](images/37755de439c2620b9482844e69b68ef5b17823c6d4c1ec5ddc693267d4de479d.jpg) +Figure 11. Cucumber. (a) One of the raw signals representing a single line in the image of a cucumber. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees well with its equivalent from the image obtained using traditional methods. + +compensated nonlinearities lead to the quality drop. + +Also, in terms of the network performance, we will try to remove the discrepancy for zero-valued elements in the ground truth and prediction signals by adding extra convolutional layers before the output to filter out the unwanted values as well as performing additional hyperparameter tuning. Additionally, we plan to speed up the processing of the stacks by lowering the amount of calculations and the size of the networks. + +Acknowledgements. Supported by the New Zealand Ministry of Business, Innovation and Employment (MBIE), Smart Ideas grant "Extending the Boundaries of Digital Signal Processing: AI-powered Fourier Transformation Alternative (E7943)". We also thank Oliver Batchelor, François Bissey, and Piotr Kolenderski for sharing their computational resources. + +# References + +[1] Kai Wang and Zhihua Ding. Spectral calibration in spectral domain optical coherence tomography. Chinese Optics Letters, 6(12):902-904, 2008. 2, 6 +[2] Gongpu Lan and Guoqiang Li. Design of ak-space spectrometer for ultra-broad waveband spectral domain optical coherence tomography. Scientific reports, 7(1):42353, 2017. 2 +[3] Ireneusz Grulkowski, Silvestre Manzanera, Lukasz Cwiklinski, Franciszek Sobczuk, Karol Karnowski, and Pablo Artal. Swept source optical coherence tomography and tunable lens technology for comprehensive imaging and biometry of the whole eye. Optica, 5(1):52-59, 2018. 2 +[4] Timothy R Hillman and David D Sampson. The effect of water dispersion and absorption on axial resolution in ultrahigh-resolution optical coherence tomography. Optics express, 13 (6):1860-1874, 2005. 2 +[5] Siyu Chen, Benjamin Potsaid, Yan Li, Junhong Lin, Yun- chan Hwang, Eric M Moult, Jason Zhang, David Huang, and James G Fujimoto. High speed, long range, deep penetration swept source oct for structural and angiographic imaging of the anterior eye. Scientific reports, 12(1):992, 2022. 2 +[6] Jiangjie Huang, Jinyu Fan, Yi He, and Guohua Shi. Physical compensation method for dispersion of multiple materials in swept source optical coherence tomography. Journal of Biophotonics, 16(10):e202300167, 2023. +[7] Daniel L Marks, Amy L Oldenburg, J Joshua Reynolds, and Stephen A Boppart. Digital algorithm for dispersion correction in optical coherence tomography for homogeneous and stratified media. Applied optics, 42(2):204-217, 2003. +[8] Liuhua Pan, Xiangzhao Wang, Zhongliang Li, Xiangyang Zhang, Yang Bu, Nan Nan, Yan Chen, Xuan Wang, and Fengzhao Dai. Depth-dependent dispersion compensation for full-depth oct image. Optics express, 25(9):10345-10354, 2017. 2 +[9] Shaiban Ahmed, David Le, Taeyoon Son, Tobiloba Adejumo, Guangying Ma, and Xincheng Yao. Asc-net: an opensource deep learning network for automated dispersion compensation in oct. Frontiers in Medicine, 9:864879, 2022. 2 +[10] Sylwia M Kolenderska, Bastian Brauer, and Frédérique Vanholsbeeck. Dispersion mapping as a simple postprocessing step for fourier domain optical coherence tomography data. Scientific reports, 8(1):9244, 2018. 2 +[11] Christos Photiou and Costas Pitris. Comparison of tissue dispersion measurement techniques based on optical coherence tomography. Journal of biomedical optics, 24(4):046003-046003, 2019. +[12] Krzysztof A Maliszewski, Magdalena A Urbańska, Piotr Kolenderski, Varvara Vetrova, and Sylwia M Kolenderska. Extracting group velocity dispersion values using quantum-mimic optical coherence tomography and machine learning. Scientific Reports, 13(1):6596, 2023. 2, 4 +[13] Christos Photiou, Evgenia Bousi, Ioanna Zouvani, and Costas Pitris. Using speckle to measure tissue dispersion in optical coherence tomography. Biomedical optics express, 8 (5):2528-2535, 2017. 2 + +[14] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4 +[15] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623-2631, 2019. 4 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4 +[17] Jayme Garcia Arnal Barbedo. Impact of dataset size and variety on the effectiveness of deep learning and transfer learning for plant disease classification. Computers and electronics in agriculture, 153:46-53, 2018. 4 \ No newline at end of file diff --git a/2024/Selective Nonlinearities Removal from Digital Signals/images.zip b/2024/Selective Nonlinearities Removal from Digital Signals/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f4705f7102a089aedcafb37bfeeb996fe415937e --- /dev/null +++ b/2024/Selective Nonlinearities Removal from Digital Signals/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08f2cfc9e565a88779f56f900530d4d6d3e3d14c5b04ab877b2f58345c3224de +size 719809 diff --git a/2024/Selective Nonlinearities Removal from Digital Signals/layout.json b/2024/Selective Nonlinearities Removal from Digital Signals/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..48f382e0a0f7ca280c2a33e85ebbbf4be914df6e --- /dev/null +++ b/2024/Selective Nonlinearities Removal from Digital Signals/layout.json @@ -0,0 +1,6156 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 103, + 457, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 103, + 457, + 121 + ], + "spans": [ + { + "bbox": [ + 137, + 103, + 457, + 121 + ], + "type": "text", + "content": "Selective nonlinearities removal from digital signals" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 144, + 449, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 144, + 449, + 199 + ], + "spans": [ + { + "bbox": [ + 143, + 144, + 449, + 199 + ], + "type": "text", + "content": "Krzysztof A. Maliszewski \nSchool of Mathematics and Statistics, University of Canterbury \nChristchurch, New Zealand \nkrzysztof.maliszewski@pg.canterbury.ac.nz" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 205, + 447, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 205, + 447, + 246 + ], + "spans": [ + { + "bbox": [ + 146, + 205, + 447, + 246 + ], + "type": "text", + "content": "Magdalena A. Urbańska \nSchool of Food and Advanced Technology, Massey University Palmerston North, New Zealand" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 251, + 357, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 251, + 357, + 260 + ], + "spans": [ + { + "bbox": [ + 235, + 251, + 357, + 260 + ], + "type": "text", + "content": "murbanska@massey.ac.nz" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 267, + 449, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 267, + 449, + 308 + ], + "spans": [ + { + "bbox": [ + 143, + 267, + 449, + 308 + ], + "type": "text", + "content": "Varvara Vetrova \nSchool of Mathematics and Statistics, University of Canterbury \nChristchurch, New Zealand" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 312, + 383, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 312, + 383, + 322 + ], + "spans": [ + { + "bbox": [ + 209, + 312, + 383, + 322 + ], + "type": "text", + "content": "varvara.vetrova@canterbury.ac.nz" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 329, + 461, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 461, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 461, + 384 + ], + "type": "text", + "content": "Sylwia M. Kolenderska \nInstitute of Physics, Faculty of Physics, Astronomy and Informatics \nNicolaus Copernicus University in Torun \nul. Grudziadzka 5, 87-100 Torun, Poland" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 385, + 462, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 462, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 462, + 399 + ], + "type": "text", + "content": "School of Physical and Chemical Sciences, University of Canterbury" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 399, + 363, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 399, + 363, + 411 + ], + "spans": [ + { + "bbox": [ + 228, + 399, + 363, + 411 + ], + "type": "text", + "content": "Christchurch, New Zealand" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 227, + 415, + 364, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 415, + 364, + 425 + ], + "spans": [ + { + "bbox": [ + 227, + 415, + 364, + 425 + ], + "type": "text", + "content": "skol745@aucklanduni.ac.nz" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 454, + 192, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 454, + 192, + 467 + ], + "spans": [ + { + "bbox": [ + 143, + 454, + 192, + 467 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "content": "Many instruments performing optical and non-optical imaging and sensing, such as Optical Coherence Tomography (OCT), Magnetic Resonance Imaging or Fourier-transform spectrometry, produce digital signals containing modulations, sine-like components, which only after Fourier transformation give information about the structure or characteristics of the investigated object. Due to the fundamental physics-related limitations of such methods, the distribution of these signal components is often nonlinear and, when not properly compensated, leads to the resolution, precision or quality drop in the final image. Here, we propose an innovative approach that has the potential to allow cleaning of the signal from the nonlinearities but most of all, it now allows to switch the given order off, leaving all others intact. The latter provides a tool for more in-depth analysis of the nonlinearity-inducing properties of the investigated object, which can lead to applications in early disease detection or more sensitive sensing of chemical compounds. We consider OCT signals and nonlinear" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 455, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 576 + ], + "type": "text", + "content": "ities up to the third order. In our approach, we propose two neural networks: one to remove solely the second-order nonlinearity and the other for removing solely the third-order nonlinearity. The input of the networks is a novel two-dimensional data structure with all the information needed for the network to infer a nonlinearity-free signal. We describe the developed networks and present the results for second-order and third-order nonlinearity removal in OCT data representing the images of various objects: a mirror, glass, and fruits." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Optical Coherence Tomography (OCT), Magnetic Resonance Imaging (MRI) or Fourier-transform Spectrometry (FTS) belong to a class of methods where useful, interpretable signal is obtained in an indirect way, more specifically by Fourier transforming the raw signal acquired by the instrument and taking its absolute value. On the one hand, such an indirect approach makes OCT-based imaging and FTS-based sensing faster and enables MRI, but on" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25028" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 182, + 131 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 182, + 131 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 182, + 131 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 182, + 131 + ], + "type": "image", + "image_path": "048039b52eca5292a5bd2bb86f3c0b5b328a9d8ae45e104c6bbef271fe54bd27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 70, + 267, + 132 + ], + "blocks": [ + { + "bbox": [ + 182, + 70, + 267, + 132 + ], + "lines": [ + { + "bbox": [ + 182, + 70, + 267, + 132 + ], + "spans": [ + { + "bbox": [ + 182, + 70, + 267, + 132 + ], + "type": "image", + "image_path": "be826ae52a302df15bba60a416d8760e8e10906ce0eb7c235ab1aa530d22e642.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 65, + 133, + 182, + 195 + ], + "blocks": [ + { + "bbox": [ + 65, + 133, + 182, + 195 + ], + "lines": [ + { + "bbox": [ + 65, + 133, + 182, + 195 + ], + "spans": [ + { + "bbox": [ + 65, + 133, + 182, + 195 + ], + "type": "image", + "image_path": "994652ee43083019d4be515be09da064c3aa67f81ac026118a4e08fbf465a239.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 133, + 266, + 196 + ], + "blocks": [ + { + "bbox": [ + 182, + 133, + 266, + 196 + ], + "lines": [ + { + "bbox": [ + 182, + 133, + 266, + 196 + ], + "spans": [ + { + "bbox": [ + 182, + 133, + 266, + 196 + ], + "type": "image", + "image_path": "e654a35bc7c766314fad7c929a8f446db8507bac22b1e21bff53076cc67193de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 66, + 200, + 182, + 262 + ], + "blocks": [ + { + "bbox": [ + 66, + 200, + 182, + 262 + ], + "lines": [ + { + "bbox": [ + 66, + 200, + 182, + 262 + ], + "spans": [ + { + "bbox": [ + 66, + 200, + 182, + 262 + ], + "type": "image", + "image_path": "8c5146151ad4384e5a1760f5c5636843a608be79cfd3c5da55fd065bb8c8129f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 182, + 200, + 266, + 262 + ], + "blocks": [ + { + "bbox": [ + 182, + 200, + 266, + 262 + ], + "lines": [ + { + "bbox": [ + 182, + 200, + 266, + 262 + ], + "spans": [ + { + "bbox": [ + 182, + 200, + 266, + 262 + ], + "type": "image", + "image_path": "86493d112cadd76d5ffdf22b70498c898eac0adc23229436d8ea0281a32b32c1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 66, + 266, + 182, + 327 + ], + "blocks": [ + { + "bbox": [ + 66, + 266, + 182, + 327 + ], + "lines": [ + { + "bbox": [ + 66, + 266, + 182, + 327 + ], + "spans": [ + { + "bbox": [ + 66, + 266, + 182, + 327 + ], + "type": "image", + "image_path": "a2e5f6d1da725eacad35c8bddf6c4afb89b654619651969ead4300c7356acfa2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 287, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 426 + ], + "type": "text", + "content": "Figure 1. Effects of signal nonlinearity on the signal's Fourier transform. (a) A signal with a linear modulation distribution Fourier transform to (b) a narrow peak. (c) If the nonlinearity of the modulation is of the second order, (d) the peak gets broadened. (e) If the nonlinearity is of the third order, (f) the peak gets distorted. In OCT, this will result in deteriorated image quality. (g) In practice, both nonlinearities appear at once, resulting in (h) the peak being elongated in the direction of the third-order distortion." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 182, + 266, + 266, + 328 + ], + "blocks": [ + { + "bbox": [ + 182, + 266, + 266, + 328 + ], + "lines": [ + { + "bbox": [ + 182, + 266, + 266, + 328 + ], + "spans": [ + { + "bbox": [ + 182, + 266, + 266, + 328 + ], + "type": "image", + "image_path": "9b34c1c2f3bda77e01e581f952567a4d157e3a5c92bf230b8d329d074952c464.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 581 + ], + "type": "text", + "content": "the other hand, it makes room for errors related to the limitations of Fourier transformation. One such limitation is the Fourier transformation's sensitivity to nonlinearities in the input signal. In the simplest case of a signal being a single-frequency modulation, a second-order nonlinear distribution will Fourier transform to a broader peak (compare Fig. 1a,b with c,d). If this distribution is of a third-order nonlinear character, the peak in the Fourier transform will get distorted (Fig. 1e,f). In practice, both nonlinearities appear at once, the second-order one bigger than the third-order one, leading to a broad peak elongated in the direction of the third-order nonlinearity distortion (Fig. 1g,h)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "The influence of nonlinearities is best visible in OCT, where the Fourier transform of the raw signal constitutes one line of the image. Since the width of the peak represents the depth resolution (i.e. how small a detail can be observed), and the peak's shape correlates with the image quality, nonlinearities are generally thought of as detrimental and various techniques are devised to remove them [1, 2]. These techniques do not achieve absolute removal of nonlinearity, preventing single-shot imaging of bulk objects, such as the entire eye [3, 4] and consequently necessitating elaborate workaround solutions [5-8]. A Machine" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "Learning solution for nonlinearity removal was proposed as well [9] – although robust and reliable, it is biased towards specific kinds of objects with specific nonlinear properties." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 109, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 204 + ], + "type": "text", + "content": "On the other hand, the inherent optical properties that produce nonlinearities in the imaged objects have been shown to enable the characterisation of the objects [10-12] and correlate the change of their nonlinear character with the progression of diseases [13]. Unfortunately, the current methods for extracting the second-order nonlinearity are very object-specific as well as error-prone preventing them from being applied in real-world scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 205, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 456 + ], + "type": "text", + "content": "We present an approach that enables flexible and comfortable management of the signal nonlinearities, letting the user study one order in the absence of the others, a capability previously impossible to achieve using current standard or Machine Learning methods. In our approach, for each nonlinearity order, a neural network is trained whose input is a special two-dimensional data structure, a stack. Each row of the stack represents an amplitude of the Fourier transform of an input signal multiplied by a different nonlinearity-correcting factor. Consequently, it contains nonlinearity-free regions which the network identifies and uses to build a nonlinearity-free Fourier transform amplitude. We present two networks, one to remove the second-order nonlinearity while keeping the third-order one intact, and the second network to remove the third-order nonlinearity while keeping the second-order one intact. We test our approach on OCT signals that naturally contain substantial amounts of second-order nonlinearity and small amounts of third-order nonlinearity. The tests on both computer-generated and experimental signals indicate the robustness of our approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 457, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 636 + ], + "type": "text", + "content": "We start by describing our approach in Section 2, including the creation of the stacks in Subsection 2.1, their application in the training processes in Subsection 2.2, presentation of the used architecture and methodology in Subsection 2.3, and the description of the datasets (Subsection 2.4) used for training of our models. In Section 3, we report the performance of the networks in removing the respective nonlinearity, complete with the analysis of the training process and the subsequent tests on computer-generated data (Subsection 3.1) and experimental data (Subsections 3.2, 3.3 and 3.4). In the final sections, we summarise the obtained results (Section 4), pointing to the strong points of our approach, identifying the observed and potential shortcomings, and discussing potential solutions and the scope of future work (Section 5)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 650, + 364, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 364, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 364, + 663 + ], + "type": "text", + "content": "2. Concept" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 670, + 454, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 454, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 454, + 683 + ], + "type": "text", + "content": "2.1. Amplitude stacks as inputs" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our networks rely on a two-dimensional data structure as inputs: a stack (Fig. 2). The rows of a stack are created" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25029" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 71, + 264, + 123 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 90, + 80 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 90, + 80 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 90, + 80 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 71, + 264, + 123 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 264, + 123 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 264, + 123 + ], + "type": "image", + "image_path": "a16dbcc707930011ceab6b8624b3f09ee051c45d8fa50269d9bbf99f0758b4ab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 127, + 161, + 137 + ], + "lines": [ + { + "bbox": [ + 151, + 127, + 161, + 137 + ], + "spans": [ + { + "bbox": [ + 151, + 127, + 161, + 137 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 141, + 143, + 225 + ], + "blocks": [ + { + "bbox": [ + 52, + 128, + 61, + 137 + ], + "lines": [ + { + "bbox": [ + 52, + 128, + 61, + 137 + ], + "spans": [ + { + "bbox": [ + 52, + 128, + 61, + 137 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 141, + 143, + 225 + ], + "lines": [ + { + "bbox": [ + 51, + 141, + 143, + 225 + ], + "spans": [ + { + "bbox": [ + 51, + 141, + 143, + 225 + ], + "type": "image", + "image_path": "2c1c30731d0abafa704dbb5a36e90f55a1ff49d6b1a8911bb320c5c3b121e37f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 165, + 128, + 282, + 234 + ], + "blocks": [ + { + "bbox": [ + 165, + 128, + 282, + 234 + ], + "lines": [ + { + "bbox": [ + 165, + 128, + 282, + 234 + ], + "spans": [ + { + "bbox": [ + 165, + 128, + 282, + 234 + ], + "type": "image", + "image_path": "97def600b40e5ba6e9df3cd0024dc57242b310ca75f8f7bf4948020d928c7f75.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 248, + 143, + 340 + ], + "blocks": [ + { + "bbox": [ + 51, + 237, + 61, + 247 + ], + "lines": [ + { + "bbox": [ + 51, + 237, + 61, + 247 + ], + "spans": [ + { + "bbox": [ + 51, + 237, + 61, + 247 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 248, + 143, + 340 + ], + "lines": [ + { + "bbox": [ + 50, + 248, + 143, + 340 + ], + "spans": [ + { + "bbox": [ + 50, + 248, + 143, + 340 + ], + "type": "image", + "image_path": "ac0c70b4f65a46e82a7e214d19f98664a95e42d451f10d7b5444245b8ed4d0f9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 162, + 240, + 282, + 345 + ], + "blocks": [ + { + "bbox": [ + 151, + 237, + 161, + 247 + ], + "lines": [ + { + "bbox": [ + 151, + 237, + 161, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 237, + 161, + 247 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 240, + 282, + 345 + ], + "lines": [ + { + "bbox": [ + 162, + 240, + 282, + 345 + ], + "spans": [ + { + "bbox": [ + 162, + 240, + 282, + 345 + ], + "type": "image", + "image_path": "ab7a219ead3eadb8a1081c5f6d5f309fb6fea3665d9b981175de4c8830e8fad7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 453 + ], + "type": "text", + "content": "Figure 2. Calculation of the input stacks. (a) The raw signal is (b) multiplied by 32 different complex exponents representing 32 different levels of the second-order nonlinearity: " + }, + { + "bbox": [ + 46, + 354, + 287, + 453 + ], + "type": "inline_equation", + "content": "C_0, C_1, C_2, \\ldots, C_{31}" + }, + { + "bbox": [ + 46, + 354, + 287, + 453 + ], + "type": "text", + "content": ". Each row is then Fourier transformed. (c) The 2nd-order stack is created from the absolute value of the complex Fourier transforms. Similarly, (d) the raw signal is multiplied by 32 different complex exponents representing 32 different levels of the third-order nonlinearity: " + }, + { + "bbox": [ + 46, + 354, + 287, + 453 + ], + "type": "inline_equation", + "content": "D_0, D_1, D_2, \\ldots, D_{31}" + }, + { + "bbox": [ + 46, + 354, + 287, + 453 + ], + "type": "text", + "content": " and then Fourier transformed to build (e) a 3rd-order stack." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": "in the following way. First, the raw signal is multiplied (Fig. 2a) by a complex exponent with a different quadratic argument in the case of the stack used for removing the second-order nonlinearity (Fig. 2b) or with a different cubic argument in the case of the stack for the third-order nonlinearity removal (Fig. 2d). Coefficients in the quadratic arguments, " + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "inline_equation", + "content": "(C_0,\\dots,C_{31}" + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": " in Fig. 2b), and coefficients in the cubic arguments " + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "inline_equation", + "content": "(D_0,\\dots,D_{31}" + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": " in Fig. 2d), are chosen to cover a wide range of the second- and third-order nonlinearity, ensuring that the nonlinearities present in the signal are fully compensated. Then, each row is Fourier transformed, and an absolute value is calculated to obtain a 2nd-order stack (Fig. 2c) or a 3rd-order stack (Fig. 2e). The range of the nonlinear coefficients is symmetrical around 0 and broad enough to ensure that each modulation's nonlinearity in the raw signal is compensated. In the 2nd-order stack, these will be the places with the smallest width (marked 1, 2, 3 and 4 in Fig. 3b) and in the 3rd-order stack, the places containing no distortions (marked 1, 2, 3 and 4 in Fig. 3c)." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 375 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 375 + ], + "type": "image", + "image_path": "f782900b0485a1b983e5c1474a319e8adffe2eef9e06b102fec9a952f84f0124.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": "Figure 3. (a) The Fourier transform amplitude of a raw signal, presented in Fig. 2a, shows four peaks affected by a different amount and type of nonlinearity. The second-order nonlinearity is removed with Network #1 trained on (b) the 2nd-order stack as input and (d) Fourier transform amplitude free of the second-order nonlinearity as output. Similarly, the third-order nonlinearity is removed with Network #2 trained on (b) the 3rd-order stack as input and (d) Fourier transform amplitude free of the third-order nonlinearity as output. As a result, (f) the second-order nonlinearity is removed leaving the third-order one (see peak 4) using Network #1, (g) the third-order nonlinearity is removed leaving the second-order one intact (see peak 4) using Network #2." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 536, + 481, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 481, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 481, + 549 + ], + "type": "text", + "content": "2.2. Nonlinearity-removing networks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "The approach for removing the second-order nonlinearity and the approach for removing the third-order nonlinearity are similar, presented schematically in Fig. 3. The input of the network for the second-order nonlinearity removal, Network #1, is the 2nd-order stack (Fig. 3b) and the output is a Fourier transform amplitude free of second-order nonlinearity (Fig. 3d). Because the input 2nd-order stack contains regions corresponding to perfectly removed second-order nonlinearity, the training process can be viewed as learning how to identify these regions and then stitching them all up. Similarly, the network for removing third-order nonlinearity, Network #2, takes the 3rd-order stack as input (Fig. 3c) and outputs a Fourier transform amplitude free" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25030" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 282, + 232 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 282, + 232 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 282, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 282, + 232 + ], + "type": "image", + "image_path": "c182229d79c02093451922524225c06d0d70644057fd459307819d87e3aecfa5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 243, + 282, + 255 + ], + "lines": [ + { + "bbox": [ + 53, + 243, + 282, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 282, + 255 + ], + "type": "text", + "content": "Figure 4. Architecture of our nonlinearity-removing networks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 361 + ], + "type": "text", + "content": "from the 3rd-order nonlinearity (Fig. 3e). Input for both networks has a shape of 32 by 1024 pixels with a single channel whereas output is a vector of size 1024. When compared to the Fourier transform amplitude of the raw signal (Fig. 3a), the signal returned by Network #1 (Fig. 3f) incorporates only the third-order nonlinearity, and the signal returned by Network #2 (Fig. 3g) contains only the second-order nonlinearity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 132, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 132, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 132, + 383 + ], + "type": "text", + "content": "2.3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "content": "Our nonlinearity-removing networks are inspired by the U-Net architecture [14]. U-Net stands out for its use of a contracting path to capture contextual information, an expanding path for precise feature localisation, and the incorporation of skip connections that preserve fine-grained details, particularly beneficial for precise object delineation. We leverage these capabilities to identify and extract nonlinearity-related features from the stacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 484, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 593 + ], + "type": "text", + "content": "While the classical U-Net deals with 2D data structures for both input and output, in our case, we perform dimension reduction to output 1D signals by incorporating a max-pooling layer. We further refined the architecture by optimising hyperparameters with the validation dataset using Optuna [15]. As a result of the tuning, we adopted the Mean Absolute Error (MAE) loss function and the Adam optimiser [16] with a set learning rate of 0.0002. Our final network architecture is presented in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 599, + 111, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 111, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 111, + 611 + ], + "type": "text", + "content": "2.4. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Our datasets consist of noiseless computer-generated signals that are Gaussian-shaped, and diverse and randomised in terms of their content. If put in the OCT-imaging-related terms, they represent a broad range of possible imaged object configurations. These objects are simulated to have a random number of interfaces (from 2 to 12) positioned at arbitrary locations, with each object layer exhibiting varying random reflectivity and second- and third-order nonlin" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 545, + 203 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 545, + 203 + ], + "type": "image", + "image_path": "6ba29dd818ac660451fabe13244685a6b03cf21effed3cba4580072f9d54318b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 215, + 547, + 272 + ], + "lines": [ + { + "bbox": [ + 304, + 215, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 547, + 272 + ], + "type": "text", + "content": "Figure 5. Network #1 and Network #2 performance measured with the Mean Absolute Error (MAE) loss function across 30 epochs of training. Blue lines depict Network #1, while orange lines showcase the Network #2 performance. Dotted vertical lines highlight epochs exhibiting the best performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 282, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 547, + 354 + ], + "type": "text", + "content": "earity values. These object parameters are used to synthesise modulatory signals equivalent to those acquired in an OCT measurement (see example signals in Fig. 1a, c, e, g or Fig. 2a). Next, the input stacks and outputs, the Fourier transform amplitudes of signals free of the respective nonlinearity, are calculated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "type": "text", + "content": "The risk of overfitting was minimised by using the training datasets with an optimised size [17]. In our research, this was found to be between 160,000 and 240,000 objects. The validation and test datasets include 14,993 objects. The distribution is uniform across the objects with a specific number of interfaces, resulting in an equal allocation of 1,363 data samples for each distinct object type." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 486 + ], + "type": "text", + "content": "Before starting the training, as part of the preprocessing stage, both the input and output data were normalised through Min-Max scaling, ensuring that values fell within the standardised range of 0 to 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 496, + 388, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 388, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 388, + 509 + ], + "type": "text", + "content": "3. Performance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 515, + 412, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 412, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 412, + 528 + ], + "type": "text", + "content": "3.1. In silico data tests" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": "Each nonlinearity-removing neural network underwent training using a training dataset consisting of 200,000 objects. Within each mini-batch, 8 input stacks were processed. The proposed algorithm requires 39 ms to process a stack on a Quadro RTX 6000 GPU. Network #1 (Fig. 5, blue lines) and Network #2 (Fig. 5, orange lines) both displayed convergence within just a few epochs (Fig. 5) and were trained until the thirtieth epoch." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "To better evaluate the networks and results, we employed the Goodness-of-Fit Approximity (GoF) metric introduced in [12]. The GoF metric assesses the accuracy of predictions by calculating the percentage of values that lie within a predefined distance from the corresponding ground truth values. For this GoF metric, we define an error threshold as a ratio between the predefined distance and the value range" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25031" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 287, + 257 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 287, + 257 + ], + "type": "image", + "image_path": "2b01f51ee0a3c226f49bad1848ea6588ba9e68bd73cf0c597b6880e5ba54f49d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "type": "text", + "content": "Figure 6. Networks performance assessed through the metric of GoF, measuring the degree to which predictions align with the ground truth within " + }, + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "type": "text", + "content": " error in the test dataset comprising 14,993 data samples for (a) Network #1 and (b) Network #2. The graphs at the top (a1,a2) are zoomed-in areas of the top area of (b1,b2) the whole graphs presented." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 349, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 372 + ], + "type": "text", + "content": "span. In our case, where the values range from 0 to 1, the distance of 0.01 means " + }, + { + "bbox": [ + 46, + 349, + 287, + 372 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 349, + 287, + 372 + ], + "type": "text", + "content": " error threshold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "type": "text", + "content": "Based on the post-training evaluation using a " + }, + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "type": "text", + "content": " error threshold and the test dataset, Network #1 started to overfit at epoch 25 (Fig. 5, indicated by the blue dotted vertical line). In contrast, Network #2 demonstrated resilience against overfitting and achieved its best results in the 28th epoch (Fig. 5, marked by the red dotted vertical line)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "content": "Using the test dataset and GoF with a " + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "content": " error threshold, both networks achieved a remarkable mean GoF value of " + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "inline_equation", + "content": "99.98\\%" + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "content": ", which means that " + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "inline_equation", + "content": "99.98\\%" + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "content": " of the predictions differed from the ground truth by less than " + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "content": ", with only one example below " + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "content": " GoF. In light of this, to be able to carry out a meaningful analysis, the error threshold was further reduced to " + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 448, + 288, + 544 + ], + "type": "text", + "content": " and a boxplot graph was plotted for each model in Fig. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Both networks, Network #1 (Fig. 6a) and Network #2 (Fig. 6b), demonstrated very good performance even for error threshold " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": ". Only approximately " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "9.52\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "3.07\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " of the predictions from Network #1 and Network #2, respectively, fell below the " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " GoF for such a small error, as showed in Tab. 1 that also includes GoF values for objects with distinct number of layers. Fig. 7 offers an in-depth analysis of the worst-predicted examples for the Network #1 (Fig. 7a) and the Network #2 (Fig. 7b) showcasing GoF values of " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "65.5\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "45.5\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": ", respectively. Predictions of peak amplitudes in the nonlinearity-free Fourier transform amplitude (Fig. 7(3), orange line) closely align the expected outcomes (blue line) deviating only by the error threshold level. Also, the networks predict values close to" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 541, + 363 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 541, + 363 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 541, + 363 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 541, + 363 + ], + "type": "image", + "image_path": "4bb2ede470abb4d7063722526af6ed5e95e5fd76a7ad78f7b5a87bbcc63ddd95.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 372, + 547, + 452 + ], + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 452 + ], + "type": "text", + "content": "Figure 7. Predictions with very low GoF for " + }, + { + "bbox": [ + 304, + 372, + 547, + 452 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 372, + 547, + 452 + ], + "type": "text", + "content": " error threshold for (a) Network #1 and (b) Network #2. Based on the input stack (1), models predict (2, orange line, PR) amplitudes free of the respective nonlinearity that exhibit a strong correlation with the ground truth (2, blue line, GT). (3) Nonlinearity-free peaks are recovered with high fidelity, but (4) the models sometimes struggle with zero-valued places." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "content": "zero rather than precisely zero, contributing to worse GoF metric results. One should keep in mind here that these discrepancies are extremely low: at the level of around 2e-3 as presented in Fig. 7a4 and b4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 527, + 547, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 547, + 552 + ], + "type": "text", + "content": "3.2. Experimental signals: mirror at different depths" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "Networks were tested on the signals obtained through an OCT imaging of a mirror placed at 11 different depth positions. Each acquired signal Fourier transforms to a peak whose amount of second-order-nonlinearity-related broadening and third-order-nonlinearity-related distortion increases with depth (Fig. 8a). Due to their simultaneous appearance, the third-order nonlinearity manifests itself in the form of peak elongation. The depth-dependent deterioration is due to the fundamental limitations of OCT's spectrometer-based light detection. Additionally, the second- and third-order nonlinearities originating from the OCT's interferometer add up - depth-independently - to that effect, further worsening the outcome. Typically," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25032" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 91, + 493, + 165 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "type": "text", + "content": "Table 1. The number of cases where GoF with error threshold " + }, + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "type": "inline_equation", + "content": "{0.1}\\%" + }, + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "type": "text", + "content": " falls below " + }, + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "type": "inline_equation", + "content": "{95}\\%" + }, + { + "bbox": [ + 47, + 70, + 545, + 82 + ], + "type": "text", + "content": " for objects with a specific number of reflective layers." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 91, + 493, + 165 + ], + "lines": [ + { + "bbox": [ + 97, + 91, + 493, + 165 + ], + "spans": [ + { + "bbox": [ + 97, + 91, + 493, + 165 + ], + "type": "table", + "html": "
Net.Interfaces23456789101112Totals
#1# below 95%3842601931381058779554544371381
# in dataset1363136313631363136313631363136313631363136314993
% below28.1719.0814.1610.127.706.385.804.043.303.322.719.52
#2# below 95%10782575438323122131311460
# in dataset1363136313631363136313631363136313631363136314993
% below7.856.024.183.962.792.352.271.610.950.950.813.07
", + "image_path": "cd91d46d5b548845b89f48297f4cb25440891e888f5ff2eb1a99a27148859cd9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 100, + 172, + 242, + 236 + ], + "blocks": [ + { + "bbox": [ + 100, + 172, + 242, + 236 + ], + "lines": [ + { + "bbox": [ + 100, + 172, + 242, + 236 + ], + "spans": [ + { + "bbox": [ + 100, + 172, + 242, + 236 + ], + "type": "image", + "image_path": "5f1e9e381f89c8ff7b700fe887e4032e6521a8952802db1eda91af5e979aae43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 100, + 237, + 244, + 301 + ], + "blocks": [ + { + "bbox": [ + 100, + 237, + 244, + 301 + ], + "lines": [ + { + "bbox": [ + 100, + 237, + 244, + 301 + ], + "spans": [ + { + "bbox": [ + 100, + 237, + 244, + 301 + ], + "type": "image", + "image_path": "5d0ee500d14f3736bdb0d46ea3ec54d1ba04ec0b079b13b39a99eb207f80b209.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 100, + 303, + 244, + 381 + ], + "blocks": [ + { + "bbox": [ + 100, + 303, + 244, + 381 + ], + "lines": [ + { + "bbox": [ + 100, + 303, + 244, + 381 + ], + "spans": [ + { + "bbox": [ + 100, + 303, + 244, + 381 + ], + "type": "image", + "image_path": "2620ec16a20ec7e8d3b0ad08f3edc044a095f704114d0f43f055873e201fabcb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 384, + 169, + 460 + ], + "blocks": [ + { + "bbox": [ + 67, + 384, + 169, + 460 + ], + "lines": [ + { + "bbox": [ + 67, + 384, + 169, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 384, + 169, + 460 + ], + "type": "image", + "image_path": "e928aad4622d3ffb33a2b0c9c743c1e900eeab944af0122a2bb452dd9a9b981e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 470, + 287, + 580 + ], + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 580 + ], + "type": "text", + "content": "Figure 8. Mirror at 11 different depths. (a) Fourier transforming raw signals results in peaks that are both broadened by the second-order nonlinearity and distorted by the third-order nonlinearity. The deeper the peak, the bigger the nonlinearity effect, a behaviour characteristic of OCT signals acquired with a spectrometer. (b) Network #1 removes the second-order nonlinearity, leaving the third-order one intact. (c) Network #2 removes the third-order nonlinearity, leaving the third-order one intact. (d, e) The traditional linearisation (grey line) is not as precise as our approach (green lines)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 170, + 384, + 268, + 460 + ], + "blocks": [ + { + "bbox": [ + 170, + 384, + 268, + 460 + ], + "lines": [ + { + "bbox": [ + 170, + 384, + 268, + 460 + ], + "spans": [ + { + "bbox": [ + 170, + 384, + 268, + 460 + ], + "type": "image", + "image_path": "3ab486d8fc1f593adf72fef4b6ad47350d57db4cb81d0209bbc379694d983211.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "text", + "content": "spectrometer- and interferometer-related nonlinearities can be satisfactorily removed in a calibration step before the measurement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "The application of Network #1 removes the second-order nonlinearity in its entirety and consequently, provides a dramatic reduction of peaks' widths (Fig. 8b). Because the third-order nonlinearity remains intact, the peaks' shapes incorporate the characteristic \"tail\". On the other hand, Network #2 removes the third-order nonlinearity, leading to a complete reduction of the elongation (Fig. 8c)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 173, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 568 + ], + "type": "text", + "content": "One observes the superiority of our approach when the peaks obtained with Network #1 are compared to the peaks obtained as a result of traditional linearisation methods which remove all nonlinearity orders. Here, we used the method in [1] where two raw signals corresponding to the mirror placed at two different locations are used. First, their phases are subtracted to remove the constant interferometer-related contribution, enabling the calculation of the correction vector for spectrometer-related nonlinearities. Once the spectrometer's nonlinearity is removed, one of the initial spectra is used to generate an interferometer-nonlinearities-compensating vector. The peaks located close to 0 depth index obtained using our approach and the standard linearisation method (Fig. 8d, green and grey lines) are similar in shape: the network-outputted peak is higher which is very favourable - it means a higher signal-to-noise ratio, which is associated with better quality imaging. One observes a small \"tail\" at the side, most probably due to a small amount of third-order nonlinearity present at such shallow depths, whereas the peak obtained using the traditional methods is a smooth Gaussian. More interestingly, for a peak at a substantial depth (Fig. 8e), where the third-order nonlinearity is naturally much bigger (let us remind that this depth-increasing nonlinearity is tied to OCT spectrometer), the \"tail\" is much more pronounced in case of our network but not present in case of the traditional method. In the latter case, the peak is elongated which suggests the presence of both second- and third-order nonlinearities. This leads to the conclusion that traditional linearisation, although very satisfactory in practical applications, still has room for improvement. In fact, despite 30 years of OCT in science and industry, new methods for more precise OCT data linearisation are still being developed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 575, + 354, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 354, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 354, + 586 + ], + "type": "text", + "content": "3.3. Glass" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Another test was performed on data corresponding to a " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "300\\mu \\mathrm{m}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " thick sapphire glass put at an angle obtained with the same instrument. This data consisted of 256 raw OCT signals which represent a different transversal location on the glass. When Fourier transformed one after another, they form a 2D depth image of the object (Fig. 9a, 1 and 2 denote the front and back surface of the glass, the third line is an artefact placed at the depth equal to the thickness of the glass, inherent to OCT). Without linearisation, the image elements are broadened the deeper they are. When the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25033" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 72, + 225, + 154 + ], + "blocks": [ + { + "bbox": [ + 118, + 72, + 225, + 154 + ], + "lines": [ + { + "bbox": [ + 118, + 72, + 225, + 154 + ], + "spans": [ + { + "bbox": [ + 118, + 72, + 225, + 154 + ], + "type": "image", + "image_path": "db05d6db814cd63f778520fd8cbf79e2753d3aca766d02d665146c796554e892.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 155, + 167, + 234 + ], + "blocks": [ + { + "bbox": [ + 61, + 155, + 167, + 234 + ], + "lines": [ + { + "bbox": [ + 61, + 155, + 167, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 155, + 167, + 234 + ], + "type": "image", + "image_path": "29b29f391d403a08ee4c1fb4296adb274529a948df5f08f0247909f1669380e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 154, + 272, + 233 + ], + "blocks": [ + { + "bbox": [ + 168, + 154, + 272, + 233 + ], + "lines": [ + { + "bbox": [ + 168, + 154, + 272, + 233 + ], + "spans": [ + { + "bbox": [ + 168, + 154, + 272, + 233 + ], + "type": "image", + "image_path": "3c3bc75914cc1ad76b547fa132501ec56d85bbcd9eab0f85bdfb679050fd43b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 235, + 166, + 312 + ], + "blocks": [ + { + "bbox": [ + 60, + 235, + 166, + 312 + ], + "lines": [ + { + "bbox": [ + 60, + 235, + 166, + 312 + ], + "spans": [ + { + "bbox": [ + 60, + 235, + 166, + 312 + ], + "type": "image", + "image_path": "b06b3c978a03e9edc2508a382adca684a7507d4598e03555fd20d3dc58c5f8d7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 167, + 235, + 272, + 312 + ], + "blocks": [ + { + "bbox": [ + 167, + 235, + 272, + 312 + ], + "lines": [ + { + "bbox": [ + 167, + 235, + 272, + 312 + ], + "spans": [ + { + "bbox": [ + 167, + 235, + 272, + 312 + ], + "type": "image", + "image_path": "5832df6ccea4e3ea07b2dd0e74e603d312d1e31cd18c06eb847279c016401075.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 92, + 315, + 239, + 391 + ], + "blocks": [ + { + "bbox": [ + 92, + 315, + 239, + 391 + ], + "lines": [ + { + "bbox": [ + 92, + 315, + 239, + 391 + ], + "spans": [ + { + "bbox": [ + 92, + 315, + 239, + 391 + ], + "type": "image", + "image_path": "d199ff707b3a1e13fa742a2a2e0da1df1dd4a3f1bd2c4dba222464fb32f3630b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 400, + 287, + 553 + ], + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 553 + ], + "type": "text", + "content": "Figure 9. Sapphire glass. (a) Fourier transformation of raw signals gives an OCT image where the quality deteriorates with depth. (b) Network #1 removes the second-order nonlinearity which accounts for the most of quality deterioration, (c) Due to a small amount of the third-order nonlinearity, the application of Network #2 does not improve the image quality. (d), Image line #152 from Network #1 (green line) is compared with its raw image equivalent (grey line) to show dramatic peaks' width reduction, (e) The image line #152 from Network #2 output (green line) and its raw image equivalent (grey line) show no significant difference except for a more favourable shape for the former. (f) Network #1 (green line) performs quite well when compared to the traditional method output (black line). 1 and 2 mark the front and back surface of the imaged glass, the middle line is an artefact." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 653 + ], + "type": "text", + "content": "same data is processed with Network #1, the broadening is removed from all the image elements (see the image in Fig. 9b and the 152nd image line in Fig. 9d). Due to very small amounts of the third-order nonlinearity, Network #2 does not bring perceptible improvements (Fig. 9c), but under a closer inspection of a single image line (Fig. 9e), one can observe some improvement in the peak height and a possible side elongation reduction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Fig. 9f compares the output of Network #1 (green line) and the traditional linearisation method (black line), indicating that this model could be a satisfactory alternative for the traditional image calculation algorithm when the third-order nonlinearities are negligible. It can be used as such an alter" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "native for the OCT laboratory instrument which was used here to obtain the data and which is nevertheless a fairly good representative of a standard OCT instrument. We note that Network #1 improved the shape of the artefact (the middle peak in Fig. 9f), which - although generally irrelevant for OCT imaging - is still something noteworthy as such a feat is beyond the capabilities of traditional linearisation methods, showing that our approach, unlike the traditional ones, removes the nonlinearities nondiscriminatively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 186, + 428, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 186, + 428, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 428, + 198 + ], + "type": "text", + "content": "3.4. Grape and cucumber" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 205, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 276 + ], + "type": "text", + "content": "The final tests were performed on the data corresponding to a grape (Fig. 10) and a cucumber (Fig. 11). The former represents a high signal-to-noise ratio (highly visible modulation in the signal as depicted in Fig. 10a) and the latter a low signal-to-noise ratio (poorly visible modulation in the signal as depicted in Fig. 11a)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 419 + ], + "type": "text", + "content": "Fourier transformation of raw signals gives images of a low quality that further deteriorates with depth (Fig. 10b, Fig. 11b). Again, the improvement obtained when applying Network #1 (Fig. 10c, Fig. 11c) is comparable to that obtained with traditional methods (Fig. 10d, Fig. 11d). Under closer inspection (Fig. 10e, Fig. 11e), one notices that Network #1 output is less detailed for the higher signal-to-noise grape data - most probably due to the remaining third-order nonlinearities - and loses some details for the lower signal-to-noise ratio cucumber data. The latter loss is the direct consequence of training data representing predominantly high signal-to-noise ratio signals." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 420, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 456 + ], + "type": "text", + "content": "Just as in the case of the glass data, due to small amounts of third-order nonlinearity, Network #2 did not provide visually significant improvement and therefore, was omitted." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 466, + 448, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 448, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 448, + 479 + ], + "type": "text", + "content": "4. Summary and discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": "Two models were demonstrated: one for removing second-order nonlinearity from digital signals and leaving the third-order one intact, and another one for removing the third-order nonlinearity but leaving the second-order one intact. Their nonlinearity-removing capabilities were extensively analysed using computer-generated signals as well as experimental signals acquired with a laboratory Optical Coherence Tomography (OCT) device. Tests on the OCT signals, which corresponded to objects with varying levels of complication: mirror, sapphire, grape and cucumber, allowed us to evaluate the networks in terms of their applicability in a world-relevant technology. Whereas the tests and analysis done on the computer-generated signals revealed extraordinary performance in terms of similarity of the ground truth and prediction signals, the tests on OCT-acquired data indicated the networks' advantageous operation when compared to approaches traditionally used in OCT." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "During training, both networks converged exceptionally fast and reached remarkable loss function values. Nevertheless" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25034" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 265, + 132 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 265, + 132 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 265, + 132 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 265, + 132 + ], + "type": "image", + "image_path": "3e91691d4c6ee5a8513471026196bc126a65485cb245696d7a3e87b0a4c6aa30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 133, + 272, + 200 + ], + "blocks": [ + { + "bbox": [ + 61, + 133, + 272, + 200 + ], + "lines": [ + { + "bbox": [ + 61, + 133, + 272, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 133, + 272, + 200 + ], + "type": "image", + "image_path": "11cef7ae94bbb341424bf00dabf5797ce05db10ee810bf72a9a19dd6b73afea2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 201, + 272, + 269 + ], + "blocks": [ + { + "bbox": [ + 62, + 201, + 272, + 269 + ], + "lines": [ + { + "bbox": [ + 62, + 201, + 272, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 201, + 272, + 269 + ], + "type": "image", + "image_path": "ff57d1cd76565e64994e4471ce3345eff28d89eb9e775d37f069acc6a2163995.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 270, + 272, + 354 + ], + "blocks": [ + { + "bbox": [ + 62, + 270, + 272, + 354 + ], + "lines": [ + { + "bbox": [ + 62, + 270, + 272, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 272, + 354 + ], + "type": "image", + "image_path": "cfb73ce259b04d136b209571ff4afd7e95e70ca39b85bd5278c4fcc5ea31b04b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 356, + 266, + 417 + ], + "blocks": [ + { + "bbox": [ + 62, + 356, + 266, + 417 + ], + "lines": [ + { + "bbox": [ + 62, + 356, + 266, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 356, + 266, + 417 + ], + "type": "image", + "image_path": "5fc46eb2c470196332d40602656d8a40560b948ae526f2e6ee8b773941028d93.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 426, + 287, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 504 + ], + "type": "text", + "content": "Figure 10. Grape. (a) One of the raw signals representing a single line in the image of a grape. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image when compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees very well with its equivalent from the image obtained using traditional methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 512, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 596 + ], + "type": "text", + "content": "less, we have identified possible improvements: the closer inspection of the predictions revealed non-zero values that are zero-valued areas in the ground-truth signals. Another current shortcoming of both networks is their time of processing an entire OCT image; consequently, for now, the processing speed is insufficient for the real-time processing of data coming straight from OCT devices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 127, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 127, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 127, + 621 + ], + "type": "text", + "content": "5. Future work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "The future work will mainly consist in developing an approach that enables one to use both networks in a sequence so that both the second- and third-order nonlinearity removal is achieved. Since, as was shown here, the networks remove the nonlinearities completely, their sequential use will provide a very valuable tool for extremely precise data linearisation, especially attractive for OCT where any un" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 74, + 529, + 133 + ], + "blocks": [ + { + "bbox": [ + 320, + 74, + 529, + 133 + ], + "lines": [ + { + "bbox": [ + 320, + 74, + 529, + 133 + ], + "spans": [ + { + "bbox": [ + 320, + 74, + 529, + 133 + ], + "type": "image", + "image_path": "62313f48081c7d2fb57f479ede3d9b1be037e21aee669caf260aace6251bb469.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 134, + 533, + 201 + ], + "blocks": [ + { + "bbox": [ + 320, + 134, + 533, + 201 + ], + "lines": [ + { + "bbox": [ + 320, + 134, + 533, + 201 + ], + "spans": [ + { + "bbox": [ + 320, + 134, + 533, + 201 + ], + "type": "image", + "image_path": "0f91b7b15067482bb19a0dc332343e88654fea8d32a1418a4276c70135d4a55c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 202, + 533, + 270 + ], + "blocks": [ + { + "bbox": [ + 320, + 202, + 533, + 270 + ], + "lines": [ + { + "bbox": [ + 320, + 202, + 533, + 270 + ], + "spans": [ + { + "bbox": [ + 320, + 202, + 533, + 270 + ], + "type": "image", + "image_path": "108a614d87b771b1ecca5c1e048a8402bddcf6819fbad750134be4c6db4038a4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 271, + 531, + 354 + ], + "blocks": [ + { + "bbox": [ + 320, + 271, + 531, + 354 + ], + "lines": [ + { + "bbox": [ + 320, + 271, + 531, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 271, + 531, + 354 + ], + "type": "image", + "image_path": "a122f416a3da4332676b271f026a756ad0aae55a4e44d28b77efac3cdaeacc00.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 357, + 527, + 418 + ], + "blocks": [ + { + "bbox": [ + 321, + 357, + 527, + 418 + ], + "lines": [ + { + "bbox": [ + 321, + 357, + 527, + 418 + ], + "spans": [ + { + "bbox": [ + 321, + 357, + 527, + 418 + ], + "type": "image", + "image_path": "37755de439c2620b9482844e69b68ef5b17823c6d4c1ec5ddc693267d4de479d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 427, + 545, + 504 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 504 + ], + "type": "text", + "content": "Figure 11. Cucumber. (a) One of the raw signals representing a single line in the image of a cucumber. (b) Fourier transform of raw signals gives a low-quality image. (c) Network #1 removes the second-order nonlinearity providing a high-quality image compared with (d) the image obtained using traditional linearisation methods. (e) The prediction for image line #200 agrees well with its equivalent from the image obtained using traditional methods." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 511, + 515, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 511, + 515, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 511, + 515, + 523 + ], + "type": "text", + "content": "compensated nonlinearities lead to the quality drop." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 523, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 617 + ], + "type": "text", + "content": "Also, in terms of the network performance, we will try to remove the discrepancy for zero-valued elements in the ground truth and prediction signals by adding extra convolutional layers before the output to filter out the unwanted values as well as performing additional hyperparameter tuning. Additionally, we plan to speed up the processing of the stacks by lowering the amount of calculations and the size of the networks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 701 + ], + "type": "text", + "content": "Acknowledgements. Supported by the New Zealand Ministry of Business, Innovation and Employment (MBIE), Smart Ideas grant \"Extending the Boundaries of Digital Signal Processing: AI-powered Fourier Transformation Alternative (E7943)\". We also thank Oliver Batchelor, François Bissey, and Piotr Kolenderski for sharing their computational resources." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25035" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Kai Wang and Zhihua Ding. Spectral calibration in spectral domain optical coherence tomography. Chinese Optics Letters, 6(12):902-904, 2008. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 167 + ], + "type": "text", + "content": "[2] Gongpu Lan and Guoqiang Li. Design of ak-space spectrometer for ultra-broad waveband spectral domain optical coherence tomography. Scientific reports, 7(1):42353, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 286, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 225 + ], + "type": "text", + "content": "[3] Ireneusz Grulkowski, Silvestre Manzanera, Lukasz Cwiklinski, Franciszek Sobczuk, Karol Karnowski, and Pablo Artal. Swept source optical coherence tomography and tunable lens technology for comprehensive imaging and biometry of the whole eye. Optica, 5(1):52-59, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 286, + 270 + ], + "type": "text", + "content": "[4] Timothy R Hillman and David D Sampson. The effect of water dispersion and absorption on axial resolution in ultrahigh-resolution optical coherence tomography. Optics express, 13 (6):1860-1874, 2005. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 286, + 327 + ], + "type": "text", + "content": "[5] Siyu Chen, Benjamin Potsaid, Yan Li, Junhong Lin, Yun- chan Hwang, Eric M Moult, Jason Zhang, David Huang, and James G Fujimoto. High speed, long range, deep penetration swept source oct for structural and angiographic imaging of the anterior eye. Scientific reports, 12(1):992, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 286, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 286, + 372 + ], + "type": "text", + "content": "[6] Jiangjie Huang, Jinyu Fan, Yi He, and Guohua Shi. Physical compensation method for dispersion of multiple materials in swept source optical coherence tomography. Journal of Biophotonics, 16(10):e202300167, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 286, + 418 + ], + "type": "text", + "content": "[7] Daniel L Marks, Amy L Oldenburg, J Joshua Reynolds, and Stephen A Boppart. Digital algorithm for dispersion correction in optical coherence tomography for homogeneous and stratified media. Applied optics, 42(2):204-217, 2003." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 286, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 286, + 474 + ], + "type": "text", + "content": "[8] Liuhua Pan, Xiangzhao Wang, Zhongliang Li, Xiangyang Zhang, Yang Bu, Nan Nan, Yan Chen, Xuan Wang, and Fengzhao Dai. Depth-dependent dispersion compensation for full-depth oct image. Optics express, 25(9):10345-10354, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 286, + 520 + ], + "type": "text", + "content": "[9] Shaiban Ahmed, David Le, Taeyoon Son, Tobiloba Adejumo, Guangying Ma, and Xincheng Yao. Asc-net: an opensource deep learning network for automated dispersion compensation in oct. Frontiers in Medicine, 9:864879, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "text", + "content": "[10] Sylwia M Kolenderska, Bastian Brauer, and Frédérique Vanholsbeeck. Dispersion mapping as a simple postprocessing step for fourier domain optical coherence tomography data. Scientific reports, 8(1):9244, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "text", + "content": "[11] Christos Photiou and Costas Pitris. Comparison of tissue dispersion measurement techniques based on optical coherence tomography. Journal of biomedical optics, 24(4):046003-046003, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "text", + "content": "[12] Krzysztof A Maliszewski, Magdalena A Urbańska, Piotr Kolenderski, Varvara Vetrova, and Sylwia M Kolenderska. Extracting group velocity dispersion values using quantum-mimic optical coherence tomography and machine learning. Scientific Reports, 13(1):6596, 2023. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[13] Christos Photiou, Evgenia Bousi, Ioanna Zouvani, and Costas Pitris. Using speckle to measure tissue dispersion in optical coherence tomography. Biomedical optics express, 8 (5):2528-2535, 2017. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 273 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[14] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 194 + ], + "type": "text", + "content": "[15] Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623-2631, 2019. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 196, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 227 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 229, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 273 + ], + "type": "text", + "content": "[17] Jayme Garcia Arnal Barbedo. Impact of dataset size and variety on the effectiveness of deep learning and transfer learning for plant disease classification. Computers and electronics in agriculture, 153:46-53, 2018. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25036" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_content_list.json b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d380fbdd5661a57d415cc3248f8c11ba680cbfc6 --- /dev/null +++ b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_content_list.json @@ -0,0 +1,1713 @@ +[ + { + "type": "text", + "text": "Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianqi Wang*, Gangwei Xu*, Hao Jia, Xin Yang†", + "bbox": [ + 282, + 179, + 691, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huazhong University of Science and Technology", + "bbox": [ + 287, + 205, + 679, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xianqiw, gwxu, haojia, xinyang2014}@hust.edu.cn", + "bbox": [ + 269, + 224, + 694, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 275, + 313, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stereo matching methods based on iterative optimization, like RAFT-Stereo and IGEV-Stereo, have evolved into a cornerstone in the field of stereo matching. However, these methods struggle to simultaneously capture high-frequency information in edges and low-frequency information in smooth regions due to the fixed receptive field. As a result, they tend to lose details, blur edges, and produce false matches in textureless areas. In this paper, we propose Selective Recurrent Unit (SRU), a novel iterative update operator for stereo matching. The SRU module can adaptively fuse hidden disparity information at multiple frequencies for edge and smooth regions. To perform adaptive fusion, we introduce a new Contextual Spatial Attention (CSA) module to generate attention maps as fusion weights. The SRU empowers the network to aggregate hidden disparity information across multiple frequencies, mitigating the risk of vital hidden disparity information loss during iterative processes. To verify SRU's universality, we apply it to representative iterative stereo matching methods, collectively referred to as Selective-Stereo. Our Selective-Stereo ranks $1^{st}$ on KITTI 2012, KITTI 2015, ETH3D, and Middlebury leaderboards among all published methods. Code is available at https://github.com/Windsrain/Selective-Stereo.", + "bbox": [ + 75, + 308, + 473, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stereo matching is a fundamental area of research in computer vision. It explores the calculation of displacement, referred to as disparity, between matching points in a pair of rectified images. This technique plays a significant role in various applications, including 3D reconstruction and autonomous driving.", + "bbox": [ + 75, + 726, + 468, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advancement of deep learning, learning-based stereo matching [4, 14, 15, 17, 34, 35] have progressively displaced traditional methods and significantly enhancing", + "bbox": [ + 75, + 816, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/002f571a9e974f670eefa979e7102f82b607bc7516ea8181d557c1da34de0207.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 273, + 694, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17257d6328d2e82198ac8b669786cf3ae5f1b0c281975362c037f1fa96310447.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 273, + 880, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35c241f98aded3ac51826d196765ea35f3e1df09c7547a1e98fbe1866111d9d4.jpg", + "image_caption": [ + "Left Image" + ], + "image_footnote": [], + "bbox": [ + 501, + 425, + 627, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86586d706209ab26d3f7c432827e7e78a0b4da09656fc674f6fb6062ea4a65e9.jpg", + "image_caption": [ + "RAFT-Stereo [17]" + ], + "image_footnote": [], + "bbox": [ + 629, + 425, + 754, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea01c757381138d4a041dccfea298fefea82731edfbf36cb4429944d9f4fea04.jpg", + "image_caption": [ + "Selective-RAFT" + ], + "image_footnote": [], + "bbox": [ + 756, + 425, + 883, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df2a9b545d1e25c01d556e5c69a4ee46d374a5102b5cf1a77c512bd5fdece4a1.jpg", + "image_caption": [ + "Left Image" + ], + "image_footnote": [], + "bbox": [ + 501, + 512, + 627, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/654fc3d6c856316bfefaef6812f4570fc5eb8686ca537e97cfd4940ed66cf325.jpg", + "image_caption": [ + "IGEV-Stereo [34]" + ], + "image_footnote": [], + "bbox": [ + 630, + 511, + 754, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b1081ab24d0706573ac72907b023494f8ea2c431be679bcd90c9167b82806ca.jpg", + "image_caption": [ + "Selective-IGEV", + "Figure 1. Row 1: Comparisons with state-of-the-art stereo methods on KITTI 2012 [13] and KITTI 2015 [21], ETH3D [24] and Middlebury [23] leaderboards. Row 2: Visual comparison with RAFT-Stereo on ETH3D. Row 3: Visual comparison with IGEV-Stereo on Middlebury. Our method distinguishes subtle details and sharp edges and performs well in weak texture regions." + ], + "image_footnote": [], + "bbox": [ + 756, + 511, + 883, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the accuracy of disparity estimation. Initially, aggregation-based methods [7, 32, 41] led the development of stereo matching algorithms. These methods begin by defining a maximum range of disparity, constructing a 4D cost volume using feature maps, and subsequently employing 3D CNN to filter the volume and derive the final disparity map. Such methods focus on filtering the initially coarse cost volume, thus effectively aggregating geometry information. However, cost aggregation requires a large number of convolutions, resulting in high computational costs, making it difficult to be applied to high-resolution images.", + "bbox": [ + 496, + 699, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, a novel class of methods based on iterative optimization [11, 16, 17, 44] has been gaining prominence and", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 205, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "achieving state-of-the-art performance on several leaderboards. These methods begin by constructing an all-pairs cost volume, indexing a local cost volume from the original cost volume, and subsequently employing recurrent units [10] to calculate disparity residuals and update the disparity prediction. One major advantage of these methods is their ability to capture all candidate matching points without predefining the range of disparities. Additionally, these methods don't need to aggregate the cost volume using a large number of redundant convolutions. Instead, a continuous update of the disparity prediction is achieved through lightweight recurrent units during iterations. Therefore, these methods are capable of processing high-resolution images.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, iterative methods encounter several challenges. Firstly, the all-pairs cost volume includes considerable noisy information [34], potentially causing the loss of crucial information when iterating the hidden information. Besides, as the network iterates, the hidden information increasingly incorporates global low-frequency information while losing local high-frequency information like edges and thin objects [44]. Secondly, the existing recurrent units possess a fixed receptive field, leading the network to solely concentrate on information at the current frequency and ignore other frequencies, such as detailed, edge, and textureless information.", + "bbox": [ + 75, + 306, + 472, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Selective Recurrent Unit (SRU) to address the limitations of traditional recurrent units. As Chen et al. [5] mentions features contain information at different frequencies, high-frequency information describes rapidly changing fine details, while low-frequency information describes smoothly changing structures. Unlike traditional recurrent units that treat information at different frequencies equally, our SRU incorporates multiple branches of GRU, each with a distinct kernel size representing different receptive fields. The hidden information obtained from each GRU branch is fused and then fed into the next iteration. This fusion enables the capture of information from different receptive fields at different frequencies, while also performing secondary filtering to reduce noise information from local cost volume. To further enhance the fusion process, we propose a Contextual Spatial Attention (CSA) module to utilize the context information. Instead of simply summarizing information from different branches, CSA introduces attention maps extracted from the context information. After doing so, information captured by small kernels has large weights in regions like edge, while information captured by large kernels has large weights in regions like low-texture. These attention maps determine the weight of fusion, allowing the network to adaptively select suitable information based on different image regions. Besides, we prove the effectiveness and universality of our module by transferring it to different iterative networks. All networks", + "bbox": [ + 75, + 493, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are collectively referred to as Selective-Stereo. By doing so, we consistently improve the performance of these networks without introducing a significant increase in parameters and time.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of our method on several stereo benchmarks. On Scene Flow [20], our Selective-RAFT reaches the state-of-the-art EPE of 0.47, and our Selective-IGEV even achieves a new state-of-the-art EPE of 0.44. And as shown in Fig. 1, our Selective-RAFT surpasses RAFT-Stereo by a large margin and achieves competitive performance compared with the state-of-the-art methods on KITTI [13, 21] leaderboards. Our Selective-IGEV ranks $1^{st}$ on KITTI, ETH3D [24], and Middlebury [23] leaderboards among all published methods.", + "bbox": [ + 496, + 152, + 893, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 517, + 303, + 880, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel iterative update operator SRU for iterative stereo matching methods.", + "- We introduce a new Contextual Spatial Attention module that generates attention maps for adaptively fusing hidden disparity information at multiple frequencies.", + "- We verify the universality of our SRU on several iterative stereo matching methods.", + "- Our method outperforms existing published methods on public leaderboards such as KITTI, ETH3D, and Middlebury." + ], + "bbox": [ + 500, + 319, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 482, + 640, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aggregation-based methods in stereo matching. Several aggregation-based methods [4, 7, 8, 14, 15, 25, 35-37, 41] have shown significant progress in the domain of stereo matching in recent years. DispNet [20] establishes the groundwork for subsequent network architecture. GCNet [15] proposes a 4D concatenate cost volume, which is subsequently regularized using 3D CNNs. Additionally, it also introduces the soft argmin function for disparity regression, resulting in a significant influence on subsequent methods. PSMNet [4] proposes a stacked hourglass 3D CNN, which improves the cost aggregation stage to enhance the network's ability to capture context information. GwcNet [14] proposes Group-wise Correlation Volume, which combines the advantages of correlation and concatenation volume. GA-Net [41] designs a semi-global guided aggregation layer and a local guided aggregation layer, inspired by the traditional semi-global matching algorithm, to further assist in aggregating global and geometry information in the network. Building upon Group-wise Correlation Volume, ACVNet [32] proposes Attention Concat Volume, which uses attention weights to suppress redundant information and maintain sufficient information for matching.", + "bbox": [ + 496, + 508, + 893, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Iterative-based methods in stereo matching. In recent years, many iterative methods [17, 27, 33, 34], spearheaded by RAFT [17], have gradually become the mainstream of research. RAFT-Stereo [17] builds upon the optical flow", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19702", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d33e626a3ec484bfdcd1d10a05b2a81f673cd9317a19877d45028bf49006b2de.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed Selective-Stereo (Selective-RAFT version). The Contextual Spatial Attention (CSA) module extracts attention maps from context information as a guide for Selective Recurrent Units (SRUs). Then the network iteratively updates the disparity using local cost volumes retrieved from the correlation pyramid and attention maps given by CSA through SRUs." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "method RAFT [27] by introducing an all-pairs cost volume pyramid that maintains high resolution. It extracts local correlation features from this pyramid, performs iterative disparity updates using GRU-based update operators, and incorporates a multi-level GRU to expand the receptive field. On this basis, IGEV-Stereo [34] asserts that the initial cost volume is excessively coarse. To alleviate the need for iterations and reduce time overhead, it proposes to use a lightweight cost aggregation network before iterations. CREstereo [16] designs a hierarchical network in a coarse-to-fine manner, as well as a stacked cascaded architecture for inference in place of the original single-resolution iterative structure. DLNR [44] proposes the use of LSTM as a replacement for GRU, providing the advantage of decoupling the update of hidden states from disparity prediction.", + "bbox": [ + 75, + 500, + 472, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frequency information application in vision. There are several works[5, 6, 39] that focus on using frequency information in computer vision. Chen et al. [5] propose the octave convolution to factorize the mixed feature maps by their frequencies. Xu et al. [39] propose a method of learning in the frequency domain and suggest that CNN models are more sensitive to low-frequency channels than high-frequency. DSGAN [12] introduces the frequency separation into super-resolution. LITv2 [22] proposes to disentangle the high/low-frequency patterns in an attention layer.", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 497, + 589, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the overall architecture of Selective-Stereo. Because our method can be plugged into different networks, we take Selective-RAFT (Fig. 2) as an example and focus on illustrating its key components.", + "bbox": [ + 496, + 527, + 892, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Extraction", + "text_level": 1, + "bbox": [ + 500, + 608, + 681, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure fair comparisons, Selective-RAFT maintains consistency with RAFT-Stereo [17] by employing its feature extraction network. Feature extraction comprises two main components: feature network and context network.", + "bbox": [ + 496, + 635, + 890, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Network. Given the left and the right images $\\mathbf{I}_{l(r)}\\in \\mathbb{R}^{3\\times H\\times W}$ , we first downsample them to $1 / 2$ resolution using a $7\\times 7$ convolutional layer. Then, a series of residual blocks is employed to extract features and we apply another downsampling layer to get features at $1 / 4$ resolution in the middle. Lastly, a $1\\times 1$ convolutional layer is applied to get the final left and right features $\\mathbf{f}$ , $\\mathbf{g}\\in \\mathbb{R}^{C\\times \\frac{H}{4}\\times \\frac{W}{4}}$ with suitable dimensions.", + "bbox": [ + 496, + 700, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Context Network. Its architecture remains consistent with the feature network, and it adds a series of residual blocks and two additional downsampling layers, obtaining multi-level context features $\\mathbf{f}_i^c$ ( $i = 1, 2, 3$ ) at $1/4, 1/8, 1/16$ resolutions. Then we can get the initial hidden and the con", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text information:", + "bbox": [ + 76, + 92, + 192, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {i} = \\tanh \\left(\\mathbf {f} _ {i} ^ {c}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 102, + 468, + 126 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i} = \\operatorname {R e L U} \\left(\\mathbf {f} _ {i} ^ {c}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 122, + 325, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cost Volume Construction", + "text_level": 1, + "bbox": [ + 76, + 154, + 316, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the left and the right features $\\mathbf{f},\\mathbf{g}$ , we first construct an all-pairs correlate cost volume:", + "bbox": [ + 76, + 176, + 468, + 207 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {i j k} = \\sum_ {h} \\mathbf {f} _ {h i j} \\cdot \\mathbf {g} _ {h i k}, \\mathbf {C} \\in \\mathbb {R} ^ {\\frac {H}{4} \\times \\frac {W}{4} \\times \\frac {W}{4}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 217, + 468, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we construct a 4-level correlation pyramid $\\{\\mathbf{C}_i\\}$ ( $i = 1,2,3,4$ ) by using 1D average pooling with a kernel size of 2 and a stride of 2 at the last dimension.", + "bbox": [ + 76, + 261, + 470, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Contextual Spatial Attention Module", + "text_level": 1, + "bbox": [ + 76, + 316, + 397, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To help information from different receptive fields and frequencies fuse, the Contextual Spatial Attention (CSA) module extracts multi-level attention maps from context information as guidance. As illustrated in Fig. 3, CSA can be divided into two submodules: Channel Attention Enhancement (CAE) and Spatial Attention Extractor (SAE). These submodules are derived from CBAM [31] and we simplify them to better adapt to stereo matching.", + "bbox": [ + 76, + 339, + 468, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel Attention Enhancement. Given a context information map $\\mathbf{c} \\in \\mathbb{R}^{C \\times H \\times W}$ , we first use an average-pooling and a max-pooling operation on the spatial dimension to get two maps $\\mathbf{f}_{avg}, \\mathbf{f}_{max} \\in \\mathbb{R}^{C \\times 1 \\times 1}$ . Then we use two convolutional layers to perform feature transformation on these maps separately. After that, we add these two maps together and use the sigmoid function to convert them into weights $\\mathbf{M}_c \\in \\mathbb{R}^{C \\times 1 \\times 1}$ between 0 and 1. Lastly, using an element-wise product, the initial map can capture which channel map has high feature values to be enhanced, and which channel map has low feature values to be suppressed.", + "bbox": [ + 76, + 460, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Attention Extractor. After the CAE module, we continue to use the same pooling operations, but now we pool on the channel dimension. Then we concatenate these pooling maps to form a map in $\\mathbb{R}^{2\\times H\\times W}$ and use one convolutional layer with a sigmoid function to generate the final attention map. Reviewing previous operations, this attention map has high weights in regions needing high-frequency information because this information possesses high feature values in the context information. Similarly, it has low weights in regions needing low-frequency information. In general, the attention map can explicitly distinguish regions that need information at different frequencies.", + "bbox": [ + 76, + 627, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Selective Recurrent Unit", + "text_level": 1, + "bbox": [ + 76, + 816, + 302, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To capture information at different frequencies, Selective Recurrent Unit (SRU) uses attention maps extracted by CSA to fuse hidden information derived from GRUs with different kernel sizes.", + "bbox": [ + 76, + 839, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-level update structure. As illustrated in Fig. 4, SRUs at $1/8$ , $1/16$ resolutions take the attention map, context information, hidden information at the same resolution, and the hidden information at adjacent resolutions as inputs. At $1/4$ resolution, SRUs take disparity, and local cost volume as additional inputs, and then their outputs will go through two convolutional layers to generate disparity residuals. The local cost volume is derived from the all-pairs correlation pyramid in the same way as RAFT-Stereo [17]. At last, disparities at $1/4$ resolution will be upsampled into full resolution using the convex combination.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SRU's architecture. A single GRU can be defined as follows:", + "bbox": [ + 500, + 257, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {k} = \\sigma (\\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {z})),\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 296, + 769, + 313 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} r _ {k} = \\sigma (\\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {r})), \\\\ \\tilde {I} _ {\\text {一}} = \\operatorname {c o v} (G _ {m} ([ \\circ_ {0} ]), W _ {r})) \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 315, + 890, + 343 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {h} _ {k} = \\tanh (\\operatorname {C o n v} ([ r _ {k} \\odot h _ {k - 1}, x _ {k} ], W _ {h})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 330, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} = \\left(1 - z _ {k}\\right) \\odot h _ {k - 1} + z _ {k} \\odot \\tilde {h} _ {k}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 357, + 781, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{k}$ is the concatenation of disparity, correlation, hidden information, and context information previously defined. Unlike RAFT-Stereo [17] that divide the context information into $c_{z}, c_{r}, c_{h}$ , we add it into $x_{k}$ because using convolutions with different kernel sizes can fully utilize context information.", + "bbox": [ + 496, + 386, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, a single SRU can be defined as follows:", + "bbox": [ + 498, + 478, + 890, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} = \\mathbf {A} \\odot h _ {k} ^ {s} + (1 - \\mathbf {A}) \\odot h _ {k} ^ {l} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 507, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ denotes the attention map derived from CSA at the same resolution, $h_k^s$ denotes the GRU with smaller kernel sizes and $h_k^l$ denotes the larger one.", + "bbox": [ + 496, + 532, + 890, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As Sec. 3.3 mentioned, the attention map has high weights in regions needing high-frequency information. Therefore, the GRU with smaller kernel sizes that can capture high-frequency information like edge, and thin objects should do element-wise products with the attention map directly, and the GRU with larger kernel sizes should do element-wise products with the contrary attention map.", + "bbox": [ + 496, + 578, + 890, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Receptive fields analysis. The repetitive fields computing formula [1] can be defined as follows:", + "bbox": [ + 498, + 685, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {0} = \\sum_ {l = 1} ^ {L} \\left(\\left(k _ {l} - 1\\right) \\prod_ {i = 1} ^ {l - 1} s _ {i}\\right) + 1 \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 726, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k_{l}$ denotes the kernel size, $s_{i}$ denotes the stride size, and $r_0$ denotes the whole network.", + "bbox": [ + 498, + 779, + 888, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a multi-level structure like Fig. 4, if we take the $1/4$ resolution as the basis, and the downsampling operations can be regarded as a convolution with kernel size 3, stride size 2, this structure's receptive fields are $k$ , $2k + 3$ , $3k + 6$ . That means it only has 3 fixed receptive fields in total.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19704", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/64eb6c020445a9fe0255b3ca8b579a8dc09c2bd8e985e1439137001b300ae8db.jpg", + "image_caption": [ + "Figure 3. The architecture of proposed modules. Left: Contextual Spatial Attention (CSA) module. Right: Selective Recurrent Unit (SRU)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 607, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/174b196e660c161d4133595da944539a3c1c14aafe2e1b8dfe811c7b1b1c1d47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 89, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f616a01686288c415c2340eb03c56fd9b28f1b0d5cf4fa3948f9d77b21e3e579.jpg", + "image_caption": [ + "Figure 4. Multi-level SRU. Information is passed between SRUs at adjacent resolutions. Dashed arrows represent upsampling and downsampling operations. At $1/4$ resolution, disparity and local cost volume will be additional information put into SRUs." + ], + "image_footnote": [], + "bbox": [ + 109, + 276, + 439, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If we replace GRUs with our SRUs with a small kernel size $s$ and a large kernel size $l$ , the multi-level structure will have 6 receptive fields initially. Besides, pixels in hidden information are affected by different receptive fields during fusion, and the fusion is influenced by attention maps adaptively. In general, the multi-level SRU holds dynamic receptive fields, and it enables itself to capture information at different frequencies.", + "bbox": [ + 75, + 478, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 608, + 222, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We supervise our network on the L1 distance between all predicted disparities $\\{\\mathbf{d}_i\\}_{i=1}^N$ and the ground truth disparity $\\mathbf{d}_{gt}$ with increasing weights. The total loss is defined as:", + "bbox": [ + 75, + 632, + 468, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\gamma^ {N - i} | | \\mathbf {d} _ {i} - \\mathbf {d} _ {g t} | | _ {1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 688, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma = 0.9$ , and $N$ is the number of iterations.", + "bbox": [ + 76, + 739, + 406, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene Flow [20] is a synthetic dataset including 35,454 training pairs and 4,370 testing pairs with dense disparity maps. For training and testing, we use the finalpass version, because it contains more realistic and difficult effects than the cleanpass version. KITTI 2012 [13] and KITTI 2015 [21] are datasets for real-world driving scenes. KITTI 2012 contains 194 training pairs and 195 testing pairs, and", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KITTI 2015 contains 200 training pairs and 200 testing pairs. ETH3D [24] is a collection of gray-scale stereo pairs containing 27 training pairs and 20 testing pairs for indoor and outdoor scenes. Middlebury [23] is a high-resolution dataset containing 15 training pairs and 15 testing pairs for indoor scenes.", + "bbox": [ + 498, + 279, + 892, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 383, + 718, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our Selective-Stereo with PyTorch and the model is trained on NVIDIA RTX 3090 GPUs. For all experiments, we use the AdamW [19] optimizer and clip gradients to the range [-1, 1]. We use the one-cycle learning rate schedule with a learning rate of 2e-4. We first train our model on Scene Flow with a batch size of 8 for 200k steps as the pretrained model. The crop size is $320 \\times 720$ , and we use 22 update iterations during training.", + "bbox": [ + 496, + 407, + 892, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 541, + 653, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our model in different settings to verify our proposed modules in several aspects. All results use 32 update iterations.", + "bbox": [ + 496, + 565, + 890, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effectiveness of proposed modules. To verify the effectiveness of our proposed modules, we take RAFT-Stereo [17] as the baseline and replace its GRUs with our SRUs. As shown in Tab. 1, the proposed SRU can improve the accuracy even without CSA. It means that if we just sum up the information from different branches, the growth of receptive fields can be beneficial for inference. If we add our CSA but invert the weights of the attention maps, the effect even decreases. That validates that our CSA's attention maps do indeed reflect the weights of information at different frequencies in regions. Therefore, if we add CSA normally, the full model (Selective-RAFT) can achieve the best performance with only a $4\\%$ increase in parameters.", + "bbox": [ + 496, + 612, + 892, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Universality of proposed modules. To verify the universality of our proposed modules, we take three typical iterative stereo matching methods as the baseline and replace their GRUs with SRUs. Especially, in DLNR [44], the recurrent units are LSTMs but not GRUs, so we just replace GRUs inside SRUs with LSTMs to make a fair comparison.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19705", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/65eb3996d08f431b68de98e4b69bb12ee8ffe08f7cb4f15f2c0b080f68886111.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelGRUSRUCSA (Contrary)CSAEPE (px)>1px (%)Param (M)
Baseline (RAFT-Stereo)✓0.536.0811.12
SRU✓0.505.3811.65
SRU+CSA (Contrary)✓✓0.505.5811.65
Full model (Selective-RAFT)✓✓0.475.3211.65
", + "bbox": [ + 173, + 88, + 802, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24e5c71fca52740b8f87028bb5ceef85a4a2a280161b03979dcaab001a1af1ef.jpg", + "image_caption": [ + "Left Image" + ], + "image_footnote": [], + "bbox": [ + 120, + 236, + 362, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38290070a432c24a353b6772cfc56e730c36a518e438e39b5a804646eeb69b56.jpg", + "image_caption": [ + "IGEV-Stereo" + ], + "image_footnote": [], + "bbox": [ + 364, + 237, + 604, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08e2403e1370b9fb3d53a974377c462a39d86b1bb6a71ce008517a6ca7060f26.jpg", + "image_caption": [ + "Selective-IGEV" + ], + "image_footnote": [], + "bbox": [ + 604, + 237, + 849, + 412 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ed11bcac44344f5eda4484aab549be34955f9a741ff61e4c3f23de38989b25bc.jpg", + "table_caption": [ + "Table 1. Ablation study of the effectiveness of proposed modules on the Scene Flow test set. SRU denotes Selective Recurrent Unit, and CSA denotes Contextual Spatial Attention. Contrary means we invert the weights of the attention maps. The baseline is RAFT-Stereo." + ], + "table_footnote": [], + "table_body": "
ModelEPE (px)>1px (%)Param (M)
RAFT-Stereo [17]0.536.0811.12
Selective-RAFT0.475.3211.65
IGEV-Stereo [34]0.475.2112.60
Selective-IGEV0.444.9813.14
DLNR [44]0.495.0657.37
Selective-DLNR0.464.7358.09
", + "bbox": [ + 102, + 463, + 439, + 553 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/766495ecfd3a9e5cb6d9216e6ffd4796a3212f03f0bb330e3f57cf0433c126e7.jpg", + "table_caption": [ + "Table 2. Ablation study of the universality of proposed modules." + ], + "table_footnote": [], + "table_body": "
ModelNumber of Iterations
1234832
RAFT-Stereo [17]2.081.130.870.750.580.53
Selective-RAFT1.951.060.810.690.530.47
IGEV-Stereo [34]0.660.620.580.550.500.47
Selective-IGEV0.650.600.560.530.480.44
", + "bbox": [ + 84, + 582, + 457, + 659 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/094b275ef2a55acc698057ff28172b2c6e8bf606417ed76114c3eda3eed44a1d.jpg", + "table_caption": [ + "Table 3. Ablation study of the number of iterations." + ], + "table_footnote": [], + "table_body": "
Kernel SizesEPE (px)>1px (%)
1 × 1 + 1 × 50.485.41
3 × 3 + 1 × 50.485.30
1 × 1 + 3 × 30.475.32
", + "bbox": [ + 133, + 686, + 415, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Ablation study of the size of convolutional kernels.", + "bbox": [ + 94, + 763, + 450, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, all methods have a significant improvement in the EPE metrics on Scene Flow, and the insertion of modules only results in a slight increase in parameters. Besides, as shown in Fig. 6, the CSA module generates different attention maps in different networks. In Selective-RAFT, because the cost volume contains a large amount of noisy information, the network needs more large kernels to", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43de67bc4c680a303a7cea9c03bc11cd9d0e4e30f6d6850039e89935c548382b.jpg", + "image_caption": [ + "Figure 5. Qualitative results on the test set of KITTI. Our Selective-IGEV outperforms IGEV in detailed and weak texture regions.", + "Left Image", + "Figure 6. Visualization of the attention map of different networks." + ], + "image_footnote": [], + "bbox": [ + 501, + 464, + 630, + 532 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86638174669e0c84281a09b9c1658f3aebe77890f5641e689ec5e590eef68519.jpg", + "image_caption": [ + "Selective-RAFT" + ], + "image_footnote": [], + "bbox": [ + 632, + 464, + 759, + 532 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/510077035b70fddc24befb840a93f9f329a30fbf3447277e7b2bd6918fcdcce1.jpg", + "image_caption": [ + "Selective-IGEV" + ], + "image_footnote": [], + "bbox": [ + 761, + 464, + 890, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "filter the local cost volume. On the contrary, the cost volume has already been aggregated in Selective-IGEV, so the network tends to maintain high-frequency information using small kernels. Moreover, the cost volume in Selective-IGEV faces an over-smooth problem [34], and that's why the attention map tends to increase the weights of large kernels to recover edge regions. In general, the CSA module shows different tendencies in different networks, which is a reflection of its adaptive ability.", + "bbox": [ + 496, + 592, + 890, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of iterations. Our Selective-Stereo can achieve better performance with a smaller number of iterations. As shown in Tab. 3, our Selective-RAFT get the same performance with only 8 iterations compared to RAFT-Stereo [17], and for IGEV-Stereo [34], our Selective-IGEV also get a slight improvement with a few iterations. It shows that our modules can make secondary filtering to reduce noisy information from the initial cost volume.", + "bbox": [ + 496, + 731, + 893, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Size of convolution kernels. We verify different kernel sizes on Scene Flow as shown in Tab. 4. At last, we choose the combination of $1 \\times 1$ and $3 \\times 3$ as our default config-", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19706", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5e6c598556672dd025b6cbd792710d139ffae009bc867c76676becc88d30ac69.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCSPNLEASTereoLaC + GANetACVNetIGEV-StereoSelective-RAFT (Ours)Selective-IGEV (Ours)
EPE (px)0.780.780.720.480.470.470.44
", + "bbox": [ + 119, + 88, + 846, + 116 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/638230c846d3239a202260d0bc644636df2b1def048f131b6b59d3047d5f5bb6.jpg", + "table_caption": [ + "Table 5. Quantitative evaluation on Scene Flow test set." + ], + "table_footnote": [], + "table_body": "
MethodKITTI 2012KITTI 2015Run-time (s)
2-noc2-all3-noc3-allEPE-nocEPE-allD1-bgD1-fgD1-all
AcfNet [42]1.832.351.171.540.50.51.513.801.890.48
LEAStereo [9]1.902.391.131.450.50.51.402.911.650.30
ACVNet [32]1.832.351.131.470.40.51.373.071.650.20
RAFT-Stereo [17]1.922.421.301.660.40.51.583.051.820.38
PCWNet [25]1.692.181.041.370.40.51.583.051.820.38
LaC + GANet [18]1.722.261.051.420.40.51.442.831.671.80
CREStereo [16]1.722.181.141.460.40.51.452.861.690.41
IGEV-Stereo [34]1.712.171.121.440.40.41.382.671.590.18
Selective-RAFT (Ours)1.642.091.101.430.40.51.412.711.630.45
Selective-IGEV (Ours)1.592.051.071.380.40.41.332.611.550.24
", + "bbox": [ + 84, + 142, + 883, + 328 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a89ce0fc7a0482c8e5d7133e9995d7c5464154357d76b00af37fdf3bbf12690.jpg", + "table_caption": [ + "Table 6. Quantitative evaluation on KITTI 2012 and KITTI 2015." + ], + "table_footnote": [], + "table_body": "
MethodEdgesNon-Edges
EPE>1pxEPE>1px
RAFT-Stereo [17]3.2129.160.536.53
Selective-RAFT2.4021.630.404.65
IGEV-Stereo [34]2.2320.420.414.58
Selective-IGEV2.1820.010.384.35
", + "bbox": [ + 96, + 363, + 450, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Quantitative evaluation on Scene Flow test set in different regions.", + "bbox": [ + 75, + 470, + 468, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "uration, because it achieves a competitive performance and reduces computational costs.", + "bbox": [ + 75, + 513, + 468, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 556, + 377, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All fine-tuned models use the model pretrained on Scene Flow. Different target datasets use different finetune strategies. We validate two models called Selective-RAFT and Selective-IGEV using RAFT-Stereo [17] and IGEV-Stereo [34] as the baseline respectively.", + "bbox": [ + 75, + 580, + 468, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Flow. As shown in Tab. 5, we achieve a new state-of-the-art EPE of 0.44 on Scene Flow with Selective-IGEV, which surpasses $\\mathrm{LaC} + \\mathrm{GANet}$ [18] by $38.89\\%$ . Besides, our Selective-RAFT also achieves a competitive EPE of 0.47 compared to IGEV-Stereo [34] with smaller parameters. To validate the ability to fuse information by regions of our modules, we then split Scene Flow test set into two regions: edge regions and non-edge regions using the Canny operator. As shown in Tab. 7, our Selective-RAFT outperforms RAFT-Stereo [17] by $25.23\\%$ and $24.53\\%$ in edge regions and non-edge regions. Due to IGEV-Stereo's aggregated cost volume [34], there's only a slight improvement in edge regions, but our Selective-IGEV still outperforms it by $7.32\\%$ in non-edge regions.", + "bbox": [ + 75, + 657, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KITTI. We finetune our model on the mixed dataset of KITTI 2012 and KITTI 2015 with a batch size of 8 for 50k", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "steps. Then we evaluate our Selective-Stereo on the test set of KITTI 2012 and KITTI 2015. As shown in Tab. 6, we achieve the best performance among all published methods for almost all metrics. On KITTI 2012, our Selective-RAFT outperforms RAFT-Stereo [17] by $14.58\\%$ and $13.64\\%$ on 2-noc and 2-all metrics, and our Selective-IGEV ranks $1^{st}$ on these metrics. On KITTI 2015, our Selective-RAFT outperforms RAFT-Stereo [17] by $10.44\\%$ on the D1-all metric, and our Selective-IGEV ranks $1^{st}$ on all metrics with only 16 iterations same as IGEV-Stereo [34]. As shown in Fig. 5, our Selective-IGEV outperforms IGEV-Stereo [34] in detailed and textureless regions.", + "bbox": [ + 496, + 367, + 890, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ETH3D. Following CREStereo [16] and GMStereo [38], we use a collection of several public stereo datasets for training. The crop size is $384 \\times 512$ and we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Sintel Stereo [3], InStereo2k [2] and ETH3D [24] datasets for 300k steps. Then we finetune it on the mixed CREStereo Dataset [16], InStereo2k [2] and ETH3D [24] datasets with for another 90k steps. As shown in Tab. 8, our Selective-RAFT outperforms RAFT-Stereo [17] by $17.90\\%$ on Bad 0.5 metric, and our Selective-IGEV achieves the best performance among all published methods for almost all metrics.", + "bbox": [ + 496, + 550, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Middlebury. Also following CREStereo [16] and GMStereo [38], we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Falling Things [28], In-Stereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of $384 \\times 512$ for 200k steps. Then we finetune it on the mixed CREStereo Dataset [16], Falling Things [28], InStereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of $384 \\times 768$ with a batch size of 8 for another 100k steps. As shown in Tab. 8, our Selective-IGEV achieves the best performance among", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19707", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a0e93285602ec2cf21927260d1ee01f70b2198a555628782c81d56dc1c9399c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodETH3DMiddlebury
Bad 1.0Bad 0.5Bad 4.0AvgErrBad 2.0Bad 1.0Bad 4.0AvgErr
CroCo-Stereo [30]0.993.270.130.147.2916.94.181.76
GMStereo [38]1.835.940.080.197.1423.62.961.31
HITNet [26]2.797.830.190.206.4613.33.811.71
IGEV-Stereo [34]1.123.520.110.144.839.413.332.89
RAFT-Stereo [17]2.447.040.150.184.749.372.751.27
CREStereo [16]0.983.580.100.133.718.252.041.15
EAI-Stereo [43]2.315.210.700.213.687.812.141.09
DLNR [44]----3.206.821.891.06
Selective-RAFT (Ours)1.695.780.130.17----
Selective-IGEV (Ours)1.233.060.050.122.516.531.360.91
", + "bbox": [ + 119, + 88, + 851, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Quantitative evaluation on ETH3D and Middlebury benchmarks. Note: Middlebury only allows one publish per paper, so we only publish our Selective-IGEV.", + "bbox": [ + 75, + 284, + 892, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/696c137fe7da6974ba62c2d51b1762e3f0a790406419b4f71189d176ab67387e.jpg", + "image_caption": [ + "Figure 7. Qualitative results on the test set of Middlebury. The third column is the visualization of attention maps generated by the CSA module. Our Selective-IGEV outperforms IGEV in large textureless and thin object regions." + ], + "image_footnote": [], + "bbox": [ + 138, + 313, + 831, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all published methods. As shown in Fig 7, compared to IGEV-Stereo [34], our Selective-IGEV performs better in textureless, and detailed regions. The third column in Fig 7 is the visualization of attention maps generated by the CSA module. It shows that attention maps can surely split regions that require information at different frequencies.", + "bbox": [ + 75, + 654, + 468, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 767, + 194, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Selective-Stereo, a novel iterative stereo matching method. The proposed Contextual Spatial Attention module and Selective Recurrent Unit help the network capture information at different frequencies for edge and smooth regions. Our Selective-Stereo ranks $1^{st}$ on KITTI, ETH3D, and Middlebury in almost all metrics among all published methods. It shows an ability to fuse information", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "at different frequencies adaptively for edge and smooth regions with the help of attention maps extracted by CSA.", + "bbox": [ + 496, + 654, + 890, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, our method still faces some challenges. Firstly, although our method can fuse information adaptively using attention maps, the SRU's receptive field is still limited by predefined values. Secondly, adding branches or increasing the sizes of convolutional kernels leads to high memory and time costs, so we will explore the combination of lightweight convolutions and our method to reduce memory costs. Lastly, it's also a good direction to do research on the combination of convolutions and self-attention due to their different advantages and receptive fields.", + "bbox": [ + 496, + 685, + 892, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research is supported by National Natural Science Foundation of China (62122029, 62061160490, U20B200007).", + "bbox": [ + 496, + 851, + 892, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 4 (11):e21, 2019. 4", + "[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 7", + "[3] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7", + "[4] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1, 2", + "[5] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Shuicheng Yan, and Jiashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3435-3444, 2019. 2, 3", + "[6] Yajie Chen, Xin Yang, and Xiang Bai. Confidence-weighted mutual supervision on dual networks for unsupervised cross-modality image segmentation. Science China Information Sciences, 66(11):210104, 2023. 3", + "[7] Junda Cheng, Xin Yang, Yuechuan Pu, and Peng Guo. Region separable stereo matching. IEEE Transactions on Multimedia, 2022. 1, 2", + "[8] Junda Cheng, Gangwei Xu, Peng Guo, and Xin Yang. Coatsnet: Fully exploiting convolution and attention for stereo matching by region separation. International Journal of Computer Vision, pages 1-18, 2023. 2", + "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33:22158-22169, 2020. 7", + "[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 2", + "[11] Miaojie Feng, Junda Cheng, Hao Jia, Longliang Liu, Gangwei Xu, and Xin Yang. Mc-stereo: Multi-peak lookup and cascade search range for stereo matching. arXiv preprint arXiv:2311.02340, 2023. 1", + "[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 3", + "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2, 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3273-3282, 2019. 1, 2", + "[15] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision, pages 66-75, 2017. 1, 2", + "[16] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1, 3, 7, 8", + "[17] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[18] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1647-1655, 2022. 7", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[20] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 2, 5, 7", + "[21] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 5", + "[22] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Fast vision transformers with hilo attention. Advances in Neural Information Processing Systems, 35:14541-14554, 2022. 3", + "[23] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 1, 2, 5, 7", + "[24] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 1, 2, 5, 7", + "[25] Zhelun Shen, Yuchao Dai, Xibin Song, Zhibo Rao, Dingfu Zhou, and Liangjun Zhang. Pcw-net: Pyramid combination and warping cost volume for stereo matching. In European Conference on Computer Vision, pages 280–297. Springer, 2022. 2, 7", + "[26] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19709", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 8", + "[27] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 2, 3", + "[28] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 7", + "[29] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 7", + "[30] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 8", + "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 4", + "[32] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 1, 2, 7", + "[33] Gangwei Xu, Shujun Chen, Hao Jia, Miaojie Feng, and Xin Yang. Memory-efficient optical flow via radius-distribution orthogonal cost volume. arXiv preprint arXiv:2312.03790, 2023. 2", + "[34] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 3, 6, 7, 8", + "[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2", + "[36] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023.", + "[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1959-1968, 2020. 2", + "[38] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow," + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7, 8", + "[39] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1740-1749, 2020. 3", + "[40] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5515-5524, 2019. 7", + "[41] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 1, 2", + "[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12926-12934, 2020. 7", + "[43] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Yong Zhao, Yitong Yang, and Ting Ouyang. Eai-stereo: Error aware iterative network for stereo matching. In Proceedings of the Asian Conference on Computer Vision, pages 315-332, 2022. 8", + "[44] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Jie Chen, Yitong Yang, and Yong Zhao. High-frequency stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1327-1336, 2023. 1, 2, 3, 5, 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 527 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19710", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_model.json b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f2430fc8ecddfc55bd2233962b0df9b45cc8f19b --- /dev/null +++ b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.892, + 0.155 + ], + "angle": 0, + "content": "Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.18, + 0.692, + 0.201 + ], + "angle": 0, + "content": "Xianqi Wang*, Gangwei Xu*, Hao Jia, Xin Yang†" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.206, + 0.68, + 0.224 + ], + "angle": 0, + "content": "Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.226, + 0.695, + 0.242 + ], + "angle": 0, + "content": "{xianqiw, gwxu, haojia, xinyang2014}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.276, + 0.314, + 0.293 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.474, + 0.658 + ], + "angle": 0, + "content": "Stereo matching methods based on iterative optimization, like RAFT-Stereo and IGEV-Stereo, have evolved into a cornerstone in the field of stereo matching. However, these methods struggle to simultaneously capture high-frequency information in edges and low-frequency information in smooth regions due to the fixed receptive field. As a result, they tend to lose details, blur edges, and produce false matches in textureless areas. In this paper, we propose Selective Recurrent Unit (SRU), a novel iterative update operator for stereo matching. The SRU module can adaptively fuse hidden disparity information at multiple frequencies for edge and smooth regions. To perform adaptive fusion, we introduce a new Contextual Spatial Attention (CSA) module to generate attention maps as fusion weights. The SRU empowers the network to aggregate hidden disparity information across multiple frequencies, mitigating the risk of vital hidden disparity information loss during iterative processes. To verify SRU's universality, we apply it to representative iterative stereo matching methods, collectively referred to as Selective-Stereo. Our Selective-Stereo ranks \\(1^{st}\\) on KITTI 2012, KITTI 2015, ETH3D, and Middlebury leaderboards among all published methods. Code is available at https://github.com/Windsrain/Selective-Stereo." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.701, + 0.21, + 0.717 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.818 + ], + "angle": 0, + "content": "Stereo matching is a fundamental area of research in computer vision. It explores the calculation of displacement, referred to as disparity, between matching points in a pair of rectified images. This technique plays a significant role in various applications, including 3D reconstruction and autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.471, + 0.866 + ], + "angle": 0, + "content": "With the advancement of deep learning, learning-based stereo matching [4, 14, 15, 17, 34, 35] have progressively displaced traditional methods and significantly enhancing" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.275, + 0.695, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.275, + 0.882, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.426, + 0.628, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.495, + 0.596, + 0.506 + ], + "angle": 0, + "content": "Left Image" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.426, + 0.756, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.495, + 0.742, + 0.506 + ], + "angle": 0, + "content": "RAFT-Stereo [17]" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.426, + 0.885, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.495, + 0.865, + 0.505 + ], + "angle": 0, + "content": "Selective-RAFT" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.513, + 0.629, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.581, + 0.596, + 0.593 + ], + "angle": 0, + "content": "Left Image" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.512, + 0.756, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.581, + 0.74, + 0.593 + ], + "angle": 0, + "content": "IGEV-Stereo [34]" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.512, + 0.885, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.581, + 0.863, + 0.592 + ], + "angle": 0, + "content": "Selective-IGEV" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.594, + 0.894, + 0.677 + ], + "angle": 0, + "content": "Figure 1. Row 1: Comparisons with state-of-the-art stereo methods on KITTI 2012 [13] and KITTI 2015 [21], ETH3D [24] and Middlebury [23] leaderboards. Row 2: Visual comparison with RAFT-Stereo on ETH3D. Row 3: Visual comparison with IGEV-Stereo on Middlebury. Our method distinguishes subtle details and sharp edges and performs well in weak texture regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.893, + 0.867 + ], + "angle": 0, + "content": "the accuracy of disparity estimation. Initially, aggregation-based methods [7, 32, 41] led the development of stereo matching algorithms. These methods begin by defining a maximum range of disparity, constructing a 4D cost volume using feature maps, and subsequently employing 3D CNN to filter the volume and derive the final disparity map. Such methods focus on filtering the initially coarse cost volume, thus effectively aggregating geometry information. However, cost aggregation requires a large number of convolutions, resulting in high computational costs, making it difficult to be applied to high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Recently, a novel class of methods based on iterative optimization [11, 16, 17, 44] has been gaining prominence and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.207, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.871, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19701" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "achieving state-of-the-art performance on several leaderboards. These methods begin by constructing an all-pairs cost volume, indexing a local cost volume from the original cost volume, and subsequently employing recurrent units [10] to calculate disparity residuals and update the disparity prediction. One major advantage of these methods is their ability to capture all candidate matching points without predefining the range of disparities. Additionally, these methods don't need to aggregate the cost volume using a large number of redundant convolutions. Instead, a continuous update of the disparity prediction is achieved through lightweight recurrent units during iterations. Therefore, these methods are capable of processing high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.473, + 0.489 + ], + "angle": 0, + "content": "However, iterative methods encounter several challenges. Firstly, the all-pairs cost volume includes considerable noisy information [34], potentially causing the loss of crucial information when iterating the hidden information. Besides, as the network iterates, the hidden information increasingly incorporates global low-frequency information while losing local high-frequency information like edges and thin objects [44]. Secondly, the existing recurrent units possess a fixed receptive field, leading the network to solely concentrate on information at the current frequency and ignore other frequencies, such as detailed, edge, and textureless information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose Selective Recurrent Unit (SRU) to address the limitations of traditional recurrent units. As Chen et al. [5] mentions features contain information at different frequencies, high-frequency information describes rapidly changing fine details, while low-frequency information describes smoothly changing structures. Unlike traditional recurrent units that treat information at different frequencies equally, our SRU incorporates multiple branches of GRU, each with a distinct kernel size representing different receptive fields. The hidden information obtained from each GRU branch is fused and then fed into the next iteration. This fusion enables the capture of information from different receptive fields at different frequencies, while also performing secondary filtering to reduce noise information from local cost volume. To further enhance the fusion process, we propose a Contextual Spatial Attention (CSA) module to utilize the context information. Instead of simply summarizing information from different branches, CSA introduces attention maps extracted from the context information. After doing so, information captured by small kernels has large weights in regions like edge, while information captured by large kernels has large weights in regions like low-texture. These attention maps determine the weight of fusion, allowing the network to adaptively select suitable information based on different image regions. Besides, we prove the effectiveness and universality of our module by transferring it to different iterative networks. All networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "are collectively referred to as Selective-Stereo. By doing so, we consistently improve the performance of these networks without introducing a significant increase in parameters and time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.895, + 0.303 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of our method on several stereo benchmarks. On Scene Flow [20], our Selective-RAFT reaches the state-of-the-art EPE of 0.47, and our Selective-IGEV even achieves a new state-of-the-art EPE of 0.44. And as shown in Fig. 1, our Selective-RAFT surpasses RAFT-Stereo by a large margin and achieves competitive performance compared with the state-of-the-art methods on KITTI [13, 21] leaderboards. Our Selective-IGEV ranks \\(1^{st}\\) on KITTI, ETH3D [24], and Middlebury [23] leaderboards among all published methods." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.304, + 0.882, + 0.318 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.32, + 0.89, + 0.348 + ], + "angle": 0, + "content": "- We propose a novel iterative update operator SRU for iterative stereo matching methods." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.349, + 0.891, + 0.393 + ], + "angle": 0, + "content": "- We introduce a new Contextual Spatial Attention module that generates attention maps for adaptively fusing hidden disparity information at multiple frequencies." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.394, + 0.89, + 0.422 + ], + "angle": 0, + "content": "- We verify the universality of our SRU on several iterative stereo matching methods." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.424, + 0.891, + 0.469 + ], + "angle": 0, + "content": "- Our method outperforms existing published methods on public leaderboards such as KITTI, ETH3D, and Middlebury." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.32, + 0.891, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.483, + 0.642, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.895, + 0.841 + ], + "angle": 0, + "content": "Aggregation-based methods in stereo matching. Several aggregation-based methods [4, 7, 8, 14, 15, 25, 35-37, 41] have shown significant progress in the domain of stereo matching in recent years. DispNet [20] establishes the groundwork for subsequent network architecture. GCNet [15] proposes a 4D concatenate cost volume, which is subsequently regularized using 3D CNNs. Additionally, it also introduces the soft argmin function for disparity regression, resulting in a significant influence on subsequent methods. PSMNet [4] proposes a stacked hourglass 3D CNN, which improves the cost aggregation stage to enhance the network's ability to capture context information. GwcNet [14] proposes Group-wise Correlation Volume, which combines the advantages of correlation and concatenation volume. GA-Net [41] designs a semi-global guided aggregation layer and a local guided aggregation layer, inspired by the traditional semi-global matching algorithm, to further assist in aggregating global and geometry information in the network. Building upon Group-wise Correlation Volume, ACVNet [32] proposes Attention Concat Volume, which uses attention weights to suppress redundant information and maintain sufficient information for matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Iterative-based methods in stereo matching. In recent years, many iterative methods [17, 27, 33, 34], spearheaded by RAFT [17], have gradually become the mainstream of research. RAFT-Stereo [17] builds upon the optical flow" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19702" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.443, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed Selective-Stereo (Selective-RAFT version). The Contextual Spatial Attention (CSA) module extracts attention maps from context information as a guide for Selective Recurrent Units (SRUs). Then the network iteratively updates the disparity using local cost volumes retrieved from the correlation pyramid and attention maps given by CSA through SRUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.473, + 0.727 + ], + "angle": 0, + "content": "method RAFT [27] by introducing an all-pairs cost volume pyramid that maintains high resolution. It extracts local correlation features from this pyramid, performs iterative disparity updates using GRU-based update operators, and incorporates a multi-level GRU to expand the receptive field. On this basis, IGEV-Stereo [34] asserts that the initial cost volume is excessively coarse. To alleviate the need for iterations and reduce time overhead, it proposes to use a lightweight cost aggregation network before iterations. CREstereo [16] designs a hierarchical network in a coarse-to-fine manner, as well as a stacked cascaded architecture for inference in place of the original single-resolution iterative structure. DLNR [44] proposes the use of LSTM as a replacement for GRU, providing the advantage of decoupling the update of hidden states from disparity prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Frequency information application in vision. There are several works[5, 6, 39] that focus on using frequency information in computer vision. Chen et al. [5] propose the octave convolution to factorize the mixed feature maps by their frequencies. Xu et al. [39] propose a method of learning in the frequency domain and suggest that CNN models are more sensitive to low-frequency channels than high-frequency. DSGAN [12] introduces the frequency separation into super-resolution. LITv2 [22] proposes to disentangle the high/low-frequency patterns in an attention layer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.591, + 0.514 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.893, + 0.588 + ], + "angle": 0, + "content": "In this section, we present the overall architecture of Selective-Stereo. Because our method can be plugged into different networks, we take Selective-RAFT (Fig. 2) as an example and focus on illustrating its key components." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.609, + 0.682, + 0.624 + ], + "angle": 0, + "content": "3.1. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.697 + ], + "angle": 0, + "content": "To ensure fair comparisons, Selective-RAFT maintains consistency with RAFT-Stereo [17] by employing its feature extraction network. Feature extraction comprises two main components: feature network and context network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Feature Network. Given the left and the right images \\(\\mathbf{I}_{l(r)}\\in \\mathbb{R}^{3\\times H\\times W}\\), we first downsample them to \\(1 / 2\\) resolution using a \\(7\\times 7\\) convolutional layer. Then, a series of residual blocks is employed to extract features and we apply another downsampling layer to get features at \\(1 / 4\\) resolution in the middle. Lastly, a \\(1\\times 1\\) convolutional layer is applied to get the final left and right features \\(\\mathbf{f}\\), \\(\\mathbf{g}\\in \\mathbb{R}^{C\\times \\frac{H}{4}\\times \\frac{W}{4}}\\) with suitable dimensions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Context Network. Its architecture remains consistent with the feature network, and it adds a series of residual blocks and two additional downsampling layers, obtaining multi-level context features \\(\\mathbf{f}_i^c\\) (\\(i = 1, 2, 3\\)) at \\(1/4, 1/8, 1/16\\) resolutions. Then we can get the initial hidden and the con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19703" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.193, + 0.105 + ], + "angle": 0, + "content": "text information:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.103, + 0.47, + 0.127 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {i} = \\tanh \\left(\\mathbf {f} _ {i} ^ {c}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.123, + 0.326, + 0.139 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i} = \\operatorname {R e L U} \\left(\\mathbf {f} _ {i} ^ {c}\\right)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.155, + 0.317, + 0.169 + ], + "angle": 0, + "content": "3.2. Cost Volume Construction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.178, + 0.47, + 0.208 + ], + "angle": 0, + "content": "Given the left and the right features \\(\\mathbf{f},\\mathbf{g}\\), we first construct an all-pairs correlate cost volume:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.218, + 0.469, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {i j k} = \\sum_ {h} \\mathbf {f} _ {h i j} \\cdot \\mathbf {g} _ {h i k}, \\mathbf {C} \\in \\mathbb {R} ^ {\\frac {H}{4} \\times \\frac {W}{4} \\times \\frac {W}{4}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.262, + 0.471, + 0.307 + ], + "angle": 0, + "content": "Then we construct a 4-level correlation pyramid \\(\\{\\mathbf{C}_i\\}\\) (\\(i = 1,2,3,4\\)) by using 1D average pooling with a kernel size of 2 and a stride of 2 at the last dimension." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.317, + 0.398, + 0.332 + ], + "angle": 0, + "content": "3.3. Contextual Spatial Attention Module" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.469, + 0.46 + ], + "angle": 0, + "content": "To help information from different receptive fields and frequencies fuse, the Contextual Spatial Attention (CSA) module extracts multi-level attention maps from context information as guidance. As illustrated in Fig. 3, CSA can be divided into two submodules: Channel Attention Enhancement (CAE) and Spatial Attention Extractor (SAE). These submodules are derived from CBAM [31] and we simplify them to better adapt to stereo matching." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.462, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Channel Attention Enhancement. Given a context information map \\(\\mathbf{c} \\in \\mathbb{R}^{C \\times H \\times W}\\), we first use an average-pooling and a max-pooling operation on the spatial dimension to get two maps \\(\\mathbf{f}_{avg}, \\mathbf{f}_{max} \\in \\mathbb{R}^{C \\times 1 \\times 1}\\). Then we use two convolutional layers to perform feature transformation on these maps separately. After that, we add these two maps together and use the sigmoid function to convert them into weights \\(\\mathbf{M}_c \\in \\mathbb{R}^{C \\times 1 \\times 1}\\) between 0 and 1. Lastly, using an element-wise product, the initial map can capture which channel map has high feature values to be enhanced, and which channel map has low feature values to be suppressed." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Spatial Attention Extractor. After the CAE module, we continue to use the same pooling operations, but now we pool on the channel dimension. Then we concatenate these pooling maps to form a map in \\(\\mathbb{R}^{2\\times H\\times W}\\) and use one convolutional layer with a sigmoid function to generate the final attention map. Reviewing previous operations, this attention map has high weights in regions needing high-frequency information because this information possesses high feature values in the context information. Similarly, it has low weights in regions needing low-frequency information. In general, the attention map can explicitly distinguish regions that need information at different frequencies." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.818, + 0.303, + 0.832 + ], + "angle": 0, + "content": "3.4. Selective Recurrent Unit" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "To capture information at different frequencies, Selective Recurrent Unit (SRU) uses attention maps extracted by CSA to fuse hidden information derived from GRUs with different kernel sizes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "Multi-level update structure. As illustrated in Fig. 4, SRUs at \\(1/8\\), \\(1/16\\) resolutions take the attention map, context information, hidden information at the same resolution, and the hidden information at adjacent resolutions as inputs. At \\(1/4\\) resolution, SRUs take disparity, and local cost volume as additional inputs, and then their outputs will go through two convolutional layers to generate disparity residuals. The local cost volume is derived from the all-pairs correlation pyramid in the same way as RAFT-Stereo [17]. At last, disparities at \\(1/4\\) resolution will be upsampled into full resolution using the convex combination." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.258, + 0.892, + 0.287 + ], + "angle": 0, + "content": "SRU's architecture. A single GRU can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.297, + 0.771, + 0.314 + ], + "angle": 0, + "content": "\\[\nz _ {k} = \\sigma (\\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {z})),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.316, + 0.892, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} r _ {k} = \\sigma (\\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {r})), \\\\ \\tilde {I} _ {\\text {一}} = \\operatorname {c o v} (G _ {m} ([ \\circ_ {0} ]), W _ {r})) \\end{array} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.331, + 0.891, + 0.355 + ], + "angle": 0, + "content": "\\[\n\\tilde {h} _ {k} = \\tanh (\\operatorname {C o n v} ([ r _ {k} \\odot h _ {k - 1}, x _ {k} ], W _ {h})), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.358, + 0.782, + 0.376 + ], + "angle": 0, + "content": "\\[\nh _ {k} = \\left(1 - z _ {k}\\right) \\odot h _ {k - 1} + z _ {k} \\odot \\tilde {h} _ {k}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.476 + ], + "angle": 0, + "content": "where \\( x_{k} \\) is the concatenation of disparity, correlation, hidden information, and context information previously defined. Unlike RAFT-Stereo [17] that divide the context information into \\( c_{z}, c_{r}, c_{h} \\), we add it into \\( x_{k} \\) because using convolutions with different kernel sizes can fully utilize context information." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.892, + 0.507 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, a single SRU can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.508, + 0.892, + 0.527 + ], + "angle": 0, + "content": "\\[\nh _ {k} = \\mathbf {A} \\odot h _ {k} ^ {s} + (1 - \\mathbf {A}) \\odot h _ {k} ^ {l} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.578 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\) denotes the attention map derived from CSA at the same resolution, \\(h_k^s\\) denotes the GRU with smaller kernel sizes and \\(h_k^l\\) denotes the larger one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.685 + ], + "angle": 0, + "content": "As Sec. 3.3 mentioned, the attention map has high weights in regions needing high-frequency information. Therefore, the GRU with smaller kernel sizes that can capture high-frequency information like edge, and thin objects should do element-wise products with the attention map directly, and the GRU with larger kernel sizes should do element-wise products with the contrary attention map." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.686, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Receptive fields analysis. The repetitive fields computing formula [1] can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.727, + 0.892, + 0.769 + ], + "angle": 0, + "content": "\\[\nr _ {0} = \\sum_ {l = 1} ^ {L} \\left(\\left(k _ {l} - 1\\right) \\prod_ {i = 1} ^ {l - 1} s _ {i}\\right) + 1 \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.89, + 0.809 + ], + "angle": 0, + "content": "where \\( k_{l} \\) denotes the kernel size, \\( s_{i} \\) denotes the stride size, and \\( r_0 \\) denotes the whole network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Given a multi-level structure like Fig. 4, if we take the \\(1/4\\) resolution as the basis, and the downsampling operations can be regarded as a convolution with kernel size 3, stride size 2, this structure's receptive fields are \\(k\\), \\(2k + 3\\), \\(3k + 6\\). That means it only has 3 fixed receptive fields in total." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19704" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.608, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.09, + 0.891, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.893, + 0.267 + ], + "angle": 0, + "content": "Figure 3. The architecture of proposed modules. Left: Contextual Spatial Attention (CSA) module. Right: Selective Recurrent Unit (SRU)." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.277, + 0.441, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.409, + 0.471, + 0.466 + ], + "angle": 0, + "content": "Figure 4. Multi-level SRU. Information is passed between SRUs at adjacent resolutions. Dashed arrows represent upsampling and downsampling operations. At \\(1/4\\) resolution, disparity and local cost volume will be additional information put into SRUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.601 + ], + "angle": 0, + "content": "If we replace GRUs with our SRUs with a small kernel size \\( s \\) and a large kernel size \\( l \\), the multi-level structure will have 6 receptive fields initially. Besides, pixels in hidden information are affected by different receptive fields during fusion, and the fusion is influenced by attention maps adaptively. In general, the multi-level SRU holds dynamic receptive fields, and it enables itself to capture information at different frequencies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.609, + 0.223, + 0.625 + ], + "angle": 0, + "content": "3.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.679 + ], + "angle": 0, + "content": "We supervise our network on the L1 distance between all predicted disparities \\(\\{\\mathbf{d}_i\\}_{i=1}^N\\) and the ground truth disparity \\(\\mathbf{d}_{gt}\\) with increasing weights. The total loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.689, + 0.47, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\gamma^ {N - i} | | \\mathbf {d} _ {i} - \\mathbf {d} _ {g t} | | _ {1} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.741, + 0.408, + 0.756 + ], + "angle": 0, + "content": "where \\(\\gamma = 0.9\\), and \\(N\\) is the number of iterations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.21, + 0.788 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Scene Flow [20] is a synthetic dataset including 35,454 training pairs and 4,370 testing pairs with dense disparity maps. For training and testing, we use the finalpass version, because it contains more realistic and difficult effects than the cleanpass version. KITTI 2012 [13] and KITTI 2015 [21] are datasets for real-world driving scenes. KITTI 2012 contains 194 training pairs and 195 testing pairs, and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.28, + 0.893, + 0.37 + ], + "angle": 0, + "content": "KITTI 2015 contains 200 training pairs and 200 testing pairs. ETH3D [24] is a collection of gray-scale stereo pairs containing 27 training pairs and 20 testing pairs for indoor and outdoor scenes. Middlebury [23] is a high-resolution dataset containing 15 training pairs and 15 testing pairs for indoor scenes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.384, + 0.719, + 0.4 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.893, + 0.53 + ], + "angle": 0, + "content": "We implement our Selective-Stereo with PyTorch and the model is trained on NVIDIA RTX 3090 GPUs. For all experiments, we use the AdamW [19] optimizer and clip gradients to the range [-1, 1]. We use the one-cycle learning rate schedule with a learning rate of 2e-4. We first train our model on Scene Flow with a batch size of 8 for 200k steps as the pretrained model. The crop size is \\(320 \\times 720\\), and we use 22 update iterations during training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.542, + 0.654, + 0.559 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.612 + ], + "angle": 0, + "content": "In this section, we evaluate our model in different settings to verify our proposed modules in several aspects. All results use 32 update iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.809 + ], + "angle": 0, + "content": "Effectiveness of proposed modules. To verify the effectiveness of our proposed modules, we take RAFT-Stereo [17] as the baseline and replace its GRUs with our SRUs. As shown in Tab. 1, the proposed SRU can improve the accuracy even without CSA. It means that if we just sum up the information from different branches, the growth of receptive fields can be beneficial for inference. If we add our CSA but invert the weights of the attention maps, the effect even decreases. That validates that our CSA's attention maps do indeed reflect the weights of information at different frequencies in regions. Therefore, if we add CSA normally, the full model (Selective-RAFT) can achieve the best performance with only a \\(4\\%\\) increase in parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Universality of proposed modules. To verify the universality of our proposed modules, we take three typical iterative stereo matching methods as the baseline and replace their GRUs with SRUs. Especially, in DLNR [44], the recurrent units are LSTMs but not GRUs, so we just replace GRUs inside SRUs with LSTMs to make a fair comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19705" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.174, + 0.089, + 0.803, + 0.187 + ], + "angle": 0, + "content": "
ModelGRUSRUCSA (Contrary)CSAEPE (px)>1px (%)Param (M)
Baseline (RAFT-Stereo)✓0.536.0811.12
SRU✓0.505.3811.65
SRU+CSA (Contrary)✓✓0.505.5811.65
Full model (Selective-RAFT)✓✓0.475.3211.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.195, + 0.895, + 0.224 + ], + "angle": 0, + "content": "Table 1. Ablation study of the effectiveness of proposed modules on the Scene Flow test set. SRU denotes Selective Recurrent Unit, and CSA denotes Contextual Spatial Attention. Contrary means we invert the weights of the attention maps. The baseline is RAFT-Stereo." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.237, + 0.363, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.415, + 0.274, + 0.428 + ], + "angle": 0, + "content": "Left Image" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.238, + 0.605, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.415, + 0.531, + 0.426 + ], + "angle": 0, + "content": "IGEV-Stereo" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.238, + 0.85, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.415, + 0.783, + 0.426 + ], + "angle": 0, + "content": "Selective-IGEV" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.439, + 0.871, + 0.455 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on the test set of KITTI. Our Selective-IGEV outperforms IGEV in detailed and weak texture regions." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.464, + 0.441, + 0.554 + ], + "angle": 0, + "content": "
ModelEPE (px)>1px (%)Param (M)
RAFT-Stereo [17]0.536.0811.12
Selective-RAFT0.475.3211.65
IGEV-Stereo [34]0.475.2112.60
Selective-IGEV0.444.9813.14
DLNR [44]0.495.0657.37
Selective-DLNR0.464.7358.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.558, + 0.465, + 0.573 + ], + "angle": 0, + "content": "Table 2. Ablation study of the universality of proposed modules." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.583, + 0.458, + 0.66 + ], + "angle": 0, + "content": "
ModelNumber of Iterations
1234832
RAFT-Stereo [17]2.081.130.870.750.580.53
Selective-RAFT1.951.060.810.690.530.47
IGEV-Stereo [34]0.660.620.580.550.500.47
Selective-IGEV0.650.600.560.530.480.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.664, + 0.427, + 0.677 + ], + "angle": 0, + "content": "Table 3. Ablation study of the number of iterations." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.688, + 0.416, + 0.753 + ], + "angle": 0, + "content": "
Kernel SizesEPE (px)>1px (%)
1 × 1 + 1 × 50.485.41
3 × 3 + 1 × 50.485.30
1 × 1 + 3 × 30.475.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.764, + 0.452, + 0.779 + ], + "angle": 0, + "content": "Table 4. Ablation study of the size of convolutional kernels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As shown in Tab. 2, all methods have a significant improvement in the EPE metrics on Scene Flow, and the insertion of modules only results in a slight increase in parameters. Besides, as shown in Fig. 6, the CSA module generates different attention maps in different networks. In Selective-RAFT, because the cost volume contains a large amount of noisy information, the network needs more large kernels to" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.465, + 0.631, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.535, + 0.596, + 0.547 + ], + "angle": 0, + "content": "Left Image" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.465, + 0.761, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.535, + 0.738, + 0.546 + ], + "angle": 0, + "content": "Selective-RAFT" + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.465, + 0.891, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.535, + 0.866, + 0.546 + ], + "angle": 0, + "content": "Selective-IGEV" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.558, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Figure 6. Visualization of the attention map of different networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.729 + ], + "angle": 0, + "content": "filter the local cost volume. On the contrary, the cost volume has already been aggregated in Selective-IGEV, so the network tends to maintain high-frequency information using small kernels. Moreover, the cost volume in Selective-IGEV faces an over-smooth problem [34], and that's why the attention map tends to increase the weights of large kernels to recover edge regions. In general, the CSA module shows different tendencies in different networks, which is a reflection of its adaptive ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.894, + 0.852 + ], + "angle": 0, + "content": "Number of iterations. Our Selective-Stereo can achieve better performance with a smaller number of iterations. As shown in Tab. 3, our Selective-RAFT get the same performance with only 8 iterations compared to RAFT-Stereo [17], and for IGEV-Stereo [34], our Selective-IGEV also get a slight improvement with a few iterations. It shows that our modules can make secondary filtering to reduce noisy information from the initial cost volume." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Size of convolution kernels. We verify different kernel sizes on Scene Flow as shown in Tab. 4. At last, we choose the combination of \\(1 \\times 1\\) and \\(3 \\times 3\\) as our default config-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19706" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.848, + 0.117 + ], + "angle": 0, + "content": "
MethodCSPNLEASTereoLaC + GANetACVNetIGEV-StereoSelective-RAFT (Ours)Selective-IGEV (Ours)
EPE (px)0.780.780.720.480.470.470.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.32, + 0.121, + 0.651, + 0.134 + ], + "angle": 0, + "content": "Table 5. Quantitative evaluation on Scene Flow test set." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.143, + 0.885, + 0.329 + ], + "angle": 0, + "content": "
MethodKITTI 2012KITTI 2015Run-time (s)
2-noc2-all3-noc3-allEPE-nocEPE-allD1-bgD1-fgD1-all
AcfNet [42]1.832.351.171.540.50.51.513.801.890.48
LEAStereo [9]1.902.391.131.450.50.51.402.911.650.30
ACVNet [32]1.832.351.131.470.40.51.373.071.650.20
RAFT-Stereo [17]1.922.421.301.660.40.51.583.051.820.38
PCWNet [25]1.692.181.041.370.40.51.583.051.820.38
LaC + GANet [18]1.722.261.051.420.40.51.442.831.671.80
CREStereo [16]1.722.181.141.460.40.51.452.861.690.41
IGEV-Stereo [34]1.712.171.121.440.40.41.382.671.590.18
Selective-RAFT (Ours)1.642.091.101.430.40.51.412.711.630.45
Selective-IGEV (Ours)1.592.051.071.380.40.41.332.611.550.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.339, + 0.68, + 0.353 + ], + "angle": 0, + "content": "Table 6. Quantitative evaluation on KITTI 2012 and KITTI 2015." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.364, + 0.451, + 0.461 + ], + "angle": 0, + "content": "
MethodEdgesNon-Edges
EPE>1pxEPE>1px
RAFT-Stereo [17]3.2129.160.536.53
Selective-RAFT2.4021.630.404.65
IGEV-Stereo [34]2.2320.420.414.58
Selective-IGEV2.1820.010.384.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.471, + 0.47, + 0.499 + ], + "angle": 0, + "content": "Table 7. Quantitative evaluation on Scene Flow test set in different regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.545 + ], + "angle": 0, + "content": "uration, because it achieves a competitive performance and reduces computational costs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.379, + 0.573 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.658 + ], + "angle": 0, + "content": "All fine-tuned models use the model pretrained on Scene Flow. Different target datasets use different finetune strategies. We validate two models called Selective-RAFT and Selective-IGEV using RAFT-Stereo [17] and IGEV-Stereo [34] as the baseline respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Scene Flow. As shown in Tab. 5, we achieve a new state-of-the-art EPE of 0.44 on Scene Flow with Selective-IGEV, which surpasses \\(\\mathrm{LaC} + \\mathrm{GANet}\\) [18] by \\(38.89\\%\\). Besides, our Selective-RAFT also achieves a competitive EPE of 0.47 compared to IGEV-Stereo [34] with smaller parameters. To validate the ability to fuse information by regions of our modules, we then split Scene Flow test set into two regions: edge regions and non-edge regions using the Canny operator. As shown in Tab. 7, our Selective-RAFT outperforms RAFT-Stereo [17] by \\(25.23\\%\\) and \\(24.53\\%\\) in edge regions and non-edge regions. Due to IGEV-Stereo's aggregated cost volume [34], there's only a slight improvement in edge regions, but our Selective-IGEV still outperforms it by \\(7.32\\%\\) in non-edge regions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "KITTI. We finetune our model on the mixed dataset of KITTI 2012 and KITTI 2015 with a batch size of 8 for 50k" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.549 + ], + "angle": 0, + "content": "steps. Then we evaluate our Selective-Stereo on the test set of KITTI 2012 and KITTI 2015. As shown in Tab. 6, we achieve the best performance among all published methods for almost all metrics. On KITTI 2012, our Selective-RAFT outperforms RAFT-Stereo [17] by \\(14.58\\%\\) and \\(13.64\\%\\) on 2-noc and 2-all metrics, and our Selective-IGEV ranks \\(1^{st}\\) on these metrics. On KITTI 2015, our Selective-RAFT outperforms RAFT-Stereo [17] by \\(10.44\\%\\) on the D1-all metric, and our Selective-IGEV ranks \\(1^{st}\\) on all metrics with only 16 iterations same as IGEV-Stereo [34]. As shown in Fig. 5, our Selective-IGEV outperforms IGEV-Stereo [34] in detailed and textureless regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.732 + ], + "angle": 0, + "content": "ETH3D. Following CREStereo [16] and GMStereo [38], we use a collection of several public stereo datasets for training. The crop size is \\(384 \\times 512\\) and we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Sintel Stereo [3], InStereo2k [2] and ETH3D [24] datasets for 300k steps. Then we finetune it on the mixed CREStereo Dataset [16], InStereo2k [2] and ETH3D [24] datasets with for another 90k steps. As shown in Tab. 8, our Selective-RAFT outperforms RAFT-Stereo [17] by \\(17.90\\%\\) on Bad 0.5 metric, and our Selective-IGEV achieves the best performance among all published methods for almost all metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Middlebury. Also following CREStereo [16] and GMStereo [38], we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Falling Things [28], In-Stereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of \\(384 \\times 512\\) for 200k steps. Then we finetune it on the mixed CREStereo Dataset [16], Falling Things [28], InStereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of \\(384 \\times 768\\) with a batch size of 8 for another 100k steps. As shown in Tab. 8, our Selective-IGEV achieves the best performance among" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19707" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.852, + 0.274 + ], + "angle": 0, + "content": "
MethodETH3DMiddlebury
Bad 1.0Bad 0.5Bad 4.0AvgErrBad 2.0Bad 1.0Bad 4.0AvgErr
CroCo-Stereo [30]0.993.270.130.147.2916.94.181.76
GMStereo [38]1.835.940.080.197.1423.62.961.31
HITNet [26]2.797.830.190.206.4613.33.811.71
IGEV-Stereo [34]1.123.520.110.144.839.413.332.89
RAFT-Stereo [17]2.447.040.150.184.749.372.751.27
CREStereo [16]0.983.580.100.133.718.252.041.15
EAI-Stereo [43]2.315.210.700.213.687.812.141.09
DLNR [44]----3.206.821.891.06
Selective-RAFT (Ours)1.695.780.130.17----
Selective-IGEV (Ours)1.233.060.050.122.516.531.360.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.285, + 0.893, + 0.311 + ], + "angle": 0, + "content": "Table 8. Quantitative evaluation on ETH3D and Middlebury benchmarks. Note: Middlebury only allows one publish per paper, so we only publish our Selective-IGEV." + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.314, + 0.832, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.613, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on the test set of Middlebury. The third column is the visualization of attention maps generated by the CSA module. Our Selective-IGEV outperforms IGEV in large textureless and thin object regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.747 + ], + "angle": 0, + "content": "all published methods. As shown in Fig 7, compared to IGEV-Stereo [34], our Selective-IGEV performs better in textureless, and detailed regions. The third column in Fig 7 is the visualization of attention maps generated by the CSA module. It shows that attention maps can surely split regions that require information at different frequencies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.768, + 0.196, + 0.784 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We propose Selective-Stereo, a novel iterative stereo matching method. The proposed Contextual Spatial Attention module and Selective Recurrent Unit help the network capture information at different frequencies for edge and smooth regions. Our Selective-Stereo ranks \\(1^{st}\\) on KITTI, ETH3D, and Middlebury in almost all metrics among all published methods. It shows an ability to fuse information" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.686 + ], + "angle": 0, + "content": "at different frequencies adaptively for edge and smooth regions with the help of attention maps extracted by CSA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.893, + 0.837 + ], + "angle": 0, + "content": "However, our method still faces some challenges. Firstly, although our method can fuse information adaptively using attention maps, the SRU's receptive field is still limited by predefined values. Secondly, adding branches or increasing the sizes of convolutional kernels leads to high memory and time costs, so we will explore the combination of lightweight convolutions and our method to reduce memory costs. Lastly, it's also a good direction to do research on the combination of convolutions and self-attention due to their different advantages and receptive fields." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.893, + 0.898 + ], + "angle": 0, + "content": "Acknowledgement. This research is supported by National Natural Science Foundation of China (62122029, 62061160490, U20B200007)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19708" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 4 (11):e21, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[4] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[5] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Shuicheng Yan, and Jiashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3435-3444, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Yajie Chen, Xin Yang, and Xiang Bai. Confidence-weighted mutual supervision on dual networks for unsupervised cross-modality image segmentation. Science China Information Sciences, 66(11):210104, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Junda Cheng, Xin Yang, Yuechuan Pu, and Peng Guo. Region separable stereo matching. IEEE Transactions on Multimedia, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.472, + 0.592 + ], + "angle": 0, + "content": "[8] Junda Cheng, Gangwei Xu, Peng Guo, and Xin Yang. Coatsnet: Fully exploiting convolution and attention for stereo matching by region separation. International Journal of Computer Vision, pages 1-18, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33:22158-22169, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[11] Miaojie Feng, Junda Cheng, Hao Jia, Longliang Liu, Gangwei Xu, and Xin Yang. Mc-stereo: Multi-peak lookup and cascade search range for stereo matching. arXiv preprint arXiv:2311.02340, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3273-3282, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[15] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision, pages 66-75, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[16] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[17] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[18] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1647-1655, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[20] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[21] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[22] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Fast vision transformers with hilo attention. Advances in Neural Information Processing Systems, 35:14541-14554, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[23] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[24] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[25] Zhelun Shen, Yuchao Dai, Xibin Song, Zhibo Rao, Dingfu Zhou, and Liangjun Zhang. Pcw-net: Pyramid combination and warping cost volume for stereo matching. In European Conference on Computer Vision, pages 280–297. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[28] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[29] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[30] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[32] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[33] Gangwei Xu, Shujun Chen, Hao Jia, Miaojie Feng, and Xin Yang. Memory-efficient optical flow via radius-distribution orthogonal cost volume. arXiv preprint arXiv:2312.03790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[34] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[36] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1959-1968, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[39] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1740-1749, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[40] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5515-5524, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[41] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12926-12934, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[43] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Yong Zhao, Yitong Yang, and Ting Ouyang. Eai-stereo: Error aware iterative network for stereo matching. In Proceedings of the Asian Conference on Computer Vision, pages 315-332, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[44] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Jie Chen, Yitong Yang, and Yong Zhao. High-frequency stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1327-1336, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19710" + } + ] +] \ No newline at end of file diff --git a/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_origin.pdf b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f3911ea60fbfb16771243fbc72e76f99dfbe06e2 --- /dev/null +++ b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/584de1de-5f52-4125-8106-326f8d8f0441_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c94236a7f27f61bab11be032870ed48930382893e7ca8728bd6f51b0745f225 +size 7260964 diff --git a/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/full.md b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..635ebe6b858b31fbad0bf8221fb807c18e14c41a --- /dev/null +++ b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/full.md @@ -0,0 +1,341 @@ +# Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching + +Xianqi Wang*, Gangwei Xu*, Hao Jia, Xin Yang† + +Huazhong University of Science and Technology + +{xianqiw, gwxu, haojia, xinyang2014}@hust.edu.cn + +# Abstract + +Stereo matching methods based on iterative optimization, like RAFT-Stereo and IGEV-Stereo, have evolved into a cornerstone in the field of stereo matching. However, these methods struggle to simultaneously capture high-frequency information in edges and low-frequency information in smooth regions due to the fixed receptive field. As a result, they tend to lose details, blur edges, and produce false matches in textureless areas. In this paper, we propose Selective Recurrent Unit (SRU), a novel iterative update operator for stereo matching. The SRU module can adaptively fuse hidden disparity information at multiple frequencies for edge and smooth regions. To perform adaptive fusion, we introduce a new Contextual Spatial Attention (CSA) module to generate attention maps as fusion weights. The SRU empowers the network to aggregate hidden disparity information across multiple frequencies, mitigating the risk of vital hidden disparity information loss during iterative processes. To verify SRU's universality, we apply it to representative iterative stereo matching methods, collectively referred to as Selective-Stereo. Our Selective-Stereo ranks $1^{st}$ on KITTI 2012, KITTI 2015, ETH3D, and Middlebury leaderboards among all published methods. Code is available at https://github.com/Windsrain/Selective-Stereo. + +# 1. Introduction + +Stereo matching is a fundamental area of research in computer vision. It explores the calculation of displacement, referred to as disparity, between matching points in a pair of rectified images. This technique plays a significant role in various applications, including 3D reconstruction and autonomous driving. + +With the advancement of deep learning, learning-based stereo matching [4, 14, 15, 17, 34, 35] have progressively displaced traditional methods and significantly enhancing + +![](images/002f571a9e974f670eefa979e7102f82b607bc7516ea8181d557c1da34de0207.jpg) + +![](images/17257d6328d2e82198ac8b669786cf3ae5f1b0c281975362c037f1fa96310447.jpg) + +![](images/35c241f98aded3ac51826d196765ea35f3e1df09c7547a1e98fbe1866111d9d4.jpg) +Left Image + +![](images/86586d706209ab26d3f7c432827e7e78a0b4da09656fc674f6fb6062ea4a65e9.jpg) +RAFT-Stereo [17] + +![](images/ea01c757381138d4a041dccfea298fefea82731edfbf36cb4429944d9f4fea04.jpg) +Selective-RAFT + +![](images/df2a9b545d1e25c01d556e5c69a4ee46d374a5102b5cf1a77c512bd5fdece4a1.jpg) +Left Image + +![](images/654fc3d6c856316bfefaef6812f4570fc5eb8686ca537e97cfd4940ed66cf325.jpg) +IGEV-Stereo [34] + +![](images/8b1081ab24d0706573ac72907b023494f8ea2c431be679bcd90c9167b82806ca.jpg) +Selective-IGEV +Figure 1. Row 1: Comparisons with state-of-the-art stereo methods on KITTI 2012 [13] and KITTI 2015 [21], ETH3D [24] and Middlebury [23] leaderboards. Row 2: Visual comparison with RAFT-Stereo on ETH3D. Row 3: Visual comparison with IGEV-Stereo on Middlebury. Our method distinguishes subtle details and sharp edges and performs well in weak texture regions. + +the accuracy of disparity estimation. Initially, aggregation-based methods [7, 32, 41] led the development of stereo matching algorithms. These methods begin by defining a maximum range of disparity, constructing a 4D cost volume using feature maps, and subsequently employing 3D CNN to filter the volume and derive the final disparity map. Such methods focus on filtering the initially coarse cost volume, thus effectively aggregating geometry information. However, cost aggregation requires a large number of convolutions, resulting in high computational costs, making it difficult to be applied to high-resolution images. + +Recently, a novel class of methods based on iterative optimization [11, 16, 17, 44] has been gaining prominence and + +achieving state-of-the-art performance on several leaderboards. These methods begin by constructing an all-pairs cost volume, indexing a local cost volume from the original cost volume, and subsequently employing recurrent units [10] to calculate disparity residuals and update the disparity prediction. One major advantage of these methods is their ability to capture all candidate matching points without predefining the range of disparities. Additionally, these methods don't need to aggregate the cost volume using a large number of redundant convolutions. Instead, a continuous update of the disparity prediction is achieved through lightweight recurrent units during iterations. Therefore, these methods are capable of processing high-resolution images. + +However, iterative methods encounter several challenges. Firstly, the all-pairs cost volume includes considerable noisy information [34], potentially causing the loss of crucial information when iterating the hidden information. Besides, as the network iterates, the hidden information increasingly incorporates global low-frequency information while losing local high-frequency information like edges and thin objects [44]. Secondly, the existing recurrent units possess a fixed receptive field, leading the network to solely concentrate on information at the current frequency and ignore other frequencies, such as detailed, edge, and textureless information. + +In this paper, we propose Selective Recurrent Unit (SRU) to address the limitations of traditional recurrent units. As Chen et al. [5] mentions features contain information at different frequencies, high-frequency information describes rapidly changing fine details, while low-frequency information describes smoothly changing structures. Unlike traditional recurrent units that treat information at different frequencies equally, our SRU incorporates multiple branches of GRU, each with a distinct kernel size representing different receptive fields. The hidden information obtained from each GRU branch is fused and then fed into the next iteration. This fusion enables the capture of information from different receptive fields at different frequencies, while also performing secondary filtering to reduce noise information from local cost volume. To further enhance the fusion process, we propose a Contextual Spatial Attention (CSA) module to utilize the context information. Instead of simply summarizing information from different branches, CSA introduces attention maps extracted from the context information. After doing so, information captured by small kernels has large weights in regions like edge, while information captured by large kernels has large weights in regions like low-texture. These attention maps determine the weight of fusion, allowing the network to adaptively select suitable information based on different image regions. Besides, we prove the effectiveness and universality of our module by transferring it to different iterative networks. All networks + +are collectively referred to as Selective-Stereo. By doing so, we consistently improve the performance of these networks without introducing a significant increase in parameters and time. + +We demonstrate the effectiveness of our method on several stereo benchmarks. On Scene Flow [20], our Selective-RAFT reaches the state-of-the-art EPE of 0.47, and our Selective-IGEV even achieves a new state-of-the-art EPE of 0.44. And as shown in Fig. 1, our Selective-RAFT surpasses RAFT-Stereo by a large margin and achieves competitive performance compared with the state-of-the-art methods on KITTI [13, 21] leaderboards. Our Selective-IGEV ranks $1^{st}$ on KITTI, ETH3D [24], and Middlebury [23] leaderboards among all published methods. + +Our main contributions can be summarized as follows: + +- We propose a novel iterative update operator SRU for iterative stereo matching methods. +- We introduce a new Contextual Spatial Attention module that generates attention maps for adaptively fusing hidden disparity information at multiple frequencies. +- We verify the universality of our SRU on several iterative stereo matching methods. +- Our method outperforms existing published methods on public leaderboards such as KITTI, ETH3D, and Middlebury. + +# 2. Related Work + +Aggregation-based methods in stereo matching. Several aggregation-based methods [4, 7, 8, 14, 15, 25, 35-37, 41] have shown significant progress in the domain of stereo matching in recent years. DispNet [20] establishes the groundwork for subsequent network architecture. GCNet [15] proposes a 4D concatenate cost volume, which is subsequently regularized using 3D CNNs. Additionally, it also introduces the soft argmin function for disparity regression, resulting in a significant influence on subsequent methods. PSMNet [4] proposes a stacked hourglass 3D CNN, which improves the cost aggregation stage to enhance the network's ability to capture context information. GwcNet [14] proposes Group-wise Correlation Volume, which combines the advantages of correlation and concatenation volume. GA-Net [41] designs a semi-global guided aggregation layer and a local guided aggregation layer, inspired by the traditional semi-global matching algorithm, to further assist in aggregating global and geometry information in the network. Building upon Group-wise Correlation Volume, ACVNet [32] proposes Attention Concat Volume, which uses attention weights to suppress redundant information and maintain sufficient information for matching. + +Iterative-based methods in stereo matching. In recent years, many iterative methods [17, 27, 33, 34], spearheaded by RAFT [17], have gradually become the mainstream of research. RAFT-Stereo [17] builds upon the optical flow + +![](images/d33e626a3ec484bfdcd1d10a05b2a81f673cd9317a19877d45028bf49006b2de.jpg) +Figure 2. Overview of our proposed Selective-Stereo (Selective-RAFT version). The Contextual Spatial Attention (CSA) module extracts attention maps from context information as a guide for Selective Recurrent Units (SRUs). Then the network iteratively updates the disparity using local cost volumes retrieved from the correlation pyramid and attention maps given by CSA through SRUs. + +method RAFT [27] by introducing an all-pairs cost volume pyramid that maintains high resolution. It extracts local correlation features from this pyramid, performs iterative disparity updates using GRU-based update operators, and incorporates a multi-level GRU to expand the receptive field. On this basis, IGEV-Stereo [34] asserts that the initial cost volume is excessively coarse. To alleviate the need for iterations and reduce time overhead, it proposes to use a lightweight cost aggregation network before iterations. CREstereo [16] designs a hierarchical network in a coarse-to-fine manner, as well as a stacked cascaded architecture for inference in place of the original single-resolution iterative structure. DLNR [44] proposes the use of LSTM as a replacement for GRU, providing the advantage of decoupling the update of hidden states from disparity prediction. + +Frequency information application in vision. There are several works[5, 6, 39] that focus on using frequency information in computer vision. Chen et al. [5] propose the octave convolution to factorize the mixed feature maps by their frequencies. Xu et al. [39] propose a method of learning in the frequency domain and suggest that CNN models are more sensitive to low-frequency channels than high-frequency. DSGAN [12] introduces the frequency separation into super-resolution. LITv2 [22] proposes to disentangle the high/low-frequency patterns in an attention layer. + +# 3. Method + +In this section, we present the overall architecture of Selective-Stereo. Because our method can be plugged into different networks, we take Selective-RAFT (Fig. 2) as an example and focus on illustrating its key components. + +# 3.1. Feature Extraction + +To ensure fair comparisons, Selective-RAFT maintains consistency with RAFT-Stereo [17] by employing its feature extraction network. Feature extraction comprises two main components: feature network and context network. + +Feature Network. Given the left and the right images $\mathbf{I}_{l(r)}\in \mathbb{R}^{3\times H\times W}$ , we first downsample them to $1 / 2$ resolution using a $7\times 7$ convolutional layer. Then, a series of residual blocks is employed to extract features and we apply another downsampling layer to get features at $1 / 4$ resolution in the middle. Lastly, a $1\times 1$ convolutional layer is applied to get the final left and right features $\mathbf{f}$ , $\mathbf{g}\in \mathbb{R}^{C\times \frac{H}{4}\times \frac{W}{4}}$ with suitable dimensions. + +Context Network. Its architecture remains consistent with the feature network, and it adds a series of residual blocks and two additional downsampling layers, obtaining multi-level context features $\mathbf{f}_i^c$ ( $i = 1, 2, 3$ ) at $1/4, 1/8, 1/16$ resolutions. Then we can get the initial hidden and the con + +text information: + +$$ +\mathbf {h} _ {i} = \tanh \left(\mathbf {f} _ {i} ^ {c}\right) \tag {1} +$$ + +$$ +\mathbf {c} _ {i} = \operatorname {R e L U} \left(\mathbf {f} _ {i} ^ {c}\right) +$$ + +# 3.2. Cost Volume Construction + +Given the left and the right features $\mathbf{f},\mathbf{g}$ , we first construct an all-pairs correlate cost volume: + +$$ +\mathbf {C} _ {i j k} = \sum_ {h} \mathbf {f} _ {h i j} \cdot \mathbf {g} _ {h i k}, \mathbf {C} \in \mathbb {R} ^ {\frac {H}{4} \times \frac {W}{4} \times \frac {W}{4}} \tag {2} +$$ + +Then we construct a 4-level correlation pyramid $\{\mathbf{C}_i\}$ ( $i = 1,2,3,4$ ) by using 1D average pooling with a kernel size of 2 and a stride of 2 at the last dimension. + +# 3.3. Contextual Spatial Attention Module + +To help information from different receptive fields and frequencies fuse, the Contextual Spatial Attention (CSA) module extracts multi-level attention maps from context information as guidance. As illustrated in Fig. 3, CSA can be divided into two submodules: Channel Attention Enhancement (CAE) and Spatial Attention Extractor (SAE). These submodules are derived from CBAM [31] and we simplify them to better adapt to stereo matching. + +Channel Attention Enhancement. Given a context information map $\mathbf{c} \in \mathbb{R}^{C \times H \times W}$ , we first use an average-pooling and a max-pooling operation on the spatial dimension to get two maps $\mathbf{f}_{avg}, \mathbf{f}_{max} \in \mathbb{R}^{C \times 1 \times 1}$ . Then we use two convolutional layers to perform feature transformation on these maps separately. After that, we add these two maps together and use the sigmoid function to convert them into weights $\mathbf{M}_c \in \mathbb{R}^{C \times 1 \times 1}$ between 0 and 1. Lastly, using an element-wise product, the initial map can capture which channel map has high feature values to be enhanced, and which channel map has low feature values to be suppressed. + +Spatial Attention Extractor. After the CAE module, we continue to use the same pooling operations, but now we pool on the channel dimension. Then we concatenate these pooling maps to form a map in $\mathbb{R}^{2\times H\times W}$ and use one convolutional layer with a sigmoid function to generate the final attention map. Reviewing previous operations, this attention map has high weights in regions needing high-frequency information because this information possesses high feature values in the context information. Similarly, it has low weights in regions needing low-frequency information. In general, the attention map can explicitly distinguish regions that need information at different frequencies. + +# 3.4. Selective Recurrent Unit + +To capture information at different frequencies, Selective Recurrent Unit (SRU) uses attention maps extracted by CSA to fuse hidden information derived from GRUs with different kernel sizes. + +Multi-level update structure. As illustrated in Fig. 4, SRUs at $1/8$ , $1/16$ resolutions take the attention map, context information, hidden information at the same resolution, and the hidden information at adjacent resolutions as inputs. At $1/4$ resolution, SRUs take disparity, and local cost volume as additional inputs, and then their outputs will go through two convolutional layers to generate disparity residuals. The local cost volume is derived from the all-pairs correlation pyramid in the same way as RAFT-Stereo [17]. At last, disparities at $1/4$ resolution will be upsampled into full resolution using the convex combination. + +SRU's architecture. A single GRU can be defined as follows: + +$$ +z _ {k} = \sigma (\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {z})), +$$ + +$$ +\begin{array}{l} r _ {k} = \sigma (\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {r})), \\ \tilde {I} _ {\text {一}} = \operatorname {c o v} (G _ {m} ([ \circ_ {0} ]), W _ {r})) \end{array} \tag {3} +$$ + +$$ +\tilde {h} _ {k} = \tanh (\operatorname {C o n v} ([ r _ {k} \odot h _ {k - 1}, x _ {k} ], W _ {h})), \tag {3} +$$ + +$$ +h _ {k} = \left(1 - z _ {k}\right) \odot h _ {k - 1} + z _ {k} \odot \tilde {h} _ {k} +$$ + +where $x_{k}$ is the concatenation of disparity, correlation, hidden information, and context information previously defined. Unlike RAFT-Stereo [17] that divide the context information into $c_{z}, c_{r}, c_{h}$ , we add it into $x_{k}$ because using convolutions with different kernel sizes can fully utilize context information. + +As illustrated in Fig. 3, a single SRU can be defined as follows: + +$$ +h _ {k} = \mathbf {A} \odot h _ {k} ^ {s} + (1 - \mathbf {A}) \odot h _ {k} ^ {l} \tag {4} +$$ + +where $\mathbf{A}$ denotes the attention map derived from CSA at the same resolution, $h_k^s$ denotes the GRU with smaller kernel sizes and $h_k^l$ denotes the larger one. + +As Sec. 3.3 mentioned, the attention map has high weights in regions needing high-frequency information. Therefore, the GRU with smaller kernel sizes that can capture high-frequency information like edge, and thin objects should do element-wise products with the attention map directly, and the GRU with larger kernel sizes should do element-wise products with the contrary attention map. + +Receptive fields analysis. The repetitive fields computing formula [1] can be defined as follows: + +$$ +r _ {0} = \sum_ {l = 1} ^ {L} \left(\left(k _ {l} - 1\right) \prod_ {i = 1} ^ {l - 1} s _ {i}\right) + 1 \tag {5} +$$ + +where $k_{l}$ denotes the kernel size, $s_{i}$ denotes the stride size, and $r_0$ denotes the whole network. + +Given a multi-level structure like Fig. 4, if we take the $1/4$ resolution as the basis, and the downsampling operations can be regarded as a convolution with kernel size 3, stride size 2, this structure's receptive fields are $k$ , $2k + 3$ , $3k + 6$ . That means it only has 3 fixed receptive fields in total. + +![](images/64eb6c020445a9fe0255b3ca8b579a8dc09c2bd8e985e1439137001b300ae8db.jpg) +Figure 3. The architecture of proposed modules. Left: Contextual Spatial Attention (CSA) module. Right: Selective Recurrent Unit (SRU). + +![](images/174b196e660c161d4133595da944539a3c1c14aafe2e1b8dfe811c7b1b1c1d47.jpg) + +![](images/f616a01686288c415c2340eb03c56fd9b28f1b0d5cf4fa3948f9d77b21e3e579.jpg) +Figure 4. Multi-level SRU. Information is passed between SRUs at adjacent resolutions. Dashed arrows represent upsampling and downsampling operations. At $1/4$ resolution, disparity and local cost volume will be additional information put into SRUs. + +If we replace GRUs with our SRUs with a small kernel size $s$ and a large kernel size $l$ , the multi-level structure will have 6 receptive fields initially. Besides, pixels in hidden information are affected by different receptive fields during fusion, and the fusion is influenced by attention maps adaptively. In general, the multi-level SRU holds dynamic receptive fields, and it enables itself to capture information at different frequencies. + +# 3.5. Loss Function + +We supervise our network on the L1 distance between all predicted disparities $\{\mathbf{d}_i\}_{i=1}^N$ and the ground truth disparity $\mathbf{d}_{gt}$ with increasing weights. The total loss is defined as: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {N} \gamma^ {N - i} | | \mathbf {d} _ {i} - \mathbf {d} _ {g t} | | _ {1} \tag {6} +$$ + +where $\gamma = 0.9$ , and $N$ is the number of iterations. + +# 4. Experiments + +Scene Flow [20] is a synthetic dataset including 35,454 training pairs and 4,370 testing pairs with dense disparity maps. For training and testing, we use the finalpass version, because it contains more realistic and difficult effects than the cleanpass version. KITTI 2012 [13] and KITTI 2015 [21] are datasets for real-world driving scenes. KITTI 2012 contains 194 training pairs and 195 testing pairs, and + +KITTI 2015 contains 200 training pairs and 200 testing pairs. ETH3D [24] is a collection of gray-scale stereo pairs containing 27 training pairs and 20 testing pairs for indoor and outdoor scenes. Middlebury [23] is a high-resolution dataset containing 15 training pairs and 15 testing pairs for indoor scenes. + +# 4.1. Implementation Details + +We implement our Selective-Stereo with PyTorch and the model is trained on NVIDIA RTX 3090 GPUs. For all experiments, we use the AdamW [19] optimizer and clip gradients to the range [-1, 1]. We use the one-cycle learning rate schedule with a learning rate of 2e-4. We first train our model on Scene Flow with a batch size of 8 for 200k steps as the pretrained model. The crop size is $320 \times 720$ , and we use 22 update iterations during training. + +# 4.2. Ablation Study + +In this section, we evaluate our model in different settings to verify our proposed modules in several aspects. All results use 32 update iterations. + +Effectiveness of proposed modules. To verify the effectiveness of our proposed modules, we take RAFT-Stereo [17] as the baseline and replace its GRUs with our SRUs. As shown in Tab. 1, the proposed SRU can improve the accuracy even without CSA. It means that if we just sum up the information from different branches, the growth of receptive fields can be beneficial for inference. If we add our CSA but invert the weights of the attention maps, the effect even decreases. That validates that our CSA's attention maps do indeed reflect the weights of information at different frequencies in regions. Therefore, if we add CSA normally, the full model (Selective-RAFT) can achieve the best performance with only a $4\%$ increase in parameters. + +Universality of proposed modules. To verify the universality of our proposed modules, we take three typical iterative stereo matching methods as the baseline and replace their GRUs with SRUs. Especially, in DLNR [44], the recurrent units are LSTMs but not GRUs, so we just replace GRUs inside SRUs with LSTMs to make a fair comparison. + +
ModelGRUSRUCSA (Contrary)CSAEPE (px)>1px (%)Param (M)
Baseline (RAFT-Stereo)✓0.536.0811.12
SRU✓0.505.3811.65
SRU+CSA (Contrary)✓✓0.505.5811.65
Full model (Selective-RAFT)✓✓0.475.3211.65
+ +![](images/24e5c71fca52740b8f87028bb5ceef85a4a2a280161b03979dcaab001a1af1ef.jpg) +Left Image + +![](images/38290070a432c24a353b6772cfc56e730c36a518e438e39b5a804646eeb69b56.jpg) +IGEV-Stereo + +![](images/08e2403e1370b9fb3d53a974377c462a39d86b1bb6a71ce008517a6ca7060f26.jpg) +Selective-IGEV + +Table 1. Ablation study of the effectiveness of proposed modules on the Scene Flow test set. SRU denotes Selective Recurrent Unit, and CSA denotes Contextual Spatial Attention. Contrary means we invert the weights of the attention maps. The baseline is RAFT-Stereo. + +
ModelEPE (px)>1px (%)Param (M)
RAFT-Stereo [17]0.536.0811.12
Selective-RAFT0.475.3211.65
IGEV-Stereo [34]0.475.2112.60
Selective-IGEV0.444.9813.14
DLNR [44]0.495.0657.37
Selective-DLNR0.464.7358.09
+ +Table 2. Ablation study of the universality of proposed modules. + +
ModelNumber of Iterations
1234832
RAFT-Stereo [17]2.081.130.870.750.580.53
Selective-RAFT1.951.060.810.690.530.47
IGEV-Stereo [34]0.660.620.580.550.500.47
Selective-IGEV0.650.600.560.530.480.44
+ +Table 3. Ablation study of the number of iterations. + +
Kernel SizesEPE (px)>1px (%)
1 × 1 + 1 × 50.485.41
3 × 3 + 1 × 50.485.30
1 × 1 + 3 × 30.475.32
+ +Table 4. Ablation study of the size of convolutional kernels. + +As shown in Tab. 2, all methods have a significant improvement in the EPE metrics on Scene Flow, and the insertion of modules only results in a slight increase in parameters. Besides, as shown in Fig. 6, the CSA module generates different attention maps in different networks. In Selective-RAFT, because the cost volume contains a large amount of noisy information, the network needs more large kernels to + +![](images/43de67bc4c680a303a7cea9c03bc11cd9d0e4e30f6d6850039e89935c548382b.jpg) +Figure 5. Qualitative results on the test set of KITTI. Our Selective-IGEV outperforms IGEV in detailed and weak texture regions. +Left Image +Figure 6. Visualization of the attention map of different networks. + +![](images/86638174669e0c84281a09b9c1658f3aebe77890f5641e689ec5e590eef68519.jpg) +Selective-RAFT + +![](images/510077035b70fddc24befb840a93f9f329a30fbf3447277e7b2bd6918fcdcce1.jpg) +Selective-IGEV + +filter the local cost volume. On the contrary, the cost volume has already been aggregated in Selective-IGEV, so the network tends to maintain high-frequency information using small kernels. Moreover, the cost volume in Selective-IGEV faces an over-smooth problem [34], and that's why the attention map tends to increase the weights of large kernels to recover edge regions. In general, the CSA module shows different tendencies in different networks, which is a reflection of its adaptive ability. + +Number of iterations. Our Selective-Stereo can achieve better performance with a smaller number of iterations. As shown in Tab. 3, our Selective-RAFT get the same performance with only 8 iterations compared to RAFT-Stereo [17], and for IGEV-Stereo [34], our Selective-IGEV also get a slight improvement with a few iterations. It shows that our modules can make secondary filtering to reduce noisy information from the initial cost volume. + +Size of convolution kernels. We verify different kernel sizes on Scene Flow as shown in Tab. 4. At last, we choose the combination of $1 \times 1$ and $3 \times 3$ as our default config- + +
MethodCSPNLEASTereoLaC + GANetACVNetIGEV-StereoSelective-RAFT (Ours)Selective-IGEV (Ours)
EPE (px)0.780.780.720.480.470.470.44
+ +Table 5. Quantitative evaluation on Scene Flow test set. + +
MethodKITTI 2012KITTI 2015Run-time (s)
2-noc2-all3-noc3-allEPE-nocEPE-allD1-bgD1-fgD1-all
AcfNet [42]1.832.351.171.540.50.51.513.801.890.48
LEAStereo [9]1.902.391.131.450.50.51.402.911.650.30
ACVNet [32]1.832.351.131.470.40.51.373.071.650.20
RAFT-Stereo [17]1.922.421.301.660.40.51.583.051.820.38
PCWNet [25]1.692.181.041.370.40.51.583.051.820.38
LaC + GANet [18]1.722.261.051.420.40.51.442.831.671.80
CREStereo [16]1.722.181.141.460.40.51.452.861.690.41
IGEV-Stereo [34]1.712.171.121.440.40.41.382.671.590.18
Selective-RAFT (Ours)1.642.091.101.430.40.51.412.711.630.45
Selective-IGEV (Ours)1.592.051.071.380.40.41.332.611.550.24
+ +Table 6. Quantitative evaluation on KITTI 2012 and KITTI 2015. + +
MethodEdgesNon-Edges
EPE>1pxEPE>1px
RAFT-Stereo [17]3.2129.160.536.53
Selective-RAFT2.4021.630.404.65
IGEV-Stereo [34]2.2320.420.414.58
Selective-IGEV2.1820.010.384.35
+ +Table 7. Quantitative evaluation on Scene Flow test set in different regions. + +uration, because it achieves a competitive performance and reduces computational costs. + +# 4.3. Comparisons with State-of-the-art + +All fine-tuned models use the model pretrained on Scene Flow. Different target datasets use different finetune strategies. We validate two models called Selective-RAFT and Selective-IGEV using RAFT-Stereo [17] and IGEV-Stereo [34] as the baseline respectively. + +Scene Flow. As shown in Tab. 5, we achieve a new state-of-the-art EPE of 0.44 on Scene Flow with Selective-IGEV, which surpasses $\mathrm{LaC} + \mathrm{GANet}$ [18] by $38.89\%$ . Besides, our Selective-RAFT also achieves a competitive EPE of 0.47 compared to IGEV-Stereo [34] with smaller parameters. To validate the ability to fuse information by regions of our modules, we then split Scene Flow test set into two regions: edge regions and non-edge regions using the Canny operator. As shown in Tab. 7, our Selective-RAFT outperforms RAFT-Stereo [17] by $25.23\%$ and $24.53\%$ in edge regions and non-edge regions. Due to IGEV-Stereo's aggregated cost volume [34], there's only a slight improvement in edge regions, but our Selective-IGEV still outperforms it by $7.32\%$ in non-edge regions. + +KITTI. We finetune our model on the mixed dataset of KITTI 2012 and KITTI 2015 with a batch size of 8 for 50k + +steps. Then we evaluate our Selective-Stereo on the test set of KITTI 2012 and KITTI 2015. As shown in Tab. 6, we achieve the best performance among all published methods for almost all metrics. On KITTI 2012, our Selective-RAFT outperforms RAFT-Stereo [17] by $14.58\%$ and $13.64\%$ on 2-noc and 2-all metrics, and our Selective-IGEV ranks $1^{st}$ on these metrics. On KITTI 2015, our Selective-RAFT outperforms RAFT-Stereo [17] by $10.44\%$ on the D1-all metric, and our Selective-IGEV ranks $1^{st}$ on all metrics with only 16 iterations same as IGEV-Stereo [34]. As shown in Fig. 5, our Selective-IGEV outperforms IGEV-Stereo [34] in detailed and textureless regions. + +ETH3D. Following CREStereo [16] and GMStereo [38], we use a collection of several public stereo datasets for training. The crop size is $384 \times 512$ and we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Sintel Stereo [3], InStereo2k [2] and ETH3D [24] datasets for 300k steps. Then we finetune it on the mixed CREStereo Dataset [16], InStereo2k [2] and ETH3D [24] datasets with for another 90k steps. As shown in Tab. 8, our Selective-RAFT outperforms RAFT-Stereo [17] by $17.90\%$ on Bad 0.5 metric, and our Selective-IGEV achieves the best performance among all published methods for almost all metrics. + +Middlebury. Also following CREStereo [16] and GMStereo [38], we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Falling Things [28], In-Stereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of $384 \times 512$ for 200k steps. Then we finetune it on the mixed CREStereo Dataset [16], Falling Things [28], InStereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of $384 \times 768$ with a batch size of 8 for another 100k steps. As shown in Tab. 8, our Selective-IGEV achieves the best performance among + +
MethodETH3DMiddlebury
Bad 1.0Bad 0.5Bad 4.0AvgErrBad 2.0Bad 1.0Bad 4.0AvgErr
CroCo-Stereo [30]0.993.270.130.147.2916.94.181.76
GMStereo [38]1.835.940.080.197.1423.62.961.31
HITNet [26]2.797.830.190.206.4613.33.811.71
IGEV-Stereo [34]1.123.520.110.144.839.413.332.89
RAFT-Stereo [17]2.447.040.150.184.749.372.751.27
CREStereo [16]0.983.580.100.133.718.252.041.15
EAI-Stereo [43]2.315.210.700.213.687.812.141.09
DLNR [44]----3.206.821.891.06
Selective-RAFT (Ours)1.695.780.130.17----
Selective-IGEV (Ours)1.233.060.050.122.516.531.360.91
+ +Table 8. Quantitative evaluation on ETH3D and Middlebury benchmarks. Note: Middlebury only allows one publish per paper, so we only publish our Selective-IGEV. + +![](images/696c137fe7da6974ba62c2d51b1762e3f0a790406419b4f71189d176ab67387e.jpg) +Figure 7. Qualitative results on the test set of Middlebury. The third column is the visualization of attention maps generated by the CSA module. Our Selective-IGEV outperforms IGEV in large textureless and thin object regions. + +all published methods. As shown in Fig 7, compared to IGEV-Stereo [34], our Selective-IGEV performs better in textureless, and detailed regions. The third column in Fig 7 is the visualization of attention maps generated by the CSA module. It shows that attention maps can surely split regions that require information at different frequencies. + +# 5. Conclusion + +We propose Selective-Stereo, a novel iterative stereo matching method. The proposed Contextual Spatial Attention module and Selective Recurrent Unit help the network capture information at different frequencies for edge and smooth regions. Our Selective-Stereo ranks $1^{st}$ on KITTI, ETH3D, and Middlebury in almost all metrics among all published methods. It shows an ability to fuse information + +at different frequencies adaptively for edge and smooth regions with the help of attention maps extracted by CSA. + +However, our method still faces some challenges. Firstly, although our method can fuse information adaptively using attention maps, the SRU's receptive field is still limited by predefined values. Secondly, adding branches or increasing the sizes of convolutional kernels leads to high memory and time costs, so we will explore the combination of lightweight convolutions and our method to reduce memory costs. Lastly, it's also a good direction to do research on the combination of convolutions and self-attention due to their different advantages and receptive fields. + +Acknowledgement. This research is supported by National Natural Science Foundation of China (62122029, 62061160490, U20B200007). + +# References + +[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 4 (11):e21, 2019. 4 +[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 7 +[3] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7 +[4] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1, 2 +[5] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Shuicheng Yan, and Jiashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3435-3444, 2019. 2, 3 +[6] Yajie Chen, Xin Yang, and Xiang Bai. Confidence-weighted mutual supervision on dual networks for unsupervised cross-modality image segmentation. Science China Information Sciences, 66(11):210104, 2023. 3 +[7] Junda Cheng, Xin Yang, Yuechuan Pu, and Peng Guo. Region separable stereo matching. IEEE Transactions on Multimedia, 2022. 1, 2 +[8] Junda Cheng, Gangwei Xu, Peng Guo, and Xin Yang. Coatsnet: Fully exploiting convolution and attention for stereo matching by region separation. International Journal of Computer Vision, pages 1-18, 2023. 2 +[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33:22158-22169, 2020. 7 +[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 2 +[11] Miaojie Feng, Junda Cheng, Hao Jia, Longliang Liu, Gangwei Xu, and Xin Yang. Mc-stereo: Multi-peak lookup and cascade search range for stereo matching. arXiv preprint arXiv:2311.02340, 2023. 1 +[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 3 +[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2, 5 + +[14] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3273-3282, 2019. 1, 2 +[15] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision, pages 66-75, 2017. 1, 2 +[16] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1, 3, 7, 8 +[17] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[18] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1647-1655, 2022. 7 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[20] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 2, 5, 7 +[21] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 5 +[22] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Fast vision transformers with hilo attention. Advances in Neural Information Processing Systems, 35:14541-14554, 2022. 3 +[23] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 1, 2, 5, 7 +[24] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 1, 2, 5, 7 +[25] Zhelun Shen, Yuchao Dai, Xibin Song, Zhibo Rao, Dingfu Zhou, and Liangjun Zhang. Pcw-net: Pyramid combination and warping cost volume for stereo matching. In European Conference on Computer Vision, pages 280–297. Springer, 2022. 2, 7 +[26] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo + +matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 8 +[27] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 2, 3 +[28] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 7 +[29] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 7 +[30] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 8 +[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 4 +[32] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 1, 2, 7 +[33] Gangwei Xu, Shujun Chen, Hao Jia, Miaojie Feng, and Xin Yang. Memory-efficient optical flow via radius-distribution orthogonal cost volume. arXiv preprint arXiv:2312.03790, 2023. 2 +[34] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 3, 6, 7, 8 +[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2 +[36] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023. +[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1959-1968, 2020. 2 +[38] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, + +stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7, 8 +[39] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1740-1749, 2020. 3 +[40] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5515-5524, 2019. 7 +[41] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 1, 2 +[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12926-12934, 2020. 7 +[43] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Yong Zhao, Yitong Yang, and Ting Ouyang. Eai-stereo: Error aware iterative network for stereo matching. In Proceedings of the Asian Conference on Computer Vision, pages 315-332, 2022. 8 +[44] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Jie Chen, Yitong Yang, and Yong Zhao. High-frequency stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1327-1336, 2023. 1, 2, 3, 5, 6, 8 \ No newline at end of file diff --git a/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/images.zip b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa17312c576f6f5b0f44db4a50298578dc6b2e3a --- /dev/null +++ b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3191e8408a50560675d2775479e6c1a3f5f2d961a535064cbc47ea7ff34e8e6 +size 811129 diff --git a/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/layout.json b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0561a80afb172a4f25895a407cec2b1fb2e252b8 --- /dev/null +++ b/2024/Selective-Stereo_ Adaptive Frequency Information Selection for Stereo Matching/layout.json @@ -0,0 +1,8599 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "text", + "content": "Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 142, + 423, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 142, + 423, + 159 + ], + "spans": [ + { + "bbox": [ + 173, + 142, + 423, + 159 + ], + "type": "text", + "content": "Xianqi Wang*, Gangwei Xu*, Hao Jia, Xin Yang†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 163, + 416, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 163, + 416, + 177 + ], + "spans": [ + { + "bbox": [ + 176, + 163, + 416, + 177 + ], + "type": "text", + "content": "Huazhong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 165, + 178, + 425, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 178, + 425, + 191 + ], + "spans": [ + { + "bbox": [ + 165, + 178, + 425, + 191 + ], + "type": "text", + "content": "{xianqiw, gwxu, haojia, xinyang2014}@hust.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 218, + 192, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 218, + 192, + 232 + ], + "spans": [ + { + "bbox": [ + 143, + 218, + 192, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 244, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 290, + 521 + ], + "type": "text", + "content": "Stereo matching methods based on iterative optimization, like RAFT-Stereo and IGEV-Stereo, have evolved into a cornerstone in the field of stereo matching. However, these methods struggle to simultaneously capture high-frequency information in edges and low-frequency information in smooth regions due to the fixed receptive field. As a result, they tend to lose details, blur edges, and produce false matches in textureless areas. In this paper, we propose Selective Recurrent Unit (SRU), a novel iterative update operator for stereo matching. The SRU module can adaptively fuse hidden disparity information at multiple frequencies for edge and smooth regions. To perform adaptive fusion, we introduce a new Contextual Spatial Attention (CSA) module to generate attention maps as fusion weights. The SRU empowers the network to aggregate hidden disparity information across multiple frequencies, mitigating the risk of vital hidden disparity information loss during iterative processes. To verify SRU's universality, we apply it to representative iterative stereo matching methods, collectively referred to as Selective-Stereo. Our Selective-Stereo ranks " + }, + { + "bbox": [ + 46, + 244, + 290, + 521 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 244, + 290, + 521 + ], + "type": "text", + "content": " on KITTI 2012, KITTI 2015, ETH3D, and Middlebury leaderboards among all published methods. Code is available at https://github.com/Windsrain/Selective-Stereo." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 575, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 647 + ], + "type": "text", + "content": "Stereo matching is a fundamental area of research in computer vision. It explores the calculation of displacement, referred to as disparity, between matching points in a pair of rectified images. This technique plays a significant role in various applications, including 3D reconstruction and autonomous driving." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 647, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 288, + 685 + ], + "type": "text", + "content": "With the advancement of deep learning, learning-based stereo matching [4, 14, 15, 17, 34, 35] have progressively displaced traditional methods and significantly enhancing" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 217, + 425, + 332 + ], + "blocks": [ + { + "bbox": [ + 306, + 217, + 425, + 332 + ], + "lines": [ + { + "bbox": [ + 306, + 217, + 425, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 425, + 332 + ], + "type": "image", + "image_path": "002f571a9e974f670eefa979e7102f82b607bc7516ea8181d557c1da34de0207.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 217, + 539, + 332 + ], + "blocks": [ + { + "bbox": [ + 425, + 217, + 539, + 332 + ], + "lines": [ + { + "bbox": [ + 425, + 217, + 539, + 332 + ], + "spans": [ + { + "bbox": [ + 425, + 217, + 539, + 332 + ], + "type": "image", + "image_path": "17257d6328d2e82198ac8b669786cf3ae5f1b0c281975362c037f1fa96310447.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 337, + 384, + 389 + ], + "blocks": [ + { + "bbox": [ + 307, + 337, + 384, + 389 + ], + "lines": [ + { + "bbox": [ + 307, + 337, + 384, + 389 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 384, + 389 + ], + "type": "image", + "image_path": "35c241f98aded3ac51826d196765ea35f3e1df09c7547a1e98fbe1866111d9d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 392, + 364, + 400 + ], + "lines": [ + { + "bbox": [ + 327, + 392, + 364, + 400 + ], + "spans": [ + { + "bbox": [ + 327, + 392, + 364, + 400 + ], + "type": "text", + "content": "Left Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 385, + 337, + 462, + 389 + ], + "blocks": [ + { + "bbox": [ + 385, + 337, + 462, + 389 + ], + "lines": [ + { + "bbox": [ + 385, + 337, + 462, + 389 + ], + "spans": [ + { + "bbox": [ + 385, + 337, + 462, + 389 + ], + "type": "image", + "image_path": "86586d706209ab26d3f7c432827e7e78a0b4da09656fc674f6fb6062ea4a65e9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 392, + 454, + 400 + ], + "lines": [ + { + "bbox": [ + 394, + 392, + 454, + 400 + ], + "spans": [ + { + "bbox": [ + 394, + 392, + 454, + 400 + ], + "type": "text", + "content": "RAFT-Stereo [17]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 463, + 337, + 541, + 389 + ], + "blocks": [ + { + "bbox": [ + 463, + 337, + 541, + 389 + ], + "lines": [ + { + "bbox": [ + 463, + 337, + 541, + 389 + ], + "spans": [ + { + "bbox": [ + 463, + 337, + 541, + 389 + ], + "type": "image", + "image_path": "ea01c757381138d4a041dccfea298fefea82731edfbf36cb4429944d9f4fea04.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 392, + 529, + 399 + ], + "lines": [ + { + "bbox": [ + 475, + 392, + 529, + 399 + ], + "spans": [ + { + "bbox": [ + 475, + 392, + 529, + 399 + ], + "type": "text", + "content": "Selective-RAFT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 406, + 384, + 458 + ], + "blocks": [ + { + "bbox": [ + 307, + 406, + 384, + 458 + ], + "lines": [ + { + "bbox": [ + 307, + 406, + 384, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 384, + 458 + ], + "type": "image", + "image_path": "df2a9b545d1e25c01d556e5c69a4ee46d374a5102b5cf1a77c512bd5fdece4a1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 460, + 364, + 469 + ], + "lines": [ + { + "bbox": [ + 327, + 460, + 364, + 469 + ], + "spans": [ + { + "bbox": [ + 327, + 460, + 364, + 469 + ], + "type": "text", + "content": "Left Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 386, + 405, + 462, + 458 + ], + "blocks": [ + { + "bbox": [ + 386, + 405, + 462, + 458 + ], + "lines": [ + { + "bbox": [ + 386, + 405, + 462, + 458 + ], + "spans": [ + { + "bbox": [ + 386, + 405, + 462, + 458 + ], + "type": "image", + "image_path": "654fc3d6c856316bfefaef6812f4570fc5eb8686ca537e97cfd4940ed66cf325.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 460, + 452, + 469 + ], + "lines": [ + { + "bbox": [ + 395, + 460, + 452, + 469 + ], + "spans": [ + { + "bbox": [ + 395, + 460, + 452, + 469 + ], + "type": "text", + "content": "IGEV-Stereo [34]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 463, + 405, + 541, + 458 + ], + "blocks": [ + { + "bbox": [ + 463, + 405, + 541, + 458 + ], + "lines": [ + { + "bbox": [ + 463, + 405, + 541, + 458 + ], + "spans": [ + { + "bbox": [ + 463, + 405, + 541, + 458 + ], + "type": "image", + "image_path": "8b1081ab24d0706573ac72907b023494f8ea2c431be679bcd90c9167b82806ca.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 460, + 528, + 468 + ], + "lines": [ + { + "bbox": [ + 475, + 460, + 528, + 468 + ], + "spans": [ + { + "bbox": [ + 475, + 460, + 528, + 468 + ], + "type": "text", + "content": "Selective-IGEV" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 470, + 547, + 536 + ], + "lines": [ + { + "bbox": [ + 305, + 470, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 547, + 536 + ], + "type": "text", + "content": "Figure 1. Row 1: Comparisons with state-of-the-art stereo methods on KITTI 2012 [13] and KITTI 2015 [21], ETH3D [24] and Middlebury [23] leaderboards. Row 2: Visual comparison with RAFT-Stereo on ETH3D. Row 3: Visual comparison with IGEV-Stereo on Middlebury. Our method distinguishes subtle details and sharp edges and performs well in weak texture regions." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 554, + 546, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 546, + 686 + ], + "type": "text", + "content": "the accuracy of disparity estimation. Initially, aggregation-based methods [7, 32, 41] led the development of stereo matching algorithms. These methods begin by defining a maximum range of disparity, constructing a 4D cost volume using feature maps, and subsequently employing 3D CNN to filter the volume and derive the final disparity map. Such methods focus on filtering the initially coarse cost volume, thus effectively aggregating geometry information. However, cost aggregation requires a large number of convolutions, resulting in high computational costs, making it difficult to be applied to high-resolution images." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Recently, a novel class of methods based on iterative optimization [11, 16, 17, 44] has been gaining prominence and" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19701" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "achieving state-of-the-art performance on several leaderboards. These methods begin by constructing an all-pairs cost volume, indexing a local cost volume from the original cost volume, and subsequently employing recurrent units [10] to calculate disparity residuals and update the disparity prediction. One major advantage of these methods is their ability to capture all candidate matching points without predefining the range of disparities. Additionally, these methods don't need to aggregate the cost volume using a large number of redundant convolutions. Instead, a continuous update of the disparity prediction is achieved through lightweight recurrent units during iterations. Therefore, these methods are capable of processing high-resolution images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 387 + ], + "type": "text", + "content": "However, iterative methods encounter several challenges. Firstly, the all-pairs cost volume includes considerable noisy information [34], potentially causing the loss of crucial information when iterating the hidden information. Besides, as the network iterates, the hidden information increasingly incorporates global low-frequency information while losing local high-frequency information like edges and thin objects [44]. Secondly, the existing recurrent units possess a fixed receptive field, leading the network to solely concentrate on information at the current frequency and ignore other frequencies, such as detailed, edge, and textureless information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": "In this paper, we propose Selective Recurrent Unit (SRU) to address the limitations of traditional recurrent units. As Chen et al. [5] mentions features contain information at different frequencies, high-frequency information describes rapidly changing fine details, while low-frequency information describes smoothly changing structures. Unlike traditional recurrent units that treat information at different frequencies equally, our SRU incorporates multiple branches of GRU, each with a distinct kernel size representing different receptive fields. The hidden information obtained from each GRU branch is fused and then fed into the next iteration. This fusion enables the capture of information from different receptive fields at different frequencies, while also performing secondary filtering to reduce noise information from local cost volume. To further enhance the fusion process, we propose a Contextual Spatial Attention (CSA) module to utilize the context information. Instead of simply summarizing information from different branches, CSA introduces attention maps extracted from the context information. After doing so, information captured by small kernels has large weights in regions like edge, while information captured by large kernels has large weights in regions like low-texture. These attention maps determine the weight of fusion, allowing the network to adaptively select suitable information based on different image regions. Besides, we prove the effectiveness and universality of our module by transferring it to different iterative networks. All networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "are collectively referred to as Selective-Stereo. By doing so, we consistently improve the performance of these networks without introducing a significant increase in parameters and time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 121, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 239 + ], + "type": "text", + "content": "We demonstrate the effectiveness of our method on several stereo benchmarks. On Scene Flow [20], our Selective-RAFT reaches the state-of-the-art EPE of 0.47, and our Selective-IGEV even achieves a new state-of-the-art EPE of 0.44. And as shown in Fig. 1, our Selective-RAFT surpasses RAFT-Stereo by a large margin and achieves competitive performance compared with the state-of-the-art methods on KITTI [13, 21] leaderboards. Our Selective-IGEV ranks " + }, + { + "bbox": [ + 304, + 121, + 547, + 239 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 121, + 547, + 239 + ], + "type": "text", + "content": " on KITTI, ETH3D [24], and Middlebury [23] leaderboards among all published methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 240, + 539, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 539, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 539, + 251 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 253, + 545, + 371 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 253, + 544, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 544, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 544, + 275 + ], + "type": "text", + "content": "- We propose a novel iterative update operator SRU for iterative stereo matching methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "text", + "content": "- We introduce a new Contextual Spatial Attention module that generates attention maps for adaptively fusing hidden disparity information at multiple frequencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 312, + 544, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 544, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 544, + 334 + ], + "type": "text", + "content": "- We verify the universality of our SRU on several iterative stereo matching methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 335, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 545, + 371 + ], + "type": "text", + "content": "- Our method outperforms existing published methods on public leaderboards such as KITTI, ETH3D, and Middlebury." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 392, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 403, + 547, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 666 + ], + "type": "text", + "content": "Aggregation-based methods in stereo matching. Several aggregation-based methods [4, 7, 8, 14, 15, 25, 35-37, 41] have shown significant progress in the domain of stereo matching in recent years. DispNet [20] establishes the groundwork for subsequent network architecture. GCNet [15] proposes a 4D concatenate cost volume, which is subsequently regularized using 3D CNNs. Additionally, it also introduces the soft argmin function for disparity regression, resulting in a significant influence on subsequent methods. PSMNet [4] proposes a stacked hourglass 3D CNN, which improves the cost aggregation stage to enhance the network's ability to capture context information. GwcNet [14] proposes Group-wise Correlation Volume, which combines the advantages of correlation and concatenation volume. GA-Net [41] designs a semi-global guided aggregation layer and a local guided aggregation layer, inspired by the traditional semi-global matching algorithm, to further assist in aggregating global and geometry information in the network. Building upon Group-wise Correlation Volume, ACVNet [32] proposes Attention Concat Volume, which uses attention weights to suppress redundant information and maintain sufficient information for matching." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Iterative-based methods in stereo matching. In recent years, many iterative methods [17, 27, 33, 34], spearheaded by RAFT [17], have gradually become the mainstream of research. RAFT-Stereo [17] builds upon the optical flow" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19702" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 344 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 344 + ], + "type": "image", + "image_path": "d33e626a3ec484bfdcd1d10a05b2a81f673cd9317a19877d45028bf49006b2de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed Selective-Stereo (Selective-RAFT version). The Contextual Spatial Attention (CSA) module extracts attention maps from context information as a guide for Selective Recurrent Units (SRUs). Then the network iteratively updates the disparity using local cost volumes retrieved from the correlation pyramid and attention maps given by CSA through SRUs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 396, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 575 + ], + "type": "text", + "content": "method RAFT [27] by introducing an all-pairs cost volume pyramid that maintains high resolution. It extracts local correlation features from this pyramid, performs iterative disparity updates using GRU-based update operators, and incorporates a multi-level GRU to expand the receptive field. On this basis, IGEV-Stereo [34] asserts that the initial cost volume is excessively coarse. To alleviate the need for iterations and reduce time overhead, it proposes to use a lightweight cost aggregation network before iterations. CREstereo [16] designs a hierarchical network in a coarse-to-fine manner, as well as a stacked cascaded architecture for inference in place of the original single-resolution iterative structure. DLNR [44] proposes the use of LSTM as a replacement for GRU, providing the advantage of decoupling the update of hidden states from disparity prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "Frequency information application in vision. There are several works[5, 6, 39] that focus on using frequency information in computer vision. Chen et al. [5] propose the octave convolution to factorize the mixed feature maps by their frequencies. Xu et al. [39] propose a method of learning in the frequency domain and suggest that CNN models are more sensitive to low-frequency channels than high-frequency. DSGAN [12] introduces the frequency separation into super-resolution. LITv2 [22] proposes to disentangle the high/low-frequency patterns in an attention layer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 394, + 361, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 361, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 361, + 407 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 418, + 546, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 465 + ], + "type": "text", + "content": "In this section, we present the overall architecture of Selective-Stereo. Because our method can be plugged into different networks, we take Selective-RAFT (Fig. 2) as an example and focus on illustrating its key components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 482, + 417, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 417, + 494 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 417, + 494 + ], + "type": "text", + "content": "3.1. Feature Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 552 + ], + "type": "text", + "content": "To ensure fair comparisons, Selective-RAFT maintains consistency with RAFT-Stereo [17] by employing its feature extraction network. Feature extraction comprises two main components: feature network and context network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": "Feature Network. Given the left and the right images " + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{l(r)}\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": ", we first downsample them to " + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "inline_equation", + "content": "1 / 2" + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": " resolution using a " + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "inline_equation", + "content": "7\\times 7" + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": " convolutional layer. Then, a series of residual blocks is employed to extract features and we apply another downsampling layer to get features at " + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "inline_equation", + "content": "1 / 4" + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": " resolution in the middle. Lastly, a " + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": " convolutional layer is applied to get the final left and right features " + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{g}\\in \\mathbb{R}^{C\\times \\frac{H}{4}\\times \\frac{W}{4}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 650 + ], + "type": "text", + "content": " with suitable dimensions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "Context Network. Its architecture remains consistent with the feature network, and it adds a series of residual blocks and two additional downsampling layers, obtaining multi-level context features " + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^c" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "i = 1, 2, 3" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": ") at " + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "1/4, 1/8, 1/16" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": " resolutions. Then we can get the initial hidden and the con" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19703" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 118, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 118, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 118, + 83 + ], + "type": "text", + "content": "text information:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 81, + 287, + 100 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 81, + 287, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 81, + 287, + 100 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {i} = \\tanh \\left(\\mathbf {f} _ {i} ^ {c}\\right) \\tag {1}", + "image_path": "4b6a06fb2118b239ae31a11feb8180cb281e8db2a36970abc3a5ee70e780c8cb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 137, + 97, + 199, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 97, + 199, + 110 + ], + "spans": [ + { + "bbox": [ + 137, + 97, + 199, + 110 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i} = \\operatorname {R e L U} \\left(\\mathbf {f} _ {i} ^ {c}\\right)", + "image_path": "8b7d5cf5216ca482aee22f614d609eac8214be973bd4691865c63014531ec797.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 122, + 194, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 194, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 194, + 133 + ], + "type": "text", + "content": "3.2. Cost Volume Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "text", + "content": "Given the left and the right features " + }, + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{f},\\mathbf{g}" + }, + { + "bbox": [ + 47, + 140, + 287, + 164 + ], + "type": "text", + "content": ", we first construct an all-pairs correlate cost volume:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 172, + 287, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 172, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 83, + 172, + 287, + 199 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {i j k} = \\sum_ {h} \\mathbf {f} _ {h i j} \\cdot \\mathbf {g} _ {h i k}, \\mathbf {C} \\in \\mathbb {R} ^ {\\frac {H}{4} \\times \\frac {W}{4} \\times \\frac {W}{4}} \\tag {2}", + "image_path": "e145072c891df3939de8dde55afcf4f4d58fd965ad5ee0f194c22a4f01cf8670.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 207, + 288, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 288, + 243 + ], + "type": "text", + "content": "Then we construct a 4-level correlation pyramid " + }, + { + "bbox": [ + 47, + 207, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_i\\}" + }, + { + "bbox": [ + 47, + 207, + 288, + 243 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 207, + 288, + 243 + ], + "type": "inline_equation", + "content": "i = 1,2,3,4" + }, + { + "bbox": [ + 47, + 207, + 288, + 243 + ], + "type": "text", + "content": ") by using 1D average pooling with a kernel size of 2 and a stride of 2 at the last dimension." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 251, + 243, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 243, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 243, + 262 + ], + "type": "text", + "content": "3.3. Contextual Spatial Attention Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 269, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 287, + 364 + ], + "type": "text", + "content": "To help information from different receptive fields and frequencies fuse, the Contextual Spatial Attention (CSA) module extracts multi-level attention maps from context information as guidance. As illustrated in Fig. 3, CSA can be divided into two submodules: Channel Attention Enhancement (CAE) and Spatial Attention Extractor (SAE). These submodules are derived from CBAM [31] and we simplify them to better adapt to stereo matching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "text", + "content": "Channel Attention Enhancement. Given a context information map " + }, + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "text", + "content": ", we first use an average-pooling and a max-pooling operation on the spatial dimension to get two maps " + }, + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{avg}, \\mathbf{f}_{max} \\in \\mathbb{R}^{C \\times 1 \\times 1}" + }, + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "text", + "content": ". Then we use two convolutional layers to perform feature transformation on these maps separately. After that, we add these two maps together and use the sigmoid function to convert them into weights " + }, + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_c \\in \\mathbb{R}^{C \\times 1 \\times 1}" + }, + { + "bbox": [ + 47, + 365, + 287, + 496 + ], + "type": "text", + "content": " between 0 and 1. Lastly, using an element-wise product, the initial map can capture which channel map has high feature values to be enhanced, and which channel map has low feature values to be suppressed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 497, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 640 + ], + "type": "text", + "content": "Spatial Attention Extractor. After the CAE module, we continue to use the same pooling operations, but now we pool on the channel dimension. Then we concatenate these pooling maps to form a map in " + }, + { + "bbox": [ + 47, + 497, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2\\times H\\times W}" + }, + { + "bbox": [ + 47, + 497, + 287, + 640 + ], + "type": "text", + "content": " and use one convolutional layer with a sigmoid function to generate the final attention map. Reviewing previous operations, this attention map has high weights in regions needing high-frequency information because this information possesses high feature values in the context information. Similarly, it has low weights in regions needing low-frequency information. In general, the attention map can explicitly distinguish regions that need information at different frequencies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 185, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 185, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 185, + 658 + ], + "type": "text", + "content": "3.4. Selective Recurrent Unit" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 712 + ], + "type": "text", + "content": "To capture information at different frequencies, Selective Recurrent Unit (SRU) uses attention maps extracted by CSA to fuse hidden information derived from GRUs with different kernel sizes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "Multi-level update structure. As illustrated in Fig. 4, SRUs at " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "1/16" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " resolutions take the attention map, context information, hidden information at the same resolution, and the hidden information at adjacent resolutions as inputs. At " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " resolution, SRUs take disparity, and local cost volume as additional inputs, and then their outputs will go through two convolutional layers to generate disparity residuals. The local cost volume is derived from the all-pairs correlation pyramid in the same way as RAFT-Stereo [17]. At last, disparities at " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " resolution will be upsampled into full resolution using the convex combination." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "type": "text", + "content": "SRU's architecture. A single GRU can be defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 343, + 235, + 471, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 235, + 471, + 248 + ], + "spans": [ + { + "bbox": [ + 343, + 235, + 471, + 248 + ], + "type": "interline_equation", + "content": "z _ {k} = \\sigma (\\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {z})),", + "image_path": "ef2ac7805f97cc6deeebbe98b3bb3b50e2c0c778121e224efed918785a532cc9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 250, + 545, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 250, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 343, + 250, + 545, + 272 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} r _ {k} = \\sigma (\\operatorname {C o n v} ([ h _ {k - 1}, x _ {k} ], W _ {r})), \\\\ \\tilde {I} _ {\\text {一}} = \\operatorname {c o v} (G _ {m} ([ \\circ_ {0} ]), W _ {r})) \\end{array} \\tag {3}", + "image_path": "d0717af6506b34534571dd824858f5bb4a30ba66838ad9c56cb707a28957420a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 343, + 262, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 262, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 343, + 262, + 545, + 281 + ], + "type": "interline_equation", + "content": "\\tilde {h} _ {k} = \\tanh (\\operatorname {C o n v} ([ r _ {k} \\odot h _ {k - 1}, x _ {k} ], W _ {h})), \\tag {3}", + "image_path": "1671f42e32a49e6b0ea89226952830c69bb35115a4dbe008071bc071ea54db73.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 283, + 478, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 283, + 478, + 297 + ], + "spans": [ + { + "bbox": [ + 343, + 283, + 478, + 297 + ], + "type": "interline_equation", + "content": "h _ {k} = \\left(1 - z _ {k}\\right) \\odot h _ {k - 1} + z _ {k} \\odot \\tilde {h} _ {k}", + "image_path": "029d5dffe48a3d78ef59e2c880334e2010031d546fc2feffc1dad44526dbd482.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "text", + "content": " is the concatenation of disparity, correlation, hidden information, and context information previously defined. Unlike RAFT-Stereo [17] that divide the context information into " + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "inline_equation", + "content": "c_{z}, c_{r}, c_{h}" + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "text", + "content": ", we add it into " + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "text", + "content": " because using convolutions with different kernel sizes can fully utilize context information." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 379, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 545, + 401 + ], + "type": "text", + "content": "As illustrated in Fig. 3, a single SRU can be defined as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 361, + 402, + 545, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 402, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 361, + 402, + 545, + 417 + ], + "type": "interline_equation", + "content": "h _ {k} = \\mathbf {A} \\odot h _ {k} ^ {s} + (1 - \\mathbf {A}) \\odot h _ {k} ^ {l} \\tag {4}", + "image_path": "97b34d0de0b230fab92c163e29f67ef514e249458dcdb03e3b835ac8b08ed46d.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "text", + "content": " denotes the attention map derived from CSA at the same resolution, " + }, + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "inline_equation", + "content": "h_k^s" + }, + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "text", + "content": " denotes the GRU with smaller kernel sizes and " + }, + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "inline_equation", + "content": "h_k^l" + }, + { + "bbox": [ + 304, + 422, + 545, + 457 + ], + "type": "text", + "content": " denotes the larger one." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "type": "text", + "content": "As Sec. 3.3 mentioned, the attention map has high weights in regions needing high-frequency information. Therefore, the GRU with smaller kernel sizes that can capture high-frequency information like edge, and thin objects should do element-wise products with the attention map directly, and the GRU with larger kernel sizes should do element-wise products with the contrary attention map." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 543, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 543, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 543, + 545, + 567 + ], + "type": "text", + "content": "Receptive fields analysis. The repetitive fields computing formula [1] can be defined as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 364, + 575, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 575, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 364, + 575, + 545, + 609 + ], + "type": "interline_equation", + "content": "r _ {0} = \\sum_ {l = 1} ^ {L} \\left(\\left(k _ {l} - 1\\right) \\prod_ {i = 1} ^ {l - 1} s _ {i}\\right) + 1 \\tag {5}", + "image_path": "4c7f0828d70c7655aa663a853d66aacce01bb807854eaca0ee1b3f102006edb3.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "text", + "content": " denotes the kernel size, " + }, + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "text", + "content": " denotes the stride size, and " + }, + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 305, + 617, + 544, + 640 + ], + "type": "text", + "content": " denotes the whole network." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Given a multi-level structure like Fig. 4, if we take the " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": " resolution as the basis, and the downsampling operations can be regarded as a convolution with kernel size 3, stride size 2, this structure's receptive fields are " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "2k + 3" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "3k + 6" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": ". That means it only has 3 fixed receptive fields in total." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19704" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 372, + 190 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 372, + 190 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 372, + 190 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 372, + 190 + ], + "type": "image", + "image_path": "64eb6c020445a9fe0255b3ca8b579a8dc09c2bd8e985e1439137001b300ae8db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 546, + 211 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 546, + 211 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 546, + 211 + ], + "type": "text", + "content": "Figure 3. The architecture of proposed modules. Left: Contextual Spatial Attention (CSA) module. Right: Selective Recurrent Unit (SRU)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 374, + 71, + 545, + 190 + ], + "blocks": [ + { + "bbox": [ + 374, + 71, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 374, + 71, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 374, + 71, + 545, + 190 + ], + "type": "image", + "image_path": "174b196e660c161d4133595da944539a3c1c14aafe2e1b8dfe811c7b1b1c1d47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 67, + 219, + 269, + 316 + ], + "blocks": [ + { + "bbox": [ + 67, + 219, + 269, + 316 + ], + "lines": [ + { + "bbox": [ + 67, + 219, + 269, + 316 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 269, + 316 + ], + "type": "image", + "image_path": "f616a01686288c415c2340eb03c56fd9b28f1b0d5cf4fa3948f9d77b21e3e579.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "type": "text", + "content": "Figure 4. Multi-level SRU. Information is passed between SRUs at adjacent resolutions. Dashed arrows represent upsampling and downsampling operations. At " + }, + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 46, + 323, + 288, + 369 + ], + "type": "text", + "content": " resolution, disparity and local cost volume will be additional information put into SRUs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 379, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 475 + ], + "type": "text", + "content": "If we replace GRUs with our SRUs with a small kernel size " + }, + { + "bbox": [ + 46, + 379, + 287, + 475 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 379, + 287, + 475 + ], + "type": "text", + "content": " and a large kernel size " + }, + { + "bbox": [ + 46, + 379, + 287, + 475 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 379, + 287, + 475 + ], + "type": "text", + "content": ", the multi-level structure will have 6 receptive fields initially. Besides, pixels in hidden information are affected by different receptive fields during fusion, and the fusion is influenced by attention maps adaptively. In general, the multi-level SRU holds dynamic receptive fields, and it enables itself to capture information at different frequencies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 482, + 136, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 136, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 136, + 495 + ], + "type": "text", + "content": "3.5. Loss Function" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": "We supervise our network on the L1 distance between all predicted disparities " + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{d}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": " and the ground truth disparity " + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{gt}" + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": " with increasing weights. The total loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 545, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 110, + 545, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\gamma^ {N - i} | | \\mathbf {d} _ {i} - \\mathbf {d} _ {g t} | | _ {1} \\tag {6}", + "image_path": "8081ec04bdfc115a786ab0a776088ffbd75492594cd8b97640451206bb25d510.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 586, + 249, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 249, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 249, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 586, + 249, + 598 + ], + "type": "inline_equation", + "content": "\\gamma = 0.9" + }, + { + "bbox": [ + 47, + 586, + 249, + 598 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 586, + 249, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 586, + 249, + 598 + ], + "type": "text", + "content": " is the number of iterations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 609, + 128, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 624 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "Scene Flow [20] is a synthetic dataset including 35,454 training pairs and 4,370 testing pairs with dense disparity maps. For training and testing, we use the finalpass version, because it contains more realistic and difficult effects than the cleanpass version. KITTI 2012 [13] and KITTI 2015 [21] are datasets for real-world driving scenes. KITTI 2012 contains 194 training pairs and 195 testing pairs, and" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 221, + 546, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 546, + 293 + ], + "type": "text", + "content": "KITTI 2015 contains 200 training pairs and 200 testing pairs. ETH3D [24] is a collection of gray-scale stereo pairs containing 27 training pairs and 20 testing pairs for indoor and outdoor scenes. Middlebury [23] is a high-resolution dataset containing 15 training pairs and 15 testing pairs for indoor scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 304, + 440, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 440, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 440, + 316 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 323, + 546, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 419 + ], + "type": "text", + "content": "We implement our Selective-Stereo with PyTorch and the model is trained on NVIDIA RTX 3090 GPUs. For all experiments, we use the AdamW [19] optimizer and clip gradients to the range [-1, 1]. We use the one-cycle learning rate schedule with a learning rate of 2e-4. We first train our model on Scene Flow with a batch size of 8 for 200k steps as the pretrained model. The crop size is " + }, + { + "bbox": [ + 304, + 323, + 546, + 419 + ], + "type": "inline_equation", + "content": "320 \\times 720" + }, + { + "bbox": [ + 304, + 323, + 546, + 419 + ], + "type": "text", + "content": ", and we use 22 update iterations during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 429, + 400, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 400, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 400, + 442 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 448, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 484 + ], + "type": "text", + "content": "In this section, we evaluate our model in different settings to verify our proposed modules in several aspects. All results use 32 update iterations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 485, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 640 + ], + "type": "text", + "content": "Effectiveness of proposed modules. To verify the effectiveness of our proposed modules, we take RAFT-Stereo [17] as the baseline and replace its GRUs with our SRUs. As shown in Tab. 1, the proposed SRU can improve the accuracy even without CSA. It means that if we just sum up the information from different branches, the growth of receptive fields can be beneficial for inference. If we add our CSA but invert the weights of the attention maps, the effect even decreases. That validates that our CSA's attention maps do indeed reflect the weights of information at different frequencies in regions. Therefore, if we add CSA normally, the full model (Selective-RAFT) can achieve the best performance with only a " + }, + { + "bbox": [ + 304, + 485, + 546, + 640 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 485, + 546, + 640 + ], + "type": "text", + "content": " increase in parameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Universality of proposed modules. To verify the universality of our proposed modules, we take three typical iterative stereo matching methods as the baseline and replace their GRUs with SRUs. Especially, in DLNR [44], the recurrent units are LSTMs but not GRUs, so we just replace GRUs inside SRUs with LSTMs to make a fair comparison." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19705" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 70, + 491, + 148 + ], + "blocks": [ + { + "bbox": [ + 106, + 70, + 491, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 491, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 491, + 148 + ], + "type": "table", + "html": "
ModelGRUSRUCSA (Contrary)CSAEPE (px)>1px (%)Param (M)
Baseline (RAFT-Stereo)✓0.536.0811.12
SRU✓0.505.3811.65
SRU+CSA (Contrary)✓✓0.505.5811.65
Full model (Selective-RAFT)✓✓0.475.3211.65
", + "image_path": "65eb3996d08f431b68de98e4b69bb12ee8ffe08f7cb4f15f2c0b080f68886111.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 187, + 222, + 327 + ], + "blocks": [ + { + "bbox": [ + 74, + 187, + 222, + 327 + ], + "lines": [ + { + "bbox": [ + 74, + 187, + 222, + 327 + ], + "spans": [ + { + "bbox": [ + 74, + 187, + 222, + 327 + ], + "type": "image", + "image_path": "24e5c71fca52740b8f87028bb5ceef85a4a2a280161b03979dcaab001a1af1ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 328, + 167, + 338 + ], + "lines": [ + { + "bbox": [ + 129, + 328, + 167, + 338 + ], + "spans": [ + { + "bbox": [ + 129, + 328, + 167, + 338 + ], + "type": "text", + "content": "Left Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 188, + 370, + 327 + ], + "blocks": [ + { + "bbox": [ + 223, + 188, + 370, + 327 + ], + "lines": [ + { + "bbox": [ + 223, + 188, + 370, + 327 + ], + "spans": [ + { + "bbox": [ + 223, + 188, + 370, + 327 + ], + "type": "image", + "image_path": "38290070a432c24a353b6772cfc56e730c36a518e438e39b5a804646eeb69b56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 328, + 324, + 337 + ], + "lines": [ + { + "bbox": [ + 279, + 328, + 324, + 337 + ], + "spans": [ + { + "bbox": [ + 279, + 328, + 324, + 337 + ], + "type": "text", + "content": "IGEV-Stereo" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 188, + 520, + 327 + ], + "blocks": [ + { + "bbox": [ + 370, + 188, + 520, + 327 + ], + "lines": [ + { + "bbox": [ + 370, + 188, + 520, + 327 + ], + "spans": [ + { + "bbox": [ + 370, + 188, + 520, + 327 + ], + "type": "image", + "image_path": "08e2403e1370b9fb3d53a974377c462a39d86b1bb6a71ce008517a6ca7060f26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 328, + 479, + 337 + ], + "lines": [ + { + "bbox": [ + 424, + 328, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 424, + 328, + 479, + 337 + ], + "type": "text", + "content": "Selective-IGEV" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 63, + 367, + 269, + 438 + ], + "blocks": [ + { + "bbox": [ + 46, + 154, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 547, + 177 + ], + "type": "text", + "content": "Table 1. Ablation study of the effectiveness of proposed modules on the Scene Flow test set. SRU denotes Selective Recurrent Unit, and CSA denotes Contextual Spatial Attention. Contrary means we invert the weights of the attention maps. The baseline is RAFT-Stereo." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 367, + 269, + 438 + ], + "lines": [ + { + "bbox": [ + 63, + 367, + 269, + 438 + ], + "spans": [ + { + "bbox": [ + 63, + 367, + 269, + 438 + ], + "type": "table", + "html": "
ModelEPE (px)>1px (%)Param (M)
RAFT-Stereo [17]0.536.0811.12
Selective-RAFT0.475.3211.65
IGEV-Stereo [34]0.475.2112.60
Selective-IGEV0.444.9813.14
DLNR [44]0.495.0657.37
Selective-DLNR0.464.7358.09
", + "image_path": "ed11bcac44344f5eda4484aab549be34955f9a741ff61e4c3f23de38989b25bc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 52, + 461, + 280, + 522 + ], + "blocks": [ + { + "bbox": [ + 50, + 441, + 284, + 453 + ], + "lines": [ + { + "bbox": [ + 50, + 441, + 284, + 453 + ], + "spans": [ + { + "bbox": [ + 50, + 441, + 284, + 453 + ], + "type": "text", + "content": "Table 2. Ablation study of the universality of proposed modules." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 461, + 280, + 522 + ], + "lines": [ + { + "bbox": [ + 52, + 461, + 280, + 522 + ], + "spans": [ + { + "bbox": [ + 52, + 461, + 280, + 522 + ], + "type": "table", + "html": "
ModelNumber of Iterations
1234832
RAFT-Stereo [17]2.081.130.870.750.580.53
Selective-RAFT1.951.060.810.690.530.47
IGEV-Stereo [34]0.660.620.580.550.500.47
Selective-IGEV0.650.600.560.530.480.44
", + "image_path": "766495ecfd3a9e5cb6d9216e6ffd4796a3212f03f0bb330e3f57cf0433c126e7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 82, + 544, + 254, + 596 + ], + "blocks": [ + { + "bbox": [ + 73, + 525, + 261, + 536 + ], + "lines": [ + { + "bbox": [ + 73, + 525, + 261, + 536 + ], + "spans": [ + { + "bbox": [ + 73, + 525, + 261, + 536 + ], + "type": "text", + "content": "Table 3. Ablation study of the number of iterations." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 544, + 254, + 596 + ], + "lines": [ + { + "bbox": [ + 82, + 544, + 254, + 596 + ], + "spans": [ + { + "bbox": [ + 82, + 544, + 254, + 596 + ], + "type": "table", + "html": "
Kernel SizesEPE (px)>1px (%)
1 × 1 + 1 × 50.485.41
3 × 3 + 1 × 50.485.30
1 × 1 + 3 × 30.475.32
", + "image_path": "094b275ef2a55acc698057ff28172b2c6e8bf606417ed76114c3eda3eed44a1d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 605, + 276, + 616 + ], + "lines": [ + { + "bbox": [ + 58, + 605, + 276, + 616 + ], + "spans": [ + { + "bbox": [ + 58, + 605, + 276, + 616 + ], + "type": "text", + "content": "Table 4. Ablation study of the size of convolutional kernels." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "As shown in Tab. 2, all methods have a significant improvement in the EPE metrics on Scene Flow, and the insertion of modules only results in a slight increase in parameters. Besides, as shown in Fig. 6, the CSA module generates different attention maps in different networks. In Selective-RAFT, because the cost volume contains a large amount of noisy information, the network needs more large kernels to" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 368, + 386, + 422 + ], + "blocks": [ + { + "bbox": [ + 60, + 347, + 533, + 360 + ], + "lines": [ + { + "bbox": [ + 60, + 347, + 533, + 360 + ], + "spans": [ + { + "bbox": [ + 60, + 347, + 533, + 360 + ], + "type": "text", + "content": "Figure 5. Qualitative results on the test set of KITTI. Our Selective-IGEV outperforms IGEV in detailed and weak texture regions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 368, + 386, + 422 + ], + "lines": [ + { + "bbox": [ + 307, + 368, + 386, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 386, + 422 + ], + "type": "image", + "image_path": "43de67bc4c680a303a7cea9c03bc11cd9d0e4e30f6d6850039e89935c548382b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 423, + 364, + 433 + ], + "lines": [ + { + "bbox": [ + 329, + 423, + 364, + 433 + ], + "spans": [ + { + "bbox": [ + 329, + 423, + 364, + 433 + ], + "type": "text", + "content": "Left Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 441, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 306, + 441, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 545, + 453 + ], + "type": "text", + "content": "Figure 6. Visualization of the attention map of different networks." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 387, + 368, + 465, + 422 + ], + "blocks": [ + { + "bbox": [ + 387, + 368, + 465, + 422 + ], + "lines": [ + { + "bbox": [ + 387, + 368, + 465, + 422 + ], + "spans": [ + { + "bbox": [ + 387, + 368, + 465, + 422 + ], + "type": "image", + "image_path": "86638174669e0c84281a09b9c1658f3aebe77890f5641e689ec5e590eef68519.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 423, + 451, + 432 + ], + "lines": [ + { + "bbox": [ + 400, + 423, + 451, + 432 + ], + "spans": [ + { + "bbox": [ + 400, + 423, + 451, + 432 + ], + "type": "text", + "content": "Selective-RAFT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 466, + 368, + 545, + 422 + ], + "blocks": [ + { + "bbox": [ + 466, + 368, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 466, + 368, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 466, + 368, + 545, + 422 + ], + "type": "image", + "image_path": "510077035b70fddc24befb840a93f9f329a30fbf3447277e7b2bd6918fcdcce1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 423, + 529, + 432 + ], + "lines": [ + { + "bbox": [ + 481, + 423, + 529, + 432 + ], + "spans": [ + { + "bbox": [ + 481, + 423, + 529, + 432 + ], + "type": "text", + "content": "Selective-IGEV" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 469, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 577 + ], + "type": "text", + "content": "filter the local cost volume. On the contrary, the cost volume has already been aggregated in Selective-IGEV, so the network tends to maintain high-frequency information using small kernels. Moreover, the cost volume in Selective-IGEV faces an over-smooth problem [34], and that's why the attention map tends to increase the weights of large kernels to recover edge regions. In general, the CSA module shows different tendencies in different networks, which is a reflection of its adaptive ability." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 579, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 674 + ], + "type": "text", + "content": "Number of iterations. Our Selective-Stereo can achieve better performance with a smaller number of iterations. As shown in Tab. 3, our Selective-RAFT get the same performance with only 8 iterations compared to RAFT-Stereo [17], and for IGEV-Stereo [34], our Selective-IGEV also get a slight improvement with a few iterations. It shows that our modules can make secondary filtering to reduce noisy information from the initial cost volume." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Size of convolution kernels. We verify different kernel sizes on Scene Flow as shown in Tab. 4. At last, we choose the combination of " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " as our default config-" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19706" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 518, + 92 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 518, + 92 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 518, + 92 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 518, + 92 + ], + "type": "table", + "html": "
MethodCSPNLEASTereoLaC + GANetACVNetIGEV-StereoSelective-RAFT (Ours)Selective-IGEV (Ours)
EPE (px)0.780.780.720.480.470.470.44
", + "image_path": "5e6c598556672dd025b6cbd792710d139ffae009bc867c76676becc88d30ac69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 113, + 541, + 260 + ], + "blocks": [ + { + "bbox": [ + 195, + 95, + 398, + 106 + ], + "lines": [ + { + "bbox": [ + 195, + 95, + 398, + 106 + ], + "spans": [ + { + "bbox": [ + 195, + 95, + 398, + 106 + ], + "type": "text", + "content": "Table 5. Quantitative evaluation on Scene Flow test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 113, + 541, + 260 + ], + "lines": [ + { + "bbox": [ + 52, + 113, + 541, + 260 + ], + "spans": [ + { + "bbox": [ + 52, + 113, + 541, + 260 + ], + "type": "table", + "html": "
MethodKITTI 2012KITTI 2015Run-time (s)
2-noc2-all3-noc3-allEPE-nocEPE-allD1-bgD1-fgD1-all
AcfNet [42]1.832.351.171.540.50.51.513.801.890.48
LEAStereo [9]1.902.391.131.450.50.51.402.911.650.30
ACVNet [32]1.832.351.131.470.40.51.373.071.650.20
RAFT-Stereo [17]1.922.421.301.660.40.51.583.051.820.38
PCWNet [25]1.692.181.041.370.40.51.583.051.820.38
LaC + GANet [18]1.722.261.051.420.40.51.442.831.671.80
CREStereo [16]1.722.181.141.460.40.51.452.861.690.41
IGEV-Stereo [34]1.712.171.121.440.40.41.382.671.590.18
Selective-RAFT (Ours)1.642.091.101.430.40.51.412.711.630.45
Selective-IGEV (Ours)1.592.051.071.380.40.41.332.611.550.24
", + "image_path": "638230c846d3239a202260d0bc644636df2b1def048f131b6b59d3047d5f5bb6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 288, + 276, + 365 + ], + "blocks": [ + { + "bbox": [ + 177, + 268, + 416, + 279 + ], + "lines": [ + { + "bbox": [ + 177, + 268, + 416, + 279 + ], + "spans": [ + { + "bbox": [ + 177, + 268, + 416, + 279 + ], + "type": "text", + "content": "Table 6. Quantitative evaluation on KITTI 2012 and KITTI 2015." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 288, + 276, + 365 + ], + "lines": [ + { + "bbox": [ + 59, + 288, + 276, + 365 + ], + "spans": [ + { + "bbox": [ + 59, + 288, + 276, + 365 + ], + "type": "table", + "html": "
MethodEdgesNon-Edges
EPE>1pxEPE>1px
RAFT-Stereo [17]3.2129.160.536.53
Selective-RAFT2.4021.630.404.65
IGEV-Stereo [34]2.2320.420.414.58
Selective-IGEV2.1820.010.384.35
", + "image_path": "4a89ce0fc7a0482c8e5d7133e9995d7c5464154357d76b00af37fdf3bbf12690.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 395 + ], + "type": "text", + "content": "Table 7. Quantitative evaluation on Scene Flow test set in different regions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "type": "text", + "content": "uration, because it achieves a competitive performance and reduces computational costs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 441, + 231, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 231, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 231, + 453 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 460, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 521 + ], + "type": "text", + "content": "All fine-tuned models use the model pretrained on Scene Flow. Different target datasets use different finetune strategies. We validate two models called Selective-RAFT and Selective-IGEV using RAFT-Stereo [17] and IGEV-Stereo [34] as the baseline respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": "Scene Flow. As shown in Tab. 5, we achieve a new state-of-the-art EPE of 0.44 on Scene Flow with Selective-IGEV, which surpasses " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{LaC} + \\mathrm{GANet}" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " [18] by " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "38.89\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": ". Besides, our Selective-RAFT also achieves a competitive EPE of 0.47 compared to IGEV-Stereo [34] with smaller parameters. To validate the ability to fuse information by regions of our modules, we then split Scene Flow test set into two regions: edge regions and non-edge regions using the Canny operator. As shown in Tab. 7, our Selective-RAFT outperforms RAFT-Stereo [17] by " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "25.23\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "24.53\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " in edge regions and non-edge regions. Due to IGEV-Stereo's aggregated cost volume [34], there's only a slight improvement in edge regions, but our Selective-IGEV still outperforms it by " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "7.32\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " in non-edge regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "KITTI. We finetune our model on the mixed dataset of KITTI 2012 and KITTI 2015 with a batch size of 8 for 50k" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "text", + "content": "steps. Then we evaluate our Selective-Stereo on the test set of KITTI 2012 and KITTI 2015. As shown in Tab. 6, we achieve the best performance among all published methods for almost all metrics. On KITTI 2012, our Selective-RAFT outperforms RAFT-Stereo [17] by " + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "inline_equation", + "content": "14.58\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "inline_equation", + "content": "13.64\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "text", + "content": " on 2-noc and 2-all metrics, and our Selective-IGEV ranks " + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "text", + "content": " on these metrics. On KITTI 2015, our Selective-RAFT outperforms RAFT-Stereo [17] by " + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "inline_equation", + "content": "10.44\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "text", + "content": " on the D1-all metric, and our Selective-IGEV ranks " + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 291, + 545, + 434 + ], + "type": "text", + "content": " on all metrics with only 16 iterations same as IGEV-Stereo [34]. As shown in Fig. 5, our Selective-IGEV outperforms IGEV-Stereo [34] in detailed and textureless regions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "text", + "content": "ETH3D. Following CREStereo [16] and GMStereo [38], we use a collection of several public stereo datasets for training. The crop size is " + }, + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "inline_equation", + "content": "384 \\times 512" + }, + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "text", + "content": " and we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Sintel Stereo [3], InStereo2k [2] and ETH3D [24] datasets for 300k steps. Then we finetune it on the mixed CREStereo Dataset [16], InStereo2k [2] and ETH3D [24] datasets with for another 90k steps. As shown in Tab. 8, our Selective-RAFT outperforms RAFT-Stereo [17] by " + }, + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "inline_equation", + "content": "17.90\\%" + }, + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "text", + "content": " on Bad 0.5 metric, and our Selective-IGEV achieves the best performance among all published methods for almost all metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Middlebury. Also following CREStereo [16] and GMStereo [38], we first finetune the Scene Flow pretrained model on the mixed Tartan Air [29], CREStereo Dataset [16], Scene Flow [20], Falling Things [28], In-Stereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "384 \\times 512" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " for 200k steps. Then we finetune it on the mixed CREStereo Dataset [16], Falling Things [28], InStereo2k [2], CARLA HR-VS [40] and Middlebury [23] datasets using a crop size of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "384 \\times 768" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " with a batch size of 8 for another 100k steps. As shown in Tab. 8, our Selective-IGEV achieves the best performance among" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19707" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 521, + 217 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 521, + 217 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 521, + 217 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 521, + 217 + ], + "type": "table", + "html": "
MethodETH3DMiddlebury
Bad 1.0Bad 0.5Bad 4.0AvgErrBad 2.0Bad 1.0Bad 4.0AvgErr
CroCo-Stereo [30]0.993.270.130.147.2916.94.181.76
GMStereo [38]1.835.940.080.197.1423.62.961.31
HITNet [26]2.797.830.190.206.4613.33.811.71
IGEV-Stereo [34]1.123.520.110.144.839.413.332.89
RAFT-Stereo [17]2.447.040.150.184.749.372.751.27
CREStereo [16]0.983.580.100.133.718.252.041.15
EAI-Stereo [43]2.315.210.700.213.687.812.141.09
DLNR [44]----3.206.821.891.06
Selective-RAFT (Ours)1.695.780.130.17----
Selective-IGEV (Ours)1.233.060.050.122.516.531.360.91
", + "image_path": "3a0e93285602ec2cf21927260d1ee01f70b2198a555628782c81d56dc1c9399c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 225, + 546, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 546, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 546, + 246 + ], + "type": "text", + "content": "Table 8. Quantitative evaluation on ETH3D and Middlebury benchmarks. Note: Middlebury only allows one publish per paper, so we only publish our Selective-IGEV." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 85, + 248, + 509, + 476 + ], + "blocks": [ + { + "bbox": [ + 85, + 248, + 509, + 476 + ], + "lines": [ + { + "bbox": [ + 85, + 248, + 509, + 476 + ], + "spans": [ + { + "bbox": [ + 85, + 248, + 509, + 476 + ], + "type": "image", + "image_path": "696c137fe7da6974ba62c2d51b1762e3f0a790406419b4f71189d176ab67387e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "type": "text", + "content": "Figure 7. Qualitative results on the test set of Middlebury. The third column is the visualization of attention maps generated by the CSA module. Our Selective-IGEV outperforms IGEV in large textureless and thin object regions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 518, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 591 + ], + "type": "text", + "content": "all published methods. As shown in Fig 7, compared to IGEV-Stereo [34], our Selective-IGEV performs better in textureless, and detailed regions. The third column in Fig 7 is the visualization of attention maps generated by the CSA module. It shows that attention maps can surely split regions that require information at different frequencies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 608, + 119, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 119, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 119, + 620 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "We propose Selective-Stereo, a novel iterative stereo matching method. The proposed Contextual Spatial Attention module and Selective Recurrent Unit help the network capture information at different frequencies for edge and smooth regions. Our Selective-Stereo ranks " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " on KITTI, ETH3D, and Middlebury in almost all metrics among all published methods. It shows an ability to fuse information" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 518, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 543 + ], + "type": "text", + "content": "at different frequencies adaptively for edge and smooth regions with the help of attention maps extracted by CSA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 543, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 546, + 662 + ], + "type": "text", + "content": "However, our method still faces some challenges. Firstly, although our method can fuse information adaptively using attention maps, the SRU's receptive field is still limited by predefined values. Secondly, adding branches or increasing the sizes of convolutional kernels leads to high memory and time costs, so we will explore the combination of lightweight convolutions and our method to reduce memory costs. Lastly, it's also a good direction to do research on the combination of convolutions and self-attention due to their different advantages and receptive fields." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 674, + 546, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 546, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 546, + 711 + ], + "type": "text", + "content": "Acknowledgement. This research is supported by National Natural Science Foundation of China (62122029, 62061160490, U20B200007)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19708" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 4 (11):e21, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "text", + "content": "[3] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "type": "text", + "content": "[4] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 346 + ], + "type": "text", + "content": "[5] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Shuicheng Yan, and Jiashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3435-3444, 2019. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "text", + "content": "[6] Yajie Chen, Xin Yang, and Xiang Bai. Confidence-weighted mutual supervision on dual networks for unsupervised cross-modality image segmentation. Science China Information Sciences, 66(11):210104, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "type": "text", + "content": "[7] Junda Cheng, Xin Yang, Yuechuan Pu, and Peng Guo. Region separable stereo matching. IEEE Transactions on Multimedia, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[8] Junda Cheng, Gangwei Xu, Peng Guo, and Xin Yang. Coatsnet: Fully exploiting convolution and attention for stereo matching by region separation. International Journal of Computer Vision, pages 1-18, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "type": "text", + "content": "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33:22158-22169, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "text", + "content": "[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[11] Miaojie Feng, Junda Cheng, Hao Jia, Longliang Liu, Gangwei Xu, and Xin Yang. Mc-stereo: Multi-peak lookup and cascade search range for stereo matching. arXiv preprint arXiv:2311.02340, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "text", + "content": "[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3273-3282, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[15] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision, pages 66-75, 2017. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 238 + ], + "type": "text", + "content": "[16] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1, 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "type": "text", + "content": "[17] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 282, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 327 + ], + "type": "text", + "content": "[18] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1647-1655, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 349 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 547, + 415 + ], + "type": "text", + "content": "[20] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 2, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "text", + "content": "[21] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 492 + ], + "type": "text", + "content": "[22] Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Fast vision transformers with hilo attention. Advances in Neural Information Processing Systems, 35:14541-14554, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 492, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 559 + ], + "type": "text", + "content": "[23] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 1, 2, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "type": "text", + "content": "[24] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 1, 2, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 680 + ], + "type": "text", + "content": "[25] Zhelun Shen, Yuchao Dai, Xibin Song, Zhibo Rao, Dingfu Zhou, and Liangjun Zhang. Pcw-net: Pyramid combination and warping cost volume for stereo matching. In European Conference on Computer Vision, pages 280–297. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19709" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[27] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[28] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "type": "text", + "content": "[29] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "type": "text", + "content": "[30] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "text", + "content": "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "text", + "content": "[32] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 508 + ], + "type": "text", + "content": "[33] Gangwei Xu, Shujun Chen, Hao Jia, Miaojie Feng, and Xin Yang. Memory-efficient optical flow via radius-distribution orthogonal cost volume. arXiv preprint arXiv:2312.03790, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 511, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 564 + ], + "type": "text", + "content": "[34] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[35] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "text", + "content": "[36] Gangwei Xu, Huan Zhou, and Xin Yang. Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction. arXiv preprint arXiv:2301.02789, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[37] Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1959-1968, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[38] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 418 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[39] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1740-1749, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[40] Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5515-5524, 2019. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[41] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "type": "text", + "content": "[42] Youmin Zhang, Yimin Chen, Xiao Bai, Suihanjin Yu, Kun Yu, Zhiwei Li, and Kuiyuan Yang. Adaptive unimodal cost volume filtering for deep stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12926-12934, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 361 + ], + "type": "text", + "content": "[43] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Yong Zhao, Yitong Yang, and Ting Ouyang. Eai-stereo: Error aware iterative network for stereo matching. In Proceedings of the Asian Conference on Computer Vision, pages 315-332, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 364, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 545, + 418 + ], + "type": "text", + "content": "[44] Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Jie Chen, Yitong Yang, and Yong Zhao. High-frequency stereo matching network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1327-1336, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19710" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_content_list.json b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c4c936600bc2a913a4273202255070a67229e40 --- /dev/null +++ b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_content_list.json @@ -0,0 +1,1639 @@ +[ + { + "type": "text", + "text": "Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization", + "text_level": 1, + "bbox": [ + 130, + 128, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jimyeong Kim1 Jungwon Park1 Wonjong Rhee1,2,3 \nDepartment of Intelligence and Information1 & IPAI2 & RICS3, Seoul National University \n{wlaud1001, quoded97, wrhee}@snu.ac.kr", + "bbox": [ + 127, + 202, + 841, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In text-to-image personalization, a timely and crucial challenge is the tendency of generated images overfitting to the biases present in the reference images. We initiate our study with a comprehensive categorization of the biases into background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. These biases manifest in the generated images due to their entanglement into the subject embedding. This undesired embedding entanglement not only results in the reflection of biases from the reference images into the generated images but also notably diminishes the alignment of the generated images with the given generation prompt. To address this challenge, we propose SID (Selectively Informative Description), a text description strategy that deviates from the prevalent approach of only characterizing the subject's class identification. SID is generated utilizing multimodal GPT-4 and can be seamlessly integrated into optimization-based models. We present comprehensive experimental results along with analyses of cross-attention maps, subject-alignment, non-subject-disentanglement, and text-alignment.", + "bbox": [ + 75, + 324, + 473, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 661, + 209, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models [5, 7, 11, 25, 30, 35, 40, 42, 43, 46] have demonstrated remarkable generation capabilities that align with textual descriptions across various applications. One such application is text-to-image personalization [12, 24, 45], where models are tailored to generate novel renditions of subjects described in a few reference images. Recent works, such as DreamBooth [45] or Custom Diffusion [24], stand out for their exceptional personalized generation results. They achieve this by fine-tuning a pretrained text-to-image diffusion model with a small number of reference images using a text description in the format of \"a [v] [class name]\" or \"photo of a [v] [class name].\" In this context, [v] represents the subject's unique identifier, which uses a rare token with minimal semantic significance, while", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the class descriptor [class name] represents a coarse category for the subject.", + "bbox": [ + 498, + 292, + 890, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent developments in text-to-image personalization can be categorized into optimization-based approaches [12, 15, 24, 45, 57] and encoder-based approaches [21, 26, 50, 62]. For both types of approaches, a problem typically addressed as 'overfitting' has emerged [2, 15, 21, 26, 45, 57]. It is a phenomenon where objects, other than the subject of interest, in the reference images affect the generated image in an undesired way. For the sake of clarity, we refer to such an entity as 'non-subject' or 'undesired object'. The cause of this phenomenon is closely related to the embedding of the identifier token [v] for optimization-based approaches and the embedding of the subject for encoder-based models, where the information related to any non-subject seeps into the embeddings. Rather than referring to it as overfitting, we formally address this phenomenon as undesired embedding entanglement.", + "bbox": [ + 496, + 324, + 892, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The phenomenon of undesired embedding entanglement has been noticed, but it has been only superficially understood so far. A potential explanation for the limited understanding is the rarity of this phenomenon in straightforward scenarios. Instances where reference images encompass diverse subject views and minimal undesired objects are much less prone to this phenomenon. However, such an ideal scenario is highly unlikely in real-world applications. Therefore, it is crucial to comprehensively understand the biases and develop counter methods to address them effectively. In an effort to gain a more thorough understanding of the prevalent biases contributing to these entanglements, we conducted exhaustive paper reviews followed by extensive experiments utilizing state-of-the-art models. The resulting identification of the five most commonly encountered biases is presented in Fig. 1. While we have chosen DreamBooth [45] for generating the examples in Fig. 1, additional examples for Custom Diffusion [24], SVDiff [15], ELITE [62] and BLIP-Diffusion [26] will be provided later.", + "bbox": [ + 496, + 566, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In prior studies, several models [2, 21, 26, 50, 62] have employed segmentation masks. However, depending exclusively on segmentation masks comes with limitations, as we", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8312", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/742751729213fb74dcc71bfb43b08cf6534b513a477aa7acb8c96feade00a8d4.jpg", + "image_caption": [ + "Figure 1. Five key biases – background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. The first four rows depict scenarios with multiple reference images, while the last row illustrates a single reference image scenario. The pose bias is particularly prone to manifest in scenarios involving a single reference image, although it can also occur when multiple reference images depict the subject in poses that are similar. In the generation prompt, the subject of interest is highlighted in red. The integration of our method into DreamBooth [45] effectively resolves embedding entanglements associated with the five key biases (rightmost column)." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 854, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "will demonstrate with additional examples in subsequent sections. Notably, these limitations become apparent in the cases of substance, tied-object, and pose biases. To tackle all of the key biases, we introduce SID (Selectively Informative Description) that requires only modifications in the train text descriptions of the reference images. The idea is very simple. Starting with an optimization-based model, we diverge from the prevalent approach of only characterizing the subject's class identification as observed in the existing works. Our approach involves integrating informative specifications of the \"undesired objects\" into the train descriptions. This modification significantly diminishes the probability of undesired entanglement between the subject embedding [v] and the non-subject information in the reference images. As the text-to-image diffusion models are trained for alignments, adding text descriptions that match non-subjects can greatly help the non-subject parts in the reference images to align with the corresponding text descriptions. This prevents the undesired objects from acci", + "bbox": [ + 75, + 607, + 472, + 895 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dently becoming aligned with [v]. In contrast, such an accidental alignment can quite easily occur in the absence of the matching text descriptions. Our method is selective because we deliberately avoid incorporating informative specifications of the \"subject\" itself into the train descriptions. Such incorporation could negatively affect the preservation of the subject's identity, as will be elucidated in Sec. 3.1. SID can be readily integrated with any optimization-based model, such as DreamBooth [45], Custom Diffusion [24], or SVDiff [15]. For the automated generation of SID, we utilize the multimodal GPT-4 [37] by guiding it with an appropriate instruction. Surprisingly, the proposed SID integration significantly reduces embedding entanglement, especially in critically biased scenarios.", + "bbox": [ + 496, + 607, + 893, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 833, + 640, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image diffusion models. Diffusion models [19, 54, 55] have demonstrated remarkable generative capabilities across a wide range of modalities, including im", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8313", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ages [10, 35, 46], audios [8, 22, 28], videos [18, 32, 66], and 3D shapes [23, 68, 69]. Within this diverse spectrum, prominent text-to-image diffusion models, such as Stable Diffusion [43], Imagen [46], and DALL-E 2 [42], have exhibited exceptional proficiency in the synthesis of high-quality images closely aligned with textual descriptions. Subsequently, significant efforts [5, 7, 11, 25, 30, 40] have been dedicated to enhancing the text-to-image alignment in the field of text-to-image generation. Our study, based on the Stable Diffusion model [43], improves the alignment between textual descriptions and images in personalized image synthesis, focusing on accurately preserving subject identity without any entanglement with undesired objects.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language models. The vision-language model (VLM) handles tasks like text-to-image retrieval [61], visual question answering [1], and image captioning [65]. VLMs can also play a crucial role as a building block, as seen with CLIP [41] in tasks such as text-to-image generation [35, 42, 43], image segmentation [31], image retrieval [4], and more [34, 38, 51, 52]. In our work, we need an image captioning VLM capable of closely following detailed instructions. While models like BLIP-2 [27] and OpenFlamingo [3] offer image captioning with text conditioning, they often struggle to adhere closely to detailed instructions. In contrast, large models such as LLaVA [29] and multi-modal GPT-4 [37] have demonstrated their ability to generate precise captions closely aligned with detailed instructions. Among these options, multi-modal GPT-4 exhibits the most powerful capability in generating instruction-following captions.", + "bbox": [ + 75, + 291, + 472, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Personalized image synthesis. Personalizing text-to-image diffusion models [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62] has become a central focus in text-to-image generation. These models generate high-quality, diverse renditions of a subject using either a few reference images (usually 3 to 7) or just a single reference image. These models can be broadly categorized into two groups: optimization-based and encoder-based models. Optimization-based models [2, 12, 15, 24, 45, 57], like DreamBooth [45], involve per-subject optimization, where they fine-tune identifier tokens or segments of diffusion models to encode a subject's identity. On the other hand, encoder-based models [13, 21, 26, 50, 62] pre-train separate identity encoders and utilize them to encode a subject's identity without the need for per-subject optimization.", + "bbox": [ + 75, + 553, + 472, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 789, + 169, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed method can be readily integrated with any optimization-based models that utilize description of the reference images. The overall process, presented in Fig. 2, entails the generation of SID (Selectively Informative Description) for each reference image using an instruction-following vision-language model (VLM). These SIDs are", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c8ac0c54cb061e34445d85e41d129667892b555d87d97f6e1b77b4527655688.jpg", + "image_caption": [ + "Figure 2. Personalization with SID. We propose integrating SID (Selectively Informative Description) into the per-subject optimization, where an instruction-following VLM (Vison-Language Model) is utilized to generate a selectively informative description for each reference image." + ], + "image_footnote": [], + "bbox": [ + 539, + 88, + 854, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "integrated into the reconstruction learning process as part of the per-subject optimization. Depending on the underlying optimization-based model, additional loss functions may be incorporated together with the main reconstruction loss. The additional loss functions, such as the class prior preservation loss [45] introduced in DreamBooth, are applied in the same manner as in their respective baseline models without any modification. The effect of SID is addressed in Sec. 3.1. The choice of VLM and its accompanying instruction are addressed in Sec. 3.2.", + "bbox": [ + 496, + 353, + 893, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SID for reducing embedding entanglement", + "text_level": 1, + "bbox": [ + 498, + 517, + 864, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The existing text-to-image personalization approaches [12, 15, 24, 45], especially optimization-based methods like DreamBooth [45], often rely on simplistic textual descriptions that closely follow the format of \"a [v] [class name]\", where [v] denotes the unique identifier for the subject, in order to extract the subject's identity. In our method, our goal is to come up with train descriptions that have an appropriate level of information such that we can effectively isolate the subject of interest even when the reference images are biased. According to the findings in DreamBooth [45], specifying the subject's correct class in the train description is instrumental in accurately preserving the subject's identity. With the correct class information, the model's prior knowledge of the subject's class can significantly enhance the editing capabilities of the model. Based on the findings, we consider the case where the train description only contains the subject's class identification as the baseline and subsequently explore additionally informative cases that are listed in Tab. 1. In Tab. 1, the term undesired object generally encompasses any object within reference images, excluding the subject of interest. For the examples in Fig. 1, the undesired objects can be readily recognized for background, nearby-object, and tied-object biases. For the substance bias, that is applicable only when performing a style", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8314", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2fdc6611ce90fa031679e88d855bd141975442390481107201e5aed552909ee0.jpg", + "image_caption": [ + "Figure 3. Two examples for comparing the four cases of descriptions. For each case, the choice of train description follows the guidelines in Tab. 1. The common generation prompt for each example is shown below the generated images. Additional examples can be found in Appendix E.1." + ], + "image_footnote": [], + "bbox": [ + 143, + 90, + 305, + 439 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f8bbc6b5934b25badf2759903c2c465ce87fde9463ff35ae435662d0530a098.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 90, + 821, + 439 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0c01220a50720b5ab752e446f9b172801b00a1360f89f11e3fb3ac52db7db6d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DescriptionsSubjectUndesired objects
Class IdentificationInformative SpecificationsClass IdentificationInformative Specifications
Case 1 (Baseline)OXXX
Case 2OXOX
Case 3 (Ours)OXOO
Case 4OOOO
", + "bbox": [ + 78, + 503, + 470, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Four cases of train descriptions. The four cases are categorized based on whether they provide class identification only or additional informative specifications of the subject and undesired objects. We define $SID$ as the description of Case 3 (Ours).", + "bbox": [ + 75, + 599, + 468, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "re-contextualization, all the objects in the reference images are considered to be undesirable because the intention is to transfer the style only. For pose bias, it is inherent to the subject itself.", + "bbox": [ + 75, + 669, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the four cases in Tab. 1, two examples are investigated in Fig. 3 where two sets of the resulting images are shown. For the simplicity of this investigation, reference images were chosen to share the same set of undesired objects and thus allowing us to use a single train description for all reference images. Case 1 (Baseline) uses the format \"a [v] [class name]\", where only the subject's class identification is included in the train description. In this case, the model often struggles with the bias in the reference images (e.g., the subject can co-appear with undesired objects). This is because of the embedding entanglement in [v] as will be", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "analyze further in Sec. 5. To reduce the undesired embedding entanglement, we introduce Case 2 where the train description contains the class identifications of the undesired objects. Case 3 is similar to Case 2, but the train description additionally contain informative specifications of the undesired objects. As shown in Fig. 3, both Case 2 and Case 3 can provide a significant improvement over Case 1. Case 3, however, often outperforms Case 2 in terms of entanglement reduction as is the case for the two examples in Fig. 3.", + "bbox": [ + 496, + 507, + 890, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the subject itself, it is also possible to include informative specifications of the subject. Case 4 corresponds to this possibility. With the inclusion, however, the subject details described in the informative specifications are disentangled from the [v]. This occurs because of the alignment between the subject's informative specifications in the text and the subject details in the images. Because preserving the subject details in the [v] is crucial in personalized image synthesis, we have opted for Case 3 as our definition of SID.", + "bbox": [ + 496, + 641, + 892, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. VLM for generating SID", + "text_level": 1, + "bbox": [ + 498, + 789, + 725, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generating SID for each reference image can be demanding and time-consuming for humans. Therefore, we have chosen to utilize image captioning VLM for an automatic generation of SID. To identify a suitable VLM capable of effectively generating SIDs, we evaluated three well-known image captioning VLMs: BLIP-2 [27], LLaVA [29], and", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8315", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instruction for GPT-4 & LLaVA", + "text_level": 1, + "bbox": [ + 104, + 93, + 261, + 102 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"Describe the image in one sentence in detail. Please start", + "bbox": [ + 104, + 104, + 416, + 111 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with \"A {class}\". *You should not describe the distinct", + "bbox": [ + 104, + 111, + 395, + 118 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features of the {class} itself.\\*", + "bbox": [ + 104, + 119, + 281, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input text for BLIP-2", + "text_level": 1, + "bbox": [ + 106, + 132, + 209, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"A {class}\"", + "bbox": [ + 106, + 142, + 163, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c7e53139580e932fcb5376e4e9773760f95fd09aa77a2bf9be83ffb0abe5ac7.jpg", + "image_caption": [ + "Figure 4. Comparison of three instruction-following VLMs for generating SID. For the reference image of a cat and the instructions shown in the top, the three VLMs generate image captions shown in the right side of the image. Subsequently, the unique identifier [v] is inserted and the resulting captions are used for conditioning the diffusion model as the train descriptions. For painting/cartoon style re-contextualization, we used a slightly different instruction, as detailed in Appendix C.1. Additional fifteen examples for VLM-generated SIDs are available in Appendix E.2." + ], + "image_footnote": [], + "bbox": [ + 99, + 152, + 217, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal GPT-4", + "text_level": 1, + "bbox": [ + 222, + 154, + 315, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A cat is perched on a tiled surface with a metal railing", + "bbox": [ + 222, + 165, + 437, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "beside it and a backdrop of stairs.", + "bbox": [ + 222, + 172, + 356, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaVA", + "text_level": 1, + "bbox": [ + 222, + 189, + 263, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A small brown and white cat is sitting on a stair railing.", + "bbox": [ + 222, + 200, + 441, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "looking down at the floor.", + "bbox": [ + 222, + 208, + 326, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BLIP-2", + "text_level": 1, + "bbox": [ + 222, + 224, + 261, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A cat is sitting on the stairs looking up.", + "bbox": [ + 222, + 234, + 377, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "multi-modal GPT-4 [37]. In Fig. 4, we illustrate a caption generation example in which the subject is explicitly defined as a cat within the image. For BLIP-2, it struggles to comprehend detailed instructions and it can often produce sequences of meaningless characters. Therefore, we exclusively conditioned BLIP-2 for the subject of interest while providing the detailed instructions to both LLaVA and GPT-4 for generating SID.", + "bbox": [ + 75, + 388, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Captions generated by BLIP-2 significantly fall short of the desired SID, lacking informative specifications, i.e., detailed features, of any undesired objects (objects other than the cat in Fig. 4 example) in the images. Although LLaVA generates descriptions better aligned with the instructions, it still falls substantially short, tending to describe the subject excessively or fail to provide sufficient details about undesired objects. Compared to the other two, multi-modal GPT-4 stands out for its ability to follow instructions, delivering descriptions that closely align with the concept of SID. The quantitative comparisons between these VLMs are shown in Appendix E.2, where multi-modal GPT-4 shows the best subject and text alignment. Therefore, we have chosen multi-modal GPT-4 as the VLM for SID generation.", + "bbox": [ + 75, + 510, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have performed comprehensive experiments to verify the enhancement resulting from SID. For the experiment datasets, images from previous works [12, 24, 45] and three websites [39, 59, 63] were examined. As the main experiment, we have integrated SID with four optimization-based models - DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion (TI) [12]. The superior effectiveness of SID is readily apparent across a broad spectrum of examples as shown in Fig. 5. In the presence of a single reference image only, we have also compared", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SID-integrated DreamBooth (Ours) with two encoder-based models, ELITE [62] and BLIP-Diffusion [26], along with DreamBooth in Fig. 6. Implementation details for each model are provided in Appendix C.2.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Analysis of cross-attention map", + "text_level": 1, + "bbox": [ + 500, + 157, + 785, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cross-attention map of each text token is known to highlight the pixels the text token describes [17, 56]. We visualized averaged cross-attention maps for the identifier [v], which is trained to encode the subject's identity. The cross-attention maps are averaged over all the timestamps, layers, and heads to show the single image that can be overlayed on the generated image, following the previous convention [56]. In Fig. 7, we compared cross-attention maps for DreamBooth and SID-integrated DreamBooth across four distinct biases: background, nearby-object, tied-object, and substance biases. While DreamBooth's identifier tends to erroneously focus on undesired objects across all biased scenarios, our method consistently displays highly accurate attention focused on the subject. To extend this analysis, we conducted a comparison between the four cases of train descriptions (detailed in Fig. 3, top row) in Fig. E.2. These analyses confirm that integrating SID does reduce undesired embedding entanglements.", + "bbox": [ + 496, + 176, + 890, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Analysis of three key measures", + "text_level": 1, + "bbox": [ + 500, + 459, + 781, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-alignment and text-alignment have become popular measures in recent personalized image synthesis [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62]. However, we have found that the widely used image-alignment score tends to be significantly influenced by background entanglement, making it an inappropriate measure for analyzing undesired embedding entanglements. Therefore, we first define two customized measures, subject-alignment and non-subject-disentanglement, by adapting image-alignment specifically for text-to-image personalization. Further details can be found in Appendix A, where we demonstrate the crucial importance of these two measures in assessing subject preservation and reduction in undesired embedding entanglement. Using the two measures, along with text-alignment, we conduct a quantitative analysis of SID.", + "bbox": [ + 496, + 479, + 890, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Customizing image-alignment measure. Consider a set of reference images $R = \\{r_1, r_2, \\dots, r_N\\}$ and a set of generated images $G = \\{g_1, g_2, \\dots, g_M\\}$ . Each image in $G$ is generated using $R$ as the reference images with a common generation prompt $p$ . Define $f_i: \\mathcal{X}_{image} \\to \\mathbb{R}^D$ and $f_t: \\mathcal{X}_{text} \\to \\mathbb{R}^D$ as the CLIP image and text encoders [41] followed by a unit-norm normalization, respectively. Also, define a subject segmentation function $s: \\mathcal{X}_{image} \\to \\{0, 1\\}^{H \\times W}$ , implemented as Grounded-SAM [14] conditioned on the subject's class name. A mask over the image resolution space is output by $s$ where 1 indicates the subject's pixel and 0 indicates the non-subject's pixel.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8316", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/29a71ef2e2182c12642957e543448e7a9b3731e88cbaf3a4795946e68101b429.jpg", + "image_caption": [ + "Figure 5. Enhancement by SID. For four optimization-based models (DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion [12]), the baseline results are shown together with SID-integrated results. SID-integration effectively resolves entanglement issues in scenarios with high biases, represented by indoor background (1st row), nearby potted plant (2nd row), filled-in blueberries (3rd row), and sunflower substances (last row). Additional examples can be found in Appendix E.4." + ], + "image_footnote": [], + "bbox": [ + 151, + 88, + 808, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71e3fbc91c896c74fc1f0bd35be6130f4a47e4a5d52f18288a77db2d156182b8.jpg", + "image_caption": [ + "Figure 6. Enhancement by SID for a single reference image. Two encoder-based models, ELITE [62] and BLIP-Diffusion [26], are compared with DreamBooth [45] and our SID-integrated DreamBooth. Notably, in the second row, our approach effectively addresses pose bias. Additional examples can be found in Appendix E.5." + ], + "image_footnote": [], + "bbox": [ + 122, + 609, + 436, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The subject-alignment $\\mathcal{SA}$ is defined as the average pairwise cosine similarity between generated images and subject segments of reference images in the CLIP-embedding space. The non-subject-disentanglement $\\mathcal{NSD}$ is defined as the 1 minus the non-subject segment similarity in CLIP-embedding space. They can be formally expressed as:", + "bbox": [ + 498, + 609, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S A} = \\quad A v g \\left(f _ {i} \\left(r _ {n} \\odot s \\left(r _ {n}\\right)\\right) ^ {\\intercal} \\cdot f _ {i} \\left(g _ {m}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 715, + 890, + 732 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N S D} = 1 - \\operatorname {A v g} \\left(f _ {i} \\left(r _ {n} \\odot (1 - s (r _ {n}))\\right) ^ {\\top} \\cdot f _ {i} \\left(g _ {m}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 734, + 890, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Avg(\\cdot)$ denotes the average operation over all possible pair selections between $n = 1, \\ldots, N$ and $m = 1, \\ldots, M$ . Additionally, we evaluate text-alignment $\\mathcal{T}\\mathcal{A}$ shown below.", + "bbox": [ + 498, + 766, + 893, + 825 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\mathcal {A} = A v g \\left(f _ {t} (p) ^ {\\mathrm {T}} \\cdot f _ {i} \\left(g _ {m}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 840, + 890, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The identifier [v] is removed from the generation prompt $p$ when evaluating $\\mathcal{T}\\mathcal{A}$ .", + "bbox": [ + 498, + 875, + 890, + 905 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8317", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e15a669baf562ad3d9ace55fac4d7bdd29d3fba0f9b85ce28738060f048bfe8.jpg", + "image_caption": [ + "Figure 7. Cross-attention map. Averaged cross-attention maps for the identifier [v] are shown. They are overlayed on the generated images. The undesirably entangled embeddings of DreamBooth [45] can be confirmed from the erroneous focusing on the undesirable objects. On the contrary, SID-integrated DreamBooth accurately highlights the subjects of interest: dog, perfume, girl, and painting style. Image credit: David Revoy (3rd row)" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 459, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of SID. The three measures have straightforward interpretations. A higher subject-alignment indicates superior preservation of the subject's identity. Elevated non-subject-disentanglement values signify a reduction in undesired embedding entanglements. Increased text-alignment suggests that the generated images align more closely with the generation prompt. Using the three measures, we have evaluated the three optimization-based models and their SID-integrated counterparts as shown in Fig. 8. For the single reference images, we evaluated ELITE [62] and BLIP-Diffusion [26] together with the three optimization-based models and the results are shown in Fig. 9. From the figures, it can be observed that all of the best performing, or Pareto optimal, models are the SID-integrated models. The average improvements of the three key metrics exhibit positive values with a substantial margin in all cases, except for subject alignment in Fig. 9. A potential explanation for this observation is the occurrence of subject overfitting in the existing models, particularly evident in cases such as pose bias. The implementation of SID mitigates subject overfitting, even when only a single reference image is provided. For the evaluation, we have utilized commonly used personalization datasets [12, 24, 45] to generate 7500 images. The details are articulated in Appendix C.4.", + "bbox": [ + 75, + 488, + 472, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human evaluation. We further assess the effectiveness of SID over its baseline by conducting human evaluation. Human evaluation was performed involving 130 partici", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74159a49e297b735dd6f6f8e0597b4e36fd351fa1a18050b75ddbaa8ccb98660.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 692, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a166aa1d6b163a53322304558c161bcb5c28e9aad34025b10c74e041d9e3b7b.jpg", + "image_caption": [ + "Figure 8. Pair-wise metric visualization for multiple reference images. The best performing ones are the SID-integrated models and they form the Pareto boundary." + ], + "image_footnote": [], + "bbox": [ + 544, + 227, + 859, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3856ebb230db49d24dee7641c029e4bd476c9b1723f2011f1d808a70f4521cd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 90, + 875, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c41e3ffff44f26f064264c59be439782923402cbb0b1d45142fcdfa63af8a71.jpg", + "image_caption": [ + "Figure 9. Pair-wise metric visualization for single reference image. The best performing ones are the SID-integrated optimization-based models and they form the Pareto boundary. They also outperform the two encoder-based models." + ], + "image_footnote": [], + "bbox": [ + 509, + 306, + 692, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21dd72467e307712903dc2c1f60d676c989f6132aed7cac1868d80ff6bf3f6c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 445, + 859, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9353a6e4fbcf1a58a6c7a16b4d0926cded8b2cf43d01ea9557c2b073a5557eea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 308, + 875, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pants, 10 subjects per participant, and 4 questions per subject (a total of 5,200 responses). Tab. B.2 shows that the survey results for subject-alignment, non-subject disentanglement, and text-alignment are consistent with the metric analysis in Fig. 8, supporting our metric analysis. The results and additional details are provided in Appendix B", + "bbox": [ + 496, + 530, + 890, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 500, + 638, + 612, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1. Negative prompt and segmentation", + "text_level": 1, + "bbox": [ + 500, + 664, + 803, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Negative prompt for classifier-free guidance [20] and segmentation can be considered as the alternatives of SID. Negative prompt is typically used to prevent the generation of unwanted features [49, 58, 64]. Segmentation mask has been used in several previous works for the purpose of isolating subject [2, 21, 26, 50, 62]. Instead of adopting SID, negative prompt can be used during inference or the subject's segmentation mask can be applied before per-subject optimization. In Fig. 10, SID-integration is compared to these two approaches. Negative prompts exhibit unsatisfactory disentanglement, suggesting that severely entangled representations cannot be easily unraveled during inference. Segmentation masks also exhibit its own limitations, and face challenges in dynamically editing the pose", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8318", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ffc763920d69243daeae35b6e31bc79c977ecaf48d607567888e90e9999d4a2.jpg", + "image_caption": [ + "Reference images", + "dog" + ], + "image_footnote": [], + "bbox": [ + 99, + 106, + 209, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a483685c79bdf40621fb8ff4006c01affa0c2117355e65573d8aee3fa0f28ae.jpg", + "image_caption": [ + "Negative prompt", + "A [v] dog sitting in front of the Eiffel Tower." + ], + "image_footnote": [], + "bbox": [ + 230, + 109, + 379, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8db23fd8229ccf40eb8ae5bde0af3f73cc0f211aa57ebb76b732376a201a6e70.jpg", + "image_caption": [ + "SID", + "t of the Eiffel Tower." + ], + "image_footnote": [], + "bbox": [ + 390, + 90, + 537, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb4ad1169d22d74e3890604c713e7476b55d1a4e9184335f883bbfd02c52c5a8.jpg", + "image_caption": [ + "Segmentation", + "A [v] dog howling at the moon on a rocky, barren landscape." + ], + "image_footnote": [], + "bbox": [ + 558, + 109, + 707, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f17f27c812c3a173a35581692984016aab0c9d03a214480e833edc682ba72832.jpg", + "image_caption": [ + "SID" + ], + "image_footnote": [], + "bbox": [ + 718, + 109, + 864, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bbdd96fdfd3f9532ec9104d039eaf6556f6dff9d3ac5cc4b2d896aafed791c5.jpg", + "image_caption": [ + "perfume", + "Figure 10. Comparison with negative prompt and segmentation. DreamBooth [45] is used as the base model. Compared to the two alternatives, SID-integration demonstrates superior non-subject-disentanglement and text-alignment. Negative prompts used: “sitting on the fluffy blanket and green couch” in the upper row example, “next to a black Prada purse with silver hardware on a bed of white sheets” in the lower row example. Additional examples can be found in Appendix E.6" + ], + "image_footnote": [], + "bbox": [ + 99, + 210, + 207, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1906ca109883cf2b4df74b69bd5a387a73be2ef3bc6de583217c68f387ac66c1.jpg", + "image_caption": [ + "A [v] perfume on a park bench." + ], + "image_footnote": [], + "bbox": [ + 230, + 215, + 379, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff6c1cf1ddd15c88d66f485ccafcf3a376fd94f5a6705d34a63633b9a09e54ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 215, + 539, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c5109f67ed28acf180985a4fe20d75795baad51048944ecd680774695019b95.jpg", + "image_caption": [ + "A [v] perfume on a windowsill of an old cottage, overlooking a meadow." + ], + "image_footnote": [], + "bbox": [ + 560, + 215, + 707, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab5f3f350627f168133db623865ac9d6db6a695b26e6f4788a9567a8a4df7c34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 215, + 864, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/030827ed414845371adad59f08cbf421836c0e863123bc1eb3add85c7923eb0f.jpg", + "image_caption": [ + "Reference images" + ], + "image_footnote": [], + "bbox": [ + 80, + 404, + 151, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ab1907664b6b52eac2f8efa6eb11d5c24b59d9b50bb6b1e3e3965277fca237d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 405, + 225, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe6f37e8b8e2e2376dadbb669012d8b4fe3186b71f1d8e52dd221f8871f44dfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 405, + 299, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f79ac9163b52eb4371c483fcbcf98d3c815837de3e7c7a146203dbf8187a974f.jpg", + "image_caption": [ + "Figure 11. Subject editing. Two examples are shown for editing subject's color. DreamBooth [45] erroneously changes the color of non-subject elements (sofa in the upper row example, background in the lower row example), while SID-integrated DreamBooth accurately edits the subject (dog)." + ], + "image_footnote": [], + "bbox": [ + 80, + 462, + 150, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2fece19a93670996d598e228c477fb785366186b9382e4760fa00ab6b230445b.jpg", + "image_caption": [ + "dog" + ], + "image_footnote": [], + "bbox": [ + 153, + 462, + 225, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30709109dbeb4a2149ab014b19331b0ed9c7a27e7b86206d6e1c4ae1db241686.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 462, + 299, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e94f49cbbc58a9e23bca4e7794a0f3ef936d67e511401a667626480906162ab6.jpg", + "image_caption": [ + "DreamBooth" + ], + "image_footnote": [], + "bbox": [ + 318, + 402, + 390, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51b4cba7f5f16f996be2c3f65e356adfdec4c60f390e74671e0a7353ec491ae8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 462, + 390, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95e92908269282786772cc715f14ef776bda051f39305502f8ea8720825461e8.jpg", + "image_caption": [ + "DreamBooth + SID" + ], + "image_footnote": [], + "bbox": [ + 395, + 404, + 467, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a07a0243035f6f47c7c2723e8cbba4f6f2c9c7721b70d46319113ddd7997ab70.jpg", + "image_caption": [ + "A purple [v] dog" + ], + "image_footnote": [], + "bbox": [ + 395, + 462, + 467, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the subject or faithfully following the generation prompt. SID shows higher non-subject-disentanglement and text-alignment than the two alternatives.", + "bbox": [ + 75, + 626, + 467, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. Enhancing subject editing", + "text_level": 1, + "bbox": [ + 76, + 676, + 313, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The identifier [v], specifically trained to encode the subject's identity, enables us to modify the properties of the subject through a textual guidance. Undesired embedding entanglements, however, can lead to misaligned modifications. In Fig. 11, an example of subject's color editing is shown. DreamBooth [45], influenced by embedding entanglements, fails to edit the subject and instead modifies the color of an undesired object or background. In contrast, SID-integrated DreamBooth precisely modifies the subject itself, following the provided color guidance.", + "bbox": [ + 75, + 694, + 468, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.3. Limitations of SID", + "text_level": 1, + "bbox": [ + 76, + 849, + 256, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A limitation of our approach is associated with the imperfections of VLM. While the multi-modal GPT-4 [37] gen", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "erally generates descriptions that closely align with the provided instructions, occasional failures do occur. Although these instances are rare and may not be easily generalized, we have listed some of these failure cases in Appendix D.1. Another limitation is the undesired entanglement of the subject's strong facial expressions. We discovered that this failure arises when the VLM-generated SID lacks informative specifications of facial expressions. This issue can be addressed by including undesired information in the SID, a solution detailed further in Appendix D.2. Finally, SID is exclusively integrated with optimization-based models in this study. Yet, there is potential for its integration into encoder-based models, particularly during encoder pre-training.", + "bbox": [ + 496, + 390, + 890, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 602, + 617, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we have introduced a robust strategy to mitigate undesired embedding entanglement in text-to-image personalization. Beginning with the identification of five key biases, we have proposed SID (Selectively Informative Description) as an effective solution to address these biases. Our cross-attention analysis demonstrates the successful removal of entanglements, while alignment analysis indicates notable enhancements in non-subject-disentanglement and text-alignment. The proposed SID strategy holds potential applicability to other multi-modal applications where managing embedding entanglements is crucial.", + "bbox": [ + 496, + 626, + 890, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the following grants funded by the Korea government: NRF (NRF2020R1A2C2007139, NRF-2022R1A6A1A03063039) and IITP ([NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], [No. RS2023-00235293]).", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8319", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015.", + "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023.", + "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023.", + "[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022.", + "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023.", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021.", + "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023.", + "[8] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020.", + "[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023.", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021.", + "[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022.", + "[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.", + "[13] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023.", + "[14] Grounded-SAM Contributors. Grounded-Segment-" + ], + "bbox": [ + 78, + 114, + 470, + 887 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Anything. https://github.com/IDEA-Research/Grounded-Segment-Anything，Apr.2023.", + "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023.", + "[16] Haotian Liu. liuhaotian/llava-v1-0719-336px-lora-merge- vicuna-13b-v1.3. https://huggingface.co/ liuhaotian/llava-v1-0719-336px-lora- merge-vicuna-13b-v1.3, 2023.", + "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.", + "[18] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022.", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020.", + "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022.", + "[21] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023.", + "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020.", + "[23] Juil Koo, Seungwoo Yoo, Minh Hieu Nguyen, and Minhyuk Sung. Salad: Part-level latent diffusion for 3d shape generation and manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14441-14451, 2023.", + "[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023.", + "[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023.", + "[26] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023.", + "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023.", + "[28] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audi-Oldm: Text-to-audio generation with latent diffusion models." + ], + "bbox": [ + 503, + 93, + 890, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8320", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "arXiv preprint arXiv:2301.12503, 2023.", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023.", + "[30] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022.", + "[31] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022.", + "[32] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023.", + "[33] Makoto Shing. svdiff-pytorch. https://github.com/mkshing/svdiff-pytorch, 2023.", + "[34] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022.", + "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.", + "[36] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023.", + "[37] OpenAI. Gpt-4 technical report, 2023.", + "[38] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021.", + "[39] Peper and Carrot. https://www.peppercarrot.com/.", + "[40] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427, 2023.", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.", + "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022.", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen" + ], + "bbox": [ + 78, + 90, + 468, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015.", + "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023.", + "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022.", + "[47] Salesforce. LAVIS. https://github.com/salesforce/LAVIS/tree/main/projects/blip-diffusion, 2023.", + "[48] Salesforce. Salesforce/blip2-opt-2.7b. https://huggingface.co/Salesforce/blip2-opt-2.7b, 2023.", + "[49] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023.", + "[50] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023.", + "[51] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022.", + "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.", + "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.", + "[54] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019.", + "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.", + "[56] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022.", + "[57] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023.", + "[58] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven" + ], + "bbox": [ + 503, + 92, + 890, + 891 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8321", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023.", + "[59] Unsplash. https://unsplash.com/.", + "[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models.", + "[61] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5764-5773, 2019.", + "[62] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023.", + "[63] Wikiart. https://www.wikiart.org/.", + "[64] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023.", + "[65] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015.", + "[66] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022.", + "[67] Yuxiang Wei. ELITE. https://github.com/csyxwei/ELITE, 2023.", + "[68] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022.", + "[69] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 646 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8322", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_model.json b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..050147c1075627a4ad6d0563875e92f2cc67195d --- /dev/null +++ b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.13, + 0.842, + 0.177 + ], + "angle": 0, + "content": "Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.203, + 0.842, + 0.258 + ], + "angle": 0, + "content": "Jimyeong Kim1 Jungwon Park1 Wonjong Rhee1,2,3 \nDepartment of Intelligence and Information1 & IPAI2 & RICS3, Seoul National University \n{wlaud1001, quoded97, wrhee}@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.629 + ], + "angle": 0, + "content": "In text-to-image personalization, a timely and crucial challenge is the tendency of generated images overfitting to the biases present in the reference images. We initiate our study with a comprehensive categorization of the biases into background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. These biases manifest in the generated images due to their entanglement into the subject embedding. This undesired embedding entanglement not only results in the reflection of biases from the reference images into the generated images but also notably diminishes the alignment of the generated images with the given generation prompt. To address this challenge, we propose SID (Selectively Informative Description), a text description strategy that deviates from the prevalent approach of only characterizing the subject's class identification. SID is generated utilizing multimodal GPT-4 and can be seamlessly integrated into optimization-based models. We present comprehensive experimental results along with analyses of cross-attention maps, subject-alignment, non-subject-disentanglement, and text-alignment." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.662, + 0.21, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Text-to-image diffusion models [5, 7, 11, 25, 30, 35, 40, 42, 43, 46] have demonstrated remarkable generation capabilities that align with textual descriptions across various applications. One such application is text-to-image personalization [12, 24, 45], where models are tailored to generate novel renditions of subjects described in a few reference images. Recent works, such as DreamBooth [45] or Custom Diffusion [24], stand out for their exceptional personalized generation results. They achieve this by fine-tuning a pretrained text-to-image diffusion model with a small number of reference images using a text description in the format of \"a [v] [class name]\" or \"photo of a [v] [class name].\" In this context, [v] represents the subject's unique identifier, which uses a rare token with minimal semantic significance, while" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.293, + 0.892, + 0.323 + ], + "angle": 0, + "content": "the class descriptor [class name] represents a coarse category for the subject." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.893, + 0.567 + ], + "angle": 0, + "content": "Recent developments in text-to-image personalization can be categorized into optimization-based approaches [12, 15, 24, 45, 57] and encoder-based approaches [21, 26, 50, 62]. For both types of approaches, a problem typically addressed as 'overfitting' has emerged [2, 15, 21, 26, 45, 57]. It is a phenomenon where objects, other than the subject of interest, in the reference images affect the generated image in an undesired way. For the sake of clarity, we refer to such an entity as 'non-subject' or 'undesired object'. The cause of this phenomenon is closely related to the embedding of the identifier token [v] for optimization-based approaches and the embedding of the subject for encoder-based models, where the information related to any non-subject seeps into the embeddings. Rather than referring to it as overfitting, we formally address this phenomenon as undesired embedding entanglement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.854 + ], + "angle": 0, + "content": "The phenomenon of undesired embedding entanglement has been noticed, but it has been only superficially understood so far. A potential explanation for the limited understanding is the rarity of this phenomenon in straightforward scenarios. Instances where reference images encompass diverse subject views and minimal undesired objects are much less prone to this phenomenon. However, such an ideal scenario is highly unlikely in real-world applications. Therefore, it is crucial to comprehensively understand the biases and develop counter methods to address them effectively. In an effort to gain a more thorough understanding of the prevalent biases contributing to these entanglements, we conducted exhaustive paper reviews followed by extensive experiments utilizing state-of-the-art models. The resulting identification of the five most commonly encountered biases is presented in Fig. 1. While we have chosen DreamBooth [45] for generating the examples in Fig. 1, additional examples for Custom Diffusion [24], SVDiff [15], ELITE [62] and BLIP-Diffusion [26] will be provided later." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In prior studies, several models [2, 21, 26, 50, 62] have employed segmentation masks. However, depending exclusively on segmentation masks comes with limitations, as we" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8312" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.855, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.529, + 0.896, + 0.601 + ], + "angle": 0, + "content": "Figure 1. Five key biases – background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. The first four rows depict scenarios with multiple reference images, while the last row illustrates a single reference image scenario. The pose bias is particularly prone to manifest in scenarios involving a single reference image, although it can also occur when multiple reference images depict the subject in poses that are similar. In the generation prompt, the subject of interest is highlighted in red. The integration of our method into DreamBooth [45] effectively resolves embedding entanglements associated with the five key biases (rightmost column)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.473, + 0.896 + ], + "angle": 0, + "content": "will demonstrate with additional examples in subsequent sections. Notably, these limitations become apparent in the cases of substance, tied-object, and pose biases. To tackle all of the key biases, we introduce SID (Selectively Informative Description) that requires only modifications in the train text descriptions of the reference images. The idea is very simple. Starting with an optimization-based model, we diverge from the prevalent approach of only characterizing the subject's class identification as observed in the existing works. Our approach involves integrating informative specifications of the \"undesired objects\" into the train descriptions. This modification significantly diminishes the probability of undesired entanglement between the subject embedding [v] and the non-subject information in the reference images. As the text-to-image diffusion models are trained for alignments, adding text descriptions that match non-subjects can greatly help the non-subject parts in the reference images to align with the corresponding text descriptions. This prevents the undesired objects from acci" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.895, + 0.82 + ], + "angle": 0, + "content": "dently becoming aligned with [v]. In contrast, such an accidental alignment can quite easily occur in the absence of the matching text descriptions. Our method is selective because we deliberately avoid incorporating informative specifications of the \"subject\" itself into the train descriptions. Such incorporation could negatively affect the preservation of the subject's identity, as will be elucidated in Sec. 3.1. SID can be readily integrated with any optimization-based model, such as DreamBooth [45], Custom Diffusion [24], or SVDiff [15]. For the automated generation of SID, we utilize the multimodal GPT-4 [37] by guiding it with an appropriate instruction. Surprisingly, the proposed SID integration significantly reduces embedding entanglement, especially in critically biased scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.834, + 0.642, + 0.85 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Text-to-image diffusion models. Diffusion models [19, 54, 55] have demonstrated remarkable generative capabilities across a wide range of modalities, including im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8313" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.288 + ], + "angle": 0, + "content": "ages [10, 35, 46], audios [8, 22, 28], videos [18, 32, 66], and 3D shapes [23, 68, 69]. Within this diverse spectrum, prominent text-to-image diffusion models, such as Stable Diffusion [43], Imagen [46], and DALL-E 2 [42], have exhibited exceptional proficiency in the synthesis of high-quality images closely aligned with textual descriptions. Subsequently, significant efforts [5, 7, 11, 25, 30, 40] have been dedicated to enhancing the text-to-image alignment in the field of text-to-image generation. Our study, based on the Stable Diffusion model [43], improves the alignment between textual descriptions and images in personalized image synthesis, focusing on accurately preserving subject identity without any entanglement with undesired objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.55 + ], + "angle": 0, + "content": "Vision-language models. The vision-language model (VLM) handles tasks like text-to-image retrieval [61], visual question answering [1], and image captioning [65]. VLMs can also play a crucial role as a building block, as seen with CLIP [41] in tasks such as text-to-image generation [35, 42, 43], image segmentation [31], image retrieval [4], and more [34, 38, 51, 52]. In our work, we need an image captioning VLM capable of closely following detailed instructions. While models like BLIP-2 [27] and OpenFlamingo [3] offer image captioning with text conditioning, they often struggle to adhere closely to detailed instructions. In contrast, large models such as LLaVA [29] and multi-modal GPT-4 [37] have demonstrated their ability to generate precise captions closely aligned with detailed instructions. Among these options, multi-modal GPT-4 exhibits the most powerful capability in generating instruction-following captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.782 + ], + "angle": 0, + "content": "Personalized image synthesis. Personalizing text-to-image diffusion models [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62] has become a central focus in text-to-image generation. These models generate high-quality, diverse renditions of a subject using either a few reference images (usually 3 to 7) or just a single reference image. These models can be broadly categorized into two groups: optimization-based and encoder-based models. Optimization-based models [2, 12, 15, 24, 45, 57], like DreamBooth [45], involve per-subject optimization, where they fine-tune identifier tokens or segments of diffusion models to encode a subject's identity. On the other hand, encoder-based models [13, 21, 26, 50, 62] pre-train separate identity encoders and utilize them to encode a subject's identity without the need for per-subject optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.17, + 0.805 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Our proposed method can be readily integrated with any optimization-based models that utilize description of the reference images. The overall process, presented in Fig. 2, entails the generation of SID (Selectively Informative Description) for each reference image using an instruction-following vision-language model (VLM). These SIDs are" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.089, + 0.855, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.277, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Personalization with SID. We propose integrating SID (Selectively Informative Description) into the per-subject optimization, where an instruction-following VLM (Vison-Language Model) is utilized to generate a selectively informative description for each reference image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.895, + 0.506 + ], + "angle": 0, + "content": "integrated into the reconstruction learning process as part of the per-subject optimization. Depending on the underlying optimization-based model, additional loss functions may be incorporated together with the main reconstruction loss. The additional loss functions, such as the class prior preservation loss [45] introduced in DreamBooth, are applied in the same manner as in their respective baseline models without any modification. The effect of SID is addressed in Sec. 3.1. The choice of VLM and its accompanying instruction are addressed in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.518, + 0.866, + 0.535 + ], + "angle": 0, + "content": "3.1. SID for reducing embedding entanglement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The existing text-to-image personalization approaches [12, 15, 24, 45], especially optimization-based methods like DreamBooth [45], often rely on simplistic textual descriptions that closely follow the format of \"a [v] [class name]\", where [v] denotes the unique identifier for the subject, in order to extract the subject's identity. In our method, our goal is to come up with train descriptions that have an appropriate level of information such that we can effectively isolate the subject of interest even when the reference images are biased. According to the findings in DreamBooth [45], specifying the subject's correct class in the train description is instrumental in accurately preserving the subject's identity. With the correct class information, the model's prior knowledge of the subject's class can significantly enhance the editing capabilities of the model. Based on the findings, we consider the case where the train description only contains the subject's class identification as the baseline and subsequently explore additionally informative cases that are listed in Tab. 1. In Tab. 1, the term undesired object generally encompasses any object within reference images, excluding the subject of interest. For the examples in Fig. 1, the undesired objects can be readily recognized for background, nearby-object, and tied-object biases. For the substance bias, that is applicable only when performing a style" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8314" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.092, + 0.307, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.092, + 0.822, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.453, + 0.895, + 0.496 + ], + "angle": 0, + "content": "Figure 3. Two examples for comparing the four cases of descriptions. For each case, the choice of train description follows the guidelines in Tab. 1. The common generation prompt for each example is shown below the generated images. Additional examples can be found in Appendix E.1." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.504, + 0.472, + 0.593 + ], + "angle": 0, + "content": "
DescriptionsSubjectUndesired objects
Class IdentificationInformative SpecificationsClass IdentificationInformative Specifications
Case 1 (Baseline)OXXX
Case 2OXOX
Case 3 (Ours)OXOO
Case 4OOOO
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Table 1. Four cases of train descriptions. The four cases are categorized based on whether they provide class identification only or additional informative specifications of the subject and undesired objects. We define \\(SID\\) as the description of Case 3 (Ours)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.731 + ], + "angle": 0, + "content": "re-contextualization, all the objects in the reference images are considered to be undesirable because the intention is to transfer the style only. For pose bias, it is inherent to the subject itself." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For the four cases in Tab. 1, two examples are investigated in Fig. 3 where two sets of the resulting images are shown. For the simplicity of this investigation, reference images were chosen to share the same set of undesired objects and thus allowing us to use a single train description for all reference images. Case 1 (Baseline) uses the format \"a [v] [class name]\", where only the subject's class identification is included in the train description. In this case, the model often struggles with the bias in the reference images (e.g., the subject can co-appear with undesired objects). This is because of the embedding entanglement in [v] as will be" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.642 + ], + "angle": 0, + "content": "analyze further in Sec. 5. To reduce the undesired embedding entanglement, we introduce Case 2 where the train description contains the class identifications of the undesired objects. Case 3 is similar to Case 2, but the train description additionally contain informative specifications of the undesired objects. As shown in Fig. 3, both Case 2 and Case 3 can provide a significant improvement over Case 1. Case 3, however, often outperforms Case 2 in terms of entanglement reduction as is the case for the two examples in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.893, + 0.78 + ], + "angle": 0, + "content": "For the subject itself, it is also possible to include informative specifications of the subject. Case 4 corresponds to this possibility. With the inclusion, however, the subject details described in the informative specifications are disentangled from the [v]. This occurs because of the alignment between the subject's informative specifications in the text and the subject details in the images. Because preserving the subject details in the [v] is crucial in personalized image synthesis, we have opted for Case 3 as our definition of SID." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.79, + 0.727, + 0.807 + ], + "angle": 0, + "content": "3.2. VLM for generating SID" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Generating SID for each reference image can be demanding and time-consuming for humans. Therefore, we have chosen to utilize image captioning VLM for an automatic generation of SID. To identify a suitable VLM capable of effectively generating SIDs, we evaluated three well-known image captioning VLMs: BLIP-2 [27], LLaVA [29], and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8315" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.106, + 0.094, + 0.262, + 0.103 + ], + "angle": 0, + "content": "Instruction for GPT-4 & LLaVA" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.105, + 0.418, + 0.112 + ], + "angle": 0, + "content": "\"Describe the image in one sentence in detail. Please start" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.112, + 0.397, + 0.119 + ], + "angle": 0, + "content": "with \"A {class}\". *You should not describe the distinct" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.12, + 0.282, + 0.128 + ], + "angle": 0, + "content": "features of the {class} itself.\\*" + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.133, + 0.21, + 0.142 + ], + "angle": 0, + "content": "Input text for BLIP-2" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.143, + 0.165, + 0.151 + ], + "angle": 0, + "content": "\"A {class}\"" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.153, + 0.218, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.155, + 0.316, + 0.165 + ], + "angle": 0, + "content": "Multimodal GPT-4" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.166, + 0.439, + 0.174 + ], + "angle": 0, + "content": "A cat is perched on a tiled surface with a metal railing" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.174, + 0.357, + 0.184 + ], + "angle": 0, + "content": "beside it and a backdrop of stairs." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.19, + 0.264, + 0.199 + ], + "angle": 0, + "content": "LLaVA" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.201, + 0.442, + 0.209 + ], + "angle": 0, + "content": "A small brown and white cat is sitting on a stair railing." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.209, + 0.327, + 0.218 + ], + "angle": 0, + "content": "looking down at the floor." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.225, + 0.262, + 0.234 + ], + "angle": 0, + "content": "BLIP-2" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.236, + 0.378, + 0.245 + ], + "angle": 0, + "content": "A cat is sitting on the stairs looking up." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.254, + 0.47, + 0.379 + ], + "angle": 0, + "content": "Figure 4. Comparison of three instruction-following VLMs for generating SID. For the reference image of a cat and the instructions shown in the top, the three VLMs generate image captions shown in the right side of the image. Subsequently, the unique identifier [v] is inserted and the resulting captions are used for conditioning the diffusion model as the train descriptions. For painting/cartoon style re-contextualization, we used a slightly different instruction, as detailed in Appendix C.1. Additional fifteen examples for VLM-generated SIDs are available in Appendix E.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.469, + 0.51 + ], + "angle": 0, + "content": "multi-modal GPT-4 [37]. In Fig. 4, we illustrate a caption generation example in which the subject is explicitly defined as a cat within the image. For BLIP-2, it struggles to comprehend detailed instructions and it can often produce sequences of meaningless characters. Therefore, we exclusively conditioned BLIP-2 for the subject of interest while providing the detailed instructions to both LLaVA and GPT-4 for generating SID." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Captions generated by BLIP-2 significantly fall short of the desired SID, lacking informative specifications, i.e., detailed features, of any undesired objects (objects other than the cat in Fig. 4 example) in the images. Although LLaVA generates descriptions better aligned with the instructions, it still falls substantially short, tending to describe the subject excessively or fail to provide sufficient details about undesired objects. Compared to the other two, multi-modal GPT-4 stands out for its ability to follow instructions, delivering descriptions that closely align with the concept of SID. The quantitative comparisons between these VLMs are shown in Appendix E.2, where multi-modal GPT-4 shows the best subject and text alignment. Therefore, we have chosen multi-modal GPT-4 as the VLM for SID generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.21, + 0.748 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We have performed comprehensive experiments to verify the enhancement resulting from SID. For the experiment datasets, images from previous works [12, 24, 45] and three websites [39, 59, 63] were examined. As the main experiment, we have integrated SID with four optimization-based models - DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion (TI) [12]. The superior effectiveness of SID is readily apparent across a broad spectrum of examples as shown in Fig. 5. In the presence of a single reference image only, we have also compared" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "SID-integrated DreamBooth (Ours) with two encoder-based models, ELITE [62] and BLIP-Diffusion [26], along with DreamBooth in Fig. 6. Implementation details for each model are provided in Appendix C.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.159, + 0.787, + 0.176 + ], + "angle": 0, + "content": "5. Analysis of cross-attention map" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.178, + 0.892, + 0.45 + ], + "angle": 0, + "content": "The cross-attention map of each text token is known to highlight the pixels the text token describes [17, 56]. We visualized averaged cross-attention maps for the identifier [v], which is trained to encode the subject's identity. The cross-attention maps are averaged over all the timestamps, layers, and heads to show the single image that can be overlayed on the generated image, following the previous convention [56]. In Fig. 7, we compared cross-attention maps for DreamBooth and SID-integrated DreamBooth across four distinct biases: background, nearby-object, tied-object, and substance biases. While DreamBooth's identifier tends to erroneously focus on undesired objects across all biased scenarios, our method consistently displays highly accurate attention focused on the subject. To extend this analysis, we conducted a comparison between the four cases of train descriptions (detailed in Fig. 3, top row) in Fig. E.2. These analyses confirm that integrating SID does reduce undesired embedding entanglements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.782, + 0.476 + ], + "angle": 0, + "content": "6. Analysis of three key measures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Image-alignment and text-alignment have become popular measures in recent personalized image synthesis [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62]. However, we have found that the widely used image-alignment score tends to be significantly influenced by background entanglement, making it an inappropriate measure for analyzing undesired embedding entanglements. Therefore, we first define two customized measures, subject-alignment and non-subject-disentanglement, by adapting image-alignment specifically for text-to-image personalization. Further details can be found in Appendix A, where we demonstrate the crucial importance of these two measures in assessing subject preservation and reduction in undesired embedding entanglement. Using the two measures, along with text-alignment, we conduct a quantitative analysis of SID." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Customizing image-alignment measure. Consider a set of reference images \\( R = \\{r_1, r_2, \\dots, r_N\\} \\) and a set of generated images \\( G = \\{g_1, g_2, \\dots, g_M\\} \\). Each image in \\( G \\) is generated using \\( R \\) as the reference images with a common generation prompt \\( p \\). Define \\( f_i: \\mathcal{X}_{image} \\to \\mathbb{R}^D \\) and \\( f_t: \\mathcal{X}_{text} \\to \\mathbb{R}^D \\) as the CLIP image and text encoders [41] followed by a unit-norm normalization, respectively. Also, define a subject segmentation function \\( s: \\mathcal{X}_{image} \\to \\{0, 1\\}^{H \\times W} \\), implemented as Grounded-SAM [14] conditioned on the subject's class name. A mask over the image resolution space is output by \\( s \\) where 1 indicates the subject's pixel and 0 indicates the non-subject's pixel." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8316" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.089, + 0.81, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.546, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Figure 5. Enhancement by SID. For four optimization-based models (DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion [12]), the baseline results are shown together with SID-integrated results. SID-integration effectively resolves entanglement issues in scenarios with high biases, represented by indoor background (1st row), nearby potted plant (2nd row), filled-in blueberries (3rd row), and sunflower substances (last row). Additional examples can be found in Appendix E.4." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.61, + 0.437, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.821, + 0.473, + 0.906 + ], + "angle": 0, + "content": "Figure 6. Enhancement by SID for a single reference image. Two encoder-based models, ELITE [62] and BLIP-Diffusion [26], are compared with DreamBooth [45] and our SID-integrated DreamBooth. Notably, in the second row, our approach effectively addresses pose bias. Additional examples can be found in Appendix E.5." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.892, + 0.701 + ], + "angle": 0, + "content": "The subject-alignment \\(\\mathcal{SA}\\) is defined as the average pairwise cosine similarity between generated images and subject segments of reference images in the CLIP-embedding space. The non-subject-disentanglement \\(\\mathcal{NSD}\\) is defined as the 1 minus the non-subject segment similarity in CLIP-embedding space. They can be formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.716, + 0.892, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S A} = \\quad A v g \\left(f _ {i} \\left(r _ {n} \\odot s \\left(r _ {n}\\right)\\right) ^ {\\intercal} \\cdot f _ {i} \\left(g _ {m}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.735, + 0.892, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N S D} = 1 - \\operatorname {A v g} \\left(f _ {i} \\left(r _ {n} \\odot (1 - s (r _ {n}))\\right) ^ {\\top} \\cdot f _ {i} \\left(g _ {m}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.767, + 0.894, + 0.827 + ], + "angle": 0, + "content": "where \\(Avg(\\cdot)\\) denotes the average operation over all possible pair selections between \\(n = 1, \\ldots, N\\) and \\(m = 1, \\ldots, M\\). Additionally, we evaluate text-alignment \\(\\mathcal{T}\\mathcal{A}\\) shown below." + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.842, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\mathcal {A} = A v g \\left(f _ {t} (p) ^ {\\mathrm {T}} \\cdot f _ {i} \\left(g _ {m}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.875, + 0.892, + 0.906 + ], + "angle": 0, + "content": "The identifier [v] is removed from the generation prompt \\( p \\) when evaluating \\( \\mathcal{T}\\mathcal{A} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8317" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.46, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.48 + ], + "angle": 0, + "content": "Figure 7. Cross-attention map. Averaged cross-attention maps for the identifier [v] are shown. They are overlayed on the generated images. The undesirably entangled embeddings of DreamBooth [45] can be confirmed from the erroneous focusing on the undesirable objects. On the contrary, SID-integrated DreamBooth accurately highlights the subjects of interest: dog, perfume, girl, and painting style. Image credit: David Revoy (3rd row)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.473, + 0.852 + ], + "angle": 0, + "content": "Analysis of SID. The three measures have straightforward interpretations. A higher subject-alignment indicates superior preservation of the subject's identity. Elevated non-subject-disentanglement values signify a reduction in undesired embedding entanglements. Increased text-alignment suggests that the generated images align more closely with the generation prompt. Using the three measures, we have evaluated the three optimization-based models and their SID-integrated counterparts as shown in Fig. 8. For the single reference images, we evaluated ELITE [62] and BLIP-Diffusion [26] together with the three optimization-based models and the results are shown in Fig. 9. From the figures, it can be observed that all of the best performing, or Pareto optimal, models are the SID-integrated models. The average improvements of the three key metrics exhibit positive values with a substantial margin in all cases, except for subject alignment in Fig. 9. A potential explanation for this observation is the occurrence of subject overfitting in the existing models, particularly evident in cases such as pose bias. The implementation of SID mitigates subject overfitting, even when only a single reference image is provided. For the evaluation, we have utilized commonly used personalization datasets [12, 24, 45] to generate 7500 images. The details are articulated in Appendix C.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Human evaluation. We further assess the effectiveness of SID over its baseline by conducting human evaluation. Human evaluation was performed involving 130 partici" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.091, + 0.693, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.228, + 0.86, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.091, + 0.877, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Figure 8. Pair-wise metric visualization for multiple reference images. The best performing ones are the SID-integrated models and they form the Pareto boundary." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.308, + 0.693, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.446, + 0.86, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.309, + 0.877, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.465, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Figure 9. Pair-wise metric visualization for single reference image. The best performing ones are the SID-integrated optimization-based models and they form the Pareto boundary. They also outperform the two encoder-based models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.622 + ], + "angle": 0, + "content": "pants, 10 subjects per participant, and 4 questions per subject (a total of 5,200 responses). Tab. B.2 shows that the survey results for subject-alignment, non-subject disentanglement, and text-alignment are consistent with the metric analysis in Fig. 8, supporting our metric analysis. The results and additional details are provided in Appendix B" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.614, + 0.655 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.805, + 0.682 + ], + "angle": 0, + "content": "7.1. Negative prompt and segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Negative prompt for classifier-free guidance [20] and segmentation can be considered as the alternatives of SID. Negative prompt is typically used to prevent the generation of unwanted features [49, 58, 64]. Segmentation mask has been used in several previous works for the purpose of isolating subject [2, 21, 26, 50, 62]. Instead of adopting SID, negative prompt can be used during inference or the subject's segmentation mask can be applied before per-subject optimization. In Fig. 10, SID-integration is compared to these two approaches. Negative prompts exhibit unsatisfactory disentanglement, suggesting that severely entangled representations cannot be easily unraveled during inference. Segmentation masks also exhibit its own limitations, and face challenges in dynamically editing the pose" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8318" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.095, + 0.203, + 0.106 + ], + "angle": 0, + "content": "Reference images" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.107, + 0.21, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.193, + 0.164, + 0.202 + ], + "angle": 0, + "content": "dog" + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.093, + 0.351, + 0.107 + ], + "angle": 0, + "content": "Negative prompt" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.111, + 0.38, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.189, + 0.487, + 0.2 + ], + "angle": 0, + "content": "A [v] dog sitting in front of the Eiffel Tower." + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.092, + 0.539, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.095, + 0.476, + 0.105 + ], + "angle": 0, + "content": "SID" + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.19, + 0.539, + 0.2 + ], + "angle": 0, + "content": "t of the Eiffel Tower." + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.094, + 0.67, + 0.106 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.111, + 0.708, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.189, + 0.851, + 0.2 + ], + "angle": 0, + "content": "A [v] dog howling at the moon on a rocky, barren landscape." + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.093, + 0.804, + 0.104 + ], + "angle": 0, + "content": "SID" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.111, + 0.866, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.211, + 0.208, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.299, + 0.174, + 0.308 + ], + "angle": 0, + "content": "perfume" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.216, + 0.38, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.295, + 0.456, + 0.306 + ], + "angle": 0, + "content": "A [v] perfume on a park bench." + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.216, + 0.54, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.216, + 0.708, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.216, + 0.866, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.293, + 0.824, + 0.313 + ], + "angle": 0, + "content": "A [v] perfume on a windowsill of an old cottage, overlooking a meadow." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.894, + 0.376 + ], + "angle": 0, + "content": "Figure 10. Comparison with negative prompt and segmentation. DreamBooth [45] is used as the base model. Compared to the two alternatives, SID-integration demonstrates superior non-subject-disentanglement and text-alignment. Negative prompts used: “sitting on the fluffy blanket and green couch” in the upper row example, “next to a black Prada purse with silver hardware on a bed of white sheets” in the lower row example. Additional examples can be found in Appendix E.6" + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.394, + 0.236, + 0.404 + ], + "angle": 0, + "content": "Reference images" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.405, + 0.152, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.406, + 0.226, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.406, + 0.3, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.463, + 0.151, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.463, + 0.226, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.463, + 0.3, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.521, + 0.2, + 0.53 + ], + "angle": 0, + "content": "dog" + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.394, + 0.38, + 0.402 + ], + "angle": 0, + "content": "DreamBooth" + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.404, + 0.392, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.463, + 0.391, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.521, + 0.434, + 0.531 + ], + "angle": 0, + "content": "A purple [v] dog" + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.394, + 0.466, + 0.404 + ], + "angle": 0, + "content": "DreamBooth + SID" + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.405, + 0.468, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.463, + 0.468, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.545, + 0.469, + 0.614 + ], + "angle": 0, + "content": "Figure 11. Subject editing. Two examples are shown for editing subject's color. DreamBooth [45] erroneously changes the color of non-subject elements (sofa in the upper row example, background in the lower row example), while SID-integrated DreamBooth accurately edits the subject (dog)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.468, + 0.673 + ], + "angle": 0, + "content": "of the subject or faithfully following the generation prompt. SID shows higher non-subject-disentanglement and text-alignment than the two alternatives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.314, + 0.694 + ], + "angle": 0, + "content": "7.2. Enhancing subject editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.469, + 0.845 + ], + "angle": 0, + "content": "The identifier [v], specifically trained to encode the subject's identity, enables us to modify the properties of the subject through a textual guidance. Undesired embedding entanglements, however, can lead to misaligned modifications. In Fig. 11, an example of subject's color editing is shown. DreamBooth [45], influenced by embedding entanglements, fails to edit the subject and instead modifies the color of an undesired object or background. In contrast, SID-integrated DreamBooth precisely modifies the subject itself, following the provided color guidance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.851, + 0.257, + 0.866 + ], + "angle": 0, + "content": "7.3. Limitations of SID" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "A limitation of our approach is associated with the imperfections of VLM. While the multi-modal GPT-4 [37] gen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.588 + ], + "angle": 0, + "content": "erally generates descriptions that closely align with the provided instructions, occasional failures do occur. Although these instances are rare and may not be easily generalized, we have listed some of these failure cases in Appendix D.1. Another limitation is the undesired entanglement of the subject's strong facial expressions. We discovered that this failure arises when the VLM-generated SID lacks informative specifications of facial expressions. This issue can be addressed by including undesired information in the SID, a solution detailed further in Appendix D.2. Finally, SID is exclusively integrated with optimization-based models in this study. Yet, there is potential for its integration into encoder-based models, particularly during encoder pre-training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.619, + 0.618 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.792 + ], + "angle": 0, + "content": "In this study, we have introduced a robust strategy to mitigate undesired embedding entanglement in text-to-image personalization. Beginning with the identification of five key biases, we have proposed SID (Selectively Informative Description) as an effective solution to address these biases. Our cross-attention analysis demonstrates the successful removal of entanglements, while alignment analysis indicates notable enhancements in non-subject-disentanglement and text-alignment. The proposed SID strategy holds potential applicability to other multi-modal applications where managing embedding entanglements is crucial." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the following grants funded by the Korea government: NRF (NRF2020R1A2C2007139, NRF-2022R1A6A1A03063039) and IITP ([NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], [No. RS2023-00235293])." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8319" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.184, + 0.471, + 0.237 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.238, + 0.471, + 0.307 + ], + "angle": 0, + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[8] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.47, + 0.753 + ], + "angle": 0, + "content": "[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.753, + 0.47, + 0.821 + ], + "angle": 0, + "content": "[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.821, + 0.47, + 0.875 + ], + "angle": 0, + "content": "[13] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.875, + 0.47, + 0.888 + ], + "angle": 0, + "content": "[14] Grounded-SAM Contributors. Grounded-Segment-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Anything. https://github.com/IDEA-Research/Grounded-Segment-Anything，Apr.2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.134, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.187, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[16] Haotian Liu. liuhaotian/llava-v1-0719-336px-lora-merge- vicuna-13b-v1.3. https://huggingface.co/ liuhaotian/llava-v1-0719-336px-lora- merge-vicuna-13b-v1.3, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.242, + 0.892, + 0.296 + ], + "angle": 0, + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[18] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[21] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[23] Juil Koo, Seungwoo Yoo, Minh Hieu Nguyen, and Minhyuk Sung. Salad: Part-level latent diffusion for 3d shape generation and manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14441-14451, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.607, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[26] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.797, + 0.892, + 0.851 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.851, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[28] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audi-Oldm: Text-to-audio generation with latent diffusion models." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8320" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.092, + 0.35, + 0.106 + ], + "angle": 0, + "content": "arXiv preprint arXiv:2301.12503, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.106, + 0.469, + 0.145 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.146, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[30] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.2, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[31] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.469, + 0.336 + ], + "angle": 0, + "content": "[32] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.336, + 0.468, + 0.363 + ], + "angle": 0, + "content": "[33] Makoto Shing. svdiff-pytorch. https://github.com/mkshing/svdiff-pytorch, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[34] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.417, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.485, + 0.468, + 0.512 + ], + "angle": 0, + "content": "[36] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.341, + 0.524 + ], + "angle": 0, + "content": "[37] OpenAI. Gpt-4 technical report, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[38] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.467, + 0.617 + ], + "angle": 0, + "content": "[39] Peper and Carrot. https://www.peppercarrot.com/." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[40] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.739 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.74, + 0.469, + 0.793 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.794, + 0.469, + 0.862 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.862, + 0.469, + 0.89 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "tation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.148, + 0.892, + 0.229 + ], + "angle": 0, + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.229, + 0.892, + 0.311 + ], + "angle": 0, + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.311, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[47] Salesforce. LAVIS. https://github.com/salesforce/LAVIS/tree/main/projects/blip-diffusion, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.352, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[48] Salesforce. Salesforce/blip2-opt-2.7b. https://huggingface.co/Salesforce/blip2-opt-2.7b, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[49] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[50] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.5, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[51] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[54] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.809 + ], + "angle": 0, + "content": "[56] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.81, + 0.892, + 0.865 + ], + "angle": 0, + "content": "[57] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.865, + 0.892, + 0.892 + ], + "angle": 0, + "content": "[58] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8321" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.368, + 0.146 + ], + "angle": 0, + "content": "[59] Unsplash. https://unsplash.com/." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.147, + 0.469, + 0.199 + ], + "angle": 0, + "content": "[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.201, + 0.469, + 0.268 + ], + "angle": 0, + "content": "[61] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5764-5773, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.269, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[62] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.323, + 0.383, + 0.335 + ], + "angle": 0, + "content": "[63] Wikiart. https://www.wikiart.org/." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.336, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[64] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.403, + 0.469, + 0.47 + ], + "angle": 0, + "content": "[65] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.471, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[66] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.468, + 0.537 + ], + "angle": 0, + "content": "[67] Yuxiang Wei. ELITE. https://github.com/csyxwei/ELITE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.538, + 0.469, + 0.591 + ], + "angle": 0, + "content": "[68] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.592, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[69] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8322" + } + ] +] \ No newline at end of file diff --git a/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_origin.pdf b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8f7f3bcc92e747968e856f200f03f60937aa6a3 --- /dev/null +++ b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/6ad8fbbe-ab46-41fb-8d30-63234684924c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1beba901919376728ef6cdffa5f5c2281b533cbe9d627759eddbf2627c4f98e +size 2735409 diff --git a/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/full.md b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..032b8519297f52a1ba549cb0dc9c2301c2efc852 --- /dev/null +++ b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/full.md @@ -0,0 +1,333 @@ +# Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization + +Jimyeong Kim1 Jungwon Park1 Wonjong Rhee1,2,3 +Department of Intelligence and Information1 & IPAI2 & RICS3, Seoul National University +{wlaud1001, quoded97, wrhee}@snu.ac.kr + +# Abstract + +In text-to-image personalization, a timely and crucial challenge is the tendency of generated images overfitting to the biases present in the reference images. We initiate our study with a comprehensive categorization of the biases into background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. These biases manifest in the generated images due to their entanglement into the subject embedding. This undesired embedding entanglement not only results in the reflection of biases from the reference images into the generated images but also notably diminishes the alignment of the generated images with the given generation prompt. To address this challenge, we propose SID (Selectively Informative Description), a text description strategy that deviates from the prevalent approach of only characterizing the subject's class identification. SID is generated utilizing multimodal GPT-4 and can be seamlessly integrated into optimization-based models. We present comprehensive experimental results along with analyses of cross-attention maps, subject-alignment, non-subject-disentanglement, and text-alignment. + +# 1. Introduction + +Text-to-image diffusion models [5, 7, 11, 25, 30, 35, 40, 42, 43, 46] have demonstrated remarkable generation capabilities that align with textual descriptions across various applications. One such application is text-to-image personalization [12, 24, 45], where models are tailored to generate novel renditions of subjects described in a few reference images. Recent works, such as DreamBooth [45] or Custom Diffusion [24], stand out for their exceptional personalized generation results. They achieve this by fine-tuning a pretrained text-to-image diffusion model with a small number of reference images using a text description in the format of "a [v] [class name]" or "photo of a [v] [class name]." In this context, [v] represents the subject's unique identifier, which uses a rare token with minimal semantic significance, while + +the class descriptor [class name] represents a coarse category for the subject. + +Recent developments in text-to-image personalization can be categorized into optimization-based approaches [12, 15, 24, 45, 57] and encoder-based approaches [21, 26, 50, 62]. For both types of approaches, a problem typically addressed as 'overfitting' has emerged [2, 15, 21, 26, 45, 57]. It is a phenomenon where objects, other than the subject of interest, in the reference images affect the generated image in an undesired way. For the sake of clarity, we refer to such an entity as 'non-subject' or 'undesired object'. The cause of this phenomenon is closely related to the embedding of the identifier token [v] for optimization-based approaches and the embedding of the subject for encoder-based models, where the information related to any non-subject seeps into the embeddings. Rather than referring to it as overfitting, we formally address this phenomenon as undesired embedding entanglement. + +The phenomenon of undesired embedding entanglement has been noticed, but it has been only superficially understood so far. A potential explanation for the limited understanding is the rarity of this phenomenon in straightforward scenarios. Instances where reference images encompass diverse subject views and minimal undesired objects are much less prone to this phenomenon. However, such an ideal scenario is highly unlikely in real-world applications. Therefore, it is crucial to comprehensively understand the biases and develop counter methods to address them effectively. In an effort to gain a more thorough understanding of the prevalent biases contributing to these entanglements, we conducted exhaustive paper reviews followed by extensive experiments utilizing state-of-the-art models. The resulting identification of the five most commonly encountered biases is presented in Fig. 1. While we have chosen DreamBooth [45] for generating the examples in Fig. 1, additional examples for Custom Diffusion [24], SVDiff [15], ELITE [62] and BLIP-Diffusion [26] will be provided later. + +In prior studies, several models [2, 21, 26, 50, 62] have employed segmentation masks. However, depending exclusively on segmentation masks comes with limitations, as we + +![](images/742751729213fb74dcc71bfb43b08cf6534b513a477aa7acb8c96feade00a8d4.jpg) +Figure 1. Five key biases – background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. The first four rows depict scenarios with multiple reference images, while the last row illustrates a single reference image scenario. The pose bias is particularly prone to manifest in scenarios involving a single reference image, although it can also occur when multiple reference images depict the subject in poses that are similar. In the generation prompt, the subject of interest is highlighted in red. The integration of our method into DreamBooth [45] effectively resolves embedding entanglements associated with the five key biases (rightmost column). + +will demonstrate with additional examples in subsequent sections. Notably, these limitations become apparent in the cases of substance, tied-object, and pose biases. To tackle all of the key biases, we introduce SID (Selectively Informative Description) that requires only modifications in the train text descriptions of the reference images. The idea is very simple. Starting with an optimization-based model, we diverge from the prevalent approach of only characterizing the subject's class identification as observed in the existing works. Our approach involves integrating informative specifications of the "undesired objects" into the train descriptions. This modification significantly diminishes the probability of undesired entanglement between the subject embedding [v] and the non-subject information in the reference images. As the text-to-image diffusion models are trained for alignments, adding text descriptions that match non-subjects can greatly help the non-subject parts in the reference images to align with the corresponding text descriptions. This prevents the undesired objects from acci + +dently becoming aligned with [v]. In contrast, such an accidental alignment can quite easily occur in the absence of the matching text descriptions. Our method is selective because we deliberately avoid incorporating informative specifications of the "subject" itself into the train descriptions. Such incorporation could negatively affect the preservation of the subject's identity, as will be elucidated in Sec. 3.1. SID can be readily integrated with any optimization-based model, such as DreamBooth [45], Custom Diffusion [24], or SVDiff [15]. For the automated generation of SID, we utilize the multimodal GPT-4 [37] by guiding it with an appropriate instruction. Surprisingly, the proposed SID integration significantly reduces embedding entanglement, especially in critically biased scenarios. + +# 2. Related Work + +Text-to-image diffusion models. Diffusion models [19, 54, 55] have demonstrated remarkable generative capabilities across a wide range of modalities, including im + +ages [10, 35, 46], audios [8, 22, 28], videos [18, 32, 66], and 3D shapes [23, 68, 69]. Within this diverse spectrum, prominent text-to-image diffusion models, such as Stable Diffusion [43], Imagen [46], and DALL-E 2 [42], have exhibited exceptional proficiency in the synthesis of high-quality images closely aligned with textual descriptions. Subsequently, significant efforts [5, 7, 11, 25, 30, 40] have been dedicated to enhancing the text-to-image alignment in the field of text-to-image generation. Our study, based on the Stable Diffusion model [43], improves the alignment between textual descriptions and images in personalized image synthesis, focusing on accurately preserving subject identity without any entanglement with undesired objects. + +Vision-language models. The vision-language model (VLM) handles tasks like text-to-image retrieval [61], visual question answering [1], and image captioning [65]. VLMs can also play a crucial role as a building block, as seen with CLIP [41] in tasks such as text-to-image generation [35, 42, 43], image segmentation [31], image retrieval [4], and more [34, 38, 51, 52]. In our work, we need an image captioning VLM capable of closely following detailed instructions. While models like BLIP-2 [27] and OpenFlamingo [3] offer image captioning with text conditioning, they often struggle to adhere closely to detailed instructions. In contrast, large models such as LLaVA [29] and multi-modal GPT-4 [37] have demonstrated their ability to generate precise captions closely aligned with detailed instructions. Among these options, multi-modal GPT-4 exhibits the most powerful capability in generating instruction-following captions. + +Personalized image synthesis. Personalizing text-to-image diffusion models [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62] has become a central focus in text-to-image generation. These models generate high-quality, diverse renditions of a subject using either a few reference images (usually 3 to 7) or just a single reference image. These models can be broadly categorized into two groups: optimization-based and encoder-based models. Optimization-based models [2, 12, 15, 24, 45, 57], like DreamBooth [45], involve per-subject optimization, where they fine-tune identifier tokens or segments of diffusion models to encode a subject's identity. On the other hand, encoder-based models [13, 21, 26, 50, 62] pre-train separate identity encoders and utilize them to encode a subject's identity without the need for per-subject optimization. + +# 3. Method + +Our proposed method can be readily integrated with any optimization-based models that utilize description of the reference images. The overall process, presented in Fig. 2, entails the generation of SID (Selectively Informative Description) for each reference image using an instruction-following vision-language model (VLM). These SIDs are + +![](images/2c8ac0c54cb061e34445d85e41d129667892b555d87d97f6e1b77b4527655688.jpg) +Figure 2. Personalization with SID. We propose integrating SID (Selectively Informative Description) into the per-subject optimization, where an instruction-following VLM (Vison-Language Model) is utilized to generate a selectively informative description for each reference image. + +integrated into the reconstruction learning process as part of the per-subject optimization. Depending on the underlying optimization-based model, additional loss functions may be incorporated together with the main reconstruction loss. The additional loss functions, such as the class prior preservation loss [45] introduced in DreamBooth, are applied in the same manner as in their respective baseline models without any modification. The effect of SID is addressed in Sec. 3.1. The choice of VLM and its accompanying instruction are addressed in Sec. 3.2. + +# 3.1. SID for reducing embedding entanglement + +The existing text-to-image personalization approaches [12, 15, 24, 45], especially optimization-based methods like DreamBooth [45], often rely on simplistic textual descriptions that closely follow the format of "a [v] [class name]", where [v] denotes the unique identifier for the subject, in order to extract the subject's identity. In our method, our goal is to come up with train descriptions that have an appropriate level of information such that we can effectively isolate the subject of interest even when the reference images are biased. According to the findings in DreamBooth [45], specifying the subject's correct class in the train description is instrumental in accurately preserving the subject's identity. With the correct class information, the model's prior knowledge of the subject's class can significantly enhance the editing capabilities of the model. Based on the findings, we consider the case where the train description only contains the subject's class identification as the baseline and subsequently explore additionally informative cases that are listed in Tab. 1. In Tab. 1, the term undesired object generally encompasses any object within reference images, excluding the subject of interest. For the examples in Fig. 1, the undesired objects can be readily recognized for background, nearby-object, and tied-object biases. For the substance bias, that is applicable only when performing a style + +![](images/2fdc6611ce90fa031679e88d855bd141975442390481107201e5aed552909ee0.jpg) +Figure 3. Two examples for comparing the four cases of descriptions. For each case, the choice of train description follows the guidelines in Tab. 1. The common generation prompt for each example is shown below the generated images. Additional examples can be found in Appendix E.1. + +![](images/3f8bbc6b5934b25badf2759903c2c465ce87fde9463ff35ae435662d0530a098.jpg) + +
DescriptionsSubjectUndesired objects
Class IdentificationInformative SpecificationsClass IdentificationInformative Specifications
Case 1 (Baseline)OXXX
Case 2OXOX
Case 3 (Ours)OXOO
Case 4OOOO
+ +Table 1. Four cases of train descriptions. The four cases are categorized based on whether they provide class identification only or additional informative specifications of the subject and undesired objects. We define $SID$ as the description of Case 3 (Ours). + +re-contextualization, all the objects in the reference images are considered to be undesirable because the intention is to transfer the style only. For pose bias, it is inherent to the subject itself. + +For the four cases in Tab. 1, two examples are investigated in Fig. 3 where two sets of the resulting images are shown. For the simplicity of this investigation, reference images were chosen to share the same set of undesired objects and thus allowing us to use a single train description for all reference images. Case 1 (Baseline) uses the format "a [v] [class name]", where only the subject's class identification is included in the train description. In this case, the model often struggles with the bias in the reference images (e.g., the subject can co-appear with undesired objects). This is because of the embedding entanglement in [v] as will be + +analyze further in Sec. 5. To reduce the undesired embedding entanglement, we introduce Case 2 where the train description contains the class identifications of the undesired objects. Case 3 is similar to Case 2, but the train description additionally contain informative specifications of the undesired objects. As shown in Fig. 3, both Case 2 and Case 3 can provide a significant improvement over Case 1. Case 3, however, often outperforms Case 2 in terms of entanglement reduction as is the case for the two examples in Fig. 3. + +For the subject itself, it is also possible to include informative specifications of the subject. Case 4 corresponds to this possibility. With the inclusion, however, the subject details described in the informative specifications are disentangled from the [v]. This occurs because of the alignment between the subject's informative specifications in the text and the subject details in the images. Because preserving the subject details in the [v] is crucial in personalized image synthesis, we have opted for Case 3 as our definition of SID. + +# 3.2. VLM for generating SID + +Generating SID for each reference image can be demanding and time-consuming for humans. Therefore, we have chosen to utilize image captioning VLM for an automatic generation of SID. To identify a suitable VLM capable of effectively generating SIDs, we evaluated three well-known image captioning VLMs: BLIP-2 [27], LLaVA [29], and + +# Instruction for GPT-4 & LLaVA + +"Describe the image in one sentence in detail. Please start + +with "A {class}". *You should not describe the distinct + +features of the {class} itself.\* + +# Input text for BLIP-2 + +"A {class}" + +![](images/3c7e53139580e932fcb5376e4e9773760f95fd09aa77a2bf9be83ffb0abe5ac7.jpg) +Figure 4. Comparison of three instruction-following VLMs for generating SID. For the reference image of a cat and the instructions shown in the top, the three VLMs generate image captions shown in the right side of the image. Subsequently, the unique identifier [v] is inserted and the resulting captions are used for conditioning the diffusion model as the train descriptions. For painting/cartoon style re-contextualization, we used a slightly different instruction, as detailed in Appendix C.1. Additional fifteen examples for VLM-generated SIDs are available in Appendix E.2. + +# Multimodal GPT-4 + +A cat is perched on a tiled surface with a metal railing + +beside it and a backdrop of stairs. + +# LLaVA + +A small brown and white cat is sitting on a stair railing. + +looking down at the floor. + +# BLIP-2 + +A cat is sitting on the stairs looking up. + +multi-modal GPT-4 [37]. In Fig. 4, we illustrate a caption generation example in which the subject is explicitly defined as a cat within the image. For BLIP-2, it struggles to comprehend detailed instructions and it can often produce sequences of meaningless characters. Therefore, we exclusively conditioned BLIP-2 for the subject of interest while providing the detailed instructions to both LLaVA and GPT-4 for generating SID. + +Captions generated by BLIP-2 significantly fall short of the desired SID, lacking informative specifications, i.e., detailed features, of any undesired objects (objects other than the cat in Fig. 4 example) in the images. Although LLaVA generates descriptions better aligned with the instructions, it still falls substantially short, tending to describe the subject excessively or fail to provide sufficient details about undesired objects. Compared to the other two, multi-modal GPT-4 stands out for its ability to follow instructions, delivering descriptions that closely align with the concept of SID. The quantitative comparisons between these VLMs are shown in Appendix E.2, where multi-modal GPT-4 shows the best subject and text alignment. Therefore, we have chosen multi-modal GPT-4 as the VLM for SID generation. + +# 4. Experiments + +We have performed comprehensive experiments to verify the enhancement resulting from SID. For the experiment datasets, images from previous works [12, 24, 45] and three websites [39, 59, 63] were examined. As the main experiment, we have integrated SID with four optimization-based models - DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion (TI) [12]. The superior effectiveness of SID is readily apparent across a broad spectrum of examples as shown in Fig. 5. In the presence of a single reference image only, we have also compared + +SID-integrated DreamBooth (Ours) with two encoder-based models, ELITE [62] and BLIP-Diffusion [26], along with DreamBooth in Fig. 6. Implementation details for each model are provided in Appendix C.2. + +# 5. Analysis of cross-attention map + +The cross-attention map of each text token is known to highlight the pixels the text token describes [17, 56]. We visualized averaged cross-attention maps for the identifier [v], which is trained to encode the subject's identity. The cross-attention maps are averaged over all the timestamps, layers, and heads to show the single image that can be overlayed on the generated image, following the previous convention [56]. In Fig. 7, we compared cross-attention maps for DreamBooth and SID-integrated DreamBooth across four distinct biases: background, nearby-object, tied-object, and substance biases. While DreamBooth's identifier tends to erroneously focus on undesired objects across all biased scenarios, our method consistently displays highly accurate attention focused on the subject. To extend this analysis, we conducted a comparison between the four cases of train descriptions (detailed in Fig. 3, top row) in Fig. E.2. These analyses confirm that integrating SID does reduce undesired embedding entanglements. + +# 6. Analysis of three key measures + +Image-alignment and text-alignment have become popular measures in recent personalized image synthesis [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62]. However, we have found that the widely used image-alignment score tends to be significantly influenced by background entanglement, making it an inappropriate measure for analyzing undesired embedding entanglements. Therefore, we first define two customized measures, subject-alignment and non-subject-disentanglement, by adapting image-alignment specifically for text-to-image personalization. Further details can be found in Appendix A, where we demonstrate the crucial importance of these two measures in assessing subject preservation and reduction in undesired embedding entanglement. Using the two measures, along with text-alignment, we conduct a quantitative analysis of SID. + +Customizing image-alignment measure. Consider a set of reference images $R = \{r_1, r_2, \dots, r_N\}$ and a set of generated images $G = \{g_1, g_2, \dots, g_M\}$ . Each image in $G$ is generated using $R$ as the reference images with a common generation prompt $p$ . Define $f_i: \mathcal{X}_{image} \to \mathbb{R}^D$ and $f_t: \mathcal{X}_{text} \to \mathbb{R}^D$ as the CLIP image and text encoders [41] followed by a unit-norm normalization, respectively. Also, define a subject segmentation function $s: \mathcal{X}_{image} \to \{0, 1\}^{H \times W}$ , implemented as Grounded-SAM [14] conditioned on the subject's class name. A mask over the image resolution space is output by $s$ where 1 indicates the subject's pixel and 0 indicates the non-subject's pixel. + +![](images/29a71ef2e2182c12642957e543448e7a9b3731e88cbaf3a4795946e68101b429.jpg) +Figure 5. Enhancement by SID. For four optimization-based models (DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion [12]), the baseline results are shown together with SID-integrated results. SID-integration effectively resolves entanglement issues in scenarios with high biases, represented by indoor background (1st row), nearby potted plant (2nd row), filled-in blueberries (3rd row), and sunflower substances (last row). Additional examples can be found in Appendix E.4. + +![](images/71e3fbc91c896c74fc1f0bd35be6130f4a47e4a5d52f18288a77db2d156182b8.jpg) +Figure 6. Enhancement by SID for a single reference image. Two encoder-based models, ELITE [62] and BLIP-Diffusion [26], are compared with DreamBooth [45] and our SID-integrated DreamBooth. Notably, in the second row, our approach effectively addresses pose bias. Additional examples can be found in Appendix E.5. + +The subject-alignment $\mathcal{SA}$ is defined as the average pairwise cosine similarity between generated images and subject segments of reference images in the CLIP-embedding space. The non-subject-disentanglement $\mathcal{NSD}$ is defined as the 1 minus the non-subject segment similarity in CLIP-embedding space. They can be formally expressed as: + +$$ +\mathcal {S A} = \quad A v g \left(f _ {i} \left(r _ {n} \odot s \left(r _ {n}\right)\right) ^ {\intercal} \cdot f _ {i} \left(g _ {m}\right)\right), \tag {1} +$$ + +$$ +\mathcal {N S D} = 1 - \operatorname {A v g} \left(f _ {i} \left(r _ {n} \odot (1 - s (r _ {n}))\right) ^ {\top} \cdot f _ {i} \left(g _ {m}\right)\right), \tag {2} +$$ + +where $Avg(\cdot)$ denotes the average operation over all possible pair selections between $n = 1, \ldots, N$ and $m = 1, \ldots, M$ . Additionally, we evaluate text-alignment $\mathcal{T}\mathcal{A}$ shown below. + +$$ +\mathcal {T} \mathcal {A} = A v g \left(f _ {t} (p) ^ {\mathrm {T}} \cdot f _ {i} \left(g _ {m}\right)\right). \tag {3} +$$ + +The identifier [v] is removed from the generation prompt $p$ when evaluating $\mathcal{T}\mathcal{A}$ . + +![](images/9e15a669baf562ad3d9ace55fac4d7bdd29d3fba0f9b85ce28738060f048bfe8.jpg) +Figure 7. Cross-attention map. Averaged cross-attention maps for the identifier [v] are shown. They are overlayed on the generated images. The undesirably entangled embeddings of DreamBooth [45] can be confirmed from the erroneous focusing on the undesirable objects. On the contrary, SID-integrated DreamBooth accurately highlights the subjects of interest: dog, perfume, girl, and painting style. Image credit: David Revoy (3rd row) + +Analysis of SID. The three measures have straightforward interpretations. A higher subject-alignment indicates superior preservation of the subject's identity. Elevated non-subject-disentanglement values signify a reduction in undesired embedding entanglements. Increased text-alignment suggests that the generated images align more closely with the generation prompt. Using the three measures, we have evaluated the three optimization-based models and their SID-integrated counterparts as shown in Fig. 8. For the single reference images, we evaluated ELITE [62] and BLIP-Diffusion [26] together with the three optimization-based models and the results are shown in Fig. 9. From the figures, it can be observed that all of the best performing, or Pareto optimal, models are the SID-integrated models. The average improvements of the three key metrics exhibit positive values with a substantial margin in all cases, except for subject alignment in Fig. 9. A potential explanation for this observation is the occurrence of subject overfitting in the existing models, particularly evident in cases such as pose bias. The implementation of SID mitigates subject overfitting, even when only a single reference image is provided. For the evaluation, we have utilized commonly used personalization datasets [12, 24, 45] to generate 7500 images. The details are articulated in Appendix C.4. + +Human evaluation. We further assess the effectiveness of SID over its baseline by conducting human evaluation. Human evaluation was performed involving 130 partici + +![](images/74159a49e297b735dd6f6f8e0597b4e36fd351fa1a18050b75ddbaa8ccb98660.jpg) + +![](images/3a166aa1d6b163a53322304558c161bcb5c28e9aad34025b10c74e041d9e3b7b.jpg) +Figure 8. Pair-wise metric visualization for multiple reference images. The best performing ones are the SID-integrated models and they form the Pareto boundary. + +![](images/3856ebb230db49d24dee7641c029e4bd476c9b1723f2011f1d808a70f4521cd0.jpg) + +![](images/9c41e3ffff44f26f064264c59be439782923402cbb0b1d45142fcdfa63af8a71.jpg) +Figure 9. Pair-wise metric visualization for single reference image. The best performing ones are the SID-integrated optimization-based models and they form the Pareto boundary. They also outperform the two encoder-based models. + +![](images/21dd72467e307712903dc2c1f60d676c989f6132aed7cac1868d80ff6bf3f6c3.jpg) + +![](images/9353a6e4fbcf1a58a6c7a16b4d0926cded8b2cf43d01ea9557c2b073a5557eea.jpg) + +pants, 10 subjects per participant, and 4 questions per subject (a total of 5,200 responses). Tab. B.2 shows that the survey results for subject-alignment, non-subject disentanglement, and text-alignment are consistent with the metric analysis in Fig. 8, supporting our metric analysis. The results and additional details are provided in Appendix B + +# 7. Discussion + +# 7.1. Negative prompt and segmentation + +Negative prompt for classifier-free guidance [20] and segmentation can be considered as the alternatives of SID. Negative prompt is typically used to prevent the generation of unwanted features [49, 58, 64]. Segmentation mask has been used in several previous works for the purpose of isolating subject [2, 21, 26, 50, 62]. Instead of adopting SID, negative prompt can be used during inference or the subject's segmentation mask can be applied before per-subject optimization. In Fig. 10, SID-integration is compared to these two approaches. Negative prompts exhibit unsatisfactory disentanglement, suggesting that severely entangled representations cannot be easily unraveled during inference. Segmentation masks also exhibit its own limitations, and face challenges in dynamically editing the pose + +![](images/2ffc763920d69243daeae35b6e31bc79c977ecaf48d607567888e90e9999d4a2.jpg) +Reference images +dog + +![](images/8a483685c79bdf40621fb8ff4006c01affa0c2117355e65573d8aee3fa0f28ae.jpg) +Negative prompt +A [v] dog sitting in front of the Eiffel Tower. + +![](images/8db23fd8229ccf40eb8ae5bde0af3f73cc0f211aa57ebb76b732376a201a6e70.jpg) +SID +t of the Eiffel Tower. + +![](images/bb4ad1169d22d74e3890604c713e7476b55d1a4e9184335f883bbfd02c52c5a8.jpg) +Segmentation +A [v] dog howling at the moon on a rocky, barren landscape. + +![](images/f17f27c812c3a173a35581692984016aab0c9d03a214480e833edc682ba72832.jpg) +SID + +![](images/6bbdd96fdfd3f9532ec9104d039eaf6556f6dff9d3ac5cc4b2d896aafed791c5.jpg) +perfume +Figure 10. Comparison with negative prompt and segmentation. DreamBooth [45] is used as the base model. Compared to the two alternatives, SID-integration demonstrates superior non-subject-disentanglement and text-alignment. Negative prompts used: “sitting on the fluffy blanket and green couch” in the upper row example, “next to a black Prada purse with silver hardware on a bed of white sheets” in the lower row example. Additional examples can be found in Appendix E.6 + +![](images/1906ca109883cf2b4df74b69bd5a387a73be2ef3bc6de583217c68f387ac66c1.jpg) +A [v] perfume on a park bench. + +![](images/ff6c1cf1ddd15c88d66f485ccafcf3a376fd94f5a6705d34a63633b9a09e54ef.jpg) + +![](images/4c5109f67ed28acf180985a4fe20d75795baad51048944ecd680774695019b95.jpg) +A [v] perfume on a windowsill of an old cottage, overlooking a meadow. + +![](images/ab5f3f350627f168133db623865ac9d6db6a695b26e6f4788a9567a8a4df7c34.jpg) + +![](images/030827ed414845371adad59f08cbf421836c0e863123bc1eb3add85c7923eb0f.jpg) +Reference images + +![](images/3ab1907664b6b52eac2f8efa6eb11d5c24b59d9b50bb6b1e3e3965277fca237d.jpg) + +![](images/fe6f37e8b8e2e2376dadbb669012d8b4fe3186b71f1d8e52dd221f8871f44dfe.jpg) + +![](images/f79ac9163b52eb4371c483fcbcf98d3c815837de3e7c7a146203dbf8187a974f.jpg) +Figure 11. Subject editing. Two examples are shown for editing subject's color. DreamBooth [45] erroneously changes the color of non-subject elements (sofa in the upper row example, background in the lower row example), while SID-integrated DreamBooth accurately edits the subject (dog). + +![](images/2fece19a93670996d598e228c477fb785366186b9382e4760fa00ab6b230445b.jpg) +dog + +![](images/30709109dbeb4a2149ab014b19331b0ed9c7a27e7b86206d6e1c4ae1db241686.jpg) + +![](images/e94f49cbbc58a9e23bca4e7794a0f3ef936d67e511401a667626480906162ab6.jpg) +DreamBooth + +![](images/51b4cba7f5f16f996be2c3f65e356adfdec4c60f390e74671e0a7353ec491ae8.jpg) + +![](images/95e92908269282786772cc715f14ef776bda051f39305502f8ea8720825461e8.jpg) +DreamBooth + SID + +![](images/a07a0243035f6f47c7c2723e8cbba4f6f2c9c7721b70d46319113ddd7997ab70.jpg) +A purple [v] dog + +of the subject or faithfully following the generation prompt. SID shows higher non-subject-disentanglement and text-alignment than the two alternatives. + +# 7.2. Enhancing subject editing + +The identifier [v], specifically trained to encode the subject's identity, enables us to modify the properties of the subject through a textual guidance. Undesired embedding entanglements, however, can lead to misaligned modifications. In Fig. 11, an example of subject's color editing is shown. DreamBooth [45], influenced by embedding entanglements, fails to edit the subject and instead modifies the color of an undesired object or background. In contrast, SID-integrated DreamBooth precisely modifies the subject itself, following the provided color guidance. + +# 7.3. Limitations of SID + +A limitation of our approach is associated with the imperfections of VLM. While the multi-modal GPT-4 [37] gen + +erally generates descriptions that closely align with the provided instructions, occasional failures do occur. Although these instances are rare and may not be easily generalized, we have listed some of these failure cases in Appendix D.1. Another limitation is the undesired entanglement of the subject's strong facial expressions. We discovered that this failure arises when the VLM-generated SID lacks informative specifications of facial expressions. This issue can be addressed by including undesired information in the SID, a solution detailed further in Appendix D.2. Finally, SID is exclusively integrated with optimization-based models in this study. Yet, there is potential for its integration into encoder-based models, particularly during encoder pre-training. + +# 8. Conclusion + +In this study, we have introduced a robust strategy to mitigate undesired embedding entanglement in text-to-image personalization. Beginning with the identification of five key biases, we have proposed SID (Selectively Informative Description) as an effective solution to address these biases. Our cross-attention analysis demonstrates the successful removal of entanglements, while alignment analysis indicates notable enhancements in non-subject-disentanglement and text-alignment. The proposed SID strategy holds potential applicability to other multi-modal applications where managing embedding entanglements is crucial. + +Acknowledgement This work was supported by the following grants funded by the Korea government: NRF (NRF2020R1A2C2007139, NRF-2022R1A6A1A03063039) and IITP ([NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], [No. RS2023-00235293]). + +# References + +[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. +[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023. +[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. +[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022. +[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. +[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. +[8] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020. +[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023. +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. +[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. +[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. +[13] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. +[14] Grounded-SAM Contributors. Grounded-Segment- + +Anything. https://github.com/IDEA-Research/Grounded-Segment-Anything，Apr.2023. +[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023. +[16] Haotian Liu. liuhaotian/llava-v1-0719-336px-lora-merge- vicuna-13b-v1.3. https://huggingface.co/ liuhaotian/llava-v1-0719-336px-lora- merge-vicuna-13b-v1.3, 2023. +[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. +[18] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. +[21] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. +[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. +[23] Juil Koo, Seungwoo Yoo, Minh Hieu Nguyen, and Minhyuk Sung. Salad: Part-level latent diffusion for 3d shape generation and manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14441-14451, 2023. +[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. +[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. +[26] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023. +[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. +[28] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audi-Oldm: Text-to-audio generation with latent diffusion models. + +arXiv preprint arXiv:2301.12503, 2023. +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. +[30] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. +[31] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. +[32] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023. +[33] Makoto Shing. svdiff-pytorch. https://github.com/mkshing/svdiff-pytorch, 2023. +[34] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. +[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. +[36] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023. +[37] OpenAI. Gpt-4 technical report, 2023. +[38] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. +[39] Peper and Carrot. https://www.peppercarrot.com/. +[40] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427, 2023. +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen + +tation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. +[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. +[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. +[47] Salesforce. LAVIS. https://github.com/salesforce/LAVIS/tree/main/projects/blip-diffusion, 2023. +[48] Salesforce. Salesforce/blip2-opt-2.7b. https://huggingface.co/Salesforce/blip2-opt-2.7b, 2023. +[49] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. +[50] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. +[51] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022. +[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. +[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. +[54] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. +[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. +[56] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022. +[57] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. +[58] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven + +image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. +[59] Unsplash. https://unsplash.com/. +[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. +[61] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5764-5773, 2019. +[62] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. +[63] Wikiart. https://www.wikiart.org/. +[64] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. +[65] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015. +[66] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022. +[67] Yuxiang Wei. ELITE. https://github.com/csyxwei/ELITE, 2023. +[68] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. +[69] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021. \ No newline at end of file diff --git a/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/images.zip b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aebc972e712c0d114c07de906c32817454d548ea --- /dev/null +++ b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d91c3cff40c0f829cd0d89a0c228a3cd237286e12ec3e233928025276e4ced37 +size 945555 diff --git a/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/layout.json b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4451edaaa4b1b6e4649ffb58eba193db9a95b692 --- /dev/null +++ b/2024/Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization/layout.json @@ -0,0 +1,8373 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 102, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 102, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 102, + 515, + 140 + ], + "type": "text", + "content": "Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 515, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 515, + 204 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 515, + 204 + ], + "type": "text", + "content": "Jimyeong Kim1 Jungwon Park1 Wonjong Rhee1,2,3 \nDepartment of Intelligence and Information1 & IPAI2 & RICS3, Seoul National University \n{wlaud1001, quoded97, wrhee}@snu.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 257, + 290, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 498 + ], + "type": "text", + "content": "In text-to-image personalization, a timely and crucial challenge is the tendency of generated images overfitting to the biases present in the reference images. We initiate our study with a comprehensive categorization of the biases into background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. These biases manifest in the generated images due to their entanglement into the subject embedding. This undesired embedding entanglement not only results in the reflection of biases from the reference images into the generated images but also notably diminishes the alignment of the generated images with the given generation prompt. To address this challenge, we propose SID (Selectively Informative Description), a text description strategy that deviates from the prevalent approach of only characterizing the subject's class identification. SID is generated utilizing multimodal GPT-4 and can be seamlessly integrated into optimization-based models. We present comprehensive experimental results along with analyses of cross-attention maps, subject-alignment, non-subject-disentanglement, and text-alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Text-to-image diffusion models [5, 7, 11, 25, 30, 35, 40, 42, 43, 46] have demonstrated remarkable generation capabilities that align with textual descriptions across various applications. One such application is text-to-image personalization [12, 24, 45], where models are tailored to generate novel renditions of subjects described in a few reference images. Recent works, such as DreamBooth [45] or Custom Diffusion [24], stand out for their exceptional personalized generation results. They achieve this by fine-tuning a pretrained text-to-image diffusion model with a small number of reference images using a text description in the format of \"a [v] [class name]\" or \"photo of a [v] [class name].\" In this context, [v] represents the subject's unique identifier, which uses a rare token with minimal semantic significance, while" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "type": "text", + "content": "the class descriptor [class name] represents a coarse category for the subject." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 257, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 546, + 449 + ], + "type": "text", + "content": "Recent developments in text-to-image personalization can be categorized into optimization-based approaches [12, 15, 24, 45, 57] and encoder-based approaches [21, 26, 50, 62]. For both types of approaches, a problem typically addressed as 'overfitting' has emerged [2, 15, 21, 26, 45, 57]. It is a phenomenon where objects, other than the subject of interest, in the reference images affect the generated image in an undesired way. For the sake of clarity, we refer to such an entity as 'non-subject' or 'undesired object'. The cause of this phenomenon is closely related to the embedding of the identifier token [v] for optimization-based approaches and the embedding of the subject for encoder-based models, where the information related to any non-subject seeps into the embeddings. Rather than referring to it as overfitting, we formally address this phenomenon as undesired embedding entanglement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "content": "The phenomenon of undesired embedding entanglement has been noticed, but it has been only superficially understood so far. A potential explanation for the limited understanding is the rarity of this phenomenon in straightforward scenarios. Instances where reference images encompass diverse subject views and minimal undesired objects are much less prone to this phenomenon. However, such an ideal scenario is highly unlikely in real-world applications. Therefore, it is crucial to comprehensively understand the biases and develop counter methods to address them effectively. In an effort to gain a more thorough understanding of the prevalent biases contributing to these entanglements, we conducted exhaustive paper reviews followed by extensive experiments utilizing state-of-the-art models. The resulting identification of the five most commonly encountered biases is presented in Fig. 1. While we have chosen DreamBooth [45] for generating the examples in Fig. 1, additional examples for Custom Diffusion [24], SVDiff [15], ELITE [62] and BLIP-Diffusion [26] will be provided later." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "In prior studies, several models [2, 21, 26, 50, 62] have employed segmentation masks. However, depending exclusively on segmentation masks comes with limitations, as we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8312" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 523, + 410 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 523, + 410 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 523, + 410 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 523, + 410 + ], + "type": "image", + "image_path": "742751729213fb74dcc71bfb43b08cf6534b513a477aa7acb8c96feade00a8d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 418, + 548, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 548, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 548, + 475 + ], + "type": "text", + "content": "Figure 1. Five key biases – background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. The first four rows depict scenarios with multiple reference images, while the last row illustrates a single reference image scenario. The pose bias is particularly prone to manifest in scenarios involving a single reference image, although it can also occur when multiple reference images depict the subject in poses that are similar. In the generation prompt, the subject of interest is highlighted in red. The integration of our method into DreamBooth [45] effectively resolves embedding entanglements associated with the five key biases (rightmost column)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 481, + 289, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 289, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 289, + 709 + ], + "type": "text", + "content": "will demonstrate with additional examples in subsequent sections. Notably, these limitations become apparent in the cases of substance, tied-object, and pose biases. To tackle all of the key biases, we introduce SID (Selectively Informative Description) that requires only modifications in the train text descriptions of the reference images. The idea is very simple. Starting with an optimization-based model, we diverge from the prevalent approach of only characterizing the subject's class identification as observed in the existing works. Our approach involves integrating informative specifications of the \"undesired objects\" into the train descriptions. This modification significantly diminishes the probability of undesired entanglement between the subject embedding [v] and the non-subject information in the reference images. As the text-to-image diffusion models are trained for alignments, adding text descriptions that match non-subjects can greatly help the non-subject parts in the reference images to align with the corresponding text descriptions. This prevents the undesired objects from acci" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 481, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 547, + 649 + ], + "type": "text", + "content": "dently becoming aligned with [v]. In contrast, such an accidental alignment can quite easily occur in the absence of the matching text descriptions. Our method is selective because we deliberately avoid incorporating informative specifications of the \"subject\" itself into the train descriptions. Such incorporation could negatively affect the preservation of the subject's identity, as will be elucidated in Sec. 3.1. SID can be readily integrated with any optimization-based model, such as DreamBooth [45], Custom Diffusion [24], or SVDiff [15]. For the automated generation of SID, we utilize the multimodal GPT-4 [37] by guiding it with an appropriate instruction. Surprisingly, the proposed SID integration significantly reduces embedding entanglement, especially in critically biased scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 660, + 392, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 660, + 392, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 660, + 392, + 673 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Text-to-image diffusion models. Diffusion models [19, 54, 55] have demonstrated remarkable generative capabilities across a wide range of modalities, including im" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8313" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "ages [10, 35, 46], audios [8, 22, 28], videos [18, 32, 66], and 3D shapes [23, 68, 69]. Within this diverse spectrum, prominent text-to-image diffusion models, such as Stable Diffusion [43], Imagen [46], and DALL-E 2 [42], have exhibited exceptional proficiency in the synthesis of high-quality images closely aligned with textual descriptions. Subsequently, significant efforts [5, 7, 11, 25, 30, 40] have been dedicated to enhancing the text-to-image alignment in the field of text-to-image generation. Our study, based on the Stable Diffusion model [43], improves the alignment between textual descriptions and images in personalized image synthesis, focusing on accurately preserving subject identity without any entanglement with undesired objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 289, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 435 + ], + "type": "text", + "content": "Vision-language models. The vision-language model (VLM) handles tasks like text-to-image retrieval [61], visual question answering [1], and image captioning [65]. VLMs can also play a crucial role as a building block, as seen with CLIP [41] in tasks such as text-to-image generation [35, 42, 43], image segmentation [31], image retrieval [4], and more [34, 38, 51, 52]. In our work, we need an image captioning VLM capable of closely following detailed instructions. While models like BLIP-2 [27] and OpenFlamingo [3] offer image captioning with text conditioning, they often struggle to adhere closely to detailed instructions. In contrast, large models such as LLaVA [29] and multi-modal GPT-4 [37] have demonstrated their ability to generate precise captions closely aligned with detailed instructions. Among these options, multi-modal GPT-4 exhibits the most powerful capability in generating instruction-following captions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 438, + 289, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 619 + ], + "type": "text", + "content": "Personalized image synthesis. Personalizing text-to-image diffusion models [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62] has become a central focus in text-to-image generation. These models generate high-quality, diverse renditions of a subject using either a few reference images (usually 3 to 7) or just a single reference image. These models can be broadly categorized into two groups: optimization-based and encoder-based models. Optimization-based models [2, 12, 15, 24, 45, 57], like DreamBooth [45], involve per-subject optimization, where they fine-tune identifier tokens or segments of diffusion models to encode a subject's identity. On the other hand, encoder-based models [13, 21, 26, 50, 62] pre-train separate identity encoders and utilize them to encode a subject's identity without the need for per-subject optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 625, + 104, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 104, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 104, + 637 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Our proposed method can be readily integrated with any optimization-based models that utilize description of the reference images. The overall process, presented in Fig. 2, entails the generation of SID (Selectively Informative Description) for each reference image using an instruction-following vision-language model (VLM). These SIDs are" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 330, + 70, + 523, + 215 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 523, + 215 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 523, + 215 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 523, + 215 + ], + "type": "image", + "image_path": "2c8ac0c54cb061e34445d85e41d129667892b555d87d97f6e1b77b4527655688.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 219, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 304, + 219, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 547, + 275 + ], + "type": "text", + "content": "Figure 2. Personalization with SID. We propose integrating SID (Selectively Informative Description) into the per-subject optimization, where an instruction-following VLM (Vison-Language Model) is utilized to generate a selectively informative description for each reference image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 280, + 547, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 547, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 547, + 400 + ], + "type": "text", + "content": "integrated into the reconstruction learning process as part of the per-subject optimization. Depending on the underlying optimization-based model, additional loss functions may be incorporated together with the main reconstruction loss. The additional loss functions, such as the class prior preservation loss [45] introduced in DreamBooth, are applied in the same manner as in their respective baseline models without any modification. The effect of SID is addressed in Sec. 3.1. The choice of VLM and its accompanying instruction are addressed in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 410, + 529, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 529, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 529, + 423 + ], + "type": "text", + "content": "3.1. SID for reducing embedding entanglement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "The existing text-to-image personalization approaches [12, 15, 24, 45], especially optimization-based methods like DreamBooth [45], often rely on simplistic textual descriptions that closely follow the format of \"a [v] [class name]\", where [v] denotes the unique identifier for the subject, in order to extract the subject's identity. In our method, our goal is to come up with train descriptions that have an appropriate level of information such that we can effectively isolate the subject of interest even when the reference images are biased. According to the findings in DreamBooth [45], specifying the subject's correct class in the train description is instrumental in accurately preserving the subject's identity. With the correct class information, the model's prior knowledge of the subject's class can significantly enhance the editing capabilities of the model. Based on the findings, we consider the case where the train description only contains the subject's class identification as the baseline and subsequently explore additionally informative cases that are listed in Tab. 1. In Tab. 1, the term undesired object generally encompasses any object within reference images, excluding the subject of interest. For the examples in Fig. 1, the undesired objects can be readily recognized for background, nearby-object, and tied-object biases. For the substance bias, that is applicable only when performing a style" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8314" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 72, + 187, + 348 + ], + "blocks": [ + { + "bbox": [ + 88, + 72, + 187, + 348 + ], + "lines": [ + { + "bbox": [ + 88, + 72, + 187, + 348 + ], + "spans": [ + { + "bbox": [ + 88, + 72, + 187, + 348 + ], + "type": "image", + "image_path": "2fdc6611ce90fa031679e88d855bd141975442390481107201e5aed552909ee0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 358, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 547, + 392 + ], + "type": "text", + "content": "Figure 3. Two examples for comparing the four cases of descriptions. For each case, the choice of train description follows the guidelines in Tab. 1. The common generation prompt for each example is shown below the generated images. Additional examples can be found in Appendix E.1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 72, + 503, + 348 + ], + "blocks": [ + { + "bbox": [ + 195, + 72, + 503, + 348 + ], + "lines": [ + { + "bbox": [ + 195, + 72, + 503, + 348 + ], + "spans": [ + { + "bbox": [ + 195, + 72, + 503, + 348 + ], + "type": "image", + "image_path": "3f8bbc6b5934b25badf2759903c2c465ce87fde9463ff35ae435662d0530a098.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 399, + 288, + 469 + ], + "blocks": [ + { + "bbox": [ + 48, + 399, + 288, + 469 + ], + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 469 + ], + "type": "table", + "html": "
DescriptionsSubjectUndesired objects
Class IdentificationInformative SpecificationsClass IdentificationInformative Specifications
Case 1 (Baseline)OXXX
Case 2OXOX
Case 3 (Ours)OXOO
Case 4OOOO
", + "image_path": "0c01220a50720b5ab752e446f9b172801b00a1360f89f11e3fb3ac52db7db6d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 475, + 287, + 519 + ], + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 519 + ], + "type": "text", + "content": "Table 1. Four cases of train descriptions. The four cases are categorized based on whether they provide class identification only or additional informative specifications of the subject and undesired objects. We define " + }, + { + "bbox": [ + 46, + 475, + 287, + 519 + ], + "type": "inline_equation", + "content": "SID" + }, + { + "bbox": [ + 46, + 475, + 287, + 519 + ], + "type": "text", + "content": " as the description of Case 3 (Ours)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 530, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 578 + ], + "type": "text", + "content": "re-contextualization, all the objects in the reference images are considered to be undesirable because the intention is to transfer the style only. For pose bias, it is inherent to the subject itself." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "For the four cases in Tab. 1, two examples are investigated in Fig. 3 where two sets of the resulting images are shown. For the simplicity of this investigation, reference images were chosen to share the same set of undesired objects and thus allowing us to use a single train description for all reference images. Case 1 (Baseline) uses the format \"a [v] [class name]\", where only the subject's class identification is included in the train description. In this case, the model often struggles with the bias in the reference images (e.g., the subject can co-appear with undesired objects). This is because of the embedding entanglement in [v] as will be" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 402, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 508 + ], + "type": "text", + "content": "analyze further in Sec. 5. To reduce the undesired embedding entanglement, we introduce Case 2 where the train description contains the class identifications of the undesired objects. Case 3 is similar to Case 2, but the train description additionally contain informative specifications of the undesired objects. As shown in Fig. 3, both Case 2 and Case 3 can provide a significant improvement over Case 1. Case 3, however, often outperforms Case 2 in terms of entanglement reduction as is the case for the two examples in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 508, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 546, + 617 + ], + "type": "text", + "content": "For the subject itself, it is also possible to include informative specifications of the subject. Case 4 corresponds to this possibility. With the inclusion, however, the subject details described in the informative specifications are disentangled from the [v]. This occurs because of the alignment between the subject's informative specifications in the text and the subject details in the images. Because preserving the subject details in the [v] is crucial in personalized image synthesis, we have opted for Case 3 as our definition of SID." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 625, + 444, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 625, + 444, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 625, + 444, + 639 + ], + "type": "text", + "content": "3.2. VLM for generating SID" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Generating SID for each reference image can be demanding and time-consuming for humans. Therefore, we have chosen to utilize image captioning VLM for an automatic generation of SID. To identify a suitable VLM capable of effectively generating SIDs, we evaluated three well-known image captioning VLMs: BLIP-2 [27], LLaVA [29], and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8315" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 74, + 160, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 74, + 160, + 81 + ], + "spans": [ + { + "bbox": [ + 64, + 74, + 160, + 81 + ], + "type": "text", + "content": "Instruction for GPT-4 & LLaVA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 83, + 255, + 88 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 83, + 255, + 88 + ], + "spans": [ + { + "bbox": [ + 64, + 83, + 255, + 88 + ], + "type": "text", + "content": "\"Describe the image in one sentence in detail. Please start" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 88, + 242, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 88, + 242, + 94 + ], + "spans": [ + { + "bbox": [ + 64, + 88, + 242, + 94 + ], + "type": "text", + "content": "with \"A {class}\". *You should not describe the distinct" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 95, + 172, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 95, + 172, + 101 + ], + "spans": [ + { + "bbox": [ + 64, + 95, + 172, + 101 + ], + "type": "text", + "content": "features of the {class} itself.\\*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 105, + 128, + 112 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 105, + 128, + 112 + ], + "spans": [ + { + "bbox": [ + 65, + 105, + 128, + 112 + ], + "type": "text", + "content": "Input text for BLIP-2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 113, + 100, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 113, + 100, + 119 + ], + "spans": [ + { + "bbox": [ + 65, + 113, + 100, + 119 + ], + "type": "text", + "content": "\"A {class}\"" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 121, + 133, + 183 + ], + "blocks": [ + { + "bbox": [ + 61, + 121, + 133, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 121, + 133, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 121, + 133, + 183 + ], + "type": "image", + "image_path": "3c7e53139580e932fcb5376e4e9773760f95fd09aa77a2bf9be83ffb0abe5ac7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 201, + 287, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 201, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 287, + 300 + ], + "type": "text", + "content": "Figure 4. Comparison of three instruction-following VLMs for generating SID. For the reference image of a cat and the instructions shown in the top, the three VLMs generate image captions shown in the right side of the image. Subsequently, the unique identifier [v] is inserted and the resulting captions are used for conditioning the diffusion model as the train descriptions. For painting/cartoon style re-contextualization, we used a slightly different instruction, as detailed in Appendix C.1. Additional fifteen examples for VLM-generated SIDs are available in Appendix E.2." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 122, + 193, + 130 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 122, + 193, + 130 + ], + "spans": [ + { + "bbox": [ + 136, + 122, + 193, + 130 + ], + "type": "text", + "content": "Multimodal GPT-4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 131, + 268, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 131, + 268, + 137 + ], + "spans": [ + { + "bbox": [ + 136, + 131, + 268, + 137 + ], + "type": "text", + "content": "A cat is perched on a tiled surface with a metal railing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 137, + 218, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 137, + 218, + 145 + ], + "spans": [ + { + "bbox": [ + 136, + 137, + 218, + 145 + ], + "type": "text", + "content": "beside it and a backdrop of stairs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 150, + 161, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 150, + 161, + 157 + ], + "spans": [ + { + "bbox": [ + 136, + 150, + 161, + 157 + ], + "type": "text", + "content": "LLaVA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 159, + 270, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 159, + 270, + 165 + ], + "spans": [ + { + "bbox": [ + 136, + 159, + 270, + 165 + ], + "type": "text", + "content": "A small brown and white cat is sitting on a stair railing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 165, + 200, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 165, + 200, + 172 + ], + "spans": [ + { + "bbox": [ + 136, + 165, + 200, + 172 + ], + "type": "text", + "content": "looking down at the floor." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 136, + 178, + 160, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 178, + 160, + 185 + ], + "spans": [ + { + "bbox": [ + 136, + 178, + 160, + 185 + ], + "type": "text", + "content": "BLIP-2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 136, + 186, + 231, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 231, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 231, + 194 + ], + "type": "text", + "content": "A cat is sitting on the stairs looking up." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "content": "multi-modal GPT-4 [37]. In Fig. 4, we illustrate a caption generation example in which the subject is explicitly defined as a cat within the image. For BLIP-2, it struggles to comprehend detailed instructions and it can often produce sequences of meaningless characters. Therefore, we exclusively conditioned BLIP-2 for the subject of interest while providing the detailed instructions to both LLaVA and GPT-4 for generating SID." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 404, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 571 + ], + "type": "text", + "content": "Captions generated by BLIP-2 significantly fall short of the desired SID, lacking informative specifications, i.e., detailed features, of any undesired objects (objects other than the cat in Fig. 4 example) in the images. Although LLaVA generates descriptions better aligned with the instructions, it still falls substantially short, tending to describe the subject excessively or fail to provide sufficient details about undesired objects. Compared to the other two, multi-modal GPT-4 stands out for its ability to follow instructions, delivering descriptions that closely align with the concept of SID. The quantitative comparisons between these VLMs are shown in Appendix E.2, where multi-modal GPT-4 shows the best subject and text alignment. Therefore, we have chosen multi-modal GPT-4 as the VLM for SID generation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 579, + 128, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 592 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "We have performed comprehensive experiments to verify the enhancement resulting from SID. For the experiment datasets, images from previous works [12, 24, 45] and three websites [39, 59, 63] were examined. As the main experiment, we have integrated SID with four optimization-based models - DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion (TI) [12]. The superior effectiveness of SID is readily apparent across a broad spectrum of examples as shown in Fig. 5. In the presence of a single reference image only, we have also compared" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "SID-integrated DreamBooth (Ours) with two encoder-based models, ELITE [62] and BLIP-Diffusion [26], along with DreamBooth in Fig. 6. Implementation details for each model are provided in Appendix C.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 125, + 481, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 125, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 125, + 481, + 139 + ], + "type": "text", + "content": "5. Analysis of cross-attention map" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 140, + 545, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 356 + ], + "type": "text", + "content": "The cross-attention map of each text token is known to highlight the pixels the text token describes [17, 56]. We visualized averaged cross-attention maps for the identifier [v], which is trained to encode the subject's identity. The cross-attention maps are averaged over all the timestamps, layers, and heads to show the single image that can be overlayed on the generated image, following the previous convention [56]. In Fig. 7, we compared cross-attention maps for DreamBooth and SID-integrated DreamBooth across four distinct biases: background, nearby-object, tied-object, and substance biases. While DreamBooth's identifier tends to erroneously focus on undesired objects across all biased scenarios, our method consistently displays highly accurate attention focused on the subject. To extend this analysis, we conducted a comparison between the four cases of train descriptions (detailed in Fig. 3, top row) in Fig. E.2. These analyses confirm that integrating SID does reduce undesired embedding entanglements." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 364, + 478, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 478, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 478, + 376 + ], + "type": "text", + "content": "6. Analysis of three key measures" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 380, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 561 + ], + "type": "text", + "content": "Image-alignment and text-alignment have become popular measures in recent personalized image synthesis [2, 9, 12, 13, 21, 24, 26, 45, 50, 57, 62]. However, we have found that the widely used image-alignment score tends to be significantly influenced by background entanglement, making it an inappropriate measure for analyzing undesired embedding entanglements. Therefore, we first define two customized measures, subject-alignment and non-subject-disentanglement, by adapting image-alignment specifically for text-to-image personalization. Further details can be found in Appendix A, where we demonstrate the crucial importance of these two measures in assessing subject preservation and reduction in undesired embedding entanglement. Using the two measures, along with text-alignment, we conduct a quantitative analysis of SID." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Customizing image-alignment measure. Consider a set of reference images " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "R = \\{r_1, r_2, \\dots, r_N\\}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and a set of generated images " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "G = \\{g_1, g_2, \\dots, g_M\\}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". Each image in " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " is generated using " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " as the reference images with a common generation prompt " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". Define " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_i: \\mathcal{X}_{image} \\to \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_t: \\mathcal{X}_{text} \\to \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " as the CLIP image and text encoders [41] followed by a unit-norm normalization, respectively. Also, define a subject segmentation function " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "s: \\mathcal{X}_{image} \\to \\{0, 1\\}^{H \\times W}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", implemented as Grounded-SAM [14] conditioned on the subject's class name. A mask over the image resolution space is output by " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " where 1 indicates the subject's pixel and 0 indicates the non-subject's pixel." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8316" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 70, + 495, + 430 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 495, + 430 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 495, + 430 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 495, + 430 + ], + "type": "image", + "image_path": "29a71ef2e2182c12642957e543448e7a9b3731e88cbaf3a4795946e68101b429.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 432, + 547, + 478 + ], + "lines": [ + { + "bbox": [ + 46, + 432, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 547, + 478 + ], + "type": "text", + "content": "Figure 5. Enhancement by SID. For four optimization-based models (DreamBooth [45], Custom Diffusion [24], SVDiff [15], and Textual Inversion [12]), the baseline results are shown together with SID-integrated results. SID-integration effectively resolves entanglement issues in scenarios with high biases, represented by indoor background (1st row), nearby potted plant (2nd row), filled-in blueberries (3rd row), and sunflower substances (last row). Additional examples can be found in Appendix E.4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 483, + 267, + 644 + ], + "blocks": [ + { + "bbox": [ + 75, + 483, + 267, + 644 + ], + "lines": [ + { + "bbox": [ + 75, + 483, + 267, + 644 + ], + "spans": [ + { + "bbox": [ + 75, + 483, + 267, + 644 + ], + "type": "image", + "image_path": "71e3fbc91c896c74fc1f0bd35be6130f4a47e4a5d52f18288a77db2d156182b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 650, + 289, + 717 + ], + "lines": [ + { + "bbox": [ + 46, + 650, + 289, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 650, + 289, + 717 + ], + "type": "text", + "content": "Figure 6. Enhancement by SID for a single reference image. Two encoder-based models, ELITE [62] and BLIP-Diffusion [26], are compared with DreamBooth [45] and our SID-integrated DreamBooth. Notably, in the second row, our approach effectively addresses pose bias. Additional examples can be found in Appendix E.5." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "content": "The subject-alignment " + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{SA}" + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "content": " is defined as the average pairwise cosine similarity between generated images and subject segments of reference images in the CLIP-embedding space. The non-subject-disentanglement " + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{NSD}" + }, + { + "bbox": [ + 305, + 483, + 545, + 555 + ], + "type": "text", + "content": " is defined as the 1 minus the non-subject segment similarity in CLIP-embedding space. They can be formally expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 567, + 545, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 567, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 323, + 567, + 545, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {S A} = \\quad A v g \\left(f _ {i} \\left(r _ {n} \\odot s \\left(r _ {n}\\right)\\right) ^ {\\intercal} \\cdot f _ {i} \\left(g _ {m}\\right)\\right), \\tag {1}", + "image_path": "6c523dd3fb1e720f0291a4a488271dadfba7224710dda4b3a926741bbeee3b7d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 582, + 545, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 545, + 595 + ], + "type": "interline_equation", + "content": "\\mathcal {N S D} = 1 - \\operatorname {A v g} \\left(f _ {i} \\left(r _ {n} \\odot (1 - s (r _ {n}))\\right) ^ {\\top} \\cdot f _ {i} \\left(g _ {m}\\right)\\right), \\tag {2}", + "image_path": "431645f1fc725e937dd1c22a8e641b73228cad528a2915c8ffdc9ff7192a7eae.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "inline_equation", + "content": "Avg(\\cdot)" + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "text", + "content": " denotes the average operation over all possible pair selections between " + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "inline_equation", + "content": "n = 1, \\ldots, N" + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "inline_equation", + "content": "m = 1, \\ldots, M" + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "text", + "content": ". Additionally, we evaluate text-alignment " + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{A}" + }, + { + "bbox": [ + 305, + 607, + 547, + 654 + ], + "type": "text", + "content": " shown below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 364, + 666, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 666, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 364, + 666, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\mathcal {A} = A v g \\left(f _ {t} (p) ^ {\\mathrm {T}} \\cdot f _ {i} \\left(g _ {m}\\right)\\right). \\tag {3}", + "image_path": "67b4c964a95df7aa668d21e108af204def3a9638317eaac222b7aa1775e2c113.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 693, + 545, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 693, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 305, + 693, + 545, + 717 + ], + "type": "text", + "content": "The identifier [v] is removed from the generation prompt " + }, + { + "bbox": [ + 305, + 693, + 545, + 717 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 693, + 545, + 717 + ], + "type": "text", + "content": " when evaluating " + }, + { + "bbox": [ + 305, + 693, + 545, + 717 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{A}" + }, + { + "bbox": [ + 305, + 693, + 545, + 717 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8317" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 281, + 296 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 281, + 296 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 281, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 281, + 296 + ], + "type": "image", + "image_path": "9e15a669baf562ad3d9ace55fac4d7bdd29d3fba0f9b85ce28738060f048bfe8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 302, + 287, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 380 + ], + "type": "text", + "content": "Figure 7. Cross-attention map. Averaged cross-attention maps for the identifier [v] are shown. They are overlayed on the generated images. The undesirably entangled embeddings of DreamBooth [45] can be confirmed from the erroneous focusing on the undesirable objects. On the contrary, SID-integrated DreamBooth accurately highlights the subjects of interest: dog, perfume, girl, and painting style. Image credit: David Revoy (3rd row)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 387, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 289, + 674 + ], + "type": "text", + "content": "Analysis of SID. The three measures have straightforward interpretations. A higher subject-alignment indicates superior preservation of the subject's identity. Elevated non-subject-disentanglement values signify a reduction in undesired embedding entanglements. Increased text-alignment suggests that the generated images align more closely with the generation prompt. Using the three measures, we have evaluated the three optimization-based models and their SID-integrated counterparts as shown in Fig. 8. For the single reference images, we evaluated ELITE [62] and BLIP-Diffusion [26] together with the three optimization-based models and the results are shown in Fig. 9. From the figures, it can be observed that all of the best performing, or Pareto optimal, models are the SID-integrated models. The average improvements of the three key metrics exhibit positive values with a substantial margin in all cases, except for subject alignment in Fig. 9. A potential explanation for this observation is the occurrence of subject overfitting in the existing models, particularly evident in cases such as pose bias. The implementation of SID mitigates subject overfitting, even when only a single reference image is provided. For the evaluation, we have utilized commonly used personalization datasets [12, 24, 45] to generate 7500 images. The details are articulated in Appendix C.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Human evaluation. We further assess the effectiveness of SID over its baseline by conducting human evaluation. Human evaluation was performed involving 130 partici" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 424, + 177 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 424, + 177 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 424, + 177 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 424, + 177 + ], + "type": "image", + "image_path": "74159a49e297b735dd6f6f8e0597b4e36fd351fa1a18050b75ddbaa8ccb98660.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 333, + 180, + 526, + 190 + ], + "blocks": [ + { + "bbox": [ + 333, + 180, + 526, + 190 + ], + "lines": [ + { + "bbox": [ + 333, + 180, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 333, + 180, + 526, + 190 + ], + "type": "image", + "image_path": "3a166aa1d6b163a53322304558c161bcb5c28e9aad34025b10c74e041d9e3b7b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 229 + ], + "type": "text", + "content": "Figure 8. Pair-wise metric visualization for multiple reference images. The best performing ones are the SID-integrated models and they form the Pareto boundary." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 428, + 72, + 536, + 176 + ], + "blocks": [ + { + "bbox": [ + 428, + 72, + 536, + 176 + ], + "lines": [ + { + "bbox": [ + 428, + 72, + 536, + 176 + ], + "spans": [ + { + "bbox": [ + 428, + 72, + 536, + 176 + ], + "type": "image", + "image_path": "3856ebb230db49d24dee7641c029e4bd476c9b1723f2011f1d808a70f4521cd0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 243, + 424, + 348 + ], + "blocks": [ + { + "bbox": [ + 312, + 243, + 424, + 348 + ], + "lines": [ + { + "bbox": [ + 312, + 243, + 424, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 424, + 348 + ], + "type": "image", + "image_path": "9c41e3ffff44f26f064264c59be439782923402cbb0b1d45142fcdfa63af8a71.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 368, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 306, + 368, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 545, + 411 + ], + "type": "text", + "content": "Figure 9. Pair-wise metric visualization for single reference image. The best performing ones are the SID-integrated optimization-based models and they form the Pareto boundary. They also outperform the two encoder-based models." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 334, + 353, + 526, + 361 + ], + "blocks": [ + { + "bbox": [ + 334, + 353, + 526, + 361 + ], + "lines": [ + { + "bbox": [ + 334, + 353, + 526, + 361 + ], + "spans": [ + { + "bbox": [ + 334, + 353, + 526, + 361 + ], + "type": "image", + "image_path": "21dd72467e307712903dc2c1f60d676c989f6132aed7cac1868d80ff6bf3f6c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 428, + 244, + 536, + 348 + ], + "blocks": [ + { + "bbox": [ + 428, + 244, + 536, + 348 + ], + "lines": [ + { + "bbox": [ + 428, + 244, + 536, + 348 + ], + "spans": [ + { + "bbox": [ + 428, + 244, + 536, + 348 + ], + "type": "image", + "image_path": "9353a6e4fbcf1a58a6c7a16b4d0926cded8b2cf43d01ea9557c2b073a5557eea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 420, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 492 + ], + "type": "text", + "content": "pants, 10 subjects per participant, and 4 questions per subject (a total of 5,200 responses). Tab. B.2 shows that the survey results for subject-alignment, non-subject disentanglement, and text-alignment are consistent with the metric analysis in Fig. 8, supporting our metric analysis. The results and additional details are provided in Appendix B" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 506, + 375, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 375, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 375, + 518 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 526, + 492, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 492, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 492, + 540 + ], + "type": "text", + "content": "7.1. Negative prompt and segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Negative prompt for classifier-free guidance [20] and segmentation can be considered as the alternatives of SID. Negative prompt is typically used to prevent the generation of unwanted features [49, 58, 64]. Segmentation mask has been used in several previous works for the purpose of isolating subject [2, 21, 26, 50, 62]. Instead of adopting SID, negative prompt can be used during inference or the subject's segmentation mask can be applied before per-subject optimization. In Fig. 10, SID-integration is compared to these two approaches. Negative prompts exhibit unsatisfactory disentanglement, suggesting that severely entangled representations cannot be easily unraveled during inference. Segmentation masks also exhibit its own limitations, and face challenges in dynamically editing the pose" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8318" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 84, + 128, + 151 + ], + "blocks": [ + { + "bbox": [ + 63, + 75, + 124, + 83 + ], + "lines": [ + { + "bbox": [ + 63, + 75, + 124, + 83 + ], + "spans": [ + { + "bbox": [ + 63, + 75, + 124, + 83 + ], + "type": "text", + "content": "Reference images" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 84, + 128, + 151 + ], + "lines": [ + { + "bbox": [ + 61, + 84, + 128, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 84, + 128, + 151 + ], + "type": "image", + "image_path": "2ffc763920d69243daeae35b6e31bc79c977ecaf48d607567888e90e9999d4a2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 152, + 100, + 159 + ], + "lines": [ + { + "bbox": [ + 88, + 152, + 100, + 159 + ], + "spans": [ + { + "bbox": [ + 88, + 152, + 100, + 159 + ], + "type": "text", + "content": "dog" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 141, + 87, + 232, + 148 + ], + "blocks": [ + { + "bbox": [ + 141, + 73, + 214, + 84 + ], + "lines": [ + { + "bbox": [ + 141, + 73, + 214, + 84 + ], + "spans": [ + { + "bbox": [ + 141, + 73, + 214, + 84 + ], + "type": "text", + "content": "Negative prompt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 87, + 232, + 148 + ], + "lines": [ + { + "bbox": [ + 141, + 87, + 232, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 87, + 232, + 148 + ], + "type": "image", + "image_path": "8a483685c79bdf40621fb8ff4006c01affa0c2117355e65573d8aee3fa0f28ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 149, + 298, + 158 + ], + "lines": [ + { + "bbox": [ + 172, + 149, + 298, + 158 + ], + "spans": [ + { + "bbox": [ + 172, + 149, + 298, + 158 + ], + "type": "text", + "content": "A [v] dog sitting in front of the Eiffel Tower." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 72, + 329, + 148 + ], + "blocks": [ + { + "bbox": [ + 239, + 72, + 329, + 148 + ], + "lines": [ + { + "bbox": [ + 239, + 72, + 329, + 148 + ], + "spans": [ + { + "bbox": [ + 239, + 72, + 329, + 148 + ], + "type": "image", + "image_path": "8db23fd8229ccf40eb8ae5bde0af3f73cc0f211aa57ebb76b732376a201a6e70.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 75, + 291, + 83 + ], + "lines": [ + { + "bbox": [ + 277, + 75, + 291, + 83 + ], + "spans": [ + { + "bbox": [ + 277, + 75, + 291, + 83 + ], + "type": "text", + "content": "SID" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 150, + 329, + 158 + ], + "lines": [ + { + "bbox": [ + 239, + 150, + 329, + 158 + ], + "spans": [ + { + "bbox": [ + 239, + 150, + 329, + 158 + ], + "type": "text", + "content": "t of the Eiffel Tower." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 87, + 433, + 148 + ], + "blocks": [ + { + "bbox": [ + 363, + 74, + 410, + 83 + ], + "lines": [ + { + "bbox": [ + 363, + 74, + 410, + 83 + ], + "spans": [ + { + "bbox": [ + 363, + 74, + 410, + 83 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 87, + 433, + 148 + ], + "lines": [ + { + "bbox": [ + 342, + 87, + 433, + 148 + ], + "spans": [ + { + "bbox": [ + 342, + 87, + 433, + 148 + ], + "type": "image", + "image_path": "bb4ad1169d22d74e3890604c713e7476b55d1a4e9184335f883bbfd02c52c5a8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 149, + 520, + 158 + ], + "lines": [ + { + "bbox": [ + 351, + 149, + 520, + 158 + ], + "spans": [ + { + "bbox": [ + 351, + 149, + 520, + 158 + ], + "type": "text", + "content": "A [v] dog howling at the moon on a rocky, barren landscape." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 87, + 529, + 148 + ], + "blocks": [ + { + "bbox": [ + 477, + 73, + 492, + 82 + ], + "lines": [ + { + "bbox": [ + 477, + 73, + 492, + 82 + ], + "spans": [ + { + "bbox": [ + 477, + 73, + 492, + 82 + ], + "type": "text", + "content": "SID" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 87, + 529, + 148 + ], + "lines": [ + { + "bbox": [ + 440, + 87, + 529, + 148 + ], + "spans": [ + { + "bbox": [ + 440, + 87, + 529, + 148 + ], + "type": "image", + "image_path": "f17f27c812c3a173a35581692984016aab0c9d03a214480e833edc682ba72832.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 61, + 167, + 127, + 234 + ], + "blocks": [ + { + "bbox": [ + 61, + 167, + 127, + 234 + ], + "lines": [ + { + "bbox": [ + 61, + 167, + 127, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 167, + 127, + 234 + ], + "type": "image", + "image_path": "6bbdd96fdfd3f9532ec9104d039eaf6556f6dff9d3ac5cc4b2d896aafed791c5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 236, + 106, + 243 + ], + "lines": [ + { + "bbox": [ + 82, + 236, + 106, + 243 + ], + "spans": [ + { + "bbox": [ + 82, + 236, + 106, + 243 + ], + "type": "text", + "content": "perfume" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 252, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 297 + ], + "type": "text", + "content": "Figure 10. Comparison with negative prompt and segmentation. DreamBooth [45] is used as the base model. Compared to the two alternatives, SID-integration demonstrates superior non-subject-disentanglement and text-alignment. Negative prompts used: “sitting on the fluffy blanket and green couch” in the upper row example, “next to a black Prada purse with silver hardware on a bed of white sheets” in the lower row example. Additional examples can be found in Appendix E.6" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 141, + 171, + 232, + 232 + ], + "blocks": [ + { + "bbox": [ + 141, + 171, + 232, + 232 + ], + "lines": [ + { + "bbox": [ + 141, + 171, + 232, + 232 + ], + "spans": [ + { + "bbox": [ + 141, + 171, + 232, + 232 + ], + "type": "image", + "image_path": "1906ca109883cf2b4df74b69bd5a387a73be2ef3bc6de583217c68f387ac66c1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 233, + 279, + 242 + ], + "lines": [ + { + "bbox": [ + 191, + 233, + 279, + 242 + ], + "spans": [ + { + "bbox": [ + 191, + 233, + 279, + 242 + ], + "type": "text", + "content": "A [v] perfume on a park bench." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 239, + 171, + 330, + 232 + ], + "blocks": [ + { + "bbox": [ + 239, + 171, + 330, + 232 + ], + "lines": [ + { + "bbox": [ + 239, + 171, + 330, + 232 + ], + "spans": [ + { + "bbox": [ + 239, + 171, + 330, + 232 + ], + "type": "image", + "image_path": "ff6c1cf1ddd15c88d66f485ccafcf3a376fd94f5a6705d34a63633b9a09e54ef.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 343, + 171, + 433, + 231 + ], + "blocks": [ + { + "bbox": [ + 343, + 171, + 433, + 231 + ], + "lines": [ + { + "bbox": [ + 343, + 171, + 433, + 231 + ], + "spans": [ + { + "bbox": [ + 343, + 171, + 433, + 231 + ], + "type": "image", + "image_path": "4c5109f67ed28acf180985a4fe20d75795baad51048944ecd680774695019b95.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 232, + 504, + 247 + ], + "lines": [ + { + "bbox": [ + 368, + 232, + 504, + 247 + ], + "spans": [ + { + "bbox": [ + 368, + 232, + 504, + 247 + ], + "type": "text", + "content": "A [v] perfume on a windowsill of an old cottage, overlooking a meadow." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 440, + 171, + 529, + 231 + ], + "blocks": [ + { + "bbox": [ + 440, + 171, + 529, + 231 + ], + "lines": [ + { + "bbox": [ + 440, + 171, + 529, + 231 + ], + "spans": [ + { + "bbox": [ + 440, + 171, + 529, + 231 + ], + "type": "image", + "image_path": "ab5f3f350627f168133db623865ac9d6db6a695b26e6f4788a9567a8a4df7c34.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 49, + 320, + 93, + 365 + ], + "blocks": [ + { + "bbox": [ + 87, + 312, + 144, + 319 + ], + "lines": [ + { + "bbox": [ + 87, + 312, + 144, + 319 + ], + "spans": [ + { + "bbox": [ + 87, + 312, + 144, + 319 + ], + "type": "text", + "content": "Reference images" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 320, + 93, + 365 + ], + "lines": [ + { + "bbox": [ + 49, + 320, + 93, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 93, + 365 + ], + "type": "image", + "image_path": "030827ed414845371adad59f08cbf421836c0e863123bc1eb3add85c7923eb0f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 93, + 321, + 138, + 365 + ], + "blocks": [ + { + "bbox": [ + 93, + 321, + 138, + 365 + ], + "lines": [ + { + "bbox": [ + 93, + 321, + 138, + 365 + ], + "spans": [ + { + "bbox": [ + 93, + 321, + 138, + 365 + ], + "type": "image", + "image_path": "3ab1907664b6b52eac2f8efa6eb11d5c24b59d9b50bb6b1e3e3965277fca237d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 139, + 321, + 183, + 365 + ], + "blocks": [ + { + "bbox": [ + 139, + 321, + 183, + 365 + ], + "lines": [ + { + "bbox": [ + 139, + 321, + 183, + 365 + ], + "spans": [ + { + "bbox": [ + 139, + 321, + 183, + 365 + ], + "type": "image", + "image_path": "fe6f37e8b8e2e2376dadbb669012d8b4fe3186b71f1d8e52dd221f8871f44dfe.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 49, + 366, + 92, + 411 + ], + "blocks": [ + { + "bbox": [ + 49, + 366, + 92, + 411 + ], + "lines": [ + { + "bbox": [ + 49, + 366, + 92, + 411 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 92, + 411 + ], + "type": "image", + "image_path": "f79ac9163b52eb4371c483fcbcf98d3c815837de3e7c7a146203dbf8187a974f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 431, + 287, + 486 + ], + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 486 + ], + "type": "text", + "content": "Figure 11. Subject editing. Two examples are shown for editing subject's color. DreamBooth [45] erroneously changes the color of non-subject elements (sofa in the upper row example, background in the lower row example), while SID-integrated DreamBooth accurately edits the subject (dog)." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 94, + 366, + 138, + 411 + ], + "blocks": [ + { + "bbox": [ + 94, + 366, + 138, + 411 + ], + "lines": [ + { + "bbox": [ + 94, + 366, + 138, + 411 + ], + "spans": [ + { + "bbox": [ + 94, + 366, + 138, + 411 + ], + "type": "image", + "image_path": "2fece19a93670996d598e228c477fb785366186b9382e4760fa00ab6b230445b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 412, + 122, + 419 + ], + "lines": [ + { + "bbox": [ + 110, + 412, + 122, + 419 + ], + "spans": [ + { + "bbox": [ + 110, + 412, + 122, + 419 + ], + "type": "text", + "content": "dog" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 139, + 366, + 183, + 411 + ], + "blocks": [ + { + "bbox": [ + 139, + 366, + 183, + 411 + ], + "lines": [ + { + "bbox": [ + 139, + 366, + 183, + 411 + ], + "spans": [ + { + "bbox": [ + 139, + 366, + 183, + 411 + ], + "type": "image", + "image_path": "30709109dbeb4a2149ab014b19331b0ed9c7a27e7b86206d6e1c4ae1db241686.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 195, + 319, + 239, + 365 + ], + "blocks": [ + { + "bbox": [ + 204, + 312, + 232, + 318 + ], + "lines": [ + { + "bbox": [ + 204, + 312, + 232, + 318 + ], + "spans": [ + { + "bbox": [ + 204, + 312, + 232, + 318 + ], + "type": "text", + "content": "DreamBooth" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 319, + 239, + 365 + ], + "lines": [ + { + "bbox": [ + 195, + 319, + 239, + 365 + ], + "spans": [ + { + "bbox": [ + 195, + 319, + 239, + 365 + ], + "type": "image", + "image_path": "e94f49cbbc58a9e23bca4e7794a0f3ef936d67e511401a667626480906162ab6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 195, + 366, + 239, + 410 + ], + "blocks": [ + { + "bbox": [ + 195, + 366, + 239, + 410 + ], + "lines": [ + { + "bbox": [ + 195, + 366, + 239, + 410 + ], + "spans": [ + { + "bbox": [ + 195, + 366, + 239, + 410 + ], + "type": "image", + "image_path": "51b4cba7f5f16f996be2c3f65e356adfdec4c60f390e74671e0a7353ec491ae8.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 242, + 320, + 286, + 365 + ], + "blocks": [ + { + "bbox": [ + 243, + 312, + 285, + 319 + ], + "lines": [ + { + "bbox": [ + 243, + 312, + 285, + 319 + ], + "spans": [ + { + "bbox": [ + 243, + 312, + 285, + 319 + ], + "type": "text", + "content": "DreamBooth + SID" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 320, + 286, + 365 + ], + "lines": [ + { + "bbox": [ + 242, + 320, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 242, + 320, + 286, + 365 + ], + "type": "image", + "image_path": "95e92908269282786772cc715f14ef776bda051f39305502f8ea8720825461e8.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 242, + 366, + 286, + 411 + ], + "blocks": [ + { + "bbox": [ + 216, + 412, + 265, + 420 + ], + "lines": [ + { + "bbox": [ + 216, + 412, + 265, + 420 + ], + "spans": [ + { + "bbox": [ + 216, + 412, + 265, + 420 + ], + "type": "text", + "content": "A purple [v] dog" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 366, + 286, + 411 + ], + "lines": [ + { + "bbox": [ + 242, + 366, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 242, + 366, + 286, + 411 + ], + "type": "image", + "image_path": "a07a0243035f6f47c7c2723e8cbba4f6f2c9c7721b70d46319113ddd7997ab70.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 496, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 286, + 533 + ], + "type": "text", + "content": "of the subject or faithfully following the generation prompt. SID shows higher non-subject-disentanglement and text-alignment than the two alternatives." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 47, + 536, + 192, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 192, + 549 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 192, + 549 + ], + "type": "text", + "content": "7.2. Enhancing subject editing" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 550, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 669 + ], + "type": "text", + "content": "The identifier [v], specifically trained to encode the subject's identity, enables us to modify the properties of the subject through a textual guidance. Undesired embedding entanglements, however, can lead to misaligned modifications. In Fig. 11, an example of subject's color editing is shown. DreamBooth [45], influenced by embedding entanglements, fails to edit the subject and instead modifies the color of an undesired object or background. In contrast, SID-integrated DreamBooth precisely modifies the subject itself, following the provided color guidance." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 47, + 673, + 157, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 157, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 157, + 685 + ], + "type": "text", + "content": "7.3. Limitations of SID" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "A limitation of our approach is associated with the imperfections of VLM. While the multi-modal GPT-4 [37] gen" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 465 + ], + "type": "text", + "content": "erally generates descriptions that closely align with the provided instructions, occasional failures do occur. Although these instances are rare and may not be easily generalized, we have listed some of these failure cases in Appendix D.1. Another limitation is the undesired entanglement of the subject's strong facial expressions. We discovered that this failure arises when the VLM-generated SID lacks informative specifications of facial expressions. This issue can be addressed by including undesired information in the SID, a solution detailed further in Appendix D.2. Finally, SID is exclusively integrated with optimization-based models in this study. Yet, there is potential for its integration into encoder-based models, particularly during encoder pre-training." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 378, + 489 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 496, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 627 + ], + "type": "text", + "content": "In this study, we have introduced a robust strategy to mitigate undesired embedding entanglement in text-to-image personalization. Beginning with the identification of five key biases, we have proposed SID (Selectively Informative Description) as an effective solution to address these biases. Our cross-attention analysis demonstrates the successful removal of entanglements, while alignment analysis indicates notable enhancements in non-subject-disentanglement and text-alignment. The proposed SID strategy holds potential applicability to other multi-modal applications where managing embedding entanglements is crucial." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the following grants funded by the Korea government: NRF (NRF2020R1A2C2007139, NRF-2022R1A6A1A03063039) and IITP ([NO.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University)], [No. RS2023-00235293])." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8319" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 145, + 288, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 145, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 145, + 288, + 187 + ], + "type": "text", + "content": "[2] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 188, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 188, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 288, + 243 + ], + "type": "text", + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 297 + ], + "type": "text", + "content": "[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 328 + ], + "type": "text", + "content": "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 382 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 382, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 288, + 425 + ], + "type": "text", + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "text", + "content": "[8] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "type": "text", + "content": "[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Rui, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. arXiv preprint arXiv:2304.00186, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 542, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 287, + 596 + ], + "type": "text", + "content": "[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 596, + 287, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 596, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 48, + 596, + 287, + 650 + ], + "type": "text", + "content": "[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 650, + 287, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 650, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 650, + 287, + 693 + ], + "type": "text", + "content": "[13] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 693, + 287, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 287, + 703 + ], + "type": "text", + "content": "[14] Grounded-SAM Contributors. Grounded-Segment-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 105 + ], + "type": "text", + "content": "Anything. https://github.com/IDEA-Research/Grounded-Segment-Anything，Apr.2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 106, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 148 + ], + "type": "text", + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 148, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 545, + 191 + ], + "type": "text", + "content": "[16] Haotian Liu. liuhaotian/llava-v1-0719-336px-lora-merge- vicuna-13b-v1.3. https://huggingface.co/ liuhaotian/llava-v1-0719-336px-lora- merge-vicuna-13b-v1.3, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 191, + 545, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 191, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 191, + 545, + 234 + ], + "type": "text", + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 234, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 545, + 288 + ], + "type": "text", + "content": "[18] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 288, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 320 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 320, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 341 + ], + "type": "text", + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 341, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 545, + 395 + ], + "type": "text", + "content": "[21] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 395, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 545, + 426 + ], + "type": "text", + "content": "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[23] Juil Koo, Seungwoo Yoo, Minh Hieu Nguyen, and Minhyuk Sung. Salad: Part-level latent diffusion for 3d shape generation and manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14441-14451, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 480, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 480, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 480, + 545, + 533 + ], + "type": "text", + "content": "[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 533, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 587 + ], + "type": "text", + "content": "[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 587, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 630 + ], + "type": "text", + "content": "[26] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 631, + 545, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 631, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 308, + 631, + 545, + 673 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 673, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 673, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 673, + 545, + 707 + ], + "type": "text", + "content": "[28] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audi-Oldm: Text-to-audio generation with latent diffusion models." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8320" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 72, + 214, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 214, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 214, + 83 + ], + "type": "text", + "content": "arXiv preprint arXiv:2301.12503, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 83, + 287, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 83, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 287, + 114 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 115, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 115, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 115, + 287, + 158 + ], + "type": "text", + "content": "[30] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "type": "text", + "content": "[31] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 287, + 266 + ], + "type": "text", + "content": "[32] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 266, + 286, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 286, + 287 + ], + "type": "text", + "content": "[33] Makoto Shing. svdiff-pytorch. https://github.com/mkshing/svdiff-pytorch, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "text", + "content": "[34] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 330, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 287, + 384 + ], + "type": "text", + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 384, + 286, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 286, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 286, + 405 + ], + "type": "text", + "content": "[36] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 405, + 208, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 208, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 208, + 415 + ], + "type": "text", + "content": "[37] OpenAI. Gpt-4 technical report, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "type": "text", + "content": "[38] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 468, + 285, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 285, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 285, + 488 + ], + "type": "text", + "content": "[39] Peper and Carrot. https://www.peppercarrot.com/." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "text", + "content": "[40] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. arXiv preprint arXiv:2306.05427, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 521, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 585 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 628, + 287, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 287, + 682 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 706 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "text", + "content": "tation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 117, + 545, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 117, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 308, + 117, + 545, + 181 + ], + "type": "text", + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 181, + 545, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 181, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 308, + 181, + 545, + 246 + ], + "type": "text", + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 246, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 246, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 545, + 278 + ], + "type": "text", + "content": "[47] Salesforce. LAVIS. https://github.com/salesforce/LAVIS/tree/main/projects/blip-diffusion, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 278, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 278, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 308, + 278, + 545, + 310 + ], + "type": "text", + "content": "[48] Salesforce. Salesforce/blip2-opt-2.7b. https://huggingface.co/Salesforce/blip2-opt-2.7b, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "text", + "content": "[49] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 364, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 545, + 396 + ], + "type": "text", + "content": "[50] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 396, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 545, + 427 + ], + "type": "text", + "content": "[51] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "text", + "content": "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 481, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 513 + ], + "type": "text", + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 513, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 545 + ], + "type": "text", + "content": "[54] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 545, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 587 + ], + "type": "text", + "content": "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 587, + 545, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 640 + ], + "type": "text", + "content": "[56] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 641, + 545, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 641, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 308, + 641, + 545, + 685 + ], + "type": "text", + "content": "[57] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 685, + 545, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 685, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 308, + 685, + 545, + 706 + ], + "type": "text", + "content": "[58] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8321" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 512 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 225, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 225, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 225, + 115 + ], + "type": "text", + "content": "[59] Unsplash. https://unsplash.com/." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 116, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 116, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 49, + 116, + 287, + 157 + ], + "type": "text", + "content": "[60] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 159, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 159, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 49, + 159, + 287, + 212 + ], + "type": "text", + "content": "[61] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5764-5773, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 213, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 213, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 287, + 255 + ], + "type": "text", + "content": "[62] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 255, + 234, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 255, + 234, + 265 + ], + "spans": [ + { + "bbox": [ + 49, + 255, + 234, + 265 + ], + "type": "text", + "content": "[63] Wikiart. https://www.wikiart.org/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 266, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 266, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 266, + 287, + 319 + ], + "type": "text", + "content": "[64] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 319, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 287, + 372 + ], + "type": "text", + "content": "[65] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 373, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 373, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 49, + 373, + 287, + 404 + ], + "type": "text", + "content": "[66] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 405, + 286, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 286, + 425 + ], + "type": "text", + "content": "[67] Yuxiang Wei. ELITE. https://github.com/csyxwei/ELITE, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "type": "text", + "content": "[68] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 468, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 287, + 512 + ], + "type": "text", + "content": "[69] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826-5835, 2021." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8322" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_content_list.json b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bef6472947692ac06572b1ba499b2c7d9901a82a --- /dev/null +++ b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_content_list.json @@ -0,0 +1,1712 @@ +[ + { + "type": "text", + "text": "Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution", + "text_level": 1, + "bbox": [ + 107, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingping Zheng $^{1}$ , Ling Zheng $^{3}$ , Yuanfan Guo $^{2}$ , Ying Li $^{1*}$ , Songcen Xu $^{2}$ , Jiankang Deng $^{2}$ , Hang Xu $^{2}$", + "bbox": [ + 91, + 179, + 880, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northwestern Polytechnical University $^{2}$ Huawei Noah's Ark Lab $^{3}$ Tsinghua-Fuzhou Institute for Data Technology", + "bbox": [ + 129, + 202, + 839, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 253, + 313, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Artifact-free super-resolution (SR) aims to translate low-resolution images into their high-resolution counterparts with a strict integrity of the original content, eliminating any distortions or synthetic details. While traditional diffusion-based SR techniques have demonstrated remarkable abilities to enhance image detail, they are prone to artifact introduction during iterative procedures. Such artifacts, ranging from trivial noise to unauthentic textures, deviate from the true structure of the source image, thus challenging the integrity of the super-resolution process. In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free method that delves into the latent space to effectively identify and mitigate the propagation of artifacts. Our SARGD begins by using an artifact detector to identify implausible pixels, creating a binary mask that highlights artifacts. Following this, the Reality Guidance Refinement (RGR) process refines artifacts by integrating this mask with realistic latent representations, improving alignment with the original image. Nonetheless, initial realistic-latent representations from lower-quality images result in over-smoothing in the final output. To address this, we introduce a Self-Adaptive Guidance (SAG) mechanism. It dynamically computes a reality score, enhancing the sharpness of the realistic latent. These alternating mechanisms collectively achieve artifact-free super-resolution. Extensive experiments demonstrate the superiority of our method, delivering detailed artifact-free high-resolution images while reducing sampling steps by $2 \\times$ . We release our code at https://github.com/ProAirVerse/Self-Adaptive-Guidance-Diffusion.git.", + "bbox": [ + 76, + 286, + 473, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 771, + 209, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Super-resolution (SR) techniques strive to restore high-resolution images from their corresponding low-resolution counterparts [8, 29]. This process is inherently ill-posed since the same LR image can yield many plausible HR solutions. Traditional deep learning SR methods, leverag", + "bbox": [ + 75, + 797, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4bc0e146e9d72cf6e9167554118e0cf422a8e20cee50d253a888196a3075801.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 255, + 583, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61918e4d994c8fcd04bb49380a6364cafa7e4125f75448ac3b62f3899d40837d.jpg", + "image_caption": [ + "(a) Variation of Artifacts within StableSR" + ], + "image_footnote": [], + "bbox": [ + 584, + 257, + 658, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10c72e833a6efd78c3de57131df0fcf4ec3bf7e204670035b6db38179fa226b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 256, + 733, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1941dd32834c13a109fa996d392e0e40a6d0d8a0705dbfb89d4792dbd9b18cfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 257, + 807, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7dd435689f20a6d3d9c8e9954231b535e63e905a920bb27e4308fcc90ed593d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 257, + 882, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/914bf025e0b8d72ea5f09722a1c206a393ae764afc303d4d4acbb19fc8e97268.jpg", + "image_caption": [ + "Figure 1. Visual comparison of artifact variability between (a) StableSR and (b) Our SARGD method. In the lower panel, the left image is the decoded output, while the right image is the artifact mask. Regions containing artifacts are highlighted in red, utilizing a binary map to serve as the artifact mask. Our SARGD method exhibits a superior capability in reducing artifacts." + ], + "image_footnote": [], + "bbox": [ + 509, + 361, + 583, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24e794b3084e1e59c85d0c8c1380d602d7a87f77a78795be268d5e580938e193.jpg", + "image_caption": [ + "(b) Variation of Artifacts within Our SARGD" + ], + "image_footnote": [], + "bbox": [ + 584, + 361, + 658, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/059b17e30a51c8119adc7823dcd04ef72843909874b928bbeeadb49015c44df0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 361, + 733, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c203cbce8b899368523e55fe7ef36be34de158b6f3297a981f435e2b0f208b8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 361, + 807, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03022f297f97a931630f37e10115df87b7bf97548f27eaae120e5fec916dcae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 361, + 882, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing techniques like squeeze layers [49], transposed convolutions [38], and sub-pixel convolutions [29], bind themselves to fixed upsampling scales, limiting output resolution adjustments without architectural changes. This causes inflexibility in real-world applications. To tackle the inherent one-to-many inverse challenge in SR, it is intuitive to envision the solution space as a distribution. Yet, this perspective often prompts generative models to introduce details that deviate from the authentic content of the original low-resolution image. As a result, a wide range of generative-based SR approaches have emerged to address this intrinsically ill-posed problem, particularly in diffusion models.", + "bbox": [ + 496, + 551, + 892, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based SR (Diffusion-SR for short) has exhibited significant potential in improving the resolution of images from their low-resolution counterparts. SRDiff [17] is the first diffusion-based model for image super-resolution, progressively transforming the Gaussian noise into a high-resolution (HR) image conditioned on its low-resolution (LR) input through a Markov chain. Building on the foundational concept of denoising score matching [31], Saharia et al. [28] propose SR3 model. This model advances the field of super-resolution by learning an empirical data distribution and leveraging U-Net architec", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25806", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ture [27] to progressively denoise and refine the image output. However, these methods fail to preserve accurate pixelwise image structures, often necessitating further training or additional skipped connections for detail reproduction. While the pixel-aware stable diffusion network proposed by Yang et al. [37] significantly improves the robustness of the super-resolution process, it concurrently introduces the challenge of generating implausible content. In contrast, StableSR [32], the current state-of-the-art, leverages pre-trained diffusion priors to retain image details, avoiding assumptions about image degradation.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, as depicted in Figure 1, contemporary diffusion-based StableSR models invariably generate implausible content, frequently resulting in perceptual artifacts within specific areas of the image. Indeed, these artifacts frequently stem from the fundamental use of Denoising Diffusion Probabilistic Models (DDPMs) [13, 30] in modern diffusion models. Specifically, DDPMs, categorized as Partial Differential Equation (PDE)-based models [6], are designed to diffuse pixel intensities in order to reduce noise levels. They can enhance the preservation of edges and textural features by integrating non-linear terms, thus improving image details. However, despite these advancements, the introduction of non-linearity can occasionally yield unexpected outcomes, particularly in areas with intricate textures or noise patterns. This unintended consequence can result in a loss of fine details and a degradation of essential image elements such as edges and textures, especially when the diffusion process is overly aggressive or inadequately constrained. Additionally, any deficiencies in the model's architecture or parameter settings can be amplified during diffusion, aggravating artifact problems and producing results that deviate from the desired photorealistic quality [36]. These challenges pose a significant obstacle to the preservation of the authentic detail and clarity expected in high-resolution images.", + "bbox": [ + 75, + 261, + 472, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a pioneering training-free approach, aiming to produce \"artifact-free\" SR images. In particular, SARGD operates through two stages: Reality-Guided Refinement (RGR) to reduce artifacts and Self-Adaptive Guidance (SAG) to boost image fidelity. During the initial RGR phase, an artifact detector is employed to identify unrealistic pixels within the latent space at each inference step, leading to a binary artifact mask. This mask collaborates with a realistic latent guidance extracted from an upscaled LR image to refine artifacts and enhance intrinsic details. However, depending on upscaled LR inputs for guidance can potentially result in over-smoothed images, as it may overlook high-frequency details that are essential for SR. To alleviate this limitation, the subsequent SAG phase introduces a 'reality score to self-adjust the authenticity of the latent guidance, significantly boosting the accuracy of", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "detail and texture representation. Two phases alternate to progressively achieve high-fidelity super-resolution, effectively eliminating artifacts in the diffusion process. To sum up, the contributions of this paper are as follows:", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a Reality-Guided Refinement (RGR) strategy, designed to identify and rectify artifacts during the diffusion process with the aid of a reality guidance mechanism, effectively eliminating such artifacts.", + "- We propose a Self-Adaptive Guidance (SAG) mechanism that iteratively enhances the realistic-latent guidance, addressing over-smoothing issues and improving the fidelity and authenticity of the reference latent guidance.", + "- We are the first to develop the Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free approach that sets a new benchmark by effectively dealing with artifacts and over-smoothing issues. Additionally, it outperforms StableSR in image quality while reducing the inference time by $2 \\times$ for super-resolution tasks." + ], + "bbox": [ + 500, + 152, + 890, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 378, + 640, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we summarize past super-resolution (SR) techniques, explore diffusion-based models in SR, and address artifact detection and mitigation in the SR process.", + "bbox": [ + 496, + 404, + 890, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Super-Resolution. Single image super-resolution (SISR) is a process designed to construct high-resolution (HR) images from their low-resolution (LR) counterparts, which is inherently a complex task due to the multiple HR images that can correspond to a single LR image. The literature has seen a variety of solutions aiming to learn mappings from LR to HR images. These solutions fall into two primary groups: PSNR-oriented methods and generative model-based methods. PSNR-oriented methods [8, 9, 18, 20, 29, 48] employ L1 or L2 norms as objectives, achieving notable PSNR metrics. However, the reliance on L1/L2 losses tends to predict an average of all possible HR images [15, 16, 22], resulting in oversmoothed outputs. In contrast, generative model-based methods [2, 12, 16, 17, 19, 22, 23, 33, 34, 42, 45] seek to address the ill-posed nature of the SR problem by learning the distribution of potential HR images through techniques such as GAN-based, diffusion-based, flow-based, and AR-based super-resolution. These models aim to generate HR images that not only have high pixel accuracy but also possess the rich textural details necessary for visual realism.", + "bbox": [ + 496, + 455, + 892, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion probabilistic models [13, 30] are a class of generative models that utilize a Markov chain to convert latent variables from simple distributions (e.g., Gaussian) into data with complex distributions. Recognizing the potential of these models to address \"one-to-many\" mapping problems, Li et al. [17] have applied them to the task of super-resolution, generating a variety of high-resolution images from a single low-resolution input. This", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25807", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c10dfa7f22ffd95d0957a7538d16645962da4fb93305e5c04c798041f0dfc69.jpg", + "image_caption": [ + "Figure 2. Self-Adaptive Reality-Guided Diffusion (SARGD) for artifact-free super-resolution. Our proposed SARGD is a training-free approach that consists of two principal components: 1) a Reality-Guided Refinement (RGR) that identifies and corrects artifacts within the latent representation by using a realistic latent as a guide to maintain the inherent details of the original image during the diffusion process, and 2) a Self-Adaptive Guidance (SAG) mechanism that enhances the fidelity of the initial realistic latent guidance, derived from upscaled low-resolution images, thereby effectively addressing the issue of over-smoothing in the final outputs." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 885, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "application not only yields diverse outcomes but also simultaneously addresses challenges related to over smoothing, mode collapse, and the extensive computational footprint that are often encountered in super-resolution tasks. A prevalent method [26, 28] in diffusion-based SR involves integrating a low-resolution image directly into the input of existing diffusion models (e.g., DDPM [13]), and retraining the model with super-resolution data from scratch. Another approach [4, 5, 40] leverages an unconditional pretrained diffusion model as a prior and adapts its reverse diffusion process to synthesize the desired high-resolution image. Nonetheless, both strategies are constrained by the intrinsic Markov chain sequence of DDPM, which can potentially lead to artifacts during inference, thus affecting the quality of the resulting high-resolution image.", + "bbox": [ + 75, + 398, + 472, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Artifact Detection Methods. Despite impressive improvement, diffusion-based super-resolution methods tend to present undesired outcomes, known as artifacts. The artifact detection approach aims to probe the perceptually implausible contents rooted in the generated images. Existing deep-learning-based image artifact detection methods can be divided into spatial and frequency domain-based approaches. The former approaches [39] focus on capturing differences in texture between real and generated images. Liu et al. [21] leverage Gram matrices to capture the difference in texture between real and synthetic images. Dang et al. [7] adopt the attention mechanism to improve the accuracy of artifact detection by highlighting informative regions. Zhao et al. [50] utilize the attention mechanism to amplify subtle differences in the shallow layers, improving the detection performance. The latter approaches [46] study the artifacts in the frequency domain. Durall et al. [10] classify generated images according to the characteristics", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the artifacts in high-frequency components. Dzanic et al. [11] demonstrate that image artifacts could be detected by leveraging the degree of partial decay at high frequencies. Frank et al. [11] find that the use of upsampling operations causes the artifacts in generated images, and employs the DCT transforms to detect them.", + "bbox": [ + 496, + 398, + 893, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 502, + 635, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose a novel training-free method called \"Self-Adaptive Reality-Guided Diffusion\" (SARGD) to tackle the ongoing challenge of artifact suppression in diffusion-based super-resolution. This approach, depicted in Figure 2, introduces two essential mechanisms.", + "bbox": [ + 496, + 527, + 893, + 602 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Reality-Guided Refinement (RGR), mitigates artifact emergence during inference by employing guidance from a realistic latent space, thereby preserving the original characteristics of the images.", + "- Self-Adaptive Guidance (SAG), enhances the authenticity of the reference realistic latent representation, effectively addressing the over-smoothing tendencies associated with the initial realistic latent representation." + ], + "bbox": [ + 500, + 603, + 893, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 732, + 640, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based super-resolution (SR) techniques [17, 28, 32] have emerged as a compelling paradigm for the enhancement of image resolutions. Consider the diffusion process as a function that aims to translate a given low-resolution (LR) image into a high-resolution (HR) image, each successive HR image at time $t + 1$ is obtained by refining the current estimate, factoring in the original LR image. This refinement is mathematically represented as:", + "bbox": [ + 496, + 755, + 893, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {H R} ^ {(t + 1)} = \\mathbf {I} _ {H R} ^ {(t)} + f _ {\\theta} (\\mathbf {I} _ {H R} ^ {(t)}, \\mathbf {I} _ {L R}), \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 883, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25808", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/492e361481c399e45d8b42396f08e58bbb67f5d74a33f924507e52bcbb0c5ed1.jpg", + "image_caption": [ + "Figure 3. Overview of the Reality-Guided Refinement (RGR) workflow: 1) Decoding the current latent into an RGB image and using an artifact detector to create a binary mask identifying areas with artifacts; and 2) Utilizing realistic latent guidance to refine the masked regions, enhancing the image's fidelity and authenticity." + ], + "image_footnote": [], + "bbox": [ + 96, + 95, + 452, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_{HR}^{(t)}$ is the current state of the high-resolution image at iteration $t$ . $f_{\\theta}$ is the diffusion function parameterized by $\\theta$ , which updates the current high-resolution state based on both its current state and the given low-resolution image. $t$ is the current iteration during the inference.", + "bbox": [ + 75, + 354, + 468, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the iterative diffusion process for super-resolution, the input LR-HR image pairs in the training set are used to train the diffusion models with the total diffusion step T. The diffusion process initially leverages a pre-trained visual encoder to convert the upsampled LR image into a latent $z_{t} = \\mathcal{E}(up(\\mathbf{I}_{LR}))$ , and gradually adding the Gaussian noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ into it, resulting the noised latent variable $x$ . The reverse diffusion process aims to guide the generation to the corresponding HR image through iterative refinement of the noised latent $x$ , taking the form of:", + "bbox": [ + 75, + 434, + 468, + 584 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t - 1} \\leftarrow \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {y} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\gamma_ {t}}} f _ {\\theta} \\left(\\boldsymbol {x}, \\boldsymbol {y} _ {t}, \\gamma_ {t}\\right)\\right) + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 595, + 468, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{t} \\in (0,1)$ is the corresponding coefficient and $\\gamma_{t}$ is the variance of the noise at iteration $t$ . Lastly, the decoder $\\mathcal{D}$ is employed to project the resulting latent into the final high-resolution image.", + "bbox": [ + 75, + 638, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Reality-Guided Refinement (RGR)", + "text_level": 1, + "bbox": [ + 76, + 709, + 379, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome the prevalent dilemma of aggravated noise and the emergence of artifacts in images subjected to superresolution (SR) via diffusion-based techniques, we propose a pioneering training-free approach: Reality-Guided Refinement (RGR). This method couples an artifact detector with a latent refinement to address these issues effectively. As illustrated in Figure 3, a pixel-level segmentation model is initially applied to locate any artifacts on the current latent variable. Once detected, these imperfections are adeptly rectified through a sophisticated procedure of realistic-latent guidance refinement. This diffusion process", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ensures that the resulting images retain the original, authentic textures and intricate details, effectively circumventing the usual distortions and artifacts associated with traditional diffusion-based upsampling techniques for SR tasks.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Artifact Detection in Diffusion Process. In the context of pixel-sensitive super-resolution, it is crucial to identify the fine-grained artifacts during the diffusion process. Hence, a robust artifact detection system should precisely pinpoint these implausible pixels within the latent space for effective identification. To achieve this, we utilize the established artifact detection method named PAL [43] to isolate these imperfections in the latent space. The process entails transforming the latent variable into the image domain at each iteration using the decoder, depicted by the following equation:", + "bbox": [ + 496, + 155, + 890, + 321 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {H R} ^ {(t - 1)} = \\mathcal {D} \\left(\\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {x}, t)\\right) + \\sigma_ {\\theta} \\left(x _ {t}, t\\right) \\boldsymbol {\\epsilon})\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 330, + 890, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the resulting image is input into the PAL detector, yielding a binary mask $E_{A}$ that denotes the presence of artifacts (referenced as line 12 in Algorithm 1). This process can be formulated as:", + "bbox": [ + 496, + 372, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {A} ^ {(t - 1)} = \\mathcal {A} \\left(\\mathbf {I} _ {H R} ^ {(t - 1)}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 439, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the resultant artifact mask is resized to match the scale of the latent variable, maintaining the proportionality between the two.", + "bbox": [ + 496, + 467, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Refinement via Reality Guidance. For the removal of artifacts during sampling, we refine the latent variable via a proposed training-free RGR. Following Algorithm 1, the RGR inference, consisting of T steps, begins by taking an upscaled low-resolution image $up(\\mathbf{I}_{LR})$ as the starting point. A noise vector $\\epsilon$ is then drawn from a standard Gaussian distribution. The process iterates backward from $t = T$ , with each step generating a latent variable $\\boldsymbol{x}_{t-1}$ in accordance with DDPM [13] framework. Once artifacts $E_A$ are identified within the current latent $\\boldsymbol{x}_{t-1}$ , we proceed to refine these imperfections through reality guidance. This correction process unfolds as follows:", + "bbox": [ + 496, + 516, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t - 1} = \\boldsymbol {x} _ {t - 1} \\times (1 - E _ {A}) + \\boldsymbol {x} _ {r} \\times E _ {A}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 707, + 890, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This process allows us to effectively eliminate artifacts from the latent space. Notably, in this formula, we employ the initially encoded latent, which has not undergone the denoising process, as a source of realistic latent guidance, since it retains the intrinsic structural integrity of the image.", + "bbox": [ + 496, + 732, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Self-Adaptive Guidance (SAG)", + "text_level": 1, + "bbox": [ + 500, + 816, + 769, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, relying on the initial latent space as a guide for artifact suppression during inference often results in oversmoothing results, primarily attributed to reliance on upscaled low-resolution images. To address this, we propose a", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25809", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "novel Self-Adaptive Guidance (SAG) renewal mechanism, an automated system that perpetually refines realistic latent representations, ensuring the guidance is continually calibrated to enhance realistic detail. This mechanism is indispensable for diffusion-based super-resolution, aiming to yield high-resolution and high-fidelity images that strike a balance between details and sharpness.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reality Score Computation. To regulate the update mechanism of reality guidance in image super-resolution, we introduce an innovative metric known as the reality score. This metric quantifies the realism of latent guidance. As delineated in Algorithm 1, we begin with an artifact-free latent variable $\\pmb{x}_{t-1}$ and utilize our decoding function to synthesize a realistic image $\\mathbf{I}_r = \\mathcal{D}(\\pmb{x}_{t-1})$ . Subsequently, we deploy a realistic-region function $\\mathcal{R}$ by inverting the output of PAL [43] on $\\mathbf{I}_r$ , resulting in a binary reality mask $M_R$ with dimensions corresponding to the channels, height, and width of the latent $C \\times H \\times W$ , respectively. The reality score $s_r$ at the current $t-1$ step is calculated as the normalized sum of the reality mask across the spatial dimensions of the latent, indicating the distribution of realistic features. The formula is given by:", + "bbox": [ + 75, + 199, + 472, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {r} ^ {t - 1} = \\mathcal {S} \\left(M _ {R}\\right) = \\frac {1}{H \\times W} \\sum_ {i = 0} ^ {H - 1} \\sum_ {j = 0} ^ {W - 1} M _ {R} ^ {(i j)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 431, + 468, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Herein, $H$ and $W$ are the height and width of the latent, and $M_R^{(ij)}$ indicates the reality measure at each pixel. This approach ensures a quantitative evaluation of the SR process, focusing on producing images that boast both high resolution and a high degree of fidelity to reality.", + "bbox": [ + 76, + 474, + 468, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-Adaptive Reality Guidance. Self-adaptability in the realistic latent space refers to the ability of the system to adjust the reality guidance to correct the details and textures that are characteristic of high-quality images. This is especially important to mitigate the smoothing effect of the initial realistic-latent reliance on bicubic upsampling. As outlined in Algorithm 1, we begin by considering a realistic image $\\mathbf{I}_r$ at the current step, decoded from an artifact-free latent variable $x_{t - 1}$ . This artifact-free image is then encoded into a realistic latent $x_{r}^{t - 1} = \\mathcal{E}(\\mathbf{I}_r)$ . Grounded on the computed reality score, self-adaptability is characterized as a guidance refinement process. During this process, the realistic latent $\\pmb{x}_r$ at the current $t - 1$ step is iteratively refined to converge towards an ideal latent representation $\\pmb{x}_r$ , precisely aligning with the high-resolution output IHR. The equation below illustrates this adaptive mechanism:", + "bbox": [ + 75, + 556, + 472, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {x} _ {r} ^ {t - 1}\\right) = \\boldsymbol {x} _ {r} \\times \\left(1 - M _ {R}\\right) + \\boldsymbol {x} _ {r} ^ {t - 1} \\times M _ {R}, \\text {i f} \\boldsymbol {s} _ {r} ^ {t - 1} > \\boldsymbol {s} _ {r}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 801, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_r$ represents the reality score associated with $xr$ , and $M_R$ is the binary reality mask corresponding to $xr^{t-1}$ . This self-adaptive mechanism ensures that the guidance is autonomously updated, leading to enhanced superresolution outcomes.", + "bbox": [ + 76, + 825, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Self-Adaptive Reality-Guided Diffusion" + ], + "code_body": "1: Input: LR image $\\mathbf{I}_{LR}$ , and total diffusion steps $T$ \n2: Load: Encoder $\\mathcal{E}$ , artifact detector $\\mathcal{A}$ and LR decoder $\\mathcal{D}$ \n3: Step 1: Initialization \n4: Upscale LR image as $up(\\mathbf{I}_{LR})$ \n5: Encode the upsampled image as $\\pmb{x} = \\mathcal{E}(up(\\mathbf{I}_{LR}))$ \n6: Initialize the $\\pmb{x}$ as a realistic latent $\\pmb{x}_r$ and set it as guidance \n7: Compute the Realty score of the realistic latent $\\pmb{s}_r$ \n8: Step 2: Sampling \n9: for $t = T, \\dots, 1$ do \n10: Sample $\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ if $t > 1$ , else $\\epsilon = 0$ \n11: Compute the latent variable at the current step $\\pmb{x}_{t-1} = \\frac{1}{\\sqrt{\\alpha_t}} (\\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\alpha_t}} \\epsilon_\\theta (\\pmb{x}_t, \\pmb{x}, t)) + \\sigma_\\theta (\\pmb{x}_t, t) \\epsilon$ \n12: Reality-Guided Refinement \n13: Detect artifacts of the current latent $E_A = \\mathcal{A}(\\mathcal{D}(\\pmb{x}_{t-1}))$ \n14: Refine the latent $\\pmb{x}_{t-1} = \\pmb{x}_{t-1} \\times (1 - E_A) + \\pmb{x}_r \\times E_A$ \n15: Reality Score Computation \n16: Decode the refined latent into an image $\\mathbf{I}_r = \\mathcal{D}(\\pmb{x}_{t-1})$ \n17: Generate the current binary reality map $M_R = \\mathcal{R}(\\mathbf{I}_r)$ \n18: Calculate the current reality score $s_r^{t-1} = S(M_R)$ \n19: Self-Adaptive Guidance \n20: Encode the current realistic latent $\\pmb{x}_r^{t-1} = \\mathcal{E}(\\mathbf{I}_r)$ \n21: Update the guidance $\\pmb{x}_r = \\mathcal{G}(\\pmb{x}_r, \\pmb{x}_r^{t-1})$ if $s_r^{t-1} > s_r$ \n22: Update the reality score $\\pmb{s}_r = s_r^{t-1}$ if $s_r^{t-1} > s_r$ \n23: end for \n24: return the artifact-free SR $\\mathbf{L}_{HR} = \\mathcal{D}(\\pmb{x}_0)$", + "bbox": [ + 503, + 107, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 491, + 705, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present our experimental results, discuss insights from ablation studies, and delve into the broader implications of our findings.", + "bbox": [ + 498, + 516, + 893, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 500, + 571, + 696, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing Datasets. Our proposed SARGD method is rigorously assessed for arbitrary-scale single image superresolution (SR) across a comprehensive set of benchmark datasets, including Set5 [1], Set14 [41], B100 [24], Urban100 [14], and Manga109 [25]. The Set5 and Set14 datasets serve as standard benchmarks within the super-resolution domain, comprising five and fourteen high-resolution images, respectively. The B100 dataset is a collection of 100 natural images used for testing the versatility of super-resolution algorithms across various real-world scenarios. Urban100 comprises 100 detailed urban images, challenging super-resolution algorithms to accurately enhance intricate structures. Manga109 contains 109 high-resolution manga images, testing the edge preservation capabilities of super-resolution models.", + "bbox": [ + 496, + 594, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We utilize established perceptual metrics including PSNR, SSIM [35], LPIPS [44], and DISTS [47] to conduct a comprehensive comparison with existing diffusion-based SR approaches. PSNR and SSIM, particularly when assessed on the Y channel of the YCbCr", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25810", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/724bd76a06f65e7086b4a51c6394dee03fa3dd69b47ac5c2177b4417ee1c07a7.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with the state-of-the-art Diffusion-SR methods across benchmark datasets for super-resolution at $\\times 2$ , $\\times 3$ , and $\\times 4$ scales. The dagger symbol ( $\\dagger$ ) signifies LDM trained on the identical dataset as used for StableSR. Bold highlights the best performance. All evaluations are conducted on a 32G GPU. Our training-free SARGD attains the most favorable results." + ], + "table_footnote": [], + "table_body": "
BenchmarksSet5 [1]Set14 [41]B100 [24]Urban100 [14]Manga109 [25]
MetricsMethods×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑LDM†27.8627.9428.3026.4226.8127.2126.9127.5427.6725.1825.7225.9126.4926.9627.04
StableSR28.7829.0929.3627.2627.6928.0427.6128.2828.4026.0026.4426.5527.5828.0127.94
Ours32.6433.0632.2731.4030.9430.0130.9230.8430.2329.3428.8427.9331.2131.1330.23
SSIM ↑LDM†0.7650.7770.7690.6730.6880.6960.6510.6720.6790.7120.7190.7190.8140.8210.815
StableSR0.8050.8140.8120.7160.7310.7330.6940.7140.7170.7450.7480.7440.8430.8470.838
Ours0.8990.8970.8710.8420.8230.7780.8170.8000.7630.8490.8170.7710.9090.8970.851
LPIPS ↓LDM†0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1670.1760.1860.1320.1280.138
StableSR0.1650.1530.1500.1970.1940.2070.2000.2140.2260.1500.1590.1660.1180.1150.123
Ours0.1180.1260.1310.1390.1490.2040.1640.1180.2240.1170.1380.1770.0830.0850.128
DISTS ↓LDM†0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1290.1340.1410.1020.0980.103
StableSR0.1670.1650.1570.1520.1510.1500.1600.1600.1640.1270.1280.1310.1020.0970.099
Ours0.1360.1340.1420.1260.1210.1560.1320.1400.1560.1110.1190.1390.0860.0830.096
", + "bbox": [ + 78, + 131, + 893, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ad272e0cd85cd67a7aed2f9deade13911e932c9f18efcb133a02145218784736.jpg", + "image_caption": [ + "Figure 4. Visual comparison with diffusion-SR methods for $\\times 2$ , $\\times 3$ , and $\\times 4$ super-resolution, including (a) Bicubic upsampling, (b) StableSR, and (c) our SARGD. The red solid-lined boxes represent the ground truth (GT), focusing on regions zoomed for detailed inspection. The green boxes illustrate how our SARGD method preserves significantly more detail and clarity compared to the alternatives." + ], + "image_footnote": [], + "bbox": [ + 83, + 325, + 254, + 521 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d2a6a77e1420eb6a5740c9e344ed9dc79dc6fab9c6ffe71b3656b2425aef86d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 325, + 467, + 521 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d37f62d23948c63574a0bf839a6711c9b04d00d6eef864cf2e50a5ebd4fba99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 325, + 651, + 521 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae5a729cba7aeb2e7e619f6c41390e110e92d127c8726f020b3b1962f972fe15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 325, + 887, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "color space, are fidelity-focused measures traditionally favored in super-resolution tasks. Typically, higher values of PSNR and SSIM correspond to better performance. Nevertheless, these metrics do not adequately assess the quality of generative models. Therefore, we also incorporate LPIPS and DISTS in our assessment suite, as they have been acknowledged for their effectiveness in evaluating the perceptual quality of images produced by generative models. Generally, lower scores in LPIPS and DISTS are indicative of superior perceptual image quality.", + "bbox": [ + 75, + 574, + 472, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adhere to the experimental protocol established by previous research [32]. All of our experiments are conducted using PyTorch and tested on a consistent computing platform. The uniform model trained by Zhang et al. [43] is employed as an artifact detector to generate the artifact mask and compute the reality score. For the inference phase, we set the number of denoising steps to 200. Following established protocols [3], we derive low-resolution test samples using the imresize function in Matlab, employing bicubic interpolation with downsampling scales of $\\times 2$ , $\\times 3$ , and $\\times 4$ , in line with the precedents", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set by prior research. In diffusion-based super-resolution methods, it is standard practice to upsample the inputs to the same size as the outputs prior to the inference stage.", + "bbox": [ + 496, + 574, + 892, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 498, + 628, + 799, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the effectiveness of our training-free SARGD, we conducted comparative analyses with various state-of-the-art methods, such as LDM [26] and StableSR [32]. To ensure a fair comparison, LDM is finetuned using the same training configurations and dataset as those employed for StableSR. For other methods, we directly use the official codes and pre-trained models for testing. Comprehensive results, along with additional comparative analyses, are available in the supplementary material.", + "bbox": [ + 496, + 652, + 890, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Table 1 demonstrates the outstanding capability of our proposed Self-Adaptive Reality-Guided Diffusion (SARGD) method in producing artifact-free super-resolution images. Notably, our method significantly surpasses the previous StableSR model in terms of PSNR, with improvements of 3.86, 3.97, and 2.91 on the Set5 dataset at upscaling factors of $\\times 2$ , $\\times 3$ , and $\\times 4$ , respec", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25811", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/71dc42521358eb0aad41260bd470efb1e24a43e901b4a6b4e41d2dee6afa5db5.jpg", + "table_caption": [ + "Table 2. Analysis of SARGD inference strategies. The 'Baseline' model represents the absence of Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). 'RGR' utilizes initial realistic latent guidance, while 'SAG' employs the self-adaptive guidance mechanism. All experiments are conducted under identical settings. The best scores are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MetricRGRSAGSet5 [1]Set14 [41]B100 [24]Manga109 [25]
×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑XX27.8627.9428.3026.4226.8127.2126.9127.5427.6726.4926.9627.04
✓X29.8230.3930.27 (+ 6.96)28.1627.9927.79 (+ 2.13)28.0828.4228.30 (+ 2.28)25.1026.4127.88 (+ 3.11)
✓✓32.4232.9331.79 (+ 12.3)30.1630.3329.28 (+ 7.61)30.2930.3829.92 (+ 8.13)30.4030.5829.23 (+ 8.10)
SSIM ↑XX0.7650.7770.7690.6730.6880.6960.6510.6720.6790.8140.8210.815
✓X0.8230.8430.824 (+ 7.15)0.7280.7320.730 (+ 4.89)0.7020.7110.704 (+ 3.68)0.7750.8080.822 (+ 0.86)
✓✓0.8970.8950.870 (+ 13.1)0.8210.8150.770 (+ 10.6)0.8070.7920.760 (+ 11.9)0.9000.8930.851 (+ 4.41)
LPIPS ↓XX0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1320.1280.138
✓X0.1440.2120.161 (- 1.23)0.2270.2700.242 (+ 2.54)0.2010.2480.292 (+ 15.8)0.1280.1390.164 (+ 18.9)
✓✓0.1220.1410.131 (- 19.6)0.1440.1520.208 (- 11.9)0.1720.2010.228 (- 9.52)0.0870.0870.132 (- 4.35)
DISTS ↓XX0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1020.0980.103
✓X0.1450.1730.192 (+ 19.3)0.1600.1790.214 (+ 32.1)0.1910.2130.238 (+ 35.2)0.0940.0980.111 (+ 7.77)
✓✓0.1370.1410.145 (- 9.94)0.1270.1290.158 (- 2.47)0.1380.1410.158 (- 10.2)0.0910.0840.099 (- 3.88)
", + "bbox": [ + 78, + 132, + 893, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12d10dc9c40277321165aabb47ba116991c2ab77942ed4584d05d1cb78666cfa.jpg", + "image_caption": [ + "(a) LDM", + "Figure 5. Visual comparison of SARGD components. Our SARGD exhibits the best outcomes for $\\times 3$ super-resolution." + ], + "image_footnote": [], + "bbox": [ + 86, + 330, + 181, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/571b3e1dd456dbb7d610354ded5fc471792cc9207160ec9c3619e894c6d3477a.jpg", + "image_caption": [ + "(b) RGR" + ], + "image_footnote": [], + "bbox": [ + 183, + 330, + 277, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c48d19254526f5161ff5d601abee239e53d482a0a4ae603d33d34c7096facc2.jpg", + "image_caption": [ + "(c) SARGD" + ], + "image_footnote": [], + "bbox": [ + 279, + 330, + 370, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03de000f9deda631eae7bb3d11cae20a419d579d9ad6747f1f5781e6f75e3a07.jpg", + "image_caption": [ + "(d) GT" + ], + "image_footnote": [], + "bbox": [ + 372, + 330, + 462, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tively. Within the SSIM metric, our SARGD method realizes significant improvements (e.g. the increase from 0.716 to 0.842 on Set14 and from 0.694 to 0.817 on B100), illustrating a substantial enhancement in structural fidelity. Additionally, in terms of the DISTS metric, SARGD significantly reduces the score from 0.102 to 0.086 in Manga109, reflecting an improved alignment with human perceptual quality. These improvements establish our SARGD as a leading approach in the field of super-resolution, setting a new standard for image clarity and detail preservation.", + "bbox": [ + 75, + 492, + 468, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization. Figure 4 offers a visual comparison between our SARGD and StableSR [32], further highlighting the superior performance of our method. Compared to columns (a) and (c), bicubic upscaling maintains satisfactory detail but often leads to an over-smoothing issue. While StableSR succeeds in generating high-resolution images, its dependence on conventional DDPM methods makes it susceptible to introducing additional artifacts. When compared to columns (b) and (c), our approach successfully produces images with high authenticity, thereby avoiding the oversmoothing problem commonly seen in other methods.", + "bbox": [ + 75, + 652, + 467, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 830, + 230, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we thoroughly evaluate and analyze the performance of the proposed training-free SARGD method, primarily focusing on its effectiveness in denoising.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/763d9675289c249a578baaa067e31384947f35679c9553edcf2bede32113e73d.jpg", + "table_caption": [ + "Table 3. Comparison of varied denoising approaches within RGR across Set14, B100, and Urban100 datasets for $\\times 2$ , $\\times 3$ , $\\times 4$ upscaling. $\\mathbf{D}_{1s}^{*}$ refers to artifact refinement in the latent space at each step, $\\mathbf{D}_{10s}^{*}$ at every ten steps, while $\\mathbf{D}_{100s}^{-}$ and $\\mathbf{D}_{100s}^{+}$ indicate refinement before and after 100 steps, respectively." + ], + "table_footnote": [], + "table_body": "
ScaleTypeSet14B100Urban100
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
×2D*1s28.160.72828.080.70225.280.708
D*10s27.990.72528.050.69225.110.703
D*100s26.790.68527.320.67224.420.676
D*100s28.420.73928.110.71525.610.716
×3D*1s27.990.73228.420.71126.250.751
D*10s27.830.72628.470.70526.140.747
D-100s26.640.68627.660.68825.780.731
D*100s28.020.73028.540.71826.590.759
×4D*1s27.790.73028.300.70425.920.711
D*10s27.830.72928.480.70225.880.708
D*100s26.940.70827.830.68925.470.702
D*100s28.330.73428.510.71425.960.713
", + "bbox": [ + 501, + 401, + 890, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Improvement. Table 2 presents the effects of our proposed modules, Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). The baseline corresponds to the LDM configured and trained similarly to StableSR [32]. Analysis of the initial two columns reveals that the integration of RGR notably enhances performance, with gains of $6.96\\%$ , $2.13\\%$ , $2.28\\%$ , and $3.11\\%$ observed in $\\times 4$ super-resolution on the Set5, Set14, B100, and Manga109 datasets, respectively. The integration of RGR into the DDPM architecture leads to an increase in SSIM scores, validating the effectiveness of RGR in enhancing image quality. The visual results in Figure 5 show that while RGR manages to maintain the essential attributes of the original images, it also leads to over-smoothing in the super-resolved outputs. By contrasting figures (b) and (c), it is evident that incorporating our SAG during inference enhances image quality and successfully addresses the issue of oversmoothing, thus confirming the effectiveness of SAG.", + "bbox": [ + 496, + 578, + 890, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Varied Denoising Strategies within RGR. Table 3 demonstrates that the artifact refinement in the last 100 steps achieves the best results. Analyzing the first two", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25812", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b855eb6f66437da5822502616b81cb1266562f21d7054af6af2011c3129453af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 90, + 127, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09e65b2905f1b8a5a06441c92223fc58e335199d3eef1d134feb451142bc1311.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 83, + 150, + 127, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/519ead637caa4c739720a1cd6302b5fe0b3e928d839bf5cce7f590ea434a199c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 90, + 196, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/342a9bc91e141a7166a655e9d975f87fa67f18a82f630e3a200385c70318e7c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 90, + 264, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1294fda6971f511152a1ae26fa277dfa7e09d924dc1e5af300f0b7206ac832ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 90, + 328, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ec6ca6017bff8d74f11667bd5076386b2ec30793fab631b09a7f7d88d253620.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 90, + 398, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f51e3577e9ce3ab45b88251cd23304fe1299f9a71f38722285c31412bd436559.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 90, + 462, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/241c278ff9eb4f7ea97649eaa94cc2c28eb799fab1adb119022fd5a46be56fe3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 129, + 150, + 196, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8e17ff669863658261184bb4b1d9ab4483abd3e3445741f5e6c164927faeeda.jpg", + "image_caption": [ + "Figure 6. Visual comparison of RGR denoising strategies impacting over-smoothing. They are represented as $\\mathbf{D}_{1s}^{*}$ for continuous artifact refinement, $\\mathbf{D}_{10s}^{*}$ for refinement every ten steps, and $\\mathbf{D}_{100s}^{-}$ vs. $\\mathbf{D}_{100s}^{+}$ for pre- and post-100-step refinement." + ], + "image_footnote": [], + "bbox": [ + 197, + 150, + 263, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99c968cf4b358f1424b8f978e6be429e276965b14e8c9c1935936f0d497b4d55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 150, + 328, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ab2f358852206a613e09678db601c419aaecb842199232e99b9969d86da3a59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 150, + 398, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff6fb3f8f10b16674b63db3179c1e42d72fe73f8f91fb2f7586fae1a1f2db3c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 150, + 462, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37cd79634d83bee20d3a35195f16c12e38745de249bbd04c32951f0826a936c6.jpg", + "table_caption": [ + "Table 4. Comparison of different realistic latent udpate strategies. $\\mathbf{R}_{1s}^{*}$ refers to realistic-latent enhancement at each step and $\\mathbf{R}_{100s}^{+}$ indicates enhancement after 100 steps, respectively." + ], + "table_footnote": [], + "table_body": "
ScaleTypeSet14B100
PSNR ↑SSIM ↑DISTS ↓PSNR ↑SSIM ↑DISTS ↓
×2R*1s30.160.8210.12730.290.8070.138
R+100s31.400.8420.12630.920.8170.132
×3R*1s30.330.8150.12930.380.7920.141
R+100s30.940.8230.12130.840.8000.140
×4R*1s29.280.7700.15829.920.7600.158
R+100s30.010.7780.15630.230.7630.156
", + "bbox": [ + 78, + 325, + 467, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lines per scale shows that refining artifacts at each step improves outcomes over ten-step intervals, as seen with higher PSNRs (28.16 vs. 27.99 on Set4 at $\\times 2$ ), due to variable artifact patterns in the latent space. Comparing artifact refinement strategies indicates that post-100-step processing yields greater enhancements, as demonstrated by a PSNR increase from 26.94 to 28.33 at $\\times 4$ super-resolution. Observations from panels (d) and (c) in Figure 6 suggest that artifact refinement early in the noisy latent phase causes more over-smoothing than later-stage refinement.", + "bbox": [ + 75, + 431, + 467, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Approaches to Updating the Realistic Latent in SAG. Table 4 presents a comparative analysis of different strategies for updating the reference realistic-latent within SAG. Notably, updating the realistic latent after 100 steps yields superior results across all metrics when compared to enhancements at every step. For instance, this strategy leads to an increase in PSNR from 30.16 to $31.40$ at $\\times 2$ , from 30.33 to $30.94$ at $\\times 3$ , and from 29.28 to $30.1$ at $\\times 4$ on Set14. These findings suggest that enhancing the reference latent later in the process correlates with improved performance.", + "bbox": [ + 75, + 590, + 467, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Artifact Detection on SARGD. Table 5 contrasts SARGD's performance with and without the implementation of artifact detection in the latent space during inference, providing a line-by-line comparative analysis. It is observed that the elimination of perceptual artifacts in the latent space significantly enhances super-resolution quality, as evidenced by improved metrics such as PSNR (e.g., 30.84 vs. 30.06 on B100 at $\\times 3$ scaling) and DISTS (e.g., 0.219 vs. 0.140 on B100 at $\\times 3$ scaling), highlighting the benefits of integrating artifact detection within SARGD.", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/249495fc7fcca65038872fff38c57dd9948cd71f3c2ca71b00ad688e74129bf7.jpg", + "table_caption": [ + "Table 5. Performance of SARGD with artifact detection ( $\\mathcal{A}$ ) versus without in latent space across $\\times 2$ , $\\times 3$ , and $\\times 4$ scales. Here, $\\mathcal{X}$ signifies the substitution of RGR with a direct sum of the latent and the initial realistic reference latent." + ], + "table_footnote": [], + "table_body": "
ScaleASet14B100
PSNR↑SSIM↑DISTS↓PSNR↑SSIM↑DISTS↓
×2X31.240.8250.13830.560.8010.158
✓31.400.8420.12630.920.8170.132
×3X30.470.7960.19330.060.7830.219
✓30.940.8230.12130.840.8000.140
×4X29.980.7730.23430.150.7530.260
✓30.010.7780.15630.230.7630.156
", + "bbox": [ + 501, + 146, + 890, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f2bfa1c2777dc2c04b3a91f0b84c8e0043225007b4b2369ea8826e7fd3b6049.jpg", + "image_caption": [ + "Figure 7. Performance comparison of SARGD using different time steps on Set14 for $\\times 2$ , $\\times 3$ , $\\times 4$ upscaling using PSNR. Our SARGD reaches peak score between 75 and 100 steps, effectively reducing the total inference steps by half, from 200 to 100." + ], + "image_footnote": [], + "bbox": [ + 506, + 258, + 870, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact on Inference Steps for SARGD. Figure 7 illustrates that our proposed SARGD not only enhances image quality but also decreases the number of total inference steps required. Specifically, the performance of SARGD in super-resolution steadily improves from the start and achieves optimal PSNR scores between 75 and 100 steps, such as a PSNR of 31.32 at 100 steps compared to 30.16 at 200 steps, which effectively cuts the inference time in half. This suggests that the removal of artifacts during the diffusion phase serves to improve image quality in SR while simultaneously boosting the speed of the inference process.", + "bbox": [ + 496, + 477, + 890, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 660, + 625, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we tackle the issue of artifacts that arise in diffusion-based super-resolution. Our SARGD incorporates two key mechanisms: Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). RGR is designed to identify and correct artifacts within the latent space by utilizing a realistic latent as a reference, thereby improving the image's fidelity. To overcome the over-smoothing effect that comes from employing an initial realistic latent for guidance, SAG is introduced to refine the reference latent, improving the quality of the super-resolved images.", + "bbox": [ + 496, + 686, + 890, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (62271400), and the Shaanxi Provincial Key R&D Program, China (2023-GHZD-02).", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 6, 7", + "[2] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 2", + "[3] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 6", + "[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. in 2021 IEEE. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356, 2021. 3", + "[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12413-12422, 2022. 3", + "[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations (ICLR), 2023. 2", + "[7] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, 2020. 3", + "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 38(2):295-307, 2015. 1, 2", + "[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 391-407. Springer, 2016. 2", + "[10] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7890–7899, 2020. 3", + "[11] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in Neural Information Processing Systems (NIPS), 33:3022-3032, 2020. 3", + "[12] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1909-1918, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems (NIPS), 33:6840-6851, 2020. 2, 3, 4", + "[14] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6", + "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711, 2016. 2", + "[16] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4681-4690, 2017. 2", + "[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 1, 2, 3", + "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using wwin transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1833-1844, 2021. 2", + "[19] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4076-4085, 2021. 2", + "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 136-144, 2017. 2", + "[21] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8060-8069, 2020. 3", + "[22] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 715-732, 2020. 2", + "[23] Andreas Lugmayr, Martin Danelljan, Fisher Yu, Luc Van Gool, and Radu Timofte. Normalizing flow as a flexible fidelity objective for photo-realistic super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1756-1765, 2022. 2", + "[24] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25814", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 416-423, 2001. 5, 6, 7", + "[25] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 5, 6, 7", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 6", + "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 2", + "[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4): 4713-4726, 2022. 1, 3", + "[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1, 2", + "[30] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICLR), pages 2256-2265, 2015. 2", + "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 1", + "[32] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 6, 7", + "[33] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 2", + "[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV) workshops, pages 0–0, 2018. 2", + "[35] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5", + "[36] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2:" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2", + "[37] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2", + "[38] Xin Yang, Haiyang Mei, Jiqing Zhang, Ke Xu, Baocai Yin, Qiang Zhang, and Xiaopeng Wei. Drfn: Deep recurrent fusion network for single-image super-resolution with large factors. IEEE Transactions on Multimedia (TMM), 21(2): 328-337, 2018. 1", + "[39] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7556-7566, 2019. 3", + "[40] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 3", + "[41] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International Conference on Curves and Surfaces, pages 711-730. Springer, 2012. 5, 6, 7", + "[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 2", + "[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7579-7590, 2023. 4, 5, 6", + "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5", + "[45] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3096-3105, 2019. 2", + "[46] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In IEEE international Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 3", + "[47] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 286-301, 2018. 5", + "[48] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25815", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yulun Zhang, Kai Li, Kunpeng Li, and Yun Fu. Mr image super-resolution with squeeze and excitation reasoning attention network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13425-13434, 2021. 1", + "[50] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2185–2194, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25816", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_model.json b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a312715836c15adf39d7fbfb19bc04f5bb5cff1e --- /dev/null +++ b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.18, + 0.882, + 0.201 + ], + "angle": 0, + "content": "Qingping Zheng\\(^{1}\\), Ling Zheng\\(^{3}\\), Yuanfan Guo\\(^{2}\\), Ying Li\\(^{1*}\\), Songcen Xu\\(^{2}\\), Jiankang Deng\\(^{2}\\), Hang Xu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.203, + 0.841, + 0.22 + ], + "angle": 0, + "content": "\\(^{1}\\)Northwestern Polytechnical University \\(^{2}\\)Huawei Noah's Ark Lab \\(^{3}\\)Tsinghua-Fuzhou Institute for Data Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.254, + 0.314, + 0.27 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.287, + 0.474, + 0.741 + ], + "angle": 0, + "content": "Artifact-free super-resolution (SR) aims to translate low-resolution images into their high-resolution counterparts with a strict integrity of the original content, eliminating any distortions or synthetic details. While traditional diffusion-based SR techniques have demonstrated remarkable abilities to enhance image detail, they are prone to artifact introduction during iterative procedures. Such artifacts, ranging from trivial noise to unauthentic textures, deviate from the true structure of the source image, thus challenging the integrity of the super-resolution process. In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free method that delves into the latent space to effectively identify and mitigate the propagation of artifacts. Our SARGD begins by using an artifact detector to identify implausible pixels, creating a binary mask that highlights artifacts. Following this, the Reality Guidance Refinement (RGR) process refines artifacts by integrating this mask with realistic latent representations, improving alignment with the original image. Nonetheless, initial realistic-latent representations from lower-quality images result in over-smoothing in the final output. To address this, we introduce a Self-Adaptive Guidance (SAG) mechanism. It dynamically computes a reality score, enhancing the sharpness of the realistic latent. These alternating mechanisms collectively achieve artifact-free super-resolution. Extensive experiments demonstrate the superiority of our method, delivering detailed artifact-free high-resolution images while reducing sampling steps by \\(2 \\times\\). We release our code at https://github.com/ProAirVerse/Self-Adaptive-Guidance-Diffusion.git." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.21, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Super-resolution (SR) techniques strive to restore high-resolution images from their corresponding low-resolution counterparts [8, 29]. This process is inherently ill-posed since the same LR image can yield many plausible HR solutions. Traditional deep learning SR methods, leverag" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.256, + 0.584, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.258, + 0.659, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.257, + 0.734, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.258, + 0.808, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.258, + 0.883, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.351, + 0.792, + 0.361 + ], + "angle": 0, + "content": "(a) Variation of Artifacts within StableSR" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.362, + 0.584, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.362, + 0.659, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.362, + 0.734, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.362, + 0.808, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.362, + 0.883, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.447, + 0.8, + 0.457 + ], + "angle": 0, + "content": "(b) Variation of Artifacts within Our SARGD" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.458, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Figure 1. Visual comparison of artifact variability between (a) StableSR and (b) Our SARGD method. In the lower panel, the left image is the decoded output, while the right image is the artifact mask. Regions containing artifacts are highlighted in red, utilizing a binary map to serve as the artifact mask. Our SARGD method exhibits a superior capability in reducing artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.734 + ], + "angle": 0, + "content": "ing techniques like squeeze layers [49], transposed convolutions [38], and sub-pixel convolutions [29], bind themselves to fixed upsampling scales, limiting output resolution adjustments without architectural changes. This causes inflexibility in real-world applications. To tackle the inherent one-to-many inverse challenge in SR, it is intuitive to envision the solution space as a distribution. Yet, this perspective often prompts generative models to introduce details that deviate from the authentic content of the original low-resolution image. As a result, a wide range of generative-based SR approaches have emerged to address this intrinsically ill-posed problem, particularly in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Diffusion-based SR (Diffusion-SR for short) has exhibited significant potential in improving the resolution of images from their low-resolution counterparts. SRDiff [17] is the first diffusion-based model for image super-resolution, progressively transforming the Gaussian noise into a high-resolution (HR) image conditioned on its low-resolution (LR) input through a Markov chain. Building on the foundational concept of denoising score matching [31], Saharia et al. [28] propose SR3 model. This model advances the field of super-resolution by learning an empirical data distribution and leveraging U-Net architec" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "ture [27] to progressively denoise and refine the image output. However, these methods fail to preserve accurate pixelwise image structures, often necessitating further training or additional skipped connections for detail reproduction. While the pixel-aware stable diffusion network proposed by Yang et al. [37] significantly improves the robustness of the super-resolution process, it concurrently introduces the challenge of generating implausible content. In contrast, StableSR [32], the current state-of-the-art, leverages pre-trained diffusion priors to retain image details, avoiding assumptions about image degradation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.641 + ], + "angle": 0, + "content": "Nevertheless, as depicted in Figure 1, contemporary diffusion-based StableSR models invariably generate implausible content, frequently resulting in perceptual artifacts within specific areas of the image. Indeed, these artifacts frequently stem from the fundamental use of Denoising Diffusion Probabilistic Models (DDPMs) [13, 30] in modern diffusion models. Specifically, DDPMs, categorized as Partial Differential Equation (PDE)-based models [6], are designed to diffuse pixel intensities in order to reduce noise levels. They can enhance the preservation of edges and textural features by integrating non-linear terms, thus improving image details. However, despite these advancements, the introduction of non-linearity can occasionally yield unexpected outcomes, particularly in areas with intricate textures or noise patterns. This unintended consequence can result in a loss of fine details and a degradation of essential image elements such as edges and textures, especially when the diffusion process is overly aggressive or inadequately constrained. Additionally, any deficiencies in the model's architecture or parameter settings can be amplified during diffusion, aggravating artifact problems and producing results that deviate from the desired photorealistic quality [36]. These challenges pose a significant obstacle to the preservation of the authentic detail and clarity expected in high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a pioneering training-free approach, aiming to produce \"artifact-free\" SR images. In particular, SARGD operates through two stages: Reality-Guided Refinement (RGR) to reduce artifacts and Self-Adaptive Guidance (SAG) to boost image fidelity. During the initial RGR phase, an artifact detector is employed to identify unrealistic pixels within the latent space at each inference step, leading to a binary artifact mask. This mask collaborates with a realistic latent guidance extracted from an upscaled LR image to refine artifacts and enhance intrinsic details. However, depending on upscaled LR inputs for guidance can potentially result in over-smoothed images, as it may overlook high-frequency details that are essential for SR. To alleviate this limitation, the subsequent SAG phase introduces a 'reality score to self-adjust the authenticity of the latent guidance, significantly boosting the accuracy of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "detail and texture representation. Two phases alternate to progressively achieve high-fidelity super-resolution, effectively eliminating artifacts in the diffusion process. To sum up, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.213 + ], + "angle": 0, + "content": "- We introduce a Reality-Guided Refinement (RGR) strategy, designed to identify and rectify artifacts during the diffusion process with the aid of a reality guidance mechanism, effectively eliminating such artifacts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.213, + 0.892, + 0.273 + ], + "angle": 0, + "content": "- We propose a Self-Adaptive Guidance (SAG) mechanism that iteratively enhances the realistic-latent guidance, addressing over-smoothing issues and improving the fidelity and authenticity of the reference latent guidance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.274, + 0.892, + 0.364 + ], + "angle": 0, + "content": "- We are the first to develop the Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free approach that sets a new benchmark by effectively dealing with artifacts and over-smoothing issues. Additionally, it outperforms StableSR in image quality while reducing the inference time by \\(2 \\times\\) for super-resolution tasks." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.379, + 0.642, + 0.394 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.451 + ], + "angle": 0, + "content": "In this section, we summarize past super-resolution (SR) techniques, explore diffusion-based models in SR, and address artifact detection and mitigation in the SR process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.775 + ], + "angle": 0, + "content": "Image Super-Resolution. Single image super-resolution (SISR) is a process designed to construct high-resolution (HR) images from their low-resolution (LR) counterparts, which is inherently a complex task due to the multiple HR images that can correspond to a single LR image. The literature has seen a variety of solutions aiming to learn mappings from LR to HR images. These solutions fall into two primary groups: PSNR-oriented methods and generative model-based methods. PSNR-oriented methods [8, 9, 18, 20, 29, 48] employ L1 or L2 norms as objectives, achieving notable PSNR metrics. However, the reliance on L1/L2 losses tends to predict an average of all possible HR images [15, 16, 22], resulting in oversmoothed outputs. In contrast, generative model-based methods [2, 12, 16, 17, 19, 22, 23, 33, 34, 42, 45] seek to address the ill-posed nature of the SR problem by learning the distribution of potential HR images through techniques such as GAN-based, diffusion-based, flow-based, and AR-based super-resolution. These models aim to generate HR images that not only have high pixel accuracy but also possess the rich textural details necessary for visual realism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion probabilistic models [13, 30] are a class of generative models that utilize a Markov chain to convert latent variables from simple distributions (e.g., Gaussian) into data with complex distributions. Recognizing the potential of these models to address \"one-to-many\" mapping problems, Li et al. [17] have applied them to the task of super-resolution, generating a variety of high-resolution images from a single low-resolution input. This" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.887, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.896, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Self-Adaptive Reality-Guided Diffusion (SARGD) for artifact-free super-resolution. Our proposed SARGD is a training-free approach that consists of two principal components: 1) a Reality-Guided Refinement (RGR) that identifies and corrects artifacts within the latent representation by using a realistic latent as a guide to maintain the inherent details of the original image during the diffusion process, and 2) a Self-Adaptive Guidance (SAG) mechanism that enhances the fidelity of the initial realistic latent guidance, derived from upscaled low-resolution images, thereby effectively addressing the issue of over-smoothing in the final outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.473, + 0.627 + ], + "angle": 0, + "content": "application not only yields diverse outcomes but also simultaneously addresses challenges related to over smoothing, mode collapse, and the extensive computational footprint that are often encountered in super-resolution tasks. A prevalent method [26, 28] in diffusion-based SR involves integrating a low-resolution image directly into the input of existing diffusion models (e.g., DDPM [13]), and retraining the model with super-resolution data from scratch. Another approach [4, 5, 40] leverages an unconditional pretrained diffusion model as a prior and adapts its reverse diffusion process to synthesize the desired high-resolution image. Nonetheless, both strategies are constrained by the intrinsic Markov chain sequence of DDPM, which can potentially lead to artifacts during inference, thus affecting the quality of the resulting high-resolution image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Artifact Detection Methods. Despite impressive improvement, diffusion-based super-resolution methods tend to present undesired outcomes, known as artifacts. The artifact detection approach aims to probe the perceptually implausible contents rooted in the generated images. Existing deep-learning-based image artifact detection methods can be divided into spatial and frequency domain-based approaches. The former approaches [39] focus on capturing differences in texture between real and generated images. Liu et al. [21] leverage Gram matrices to capture the difference in texture between real and synthetic images. Dang et al. [7] adopt the attention mechanism to improve the accuracy of artifact detection by highlighting informative regions. Zhao et al. [50] utilize the attention mechanism to amplify subtle differences in the shallow layers, improving the detection performance. The latter approaches [46] study the artifacts in the frequency domain. Durall et al. [10] classify generated images according to the characteristics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.492 + ], + "angle": 0, + "content": "of the artifacts in high-frequency components. Dzanic et al. [11] demonstrate that image artifacts could be detected by leveraging the degree of partial decay at high frequencies. Frank et al. [11] find that the use of upsampling operations causes the artifacts in generated images, and employs the DCT transforms to detect them." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.503, + 0.637, + 0.521 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.895, + 0.603 + ], + "angle": 0, + "content": "In this section, we propose a novel training-free method called \"Self-Adaptive Reality-Guided Diffusion\" (SARGD) to tackle the ongoing challenge of artifact suppression in diffusion-based super-resolution. This approach, depicted in Figure 2, introduces two essential mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.604, + 0.894, + 0.663 + ], + "angle": 0, + "content": "- Reality-Guided Refinement (RGR), mitigates artifact emergence during inference by employing guidance from a realistic latent space, thereby preserving the original characteristics of the images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.664, + 0.894, + 0.725 + ], + "angle": 0, + "content": "- Self-Adaptive Guidance (SAG), enhances the authenticity of the reference realistic latent representation, effectively addressing the over-smoothing tendencies associated with the initial realistic latent representation." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.604, + 0.894, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.733, + 0.642, + 0.748 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.895, + 0.879 + ], + "angle": 0, + "content": "Diffusion-based super-resolution (SR) techniques [17, 28, 32] have emerged as a compelling paradigm for the enhancement of image resolutions. Consider the diffusion process as a function that aims to translate a given low-resolution (LR) image into a high-resolution (HR) image, each successive HR image at time \\( t + 1 \\) is obtained by refining the current estimate, factoring in the original LR image. This refinement is mathematically represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.884, + 0.893, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {H R} ^ {(t + 1)} = \\mathbf {I} _ {H R} ^ {(t)} + f _ {\\theta} (\\mathbf {I} _ {H R} ^ {(t)}, \\mathbf {I} _ {L R}), \\qquad (1)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.096, + 0.454, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Overview of the Reality-Guided Refinement (RGR) workflow: 1) Decoding the current latent into an RGB image and using an artifact detector to create a binary mask identifying areas with artifacts; and 2) Utilizing realistic latent guidance to refine the masked regions, enhancing the image's fidelity and authenticity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.434 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_{HR}^{(t)}\\) is the current state of the high-resolution image at iteration \\(t\\). \\(f_{\\theta}\\) is the diffusion function parameterized by \\(\\theta\\), which updates the current high-resolution state based on both its current state and the given low-resolution image. \\(t\\) is the current iteration during the inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.585 + ], + "angle": 0, + "content": "In the iterative diffusion process for super-resolution, the input LR-HR image pairs in the training set are used to train the diffusion models with the total diffusion step T. The diffusion process initially leverages a pre-trained visual encoder to convert the upsampled LR image into a latent \\( z_{t} = \\mathcal{E}(up(\\mathbf{I}_{LR})) \\), and gradually adding the Gaussian noise \\( \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) into it, resulting the noised latent variable \\( x \\). The reverse diffusion process aims to guide the generation to the corresponding HR image through iterative refinement of the noised latent \\( x \\), taking the form of:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.596, + 0.469, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t - 1} \\leftarrow \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {y} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\gamma_ {t}}} f _ {\\theta} \\left(\\boldsymbol {x}, \\boldsymbol {y} _ {t}, \\gamma_ {t}\\right)\\right) + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.7 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} \\in (0,1)\\) is the corresponding coefficient and \\(\\gamma_{t}\\) is the variance of the noise at iteration \\(t\\). Lastly, the decoder \\(\\mathcal{D}\\) is employed to project the resulting latent into the final high-resolution image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.38, + 0.726 + ], + "angle": 0, + "content": "3.2. Reality-Guided Refinement (RGR)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To overcome the prevalent dilemma of aggravated noise and the emergence of artifacts in images subjected to superresolution (SR) via diffusion-based techniques, we propose a pioneering training-free approach: Reality-Guided Refinement (RGR). This method couples an artifact detector with a latent refinement to address these issues effectively. As illustrated in Figure 3, a pixel-level segmentation model is initially applied to locate any artifacts on the current latent variable. Once detected, these imperfections are adeptly rectified through a sophisticated procedure of realistic-latent guidance refinement. This diffusion process" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "ensures that the resulting images retain the original, authentic textures and intricate details, effectively circumventing the usual distortions and artifacts associated with traditional diffusion-based upsampling techniques for SR tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.322 + ], + "angle": 0, + "content": "Artifact Detection in Diffusion Process. In the context of pixel-sensitive super-resolution, it is crucial to identify the fine-grained artifacts during the diffusion process. Hence, a robust artifact detection system should precisely pinpoint these implausible pixels within the latent space for effective identification. To achieve this, we utilize the established artifact detection method named PAL [43] to isolate these imperfections in the latent space. The process entails transforming the latent variable into the image domain at each iteration using the decoder, depicted by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.331, + 0.892, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {H R} ^ {(t - 1)} = \\mathcal {D} \\left(\\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {x}, t)\\right) + \\sigma_ {\\theta} \\left(x _ {t}, t\\right) \\boldsymbol {\\epsilon})\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Subsequently, the resulting image is input into the PAL detector, yielding a binary mask \\( E_{A} \\) that denotes the presence of artifacts (referenced as line 12 in Algorithm 1). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.44, + 0.892, + 0.459 + ], + "angle": 0, + "content": "\\[\nE _ {A} ^ {(t - 1)} = \\mathcal {A} \\left(\\mathbf {I} _ {H R} ^ {(t - 1)}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.512 + ], + "angle": 0, + "content": "Finally, the resultant artifact mask is resized to match the scale of the latent variable, maintaining the proportionality between the two." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Refinement via Reality Guidance. For the removal of artifacts during sampling, we refine the latent variable via a proposed training-free RGR. Following Algorithm 1, the RGR inference, consisting of T steps, begins by taking an upscaled low-resolution image \\(up(\\mathbf{I}_{LR})\\) as the starting point. A noise vector \\(\\epsilon\\) is then drawn from a standard Gaussian distribution. The process iterates backward from \\(t = T\\), with each step generating a latent variable \\(\\boldsymbol{x}_{t-1}\\) in accordance with DDPM [13] framework. Once artifacts \\(E_A\\) are identified within the current latent \\(\\boldsymbol{x}_{t-1}\\), we proceed to refine these imperfections through reality guidance. This correction process unfolds as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.708, + 0.892, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t - 1} = \\boldsymbol {x} _ {t - 1} \\times (1 - E _ {A}) + \\boldsymbol {x} _ {r} \\times E _ {A}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.809 + ], + "angle": 0, + "content": "This process allows us to effectively eliminate artifacts from the latent space. Notably, in this formula, we employ the initially encoded latent, which has not undergone the denoising process, as a source of realistic latent guidance, since it retains the intrinsic structural integrity of the image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.771, + 0.834 + ], + "angle": 0, + "content": "3.3. Self-Adaptive Guidance (SAG)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, relying on the initial latent space as a guide for artifact suppression during inference often results in oversmoothing results, primarily attributed to reliance on upscaled low-resolution images. To address this, we propose a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "novel Self-Adaptive Guidance (SAG) renewal mechanism, an automated system that perpetually refines realistic latent representations, ensuring the guidance is continually calibrated to enhance realistic detail. This mechanism is indispensable for diffusion-based super-resolution, aiming to yield high-resolution and high-fidelity images that strike a balance between details and sharpness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.473, + 0.428 + ], + "angle": 0, + "content": "Reality Score Computation. To regulate the update mechanism of reality guidance in image super-resolution, we introduce an innovative metric known as the reality score. This metric quantifies the realism of latent guidance. As delineated in Algorithm 1, we begin with an artifact-free latent variable \\( \\pmb{x}_{t-1} \\) and utilize our decoding function to synthesize a realistic image \\( \\mathbf{I}_r = \\mathcal{D}(\\pmb{x}_{t-1}) \\). Subsequently, we deploy a realistic-region function \\( \\mathcal{R} \\) by inverting the output of PAL [43] on \\( \\mathbf{I}_r \\), resulting in a binary reality mask \\( M_R \\) with dimensions corresponding to the channels, height, and width of the latent \\( C \\times H \\times W \\), respectively. The reality score \\( s_r \\) at the current \\( t-1 \\) step is calculated as the normalized sum of the reality mask across the spatial dimensions of the latent, indicating the distribution of realistic features. The formula is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.433, + 0.47, + 0.471 + ], + "angle": 0, + "content": "\\[\ns _ {r} ^ {t - 1} = \\mathcal {S} \\left(M _ {R}\\right) = \\frac {1}{H \\times W} \\sum_ {i = 0} ^ {H - 1} \\sum_ {j = 0} ^ {W - 1} M _ {R} ^ {(i j)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.47, + 0.554 + ], + "angle": 0, + "content": "Herein, \\( H \\) and \\( W \\) are the height and width of the latent, and \\( M_R^{(ij)} \\) indicates the reality measure at each pixel. This approach ensures a quantitative evaluation of the SR process, focusing on producing images that boast both high resolution and a high degree of fidelity to reality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.473, + 0.798 + ], + "angle": 0, + "content": "Self-Adaptive Reality Guidance. Self-adaptability in the realistic latent space refers to the ability of the system to adjust the reality guidance to correct the details and textures that are characteristic of high-quality images. This is especially important to mitigate the smoothing effect of the initial realistic-latent reliance on bicubic upsampling. As outlined in Algorithm 1, we begin by considering a realistic image \\(\\mathbf{I}_r\\) at the current step, decoded from an artifact-free latent variable \\(x_{t - 1}\\). This artifact-free image is then encoded into a realistic latent \\(x_{r}^{t - 1} = \\mathcal{E}(\\mathbf{I}_r)\\). Grounded on the computed reality score, self-adaptability is characterized as a guidance refinement process. During this process, the realistic latent \\(\\pmb{x}_r\\) at the current \\(t - 1\\) step is iteratively refined to converge towards an ideal latent representation \\(\\pmb{x}_r\\), precisely aligning with the high-resolution output IHR. The equation below illustrates this adaptive mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.803, + 0.47, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {x} _ {r} ^ {t - 1}\\right) = \\boldsymbol {x} _ {r} \\times \\left(1 - M _ {R}\\right) + \\boldsymbol {x} _ {r} ^ {t - 1} \\times M _ {R}, \\text {i f} \\boldsymbol {s} _ {r} ^ {t - 1} > \\boldsymbol {s} _ {r}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": "where \\( s_r \\) represents the reality score associated with \\( xr \\), and \\( M_R \\) is the binary reality mask corresponding to \\( xr^{t-1} \\). This self-adaptive mechanism ensures that the guidance is autonomously updated, leading to enhanced superresolution outcomes." + }, + { + "type": "code_caption", + "bbox": [ + 0.503, + 0.092, + 0.856, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Self-Adaptive Reality-Guided Diffusion" + }, + { + "type": "algorithm", + "bbox": [ + 0.504, + 0.108, + 0.892, + 0.467 + ], + "angle": 0, + "content": "1: Input: LR image \\(\\mathbf{I}_{LR}\\), and total diffusion steps \\(T\\) \n2: Load: Encoder \\(\\mathcal{E}\\), artifact detector \\(\\mathcal{A}\\) and LR decoder \\(\\mathcal{D}\\) \n3: Step 1: Initialization \n4: Upscale LR image as \\(up(\\mathbf{I}_{LR})\\) \n5: Encode the upsampled image as \\(\\pmb{x} = \\mathcal{E}(up(\\mathbf{I}_{LR}))\\) \n6: Initialize the \\(\\pmb{x}\\) as a realistic latent \\(\\pmb{x}_r\\) and set it as guidance \n7: Compute the Realty score of the realistic latent \\(\\pmb{s}_r\\) \n8: Step 2: Sampling \n9: for \\(t = T, \\dots, 1\\) do \n10: Sample \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) if \\(t > 1\\), else \\(\\epsilon = 0\\) \n11: Compute the latent variable at the current step \\(\\pmb{x}_{t-1} = \\frac{1}{\\sqrt{\\alpha_t}} (\\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\alpha_t}} \\epsilon_\\theta (\\pmb{x}_t, \\pmb{x}, t)) + \\sigma_\\theta (\\pmb{x}_t, t) \\epsilon\\) \n12: Reality-Guided Refinement \n13: Detect artifacts of the current latent \\(E_A = \\mathcal{A}(\\mathcal{D}(\\pmb{x}_{t-1}))\\) \n14: Refine the latent \\(\\pmb{x}_{t-1} = \\pmb{x}_{t-1} \\times (1 - E_A) + \\pmb{x}_r \\times E_A\\) \n15: Reality Score Computation \n16: Decode the refined latent into an image \\(\\mathbf{I}_r = \\mathcal{D}(\\pmb{x}_{t-1})\\) \n17: Generate the current binary reality map \\(M_R = \\mathcal{R}(\\mathbf{I}_r)\\) \n18: Calculate the current reality score \\(s_r^{t-1} = S(M_R)\\) \n19: Self-Adaptive Guidance \n20: Encode the current realistic latent \\(\\pmb{x}_r^{t-1} = \\mathcal{E}(\\mathbf{I}_r)\\) \n21: Update the guidance \\(\\pmb{x}_r = \\mathcal{G}(\\pmb{x}_r, \\pmb{x}_r^{t-1})\\) if \\(s_r^{t-1} > s_r\\) \n22: Update the reality score \\(\\pmb{s}_r = s_r^{t-1}\\) if \\(s_r^{t-1} > s_r\\) \n23: end for \n24: return the artifact-free SR \\(\\mathbf{L}_{HR} = \\mathcal{D}(\\pmb{x}_0)\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.492, + 0.706, + 0.509 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.517, + 0.894, + 0.563 + ], + "angle": 0, + "content": "In this section, we present our experimental results, discuss insights from ablation studies, and delve into the broader implications of our findings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.697, + 0.589 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Testing Datasets. Our proposed SARGD method is rigorously assessed for arbitrary-scale single image superresolution (SR) across a comprehensive set of benchmark datasets, including Set5 [1], Set14 [41], B100 [24], Urban100 [14], and Manga109 [25]. The Set5 and Set14 datasets serve as standard benchmarks within the super-resolution domain, comprising five and fourteen high-resolution images, respectively. The B100 dataset is a collection of 100 natural images used for testing the versatility of super-resolution algorithms across various real-world scenarios. Urban100 comprises 100 detailed urban images, challenging super-resolution algorithms to accurately enhance intricate structures. Manga109 contains 109 high-resolution manga images, testing the edge preservation capabilities of super-resolution models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We utilize established perceptual metrics including PSNR, SSIM [35], LPIPS [44], and DISTS [47] to conduct a comprehensive comparison with existing diffusion-based SR approaches. PSNR and SSIM, particularly when assessed on the Y channel of the YCbCr" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25810" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with the state-of-the-art Diffusion-SR methods across benchmark datasets for super-resolution at \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\) scales. The dagger symbol (\\(\\dagger\\)) signifies LDM trained on the identical dataset as used for StableSR. Bold highlights the best performance. All evaluations are conducted on a 32G GPU. Our training-free SARGD attains the most favorable results." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.132, + 0.895, + 0.313 + ], + "angle": 0, + "content": "
BenchmarksSet5 [1]Set14 [41]B100 [24]Urban100 [14]Manga109 [25]
MetricsMethods×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑LDM†27.8627.9428.3026.4226.8127.2126.9127.5427.6725.1825.7225.9126.4926.9627.04
StableSR28.7829.0929.3627.2627.6928.0427.6128.2828.4026.0026.4426.5527.5828.0127.94
Ours32.6433.0632.2731.4030.9430.0130.9230.8430.2329.3428.8427.9331.2131.1330.23
SSIM ↑LDM†0.7650.7770.7690.6730.6880.6960.6510.6720.6790.7120.7190.7190.8140.8210.815
StableSR0.8050.8140.8120.7160.7310.7330.6940.7140.7170.7450.7480.7440.8430.8470.838
Ours0.8990.8970.8710.8420.8230.7780.8170.8000.7630.8490.8170.7710.9090.8970.851
LPIPS ↓LDM†0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1670.1760.1860.1320.1280.138
StableSR0.1650.1530.1500.1970.1940.2070.2000.2140.2260.1500.1590.1660.1180.1150.123
Ours0.1180.1260.1310.1390.1490.2040.1640.1180.2240.1170.1380.1770.0830.0850.128
DISTS ↓LDM†0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1290.1340.1410.1020.0980.103
StableSR0.1670.1650.1570.1520.1510.1500.1600.1600.1640.1270.1280.1310.1020.0970.099
Ours0.1360.1340.1420.1260.1210.1560.1320.1400.1560.1110.1190.1390.0860.0830.096
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.327, + 0.256, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.327, + 0.468, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.327, + 0.653, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.327, + 0.888, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.525, + 0.893, + 0.568 + ], + "angle": 0, + "content": "Figure 4. Visual comparison with diffusion-SR methods for \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\) super-resolution, including (a) Bicubic upsampling, (b) StableSR, and (c) our SARGD. The red solid-lined boxes represent the ground truth (GT), focusing on regions zoomed for detailed inspection. The green boxes illustrate how our SARGD method preserves significantly more detail and clarity compared to the alternatives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.473, + 0.728 + ], + "angle": 0, + "content": "color space, are fidelity-focused measures traditionally favored in super-resolution tasks. Typically, higher values of PSNR and SSIM correspond to better performance. Nevertheless, these metrics do not adequately assess the quality of generative models. Therefore, we also incorporate LPIPS and DISTS in our assessment suite, as they have been acknowledged for their effectiveness in evaluating the perceptual quality of images produced by generative models. Generally, lower scores in LPIPS and DISTS are indicative of superior perceptual image quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We adhere to the experimental protocol established by previous research [32]. All of our experiments are conducted using PyTorch and tested on a consistent computing platform. The uniform model trained by Zhang et al. [43] is employed as an artifact detector to generate the artifact mask and compute the reality score. For the inference phase, we set the number of denoising steps to 200. Following established protocols [3], we derive low-resolution test samples using the imresize function in Matlab, employing bicubic interpolation with downsampling scales of \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\), in line with the precedents" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.893, + 0.622 + ], + "angle": 0, + "content": "set by prior research. In diffusion-based super-resolution methods, it is standard practice to upsample the inputs to the same size as the outputs prior to the inference stage." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.63, + 0.8, + 0.647 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.79 + ], + "angle": 0, + "content": "To verify the effectiveness of our training-free SARGD, we conducted comparative analyses with various state-of-the-art methods, such as LDM [26] and StableSR [32]. To ensure a fair comparison, LDM is finetuned using the same training configurations and dataset as those employed for StableSR. For other methods, we directly use the official codes and pre-trained models for testing. Comprehensive results, along with additional comparative analyses, are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Quantitative Results. Table 1 demonstrates the outstanding capability of our proposed Self-Adaptive Reality-Guided Diffusion (SARGD) method in producing artifact-free super-resolution images. Notably, our method significantly surpasses the previous StableSR model in terms of PSNR, with improvements of 3.86, 3.97, and 2.91 on the Set5 dataset at upscaling factors of \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\), respec" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "25811" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 2. Analysis of SARGD inference strategies. The 'Baseline' model represents the absence of Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). 'RGR' utilizes initial realistic latent guidance, while 'SAG' employs the self-adaptive guidance mechanism. All experiments are conducted under identical settings. The best scores are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.133, + 0.894, + 0.32 + ], + "angle": 0, + "content": "
MetricRGRSAGSet5 [1]Set14 [41]B100 [24]Manga109 [25]
×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑XX27.8627.9428.3026.4226.8127.2126.9127.5427.6726.4926.9627.04
✓X29.8230.3930.27 (+ 6.96)28.1627.9927.79 (+ 2.13)28.0828.4228.30 (+ 2.28)25.1026.4127.88 (+ 3.11)
✓✓32.4232.9331.79 (+ 12.3)30.1630.3329.28 (+ 7.61)30.2930.3829.92 (+ 8.13)30.4030.5829.23 (+ 8.10)
SSIM ↑XX0.7650.7770.7690.6730.6880.6960.6510.6720.6790.8140.8210.815
✓X0.8230.8430.824 (+ 7.15)0.7280.7320.730 (+ 4.89)0.7020.7110.704 (+ 3.68)0.7750.8080.822 (+ 0.86)
✓✓0.8970.8950.870 (+ 13.1)0.8210.8150.770 (+ 10.6)0.8070.7920.760 (+ 11.9)0.9000.8930.851 (+ 4.41)
LPIPS ↓XX0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1320.1280.138
✓X0.1440.2120.161 (- 1.23)0.2270.2700.242 (+ 2.54)0.2010.2480.292 (+ 15.8)0.1280.1390.164 (+ 18.9)
✓✓0.1220.1410.131 (- 19.6)0.1440.1520.208 (- 11.9)0.1720.2010.228 (- 9.52)0.0870.0870.132 (- 4.35)
DISTS ↓XX0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1020.0980.103
✓X0.1450.1730.192 (+ 19.3)0.1600.1790.214 (+ 32.1)0.1910.2130.238 (+ 35.2)0.0940.0980.111 (+ 7.77)
✓✓0.1370.1410.145 (- 9.94)0.1270.1290.158 (- 2.47)0.1380.1410.158 (- 10.2)0.0910.0840.099 (- 3.88)
" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.332, + 0.182, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.436, + 0.158, + 0.445 + ], + "angle": 0, + "content": "(a) LDM" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.332, + 0.278, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.437, + 0.254, + 0.446 + ], + "angle": 0, + "content": "(b) RGR" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.332, + 0.371, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.437, + 0.355, + 0.446 + ], + "angle": 0, + "content": "(c) SARGD" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.332, + 0.464, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.436, + 0.438, + 0.445 + ], + "angle": 0, + "content": "(d) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.448, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of SARGD components. Our SARGD exhibits the best outcomes for \\(\\times 3\\) super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.645 + ], + "angle": 0, + "content": "tively. Within the SSIM metric, our SARGD method realizes significant improvements (e.g. the increase from 0.716 to 0.842 on Set14 and from 0.694 to 0.817 on B100), illustrating a substantial enhancement in structural fidelity. Additionally, in terms of the DISTS metric, SARGD significantly reduces the score from 0.102 to 0.086 in Manga109, reflecting an improved alignment with human perceptual quality. These improvements establish our SARGD as a leading approach in the field of super-resolution, setting a new standard for image clarity and detail preservation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.468, + 0.819 + ], + "angle": 0, + "content": "Visualization. Figure 4 offers a visual comparison between our SARGD and StableSR [32], further highlighting the superior performance of our method. Compared to columns (a) and (c), bicubic upscaling maintains satisfactory detail but often leads to an over-smoothing issue. While StableSR succeeds in generating high-resolution images, its dependence on conventional DDPM methods makes it susceptible to introducing additional artifacts. When compared to columns (b) and (c), our approach successfully produces images with high authenticity, thereby avoiding the oversmoothing problem commonly seen in other methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.231, + 0.848 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this section, we thoroughly evaluate and analyze the performance of the proposed training-free SARGD method, primarily focusing on its effectiveness in denoising." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Table 3. Comparison of varied denoising approaches within RGR across Set14, B100, and Urban100 datasets for \\( \\times 2 \\), \\( \\times 3 \\), \\( \\times 4 \\) upscaling. \\( \\mathbf{D}_{1s}^{*} \\) refers to artifact refinement in the latent space at each step, \\( \\mathbf{D}_{10s}^{*} \\) at every ten steps, while \\( \\mathbf{D}_{100s}^{-} \\) and \\( \\mathbf{D}_{100s}^{+} \\) indicate refinement before and after 100 steps, respectively." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.402, + 0.892, + 0.566 + ], + "angle": 0, + "content": "
ScaleTypeSet14B100Urban100
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
×2D*1s28.160.72828.080.70225.280.708
D*10s27.990.72528.050.69225.110.703
D*100s26.790.68527.320.67224.420.676
D*100s28.420.73928.110.71525.610.716
×3D*1s27.990.73228.420.71126.250.751
D*10s27.830.72628.470.70526.140.747
D-100s26.640.68627.660.68825.780.731
D*100s28.020.73028.540.71826.590.759
×4D*1s27.790.73028.300.70425.920.711
D*10s27.830.72928.480.70225.880.708
D*100s26.940.70827.830.68925.470.702
D*100s28.330.73428.510.71425.960.713
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Analysis of Improvement. Table 2 presents the effects of our proposed modules, Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). The baseline corresponds to the LDM configured and trained similarly to StableSR [32]. Analysis of the initial two columns reveals that the integration of RGR notably enhances performance, with gains of \\(6.96\\%\\), \\(2.13\\%\\), \\(2.28\\%\\), and \\(3.11\\%\\) observed in \\(\\times 4\\) super-resolution on the Set5, Set14, B100, and Manga109 datasets, respectively. The integration of RGR into the DDPM architecture leads to an increase in SSIM scores, validating the effectiveness of RGR in enhancing image quality. The visual results in Figure 5 show that while RGR manages to maintain the essential attributes of the original images, it also leads to over-smoothing in the super-resolved outputs. By contrasting figures (b) and (c), it is evident that incorporating our SAG during inference enhances image quality and successfully addresses the issue of oversmoothing, thus confirming the effectiveness of SAG." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of Varied Denoising Strategies within RGR. Table 3 demonstrates that the artifact refinement in the last 100 steps achieves the best results. Analyzing the first two" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.129, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.151, + 0.129, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.092, + 0.197, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.092, + 0.265, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.092, + 0.329, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.092, + 0.4, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.092, + 0.464, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.207, + 0.128, + 0.218 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.13, + 0.151, + 0.197, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.151, + 0.264, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.151, + 0.329, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.151, + 0.4, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.151, + 0.464, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.22, + 0.468, + 0.276 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of RGR denoising strategies impacting over-smoothing. They are represented as \\(\\mathbf{D}_{1s}^{*}\\) for continuous artifact refinement, \\(\\mathbf{D}_{10s}^{*}\\) for refinement every ten steps, and \\(\\mathbf{D}_{100s}^{-}\\) vs. \\(\\mathbf{D}_{100s}^{+}\\) for pre- and post-100-step refinement." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.282, + 0.468, + 0.325 + ], + "angle": 0, + "content": "Table 4. Comparison of different realistic latent udpate strategies. \\( \\mathbf{R}_{1s}^{*} \\) refers to realistic-latent enhancement at each step and \\( \\mathbf{R}_{100s}^{+} \\) indicates enhancement after 100 steps, respectively." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.326, + 0.468, + 0.424 + ], + "angle": 0, + "content": "
ScaleTypeSet14B100
PSNR ↑SSIM ↑DISTS ↓PSNR ↑SSIM ↑DISTS ↓
×2R*1s30.160.8210.12730.290.8070.138
R+100s31.400.8420.12630.920.8170.132
×3R*1s30.330.8150.12930.380.7920.141
R+100s30.940.8230.12130.840.8000.140
×4R*1s29.280.7700.15829.920.7600.158
R+100s30.010.7780.15630.230.7630.156
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.468, + 0.582 + ], + "angle": 0, + "content": "lines per scale shows that refining artifacts at each step improves outcomes over ten-step intervals, as seen with higher PSNRs (28.16 vs. 27.99 on Set4 at \\(\\times 2\\)), due to variable artifact patterns in the latent space. Comparing artifact refinement strategies indicates that post-100-step processing yields greater enhancements, as demonstrated by a PSNR increase from 26.94 to 28.33 at \\(\\times 4\\) super-resolution. Observations from panels (d) and (c) in Figure 6 suggest that artifact refinement early in the noisy latent phase causes more over-smoothing than later-stage refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.468, + 0.743 + ], + "angle": 0, + "content": "Approaches to Updating the Realistic Latent in SAG. Table 4 presents a comparative analysis of different strategies for updating the reference realistic-latent within SAG. Notably, updating the realistic latent after 100 steps yields superior results across all metrics when compared to enhancements at every step. For instance, this strategy leads to an increase in PSNR from 30.16 to \\(31.40\\) at \\(\\times 2\\), from 30.33 to \\(30.94\\) at \\(\\times 3\\), and from 29.28 to \\(30.1\\) at \\(\\times 4\\) on Set14. These findings suggest that enhancing the reference latent later in the process correlates with improved performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Influence of Artifact Detection on SARGD. Table 5 contrasts SARGD's performance with and without the implementation of artifact detection in the latent space during inference, providing a line-by-line comparative analysis. It is observed that the elimination of perceptual artifacts in the latent space significantly enhances super-resolution quality, as evidenced by improved metrics such as PSNR (e.g., 30.84 vs. 30.06 on B100 at \\(\\times 3\\) scaling) and DISTS (e.g., 0.219 vs. 0.140 on B100 at \\(\\times 3\\) scaling), highlighting the benefits of integrating artifact detection within SARGD." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.144 + ], + "angle": 0, + "content": "Table 5. Performance of SARGD with artifact detection (\\( \\mathcal{A} \\)) versus without in latent space across \\( \\times 2 \\), \\( \\times 3 \\), and \\( \\times 4 \\) scales. Here, \\( \\mathcal{X} \\) signifies the substitution of RGR with a direct sum of the latent and the initial realistic reference latent." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.147, + 0.892, + 0.257 + ], + "angle": 0, + "content": "
ScaleASet14B100
PSNR↑SSIM↑DISTS↓PSNR↑SSIM↑DISTS↓
×2X31.240.8250.13830.560.8010.158
✓31.400.8420.12630.920.8170.132
×3X30.470.7960.19330.060.7830.219
✓30.940.8230.12130.840.8000.140
×4X29.980.7730.23430.150.7530.260
✓30.010.7780.15630.230.7630.156
" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.26, + 0.871, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Figure 7. Performance comparison of SARGD using different time steps on Set14 for \\(\\times 2\\), \\(\\times 3\\), \\(\\times 4\\) upscaling using PSNR. Our SARGD reaches peak score between 75 and 100 steps, effectively reducing the total inference steps by half, from 200 to 100." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Impact on Inference Steps for SARGD. Figure 7 illustrates that our proposed SARGD not only enhances image quality but also decreases the number of total inference steps required. Specifically, the performance of SARGD in super-resolution steadily improves from the start and achieves optimal PSNR scores between 75 and 100 steps, such as a PSNR of 31.32 at 100 steps compared to 30.16 at 200 steps, which effectively cuts the inference time in half. This suggests that the removal of artifacts during the diffusion phase serves to improve image quality in SR while simultaneously boosting the speed of the inference process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.661, + 0.627, + 0.677 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.839 + ], + "angle": 0, + "content": "In this study, we tackle the issue of artifacts that arise in diffusion-based super-resolution. Our SARGD incorporates two key mechanisms: Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). RGR is designed to identify and correct artifacts within the latent space by utilizing a realistic latent as a reference, thereby improving the image's fidelity. To overcome the over-smoothing effect that comes from employing an initial realistic latent for guidance, SAG is introduced to refine the reference latent, improving the quality of the super-resolved images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (62271400), and the Shaanxi Provincial Key R&D Program, China (2023-GHZD-02)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. in 2021 IEEE. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12413-12422, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations (ICLR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[7] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 38(2):295-307, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.646, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 391-407. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[10] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7890–7899, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[11] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in Neural Information Processing Systems (NIPS), 33:3022-3032, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1909-1918, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems (NIPS), 33:6840-6851, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[16] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using wwin transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1833-1844, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[19] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4076-4085, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 136-144, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[21] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8060-8069, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[22] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 715-732, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[23] Andreas Lugmayr, Martin Danelljan, Fisher Yu, Luc Van Gool, and Radu Timofte. Normalizing flow as a flexible fidelity objective for photo-realistic super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1756-1765, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 416-423, 2001. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[25] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4): 4713-4726, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[30] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICLR), pages 2256-2265, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[32] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[33] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV) workshops, pages 0–0, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[35] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[36] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[37] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[38] Xin Yang, Haiyang Mei, Jiqing Zhang, Ke Xu, Baocai Yin, Qiang Zhang, and Xiaopeng Wei. Drfn: Deep recurrent fusion network for single-image super-resolution with large factors. IEEE Transactions on Multimedia (TMM), 21(2): 328-337, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[39] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7556-7566, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[40] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[41] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International Conference on Curves and Surfaces, pages 711-730. Springer, 2012. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7579-7590, 2023. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[45] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3096-3105, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[46] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In IEEE international Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[47] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 286-301, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[48] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[49] Yulun Zhang, Kai Li, Kunpeng Li, and Yun Fu. Mr image super-resolution with squeeze and excitation reasoning attention network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13425-13434, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[50] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2185–2194, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25816" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_origin.pdf b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..14a3d54723fdcceb86617c6ed2b8b5940616640a --- /dev/null +++ b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/c6d65069-6c73-46fe-956d-1a0443a4ed2e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5c1d31a23ecf32aec08871a32d0dbb389e9f16ac7ad4e08ac5708f18c3c10e6 +size 1789286 diff --git a/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/full.md b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a2d8447ea2d4fc0653ec3b7388d83c540a3b59e2 --- /dev/null +++ b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/full.md @@ -0,0 +1,351 @@ +# Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution + +Qingping Zheng $^{1}$ , Ling Zheng $^{3}$ , Yuanfan Guo $^{2}$ , Ying Li $^{1*}$ , Songcen Xu $^{2}$ , Jiankang Deng $^{2}$ , Hang Xu $^{2}$ + +$^{1}$ Northwestern Polytechnical University $^{2}$ Huawei Noah's Ark Lab $^{3}$ Tsinghua-Fuzhou Institute for Data Technology + +# Abstract + +Artifact-free super-resolution (SR) aims to translate low-resolution images into their high-resolution counterparts with a strict integrity of the original content, eliminating any distortions or synthetic details. While traditional diffusion-based SR techniques have demonstrated remarkable abilities to enhance image detail, they are prone to artifact introduction during iterative procedures. Such artifacts, ranging from trivial noise to unauthentic textures, deviate from the true structure of the source image, thus challenging the integrity of the super-resolution process. In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free method that delves into the latent space to effectively identify and mitigate the propagation of artifacts. Our SARGD begins by using an artifact detector to identify implausible pixels, creating a binary mask that highlights artifacts. Following this, the Reality Guidance Refinement (RGR) process refines artifacts by integrating this mask with realistic latent representations, improving alignment with the original image. Nonetheless, initial realistic-latent representations from lower-quality images result in over-smoothing in the final output. To address this, we introduce a Self-Adaptive Guidance (SAG) mechanism. It dynamically computes a reality score, enhancing the sharpness of the realistic latent. These alternating mechanisms collectively achieve artifact-free super-resolution. Extensive experiments demonstrate the superiority of our method, delivering detailed artifact-free high-resolution images while reducing sampling steps by $2 \times$ . We release our code at https://github.com/ProAirVerse/Self-Adaptive-Guidance-Diffusion.git. + +# 1. Introduction + +Super-resolution (SR) techniques strive to restore high-resolution images from their corresponding low-resolution counterparts [8, 29]. This process is inherently ill-posed since the same LR image can yield many plausible HR solutions. Traditional deep learning SR methods, leverag + +![](images/d4bc0e146e9d72cf6e9167554118e0cf422a8e20cee50d253a888196a3075801.jpg) + +![](images/61918e4d994c8fcd04bb49380a6364cafa7e4125f75448ac3b62f3899d40837d.jpg) +(a) Variation of Artifacts within StableSR + +![](images/10c72e833a6efd78c3de57131df0fcf4ec3bf7e204670035b6db38179fa226b7.jpg) + +![](images/1941dd32834c13a109fa996d392e0e40a6d0d8a0705dbfb89d4792dbd9b18cfd.jpg) + +![](images/7dd435689f20a6d3d9c8e9954231b535e63e905a920bb27e4308fcc90ed593d8.jpg) + +![](images/914bf025e0b8d72ea5f09722a1c206a393ae764afc303d4d4acbb19fc8e97268.jpg) +Figure 1. Visual comparison of artifact variability between (a) StableSR and (b) Our SARGD method. In the lower panel, the left image is the decoded output, while the right image is the artifact mask. Regions containing artifacts are highlighted in red, utilizing a binary map to serve as the artifact mask. Our SARGD method exhibits a superior capability in reducing artifacts. + +![](images/24e794b3084e1e59c85d0c8c1380d602d7a87f77a78795be268d5e580938e193.jpg) +(b) Variation of Artifacts within Our SARGD + +![](images/059b17e30a51c8119adc7823dcd04ef72843909874b928bbeeadb49015c44df0.jpg) + +![](images/c203cbce8b899368523e55fe7ef36be34de158b6f3297a981f435e2b0f208b8f.jpg) + +![](images/03022f297f97a931630f37e10115df87b7bf97548f27eaae120e5fec916dcae0.jpg) + +ing techniques like squeeze layers [49], transposed convolutions [38], and sub-pixel convolutions [29], bind themselves to fixed upsampling scales, limiting output resolution adjustments without architectural changes. This causes inflexibility in real-world applications. To tackle the inherent one-to-many inverse challenge in SR, it is intuitive to envision the solution space as a distribution. Yet, this perspective often prompts generative models to introduce details that deviate from the authentic content of the original low-resolution image. As a result, a wide range of generative-based SR approaches have emerged to address this intrinsically ill-posed problem, particularly in diffusion models. + +Diffusion-based SR (Diffusion-SR for short) has exhibited significant potential in improving the resolution of images from their low-resolution counterparts. SRDiff [17] is the first diffusion-based model for image super-resolution, progressively transforming the Gaussian noise into a high-resolution (HR) image conditioned on its low-resolution (LR) input through a Markov chain. Building on the foundational concept of denoising score matching [31], Saharia et al. [28] propose SR3 model. This model advances the field of super-resolution by learning an empirical data distribution and leveraging U-Net architec + +ture [27] to progressively denoise and refine the image output. However, these methods fail to preserve accurate pixelwise image structures, often necessitating further training or additional skipped connections for detail reproduction. While the pixel-aware stable diffusion network proposed by Yang et al. [37] significantly improves the robustness of the super-resolution process, it concurrently introduces the challenge of generating implausible content. In contrast, StableSR [32], the current state-of-the-art, leverages pre-trained diffusion priors to retain image details, avoiding assumptions about image degradation. + +Nevertheless, as depicted in Figure 1, contemporary diffusion-based StableSR models invariably generate implausible content, frequently resulting in perceptual artifacts within specific areas of the image. Indeed, these artifacts frequently stem from the fundamental use of Denoising Diffusion Probabilistic Models (DDPMs) [13, 30] in modern diffusion models. Specifically, DDPMs, categorized as Partial Differential Equation (PDE)-based models [6], are designed to diffuse pixel intensities in order to reduce noise levels. They can enhance the preservation of edges and textural features by integrating non-linear terms, thus improving image details. However, despite these advancements, the introduction of non-linearity can occasionally yield unexpected outcomes, particularly in areas with intricate textures or noise patterns. This unintended consequence can result in a loss of fine details and a degradation of essential image elements such as edges and textures, especially when the diffusion process is overly aggressive or inadequately constrained. Additionally, any deficiencies in the model's architecture or parameter settings can be amplified during diffusion, aggravating artifact problems and producing results that deviate from the desired photorealistic quality [36]. These challenges pose a significant obstacle to the preservation of the authentic detail and clarity expected in high-resolution images. + +In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a pioneering training-free approach, aiming to produce "artifact-free" SR images. In particular, SARGD operates through two stages: Reality-Guided Refinement (RGR) to reduce artifacts and Self-Adaptive Guidance (SAG) to boost image fidelity. During the initial RGR phase, an artifact detector is employed to identify unrealistic pixels within the latent space at each inference step, leading to a binary artifact mask. This mask collaborates with a realistic latent guidance extracted from an upscaled LR image to refine artifacts and enhance intrinsic details. However, depending on upscaled LR inputs for guidance can potentially result in over-smoothed images, as it may overlook high-frequency details that are essential for SR. To alleviate this limitation, the subsequent SAG phase introduces a 'reality score to self-adjust the authenticity of the latent guidance, significantly boosting the accuracy of + +detail and texture representation. Two phases alternate to progressively achieve high-fidelity super-resolution, effectively eliminating artifacts in the diffusion process. To sum up, the contributions of this paper are as follows: + +- We introduce a Reality-Guided Refinement (RGR) strategy, designed to identify and rectify artifacts during the diffusion process with the aid of a reality guidance mechanism, effectively eliminating such artifacts. +- We propose a Self-Adaptive Guidance (SAG) mechanism that iteratively enhances the realistic-latent guidance, addressing over-smoothing issues and improving the fidelity and authenticity of the reference latent guidance. +- We are the first to develop the Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free approach that sets a new benchmark by effectively dealing with artifacts and over-smoothing issues. Additionally, it outperforms StableSR in image quality while reducing the inference time by $2 \times$ for super-resolution tasks. + +# 2. Related Work + +In this section, we summarize past super-resolution (SR) techniques, explore diffusion-based models in SR, and address artifact detection and mitigation in the SR process. + +Image Super-Resolution. Single image super-resolution (SISR) is a process designed to construct high-resolution (HR) images from their low-resolution (LR) counterparts, which is inherently a complex task due to the multiple HR images that can correspond to a single LR image. The literature has seen a variety of solutions aiming to learn mappings from LR to HR images. These solutions fall into two primary groups: PSNR-oriented methods and generative model-based methods. PSNR-oriented methods [8, 9, 18, 20, 29, 48] employ L1 or L2 norms as objectives, achieving notable PSNR metrics. However, the reliance on L1/L2 losses tends to predict an average of all possible HR images [15, 16, 22], resulting in oversmoothed outputs. In contrast, generative model-based methods [2, 12, 16, 17, 19, 22, 23, 33, 34, 42, 45] seek to address the ill-posed nature of the SR problem by learning the distribution of potential HR images through techniques such as GAN-based, diffusion-based, flow-based, and AR-based super-resolution. These models aim to generate HR images that not only have high pixel accuracy but also possess the rich textural details necessary for visual realism. + +Diffusion Models. Diffusion probabilistic models [13, 30] are a class of generative models that utilize a Markov chain to convert latent variables from simple distributions (e.g., Gaussian) into data with complex distributions. Recognizing the potential of these models to address "one-to-many" mapping problems, Li et al. [17] have applied them to the task of super-resolution, generating a variety of high-resolution images from a single low-resolution input. This + +![](images/1c10dfa7f22ffd95d0957a7538d16645962da4fb93305e5c04c798041f0dfc69.jpg) +Figure 2. Self-Adaptive Reality-Guided Diffusion (SARGD) for artifact-free super-resolution. Our proposed SARGD is a training-free approach that consists of two principal components: 1) a Reality-Guided Refinement (RGR) that identifies and corrects artifacts within the latent representation by using a realistic latent as a guide to maintain the inherent details of the original image during the diffusion process, and 2) a Self-Adaptive Guidance (SAG) mechanism that enhances the fidelity of the initial realistic latent guidance, derived from upscaled low-resolution images, thereby effectively addressing the issue of over-smoothing in the final outputs. + +application not only yields diverse outcomes but also simultaneously addresses challenges related to over smoothing, mode collapse, and the extensive computational footprint that are often encountered in super-resolution tasks. A prevalent method [26, 28] in diffusion-based SR involves integrating a low-resolution image directly into the input of existing diffusion models (e.g., DDPM [13]), and retraining the model with super-resolution data from scratch. Another approach [4, 5, 40] leverages an unconditional pretrained diffusion model as a prior and adapts its reverse diffusion process to synthesize the desired high-resolution image. Nonetheless, both strategies are constrained by the intrinsic Markov chain sequence of DDPM, which can potentially lead to artifacts during inference, thus affecting the quality of the resulting high-resolution image. + +Artifact Detection Methods. Despite impressive improvement, diffusion-based super-resolution methods tend to present undesired outcomes, known as artifacts. The artifact detection approach aims to probe the perceptually implausible contents rooted in the generated images. Existing deep-learning-based image artifact detection methods can be divided into spatial and frequency domain-based approaches. The former approaches [39] focus on capturing differences in texture between real and generated images. Liu et al. [21] leverage Gram matrices to capture the difference in texture between real and synthetic images. Dang et al. [7] adopt the attention mechanism to improve the accuracy of artifact detection by highlighting informative regions. Zhao et al. [50] utilize the attention mechanism to amplify subtle differences in the shallow layers, improving the detection performance. The latter approaches [46] study the artifacts in the frequency domain. Durall et al. [10] classify generated images according to the characteristics + +of the artifacts in high-frequency components. Dzanic et al. [11] demonstrate that image artifacts could be detected by leveraging the degree of partial decay at high frequencies. Frank et al. [11] find that the use of upsampling operations causes the artifacts in generated images, and employs the DCT transforms to detect them. + +# 3. Methodology + +In this section, we propose a novel training-free method called "Self-Adaptive Reality-Guided Diffusion" (SARGD) to tackle the ongoing challenge of artifact suppression in diffusion-based super-resolution. This approach, depicted in Figure 2, introduces two essential mechanisms. + +- Reality-Guided Refinement (RGR), mitigates artifact emergence during inference by employing guidance from a realistic latent space, thereby preserving the original characteristics of the images. +- Self-Adaptive Guidance (SAG), enhances the authenticity of the reference realistic latent representation, effectively addressing the over-smoothing tendencies associated with the initial realistic latent representation. + +# 3.1. Preliminaries + +Diffusion-based super-resolution (SR) techniques [17, 28, 32] have emerged as a compelling paradigm for the enhancement of image resolutions. Consider the diffusion process as a function that aims to translate a given low-resolution (LR) image into a high-resolution (HR) image, each successive HR image at time $t + 1$ is obtained by refining the current estimate, factoring in the original LR image. This refinement is mathematically represented as: + +$$ +\mathbf {I} _ {H R} ^ {(t + 1)} = \mathbf {I} _ {H R} ^ {(t)} + f _ {\theta} (\mathbf {I} _ {H R} ^ {(t)}, \mathbf {I} _ {L R}), \qquad (1) +$$ + +![](images/492e361481c399e45d8b42396f08e58bbb67f5d74a33f924507e52bcbb0c5ed1.jpg) +Figure 3. Overview of the Reality-Guided Refinement (RGR) workflow: 1) Decoding the current latent into an RGB image and using an artifact detector to create a binary mask identifying areas with artifacts; and 2) Utilizing realistic latent guidance to refine the masked regions, enhancing the image's fidelity and authenticity. + +where $\mathbf{I}_{HR}^{(t)}$ is the current state of the high-resolution image at iteration $t$ . $f_{\theta}$ is the diffusion function parameterized by $\theta$ , which updates the current high-resolution state based on both its current state and the given low-resolution image. $t$ is the current iteration during the inference. + +In the iterative diffusion process for super-resolution, the input LR-HR image pairs in the training set are used to train the diffusion models with the total diffusion step T. The diffusion process initially leverages a pre-trained visual encoder to convert the upsampled LR image into a latent $z_{t} = \mathcal{E}(up(\mathbf{I}_{LR}))$ , and gradually adding the Gaussian noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ into it, resulting the noised latent variable $x$ . The reverse diffusion process aims to guide the generation to the corresponding HR image through iterative refinement of the noised latent $x$ , taking the form of: + +$$ +\boldsymbol {y} _ {t - 1} \leftarrow \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {y} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \gamma_ {t}}} f _ {\theta} \left(\boldsymbol {x}, \boldsymbol {y} _ {t}, \gamma_ {t}\right)\right) + \sqrt {1 - \alpha_ {t}} \boldsymbol {\epsilon} _ {t}, \tag {2} +$$ + +where $\alpha_{t} \in (0,1)$ is the corresponding coefficient and $\gamma_{t}$ is the variance of the noise at iteration $t$ . Lastly, the decoder $\mathcal{D}$ is employed to project the resulting latent into the final high-resolution image. + +# 3.2. Reality-Guided Refinement (RGR) + +To overcome the prevalent dilemma of aggravated noise and the emergence of artifacts in images subjected to superresolution (SR) via diffusion-based techniques, we propose a pioneering training-free approach: Reality-Guided Refinement (RGR). This method couples an artifact detector with a latent refinement to address these issues effectively. As illustrated in Figure 3, a pixel-level segmentation model is initially applied to locate any artifacts on the current latent variable. Once detected, these imperfections are adeptly rectified through a sophisticated procedure of realistic-latent guidance refinement. This diffusion process + +ensures that the resulting images retain the original, authentic textures and intricate details, effectively circumventing the usual distortions and artifacts associated with traditional diffusion-based upsampling techniques for SR tasks. + +Artifact Detection in Diffusion Process. In the context of pixel-sensitive super-resolution, it is crucial to identify the fine-grained artifacts during the diffusion process. Hence, a robust artifact detection system should precisely pinpoint these implausible pixels within the latent space for effective identification. To achieve this, we utilize the established artifact detection method named PAL [43] to isolate these imperfections in the latent space. The process entails transforming the latent variable into the image domain at each iteration using the decoder, depicted by the following equation: + +$$ +\mathbf {I} _ {H R} ^ {(t - 1)} = \mathcal {D} \left(\frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {x} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (\boldsymbol {x} _ {t}, \boldsymbol {x}, t)\right) + \sigma_ {\theta} \left(x _ {t}, t\right) \boldsymbol {\epsilon})\right), \tag {3} +$$ + +Subsequently, the resulting image is input into the PAL detector, yielding a binary mask $E_{A}$ that denotes the presence of artifacts (referenced as line 12 in Algorithm 1). This process can be formulated as: + +$$ +E _ {A} ^ {(t - 1)} = \mathcal {A} \left(\mathbf {I} _ {H R} ^ {(t - 1)}\right). \tag {4} +$$ + +Finally, the resultant artifact mask is resized to match the scale of the latent variable, maintaining the proportionality between the two. + +Refinement via Reality Guidance. For the removal of artifacts during sampling, we refine the latent variable via a proposed training-free RGR. Following Algorithm 1, the RGR inference, consisting of T steps, begins by taking an upscaled low-resolution image $up(\mathbf{I}_{LR})$ as the starting point. A noise vector $\epsilon$ is then drawn from a standard Gaussian distribution. The process iterates backward from $t = T$ , with each step generating a latent variable $\boldsymbol{x}_{t-1}$ in accordance with DDPM [13] framework. Once artifacts $E_A$ are identified within the current latent $\boldsymbol{x}_{t-1}$ , we proceed to refine these imperfections through reality guidance. This correction process unfolds as follows: + +$$ +\boldsymbol {x} _ {t - 1} = \boldsymbol {x} _ {t - 1} \times (1 - E _ {A}) + \boldsymbol {x} _ {r} \times E _ {A}. \tag {5} +$$ + +This process allows us to effectively eliminate artifacts from the latent space. Notably, in this formula, we employ the initially encoded latent, which has not undergone the denoising process, as a source of realistic latent guidance, since it retains the intrinsic structural integrity of the image. + +# 3.3. Self-Adaptive Guidance (SAG) + +However, relying on the initial latent space as a guide for artifact suppression during inference often results in oversmoothing results, primarily attributed to reliance on upscaled low-resolution images. To address this, we propose a + +novel Self-Adaptive Guidance (SAG) renewal mechanism, an automated system that perpetually refines realistic latent representations, ensuring the guidance is continually calibrated to enhance realistic detail. This mechanism is indispensable for diffusion-based super-resolution, aiming to yield high-resolution and high-fidelity images that strike a balance between details and sharpness. + +Reality Score Computation. To regulate the update mechanism of reality guidance in image super-resolution, we introduce an innovative metric known as the reality score. This metric quantifies the realism of latent guidance. As delineated in Algorithm 1, we begin with an artifact-free latent variable $\pmb{x}_{t-1}$ and utilize our decoding function to synthesize a realistic image $\mathbf{I}_r = \mathcal{D}(\pmb{x}_{t-1})$ . Subsequently, we deploy a realistic-region function $\mathcal{R}$ by inverting the output of PAL [43] on $\mathbf{I}_r$ , resulting in a binary reality mask $M_R$ with dimensions corresponding to the channels, height, and width of the latent $C \times H \times W$ , respectively. The reality score $s_r$ at the current $t-1$ step is calculated as the normalized sum of the reality mask across the spatial dimensions of the latent, indicating the distribution of realistic features. The formula is given by: + +$$ +s _ {r} ^ {t - 1} = \mathcal {S} \left(M _ {R}\right) = \frac {1}{H \times W} \sum_ {i = 0} ^ {H - 1} \sum_ {j = 0} ^ {W - 1} M _ {R} ^ {(i j)}. \tag {6} +$$ + +Herein, $H$ and $W$ are the height and width of the latent, and $M_R^{(ij)}$ indicates the reality measure at each pixel. This approach ensures a quantitative evaluation of the SR process, focusing on producing images that boast both high resolution and a high degree of fidelity to reality. + +Self-Adaptive Reality Guidance. Self-adaptability in the realistic latent space refers to the ability of the system to adjust the reality guidance to correct the details and textures that are characteristic of high-quality images. This is especially important to mitigate the smoothing effect of the initial realistic-latent reliance on bicubic upsampling. As outlined in Algorithm 1, we begin by considering a realistic image $\mathbf{I}_r$ at the current step, decoded from an artifact-free latent variable $x_{t - 1}$ . This artifact-free image is then encoded into a realistic latent $x_{r}^{t - 1} = \mathcal{E}(\mathbf{I}_r)$ . Grounded on the computed reality score, self-adaptability is characterized as a guidance refinement process. During this process, the realistic latent $\pmb{x}_r$ at the current $t - 1$ step is iteratively refined to converge towards an ideal latent representation $\pmb{x}_r$ , precisely aligning with the high-resolution output IHR. The equation below illustrates this adaptive mechanism: + +$$ +\mathcal {G} \left(\boldsymbol {x} _ {r}, \boldsymbol {x} _ {r} ^ {t - 1}\right) = \boldsymbol {x} _ {r} \times \left(1 - M _ {R}\right) + \boldsymbol {x} _ {r} ^ {t - 1} \times M _ {R}, \text {i f} \boldsymbol {s} _ {r} ^ {t - 1} > \boldsymbol {s} _ {r}, \tag {7} +$$ + +where $s_r$ represents the reality score associated with $xr$ , and $M_R$ is the binary reality mask corresponding to $xr^{t-1}$ . This self-adaptive mechanism ensures that the guidance is autonomously updated, leading to enhanced superresolution outcomes. + +Algorithm 1 Self-Adaptive Reality-Guided Diffusion +1: Input: LR image $\mathbf{I}_{LR}$ , and total diffusion steps $T$ +2: Load: Encoder $\mathcal{E}$ , artifact detector $\mathcal{A}$ and LR decoder $\mathcal{D}$ +3: Step 1: Initialization +4: Upscale LR image as $up(\mathbf{I}_{LR})$ +5: Encode the upsampled image as $\pmb{x} = \mathcal{E}(up(\mathbf{I}_{LR}))$ +6: Initialize the $\pmb{x}$ as a realistic latent $\pmb{x}_r$ and set it as guidance +7: Compute the Realty score of the realistic latent $\pmb{s}_r$ +8: Step 2: Sampling +9: for $t = T, \dots, 1$ do +10: Sample $\epsilon \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ if $t > 1$ , else $\epsilon = 0$ +11: Compute the latent variable at the current step $\pmb{x}_{t-1} = \frac{1}{\sqrt{\alpha_t}} (\pmb{x}_t - \frac{1 - \alpha_t}{\sqrt{1 - \alpha_t}} \epsilon_\theta (\pmb{x}_t, \pmb{x}, t)) + \sigma_\theta (\pmb{x}_t, t) \epsilon$ +12: Reality-Guided Refinement +13: Detect artifacts of the current latent $E_A = \mathcal{A}(\mathcal{D}(\pmb{x}_{t-1}))$ +14: Refine the latent $\pmb{x}_{t-1} = \pmb{x}_{t-1} \times (1 - E_A) + \pmb{x}_r \times E_A$ +15: Reality Score Computation +16: Decode the refined latent into an image $\mathbf{I}_r = \mathcal{D}(\pmb{x}_{t-1})$ +17: Generate the current binary reality map $M_R = \mathcal{R}(\mathbf{I}_r)$ +18: Calculate the current reality score $s_r^{t-1} = S(M_R)$ +19: Self-Adaptive Guidance +20: Encode the current realistic latent $\pmb{x}_r^{t-1} = \mathcal{E}(\mathbf{I}_r)$ +21: Update the guidance $\pmb{x}_r = \mathcal{G}(\pmb{x}_r, \pmb{x}_r^{t-1})$ if $s_r^{t-1} > s_r$ +22: Update the reality score $\pmb{s}_r = s_r^{t-1}$ if $s_r^{t-1} > s_r$ +23: end for +24: return the artifact-free SR $\mathbf{L}_{HR} = \mathcal{D}(\pmb{x}_0)$ + +# 4. Experimental Results + +In this section, we present our experimental results, discuss insights from ablation studies, and delve into the broader implications of our findings. + +# 4.1. Experimental Setups + +Testing Datasets. Our proposed SARGD method is rigorously assessed for arbitrary-scale single image superresolution (SR) across a comprehensive set of benchmark datasets, including Set5 [1], Set14 [41], B100 [24], Urban100 [14], and Manga109 [25]. The Set5 and Set14 datasets serve as standard benchmarks within the super-resolution domain, comprising five and fourteen high-resolution images, respectively. The B100 dataset is a collection of 100 natural images used for testing the versatility of super-resolution algorithms across various real-world scenarios. Urban100 comprises 100 detailed urban images, challenging super-resolution algorithms to accurately enhance intricate structures. Manga109 contains 109 high-resolution manga images, testing the edge preservation capabilities of super-resolution models. + +Evaluation Metrics. We utilize established perceptual metrics including PSNR, SSIM [35], LPIPS [44], and DISTS [47] to conduct a comprehensive comparison with existing diffusion-based SR approaches. PSNR and SSIM, particularly when assessed on the Y channel of the YCbCr + +Table 1. Quantitative comparison with the state-of-the-art Diffusion-SR methods across benchmark datasets for super-resolution at $\times 2$ , $\times 3$ , and $\times 4$ scales. The dagger symbol ( $\dagger$ ) signifies LDM trained on the identical dataset as used for StableSR. Bold highlights the best performance. All evaluations are conducted on a 32G GPU. Our training-free SARGD attains the most favorable results. + +
BenchmarksSet5 [1]Set14 [41]B100 [24]Urban100 [14]Manga109 [25]
MetricsMethods×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑LDM†27.8627.9428.3026.4226.8127.2126.9127.5427.6725.1825.7225.9126.4926.9627.04
StableSR28.7829.0929.3627.2627.6928.0427.6128.2828.4026.0026.4426.5527.5828.0127.94
Ours32.6433.0632.2731.4030.9430.0130.9230.8430.2329.3428.8427.9331.2131.1330.23
SSIM ↑LDM†0.7650.7770.7690.6730.6880.6960.6510.6720.6790.7120.7190.7190.8140.8210.815
StableSR0.8050.8140.8120.7160.7310.7330.6940.7140.7170.7450.7480.7440.8430.8470.838
Ours0.8990.8970.8710.8420.8230.7780.8170.8000.7630.8490.8170.7710.9090.8970.851
LPIPS ↓LDM†0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1670.1760.1860.1320.1280.138
StableSR0.1650.1530.1500.1970.1940.2070.2000.2140.2260.1500.1590.1660.1180.1150.123
Ours0.1180.1260.1310.1390.1490.2040.1640.1180.2240.1170.1380.1770.0830.0850.128
DISTS ↓LDM†0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1290.1340.1410.1020.0980.103
StableSR0.1670.1650.1570.1520.1510.1500.1600.1600.1640.1270.1280.1310.1020.0970.099
Ours0.1360.1340.1420.1260.1210.1560.1320.1400.1560.1110.1190.1390.0860.0830.096
+ +![](images/ad272e0cd85cd67a7aed2f9deade13911e932c9f18efcb133a02145218784736.jpg) +Figure 4. Visual comparison with diffusion-SR methods for $\times 2$ , $\times 3$ , and $\times 4$ super-resolution, including (a) Bicubic upsampling, (b) StableSR, and (c) our SARGD. The red solid-lined boxes represent the ground truth (GT), focusing on regions zoomed for detailed inspection. The green boxes illustrate how our SARGD method preserves significantly more detail and clarity compared to the alternatives. + +![](images/4d2a6a77e1420eb6a5740c9e344ed9dc79dc6fab9c6ffe71b3656b2425aef86d.jpg) + +![](images/4d37f62d23948c63574a0bf839a6711c9b04d00d6eef864cf2e50a5ebd4fba99.jpg) + +![](images/ae5a729cba7aeb2e7e619f6c41390e110e92d127c8726f020b3b1962f972fe15.jpg) + +color space, are fidelity-focused measures traditionally favored in super-resolution tasks. Typically, higher values of PSNR and SSIM correspond to better performance. Nevertheless, these metrics do not adequately assess the quality of generative models. Therefore, we also incorporate LPIPS and DISTS in our assessment suite, as they have been acknowledged for their effectiveness in evaluating the perceptual quality of images produced by generative models. Generally, lower scores in LPIPS and DISTS are indicative of superior perceptual image quality. + +Implementation Details. We adhere to the experimental protocol established by previous research [32]. All of our experiments are conducted using PyTorch and tested on a consistent computing platform. The uniform model trained by Zhang et al. [43] is employed as an artifact detector to generate the artifact mask and compute the reality score. For the inference phase, we set the number of denoising steps to 200. Following established protocols [3], we derive low-resolution test samples using the imresize function in Matlab, employing bicubic interpolation with downsampling scales of $\times 2$ , $\times 3$ , and $\times 4$ , in line with the precedents + +set by prior research. In diffusion-based super-resolution methods, it is standard practice to upsample the inputs to the same size as the outputs prior to the inference stage. + +# 4.2. Comparison with State-of-the-Art + +To verify the effectiveness of our training-free SARGD, we conducted comparative analyses with various state-of-the-art methods, such as LDM [26] and StableSR [32]. To ensure a fair comparison, LDM is finetuned using the same training configurations and dataset as those employed for StableSR. For other methods, we directly use the official codes and pre-trained models for testing. Comprehensive results, along with additional comparative analyses, are available in the supplementary material. + +Quantitative Results. Table 1 demonstrates the outstanding capability of our proposed Self-Adaptive Reality-Guided Diffusion (SARGD) method in producing artifact-free super-resolution images. Notably, our method significantly surpasses the previous StableSR model in terms of PSNR, with improvements of 3.86, 3.97, and 2.91 on the Set5 dataset at upscaling factors of $\times 2$ , $\times 3$ , and $\times 4$ , respec + +Table 2. Analysis of SARGD inference strategies. The 'Baseline' model represents the absence of Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). 'RGR' utilizes initial realistic latent guidance, while 'SAG' employs the self-adaptive guidance mechanism. All experiments are conducted under identical settings. The best scores are highlighted in bold. + +
MetricRGRSAGSet5 [1]Set14 [41]B100 [24]Manga109 [25]
×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑XX27.8627.9428.3026.4226.8127.2126.9127.5427.6726.4926.9627.04
✓X29.8230.3930.27 (+ 6.96)28.1627.9927.79 (+ 2.13)28.0828.4228.30 (+ 2.28)25.1026.4127.88 (+ 3.11)
✓✓32.4232.9331.79 (+ 12.3)30.1630.3329.28 (+ 7.61)30.2930.3829.92 (+ 8.13)30.4030.5829.23 (+ 8.10)
SSIM ↑XX0.7650.7770.7690.6730.6880.6960.6510.6720.6790.8140.8210.815
✓X0.8230.8430.824 (+ 7.15)0.7280.7320.730 (+ 4.89)0.7020.7110.704 (+ 3.68)0.7750.8080.822 (+ 0.86)
✓✓0.8970.8950.870 (+ 13.1)0.8210.8150.770 (+ 10.6)0.8070.7920.760 (+ 11.9)0.9000.8930.851 (+ 4.41)
LPIPS ↓XX0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1320.1280.138
✓X0.1440.2120.161 (- 1.23)0.2270.2700.242 (+ 2.54)0.2010.2480.292 (+ 15.8)0.1280.1390.164 (+ 18.9)
✓✓0.1220.1410.131 (- 19.6)0.1440.1520.208 (- 11.9)0.1720.2010.228 (- 9.52)0.0870.0870.132 (- 4.35)
DISTS ↓XX0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1020.0980.103
✓X0.1450.1730.192 (+ 19.3)0.1600.1790.214 (+ 32.1)0.1910.2130.238 (+ 35.2)0.0940.0980.111 (+ 7.77)
✓✓0.1370.1410.145 (- 9.94)0.1270.1290.158 (- 2.47)0.1380.1410.158 (- 10.2)0.0910.0840.099 (- 3.88)
+ +![](images/12d10dc9c40277321165aabb47ba116991c2ab77942ed4584d05d1cb78666cfa.jpg) +(a) LDM +Figure 5. Visual comparison of SARGD components. Our SARGD exhibits the best outcomes for $\times 3$ super-resolution. + +![](images/571b3e1dd456dbb7d610354ded5fc471792cc9207160ec9c3619e894c6d3477a.jpg) +(b) RGR + +![](images/5c48d19254526f5161ff5d601abee239e53d482a0a4ae603d33d34c7096facc2.jpg) +(c) SARGD + +![](images/03de000f9deda631eae7bb3d11cae20a419d579d9ad6747f1f5781e6f75e3a07.jpg) +(d) GT + +tively. Within the SSIM metric, our SARGD method realizes significant improvements (e.g. the increase from 0.716 to 0.842 on Set14 and from 0.694 to 0.817 on B100), illustrating a substantial enhancement in structural fidelity. Additionally, in terms of the DISTS metric, SARGD significantly reduces the score from 0.102 to 0.086 in Manga109, reflecting an improved alignment with human perceptual quality. These improvements establish our SARGD as a leading approach in the field of super-resolution, setting a new standard for image clarity and detail preservation. + +Visualization. Figure 4 offers a visual comparison between our SARGD and StableSR [32], further highlighting the superior performance of our method. Compared to columns (a) and (c), bicubic upscaling maintains satisfactory detail but often leads to an over-smoothing issue. While StableSR succeeds in generating high-resolution images, its dependence on conventional DDPM methods makes it susceptible to introducing additional artifacts. When compared to columns (b) and (c), our approach successfully produces images with high authenticity, thereby avoiding the oversmoothing problem commonly seen in other methods. + +# 4.3. Ablation Study + +In this section, we thoroughly evaluate and analyze the performance of the proposed training-free SARGD method, primarily focusing on its effectiveness in denoising. + +Table 3. Comparison of varied denoising approaches within RGR across Set14, B100, and Urban100 datasets for $\times 2$ , $\times 3$ , $\times 4$ upscaling. $\mathbf{D}_{1s}^{*}$ refers to artifact refinement in the latent space at each step, $\mathbf{D}_{10s}^{*}$ at every ten steps, while $\mathbf{D}_{100s}^{-}$ and $\mathbf{D}_{100s}^{+}$ indicate refinement before and after 100 steps, respectively. + +
ScaleTypeSet14B100Urban100
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
×2D*1s28.160.72828.080.70225.280.708
D*10s27.990.72528.050.69225.110.703
D*100s26.790.68527.320.67224.420.676
D*100s28.420.73928.110.71525.610.716
×3D*1s27.990.73228.420.71126.250.751
D*10s27.830.72628.470.70526.140.747
D-100s26.640.68627.660.68825.780.731
D*100s28.020.73028.540.71826.590.759
×4D*1s27.790.73028.300.70425.920.711
D*10s27.830.72928.480.70225.880.708
D*100s26.940.70827.830.68925.470.702
D*100s28.330.73428.510.71425.960.713
+ +Analysis of Improvement. Table 2 presents the effects of our proposed modules, Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). The baseline corresponds to the LDM configured and trained similarly to StableSR [32]. Analysis of the initial two columns reveals that the integration of RGR notably enhances performance, with gains of $6.96\%$ , $2.13\%$ , $2.28\%$ , and $3.11\%$ observed in $\times 4$ super-resolution on the Set5, Set14, B100, and Manga109 datasets, respectively. The integration of RGR into the DDPM architecture leads to an increase in SSIM scores, validating the effectiveness of RGR in enhancing image quality. The visual results in Figure 5 show that while RGR manages to maintain the essential attributes of the original images, it also leads to over-smoothing in the super-resolved outputs. By contrasting figures (b) and (c), it is evident that incorporating our SAG during inference enhances image quality and successfully addresses the issue of oversmoothing, thus confirming the effectiveness of SAG. + +Effect of Varied Denoising Strategies within RGR. Table 3 demonstrates that the artifact refinement in the last 100 steps achieves the best results. Analyzing the first two + +![](images/b855eb6f66437da5822502616b81cb1266562f21d7054af6af2011c3129453af.jpg) + +![](images/09e65b2905f1b8a5a06441c92223fc58e335199d3eef1d134feb451142bc1311.jpg) +(a) Image + +![](images/519ead637caa4c739720a1cd6302b5fe0b3e928d839bf5cce7f590ea434a199c.jpg) + +![](images/342a9bc91e141a7166a655e9d975f87fa67f18a82f630e3a200385c70318e7c3.jpg) + +![](images/1294fda6971f511152a1ae26fa277dfa7e09d924dc1e5af300f0b7206ac832ed.jpg) + +![](images/2ec6ca6017bff8d74f11667bd5076386b2ec30793fab631b09a7f7d88d253620.jpg) + +![](images/f51e3577e9ce3ab45b88251cd23304fe1299f9a71f38722285c31412bd436559.jpg) + +![](images/241c278ff9eb4f7ea97649eaa94cc2c28eb799fab1adb119022fd5a46be56fe3.jpg) + +![](images/c8e17ff669863658261184bb4b1d9ab4483abd3e3445741f5e6c164927faeeda.jpg) +Figure 6. Visual comparison of RGR denoising strategies impacting over-smoothing. They are represented as $\mathbf{D}_{1s}^{*}$ for continuous artifact refinement, $\mathbf{D}_{10s}^{*}$ for refinement every ten steps, and $\mathbf{D}_{100s}^{-}$ vs. $\mathbf{D}_{100s}^{+}$ for pre- and post-100-step refinement. + +![](images/99c968cf4b358f1424b8f978e6be429e276965b14e8c9c1935936f0d497b4d55.jpg) + +![](images/7ab2f358852206a613e09678db601c419aaecb842199232e99b9969d86da3a59.jpg) + +![](images/ff6fb3f8f10b16674b63db3179c1e42d72fe73f8f91fb2f7586fae1a1f2db3c6.jpg) + +Table 4. Comparison of different realistic latent udpate strategies. $\mathbf{R}_{1s}^{*}$ refers to realistic-latent enhancement at each step and $\mathbf{R}_{100s}^{+}$ indicates enhancement after 100 steps, respectively. + +
ScaleTypeSet14B100
PSNR ↑SSIM ↑DISTS ↓PSNR ↑SSIM ↑DISTS ↓
×2R*1s30.160.8210.12730.290.8070.138
R+100s31.400.8420.12630.920.8170.132
×3R*1s30.330.8150.12930.380.7920.141
R+100s30.940.8230.12130.840.8000.140
×4R*1s29.280.7700.15829.920.7600.158
R+100s30.010.7780.15630.230.7630.156
+ +lines per scale shows that refining artifacts at each step improves outcomes over ten-step intervals, as seen with higher PSNRs (28.16 vs. 27.99 on Set4 at $\times 2$ ), due to variable artifact patterns in the latent space. Comparing artifact refinement strategies indicates that post-100-step processing yields greater enhancements, as demonstrated by a PSNR increase from 26.94 to 28.33 at $\times 4$ super-resolution. Observations from panels (d) and (c) in Figure 6 suggest that artifact refinement early in the noisy latent phase causes more over-smoothing than later-stage refinement. + +Approaches to Updating the Realistic Latent in SAG. Table 4 presents a comparative analysis of different strategies for updating the reference realistic-latent within SAG. Notably, updating the realistic latent after 100 steps yields superior results across all metrics when compared to enhancements at every step. For instance, this strategy leads to an increase in PSNR from 30.16 to $31.40$ at $\times 2$ , from 30.33 to $30.94$ at $\times 3$ , and from 29.28 to $30.1$ at $\times 4$ on Set14. These findings suggest that enhancing the reference latent later in the process correlates with improved performance. + +Influence of Artifact Detection on SARGD. Table 5 contrasts SARGD's performance with and without the implementation of artifact detection in the latent space during inference, providing a line-by-line comparative analysis. It is observed that the elimination of perceptual artifacts in the latent space significantly enhances super-resolution quality, as evidenced by improved metrics such as PSNR (e.g., 30.84 vs. 30.06 on B100 at $\times 3$ scaling) and DISTS (e.g., 0.219 vs. 0.140 on B100 at $\times 3$ scaling), highlighting the benefits of integrating artifact detection within SARGD. + +Table 5. Performance of SARGD with artifact detection ( $\mathcal{A}$ ) versus without in latent space across $\times 2$ , $\times 3$ , and $\times 4$ scales. Here, $\mathcal{X}$ signifies the substitution of RGR with a direct sum of the latent and the initial realistic reference latent. + +
ScaleASet14B100
PSNR↑SSIM↑DISTS↓PSNR↑SSIM↑DISTS↓
×2X31.240.8250.13830.560.8010.158
✓31.400.8420.12630.920.8170.132
×3X30.470.7960.19330.060.7830.219
✓30.940.8230.12130.840.8000.140
×4X29.980.7730.23430.150.7530.260
✓30.010.7780.15630.230.7630.156
+ +![](images/5f2bfa1c2777dc2c04b3a91f0b84c8e0043225007b4b2369ea8826e7fd3b6049.jpg) +Figure 7. Performance comparison of SARGD using different time steps on Set14 for $\times 2$ , $\times 3$ , $\times 4$ upscaling using PSNR. Our SARGD reaches peak score between 75 and 100 steps, effectively reducing the total inference steps by half, from 200 to 100. + +Impact on Inference Steps for SARGD. Figure 7 illustrates that our proposed SARGD not only enhances image quality but also decreases the number of total inference steps required. Specifically, the performance of SARGD in super-resolution steadily improves from the start and achieves optimal PSNR scores between 75 and 100 steps, such as a PSNR of 31.32 at 100 steps compared to 30.16 at 200 steps, which effectively cuts the inference time in half. This suggests that the removal of artifacts during the diffusion phase serves to improve image quality in SR while simultaneously boosting the speed of the inference process. + +# 5. Conclusions + +In this study, we tackle the issue of artifacts that arise in diffusion-based super-resolution. Our SARGD incorporates two key mechanisms: Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). RGR is designed to identify and correct artifacts within the latent space by utilizing a realistic latent as a reference, thereby improving the image's fidelity. To overcome the over-smoothing effect that comes from employing an initial realistic latent for guidance, SAG is introduced to refine the reference latent, improving the quality of the super-resolved images. + +Acknowledgements. This work is supported by the National Natural Science Foundation of China (62271400), and the Shaanxi Provincial Key R&D Program, China (2023-GHZD-02). + +# References + +[1] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 6, 7 +[2] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 2 +[3] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 6 +[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. in 2021 IEEE. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356, 2021. 3 +[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12413-12422, 2022. 3 +[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations (ICLR), 2023. 2 +[7] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, 2020. 3 +[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 38(2):295-307, 2015. 1, 2 +[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 391-407. Springer, 2016. 2 +[10] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7890–7899, 2020. 3 +[11] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in Neural Information Processing Systems (NIPS), 33:3022-3032, 2020. 3 +[12] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1909-1918, 2022. 2 + +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems (NIPS), 33:6840-6851, 2020. 2, 3, 4 +[14] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6 +[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711, 2016. 2 +[16] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4681-4690, 2017. 2 +[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 1, 2, 3 +[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using wwin transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1833-1844, 2021. 2 +[19] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4076-4085, 2021. 2 +[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 136-144, 2017. 2 +[21] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8060-8069, 2020. 3 +[22] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 715-732, 2020. 2 +[23] Andreas Lugmayr, Martin Danelljan, Fisher Yu, Luc Van Gool, and Radu Timofte. Normalizing flow as a flexible fidelity objective for photo-realistic super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1756-1765, 2022. 2 +[24] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms + +and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 416-423, 2001. 5, 6, 7 +[25] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 5, 6, 7 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 6 +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 2 +[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4): 4713-4726, 2022. 1, 3 +[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1, 2 +[30] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICLR), pages 2256-2265, 2015. 2 +[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 1 +[32] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 6, 7 +[33] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 2 +[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV) workshops, pages 0–0, 2018. 2 +[35] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5 +[36] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: + +Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2 +[37] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2 +[38] Xin Yang, Haiyang Mei, Jiqing Zhang, Ke Xu, Baocai Yin, Qiang Zhang, and Xiaopeng Wei. Drfn: Deep recurrent fusion network for single-image super-resolution with large factors. IEEE Transactions on Multimedia (TMM), 21(2): 328-337, 2018. 1 +[39] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7556-7566, 2019. 3 +[40] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 3 +[41] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International Conference on Curves and Surfaces, pages 711-730. Springer, 2012. 5, 6, 7 +[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 2 +[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7579-7590, 2023. 4, 5, 6 +[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5 +[45] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3096-3105, 2019. 2 +[46] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In IEEE international Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 3 +[47] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 286-301, 2018. 5 +[48] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2 + +[49] Yulun Zhang, Kai Li, Kunpeng Li, and Yun Fu. Mr image super-resolution with squeeze and excitation reasoning attention network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13425-13434, 2021. 1 +[50] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2185–2194, 2021. 3 \ No newline at end of file diff --git a/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/images.zip b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..27df4b23154bdcfdc6df5cd95cc7f1b965542e5a --- /dev/null +++ b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:764cbb9f3eebf01360c57c467e29cc9ac32389831e3ff0766a8583cc5eb5c38f +size 786063 diff --git a/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/layout.json b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2c458566f0e44e94f7a2b230d1512dc28045d81 --- /dev/null +++ b/2024/Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution/layout.json @@ -0,0 +1,10074 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 526, + 121 + ], + "type": "text", + "content": "Self-Adaptive Reality-Guided Diffusion for Artifact-Free Super-Resolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "content": "Qingping Zheng" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "content": ", Ling Zheng" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "content": ", Yuanfan Guo" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "content": ", Ying Li" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "content": ", Songcen Xu" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "content": ", Jiankang Deng" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "text", + "content": ", Hang Xu" + }, + { + "bbox": [ + 56, + 142, + 539, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "spans": [ + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "type": "text", + "content": "Northwestern Polytechnical University " + }, + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab " + }, + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 160, + 514, + 174 + ], + "type": "text", + "content": "Tsinghua-Fuzhou Institute for Data Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 201, + 192, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 192, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 192, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 227, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 290, + 586 + ], + "type": "text", + "content": "Artifact-free super-resolution (SR) aims to translate low-resolution images into their high-resolution counterparts with a strict integrity of the original content, eliminating any distortions or synthetic details. While traditional diffusion-based SR techniques have demonstrated remarkable abilities to enhance image detail, they are prone to artifact introduction during iterative procedures. Such artifacts, ranging from trivial noise to unauthentic textures, deviate from the true structure of the source image, thus challenging the integrity of the super-resolution process. In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free method that delves into the latent space to effectively identify and mitigate the propagation of artifacts. Our SARGD begins by using an artifact detector to identify implausible pixels, creating a binary mask that highlights artifacts. Following this, the Reality Guidance Refinement (RGR) process refines artifacts by integrating this mask with realistic latent representations, improving alignment with the original image. Nonetheless, initial realistic-latent representations from lower-quality images result in over-smoothing in the final output. To address this, we introduce a Self-Adaptive Guidance (SAG) mechanism. It dynamically computes a reality score, enhancing the sharpness of the realistic latent. These alternating mechanisms collectively achieve artifact-free super-resolution. Extensive experiments demonstrate the superiority of our method, delivering detailed artifact-free high-resolution images while reducing sampling steps by " + }, + { + "bbox": [ + 47, + 227, + 290, + 586 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 47, + 227, + 290, + 586 + ], + "type": "text", + "content": ". We release our code at https://github.com/ProAirVerse/Self-Adaptive-Guidance-Diffusion.git." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 611, + 128, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 128, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 128, + 623 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "content": "Super-resolution (SR) techniques strive to restore high-resolution images from their corresponding low-resolution counterparts [8, 29]. This process is inherently ill-posed since the same LR image can yield many plausible HR solutions. Traditional deep learning SR methods, leverag" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 202, + 357, + 277 + ], + "blocks": [ + { + "bbox": [ + 312, + 202, + 357, + 277 + ], + "lines": [ + { + "bbox": [ + 312, + 202, + 357, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 202, + 357, + 277 + ], + "type": "image", + "image_path": "d4bc0e146e9d72cf6e9167554118e0cf422a8e20cee50d253a888196a3075801.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 358, + 204, + 403, + 277 + ], + "blocks": [ + { + "bbox": [ + 358, + 204, + 403, + 277 + ], + "lines": [ + { + "bbox": [ + 358, + 204, + 403, + 277 + ], + "spans": [ + { + "bbox": [ + 358, + 204, + 403, + 277 + ], + "type": "image", + "image_path": "61918e4d994c8fcd04bb49380a6364cafa7e4125f75448ac3b62f3899d40837d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 277, + 484, + 285 + ], + "lines": [ + { + "bbox": [ + 369, + 277, + 484, + 285 + ], + "spans": [ + { + "bbox": [ + 369, + 277, + 484, + 285 + ], + "type": "text", + "content": "(a) Variation of Artifacts within StableSR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 403, + 203, + 449, + 277 + ], + "blocks": [ + { + "bbox": [ + 403, + 203, + 449, + 277 + ], + "lines": [ + { + "bbox": [ + 403, + 203, + 449, + 277 + ], + "spans": [ + { + "bbox": [ + 403, + 203, + 449, + 277 + ], + "type": "image", + "image_path": "10c72e833a6efd78c3de57131df0fcf4ec3bf7e204670035b6db38179fa226b7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 449, + 204, + 494, + 277 + ], + "blocks": [ + { + "bbox": [ + 449, + 204, + 494, + 277 + ], + "lines": [ + { + "bbox": [ + 449, + 204, + 494, + 277 + ], + "spans": [ + { + "bbox": [ + 449, + 204, + 494, + 277 + ], + "type": "image", + "image_path": "1941dd32834c13a109fa996d392e0e40a6d0d8a0705dbfb89d4792dbd9b18cfd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 495, + 204, + 540, + 277 + ], + "blocks": [ + { + "bbox": [ + 495, + 204, + 540, + 277 + ], + "lines": [ + { + "bbox": [ + 495, + 204, + 540, + 277 + ], + "spans": [ + { + "bbox": [ + 495, + 204, + 540, + 277 + ], + "type": "image", + "image_path": "7dd435689f20a6d3d9c8e9954231b535e63e905a920bb27e4308fcc90ed593d8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 286, + 357, + 354 + ], + "blocks": [ + { + "bbox": [ + 312, + 286, + 357, + 354 + ], + "lines": [ + { + "bbox": [ + 312, + 286, + 357, + 354 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 357, + 354 + ], + "type": "image", + "image_path": "914bf025e0b8d72ea5f09722a1c206a393ae764afc303d4d4acbb19fc8e97268.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 362, + 545, + 429 + ], + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 429 + ], + "type": "text", + "content": "Figure 1. Visual comparison of artifact variability between (a) StableSR and (b) Our SARGD method. In the lower panel, the left image is the decoded output, while the right image is the artifact mask. Regions containing artifacts are highlighted in red, utilizing a binary map to serve as the artifact mask. Our SARGD method exhibits a superior capability in reducing artifacts." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 358, + 286, + 403, + 354 + ], + "blocks": [ + { + "bbox": [ + 358, + 286, + 403, + 354 + ], + "lines": [ + { + "bbox": [ + 358, + 286, + 403, + 354 + ], + "spans": [ + { + "bbox": [ + 358, + 286, + 403, + 354 + ], + "type": "image", + "image_path": "24e794b3084e1e59c85d0c8c1380d602d7a87f77a78795be268d5e580938e193.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 354, + 489, + 361 + ], + "lines": [ + { + "bbox": [ + 363, + 354, + 489, + 361 + ], + "spans": [ + { + "bbox": [ + 363, + 354, + 489, + 361 + ], + "type": "text", + "content": "(b) Variation of Artifacts within Our SARGD" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 403, + 286, + 449, + 354 + ], + "blocks": [ + { + "bbox": [ + 403, + 286, + 449, + 354 + ], + "lines": [ + { + "bbox": [ + 403, + 286, + 449, + 354 + ], + "spans": [ + { + "bbox": [ + 403, + 286, + 449, + 354 + ], + "type": "image", + "image_path": "059b17e30a51c8119adc7823dcd04ef72843909874b928bbeeadb49015c44df0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 449, + 286, + 494, + 354 + ], + "blocks": [ + { + "bbox": [ + 449, + 286, + 494, + 354 + ], + "lines": [ + { + "bbox": [ + 449, + 286, + 494, + 354 + ], + "spans": [ + { + "bbox": [ + 449, + 286, + 494, + 354 + ], + "type": "image", + "image_path": "c203cbce8b899368523e55fe7ef36be34de158b6f3297a981f435e2b0f208b8f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 495, + 286, + 540, + 354 + ], + "blocks": [ + { + "bbox": [ + 495, + 286, + 540, + 354 + ], + "lines": [ + { + "bbox": [ + 495, + 286, + 540, + 354 + ], + "spans": [ + { + "bbox": [ + 495, + 286, + 540, + 354 + ], + "type": "image", + "image_path": "03022f297f97a931630f37e10115df87b7bf97548f27eaae120e5fec916dcae0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 437, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 581 + ], + "type": "text", + "content": "ing techniques like squeeze layers [49], transposed convolutions [38], and sub-pixel convolutions [29], bind themselves to fixed upsampling scales, limiting output resolution adjustments without architectural changes. This causes inflexibility in real-world applications. To tackle the inherent one-to-many inverse challenge in SR, it is intuitive to envision the solution space as a distribution. Yet, this perspective often prompts generative models to introduce details that deviate from the authentic content of the original low-resolution image. As a result, a wide range of generative-based SR approaches have emerged to address this intrinsically ill-posed problem, particularly in diffusion models." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Diffusion-based SR (Diffusion-SR for short) has exhibited significant potential in improving the resolution of images from their low-resolution counterparts. SRDiff [17] is the first diffusion-based model for image super-resolution, progressively transforming the Gaussian noise into a high-resolution (HR) image conditioned on its low-resolution (LR) input through a Markov chain. Building on the foundational concept of denoising score matching [31], Saharia et al. [28] propose SR3 model. This model advances the field of super-resolution by learning an empirical data distribution and leveraging U-Net architec" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25806" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "ture [27] to progressively denoise and refine the image output. However, these methods fail to preserve accurate pixelwise image structures, often necessitating further training or additional skipped connections for detail reproduction. While the pixel-aware stable diffusion network proposed by Yang et al. [37] significantly improves the robustness of the super-resolution process, it concurrently introduces the challenge of generating implausible content. In contrast, StableSR [32], the current state-of-the-art, leverages pre-trained diffusion priors to retain image details, avoiding assumptions about image degradation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 507 + ], + "type": "text", + "content": "Nevertheless, as depicted in Figure 1, contemporary diffusion-based StableSR models invariably generate implausible content, frequently resulting in perceptual artifacts within specific areas of the image. Indeed, these artifacts frequently stem from the fundamental use of Denoising Diffusion Probabilistic Models (DDPMs) [13, 30] in modern diffusion models. Specifically, DDPMs, categorized as Partial Differential Equation (PDE)-based models [6], are designed to diffuse pixel intensities in order to reduce noise levels. They can enhance the preservation of edges and textural features by integrating non-linear terms, thus improving image details. However, despite these advancements, the introduction of non-linearity can occasionally yield unexpected outcomes, particularly in areas with intricate textures or noise patterns. This unintended consequence can result in a loss of fine details and a degradation of essential image elements such as edges and textures, especially when the diffusion process is overly aggressive or inadequately constrained. Additionally, any deficiencies in the model's architecture or parameter settings can be amplified during diffusion, aggravating artifact problems and producing results that deviate from the desired photorealistic quality [36]. These challenges pose a significant obstacle to the preservation of the authentic detail and clarity expected in high-resolution images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "In this work, we propose Self-Adaptive Reality-Guided Diffusion (SARGD), a pioneering training-free approach, aiming to produce \"artifact-free\" SR images. In particular, SARGD operates through two stages: Reality-Guided Refinement (RGR) to reduce artifacts and Self-Adaptive Guidance (SAG) to boost image fidelity. During the initial RGR phase, an artifact detector is employed to identify unrealistic pixels within the latent space at each inference step, leading to a binary artifact mask. This mask collaborates with a realistic latent guidance extracted from an upscaled LR image to refine artifacts and enhance intrinsic details. However, depending on upscaled LR inputs for guidance can potentially result in over-smoothed images, as it may overlook high-frequency details that are essential for SR. To alleviate this limitation, the subsequent SAG phase introduces a 'reality score to self-adjust the authenticity of the latent guidance, significantly boosting the accuracy of" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "detail and texture representation. Two phases alternate to progressively achieve high-fidelity super-resolution, effectively eliminating artifacts in the diffusion process. To sum up, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 121, + 545, + 288 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 121, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 168 + ], + "type": "text", + "content": "- We introduce a Reality-Guided Refinement (RGR) strategy, designed to identify and rectify artifacts during the diffusion process with the aid of a reality guidance mechanism, effectively eliminating such artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 168, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 545, + 216 + ], + "type": "text", + "content": "- We propose a Self-Adaptive Guidance (SAG) mechanism that iteratively enhances the realistic-latent guidance, addressing over-smoothing issues and improving the fidelity and authenticity of the reference latent guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 217, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 545, + 288 + ], + "type": "text", + "content": "- We are the first to develop the Self-Adaptive Reality-Guided Diffusion (SARGD), a training-free approach that sets a new benchmark by effectively dealing with artifacts and over-smoothing issues. Additionally, it outperforms StableSR in image quality while reducing the inference time by " + }, + { + "bbox": [ + 306, + 217, + 545, + 288 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 306, + 217, + 545, + 288 + ], + "type": "text", + "content": " for super-resolution tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 300, + 392, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 392, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 392, + 312 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "type": "text", + "content": "In this section, we summarize past super-resolution (SR) techniques, explore diffusion-based models in SR, and address artifact detection and mitigation in the SR process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 546, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 613 + ], + "type": "text", + "content": "Image Super-Resolution. Single image super-resolution (SISR) is a process designed to construct high-resolution (HR) images from their low-resolution (LR) counterparts, which is inherently a complex task due to the multiple HR images that can correspond to a single LR image. The literature has seen a variety of solutions aiming to learn mappings from LR to HR images. These solutions fall into two primary groups: PSNR-oriented methods and generative model-based methods. PSNR-oriented methods [8, 9, 18, 20, 29, 48] employ L1 or L2 norms as objectives, achieving notable PSNR metrics. However, the reliance on L1/L2 losses tends to predict an average of all possible HR images [15, 16, 22], resulting in oversmoothed outputs. In contrast, generative model-based methods [2, 12, 16, 17, 19, 22, 23, 33, 34, 42, 45] seek to address the ill-posed nature of the SR problem by learning the distribution of potential HR images through techniques such as GAN-based, diffusion-based, flow-based, and AR-based super-resolution. These models aim to generate HR images that not only have high pixel accuracy but also possess the rich textural details necessary for visual realism." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Diffusion Models. Diffusion probabilistic models [13, 30] are a class of generative models that utilize a Markov chain to convert latent variables from simple distributions (e.g., Gaussian) into data with complex distributions. Recognizing the potential of these models to address \"one-to-many\" mapping problems, Li et al. [17] have applied them to the task of super-resolution, generating a variety of high-resolution images from a single low-resolution input. This" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25807" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 542, + 247 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 542, + 247 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 542, + 247 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 542, + 247 + ], + "type": "image", + "image_path": "1c10dfa7f22ffd95d0957a7538d16645962da4fb93305e5c04c798041f0dfc69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 548, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 548, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 548, + 304 + ], + "type": "text", + "content": "Figure 2. Self-Adaptive Reality-Guided Diffusion (SARGD) for artifact-free super-resolution. Our proposed SARGD is a training-free approach that consists of two principal components: 1) a Reality-Guided Refinement (RGR) that identifies and corrects artifacts within the latent representation by using a realistic latent as a guide to maintain the inherent details of the original image during the diffusion process, and 2) a Self-Adaptive Guidance (SAG) mechanism that enhances the fidelity of the initial realistic latent guidance, derived from upscaled low-resolution images, thereby effectively addressing the issue of over-smoothing in the final outputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 289, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 496 + ], + "type": "text", + "content": "application not only yields diverse outcomes but also simultaneously addresses challenges related to over smoothing, mode collapse, and the extensive computational footprint that are often encountered in super-resolution tasks. A prevalent method [26, 28] in diffusion-based SR involves integrating a low-resolution image directly into the input of existing diffusion models (e.g., DDPM [13]), and retraining the model with super-resolution data from scratch. Another approach [4, 5, 40] leverages an unconditional pretrained diffusion model as a prior and adapts its reverse diffusion process to synthesize the desired high-resolution image. Nonetheless, both strategies are constrained by the intrinsic Markov chain sequence of DDPM, which can potentially lead to artifacts during inference, thus affecting the quality of the resulting high-resolution image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "Artifact Detection Methods. Despite impressive improvement, diffusion-based super-resolution methods tend to present undesired outcomes, known as artifacts. The artifact detection approach aims to probe the perceptually implausible contents rooted in the generated images. Existing deep-learning-based image artifact detection methods can be divided into spatial and frequency domain-based approaches. The former approaches [39] focus on capturing differences in texture between real and generated images. Liu et al. [21] leverage Gram matrices to capture the difference in texture between real and synthetic images. Dang et al. [7] adopt the attention mechanism to improve the accuracy of artifact detection by highlighting informative regions. Zhao et al. [50] utilize the attention mechanism to amplify subtle differences in the shallow layers, improving the detection performance. The latter approaches [46] study the artifacts in the frequency domain. Durall et al. [10] classify generated images according to the characteristics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 316, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 389 + ], + "type": "text", + "content": "of the artifacts in high-frequency components. Dzanic et al. [11] demonstrate that image artifacts could be detected by leveraging the degree of partial decay at high frequencies. Frank et al. [11] find that the use of upsampling operations causes the artifacts in generated images, and employs the DCT transforms to detect them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 398, + 389, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 398, + 389, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 389, + 412 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 418, + 547, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 477 + ], + "type": "text", + "content": "In this section, we propose a novel training-free method called \"Self-Adaptive Reality-Guided Diffusion\" (SARGD) to tackle the ongoing challenge of artifact suppression in diffusion-based super-resolution. This approach, depicted in Figure 2, introduces two essential mechanisms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 478, + 547, + 574 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 478, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 547, + 525 + ], + "type": "text", + "content": "- Reality-Guided Refinement (RGR), mitigates artifact emergence during inference by employing guidance from a realistic latent space, thereby preserving the original characteristics of the images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 525, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 547, + 574 + ], + "type": "text", + "content": "- Self-Adaptive Guidance (SAG), enhances the authenticity of the reference realistic latent representation, effectively addressing the over-smoothing tendencies associated with the initial realistic latent representation." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 580, + 392, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 580, + 392, + 592 + ], + "spans": [ + { + "bbox": [ + 306, + 580, + 392, + 592 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 598, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 547, + 696 + ], + "type": "text", + "content": "Diffusion-based super-resolution (SR) techniques [17, 28, 32] have emerged as a compelling paradigm for the enhancement of image resolutions. Consider the diffusion process as a function that aims to translate a given low-resolution (LR) image into a high-resolution (HR) image, each successive HR image at time " + }, + { + "bbox": [ + 304, + 598, + 547, + 696 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 598, + 547, + 696 + ], + "type": "text", + "content": " is obtained by refining the current estimate, factoring in the original LR image. This refinement is mathematically represented as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 700, + 546, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 700, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 365, + 700, + 546, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {H R} ^ {(t + 1)} = \\mathbf {I} _ {H R} ^ {(t)} + f _ {\\theta} (\\mathbf {I} _ {H R} ^ {(t)}, \\mathbf {I} _ {L R}), \\qquad (1)", + "image_path": "41db9eae9862b9278381ceff8b0131cfca79b1cf661f89d670ed5b2af7323296.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25808" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 76, + 277, + 208 + ], + "blocks": [ + { + "bbox": [ + 59, + 76, + 277, + 208 + ], + "lines": [ + { + "bbox": [ + 59, + 76, + 277, + 208 + ], + "spans": [ + { + "bbox": [ + 59, + 76, + 277, + 208 + ], + "type": "image", + "image_path": "492e361481c399e45d8b42396f08e58bbb67f5d74a33f924507e52bcbb0c5ed1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 270 + ], + "type": "text", + "content": "Figure 3. Overview of the Reality-Guided Refinement (RGR) workflow: 1) Decoding the current latent into an RGB image and using an artifact detector to create a binary mask identifying areas with artifacts; and 2) Utilizing realistic latent guidance to refine the masked regions, enhancing the image's fidelity and authenticity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{HR}^{(t)}" + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "text", + "content": " is the current state of the high-resolution image at iteration " + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "text", + "content": " is the diffusion function parameterized by " + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "text", + "content": ", which updates the current high-resolution state based on both its current state and the given low-resolution image. " + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 281, + 287, + 343 + ], + "type": "text", + "content": " is the current iteration during the inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "text", + "content": "In the iterative diffusion process for super-resolution, the input LR-HR image pairs in the training set are used to train the diffusion models with the total diffusion step T. The diffusion process initially leverages a pre-trained visual encoder to convert the upsampled LR image into a latent " + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "inline_equation", + "content": "z_{t} = \\mathcal{E}(up(\\mathbf{I}_{LR}))" + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "text", + "content": ", and gradually adding the Gaussian noise " + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "text", + "content": " into it, resulting the noised latent variable " + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "text", + "content": ". The reverse diffusion process aims to guide the generation to the corresponding HR image through iterative refinement of the noised latent " + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 344, + 287, + 463 + ], + "type": "text", + "content": ", taking the form of:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 472, + 287, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 472, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 52, + 472, + 287, + 505 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t - 1} \\leftarrow \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {y} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\gamma_ {t}}} f _ {\\theta} \\left(\\boldsymbol {x}, \\boldsymbol {y} _ {t}, \\gamma_ {t}\\right)\\right) + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {2}", + "image_path": "0c948569d13762dc9408dfa3daf9442e2cb7c94ad7305af0c535c47eb7bb49dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\in (0,1)" + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "content": " is the corresponding coefficient and " + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "content": " is the variance of the noise at iteration " + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "content": ". Lastly, the decoder " + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "content": " is employed to project the resulting latent into the final high-resolution image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 232, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 232, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 232, + 574 + ], + "type": "text", + "content": "3.2. Reality-Guided Refinement (RGR)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "To overcome the prevalent dilemma of aggravated noise and the emergence of artifacts in images subjected to superresolution (SR) via diffusion-based techniques, we propose a pioneering training-free approach: Reality-Guided Refinement (RGR). This method couples an artifact detector with a latent refinement to address these issues effectively. As illustrated in Figure 3, a pixel-level segmentation model is initially applied to locate any artifacts on the current latent variable. Once detected, these imperfections are adeptly rectified through a sophisticated procedure of realistic-latent guidance refinement. This diffusion process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "ensures that the resulting images retain the original, authentic textures and intricate details, effectively circumventing the usual distortions and artifacts associated with traditional diffusion-based upsampling techniques for SR tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 123, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 255 + ], + "type": "text", + "content": "Artifact Detection in Diffusion Process. In the context of pixel-sensitive super-resolution, it is crucial to identify the fine-grained artifacts during the diffusion process. Hence, a robust artifact detection system should precisely pinpoint these implausible pixels within the latent space for effective identification. To achieve this, we utilize the established artifact detection method named PAL [43] to isolate these imperfections in the latent space. The process entails transforming the latent variable into the image domain at each iteration using the decoder, depicted by the following equation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 262, + 545, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 545, + 294 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {H R} ^ {(t - 1)} = \\mathcal {D} \\left(\\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {x}, t)\\right) + \\sigma_ {\\theta} \\left(x _ {t}, t\\right) \\boldsymbol {\\epsilon})\\right), \\tag {3}", + "image_path": "7c6e57cccd4c66b3f20a7c548815f2dd3af64d85046634044ae18f159172c043.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 295, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 342 + ], + "type": "text", + "content": "Subsequently, the resulting image is input into the PAL detector, yielding a binary mask " + }, + { + "bbox": [ + 304, + 295, + 545, + 342 + ], + "type": "inline_equation", + "content": "E_{A}" + }, + { + "bbox": [ + 304, + 295, + 545, + 342 + ], + "type": "text", + "content": " that denotes the presence of artifacts (referenced as line 12 in Algorithm 1). This process can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 385, + 348, + 545, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 348, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 385, + 348, + 545, + 363 + ], + "type": "interline_equation", + "content": "E _ {A} ^ {(t - 1)} = \\mathcal {A} \\left(\\mathbf {I} _ {H R} ^ {(t - 1)}\\right). \\tag {4}", + "image_path": "7c3e867c3019899e64334d525f856cf8f59a02d708030706adbb1eefe7cb44fc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 370, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 405 + ], + "type": "text", + "content": "Finally, the resultant artifact mask is resized to match the scale of the latent variable, maintaining the proportionality between the two." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "content": "Refinement via Reality Guidance. For the removal of artifacts during sampling, we refine the latent variable via a proposed training-free RGR. Following Algorithm 1, the RGR inference, consisting of T steps, begins by taking an upscaled low-resolution image " + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "inline_equation", + "content": "up(\\mathbf{I}_{LR})" + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "content": " as the starting point. A noise vector " + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "content": " is then drawn from a standard Gaussian distribution. The process iterates backward from " + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "content": ", with each step generating a latent variable " + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t-1}" + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "content": " in accordance with DDPM [13] framework. Once artifacts " + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "inline_equation", + "content": "E_A" + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "content": " are identified within the current latent " + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t-1}" + }, + { + "bbox": [ + 304, + 409, + 545, + 553 + ], + "type": "text", + "content": ", we proceed to refine these imperfections through reality guidance. This correction process unfolds as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 560, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 560, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 351, + 560, + 545, + 572 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t - 1} = \\boldsymbol {x} _ {t - 1} \\times (1 - E _ {A}) + \\boldsymbol {x} _ {r} \\times E _ {A}. \\tag {5}", + "image_path": "63b9963fef9f571da39e4588938a0f529a7fdd3c0d95988c705ce2629fc9090a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": "This process allows us to effectively eliminate artifacts from the latent space. Notably, in this formula, we employ the initially encoded latent, which has not undergone the denoising process, as a source of realistic latent guidance, since it retains the intrinsic structural integrity of the image." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 471, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 471, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 471, + 660 + ], + "type": "text", + "content": "3.3. Self-Adaptive Guidance (SAG)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "However, relying on the initial latent space as a guide for artifact suppression during inference often results in oversmoothing results, primarily attributed to reliance on upscaled low-resolution images. To address this, we propose a" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25809" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "novel Self-Adaptive Guidance (SAG) renewal mechanism, an automated system that perpetually refines realistic latent representations, ensuring the guidance is continually calibrated to enhance realistic detail. This mechanism is indispensable for diffusion-based super-resolution, aiming to yield high-resolution and high-fidelity images that strike a balance between details and sharpness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": "Reality Score Computation. To regulate the update mechanism of reality guidance in image super-resolution, we introduce an innovative metric known as the reality score. This metric quantifies the realism of latent guidance. As delineated in Algorithm 1, we begin with an artifact-free latent variable " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1}" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": " and utilize our decoding function to synthesize a realistic image " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r = \\mathcal{D}(\\pmb{x}_{t-1})" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": ". Subsequently, we deploy a realistic-region function " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": " by inverting the output of PAL [43] on " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": ", resulting in a binary reality mask " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "M_R" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": " with dimensions corresponding to the channels, height, and width of the latent " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "C \\times H \\times W" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": ", respectively. The reality score " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": " at the current " + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 46, + 158, + 289, + 338 + ], + "type": "text", + "content": " step is calculated as the normalized sum of the reality mask across the spatial dimensions of the latent, indicating the distribution of realistic features. The formula is given by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 342, + 287, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 342, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 83, + 342, + 287, + 373 + ], + "type": "interline_equation", + "content": "s _ {r} ^ {t - 1} = \\mathcal {S} \\left(M _ {R}\\right) = \\frac {1}{H \\times W} \\sum_ {i = 0} ^ {H - 1} \\sum_ {j = 0} ^ {W - 1} M _ {R} ^ {(i j)}. \\tag {6}", + "image_path": "13147161753ef47eaac01ae231425c95dca7cf64d11a18da0b227933f07f94c3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "text", + "content": "Herein, " + }, + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "text", + "content": " are the height and width of the latent, and " + }, + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "inline_equation", + "content": "M_R^{(ij)}" + }, + { + "bbox": [ + 47, + 376, + 287, + 438 + ], + "type": "text", + "content": " indicates the reality measure at each pixel. This approach ensures a quantitative evaluation of the SR process, focusing on producing images that boast both high resolution and a high degree of fidelity to reality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "content": "Self-Adaptive Reality Guidance. Self-adaptability in the realistic latent space refers to the ability of the system to adjust the reality guidance to correct the details and textures that are characteristic of high-quality images. This is especially important to mitigate the smoothing effect of the initial realistic-latent reliance on bicubic upsampling. As outlined in Algorithm 1, we begin by considering a realistic image " + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "content": " at the current step, decoded from an artifact-free latent variable " + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "content": ". This artifact-free image is then encoded into a realistic latent " + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "inline_equation", + "content": "x_{r}^{t - 1} = \\mathcal{E}(\\mathbf{I}_r)" + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "content": ". Grounded on the computed reality score, self-adaptability is characterized as a guidance refinement process. During this process, the realistic latent " + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r" + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "content": " at the current " + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "content": " step is iteratively refined to converge towards an ideal latent representation " + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r" + }, + { + "bbox": [ + 46, + 441, + 289, + 632 + ], + "type": "text", + "content": ", precisely aligning with the high-resolution output IHR. The equation below illustrates this adaptive mechanism:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 635, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 635, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 52, + 635, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {G} \\left(\\boldsymbol {x} _ {r}, \\boldsymbol {x} _ {r} ^ {t - 1}\\right) = \\boldsymbol {x} _ {r} \\times \\left(1 - M _ {R}\\right) + \\boldsymbol {x} _ {r} ^ {t - 1} \\times M _ {R}, \\text {i f} \\boldsymbol {s} _ {r} ^ {t - 1} > \\boldsymbol {s} _ {r}, \\tag {7}", + "image_path": "c8999a18321290b8e50fd39189ef0a768a7ba95d4df2363cb36258e0f65ad3c9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": " represents the reality score associated with " + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "xr" + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "M_R" + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": " is the binary reality mask corresponding to " + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "xr^{t-1}" + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": ". This self-adaptive mechanism ensures that the guidance is autonomously updated, leading to enhanced superresolution outcomes." + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 308, + 85, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 523, + 83 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 523, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 523, + 83 + ], + "type": "text", + "content": "Algorithm 1 Self-Adaptive Reality-Guided Diffusion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": "1: Input: LR image " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{LR}" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": ", and total diffusion steps " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n2: Load: Encoder " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": ", artifact detector " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " and LR decoder " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n3: Step 1: Initialization \n4: Upscale LR image as " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "up(\\mathbf{I}_{LR})" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n5: Encode the upsampled image as " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x} = \\mathcal{E}(up(\\mathbf{I}_{LR}))" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n6: Initialize the " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " as a realistic latent " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " and set it as guidance \n7: Compute the Realty score of the realistic latent " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{s}_r" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n8: Step 2: Sampling \n9: for " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " do \n10: Sample " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": ", else " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon = 0" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n11: Compute the latent variable at the current step " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1} = \\frac{1}{\\sqrt{\\alpha_t}} (\\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\alpha_t}} \\epsilon_\\theta (\\pmb{x}_t, \\pmb{x}, t)) + \\sigma_\\theta (\\pmb{x}_t, t) \\epsilon" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n12: Reality-Guided Refinement \n13: Detect artifacts of the current latent " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "E_A = \\mathcal{A}(\\mathcal{D}(\\pmb{x}_{t-1}))" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n14: Refine the latent " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1} = \\pmb{x}_{t-1} \\times (1 - E_A) + \\pmb{x}_r \\times E_A" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n15: Reality Score Computation \n16: Decode the refined latent into an image " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r = \\mathcal{D}(\\pmb{x}_{t-1})" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n17: Generate the current binary reality map " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "M_R = \\mathcal{R}(\\mathbf{I}_r)" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n18: Calculate the current reality score " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "s_r^{t-1} = S(M_R)" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n19: Self-Adaptive Guidance \n20: Encode the current realistic latent " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r^{t-1} = \\mathcal{E}(\\mathbf{I}_r)" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n21: Update the guidance " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r = \\mathcal{G}(\\pmb{x}_r, \\pmb{x}_r^{t-1})" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "s_r^{t-1} > s_r" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n22: Update the reality score " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{s}_r = s_r^{t-1}" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "s_r^{t-1} > s_r" + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "text", + "content": " \n23: end for \n24: return the artifact-free SR " + }, + { + "bbox": [ + 308, + 85, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{HR} = \\mathcal{D}(\\pmb{x}_0)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 389, + 432, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 432, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 432, + 403 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 409, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 409, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 547, + 445 + ], + "type": "text", + "content": "In this section, we present our experimental results, discuss insights from ablation studies, and delve into the broader implications of our findings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 453, + 426, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 426, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 426, + 466 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 651 + ], + "type": "text", + "content": "Testing Datasets. Our proposed SARGD method is rigorously assessed for arbitrary-scale single image superresolution (SR) across a comprehensive set of benchmark datasets, including Set5 [1], Set14 [41], B100 [24], Urban100 [14], and Manga109 [25]. The Set5 and Set14 datasets serve as standard benchmarks within the super-resolution domain, comprising five and fourteen high-resolution images, respectively. The B100 dataset is a collection of 100 natural images used for testing the versatility of super-resolution algorithms across various real-world scenarios. Urban100 comprises 100 detailed urban images, challenging super-resolution algorithms to accurately enhance intricate structures. Manga109 contains 109 high-resolution manga images, testing the edge preservation capabilities of super-resolution models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We utilize established perceptual metrics including PSNR, SSIM [35], LPIPS [44], and DISTS [47] to conduct a comprehensive comparison with existing diffusion-based SR approaches. PSNR and SSIM, particularly when assessed on the Y channel of the YCbCr" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25810" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 104, + 547, + 247 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with the state-of-the-art Diffusion-SR methods across benchmark datasets for super-resolution at " + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "text", + "content": " scales. The dagger symbol (" + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 71, + 547, + 103 + ], + "type": "text", + "content": ") signifies LDM trained on the identical dataset as used for StableSR. Bold highlights the best performance. All evaluations are conducted on a 32G GPU. Our training-free SARGD attains the most favorable results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 104, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 48, + 104, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 104, + 547, + 247 + ], + "type": "table", + "html": "
BenchmarksSet5 [1]Set14 [41]B100 [24]Urban100 [14]Manga109 [25]
MetricsMethods×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑LDM†27.8627.9428.3026.4226.8127.2126.9127.5427.6725.1825.7225.9126.4926.9627.04
StableSR28.7829.0929.3627.2627.6928.0427.6128.2828.4026.0026.4426.5527.5828.0127.94
Ours32.6433.0632.2731.4030.9430.0130.9230.8430.2329.3428.8427.9331.2131.1330.23
SSIM ↑LDM†0.7650.7770.7690.6730.6880.6960.6510.6720.6790.7120.7190.7190.8140.8210.815
StableSR0.8050.8140.8120.7160.7310.7330.6940.7140.7170.7450.7480.7440.8430.8470.838
Ours0.8990.8970.8710.8420.8230.7780.8170.8000.7630.8490.8170.7710.9090.8970.851
LPIPS ↓LDM†0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1670.1760.1860.1320.1280.138
StableSR0.1650.1530.1500.1970.1940.2070.2000.2140.2260.1500.1590.1660.1180.1150.123
Ours0.1180.1260.1310.1390.1490.2040.1640.1180.2240.1170.1380.1770.0830.0850.128
DISTS ↓LDM†0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1290.1340.1410.1020.0980.103
StableSR0.1670.1650.1570.1520.1510.1500.1600.1600.1640.1270.1280.1310.1020.0970.099
Ours0.1360.1340.1420.1260.1210.1560.1320.1400.1560.1110.1190.1390.0860.0830.096
", + "image_path": "724bd76a06f65e7086b4a51c6394dee03fa3dd69b47ac5c2177b4417ee1c07a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 258, + 156, + 413 + ], + "blocks": [ + { + "bbox": [ + 51, + 258, + 156, + 413 + ], + "lines": [ + { + "bbox": [ + 51, + 258, + 156, + 413 + ], + "spans": [ + { + "bbox": [ + 51, + 258, + 156, + 413 + ], + "type": "image", + "image_path": "ad272e0cd85cd67a7aed2f9deade13911e932c9f18efcb133a02145218784736.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "type": "text", + "content": "Figure 4. Visual comparison with diffusion-SR methods for " + }, + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 415, + 546, + 449 + ], + "type": "text", + "content": " super-resolution, including (a) Bicubic upsampling, (b) StableSR, and (c) our SARGD. The red solid-lined boxes represent the ground truth (GT), focusing on regions zoomed for detailed inspection. The green boxes illustrate how our SARGD method preserves significantly more detail and clarity compared to the alternatives." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 156, + 258, + 286, + 413 + ], + "blocks": [ + { + "bbox": [ + 156, + 258, + 286, + 413 + ], + "lines": [ + { + "bbox": [ + 156, + 258, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 156, + 258, + 286, + 413 + ], + "type": "image", + "image_path": "4d2a6a77e1420eb6a5740c9e344ed9dc79dc6fab9c6ffe71b3656b2425aef86d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 286, + 258, + 399, + 413 + ], + "blocks": [ + { + "bbox": [ + 286, + 258, + 399, + 413 + ], + "lines": [ + { + "bbox": [ + 286, + 258, + 399, + 413 + ], + "spans": [ + { + "bbox": [ + 286, + 258, + 399, + 413 + ], + "type": "image", + "image_path": "4d37f62d23948c63574a0bf839a6711c9b04d00d6eef864cf2e50a5ebd4fba99.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 258, + 543, + 413 + ], + "blocks": [ + { + "bbox": [ + 399, + 258, + 543, + 413 + ], + "lines": [ + { + "bbox": [ + 399, + 258, + 543, + 413 + ], + "spans": [ + { + "bbox": [ + 399, + 258, + 543, + 413 + ], + "type": "image", + "image_path": "ae5a729cba7aeb2e7e619f6c41390e110e92d127c8726f020b3b1962f972fe15.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 455, + 289, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 289, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 289, + 576 + ], + "type": "text", + "content": "color space, are fidelity-focused measures traditionally favored in super-resolution tasks. Typically, higher values of PSNR and SSIM correspond to better performance. Nevertheless, these metrics do not adequately assess the quality of generative models. Therefore, we also incorporate LPIPS and DISTS in our assessment suite, as they have been acknowledged for their effectiveness in evaluating the perceptual quality of images produced by generative models. Generally, lower scores in LPIPS and DISTS are indicative of superior perceptual image quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Implementation Details. We adhere to the experimental protocol established by previous research [32]. All of our experiments are conducted using PyTorch and tested on a consistent computing platform. The uniform model trained by Zhang et al. [43] is employed as an artifact detector to generate the artifact mask and compute the reality score. For the inference phase, we set the number of denoising steps to 200. Following established protocols [3], we derive low-resolution test samples using the imresize function in Matlab, employing bicubic interpolation with downsampling scales of " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", in line with the precedents" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 455, + 546, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 546, + 492 + ], + "type": "text", + "content": "set by prior research. In diffusion-based super-resolution methods, it is standard practice to upsample the inputs to the same size as the outputs prior to the inference stage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 498, + 489, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 489, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 489, + 512 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 517, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 625 + ], + "type": "text", + "content": "To verify the effectiveness of our training-free SARGD, we conducted comparative analyses with various state-of-the-art methods, such as LDM [26] and StableSR [32]. To ensure a fair comparison, LDM is finetuned using the same training configurations and dataset as those employed for StableSR. For other methods, we directly use the official codes and pre-trained models for testing. Comprehensive results, along with additional comparative analyses, are available in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Quantitative Results. Table 1 demonstrates the outstanding capability of our proposed Self-Adaptive Reality-Guided Diffusion (SARGD) method in producing artifact-free super-resolution images. Notably, our method significantly surpasses the previous StableSR model in terms of PSNR, with improvements of 3.86, 3.97, and 2.91 on the Set5 dataset at upscaling factors of " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ", respec" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25811" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 105, + 547, + 253 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "type": "text", + "content": "Table 2. Analysis of SARGD inference strategies. The 'Baseline' model represents the absence of Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). 'RGR' utilizes initial realistic latent guidance, while 'SAG' employs the self-adaptive guidance mechanism. All experiments are conducted under identical settings. The best scores are highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 105, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 48, + 105, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 105, + 547, + 253 + ], + "type": "table", + "html": "
MetricRGRSAGSet5 [1]Set14 [41]B100 [24]Manga109 [25]
×2×3×4×2×3×4×2×3×4×2×3×4
PSNR ↑XX27.8627.9428.3026.4226.8127.2126.9127.5427.6726.4926.9627.04
✓X29.8230.3930.27 (+ 6.96)28.1627.9927.79 (+ 2.13)28.0828.4228.30 (+ 2.28)25.1026.4127.88 (+ 3.11)
✓✓32.4232.9331.79 (+ 12.3)30.1630.3329.28 (+ 7.61)30.2930.3829.92 (+ 8.13)30.4030.5829.23 (+ 8.10)
SSIM ↑XX0.7650.7770.7690.6730.6880.6960.6510.6720.6790.8140.8210.815
✓X0.8230.8430.824 (+ 7.15)0.7280.7320.730 (+ 4.89)0.7020.7110.704 (+ 3.68)0.7750.8080.822 (+ 0.86)
✓✓0.8970.8950.870 (+ 13.1)0.8210.8150.770 (+ 10.6)0.8070.7920.760 (+ 11.9)0.9000.8930.851 (+ 4.41)
LPIPS ↓XX0.1760.1720.1630.2170.2140.2360.2220.2380.2520.1320.1280.138
✓X0.1440.2120.161 (- 1.23)0.2270.2700.242 (+ 2.54)0.2010.2480.292 (+ 15.8)0.1280.1390.164 (+ 18.9)
✓✓0.1220.1410.131 (- 19.6)0.1440.1520.208 (- 11.9)0.1720.2010.228 (- 9.52)0.0870.0870.132 (- 4.35)
DISTS ↓XX0.1720.1730.1610.1570.1580.1620.1670.1710.1760.1020.0980.103
✓X0.1450.1730.192 (+ 19.3)0.1600.1790.214 (+ 32.1)0.1910.2130.238 (+ 35.2)0.0940.0980.111 (+ 7.77)
✓✓0.1370.1410.145 (- 9.94)0.1270.1290.158 (- 2.47)0.1380.1410.158 (- 10.2)0.0910.0840.099 (- 3.88)
", + "image_path": "71dc42521358eb0aad41260bd470efb1e24a43e901b4a6b4e41d2dee6afa5db5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 262, + 111, + 345 + ], + "blocks": [ + { + "bbox": [ + 53, + 262, + 111, + 345 + ], + "lines": [ + { + "bbox": [ + 53, + 262, + 111, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 111, + 345 + ], + "type": "image", + "image_path": "12d10dc9c40277321165aabb47ba116991c2ab77942ed4584d05d1cb78666cfa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 345, + 96, + 352 + ], + "lines": [ + { + "bbox": [ + 71, + 345, + 96, + 352 + ], + "spans": [ + { + "bbox": [ + 71, + 345, + 96, + 352 + ], + "type": "text", + "content": "(a) LDM" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "type": "text", + "content": "Figure 5. Visual comparison of SARGD components. Our SARGD exhibits the best outcomes for " + }, + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 47, + 354, + 287, + 376 + ], + "type": "text", + "content": " super-resolution." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 112, + 262, + 170, + 345 + ], + "blocks": [ + { + "bbox": [ + 112, + 262, + 170, + 345 + ], + "lines": [ + { + "bbox": [ + 112, + 262, + 170, + 345 + ], + "spans": [ + { + "bbox": [ + 112, + 262, + 170, + 345 + ], + "type": "image", + "image_path": "571b3e1dd456dbb7d610354ded5fc471792cc9207160ec9c3619e894c6d3477a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 346, + 155, + 353 + ], + "lines": [ + { + "bbox": [ + 129, + 346, + 155, + 353 + ], + "spans": [ + { + "bbox": [ + 129, + 346, + 155, + 353 + ], + "type": "text", + "content": "(b) RGR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 171, + 262, + 227, + 345 + ], + "blocks": [ + { + "bbox": [ + 171, + 262, + 227, + 345 + ], + "lines": [ + { + "bbox": [ + 171, + 262, + 227, + 345 + ], + "spans": [ + { + "bbox": [ + 171, + 262, + 227, + 345 + ], + "type": "image", + "image_path": "5c48d19254526f5161ff5d601abee239e53d482a0a4ae603d33d34c7096facc2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 346, + 217, + 353 + ], + "lines": [ + { + "bbox": [ + 184, + 346, + 217, + 353 + ], + "spans": [ + { + "bbox": [ + 184, + 346, + 217, + 353 + ], + "type": "text", + "content": "(c) SARGD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 262, + 283, + 345 + ], + "blocks": [ + { + "bbox": [ + 228, + 262, + 283, + 345 + ], + "lines": [ + { + "bbox": [ + 228, + 262, + 283, + 345 + ], + "spans": [ + { + "bbox": [ + 228, + 262, + 283, + 345 + ], + "type": "image", + "image_path": "03de000f9deda631eae7bb3d11cae20a419d579d9ad6747f1f5781e6f75e3a07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 345, + 268, + 352 + ], + "lines": [ + { + "bbox": [ + 247, + 345, + 268, + 352 + ], + "spans": [ + { + "bbox": [ + 247, + 345, + 268, + 352 + ], + "type": "text", + "content": "(d) GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": "tively. Within the SSIM metric, our SARGD method realizes significant improvements (e.g. the increase from 0.716 to 0.842 on Set14 and from 0.694 to 0.817 on B100), illustrating a substantial enhancement in structural fidelity. Additionally, in terms of the DISTS metric, SARGD significantly reduces the score from 0.102 to 0.086 in Manga109, reflecting an improved alignment with human perceptual quality. These improvements establish our SARGD as a leading approach in the field of super-resolution, setting a new standard for image clarity and detail preservation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 517, + 286, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 286, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 286, + 648 + ], + "type": "text", + "content": "Visualization. Figure 4 offers a visual comparison between our SARGD and StableSR [32], further highlighting the superior performance of our method. Compared to columns (a) and (c), bicubic upscaling maintains satisfactory detail but often leads to an over-smoothing issue. While StableSR succeeds in generating high-resolution images, its dependence on conventional DDPM methods makes it susceptible to introducing additional artifacts. When compared to columns (b) and (c), our approach successfully produces images with high authenticity, thereby avoiding the oversmoothing problem commonly seen in other methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 658, + 141, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 141, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 141, + 671 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "In this section, we thoroughly evaluate and analyze the performance of the proposed training-free SARGD method, primarily focusing on its effectiveness in denoising." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 307, + 318, + 545, + 448 + ], + "blocks": [ + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": "Table 3. Comparison of varied denoising approaches within RGR across Set14, B100, and Urban100 datasets for " + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": " upscaling. " + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{1s}^{*}" + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": " refers to artifact refinement in the latent space at each step, " + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{10s}^{*}" + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": " at every ten steps, while " + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{100s}^{-}" + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{100s}^{+}" + }, + { + "bbox": [ + 305, + 262, + 545, + 316 + ], + "type": "text", + "content": " indicate refinement before and after 100 steps, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 318, + 545, + 448 + ], + "lines": [ + { + "bbox": [ + 307, + 318, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 545, + 448 + ], + "type": "table", + "html": "
ScaleTypeSet14B100Urban100
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
×2D*1s28.160.72828.080.70225.280.708
D*10s27.990.72528.050.69225.110.703
D*100s26.790.68527.320.67224.420.676
D*100s28.420.73928.110.71525.610.716
×3D*1s27.990.73228.420.71126.250.751
D*10s27.830.72628.470.70526.140.747
D-100s26.640.68627.660.68825.780.731
D*100s28.020.73028.540.71826.590.759
×4D*1s27.790.73028.300.70425.920.711
D*10s27.830.72928.480.70225.880.708
D*100s26.940.70827.830.68925.470.702
D*100s28.330.73428.510.71425.960.713
", + "image_path": "763d9675289c249a578baaa067e31384947f35679c9553edcf2bede32113e73d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "text", + "content": "Analysis of Improvement. Table 2 presents the effects of our proposed modules, Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). The baseline corresponds to the LDM configured and trained similarly to StableSR [32]. Analysis of the initial two columns reveals that the integration of RGR notably enhances performance, with gains of " + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "inline_equation", + "content": "6.96\\%" + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "inline_equation", + "content": "2.13\\%" + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "inline_equation", + "content": "2.28\\%" + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "inline_equation", + "content": "3.11\\%" + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "text", + "content": " observed in " + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 458, + 545, + 674 + ], + "type": "text", + "content": " super-resolution on the Set5, Set14, B100, and Manga109 datasets, respectively. The integration of RGR into the DDPM architecture leads to an increase in SSIM scores, validating the effectiveness of RGR in enhancing image quality. The visual results in Figure 5 show that while RGR manages to maintain the essential attributes of the original images, it also leads to over-smoothing in the super-resolved outputs. By contrasting figures (b) and (c), it is evident that incorporating our SAG during inference enhances image quality and successfully addresses the issue of oversmoothing, thus confirming the effectiveness of SAG." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Effect of Varied Denoising Strategies within RGR. Table 3 demonstrates that the artifact refinement in the last 100 steps achieves the best results. Analyzing the first two" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25812" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 78, + 119 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 78, + 119 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 78, + 119 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 78, + 119 + ], + "type": "image", + "image_path": "b855eb6f66437da5822502616b81cb1266562f21d7054af6af2011c3129453af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 119, + 78, + 163 + ], + "blocks": [ + { + "bbox": [ + 51, + 119, + 78, + 163 + ], + "lines": [ + { + "bbox": [ + 51, + 119, + 78, + 163 + ], + "spans": [ + { + "bbox": [ + 51, + 119, + 78, + 163 + ], + "type": "image", + "image_path": "09e65b2905f1b8a5a06441c92223fc58e335199d3eef1d134feb451142bc1311.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 163, + 78, + 172 + ], + "lines": [ + { + "bbox": [ + 51, + 163, + 78, + 172 + ], + "spans": [ + { + "bbox": [ + 51, + 163, + 78, + 172 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 78, + 72, + 120, + 118 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 120, + 118 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 120, + 118 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 120, + 118 + ], + "type": "image", + "image_path": "519ead637caa4c739720a1cd6302b5fe0b3e928d839bf5cce7f590ea434a199c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 121, + 72, + 162, + 118 + ], + "blocks": [ + { + "bbox": [ + 121, + 72, + 162, + 118 + ], + "lines": [ + { + "bbox": [ + 121, + 72, + 162, + 118 + ], + "spans": [ + { + "bbox": [ + 121, + 72, + 162, + 118 + ], + "type": "image", + "image_path": "342a9bc91e141a7166a655e9d975f87fa67f18a82f630e3a200385c70318e7c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 162, + 72, + 201, + 118 + ], + "blocks": [ + { + "bbox": [ + 162, + 72, + 201, + 118 + ], + "lines": [ + { + "bbox": [ + 162, + 72, + 201, + 118 + ], + "spans": [ + { + "bbox": [ + 162, + 72, + 201, + 118 + ], + "type": "image", + "image_path": "1294fda6971f511152a1ae26fa277dfa7e09d924dc1e5af300f0b7206ac832ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 201, + 72, + 244, + 118 + ], + "blocks": [ + { + "bbox": [ + 201, + 72, + 244, + 118 + ], + "lines": [ + { + "bbox": [ + 201, + 72, + 244, + 118 + ], + "spans": [ + { + "bbox": [ + 201, + 72, + 244, + 118 + ], + "type": "image", + "image_path": "2ec6ca6017bff8d74f11667bd5076386b2ec30793fab631b09a7f7d88d253620.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 244, + 72, + 283, + 118 + ], + "blocks": [ + { + "bbox": [ + 244, + 72, + 283, + 118 + ], + "lines": [ + { + "bbox": [ + 244, + 72, + 283, + 118 + ], + "spans": [ + { + "bbox": [ + 244, + 72, + 283, + 118 + ], + "type": "image", + "image_path": "f51e3577e9ce3ab45b88251cd23304fe1299f9a71f38722285c31412bd436559.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 79, + 119, + 120, + 163 + ], + "blocks": [ + { + "bbox": [ + 79, + 119, + 120, + 163 + ], + "lines": [ + { + "bbox": [ + 79, + 119, + 120, + 163 + ], + "spans": [ + { + "bbox": [ + 79, + 119, + 120, + 163 + ], + "type": "image", + "image_path": "241c278ff9eb4f7ea97649eaa94cc2c28eb799fab1adb119022fd5a46be56fe3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 121, + 119, + 161, + 163 + ], + "blocks": [ + { + "bbox": [ + 121, + 119, + 161, + 163 + ], + "lines": [ + { + "bbox": [ + 121, + 119, + 161, + 163 + ], + "spans": [ + { + "bbox": [ + 121, + 119, + 161, + 163 + ], + "type": "image", + "image_path": "c8e17ff669863658261184bb4b1d9ab4483abd3e3445741f5e6c164927faeeda.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "lines": [ + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "text", + "content": "Figure 6. Visual comparison of RGR denoising strategies impacting over-smoothing. They are represented as " + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{1s}^{*}" + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "text", + "content": " for continuous artifact refinement, " + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{10s}^{*}" + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "text", + "content": " for refinement every ten steps, and " + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{100s}^{-}" + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{100s}^{+}" + }, + { + "bbox": [ + 47, + 174, + 286, + 218 + ], + "type": "text", + "content": " for pre- and post-100-step refinement." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 162, + 119, + 201, + 163 + ], + "blocks": [ + { + "bbox": [ + 162, + 119, + 201, + 163 + ], + "lines": [ + { + "bbox": [ + 162, + 119, + 201, + 163 + ], + "spans": [ + { + "bbox": [ + 162, + 119, + 201, + 163 + ], + "type": "image", + "image_path": "99c968cf4b358f1424b8f978e6be429e276965b14e8c9c1935936f0d497b4d55.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 201, + 119, + 244, + 163 + ], + "blocks": [ + { + "bbox": [ + 201, + 119, + 244, + 163 + ], + "lines": [ + { + "bbox": [ + 201, + 119, + 244, + 163 + ], + "spans": [ + { + "bbox": [ + 201, + 119, + 244, + 163 + ], + "type": "image", + "image_path": "7ab2f358852206a613e09678db601c419aaecb842199232e99b9969d86da3a59.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 244, + 119, + 283, + 163 + ], + "blocks": [ + { + "bbox": [ + 244, + 119, + 283, + 163 + ], + "lines": [ + { + "bbox": [ + 244, + 119, + 283, + 163 + ], + "spans": [ + { + "bbox": [ + 244, + 119, + 283, + 163 + ], + "type": "image", + "image_path": "ff6fb3f8f10b16674b63db3179c1e42d72fe73f8f91fb2f7586fae1a1f2db3c6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 48, + 258, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 47, + 223, + 286, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 223, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 286, + 257 + ], + "type": "text", + "content": "Table 4. Comparison of different realistic latent udpate strategies. " + }, + { + "bbox": [ + 47, + 223, + 286, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{1s}^{*}" + }, + { + "bbox": [ + 47, + 223, + 286, + 257 + ], + "type": "text", + "content": " refers to realistic-latent enhancement at each step and " + }, + { + "bbox": [ + 47, + 223, + 286, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{100s}^{+}" + }, + { + "bbox": [ + 47, + 223, + 286, + 257 + ], + "type": "text", + "content": " indicates enhancement after 100 steps, respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 258, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 48, + 258, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 286, + 335 + ], + "type": "table", + "html": "
ScaleTypeSet14B100
PSNR ↑SSIM ↑DISTS ↓PSNR ↑SSIM ↑DISTS ↓
×2R*1s30.160.8210.12730.290.8070.138
R+100s31.400.8420.12630.920.8170.132
×3R*1s30.330.8150.12930.380.7920.141
R+100s30.940.8230.12130.840.8000.140
×4R*1s29.280.7700.15829.920.7600.158
R+100s30.010.7780.15630.230.7630.156
", + "image_path": "37cd79634d83bee20d3a35195f16c12e38745de249bbd04c32951f0826a936c6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 342, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 460 + ], + "type": "text", + "content": "lines per scale shows that refining artifacts at each step improves outcomes over ten-step intervals, as seen with higher PSNRs (28.16 vs. 27.99 on Set4 at " + }, + { + "bbox": [ + 46, + 342, + 286, + 460 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 46, + 342, + 286, + 460 + ], + "type": "text", + "content": "), due to variable artifact patterns in the latent space. Comparing artifact refinement strategies indicates that post-100-step processing yields greater enhancements, as demonstrated by a PSNR increase from 26.94 to 28.33 at " + }, + { + "bbox": [ + 46, + 342, + 286, + 460 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 342, + 286, + 460 + ], + "type": "text", + "content": " super-resolution. Observations from panels (d) and (c) in Figure 6 suggest that artifact refinement early in the noisy latent phase causes more over-smoothing than later-stage refinement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "content": "Approaches to Updating the Realistic Latent in SAG. Table 4 presents a comparative analysis of different strategies for updating the reference realistic-latent within SAG. Notably, updating the realistic latent after 100 steps yields superior results across all metrics when compared to enhancements at every step. For instance, this strategy leads to an increase in PSNR from 30.16 to " + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "inline_equation", + "content": "31.40" + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "content": ", from 30.33 to " + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "inline_equation", + "content": "30.94" + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "content": ", and from 29.28 to " + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "inline_equation", + "content": "30.1" + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 468, + 286, + 588 + ], + "type": "text", + "content": " on Set14. These findings suggest that enhancing the reference latent later in the process correlates with improved performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Influence of Artifact Detection on SARGD. Table 5 contrasts SARGD's performance with and without the implementation of artifact detection in the latent space during inference, providing a line-by-line comparative analysis. It is observed that the elimination of perceptual artifacts in the latent space significantly enhances super-resolution quality, as evidenced by improved metrics such as PSNR (e.g., 30.84 vs. 30.06 on B100 at " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " scaling) and DISTS (e.g., 0.219 vs. 0.140 on B100 at " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " scaling), highlighting the benefits of integrating artifact detection within SARGD." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 307, + 116, + 545, + 203 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": "Table 5. Performance of SARGD with artifact detection (" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": ") versus without in latent space across " + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": " scales. Here, " + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": " signifies the substitution of RGR with a direct sum of the latent and the initial realistic reference latent." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 116, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 203 + ], + "type": "table", + "html": "
ScaleASet14B100
PSNR↑SSIM↑DISTS↓PSNR↑SSIM↑DISTS↓
×2X31.240.8250.13830.560.8010.158
✓31.400.8420.12630.920.8170.132
×3X30.470.7960.19330.060.7830.219
✓30.940.8230.12130.840.8000.140
×4X29.980.7730.23430.150.7530.260
✓30.010.7780.15630.230.7630.156
", + "image_path": "249495fc7fcca65038872fff38c57dd9948cd71f3c2ca71b00ad688e74129bf7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 310, + 205, + 533, + 326 + ], + "blocks": [ + { + "bbox": [ + 310, + 205, + 533, + 326 + ], + "lines": [ + { + "bbox": [ + 310, + 205, + 533, + 326 + ], + "spans": [ + { + "bbox": [ + 310, + 205, + 533, + 326 + ], + "type": "image", + "image_path": "5f2bfa1c2777dc2c04b3a91f0b84c8e0043225007b4b2369ea8826e7fd3b6049.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "type": "text", + "content": "Figure 7. Performance comparison of SARGD using different time steps on Set14 for " + }, + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 305, + 325, + 545, + 369 + ], + "type": "text", + "content": " upscaling using PSNR. Our SARGD reaches peak score between 75 and 100 steps, effectively reducing the total inference steps by half, from 200 to 100." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "type": "text", + "content": "Impact on Inference Steps for SARGD. Figure 7 illustrates that our proposed SARGD not only enhances image quality but also decreases the number of total inference steps required. Specifically, the performance of SARGD in super-resolution steadily improves from the start and achieves optimal PSNR scores between 75 and 100 steps, such as a PSNR of 31.32 at 100 steps compared to 30.16 at 200 steps, which effectively cuts the inference time in half. This suggests that the removal of artifacts during the diffusion phase serves to improve image quality in SR while simultaneously boosting the speed of the inference process." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 523, + 383, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 383, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 383, + 536 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 544, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 664 + ], + "type": "text", + "content": "In this study, we tackle the issue of artifacts that arise in diffusion-based super-resolution. Our SARGD incorporates two key mechanisms: Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). RGR is designed to identify and correct artifacts within the latent space by utilizing a realistic latent as a reference, thereby improving the image's fidelity. To overcome the over-smoothing effect that comes from employing an initial realistic latent for guidance, SAG is introduced to refine the reference latent, improving the quality of the super-resolved images." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (62271400), and the Shaanxi Provincial Key R&D Program, China (2023-GHZD-02)." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25813" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "text", + "content": "[2] Adrian Bulat, Jing Yang, and Georgios Tzimiropoulos. To learn image super-resolution, use a gan to learn how to do image degradation first. In ECCV, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "text", + "content": "[3] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 282 + ], + "type": "text", + "content": "[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. in 2021 IEEE. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 349 + ], + "type": "text", + "content": "[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12413-12422, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 406 + ], + "type": "text", + "content": "[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations (ICLR), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 462 + ], + "type": "text", + "content": "[7] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 509 + ], + "type": "text", + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 38(2):295-307, 2015. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 511, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 511, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 511, + 287, + 554 + ], + "type": "text", + "content": "[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 391-407. Springer, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "text", + "content": "[10] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7890–7899, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "text", + "content": "[11] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in Neural Information Processing Systems (NIPS), 33:3022-3032, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1909-1918, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems (NIPS), 33:6840-6851, 2020. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[14] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 285 + ], + "type": "text", + "content": "[16] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 332, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 386 + ], + "type": "text", + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using wwin transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1833-1844, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "text", + "content": "[19] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4076-4085, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 136-144, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[21] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8060-8069, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 609 + ], + "type": "text", + "content": "[22] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 715-732, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "type": "text", + "content": "[23] Andreas Lugmayr, Martin Danelljan, Fisher Yu, Luc Van Gool, and Radu Timofte. Normalizing flow as a flexible fidelity objective for photo-realistic super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1756-1765, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[24] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25814" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 416-423, 2001. 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 160 + ], + "type": "text", + "content": "[25] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "text", + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "type": "text", + "content": "[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4): 4713-4726, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 403 + ], + "type": "text", + "content": "[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "text", + "content": "[30] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICLR), pages 2256-2265, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "text", + "content": "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "text", + "content": "[32] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "text", + "content": "[33] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv:2211.02831, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "text", + "content": "[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV) workshops, pages 0–0, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[35] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[36] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[37] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 215 + ], + "type": "text", + "content": "[38] Xin Yang, Haiyang Mei, Jiqing Zhang, Ke Xu, Baocai Yin, Qiang Zhang, and Xiaopeng Wei. Drfn: Deep recurrent fusion network for single-image super-resolution with large factors. IEEE Transactions on Multimedia (TMM), 21(2): 328-337, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "text", + "content": "[39] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7556-7566, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[40] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 338 + ], + "type": "text", + "content": "[41] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International Conference on Curves and Surfaces, pages 711-730. Springer, 2012. 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "type": "text", + "content": "[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Jingyu Liu, Jiankang Deng, Wei Liu, and Stefanos Zafeiriou. Edfaceceleb-1 m: Benchmarking face hallucination with a million-scale dataset. TPAMI, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 383, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 447 + ], + "type": "text", + "content": "[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7579-7590, 2023. 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 503 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "text", + "content": "[45] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3096-3105, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 603 + ], + "type": "text", + "content": "[46] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In IEEE international Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 657 + ], + "type": "text", + "content": "[47] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 286-301, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[48] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25815" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 183 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[49] Yulun Zhang, Kai Li, Kunpeng Li, and Yun Fu. Mr image super-resolution with squeeze and excitation reasoning attention network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13425-13434, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[50] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2185–2194, 2021. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25816" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_content_list.json b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..00c3ba5344904381e69d0ee0e6ebc95ebdfe5ae7 --- /dev/null +++ b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_content_list.json @@ -0,0 +1,1885 @@ +[ + { + "type": "text", + "text": "Self-Calibrating Vicinal Risk Minimisation for Model Calibration", + "text_level": 1, + "bbox": [ + 153, + 130, + 815, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiawei Liu, Changkun Ye, Ruikai Cui, Nick Barnes \nThe Australian National University \nCanberra, Australia", + "bbox": [ + 282, + 180, + 684, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jiawei.liu3, changkun.ye, ruikai.cui, nick.barnes}@anu.edu.au", + "bbox": [ + 209, + 236, + 754, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 310, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model calibration, measuring the alignment between the prediction accuracy and model confidence, is an important metric reflecting model trustworthiness. Existing dense binary classification methods, without proper regularisation of model confidence, are prone to being over-confident. To calibrate Deep Neural Networks (DNNs), we propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) that explores the vicinity space of labeled data, where vicinal images that are farther away from labeled images adopt the groundtruth label with decreasing label confidence. We prove that in the logistic regression problem, SCVRM can be seen as a Vicinal Risk Minimisation plus a regularisation term that penalises the over-confident predictions. In practical implementation, SCVRM is approximated using Monte Carlo sampling that samples additional augmented training images and labels from the vicinal distributions. Experimental results demonstrate that SCVRM can significantly enhance model calibration for different dense classification tasks on both in-distribution and out-of-distribution data. Code is available at https://github.com/Carlisle-Liu/SCVRM.", + "bbox": [ + 75, + 316, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 207, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Binary dense classification tasks [10, 54, 64] have advanced significantly since the debut of Deep Neural Networks (DNNs) associated with complex network architectures and large numbers of trainable parameters. Increasing model complexity has been shown to negatively impact model calibration [14] which has remained under investigated in the binary dense classification [29-31]. Miscalibration, a mis-alignment between model confidence and prediction accuracy [38], is an undesirable qualify that hinders the deployment of DNNs, especially in safety critical applications. In this work, we study the model calibration for dense binary classification models.", + "bbox": [ + 75, + 688, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most commonly observed mis-calibration problem is over-confidence, where the model confidence is signifi-", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e81e05f8b4d1f7983d14bd087029ca2986607db870f244ea29fb2da5a640a2ac.jpg", + "image_caption": [ + "Figure 1. An illustration of Mixup [69] and our proposed SCVRM. Circles with solid boundary $\\{(x_i, y_i)\\}_{i=1}^4$ are labeled training images. Vicinal images of Mixup (circles with dashed boundary) are only distributed along the vectors connecting the labeled images. On the other hand, in SCVRM, the label of vicinal image adopts the groundtruth category (shown in different colours) of the closest labeled image, but with label confidence (shown in colour intensity) reduced monotonically with increasing Euclidean distance between vicinal and labeled images. The example sample $\\otimes$ , with Mixup/SCVRM vicinal image and augmented label shown on the right, is selected at the same relative position to $(x_4, y_4)$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 287, + 890, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cantly higher than the accuracy of its predictions on a cohort of samples [14]. Increasing research interest has been dedicated to study model calibration methods that regularise the probability associated with the prediction to be meaningful in reflecting the chance of the prediction being correct. The existing methods can be roughly categorised into three groups: (i) training objective based [2, 8, 13, 21, 23, 38, 45]; (ii) post-processing based [14, 19, 47]; and (iii) label augmentation based [31, 39, 43]. The first two categories focus on penalising over-confident predictions while the third cat", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3335", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "egory directly moderates the confidence of training labels for the DNN models to train on. They do not explore the image space in improving the model calibration.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vicinal Risk Minimisation (VRM) extends Empirical Risk Minimisation (ERM) by introducing a vicinal distribution around each labeled data in the image space [3, 52]. Mixup [69], a variant of VRM, assumes that vicinal images are only distributed along the vectors connecting the labeled data (pair of image and label) pairs. In addition, instead of letting vicinal images adopt the hard groundtruth label of the nearest labeled image, Mixup assigns smoothed versions of nearest groundtruth labels to the vicinal images. In practice, it samples augmented data through the convex combination of labeled data pairs: $(\\tilde{x} = \\lambda x_{i} + (1 - \\lambda)x_{i},\\tilde{y} = \\lambda y_{i} + (1 - \\lambda)y_{j})$ , where the combination factor is sampled from a beta distribution: $\\lambda \\in \\mathrm{Beta}(\\alpha ,\\alpha)$ . Followup works [69, 73] show that Mixup is effective in improving model calibration in the image classification task.", + "bbox": [ + 75, + 138, + 468, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrate the DNNs. Different from Mixup [69], the vicinal distribution of SCVRM can use any distribution that sufficiently covers the vicinity space with reasonable probability density. We consider the case where the vicinal distribution is assumed to be Gaussian with standard deviation being a random variable following a uniform distribution. In addition, the labels assigned to the vicinal images are softened versions of the groundtruth label associated with the labeled image at the distribution centre, where the strength of softening is proportional to the L2 distance between the vicinal and labeled images. As shown in Fig. 1, compared with Mixup [69], our proposed SCVRM has the following advantages: (1) the vicinity image space is not restricted by the pair-wise spatial relations of training images; (2) consistent label confidence across the pixels in the augmented label for dense prediction; and (3) defines confidence boundaries beyond the training distribution to better handle out-of-distribution samples.", + "bbox": [ + 75, + 368, + 468, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a labeled training image set, SCVRM can be understood by the following principals: (1) the human visual system is invariant to small random image changes [12] so that images undergoing up to a certain level of transformation should remain correctly classified [3, 20, 52]. This is also in line with adversarial robustness approaches that require images under small perturbations to maintain their classifications [7]. However, they are no longer labeled training data, so their label confidence should be slightly reduced. (2) it is understood from information theory [16] that we have no information at extreme distance where the label should assume a uniform categorical distribution with maximum entropy. Following the principal of VRM, we propose this transition should be accomplished by a smooth vicinal transition from the exact label for the labeled images to a uniform categorical distribution at large distance from", + "bbox": [ + 75, + 657, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the labeled data. (See Fig. 2 for examples.)", + "bbox": [ + 498, + 90, + 785, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarise our contributions as: (1) propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM), where the labels associated with vicinal images have reduced confidence with increasing Euclidean distance between vicinal and labelled images, to calibrate DNN models for the dense binary classification task; (2) we approximate the vicinal distribution with a Gaussian distribution whose standard deviation is a random variable following a uniform distribution, denoted as SCVRM-G; (3) we supply an example showing that SCVRM is equivalent to VRM and a regularisation term on prediction confidence under a simplified model; (4) We realise SCVRM as a data augmentation technique, where augmented data are sampled from the vicinal distributions with Monte Carlo methods. Finally, we show state-of-the-art calibration results for salient object detection (main paper); and camouflage object and smoke detection, and semantic segmentation (Supp. 12 and Supp. 13).", + "bbox": [ + 496, + 106, + 890, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 376, + 648, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Calibration: aims to minimise the distributional gap between model confidence and prediction accuracy. Existing methods propose to align confidence with accuracy through (i) post-processing techniques, e.g., Temperature Scaling (TS) [14, 19], Platt Scaling [41, 47], Dirichlet Scaling [22], Bayesian Binning [40, 67], Isotropic Regression [68], and Mix-and-Match [70]; and (ii) objective functions including Brier Loss [2, 8], Confidence Penalty [45], Maximum Mean Calibration Error (MMCE) [24], Soft Calibration Objective [21] and Focal Loss [13, 38]. These methods emphasise on suppressing over-confident predictions to alleviate the mis-calibration issue.", + "bbox": [ + 496, + 401, + 890, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research efforts have also been dedicated to investigate the effect of data augmentation methods on the calibration degree. Label augmentation methods align the confidence distribution of the target label, to which the model predictions converge through optimisation, with the prediction accuracy distribution. Label Smoothing (LS) [39, 43] directly softens the target label probability, preventing the model from producing over-confident predictions. [31] presents an alternative label augmentation solution that stochastically perturbs the groundtruth label and aligns the confidence distribution of expected label with the prediction accuracy distribution. On the other hand, Mixup [69], which simultaneously augments both the input image and its corresponding groundtruth label conditioned on the relative position between training data, has also been demonstrated to improve the model calibration degree [51].", + "bbox": [ + 496, + 583, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vicinal Risk Minimisation: is proposed by [3, 52] to explore the vicinity of labeled data in hope of achieving a better approximation of the expected risk. The vicinity space is approximated with a vicinal distribution that can be either estimated with sufficient unlabeled data, or otherwise", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3336", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "assumed to follow certain prior distribution, e.g. Gaussian distribution [3, 52]. Mixup [69] is an extension of VRM that restricts the vicinal data to be distributed along the vectors connecting the data pairs through convex combination. Further, Mixup applies the convex combination in both image and label spaces, resulting in the vicinal samples having a softened version of the groundtruth label associated with the nearest labeled image. VRM has also been applied in randomised smoothing [7, 25, 66] to achieve certified adversarial robustness, where a Gaussian kernel with fixed variance is employed to approximate the vicinal distribution.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Salient Object Detection: is a dense binary classification task. Conventional methods primarily depend on hard-crafted features to identify salient objects [5, 17, 18, 44]. Early DNN based approaches employ learned features of local regions, e.g., super-pixel, object proposal and image patch, which need to be further sequentially processed [15, 26, 53]. The advent of Fully Convolutional Network (FCN) shifted the research focus onto the network architecture design for better pixel-wise predictions. Multi-level feature aggregation takes full advantage of spatial cues imbued in low-level features and semantic information embodied in high-level features [42, 72, 78]. Further, attention mechanisms, including spatial attention, channel attention, etc., have been leveraged to explore intra- and inter-feature map correlations [33, 46, 77], with some works directly adopting transformer backbones with self-attention [50, 71]. A different direction investigates auxiliary cues, e.g. boundary and edge [59, 62], fixation [57].", + "bbox": [ + 75, + 258, + 472, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 547, + 202, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Settings and Notations", + "text_level": 1, + "bbox": [ + 76, + 571, + 285, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This papers deals primarily with dense binary classification over the image space $\\mathbb{R}^d$ with the corresponding label set of $\\mathcal{V}\\in \\{0,1\\} ^d$ , where $d$ is the dimension of image/label space, \"0\" and \"1\" represent the two categories. $\\mathcal{V}_s\\in [0,1]^d$ denotes a soft label space, clearly we have $\\mathcal{V}\\subset \\mathcal{V}_s$ . Let $\\mathcal{D} = \\{(x_i,y_i)\\}_{i = 1}^N$ be a finite dataset with image and label pairs $(x_{i},y_{i})$ sampled i.i.d. from the joint distribution $p(x,y)$ defined on $\\mathcal{X}\\times \\mathcal{Y}$ . The task is to obtain an optimal classifier $f^{*}\\in \\mathcal{H}$ in the Hypothesis space $\\mathcal{H}\\subset \\{f:\\mathcal{X}\\to \\mathcal{Y}\\}$ that maximises the prediction accuracy and model calibration performance. The dimensionality or spatial index is omitted for simplicity wherever it is clear.", + "bbox": [ + 75, + 597, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model calibration degree over the joint distribution $p(x,y)$ can be evaluated in terms of Expected Calibration Error (ECE) [14, 23]. Let $S(x) \\in (0,1)$ be the Sigmoid-activation value before classification: $f(x) = \\mathbb{1}(S(x) > 0.5)$ , where $\\mathbb{1}(\\cdot)$ is an indicator function. The prediction confidence $c$ and accuracy $a$ can be defined as: $c = |S(x) - 0.5| + 0.5$ , and $a = \\mathbb{1}(f(x) = y)$ respectively. We further use $p_{f,\\mathcal{D}}(c,a)$ to denote the joint dis", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ribution of prediction confidence and prediction accuracy of model, $f(\\cdot)$ , on dataset, $\\mathcal{D}$ . Then, ECE can be defined as: $\\mathrm{ECE}(p_{f,\\mathcal{D}}) = \\mathbb{E}_{p_f,\\mathcal{D}(c)}\\bigl [|\\mathbb{E}_{p_f,\\mathcal{D}(a|c)}[a] - c|\\bigr ]$ , where $\\mathbb{P}_{f,\\mathcal{D}}(c)$ is a marginal distribution on prediction confidence, and $\\mathbb{E}_{\\mathbb{P}_{\\theta ,\\mathcal{D}}(a|c)}$ is a conditional distribution of prediction accuracy. For a well calibrated model, we should have: $p_{f,\\mathcal{D}}(f(x)_i = y_i|c = r) = r$ $\\forall r\\in [0,1]$", + "bbox": [ + 496, + 90, + 890, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Vicinal Risk Minimisation", + "text_level": 1, + "bbox": [ + 500, + 205, + 736, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We may formulate a learning problem as searching for model $f \\in \\mathcal{H}$ that minimises expected risk $R(f)$ over loss function $\\ell(f(x), y)$ , which can be written as:", + "bbox": [ + 496, + 229, + 890, + 275 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR (f) = \\int_ {\\mathcal {X} \\times \\mathcal {Y}} \\ell (f (x), y) \\cdot p (x, y) d x d y. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 285, + 890, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a dataset $\\mathcal{D} = \\{(x_i, y_i)\\}_i^N$ , the Empirical Risk Minimisation (ERM) approach approximates $R(f)$ by:", + "bbox": [ + 496, + 329, + 890, + 361 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R _ {E} (f) = \\int_ {\\mathcal {X} \\times \\mathcal {Y}} \\ell (f (x), y) \\cdot \\frac {1}{N} \\delta_ {x _ {i}, y _ {i}} (x, y) d x d y \\tag {2} \\\\ = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell \\left(f \\left(x _ {i}\\right), y _ {i}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 371, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta_{x_i,y_i}(x,y)$ is the Dirac delta distribution over $(x,y)$ that can only adopts the value $(x_i,y_i)$ .", + "bbox": [ + 496, + 458, + 890, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on ERM, the Vicinal Risk Minimisation (VRM) approach [3, 52] approximates $R(f)$ by", + "bbox": [ + 496, + 489, + 890, + 520 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {V} (f) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(\\tilde {x} _ {i}\\right), y _ {i}\\right) p \\left(\\tilde {x} _ {i} \\mid x _ {i}\\right) d \\tilde {x}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 531, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(\\tilde{x}_i|x_i)$ denotes the Probability Density Function (PDF) of vicinal images $\\tilde{x}_i$ given labeled image $x_i$ .", + "bbox": [ + 496, + 583, + 890, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, VRM proposes to use isotropic Gaussian distributions as a choice of $p(\\tilde{x}_i|x_i)$ . Specifically, VRM defines $\\tilde{x}_i = x_i + \\epsilon, \\epsilon \\sim \\mathcal{N}_d(0,\\sigma^2 I_d)$ , where $I_{d}$ is a $d \\times d$ identity matrix and the variance $\\sigma^2$ is a fixed hyperparameter. The corresponding $R_V(f)$ can then be written as:", + "bbox": [ + 496, + 614, + 890, + 689 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {V - G} (f; \\sigma) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), y _ {i}\\right) p (\\epsilon) d \\epsilon , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 700, + 890, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(\\epsilon) = \\frac{1}{\\sqrt{(2\\pi\\sigma^2)^d}}\\exp \\left(-\\frac{\\|\\epsilon\\|_2^2}{2\\sigma^2}\\right)$ is the PDF of the additive Gaussian noise $\\epsilon$", + "bbox": [ + 496, + 753, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 806, + 589, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Self-Calibrating Vicinal Risk Minimisation", + "text_level": 1, + "bbox": [ + 500, + 830, + 864, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is based on the principals that samples in the proximity of labeled data should inherit their groundtruth labels with reduced label confidence, while samples in the", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3337", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af90f54dffbdbda5bdf230f2f940cbbcf20d33c3159b9bfa02dc27eb4e3b1865.jpg", + "image_caption": [ + "Figure 2. Examples of vicinal images and assigned labels under various standard deviation $\\sigma$ values. $\\sigma = 0$ denotes the labeled data $(x,y)$ . Vicinal images are generated via adding Gaussian noise $\\epsilon \\sim \\mathcal{N}(0,\\sigma^2 I_d)$ to the labeled image as defined in Eq. 7. Their assigned labels, being the softer versions of the groundtruth label $y$ , are computed with the Eq. 8 and Eq. 9, whose plot are shown on the right." + ], + "image_footnote": [], + "bbox": [ + 81, + 102, + 893, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "extreme distance should have a near uniform categorical distribution for their labels. As an extension of VRM, we reduce label confidence for vicinal images based on the L2-distance between vicinal and labeled images. Extending the definition of VRM [3, 52], we propose a novel Self-Calibrating Vicinal Risk Minimisation (SCVRM):", + "bbox": [ + 75, + 272, + 472, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1. (SCVR) Let $\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^N$ . For $f \\in \\mathcal{H}$ and $\\ell : f(\\mathcal{X}) \\times \\mathcal{Y}_s \\to \\mathbb{R}$ , Self-Calibrating Vicinal Risk is defined as:", + "bbox": [ + 75, + 371, + 470, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {S C} (f) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(\\tilde {x} _ {i}\\right), g \\left(y _ {i}, \\| \\tilde {x} _ {i} - x _ {i} \\| _ {2}\\right)\\right) \\times \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 428, + 468, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\tilde {x} _ {i} | x _ {i}) d \\tilde {x} _ {i},\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 472, + 320, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g:\\mathcal{Y}\\times \\mathbb{R}_{+}\\to \\mathcal{Y}_{s}$", + "bbox": [ + 76, + 497, + 250, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The difference between VRM (Eq. 3) and SCVRM lies in the labels associated with the vicinal images. VRM assigns the exact groundtruth label $y_{i}$ paired with labeled image $x_{i}$ to all its vicinal images $\\tilde{x}_i$ , whereas we use softened versions of groundtruth label $g(y_{i},\\| \\tilde{x}_{i} - x_{i}\\|_{2})$ , whose label confidence reduces with increases in the L2-distance between the vicinal and labeled images. Following VRM to represent the distribution of vicinal images conditioned on the labeled image $p(\\tilde{x}|x)$ with an isotropic Gaussian distribution with a fixed variance, our SCVRM can be written as:", + "bbox": [ + 75, + 521, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {S C - V} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), g \\left(y _ {i}; \\| \\epsilon \\| _ {2}\\right)\\right) p (\\epsilon) d \\epsilon , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 695, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where L2-norm of Gaussian noise equals the L2-distance between vicinal and labeled images $\\| \\epsilon \\| _2 = \\| \\tilde{x}_i - x_i\\| _2$", + "bbox": [ + 75, + 750, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that the isotropic Gaussian noise $\\epsilon \\sim \\mathcal{N}_d(0,\\sigma^2 I_d)$ used in the VRM may face the problem of a bubbling effect, that results in the probability density of an isotropic Gaussian $\\mathcal{N}_d(\\mu ,\\sigma^2 I_d)$ concentrating on a thin spherical shell centred on $\\mu \\in \\mathbb{R}^d$ with radius $\\sigma \\sqrt{d}$ in high dimensional space $d\\gg 1$ [1, 36]. Following this observation: (a) we propose to let the scale of the standard deviation of the isotropic Gaussian distribution be a random", + "bbox": [ + 75, + 780, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "variable following a uniform distribution $\\sigma \\sim \\mathcal{U}(0,\\gamma]$ to prevent vicinal images $\\tilde{x}$ from being distributed only near a hypersphere centring on $x_{i}$ ; and (b) we approximate the L2-distance between the vicinal and labeled images with a dummy variable $\\| \\tilde{\\epsilon}\\| _2\\approx \\| \\epsilon \\| _2$ which is set to $\\| \\tilde{\\epsilon}\\| _2 = \\sigma \\sqrt{d}$ (See Supp. 8.1 for experimental justifications). Our implementation of SCVRM can then be defined as:", + "bbox": [ + 496, + 272, + 893, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {S C - G} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), g \\left(y _ {i}; \\| \\tilde {\\epsilon} \\| _ {2}\\right)\\right) \\times \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 387, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\epsilon) p (\\sigma) d \\epsilon d \\sigma ,\n$$\n", + "text_format": "latex", + "bbox": [ + 671, + 431, + 772, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p(\\sigma) = 1 / \\gamma$ is the PDF of the standard deviation $\\sigma$ .", + "bbox": [ + 500, + 458, + 883, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are many choices for the projection function from hard label space to soft label space. The most straightforward solution is to apply the label smoothing [39] function: $\\mathrm{LS}(y,\\beta) = (1 - \\beta) + \\frac{\\beta}{K},\\beta \\in [0,1]$ where $K = 2$ denotes the number of classes in a binary task. Then our projection function can be defined as:", + "bbox": [ + 496, + 474, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\mathrm {L S}} (y; \\varphi (\\| \\tilde {\\epsilon} \\| _ {2})) = \\left(1 - \\varphi (\\| \\tilde {\\epsilon} \\| _ {2})\\right) y + \\frac {\\varphi (\\| \\tilde {\\epsilon} \\| _ {2})}{2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 575, + 890, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we define $\\varphi (\\cdot):\\mathbb{R}_0^+\\to [0,1)$ to be a Gaussian function $\\varphi_G(\\cdot ;\\cdot)$ , that is formulated as:", + "bbox": [ + 496, + 616, + 890, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {G} \\left(\\| \\tilde {\\epsilon} \\| _ {2}; \\eta\\right) = 1 - \\exp \\left(- \\frac {\\| \\tilde {\\epsilon} \\| _ {2}}{\\eta}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 657, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is a hyperparameter that scales the value of $\\| \\tilde{\\epsilon}\\| _2$ , and we empirically set $\\eta = \\sqrt{d}$ . Fig. 2 illustrates the resultant groundtruth map assigned to vicinal images sampled under various standard deviations. Please note that the function $\\varphi$ can be an arbitrary function that satisfies: $\\varphi :\\mathbb{R}_0^+ \\to [0,1)$ . We investigate other options in Supp. 8.2.", + "bbox": [ + 496, + 700, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Connection between SCVRM and VRM", + "text_level": 1, + "bbox": [ + 498, + 801, + 843, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We analyse the relationship between VRM and our proposed SCVRM in a simple setting, where we find mathematical connections between the two approaches. Specifically, we consider the case of binary classification tasks, where our $R_{SC - G}(f)$ in Eq. 7 adopts a single-layer logistic", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3338", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "regression model and the binary cross entropy loss. In this setting, we prove that if $g(y_{i},\\| \\tilde{\\epsilon}\\|_{2})$ with $\\| \\tilde{\\epsilon}\\| _2 = \\sigma \\sqrt{d}$ is a label smoothing function, $R_{SC - G}(f)$ can be written as a combination of the Vicinal Risk $R_{V - G}(f;\\sigma)$ in Eq. 4 and a regularization term that penalises overconfident predictions of the model $f$ (See proof in Supp. 7.1).", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Example 1. In $R_{SC - G}(f)$ defined in Eq. 7, suppose:", + "bbox": [ + 76, + 199, + 434, + 214 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Data: $\\{(x_i, y_i)\\}_{i=1}^N$ with $x_i \\in \\mathbb{R}^d$ and $y_i \\in \\{0, 1\\}$ ,", + "- Model: $f(x) = 1 / (1 + e^{-w^T x})$ with $w \\in \\mathbb{R}^d$ ,", + "- Loss: $\\ell(f(x), y) = -y \\log f(x) - (1 - y) \\log (1 - f(x))$ ,", + "- $g(y, \\| \\tilde{\\epsilon} \\|_2) = (1 - 2\\sigma)y + \\sigma, \\sigma \\sim \\mathcal{U}(0, \\gamma]$ and $\\gamma \\in (0, \\frac{1}{2}]$ . then we have:" + ], + "bbox": [ + 76, + 215, + 467, + 295 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {S C - G} (f) = \\int_ {0} ^ {\\gamma} \\frac {1}{\\gamma} R _ {V - G} (f; \\sigma) d \\sigma + \\tau (f), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 309, + 468, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 357, + 122, + 369 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau (f) = \\frac {\\gamma}{2 N} \\sum_ {i = 1} ^ {N} \\left(2 y _ {i} - 1\\right) \\cdot w ^ {T} x _ {i}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 375, + 468, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the first term of the RHS in Eq. 10 is equivalent to introducing our design of $\\sigma \\sim \\mathcal{U}(0,\\gamma]$ into $R_{V - G}(f;\\sigma)$ which is the VRM with Gaussian kernel defined in Eq. 4.", + "bbox": [ + 76, + 429, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the label satisfies $y_{i} \\in \\{0,1\\}$ , therefore in the term $\\tau (f)$ in Eq. 11 we have:", + "bbox": [ + 76, + 491, + 468, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(2 y _ {i} - 1\\right) \\cdot w ^ {T} x _ {i} = \\left\\{ \\begin{array}{l l} w ^ {T} x _ {i}, & y _ {i} = 1, \\\\ - w ^ {T} x _ {i}, & y _ {i} = 0. \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 537, + 468, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this sense, the term $\\tau(f)$ penalises high $w^T x_i$ when $y_i = 1$ and penalises low $w^T x_i$ when $y_i = 0$ . Since high $w^T x$ will lead to $f(x) = 1 / (1 + e^{w^T x})$ approaching 1 and low $w^T x$ will lead to $f(x)$ approaching 0, $\\tau(f)$ actually penalizes overconfident predictions of the model $f$ .", + "bbox": [ + 76, + 597, + 468, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. In this simplified case, minimising the Self-Calibrating Vicinal Risk defined in Eq. 7 equivalently minimises the Vicinal Risk defined in Eq. 4 incorporating our design of $\\sigma \\in \\mathcal{U}(0,\\gamma]$ and a regularisation term defined in Eq. 11, where the former improves the model classification accuracy the latter prevents the model from becoming overconfident on vicinal samples.", + "bbox": [ + 76, + 691, + 468, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Practical Model", + "text_level": 1, + "bbox": [ + 76, + 814, + 236, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practical implementation, we approximate the intractable SCVRM defined in Eq. 7 with the Monte-Carlo (MC) Sampling. More specifically, it is formulated as a data augmentation technique where, based on the labeled dataset", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{D} = \\{(x_{1},y_{i})\\}_{i = 1}^{N}$ , we sample an augmented dataset:", + "bbox": [ + 500, + 90, + 854, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} _ {s c} = \\bigcup_ {i = 1} ^ {N} \\left\\{(\\tilde {x} _ {i} ^ {j}, \\tilde {y} _ {i} ^ {j}) \\mid \\tilde {x} _ {i} ^ {j} = x _ {i} + \\epsilon_ {i} ^ {j}, \\epsilon_ {i} ^ {j} \\sim_ {i. i. d} \\mathcal {N} (0, \\sigma_ {i, j} ^ {2} I _ {d}), \\right. \\\\ \\sigma_ {i, j} \\sim_ {i. i. d} \\mathcal {U} (0, \\gamma ], \\\\ \\left. \\tilde {y} _ {i} ^ {j} = g _ {\\mathrm {L S}} \\left(y _ {i}; \\varphi_ {G} (\\| \\tilde {\\epsilon} \\| _ {2}; \\eta)\\right) \\right\\} _ {j = 1} ^ {M}, \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 114, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{x}_i^j$ is the $j^{th}$ augmented image sampled from the vicinity of the $i^{th}$ labeled image, $\\tilde{y}_i^j$ is a smoothed version of the groundtruth label $y_i$ , $M$ is a hyperparameter being the number of augmented data from the vicinity of each labeled data, and we set $\\| \\tilde{\\epsilon} \\|_2 = \\sigma \\sqrt{d}$ to approximate the L2-distance between the vicinal and labeled images $\\sigma \\sqrt{d} \\approx \\| \\epsilon \\|_2$ which relieves the computational burden of computing an L2-norm in high-dimensional space. The model is trained on both the labeled dataset $\\mathcal{D}$ and the augmented dataset $\\mathcal{D}_{sc}$ using a Binary Cross Entropy loss: $\\mathcal{L}_{bce}(f(x), y) = -y \\log(f(x)) - (1 - y) \\log(1 - f(x))$ . The training loss can be defined as:", + "bbox": [ + 496, + 220, + 890, + 405 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {b c e} \\left(f \\left(x _ {i}\\right), y _ {i}\\right) + \\sum_ {j = 1} ^ {M} \\mathcal {L} _ {b c e} \\left(f \\left(\\tilde {x} _ {i} ^ {j}\\right), \\tilde {y} _ {i} ^ {j}\\right)\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 412, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the augmented dataset defined in Eq. 13 is resampled after each training epoch.", + "bbox": [ + 498, + 469, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 500, + 511, + 733, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main paper presents results on Salient Object Detection. Results for other dense classification tasks, e.g., Camouflaged Object Detection (binary) and Smoke Detection (binary), and semantic segmentation (multi-class) can be found in Supp. 12 and 13 respectively.", + "bbox": [ + 496, + 537, + 890, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 619, + 718, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics: We employ bin-based Expected Calibration Error (denoted $\\mathrm{ECE}_{\\mathrm{EW}}$ [14]) and Over-confidence Error $(\\mathrm{OE}_{\\mathrm{EW}}$ [51]) with $B = 10$ bins, to evaluate model calibration. (See Supp. 9.1 for implementation details.)", + "bbox": [ + 496, + 643, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: Following [31] we divide DUTS-TR [54] into a train $(\\mathcal{D}_{\\mathrm{TR}})$ and validation set $(\\mathcal{D}_{\\mathrm{VAL}})$ , 9,553 and 1,000 images respectively. We evaluate model calibration degree using SOD test datasets, DUTS-TE [54], DUT-OMRON [65], PASCAL-S [28], SOD [37], ECSSD [63], HKU-IS [26].", + "bbox": [ + 496, + 704, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Architecture: The U-Net based model consists of a ResNet50 encoder and a decoder that are initialised with ImageNet pretrained weights and by default respectively. The implementation uses the Pytorch framework. Experiments with different encoders, e.g. VGG16 [49] and Swin transformer [34] are detailed in Supp. 11.", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperparameters: Optimal calibration performance of model is achieved by setting $\\gamma = 2$ , $\\eta = \\sqrt{d}$ and $M = 3$ .", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/025b00b06ec701f7939af55cba32fd927eb29f88209a76da9f6285dc3b1479fd.jpg", + "table_caption": [ + "Table 1. Salient object detection model calibration benchmark. Results are evaluated with ${\\mathrm{{ECE}}}_{\\mathrm{{EW}}}$ and ${\\mathrm{{OE}}}_{\\mathrm{{EW}}}$ with 10 bins (units in %)." + ], + "table_footnote": [], + "table_body": "
MethodsDUTS-TE [54]DUT-OMRON [65]PASCAL-S [28]SOD [37]ECSSD [63]HKU-IS [26]
CategoryNameYearECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓
SOD MethodsMSRNet [27]20172.572.343.323.163.443.236.426.140.970.940.920.87
SRM [55]20174.023.724.193.964.884.599.939.582.532.351.861.72
Amulet [74]20175.675.285.845.495.765.4310.039.592.562.421.981.87
BMPM [72]20183.743.524.524.374.884.688.167.931.951.891.581.53
DGRL [56]20184.123.864.414.215.014.778.448.202.132.021.631.53
PAGR [75]20184.043.795.144.965.645.3712.1711.872.842.701.621.54
PiCANet [33]20185.124.904.844.708.147.9210.5010.303.483.392.552.47
CPD [61]20193.973.784.204.065.375.179.659.392.292.191.991.90
BASNet [48]20195.004.864.934.836.506.3610.4010.272.742.702.302.26
EGNet [76]20193.333.143.663.505.425.198.047.791.981.881.471.40
AFNet [11]20193.953.744.254.095.064.848.158.022.382.271.871.78
PoolNet [32]20193.333.123.863.705.325.078.147.872.001.901.821.75
GCPANet [4]20203.182.993.993.844.163.977.056.881.611.541.271.21
MINet [42]20203.653.484.454.294.944.758.017.892.132.031.741.65
F3Met [58]20203.673.504.254.104.854.677.957.782.262.161.921.83
EBMGSD [71]20213.453.294.113.954.794.617.487.302.142.051.791.70
ICON [79]20212.892.763.843.714.083.956.706.551.561.491.381.32
PFSNet [35]20212.942.723.953.814.454.277.597.392.412.252.061.96
EDN [60]20223.623.474.023.904.894.748.818.662.202.131.651.58
Model Calibration MethodsBrier Loss [2]19502.772.583.553.383.903.706.406.161.371.301.040.99
Temperature Scaling [14]20172.532.343.183.033.563.366.326.050.960.930.830.70
MMCE [23]20182.862.673.563.414.003.816.856.631.411.351.181.13
Label Smoothing [39]20192.001.792.892.713.042.835.975.690.830.680.820.47
Mixup [51]20192.452.253.413.233.132.995.825.701.411.183.830.05
Focal Loss [38]20202.252.083.102.823.403.136.215.981.411.031.240.77
AdaFocal [13]20221.611.412.311.842.532.275.885.471.630.791.350.52
ASLP [31]20231.401.221.991.832.312.105.505.170.480.200.790.17
OursSCVRM20230.780.611.641.491.911.753.903.600.440.190.780.10
", + "bbox": [ + 99, + 104, + 867, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Optimisation Details:** The model is optimised for 30 epochs with an initial learning rate of $2.5 \\times 10^{-5}$ that decays by a factor of 0.9 per epoch after the $10^{\\text{th}}$ epoch. During each epoch, $M = 3$ augmented data are sampled from each vicinity space. Image size is set to $3 \\times 384 \\times 384$ and the batch size is 8. Basic data augmentation techniques including random flipping, translation and cropping are applied.", + "bbox": [ + 75, + 488, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Model Calibration Degree Performance", + "text_level": 1, + "bbox": [ + 76, + 604, + 416, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The calibration degrees of existing SOD and model calibration methods, and our proposed SCVRM, evaluated in terms of $\\mathrm{ECE}_{\\mathrm{EW}}$ and $\\mathrm{OE}_{\\mathrm{EW}}$ , are presented in Tab. 1. It can be observed that SOD methods without regularisations on prediction confidence are, in general, less well calibrated than the models equipped with calibration methods. On the other hand, our proposed SCVRM achieves the lowest calibration errors across all six testing datasets among the model calibration methods. Largest improvements over the second-best model are obtained on DUTS-TE, DUTOMRON, PASCAL-S and SOD datasets, reducing the ECE by $44.3\\%$ , $17.6\\%$ , $17.3\\%$ and $29.1\\%$ respectively. There is little room for improvement on ECSSD and HKU-IS, where ASLP is already well calibrated. On these datasets, our proposed SCVRM maintains the high calibration degrees that are comparable with those of ASLP.", + "bbox": [ + 75, + 628, + 468, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 depicts the joint distribution of prediction confidence and prediction accuracy of some model calibration", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods on the DUTS-TE [54] dataset. The joint distribution of better calibrated models become more closely aligned with the oracle line, which indicates a perfectly calibrated model. Our SCVRM has its joint distribution almost completely aligned with the oracle line. In comparison, Mixup is more over-confident with the high-density area of its joint distribution slightly to the bottom-right of the oracle and the low-density area deviating even more. This further demonstrates the effectiveness of exploring a continuous vicinal image space with softened labels.", + "bbox": [ + 496, + 488, + 890, + 640 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b3b4e0efa190b136c58b47eb3a945ac29ceabcacf009275d29315c89fcb30283.jpg", + "table_caption": [ + "Table 2. Existing model calibration methods and our proposed LSR evaluated on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6] in terms of $\\mathrm{ECE_{EW}}$ and $\\mathrm{OE_{EW}}$ with 10 bins, and Accuracy (ACC)." + ], + "table_footnote": [], + "table_body": "
MethodEvaluation (%)
ECE ↓OE ↓ACC ↑
Baseline52.3651.0541.88
Brier Loss [2]38.8537.1853.62
Temperature Scaling [14]51.9550.4641.59
Label Smoothing [39]37.2235.4855.41
MMCE [23]40.6439.6754.39
Mixup [51]31.0729.1058.71
Focal Loss [38]40.0138.4349.71
AdaFocal [13]27.5525.0755.39
ASLP [31]18.3116.3761.93
SCVRM11.938.2683.93
", + "bbox": [ + 532, + 713, + 856, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3340", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75d40a443d9230b863265fce2aec65a702d3df8520ceead1bef048033897eafa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 85, + 207, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34690ba4109329e8870796193002f67926e9ea45dc6da557728ef5259b2836b8.jpg", + "image_caption": [ + "Label Smoothing" + ], + "image_footnote": [], + "bbox": [ + 80, + 186, + 204, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0063631d6ffe975c7f838625067b44c181bf26292f35b3814b432614bdabfe7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 334, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0af0452b9afa5f5c7463836c67f9b093e74f987e9c626f9bc648dff63821d3b.jpg", + "image_caption": [ + "Mixup" + ], + "image_footnote": [], + "bbox": [ + 210, + 186, + 334, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d96243020edbfbf5c14bf16471cdce86128e1b11d0fbc4f79bbf6bae7252b463.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 88, + 467, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec2c4acb75502d4332ecade00a976d02259070b961a444ebd9cb628323534911.jpg", + "image_caption": [ + "Focal Loss" + ], + "image_footnote": [], + "bbox": [ + 339, + 186, + 465, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34b16694490a161de749c850e2d4875770bf05a854b8b92e1c41fb225c247b31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 468, + 88, + 596, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19f7c10abcebe2260b69d4a1955c76be862c4ae2c7c51f880586011a5e687f54.jpg", + "image_caption": [ + "AdaFocal" + ], + "image_footnote": [], + "bbox": [ + 468, + 186, + 594, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b21974af960421bf50b412a41f2f84c39dfdeb5d23aec1f741dfbbdc94464422.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 88, + 725, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/713e02010444d53cf5f8bdbe53cb04a4b0eae8aff31f68a5b7c38e1f749c1748.jpg", + "image_caption": [ + "ASLP" + ], + "image_footnote": [], + "bbox": [ + 599, + 186, + 723, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dca3a226ecc5e10324e1ee1afc8d5436bdede99b11e931d76293f54acf39e44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 88, + 856, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/202c649884647845e5028e1493cf92f79d8d66cadf023451391a9506f153c0f0.jpg", + "image_caption": [ + "SCVRM" + ], + "image_footnote": [], + "bbox": [ + 730, + 186, + 856, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3dd3d9bb05ed6b3591b33ff7ac8903d61f00f10812546d526d7cac7abce114b.jpg", + "image_caption": [ + "Figure 3. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on DUTS-TE [54]. The dashed red diagonal line represents the perfectly calibrated oracle model.", + "Figure 4. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6]. The dashed red diagonal line represents the perfectly calibrated oracle model." + ], + "image_footnote": [], + "bbox": [ + 81, + 340, + 464, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Model Calibration on Out-of-Distribution Data", + "text_level": 1, + "bbox": [ + 76, + 738, + 468, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The calibrating ability of existing model calibration techniques and our proposed SCVRM on Out-of-Distribution (OoD) samples are evaluated on the 500 texture images [31] selected from the Describable Texture Dataset [6]. These texture images, not including any salient (foreground) object, demonstrates certain level of distributional shift from both training and testing distributions of SOD. As shown in Tab. 2, SCVRM achieves the lowest calibration errors on OoD data. Compared with the second-best model, it", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reduces ECE by $34.8\\%$ and OE by $49.5\\%$ . The improvements are also reflected in the joint distribution plot in Fig. 4, where the distribution area of SCVRM is significantly better aligned with the oracle line. Further, SCVRM significantly outperforms Mixup in terms of both calibration and classification. Its superior performance can be attributed to more effective utilisation of vicinity space, and a distance-based label augmentation technique that yields consistent label confidence across the pixels.", + "bbox": [ + 496, + 343, + 890, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 494, + 653, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate the effect of SCVRM and its hyperparameters including $\\gamma$ , $\\eta$ and $M$ which adopt the default setting specified in Sec. 5.1 unless stated otherwise.", + "bbox": [ + 496, + 518, + 890, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of SCVRM: We compare our SCVRM with ERM, the baseline model, VRM with a fixed standard deviation, and $\\mathrm{VRM}^*$ with our design of $\\sigma \\sim \\mathcal{U}(0,\\gamma]$ (see implementation detail in Supp. 8.3). Tab. 3 shows that VRM (see implementation detail in Supp. 8.3) does not reduce calibration error over ERM in general, producing mixed impacts across the testing datasets. Introducing our design of $\\sigma \\sim \\mathcal{U}(0,\\gamma]$ also does not alleviate VRM's over-confidence. On the other hand, SCVRM significantly improves the model calibration over ERM, VRM and $\\mathrm{VRM}^*$ . In addition, SCVRM also achieves improved classification accuracy than ERM (see Supp. 10.1), which can be attributed to more effective utilisation of vicinity space with our design of $\\sigma \\sim \\mathcal{U}(0,\\gamma]$ . It also enables $\\mathrm{VRM}^*$ to achieve better classification accuracy compared to VRM (see Supp. 10.1).", + "bbox": [ + 496, + 566, + 892, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of $\\gamma$ : The hyperparameter $\\gamma$ suggests the exploration radius around the training images. As illustrated in Fig. 5a, SCVRM can reach optimal calibration performance in a wide range $\\gamma = [1, 10]$ . When $\\gamma$ is too small, e.g. 0.5, SCVRM achieves limited improvements on model calibration as the vicinal images are too close to the labeled images, and their corresponding augmented labels retain rel", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3341", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3957c7aff43246ea045e21e408b22d70f3022a4f61e9d3264ea043f46a1ec5ba.jpg", + "table_caption": [ + "Table 3. Ablation study on SCVRM." + ], + "table_footnote": [ + "$^{1}$ VRM* incorporates our design of $\\sigma \\sim \\mathcal{U}(0, \\gamma]$ ." + ], + "table_body": "
MethodECE ↓ DUTS-TE [54]ECE ↓ DUT-OMRON [65]ECE ↓ PASCAL-S [28]
ERM (Baseline)3.052.893.803.684.143.98
VRM3.543.374.354.214.394.24
VRM*13.183.033.903.794.133.98
SCVRM0.780.611.641.491.911.75
MethodSOD [37]ECSSD [63]HKU-IS [26]
ERM (Baseline)7.076.851.821.761.381.34
VRM6.606.421.671.621.301.26
VRM*17.357.161.721.681.281.23
SCVRM3.903.600.440.190.780.10
", + "bbox": [ + 84, + 102, + 455, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "atively high label confidences (see Fig. 2). As $\\gamma$ grows, the augmented vicinal images can get sufficiently far away from the labeled images to set up label confidence contours of various levels and obtains optimal model calibration.", + "bbox": [ + 75, + 271, + 468, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of $\\eta$ : The hyperparameter $\\eta$ is part of the Gaussian equation (Eq. 9) that affects the strength of smoothing factor. (See Supp. 8.4 for example vicinal data under different $\\eta$ .) At $\\eta = 0.1\\sqrt{d}$ , the resultant augmented label is overly softened, leading to the trained model becoming under-confirmed1. When $\\eta$ is too large, e.g. $5\\sqrt{d}$ and $10.0\\sqrt{d}$ , vicinal images all retain a relatively high label confidence, resulting in insufficient calibration regularisation. We find that $\\eta$ works well in $[0.5\\sqrt{d}, 2\\sqrt{d}]$ , leading consistently to near-optimal calibration results.", + "bbox": [ + 75, + 332, + 468, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of $M$ : We ablate the number of augmented images sampled from the vicinal distribution of each labeled image per training epoch, setting $M = \\{1,2,3,4,5\\}$ . Results in Fig. 5c show that SCVRM is not very sensitive to the number of augmented data sampled $M$ .", + "bbox": [ + 75, + 483, + 468, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 568, + 194, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further demonstrate the effectiveness and generalisation ability of SCVRM in calibrating DNNs via additional experiments. The default training setting specified in Sec. 5.1 is adopted unless specified otherwise.", + "bbox": [ + 75, + 590, + 468, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness in Multi-Class Dense Classification: SCVRM can also be generalised to semantic segmentation [9]. We demonstrate its effectiveness in calibrating multi-class dense classification model (See Supp. 13).", + "bbox": [ + 75, + 651, + 468, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness in Other Binary Dense Classification Tasks: We verify the effectiveness of SCVRM in improving model calibration degree and dense classification accuracy of DNNs on additional binary dense classification tasks, such as Camouflaged Object Detection [10] and Smoke Detection [64] (See Supp. 12).", + "bbox": [ + 75, + 710, + 468, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness with Different Base Models: SCVRM is also effective in calibrating different base models while improving their dense classification accuracy. across a range of base models. We show its compatibility with VGG16 [49] and Swin transformer [34] (See Supp. 11).", + "bbox": [ + 75, + 801, + 468, + 878 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44b8be54fed88d2862b51a1b37474efd13af95c926442cfbde667a56c74c69d6.jpg", + "image_caption": [ + "(a) $\\gamma$ (horizontal axis)." + ], + "image_footnote": [], + "bbox": [ + 509, + 90, + 880, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af82a35d29d59ba54592d502afba38240a671342bc6fee1380c48ac000e26fdc.jpg", + "image_caption": [ + "(b) $\\eta$ (horizontal axis, and units in $\\sqrt{d}$ )." + ], + "image_footnote": [], + "bbox": [ + 511, + 203, + 879, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e80b5bed574d74890f5569a8e5f0d65898ee1bdb8c87a0257c33628fed628f99.jpg", + "image_caption": [ + "(c) $M$ (horizontal axis).", + "Figure 5. $\\mathrm{ECE}_{\\mathrm{EW}}$ and $\\mathrm{OE}_{\\mathrm{EW}}$ scores on the DUTS-TE [54] testing dataset under different choices of hyperparameters: (a) $\\gamma = \\{0.5,1.0,2.0,3.0,5.0,10.0\\}$ ; and (b) $\\eta = \\{i\\sqrt{d} | i = 0.1,0.5,1.0,2.0,5.0,10.0,20.0\\}$ ; and (c) $M = \\{1,2,3,4,5\\}$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 316, + 879, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness on Existing SOD Models: SCVRM can be utilised by various existing SOD models, e.g., EBMGSOD [71], EDN [60] and ICON [79] (See Supp.14).", + "bbox": [ + 498, + 489, + 890, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 549, + 617, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrated DNNs via exploring the vicinity space of labeled data. Vicinal images adopt the groundtruth label of the labeled image at the centre of vicinal distribution, but with diminishing label confidence as they get farther away. In a simplified setting, SCVRM is proved equivalent to a Vicinal Risk Minimisation plus a regularisation term, where the former improves model classification accuracy and the later penalises over-confident predictions. In practice, SCVRM is implemented as a data augmentation technique where MC sampling is applied to sample augmented data from the vicinal distribution. Experimental results on various dense classification tasks demonstrate the effectiveness of SCVRM in improving not only model calibration, but also dense classification accuracy. We also thoroughly study its compatibility with different backbone models and existing methods.", + "bbox": [ + 496, + 574, + 890, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 845, + 660, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was in-part supported by the ANU-Optus Bushfire Research Center of Excellence.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1Under-confidence occurs when ECE - OE > OE.", + "bbox": [ + 94, + 886, + 367, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3342", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 4", + "[2] GLENN W BRIER. Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1):1-3, 1950. 1, 2, 6", + "[3] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 2, 3, 4", + "[4] Zuyao Chen, Qianqian Xu, Runmin Cong, and Qingming Huang. Global context-aware progressive aggregation network for salient object detection. In AAAI, pages 10599-10606, 2020. 6", + "[5] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 37(3):569-582, 2014. 3", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6, 7", + "[7] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In ICML, pages 1310-1320. PMLR, 2019. 2, 3", + "[8] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 2", + "[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 88(2):303-338, 2010. 8", + "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 8", + "[11] Mengyang Feng, Huchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In CVPR, pages 1623-1632, 2019. 6", + "[12] Masoud Ghodrati, Amirhossein Farzmahdi, Karim Rajaei, Reza Ebrahimpour, and Seyed-Mahdi Khaligh-Razavi. Feedforward object-vision models only tolerate small image variations compared to human. Frontiers in computational neuroscience, 8:74, 2014. 2", + "[13] Arindam Ghosh, Thomas Schaaf, and Matthew R. Gormley. Adafocal: Calibration-aware adaptive focal loss. In NeurIPS, 2022. 1, 2, 6", + "[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, pages 1321-1330. PMLR, 2017. 1, 2, 3, 5, 6", + "[15] Shengfeng He, Rynson WH Lau, Wenxi Liu, Zhe Huang, and Qingxiong Yang. Supercnn: A superpixelwise convolutional neural network for salient object detection. IJCV, 115:330-344, 2015. 3", + "[16] Edwin T Jaynes. Information theory and statistical mechanics. Physical review, 106(4):620, 1957. 2", + "[17] Huaizu Jiang, Jingdong Wang, Zejian Yuan, Yang Wu, Nanning Zheng, and Shipeng Li. Salient object detection: A discriminative regional feature integration approach. In CVPR, pages 2083-2090, 2013. 3" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Zhuolin Jiang and Larry S Davis. Submodular salient region detection. In CVPR, pages 2043-2050, 2013. 3", + "[19] Tom Joy, Francesco Pinto, Ser-Nam Lim, Philip HS Torr, and Puneet K Dokania. Sample-dependent adaptive temperature scaling for improved calibration. In AAAI, pages 14919-14926, 2023. 1, 2", + "[20] Christoph Kamann and Carsten Rother. Benchmarking the robustness of semantic segmentation models. In CVPR, pages 8828-8838, 2020. 2", + "[21] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, pages 29768-29779, 2021. 1, 2", + "[22] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. In NeurIPS, 2019. 2", + "[23] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 1, 3, 6", + "[24] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 2", + "[25] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE symposium on security and privacy (SP), pages 656-672. IEEE, 2019. 3", + "[26] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 3, 5, 6, 8", + "[27] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, pages 2386-2395, 2017. 6", + "[28] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 6, 8", + "[29] Jiawei Liu, Jing Zhang, and Nick Barnes. Modeling aleatorc uncertainty for camouflaged object detection. In WACV, pages 1445-1454, 2022. 1", + "[30] Jiawei Liu, Jing Zhang, Ruikai Cui, Kaihao Zhang, Weihao Li, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022.", + "[31] Jiawei Liu, Changkun Ye, Shan Wang, Ruikai Cui, Jing Zhang, Kaihao Zhang, and Nick Barnes. Model calibration in dense classification with adaptive label perturbation. In ICCV, pages 1173-1184, 2023. 1, 2, 5, 6, 7", + "[32] Jiang-Jiang Liu, Qibin Hou, Zhi-Ang Liu, and Ming-Ming Cheng. Poolnet+: Exploring the potential of pooling for salient object detection. IEEE TPAMI, 45(1):887-904, 2022. 6", + "[33] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 3, 6", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3343", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 8", + "[35] Mingcan Ma, Changqun Xia, and Jia Li. Pyramidal feature shrinking for salient object detection. In AAAI, pages 2311-2318, 2021. 6", + "[36] David JC MacKay. Information theory, inference and learning algorithms. Cambridge university press, 2003. 4", + "[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 5, 6, 8", + "[38] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, pages 15288-15299, 2020. 1, 2, 6", + "[39] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 2, 4, 6", + "[40] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 2", + "[41] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In ICML, pages 625-632, 2005. 2", + "[42] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Hutchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 3, 6", + "[43] Hyekang Park, Jongyoun Noh, Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Acls: Adaptive and conditional label smoothing for network calibration. In ICCV, pages 3936-3945, 2023. 1, 2", + "[44] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740. IEEE, 2012. 3", + "[45] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 2", + "[46] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 3", + "[47] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 1, 2", + "[48] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 6", + "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5, 8", + "[50] Avishek Siris, Jianbo Jiao, Gary K.L. Tam, Xianghua Xie, and Rynson W.H. Lau. Scene context-aware salient object detection. In ICCV, pages 4156-4166, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurlPS, 2019. 2, 5, 6", + "[52] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 2, 3, 4", + "[53] Lijun Wang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Deep networks for saliency detection via local estimation and global search. In CVPR, pages 3183-3192, 2015. 3", + "[54] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 1, 5, 6, 7, 8", + "[55] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 6", + "[56] Tiantian Wang, Lihe Zhang, Shuo Wang, Huchuan Lu, Gang Yang, Xiang Ruan, and Ali Borji. Detect globally, refine locally: A novel approach to saliency detection. In CVPR, pages 3127-3135, 2018. 6", + "[57] Wenguan Wang, Jianbing Shen, Xingping Dong, Ali Borji, and Ruigang Yang. Inferring salient objects from human fixations. IEEE TPAMI, 42(8):1913-1927, 2019. 3", + "[58] Jun Wei, Shuhui Wang, and Qingming Huang. $\\mathbf{F}^3$ net: fusion, feedback and focus for salient object detection. In AAAI, pages 12321-12328, 2020. 6", + "[59] Runmin Wu, Mengyang Feng, Wenlong Guan, Dong Wang, Huchuan Lu, and Errui Ding. A mutual learning method for salient object detection with intertwined multi-supervision. In CVPR, pages 8150-8159, 2019. 3", + "[60] Yu-Huan Wu, Yun Liu, Le Zhang, Ming-Ming Cheng, and Bo Ren. Edn: Salient object detection via extremely-downsampled network. IEEE TIP, 2022. 6, 8", + "[61] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6", + "[62] Zhe Wu, Li Su, and Qingming Huang. Stacked cross refinement network for edge-aware salient object detection. In ICCV, pages 7264-7273, 2019. 3", + "[63] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 6, 8", + "[64] Siyuan Yan, Jing Zhang, and Nick Barnes. Transmission-guided bayesian generative model for smoke segmentation. In AAAI, pages 3009-3017, 2022. 1, 8", + "[65] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173. IEEE, 2013. 5, 6, 8", + "[66] Greg Yang, Tony Duan, J Edward Hu, Hadi Salman, Ilya Razenshteyn, and Jerry Li. Randomized smoothing of all shapes and sizes. In ICML, pages 10693-10705. PMLR, 2020. 3", + "[67] Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML, pages 609-616, 2001. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3344", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In ACM SIGKDD Int. Conf. Knowledge Disc. Data Min., pages 694-699, 2002. 2", + "[69] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3", + "[70] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In ICML. PMLR, 2020. 2", + "[71] Jing Zhang, Jianwen Xie, Nick Barnes, and Ping Li. Learning generative vision transformer with energy-based latent space for saliency prediction. In NeurIPS, pages 15448-15463, 2021. 3, 6, 8", + "[72] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, pages 1741-1750, 2018. 3, 6", + "[73] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, and James Zou. When and how mixup improves calibration. In ICML, pages 26135-26160. PMLR, 2022. 2", + "[74] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In ICCV, pages 202-211, 2017. 6", + "[75] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Hutchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 6", + "[76] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 6", + "[77] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, pages 3085-3094, 2019. 3", + "[78] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51. Springer, 2020. 3", + "[79] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 686 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3345", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_model.json b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..67cf79e80281ea08519992dc98e64aec6eda5e7a --- /dev/null +++ b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_model.json @@ -0,0 +1,2818 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.816, + 0.152 + ], + "angle": 0, + "content": "Self-Calibrating Vicinal Risk Minimisation for Model Calibration" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.181, + 0.686, + 0.233 + ], + "angle": 0, + "content": "Jiawei Liu, Changkun Ye, Ruikai Cui, Nick Barnes \nThe Australian National University \nCanberra, Australia" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.237, + 0.756, + 0.252 + ], + "angle": 0, + "content": "{jiawei.liu3, changkun.ye, ruikai.cui, nick.barnes}@anu.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.312, + 0.302 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.635 + ], + "angle": 0, + "content": "Model calibration, measuring the alignment between the prediction accuracy and model confidence, is an important metric reflecting model trustworthiness. Existing dense binary classification methods, without proper regularisation of model confidence, are prone to being over-confident. To calibrate Deep Neural Networks (DNNs), we propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) that explores the vicinity space of labeled data, where vicinal images that are farther away from labeled images adopt the groundtruth label with decreasing label confidence. We prove that in the logistic regression problem, SCVRM can be seen as a Vicinal Risk Minimisation plus a regularisation term that penalises the over-confident predictions. In practical implementation, SCVRM is approximated using Monte Carlo sampling that samples additional augmented training images and labels from the vicinal distributions. Experimental results demonstrate that SCVRM can significantly enhance model calibration for different dense classification tasks on both in-distribution and out-of-distribution data. Code is available at https://github.com/Carlisle-Liu/SCVRM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.208, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Binary dense classification tasks [10, 54, 64] have advanced significantly since the debut of Deep Neural Networks (DNNs) associated with complex network architectures and large numbers of trainable parameters. Increasing model complexity has been shown to negatively impact model calibration [14] which has remained under investigated in the binary dense classification [29-31]. Miscalibration, a mis-alignment between model confidence and prediction accuracy [38], is an undesirable qualify that hinders the deployment of DNNs, especially in safety critical applications. In this work, we study the model calibration for dense binary classification models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "The most commonly observed mis-calibration problem is over-confidence, where the model confidence is signifi-" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.288, + 0.891, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Figure 1. An illustration of Mixup [69] and our proposed SCVRM. Circles with solid boundary \\(\\{(x_i, y_i)\\}_{i=1}^4\\) are labeled training images. Vicinal images of Mixup (circles with dashed boundary) are only distributed along the vectors connecting the labeled images. On the other hand, in SCVRM, the label of vicinal image adopts the groundtruth category (shown in different colours) of the closest labeled image, but with label confidence (shown in colour intensity) reduced monotonically with increasing Euclidean distance between vicinal and labeled images. The example sample \\(\\otimes\\), with Mixup/SCVRM vicinal image and augmented label shown on the right, is selected at the same relative position to \\((x_4, y_4)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "cantly higher than the accuracy of its predictions on a cohort of samples [14]. Increasing research interest has been dedicated to study model calibration methods that regularise the probability associated with the prediction to be meaningful in reflecting the chance of the prediction being correct. The existing methods can be roughly categorised into three groups: (i) training objective based [2, 8, 13, 21, 23, 38, 45]; (ii) post-processing based [14, 19, 47]; and (iii) label augmentation based [31, 39, 43]. The first two categories focus on penalising over-confident predictions while the third cat" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3335" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "egory directly moderates the confidence of training labels for the DNN models to train on. They do not explore the image space in improving the model calibration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.367 + ], + "angle": 0, + "content": "Vicinal Risk Minimisation (VRM) extends Empirical Risk Minimisation (ERM) by introducing a vicinal distribution around each labeled data in the image space [3, 52]. Mixup [69], a variant of VRM, assumes that vicinal images are only distributed along the vectors connecting the labeled data (pair of image and label) pairs. In addition, instead of letting vicinal images adopt the hard groundtruth label of the nearest labeled image, Mixup assigns smoothed versions of nearest groundtruth labels to the vicinal images. In practice, it samples augmented data through the convex combination of labeled data pairs: \\((\\tilde{x} = \\lambda x_{i} + (1 - \\lambda)x_{i},\\tilde{y} = \\lambda y_{i} + (1 - \\lambda)y_{j})\\) , where the combination factor is sampled from a beta distribution: \\(\\lambda \\in \\mathrm{Beta}(\\alpha ,\\alpha)\\) . Followup works [69, 73] show that Mixup is effective in improving model calibration in the image classification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.656 + ], + "angle": 0, + "content": "We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrate the DNNs. Different from Mixup [69], the vicinal distribution of SCVRM can use any distribution that sufficiently covers the vicinity space with reasonable probability density. We consider the case where the vicinal distribution is assumed to be Gaussian with standard deviation being a random variable following a uniform distribution. In addition, the labels assigned to the vicinal images are softened versions of the groundtruth label associated with the labeled image at the distribution centre, where the strength of softening is proportional to the L2 distance between the vicinal and labeled images. As shown in Fig. 1, compared with Mixup [69], our proposed SCVRM has the following advantages: (1) the vicinity image space is not restricted by the pair-wise spatial relations of training images; (2) consistent label confidence across the pixels in the augmented label for dense prediction; and (3) defines confidence boundaries beyond the training distribution to better handle out-of-distribution samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Given a labeled training image set, SCVRM can be understood by the following principals: (1) the human visual system is invariant to small random image changes [12] so that images undergoing up to a certain level of transformation should remain correctly classified [3, 20, 52]. This is also in line with adversarial robustness approaches that require images under small perturbations to maintain their classifications [7]. However, they are no longer labeled training data, so their label confidence should be slightly reduced. (2) it is understood from information theory [16] that we have no information at extreme distance where the label should assume a uniform categorical distribution with maximum entropy. Following the principal of VRM, we propose this transition should be accomplished by a smooth vicinal transition from the exact label for the labeled images to a uniform categorical distribution at large distance from" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.787, + 0.107 + ], + "angle": 0, + "content": "the labeled data. (See Fig. 2 for examples.)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.364 + ], + "angle": 0, + "content": "We summarise our contributions as: (1) propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM), where the labels associated with vicinal images have reduced confidence with increasing Euclidean distance between vicinal and labelled images, to calibrate DNN models for the dense binary classification task; (2) we approximate the vicinal distribution with a Gaussian distribution whose standard deviation is a random variable following a uniform distribution, denoted as SCVRM-G; (3) we supply an example showing that SCVRM is equivalent to VRM and a regularisation term on prediction confidence under a simplified model; (4) We realise SCVRM as a data augmentation technique, where augmented data are sampled from the vicinal distributions with Monte Carlo methods. Finally, we show state-of-the-art calibration results for salient object detection (main paper); and camouflage object and smoke detection, and semantic segmentation (Supp. 12 and Supp. 13)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.377, + 0.65, + 0.393 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Model Calibration: aims to minimise the distributional gap between model confidence and prediction accuracy. Existing methods propose to align confidence with accuracy through (i) post-processing techniques, e.g., Temperature Scaling (TS) [14, 19], Platt Scaling [41, 47], Dirichlet Scaling [22], Bayesian Binning [40, 67], Isotropic Regression [68], and Mix-and-Match [70]; and (ii) objective functions including Brier Loss [2, 8], Confidence Penalty [45], Maximum Mean Calibration Error (MMCE) [24], Soft Calibration Objective [21] and Focal Loss [13, 38]. These methods emphasise on suppressing over-confident predictions to alleviate the mis-calibration issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Research efforts have also been dedicated to investigate the effect of data augmentation methods on the calibration degree. Label augmentation methods align the confidence distribution of the target label, to which the model predictions converge through optimisation, with the prediction accuracy distribution. Label Smoothing (LS) [39, 43] directly softens the target label probability, preventing the model from producing over-confident predictions. [31] presents an alternative label augmentation solution that stochastically perturbs the groundtruth label and aligns the confidence distribution of expected label with the prediction accuracy distribution. On the other hand, Mixup [69], which simultaneously augments both the input image and its corresponding groundtruth label conditioned on the relative position between training data, has also been demonstrated to improve the model calibration degree [51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Vicinal Risk Minimisation: is proposed by [3, 52] to explore the vicinity of labeled data in hope of achieving a better approximation of the expected risk. The vicinity space is approximated with a vicinal distribution that can be either estimated with sufficient unlabeled data, or otherwise" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3336" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "assumed to follow certain prior distribution, e.g. Gaussian distribution [3, 52]. Mixup [69] is an extension of VRM that restricts the vicinal data to be distributed along the vectors connecting the data pairs through convex combination. Further, Mixup applies the convex combination in both image and label spaces, resulting in the vicinal samples having a softened version of the groundtruth label associated with the nearest labeled image. VRM has also been applied in randomised smoothing [7, 25, 66] to achieve certified adversarial robustness, where a Gaussian kernel with fixed variance is employed to approximate the vicinal distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.532 + ], + "angle": 0, + "content": "Salient Object Detection: is a dense binary classification task. Conventional methods primarily depend on hard-crafted features to identify salient objects [5, 17, 18, 44]. Early DNN based approaches employ learned features of local regions, e.g., super-pixel, object proposal and image patch, which need to be further sequentially processed [15, 26, 53]. The advent of Fully Convolutional Network (FCN) shifted the research focus onto the network architecture design for better pixel-wise predictions. Multi-level feature aggregation takes full advantage of spatial cues imbued in low-level features and semantic information embodied in high-level features [42, 72, 78]. Further, attention mechanisms, including spatial attention, channel attention, etc., have been leveraged to explore intra- and inter-feature map correlations [33, 46, 77], with some works directly adopting transformer backbones with self-attention [50, 71]. A different direction investigates auxiliary cues, e.g. boundary and edge [59, 62], fixation [57]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.548, + 0.204, + 0.565 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.287, + 0.589 + ], + "angle": 0, + "content": "3.1. Settings and Notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.78 + ], + "angle": 0, + "content": "This papers deals primarily with dense binary classification over the image space \\(\\mathbb{R}^d\\) with the corresponding label set of \\(\\mathcal{V}\\in \\{0,1\\} ^d\\), where \\(d\\) is the dimension of image/label space, \"0\" and \"1\" represent the two categories. \\(\\mathcal{V}_s\\in [0,1]^d\\) denotes a soft label space, clearly we have \\(\\mathcal{V}\\subset \\mathcal{V}_s\\). Let \\(\\mathcal{D} = \\{(x_i,y_i)\\}_{i = 1}^N\\) be a finite dataset with image and label pairs \\((x_{i},y_{i})\\) sampled i.i.d. from the joint distribution \\(p(x,y)\\) defined on \\(\\mathcal{X}\\times \\mathcal{Y}\\). The task is to obtain an optimal classifier \\(f^{*}\\in \\mathcal{H}\\) in the Hypothesis space \\(\\mathcal{H}\\subset \\{f:\\mathcal{X}\\to \\mathcal{Y}\\}\\) that maximises the prediction accuracy and model calibration performance. The dimensionality or spatial index is omitted for simplicity wherever it is clear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The model calibration degree over the joint distribution \\( p(x,y) \\) can be evaluated in terms of Expected Calibration Error (ECE) [14, 23]. Let \\( S(x) \\in (0,1) \\) be the Sigmoid-activation value before classification: \\( f(x) = \\mathbb{1}(S(x) > 0.5) \\), where \\( \\mathbb{1}(\\cdot) \\) is an indicator function. The prediction confidence \\( c \\) and accuracy \\( a \\) can be defined as: \\( c = |S(x) - 0.5| + 0.5 \\), and \\( a = \\mathbb{1}(f(x) = y) \\) respectively. We further use \\( p_{f,\\mathcal{D}}(c,a) \\) to denote the joint dis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.199 + ], + "angle": 0, + "content": "ribution of prediction confidence and prediction accuracy of model, \\(f(\\cdot)\\) , on dataset, \\(\\mathcal{D}\\) . Then, ECE can be defined as: \\(\\mathrm{ECE}(p_{f,\\mathcal{D}}) = \\mathbb{E}_{p_f,\\mathcal{D}(c)}\\bigl [|\\mathbb{E}_{p_f,\\mathcal{D}(a|c)}[a] - c|\\bigr ]\\) , where \\(\\mathbb{P}_{f,\\mathcal{D}}(c)\\) is a marginal distribution on prediction confidence, and \\(\\mathbb{E}_{\\mathbb{P}_{\\theta ,\\mathcal{D}}(a|c)}\\) is a conditional distribution of prediction accuracy. For a well calibrated model, we should have: \\(p_{f,\\mathcal{D}}(f(x)_i = y_i|c = r) = r\\) \\(\\forall r\\in [0,1]\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.207, + 0.738, + 0.221 + ], + "angle": 0, + "content": "3.2. Vicinal Risk Minimisation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.276 + ], + "angle": 0, + "content": "We may formulate a learning problem as searching for model \\( f \\in \\mathcal{H} \\) that minimises expected risk \\( R(f) \\) over loss function \\( \\ell(f(x), y) \\), which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.286, + 0.892, + 0.32 + ], + "angle": 0, + "content": "\\[\nR (f) = \\int_ {\\mathcal {X} \\times \\mathcal {Y}} \\ell (f (x), y) \\cdot p (x, y) d x d y. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Given a dataset \\(\\mathcal{D} = \\{(x_i, y_i)\\}_i^N\\), the Empirical Risk Minimisation (ERM) approach approximates \\(R(f)\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.372, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R _ {E} (f) = \\int_ {\\mathcal {X} \\times \\mathcal {Y}} \\ell (f (x), y) \\cdot \\frac {1}{N} \\delta_ {x _ {i}, y _ {i}} (x, y) d x d y \\tag {2} \\\\ = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell \\left(f \\left(x _ {i}\\right), y _ {i}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.49 + ], + "angle": 0, + "content": "where \\(\\delta_{x_i,y_i}(x,y)\\) is the Dirac delta distribution over \\((x,y)\\) that can only adopts the value \\((x_i,y_i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Based on ERM, the Vicinal Risk Minimisation (VRM) approach [3, 52] approximates \\( R(f) \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.532, + 0.892, + 0.572 + ], + "angle": 0, + "content": "\\[\nR _ {V} (f) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(\\tilde {x} _ {i}\\right), y _ {i}\\right) p \\left(\\tilde {x} _ {i} \\mid x _ {i}\\right) d \\tilde {x}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.614 + ], + "angle": 0, + "content": "where \\( p(\\tilde{x}_i|x_i) \\) denotes the Probability Density Function (PDF) of vicinal images \\( \\tilde{x}_i \\) given labeled image \\( x_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.69 + ], + "angle": 0, + "content": "In practice, VRM proposes to use isotropic Gaussian distributions as a choice of \\( p(\\tilde{x}_i|x_i) \\). Specifically, VRM defines \\( \\tilde{x}_i = x_i + \\epsilon, \\epsilon \\sim \\mathcal{N}_d(0,\\sigma^2 I_d) \\), where \\( I_{d} \\) is a \\( d \\times d \\) identity matrix and the variance \\( \\sigma^2 \\) is a fixed hyperparameter. The corresponding \\( R_V(f) \\) can then be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.701, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\nR _ {V - G} (f; \\sigma) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), y _ {i}\\right) p (\\epsilon) d \\epsilon , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.892, + 0.793 + ], + "angle": 0, + "content": "where \\(p(\\epsilon) = \\frac{1}{\\sqrt{(2\\pi\\sigma^2)^d}}\\exp \\left(-\\frac{\\|\\epsilon\\|_2^2}{2\\sigma^2}\\right)\\) is the PDF of the additive Gaussian noise \\(\\epsilon\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.591, + 0.823 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.866, + 0.848 + ], + "angle": 0, + "content": "4.1. Self-Calibrating Vicinal Risk Minimisation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method is based on the principals that samples in the proximity of labeled data should inherit their groundtruth labels with reduced label confidence, while samples in the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3337" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.103, + 0.895, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.895, + 0.267 + ], + "angle": 0, + "content": "Figure 2. Examples of vicinal images and assigned labels under various standard deviation \\(\\sigma\\) values. \\(\\sigma = 0\\) denotes the labeled data \\((x,y)\\). Vicinal images are generated via adding Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(0,\\sigma^2 I_d)\\) to the labeled image as defined in Eq. 7. Their assigned labels, being the softer versions of the groundtruth label \\(y\\), are computed with the Eq. 8 and Eq. 9, whose plot are shown on the right." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.366 + ], + "angle": 0, + "content": "extreme distance should have a near uniform categorical distribution for their labels. As an extension of VRM, we reduce label confidence for vicinal images based on the L2-distance between vicinal and labeled images. Extending the definition of VRM [3, 52], we propose a novel Self-Calibrating Vicinal Risk Minimisation (SCVRM):" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.472, + 0.42 + ], + "angle": 0, + "content": "Definition 1. (SCVR) Let \\(\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^N\\). For \\(f \\in \\mathcal{H}\\) and \\(\\ell : f(\\mathcal{X}) \\times \\mathcal{Y}_s \\to \\mathbb{R}\\), Self-Calibrating Vicinal Risk is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.429, + 0.47, + 0.469 + ], + "angle": 0, + "content": "\\[\nR _ {S C} (f) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(\\tilde {x} _ {i}\\right), g \\left(y _ {i}, \\| \\tilde {x} _ {i} - x _ {i} \\| _ {2}\\right)\\right) \\times \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.473, + 0.321, + 0.489 + ], + "angle": 0, + "content": "\\[\np (\\tilde {x} _ {i} | x _ {i}) d \\tilde {x} _ {i},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.498, + 0.251, + 0.515 + ], + "angle": 0, + "content": "where \\(g:\\mathcal{Y}\\times \\mathbb{R}_{+}\\to \\mathcal{Y}_{s}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.675 + ], + "angle": 0, + "content": "The difference between VRM (Eq. 3) and SCVRM lies in the labels associated with the vicinal images. VRM assigns the exact groundtruth label \\( y_{i} \\) paired with labeled image \\( x_{i} \\) to all its vicinal images \\( \\tilde{x}_i \\), whereas we use softened versions of groundtruth label \\( g(y_{i},\\| \\tilde{x}_{i} - x_{i}\\|_{2}) \\), whose label confidence reduces with increases in the L2-distance between the vicinal and labeled images. Following VRM to represent the distribution of vicinal images conditioned on the labeled image \\( p(\\tilde{x}|x) \\) with an isotropic Gaussian distribution with a fixed variance, our SCVRM can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.696, + 0.47, + 0.749 + ], + "angle": 0, + "content": "\\[\nR _ {S C - V} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), g \\left(y _ {i}; \\| \\epsilon \\| _ {2}\\right)\\right) p (\\epsilon) d \\epsilon , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.781 + ], + "angle": 0, + "content": "where L2-norm of Gaussian noise equals the L2-distance between vicinal and labeled images \\(\\| \\epsilon \\| _2 = \\| \\tilde{x}_i - x_i\\| _2\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We observe that the isotropic Gaussian noise \\(\\epsilon \\sim \\mathcal{N}_d(0,\\sigma^2 I_d)\\) used in the VRM may face the problem of a bubbling effect, that results in the probability density of an isotropic Gaussian \\(\\mathcal{N}_d(\\mu ,\\sigma^2 I_d)\\) concentrating on a thin spherical shell centred on \\(\\mu \\in \\mathbb{R}^d\\) with radius \\(\\sigma \\sqrt{d}\\) in high dimensional space \\(d\\gg 1\\) [1, 36]. Following this observation: (a) we propose to let the scale of the standard deviation of the isotropic Gaussian distribution be a random" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.895, + 0.38 + ], + "angle": 0, + "content": "variable following a uniform distribution \\(\\sigma \\sim \\mathcal{U}(0,\\gamma]\\) to prevent vicinal images \\(\\tilde{x}\\) from being distributed only near a hypersphere centring on \\(x_{i}\\); and (b) we approximate the L2-distance between the vicinal and labeled images with a dummy variable \\(\\| \\tilde{\\epsilon}\\| _2\\approx \\| \\epsilon \\| _2\\) which is set to \\(\\| \\tilde{\\epsilon}\\| _2 = \\sigma \\sqrt{d}\\) (See Supp. 8.1 for experimental justifications). Our implementation of SCVRM can then be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.388, + 0.892, + 0.428 + ], + "angle": 0, + "content": "\\[\nR _ {S C - G} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), g \\left(y _ {i}; \\| \\tilde {\\epsilon} \\| _ {2}\\right)\\right) \\times \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.673, + 0.432, + 0.774, + 0.449 + ], + "angle": 0, + "content": "\\[\np (\\epsilon) p (\\sigma) d \\epsilon d \\sigma ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.459, + 0.884, + 0.475 + ], + "angle": 0, + "content": "where \\(p(\\sigma) = 1 / \\gamma\\) is the PDF of the standard deviation \\(\\sigma\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.567 + ], + "angle": 0, + "content": "There are many choices for the projection function from hard label space to soft label space. The most straightforward solution is to apply the label smoothing [39] function: \\(\\mathrm{LS}(y,\\beta) = (1 - \\beta) + \\frac{\\beta}{K},\\beta \\in [0,1]\\) where \\(K = 2\\) denotes the number of classes in a binary task. Then our projection function can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.576, + 0.892, + 0.607 + ], + "angle": 0, + "content": "\\[\ng _ {\\mathrm {L S}} (y; \\varphi (\\| \\tilde {\\epsilon} \\| _ {2})) = \\left(1 - \\varphi (\\| \\tilde {\\epsilon} \\| _ {2})\\right) y + \\frac {\\varphi (\\| \\tilde {\\epsilon} \\| _ {2})}{2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.648 + ], + "angle": 0, + "content": "where we define \\(\\varphi (\\cdot):\\mathbb{R}_0^+\\to [0,1)\\) to be a Gaussian function \\(\\varphi_G(\\cdot ;\\cdot)\\), that is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.658, + 0.892, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {G} \\left(\\| \\tilde {\\epsilon} \\| _ {2}; \\eta\\right) = 1 - \\exp \\left(- \\frac {\\| \\tilde {\\epsilon} \\| _ {2}}{\\eta}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.795 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a hyperparameter that scales the value of \\(\\| \\tilde{\\epsilon}\\| _2\\), and we empirically set \\(\\eta = \\sqrt{d}\\). Fig. 2 illustrates the resultant groundtruth map assigned to vicinal images sampled under various standard deviations. Please note that the function \\(\\varphi\\) can be an arbitrary function that satisfies: \\(\\varphi :\\mathbb{R}_0^+ \\to [0,1)\\). We investigate other options in Supp. 8.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.802, + 0.844, + 0.817 + ], + "angle": 0, + "content": "4.2. Connection between SCVRM and VRM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We analyse the relationship between VRM and our proposed SCVRM in a simple setting, where we find mathematical connections between the two approaches. Specifically, we consider the case of binary classification tasks, where our \\( R_{SC - G}(f) \\) in Eq. 7 adopts a single-layer logistic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3338" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "regression model and the binary cross entropy loss. In this setting, we prove that if \\( g(y_{i},\\| \\tilde{\\epsilon}\\|_{2}) \\) with \\( \\| \\tilde{\\epsilon}\\| _2 = \\sigma \\sqrt{d} \\) is a label smoothing function, \\( R_{SC - G}(f) \\) can be written as a combination of the Vicinal Risk \\( R_{V - G}(f;\\sigma) \\) in Eq. 4 and a regularization term that penalises overconfident predictions of the model \\( f \\) (See proof in Supp. 7.1)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.2, + 0.435, + 0.215 + ], + "angle": 0, + "content": "Example 1. In \\( R_{SC - G}(f) \\) defined in Eq. 7, suppose:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.216, + 0.425, + 0.232 + ], + "angle": 0, + "content": "- Data: \\(\\{(x_i, y_i)\\}_{i=1}^N\\) with \\(x_i \\in \\mathbb{R}^d\\) and \\(y_i \\in \\{0, 1\\}\\)," + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.233, + 0.395, + 0.248 + ], + "angle": 0, + "content": "- Model: \\( f(x) = 1 / (1 + e^{-w^T x}) \\) with \\( w \\in \\mathbb{R}^d \\)," + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.249, + 0.468, + 0.265 + ], + "angle": 0, + "content": "- Loss: \\(\\ell(f(x), y) = -y \\log f(x) - (1 - y) \\log (1 - f(x))\\)," + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.265, + 0.468, + 0.296 + ], + "angle": 0, + "content": "- \\(g(y, \\| \\tilde{\\epsilon} \\|_2) = (1 - 2\\sigma)y + \\sigma, \\sigma \\sim \\mathcal{U}(0, \\gamma]\\) and \\(\\gamma \\in (0, \\frac{1}{2}]\\). then we have:" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.216, + 0.468, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.31, + 0.47, + 0.343 + ], + "angle": 0, + "content": "\\[\nR _ {S C - G} (f) = \\int_ {0} ^ {\\gamma} \\frac {1}{\\gamma} R _ {V - G} (f; \\sigma) d \\sigma + \\tau (f), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.358, + 0.124, + 0.371 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.376, + 0.469, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\tau (f) = \\frac {\\gamma}{2 N} \\sum_ {i = 1} ^ {N} \\left(2 y _ {i} - 1\\right) \\cdot w ^ {T} x _ {i}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.43, + 0.47, + 0.476 + ], + "angle": 0, + "content": "and the first term of the RHS in Eq. 10 is equivalent to introducing our design of \\(\\sigma \\sim \\mathcal{U}(0,\\gamma]\\) into \\(R_{V - G}(f;\\sigma)\\) which is the VRM with Gaussian kernel defined in Eq. 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.47, + 0.523 + ], + "angle": 0, + "content": "Note that the label satisfies \\( y_{i} \\in \\{0,1\\} \\), therefore in the term \\( \\tau (f) \\) in Eq. 11 we have:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.539, + 0.469, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\left(2 y _ {i} - 1\\right) \\cdot w ^ {T} x _ {i} = \\left\\{ \\begin{array}{l l} w ^ {T} x _ {i}, & y _ {i} = 1, \\\\ - w ^ {T} x _ {i}, & y _ {i} = 0. \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.47, + 0.676 + ], + "angle": 0, + "content": "In this sense, the term \\(\\tau(f)\\) penalises high \\(w^T x_i\\) when \\(y_i = 1\\) and penalises low \\(w^T x_i\\) when \\(y_i = 0\\). Since high \\(w^T x\\) will lead to \\(f(x) = 1 / (1 + e^{w^T x})\\) approaching 1 and low \\(w^T x\\) will lead to \\(f(x)\\) approaching 0, \\(\\tau(f)\\) actually penalizes overconfident predictions of the model \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.692, + 0.47, + 0.799 + ], + "angle": 0, + "content": "Remark 1. In this simplified case, minimising the Self-Calibrating Vicinal Risk defined in Eq. 7 equivalently minimises the Vicinal Risk defined in Eq. 4 incorporating our design of \\(\\sigma \\in \\mathcal{U}(0,\\gamma]\\) and a regularisation term defined in Eq. 11, where the former improves the model classification accuracy the latter prevents the model from becoming overconfident on vicinal samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.237, + 0.83 + ], + "angle": 0, + "content": "4.3. Practical Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In practical implementation, we approximate the intractable SCVRM defined in Eq. 7 with the Monte-Carlo (MC) Sampling. More specifically, it is formulated as a data augmentation technique where, based on the labeled dataset" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.855, + 0.108 + ], + "angle": 0, + "content": "\\(\\mathcal{D} = \\{(x_{1},y_{i})\\}_{i = 1}^{N}\\), we sample an augmented dataset:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.116, + 0.891, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} _ {s c} = \\bigcup_ {i = 1} ^ {N} \\left\\{(\\tilde {x} _ {i} ^ {j}, \\tilde {y} _ {i} ^ {j}) \\mid \\tilde {x} _ {i} ^ {j} = x _ {i} + \\epsilon_ {i} ^ {j}, \\epsilon_ {i} ^ {j} \\sim_ {i. i. d} \\mathcal {N} (0, \\sigma_ {i, j} ^ {2} I _ {d}), \\right. \\\\ \\sigma_ {i, j} \\sim_ {i. i. d} \\mathcal {U} (0, \\gamma ], \\\\ \\left. \\tilde {y} _ {i} ^ {j} = g _ {\\mathrm {L S}} \\left(y _ {i}; \\varphi_ {G} (\\| \\tilde {\\epsilon} \\| _ {2}; \\eta)\\right) \\right\\} _ {j = 1} ^ {M}, \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.892, + 0.406 + ], + "angle": 0, + "content": "where \\(\\tilde{x}_i^j\\) is the \\(j^{th}\\) augmented image sampled from the vicinity of the \\(i^{th}\\) labeled image, \\(\\tilde{y}_i^j\\) is a smoothed version of the groundtruth label \\(y_i\\), \\(M\\) is a hyperparameter being the number of augmented data from the vicinity of each labeled data, and we set \\(\\| \\tilde{\\epsilon} \\|_2 = \\sigma \\sqrt{d}\\) to approximate the L2-distance between the vicinal and labeled images \\(\\sigma \\sqrt{d} \\approx \\| \\epsilon \\|_2\\) which relieves the computational burden of computing an L2-norm in high-dimensional space. The model is trained on both the labeled dataset \\(\\mathcal{D}\\) and the augmented dataset \\(\\mathcal{D}_{sc}\\) using a Binary Cross Entropy loss: \\(\\mathcal{L}_{bce}(f(x), y) = -y \\log(f(x)) - (1 - y) \\log(1 - f(x))\\). The training loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.413, + 0.892, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {b c e} \\left(f \\left(x _ {i}\\right), y _ {i}\\right) + \\sum_ {j = 1} ^ {M} \\mathcal {L} _ {b c e} \\left(f \\left(\\tilde {x} _ {i} ^ {j}\\right), \\tilde {y} _ {i} ^ {j}\\right)\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.5 + ], + "angle": 0, + "content": "Note that the augmented dataset defined in Eq. 13 is resampled after each training epoch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.734, + 0.529 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.614 + ], + "angle": 0, + "content": "The main paper presents results on Salient Object Detection. Results for other dense classification tasks, e.g., Camouflaged Object Detection (binary) and Smoke Detection (binary), and semantic segmentation (multi-class) can be found in Supp. 12 and 13 respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.719, + 0.637 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Evaluation Metrics: We employ bin-based Expected Calibration Error (denoted \\(\\mathrm{ECE}_{\\mathrm{EW}}\\) [14]) and Over-confidence Error \\((\\mathrm{OE}_{\\mathrm{EW}}\\) [51]) with \\(B = 10\\) bins, to evaluate model calibration. (See Supp. 9.1 for implementation details.)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Datasets: Following [31] we divide DUTS-TR [54] into a train \\((\\mathcal{D}_{\\mathrm{TR}})\\) and validation set \\((\\mathcal{D}_{\\mathrm{VAL}})\\), 9,553 and 1,000 images respectively. We evaluate model calibration degree using SOD test datasets, DUTS-TE [54], DUT-OMRON [65], PASCAL-S [28], SOD [37], ECSSD [63], HKU-IS [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Model Architecture: The U-Net based model consists of a ResNet50 encoder and a decoder that are initialised with ImageNet pretrained weights and by default respectively. The implementation uses the Pytorch framework. Experiments with different encoders, e.g. VGG16 [49] and Swin transformer [34] are detailed in Supp. 11." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Hyperparameters: Optimal calibration performance of model is achieved by setting \\(\\gamma = 2\\), \\(\\eta = \\sqrt{d}\\) and \\(M = 3\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3339" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.09, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Table 1. Salient object detection model calibration benchmark. Results are evaluated with \\( {\\mathrm{{ECE}}}_{\\mathrm{{EW}}} \\) and \\( {\\mathrm{{OE}}}_{\\mathrm{{EW}}} \\) with 10 bins (units in %)." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.105, + 0.868, + 0.47 + ], + "angle": 0, + "content": "
MethodsDUTS-TE [54]DUT-OMRON [65]PASCAL-S [28]SOD [37]ECSSD [63]HKU-IS [26]
CategoryNameYearECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓
SOD MethodsMSRNet [27]20172.572.343.323.163.443.236.426.140.970.940.920.87
SRM [55]20174.023.724.193.964.884.599.939.582.532.351.861.72
Amulet [74]20175.675.285.845.495.765.4310.039.592.562.421.981.87
BMPM [72]20183.743.524.524.374.884.688.167.931.951.891.581.53
DGRL [56]20184.123.864.414.215.014.778.448.202.132.021.631.53
PAGR [75]20184.043.795.144.965.645.3712.1711.872.842.701.621.54
PiCANet [33]20185.124.904.844.708.147.9210.5010.303.483.392.552.47
CPD [61]20193.973.784.204.065.375.179.659.392.292.191.991.90
BASNet [48]20195.004.864.934.836.506.3610.4010.272.742.702.302.26
EGNet [76]20193.333.143.663.505.425.198.047.791.981.881.471.40
AFNet [11]20193.953.744.254.095.064.848.158.022.382.271.871.78
PoolNet [32]20193.333.123.863.705.325.078.147.872.001.901.821.75
GCPANet [4]20203.182.993.993.844.163.977.056.881.611.541.271.21
MINet [42]20203.653.484.454.294.944.758.017.892.132.031.741.65
F3Met [58]20203.673.504.254.104.854.677.957.782.262.161.921.83
EBMGSD [71]20213.453.294.113.954.794.617.487.302.142.051.791.70
ICON [79]20212.892.763.843.714.083.956.706.551.561.491.381.32
PFSNet [35]20212.942.723.953.814.454.277.597.392.412.252.061.96
EDN [60]20223.623.474.023.904.894.748.818.662.202.131.651.58
Model Calibration MethodsBrier Loss [2]19502.772.583.553.383.903.706.406.161.371.301.040.99
Temperature Scaling [14]20172.532.343.183.033.563.366.326.050.960.930.830.70
MMCE [23]20182.862.673.563.414.003.816.856.631.411.351.181.13
Label Smoothing [39]20192.001.792.892.713.042.835.975.690.830.680.820.47
Mixup [51]20192.452.253.413.233.132.995.825.701.411.183.830.05
Focal Loss [38]20202.252.083.102.823.403.136.215.981.411.031.240.77
AdaFocal [13]20221.611.412.311.842.532.275.885.471.630.791.350.52
ASLP [31]20231.401.221.991.832.312.105.505.170.480.200.790.17
OursSCVRM20230.780.611.641.491.911.753.903.600.440.190.780.10
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.596 + ], + "angle": 0, + "content": "**Optimisation Details:** The model is optimised for 30 epochs with an initial learning rate of \\(2.5 \\times 10^{-5}\\) that decays by a factor of 0.9 per epoch after the \\(10^{\\text{th}}\\) epoch. During each epoch, \\(M = 3\\) augmented data are sampled from each vicinity space. Image size is set to \\(3 \\times 384 \\times 384\\) and the batch size is 8. Basic data augmentation techniques including random flipping, translation and cropping are applied." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.418, + 0.621 + ], + "angle": 0, + "content": "5.2. Model Calibration Degree Performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.869 + ], + "angle": 0, + "content": "The calibration degrees of existing SOD and model calibration methods, and our proposed SCVRM, evaluated in terms of \\(\\mathrm{ECE}_{\\mathrm{EW}}\\) and \\(\\mathrm{OE}_{\\mathrm{EW}}\\), are presented in Tab. 1. It can be observed that SOD methods without regularisations on prediction confidence are, in general, less well calibrated than the models equipped with calibration methods. On the other hand, our proposed SCVRM achieves the lowest calibration errors across all six testing datasets among the model calibration methods. Largest improvements over the second-best model are obtained on DUTS-TE, DUTOMRON, PASCAL-S and SOD datasets, reducing the ECE by \\(44.3\\%\\), \\(17.6\\%\\), \\(17.3\\%\\) and \\(29.1\\%\\) respectively. There is little room for improvement on ECSSD and HKU-IS, where ASLP is already well calibrated. On these datasets, our proposed SCVRM maintains the high calibration degrees that are comparable with those of ASLP." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 depicts the joint distribution of prediction confidence and prediction accuracy of some model calibration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.641 + ], + "angle": 0, + "content": "methods on the DUTS-TE [54] dataset. The joint distribution of better calibrated models become more closely aligned with the oracle line, which indicates a perfectly calibrated model. Our SCVRM has its joint distribution almost completely aligned with the oracle line. In comparison, Mixup is more over-confident with the high-density area of its joint distribution slightly to the bottom-right of the oracle and the low-density area deviating even more. This further demonstrates the effectiveness of exploring a continuous vicinal image space with softened labels." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Table 2. Existing model calibration methods and our proposed LSR evaluated on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6] in terms of \\(\\mathrm{ECE_{EW}}\\) and \\(\\mathrm{OE_{EW}}\\) with 10 bins, and Accuracy (ACC)." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.714, + 0.857, + 0.894 + ], + "angle": 0, + "content": "
MethodEvaluation (%)
ECE ↓OE ↓ACC ↑
Baseline52.3651.0541.88
Brier Loss [2]38.8537.1853.62
Temperature Scaling [14]51.9550.4641.59
Label Smoothing [39]37.2235.4855.41
MMCE [23]40.6439.6754.39
Mixup [51]31.0729.1058.71
Focal Loss [38]40.0138.4349.71
AdaFocal [13]27.5525.0755.39
ASLP [31]18.3116.3761.93
SCVRM11.938.2683.93
" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.087, + 0.209, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.187, + 0.205, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.284, + 0.186, + 0.295 + ], + "angle": 0, + "content": "Label Smoothing" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.089, + 0.336, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.187, + 0.336, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.284, + 0.29, + 0.294 + ], + "angle": 0, + "content": "Mixup" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.089, + 0.468, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.187, + 0.467, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.284, + 0.432, + 0.294 + ], + "angle": 0, + "content": "Focal Loss" + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.089, + 0.597, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.187, + 0.596, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.284, + 0.558, + 0.293 + ], + "angle": 0, + "content": "AdaFocal" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.089, + 0.727, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.187, + 0.725, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.284, + 0.679, + 0.293 + ], + "angle": 0, + "content": "ASLP" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.089, + 0.857, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.187, + 0.857, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.284, + 0.815, + 0.293 + ], + "angle": 0, + "content": "SCVRM" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.892, + 0.331 + ], + "angle": 0, + "content": "Figure 3. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on DUTS-TE [54]. The dashed red diagonal line represents the perfectly calibrated oracle model." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.341, + 0.465, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Figure 4. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6]. The dashed red diagonal line represents the perfectly calibrated oracle model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.47, + 0.755 + ], + "angle": 0, + "content": "5.3. Model Calibration on Out-of-Distribution Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The calibrating ability of existing model calibration techniques and our proposed SCVRM on Out-of-Distribution (OoD) samples are evaluated on the 500 texture images [31] selected from the Describable Texture Dataset [6]. These texture images, not including any salient (foreground) object, demonstrates certain level of distributional shift from both training and testing distributions of SOD. As shown in Tab. 2, SCVRM achieves the lowest calibration errors on OoD data. Compared with the second-best model, it" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.482 + ], + "angle": 0, + "content": "reduces ECE by \\(34.8\\%\\) and OE by \\(49.5\\%\\). The improvements are also reflected in the joint distribution plot in Fig. 4, where the distribution area of SCVRM is significantly better aligned with the oracle line. Further, SCVRM significantly outperforms Mixup in terms of both calibration and classification. Its superior performance can be attributed to more effective utilisation of vicinity space, and a distance-based label augmentation technique that yields consistent label confidence across the pixels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.654, + 0.511 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.565 + ], + "angle": 0, + "content": "We investigate the effect of SCVRM and its hyperparameters including \\(\\gamma\\), \\(\\eta\\) and \\(M\\) which adopt the default setting specified in Sec. 5.1 unless stated otherwise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Effect of SCVRM: We compare our SCVRM with ERM, the baseline model, VRM with a fixed standard deviation, and \\(\\mathrm{VRM}^*\\) with our design of \\(\\sigma \\sim \\mathcal{U}(0,\\gamma]\\) (see implementation detail in Supp. 8.3). Tab. 3 shows that VRM (see implementation detail in Supp. 8.3) does not reduce calibration error over ERM in general, producing mixed impacts across the testing datasets. Introducing our design of \\(\\sigma \\sim \\mathcal{U}(0,\\gamma]\\) also does not alleviate VRM's over-confidence. On the other hand, SCVRM significantly improves the model calibration over ERM, VRM and \\(\\mathrm{VRM}^*\\). In addition, SCVRM also achieves improved classification accuracy than ERM (see Supp. 10.1), which can be attributed to more effective utilisation of vicinity space with our design of \\(\\sigma \\sim \\mathcal{U}(0,\\gamma]\\). It also enables \\(\\mathrm{VRM}^*\\) to achieve better classification accuracy compared to VRM (see Supp. 10.1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of \\(\\gamma\\): The hyperparameter \\(\\gamma\\) suggests the exploration radius around the training images. As illustrated in Fig. 5a, SCVRM can reach optimal calibration performance in a wide range \\(\\gamma = [1, 10]\\). When \\(\\gamma\\) is too small, e.g. 0.5, SCVRM achieves limited improvements on model calibration as the vicinal images are too close to the labeled images, and their corresponding augmented labels retain rel" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3341" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.164, + 0.09, + 0.383, + 0.103 + ], + "angle": 0, + "content": "Table 3. Ablation study on SCVRM." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.103, + 0.457, + 0.254 + ], + "angle": 0, + "content": "
MethodECE ↓ DUTS-TE [54]ECE ↓ DUT-OMRON [65]ECE ↓ PASCAL-S [28]
ERM (Baseline)3.052.893.803.684.143.98
VRM3.543.374.354.214.394.24
VRM*13.183.033.903.794.133.98
SCVRM0.780.611.641.491.911.75
MethodSOD [37]ECSSD [63]HKU-IS [26]
ERM (Baseline)7.076.851.821.761.381.34
VRM6.606.421.671.621.301.26
VRM*17.357.161.721.681.281.23
SCVRM3.903.600.440.190.780.10
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.094, + 0.255, + 0.332, + 0.267 + ], + "angle": 0, + "content": "\\(^{1}\\) VRM* incorporates our design of \\(\\sigma \\sim \\mathcal{U}(0, \\gamma]\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.469, + 0.332 + ], + "angle": 0, + "content": "atively high label confidences (see Fig. 2). As \\(\\gamma\\) grows, the augmented vicinal images can get sufficiently far away from the labeled images to set up label confidence contours of various levels and obtains optimal model calibration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.483 + ], + "angle": 0, + "content": "Effect of \\(\\eta\\): The hyperparameter \\(\\eta\\) is part of the Gaussian equation (Eq. 9) that affects the strength of smoothing factor. (See Supp. 8.4 for example vicinal data under different \\(\\eta\\).) At \\(\\eta = 0.1\\sqrt{d}\\), the resultant augmented label is overly softened, leading to the trained model becoming under-confirmed1. When \\(\\eta\\) is too large, e.g. \\(5\\sqrt{d}\\) and \\(10.0\\sqrt{d}\\), vicinal images all retain a relatively high label confidence, resulting in insufficient calibration regularisation. We find that \\(\\eta\\) works well in \\([0.5\\sqrt{d}, 2\\sqrt{d}]\\), leading consistently to near-optimal calibration results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.469, + 0.56 + ], + "angle": 0, + "content": "Effect of \\(M\\): We ablate the number of augmented images sampled from the vicinal distribution of each labeled image per training epoch, setting \\(M = \\{1,2,3,4,5\\}\\). Results in Fig. 5c show that SCVRM is not very sensitive to the number of augmented data sampled \\(M\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.569, + 0.195, + 0.582 + ], + "angle": 0, + "content": "5.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.469, + 0.652 + ], + "angle": 0, + "content": "We further demonstrate the effectiveness and generalisation ability of SCVRM in calibrating DNNs via additional experiments. The default training setting specified in Sec. 5.1 is adopted unless specified otherwise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.469, + 0.712 + ], + "angle": 0, + "content": "Effectiveness in Multi-Class Dense Classification: SCVRM can also be generalised to semantic segmentation [9]. We demonstrate its effectiveness in calibrating multi-class dense classification model (See Supp. 13)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.803 + ], + "angle": 0, + "content": "Effectiveness in Other Binary Dense Classification Tasks: We verify the effectiveness of SCVRM in improving model calibration degree and dense classification accuracy of DNNs on additional binary dense classification tasks, such as Camouflaged Object Detection [10] and Smoke Detection [64] (See Supp. 12)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.469, + 0.879 + ], + "angle": 0, + "content": "Effectiveness with Different Base Models: SCVRM is also effective in calibrating different base models while improving their dense classification accuracy. across a range of base models. We show its compatibility with VGG16 [49] and Swin transformer [34] (See Supp. 11)." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.092, + 0.882, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.19, + 0.757, + 0.201 + ], + "angle": 0, + "content": "(a) \\(\\gamma\\) (horizontal axis)." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.204, + 0.88, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.302, + 0.802, + 0.314 + ], + "angle": 0, + "content": "(b) \\(\\eta\\) (horizontal axis, and units in \\(\\sqrt{d}\\))." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.317, + 0.88, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.414, + 0.76, + 0.426 + ], + "angle": 0, + "content": "(c) \\(M\\) (horizontal axis)." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Figure 5. \\(\\mathrm{ECE}_{\\mathrm{EW}}\\) and \\(\\mathrm{OE}_{\\mathrm{EW}}\\) scores on the DUTS-TE [54] testing dataset under different choices of hyperparameters: (a) \\(\\gamma = \\{0.5,1.0,2.0,3.0,5.0,10.0\\}\\); and (b) \\(\\eta = \\{i\\sqrt{d} | i = 0.1,0.5,1.0,2.0,5.0,10.0,20.0\\}\\); and (c) \\(M = \\{1,2,3,4,5\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Effectiveness on Existing SOD Models: SCVRM can be utilised by various existing SOD models, e.g., EBMGSOD [71], EDN [60] and ICON [79] (See Supp.14)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.619, + 0.565 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.832 + ], + "angle": 0, + "content": "We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrated DNNs via exploring the vicinity space of labeled data. Vicinal images adopt the groundtruth label of the labeled image at the centre of vicinal distribution, but with diminishing label confidence as they get farther away. In a simplified setting, SCVRM is proved equivalent to a Vicinal Risk Minimisation plus a regularisation term, where the former improves model classification accuracy and the later penalises over-confident predictions. In practice, SCVRM is implemented as a data augmentation technique where MC sampling is applied to sample augmented data from the vicinal distribution. Experimental results on various dense classification tasks demonstrate the effectiveness of SCVRM in improving not only model calibration, but also dense classification accuracy. We also thoroughly study its compatibility with different backbone models and existing methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.661, + 0.863 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This research was in-part supported by the ANU-Optus Bushfire Research Center of Excellence." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.368, + 0.9 + ], + "angle": 0, + "content": "1Under-confidence occurs when ECE - OE > OE." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3342" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.185 + ], + "angle": 0, + "content": "[2] GLENN W BRIER. Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1):1-3, 1950. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.229 + ], + "angle": 0, + "content": "[3] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.468, + 0.285 + ], + "angle": 0, + "content": "[4] Zuyao Chen, Qianqian Xu, Runmin Cong, and Qingming Huang. Global context-aware progressive aggregation network for salient object detection. In AAAI, pages 10599-10606, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.468, + 0.328 + ], + "angle": 0, + "content": "[5] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 37(3):569-582, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.468, + 0.372 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.468, + 0.415 + ], + "angle": 0, + "content": "[7] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In ICML, pages 1310-1320. PMLR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.468, + 0.47 + ], + "angle": 0, + "content": "[8] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.468, + 0.515 + ], + "angle": 0, + "content": "[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 88(2):303-338, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.468, + 0.558 + ], + "angle": 0, + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.468, + 0.601 + ], + "angle": 0, + "content": "[11] Mengyang Feng, Huchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In CVPR, pages 1623-1632, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[12] Masoud Ghodrati, Amirhossein Farzmahdi, Karim Rajaei, Reza Ebrahimpour, and Seyed-Mahdi Khaligh-Razavi. Feedforward object-vision models only tolerate small image variations compared to human. Frontiers in computational neuroscience, 8:74, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[13] Arindam Ghosh, Thomas Schaaf, and Matthew R. Gormley. Adafocal: Calibration-aware adaptive focal loss. In NeurIPS, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, pages 1321-1330. PMLR, 2017. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.468, + 0.813 + ], + "angle": 0, + "content": "[15] Shengfeng He, Rynson WH Lau, Wenxi Liu, Zhe Huang, and Qingxiong Yang. Supercnn: A superpixelwise convolutional neural network for salient object detection. IJCV, 115:330-344, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[16] Edwin T Jaynes. Information theory and statistical mechanics. Physical review, 106(4):620, 1957. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Huaizu Jiang, Jingdong Wang, Zejian Yuan, Yang Wu, Nanning Zheng, and Shipeng Li. Salient object detection: A discriminative regional feature integration approach. In CVPR, pages 2083-2090, 2013. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[18] Zhuolin Jiang and Larry S Davis. Submodular salient region detection. In CVPR, pages 2043-2050, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[19] Tom Joy, Francesco Pinto, Ser-Nam Lim, Philip HS Torr, and Puneet K Dokania. Sample-dependent adaptive temperature scaling for improved calibration. In AAAI, pages 14919-14926, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[20] Christoph Kamann and Carsten Rother. Benchmarking the robustness of semantic segmentation models. In CVPR, pages 8828-8838, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[21] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, pages 29768-29779, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[22] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[23] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[24] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[25] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE symposium on security and privacy (SP), pages 656-672. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[26] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[27] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, pages 2386-2395, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[28] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[29] Jiawei Liu, Jing Zhang, and Nick Barnes. Modeling aleatorc uncertainty for camouflaged object detection. In WACV, pages 1445-1454, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[30] Jiawei Liu, Jing Zhang, Ruikai Cui, Kaihao Zhang, Weihao Li, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[31] Jiawei Liu, Changkun Ye, Shan Wang, Ruikai Cui, Jing Zhang, Kaihao Zhang, and Nick Barnes. Model calibration in dense classification with adaptive label perturbation. In ICCV, pages 1173-1184, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[32] Jiang-Jiang Liu, Qibin Hou, Zhi-Ang Liu, and Ming-Ming Cheng. Poolnet+: Exploring the potential of pooling for salient object detection. IEEE TPAMI, 45(1):887-904, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[33] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3343" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[35] Mingcan Ma, Changqun Xia, and Jia Li. Pyramidal feature shrinking for salient object detection. In AAAI, pages 2311-2318, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[36] David JC MacKay. Information theory, inference and learning algorithms. Cambridge university press, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[38] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, pages 15288-15299, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[39] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[40] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[41] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In ICML, pages 625-632, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[42] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Hutchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[43] Hyekang Park, Jongyoun Noh, Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Acls: Adaptive and conditional label smoothing for network calibration. In ICCV, pages 3936-3945, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[44] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[45] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[46] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[47] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[48] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[50] Avishek Siris, Jianbo Jiao, Gary K.L. Tam, Xianghua Xie, and Rynson W.H. Lau. Scene context-aware salient object detection. In ICCV, pages 4156-4166, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[51] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurlPS, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[52] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[53] Lijun Wang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Deep networks for saliency detection via local estimation and global search. In CVPR, pages 3183-3192, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.231, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[54] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[55] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.343, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[56] Tiantian Wang, Lihe Zhang, Shuo Wang, Huchuan Lu, Gang Yang, Xiang Ruan, and Ali Borji. Detect globally, refine locally: A novel approach to saliency detection. In CVPR, pages 3127-3135, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[57] Wenguan Wang, Jianbing Shen, Xingping Dong, Ali Borji, and Ruigang Yang. Inferring salient objects from human fixations. IEEE TPAMI, 42(8):1913-1927, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[58] Jun Wei, Shuhui Wang, and Qingming Huang. \\(\\mathbf{F}^3\\) net: fusion, feedback and focus for salient object detection. In AAAI, pages 12321-12328, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[59] Runmin Wu, Mengyang Feng, Wenlong Guan, Dong Wang, Huchuan Lu, and Errui Ding. A mutual learning method for salient object detection with intertwined multi-supervision. In CVPR, pages 8150-8159, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[60] Yu-Huan Wu, Yun Liu, Le Zhang, Ming-Ming Cheng, and Bo Ren. Edn: Salient object detection via extremely-downsampled network. IEEE TIP, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[61] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[62] Zhe Wu, Li Su, and Qingming Huang. Stacked cross refinement network for edge-aware salient object detection. In ICCV, pages 7264-7273, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[63] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[64] Siyuan Yan, Jing Zhang, and Nick Barnes. Transmission-guided bayesian generative model for smoke segmentation. In AAAI, pages 3009-3017, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[65] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173. IEEE, 2013. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[66] Greg Yang, Tony Duan, J Edward Hu, Hadi Salman, Ilya Razenshteyn, and Jerry Li. Randomized smoothing of all shapes and sizes. In ICML, pages 10693-10705. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[67] Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML, pages 609-616, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3344" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[68] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In ACM SIGKDD Int. Conf. Knowledge Disc. Data Min., pages 694-699, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[69] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[70] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In ICML. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[71] Jing Zhang, Jianwen Xie, Nick Barnes, and Ping Li. Learning generative vision transformer with energy-based latent space for saliency prediction. In NeurIPS, pages 15448-15463, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.333 + ], + "angle": 0, + "content": "[72] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, pages 1741-1750, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.468, + 0.376 + ], + "angle": 0, + "content": "[73] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, and James Zou. When and how mixup improves calibration. In ICML, pages 26135-26160. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[74] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In ICCV, pages 202-211, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[75] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Hutchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[76] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[77] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, pages 3085-3094, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.468, + 0.643 + ], + "angle": 0, + "content": "[78] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[79] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "3345" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_origin.pdf b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..979bf8fe181a955295fe0e0c116efe2504ce1cc2 --- /dev/null +++ b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/d8d13c75-1d94-496a-b79e-a2d1b272921f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e08a8795fb354bffdf43e763c296e3ded8233914a43a0331ce7869308784d158 +size 2397886 diff --git a/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/full.md b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8d9922a748ce57f1debadc4510283b7f4a5f8d8a --- /dev/null +++ b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/full.md @@ -0,0 +1,409 @@ +# Self-Calibrating Vicinal Risk Minimisation for Model Calibration + +Jiawei Liu, Changkun Ye, Ruikai Cui, Nick Barnes +The Australian National University +Canberra, Australia + +{jiawei.liu3, changkun.ye, ruikai.cui, nick.barnes}@anu.edu.au + +# Abstract + +Model calibration, measuring the alignment between the prediction accuracy and model confidence, is an important metric reflecting model trustworthiness. Existing dense binary classification methods, without proper regularisation of model confidence, are prone to being over-confident. To calibrate Deep Neural Networks (DNNs), we propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) that explores the vicinity space of labeled data, where vicinal images that are farther away from labeled images adopt the groundtruth label with decreasing label confidence. We prove that in the logistic regression problem, SCVRM can be seen as a Vicinal Risk Minimisation plus a regularisation term that penalises the over-confident predictions. In practical implementation, SCVRM is approximated using Monte Carlo sampling that samples additional augmented training images and labels from the vicinal distributions. Experimental results demonstrate that SCVRM can significantly enhance model calibration for different dense classification tasks on both in-distribution and out-of-distribution data. Code is available at https://github.com/Carlisle-Liu/SCVRM. + +# 1. Introduction + +Binary dense classification tasks [10, 54, 64] have advanced significantly since the debut of Deep Neural Networks (DNNs) associated with complex network architectures and large numbers of trainable parameters. Increasing model complexity has been shown to negatively impact model calibration [14] which has remained under investigated in the binary dense classification [29-31]. Miscalibration, a mis-alignment between model confidence and prediction accuracy [38], is an undesirable qualify that hinders the deployment of DNNs, especially in safety critical applications. In this work, we study the model calibration for dense binary classification models. + +The most commonly observed mis-calibration problem is over-confidence, where the model confidence is signifi- + +![](images/e81e05f8b4d1f7983d14bd087029ca2986607db870f244ea29fb2da5a640a2ac.jpg) +Figure 1. An illustration of Mixup [69] and our proposed SCVRM. Circles with solid boundary $\{(x_i, y_i)\}_{i=1}^4$ are labeled training images. Vicinal images of Mixup (circles with dashed boundary) are only distributed along the vectors connecting the labeled images. On the other hand, in SCVRM, the label of vicinal image adopts the groundtruth category (shown in different colours) of the closest labeled image, but with label confidence (shown in colour intensity) reduced monotonically with increasing Euclidean distance between vicinal and labeled images. The example sample $\otimes$ , with Mixup/SCVRM vicinal image and augmented label shown on the right, is selected at the same relative position to $(x_4, y_4)$ . + +cantly higher than the accuracy of its predictions on a cohort of samples [14]. Increasing research interest has been dedicated to study model calibration methods that regularise the probability associated with the prediction to be meaningful in reflecting the chance of the prediction being correct. The existing methods can be roughly categorised into three groups: (i) training objective based [2, 8, 13, 21, 23, 38, 45]; (ii) post-processing based [14, 19, 47]; and (iii) label augmentation based [31, 39, 43]. The first two categories focus on penalising over-confident predictions while the third cat + +egory directly moderates the confidence of training labels for the DNN models to train on. They do not explore the image space in improving the model calibration. + +Vicinal Risk Minimisation (VRM) extends Empirical Risk Minimisation (ERM) by introducing a vicinal distribution around each labeled data in the image space [3, 52]. Mixup [69], a variant of VRM, assumes that vicinal images are only distributed along the vectors connecting the labeled data (pair of image and label) pairs. In addition, instead of letting vicinal images adopt the hard groundtruth label of the nearest labeled image, Mixup assigns smoothed versions of nearest groundtruth labels to the vicinal images. In practice, it samples augmented data through the convex combination of labeled data pairs: $(\tilde{x} = \lambda x_{i} + (1 - \lambda)x_{i},\tilde{y} = \lambda y_{i} + (1 - \lambda)y_{j})$ , where the combination factor is sampled from a beta distribution: $\lambda \in \mathrm{Beta}(\alpha ,\alpha)$ . Followup works [69, 73] show that Mixup is effective in improving model calibration in the image classification task. + +We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrate the DNNs. Different from Mixup [69], the vicinal distribution of SCVRM can use any distribution that sufficiently covers the vicinity space with reasonable probability density. We consider the case where the vicinal distribution is assumed to be Gaussian with standard deviation being a random variable following a uniform distribution. In addition, the labels assigned to the vicinal images are softened versions of the groundtruth label associated with the labeled image at the distribution centre, where the strength of softening is proportional to the L2 distance between the vicinal and labeled images. As shown in Fig. 1, compared with Mixup [69], our proposed SCVRM has the following advantages: (1) the vicinity image space is not restricted by the pair-wise spatial relations of training images; (2) consistent label confidence across the pixels in the augmented label for dense prediction; and (3) defines confidence boundaries beyond the training distribution to better handle out-of-distribution samples. + +Given a labeled training image set, SCVRM can be understood by the following principals: (1) the human visual system is invariant to small random image changes [12] so that images undergoing up to a certain level of transformation should remain correctly classified [3, 20, 52]. This is also in line with adversarial robustness approaches that require images under small perturbations to maintain their classifications [7]. However, they are no longer labeled training data, so their label confidence should be slightly reduced. (2) it is understood from information theory [16] that we have no information at extreme distance where the label should assume a uniform categorical distribution with maximum entropy. Following the principal of VRM, we propose this transition should be accomplished by a smooth vicinal transition from the exact label for the labeled images to a uniform categorical distribution at large distance from + +the labeled data. (See Fig. 2 for examples.) + +We summarise our contributions as: (1) propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM), where the labels associated with vicinal images have reduced confidence with increasing Euclidean distance between vicinal and labelled images, to calibrate DNN models for the dense binary classification task; (2) we approximate the vicinal distribution with a Gaussian distribution whose standard deviation is a random variable following a uniform distribution, denoted as SCVRM-G; (3) we supply an example showing that SCVRM is equivalent to VRM and a regularisation term on prediction confidence under a simplified model; (4) We realise SCVRM as a data augmentation technique, where augmented data are sampled from the vicinal distributions with Monte Carlo methods. Finally, we show state-of-the-art calibration results for salient object detection (main paper); and camouflage object and smoke detection, and semantic segmentation (Supp. 12 and Supp. 13). + +# 2. Related Works + +Model Calibration: aims to minimise the distributional gap between model confidence and prediction accuracy. Existing methods propose to align confidence with accuracy through (i) post-processing techniques, e.g., Temperature Scaling (TS) [14, 19], Platt Scaling [41, 47], Dirichlet Scaling [22], Bayesian Binning [40, 67], Isotropic Regression [68], and Mix-and-Match [70]; and (ii) objective functions including Brier Loss [2, 8], Confidence Penalty [45], Maximum Mean Calibration Error (MMCE) [24], Soft Calibration Objective [21] and Focal Loss [13, 38]. These methods emphasise on suppressing over-confident predictions to alleviate the mis-calibration issue. + +Research efforts have also been dedicated to investigate the effect of data augmentation methods on the calibration degree. Label augmentation methods align the confidence distribution of the target label, to which the model predictions converge through optimisation, with the prediction accuracy distribution. Label Smoothing (LS) [39, 43] directly softens the target label probability, preventing the model from producing over-confident predictions. [31] presents an alternative label augmentation solution that stochastically perturbs the groundtruth label and aligns the confidence distribution of expected label with the prediction accuracy distribution. On the other hand, Mixup [69], which simultaneously augments both the input image and its corresponding groundtruth label conditioned on the relative position between training data, has also been demonstrated to improve the model calibration degree [51]. + +Vicinal Risk Minimisation: is proposed by [3, 52] to explore the vicinity of labeled data in hope of achieving a better approximation of the expected risk. The vicinity space is approximated with a vicinal distribution that can be either estimated with sufficient unlabeled data, or otherwise + +assumed to follow certain prior distribution, e.g. Gaussian distribution [3, 52]. Mixup [69] is an extension of VRM that restricts the vicinal data to be distributed along the vectors connecting the data pairs through convex combination. Further, Mixup applies the convex combination in both image and label spaces, resulting in the vicinal samples having a softened version of the groundtruth label associated with the nearest labeled image. VRM has also been applied in randomised smoothing [7, 25, 66] to achieve certified adversarial robustness, where a Gaussian kernel with fixed variance is employed to approximate the vicinal distribution. + +Salient Object Detection: is a dense binary classification task. Conventional methods primarily depend on hard-crafted features to identify salient objects [5, 17, 18, 44]. Early DNN based approaches employ learned features of local regions, e.g., super-pixel, object proposal and image patch, which need to be further sequentially processed [15, 26, 53]. The advent of Fully Convolutional Network (FCN) shifted the research focus onto the network architecture design for better pixel-wise predictions. Multi-level feature aggregation takes full advantage of spatial cues imbued in low-level features and semantic information embodied in high-level features [42, 72, 78]. Further, attention mechanisms, including spatial attention, channel attention, etc., have been leveraged to explore intra- and inter-feature map correlations [33, 46, 77], with some works directly adopting transformer backbones with self-attention [50, 71]. A different direction investigates auxiliary cues, e.g. boundary and edge [59, 62], fixation [57]. + +# 3. Preliminary + +# 3.1. Settings and Notations + +This papers deals primarily with dense binary classification over the image space $\mathbb{R}^d$ with the corresponding label set of $\mathcal{V}\in \{0,1\} ^d$ , where $d$ is the dimension of image/label space, "0" and "1" represent the two categories. $\mathcal{V}_s\in [0,1]^d$ denotes a soft label space, clearly we have $\mathcal{V}\subset \mathcal{V}_s$ . Let $\mathcal{D} = \{(x_i,y_i)\}_{i = 1}^N$ be a finite dataset with image and label pairs $(x_{i},y_{i})$ sampled i.i.d. from the joint distribution $p(x,y)$ defined on $\mathcal{X}\times \mathcal{Y}$ . The task is to obtain an optimal classifier $f^{*}\in \mathcal{H}$ in the Hypothesis space $\mathcal{H}\subset \{f:\mathcal{X}\to \mathcal{Y}\}$ that maximises the prediction accuracy and model calibration performance. The dimensionality or spatial index is omitted for simplicity wherever it is clear. + +The model calibration degree over the joint distribution $p(x,y)$ can be evaluated in terms of Expected Calibration Error (ECE) [14, 23]. Let $S(x) \in (0,1)$ be the Sigmoid-activation value before classification: $f(x) = \mathbb{1}(S(x) > 0.5)$ , where $\mathbb{1}(\cdot)$ is an indicator function. The prediction confidence $c$ and accuracy $a$ can be defined as: $c = |S(x) - 0.5| + 0.5$ , and $a = \mathbb{1}(f(x) = y)$ respectively. We further use $p_{f,\mathcal{D}}(c,a)$ to denote the joint dis + +ribution of prediction confidence and prediction accuracy of model, $f(\cdot)$ , on dataset, $\mathcal{D}$ . Then, ECE can be defined as: $\mathrm{ECE}(p_{f,\mathcal{D}}) = \mathbb{E}_{p_f,\mathcal{D}(c)}\bigl [|\mathbb{E}_{p_f,\mathcal{D}(a|c)}[a] - c|\bigr ]$ , where $\mathbb{P}_{f,\mathcal{D}}(c)$ is a marginal distribution on prediction confidence, and $\mathbb{E}_{\mathbb{P}_{\theta ,\mathcal{D}}(a|c)}$ is a conditional distribution of prediction accuracy. For a well calibrated model, we should have: $p_{f,\mathcal{D}}(f(x)_i = y_i|c = r) = r$ $\forall r\in [0,1]$ + +# 3.2. Vicinal Risk Minimisation + +We may formulate a learning problem as searching for model $f \in \mathcal{H}$ that minimises expected risk $R(f)$ over loss function $\ell(f(x), y)$ , which can be written as: + +$$ +R (f) = \int_ {\mathcal {X} \times \mathcal {Y}} \ell (f (x), y) \cdot p (x, y) d x d y. \tag {1} +$$ + +Given a dataset $\mathcal{D} = \{(x_i, y_i)\}_i^N$ , the Empirical Risk Minimisation (ERM) approach approximates $R(f)$ by: + +$$ +\begin{array}{l} R _ {E} (f) = \int_ {\mathcal {X} \times \mathcal {Y}} \ell (f (x), y) \cdot \frac {1}{N} \delta_ {x _ {i}, y _ {i}} (x, y) d x d y \tag {2} \\ = \frac {1}{N} \sum_ {i = 1} ^ {N} \ell \left(f \left(x _ {i}\right), y _ {i}\right), \\ \end{array} +$$ + +where $\delta_{x_i,y_i}(x,y)$ is the Dirac delta distribution over $(x,y)$ that can only adopts the value $(x_i,y_i)$ . + +Based on ERM, the Vicinal Risk Minimisation (VRM) approach [3, 52] approximates $R(f)$ by + +$$ +R _ {V} (f) = \frac {1}{N} \sum_ {i = 1} ^ {N} \int \ell \left(f \left(\tilde {x} _ {i}\right), y _ {i}\right) p \left(\tilde {x} _ {i} \mid x _ {i}\right) d \tilde {x}, \tag {3} +$$ + +where $p(\tilde{x}_i|x_i)$ denotes the Probability Density Function (PDF) of vicinal images $\tilde{x}_i$ given labeled image $x_i$ . + +In practice, VRM proposes to use isotropic Gaussian distributions as a choice of $p(\tilde{x}_i|x_i)$ . Specifically, VRM defines $\tilde{x}_i = x_i + \epsilon, \epsilon \sim \mathcal{N}_d(0,\sigma^2 I_d)$ , where $I_{d}$ is a $d \times d$ identity matrix and the variance $\sigma^2$ is a fixed hyperparameter. The corresponding $R_V(f)$ can then be written as: + +$$ +R _ {V - G} (f; \sigma) = \frac {1}{N} \sum_ {i = 1} ^ {N} \int \ell \left(f \left(x _ {i} + \epsilon\right), y _ {i}\right) p (\epsilon) d \epsilon , \tag {4} +$$ + +where $p(\epsilon) = \frac{1}{\sqrt{(2\pi\sigma^2)^d}}\exp \left(-\frac{\|\epsilon\|_2^2}{2\sigma^2}\right)$ is the PDF of the additive Gaussian noise $\epsilon$ + +# 4. Method + +# 4.1. Self-Calibrating Vicinal Risk Minimisation + +Our method is based on the principals that samples in the proximity of labeled data should inherit their groundtruth labels with reduced label confidence, while samples in the + +![](images/af90f54dffbdbda5bdf230f2f940cbbcf20d33c3159b9bfa02dc27eb4e3b1865.jpg) +Figure 2. Examples of vicinal images and assigned labels under various standard deviation $\sigma$ values. $\sigma = 0$ denotes the labeled data $(x,y)$ . Vicinal images are generated via adding Gaussian noise $\epsilon \sim \mathcal{N}(0,\sigma^2 I_d)$ to the labeled image as defined in Eq. 7. Their assigned labels, being the softer versions of the groundtruth label $y$ , are computed with the Eq. 8 and Eq. 9, whose plot are shown on the right. + +extreme distance should have a near uniform categorical distribution for their labels. As an extension of VRM, we reduce label confidence for vicinal images based on the L2-distance between vicinal and labeled images. Extending the definition of VRM [3, 52], we propose a novel Self-Calibrating Vicinal Risk Minimisation (SCVRM): + +Definition 1. (SCVR) Let $\mathcal{D} = \{(x_i, y_i)\}_{i=1}^N$ . For $f \in \mathcal{H}$ and $\ell : f(\mathcal{X}) \times \mathcal{Y}_s \to \mathbb{R}$ , Self-Calibrating Vicinal Risk is defined as: + +$$ +R _ {S C} (f) = \frac {1}{N} \sum_ {i = 1} ^ {N} \int \ell \left(f \left(\tilde {x} _ {i}\right), g \left(y _ {i}, \| \tilde {x} _ {i} - x _ {i} \| _ {2}\right)\right) \times \tag {5} +$$ + +$$ +p (\tilde {x} _ {i} | x _ {i}) d \tilde {x} _ {i}, +$$ + +where $g:\mathcal{Y}\times \mathbb{R}_{+}\to \mathcal{Y}_{s}$ + +The difference between VRM (Eq. 3) and SCVRM lies in the labels associated with the vicinal images. VRM assigns the exact groundtruth label $y_{i}$ paired with labeled image $x_{i}$ to all its vicinal images $\tilde{x}_i$ , whereas we use softened versions of groundtruth label $g(y_{i},\| \tilde{x}_{i} - x_{i}\|_{2})$ , whose label confidence reduces with increases in the L2-distance between the vicinal and labeled images. Following VRM to represent the distribution of vicinal images conditioned on the labeled image $p(\tilde{x}|x)$ with an isotropic Gaussian distribution with a fixed variance, our SCVRM can be written as: + +$$ +R _ {S C - V} (f) = \frac {1}{N} \sum_ {i} ^ {N} \int \ell \left(f \left(x _ {i} + \epsilon\right), g \left(y _ {i}; \| \epsilon \| _ {2}\right)\right) p (\epsilon) d \epsilon , \tag {6} +$$ + +where L2-norm of Gaussian noise equals the L2-distance between vicinal and labeled images $\| \epsilon \| _2 = \| \tilde{x}_i - x_i\| _2$ + +We observe that the isotropic Gaussian noise $\epsilon \sim \mathcal{N}_d(0,\sigma^2 I_d)$ used in the VRM may face the problem of a bubbling effect, that results in the probability density of an isotropic Gaussian $\mathcal{N}_d(\mu ,\sigma^2 I_d)$ concentrating on a thin spherical shell centred on $\mu \in \mathbb{R}^d$ with radius $\sigma \sqrt{d}$ in high dimensional space $d\gg 1$ [1, 36]. Following this observation: (a) we propose to let the scale of the standard deviation of the isotropic Gaussian distribution be a random + +variable following a uniform distribution $\sigma \sim \mathcal{U}(0,\gamma]$ to prevent vicinal images $\tilde{x}$ from being distributed only near a hypersphere centring on $x_{i}$ ; and (b) we approximate the L2-distance between the vicinal and labeled images with a dummy variable $\| \tilde{\epsilon}\| _2\approx \| \epsilon \| _2$ which is set to $\| \tilde{\epsilon}\| _2 = \sigma \sqrt{d}$ (See Supp. 8.1 for experimental justifications). Our implementation of SCVRM can then be defined as: + +$$ +R _ {S C - G} (f) = \frac {1}{N} \sum_ {i} ^ {N} \int \ell \left(f \left(x _ {i} + \epsilon\right), g \left(y _ {i}; \| \tilde {\epsilon} \| _ {2}\right)\right) \times \tag {7} +$$ + +$$ +p (\epsilon) p (\sigma) d \epsilon d \sigma , +$$ + +where $p(\sigma) = 1 / \gamma$ is the PDF of the standard deviation $\sigma$ . + +There are many choices for the projection function from hard label space to soft label space. The most straightforward solution is to apply the label smoothing [39] function: $\mathrm{LS}(y,\beta) = (1 - \beta) + \frac{\beta}{K},\beta \in [0,1]$ where $K = 2$ denotes the number of classes in a binary task. Then our projection function can be defined as: + +$$ +g _ {\mathrm {L S}} (y; \varphi (\| \tilde {\epsilon} \| _ {2})) = \left(1 - \varphi (\| \tilde {\epsilon} \| _ {2})\right) y + \frac {\varphi (\| \tilde {\epsilon} \| _ {2})}{2}, \tag {8} +$$ + +where we define $\varphi (\cdot):\mathbb{R}_0^+\to [0,1)$ to be a Gaussian function $\varphi_G(\cdot ;\cdot)$ , that is formulated as: + +$$ +\varphi_ {G} \left(\| \tilde {\epsilon} \| _ {2}; \eta\right) = 1 - \exp \left(- \frac {\| \tilde {\epsilon} \| _ {2}}{\eta}\right), \tag {9} +$$ + +where $\eta$ is a hyperparameter that scales the value of $\| \tilde{\epsilon}\| _2$ , and we empirically set $\eta = \sqrt{d}$ . Fig. 2 illustrates the resultant groundtruth map assigned to vicinal images sampled under various standard deviations. Please note that the function $\varphi$ can be an arbitrary function that satisfies: $\varphi :\mathbb{R}_0^+ \to [0,1)$ . We investigate other options in Supp. 8.2. + +# 4.2. Connection between SCVRM and VRM + +We analyse the relationship between VRM and our proposed SCVRM in a simple setting, where we find mathematical connections between the two approaches. Specifically, we consider the case of binary classification tasks, where our $R_{SC - G}(f)$ in Eq. 7 adopts a single-layer logistic + +regression model and the binary cross entropy loss. In this setting, we prove that if $g(y_{i},\| \tilde{\epsilon}\|_{2})$ with $\| \tilde{\epsilon}\| _2 = \sigma \sqrt{d}$ is a label smoothing function, $R_{SC - G}(f)$ can be written as a combination of the Vicinal Risk $R_{V - G}(f;\sigma)$ in Eq. 4 and a regularization term that penalises overconfident predictions of the model $f$ (See proof in Supp. 7.1). + +Example 1. In $R_{SC - G}(f)$ defined in Eq. 7, suppose: + +- Data: $\{(x_i, y_i)\}_{i=1}^N$ with $x_i \in \mathbb{R}^d$ and $y_i \in \{0, 1\}$ , +- Model: $f(x) = 1 / (1 + e^{-w^T x})$ with $w \in \mathbb{R}^d$ , +- Loss: $\ell(f(x), y) = -y \log f(x) - (1 - y) \log (1 - f(x))$ , +- $g(y, \| \tilde{\epsilon} \|_2) = (1 - 2\sigma)y + \sigma, \sigma \sim \mathcal{U}(0, \gamma]$ and $\gamma \in (0, \frac{1}{2}]$ . then we have: + +$$ +R _ {S C - G} (f) = \int_ {0} ^ {\gamma} \frac {1}{\gamma} R _ {V - G} (f; \sigma) d \sigma + \tau (f), \tag {10} +$$ + +where + +$$ +\tau (f) = \frac {\gamma}{2 N} \sum_ {i = 1} ^ {N} \left(2 y _ {i} - 1\right) \cdot w ^ {T} x _ {i}, \tag {11} +$$ + +and the first term of the RHS in Eq. 10 is equivalent to introducing our design of $\sigma \sim \mathcal{U}(0,\gamma]$ into $R_{V - G}(f;\sigma)$ which is the VRM with Gaussian kernel defined in Eq. 4. + +Note that the label satisfies $y_{i} \in \{0,1\}$ , therefore in the term $\tau (f)$ in Eq. 11 we have: + +$$ +\left(2 y _ {i} - 1\right) \cdot w ^ {T} x _ {i} = \left\{ \begin{array}{l l} w ^ {T} x _ {i}, & y _ {i} = 1, \\ - w ^ {T} x _ {i}, & y _ {i} = 0. \end{array} \right. \tag {12} +$$ + +In this sense, the term $\tau(f)$ penalises high $w^T x_i$ when $y_i = 1$ and penalises low $w^T x_i$ when $y_i = 0$ . Since high $w^T x$ will lead to $f(x) = 1 / (1 + e^{w^T x})$ approaching 1 and low $w^T x$ will lead to $f(x)$ approaching 0, $\tau(f)$ actually penalizes overconfident predictions of the model $f$ . + +Remark 1. In this simplified case, minimising the Self-Calibrating Vicinal Risk defined in Eq. 7 equivalently minimises the Vicinal Risk defined in Eq. 4 incorporating our design of $\sigma \in \mathcal{U}(0,\gamma]$ and a regularisation term defined in Eq. 11, where the former improves the model classification accuracy the latter prevents the model from becoming overconfident on vicinal samples. + +# 4.3. Practical Model + +In practical implementation, we approximate the intractable SCVRM defined in Eq. 7 with the Monte-Carlo (MC) Sampling. More specifically, it is formulated as a data augmentation technique where, based on the labeled dataset + +$\mathcal{D} = \{(x_{1},y_{i})\}_{i = 1}^{N}$ , we sample an augmented dataset: + +$$ +\begin{array}{l} \mathcal {D} _ {s c} = \bigcup_ {i = 1} ^ {N} \left\{(\tilde {x} _ {i} ^ {j}, \tilde {y} _ {i} ^ {j}) \mid \tilde {x} _ {i} ^ {j} = x _ {i} + \epsilon_ {i} ^ {j}, \epsilon_ {i} ^ {j} \sim_ {i. i. d} \mathcal {N} (0, \sigma_ {i, j} ^ {2} I _ {d}), \right. \\ \sigma_ {i, j} \sim_ {i. i. d} \mathcal {U} (0, \gamma ], \\ \left. \tilde {y} _ {i} ^ {j} = g _ {\mathrm {L S}} \left(y _ {i}; \varphi_ {G} (\| \tilde {\epsilon} \| _ {2}; \eta)\right) \right\} _ {j = 1} ^ {M}, \tag {13} \\ \end{array} +$$ + +where $\tilde{x}_i^j$ is the $j^{th}$ augmented image sampled from the vicinity of the $i^{th}$ labeled image, $\tilde{y}_i^j$ is a smoothed version of the groundtruth label $y_i$ , $M$ is a hyperparameter being the number of augmented data from the vicinity of each labeled data, and we set $\| \tilde{\epsilon} \|_2 = \sigma \sqrt{d}$ to approximate the L2-distance between the vicinal and labeled images $\sigma \sqrt{d} \approx \| \epsilon \|_2$ which relieves the computational burden of computing an L2-norm in high-dimensional space. The model is trained on both the labeled dataset $\mathcal{D}$ and the augmented dataset $\mathcal{D}_{sc}$ using a Binary Cross Entropy loss: $\mathcal{L}_{bce}(f(x), y) = -y \log(f(x)) - (1 - y) \log(1 - f(x))$ . The training loss can be defined as: + +$$ +\mathcal {L} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\mathcal {L} _ {b c e} \left(f \left(x _ {i}\right), y _ {i}\right) + \sum_ {j = 1} ^ {M} \mathcal {L} _ {b c e} \left(f \left(\tilde {x} _ {i} ^ {j}\right), \tilde {y} _ {i} ^ {j}\right)\right). \tag {14} +$$ + +Note that the augmented dataset defined in Eq. 13 is resampled after each training epoch. + +# 5. Experiments and Results + +The main paper presents results on Salient Object Detection. Results for other dense classification tasks, e.g., Camouflaged Object Detection (binary) and Smoke Detection (binary), and semantic segmentation (multi-class) can be found in Supp. 12 and 13 respectively. + +# 5.1. Implementation Details + +Evaluation Metrics: We employ bin-based Expected Calibration Error (denoted $\mathrm{ECE}_{\mathrm{EW}}$ [14]) and Over-confidence Error $(\mathrm{OE}_{\mathrm{EW}}$ [51]) with $B = 10$ bins, to evaluate model calibration. (See Supp. 9.1 for implementation details.) + +Datasets: Following [31] we divide DUTS-TR [54] into a train $(\mathcal{D}_{\mathrm{TR}})$ and validation set $(\mathcal{D}_{\mathrm{VAL}})$ , 9,553 and 1,000 images respectively. We evaluate model calibration degree using SOD test datasets, DUTS-TE [54], DUT-OMRON [65], PASCAL-S [28], SOD [37], ECSSD [63], HKU-IS [26]. + +Model Architecture: The U-Net based model consists of a ResNet50 encoder and a decoder that are initialised with ImageNet pretrained weights and by default respectively. The implementation uses the Pytorch framework. Experiments with different encoders, e.g. VGG16 [49] and Swin transformer [34] are detailed in Supp. 11. + +Hyperparameters: Optimal calibration performance of model is achieved by setting $\gamma = 2$ , $\eta = \sqrt{d}$ and $M = 3$ . + +Table 1. Salient object detection model calibration benchmark. Results are evaluated with ${\mathrm{{ECE}}}_{\mathrm{{EW}}}$ and ${\mathrm{{OE}}}_{\mathrm{{EW}}}$ with 10 bins (units in %). + +
MethodsDUTS-TE [54]DUT-OMRON [65]PASCAL-S [28]SOD [37]ECSSD [63]HKU-IS [26]
CategoryNameYearECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓
SOD MethodsMSRNet [27]20172.572.343.323.163.443.236.426.140.970.940.920.87
SRM [55]20174.023.724.193.964.884.599.939.582.532.351.861.72
Amulet [74]20175.675.285.845.495.765.4310.039.592.562.421.981.87
BMPM [72]20183.743.524.524.374.884.688.167.931.951.891.581.53
DGRL [56]20184.123.864.414.215.014.778.448.202.132.021.631.53
PAGR [75]20184.043.795.144.965.645.3712.1711.872.842.701.621.54
PiCANet [33]20185.124.904.844.708.147.9210.5010.303.483.392.552.47
CPD [61]20193.973.784.204.065.375.179.659.392.292.191.991.90
BASNet [48]20195.004.864.934.836.506.3610.4010.272.742.702.302.26
EGNet [76]20193.333.143.663.505.425.198.047.791.981.881.471.40
AFNet [11]20193.953.744.254.095.064.848.158.022.382.271.871.78
PoolNet [32]20193.333.123.863.705.325.078.147.872.001.901.821.75
GCPANet [4]20203.182.993.993.844.163.977.056.881.611.541.271.21
MINet [42]20203.653.484.454.294.944.758.017.892.132.031.741.65
F3Met [58]20203.673.504.254.104.854.677.957.782.262.161.921.83
EBMGSD [71]20213.453.294.113.954.794.617.487.302.142.051.791.70
ICON [79]20212.892.763.843.714.083.956.706.551.561.491.381.32
PFSNet [35]20212.942.723.953.814.454.277.597.392.412.252.061.96
EDN [60]20223.623.474.023.904.894.748.818.662.202.131.651.58
Model Calibration MethodsBrier Loss [2]19502.772.583.553.383.903.706.406.161.371.301.040.99
Temperature Scaling [14]20172.532.343.183.033.563.366.326.050.960.930.830.70
MMCE [23]20182.862.673.563.414.003.816.856.631.411.351.181.13
Label Smoothing [39]20192.001.792.892.713.042.835.975.690.830.680.820.47
Mixup [51]20192.452.253.413.233.132.995.825.701.411.183.830.05
Focal Loss [38]20202.252.083.102.823.403.136.215.981.411.031.240.77
AdaFocal [13]20221.611.412.311.842.532.275.885.471.630.791.350.52
ASLP [31]20231.401.221.991.832.312.105.505.170.480.200.790.17
OursSCVRM20230.780.611.641.491.911.753.903.600.440.190.780.10
+ +**Optimisation Details:** The model is optimised for 30 epochs with an initial learning rate of $2.5 \times 10^{-5}$ that decays by a factor of 0.9 per epoch after the $10^{\text{th}}$ epoch. During each epoch, $M = 3$ augmented data are sampled from each vicinity space. Image size is set to $3 \times 384 \times 384$ and the batch size is 8. Basic data augmentation techniques including random flipping, translation and cropping are applied. + +# 5.2. Model Calibration Degree Performance + +The calibration degrees of existing SOD and model calibration methods, and our proposed SCVRM, evaluated in terms of $\mathrm{ECE}_{\mathrm{EW}}$ and $\mathrm{OE}_{\mathrm{EW}}$ , are presented in Tab. 1. It can be observed that SOD methods without regularisations on prediction confidence are, in general, less well calibrated than the models equipped with calibration methods. On the other hand, our proposed SCVRM achieves the lowest calibration errors across all six testing datasets among the model calibration methods. Largest improvements over the second-best model are obtained on DUTS-TE, DUTOMRON, PASCAL-S and SOD datasets, reducing the ECE by $44.3\%$ , $17.6\%$ , $17.3\%$ and $29.1\%$ respectively. There is little room for improvement on ECSSD and HKU-IS, where ASLP is already well calibrated. On these datasets, our proposed SCVRM maintains the high calibration degrees that are comparable with those of ASLP. + +Fig. 3 depicts the joint distribution of prediction confidence and prediction accuracy of some model calibration + +methods on the DUTS-TE [54] dataset. The joint distribution of better calibrated models become more closely aligned with the oracle line, which indicates a perfectly calibrated model. Our SCVRM has its joint distribution almost completely aligned with the oracle line. In comparison, Mixup is more over-confident with the high-density area of its joint distribution slightly to the bottom-right of the oracle and the low-density area deviating even more. This further demonstrates the effectiveness of exploring a continuous vicinal image space with softened labels. + +Table 2. Existing model calibration methods and our proposed LSR evaluated on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6] in terms of $\mathrm{ECE_{EW}}$ and $\mathrm{OE_{EW}}$ with 10 bins, and Accuracy (ACC). + +
MethodEvaluation (%)
ECE ↓OE ↓ACC ↑
Baseline52.3651.0541.88
Brier Loss [2]38.8537.1853.62
Temperature Scaling [14]51.9550.4641.59
Label Smoothing [39]37.2235.4855.41
MMCE [23]40.6439.6754.39
Mixup [51]31.0729.1058.71
Focal Loss [38]40.0138.4349.71
AdaFocal [13]27.5525.0755.39
ASLP [31]18.3116.3761.93
SCVRM11.938.2683.93
+ +![](images/75d40a443d9230b863265fce2aec65a702d3df8520ceead1bef048033897eafa.jpg) + +![](images/34690ba4109329e8870796193002f67926e9ea45dc6da557728ef5259b2836b8.jpg) +Label Smoothing + +![](images/a0063631d6ffe975c7f838625067b44c181bf26292f35b3814b432614bdabfe7.jpg) + +![](images/f0af0452b9afa5f5c7463836c67f9b093e74f987e9c626f9bc648dff63821d3b.jpg) +Mixup + +![](images/d96243020edbfbf5c14bf16471cdce86128e1b11d0fbc4f79bbf6bae7252b463.jpg) + +![](images/ec2c4acb75502d4332ecade00a976d02259070b961a444ebd9cb628323534911.jpg) +Focal Loss + +![](images/34b16694490a161de749c850e2d4875770bf05a854b8b92e1c41fb225c247b31.jpg) + +![](images/19f7c10abcebe2260b69d4a1955c76be862c4ae2c7c51f880586011a5e687f54.jpg) +AdaFocal + +![](images/b21974af960421bf50b412a41f2f84c39dfdeb5d23aec1f741dfbbdc94464422.jpg) + +![](images/713e02010444d53cf5f8bdbe53cb04a4b0eae8aff31f68a5b7c38e1f749c1748.jpg) +ASLP + +![](images/6dca3a226ecc5e10324e1ee1afc8d5436bdede99b11e931d76293f54acf39e44.jpg) + +![](images/202c649884647845e5028e1493cf92f79d8d66cadf023451391a9506f153c0f0.jpg) +SCVRM + +![](images/e3dd3d9bb05ed6b3591b33ff7ac8903d61f00f10812546d526d7cac7abce114b.jpg) +Figure 3. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on DUTS-TE [54]. The dashed red diagonal line represents the perfectly calibrated oracle model. +Figure 4. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6]. The dashed red diagonal line represents the perfectly calibrated oracle model. + +# 5.3. Model Calibration on Out-of-Distribution Data + +The calibrating ability of existing model calibration techniques and our proposed SCVRM on Out-of-Distribution (OoD) samples are evaluated on the 500 texture images [31] selected from the Describable Texture Dataset [6]. These texture images, not including any salient (foreground) object, demonstrates certain level of distributional shift from both training and testing distributions of SOD. As shown in Tab. 2, SCVRM achieves the lowest calibration errors on OoD data. Compared with the second-best model, it + +reduces ECE by $34.8\%$ and OE by $49.5\%$ . The improvements are also reflected in the joint distribution plot in Fig. 4, where the distribution area of SCVRM is significantly better aligned with the oracle line. Further, SCVRM significantly outperforms Mixup in terms of both calibration and classification. Its superior performance can be attributed to more effective utilisation of vicinity space, and a distance-based label augmentation technique that yields consistent label confidence across the pixels. + +# 5.4. Ablation Study + +We investigate the effect of SCVRM and its hyperparameters including $\gamma$ , $\eta$ and $M$ which adopt the default setting specified in Sec. 5.1 unless stated otherwise. + +Effect of SCVRM: We compare our SCVRM with ERM, the baseline model, VRM with a fixed standard deviation, and $\mathrm{VRM}^*$ with our design of $\sigma \sim \mathcal{U}(0,\gamma]$ (see implementation detail in Supp. 8.3). Tab. 3 shows that VRM (see implementation detail in Supp. 8.3) does not reduce calibration error over ERM in general, producing mixed impacts across the testing datasets. Introducing our design of $\sigma \sim \mathcal{U}(0,\gamma]$ also does not alleviate VRM's over-confidence. On the other hand, SCVRM significantly improves the model calibration over ERM, VRM and $\mathrm{VRM}^*$ . In addition, SCVRM also achieves improved classification accuracy than ERM (see Supp. 10.1), which can be attributed to more effective utilisation of vicinity space with our design of $\sigma \sim \mathcal{U}(0,\gamma]$ . It also enables $\mathrm{VRM}^*$ to achieve better classification accuracy compared to VRM (see Supp. 10.1). + +Effect of $\gamma$ : The hyperparameter $\gamma$ suggests the exploration radius around the training images. As illustrated in Fig. 5a, SCVRM can reach optimal calibration performance in a wide range $\gamma = [1, 10]$ . When $\gamma$ is too small, e.g. 0.5, SCVRM achieves limited improvements on model calibration as the vicinal images are too close to the labeled images, and their corresponding augmented labels retain rel + +Table 3. Ablation study on SCVRM. + +
MethodECE ↓ DUTS-TE [54]ECE ↓ DUT-OMRON [65]ECE ↓ PASCAL-S [28]
ERM (Baseline)3.052.893.803.684.143.98
VRM3.543.374.354.214.394.24
VRM*13.183.033.903.794.133.98
SCVRM0.780.611.641.491.911.75
MethodSOD [37]ECSSD [63]HKU-IS [26]
ERM (Baseline)7.076.851.821.761.381.34
VRM6.606.421.671.621.301.26
VRM*17.357.161.721.681.281.23
SCVRM3.903.600.440.190.780.10
+ +$^{1}$ VRM* incorporates our design of $\sigma \sim \mathcal{U}(0, \gamma]$ . + +atively high label confidences (see Fig. 2). As $\gamma$ grows, the augmented vicinal images can get sufficiently far away from the labeled images to set up label confidence contours of various levels and obtains optimal model calibration. + +Effect of $\eta$ : The hyperparameter $\eta$ is part of the Gaussian equation (Eq. 9) that affects the strength of smoothing factor. (See Supp. 8.4 for example vicinal data under different $\eta$ .) At $\eta = 0.1\sqrt{d}$ , the resultant augmented label is overly softened, leading to the trained model becoming under-confirmed1. When $\eta$ is too large, e.g. $5\sqrt{d}$ and $10.0\sqrt{d}$ , vicinal images all retain a relatively high label confidence, resulting in insufficient calibration regularisation. We find that $\eta$ works well in $[0.5\sqrt{d}, 2\sqrt{d}]$ , leading consistently to near-optimal calibration results. + +Effect of $M$ : We ablate the number of augmented images sampled from the vicinal distribution of each labeled image per training epoch, setting $M = \{1,2,3,4,5\}$ . Results in Fig. 5c show that SCVRM is not very sensitive to the number of augmented data sampled $M$ . + +# 5.5. Discussion + +We further demonstrate the effectiveness and generalisation ability of SCVRM in calibrating DNNs via additional experiments. The default training setting specified in Sec. 5.1 is adopted unless specified otherwise. + +Effectiveness in Multi-Class Dense Classification: SCVRM can also be generalised to semantic segmentation [9]. We demonstrate its effectiveness in calibrating multi-class dense classification model (See Supp. 13). + +Effectiveness in Other Binary Dense Classification Tasks: We verify the effectiveness of SCVRM in improving model calibration degree and dense classification accuracy of DNNs on additional binary dense classification tasks, such as Camouflaged Object Detection [10] and Smoke Detection [64] (See Supp. 12). + +Effectiveness with Different Base Models: SCVRM is also effective in calibrating different base models while improving their dense classification accuracy. across a range of base models. We show its compatibility with VGG16 [49] and Swin transformer [34] (See Supp. 11). + +![](images/44b8be54fed88d2862b51a1b37474efd13af95c926442cfbde667a56c74c69d6.jpg) +(a) $\gamma$ (horizontal axis). + +![](images/af82a35d29d59ba54592d502afba38240a671342bc6fee1380c48ac000e26fdc.jpg) +(b) $\eta$ (horizontal axis, and units in $\sqrt{d}$ ). + +![](images/e80b5bed574d74890f5569a8e5f0d65898ee1bdb8c87a0257c33628fed628f99.jpg) +(c) $M$ (horizontal axis). +Figure 5. $\mathrm{ECE}_{\mathrm{EW}}$ and $\mathrm{OE}_{\mathrm{EW}}$ scores on the DUTS-TE [54] testing dataset under different choices of hyperparameters: (a) $\gamma = \{0.5,1.0,2.0,3.0,5.0,10.0\}$ ; and (b) $\eta = \{i\sqrt{d} | i = 0.1,0.5,1.0,2.0,5.0,10.0,20.0\}$ ; and (c) $M = \{1,2,3,4,5\}$ . + +Effectiveness on Existing SOD Models: SCVRM can be utilised by various existing SOD models, e.g., EBMGSOD [71], EDN [60] and ICON [79] (See Supp.14). + +# 6. Conclusion + +We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrated DNNs via exploring the vicinity space of labeled data. Vicinal images adopt the groundtruth label of the labeled image at the centre of vicinal distribution, but with diminishing label confidence as they get farther away. In a simplified setting, SCVRM is proved equivalent to a Vicinal Risk Minimisation plus a regularisation term, where the former improves model classification accuracy and the later penalises over-confident predictions. In practice, SCVRM is implemented as a data augmentation technique where MC sampling is applied to sample augmented data from the vicinal distribution. Experimental results on various dense classification tasks demonstrate the effectiveness of SCVRM in improving not only model calibration, but also dense classification accuracy. We also thoroughly study its compatibility with different backbone models and existing methods. + +# Acknowledgement + +This research was in-part supported by the ANU-Optus Bushfire Research Center of Excellence. + +# References + +[1] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 4 +[2] GLENN W BRIER. Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1):1-3, 1950. 1, 2, 6 +[3] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 2, 3, 4 +[4] Zuyao Chen, Qianqian Xu, Runmin Cong, and Qingming Huang. Global context-aware progressive aggregation network for salient object detection. In AAAI, pages 10599-10606, 2020. 6 +[5] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 37(3):569-582, 2014. 3 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6, 7 +[7] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In ICML, pages 1310-1320. PMLR, 2019. 2, 3 +[8] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 2 +[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 88(2):303-338, 2010. 8 +[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 8 +[11] Mengyang Feng, Huchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In CVPR, pages 1623-1632, 2019. 6 +[12] Masoud Ghodrati, Amirhossein Farzmahdi, Karim Rajaei, Reza Ebrahimpour, and Seyed-Mahdi Khaligh-Razavi. Feedforward object-vision models only tolerate small image variations compared to human. Frontiers in computational neuroscience, 8:74, 2014. 2 +[13] Arindam Ghosh, Thomas Schaaf, and Matthew R. Gormley. Adafocal: Calibration-aware adaptive focal loss. In NeurIPS, 2022. 1, 2, 6 +[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, pages 1321-1330. PMLR, 2017. 1, 2, 3, 5, 6 +[15] Shengfeng He, Rynson WH Lau, Wenxi Liu, Zhe Huang, and Qingxiong Yang. Supercnn: A superpixelwise convolutional neural network for salient object detection. IJCV, 115:330-344, 2015. 3 +[16] Edwin T Jaynes. Information theory and statistical mechanics. Physical review, 106(4):620, 1957. 2 +[17] Huaizu Jiang, Jingdong Wang, Zejian Yuan, Yang Wu, Nanning Zheng, and Shipeng Li. Salient object detection: A discriminative regional feature integration approach. In CVPR, pages 2083-2090, 2013. 3 + +[18] Zhuolin Jiang and Larry S Davis. Submodular salient region detection. In CVPR, pages 2043-2050, 2013. 3 +[19] Tom Joy, Francesco Pinto, Ser-Nam Lim, Philip HS Torr, and Puneet K Dokania. Sample-dependent adaptive temperature scaling for improved calibration. In AAAI, pages 14919-14926, 2023. 1, 2 +[20] Christoph Kamann and Carsten Rother. Benchmarking the robustness of semantic segmentation models. In CVPR, pages 8828-8838, 2020. 2 +[21] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, pages 29768-29779, 2021. 1, 2 +[22] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. In NeurIPS, 2019. 2 +[23] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 1, 3, 6 +[24] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 2 +[25] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE symposium on security and privacy (SP), pages 656-672. IEEE, 2019. 3 +[26] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 3, 5, 6, 8 +[27] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, pages 2386-2395, 2017. 6 +[28] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 6, 8 +[29] Jiawei Liu, Jing Zhang, and Nick Barnes. Modeling aleatorc uncertainty for camouflaged object detection. In WACV, pages 1445-1454, 2022. 1 +[30] Jiawei Liu, Jing Zhang, Ruikai Cui, Kaihao Zhang, Weihao Li, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022. +[31] Jiawei Liu, Changkun Ye, Shan Wang, Ruikai Cui, Jing Zhang, Kaihao Zhang, and Nick Barnes. Model calibration in dense classification with adaptive label perturbation. In ICCV, pages 1173-1184, 2023. 1, 2, 5, 6, 7 +[32] Jiang-Jiang Liu, Qibin Hou, Zhi-Ang Liu, and Ming-Ming Cheng. Poolnet+: Exploring the potential of pooling for salient object detection. IEEE TPAMI, 45(1):887-904, 2022. 6 +[33] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 3, 6 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: + +Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 8 +[35] Mingcan Ma, Changqun Xia, and Jia Li. Pyramidal feature shrinking for salient object detection. In AAAI, pages 2311-2318, 2021. 6 +[36] David JC MacKay. Information theory, inference and learning algorithms. Cambridge university press, 2003. 4 +[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 5, 6, 8 +[38] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, pages 15288-15299, 2020. 1, 2, 6 +[39] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 2, 4, 6 +[40] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 2 +[41] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In ICML, pages 625-632, 2005. 2 +[42] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Hutchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 3, 6 +[43] Hyekang Park, Jongyoun Noh, Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Acls: Adaptive and conditional label smoothing for network calibration. In ICCV, pages 3936-3945, 2023. 1, 2 +[44] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740. IEEE, 2012. 3 +[45] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 2 +[46] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 3 +[47] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 1, 2 +[48] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 6 +[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5, 8 +[50] Avishek Siris, Jianbo Jiao, Gary K.L. Tam, Xianghua Xie, and Rynson W.H. Lau. Scene context-aware salient object detection. In ICCV, pages 4156-4166, 2021. 3 + +[51] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurlPS, 2019. 2, 5, 6 +[52] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 2, 3, 4 +[53] Lijun Wang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Deep networks for saliency detection via local estimation and global search. In CVPR, pages 3183-3192, 2015. 3 +[54] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 1, 5, 6, 7, 8 +[55] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 6 +[56] Tiantian Wang, Lihe Zhang, Shuo Wang, Huchuan Lu, Gang Yang, Xiang Ruan, and Ali Borji. Detect globally, refine locally: A novel approach to saliency detection. In CVPR, pages 3127-3135, 2018. 6 +[57] Wenguan Wang, Jianbing Shen, Xingping Dong, Ali Borji, and Ruigang Yang. Inferring salient objects from human fixations. IEEE TPAMI, 42(8):1913-1927, 2019. 3 +[58] Jun Wei, Shuhui Wang, and Qingming Huang. $\mathbf{F}^3$ net: fusion, feedback and focus for salient object detection. In AAAI, pages 12321-12328, 2020. 6 +[59] Runmin Wu, Mengyang Feng, Wenlong Guan, Dong Wang, Huchuan Lu, and Errui Ding. A mutual learning method for salient object detection with intertwined multi-supervision. In CVPR, pages 8150-8159, 2019. 3 +[60] Yu-Huan Wu, Yun Liu, Le Zhang, Ming-Ming Cheng, and Bo Ren. Edn: Salient object detection via extremely-downsampled network. IEEE TIP, 2022. 6, 8 +[61] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6 +[62] Zhe Wu, Li Su, and Qingming Huang. Stacked cross refinement network for edge-aware salient object detection. In ICCV, pages 7264-7273, 2019. 3 +[63] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 6, 8 +[64] Siyuan Yan, Jing Zhang, and Nick Barnes. Transmission-guided bayesian generative model for smoke segmentation. In AAAI, pages 3009-3017, 2022. 1, 8 +[65] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173. IEEE, 2013. 5, 6, 8 +[66] Greg Yang, Tony Duan, J Edward Hu, Hadi Salman, Ilya Razenshteyn, and Jerry Li. Randomized smoothing of all shapes and sizes. In ICML, pages 10693-10705. PMLR, 2020. 3 +[67] Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML, pages 609-616, 2001. 2 + +[68] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In ACM SIGKDD Int. Conf. Knowledge Disc. Data Min., pages 694-699, 2002. 2 +[69] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3 +[70] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In ICML. PMLR, 2020. 2 +[71] Jing Zhang, Jianwen Xie, Nick Barnes, and Ping Li. Learning generative vision transformer with energy-based latent space for saliency prediction. In NeurIPS, pages 15448-15463, 2021. 3, 6, 8 +[72] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, pages 1741-1750, 2018. 3, 6 +[73] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, and James Zou. When and how mixup improves calibration. In ICML, pages 26135-26160. PMLR, 2022. 2 +[74] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In ICCV, pages 202-211, 2017. 6 +[75] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Hutchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 6 +[76] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 6 +[77] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, pages 3085-3094, 2019. 3 +[78] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51. Springer, 2020. 3 +[79] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 6, 8 \ No newline at end of file diff --git a/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/images.zip b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0870b6fdb36eda9bf9d403f27ca1854bc2026f6e --- /dev/null +++ b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3acda6e1b92720c0fd60f29ff99be293e95d1997aeb4feb9d2e766aed109411b +size 590379 diff --git a/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/layout.json b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c86111f76e9821862bc3536d86b00e8e2a4bcc09 --- /dev/null +++ b/2024/Self-Calibrating Vicinal Risk Minimisation for Model Calibration/layout.json @@ -0,0 +1,12456 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 499, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 499, + 120 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 499, + 120 + ], + "type": "text", + "content": "Self-Calibrating Vicinal Risk Minimisation for Model Calibration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 143, + 419, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 143, + 419, + 184 + ], + "spans": [ + { + "bbox": [ + 173, + 143, + 419, + 184 + ], + "type": "text", + "content": "Jiawei Liu, Changkun Ye, Ruikai Cui, Nick Barnes \nThe Australian National University \nCanberra, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 187, + 462, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 187, + 462, + 199 + ], + "spans": [ + { + "bbox": [ + 128, + 187, + 462, + 199 + ], + "type": "text", + "content": "{jiawei.liu3, changkun.ye, ruikai.cui, nick.barnes}@anu.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 251, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 502 + ], + "type": "text", + "content": "Model calibration, measuring the alignment between the prediction accuracy and model confidence, is an important metric reflecting model trustworthiness. Existing dense binary classification methods, without proper regularisation of model confidence, are prone to being over-confident. To calibrate Deep Neural Networks (DNNs), we propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) that explores the vicinity space of labeled data, where vicinal images that are farther away from labeled images adopt the groundtruth label with decreasing label confidence. We prove that in the logistic regression problem, SCVRM can be seen as a Vicinal Risk Minimisation plus a regularisation term that penalises the over-confident predictions. In practical implementation, SCVRM is approximated using Monte Carlo sampling that samples additional augmented training images and labels from the vicinal distributions. Experimental results demonstrate that SCVRM can significantly enhance model calibration for different dense classification tasks on both in-distribution and out-of-distribution data. Code is available at https://github.com/Carlisle-Liu/SCVRM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 525, + 127, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 127, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 127, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "content": "Binary dense classification tasks [10, 54, 64] have advanced significantly since the debut of Deep Neural Networks (DNNs) associated with complex network architectures and large numbers of trainable parameters. Increasing model complexity has been shown to negatively impact model calibration [14] which has remained under investigated in the binary dense classification [29-31]. Miscalibration, a mis-alignment between model confidence and prediction accuracy [38], is an undesirable qualify that hinders the deployment of DNNs, especially in safety critical applications. In this work, we study the model calibration for dense binary classification models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "The most commonly observed mis-calibration problem is over-confidence, where the model confidence is signifi-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 228, + 545, + 451 + ], + "blocks": [ + { + "bbox": [ + 310, + 228, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 310, + 228, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 310, + 228, + 545, + 451 + ], + "type": "image", + "image_path": "e81e05f8b4d1f7983d14bd087029ca2986607db870f244ea29fb2da5a640a2ac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "type": "text", + "content": "Figure 1. An illustration of Mixup [69] and our proposed SCVRM. Circles with solid boundary " + }, + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\{(x_i, y_i)\\}_{i=1}^4" + }, + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "type": "text", + "content": " are labeled training images. Vicinal images of Mixup (circles with dashed boundary) are only distributed along the vectors connecting the labeled images. On the other hand, in SCVRM, the label of vicinal image adopts the groundtruth category (shown in different colours) of the closest labeled image, but with label confidence (shown in colour intensity) reduced monotonically with increasing Euclidean distance between vicinal and labeled images. The example sample " + }, + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "type": "text", + "content": ", with Mixup/SCVRM vicinal image and augmented label shown on the right, is selected at the same relative position to " + }, + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "type": "inline_equation", + "content": "(x_4, y_4)" + }, + { + "bbox": [ + 304, + 457, + 545, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "cantly higher than the accuracy of its predictions on a cohort of samples [14]. Increasing research interest has been dedicated to study model calibration methods that regularise the probability associated with the prediction to be meaningful in reflecting the chance of the prediction being correct. The existing methods can be roughly categorised into three groups: (i) training objective based [2, 8, 13, 21, 23, 38, 45]; (ii) post-processing based [14, 19, 47]; and (iii) label augmentation based [31, 39, 43]. The first two categories focus on penalising over-confident predictions while the third cat" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3335" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "egory directly moderates the confidence of training labels for the DNN models to train on. They do not explore the image space in improving the model calibration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 290 + ], + "type": "text", + "content": "Vicinal Risk Minimisation (VRM) extends Empirical Risk Minimisation (ERM) by introducing a vicinal distribution around each labeled data in the image space [3, 52]. Mixup [69], a variant of VRM, assumes that vicinal images are only distributed along the vectors connecting the labeled data (pair of image and label) pairs. In addition, instead of letting vicinal images adopt the hard groundtruth label of the nearest labeled image, Mixup assigns smoothed versions of nearest groundtruth labels to the vicinal images. In practice, it samples augmented data through the convex combination of labeled data pairs: " + }, + { + "bbox": [ + 46, + 110, + 287, + 290 + ], + "type": "inline_equation", + "content": "(\\tilde{x} = \\lambda x_{i} + (1 - \\lambda)x_{i},\\tilde{y} = \\lambda y_{i} + (1 - \\lambda)y_{j})" + }, + { + "bbox": [ + 46, + 110, + 287, + 290 + ], + "type": "text", + "content": " , where the combination factor is sampled from a beta distribution: " + }, + { + "bbox": [ + 46, + 110, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathrm{Beta}(\\alpha ,\\alpha)" + }, + { + "bbox": [ + 46, + 110, + 287, + 290 + ], + "type": "text", + "content": " . Followup works [69, 73] show that Mixup is effective in improving model calibration in the image classification task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 519 + ], + "type": "text", + "content": "We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrate the DNNs. Different from Mixup [69], the vicinal distribution of SCVRM can use any distribution that sufficiently covers the vicinity space with reasonable probability density. We consider the case where the vicinal distribution is assumed to be Gaussian with standard deviation being a random variable following a uniform distribution. In addition, the labels assigned to the vicinal images are softened versions of the groundtruth label associated with the labeled image at the distribution centre, where the strength of softening is proportional to the L2 distance between the vicinal and labeled images. As shown in Fig. 1, compared with Mixup [69], our proposed SCVRM has the following advantages: (1) the vicinity image space is not restricted by the pair-wise spatial relations of training images; (2) consistent label confidence across the pixels in the augmented label for dense prediction; and (3) defines confidence boundaries beyond the training distribution to better handle out-of-distribution samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": "Given a labeled training image set, SCVRM can be understood by the following principals: (1) the human visual system is invariant to small random image changes [12] so that images undergoing up to a certain level of transformation should remain correctly classified [3, 20, 52]. This is also in line with adversarial robustness approaches that require images under small perturbations to maintain their classifications [7]. However, they are no longer labeled training data, so their label confidence should be slightly reduced. (2) it is understood from information theory [16] that we have no information at extreme distance where the label should assume a uniform categorical distribution with maximum entropy. Following the principal of VRM, we propose this transition should be accomplished by a smooth vicinal transition from the exact label for the labeled images to a uniform categorical distribution at large distance from" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 481, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 481, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 481, + 84 + ], + "type": "text", + "content": "the labeled data. (See Fig. 2 for examples.)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 84, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 288 + ], + "type": "text", + "content": "We summarise our contributions as: (1) propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM), where the labels associated with vicinal images have reduced confidence with increasing Euclidean distance between vicinal and labelled images, to calibrate DNN models for the dense binary classification task; (2) we approximate the vicinal distribution with a Gaussian distribution whose standard deviation is a random variable following a uniform distribution, denoted as SCVRM-G; (3) we supply an example showing that SCVRM is equivalent to VRM and a regularisation term on prediction confidence under a simplified model; (4) We realise SCVRM as a data augmentation technique, where augmented data are sampled from the vicinal distributions with Monte Carlo methods. Finally, we show state-of-the-art calibration results for salient object detection (main paper); and camouflage object and smoke detection, and semantic segmentation (Supp. 12 and Supp. 13)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 298, + 397, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 397, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 397, + 311 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 318, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 461 + ], + "type": "text", + "content": "Model Calibration: aims to minimise the distributional gap between model confidence and prediction accuracy. Existing methods propose to align confidence with accuracy through (i) post-processing techniques, e.g., Temperature Scaling (TS) [14, 19], Platt Scaling [41, 47], Dirichlet Scaling [22], Bayesian Binning [40, 67], Isotropic Regression [68], and Mix-and-Match [70]; and (ii) objective functions including Brier Loss [2, 8], Confidence Penalty [45], Maximum Mean Calibration Error (MMCE) [24], Soft Calibration Objective [21] and Focal Loss [13, 38]. These methods emphasise on suppressing over-confident predictions to alleviate the mis-calibration issue." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "type": "text", + "content": "Research efforts have also been dedicated to investigate the effect of data augmentation methods on the calibration degree. Label augmentation methods align the confidence distribution of the target label, to which the model predictions converge through optimisation, with the prediction accuracy distribution. Label Smoothing (LS) [39, 43] directly softens the target label probability, preventing the model from producing over-confident predictions. [31] presents an alternative label augmentation solution that stochastically perturbs the groundtruth label and aligns the confidence distribution of expected label with the prediction accuracy distribution. On the other hand, Mixup [69], which simultaneously augments both the input image and its corresponding groundtruth label conditioned on the relative position between training data, has also been demonstrated to improve the model calibration degree [51]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Vicinal Risk Minimisation: is proposed by [3, 52] to explore the vicinity of labeled data in hope of achieving a better approximation of the expected risk. The vicinity space is approximated with a vicinal distribution that can be either estimated with sufficient unlabeled data, or otherwise" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3336" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "assumed to follow certain prior distribution, e.g. Gaussian distribution [3, 52]. Mixup [69] is an extension of VRM that restricts the vicinal data to be distributed along the vectors connecting the data pairs through convex combination. Further, Mixup applies the convex combination in both image and label spaces, resulting in the vicinal samples having a softened version of the groundtruth label associated with the nearest labeled image. VRM has also been applied in randomised smoothing [7, 25, 66] to achieve certified adversarial robustness, where a Gaussian kernel with fixed variance is employed to approximate the vicinal distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 421 + ], + "type": "text", + "content": "Salient Object Detection: is a dense binary classification task. Conventional methods primarily depend on hard-crafted features to identify salient objects [5, 17, 18, 44]. Early DNN based approaches employ learned features of local regions, e.g., super-pixel, object proposal and image patch, which need to be further sequentially processed [15, 26, 53]. The advent of Fully Convolutional Network (FCN) shifted the research focus onto the network architecture design for better pixel-wise predictions. Multi-level feature aggregation takes full advantage of spatial cues imbued in low-level features and semantic information embodied in high-level features [42, 72, 78]. Further, attention mechanisms, including spatial attention, channel attention, etc., have been leveraged to explore intra- and inter-feature map correlations [33, 46, 77], with some works directly adopting transformer backbones with self-attention [50, 71]. A different direction investigates auxiliary cues, e.g. boundary and edge [59, 62], fixation [57]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 434, + 124, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 124, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 124, + 447 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 175, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 175, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 175, + 466 + ], + "type": "text", + "content": "3.1. Settings and Notations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": "This papers deals primarily with dense binary classification over the image space " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " with the corresponding label set of " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\in \\{0,1\\} ^d" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " is the dimension of image/label space, \"0\" and \"1\" represent the two categories. " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_s\\in [0,1]^d" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " denotes a soft label space, clearly we have " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\subset \\mathcal{V}_s" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i,y_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " be a finite dataset with image and label pairs " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " sampled i.i.d. from the joint distribution " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "p(x,y)" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " defined on " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{X}\\times \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": ". The task is to obtain an optimal classifier " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "f^{*}\\in \\mathcal{H}" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " in the Hypothesis space " + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{H}\\subset \\{f:\\mathcal{X}\\to \\mathcal{Y}\\}" + }, + { + "bbox": [ + 46, + 473, + 287, + 617 + ], + "type": "text", + "content": " that maximises the prediction accuracy and model calibration performance. The dimensionality or spatial index is omitted for simplicity wherever it is clear." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "The model calibration degree over the joint distribution " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "p(x,y)" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " can be evaluated in terms of Expected Calibration Error (ECE) [14, 23]. Let " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "S(x) \\in (0,1)" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " be the Sigmoid-activation value before classification: " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "f(x) = \\mathbb{1}(S(x) > 0.5)" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is an indicator function. The prediction confidence " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " and accuracy " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " can be defined as: " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "c = |S(x) - 0.5| + 0.5" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "a = \\mathbb{1}(f(x) = y)" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " respectively. We further use " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "p_{f,\\mathcal{D}}(c,a)" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " to denote the joint dis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": "ribution of prediction confidence and prediction accuracy of model, " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " , on dataset, " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " . Then, ECE can be defined as: " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{ECE}(p_{f,\\mathcal{D}}) = \\mathbb{E}_{p_f,\\mathcal{D}(c)}\\bigl [|\\mathbb{E}_{p_f,\\mathcal{D}(a|c)}[a] - c|\\bigr ]" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{f,\\mathcal{D}}(c)" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " is a marginal distribution on prediction confidence, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbb{P}_{\\theta ,\\mathcal{D}}(a|c)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " is a conditional distribution of prediction accuracy. For a well calibrated model, we should have: " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "p_{f,\\mathcal{D}}(f(x)_i = y_i|c = r) = r" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\forall r\\in [0,1]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 163, + 451, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 163, + 451, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 451, + 175 + ], + "type": "text", + "content": "3.2. Vicinal Risk Minimisation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "text", + "content": "We may formulate a learning problem as searching for model " + }, + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{H}" + }, + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "text", + "content": " that minimises expected risk " + }, + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "inline_equation", + "content": "R(f)" + }, + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "text", + "content": " over loss function " + }, + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "inline_equation", + "content": "\\ell(f(x), y)" + }, + { + "bbox": [ + 304, + 182, + 545, + 218 + ], + "type": "text", + "content": ", which can be written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 226, + 545, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 226, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 343, + 226, + 545, + 253 + ], + "type": "interline_equation", + "content": "R (f) = \\int_ {\\mathcal {X} \\times \\mathcal {Y}} \\ell (f (x), y) \\cdot p (x, y) d x d y. \\tag {1}", + "image_path": "da94fc78b34768609ceef767fa1ee0c23eaab354947fcf065b5489fa15ea5a79.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 261, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 286 + ], + "type": "text", + "content": "Given a dataset " + }, + { + "bbox": [ + 304, + 261, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, y_i)\\}_i^N" + }, + { + "bbox": [ + 304, + 261, + 545, + 286 + ], + "type": "text", + "content": ", the Empirical Risk Minimisation (ERM) approach approximates " + }, + { + "bbox": [ + 304, + 261, + 545, + 286 + ], + "type": "inline_equation", + "content": "R(f)" + }, + { + "bbox": [ + 304, + 261, + 545, + 286 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 294, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 294, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 321, + 294, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R _ {E} (f) = \\int_ {\\mathcal {X} \\times \\mathcal {Y}} \\ell (f (x), y) \\cdot \\frac {1}{N} \\delta_ {x _ {i}, y _ {i}} (x, y) d x d y \\tag {2} \\\\ = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell \\left(f \\left(x _ {i}\\right), y _ {i}\\right), \\\\ \\end{array}", + "image_path": "4ecd0ec955c852d4421ae4cbb52c93f649bc608af2736af58d6a1e4a18519abc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\delta_{x_i,y_i}(x,y)" + }, + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "text", + "content": " is the Dirac delta distribution over " + }, + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "text", + "content": " that can only adopts the value " + }, + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "inline_equation", + "content": "(x_i,y_i)" + }, + { + "bbox": [ + 304, + 363, + 545, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 388, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 412 + ], + "type": "text", + "content": "Based on ERM, the Vicinal Risk Minimisation (VRM) approach [3, 52] approximates " + }, + { + "bbox": [ + 304, + 388, + 545, + 412 + ], + "type": "inline_equation", + "content": "R(f)" + }, + { + "bbox": [ + 304, + 388, + 545, + 412 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 421, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 335, + 421, + 545, + 453 + ], + "type": "interline_equation", + "content": "R _ {V} (f) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(\\tilde {x} _ {i}\\right), y _ {i}\\right) p \\left(\\tilde {x} _ {i} \\mid x _ {i}\\right) d \\tilde {x}, \\tag {3}", + "image_path": "1cf5e3eddfa60a1de863fb4af37448fa3caf3aadca5687d749b0e4b496e67cad.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "inline_equation", + "content": "p(\\tilde{x}_i|x_i)" + }, + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "text", + "content": " denotes the Probability Density Function (PDF) of vicinal images " + }, + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "text", + "content": " given labeled image " + }, + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 462, + 545, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": "In practice, VRM proposes to use isotropic Gaussian distributions as a choice of " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "p(\\tilde{x}_i|x_i)" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": ". Specifically, VRM defines " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i = x_i + \\epsilon, \\epsilon \\sim \\mathcal{N}_d(0,\\sigma^2 I_d)" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "d \\times d" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " identity matrix and the variance " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " is a fixed hyperparameter. The corresponding " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "R_V(f)" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " can then be written as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 555, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 555, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 319, + 555, + 545, + 587 + ], + "type": "interline_equation", + "content": "R _ {V - G} (f; \\sigma) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), y _ {i}\\right) p (\\epsilon) d \\epsilon , \\tag {4}", + "image_path": "fe892a0a6801eb482a1d69c2d69f8f29a21d3b4037dc62d588a9037471bb7996.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 597, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 597, + 545, + 628 + ], + "type": "inline_equation", + "content": "p(\\epsilon) = \\frac{1}{\\sqrt{(2\\pi\\sigma^2)^d}}\\exp \\left(-\\frac{\\|\\epsilon\\|_2^2}{2\\sigma^2}\\right)" + }, + { + "bbox": [ + 304, + 597, + 545, + 628 + ], + "type": "text", + "content": " is the PDF of the additive Gaussian noise " + }, + { + "bbox": [ + 304, + 597, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 639, + 361, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 361, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 361, + 651 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 658, + 529, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 529, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 529, + 671 + ], + "type": "text", + "content": "4.1. Self-Calibrating Vicinal Risk Minimisation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Our method is based on the principals that samples in the proximity of labeled data should inherit their groundtruth labels with reduced label confidence, while samples in the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3337" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 81, + 547, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 81, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 81, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 81, + 547, + 173 + ], + "type": "image", + "image_path": "af90f54dffbdbda5bdf230f2f940cbbcf20d33c3159b9bfa02dc27eb4e3b1865.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "text", + "content": "Figure 2. Examples of vicinal images and assigned labels under various standard deviation " + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "text", + "content": " values. " + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\sigma = 0" + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "text", + "content": " denotes the labeled data " + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "text", + "content": ". Vicinal images are generated via adding Gaussian noise " + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\sigma^2 I_d)" + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "text", + "content": " to the labeled image as defined in Eq. 7. Their assigned labels, being the softer versions of the groundtruth label " + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 177, + 547, + 211 + ], + "type": "text", + "content": ", are computed with the Eq. 8 and Eq. 9, whose plot are shown on the right." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 289 + ], + "type": "text", + "content": "extreme distance should have a near uniform categorical distribution for their labels. As an extension of VRM, we reduce label confidence for vicinal images based on the L2-distance between vicinal and labeled images. Extending the definition of VRM [3, 52], we propose a novel Self-Calibrating Vicinal Risk Minimisation (SCVRM):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "text", + "content": "Definition 1. (SCVR) Let " + }, + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{H}" + }, + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "inline_equation", + "content": "\\ell : f(\\mathcal{X}) \\times \\mathcal{Y}_s \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 294, + 288, + 332 + ], + "type": "text", + "content": ", Self-Calibrating Vicinal Risk is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 339, + 287, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 287, + 371 + ], + "type": "interline_equation", + "content": "R _ {S C} (f) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\int \\ell \\left(f \\left(\\tilde {x} _ {i}\\right), g \\left(y _ {i}, \\| \\tilde {x} _ {i} - x _ {i} \\| _ {2}\\right)\\right) \\times \\tag {5}", + "image_path": "8fb9d0a03ff77cbb02c2c8132460894c5de21803996ecc22760ae70871323b73.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 374, + 196, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 374, + 196, + 387 + ], + "spans": [ + { + "bbox": [ + 143, + 374, + 196, + 387 + ], + "type": "interline_equation", + "content": "p (\\tilde {x} _ {i} | x _ {i}) d \\tilde {x} _ {i},", + "image_path": "38e4872c41813de2fa702ac3dd70d6c0de59b3aaeb653320f2a71b893dee83eb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 394, + 153, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 153, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 153, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 394, + 153, + 407 + ], + "type": "inline_equation", + "content": "g:\\mathcal{Y}\\times \\mathbb{R}_{+}\\to \\mathcal{Y}_{s}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "text", + "content": "The difference between VRM (Eq. 3) and SCVRM lies in the labels associated with the vicinal images. VRM assigns the exact groundtruth label " + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "text", + "content": " paired with labeled image " + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "text", + "content": " to all its vicinal images " + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "text", + "content": ", whereas we use softened versions of groundtruth label " + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "inline_equation", + "content": "g(y_{i},\\| \\tilde{x}_{i} - x_{i}\\|_{2})" + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "text", + "content": ", whose label confidence reduces with increases in the L2-distance between the vicinal and labeled images. Following VRM to represent the distribution of vicinal images conditioned on the labeled image " + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "inline_equation", + "content": "p(\\tilde{x}|x)" + }, + { + "bbox": [ + 46, + 413, + 287, + 534 + ], + "type": "text", + "content": " with an isotropic Gaussian distribution with a fixed variance, our SCVRM can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 551, + 287, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 551, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 49, + 551, + 287, + 593 + ], + "type": "interline_equation", + "content": "R _ {S C - V} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), g \\left(y _ {i}; \\| \\epsilon \\| _ {2}\\right)\\right) p (\\epsilon) d \\epsilon , \\tag {6}", + "image_path": "4afc70f0829e1c62665335933e6d337ad68148ebe8b9302a0d1ec23266405a42.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "text", + "content": "where L2-norm of Gaussian noise equals the L2-distance between vicinal and labeled images " + }, + { + "bbox": [ + 46, + 594, + 287, + 618 + ], + "type": "inline_equation", + "content": "\\| \\epsilon \\| _2 = \\| \\tilde{x}_i - x_i\\| _2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": "We observe that the isotropic Gaussian noise " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}_d(0,\\sigma^2 I_d)" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": " used in the VRM may face the problem of a bubbling effect, that results in the probability density of an isotropic Gaussian " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_d(\\mu ,\\sigma^2 I_d)" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": " concentrating on a thin spherical shell centred on " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": " with radius " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\sigma \\sqrt{d}" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": " in high dimensional space " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "d\\gg 1" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": " [1, 36]. Following this observation: (a) we propose to let the scale of the standard deviation of the isotropic Gaussian distribution be a random" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "text", + "content": "variable following a uniform distribution " + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\sigma \\sim \\mathcal{U}(0,\\gamma]" + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "text", + "content": " to prevent vicinal images " + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "text", + "content": " from being distributed only near a hypersphere centring on " + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "text", + "content": "; and (b) we approximate the L2-distance between the vicinal and labeled images with a dummy variable " + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\epsilon}\\| _2\\approx \\| \\epsilon \\| _2" + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "text", + "content": " which is set to " + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\epsilon}\\| _2 = \\sigma \\sqrt{d}" + }, + { + "bbox": [ + 304, + 216, + 547, + 300 + ], + "type": "text", + "content": " (See Supp. 8.1 for experimental justifications). Our implementation of SCVRM can then be defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 307, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 545, + 338 + ], + "type": "interline_equation", + "content": "R _ {S C - G} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} \\int \\ell \\left(f \\left(x _ {i} + \\epsilon\\right), g \\left(y _ {i}; \\| \\tilde {\\epsilon} \\| _ {2}\\right)\\right) \\times \\tag {7}", + "image_path": "9fe17b4f6b8704244cc5b0af8512e9ce2afdc4e4ad04f588629d75e07a42ba4c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 411, + 342, + 473, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 342, + 473, + 355 + ], + "spans": [ + { + "bbox": [ + 411, + 342, + 473, + 355 + ], + "type": "interline_equation", + "content": "p (\\epsilon) p (\\sigma) d \\epsilon d \\sigma ,", + "image_path": "0e0ff928198a37af3c455be70285fc1b78522a8e57db6b9299aa8e35b0c5ceee.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 363, + 541, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 363, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 363, + 541, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 363, + 541, + 376 + ], + "type": "inline_equation", + "content": "p(\\sigma) = 1 / \\gamma" + }, + { + "bbox": [ + 306, + 363, + 541, + 376 + ], + "type": "text", + "content": " is the PDF of the standard deviation " + }, + { + "bbox": [ + 306, + 363, + 541, + 376 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 306, + 363, + 541, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": "There are many choices for the projection function from hard label space to soft label space. The most straightforward solution is to apply the label smoothing [39] function: " + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{LS}(y,\\beta) = (1 - \\beta) + \\frac{\\beta}{K},\\beta \\in [0,1]" + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 376, + 545, + 449 + ], + "type": "text", + "content": " denotes the number of classes in a binary task. Then our projection function can be defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 456, + 545, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 545, + 480 + ], + "type": "interline_equation", + "content": "g _ {\\mathrm {L S}} (y; \\varphi (\\| \\tilde {\\epsilon} \\| _ {2})) = \\left(1 - \\varphi (\\| \\tilde {\\epsilon} \\| _ {2})\\right) y + \\frac {\\varphi (\\| \\tilde {\\epsilon} \\| _ {2})}{2}, \\tag {8}", + "image_path": "7ebe2ee7e08e11774fa1b5b8f95a13afae5e7ca955d406c78631d9ec5af34d14.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 488, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 513 + ], + "type": "text", + "content": "where we define " + }, + { + "bbox": [ + 304, + 488, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\varphi (\\cdot):\\mathbb{R}_0^+\\to [0,1)" + }, + { + "bbox": [ + 304, + 488, + 545, + 513 + ], + "type": "text", + "content": " to be a Gaussian function " + }, + { + "bbox": [ + 304, + 488, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\varphi_G(\\cdot ;\\cdot)" + }, + { + "bbox": [ + 304, + 488, + 545, + 513 + ], + "type": "text", + "content": ", that is formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 352, + 521, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 521, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 352, + 521, + 545, + 548 + ], + "type": "interline_equation", + "content": "\\varphi_ {G} \\left(\\| \\tilde {\\epsilon} \\| _ {2}; \\eta\\right) = 1 - \\exp \\left(- \\frac {\\| \\tilde {\\epsilon} \\| _ {2}}{\\eta}\\right), \\tag {9}", + "image_path": "77a4ffc84751d6c649b8e623df8e157730e9d3de29441421a57ce46893c9b777.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "text", + "content": " is a hyperparameter that scales the value of " + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\epsilon}\\| _2" + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "text", + "content": ", and we empirically set " + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\eta = \\sqrt{d}" + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "text", + "content": ". Fig. 2 illustrates the resultant groundtruth map assigned to vicinal images sampled under various standard deviations. Please note that the function " + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "text", + "content": " can be an arbitrary function that satisfies: " + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\varphi :\\mathbb{R}_0^+ \\to [0,1)" + }, + { + "bbox": [ + 304, + 555, + 545, + 629 + ], + "type": "text", + "content": ". We investigate other options in Supp. 8.2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 635, + 516, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 516, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 516, + 647 + ], + "type": "text", + "content": "4.2. Connection between SCVRM and VRM" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "We analyse the relationship between VRM and our proposed SCVRM in a simple setting, where we find mathematical connections between the two approaches. Specifically, we consider the case of binary classification tasks, where our " + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "inline_equation", + "content": "R_{SC - G}(f)" + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": " in Eq. 7 adopts a single-layer logistic" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3338" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "regression model and the binary cross entropy loss. In this setting, we prove that if " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "g(y_{i},\\| \\tilde{\\epsilon}\\|_{2})" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\epsilon}\\| _2 = \\sigma \\sqrt{d}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is a label smoothing function, " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "R_{SC - G}(f)" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " can be written as a combination of the Vicinal Risk " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "R_{V - G}(f;\\sigma)" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " in Eq. 4 and a regularization term that penalises overconfident predictions of the model " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " (See proof in Supp. 7.1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 158, + 266, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 266, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 266, + 170 + ], + "type": "text", + "content": "Example 1. In " + }, + { + "bbox": [ + 47, + 158, + 266, + 170 + ], + "type": "inline_equation", + "content": "R_{SC - G}(f)" + }, + { + "bbox": [ + 47, + 158, + 266, + 170 + ], + "type": "text", + "content": " defined in Eq. 7, suppose:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 171, + 286, + 234 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "text", + "content": "- Data: " + }, + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "inline_equation", + "content": "\\{(x_i, y_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "inline_equation", + "content": "y_i \\in \\{0, 1\\}" + }, + { + "bbox": [ + 47, + 171, + 260, + 183 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 184, + 241, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 241, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 241, + 196 + ], + "type": "text", + "content": "- Model: " + }, + { + "bbox": [ + 47, + 184, + 241, + 196 + ], + "type": "inline_equation", + "content": "f(x) = 1 / (1 + e^{-w^T x})" + }, + { + "bbox": [ + 47, + 184, + 241, + 196 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 184, + 241, + 196 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 184, + 241, + 196 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "type": "text", + "content": "- Loss: " + }, + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "type": "inline_equation", + "content": "\\ell(f(x), y) = -y \\log f(x) - (1 - y) \\log (1 - f(x))" + }, + { + "bbox": [ + 47, + 197, + 286, + 209 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 209, + 286, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 286, + 234 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 47, + 209, + 286, + 234 + ], + "type": "inline_equation", + "content": "g(y, \\| \\tilde{\\epsilon} \\|_2) = (1 - 2\\sigma)y + \\sigma, \\sigma \\sim \\mathcal{U}(0, \\gamma]" + }, + { + "bbox": [ + 47, + 209, + 286, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 209, + 286, + 234 + ], + "type": "inline_equation", + "content": "\\gamma \\in (0, \\frac{1}{2}]" + }, + { + "bbox": [ + 47, + 209, + 286, + 234 + ], + "type": "text", + "content": ". then we have:" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 63, + 245, + 287, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 245, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 63, + 245, + 287, + 271 + ], + "type": "interline_equation", + "content": "R _ {S C - G} (f) = \\int_ {0} ^ {\\gamma} \\frac {1}{\\gamma} R _ {V - G} (f; \\sigma) d \\sigma + \\tau (f), \\tag {10}", + "image_path": "0d4306a8529045a8ade66a7443ef214e24961babede956b7b39271005a790190.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 283, + 75, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 75, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 75, + 293 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 297, + 287, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 297, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 97, + 297, + 287, + 330 + ], + "type": "interline_equation", + "content": "\\tau (f) = \\frac {\\gamma}{2 N} \\sum_ {i = 1} ^ {N} \\left(2 y _ {i} - 1\\right) \\cdot w ^ {T} x _ {i}, \\tag {11}", + "image_path": "f29211f3cfeae816ba230155dea73ccd7b9491047ae6972b4633e38c8f4e957e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "content": "and the first term of the RHS in Eq. 10 is equivalent to introducing our design of " + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\sigma \\sim \\mathcal{U}(0,\\gamma]" + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "inline_equation", + "content": "R_{V - G}(f;\\sigma)" + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "content": " which is the VRM with Gaussian kernel defined in Eq. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "text", + "content": "Note that the label satisfies " + }, + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "inline_equation", + "content": "y_{i} \\in \\{0,1\\}" + }, + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "text", + "content": ", therefore in the term " + }, + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\tau (f)" + }, + { + "bbox": [ + 47, + 389, + 287, + 414 + ], + "type": "text", + "content": " in Eq. 11 we have:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 84, + 426, + 287, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 426, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 84, + 426, + 287, + 458 + ], + "type": "interline_equation", + "content": "\\left(2 y _ {i} - 1\\right) \\cdot w ^ {T} x _ {i} = \\left\\{ \\begin{array}{l l} w ^ {T} x _ {i}, & y _ {i} = 1, \\\\ - w ^ {T} x _ {i}, & y _ {i} = 0. \\end{array} \\right. \\tag {12}", + "image_path": "d49ecaa459f27055387192bf7a40dd7d61ba075515b4d8c2915c16a3c6f4308d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": "In this sense, the term " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\tau(f)" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " penalises high " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "w^T x_i" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "y_i = 1" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " and penalises low " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "w^T x_i" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "y_i = 0" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": ". Since high " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "w^T x" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " will lead to " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "f(x) = 1 / (1 + e^{w^T x})" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " approaching 1 and low " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "w^T x" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " will lead to " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " approaching 0, " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\tau(f)" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": " actually penalizes overconfident predictions of the model " + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 473, + 287, + 535 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 548, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 287, + 632 + ], + "type": "text", + "content": "Remark 1. In this simplified case, minimising the Self-Calibrating Vicinal Risk defined in Eq. 7 equivalently minimises the Vicinal Risk defined in Eq. 4 incorporating our design of " + }, + { + "bbox": [ + 47, + 548, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathcal{U}(0,\\gamma]" + }, + { + "bbox": [ + 47, + 548, + 287, + 632 + ], + "type": "text", + "content": " and a regularisation term defined in Eq. 11, where the former improves the model classification accuracy the latter prevents the model from becoming overconfident on vicinal samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 645, + 145, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 145, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 145, + 657 + ], + "type": "text", + "content": "4.3. Practical Model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "In practical implementation, we approximate the intractable SCVRM defined in Eq. 7 with the Monte-Carlo (MC) Sampling. More specifically, it is formulated as a data augmentation technique where, based on the labeled dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 72, + 523, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 523, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 523, + 85 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_{1},y_{i})\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 306, + 72, + 523, + 85 + ], + "type": "text", + "content": ", we sample an augmented dataset:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 91, + 545, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 91, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 91, + 545, + 175 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} _ {s c} = \\bigcup_ {i = 1} ^ {N} \\left\\{(\\tilde {x} _ {i} ^ {j}, \\tilde {y} _ {i} ^ {j}) \\mid \\tilde {x} _ {i} ^ {j} = x _ {i} + \\epsilon_ {i} ^ {j}, \\epsilon_ {i} ^ {j} \\sim_ {i. i. d} \\mathcal {N} (0, \\sigma_ {i, j} ^ {2} I _ {d}), \\right. \\\\ \\sigma_ {i, j} \\sim_ {i. i. d} \\mathcal {U} (0, \\gamma ], \\\\ \\left. \\tilde {y} _ {i} ^ {j} = g _ {\\mathrm {L S}} \\left(y _ {i}; \\varphi_ {G} (\\| \\tilde {\\epsilon} \\| _ {2}; \\eta)\\right) \\right\\} _ {j = 1} ^ {M}, \\tag {13} \\\\ \\end{array}", + "image_path": "86044338327eb53738bac961a25cd695eb273e4c0f3bc2cf0205bd44e9b6582c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^j" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " augmented image sampled from the vicinity of the " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " labeled image, " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i^j" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " is a smoothed version of the groundtruth label " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " is a hyperparameter being the number of augmented data from the vicinity of each labeled data, and we set " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\epsilon} \\|_2 = \\sigma \\sqrt{d}" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " to approximate the L2-distance between the vicinal and labeled images " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\sigma \\sqrt{d} \\approx \\| \\epsilon \\|_2" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " which relieves the computational burden of computing an L2-norm in high-dimensional space. The model is trained on both the labeled dataset " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " and the augmented dataset " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sc}" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": " using a Binary Cross Entropy loss: " + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bce}(f(x), y) = -y \\log(f(x)) - (1 - y) \\log(1 - f(x))" + }, + { + "bbox": [ + 304, + 175, + 545, + 321 + ], + "type": "text", + "content": ". The training loss can be defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 327, + 545, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 545, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {b c e} \\left(f \\left(x _ {i}\\right), y _ {i}\\right) + \\sum_ {j = 1} ^ {M} \\mathcal {L} _ {b c e} \\left(f \\left(\\tilde {x} _ {i} ^ {j}\\right), \\tilde {y} _ {i} ^ {j}\\right)\\right). \\tag {14}", + "image_path": "94e69cbcd80918ccd1812b38250d8ef74b1ad05b7deb48bb5c9bb60f01b44eab.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "type": "text", + "content": "Note that the augmented dataset defined in Eq. 13 is resampled after each training epoch." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 405, + 449, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 449, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 449, + 418 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 426, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 486 + ], + "type": "text", + "content": "The main paper presents results on Salient Object Detection. Results for other dense classification tasks, e.g., Camouflaged Object Detection (binary) and Smoke Detection (binary), and semantic segmentation (multi-class) can be found in Supp. 12 and 13 respectively." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 491, + 440, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 440, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 440, + 504 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": "Evaluation Metrics: We employ bin-based Expected Calibration Error (denoted " + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{ECE}_{\\mathrm{EW}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": " [14]) and Over-confidence Error " + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "inline_equation", + "content": "(\\mathrm{OE}_{\\mathrm{EW}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": " [51]) with " + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "inline_equation", + "content": "B = 10" + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": " bins, to evaluate model calibration. (See Supp. 9.1 for implementation details.)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "content": "Datasets: Following [31] we divide DUTS-TR [54] into a train " + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{\\mathrm{TR}})" + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "content": " and validation set " + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{\\mathrm{VAL}})" + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "content": ", 9,553 and 1,000 images respectively. We evaluate model calibration degree using SOD test datasets, DUTS-TE [54], DUT-OMRON [65], PASCAL-S [28], SOD [37], ECSSD [63], HKU-IS [26]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "Model Architecture: The U-Net based model consists of a ResNet50 encoder and a decoder that are initialised with ImageNet pretrained weights and by default respectively. The implementation uses the Pytorch framework. Experiments with different encoders, e.g. VGG16 [49] and Swin transformer [34] are detailed in Supp. 11." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Hyperparameters: Optimal calibration performance of model is achieved by setting " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\eta = \\sqrt{d}" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3339" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 83, + 531, + 372 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 539, + 82 + ], + "type": "text", + "content": "Table 1. Salient object detection model calibration benchmark. Results are evaluated with " + }, + { + "bbox": [ + 53, + 71, + 539, + 82 + ], + "type": "inline_equation", + "content": "{\\mathrm{{ECE}}}_{\\mathrm{{EW}}}" + }, + { + "bbox": [ + 53, + 71, + 539, + 82 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 71, + 539, + 82 + ], + "type": "inline_equation", + "content": "{\\mathrm{{OE}}}_{\\mathrm{{EW}}}" + }, + { + "bbox": [ + 53, + 71, + 539, + 82 + ], + "type": "text", + "content": " with 10 bins (units in %)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 83, + 531, + 372 + ], + "lines": [ + { + "bbox": [ + 61, + 83, + 531, + 372 + ], + "spans": [ + { + "bbox": [ + 61, + 83, + 531, + 372 + ], + "type": "table", + "html": "
MethodsDUTS-TE [54]DUT-OMRON [65]PASCAL-S [28]SOD [37]ECSSD [63]HKU-IS [26]
CategoryNameYearECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓ECE ↓OE ↓
SOD MethodsMSRNet [27]20172.572.343.323.163.443.236.426.140.970.940.920.87
SRM [55]20174.023.724.193.964.884.599.939.582.532.351.861.72
Amulet [74]20175.675.285.845.495.765.4310.039.592.562.421.981.87
BMPM [72]20183.743.524.524.374.884.688.167.931.951.891.581.53
DGRL [56]20184.123.864.414.215.014.778.448.202.132.021.631.53
PAGR [75]20184.043.795.144.965.645.3712.1711.872.842.701.621.54
PiCANet [33]20185.124.904.844.708.147.9210.5010.303.483.392.552.47
CPD [61]20193.973.784.204.065.375.179.659.392.292.191.991.90
BASNet [48]20195.004.864.934.836.506.3610.4010.272.742.702.302.26
EGNet [76]20193.333.143.663.505.425.198.047.791.981.881.471.40
AFNet [11]20193.953.744.254.095.064.848.158.022.382.271.871.78
PoolNet [32]20193.333.123.863.705.325.078.147.872.001.901.821.75
GCPANet [4]20203.182.993.993.844.163.977.056.881.611.541.271.21
MINet [42]20203.653.484.454.294.944.758.017.892.132.031.741.65
F3Met [58]20203.673.504.254.104.854.677.957.782.262.161.921.83
EBMGSD [71]20213.453.294.113.954.794.617.487.302.142.051.791.70
ICON [79]20212.892.763.843.714.083.956.706.551.561.491.381.32
PFSNet [35]20212.942.723.953.814.454.277.597.392.412.252.061.96
EDN [60]20223.623.474.023.904.894.748.818.662.202.131.651.58
Model Calibration MethodsBrier Loss [2]19502.772.583.553.383.903.706.406.161.371.301.040.99
Temperature Scaling [14]20172.532.343.183.033.563.366.326.050.960.930.830.70
MMCE [23]20182.862.673.563.414.003.816.856.631.411.351.181.13
Label Smoothing [39]20192.001.792.892.713.042.835.975.690.830.680.820.47
Mixup [51]20192.452.253.413.233.132.995.825.701.411.183.830.05
Focal Loss [38]20202.252.083.102.823.403.136.215.981.411.031.240.77
AdaFocal [13]20221.611.412.311.842.532.275.885.471.630.791.350.52
ASLP [31]20231.401.221.991.832.312.105.505.170.480.200.790.17
OursSCVRM20230.780.611.641.491.911.753.903.600.440.190.780.10
", + "image_path": "025b00b06ec701f7939af55cba32fd927eb29f88209a76da9f6285dc3b1479fd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": "**Optimisation Details:** The model is optimised for 30 epochs with an initial learning rate of " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " that decays by a factor of 0.9 per epoch after the " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "10^{\\text{th}}" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " epoch. During each epoch, " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " augmented data are sampled from each vicinity space. Image size is set to " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "3 \\times 384 \\times 384" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " and the batch size is 8. Basic data augmentation techniques including random flipping, translation and cropping are applied." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 479, + 255, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 255, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 255, + 491 + ], + "type": "text", + "content": "5.2. Model Calibration Degree Performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "content": "The calibration degrees of existing SOD and model calibration methods, and our proposed SCVRM, evaluated in terms of " + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{ECE}_{\\mathrm{EW}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{OE}_{\\mathrm{EW}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "content": ", are presented in Tab. 1. It can be observed that SOD methods without regularisations on prediction confidence are, in general, less well calibrated than the models equipped with calibration methods. On the other hand, our proposed SCVRM achieves the lowest calibration errors across all six testing datasets among the model calibration methods. Largest improvements over the second-best model are obtained on DUTS-TE, DUTOMRON, PASCAL-S and SOD datasets, reducing the ECE by " + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "inline_equation", + "content": "44.3\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "inline_equation", + "content": "17.6\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "inline_equation", + "content": "17.3\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "inline_equation", + "content": "29.1\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 688 + ], + "type": "text", + "content": " respectively. There is little room for improvement on ECSSD and HKU-IS, where ASLP is already well calibrated. On these datasets, our proposed SCVRM maintains the high calibration degrees that are comparable with those of ASLP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Fig. 3 depicts the joint distribution of prediction confidence and prediction accuracy of some model calibration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 387, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 507 + ], + "type": "text", + "content": "methods on the DUTS-TE [54] dataset. The joint distribution of better calibrated models become more closely aligned with the oracle line, which indicates a perfectly calibrated model. Our SCVRM has its joint distribution almost completely aligned with the oracle line. In comparison, Mixup is more over-confident with the high-density area of its joint distribution slightly to the bottom-right of the oracle and the low-density area deviating even more. This further demonstrates the effectiveness of exploring a continuous vicinal image space with softened labels." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 565, + 524, + 708 + ], + "blocks": [ + { + "bbox": [ + 304, + 521, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 564 + ], + "type": "text", + "content": "Table 2. Existing model calibration methods and our proposed LSR evaluated on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6] in terms of " + }, + { + "bbox": [ + 304, + 521, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathrm{ECE_{EW}}" + }, + { + "bbox": [ + 304, + 521, + 545, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathrm{OE_{EW}}" + }, + { + "bbox": [ + 304, + 521, + 545, + 564 + ], + "type": "text", + "content": " with 10 bins, and Accuracy (ACC)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 565, + 524, + 708 + ], + "lines": [ + { + "bbox": [ + 326, + 565, + 524, + 708 + ], + "spans": [ + { + "bbox": [ + 326, + 565, + 524, + 708 + ], + "type": "table", + "html": "
MethodEvaluation (%)
ECE ↓OE ↓ACC ↑
Baseline52.3651.0541.88
Brier Loss [2]38.8537.1853.62
Temperature Scaling [14]51.9550.4641.59
Label Smoothing [39]37.2235.4855.41
MMCE [23]40.6439.6754.39
Mixup [51]31.0729.1058.71
Focal Loss [38]40.0138.4349.71
AdaFocal [13]27.5525.0755.39
ASLP [31]18.3116.3761.93
SCVRM11.938.2683.93
", + "image_path": "b3b4e0efa190b136c58b47eb3a945ac29ceabcacf009275d29315c89fcb30283.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3340" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 127, + 148 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 127, + 148 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 127, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 127, + 148 + ], + "type": "image", + "image_path": "75d40a443d9230b863265fce2aec65a702d3df8520ceead1bef048033897eafa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 148, + 125, + 222 + ], + "blocks": [ + { + "bbox": [ + 49, + 148, + 125, + 222 + ], + "lines": [ + { + "bbox": [ + 49, + 148, + 125, + 222 + ], + "spans": [ + { + "bbox": [ + 49, + 148, + 125, + 222 + ], + "type": "image", + "image_path": "34690ba4109329e8870796193002f67926e9ea45dc6da557728ef5259b2836b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 224, + 113, + 233 + ], + "lines": [ + { + "bbox": [ + 61, + 224, + 113, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 113, + 233 + ], + "type": "text", + "content": "Label Smoothing" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 205, + 148 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 205, + 148 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 205, + 148 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 205, + 148 + ], + "type": "image", + "image_path": "a0063631d6ffe975c7f838625067b44c181bf26292f35b3814b432614bdabfe7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 148, + 205, + 220 + ], + "blocks": [ + { + "bbox": [ + 129, + 148, + 205, + 220 + ], + "lines": [ + { + "bbox": [ + 129, + 148, + 205, + 220 + ], + "spans": [ + { + "bbox": [ + 129, + 148, + 205, + 220 + ], + "type": "image", + "image_path": "f0af0452b9afa5f5c7463836c67f9b093e74f987e9c626f9bc648dff63821d3b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 224, + 177, + 232 + ], + "lines": [ + { + "bbox": [ + 157, + 224, + 177, + 232 + ], + "spans": [ + { + "bbox": [ + 157, + 224, + 177, + 232 + ], + "type": "text", + "content": "Mixup" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 208, + 70, + 286, + 148 + ], + "blocks": [ + { + "bbox": [ + 208, + 70, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 208, + 70, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 208, + 70, + 286, + 148 + ], + "type": "image", + "image_path": "d96243020edbfbf5c14bf16471cdce86128e1b11d0fbc4f79bbf6bae7252b463.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 208, + 148, + 285, + 222 + ], + "blocks": [ + { + "bbox": [ + 208, + 148, + 285, + 222 + ], + "lines": [ + { + "bbox": [ + 208, + 148, + 285, + 222 + ], + "spans": [ + { + "bbox": [ + 208, + 148, + 285, + 222 + ], + "type": "image", + "image_path": "ec2c4acb75502d4332ecade00a976d02259070b961a444ebd9cb628323534911.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 224, + 264, + 232 + ], + "lines": [ + { + "bbox": [ + 229, + 224, + 264, + 232 + ], + "spans": [ + { + "bbox": [ + 229, + 224, + 264, + 232 + ], + "type": "text", + "content": "Focal Loss" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 287, + 70, + 365, + 148 + ], + "blocks": [ + { + "bbox": [ + 287, + 70, + 365, + 148 + ], + "lines": [ + { + "bbox": [ + 287, + 70, + 365, + 148 + ], + "spans": [ + { + "bbox": [ + 287, + 70, + 365, + 148 + ], + "type": "image", + "image_path": "34b16694490a161de749c850e2d4875770bf05a854b8b92e1c41fb225c247b31.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 287, + 148, + 364, + 222 + ], + "blocks": [ + { + "bbox": [ + 287, + 148, + 364, + 222 + ], + "lines": [ + { + "bbox": [ + 287, + 148, + 364, + 222 + ], + "spans": [ + { + "bbox": [ + 287, + 148, + 364, + 222 + ], + "type": "image", + "image_path": "19f7c10abcebe2260b69d4a1955c76be862c4ae2c7c51f880586011a5e687f54.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 224, + 341, + 232 + ], + "lines": [ + { + "bbox": [ + 310, + 224, + 341, + 232 + ], + "spans": [ + { + "bbox": [ + 310, + 224, + 341, + 232 + ], + "type": "text", + "content": "AdaFocal" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 366, + 70, + 444, + 148 + ], + "blocks": [ + { + "bbox": [ + 366, + 70, + 444, + 148 + ], + "lines": [ + { + "bbox": [ + 366, + 70, + 444, + 148 + ], + "spans": [ + { + "bbox": [ + 366, + 70, + 444, + 148 + ], + "type": "image", + "image_path": "b21974af960421bf50b412a41f2f84c39dfdeb5d23aec1f741dfbbdc94464422.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 367, + 148, + 443, + 221 + ], + "blocks": [ + { + "bbox": [ + 367, + 148, + 443, + 221 + ], + "lines": [ + { + "bbox": [ + 367, + 148, + 443, + 221 + ], + "spans": [ + { + "bbox": [ + 367, + 148, + 443, + 221 + ], + "type": "image", + "image_path": "713e02010444d53cf5f8bdbe53cb04a4b0eae8aff31f68a5b7c38e1f749c1748.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 224, + 415, + 232 + ], + "lines": [ + { + "bbox": [ + 396, + 224, + 415, + 232 + ], + "spans": [ + { + "bbox": [ + 396, + 224, + 415, + 232 + ], + "type": "text", + "content": "ASLP" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 446, + 70, + 524, + 148 + ], + "blocks": [ + { + "bbox": [ + 446, + 70, + 524, + 148 + ], + "lines": [ + { + "bbox": [ + 446, + 70, + 524, + 148 + ], + "spans": [ + { + "bbox": [ + 446, + 70, + 524, + 148 + ], + "type": "image", + "image_path": "6dca3a226ecc5e10324e1ee1afc8d5436bdede99b11e931d76293f54acf39e44.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 447, + 148, + 524, + 221 + ], + "blocks": [ + { + "bbox": [ + 447, + 148, + 524, + 221 + ], + "lines": [ + { + "bbox": [ + 447, + 148, + 524, + 221 + ], + "spans": [ + { + "bbox": [ + 447, + 148, + 524, + 221 + ], + "type": "image", + "image_path": "202c649884647845e5028e1493cf92f79d8d66cadf023451391a9506f153c0f0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 224, + 498, + 232 + ], + "lines": [ + { + "bbox": [ + 473, + 224, + 498, + 232 + ], + "spans": [ + { + "bbox": [ + 473, + 224, + 498, + 232 + ], + "type": "text", + "content": "SCVRM" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 50, + 270, + 284, + 514 + ], + "blocks": [ + { + "bbox": [ + 46, + 239, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 545, + 262 + ], + "type": "text", + "content": "Figure 3. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on DUTS-TE [54]. The dashed red diagonal line represents the perfectly calibrated oracle model." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 270, + 284, + 514 + ], + "lines": [ + { + "bbox": [ + 50, + 270, + 284, + 514 + ], + "spans": [ + { + "bbox": [ + 50, + 270, + 284, + 514 + ], + "type": "image", + "image_path": "e3dd3d9bb05ed6b3591b33ff7ac8903d61f00f10812546d526d7cac7abce114b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 516, + 287, + 571 + ], + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 571 + ], + "type": "text", + "content": "Figure 4. Joint distribution of prediction accuracy (vertical axis) and prediction confidence (horizontal axis) of model calibration methods on the 500 Out-of-Distribution texture images [31] selected from the Describable Texture Dataset [6]. The dashed red diagonal line represents the perfectly calibrated oracle model." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 585, + 287, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 287, + 597 + ], + "type": "text", + "content": "5.3. Model Calibration on Out-of-Distribution Data" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "The calibrating ability of existing model calibration techniques and our proposed SCVRM on Out-of-Distribution (OoD) samples are evaluated on the 500 texture images [31] selected from the Describable Texture Dataset [6]. These texture images, not including any salient (foreground) object, demonstrates certain level of distributional shift from both training and testing distributions of SOD. As shown in Tab. 2, SCVRM achieves the lowest calibration errors on OoD data. Compared with the second-best model, it" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 272, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 381 + ], + "type": "text", + "content": "reduces ECE by " + }, + { + "bbox": [ + 304, + 272, + 545, + 381 + ], + "type": "inline_equation", + "content": "34.8\\%" + }, + { + "bbox": [ + 304, + 272, + 545, + 381 + ], + "type": "text", + "content": " and OE by " + }, + { + "bbox": [ + 304, + 272, + 545, + 381 + ], + "type": "inline_equation", + "content": "49.5\\%" + }, + { + "bbox": [ + 304, + 272, + 545, + 381 + ], + "type": "text", + "content": ". The improvements are also reflected in the joint distribution plot in Fig. 4, where the distribution area of SCVRM is significantly better aligned with the oracle line. Further, SCVRM significantly outperforms Mixup in terms of both calibration and classification. Its superior performance can be attributed to more effective utilisation of vicinity space, and a distance-based label augmentation technique that yields consistent label confidence across the pixels." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 392, + 400, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 400, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 400, + 404 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "text", + "content": "We investigate the effect of SCVRM and its hyperparameters including " + }, + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 411, + 545, + 447 + ], + "type": "text", + "content": " which adopt the default setting specified in Sec. 5.1 unless stated otherwise." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": "Effect of SCVRM: We compare our SCVRM with ERM, the baseline model, VRM with a fixed standard deviation, and " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{VRM}^*" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " with our design of " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\sigma \\sim \\mathcal{U}(0,\\gamma]" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " (see implementation detail in Supp. 8.3). Tab. 3 shows that VRM (see implementation detail in Supp. 8.3) does not reduce calibration error over ERM in general, producing mixed impacts across the testing datasets. Introducing our design of " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\sigma \\sim \\mathcal{U}(0,\\gamma]" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " also does not alleviate VRM's over-confidence. On the other hand, SCVRM significantly improves the model calibration over ERM, VRM and " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{VRM}^*" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": ". In addition, SCVRM also achieves improved classification accuracy than ERM (see Supp. 10.1), which can be attributed to more effective utilisation of vicinity space with our design of " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\sigma \\sim \\mathcal{U}(0,\\gamma]" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": ". It also enables " + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{VRM}^*" + }, + { + "bbox": [ + 304, + 449, + 546, + 628 + ], + "type": "text", + "content": " to achieve better classification accuracy compared to VRM (see Supp. 10.1)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ": The hyperparameter " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " suggests the exploration radius around the training images. As illustrated in Fig. 5a, SCVRM can reach optimal calibration performance in a wide range " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = [1, 10]" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is too small, e.g. 0.5, SCVRM achieves limited improvements on model calibration as the vicinal images are too close to the labeled images, and their corresponding augmented labels retain rel" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3341" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 81, + 279, + 201 + ], + "blocks": [ + { + "bbox": [ + 100, + 71, + 234, + 81 + ], + "lines": [ + { + "bbox": [ + 100, + 71, + 234, + 81 + ], + "spans": [ + { + "bbox": [ + 100, + 71, + 234, + 81 + ], + "type": "text", + "content": "Table 3. Ablation study on SCVRM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 81, + 279, + 201 + ], + "lines": [ + { + "bbox": [ + 52, + 81, + 279, + 201 + ], + "spans": [ + { + "bbox": [ + 52, + 81, + 279, + 201 + ], + "type": "table", + "html": "
MethodECE ↓ DUTS-TE [54]ECE ↓ DUT-OMRON [65]ECE ↓ PASCAL-S [28]
ERM (Baseline)3.052.893.803.684.143.98
VRM3.543.374.354.214.394.24
VRM*13.183.033.903.794.133.98
SCVRM0.780.611.641.491.911.75
MethodSOD [37]ECSSD [63]HKU-IS [26]
ERM (Baseline)7.076.851.821.761.381.34
VRM6.606.421.671.621.301.26
VRM*17.357.161.721.681.281.23
SCVRM3.903.600.440.190.780.10
", + "image_path": "3957c7aff43246ea045e21e408b22d70f3022a4f61e9d3264ea043f46a1ec5ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 57, + 201, + 203, + 211 + ], + "lines": [ + { + "bbox": [ + 57, + 201, + 203, + 211 + ], + "spans": [ + { + "bbox": [ + 57, + 201, + 203, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 57, + 201, + 203, + 211 + ], + "type": "text", + "content": " VRM* incorporates our design of " + }, + { + "bbox": [ + 57, + 201, + 203, + 211 + ], + "type": "inline_equation", + "content": "\\sigma \\sim \\mathcal{U}(0, \\gamma]" + }, + { + "bbox": [ + 57, + 201, + 203, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "text", + "content": "atively high label confidences (see Fig. 2). As " + }, + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "text", + "content": " grows, the augmented vicinal images can get sufficiently far away from the labeled images to set up label confidence contours of various levels and obtains optimal model calibration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": ": The hyperparameter " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": " is part of the Gaussian equation (Eq. 9) that affects the strength of smoothing factor. (See Supp. 8.4 for example vicinal data under different " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": ".) At " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\eta = 0.1\\sqrt{d}" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": ", the resultant augmented label is overly softened, leading to the trained model becoming under-confirmed1. When " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": " is too large, e.g. " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "5\\sqrt{d}" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "10.0\\sqrt{d}" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": ", vicinal images all retain a relatively high label confidence, resulting in insufficient calibration regularisation. We find that " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": " works well in " + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "inline_equation", + "content": "[0.5\\sqrt{d}, 2\\sqrt{d}]" + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": ", leading consistently to near-optimal calibration results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "text", + "content": ": We ablate the number of augmented images sampled from the vicinal distribution of each labeled image per training epoch, setting " + }, + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "inline_equation", + "content": "M = \\{1,2,3,4,5\\}" + }, + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "text", + "content": ". Results in Fig. 5c show that SCVRM is not very sensitive to the number of augmented data sampled " + }, + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 383, + 287, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 450, + 119, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 119, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 119, + 460 + ], + "type": "text", + "content": "5.5. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 468, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 516 + ], + "type": "text", + "content": "We further demonstrate the effectiveness and generalisation ability of SCVRM in calibrating DNNs via additional experiments. The default training setting specified in Sec. 5.1 is adopted unless specified otherwise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 516, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 563 + ], + "type": "text", + "content": "Effectiveness in Multi-Class Dense Classification: SCVRM can also be generalised to semantic segmentation [9]. We demonstrate its effectiveness in calibrating multi-class dense classification model (See Supp. 13)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 563, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 635 + ], + "type": "text", + "content": "Effectiveness in Other Binary Dense Classification Tasks: We verify the effectiveness of SCVRM in improving model calibration degree and dense classification accuracy of DNNs on additional binary dense classification tasks, such as Camouflaged Object Detection [10] and Smoke Detection [64] (See Supp. 12)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "Effectiveness with Different Base Models: SCVRM is also effective in calibrating different base models while improving their dense classification accuracy. across a range of base models. We show its compatibility with VGG16 [49] and Swin transformer [34] (See Supp. 11)." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 539, + 150 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 539, + 150 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 539, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 539, + 150 + ], + "type": "image", + "image_path": "44b8be54fed88d2862b51a1b37474efd13af95c926442cfbde667a56c74c69d6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 150, + 463, + 159 + ], + "lines": [ + { + "bbox": [ + 390, + 150, + 463, + 159 + ], + "spans": [ + { + "bbox": [ + 390, + 150, + 463, + 159 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 390, + 150, + 463, + 159 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 390, + 150, + 463, + 159 + ], + "type": "text", + "content": " (horizontal axis)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 161, + 538, + 239 + ], + "blocks": [ + { + "bbox": [ + 313, + 161, + 538, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 161, + 538, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 538, + 239 + ], + "type": "image", + "image_path": "af82a35d29d59ba54592d502afba38240a671342bc6fee1380c48ac000e26fdc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 239, + 490, + 248 + ], + "lines": [ + { + "bbox": [ + 362, + 239, + 490, + 248 + ], + "spans": [ + { + "bbox": [ + 362, + 239, + 490, + 248 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 362, + 239, + 490, + 248 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 362, + 239, + 490, + 248 + ], + "type": "text", + "content": " (horizontal axis, and units in " + }, + { + "bbox": [ + 362, + 239, + 490, + 248 + ], + "type": "inline_equation", + "content": "\\sqrt{d}" + }, + { + "bbox": [ + 362, + 239, + 490, + 248 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 251, + 538, + 327 + ], + "blocks": [ + { + "bbox": [ + 317, + 251, + 538, + 327 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 538, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 538, + 327 + ], + "type": "image", + "image_path": "e80b5bed574d74890f5569a8e5f0d65898ee1bdb8c87a0257c33628fed628f99.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 327, + 465, + 337 + ], + "lines": [ + { + "bbox": [ + 387, + 327, + 465, + 337 + ], + "spans": [ + { + "bbox": [ + 387, + 327, + 465, + 337 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 387, + 327, + 465, + 337 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 387, + 327, + 465, + 337 + ], + "type": "text", + "content": " (horizontal axis)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "text", + "content": "Figure 5. " + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{ECE}_{\\mathrm{EW}}" + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{OE}_{\\mathrm{EW}}" + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "text", + "content": " scores on the DUTS-TE [54] testing dataset under different choices of hyperparameters: (a) " + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\gamma = \\{0.5,1.0,2.0,3.0,5.0,10.0\\}" + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "text", + "content": "; and (b) " + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\eta = \\{i\\sqrt{d} | i = 0.1,0.5,1.0,2.0,5.0,10.0,20.0\\}" + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "text", + "content": "; and (c) " + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "inline_equation", + "content": "M = \\{1,2,3,4,5\\}" + }, + { + "bbox": [ + 305, + 338, + 545, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "content": "Effectiveness on Existing SOD Models: SCVRM can be utilised by various existing SOD models, e.g., EBMGSOD [71], EDN [60] and ICON [79] (See Supp.14)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 455, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 658 + ], + "type": "text", + "content": "We propose a Self-Calibrating Vicinal Risk Minimisation (SCVRM) to calibrated DNNs via exploring the vicinity space of labeled data. Vicinal images adopt the groundtruth label of the labeled image at the centre of vicinal distribution, but with diminishing label confidence as they get farther away. In a simplified setting, SCVRM is proved equivalent to a Vicinal Risk Minimisation plus a regularisation term, where the former improves model classification accuracy and the later penalises over-confident predictions. In practice, SCVRM is implemented as a data augmentation technique where MC sampling is applied to sample augmented data from the vicinal distribution. Experimental results on various dense classification tasks demonstrate the effectiveness of SCVRM in improving not only model calibration, but also dense classification accuracy. We also thoroughly study its compatibility with different backbone models and existing methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 670, + 404, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 404, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 404, + 683 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "This research was in-part supported by the ANU-Optus Bushfire Research Center of Excellence." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 225, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 225, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 225, + 712 + ], + "type": "text", + "content": "1Under-confidence occurs when ECE - OE > OE." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3342" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "type": "text", + "content": "[2] GLENN W BRIER. Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1):1-3, 1950. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 181 + ], + "type": "text", + "content": "[3] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. Advances in neural information processing systems, 13, 2000. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 182, + 286, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 286, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 286, + 225 + ], + "type": "text", + "content": "[4] Zuyao Chen, Qianqian Xu, Runmin Cong, and Qingming Huang. Global context-aware progressive aggregation network for salient object detection. In AAAI, pages 10599-10606, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 227, + 286, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 286, + 259 + ], + "type": "text", + "content": "[5] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE TPAMI, 37(3):569-582, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 261, + 286, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 286, + 294 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 296, + 286, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 286, + 328 + ], + "type": "text", + "content": "[7] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In ICML, pages 1310-1320. PMLR, 2019. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "type": "text", + "content": "[8] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "type": "text", + "content": "[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 88(2):303-338, 2010. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 409, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 286, + 441 + ], + "type": "text", + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 1, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 443, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 286, + 475 + ], + "type": "text", + "content": "[11] Mengyang Feng, Huchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In CVPR, pages 1623-1632, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 477, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 286, + 531 + ], + "type": "text", + "content": "[12] Masoud Ghodrati, Amirhossein Farzmahdi, Karim Rajaei, Reza Ebrahimpour, and Seyed-Mahdi Khaligh-Razavi. Feedforward object-vision models only tolerate small image variations compared to human. Frontiers in computational neuroscience, 8:74, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 533, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 286, + 565 + ], + "type": "text", + "content": "[13] Arindam Ghosh, Thomas Schaaf, and Matthew R. Gormley. Adafocal: Calibration-aware adaptive focal loss. In NeurIPS, 2022. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "type": "text", + "content": "[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, pages 1321-1330. PMLR, 2017. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 601, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 643 + ], + "type": "text", + "content": "[15] Shengfeng He, Rynson WH Lau, Wenxi Liu, Zhe Huang, and Qingxiong Yang. Supercnn: A superpixelwise convolutional neural network for salient object detection. IJCV, 115:330-344, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 646, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 668 + ], + "type": "text", + "content": "[16] Edwin T Jaynes. Information theory and statistical mechanics. Physical review, 106(4):620, 1957. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[17] Huaizu Jiang, Jingdong Wang, Zejian Yuan, Yang Wu, Nanning Zheng, and Shipeng Li. Salient object detection: A discriminative regional feature integration approach. In CVPR, pages 2083-2090, 2013. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[18] Zhuolin Jiang and Larry S Davis. Submodular salient region detection. In CVPR, pages 2043-2050, 2013. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[19] Tom Joy, Francesco Pinto, Ser-Nam Lim, Philip HS Torr, and Puneet K Dokania. Sample-dependent adaptive temperature scaling for improved calibration. In AAAI, pages 14919-14926, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[20] Christoph Kamann and Carsten Rother. Benchmarking the robustness of semantic segmentation models. In CVPR, pages 8828-8838, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[21] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, pages 29768-29779, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[22] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "text", + "content": "[23] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[24] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, pages 2805-2814. PMLR, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 343, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 396 + ], + "type": "text", + "content": "[25] Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE symposium on security and privacy (SP), pages 656-672. IEEE, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "text", + "content": "[26] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, pages 5455-5463, 2015. 3, 5, 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "text", + "content": "[27] Guanbin Li, Yuan Xie, Liang Lin, and Yizhou Yu. Instance-level salient object segmentation. In CVPR, pages 2386-2395, 2017. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "text", + "content": "[28] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, pages 280-287, 2014. 5, 6, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 533 + ], + "type": "text", + "content": "[29] Jiawei Liu, Jing Zhang, and Nick Barnes. Modeling aleatorc uncertainty for camouflaged object detection. In WACV, pages 1445-1454, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "text", + "content": "[30] Jiawei Liu, Jing Zhang, Ruikai Cui, Kaihao Zhang, Weihao Li, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "text", + "content": "[31] Jiawei Liu, Changkun Ye, Shan Wang, Ruikai Cui, Jing Zhang, Kaihao Zhang, and Nick Barnes. Model calibration in dense classification with adaptive label perturbation. In ICCV, pages 1173-1184, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "text", + "content": "[32] Jiang-Jiang Liu, Qibin Hou, Zhi-Ang Liu, and Ming-Ming Cheng. Poolnet+: Exploring the potential of pooling for salient object detection. IEEE TPAMI, 45(1):887-904, 2022. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[33] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In CVPR, pages 3089-3098, 2018. 3, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3343" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 127 + ], + "type": "text", + "content": "[35] Mingcan Ma, Changqun Xia, and Jia Li. Pyramidal feature shrinking for salient object detection. In AAAI, pages 2311-2318, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 128, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 150 + ], + "type": "text", + "content": "[36] David JC MacKay. Information theory, inference and learning algorithms. Cambridge university press, 2003. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 151, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 217 + ], + "type": "text", + "content": "[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[38] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, pages 15288-15299, 2020. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "text", + "content": "[39] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 2, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 293, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 326 + ], + "type": "text", + "content": "[40] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "text", + "content": "[41] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In ICML, pages 625-632, 2005. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[42] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Hutchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[43] Hyekang Park, Jongyoun Noh, Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Acls: Adaptive and conditional label smoothing for network calibration. In ICCV, pages 3936-3945, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[44] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In CVPR, pages 733-740. IEEE, 2012. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "text", + "content": "[45] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "text", + "content": "[46] Yongri Piao, Wei Ji, Jingjing Li, Miao Zhang, and Huchuan Lu. Depth-induced multi-scale recurrent attention network for saliency detection. In ICCV, pages 7254-7263, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "text", + "content": "[47] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[48] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In CVPR, pages 7479-7489, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[50] Avishek Siris, Jianbo Jiao, Gary K.L. Tam, Xianghua Xie, and Rynson W.H. Lau. Scene context-aware salient object detection. In ICCV, pages 4156-4166, 2021. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[51] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurlPS, 2019. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "text", + "content": "[52] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[53] Lijun Wang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Deep networks for saliency detection via local estimation and global search. In CVPR, pages 3183-3192, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 182, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 182, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 182, + 545, + 227 + ], + "type": "text", + "content": "[54] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, pages 136-145, 2017. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[55] Tiantian Wang, Ali Borji, Lihe Zhang, Pingping Zhang, and Huchuan Lu. A stagewise refinement model for detecting salient objects in images. In ICCV, pages 4019-4028, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "type": "text", + "content": "[56] Tiantian Wang, Lihe Zhang, Shuo Wang, Huchuan Lu, Gang Yang, Xiang Ruan, and Ali Borji. Detect globally, refine locally: A novel approach to saliency detection. In CVPR, pages 3127-3135, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[57] Wenguan Wang, Jianbing Shen, Xingping Dong, Ali Borji, and Ruigang Yang. Inferring salient objects from human fixations. IEEE TPAMI, 42(8):1913-1927, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "text", + "content": "[58] Jun Wei, Shuhui Wang, and Qingming Huang. " + }, + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^3" + }, + { + "bbox": [ + 308, + 349, + 545, + 381 + ], + "type": "text", + "content": " net: fusion, feedback and focus for salient object detection. In AAAI, pages 12321-12328, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "text", + "content": "[59] Runmin Wu, Mengyang Feng, Wenlong Guan, Dong Wang, Huchuan Lu, and Errui Ding. A mutual learning method for salient object detection with intertwined multi-supervision. In CVPR, pages 8150-8159, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[60] Yu-Huan Wu, Yun Liu, Le Zhang, Ming-Ming Cheng, and Bo Ren. Edn: Salient object detection via extremely-downsampled network. IEEE TIP, 2022. 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[61] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "text", + "content": "[62] Zhe Wu, Li Su, and Qingming Huang. Stacked cross refinement network for edge-aware salient object detection. In ICCV, pages 7264-7273, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 525, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 558 + ], + "type": "text", + "content": "[63] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In CVPR, pages 1155-1162, 2013. 5, 6, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "text", + "content": "[64] Siyuan Yan, Jing Zhang, and Nick Barnes. Transmission-guided bayesian generative model for smoke segmentation. In AAAI, pages 3009-3017, 2022. 1, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[65] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, pages 3166-3173. IEEE, 2013. 5, 6, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[66] Greg Yang, Tony Duan, J Edward Hu, Hadi Salman, Ilya Razenshteyn, and Jerry Li. Randomized smoothing of all shapes and sizes. In ICML, pages 10693-10705. PMLR, 2020. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[67] Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML, pages 609-616, 2001. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3344" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 544 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[68] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In ACM SIGKDD Int. Conf. Knowledge Disc. Data Min., pages 694-699, 2002. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[69] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "text", + "content": "[70] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In ICML. PMLR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "text", + "content": "[71] Jing Zhang, Jianwen Xie, Nick Barnes, and Ping Li. Learning generative vision transformer with energy-based latent space for saliency prediction. In NeurIPS, pages 15448-15463, 2021. 3, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 263 + ], + "type": "text", + "content": "[72] Lu Zhang, Ju Dai, Huchuan Lu, You He, and Gang Wang. A bi-directional message passing model for salient object detection. In CVPR, pages 1741-1750, 2018. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 286, + 297 + ], + "type": "text", + "content": "[73] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, and James Zou. When and how mixup improves calibration. In ICML, pages 26135-26160. PMLR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 299, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 286, + 341 + ], + "type": "text", + "content": "[74] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In ICCV, pages 202-211, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 286, + 385 + ], + "type": "text", + "content": "[75] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Hutchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In CVPR, pages 714-722, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 388, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 430 + ], + "type": "text", + "content": "[76] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, pages 8779-8788, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 433, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 286, + 464 + ], + "type": "text", + "content": "[77] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In CVPR, pages 3085-3094, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 467, + 286, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 286, + 509 + ], + "type": "text", + "content": "[78] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In ECCV, pages 35-51. Springer, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 511, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 286, + 544 + ], + "type": "text", + "content": "[79] Mingchen Zhuge, Deng-Ping Fan, Nian Liu, Dingwen Zhang, Dong Xu, and Ling Shao. Salient object detection via integrity learning. IEEE TPAMI, 2022. 6, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3345" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_content_list.json b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cbfa6298c9251c6835147ca6b19b79d97eaa38b3 --- /dev/null +++ b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_content_list.json @@ -0,0 +1,1472 @@ +[ + { + "type": "text", + "text": "Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 109, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hang Li $^{1,4,5}$ Chengzhi Shen $^{3}$ Philip Torr $^{2}$ Volker Tresp $^{1,4}$ Jindong Gu $^{2*}$", + "bbox": [ + 125, + 202, + 844, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ LMU Munich, Germany $^{2}$ University of Oxford, UK $^{3}$ Technical University of Munich, Germany $^{4}$ Munich Center for Machine Learning, Germany $^{5}$ Siemens AG, Germany", + "bbox": [ + 89, + 220, + 880, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based models have gained significant popularity for text-to-image generation due to their exceptional image-generation capabilities. A risk with these models is the potential generation of inappropriate content, such as biased or harmful images. However, the underlying reasons for generating such undesired content from the perspective of the diffusion model's internal representation remain unclear. Previous work interprets vectors in an interpretable latent space of diffusion models as semantic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation. Extensive experiments have been conducted to verify the effectiveness of our mitigation approach, namely, for fair generation, safe generation, and responsible text-enhancing generation. Project page: https://interpretdiffusion.github.io.", + "bbox": [ + 76, + 321, + 473, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advances in vision language models have sparked increasing interest in ensuring their safety and responsible use [7, 23, 24]. In particular, text-to-image diffusion models, which have exhibited remarkable performance in creating images from text prompts [13, 15, 20, 31, 35, 37, 49], raise concerns about the risks of generating inappropriate content. The generated images may exhibit biases and unsafe elements, including instances of gender discrimination or the depiction of violent scenes that could be harmful to children. Recent research efforts have focused on introducing safety mechanisms to mitigate these issues, such as filtering out inappropriate text input, detecting inappropriate images with a safety guard classifier [4, 32, 36, 38] and", + "bbox": [ + 76, + 680, + 470, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "building safe diffusion models [5, 6, 17]. However, the underlying mechanism of how diffusion models generate inappropriate content remains poorly understood. In this work, we aim to explore the following questions. 1) Are there any internal representations associated with these inappropriate concepts in the diffusion model-based generation process? 2) Can we manipulate representations to avoid inappropriate content corresponding to a given concept, i.e., to achieve responsible image generation?", + "bbox": [ + 500, + 292, + 892, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To understand the image generation process of diffusion models, previous work has identified the bottleneck layer of the U-Net as a semantic representation space, dubbed $h$ -space [18]. They demonstrated that a vector in the $h$ -space can be associated with a specific semantic concept in the generated image. Manipulating the vector in the space can alter the generated image in a semantically meaningful way, such as adding a smile to a face. Several approaches [9, 18, 30] have been proposed to discover meaningful directions in this $h$ -space. For instance, an approach in [9] uses PCA to identify a set of latent directions that may represent semantic concepts.", + "bbox": [ + 496, + 429, + 893, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing approaches to identifying interpretable latent vectors are limited. In unsupervised approaches [9, 30], it is not clear to which semantic concepts those identified vectors correspond. The found vectors must be interpreted with humans in a loop. Furthermore, the number of interpretable directions depends on the training data [9, 30]. It is highly likely that some target concepts may not be found in the discovered directions, especially those related to fairness and safety. Supervised approaches [9, 18] have also been explored to identify target concepts. These methods require training external attribute classifiers supervised by human annotations. Additionally, the quality of the identified vectors is sensitive to the classifier's performance. Furthermore, new concepts require the training of new classifiers. Overall, existing interpretation methods cannot be easily applied to identify the corresponding semantic vector for a given inappropriate concept.", + "bbox": [ + 496, + 611, + 895, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a self-discovery approach to find interpretable latent directions in the $h$ -space for user", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12006", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "defined concepts. We learn a latent vector that effectively represents the concept by leveraging the model's acquired semantic knowledge in its internal representations. Initially, images are generated using specific text prompts related to the concept. The images are then used in a de-noising process where the frozen pretrained diffusion model reconstructs these images from noise, guided by a modified text prompt that omits the desired concept, and our introduced latent vector. By minimizing the reconstruction loss, the vector learns to represent the given concept. Our self-discovery approach eliminates the need for external models like CLIP text encoder [34] or dedicated attribute classifiers trained on human-labeled datasets. We identify ethical-related latent vectors and demonstrate their applications in responsible text-to-image generation: 1) fairness by sampling an ethical concept, e.g., gender, in the latent space, which generates images with unbiased attributes and aligned with the prompt. 2) safety generation by incorporating safety-related concepts, e.g., one that eliminates the nudity content, into $h$ -space to prevent the model from generating such harmful content. 3) responsible guidance, where we first discover responsible concepts in the text prompt and enhance the expression of those ethical concepts.", + "bbox": [ + 76, + 90, + 472, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous approaches enhance responsible image generation from different perspectives. Concretely, [2, 5, 6, 8, 17] fine-tune the diffusion models or text embeddings to unlearn harmful concepts, and [39] applies classifier-free guidance to steer the generation away from unsafe concepts. Despite the mitigation mechanism of previous approaches, diffusion models still suffer from inappropriate content generation [39, 50]. Unlike previous work, in this work, we provide a new perspective to mitigate the inappropriate generation, namely, finding and manipulating concepts in an interpretable latent space. Our work can be easily combined with previous mitigation approaches to further enhance responsible text-to-image generation.", + "bbox": [ + 76, + 441, + 472, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted extensive experiments on fairness, safety, and responsible guidance-enhancing generation. Our model consistently produces images with a balanced representation across societal groups. Further, we successfully mitigate harmful content for inappropriate prompts. In addition, our approach synergistically improves the performance of responsible image generation when combined with existing methods. Furthermore, we enhance text guidance to generate fair and safe content for responsible prompts.", + "bbox": [ + 76, + 638, + 472, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 777, + 423, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a self-discovery method for identifying interpretable directions in the diffusion latent space. Our approach can find a vector that represents any desired concept, without the need for labeled data or external models.", + "- With the discovered vectors, we propose a straightforward yet effective approach to enhance responsible generation, including fair generation, safe generation, and re" + ], + "bbox": [ + 76, + 795, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "sponsored text-enhancing generation.", + "- Extensive experiments are conducted to validate the effectiveness of our approach." + ], + "bbox": [ + 500, + 92, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 156, + 640, + 171 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Responsible Alignment of Diffusion Models Various approaches have been proposed to mitigate the generation of biased and unsafe content in diffusion models. A straightforward method involves refining the training dataset to remove biased and inappropriate content, exemplified by Stable Diffusion (SD) v2 [37]. Such approaches can be computationally intensive, may not fully eliminate harmful content [5], and could degrade the model's performance [39]. An alternative is to detect and filter out inappropriate words from the input prompts [1, 2, 27]. However, this fails to address non-explicit phrases that can still yield inappropriate outputs. Another line of approaches involves finetuning the parameters of pretrained models, aiming to remove the model's representation capability of generating such inappropriate concepts [5, 17]. However, they are sensitive to the adaptation process and may result in the degradation of the original models [6, 10, 28, 29, 48]. Moreover, such approaches require a potentially exhaustive list of words that introduce biases and harmful concepts [5, 6, 29]. Training-free approaches utilize classifier-free guidance to direct the generated images away from undesirable content during inference [1, 3, 39, 39, 47]. While they modify the noise space using text-based guidance through cross-attention mechanisms, we adopt a similar conditioning strategy to manipulate the generation for frozen pretrained models in the semantic latent space. As an orthogonal approach to the existing literature, we mitigate the inappropriate content by finding the corresponding latent directions in the U-Net bottleneck layer and suppressing their activations.", + "bbox": [ + 496, + 183, + 893, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpreting Diffusion Models To understand the working mechanisms of diffusion models, recent works mainly focus on investigating text guidance for conditional diffusion models [11, 16, 22, 26, 29, 43, 46], or analyzing the internal representations in diffusion models' intermediate layer activations [9, 18, 30, 40, 44]. We focus on elucidating the internal representations learned within the diffusion model, in line with prior works [18]. Some work [19, 33, 45, 51] proposes to create a semantic space in diffusion models by employing an autoencoder to encode the image into a semantic vector that guides the decoding process. However, their approaches require adapting the parameters of the autoencoder or even the entire framework.", + "bbox": [ + 496, + 626, + 893, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The seminal work [18] reveals that the bottleneck layer of U-Net architecture already exhibits properties suitable for a semantic representation space. They identified disentangled representations associated with the semantics of the generated image and demonstrated that those latent di", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12007", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb13d88ae470191c1fc694af0e667ae33a32f125a035adac5eec27731b720cf6.jpg", + "image_caption": [ + "Figure 1. Optimization framework to discover a semantic vector for a given concept. The top line shows that an image is firstly generated by the pretrained Stable Diffusion model for the prompt \"a female face\". The bottom part shows the optimization process for finding the concept for \"female\" in the semantic $h$ -space. The concept vector is used to reconstruct the image along with a modified prompt \"a face\", under an iterative denoising process. With the pretrained diffusion model frozen, the gradients of the reconstruction loss can solely update the latent vector to represent the missing gender information. After convergence, the latent vector is aligned with the U-Net's internal representation of the \"female\" concept, which can be used to guide new image generation." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rections are identical to different images. However, their approach relies on the CLIP classifier and paired source-target images and edits, making it inefficient. Another work proposes a PCA-based decomposition method on the latent space and finds interpretable attribute directions using the top right-hand singular vectors of the Jacobian. Additionally, [30] uses Riemannian metrics to define more accurate and meaningful directions. However, these approaches require manual interpretation to identify the editing effect of each component. Our approach differs from the supervised approach in [9] by enabling the efficient discovery of latent directions for any given target concept without requiring a data collection process or training external classifiers.", + "bbox": [ + 75, + 506, + 472, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 718, + 187, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section first introduces our optimization method to find interpretable directions in diffusion models' $h$ -space. In the second part, we show how to utilize discovered concepts in the inference process for responsible generation, including fairness, safety, and text-enhancing generation.", + "bbox": [ + 75, + 743, + 468, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Finding a Semantic Concept", + "text_level": 1, + "bbox": [ + 76, + 830, + 333, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models are generative models that generate samples from Gaussian noise through a denoising process [13, 41, 42]. Starting from a random vector $x_{T} \\sim \\mathcal{N}(0,1)$ of the", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "same dimension as the image, the model estimates a noise value at each time step to subtract from the current vector to obtain a denoised image, denoted as $x_{t - 1} = x_t - \\epsilon_\\theta (x_t,t)$ , where $\\epsilon_{\\theta}$ represents the U-Net of the diffusion model. A clean image $x_0$ is obtained at the end of this denoising process. The training of diffusion models involves a forward process that iteratively adds noise to images from the data, denoted as $x_{t} = x_{t - 1} + \\epsilon_{t}$ , with $\\epsilon_{t}\\sim \\mathcal{N}(0,1)$ . The training loss includes predicting noises for different steps,", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL = \\sum_ {x \\sim \\mathcal {D}} \\sum_ {t \\sim [ 0, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 241, + 890, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent work identified a semantic space in the diffusion model, the activations of U-Net's bottleneck layer $h$ , as shown in Figure 1. The activations in $h$ -space leads to the generation of a less noised image for the next timestep $x_{t - 1}$ . This space exhibits semantic structures and is easy to interpret. Activating a specific vector in the U-Net bottleneck layer leads to the image having a certain attribute. However, existing approaches cannot find the vector for an arbitrarily given concept. Our goal is to find such vectors.", + "bbox": [ + 496, + 290, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we utilize the text-to-image conditional diffusion model which can generate images from a given text input. The prediction function in Eq. 1 becomes $\\epsilon_{\\theta}(x_t,\\pi (y),t)$ where $\\pi (y)$ is the encodings of the input text $y$ . The equation specifies a conditional distribution that drives the generation of the image towards data regions that are highly likely given the input text [12]. To discover an interpretable direction, we leverage the pre-trained model to generate a set of images using dedicated prompts related to that concept. For example, to find the latent direction of the concept \"female\", we first generate a set of images $x^{+}$ with a descriptive prompt $y^{+}$ \"a photo of a female face\". Then, a concept vector is optimized for the conditional generation where the original prompt has been modified into $y^{-}$ \"a photo of a face\", eliminating gender information. The concept vector $c\\in \\mathbb{R}^D$ is randomly initialized in the latent space, where $D$ is the dimension of $h$ -space, and is optimized to minimize the reconstruction error. Since the pre-trained diffusion model is frozen, the model has to utilize the extra condition $c$ to compensate for the missing information not in the text condition but in the image. The concept vector $c$ will be forced to represent the missing information from the input text to produce an image with the lowest reconstruction error. After convergence, that vector $c$ is expected to represent the gender information \"female\". In this way, we discover a set of vectors that represent target concepts, such as gender, safety, and facial expressions.", + "bbox": [ + 496, + 428, + 892, + 837 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1The decoding of $x_{t-1}$ depends on other variables due to the presence of skip connections. For simplicity, we omit this consideration as the skip connection seems less significant in encoding compact semantic information, as supported in previous findings [9, 14].", + "bbox": [ + 496, + 849, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12008", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/838c5aac9ca7fa1df74e1b4ad20b0cfb88c3b151585475b7d9ac40a01a6349fd.jpg", + "image_caption": [ + "Figure 2. Fair Generation. Top: images generated from the prompt \"doctor\" are biased toward males. Bottom: we sample a learned male or female concept with equal probability for generating the doctors. The doctors now have fair gender. Images are generated from different random seeds." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 449, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, the optimal $c^*$ for a given concept is found by", + "bbox": [ + 76, + 321, + 447, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {*} = \\arg \\min _ {c} \\sum_ {x, y \\sim \\mathcal {D}} \\sum_ {t \\sim [ 0, T ]} \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t} ^ {+}, t, \\pi \\left(y ^ {-}\\right), c\\right) \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 349, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$x_{t}^{+}$ denotes the noised version of the original image generated with $y^{+}$ , $c$ represents the target concept. $\\epsilon_{\\theta}$ denotes the U-Net that linearly adds an additional concept vector $c$ to its $h$ -space, at each decoding timestep. Regarding implementation, the $h$ -space is the flattened activations after the middle bottleneck layer of the U-Net. The pseudo-code for this training pipeline is in Appendix A.1.", + "bbox": [ + 75, + 397, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We learn a single vector for each concept for all timesteps, as the latent direction remains approximately consistent across different timesteps [18]. Moreover, we restrict the operation to linearity to demonstrate the power of this latent space. Notably, the learned vector generalizes effectively to new images [9] and diverse prompts [30]. For instance, a \"male\" concept learned with the base prompt \"person\" can be used in different contexts, such as \"doctor\" or \"manager\", as shown in the next section. Additionally, the concepts can be optimized jointly or independently, with the experimental section demonstrating the impact of concept composition. A key strength of our approach is utilizing synthesis by diffusion models to collect data, eliminating the need for human labeling and training of guiding classifiers. Nevertheless, our approach can be applied to realistic datasets with annotated attributes.", + "bbox": [ + 75, + 503, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Responsible Generation with Self-discovered Interpretable Latent Direction", + "text_level": 1, + "bbox": [ + 76, + 755, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we utilize the identified directions to manipulate the latent activation in the latent space for fair, safe, and enhanced responsible generation.", + "bbox": [ + 75, + 794, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fair Generation Method A text prompt contains words that lead to the generation of biased societal groups. We aim to generate images with evenly distributed attributes for a given text prompt. For example, for the prompt \"doctor\",", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f362d56a6aae7352e4d2e42b09b3e1058cafd58ca0ce352f396f4b568daf7457.jpg", + "image_caption": [ + "Figure 3. Safe Generation. When the user's prompt contains implicit references to nudity, the original model (shown in the top row) generates an inappropriate image, as the added blurriness indicates. In contrast, our approach generates an image for the same prompt by setting a safety-related concept in $h$ -space, identified in the previous section. The vector anti-sexual concept represents the direction to suppress nudity content, effectively eliminating inappropriate content while maintaining fidelity to the prompt." + ], + "image_footnote": [], + "bbox": [ + 521, + 87, + 870, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we aim to generate an image of a male doctor with a $50\\%$ probability and a female doctor with a $50\\%$ probability. For that, we learn a set of semantic concepts representing different societal groups using the approach in the previous section. For inference, a concept vector is sampled from the learned concepts in the societal group with equal probability, e.g., the C-male and C-female concept vectors for gender are chosen with fair chance. The inference process is fixed as before, except that the sampled vector is added to the original activations in $h$ -space at each decoding step, denoted by", + "bbox": [ + 496, + 364, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh \\leftarrow h + c \\sim \\operatorname {C a t e g o r i c a l} \\left(p _ {k}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 542, + 890, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_k$ represents the probability of sampling a particular attribute $c_k$ from the societal group with $C$ distinct attributes. For the fair generation, $p_k = 1 / C$ . Guided by this sampled concept vector, the generated image is expected to be a male doctor if the C-male concept is sampled or a female doctor otherwise. This allows the generated images to have an equal number of attributes, e.g., an equal number of male and female doctors, shown in Figure 2.", + "bbox": [ + 496, + 568, + 890, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Safe Generation Method For safety generation, we consider text prompts that contain explicit or implicit references to inappropriate content, which we aim to eliminate. An example of such a prompt is illustrated in Figure 3, where the phrase \"a gorgeous woman\" may indirectly lead to the generation of nudity. We identify a collection of safety-related concepts, such as anti-sexual, to achieve safe generation.", + "bbox": [ + 496, + 689, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we learn the opposite latent direction of an inappropriate concept, leveraging the negative prompt technique. For instance, the training images are generated by the prompt $y^{+}$ \"a gorgeous person\" with a negative prompt \"sexual\", which effectively instructs the Stable Diffusion to generate safe images without sexual content. The concept vector is then optimized on those training images that depict", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12009", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "safe content. For that, the input prompt $y^{-}$ is set to \"a gorgeous person\" but without the negative prompt \"sexual\". In this way, the concept vector directly learns the concept of \"anti-sexual\". The reason for adopting this strategy is the difficulty of listing all the opposite concepts of sexuality, e.g., \"dressed\", \"clothes\", or more. An alternative approach is to learn the concept of sexuality directly and apply negation during generation, which we found less effective. More details regarding the negative prompt are in Appendix A.2.", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the learning process, we maintain all aspects of inference unchanged, except for adding the learned vector to the original activations at the bottleneck layer, formally as", + "bbox": [ + 75, + 227, + 468, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh \\leftarrow h + c _ {s}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 280, + 468, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $c_{s}$ refers to a safety concept, such as \"anti-sexual\", which represents the opposite of sexual content. This strengthens the expression of safe concepts in the generated images so they are devoid of harmful content. Figure 3 illustrates the impact of including the anti-sexual vector, resulting in a visually appealing person with appropriate clothes. Responsible Text-enhancing Generation Method Even when a prompt is intentionally designed to promote safety, the generative models may struggle to accurately incorporate all the concepts defined in the prompt. For instance, consider a text prompt like \"an exciting Halloween party, no violence\". The generative model may encounter difficulties in faithfully representing each responsible concept from the prompt, e.g., its poor understanding of negation on \"violence\" may result in the generation of inappropriate content.", + "bbox": [ + 75, + 303, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we utilize our self-discovery approach to learn concepts such as gender, race, and safety. To enhance the generation of responsible prompts, we extract safety-related content from the text and leverage our learned ethical-related concepts to reinforce the expression of desired visual features. During inference, we apply the extracted concepts $c(y)$ from the prompt to the original activations, denoted as", + "bbox": [ + 75, + 545, + 468, + 665 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh \\leftarrow h + c (y) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 674, + 468, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For example, as shown in Figure 4, the concept of \"no violence\" from the text prompt activates our learned \"anti-violence\" concept during inference. By directly manipulating the semantic space, our approach introduces the desired attributes to the generated image. Compared to the original generated image, the anti-violence concept effectively mitigates the presence of violent content and makes the generated images more appropriate.", + "bbox": [ + 75, + 696, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 830, + 209, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the performance of our methods on fair generation, safe generation, and enhanced responsible generation. Additionally, we demonstrate the general", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e5b7fe4746436af20cd134c77d7aada804c178c8a0105c8031b89fa9d7efa521.jpg", + "image_caption": [ + "Figure 4. Responsible text-enhancing generation. The original model may fail to capture the safety concepts specified in the text, such as \"no violence\". We propose extracting those safety concepts from the given prompt and activating the safety directions during generation. The bottom image demonstrates that incorporating our safety concepts can enhance the text guidance of the original prompt." + ], + "image_footnote": [], + "bbox": [ + 521, + 87, + 870, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "properties of our discovered concept vectors, highlighting their generalization capability and compositionality.", + "bbox": [ + 498, + 343, + 890, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Fair Generation", + "text_level": 1, + "bbox": [ + 500, + 386, + 661, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task The goal is to increase the diversity of societal groups in the generated images, particularly in professions where existing models exhibit gender and racial bias.", + "bbox": [ + 498, + 410, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Following [6, 29], our approach is evaluated on the Winobias [52] benchmark that comprises 36 professions known to exhibit gender biases. We generate five different prompts for each profession using various templates, such as \"a portrait of a doctor\". Additionally, we create a hard dataset by augmenting the existing dataset with prompts that are more likely to induce stereotypes. This extended dataset adds the term \"successful\" to each original prompt, such as \"a portrait of a successful doctor\". The word \"successful\" often leads to the generation of male images and thus constitutes challenges for a fair generation [6]. Appendix B.1 contains the complete list of prompts.", + "bbox": [ + 496, + 455, + 892, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric The CLIP classifier is employed to predict attributes by measuring the similarity between the text embedding of a concept (e.g., female, male) and the embedding of the generated image. We utilize the deviation ratio [6, 29] to quantify the imbalance of different attributes. To accommodate an arbitrary number of attributes, the metric is modified as $\\Delta = \\max_{c\\in C}\\frac{|N_c / N - 1 / C|}{1 - 1 / C}$ , where $C$ is the total number of attributes within a societal group, $N$ is the total number of generated images, and $N_{c}$ denotes the number of images whose maximum predicted attribute equals $c$ . In particular, we test the gender, male, female, and racial, black, white, Asian, biases associated with the professions. These races are selected as the CLIP classifier has relatively reliable predictions on these attributes. During the evaluation, 150 images were generated for each profession.", + "bbox": [ + 496, + 638, + 893, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Approach Setting In all experiments, we use the Stable Diffusion v1.4 checkpoint and set the guidance scale to 7.5", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12010", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4434228768e8c00e84057e02eeeb376e4346343c7d3c35a42a7d610801504a9c.jpg", + "image_caption": [ + "(a) Images generated by $\\mathbf{SD}$" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 477, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89c656c08923f91da03787a4b3eac60e505e42f572dbe9fb284e3d9c88324882.jpg", + "image_caption": [ + "(b) Images generated by Ours", + "Figure 5. Gender fairness generation. From the prompt \"a photo of a doctor\", the original SD exhibits significant gender bias, as shown on the left side. Our approach with uniformly sampled gender vectors represents genders equally in the generated images." + ], + "image_footnote": [], + "bbox": [ + 493, + 90, + 888, + 213 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/75f931260144ccacae93b4fbff2fae0b6bab824786af5db068bab2b2c88b1e81.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset\nMethodGenderGender+\nUCERaceRace+\nUCE
SDUCEOursSDUCEOursSDUCEOursSDUCEOurs
Analyst0.700.200.020.540.040.020.820.290.240.770.200.41
CEO0.920.280.060.900.580.060.380.130.220.310.080.22
Laborer1.000.090.120.980.080.140.330.400.240.530.380.20
Secretary0.640.100.360.920.960.460.370.350.240.550.430.32
Teacher0.300.060.040.480.160.100.510.100.040.260.230.21
Winobias[52]0.680.220.170.700.520.230.560.210.230.480.350.20
", + "bbox": [ + 94, + 282, + 874, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Fair generation quantified by the deviation ratio $\\left( {0 \\leq \\Delta \\leq 1}\\right)$ . Lower values indicate better performance. The left side of the table presents the results on gender attributes, whereas the right side quantifies the racial bias. \"Gender+/Race+\" refers to the extended Winobias dataset, which is more challenging, as described in Subsection 4.1. Our approach leads to unbiased generation for biased prompts and is robust to diverse sources of bias in the prompt.", + "bbox": [ + 75, + 422, + 892, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for text-to-image generation. We find five concept vectors using a base prompt \"person\", e.g., $y^{+} =$ \"a photo of a woman\" and $y^{-} =$ \"a photo of a person\" to learn the concept \"female\". The concept vectors are optimized for 10K steps on 1K synthesized images for each concept. During inference, we directly employ the learned vector without any scaling. Unlike the baseline approach UCE [6], which needs to debias each profession in Winobias, Our approach is trained solely on the \"person\" prompt to learn the male and female concept that generalizes to all different professions. For comparison, we report UCE's published scores when available and otherwise use their released code to train the model.", + "bbox": [ + 75, + 506, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results and Analysis Table 1 reveals that our approach is significantly better than the original SD and outperforms the state-of-the-art debiasing approach UCE. The professions on the table are randomly selected from the complete list of 36 professions (see Appendix B.2). Figure 5 compares images generated from our approach and those from the original SD. Further, we highlight the generalization capability of our approach to different text prompts using the extended Winobias dataset, as shown in the second and fourth column blocks of Table 1. Despite the presence of bias in the text prompts, our approach consistently performs well as it directly operates on the latent visual space. In contrast, UCE performs poorly on this challenging dataset, as it relies on", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "debiasing each word in the prompt. The effectiveness of UCE is easily weakened by biased words that are not included in its training set. Additionally, we demonstrate that the quality of images generated by our approach remains consistent with the original SD and UCE in Appendix B.3.", + "bbox": [ + 496, + 506, + 892, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Safe Generation", + "text_level": 1, + "bbox": [ + 500, + 593, + 660, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task This section focuses on generating images that eliminate harmful content specified in inappropriate prompts. As an orthogonal approach to existing methods, our approach is combined with current safety methods, including SLD [39] and ESD [5], to eliminate inappropriate generation further.", + "bbox": [ + 496, + 616, + 890, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and Evaluation Metric The I2P benchmark [39] is a collection of 4703 inappropriate prompts from real-world user prompts. The inappropriateness covers seven categories, including, e.g., illegal activity, sexual, and violence. For evaluation, the Nudenet detector and Q16 classifier are used to detect nudity or violent content in an image. An image is classified as inappropriate if any of the classifiers predicts a positive [5]. Five images are generated for each prompt for evaluation.", + "bbox": [ + 496, + 693, + 890, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Approach Setting We find that optimizing a single concept vector for \"safety\" is challenging. Therefore, we learn the concept vector for each inappropriate concept defined in the", + "bbox": [ + 496, + 830, + 890, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://github.com/notAI-tech/NudeNet", + "bbox": [ + 514, + 886, + 816, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c5cb060bbff5193c99c85aa90f583567e86f469ccae81f7abfabecef632263df.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryHarassmentHateIllegalSelf-harmSexualShockingViolenceI2P[39]
SD0.34 ±0.0190.41 ±0.0320.34 ±0.0180.44 ±0.0190.38 ±0.0160.51 ±0.0170.44 ±0.0180.41 ±0.007
Ours-SD0.18 ±0.0150.29 ±0.0300.23 ±0.0160.28 ±0.0170.22 ±0.0140.36 ±0.0170.30 ±0.0170.27 ±0.006
SLD[39]0.15 ±0.0140.18 ±0.0250.17 ±0.0150.19 ±0.0150.15 ±0.0120.32 ±0.0160.21 ±0.0150.20 ±0.006
Ours-SLD0.14 ±0.0140.20 ±0.0270.14 ±0.0130.14 ±0.0130.09 ±0.0100.25 ±0.0150.16 ±0.0130.16 ±0.005
ESD[5]0.27 ±0.0180.32 ±0.0310.33 ±0.0180.35 ±0.0180.18 ±0.0130.41 ±0.0170.41 ±0.0180.32 ±0.007
Ours-ESD0.26 ±0.0170.29 ±0.0300.25 ±0.0170.26 ±0.0170.13 ±0.0110.41 ±0.0170.30 ±0.0170.27 ±0.006
", + "bbox": [ + 78, + 88, + 895, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "I2P dataset, e.g., \"anti-sexual\". The identified concept vectors are linearly combined as the final vector to modulate the generation process. The effect of linearly combining concept vectors is further discussed in the next section. Additionally, certain concepts are rather abstract and include diverse visual categories, e.g., \"hate\". Adding these concepts improves safety yet at a higher cost of image quality degradation. Therefore, we utilize two prevailing concepts, anti-sexual and anti-violence, which may underline other concepts, as our final model. For baselines, we use SLD [39] and ESD [5] with their released model weights. Detailed settings and hyperparameters are in Appendix C.", + "bbox": [ + 75, + 282, + 468, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results and Analysis Table 2 demonstrates the effectiveness of our approach in eliminating inappropriate content. Notably, our safety vector can suppress inappropriate concepts that existing approaches failed to eliminate, as evidenced by the improved performance compared to existing approaches. In particular, we observe significant improvement in eliminating nudity with the combined approach of SLD and our proposed method, with a relative improvement of $40\\%$ . Additionally, even after retraining the model weights to eliminate nudity, ESD still generates inappropriate images, which our approach can mitigate.", + "bbox": [ + 75, + 465, + 470, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Enhancing Responsible Text Guidance", + "text_level": 1, + "bbox": [ + 76, + 645, + 410, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task For user prompts classified as responsible text, we aim to accurately represent the responsible phrases in the prompt in the generated image.", + "bbox": [ + 75, + 670, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset We created a dataset of 200 prompts that explicitly include responsible concepts. In particular, the dataset focuses on gender and race fairness, as well as the removal of sexual and violent content. For instance, the prompt \"a fair-gender doctor is operating a surgery\" assesses the model's ability to generate images with balanced genders based on the text cue \"fair-gender\". Additionally, prompts such as \"a picture of a loved couple, without sexual content\" address safety concerns. More dataset details are in Appendix D.", + "bbox": [ + 75, + 717, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results and Analysis We utilized previously introduced metrics for evaluation, i.e., the deviation ratio for fairness, NudeNet for sexual content, and Q16 for violent content.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b3e18b7dda0cce3bb1e70cec1b29f536066e62e9298a0ff6f73e47a19d31972c.jpg", + "table_caption": [ + "Table 2. The proportion of images classified as inappropriate on the I2P benchmark. In each block of results, the first row shows the performance of the original method, while the second row represents adding our concept vector to the corresponding baseline model. Our identified safety-related vector can be combined with existing safety approaches to mitigate inappropriate content generation." + ], + "table_footnote": [], + "table_body": "
ModelGenderRaceSexualViolence
SD0.21990.16000.43000.4551
Ours-SD0.14330.13990.26400.3204
", + "bbox": [ + 508, + 279, + 880, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. For prompts containing responsible concepts, the original SD may fail to follow the prompts faithfully. Our approach effectively enhances responsible text-guidance generation.", + "bbox": [ + 498, + 335, + 893, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For evaluation, 3500 images are generated for the dataset. For our approach, we provide the model with the corresponding concept associated with the input prompt. For example, if a prompt mentions \"no sexual\", the anti-sexual vector is added to the generation process. Table 3 compares our approach with the original model, which does not use the safety concepts. Our approach effectively enhances the text guidance for responsible instructions.", + "bbox": [ + 496, + 405, + 890, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99de2f3204c9905559f8260432c09c4dc6c811f61596c94c03e0b96ae884caae.jpg", + "image_caption": [ + "Figure 6. Concept interpolation. The first column displays the original image generated by SD. The following columns show images generated by the same random seeds, but with concept vector scales linearly increasing from 0.2 to 0.8." + ], + "image_footnote": [], + "bbox": [ + 500, + 540, + 890, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Semantic Concepts", + "text_level": 1, + "bbox": [ + 500, + 770, + 684, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In previous experiments, we have demonstrated the specific applications of our identified concept vectors for responsible generation. This subsection introduces the general properties of discovered vectors related to the semantic space.", + "bbox": [ + 496, + 794, + 890, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpolation Figure 6 illustrates the impact of manipulating image semantics by linearly controlling the strength of the concept vector, denoted as $\\lambda$ in the equation $h\\gets h + \\lambda c$", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12012", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dee4388c5d867e1c32ea30c301707d7bc96abb9a306fc1e5e32c8539a0d38d22.jpg", + "image_caption": [ + "Figure 7. Multiple concepts composition. The concept vectors of gender, age, and race were learned independently. Linearly adding latent vectors can generate images with corresponding semantics." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The image is gradually modified to the introduced concept by adjusting the added vector's strength. The smooth transition indicates that the discovered vector represents the target semantic concept while remaining approximately disentangled from other semantic factors. Appendix E.1 presents more examples of concept manipulation and enhanced fidelity by post-hoc interpolation methods [25].", + "bbox": [ + 75, + 310, + 468, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Composition Figure 7 showcases the composability of learned concept vectors, which were trained independently. Images are generated from the prompt \"a photo of a doctor\". By linearly combining these concept vectors, we can control the corresponding attributes in the generated image. Appendix E.2 provides a quantitative evaluation.", + "bbox": [ + 75, + 419, + 472, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ddbae4d8953f1c74dc4dfa3dcaaafec76a56f77dd8b17a0caecdefd33c9b087.jpg", + "image_caption": [ + "Figure 8. General semantic concepts identified by our approach. Top: The concept of \"running\" is learned from dog images and can be generalized to different objects. Bottom: The concept vector of \"glasses\" enhances the prompt \"an elephant wearing glasses\"." + ], + "image_footnote": [], + "bbox": [ + 107, + 525, + 436, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization Figure 8 illustrates the generalization capability of our discovered concept vector to universal semantic concepts. We train the latent vector for the concept \"running\" on generated dog images and test its effect on other objects using prompts such as \"a photo of a cat\". Each pair of images in the figure is generated from the same random seed. Although the vector of \"running\" was learned from dogs, it successfully extends to different animals and even humans. Additionally, our approach enhances the original text guidance for the prompt \"an ele", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5bacde55c5f5b2db2c4d2d0c013fc0879393cf46b89fdc1db8058d7e7aa938fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSDSLD*SLDESD*ESDOurs-SD
FID14.0916.9019.3513.6815.3615.98
CLIP31.33-30.41-30.0531.03
", + "bbox": [ + 503, + 88, + 890, + 141 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Evaluation of the quality of generated images on the COCO-30K [21] dataset using FID for image fidelity and CLIP Score for semantic alignment with input text. Various safety approaches have approximately the same level of image quality as the original SD. Numbers reported from the corresponding papers are denoted with * .", + "bbox": [ + 496, + 150, + 893, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "phant wearing glasses\". The original SD cannot produce accurate images, as shown in the first and third images on the bottom. The correct images can be generated by adding the concept vector \"glasses\" in $h$ -space. More visualizations are in Appendix E.3.", + "bbox": [ + 496, + 261, + 890, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact on Image Quality Additionally, we find that the quality of generated images remains approximately the same level as the original SD, as shown in Table 4. The observed differences in the reported scores and these in our experiments can be attributed to the randomness during image generation and caption sampling, which aligns with the inconsistencies reported in other studies [5].", + "bbox": [ + 496, + 340, + 892, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity to Hyperparameters In Appendix F, we investigate the sensitivity of our approach to hyperparameters, finding that it is less affected by factors such as the number of training images or different input prompts. Additionally, we demonstrate that our approach can leverage existing datasets to discover concept vectors.", + "bbox": [ + 496, + 449, + 893, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 554, + 619, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we introduced a self-discovery approach to identify semantic concepts in the latent space of text-to-image diffusion models. Our research findings highlight that the generation of inappropriate content can be attributed to ethical-related concepts present in the internal semantic space of diffusion models. Leveraging these concept vectors, we enable responsible generation, including promoting equality among societal groups, eliminating inappropriate content, and enhancing text guidance for responsible prompts. Through extensive experiments, we have demonstrated the effectiveness and superiority of our proposed approach. Our work contributes to the understanding of internal representations in diffusion models and facilitates the generation of responsible content, maximizing the utility of high-quality text-to-image generation.", + "bbox": [ + 496, + 579, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1, EPSRC/MURI grant: EP/N019474/1. We thank the Royal Academy of Engineering and FiveAI. This work is also founded by the German Federal Ministry of Education and Research and the Bavarian State Ministry for Science and the Arts.", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12013", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2", + "[2] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2", + "[3] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2", + "[4] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/ logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 1", + "[5] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8", + "[6] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 5, 6", + "[7] Jindong Gu, Ahmad Beirami, Xuezhi Wang, Alex Beutel, Philip Torr, and Yao Qin. Towards robust prompts on vision-language models. arXiv preprint arXiv:2304.08479, 2023. 1", + "[8] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 2", + "[9] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. arXiv preprint arXiv:2303.11073, 2023. 1, 2, 3, 4", + "[10] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. arXiv preprint arXiv:2305.10120, 2023. 2", + "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 12", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 12", + "[14] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 3", + "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 1", + "[16] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2", + "[17] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2", + "[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4", + "[19] Yipeng Leng, Qiangjuan Huang, Zhiyuan Wang, Yangyang Liu, and Haoyu Zhang. Diffusegae: Controllable and high-fidelity image manipulation from disentangled representation. arXiv preprint arXiv:2307.05899, 2023. 2", + "[20] Hang Li, Jindong Gu, Rajat Koner, Sahand Sharifzadeh, and Volker Tresp. Do dall-e and flamingo understand each other? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1999-2010, 2023. 1", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8", + "[22] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2", + "[23] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In The Twelfth International Conference on Learning Representations, 2023. 1", + "[24] Avery Ma, Amir-massoud Farahmand, Yangchen Pan, Philip Torr, and Jindong Gu. Improving adversarial transferability via model alignment. arXiv preprint arXiv:2311.18495, 2023. 1", + "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 8, 15", + "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12014", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Minheng Ni, Chenfei Wu, Xiaodong Wang, Shengming Yin, Lijuan Wang, Zicheng Liu, and Nan Duan. Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785, 2023. 2", + "[28] Zixuan Ni, Longhui Wei, Jiacheng Li, Siliang Tang, Yueting Zhuang, and Qi Tian. Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8900-8909, 2023. 2", + "[29] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7053-7061, 2023. 2, 5", + "[30] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4", + "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1", + "[32] Vinay Uday Prabhu and Abeba Birhane. Large image datasets: A pyrrhic win for computer vision? arXiv preprint arXiv:2006.16923, 2020. 1", + "[33] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[36] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 13", + "[38] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350–1361, 2022. 1, 6", + "[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2, 6, 7", + "[40] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 2", + "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3", + "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[43] Matthew Trager, Pramuditha Perera, Luca Zancato, Alessandro Achille, Parminder Bhatia, and Stefano Soatto. Linear spaces of meanings: Compositional structures in vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15395-15404, 2023. 2", + "[44] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2", + "[45] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. arXiv preprint arXiv:2306.08757, 2023. 2", + "[46] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 2", + "[47] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2", + "[48] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2", + "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1", + "[50] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868, 2023. 2", + "[51] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in Neural Information Processing Systems, 35:22117-22130, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12015", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[52] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. arXiv preprint arXiv:1804.06876, 2018. 5, 6", + "bbox": [ + 78, + 90, + 470, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12016", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_model.json b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b69b223757e3c2c7776fafdc67022a2411d0179a --- /dev/null +++ b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_model.json @@ -0,0 +1,2103 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.203, + 0.845, + 0.221 + ], + "angle": 0, + "content": "Hang Li\\(^{1,4,5}\\) Chengzhi Shen\\(^{3}\\) Philip Torr\\(^{2}\\) Volker Tresp\\(^{1,4}\\) Jindong Gu\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.882, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)LMU Munich, Germany \\(^{2}\\)University of Oxford, UK \\(^{3}\\)Technical University of Munich, Germany \\(^{4}\\)Munich Center for Machine Learning, Germany \\(^{5}\\)Siemens AG, Germany" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.323, + 0.474, + 0.626 + ], + "angle": 0, + "content": "Diffusion-based models have gained significant popularity for text-to-image generation due to their exceptional image-generation capabilities. A risk with these models is the potential generation of inappropriate content, such as biased or harmful images. However, the underlying reasons for generating such undesired content from the perspective of the diffusion model's internal representation remain unclear. Previous work interprets vectors in an interpretable latent space of diffusion models as semantic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation. Extensive experiments have been conducted to verify the effectiveness of our mitigation approach, namely, for fair generation, safe generation, and responsible text-enhancing generation. Project page: https://interpretdiffusion.github.io." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.672 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.471, + 0.878 + ], + "angle": 0, + "content": "The rapid advances in vision language models have sparked increasing interest in ensuring their safety and responsible use [7, 23, 24]. In particular, text-to-image diffusion models, which have exhibited remarkable performance in creating images from text prompts [13, 15, 20, 31, 35, 37, 49], raise concerns about the risks of generating inappropriate content. The generated images may exhibit biases and unsafe elements, including instances of gender discrimination or the depiction of violent scenes that could be harmful to children. Recent research efforts have focused on introducing safety mechanisms to mitigate these issues, such as filtering out inappropriate text input, detecting inappropriate images with a safety guard classifier [4, 32, 36, 38] and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.893, + 0.43 + ], + "angle": 0, + "content": "building safe diffusion models [5, 6, 17]. However, the underlying mechanism of how diffusion models generate inappropriate content remains poorly understood. In this work, we aim to explore the following questions. 1) Are there any internal representations associated with these inappropriate concepts in the diffusion model-based generation process? 2) Can we manipulate representations to avoid inappropriate content corresponding to a given concept, i.e., to achieve responsible image generation?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.895, + 0.612 + ], + "angle": 0, + "content": "To understand the image generation process of diffusion models, previous work has identified the bottleneck layer of the U-Net as a semantic representation space, dubbed \\( h \\)-space [18]. They demonstrated that a vector in the \\( h \\)-space can be associated with a specific semantic concept in the generated image. Manipulating the vector in the space can alter the generated image in a semantically meaningful way, such as adding a smile to a face. Several approaches [9, 18, 30] have been proposed to discover meaningful directions in this \\( h \\)-space. For instance, an approach in [9] uses PCA to identify a set of latent directions that may represent semantic concepts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.896, + 0.87 + ], + "angle": 0, + "content": "However, existing approaches to identifying interpretable latent vectors are limited. In unsupervised approaches [9, 30], it is not clear to which semantic concepts those identified vectors correspond. The found vectors must be interpreted with humans in a loop. Furthermore, the number of interpretable directions depends on the training data [9, 30]. It is highly likely that some target concepts may not be found in the discovered directions, especially those related to fairness and safety. Supervised approaches [9, 18] have also been explored to identify target concepts. These methods require training external attribute classifiers supervised by human annotations. Additionally, the quality of the identified vectors is sensitive to the classifier's performance. Furthermore, new concepts require the training of new classifiers. Overall, existing interpretation methods cannot be easily applied to identify the corresponding semantic vector for a given inappropriate concept." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose a self-discovery approach to find interpretable latent directions in the \\(h\\)-space for user" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12006" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.439 + ], + "angle": 0, + "content": "defined concepts. We learn a latent vector that effectively represents the concept by leveraging the model's acquired semantic knowledge in its internal representations. Initially, images are generated using specific text prompts related to the concept. The images are then used in a de-noising process where the frozen pretrained diffusion model reconstructs these images from noise, guided by a modified text prompt that omits the desired concept, and our introduced latent vector. By minimizing the reconstruction loss, the vector learns to represent the given concept. Our self-discovery approach eliminates the need for external models like CLIP text encoder [34] or dedicated attribute classifiers trained on human-labeled datasets. We identify ethical-related latent vectors and demonstrate their applications in responsible text-to-image generation: 1) fairness by sampling an ethical concept, e.g., gender, in the latent space, which generates images with unbiased attributes and aligned with the prompt. 2) safety generation by incorporating safety-related concepts, e.g., one that eliminates the nudity content, into \\( h \\)-space to prevent the model from generating such harmful content. 3) responsible guidance, where we first discover responsible concepts in the text prompt and enhance the expression of those ethical concepts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.473, + 0.639 + ], + "angle": 0, + "content": "Previous approaches enhance responsible image generation from different perspectives. Concretely, [2, 5, 6, 8, 17] fine-tune the diffusion models or text embeddings to unlearn harmful concepts, and [39] applies classifier-free guidance to steer the generation away from unsafe concepts. Despite the mitigation mechanism of previous approaches, diffusion models still suffer from inappropriate content generation [39, 50]. Unlike previous work, in this work, we provide a new perspective to mitigate the inappropriate generation, namely, finding and manipulating concepts in an interpretable latent space. Our work can be easily combined with previous mitigation approaches to further enhance responsible text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.473, + 0.777 + ], + "angle": 0, + "content": "We conducted extensive experiments on fairness, safety, and responsible guidance-enhancing generation. Our model consistently produces images with a balanced representation across societal groups. Further, we successfully mitigate harmful content for inappropriate prompts. In addition, our approach synergistically improves the performance of responsible image generation when combined with existing methods. Furthermore, we enhance text guidance to generate fair and safe content for responsible prompts." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.778, + 0.424, + 0.793 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.47, + 0.856 + ], + "angle": 0, + "content": "- We propose a self-discovery method for identifying interpretable directions in the diffusion latent space. Our approach can find a vector that represents any desired concept, without the need for labeled data or external models." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.857, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- With the discovered vectors, we propose a straightforward yet effective approach to enhance responsible generation, including fair generation, safe generation, and re" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.759, + 0.107 + ], + "angle": 0, + "content": "sponsored text-enhancing generation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.138 + ], + "angle": 0, + "content": "- Extensive experiments are conducted to validate the effectiveness of our approach." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.093, + 0.892, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.157, + 0.642, + 0.172 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.895, + 0.624 + ], + "angle": 0, + "content": "Responsible Alignment of Diffusion Models Various approaches have been proposed to mitigate the generation of biased and unsafe content in diffusion models. A straightforward method involves refining the training dataset to remove biased and inappropriate content, exemplified by Stable Diffusion (SD) v2 [37]. Such approaches can be computationally intensive, may not fully eliminate harmful content [5], and could degrade the model's performance [39]. An alternative is to detect and filter out inappropriate words from the input prompts [1, 2, 27]. However, this fails to address non-explicit phrases that can still yield inappropriate outputs. Another line of approaches involves finetuning the parameters of pretrained models, aiming to remove the model's representation capability of generating such inappropriate concepts [5, 17]. However, they are sensitive to the adaptation process and may result in the degradation of the original models [6, 10, 28, 29, 48]. Moreover, such approaches require a potentially exhaustive list of words that introduce biases and harmful concepts [5, 6, 29]. Training-free approaches utilize classifier-free guidance to direct the generated images away from undesirable content during inference [1, 3, 39, 39, 47]. While they modify the noise space using text-based guidance through cross-attention mechanisms, we adopt a similar conditioning strategy to manipulate the generation for frozen pretrained models in the semantic latent space. As an orthogonal approach to the existing literature, we mitigate the inappropriate content by finding the corresponding latent directions in the U-Net bottleneck layer and suppressing their activations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.895, + 0.824 + ], + "angle": 0, + "content": "Interpreting Diffusion Models To understand the working mechanisms of diffusion models, recent works mainly focus on investigating text guidance for conditional diffusion models [11, 16, 22, 26, 29, 43, 46], or analyzing the internal representations in diffusion models' intermediate layer activations [9, 18, 30, 40, 44]. We focus on elucidating the internal representations learned within the diffusion model, in line with prior works [18]. Some work [19, 33, 45, 51] proposes to create a semantic space in diffusion models by employing an autoencoder to encode the image into a semantic vector that guides the decoding process. However, their approaches require adapting the parameters of the autoencoder or even the entire framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The seminal work [18] reveals that the bottleneck layer of U-Net architecture already exhibits properties suitable for a semantic representation space. They identified disentangled representations associated with the semantics of the generated image and demonstrated that those latent di" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12007" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.473, + 0.479 + ], + "angle": 0, + "content": "Figure 1. Optimization framework to discover a semantic vector for a given concept. The top line shows that an image is firstly generated by the pretrained Stable Diffusion model for the prompt \"a female face\". The bottom part shows the optimization process for finding the concept for \"female\" in the semantic \\( h \\)-space. The concept vector is used to reconstruct the image along with a modified prompt \"a face\", under an iterative denoising process. With the pretrained diffusion model frozen, the gradients of the reconstruction loss can solely update the latent vector to represent the missing gender information. After convergence, the latent vector is aligned with the U-Net's internal representation of the \"female\" concept, which can be used to guide new image generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.473, + 0.704 + ], + "angle": 0, + "content": "rections are identical to different images. However, their approach relies on the CLIP classifier and paired source-target images and edits, making it inefficient. Another work proposes a PCA-based decomposition method on the latent space and finds interpretable attribute directions using the top right-hand singular vectors of the Jacobian. Additionally, [30] uses Riemannian metrics to define more accurate and meaningful directions. However, these approaches require manual interpretation to identify the editing effect of each component. Our approach differs from the supervised approach in [9] by enabling the efficient discovery of latent directions for any given target concept without requiring a data collection process or training external classifiers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.719, + 0.188, + 0.736 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.82 + ], + "angle": 0, + "content": "This section first introduces our optimization method to find interpretable directions in diffusion models' \\( h \\)-space. In the second part, we show how to utilize discovered concepts in the inference process for responsible generation, including fairness, safety, and text-enhancing generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.334, + 0.849 + ], + "angle": 0, + "content": "3.1. Finding a Semantic Concept" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Diffusion models are generative models that generate samples from Gaussian noise through a denoising process [13, 41, 42]. Starting from a random vector \\( x_{T} \\sim \\mathcal{N}(0,1) \\) of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "same dimension as the image, the model estimates a noise value at each time step to subtract from the current vector to obtain a denoised image, denoted as \\( x_{t - 1} = x_t - \\epsilon_\\theta (x_t,t) \\), where \\( \\epsilon_{\\theta} \\) represents the U-Net of the diffusion model. A clean image \\( x_0 \\) is obtained at the end of this denoising process. The training of diffusion models involves a forward process that iteratively adds noise to images from the data, denoted as \\( x_{t} = x_{t - 1} + \\epsilon_{t} \\), with \\( \\epsilon_{t}\\sim \\mathcal{N}(0,1) \\). The training loss includes predicting noises for different steps," + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.242, + 0.892, + 0.276 + ], + "angle": 0, + "content": "\\[\nL = \\sum_ {x \\sim \\mathcal {D}} \\sum_ {t \\sim [ 0, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Recent work identified a semantic space in the diffusion model, the activations of U-Net's bottleneck layer \\( h \\), as shown in Figure 1. The activations in \\( h \\)-space leads to the generation of a less noised image for the next timestep \\( x_{t - 1} \\). This space exhibits semantic structures and is easy to interpret. Activating a specific vector in the U-Net bottleneck layer leads to the image having a certain attribute. However, existing approaches cannot find the vector for an arbitrarily given concept. Our goal is to find such vectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.893, + 0.838 + ], + "angle": 0, + "content": "To this end, we utilize the text-to-image conditional diffusion model which can generate images from a given text input. The prediction function in Eq. 1 becomes \\(\\epsilon_{\\theta}(x_t,\\pi (y),t)\\) where \\(\\pi (y)\\) is the encodings of the input text \\(y\\). The equation specifies a conditional distribution that drives the generation of the image towards data regions that are highly likely given the input text [12]. To discover an interpretable direction, we leverage the pre-trained model to generate a set of images using dedicated prompts related to that concept. For example, to find the latent direction of the concept \"female\", we first generate a set of images \\(x^{+}\\) with a descriptive prompt \\(y^{+}\\) \"a photo of a female face\". Then, a concept vector is optimized for the conditional generation where the original prompt has been modified into \\(y^{-}\\) \"a photo of a face\", eliminating gender information. The concept vector \\(c\\in \\mathbb{R}^D\\) is randomly initialized in the latent space, where \\(D\\) is the dimension of \\(h\\)-space, and is optimized to minimize the reconstruction error. Since the pre-trained diffusion model is frozen, the model has to utilize the extra condition \\(c\\) to compensate for the missing information not in the text condition but in the image. The concept vector \\(c\\) will be forced to represent the missing information from the input text to produce an image with the lowest reconstruction error. After convergence, that vector \\(c\\) is expected to represent the gender information \"female\". In this way, we discover a set of vectors that represent target concepts, such as gender, safety, and facial expressions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.851, + 0.894, + 0.901 + ], + "angle": 0, + "content": "1The decoding of \\( x_{t-1} \\) depends on other variables due to the presence of skip connections. For simplicity, we omit this consideration as the skip connection seems less significant in encoding compact semantic information, as supported in previous findings [9, 14]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12008" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.088, + 0.45, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.471, + 0.295 + ], + "angle": 0, + "content": "Figure 2. Fair Generation. Top: images generated from the prompt \"doctor\" are biased toward males. Bottom: we sample a learned male or female concept with equal probability for generating the doctors. The doctors now have fair gender. Images are generated from different random seeds." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.323, + 0.449, + 0.339 + ], + "angle": 0, + "content": "Formally, the optimal \\( c^* \\) for a given concept is found by" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.35, + 0.469, + 0.385 + ], + "angle": 0, + "content": "\\[\nc ^ {*} = \\arg \\min _ {c} \\sum_ {x, y \\sim \\mathcal {D}} \\sum_ {t \\sim [ 0, T ]} \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t} ^ {+}, t, \\pi \\left(y ^ {-}\\right), c\\right) \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.469, + 0.504 + ], + "angle": 0, + "content": "\\(x_{t}^{+}\\) denotes the noised version of the original image generated with \\(y^{+}\\), \\(c\\) represents the target concept. \\(\\epsilon_{\\theta}\\) denotes the U-Net that linearly adds an additional concept vector \\(c\\) to its \\(h\\)-space, at each decoding timestep. Regarding implementation, the \\(h\\)-space is the flattened activations after the middle bottleneck layer of the U-Net. The pseudo-code for this training pipeline is in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.745 + ], + "angle": 0, + "content": "We learn a single vector for each concept for all timesteps, as the latent direction remains approximately consistent across different timesteps [18]. Moreover, we restrict the operation to linearity to demonstrate the power of this latent space. Notably, the learned vector generalizes effectively to new images [9] and diverse prompts [30]. For instance, a \"male\" concept learned with the base prompt \"person\" can be used in different contexts, such as \"doctor\" or \"manager\", as shown in the next section. Additionally, the concepts can be optimized jointly or independently, with the experimental section demonstrating the impact of concept composition. A key strength of our approach is utilizing synthesis by diffusion models to collect data, eliminating the need for human labeling and training of guiding classifiers. Nevertheless, our approach can be applied to realistic datasets with annotated attributes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.469, + 0.788 + ], + "angle": 0, + "content": "3.2. Responsible Generation with Self-discovered Interpretable Latent Direction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.84 + ], + "angle": 0, + "content": "In this subsection, we utilize the identified directions to manipulate the latent activation in the latent space for fair, safe, and enhanced responsible generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Fair Generation Method A text prompt contains words that lead to the generation of biased societal groups. We aim to generate images with evenly distributed attributes for a given text prompt. For example, for the prompt \"doctor\"," + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.088, + 0.872, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.341 + ], + "angle": 0, + "content": "Figure 3. Safe Generation. When the user's prompt contains implicit references to nudity, the original model (shown in the top row) generates an inappropriate image, as the added blurriness indicates. In contrast, our approach generates an image for the same prompt by setting a safety-related concept in \\( h \\)-space, identified in the previous section. The vector anti-sexual concept represents the direction to suppress nudity content, effectively eliminating inappropriate content while maintaining fidelity to the prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.532 + ], + "angle": 0, + "content": "we aim to generate an image of a male doctor with a \\(50\\%\\) probability and a female doctor with a \\(50\\%\\) probability. For that, we learn a set of semantic concepts representing different societal groups using the approach in the previous section. For inference, a concept vector is sampled from the learned concepts in the societal group with equal probability, e.g., the C-male and C-female concept vectors for gender are chosen with fair chance. The inference process is fixed as before, except that the sampled vector is added to the original activations in \\(h\\)-space at each decoding step, denoted by" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.543, + 0.891, + 0.559 + ], + "angle": 0, + "content": "\\[\nh \\leftarrow h + c \\sim \\operatorname {C a t e g o r i c a l} \\left(p _ {k}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.689 + ], + "angle": 0, + "content": "where \\( p_k \\) represents the probability of sampling a particular attribute \\( c_k \\) from the societal group with \\( C \\) distinct attributes. For the fair generation, \\( p_k = 1 / C \\). Guided by this sampled concept vector, the generated image is expected to be a male doctor if the C-male concept is sampled or a female doctor otherwise. This allows the generated images to have an equal number of attributes, e.g., an equal number of male and female doctors, shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Safe Generation Method For safety generation, we consider text prompts that contain explicit or implicit references to inappropriate content, which we aim to eliminate. An example of such a prompt is illustrated in Figure 3, where the phrase \"a gorgeous woman\" may indirectly lead to the generation of nudity. We identify a collection of safety-related concepts, such as anti-sexual, to achieve safe generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Specifically, we learn the opposite latent direction of an inappropriate concept, leveraging the negative prompt technique. For instance, the training images are generated by the prompt \\( y^{+} \\) \"a gorgeous person\" with a negative prompt \"sexual\", which effectively instructs the Stable Diffusion to generate safe images without sexual content. The concept vector is then optimized on those training images that depict" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12009" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "safe content. For that, the input prompt \\( y^{-} \\) is set to \"a gorgeous person\" but without the negative prompt \"sexual\". In this way, the concept vector directly learns the concept of \"anti-sexual\". The reason for adopting this strategy is the difficulty of listing all the opposite concepts of sexuality, e.g., \"dressed\", \"clothes\", or more. An alternative approach is to learn the concept of sexuality directly and apply negation during generation, which we found less effective. More details regarding the negative prompt are in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.273 + ], + "angle": 0, + "content": "After the learning process, we maintain all aspects of inference unchanged, except for adding the learned vector to the original activations at the bottleneck layer, formally as" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.281, + 0.47, + 0.297 + ], + "angle": 0, + "content": "\\[\nh \\leftarrow h + c _ {s}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.544 + ], + "angle": 0, + "content": "Here, \\( c_{s} \\) refers to a safety concept, such as \"anti-sexual\", which represents the opposite of sexual content. This strengthens the expression of safe concepts in the generated images so they are devoid of harmful content. Figure 3 illustrates the impact of including the anti-sexual vector, resulting in a visually appealing person with appropriate clothes. Responsible Text-enhancing Generation Method Even when a prompt is intentionally designed to promote safety, the generative models may struggle to accurately incorporate all the concepts defined in the prompt. For instance, consider a text prompt like \"an exciting Halloween party, no violence\". The generative model may encounter difficulties in faithfully representing each responsible concept from the prompt, e.g., its poor understanding of negation on \"violence\" may result in the generation of inappropriate content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.666 + ], + "angle": 0, + "content": "To address this issue, we utilize our self-discovery approach to learn concepts such as gender, race, and safety. To enhance the generation of responsible prompts, we extract safety-related content from the text and leverage our learned ethical-related concepts to reinforce the expression of desired visual features. During inference, we apply the extracted concepts \\( c(y) \\) from the prompt to the original activations, denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.675, + 0.469, + 0.691 + ], + "angle": 0, + "content": "\\[\nh \\leftarrow h + c (y) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.82 + ], + "angle": 0, + "content": "For example, as shown in Figure 4, the concept of \"no violence\" from the text prompt activates our learned \"anti-violence\" concept during inference. By directly manipulating the semantic space, our approach introduces the desired attributes to the generated image. Compared to the original generated image, the anti-violence concept effectively mitigates the presence of violent content and makes the generated images more appropriate." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.21, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we present the performance of our methods on fair generation, safe generation, and enhanced responsible generation. Additionally, we demonstrate the general" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.088, + 0.872, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Responsible text-enhancing generation. The original model may fail to capture the safety concepts specified in the text, such as \"no violence\". We propose extracting those safety concepts from the given prompt and activating the safety directions during generation. The bottom image demonstrates that incorporating our safety concepts can enhance the text guidance of the original prompt." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.344, + 0.892, + 0.374 + ], + "angle": 0, + "content": "properties of our discovered concept vectors, highlighting their generalization capability and compositionality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.662, + 0.403 + ], + "angle": 0, + "content": "4.1. Fair Generation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Task The goal is to increase the diversity of societal groups in the generated images, particularly in professions where existing models exhibit gender and racial bias." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.639 + ], + "angle": 0, + "content": "Dataset Following [6, 29], our approach is evaluated on the Winobias [52] benchmark that comprises 36 professions known to exhibit gender biases. We generate five different prompts for each profession using various templates, such as \"a portrait of a doctor\". Additionally, we create a hard dataset by augmenting the existing dataset with prompts that are more likely to induce stereotypes. This extended dataset adds the term \"successful\" to each original prompt, such as \"a portrait of a successful doctor\". The word \"successful\" often leads to the generation of male images and thus constitutes challenges for a fair generation [6]. Appendix B.1 contains the complete list of prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.894, + 0.87 + ], + "angle": 0, + "content": "Evaluation Metric The CLIP classifier is employed to predict attributes by measuring the similarity between the text embedding of a concept (e.g., female, male) and the embedding of the generated image. We utilize the deviation ratio [6, 29] to quantify the imbalance of different attributes. To accommodate an arbitrary number of attributes, the metric is modified as \\(\\Delta = \\max_{c\\in C}\\frac{|N_c / N - 1 / C|}{1 - 1 / C}\\), where \\(C\\) is the total number of attributes within a societal group, \\(N\\) is the total number of generated images, and \\(N_{c}\\) denotes the number of images whose maximum predicted attribute equals \\(c\\). In particular, we test the gender, male, female, and racial, black, white, Asian, biases associated with the professions. These races are selected as the CLIP classifier has relatively reliable predictions on these attributes. During the evaluation, 150 images were generated for each profession." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Approach Setting In all experiments, we use the Stable Diffusion v1.4 checkpoint and set the guidance scale to 7.5" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.478, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.217, + 0.354, + 0.229 + ], + "angle": 0, + "content": "(a) Images generated by \\(\\mathbf{SD}\\)" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.091, + 0.89, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.217, + 0.772, + 0.229 + ], + "angle": 0, + "content": "(b) Images generated by Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.241, + 0.893, + 0.271 + ], + "angle": 0, + "content": "Figure 5. Gender fairness generation. From the prompt \"a photo of a doctor\", the original SD exhibits significant gender bias, as shown on the left side. Our approach with uniformly sampled gender vectors represents genders equally in the generated images." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.283, + 0.875, + 0.414 + ], + "angle": 0, + "content": "
Dataset\nMethodGenderGender+\nUCERaceRace+\nUCE
SDUCEOursSDUCEOursSDUCEOursSDUCEOurs
Analyst0.700.200.020.540.040.020.820.290.240.770.200.41
CEO0.920.280.060.900.580.060.380.130.220.310.080.22
Laborer1.000.090.120.980.080.140.330.400.240.530.380.20
Secretary0.640.100.360.920.960.460.370.350.240.550.430.32
Teacher0.300.060.040.480.160.100.510.100.040.260.230.21
Winobias[52]0.680.220.170.700.520.230.560.210.230.480.350.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.424, + 0.893, + 0.482 + ], + "angle": 0, + "content": "Table 1. Fair generation quantified by the deviation ratio \\( \\left( {0 \\leq \\Delta \\leq 1}\\right) \\) . Lower values indicate better performance. The left side of the table presents the results on gender attributes, whereas the right side quantifies the racial bias. \"Gender+/Race+\" refers to the extended Winobias dataset, which is more challenging, as described in Subsection 4.1. Our approach leads to unbiased generation for biased prompts and is robust to diverse sources of bias in the prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.702 + ], + "angle": 0, + "content": "for text-to-image generation. We find five concept vectors using a base prompt \"person\", e.g., \\( y^{+} = \\) \"a photo of a woman\" and \\( y^{-} = \\) \"a photo of a person\" to learn the concept \"female\". The concept vectors are optimized for 10K steps on 1K synthesized images for each concept. During inference, we directly employ the learned vector without any scaling. Unlike the baseline approach UCE [6], which needs to debias each profession in Winobias, Our approach is trained solely on the \"person\" prompt to learn the male and female concept that generalizes to all different professions. For comparison, we report UCE's published scores when available and otherwise use their released code to train the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results and Analysis Table 1 reveals that our approach is significantly better than the original SD and outperforms the state-of-the-art debiasing approach UCE. The professions on the table are randomly selected from the complete list of 36 professions (see Appendix B.2). Figure 5 compares images generated from our approach and those from the original SD. Further, we highlight the generalization capability of our approach to different text prompts using the extended Winobias dataset, as shown in the second and fourth column blocks of Table 1. Despite the presence of bias in the text prompts, our approach consistently performs well as it directly operates on the latent visual space. In contrast, UCE performs poorly on this challenging dataset, as it relies on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.584 + ], + "angle": 0, + "content": "debiasing each word in the prompt. The effectiveness of UCE is easily weakened by biased words that are not included in its training set. Additionally, we demonstrate that the quality of images generated by our approach remains consistent with the original SD and UCE in Appendix B.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.661, + 0.608 + ], + "angle": 0, + "content": "4.2. Safe Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Task This section focuses on generating images that eliminate harmful content specified in inappropriate prompts. As an orthogonal approach to existing methods, our approach is combined with current safety methods, including SLD [39] and ESD [5], to eliminate inappropriate generation further." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Dataset and Evaluation Metric The I2P benchmark [39] is a collection of 4703 inappropriate prompts from real-world user prompts. The inappropriateness covers seven categories, including, e.g., illegal activity, sexual, and violence. For evaluation, the Nudenet detector and Q16 classifier are used to detect nudity or violent content in an image. An image is classified as inappropriate if any of the classifiers predicts a positive [5]. Five images are generated for each prompt for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.831, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Approach Setting We find that optimizing a single concept vector for \"safety\" is challenging. Therefore, we learn the concept vector for each inappropriate concept defined in the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.818, + 0.9 + ], + "angle": 0, + "content": "2https://github.com/notAI-tech/NudeNet" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12011" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.896, + 0.204 + ], + "angle": 0, + "content": "
CategoryHarassmentHateIllegalSelf-harmSexualShockingViolenceI2P[39]
SD0.34 ±0.0190.41 ±0.0320.34 ±0.0180.44 ±0.0190.38 ±0.0160.51 ±0.0170.44 ±0.0180.41 ±0.007
Ours-SD0.18 ±0.0150.29 ±0.0300.23 ±0.0160.28 ±0.0170.22 ±0.0140.36 ±0.0170.30 ±0.0170.27 ±0.006
SLD[39]0.15 ±0.0140.18 ±0.0250.17 ±0.0150.19 ±0.0150.15 ±0.0120.32 ±0.0160.21 ±0.0150.20 ±0.006
Ours-SLD0.14 ±0.0140.20 ±0.0270.14 ±0.0130.14 ±0.0130.09 ±0.0100.25 ±0.0150.16 ±0.0130.16 ±0.005
ESD[5]0.27 ±0.0180.32 ±0.0310.33 ±0.0180.35 ±0.0180.18 ±0.0130.41 ±0.0170.41 ±0.0180.32 ±0.007
Ours-ESD0.26 ±0.0170.29 ±0.0300.25 ±0.0170.26 ±0.0170.13 ±0.0110.41 ±0.0170.30 ±0.0170.27 ±0.006
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.214, + 0.895, + 0.258 + ], + "angle": 0, + "content": "Table 2. The proportion of images classified as inappropriate on the I2P benchmark. In each block of results, the first row shows the performance of the original method, while the second row represents adding our concept vector to the corresponding baseline model. Our identified safety-related vector can be combined with existing safety approaches to mitigate inappropriate content generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.47, + 0.465 + ], + "angle": 0, + "content": "I2P dataset, e.g., \"anti-sexual\". The identified concept vectors are linearly combined as the final vector to modulate the generation process. The effect of linearly combining concept vectors is further discussed in the next section. Additionally, certain concepts are rather abstract and include diverse visual categories, e.g., \"hate\". Adding these concepts improves safety yet at a higher cost of image quality degradation. Therefore, we utilize two prevailing concepts, anti-sexual and anti-violence, which may underline other concepts, as our final model. For baselines, we use SLD [39] and ESD [5] with their released model weights. Detailed settings and hyperparameters are in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.471, + 0.632 + ], + "angle": 0, + "content": "Results and Analysis Table 2 demonstrates the effectiveness of our approach in eliminating inappropriate content. Notably, our safety vector can suppress inappropriate concepts that existing approaches failed to eliminate, as evidenced by the improved performance compared to existing approaches. In particular, we observe significant improvement in eliminating nudity with the combined approach of SLD and our proposed method, with a relative improvement of \\(40\\%\\). Additionally, even after retraining the model weights to eliminate nudity, ESD still generates inappropriate images, which our approach can mitigate." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.646, + 0.411, + 0.663 + ], + "angle": 0, + "content": "4.3. Enhancing Responsible Text Guidance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Task For user prompts classified as responsible text, we aim to accurately represent the responsible phrases in the prompt in the generated image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Dataset We created a dataset of 200 prompts that explicitly include responsible concepts. In particular, the dataset focuses on gender and race fairness, as well as the removal of sexual and violent content. For instance, the prompt \"a fair-gender doctor is operating a surgery\" assesses the model's ability to generate images with balanced genders based on the text cue \"fair-gender\". Additionally, prompts such as \"a picture of a loved couple, without sexual content\" address safety concerns. More dataset details are in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Results and Analysis We utilized previously introduced metrics for evaluation, i.e., the deviation ratio for fairness, NudeNet for sexual content, and Q16 for violent content." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.28, + 0.882, + 0.332 + ], + "angle": 0, + "content": "
ModelGenderRaceSexualViolence
SD0.21990.16000.43000.4551
Ours-SD0.14330.13990.26400.3204
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.336, + 0.895, + 0.38 + ], + "angle": 0, + "content": "Table 3. For prompts containing responsible concepts, the original SD may fail to follow the prompts faithfully. Our approach effectively enhances responsible text-guidance generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.527 + ], + "angle": 0, + "content": "For evaluation, 3500 images are generated for the dataset. For our approach, we provide the model with the corresponding concept associated with the input prompt. For example, if a prompt mentions \"no sexual\", the anti-sexual vector is added to the generation process. Table 3 compares our approach with the original model, which does not use the safety concepts. Our approach effectively enhances the text guidance for responsible instructions." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.541, + 0.892, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.701, + 0.895, + 0.758 + ], + "angle": 0, + "content": "Figure 6. Concept interpolation. The first column displays the original image generated by SD. The following columns show images generated by the same random seeds, but with concept vector scales linearly increasing from 0.2 to 0.8." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.685, + 0.788 + ], + "angle": 0, + "content": "4.4. Semantic Concepts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.856 + ], + "angle": 0, + "content": "In previous experiments, we have demonstrated the specific applications of our identified concept vectors for responsible generation. This subsection introduces the general properties of discovered vectors related to the semantic space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Interpolation Figure 6 illustrates the impact of manipulating image semantics by linearly controlling the strength of the concept vector, denoted as \\(\\lambda\\) in the equation \\(h\\gets h + \\lambda c\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12012" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.472, + 0.288 + ], + "angle": 0, + "content": "Figure 7. Multiple concepts composition. The concept vectors of gender, age, and race were learned independently. Linearly adding latent vectors can generate images with corresponding semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.418 + ], + "angle": 0, + "content": "The image is gradually modified to the introduced concept by adjusting the added vector's strength. The smooth transition indicates that the discovered vector represents the target semantic concept while remaining approximately disentangled from other semantic factors. Appendix E.1 presents more examples of concept manipulation and enhanced fidelity by post-hoc interpolation methods [25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.473, + 0.513 + ], + "angle": 0, + "content": "Composition Figure 7 showcases the composability of learned concept vectors, which were trained independently. Images are generated from the prompt \"a photo of a doctor\". By linearly combining these concept vectors, we can control the corresponding attributes in the generated image. Appendix E.2 provides a quantitative evaluation." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.526, + 0.437, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.675, + 0.472, + 0.734 + ], + "angle": 0, + "content": "Figure 8. General semantic concepts identified by our approach. Top: The concept of \"running\" is learned from dog images and can be generalized to different objects. Bottom: The concept vector of \"glasses\" enhances the prompt \"an elephant wearing glasses\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Generalization Figure 8 illustrates the generalization capability of our discovered concept vector to universal semantic concepts. We train the latent vector for the concept \"running\" on generated dog images and test its effect on other objects using prompts such as \"a photo of a cat\". Each pair of images in the figure is generated from the same random seed. Although the vector of \"running\" was learned from dogs, it successfully extends to different animals and even humans. Additionally, our approach enhances the original text guidance for the prompt \"an ele" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.089, + 0.891, + 0.142 + ], + "angle": 0, + "content": "
ModelSDSLD*SLDESD*ESDOurs-SD
FID14.0916.9019.3513.6815.3615.98
CLIP31.33-30.41-30.0531.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.151, + 0.894, + 0.236 + ], + "angle": 0, + "content": "Table 4. Evaluation of the quality of generated images on the COCO-30K [21] dataset using FID for image fidelity and CLIP Score for semantic alignment with input text. Various safety approaches have approximately the same level of image quality as the original SD. Numbers reported from the corresponding papers are denoted with * ." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.338 + ], + "angle": 0, + "content": "phant wearing glasses\". The original SD cannot produce accurate images, as shown in the first and third images on the bottom. The correct images can be generated by adding the concept vector \"glasses\" in \\( h \\)-space. More visualizations are in Appendix E.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.893, + 0.448 + ], + "angle": 0, + "content": "Impact on Image Quality Additionally, we find that the quality of generated images remains approximately the same level as the original SD, as shown in Table 4. The observed differences in the reported scores and these in our experiments can be attributed to the randomness during image generation and caption sampling, which aligns with the inconsistencies reported in other studies [5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.894, + 0.542 + ], + "angle": 0, + "content": "Sensitivity to Hyperparameters In Appendix F, we investigate the sensitivity of our approach to hyperparameters, finding that it is less affected by factors such as the number of training images or different input prompts. Additionally, we demonstrate that our approach can leverage existing datasets to discover concept vectors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.62, + 0.57 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In this study, we introduced a self-discovery approach to identify semantic concepts in the latent space of text-to-image diffusion models. Our research findings highlight that the generation of inappropriate content can be attributed to ethical-related concepts present in the internal semantic space of diffusion models. Leveraging these concept vectors, we enable responsible generation, including promoting equality among societal groups, eliminating inappropriate content, and enhancing text guidance for responsible prompts. Through extensive experiments, we have demonstrated the effectiveness and superiority of our proposed approach. Our work contributes to the understanding of internal representations in diffusion models and facilitates the generation of responsible content, maximizing the utility of high-quality text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1, EPSRC/MURI grant: EP/N019474/1. We thank the Royal Academy of Engineering and FiveAI. This work is also founded by the German Federal Ministry of Education and Research and the Bavarian State Ministry for Science and the Arts." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12013" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.299 + ], + "angle": 0, + "content": "[3] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[4] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/ logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[5] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.471, + 0.512 + ], + "angle": 0, + "content": "[6] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.514, + 0.471, + 0.569 + ], + "angle": 0, + "content": "[7] Jindong Gu, Ahmad Beirami, Xuezhi Wang, Alex Beutel, Philip Torr, and Yao Qin. Towards robust prompts on vision-language models. arXiv preprint arXiv:2304.08479, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.471, + 0.64 + ], + "angle": 0, + "content": "[8] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.642, + 0.471, + 0.697 + ], + "angle": 0, + "content": "[9] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. arXiv preprint arXiv:2303.11073, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.7, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[10] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. arXiv preprint arXiv:2305.10120, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.743, + 0.471, + 0.798 + ], + "angle": 0, + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.8, + 0.471, + 0.842 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.844, + 0.471, + 0.885 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.887, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.895, + 0.319 + ], + "angle": 0, + "content": "[17] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[19] Yipeng Leng, Qiangjuan Huang, Zhiyuan Wang, Yangyang Liu, and Haoyu Zhang. Diffusegae: Controllable and high-fidelity image manipulation from disentangled representation. arXiv preprint arXiv:2307.05899, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[20] Hang Li, Jindong Gu, Rajat Koner, Sahand Sharifzadeh, and Volker Tresp. Do dall-e and flamingo understand each other? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1999-2010, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[22] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[23] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In The Twelfth International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[24] Avery Ma, Amir-massoud Farahmand, Yangchen Pan, Philip Torr, and Jindong Gu. Improving adversarial transferability via model alignment. arXiv preprint arXiv:2311.18495, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12014" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Minheng Ni, Chenfei Wu, Xiaodong Wang, Shengming Yin, Lijuan Wang, Zicheng Liu, and Nan Duan. Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[28] Zixuan Ni, Longhui Wei, Jiacheng Li, Siliang Tang, Yueting Zhuang, and Qi Tian. Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8900-8909, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[29] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7053-7061, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[30] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.399, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[32] Vinay Uday Prabhu and Abeba Birhane. Large image datasets: A pyrrhic win for computer vision? arXiv preprint arXiv:2006.16923, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[33] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[36] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[38] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350–1361, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.858, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[40] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.162, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[43] Matthew Trager, Pramuditha Perera, Luca Zancato, Alessandro Achille, Parminder Bhatia, and Stefano Soatto. Linear spaces of meanings: Compositional structures in vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15395-15404, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[44] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[45] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. arXiv preprint arXiv:2306.08757, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.51, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[46] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[47] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[48] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[50] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in Neural Information Processing Systems, 35:22117-22130, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12015" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.148 + ], + "angle": 0, + "content": "[52] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. arXiv preprint arXiv:1804.06876, 2018. 5, 6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12016" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_origin.pdf b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7178bfab514338eb0c73178e98b4c1225b1a8716 --- /dev/null +++ b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/3095ee86-791c-49b2-8a2e-753d19e700c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2be7c7f6752e03f3b28776eaf3324879a8a8a67b825da99fcc03e794dfe50d8c +size 5608460 diff --git a/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/full.md b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ecd9c1cd459777b6ce2e2791d7fabd607eaed892 --- /dev/null +++ b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/full.md @@ -0,0 +1,285 @@ +# Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation + +Hang Li $^{1,4,5}$ Chengzhi Shen $^{3}$ Philip Torr $^{2}$ Volker Tresp $^{1,4}$ Jindong Gu $^{2*}$ + +$^{1}$ LMU Munich, Germany $^{2}$ University of Oxford, UK $^{3}$ Technical University of Munich, Germany $^{4}$ Munich Center for Machine Learning, Germany $^{5}$ Siemens AG, Germany + +# Abstract + +Diffusion-based models have gained significant popularity for text-to-image generation due to their exceptional image-generation capabilities. A risk with these models is the potential generation of inappropriate content, such as biased or harmful images. However, the underlying reasons for generating such undesired content from the perspective of the diffusion model's internal representation remain unclear. Previous work interprets vectors in an interpretable latent space of diffusion models as semantic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation. Extensive experiments have been conducted to verify the effectiveness of our mitigation approach, namely, for fair generation, safe generation, and responsible text-enhancing generation. Project page: https://interpretdiffusion.github.io. + +# 1. Introduction + +The rapid advances in vision language models have sparked increasing interest in ensuring their safety and responsible use [7, 23, 24]. In particular, text-to-image diffusion models, which have exhibited remarkable performance in creating images from text prompts [13, 15, 20, 31, 35, 37, 49], raise concerns about the risks of generating inappropriate content. The generated images may exhibit biases and unsafe elements, including instances of gender discrimination or the depiction of violent scenes that could be harmful to children. Recent research efforts have focused on introducing safety mechanisms to mitigate these issues, such as filtering out inappropriate text input, detecting inappropriate images with a safety guard classifier [4, 32, 36, 38] and + +building safe diffusion models [5, 6, 17]. However, the underlying mechanism of how diffusion models generate inappropriate content remains poorly understood. In this work, we aim to explore the following questions. 1) Are there any internal representations associated with these inappropriate concepts in the diffusion model-based generation process? 2) Can we manipulate representations to avoid inappropriate content corresponding to a given concept, i.e., to achieve responsible image generation? + +To understand the image generation process of diffusion models, previous work has identified the bottleneck layer of the U-Net as a semantic representation space, dubbed $h$ -space [18]. They demonstrated that a vector in the $h$ -space can be associated with a specific semantic concept in the generated image. Manipulating the vector in the space can alter the generated image in a semantically meaningful way, such as adding a smile to a face. Several approaches [9, 18, 30] have been proposed to discover meaningful directions in this $h$ -space. For instance, an approach in [9] uses PCA to identify a set of latent directions that may represent semantic concepts. + +However, existing approaches to identifying interpretable latent vectors are limited. In unsupervised approaches [9, 30], it is not clear to which semantic concepts those identified vectors correspond. The found vectors must be interpreted with humans in a loop. Furthermore, the number of interpretable directions depends on the training data [9, 30]. It is highly likely that some target concepts may not be found in the discovered directions, especially those related to fairness and safety. Supervised approaches [9, 18] have also been explored to identify target concepts. These methods require training external attribute classifiers supervised by human annotations. Additionally, the quality of the identified vectors is sensitive to the classifier's performance. Furthermore, new concepts require the training of new classifiers. Overall, existing interpretation methods cannot be easily applied to identify the corresponding semantic vector for a given inappropriate concept. + +In this work, we propose a self-discovery approach to find interpretable latent directions in the $h$ -space for user + +defined concepts. We learn a latent vector that effectively represents the concept by leveraging the model's acquired semantic knowledge in its internal representations. Initially, images are generated using specific text prompts related to the concept. The images are then used in a de-noising process where the frozen pretrained diffusion model reconstructs these images from noise, guided by a modified text prompt that omits the desired concept, and our introduced latent vector. By minimizing the reconstruction loss, the vector learns to represent the given concept. Our self-discovery approach eliminates the need for external models like CLIP text encoder [34] or dedicated attribute classifiers trained on human-labeled datasets. We identify ethical-related latent vectors and demonstrate their applications in responsible text-to-image generation: 1) fairness by sampling an ethical concept, e.g., gender, in the latent space, which generates images with unbiased attributes and aligned with the prompt. 2) safety generation by incorporating safety-related concepts, e.g., one that eliminates the nudity content, into $h$ -space to prevent the model from generating such harmful content. 3) responsible guidance, where we first discover responsible concepts in the text prompt and enhance the expression of those ethical concepts. + +Previous approaches enhance responsible image generation from different perspectives. Concretely, [2, 5, 6, 8, 17] fine-tune the diffusion models or text embeddings to unlearn harmful concepts, and [39] applies classifier-free guidance to steer the generation away from unsafe concepts. Despite the mitigation mechanism of previous approaches, diffusion models still suffer from inappropriate content generation [39, 50]. Unlike previous work, in this work, we provide a new perspective to mitigate the inappropriate generation, namely, finding and manipulating concepts in an interpretable latent space. Our work can be easily combined with previous mitigation approaches to further enhance responsible text-to-image generation. + +We conducted extensive experiments on fairness, safety, and responsible guidance-enhancing generation. Our model consistently produces images with a balanced representation across societal groups. Further, we successfully mitigate harmful content for inappropriate prompts. In addition, our approach synergistically improves the performance of responsible image generation when combined with existing methods. Furthermore, we enhance text guidance to generate fair and safe content for responsible prompts. + +Our contributions can be summarized as follows: + +- We propose a self-discovery method for identifying interpretable directions in the diffusion latent space. Our approach can find a vector that represents any desired concept, without the need for labeled data or external models. +- With the discovered vectors, we propose a straightforward yet effective approach to enhance responsible generation, including fair generation, safe generation, and re + +sponsored text-enhancing generation. +- Extensive experiments are conducted to validate the effectiveness of our approach. + +# 2. Related Work + +Responsible Alignment of Diffusion Models Various approaches have been proposed to mitigate the generation of biased and unsafe content in diffusion models. A straightforward method involves refining the training dataset to remove biased and inappropriate content, exemplified by Stable Diffusion (SD) v2 [37]. Such approaches can be computationally intensive, may not fully eliminate harmful content [5], and could degrade the model's performance [39]. An alternative is to detect and filter out inappropriate words from the input prompts [1, 2, 27]. However, this fails to address non-explicit phrases that can still yield inappropriate outputs. Another line of approaches involves finetuning the parameters of pretrained models, aiming to remove the model's representation capability of generating such inappropriate concepts [5, 17]. However, they are sensitive to the adaptation process and may result in the degradation of the original models [6, 10, 28, 29, 48]. Moreover, such approaches require a potentially exhaustive list of words that introduce biases and harmful concepts [5, 6, 29]. Training-free approaches utilize classifier-free guidance to direct the generated images away from undesirable content during inference [1, 3, 39, 39, 47]. While they modify the noise space using text-based guidance through cross-attention mechanisms, we adopt a similar conditioning strategy to manipulate the generation for frozen pretrained models in the semantic latent space. As an orthogonal approach to the existing literature, we mitigate the inappropriate content by finding the corresponding latent directions in the U-Net bottleneck layer and suppressing their activations. + +Interpreting Diffusion Models To understand the working mechanisms of diffusion models, recent works mainly focus on investigating text guidance for conditional diffusion models [11, 16, 22, 26, 29, 43, 46], or analyzing the internal representations in diffusion models' intermediate layer activations [9, 18, 30, 40, 44]. We focus on elucidating the internal representations learned within the diffusion model, in line with prior works [18]. Some work [19, 33, 45, 51] proposes to create a semantic space in diffusion models by employing an autoencoder to encode the image into a semantic vector that guides the decoding process. However, their approaches require adapting the parameters of the autoencoder or even the entire framework. + +The seminal work [18] reveals that the bottleneck layer of U-Net architecture already exhibits properties suitable for a semantic representation space. They identified disentangled representations associated with the semantics of the generated image and demonstrated that those latent di + +![](images/bb13d88ae470191c1fc694af0e667ae33a32f125a035adac5eec27731b720cf6.jpg) +Figure 1. Optimization framework to discover a semantic vector for a given concept. The top line shows that an image is firstly generated by the pretrained Stable Diffusion model for the prompt "a female face". The bottom part shows the optimization process for finding the concept for "female" in the semantic $h$ -space. The concept vector is used to reconstruct the image along with a modified prompt "a face", under an iterative denoising process. With the pretrained diffusion model frozen, the gradients of the reconstruction loss can solely update the latent vector to represent the missing gender information. After convergence, the latent vector is aligned with the U-Net's internal representation of the "female" concept, which can be used to guide new image generation. + +rections are identical to different images. However, their approach relies on the CLIP classifier and paired source-target images and edits, making it inefficient. Another work proposes a PCA-based decomposition method on the latent space and finds interpretable attribute directions using the top right-hand singular vectors of the Jacobian. Additionally, [30] uses Riemannian metrics to define more accurate and meaningful directions. However, these approaches require manual interpretation to identify the editing effect of each component. Our approach differs from the supervised approach in [9] by enabling the efficient discovery of latent directions for any given target concept without requiring a data collection process or training external classifiers. + +# 3. Approach + +This section first introduces our optimization method to find interpretable directions in diffusion models' $h$ -space. In the second part, we show how to utilize discovered concepts in the inference process for responsible generation, including fairness, safety, and text-enhancing generation. + +# 3.1. Finding a Semantic Concept + +Diffusion models are generative models that generate samples from Gaussian noise through a denoising process [13, 41, 42]. Starting from a random vector $x_{T} \sim \mathcal{N}(0,1)$ of the + +same dimension as the image, the model estimates a noise value at each time step to subtract from the current vector to obtain a denoised image, denoted as $x_{t - 1} = x_t - \epsilon_\theta (x_t,t)$ , where $\epsilon_{\theta}$ represents the U-Net of the diffusion model. A clean image $x_0$ is obtained at the end of this denoising process. The training of diffusion models involves a forward process that iteratively adds noise to images from the data, denoted as $x_{t} = x_{t - 1} + \epsilon_{t}$ , with $\epsilon_{t}\sim \mathcal{N}(0,1)$ . The training loss includes predicting noises for different steps, + +$$ +L = \sum_ {x \sim \mathcal {D}} \sum_ {t \sim [ 0, T ]} \| \epsilon - \epsilon_ {\theta} (x _ {t}, t) \| ^ {2}. \tag {1} +$$ + +Recent work identified a semantic space in the diffusion model, the activations of U-Net's bottleneck layer $h$ , as shown in Figure 1. The activations in $h$ -space leads to the generation of a less noised image for the next timestep $x_{t - 1}$ . This space exhibits semantic structures and is easy to interpret. Activating a specific vector in the U-Net bottleneck layer leads to the image having a certain attribute. However, existing approaches cannot find the vector for an arbitrarily given concept. Our goal is to find such vectors. + +To this end, we utilize the text-to-image conditional diffusion model which can generate images from a given text input. The prediction function in Eq. 1 becomes $\epsilon_{\theta}(x_t,\pi (y),t)$ where $\pi (y)$ is the encodings of the input text $y$ . The equation specifies a conditional distribution that drives the generation of the image towards data regions that are highly likely given the input text [12]. To discover an interpretable direction, we leverage the pre-trained model to generate a set of images using dedicated prompts related to that concept. For example, to find the latent direction of the concept "female", we first generate a set of images $x^{+}$ with a descriptive prompt $y^{+}$ "a photo of a female face". Then, a concept vector is optimized for the conditional generation where the original prompt has been modified into $y^{-}$ "a photo of a face", eliminating gender information. The concept vector $c\in \mathbb{R}^D$ is randomly initialized in the latent space, where $D$ is the dimension of $h$ -space, and is optimized to minimize the reconstruction error. Since the pre-trained diffusion model is frozen, the model has to utilize the extra condition $c$ to compensate for the missing information not in the text condition but in the image. The concept vector $c$ will be forced to represent the missing information from the input text to produce an image with the lowest reconstruction error. After convergence, that vector $c$ is expected to represent the gender information "female". In this way, we discover a set of vectors that represent target concepts, such as gender, safety, and facial expressions. + +![](images/838c5aac9ca7fa1df74e1b4ad20b0cfb88c3b151585475b7d9ac40a01a6349fd.jpg) +Figure 2. Fair Generation. Top: images generated from the prompt "doctor" are biased toward males. Bottom: we sample a learned male or female concept with equal probability for generating the doctors. The doctors now have fair gender. Images are generated from different random seeds. + +Formally, the optimal $c^*$ for a given concept is found by + +$$ +c ^ {*} = \arg \min _ {c} \sum_ {x, y \sim \mathcal {D}} \sum_ {t \sim [ 0, T ]} \| \epsilon - \epsilon_ {\theta} \left(x _ {t} ^ {+}, t, \pi \left(y ^ {-}\right), c\right) \| ^ {2}, \tag {2} +$$ + +$x_{t}^{+}$ denotes the noised version of the original image generated with $y^{+}$ , $c$ represents the target concept. $\epsilon_{\theta}$ denotes the U-Net that linearly adds an additional concept vector $c$ to its $h$ -space, at each decoding timestep. Regarding implementation, the $h$ -space is the flattened activations after the middle bottleneck layer of the U-Net. The pseudo-code for this training pipeline is in Appendix A.1. + +We learn a single vector for each concept for all timesteps, as the latent direction remains approximately consistent across different timesteps [18]. Moreover, we restrict the operation to linearity to demonstrate the power of this latent space. Notably, the learned vector generalizes effectively to new images [9] and diverse prompts [30]. For instance, a "male" concept learned with the base prompt "person" can be used in different contexts, such as "doctor" or "manager", as shown in the next section. Additionally, the concepts can be optimized jointly or independently, with the experimental section demonstrating the impact of concept composition. A key strength of our approach is utilizing synthesis by diffusion models to collect data, eliminating the need for human labeling and training of guiding classifiers. Nevertheless, our approach can be applied to realistic datasets with annotated attributes. + +# 3.2. Responsible Generation with Self-discovered Interpretable Latent Direction + +In this subsection, we utilize the identified directions to manipulate the latent activation in the latent space for fair, safe, and enhanced responsible generation. + +Fair Generation Method A text prompt contains words that lead to the generation of biased societal groups. We aim to generate images with evenly distributed attributes for a given text prompt. For example, for the prompt "doctor", + +![](images/f362d56a6aae7352e4d2e42b09b3e1058cafd58ca0ce352f396f4b568daf7457.jpg) +Figure 3. Safe Generation. When the user's prompt contains implicit references to nudity, the original model (shown in the top row) generates an inappropriate image, as the added blurriness indicates. In contrast, our approach generates an image for the same prompt by setting a safety-related concept in $h$ -space, identified in the previous section. The vector anti-sexual concept represents the direction to suppress nudity content, effectively eliminating inappropriate content while maintaining fidelity to the prompt. + +we aim to generate an image of a male doctor with a $50\%$ probability and a female doctor with a $50\%$ probability. For that, we learn a set of semantic concepts representing different societal groups using the approach in the previous section. For inference, a concept vector is sampled from the learned concepts in the societal group with equal probability, e.g., the C-male and C-female concept vectors for gender are chosen with fair chance. The inference process is fixed as before, except that the sampled vector is added to the original activations in $h$ -space at each decoding step, denoted by + +$$ +h \leftarrow h + c \sim \operatorname {C a t e g o r i c a l} \left(p _ {k}\right), \tag {3} +$$ + +where $p_k$ represents the probability of sampling a particular attribute $c_k$ from the societal group with $C$ distinct attributes. For the fair generation, $p_k = 1 / C$ . Guided by this sampled concept vector, the generated image is expected to be a male doctor if the C-male concept is sampled or a female doctor otherwise. This allows the generated images to have an equal number of attributes, e.g., an equal number of male and female doctors, shown in Figure 2. + +Safe Generation Method For safety generation, we consider text prompts that contain explicit or implicit references to inappropriate content, which we aim to eliminate. An example of such a prompt is illustrated in Figure 3, where the phrase "a gorgeous woman" may indirectly lead to the generation of nudity. We identify a collection of safety-related concepts, such as anti-sexual, to achieve safe generation. + +Specifically, we learn the opposite latent direction of an inappropriate concept, leveraging the negative prompt technique. For instance, the training images are generated by the prompt $y^{+}$ "a gorgeous person" with a negative prompt "sexual", which effectively instructs the Stable Diffusion to generate safe images without sexual content. The concept vector is then optimized on those training images that depict + +safe content. For that, the input prompt $y^{-}$ is set to "a gorgeous person" but without the negative prompt "sexual". In this way, the concept vector directly learns the concept of "anti-sexual". The reason for adopting this strategy is the difficulty of listing all the opposite concepts of sexuality, e.g., "dressed", "clothes", or more. An alternative approach is to learn the concept of sexuality directly and apply negation during generation, which we found less effective. More details regarding the negative prompt are in Appendix A.2. + +After the learning process, we maintain all aspects of inference unchanged, except for adding the learned vector to the original activations at the bottleneck layer, formally as + +$$ +h \leftarrow h + c _ {s}. \tag {4} +$$ + +Here, $c_{s}$ refers to a safety concept, such as "anti-sexual", which represents the opposite of sexual content. This strengthens the expression of safe concepts in the generated images so they are devoid of harmful content. Figure 3 illustrates the impact of including the anti-sexual vector, resulting in a visually appealing person with appropriate clothes. Responsible Text-enhancing Generation Method Even when a prompt is intentionally designed to promote safety, the generative models may struggle to accurately incorporate all the concepts defined in the prompt. For instance, consider a text prompt like "an exciting Halloween party, no violence". The generative model may encounter difficulties in faithfully representing each responsible concept from the prompt, e.g., its poor understanding of negation on "violence" may result in the generation of inappropriate content. + +To address this issue, we utilize our self-discovery approach to learn concepts such as gender, race, and safety. To enhance the generation of responsible prompts, we extract safety-related content from the text and leverage our learned ethical-related concepts to reinforce the expression of desired visual features. During inference, we apply the extracted concepts $c(y)$ from the prompt to the original activations, denoted as + +$$ +h \leftarrow h + c (y) \tag {5} +$$ + +For example, as shown in Figure 4, the concept of "no violence" from the text prompt activates our learned "anti-violence" concept during inference. By directly manipulating the semantic space, our approach introduces the desired attributes to the generated image. Compared to the original generated image, the anti-violence concept effectively mitigates the presence of violent content and makes the generated images more appropriate. + +# 4. Experiments + +In this section, we present the performance of our methods on fair generation, safe generation, and enhanced responsible generation. Additionally, we demonstrate the general + +![](images/e5b7fe4746436af20cd134c77d7aada804c178c8a0105c8031b89fa9d7efa521.jpg) +Figure 4. Responsible text-enhancing generation. The original model may fail to capture the safety concepts specified in the text, such as "no violence". We propose extracting those safety concepts from the given prompt and activating the safety directions during generation. The bottom image demonstrates that incorporating our safety concepts can enhance the text guidance of the original prompt. + +properties of our discovered concept vectors, highlighting their generalization capability and compositionality. + +# 4.1. Fair Generation + +Task The goal is to increase the diversity of societal groups in the generated images, particularly in professions where existing models exhibit gender and racial bias. + +Dataset Following [6, 29], our approach is evaluated on the Winobias [52] benchmark that comprises 36 professions known to exhibit gender biases. We generate five different prompts for each profession using various templates, such as "a portrait of a doctor". Additionally, we create a hard dataset by augmenting the existing dataset with prompts that are more likely to induce stereotypes. This extended dataset adds the term "successful" to each original prompt, such as "a portrait of a successful doctor". The word "successful" often leads to the generation of male images and thus constitutes challenges for a fair generation [6]. Appendix B.1 contains the complete list of prompts. + +Evaluation Metric The CLIP classifier is employed to predict attributes by measuring the similarity between the text embedding of a concept (e.g., female, male) and the embedding of the generated image. We utilize the deviation ratio [6, 29] to quantify the imbalance of different attributes. To accommodate an arbitrary number of attributes, the metric is modified as $\Delta = \max_{c\in C}\frac{|N_c / N - 1 / C|}{1 - 1 / C}$ , where $C$ is the total number of attributes within a societal group, $N$ is the total number of generated images, and $N_{c}$ denotes the number of images whose maximum predicted attribute equals $c$ . In particular, we test the gender, male, female, and racial, black, white, Asian, biases associated with the professions. These races are selected as the CLIP classifier has relatively reliable predictions on these attributes. During the evaluation, 150 images were generated for each profession. + +Approach Setting In all experiments, we use the Stable Diffusion v1.4 checkpoint and set the guidance scale to 7.5 + +![](images/4434228768e8c00e84057e02eeeb376e4346343c7d3c35a42a7d610801504a9c.jpg) +(a) Images generated by $\mathbf{SD}$ + +![](images/89c656c08923f91da03787a4b3eac60e505e42f572dbe9fb284e3d9c88324882.jpg) +(b) Images generated by Ours +Figure 5. Gender fairness generation. From the prompt "a photo of a doctor", the original SD exhibits significant gender bias, as shown on the left side. Our approach with uniformly sampled gender vectors represents genders equally in the generated images. + +
Dataset +MethodGenderGender+ +UCERaceRace+ +UCE
SDUCEOursSDUCEOursSDUCEOursSDUCEOurs
Analyst0.700.200.020.540.040.020.820.290.240.770.200.41
CEO0.920.280.060.900.580.060.380.130.220.310.080.22
Laborer1.000.090.120.980.080.140.330.400.240.530.380.20
Secretary0.640.100.360.920.960.460.370.350.240.550.430.32
Teacher0.300.060.040.480.160.100.510.100.040.260.230.21
Winobias[52]0.680.220.170.700.520.230.560.210.230.480.350.20
+ +Table 1. Fair generation quantified by the deviation ratio $\left( {0 \leq \Delta \leq 1}\right)$ . Lower values indicate better performance. The left side of the table presents the results on gender attributes, whereas the right side quantifies the racial bias. "Gender+/Race+" refers to the extended Winobias dataset, which is more challenging, as described in Subsection 4.1. Our approach leads to unbiased generation for biased prompts and is robust to diverse sources of bias in the prompt. + +for text-to-image generation. We find five concept vectors using a base prompt "person", e.g., $y^{+} =$ "a photo of a woman" and $y^{-} =$ "a photo of a person" to learn the concept "female". The concept vectors are optimized for 10K steps on 1K synthesized images for each concept. During inference, we directly employ the learned vector without any scaling. Unlike the baseline approach UCE [6], which needs to debias each profession in Winobias, Our approach is trained solely on the "person" prompt to learn the male and female concept that generalizes to all different professions. For comparison, we report UCE's published scores when available and otherwise use their released code to train the model. + +Results and Analysis Table 1 reveals that our approach is significantly better than the original SD and outperforms the state-of-the-art debiasing approach UCE. The professions on the table are randomly selected from the complete list of 36 professions (see Appendix B.2). Figure 5 compares images generated from our approach and those from the original SD. Further, we highlight the generalization capability of our approach to different text prompts using the extended Winobias dataset, as shown in the second and fourth column blocks of Table 1. Despite the presence of bias in the text prompts, our approach consistently performs well as it directly operates on the latent visual space. In contrast, UCE performs poorly on this challenging dataset, as it relies on + +debiasing each word in the prompt. The effectiveness of UCE is easily weakened by biased words that are not included in its training set. Additionally, we demonstrate that the quality of images generated by our approach remains consistent with the original SD and UCE in Appendix B.3. + +# 4.2. Safe Generation + +Task This section focuses on generating images that eliminate harmful content specified in inappropriate prompts. As an orthogonal approach to existing methods, our approach is combined with current safety methods, including SLD [39] and ESD [5], to eliminate inappropriate generation further. + +Dataset and Evaluation Metric The I2P benchmark [39] is a collection of 4703 inappropriate prompts from real-world user prompts. The inappropriateness covers seven categories, including, e.g., illegal activity, sexual, and violence. For evaluation, the Nudenet detector and Q16 classifier are used to detect nudity or violent content in an image. An image is classified as inappropriate if any of the classifiers predicts a positive [5]. Five images are generated for each prompt for evaluation. + +Approach Setting We find that optimizing a single concept vector for "safety" is challenging. Therefore, we learn the concept vector for each inappropriate concept defined in the + +
CategoryHarassmentHateIllegalSelf-harmSexualShockingViolenceI2P[39]
SD0.34 ±0.0190.41 ±0.0320.34 ±0.0180.44 ±0.0190.38 ±0.0160.51 ±0.0170.44 ±0.0180.41 ±0.007
Ours-SD0.18 ±0.0150.29 ±0.0300.23 ±0.0160.28 ±0.0170.22 ±0.0140.36 ±0.0170.30 ±0.0170.27 ±0.006
SLD[39]0.15 ±0.0140.18 ±0.0250.17 ±0.0150.19 ±0.0150.15 ±0.0120.32 ±0.0160.21 ±0.0150.20 ±0.006
Ours-SLD0.14 ±0.0140.20 ±0.0270.14 ±0.0130.14 ±0.0130.09 ±0.0100.25 ±0.0150.16 ±0.0130.16 ±0.005
ESD[5]0.27 ±0.0180.32 ±0.0310.33 ±0.0180.35 ±0.0180.18 ±0.0130.41 ±0.0170.41 ±0.0180.32 ±0.007
Ours-ESD0.26 ±0.0170.29 ±0.0300.25 ±0.0170.26 ±0.0170.13 ±0.0110.41 ±0.0170.30 ±0.0170.27 ±0.006
+ +I2P dataset, e.g., "anti-sexual". The identified concept vectors are linearly combined as the final vector to modulate the generation process. The effect of linearly combining concept vectors is further discussed in the next section. Additionally, certain concepts are rather abstract and include diverse visual categories, e.g., "hate". Adding these concepts improves safety yet at a higher cost of image quality degradation. Therefore, we utilize two prevailing concepts, anti-sexual and anti-violence, which may underline other concepts, as our final model. For baselines, we use SLD [39] and ESD [5] with their released model weights. Detailed settings and hyperparameters are in Appendix C. + +Results and Analysis Table 2 demonstrates the effectiveness of our approach in eliminating inappropriate content. Notably, our safety vector can suppress inappropriate concepts that existing approaches failed to eliminate, as evidenced by the improved performance compared to existing approaches. In particular, we observe significant improvement in eliminating nudity with the combined approach of SLD and our proposed method, with a relative improvement of $40\%$ . Additionally, even after retraining the model weights to eliminate nudity, ESD still generates inappropriate images, which our approach can mitigate. + +# 4.3. Enhancing Responsible Text Guidance + +Task For user prompts classified as responsible text, we aim to accurately represent the responsible phrases in the prompt in the generated image. + +Dataset We created a dataset of 200 prompts that explicitly include responsible concepts. In particular, the dataset focuses on gender and race fairness, as well as the removal of sexual and violent content. For instance, the prompt "a fair-gender doctor is operating a surgery" assesses the model's ability to generate images with balanced genders based on the text cue "fair-gender". Additionally, prompts such as "a picture of a loved couple, without sexual content" address safety concerns. More dataset details are in Appendix D. + +Results and Analysis We utilized previously introduced metrics for evaluation, i.e., the deviation ratio for fairness, NudeNet for sexual content, and Q16 for violent content. + +Table 2. The proportion of images classified as inappropriate on the I2P benchmark. In each block of results, the first row shows the performance of the original method, while the second row represents adding our concept vector to the corresponding baseline model. Our identified safety-related vector can be combined with existing safety approaches to mitigate inappropriate content generation. + +
ModelGenderRaceSexualViolence
SD0.21990.16000.43000.4551
Ours-SD0.14330.13990.26400.3204
+ +Table 3. For prompts containing responsible concepts, the original SD may fail to follow the prompts faithfully. Our approach effectively enhances responsible text-guidance generation. + +For evaluation, 3500 images are generated for the dataset. For our approach, we provide the model with the corresponding concept associated with the input prompt. For example, if a prompt mentions "no sexual", the anti-sexual vector is added to the generation process. Table 3 compares our approach with the original model, which does not use the safety concepts. Our approach effectively enhances the text guidance for responsible instructions. + +![](images/99de2f3204c9905559f8260432c09c4dc6c811f61596c94c03e0b96ae884caae.jpg) +Figure 6. Concept interpolation. The first column displays the original image generated by SD. The following columns show images generated by the same random seeds, but with concept vector scales linearly increasing from 0.2 to 0.8. + +# 4.4. Semantic Concepts + +In previous experiments, we have demonstrated the specific applications of our identified concept vectors for responsible generation. This subsection introduces the general properties of discovered vectors related to the semantic space. + +Interpolation Figure 6 illustrates the impact of manipulating image semantics by linearly controlling the strength of the concept vector, denoted as $\lambda$ in the equation $h\gets h + \lambda c$ + +![](images/dee4388c5d867e1c32ea30c301707d7bc96abb9a306fc1e5e32c8539a0d38d22.jpg) +Figure 7. Multiple concepts composition. The concept vectors of gender, age, and race were learned independently. Linearly adding latent vectors can generate images with corresponding semantics. + +The image is gradually modified to the introduced concept by adjusting the added vector's strength. The smooth transition indicates that the discovered vector represents the target semantic concept while remaining approximately disentangled from other semantic factors. Appendix E.1 presents more examples of concept manipulation and enhanced fidelity by post-hoc interpolation methods [25]. + +Composition Figure 7 showcases the composability of learned concept vectors, which were trained independently. Images are generated from the prompt "a photo of a doctor". By linearly combining these concept vectors, we can control the corresponding attributes in the generated image. Appendix E.2 provides a quantitative evaluation. + +![](images/0ddbae4d8953f1c74dc4dfa3dcaaafec76a56f77dd8b17a0caecdefd33c9b087.jpg) +Figure 8. General semantic concepts identified by our approach. Top: The concept of "running" is learned from dog images and can be generalized to different objects. Bottom: The concept vector of "glasses" enhances the prompt "an elephant wearing glasses". + +Generalization Figure 8 illustrates the generalization capability of our discovered concept vector to universal semantic concepts. We train the latent vector for the concept "running" on generated dog images and test its effect on other objects using prompts such as "a photo of a cat". Each pair of images in the figure is generated from the same random seed. Although the vector of "running" was learned from dogs, it successfully extends to different animals and even humans. Additionally, our approach enhances the original text guidance for the prompt "an ele + +
ModelSDSLD*SLDESD*ESDOurs-SD
FID14.0916.9019.3513.6815.3615.98
CLIP31.33-30.41-30.0531.03
+ +Table 4. Evaluation of the quality of generated images on the COCO-30K [21] dataset using FID for image fidelity and CLIP Score for semantic alignment with input text. Various safety approaches have approximately the same level of image quality as the original SD. Numbers reported from the corresponding papers are denoted with * . + +phant wearing glasses". The original SD cannot produce accurate images, as shown in the first and third images on the bottom. The correct images can be generated by adding the concept vector "glasses" in $h$ -space. More visualizations are in Appendix E.3. + +Impact on Image Quality Additionally, we find that the quality of generated images remains approximately the same level as the original SD, as shown in Table 4. The observed differences in the reported scores and these in our experiments can be attributed to the randomness during image generation and caption sampling, which aligns with the inconsistencies reported in other studies [5]. + +Sensitivity to Hyperparameters In Appendix F, we investigate the sensitivity of our approach to hyperparameters, finding that it is less affected by factors such as the number of training images or different input prompts. Additionally, we demonstrate that our approach can leverage existing datasets to discover concept vectors. + +# 5. Conclusion + +In this study, we introduced a self-discovery approach to identify semantic concepts in the latent space of text-to-image diffusion models. Our research findings highlight that the generation of inappropriate content can be attributed to ethical-related concepts present in the internal semantic space of diffusion models. Leveraging these concept vectors, we enable responsible generation, including promoting equality among societal groups, eliminating inappropriate content, and enhancing text guidance for responsible prompts. Through extensive experiments, we have demonstrated the effectiveness and superiority of our proposed approach. Our work contributes to the understanding of internal representations in diffusion models and facilitates the generation of responsible content, maximizing the utility of high-quality text-to-image generation. + +Acknowledgement This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1, EPSRC/MURI grant: EP/N019474/1. We thank the Royal Academy of Engineering and FiveAI. This work is also founded by the German Federal Ministry of Education and Research and the Bavarian State Ministry for Science and the Arts. + +# References + +[1] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2 +[2] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2 +[3] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2 +[4] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/ logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 1 +[5] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8 +[6] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 5, 6 +[7] Jindong Gu, Ahmad Beirami, Xuezhi Wang, Alex Beutel, Philip Torr, and Yao Qin. Towards robust prompts on vision-language models. arXiv preprint arXiv:2304.08479, 2023. 1 +[8] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 2 +[9] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. arXiv preprint arXiv:2303.11073, 2023. 1, 2, 3, 4 +[10] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. arXiv preprint arXiv:2305.10120, 2023. 2 +[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 12 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 12 +[14] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training- + +free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 3 +[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 1 +[16] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2 +[17] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2 +[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4 +[19] Yipeng Leng, Qiangjuan Huang, Zhiyuan Wang, Yangyang Liu, and Haoyu Zhang. Diffusegae: Controllable and high-fidelity image manipulation from disentangled representation. arXiv preprint arXiv:2307.05899, 2023. 2 +[20] Hang Li, Jindong Gu, Rajat Koner, Sahand Sharifzadeh, and Volker Tresp. Do dall-e and flamingo understand each other? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1999-2010, 2023. 1 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8 +[22] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2 +[23] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In The Twelfth International Conference on Learning Representations, 2023. 1 +[24] Avery Ma, Amir-massoud Farahmand, Yangchen Pan, Philip Torr, and Jindong Gu. Improving adversarial transferability via model alignment. arXiv preprint arXiv:2311.18495, 2023. 1 +[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 8, 15 +[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2 + +[27] Minheng Ni, Chenfei Wu, Xiaodong Wang, Shengming Yin, Lijuan Wang, Zicheng Liu, and Nan Duan. Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785, 2023. 2 +[28] Zixuan Ni, Longhui Wei, Jiacheng Li, Siliang Tang, Yueting Zhuang, and Qi Tian. Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8900-8909, 2023. 2 +[29] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7053-7061, 2023. 2, 5 +[30] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4 +[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1 +[32] Vinay Uday Prabhu and Abeba Birhane. Large image datasets: A pyrrhic win for computer vision? arXiv preprint arXiv:2006.16923, 2020. 1 +[33] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[36] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 13 +[38] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350–1361, 2022. 1, 6 +[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2, 6, 7 +[40] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 2 +[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[43] Matthew Trager, Pramuditha Perera, Luca Zancato, Alessandro Achille, Parminder Bhatia, and Stefano Soatto. Linear spaces of meanings: Compositional structures in vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15395-15404, 2023. 2 +[44] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2 +[45] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. arXiv preprint arXiv:2306.08757, 2023. 2 +[46] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 2 +[47] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2 +[48] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2 +[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[50] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868, 2023. 2 +[51] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in Neural Information Processing Systems, 35:22117-22130, 2022. 2 + +[52] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. arXiv preprint arXiv:1804.06876, 2018. 5, 6 \ No newline at end of file diff --git a/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/images.zip b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e645f99086a27e6f0a04e415402809b559ff34e5 --- /dev/null +++ b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd863d291bca28ca8ca3366baf6ccb29c23dd200e99b61666aeb51e470b5fd36 +size 477627 diff --git a/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/layout.json b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..26308d4b8e3269aa7aebf155694adc2abdf2d1f3 --- /dev/null +++ b/2024/Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation/layout.json @@ -0,0 +1,7805 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "text", + "content": "Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "spans": [ + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "text", + "content": "Hang Li" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{1,4,5}" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "text", + "content": " Chengzhi Shen" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "text", + "content": " Philip Torr" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "text", + "content": " Volker Tresp" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "text", + "content": " Jindong Gu" + }, + { + "bbox": [ + 77, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "text", + "content": "LMU Munich, Germany " + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "text", + "content": "University of Oxford, UK " + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "text", + "content": "Technical University of Munich, Germany " + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "text", + "content": "Munich Center for Machine Learning, Germany " + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 55, + 175, + 539, + 204 + ], + "type": "text", + "content": "Siemens AG, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 255, + 290, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 290, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 290, + 495 + ], + "type": "text", + "content": "Diffusion-based models have gained significant popularity for text-to-image generation due to their exceptional image-generation capabilities. A risk with these models is the potential generation of inappropriate content, such as biased or harmful images. However, the underlying reasons for generating such undesired content from the perspective of the diffusion model's internal representation remain unclear. Previous work interprets vectors in an interpretable latent space of diffusion models as semantic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation. Extensive experiments have been conducted to verify the effectiveness of our mitigation approach, namely, for fair generation, safe generation, and responsible text-enhancing generation. Project page: https://interpretdiffusion.github.io." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 539, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 288, + 695 + ], + "type": "text", + "content": "The rapid advances in vision language models have sparked increasing interest in ensuring their safety and responsible use [7, 23, 24]. In particular, text-to-image diffusion models, which have exhibited remarkable performance in creating images from text prompts [13, 15, 20, 31, 35, 37, 49], raise concerns about the risks of generating inappropriate content. The generated images may exhibit biases and unsafe elements, including instances of gender discrimination or the depiction of violent scenes that could be harmful to children. Recent research efforts have focused on introducing safety mechanisms to mitigate these issues, such as filtering out inappropriate text input, detecting inappropriate images with a safety guard classifier [4, 32, 36, 38] and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 232, + 546, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 546, + 340 + ], + "type": "text", + "content": "building safe diffusion models [5, 6, 17]. However, the underlying mechanism of how diffusion models generate inappropriate content remains poorly understood. In this work, we aim to explore the following questions. 1) Are there any internal representations associated with these inappropriate concepts in the diffusion model-based generation process? 2) Can we manipulate representations to avoid inappropriate content corresponding to a given concept, i.e., to achieve responsible image generation?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "text", + "content": "To understand the image generation process of diffusion models, previous work has identified the bottleneck layer of the U-Net as a semantic representation space, dubbed " + }, + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "text", + "content": "-space [18]. They demonstrated that a vector in the " + }, + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "text", + "content": "-space can be associated with a specific semantic concept in the generated image. Manipulating the vector in the space can alter the generated image in a semantically meaningful way, such as adding a smile to a face. Several approaches [9, 18, 30] have been proposed to discover meaningful directions in this " + }, + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 340, + 547, + 484 + ], + "type": "text", + "content": "-space. For instance, an approach in [9] uses PCA to identify a set of latent directions that may represent semantic concepts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 484, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 548, + 689 + ], + "type": "text", + "content": "However, existing approaches to identifying interpretable latent vectors are limited. In unsupervised approaches [9, 30], it is not clear to which semantic concepts those identified vectors correspond. The found vectors must be interpreted with humans in a loop. Furthermore, the number of interpretable directions depends on the training data [9, 30]. It is highly likely that some target concepts may not be found in the discovered directions, especially those related to fairness and safety. Supervised approaches [9, 18] have also been explored to identify target concepts. These methods require training external attribute classifiers supervised by human annotations. Additionally, the quality of the identified vectors is sensitive to the classifier's performance. Furthermore, new concepts require the training of new classifiers. Overall, existing interpretation methods cannot be easily applied to identify the corresponding semantic vector for a given inappropriate concept." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "In this work, we propose a self-discovery approach to find interpretable latent directions in the " + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "-space for user" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12006" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 347 + ], + "type": "text", + "content": "defined concepts. We learn a latent vector that effectively represents the concept by leveraging the model's acquired semantic knowledge in its internal representations. Initially, images are generated using specific text prompts related to the concept. The images are then used in a de-noising process where the frozen pretrained diffusion model reconstructs these images from noise, guided by a modified text prompt that omits the desired concept, and our introduced latent vector. By minimizing the reconstruction loss, the vector learns to represent the given concept. Our self-discovery approach eliminates the need for external models like CLIP text encoder [34] or dedicated attribute classifiers trained on human-labeled datasets. We identify ethical-related latent vectors and demonstrate their applications in responsible text-to-image generation: 1) fairness by sampling an ethical concept, e.g., gender, in the latent space, which generates images with unbiased attributes and aligned with the prompt. 2) safety generation by incorporating safety-related concepts, e.g., one that eliminates the nudity content, into " + }, + { + "bbox": [ + 47, + 72, + 289, + 347 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 72, + 289, + 347 + ], + "type": "text", + "content": "-space to prevent the model from generating such harmful content. 3) responsible guidance, where we first discover responsible concepts in the text prompt and enhance the expression of those ethical concepts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 350, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 289, + 506 + ], + "type": "text", + "content": "Previous approaches enhance responsible image generation from different perspectives. Concretely, [2, 5, 6, 8, 17] fine-tune the diffusion models or text embeddings to unlearn harmful concepts, and [39] applies classifier-free guidance to steer the generation away from unsafe concepts. Despite the mitigation mechanism of previous approaches, diffusion models still suffer from inappropriate content generation [39, 50]. Unlike previous work, in this work, we provide a new perspective to mitigate the inappropriate generation, namely, finding and manipulating concepts in an interpretable latent space. Our work can be easily combined with previous mitigation approaches to further enhance responsible text-to-image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 506, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 289, + 615 + ], + "type": "text", + "content": "We conducted extensive experiments on fairness, safety, and responsible guidance-enhancing generation. Our model consistently produces images with a balanced representation across societal groups. Further, we successfully mitigate harmful content for inappropriate prompts. In addition, our approach synergistically improves the performance of responsible image generation when combined with existing methods. Furthermore, we enhance text guidance to generate fair and safe content for responsible prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 616, + 259, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 616, + 259, + 628 + ], + "spans": [ + { + "bbox": [ + 59, + 616, + 259, + 628 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 630, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "text", + "content": "- We propose a self-discovery method for identifying interpretable directions in the diffusion latent space. Our approach can find a vector that represents any desired concept, without the need for labeled data or external models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 678, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 678, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 678, + 287, + 714 + ], + "type": "text", + "content": "- With the discovered vectors, we propose a straightforward yet effective approach to enhance responsible generation, including fair generation, safe generation, and re" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 73, + 545, + 109 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 73, + 464, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 464, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 464, + 84 + ], + "type": "text", + "content": "sponsored text-enhancing generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 84, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 545, + 109 + ], + "type": "text", + "content": "- Extensive experiments are conducted to validate the effectiveness of our approach." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 124, + 392, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 124, + 392, + 136 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 392, + 136 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 145, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 547, + 494 + ], + "type": "text", + "content": "Responsible Alignment of Diffusion Models Various approaches have been proposed to mitigate the generation of biased and unsafe content in diffusion models. A straightforward method involves refining the training dataset to remove biased and inappropriate content, exemplified by Stable Diffusion (SD) v2 [37]. Such approaches can be computationally intensive, may not fully eliminate harmful content [5], and could degrade the model's performance [39]. An alternative is to detect and filter out inappropriate words from the input prompts [1, 2, 27]. However, this fails to address non-explicit phrases that can still yield inappropriate outputs. Another line of approaches involves finetuning the parameters of pretrained models, aiming to remove the model's representation capability of generating such inappropriate concepts [5, 17]. However, they are sensitive to the adaptation process and may result in the degradation of the original models [6, 10, 28, 29, 48]. Moreover, such approaches require a potentially exhaustive list of words that introduce biases and harmful concepts [5, 6, 29]. Training-free approaches utilize classifier-free guidance to direct the generated images away from undesirable content during inference [1, 3, 39, 39, 47]. While they modify the noise space using text-based guidance through cross-attention mechanisms, we adopt a similar conditioning strategy to manipulate the generation for frozen pretrained models in the semantic latent space. As an orthogonal approach to the existing literature, we mitigate the inappropriate content by finding the corresponding latent directions in the U-Net bottleneck layer and suppressing their activations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 496, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 652 + ], + "type": "text", + "content": "Interpreting Diffusion Models To understand the working mechanisms of diffusion models, recent works mainly focus on investigating text guidance for conditional diffusion models [11, 16, 22, 26, 29, 43, 46], or analyzing the internal representations in diffusion models' intermediate layer activations [9, 18, 30, 40, 44]. We focus on elucidating the internal representations learned within the diffusion model, in line with prior works [18]. Some work [19, 33, 45, 51] proposes to create a semantic space in diffusion models by employing an autoencoder to encode the image into a semantic vector that guides the decoding process. However, their approaches require adapting the parameters of the autoencoder or even the entire framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "The seminal work [18] reveals that the bottleneck layer of U-Net architecture already exhibits properties suitable for a semantic representation space. They identified disentangled representations associated with the semantics of the generated image and demonstrated that those latent di" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12007" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 236 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 236 + ], + "type": "image", + "image_path": "bb13d88ae470191c1fc694af0e667ae33a32f125a035adac5eec27731b720cf6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 246, + 289, + 379 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 289, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 289, + 379 + ], + "type": "text", + "content": "Figure 1. Optimization framework to discover a semantic vector for a given concept. The top line shows that an image is firstly generated by the pretrained Stable Diffusion model for the prompt \"a female face\". The bottom part shows the optimization process for finding the concept for \"female\" in the semantic " + }, + { + "bbox": [ + 46, + 246, + 289, + 379 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 246, + 289, + 379 + ], + "type": "text", + "content": "-space. The concept vector is used to reconstruct the image along with a modified prompt \"a face\", under an iterative denoising process. With the pretrained diffusion model frozen, the gradients of the reconstruction loss can solely update the latent vector to represent the missing gender information. After convergence, the latent vector is aligned with the U-Net's internal representation of the \"female\" concept, which can be used to guide new image generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 401, + 289, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 289, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 289, + 557 + ], + "type": "text", + "content": "rections are identical to different images. However, their approach relies on the CLIP classifier and paired source-target images and edits, making it inefficient. Another work proposes a PCA-based decomposition method on the latent space and finds interpretable attribute directions using the top right-hand singular vectors of the Jacobian. Additionally, [30] uses Riemannian metrics to define more accurate and meaningful directions. However, these approaches require manual interpretation to identify the editing effect of each component. Our approach differs from the supervised approach in [9] by enabling the efficient discovery of latent directions for any given target concept without requiring a data collection process or training external classifiers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 569, + 115, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 115, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 115, + 582 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "content": "This section first introduces our optimization method to find interpretable directions in diffusion models' " + }, + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "content": "-space. In the second part, we show how to utilize discovered concepts in the inference process for responsible generation, including fairness, safety, and text-enhancing generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 204, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 204, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 204, + 672 + ], + "type": "text", + "content": "3.1. Finding a Semantic Concept" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Diffusion models are generative models that generate samples from Gaussian noise through a denoising process [13, 41, 42]. Starting from a random vector " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " of the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "same dimension as the image, the model estimates a noise value at each time step to subtract from the current vector to obtain a denoised image, denoted as " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "x_{t - 1} = x_t - \\epsilon_\\theta (x_t,t)" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " represents the U-Net of the diffusion model. A clean image " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " is obtained at the end of this denoising process. The training of diffusion models involves a forward process that iteratively adds noise to images from the data, denoted as " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "x_{t} = x_{t - 1} + \\epsilon_{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ". The training loss includes predicting noises for different steps," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 357, + 191, + 545, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 191, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 357, + 191, + 545, + 218 + ], + "type": "interline_equation", + "content": "L = \\sum_ {x \\sim \\mathcal {D}} \\sum_ {t \\sim [ 0, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2}. \\tag {1}", + "image_path": "d7bc475bbe5f3dfec50bcb5587548919c289db080a4f5b7ea6cfc90ea7d2a5a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": "Recent work identified a semantic space in the diffusion model, the activations of U-Net's bottleneck layer " + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": ", as shown in Figure 1. The activations in " + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": "-space leads to the generation of a less noised image for the next timestep " + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": ". This space exhibits semantic structures and is easy to interpret. Activating a specific vector in the U-Net bottleneck layer leads to the image having a certain attribute. However, existing approaches cannot find the vector for an arbitrarily given concept. Our goal is to find such vectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": "To this end, we utilize the text-to-image conditional diffusion model which can generate images from a given text input. The prediction function in Eq. 1 becomes " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,\\pi (y),t)" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "\\pi (y)" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " is the encodings of the input text " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": ". The equation specifies a conditional distribution that drives the generation of the image towards data regions that are highly likely given the input text [12]. To discover an interpretable direction, we leverage the pre-trained model to generate a set of images using dedicated prompts related to that concept. For example, to find the latent direction of the concept \"female\", we first generate a set of images " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "x^{+}" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " with a descriptive prompt " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "y^{+}" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " \"a photo of a female face\". Then, a concept vector is optimized for the conditional generation where the original prompt has been modified into " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "y^{-}" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " \"a photo of a face\", eliminating gender information. The concept vector " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " is randomly initialized in the latent space, where " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " is the dimension of " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": "-space, and is optimized to minimize the reconstruction error. Since the pre-trained diffusion model is frozen, the model has to utilize the extra condition " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " to compensate for the missing information not in the text condition but in the image. The concept vector " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " will be forced to represent the missing information from the input text to produce an image with the lowest reconstruction error. After convergence, that vector " + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 339, + 546, + 663 + ], + "type": "text", + "content": " is expected to represent the gender information \"female\". In this way, we discover a set of vectors that represent target concepts, such as gender, safety, and facial expressions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 673, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 673, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 547, + 713 + ], + "type": "text", + "content": "1The decoding of " + }, + { + "bbox": [ + 304, + 673, + 547, + 713 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 304, + 673, + 547, + 713 + ], + "type": "text", + "content": " depends on other variables due to the presence of skip connections. For simplicity, we omit this consideration as the skip connection seems less significant in encoding compact semantic information, as supported in previous findings [9, 14]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12008" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 275, + 167 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 275, + 167 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 275, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 275, + 167 + ], + "type": "image", + "image_path": "838c5aac9ca7fa1df74e1b4ad20b0cfb88c3b151585475b7d9ac40a01a6349fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 288, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 288, + 233 + ], + "type": "text", + "content": "Figure 2. Fair Generation. Top: images generated from the prompt \"doctor\" are biased toward males. Bottom: we sample a learned male or female concept with equal probability for generating the doctors. The doctors now have fair gender. Images are generated from different random seeds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 255, + 274, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 274, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 274, + 268 + ], + "type": "text", + "content": "Formally, the optimal " + }, + { + "bbox": [ + 47, + 255, + 274, + 268 + ], + "type": "inline_equation", + "content": "c^*" + }, + { + "bbox": [ + 47, + 255, + 274, + 268 + ], + "type": "text", + "content": " for a given concept is found by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 277, + 287, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 277, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 51, + 277, + 287, + 304 + ], + "type": "interline_equation", + "content": "c ^ {*} = \\arg \\min _ {c} \\sum_ {x, y \\sim \\mathcal {D}} \\sum_ {t \\sim [ 0, T ]} \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t} ^ {+}, t, \\pi \\left(y ^ {-}\\right), c\\right) \\| ^ {2}, \\tag {2}", + "image_path": "0bd907ebab72cee7b863198ec85b6fdf2c91204f52c2d25299fc7a2bab3b6a23.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "x_{t}^{+}" + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": " denotes the noised version of the original image generated with " + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "y^{+}" + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": " represents the target concept. " + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": " denotes the U-Net that linearly adds an additional concept vector " + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": " to its " + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": "-space, at each decoding timestep. Regarding implementation, the " + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": "-space is the flattened activations after the middle bottleneck layer of the U-Net. The pseudo-code for this training pipeline is in Appendix A.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 590 + ], + "type": "text", + "content": "We learn a single vector for each concept for all timesteps, as the latent direction remains approximately consistent across different timesteps [18]. Moreover, we restrict the operation to linearity to demonstrate the power of this latent space. Notably, the learned vector generalizes effectively to new images [9] and diverse prompts [30]. For instance, a \"male\" concept learned with the base prompt \"person\" can be used in different contexts, such as \"doctor\" or \"manager\", as shown in the next section. Additionally, the concepts can be optimized jointly or independently, with the experimental section demonstrating the impact of concept composition. A key strength of our approach is utilizing synthesis by diffusion models to collect data, eliminating the need for human labeling and training of guiding classifiers. Nevertheless, our approach can be applied to realistic datasets with annotated attributes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "text", + "content": "3.2. Responsible Generation with Self-discovered Interpretable Latent Direction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "type": "text", + "content": "In this subsection, we utilize the identified directions to manipulate the latent activation in the latent space for fair, safe, and enhanced responsible generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Fair Generation Method A text prompt contains words that lead to the generation of biased societal groups. We aim to generate images with evenly distributed attributes for a given text prompt. For example, for the prompt \"doctor\"," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 69, + 533, + 170 + ], + "blocks": [ + { + "bbox": [ + 319, + 69, + 533, + 170 + ], + "lines": [ + { + "bbox": [ + 319, + 69, + 533, + 170 + ], + "spans": [ + { + "bbox": [ + 319, + 69, + 533, + 170 + ], + "type": "image", + "image_path": "f362d56a6aae7352e4d2e42b09b3e1058cafd58ca0ce352f396f4b568daf7457.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 181, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 270 + ], + "type": "text", + "content": "Figure 3. Safe Generation. When the user's prompt contains implicit references to nudity, the original model (shown in the top row) generates an inappropriate image, as the added blurriness indicates. In contrast, our approach generates an image for the same prompt by setting a safety-related concept in " + }, + { + "bbox": [ + 304, + 181, + 545, + 270 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 181, + 545, + 270 + ], + "type": "text", + "content": "-space, identified in the previous section. The vector anti-sexual concept represents the direction to suppress nudity content, effectively eliminating inappropriate content while maintaining fidelity to the prompt." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "text", + "content": "we aim to generate an image of a male doctor with a " + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "text", + "content": " probability and a female doctor with a " + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "text", + "content": " probability. For that, we learn a set of semantic concepts representing different societal groups using the approach in the previous section. For inference, a concept vector is sampled from the learned concepts in the societal group with equal probability, e.g., the C-male and C-female concept vectors for gender are chosen with fair chance. The inference process is fixed as before, except that the sampled vector is added to the original activations in " + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 289, + 545, + 421 + ], + "type": "text", + "content": "-space at each decoding step, denoted by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 360, + 430, + 545, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 430, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 360, + 430, + 545, + 442 + ], + "type": "interline_equation", + "content": "h \\leftarrow h + c \\sim \\operatorname {C a t e g o r i c a l} \\left(p _ {k}\\right), \\tag {3}", + "image_path": "a95f13fc5f0f92605935a16a58cb6a2d5b8aad26bd4b3b3fb0e3cc0e1c790fa6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "content": " represents the probability of sampling a particular attribute " + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "content": " from the societal group with " + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "content": " distinct attributes. For the fair generation, " + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "inline_equation", + "content": "p_k = 1 / C" + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "content": ". Guided by this sampled concept vector, the generated image is expected to be a male doctor if the C-male concept is sampled or a female doctor otherwise. This allows the generated images to have an equal number of attributes, e.g., an equal number of male and female doctors, shown in Figure 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "content": "Safe Generation Method For safety generation, we consider text prompts that contain explicit or implicit references to inappropriate content, which we aim to eliminate. An example of such a prompt is illustrated in Figure 3, where the phrase \"a gorgeous woman\" may indirectly lead to the generation of nudity. We identify a collection of safety-related concepts, such as anti-sexual, to achieve safe generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Specifically, we learn the opposite latent direction of an inappropriate concept, leveraging the negative prompt technique. For instance, the training images are generated by the prompt " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "y^{+}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " \"a gorgeous person\" with a negative prompt \"sexual\", which effectively instructs the Stable Diffusion to generate safe images without sexual content. The concept vector is then optimized on those training images that depict" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12009" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "safe content. For that, the input prompt " + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "inline_equation", + "content": "y^{-}" + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": " is set to \"a gorgeous person\" but without the negative prompt \"sexual\". In this way, the concept vector directly learns the concept of \"anti-sexual\". The reason for adopting this strategy is the difficulty of listing all the opposite concepts of sexuality, e.g., \"dressed\", \"clothes\", or more. An alternative approach is to learn the concept of sexuality directly and apply negation during generation, which we found less effective. More details regarding the negative prompt are in Appendix A.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 216 + ], + "type": "text", + "content": "After the learning process, we maintain all aspects of inference unchanged, except for adding the learned vector to the original activations at the bottleneck layer, formally as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 222, + 287, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 222, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 140, + 222, + 287, + 235 + ], + "type": "interline_equation", + "content": "h \\leftarrow h + c _ {s}. \\tag {4}", + "image_path": "286b3946a31138597b9e38f09bef7b2df6b0007694644052517c8e976e7d7ff5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 430 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 240, + 287, + 430 + ], + "type": "inline_equation", + "content": "c_{s}" + }, + { + "bbox": [ + 46, + 240, + 287, + 430 + ], + "type": "text", + "content": " refers to a safety concept, such as \"anti-sexual\", which represents the opposite of sexual content. This strengthens the expression of safe concepts in the generated images so they are devoid of harmful content. Figure 3 illustrates the impact of including the anti-sexual vector, resulting in a visually appealing person with appropriate clothes. Responsible Text-enhancing Generation Method Even when a prompt is intentionally designed to promote safety, the generative models may struggle to accurately incorporate all the concepts defined in the prompt. For instance, consider a text prompt like \"an exciting Halloween party, no violence\". The generative model may encounter difficulties in faithfully representing each responsible concept from the prompt, e.g., its poor understanding of negation on \"violence\" may result in the generation of inappropriate content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "text", + "content": "To address this issue, we utilize our self-discovery approach to learn concepts such as gender, race, and safety. To enhance the generation of responsible prompts, we extract safety-related content from the text and leverage our learned ethical-related concepts to reinforce the expression of desired visual features. During inference, we apply the extracted concepts " + }, + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "inline_equation", + "content": "c(y)" + }, + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "text", + "content": " from the prompt to the original activations, denoted as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 534, + 287, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 534, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 136, + 534, + 287, + 547 + ], + "type": "interline_equation", + "content": "h \\leftarrow h + c (y) \\tag {5}", + "image_path": "ce927e399757ac06f36e86b6968020ee62035185801e964aa123ed3b9a0c04a5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 649 + ], + "type": "text", + "content": "For example, as shown in Figure 4, the concept of \"no violence\" from the text prompt activates our learned \"anti-violence\" concept during inference. By directly manipulating the semantic space, our approach introduces the desired attributes to the generated image. Compared to the original generated image, the anti-violence concept effectively mitigates the presence of violent content and makes the generated images more appropriate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 128, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 128, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 128, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "In this section, we present the performance of our methods on fair generation, safe generation, and enhanced responsible generation. Additionally, we demonstrate the general" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 69, + 533, + 162 + ], + "blocks": [ + { + "bbox": [ + 319, + 69, + 533, + 162 + ], + "lines": [ + { + "bbox": [ + 319, + 69, + 533, + 162 + ], + "spans": [ + { + "bbox": [ + 319, + 69, + 533, + 162 + ], + "type": "image", + "image_path": "e5b7fe4746436af20cd134c77d7aada804c178c8a0105c8031b89fa9d7efa521.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 171, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 250 + ], + "type": "text", + "content": "Figure 4. Responsible text-enhancing generation. The original model may fail to capture the safety concepts specified in the text, such as \"no violence\". We propose extracting those safety concepts from the given prompt and activating the safety directions during generation. The bottom image demonstrates that incorporating our safety concepts can enhance the text guidance of the original prompt." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 272, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 272, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 545, + 296 + ], + "type": "text", + "content": "properties of our discovered concept vectors, highlighting their generalization capability and compositionality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 306, + 405, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 405, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 405, + 319 + ], + "type": "text", + "content": "4.1. Fair Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "content": "Task The goal is to increase the diversity of societal groups in the generated images, particularly in professions where existing models exhibit gender and racial bias." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 506 + ], + "type": "text", + "content": "Dataset Following [6, 29], our approach is evaluated on the Winobias [52] benchmark that comprises 36 professions known to exhibit gender biases. We generate five different prompts for each profession using various templates, such as \"a portrait of a doctor\". Additionally, we create a hard dataset by augmenting the existing dataset with prompts that are more likely to induce stereotypes. This extended dataset adds the term \"successful\" to each original prompt, such as \"a portrait of a successful doctor\". The word \"successful\" often leads to the generation of male images and thus constitutes challenges for a fair generation [6]. Appendix B.1 contains the complete list of prompts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "text", + "content": "Evaluation Metric The CLIP classifier is employed to predict attributes by measuring the similarity between the text embedding of a concept (e.g., female, male) and the embedding of the generated image. We utilize the deviation ratio [6, 29] to quantify the imbalance of different attributes. To accommodate an arbitrary number of attributes, the metric is modified as " + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\Delta = \\max_{c\\in C}\\frac{|N_c / N - 1 / C|}{1 - 1 / C}" + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "text", + "content": " is the total number of attributes within a societal group, " + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "text", + "content": " is the total number of generated images, and " + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "text", + "content": " denotes the number of images whose maximum predicted attribute equals " + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 506, + 547, + 689 + ], + "type": "text", + "content": ". In particular, we test the gender, male, female, and racial, black, white, Asian, biases associated with the professions. These races are selected as the CLIP classifier has relatively reliable predictions on these attributes. During the evaluation, 150 images were generated for each profession." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Approach Setting In all experiments, we use the Stable Diffusion v1.4 checkpoint and set the guidance scale to 7.5" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12010" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 292, + 169 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 292, + 169 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 292, + 169 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 292, + 169 + ], + "type": "image", + "image_path": "4434228768e8c00e84057e02eeeb376e4346343c7d3c35a42a7d610801504a9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 171, + 216, + 181 + ], + "lines": [ + { + "bbox": [ + 125, + 171, + 216, + 181 + ], + "spans": [ + { + "bbox": [ + 125, + 171, + 216, + 181 + ], + "type": "text", + "content": "(a) Images generated by " + }, + { + "bbox": [ + 125, + 171, + 216, + 181 + ], + "type": "inline_equation", + "content": "\\mathbf{SD}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 302, + 72, + 544, + 169 + ], + "blocks": [ + { + "bbox": [ + 302, + 72, + 544, + 169 + ], + "lines": [ + { + "bbox": [ + 302, + 72, + 544, + 169 + ], + "spans": [ + { + "bbox": [ + 302, + 72, + 544, + 169 + ], + "type": "image", + "image_path": "89c656c08923f91da03787a4b3eac60e505e42f572dbe9fb284e3d9c88324882.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 171, + 472, + 181 + ], + "lines": [ + { + "bbox": [ + 374, + 171, + 472, + 181 + ], + "spans": [ + { + "bbox": [ + 374, + 171, + 472, + 181 + ], + "type": "text", + "content": "(b) Images generated by Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 190, + 546, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 546, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 546, + 214 + ], + "type": "text", + "content": "Figure 5. Gender fairness generation. From the prompt \"a photo of a doctor\", the original SD exhibits significant gender bias, as shown on the left side. Our approach with uniformly sampled gender vectors represents genders equally in the generated images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 224, + 535, + 327 + ], + "blocks": [ + { + "bbox": [ + 58, + 224, + 535, + 327 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 535, + 327 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 535, + 327 + ], + "type": "table", + "html": "
Dataset\nMethodGenderGender+\nUCERaceRace+\nUCE
SDUCEOursSDUCEOursSDUCEOursSDUCEOurs
Analyst0.700.200.020.540.040.020.820.290.240.770.200.41
CEO0.920.280.060.900.580.060.380.130.220.310.080.22
Laborer1.000.090.120.980.080.140.330.400.240.530.380.20
Secretary0.640.100.360.920.960.460.370.350.240.550.430.32
Teacher0.300.060.040.480.160.100.510.100.040.260.230.21
Winobias[52]0.680.220.170.700.520.230.560.210.230.480.350.20
", + "image_path": "75f931260144ccacae93b4fbff2fae0b6bab824786af5db068bab2b2c88b1e81.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 546, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 546, + 381 + ], + "type": "text", + "content": "Table 1. Fair generation quantified by the deviation ratio " + }, + { + "bbox": [ + 46, + 335, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\left( {0 \\leq \\Delta \\leq 1}\\right)" + }, + { + "bbox": [ + 46, + 335, + 546, + 381 + ], + "type": "text", + "content": " . Lower values indicate better performance. The left side of the table presents the results on gender attributes, whereas the right side quantifies the racial bias. \"Gender+/Race+\" refers to the extended Winobias dataset, which is more challenging, as described in Subsection 4.1. Our approach leads to unbiased generation for biased prompts and is robust to diverse sources of bias in the prompt." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": "for text-to-image generation. We find five concept vectors using a base prompt \"person\", e.g., " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "y^{+} =" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " \"a photo of a woman\" and " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "y^{-} =" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " \"a photo of a person\" to learn the concept \"female\". The concept vectors are optimized for 10K steps on 1K synthesized images for each concept. During inference, we directly employ the learned vector without any scaling. Unlike the baseline approach UCE [6], which needs to debias each profession in Winobias, Our approach is trained solely on the \"person\" prompt to learn the male and female concept that generalizes to all different professions. For comparison, we report UCE's published scores when available and otherwise use their released code to train the model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Results and Analysis Table 1 reveals that our approach is significantly better than the original SD and outperforms the state-of-the-art debiasing approach UCE. The professions on the table are randomly selected from the complete list of 36 professions (see Appendix B.2). Figure 5 compares images generated from our approach and those from the original SD. Further, we highlight the generalization capability of our approach to different text prompts using the extended Winobias dataset, as shown in the second and fourth column blocks of Table 1. Despite the presence of bias in the text prompts, our approach consistently performs well as it directly operates on the latent visual space. In contrast, UCE performs poorly on this challenging dataset, as it relies on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 401, + 546, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 462 + ], + "type": "text", + "content": "debiasing each word in the prompt. The effectiveness of UCE is easily weakened by biased words that are not included in its training set. Additionally, we demonstrate that the quality of images generated by our approach remains consistent with the original SD and UCE in Appendix B.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 470, + 404, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 404, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 404, + 481 + ], + "type": "text", + "content": "4.2. Safe Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 488, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 548 + ], + "type": "text", + "content": "Task This section focuses on generating images that eliminate harmful content specified in inappropriate prompts. As an orthogonal approach to existing methods, our approach is combined with current safety methods, including SLD [39] and ESD [5], to eliminate inappropriate generation further." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 549, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 657 + ], + "type": "text", + "content": "Dataset and Evaluation Metric The I2P benchmark [39] is a collection of 4703 inappropriate prompts from real-world user prompts. The inappropriateness covers seven categories, including, e.g., illegal activity, sexual, and violence. For evaluation, the Nudenet detector and Q16 classifier are used to detect nudity or violent content in an image. An image is classified as inappropriate if any of the classifiers predicts a positive [5]. Five images are generated for each prompt for evaluation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 658, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 545, + 693 + ], + "type": "text", + "content": "Approach Setting We find that optimizing a single concept vector for \"safety\" is challenging. Therefore, we learn the concept vector for each inappropriate concept defined in the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 500, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 500, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 500, + 712 + ], + "type": "text", + "content": "2https://github.com/notAI-tech/NudeNet" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12011" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 548, + 161 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 548, + 161 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 548, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 548, + 161 + ], + "type": "table", + "html": "
CategoryHarassmentHateIllegalSelf-harmSexualShockingViolenceI2P[39]
SD0.34 ±0.0190.41 ±0.0320.34 ±0.0180.44 ±0.0190.38 ±0.0160.51 ±0.0170.44 ±0.0180.41 ±0.007
Ours-SD0.18 ±0.0150.29 ±0.0300.23 ±0.0160.28 ±0.0170.22 ±0.0140.36 ±0.0170.30 ±0.0170.27 ±0.006
SLD[39]0.15 ±0.0140.18 ±0.0250.17 ±0.0150.19 ±0.0150.15 ±0.0120.32 ±0.0160.21 ±0.0150.20 ±0.006
Ours-SLD0.14 ±0.0140.20 ±0.0270.14 ±0.0130.14 ±0.0130.09 ±0.0100.25 ±0.0150.16 ±0.0130.16 ±0.005
ESD[5]0.27 ±0.0180.32 ±0.0310.33 ±0.0180.35 ±0.0180.18 ±0.0130.41 ±0.0170.41 ±0.0180.32 ±0.007
Ours-ESD0.26 ±0.0170.29 ±0.0300.25 ±0.0170.26 ±0.0170.13 ±0.0110.41 ±0.0170.30 ±0.0170.27 ±0.006
", + "image_path": "c5cb060bbff5193c99c85aa90f583567e86f469ccae81f7abfabecef632263df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 224, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 368 + ], + "type": "text", + "content": "I2P dataset, e.g., \"anti-sexual\". The identified concept vectors are linearly combined as the final vector to modulate the generation process. The effect of linearly combining concept vectors is further discussed in the next section. Additionally, certain concepts are rather abstract and include diverse visual categories, e.g., \"hate\". Adding these concepts improves safety yet at a higher cost of image quality degradation. Therefore, we utilize two prevailing concepts, anti-sexual and anti-violence, which may underline other concepts, as our final model. For baselines, we use SLD [39] and ESD [5] with their released model weights. Detailed settings and hyperparameters are in Appendix C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 288, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 500 + ], + "type": "text", + "content": "Results and Analysis Table 2 demonstrates the effectiveness of our approach in eliminating inappropriate content. Notably, our safety vector can suppress inappropriate concepts that existing approaches failed to eliminate, as evidenced by the improved performance compared to existing approaches. In particular, we observe significant improvement in eliminating nudity with the combined approach of SLD and our proposed method, with a relative improvement of " + }, + { + "bbox": [ + 46, + 369, + 288, + 500 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 369, + 288, + 500 + ], + "type": "text", + "content": ". Additionally, even after retraining the model weights to eliminate nudity, ESD still generates inappropriate images, which our approach can mitigate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 511, + 251, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 251, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 251, + 525 + ], + "type": "text", + "content": "4.3. Enhancing Responsible Text Guidance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "text", + "content": "Task For user prompts classified as responsible text, we aim to accurately represent the responsible phrases in the prompt in the generated image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 676 + ], + "type": "text", + "content": "Dataset We created a dataset of 200 prompts that explicitly include responsible concepts. In particular, the dataset focuses on gender and race fairness, as well as the removal of sexual and violent content. For instance, the prompt \"a fair-gender doctor is operating a surgery\" assesses the model's ability to generate images with balanced genders based on the text cue \"fair-gender\". Additionally, prompts such as \"a picture of a loved couple, without sexual content\" address safety concerns. More dataset details are in Appendix D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Results and Analysis We utilized previously introduced metrics for evaluation, i.e., the deviation ratio for fairness, NudeNet for sexual content, and Q16 for violent content." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 311, + 221, + 539, + 262 + ], + "blocks": [ + { + "bbox": [ + 46, + 169, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 547, + 204 + ], + "type": "text", + "content": "Table 2. The proportion of images classified as inappropriate on the I2P benchmark. In each block of results, the first row shows the performance of the original method, while the second row represents adding our concept vector to the corresponding baseline model. Our identified safety-related vector can be combined with existing safety approaches to mitigate inappropriate content generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 221, + 539, + 262 + ], + "lines": [ + { + "bbox": [ + 311, + 221, + 539, + 262 + ], + "spans": [ + { + "bbox": [ + 311, + 221, + 539, + 262 + ], + "type": "table", + "html": "
ModelGenderRaceSexualViolence
SD0.21990.16000.43000.4551
Ours-SD0.14330.13990.26400.3204
", + "image_path": "b3e18b7dda0cce3bb1e70cec1b29f536066e62e9298a0ff6f73e47a19d31972c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 266, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 547, + 300 + ], + "type": "text", + "content": "Table 3. For prompts containing responsible concepts, the original SD may fail to follow the prompts faithfully. Our approach effectively enhances responsible text-guidance generation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 417 + ], + "type": "text", + "content": "For evaluation, 3500 images are generated for the dataset. For our approach, we provide the model with the corresponding concept associated with the input prompt. For example, if a prompt mentions \"no sexual\", the anti-sexual vector is added to the generation process. Table 3 compares our approach with the original model, which does not use the safety concepts. Our approach effectively enhances the text guidance for responsible instructions." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 428, + 545, + 548 + ], + "blocks": [ + { + "bbox": [ + 306, + 428, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 306, + 428, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 545, + 548 + ], + "type": "image", + "image_path": "99de2f3204c9905559f8260432c09c4dc6c811f61596c94c03e0b96ae884caae.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 555, + 547, + 600 + ], + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 600 + ], + "type": "text", + "content": "Figure 6. Concept interpolation. The first column displays the original image generated by SD. The following columns show images generated by the same random seeds, but with concept vector scales linearly increasing from 0.2 to 0.8." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 610, + 419, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 419, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 419, + 624 + ], + "type": "text", + "content": "4.4. Semantic Concepts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": "In previous experiments, we have demonstrated the specific applications of our identified concept vectors for responsible generation. This subsection introduces the general properties of discovered vectors related to the semantic space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Interpolation Figure 6 illustrates the impact of manipulating image semantics by linearly controlling the strength of the concept vector, denoted as " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " in the equation " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "h\\gets h + \\lambda c" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12012" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 185 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 185 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 185 + ], + "type": "image", + "image_path": "dee4388c5d867e1c32ea30c301707d7bc96abb9a306fc1e5e32c8539a0d38d22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 288, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 228 + ], + "type": "text", + "content": "Figure 7. Multiple concepts composition. The concept vectors of gender, age, and race were learned independently. Linearly adding latent vectors can generate images with corresponding semantics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 246, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 331 + ], + "type": "text", + "content": "The image is gradually modified to the introduced concept by adjusting the added vector's strength. The smooth transition indicates that the discovered vector represents the target semantic concept while remaining approximately disentangled from other semantic factors. Appendix E.1 presents more examples of concept manipulation and enhanced fidelity by post-hoc interpolation methods [25]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 332, + 289, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 289, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 289, + 406 + ], + "type": "text", + "content": "Composition Figure 7 showcases the composability of learned concept vectors, which were trained independently. Images are generated from the prompt \"a photo of a doctor\". By linearly combining these concept vectors, we can control the corresponding attributes in the generated image. Appendix E.2 provides a quantitative evaluation." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 66, + 416, + 267, + 528 + ], + "blocks": [ + { + "bbox": [ + 66, + 416, + 267, + 528 + ], + "lines": [ + { + "bbox": [ + 66, + 416, + 267, + 528 + ], + "spans": [ + { + "bbox": [ + 66, + 416, + 267, + 528 + ], + "type": "image", + "image_path": "0ddbae4d8953f1c74dc4dfa3dcaaafec76a56f77dd8b17a0caecdefd33c9b087.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 534, + 288, + 581 + ], + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 581 + ], + "type": "text", + "content": "Figure 8. General semantic concepts identified by our approach. Top: The concept of \"running\" is learned from dog images and can be generalized to different objects. Bottom: The concept vector of \"glasses\" enhances the prompt \"an elephant wearing glasses\"." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Generalization Figure 8 illustrates the generalization capability of our discovered concept vector to universal semantic concepts. We train the latent vector for the concept \"running\" on generated dog images and test its effect on other objects using prompts such as \"a photo of a cat\". Each pair of images in the figure is generated from the same random seed. Although the vector of \"running\" was learned from dogs, it successfully extends to different animals and even humans. Additionally, our approach enhances the original text guidance for the prompt \"an ele" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 70, + 545, + 112 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 545, + 112 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 545, + 112 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 545, + 112 + ], + "type": "table", + "html": "
ModelSDSLD*SLDESD*ESDOurs-SD
FID14.0916.9019.3513.6815.3615.98
CLIP31.33-30.41-30.0531.03
", + "image_path": "5bacde55c5f5b2db2c4d2d0c013fc0879393cf46b89fdc1db8058d7e7aa938fc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 119, + 547, + 186 + ], + "lines": [ + { + "bbox": [ + 304, + 119, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 119, + 547, + 186 + ], + "type": "text", + "content": "Table 4. Evaluation of the quality of generated images on the COCO-30K [21] dataset using FID for image fidelity and CLIP Score for semantic alignment with input text. Various safety approaches have approximately the same level of image quality as the original SD. Numbers reported from the corresponding papers are denoted with * ." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 207, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 267 + ], + "type": "text", + "content": "phant wearing glasses\". The original SD cannot produce accurate images, as shown in the first and third images on the bottom. The correct images can be generated by adding the concept vector \"glasses\" in " + }, + { + "bbox": [ + 304, + 207, + 545, + 267 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 207, + 545, + 267 + ], + "type": "text", + "content": "-space. More visualizations are in Appendix E.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 270, + 546, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 546, + 354 + ], + "type": "text", + "content": "Impact on Image Quality Additionally, we find that the quality of generated images remains approximately the same level as the original SD, as shown in Table 4. The observed differences in the reported scores and these in our experiments can be attributed to the randomness during image generation and caption sampling, which aligns with the inconsistencies reported in other studies [5]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 356, + 547, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 429 + ], + "type": "text", + "content": "Sensitivity to Hyperparameters In Appendix F, we investigate the sensitivity of our approach to hyperparameters, finding that it is less affected by factors such as the number of training images or different input prompts. Additionally, we demonstrate that our approach can leverage existing datasets to discover concept vectors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 439, + 379, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 379, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 379, + 451 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 459, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 639 + ], + "type": "text", + "content": "In this study, we introduced a self-discovery approach to identify semantic concepts in the latent space of text-to-image diffusion models. Our research findings highlight that the generation of inappropriate content can be attributed to ethical-related concepts present in the internal semantic space of diffusion models. Leveraging these concept vectors, we enable responsible generation, including promoting equality among societal groups, eliminating inappropriate content, and enhancing text guidance for responsible prompts. Through extensive experiments, we have demonstrated the effectiveness and superiority of our proposed approach. Our work contributes to the understanding of internal representations in diffusion models and facilitates the generation of responsible content, maximizing the utility of high-quality text-to-image generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement This work is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1, EPSRC/MURI grant: EP/N019474/1. We thank the Royal Academy of Engineering and FiveAI. This work is also founded by the German Federal Ministry of Education and Research and the Bavarian State Ministry for Science and the Arts." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12013" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "type": "text", + "content": "[3] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 288, + 315 + ], + "type": "text", + "content": "[4] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable detection of offensive and non-compliant content/ logo in product images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2247-2256, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 288, + 360 + ], + "type": "text", + "content": "[5] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 361, + 288, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 288, + 405 + ], + "type": "text", + "content": "[6] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 450 + ], + "type": "text", + "content": "[7] Jindong Gu, Ahmad Beirami, Xuezhi Wang, Alex Beutel, Philip Torr, and Yao Qin. Towards robust prompts on vision-language models. arXiv preprint arXiv:2304.08479, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 288, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 288, + 506 + ], + "type": "text", + "content": "[8] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 508, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 508, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 508, + 288, + 552 + ], + "type": "text", + "content": "[9] René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. arXiv preprint arXiv:2303.11073, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 288, + 586 + ], + "type": "text", + "content": "[10] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. arXiv preprint arXiv:2305.10120, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 588, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 288, + 632 + ], + "type": "text", + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 633, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 633, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 633, + 288, + 666 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 668, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 288, + 700 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "text", + "content": "[14] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[16] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "text", + "content": "[17] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[18] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "text", + "content": "[19] Yipeng Leng, Qiangjuan Huang, Zhiyuan Wang, Yangyang Liu, and Haoyu Zhang. Diffusegae: Controllable and high-fidelity image manipulation from disentangled representation. arXiv preprint arXiv:2307.05899, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 344, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 388 + ], + "type": "text", + "content": "[20] Hang Li, Jindong Gu, Rajat Koner, Sahand Sharifzadeh, and Volker Tresp. Do dall-e and flamingo understand each other? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1999-2010, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 389, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 545, + 455 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[22] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[23] Haochen Luo, Jindong Gu, Fengyuan Liu, and Philip Torr. An image is worth 1000 lies: Transferability of adversarial images across prompts on vision-language models. In The Twelfth International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "text", + "content": "[24] Avery Ma, Amir-massoud Farahmand, Yangchen Pan, Philip Torr, and Jindong Gu. Improving adversarial transferability via model alignment. arXiv preprint arXiv:2311.18495, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "text", + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 8, 15" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "type": "text", + "content": "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12014" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Minheng Ni, Chenfei Wu, Xiaodong Wang, Shengming Yin, Lijuan Wang, Zicheng Liu, and Nan Duan. Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "text", + "content": "[28] Zixuan Ni, Longhui Wei, Jiacheng Li, Siliang Tang, Yueting Zhuang, and Qi Tian. Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8900-8909, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[29] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7053-7061, 2023. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "text", + "content": "[30] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "type": "text", + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[32] Vinay Uday Prabhu and Abeba Birhane. Large image datasets: A pyrrhic win for computer vision? arXiv preprint arXiv:2006.16923, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 413 + ], + "type": "text", + "content": "[33] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 480 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "text", + "content": "[36] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 678 + ], + "type": "text", + "content": "[38] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1350–1361, 2022. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 679, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 679, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 679, + 287, + 714 + ], + "type": "text", + "content": "[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 127 + ], + "type": "text", + "content": "[40] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. arXiv preprint arXiv:2309.11497, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 128, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 128, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 128, + 545, + 182 + ], + "type": "text", + "content": "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 183, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 545, + 228 + ], + "type": "text", + "content": "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 293 + ], + "type": "text", + "content": "[43] Matthew Trager, Pramuditha Perera, Luca Zancato, Alessandro Achille, Parminder Bhatia, and Stefano Soatto. Linear spaces of meanings: Compositional structures in vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15395-15404, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 293, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 348 + ], + "type": "text", + "content": "[44] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "type": "text", + "content": "[45] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. arXiv preprint arXiv:2306.08757, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 403, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 403, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 403, + 545, + 470 + ], + "type": "text", + "content": "[46] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[47] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "text", + "content": "[48] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 569, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 613 + ], + "type": "text", + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "text", + "content": "[50] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[51] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in Neural Information Processing Systems, 35:22117-22130, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12015" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "text", + "content": "[52] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. arXiv preprint arXiv:1804.06876, 2018. 5, 6" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12016" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_content_list.json b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ecff4b0e0433de4bf9f7709c6d8be4b654a5589b --- /dev/null +++ b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_content_list.json @@ -0,0 +1,1507 @@ +[ + { + "type": "text", + "text": "Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors", + "text_level": 1, + "bbox": [ + 101, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nicolae-Cătălin Ristea $^{1,2,\\diamond}$ , Florinel-Alin Croitoru $^{1,\\diamond}$ , Radu Tudor Ionescu $^{1,3,*}$ , Marius Popescu $^{1,3}$ , Fahad Shahbaz Khan $^{4,5}$ , Mubarak Shah $^{6}$", + "bbox": [ + 99, + 167, + 867, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Bucharest, Romania, 2NUST Politehnica Bucharest, Romania, 3SecurifAI, Romania, 4MBZ University of Artificial Intelligence, UAE, 5Linköping University, Sweden, 6University of Central Florida, US", + "bbox": [ + 179, + 205, + 787, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 282, + 313, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose an efficient abnormal event detection model based on a lightweight masked auto-encoder (AE) applied at the video frame level. The novelty of the proposed model is threefold. First, we introduce an approach to weight tokens based on motion gradients, thus shifting the focus from the static background scene to the foreground objects. Second, we integrate a teacher decoder and a student decoder into our architecture, leveraging the discrepancy between the outputs given by the two decoders to improve anomaly detection. Third, we generate synthetic abnormal events to augment the training videos, and task the masked AE model to jointly reconstruct the original frames (without anomalies) and the corresponding pixel-level anomaly maps. Our design leads to an efficient and effective model, as demonstrated by the extensive experiments carried out on four benchmarks: Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The empirical results show that our model achieves an excellent trade-off between speed and accuracy, obtaining competitive AUC scores, while processing 1655 FPS. Hence, our model is between 8 and 70 times faster than competing methods. We also conduct an ablation study to justify our design. Our code is freely available at: https://github.com/ristea/aed-mae.", + "bbox": [ + 75, + 303, + 473, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 659, + 209, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, research on abnormal event detection in video gained significant traction [1, 10, 17, 18, 26-28, 36, 38, 43, 44, 49, 52, 57, 58, 61, 62, 65, 69, 76, 78, 80, 83, 87, 90, 95, 97-100], due to its utter importance in video surveillance. Despite the growing interest, video anomaly detection remains a complex task, owing its complexity to the fact that abnormal situations are context-dependent and do not occur very often. This makes it very difficult to collect a representative set of abnormal events for training state-of-the-art deep learning models in a fully supervised manner. To showcase the rarity and reliance on context of anomalies, we refer to the vehicle ramming attacks carried out by terrorists against pedestrians. As soon as a car is steered", + "bbox": [ + 75, + 680, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1873e97897d64d12c82b39939d5b3a6ded4ec79e35d3f775ddeb74fa8f4e1379.jpg", + "image_caption": [ + "Figure 1. Our masked auto-encoder for abnormal event detection based on self-distillation. At training time, some video frames are augmented with synthetic anomalies. The teacher decoder learns to reconstruct original frames (without anomalies) and predict anomaly maps. The student decoder learns to reproduce the teacher's output. Motion gradients are aggregated at the token level and used as weights for the reconstruction loss. Red dashed lines represent steps executed only during training." + ], + "image_footnote": [], + "bbox": [ + 504, + 282, + 890, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on the sidewalk, it becomes an abnormal event. Hence, the place where the car is driven (street versus sidewalk) determines the normal or abnormal label of the action, i.e. the label depends on context. Furthermore, there are less than 200 vehicle ramming attacks registered to date1, confirming", + "bbox": [ + 496, + 790, + 893, + 868 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*corresp. author: raducu.ionescu@gmail.com; equal contribution", + "bbox": [ + 94, + 886, + 450, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://en.wikipedia.org/wiki/Vehicle-rammingattack", + "bbox": [ + 500, + 875, + 892, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15984", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ac81caecf002633eda4ec46e306012c427e194757851963aa694c74f90abdbf.jpg", + "image_caption": [ + "Figure 2. Performance versus speed trade-offs for our self-distilled masked AE and several state-of-the-art methods [26-28, 47, 49, 60, 61, 69, 84] (with open-sourced code), on the Avenue data set. The running times of all methods are measured on a computer with one Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 472, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the scarcity of such events (even less are caught on video).", + "bbox": [ + 75, + 477, + 464, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since training anomaly detectors under a fully supervised setting is not possible, most studies dealing with abnormal event detection took a distinct path, proposing variations of outlier detection methods [3, 13, 14, 17, 21, 22, 29, 37, 41, 43, 45, 47, 51, 53, 55, 56, 61-68, 71, 80, 90, 92, 102-105]. Such methods treat abnormal event detection as an outlier detection task, where a normality model trained on normal events is applied on both normal and abnormal events during inference, labeling events deviating from the learned model as abnormal. Different from the mainstream path based on outlier detection, we propose an approach to augment each training video scene with synthetic anomalies, by randomly superimposing temporal action segments from the synthetic UBnormal data set [1] on our real-world data sets. We thus introduce synthetic anomalies at training time, enabling our model to learn in an open-set supervised manner. Additionally, we force our model to reconstruct the original training frames (without anomalies) to limit its ability to reconstruct anomalies, hence generating higher errors when anomalies occur.", + "bbox": [ + 73, + 492, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A large body of work on video anomaly detection has focused on employing auto-encoders (AEs) to address the task [5, 22, 27-29, 36, 49, 81, 85], relying on the poor reconstruction capabilities of these models on out-of-distribution data. Since training is carried out only on normal examples, it is expected for AEs to exhibit high reconstruction errors when anomalies occur. However, several researchers", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "observed that AEs generalize too well [5, 36], being able to reconstruct anomalies with very high precision. Thus, to better leverage the reconstruction error of AEs in anomaly detection, researchers explored a few alternatives, from the use of dummy [36] or pseudo-anomalies [5, 27] to the integration of memory modules [28, 49, 61]. With the same purpose in mind, we propose to employ masked auto-encoders [30] in anomaly detection, introducing new ways to regulate their generalization capacity. Indeed, we go beyond employing the standard masked AE framework, and propose three novel changes to enhance the anomaly detection performance of our model. First, we propose to weight tokens based on the magnitude of motion gradients, raising the importance of tokens with higher motion in the reconstruction loss. This makes our model focus on reconstructing tokens with high motion, and avoid reconstructing the background scene, which is typically static for surveillance cameras. Second, we attach a classification head to discriminate between normal and pseudo-abnormal instances in the latent encoding space. Third, we integrate a teacher decoder and a student decoder into our masked AE architecture, where the student decoder learns to distill knowledge from the already optimized teacher. To reduce our processing time, we use a shared encoder for the teacher and student models, leading to a process known as self-distillation [101]. During the self-distillation process, the shared encoder is frozen. We leverage the discrepancy between the outputs given by the teacher and student decoders along with the reconstruction error of the teacher to boost anomaly detection performance. Our entire framework, integrating these components into a meticulous design, is shown in Figure 1.", + "bbox": [ + 496, + 90, + 893, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State-of-the-art deep anomaly detectors [10, 26, 27, 84] typically rely on a costly object detection method to increase precision, limiting the processing bandwidth to one video stream per GPU, at around 20-30 FPS. However, for real-world video surveillance, e.g. monitoring an entire city with hundreds or thousands of cameras, the processing costs of object-centric video anomaly detectors are simply too high, given their power consumption and that one GPU can cost around $2,000. To this end, we turn our attention to developing a lightweight model (6 transformer blocks, 3M parameters), capable of processing around 66 video streams at 25 FPS, significantly reducing the processing costs. Different from competing models performing anomaly detection at the object [10, 18, 26, 27, 36, 84, 95] or spatio-temporal cube [15, 21, 35, 37, 41, 48, 51, 53, 55, 60, 71, 75, 103] levels, we present a model that takes whole video frames as input, which is significantly more efficient (see Figure 2).", + "bbox": [ + 496, + 563, + 893, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We carry out comprehensive experiments on four benchmarks: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. The empirical results show that our method is 8 to 70 times faster than competing methods [26-28, 47, 49, 60, 61, 69, 84], while achieving comparable ac", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15985", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "curacy levels. Aside from the main results, we conduct an ablation study showing that our novel design choices are supported by empirical evidence.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contribution is threefold:", + "bbox": [ + 96, + 138, + 374, + 151 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a lightweight masked auto-encoder for anomaly detection in video, which learns to reconstruct tokens with higher motion magnitude.", + "- We introduce a self-distillation training pipeline, leveraging the discrepancy between teacher and student decoders to obtain a significant accuracy boost for our highly efficient model (due to the shared encoder).", + "- To further boost the performance of our model, we introduce a data augmentation approach based on superimposing synthetic anomalies on normal training videos, which enables the masked AE model to learn with open-set supervision." + ], + "bbox": [ + 76, + 154, + 468, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 345, + 217, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anomaly detection in video is typically formulated as a one-class learning problem, where only normal data is available at training time. During test time, both normal and abnormal examples are present [59, 65]. There are several categories of anomaly detection approaches, including dictionary learning methods [13, 14, 21, 51, 68, 89], probabilistic models [2, 3, 25, 31, 41, 55, 56, 73, 91], change detection frameworks [15, 35, 48, 58], distance-based models [36, 37, 62, 63, 67, 71, 72, 75, 76, 79, 81] and reconstruction-based approaches [23, 28, 29, 46, 47, 53, 57, 61, 66, 69, 80]. Considering that reconstruction-based methods often reach state-of-the-art performance in anomaly detection [27, 69], a large body of works used the reconstruction-based paradigm in the past few years. To this end, we adopt this paradigm in our study.", + "bbox": [ + 75, + 368, + 468, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With respect to the level at which the anomaly detection is carried out, methods can be categorized into spatiotemporal cube-level methods [37, 47, 48, 51, 55, 56, 66, 67, 71, 75, 76, 96, 103], frame-level methods [47, 66, 67], and object-level methods [10, 18, 19, 26, 27, 36, 49, 84, 95].", + "bbox": [ + 75, + 595, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame-level and cube-level methods. Before the deep learning era, preliminary abnormal event detection models commonly relied on taking short video sequences and dividing them into spatio-temporal cuboids [15, 21, 41, 51, 55, 75, 103]. The cubes are then considered as independent examples, being passed as input to a machine learning model. This mainstream practice continued during the deep learning period [29, 35, 37, 43, 44, 53, 60, 62, 63, 71, 76, 100], when deep networks have been used to extract features [35, 37, 48, 53] or learn [28, 29, 43, 44, 53, 60, 62-64, 71, 89, 100] from these spatio-temporal cubes.", + "bbox": [ + 75, + 672, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the same time, some studies considered using entire video frames as input [47, 66, 67]. For example, Liu et al. [47] proposed an effective algorithm, which learns to reconstruct the next frame of a short video sequence. A", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "more complex approach is proposed by Ravanbakhsh et al. [67], who employ optical-flow reconstruction to predict the anomalous regions from an input image. In a different study, Ravanbakhsh et al. [66] proposed to detect anomalies at the frame level via generative adversarial networks.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame-level and cube-level methods have a common characteristic, namely their relatively high processing speed due to the reasonably fast preprocessing steps, as opposed to object-centric methods. Still, frame-level methods hold a stronger advantage in terms of time, since cube-level methods need to process each cube as an independent example. Indeed, it is more efficient to process a mini-batch of frames rather than several mini-batches of spatio-temporal cubes. However, cube-level methods often outperform frame-level methods. To this end, we propose a masked AE that takes whole frames as input, yet learns interactions between video patches, thus integrating the best of both worlds.", + "bbox": [ + 496, + 169, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To boost the performance of frame-level or cube-level methods, researchers explored the inclusion of various components, such as memory modules [28, 61] or masked convolutional blocks [69]. Although integrating additional modules into the framework leads to accuracy gains, the procedure often comes with efficiency drawbacks. In contrast, our goal is to achieve a superior trade-off between performance and speed, with a higher focus on efficiency. As such, we design a lightweight masked AE based on convolutional vision transformer (CvT) blocks [88], and propose several upgrades resulting in a minimal time overhead. For example, we employ knowledge distillation to leverage the discrepancy between the teacher and the student models. However, to keep the processing time to the bare minimum, we resort to self-distillation [101] and use a shared encoder for the teacher and student networks.", + "bbox": [ + 496, + 352, + 890, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object-level methods. To reduce the number of false positive detections often observed for other methods, some recent studies [10, 26, 27, 36, 49, 84, 95] proposed to look for anomalous objects rather than anomalous frames or cubes. Object-centric methods use the prior information from an object detector, enabling the anomaly detector to focus only on objects. This kind of framework boosts the accuracy by significant margins, currently reaching state-of-the-art performance [10, 84]. However, a considerable drawback is that the inference speed of the whole framework is directly conditioned by the object detector's speed, which is often much lower than that of the anomaly detection network [26, 36]. Hence, the processing time is significantly limited. In contrast, we perform anomaly detection at the frame level, obtaining an inference speed that is between 32 to 70 times faster than object-centric models [10, 26, 27, 36, 49, 84, 95].", + "bbox": [ + 496, + 595, + 890, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked auto-encoders in anomaly detection. He et al. [30] proposed masked auto-encoders as a pretraining method to obtain strong backbones for downstream tasks.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15986", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since then, the method has been adopted in various fields, e.g. video processing [24] or multimodal learning [7], with remarkable results. We elaborate the connection to seemingly related masked AEs in the supplementary [11, 94]. The masking framework has also been used for anomaly detection in medical [34] and industrial [39] images. To the best of our knowledge, we are the first to propose a masked transformer-based auto-encoder for video anomaly detection. Moreover, we go beyond applying standard masked AEs, proposing several modifications leading to superior performance levels: emphasizing tokens with higher motion, augmenting training videos with synthetic anomalies, and employing self-distillation.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge distillation in anomaly detection. Knowledge distillation [6, 32] was originally designed to compress one or multiple large models (teachers) into a lighter neural network (student). Recently adopted in anomaly detection [8, 12, 16, 26, 74, 86], knowledge distillation was deemed useful due to the possibility of leveraging the representation discrepancy between the teacher and the student networks, which is larger in the case of anomalies. For example, Bergmann et al. [8] trained an ensemble of student networks on normal data to reproduce the output of a deep feature extractor, which is pretrained on ImageNet [70]. The authors use the difference between the teacher label and the mean over student labels to detect abnormal pixels. Salehi et al. [74] employed a more thorough distillation process, called hint learning, in which the multi-level features of a teacher pretrained on ImageNet are distilled into a clone.", + "bbox": [ + 75, + 287, + 472, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most studies based on knowledge distillation applied the framework to image anomaly detection [8, 12, 16, 74]. With few exceptions [26, 86], knowledge distillation in video anomaly detection remains largely unexplored. Wang et al. [86] employed the teacher-student training paradigm to learn from unlabeled video samples in a self-supervised manner. Georgescu et al. [26] integrated knowledge distillation as a proxy task into a multi-task learning framework for video anomaly detection.", + "bbox": [ + 75, + 531, + 472, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distinct from the aforementioned studies, to our knowledge, we are the first to introduce a variant of self-distillation in anomaly detection. Self-distillation [101] attaches multiple classification heads at various depths to boost the classification performance of a neural classifier. In contrast, we integrate self-distillation into a masked AE, employing two decoders of different depths. Due to the shared encoder, we are able to leverage the reconstruction discrepancy between the teacher and the student with a minimal computational overhead.", + "bbox": [ + 75, + 669, + 472, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 832, + 169, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. We introduce a lightweight teacher-student transformer-based masked AE, which employs a two-stage training pipeline. In the first stage, we optimize a teacher", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "masked AE via a reconstruction loss that employs a novel weighting mechanism based on motion gradients. In the second stage, we optimize the last (and only) decoder block of a student masked AE, which shares most of the backbone (kept frozen) with its teacher, to preserve efficiency. Next, we describe how to create training videos with synthetic anomalies and train the masked AEs to jointly predict the anomaly maps and overlook (not reconstruct) the anomalies from training frames. Lastly, we introduce a classification head to distinguish between frames with and without synthetic anomalies, which further boosts the performance of our method, with a marginal computational overhead.", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. Our masked AE pursues the architectural principles proposed in [30]. Hence, the entire architecture is formed of visual transformer blocks. In contrast to He et al. [30], we replace the ViT [20] blocks with CvT blocks [88], aiming for higher efficiency. Our processing starts by dividing the input images into non-overlapping tokens and removing a certain number of tokens. The encoder embeds the remaining tokens via convolutional projection layers, and the result is processed by transformer blocks. The decoder operates on a complete set of tokens, those removed being replaced with mask tokens. Its architecture is symmetric to that of the encoder. For efficiency reasons, we only use three blocks for the encoder and three blocks for the decoder. Each block is equipped with four attention heads. To achieve further speed gains, we replace all dense layers inside the CvT blocks with pointwise convolutions. We consider the architecture described so far as a teacher network. A student decoder branches out from the teacher after the first transformer block of the main decoder, adding only one extra transformer block (as shown in Figure 1).", + "bbox": [ + 496, + 273, + 893, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion gradient weighting. Masked AEs [30] have been originally applied on natural images. In this context, reconstructing randomly masked tokens is a viable solution, since images have high foreground and background variations. However, abnormal event detection data sets [2, 51, 53, 55] contain videos from fixed cameras with static backgrounds [65]. Learning to reconstruct the static background via masked AEs is both trivial and useless. Hence, naively training masked AEs to reconstruct randomly masked tokens in video anomaly detection is suboptimal. To this end, we propose to take into account the magnitude of the motion gradients when computing the reconstruction loss.", + "bbox": [ + 496, + 577, + 893, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\pmb{x}_t\\in \\mathbb{R}^{h\\times w\\times c}$ be the video frame at index $t$ . Let $n$ be the number of non-overlapping visual tokens (patches) of size $d\\times d\\times c$ from each frame $\\pmb{x}_t$ , where $c$ is the number of input channels, and $d$ is a hyperparameter that directly determines $n$ . Let $\\left\\{\\pmb{p}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}$ denote the set of tokens in frame $\\pmb{x}_t$ , and $\\left\\{\\hat{\\pmb{p}}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}$ the corresponding set of reconstructed tokens.", + "bbox": [ + 496, + 758, + 893, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following Ionescu et al. [36], we estimate the motion", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15987", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "gradient map $\\pmb{g}_t$ of frame $\\pmb{x}_t$ by computing the absolute difference between consecutive frames, which are previously filtered with a $3\\times 3$ median filter. Next, we divide the gradient magnitude map $\\pmb{g}_t$ into non-overlapping patches, obtaining the set of gradient patches $\\left\\{\\pmb{r}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}$ . Inside each gradient patch, we compute the maximum gradient magnitude per channel. Then, we compute the channel-wise mean over the maximum gradient magnitudes, as follows:", + "bbox": [ + 76, + 90, + 472, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} ^ {(t)} = \\frac {1}{c} \\sum_ {l = 1} ^ {c} \\max _ {j, k} \\left\\{\\boldsymbol {r} _ {i j k l} ^ {(t)} \\right\\}, \\forall j, k \\in \\{1, \\dots , d \\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 234, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we compute the token-wise weights for the reconstruction loss as follows:", + "bbox": [ + 76, + 275, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {(t)} = \\frac {m _ {i} ^ {(t)}}{\\sum_ {j = 1} ^ {n} m _ {j} ^ {(t)}}, \\forall i \\in \\{1, \\dots , n \\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 305, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Introducing the resulting weights $w_{i}^{(t)}$ into the conventional token-level reconstruction loss leads to an objective that pushes the masked AE to focus on reconstructing the patches with high motion magnitude. Formally, our weighted mean squared error loss is given by:", + "bbox": [ + 76, + 351, + 468, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {w M S E}} \\left(\\boldsymbol {x} _ {t}, \\theta_ {T}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \\cdot \\left\\| \\boldsymbol {p} _ {i} ^ {(t)} - \\hat {\\boldsymbol {p}} _ {i} ^ {(t)} \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 431, + 468, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_T$ are the weights our teacher masked AE. Although our reconstruction loss focuses on tokens with high motion, the masked tokens are still chosen randomly.", + "bbox": [ + 76, + 470, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-distillation. Knowledge distillation has already shown its utility in anomaly detection [8, 12, 16, 26, 74]. Intuitively, since the teacher and student models are both trained on normal data, their reconstructions should be very similar for normal test samples. However, their behavior is not guaranteed to be similar on abnormal examples. Therefore, the magnitude of the teacher-student output gap (discrepancy) can serve as a means to quantify the anomaly level of a given sample. Unfortunately, this approach implies using both teacher and student models during inference, virtually splitting our processing speed in half. To slash the additional burden of using another model during inference, we propose to employ a novel variant of self-distillation with a shared encoder and two decoders, a teacher and a student. More precisely, the student branches out from the original architecture after the first transformer block of the teacher decoder, essentially adding only one transformer block.", + "bbox": [ + 76, + 518, + 468, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training process is carried out in two stages. In the first phase, the teacher is trained with the loss defined in Eq. (3). In the second phase, we freeze the weights of the shared backbone and train only the student decoder via self-distillation. The self-distillation loss is similar to the one defined in Eq. (3). The main difference is that instead of reconstructing the patches from the real image, the student learns to reconstruct the ones produced by the teacher. Let", + "bbox": [ + 76, + 779, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5b99fc58c65f8eb9dc5e443924c58b9508c05d837d8239da49b564db78a86a2.jpg", + "image_caption": [ + "Figure 3. Four synthetic anomalies (with red contours) taken from the UBnormal data set [1] and overlaid on training frames from Avenue [51]. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 544, + 90, + 849, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\left\\{\\tilde{\\pmb{p}}_i^{(t)}\\right\\}_{i = 1}^n \\in \\mathbb{R}^{d\\times d\\times c}$ denote the patches reconstructed by the student. Then, the self-distillation loss can be expressed as follows:", + "bbox": [ + 500, + 273, + 890, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S D}} \\left(\\hat {\\boldsymbol {x}} _ {t}, \\theta_ {S}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \\cdot \\left\\| \\hat {\\boldsymbol {p}} _ {i} ^ {(t)} - \\tilde {\\boldsymbol {p}} _ {i} ^ {(t)} \\right\\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 330, + 890, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{x}}_t$ is the frame reconstructed by the teacher, and $\\theta_S$ are the weights of the student decoder. Notice that we keep the motion gradient weights $w_i^{(t)}$ during self-distillation.", + "bbox": [ + 498, + 369, + 890, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic anomalies. As observed in other studies [5, 36], AEs tend to generalize too well to out-of-distribution data. This behavior is not desired in anomaly detection, since methods based on AEs rely on having high reconstruction errors for abnormal examples and low reconstruction errors for normal ones. To this end, we propose to augment the training videos with abnormal events. Since collecting abnormal training examples from the real-world is not possible, we resort to adding synthetic (virtual) anomalies. We leverage the recently introduced UBnormal data set [1] and its accurate pixel-level annotations to crop out abnormal events and blend them in our training videos, while ensuring the temporal consistency of the added events. The resulting examples, some depicted in Figure 3, are used to augment the training set with extra data.", + "bbox": [ + 498, + 420, + 892, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The synthetic examples help our model in three ways. First, in the reconstruction loss, we consider the original training frames (without superimposed anomalies) as the ground-truth, essentially forcing our model to overlook the anomalies. Formally, in Eq. (3), we use the patches $\\left\\{\\pmb{p}_i^{(t)}\\right\\}_{i = 1}^n$ from the normal version of frame $\\pmb{x}_t$ . Second, we add the anomaly map as an additional channel to our target image. In the anomaly map, we set normal pixels to 0 and abnormal pixels to 1. This change implies that, in Eq. (3) and Eq. (4), all patches will have an additional channel. Third, we use the ground-truth anomaly map to enhance the weights defined in Eq. (2). The added synthetic anomalies do not necessarily yield motion gradients with high magnitude. Hence, it is possible to have low weights in Eq. (3) and Eq. (4) for patches that correspond to anomaly regions. This is not desirable if we want the model to detect anom", + "bbox": [ + 498, + 651, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15988", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lies. To this end, we propose to add the anomaly maps and the gradients together, before computing the weights as in Eq. (2). Formally, in Eq. (1), we replace $\\left\\{\\pmb{r}_i^{(t)}\\right\\}_{i=1}^n$ with $\\left\\{\\pmb{r}_i^{(t)} + \\pmb{a}_i^{(t)}\\right\\}_{i=1}^n$ , where $\\left\\{\\pmb{a}_i^{(t)}\\right\\}_{i=1}^n$ is the set of the patches extracted from the anomaly map.", + "bbox": [ + 76, + 90, + 468, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classification head. We further harness the synthetic anomalies to train a classification head applied on the final [CLS] token of the shared encoder. The head is trained to discriminate between frames with and without synthetic anomalies. This head is trained using binary cross-entropy:", + "bbox": [ + 76, + 184, + 467, + 260 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} \\left(\\hat {\\boldsymbol {x}} _ {t}, \\theta_ {E}\\right) = - y _ {t} \\cdot \\log \\left(\\hat {y} _ {t}\\right) - \\left(1 - y _ {t}\\right) \\cdot \\log \\left(1 - \\hat {y} _ {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 263, + 467, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y_{t} \\in \\{0,1\\}$ is 1 if the frame contains an anomaly and 0 otherwise, $\\hat{y}_t$ is the prediction, and $\\theta_E$ represents the set of weights of the shared encoder.", + "bbox": [ + 76, + 286, + 467, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. During inference, we pass each frame $\\pmb{x}_t$ through both teacher and student models to obtain the reconstructed frames $\\hat{\\pmb{x}}_t$ and $\\tilde{\\pmb{x}}_t$ , respectively. Then, we compute output pixel-level anomaly map as:", + "bbox": [ + 76, + 330, + 467, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {o} _ {t} = \\alpha \\cdot \\| \\boldsymbol {x} _ {t} - \\hat {\\boldsymbol {x}} _ {t} \\| _ {2} ^ {2} + \\beta \\cdot \\| \\hat {\\boldsymbol {x}} _ {t} - \\tilde {\\boldsymbol {x}} _ {t} \\| _ {2} ^ {2} + \\gamma \\cdot \\hat {y} _ {t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 393, + 467, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ , $\\beta$ and $\\gamma$ are hyperparameters that control the importance of the individual anomaly score components. Following [15, 35], we apply spatio-temporal 3D filtering to smooth the anomaly volumes. To obtain the frame-level anomaly scores, we keep the maximum value from each output map $\\pmb{o}_t$ and subsequently apply another temporal Gaussian filter to smooth the values.", + "bbox": [ + 75, + 414, + 467, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 527, + 207, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 547, + 266, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data sets. We verify the performance of our method on four data sets for video anomaly detection: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. ShanghaiTech is the largest data set, with 270K frames for training and about 50K for testing. UBnormal is the second largest, with about 116K training frames and 93K testing frames. Avenue is a popular benchmark containing 15K frames for training and another 15K for testing. UCSD Ped2 holds a total of 4.5K frames, out of which 2.5K are used for training. UBnormal is a benchmark that uses an open set evaluation, where training and test anomalies belong to disjoint category sets. For the other three data sets, the training videos contain only normal events, and the test ones include both normal and abnormal scenarios. To augment the normal training videos, we sample abnormal events from the UBnormal data set [1]. UBnormal is a synthetic (virtual) data set containing anomalies simulated by video game characters, which alleviates the burden of collecting anomalies from the real world. The probability of augmenting a frame from Avenue, ShanghaiTech or UCSD Ped2 is 0.25.", + "bbox": [ + 75, + 568, + 467, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We evaluate all models following recent related works [1, 27, 69], considering both micro and macro", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/023de1281037ea04993729da30d19d250b91671fbfb8bb6cd0ab0aea2ffca105.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodAvenueShanghaiUBnormalPed2FPS
MicroMacroMicroMacroMicroMacroMicroMacro
Object-centric[10]91.692.583.890.562.186.5--20
[18]86.4-71.6---97.8--
[26]91.592.882.490.255.484.597.599.851
[27]92.390.482.789.361.385.698.799.724
[33]--85.9-71.8----
[36]87.490.478.784.9--94.397.8-
[40]80.2-73.7-50.7----
[49]89.993.574.283.2--99.3-12
[50]91.892.383.887.8-----
[54]+[10]91.692.483.690.6----20
[54]+[27]93.291.883.389.3----31
[54]+[49]89.593.675.283.8----10
[69]+[27]92.991.983.689.5----31
[69]+[49]90.992.275.583.7----10
[84]92.2-84.3---99.0-35
[95]89.6-74.8---97.3--
Frame or cube level[4]87.1-75.9---96.5--
[5]84.7-73.7---98.4--
[9]----68.580.3--37
[28]83.3-71.2---94.1-35
[37]88.9--------
[43]90.0-----96.6--
[47]85.181.772.880.6--95.4-28
[48]84.4-----87.5--
[54]+[47]89.184.874.683.3----26
[54]+[61]86.486.370.680.3----94
[57]86.9-----96.2--
[60]85.3-72.2---96.3-195
[61]82.886.868.379.7--97.0-101
[62]72.0-----88.3--
[63]87.2-----93.0--
[66]------93.5--
[67]------88.4--
[69]+[47]87.384.574.582.9----26
[69]+[61]84.888.669.880.2----95
[76]84.6--------
[77]---76.550.376.8--56
[78]89.6-74.7------
[80]85.1-73.0---96.3--
[82]--76.1------
[87]87.0-79.3------
[89]--80.4------
[90]86.6-----96.9--
[93]90.1-78.6-62.7----
[96]90.2-----97.3--
[100]--78.9------
[103]------91.0--
Ours91.390.979.184.758.581.495.498.41655
", + "bbox": [ + 514, + 88, + 870, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Micro and macro AUC scores (in %) of several state-of-the-art frame-level, cube-level and object-level methods versus our self-distilled masked AE on Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The top three scores for each category of methods are shown in red, green, and blue. All reported running times (including those of the baselines) are measured on a machine with an Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM.", + "bbox": [ + 498, + 694, + 890, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AUC metrics. The area under the ROC curve (AUC) expresses the overlap between the ground-truth frame-level annotations and the anomaly scores predicted by a model, at multiple thresholds. At a given threshold, a frame is labeled as abnormal if the predicted anomaly score is above the threshold. For the micro AUC, the test frames from all", + "bbox": [ + 498, + 809, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15989", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4756055ee566935d5e1262423f127167da719791fff36bc6342a50cd1feb2f7b.jpg", + "image_caption": [ + "Figure 4. Predictions for test video 04 from Avenue. The abnormal bounding boxes are given by the convex hull of the patches labeled as abnormal. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 450, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "videos are concatenated before computing the AUC over all frames. For the macro AUC, the AUC of each test video is first computed, and the resulting AUC scores are averaged to obtain a single value.", + "bbox": [ + 76, + 328, + 467, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameters. The encoder module is formed of three CvT blocks, each with a projection size of 256 and four attention heads. The teacher decoder contains three CvT blocks, while the student decoder contains only one block. All decoder blocks have four attention heads and a projection dimension of 128. Since the data sets have different input resolutions and objects vary in size, we adapt the patch size to each data set. Thus, we set the patch size to $16 \\times 16$ on Avenue, $8 \\times 8$ on ShanghaiTech and UBnormal, and $4 \\times 4$ on UCSD Ped2. Regardless of the data set, the teacher network is trained for 100 epochs, while the student is trained for 40 epochs. We optimize the networks with Adam [42], using a learning rate of $10^{-4}$ and mini-batches of 100 samples. The hyperparameters in Eq. (6) are set to $\\alpha = 0.4$ , $\\beta = 0.3$ and $\\gamma = 0.3$ , for all data sets.", + "bbox": [ + 76, + 388, + 467, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 622, + 171, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present our results in Table 1 and discuss them below.", + "bbox": [ + 76, + 642, + 457, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Avenue. Our method obtains a micro AUC score of $91.3\\%$ on Avenue, being only $1.9\\%$ below the state-of-the-art object-centric method. Remarkably, in the category of frame-level methods, we reach the best micro and macro AUC scores. Taking into account that our method is much faster than all other methods, we consider that its performance is remarkable. In Figure 4, we illustrate the anomaly scores for test video 04 from Avenue. Here, our model is close to perfect, highlighting its ability to capture anomalies, such as people running.", + "bbox": [ + 75, + 657, + 467, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ShanghaiTech. On ShanghaiTech, our method reaches the top macro AUC score and the third-best micro AUC score, when compared with other frame-level frameworks. Object-centric methods generally surpass frame-level methods, but the former methods have much lower processing speeds (see Figure 2).", + "bbox": [ + 76, + 810, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a5620e1311180f04a6ffc7974c630c7e64969b4dca5220d06ab0d393a0a02b7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Motion weightsSelf-distillationSynthetic dataAnomaly mapsClassif. headAvenueShanghai
AUC
MicroMacroMicroMacro
84.085.669.780.1
✓84.886.371.380.9
✓✓88.586.076.383.8
✓✓✓88.586.977.083.0
✓✓✓✓90.589.677.384.2
✓✓✓✓✓91.390.979.184.7
", + "bbox": [ + 501, + 88, + 890, + 199 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fe0eac877265251638db5585abb989587e31994feae634d90ac2c339a69c41f1.jpg", + "table_caption": [ + "Table 2. Impact of each novel component on the micro and macro AUC scores (in %), on Avenue [51] and ShanghaiTech [53]." + ], + "table_footnote": [], + "table_body": "
StrategyAvenueShanghai
AUC
MicroMacroMicroMacro
Teacher84.085.674.882.3
Teacher + Student85.485.875.182.1
Teacher + Teacher-Student Difference88.586.076.383.8
Teacher + Student + Teacher-Student Difference86.985.875.883.6
", + "bbox": [ + 501, + 244, + 890, + 332 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b9f449051c8e619c72a857116019211bbec35abeceddfd8db0d0376332db5bdd.jpg", + "table_caption": [ + "Table 3. Impact of strategies to combine the outputs of the teacher and student models on Avenue [51] and ShanghaiTech [53]. These results do not include the synthetic anomalies and the classification head." + ], + "table_footnote": [], + "table_body": "
Data setMeasurePercentage of synthetic data
0%25%50%75%
AvenueMicro AUC88.591.390.689.9
Macro AUC86.090.989.487.7
ShanghaiMicro AUC76.379.177.977.7
Macro AUC83.884.784.484.1
", + "bbox": [ + 521, + 405, + 864, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Impact of varying the proportion of synthetic anomalies on the Avenue [51] and ShanghaiTech [53] data sets.", + "bbox": [ + 500, + 484, + 890, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on UBnormal. In terms of the micro AUC, the best frame-level method on UBnormal is TimeSformer [9], which benefits from large-scale pretraining. Notably, our method obtains a higher macro AUC than TimeSformer. The micro AUC of our method is fairly close to the micro AUC levels of the better object-centric approaches. In terms of speed, our method is significantly faster than all the other methods reporting results on UBnormal.", + "bbox": [ + 498, + 534, + 890, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on UCSD Ped2. Our framework obtains a micro AUC of $95.4\\%$ , being $3.9\\%$ below the state-of-the-art performance of Liu et al. [49]. While being slightly below in terms of micro AUC, our macro AUC score is remarkably on par with the object-centric method of Ionescu et al. [36], and only $1.4\\%$ below that of Georgescu et al. [26]. Considering that our FPS is more than 30 times higher compared with these methods [26, 36] on the same GPU, our framework provides a clearly superior accuracy-speed trade-off.", + "bbox": [ + 498, + 657, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study. In Table 2, we illustrate the impact of each novel component on our model's performance. The first model is a vanilla masked AE, which obtains a rather low performance on both Avenue and ShanghaiTech. Each and every component contributes towards boosting the performance of the vanilla model. We observe that self-distillation gives the highest boost in terms of the micro AUC. How-", + "bbox": [ + 498, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15990", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f893c7daf0da06f86eed77efa86fcde04a3a3a9f321e024967f37e01b4f63141.jpg", + "image_caption": [ + "Figure 5. Examples of frames and anomaly maps reconstructed by our teacher. The first four columns correspond to abnormal examples from the Avenue and ShanghaiTech data sets, while the last column shows a normal example. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 176, + 85, + 795, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ever, to surpass the $90\\%$ milestone on Avenue, it is mandatory to introduce the prediction of anomaly maps in the learning task. An additional boost is given by our classification head.", + "bbox": [ + 75, + 354, + 470, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The self-distillation procedure gives us a few possible strategies to combine the outputs of the teacher and the student. Hence, we investigate this aspect and report the results in Table 3. We observe that the best micro AUC is obtained when we combine the teacher reconstruction error with the teacher-student discrepancy.", + "bbox": [ + 75, + 415, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proportion of synthetic examples per mini-batch is another aspect that can influence our model's performance. In Table 4, we report the micro and macro AUC scores for three possible augmentation levels. We observe that augmentation is always useful, but, for a better outcome, it requires a moderate percentage (25%).", + "bbox": [ + 75, + 506, + 468, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance-speed trade-off. In Figure 2, we compare our model with several other methods in terms of the performance-speed trade-off. This comparison undoubtedly shows that our method reaches a far better processing speed, while achieving fairly good performance. To strengthen this observation, we also compare the methods in terms of GFLOPs and number of parameters in Table 5. We underline that the method of Gong et al. [28] might seem small in terms of the number of parameters, but it is slowed down by its input, which is formed of a cuboid constructed by stacking 16 consecutive frames. Moreover, the method relies on a memory module, where each memory slot records the features of one pixel in the activation maps. Although their method has twice as many parameters as our own, the large input volume and the sizable memory bank reduce the speed to 35 FPS. With an FPS of 1655, our method proves to be significantly lighter than all its competitors.", + "bbox": [ + 75, + 598, + 470, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. In Figure 5, we present the frames and anomaly maps reconstructed by the teacher in four abnormal scenarios from Avenue and ShanghaiTech. Moreover,", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa82f89c710e1c30f7b886acaeda1919d07cfe35474f4e129816cf06483b4e18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGFLOPs ↓#Params (M) ↓FPS ↑
Georgescu et al. [26]107.96551
Georgescu et al. [27]121.66724
Liu et al. [49]179.532012
Park et al. [60]8464195
Gong et al. [28]55.2635
Ours0.831655
", + "bbox": [ + 540, + 349, + 846, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparing methods in terms of floating point operations (GFLOPs), number of parameters, and FPS.", + "bbox": [ + 500, + 441, + 890, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the fifth column, we illustrate the behavior of the teacher in a normal scenario. In all four abnormal cases, the reconstruction error is visibly higher for anomalous regions. This effect is mostly due to our training procedure based on synthetic data augmentation. The most obvious example is in the fourth column, where the bicycle seen in a pedestrian area is almost entirely removed from the output. Additionally, the predicted anomaly maps are well aligned with the ground-truth ones.", + "bbox": [ + 496, + 489, + 890, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 636, + 619, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed a lightweight masked autoencoder (3M parameters, 0.8 GFLOPs) for video anomaly detection, which learns to reconstruct tokens with high motion gradients. Our framework is based on self-distillation, leveraging the discrepancy between teacher and student decoders for anomaly detection. Moreover, we boost the performance of our model by introducing a data augmentation technique based on overlapping synthetic anomalies on normal training data. Our highly efficient framework reached an unprecedented speed of 1655 FPS, with a minimal performance gap with respect to the state-of-the-art object-centric approaches.", + "bbox": [ + 496, + 657, + 890, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. Work supported by a grant of the Romanian Ministry of Education and Research, CNCS-UEFISCDI, project number PN-III-P2-2.1-PED-2021-0195, within PNCDI III.", + "bbox": [ + 496, + 839, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15991", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1]Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. UBnormal: New Benchmark for Supervised Open-Set Video Anomaly Detection. In Proceedings of CVPR, pages 20143-20153, 2022. 1, 2, 5, 6", + "[2] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust Real-Time Unusual Event Detection Using Multiple Fixed-Location Monitors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):555-560, 2008. 3, 4", + "[3] Borislav Antic and Bjorn Ommer. Video parsing for abnormality detection. In Proceedings of ICCV, pages 2415-2422, 2011. 2, 3", + "[4] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. In Proceedings of BMVC, 2021. 6", + "[5] Marcella Astrid, Muhammad Zaigham Zaheer, and Seung-Ik Lee. Synthetic Temporal Anomaly Guided End-to-End Video Anomaly Detection. In Proceedings of ICCVW, pages 207-214, 2021. 2, 5, 6", + "[6] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In Proceedings of NIPS, pages 2654-2662, 2014. 4", + "[7] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. MultiMAE: Multi-modal multi-task masked autoencoders. In Proceedings of ECCV, pages 348-367. Springer, 2022. 4", + "[8] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed Students: Student-Teacher Anomaly Detection With Discriminative Latent Embeddings. In Proceedings of CVPR, pages 4183–4192, 2020. 4, 5", + "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In Proceedings of ICML, 2021. 6, 7", + "[10] Antonio Bărbălău, Radu Tudor Ionescu, Mariana-Iuliana Georgescu, Jacob Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. SSMTL++: Revisiting Self-Supervised Multi-Task Learning for Video Anomaly Detection. Computer Vision and Image Understanding, 229: 103656, 2023. 1, 2, 3, 6", + "[11] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. SdAE: Self-distillated Masked Autoencoder. In Proceedings of ECCV, pages 108-124, 2022. 4", + "[12] Hekai Cheng, Lu Yang, and Zulong Liu. Relation-based knowledge distillation for anomaly detection. In Proceedings of PRCV, pages 105-116, 2021. 4, 5", + "[13] Kai-Wen Cheng, Yie-Tarng Chen, and Wen-Hsien Fang. Video anomaly detection and localization using hierarchical feature representation and Gaussian process regression. In Proceedings of CVPR, pages 2909-2917, 2015. 2, 3", + "[14] Y. Cong, J. Yuan, and J. Liu. Sparse reconstruction cost for" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "abnormal event detection. In Proceedings of CVPR, pages 3449-3456, 2011. 2, 3", + "[15] Allison Del Giorno, J. Andrew Bagnell, and Martial Hebert. A Discriminative Framework for Anomaly Detection in Large Videos. In Proceedings of ECCV, pages 334-349, 2016. 2, 3, 6", + "[16] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of CVPR, pages 9737-9746, 2022. 4, 5", + "[17] Fei Dong, Yu Zhang, and Xiushan Nie. Dual Discriminator Generative Adversarial Network for Video Anomaly Detection. IEEE Access, 8:88170-88176, 2020. 1, 2", + "[18] Keval Doshi and Yasin Yilmaz. Any-Shot Sequential Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 934-935, 2020. 1, 2, 3, 6", + "[19] Keval Doshi and Yasin Yilmaz. Continual Learning for Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 254-255, 2020. 3", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of ICLR, 2021. 4", + "[21] Jayanta K. Dutta and Bonny Banerjee. Online Detection of Abnormal Events Using Incremental Coding Length. In Proceedings of AAAI, pages 3755-3761, 2015. 2, 3", + "[22] Yaxiang Fan, Gongjian Wen, Deren Li, Shaohua Qiu, Martin D. Levine, and Fei Xiao. Video anomaly detection and localization via Gaussian Mixture Fully Convolutional Variational Autoencoder. Computer Vision and Image Understanding, 195:102920, 2020. 2", + "[23] Ye Fei, Chaoqin Huang, Cao Jinkun, Maosen Li, Ya Zhang, and Cewu Lu. Attribute Restoration Framework for Anomaly Detection. IEEE Transactions on Multimedia, 24: 116-127, 2022. 3", + "[24] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In Proceedings of NeurIPS, 2022. 4", + "[25] Yachuang Feng, Yuan Yuan, and Xiaoqiang Lu. Learning deep event models for crowd anomaly detection. Neurocomputing, 219:548-556, 2017. 3", + "[26] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly Detection in Video via Self-Supervised and Multi-Task Learning. In Proceedings of CVPR, pages 12742–12752, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[27] Mariana Iuliana Georgescu, Radu Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A Background-Agnostic Framework with Adversarial Training for Abnormal Event Detection in Video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4505-4523, 2022. 2, 3, 6, 8", + "[28] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton Van Den Hengel. Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15992", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Anomaly Detection. In Proceedings of ICCV, pages 1705-1714, 2019. 1, 2, 3, 6, 8", + "[29] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In Proceedings of CVPR, pages 733-742, 2016. 2, 3", + "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Proceedings of CVPR, pages 16000-16009, 2022. 2, 3, 4", + "[31] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint Detection and Recounting of Abnormal Events by Learning Deep Generic Knowledge. In Proceedings of ICCV, pages 3639-3647, 2017. 3", + "[32] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the Knowledge in a Neural Network. In Proceedings of NIPS Deep Learning and Representation Learning Workshop, 2014. 4", + "[33] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In Proceedings of ICCV, pages 13545-13554, 2023. 6", + "[34] Chaoqin Huang, Qinwei Xu, Yanfeng Wang, Yu Wang, and Ya Zhang. Self-Supervised Masking for Unsupervised Anomaly Detection and Localization. IEEE Transactions on Multimedia, 2022. 4", + "[35] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In Proceedings of ICCV, pages 2895-2903, 2017. 2, 3, 6", + "[36] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-Centric AutoEncoders and Dummy Anomalies for Abnormal Event Detection in Video. In Proceedings of CVPR, pages 7842-7851, 2019. 1, 2, 3, 4, 5, 6, 7", + "[37] Radu Tudor Ionescu, Sorina Smeureanu, Marius Popescu, and Bogdan Alexe. Detecting abnormal events in video using Narrowed Normality Clusters. In Proceedings of WACV, pages 1951-1960, 2019. 2, 3, 6", + "[38] Xiangli Ji, Bairong Li, and Yuesheng Zhu. TAM-Net: Temporal Enhanced Appearance-to-Motion Generative Network for Video Anomaly Detection. In Proceedings of IJCNN, pages 1–8, 2020. 1", + "[39] Jielin Jiang, Jiale Zhu, Muhammad Bilal, Yan Cui, Neeraj Kumar, Ruihan Dou, Feng Su, and Xiaolong Xu. Masked Swin Transformer Unet for Industrial Anomaly Detection. IEEE Transactions on Industrial Informatics, 19(2):2200-2209, 2022. 4", + "[40] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. BiPOCO: Bi-Directional Trajectory Prediction with Pose Constraints for Pedestrian Anomaly Detection. In Proceedings of SL4AD, 2022. 6", + "[41] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: A space-time MRF for detecting abnormal activities with incremental updates. In Proceedings of CVPR, pages 2921-2928, 2009. 2, 3", + "[42] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of ICLR, 2015. 7" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: Bidirectional Multi-Scale Aggregation Networks for Abnormal Event Detection. IEEE Transactions on Image Processing, 29:2395-2408, 2019. 1, 2, 3, 6", + "[44] Guoqiu Li, Guanxiong Cai, Xingyu Zeng, and Rui Zhao. Scale-Aware Spatio-Temporal Relation Learning for Video Anomaly Detection. In Proceedings of ECCV, pages 333-350, 2022. 1, 3", + "[45] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(1):18-32, 2014. 2", + "[46] Zhenyu Li, Ning Li, Kaitao Jiang, Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Superpixel Masking and Inpainting for Self-Supervised Anomaly Detection. In Proceedings of BMVC, 2020. 3", + "[47] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future Frame Prediction for Anomaly Detection - A New Baseline. In Proceedings of CVPR, pages 6536-6545, 2018. 2, 3, 6", + "[48] Yusha Liu, Chun-Liang Li, and Barnabaás Póczos. Classifier Two-Sample Test for Video Anomaly Detections. In Proceedings of BMVC, 2018. 2, 3, 6", + "[49] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction. In Proceedings of ICCV, pages 13588-13597, 2021. 1, 2, 3, 6, 7, 8", + "[50] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating Anomalies for Video Anomaly Detection With Prompt-Based Feature Mapping. In Proceedings of CVPR, pages 24500-24510, 2023. 6", + "[51] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal Event Detection at 150 FPS in MATLAB. In Proceedings of ICCV, pages 2720-2727, 2013. 2, 3, 4, 5, 6, 7", + "[52] Yiwei Lu, Frank Yu, Mahesh Kumar, Krishna Reddy, and Yang Wang. Few-Shot Scene-Adaptive Anomaly Detection. In Proceedings of ECCV, pages 125–141, 2020. 1", + "[53] Weixin Luo, Wen Liu, and Shenghua Gao. A Revisit of Sparse Coding Based Anomaly Detection in Stacked RNN Framework. In Proceedings of ICCV, pages 341-349, 2017. 2, 3, 4, 6, 7", + "[54] Neelu Madan, Nicolae-Catalin Ristea, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised masked convolutional transformer block for anomaly detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):525-542, 2024. 6", + "[55] Vijay Mahadevan, Wei-Xin LI, Viral Bhalodia, and Nuno Vasconcelos. Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 1975-1981, 2010. 2, 3, 4, 6", + "[56] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of CVPR, pages 935-942, 2009. 2, 3", + "[57] Trong-Nguyen Nguyen and Jean Meunier. Anomaly Detection in Video Sequence With Appearance-Motion Correspondence. In Proceedings of ICCV, pages 1273-1283, 2019. 1, 3, 6" + ], + "bbox": [ + 509, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15993", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Guansong Pang, Cheng Yan, Chunhua Shen, Anton van den Hengel, and Xiao Bai. Self-trained Deep Ordinal Regression for End-to-End Video Anomaly Detection. In Proceedings of CVPR, pages 12173–12182, 2020. 1, 3", + "[59] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM Computing Surveys, 54(2):1-38, 2021. 3", + "[60] Chaewon Park, MyeongAh Cho, Minhyeok Lee, and Sangyoun Lee. FastAno: Fast anomaly detection via spatiotemporal patch transformation. In Proceedings of WACV, pages 2249-2259, 2022. 2, 3, 6, 8", + "[61] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection. In Proceedings of CVPR, pages 14372-14381, 2020. 1, 2, 3, 6", + "[62] Bharathkumar Ramachandra and Michael Jones. Street Scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of WACV, pages 2569-2578, 2020. 1, 3, 6", + "[63] Bharathkumar Ramachandra, Michael Jones, and Ranga Vatsavai. Learning a distance function with a Siamese network to localize anomalies in videos. In Proceedings of WACV, pages 2598-2607, 2020. 3, 6", + "[64] Bharathkumar Ramachandra, Michael Jones, and Ranga Raju Vatsavai. Perceptual metric learning for video anomaly detection. Machine Vision and Applications, 32:1432-1769, 2021. 3", + "[65] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. A Survey of Single-Scene Video Anomaly Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2293-2312, 2022. 1, 3, 4", + "[66] Mahdyar Ravanbakhsh, Moin Nabi, Enver Sangineto, Lucio Marcenaro, Carlo Regazzoni, and Nicu Sebe. Abnormal Event Detection in Videos using Generative Adversarial Nets. In Proceedings of ICIP, pages 1577-1581, 2017. 3, 6", + "[67] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-Play CNN for Crowd Motion Analysis: An Application in Abnormal Event Detection. In Proceedings of WACV, pages 1689–1698, 2018. 3, 6", + "[68] Huamin Ren, Weifeng Liu, Soren Ingvor Olsen, Sergio Escalera, and Thomas B. Moeslund. Unsupervised Behavior-Specific Dictionary Learning for Abnormal Event Detection. In Proceedings of BMVC, pages 28.1-28.13, 2015. 2, 3", + "[69] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection. In Proceedings of CVPR, pages 13576-13586, 2022. 1, 2, 3, 6", + "[70] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, Karpathy A., A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 4" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, and Reinhard Klette. Deep-Cascade: Cascading 3D Deep Neural Networks for Fast Anomaly Detection and Localization in Crowded Scenes. IEEE Transactions on Image Processing, 26(4):1992-2004, 2017. 2, 3", + "[72] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, Zahra Moayed, and Reinhard Klette. Deep-Anomaly: Fully Convolutional Neural Network for Fast Anomaly Detection in Crowded Scenes. Computer Vision and Image Understanding, 172:88-97, 2018. 3", + "[73] Babak Saleh, Ali Farhadi, and Ahmed Elgammal. Object-Centric Anomaly Detection by Attribute-Based Reasoning. In Proceedings of CVPR, pages 787-794, 2013. 3", + "[74] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution Knowledge Distillation for Anomaly Detection. In Proceedings of CVPR, pages 14902-14912, 2021. 4, 5", + "[75] Venkatesh Saligrama and Zhu Chen. Video anomaly detection based on local statistical aggregates. In Proceedings of CVPR, pages 2112-2119, 2012. 2, 3", + "[76] Sorina Smeureanu, Radu Tudor Ionescu, Marius Popescu, and Bogdan Alexe. Deep Appearance Features for Abnormal Behavior Detection in Video. In Proceedings of ICIAP, pages 779-789, 2017. 1, 3, 6", + "[77] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In Proceedings of CVPR, pages 6479–6488, 2018. 6", + "[78] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-Aware Context Reasoning for Unsupervised Abnormal Event Detection in Videos. In Proceedings of ACMMM, pages 184-192, 2020. 1, 6", + "[79] Qianru Sun, Hong Liu, and Tatsuya Harada. Online growing neural gas for anomaly detection in changing surveillance scenes. Pattern Recognition, 64(C):187-201, 2017. 3", + "[80] Yao Tang, Lin Zhao, Shanshan Zhang, Chen Gong, Guangyu Li, and Jian Yang. Integrating prediction and reconstruction for anomaly detection. Pattern Recognition Letters, 129:123-130, 2020. 1, 2, 3, 6", + "[81] Hanh T.M. Tran and David Hogg. Anomaly Detection using a Convolutional Winner-Take-All Autoencoder. In Proceedings of BMVC, 2017. 2, 3", + "[82] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised video anomaly detection. In Proceedings of ICIP, pages 2540-2544, 2023. 6", + "[83] Hung Vu, Tu Dinh Nguyen, Trung Le, Wei Luo, and Dinh Phung. Robust Anomaly Detection in Videos Using Multilevel Representations. In Proceedings of AAAI, pages 5216-5223, 2019. 1", + "[84] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Proceedings of ECCV, pages 494-511, 2022. 2, 3, 6", + "[85] L. Wang, F. Zhou, Z. Li, W. Zuo, and H. Tan. Abnormal Event Detection in Videos Using Hybrid Spatio-Temporal" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15994", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Autoencoder. In Proceedings of ICIP, pages 2276-2280, 2018. 2", + "[86] Xusheng Wang, Mingtao Pei, and Zhengang Nie. Self-trained video anomaly detection based on teacher-student model. In Proceedings of MLSP, pages 1-6, 2021. 4", + "[87] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster Attention Contrast for Video Anomaly Detection. In Proceedings of ACMMM, pages 2463-2471, 2020. 1, 6", + "[88] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvT: Introducing Convolutions to Vision Transformers. In Proceedings of ICCV, pages 22-31, 2021. 3, 4", + "[89] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In Proceedings of ECCV, pages 729-745, 2022. 3, 6", + "[90] Peng Wu, Jing Liu, and Fang Shen. A Deep One-Class Neural Network for Anomalous Event Detection in Complex Scenes. IEEE Transactions on Neural Networks and Learning Systems, 31(7):2609-2622, 2019. 1, 2, 6", + "[91] Shandong Wu, Brian E. Moore, and Mubarak Shah. Chaotic Invariants of Lagrangian Particle Trajectories for Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 2054-2060, 2010. 3", + "[92] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting Anomalous Events in Videos by Learning Deep Representations of Appearance and Motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2", + "[93] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature Prediction Diffusion Model for Video Anomaly Detection. In Proceedings of ICCV, pages 5527-5537, 2023. 6", + "[94] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 4", + "[95] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze Test Helps: Effective Video Anomaly Detection via Learning to Complete Video Events. In Proceedings of ACMMM, pages 583-591, 2020. 1, 2, 3, 6", + "[96] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Transactions on Neural Networks and Learning Systems, 33(8): 3572-3586, 2022. 3, 6", + "[97] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proceedings of CVPR, pages 19313–19322, 2022. 1", + "[98] Muhammad Zaigham Zaheer, Jin-ha Lee, Marcella Astrid, and Seung-Ik Lee. Old is Gold: Redefining the Adversarily Learned One-Class Classifier Training Paradigm. In Proceedings of CVPR, pages 14183-14193, 2020.", + "[99] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. CLAWS: Clustering Assisted" + ], + "bbox": [ + 86, + 90, + 470, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Weakly Supervised Learning with Normalcy Suppression for Anomalous Event Detection. In Proceedings of ECCV, pages 358-376, 2020.", + "[100] Muhammad Zaigham Zaheer, Arif Mahmood, Haris M. Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative Cooperative Learning for Unsupervised Video Anomaly Detection. In Proceedings of CVPR, pages 14744-14754, 2022. 1, 3, 6", + "[101] Linfeng Zhang, Chenglong Bao, and Kaisheng Ma. Self-Distillation: Towards Efficient and Compact Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4388-4403, 2022. 2, 3, 4", + "[102] Xinfeng Zhang, Su Yang, Jiulong Zhang, and Weishan Zhang. Video Anomaly Detection and Localization using Motion-field Shape Description and Homogeneity Testing. Pattern Recognition, page 107394, 2020. 2", + "[103] Ying Zhang, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Shun Sakai. Video anomaly detection based on locality sensitive hashing filters. Pattern Recognition, 59:302-311, 2016. 2, 3, 6", + "[104] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online Detection of Unusual Events in Videos via Dynamic Sparse Coding. In Proceedings of CVPR, pages 3313-3320, 2011.", + "[105] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph Convolutional Label Noise Cleaner: Train a Plug-And-Play Action Classifier for Anomaly Detection. In Proceedings of CVPR, pages 1237–1246, 2019. 2" + ], + "bbox": [ + 503, + 90, + 890, + 486 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15995", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_model.json b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5f727a5560000769d7f10d5ac9f0ee3050939852 --- /dev/null +++ b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_model.json @@ -0,0 +1,2655 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.87, + 0.154 + ], + "angle": 0, + "content": "Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.169, + 0.868, + 0.205 + ], + "angle": 0, + "content": "Nicolae-Cătălin Ristea\\(^{1,2,\\diamond}\\), Florinel-Alin Croitoru\\(^{1,\\diamond}\\), Radu Tudor Ionescu\\(^{1,3,*}\\), Marius Popescu\\(^{1,3}\\), Fahad Shahbaz Khan\\(^{4,5}\\), Mubarak Shah\\(^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.206, + 0.788, + 0.26 + ], + "angle": 0, + "content": "1University of Bucharest, Romania, 2NUST Politehnica Bucharest, Romania, 3SecurifAI, Romania, 4MBZ University of Artificial Intelligence, UAE, 5Linköping University, Sweden, 6University of Central Florida, US" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.283, + 0.314, + 0.299 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.474, + 0.652 + ], + "angle": 0, + "content": "We propose an efficient abnormal event detection model based on a lightweight masked auto-encoder (AE) applied at the video frame level. The novelty of the proposed model is threefold. First, we introduce an approach to weight tokens based on motion gradients, thus shifting the focus from the static background scene to the foreground objects. Second, we integrate a teacher decoder and a student decoder into our architecture, leveraging the discrepancy between the outputs given by the two decoders to improve anomaly detection. Third, we generate synthetic abnormal events to augment the training videos, and task the masked AE model to jointly reconstruct the original frames (without anomalies) and the corresponding pixel-level anomaly maps. Our design leads to an efficient and effective model, as demonstrated by the extensive experiments carried out on four benchmarks: Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The empirical results show that our model achieves an excellent trade-off between speed and accuracy, obtaining competitive AUC scores, while processing 1655 FPS. Hence, our model is between 8 and 70 times faster than competing methods. We also conduct an ablation study to justify our design. Our code is freely available at: https://github.com/ristea/aed-mae." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.66, + 0.21, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.877 + ], + "angle": 0, + "content": "In recent years, research on abnormal event detection in video gained significant traction [1, 10, 17, 18, 26-28, 36, 38, 43, 44, 49, 52, 57, 58, 61, 62, 65, 69, 76, 78, 80, 83, 87, 90, 95, 97-100], due to its utter importance in video surveillance. Despite the growing interest, video anomaly detection remains a complex task, owing its complexity to the fact that abnormal situations are context-dependent and do not occur very often. This makes it very difficult to collect a representative set of abnormal events for training state-of-the-art deep learning models in a fully supervised manner. To showcase the rarity and reliance on context of anomalies, we refer to the vehicle ramming attacks carried out by terrorists against pedestrians. As soon as a car is steered" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.283, + 0.891, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.67, + 0.894, + 0.782 + ], + "angle": 0, + "content": "Figure 1. Our masked auto-encoder for abnormal event detection based on self-distillation. At training time, some video frames are augmented with synthetic anomalies. The teacher decoder learns to reconstruct original frames (without anomalies) and predict anomaly maps. The student decoder learns to reproduce the teacher's output. Motion gradients are aggregated at the token level and used as weights for the reconstruction loss. Red dashed lines represent steps executed only during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.894, + 0.869 + ], + "angle": 0, + "content": "on the sidewalk, it becomes an abnormal event. Hence, the place where the car is driven (street versus sidewalk) determines the normal or abnormal label of the action, i.e. the label depends on context. Furthermore, there are less than 200 vehicle ramming attacks registered to date1, confirming" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.451, + 0.901 + ], + "angle": 0, + "content": "*corresp. author: raducu.ionescu@gmail.com; equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.893, + 0.9 + ], + "angle": 0, + "content": "1https://en.wikipedia.org/wiki/Vehicle-rammingattack" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15984" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.473, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.376, + 0.471, + 0.461 + ], + "angle": 0, + "content": "Figure 2. Performance versus speed trade-offs for our self-distilled masked AE and several state-of-the-art methods [26-28, 47, 49, 60, 61, 69, 84] (with open-sourced code), on the Avenue data set. The running times of all methods are measured on a computer with one Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.465, + 0.493 + ], + "angle": 0, + "content": "the scarcity of such events (even less are caught on video)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.493, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Since training anomaly detectors under a fully supervised setting is not possible, most studies dealing with abnormal event detection took a distinct path, proposing variations of outlier detection methods [3, 13, 14, 17, 21, 22, 29, 37, 41, 43, 45, 47, 51, 53, 55, 56, 61-68, 71, 80, 90, 92, 102-105]. Such methods treat abnormal event detection as an outlier detection task, where a normality model trained on normal events is applied on both normal and abnormal events during inference, labeling events deviating from the learned model as abnormal. Different from the mainstream path based on outlier detection, we propose an approach to augment each training video scene with synthetic anomalies, by randomly superimposing temporal action segments from the synthetic UBnormal data set [1] on our real-world data sets. We thus introduce synthetic anomalies at training time, enabling our model to learn in an open-set supervised manner. Additionally, we force our model to reconstruct the original training frames (without anomalies) to limit its ability to reconstruct anomalies, hence generating higher errors when anomalies occur." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "A large body of work on video anomaly detection has focused on employing auto-encoders (AEs) to address the task [5, 22, 27-29, 36, 49, 81, 85], relying on the poor reconstruction capabilities of these models on out-of-distribution data. Since training is carried out only on normal examples, it is expected for AEs to exhibit high reconstruction errors when anomalies occur. However, several researchers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.562 + ], + "angle": 0, + "content": "observed that AEs generalize too well [5, 36], being able to reconstruct anomalies with very high precision. Thus, to better leverage the reconstruction error of AEs in anomaly detection, researchers explored a few alternatives, from the use of dummy [36] or pseudo-anomalies [5, 27] to the integration of memory modules [28, 49, 61]. With the same purpose in mind, we propose to employ masked auto-encoders [30] in anomaly detection, introducing new ways to regulate their generalization capacity. Indeed, we go beyond employing the standard masked AE framework, and propose three novel changes to enhance the anomaly detection performance of our model. First, we propose to weight tokens based on the magnitude of motion gradients, raising the importance of tokens with higher motion in the reconstruction loss. This makes our model focus on reconstructing tokens with high motion, and avoid reconstructing the background scene, which is typically static for surveillance cameras. Second, we attach a classification head to discriminate between normal and pseudo-abnormal instances in the latent encoding space. Third, we integrate a teacher decoder and a student decoder into our masked AE architecture, where the student decoder learns to distill knowledge from the already optimized teacher. To reduce our processing time, we use a shared encoder for the teacher and student models, leading to a process known as self-distillation [101]. During the self-distillation process, the shared encoder is frozen. We leverage the discrepancy between the outputs given by the teacher and student decoders along with the reconstruction error of the teacher to boost anomaly detection performance. Our entire framework, integrating these components into a meticulous design, is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.895, + 0.822 + ], + "angle": 0, + "content": "State-of-the-art deep anomaly detectors [10, 26, 27, 84] typically rely on a costly object detection method to increase precision, limiting the processing bandwidth to one video stream per GPU, at around 20-30 FPS. However, for real-world video surveillance, e.g. monitoring an entire city with hundreds or thousands of cameras, the processing costs of object-centric video anomaly detectors are simply too high, given their power consumption and that one GPU can cost around $2,000. To this end, we turn our attention to developing a lightweight model (6 transformer blocks, 3M parameters), capable of processing around 66 video streams at 25 FPS, significantly reducing the processing costs. Different from competing models performing anomaly detection at the object [10, 18, 26, 27, 36, 84, 95] or spatio-temporal cube [15, 21, 35, 37, 41, 48, 51, 53, 55, 60, 71, 75, 103] levels, we present a model that takes whole video frames as input, which is significantly more efficient (see Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We carry out comprehensive experiments on four benchmarks: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. The empirical results show that our method is 8 to 70 times faster than competing methods [26-28, 47, 49, 60, 61, 69, 84], while achieving comparable ac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15985" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "curacy levels. Aside from the main results, we conduct an ablation study showing that our novel design choices are supported by empirical evidence." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.139, + 0.375, + 0.152 + ], + "angle": 0, + "content": "In summary, our contribution is threefold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.155, + 0.469, + 0.199 + ], + "angle": 0, + "content": "- We propose a lightweight masked auto-encoder for anomaly detection in video, which learns to reconstruct tokens with higher motion magnitude." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.2, + 0.468, + 0.259 + ], + "angle": 0, + "content": "- We introduce a self-distillation training pipeline, leveraging the discrepancy between teacher and student decoders to obtain a significant accuracy boost for our highly efficient model (due to the shared encoder)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.261, + 0.468, + 0.334 + ], + "angle": 0, + "content": "- To further boost the performance of our model, we introduce a data augmentation approach based on superimposing synthetic anomalies on normal training videos, which enables the masked AE model to learn with open-set supervision." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.155, + 0.469, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.347, + 0.218, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.469, + 0.596 + ], + "angle": 0, + "content": "Anomaly detection in video is typically formulated as a one-class learning problem, where only normal data is available at training time. During test time, both normal and abnormal examples are present [59, 65]. There are several categories of anomaly detection approaches, including dictionary learning methods [13, 14, 21, 51, 68, 89], probabilistic models [2, 3, 25, 31, 41, 55, 56, 73, 91], change detection frameworks [15, 35, 48, 58], distance-based models [36, 37, 62, 63, 67, 71, 72, 75, 76, 79, 81] and reconstruction-based approaches [23, 28, 29, 46, 47, 53, 57, 61, 66, 69, 80]. Considering that reconstruction-based methods often reach state-of-the-art performance in anomaly detection [27, 69], a large body of works used the reconstruction-based paradigm in the past few years. To this end, we adopt this paradigm in our study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.469, + 0.673 + ], + "angle": 0, + "content": "With respect to the level at which the anomaly detection is carried out, methods can be categorized into spatiotemporal cube-level methods [37, 47, 48, 51, 55, 56, 66, 67, 71, 75, 76, 96, 103], frame-level methods [47, 66, 67], and object-level methods [10, 18, 19, 26, 27, 36, 49, 84, 95]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Frame-level and cube-level methods. Before the deep learning era, preliminary abnormal event detection models commonly relied on taking short video sequences and dividing them into spatio-temporal cuboids [15, 21, 41, 51, 55, 75, 103]. The cubes are then considered as independent examples, being passed as input to a machine learning model. This mainstream practice continued during the deep learning period [29, 35, 37, 43, 44, 53, 60, 62, 63, 71, 76, 100], when deep networks have been used to extract features [35, 37, 48, 53] or learn [28, 29, 43, 44, 53, 60, 62-64, 71, 89, 100] from these spatio-temporal cubes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "At the same time, some studies considered using entire video frames as input [47, 66, 67]. For example, Liu et al. [47] proposed an effective algorithm, which learns to reconstruct the next frame of a short video sequence. A" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "more complex approach is proposed by Ravanbakhsh et al. [67], who employ optical-flow reconstruction to predict the anomalous regions from an input image. In a different study, Ravanbakhsh et al. [66] proposed to detect anomalies at the frame level via generative adversarial networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Frame-level and cube-level methods have a common characteristic, namely their relatively high processing speed due to the reasonably fast preprocessing steps, as opposed to object-centric methods. Still, frame-level methods hold a stronger advantage in terms of time, since cube-level methods need to process each cube as an independent example. Indeed, it is more efficient to process a mini-batch of frames rather than several mini-batches of spatio-temporal cubes. However, cube-level methods often outperform frame-level methods. To this end, we propose a masked AE that takes whole frames as input, yet learns interactions between video patches, thus integrating the best of both worlds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.594 + ], + "angle": 0, + "content": "To boost the performance of frame-level or cube-level methods, researchers explored the inclusion of various components, such as memory modules [28, 61] or masked convolutional blocks [69]. Although integrating additional modules into the framework leads to accuracy gains, the procedure often comes with efficiency drawbacks. In contrast, our goal is to achieve a superior trade-off between performance and speed, with a higher focus on efficiency. As such, we design a lightweight masked AE based on convolutional vision transformer (CvT) blocks [88], and propose several upgrades resulting in a minimal time overhead. For example, we employ knowledge distillation to leverage the discrepancy between the teacher and the student models. However, to keep the processing time to the bare minimum, we resort to self-distillation [101] and use a shared encoder for the teacher and student networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Object-level methods. To reduce the number of false positive detections often observed for other methods, some recent studies [10, 26, 27, 36, 49, 84, 95] proposed to look for anomalous objects rather than anomalous frames or cubes. Object-centric methods use the prior information from an object detector, enabling the anomaly detector to focus only on objects. This kind of framework boosts the accuracy by significant margins, currently reaching state-of-the-art performance [10, 84]. However, a considerable drawback is that the inference speed of the whole framework is directly conditioned by the object detector's speed, which is often much lower than that of the anomaly detection network [26, 36]. Hence, the processing time is significantly limited. In contrast, we perform anomaly detection at the frame level, obtaining an inference speed that is between 32 to 70 times faster than object-centric models [10, 26, 27, 36, 49, 84, 95]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Masked auto-encoders in anomaly detection. He et al. [30] proposed masked auto-encoders as a pretraining method to obtain strong backbones for downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15986" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "Since then, the method has been adopted in various fields, e.g. video processing [24] or multimodal learning [7], with remarkable results. We elaborate the connection to seemingly related masked AEs in the supplementary [11, 94]. The masking framework has also been used for anomaly detection in medical [34] and industrial [39] images. To the best of our knowledge, we are the first to propose a masked transformer-based auto-encoder for video anomaly detection. Moreover, we go beyond applying standard masked AEs, proposing several modifications leading to superior performance levels: emphasizing tokens with higher motion, augmenting training videos with synthetic anomalies, and employing self-distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Knowledge distillation in anomaly detection. Knowledge distillation [6, 32] was originally designed to compress one or multiple large models (teachers) into a lighter neural network (student). Recently adopted in anomaly detection [8, 12, 16, 26, 74, 86], knowledge distillation was deemed useful due to the possibility of leveraging the representation discrepancy between the teacher and the student networks, which is larger in the case of anomalies. For example, Bergmann et al. [8] trained an ensemble of student networks on normal data to reproduce the output of a deep feature extractor, which is pretrained on ImageNet [70]. The authors use the difference between the teacher label and the mean over student labels to detect abnormal pixels. Salehi et al. [74] employed a more thorough distillation process, called hint learning, in which the multi-level features of a teacher pretrained on ImageNet are distilled into a clone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.473, + 0.669 + ], + "angle": 0, + "content": "Most studies based on knowledge distillation applied the framework to image anomaly detection [8, 12, 16, 74]. With few exceptions [26, 86], knowledge distillation in video anomaly detection remains largely unexplored. Wang et al. [86] employed the teacher-student training paradigm to learn from unlabeled video samples in a self-supervised manner. Georgescu et al. [26] integrated knowledge distillation as a proxy task into a multi-task learning framework for video anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.473, + 0.821 + ], + "angle": 0, + "content": "Distinct from the aforementioned studies, to our knowledge, we are the first to introduce a variant of self-distillation in anomaly detection. Self-distillation [101] attaches multiple classification heads at various depths to boost the classification performance of a neural classifier. In contrast, we integrate self-distillation into a masked AE, employing two decoders of different depths. Due to the shared encoder, we are able to leverage the reconstruction discrepancy between the teacher and the student with a minimal computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.17, + 0.849 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Overview. We introduce a lightweight teacher-student transformer-based masked AE, which employs a two-stage training pipeline. In the first stage, we optimize a teacher" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "masked AE via a reconstruction loss that employs a novel weighting mechanism based on motion gradients. In the second stage, we optimize the last (and only) decoder block of a student masked AE, which shares most of the backbone (kept frozen) with its teacher, to preserve efficiency. Next, we describe how to create training videos with synthetic anomalies and train the masked AEs to jointly predict the anomaly maps and overlook (not reconstruct) the anomalies from training frames. Lastly, we introduce a classification head to distinguish between frames with and without synthetic anomalies, which further boosts the performance of our method, with a marginal computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Architecture. Our masked AE pursues the architectural principles proposed in [30]. Hence, the entire architecture is formed of visual transformer blocks. In contrast to He et al. [30], we replace the ViT [20] blocks with CvT blocks [88], aiming for higher efficiency. Our processing starts by dividing the input images into non-overlapping tokens and removing a certain number of tokens. The encoder embeds the remaining tokens via convolutional projection layers, and the result is processed by transformer blocks. The decoder operates on a complete set of tokens, those removed being replaced with mask tokens. Its architecture is symmetric to that of the encoder. For efficiency reasons, we only use three blocks for the encoder and three blocks for the decoder. Each block is equipped with four attention heads. To achieve further speed gains, we replace all dense layers inside the CvT blocks with pointwise convolutions. We consider the architecture described so far as a teacher network. A student decoder branches out from the teacher after the first transformer block of the main decoder, adding only one extra transformer block (as shown in Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.895, + 0.759 + ], + "angle": 0, + "content": "Motion gradient weighting. Masked AEs [30] have been originally applied on natural images. In this context, reconstructing randomly masked tokens is a viable solution, since images have high foreground and background variations. However, abnormal event detection data sets [2, 51, 53, 55] contain videos from fixed cameras with static backgrounds [65]. Learning to reconstruct the static background via masked AEs is both trivial and useless. Hence, naively training masked AEs to reconstruct randomly masked tokens in video anomaly detection is suboptimal. To this end, we propose to take into account the magnitude of the motion gradients when computing the reconstruction loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.895, + 0.884 + ], + "angle": 0, + "content": "Let \\(\\pmb{x}_t\\in \\mathbb{R}^{h\\times w\\times c}\\) be the video frame at index \\(t\\). Let \\(n\\) be the number of non-overlapping visual tokens (patches) of size \\(d\\times d\\times c\\) from each frame \\(\\pmb{x}_t\\), where \\(c\\) is the number of input channels, and \\(d\\) is a hyperparameter that directly determines \\(n\\). Let \\(\\left\\{\\pmb{p}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}\\) denote the set of tokens in frame \\(\\pmb{x}_t\\), and \\(\\left\\{\\hat{\\pmb{p}}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}\\) the corresponding set of reconstructed tokens." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Following Ionescu et al. [36], we estimate the motion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15987" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.234 + ], + "angle": 0, + "content": "gradient map \\(\\pmb{g}_t\\) of frame \\(\\pmb{x}_t\\) by computing the absolute difference between consecutive frames, which are previously filtered with a \\(3\\times 3\\) median filter. Next, we divide the gradient magnitude map \\(\\pmb{g}_t\\) into non-overlapping patches, obtaining the set of gradient patches \\(\\left\\{\\pmb{r}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}\\). Inside each gradient patch, we compute the maximum gradient magnitude per channel. Then, we compute the channel-wise mean over the maximum gradient magnitudes, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.236, + 0.47, + 0.274 + ], + "angle": 0, + "content": "\\[\nm _ {i} ^ {(t)} = \\frac {1}{c} \\sum_ {l = 1} ^ {c} \\max _ {j, k} \\left\\{\\boldsymbol {r} _ {i j k l} ^ {(t)} \\right\\}, \\forall j, k \\in \\{1, \\dots , d \\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.276, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Finally, we compute the token-wise weights for the reconstruction loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.306, + 0.47, + 0.348 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {(t)} = \\frac {m _ {i} ^ {(t)}}{\\sum_ {j = 1} ^ {n} m _ {j} ^ {(t)}}, \\forall i \\in \\{1, \\dots , n \\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Introducing the resulting weights \\( w_{i}^{(t)} \\) into the conventional token-level reconstruction loss leads to an objective that pushes the masked AE to focus on reconstructing the patches with high motion magnitude. Formally, our weighted mean squared error loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.432, + 0.469, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {w M S E}} \\left(\\boldsymbol {x} _ {t}, \\theta_ {T}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \\cdot \\left\\| \\boldsymbol {p} _ {i} ^ {(t)} - \\hat {\\boldsymbol {p}} _ {i} ^ {(t)} \\right\\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.472, + 0.469, + 0.517 + ], + "angle": 0, + "content": "where \\(\\theta_T\\) are the weights our teacher masked AE. Although our reconstruction loss focuses on tokens with high motion, the masked tokens are still chosen randomly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Self-distillation. Knowledge distillation has already shown its utility in anomaly detection [8, 12, 16, 26, 74]. Intuitively, since the teacher and student models are both trained on normal data, their reconstructions should be very similar for normal test samples. However, their behavior is not guaranteed to be similar on abnormal examples. Therefore, the magnitude of the teacher-student output gap (discrepancy) can serve as a means to quantify the anomaly level of a given sample. Unfortunately, this approach implies using both teacher and student models during inference, virtually splitting our processing speed in half. To slash the additional burden of using another model during inference, we propose to employ a novel variant of self-distillation with a shared encoder and two decoders, a teacher and a student. More precisely, the student branches out from the original architecture after the first transformer block of the teacher decoder, essentially adding only one transformer block." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our training process is carried out in two stages. In the first phase, the teacher is trained with the loss defined in Eq. (3). In the second phase, we freeze the weights of the shared backbone and train only the student decoder via self-distillation. The self-distillation loss is similar to the one defined in Eq. (3). The main difference is that instead of reconstructing the patches from the real image, the student learns to reconstruct the ones produced by the teacher. Let" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.092, + 0.851, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.216, + 0.892, + 0.257 + ], + "angle": 0, + "content": "Figure 3. Four synthetic anomalies (with red contours) taken from the UBnormal data set [1] and overlaid on training frames from Avenue [51]. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.275, + 0.892, + 0.329 + ], + "angle": 0, + "content": "\\(\\left\\{\\tilde{\\pmb{p}}_i^{(t)}\\right\\}_{i = 1}^n \\in \\mathbb{R}^{d\\times d\\times c}\\) denote the patches reconstructed by the student. Then, the self-distillation loss can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.331, + 0.892, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S D}} \\left(\\hat {\\boldsymbol {x}} _ {t}, \\theta_ {S}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \\cdot \\left\\| \\hat {\\boldsymbol {p}} _ {i} ^ {(t)} - \\tilde {\\boldsymbol {p}} _ {i} ^ {(t)} \\right\\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.419 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{x}}_t\\) is the frame reconstructed by the teacher, and \\(\\theta_S\\) are the weights of the student decoder. Notice that we keep the motion gradient weights \\(w_i^{(t)}\\) during self-distillation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.421, + 0.893, + 0.649 + ], + "angle": 0, + "content": "Synthetic anomalies. As observed in other studies [5, 36], AEs tend to generalize too well to out-of-distribution data. This behavior is not desired in anomaly detection, since methods based on AEs rely on having high reconstruction errors for abnormal examples and low reconstruction errors for normal ones. To this end, we propose to augment the training videos with abnormal events. Since collecting abnormal training examples from the real-world is not possible, we resort to adding synthetic (virtual) anomalies. We leverage the recently introduced UBnormal data set [1] and its accurate pixel-level annotations to crop out abnormal events and blend them in our training videos, while ensuring the temporal consistency of the added events. The resulting examples, some depicted in Figure 3, are used to augment the training set with extra data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.652, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The synthetic examples help our model in three ways. First, in the reconstruction loss, we consider the original training frames (without superimposed anomalies) as the ground-truth, essentially forcing our model to overlook the anomalies. Formally, in Eq. (3), we use the patches \\(\\left\\{\\pmb{p}_i^{(t)}\\right\\}_{i = 1}^n\\) from the normal version of frame \\(\\pmb{x}_t\\). Second, we add the anomaly map as an additional channel to our target image. In the anomaly map, we set normal pixels to 0 and abnormal pixels to 1. This change implies that, in Eq. (3) and Eq. (4), all patches will have an additional channel. Third, we use the ground-truth anomaly map to enhance the weights defined in Eq. (2). The added synthetic anomalies do not necessarily yield motion gradients with high magnitude. Hence, it is possible to have low weights in Eq. (3) and Eq. (4) for patches that correspond to anomaly regions. This is not desirable if we want the model to detect anom" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15988" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.184 + ], + "angle": 0, + "content": "lies. To this end, we propose to add the anomaly maps and the gradients together, before computing the weights as in Eq. (2). Formally, in Eq. (1), we replace \\(\\left\\{\\pmb{r}_i^{(t)}\\right\\}_{i=1}^n\\) with \\(\\left\\{\\pmb{r}_i^{(t)} + \\pmb{a}_i^{(t)}\\right\\}_{i=1}^n\\), where \\(\\left\\{\\pmb{a}_i^{(t)}\\right\\}_{i=1}^n\\) is the set of the patches extracted from the anomaly map." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.185, + 0.468, + 0.261 + ], + "angle": 0, + "content": "Classification head. We further harness the synthetic anomalies to train a classification head applied on the final [CLS] token of the shared encoder. The head is trained to discriminate between frames with and without synthetic anomalies. This head is trained using binary cross-entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.265, + 0.468, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} \\left(\\hat {\\boldsymbol {x}} _ {t}, \\theta_ {E}\\right) = - y _ {t} \\cdot \\log \\left(\\hat {y} _ {t}\\right) - \\left(1 - y _ {t}\\right) \\cdot \\log \\left(1 - \\hat {y} _ {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.287, + 0.468, + 0.331 + ], + "angle": 0, + "content": "where \\(y_{t} \\in \\{0,1\\}\\) is 1 if the frame contains an anomaly and 0 otherwise, \\(\\hat{y}_t\\) is the prediction, and \\(\\theta_E\\) represents the set of weights of the shared encoder." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.468, + 0.392 + ], + "angle": 0, + "content": "Inference. During inference, we pass each frame \\(\\pmb{x}_t\\) through both teacher and student models to obtain the reconstructed frames \\(\\hat{\\pmb{x}}_t\\) and \\(\\tilde{\\pmb{x}}_t\\), respectively. Then, we compute output pixel-level anomaly map as:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.395, + 0.468, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {o} _ {t} = \\alpha \\cdot \\| \\boldsymbol {x} _ {t} - \\hat {\\boldsymbol {x}} _ {t} \\| _ {2} ^ {2} + \\beta \\cdot \\| \\hat {\\boldsymbol {x}} _ {t} - \\tilde {\\boldsymbol {x}} _ {t} \\| _ {2} ^ {2} + \\gamma \\cdot \\hat {y} _ {t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.468, + 0.519 + ], + "angle": 0, + "content": "where \\(\\alpha\\), \\(\\beta\\) and \\(\\gamma\\) are hyperparameters that control the importance of the individual anomaly score components. Following [15, 35], we apply spatio-temporal 3D filtering to smooth the anomaly volumes. To obtain the frame-level anomaly scores, we keep the maximum value from each output map \\(\\pmb{o}_t\\) and subsequently apply another temporal Gaussian filter to smooth the values." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.529, + 0.209, + 0.545 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.267, + 0.566 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.871 + ], + "angle": 0, + "content": "Data sets. We verify the performance of our method on four data sets for video anomaly detection: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. ShanghaiTech is the largest data set, with 270K frames for training and about 50K for testing. UBnormal is the second largest, with about 116K training frames and 93K testing frames. Avenue is a popular benchmark containing 15K frames for training and another 15K for testing. UCSD Ped2 holds a total of 4.5K frames, out of which 2.5K are used for training. UBnormal is a benchmark that uses an open set evaluation, where training and test anomalies belong to disjoint category sets. For the other three data sets, the training videos contain only normal events, and the test ones include both normal and abnormal scenarios. To augment the normal training videos, we sample abnormal events from the UBnormal data set [1]. UBnormal is a synthetic (virtual) data set containing anomalies simulated by video game characters, which alleviates the burden of collecting anomalies from the real world. The probability of augmenting a frame from Avenue, ShanghaiTech or UCSD Ped2 is 0.25." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Evaluation. We evaluate all models following recent related works [1, 27, 69], considering both micro and macro" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.871, + 0.691 + ], + "angle": 0, + "content": "
TypeMethodAvenueShanghaiUBnormalPed2FPS
MicroMacroMicroMacroMicroMacroMicroMacro
Object-centric[10]91.692.583.890.562.186.5--20
[18]86.4-71.6---97.8--
[26]91.592.882.490.255.484.597.599.851
[27]92.390.482.789.361.385.698.799.724
[33]--85.9-71.8----
[36]87.490.478.784.9--94.397.8-
[40]80.2-73.7-50.7----
[49]89.993.574.283.2--99.3-12
[50]91.892.383.887.8-----
[54]+[10]91.692.483.690.6----20
[54]+[27]93.291.883.389.3----31
[54]+[49]89.593.675.283.8----10
[69]+[27]92.991.983.689.5----31
[69]+[49]90.992.275.583.7----10
[84]92.2-84.3---99.0-35
[95]89.6-74.8---97.3--
Frame or cube level[4]87.1-75.9---96.5--
[5]84.7-73.7---98.4--
[9]----68.580.3--37
[28]83.3-71.2---94.1-35
[37]88.9--------
[43]90.0-----96.6--
[47]85.181.772.880.6--95.4-28
[48]84.4-----87.5--
[54]+[47]89.184.874.683.3----26
[54]+[61]86.486.370.680.3----94
[57]86.9-----96.2--
[60]85.3-72.2---96.3-195
[61]82.886.868.379.7--97.0-101
[62]72.0-----88.3--
[63]87.2-----93.0--
[66]------93.5--
[67]------88.4--
[69]+[47]87.384.574.582.9----26
[69]+[61]84.888.669.880.2----95
[76]84.6--------
[77]---76.550.376.8--56
[78]89.6-74.7------
[80]85.1-73.0---96.3--
[82]--76.1------
[87]87.0-79.3------
[89]--80.4------
[90]86.6-----96.9--
[93]90.1-78.6-62.7----
[96]90.2-----97.3--
[100]--78.9------
[103]------91.0--
Ours91.390.979.184.758.581.495.498.41655
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.695, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Table 1. Micro and macro AUC scores (in %) of several state-of-the-art frame-level, cube-level and object-level methods versus our self-distilled masked AE on Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The top three scores for each category of methods are shown in red, green, and blue. All reported running times (including those of the baselines) are measured on a machine with an Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "AUC metrics. The area under the ROC curve (AUC) expresses the overlap between the ground-truth frame-level annotations and the anomaly scores predicted by a model, at multiple thresholds. At a given threshold, a frame is labeled as abnormal if the predicted anomaly score is above the threshold. For the micro AUC, the test frames from all" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15989" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.452, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.268, + 0.47, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Predictions for test video 04 from Avenue. The abnormal bounding boxes are given by the convex hull of the patches labeled as abnormal. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.329, + 0.468, + 0.388 + ], + "angle": 0, + "content": "videos are concatenated before computing the AUC over all frames. For the macro AUC, the AUC of each test video is first computed, and the resulting AUC scores are averaged to obtain a single value." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.39, + 0.468, + 0.614 + ], + "angle": 0, + "content": "Hyperparameters. The encoder module is formed of three CvT blocks, each with a projection size of 256 and four attention heads. The teacher decoder contains three CvT blocks, while the student decoder contains only one block. All decoder blocks have four attention heads and a projection dimension of 128. Since the data sets have different input resolutions and objects vary in size, we adapt the patch size to each data set. Thus, we set the patch size to \\(16 \\times 16\\) on Avenue, \\(8 \\times 8\\) on ShanghaiTech and UBnormal, and \\(4 \\times 4\\) on UCSD Ped2. Regardless of the data set, the teacher network is trained for 100 epochs, while the student is trained for 40 epochs. We optimize the networks with Adam [42], using a learning rate of \\(10^{-4}\\) and mini-batches of 100 samples. The hyperparameters in Eq. (6) are set to \\(\\alpha = 0.4\\), \\(\\beta = 0.3\\) and \\(\\gamma = 0.3\\), for all data sets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.623, + 0.172, + 0.637 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.643, + 0.458, + 0.658 + ], + "angle": 0, + "content": "We present our results in Table 1 and discuss them below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Results on Avenue. Our method obtains a micro AUC score of \\(91.3\\%\\) on Avenue, being only \\(1.9\\%\\) below the state-of-the-art object-centric method. Remarkably, in the category of frame-level methods, we reach the best micro and macro AUC scores. Taking into account that our method is much faster than all other methods, we consider that its performance is remarkable. In Figure 4, we illustrate the anomaly scores for test video 04 from Avenue. Here, our model is close to perfect, highlighting its ability to capture anomalies, such as people running." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Results on ShanghaiTech. On ShanghaiTech, our method reaches the top macro AUC score and the third-best micro AUC score, when compared with other frame-level frameworks. Object-centric methods generally surpass frame-level methods, but the former methods have much lower processing speeds (see Figure 2)." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.2 + ], + "angle": 0, + "content": "
Motion weightsSelf-distillationSynthetic dataAnomaly mapsClassif. headAvenueShanghai
AUC
MicroMacroMicroMacro
84.085.669.780.1
✓84.886.371.380.9
✓✓88.586.076.383.8
✓✓✓88.586.977.083.0
✓✓✓✓90.589.677.384.2
✓✓✓✓✓91.390.979.184.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.205, + 0.892, + 0.233 + ], + "angle": 0, + "content": "Table 2. Impact of each novel component on the micro and macro AUC scores (in %), on Avenue [51] and ShanghaiTech [53]." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.333 + ], + "angle": 0, + "content": "
StrategyAvenueShanghai
AUC
MicroMacroMicroMacro
Teacher84.085.674.882.3
Teacher + Student85.485.875.182.1
Teacher + Teacher-Student Difference88.586.076.383.8
Teacher + Student + Teacher-Student Difference86.985.875.883.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.337, + 0.892, + 0.391 + ], + "angle": 0, + "content": "Table 3. Impact of strategies to combine the outputs of the teacher and student models on Avenue [51] and ShanghaiTech [53]. These results do not include the synthetic anomalies and the classification head." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.406, + 0.866, + 0.481 + ], + "angle": 0, + "content": "
Data setMeasurePercentage of synthetic data
0%25%50%75%
AvenueMicro AUC88.591.390.689.9
Macro AUC86.090.989.487.7
ShanghaiMicro AUC76.379.177.977.7
Macro AUC83.884.784.484.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.485, + 0.891, + 0.513 + ], + "angle": 0, + "content": "Table 4. Impact of varying the proportion of synthetic anomalies on the Avenue [51] and ShanghaiTech [53] data sets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Results on UBnormal. In terms of the micro AUC, the best frame-level method on UBnormal is TimeSformer [9], which benefits from large-scale pretraining. Notably, our method obtains a higher macro AUC than TimeSformer. The micro AUC of our method is fairly close to the micro AUC levels of the better object-centric approaches. In terms of speed, our method is significantly faster than all the other methods reporting results on UBnormal." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Results on UCSD Ped2. Our framework obtains a micro AUC of \\(95.4\\%\\), being \\(3.9\\%\\) below the state-of-the-art performance of Liu et al. [49]. While being slightly below in terms of micro AUC, our macro AUC score is remarkably on par with the object-centric method of Ionescu et al. [36], and only \\(1.4\\%\\) below that of Georgescu et al. [26]. Considering that our FPS is more than 30 times higher compared with these methods [26, 36] on the same GPU, our framework provides a clearly superior accuracy-speed trade-off." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation study. In Table 2, we illustrate the impact of each novel component on our model's performance. The first model is a vanilla masked AE, which obtains a rather low performance on both Avenue and ShanghaiTech. Each and every component contributes towards boosting the performance of the vanilla model. We observe that self-distillation gives the highest boost in terms of the micro AUC. How-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15990" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.087, + 0.796, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.309, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Figure 5. Examples of frames and anomaly maps reconstructed by our teacher. The first four columns correspond to abnormal examples from the Avenue and ShanghaiTech data sets, while the last column shows a normal example. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.471, + 0.414 + ], + "angle": 0, + "content": "ever, to surpass the \\(90\\%\\) milestone on Avenue, it is mandatory to introduce the prediction of anomaly maps in the learning task. An additional boost is given by our classification head." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.506 + ], + "angle": 0, + "content": "The self-distillation procedure gives us a few possible strategies to combine the outputs of the teacher and the student. Hence, we investigate this aspect and report the results in Table 3. We observe that the best micro AUC is obtained when we combine the teacher reconstruction error with the teacher-student discrepancy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.599 + ], + "angle": 0, + "content": "The proportion of synthetic examples per mini-batch is another aspect that can influence our model's performance. In Table 4, we report the micro and macro AUC scores for three possible augmentation levels. We observe that augmentation is always useful, but, for a better outcome, it requires a moderate percentage (25%)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.854 + ], + "angle": 0, + "content": "Performance-speed trade-off. In Figure 2, we compare our model with several other methods in terms of the performance-speed trade-off. This comparison undoubtedly shows that our method reaches a far better processing speed, while achieving fairly good performance. To strengthen this observation, we also compare the methods in terms of GFLOPs and number of parameters in Table 5. We underline that the method of Gong et al. [28] might seem small in terms of the number of parameters, but it is slowed down by its input, which is formed of a cuboid constructed by stacking 16 consecutive frames. Moreover, the method relies on a memory module, where each memory slot records the features of one pixel in the activation maps. Although their method has twice as many parameters as our own, the large input volume and the sizable memory bank reduce the speed to 35 FPS. With an FPS of 1655, our method proves to be significantly lighter than all its competitors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Qualitative results. In Figure 5, we present the frames and anomaly maps reconstructed by the teacher in four abnormal scenarios from Avenue and ShanghaiTech. Moreover," + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.351, + 0.848, + 0.438 + ], + "angle": 0, + "content": "
MethodGFLOPs ↓#Params (M) ↓FPS ↑
Georgescu et al. [26]107.96551
Georgescu et al. [27]121.66724
Liu et al. [49]179.532012
Park et al. [60]8464195
Gong et al. [28]55.2635
Ours0.831655
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.442, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Table 5. Comparing methods in terms of floating point operations (GFLOPs), number of parameters, and FPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.627 + ], + "angle": 0, + "content": "in the fifth column, we illustrate the behavior of the teacher in a normal scenario. In all four abnormal cases, the reconstruction error is visibly higher for anomalous regions. This effect is mostly due to our training procedure based on synthetic data augmentation. The most obvious example is in the fourth column, where the bicycle seen in a pedestrian area is almost entirely removed from the output. Additionally, the predicted anomaly maps are well aligned with the ground-truth ones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.62, + 0.652 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In this work, we proposed a lightweight masked autoencoder (3M parameters, 0.8 GFLOPs) for video anomaly detection, which learns to reconstruct tokens with high motion gradients. Our framework is based on self-distillation, leveraging the discrepancy between teacher and student decoders for anomaly detection. Moreover, we boost the performance of our model by introducing a data augmentation technique based on overlapping synthetic anomalies on normal training data. Our highly efficient framework reached an unprecedented speed of 1655 FPS, with a minimal performance gap with respect to the state-of-the-art object-centric approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. Work supported by a grant of the Romanian Ministry of Education and Research, CNCS-UEFISCDI, project number PN-III-P2-2.1-PED-2021-0195, within PNCDI III." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15991" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1]Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. UBnormal: New Benchmark for Supervised Open-Set Video Anomaly Detection. In Proceedings of CVPR, pages 20143-20153, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.2, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust Real-Time Unusual Event Detection Using Multiple Fixed-Location Monitors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):555-560, 2008. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.271, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[3] Borislav Antic and Bjorn Ommer. Video parsing for abnormality detection. In Proceedings of ICCV, pages 2415-2422, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.314, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. In Proceedings of BMVC, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.359, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[5] Marcella Astrid, Muhammad Zaigham Zaheer, and Seung-Ik Lee. Synthetic Temporal Anomaly Guided End-to-End Video Anomaly Detection. In Proceedings of ICCVW, pages 207-214, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.415, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[6] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In Proceedings of NIPS, pages 2654-2662, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.459, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[7] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. MultiMAE: Multi-modal multi-task masked autoencoders. In Proceedings of ECCV, pages 348-367. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.516, + 0.471, + 0.584 + ], + "angle": 0, + "content": "[8] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed Students: Student-Teacher Anomaly Detection With Discriminative Latent Embeddings. In Proceedings of CVPR, pages 4183–4192, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.587, + 0.471, + 0.628 + ], + "angle": 0, + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In Proceedings of ICML, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.727 + ], + "angle": 0, + "content": "[10] Antonio Bărbălău, Radu Tudor Ionescu, Mariana-Iuliana Georgescu, Jacob Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. SSMTL++: Revisiting Self-Supervised Multi-Task Learning for Video Anomaly Detection. Computer Vision and Image Understanding, 229: 103656, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.729, + 0.471, + 0.784 + ], + "angle": 0, + "content": "[11] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. SdAE: Self-distillated Masked Autoencoder. In Proceedings of ECCV, pages 108-124, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.786, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[12] Hekai Cheng, Lu Yang, and Zulong Liu. Relation-based knowledge distillation for anomaly detection. In Proceedings of PRCV, pages 105-116, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.83, + 0.47, + 0.885 + ], + "angle": 0, + "content": "[13] Kai-Wen Cheng, Yie-Tarng Chen, and Wen-Hsien Fang. Video anomaly detection and localization using hierarchical feature representation and Gaussian process regression. In Proceedings of CVPR, pages 2909-2917, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.887, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[14] Y. Cong, J. Yuan, and J. Liu. Sparse reconstruction cost for" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "abnormal event detection. In Proceedings of CVPR, pages 3449-3456, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Allison Del Giorno, J. Andrew Bagnell, and Martial Hebert. A Discriminative Framework for Anomaly Detection in Large Videos. In Proceedings of ECCV, pages 334-349, 2016. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[16] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of CVPR, pages 9737-9746, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.222, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[17] Fei Dong, Yu Zhang, and Xiushan Nie. Dual Discriminator Generative Adversarial Network for Video Anomaly Detection. IEEE Access, 8:88170-88176, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.264, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[18] Keval Doshi and Yasin Yilmaz. Any-Shot Sequential Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 934-935, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.308, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[19] Keval Doshi and Yasin Yilmaz. Continual Learning for Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 254-255, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.351, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.435, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[21] Jayanta K. Dutta and Bonny Banerjee. Online Detection of Abnormal Events Using Incremental Coding Length. In Proceedings of AAAI, pages 3755-3761, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.478, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[22] Yaxiang Fan, Gongjian Wen, Deren Li, Shaohua Qiu, Martin D. Levine, and Fei Xiao. Video anomaly detection and localization via Gaussian Mixture Fully Convolutional Variational Autoencoder. Computer Vision and Image Understanding, 195:102920, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[23] Ye Fei, Chaoqin Huang, Cao Jinkun, Maosen Li, Ya Zhang, and Cewu Lu. Attribute Restoration Framework for Anomaly Detection. IEEE Transactions on Multimedia, 24: 116-127, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.605, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[24] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In Proceedings of NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.648, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[25] Yachuang Feng, Yuan Yuan, and Xiaoqiang Lu. Learning deep event models for crowd anomaly detection. Neurocomputing, 219:548-556, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.691, + 0.894, + 0.759 + ], + "angle": 0, + "content": "[26] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly Detection in Video via Self-Supervised and Multi-Task Learning. In Proceedings of CVPR, pages 12742–12752, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.761, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[27] Mariana Iuliana Georgescu, Radu Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A Background-Agnostic Framework with Adversarial Training for Abnormal Event Detection in Video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4505-4523, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[28] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton Van Den Hengel. Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15992" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Anomaly Detection. In Proceedings of ICCV, pages 1705-1714, 2019. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.178 + ], + "angle": 0, + "content": "[29] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In Proceedings of CVPR, pages 733-742, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Proceedings of CVPR, pages 16000-16009, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.235, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[31] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint Detection and Recounting of Abnormal Events by Learning Deep Generic Knowledge. In Proceedings of ICCV, pages 3639-3647, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.292, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[32] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the Knowledge in a Neural Network. In Proceedings of NIPS Deep Learning and Representation Learning Workshop, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.349, + 0.472, + 0.39 + ], + "angle": 0, + "content": "[33] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In Proceedings of ICCV, pages 13545-13554, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.392, + 0.472, + 0.447 + ], + "angle": 0, + "content": "[34] Chaoqin Huang, Qinwei Xu, Yanfeng Wang, Yu Wang, and Ya Zhang. Self-Supervised Masking for Unsupervised Anomaly Detection and Localization. IEEE Transactions on Multimedia, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.449, + 0.472, + 0.503 + ], + "angle": 0, + "content": "[35] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In Proceedings of ICCV, pages 2895-2903, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.505, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[36] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-Centric AutoEncoders and Dummy Anomalies for Abnormal Event Detection in Video. In Proceedings of CVPR, pages 7842-7851, 2019. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.576, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[37] Radu Tudor Ionescu, Sorina Smeureanu, Marius Popescu, and Bogdan Alexe. Detecting abnormal events in video using Narrowed Normality Clusters. In Proceedings of WACV, pages 1951-1960, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.633, + 0.472, + 0.688 + ], + "angle": 0, + "content": "[38] Xiangli Ji, Bairong Li, and Yuesheng Zhu. TAM-Net: Temporal Enhanced Appearance-to-Motion Generative Network for Video Anomaly Detection. In Proceedings of IJCNN, pages 1–8, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[39] Jielin Jiang, Jiale Zhu, Muhammad Bilal, Yan Cui, Neeraj Kumar, Ruihan Dou, Feng Su, and Xiaolong Xu. Masked Swin Transformer Unet for Industrial Anomaly Detection. IEEE Transactions on Industrial Informatics, 19(2):2200-2209, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[40] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. BiPOCO: Bi-Directional Trajectory Prediction with Pose Constraints for Pedestrian Anomaly Detection. In Proceedings of SL4AD, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[41] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: A space-time MRF for detecting abnormal activities with incremental updates. In Proceedings of CVPR, pages 2921-2928, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[42] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of ICLR, 2015. 7" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: Bidirectional Multi-Scale Aggregation Networks for Abnormal Event Detection. IEEE Transactions on Image Processing, 29:2395-2408, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[44] Guoqiu Li, Guanxiong Cai, Xingyu Zeng, and Rui Zhao. Scale-Aware Spatio-Temporal Relation Learning for Video Anomaly Detection. In Proceedings of ECCV, pages 333-350, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[45] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(1):18-32, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[46] Zhenyu Li, Ning Li, Kaitao Jiang, Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Superpixel Masking and Inpainting for Self-Supervised Anomaly Detection. In Proceedings of BMVC, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[47] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future Frame Prediction for Anomaly Detection - A New Baseline. In Proceedings of CVPR, pages 6536-6545, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[48] Yusha Liu, Chun-Liang Li, and Barnabaás Póczos. Classifier Two-Sample Test for Video Anomaly Detections. In Proceedings of BMVC, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.414, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[49] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction. In Proceedings of ICCV, pages 13588-13597, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[50] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating Anomalies for Video Anomaly Detection With Prompt-Based Feature Mapping. In Proceedings of CVPR, pages 24500-24510, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.54, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[51] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal Event Detection at 150 FPS in MATLAB. In Proceedings of ICCV, pages 2720-2727, 2013. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[52] Yiwei Lu, Frank Yu, Mahesh Kumar, Krishna Reddy, and Yang Wang. Few-Shot Scene-Adaptive Anomaly Detection. In Proceedings of ECCV, pages 125–141, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[53] Weixin Luo, Wen Liu, and Shenghua Gao. A Revisit of Sparse Coding Based Anomaly Detection in Stacked RNN Framework. In Proceedings of ICCV, pages 341-349, 2017. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.679, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[54] Neelu Madan, Nicolae-Catalin Ristea, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised masked convolutional transformer block for anomaly detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):525-542, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.762, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[55] Vijay Mahadevan, Wei-Xin LI, Viral Bhalodia, and Nuno Vasconcelos. Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 1975-1981, 2010. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[56] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of CVPR, pages 935-942, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[57] Trong-Nguyen Nguyen and Jean Meunier. Anomaly Detection in Video Sequence With Appearance-Motion Correspondence. In Proceedings of ICCV, pages 1273-1283, 2019. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15993" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[58] Guansong Pang, Cheng Yan, Chunhua Shen, Anton van den Hengel, and Xiao Bai. Self-trained Deep Ordinal Regression for End-to-End Video Anomaly Detection. In Proceedings of CVPR, pages 12173–12182, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.15, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[59] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM Computing Surveys, 54(2):1-38, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.193, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[60] Chaewon Park, MyeongAh Cho, Minhyeok Lee, and Sangyoun Lee. FastAno: Fast anomaly detection via spatiotemporal patch transformation. In Proceedings of WACV, pages 2249-2259, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.25, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[61] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection. In Proceedings of CVPR, pages 14372-14381, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.471, + 0.362 + ], + "angle": 0, + "content": "[62] Bharathkumar Ramachandra and Michael Jones. Street Scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of WACV, pages 2569-2578, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.364, + 0.472, + 0.419 + ], + "angle": 0, + "content": "[63] Bharathkumar Ramachandra, Michael Jones, and Ranga Vatsavai. Learning a distance function with a Siamese network to localize anomalies in videos. In Proceedings of WACV, pages 2598-2607, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.421, + 0.471, + 0.475 + ], + "angle": 0, + "content": "[64] Bharathkumar Ramachandra, Michael Jones, and Ranga Raju Vatsavai. Perceptual metric learning for video anomaly detection. Machine Vision and Applications, 32:1432-1769, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.478, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[65] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. A Survey of Single-Scene Video Anomaly Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2293-2312, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.549, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[66] Mahdyar Ravanbakhsh, Moin Nabi, Enver Sangineto, Lucio Marcenaro, Carlo Regazzoni, and Nicu Sebe. Abnormal Event Detection in Videos using Generative Adversarial Nets. In Proceedings of ICIP, pages 1577-1581, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[67] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-Play CNN for Crowd Motion Analysis: An Application in Abnormal Event Detection. In Proceedings of WACV, pages 1689–1698, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.69, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[68] Huamin Ren, Weifeng Liu, Soren Ingvor Olsen, Sergio Escalera, and Thomas B. Moeslund. Unsupervised Behavior-Specific Dictionary Learning for Abnormal Event Detection. In Proceedings of BMVC, pages 28.1-28.13, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[69] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection. In Proceedings of CVPR, pages 13576-13586, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[70] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, Karpathy A., A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 4" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[71] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, and Reinhard Klette. Deep-Cascade: Cascading 3D Deep Neural Networks for Fast Anomaly Detection and Localization in Crowded Scenes. IEEE Transactions on Image Processing, 26(4):1992-2004, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[72] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, Zahra Moayed, and Reinhard Klette. Deep-Anomaly: Fully Convolutional Neural Network for Fast Anomaly Detection in Crowded Scenes. Computer Vision and Image Understanding, 172:88-97, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.234, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[73] Babak Saleh, Ali Farhadi, and Ahmed Elgammal. Object-Centric Anomaly Detection by Attribute-Based Reasoning. In Proceedings of CVPR, pages 787-794, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.278, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[74] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution Knowledge Distillation for Anomaly Detection. In Proceedings of CVPR, pages 14902-14912, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.348, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[75] Venkatesh Saligrama and Zhu Chen. Video anomaly detection based on local statistical aggregates. In Proceedings of CVPR, pages 2112-2119, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[76] Sorina Smeureanu, Radu Tudor Ionescu, Marius Popescu, and Bogdan Alexe. Deep Appearance Features for Abnormal Behavior Detection in Video. In Proceedings of ICIAP, pages 779-789, 2017. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[77] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In Proceedings of CVPR, pages 6479–6488, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[78] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-Aware Context Reasoning for Unsupervised Abnormal Event Detection in Videos. In Proceedings of ACMMM, pages 184-192, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.547, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[79] Qianru Sun, Hong Liu, and Tatsuya Harada. Online growing neural gas for anomaly detection in changing surveillance scenes. Pattern Recognition, 64(C):187-201, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[80] Yao Tang, Lin Zhao, Shanshan Zhang, Chen Gong, Guangyu Li, and Jian Yang. Integrating prediction and reconstruction for anomaly detection. Pattern Recognition Letters, 129:123-130, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[81] Hanh T.M. Tran and David Hogg. Anomaly Detection using a Convolutional Winner-Take-All Autoencoder. In Proceedings of BMVC, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[82] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised video anomaly detection. In Proceedings of ICIP, pages 2540-2544, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.761, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[83] Hung Vu, Tu Dinh Nguyen, Trung Le, Wei Luo, and Dinh Phung. Robust Anomaly Detection in Videos Using Multilevel Representations. In Proceedings of AAAI, pages 5216-5223, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[84] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Proceedings of ECCV, pages 494-511, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[85] L. Wang, F. Zhou, Z. Li, W. Zuo, and H. Tan. Abnormal Event Detection in Videos Using Hybrid Spatio-Temporal" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15994" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Autoencoder. In Proceedings of ICIP, pages 2276-2280, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[86] Xusheng Wang, Mingtao Pei, and Zhengang Nie. Self-trained video anomaly detection based on teacher-student model. In Proceedings of MLSP, pages 1-6, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.165, + 0.471, + 0.207 + ], + "angle": 0, + "content": "[87] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster Attention Contrast for Video Anomaly Detection. In Proceedings of ACMMM, pages 2463-2471, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.208, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[88] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvT: Introducing Convolutions to Vision Transformers. In Proceedings of ICCV, pages 22-31, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.264, + 0.471, + 0.32 + ], + "angle": 0, + "content": "[89] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In Proceedings of ECCV, pages 729-745, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.321, + 0.471, + 0.377 + ], + "angle": 0, + "content": "[90] Peng Wu, Jing Liu, and Fang Shen. A Deep One-Class Neural Network for Anomalous Event Detection in Complex Scenes. IEEE Transactions on Neural Networks and Learning Systems, 31(7):2609-2622, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.378, + 0.471, + 0.434 + ], + "angle": 0, + "content": "[91] Shandong Wu, Brian E. Moore, and Mubarak Shah. Chaotic Invariants of Lagrangian Particle Trajectories for Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 2054-2060, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.435, + 0.471, + 0.49 + ], + "angle": 0, + "content": "[92] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting Anomalous Events in Videos by Learning Deep Representations of Appearance and Motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.491, + 0.471, + 0.546 + ], + "angle": 0, + "content": "[93] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature Prediction Diffusion Model for Video Anomaly Detection. In Proceedings of ICCV, pages 5527-5537, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.548, + 0.471, + 0.617 + ], + "angle": 0, + "content": "[94] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.618, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[95] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze Test Helps: Effective Video Anomaly Detection via Learning to Complete Video Events. In Proceedings of ACMMM, pages 583-591, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[96] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Transactions on Neural Networks and Learning Systems, 33(8): 3572-3586, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[97] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proceedings of CVPR, pages 19313–19322, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[98] Muhammad Zaigham Zaheer, Jin-ha Lee, Marcella Astrid, and Seung-Ik Lee. Old is Gold: Redefining the Adversarily Learned One-Class Classifier Training Paradigm. In Proceedings of CVPR, pages 14183-14193, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[99] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. CLAWS: Clustering Assisted" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Weakly Supervised Learning with Normalcy Suppression for Anomalous Event Detection. In Proceedings of ECCV, pages 358-376, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[100] Muhammad Zaigham Zaheer, Arif Mahmood, Haris M. Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative Cooperative Learning for Unsupervised Video Anomaly Detection. In Proceedings of CVPR, pages 14744-14754, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[101] Linfeng Zhang, Chenglong Bao, and Kaisheng Ma. Self-Distillation: Towards Efficient and Compact Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4388-4403, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[102] Xinfeng Zhang, Su Yang, Jiulong Zhang, and Weishan Zhang. Video Anomaly Detection and Localization using Motion-field Shape Description and Homogeneity Testing. Pattern Recognition, page 107394, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[103] Ying Zhang, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Shun Sakai. Video anomaly detection based on locality sensitive hashing filters. Pattern Recognition, 59:302-311, 2016. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[104] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online Detection of Unusual Events in Videos via Dynamic Sparse Coding. In Proceedings of CVPR, pages 3313-3320, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[105] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph Convolutional Label Noise Cleaner: Train a Plug-And-Play Action Classifier for Anomaly Detection. In Proceedings of CVPR, pages 1237–1246, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15995" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_origin.pdf b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9581052bd608e664b8a90e5fa4622814c1b9f274 --- /dev/null +++ b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/b79b2889-9304-49a2-9631-466d8e496991_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfe25ab9a57d5406af03294b24ebc0028e59f3b3572286ddefa4305f602de4bc +size 9079306 diff --git a/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/full.md b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75ffbda64cfb005dcf731afa69db088734ecdeba --- /dev/null +++ b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/full.md @@ -0,0 +1,331 @@ +# Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors + +Nicolae-Cătălin Ristea $^{1,2,\diamond}$ , Florinel-Alin Croitoru $^{1,\diamond}$ , Radu Tudor Ionescu $^{1,3,*}$ , Marius Popescu $^{1,3}$ , Fahad Shahbaz Khan $^{4,5}$ , Mubarak Shah $^{6}$ + +1University of Bucharest, Romania, 2NUST Politehnica Bucharest, Romania, 3SecurifAI, Romania, 4MBZ University of Artificial Intelligence, UAE, 5Linköping University, Sweden, 6University of Central Florida, US + +# Abstract + +We propose an efficient abnormal event detection model based on a lightweight masked auto-encoder (AE) applied at the video frame level. The novelty of the proposed model is threefold. First, we introduce an approach to weight tokens based on motion gradients, thus shifting the focus from the static background scene to the foreground objects. Second, we integrate a teacher decoder and a student decoder into our architecture, leveraging the discrepancy between the outputs given by the two decoders to improve anomaly detection. Third, we generate synthetic abnormal events to augment the training videos, and task the masked AE model to jointly reconstruct the original frames (without anomalies) and the corresponding pixel-level anomaly maps. Our design leads to an efficient and effective model, as demonstrated by the extensive experiments carried out on four benchmarks: Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The empirical results show that our model achieves an excellent trade-off between speed and accuracy, obtaining competitive AUC scores, while processing 1655 FPS. Hence, our model is between 8 and 70 times faster than competing methods. We also conduct an ablation study to justify our design. Our code is freely available at: https://github.com/ristea/aed-mae. + +# 1. Introduction + +In recent years, research on abnormal event detection in video gained significant traction [1, 10, 17, 18, 26-28, 36, 38, 43, 44, 49, 52, 57, 58, 61, 62, 65, 69, 76, 78, 80, 83, 87, 90, 95, 97-100], due to its utter importance in video surveillance. Despite the growing interest, video anomaly detection remains a complex task, owing its complexity to the fact that abnormal situations are context-dependent and do not occur very often. This makes it very difficult to collect a representative set of abnormal events for training state-of-the-art deep learning models in a fully supervised manner. To showcase the rarity and reliance on context of anomalies, we refer to the vehicle ramming attacks carried out by terrorists against pedestrians. As soon as a car is steered + +![](images/1873e97897d64d12c82b39939d5b3a6ded4ec79e35d3f775ddeb74fa8f4e1379.jpg) +Figure 1. Our masked auto-encoder for abnormal event detection based on self-distillation. At training time, some video frames are augmented with synthetic anomalies. The teacher decoder learns to reconstruct original frames (without anomalies) and predict anomaly maps. The student decoder learns to reproduce the teacher's output. Motion gradients are aggregated at the token level and used as weights for the reconstruction loss. Red dashed lines represent steps executed only during training. + +on the sidewalk, it becomes an abnormal event. Hence, the place where the car is driven (street versus sidewalk) determines the normal or abnormal label of the action, i.e. the label depends on context. Furthermore, there are less than 200 vehicle ramming attacks registered to date1, confirming + +![](images/9ac81caecf002633eda4ec46e306012c427e194757851963aa694c74f90abdbf.jpg) +Figure 2. Performance versus speed trade-offs for our self-distilled masked AE and several state-of-the-art methods [26-28, 47, 49, 60, 61, 69, 84] (with open-sourced code), on the Avenue data set. The running times of all methods are measured on a computer with one Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM. Best viewed in color. + +the scarcity of such events (even less are caught on video). + +Since training anomaly detectors under a fully supervised setting is not possible, most studies dealing with abnormal event detection took a distinct path, proposing variations of outlier detection methods [3, 13, 14, 17, 21, 22, 29, 37, 41, 43, 45, 47, 51, 53, 55, 56, 61-68, 71, 80, 90, 92, 102-105]. Such methods treat abnormal event detection as an outlier detection task, where a normality model trained on normal events is applied on both normal and abnormal events during inference, labeling events deviating from the learned model as abnormal. Different from the mainstream path based on outlier detection, we propose an approach to augment each training video scene with synthetic anomalies, by randomly superimposing temporal action segments from the synthetic UBnormal data set [1] on our real-world data sets. We thus introduce synthetic anomalies at training time, enabling our model to learn in an open-set supervised manner. Additionally, we force our model to reconstruct the original training frames (without anomalies) to limit its ability to reconstruct anomalies, hence generating higher errors when anomalies occur. + +A large body of work on video anomaly detection has focused on employing auto-encoders (AEs) to address the task [5, 22, 27-29, 36, 49, 81, 85], relying on the poor reconstruction capabilities of these models on out-of-distribution data. Since training is carried out only on normal examples, it is expected for AEs to exhibit high reconstruction errors when anomalies occur. However, several researchers + +observed that AEs generalize too well [5, 36], being able to reconstruct anomalies with very high precision. Thus, to better leverage the reconstruction error of AEs in anomaly detection, researchers explored a few alternatives, from the use of dummy [36] or pseudo-anomalies [5, 27] to the integration of memory modules [28, 49, 61]. With the same purpose in mind, we propose to employ masked auto-encoders [30] in anomaly detection, introducing new ways to regulate their generalization capacity. Indeed, we go beyond employing the standard masked AE framework, and propose three novel changes to enhance the anomaly detection performance of our model. First, we propose to weight tokens based on the magnitude of motion gradients, raising the importance of tokens with higher motion in the reconstruction loss. This makes our model focus on reconstructing tokens with high motion, and avoid reconstructing the background scene, which is typically static for surveillance cameras. Second, we attach a classification head to discriminate between normal and pseudo-abnormal instances in the latent encoding space. Third, we integrate a teacher decoder and a student decoder into our masked AE architecture, where the student decoder learns to distill knowledge from the already optimized teacher. To reduce our processing time, we use a shared encoder for the teacher and student models, leading to a process known as self-distillation [101]. During the self-distillation process, the shared encoder is frozen. We leverage the discrepancy between the outputs given by the teacher and student decoders along with the reconstruction error of the teacher to boost anomaly detection performance. Our entire framework, integrating these components into a meticulous design, is shown in Figure 1. + +State-of-the-art deep anomaly detectors [10, 26, 27, 84] typically rely on a costly object detection method to increase precision, limiting the processing bandwidth to one video stream per GPU, at around 20-30 FPS. However, for real-world video surveillance, e.g. monitoring an entire city with hundreds or thousands of cameras, the processing costs of object-centric video anomaly detectors are simply too high, given their power consumption and that one GPU can cost around $2,000. To this end, we turn our attention to developing a lightweight model (6 transformer blocks, 3M parameters), capable of processing around 66 video streams at 25 FPS, significantly reducing the processing costs. Different from competing models performing anomaly detection at the object [10, 18, 26, 27, 36, 84, 95] or spatio-temporal cube [15, 21, 35, 37, 41, 48, 51, 53, 55, 60, 71, 75, 103] levels, we present a model that takes whole video frames as input, which is significantly more efficient (see Figure 2). + +We carry out comprehensive experiments on four benchmarks: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. The empirical results show that our method is 8 to 70 times faster than competing methods [26-28, 47, 49, 60, 61, 69, 84], while achieving comparable ac + +curacy levels. Aside from the main results, we conduct an ablation study showing that our novel design choices are supported by empirical evidence. + +In summary, our contribution is threefold: + +- We propose a lightweight masked auto-encoder for anomaly detection in video, which learns to reconstruct tokens with higher motion magnitude. +- We introduce a self-distillation training pipeline, leveraging the discrepancy between teacher and student decoders to obtain a significant accuracy boost for our highly efficient model (due to the shared encoder). +- To further boost the performance of our model, we introduce a data augmentation approach based on superimposing synthetic anomalies on normal training videos, which enables the masked AE model to learn with open-set supervision. + +# 2. Related Work + +Anomaly detection in video is typically formulated as a one-class learning problem, where only normal data is available at training time. During test time, both normal and abnormal examples are present [59, 65]. There are several categories of anomaly detection approaches, including dictionary learning methods [13, 14, 21, 51, 68, 89], probabilistic models [2, 3, 25, 31, 41, 55, 56, 73, 91], change detection frameworks [15, 35, 48, 58], distance-based models [36, 37, 62, 63, 67, 71, 72, 75, 76, 79, 81] and reconstruction-based approaches [23, 28, 29, 46, 47, 53, 57, 61, 66, 69, 80]. Considering that reconstruction-based methods often reach state-of-the-art performance in anomaly detection [27, 69], a large body of works used the reconstruction-based paradigm in the past few years. To this end, we adopt this paradigm in our study. + +With respect to the level at which the anomaly detection is carried out, methods can be categorized into spatiotemporal cube-level methods [37, 47, 48, 51, 55, 56, 66, 67, 71, 75, 76, 96, 103], frame-level methods [47, 66, 67], and object-level methods [10, 18, 19, 26, 27, 36, 49, 84, 95]. + +Frame-level and cube-level methods. Before the deep learning era, preliminary abnormal event detection models commonly relied on taking short video sequences and dividing them into spatio-temporal cuboids [15, 21, 41, 51, 55, 75, 103]. The cubes are then considered as independent examples, being passed as input to a machine learning model. This mainstream practice continued during the deep learning period [29, 35, 37, 43, 44, 53, 60, 62, 63, 71, 76, 100], when deep networks have been used to extract features [35, 37, 48, 53] or learn [28, 29, 43, 44, 53, 60, 62-64, 71, 89, 100] from these spatio-temporal cubes. + +At the same time, some studies considered using entire video frames as input [47, 66, 67]. For example, Liu et al. [47] proposed an effective algorithm, which learns to reconstruct the next frame of a short video sequence. A + +more complex approach is proposed by Ravanbakhsh et al. [67], who employ optical-flow reconstruction to predict the anomalous regions from an input image. In a different study, Ravanbakhsh et al. [66] proposed to detect anomalies at the frame level via generative adversarial networks. + +Frame-level and cube-level methods have a common characteristic, namely their relatively high processing speed due to the reasonably fast preprocessing steps, as opposed to object-centric methods. Still, frame-level methods hold a stronger advantage in terms of time, since cube-level methods need to process each cube as an independent example. Indeed, it is more efficient to process a mini-batch of frames rather than several mini-batches of spatio-temporal cubes. However, cube-level methods often outperform frame-level methods. To this end, we propose a masked AE that takes whole frames as input, yet learns interactions between video patches, thus integrating the best of both worlds. + +To boost the performance of frame-level or cube-level methods, researchers explored the inclusion of various components, such as memory modules [28, 61] or masked convolutional blocks [69]. Although integrating additional modules into the framework leads to accuracy gains, the procedure often comes with efficiency drawbacks. In contrast, our goal is to achieve a superior trade-off between performance and speed, with a higher focus on efficiency. As such, we design a lightweight masked AE based on convolutional vision transformer (CvT) blocks [88], and propose several upgrades resulting in a minimal time overhead. For example, we employ knowledge distillation to leverage the discrepancy between the teacher and the student models. However, to keep the processing time to the bare minimum, we resort to self-distillation [101] and use a shared encoder for the teacher and student networks. + +Object-level methods. To reduce the number of false positive detections often observed for other methods, some recent studies [10, 26, 27, 36, 49, 84, 95] proposed to look for anomalous objects rather than anomalous frames or cubes. Object-centric methods use the prior information from an object detector, enabling the anomaly detector to focus only on objects. This kind of framework boosts the accuracy by significant margins, currently reaching state-of-the-art performance [10, 84]. However, a considerable drawback is that the inference speed of the whole framework is directly conditioned by the object detector's speed, which is often much lower than that of the anomaly detection network [26, 36]. Hence, the processing time is significantly limited. In contrast, we perform anomaly detection at the frame level, obtaining an inference speed that is between 32 to 70 times faster than object-centric models [10, 26, 27, 36, 49, 84, 95]. + +Masked auto-encoders in anomaly detection. He et al. [30] proposed masked auto-encoders as a pretraining method to obtain strong backbones for downstream tasks. + +Since then, the method has been adopted in various fields, e.g. video processing [24] or multimodal learning [7], with remarkable results. We elaborate the connection to seemingly related masked AEs in the supplementary [11, 94]. The masking framework has also been used for anomaly detection in medical [34] and industrial [39] images. To the best of our knowledge, we are the first to propose a masked transformer-based auto-encoder for video anomaly detection. Moreover, we go beyond applying standard masked AEs, proposing several modifications leading to superior performance levels: emphasizing tokens with higher motion, augmenting training videos with synthetic anomalies, and employing self-distillation. + +Knowledge distillation in anomaly detection. Knowledge distillation [6, 32] was originally designed to compress one or multiple large models (teachers) into a lighter neural network (student). Recently adopted in anomaly detection [8, 12, 16, 26, 74, 86], knowledge distillation was deemed useful due to the possibility of leveraging the representation discrepancy between the teacher and the student networks, which is larger in the case of anomalies. For example, Bergmann et al. [8] trained an ensemble of student networks on normal data to reproduce the output of a deep feature extractor, which is pretrained on ImageNet [70]. The authors use the difference between the teacher label and the mean over student labels to detect abnormal pixels. Salehi et al. [74] employed a more thorough distillation process, called hint learning, in which the multi-level features of a teacher pretrained on ImageNet are distilled into a clone. + +Most studies based on knowledge distillation applied the framework to image anomaly detection [8, 12, 16, 74]. With few exceptions [26, 86], knowledge distillation in video anomaly detection remains largely unexplored. Wang et al. [86] employed the teacher-student training paradigm to learn from unlabeled video samples in a self-supervised manner. Georgescu et al. [26] integrated knowledge distillation as a proxy task into a multi-task learning framework for video anomaly detection. + +Distinct from the aforementioned studies, to our knowledge, we are the first to introduce a variant of self-distillation in anomaly detection. Self-distillation [101] attaches multiple classification heads at various depths to boost the classification performance of a neural classifier. In contrast, we integrate self-distillation into a masked AE, employing two decoders of different depths. Due to the shared encoder, we are able to leverage the reconstruction discrepancy between the teacher and the student with a minimal computational overhead. + +# 3. Method + +Overview. We introduce a lightweight teacher-student transformer-based masked AE, which employs a two-stage training pipeline. In the first stage, we optimize a teacher + +masked AE via a reconstruction loss that employs a novel weighting mechanism based on motion gradients. In the second stage, we optimize the last (and only) decoder block of a student masked AE, which shares most of the backbone (kept frozen) with its teacher, to preserve efficiency. Next, we describe how to create training videos with synthetic anomalies and train the masked AEs to jointly predict the anomaly maps and overlook (not reconstruct) the anomalies from training frames. Lastly, we introduce a classification head to distinguish between frames with and without synthetic anomalies, which further boosts the performance of our method, with a marginal computational overhead. + +Architecture. Our masked AE pursues the architectural principles proposed in [30]. Hence, the entire architecture is formed of visual transformer blocks. In contrast to He et al. [30], we replace the ViT [20] blocks with CvT blocks [88], aiming for higher efficiency. Our processing starts by dividing the input images into non-overlapping tokens and removing a certain number of tokens. The encoder embeds the remaining tokens via convolutional projection layers, and the result is processed by transformer blocks. The decoder operates on a complete set of tokens, those removed being replaced with mask tokens. Its architecture is symmetric to that of the encoder. For efficiency reasons, we only use three blocks for the encoder and three blocks for the decoder. Each block is equipped with four attention heads. To achieve further speed gains, we replace all dense layers inside the CvT blocks with pointwise convolutions. We consider the architecture described so far as a teacher network. A student decoder branches out from the teacher after the first transformer block of the main decoder, adding only one extra transformer block (as shown in Figure 1). + +Motion gradient weighting. Masked AEs [30] have been originally applied on natural images. In this context, reconstructing randomly masked tokens is a viable solution, since images have high foreground and background variations. However, abnormal event detection data sets [2, 51, 53, 55] contain videos from fixed cameras with static backgrounds [65]. Learning to reconstruct the static background via masked AEs is both trivial and useless. Hence, naively training masked AEs to reconstruct randomly masked tokens in video anomaly detection is suboptimal. To this end, we propose to take into account the magnitude of the motion gradients when computing the reconstruction loss. + +Let $\pmb{x}_t\in \mathbb{R}^{h\times w\times c}$ be the video frame at index $t$ . Let $n$ be the number of non-overlapping visual tokens (patches) of size $d\times d\times c$ from each frame $\pmb{x}_t$ , where $c$ is the number of input channels, and $d$ is a hyperparameter that directly determines $n$ . Let $\left\{\pmb{p}_i^{(t)}\right\}_{i = 1}^n\in \mathbb{R}^{d\times d\times c}$ denote the set of tokens in frame $\pmb{x}_t$ , and $\left\{\hat{\pmb{p}}_i^{(t)}\right\}_{i = 1}^n\in \mathbb{R}^{d\times d\times c}$ the corresponding set of reconstructed tokens. + +Following Ionescu et al. [36], we estimate the motion + +gradient map $\pmb{g}_t$ of frame $\pmb{x}_t$ by computing the absolute difference between consecutive frames, which are previously filtered with a $3\times 3$ median filter. Next, we divide the gradient magnitude map $\pmb{g}_t$ into non-overlapping patches, obtaining the set of gradient patches $\left\{\pmb{r}_i^{(t)}\right\}_{i = 1}^n\in \mathbb{R}^{d\times d\times c}$ . Inside each gradient patch, we compute the maximum gradient magnitude per channel. Then, we compute the channel-wise mean over the maximum gradient magnitudes, as follows: + +$$ +m _ {i} ^ {(t)} = \frac {1}{c} \sum_ {l = 1} ^ {c} \max _ {j, k} \left\{\boldsymbol {r} _ {i j k l} ^ {(t)} \right\}, \forall j, k \in \{1, \dots , d \}. \tag {1} +$$ + +Finally, we compute the token-wise weights for the reconstruction loss as follows: + +$$ +w _ {i} ^ {(t)} = \frac {m _ {i} ^ {(t)}}{\sum_ {j = 1} ^ {n} m _ {j} ^ {(t)}}, \forall i \in \{1, \dots , n \}. \tag {2} +$$ + +Introducing the resulting weights $w_{i}^{(t)}$ into the conventional token-level reconstruction loss leads to an objective that pushes the masked AE to focus on reconstructing the patches with high motion magnitude. Formally, our weighted mean squared error loss is given by: + +$$ +\mathcal {L} _ {\mathrm {w M S E}} \left(\boldsymbol {x} _ {t}, \theta_ {T}\right) = \frac {1}{n} \sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \cdot \left\| \boldsymbol {p} _ {i} ^ {(t)} - \hat {\boldsymbol {p}} _ {i} ^ {(t)} \right\| _ {2} ^ {2}, \tag {3} +$$ + +where $\theta_T$ are the weights our teacher masked AE. Although our reconstruction loss focuses on tokens with high motion, the masked tokens are still chosen randomly. + +Self-distillation. Knowledge distillation has already shown its utility in anomaly detection [8, 12, 16, 26, 74]. Intuitively, since the teacher and student models are both trained on normal data, their reconstructions should be very similar for normal test samples. However, their behavior is not guaranteed to be similar on abnormal examples. Therefore, the magnitude of the teacher-student output gap (discrepancy) can serve as a means to quantify the anomaly level of a given sample. Unfortunately, this approach implies using both teacher and student models during inference, virtually splitting our processing speed in half. To slash the additional burden of using another model during inference, we propose to employ a novel variant of self-distillation with a shared encoder and two decoders, a teacher and a student. More precisely, the student branches out from the original architecture after the first transformer block of the teacher decoder, essentially adding only one transformer block. + +Our training process is carried out in two stages. In the first phase, the teacher is trained with the loss defined in Eq. (3). In the second phase, we freeze the weights of the shared backbone and train only the student decoder via self-distillation. The self-distillation loss is similar to the one defined in Eq. (3). The main difference is that instead of reconstructing the patches from the real image, the student learns to reconstruct the ones produced by the teacher. Let + +![](images/c5b99fc58c65f8eb9dc5e443924c58b9508c05d837d8239da49b564db78a86a2.jpg) +Figure 3. Four synthetic anomalies (with red contours) taken from the UBnormal data set [1] and overlaid on training frames from Avenue [51]. Best viewed in color. + +$\left\{\tilde{\pmb{p}}_i^{(t)}\right\}_{i = 1}^n \in \mathbb{R}^{d\times d\times c}$ denote the patches reconstructed by the student. Then, the self-distillation loss can be expressed as follows: + +$$ +\mathcal {L} _ {\mathrm {S D}} \left(\hat {\boldsymbol {x}} _ {t}, \theta_ {S}\right) = \frac {1}{n} \sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \cdot \left\| \hat {\boldsymbol {p}} _ {i} ^ {(t)} - \tilde {\boldsymbol {p}} _ {i} ^ {(t)} \right\| _ {2} ^ {2}, \tag {4} +$$ + +where $\hat{\pmb{x}}_t$ is the frame reconstructed by the teacher, and $\theta_S$ are the weights of the student decoder. Notice that we keep the motion gradient weights $w_i^{(t)}$ during self-distillation. + +Synthetic anomalies. As observed in other studies [5, 36], AEs tend to generalize too well to out-of-distribution data. This behavior is not desired in anomaly detection, since methods based on AEs rely on having high reconstruction errors for abnormal examples and low reconstruction errors for normal ones. To this end, we propose to augment the training videos with abnormal events. Since collecting abnormal training examples from the real-world is not possible, we resort to adding synthetic (virtual) anomalies. We leverage the recently introduced UBnormal data set [1] and its accurate pixel-level annotations to crop out abnormal events and blend them in our training videos, while ensuring the temporal consistency of the added events. The resulting examples, some depicted in Figure 3, are used to augment the training set with extra data. + +The synthetic examples help our model in three ways. First, in the reconstruction loss, we consider the original training frames (without superimposed anomalies) as the ground-truth, essentially forcing our model to overlook the anomalies. Formally, in Eq. (3), we use the patches $\left\{\pmb{p}_i^{(t)}\right\}_{i = 1}^n$ from the normal version of frame $\pmb{x}_t$ . Second, we add the anomaly map as an additional channel to our target image. In the anomaly map, we set normal pixels to 0 and abnormal pixels to 1. This change implies that, in Eq. (3) and Eq. (4), all patches will have an additional channel. Third, we use the ground-truth anomaly map to enhance the weights defined in Eq. (2). The added synthetic anomalies do not necessarily yield motion gradients with high magnitude. Hence, it is possible to have low weights in Eq. (3) and Eq. (4) for patches that correspond to anomaly regions. This is not desirable if we want the model to detect anom + +lies. To this end, we propose to add the anomaly maps and the gradients together, before computing the weights as in Eq. (2). Formally, in Eq. (1), we replace $\left\{\pmb{r}_i^{(t)}\right\}_{i=1}^n$ with $\left\{\pmb{r}_i^{(t)} + \pmb{a}_i^{(t)}\right\}_{i=1}^n$ , where $\left\{\pmb{a}_i^{(t)}\right\}_{i=1}^n$ is the set of the patches extracted from the anomaly map. + +Classification head. We further harness the synthetic anomalies to train a classification head applied on the final [CLS] token of the shared encoder. The head is trained to discriminate between frames with and without synthetic anomalies. This head is trained using binary cross-entropy: + +$$ +\mathcal {L} _ {\mathrm {C E}} \left(\hat {\boldsymbol {x}} _ {t}, \theta_ {E}\right) = - y _ {t} \cdot \log \left(\hat {y} _ {t}\right) - \left(1 - y _ {t}\right) \cdot \log \left(1 - \hat {y} _ {t}\right), \tag {5} +$$ + +where $y_{t} \in \{0,1\}$ is 1 if the frame contains an anomaly and 0 otherwise, $\hat{y}_t$ is the prediction, and $\theta_E$ represents the set of weights of the shared encoder. + +Inference. During inference, we pass each frame $\pmb{x}_t$ through both teacher and student models to obtain the reconstructed frames $\hat{\pmb{x}}_t$ and $\tilde{\pmb{x}}_t$ , respectively. Then, we compute output pixel-level anomaly map as: + +$$ +\boldsymbol {o} _ {t} = \alpha \cdot \| \boldsymbol {x} _ {t} - \hat {\boldsymbol {x}} _ {t} \| _ {2} ^ {2} + \beta \cdot \| \hat {\boldsymbol {x}} _ {t} - \tilde {\boldsymbol {x}} _ {t} \| _ {2} ^ {2} + \gamma \cdot \hat {y} _ {t}, \tag {6} +$$ + +where $\alpha$ , $\beta$ and $\gamma$ are hyperparameters that control the importance of the individual anomaly score components. Following [15, 35], we apply spatio-temporal 3D filtering to smooth the anomaly volumes. To obtain the frame-level anomaly scores, we keep the maximum value from each output map $\pmb{o}_t$ and subsequently apply another temporal Gaussian filter to smooth the values. + +# 4. Experiments + +# 4.1. Experimental Setup + +Data sets. We verify the performance of our method on four data sets for video anomaly detection: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. ShanghaiTech is the largest data set, with 270K frames for training and about 50K for testing. UBnormal is the second largest, with about 116K training frames and 93K testing frames. Avenue is a popular benchmark containing 15K frames for training and another 15K for testing. UCSD Ped2 holds a total of 4.5K frames, out of which 2.5K are used for training. UBnormal is a benchmark that uses an open set evaluation, where training and test anomalies belong to disjoint category sets. For the other three data sets, the training videos contain only normal events, and the test ones include both normal and abnormal scenarios. To augment the normal training videos, we sample abnormal events from the UBnormal data set [1]. UBnormal is a synthetic (virtual) data set containing anomalies simulated by video game characters, which alleviates the burden of collecting anomalies from the real world. The probability of augmenting a frame from Avenue, ShanghaiTech or UCSD Ped2 is 0.25. + +Evaluation. We evaluate all models following recent related works [1, 27, 69], considering both micro and macro + +
TypeMethodAvenueShanghaiUBnormalPed2FPS
MicroMacroMicroMacroMicroMacroMicroMacro
Object-centric[10]91.692.583.890.562.186.5--20
[18]86.4-71.6---97.8--
[26]91.592.882.490.255.484.597.599.851
[27]92.390.482.789.361.385.698.799.724
[33]--85.9-71.8----
[36]87.490.478.784.9--94.397.8-
[40]80.2-73.7-50.7----
[49]89.993.574.283.2--99.3-12
[50]91.892.383.887.8-----
[54]+[10]91.692.483.690.6----20
[54]+[27]93.291.883.389.3----31
[54]+[49]89.593.675.283.8----10
[69]+[27]92.991.983.689.5----31
[69]+[49]90.992.275.583.7----10
[84]92.2-84.3---99.0-35
[95]89.6-74.8---97.3--
Frame or cube level[4]87.1-75.9---96.5--
[5]84.7-73.7---98.4--
[9]----68.580.3--37
[28]83.3-71.2---94.1-35
[37]88.9--------
[43]90.0-----96.6--
[47]85.181.772.880.6--95.4-28
[48]84.4-----87.5--
[54]+[47]89.184.874.683.3----26
[54]+[61]86.486.370.680.3----94
[57]86.9-----96.2--
[60]85.3-72.2---96.3-195
[61]82.886.868.379.7--97.0-101
[62]72.0-----88.3--
[63]87.2-----93.0--
[66]------93.5--
[67]------88.4--
[69]+[47]87.384.574.582.9----26
[69]+[61]84.888.669.880.2----95
[76]84.6--------
[77]---76.550.376.8--56
[78]89.6-74.7------
[80]85.1-73.0---96.3--
[82]--76.1------
[87]87.0-79.3------
[89]--80.4------
[90]86.6-----96.9--
[93]90.1-78.6-62.7----
[96]90.2-----97.3--
[100]--78.9------
[103]------91.0--
Ours91.390.979.184.758.581.495.498.41655
+ +Table 1. Micro and macro AUC scores (in %) of several state-of-the-art frame-level, cube-level and object-level methods versus our self-distilled masked AE on Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The top three scores for each category of methods are shown in red, green, and blue. All reported running times (including those of the baselines) are measured on a machine with an Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM. + +AUC metrics. The area under the ROC curve (AUC) expresses the overlap between the ground-truth frame-level annotations and the anomaly scores predicted by a model, at multiple thresholds. At a given threshold, a frame is labeled as abnormal if the predicted anomaly score is above the threshold. For the micro AUC, the test frames from all + +![](images/4756055ee566935d5e1262423f127167da719791fff36bc6342a50cd1feb2f7b.jpg) +Figure 4. Predictions for test video 04 from Avenue. The abnormal bounding boxes are given by the convex hull of the patches labeled as abnormal. Best viewed in color. + +videos are concatenated before computing the AUC over all frames. For the macro AUC, the AUC of each test video is first computed, and the resulting AUC scores are averaged to obtain a single value. + +Hyperparameters. The encoder module is formed of three CvT blocks, each with a projection size of 256 and four attention heads. The teacher decoder contains three CvT blocks, while the student decoder contains only one block. All decoder blocks have four attention heads and a projection dimension of 128. Since the data sets have different input resolutions and objects vary in size, we adapt the patch size to each data set. Thus, we set the patch size to $16 \times 16$ on Avenue, $8 \times 8$ on ShanghaiTech and UBnormal, and $4 \times 4$ on UCSD Ped2. Regardless of the data set, the teacher network is trained for 100 epochs, while the student is trained for 40 epochs. We optimize the networks with Adam [42], using a learning rate of $10^{-4}$ and mini-batches of 100 samples. The hyperparameters in Eq. (6) are set to $\alpha = 0.4$ , $\beta = 0.3$ and $\gamma = 0.3$ , for all data sets. + +# 4.2. Results + +We present our results in Table 1 and discuss them below. + +Results on Avenue. Our method obtains a micro AUC score of $91.3\%$ on Avenue, being only $1.9\%$ below the state-of-the-art object-centric method. Remarkably, in the category of frame-level methods, we reach the best micro and macro AUC scores. Taking into account that our method is much faster than all other methods, we consider that its performance is remarkable. In Figure 4, we illustrate the anomaly scores for test video 04 from Avenue. Here, our model is close to perfect, highlighting its ability to capture anomalies, such as people running. + +Results on ShanghaiTech. On ShanghaiTech, our method reaches the top macro AUC score and the third-best micro AUC score, when compared with other frame-level frameworks. Object-centric methods generally surpass frame-level methods, but the former methods have much lower processing speeds (see Figure 2). + +
Motion weightsSelf-distillationSynthetic dataAnomaly mapsClassif. headAvenueShanghai
AUC
MicroMacroMicroMacro
84.085.669.780.1
✓84.886.371.380.9
✓✓88.586.076.383.8
✓✓✓88.586.977.083.0
✓✓✓✓90.589.677.384.2
✓✓✓✓✓91.390.979.184.7
+ +Table 2. Impact of each novel component on the micro and macro AUC scores (in %), on Avenue [51] and ShanghaiTech [53]. + +
StrategyAvenueShanghai
AUC
MicroMacroMicroMacro
Teacher84.085.674.882.3
Teacher + Student85.485.875.182.1
Teacher + Teacher-Student Difference88.586.076.383.8
Teacher + Student + Teacher-Student Difference86.985.875.883.6
+ +Table 3. Impact of strategies to combine the outputs of the teacher and student models on Avenue [51] and ShanghaiTech [53]. These results do not include the synthetic anomalies and the classification head. + +
Data setMeasurePercentage of synthetic data
0%25%50%75%
AvenueMicro AUC88.591.390.689.9
Macro AUC86.090.989.487.7
ShanghaiMicro AUC76.379.177.977.7
Macro AUC83.884.784.484.1
+ +Table 4. Impact of varying the proportion of synthetic anomalies on the Avenue [51] and ShanghaiTech [53] data sets. + +Results on UBnormal. In terms of the micro AUC, the best frame-level method on UBnormal is TimeSformer [9], which benefits from large-scale pretraining. Notably, our method obtains a higher macro AUC than TimeSformer. The micro AUC of our method is fairly close to the micro AUC levels of the better object-centric approaches. In terms of speed, our method is significantly faster than all the other methods reporting results on UBnormal. + +Results on UCSD Ped2. Our framework obtains a micro AUC of $95.4\%$ , being $3.9\%$ below the state-of-the-art performance of Liu et al. [49]. While being slightly below in terms of micro AUC, our macro AUC score is remarkably on par with the object-centric method of Ionescu et al. [36], and only $1.4\%$ below that of Georgescu et al. [26]. Considering that our FPS is more than 30 times higher compared with these methods [26, 36] on the same GPU, our framework provides a clearly superior accuracy-speed trade-off. + +Ablation study. In Table 2, we illustrate the impact of each novel component on our model's performance. The first model is a vanilla masked AE, which obtains a rather low performance on both Avenue and ShanghaiTech. Each and every component contributes towards boosting the performance of the vanilla model. We observe that self-distillation gives the highest boost in terms of the micro AUC. How- + +![](images/f893c7daf0da06f86eed77efa86fcde04a3a3a9f321e024967f37e01b4f63141.jpg) +Figure 5. Examples of frames and anomaly maps reconstructed by our teacher. The first four columns correspond to abnormal examples from the Avenue and ShanghaiTech data sets, while the last column shows a normal example. Best viewed in color. + +ever, to surpass the $90\%$ milestone on Avenue, it is mandatory to introduce the prediction of anomaly maps in the learning task. An additional boost is given by our classification head. + +The self-distillation procedure gives us a few possible strategies to combine the outputs of the teacher and the student. Hence, we investigate this aspect and report the results in Table 3. We observe that the best micro AUC is obtained when we combine the teacher reconstruction error with the teacher-student discrepancy. + +The proportion of synthetic examples per mini-batch is another aspect that can influence our model's performance. In Table 4, we report the micro and macro AUC scores for three possible augmentation levels. We observe that augmentation is always useful, but, for a better outcome, it requires a moderate percentage (25%). + +Performance-speed trade-off. In Figure 2, we compare our model with several other methods in terms of the performance-speed trade-off. This comparison undoubtedly shows that our method reaches a far better processing speed, while achieving fairly good performance. To strengthen this observation, we also compare the methods in terms of GFLOPs and number of parameters in Table 5. We underline that the method of Gong et al. [28] might seem small in terms of the number of parameters, but it is slowed down by its input, which is formed of a cuboid constructed by stacking 16 consecutive frames. Moreover, the method relies on a memory module, where each memory slot records the features of one pixel in the activation maps. Although their method has twice as many parameters as our own, the large input volume and the sizable memory bank reduce the speed to 35 FPS. With an FPS of 1655, our method proves to be significantly lighter than all its competitors. + +Qualitative results. In Figure 5, we present the frames and anomaly maps reconstructed by the teacher in four abnormal scenarios from Avenue and ShanghaiTech. Moreover, + +
MethodGFLOPs ↓#Params (M) ↓FPS ↑
Georgescu et al. [26]107.96551
Georgescu et al. [27]121.66724
Liu et al. [49]179.532012
Park et al. [60]8464195
Gong et al. [28]55.2635
Ours0.831655
+ +Table 5. Comparing methods in terms of floating point operations (GFLOPs), number of parameters, and FPS. + +in the fifth column, we illustrate the behavior of the teacher in a normal scenario. In all four abnormal cases, the reconstruction error is visibly higher for anomalous regions. This effect is mostly due to our training procedure based on synthetic data augmentation. The most obvious example is in the fourth column, where the bicycle seen in a pedestrian area is almost entirely removed from the output. Additionally, the predicted anomaly maps are well aligned with the ground-truth ones. + +# 5. Conclusion + +In this work, we proposed a lightweight masked autoencoder (3M parameters, 0.8 GFLOPs) for video anomaly detection, which learns to reconstruct tokens with high motion gradients. Our framework is based on self-distillation, leveraging the discrepancy between teacher and student decoders for anomaly detection. Moreover, we boost the performance of our model by introducing a data augmentation technique based on overlapping synthetic anomalies on normal training data. Our highly efficient framework reached an unprecedented speed of 1655 FPS, with a minimal performance gap with respect to the state-of-the-art object-centric approaches. + +Acknowledgments. Work supported by a grant of the Romanian Ministry of Education and Research, CNCS-UEFISCDI, project number PN-III-P2-2.1-PED-2021-0195, within PNCDI III. + +# References + +[1]Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. UBnormal: New Benchmark for Supervised Open-Set Video Anomaly Detection. In Proceedings of CVPR, pages 20143-20153, 2022. 1, 2, 5, 6 +[2] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust Real-Time Unusual Event Detection Using Multiple Fixed-Location Monitors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):555-560, 2008. 3, 4 +[3] Borislav Antic and Bjorn Ommer. Video parsing for abnormality detection. In Proceedings of ICCV, pages 2415-2422, 2011. 2, 3 +[4] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. In Proceedings of BMVC, 2021. 6 +[5] Marcella Astrid, Muhammad Zaigham Zaheer, and Seung-Ik Lee. Synthetic Temporal Anomaly Guided End-to-End Video Anomaly Detection. In Proceedings of ICCVW, pages 207-214, 2021. 2, 5, 6 +[6] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In Proceedings of NIPS, pages 2654-2662, 2014. 4 +[7] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. MultiMAE: Multi-modal multi-task masked autoencoders. In Proceedings of ECCV, pages 348-367. Springer, 2022. 4 +[8] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed Students: Student-Teacher Anomaly Detection With Discriminative Latent Embeddings. In Proceedings of CVPR, pages 4183–4192, 2020. 4, 5 +[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In Proceedings of ICML, 2021. 6, 7 +[10] Antonio Bărbălău, Radu Tudor Ionescu, Mariana-Iuliana Georgescu, Jacob Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. SSMTL++: Revisiting Self-Supervised Multi-Task Learning for Video Anomaly Detection. Computer Vision and Image Understanding, 229: 103656, 2023. 1, 2, 3, 6 +[11] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. SdAE: Self-distillated Masked Autoencoder. In Proceedings of ECCV, pages 108-124, 2022. 4 +[12] Hekai Cheng, Lu Yang, and Zulong Liu. Relation-based knowledge distillation for anomaly detection. In Proceedings of PRCV, pages 105-116, 2021. 4, 5 +[13] Kai-Wen Cheng, Yie-Tarng Chen, and Wen-Hsien Fang. Video anomaly detection and localization using hierarchical feature representation and Gaussian process regression. In Proceedings of CVPR, pages 2909-2917, 2015. 2, 3 +[14] Y. Cong, J. Yuan, and J. Liu. Sparse reconstruction cost for + +abnormal event detection. In Proceedings of CVPR, pages 3449-3456, 2011. 2, 3 +[15] Allison Del Giorno, J. Andrew Bagnell, and Martial Hebert. A Discriminative Framework for Anomaly Detection in Large Videos. In Proceedings of ECCV, pages 334-349, 2016. 2, 3, 6 +[16] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of CVPR, pages 9737-9746, 2022. 4, 5 +[17] Fei Dong, Yu Zhang, and Xiushan Nie. Dual Discriminator Generative Adversarial Network for Video Anomaly Detection. IEEE Access, 8:88170-88176, 2020. 1, 2 +[18] Keval Doshi and Yasin Yilmaz. Any-Shot Sequential Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 934-935, 2020. 1, 2, 3, 6 +[19] Keval Doshi and Yasin Yilmaz. Continual Learning for Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 254-255, 2020. 3 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of ICLR, 2021. 4 +[21] Jayanta K. Dutta and Bonny Banerjee. Online Detection of Abnormal Events Using Incremental Coding Length. In Proceedings of AAAI, pages 3755-3761, 2015. 2, 3 +[22] Yaxiang Fan, Gongjian Wen, Deren Li, Shaohua Qiu, Martin D. Levine, and Fei Xiao. Video anomaly detection and localization via Gaussian Mixture Fully Convolutional Variational Autoencoder. Computer Vision and Image Understanding, 195:102920, 2020. 2 +[23] Ye Fei, Chaoqin Huang, Cao Jinkun, Maosen Li, Ya Zhang, and Cewu Lu. Attribute Restoration Framework for Anomaly Detection. IEEE Transactions on Multimedia, 24: 116-127, 2022. 3 +[24] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In Proceedings of NeurIPS, 2022. 4 +[25] Yachuang Feng, Yuan Yuan, and Xiaoqiang Lu. Learning deep event models for crowd anomaly detection. Neurocomputing, 219:548-556, 2017. 3 +[26] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly Detection in Video via Self-Supervised and Multi-Task Learning. In Proceedings of CVPR, pages 12742–12752, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[27] Mariana Iuliana Georgescu, Radu Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A Background-Agnostic Framework with Adversarial Training for Abnormal Event Detection in Video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4505-4523, 2022. 2, 3, 6, 8 +[28] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton Van Den Hengel. Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised + +Anomaly Detection. In Proceedings of ICCV, pages 1705-1714, 2019. 1, 2, 3, 6, 8 +[29] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In Proceedings of CVPR, pages 733-742, 2016. 2, 3 +[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Proceedings of CVPR, pages 16000-16009, 2022. 2, 3, 4 +[31] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint Detection and Recounting of Abnormal Events by Learning Deep Generic Knowledge. In Proceedings of ICCV, pages 3639-3647, 2017. 3 +[32] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the Knowledge in a Neural Network. In Proceedings of NIPS Deep Learning and Representation Learning Workshop, 2014. 4 +[33] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In Proceedings of ICCV, pages 13545-13554, 2023. 6 +[34] Chaoqin Huang, Qinwei Xu, Yanfeng Wang, Yu Wang, and Ya Zhang. Self-Supervised Masking for Unsupervised Anomaly Detection and Localization. IEEE Transactions on Multimedia, 2022. 4 +[35] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In Proceedings of ICCV, pages 2895-2903, 2017. 2, 3, 6 +[36] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-Centric AutoEncoders and Dummy Anomalies for Abnormal Event Detection in Video. In Proceedings of CVPR, pages 7842-7851, 2019. 1, 2, 3, 4, 5, 6, 7 +[37] Radu Tudor Ionescu, Sorina Smeureanu, Marius Popescu, and Bogdan Alexe. Detecting abnormal events in video using Narrowed Normality Clusters. In Proceedings of WACV, pages 1951-1960, 2019. 2, 3, 6 +[38] Xiangli Ji, Bairong Li, and Yuesheng Zhu. TAM-Net: Temporal Enhanced Appearance-to-Motion Generative Network for Video Anomaly Detection. In Proceedings of IJCNN, pages 1–8, 2020. 1 +[39] Jielin Jiang, Jiale Zhu, Muhammad Bilal, Yan Cui, Neeraj Kumar, Ruihan Dou, Feng Su, and Xiaolong Xu. Masked Swin Transformer Unet for Industrial Anomaly Detection. IEEE Transactions on Industrial Informatics, 19(2):2200-2209, 2022. 4 +[40] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. BiPOCO: Bi-Directional Trajectory Prediction with Pose Constraints for Pedestrian Anomaly Detection. In Proceedings of SL4AD, 2022. 6 +[41] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: A space-time MRF for detecting abnormal activities with incremental updates. In Proceedings of CVPR, pages 2921-2928, 2009. 2, 3 +[42] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of ICLR, 2015. 7 + +[43] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: Bidirectional Multi-Scale Aggregation Networks for Abnormal Event Detection. IEEE Transactions on Image Processing, 29:2395-2408, 2019. 1, 2, 3, 6 +[44] Guoqiu Li, Guanxiong Cai, Xingyu Zeng, and Rui Zhao. Scale-Aware Spatio-Temporal Relation Learning for Video Anomaly Detection. In Proceedings of ECCV, pages 333-350, 2022. 1, 3 +[45] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(1):18-32, 2014. 2 +[46] Zhenyu Li, Ning Li, Kaitao Jiang, Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Superpixel Masking and Inpainting for Self-Supervised Anomaly Detection. In Proceedings of BMVC, 2020. 3 +[47] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future Frame Prediction for Anomaly Detection - A New Baseline. In Proceedings of CVPR, pages 6536-6545, 2018. 2, 3, 6 +[48] Yusha Liu, Chun-Liang Li, and Barnabaás Póczos. Classifier Two-Sample Test for Video Anomaly Detections. In Proceedings of BMVC, 2018. 2, 3, 6 +[49] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction. In Proceedings of ICCV, pages 13588-13597, 2021. 1, 2, 3, 6, 7, 8 +[50] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating Anomalies for Video Anomaly Detection With Prompt-Based Feature Mapping. In Proceedings of CVPR, pages 24500-24510, 2023. 6 +[51] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal Event Detection at 150 FPS in MATLAB. In Proceedings of ICCV, pages 2720-2727, 2013. 2, 3, 4, 5, 6, 7 +[52] Yiwei Lu, Frank Yu, Mahesh Kumar, Krishna Reddy, and Yang Wang. Few-Shot Scene-Adaptive Anomaly Detection. In Proceedings of ECCV, pages 125–141, 2020. 1 +[53] Weixin Luo, Wen Liu, and Shenghua Gao. A Revisit of Sparse Coding Based Anomaly Detection in Stacked RNN Framework. In Proceedings of ICCV, pages 341-349, 2017. 2, 3, 4, 6, 7 +[54] Neelu Madan, Nicolae-Catalin Ristea, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised masked convolutional transformer block for anomaly detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):525-542, 2024. 6 +[55] Vijay Mahadevan, Wei-Xin LI, Viral Bhalodia, and Nuno Vasconcelos. Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 1975-1981, 2010. 2, 3, 4, 6 +[56] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of CVPR, pages 935-942, 2009. 2, 3 +[57] Trong-Nguyen Nguyen and Jean Meunier. Anomaly Detection in Video Sequence With Appearance-Motion Correspondence. In Proceedings of ICCV, pages 1273-1283, 2019. 1, 3, 6 + +[58] Guansong Pang, Cheng Yan, Chunhua Shen, Anton van den Hengel, and Xiao Bai. Self-trained Deep Ordinal Regression for End-to-End Video Anomaly Detection. In Proceedings of CVPR, pages 12173–12182, 2020. 1, 3 +[59] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM Computing Surveys, 54(2):1-38, 2021. 3 +[60] Chaewon Park, MyeongAh Cho, Minhyeok Lee, and Sangyoun Lee. FastAno: Fast anomaly detection via spatiotemporal patch transformation. In Proceedings of WACV, pages 2249-2259, 2022. 2, 3, 6, 8 +[61] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection. In Proceedings of CVPR, pages 14372-14381, 2020. 1, 2, 3, 6 +[62] Bharathkumar Ramachandra and Michael Jones. Street Scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of WACV, pages 2569-2578, 2020. 1, 3, 6 +[63] Bharathkumar Ramachandra, Michael Jones, and Ranga Vatsavai. Learning a distance function with a Siamese network to localize anomalies in videos. In Proceedings of WACV, pages 2598-2607, 2020. 3, 6 +[64] Bharathkumar Ramachandra, Michael Jones, and Ranga Raju Vatsavai. Perceptual metric learning for video anomaly detection. Machine Vision and Applications, 32:1432-1769, 2021. 3 +[65] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. A Survey of Single-Scene Video Anomaly Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2293-2312, 2022. 1, 3, 4 +[66] Mahdyar Ravanbakhsh, Moin Nabi, Enver Sangineto, Lucio Marcenaro, Carlo Regazzoni, and Nicu Sebe. Abnormal Event Detection in Videos using Generative Adversarial Nets. In Proceedings of ICIP, pages 1577-1581, 2017. 3, 6 +[67] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-Play CNN for Crowd Motion Analysis: An Application in Abnormal Event Detection. In Proceedings of WACV, pages 1689–1698, 2018. 3, 6 +[68] Huamin Ren, Weifeng Liu, Soren Ingvor Olsen, Sergio Escalera, and Thomas B. Moeslund. Unsupervised Behavior-Specific Dictionary Learning for Abnormal Event Detection. In Proceedings of BMVC, pages 28.1-28.13, 2015. 2, 3 +[69] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection. In Proceedings of CVPR, pages 13576-13586, 2022. 1, 2, 3, 6 +[70] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, Karpathy A., A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 4 + +[71] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, and Reinhard Klette. Deep-Cascade: Cascading 3D Deep Neural Networks for Fast Anomaly Detection and Localization in Crowded Scenes. IEEE Transactions on Image Processing, 26(4):1992-2004, 2017. 2, 3 +[72] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, Zahra Moayed, and Reinhard Klette. Deep-Anomaly: Fully Convolutional Neural Network for Fast Anomaly Detection in Crowded Scenes. Computer Vision and Image Understanding, 172:88-97, 2018. 3 +[73] Babak Saleh, Ali Farhadi, and Ahmed Elgammal. Object-Centric Anomaly Detection by Attribute-Based Reasoning. In Proceedings of CVPR, pages 787-794, 2013. 3 +[74] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution Knowledge Distillation for Anomaly Detection. In Proceedings of CVPR, pages 14902-14912, 2021. 4, 5 +[75] Venkatesh Saligrama and Zhu Chen. Video anomaly detection based on local statistical aggregates. In Proceedings of CVPR, pages 2112-2119, 2012. 2, 3 +[76] Sorina Smeureanu, Radu Tudor Ionescu, Marius Popescu, and Bogdan Alexe. Deep Appearance Features for Abnormal Behavior Detection in Video. In Proceedings of ICIAP, pages 779-789, 2017. 1, 3, 6 +[77] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In Proceedings of CVPR, pages 6479–6488, 2018. 6 +[78] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-Aware Context Reasoning for Unsupervised Abnormal Event Detection in Videos. In Proceedings of ACMMM, pages 184-192, 2020. 1, 6 +[79] Qianru Sun, Hong Liu, and Tatsuya Harada. Online growing neural gas for anomaly detection in changing surveillance scenes. Pattern Recognition, 64(C):187-201, 2017. 3 +[80] Yao Tang, Lin Zhao, Shanshan Zhang, Chen Gong, Guangyu Li, and Jian Yang. Integrating prediction and reconstruction for anomaly detection. Pattern Recognition Letters, 129:123-130, 2020. 1, 2, 3, 6 +[81] Hanh T.M. Tran and David Hogg. Anomaly Detection using a Convolutional Winner-Take-All Autoencoder. In Proceedings of BMVC, 2017. 2, 3 +[82] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised video anomaly detection. In Proceedings of ICIP, pages 2540-2544, 2023. 6 +[83] Hung Vu, Tu Dinh Nguyen, Trung Le, Wei Luo, and Dinh Phung. Robust Anomaly Detection in Videos Using Multilevel Representations. In Proceedings of AAAI, pages 5216-5223, 2019. 1 +[84] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Proceedings of ECCV, pages 494-511, 2022. 2, 3, 6 +[85] L. Wang, F. Zhou, Z. Li, W. Zuo, and H. Tan. Abnormal Event Detection in Videos Using Hybrid Spatio-Temporal + +Autoencoder. In Proceedings of ICIP, pages 2276-2280, 2018. 2 +[86] Xusheng Wang, Mingtao Pei, and Zhengang Nie. Self-trained video anomaly detection based on teacher-student model. In Proceedings of MLSP, pages 1-6, 2021. 4 +[87] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster Attention Contrast for Video Anomaly Detection. In Proceedings of ACMMM, pages 2463-2471, 2020. 1, 6 +[88] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvT: Introducing Convolutions to Vision Transformers. In Proceedings of ICCV, pages 22-31, 2021. 3, 4 +[89] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In Proceedings of ECCV, pages 729-745, 2022. 3, 6 +[90] Peng Wu, Jing Liu, and Fang Shen. A Deep One-Class Neural Network for Anomalous Event Detection in Complex Scenes. IEEE Transactions on Neural Networks and Learning Systems, 31(7):2609-2622, 2019. 1, 2, 6 +[91] Shandong Wu, Brian E. Moore, and Mubarak Shah. Chaotic Invariants of Lagrangian Particle Trajectories for Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 2054-2060, 2010. 3 +[92] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting Anomalous Events in Videos by Learning Deep Representations of Appearance and Motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2 +[93] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature Prediction Diffusion Model for Video Anomaly Detection. In Proceedings of ICCV, pages 5527-5537, 2023. 6 +[94] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 4 +[95] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze Test Helps: Effective Video Anomaly Detection via Learning to Complete Video Events. In Proceedings of ACMMM, pages 583-591, 2020. 1, 2, 3, 6 +[96] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Transactions on Neural Networks and Learning Systems, 33(8): 3572-3586, 2022. 3, 6 +[97] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proceedings of CVPR, pages 19313–19322, 2022. 1 +[98] Muhammad Zaigham Zaheer, Jin-ha Lee, Marcella Astrid, and Seung-Ik Lee. Old is Gold: Redefining the Adversarily Learned One-Class Classifier Training Paradigm. In Proceedings of CVPR, pages 14183-14193, 2020. +[99] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. CLAWS: Clustering Assisted + +Weakly Supervised Learning with Normalcy Suppression for Anomalous Event Detection. In Proceedings of ECCV, pages 358-376, 2020. +[100] Muhammad Zaigham Zaheer, Arif Mahmood, Haris M. Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative Cooperative Learning for Unsupervised Video Anomaly Detection. In Proceedings of CVPR, pages 14744-14754, 2022. 1, 3, 6 +[101] Linfeng Zhang, Chenglong Bao, and Kaisheng Ma. Self-Distillation: Towards Efficient and Compact Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4388-4403, 2022. 2, 3, 4 +[102] Xinfeng Zhang, Su Yang, Jiulong Zhang, and Weishan Zhang. Video Anomaly Detection and Localization using Motion-field Shape Description and Homogeneity Testing. Pattern Recognition, page 107394, 2020. 2 +[103] Ying Zhang, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Shun Sakai. Video anomaly detection based on locality sensitive hashing filters. Pattern Recognition, 59:302-311, 2016. 2, 3, 6 +[104] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online Detection of Unusual Events in Videos via Dynamic Sparse Coding. In Proceedings of CVPR, pages 3313-3320, 2011. +[105] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph Convolutional Label Noise Cleaner: Train a Plug-And-Play Action Classifier for Anomaly Detection. In Proceedings of CVPR, pages 1237–1246, 2019. 2 \ No newline at end of file diff --git a/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/images.zip b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..86cb1de211f55b5d04263da2c18a8f9934e48a5c --- /dev/null +++ b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17fda2fce85f31bc16ed003beb6fb6e489cd35ed91156deb3f5a9a1a74c6aa79 +size 497367 diff --git a/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/layout.json b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b30c547a65fc8e3a23924ebe5a46a62742f5de08 --- /dev/null +++ b/2024/Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors/layout.json @@ -0,0 +1,9085 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "text", + "content": "Self-Distilled Masked Auto-Encoders are Efficient Video Anomaly Detectors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "text", + "content": "Nicolae-Cătălin Ristea" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "inline_equation", + "content": "^{1,2,\\diamond}" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "text", + "content": ", Florinel-Alin Croitoru" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "inline_equation", + "content": "^{1,\\diamond}" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "text", + "content": ", Radu Tudor Ionescu" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "inline_equation", + "content": "^{1,3,*}" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "text", + "content": ", Marius Popescu" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "text", + "content": ", Fahad Shahbaz Khan" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "text", + "content": ", Mubarak Shah" + }, + { + "bbox": [ + 61, + 133, + 531, + 162 + ], + "type": "inline_equation", + "content": "^{6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 163, + 482, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 163, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 110, + 163, + 482, + 205 + ], + "type": "text", + "content": "1University of Bucharest, Romania, 2NUST Politehnica Bucharest, Romania, 3SecurifAI, Romania, 4MBZ University of Artificial Intelligence, UAE, 5Linköping University, Sweden, 6University of Central Florida, US" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 224, + 192, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 224, + 192, + 236 + ], + "spans": [ + { + "bbox": [ + 143, + 224, + 192, + 236 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 240, + 290, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 290, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 290, + 516 + ], + "type": "text", + "content": "We propose an efficient abnormal event detection model based on a lightweight masked auto-encoder (AE) applied at the video frame level. The novelty of the proposed model is threefold. First, we introduce an approach to weight tokens based on motion gradients, thus shifting the focus from the static background scene to the foreground objects. Second, we integrate a teacher decoder and a student decoder into our architecture, leveraging the discrepancy between the outputs given by the two decoders to improve anomaly detection. Third, we generate synthetic abnormal events to augment the training videos, and task the masked AE model to jointly reconstruct the original frames (without anomalies) and the corresponding pixel-level anomaly maps. Our design leads to an efficient and effective model, as demonstrated by the extensive experiments carried out on four benchmarks: Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The empirical results show that our model achieves an excellent trade-off between speed and accuracy, obtaining competitive AUC scores, while processing 1655 FPS. Hence, our model is between 8 and 70 times faster than competing methods. We also conduct an ablation study to justify our design. Our code is freely available at: https://github.com/ristea/aed-mae." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 128, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "type": "text", + "content": "In recent years, research on abnormal event detection in video gained significant traction [1, 10, 17, 18, 26-28, 36, 38, 43, 44, 49, 52, 57, 58, 61, 62, 65, 69, 76, 78, 80, 83, 87, 90, 95, 97-100], due to its utter importance in video surveillance. Despite the growing interest, video anomaly detection remains a complex task, owing its complexity to the fact that abnormal situations are context-dependent and do not occur very often. This makes it very difficult to collect a representative set of abnormal events for training state-of-the-art deep learning models in a fully supervised manner. To showcase the rarity and reliance on context of anomalies, we refer to the vehicle ramming attacks carried out by terrorists against pedestrians. As soon as a car is steered" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 224, + 545, + 529 + ], + "blocks": [ + { + "bbox": [ + 309, + 224, + 545, + 529 + ], + "lines": [ + { + "bbox": [ + 309, + 224, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 309, + 224, + 545, + 529 + ], + "type": "image", + "image_path": "1873e97897d64d12c82b39939d5b3a6ded4ec79e35d3f775ddeb74fa8f4e1379.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 530, + 547, + 619 + ], + "lines": [ + { + "bbox": [ + 304, + 530, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 547, + 619 + ], + "type": "text", + "content": "Figure 1. Our masked auto-encoder for abnormal event detection based on self-distillation. At training time, some video frames are augmented with synthetic anomalies. The teacher decoder learns to reconstruct original frames (without anomalies) and predict anomaly maps. The student decoder learns to reproduce the teacher's output. Motion gradients are aggregated at the token level and used as weights for the reconstruction loss. Red dashed lines represent steps executed only during training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 626, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 547, + 688 + ], + "type": "text", + "content": "on the sidewalk, it becomes an abnormal event. Hence, the place where the car is driven (street versus sidewalk) determines the normal or abnormal label of the action, i.e. the label depends on context. Furthermore, there are less than 200 vehicle ramming attacks registered to date1, confirming" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 276, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 276, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 276, + 713 + ], + "type": "text", + "content": "*corresp. author: raducu.ionescu@gmail.com; equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 693, + 546, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 546, + 712 + ], + "type": "text", + "content": "1https://en.wikipedia.org/wiki/Vehicle-rammingattack" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15984" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 289, + 297 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 297 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 297 + ], + "type": "image", + "image_path": "9ac81caecf002633eda4ec46e306012c427e194757851963aa694c74f90abdbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 288, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 365 + ], + "type": "text", + "content": "Figure 2. Performance versus speed trade-offs for our self-distilled masked AE and several state-of-the-art methods [26-28, 47, 49, 60, 61, 69, 84] (with open-sourced code), on the Avenue data set. The running times of all methods are measured on a computer with one Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 378, + 284, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 284, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 284, + 390 + ], + "type": "text", + "content": "the scarcity of such events (even less are caught on video)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 390, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 390, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 45, + 390, + 287, + 628 + ], + "type": "text", + "content": "Since training anomaly detectors under a fully supervised setting is not possible, most studies dealing with abnormal event detection took a distinct path, proposing variations of outlier detection methods [3, 13, 14, 17, 21, 22, 29, 37, 41, 43, 45, 47, 51, 53, 55, 56, 61-68, 71, 80, 90, 92, 102-105]. Such methods treat abnormal event detection as an outlier detection task, where a normality model trained on normal events is applied on both normal and abnormal events during inference, labeling events deviating from the learned model as abnormal. Different from the mainstream path based on outlier detection, we propose an approach to augment each training video scene with synthetic anomalies, by randomly superimposing temporal action segments from the synthetic UBnormal data set [1] on our real-world data sets. We thus introduce synthetic anomalies at training time, enabling our model to learn in an open-set supervised manner. Additionally, we force our model to reconstruct the original training frames (without anomalies) to limit its ability to reconstruct anomalies, hence generating higher errors when anomalies occur." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "A large body of work on video anomaly detection has focused on employing auto-encoders (AEs) to address the task [5, 22, 27-29, 36, 49, 81, 85], relying on the poor reconstruction capabilities of these models on out-of-distribution data. Since training is carried out only on normal examples, it is expected for AEs to exhibit high reconstruction errors when anomalies occur. However, several researchers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "text", + "content": "observed that AEs generalize too well [5, 36], being able to reconstruct anomalies with very high precision. Thus, to better leverage the reconstruction error of AEs in anomaly detection, researchers explored a few alternatives, from the use of dummy [36] or pseudo-anomalies [5, 27] to the integration of memory modules [28, 49, 61]. With the same purpose in mind, we propose to employ masked auto-encoders [30] in anomaly detection, introducing new ways to regulate their generalization capacity. Indeed, we go beyond employing the standard masked AE framework, and propose three novel changes to enhance the anomaly detection performance of our model. First, we propose to weight tokens based on the magnitude of motion gradients, raising the importance of tokens with higher motion in the reconstruction loss. This makes our model focus on reconstructing tokens with high motion, and avoid reconstructing the background scene, which is typically static for surveillance cameras. Second, we attach a classification head to discriminate between normal and pseudo-abnormal instances in the latent encoding space. Third, we integrate a teacher decoder and a student decoder into our masked AE architecture, where the student decoder learns to distill knowledge from the already optimized teacher. To reduce our processing time, we use a shared encoder for the teacher and student models, leading to a process known as self-distillation [101]. During the self-distillation process, the shared encoder is frozen. We leverage the discrepancy between the outputs given by the teacher and student decoders along with the reconstruction error of the teacher to boost anomaly detection performance. Our entire framework, integrating these components into a meticulous design, is shown in Figure 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 446, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 651 + ], + "type": "text", + "content": "State-of-the-art deep anomaly detectors [10, 26, 27, 84] typically rely on a costly object detection method to increase precision, limiting the processing bandwidth to one video stream per GPU, at around 20-30 FPS. However, for real-world video surveillance, e.g. monitoring an entire city with hundreds or thousands of cameras, the processing costs of object-centric video anomaly detectors are simply too high, given their power consumption and that one GPU can cost around $2,000. To this end, we turn our attention to developing a lightweight model (6 transformer blocks, 3M parameters), capable of processing around 66 video streams at 25 FPS, significantly reducing the processing costs. Different from competing models performing anomaly detection at the object [10, 18, 26, 27, 36, 84, 95] or spatio-temporal cube [15, 21, 35, 37, 41, 48, 51, 53, 55, 60, 71, 75, 103] levels, we present a model that takes whole video frames as input, which is significantly more efficient (see Figure 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "We carry out comprehensive experiments on four benchmarks: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. The empirical results show that our method is 8 to 70 times faster than competing methods [26-28, 47, 49, 60, 61, 69, 84], while achieving comparable ac" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15985" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "curacy levels. Aside from the main results, we conduct an ablation study showing that our novel design choices are supported by empirical evidence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 110, + 229, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 110, + 229, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 110, + 229, + 120 + ], + "type": "text", + "content": "In summary, our contribution is threefold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 122, + 287, + 264 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 122, + 287, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 287, + 157 + ], + "type": "text", + "content": "- We propose a lightweight masked auto-encoder for anomaly detection in video, which learns to reconstruct tokens with higher motion magnitude." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 158, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 286, + 205 + ], + "type": "text", + "content": "- We introduce a self-distillation training pipeline, leveraging the discrepancy between teacher and student decoders to obtain a significant accuracy boost for our highly efficient model (due to the shared encoder)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 206, + 286, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 286, + 264 + ], + "type": "text", + "content": "- To further boost the performance of our model, we introduce a data augmentation approach based on superimposing synthetic anomalies on normal training videos, which enables the masked AE model to learn with open-set supervision." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 274, + 133, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 133, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 133, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 292, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 472 + ], + "type": "text", + "content": "Anomaly detection in video is typically formulated as a one-class learning problem, where only normal data is available at training time. During test time, both normal and abnormal examples are present [59, 65]. There are several categories of anomaly detection approaches, including dictionary learning methods [13, 14, 21, 51, 68, 89], probabilistic models [2, 3, 25, 31, 41, 55, 56, 73, 91], change detection frameworks [15, 35, 48, 58], distance-based models [36, 37, 62, 63, 67, 71, 72, 75, 76, 79, 81] and reconstruction-based approaches [23, 28, 29, 46, 47, 53, 57, 61, 66, 69, 80]. Considering that reconstruction-based methods often reach state-of-the-art performance in anomaly detection [27, 69], a large body of works used the reconstruction-based paradigm in the past few years. To this end, we adopt this paradigm in our study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "content": "With respect to the level at which the anomaly detection is carried out, methods can be categorized into spatiotemporal cube-level methods [37, 47, 48, 51, 55, 56, 66, 67, 71, 75, 76, 96, 103], frame-level methods [47, 66, 67], and object-level methods [10, 18, 19, 26, 27, 36, 49, 84, 95]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": "Frame-level and cube-level methods. Before the deep learning era, preliminary abnormal event detection models commonly relied on taking short video sequences and dividing them into spatio-temporal cuboids [15, 21, 41, 51, 55, 75, 103]. The cubes are then considered as independent examples, being passed as input to a machine learning model. This mainstream practice continued during the deep learning period [29, 35, 37, 43, 44, 53, 60, 62, 63, 71, 76, 100], when deep networks have been used to extract features [35, 37, 48, 53] or learn [28, 29, 43, 44, 53, 60, 62-64, 71, 89, 100] from these spatio-temporal cubes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "At the same time, some studies considered using entire video frames as input [47, 66, 67]. For example, Liu et al. [47] proposed an effective algorithm, which learns to reconstruct the next frame of a short video sequence. A" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "more complex approach is proposed by Ravanbakhsh et al. [67], who employ optical-flow reconstruction to predict the anomalous regions from an input image. In a different study, Ravanbakhsh et al. [66] proposed to detect anomalies at the frame level via generative adversarial networks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 134, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 545, + 277 + ], + "type": "text", + "content": "Frame-level and cube-level methods have a common characteristic, namely their relatively high processing speed due to the reasonably fast preprocessing steps, as opposed to object-centric methods. Still, frame-level methods hold a stronger advantage in terms of time, since cube-level methods need to process each cube as an independent example. Indeed, it is more efficient to process a mini-batch of frames rather than several mini-batches of spatio-temporal cubes. However, cube-level methods often outperform frame-level methods. To this end, we propose a masked AE that takes whole frames as input, yet learns interactions between video patches, thus integrating the best of both worlds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 279, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 470 + ], + "type": "text", + "content": "To boost the performance of frame-level or cube-level methods, researchers explored the inclusion of various components, such as memory modules [28, 61] or masked convolutional blocks [69]. Although integrating additional modules into the framework leads to accuracy gains, the procedure often comes with efficiency drawbacks. In contrast, our goal is to achieve a superior trade-off between performance and speed, with a higher focus on efficiency. As such, we design a lightweight masked AE based on convolutional vision transformer (CvT) blocks [88], and propose several upgrades resulting in a minimal time overhead. For example, we employ knowledge distillation to leverage the discrepancy between the teacher and the student models. However, to keep the processing time to the bare minimum, we resort to self-distillation [101] and use a shared encoder for the teacher and student networks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 675 + ], + "type": "text", + "content": "Object-level methods. To reduce the number of false positive detections often observed for other methods, some recent studies [10, 26, 27, 36, 49, 84, 95] proposed to look for anomalous objects rather than anomalous frames or cubes. Object-centric methods use the prior information from an object detector, enabling the anomaly detector to focus only on objects. This kind of framework boosts the accuracy by significant margins, currently reaching state-of-the-art performance [10, 84]. However, a considerable drawback is that the inference speed of the whole framework is directly conditioned by the object detector's speed, which is often much lower than that of the anomaly detection network [26, 36]. Hence, the processing time is significantly limited. In contrast, we perform anomaly detection at the frame level, obtaining an inference speed that is between 32 to 70 times faster than object-centric models [10, 26, 27, 36, 49, 84, 95]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Masked auto-encoders in anomaly detection. He et al. [30] proposed masked auto-encoders as a pretraining method to obtain strong backbones for downstream tasks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15986" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "Since then, the method has been adopted in various fields, e.g. video processing [24] or multimodal learning [7], with remarkable results. We elaborate the connection to seemingly related masked AEs in the supplementary [11, 94]. The masking framework has also been used for anomaly detection in medical [34] and industrial [39] images. To the best of our knowledge, we are the first to propose a masked transformer-based auto-encoder for video anomaly detection. Moreover, we go beyond applying standard masked AEs, proposing several modifications leading to superior performance levels: emphasizing tokens with higher motion, augmenting training videos with synthetic anomalies, and employing self-distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 420 + ], + "type": "text", + "content": "Knowledge distillation in anomaly detection. Knowledge distillation [6, 32] was originally designed to compress one or multiple large models (teachers) into a lighter neural network (student). Recently adopted in anomaly detection [8, 12, 16, 26, 74, 86], knowledge distillation was deemed useful due to the possibility of leveraging the representation discrepancy between the teacher and the student networks, which is larger in the case of anomalies. For example, Bergmann et al. [8] trained an ensemble of student networks on normal data to reproduce the output of a deep feature extractor, which is pretrained on ImageNet [70]. The authors use the difference between the teacher label and the mean over student labels to detect abnormal pixels. Salehi et al. [74] employed a more thorough distillation process, called hint learning, in which the multi-level features of a teacher pretrained on ImageNet are distilled into a clone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 421, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 289, + 529 + ], + "type": "text", + "content": "Most studies based on knowledge distillation applied the framework to image anomaly detection [8, 12, 16, 74]. With few exceptions [26, 86], knowledge distillation in video anomaly detection remains largely unexplored. Wang et al. [86] employed the teacher-student training paradigm to learn from unlabeled video samples in a self-supervised manner. Georgescu et al. [26] integrated knowledge distillation as a proxy task into a multi-task learning framework for video anomaly detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 530, + 289, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 289, + 650 + ], + "type": "text", + "content": "Distinct from the aforementioned studies, to our knowledge, we are the first to introduce a variant of self-distillation in anomaly detection. Self-distillation [101] attaches multiple classification heads at various depths to boost the classification performance of a neural classifier. In contrast, we integrate self-distillation into a masked AE, employing two decoders of different depths. Due to the shared encoder, we are able to leverage the reconstruction discrepancy between the teacher and the student with a minimal computational overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 659, + 104, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 104, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 104, + 672 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Overview. We introduce a lightweight teacher-student transformer-based masked AE, which employs a two-stage training pipeline. In the first stage, we optimize a teacher" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "masked AE via a reconstruction loss that employs a novel weighting mechanism based on motion gradients. In the second stage, we optimize the last (and only) decoder block of a student masked AE, which shares most of the backbone (kept frozen) with its teacher, to preserve efficiency. Next, we describe how to create training videos with synthetic anomalies and train the masked AEs to jointly predict the anomaly maps and overlook (not reconstruct) the anomalies from training frames. Lastly, we introduce a classification head to distinguish between frames with and without synthetic anomalies, which further boosts the performance of our method, with a marginal computational overhead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 217, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 456 + ], + "type": "text", + "content": "Architecture. Our masked AE pursues the architectural principles proposed in [30]. Hence, the entire architecture is formed of visual transformer blocks. In contrast to He et al. [30], we replace the ViT [20] blocks with CvT blocks [88], aiming for higher efficiency. Our processing starts by dividing the input images into non-overlapping tokens and removing a certain number of tokens. The encoder embeds the remaining tokens via convolutional projection layers, and the result is processed by transformer blocks. The decoder operates on a complete set of tokens, those removed being replaced with mask tokens. Its architecture is symmetric to that of the encoder. For efficiency reasons, we only use three blocks for the encoder and three blocks for the decoder. Each block is equipped with four attention heads. To achieve further speed gains, we replace all dense layers inside the CvT blocks with pointwise convolutions. We consider the architecture described so far as a teacher network. A student decoder branches out from the teacher after the first transformer block of the main decoder, adding only one extra transformer block (as shown in Figure 1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 457, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 601 + ], + "type": "text", + "content": "Motion gradient weighting. Masked AEs [30] have been originally applied on natural images. In this context, reconstructing randomly masked tokens is a viable solution, since images have high foreground and background variations. However, abnormal event detection data sets [2, 51, 53, 55] contain videos from fixed cameras with static backgrounds [65]. Learning to reconstruct the static background via masked AEs is both trivial and useless. Hence, naively training masked AEs to reconstruct randomly masked tokens in video anomaly detection is suboptimal. To this end, we propose to take into account the magnitude of the motion gradients when computing the reconstruction loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": " be the video frame at index " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": " be the number of non-overlapping visual tokens (patches) of size " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "d\\times d\\times c" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": " from each frame " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": " is the number of input channels, and " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": " is a hyperparameter that directly determines " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{p}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": " denote the set of tokens in frame " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\left\\{\\hat{\\pmb{p}}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}" + }, + { + "bbox": [ + 304, + 601, + 547, + 700 + ], + "type": "text", + "content": " the corresponding set of reconstructed tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Following Ionescu et al. [36], we estimate the motion" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15987" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "text", + "content": "gradient map " + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t" + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "text", + "content": " of frame " + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "text", + "content": " by computing the absolute difference between consecutive frames, which are previously filtered with a " + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "text", + "content": " median filter. Next, we divide the gradient magnitude map " + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t" + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "text", + "content": " into non-overlapping patches, obtaining the set of gradient patches " + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{r}_i^{(t)}\\right\\}_{i = 1}^n\\in \\mathbb{R}^{d\\times d\\times c}" + }, + { + "bbox": [ + 47, + 72, + 289, + 185 + ], + "type": "text", + "content": ". Inside each gradient patch, we compute the maximum gradient magnitude per channel. Then, we compute the channel-wise mean over the maximum gradient magnitudes, as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 186, + 287, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 72, + 186, + 287, + 217 + ], + "type": "interline_equation", + "content": "m _ {i} ^ {(t)} = \\frac {1}{c} \\sum_ {l = 1} ^ {c} \\max _ {j, k} \\left\\{\\boldsymbol {r} _ {i j k l} ^ {(t)} \\right\\}, \\forall j, k \\in \\{1, \\dots , d \\}. \\tag {1}", + "image_path": "701569b1d7ebb4e851f4177fa25cae7560c52aad9177a75c6c0482fd140ab960.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 218, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 287, + 240 + ], + "type": "text", + "content": "Finally, we compute the token-wise weights for the reconstruction loss as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 242, + 287, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 242, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 93, + 242, + 287, + 275 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {(t)} = \\frac {m _ {i} ^ {(t)}}{\\sum_ {j = 1} ^ {n} m _ {j} ^ {(t)}}, \\forall i \\in \\{1, \\dots , n \\}. \\tag {2}", + "image_path": "290519d594c6140e77dbe643a92469852ed487847bd82b6743fc4087ec7d5f4a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 278, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 340 + ], + "type": "text", + "content": "Introducing the resulting weights " + }, + { + "bbox": [ + 47, + 278, + 287, + 340 + ], + "type": "inline_equation", + "content": "w_{i}^{(t)}" + }, + { + "bbox": [ + 47, + 278, + 287, + 340 + ], + "type": "text", + "content": " into the conventional token-level reconstruction loss leads to an objective that pushes the masked AE to focus on reconstructing the patches with high motion magnitude. Formally, our weighted mean squared error loss is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 342, + 287, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 342, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 74, + 342, + 287, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {w M S E}} \\left(\\boldsymbol {x} _ {t}, \\theta_ {T}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \\cdot \\left\\| \\boldsymbol {p} _ {i} ^ {(t)} - \\hat {\\boldsymbol {p}} _ {i} ^ {(t)} \\right\\| _ {2} ^ {2}, \\tag {3}", + "image_path": "cc2c1adb7002359648ce290bb2482cf78c86f20b615418ceace4910185b88cc7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\theta_T" + }, + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "type": "text", + "content": " are the weights our teacher masked AE. Although our reconstruction loss focuses on tokens with high motion, the masked tokens are still chosen randomly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 411, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 615 + ], + "type": "text", + "content": "Self-distillation. Knowledge distillation has already shown its utility in anomaly detection [8, 12, 16, 26, 74]. Intuitively, since the teacher and student models are both trained on normal data, their reconstructions should be very similar for normal test samples. However, their behavior is not guaranteed to be similar on abnormal examples. Therefore, the magnitude of the teacher-student output gap (discrepancy) can serve as a means to quantify the anomaly level of a given sample. Unfortunately, this approach implies using both teacher and student models during inference, virtually splitting our processing speed in half. To slash the additional burden of using another model during inference, we propose to employ a novel variant of self-distillation with a shared encoder and two decoders, a teacher and a student. More precisely, the student branches out from the original architecture after the first transformer block of the teacher decoder, essentially adding only one transformer block." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": "Our training process is carried out in two stages. In the first phase, the teacher is trained with the loss defined in Eq. (3). In the second phase, we freeze the weights of the shared backbone and train only the student decoder via self-distillation. The self-distillation loss is similar to the one defined in Eq. (3). The main difference is that instead of reconstructing the patches from the real image, the student learns to reconstruct the ones produced by the teacher. Let" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 72, + 520, + 168 + ], + "blocks": [ + { + "bbox": [ + 333, + 72, + 520, + 168 + ], + "lines": [ + { + "bbox": [ + 333, + 72, + 520, + 168 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 520, + 168 + ], + "type": "image", + "image_path": "c5b99fc58c65f8eb9dc5e443924c58b9508c05d837d8239da49b564db78a86a2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 171, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 306, + 171, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 545, + 203 + ], + "type": "text", + "content": "Figure 3. Four synthetic anomalies (with red contours) taken from the UBnormal data set [1] and overlaid on training frames from Avenue [51]. Best viewed in color." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 217, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\left\\{\\tilde{\\pmb{p}}_i^{(t)}\\right\\}_{i = 1}^n \\in \\mathbb{R}^{d\\times d\\times c}" + }, + { + "bbox": [ + 306, + 217, + 545, + 260 + ], + "type": "text", + "content": " denote the patches reconstructed by the student. Then, the self-distillation loss can be expressed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 262, + 545, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 262, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 339, + 262, + 545, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S D}} \\left(\\hat {\\boldsymbol {x}} _ {t}, \\theta_ {S}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} w _ {i} ^ {(t)} \\cdot \\left\\| \\hat {\\boldsymbol {p}} _ {i} ^ {(t)} - \\tilde {\\boldsymbol {p}} _ {i} ^ {(t)} \\right\\| _ {2} ^ {2}, \\tag {4}", + "image_path": "19b47cae0b0ce8fec351472c5e288910224910a30c544eaa77ab88e2ecb3358d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_t" + }, + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "text", + "content": " is the frame reconstructed by the teacher, and " + }, + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\theta_S" + }, + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "text", + "content": " are the weights of the student decoder. Notice that we keep the motion gradient weights " + }, + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "inline_equation", + "content": "w_i^{(t)}" + }, + { + "bbox": [ + 305, + 293, + 545, + 331 + ], + "type": "text", + "content": " during self-distillation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 333, + 546, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 333, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 546, + 514 + ], + "type": "text", + "content": "Synthetic anomalies. As observed in other studies [5, 36], AEs tend to generalize too well to out-of-distribution data. This behavior is not desired in anomaly detection, since methods based on AEs rely on having high reconstruction errors for abnormal examples and low reconstruction errors for normal ones. To this end, we propose to augment the training videos with abnormal events. Since collecting abnormal training examples from the real-world is not possible, we resort to adding synthetic (virtual) anomalies. We leverage the recently introduced UBnormal data set [1] and its accurate pixel-level annotations to crop out abnormal events and blend them in our training videos, while ensuring the temporal consistency of the added events. The resulting examples, some depicted in Figure 3, are used to augment the training set with extra data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 516, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 516, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 516, + 547, + 713 + ], + "type": "text", + "content": "The synthetic examples help our model in three ways. First, in the reconstruction loss, we consider the original training frames (without superimposed anomalies) as the ground-truth, essentially forcing our model to overlook the anomalies. Formally, in Eq. (3), we use the patches " + }, + { + "bbox": [ + 305, + 516, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{p}_i^{(t)}\\right\\}_{i = 1}^n" + }, + { + "bbox": [ + 305, + 516, + 547, + 713 + ], + "type": "text", + "content": " from the normal version of frame " + }, + { + "bbox": [ + 305, + 516, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 305, + 516, + 547, + 713 + ], + "type": "text", + "content": ". Second, we add the anomaly map as an additional channel to our target image. In the anomaly map, we set normal pixels to 0 and abnormal pixels to 1. This change implies that, in Eq. (3) and Eq. (4), all patches will have an additional channel. Third, we use the ground-truth anomaly map to enhance the weights defined in Eq. (2). The added synthetic anomalies do not necessarily yield motion gradients with high magnitude. Hence, it is possible to have low weights in Eq. (3) and Eq. (4) for patches that correspond to anomaly regions. This is not desirable if we want the model to detect anom" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15988" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "text", + "content": "lies. To this end, we propose to add the anomaly maps and the gradients together, before computing the weights as in Eq. (2). Formally, in Eq. (1), we replace " + }, + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{r}_i^{(t)}\\right\\}_{i=1}^n" + }, + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{r}_i^{(t)} + \\pmb{a}_i^{(t)}\\right\\}_{i=1}^n" + }, + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{a}_i^{(t)}\\right\\}_{i=1}^n" + }, + { + "bbox": [ + 47, + 72, + 287, + 145 + ], + "type": "text", + "content": " is the set of the patches extracted from the anomaly map." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 146, + 286, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 146, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 286, + 206 + ], + "type": "text", + "content": "Classification head. We further harness the synthetic anomalies to train a classification head applied on the final [CLS] token of the shared encoder. The head is trained to discriminate between frames with and without synthetic anomalies. This head is trained using binary cross-entropy:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 209, + 286, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 286, + 223 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} \\left(\\hat {\\boldsymbol {x}} _ {t}, \\theta_ {E}\\right) = - y _ {t} \\cdot \\log \\left(\\hat {y} _ {t}\\right) - \\left(1 - y _ {t}\\right) \\cdot \\log \\left(1 - \\hat {y} _ {t}\\right), \\tag {5}", + "image_path": "103be4f3361a9b43127485fa30c2dbf61144a1a3f4cbef0d530b4e4a5c609586.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "inline_equation", + "content": "y_{t} \\in \\{0,1\\}" + }, + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "text", + "content": " is 1 if the frame contains an anomaly and 0 otherwise, " + }, + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "text", + "content": " is the prediction, and " + }, + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\theta_E" + }, + { + "bbox": [ + 47, + 227, + 286, + 262 + ], + "type": "text", + "content": " represents the set of weights of the shared encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "text", + "content": "Inference. During inference, we pass each frame " + }, + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "text", + "content": " through both teacher and student models to obtain the reconstructed frames " + }, + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_t" + }, + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_t" + }, + { + "bbox": [ + 47, + 262, + 286, + 310 + ], + "type": "text", + "content": ", respectively. Then, we compute output pixel-level anomaly map as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 312, + 286, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 312, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 312, + 286, + 326 + ], + "type": "interline_equation", + "content": "\\boldsymbol {o} _ {t} = \\alpha \\cdot \\| \\boldsymbol {x} _ {t} - \\hat {\\boldsymbol {x}} _ {t} \\| _ {2} ^ {2} + \\beta \\cdot \\| \\hat {\\boldsymbol {x}} _ {t} - \\tilde {\\boldsymbol {x}} _ {t} \\| _ {2} ^ {2} + \\gamma \\cdot \\hat {y} _ {t}, \\tag {6}", + "image_path": "e160d3cf7c2852a388224ec164f18805ed2675ed2729b53ae9561013b19bf3c7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": " are hyperparameters that control the importance of the individual anomaly score components. Following [15, 35], we apply spatio-temporal 3D filtering to smooth the anomaly volumes. To obtain the frame-level anomaly scores, we keep the maximum value from each output map " + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{o}_t" + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": " and subsequently apply another temporal Gaussian filter to smooth the values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 418, + 127, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 127, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 127, + 431 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 434, + 163, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 163, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 163, + 448 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 450, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 689 + ], + "type": "text", + "content": "Data sets. We verify the performance of our method on four data sets for video anomaly detection: Avenue [51], ShanghaiTech [53], UBnormal [1] and UCSD Ped2 [55]. ShanghaiTech is the largest data set, with 270K frames for training and about 50K for testing. UBnormal is the second largest, with about 116K training frames and 93K testing frames. Avenue is a popular benchmark containing 15K frames for training and another 15K for testing. UCSD Ped2 holds a total of 4.5K frames, out of which 2.5K are used for training. UBnormal is a benchmark that uses an open set evaluation, where training and test anomalies belong to disjoint category sets. For the other three data sets, the training videos contain only normal events, and the test ones include both normal and abnormal scenarios. To augment the normal training videos, we sample abnormal events from the UBnormal data set [1]. UBnormal is a synthetic (virtual) data set containing anomalies simulated by video game characters, which alleviates the burden of collecting anomalies from the real world. The probability of augmenting a frame from Avenue, ShanghaiTech or UCSD Ped2 is 0.25." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Evaluation. We evaluate all models following recent related works [1, 27, 69], considering both micro and macro" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 533, + 547 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 533, + 547 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 533, + 547 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 533, + 547 + ], + "type": "table", + "html": "
TypeMethodAvenueShanghaiUBnormalPed2FPS
MicroMacroMicroMacroMicroMacroMicroMacro
Object-centric[10]91.692.583.890.562.186.5--20
[18]86.4-71.6---97.8--
[26]91.592.882.490.255.484.597.599.851
[27]92.390.482.789.361.385.698.799.724
[33]--85.9-71.8----
[36]87.490.478.784.9--94.397.8-
[40]80.2-73.7-50.7----
[49]89.993.574.283.2--99.3-12
[50]91.892.383.887.8-----
[54]+[10]91.692.483.690.6----20
[54]+[27]93.291.883.389.3----31
[54]+[49]89.593.675.283.8----10
[69]+[27]92.991.983.689.5----31
[69]+[49]90.992.275.583.7----10
[84]92.2-84.3---99.0-35
[95]89.6-74.8---97.3--
Frame or cube level[4]87.1-75.9---96.5--
[5]84.7-73.7---98.4--
[9]----68.580.3--37
[28]83.3-71.2---94.1-35
[37]88.9--------
[43]90.0-----96.6--
[47]85.181.772.880.6--95.4-28
[48]84.4-----87.5--
[54]+[47]89.184.874.683.3----26
[54]+[61]86.486.370.680.3----94
[57]86.9-----96.2--
[60]85.3-72.2---96.3-195
[61]82.886.868.379.7--97.0-101
[62]72.0-----88.3--
[63]87.2-----93.0--
[66]------93.5--
[67]------88.4--
[69]+[47]87.384.574.582.9----26
[69]+[61]84.888.669.880.2----95
[76]84.6--------
[77]---76.550.376.8--56
[78]89.6-74.7------
[80]85.1-73.0---96.3--
[82]--76.1------
[87]87.0-79.3------
[89]--80.4------
[90]86.6-----96.9--
[93]90.1-78.6-62.7----
[96]90.2-----97.3--
[100]--78.9------
[103]------91.0--
Ours91.390.979.184.758.581.495.498.41655
", + "image_path": "023de1281037ea04993729da30d19d250b91671fbfb8bb6cd0ab0aea2ffca105.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 550, + 545, + 625 + ], + "lines": [ + { + "bbox": [ + 305, + 550, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 545, + 625 + ], + "type": "text", + "content": "Table 1. Micro and macro AUC scores (in %) of several state-of-the-art frame-level, cube-level and object-level methods versus our self-distilled masked AE on Avenue, ShanghaiTech, UBnormal and UCSD Ped2. The top three scores for each category of methods are shown in red, green, and blue. All reported running times (including those of the baselines) are measured on a machine with an Nvidia GeForce GTX 3090 GPU with 24 GB of VRAM." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "content": "AUC metrics. The area under the ROC curve (AUC) expresses the overlap between the ground-truth frame-level annotations and the anomaly scores predicted by a model, at multiple thresholds. At a given threshold, a frame is labeled as abnormal if the predicted anomaly score is above the threshold. For the micro AUC, the test frames from all" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15989" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 276, + 209 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 276, + 209 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 276, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 276, + 209 + ], + "type": "image", + "image_path": "4756055ee566935d5e1262423f127167da719791fff36bc6342a50cd1feb2f7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 212, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 47, + 212, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 287, + 243 + ], + "type": "text", + "content": "Figure 4. Predictions for test video 04 from Avenue. The abnormal bounding boxes are given by the convex hull of the patches labeled as abnormal. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 260, + 286, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 286, + 307 + ], + "type": "text", + "content": "videos are concatenated before computing the AUC over all frames. For the macro AUC, the AUC of each test video is first computed, and the resulting AUC scores are averaged to obtain a single value." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": "Hyperparameters. The encoder module is formed of three CvT blocks, each with a projection size of 256 and four attention heads. The teacher decoder contains three CvT blocks, while the student decoder contains only one block. All decoder blocks have four attention heads and a projection dimension of 128. Since the data sets have different input resolutions and objects vary in size, we adapt the patch size to each data set. Thus, we set the patch size to " + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": " on Avenue, " + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": " on ShanghaiTech and UBnormal, and " + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": " on UCSD Ped2. Regardless of the data set, the teacher network is trained for 100 epochs, while the student is trained for 40 epochs. We optimize the networks with Adam [42], using a learning rate of " + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": " and mini-batches of 100 samples. The hyperparameters in Eq. (6) are set to " + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "inline_equation", + "content": "\\alpha = 0.4" + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "inline_equation", + "content": "\\beta = 0.3" + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "inline_equation", + "content": "\\gamma = 0.3" + }, + { + "bbox": [ + 47, + 308, + 286, + 486 + ], + "type": "text", + "content": ", for all data sets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 493, + 105, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 105, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 105, + 504 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 509, + 280, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 280, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 280, + 521 + ], + "type": "text", + "content": "We present our results in Table 1 and discuss them below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 641 + ], + "type": "text", + "content": "Results on Avenue. Our method obtains a micro AUC score of " + }, + { + "bbox": [ + 46, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "91.3\\%" + }, + { + "bbox": [ + 46, + 521, + 286, + 641 + ], + "type": "text", + "content": " on Avenue, being only " + }, + { + "bbox": [ + 46, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 46, + 521, + 286, + 641 + ], + "type": "text", + "content": " below the state-of-the-art object-centric method. Remarkably, in the category of frame-level methods, we reach the best micro and macro AUC scores. Taking into account that our method is much faster than all other methods, we consider that its performance is remarkable. In Figure 4, we illustrate the anomaly scores for test video 04 from Avenue. Here, our model is close to perfect, highlighting its ability to capture anomalies, such as people running." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "type": "text", + "content": "Results on ShanghaiTech. On ShanghaiTech, our method reaches the top macro AUC score and the third-best micro AUC score, when compared with other frame-level frameworks. Object-centric methods generally surpass frame-level methods, but the former methods have much lower processing speeds (see Figure 2)." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 158 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 158 + ], + "type": "table", + "html": "
Motion weightsSelf-distillationSynthetic dataAnomaly mapsClassif. headAvenueShanghai
AUC
MicroMacroMicroMacro
84.085.669.780.1
✓84.886.371.380.9
✓✓88.586.076.383.8
✓✓✓88.586.977.083.0
✓✓✓✓90.589.677.384.2
✓✓✓✓✓91.390.979.184.7
", + "image_path": "a5620e1311180f04a6ffc7974c630c7e64969b4dca5220d06ab0d393a0a02b7b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 194, + 545, + 263 + ], + "blocks": [ + { + "bbox": [ + 306, + 162, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 306, + 162, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 306, + 162, + 545, + 184 + ], + "type": "text", + "content": "Table 2. Impact of each novel component on the micro and macro AUC scores (in %), on Avenue [51] and ShanghaiTech [53]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 194, + 545, + 263 + ], + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 263 + ], + "type": "table", + "html": "
StrategyAvenueShanghai
AUC
MicroMacroMicroMacro
Teacher84.085.674.882.3
Teacher + Student85.485.875.182.1
Teacher + Teacher-Student Difference88.586.076.383.8
Teacher + Student + Teacher-Student Difference86.985.875.883.6
", + "image_path": "fe0eac877265251638db5585abb989587e31994feae634d90ac2c339a69c41f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 319, + 321, + 529, + 380 + ], + "blocks": [ + { + "bbox": [ + 305, + 266, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 309 + ], + "type": "text", + "content": "Table 3. Impact of strategies to combine the outputs of the teacher and student models on Avenue [51] and ShanghaiTech [53]. These results do not include the synthetic anomalies and the classification head." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 321, + 529, + 380 + ], + "lines": [ + { + "bbox": [ + 319, + 321, + 529, + 380 + ], + "spans": [ + { + "bbox": [ + 319, + 321, + 529, + 380 + ], + "type": "table", + "html": "
Data setMeasurePercentage of synthetic data
0%25%50%75%
AvenueMicro AUC88.591.390.689.9
Macro AUC86.090.989.487.7
ShanghaiMicro AUC76.379.177.977.7
Macro AUC83.884.784.484.1
", + "image_path": "b9f449051c8e619c72a857116019211bbec35abeceddfd8db0d0376332db5bdd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 384, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 306, + 384, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 545, + 406 + ], + "type": "text", + "content": "Table 4. Impact of varying the proportion of synthetic anomalies on the Avenue [51] and ShanghaiTech [53] data sets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 423, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 519 + ], + "type": "text", + "content": "Results on UBnormal. In terms of the micro AUC, the best frame-level method on UBnormal is TimeSformer [9], which benefits from large-scale pretraining. Notably, our method obtains a higher macro AUC than TimeSformer. The micro AUC of our method is fairly close to the micro AUC levels of the better object-centric approaches. In terms of speed, our method is significantly faster than all the other methods reporting results on UBnormal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": "Results on UCSD Ped2. Our framework obtains a micro AUC of " + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "95.4\\%" + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": ", being " + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": " below the state-of-the-art performance of Liu et al. [49]. While being slightly below in terms of micro AUC, our macro AUC score is remarkably on par with the object-centric method of Ionescu et al. [36], and only " + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": " below that of Georgescu et al. [26]. Considering that our FPS is more than 30 times higher compared with these methods [26, 36] on the same GPU, our framework provides a clearly superior accuracy-speed trade-off." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "text", + "content": "Ablation study. In Table 2, we illustrate the impact of each novel component on our model's performance. The first model is a vanilla masked AE, which obtains a rather low performance on both Avenue and ShanghaiTech. Each and every component contributes towards boosting the performance of the vanilla model. We observe that self-distillation gives the highest boost in terms of the micro AUC. How-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15990" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 68, + 487, + 241 + ], + "blocks": [ + { + "bbox": [ + 108, + 68, + 487, + 241 + ], + "lines": [ + { + "bbox": [ + 108, + 68, + 487, + 241 + ], + "spans": [ + { + "bbox": [ + 108, + 68, + 487, + 241 + ], + "type": "image", + "image_path": "f893c7daf0da06f86eed77efa86fcde04a3a3a9f321e024967f37e01b4f63141.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 244, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 547, + 268 + ], + "type": "text", + "content": "Figure 5. Examples of frames and anomaly maps reconstructed by our teacher. The first four columns correspond to abnormal examples from the Avenue and ShanghaiTech data sets, while the last column shows a normal example. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 288, + 327 + ], + "type": "text", + "content": "ever, to surpass the " + }, + { + "bbox": [ + 46, + 281, + 288, + 327 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 281, + 288, + 327 + ], + "type": "text", + "content": " milestone on Avenue, it is mandatory to introduce the prediction of anomaly maps in the learning task. An additional boost is given by our classification head." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": "The self-distillation procedure gives us a few possible strategies to combine the outputs of the teacher and the student. Hence, we investigate this aspect and report the results in Table 3. We observe that the best micro AUC is obtained when we combine the teacher reconstruction error with the teacher-student discrepancy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 474 + ], + "type": "text", + "content": "The proportion of synthetic examples per mini-batch is another aspect that can influence our model's performance. In Table 4, we report the micro and macro AUC scores for three possible augmentation levels. We observe that augmentation is always useful, but, for a better outcome, it requires a moderate percentage (25%)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 676 + ], + "type": "text", + "content": "Performance-speed trade-off. In Figure 2, we compare our model with several other methods in terms of the performance-speed trade-off. This comparison undoubtedly shows that our method reaches a far better processing speed, while achieving fairly good performance. To strengthen this observation, we also compare the methods in terms of GFLOPs and number of parameters in Table 5. We underline that the method of Gong et al. [28] might seem small in terms of the number of parameters, but it is slowed down by its input, which is formed of a cuboid constructed by stacking 16 consecutive frames. Moreover, the method relies on a memory module, where each memory slot records the features of one pixel in the activation maps. Although their method has twice as many parameters as our own, the large input volume and the sizable memory bank reduce the speed to 35 FPS. With an FPS of 1655, our method proves to be significantly lighter than all its competitors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Qualitative results. In Figure 5, we present the frames and anomaly maps reconstructed by the teacher in four abnormal scenarios from Avenue and ShanghaiTech. Moreover," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 331, + 277, + 518, + 346 + ], + "blocks": [ + { + "bbox": [ + 331, + 277, + 518, + 346 + ], + "lines": [ + { + "bbox": [ + 331, + 277, + 518, + 346 + ], + "spans": [ + { + "bbox": [ + 331, + 277, + 518, + 346 + ], + "type": "table", + "html": "
MethodGFLOPs ↓#Params (M) ↓FPS ↑
Georgescu et al. [26]107.96551
Georgescu et al. [27]121.66724
Liu et al. [49]179.532012
Park et al. [60]8464195
Gong et al. [28]55.2635
Ours0.831655
", + "image_path": "aa82f89c710e1c30f7b886acaeda1919d07cfe35474f4e129816cf06483b4e18.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 350, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 306, + 350, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 545, + 372 + ], + "type": "text", + "content": "Table 5. Comparing methods in terms of floating point operations (GFLOPs), number of parameters, and FPS." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": "in the fifth column, we illustrate the behavior of the teacher in a normal scenario. In all four abnormal cases, the reconstruction error is visibly higher for anomalous regions. This effect is mostly due to our training procedure based on synthetic data augmentation. The most obvious example is in the fourth column, where the bicycle seen in a pedestrian area is almost entirely removed from the output. Additionally, the predicted anomaly maps are well aligned with the ground-truth ones." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 504, + 379, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 379, + 516 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "In this work, we proposed a lightweight masked autoencoder (3M parameters, 0.8 GFLOPs) for video anomaly detection, which learns to reconstruct tokens with high motion gradients. Our framework is based on self-distillation, leveraging the discrepancy between teacher and student decoders for anomaly detection. Moreover, we boost the performance of our model by introducing a data augmentation technique based on overlapping synthetic anomalies on normal training data. Our highly efficient framework reached an unprecedented speed of 1655 FPS, with a minimal performance gap with respect to the state-of-the-art object-centric approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgments. Work supported by a grant of the Romanian Ministry of Education and Research, CNCS-UEFISCDI, project number PN-III-P2-2.1-PED-2021-0195, within PNCDI III." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15991" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1]Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. UBnormal: New Benchmark for Supervised Open-Set Video Anomaly Detection. In Proceedings of CVPR, pages 20143-20153, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 158, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 288, + 213 + ], + "type": "text", + "content": "[2] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust Real-Time Unusual Event Detection Using Multiple Fixed-Location Monitors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):555-560, 2008. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 214, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 214, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 57, + 214, + 288, + 247 + ], + "type": "text", + "content": "[3] Borislav Antic and Bjorn Ommer. Video parsing for abnormality detection. In Proceedings of ICCV, pages 2415-2422, 2011. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 248, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 248, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 57, + 248, + 288, + 282 + ], + "type": "text", + "content": "[4] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. In Proceedings of BMVC, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 284, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 284, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 57, + 284, + 288, + 327 + ], + "type": "text", + "content": "[5] Marcella Astrid, Muhammad Zaigham Zaheer, and Seung-Ik Lee. Synthetic Temporal Anomaly Guided End-to-End Video Anomaly Detection. In Proceedings of ICCVW, pages 207-214, 2021. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 328, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 328, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 57, + 328, + 288, + 360 + ], + "type": "text", + "content": "[6] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In Proceedings of NIPS, pages 2654-2662, 2014. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 363, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 363, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 363, + 288, + 407 + ], + "type": "text", + "content": "[7] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. MultiMAE: Multi-modal multi-task masked autoencoders. In Proceedings of ECCV, pages 348-367. Springer, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 408, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 408, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 57, + 408, + 288, + 462 + ], + "type": "text", + "content": "[8] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed Students: Student-Teacher Anomaly Detection With Discriminative Latent Embeddings. In Proceedings of CVPR, pages 4183–4192, 2020. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 464, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 464, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 288, + 497 + ], + "type": "text", + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is Space-Time Attention All You Need for Video Understanding? In Proceedings of ICML, 2021. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 499, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 575 + ], + "type": "text", + "content": "[10] Antonio Bărbălău, Radu Tudor Ionescu, Mariana-Iuliana Georgescu, Jacob Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. SSMTL++: Revisiting Self-Supervised Multi-Task Learning for Video Anomaly Detection. Computer Vision and Image Understanding, 229: 103656, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 577, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 577, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 53, + 577, + 288, + 620 + ], + "type": "text", + "content": "[11] Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. SdAE: Self-distillated Masked Autoencoder. In Proceedings of ECCV, pages 108-124, 2022. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 622, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 622, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 53, + 622, + 288, + 655 + ], + "type": "text", + "content": "[12] Hekai Cheng, Lu Yang, and Zulong Liu. Relation-based knowledge distillation for anomaly detection. In Proceedings of PRCV, pages 105-116, 2021. 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 657, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 657, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 53, + 657, + 287, + 700 + ], + "type": "text", + "content": "[13] Kai-Wen Cheng, Yie-Tarng Chen, and Wen-Hsien Fang. Video anomaly detection and localization using hierarchical feature representation and Gaussian process regression. In Proceedings of CVPR, pages 2909-2917, 2015. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 702, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 702, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 702, + 288, + 713 + ], + "type": "text", + "content": "[14] Y. Cong, J. Yuan, and J. Liu. Sparse reconstruction cost for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "abnormal event detection. In Proceedings of CVPR, pages 3449-3456, 2011. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "text", + "content": "[15] Allison Del Giorno, J. Andrew Bagnell, and Martial Hebert. A Discriminative Framework for Anomaly Detection in Large Videos. In Proceedings of ECCV, pages 334-349, 2016. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 545, + 174 + ], + "type": "text", + "content": "[16] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of CVPR, pages 9737-9746, 2022. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "type": "text", + "content": "[17] Fei Dong, Yu Zhang, and Xiushan Nie. Dual Discriminator Generative Adversarial Network for Video Anomaly Detection. IEEE Access, 8:88170-88176, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 242 + ], + "type": "text", + "content": "[18] Keval Doshi and Yasin Yilmaz. Any-Shot Sequential Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 934-935, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 243, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 243, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 545, + 276 + ], + "type": "text", + "content": "[19] Keval Doshi and Yasin Yilmaz. Continual Learning for Anomaly Detection in Surveillance Videos. In Proceedings of CVPRW, pages 254-255, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 277, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 277, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 312, + 277, + 545, + 342 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of ICLR, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 344, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 344, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 312, + 344, + 545, + 376 + ], + "type": "text", + "content": "[21] Jayanta K. Dutta and Bonny Banerjee. Online Detection of Abnormal Events Using Incremental Coding Length. In Proceedings of AAAI, pages 3755-3761, 2015. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 378, + 545, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 378, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 312, + 378, + 545, + 433 + ], + "type": "text", + "content": "[22] Yaxiang Fan, Gongjian Wen, Deren Li, Shaohua Qiu, Martin D. Levine, and Fei Xiao. Video anomaly detection and localization via Gaussian Mixture Fully Convolutional Variational Autoencoder. Computer Vision and Image Understanding, 195:102920, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 545, + 477 + ], + "type": "text", + "content": "[23] Ye Fei, Chaoqin Huang, Cao Jinkun, Maosen Li, Ya Zhang, and Cewu Lu. Attribute Restoration Framework for Anomaly Detection. IEEE Transactions on Multimedia, 24: 116-127, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 479, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 479, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 545, + 511 + ], + "type": "text", + "content": "[24] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked Autoencoders As Spatiotemporal Learners. In Proceedings of NeurIPS, 2022. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 513, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 513, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 312, + 513, + 545, + 545 + ], + "type": "text", + "content": "[25] Yachuang Feng, Yuan Yuan, and Xiaoqiang Lu. Learning deep event models for crowd anomaly detection. Neurocomputing, 219:548-556, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 547, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 547, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 547, + 547, + 601 + ], + "type": "text", + "content": "[26] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly Detection in Video via Self-Supervised and Multi-Task Learning. In Proceedings of CVPR, pages 12742–12752, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 602, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 602, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 602, + 547, + 667 + ], + "type": "text", + "content": "[27] Mariana Iuliana Georgescu, Radu Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A Background-Agnostic Framework with Adversarial Training for Abnormal Event Detection in Video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4505-4523, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "type": "text", + "content": "[28] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton Van Den Hengel. Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15992" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "Anomaly Detection. In Proceedings of ICCV, pages 1705-1714, 2019. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 140 + ], + "type": "text", + "content": "[29] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In Proceedings of CVPR, pages 733-742, 2016. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 288, + 185 + ], + "type": "text", + "content": "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Proceedings of CVPR, pages 16000-16009, 2022. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 229 + ], + "type": "text", + "content": "[31] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint Detection and Recounting of Abnormal Events by Learning Deep Generic Knowledge. In Proceedings of ICCV, pages 3639-3647, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "text", + "content": "[32] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the Knowledge in a Neural Network. In Proceedings of NIPS Deep Learning and Representation Learning Workshop, 2014. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "type": "text", + "content": "[33] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In Proceedings of ICCV, pages 13545-13554, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "type": "text", + "content": "[34] Chaoqin Huang, Qinwei Xu, Yanfeng Wang, Yu Wang, and Ya Zhang. Self-Supervised Masking for Unsupervised Anomaly Detection and Localization. IEEE Transactions on Multimedia, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 355, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 355, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 53, + 355, + 288, + 398 + ], + "type": "text", + "content": "[35] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In Proceedings of ICCV, pages 2895-2903, 2017. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 454 + ], + "type": "text", + "content": "[36] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-Centric AutoEncoders and Dummy Anomalies for Abnormal Event Detection in Video. In Proceedings of CVPR, pages 7842-7851, 2019. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 288, + 499 + ], + "type": "text", + "content": "[37] Radu Tudor Ionescu, Sorina Smeureanu, Marius Popescu, and Bogdan Alexe. Detecting abnormal events in video using Narrowed Normality Clusters. In Proceedings of WACV, pages 1951-1960, 2019. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 501, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 288, + 544 + ], + "type": "text", + "content": "[38] Xiangli Ji, Bairong Li, and Yuesheng Zhu. TAM-Net: Temporal Enhanced Appearance-to-Motion Generative Network for Video Anomaly Detection. In Proceedings of IJCNN, pages 1–8, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 545, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 599 + ], + "type": "text", + "content": "[39] Jielin Jiang, Jiale Zhu, Muhammad Bilal, Yan Cui, Neeraj Kumar, Ruihan Dou, Feng Su, and Xiaolong Xu. Masked Swin Transformer Unet for Industrial Anomaly Detection. IEEE Transactions on Industrial Informatics, 19(2):2200-2209, 2022. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "text", + "content": "[40] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. BiPOCO: Bi-Directional Trajectory Prediction with Pose Constraints for Pedestrian Anomaly Detection. In Proceedings of SL4AD, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 689 + ], + "type": "text", + "content": "[41] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: A space-time MRF for detecting abnormal activities with incremental updates. In Proceedings of CVPR, pages 2921-2928, 2009. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "text", + "content": "[42] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of ICLR, 2015. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 116 + ], + "type": "text", + "content": "[43] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. BMAN: Bidirectional Multi-Scale Aggregation Networks for Abnormal Event Detection. IEEE Transactions on Image Processing, 29:2395-2408, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "type": "text", + "content": "[44] Guoqiu Li, Guanxiong Cai, Xingyu Zeng, and Rui Zhao. Scale-Aware Spatio-Temporal Relation Learning for Video Anomaly Detection. In Proceedings of ECCV, pages 333-350, 2022. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "text", + "content": "[45] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(1):18-32, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "text", + "content": "[46] Zhenyu Li, Ning Li, Kaitao Jiang, Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Superpixel Masking and Inpainting for Self-Supervised Anomaly Detection. In Proceedings of BMVC, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "type": "text", + "content": "[47] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future Frame Prediction for Anomaly Detection - A New Baseline. In Proceedings of CVPR, pages 6536-6545, 2018. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 545, + 327 + ], + "type": "text", + "content": "[48] Yusha Liu, Chun-Liang Li, and Barnabaás Póczos. Classifier Two-Sample Test for Video Anomaly Detections. In Proceedings of BMVC, 2018. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 545, + 382 + ], + "type": "text", + "content": "[49] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction. In Proceedings of ICCV, pages 13588-13597, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 545, + 426 + ], + "type": "text", + "content": "[50] Zuhao Liu, Xiao-Ming Wu, Dian Zheng, Kun-Yu Lin, and Wei-Shi Zheng. Generating Anomalies for Video Anomaly Detection With Prompt-Based Feature Mapping. In Proceedings of CVPR, pages 24500-24510, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 427, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 427, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 427, + 545, + 459 + ], + "type": "text", + "content": "[51] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal Event Detection at 150 FPS in MATLAB. In Proceedings of ICCV, pages 2720-2727, 2013. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 492 + ], + "type": "text", + "content": "[52] Yiwei Lu, Frank Yu, Mahesh Kumar, Krishna Reddy, and Yang Wang. Few-Shot Scene-Adaptive Anomaly Detection. In Proceedings of ECCV, pages 125–141, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 545, + 536 + ], + "type": "text", + "content": "[53] Weixin Luo, Wen Liu, and Shenghua Gao. A Revisit of Sparse Coding Based Anomaly Detection in Stacked RNN Framework. In Proceedings of ICCV, pages 341-349, 2017. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 537, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 601 + ], + "type": "text", + "content": "[54] Neelu Madan, Nicolae-Catalin Ristea, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised masked convolutional transformer block for anomaly detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):525-542, 2024. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 603, + 545, + 635 + ], + "type": "text", + "content": "[55] Vijay Mahadevan, Wei-Xin LI, Viral Bhalodia, and Nuno Vasconcelos. Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 1975-1981, 2010. 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 669 + ], + "type": "text", + "content": "[56] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of CVPR, pages 935-942, 2009. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "text", + "content": "[57] Trong-Nguyen Nguyen and Jean Meunier. Anomaly Detection in Video Sequence With Appearance-Motion Correspondence. In Proceedings of ICCV, pages 1273-1283, 2019. 1, 3, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15993" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "text", + "content": "[58] Guansong Pang, Cheng Yan, Chunhua Shen, Anton van den Hengel, and Xiao Bai. Self-trained Deep Ordinal Regression for End-to-End Video Anomaly Detection. In Proceedings of CVPR, pages 12173–12182, 2020. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 152 + ], + "type": "text", + "content": "[59] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM Computing Surveys, 54(2):1-38, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 196 + ], + "type": "text", + "content": "[60] Chaewon Park, MyeongAh Cho, Minhyeok Lee, and Sangyoun Lee. FastAno: Fast anomaly detection via spatiotemporal patch transformation. In Proceedings of WACV, pages 2249-2259, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "type": "text", + "content": "[61] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection. In Proceedings of CVPR, pages 14372-14381, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 286 + ], + "type": "text", + "content": "[62] Bharathkumar Ramachandra and Michael Jones. Street Scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of WACV, pages 2569-2578, 2020. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 288, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 288, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 288, + 288, + 331 + ], + "type": "text", + "content": "[63] Bharathkumar Ramachandra, Michael Jones, and Ranga Vatsavai. Learning a distance function with a Siamese network to localize anomalies in videos. In Proceedings of WACV, pages 2598-2607, 2020. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 333, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 333, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 333, + 288, + 376 + ], + "type": "text", + "content": "[64] Bharathkumar Ramachandra, Michael Jones, and Ranga Raju Vatsavai. Perceptual metric learning for video anomaly detection. Machine Vision and Applications, 32:1432-1769, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 378, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 378, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 378, + 287, + 432 + ], + "type": "text", + "content": "[65] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. A Survey of Single-Scene Video Anomaly Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2293-2312, 2022. 1, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 434, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 287, + 488 + ], + "type": "text", + "content": "[66] Mahdyar Ravanbakhsh, Moin Nabi, Enver Sangineto, Lucio Marcenaro, Carlo Regazzoni, and Nicu Sebe. Abnormal Event Detection in Videos using Generative Adversarial Nets. In Proceedings of ICIP, pages 1577-1581, 2017. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 287, + 544 + ], + "type": "text", + "content": "[67] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-Play CNN for Crowd Motion Analysis: An Application in Abnormal Event Detection. In Proceedings of WACV, pages 1689–1698, 2018. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 546, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 287, + 600 + ], + "type": "text", + "content": "[68] Huamin Ren, Weifeng Liu, Soren Ingvor Olsen, Sergio Escalera, and Thomas B. Moeslund. Unsupervised Behavior-Specific Dictionary Learning for Abnormal Event Detection. In Proceedings of BMVC, pages 28.1-28.13, 2015. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 657 + ], + "type": "text", + "content": "[69] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection. In Proceedings of CVPR, pages 13576-13586, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 287, + 713 + ], + "type": "text", + "content": "[70] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, Karpathy A., A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "type": "text", + "content": "[71] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, and Reinhard Klette. Deep-Cascade: Cascading 3D Deep Neural Networks for Fast Anomaly Detection and Localization in Crowded Scenes. IEEE Transactions on Image Processing, 26(4):1992-2004, 2017. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 184 + ], + "type": "text", + "content": "[72] Mohammad Sabokrou, Mohsen Fayyaz, Mahmood Fathy, Zahra Moayed, and Reinhard Klette. Deep-Anomaly: Fully Convolutional Neural Network for Fast Anomaly Detection in Crowded Scenes. Computer Vision and Image Understanding, 172:88-97, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 185, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 545, + 217 + ], + "type": "text", + "content": "[73] Babak Saleh, Ali Farhadi, and Ahmed Elgammal. Object-Centric Anomaly Detection by Attribute-Based Reasoning. In Proceedings of CVPR, pages 787-794, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 220, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 545, + 273 + ], + "type": "text", + "content": "[74] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution Knowledge Distillation for Anomaly Detection. In Proceedings of CVPR, pages 14902-14912, 2021. 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 275, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 275, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 275, + 547, + 308 + ], + "type": "text", + "content": "[75] Venkatesh Saligrama and Zhu Chen. Video anomaly detection based on local statistical aggregates. In Proceedings of CVPR, pages 2112-2119, 2012. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 309, + 545, + 353 + ], + "type": "text", + "content": "[76] Sorina Smeureanu, Radu Tudor Ionescu, Marius Popescu, and Bogdan Alexe. Deep Appearance Features for Abnormal Behavior Detection in Video. In Proceedings of ICIAP, pages 779-789, 2017. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 387 + ], + "type": "text", + "content": "[77] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In Proceedings of CVPR, pages 6479–6488, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 545, + 430 + ], + "type": "text", + "content": "[78] Che Sun, Yunde Jia, Yao Hu, and Yuwei Wu. Scene-Aware Context Reasoning for Unsupervised Abnormal Event Detection in Videos. In Proceedings of ACMMM, pages 184-192, 2020. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 433, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 433, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 312, + 433, + 545, + 475 + ], + "type": "text", + "content": "[79] Qianru Sun, Hong Liu, and Tatsuya Harada. Online growing neural gas for anomaly detection in changing surveillance scenes. Pattern Recognition, 64(C):187-201, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 521 + ], + "type": "text", + "content": "[80] Yao Tang, Lin Zhao, Shanshan Zhang, Chen Gong, Guangyu Li, and Jian Yang. Integrating prediction and reconstruction for anomaly detection. Pattern Recognition Letters, 129:123-130, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 545, + 555 + ], + "type": "text", + "content": "[81] Hanh T.M. Tran and David Hogg. Anomaly Detection using a Convolutional Winner-Take-All Autoencoder. In Proceedings of BMVC, 2017. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "type": "text", + "content": "[82] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised video anomaly detection. In Proceedings of ICIP, pages 2540-2544, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 602, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 602, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 602, + 545, + 644 + ], + "type": "text", + "content": "[83] Hung Vu, Tu Dinh Nguyen, Trung Le, Wei Luo, and Dinh Phung. Robust Anomaly Detection in Videos Using Multilevel Representations. In Proceedings of AAAI, pages 5216-5223, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 690 + ], + "type": "text", + "content": "[84] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Proceedings of ECCV, pages 494-511, 2022. 2, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "text", + "content": "[85] L. Wang, F. Zhou, Z. Li, W. Zuo, and H. Tan. Abnormal Event Detection in Videos Using Hybrid Spatio-Temporal" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15994" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 95 + ], + "type": "text", + "content": "Autoencoder. In Proceedings of ICIP, pages 2276-2280, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 129 + ], + "type": "text", + "content": "[86] Xusheng Wang, Mingtao Pei, and Zhengang Nie. Self-trained video anomaly detection based on teacher-student model. In Proceedings of MLSP, pages 1-6, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 130, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 130, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 288, + 163 + ], + "type": "text", + "content": "[87] Ziming Wang, Yuexian Zou, and Zeming Zhang. Cluster Attention Contrast for Video Anomaly Detection. In Proceedings of ACMMM, pages 2463-2471, 2020. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 164, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 164, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 53, + 164, + 288, + 208 + ], + "type": "text", + "content": "[88] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. CvT: Introducing Convolutions to Vision Transformers. In Proceedings of ICCV, pages 22-31, 2021. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 253 + ], + "type": "text", + "content": "[89] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In Proceedings of ECCV, pages 729-745, 2022. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 254, + 288, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 254, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 53, + 254, + 288, + 298 + ], + "type": "text", + "content": "[90] Peng Wu, Jing Liu, and Fang Shen. A Deep One-Class Neural Network for Anomalous Event Detection in Complex Scenes. IEEE Transactions on Neural Networks and Learning Systems, 31(7):2609-2622, 2019. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 299, + 288, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 288, + 343 + ], + "type": "text", + "content": "[91] Shandong Wu, Brian E. Moore, and Mubarak Shah. Chaotic Invariants of Lagrangian Particle Trajectories for Anomaly Detection in Crowded Scenes. In Proceedings of CVPR, pages 2054-2060, 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 344, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 288, + 388 + ], + "type": "text", + "content": "[92] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting Anomalous Events in Videos by Learning Deep Representations of Appearance and Motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "type": "text", + "content": "[93] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature Prediction Diffusion Model for Video Anomaly Detection. In Proceedings of ICCV, pages 5527-5537, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 434, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 488 + ], + "type": "text", + "content": "[94] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "type": "text", + "content": "[95] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze Test Helps: Effective Video Anomaly Detection via Learning to Complete Video Events. In Proceedings of ACMMM, pages 583-591, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 545, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 600 + ], + "type": "text", + "content": "[96] Jongmin Yu, Younkwan Lee, Kin Choong Yow, Moongu Jeon, and Witold Pedrycz. Abnormal event detection and localization via adversarial event prediction. IEEE Transactions on Neural Networks and Learning Systems, 33(8): 3572-3586, 2022. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "text", + "content": "[97] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proceedings of CVPR, pages 19313–19322, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "text", + "content": "[98] Muhammad Zaigham Zaheer, Jin-ha Lee, Marcella Astrid, and Seung-Ik Lee. Old is Gold: Redefining the Adversarily Learned One-Class Classifier Training Paradigm. In Proceedings of CVPR, pages 14183-14193, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "text", + "content": "[99] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. CLAWS: Clustering Assisted" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 545, + 385 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "type": "text", + "content": "Weakly Supervised Learning with Normalcy Suppression for Anomalous Event Detection. In Proceedings of ECCV, pages 358-376, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "text", + "content": "[100] Muhammad Zaigham Zaheer, Arif Mahmood, Haris M. Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative Cooperative Learning for Unsupervised Video Anomaly Detection. In Proceedings of CVPR, pages 14744-14754, 2022. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[101] Linfeng Zhang, Chenglong Bao, and Kaisheng Ma. Self-Distillation: Towards Efficient and Compact Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4388-4403, 2022. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 207, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 252 + ], + "type": "text", + "content": "[102] Xinfeng Zhang, Su Yang, Jiulong Zhang, and Weishan Zhang. Video Anomaly Detection and Localization using Motion-field Shape Description and Homogeneity Testing. Pattern Recognition, page 107394, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "text", + "content": "[103] Ying Zhang, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Shun Sakai. Video anomaly detection based on locality sensitive hashing filters. Pattern Recognition, 59:302-311, 2016. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 297, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 331 + ], + "type": "text", + "content": "[104] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online Detection of Unusual Events in Videos via Dynamic Sparse Coding. In Proceedings of CVPR, pages 3313-3320, 2011." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "text", + "content": "[105] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph Convolutional Label Noise Cleaner: Train a Plug-And-Play Action Classifier for Anomaly Detection. In Proceedings of CVPR, pages 1237–1246, 2019. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15995" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_content_list.json b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0d059536f3cda4614d2404cec14672d4ec988f52 --- /dev/null +++ b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_content_list.json @@ -0,0 +1,1419 @@ +[ + { + "type": "text", + "text": "Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations", + "text_level": 1, + "bbox": [ + 94, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kewei Wang $^{1,2}$ , Yizheng Wu $^{1,2}$ , Jun CEN $^{2}$ , Zhiyu Pan $^{1}$ , Xingyi Li $^{1,2}$ , Zhe Wang $^{3}$ , Zhiguo Cao $^{1}$ , Guosheng Lin $^{2*}$ $^{1}$ School of AIA, Huazhong University of Science and Technology \n $^{2}$ S-Lab, Nanyang Technological University $^{3}$ SenseTime Research", + "bbox": [ + 86, + 200, + 965, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The perception of motion behavior in a dynamic environment holds significant importance for autonomous driving systems, wherein class-agnostic motion prediction methods directly predict the motion of the entire point cloud. While most existing methods rely on fully-supervised learning, the manual labeling of point cloud data is laborious and time-consuming. Therefore, several annotation-efficient methods have been proposed to address this challenge. Although effective, these methods rely on weak annotations or additional multi-modal data like images, and the potential benefits inherent in the point cloud sequence are still underexplored. To this end, we explore the feasibility of self-supervised motion prediction with only unlabeled Li-DAR point clouds. Initially, we employ an optimal transport solver to establish coarse correspondences between current and future point clouds as the coarse pseudo motion labels. Training models directly using such coarse labels leads to noticeable spatial and temporal prediction inconsistencies. To mitigate these issues, we introduce three simple spatial and temporal regularization losses, which facilitate the self-supervised training process effectively. Experimental results demonstrate the significant superiority of our approach over the state-of-the-art self-supervised methods. Code will be available at https://github.com/kwwcv/SelfMotion.", + "bbox": [ + 76, + 323, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motion perception provides key information for autonomous driving. Classical approaches formulate the motion perception task as object-level trajectory prediction [3, 5, 7, 17, 34]. However, these approaches may fail if agents are not successfully detected by the detectors, especially when encountering unseen categories in open-set scenarios [31]. To compensate for trajectory prediction and", + "bbox": [ + 75, + 757, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce930b58cf4438ef67882a4cc13a3aee263dac2a39d9c54cd32655cedbb0fed1.jpg", + "image_caption": [ + "(a) Static" + ], + "image_footnote": [], + "bbox": [ + 504, + 292, + 624, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3861d21624ce41455162fd11497d55066089487ded343635fecb8763b38fb167.jpg", + "image_caption": [ + "(b) Slow" + ], + "image_footnote": [], + "bbox": [ + 635, + 294, + 754, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08ee62acdb6ef05645e69957b99eef61d44dd42491849ff918e08dceb654dd3c.jpg", + "image_caption": [ + "(c) Fast", + "Figure 1. Performance comparison over static, slow, and fast speed levels between self-supervised PillarMotion [19] and our approach on the nuScenes dataset. The dashed line represents the performance of fully-supervised MotionNet [31]. Our proposed self-supervised approach outperforms the PillarMotion which uses additional image data by a large margin and substantially narrows the performance gap with fully-supervised results." + ], + "image_footnote": [], + "bbox": [ + 767, + 294, + 883, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "provide backup, class-agnostic motion prediction was studied [15, 25, 29-31]. Currently, these approaches predict the motions of the entire environment from point clouds directly instead of for each agent.", + "bbox": [ + 496, + 536, + 893, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While fully-supervised class-agnostic motion prediction methods have shown success, they typically require a vast amount of annotated point cloud data, which can be both costly and time-consuming to acquire. To overcome this limitation, WeakMotionNet [15] proposes a weakly supervised approach using only a limited number of foreground-background semantic annotations. WeakMotionNet also introduces a consistency-aware chamfer loss that uses the difference between correspondences from different frames as loss weights. Although effective, it requires establishing correspondences multiple times and is still based on matching results between individual frames, rather than taking advantage of the temporal information within the input sequences. Without using any annotations, PillarMotion [19] introduces a self-supervised training approach that incorporates regularization of 2D and 3D flows derived from paired image and point cloud data. However, PillarMotion relies on multi-modality pairs to regularize motion learning, making the training process less flexible, and a considerable performance gap still exists between self-supervised", + "bbox": [ + 496, + 598, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. gslin@ntu.edu.sg, {wangkewei, zgcao} @hust.edu.cn,", + "bbox": [ + 76, + 875, + 362, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14638", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and fully-supervised methods. Therefore, we explore the spatial and temporal information from the point cloud sequences themselves to train a better class-agnostic motion prediction model with only unlabeled LiDAR data in this paper.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach primarily involves pseudo label generation and spatial and temporal regularizations. To generate pseudo motion labels, we first solve a matching problem to find correspondence between current and future point clouds. However, the correctness of the matching correspondence cannot be guaranteed due to the occlusion, background artifacts, noise, etc. Self-supervised learning from these coarse matching correspondences (pseudo labels) leads to two major failure cases: (1) predicted motions from the same rigid object are spatially diverse; (2) the background cells that should be static are predicted with large motions. To alleviate these issues, we introduce three simple but effective spatial and temporal regularizations from the point cloud motion prediction itself to facilitate self-supervised learning. Specifically, cluster consistency regularization, backward consistency regularization, and forward consistency regularization are proposed to encourage spatial consistency, penalize incorrect predictions (e.g., large motion for static cells), and smooth the prediction learned from pseudo labels of different time horizons, respectively.", + "bbox": [ + 75, + 167, + 472, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments demonstrate that our proposed spatial and temporal consistency regularization terms work effectively together, significantly boosting the performance of self-supervised motion prediction. The proposed approach turns out with excellent performance, as illustrated in Fig. 1. On the nuScenes dataset, our approach outperforms previous methods by a substantial margin. Our contributions can be summarized as follows:", + "bbox": [ + 75, + 484, + 468, + 602 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel approach for self-supervised class-diagnostic motion prediction that incorporates an optimal transport solver and three simple yet effective spatial and temporal regularization losses.", + "- Our approach is model-independent, and does not require any additional models or multi-modal data.", + "- Our method surpasses the state-of-the-art (SOTA) self-supervised methods by a large margin and significantly reduces the performance gap with the fully-supervised methods." + ], + "bbox": [ + 76, + 604, + 468, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 768, + 225, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Class-agnostic motion prediction. Classical motion prediction approaches aim to predict the future trajectories of agents based on past observations. The entire system typically involves object detection [11, 20, 35], tracking, and forecasting [3, 5, 7, 17, 33, 34]. However, the trajectory prediction relies on the object detector which may fail to detect objects with unknown classes in open-set scenarios [31]. To", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enhance system redundancy, especially in cases where objects may be not successfully detected and potentially lead to a failure in object-level trajectory forecasting [31], class-agnostic motion prediction methods predict the motion of entire scenes directly instead of individually for each agent. These methods represent the environment state based on BEV maps discretized from point clouds and aim to predict the 2D displacement vector for each BEV cell along the ground plane. MotionNet [31] is one of the pioneers in this task, proposing to perform joint perception and motion prediction based on a spatial-temporal pyramid network. LSTM-ED [25] introduce convolutional LSTM [26] to aggregate temporal context. BE-STI [30] leverages semantic cues by training an additional semantic decoder to guide motion prediction.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the reliance on manual annotations, WeakMotionNet [15] proposes to train the MotionNet in a weakly-supervised manner with limited foreground and background semantic annotations. PillarMotion [19] proposes to train the model in a self-supervised manner without any annotations by cross-sensor motion regularization. Although effective, a substantial performance gap persists between the self-supervised PillarMotion and the fully-supervised model even using additional image data. Therefore, an open question remains about whether we can more efficiently train a self-supervised motion prediction model with higher performance.", + "bbox": [ + 496, + 316, + 890, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene flow estimation. Scene flow estimation [1, 8, 18, 28, 32] from point clouds is an alternative way to understand the dense class-agnostic 3D motion field. In self-supervised scene flow estimation, several methods rely on establishing point correspondences between source and target point clouds through point matching. They treat the 3D coordinate difference between each matching pair as a pseudo scene flow label [10, 13, 14, 21, 24, 32]. Many methods also utilize the chamfer distance loss [9] to drive self-supervised scene flow learning [6, 10, 16, 23, 32].", + "bbox": [ + 496, + 500, + 890, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While similar to the class-agnostic motion prediction task, scene flow estimation majorly differs in the following ways: (1) The objectiveness of the two tasks is different [15, 31]. Scene flow estimation methods focus on estimating the motion flow between two observed point clouds, while motion prediction methods aim to forecast future displacements based on past observations; (2) Estimating dense 3D flow is time-consuming, which is not feasible for self-driving systems [15]. Furthermore, directly applying scene flow estimation on real LiDAR point clouds poses challenges due to the lack of one-to-one correspondences [19, 30].", + "bbox": [ + 496, + 650, + 890, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 844, + 589, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we begin by introducing the problem formulation of class-agnostic motion prediction. Subsequently,", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14639", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0193a72f083ec9645e4381c1c8f3cfe0aff63792de326a79e20cf656354cdba3.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed approach. Without ground truth labels, we first generate pseudo labels by matching. We then introduce cluster, forward, and backward regularization losses to facilitate self-supervised motion learning." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 880, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we delve into a comprehensive presentation and discussion of each component comprising our approach. The overview of our approach is illustrated in Fig. 2.", + "bbox": [ + 75, + 363, + 468, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 417, + 279, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a temporal sequence of LiDAR point cloud frames, past frames are first synchronized to the current coordinate system [31]. We denote synchronized point cloud captured at time $t$ as $P^t$ . Following [31], we discretized $P^t$ into dense voxels $V^t \\in \\{0,1\\}^{H \\times W \\times C}$ , where 0 indicates the voxel is empty, 1 indicates the voxel is occupied by at least 1 point, and $H, W, C$ are the voxel numbers along $X, Y, Z$ axis respectively. Then we represent the 3D voxel lattice as a 2D pseudo-image with the $C$ corresponding to the image channel and view $V^t$ as a virtual BEV map. Then the motion field $M \\in \\mathbb{R}^{H \\times W \\times 2}$ in the BEV map is defined as the 2D displacement of each grid cell to the future timestamp. Taking sequence $\\vec{\\mathcal{V}} = \\{V^t\\}_{t=1}^T$ as input, the task aims to predict the motion fields $\\vec{\\mathcal{M}} = \\{M^T \\rightarrow T + t\\}_{t=1}^{T'}$ , where $T$ and $T'$ are the number of past and future timestamps, respectively.", + "bbox": [ + 75, + 440, + 473, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 76, + 696, + 187, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, our approach mainly involves pseudo label generation and self-supervised training. To generate pseudo labels, we initially solve the matching problem between two consecutive frames. Firstly, we remove the majority of ground plane points from point clouds by ground segmentation algorithm [12] to reduce noise for better matching. Subsequently, we feed the raw BEV map sequence into MotionNet [31] and train the model using the generated pseudo labels through supervised loss $\\mathcal{L}_{sup}$ . As the quality of pseudo labels is low at the initial stage of training, we introduce three spatial and temporal consistency regularization losses, named cluster consistency loss", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{L}_c$ , forward consistency loss $\\mathcal{L}_f$ , and backward consistency loss $\\mathcal{L}_b$ , to facilitate the self-supervised learning.", + "bbox": [ + 498, + 363, + 890, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Pseudo Label Generation", + "text_level": 1, + "bbox": [ + 500, + 402, + 732, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike fully-supervised learning, where ground truth labels are available, self-supervised methods learn from data itself without annotations. In this condition, one alternative approach is to generate pseudo labels to provide the necessary supervision for training. Considering motion can be defined as the variation in coordinates between correspondent points in two consecutive frames, the process of generating pseudo labels can be viewed as a matching problem. We denote $B^{T} = \\{b_{i}^{T}\\in \\mathbb{R}^{2}\\}_{i = 1}^{N_{T}}$ as the 2D coordinates of BEV cells on the current virtual BEV map $V^T$ , where $N_{T}$ is the number of non-empty cells. Then we perform matching between source $B^{T}$ and target $B^{T + t} = \\{b_j^{T + t}\\in \\mathbb{R}^2\\}_{j = 1}^{N_{T + t}}$ . Ideally, if there exists one-to-one correspondence between $B^{T}$ and $B^{T + t}$ , $B^{T}$ can be projected into $B^{T + t}$ and fully occupy $B^{T + t}$ :", + "bbox": [ + 496, + 425, + 892, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nB ^ {T} + M = \\pi B ^ {T + t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 662, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pi \\in \\{0,1\\}^{N_T\\times N_{T + t}}$ is a permutation matrix indicating correspondence between $B^T$ and $B^{T + t}$ , and $M$ is the motion flow. The optimal soft permutation matrix $\\pi^{*}\\in [0,1]^{N_{T}\\times N_{T + t}}$ can be computed by solving an optimal transport problem:", + "bbox": [ + 496, + 690, + 890, + 767 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\pi^ {*} = \\underset {\\pi} {\\arg \\min } \\sum_ {i, j} C _ {i j} \\pi_ {i j} \\\\ s. t. \\quad \\pi \\mathbf {1} _ {N _ {T + t}} = \\frac {1}{N _ {T}} \\mathbf {1} _ {N _ {T}}, \\pi^ {T} \\mathbf {1} _ {N _ {T}} = \\frac {1}{N _ {T + t}} \\mathbf {1} _ {N _ {T + t}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 777, + 856, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_{N_T} \\in \\mathbb{R}^{N_T}$ and $\\mathbf{1}_{N_{T + t}} \\in \\mathbb{R}^{N_{T + t}}$ are vectors with all elements are 1. $C$ is the transport cost matrix and $\\pi^T$ is the transpose of $\\pi$ . This problem can be solved by the Sinkhon", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14640", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2adb0d29d71a5795688de1f5d6d89a768ff3f99602cfa1f4a5a880c11068717a.jpg", + "image_caption": [ + "Figure 3. Prediction results with inconsistency. (a) Ground truth and pseudo labels. (b) Predictions from the same object are inconsistent. (c) Predictions for consecutive future timestamps are inconsistent. (d) Forward and backward predictions (e.g., $M^{T \\to T + 1}$ and $M^{T \\to T - 1}$ ) are inconsistent. Inconsistency regions are highlighted by red circles. The blue arrow denotes the future displacement (motion)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "algorithm [4]. We use a pre-warping operation to warp the source BEV cells $B^{T}$ by the predicted motion $M^{T\\rightarrow T + t}$ , and find the correspondence between the warped BEV cells $\\hat{B}^T = B^T +M^{T\\rightarrow T + t}$ and target $B^{T + t}$ to calculate $C$ . In this way, with better predictions, the quality of pseudo labels keeps improving during training. The cost matrix $C$ is obtained by computing the pairwise distance between coordinates:", + "bbox": [ + 75, + 297, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {i j} = 1 - \\exp \\left(- \\frac {\\left| \\left| \\hat {b} _ {i} ^ {T} - b _ {j} ^ {T + t} \\right| \\right| ^ {2}}{\\theta_ {c}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 424, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{c}$ is a temperature parameter. According to Eq.1, the pseudo motion labels can be obtained by:", + "bbox": [ + 76, + 465, + 468, + 497 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} ^ {T \\rightarrow T + t} = \\pi^ {*} B ^ {T + t} - B ^ {T}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 503, + 468, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the pseudo motion labels, the supervised training loss can be computed by:", + "bbox": [ + 76, + 530, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s u p} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\operatorname {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, \\hat {M} ^ {T \\rightarrow T + t}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 570, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During pseudo label generation, we first employ a training-free ground segmentation algorithm [12] to remove the majority of ground plane points to reduce the negative influence of ground points for matching and set the motion labels of these removed points as zeros. $\\theta_{c}$ is set as 3 in the paper.", + "bbox": [ + 75, + 619, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Spatial-Temporal Consistency Regularizations", + "text_level": 1, + "bbox": [ + 76, + 717, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1 Motivations", + "text_level": 1, + "bbox": [ + 76, + 739, + 215, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to factors such as occlusion, ground points, and sensors moving, establishing correct correspondences between source and target BEV maps is challenging. Despite removing ground points by the ground segmentation algorithm, noises without correspondences remain. The existence of these noises and cells without correspondences significantly impairs the solution to optimal transport problems. Consequently, the generated pseudo labels are inaccurate, leading to suboptimal outcomes.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For instance, certain cells corresponding to static background regions may get apparent moving labels (e.g., circle A in Fig. 3 (a)), while cells associated with moving objects might possess static or inaccurate moving labels (e.g., circle B in Fig. 3 (a)). Training the model only using such low-quality pseudo labels yields unsatisfactory results including the spatial and temporal inconsistent predictions, as shown in Fig. 3 (b)(c)(d). Therefore, we introduce three spatial and temporal consistency regularization terms. These regularization terms aim to penalize inaccurate predictions and propagate accurate ones. And through the pre-warping operation, the quality of the pseudo labels can improve iteratively as the accuracy of predictions increases.", + "bbox": [ + 496, + 297, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.2 Spatial cluster consistency regularization.", + "text_level": 1, + "bbox": [ + 498, + 513, + 841, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In real-world traffic scenarios, the majority of objects, such as cars and trucks, exhibit a high degree of rigidity. It is intuitive to expect that cells belonging to the same rigid object should display consistent predicted motions. Unfortunately, the generated pseudo labels for cells within the same rigid object are inconsistent (e.g., circle B in Fig. 3 (a)). Consequently, directly learning from these pseudo labels may result in situations like half of the object being predicted with moving status but the other half being predicted with static status (e.g., circle B in Fig. 3 (b)). Therefore, encouraging the consistency of predicted motion within the same object becomes an important factor in enhancing the performance of the self-supervised model.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the unavailability of instance segmentation labels, one potential solution is to employ K-Nearest-Neighbors (KNN) consistency loss [32]. Although this loss helps encourage local smoothness, it does not consider instance-level consistency, especially in the case of large instances. To this end, we utilize cluster consistency loss to perform the spatial consistency constraint. As cells belonging to the same object are often in close proximity to one another in the virtual BEV map, objects can be coarsely represented as compact clusters. Therefore, we first utilize a breadth-first clustering algorithm to cluster every single cell within", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14641", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a distance $d_{c}$ (The detailed clustering process is in the Supplementary). Then for each cluster, we encourage spatial consistency. As illustrated in Sec. 3.3, we perform clustering on 2D BEV maps to reduce the computational cost. The cluster consistency loss is formulated as:", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{| S |} \\sum_ {s _ {i} \\in S} \\frac {1}{| s _ {i} | ^ {2}} \\sum_ {b _ {i}, b _ {j} \\in s _ {i}} \\left(\\left| \\left| \\overrightarrow {\\mathcal {M}} \\left(b _ {i}\\right) - \\overrightarrow {\\mathcal {M}} \\left(b _ {j}\\right) \\right| \\right|\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 172, + 468, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S = \\{s_i\\}_{i=1}^{N_s}$ , $s_i$ indicates a cluster, $N_s$ is the cluster numbers in the BEV map and $\\overrightarrow{\\mathcal{M}}(b)$ is the motion field at cell $b$ . $d_c$ is set as 3 in the paper.", + "bbox": [ + 76, + 220, + 468, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3 Temporal forward consistency regularization.", + "text_level": 1, + "bbox": [ + 76, + 287, + 447, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From supervised loss $\\mathcal{L}_{sup}$ , the motions are learned individually from the pseudo labels for each timestamp. However, since the foreground and noise distributions of future BEV maps are different (e.g. $B^{T + 1}$ and $B^{T + 2}$ ), we will get inconsistent pseudo motion labels between $\\hat{M}^{T\\rightarrow T + 1}$ and $\\hat{M}^{T\\rightarrow T + 2}$ . As the model learns from these inconsistent pseudo labels, it will produce inconsistent predictions. For example, the predictions of static cells in $M^{T\\rightarrow T + 1}$ are wrong but correct in $M^{T\\rightarrow T + 2}$ (circle A in Fig. 3 (c)). As the object displacements in two consecutive short time windows are supposed to be consistent, we regularize forward predictions as:", + "bbox": [ + 75, + 311, + 470, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\sum_ {t = 1} ^ {T ^ {\\prime} - 1} \\operatorname {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, \\frac {t}{t + 1} M ^ {T \\rightarrow T + t + 1}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 500, + 468, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With forward consistency loss $\\mathcal{L}_f$ , the predicted motions of adjacent timestamps act as mutual regularization, which is important for self-supervised motion learning as the pseudo labels are temporally inconsistent. For example, $M^{T\\rightarrow T + 2}$ can provide more accurate information for $M^{T\\rightarrow T + 1}$ in Fig. 3 (c) case.", + "bbox": [ + 75, + 555, + 468, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.4 Temporal backward consistency regularization.", + "text_level": 1, + "bbox": [ + 76, + 662, + 460, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When generating pseudo labels, it is inevitable to assign noise and some ground source cells to the wrong target cells due to the absence of corresponding cells in the target point clouds. Wrong pseudo labels consequently lead the trained model to predict incorrect motion, especially wrongly predicting large displacements for background cells (e.g. circle A in left Fig. 3 (d)). This creates a contradiction between the model's expected behavior (which is to predict static motion for background cells based on the past frames) and the actual predictions learned directly from incorrect pseudo labels. Inspired by the contradiction, we propose backward consistency loss to penalize incorrect predictions.", + "bbox": [ + 75, + 686, + 468, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, besides forward sequence $\\vec{\\nu} = \\{V^1,V^2,\\dots,V^T\\}$ , we also feed additional backward", + "bbox": [ + 76, + 867, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sequence $\\overleftarrow{V} = \\{V^{2T - 1}, V^{2T - 2}, \\dots, V^T\\}$ to the network to predict the backward motion $\\overleftarrow{\\mathcal{M}}$ . Compared to the forward sequence, the backward sequence has entirely different past frames. During training, without being directly constrained by pseudo labels like the forward motion, the backward motion reflects the model's general ability to perceive motion cues hidden in the temporal sequences. Considering that the forward and backward speeds of an object are opposites at the same moment, we formulate the backward regularization term as:", + "bbox": [ + 496, + 88, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\exp \\left(- \\frac {t}{\\theta_ {b}}\\right) \\text {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, - M ^ {T \\rightarrow T - t}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 252, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{b}$ is the temperature parameter. We use the $\\exp \\left(-\\frac{t}{\\theta_b}\\right)$ term to dynamically adjust the loss weights, taking into account that as the time horizon increases, accurately predicting motion becomes more challenging, and the forward-backward divergence becomes less reliable. This inconsistency becomes noticeable in cells with incorrect predictions (Fig. 4), which demonstrates that backward regularization can effectively reflect the accuracy of the predictions. Thus, the backward regularization $\\mathcal{L}_b$ generally penalizes incorrect predictions while having less impact on correct ones. $\\theta_{b}$ is set as 10 in the paper. In Sec. 4, we will show that the backward regularization effectively benefits the self-supervised motion learning process.", + "bbox": [ + 496, + 306, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall loss functions. To summarize, the overall training objective is:", + "bbox": [ + 496, + 502, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p}} + \\alpha \\mathcal {L} _ {c} + \\beta \\mathcal {L} _ {f} + \\gamma \\mathcal {L} _ {b}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 539, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta$ , and $\\gamma$ are hyper-parameters, balancing the contribution of the different loss terms. It is worth noting that all the regularization terms are only used for training.", + "bbox": [ + 496, + 561, + 890, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 618, + 633, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we begin by comparing our methods with state-of-the-art (SOTA) motion prediction methods. Subsequently, we conduct ablation studies to analyze the effectiveness of each individual component.", + "bbox": [ + 496, + 643, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We mainly evaluate the proposed approach on the large-scale self-driving dataset: nuScenes [2], which contains 850 scenes with annotations. For fair comparisons, we follow previous works [19, 30, 31] to use 500 scenes for training, 100 scenes for validation, and 250 scenes for testing.", + "bbox": [ + 496, + 704, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We follow the same pre-process setting in [31], where the point clouds are cropped in the range of $[-32\\mathrm{m}, 32\\mathrm{m}] \\times [-32\\mathrm{m}, 32\\mathrm{m}] \\times [-3\\mathrm{m}, 2\\mathrm{m}]$ and the voxel size is set to $0.25\\mathrm{m} \\times 0.25\\mathrm{m} \\times 0.4\\mathrm{m}$ along XYZ axis. We take the current sweep and the past four sweeps as input. We use MotionNet as the baseline model for its effectiveness and flexibility following [15, 19]. We train the", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14642", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/91420f9252144d9afcc5ab9b75cf3b0c68008782ef56af7ab5e038b4ed307c80.jpg", + "table_caption": [ + "Table 1. Performance comparison on nuScenes dataset. Full., Weak., and Self. refer to fully-supervised, weakly-supervised, and selfsupervised training, respectively. Our approach surpasses state-of-the-arts by a large margin." + ], + "table_footnote": [], + "table_body": "
MethodSup.ModalityYearsStaticSpeed ≤ 5m/s (Slow)Speed ≥ 5m/s (Fast)
Mean↓Median↓Mean↓Median↓Mean↓Median↓
Static-LiDAR-000.61110.09718.65178.1412
FlowNet3D [18]Full.LiDARCVPR20180.041000.81830.17828.52618.0230
HPLFlowNet [8]Full.LiDARCVPR20190.00410.00020.44580.09604.32062.4881
LSTM-ED [25]Full.LiDARICRA20190.035800.35510.10441.58851.0003
MotionNet [31]Full.LiDARCVPR20200.025600.25650.09621.07440.7332
BE-STI [30]Full.LiDARCVPR20220.022000.21150.09290.75110.5413
WeakMotion [15]Weak.LiDARCVPR20230.055800.43370.13051.78231.0887
PointPWC [32]Self.LiDARECCV20200.526400.94230.32253.95302.5149
RigidFlow [14]Self.LiDARCVPR20220.636101.18360.56663.94392.5781
PillarMotion [19]Self.LiDAR+ImageCVPR20210.16200.00100.69720.17583.55042.0844
WeakMotion [15]Self.LiDARCVPR20230.191900.45100.11502.97151.8822
OursSelf.LiDAR-0.041900.32130.10612.29431.0508
", + "bbox": [ + 80, + 127, + 890, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model for 20 epochs and set the learning rate to be 0.002. Adam is used as the optimizer. We implement our model in Pytorch [22] with a single RTX 3090 GPU. Hyperparameters $\\alpha$ , $\\beta$ , and $\\gamma$ are set as 0.05, 0.1, and 1 respectively. Following previous works, we take the sequence with 5 frames ( $T = 5$ ) as input and show the results of the next 1s ( $T' = 5$ ) in the following tables. The time interval between each frame is 0.2s.", + "bbox": [ + 75, + 368, + 468, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Following previous works [19, 30, 31], we evaluate the mean and median errors at different speed levels by dividing the grid cells into 3 groups according to the ground truth speeds: static, slow ( $\\leq 5\\mathrm{m/s}$ ) and fast ( $\\geq 5\\mathrm{m/s}$ ). Errors are measured by $L_{2}$ distances between the predicted displacements and the ground truth displacements.", + "bbox": [ + 75, + 489, + 468, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 604, + 444, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1, we show the results of the following methods: (1) A static model that assumes a static environment. (2) Scene flow estimation methods, including FlowNet3D [18], HPLFlowNet [8], PointPWC [32], and RigidFlow [14]. Following [31], we employ scene flow estimation methods by assuming linear dynamics. First, we predict the motion by estimating the flow from the current time to the past 0.4s. Then, we project the predicted flow onto the BEV map and to the future 1.0s for comparison. (3) Class-agnostic motion prediction methods, including LSTM-ED [25], MotionNet [31], PillarMotion [19], BE-STI [30], and WeakMotionNet [15].", + "bbox": [ + 75, + 628, + 467, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to fully-supervised models, our approach outperforms FlowNet3D and HPLFlowNet at both slow and fast speed levels, and it surpasses the weakly-supervised WeakMotionNet (with $1\\%$ foreground and background masks [15]) at static and slow speed levels. In comparison with self-supervised methods, our approach achieves", + "bbox": [ + 75, + 810, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "superior performance to PillarMotion, with significant improvements of $74.1\\% (-0.1201\\mathrm{m})$ , $53.9\\%$ $(-0.3759\\mathrm{m})$ , and $35.4\\% (-1.2561\\mathrm{m})$ at static, slow, and fast speed levels, respectively. To further show the superiority of the proposed approach, we implement recent WeakMotionNet in a self-supervised manner by replacing its pre-segmentation network with the same ground segmentation algorithm [12] used in our approach. While self-supervised WeakMotionNet outperforms PillarMotion at slow and fast speed levels, our approach still surpasses it by a large margin.", + "bbox": [ + 496, + 368, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 530, + 663, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we evaluate the effectiveness of each component of our approach. As shown in Table 2, the base model trained only with pseudo labels does not perform well at all speed levels (Row 1).", + "bbox": [ + 496, + 553, + 890, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of cluster consistency regularization. When cluster consistency loss is not applied, the model exhibits noticeably poorer performance at the fast speed level (Compare Row 1 and Row 6). This disparity in performance can be attributed to the way the cost matrix for the optimal transport problem is computed based on the distance between cells. Consequently, the solver tends to find correspondences between cells that are in close proximity to each other. While this proximity-based approach has its merits, it presents challenges in generating long displacement pseudo labels that the model can learn from. As a result, during the early stages of training, there is a scarcity of correct pseudo labels for fast motion. By imposing constraints on cluster consistency, the model is encouraged to predict faster motion, even if it corresponds to a short displacement pseudo label. This is possible due to the presence of fast displacement predictions within the same cluster. With larger displacement and pre-warping, the optimal transport solver is more likely to find correct correspondence between two", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14643", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b2195ae128900a7fe7652bb7190d3a6fbcad797531284938df9367ee55a54400.jpg", + "table_caption": [ + "Table 2. Ablation study for different combinations of proposed regularization terms. Bold fonts and underlines indicate the best and the second-best performance, respectively." + ], + "table_footnote": [], + "table_body": "
#Row Lsup LcLfLbStaticSlowFast
Mean Error ↓
1✓0.16500.55403.4503
2✓✓0.04130.32492.7587
3✓✓0.11970.42643.2072
4✓✓✓0.04990.32772.5753
5✓✓✓0.05020.34322.5154
6✓✓0.17050.41542.7027
7✓✓✓0.14500.40872.6509
8✓✓✓0.04190.32132.2943
", + "bbox": [ + 81, + 141, + 467, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "points at long distances. In Table 3, we present the results of KNN consistency loss and cluster consistency loss, both of which function as spatial smooth losses. Compared to $\\mathcal{L}_{KNN}$ , $\\mathcal{L}_c$ bring more benefits comprehensively.", + "bbox": [ + 75, + 373, + 468, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of temporal consistency regularization. Both forward and backward regularization terms boost the performances across all speed levels (Compare Row 5 and Row 7 with Row 8 respectively). To further illustrate the impact of backward regularization, we separately feed forward and backward sequences of all training samples into the model only trained with $\\mathcal{L}_{sup}$ to demonstrate prediction divergence $(|\\overrightarrow{\\mathcal{M}} -\\overleftarrow{\\mathcal{M}} |)$ . As depicted in Fig. 4, there is a positive correlation between the forward-backward prediction divergence and prediction accuracy. The divergence is large when the prediction is inaccurate and vice versa, which indicates that backward regularization will have a stronger impact on inaccurate predictions but a weaker influence on accurate ones. By using the forward-backward divergence as a loss term, the model is penalized for making inaccurate predictions. Results for various $\\theta_{b}$ are shown in Table 4, which indicates that the weight term can further improve the performance.", + "bbox": [ + 75, + 439, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. We present the result visualizations of our approach in Fig. 5 to qualitatively show the impacts of each regularization term. When directly training with pseudo labels $\\mathcal{L}_{sup}$ , the model tends to produce inconsistent predictions for the same object and predict large motions for background cells. The third column illustrates that cluster consistency loss $\\mathcal{L}_c$ effectively benefits spatial consistency. With the addition of forward regularization $\\mathcal{L}_f$ , the prediction results are further improved, but there are still noticeable noises in the background cells. Finally, by incorporating backward regularization $\\mathcal{L}_b$ , the noise predictions in the background cells are significantly suppressed.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/051b30ff7754e176e3b6ba38ef9294a4d619ce3f1247347a71d963700f16cce4.jpg", + "table_caption": [ + "Table 3. Impact of different spatial regularizations. Cluster-level regularization performs better than the KNN-based regularization. Bold fonts and underlines indicate the best and the second-best performance, respectively." + ], + "table_footnote": [], + "table_body": "
MethodStaticSlowFast
Mean Errors ↓
Baseline0.05020.34322.5154
+KLNNK=30.03710.33702.5362
K=50.04900.33782.4568
K=100.07310.34052.5116
+Lc(ours)dc=20.05260.34382.1814
dc=30.04190.32132.2943
dc=40.04590.33182.4434
", + "bbox": [ + 514, + 155, + 879, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fba3aa430a63499204a839452aaeb0135138206cdbd55b4c8fdbc835380013a3.jpg", + "table_caption": [ + "Table 4. Varying $\\theta_{b}$ for backward consistency loss. - indicates the results without the $\\exp\\left(-\\frac{t}{\\theta_{b}}\\right)$ term. Bold fonts and underlines indicate the best and the second-best performance, respectively." + ], + "table_footnote": [], + "table_body": "
θbStaticSlowFast
Mean Errors ↓
50.04820.32642.3580
100.04190.32132.2943
150.02930.30912.4149
-0.03820.32982.4596
", + "bbox": [ + 578, + 383, + 815, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e27d30d153bcee99a0f3e7f3fe3620f0aed3505263e569ca21cc2dc8a2122d2f.jpg", + "image_caption": [ + "Figure 4. Forward-backward divergence of the training set. We analyze the prediction errors of all samples in the training set and their corresponding forward-backward divergences. The Forward-backward divergence is positively correlated with the prediction error." + ], + "image_footnote": [], + "bbox": [ + 521, + 522, + 864, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on Waymo Dataset", + "text_level": 1, + "bbox": [ + 500, + 782, + 741, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following WeakMotionNet [15], we further evaluate our approach on the Waymo dataset [27]. As shown in Table 5, similar to the results on the nuScenes dataset, our approach achieves mean errors of around $0.05\\mathrm{m}$ , $0.4\\mathrm{m}$ , and $2.3\\mathrm{m}$ at static, slow, and fast speed levels, respectively, outperforming self-supervised WeakMotionNet.", + "bbox": [ + 498, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14644", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03e6c58ef98f461137f958f125bfc4be64ebe2489e2a7494bf2d6d9436acbc14.jpg", + "image_caption": [ + "Figure 5. Qualitative results of the proposed self-supervised approach. The future displacements $(M^{T\\rightarrow T + 5})$ are depicted using the color wheel representation. (Zoom in for the best view)" + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 867, + 349 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9a66f31a4bf185aa51b128715ead715efae984d74f1c46e6bec346db08dbaf7a.jpg", + "table_caption": [ + "Table 5. Motion prediction results on the Waymo Dataset. Our method still outperforms self-supervised WeakMotion by a large margin." + ], + "table_footnote": [], + "table_body": "
MethodSup.StaticSlowFast
Mean Errors ↓
MotionNet [31]Full.0.02630.26200.9493
WeakMotion [15]Weak.0.02970.34581.5655
WeakMotion [15]Self.0.13450.58333.0180
OursSelf.0.05150.44402.3692
", + "bbox": [ + 78, + 460, + 473, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Discussions", + "text_level": 1, + "bbox": [ + 76, + 603, + 204, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground Segmentation. We evaluate the performance of the ground segmentation algorithm [12] on the nuScenes dataset, and get a $95\\%$ precision rate while a $92\\%$ recall rate, indicating the majority of ground points can be correctly removed. Although almost all ground points can be successfully removed, the self-supervised training process still suffers from remaining ground points, noises, and points without correspondences. The issues are alleviated using our proposed regularizations, as shown in Fig. 2.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clustering. In the paper, we utilize the breadth-first clustering (BFS) algorithm to demonstrate the effectiveness of the cluster consistency loss. We evaluate the clustering results of BEV cells on the nuScenes dataset and observe a $9.6\\%$ under-clustering rate and an $11.1\\%$ over-clustering rate. It indicates that although effective, the BFS algorithm still wrongly clusters cells from different instances in some cases. Therefore, fellow researchers can consider utilizing or developing more advanced clustering strategies, which", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "may lead to better performances.", + "bbox": [ + 498, + 410, + 718, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency. During training, ground segmentation and clustering take around 5ms and 10ms, respectively. In practice, these processes can be employed as data pre-processing before training. For the optimal transport (OT) solver, we implement it with Pytorch and operate in batch. It takes less than 1ms to solve the OT problem. During inference, given that our approach is model-independent, the inference time is based on the model used. When applying MotionNet, it takes 10ms for point cloud transformation and voxelization and another 10ms for prediction [31].", + "bbox": [ + 496, + 425, + 893, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 616, + 619, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel approach for self-supervised motion prediction with only unlabeled LiDAR point clouds. We utilize optimal transport to generate pseudo labels and propose three novel spatial and temporal regularization terms to facilitate self-supervised motion learning. Extensive experiments on the large-scale nuScenes dataset illustrate that our approach significantly surpasses previous self-supervised methods and shallows the gap with fully-supervised ones.", + "bbox": [ + 496, + 650, + 893, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and the MoE AcRF Tier 1 grant (RG14/22).", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14645", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aseem Behl, Despoina Paschalidou, Simon Donne, and Andreas Geiger. Pointflownet: Learning representations for rigid motion estimation from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5", + "[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2", + "[4] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Neural Information Processing Systems (NeurIPS), 2013. 4", + "[5] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2", + "[6] Guanting Dong, Yueyi Zhang, Hanlin Li, Xiaoyan Sun, and Zhiwei Xiong. Exploiting rigidity constraints for lidar scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[7] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2", + "[8] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6", + "[9] Marcel Van Herk. Image registration using chamfer matching. Handbook of Medical Imaging, 2000. 2", + "[10] Yair Kittenplon, Yonina C Eldar, and Dan Raviv. Flow-step3d: Model unrolling for self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[12] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 3, 4, 6, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ruibo Li, Guosheng Lin, and Lihua Xie. Self-point-flow: Self-supervised scene flow estimation from point clouds with optimal transport and random walk. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[14] Ruibo Li, Chi Zhang, Guosheng Lin, Zhe Wang, and Chunhua Shen. Rigidflow: Self-supervised scene flow learning on point clouds by local rigidity prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6", + "[15] Ruibo Li, Hanyu Shi, Ziang Fu, Zhe Wang, and Guosheng Lin. Weakly supervised class-agnostic motion prediction for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 7, 8", + "[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Neural Information Processing Systems (NeurIPS), 2021. 2", + "[17] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2", + "[18] Xingyu Liu, C. Qi, and Leonidas J. Guibas. Flownet3d: Learning scene flow in 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6", + "[19] Chenxu Luo, Xiaodong Yang, and Alan Loddon Yuille. Self-supervised pillar motion learning for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5, 6", + "[20] Wenjie Luo, Binh Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[21] Himangi Mittal, Brian Okorn, and David Held. Just go with the flow: Self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 6", + "[23] Jhony Kaesemodel Pontes, James Hays, and Simon Lucey. Scene flow from point clouds with or without learning. In International Conference on 3D Vision (3DV), 2020. 2", + "[24] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision (ECCV), 2020. 2", + "[25] Marcel Schreiber, Stefan Hoermann, and Klaus Dietmayer. Long-term occupancy grid prediction using recurrent neural networks. In IEEE international conference on robotics and automation (ICRA), 2019. 1, 2, 6", + "[26] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang chun Woo. Convolutional LSTM network: A machine learning approach for precipita" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14646", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion nowcasting. In Neural Information Processing Systems (NeurIPS), 2015. 2", + "[27] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7", + "[28] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. Hierarchical attention learning of scene flow in 3d point clouds. IEEE Transactions on Image Processing (TIP), 2021. 2", + "[29] Kewei Wang, Yizheng Wu, Zhiyu Pan, Xingyi Li, Ke Xian, Zhe Wang, Zhiguo Cao, and Guosheng Lin. Semi-supervised class-agnostic motion prediction with pseudo label regeneration and bevmix, 2023. 1", + "[30] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6", + "[31] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 5, 6, 8", + "[32] Wenxuan Wu, Zhi Yuan Wang, Zhuwen Li, Wei Liu, and Li Fuxin. Pointpwc-net: Cost volume on point clouds for (self-) supervised scene flow estimation. In European Conference on Computer Vision (ECCV), 2020. 2, 4, 6", + "[33] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In IEEE Conference on International Conference on Computer Vision (ICCV), 2023. 2", + "[34] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yun-ting Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. Tnt: Target-driven trajectory prediction. In Conference on Robot Learning (CoRL), 2020. 1, 2", + "[35] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + ], + "bbox": [ + 78, + 92, + 468, + 724 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14647", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_model.json b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..44da7a3899f0ff9cf5e8bfd0cde94baf19244d16 --- /dev/null +++ b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_model.json @@ -0,0 +1,1881 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.202, + 0.967, + 0.258 + ], + "angle": 0, + "content": "Kewei Wang\\(^{1,2}\\), Yizheng Wu\\(^{1,2}\\), Jun CEN\\(^{2}\\), Zhiyu Pan\\(^{1}\\), Xingyi Li\\(^{1,2}\\), Zhe Wang\\(^{3}\\), Zhiguo Cao\\(^{1}\\), Guosheng Lin\\(^{2*}\\) \n\\(^{1}\\) School of AIA, Huazhong University of Science and Technology \n\\(^{2}\\) S-Lab, Nanyang Technological University \\(^{3}\\) SenseTime Research" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.474, + 0.702 + ], + "angle": 0, + "content": "The perception of motion behavior in a dynamic environment holds significant importance for autonomous driving systems, wherein class-agnostic motion prediction methods directly predict the motion of the entire point cloud. While most existing methods rely on fully-supervised learning, the manual labeling of point cloud data is laborious and time-consuming. Therefore, several annotation-efficient methods have been proposed to address this challenge. Although effective, these methods rely on weak annotations or additional multi-modal data like images, and the potential benefits inherent in the point cloud sequence are still underexplored. To this end, we explore the feasibility of self-supervised motion prediction with only unlabeled Li-DAR point clouds. Initially, we employ an optimal transport solver to establish coarse correspondences between current and future point clouds as the coarse pseudo motion labels. Training models directly using such coarse labels leads to noticeable spatial and temporal prediction inconsistencies. To mitigate these issues, we introduce three simple spatial and temporal regularization losses, which facilitate the self-supervised training process effectively. Experimental results demonstrate the significant superiority of our approach over the state-of-the-art self-supervised methods. Code will be available at https://github.com/kwwcv/SelfMotion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.732, + 0.21, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Motion perception provides key information for autonomous driving. Classical approaches formulate the motion perception task as object-level trajectory prediction [3, 5, 7, 17, 34]. However, these approaches may fail if agents are not successfully detected by the detectors, especially when encountering unseen categories in open-set scenarios [31]. To compensate for trajectory prediction and" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.293, + 0.625, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.399, + 0.591, + 0.41 + ], + "angle": 0, + "content": "(a) Static" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.295, + 0.755, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.399, + 0.721, + 0.41 + ], + "angle": 0, + "content": "(b) Slow" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.295, + 0.884, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.399, + 0.849, + 0.41 + ], + "angle": 0, + "content": "(c) Fast" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.422, + 0.895, + 0.52 + ], + "angle": 0, + "content": "Figure 1. Performance comparison over static, slow, and fast speed levels between self-supervised PillarMotion [19] and our approach on the nuScenes dataset. The dashed line represents the performance of fully-supervised MotionNet [31]. Our proposed self-supervised approach outperforms the PillarMotion which uses additional image data by a large margin and substantially narrows the performance gap with fully-supervised results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.598 + ], + "angle": 0, + "content": "provide backup, class-agnostic motion prediction was studied [15, 25, 29-31]. Currently, these approaches predict the motions of the entire environment from point clouds directly instead of for each agent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.896, + 0.903 + ], + "angle": 0, + "content": "While fully-supervised class-agnostic motion prediction methods have shown success, they typically require a vast amount of annotated point cloud data, which can be both costly and time-consuming to acquire. To overcome this limitation, WeakMotionNet [15] proposes a weakly supervised approach using only a limited number of foreground-background semantic annotations. WeakMotionNet also introduces a consistency-aware chamfer loss that uses the difference between correspondences from different frames as loss weights. Although effective, it requires establishing correspondences multiple times and is still based on matching results between individual frames, rather than taking advantage of the temporal information within the input sequences. Without using any annotations, PillarMotion [19] introduces a self-supervised training approach that incorporates regularization of 2D and 3D flows derived from paired image and point cloud data. However, PillarMotion relies on multi-modality pairs to regularize motion learning, making the training process less flexible, and a considerable performance gap still exists between self-supervised" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.364, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. gslin@ntu.edu.sg, {wangkewei, zgcao} @hust.edu.cn," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14638" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "and fully-supervised methods. Therefore, we explore the spatial and temporal information from the point cloud sequences themselves to train a better class-agnostic motion prediction model with only unlabeled LiDAR data in this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.473, + 0.484 + ], + "angle": 0, + "content": "Our approach primarily involves pseudo label generation and spatial and temporal regularizations. To generate pseudo motion labels, we first solve a matching problem to find correspondence between current and future point clouds. However, the correctness of the matching correspondence cannot be guaranteed due to the occlusion, background artifacts, noise, etc. Self-supervised learning from these coarse matching correspondences (pseudo labels) leads to two major failure cases: (1) predicted motions from the same rigid object are spatially diverse; (2) the background cells that should be static are predicted with large motions. To alleviate these issues, we introduce three simple but effective spatial and temporal regularizations from the point cloud motion prediction itself to facilitate self-supervised learning. Specifically, cluster consistency regularization, backward consistency regularization, and forward consistency regularization are proposed to encourage spatial consistency, penalize incorrect predictions (e.g., large motion for static cells), and smooth the prediction learned from pseudo labels of different time horizons, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Experiments demonstrate that our proposed spatial and temporal consistency regularization terms work effectively together, significantly boosting the performance of self-supervised motion prediction. The proposed approach turns out with excellent performance, as illustrated in Fig. 1. On the nuScenes dataset, our approach outperforms previous methods by a substantial margin. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.605, + 0.469, + 0.665 + ], + "angle": 0, + "content": "- We propose a novel approach for self-supervised class-diagnostic motion prediction that incorporates an optimal transport solver and three simple yet effective spatial and temporal regularization losses." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.666, + 0.469, + 0.695 + ], + "angle": 0, + "content": "- Our approach is model-independent, and does not require any additional models or multi-modal data." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.696, + 0.469, + 0.755 + ], + "angle": 0, + "content": "- Our method surpasses the state-of-the-art (SOTA) self-supervised methods by a large margin and significantly reduces the performance gap with the fully-supervised methods." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.605, + 0.469, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.227, + 0.785 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Class-agnostic motion prediction. Classical motion prediction approaches aim to predict the future trajectories of agents based on past observations. The entire system typically involves object detection [11, 20, 35], tracking, and forecasting [3, 5, 7, 17, 33, 34]. However, the trajectory prediction relies on the object detector which may fail to detect objects with unknown classes in open-set scenarios [31]. To" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.317 + ], + "angle": 0, + "content": "enhance system redundancy, especially in cases where objects may be not successfully detected and potentially lead to a failure in object-level trajectory forecasting [31], class-agnostic motion prediction methods predict the motion of entire scenes directly instead of individually for each agent. These methods represent the environment state based on BEV maps discretized from point clouds and aim to predict the 2D displacement vector for each BEV cell along the ground plane. MotionNet [31] is one of the pioneers in this task, proposing to perform joint perception and motion prediction based on a spatial-temporal pyramid network. LSTM-ED [25] introduce convolutional LSTM [26] to aggregate temporal context. BE-STI [30] leverages semantic cues by training an additional semantic decoder to guide motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.499 + ], + "angle": 0, + "content": "To alleviate the reliance on manual annotations, WeakMotionNet [15] proposes to train the MotionNet in a weakly-supervised manner with limited foreground and background semantic annotations. PillarMotion [19] proposes to train the model in a self-supervised manner without any annotations by cross-sensor motion regularization. Although effective, a substantial performance gap persists between the self-supervised PillarMotion and the fully-supervised model even using additional image data. Therefore, an open question remains about whether we can more efficiently train a self-supervised motion prediction model with higher performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.65 + ], + "angle": 0, + "content": "Scene flow estimation. Scene flow estimation [1, 8, 18, 28, 32] from point clouds is an alternative way to understand the dense class-agnostic 3D motion field. In self-supervised scene flow estimation, several methods rely on establishing point correspondences between source and target point clouds through point matching. They treat the 3D coordinate difference between each matching pair as a pseudo scene flow label [10, 13, 14, 21, 24, 32]. Many methods also utilize the chamfer distance loss [9] to drive self-supervised scene flow learning [6, 10, 16, 23, 32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.831 + ], + "angle": 0, + "content": "While similar to the class-agnostic motion prediction task, scene flow estimation majorly differs in the following ways: (1) The objectiveness of the two tasks is different [15, 31]. Scene flow estimation methods focus on estimating the motion flow between two observed point clouds, while motion prediction methods aim to forecast future displacements based on past observations; (2) Estimating dense 3D flow is time-consuming, which is not feasible for self-driving systems [15]. Furthermore, directly applying scene flow estimation on real LiDAR point clouds poses challenges due to the lack of one-to-one correspondences [19, 30]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.591, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we begin by introducing the problem formulation of class-agnostic motion prediction. Subsequently," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14639" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.881, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.322, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed approach. Without ground truth labels, we first generate pseudo labels by matching. We then introduce cluster, forward, and backward regularization losses to facilitate self-supervised motion learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.41 + ], + "angle": 0, + "content": "we delve into a comprehensive presentation and discussion of each component comprising our approach. The overview of our approach is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.419, + 0.281, + 0.433 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.474, + 0.689 + ], + "angle": 0, + "content": "Given a temporal sequence of LiDAR point cloud frames, past frames are first synchronized to the current coordinate system [31]. We denote synchronized point cloud captured at time \\( t \\) as \\( P^t \\). Following [31], we discretized \\( P^t \\) into dense voxels \\( V^t \\in \\{0,1\\}^{H \\times W \\times C} \\), where 0 indicates the voxel is empty, 1 indicates the voxel is occupied by at least 1 point, and \\( H, W, C \\) are the voxel numbers along \\( X, Y, Z \\) axis respectively. Then we represent the 3D voxel lattice as a 2D pseudo-image with the \\( C \\) corresponding to the image channel and view \\( V^t \\) as a virtual BEV map. Then the motion field \\( M \\in \\mathbb{R}^{H \\times W \\times 2} \\) in the BEV map is defined as the 2D displacement of each grid cell to the future timestamp. Taking sequence \\( \\vec{\\mathcal{V}} = \\{V^t\\}_{t=1}^T \\) as input, the task aims to predict the motion fields \\( \\vec{\\mathcal{M}} = \\{M^T \\rightarrow T + t\\}_{t=1}^{T'} \\), where \\( T \\) and \\( T' \\) are the number of past and future timestamps, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.188, + 0.71 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As depicted in Fig. 2, our approach mainly involves pseudo label generation and self-supervised training. To generate pseudo labels, we initially solve the matching problem between two consecutive frames. Firstly, we remove the majority of ground plane points from point clouds by ground segmentation algorithm [12] to reduce noise for better matching. Subsequently, we feed the raw BEV map sequence into MotionNet [31] and train the model using the generated pseudo labels through supervised loss \\(\\mathcal{L}_{sup}\\). As the quality of pseudo labels is low at the initial stage of training, we introduce three spatial and temporal consistency regularization losses, named cluster consistency loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.364, + 0.892, + 0.395 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_c\\), forward consistency loss \\(\\mathcal{L}_f\\), and backward consistency loss \\(\\mathcal{L}_b\\), to facilitate the self-supervised learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.403, + 0.733, + 0.418 + ], + "angle": 0, + "content": "3.3. Pseudo Label Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Unlike fully-supervised learning, where ground truth labels are available, self-supervised methods learn from data itself without annotations. In this condition, one alternative approach is to generate pseudo labels to provide the necessary supervision for training. Considering motion can be defined as the variation in coordinates between correspondent points in two consecutive frames, the process of generating pseudo labels can be viewed as a matching problem. We denote \\( B^{T} = \\{b_{i}^{T}\\in \\mathbb{R}^{2}\\}_{i = 1}^{N_{T}} \\) as the 2D coordinates of BEV cells on the current virtual BEV map \\( V^T \\), where \\( N_{T} \\) is the number of non-empty cells. Then we perform matching between source \\( B^{T} \\) and target \\( B^{T + t} = \\{b_j^{T + t}\\in \\mathbb{R}^2\\}_{j = 1}^{N_{T + t}} \\). Ideally, if there exists one-to-one correspondence between \\( B^{T} \\) and \\( B^{T + t} \\), \\( B^{T} \\) can be projected into \\( B^{T + t} \\) and fully occupy \\( B^{T + t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.663, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\nB ^ {T} + M = \\pi B ^ {T + t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.768 + ], + "angle": 0, + "content": "where \\(\\pi \\in \\{0,1\\}^{N_T\\times N_{T + t}}\\) is a permutation matrix indicating correspondence between \\(B^T\\) and \\(B^{T + t}\\), and \\(M\\) is the motion flow. The optimal soft permutation matrix \\(\\pi^{*}\\in [0,1]^{N_{T}\\times N_{T + t}}\\) can be computed by solving an optimal transport problem:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.778, + 0.857, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\pi^ {*} = \\underset {\\pi} {\\arg \\min } \\sum_ {i, j} C _ {i j} \\pi_ {i j} \\\\ s. t. \\quad \\pi \\mathbf {1} _ {N _ {T + t}} = \\frac {1}{N _ {T}} \\mathbf {1} _ {N _ {T}}, \\pi^ {T} \\mathbf {1} _ {N _ {T}} = \\frac {1}{N _ {T + t}} \\mathbf {1} _ {N _ {T + t}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_{N_T} \\in \\mathbb{R}^{N_T}\\) and \\(\\mathbf{1}_{N_{T + t}} \\in \\mathbb{R}^{N_{T + t}}\\) are vectors with all elements are 1. \\(C\\) is the transport cost matrix and \\(\\pi^T\\) is the transpose of \\(\\pi\\). This problem can be solved by the Sinkhon" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.892, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Figure 3. Prediction results with inconsistency. (a) Ground truth and pseudo labels. (b) Predictions from the same object are inconsistent. (c) Predictions for consecutive future timestamps are inconsistent. (d) Forward and backward predictions (e.g., \\( M^{T \\to T + 1} \\) and \\( M^{T \\to T - 1} \\)) are inconsistent. Inconsistency regions are highlighted by red circles. The blue arrow denotes the future displacement (motion)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.47, + 0.418 + ], + "angle": 0, + "content": "algorithm [4]. We use a pre-warping operation to warp the source BEV cells \\( B^{T} \\) by the predicted motion \\( M^{T\\rightarrow T + t} \\), and find the correspondence between the warped BEV cells \\( \\hat{B}^T = B^T +M^{T\\rightarrow T + t} \\) and target \\( B^{T + t} \\) to calculate \\( C \\). In this way, with better predictions, the quality of pseudo labels keeps improving during training. The cost matrix \\( C \\) is obtained by computing the pairwise distance between coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.425, + 0.469, + 0.461 + ], + "angle": 0, + "content": "\\[\nC _ {i j} = 1 - \\exp \\left(- \\frac {\\left| \\left| \\hat {b} _ {i} ^ {T} - b _ {j} ^ {T + t} \\right| \\right| ^ {2}}{\\theta_ {c}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.467, + 0.47, + 0.498 + ], + "angle": 0, + "content": "where \\(\\theta_{c}\\) is a temperature parameter. According to Eq.1, the pseudo motion labels can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.505, + 0.469, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\hat {M} ^ {T \\rightarrow T + t} = \\pi^ {*} B ^ {T + t} - B ^ {T}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.469, + 0.562 + ], + "angle": 0, + "content": "With the pseudo motion labels, the supervised training loss can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.571, + 0.469, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s u p} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\operatorname {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, \\hat {M} ^ {T \\rightarrow T + t}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.469, + 0.71 + ], + "angle": 0, + "content": "During pseudo label generation, we first employ a training-free ground segmentation algorithm [12] to remove the majority of ground plane points to reduce the negative influence of ground points for matching and set the motion labels of these removed points as zeros. \\(\\theta_{c}\\) is set as 3 in the paper." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.718, + 0.469, + 0.736 + ], + "angle": 0, + "content": "3.4. Spatial-Temporal Consistency Regularizations" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.216, + 0.755 + ], + "angle": 0, + "content": "3.4.1 Motivations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Due to factors such as occlusion, ground points, and sensors moving, establishing correct correspondences between source and target BEV maps is challenging. Despite removing ground points by the ground segmentation algorithm, noises without correspondences remain. The existence of these noises and cells without correspondences significantly impairs the solution to optimal transport problems. Consequently, the generated pseudo labels are inaccurate, leading to suboptimal outcomes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.495 + ], + "angle": 0, + "content": "For instance, certain cells corresponding to static background regions may get apparent moving labels (e.g., circle A in Fig. 3 (a)), while cells associated with moving objects might possess static or inaccurate moving labels (e.g., circle B in Fig. 3 (a)). Training the model only using such low-quality pseudo labels yields unsatisfactory results including the spatial and temporal inconsistent predictions, as shown in Fig. 3 (b)(c)(d). Therefore, we introduce three spatial and temporal consistency regularization terms. These regularization terms aim to penalize inaccurate predictions and propagate accurate ones. And through the pre-warping operation, the quality of the pseudo labels can improve iteratively as the accuracy of predictions increases." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.843, + 0.53 + ], + "angle": 0, + "content": "3.4.2 Spatial cluster consistency regularization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.734 + ], + "angle": 0, + "content": "In real-world traffic scenarios, the majority of objects, such as cars and trucks, exhibit a high degree of rigidity. It is intuitive to expect that cells belonging to the same rigid object should display consistent predicted motions. Unfortunately, the generated pseudo labels for cells within the same rigid object are inconsistent (e.g., circle B in Fig. 3 (a)). Consequently, directly learning from these pseudo labels may result in situations like half of the object being predicted with moving status but the other half being predicted with static status (e.g., circle B in Fig. 3 (b)). Therefore, encouraging the consistency of predicted motion within the same object becomes an important factor in enhancing the performance of the self-supervised model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Due to the unavailability of instance segmentation labels, one potential solution is to employ K-Nearest-Neighbors (KNN) consistency loss [32]. Although this loss helps encourage local smoothness, it does not consider instance-level consistency, especially in the case of large instances. To this end, we utilize cluster consistency loss to perform the spatial consistency constraint. As cells belonging to the same object are often in close proximity to one another in the virtual BEV map, objects can be coarsely represented as compact clusters. Therefore, we first utilize a breadth-first clustering algorithm to cluster every single cell within" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14641" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.168 + ], + "angle": 0, + "content": "a distance \\( d_{c} \\) (The detailed clustering process is in the Supplementary). Then for each cluster, we encourage spatial consistency. As illustrated in Sec. 3.3, we perform clustering on 2D BEV maps to reduce the computational cost. The cluster consistency loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.174, + 0.47, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{| S |} \\sum_ {s _ {i} \\in S} \\frac {1}{| s _ {i} | ^ {2}} \\sum_ {b _ {i}, b _ {j} \\in s _ {i}} \\left(\\left| \\left| \\overrightarrow {\\mathcal {M}} \\left(b _ {i}\\right) - \\overrightarrow {\\mathcal {M}} \\left(b _ {j}\\right) \\right| \\right|\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.221, + 0.47, + 0.271 + ], + "angle": 0, + "content": "where \\( S = \\{s_i\\}_{i=1}^{N_s} \\), \\( s_i \\) indicates a cluster, \\( N_s \\) is the cluster numbers in the BEV map and \\( \\overrightarrow{\\mathcal{M}}(b) \\) is the motion field at cell \\( b \\). \\( d_c \\) is set as 3 in the paper." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.288, + 0.449, + 0.304 + ], + "angle": 0, + "content": "3.4.3 Temporal forward consistency regularization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.493 + ], + "angle": 0, + "content": "From supervised loss \\(\\mathcal{L}_{sup}\\), the motions are learned individually from the pseudo labels for each timestamp. However, since the foreground and noise distributions of future BEV maps are different (e.g. \\(B^{T + 1}\\) and \\(B^{T + 2}\\)), we will get inconsistent pseudo motion labels between \\(\\hat{M}^{T\\rightarrow T + 1}\\) and \\(\\hat{M}^{T\\rightarrow T + 2}\\). As the model learns from these inconsistent pseudo labels, it will produce inconsistent predictions. For example, the predictions of static cells in \\(M^{T\\rightarrow T + 1}\\) are wrong but correct in \\(M^{T\\rightarrow T + 2}\\) (circle A in Fig. 3 (c)). As the object displacements in two consecutive short time windows are supposed to be consistent, we regularize forward predictions as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.501, + 0.469, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\sum_ {t = 1} ^ {T ^ {\\prime} - 1} \\operatorname {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, \\frac {t}{t + 1} M ^ {T \\rightarrow T + t + 1}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.646 + ], + "angle": 0, + "content": "With forward consistency loss \\(\\mathcal{L}_f\\), the predicted motions of adjacent timestamps act as mutual regularization, which is important for self-supervised motion learning as the pseudo labels are temporally inconsistent. For example, \\(M^{T\\rightarrow T + 2}\\) can provide more accurate information for \\(M^{T\\rightarrow T + 1}\\) in Fig. 3 (c) case." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.462, + 0.68 + ], + "angle": 0, + "content": "3.4.4 Temporal backward consistency regularization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.868 + ], + "angle": 0, + "content": "When generating pseudo labels, it is inevitable to assign noise and some ground source cells to the wrong target cells due to the absence of corresponding cells in the target point clouds. Wrong pseudo labels consequently lead the trained model to predict incorrect motion, especially wrongly predicting large displacements for background cells (e.g. circle A in left Fig. 3 (d)). This creates a contradiction between the model's expected behavior (which is to predict static motion for background cells based on the past frames) and the actual predictions learned directly from incorrect pseudo labels. Inspired by the contradiction, we propose backward consistency loss to penalize incorrect predictions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.868, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Specifically, besides forward sequence \\(\\vec{\\nu} = \\{V^1,V^2,\\dots,V^T\\}\\), we also feed additional backward" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.246 + ], + "angle": 0, + "content": "sequence \\(\\overleftarrow{V} = \\{V^{2T - 1}, V^{2T - 2}, \\dots, V^T\\}\\) to the network to predict the backward motion \\(\\overleftarrow{\\mathcal{M}}\\). Compared to the forward sequence, the backward sequence has entirely different past frames. During training, without being directly constrained by pseudo labels like the forward motion, the backward motion reflects the model's general ability to perceive motion cues hidden in the temporal sequences. Considering that the forward and backward speeds of an object are opposites at the same moment, we formulate the backward regularization term as:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.253, + 0.892, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\exp \\left(- \\frac {t}{\\theta_ {b}}\\right) \\text {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, - M ^ {T \\rightarrow T - t}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.502 + ], + "angle": 0, + "content": "where \\(\\theta_{b}\\) is the temperature parameter. We use the \\(\\exp \\left(-\\frac{t}{\\theta_b}\\right)\\) term to dynamically adjust the loss weights, taking into account that as the time horizon increases, accurately predicting motion becomes more challenging, and the forward-backward divergence becomes less reliable. This inconsistency becomes noticeable in cells with incorrect predictions (Fig. 4), which demonstrates that backward regularization can effectively reflect the accuracy of the predictions. Thus, the backward regularization \\(\\mathcal{L}_b\\) generally penalizes incorrect predictions while having less impact on correct ones. \\(\\theta_{b}\\) is set as 10 in the paper. In Sec. 4, we will show that the backward regularization effectively benefits the self-supervised motion learning process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Overall loss functions. To summarize, the overall training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.54, + 0.892, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p}} + \\alpha \\mathcal {L} _ {c} + \\beta \\mathcal {L} _ {f} + \\gamma \\mathcal {L} _ {b}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.608 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta\\), and \\(\\gamma\\) are hyper-parameters, balancing the contribution of the different loss terms. It is worth noting that all the regularization terms are only used for training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.634, + 0.636 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.704 + ], + "angle": 0, + "content": "In this section, we begin by comparing our methods with state-of-the-art (SOTA) motion prediction methods. Subsequently, we conduct ablation studies to analyze the effectiveness of each individual component." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Dataset. We mainly evaluate the proposed approach on the large-scale self-driving dataset: nuScenes [2], which contains 850 scenes with annotations. For fair comparisons, we follow previous works [19, 30, 31] to use 500 scenes for training, 100 scenes for validation, and 250 scenes for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We follow the same pre-process setting in [31], where the point clouds are cropped in the range of \\([-32\\mathrm{m}, 32\\mathrm{m}] \\times [-32\\mathrm{m}, 32\\mathrm{m}] \\times [-3\\mathrm{m}, 2\\mathrm{m}]\\) and the voxel size is set to \\(0.25\\mathrm{m} \\times 0.25\\mathrm{m} \\times 0.4\\mathrm{m}\\) along XYZ axis. We take the current sweep and the past four sweeps as input. We use MotionNet as the baseline model for its effectiveness and flexibility following [15, 19]. We train the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14642" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performance comparison on nuScenes dataset. Full., Weak., and Self. refer to fully-supervised, weakly-supervised, and selfsupervised training, respectively. Our approach surpasses state-of-the-arts by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.128, + 0.892, + 0.344 + ], + "angle": 0, + "content": "
MethodSup.ModalityYearsStaticSpeed ≤ 5m/s (Slow)Speed ≥ 5m/s (Fast)
Mean↓Median↓Mean↓Median↓Mean↓Median↓
Static-LiDAR-000.61110.09718.65178.1412
FlowNet3D [18]Full.LiDARCVPR20180.041000.81830.17828.52618.0230
HPLFlowNet [8]Full.LiDARCVPR20190.00410.00020.44580.09604.32062.4881
LSTM-ED [25]Full.LiDARICRA20190.035800.35510.10441.58851.0003
MotionNet [31]Full.LiDARCVPR20200.025600.25650.09621.07440.7332
BE-STI [30]Full.LiDARCVPR20220.022000.21150.09290.75110.5413
WeakMotion [15]Weak.LiDARCVPR20230.055800.43370.13051.78231.0887
PointPWC [32]Self.LiDARECCV20200.526400.94230.32253.95302.5149
RigidFlow [14]Self.LiDARCVPR20220.636101.18360.56663.94392.5781
PillarMotion [19]Self.LiDAR+ImageCVPR20210.16200.00100.69720.17583.55042.0844
WeakMotion [15]Self.LiDARCVPR20230.191900.45100.11502.97151.8822
OursSelf.LiDAR-0.041900.32130.10612.29431.0508
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.489 + ], + "angle": 0, + "content": "model for 20 epochs and set the learning rate to be 0.002. Adam is used as the optimizer. We implement our model in Pytorch [22] with a single RTX 3090 GPU. Hyperparameters \\(\\alpha\\), \\(\\beta\\), and \\(\\gamma\\) are set as 0.05, 0.1, and 1 respectively. Following previous works, we take the sequence with 5 frames (\\(T = 5\\)) as input and show the results of the next 1s (\\(T' = 5\\)) in the following tables. The time interval between each frame is 0.2s." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Evaluation metrics. Following previous works [19, 30, 31], we evaluate the mean and median errors at different speed levels by dividing the grid cells into 3 groups according to the ground truth speeds: static, slow (\\(\\leq 5\\mathrm{m/s}\\)) and fast (\\(\\geq 5\\mathrm{m/s}\\)). Errors are measured by \\(L_{2}\\) distances between the predicted displacements and the ground truth displacements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.446, + 0.621 + ], + "angle": 0, + "content": "4.1. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.809 + ], + "angle": 0, + "content": "In Table 1, we show the results of the following methods: (1) A static model that assumes a static environment. (2) Scene flow estimation methods, including FlowNet3D [18], HPLFlowNet [8], PointPWC [32], and RigidFlow [14]. Following [31], we employ scene flow estimation methods by assuming linear dynamics. First, we predict the motion by estimating the flow from the current time to the past 0.4s. Then, we project the predicted flow onto the BEV map and to the future 1.0s for comparison. (3) Class-agnostic motion prediction methods, including LSTM-ED [25], MotionNet [31], PillarMotion [19], BE-STI [30], and WeakMotionNet [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Compared to fully-supervised models, our approach outperforms FlowNet3D and HPLFlowNet at both slow and fast speed levels, and it surpasses the weakly-supervised WeakMotionNet (with \\(1\\%\\) foreground and background masks [15]) at static and slow speed levels. In comparison with self-supervised methods, our approach achieves" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.522 + ], + "angle": 0, + "content": "superior performance to PillarMotion, with significant improvements of \\(74.1\\% (-0.1201\\mathrm{m})\\), \\(53.9\\%\\) \\((-0.3759\\mathrm{m})\\), and \\(35.4\\% (-1.2561\\mathrm{m})\\) at static, slow, and fast speed levels, respectively. To further show the superiority of the proposed approach, we implement recent WeakMotionNet in a self-supervised manner by replacing its pre-segmentation network with the same ground segmentation algorithm [12] used in our approach. While self-supervised WeakMotionNet outperforms PillarMotion at slow and fast speed levels, our approach still surpasses it by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.665, + 0.545 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.613 + ], + "angle": 0, + "content": "In this subsection, we evaluate the effectiveness of each component of our approach. As shown in Table 2, the base model trained only with pseudo labels does not perform well at all speed levels (Row 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of cluster consistency regularization. When cluster consistency loss is not applied, the model exhibits noticeably poorer performance at the fast speed level (Compare Row 1 and Row 6). This disparity in performance can be attributed to the way the cost matrix for the optimal transport problem is computed based on the distance between cells. Consequently, the solver tends to find correspondences between cells that are in close proximity to each other. While this proximity-based approach has its merits, it presents challenges in generating long displacement pseudo labels that the model can learn from. As a result, during the early stages of training, there is a scarcity of correct pseudo labels for fast motion. By imposing constraints on cluster consistency, the model is encouraged to predict faster motion, even if it corresponds to a short displacement pseudo label. This is possible due to the presence of fast displacement predictions within the same cluster. With larger displacement and pre-warping, the optimal transport solver is more likely to find correct correspondence between two" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14643" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.132 + ], + "angle": 0, + "content": "Table 2. Ablation study for different combinations of proposed regularization terms. Bold fonts and underlines indicate the best and the second-best performance, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.142, + 0.468, + 0.338 + ], + "angle": 0, + "content": "
#Row Lsup LcLfLbStaticSlowFast
Mean Error ↓
1✓0.16500.55403.4503
2✓✓0.04130.32492.7587
3✓✓0.11970.42643.2072
4✓✓✓0.04990.32772.5753
5✓✓✓0.05020.34322.5154
6✓✓0.17050.41542.7027
7✓✓✓0.14500.40872.6509
8✓✓✓0.04190.32132.2943
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.47, + 0.435 + ], + "angle": 0, + "content": "points at long distances. In Table 3, we present the results of KNN consistency loss and cluster consistency loss, both of which function as spatial smooth losses. Compared to \\(\\mathcal{L}_{KNN}\\), \\(\\mathcal{L}_c\\) bring more benefits comprehensively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Effectiveness of temporal consistency regularization. Both forward and backward regularization terms boost the performances across all speed levels (Compare Row 5 and Row 7 with Row 8 respectively). To further illustrate the impact of backward regularization, we separately feed forward and backward sequences of all training samples into the model only trained with \\(\\mathcal{L}_{sup}\\) to demonstrate prediction divergence \\((|\\overrightarrow{\\mathcal{M}} -\\overleftarrow{\\mathcal{M}} |)\\). As depicted in Fig. 4, there is a positive correlation between the forward-backward prediction divergence and prediction accuracy. The divergence is large when the prediction is inaccurate and vice versa, which indicates that backward regularization will have a stronger impact on inaccurate predictions but a weaker influence on accurate ones. By using the forward-backward divergence as a loss term, the model is penalized for making inaccurate predictions. Results for various \\(\\theta_{b}\\) are shown in Table 4, which indicates that the weight term can further improve the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. We present the result visualizations of our approach in Fig. 5 to qualitatively show the impacts of each regularization term. When directly training with pseudo labels \\(\\mathcal{L}_{sup}\\), the model tends to produce inconsistent predictions for the same object and predict large motions for background cells. The third column illustrates that cluster consistency loss \\(\\mathcal{L}_c\\) effectively benefits spatial consistency. With the addition of forward regularization \\(\\mathcal{L}_f\\), the prediction results are further improved, but there are still noticeable noises in the background cells. Finally, by incorporating backward regularization \\(\\mathcal{L}_b\\), the noise predictions in the background cells are significantly suppressed." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 3. Impact of different spatial regularizations. Cluster-level regularization performs better than the KNN-based regularization. Bold fonts and underlines indicate the best and the second-best performance, respectively." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.156, + 0.88, + 0.313 + ], + "angle": 0, + "content": "
MethodStaticSlowFast
Mean Errors ↓
Baseline0.05020.34322.5154
+KLNNK=30.03710.33702.5362
K=50.04900.33782.4568
K=100.07310.34052.5116
+Lc(ours)dc=20.05260.34382.1814
dc=30.04190.32132.2943
dc=40.04590.33182.4434
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.332, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Table 4. Varying \\( \\theta_{b} \\) for backward consistency loss. - indicates the results without the \\( \\exp\\left(-\\frac{t}{\\theta_{b}}\\right) \\) term. Bold fonts and underlines indicate the best and the second-best performance, respectively." + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.385, + 0.816, + 0.5 + ], + "angle": 0, + "content": "
θbStaticSlowFast
Mean Errors ↓
50.04820.32642.3580
100.04190.32132.2943
150.02930.30912.4149
-0.03820.32982.4596
" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.523, + 0.866, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.667, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Figure 4. Forward-backward divergence of the training set. We analyze the prediction errors of all samples in the training set and their corresponding forward-backward divergences. The Forward-backward divergence is positively correlated with the prediction error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.783, + 0.742, + 0.8 + ], + "angle": 0, + "content": "4.3. Results on Waymo Dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Following WeakMotionNet [15], we further evaluate our approach on the Waymo dataset [27]. As shown in Table 5, similar to the results on the nuScenes dataset, our approach achieves mean errors of around \\(0.05\\mathrm{m}\\), \\(0.4\\mathrm{m}\\), and \\(2.3\\mathrm{m}\\) at static, slow, and fast speed levels, respectively, outperforming self-supervised WeakMotionNet." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14644" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.091, + 0.868, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.361, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of the proposed self-supervised approach. The future displacements \\((M^{T\\rightarrow T + 5})\\) are depicted using the color wheel representation. (Zoom in for the best view)" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.408, + 0.475, + 0.451 + ], + "angle": 0, + "content": "Table 5. Motion prediction results on the Waymo Dataset. Our method still outperforms self-supervised WeakMotion by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.462, + 0.474, + 0.577 + ], + "angle": 0, + "content": "
MethodSup.StaticSlowFast
Mean Errors ↓
MotionNet [31]Full.0.02630.26200.9493
WeakMotion [15]Weak.0.02970.34581.5655
WeakMotion [15]Self.0.13450.58333.0180
OursSelf.0.05150.44402.3692
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.205, + 0.619 + ], + "angle": 0, + "content": "4.4. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Ground Segmentation. We evaluate the performance of the ground segmentation algorithm [12] on the nuScenes dataset, and get a \\(95\\%\\) precision rate while a \\(92\\%\\) recall rate, indicating the majority of ground points can be correctly removed. Although almost all ground points can be successfully removed, the self-supervised training process still suffers from remaining ground points, noises, and points without correspondences. The issues are alleviated using our proposed regularizations, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Clustering. In the paper, we utilize the breadth-first clustering (BFS) algorithm to demonstrate the effectiveness of the cluster consistency loss. We evaluate the clustering results of BEV cells on the nuScenes dataset and observe a \\(9.6\\%\\) under-clustering rate and an \\(11.1\\%\\) over-clustering rate. It indicates that although effective, the BFS algorithm still wrongly clusters cells from different instances in some cases. Therefore, fellow researchers can consider utilizing or developing more advanced clustering strategies, which" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.72, + 0.426 + ], + "angle": 0, + "content": "may lead to better performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Efficiency. During training, ground segmentation and clustering take around 5ms and 10ms, respectively. In practice, these processes can be employed as data pre-processing before training. For the optimal transport (OT) solver, we implement it with Pytorch and operate in batch. It takes less than 1ms to solve the OT problem. During inference, given that our approach is model-independent, the inference time is based on the model used. When applying MotionNet, it takes 10ms for point cloud transformation and voxelization and another 10ms for prediction [31]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.62, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.894, + 0.788 + ], + "angle": 0, + "content": "In this paper, we introduce a novel approach for self-supervised motion prediction with only unlabeled LiDAR point clouds. We utilize optimal transport to generate pseudo labels and propose three novel spatial and temporal regularization terms to facilitate self-supervised motion learning. Extensive experiments on the large-scale nuScenes dataset illustrate that our approach significantly surpasses previous self-supervised methods and shallows the gap with fully-supervised ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and the MoE AcRF Tier 1 grant (RG14/22)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14645" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Aseem Behl, Despoina Paschalidou, Simon Donne, and Andreas Geiger. Pointflownet: Learning representations for rigid motion estimation from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.471, + 0.355 + ], + "angle": 0, + "content": "[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[4] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Neural Information Processing Systems (NeurIPS), 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[5] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[6] Guanting Dong, Yueyi Zhang, Hanlin Li, Xiaoyan Sun, and Zhiwei Xiong. Exploiting rigidity constraints for lidar scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[7] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[8] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.673, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[9] Marcel Van Herk. Image registration using chamfer matching. Handbook of Medical Imaging, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[10] Yair Kittenplon, Yonina C Eldar, and Dan Raviv. Flow-step3d: Model unrolling for self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 3, 4, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[13] Ruibo Li, Guosheng Lin, and Lihua Xie. Self-point-flow: Self-supervised scene flow estimation from point clouds with optimal transport and random walk. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[14] Ruibo Li, Chi Zhang, Guosheng Lin, Zhe Wang, and Chunhua Shen. Rigidflow: Self-supervised scene flow learning on point clouds by local rigidity prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[15] Ruibo Li, Hanyu Shi, Ziang Fu, Zhe Wang, and Guosheng Lin. Weakly supervised class-agnostic motion prediction for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[18] Xingyu Liu, C. Qi, and Leonidas J. Guibas. Flownet3d: Learning scene flow in 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Chenxu Luo, Xiaodong Yang, and Alan Loddon Yuille. Self-supervised pillar motion learning for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[20] Wenjie Luo, Binh Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[21] Himangi Mittal, Brian Okorn, and David Held. Just go with the flow: Self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[23] Jhony Kaesemodel Pontes, James Hays, and Simon Lucey. Scene flow from point clouds with or without learning. In International Conference on 3D Vision (3DV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Marcel Schreiber, Stefan Hoermann, and Klaus Dietmayer. Long-term occupancy grid prediction using recurrent neural networks. In IEEE international conference on robotics and automation (ICRA), 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang chun Woo. Convolutional LSTM network: A machine learning approach for precipita" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14646" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "tion nowcasting. In Neural Information Processing Systems (NeurIPS), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[27] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[28] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. Hierarchical attention learning of scene flow in 3d point clouds. IEEE Transactions on Image Processing (TIP), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[29] Kewei Wang, Yizheng Wu, Zhiyu Pan, Xingyi Li, Ke Xian, Zhe Wang, Zhiguo Cao, and Guosheng Lin. Semi-supervised class-agnostic motion prediction with pseudo label regeneration and bevmix, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[30] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[31] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[32] Wenxuan Wu, Zhi Yuan Wang, Zhuwen Li, Wei Liu, and Li Fuxin. Pointpwc-net: Cost volume on point clouds for (self-) supervised scene flow estimation. In European Conference on Computer Vision (ECCV), 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[33] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In IEEE Conference on International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[34] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yun-ting Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. Tnt: Target-driven trajectory prediction. In Conference on Robot Learning (CoRL), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.725 + ], + "angle": 0, + "content": "[35] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14647" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_origin.pdf b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4a3f6e579658195a157450a424b95edf9ae5625b --- /dev/null +++ b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/b067b514-00bd-444a-bc87-e7fa056dae8d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3997bd8f5e6e9585ec2b8323e43aef71ad6d95b8024e794b589a4e56beb7fde0 +size 1532878 diff --git a/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/full.md b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..71ab97904fedf9c56de9e90e9606cf645de15bd1 --- /dev/null +++ b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/full.md @@ -0,0 +1,279 @@ +# Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations + +Kewei Wang $^{1,2}$ , Yizheng Wu $^{1,2}$ , Jun CEN $^{2}$ , Zhiyu Pan $^{1}$ , Xingyi Li $^{1,2}$ , Zhe Wang $^{3}$ , Zhiguo Cao $^{1}$ , Guosheng Lin $^{2*}$ $^{1}$ School of AIA, Huazhong University of Science and Technology + $^{2}$ S-Lab, Nanyang Technological University $^{3}$ SenseTime Research + +# Abstract + +The perception of motion behavior in a dynamic environment holds significant importance for autonomous driving systems, wherein class-agnostic motion prediction methods directly predict the motion of the entire point cloud. While most existing methods rely on fully-supervised learning, the manual labeling of point cloud data is laborious and time-consuming. Therefore, several annotation-efficient methods have been proposed to address this challenge. Although effective, these methods rely on weak annotations or additional multi-modal data like images, and the potential benefits inherent in the point cloud sequence are still underexplored. To this end, we explore the feasibility of self-supervised motion prediction with only unlabeled Li-DAR point clouds. Initially, we employ an optimal transport solver to establish coarse correspondences between current and future point clouds as the coarse pseudo motion labels. Training models directly using such coarse labels leads to noticeable spatial and temporal prediction inconsistencies. To mitigate these issues, we introduce three simple spatial and temporal regularization losses, which facilitate the self-supervised training process effectively. Experimental results demonstrate the significant superiority of our approach over the state-of-the-art self-supervised methods. Code will be available at https://github.com/kwwcv/SelfMotion. + +# 1. Introduction + +Motion perception provides key information for autonomous driving. Classical approaches formulate the motion perception task as object-level trajectory prediction [3, 5, 7, 17, 34]. However, these approaches may fail if agents are not successfully detected by the detectors, especially when encountering unseen categories in open-set scenarios [31]. To compensate for trajectory prediction and + +![](images/ce930b58cf4438ef67882a4cc13a3aee263dac2a39d9c54cd32655cedbb0fed1.jpg) +(a) Static + +![](images/3861d21624ce41455162fd11497d55066089487ded343635fecb8763b38fb167.jpg) +(b) Slow + +![](images/08ee62acdb6ef05645e69957b99eef61d44dd42491849ff918e08dceb654dd3c.jpg) +(c) Fast +Figure 1. Performance comparison over static, slow, and fast speed levels between self-supervised PillarMotion [19] and our approach on the nuScenes dataset. The dashed line represents the performance of fully-supervised MotionNet [31]. Our proposed self-supervised approach outperforms the PillarMotion which uses additional image data by a large margin and substantially narrows the performance gap with fully-supervised results. + +provide backup, class-agnostic motion prediction was studied [15, 25, 29-31]. Currently, these approaches predict the motions of the entire environment from point clouds directly instead of for each agent. + +While fully-supervised class-agnostic motion prediction methods have shown success, they typically require a vast amount of annotated point cloud data, which can be both costly and time-consuming to acquire. To overcome this limitation, WeakMotionNet [15] proposes a weakly supervised approach using only a limited number of foreground-background semantic annotations. WeakMotionNet also introduces a consistency-aware chamfer loss that uses the difference between correspondences from different frames as loss weights. Although effective, it requires establishing correspondences multiple times and is still based on matching results between individual frames, rather than taking advantage of the temporal information within the input sequences. Without using any annotations, PillarMotion [19] introduces a self-supervised training approach that incorporates regularization of 2D and 3D flows derived from paired image and point cloud data. However, PillarMotion relies on multi-modality pairs to regularize motion learning, making the training process less flexible, and a considerable performance gap still exists between self-supervised + +and fully-supervised methods. Therefore, we explore the spatial and temporal information from the point cloud sequences themselves to train a better class-agnostic motion prediction model with only unlabeled LiDAR data in this paper. + +Our approach primarily involves pseudo label generation and spatial and temporal regularizations. To generate pseudo motion labels, we first solve a matching problem to find correspondence between current and future point clouds. However, the correctness of the matching correspondence cannot be guaranteed due to the occlusion, background artifacts, noise, etc. Self-supervised learning from these coarse matching correspondences (pseudo labels) leads to two major failure cases: (1) predicted motions from the same rigid object are spatially diverse; (2) the background cells that should be static are predicted with large motions. To alleviate these issues, we introduce three simple but effective spatial and temporal regularizations from the point cloud motion prediction itself to facilitate self-supervised learning. Specifically, cluster consistency regularization, backward consistency regularization, and forward consistency regularization are proposed to encourage spatial consistency, penalize incorrect predictions (e.g., large motion for static cells), and smooth the prediction learned from pseudo labels of different time horizons, respectively. + +Experiments demonstrate that our proposed spatial and temporal consistency regularization terms work effectively together, significantly boosting the performance of self-supervised motion prediction. The proposed approach turns out with excellent performance, as illustrated in Fig. 1. On the nuScenes dataset, our approach outperforms previous methods by a substantial margin. Our contributions can be summarized as follows: + +- We propose a novel approach for self-supervised class-diagnostic motion prediction that incorporates an optimal transport solver and three simple yet effective spatial and temporal regularization losses. +- Our approach is model-independent, and does not require any additional models or multi-modal data. +- Our method surpasses the state-of-the-art (SOTA) self-supervised methods by a large margin and significantly reduces the performance gap with the fully-supervised methods. + +# 2. Related Works + +Class-agnostic motion prediction. Classical motion prediction approaches aim to predict the future trajectories of agents based on past observations. The entire system typically involves object detection [11, 20, 35], tracking, and forecasting [3, 5, 7, 17, 33, 34]. However, the trajectory prediction relies on the object detector which may fail to detect objects with unknown classes in open-set scenarios [31]. To + +enhance system redundancy, especially in cases where objects may be not successfully detected and potentially lead to a failure in object-level trajectory forecasting [31], class-agnostic motion prediction methods predict the motion of entire scenes directly instead of individually for each agent. These methods represent the environment state based on BEV maps discretized from point clouds and aim to predict the 2D displacement vector for each BEV cell along the ground plane. MotionNet [31] is one of the pioneers in this task, proposing to perform joint perception and motion prediction based on a spatial-temporal pyramid network. LSTM-ED [25] introduce convolutional LSTM [26] to aggregate temporal context. BE-STI [30] leverages semantic cues by training an additional semantic decoder to guide motion prediction. + +To alleviate the reliance on manual annotations, WeakMotionNet [15] proposes to train the MotionNet in a weakly-supervised manner with limited foreground and background semantic annotations. PillarMotion [19] proposes to train the model in a self-supervised manner without any annotations by cross-sensor motion regularization. Although effective, a substantial performance gap persists between the self-supervised PillarMotion and the fully-supervised model even using additional image data. Therefore, an open question remains about whether we can more efficiently train a self-supervised motion prediction model with higher performance. + +Scene flow estimation. Scene flow estimation [1, 8, 18, 28, 32] from point clouds is an alternative way to understand the dense class-agnostic 3D motion field. In self-supervised scene flow estimation, several methods rely on establishing point correspondences between source and target point clouds through point matching. They treat the 3D coordinate difference between each matching pair as a pseudo scene flow label [10, 13, 14, 21, 24, 32]. Many methods also utilize the chamfer distance loss [9] to drive self-supervised scene flow learning [6, 10, 16, 23, 32]. + +While similar to the class-agnostic motion prediction task, scene flow estimation majorly differs in the following ways: (1) The objectiveness of the two tasks is different [15, 31]. Scene flow estimation methods focus on estimating the motion flow between two observed point clouds, while motion prediction methods aim to forecast future displacements based on past observations; (2) Estimating dense 3D flow is time-consuming, which is not feasible for self-driving systems [15]. Furthermore, directly applying scene flow estimation on real LiDAR point clouds poses challenges due to the lack of one-to-one correspondences [19, 30]. + +# 3. Method + +In this section, we begin by introducing the problem formulation of class-agnostic motion prediction. Subsequently, + +![](images/0193a72f083ec9645e4381c1c8f3cfe0aff63792de326a79e20cf656354cdba3.jpg) +Figure 2. Overview of the proposed approach. Without ground truth labels, we first generate pseudo labels by matching. We then introduce cluster, forward, and backward regularization losses to facilitate self-supervised motion learning. + +we delve into a comprehensive presentation and discussion of each component comprising our approach. The overview of our approach is illustrated in Fig. 2. + +# 3.1. Problem Formulation + +Given a temporal sequence of LiDAR point cloud frames, past frames are first synchronized to the current coordinate system [31]. We denote synchronized point cloud captured at time $t$ as $P^t$ . Following [31], we discretized $P^t$ into dense voxels $V^t \in \{0,1\}^{H \times W \times C}$ , where 0 indicates the voxel is empty, 1 indicates the voxel is occupied by at least 1 point, and $H, W, C$ are the voxel numbers along $X, Y, Z$ axis respectively. Then we represent the 3D voxel lattice as a 2D pseudo-image with the $C$ corresponding to the image channel and view $V^t$ as a virtual BEV map. Then the motion field $M \in \mathbb{R}^{H \times W \times 2}$ in the BEV map is defined as the 2D displacement of each grid cell to the future timestamp. Taking sequence $\vec{\mathcal{V}} = \{V^t\}_{t=1}^T$ as input, the task aims to predict the motion fields $\vec{\mathcal{M}} = \{M^T \rightarrow T + t\}_{t=1}^{T'}$ , where $T$ and $T'$ are the number of past and future timestamps, respectively. + +# 3.2. Overview + +As depicted in Fig. 2, our approach mainly involves pseudo label generation and self-supervised training. To generate pseudo labels, we initially solve the matching problem between two consecutive frames. Firstly, we remove the majority of ground plane points from point clouds by ground segmentation algorithm [12] to reduce noise for better matching. Subsequently, we feed the raw BEV map sequence into MotionNet [31] and train the model using the generated pseudo labels through supervised loss $\mathcal{L}_{sup}$ . As the quality of pseudo labels is low at the initial stage of training, we introduce three spatial and temporal consistency regularization losses, named cluster consistency loss + +$\mathcal{L}_c$ , forward consistency loss $\mathcal{L}_f$ , and backward consistency loss $\mathcal{L}_b$ , to facilitate the self-supervised learning. + +# 3.3. Pseudo Label Generation + +Unlike fully-supervised learning, where ground truth labels are available, self-supervised methods learn from data itself without annotations. In this condition, one alternative approach is to generate pseudo labels to provide the necessary supervision for training. Considering motion can be defined as the variation in coordinates between correspondent points in two consecutive frames, the process of generating pseudo labels can be viewed as a matching problem. We denote $B^{T} = \{b_{i}^{T}\in \mathbb{R}^{2}\}_{i = 1}^{N_{T}}$ as the 2D coordinates of BEV cells on the current virtual BEV map $V^T$ , where $N_{T}$ is the number of non-empty cells. Then we perform matching between source $B^{T}$ and target $B^{T + t} = \{b_j^{T + t}\in \mathbb{R}^2\}_{j = 1}^{N_{T + t}}$ . Ideally, if there exists one-to-one correspondence between $B^{T}$ and $B^{T + t}$ , $B^{T}$ can be projected into $B^{T + t}$ and fully occupy $B^{T + t}$ : + +$$ +B ^ {T} + M = \pi B ^ {T + t}, \tag {1} +$$ + +where $\pi \in \{0,1\}^{N_T\times N_{T + t}}$ is a permutation matrix indicating correspondence between $B^T$ and $B^{T + t}$ , and $M$ is the motion flow. The optimal soft permutation matrix $\pi^{*}\in [0,1]^{N_{T}\times N_{T + t}}$ can be computed by solving an optimal transport problem: + +$$ +\begin{array}{l} \pi^ {*} = \underset {\pi} {\arg \min } \sum_ {i, j} C _ {i j} \pi_ {i j} \\ s. t. \quad \pi \mathbf {1} _ {N _ {T + t}} = \frac {1}{N _ {T}} \mathbf {1} _ {N _ {T}}, \pi^ {T} \mathbf {1} _ {N _ {T}} = \frac {1}{N _ {T + t}} \mathbf {1} _ {N _ {T + t}}, \\ \end{array} +$$ + +where $\mathbf{1}_{N_T} \in \mathbb{R}^{N_T}$ and $\mathbf{1}_{N_{T + t}} \in \mathbb{R}^{N_{T + t}}$ are vectors with all elements are 1. $C$ is the transport cost matrix and $\pi^T$ is the transpose of $\pi$ . This problem can be solved by the Sinkhon + +![](images/2adb0d29d71a5795688de1f5d6d89a768ff3f99602cfa1f4a5a880c11068717a.jpg) +Figure 3. Prediction results with inconsistency. (a) Ground truth and pseudo labels. (b) Predictions from the same object are inconsistent. (c) Predictions for consecutive future timestamps are inconsistent. (d) Forward and backward predictions (e.g., $M^{T \to T + 1}$ and $M^{T \to T - 1}$ ) are inconsistent. Inconsistency regions are highlighted by red circles. The blue arrow denotes the future displacement (motion). + +algorithm [4]. We use a pre-warping operation to warp the source BEV cells $B^{T}$ by the predicted motion $M^{T\rightarrow T + t}$ , and find the correspondence between the warped BEV cells $\hat{B}^T = B^T +M^{T\rightarrow T + t}$ and target $B^{T + t}$ to calculate $C$ . In this way, with better predictions, the quality of pseudo labels keeps improving during training. The cost matrix $C$ is obtained by computing the pairwise distance between coordinates: + +$$ +C _ {i j} = 1 - \exp \left(- \frac {\left| \left| \hat {b} _ {i} ^ {T} - b _ {j} ^ {T + t} \right| \right| ^ {2}}{\theta_ {c}}\right), \tag {3} +$$ + +where $\theta_{c}$ is a temperature parameter. According to Eq.1, the pseudo motion labels can be obtained by: + +$$ +\hat {M} ^ {T \rightarrow T + t} = \pi^ {*} B ^ {T + t} - B ^ {T}. \tag {4} +$$ + +With the pseudo motion labels, the supervised training loss can be computed by: + +$$ +\mathcal {L} _ {s u p} = \sum_ {t = 1} ^ {T ^ {\prime}} \operatorname {s m o o t h} _ {L 1} \left(M ^ {T \rightarrow T + t}, \hat {M} ^ {T \rightarrow T + t}\right) \tag {5} +$$ + +During pseudo label generation, we first employ a training-free ground segmentation algorithm [12] to remove the majority of ground plane points to reduce the negative influence of ground points for matching and set the motion labels of these removed points as zeros. $\theta_{c}$ is set as 3 in the paper. + +# 3.4. Spatial-Temporal Consistency Regularizations + +# 3.4.1 Motivations + +Due to factors such as occlusion, ground points, and sensors moving, establishing correct correspondences between source and target BEV maps is challenging. Despite removing ground points by the ground segmentation algorithm, noises without correspondences remain. The existence of these noises and cells without correspondences significantly impairs the solution to optimal transport problems. Consequently, the generated pseudo labels are inaccurate, leading to suboptimal outcomes. + +For instance, certain cells corresponding to static background regions may get apparent moving labels (e.g., circle A in Fig. 3 (a)), while cells associated with moving objects might possess static or inaccurate moving labels (e.g., circle B in Fig. 3 (a)). Training the model only using such low-quality pseudo labels yields unsatisfactory results including the spatial and temporal inconsistent predictions, as shown in Fig. 3 (b)(c)(d). Therefore, we introduce three spatial and temporal consistency regularization terms. These regularization terms aim to penalize inaccurate predictions and propagate accurate ones. And through the pre-warping operation, the quality of the pseudo labels can improve iteratively as the accuracy of predictions increases. + +# 3.4.2 Spatial cluster consistency regularization. + +In real-world traffic scenarios, the majority of objects, such as cars and trucks, exhibit a high degree of rigidity. It is intuitive to expect that cells belonging to the same rigid object should display consistent predicted motions. Unfortunately, the generated pseudo labels for cells within the same rigid object are inconsistent (e.g., circle B in Fig. 3 (a)). Consequently, directly learning from these pseudo labels may result in situations like half of the object being predicted with moving status but the other half being predicted with static status (e.g., circle B in Fig. 3 (b)). Therefore, encouraging the consistency of predicted motion within the same object becomes an important factor in enhancing the performance of the self-supervised model. + +Due to the unavailability of instance segmentation labels, one potential solution is to employ K-Nearest-Neighbors (KNN) consistency loss [32]. Although this loss helps encourage local smoothness, it does not consider instance-level consistency, especially in the case of large instances. To this end, we utilize cluster consistency loss to perform the spatial consistency constraint. As cells belonging to the same object are often in close proximity to one another in the virtual BEV map, objects can be coarsely represented as compact clusters. Therefore, we first utilize a breadth-first clustering algorithm to cluster every single cell within + +a distance $d_{c}$ (The detailed clustering process is in the Supplementary). Then for each cluster, we encourage spatial consistency. As illustrated in Sec. 3.3, we perform clustering on 2D BEV maps to reduce the computational cost. The cluster consistency loss is formulated as: + +$$ +\mathcal {L} _ {c} = \frac {1}{| S |} \sum_ {s _ {i} \in S} \frac {1}{| s _ {i} | ^ {2}} \sum_ {b _ {i}, b _ {j} \in s _ {i}} \left(\left| \left| \overrightarrow {\mathcal {M}} \left(b _ {i}\right) - \overrightarrow {\mathcal {M}} \left(b _ {j}\right) \right| \right|\right), \tag {6} +$$ + +where $S = \{s_i\}_{i=1}^{N_s}$ , $s_i$ indicates a cluster, $N_s$ is the cluster numbers in the BEV map and $\overrightarrow{\mathcal{M}}(b)$ is the motion field at cell $b$ . $d_c$ is set as 3 in the paper. + +# 3.4.3 Temporal forward consistency regularization. + +From supervised loss $\mathcal{L}_{sup}$ , the motions are learned individually from the pseudo labels for each timestamp. However, since the foreground and noise distributions of future BEV maps are different (e.g. $B^{T + 1}$ and $B^{T + 2}$ ), we will get inconsistent pseudo motion labels between $\hat{M}^{T\rightarrow T + 1}$ and $\hat{M}^{T\rightarrow T + 2}$ . As the model learns from these inconsistent pseudo labels, it will produce inconsistent predictions. For example, the predictions of static cells in $M^{T\rightarrow T + 1}$ are wrong but correct in $M^{T\rightarrow T + 2}$ (circle A in Fig. 3 (c)). As the object displacements in two consecutive short time windows are supposed to be consistent, we regularize forward predictions as: + +$$ +\mathcal {L} _ {f} = \sum_ {t = 1} ^ {T ^ {\prime} - 1} \operatorname {s m o o t h} _ {L 1} \left(M ^ {T \rightarrow T + t}, \frac {t}{t + 1} M ^ {T \rightarrow T + t + 1}\right). \tag {7} +$$ + +With forward consistency loss $\mathcal{L}_f$ , the predicted motions of adjacent timestamps act as mutual regularization, which is important for self-supervised motion learning as the pseudo labels are temporally inconsistent. For example, $M^{T\rightarrow T + 2}$ can provide more accurate information for $M^{T\rightarrow T + 1}$ in Fig. 3 (c) case. + +# 3.4.4 Temporal backward consistency regularization. + +When generating pseudo labels, it is inevitable to assign noise and some ground source cells to the wrong target cells due to the absence of corresponding cells in the target point clouds. Wrong pseudo labels consequently lead the trained model to predict incorrect motion, especially wrongly predicting large displacements for background cells (e.g. circle A in left Fig. 3 (d)). This creates a contradiction between the model's expected behavior (which is to predict static motion for background cells based on the past frames) and the actual predictions learned directly from incorrect pseudo labels. Inspired by the contradiction, we propose backward consistency loss to penalize incorrect predictions. + +Specifically, besides forward sequence $\vec{\nu} = \{V^1,V^2,\dots,V^T\}$ , we also feed additional backward + +sequence $\overleftarrow{V} = \{V^{2T - 1}, V^{2T - 2}, \dots, V^T\}$ to the network to predict the backward motion $\overleftarrow{\mathcal{M}}$ . Compared to the forward sequence, the backward sequence has entirely different past frames. During training, without being directly constrained by pseudo labels like the forward motion, the backward motion reflects the model's general ability to perceive motion cues hidden in the temporal sequences. Considering that the forward and backward speeds of an object are opposites at the same moment, we formulate the backward regularization term as: + +$$ +\mathcal {L} _ {b} = \sum_ {t = 1} ^ {T ^ {\prime}} \exp \left(- \frac {t}{\theta_ {b}}\right) \text {s m o o t h} _ {L 1} \left(M ^ {T \rightarrow T + t}, - M ^ {T \rightarrow T - t}\right), \tag {8} +$$ + +where $\theta_{b}$ is the temperature parameter. We use the $\exp \left(-\frac{t}{\theta_b}\right)$ term to dynamically adjust the loss weights, taking into account that as the time horizon increases, accurately predicting motion becomes more challenging, and the forward-backward divergence becomes less reliable. This inconsistency becomes noticeable in cells with incorrect predictions (Fig. 4), which demonstrates that backward regularization can effectively reflect the accuracy of the predictions. Thus, the backward regularization $\mathcal{L}_b$ generally penalizes incorrect predictions while having less impact on correct ones. $\theta_{b}$ is set as 10 in the paper. In Sec. 4, we will show that the backward regularization effectively benefits the self-supervised motion learning process. + +Overall loss functions. To summarize, the overall training objective is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {s u p}} + \alpha \mathcal {L} _ {c} + \beta \mathcal {L} _ {f} + \gamma \mathcal {L} _ {b}, \tag {9} +$$ + +where $\alpha, \beta$ , and $\gamma$ are hyper-parameters, balancing the contribution of the different loss terms. It is worth noting that all the regularization terms are only used for training. + +# 4. Experiments + +In this section, we begin by comparing our methods with state-of-the-art (SOTA) motion prediction methods. Subsequently, we conduct ablation studies to analyze the effectiveness of each individual component. + +Dataset. We mainly evaluate the proposed approach on the large-scale self-driving dataset: nuScenes [2], which contains 850 scenes with annotations. For fair comparisons, we follow previous works [19, 30, 31] to use 500 scenes for training, 100 scenes for validation, and 250 scenes for testing. + +Implementation details. We follow the same pre-process setting in [31], where the point clouds are cropped in the range of $[-32\mathrm{m}, 32\mathrm{m}] \times [-32\mathrm{m}, 32\mathrm{m}] \times [-3\mathrm{m}, 2\mathrm{m}]$ and the voxel size is set to $0.25\mathrm{m} \times 0.25\mathrm{m} \times 0.4\mathrm{m}$ along XYZ axis. We take the current sweep and the past four sweeps as input. We use MotionNet as the baseline model for its effectiveness and flexibility following [15, 19]. We train the + +Table 1. Performance comparison on nuScenes dataset. Full., Weak., and Self. refer to fully-supervised, weakly-supervised, and selfsupervised training, respectively. Our approach surpasses state-of-the-arts by a large margin. + +
MethodSup.ModalityYearsStaticSpeed ≤ 5m/s (Slow)Speed ≥ 5m/s (Fast)
Mean↓Median↓Mean↓Median↓Mean↓Median↓
Static-LiDAR-000.61110.09718.65178.1412
FlowNet3D [18]Full.LiDARCVPR20180.041000.81830.17828.52618.0230
HPLFlowNet [8]Full.LiDARCVPR20190.00410.00020.44580.09604.32062.4881
LSTM-ED [25]Full.LiDARICRA20190.035800.35510.10441.58851.0003
MotionNet [31]Full.LiDARCVPR20200.025600.25650.09621.07440.7332
BE-STI [30]Full.LiDARCVPR20220.022000.21150.09290.75110.5413
WeakMotion [15]Weak.LiDARCVPR20230.055800.43370.13051.78231.0887
PointPWC [32]Self.LiDARECCV20200.526400.94230.32253.95302.5149
RigidFlow [14]Self.LiDARCVPR20220.636101.18360.56663.94392.5781
PillarMotion [19]Self.LiDAR+ImageCVPR20210.16200.00100.69720.17583.55042.0844
WeakMotion [15]Self.LiDARCVPR20230.191900.45100.11502.97151.8822
OursSelf.LiDAR-0.041900.32130.10612.29431.0508
+ +model for 20 epochs and set the learning rate to be 0.002. Adam is used as the optimizer. We implement our model in Pytorch [22] with a single RTX 3090 GPU. Hyperparameters $\alpha$ , $\beta$ , and $\gamma$ are set as 0.05, 0.1, and 1 respectively. Following previous works, we take the sequence with 5 frames ( $T = 5$ ) as input and show the results of the next 1s ( $T' = 5$ ) in the following tables. The time interval between each frame is 0.2s. + +Evaluation metrics. Following previous works [19, 30, 31], we evaluate the mean and median errors at different speed levels by dividing the grid cells into 3 groups according to the ground truth speeds: static, slow ( $\leq 5\mathrm{m/s}$ ) and fast ( $\geq 5\mathrm{m/s}$ ). Errors are measured by $L_{2}$ distances between the predicted displacements and the ground truth displacements. + +# 4.1. Comparison with State-of-the-Art Methods + +In Table 1, we show the results of the following methods: (1) A static model that assumes a static environment. (2) Scene flow estimation methods, including FlowNet3D [18], HPLFlowNet [8], PointPWC [32], and RigidFlow [14]. Following [31], we employ scene flow estimation methods by assuming linear dynamics. First, we predict the motion by estimating the flow from the current time to the past 0.4s. Then, we project the predicted flow onto the BEV map and to the future 1.0s for comparison. (3) Class-agnostic motion prediction methods, including LSTM-ED [25], MotionNet [31], PillarMotion [19], BE-STI [30], and WeakMotionNet [15]. + +Compared to fully-supervised models, our approach outperforms FlowNet3D and HPLFlowNet at both slow and fast speed levels, and it surpasses the weakly-supervised WeakMotionNet (with $1\%$ foreground and background masks [15]) at static and slow speed levels. In comparison with self-supervised methods, our approach achieves + +superior performance to PillarMotion, with significant improvements of $74.1\% (-0.1201\mathrm{m})$ , $53.9\%$ $(-0.3759\mathrm{m})$ , and $35.4\% (-1.2561\mathrm{m})$ at static, slow, and fast speed levels, respectively. To further show the superiority of the proposed approach, we implement recent WeakMotionNet in a self-supervised manner by replacing its pre-segmentation network with the same ground segmentation algorithm [12] used in our approach. While self-supervised WeakMotionNet outperforms PillarMotion at slow and fast speed levels, our approach still surpasses it by a large margin. + +# 4.2. Ablation Studies + +In this subsection, we evaluate the effectiveness of each component of our approach. As shown in Table 2, the base model trained only with pseudo labels does not perform well at all speed levels (Row 1). + +Effectiveness of cluster consistency regularization. When cluster consistency loss is not applied, the model exhibits noticeably poorer performance at the fast speed level (Compare Row 1 and Row 6). This disparity in performance can be attributed to the way the cost matrix for the optimal transport problem is computed based on the distance between cells. Consequently, the solver tends to find correspondences between cells that are in close proximity to each other. While this proximity-based approach has its merits, it presents challenges in generating long displacement pseudo labels that the model can learn from. As a result, during the early stages of training, there is a scarcity of correct pseudo labels for fast motion. By imposing constraints on cluster consistency, the model is encouraged to predict faster motion, even if it corresponds to a short displacement pseudo label. This is possible due to the presence of fast displacement predictions within the same cluster. With larger displacement and pre-warping, the optimal transport solver is more likely to find correct correspondence between two + +Table 2. Ablation study for different combinations of proposed regularization terms. Bold fonts and underlines indicate the best and the second-best performance, respectively. + +
#Row Lsup LcLfLbStaticSlowFast
Mean Error ↓
1✓0.16500.55403.4503
2✓✓0.04130.32492.7587
3✓✓0.11970.42643.2072
4✓✓✓0.04990.32772.5753
5✓✓✓0.05020.34322.5154
6✓✓0.17050.41542.7027
7✓✓✓0.14500.40872.6509
8✓✓✓0.04190.32132.2943
+ +points at long distances. In Table 3, we present the results of KNN consistency loss and cluster consistency loss, both of which function as spatial smooth losses. Compared to $\mathcal{L}_{KNN}$ , $\mathcal{L}_c$ bring more benefits comprehensively. + +Effectiveness of temporal consistency regularization. Both forward and backward regularization terms boost the performances across all speed levels (Compare Row 5 and Row 7 with Row 8 respectively). To further illustrate the impact of backward regularization, we separately feed forward and backward sequences of all training samples into the model only trained with $\mathcal{L}_{sup}$ to demonstrate prediction divergence $(|\overrightarrow{\mathcal{M}} -\overleftarrow{\mathcal{M}} |)$ . As depicted in Fig. 4, there is a positive correlation between the forward-backward prediction divergence and prediction accuracy. The divergence is large when the prediction is inaccurate and vice versa, which indicates that backward regularization will have a stronger impact on inaccurate predictions but a weaker influence on accurate ones. By using the forward-backward divergence as a loss term, the model is penalized for making inaccurate predictions. Results for various $\theta_{b}$ are shown in Table 4, which indicates that the weight term can further improve the performance. + +Qualitative results. We present the result visualizations of our approach in Fig. 5 to qualitatively show the impacts of each regularization term. When directly training with pseudo labels $\mathcal{L}_{sup}$ , the model tends to produce inconsistent predictions for the same object and predict large motions for background cells. The third column illustrates that cluster consistency loss $\mathcal{L}_c$ effectively benefits spatial consistency. With the addition of forward regularization $\mathcal{L}_f$ , the prediction results are further improved, but there are still noticeable noises in the background cells. Finally, by incorporating backward regularization $\mathcal{L}_b$ , the noise predictions in the background cells are significantly suppressed. + +Table 3. Impact of different spatial regularizations. Cluster-level regularization performs better than the KNN-based regularization. Bold fonts and underlines indicate the best and the second-best performance, respectively. + +
MethodStaticSlowFast
Mean Errors ↓
Baseline0.05020.34322.5154
+KLNNK=30.03710.33702.5362
K=50.04900.33782.4568
K=100.07310.34052.5116
+Lc(ours)dc=20.05260.34382.1814
dc=30.04190.32132.2943
dc=40.04590.33182.4434
+ +Table 4. Varying $\theta_{b}$ for backward consistency loss. - indicates the results without the $\exp\left(-\frac{t}{\theta_{b}}\right)$ term. Bold fonts and underlines indicate the best and the second-best performance, respectively. + +
θbStaticSlowFast
Mean Errors ↓
50.04820.32642.3580
100.04190.32132.2943
150.02930.30912.4149
-0.03820.32982.4596
+ +![](images/e27d30d153bcee99a0f3e7f3fe3620f0aed3505263e569ca21cc2dc8a2122d2f.jpg) +Figure 4. Forward-backward divergence of the training set. We analyze the prediction errors of all samples in the training set and their corresponding forward-backward divergences. The Forward-backward divergence is positively correlated with the prediction error. + +# 4.3. Results on Waymo Dataset + +Following WeakMotionNet [15], we further evaluate our approach on the Waymo dataset [27]. As shown in Table 5, similar to the results on the nuScenes dataset, our approach achieves mean errors of around $0.05\mathrm{m}$ , $0.4\mathrm{m}$ , and $2.3\mathrm{m}$ at static, slow, and fast speed levels, respectively, outperforming self-supervised WeakMotionNet. + +![](images/03e6c58ef98f461137f958f125bfc4be64ebe2489e2a7494bf2d6d9436acbc14.jpg) +Figure 5. Qualitative results of the proposed self-supervised approach. The future displacements $(M^{T\rightarrow T + 5})$ are depicted using the color wheel representation. (Zoom in for the best view) + +Table 5. Motion prediction results on the Waymo Dataset. Our method still outperforms self-supervised WeakMotion by a large margin. + +
MethodSup.StaticSlowFast
Mean Errors ↓
MotionNet [31]Full.0.02630.26200.9493
WeakMotion [15]Weak.0.02970.34581.5655
WeakMotion [15]Self.0.13450.58333.0180
OursSelf.0.05150.44402.3692
+ +# 4.4. Discussions + +Ground Segmentation. We evaluate the performance of the ground segmentation algorithm [12] on the nuScenes dataset, and get a $95\%$ precision rate while a $92\%$ recall rate, indicating the majority of ground points can be correctly removed. Although almost all ground points can be successfully removed, the self-supervised training process still suffers from remaining ground points, noises, and points without correspondences. The issues are alleviated using our proposed regularizations, as shown in Fig. 2. + +Clustering. In the paper, we utilize the breadth-first clustering (BFS) algorithm to demonstrate the effectiveness of the cluster consistency loss. We evaluate the clustering results of BEV cells on the nuScenes dataset and observe a $9.6\%$ under-clustering rate and an $11.1\%$ over-clustering rate. It indicates that although effective, the BFS algorithm still wrongly clusters cells from different instances in some cases. Therefore, fellow researchers can consider utilizing or developing more advanced clustering strategies, which + +may lead to better performances. + +Efficiency. During training, ground segmentation and clustering take around 5ms and 10ms, respectively. In practice, these processes can be employed as data pre-processing before training. For the optimal transport (OT) solver, we implement it with Pytorch and operate in batch. It takes less than 1ms to solve the OT problem. During inference, given that our approach is model-independent, the inference time is based on the model used. When applying MotionNet, it takes 10ms for point cloud transformation and voxelization and another 10ms for prediction [31]. + +# 5. Conclusion + +In this paper, we introduce a novel approach for self-supervised motion prediction with only unlabeled LiDAR point clouds. We utilize optimal transport to generate pseudo labels and propose three novel spatial and temporal regularization terms to facilitate self-supervised motion learning. Extensive experiments on the large-scale nuScenes dataset illustrate that our approach significantly surpasses previous self-supervised methods and shallows the gap with fully-supervised ones. + +Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and the MoE AcRF Tier 1 grant (RG14/22). + +# References + +[1] Aseem Behl, Despoina Paschalidou, Simon Donne, and Andreas Geiger. Pointflownet: Learning representations for rigid motion estimation from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5 +[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2 +[4] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Neural Information Processing Systems (NeurIPS), 2013. 4 +[5] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2 +[6] Guanting Dong, Yueyi Zhang, Hanlin Li, Xiaoyan Sun, and Zhiwei Xiong. Exploiting rigidity constraints for lidar scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[7] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2 +[8] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6 +[9] Marcel Van Herk. Image registration using chamfer matching. Handbook of Medical Imaging, 2000. 2 +[10] Yair Kittenplon, Yonina C Eldar, and Dan Raviv. Flow-step3d: Model unrolling for self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[12] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 3, 4, 6, 8 + +[13] Ruibo Li, Guosheng Lin, and Lihua Xie. Self-point-flow: Self-supervised scene flow estimation from point clouds with optimal transport and random walk. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[14] Ruibo Li, Chi Zhang, Guosheng Lin, Zhe Wang, and Chunhua Shen. Rigidflow: Self-supervised scene flow learning on point clouds by local rigidity prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6 +[15] Ruibo Li, Hanyu Shi, Ziang Fu, Zhe Wang, and Guosheng Lin. Weakly supervised class-agnostic motion prediction for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 7, 8 +[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Neural Information Processing Systems (NeurIPS), 2021. 2 +[17] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2 +[18] Xingyu Liu, C. Qi, and Leonidas J. Guibas. Flownet3d: Learning scene flow in 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6 +[19] Chenxu Luo, Xiaodong Yang, and Alan Loddon Yuille. Self-supervised pillar motion learning for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5, 6 +[20] Wenjie Luo, Binh Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[21] Himangi Mittal, Brian Okorn, and David Held. Just go with the flow: Self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 6 +[23] Jhony Kaesemodel Pontes, James Hays, and Simon Lucey. Scene flow from point clouds with or without learning. In International Conference on 3D Vision (3DV), 2020. 2 +[24] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision (ECCV), 2020. 2 +[25] Marcel Schreiber, Stefan Hoermann, and Klaus Dietmayer. Long-term occupancy grid prediction using recurrent neural networks. In IEEE international conference on robotics and automation (ICRA), 2019. 1, 2, 6 +[26] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang chun Woo. Convolutional LSTM network: A machine learning approach for precipita + +tion nowcasting. In Neural Information Processing Systems (NeurIPS), 2015. 2 +[27] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7 +[28] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. Hierarchical attention learning of scene flow in 3d point clouds. IEEE Transactions on Image Processing (TIP), 2021. 2 +[29] Kewei Wang, Yizheng Wu, Zhiyu Pan, Xingyi Li, Ke Xian, Zhe Wang, Zhiguo Cao, and Guosheng Lin. Semi-supervised class-agnostic motion prediction with pseudo label regeneration and bevmix, 2023. 1 +[30] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6 +[31] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 5, 6, 8 +[32] Wenxuan Wu, Zhi Yuan Wang, Zhuwen Li, Wei Liu, and Li Fuxin. Pointpwc-net: Cost volume on point clouds for (self-) supervised scene flow estimation. In European Conference on Computer Vision (ECCV), 2020. 2, 4, 6 +[33] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In IEEE Conference on International Conference on Computer Vision (ICCV), 2023. 2 +[34] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yun-ting Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. Tnt: Target-driven trajectory prediction. In Conference on Robot Learning (CoRL), 2020. 1, 2 +[35] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 \ No newline at end of file diff --git a/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/images.zip b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..61c02b94edf23a8e994263f6a01385ac6bceaad7 --- /dev/null +++ b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c20ca1478c8eb18d34ed773fd8319303447cf2bc4bbf6593e29e5416a97fd46c +size 489278 diff --git a/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/layout.json b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..26e88a8f5fa81740e3b13739c59a499f733b10b1 --- /dev/null +++ b/2024/Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations/layout.json @@ -0,0 +1,8105 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "text", + "content": "Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": "Kewei Wang" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": ", Yizheng Wu" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": ", Jun CEN" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": ", Zhiyu Pan" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": ", Xingyi Li" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": ", Zhe Wang" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": ", Zhiguo Cao" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": ", Guosheng Lin" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": " School of AIA, Huazhong University of Science and Technology \n" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": " S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 159, + 591, + 204 + ], + "type": "text", + "content": " SenseTime Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "type": "text", + "content": "The perception of motion behavior in a dynamic environment holds significant importance for autonomous driving systems, wherein class-agnostic motion prediction methods directly predict the motion of the entire point cloud. While most existing methods rely on fully-supervised learning, the manual labeling of point cloud data is laborious and time-consuming. Therefore, several annotation-efficient methods have been proposed to address this challenge. Although effective, these methods rely on weak annotations or additional multi-modal data like images, and the potential benefits inherent in the point cloud sequence are still underexplored. To this end, we explore the feasibility of self-supervised motion prediction with only unlabeled Li-DAR point clouds. Initially, we employ an optimal transport solver to establish coarse correspondences between current and future point clouds as the coarse pseudo motion labels. Training models directly using such coarse labels leads to noticeable spatial and temporal prediction inconsistencies. To mitigate these issues, we introduce three simple spatial and temporal regularization losses, which facilitate the self-supervised training process effectively. Experimental results demonstrate the significant superiority of our approach over the state-of-the-art self-supervised methods. Code will be available at https://github.com/kwwcv/SelfMotion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 600, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 685 + ], + "type": "text", + "content": "Motion perception provides key information for autonomous driving. Classical approaches formulate the motion perception task as object-level trajectory prediction [3, 5, 7, 17, 34]. However, these approaches may fail if agents are not successfully detected by the detectors, especially when encountering unseen categories in open-set scenarios [31]. To compensate for trajectory prediction and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 232, + 382, + 312 + ], + "blocks": [ + { + "bbox": [ + 309, + 232, + 382, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 232, + 382, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 232, + 382, + 312 + ], + "type": "image", + "image_path": "ce930b58cf4438ef67882a4cc13a3aee263dac2a39d9c54cd32655cedbb0fed1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 316, + 361, + 324 + ], + "lines": [ + { + "bbox": [ + 331, + 316, + 361, + 324 + ], + "spans": [ + { + "bbox": [ + 331, + 316, + 361, + 324 + ], + "type": "text", + "content": "(a) Static" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 233, + 462, + 312 + ], + "blocks": [ + { + "bbox": [ + 389, + 233, + 462, + 312 + ], + "lines": [ + { + "bbox": [ + 389, + 233, + 462, + 312 + ], + "spans": [ + { + "bbox": [ + 389, + 233, + 462, + 312 + ], + "type": "image", + "image_path": "3861d21624ce41455162fd11497d55066089487ded343635fecb8763b38fb167.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 316, + 441, + 324 + ], + "lines": [ + { + "bbox": [ + 411, + 316, + 441, + 324 + ], + "spans": [ + { + "bbox": [ + 411, + 316, + 441, + 324 + ], + "type": "text", + "content": "(b) Slow" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 470, + 233, + 541, + 312 + ], + "blocks": [ + { + "bbox": [ + 470, + 233, + 541, + 312 + ], + "lines": [ + { + "bbox": [ + 470, + 233, + 541, + 312 + ], + "spans": [ + { + "bbox": [ + 470, + 233, + 541, + 312 + ], + "type": "image", + "image_path": "08ee62acdb6ef05645e69957b99eef61d44dd42491849ff918e08dceb654dd3c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 316, + 519, + 324 + ], + "lines": [ + { + "bbox": [ + 493, + 316, + 519, + 324 + ], + "spans": [ + { + "bbox": [ + 493, + 316, + 519, + 324 + ], + "type": "text", + "content": "(c) Fast" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 334, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 334, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 547, + 411 + ], + "type": "text", + "content": "Figure 1. Performance comparison over static, slow, and fast speed levels between self-supervised PillarMotion [19] and our approach on the nuScenes dataset. The dashed line represents the performance of fully-supervised MotionNet [31]. Our proposed self-supervised approach outperforms the PillarMotion which uses additional image data by a large margin and substantially narrows the performance gap with fully-supervised results." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 425, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 473 + ], + "type": "text", + "content": "provide backup, class-agnostic motion prediction was studied [15, 25, 29-31]. Currently, these approaches predict the motions of the entire environment from point clouds directly instead of for each agent." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 474, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 548, + 715 + ], + "type": "text", + "content": "While fully-supervised class-agnostic motion prediction methods have shown success, they typically require a vast amount of annotated point cloud data, which can be both costly and time-consuming to acquire. To overcome this limitation, WeakMotionNet [15] proposes a weakly supervised approach using only a limited number of foreground-background semantic annotations. WeakMotionNet also introduces a consistency-aware chamfer loss that uses the difference between correspondences from different frames as loss weights. Although effective, it requires establishing correspondences multiple times and is still based on matching results between individual frames, rather than taking advantage of the temporal information within the input sequences. Without using any annotations, PillarMotion [19] introduces a self-supervised training approach that incorporates regularization of 2D and 3D flows derived from paired image and point cloud data. However, PillarMotion relies on multi-modality pairs to regularize motion learning, making the training process less flexible, and a considerable performance gap still exists between self-supervised" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 222, + 713 + ], + "type": "text", + "content": "*Corresponding author. gslin@ntu.edu.sg, {wangkewei, zgcao} @hust.edu.cn," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14638" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "and fully-supervised methods. Therefore, we explore the spatial and temporal information from the point cloud sequences themselves to train a better class-agnostic motion prediction model with only unlabeled LiDAR data in this paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 383 + ], + "type": "text", + "content": "Our approach primarily involves pseudo label generation and spatial and temporal regularizations. To generate pseudo motion labels, we first solve a matching problem to find correspondence between current and future point clouds. However, the correctness of the matching correspondence cannot be guaranteed due to the occlusion, background artifacts, noise, etc. Self-supervised learning from these coarse matching correspondences (pseudo labels) leads to two major failure cases: (1) predicted motions from the same rigid object are spatially diverse; (2) the background cells that should be static are predicted with large motions. To alleviate these issues, we introduce three simple but effective spatial and temporal regularizations from the point cloud motion prediction itself to facilitate self-supervised learning. Specifically, cluster consistency regularization, backward consistency regularization, and forward consistency regularization are proposed to encourage spatial consistency, penalize incorrect predictions (e.g., large motion for static cells), and smooth the prediction learned from pseudo labels of different time horizons, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 477 + ], + "type": "text", + "content": "Experiments demonstrate that our proposed spatial and temporal consistency regularization terms work effectively together, significantly boosting the performance of self-supervised motion prediction. The proposed approach turns out with excellent performance, as illustrated in Fig. 1. On the nuScenes dataset, our approach outperforms previous methods by a substantial margin. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 479, + 287, + 597 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 479, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 526 + ], + "type": "text", + "content": "- We propose a novel approach for self-supervised class-diagnostic motion prediction that incorporates an optimal transport solver and three simple yet effective spatial and temporal regularization losses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 527, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 287, + 550 + ], + "type": "text", + "content": "- Our approach is model-independent, and does not require any additional models or multi-modal data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 551, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 287, + 597 + ], + "type": "text", + "content": "- Our method surpasses the state-of-the-art (SOTA) self-supervised methods by a large margin and significantly reduces the performance gap with the fully-supervised methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 609, + 138, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 138, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 138, + 621 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Class-agnostic motion prediction. Classical motion prediction approaches aim to predict the future trajectories of agents based on past observations. The entire system typically involves object detection [11, 20, 35], tracking, and forecasting [3, 5, 7, 17, 33, 34]. However, the trajectory prediction relies on the object detector which may fail to detect objects with unknown classes in open-set scenarios [31]. To" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "enhance system redundancy, especially in cases where objects may be not successfully detected and potentially lead to a failure in object-level trajectory forecasting [31], class-agnostic motion prediction methods predict the motion of entire scenes directly instead of individually for each agent. These methods represent the environment state based on BEV maps discretized from point clouds and aim to predict the 2D displacement vector for each BEV cell along the ground plane. MotionNet [31] is one of the pioneers in this task, proposing to perform joint perception and motion prediction based on a spatial-temporal pyramid network. LSTM-ED [25] introduce convolutional LSTM [26] to aggregate temporal context. BE-STI [30] leverages semantic cues by training an additional semantic decoder to guide motion prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 251, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 395 + ], + "type": "text", + "content": "To alleviate the reliance on manual annotations, WeakMotionNet [15] proposes to train the MotionNet in a weakly-supervised manner with limited foreground and background semantic annotations. PillarMotion [19] proposes to train the model in a self-supervised manner without any annotations by cross-sensor motion regularization. Although effective, a substantial performance gap persists between the self-supervised PillarMotion and the fully-supervised model even using additional image data. Therefore, an open question remains about whether we can more efficiently train a self-supervised motion prediction model with higher performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 396, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 514 + ], + "type": "text", + "content": "Scene flow estimation. Scene flow estimation [1, 8, 18, 28, 32] from point clouds is an alternative way to understand the dense class-agnostic 3D motion field. In self-supervised scene flow estimation, several methods rely on establishing point correspondences between source and target point clouds through point matching. They treat the 3D coordinate difference between each matching pair as a pseudo scene flow label [10, 13, 14, 21, 24, 32]. Many methods also utilize the chamfer distance loss [9] to drive self-supervised scene flow learning [6, 10, 16, 23, 32]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "content": "While similar to the class-agnostic motion prediction task, scene flow estimation majorly differs in the following ways: (1) The objectiveness of the two tasks is different [15, 31]. Scene flow estimation methods focus on estimating the motion flow between two observed point clouds, while motion prediction methods aim to forecast future displacements based on past observations; (2) Estimating dense 3D flow is time-consuming, which is not feasible for self-driving systems [15]. Furthermore, directly applying scene flow estimation on real LiDAR point clouds poses challenges due to the lack of one-to-one correspondences [19, 30]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this section, we begin by introducing the problem formulation of class-agnostic motion prediction. Subsequently," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14639" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 539, + 247 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 539, + 247 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 539, + 247 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 539, + 247 + ], + "type": "image", + "image_path": "0193a72f083ec9645e4381c1c8f3cfe0aff63792de326a79e20cf656354cdba3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 278 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed approach. Without ground truth labels, we first generate pseudo labels by matching. We then introduce cluster, forward, and backward regularization losses to facilitate self-supervised motion learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 324 + ], + "type": "text", + "content": "we delve into a comprehensive presentation and discussion of each component comprising our approach. The overview of our approach is illustrated in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 331, + 171, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 171, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 171, + 342 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": "Given a temporal sequence of LiDAR point cloud frames, past frames are first synchronized to the current coordinate system [31]. We denote synchronized point cloud captured at time " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "P^t" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": ". Following [31], we discretized " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "P^t" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " into dense voxels " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "V^t \\in \\{0,1\\}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": ", where 0 indicates the voxel is empty, 1 indicates the voxel is occupied by at least 1 point, and " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "H, W, C" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " are the voxel numbers along " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "X, Y, Z" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " axis respectively. Then we represent the 3D voxel lattice as a 2D pseudo-image with the " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " corresponding to the image channel and view " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "V^t" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " as a virtual BEV map. Then the motion field " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " in the BEV map is defined as the 2D displacement of each grid cell to the future timestamp. Taking sequence " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "\\vec{\\mathcal{V}} = \\{V^t\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " as input, the task aims to predict the motion fields " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "\\vec{\\mathcal{M}} = \\{M^T \\rightarrow T + t\\}_{t=1}^{T'}" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 46, + 349, + 290, + 545 + ], + "type": "text", + "content": " are the number of past and future timestamps, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 552, + 115, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 115, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 115, + 562 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "As depicted in Fig. 2, our approach mainly involves pseudo label generation and self-supervised training. To generate pseudo labels, we initially solve the matching problem between two consecutive frames. Firstly, we remove the majority of ground plane points from point clouds by ground segmentation algorithm [12] to reduce noise for better matching. Subsequently, we feed the raw BEV map sequence into MotionNet [31] and train the model using the generated pseudo labels through supervised loss " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". As the quality of pseudo labels is low at the initial stage of training, we introduce three spatial and temporal consistency regularization losses, named cluster consistency loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "text", + "content": ", forward consistency loss " + }, + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "text", + "content": ", and backward consistency loss " + }, + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "text", + "content": ", to facilitate the self-supervised learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 319, + 448, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 448, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 448, + 331 + ], + "type": "text", + "content": "3.3. Pseudo Label Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": "Unlike fully-supervised learning, where ground truth labels are available, self-supervised methods learn from data itself without annotations. In this condition, one alternative approach is to generate pseudo labels to provide the necessary supervision for training. Considering motion can be defined as the variation in coordinates between correspondent points in two consecutive frames, the process of generating pseudo labels can be viewed as a matching problem. We denote " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T} = \\{b_{i}^{T}\\in \\mathbb{R}^{2}\\}_{i = 1}^{N_{T}}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": " as the 2D coordinates of BEV cells on the current virtual BEV map " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "V^T" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": " is the number of non-empty cells. Then we perform matching between source " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T + t} = \\{b_j^{T + t}\\in \\mathbb{R}^2\\}_{j = 1}^{N_{T + t}}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": ". Ideally, if there exists one-to-one correspondence between " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T + t}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": " can be projected into " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T + t}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": " and fully occupy " + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "inline_equation", + "content": "B^{T + t}" + }, + { + "bbox": [ + 304, + 337, + 546, + 518 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 383, + 525, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 525, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 383, + 525, + 545, + 539 + ], + "type": "interline_equation", + "content": "B ^ {T} + M = \\pi B ^ {T + t}, \\tag {1}", + "image_path": "44155871ddce58adca124807ca7fe8e10e42f8ce887bdecf3d23db59fc407c54.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "inline_equation", + "content": "\\pi \\in \\{0,1\\}^{N_T\\times N_{T + t}}" + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "text", + "content": " is a permutation matrix indicating correspondence between " + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "inline_equation", + "content": "B^T" + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "inline_equation", + "content": "B^{T + t}" + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "text", + "content": " is the motion flow. The optimal soft permutation matrix " + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "inline_equation", + "content": "\\pi^{*}\\in [0,1]^{N_{T}\\times N_{T + t}}" + }, + { + "bbox": [ + 304, + 547, + 545, + 608 + ], + "type": "text", + "content": " can be computed by solving an optimal transport problem:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 616, + 524, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 616, + 524, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 616, + 524, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\pi^ {*} = \\underset {\\pi} {\\arg \\min } \\sum_ {i, j} C _ {i j} \\pi_ {i j} \\\\ s. t. \\quad \\pi \\mathbf {1} _ {N _ {T + t}} = \\frac {1}{N _ {T}} \\mathbf {1} _ {N _ {T}}, \\pi^ {T} \\mathbf {1} _ {N _ {T}} = \\frac {1}{N _ {T + t}} \\mathbf {1} _ {N _ {T + t}}, \\\\ \\end{array}", + "image_path": "d4e7ad128af482825826bcf7273e923b3812dc05a528c56f65d1b542a98c2955.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{N_T} \\in \\mathbb{R}^{N_T}" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{N_{T + t}} \\in \\mathbb{R}^{N_{T + t}}" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " are vectors with all elements are 1. " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " is the transport cost matrix and " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pi^T" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " is the transpose of " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": ". This problem can be solved by the Sinkhon" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14640" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 175 + ], + "type": "image", + "image_path": "2adb0d29d71a5795688de1f5d6d89a768ff3f99602cfa1f4a5a880c11068717a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "text", + "content": "Figure 3. Prediction results with inconsistency. (a) Ground truth and pseudo labels. (b) Predictions from the same object are inconsistent. (c) Predictions for consecutive future timestamps are inconsistent. (d) Forward and backward predictions (e.g., " + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "inline_equation", + "content": "M^{T \\to T + 1}" + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "inline_equation", + "content": "M^{T \\to T - 1}" + }, + { + "bbox": [ + 46, + 182, + 547, + 216 + ], + "type": "text", + "content": ") are inconsistent. Inconsistency regions are highlighted by red circles. The blue arrow denotes the future displacement (motion)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": "algorithm [4]. We use a pre-warping operation to warp the source BEV cells " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "B^{T}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " by the predicted motion " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "M^{T\\rightarrow T + t}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ", and find the correspondence between the warped BEV cells " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\hat{B}^T = B^T +M^{T\\rightarrow T + t}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "B^{T + t}" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " to calculate " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": ". In this way, with better predictions, the quality of pseudo labels keeps improving during training. The cost matrix " + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 236, + 287, + 331 + ], + "type": "text", + "content": " is obtained by computing the pairwise distance between coordinates:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 336, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 336, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 94, + 336, + 287, + 365 + ], + "type": "interline_equation", + "content": "C _ {i j} = 1 - \\exp \\left(- \\frac {\\left| \\left| \\hat {b} _ {i} ^ {T} - b _ {j} ^ {T + t} \\right| \\right| ^ {2}}{\\theta_ {c}}\\right), \\tag {3}", + "image_path": "b8745ad165f9f74b9342fef626ab99c56a8733a364204584caa84c7d2b720a52.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 369, + 287, + 394 + ], + "type": "inline_equation", + "content": "\\theta_{c}" + }, + { + "bbox": [ + 47, + 369, + 287, + 394 + ], + "type": "text", + "content": " is a temperature parameter. According to Eq.1, the pseudo motion labels can be obtained by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 399, + 287, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 399, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 107, + 399, + 287, + 414 + ], + "type": "interline_equation", + "content": "\\hat {M} ^ {T \\rightarrow T + t} = \\pi^ {*} B ^ {T + t} - B ^ {T}. \\tag {4}", + "image_path": "b25da10d5b4d83a1829e012145e5f38ca3f0a565161333fa4774e09980d39adf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 420, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 445 + ], + "type": "text", + "content": "With the pseudo motion labels, the supervised training loss can be computed by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 452, + 287, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 452, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 73, + 452, + 287, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s u p} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\operatorname {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, \\hat {M} ^ {T \\rightarrow T + t}\\right) \\tag {5}", + "image_path": "51d37d42486706f44c362b11577be7df9744d67aeef7686dff41458c7d328037.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "type": "text", + "content": "During pseudo label generation, we first employ a training-free ground segmentation algorithm [12] to remove the majority of ground plane points to reduce the negative influence of ground points for matching and set the motion labels of these removed points as zeros. " + }, + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "type": "inline_equation", + "content": "\\theta_{c}" + }, + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "type": "text", + "content": " is set as 3 in the paper." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 568, + 287, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 582 + ], + "type": "text", + "content": "3.4. Spatial-Temporal Consistency Regularizations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 586, + 132, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 132, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 132, + 597 + ], + "type": "text", + "content": "3.4.1 Motivations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Due to factors such as occlusion, ground points, and sensors moving, establishing correct correspondences between source and target BEV maps is challenging. Despite removing ground points by the ground segmentation algorithm, noises without correspondences remain. The existence of these noises and cells without correspondences significantly impairs the solution to optimal transport problems. Consequently, the generated pseudo labels are inaccurate, leading to suboptimal outcomes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 236, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 392 + ], + "type": "text", + "content": "For instance, certain cells corresponding to static background regions may get apparent moving labels (e.g., circle A in Fig. 3 (a)), while cells associated with moving objects might possess static or inaccurate moving labels (e.g., circle B in Fig. 3 (a)). Training the model only using such low-quality pseudo labels yields unsatisfactory results including the spatial and temporal inconsistent predictions, as shown in Fig. 3 (b)(c)(d). Therefore, we introduce three spatial and temporal consistency regularization terms. These regularization terms aim to penalize inaccurate predictions and propagate accurate ones. And through the pre-warping operation, the quality of the pseudo labels can improve iteratively as the accuracy of predictions increases." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 407, + 515, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 515, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 515, + 419 + ], + "type": "text", + "content": "3.4.2 Spatial cluster consistency regularization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "In real-world traffic scenarios, the majority of objects, such as cars and trucks, exhibit a high degree of rigidity. It is intuitive to expect that cells belonging to the same rigid object should display consistent predicted motions. Unfortunately, the generated pseudo labels for cells within the same rigid object are inconsistent (e.g., circle B in Fig. 3 (a)). Consequently, directly learning from these pseudo labels may result in situations like half of the object being predicted with moving status but the other half being predicted with static status (e.g., circle B in Fig. 3 (b)). Therefore, encouraging the consistency of predicted motion within the same object becomes an important factor in enhancing the performance of the self-supervised model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Due to the unavailability of instance segmentation labels, one potential solution is to employ K-Nearest-Neighbors (KNN) consistency loss [32]. Although this loss helps encourage local smoothness, it does not consider instance-level consistency, especially in the case of large instances. To this end, we utilize cluster consistency loss to perform the spatial consistency constraint. As cells belonging to the same object are often in close proximity to one another in the virtual BEV map, objects can be coarsely represented as compact clusters. Therefore, we first utilize a breadth-first clustering algorithm to cluster every single cell within" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14641" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "a distance " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "d_{c}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " (The detailed clustering process is in the Supplementary). Then for each cluster, we encourage spatial consistency. As illustrated in Sec. 3.3, we perform clustering on 2D BEV maps to reduce the computational cost. The cluster consistency loss is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 137, + 287, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{| S |} \\sum_ {s _ {i} \\in S} \\frac {1}{| s _ {i} | ^ {2}} \\sum_ {b _ {i}, b _ {j} \\in s _ {i}} \\left(\\left| \\left| \\overrightarrow {\\mathcal {M}} \\left(b _ {i}\\right) - \\overrightarrow {\\mathcal {M}} \\left(b _ {j}\\right) \\right| \\right|\\right), \\tag {6}", + "image_path": "6ee5afa0fd6b3b087ebcc5b2ca499b89155e653116dab45077f4c802383d6295.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "inline_equation", + "content": "S = \\{s_i\\}_{i=1}^{N_s}" + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "content": " indicates a cluster, " + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "content": " is the cluster numbers in the BEV map and " + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\overrightarrow{\\mathcal{M}}(b)" + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "content": " is the motion field at cell " + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "inline_equation", + "content": "d_c" + }, + { + "bbox": [ + 47, + 175, + 287, + 214 + ], + "type": "text", + "content": " is set as 3 in the paper." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 228, + 274, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 274, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 274, + 240 + ], + "type": "text", + "content": "3.4.3 Temporal forward consistency regularization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": "From supervised loss " + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup}" + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": ", the motions are learned individually from the pseudo labels for each timestamp. However, since the foreground and noise distributions of future BEV maps are different (e.g. " + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "inline_equation", + "content": "B^{T + 1}" + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "inline_equation", + "content": "B^{T + 2}" + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": "), we will get inconsistent pseudo motion labels between " + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "inline_equation", + "content": "\\hat{M}^{T\\rightarrow T + 1}" + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "inline_equation", + "content": "\\hat{M}^{T\\rightarrow T + 2}" + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": ". As the model learns from these inconsistent pseudo labels, it will produce inconsistent predictions. For example, the predictions of static cells in " + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "inline_equation", + "content": "M^{T\\rightarrow T + 1}" + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": " are wrong but correct in " + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "inline_equation", + "content": "M^{T\\rightarrow T + 2}" + }, + { + "bbox": [ + 46, + 247, + 288, + 390 + ], + "type": "text", + "content": " (circle A in Fig. 3 (c)). As the object displacements in two consecutive short time windows are supposed to be consistent, we regularize forward predictions as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 396, + 287, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 287, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\sum_ {t = 1} ^ {T ^ {\\prime} - 1} \\operatorname {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, \\frac {t}{t + 1} M ^ {T \\rightarrow T + t + 1}\\right). \\tag {7}", + "image_path": "46810750ae07569fabd8aecab8e1ad114e948dc7a6318d73fe1ff5cac496b2b9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "text", + "content": "With forward consistency loss " + }, + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "text", + "content": ", the predicted motions of adjacent timestamps act as mutual regularization, which is important for self-supervised motion learning as the pseudo labels are temporally inconsistent. For example, " + }, + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "inline_equation", + "content": "M^{T\\rightarrow T + 2}" + }, + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "text", + "content": " can provide more accurate information for " + }, + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "inline_equation", + "content": "M^{T\\rightarrow T + 1}" + }, + { + "bbox": [ + 46, + 440, + 287, + 511 + ], + "type": "text", + "content": " in Fig. 3 (c) case." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 525, + 282, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 282, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 282, + 538 + ], + "type": "text", + "content": "3.4.4 Temporal backward consistency regularization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 687 + ], + "type": "text", + "content": "When generating pseudo labels, it is inevitable to assign noise and some ground source cells to the wrong target cells due to the absence of corresponding cells in the target point clouds. Wrong pseudo labels consequently lead the trained model to predict incorrect motion, especially wrongly predicting large displacements for background cells (e.g. circle A in left Fig. 3 (d)). This creates a contradiction between the model's expected behavior (which is to predict static motion for background cells based on the past frames) and the actual predictions learned directly from incorrect pseudo labels. Inspired by the contradiction, we propose backward consistency loss to penalize incorrect predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 687, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 287, + 714 + ], + "type": "text", + "content": "Specifically, besides forward sequence " + }, + { + "bbox": [ + 47, + 687, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\vec{\\nu} = \\{V^1,V^2,\\dots,V^T\\}" + }, + { + "bbox": [ + 47, + 687, + 287, + 714 + ], + "type": "text", + "content": ", we also feed additional backward" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 70, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 194 + ], + "type": "text", + "content": "sequence " + }, + { + "bbox": [ + 304, + 70, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\overleftarrow{V} = \\{V^{2T - 1}, V^{2T - 2}, \\dots, V^T\\}" + }, + { + "bbox": [ + 304, + 70, + 545, + 194 + ], + "type": "text", + "content": " to the network to predict the backward motion " + }, + { + "bbox": [ + 304, + 70, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\overleftarrow{\\mathcal{M}}" + }, + { + "bbox": [ + 304, + 70, + 545, + 194 + ], + "type": "text", + "content": ". Compared to the forward sequence, the backward sequence has entirely different past frames. During training, without being directly constrained by pseudo labels like the forward motion, the backward motion reflects the model's general ability to perceive motion cues hidden in the temporal sequences. Considering that the forward and backward speeds of an object are opposites at the same moment, we formulate the backward regularization term as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 200, + 545, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 200, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 545, + 243 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\exp \\left(- \\frac {t}{\\theta_ {b}}\\right) \\text {s m o o t h} _ {L 1} \\left(M ^ {T \\rightarrow T + t}, - M ^ {T \\rightarrow T - t}\\right), \\tag {8}", + "image_path": "9fa3d47e17d1f1e2457d77c60737c41d7b750afb7acaf9e5a18717ffed0d3f2c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\theta_{b}" + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "text", + "content": " is the temperature parameter. We use the " + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\exp \\left(-\\frac{t}{\\theta_b}\\right)" + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "text", + "content": " term to dynamically adjust the loss weights, taking into account that as the time horizon increases, accurately predicting motion becomes more challenging, and the forward-backward divergence becomes less reliable. This inconsistency becomes noticeable in cells with incorrect predictions (Fig. 4), which demonstrates that backward regularization can effectively reflect the accuracy of the predictions. Thus, the backward regularization " + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "text", + "content": " generally penalizes incorrect predictions while having less impact on correct ones. " + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\theta_{b}" + }, + { + "bbox": [ + 304, + 243, + 545, + 397 + ], + "type": "text", + "content": " is set as 10 in the paper. In Sec. 4, we will show that the backward regularization effectively benefits the self-supervised motion learning process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 398, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 422 + ], + "type": "text", + "content": "Overall loss functions. To summarize, the overall training objective is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 427, + 545, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 427, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 350, + 427, + 545, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p}} + \\alpha \\mathcal {L} _ {c} + \\beta \\mathcal {L} _ {f} + \\gamma \\mathcal {L} _ {b}, \\tag {9}", + "image_path": "15308cbfe506cec7dbaed69a42d2272681f73b9eb50461d32a550cba6b76f1da.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "type": "text", + "content": " are hyper-parameters, balancing the contribution of the different loss terms. It is worth noting that all the regularization terms are only used for training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 490, + 388, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 388, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 388, + 503 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": "In this section, we begin by comparing our methods with state-of-the-art (SOTA) motion prediction methods. Subsequently, we conduct ablation studies to analyze the effectiveness of each individual component." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": "Dataset. We mainly evaluate the proposed approach on the large-scale self-driving dataset: nuScenes [2], which contains 850 scenes with annotations. For fair comparisons, we follow previous works [19, 30, 31] to use 500 scenes for training, 100 scenes for validation, and 250 scenes for testing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. We follow the same pre-process setting in [31], where the point clouds are cropped in the range of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "[-32\\mathrm{m}, 32\\mathrm{m}] \\times [-32\\mathrm{m}, 32\\mathrm{m}] \\times [-3\\mathrm{m}, 2\\mathrm{m}]" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and the voxel size is set to " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{m} \\times 0.25\\mathrm{m} \\times 0.4\\mathrm{m}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " along XYZ axis. We take the current sweep and the past four sweeps as input. We use MotionNet as the baseline model for its effectiveness and flexibility following [15, 19]. We train the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14642" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 101, + 545, + 272 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Performance comparison on nuScenes dataset. Full., Weak., and Self. refer to fully-supervised, weakly-supervised, and selfsupervised training, respectively. Our approach surpasses state-of-the-arts by a large margin." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 101, + 545, + 272 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 545, + 272 + ], + "type": "table", + "html": "
MethodSup.ModalityYearsStaticSpeed ≤ 5m/s (Slow)Speed ≥ 5m/s (Fast)
Mean↓Median↓Mean↓Median↓Mean↓Median↓
Static-LiDAR-000.61110.09718.65178.1412
FlowNet3D [18]Full.LiDARCVPR20180.041000.81830.17828.52618.0230
HPLFlowNet [8]Full.LiDARCVPR20190.00410.00020.44580.09604.32062.4881
LSTM-ED [25]Full.LiDARICRA20190.035800.35510.10441.58851.0003
MotionNet [31]Full.LiDARCVPR20200.025600.25650.09621.07440.7332
BE-STI [30]Full.LiDARCVPR20220.022000.21150.09290.75110.5413
WeakMotion [15]Weak.LiDARCVPR20230.055800.43370.13051.78231.0887
PointPWC [32]Self.LiDARECCV20200.526400.94230.32253.95302.5149
RigidFlow [14]Self.LiDARCVPR20220.636101.18360.56663.94392.5781
PillarMotion [19]Self.LiDAR+ImageCVPR20210.16200.00100.69720.17583.55042.0844
WeakMotion [15]Self.LiDARCVPR20230.191900.45100.11502.97151.8822
OursSelf.LiDAR-0.041900.32130.10612.29431.0508
", + "image_path": "91420f9252144d9afcc5ab9b75cf3b0c68008782ef56af7ab5e038b4ed307c80.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": "model for 20 epochs and set the learning rate to be 0.002. Adam is used as the optimizer. We implement our model in Pytorch [22] with a single RTX 3090 GPU. Hyperparameters " + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": " are set as 0.05, 0.1, and 1 respectively. Following previous works, we take the sequence with 5 frames (" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": ") as input and show the results of the next 1s (" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "inline_equation", + "content": "T' = 5" + }, + { + "bbox": [ + 46, + 292, + 287, + 387 + ], + "type": "text", + "content": ") in the following tables. The time interval between each frame is 0.2s." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "text", + "content": "Evaluation metrics. Following previous works [19, 30, 31], we evaluate the mean and median errors at different speed levels by dividing the grid cells into 3 groups according to the ground truth speeds: static, slow (" + }, + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\leq 5\\mathrm{m/s}" + }, + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "text", + "content": ") and fast (" + }, + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\geq 5\\mathrm{m/s}" + }, + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "text", + "content": "). Errors are measured by " + }, + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 388, + 287, + 471 + ], + "type": "text", + "content": " distances between the predicted displacements and the ground truth displacements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 272, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 272, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 272, + 491 + ], + "type": "text", + "content": "4.1. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 640 + ], + "type": "text", + "content": "In Table 1, we show the results of the following methods: (1) A static model that assumes a static environment. (2) Scene flow estimation methods, including FlowNet3D [18], HPLFlowNet [8], PointPWC [32], and RigidFlow [14]. Following [31], we employ scene flow estimation methods by assuming linear dynamics. First, we predict the motion by estimating the flow from the current time to the past 0.4s. Then, we project the predicted flow onto the BEV map and to the future 1.0s for comparison. (3) Class-agnostic motion prediction methods, including LSTM-ED [25], MotionNet [31], PillarMotion [19], BE-STI [30], and WeakMotionNet [15]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": "Compared to fully-supervised models, our approach outperforms FlowNet3D and HPLFlowNet at both slow and fast speed levels, and it surpasses the weakly-supervised WeakMotionNet (with " + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": " foreground and background masks [15]) at static and slow speed levels. In comparison with self-supervised methods, our approach achieves" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "text", + "content": "superior performance to PillarMotion, with significant improvements of " + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "inline_equation", + "content": "74.1\\% (-0.1201\\mathrm{m})" + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "inline_equation", + "content": "53.9\\%" + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "inline_equation", + "content": "(-0.3759\\mathrm{m})" + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "inline_equation", + "content": "35.4\\% (-1.2561\\mathrm{m})" + }, + { + "bbox": [ + 304, + 292, + 545, + 413 + ], + "type": "text", + "content": " at static, slow, and fast speed levels, respectively. To further show the superiority of the proposed approach, we implement recent WeakMotionNet in a self-supervised manner by replacing its pre-segmentation network with the same ground segmentation algorithm [12] used in our approach. While self-supervised WeakMotionNet outperforms PillarMotion at slow and fast speed levels, our approach still surpasses it by a large margin." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 420, + 406, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 406, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 406, + 431 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 485 + ], + "type": "text", + "content": "In this subsection, we evaluate the effectiveness of each component of our approach. As shown in Table 2, the base model trained only with pseudo labels does not perform well at all speed levels (Row 1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of cluster consistency regularization. When cluster consistency loss is not applied, the model exhibits noticeably poorer performance at the fast speed level (Compare Row 1 and Row 6). This disparity in performance can be attributed to the way the cost matrix for the optimal transport problem is computed based on the distance between cells. Consequently, the solver tends to find correspondences between cells that are in close proximity to each other. While this proximity-based approach has its merits, it presents challenges in generating long displacement pseudo labels that the model can learn from. As a result, during the early stages of training, there is a scarcity of correct pseudo labels for fast motion. By imposing constraints on cluster consistency, the model is encouraged to predict faster motion, even if it corresponds to a short displacement pseudo label. This is possible due to the presence of fast displacement predictions within the same cluster. With larger displacement and pre-warping, the optimal transport solver is more likely to find correct correspondence between two" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14643" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 112, + 286, + 267 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": "Table 2. Ablation study for different combinations of proposed regularization terms. Bold fonts and underlines indicate the best and the second-best performance, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 112, + 286, + 267 + ], + "lines": [ + { + "bbox": [ + 50, + 112, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 50, + 112, + 286, + 267 + ], + "type": "table", + "html": "
#Row Lsup LcLfLbStaticSlowFast
Mean Error ↓
1✓0.16500.55403.4503
2✓✓0.04130.32492.7587
3✓✓0.11970.42643.2072
4✓✓✓0.04990.32772.5753
5✓✓✓0.05020.34322.5154
6✓✓0.17050.41542.7027
7✓✓✓0.14500.40872.6509
8✓✓✓0.04190.32132.2943
", + "image_path": "b2195ae128900a7fe7652bb7190d3a6fbcad797531284938df9367ee55a54400.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 296, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 287, + 344 + ], + "type": "text", + "content": "points at long distances. In Table 3, we present the results of KNN consistency loss and cluster consistency loss, both of which function as spatial smooth losses. Compared to " + }, + { + "bbox": [ + 46, + 296, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KNN}" + }, + { + "bbox": [ + 46, + 296, + 287, + 344 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 296, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 296, + 287, + 344 + ], + "type": "text", + "content": " bring more benefits comprehensively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "text", + "content": "Effectiveness of temporal consistency regularization. Both forward and backward regularization terms boost the performances across all speed levels (Compare Row 5 and Row 7 with Row 8 respectively). To further illustrate the impact of backward regularization, we separately feed forward and backward sequences of all training samples into the model only trained with " + }, + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup}" + }, + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "text", + "content": " to demonstrate prediction divergence " + }, + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "inline_equation", + "content": "(|\\overrightarrow{\\mathcal{M}} -\\overleftarrow{\\mathcal{M}} |)" + }, + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "text", + "content": ". As depicted in Fig. 4, there is a positive correlation between the forward-backward prediction divergence and prediction accuracy. The divergence is large when the prediction is inaccurate and vice versa, which indicates that backward regularization will have a stronger impact on inaccurate predictions but a weaker influence on accurate ones. By using the forward-backward divergence as a loss term, the model is penalized for making inaccurate predictions. Results for various " + }, + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\theta_{b}" + }, + { + "bbox": [ + 46, + 348, + 287, + 567 + ], + "type": "text", + "content": " are shown in Table 4, which indicates that the weight term can further improve the performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Qualitative results. We present the result visualizations of our approach in Fig. 5 to qualitatively show the impacts of each regularization term. When directly training with pseudo labels " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", the model tends to produce inconsistent predictions for the same object and predict large motions for background cells. The third column illustrates that cluster consistency loss " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " effectively benefits spatial consistency. With the addition of forward regularization " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", the prediction results are further improved, but there are still noticeable noises in the background cells. Finally, by incorporating backward regularization " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", the noise predictions in the background cells are significantly suppressed." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 123, + 538, + 247 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 3. Impact of different spatial regularizations. Cluster-level regularization performs better than the KNN-based regularization. Bold fonts and underlines indicate the best and the second-best performance, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 123, + 538, + 247 + ], + "lines": [ + { + "bbox": [ + 315, + 123, + 538, + 247 + ], + "spans": [ + { + "bbox": [ + 315, + 123, + 538, + 247 + ], + "type": "table", + "html": "
MethodStaticSlowFast
Mean Errors ↓
Baseline0.05020.34322.5154
+KLNNK=30.03710.33702.5362
K=50.04900.33782.4568
K=100.07310.34052.5116
+Lc(ours)dc=20.05260.34382.1814
dc=30.04190.32132.2943
dc=40.04590.33182.4434
", + "image_path": "051b30ff7754e176e3b6ba38ef9294a4d619ce3f1247347a71d963700f16cce4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 354, + 304, + 499, + 396 + ], + "blocks": [ + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "text", + "content": "Table 4. Varying " + }, + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\theta_{b}" + }, + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "text", + "content": " for backward consistency loss. - indicates the results without the " + }, + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\exp\\left(-\\frac{t}{\\theta_{b}}\\right)" + }, + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "text", + "content": " term. Bold fonts and underlines indicate the best and the second-best performance, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 304, + 499, + 396 + ], + "lines": [ + { + "bbox": [ + 354, + 304, + 499, + 396 + ], + "spans": [ + { + "bbox": [ + 354, + 304, + 499, + 396 + ], + "type": "table", + "html": "
θbStaticSlowFast
Mean Errors ↓
50.04820.32642.3580
100.04190.32132.2943
150.02930.30912.4149
-0.03820.32982.4596
", + "image_path": "fba3aa430a63499204a839452aaeb0135138206cdbd55b4c8fdbc835380013a3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 414, + 529, + 517 + ], + "blocks": [ + { + "bbox": [ + 319, + 414, + 529, + 517 + ], + "lines": [ + { + "bbox": [ + 319, + 414, + 529, + 517 + ], + "spans": [ + { + "bbox": [ + 319, + 414, + 529, + 517 + ], + "type": "image", + "image_path": "e27d30d153bcee99a0f3e7f3fe3620f0aed3505263e569ca21cc2dc8a2122d2f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 528, + 545, + 583 + ], + "lines": [ + { + "bbox": [ + 305, + 528, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 545, + 583 + ], + "type": "text", + "content": "Figure 4. Forward-backward divergence of the training set. We analyze the prediction errors of all samples in the training set and their corresponding forward-backward divergences. The Forward-backward divergence is positively correlated with the prediction error." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 620, + 454, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 454, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 454, + 633 + ], + "type": "text", + "content": "4.3. Results on Waymo Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": "Following WeakMotionNet [15], we further evaluate our approach on the Waymo dataset [27]. As shown in Table 5, similar to the results on the nuScenes dataset, our approach achieves mean errors of around " + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{m}" + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m}" + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{m}" + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": " at static, slow, and fast speed levels, respectively, outperforming self-supervised WeakMotionNet." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14644" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 531, + 277 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 531, + 277 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 531, + 277 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 531, + 277 + ], + "type": "image", + "image_path": "03e6c58ef98f461137f958f125bfc4be64ebe2489e2a7494bf2d6d9436acbc14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 309 + ], + "type": "text", + "content": "Figure 5. Qualitative results of the proposed self-supervised approach. The future displacements " + }, + { + "bbox": [ + 46, + 285, + 547, + 309 + ], + "type": "inline_equation", + "content": "(M^{T\\rightarrow T + 5})" + }, + { + "bbox": [ + 46, + 285, + 547, + 309 + ], + "type": "text", + "content": " are depicted using the color wheel representation. (Zoom in for the best view)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 365, + 290, + 456 + ], + "blocks": [ + { + "bbox": [ + 47, + 323, + 290, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 323, + 290, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 290, + 357 + ], + "type": "text", + "content": "Table 5. Motion prediction results on the Waymo Dataset. Our method still outperforms self-supervised WeakMotion by a large margin." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 365, + 290, + 456 + ], + "lines": [ + { + "bbox": [ + 48, + 365, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 290, + 456 + ], + "type": "table", + "html": "
MethodSup.StaticSlowFast
Mean Errors ↓
MotionNet [31]Full.0.02630.26200.9493
WeakMotion [15]Weak.0.02970.34581.5655
WeakMotion [15]Self.0.13450.58333.0180
OursSelf.0.05150.44402.3692
", + "image_path": "9a66f31a4bf185aa51b128715ead715efae984d74f1c46e6bec346db08dbaf7a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 478, + 125, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 125, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 125, + 490 + ], + "type": "text", + "content": "4.4. Discussions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "Ground Segmentation. We evaluate the performance of the ground segmentation algorithm [12] on the nuScenes dataset, and get a " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " precision rate while a " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " recall rate, indicating the majority of ground points can be correctly removed. Although almost all ground points can be successfully removed, the self-supervised training process still suffers from remaining ground points, noises, and points without correspondences. The issues are alleviated using our proposed regularizations, as shown in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Clustering. In the paper, we utilize the breadth-first clustering (BFS) algorithm to demonstrate the effectiveness of the cluster consistency loss. We evaluate the clustering results of BEV cells on the nuScenes dataset and observe a " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " under-clustering rate and an " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " over-clustering rate. It indicates that although effective, the BFS algorithm still wrongly clusters cells from different instances in some cases. Therefore, fellow researchers can consider utilizing or developing more advanced clustering strategies, which" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 325, + 440, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 440, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 440, + 337 + ], + "type": "text", + "content": "may lead to better performances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 337, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 456 + ], + "type": "text", + "content": "Efficiency. During training, ground segmentation and clustering take around 5ms and 10ms, respectively. In practice, these processes can be employed as data pre-processing before training. For the optimal transport (OT) solver, we implement it with Pytorch and operate in batch. It takes less than 1ms to solve the OT problem. During inference, given that our approach is model-independent, the inference time is based on the model used. When applying MotionNet, it takes 10ms for point cloud transformation and voxelization and another 10ms for prediction [31]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 488, + 379, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 379, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 379, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 515, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 547, + 624 + ], + "type": "text", + "content": "In this paper, we introduce a novel approach for self-supervised motion prediction with only unlabeled LiDAR point clouds. We utilize optimal transport to generate pseudo labels and propose three novel spatial and temporal regularization terms to facilitate self-supervised motion learning. Extensive experiments on the large-scale nuScenes dataset illustrate that our approach significantly surpasses previous self-supervised methods and shallows the gap with fully-supervised ones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and the MoE AcRF Tier 1 grant (RG14/22)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14645" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Aseem Behl, Despoina Paschalidou, Simon Donne, and Andreas Geiger. Pointflownet: Learning representations for rigid motion estimation from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 288, + 281 + ], + "type": "text", + "content": "[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 282, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 287, + 316 + ], + "type": "text", + "content": "[4] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Neural Information Processing Systems (NeurIPS), 2013. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 384 + ], + "type": "text", + "content": "[5] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 430 + ], + "type": "text", + "content": "[6] Guanting Dong, Yueyi Zhang, Hanlin Li, Xiaoyan Sun, and Zhiwei Xiong. Exploiting rigidity constraints for lidar scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 430, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 474 + ], + "type": "text", + "content": "[7] Liangji Fang, Qinhong Jiang, Jianping Shi, and Bolei Zhou. Tpnet: Trajectory proposal network for motion prediction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 531 + ], + "type": "text", + "content": "[8] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 287, + 555 + ], + "type": "text", + "content": "[9] Marcel Van Herk. Image registration using chamfer matching. Handbook of Medical Imaging, 2000. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "text", + "content": "[10] Yair Kittenplon, Yonina C Eldar, and Dan Raviv. Flow-step3d: Model unrolling for self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "text", + "content": "[11] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 3, 4, 6, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "text", + "content": "[13] Ruibo Li, Guosheng Lin, and Lihua Xie. Self-point-flow: Self-supervised scene flow estimation from point clouds with optimal transport and random walk. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[14] Ruibo Li, Chi Zhang, Guosheng Lin, Zhe Wang, and Chunhua Shen. Rigidflow: Self-supervised scene flow learning on point clouds by local rigidity prior. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[15] Ruibo Li, Hanyu Shi, Ziang Fu, Zhe Wang, and Guosheng Lin. Weakly supervised class-agnostic motion prediction for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[17] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "text", + "content": "[18] Xingyu Liu, C. Qi, and Leonidas J. Guibas. Flownet3d: Learning scene flow in 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[19] Chenxu Luo, Xiaodong Yang, and Alan Loddon Yuille. Self-supervised pillar motion learning for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "text", + "content": "[20] Wenjie Luo, Binh Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "text", + "content": "[21] Himangi Mittal, Brian Okorn, and David Held. Just go with the flow: Self-supervised scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 511, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 566 + ], + "type": "text", + "content": "[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "text", + "content": "[23] Jhony Kaesemodel Pontes, James Hays, and Simon Lucey. Scene flow from point clouds with or without learning. In International Conference on 3D Vision (3DV), 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[24] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[25] Marcel Schreiber, Stefan Hoermann, and Klaus Dietmayer. Long-term occupancy grid prediction using recurrent neural networks. In IEEE international conference on robotics and automation (ICRA), 2019. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang chun Woo. Convolutional LSTM network: A machine learning approach for precipita" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14646" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 574 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "tion nowcasting. In Neural Information Processing Systems (NeurIPS), 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[27] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[28] Guangming Wang, Xinrui Wu, Zhe Liu, and Hesheng Wang. Hierarchical attention learning of scene flow in 3d point clouds. IEEE Transactions on Image Processing (TIP), 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[29] Kewei Wang, Yizheng Wu, Zhiyu Pan, Xingyi Li, Ke Xian, Zhe Wang, Zhiguo Cao, and Guosheng Lin. Semi-supervised class-agnostic motion prediction with pseudo label regeneration and bevmix, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 317 + ], + "type": "text", + "content": "[30] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "text", + "content": "[31] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "type": "text", + "content": "[32] Wenxuan Wu, Zhi Yuan Wang, Zhuwen Li, Wei Liu, and Li Fuxin. Pointpwc-net: Cost volume on point clouds for (self-) supervised scene flow estimation. In European Conference on Computer Vision (ECCV), 2020. 2, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 473 + ], + "type": "text", + "content": "[33] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. In IEEE Conference on International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "type": "text", + "content": "[34] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yun-ting Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. Tnt: Target-driven trajectory prediction. In Conference on Robot Learning (CoRL), 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 532, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 574 + ], + "type": "text", + "content": "[35] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14647" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_content_list.json b/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a09ff3d1b8d94ccf7f6c9c2d66ecce73c4b46593 --- /dev/null +++ b/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_content_list.json @@ -0,0 +1,1563 @@ +[ + { + "type": "text", + "text": "Self-Supervised Dual Contouring", + "text_level": 1, + "bbox": [ + 315, + 130, + 653, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ramana Sundararaman", + "bbox": [ + 241, + 181, + 429, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LIX, Ecole Polytechnique", + "bbox": [ + 233, + 199, + 441, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sundararman@lix.polytechnique.fr", + "bbox": [ + 194, + 219, + 478, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Roman Klokov", + "bbox": [ + 593, + 181, + 715, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LIX, Ecole Polytechnique", + "bbox": [ + 550, + 199, + 756, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "klokov@lix.polytechnique.fr", + "bbox": [ + 534, + 219, + 774, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maks Ovsjanikov \nLIX, Ecole Polytechnique", + "bbox": [ + 383, + 241, + 583, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "maks@lix.polytechnique.fr", + "bbox": [ + 370, + 279, + 594, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 328, + 312, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning-based isosurface extraction methods have recently emerged as a robust and efficient alternative to axiomatic techniques. However, the vast majority of such approaches rely on supervised training with axiomatically computed ground truths, thus potentially inheriting biases and data artefacts of the corresponding axiomatic methods. Steering away from such dependencies, we propose a self-supervised training scheme to the Neural Dual Contouring meshing framework, resulting in our method: Self-Supervised Dual Contouring (SDC). Instead of optimizing predicted mesh vertices with supervised training, we use two novel self-supervised loss functions that encourage the consistency between distances to the generated mesh up to the first order. Meshes reconstructed by SDC surpass existing data-driven methods in capturing intricate details while being more robust to possible irregularities in the input. Furthermore, we use the same self-supervised training objective linking inferred mesh and input SDF, to regularize the training process of Deep Implicit Networks (DINs). We demonstrate that the resulting DINs produce higher-quality implicit functions, ultimately leading to more accurate and detail-preserving surfaces compared to prior baselines for different input modalities. Finally, we demonstrate that our self-supervised losses improve meshing performance in the single-view reconstruction task by enabling joint training of predicted SDF and resulting output mesh. We open-source our code at https://github.com/Sentient07/SDC.", + "bbox": [ + 75, + 361, + 472, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 799, + 209, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Surface mesh extraction from implicit functions, often referred to as isosurfacing [30, 37, 45, 62], is a fundamental problem in computer graphics and geometry processing as the quality of reconstructed mesh impacts algorithms used in numerous downstream tasks [49]. Prominent primal ax", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e44ea2a865b510402cbcf2afe5a97db9dea85510f5c1b72ea4ccb07822bd9042.jpg", + "image_caption": [ + "Figure 1. We propose SDC, a data-driven self-supervised isosurfacing method that reconstructs better feature-preserving meshes from learnt implicit functions compared to the baseline NDC [13]. Our framework is extendable as a regularizer for training Deep Implicit Networks (DINs) for improved reconstruction." + ], + "image_footnote": [], + "bbox": [ + 501, + 333, + 877, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "iomatic methods such as Marching Cubes [37, 45], are computationally efficient but fail to reconstruct sharp features. Dual methods [22, 30, 34, 62] achieve sharper reconstructions using Quadratic Error Function (QEF) but their performance heavily relies on the data fidelity.", + "bbox": [ + 496, + 579, + 890, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following the ubiquitous success of neural methods in 2D and 3D applications, recent learning-based meshing approaches [11, 13] propose to predict mesh vertices and connectivity from various inputs in a feed-forward manner. When trained on extensive, augmented mesh collections, these methods demonstrate enhanced generalization and yield higher-quality reconstructions. However, these approaches rely either directly on the ground truth meshes [63] or on some intermediate pre-computed proxy meshes, like complex tessellation templates [11] and outputs of axiomatic approaches [13].", + "bbox": [ + 496, + 656, + 892, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Dual Contouring (NDC) [13], the most relevant method to our work, is a promising recent approach that enhances standard Dual Contouring (DC) [30, 62] with a data-driven pipeline. It aims to replicate DC-produced meshes, resulting in a framework capable of generating high-quality,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4681", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "feature-preserving meshes. However, since vertex estimation of DC follows QEF minimization, it requires precise SDF values at grid points and normals at edge intersections for training. Due to QEF minimization being ill-defined in particular scenarios [30], and susceptible to irregularities in the input SDF, NDC may exhibit a bias towards poor triangulation quality for challenging inputs and intricate details. Furthermore, its reliance on DC meshes for training restricts its use in an end-to-end scenario where both the implicit function and the explicit mesh can be optimized jointly.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose Self-Supervised Dual Contouring (SDC), a meshing framework that extends Neural Dual Contouring with two novel geometrically motivated self-supervised losses to learn the mesh vertex positions from SDF inputs while relying on sign changes to extract connectivity. The first self-supervised objective is a distance-based loss which minimizes the difference between the input implicit function values at points close to the surface and distances from the same points to the reconstructed mesh. This ensures that the predicted reconstructed mesh agrees with the input SDF. Secondly, we align the normals corresponding to reconstructed faces to those estimated from the input grid of SDFs. Since our training objective relies only on the input SDF and is agnostic to any explicit representation of the shape, e.g., based either on surface samples [26, 63] or on mesh vertices [11, 13], we refer to our approach as Self-Supervised. Together these two terms produce vertices that best explain the input grid of signed distance function values up to the first order like dual methods, while avoiding shortcomings of the quadratic error function (QEF) minimization [13, 30]. In particular, unlike DC, our method produces vertices that lie within each voxel-cell, and optimization of our self-supervised objectives leads to overall better results for input data with imperfections. As a result, we demonstrate better performance than supervised baselines in obtaining sharper meshes with negligible self-intersections and better generalization to unseen input data.", + "bbox": [ + 75, + 246, + 472, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, leveraging the connection between reconstructed meshes and input SDF grids, we propose a mesh-based regularization for training Deep Implicit Networks (DINs) - MLPs which represent zero-level sets of a shape. In a standard setup, training of DINs is completely decoupled from meshing, which is only performed a posteriori. Thus, existing DINs might produce implicit functions that do not correspond to a distance function arising from an extracted mesh. This motivates us to use SDC as a regularization to train DINs. In particular, we enforce the distance function produced by DINs during training to remain as close as possible to the distance function arising from the mesh produced by SDC. This forms the converse of our self-supervised meshing losses, but now geared to produce more coherent implicit functions. We demonstrate that this", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "regularization can be directly adopted into existing training paradigms for Deep Implicit Networks and it that leads to more plausible reconstructions.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, since SDC training does not require ground truth meshes, it can be seamlessly integrated within end-to-end training paradigms. To highlight this, we consider the task of mesh reconstruction from images where we first construct an implicit representation given an image using an existing approach [55, 60] and then reconstruct a mesh using SDC. The resulting model is trained end-to-end by jointly minimizing the standard surface reconstruction loss, our self-supervised loss and the proposed regularization for Deep Implicit Networks. The resulting approach demonstrates compelling surface reconstruction efficacy across multiple object categories from the ShapeNet dataset [9]. In summary, our contributions are as follows:", + "bbox": [ + 496, + 137, + 892, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We extend Neural Dual Contouring with two novel self-supervised loss functions ensuring consistency between input SDF and distance to output meshes, which is geometrically better motivated than the standard QEF minimization.", + "- We introduce a novel regularizer for training Deep Implicit Networks (DINs), by penalizing SDFs that do not correspond to underlying meshes.", + "- We show that SDC generalizes better than supervised meshing methods, and further demonstrate its utility in feature-preserving mesh reconstruction from images." + ], + "bbox": [ + 500, + 334, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 515, + 637, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Surface reconstruction and meshing tasks have been extensively studied within computer graphics and vision. Below, we review the methods that are most closely related to ours and refer interested readers to several surveys [5, 33, 66] for a more in-depth discussion.", + "bbox": [ + 496, + 541, + 890, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Axiomatic contouring techniques. The process of isosurface extraction from volumetric data is referred to as contouring. The most widely-used pipeline for surface reconstruction consists of two major steps. First, volumetric data is computed using a signed distance function [14, 29, 32, 51]. In the second step, a mesh is extracted from this representation, using Marching Cubes [45], Dual Contouring [30] or related approaches [17, 52, 71]. This pipeline can produce very high quality meshes, but is sensitive to the quality of input data and is not differentiable and thus does not easily fit within modern learning-based pipelines [38]. Another line of approaches is based on Delaunay triangulations, [7, 8, 36], as well as closely-related constructions [2, 3, 6, 19] which often come with strong theoretical guarantees and are typically geared towards preserving some input point set [1, 6]. Such methods, however, provide little control over the output triangulation. Moreover, classical approaches are typically not differentiable and thus cannot be easily integrated into learning pipelines.", + "bbox": [ + 496, + 625, + 892, + 912 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4682", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-driven mesh reconstruction. To address differentiability issues of classical mesh reconstruction approaches, a number of data-driven techniques have recently been proposed. This includes differentiable variants of contouring methods, such as the Marching Cubes [11, 38], dual Marching Cubes [65], Marching Tetrahedra [64], and Dual Contouring [13]. Our approach can be categorized into this family of differentiable contouring techniques. In addition, surface reconstruction has also been tackled using template-based techniques which fit a template mesh to the input [31, 39, 42], or deform an initial mesh while potentially updating its connectivity [20, 54, 72]. Furthermore, other local approaches propose to fit parameterized surface patches to points [24, 73], or to decompose space into convex sets [12, 16, 46]. Finally, several learning-based methods mesh an input point set [15, 44, 59, 63]. These approaches strongly rely on the given input point cloud and can thus be sensitive to artifacts and noise.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural fields. There has been a recent surge in methods for learning implicit functions, referred to as neural fields [74]. This includes pioneering methods for predicting grid occupancy and signed distance functions [10, 48, 55] as well as their many follow-up works [21, 61, 67, 70]. Often, training Deep Implicit Networks introduces inductive bias [58] thereby producing inexplicable surface behaviors. Multiple regularization techniques [4, 23, 40, 41, 43, 58, 67, 70] have been proposed to address that issue. They largely focus on either learning high-frequency signals [40, 70] or designing loss functions based to guide the gradient level-set [23, 41, 58]. While the latter category of methods has convergence guarantees [41], the former helps in producing detail preserving reconstruction. Our work, on the other hand, provides a novel regularization by utilizing a meshing framework to ensure that the predicted signed distance field arises from a mesh. Similar to recent works [25, 47, 56, 60, 64, 65, 76] which optimize the mesh by controlling the underlying implicit function, our approach also enables back-propagation from the mesh to the input SDF. However, our approach is non-iterative and produces a mesh in one feed-forward pass.", + "bbox": [ + 75, + 369, + 472, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 76, + 718, + 205, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first begin by providing the notations used across the paper in Section 3.1 and review the Dual Contouring and Neural Dual Contouring algorithms in Section 3.2.", + "bbox": [ + 75, + 744, + 468, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notations", + "text_level": 1, + "bbox": [ + 76, + 801, + 187, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We let $f: \\mathbb{R}^3 \\to \\mathbb{R}$ be an implicit function that determines the signed distance of a query point $x \\in \\mathbb{R}^3$ from an underlying surface $\\chi$ . We let $f_{\\theta}$ denote the implicit function produced by a neural network parameterized by $\\theta$ . We use $\\mathcal{G}$ to denote an $l \\times m \\times n$ regular grid in the Euclidean space,", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$g \\in \\mathcal{G}$ to be a node in the grid and $e_{ij} \\in \\mathcal{G}$ to be an edge between adjacent nodes $g_i, g_j$ respectively. We denote $\\mathcal{S}_{\\mathcal{G}}$ as the discretization of $f$ on $\\mathcal{G}$ i.e. SDF evaluated at every grid node. Finally, we let $\\lambda$ denote the latent vector associated with a shape computed using an encoder.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dual contouring", + "text_level": 1, + "bbox": [ + 498, + 174, + 661, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Dual Contouring (DC) [30] framework is designed to produce quadrilateral meshes from input signed distance functions discretized on a grid $S_{\\mathcal{G}}$ . It can be decomposed into two main steps: (1) construction of mesh faces $\\mathcal{F}$ ; (2) computation of mesh vertices $\\mathcal{V}$ . To construct mesh faces, DC considers adjacent nodes $g_i, g_j \\in \\mathcal{G}$ such that their signed distance are of opposite signs, $f(g_i) \\neq f(g_j)$ . Then, for those pairs, a unique quadrilateral face $q_k$ is constructed such that $q_k$ crosses the edge $e_{ij}$ . Vertices of this quadrilateral $q_k$ reside in adjacent grid cells. The faces of the resulting mesh are obtained as a union of all quadrilateral faces $\\mathcal{F} = \\bigcup_{k \\in I} q_k$ , where $I$ is a set of indices of grid edges connecting grid vertices of different signs. Final mesh vertex positions within each cell are obtained by minimizing the Quadratic Error Function (QEF). More precisely, given $p_e$ to be the intersection position between edge $e_{ij}$ and the face $q_{ij}$ (along $e_{ij}$ ) and $n_e$ to be the gradients of the SDF at positions $e_{ij}$ (or surface normals), vertices are determined for each grid cell $c$ by solving:", + "bbox": [ + 496, + 196, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv _ {c} = \\underset {x} {\\arg \\min } \\sum_ {e \\in I _ {c}} \\left[ n _ {e} * \\left(x - p _ {e}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 491, + 890, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I_{c}$ is a set of indices of intersected grid edges corresponding to cell $c$ . Recently introduced Neural Dual Contouring [13] attempts to emulate Dual Contouring via a neural network in order to reduce the inference time while achieving optimal reconstruction.", + "bbox": [ + 496, + 530, + 890, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 617, + 589, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We describe our Self-supervised Dual Contouring (SDC) loss functions in Section 4.1, discuss how to use them for regularization while training deep implicit networks in Section 4.2, and apply them for mesh reconstruction from input images in Section 4.3.", + "bbox": [ + 496, + 642, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Self-supervised dual contouring (SDC)", + "text_level": 1, + "bbox": [ + 498, + 726, + 831, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a Self-supervised training scheme for Dual Contouring, and refer to a model trained with it as to SDC. It excludes the dependence on vertex positions obtained via QEF minimization. SDC produces quad meshes from the signed distance function discretized on a grid $S_{\\mathcal{G}}$ similarly to any Dual Contouring method. We triangulate quadrilaterals by joining the top-left and bottom-right vertex as a convention. $S_{\\mathcal{G}}$ can either be measured [27] or predicted using a DIN [55] $f_{\\theta}$ . SDC consists of two main components constructing faces and vertices. To construct faces,", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4683", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we follow the Dual Contouring [30] algorithm to determine the connectivity based on signed distance value at grid nodes. A quadrilateral face is constructed in the cells incident on every edge $e_{ij} \\in \\mathcal{G}$ connecting grid nodes of different signs $\\mathrm{sign}(f(g_i)) \\neq \\mathrm{sign}(f(g_j))$ . For vertex prediction, we use a 6-layered 3D Convolutional Neural Network $h_\\phi : S_\\mathcal{G} \\to \\mathcal{P} \\in \\mathbb{R}^{(l-1) \\times (m-1) \\times (n-1) \\times 3}$ , which takes the grid of signed distance values as input and predicts a single point per each grid cell as the output. Then, we apply a masking to select grid cells, that bear node SDF values of opposite signs. Please refer to Suppl. for more details.", + "bbox": [ + 76, + 90, + 472, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SDC is trained without explicit fitting to the ground truth meshes using two loss functions. Firstly, to ensure formal correspondence of produced surfaces to input SDF grids, we propose a distance-based loss $\\mathcal{L}_D$ . This objective minimizes the differences between 1) the absolute distance values at the nodes of the discretization grid $f(g_{i})\\in S_{\\mathcal{G}}$ which are provided as inputs and 2) the distance $\\mathbf{d}^p (g_i,\\mathcal{M})$ measured from the same point $g_{i}$ to the predicted mesh $\\mathcal{M}$ . For a smooth or gently undulating surface, this distance provides a reasonable approximation of the underlying geometry [50, 57]. However, in the presence of sharp features, the local geometry changes direction rapidly. Close query points residing on different sides of a sharp feature have similar distances to the mesh but may be projected to very different locations on the actual surface. To address this, we also introduce a normal consistency loss, which aims to align the normals estimated at the nodes of the discretization grid $f(g_{i})\\in S_{\\mathcal{G}}$ and normals corresponding to the generated mesh face.", + "bbox": [ + 76, + 258, + 472, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We refer to the losses proposed above as self-supervised since the training signal for our pipeline comes from the input alone. We do not use any explicit surface discretization for training and our losses are solely based on the input implicit function values and the generated mesh. In contrast, we refer to methods that predict vertices to fit a specific mesh structure as supervised meshing methods [11, 13]. We provide visual explanations of our loss functions in Figure 2. In summary, our combined self-supervised objective for the vertex prediction network is given by:", + "bbox": [ + 76, + 547, + 472, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {M e s h}} = \\mathcal {L} _ {D} + \\alpha_ {1} \\mathcal {L} _ {N}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 715, + 468, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ is a hyperparameter to weigh $\\mathcal{L}_N$ .", + "bbox": [ + 76, + 750, + 370, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distance loss. The objective of our distance loss is to minimize the discrepancy between surfaces produced by SDC and zero-level sets defined by the input SDFs. In order to achieve this, given grid nodes $g_{i} \\in \\mathcal{G}$ , we measure the distance between the grid nodes and the mesh produced by SDC and penalize its deviation from the input absolute distance values $f(g_{i})$ . In summary, our distance loss is defined", + "bbox": [ + 76, + 794, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e34092315d8f2d81533f2c84bc32d0ad6ef55fd34585f1022ad69c68b0c71db6.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 504, + 94, + 694, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5fdcefafe6f6a567cba586466686ddbfc38a6ce383938b459c3487220e67a196.jpg", + "image_caption": [ + "(b)", + "Figure 2. Illustration of our self-supervised losses. (a) Measurements of $\\mathbf{d}^p$ between grid nodes and the closest point on a mesh face. For visualization purposes, we only show this estimation at one node. (b) Normal consistency loss is measured as the discrepancy between interpolated normals (orange) and the face normal (green). Red nodes denote negative SDF grid and green denotes positive. The constructed quad is shown in transparent purple." + ], + "image_footnote": [], + "bbox": [ + 697, + 140, + 880, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as:", + "bbox": [ + 500, + 383, + 522, + 395 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = \\sum_ {g _ {i} \\in \\mathcal {G}} \\left\\| \\operatorname {a b s} \\left(f \\left(g _ {i}\\right)\\right) - \\mathbf {d} ^ {p} \\left(g _ {i}, \\mathcal {M}\\right) \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 393, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{d}^p$ is the point-to-face distance, computed between $g_{i}$ and its projection on $\\mathcal{M}$ . For computational efficiency, we compute $\\mathcal{L}_D$ only at grid nodes that are close to the surface. Please refer to Suppl for details.", + "bbox": [ + 496, + 431, + 893, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normal consistency loss. We introduce a novel normal consistency loss, $\\mathcal{L}_N$ which aims to align the level-sets up to the first order. In particular, given the grid of scalar SDFs, we compute normals as gradients $n(g_i) \\coloneqq f'(g_i) / ||f'(g_i)||_2$ in grid nodes $g_i$ using Five-point stencil, a commonly used finite difference approach for solving partial differential equations on regular grids [68]. More details on gradient estimation is provided in the Suppl. Given an edge $e_{ij} \\in \\mathcal{G}$ connecting nodes $g_i, g_j$ of the opposite signs, we first estimate a point $p \\in e_{ij}$ along the edge where $f(p) = 0$ . The point $p$ can be represented in terms of $g_i$ and $g_j$ as $p = (1 - t)g_i + tg_j$ . Where $t$ is a parameter that determines the position of $p$ along the edge $e_{ij}$ . Solving for $p = 0$ , the parameter $t$ can be expressed in closed form as:", + "bbox": [ + 496, + 510, + 893, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt = \\frac {f \\left(g _ {i}\\right)}{f \\left(g _ {i}\\right) - f \\left(g _ {j}\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 731, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, we can estimate the normal at the point $p$ by linear interpolation. Using the previously computed parameter $t$ , the normal at $p$ can be represented as:", + "bbox": [ + 496, + 772, + 893, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn (p) = \\operatorname {s i g n} \\left(f \\left(g _ {i}\\right)\\right) (1 - t) n \\left(g _ {i}\\right) + \\operatorname {s i g n} \\left(f \\left(g _ {j}\\right)\\right) t n \\left(g _ {j}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 828, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This expression combines the known gradients at the neighboring nodes, weighted by the parameter $t$ , to estimate the gradient at a point $p$ . Since the inward normal is oriented", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4684", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "opposite to the outward normal, we multiply by the sign for consistency. As our mesh is constructed following Dual Contouring [30], we denote the quad-face constructed dual to the edge $e_{ij}$ as $q_{ij}$ and its normal as $n(q_{ij})$ . With the aim of aligning these two normals, our normal consistency loss can be written as:", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N} = 1 - \\frac {n \\left(q _ {i j}\\right) \\cdot n (p)}{\\| n \\left(q _ {i j}\\right) \\| \\| n (p) \\|.} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 186, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training augmentation. At training time, we add synthetic noise which is zero-mean Gaussian noise with a standard deviation equivalent to a third grid edge length $e_{ij}$ to the input SDF. Since SDC constructs mesh faces based on Dual Contouring, a noise that induces a sign change to the initial SDF will also lead to a topological change. More precisely, augmentation will lead to inclusion of vertices (and faces) in cells whose nodes differ in sign. Similarly, faces are removed dual to cells whose signs agree after augmentation. More training details are provided in the Suppl.", + "bbox": [ + 75, + 234, + 472, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Deep implicit network regularizer", + "text_level": 1, + "bbox": [ + 76, + 393, + 375, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deep Implicit Networks (DINs), commonly represented by Multi-Layer Perceptrons (MLP), are trained to acquire an implicit representation of a shape and are primarily utilized for various inference-based surface reconstruction tasks. In a traditional setup, DINs are trained by direct supervision of the SDF values at query points sampled close to the surface. Such a training scheme, as also observed by previous works [23, 41, 58], is known to produce inexplicable behavior, such as ambiguous level sets. To address this, we propose a novel mesh-based training regularization for DINs to ensure that the Distance Field (DF) produced by DIN agrees well with the resulting extracted mesh. Our regularization minimizes the discrepancy between the predicted DF (absolute value of SDF) and the distance computed from the mesh corresponding to the SDF extracted using SDC.", + "bbox": [ + 75, + 417, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assume that we are endowed with a learnable DIN $f_{\\theta}(\\lambda, p): p \\in \\mathbb{R}^3 \\to s_p \\in \\mathbb{R}$ , represented as an MLP, whose weights $\\theta$ encode the implicit surface of a shape conditioned by a latent vector $\\lambda$ , $p$ is a point in space such that $f_{\\theta}(\\lambda, p) = s_p$ is its signed distance from the zero-level set. Then, for a given $S_G$ - a discretization of $f_{\\theta}$ on a regular grid and $\\mathcal{M}^*(S_G)$ - a mesh that is produced following any iso-surfacing algorithm, it is not necessarily true that $\\mathrm{abs}(f_{\\theta}(p, \\lambda)) = \\mathbf{d}^p(p, \\mathcal{M}^*(S_G))$ , where $\\mathbf{d}^p$ is the point-to-face distance. While this disparity is viewed as an inherent drawback of primal iso-surfacing methods [53], for dual methods, this disparity can, in principle, be minimized due to their ability to preserve sharp details in the reconstructed meshes. This motivates us to use our SDC for iso-surface extraction and regularize the training of DINs by minimizing the aforementioned disparity. Assuming we have access to a dataset consisting of $N$ shapes and a pre-trained", + "bbox": [ + 75, + 643, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SDC with frozen weights, our regularization objective can be summarized as follows:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D R} = \\sum_ {j = 1} ^ {N} \\sum_ {g _ {i} \\in \\mathcal {G}} | | \\operatorname {a b s} \\left(f _ {\\theta} \\left(\\lambda_ {j}, g _ {i}\\right) - \\mathbf {d} ^ {p} \\left(g _ {i}, \\mathcal {M}\\right) \\right| | _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 128, + 890, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This regularization penalizes the discrepancy between the DF at grid nodes predicted by the DIN $f_{\\theta}(g_i)$ and the distance $\\mathbf{d}^p$ computed between the grid nodes and the mesh $\\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}})$ . To obtain $\\mathcal{M}$ , we first evaluate $f_{\\theta}$ over a discrete grid $\\mathcal{G}$ and obtain $\\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}})$ . Note that the above regularization is similar to the signed distance loss defined in Eq. (3), but with an important difference. Here we do not use ground-truth SDF values but instead, regularize the predicted signed distance. This regularization is applied alongside the main objective function to reconstruct the implicit surface commonly used to train DINs [18, 55] as follows:", + "bbox": [ + 498, + 181, + 890, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S D F}} = \\sum_ {j = 1} ^ {N} \\sum_ {i} ^ {K} \\left| \\left| \\left(f _ {\\theta} \\left(\\lambda_ {j}, x _ {i}\\right) - s _ {i} \\right| \\right| _ {2} ^ {2} + \\frac {1}{\\sigma^ {2}} \\left\\| \\lambda_ {j} \\right\\| _ {2} ^ {2}, \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 369, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{K}$ is the number of points with annotated SDF values per shape $j$ , $s_i$ is the ground-truth SDF value, and $\\sigma$ is a parameter used to promote compactness in latent space [55].", + "bbox": [ + 498, + 422, + 890, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining the standard training loss for DINs [55] alongside our regularization weighted with a scalar $\\alpha_{2}$ , our learning objective is formulated as follows:", + "bbox": [ + 498, + 468, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a i n}} = \\mathcal {L} _ {\\text {S D F}} + \\alpha_ {2} \\mathcal {L} _ {S D R}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 523, + 890, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the inference time, we reconstruct surfaces from point clouds by first recovering the optimal latent vectors (See Eq. (4) in [18]). Then, we predict a grid of SDF values $S_{\\mathcal{G}}$ and extract the iso-surface using our SDC.", + "bbox": [ + 498, + 551, + 890, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Joint learning of implicit surface and meshing", + "text_level": 1, + "bbox": [ + 498, + 619, + 888, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differently from existing data-driven meshing methods, SDC relies neither on explicit mesh data [59, 63] nor on axiomatically produced proxy meshes [11, 13]. This means that SDC can be used to optimize both SDF and reconstructed meshes jointly in an end-to-end manner. To demonstrate this we consider the task of predicting surface meshes from images. We first model the SDF representation of a shape given an image and then predict the mesh using SDC. In terms of the architectures, we follow MeshSDF [60]: we produce latent vectors $\\lambda$ from input images with a ResNet-18 [28] encoder and learn an implicit shape representation of shapes using DeepSDF [55] conditioned on the aforementioned latent vectors. Instead of performing test-time optimization like MeshSDF, we predict SDF values on a regular grid $f_{\\theta}(\\mathcal{G})$ and use SDC to reconstruct a mesh $\\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}})$ by a simple forward pass of our SDC network. Initially, we train the networks (1) DeepSDF (with", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4685", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an image encoder) and (2) SDC independently, then jointly fine-tune them for 200 epochs. This fine-tuning uses an image as input and the mesh $\\mathcal{M}$ from SDC as output, while minimizing the combined objective terms for surface reconstruction and meshing as follows:", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S V R}} = \\mathcal {L} _ {\\mathrm {S D F}} + \\alpha_ {3} \\mathcal {L} _ {S D R} + \\alpha_ {4} \\mathcal {L} _ {\\mathrm {D}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 180, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{SDF}}$ and $\\mathcal{L}_{\\mathrm{SDR}}$ (defined in Eq. (8) and Eq. (7) respectively) are used to reconstruct and regularize the implicit representation, while $\\mathcal{L}_D$ , defined in Eq. (3), is used to train SDC. Since the evaluated SDF might be noisy during training, we observed that using $\\mathcal{L}_N$ results in instability.", + "bbox": [ + 75, + 208, + 468, + 285 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments, results and discussion", + "text_level": 1, + "bbox": [ + 76, + 297, + 398, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the effectiveness of SDC across three tasks, namely, implicit function meshing, DIN training regularization and joint SDF and mesh prediction from images.", + "bbox": [ + 75, + 323, + 468, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use Chamfer Distance (CD), Normal Consistency (NC), Self-Intersection (SI), Edge Chamfer Distance (ECD), 3D-IoU, Precise Level-Set Discrepancy (LSD-P), and Approximate Level-Set Discrepancy (LSD-A) as metrics to compare all approaches. For measuring Level-Set discrepancy, we use two terms, Precise (LSD-P) and Approximate (LSD-A). The former is used in case where shapes are water-tight and analytical SDF can be measured. In such cases, we measure the discrepancy between the ground truth analytical SDF and the distance function measured from the generated mesh. LSD-A is used for non-watertight shapes and is measured by sampling points on the ground truth mesh and measuring its distance from the generated mesh's faces. Self-intersection is reported as the total number of intersecting triangles per mesh averaged across the test set. For the single-view reconstruction task, we also measure the Structural Similarity Index (SSIM) between the rendering of the reconstructed mesh and the ground truth mesh. We measure 3D-IoU for experiments where the generated mesh could topologically differ from the ground truth. CD, ECD, LSD are scaled by $10^{3}$ while NC, SSIM 3D IoU are in $\\%$ .We provide more details on all the metrics in the Suppl.", + "bbox": [ + 75, + 388, + 470, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Meshing analytical implicit functions", + "text_level": 1, + "bbox": [ + 76, + 746, + 398, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first consider the surface mesh reconstruction task given the ground truth SDFs of shapes discretized on a regular grid. For this task, we train SDC on 3000 shapes from the $1^{st}$ split of the ABC dataset [35] for 100 epochs. We scale each shape to fit a unit sphere and estimate the SDF on a regular grid of dimension $64^3$ using SDFGen library1. We evaluate all methods on 300 shapes from the test set of", + "bbox": [ + 75, + 770, + 468, + 876 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4766a179210be62b82f858e9b73591dee6bf9bf098d0e7189fbad6d4ebd1ee33.jpg", + "image_caption": [ + "Figure 3. We show three qualitative mesh reconstruction examples from the ABC dataset (Row 1,2) and the Thingi10k dataset (Row 3). Self-intersecting faces are highlighted in red. QEF-based learning methods NDC and UNDC show significantly higher self-intersecting faces along sharp edges while ours does not." + ], + "image_footnote": [], + "bbox": [ + 500, + 89, + 883, + 330 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e11d341c366f60d74b4630ac11f031b90070d0ff23e9ab98f98c48f5d284b781.jpg", + "image_caption": [ + "Figure 4. Qualitative examples of meshing SDF produced by NGLOD [69]. MC, NMC, and NDC produce noisy meshes while SDC produces a smoother yet feature-preserving reconstruction." + ], + "image_footnote": [], + "bbox": [ + 500, + 435, + 888, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the ABC dataset. Additionally, we evaluate on 300 shapes from the Thingi10K [77] dataset to demonstrate generalization to unseen data domains. We compare SDC against four baselines including two axiomatic baselines, namely the improved Marching Cubes (MC) [37] and the standard Dual Contouring (DC) [30] where positions of each vertex are estimated by solving the Quadratic Error Function (QEF), and three of the recent Neural Meshing techniques, Neural Marching Cubes (NMC) [11], Neural Dual Contouring [13] and FlexiCubes [65] respectively. For fairness, we compare to variants of comparable network capacity and for FlexiCubes, we compare without test-time optimization.", + "bbox": [ + 496, + 656, + 890, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We summarize our quantitative results in Table 1. Our self-supervised method SDC shows better performance and generalization to unseen data domains compared to the data-driven baselines across all metrics. While Marching", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "DatasetInput SDFMethodCD (↓)NC (↑)#SI (↓)ECD (↓)LSD-P (↓)ABCAnalyticalMC [37]4.792.104.816.1NMC [11]3.794.734.53.612.1DC [30]3.593.596.33.411.6NDC [13]3.694.343.13.511.9FlexiCubes [65]4.592.404.815.8Ours3.394.99.73.211.3Thingi10KAnalyticalMC [37]5.161.4013.315.3NMC [11]4.166.032.411.912.2DC [30]4.063.8105.611.511.8NDC [13]4.064.970.411.312.0FlexiCubes [65]5.255.3013.615.4Ours3.668.015.710.911.6Thingi10KPredicted [69]MC [37]5.856.7014.216.5NMC [11]4.763.230.112.612.9DC [30]6.254.0230.414.815.0NDC [13]4.762.842.612.813.1Ours4.265.715.912.212.4", + "bbox": [ + 78, + 88, + 468, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative mesh reconstruction results on the ABC and the Thingi10k datasets. The best scores are highlighted in bold and the second-best scores are in italics. The first two row-blocks correspond to input SDF which was computed analytically while in the last row-block, we mesh the SDF predicted by NGLOD [69]", + "bbox": [ + 75, + 253, + 468, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cubes [37] and FlexiCubes [65] do not produce any self-intersecting faces, the vertex placement is restricted for these methods, leading to inferior results with respect to other reconstruction metrics. In the qualitative examples visualized in Figure 3, we highlight an inherent issue with NDC and UNDC - they produce a considerable number of self-intersections along sharp edges. We reason this to be an undesirable trait that is inherited from the standard Dual Contouring as previously elaborated. SDC, on the other hand, bears a minimal number of self-intersections, as our training objective for mesh prediction is geometrically well-motivated. Additionally, we report quantitative and qualitative results for $128^{3}$ grid resolution in the Suppl.", + "bbox": [ + 75, + 363, + 470, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Meshing predicted implicit functions", + "text_level": 1, + "bbox": [ + 76, + 573, + 393, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Differently from the previous section, where the SDF is computed analytically, here, we use an open-source implementation (with suggested hyperparameters) of a recent deep implicit network NGLOD [69] to predict SDF values on a regular grid. We evaluate the meshing efficacy of SDC over a set of 300 shapes from the Thingi10k [77] dataset. To avoid possible biases, we chose these 300 shapes to be different from the ones used in the previous section. Our quantitative results are summarized in Table 1. Consistent with our observation from the previous section, SDC outperforms all baselines. We provide a qualitative example of meshing a noisy SDF in Figure 4. To generate the noisy SDF, we prematurely terminate NGLOD [69] fitting to the shape and evaluate different Neural Meshing methods. While all baselines produce a noisy reconstruction, SDC recovers a smoother surface while preserving sharp edges, thanks to our regularization and self-supervised losses. This suggests that our self-supervised training objectives and data augmentation are capable of generalizing to noisy imperfect predicted data without explicit training on this data.", + "bbox": [ + 75, + 598, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Regularizing implicit surface learning", + "text_level": 1, + "bbox": [ + 498, + 90, + 826, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in the previous section, SDC successfully produces sharp meshes from both analytical and parameterized implicit functions. We now employ SDC for regularizing DIN training, focusing on the task of surface reconstruction using the ShapeNet dataset [9]. In particular, we evaluate over 4 categories of objects, namely, cars, planes, tables, and cabinets with 1000 shapes per category as our training set, and set aside a separate set of 200 unseen shapes for evaluation. We compare our regularization against three possible baselines. Firstly, we train Curriculum DeepSDF (CSDF) [18] over each category separately. Secondly, we re-train the same network by enforcing the Eikonal constraint, i.e. enforcing a unit norm for SDF's gradient in a setup similar to Implicit Geometric Regularization [23]. Thirdly, we use the supervised Dual Contouring baseline NDC [13] as the mesh-based regularizer. More specifically, we replace SDC with NDC in computing the regularization introduced in Eq. (7). For each shape, we sample 400,000 points in the shape volume, aggressively near the surface following [55] to supervise the SDF prediction. For fairness, we use the same sampled points for all methods we compare.", + "bbox": [ + 496, + 113, + 890, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We summarize our quantitative results in Table 2. We observe that regularizing the network using SDC produces a noticeable improvement in reconstruction across all metrics in comparison to baselines. Moreover, we observe a poorer reconstruction when using NDC [13] for regularization. We argue that this is due to their inability to handle noisy predicted SDF values as inputs as this approach learns to emulate solution to QEF (c.f. Eq. (1)), which is ill-defined for imperfect implicit functions. Our self-supervised loss functions do not rely on the ill-defined vertices produced by QEF minimization and align produced meshes to the input SDF grids, resulting in more plausible surface predictions. We believe this difference to be the key reason behind SDC's efficacy as a regularizer for training DINs. We also show three qualitative results in Figure 5. In some examples, we also observe topological differences although our regularization does not explicitly penalize it. Since the loss function which we minimize is highly non-convex, we believe that our regularization has aided in discovering more \"plausible\" latent space which could lead to better quality of implicit surfaces. We report additional qualitative and quantitative results in Suppl demonstrating the regularizer's generalization capabilities with other DINs.", + "bbox": [ + 496, + 446, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Single view reconstruction", + "text_level": 1, + "bbox": [ + 500, + 801, + 738, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we consider the task of reconstructing surface meshes from images. We use 4 categories from the ShapeNet dataset [9], namely planes, chairs, rifles, and tables. We train and evaluate our approach and baseline (using the official codebase) on the same training and evalua", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4687", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2826841f986fab930d15f8c682bbdae6eaeabf9d4a3b8adc7e664a0b82dab2c7.jpg", + "image_caption": [ + "Figure 5. Qualitative examples comparing surface reconstruction from the ShapeNet [9] dataset. The columns compare different methods, whose training objectives differ. Our mesh-based regularization produces reasonable results compared to baselines." + ], + "image_footnote": [], + "bbox": [ + 81, + 94, + 460, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/972aa891be39f2e0fb639f899bdcbd4abdfbefe81e22a89b829b0a929e5478ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCD (↓)NC (↑)3D IoU (↑)LSD-A (↓)
CSDF [18]3.078.183.93.1
+ Eikonal [23]3.278.083.03.2
+ NDC [13]2.978.484.82.9
+ SDC2.678.986.02.7
", + "bbox": [ + 91, + 361, + 452, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion split for a fair comparison. We compare against three baselines, namely, MeshSDF [60], DISN [75], and a variant of our approach referred to as WoBW. More specifically, WoBW meshes the implicit function produced by pretrained SV-DIN using SDC, without the joint fine-tuning. For MeshSDF [60] and DISN [75] we use the official codebase for re-training and evaluation.", + "bbox": [ + 75, + 512, + 467, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative results averaged across 4 object categories are summarized in Table 3. Notably, our SDC outperforms MeshSDF [60] without additional test-time optimization fitting a mesh to the rendering. In addition, our end-to-end model consistently outperforms the WoBW baseline, implying the efficacy of our mesh-based regularization and end-to-end training. We also show three qualitative examples in Figure 6. Our SDC demonstrates an improved ability to reconstruct sharp features in the predicted meshes, resulting in surfaces more faithful to the ground truth. In the third row, we highlight an example showing significant improvements in the quality of predicted geometry. It shows that our method is capable of better thin surface reconstructions for the same inputs.", + "bbox": [ + 75, + 619, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and future work", + "text_level": 1, + "bbox": [ + 76, + 844, + 336, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced SDC, a Self-Supervised training approach for Neural Dual Contouring. Differently from previous", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/925fa32e0d5f30af812b3065663ced19b29e69726d0fc41ff2b3ed6fc2c03963.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of Single View Reconstruction (SVR) accuracy among three objects from ShapeNet [9] dataset. Input denotes the input view and GT denotes ground-truth mesh. SDC produces more sharper and plausible surfaces in comparison to baselines." + ], + "image_footnote": [], + "bbox": [ + 501, + 93, + 890, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/227f50b64fe47686f2b9b40a622f287fa6401fce05e31ec856f9dca7d0007ddf.jpg", + "table_caption": [ + "Table 2. Comparison of surface reconstruction accuracy across object categories from the ShapeNet dataset [9]. Methods are trained with different regularization (see text) while evaluated alike." + ], + "table_footnote": [], + "table_body": "
MethodCD (↓)NC (↑)SSIM (↑)LSD-A (↓)
DISN [75]13.070.282.010.5
MeshSDF [60]11.371.484.19.1
WoBW9.972.086.48.1
Ours9.172.888.57.7
", + "bbox": [ + 496, + 397, + 857, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative results for single-view reconstruction across different object categories from ShapeNet [9] dataset.", + "bbox": [ + 498, + 487, + 890, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work, we do not fit the generated mesh to an axiomatically produced proxy mesh but instead use geometrically motivated self-supervised losses which only depends on the input SDF. SDC shows consistent improvements over baselines when applied to end-to-end surface reconstruction and meshing tasks. In addition, our work forges a link between signed distance fields and meshes and shows how the former can be regularized by guiding the network to produce distance fields better corresponding to resulting meshes. Our meshing and regularization are applied on a regular grid, which limits the resolution of a reconstructed mesh. It would be interesting to explore adaptive grids within the learning framework. Finally, another related and interesting scope for future work is the exploration of strict manifoldness conditions [62] and theoretical guarantees of self-intersection free surfaces.", + "bbox": [ + 496, + 551, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, and gifts from Ansys and Adobe Research.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4688", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nina Amenta and Marshall Bern. Surface reconstruction by voronoi filtering. Discrete & Computational Geometry, 22 (4):481-504, 1999. 2", + "[2] Nina Amenta, Marshall Bern, and Manolis Kamvysselis. A new voronoi-based surface reconstruction algorithm. Proc. SIGGRAPH, pages 415-421, 1998. 2", + "[3] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. Computational Geometry, 19(2-3):127-153, 2001. 2", + "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 3", + "[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Computer Graphics Forum, 36(1): 301-329, 2017. 2", + "[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2", + "[7] Jean-Daniel Boissonnat. Geometric structures for three-dimensional shape representation. ACM Transactions on Graphics (TOG), 3(4):266-286, 1984. 2", + "[8] Jean-Daniel Boissonnat and Steve Oudot. Provably good sampling and meshing of surfaces. Graphical Models, 67 (5):405-451, 2005. 2", + "[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 7, 8", + "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3", + "[11] Zhiqin Chen and Hao Zhang. Neural marching cubes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 1, 2, 3, 4, 5, 6, 7", + "[12] Zhiqin Chen, Andrea Tagliausacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3", + "[13] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. arXiv preprint arXiv:2202.01999, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2", + "[15] Angela Dai and Matthias Nießner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5574-5583, 2019. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 31-44, 2020. 3", + "[17] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE TRANSACTIONS on Information and Systems, 74(1):214-224, 1991. 2", + "[18] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevata, and Leonidas J. Guibas. Curriculum deepsdf. CoRR, abs/2003.08593, 2020. 5, 7, 8", + "[19] Herbert Edelsbrunner and Ernst P Mücke. Three-dimensional alpha shapes. ACM Transactions on Graphics (TOG), 13(1):43-72, 1994. 2", + "[20] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 3", + "[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. arXiv preprint arXiv:1904.06447, 2019. 3", + "[22] S.F.F. Gibson. Using distance maps for accurate surface representation in sampled volumes. In IEEE Symposium on Volume Visualization (Cat. No.989EX300), pages 23-30, 1998.", + "[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 5, 7, 8", + "[24] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3", + "[25] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In European Conference on Computer Vision, 2022. 3", + "[26] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph., 39(4), 2020. 2", + "[27] John C. Hart. Sphere tracing: a geometric method for the antialiased ray tracing of implicit surfaces. The Visual Computer, 12(10):527-545, 1996. 3", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 5", + "[29] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. ACM SIGGRAPH Computer Graphics, 26 (2), 1992. 2", + "[30] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 1, 2, 3, 4, 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3", + "[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2", + "[33] Alireza Khatamian and Hamid R Arabnia. Survey on 3d surface reconstruction. Journal of Information Processing Systems, 12(3), 2016. 2", + "[34] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 1", + "[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[36] Ravikrishna Kolluri, Jonathan Richard Shewchuk, and James F O'Brien. Spectral surface reconstruction from noisy point clouds. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 11-21. ACM, 2004. 2", + "[37] Thomas Lewiner, Hélio Lopes, Antonio Wilson Vieira, and Geovan Tavares. Efficient implementation of marching cubes' cases with topological guarantees. Journal of Graphics Tools, 8(2):1-15, 2003. 1, 6, 7", + "[38] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3", + "[39] Chen-Hsuan Lin, Oliver Wang, Bryan C Russell, Eli Shechtman, Vladimir G Kim, Matthew Fisher, and Simon Lucey. Photometric mesh optimization for video-aligned 3d object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 969-978, 2019. 3", + "[40] David B. Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. arXiv preprint arXiv:0000.00000, 2021.3", + "[41] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, 2021. 3, 5", + "[42] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1886-1895, 2018. 3", + "[43] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization, 2022. 3", + "[44] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision, pages 68-84. Springer, 2020. 3", + "[45] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 2", + "[46] N. Maruani, R. Klokov, M. Ovsjanikov, P. Alliez, and M. Desbrun. Voromesh: Learning watertight surface meshes with voronoi diagrams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[47] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. arXiv preprint arXiv:2204.07159, 2022. 3", + "[48] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3", + "[49] Mark Meyer, Mathieu Desbrun, Peter Schröder, and Alan H Barr. Discrete differential-geometry operators for triangulated 2-manifolds. In Visualization and mathematics III, pages 35-57. Springer, 2003. 1", + "[50] Niloy J. Mitra, Natasha Gelfand, Helmut Pottmann, and Leonidas Guibas. Registration of point cloud data from a geometric optimization perspective. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH Symposium on Geometry Processing, page 22-31, New York, NY, USA, 2004. Association for Computing Machinery. 4", + "[51] Patrick Mullen, Fernando De Goes, Mathieu Desbrun, David Cohen-Steiner, and Pierre Alliez. Signing the unsigned: Robust surface reconstruction from raw pointsets. Computer Graphics Forum, 29(5):1733-1741, 2010. 2", + "[52] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 2", + "[53] G.M. Nielson. Dual marching cubes. In IEEE Visualization 2004, pages 489-496, 2004. 5", + "[54] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 9964-9973, 2019. 3", + "[55] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2, 3, 5, 7", + "[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 3", + "[57] Helmut Pottmann and Michael Hofer. Geometry of the squared distance function to curves and surfaces. In Mathematics and Visualization, pages 221-242. Springer Berlin Heidelberg, 2003. 4", + "[58] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4690", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 3, 5", + "[59] Marie-Julie Rakotaosaona, Paul Guerrero, Noam Algerman, Niloy J Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22-31, 2021. 3, 5", + "[60] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2, 3, 5, 8", + "[61] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 3", + "[62] Scott Schaefer, Tao Ju, and Joe Warren. Manifold dual contouring. IEEE Transactions on Visualization and Computer Graphics, 13(3):610-619, 2007. 1, 8", + "[63] Nicholas Sharp and Maks Ovsjanikov. Pointtrinet: Learned triangulation of 3d point sets. In European Conference on Computer Vision, pages 762-778. Springer, 2020. 1, 2, 3, 5", + "[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3", + "[65] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 6, 7", + "[66] Jonathan Shewchuk, Tamal K Dey, and Siu-Wing Cheng. Delaunay mesh generation. Chapman and Hall/CRC, 2016. 2", + "[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3", + "[68] G D Smith. Numerical solution of partial differential equations. Clarendon Press, Oxford, England, 3 edition, 1985. 4", + "[69] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. 2021. 6, 7", + "[70] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3", + "[71] Graham M Treece, Richard W Prager, and Andrew H Gee. Regularised marching tetrahedra: improved iso-surface extraction. Computers & Graphics, 23(4):583-598, 1999. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-67, 2018. 3", + "[73] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10130-10139, 2019. 3", + "[74] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 41(2):641-676, 2022. 3", + "[75] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. Curran Associates, Inc., 2019. 8", + "[76] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 3", + "[77] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10, 000 3d-printing models. CoRR, abs/1605.04797, 2016. 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 470 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4691", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_model.json b/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a86024d8115d63ca3d32d0288c95a0fa411d9e5a --- /dev/null +++ b/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.316, + 0.131, + 0.654, + 0.154 + ], + "angle": 0, + "content": "Self-Supervised Dual Contouring" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.182, + 0.431, + 0.198 + ], + "angle": 0, + "content": "Ramana Sundararaman" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.2, + 0.442, + 0.217 + ], + "angle": 0, + "content": "LIX, Ecole Polytechnique" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.22, + 0.48, + 0.234 + ], + "angle": 0, + "content": "sundararman@lix.polytechnique.fr" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.182, + 0.716, + 0.198 + ], + "angle": 0, + "content": "Roman Klokov" + }, + { + "type": "text", + "bbox": [ + 0.551, + 0.2, + 0.758, + 0.218 + ], + "angle": 0, + "content": "LIX, Ecole Polytechnique" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.22, + 0.775, + 0.234 + ], + "angle": 0, + "content": "klokov@lix.polytechnique.fr" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.242, + 0.584, + 0.278 + ], + "angle": 0, + "content": "Maks Ovsjanikov \nLIX, Ecole Polytechnique" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.28, + 0.596, + 0.295 + ], + "angle": 0, + "content": "maks@lix.polytechnique.fr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.329, + 0.313, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.473, + 0.771 + ], + "angle": 0, + "content": "Learning-based isosurface extraction methods have recently emerged as a robust and efficient alternative to axiomatic techniques. However, the vast majority of such approaches rely on supervised training with axiomatically computed ground truths, thus potentially inheriting biases and data artefacts of the corresponding axiomatic methods. Steering away from such dependencies, we propose a self-supervised training scheme to the Neural Dual Contouring meshing framework, resulting in our method: Self-Supervised Dual Contouring (SDC). Instead of optimizing predicted mesh vertices with supervised training, we use two novel self-supervised loss functions that encourage the consistency between distances to the generated mesh up to the first order. Meshes reconstructed by SDC surpass existing data-driven methods in capturing intricate details while being more robust to possible irregularities in the input. Furthermore, we use the same self-supervised training objective linking inferred mesh and input SDF, to regularize the training process of Deep Implicit Networks (DINs). We demonstrate that the resulting DINs produce higher-quality implicit functions, ultimately leading to more accurate and detail-preserving surfaces compared to prior baselines for different input modalities. Finally, we demonstrate that our self-supervised losses improve meshing performance in the single-view reconstruction task by enabling joint training of predicted SDF and resulting output mesh. We open-source our code at https://github.com/Sentient07/SDC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.8, + 0.21, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Surface mesh extraction from implicit functions, often referred to as isosurfacing [30, 37, 45, 62], is a fundamental problem in computer graphics and geometry processing as the quality of reconstructed mesh impacts algorithms used in numerous downstream tasks [49]. Prominent primal ax" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.334, + 0.878, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.489, + 0.893, + 0.561 + ], + "angle": 0, + "content": "Figure 1. We propose SDC, a data-driven self-supervised isosurfacing method that reconstructs better feature-preserving meshes from learnt implicit functions compared to the baseline NDC [13]. Our framework is extendable as a regularizer for training Deep Implicit Networks (DINs) for improved reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.656 + ], + "angle": 0, + "content": "iomatic methods such as Marching Cubes [37, 45], are computationally efficient but fail to reconstruct sharp features. Dual methods [22, 30, 34, 62] achieve sharper reconstructions using Quadratic Error Function (QEF) but their performance heavily relies on the data fidelity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Following the ubiquitous success of neural methods in 2D and 3D applications, recent learning-based meshing approaches [11, 13] propose to predict mesh vertices and connectivity from various inputs in a feed-forward manner. When trained on extensive, augmented mesh collections, these methods demonstrate enhanced generalization and yield higher-quality reconstructions. However, these approaches rely either directly on the ground truth meshes [63] or on some intermediate pre-computed proxy meshes, like complex tessellation templates [11] and outputs of axiomatic approaches [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Neural Dual Contouring (NDC) [13], the most relevant method to our work, is a promising recent approach that enhances standard Dual Contouring (DC) [30, 62] with a data-driven pipeline. It aims to replicate DC-produced meshes, resulting in a framework capable of generating high-quality," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4681" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "feature-preserving meshes. However, since vertex estimation of DC follows QEF minimization, it requires precise SDF values at grid points and normals at edge intersections for training. Due to QEF minimization being ill-defined in particular scenarios [30], and susceptible to irregularities in the input SDF, NDC may exhibit a bias towards poor triangulation quality for challenging inputs and intricate details. Furthermore, its reliance on DC meshes for training restricts its use in an end-to-end scenario where both the implicit function and the explicit mesh can be optimized jointly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.671 + ], + "angle": 0, + "content": "To overcome these limitations, we propose Self-Supervised Dual Contouring (SDC), a meshing framework that extends Neural Dual Contouring with two novel geometrically motivated self-supervised losses to learn the mesh vertex positions from SDF inputs while relying on sign changes to extract connectivity. The first self-supervised objective is a distance-based loss which minimizes the difference between the input implicit function values at points close to the surface and distances from the same points to the reconstructed mesh. This ensures that the predicted reconstructed mesh agrees with the input SDF. Secondly, we align the normals corresponding to reconstructed faces to those estimated from the input grid of SDFs. Since our training objective relies only on the input SDF and is agnostic to any explicit representation of the shape, e.g., based either on surface samples [26, 63] or on mesh vertices [11, 13], we refer to our approach as Self-Supervised. Together these two terms produce vertices that best explain the input grid of signed distance function values up to the first order like dual methods, while avoiding shortcomings of the quadratic error function (QEF) minimization [13, 30]. In particular, unlike DC, our method produces vertices that lie within each voxel-cell, and optimization of our self-supervised objectives leads to overall better results for input data with imperfections. As a result, we demonstrate better performance than supervised baselines in obtaining sharper meshes with negligible self-intersections and better generalization to unseen input data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In addition, leveraging the connection between reconstructed meshes and input SDF grids, we propose a mesh-based regularization for training Deep Implicit Networks (DINs) - MLPs which represent zero-level sets of a shape. In a standard setup, training of DINs is completely decoupled from meshing, which is only performed a posteriori. Thus, existing DINs might produce implicit functions that do not correspond to a distance function arising from an extracted mesh. This motivates us to use SDC as a regularization to train DINs. In particular, we enforce the distance function produced by DINs during training to remain as close as possible to the distance function arising from the mesh produced by SDC. This forms the converse of our self-supervised meshing losses, but now geared to produce more coherent implicit functions. We demonstrate that this" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "regularization can be directly adopted into existing training paradigms for Deep Implicit Networks and it that leads to more plausible reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.893, + 0.333 + ], + "angle": 0, + "content": "Finally, since SDC training does not require ground truth meshes, it can be seamlessly integrated within end-to-end training paradigms. To highlight this, we consider the task of mesh reconstruction from images where we first construct an implicit representation given an image using an existing approach [55, 60] and then reconstruct a mesh using SDC. The resulting model is trained end-to-end by jointly minimizing the standard surface reconstruction loss, our self-supervised loss and the proposed regularization for Deep Implicit Networks. The resulting approach demonstrates compelling surface reconstruction efficacy across multiple object categories from the ShapeNet dataset [9]. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.335, + 0.892, + 0.409 + ], + "angle": 0, + "content": "- We extend Neural Dual Contouring with two novel self-supervised loss functions ensuring consistency between input SDF and distance to output meshes, which is geometrically better motivated than the standard QEF minimization." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.41, + 0.892, + 0.456 + ], + "angle": 0, + "content": "- We introduce a novel regularizer for training Deep Implicit Networks (DINs), by penalizing SDFs that do not correspond to underlying meshes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.892, + 0.501 + ], + "angle": 0, + "content": "- We show that SDC generalizes better than supervised meshing methods, and further demonstrate its utility in feature-preserving mesh reconstruction from images." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.335, + 0.892, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.516, + 0.638, + 0.532 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Surface reconstruction and meshing tasks have been extensively studied within computer graphics and vision. Below, we review the methods that are most closely related to ours and refer interested readers to several surveys [5, 33, 66] for a more in-depth discussion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.893, + 0.913 + ], + "angle": 0, + "content": "Axiomatic contouring techniques. The process of isosurface extraction from volumetric data is referred to as contouring. The most widely-used pipeline for surface reconstruction consists of two major steps. First, volumetric data is computed using a signed distance function [14, 29, 32, 51]. In the second step, a mesh is extracted from this representation, using Marching Cubes [45], Dual Contouring [30] or related approaches [17, 52, 71]. This pipeline can produce very high quality meshes, but is sensitive to the quality of input data and is not differentiable and thus does not easily fit within modern learning-based pipelines [38]. Another line of approaches is based on Delaunay triangulations, [7, 8, 36], as well as closely-related constructions [2, 3, 6, 19] which often come with strong theoretical guarantees and are typically geared towards preserving some input point set [1, 6]. Such methods, however, provide little control over the output triangulation. Moreover, classical approaches are typically not differentiable and thus cannot be easily integrated into learning pipelines." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4682" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Data-driven mesh reconstruction. To address differentiability issues of classical mesh reconstruction approaches, a number of data-driven techniques have recently been proposed. This includes differentiable variants of contouring methods, such as the Marching Cubes [11, 38], dual Marching Cubes [65], Marching Tetrahedra [64], and Dual Contouring [13]. Our approach can be categorized into this family of differentiable contouring techniques. In addition, surface reconstruction has also been tackled using template-based techniques which fit a template mesh to the input [31, 39, 42], or deform an initial mesh while potentially updating its connectivity [20, 54, 72]. Furthermore, other local approaches propose to fit parameterized surface patches to points [24, 73], or to decompose space into convex sets [12, 16, 46]. Finally, several learning-based methods mesh an input point set [15, 44, 59, 63]. These approaches strongly rely on the given input point cloud and can thus be sensitive to artifacts and noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.473, + 0.706 + ], + "angle": 0, + "content": "Neural fields. There has been a recent surge in methods for learning implicit functions, referred to as neural fields [74]. This includes pioneering methods for predicting grid occupancy and signed distance functions [10, 48, 55] as well as their many follow-up works [21, 61, 67, 70]. Often, training Deep Implicit Networks introduces inductive bias [58] thereby producing inexplicable surface behaviors. Multiple regularization techniques [4, 23, 40, 41, 43, 58, 67, 70] have been proposed to address that issue. They largely focus on either learning high-frequency signals [40, 70] or designing loss functions based to guide the gradient level-set [23, 41, 58]. While the latter category of methods has convergence guarantees [41], the former helps in producing detail preserving reconstruction. Our work, on the other hand, provides a novel regularization by utilizing a meshing framework to ensure that the predicted signed distance field arises from a mesh. Similar to recent works [25, 47, 56, 60, 64, 65, 76] which optimize the mesh by controlling the underlying implicit function, our approach also enables back-propagation from the mesh to the input SDF. However, our approach is non-iterative and produces a mesh in one feed-forward pass." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.719, + 0.206, + 0.736 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.79 + ], + "angle": 0, + "content": "We first begin by providing the notations used across the paper in Section 3.1 and review the Dual Contouring and Neural Dual Contouring algorithms in Section 3.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.188, + 0.816 + ], + "angle": 0, + "content": "3.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We let \\( f: \\mathbb{R}^3 \\to \\mathbb{R} \\) be an implicit function that determines the signed distance of a query point \\( x \\in \\mathbb{R}^3 \\) from an underlying surface \\( \\chi \\). We let \\( f_{\\theta} \\) denote the implicit function produced by a neural network parameterized by \\( \\theta \\). We use \\( \\mathcal{G} \\) to denote an \\( l \\times m \\times n \\) regular grid in the Euclidean space," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "\\(g \\in \\mathcal{G}\\) to be a node in the grid and \\(e_{ij} \\in \\mathcal{G}\\) to be an edge between adjacent nodes \\(g_i, g_j\\) respectively. We denote \\(\\mathcal{S}_{\\mathcal{G}}\\) as the discretization of \\(f\\) on \\(\\mathcal{G}\\) i.e. SDF evaluated at every grid node. Finally, we let \\(\\lambda\\) denote the latent vector associated with a shape computed using an encoder." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.175, + 0.663, + 0.191 + ], + "angle": 0, + "content": "3.2. Dual contouring" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.486 + ], + "angle": 0, + "content": "The Dual Contouring (DC) [30] framework is designed to produce quadrilateral meshes from input signed distance functions discretized on a grid \\( S_{\\mathcal{G}} \\). It can be decomposed into two main steps: (1) construction of mesh faces \\( \\mathcal{F} \\); (2) computation of mesh vertices \\( \\mathcal{V} \\). To construct mesh faces, DC considers adjacent nodes \\( g_i, g_j \\in \\mathcal{G} \\) such that their signed distance are of opposite signs, \\( f(g_i) \\neq f(g_j) \\). Then, for those pairs, a unique quadrilateral face \\( q_k \\) is constructed such that \\( q_k \\) crosses the edge \\( e_{ij} \\). Vertices of this quadrilateral \\( q_k \\) reside in adjacent grid cells. The faces of the resulting mesh are obtained as a union of all quadrilateral faces \\( \\mathcal{F} = \\bigcup_{k \\in I} q_k \\), where \\( I \\) is a set of indices of grid edges connecting grid vertices of different signs. Final mesh vertex positions within each cell are obtained by minimizing the Quadratic Error Function (QEF). More precisely, given \\( p_e \\) to be the intersection position between edge \\( e_{ij} \\) and the face \\( q_{ij} \\) (along \\( e_{ij} \\)) and \\( n_e \\) to be the gradients of the SDF at positions \\( e_{ij} \\) (or surface normals), vertices are determined for each grid cell \\( c \\) by solving:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.492, + 0.892, + 0.525 + ], + "angle": 0, + "content": "\\[\nv _ {c} = \\underset {x} {\\arg \\min } \\sum_ {e \\in I _ {c}} \\left[ n _ {e} * \\left(x - p _ {e}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.607 + ], + "angle": 0, + "content": "where \\( I_{c} \\) is a set of indices of intersected grid edges corresponding to cell \\( c \\). Recently introduced Neural Dual Contouring [13] attempts to emulate Dual Contouring via a neural network in order to reduce the inference time while achieving optimal reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.591, + 0.634 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.719 + ], + "angle": 0, + "content": "We describe our Self-supervised Dual Contouring (SDC) loss functions in Section 4.1, discuss how to use them for regularization while training deep implicit networks in Section 4.2, and apply them for mesh reconstruction from input images in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.727, + 0.833, + 0.743 + ], + "angle": 0, + "content": "4.1. Self-supervised dual contouring (SDC)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce a Self-supervised training scheme for Dual Contouring, and refer to a model trained with it as to SDC. It excludes the dependence on vertex positions obtained via QEF minimization. SDC produces quad meshes from the signed distance function discretized on a grid \\(S_{\\mathcal{G}}\\) similarly to any Dual Contouring method. We triangulate quadrilaterals by joining the top-left and bottom-right vertex as a convention. \\(S_{\\mathcal{G}}\\) can either be measured [27] or predicted using a DIN [55] \\(f_{\\theta}\\). SDC consists of two main components constructing faces and vertices. To construct faces," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "we follow the Dual Contouring [30] algorithm to determine the connectivity based on signed distance value at grid nodes. A quadrilateral face is constructed in the cells incident on every edge \\( e_{ij} \\in \\mathcal{G} \\) connecting grid nodes of different signs \\( \\mathrm{sign}(f(g_i)) \\neq \\mathrm{sign}(f(g_j)) \\). For vertex prediction, we use a 6-layered 3D Convolutional Neural Network \\( h_\\phi : S_\\mathcal{G} \\to \\mathcal{P} \\in \\mathbb{R}^{(l-1) \\times (m-1) \\times (n-1) \\times 3} \\), which takes the grid of signed distance values as input and predicts a single point per each grid cell as the output. Then, we apply a masking to select grid cells, that bear node SDF values of opposite signs. Please refer to Suppl. for more details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.26, + 0.473, + 0.546 + ], + "angle": 0, + "content": "SDC is trained without explicit fitting to the ground truth meshes using two loss functions. Firstly, to ensure formal correspondence of produced surfaces to input SDF grids, we propose a distance-based loss \\(\\mathcal{L}_D\\). This objective minimizes the differences between 1) the absolute distance values at the nodes of the discretization grid \\(f(g_{i})\\in S_{\\mathcal{G}}\\) which are provided as inputs and 2) the distance \\(\\mathbf{d}^p (g_i,\\mathcal{M})\\) measured from the same point \\(g_{i}\\) to the predicted mesh \\(\\mathcal{M}\\). For a smooth or gently undulating surface, this distance provides a reasonable approximation of the underlying geometry [50, 57]. However, in the presence of sharp features, the local geometry changes direction rapidly. Close query points residing on different sides of a sharp feature have similar distances to the mesh but may be projected to very different locations on the actual surface. To address this, we also introduce a normal consistency loss, which aims to align the normals estimated at the nodes of the discretization grid \\(f(g_{i})\\in S_{\\mathcal{G}}\\) and normals corresponding to the generated mesh face." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.549, + 0.473, + 0.701 + ], + "angle": 0, + "content": "We refer to the losses proposed above as self-supervised since the training signal for our pipeline comes from the input alone. We do not use any explicit surface discretization for training and our losses are solely based on the input implicit function values and the generated mesh. In contrast, we refer to methods that predict vertices to fit a specific mesh structure as supervised meshing methods [11, 13]. We provide visual explanations of our loss functions in Figure 2. In summary, our combined self-supervised objective for the vertex prediction network is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.717, + 0.47, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {M e s h}} = \\mathcal {L} _ {D} + \\alpha_ {1} \\mathcal {L} _ {N}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.75, + 0.371, + 0.765 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) is a hyperparameter to weigh \\(\\mathcal{L}_N\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Distance loss. The objective of our distance loss is to minimize the discrepancy between surfaces produced by SDC and zero-level sets defined by the input SDFs. In order to achieve this, given grid nodes \\( g_{i} \\in \\mathcal{G} \\), we measure the distance between the grid nodes and the mesh produced by SDC and penalize its deviation from the input absolute distance values \\( f(g_{i}) \\). In summary, our distance loss is defined" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.095, + 0.695, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.236, + 0.61, + 0.248 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.141, + 0.882, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.236, + 0.805, + 0.248 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.26, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Illustration of our self-supervised losses. (a) Measurements of \\(\\mathbf{d}^p\\) between grid nodes and the closest point on a mesh face. For visualization purposes, we only show this estimation at one node. (b) Normal consistency loss is measured as the discrepancy between interpolated normals (orange) and the face normal (green). Red nodes denote negative SDF grid and green denotes positive. The constructed quad is shown in transparent purple." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.385, + 0.524, + 0.396 + ], + "angle": 0, + "content": "as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.394, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = \\sum_ {g _ {i} \\in \\mathcal {G}} \\left\\| \\operatorname {a b s} \\left(f \\left(g _ {i}\\right)\\right) - \\mathbf {d} ^ {p} \\left(g _ {i}, \\mathcal {M}\\right) \\right\\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.895, + 0.493 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}^p\\) is the point-to-face distance, computed between \\(g_{i}\\) and its projection on \\(\\mathcal{M}\\). For computational efficiency, we compute \\(\\mathcal{L}_D\\) only at grid nodes that are close to the surface. Please refer to Suppl for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.895, + 0.723 + ], + "angle": 0, + "content": "Normal consistency loss. We introduce a novel normal consistency loss, \\(\\mathcal{L}_N\\) which aims to align the level-sets up to the first order. In particular, given the grid of scalar SDFs, we compute normals as gradients \\(n(g_i) \\coloneqq f'(g_i) / ||f'(g_i)||_2\\) in grid nodes \\(g_i\\) using Five-point stencil, a commonly used finite difference approach for solving partial differential equations on regular grids [68]. More details on gradient estimation is provided in the Suppl. Given an edge \\(e_{ij} \\in \\mathcal{G}\\) connecting nodes \\(g_i, g_j\\) of the opposite signs, we first estimate a point \\(p \\in e_{ij}\\) along the edge where \\(f(p) = 0\\). The point \\(p\\) can be represented in terms of \\(g_i\\) and \\(g_j\\) as \\(p = (1 - t)g_i + tg_j\\). Where \\(t\\) is a parameter that determines the position of \\(p\\) along the edge \\(e_{ij}\\). Solving for \\(p = 0\\), the parameter \\(t\\) can be expressed in closed form as:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.732, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\nt = \\frac {f \\left(g _ {i}\\right)}{f \\left(g _ {i}\\right) - f \\left(g _ {j}\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.895, + 0.82 + ], + "angle": 0, + "content": "Similarly, we can estimate the normal at the point \\( p \\) by linear interpolation. Using the previously computed parameter \\( t \\), the normal at \\( p \\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.829, + 0.892, + 0.847 + ], + "angle": 0, + "content": "\\[\nn (p) = \\operatorname {s i g n} \\left(f \\left(g _ {i}\\right)\\right) (1 - t) n \\left(g _ {i}\\right) + \\operatorname {s i g n} \\left(f \\left(g _ {j}\\right)\\right) t n \\left(g _ {j}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "This expression combines the known gradients at the neighboring nodes, weighted by the parameter \\( t \\), to estimate the gradient at a point \\( p \\). Since the inward normal is oriented" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4684" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "opposite to the outward normal, we multiply by the sign for consistency. As our mesh is constructed following Dual Contouring [30], we denote the quad-face constructed dual to the edge \\( e_{ij} \\) as \\( q_{ij} \\) and its normal as \\( n(q_{ij}) \\). With the aim of aligning these two normals, our normal consistency loss can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.188, + 0.47, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N} = 1 - \\frac {n \\left(q _ {i j}\\right) \\cdot n (p)}{\\| n \\left(q _ {i j}\\right) \\| \\| n (p) \\|.} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.473, + 0.388 + ], + "angle": 0, + "content": "Training augmentation. At training time, we add synthetic noise which is zero-mean Gaussian noise with a standard deviation equivalent to a third grid edge length \\( e_{ij} \\) to the input SDF. Since SDC constructs mesh faces based on Dual Contouring, a noise that induces a sign change to the initial SDF will also lead to a topological change. More precisely, augmentation will lead to inclusion of vertices (and faces) in cells whose nodes differ in sign. Similarly, faces are removed dual to cells whose signs agree after augmentation. More training details are provided in the Suppl." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.395, + 0.377, + 0.412 + ], + "angle": 0, + "content": "4.2. Deep implicit network regularizer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Deep Implicit Networks (DINs), commonly represented by Multi-Layer Perceptrons (MLP), are trained to acquire an implicit representation of a shape and are primarily utilized for various inference-based surface reconstruction tasks. In a traditional setup, DINs are trained by direct supervision of the SDF values at query points sampled close to the surface. Such a training scheme, as also observed by previous works [23, 41, 58], is known to produce inexplicable behavior, such as ambiguous level sets. To address this, we propose a novel mesh-based training regularization for DINs to ensure that the Distance Field (DF) produced by DIN agrees well with the resulting extracted mesh. Our regularization minimizes the discrepancy between the predicted DF (absolute value of SDF) and the distance computed from the mesh corresponding to the SDF extracted using SDC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We assume that we are endowed with a learnable DIN \\(f_{\\theta}(\\lambda, p): p \\in \\mathbb{R}^3 \\to s_p \\in \\mathbb{R}\\), represented as an MLP, whose weights \\(\\theta\\) encode the implicit surface of a shape conditioned by a latent vector \\(\\lambda\\), \\(p\\) is a point in space such that \\(f_{\\theta}(\\lambda, p) = s_p\\) is its signed distance from the zero-level set. Then, for a given \\(S_G\\) - a discretization of \\(f_{\\theta}\\) on a regular grid and \\(\\mathcal{M}^*(S_G)\\) - a mesh that is produced following any iso-surfacing algorithm, it is not necessarily true that \\(\\mathrm{abs}(f_{\\theta}(p, \\lambda)) = \\mathbf{d}^p(p, \\mathcal{M}^*(S_G))\\), where \\(\\mathbf{d}^p\\) is the point-to-face distance. While this disparity is viewed as an inherent drawback of primal iso-surfacing methods [53], for dual methods, this disparity can, in principle, be minimized due to their ability to preserve sharp details in the reconstructed meshes. This motivates us to use our SDC for iso-surface extraction and regularize the training of DINs by minimizing the aforementioned disparity. Assuming we have access to a dataset consisting of \\(N\\) shapes and a pre-trained" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "SDC with frozen weights, our regularization objective can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.129, + 0.892, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D R} = \\sum_ {j = 1} ^ {N} \\sum_ {g _ {i} \\in \\mathcal {G}} | | \\operatorname {a b s} \\left(f _ {\\theta} \\left(\\lambda_ {j}, g _ {i}\\right) - \\mathbf {d} ^ {p} \\left(g _ {i}, \\mathcal {M}\\right) \\right| | _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.182, + 0.892, + 0.362 + ], + "angle": 0, + "content": "This regularization penalizes the discrepancy between the DF at grid nodes predicted by the DIN \\( f_{\\theta}(g_i) \\) and the distance \\( \\mathbf{d}^p \\) computed between the grid nodes and the mesh \\( \\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}}) \\). To obtain \\( \\mathcal{M} \\), we first evaluate \\( f_{\\theta} \\) over a discrete grid \\( \\mathcal{G} \\) and obtain \\( \\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}}) \\). Note that the above regularization is similar to the signed distance loss defined in Eq. (3), but with an important difference. Here we do not use ground-truth SDF values but instead, regularize the predicted signed distance. This regularization is applied alongside the main objective function to reconstruct the implicit surface commonly used to train DINs [18, 55] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.37, + 0.892, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S D F}} = \\sum_ {j = 1} ^ {N} \\sum_ {i} ^ {K} \\left| \\left| \\left(f _ {\\theta} \\left(\\lambda_ {j}, x _ {i}\\right) - s _ {i} \\right| \\right| _ {2} ^ {2} + \\frac {1}{\\sigma^ {2}} \\left\\| \\lambda_ {j} \\right\\| _ {2} ^ {2}, \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.423, + 0.892, + 0.468 + ], + "angle": 0, + "content": "where \\( \\mathrm{K} \\) is the number of points with annotated SDF values per shape \\( j \\), \\( s_i \\) is the ground-truth SDF value, and \\( \\sigma \\) is a parameter used to promote compactness in latent space [55]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Combining the standard training loss for DINs [55] alongside our regularization weighted with a scalar \\(\\alpha_{2}\\), our learning objective is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.525, + 0.892, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a i n}} = \\mathcal {L} _ {\\text {S D F}} + \\alpha_ {2} \\mathcal {L} _ {S D R}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.892, + 0.612 + ], + "angle": 0, + "content": "At the inference time, we reconstruct surfaces from point clouds by first recovering the optimal latent vectors (See Eq. (4) in [18]). Then, we predict a grid of SDF values \\(S_{\\mathcal{G}}\\) and extract the iso-surface using our SDC." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.621, + 0.889, + 0.638 + ], + "angle": 0, + "content": "4.3. Joint learning of implicit surface and meshing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Differently from existing data-driven meshing methods, SDC relies neither on explicit mesh data [59, 63] nor on axiomatically produced proxy meshes [11, 13]. This means that SDC can be used to optimize both SDF and reconstructed meshes jointly in an end-to-end manner. To demonstrate this we consider the task of predicting surface meshes from images. We first model the SDF representation of a shape given an image and then predict the mesh using SDC. In terms of the architectures, we follow MeshSDF [60]: we produce latent vectors \\(\\lambda\\) from input images with a ResNet-18 [28] encoder and learn an implicit shape representation of shapes using DeepSDF [55] conditioned on the aforementioned latent vectors. Instead of performing test-time optimization like MeshSDF, we predict SDF values on a regular grid \\(f_{\\theta}(\\mathcal{G})\\) and use SDC to reconstruct a mesh \\(\\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}})\\) by a simple forward pass of our SDC network. Initially, we train the networks (1) DeepSDF (with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "an image encoder) and (2) SDC independently, then jointly fine-tune them for 200 epochs. This fine-tuning uses an image as input and the mesh \\(\\mathcal{M}\\) from SDC as output, while minimizing the combined objective terms for surface reconstruction and meshing as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.181, + 0.47, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S V R}} = \\mathcal {L} _ {\\mathrm {S D F}} + \\alpha_ {3} \\mathcal {L} _ {S D R} + \\alpha_ {4} \\mathcal {L} _ {\\mathrm {D}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.209, + 0.47, + 0.286 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{SDF}}\\) and \\(\\mathcal{L}_{\\mathrm{SDR}}\\) (defined in Eq. (8) and Eq. (7) respectively) are used to reconstruct and regularize the implicit representation, while \\(\\mathcal{L}_D\\), defined in Eq. (3), is used to train SDC. Since the evaluated SDF might be noisy during training, we observed that using \\(\\mathcal{L}_N\\) results in instability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.299, + 0.4, + 0.316 + ], + "angle": 0, + "content": "5. Experiments, results and discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.371 + ], + "angle": 0, + "content": "We show the effectiveness of SDC across three tasks, namely, implicit function meshing, DIN training regularization and joint SDF and mesh prediction from images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.471, + 0.737 + ], + "angle": 0, + "content": "Evaluation Metrics. We use Chamfer Distance (CD), Normal Consistency (NC), Self-Intersection (SI), Edge Chamfer Distance (ECD), 3D-IoU, Precise Level-Set Discrepancy (LSD-P), and Approximate Level-Set Discrepancy (LSD-A) as metrics to compare all approaches. For measuring Level-Set discrepancy, we use two terms, Precise (LSD-P) and Approximate (LSD-A). The former is used in case where shapes are water-tight and analytical SDF can be measured. In such cases, we measure the discrepancy between the ground truth analytical SDF and the distance function measured from the generated mesh. LSD-A is used for non-watertight shapes and is measured by sampling points on the ground truth mesh and measuring its distance from the generated mesh's faces. Self-intersection is reported as the total number of intersecting triangles per mesh averaged across the test set. For the single-view reconstruction task, we also measure the Structural Similarity Index (SSIM) between the rendering of the reconstructed mesh and the ground truth mesh. We measure 3D-IoU for experiments where the generated mesh could topologically differ from the ground truth. CD, ECD, LSD are scaled by \\(10^{3}\\) while NC, SSIM 3D IoU are in \\(\\%\\) .We provide more details on all the metrics in the Suppl." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.747, + 0.399, + 0.764 + ], + "angle": 0, + "content": "5.1. Meshing analytical implicit functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.47, + 0.877 + ], + "angle": 0, + "content": "We first consider the surface mesh reconstruction task given the ground truth SDFs of shapes discretized on a regular grid. For this task, we train SDC on 3000 shapes from the \\(1^{st}\\) split of the ABC dataset [35] for 100 epochs. We scale each shape to fit a unit sphere and estimate the SDF on a regular grid of dimension \\(64^3\\) using SDFGen library1. We evaluate all methods on 300 shapes from the test set of" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.09, + 0.885, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.349, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Figure 3. We show three qualitative mesh reconstruction examples from the ABC dataset (Row 1,2) and the Thingi10k dataset (Row 3). Self-intersecting faces are highlighted in red. QEF-based learning methods NDC and UNDC show significantly higher self-intersecting faces along sharp edges while ours does not." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.436, + 0.89, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.57, + 0.893, + 0.613 + ], + "angle": 0, + "content": "Figure 4. Qualitative examples of meshing SDF produced by NGLOD [69]. MC, NMC, and NDC produce noisy meshes while SDC produces a smoother yet feature-preserving reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.838 + ], + "angle": 0, + "content": "the ABC dataset. Additionally, we evaluate on 300 shapes from the Thingi10K [77] dataset to demonstrate generalization to unseen data domains. We compare SDC against four baselines including two axiomatic baselines, namely the improved Marching Cubes (MC) [37] and the standard Dual Contouring (DC) [30] where positions of each vertex are estimated by solving the Quadratic Error Function (QEF), and three of the recent Neural Meshing techniques, Neural Marching Cubes (NMC) [11], Neural Dual Contouring [13] and FlexiCubes [65] respectively. For fairness, we compare to variants of comparable network capacity and for FlexiCubes, we compare without test-time optimization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We summarize our quantitative results in Table 1. Our self-supervised method SDC shows better performance and generalization to unseen data domains compared to the data-driven baselines across all metrics. While Marching" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.435, + 0.9 + ], + "angle": 0, + "content": "DatasetInput SDFMethodCD (↓)NC (↑)#SI (↓)ECD (↓)LSD-P (↓)ABCAnalyticalMC [37]4.792.104.816.1NMC [11]3.794.734.53.612.1DC [30]3.593.596.33.411.6NDC [13]3.694.343.13.511.9FlexiCubes [65]4.592.404.815.8Ours3.394.99.73.211.3Thingi10KAnalyticalMC [37]5.161.4013.315.3NMC [11]4.166.032.411.912.2DC [30]4.063.8105.611.511.8NDC [13]4.064.970.411.312.0FlexiCubes [65]5.255.3013.615.4Ours3.668.015.710.911.6Thingi10KPredicted [69]MC [37]5.856.7014.216.5NMC [11]4.763.230.112.612.9DC [30]6.254.0230.414.815.0NDC [13]4.762.842.612.813.1Ours4.265.715.912.212.4" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Table 1. Quantitative mesh reconstruction results on the ABC and the Thingi10k datasets. The best scores are highlighted in bold and the second-best scores are in italics. The first two row-blocks correspond to input SDF which was computed analytically while in the last row-block, we mesh the SDF predicted by NGLOD [69]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.471, + 0.562 + ], + "angle": 0, + "content": "Cubes [37] and FlexiCubes [65] do not produce any self-intersecting faces, the vertex placement is restricted for these methods, leading to inferior results with respect to other reconstruction metrics. In the qualitative examples visualized in Figure 3, we highlight an inherent issue with NDC and UNDC - they produce a considerable number of self-intersections along sharp edges. We reason this to be an undesirable trait that is inherited from the standard Dual Contouring as previously elaborated. SDC, on the other hand, bears a minimal number of self-intersections, as our training objective for mesh prediction is geometrically well-motivated. Additionally, we report quantitative and qualitative results for \\(128^{3}\\) grid resolution in the Suppl." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.395, + 0.59 + ], + "angle": 0, + "content": "5.2. Meshing predicted implicit functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Differently from the previous section, where the SDF is computed analytically, here, we use an open-source implementation (with suggested hyperparameters) of a recent deep implicit network NGLOD [69] to predict SDF values on a regular grid. We evaluate the meshing efficacy of SDC over a set of 300 shapes from the Thingi10k [77] dataset. To avoid possible biases, we chose these 300 shapes to be different from the ones used in the previous section. Our quantitative results are summarized in Table 1. Consistent with our observation from the previous section, SDC outperforms all baselines. We provide a qualitative example of meshing a noisy SDF in Figure 4. To generate the noisy SDF, we prematurely terminate NGLOD [69] fitting to the shape and evaluate different Neural Meshing methods. While all baselines produce a noisy reconstruction, SDC recovers a smoother surface while preserving sharp edges, thanks to our regularization and self-supervised losses. This suggests that our self-supervised training objectives and data augmentation are capable of generalizing to noisy imperfect predicted data without explicit training on this data." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.828, + 0.108 + ], + "angle": 0, + "content": "5.3. Regularizing implicit surface learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.446 + ], + "angle": 0, + "content": "As illustrated in the previous section, SDC successfully produces sharp meshes from both analytical and parameterized implicit functions. We now employ SDC for regularizing DIN training, focusing on the task of surface reconstruction using the ShapeNet dataset [9]. In particular, we evaluate over 4 categories of objects, namely, cars, planes, tables, and cabinets with 1000 shapes per category as our training set, and set aside a separate set of 200 unseen shapes for evaluation. We compare our regularization against three possible baselines. Firstly, we train Curriculum DeepSDF (CSDF) [18] over each category separately. Secondly, we re-train the same network by enforcing the Eikonal constraint, i.e. enforcing a unit norm for SDF's gradient in a setup similar to Implicit Geometric Regularization [23]. Thirdly, we use the supervised Dual Contouring baseline NDC [13] as the mesh-based regularizer. More specifically, we replace SDC with NDC in computing the regularization introduced in Eq. (7). For each shape, we sample 400,000 points in the shape volume, aggressively near the surface following [55] to supervise the SDF prediction. For fairness, we use the same sampled points for all methods we compare." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.793 + ], + "angle": 0, + "content": "We summarize our quantitative results in Table 2. We observe that regularizing the network using SDC produces a noticeable improvement in reconstruction across all metrics in comparison to baselines. Moreover, we observe a poorer reconstruction when using NDC [13] for regularization. We argue that this is due to their inability to handle noisy predicted SDF values as inputs as this approach learns to emulate solution to QEF (c.f. Eq. (1)), which is ill-defined for imperfect implicit functions. Our self-supervised loss functions do not rely on the ill-defined vertices produced by QEF minimization and align produced meshes to the input SDF grids, resulting in more plausible surface predictions. We believe this difference to be the key reason behind SDC's efficacy as a regularizer for training DINs. We also show three qualitative results in Figure 5. In some examples, we also observe topological differences although our regularization does not explicitly penalize it. Since the loss function which we minimize is highly non-convex, we believe that our regularization has aided in discovering more \"plausible\" latent space which could lead to better quality of implicit surfaces. We report additional qualitative and quantitative results in Suppl demonstrating the regularizer's generalization capabilities with other DINs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.74, + 0.818 + ], + "angle": 0, + "content": "5.4. Single view reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this section, we consider the task of reconstructing surface meshes from images. We use 4 categories from the ShapeNet dataset [9], namely planes, chairs, rifles, and tables. We train and evaluate our approach and baseline (using the official codebase) on the same training and evalua" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4687" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.095, + 0.462, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.339 + ], + "angle": 0, + "content": "Figure 5. Qualitative examples comparing surface reconstruction from the ShapeNet [9] dataset. The columns compare different methods, whose training objectives differ. Our mesh-based regularization produces reasonable results compared to baselines." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.362, + 0.454, + 0.441 + ], + "angle": 0, + "content": "
MethodCD (↓)NC (↑)3D IoU (↑)LSD-A (↓)
CSDF [18]3.078.183.93.1
+ Eikonal [23]3.278.083.03.2
+ NDC [13]2.978.484.82.9
+ SDC2.678.986.02.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.451, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Table 2. Comparison of surface reconstruction accuracy across object categories from the ShapeNet dataset [9]. Methods are trained with different regularization (see text) while evaluated alike." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.468, + 0.619 + ], + "angle": 0, + "content": "tion split for a fair comparison. We compare against three baselines, namely, MeshSDF [60], DISN [75], and a variant of our approach referred to as WoBW. More specifically, WoBW meshes the implicit function produced by pretrained SV-DIN using SDC, without the joint fine-tuning. For MeshSDF [60] and DISN [75] we use the official codebase for re-training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Quantitative results averaged across 4 object categories are summarized in Table 3. Notably, our SDC outperforms MeshSDF [60] without additional test-time optimization fitting a mesh to the rendering. In addition, our end-to-end model consistently outperforms the WoBW baseline, implying the efficacy of our mesh-based regularization and end-to-end training. We also show three qualitative examples in Figure 6. Our SDC demonstrates an improved ability to reconstruct sharp features in the predicted meshes, resulting in surfaces more faithful to the ground truth. In the third row, we highlight an example showing significant improvements in the quality of predicted geometry. It shows that our method is capable of better thin surface reconstructions for the same inputs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.337, + 0.861 + ], + "angle": 0, + "content": "6. Conclusion and future work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We introduced SDC, a Self-Supervised training approach for Neural Dual Contouring. Differently from previous" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.094, + 0.892, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.359 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of Single View Reconstruction (SVR) accuracy among three objects from ShapeNet [9] dataset. Input denotes the input view and GT denotes ground-truth mesh. SDC produces more sharper and plausible surfaces in comparison to baselines." + }, + { + "type": "table", + "bbox": [ + 0.497, + 0.398, + 0.858, + 0.478 + ], + "angle": 0, + "content": "
MethodCD (↓)NC (↑)SSIM (↑)LSD-A (↓)
DISN [75]13.070.282.010.5
MeshSDF [60]11.371.484.19.1
WoBW9.972.086.48.1
Ours9.172.888.57.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.488, + 0.891, + 0.516 + ], + "angle": 0, + "content": "Table 3. Quantitative results for single-view reconstruction across different object categories from ShapeNet [9] dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.794 + ], + "angle": 0, + "content": "work, we do not fit the generated mesh to an axiomatically produced proxy mesh but instead use geometrically motivated self-supervised losses which only depends on the input SDF. SDC shows consistent improvements over baselines when applied to end-to-end surface reconstruction and meshing tasks. In addition, our work forges a link between signed distance fields and meshes and shows how the former can be regularized by guiding the network to produce distance fields better corresponding to resulting meshes. Our meshing and regularization are applied on a regular grid, which limits the resolution of a reconstructed mesh. It would be interesting to explore adaptive grids within the learning framework. Finally, another related and interesting scope for future work is the exploration of strict manifoldness conditions [62] and theoretical guarantees of self-intersection free surfaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, and gifts from Ansys and Adobe Research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4688" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Nina Amenta and Marshall Bern. Surface reconstruction by voronoi filtering. Discrete & Computational Geometry, 22 (4):481-504, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.199 + ], + "angle": 0, + "content": "[2] Nina Amenta, Marshall Bern, and Manolis Kamvysselis. A new voronoi-based surface reconstruction algorithm. Proc. SIGGRAPH, pages 415-421, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.242 + ], + "angle": 0, + "content": "[3] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. Computational Geometry, 19(2-3):127-153, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Computer Graphics Forum, 36(1): 301-329, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[7] Jean-Daniel Boissonnat. Geometric structures for three-dimensional shape representation. ACM Transactions on Graphics (TOG), 3(4):266-286, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[8] Jean-Daniel Boissonnat and Steve Oudot. Provably good sampling and meshing of surfaces. Graphical Models, 67 (5):405-451, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.591 + ], + "angle": 0, + "content": "[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.592, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.649, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[11] Zhiqin Chen and Hao Zhang. Neural marching cubes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[12] Zhiqin Chen, Andrea Tagliausacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[13] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. arXiv preprint arXiv:2202.01999, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Angela Dai and Matthias Nießner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5574-5583, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[16] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 31-44, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[17] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE TRANSACTIONS on Information and Systems, 74(1):214-224, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[18] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevata, and Leonidas J. Guibas. Curriculum deepsdf. CoRR, abs/2003.08593, 2020. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[19] Herbert Edelsbrunner and Ernst P Mücke. Three-dimensional alpha shapes. ACM Transactions on Graphics (TOG), 13(1):43-72, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[20] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.367, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. arXiv preprint arXiv:1904.06447, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[22] S.F.F. Gibson. Using distance maps for accurate surface representation in sampled volumes. In IEEE Symposium on Volume Visualization (Cat. No.989EX300), pages 23-30, 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.54, + 0.894, + 0.61 + ], + "angle": 0, + "content": "[24] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[25] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.669, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[26] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph., 39(4), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.713, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[27] John C. Hart. Sphere tracing: a geometric method for the antialiased ray tracing of implicit surfaces. The Visual Computer, 12(10):527-545, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[29] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. ACM SIGGRAPH Computer Graphics, 26 (2), 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4689" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[31] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[33] Alireza Khatamian and Hamid R Arabnia. Survey on 3d surface reconstruction. Journal of Information Processing Systems, 12(3), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[34] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[36] Ravikrishna Kolluri, Jonathan Richard Shewchuk, and James F O'Brien. Spectral surface reconstruction from noisy point clouds. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 11-21. ACM, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[37] Thomas Lewiner, Hélio Lopes, Antonio Wilson Vieira, and Geovan Tavares. Efficient implementation of marching cubes' cases with topological guarantees. Journal of Graphics Tools, 8(2):1-15, 2003. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[38] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[39] Chen-Hsuan Lin, Oliver Wang, Bryan C Russell, Eli Shechtman, Vladimir G Kim, Matthew Fisher, and Simon Lucey. Photometric mesh optimization for video-aligned 3d object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 969-978, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[40] David B. Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. arXiv preprint arXiv:0000.00000, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[41] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[42] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1886-1895, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[43] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[44] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision, pages 68-84. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[45] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[46] N. Maruani, R. Klokov, M. Ovsjanikov, P. Alliez, and M. Desbrun. Voromesh: Learning watertight surface meshes with voronoi diagrams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[47] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. arXiv preprint arXiv:2204.07159, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[48] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[49] Mark Meyer, Mathieu Desbrun, Peter Schröder, and Alan H Barr. Discrete differential-geometry operators for triangulated 2-manifolds. In Visualization and mathematics III, pages 35-57. Springer, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[50] Niloy J. Mitra, Natasha Gelfand, Helmut Pottmann, and Leonidas Guibas. Registration of point cloud data from a geometric optimization perspective. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH Symposium on Geometry Processing, page 22-31, New York, NY, USA, 2004. Association for Computing Machinery. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[51] Patrick Mullen, Fernando De Goes, Mathieu Desbrun, David Cohen-Steiner, and Pierre Alliez. Signing the unsigned: Robust surface reconstruction from raw pointsets. Computer Graphics Forum, 29(5):1733-1741, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[52] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[53] G.M. Nielson. Dual marching cubes. In IEEE Visualization 2004, pages 489-496, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[54] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 9964-9973, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[55] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[57] Helmut Pottmann and Michael Hofer. Geometry of the squared distance function to curves and surfaces. In Mathematics and Visualization, pages 221-242. Springer Berlin Heidelberg, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[58] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4690" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[59] Marie-Julie Rakotaosaona, Paul Guerrero, Noam Algerman, Niloy J Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22-31, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[60] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[61] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[62] Scott Schaefer, Tao Ju, and Joe Warren. Manifold dual contouring. IEEE Transactions on Visualization and Computer Graphics, 13(3):610-619, 2007. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.419 + ], + "angle": 0, + "content": "[63] Nicholas Sharp and Maks Ovsjanikov. Pointtrinet: Learned triangulation of 3d point sets. In European Conference on Computer Vision, pages 762-778. Springer, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[65] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[66] Jonathan Shewchuk, Tamal K Dey, and Siu-Wing Cheng. Delaunay mesh generation. Chapman and Hall/CRC, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[68] G D Smith. Numerical solution of partial differential equations. Clarendon Press, Oxford, England, 3 edition, 1985. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[69] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[70] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[71] Graham M Treece, Richard W Prager, and Andrew H Gee. Regularised marching tetrahedra: improved iso-surface extraction. Computers & Graphics, 23(4):583-598, 1999. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[72] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-67, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[73] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10130-10139, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[74] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 41(2):641-676, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[75] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. Curran Associates, Inc., 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[76] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[77] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10, 000 3d-printing models. CoRR, abs/1605.04797, 2016. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4691" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_origin.pdf b/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9896cbff0628755fef38bf4a7ef09a83baaba6a3 --- /dev/null +++ b/2024/Self-Supervised Dual Contouring/0eb29fd9-574c-47d8-9f41-53c9d4c10474_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2041c5491576f42365a236770c15994de0ab1d8d438596e06e0a4351bcce31d2 +size 9366463 diff --git a/2024/Self-Supervised Dual Contouring/full.md b/2024/Self-Supervised Dual Contouring/full.md new file mode 100644 index 0000000000000000000000000000000000000000..934de30aac9a5a892c4d67b56304c4ddc86bf8d1 --- /dev/null +++ b/2024/Self-Supervised Dual Contouring/full.md @@ -0,0 +1,331 @@ +# Self-Supervised Dual Contouring + +Ramana Sundararaman + +LIX, Ecole Polytechnique + +sundararman@lix.polytechnique.fr + +Roman Klokov + +LIX, Ecole Polytechnique + +klokov@lix.polytechnique.fr + +Maks Ovsjanikov +LIX, Ecole Polytechnique + +maks@lix.polytechnique.fr + +# Abstract + +Learning-based isosurface extraction methods have recently emerged as a robust and efficient alternative to axiomatic techniques. However, the vast majority of such approaches rely on supervised training with axiomatically computed ground truths, thus potentially inheriting biases and data artefacts of the corresponding axiomatic methods. Steering away from such dependencies, we propose a self-supervised training scheme to the Neural Dual Contouring meshing framework, resulting in our method: Self-Supervised Dual Contouring (SDC). Instead of optimizing predicted mesh vertices with supervised training, we use two novel self-supervised loss functions that encourage the consistency between distances to the generated mesh up to the first order. Meshes reconstructed by SDC surpass existing data-driven methods in capturing intricate details while being more robust to possible irregularities in the input. Furthermore, we use the same self-supervised training objective linking inferred mesh and input SDF, to regularize the training process of Deep Implicit Networks (DINs). We demonstrate that the resulting DINs produce higher-quality implicit functions, ultimately leading to more accurate and detail-preserving surfaces compared to prior baselines for different input modalities. Finally, we demonstrate that our self-supervised losses improve meshing performance in the single-view reconstruction task by enabling joint training of predicted SDF and resulting output mesh. We open-source our code at https://github.com/Sentient07/SDC. + +# 1. Introduction + +Surface mesh extraction from implicit functions, often referred to as isosurfacing [30, 37, 45, 62], is a fundamental problem in computer graphics and geometry processing as the quality of reconstructed mesh impacts algorithms used in numerous downstream tasks [49]. Prominent primal ax + +![](images/e44ea2a865b510402cbcf2afe5a97db9dea85510f5c1b72ea4ccb07822bd9042.jpg) +Figure 1. We propose SDC, a data-driven self-supervised isosurfacing method that reconstructs better feature-preserving meshes from learnt implicit functions compared to the baseline NDC [13]. Our framework is extendable as a regularizer for training Deep Implicit Networks (DINs) for improved reconstruction. + +iomatic methods such as Marching Cubes [37, 45], are computationally efficient but fail to reconstruct sharp features. Dual methods [22, 30, 34, 62] achieve sharper reconstructions using Quadratic Error Function (QEF) but their performance heavily relies on the data fidelity. + +Following the ubiquitous success of neural methods in 2D and 3D applications, recent learning-based meshing approaches [11, 13] propose to predict mesh vertices and connectivity from various inputs in a feed-forward manner. When trained on extensive, augmented mesh collections, these methods demonstrate enhanced generalization and yield higher-quality reconstructions. However, these approaches rely either directly on the ground truth meshes [63] or on some intermediate pre-computed proxy meshes, like complex tessellation templates [11] and outputs of axiomatic approaches [13]. + +Neural Dual Contouring (NDC) [13], the most relevant method to our work, is a promising recent approach that enhances standard Dual Contouring (DC) [30, 62] with a data-driven pipeline. It aims to replicate DC-produced meshes, resulting in a framework capable of generating high-quality, + +feature-preserving meshes. However, since vertex estimation of DC follows QEF minimization, it requires precise SDF values at grid points and normals at edge intersections for training. Due to QEF minimization being ill-defined in particular scenarios [30], and susceptible to irregularities in the input SDF, NDC may exhibit a bias towards poor triangulation quality for challenging inputs and intricate details. Furthermore, its reliance on DC meshes for training restricts its use in an end-to-end scenario where both the implicit function and the explicit mesh can be optimized jointly. + +To overcome these limitations, we propose Self-Supervised Dual Contouring (SDC), a meshing framework that extends Neural Dual Contouring with two novel geometrically motivated self-supervised losses to learn the mesh vertex positions from SDF inputs while relying on sign changes to extract connectivity. The first self-supervised objective is a distance-based loss which minimizes the difference between the input implicit function values at points close to the surface and distances from the same points to the reconstructed mesh. This ensures that the predicted reconstructed mesh agrees with the input SDF. Secondly, we align the normals corresponding to reconstructed faces to those estimated from the input grid of SDFs. Since our training objective relies only on the input SDF and is agnostic to any explicit representation of the shape, e.g., based either on surface samples [26, 63] or on mesh vertices [11, 13], we refer to our approach as Self-Supervised. Together these two terms produce vertices that best explain the input grid of signed distance function values up to the first order like dual methods, while avoiding shortcomings of the quadratic error function (QEF) minimization [13, 30]. In particular, unlike DC, our method produces vertices that lie within each voxel-cell, and optimization of our self-supervised objectives leads to overall better results for input data with imperfections. As a result, we demonstrate better performance than supervised baselines in obtaining sharper meshes with negligible self-intersections and better generalization to unseen input data. + +In addition, leveraging the connection between reconstructed meshes and input SDF grids, we propose a mesh-based regularization for training Deep Implicit Networks (DINs) - MLPs which represent zero-level sets of a shape. In a standard setup, training of DINs is completely decoupled from meshing, which is only performed a posteriori. Thus, existing DINs might produce implicit functions that do not correspond to a distance function arising from an extracted mesh. This motivates us to use SDC as a regularization to train DINs. In particular, we enforce the distance function produced by DINs during training to remain as close as possible to the distance function arising from the mesh produced by SDC. This forms the converse of our self-supervised meshing losses, but now geared to produce more coherent implicit functions. We demonstrate that this + +regularization can be directly adopted into existing training paradigms for Deep Implicit Networks and it that leads to more plausible reconstructions. + +Finally, since SDC training does not require ground truth meshes, it can be seamlessly integrated within end-to-end training paradigms. To highlight this, we consider the task of mesh reconstruction from images where we first construct an implicit representation given an image using an existing approach [55, 60] and then reconstruct a mesh using SDC. The resulting model is trained end-to-end by jointly minimizing the standard surface reconstruction loss, our self-supervised loss and the proposed regularization for Deep Implicit Networks. The resulting approach demonstrates compelling surface reconstruction efficacy across multiple object categories from the ShapeNet dataset [9]. In summary, our contributions are as follows: + +- We extend Neural Dual Contouring with two novel self-supervised loss functions ensuring consistency between input SDF and distance to output meshes, which is geometrically better motivated than the standard QEF minimization. +- We introduce a novel regularizer for training Deep Implicit Networks (DINs), by penalizing SDFs that do not correspond to underlying meshes. +- We show that SDC generalizes better than supervised meshing methods, and further demonstrate its utility in feature-preserving mesh reconstruction from images. + +# 2. Related work + +Surface reconstruction and meshing tasks have been extensively studied within computer graphics and vision. Below, we review the methods that are most closely related to ours and refer interested readers to several surveys [5, 33, 66] for a more in-depth discussion. + +Axiomatic contouring techniques. The process of isosurface extraction from volumetric data is referred to as contouring. The most widely-used pipeline for surface reconstruction consists of two major steps. First, volumetric data is computed using a signed distance function [14, 29, 32, 51]. In the second step, a mesh is extracted from this representation, using Marching Cubes [45], Dual Contouring [30] or related approaches [17, 52, 71]. This pipeline can produce very high quality meshes, but is sensitive to the quality of input data and is not differentiable and thus does not easily fit within modern learning-based pipelines [38]. Another line of approaches is based on Delaunay triangulations, [7, 8, 36], as well as closely-related constructions [2, 3, 6, 19] which often come with strong theoretical guarantees and are typically geared towards preserving some input point set [1, 6]. Such methods, however, provide little control over the output triangulation. Moreover, classical approaches are typically not differentiable and thus cannot be easily integrated into learning pipelines. + +Data-driven mesh reconstruction. To address differentiability issues of classical mesh reconstruction approaches, a number of data-driven techniques have recently been proposed. This includes differentiable variants of contouring methods, such as the Marching Cubes [11, 38], dual Marching Cubes [65], Marching Tetrahedra [64], and Dual Contouring [13]. Our approach can be categorized into this family of differentiable contouring techniques. In addition, surface reconstruction has also been tackled using template-based techniques which fit a template mesh to the input [31, 39, 42], or deform an initial mesh while potentially updating its connectivity [20, 54, 72]. Furthermore, other local approaches propose to fit parameterized surface patches to points [24, 73], or to decompose space into convex sets [12, 16, 46]. Finally, several learning-based methods mesh an input point set [15, 44, 59, 63]. These approaches strongly rely on the given input point cloud and can thus be sensitive to artifacts and noise. + +Neural fields. There has been a recent surge in methods for learning implicit functions, referred to as neural fields [74]. This includes pioneering methods for predicting grid occupancy and signed distance functions [10, 48, 55] as well as their many follow-up works [21, 61, 67, 70]. Often, training Deep Implicit Networks introduces inductive bias [58] thereby producing inexplicable surface behaviors. Multiple regularization techniques [4, 23, 40, 41, 43, 58, 67, 70] have been proposed to address that issue. They largely focus on either learning high-frequency signals [40, 70] or designing loss functions based to guide the gradient level-set [23, 41, 58]. While the latter category of methods has convergence guarantees [41], the former helps in producing detail preserving reconstruction. Our work, on the other hand, provides a novel regularization by utilizing a meshing framework to ensure that the predicted signed distance field arises from a mesh. Similar to recent works [25, 47, 56, 60, 64, 65, 76] which optimize the mesh by controlling the underlying implicit function, our approach also enables back-propagation from the mesh to the input SDF. However, our approach is non-iterative and produces a mesh in one feed-forward pass. + +# 3. Background + +We first begin by providing the notations used across the paper in Section 3.1 and review the Dual Contouring and Neural Dual Contouring algorithms in Section 3.2. + +# 3.1. Notations + +We let $f: \mathbb{R}^3 \to \mathbb{R}$ be an implicit function that determines the signed distance of a query point $x \in \mathbb{R}^3$ from an underlying surface $\chi$ . We let $f_{\theta}$ denote the implicit function produced by a neural network parameterized by $\theta$ . We use $\mathcal{G}$ to denote an $l \times m \times n$ regular grid in the Euclidean space, + +$g \in \mathcal{G}$ to be a node in the grid and $e_{ij} \in \mathcal{G}$ to be an edge between adjacent nodes $g_i, g_j$ respectively. We denote $\mathcal{S}_{\mathcal{G}}$ as the discretization of $f$ on $\mathcal{G}$ i.e. SDF evaluated at every grid node. Finally, we let $\lambda$ denote the latent vector associated with a shape computed using an encoder. + +# 3.2. Dual contouring + +The Dual Contouring (DC) [30] framework is designed to produce quadrilateral meshes from input signed distance functions discretized on a grid $S_{\mathcal{G}}$ . It can be decomposed into two main steps: (1) construction of mesh faces $\mathcal{F}$ ; (2) computation of mesh vertices $\mathcal{V}$ . To construct mesh faces, DC considers adjacent nodes $g_i, g_j \in \mathcal{G}$ such that their signed distance are of opposite signs, $f(g_i) \neq f(g_j)$ . Then, for those pairs, a unique quadrilateral face $q_k$ is constructed such that $q_k$ crosses the edge $e_{ij}$ . Vertices of this quadrilateral $q_k$ reside in adjacent grid cells. The faces of the resulting mesh are obtained as a union of all quadrilateral faces $\mathcal{F} = \bigcup_{k \in I} q_k$ , where $I$ is a set of indices of grid edges connecting grid vertices of different signs. Final mesh vertex positions within each cell are obtained by minimizing the Quadratic Error Function (QEF). More precisely, given $p_e$ to be the intersection position between edge $e_{ij}$ and the face $q_{ij}$ (along $e_{ij}$ ) and $n_e$ to be the gradients of the SDF at positions $e_{ij}$ (or surface normals), vertices are determined for each grid cell $c$ by solving: + +$$ +v _ {c} = \underset {x} {\arg \min } \sum_ {e \in I _ {c}} \left[ n _ {e} * \left(x - p _ {e}\right) \right], \tag {1} +$$ + +where $I_{c}$ is a set of indices of intersected grid edges corresponding to cell $c$ . Recently introduced Neural Dual Contouring [13] attempts to emulate Dual Contouring via a neural network in order to reduce the inference time while achieving optimal reconstruction. + +# 4. Method + +We describe our Self-supervised Dual Contouring (SDC) loss functions in Section 4.1, discuss how to use them for regularization while training deep implicit networks in Section 4.2, and apply them for mesh reconstruction from input images in Section 4.3. + +# 4.1. Self-supervised dual contouring (SDC) + +We introduce a Self-supervised training scheme for Dual Contouring, and refer to a model trained with it as to SDC. It excludes the dependence on vertex positions obtained via QEF minimization. SDC produces quad meshes from the signed distance function discretized on a grid $S_{\mathcal{G}}$ similarly to any Dual Contouring method. We triangulate quadrilaterals by joining the top-left and bottom-right vertex as a convention. $S_{\mathcal{G}}$ can either be measured [27] or predicted using a DIN [55] $f_{\theta}$ . SDC consists of two main components constructing faces and vertices. To construct faces, + +we follow the Dual Contouring [30] algorithm to determine the connectivity based on signed distance value at grid nodes. A quadrilateral face is constructed in the cells incident on every edge $e_{ij} \in \mathcal{G}$ connecting grid nodes of different signs $\mathrm{sign}(f(g_i)) \neq \mathrm{sign}(f(g_j))$ . For vertex prediction, we use a 6-layered 3D Convolutional Neural Network $h_\phi : S_\mathcal{G} \to \mathcal{P} \in \mathbb{R}^{(l-1) \times (m-1) \times (n-1) \times 3}$ , which takes the grid of signed distance values as input and predicts a single point per each grid cell as the output. Then, we apply a masking to select grid cells, that bear node SDF values of opposite signs. Please refer to Suppl. for more details. + +SDC is trained without explicit fitting to the ground truth meshes using two loss functions. Firstly, to ensure formal correspondence of produced surfaces to input SDF grids, we propose a distance-based loss $\mathcal{L}_D$ . This objective minimizes the differences between 1) the absolute distance values at the nodes of the discretization grid $f(g_{i})\in S_{\mathcal{G}}$ which are provided as inputs and 2) the distance $\mathbf{d}^p (g_i,\mathcal{M})$ measured from the same point $g_{i}$ to the predicted mesh $\mathcal{M}$ . For a smooth or gently undulating surface, this distance provides a reasonable approximation of the underlying geometry [50, 57]. However, in the presence of sharp features, the local geometry changes direction rapidly. Close query points residing on different sides of a sharp feature have similar distances to the mesh but may be projected to very different locations on the actual surface. To address this, we also introduce a normal consistency loss, which aims to align the normals estimated at the nodes of the discretization grid $f(g_{i})\in S_{\mathcal{G}}$ and normals corresponding to the generated mesh face. + +We refer to the losses proposed above as self-supervised since the training signal for our pipeline comes from the input alone. We do not use any explicit surface discretization for training and our losses are solely based on the input implicit function values and the generated mesh. In contrast, we refer to methods that predict vertices to fit a specific mesh structure as supervised meshing methods [11, 13]. We provide visual explanations of our loss functions in Figure 2. In summary, our combined self-supervised objective for the vertex prediction network is given by: + +$$ +\mathcal {L} _ {\text {M e s h}} = \mathcal {L} _ {D} + \alpha_ {1} \mathcal {L} _ {N}, \tag {2} +$$ + +where $\alpha_{1}$ is a hyperparameter to weigh $\mathcal{L}_N$ . + +Distance loss. The objective of our distance loss is to minimize the discrepancy between surfaces produced by SDC and zero-level sets defined by the input SDFs. In order to achieve this, given grid nodes $g_{i} \in \mathcal{G}$ , we measure the distance between the grid nodes and the mesh produced by SDC and penalize its deviation from the input absolute distance values $f(g_{i})$ . In summary, our distance loss is defined + +![](images/e34092315d8f2d81533f2c84bc32d0ad6ef55fd34585f1022ad69c68b0c71db6.jpg) +(a) + +![](images/5fdcefafe6f6a567cba586466686ddbfc38a6ce383938b459c3487220e67a196.jpg) +(b) +Figure 2. Illustration of our self-supervised losses. (a) Measurements of $\mathbf{d}^p$ between grid nodes and the closest point on a mesh face. For visualization purposes, we only show this estimation at one node. (b) Normal consistency loss is measured as the discrepancy between interpolated normals (orange) and the face normal (green). Red nodes denote negative SDF grid and green denotes positive. The constructed quad is shown in transparent purple. + +as: + +$$ +\mathcal {L} _ {D} = \sum_ {g _ {i} \in \mathcal {G}} \left\| \operatorname {a b s} \left(f \left(g _ {i}\right)\right) - \mathbf {d} ^ {p} \left(g _ {i}, \mathcal {M}\right) \right\| _ {2} ^ {2}, \tag {3} +$$ + +where $\mathbf{d}^p$ is the point-to-face distance, computed between $g_{i}$ and its projection on $\mathcal{M}$ . For computational efficiency, we compute $\mathcal{L}_D$ only at grid nodes that are close to the surface. Please refer to Suppl for details. + +Normal consistency loss. We introduce a novel normal consistency loss, $\mathcal{L}_N$ which aims to align the level-sets up to the first order. In particular, given the grid of scalar SDFs, we compute normals as gradients $n(g_i) \coloneqq f'(g_i) / ||f'(g_i)||_2$ in grid nodes $g_i$ using Five-point stencil, a commonly used finite difference approach for solving partial differential equations on regular grids [68]. More details on gradient estimation is provided in the Suppl. Given an edge $e_{ij} \in \mathcal{G}$ connecting nodes $g_i, g_j$ of the opposite signs, we first estimate a point $p \in e_{ij}$ along the edge where $f(p) = 0$ . The point $p$ can be represented in terms of $g_i$ and $g_j$ as $p = (1 - t)g_i + tg_j$ . Where $t$ is a parameter that determines the position of $p$ along the edge $e_{ij}$ . Solving for $p = 0$ , the parameter $t$ can be expressed in closed form as: + +$$ +t = \frac {f \left(g _ {i}\right)}{f \left(g _ {i}\right) - f \left(g _ {j}\right)}. \tag {4} +$$ + +Similarly, we can estimate the normal at the point $p$ by linear interpolation. Using the previously computed parameter $t$ , the normal at $p$ can be represented as: + +$$ +n (p) = \operatorname {s i g n} \left(f \left(g _ {i}\right)\right) (1 - t) n \left(g _ {i}\right) + \operatorname {s i g n} \left(f \left(g _ {j}\right)\right) t n \left(g _ {j}\right). \tag {5} +$$ + +This expression combines the known gradients at the neighboring nodes, weighted by the parameter $t$ , to estimate the gradient at a point $p$ . Since the inward normal is oriented + +opposite to the outward normal, we multiply by the sign for consistency. As our mesh is constructed following Dual Contouring [30], we denote the quad-face constructed dual to the edge $e_{ij}$ as $q_{ij}$ and its normal as $n(q_{ij})$ . With the aim of aligning these two normals, our normal consistency loss can be written as: + +$$ +\mathcal {L} _ {N} = 1 - \frac {n \left(q _ {i j}\right) \cdot n (p)}{\| n \left(q _ {i j}\right) \| \| n (p) \|.} \tag {6} +$$ + +Training augmentation. At training time, we add synthetic noise which is zero-mean Gaussian noise with a standard deviation equivalent to a third grid edge length $e_{ij}$ to the input SDF. Since SDC constructs mesh faces based on Dual Contouring, a noise that induces a sign change to the initial SDF will also lead to a topological change. More precisely, augmentation will lead to inclusion of vertices (and faces) in cells whose nodes differ in sign. Similarly, faces are removed dual to cells whose signs agree after augmentation. More training details are provided in the Suppl. + +# 4.2. Deep implicit network regularizer + +Deep Implicit Networks (DINs), commonly represented by Multi-Layer Perceptrons (MLP), are trained to acquire an implicit representation of a shape and are primarily utilized for various inference-based surface reconstruction tasks. In a traditional setup, DINs are trained by direct supervision of the SDF values at query points sampled close to the surface. Such a training scheme, as also observed by previous works [23, 41, 58], is known to produce inexplicable behavior, such as ambiguous level sets. To address this, we propose a novel mesh-based training regularization for DINs to ensure that the Distance Field (DF) produced by DIN agrees well with the resulting extracted mesh. Our regularization minimizes the discrepancy between the predicted DF (absolute value of SDF) and the distance computed from the mesh corresponding to the SDF extracted using SDC. + +We assume that we are endowed with a learnable DIN $f_{\theta}(\lambda, p): p \in \mathbb{R}^3 \to s_p \in \mathbb{R}$ , represented as an MLP, whose weights $\theta$ encode the implicit surface of a shape conditioned by a latent vector $\lambda$ , $p$ is a point in space such that $f_{\theta}(\lambda, p) = s_p$ is its signed distance from the zero-level set. Then, for a given $S_G$ - a discretization of $f_{\theta}$ on a regular grid and $\mathcal{M}^*(S_G)$ - a mesh that is produced following any iso-surfacing algorithm, it is not necessarily true that $\mathrm{abs}(f_{\theta}(p, \lambda)) = \mathbf{d}^p(p, \mathcal{M}^*(S_G))$ , where $\mathbf{d}^p$ is the point-to-face distance. While this disparity is viewed as an inherent drawback of primal iso-surfacing methods [53], for dual methods, this disparity can, in principle, be minimized due to their ability to preserve sharp details in the reconstructed meshes. This motivates us to use our SDC for iso-surface extraction and regularize the training of DINs by minimizing the aforementioned disparity. Assuming we have access to a dataset consisting of $N$ shapes and a pre-trained + +SDC with frozen weights, our regularization objective can be summarized as follows: + +$$ +\mathcal {L} _ {S D R} = \sum_ {j = 1} ^ {N} \sum_ {g _ {i} \in \mathcal {G}} | | \operatorname {a b s} \left(f _ {\theta} \left(\lambda_ {j}, g _ {i}\right) - \mathbf {d} ^ {p} \left(g _ {i}, \mathcal {M}\right) \right| | _ {2} ^ {2}. \tag {7} +$$ + +This regularization penalizes the discrepancy between the DF at grid nodes predicted by the DIN $f_{\theta}(g_i)$ and the distance $\mathbf{d}^p$ computed between the grid nodes and the mesh $\mathcal{M} = \mathrm{SDC}(S_{\mathcal{G}})$ . To obtain $\mathcal{M}$ , we first evaluate $f_{\theta}$ over a discrete grid $\mathcal{G}$ and obtain $\mathcal{M} = \mathrm{SDC}(S_{\mathcal{G}})$ . Note that the above regularization is similar to the signed distance loss defined in Eq. (3), but with an important difference. Here we do not use ground-truth SDF values but instead, regularize the predicted signed distance. This regularization is applied alongside the main objective function to reconstruct the implicit surface commonly used to train DINs [18, 55] as follows: + +$$ +\mathcal {L} _ {\mathrm {S D F}} = \sum_ {j = 1} ^ {N} \sum_ {i} ^ {K} \left| \left| \left(f _ {\theta} \left(\lambda_ {j}, x _ {i}\right) - s _ {i} \right| \right| _ {2} ^ {2} + \frac {1}{\sigma^ {2}} \left\| \lambda_ {j} \right\| _ {2} ^ {2}, \right. \tag {8} +$$ + +where $\mathrm{K}$ is the number of points with annotated SDF values per shape $j$ , $s_i$ is the ground-truth SDF value, and $\sigma$ is a parameter used to promote compactness in latent space [55]. + +Combining the standard training loss for DINs [55] alongside our regularization weighted with a scalar $\alpha_{2}$ , our learning objective is formulated as follows: + +$$ +\mathcal {L} _ {\text {t r a i n}} = \mathcal {L} _ {\text {S D F}} + \alpha_ {2} \mathcal {L} _ {S D R}. \tag {9} +$$ + +At the inference time, we reconstruct surfaces from point clouds by first recovering the optimal latent vectors (See Eq. (4) in [18]). Then, we predict a grid of SDF values $S_{\mathcal{G}}$ and extract the iso-surface using our SDC. + +# 4.3. Joint learning of implicit surface and meshing + +Differently from existing data-driven meshing methods, SDC relies neither on explicit mesh data [59, 63] nor on axiomatically produced proxy meshes [11, 13]. This means that SDC can be used to optimize both SDF and reconstructed meshes jointly in an end-to-end manner. To demonstrate this we consider the task of predicting surface meshes from images. We first model the SDF representation of a shape given an image and then predict the mesh using SDC. In terms of the architectures, we follow MeshSDF [60]: we produce latent vectors $\lambda$ from input images with a ResNet-18 [28] encoder and learn an implicit shape representation of shapes using DeepSDF [55] conditioned on the aforementioned latent vectors. Instead of performing test-time optimization like MeshSDF, we predict SDF values on a regular grid $f_{\theta}(\mathcal{G})$ and use SDC to reconstruct a mesh $\mathcal{M} = \mathrm{SDC}(S_{\mathcal{G}})$ by a simple forward pass of our SDC network. Initially, we train the networks (1) DeepSDF (with + +an image encoder) and (2) SDC independently, then jointly fine-tune them for 200 epochs. This fine-tuning uses an image as input and the mesh $\mathcal{M}$ from SDC as output, while minimizing the combined objective terms for surface reconstruction and meshing as follows: + +$$ +\mathcal {L} _ {\mathrm {S V R}} = \mathcal {L} _ {\mathrm {S D F}} + \alpha_ {3} \mathcal {L} _ {S D R} + \alpha_ {4} \mathcal {L} _ {\mathrm {D}}, \tag {10} +$$ + +where $\mathcal{L}_{\mathrm{SDF}}$ and $\mathcal{L}_{\mathrm{SDR}}$ (defined in Eq. (8) and Eq. (7) respectively) are used to reconstruct and regularize the implicit representation, while $\mathcal{L}_D$ , defined in Eq. (3), is used to train SDC. Since the evaluated SDF might be noisy during training, we observed that using $\mathcal{L}_N$ results in instability. + +# 5. Experiments, results and discussion + +We show the effectiveness of SDC across three tasks, namely, implicit function meshing, DIN training regularization and joint SDF and mesh prediction from images. + +Evaluation Metrics. We use Chamfer Distance (CD), Normal Consistency (NC), Self-Intersection (SI), Edge Chamfer Distance (ECD), 3D-IoU, Precise Level-Set Discrepancy (LSD-P), and Approximate Level-Set Discrepancy (LSD-A) as metrics to compare all approaches. For measuring Level-Set discrepancy, we use two terms, Precise (LSD-P) and Approximate (LSD-A). The former is used in case where shapes are water-tight and analytical SDF can be measured. In such cases, we measure the discrepancy between the ground truth analytical SDF and the distance function measured from the generated mesh. LSD-A is used for non-watertight shapes and is measured by sampling points on the ground truth mesh and measuring its distance from the generated mesh's faces. Self-intersection is reported as the total number of intersecting triangles per mesh averaged across the test set. For the single-view reconstruction task, we also measure the Structural Similarity Index (SSIM) between the rendering of the reconstructed mesh and the ground truth mesh. We measure 3D-IoU for experiments where the generated mesh could topologically differ from the ground truth. CD, ECD, LSD are scaled by $10^{3}$ while NC, SSIM 3D IoU are in $\%$ .We provide more details on all the metrics in the Suppl. + +# 5.1. Meshing analytical implicit functions + +We first consider the surface mesh reconstruction task given the ground truth SDFs of shapes discretized on a regular grid. For this task, we train SDC on 3000 shapes from the $1^{st}$ split of the ABC dataset [35] for 100 epochs. We scale each shape to fit a unit sphere and estimate the SDF on a regular grid of dimension $64^3$ using SDFGen library1. We evaluate all methods on 300 shapes from the test set of + +![](images/4766a179210be62b82f858e9b73591dee6bf9bf098d0e7189fbad6d4ebd1ee33.jpg) +Figure 3. We show three qualitative mesh reconstruction examples from the ABC dataset (Row 1,2) and the Thingi10k dataset (Row 3). Self-intersecting faces are highlighted in red. QEF-based learning methods NDC and UNDC show significantly higher self-intersecting faces along sharp edges while ours does not. + +![](images/e11d341c366f60d74b4630ac11f031b90070d0ff23e9ab98f98c48f5d284b781.jpg) +Figure 4. Qualitative examples of meshing SDF produced by NGLOD [69]. MC, NMC, and NDC produce noisy meshes while SDC produces a smoother yet feature-preserving reconstruction. + +the ABC dataset. Additionally, we evaluate on 300 shapes from the Thingi10K [77] dataset to demonstrate generalization to unseen data domains. We compare SDC against four baselines including two axiomatic baselines, namely the improved Marching Cubes (MC) [37] and the standard Dual Contouring (DC) [30] where positions of each vertex are estimated by solving the Quadratic Error Function (QEF), and three of the recent Neural Meshing techniques, Neural Marching Cubes (NMC) [11], Neural Dual Contouring [13] and FlexiCubes [65] respectively. For fairness, we compare to variants of comparable network capacity and for FlexiCubes, we compare without test-time optimization. + +We summarize our quantitative results in Table 1. Our self-supervised method SDC shows better performance and generalization to unseen data domains compared to the data-driven baselines across all metrics. While Marching + +
DatasetInput SDFMethodCD (↓)NC (↑)#SI (↓)ECD (↓)LSD-P (↓)
ABCAnalyticalMC [37]4.792.104.816.1
NMC [11]3.794.734.53.612.1
DC [30]3.593.596.33.411.6
NDC [13]3.694.343.13.511.9
FlexiCubes [65]4.592.404.815.8
Ours3.394.99.73.211.3
Thingi10KAnalyticalMC [37]5.161.4013.315.3
NMC [11]4.166.032.411.912.2
DC [30]4.063.8105.611.511.8
NDC [13]4.064.970.411.312.0
FlexiCubes [65]5.255.3013.615.4
Ours3.668.015.710.911.6
Thingi10KPredicted [69]MC [37]5.856.7014.216.5
NMC [11]4.763.230.112.612.9
DC [30]6.254.0230.414.815.0
NDC [13]4.762.842.612.813.1
Ours4.265.715.912.212.4
+ +Table 1. Quantitative mesh reconstruction results on the ABC and the Thingi10k datasets. The best scores are highlighted in bold and the second-best scores are in italics. The first two row-blocks correspond to input SDF which was computed analytically while in the last row-block, we mesh the SDF predicted by NGLOD [69] + +Cubes [37] and FlexiCubes [65] do not produce any self-intersecting faces, the vertex placement is restricted for these methods, leading to inferior results with respect to other reconstruction metrics. In the qualitative examples visualized in Figure 3, we highlight an inherent issue with NDC and UNDC - they produce a considerable number of self-intersections along sharp edges. We reason this to be an undesirable trait that is inherited from the standard Dual Contouring as previously elaborated. SDC, on the other hand, bears a minimal number of self-intersections, as our training objective for mesh prediction is geometrically well-motivated. Additionally, we report quantitative and qualitative results for $128^{3}$ grid resolution in the Suppl. + +# 5.2. Meshing predicted implicit functions + +Differently from the previous section, where the SDF is computed analytically, here, we use an open-source implementation (with suggested hyperparameters) of a recent deep implicit network NGLOD [69] to predict SDF values on a regular grid. We evaluate the meshing efficacy of SDC over a set of 300 shapes from the Thingi10k [77] dataset. To avoid possible biases, we chose these 300 shapes to be different from the ones used in the previous section. Our quantitative results are summarized in Table 1. Consistent with our observation from the previous section, SDC outperforms all baselines. We provide a qualitative example of meshing a noisy SDF in Figure 4. To generate the noisy SDF, we prematurely terminate NGLOD [69] fitting to the shape and evaluate different Neural Meshing methods. While all baselines produce a noisy reconstruction, SDC recovers a smoother surface while preserving sharp edges, thanks to our regularization and self-supervised losses. This suggests that our self-supervised training objectives and data augmentation are capable of generalizing to noisy imperfect predicted data without explicit training on this data. + +# 5.3. Regularizing implicit surface learning + +As illustrated in the previous section, SDC successfully produces sharp meshes from both analytical and parameterized implicit functions. We now employ SDC for regularizing DIN training, focusing on the task of surface reconstruction using the ShapeNet dataset [9]. In particular, we evaluate over 4 categories of objects, namely, cars, planes, tables, and cabinets with 1000 shapes per category as our training set, and set aside a separate set of 200 unseen shapes for evaluation. We compare our regularization against three possible baselines. Firstly, we train Curriculum DeepSDF (CSDF) [18] over each category separately. Secondly, we re-train the same network by enforcing the Eikonal constraint, i.e. enforcing a unit norm for SDF's gradient in a setup similar to Implicit Geometric Regularization [23]. Thirdly, we use the supervised Dual Contouring baseline NDC [13] as the mesh-based regularizer. More specifically, we replace SDC with NDC in computing the regularization introduced in Eq. (7). For each shape, we sample 400,000 points in the shape volume, aggressively near the surface following [55] to supervise the SDF prediction. For fairness, we use the same sampled points for all methods we compare. + +We summarize our quantitative results in Table 2. We observe that regularizing the network using SDC produces a noticeable improvement in reconstruction across all metrics in comparison to baselines. Moreover, we observe a poorer reconstruction when using NDC [13] for regularization. We argue that this is due to their inability to handle noisy predicted SDF values as inputs as this approach learns to emulate solution to QEF (c.f. Eq. (1)), which is ill-defined for imperfect implicit functions. Our self-supervised loss functions do not rely on the ill-defined vertices produced by QEF minimization and align produced meshes to the input SDF grids, resulting in more plausible surface predictions. We believe this difference to be the key reason behind SDC's efficacy as a regularizer for training DINs. We also show three qualitative results in Figure 5. In some examples, we also observe topological differences although our regularization does not explicitly penalize it. Since the loss function which we minimize is highly non-convex, we believe that our regularization has aided in discovering more "plausible" latent space which could lead to better quality of implicit surfaces. We report additional qualitative and quantitative results in Suppl demonstrating the regularizer's generalization capabilities with other DINs. + +# 5.4. Single view reconstruction + +In this section, we consider the task of reconstructing surface meshes from images. We use 4 categories from the ShapeNet dataset [9], namely planes, chairs, rifles, and tables. We train and evaluate our approach and baseline (using the official codebase) on the same training and evalua + +![](images/2826841f986fab930d15f8c682bbdae6eaeabf9d4a3b8adc7e664a0b82dab2c7.jpg) +Figure 5. Qualitative examples comparing surface reconstruction from the ShapeNet [9] dataset. The columns compare different methods, whose training objectives differ. Our mesh-based regularization produces reasonable results compared to baselines. + +
MethodCD (↓)NC (↑)3D IoU (↑)LSD-A (↓)
CSDF [18]3.078.183.93.1
+ Eikonal [23]3.278.083.03.2
+ NDC [13]2.978.484.82.9
+ SDC2.678.986.02.7
+ +tion split for a fair comparison. We compare against three baselines, namely, MeshSDF [60], DISN [75], and a variant of our approach referred to as WoBW. More specifically, WoBW meshes the implicit function produced by pretrained SV-DIN using SDC, without the joint fine-tuning. For MeshSDF [60] and DISN [75] we use the official codebase for re-training and evaluation. + +Quantitative results averaged across 4 object categories are summarized in Table 3. Notably, our SDC outperforms MeshSDF [60] without additional test-time optimization fitting a mesh to the rendering. In addition, our end-to-end model consistently outperforms the WoBW baseline, implying the efficacy of our mesh-based regularization and end-to-end training. We also show three qualitative examples in Figure 6. Our SDC demonstrates an improved ability to reconstruct sharp features in the predicted meshes, resulting in surfaces more faithful to the ground truth. In the third row, we highlight an example showing significant improvements in the quality of predicted geometry. It shows that our method is capable of better thin surface reconstructions for the same inputs. + +# 6. Conclusion and future work + +We introduced SDC, a Self-Supervised training approach for Neural Dual Contouring. Differently from previous + +![](images/925fa32e0d5f30af812b3065663ced19b29e69726d0fc41ff2b3ed6fc2c03963.jpg) +Figure 6. Qualitative comparison of Single View Reconstruction (SVR) accuracy among three objects from ShapeNet [9] dataset. Input denotes the input view and GT denotes ground-truth mesh. SDC produces more sharper and plausible surfaces in comparison to baselines. + +Table 2. Comparison of surface reconstruction accuracy across object categories from the ShapeNet dataset [9]. Methods are trained with different regularization (see text) while evaluated alike. + +
MethodCD (↓)NC (↑)SSIM (↑)LSD-A (↓)
DISN [75]13.070.282.010.5
MeshSDF [60]11.371.484.19.1
WoBW9.972.086.48.1
Ours9.172.888.57.7
+ +Table 3. Quantitative results for single-view reconstruction across different object categories from ShapeNet [9] dataset. + +work, we do not fit the generated mesh to an axiomatically produced proxy mesh but instead use geometrically motivated self-supervised losses which only depends on the input SDF. SDC shows consistent improvements over baselines when applied to end-to-end surface reconstruction and meshing tasks. In addition, our work forges a link between signed distance fields and meshes and shows how the former can be regularized by guiding the network to produce distance fields better corresponding to resulting meshes. Our meshing and regularization are applied on a regular grid, which limits the resolution of a reconstructed mesh. It would be interesting to explore adaptive grids within the learning framework. Finally, another related and interesting scope for future work is the exploration of strict manifoldness conditions [62] and theoretical guarantees of self-intersection free surfaces. + +Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, and gifts from Ansys and Adobe Research. + +# References + +[1] Nina Amenta and Marshall Bern. Surface reconstruction by voronoi filtering. Discrete & Computational Geometry, 22 (4):481-504, 1999. 2 +[2] Nina Amenta, Marshall Bern, and Manolis Kamvysselis. A new voronoi-based surface reconstruction algorithm. Proc. SIGGRAPH, pages 415-421, 1998. 2 +[3] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. Computational Geometry, 19(2-3):127-153, 2001. 2 +[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 3 +[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Computer Graphics Forum, 36(1): 301-329, 2017. 2 +[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2 +[7] Jean-Daniel Boissonnat. Geometric structures for three-dimensional shape representation. ACM Transactions on Graphics (TOG), 3(4):266-286, 1984. 2 +[8] Jean-Daniel Boissonnat and Steve Oudot. Provably good sampling and meshing of surfaces. Graphical Models, 67 (5):405-451, 2005. 2 +[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 7, 8 +[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3 +[11] Zhiqin Chen and Hao Zhang. Neural marching cubes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 1, 2, 3, 4, 5, 6, 7 +[12] Zhiqin Chen, Andrea Tagliausacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3 +[13] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. arXiv preprint arXiv:2202.01999, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2 +[15] Angela Dai and Matthias Nießner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5574-5583, 2019. 3 + +[16] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 31-44, 2020. 3 +[17] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE TRANSACTIONS on Information and Systems, 74(1):214-224, 1991. 2 +[18] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevata, and Leonidas J. Guibas. Curriculum deepsdf. CoRR, abs/2003.08593, 2020. 5, 7, 8 +[19] Herbert Edelsbrunner and Ernst P Mücke. Three-dimensional alpha shapes. ACM Transactions on Graphics (TOG), 13(1):43-72, 1994. 2 +[20] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 3 +[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. arXiv preprint arXiv:1904.06447, 2019. 3 +[22] S.F.F. Gibson. Using distance maps for accurate surface representation in sampled volumes. In IEEE Symposium on Volume Visualization (Cat. No.989EX300), pages 23-30, 1998. +[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 5, 7, 8 +[24] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3 +[25] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In European Conference on Computer Vision, 2022. 3 +[26] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph., 39(4), 2020. 2 +[27] John C. Hart. Sphere tracing: a geometric method for the antialiased ray tracing of implicit surfaces. The Visual Computer, 12(10):527-545, 1996. 3 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 5 +[29] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. ACM SIGGRAPH Computer Graphics, 26 (2), 1992. 2 +[30] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 1, 2, 3, 4, 5, 6, 7 + +[31] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3 +[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2 +[33] Alireza Khatamian and Hamid R Arabnia. Survey on 3d surface reconstruction. Journal of Information Processing Systems, 12(3), 2016. 2 +[34] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 1 +[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[36] Ravikrishna Kolluri, Jonathan Richard Shewchuk, and James F O'Brien. Spectral surface reconstruction from noisy point clouds. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 11-21. ACM, 2004. 2 +[37] Thomas Lewiner, Hélio Lopes, Antonio Wilson Vieira, and Geovan Tavares. Efficient implementation of marching cubes' cases with topological guarantees. Journal of Graphics Tools, 8(2):1-15, 2003. 1, 6, 7 +[38] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3 +[39] Chen-Hsuan Lin, Oliver Wang, Bryan C Russell, Eli Shechtman, Vladimir G Kim, Matthew Fisher, and Simon Lucey. Photometric mesh optimization for video-aligned 3d object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 969-978, 2019. 3 +[40] David B. Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. arXiv preprint arXiv:0000.00000, 2021.3 +[41] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, 2021. 3, 5 +[42] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1886-1895, 2018. 3 +[43] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization, 2022. 3 +[44] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In + +European Conference on Computer Vision, pages 68-84. Springer, 2020. 3 +[45] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 2 +[46] N. Maruani, R. Klokov, M. Ovsjanikov, P. Alliez, and M. Desbrun. Voromesh: Learning watertight surface meshes with voronoi diagrams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[47] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. arXiv preprint arXiv:2204.07159, 2022. 3 +[48] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3 +[49] Mark Meyer, Mathieu Desbrun, Peter Schröder, and Alan H Barr. Discrete differential-geometry operators for triangulated 2-manifolds. In Visualization and mathematics III, pages 35-57. Springer, 2003. 1 +[50] Niloy J. Mitra, Natasha Gelfand, Helmut Pottmann, and Leonidas Guibas. Registration of point cloud data from a geometric optimization perspective. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH Symposium on Geometry Processing, page 22-31, New York, NY, USA, 2004. Association for Computing Machinery. 4 +[51] Patrick Mullen, Fernando De Goes, Mathieu Desbrun, David Cohen-Steiner, and Pierre Alliez. Signing the unsigned: Robust surface reconstruction from raw pointsets. Computer Graphics Forum, 29(5):1733-1741, 2010. 2 +[52] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 2 +[53] G.M. Nielson. Dual marching cubes. In IEEE Visualization 2004, pages 489-496, 2004. 5 +[54] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 9964-9973, 2019. 3 +[55] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2, 3, 5, 7 +[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 3 +[57] Helmut Pottmann and Michael Hofer. Geometry of the squared distance function to curves and surfaces. In Mathematics and Visualization, pages 221-242. Springer Berlin Heidelberg, 2003. 4 +[58] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction + +with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 3, 5 +[59] Marie-Julie Rakotaosaona, Paul Guerrero, Noam Algerman, Niloy J Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22-31, 2021. 3, 5 +[60] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2, 3, 5, 8 +[61] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 3 +[62] Scott Schaefer, Tao Ju, and Joe Warren. Manifold dual contouring. IEEE Transactions on Visualization and Computer Graphics, 13(3):610-619, 2007. 1, 8 +[63] Nicholas Sharp and Maks Ovsjanikov. Pointtrinet: Learned triangulation of 3d point sets. In European Conference on Computer Vision, pages 762-778. Springer, 2020. 1, 2, 3, 5 +[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3 +[65] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 6, 7 +[66] Jonathan Shewchuk, Tamal K Dey, and Siu-Wing Cheng. Delaunay mesh generation. Chapman and Hall/CRC, 2016. 2 +[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3 +[68] G D Smith. Numerical solution of partial differential equations. Clarendon Press, Oxford, England, 3 edition, 1985. 4 +[69] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. 2021. 6, 7 +[70] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3 +[71] Graham M Treece, Richard W Prager, and Andrew H Gee. Regularised marching tetrahedra: improved iso-surface extraction. Computers & Graphics, 23(4):583-598, 1999. 2 + +[72] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-67, 2018. 3 +[73] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10130-10139, 2019. 3 +[74] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 41(2):641-676, 2022. 3 +[75] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. Curran Associates, Inc., 2019. 8 +[76] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 3 +[77] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10, 000 3d-printing models. CoRR, abs/1605.04797, 2016. 6, 7 \ No newline at end of file diff --git a/2024/Self-Supervised Dual Contouring/images.zip b/2024/Self-Supervised Dual Contouring/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab9dd172d0f472ae0fb4ba59a212c7da0bdefc6a --- /dev/null +++ b/2024/Self-Supervised Dual Contouring/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3909814ba171875ed4f9336fe823d102865ed9f25b3452af0da08160eab1c40d +size 295880 diff --git a/2024/Self-Supervised Dual Contouring/layout.json b/2024/Self-Supervised Dual Contouring/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..decf2f0a43d73a10e4b6180d82e5c991303bb307 --- /dev/null +++ b/2024/Self-Supervised Dual Contouring/layout.json @@ -0,0 +1,9651 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 193, + 103, + 400, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 103, + 400, + 121 + ], + "spans": [ + { + "bbox": [ + 193, + 103, + 400, + 121 + ], + "type": "text", + "content": "Self-Supervised Dual Contouring" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 144, + 263, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 144, + 263, + 156 + ], + "spans": [ + { + "bbox": [ + 148, + 144, + 263, + 156 + ], + "type": "text", + "content": "Ramana Sundararaman" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 158, + 270, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 158, + 270, + 171 + ], + "spans": [ + { + "bbox": [ + 143, + 158, + 270, + 171 + ], + "type": "text", + "content": "LIX, Ecole Polytechnique" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 174, + 293, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 174, + 293, + 185 + ], + "spans": [ + { + "bbox": [ + 119, + 174, + 293, + 185 + ], + "type": "text", + "content": "sundararman@lix.polytechnique.fr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 363, + 144, + 438, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 144, + 438, + 156 + ], + "spans": [ + { + "bbox": [ + 363, + 144, + 438, + 156 + ], + "type": "text", + "content": "Roman Klokov" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 337, + 158, + 463, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 158, + 463, + 172 + ], + "spans": [ + { + "bbox": [ + 337, + 158, + 463, + 172 + ], + "type": "text", + "content": "LIX, Ecole Polytechnique" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 327, + 174, + 474, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 174, + 474, + 185 + ], + "spans": [ + { + "bbox": [ + 327, + 174, + 474, + 185 + ], + "type": "text", + "content": "klokov@lix.polytechnique.fr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 235, + 191, + 357, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 191, + 357, + 220 + ], + "spans": [ + { + "bbox": [ + 235, + 191, + 357, + 220 + ], + "type": "text", + "content": "Maks Ovsjanikov \nLIX, Ecole Polytechnique" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 227, + 221, + 364, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 221, + 364, + 233 + ], + "spans": [ + { + "bbox": [ + 227, + 221, + 364, + 233 + ], + "type": "text", + "content": "maks@lix.polytechnique.fr" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 260, + 191, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 260, + 191, + 273 + ], + "spans": [ + { + "bbox": [ + 143, + 260, + 191, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 286, + 289, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 289, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 289, + 610 + ], + "type": "text", + "content": "Learning-based isosurface extraction methods have recently emerged as a robust and efficient alternative to axiomatic techniques. However, the vast majority of such approaches rely on supervised training with axiomatically computed ground truths, thus potentially inheriting biases and data artefacts of the corresponding axiomatic methods. Steering away from such dependencies, we propose a self-supervised training scheme to the Neural Dual Contouring meshing framework, resulting in our method: Self-Supervised Dual Contouring (SDC). Instead of optimizing predicted mesh vertices with supervised training, we use two novel self-supervised loss functions that encourage the consistency between distances to the generated mesh up to the first order. Meshes reconstructed by SDC surpass existing data-driven methods in capturing intricate details while being more robust to possible irregularities in the input. Furthermore, we use the same self-supervised training objective linking inferred mesh and input SDF, to regularize the training process of Deep Implicit Networks (DINs). We demonstrate that the resulting DINs produce higher-quality implicit functions, ultimately leading to more accurate and detail-preserving surfaces compared to prior baselines for different input modalities. Finally, we demonstrate that our self-supervised losses improve meshing performance in the single-view reconstruction task by enabling joint training of predicted SDF and resulting output mesh. We open-source our code at https://github.com/Sentient07/SDC." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 633, + 128, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 128, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 128, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Surface mesh extraction from implicit functions, often referred to as isosurfacing [30, 37, 45, 62], is a fundamental problem in computer graphics and geometry processing as the quality of reconstructed mesh impacts algorithms used in numerous downstream tasks [49]. Prominent primal ax" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 264, + 537, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 264, + 537, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 264, + 537, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 537, + 376 + ], + "type": "image", + "image_path": "e44ea2a865b510402cbcf2afe5a97db9dea85510f5c1b72ea4ccb07822bd9042.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 387, + 546, + 444 + ], + "lines": [ + { + "bbox": [ + 305, + 387, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 546, + 444 + ], + "type": "text", + "content": "Figure 1. We propose SDC, a data-driven self-supervised isosurfacing method that reconstructs better feature-preserving meshes from learnt implicit functions compared to the baseline NDC [13]. Our framework is extendable as a regularizer for training Deep Implicit Networks (DINs) for improved reconstruction." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 459, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 519 + ], + "type": "text", + "content": "iomatic methods such as Marching Cubes [37, 45], are computationally efficient but fail to reconstruct sharp features. Dual methods [22, 30, 34, 62] achieve sharper reconstructions using Quadratic Error Function (QEF) but their performance heavily relies on the data fidelity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 520, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 546, + 652 + ], + "type": "text", + "content": "Following the ubiquitous success of neural methods in 2D and 3D applications, recent learning-based meshing approaches [11, 13] propose to predict mesh vertices and connectivity from various inputs in a feed-forward manner. When trained on extensive, augmented mesh collections, these methods demonstrate enhanced generalization and yield higher-quality reconstructions. However, these approaches rely either directly on the ground truth meshes [63] or on some intermediate pre-computed proxy meshes, like complex tessellation templates [11] and outputs of axiomatic approaches [13]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Neural Dual Contouring (NDC) [13], the most relevant method to our work, is a promising recent approach that enhances standard Dual Contouring (DC) [30, 62] with a data-driven pipeline. It aims to replicate DC-produced meshes, resulting in a framework capable of generating high-quality," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4681" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "feature-preserving meshes. However, since vertex estimation of DC follows QEF minimization, it requires precise SDF values at grid points and normals at edge intersections for training. Due to QEF minimization being ill-defined in particular scenarios [30], and susceptible to irregularities in the input SDF, NDC may exhibit a bias towards poor triangulation quality for challenging inputs and intricate details. Furthermore, its reliance on DC meshes for training restricts its use in an end-to-end scenario where both the implicit function and the explicit mesh can be optimized jointly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 531 + ], + "type": "text", + "content": "To overcome these limitations, we propose Self-Supervised Dual Contouring (SDC), a meshing framework that extends Neural Dual Contouring with two novel geometrically motivated self-supervised losses to learn the mesh vertex positions from SDF inputs while relying on sign changes to extract connectivity. The first self-supervised objective is a distance-based loss which minimizes the difference between the input implicit function values at points close to the surface and distances from the same points to the reconstructed mesh. This ensures that the predicted reconstructed mesh agrees with the input SDF. Secondly, we align the normals corresponding to reconstructed faces to those estimated from the input grid of SDFs. Since our training objective relies only on the input SDF and is agnostic to any explicit representation of the shape, e.g., based either on surface samples [26, 63] or on mesh vertices [11, 13], we refer to our approach as Self-Supervised. Together these two terms produce vertices that best explain the input grid of signed distance function values up to the first order like dual methods, while avoiding shortcomings of the quadratic error function (QEF) minimization [13, 30]. In particular, unlike DC, our method produces vertices that lie within each voxel-cell, and optimization of our self-supervised objectives leads to overall better results for input data with imperfections. As a result, we demonstrate better performance than supervised baselines in obtaining sharper meshes with negligible self-intersections and better generalization to unseen input data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "In addition, leveraging the connection between reconstructed meshes and input SDF grids, we propose a mesh-based regularization for training Deep Implicit Networks (DINs) - MLPs which represent zero-level sets of a shape. In a standard setup, training of DINs is completely decoupled from meshing, which is only performed a posteriori. Thus, existing DINs might produce implicit functions that do not correspond to a distance function arising from an extracted mesh. This motivates us to use SDC as a regularization to train DINs. In particular, we enforce the distance function produced by DINs during training to remain as close as possible to the distance function arising from the mesh produced by SDC. This forms the converse of our self-supervised meshing losses, but now geared to produce more coherent implicit functions. We demonstrate that this" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "regularization can be directly adopted into existing training paradigms for Deep Implicit Networks and it that leads to more plausible reconstructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 109, + 546, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 546, + 263 + ], + "type": "text", + "content": "Finally, since SDC training does not require ground truth meshes, it can be seamlessly integrated within end-to-end training paradigms. To highlight this, we consider the task of mesh reconstruction from images where we first construct an implicit representation given an image using an existing approach [55, 60] and then reconstruct a mesh using SDC. The resulting model is trained end-to-end by jointly minimizing the standard surface reconstruction loss, our self-supervised loss and the proposed regularization for Deep Implicit Networks. The resulting approach demonstrates compelling surface reconstruction efficacy across multiple object categories from the ShapeNet dataset [9]. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 265, + 545, + 396 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 265, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 265, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 545, + 323 + ], + "type": "text", + "content": "- We extend Neural Dual Contouring with two novel self-supervised loss functions ensuring consistency between input SDF and distance to output meshes, which is geometrically better motivated than the standard QEF minimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 324, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 545, + 361 + ], + "type": "text", + "content": "- We introduce a novel regularizer for training Deep Implicit Networks (DINs), by penalizing SDFs that do not correspond to underlying meshes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "text", + "content": "- We show that SDC generalizes better than supervised meshing methods, and further demonstrate its utility in feature-preserving mesh reconstruction from images." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 408, + 390, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 408, + 390, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 390, + 421 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 429, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 489 + ], + "type": "text", + "content": "Surface reconstruction and meshing tasks have been extensively studied within computer graphics and vision. Below, we review the methods that are most closely related to ours and refer interested readers to several surveys [5, 33, 66] for a more in-depth discussion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 546, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 546, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 546, + 723 + ], + "type": "text", + "content": "Axiomatic contouring techniques. The process of isosurface extraction from volumetric data is referred to as contouring. The most widely-used pipeline for surface reconstruction consists of two major steps. First, volumetric data is computed using a signed distance function [14, 29, 32, 51]. In the second step, a mesh is extracted from this representation, using Marching Cubes [45], Dual Contouring [30] or related approaches [17, 52, 71]. This pipeline can produce very high quality meshes, but is sensitive to the quality of input data and is not differentiable and thus does not easily fit within modern learning-based pipelines [38]. Another line of approaches is based on Delaunay triangulations, [7, 8, 36], as well as closely-related constructions [2, 3, 6, 19] which often come with strong theoretical guarantees and are typically geared towards preserving some input point set [1, 6]. Such methods, however, provide little control over the output triangulation. Moreover, classical approaches are typically not differentiable and thus cannot be easily integrated into learning pipelines." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4682" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "Data-driven mesh reconstruction. To address differentiability issues of classical mesh reconstruction approaches, a number of data-driven techniques have recently been proposed. This includes differentiable variants of contouring methods, such as the Marching Cubes [11, 38], dual Marching Cubes [65], Marching Tetrahedra [64], and Dual Contouring [13]. Our approach can be categorized into this family of differentiable contouring techniques. In addition, surface reconstruction has also been tackled using template-based techniques which fit a template mesh to the input [31, 39, 42], or deform an initial mesh while potentially updating its connectivity [20, 54, 72]. Furthermore, other local approaches propose to fit parameterized surface patches to points [24, 73], or to decompose space into convex sets [12, 16, 46]. Finally, several learning-based methods mesh an input point set [15, 44, 59, 63]. These approaches strongly rely on the given input point cloud and can thus be sensitive to artifacts and noise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": "Neural fields. There has been a recent surge in methods for learning implicit functions, referred to as neural fields [74]. This includes pioneering methods for predicting grid occupancy and signed distance functions [10, 48, 55] as well as their many follow-up works [21, 61, 67, 70]. Often, training Deep Implicit Networks introduces inductive bias [58] thereby producing inexplicable surface behaviors. Multiple regularization techniques [4, 23, 40, 41, 43, 58, 67, 70] have been proposed to address that issue. They largely focus on either learning high-frequency signals [40, 70] or designing loss functions based to guide the gradient level-set [23, 41, 58]. While the latter category of methods has convergence guarantees [41], the former helps in producing detail preserving reconstruction. Our work, on the other hand, provides a novel regularization by utilizing a meshing framework to ensure that the predicted signed distance field arises from a mesh. Similar to recent works [25, 47, 56, 60, 64, 65, 76] which optimize the mesh by controlling the underlying implicit function, our approach also enables back-propagation from the mesh to the input SDF. However, our approach is non-iterative and produces a mesh in one feed-forward pass." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 569, + 126, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 126, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 126, + 582 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "text", + "content": "We first begin by providing the notations used across the paper in Section 3.1 and review the Dual Contouring and Neural Dual Contouring algorithms in Section 3.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 115, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 115, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 115, + 646 + ], + "type": "text", + "content": "3.1. Notations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "We let " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^3 \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " be an implicit function that determines the signed distance of a query point " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " from an underlying surface " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": ". We let " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " denote the implicit function produced by a neural network parameterized by " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " to denote an " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "l \\times m \\times n" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " regular grid in the Euclidean space," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "g \\in \\mathcal{G}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " to be a node in the grid and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "e_{ij} \\in \\mathcal{G}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " to be an edge between adjacent nodes " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "g_i, g_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " respectively. We denote " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " as the discretization of " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " i.e. SDF evaluated at every grid node. Finally, we let " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " denote the latent vector associated with a shape computed using an encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 138, + 405, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 138, + 405, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 138, + 405, + 151 + ], + "type": "text", + "content": "3.2. Dual contouring" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": "The Dual Contouring (DC) [30] framework is designed to produce quadrilateral meshes from input signed distance functions discretized on a grid " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": ". It can be decomposed into two main steps: (1) construction of mesh faces " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": "; (2) computation of mesh vertices " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": ". To construct mesh faces, DC considers adjacent nodes " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "g_i, g_j \\in \\mathcal{G}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " such that their signed distance are of opposite signs, " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "f(g_i) \\neq f(g_j)" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": ". Then, for those pairs, a unique quadrilateral face " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "q_k" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " is constructed such that " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "q_k" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " crosses the edge " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": ". Vertices of this quadrilateral " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "q_k" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " reside in adjacent grid cells. The faces of the resulting mesh are obtained as a union of all quadrilateral faces " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\bigcup_{k \\in I} q_k" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " is a set of indices of grid edges connecting grid vertices of different signs. Final mesh vertex positions within each cell are obtained by minimizing the Quadratic Error Function (QEF). More precisely, given " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "p_e" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " to be the intersection position between edge " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " and the face " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " (along " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "n_e" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " to be the gradients of the SDF at positions " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " (or surface normals), vertices are determined for each grid cell " + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 156, + 545, + 384 + ], + "type": "text", + "content": " by solving:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 389, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 389, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 356, + 389, + 545, + 415 + ], + "type": "interline_equation", + "content": "v _ {c} = \\underset {x} {\\arg \\min } \\sum_ {e \\in I _ {c}} \\left[ n _ {e} * \\left(x - p _ {e}\\right) \\right], \\tag {1}", + "image_path": "93310dbb5d3a31bb07f2370bdc5704d4c5b8be589a84cc377f04957dd7740c83.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "text", + "content": " is a set of indices of intersected grid edges corresponding to cell " + }, + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "text", + "content": ". Recently introduced Neural Dual Contouring [13] attempts to emulate Dual Contouring via a neural network in order to reduce the inference time while achieving optimal reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 489, + 361, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 361, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 361, + 502 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 569 + ], + "type": "text", + "content": "We describe our Self-supervised Dual Contouring (SDC) loss functions in Section 4.1, discuss how to use them for regularization while training deep implicit networks in Section 4.2, and apply them for mesh reconstruction from input images in Section 4.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 575, + 509, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 575, + 509, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 575, + 509, + 588 + ], + "type": "text", + "content": "4.1. Self-supervised dual contouring (SDC)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We introduce a Self-supervised training scheme for Dual Contouring, and refer to a model trained with it as to SDC. It excludes the dependence on vertex positions obtained via QEF minimization. SDC produces quad meshes from the signed distance function discretized on a grid " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " similarly to any Dual Contouring method. We triangulate quadrilaterals by joining the top-left and bottom-right vertex as a convention. " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " can either be measured [27] or predicted using a DIN [55] " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". SDC consists of two main components constructing faces and vertices. To construct faces," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4683" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": "we follow the Dual Contouring [30] algorithm to determine the connectivity based on signed distance value at grid nodes. A quadrilateral face is constructed in the cells incident on every edge " + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "e_{ij} \\in \\mathcal{G}" + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": " connecting grid nodes of different signs " + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{sign}(f(g_i)) \\neq \\mathrm{sign}(f(g_j))" + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": ". For vertex prediction, we use a 6-layered 3D Convolutional Neural Network " + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "h_\\phi : S_\\mathcal{G} \\to \\mathcal{P} \\in \\mathbb{R}^{(l-1) \\times (m-1) \\times (n-1) \\times 3}" + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": ", which takes the grid of signed distance values as input and predicts a single point per each grid cell as the output. Then, we apply a masking to select grid cells, that bear node SDF values of opposite signs. Please refer to Suppl. for more details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "content": "SDC is trained without explicit fitting to the ground truth meshes using two loss functions. Firstly, to ensure formal correspondence of produced surfaces to input SDF grids, we propose a distance-based loss " + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "content": ". This objective minimizes the differences between 1) the absolute distance values at the nodes of the discretization grid " + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "inline_equation", + "content": "f(g_{i})\\in S_{\\mathcal{G}}" + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "content": " which are provided as inputs and 2) the distance " + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^p (g_i,\\mathcal{M})" + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "content": " measured from the same point " + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "content": " to the predicted mesh " + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "content": ". For a smooth or gently undulating surface, this distance provides a reasonable approximation of the underlying geometry [50, 57]. However, in the presence of sharp features, the local geometry changes direction rapidly. Close query points residing on different sides of a sharp feature have similar distances to the mesh but may be projected to very different locations on the actual surface. To address this, we also introduce a normal consistency loss, which aims to align the normals estimated at the nodes of the discretization grid " + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "inline_equation", + "content": "f(g_{i})\\in S_{\\mathcal{G}}" + }, + { + "bbox": [ + 47, + 205, + 289, + 432 + ], + "type": "text", + "content": " and normals corresponding to the generated mesh face." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 434, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 289, + 555 + ], + "type": "text", + "content": "We refer to the losses proposed above as self-supervised since the training signal for our pipeline comes from the input alone. We do not use any explicit surface discretization for training and our losses are solely based on the input implicit function values and the generated mesh. In contrast, we refer to methods that predict vertices to fit a specific mesh structure as supervised meshing methods [11, 13]. We provide visual explanations of our loss functions in Figure 2. In summary, our combined self-supervised objective for the vertex prediction network is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 567, + 287, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 567, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 118, + 567, + 287, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {M e s h}} = \\mathcal {L} _ {D} + \\alpha_ {1} \\mathcal {L} _ {N}, \\tag {2}", + "image_path": "4bfd02153a64368704438b6e051ce187a1e85c04563814ed8e6d91d18c73f646.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 594, + 227, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 227, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 227, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 594, + 227, + 605 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 47, + 594, + 227, + 605 + ], + "type": "text", + "content": " is a hyperparameter to weigh " + }, + { + "bbox": [ + 47, + 594, + 227, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_N" + }, + { + "bbox": [ + 47, + 594, + 227, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "content": "Distance loss. The objective of our distance loss is to minimize the discrepancy between surfaces produced by SDC and zero-level sets defined by the input SDFs. In order to achieve this, given grid nodes " + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "g_{i} \\in \\mathcal{G}" + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "content": ", we measure the distance between the grid nodes and the mesh produced by SDC and penalize its deviation from the input absolute distance values " + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "f(g_{i})" + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "text", + "content": ". In summary, our distance loss is defined" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 75, + 425, + 182 + ], + "blocks": [ + { + "bbox": [ + 309, + 75, + 425, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 75, + 425, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 75, + 425, + 182 + ], + "type": "image", + "image_path": "e34092315d8f2d81533f2c84bc32d0ad6ef55fd34585f1022ad69c68b0c71db6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 186, + 373, + 196 + ], + "lines": [ + { + "bbox": [ + 362, + 186, + 373, + 196 + ], + "spans": [ + { + "bbox": [ + 362, + 186, + 373, + 196 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 111, + 539, + 185 + ], + "blocks": [ + { + "bbox": [ + 427, + 111, + 539, + 185 + ], + "lines": [ + { + "bbox": [ + 427, + 111, + 539, + 185 + ], + "spans": [ + { + "bbox": [ + 427, + 111, + 539, + 185 + ], + "type": "image", + "image_path": "5fdcefafe6f6a567cba586466686ddbfc38a6ce383938b459c3487220e67a196.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 186, + 492, + 196 + ], + "lines": [ + { + "bbox": [ + 481, + 186, + 492, + 196 + ], + "spans": [ + { + "bbox": [ + 481, + 186, + 492, + 196 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 205, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 283 + ], + "type": "text", + "content": "Figure 2. Illustration of our self-supervised losses. (a) Measurements of " + }, + { + "bbox": [ + 304, + 205, + 547, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^p" + }, + { + "bbox": [ + 304, + 205, + 547, + 283 + ], + "type": "text", + "content": " between grid nodes and the closest point on a mesh face. For visualization purposes, we only show this estimation at one node. (b) Normal consistency loss is measured as the discrepancy between interpolated normals (orange) and the face normal (green). Red nodes denote negative SDF grid and green denotes positive. The constructed quad is shown in transparent purple." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 304, + 320, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 320, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 320, + 313 + ], + "type": "text", + "content": "as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 312, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 312, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 341, + 312, + 545, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = \\sum_ {g _ {i} \\in \\mathcal {G}} \\left\\| \\operatorname {a b s} \\left(f \\left(g _ {i}\\right)\\right) - \\mathbf {d} ^ {p} \\left(g _ {i}, \\mathcal {M}\\right) \\right\\| _ {2} ^ {2}, \\tag {3}", + "image_path": "b57404596a33b415aa47a8ab615fd4425a4755da020c0356ec0687ada3c405ef.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^p" + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "content": " is the point-to-face distance, computed between " + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "content": " and its projection on " + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "content": ". For computational efficiency, we compute " + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "content": " only at grid nodes that are close to the surface. Please refer to Suppl for details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": "Normal consistency loss. We introduce a novel normal consistency loss, " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_N" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " which aims to align the level-sets up to the first order. In particular, given the grid of scalar SDFs, we compute normals as gradients " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "n(g_i) \\coloneqq f'(g_i) / ||f'(g_i)||_2" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " in grid nodes " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " using Five-point stencil, a commonly used finite difference approach for solving partial differential equations on regular grids [68]. More details on gradient estimation is provided in the Suppl. Given an edge " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "e_{ij} \\in \\mathcal{G}" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " connecting nodes " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "g_i, g_j" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " of the opposite signs, we first estimate a point " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "p \\in e_{ij}" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " along the edge where " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "f(p) = 0" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": ". The point " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " can be represented in terms of " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "g_j" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "p = (1 - t)g_i + tg_j" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " is a parameter that determines the position of " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " along the edge " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": ". Solving for " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "p = 0" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": ", the parameter " + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 404, + 547, + 572 + ], + "type": "text", + "content": " can be expressed in closed form as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 384, + 579, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 579, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 384, + 579, + 545, + 605 + ], + "type": "interline_equation", + "content": "t = \\frac {f \\left(g _ {i}\\right)}{f \\left(g _ {i}\\right) - f \\left(g _ {j}\\right)}. \\tag {4}", + "image_path": "c52e7fe06e9f1c9c87e75720be4bde88d7079bb5952c8bca69dff1cea861ec1f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "text", + "content": "Similarly, we can estimate the normal at the point " + }, + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "text", + "content": " by linear interpolation. Using the previously computed parameter " + }, + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "text", + "content": ", the normal at " + }, + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 612, + 547, + 649 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 656, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 656, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 310, + 656, + 545, + 670 + ], + "type": "interline_equation", + "content": "n (p) = \\operatorname {s i g n} \\left(f \\left(g _ {i}\\right)\\right) (1 - t) n \\left(g _ {i}\\right) + \\operatorname {s i g n} \\left(f \\left(g _ {j}\\right)\\right) t n \\left(g _ {j}\\right). \\tag {5}", + "image_path": "0fe112d4c569bfc62706322be6ee6fffaedf6a291f3a7b12df94b29f76d68eba.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "This expression combines the known gradients at the neighboring nodes, weighted by the parameter " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": ", to estimate the gradient at a point " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": ". Since the inward normal is oriented" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4684" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "opposite to the outward normal, we multiply by the sign for consistency. As our mesh is constructed following Dual Contouring [30], we denote the quad-face constructed dual to the edge " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": " and its normal as " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "n(q_{ij})" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": ". With the aim of aligning these two normals, our normal consistency loss can be written as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 108, + 148, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 148, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 108, + 148, + 287, + 175 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N} = 1 - \\frac {n \\left(q _ {i j}\\right) \\cdot n (p)}{\\| n \\left(q _ {i j}\\right) \\| \\| n (p) \\|.} \\tag {6}", + "image_path": "3d2a5abdc9030f92558dedd817de97c3a84a33af618f5f8addc0b43fd2a21488.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "type": "text", + "content": "Training augmentation. At training time, we add synthetic noise which is zero-mean Gaussian noise with a standard deviation equivalent to a third grid edge length " + }, + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "type": "text", + "content": " to the input SDF. Since SDC constructs mesh faces based on Dual Contouring, a noise that induces a sign change to the initial SDF will also lead to a topological change. More precisely, augmentation will lead to inclusion of vertices (and faces) in cells whose nodes differ in sign. Similarly, faces are removed dual to cells whose signs agree after augmentation. More training details are provided in the Suppl." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 230, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 230, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 230, + 326 + ], + "type": "text", + "content": "4.2. Deep implicit network regularizer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 510 + ], + "type": "text", + "content": "Deep Implicit Networks (DINs), commonly represented by Multi-Layer Perceptrons (MLP), are trained to acquire an implicit representation of a shape and are primarily utilized for various inference-based surface reconstruction tasks. In a traditional setup, DINs are trained by direct supervision of the SDF values at query points sampled close to the surface. Such a training scheme, as also observed by previous works [23, 41, 58], is known to produce inexplicable behavior, such as ambiguous level sets. To address this, we propose a novel mesh-based training regularization for DINs to ensure that the Distance Field (DF) produced by DIN agrees well with the resulting extracted mesh. Our regularization minimizes the discrepancy between the predicted DF (absolute value of SDF) and the distance computed from the mesh corresponding to the SDF extracted using SDC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": "We assume that we are endowed with a learnable DIN " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\lambda, p): p \\in \\mathbb{R}^3 \\to s_p \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ", represented as an MLP, whose weights " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " encode the implicit surface of a shape conditioned by a latent vector " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " is a point in space such that " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\lambda, p) = s_p" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " is its signed distance from the zero-level set. Then, for a given " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "S_G" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " - a discretization of " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " on a regular grid and " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^*(S_G)" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " - a mesh that is produced following any iso-surfacing algorithm, it is not necessarily true that " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{abs}(f_{\\theta}(p, \\lambda)) = \\mathbf{d}^p(p, \\mathcal{M}^*(S_G))" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^p" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " is the point-to-face distance. While this disparity is viewed as an inherent drawback of primal iso-surfacing methods [53], for dual methods, this disparity can, in principle, be minimized due to their ability to preserve sharp details in the reconstructed meshes. This motivates us to use our SDC for iso-surface extraction and regularize the training of DINs by minimizing the aforementioned disparity. Assuming we have access to a dataset consisting of " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " shapes and a pre-trained" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "SDC with frozen weights, our regularization objective can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 102, + 545, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 102, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 102, + 545, + 137 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D R} = \\sum_ {j = 1} ^ {N} \\sum_ {g _ {i} \\in \\mathcal {G}} | | \\operatorname {a b s} \\left(f _ {\\theta} \\left(\\lambda_ {j}, g _ {i}\\right) - \\mathbf {d} ^ {p} \\left(g _ {i}, \\mathcal {M}\\right) \\right| | _ {2} ^ {2}. \\tag {7}", + "image_path": "6100b1a8c96b15f1f4ebf327b7c5a9d763320fb7520e1ad129b05fb7fc56ca9e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": "This regularization penalizes the discrepancy between the DF at grid nodes predicted by the DIN " + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "inline_equation", + "content": "f_{\\theta}(g_i)" + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": " and the distance " + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^p" + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": " computed between the grid nodes and the mesh " + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}})" + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": ". To obtain " + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": ", we first evaluate " + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": " over a discrete grid " + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}})" + }, + { + "bbox": [ + 305, + 144, + 545, + 286 + ], + "type": "text", + "content": ". Note that the above regularization is similar to the signed distance loss defined in Eq. (3), but with an important difference. Here we do not use ground-truth SDF values but instead, regularize the predicted signed distance. This regularization is applied alongside the main objective function to reconstruct the implicit surface commonly used to train DINs [18, 55] as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 545, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 545, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S D F}} = \\sum_ {j = 1} ^ {N} \\sum_ {i} ^ {K} \\left| \\left| \\left(f _ {\\theta} \\left(\\lambda_ {j}, x _ {i}\\right) - s _ {i} \\right| \\right| _ {2} ^ {2} + \\frac {1}{\\sigma^ {2}} \\left\\| \\lambda_ {j} \\right\\| _ {2} ^ {2}, \\right. \\tag {8}", + "image_path": "8798ad65a8d5c2317e8ea301167607c79e5bf38526b85ff8d64e9bd6119aea6f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "text", + "content": " is the number of points with annotated SDF values per shape " + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "text", + "content": " is the ground-truth SDF value, and " + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 335, + 545, + 370 + ], + "type": "text", + "content": " is a parameter used to promote compactness in latent space [55]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "text", + "content": "Combining the standard training loss for DINs [55] alongside our regularization weighted with a scalar " + }, + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "text", + "content": ", our learning objective is formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 415, + 545, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 415, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 367, + 415, + 545, + 428 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a i n}} = \\mathcal {L} _ {\\text {S D F}} + \\alpha_ {2} \\mathcal {L} _ {S D R}. \\tag {9}", + "image_path": "843c394e98f598e1aab3c44336d509d6f867b76f79f35f29caaf601e2119f529.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "type": "text", + "content": "At the inference time, we reconstruct surfaces from point clouds by first recovering the optimal latent vectors (See Eq. (4) in [18]). Then, we predict a grid of SDF values " + }, + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{G}}" + }, + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "type": "text", + "content": " and extract the iso-surface using our SDC." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 491, + 544, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 544, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 544, + 505 + ], + "type": "text", + "content": "4.3. Joint learning of implicit surface and meshing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Differently from existing data-driven meshing methods, SDC relies neither on explicit mesh data [59, 63] nor on axiomatically produced proxy meshes [11, 13]. This means that SDC can be used to optimize both SDF and reconstructed meshes jointly in an end-to-end manner. To demonstrate this we consider the task of predicting surface meshes from images. We first model the SDF representation of a shape given an image and then predict the mesh using SDC. In terms of the architectures, we follow MeshSDF [60]: we produce latent vectors " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " from input images with a ResNet-18 [28] encoder and learn an implicit shape representation of shapes using DeepSDF [55] conditioned on the aforementioned latent vectors. Instead of performing test-time optimization like MeshSDF, we predict SDF values on a regular grid " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathcal{G})" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and use SDC to reconstruct a mesh " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\mathrm{SDC}(S_{\\mathcal{G}})" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " by a simple forward pass of our SDC network. Initially, we train the networks (1) DeepSDF (with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4685" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "an image encoder) and (2) SDC independently, then jointly fine-tune them for 200 epochs. This fine-tuning uses an image as input and the mesh " + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": " from SDC as output, while minimizing the combined objective terms for surface reconstruction and meshing as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 143, + 287, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 143, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 91, + 143, + 287, + 156 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S V R}} = \\mathcal {L} _ {\\mathrm {S D F}} + \\alpha_ {3} \\mathcal {L} _ {S D R} + \\alpha_ {4} \\mathcal {L} _ {\\mathrm {D}}, \\tag {10}", + "image_path": "15c597399b463f7c4a194138fad64946f8b52cd8f8b920c3ce5d46dea24c4e3d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDF}}" + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDR}}" + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "text", + "content": " (defined in Eq. (8) and Eq. (7) respectively) are used to reconstruct and regularize the implicit representation, while " + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "text", + "content": ", defined in Eq. (3), is used to train SDC. Since the evaluated SDF might be noisy during training, we observed that using " + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_N" + }, + { + "bbox": [ + 46, + 165, + 287, + 226 + ], + "type": "text", + "content": " results in instability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 236, + 244, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 244, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 244, + 250 + ], + "type": "text", + "content": "5. Experiments, results and discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 256, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 293 + ], + "type": "text", + "content": "We show the effectiveness of SDC across three tasks, namely, implicit function meshing, DIN training regularization and joint SDF and mesh prediction from images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 308, + 288, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 583 + ], + "type": "text", + "content": "Evaluation Metrics. We use Chamfer Distance (CD), Normal Consistency (NC), Self-Intersection (SI), Edge Chamfer Distance (ECD), 3D-IoU, Precise Level-Set Discrepancy (LSD-P), and Approximate Level-Set Discrepancy (LSD-A) as metrics to compare all approaches. For measuring Level-Set discrepancy, we use two terms, Precise (LSD-P) and Approximate (LSD-A). The former is used in case where shapes are water-tight and analytical SDF can be measured. In such cases, we measure the discrepancy between the ground truth analytical SDF and the distance function measured from the generated mesh. LSD-A is used for non-watertight shapes and is measured by sampling points on the ground truth mesh and measuring its distance from the generated mesh's faces. Self-intersection is reported as the total number of intersecting triangles per mesh averaged across the test set. For the single-view reconstruction task, we also measure the Structural Similarity Index (SSIM) between the rendering of the reconstructed mesh and the ground truth mesh. We measure 3D-IoU for experiments where the generated mesh could topologically differ from the ground truth. CD, ECD, LSD are scaled by " + }, + { + "bbox": [ + 46, + 308, + 288, + 583 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 46, + 308, + 288, + 583 + ], + "type": "text", + "content": " while NC, SSIM 3D IoU are in " + }, + { + "bbox": [ + 46, + 308, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 46, + 308, + 288, + 583 + ], + "type": "text", + "content": " .We provide more details on all the metrics in the Suppl." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 591, + 244, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 244, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 244, + 605 + ], + "type": "text", + "content": "5.1. Meshing analytical implicit functions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "text", + "content": "We first consider the surface mesh reconstruction task given the ground truth SDFs of shapes discretized on a regular grid. For this task, we train SDC on 3000 shapes from the " + }, + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "text", + "content": " split of the ABC dataset [35] for 100 epochs. We scale each shape to fit a unit sphere and estimate the SDF on a regular grid of dimension " + }, + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "inline_equation", + "content": "64^3" + }, + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "text", + "content": " using SDFGen library1. We evaluate all methods on 300 shapes from the test set of" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 541, + 262 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 541, + 262 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 541, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 541, + 262 + ], + "type": "image", + "image_path": "4766a179210be62b82f858e9b73591dee6bf9bf098d0e7189fbad6d4ebd1ee33.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 276, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 276, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 276, + 547, + 331 + ], + "type": "text", + "content": "Figure 3. We show three qualitative mesh reconstruction examples from the ABC dataset (Row 1,2) and the Thingi10k dataset (Row 3). Self-intersecting faces are highlighted in red. QEF-based learning methods NDC and UNDC show significantly higher self-intersecting faces along sharp edges while ours does not." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 345, + 544, + 444 + ], + "blocks": [ + { + "bbox": [ + 306, + 345, + 544, + 444 + ], + "lines": [ + { + "bbox": [ + 306, + 345, + 544, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 544, + 444 + ], + "type": "image", + "image_path": "e11d341c366f60d74b4630ac11f031b90070d0ff23e9ab98f98c48f5d284b781.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 451, + 546, + 485 + ], + "lines": [ + { + "bbox": [ + 305, + 451, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 546, + 485 + ], + "type": "text", + "content": "Figure 4. Qualitative examples of meshing SDF produced by NGLOD [69]. MC, NMC, and NDC produce noisy meshes while SDC produces a smoother yet feature-preserving reconstruction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 520, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 663 + ], + "type": "text", + "content": "the ABC dataset. Additionally, we evaluate on 300 shapes from the Thingi10K [77] dataset to demonstrate generalization to unseen data domains. We compare SDC against four baselines including two axiomatic baselines, namely the improved Marching Cubes (MC) [37] and the standard Dual Contouring (DC) [30] where positions of each vertex are estimated by solving the Quadratic Error Function (QEF), and three of the recent Neural Meshing techniques, Neural Marching Cubes (NMC) [11], Neural Dual Contouring [13] and FlexiCubes [65] respectively. For fairness, we compare to variants of comparable network capacity and for FlexiCubes, we compare without test-time optimization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "We summarize our quantitative results in Table 1. Our self-supervised method SDC shows better performance and generalization to unseen data domains compared to the data-driven baselines across all metrics. While Marching" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 266, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 266, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 266, + 712 + ], + "type": "text", + "content": "DatasetInput SDFMethodCD (↓)NC (↑)#SI (↓)ECD (↓)LSD-P (↓)ABCAnalyticalMC [37]4.792.104.816.1NMC [11]3.794.734.53.612.1DC [30]3.593.596.33.411.6NDC [13]3.694.343.13.511.9FlexiCubes [65]4.592.404.815.8Ours3.394.99.73.211.3Thingi10KAnalyticalMC [37]5.161.4013.315.3NMC [11]4.166.032.411.912.2DC [30]4.063.8105.611.511.8NDC [13]4.064.970.411.312.0FlexiCubes [65]5.255.3013.615.4Ours3.668.015.710.911.6Thingi10KPredicted [69]MC [37]5.856.7014.216.5NMC [11]4.763.230.112.612.9DC [30]6.254.0230.414.815.0NDC [13]4.762.842.612.813.1Ours4.265.715.912.212.4", + "image_path": "2bb78247572fa8d55c483e6f0a74fd721aca13ea900166d76aeccf8f5b431095.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "text", + "content": "Table 1. Quantitative mesh reconstruction results on the ABC and the Thingi10k datasets. The best scores are highlighted in bold and the second-best scores are in italics. The first two row-blocks correspond to input SDF which was computed analytically while in the last row-block, we mesh the SDF predicted by NGLOD [69]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 288, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 288, + 445 + ], + "type": "text", + "content": "Cubes [37] and FlexiCubes [65] do not produce any self-intersecting faces, the vertex placement is restricted for these methods, leading to inferior results with respect to other reconstruction metrics. In the qualitative examples visualized in Figure 3, we highlight an inherent issue with NDC and UNDC - they produce a considerable number of self-intersections along sharp edges. We reason this to be an undesirable trait that is inherited from the standard Dual Contouring as previously elaborated. SDC, on the other hand, bears a minimal number of self-intersections, as our training objective for mesh prediction is geometrically well-motivated. Additionally, we report quantitative and qualitative results for " + }, + { + "bbox": [ + 46, + 288, + 288, + 445 + ], + "type": "inline_equation", + "content": "128^{3}" + }, + { + "bbox": [ + 46, + 288, + 288, + 445 + ], + "type": "text", + "content": " grid resolution in the Suppl." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 454, + 241, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 241, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 241, + 467 + ], + "type": "text", + "content": "5.2. Meshing predicted implicit functions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 715 + ], + "type": "text", + "content": "Differently from the previous section, where the SDF is computed analytically, here, we use an open-source implementation (with suggested hyperparameters) of a recent deep implicit network NGLOD [69] to predict SDF values on a regular grid. We evaluate the meshing efficacy of SDC over a set of 300 shapes from the Thingi10k [77] dataset. To avoid possible biases, we chose these 300 shapes to be different from the ones used in the previous section. Our quantitative results are summarized in Table 1. Consistent with our observation from the previous section, SDC outperforms all baselines. We provide a qualitative example of meshing a noisy SDF in Figure 4. To generate the noisy SDF, we prematurely terminate NGLOD [69] fitting to the shape and evaluate different Neural Meshing methods. While all baselines produce a noisy reconstruction, SDC recovers a smoother surface while preserving sharp edges, thanks to our regularization and self-supervised losses. This suggests that our self-supervised training objectives and data augmentation are capable of generalizing to noisy imperfect predicted data without explicit training on this data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 506, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 506, + 85 + ], + "type": "text", + "content": "5.3. Regularizing implicit surface learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 90, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 353 + ], + "type": "text", + "content": "As illustrated in the previous section, SDC successfully produces sharp meshes from both analytical and parameterized implicit functions. We now employ SDC for regularizing DIN training, focusing on the task of surface reconstruction using the ShapeNet dataset [9]. In particular, we evaluate over 4 categories of objects, namely, cars, planes, tables, and cabinets with 1000 shapes per category as our training set, and set aside a separate set of 200 unseen shapes for evaluation. We compare our regularization against three possible baselines. Firstly, we train Curriculum DeepSDF (CSDF) [18] over each category separately. Secondly, we re-train the same network by enforcing the Eikonal constraint, i.e. enforcing a unit norm for SDF's gradient in a setup similar to Implicit Geometric Regularization [23]. Thirdly, we use the supervised Dual Contouring baseline NDC [13] as the mesh-based regularizer. More specifically, we replace SDC with NDC in computing the regularization introduced in Eq. (7). For each shape, we sample 400,000 points in the shape volume, aggressively near the surface following [55] to supervise the SDF prediction. For fairness, we use the same sampled points for all methods we compare." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 354, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 628 + ], + "type": "text", + "content": "We summarize our quantitative results in Table 2. We observe that regularizing the network using SDC produces a noticeable improvement in reconstruction across all metrics in comparison to baselines. Moreover, we observe a poorer reconstruction when using NDC [13] for regularization. We argue that this is due to their inability to handle noisy predicted SDF values as inputs as this approach learns to emulate solution to QEF (c.f. Eq. (1)), which is ill-defined for imperfect implicit functions. Our self-supervised loss functions do not rely on the ill-defined vertices produced by QEF minimization and align produced meshes to the input SDF grids, resulting in more plausible surface predictions. We believe this difference to be the key reason behind SDC's efficacy as a regularizer for training DINs. We also show three qualitative results in Figure 5. In some examples, we also observe topological differences although our regularization does not explicitly penalize it. Since the loss function which we minimize is highly non-convex, we believe that our regularization has aided in discovering more \"plausible\" latent space which could lead to better quality of implicit surfaces. We report additional qualitative and quantitative results in Suppl demonstrating the regularizer's generalization capabilities with other DINs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 635, + 452, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 452, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 452, + 647 + ], + "type": "text", + "content": "5.4. Single view reconstruction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "In this section, we consider the task of reconstructing surface meshes from images. We use 4 categories from the ShapeNet dataset [9], namely planes, chairs, rifles, and tables. We train and evaluate our approach and baseline (using the official codebase) on the same training and evalua" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4687" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 75, + 282, + 213 + ], + "blocks": [ + { + "bbox": [ + 50, + 75, + 282, + 213 + ], + "lines": [ + { + "bbox": [ + 50, + 75, + 282, + 213 + ], + "spans": [ + { + "bbox": [ + 50, + 75, + 282, + 213 + ], + "type": "image", + "image_path": "2826841f986fab930d15f8c682bbdae6eaeabf9d4a3b8adc7e664a0b82dab2c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 223, + 287, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 268 + ], + "type": "text", + "content": "Figure 5. Qualitative examples comparing surface reconstruction from the ShapeNet [9] dataset. The columns compare different methods, whose training objectives differ. Our mesh-based regularization produces reasonable results compared to baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 286, + 277, + 349 + ], + "blocks": [ + { + "bbox": [ + 56, + 286, + 277, + 349 + ], + "lines": [ + { + "bbox": [ + 56, + 286, + 277, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 277, + 349 + ], + "type": "table", + "html": "
MethodCD (↓)NC (↑)3D IoU (↑)LSD-A (↓)
CSDF [18]3.078.183.93.1
+ Eikonal [23]3.278.083.03.2
+ NDC [13]2.978.484.82.9
+ SDC2.678.986.02.7
", + "image_path": "972aa891be39f2e0fb639f899bdcbd4abdfbefe81e22a89b829b0a929e5478ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 406, + 286, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 286, + 490 + ], + "type": "text", + "content": "tion split for a fair comparison. We compare against three baselines, namely, MeshSDF [60], DISN [75], and a variant of our approach referred to as WoBW. More specifically, WoBW meshes the implicit function produced by pretrained SV-DIN using SDC, without the joint fine-tuning. For MeshSDF [60] and DISN [75] we use the official codebase for re-training and evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 491, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 658 + ], + "type": "text", + "content": "Quantitative results averaged across 4 object categories are summarized in Table 3. Notably, our SDC outperforms MeshSDF [60] without additional test-time optimization fitting a mesh to the rendering. In addition, our end-to-end model consistently outperforms the WoBW baseline, implying the efficacy of our mesh-based regularization and end-to-end training. We also show three qualitative examples in Figure 6. Our SDC demonstrates an improved ability to reconstruct sharp features in the predicted meshes, resulting in surfaces more faithful to the ground truth. In the third row, we highlight an example showing significant improvements in the quality of predicted geometry. It shows that our method is capable of better thin surface reconstructions for the same inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 669, + 206, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 206, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 206, + 681 + ], + "type": "text", + "content": "6. Conclusion and future work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We introduced SDC, a Self-Supervised training approach for Neural Dual Contouring. Differently from previous" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 74, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 307, + 74, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 307, + 74, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 74, + 545, + 220 + ], + "type": "image", + "image_path": "925fa32e0d5f30af812b3065663ced19b29e69726d0fc41ff2b3ed6fc2c03963.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 230, + 545, + 284 + ], + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 284 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of Single View Reconstruction (SVR) accuracy among three objects from ShapeNet [9] dataset. Input denotes the input view and GT denotes ground-truth mesh. SDC produces more sharper and plausible surfaces in comparison to baselines." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 304, + 315, + 525, + 378 + ], + "blocks": [ + { + "bbox": [ + 46, + 357, + 287, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 390 + ], + "type": "text", + "content": "Table 2. Comparison of surface reconstruction accuracy across object categories from the ShapeNet dataset [9]. Methods are trained with different regularization (see text) while evaluated alike." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 315, + 525, + 378 + ], + "lines": [ + { + "bbox": [ + 304, + 315, + 525, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 525, + 378 + ], + "type": "table", + "html": "
MethodCD (↓)NC (↑)SSIM (↑)LSD-A (↓)
DISN [75]13.070.282.010.5
MeshSDF [60]11.371.484.19.1
WoBW9.972.086.48.1
Ours9.172.888.57.7
", + "image_path": "227f50b64fe47686f2b9b40a622f287fa6401fce05e31ec856f9dca7d0007ddf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 386, + 545, + 408 + ], + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 408 + ], + "type": "text", + "content": "Table 3. Quantitative results for single-view reconstruction across different object categories from ShapeNet [9] dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "content": "work, we do not fit the generated mesh to an axiomatically produced proxy mesh but instead use geometrically motivated self-supervised losses which only depends on the input SDF. SDC shows consistent improvements over baselines when applied to end-to-end surface reconstruction and meshing tasks. In addition, our work forges a link between signed distance fields and meshes and shows how the former can be regularized by guiding the network to produce distance fields better corresponding to resulting meshes. Our meshing and regularization are applied on a regular grid, which limits the resolution of a reconstructed mesh. It would be interesting to explore adaptive grids within the learning framework. Finally, another related and interesting scope for future work is the exploration of strict manifoldness conditions [62] and theoretical guarantees of self-intersection free surfaces." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements Parts of this work were supported by the ERC Starting Grant 758800 (EXPROTEA), ERC Consolidator Grant 101087347 (VEGA), ANR AI Chair AIGRETTE, and gifts from Ansys and Adobe Research." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4688" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Nina Amenta and Marshall Bern. Surface reconstruction by voronoi filtering. Discrete & Computational Geometry, 22 (4):481-504, 1999. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "type": "text", + "content": "[2] Nina Amenta, Marshall Bern, and Manolis Kamvysselis. A new voronoi-based surface reconstruction algorithm. Proc. SIGGRAPH, pages 415-421, 1998. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "text", + "content": "[3] Nina Amenta, Sunghee Choi, and Ravi Krishna Kolluri. The power crust, unions of balls, and the medial axis transform. Computational Geometry, 19(2-3):127-153, 2001. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19323-19332, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Computer Graphics Forum, 36(1): 301-329, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "text", + "content": "[6] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "text", + "content": "[7] Jean-Daniel Boissonnat. Geometric structures for three-dimensional shape representation. ACM Transactions on Graphics (TOG), 3(4):266-286, 1984. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "text", + "content": "[8] Jean-Daniel Boissonnat and Steve Oudot. Provably good sampling and meshing of surfaces. Graphical Models, 67 (5):405-451, 2005. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 512 + ], + "type": "text", + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 545 + ], + "type": "text", + "content": "[11] Zhiqin Chen and Hao Zhang. Neural marching cubes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 591 + ], + "type": "text", + "content": "[12] Zhiqin Chen, Andrea Tagliausacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "text", + "content": "[13] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. arXiv preprint arXiv:2202.01999, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "text", + "content": "[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[15] Angela Dai and Matthias Nießner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5574-5583, 2019. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[16] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 31-44, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[17] Akio Doi and Akio Koide. An efficient method of triangulating equi-valued surfaces by using tetrahedral cells. IEICE TRANSACTIONS on Information and Systems, 74(1):214-224, 1991. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "text", + "content": "[18] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevata, and Leonidas J. Guibas. Curriculum deepsdf. CoRR, abs/2003.08593, 2020. 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 243 + ], + "type": "text", + "content": "[19] Herbert Edelsbrunner and Ernst P Mücke. Three-dimensional alpha shapes. ACM Transactions on Graphics (TOG), 13(1):43-72, 1994. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 244, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 289 + ], + "type": "text", + "content": "[20] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 290, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 290, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 545, + 334 + ], + "type": "text", + "content": "[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. arXiv preprint arXiv:1904.06447, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 335, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 545, + 379 + ], + "type": "text", + "content": "[22] S.F.F. Gibson. Using distance maps for accurate surface representation in sampled volumes. In IEEE Symposium on Volume Visualization (Cat. No.989EX300), pages 23-30, 1998." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 381, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 426 + ], + "type": "text", + "content": "[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 5, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 427, + 547, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 427, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 547, + 483 + ], + "type": "text", + "content": "[24] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 484, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 528 + ], + "type": "text", + "content": "[25] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 529, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 529, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 529, + 545, + 563 + ], + "type": "text", + "content": "[26] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. ACM Trans. Graph., 39(4), 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 564, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 545, + 597 + ], + "type": "text", + "content": "[27] John C. Hart. Sphere tracing: a geometric method for the antialiased ray tracing of implicit surfaces. The Visual Computer, 12(10):527-545, 1996. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 600, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 622 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[29] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. ACM SIGGRAPH Computer Graphics, 26 (2), 1992. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[30] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4689" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[31] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "text", + "content": "[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "text", + "content": "[33] Alireza Khatamian and Hamid R Arabnia. Survey on 3d surface reconstruction. Journal of Information Processing Systems, 12(3), 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[34] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 306 + ], + "type": "text", + "content": "[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "text", + "content": "[36] Ravikrishna Kolluri, Jonathan Richard Shewchuk, and James F O'Brien. Spectral surface reconstruction from noisy point clouds. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 11-21. ACM, 2004. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[37] Thomas Lewiner, Hélio Lopes, Antonio Wilson Vieira, and Geovan Tavares. Efficient implementation of marching cubes' cases with topological guarantees. Journal of Graphics Tools, 8(2):1-15, 2003. 1, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "text", + "content": "[38] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "type": "text", + "content": "[39] Chen-Hsuan Lin, Oliver Wang, Bryan C Russell, Eli Shechtman, Vladimir G Kim, Matthew Fisher, and Simon Lucey. Photometric mesh optimization for video-aligned 3d object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 969-978, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[40] David B. Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. arXiv preprint arXiv:0000.00000, 2021.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "text", + "content": "[41] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, 2021. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[42] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1886-1895, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[43] Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 715 + ], + "type": "text", + "content": "[44] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision, pages 68-84. Springer, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "type": "text", + "content": "[45] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "text", + "content": "[46] N. Maruani, R. Klokov, M. Ovsjanikov, P. Alliez, and M. Desbrun. Voromesh: Learning watertight surface meshes with voronoi diagrams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "text", + "content": "[47] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. arXiv preprint arXiv:2204.07159, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "type": "text", + "content": "[48] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "text", + "content": "[49] Mark Meyer, Mathieu Desbrun, Peter Schröder, and Alan H Barr. Discrete differential-geometry operators for triangulated 2-manifolds. In Visualization and mathematics III, pages 35-57. Springer, 2003. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 310, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 376 + ], + "type": "text", + "content": "[50] Niloy J. Mitra, Natasha Gelfand, Helmut Pottmann, and Leonidas Guibas. Registration of point cloud data from a geometric optimization perspective. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH Symposium on Geometry Processing, page 22-31, New York, NY, USA, 2004. Association for Computing Machinery. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "text", + "content": "[51] Patrick Mullen, Fernando De Goes, Mathieu Desbrun, David Cohen-Steiner, and Pierre Alliez. Signing the unsigned: Robust surface reconstruction from raw pointsets. Computer Graphics Forum, 29(5):1733-1741, 2010. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 453 + ], + "type": "text", + "content": "[52] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 478 + ], + "type": "text", + "content": "[53] G.M. Nielson. Dual marching cubes. In IEEE Visualization 2004, pages 489-496, 2004. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "type": "text", + "content": "[54] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 9964-9973, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 535, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 600 + ], + "type": "text", + "content": "[55] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2, 3, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "text", + "content": "[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[57] Helmut Pottmann and Michael Hofer. Geometry of the squared distance function to curves and surfaces. In Mathematics and Visualization, pages 221-242. Springer Berlin Heidelberg, 2003. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 715 + ], + "type": "text", + "content": "[58] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4690" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[59] Marie-Julie Rakotaosaona, Paul Guerrero, Noam Algerman, Niloy J Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22-31, 2021. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "text", + "content": "[60] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2, 3, 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "text", + "content": "[61] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 297 + ], + "type": "text", + "content": "[62] Scott Schaefer, Tao Ju, and Joe Warren. Manifold dual contouring. IEEE Transactions on Visualization and Computer Graphics, 13(3):610-619, 2007. 1, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "text", + "content": "[63] Nicholas Sharp and Maks Ovsjanikov. Pointtrinet: Learned triangulation of 3d point sets. In European Conference on Computer Vision, pages 762-778. Springer, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 386 + ], + "type": "text", + "content": "[64] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 443 + ], + "type": "text", + "content": "[65] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 42(4), 2023. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 476 + ], + "type": "text", + "content": "[66] Jonathan Shewchuk, Tamal K Dey, and Siu-Wing Cheng. Delaunay mesh generation. Chapman and Hall/CRC, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "text", + "content": "[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "text", + "content": "[68] G D Smith. Numerical solution of partial differential equations. Clarendon Press, Oxford, England, 3 edition, 1985. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "text", + "content": "[69] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. 2021. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "text", + "content": "[70] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[71] Graham M Treece, Richard W Prager, and Andrew H Gee. Regularised marching tetrahedra: improved iso-surface extraction. Computers & Graphics, 23(4):583-598, 1999. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 373 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[72] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-67, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[73] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10130-10139, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "type": "text", + "content": "[74] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 41(2):641-676, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 296 + ], + "type": "text", + "content": "[75] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In Advances in Neural Information Processing Systems 32, pages 492-502. Curran Associates, Inc., 2019. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "type": "text", + "content": "[76] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 341, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 545, + 373 + ], + "type": "text", + "content": "[77] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10, 000 3d-printing models. CoRR, abs/1605.04797, 2016. 6, 7" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4691" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_content_list.json b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..20833dd44afbfbb81f90fdd68a72b59836eeee09 --- /dev/null +++ b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_content_list.json @@ -0,0 +1,1594 @@ +[ + { + "type": "text", + "text": "Self-Supervised Facial Representation Learning with Facial Region Awareness", + "text_level": 1, + "bbox": [ + 125, + 161, + 846, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheng Gao Ioannis Patras", + "bbox": [ + 354, + 208, + 614, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Queen Mary University of London, Mile End Road, London, E1 4NS", + "bbox": [ + 235, + 239, + 736, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{z.gao, i.patras}@qmul.ac.uk", + "bbox": [ + 370, + 258, + 596, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 304, + 328, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised pre-training has been proved to be effective in learning transferable representations that benefit various visual tasks. This paper asks this question: can self-supervised pre-training learn general facial representations for various facial analysis tasks? Recent efforts toward this goal are limited to treating each face image as a whole, i.e., learning consistent facial representations at the image-level, which overlooks the \"consistency of local facial representations\" (i.e., facial regions like eyes, nose, etc). In this work, we make a first attempt to propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). Specifically, we explicitly enforce the consistency of facial regions by matching the local facial representations across views, which are extracted with learned heatmaps highlighting the facial regions. Inspired by the mask prediction in supervised semantic segmentation, we obtain the heatmaps via cosine similarity between the per-pixel projection of feature maps and \"facial mask embeddings\" computed from learnable positional embeddings, which leverage the attention mechanism to globally look up the facial image for facial regions. To learn such heatmaps, we formulate the learning of facial mask embeddings as a deep clustering problem by assigning the pixel features from the feature maps to them. The transfer learning results on facial classification and regression tasks show that our FRA outperforms previous pre-trained models and more importantly, using ResNet as the unified backbone for various tasks, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks.", + "bbox": [ + 109, + 333, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 787, + 233, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human face understanding is an important and challenging topic in computer vision [41, 72] and supervised learning algorithms have shown promising results on a wide range of facial analysis tasks recently [5, 31, 55, 71]. Despite", + "bbox": [ + 111, + 810, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the impressive progress, these methods require large-scale well-annotated training data, which is expensive to collect.", + "bbox": [ + 500, + 306, + 857, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works in self-supervised representation learning for visual images have shown that self-supervised pretraining is effective in improving the performance on various downstream tasks such as image classification, object detection and segmentation as it can learn general representations from unlabeled data that could be transferred to downstream visual tasks, especially tasks with limited labeled data [18, 24, 28, 30, 58, 62, 65]. Among them, instance discrimination (including contrastive learning [10, 22, 36] and non-contrastive learning [11, 20] paradigms) has been shown to be effective in learning generalizable self-supervised features. Instance discrimination aims to learn view-invariant representations by matching the global representations between the augmented views generated by aggressive image augmentations, i.e., the image-level representations of the augmented views should be similar [10-12, 20, 22, 36]. Another self-supervised learning paradigm, masked image modeling (MIM) [23, 54] learns visual representations by reconstructing image content from a masked image, achieving excellent performance in full model fine-tuning. This leads to the question: can self-supervised pre-training learn general facial representations which benefit downstream facial analysis tasks?", + "bbox": [ + 496, + 337, + 859, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several attempts have been made to learn general facial representations for facial analysis tasks [3, 41, 72]. For example, Bulat et al. [3] directly applies the contrastive objective to facial features. FaRL [72] and MCF [59] combine contrastive learning and mask image modeling [23]. PCL [41] proposes to disentangle the pose-related and pose-unrelated features, achieving strong performance on both pose-related (regression) and pose-unrelated (classification) tasks. However, it runs the model forward and backward three times for each training step, which is time-consuming. Despite different techniques, these methods commonly treats each face image as a whole to learn consistent global representations at the image-level and overlook the \"spatial consistency of local representations\", i.e., local facial regions (e.g., eyes, nose and mouth) should also be", + "bbox": [ + 496, + 657, + 859, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 138, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 259, + 42, + 782, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2081", + "bbox": [ + 483, + 890, + 511, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "similar across the augmented views, thus limiting the generalization to downstream tasks. This brings us to the focus: learning consistent global and local representations for facial representation learning.", + "bbox": [ + 111, + 126, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We argue that in order to learn consistent local representations, the model needs to look into facial regions. Towards that goal, we predict a set of heatmaps highlighting different facial regions by leveraging learnable positional embeddings as facial queries (the feature maps as keys and values) to look up the facial image globally for facial regions, which is inspired by the mask prediction in supervised segmentation [13]. For visual images, the attention mechanism of Transformer allows the learnable positional embeddings to serve as object queries for visual pattern look-up [7, 13]. In our case (facial images), the learnable positional embeddings can be used as facial queries for facial regions (see the visualization in the supplementary material).", + "bbox": [ + 111, + 185, + 473, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, taking the consistency of facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns general facial representations by enforcing consistent global and local facial representations, based on a popular instance discrimination baseline BYOL [20] for its simplicity. Specifically, we learn consistent local facial representations by match them across augmented views, which are extracted by aggregating the feature maps using learned heatmaps highlighting the facial regions as weights. Inspired by the mask prediction in MaskFormer [13], we produce the heatmaps from a set of learnable positional embeddings, which are used as facial queries to look up the facial image for facial regions. A Transformer decoder takes as input the feature maps from the encoder and the learnable positional embeddings to output a set of \"facial mask embeddings\", each associated with a facial region. The facial mask embeddings are used to compute cosine similarity with the per-pixel projection of feature maps to produce the heatmaps. In addition, we enforce the consistency of global representations across views simultaneously so that the image-level information is preserved. In order to learn the heatmaps (facial mask embeddings), inspired by deep clustering [8] that learns to assign samples to clusters, we treat the facial mask embeddings as clusters and learn to assign pixel features from the feature maps to them. Specifically, we align the per-pixel cluster assignments of each pixel feature over the facial region clusters between the online and momentum network for the same augmented view (i.e., each pixel feature should have similar similarity distribution over the facial mask embeddings between the momentum teacher and online student). In contrast to supervised segmentation that directly uses ground-truth masks to supervise the learning of the masks (heatmaps) with a per-pixel binary mask loss, we formulate the learning of heatmaps as a deep cluster", + "bbox": [ + 111, + 369, + 473, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing [8] problem that learns to assign pixel features to clusters (facial mask embeddings) in a self-supervised manner.", + "bbox": [ + 498, + 126, + 856, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 516, + 154, + 815, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Taking into the consistency of local facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns consistent global and local facial representations.", + "- We show that the learned heatmaps can roughly discover facial regions in the supplementary material.", + "- In previous works, different backbones are adopted for different facial analysis tasks (e.g., in face alignment the common backbone is hourglass network [63] while in facial expression recognition ResNet [21] is the popular backbone). In this work, our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks.", + "- Our FRA outperforms existing self-supervised pretraining approaches (e.g., BYOL [20] and PCL [41]) on facial classification (i.e., facial expression recognition [19, 38] and facial attribute recognition [42]) and regression (i.e., face alignment [48-50, 60]) tasks. More importantly, our FRA achieves comparable (e.g., face alignment) or even better performance (e.g., facial expression recognition) compared with SOTA methods in the corresponding facial analysis tasks." + ], + "bbox": [ + 500, + 169, + 857, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 501, + 624, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual representation learning", + "text_level": 1, + "bbox": [ + 500, + 525, + 745, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As one of the main paradigms for self-supervised pretraining, instance discrimination learns representations by treating an image as a whole and enforcing the consistency of global representations at the image-level across augmented views. Generally, instance discrimination includes two paradigms: contrastive learning [10, 22, 36] and non-contrastive learning [11, 20]. Contrastive learning considers each image and its transformations as a distinct class, i.e., \"positive\" samples are pulled together while \"negative\" samples are pushed apart in the latent space. Unlike contrastive learning that relies on negative samples to avoid collapse, non-contrastive learning directly maximizes the similarity of the global representations between the augmented views without involving negative samples based on techniques like stop-gradient [11] and predictor [20]. Further works perform visual-language pre-training by applying contrastive objective to image-text pairs [29, 35, 47].", + "bbox": [ + 496, + 547, + 857, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of work, masked image modeling (MIM) learns visual representations by reconstructing image content from a masked image [1, 23, 54], which is inspired by the masked language modeling in NLP [15]. In contrast to instance discrimination, MIM achieves strong full model fine-tuning performance with Vision Transformers", + "bbox": [ + 496, + 782, + 857, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2082", + "bbox": [ + 483, + 890, + 514, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pre-trained for enough epochs. However, these works suffer from poor linear separability and are less data-efficient than instance discrimination in few-shot scenarios [1].", + "bbox": [ + 112, + 126, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Facial representation learning", + "text_level": 1, + "bbox": [ + 112, + 176, + 356, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works on facial analysis explore self-supervised learning for several face-related tasks, such as facial expression recognition [9, 53], face recognition [9, 57], facial micro-expression recognition [46], AU detection [39, 40], face alignment (facial landmark detection) [14, 64], etc. However, these methods are task-specific, i.e., tailored for a specific facial task and thus lack the ability to generalize to various facial analysis tasks [41]. Further efforts [3, 41, 72] focus on learning general facial representations with contrastive learning and mask image modeling [23, 54]. Bulat et al. [3] directly apply the contrastive objective to augmented views of the same face image, showing that general facial representation learned from pre-training can benefit various facial analysis tasks. FaRL [72] performs pre-training in a visual-linguistic manner by employing image-text contrastive learning and masked image modeling. MCF [59] leverages image-level contrastive learning and masked image modeling, along with the knowledge distilled from external ImageNet pre-trained model for facial representation learning. PCL [41] argues that directly applying the contrastive objective to face images overlooks the variances of facial poses and thus leads to pose-invariant representations, limiting the performance on pose-related tasks [66, 75]. Therefore, PCL [41] disentangles the pose-related and pose-unrelated features and then performs contrastive learning on these features, achieving strong performance on both pose-related and pose-unrelated facial analysis tasks. Despite the success, it performs forward and backward three times for each input image, which brings significant increase on computational cost. These works are commonly limited by instance discrimination and overlook the consistency of local facial regions. In contrast, inspired by supervised semantic segmentation, we learn consistent global and local facial representations by learning a set of heatmaps indicating facial regions from learnable positional embeddings, which leverage the attention mechanism to look up facial image globally for facial regions.", + "bbox": [ + 112, + 198, + 473, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Facial region discovery", + "text_level": 1, + "bbox": [ + 112, + 719, + 307, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are some approaches leveraging facial region (landmark) discovery for facial analysis [27, 43, 61]. Some focus on landmark detection by either learning a heatmap for each landmark via image reconstruction [27, 68], or performing pixel-level matching with an equivariance loss [56, 68]. Despite different techniques, these methods are task-specific, i.e., landmark detection with discovery of local information, while our method is task-agnostic, i.e., learn general facial representations for various tasks by preserving", + "bbox": [ + 112, + 741, + 470, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "global and local information in a image, region and pixel-level contrastive manner. MARLIN [4] applies masked image modeling to learn general representations for facial videos by utilizing an external face parsing algorithm to discover the facial regions (e.g., eyes, nose and mouth), which are used to guide the masking for the masked autoencoder. A closely related work SLPT [61] leverages the attention mechanism to estimate facial landmarks from initial facial landmarks estimates of the mean face through supervised learning. These works commonly rely on external supervisory signal, whether it is from ground-truths or additional algorithms. In contrast, we learn to discover the facial regions in an end-to-end self-supervised manner for facial image representation learning.", + "bbox": [ + 496, + 126, + 857, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 333, + 622, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 356, + 599, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of the proposed FRA is shown in Fig. 1. The goal is to learn consistent global and local facial representations. Toward this goal, we propose two objectives: pixel-level semantic relation and image/region-level semantic consistency. Semantic relation aligns the per-pixel cluster assignments of each pixel feature over the facial mask embeddings between the online and momentum network to learn heatmaps for facial regions (Sec. 3.2) while semantic consistency directly matches the global and local facial representations across augmented views (Sec. 3.3) with the learned heatmaps.", + "bbox": [ + 496, + 377, + 857, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantic relation", + "text_level": 1, + "bbox": [ + 500, + 539, + 656, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, our method adopts the Siamese structure of BYOL [20], a popular self-supervised pretraining baseline based on instance discrimination. Following BYOL [20], we employ two branches: the online network parameterized by $\\theta$ and the momentum network parameterized by $\\xi$ . The online network $\\theta$ consists of an encoder $E_{\\theta}$ , a global projector $H_{\\theta}^{g}$ and a local projector $H_{\\theta}^{l}$ . The momentum network $\\xi$ has the same architecture with the online network, except $\\xi$ is updated with an exponential moving average of $\\theta$ . As in BYOL [20], we also adopt additional predictors $G_{\\theta}^{g}$ and $G_{\\theta}^{l}$ on top of the projectors in the online network. Note that this is omitted for brevity in Fig. 1.", + "bbox": [ + 496, + 561, + 857, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $\\mathbf{x}$ , two random augmentations are applied to generate two augmented views $\\mathbf{x}_1 = \\mathcal{T}_1(\\mathbf{x})$ and $\\mathbf{x}_2 = \\mathcal{T}_2(\\mathbf{x})$ , following BYOL [20]. Each augmented view $\\mathbf{x}_i \\in \\{\\mathbf{x}_1, \\mathbf{x}_2\\}$ is fed into an encoder $E$ to obtain a feature map $\\mathbf{F}_i \\in \\mathbb{R}^{C \\times H \\times W}$ (before global average pooling), where $C, H, W$ are the number of channels, height and width of $\\mathbf{F}_i$ and a latent representation $\\mathbf{h}_i \\in \\{\\mathbf{h}_1, \\mathbf{h}_2\\}$ (after global average pooling), i.e., $\\mathbf{h}_1 = E_{\\theta}(\\mathbf{x}_1)$ and $\\mathbf{h}_2 = E_{\\xi}(\\mathbf{x}_2)$ . Then each latent representation $\\mathbf{h}_i$ is transformed", + "bbox": [ + 496, + 742, + 857, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2083", + "bbox": [ + 483, + 890, + 513, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4db6d3a6998f58ae0a8d163e71b2d81b1929bebbd4b443f85e50ee48549d111b.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed FRA framework. $\\odot$ denotes cosine similarity. For each input image $\\mathbf{x}$ , its augmented views $\\mathbf{x}_1$ and $\\mathbf{x}_2$ are passed into two network branches to produce the global embeddings $\\mathbf{z}_1$ and $\\mathbf{z}_2$ . In addition, we produce a set of heatmaps $\\mathbf{M}_1$ and $\\mathbf{M}_2$ indicating the local facial regions, via the correlation between the pixel features and \"facial mask embeddings\" computed from a set of learnable positional embeddings. Then we aggregate the feature map to obtain the local facial embeddings $\\{\\mathbf{z}_1^m\\}$ and $\\{\\mathbf{z}_2^m\\}$ . The semantic consistency loss is applied to global embeddings and facial embeddings to maximize the similarity across augmented views. To learn such heatmaps, i.e., facial mask embeddings, we treat the facial mask embeddings as facial region clusters and propose a semantic relation loss to align the cluster assignments of each pixel feature over the facial region clusters between the online and momentum network." + ], + "image_footnote": [], + "bbox": [ + 155, + 123, + 818, + 338 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/12e98237741f2f7ca67d9f53d279a2aae367a7bf7b91549941c016c0bf2bdb16.jpg", + "image_caption": [ + "Figure 2. Generation of heatmaps using learnable positional embeddings as facial queries and the feature maps as keys and values." + ], + "image_footnote": [], + "bbox": [ + 117, + 460, + 472, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by a global projector $H^g$ to produce a global embedding $\\mathbf{z}_i\\in \\{\\mathbf{z}_1,\\mathbf{z}_2\\}$ of dimension $\\mathbf{z}_i\\in \\mathbb{R}^D$ , i.e., $\\mathbf{z}_1 = H_\\theta^g (\\mathbf{h}_1)$ and $\\mathbf{z}_2 = H_{\\xi}^{g}(\\mathbf{h}_{2})$", + "bbox": [ + 111, + 661, + 470, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we obtain a set of heatmaps $\\mathbf{M}_i\\in \\{\\mathbf{M}_1,\\mathbf{M}_2\\}$ highlighting the facial regions from the feature map $\\mathbf{F}_i$ for each view, which is inspired by mask classification-based supervised segmentation [7, 13] that leverages attention mechanism to look up visual patterns globally. First the local projector (e.g., $H_{\\theta}^{l}$ ) is applied to project the pixel features of $\\mathbf{F}_i$ in a pixel-wise manner, mapping it to $D$ dimensions to get the dense feature map $\\mathbf{F}_i^{\\mathrm{dense}}\\in \\mathbb{R}^{D\\times H\\times W}$ . Take view $\\mathbf{x}_1$ as an example, the projected feature map can be expressed as:", + "bbox": [ + 111, + 702, + 473, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {1} ^ {\\text {d e n s e}} [ *, u, v ] = H _ {\\theta} ^ {l} \\left(\\mathbf {F} _ {1} [ *, u, v ]\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 849, + 472, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_1[*,u,v]\\in \\mathbb{R}^C$ is the pixel feature at the $(u,v)$ -th grid of $\\mathbf{F}_1$ . Then as shown in Fig. 2, inspired by supervised segmentation [13], we use a Transformer decoder followed by a MLP, which takes as input the feature map $\\mathbf{F}_i$ and $N$ learnable positional embeddings (i.e., facial queries for looking up the facial image globally for facial regions) to predict $N$ \"facial mask embeddings\" $\\mathbf{Q}_i\\in \\mathbb{R}^{N\\times D}$ of dimension $D$ , where each row associated with a facial region. Next, we compute the cosine similarity between facial mask embeddings $\\mathbf{Q}_i$ and dense feature map $\\mathbf{F}_i^{\\mathrm{dense}}$ along the channel dimension, yielding per-pixel cluster assignments $\\mathbf{S}_i\\in \\mathbb{R}^{N\\times H\\times W}$ , where $\\mathbf{S}_i[*,u,v]$ denotes the relation between the dense pixel feature $\\mathbf{F}_1^{\\mathrm{dense}}[*,u,v]$ and facial mask embeddings $\\mathbf{Q}_i$ . Finally, we normalize $\\mathbf{S}_i$ along the channel dimension with a softmax operation to encourage each channel to capture a different pattern, obtaining $N$ heatmaps $\\mathbf{M}_i\\in \\mathbb{R}^{N\\times H\\times W}$ where each vector at location $(u,v)$ is a probabilistic similarity distribution (i.e., normalized per-pixel cluster assignments) $\\mathbf{s}_1^{u,v}$ between $\\mathbf{F}_i^{\\mathrm{dense}}[*,u,v]$ and $\\mathbf{Q}_i$ . Note that $\\mathbf{M}_i$ is a set of heatmaps where each channel of $\\mathbf{M}_i$ represents a 2D heatmap $\\mathbf{M}_i^{(m)}$ .", + "bbox": [ + 496, + 460, + 859, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn such heatmaps, i.e., facial mask embeddings, inspired by deep clustering [8], we treat the facial mask embeddings as facial region clusters and align the per-pixel cluster assignments of each pixel feature over these clusters between the online and momentum network for the same augmented view, using the momentum network as momentum teacher [16, 73] to provide reliable target.", + "bbox": [ + 496, + 755, + 857, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following BYOL [20], we pass both augmented views", + "bbox": [ + 516, + 852, + 857, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2084", + "bbox": [ + 483, + 890, + 513, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{x}_1$ and $\\mathbf{x}_2$ through the online and momentum network. Take $\\mathbf{x}_1$ as an example, the online network $\\theta$ outputs the normalized per-pixel cluster assignments $\\mathbf{s}_1^{u,v}$ and the momentum network outputs normalized assignments $\\widehat{\\mathbf{s}}_1^{u,v}$ for view $\\mathbf{x}_1$ . Then we learn $\\mathbf{s}_1^{u,v}$ using $\\widehat{\\mathbf{s}}_1^{u,v}$ as guidance based on the following cross-entropy loss:", + "bbox": [ + 111, + 126, + 472, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC E \\left(\\mathbf {s} _ {1} ^ {u, v}, \\widehat {\\mathbf {s}} _ {1} ^ {u, v}\\right) = - \\sum_ {m = 1} ^ {N} \\widehat {\\mathbf {s}} _ {1} ^ {u, v} [ m ] \\log \\mathbf {s} _ {1} ^ {u, v} [ m ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 219, + 472, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For both augmented views, we define the symmetrized semantic relation objective as:", + "bbox": [ + 112, + 268, + 472, + 296 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r}} = \\frac {1}{H W} \\sum_ {u, v} \\left(C E \\left(\\mathbf {s} _ {1} ^ {u, v}, \\widehat {\\mathbf {s}} _ {1} ^ {u, v}\\right) + C E \\left(\\mathbf {s} _ {2} ^ {u, v}, \\widehat {\\mathbf {s}} _ {2} ^ {u, v}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 306, + 472, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $CE(\\mathbf{s}_2^{u,v},\\widehat{\\mathbf{s}}_2^{u,v})$ is the cross-entropy loss for view $\\mathbf{x}_2$ . We apply the Sinkhorn-Knopp normalization to the target assignments from the momentum network following [8] to avoid collapse and the mean entropy maximization (ME-MAX) regularizer [1] to maximize the entropy of the prediction to encourage full use of the clusters.", + "bbox": [ + 111, + 351, + 472, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Semantic consistency", + "text_level": 1, + "bbox": [ + 112, + 445, + 295, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we enforce the consistency of global embeddings and local facial embeddings. With the learned heatmaps $\\mathbf{M}_i$ , we generate the latent representations for the local facial regions through weighted average pooling:", + "bbox": [ + 111, + 465, + 472, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {h} _ {i} ^ {m} = \\mathbf {M} _ {i} ^ {(m)} \\otimes \\mathbf {F} _ {i} \\\\ = \\frac {1}{\\sum_ {u , v} \\mathbf {M} _ {i} [ m , u , v ]} \\sum_ {u, v} \\mathbf {M} _ {i} [ m, u, v ] \\mathbf {F} _ {i} [ *, u, v ], \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 534, + 470, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes channel-wise weighted average pooling, $\\mathbf{M}_i^{(m)}$ is the $m$ -th channel (heatmap) of $\\mathbf{M}_i$ and $\\mathbf{h}_i^m \\in \\mathbb{R}^C$ is the corresponding latent representation produced with $\\mathbf{M}_i^{(m)}$ . The facial embeddings $\\{\\mathbf{z}_1^m : \\mathbf{z}_1^m \\in \\mathbb{R}^D\\}_{m=1}^N$ and $\\{\\mathbf{z}_2^m : \\mathbf{z}_2^m \\in \\mathbb{R}^D\\}_{m=1}^N$ are obtained accordingly with the local projector $H_\\theta^l$ and $H_\\xi^l$ :", + "bbox": [ + 112, + 597, + 472, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {z} _ {1} ^ {m} = H _ {\\theta} ^ {l} \\left(\\mathbf {h} _ {1} ^ {m}\\right), \\\\ \\mathbf {z} _ {2} ^ {m} = H _ {\\xi} ^ {l} \\left(\\mathbf {h} _ {2} ^ {m}\\right). \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 696, + 470, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then match the global embeddings and local facial embeddings across views using the negative cosine similarity in BYOL [20]:", + "bbox": [ + 111, + 738, + 470, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}\\right) = - \\left(\\lambda_ {\\mathrm {c}} \\times f _ {\\mathrm {s}} \\left(G _ {\\theta} ^ {g} \\left(\\mathbf {z} _ {1}\\right), \\mathbf {z} _ {2}\\right) + \\right. \\\\ + \\left(1 - \\lambda_ {\\mathrm {c}}\\right) \\times \\frac {1}{N} \\sum_ {m = 1} ^ {N} f _ {\\mathrm {s}} \\left(G _ {\\theta} ^ {l} \\left(\\mathbf {z} _ {1} ^ {m}\\right), \\mathbf {z} _ {2} ^ {m}\\right)), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 792, + 470, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{s}}(\\mathbf{u},\\mathbf{v}) = \\frac{\\mathbf{u}^{\\top}\\mathbf{v}}{\\|\\mathbf{u}\\|_{2}\\|\\mathbf{v}\\|_{2}}$ denotes the cosine similarity between the vectors $\\mathbf{u}$ and $\\mathbf{v}$ , $\\lambda_{\\mathrm{c}}$ is the loss weight, $G_{\\theta}^{g}$ and $G_{\\theta}^{l}$ are the predictors on top of the projectors $H_{\\theta}^{g}$ and $H_{\\theta}^{l}$ , respectively. Following BYOL [20], we symmetrize the loss $\\mathcal{L}_{\\mathrm{sim}}(\\mathbf{z}_1,\\mathbf{z}_2)$ defined in Eq. (6) by passing $\\mathbf{x}_1$ through the momentum network $\\xi$ and $\\mathbf{x}_2$ through the online network $\\theta$ to compute $\\mathcal{L}_{\\mathrm{sim}}(\\mathbf{z}_2,\\mathbf{z}_1)$ . The semantic consistency objective can be expressed as follows:", + "bbox": [ + 496, + 125, + 859, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}\\right) + \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {2}, \\mathbf {z} _ {1}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 251, + 857, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall objective", + "text_level": 1, + "bbox": [ + 500, + 277, + 653, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We jointly optimize the semantic relation objective Eq. (3) and the semantic consistency objective Eq. (7), leading to the following overall objective:", + "bbox": [ + 496, + 300, + 857, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 353, + 857, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{r}}$ is the loss weight for balancing $\\mathcal{L}_{\\mathrm{c}}$ and $\\mathcal{L}_{\\mathrm{r}}$", + "bbox": [ + 500, + 378, + 813, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 407, + 619, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setups", + "text_level": 1, + "bbox": [ + 500, + 431, + 676, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Implementation details", + "text_level": 1, + "bbox": [ + 500, + 452, + 696, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the same augmentation strategy as in [20, 25]. The number of heatmaps $N$ is set to 8 empirically. The loss weight $\\lambda_{\\mathrm{c}}$ and $\\lambda_{\\mathrm{r}}$ are set to $0.5 / 0.1$ , respectively. For fair comparisons, the other hyper-parameters are kept the same as BYOL [20] in all experiments. The architecture and pretraining details are provided in the supplementary material.", + "bbox": [ + 496, + 474, + 857, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Baselines", + "text_level": 1, + "bbox": [ + 500, + 577, + 607, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our baselines are self-supervised pre-training approaches for visual images (e.g., BYOL [20] and LEWEL [25]), and pre-training approaches for facial images (e.g., Bulat et al. [3] and PCL [41]). Note that SwAV [8] is equivalent to Bulat et al. [3]. As we adopt BYOL [20] as the pretraining backbone, we compare our FRA with BYOL [20] in all experiments. We also compare our FRA with another pre-training method LEWEL [25], which learns consistent local representations for visual images. Moreover, we perform comparisons with SOTA methods in the corresponding downstream tasks.", + "bbox": [ + 496, + 599, + 857, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation protocols", + "text_level": 1, + "bbox": [ + 500, + 761, + 678, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the common practice in previous works [41, 72], we evaluate the transfer performance of the self-supervised pre-trained facial representations on several popular downstream facial analysis tasks: facial expression recognition (FER) [2, 38], facial attribute recognition (FAR) [42] and face alignment (FA) [48-50, 60]. Specifically, we use", + "bbox": [ + 496, + 782, + 857, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2085", + "bbox": [ + 483, + 890, + 513, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fc3864c927a80c34fc398a856e0110025bd0ec861ba9b11b5fb3f691daffba3a.jpg", + "table_caption": [ + "Table 1. Comparisons with weakly-supervised pre-trained vision transformer on several downstream facial analysis tasks, including facial expression recognition (AffectNet), facial attribute recognition (CelebA) and face alignment (300W)." + ], + "table_footnote": [], + "table_body": "
MethodArch.Params.Pre-training settingsDownstream performances
DatasetScaleSupervisionAffectNet Acc. ↑CelebA Acc. ↑300W NME ↓
FaRL [72]ViT-B/16 [17]86MLAION-FACE [72]20Mface image + text64.8591.883.08
FRAR50 [21]24MVGGFace2 [6]3.3Mface image66.1692.022.91
", + "bbox": [ + 133, + 159, + 838, + 252 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c4387200e8da7f4b2db8832951e592bb26971a8991dd23819c6892669337606f.jpg", + "table_caption": [ + "Table 2. Comparisons on facial expression recognition. We report the Top-1 accuracy on test set. Text denotes text supervision.†: our reproduction using the official codes." + ], + "table_footnote": [], + "table_body": "
MethodTextFERPlusRAF-DBAffectNet
Supervised
KTN [32]×90.4988.0763.97
RUL [69]×88.7588.9861.43
EAC [70]×90.0590.3565.32
Weakly-Supervised
FaRL [72]†✓88.6288.3164.85
CLEF [67]✓89.7490.0965.66
Self-supervised
MCF [59]†×88.1786.8660.98
Bulat et al. [3, 8]×--60.20
BYOL [20]×89.2589.5365.65
LEWEL [25]×85.6181.8561.20
PCL [41]×85.8785.9260.77
FRA (LP)×78.1373.8957.38
FRA (FT)×89.7889.9566.16
FRA (EAC)×90.6290.7665.85
", + "bbox": [ + 114, + 320, + 473, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the pre-trained weights to initialize the backbone of downstream tasks and then learn the backbone and task-specific head (attached to the backbone) jointly. Following [72], we report the performance with linear probe (denoted by \"LP\") and fine-tuning (denoted by \"FT\"). The details of the downstream tasks are described as follows:", + "bbox": [ + 109, + 611, + 470, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Facial expression recognition is a multi-class classification task where the goal is to categorize the emotional expressions (e.g., anger, fear and surprise) for a given face image. Three widely-used datasets are adopted: FERPlus [2], RAF-DB [38] and AffectNet [45]. For RAF-DB, we use the basic emotion subset following [32, 41, 70]. For AffectNet, we report the results with 7 emotion classes (i.e., neutral, happy, sad, surprise, fear, anger, disgust) following [32, 70].", + "bbox": [ + 109, + 696, + 470, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Facial attribute recognition is a multi-label classification task to predict various attributes (e.g., gender, age and race) of a given face image. We adopt the popular benchmark CelebA [42], which consists of more than 200K face", + "bbox": [ + 109, + 811, + 472, + 866 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1f87af4b59b60e2e6694e61b4e84346742a27bfd36b0cb63a9eeb51a98d8c3dd.jpg", + "table_caption": [ + "Table 3. Comparisons on CelebA [42] facial attribute recognition. We report the averaged accuracy over all attributes. †: our reproduction using the official codes. *: results cited from [72]." + ], + "table_footnote": [], + "table_body": "
MethodAcc. (%)
Supervised
DMM [44]91.70
SlimCNN [51]91.24
AFFAIR [34]91.45
Self-supervised
SSPL [52]91.77
Bulat et al. [3, 8]*89.65
SimCLR [10]*91.08
BYOL [20]91.56
LEWEL [25]90.69
PCL [41]91.48
MCF [59]†91.33
FRA (LP)90.86
FRA (FT)92.02
", + "bbox": [ + 576, + 320, + 782, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images with 40 facial attributes per image. Following [72], we report the averaged accuracy over all attributes.", + "bbox": [ + 498, + 571, + 856, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face alignment is a regression task to predict 2D face landmark coordinates on a face image. We use two popular benchmarks: WFLW [60] and 300W [48-50]. Following the common practice [14, 26, 74], we report normalized mean error (NME), failure rate (FR) and AUC. For 300W, we report the results on full test set, common (554 images) and challenge (135 images) splits of the test set following [26, 74].", + "bbox": [ + 496, + 601, + 857, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with weakly-supervised pretraining", + "text_level": 1, + "bbox": [ + 498, + 720, + 857, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we compare our FRA with SOTA pre-trained Transformer FaRL [72], which is a weakly-supervised model pre-trained on 20M visual-linguistic data (face image and text) with image-text contrastive learning and mask image modeling. We fine-tune both the pre-trained feature backbone and the task-specific head on the corresponding downstream facial analysis task. Our self-supervised FRA with 24M parameter ResNet-50 achieves superior perfor", + "bbox": [ + 496, + 755, + 857, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2086", + "bbox": [ + 483, + 890, + 514, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8eeb4f22084a5ced5d065f40354f57d244db8603a83d1ad70b53c94f462f46fb.jpg", + "table_caption": [ + "Table 4. Comparisons on face alignment. †: our reproduction using the official codes." + ], + "table_footnote": [], + "table_body": "
MethodVenueArch.WFLW300W (NME ↓)
NME ↓FR10% ↓AUC10% ↑FullComm.Chal.
Supervised
SLPT [61][CVPR'22]ResNet [21]4.203.040.5883.202.784.93
DTLD [33][CVPR'22]ResNet [21]4.082.76-2.962.594.50
RePFormer [37][IJCAI'22]ResNet [21]4.11--3.01--
ADNet [26][ICCV'21]Hourglass [63]4.142.720.6022.932.534.58
STAR [74][CVPR'23]Hourglass [63]4.022.320.6052.872.524.32
Self-supervised
MCF [59] (concurrent work)[ACMM'M'23]ViT [17]3.961.400.6092.982.604.51
Bulat et al. [3, 8][ECCV'22]ResNet [21]4.57--3.20--
BYOL [20][NeurIPS'20]ResNet [21]4.292.960.5793.032.664.55
LEWEL [25][CVPR'22]ResNet [21]4.524.500.5633.092.704.71
PCL [41]†[CVPR'23]ResNet [21]4.846.180.5353.352.775.12
FRAOursResNet [21]4.112.530.5912.912.604.46
", + "bbox": [ + 114, + 147, + 869, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance compared with weakly-supervised FaRL [72] with 86M parameter ViT-B/16 and text supervision on all tasks.", + "bbox": [ + 112, + 404, + 470, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Transfer learning", + "text_level": 1, + "bbox": [ + 112, + 453, + 269, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare our FRA with self-supervised pre-training approaches and SOTA methods in several downstream tasks. Please refer to the supplementary material for setup details.", + "bbox": [ + 111, + 479, + 470, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Facial expression recognition", + "text_level": 1, + "bbox": [ + 112, + 574, + 346, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results on facial expression recognition are reported in Tab. 2. We observe: (1) With the setting of fine-tuning (FT), our FRA outperforms previous self-supervised pre-training approaches for visual images (e.g., BYOL and LEWEL) and pre-training approaches tailored for facial images (e.g., PCL, MCF). In particular, our FRA using a 24M parameter ResNet-50 surpasses the concurrent work MCF [59] with 86M parameter ViT-B/16 [17]. (2) By simply learning a linear classifier on top of the encoder backbone, our FRA outperforms SOTA facial expression recognition methods with sophisticated designs (e.g., EAC [70]) on AffectNet [45], the largest facial expression recognition dataset. (3) More importantly, by using our pre-trained model to initialize the backbone of SOTA facial expression recognition method EAC [70], our variant \"FRA (EAC)\" consistently improves EAC [70] on all datasets, which suggests \"FRA (EAC)\" outperforms SOTA FER methods and demonstrates the superiority of the proposed self-supervised pre-training.", + "bbox": [ + 111, + 603, + 473, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Facial attribute recognition", + "text_level": 1, + "bbox": [ + 500, + 404, + 723, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 3, our FRA outperforms both self-supervised pre-training approaches for visual images and pre-training approaches tailored for facial images. The results on facial expression recognition and facial attribute recognition show that our FRA learns better facial representations for facial classification task.", + "bbox": [ + 498, + 426, + 857, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3 Face alignment", + "text_level": 1, + "bbox": [ + 500, + 531, + 645, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 4, despite SOTA face alignment methods (e.g., ADNet [26] and STAR [74]) commonly rely on hourglass network [63] for feature extraction, which is tailored for regression tasks like landmark detection, our method based on ResNet backbone achieves comparable performance with these SOTA methods (e.g., 2.91 vs. 2.87 on 300W). The results on classification (e.g., facial expression recognition) and regression tasks (e.g., face alignment) show that our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks.", + "bbox": [ + 498, + 554, + 857, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 717, + 645, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We pre-train the model on VGGFace2 [6] and then evaluate it on facial expression recognition (RAF-DB) and facial attribute recognition (CelebA), as described in Sec. 4.2.", + "bbox": [ + 498, + 739, + 857, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.1 Effect of different modules", + "text_level": 1, + "bbox": [ + 500, + 801, + 718, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 5, we investigate the contributions of the proposed semantic consistency loss (i.e., global consistency of whole face and local consistency of facial regions) and semantic", + "bbox": [ + 500, + 824, + 857, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2087", + "bbox": [ + 483, + 890, + 513, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb7a6e19e300c60c00cfe8c8ab9f35cfaec370acda28c91eb8bac7d839a33e90.jpg", + "table_caption": [ + "Table 5. Effect of different modules. GC denotes the global consistency for aligning images, LC denotes the local consistency for aligning facial regions with heatmaps and SR represents semantic relation for aligning pixels and heatmaps." + ], + "table_footnote": [], + "table_body": "
GCLCSRRAF-DB ↑CelebA ↑300W ↓
✓--87.8290.633.56
--✓85.3490.663.25
-✓-87.4690.783.19
✓✓-88.0590.893.26
✓✓✓88.7291.183.14
", + "bbox": [ + 122, + 184, + 460, + 282 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0a098eb2653778887895f4b4fb92d31aa0714062d0306cdfc1093bfadc4190c9.jpg", + "table_caption": [ + "Table 6. Effect of the number of heatmaps $N$ ." + ], + "table_footnote": [], + "table_body": "
N83264
RAF-DB88.7288.3688.18
CelebA91.1891.0190.98
", + "bbox": [ + 176, + 323, + 406, + 381 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/13930d0d44b24126d33ca7d88c729f4a20c51cf4c94345d58fd7db899616e067.jpg", + "table_caption": [ + "Table 7. Effect of the loss weights. Please refer to Tab. 5 for GC, LC and SR." + ], + "table_footnote": [], + "table_body": "
SettingsλcλrRAF-DBCelebA
GC1.0087.8290.63
GC + LC0.5088.0590.89
GC + LC + SR (FRA)0.50.188.7291.18
GC + LC + SR (FRA)0.50.588.4591.04
GC + LC + SR (FRA)0.51.088.0890.46
", + "bbox": [ + 119, + 434, + 467, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "relation loss to our approach. Note that the global consistency (first row) is BYOL [20]. We have the following observations: (1) The variant using all losses achieves the best results. (2) GC is essential to avoid degeneration on classification task. (3) LC or SR alone benefits regression task (landmark). Altogether, LC and SR improve BYOL [20] (GC) on both classification and regression by capturing spatial/local information, which validates our facial region awareness.", + "bbox": [ + 111, + 560, + 470, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.2 Effect of the number of heatmaps", + "text_level": 1, + "bbox": [ + 112, + 715, + 375, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 6, we study the effect of the number of heatmaps. We observe that the best setting is $N = 8$ , which is close to the facial landmarks number 5. This suggests that given enough face images for training, a suitable $N$ can encourage the model to learn face-specific patterns, which helps the transfer learning performance on various facial analysis tasks. Further increasing the number of heatmaps might force the model to look into fine-grained patterns that may not be suitable for facial tasks.", + "bbox": [ + 111, + 741, + 470, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.3 Effect of loss weights", + "text_level": 1, + "bbox": [ + 500, + 126, + 681, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 7, we ablate the weights for the semantic consistency loss and semantic relation loss. We find that setting $\\lambda_{\\mathrm{c}} = 0.5$ and $\\lambda_{\\mathrm{r}} = 0.1$ works best. When $\\lambda_{\\mathrm{c}} = 1.0$ and $\\lambda_{\\mathrm{r}} = 0$ , only the consistency of global representations is applied, and the model performs relatively worse, which suggests the importance of the consistency of local representations and the semantic relation loss. By using the semantic relation objective, the performance is significantly improved. However, when $\\lambda_{\\mathrm{r}}$ is too high, the performance degrades as the pixel-level consistency between the online and momentum network might affect the capture of image/object-level information.", + "bbox": [ + 496, + 147, + 857, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.4 Effect of Transformer decoder layers", + "text_level": 1, + "bbox": [ + 500, + 333, + 782, + 347 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5cfe6ac8ce5db9e46595577460fba60480a1351aa29350609cee8b36ae22dc8d.jpg", + "table_caption": [ + "Table 8. Effect of Transformer decoder layers. 0 decoder layer represents BYOL [20] where only the consistency of global representation is enforced." + ], + "table_footnote": [], + "table_body": "
# decoder layer0123
RAF-DB87.8288.7289.0189.06
CelebA90.6391.1891.3091.35
", + "bbox": [ + 519, + 416, + 839, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 8, we study the effect of the number of decoder layers used for heatmap prediction. We observe that a single decoder layer is able to produce decent results, showing that a 1-layer decoder is large enough to capture the facial region (landmarks) relations in face images. The performance gain diminishes as the decoder depth increases. By default, we only use 1 decoder layer for fast training.", + "bbox": [ + 496, + 487, + 857, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 598, + 607, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). We learn a set of heatmaps indicating facial regions from learnable positional embeddings, which leverages the attention mechanism to look up facial image globally for facial regions. We show that our FRA outperforms previous pre-trained models on several facial classification and regression tasks. More importantly, using ResNet as the unified backbone, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks.", + "bbox": [ + 496, + 621, + 857, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 801, + 643, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the EU H2020 AI4Media No.951911 project. We thank Zengqun Zhao for his helpful comments on facial expression recognition.", + "bbox": [ + 496, + 824, + 857, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2088", + "bbox": [ + 483, + 890, + 513, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 125, + 200, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 2, 3, 5", + "[2] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM International Conference on Multimodal Interaction, page 279-283, New York, NY, USA, 2016. Association for Computing Machinery. 5, 6", + "[3] Adrian Bulat, Shiyang Cheng, Jing Yang, Andrew Garbett, Enrique Sanchez, and Georgios Tzimiropoulos. Pre-training strategies and datasets for facial representation learning. In Computer Vision - ECCV 2022, pages 107-125, Cham, 2022. Springer Nature Switzerland. 1, 3, 5, 6, 7", + "[4] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. Marlin: Masked autoencoder for facial video representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1493-1504, 2023. 3", + "[5] Jiajiong Cao, Yingming Li, and Zhongfei Zhang. Partially shared multi-task convolutional neural network with local constraint for face attribute learning. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4290-4299, 2018. 1", + "[6] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6, 7", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020, pages 213-229, Cham, 2020. Springer International Publishing. 2, 4", + "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 5, 6, 7", + "[9] Jia-Ren Chang, Yong-Sheng Chen, and Wei-Chen Chiu. Learning facial representations from the cycle-consistency of face. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9660-9669, 2021. 3", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2, 6", + "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 1, 2" + ], + "bbox": [ + 114, + 147, + 472, + 864 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1", + "[13] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875. Curran Associates, Inc., 2021. 2, 4", + "[14] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9877-9886, 2021. 3, 6", + "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2", + "[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In Computer Vision - ECCV 2022, pages 247-264, Cham, 2022. Springer Nature Switzerland. 4", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7", + "[18] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 1", + "[19] Ian J Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, et al. Challenges in representation learning: A report on three machine learning contests. In Neural Information Processing: 20th International Conference, ICONIP 2013, Daegu, Korea, November 3-7, 2013. Proceedings, Part III 20, pages 117-124. Springer, 2013. 2", + "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2, 3, 4, 5, 6, 7, 8", + "[21] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 6, 7" + ], + "bbox": [ + 501, + 127, + 856, + 864 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2089", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 1, 2", + "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 1, 2, 3", + "[24] Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Philip Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR 2019. ICLR, 2019. 1", + "[25] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14431-14440, 2022. 5, 6, 7", + "[26] Yangyu Huang, Hao Yang, Chong Li, Jongyoo Kim, and Fangyun Wei. Adnet: Leveraging error-bias towards normal direction in face alignment. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3060-3070, 2021. 6, 7", + "[27] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. In Advances in Neural Information Processing Systems, pages 4016-4027. Curran Associates, Inc., 2018. 3", + "[28] X. Ji, A. Vedaldi, and J. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9864-9873, 2019. 1", + "[29] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[30] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Advances in Neural Information Processing Systems, pages 18661-18673. Curran Associates, Inc., 2020. 1", + "[31] Abhinav Kumar, Tim K. Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8233-8243, 2020. 1", + "[32] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via c-f labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 6", + "[33] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In 2022 IEEE/CVF Conference" + ], + "bbox": [ + 114, + 126, + 472, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), pages 4166-4175, 2022. 7", + "[34] Jianshu Li, Fang Zhao, Jiashi Feng, Sujoy Roy, Shuicheng Yan, and Terence Sim. Landmark free face attribute prediction. IEEE Transactions on Image Processing, 27(9):4651-4662, 2018. 6", + "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2", + "[36] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 1, 2", + "[37] Jinpeng Li, Haibo Jin, Shengcai Liao, Ling Shao, and Pheng-Ann Heng. Repformer: Refinement pyramid transformer for robust facial landmark detection. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 1088–1094. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 7", + "[38] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2584-2593, 2017. 2, 5, 6", + "[39] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Self-supervised representation learning from videos for facial action unit detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10916-10925, 2019. 3", + "[40] Yong Li, Jiabei Zeng, and Shiguang Shan. Learning representations for facial actions from unlabeled videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):302-317, 2022. 3", + "[41] Yuanyuan Liu, Wenbin Wang, Yibing Zhan, Shaoze Feng, Kejun Liu, and Zhe Chen. Pose-disentangled contrastive learning for self-supervised facial representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9717-9728, 2023. 1, 2, 3, 5, 6, 7", + "[42] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, 2015. 2, 5, 6", + "[43] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1557-1567, 2023. 3", + "[44] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2022. 6", + "[45] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Ma-hoor. Affectnet: A database for facial expression, valence," + ], + "bbox": [ + 501, + 127, + 856, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2090", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 6, 7", + "[46] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-bert: Bert-based facial micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1492, 2023. 3", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[48] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 2, 5, 6", + "[49] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. A semi-automatic methodology for facial landmark annotation. In 2013 IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 896-903, 2013.", + "[50] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: Database and results. Image and vision computing, 47:3-18, 2016. 2, 5, 6", + "[51] Ankit Kumar Sharma and Hassan Foroosh. Slim-cnn: A light-weight cnn for face attribute prediction. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 329-335, 2020. 6", + "[52] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11911-11920, 2021. 6", + "[53] Yuxuan Shu, Xiao Gu, Guang-Zhong Yang, and Benny P L Lo. Revisiting self-supervised contrastive learning for facial expression recognition. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 3", + "[54] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2132-2141, 2023. 1, 2, 3", + "[55] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 1", + "[56] James Thewlis, Samuel Albanie, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of landmarks by descriptor vector exchange. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6360-6370, 2019. 3" + ], + "bbox": [ + 114, + 126, + 470, + 864 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Hao Wang, Min Li, Yangyang Song, Youjian Zhang, and Liying Chi. Ucol: Unsupervised learning of discriminative facial representations via uncertainty-aware contrast. Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):2510-2518, 2023. 3", + "[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 1", + "[59] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Liang Liu, Yabiao Wang, and Chengjie Wang. Toward high quality facial representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, page 5048-5058, New York, NY, USA, 2023. Association for Computing Machinery. 1, 3, 6, 7", + "[60] Wenyan Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2138, 2018. 2, 5, 6", + "[61] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4042-4051, 2022. 3, 7", + "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1", + "[63] Jing Yang, Qingshan Liu, and Kaihua Zhang. Stacked hourglass network for robust facial landmark localisation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2025-2033, 2017. 2, 7", + "[64] Sejong Yang, Subin Jeon, Seonghyeon Nam, and Seon Joo Kim. Dense interspecies face embedding. In Advances in Neural Information Processing Systems, pages 33275-33288. Curran Associates, Inc., 2022. 3", + "[65] M. Ye, X. Zhang, P. C. Yuen, and S. Chang. Unsupervised embedding learning via invariant and spreading instance feature. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6203-6212, 2019. 1", + "[66] Lijun Yin, Xiaozhou Wei, Yi Sun, Jun Wang, and Matthew J Rosato. A 3d facial expression database for facial behavior research. In 7th international conference on automatic face and gesture recognition (FGR06), pages 211-216. IEEE, 2006. 3", + "[67] Xiang Zhang, Taoyue Wang, Xiaotian Li, Huiyuan Yang, and Lijun Yin. Weakly-supervised text-driven contrastive learning for facial behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20751-20762, 2023. 6", + "[68] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object land" + ], + "bbox": [ + 501, + 127, + 856, + 866 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2091", + "bbox": [ + 483, + 891, + 511, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "marks as structural representations. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3", + "[69] Yuhang Zhang, Chengrui Wang, and Weihong Deng. Relative uncertainty learning for facial expression recognition. In Advances in Neural Information Processing Systems, pages 17616-17627. Curran Associates, Inc., 2021. 6", + "[70] Yuhang Zhang, Chengrui Wang, Xu Ling, and Weihong Deng. Learn from all: Erasing attention consistency for noisy label facial expression recognition. In Computer Vision - ECCV 2022, pages 418-434, Cham, 2022. Springer Nature Switzerland. 6, 7", + "[71] Kaili Zhao, Wen-Sheng Chu, and Honggang Zhang. Deep region and multi-label learning for facial action unit detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3399, 2016. 1", + "[72] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18676-18688, 2022. 1, 3, 5, 6, 7", + "[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 4", + "[74] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15475-15484, 2023. 6, 7", + "[75] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 146-155, 2016. 3" + ], + "bbox": [ + 114, + 126, + 472, + 578 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2092", + "bbox": [ + 483, + 890, + 514, + 901 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_model.json b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee4161d6ac6a4cd78fc44a01660a7173ad9cf744 --- /dev/null +++ b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_model.json @@ -0,0 +1,2446 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.26, + 0.044, + 0.783, + 0.085 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.162, + 0.847, + 0.183 + ], + "angle": 0, + "content": "Self-Supervised Facial Representation Learning with Facial Region Awareness" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.209, + 0.616, + 0.224 + ], + "angle": 0, + "content": "Zheng Gao Ioannis Patras" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.241, + 0.737, + 0.257 + ], + "angle": 0, + "content": "Queen Mary University of London, Mile End Road, London, E1 4NS" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.259, + 0.597, + 0.273 + ], + "angle": 0, + "content": "{z.gao, i.patras}@qmul.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.305, + 0.329, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.334, + 0.475, + 0.765 + ], + "angle": 0, + "content": "Self-supervised pre-training has been proved to be effective in learning transferable representations that benefit various visual tasks. This paper asks this question: can self-supervised pre-training learn general facial representations for various facial analysis tasks? Recent efforts toward this goal are limited to treating each face image as a whole, i.e., learning consistent facial representations at the image-level, which overlooks the \"consistency of local facial representations\" (i.e., facial regions like eyes, nose, etc). In this work, we make a first attempt to propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). Specifically, we explicitly enforce the consistency of facial regions by matching the local facial representations across views, which are extracted with learned heatmaps highlighting the facial regions. Inspired by the mask prediction in supervised semantic segmentation, we obtain the heatmaps via cosine similarity between the per-pixel projection of feature maps and \"facial mask embeddings\" computed from learnable positional embeddings, which leverage the attention mechanism to globally look up the facial image for facial regions. To learn such heatmaps, we formulate the learning of facial mask embeddings as a deep clustering problem by assigning the pixel features from the feature maps to them. The transfer learning results on facial classification and regression tasks show that our FRA outperforms previous pre-trained models and more importantly, using ResNet as the unified backbone for various tasks, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.788, + 0.234, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.811, + 0.472, + 0.868 + ], + "angle": 0, + "content": "Human face understanding is an important and challenging topic in computer vision [41, 72] and supervised learning algorithms have shown promising results on a wide range of facial analysis tasks recently [5, 31, 55, 71]. Despite" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.307, + 0.859, + 0.334 + ], + "angle": 0, + "content": "the impressive progress, these methods require large-scale well-annotated training data, which is expensive to collect." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.86, + 0.656 + ], + "angle": 0, + "content": "Recent works in self-supervised representation learning for visual images have shown that self-supervised pretraining is effective in improving the performance on various downstream tasks such as image classification, object detection and segmentation as it can learn general representations from unlabeled data that could be transferred to downstream visual tasks, especially tasks with limited labeled data [18, 24, 28, 30, 58, 62, 65]. Among them, instance discrimination (including contrastive learning [10, 22, 36] and non-contrastive learning [11, 20] paradigms) has been shown to be effective in learning generalizable self-supervised features. Instance discrimination aims to learn view-invariant representations by matching the global representations between the augmented views generated by aggressive image augmentations, i.e., the image-level representations of the augmented views should be similar [10-12, 20, 22, 36]. Another self-supervised learning paradigm, masked image modeling (MIM) [23, 54] learns visual representations by reconstructing image content from a masked image, achieving excellent performance in full model fine-tuning. This leads to the question: can self-supervised pre-training learn general facial representations which benefit downstream facial analysis tasks?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.861, + 0.868 + ], + "angle": 0, + "content": "Several attempts have been made to learn general facial representations for facial analysis tasks [3, 41, 72]. For example, Bulat et al. [3] directly applies the contrastive objective to facial features. FaRL [72] and MCF [59] combine contrastive learning and mask image modeling [23]. PCL [41] proposes to disentangle the pose-related and pose-unrelated features, achieving strong performance on both pose-related (regression) and pose-unrelated (classification) tasks. However, it runs the model forward and backward three times for each training step, which is time-consuming. Despite different techniques, these methods commonly treats each face image as a whole to learn consistent global representations at the image-level and overlook the \"spatial consistency of local representations\", i.e., local facial regions (e.g., eyes, nose and mouth) should also be" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.513, + 0.902 + ], + "angle": 0, + "content": "2081" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.127, + 0.471, + 0.182 + ], + "angle": 0, + "content": "similar across the augmented views, thus limiting the generalization to downstream tasks. This brings us to the focus: learning consistent global and local representations for facial representation learning." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.186, + 0.474, + 0.365 + ], + "angle": 0, + "content": "We argue that in order to learn consistent local representations, the model needs to look into facial regions. Towards that goal, we predict a set of heatmaps highlighting different facial regions by leveraging learnable positional embeddings as facial queries (the feature maps as keys and values) to look up the facial image globally for facial regions, which is inspired by the mask prediction in supervised segmentation [13]. For visual images, the attention mechanism of Transformer allows the learnable positional embeddings to serve as object queries for visual pattern look-up [7, 13]. In our case (facial images), the learnable positional embeddings can be used as facial queries for facial regions (see the visualization in the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.37, + 0.474, + 0.868 + ], + "angle": 0, + "content": "In this work, taking the consistency of facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns general facial representations by enforcing consistent global and local facial representations, based on a popular instance discrimination baseline BYOL [20] for its simplicity. Specifically, we learn consistent local facial representations by match them across augmented views, which are extracted by aggregating the feature maps using learned heatmaps highlighting the facial regions as weights. Inspired by the mask prediction in MaskFormer [13], we produce the heatmaps from a set of learnable positional embeddings, which are used as facial queries to look up the facial image for facial regions. A Transformer decoder takes as input the feature maps from the encoder and the learnable positional embeddings to output a set of \"facial mask embeddings\", each associated with a facial region. The facial mask embeddings are used to compute cosine similarity with the per-pixel projection of feature maps to produce the heatmaps. In addition, we enforce the consistency of global representations across views simultaneously so that the image-level information is preserved. In order to learn the heatmaps (facial mask embeddings), inspired by deep clustering [8] that learns to assign samples to clusters, we treat the facial mask embeddings as clusters and learn to assign pixel features from the feature maps to them. Specifically, we align the per-pixel cluster assignments of each pixel feature over the facial region clusters between the online and momentum network for the same augmented view (i.e., each pixel feature should have similar similarity distribution over the facial mask embeddings between the momentum teacher and online student). In contrast to supervised segmentation that directly uses ground-truth masks to supervise the learning of the masks (heatmaps) with a per-pixel binary mask loss, we formulate the learning of heatmaps as a deep cluster" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.857, + 0.153 + ], + "angle": 0, + "content": "ing [8] problem that learns to assign pixel features to clusters (facial mask embeddings) in a self-supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.155, + 0.816, + 0.168 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.17, + 0.858, + 0.239 + ], + "angle": 0, + "content": "- Taking into the consistency of local facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns consistent global and local facial representations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.239, + 0.857, + 0.266 + ], + "angle": 0, + "content": "- We show that the learned heatmaps can roughly discover facial regions in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.267, + 0.858, + 0.362 + ], + "angle": 0, + "content": "- In previous works, different backbones are adopted for different facial analysis tasks (e.g., in face alignment the common backbone is hourglass network [63] while in facial expression recognition ResNet [21] is the popular backbone). In this work, our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.363, + 0.858, + 0.487 + ], + "angle": 0, + "content": "- Our FRA outperforms existing self-supervised pretraining approaches (e.g., BYOL [20] and PCL [41]) on facial classification (i.e., facial expression recognition [19, 38] and facial attribute recognition [42]) and regression (i.e., face alignment [48-50, 60]) tasks. More importantly, our FRA achieves comparable (e.g., face alignment) or even better performance (e.g., facial expression recognition) compared with SOTA methods in the corresponding facial analysis tasks." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.17, + 0.858, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.502, + 0.625, + 0.516 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.526, + 0.746, + 0.541 + ], + "angle": 0, + "content": "2.1. Visual representation learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.859, + 0.783 + ], + "angle": 0, + "content": "As one of the main paradigms for self-supervised pretraining, instance discrimination learns representations by treating an image as a whole and enforcing the consistency of global representations at the image-level across augmented views. Generally, instance discrimination includes two paradigms: contrastive learning [10, 22, 36] and non-contrastive learning [11, 20]. Contrastive learning considers each image and its transformations as a distinct class, i.e., \"positive\" samples are pulled together while \"negative\" samples are pushed apart in the latent space. Unlike contrastive learning that relies on negative samples to avoid collapse, non-contrastive learning directly maximizes the similarity of the global representations between the augmented views without involving negative samples based on techniques like stop-gradient [11] and predictor [20]. Further works perform visual-language pre-training by applying contrastive objective to image-text pairs [29, 35, 47]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.859, + 0.867 + ], + "angle": 0, + "content": "Another line of work, masked image modeling (MIM) learns visual representations by reconstructing image content from a masked image [1, 23, 54], which is inspired by the masked language modeling in NLP [15]. In contrast to instance discrimination, MIM achieves strong full model fine-tuning performance with Vision Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.516, + 0.902 + ], + "angle": 0, + "content": "2082" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.127, + 0.471, + 0.168 + ], + "angle": 0, + "content": "pre-trained for enough epochs. However, these works suffer from poor linear separability and are less data-efficient than instance discrimination in few-shot scenarios [1]." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.178, + 0.357, + 0.193 + ], + "angle": 0, + "content": "2.2. Facial representation learning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.199, + 0.474, + 0.712 + ], + "angle": 0, + "content": "Recent works on facial analysis explore self-supervised learning for several face-related tasks, such as facial expression recognition [9, 53], face recognition [9, 57], facial micro-expression recognition [46], AU detection [39, 40], face alignment (facial landmark detection) [14, 64], etc. However, these methods are task-specific, i.e., tailored for a specific facial task and thus lack the ability to generalize to various facial analysis tasks [41]. Further efforts [3, 41, 72] focus on learning general facial representations with contrastive learning and mask image modeling [23, 54]. Bulat et al. [3] directly apply the contrastive objective to augmented views of the same face image, showing that general facial representation learned from pre-training can benefit various facial analysis tasks. FaRL [72] performs pre-training in a visual-linguistic manner by employing image-text contrastive learning and masked image modeling. MCF [59] leverages image-level contrastive learning and masked image modeling, along with the knowledge distilled from external ImageNet pre-trained model for facial representation learning. PCL [41] argues that directly applying the contrastive objective to face images overlooks the variances of facial poses and thus leads to pose-invariant representations, limiting the performance on pose-related tasks [66, 75]. Therefore, PCL [41] disentangles the pose-related and pose-unrelated features and then performs contrastive learning on these features, achieving strong performance on both pose-related and pose-unrelated facial analysis tasks. Despite the success, it performs forward and backward three times for each input image, which brings significant increase on computational cost. These works are commonly limited by instance discrimination and overlook the consistency of local facial regions. In contrast, inspired by supervised semantic segmentation, we learn consistent global and local facial representations by learning a set of heatmaps indicating facial regions from learnable positional embeddings, which leverage the attention mechanism to look up facial image globally for facial regions." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.72, + 0.308, + 0.735 + ], + "angle": 0, + "content": "2.3. Facial region discovery" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.742, + 0.472, + 0.868 + ], + "angle": 0, + "content": "There are some approaches leveraging facial region (landmark) discovery for facial analysis [27, 43, 61]. Some focus on landmark detection by either learning a heatmap for each landmark via image reconstruction [27, 68], or performing pixel-level matching with an equivariance loss [56, 68]. Despite different techniques, these methods are task-specific, i.e., landmark detection with discovery of local information, while our method is task-agnostic, i.e., learn general facial representations for various tasks by preserving" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.858, + 0.32 + ], + "angle": 0, + "content": "global and local information in a image, region and pixel-level contrastive manner. MARLIN [4] applies masked image modeling to learn general representations for facial videos by utilizing an external face parsing algorithm to discover the facial regions (e.g., eyes, nose and mouth), which are used to guide the masking for the masked autoencoder. A closely related work SLPT [61] leverages the attention mechanism to estimate facial landmarks from initial facial landmarks estimates of the mean face through supervised learning. These works commonly rely on external supervisory signal, whether it is from ground-truths or additional algorithms. In contrast, we learn to discover the facial regions in an end-to-end self-supervised manner for facial image representation learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.623, + 0.35 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.601, + 0.37 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.859, + 0.531 + ], + "angle": 0, + "content": "The overview of the proposed FRA is shown in Fig. 1. The goal is to learn consistent global and local facial representations. Toward this goal, we propose two objectives: pixel-level semantic relation and image/region-level semantic consistency. Semantic relation aligns the per-pixel cluster assignments of each pixel feature over the facial mask embeddings between the online and momentum network to learn heatmaps for facial regions (Sec. 3.2) while semantic consistency directly matches the global and local facial representations across augmented views (Sec. 3.3) with the learned heatmaps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.657, + 0.554 + ], + "angle": 0, + "content": "3.2. Semantic relation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.859, + 0.742 + ], + "angle": 0, + "content": "As shown in Fig. 1, our method adopts the Siamese structure of BYOL [20], a popular self-supervised pretraining baseline based on instance discrimination. Following BYOL [20], we employ two branches: the online network parameterized by \\(\\theta\\) and the momentum network parameterized by \\(\\xi\\). The online network \\(\\theta\\) consists of an encoder \\(E_{\\theta}\\), a global projector \\(H_{\\theta}^{g}\\) and a local projector \\(H_{\\theta}^{l}\\). The momentum network \\(\\xi\\) has the same architecture with the online network, except \\(\\xi\\) is updated with an exponential moving average of \\(\\theta\\). As in BYOL [20], we also adopt additional predictors \\(G_{\\theta}^{g}\\) and \\(G_{\\theta}^{l}\\) on top of the projectors in the online network. Note that this is omitted for brevity in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.859, + 0.868 + ], + "angle": 0, + "content": "Given an input image \\(\\mathbf{x}\\), two random augmentations are applied to generate two augmented views \\(\\mathbf{x}_1 = \\mathcal{T}_1(\\mathbf{x})\\) and \\(\\mathbf{x}_2 = \\mathcal{T}_2(\\mathbf{x})\\), following BYOL [20]. Each augmented view \\(\\mathbf{x}_i \\in \\{\\mathbf{x}_1, \\mathbf{x}_2\\}\\) is fed into an encoder \\(E\\) to obtain a feature map \\(\\mathbf{F}_i \\in \\mathbb{R}^{C \\times H \\times W}\\) (before global average pooling), where \\(C, H, W\\) are the number of channels, height and width of \\(\\mathbf{F}_i\\) and a latent representation \\(\\mathbf{h}_i \\in \\{\\mathbf{h}_1, \\mathbf{h}_2\\}\\) (after global average pooling), i.e., \\(\\mathbf{h}_1 = E_{\\theta}(\\mathbf{x}_1)\\) and \\(\\mathbf{h}_2 = E_{\\xi}(\\mathbf{x}_2)\\). Then each latent representation \\(\\mathbf{h}_i\\) is transformed" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2083" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.124, + 0.82, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.349, + 0.862, + 0.44 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed FRA framework. \\(\\odot\\) denotes cosine similarity. For each input image \\(\\mathbf{x}\\), its augmented views \\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\) are passed into two network branches to produce the global embeddings \\(\\mathbf{z}_1\\) and \\(\\mathbf{z}_2\\). In addition, we produce a set of heatmaps \\(\\mathbf{M}_1\\) and \\(\\mathbf{M}_2\\) indicating the local facial regions, via the correlation between the pixel features and \"facial mask embeddings\" computed from a set of learnable positional embeddings. Then we aggregate the feature map to obtain the local facial embeddings \\(\\{\\mathbf{z}_1^m\\}\\) and \\(\\{\\mathbf{z}_2^m\\}\\). The semantic consistency loss is applied to global embeddings and facial embeddings to maximize the similarity across augmented views. To learn such heatmaps, i.e., facial mask embeddings, we treat the facial mask embeddings as facial region clusters and propose a semantic relation loss to align the cluster assignments of each pixel feature over the facial region clusters between the online and momentum network." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.461, + 0.473, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.611, + 0.473, + 0.639 + ], + "angle": 0, + "content": "Figure 2. Generation of heatmaps using learnable positional embeddings as facial queries and the feature maps as keys and values." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.662, + 0.472, + 0.703 + ], + "angle": 0, + "content": "by a global projector \\(H^g\\) to produce a global embedding \\(\\mathbf{z}_i\\in \\{\\mathbf{z}_1,\\mathbf{z}_2\\}\\) of dimension \\(\\mathbf{z}_i\\in \\mathbb{R}^D\\) , i.e., \\(\\mathbf{z}_1 = H_\\theta^g (\\mathbf{h}_1)\\) and \\(\\mathbf{z}_2 = H_{\\xi}^{g}(\\mathbf{h}_{2})\\)" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.703, + 0.474, + 0.841 + ], + "angle": 0, + "content": "Then we obtain a set of heatmaps \\(\\mathbf{M}_i\\in \\{\\mathbf{M}_1,\\mathbf{M}_2\\}\\) highlighting the facial regions from the feature map \\(\\mathbf{F}_i\\) for each view, which is inspired by mask classification-based supervised segmentation [7, 13] that leverages attention mechanism to look up visual patterns globally. First the local projector (e.g., \\(H_{\\theta}^{l}\\)) is applied to project the pixel features of \\(\\mathbf{F}_i\\) in a pixel-wise manner, mapping it to \\(D\\) dimensions to get the dense feature map \\(\\mathbf{F}_i^{\\mathrm{dense}}\\in \\mathbb{R}^{D\\times H\\times W}\\). Take view \\(\\mathbf{x}_1\\) as an example, the projected feature map can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.851, + 0.473, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {1} ^ {\\text {d e n s e}} [ *, u, v ] = H _ {\\theta} ^ {l} \\left(\\mathbf {F} _ {1} [ *, u, v ]\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.86, + 0.756 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_1[*,u,v]\\in \\mathbb{R}^C\\) is the pixel feature at the \\((u,v)\\) -th grid of \\(\\mathbf{F}_1\\). Then as shown in Fig. 2, inspired by supervised segmentation [13], we use a Transformer decoder followed by a MLP, which takes as input the feature map \\(\\mathbf{F}_i\\) and \\(N\\) learnable positional embeddings (i.e., facial queries for looking up the facial image globally for facial regions) to predict \\(N\\) \"facial mask embeddings\" \\(\\mathbf{Q}_i\\in \\mathbb{R}^{N\\times D}\\) of dimension \\(D\\), where each row associated with a facial region. Next, we compute the cosine similarity between facial mask embeddings \\(\\mathbf{Q}_i\\) and dense feature map \\(\\mathbf{F}_i^{\\mathrm{dense}}\\) along the channel dimension, yielding per-pixel cluster assignments \\(\\mathbf{S}_i\\in \\mathbb{R}^{N\\times H\\times W}\\), where \\(\\mathbf{S}_i[*,u,v]\\) denotes the relation between the dense pixel feature \\(\\mathbf{F}_1^{\\mathrm{dense}}[*,u,v]\\) and facial mask embeddings \\(\\mathbf{Q}_i\\). Finally, we normalize \\(\\mathbf{S}_i\\) along the channel dimension with a softmax operation to encourage each channel to capture a different pattern, obtaining \\(N\\) heatmaps \\(\\mathbf{M}_i\\in \\mathbb{R}^{N\\times H\\times W}\\) where each vector at location \\((u,v)\\) is a probabilistic similarity distribution (i.e., normalized per-pixel cluster assignments) \\(\\mathbf{s}_1^{u,v}\\) between \\(\\mathbf{F}_i^{\\mathrm{dense}}[*,u,v]\\) and \\(\\mathbf{Q}_i\\). Note that \\(\\mathbf{M}_i\\) is a set of heatmaps where each channel of \\(\\mathbf{M}_i\\) represents a 2D heatmap \\(\\mathbf{M}_i^{(m)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.859, + 0.852 + ], + "angle": 0, + "content": "To learn such heatmaps, i.e., facial mask embeddings, inspired by deep clustering [8], we treat the facial mask embeddings as facial region clusters and align the per-pixel cluster assignments of each pixel feature over these clusters between the online and momentum network for the same augmented view, using the momentum network as momentum teacher [16, 73] to provide reliable target." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.853, + 0.859, + 0.867 + ], + "angle": 0, + "content": "Following BYOL [20], we pass both augmented views" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2084" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.127, + 0.473, + 0.21 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\) through the online and momentum network. Take \\(\\mathbf{x}_1\\) as an example, the online network \\(\\theta\\) outputs the normalized per-pixel cluster assignments \\(\\mathbf{s}_1^{u,v}\\) and the momentum network outputs normalized assignments \\(\\widehat{\\mathbf{s}}_1^{u,v}\\) for view \\(\\mathbf{x}_1\\). Then we learn \\(\\mathbf{s}_1^{u,v}\\) using \\(\\widehat{\\mathbf{s}}_1^{u,v}\\) as guidance based on the following cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.22, + 0.473, + 0.259 + ], + "angle": 0, + "content": "\\[\nC E \\left(\\mathbf {s} _ {1} ^ {u, v}, \\widehat {\\mathbf {s}} _ {1} ^ {u, v}\\right) = - \\sum_ {m = 1} ^ {N} \\widehat {\\mathbf {s}} _ {1} ^ {u, v} [ m ] \\log \\mathbf {s} _ {1} ^ {u, v} [ m ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.269, + 0.473, + 0.297 + ], + "angle": 0, + "content": "For both augmented views, we define the symmetrized semantic relation objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.307, + 0.473, + 0.341 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r}} = \\frac {1}{H W} \\sum_ {u, v} \\left(C E \\left(\\mathbf {s} _ {1} ^ {u, v}, \\widehat {\\mathbf {s}} _ {1} ^ {u, v}\\right) + C E \\left(\\mathbf {s} _ {2} ^ {u, v}, \\widehat {\\mathbf {s}} _ {2} ^ {u, v}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.352, + 0.473, + 0.436 + ], + "angle": 0, + "content": "where \\(CE(\\mathbf{s}_2^{u,v},\\widehat{\\mathbf{s}}_2^{u,v})\\) is the cross-entropy loss for view \\(\\mathbf{x}_2\\). We apply the Sinkhorn-Knopp normalization to the target assignments from the momentum network following [8] to avoid collapse and the mean entropy maximization (ME-MAX) regularizer [1] to maximize the entropy of the prediction to encourage full use of the clusters." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.446, + 0.296, + 0.461 + ], + "angle": 0, + "content": "3.3. Semantic consistency" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.467, + 0.473, + 0.524 + ], + "angle": 0, + "content": "In this section, we enforce the consistency of global embeddings and local facial embeddings. With the learned heatmaps \\(\\mathbf{M}_i\\), we generate the latent representations for the local facial regions through weighted average pooling:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.535, + 0.472, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {h} _ {i} ^ {m} = \\mathbf {M} _ {i} ^ {(m)} \\otimes \\mathbf {F} _ {i} \\\\ = \\frac {1}{\\sum_ {u , v} \\mathbf {M} _ {i} [ m , u , v ]} \\sum_ {u, v} \\mathbf {M} _ {i} [ m, u, v ] \\mathbf {F} _ {i} [ *, u, v ], \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.598, + 0.473, + 0.687 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes channel-wise weighted average pooling, \\(\\mathbf{M}_i^{(m)}\\) is the \\(m\\)-th channel (heatmap) of \\(\\mathbf{M}_i\\) and \\(\\mathbf{h}_i^m \\in \\mathbb{R}^C\\) is the corresponding latent representation produced with \\(\\mathbf{M}_i^{(m)}\\). The facial embeddings \\(\\{\\mathbf{z}_1^m : \\mathbf{z}_1^m \\in \\mathbb{R}^D\\}_{m=1}^N\\) and \\(\\{\\mathbf{z}_2^m : \\mathbf{z}_2^m \\in \\mathbb{R}^D\\}_{m=1}^N\\) are obtained accordingly with the local projector \\(H_\\theta^l\\) and \\(H_\\xi^l\\):" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.698, + 0.472, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {z} _ {1} ^ {m} = H _ {\\theta} ^ {l} \\left(\\mathbf {h} _ {1} ^ {m}\\right), \\\\ \\mathbf {z} _ {2} ^ {m} = H _ {\\xi} ^ {l} \\left(\\mathbf {h} _ {2} ^ {m}\\right). \\end{array} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.739, + 0.472, + 0.782 + ], + "angle": 0, + "content": "We then match the global embeddings and local facial embeddings across views using the negative cosine similarity in BYOL [20]:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.793, + 0.472, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}\\right) = - \\left(\\lambda_ {\\mathrm {c}} \\times f _ {\\mathrm {s}} \\left(G _ {\\theta} ^ {g} \\left(\\mathbf {z} _ {1}\\right), \\mathbf {z} _ {2}\\right) + \\right. \\\\ + \\left(1 - \\lambda_ {\\mathrm {c}}\\right) \\times \\frac {1}{N} \\sum_ {m = 1} ^ {N} f _ {\\mathrm {s}} \\left(G _ {\\theta} ^ {l} \\left(\\mathbf {z} _ {1} ^ {m}\\right), \\mathbf {z} _ {2} ^ {m}\\right)), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.125, + 0.86, + 0.241 + ], + "angle": 0, + "content": "where \\(f_{\\mathrm{s}}(\\mathbf{u},\\mathbf{v}) = \\frac{\\mathbf{u}^{\\top}\\mathbf{v}}{\\|\\mathbf{u}\\|_{2}\\|\\mathbf{v}\\|_{2}}\\) denotes the cosine similarity between the vectors \\(\\mathbf{u}\\) and \\(\\mathbf{v}\\), \\(\\lambda_{\\mathrm{c}}\\) is the loss weight, \\(G_{\\theta}^{g}\\) and \\(G_{\\theta}^{l}\\) are the predictors on top of the projectors \\(H_{\\theta}^{g}\\) and \\(H_{\\theta}^{l}\\), respectively. Following BYOL [20], we symmetrize the loss \\(\\mathcal{L}_{\\mathrm{sim}}(\\mathbf{z}_1,\\mathbf{z}_2)\\) defined in Eq. (6) by passing \\(\\mathbf{x}_1\\) through the momentum network \\(\\xi\\) and \\(\\mathbf{x}_2\\) through the online network \\(\\theta\\) to compute \\(\\mathcal{L}_{\\mathrm{sim}}(\\mathbf{z}_2,\\mathbf{z}_1)\\). The semantic consistency objective can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.252, + 0.858, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}\\right) + \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {2}, \\mathbf {z} _ {1}\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.279, + 0.654, + 0.294 + ], + "angle": 0, + "content": "3.4. Overall objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.859, + 0.343 + ], + "angle": 0, + "content": "We jointly optimize the semantic relation objective Eq. (3) and the semantic consistency objective Eq. (7), leading to the following overall objective:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.354, + 0.858, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.38, + 0.814, + 0.395 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{r}}\\) is the loss weight for balancing \\(\\mathcal{L}_{\\mathrm{c}}\\) and \\(\\mathcal{L}_{\\mathrm{r}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.409, + 0.62, + 0.424 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.432, + 0.677, + 0.447 + ], + "angle": 0, + "content": "4.1. Experimental setups" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.697, + 0.467 + ], + "angle": 0, + "content": "4.1.1 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.859, + 0.56 + ], + "angle": 0, + "content": "We use the same augmentation strategy as in [20, 25]. The number of heatmaps \\(N\\) is set to 8 empirically. The loss weight \\(\\lambda_{\\mathrm{c}}\\) and \\(\\lambda_{\\mathrm{r}}\\) are set to \\(0.5 / 0.1\\), respectively. For fair comparisons, the other hyper-parameters are kept the same as BYOL [20] in all experiments. The architecture and pretraining details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.609, + 0.59 + ], + "angle": 0, + "content": "4.1.2 Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.859, + 0.751 + ], + "angle": 0, + "content": "Our baselines are self-supervised pre-training approaches for visual images (e.g., BYOL [20] and LEWEL [25]), and pre-training approaches for facial images (e.g., Bulat et al. [3] and PCL [41]). Note that SwAV [8] is equivalent to Bulat et al. [3]. As we adopt BYOL [20] as the pretraining backbone, we compare our FRA with BYOL [20] in all experiments. We also compare our FRA with another pre-training method LEWEL [25], which learns consistent local representations for visual images. Moreover, we perform comparisons with SOTA methods in the corresponding downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.762, + 0.679, + 0.777 + ], + "angle": 0, + "content": "4.2. Evaluation protocols" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.859, + 0.868 + ], + "angle": 0, + "content": "Following the common practice in previous works [41, 72], we evaluate the transfer performance of the self-supervised pre-trained facial representations on several popular downstream facial analysis tasks: facial expression recognition (FER) [2, 38], facial attribute recognition (FAR) [42] and face alignment (FA) [48-50, 60]. Specifically, we use" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2085" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.124, + 0.859, + 0.151 + ], + "angle": 0, + "content": "Table 1. Comparisons with weakly-supervised pre-trained vision transformer on several downstream facial analysis tasks, including facial expression recognition (AffectNet), facial attribute recognition (CelebA) and face alignment (300W)." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.16, + 0.839, + 0.253 + ], + "angle": 0, + "content": "
MethodArch.Params.Pre-training settingsDownstream performances
DatasetScaleSupervisionAffectNet Acc. ↑CelebA Acc. ↑300W NME ↓
FaRL [72]ViT-B/16 [17]86MLAION-FACE [72]20Mface image + text64.8591.883.08
FRAR50 [21]24MVGGFace2 [6]3.3Mface image66.1692.022.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.273, + 0.471, + 0.311 + ], + "angle": 0, + "content": "Table 2. Comparisons on facial expression recognition. We report the Top-1 accuracy on test set. Text denotes text supervision.†: our reproduction using the official codes." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.321, + 0.474, + 0.584 + ], + "angle": 0, + "content": "
MethodTextFERPlusRAF-DBAffectNet
Supervised
KTN [32]×90.4988.0763.97
RUL [69]×88.7588.9861.43
EAC [70]×90.0590.3565.32
Weakly-Supervised
FaRL [72]†✓88.6288.3164.85
CLEF [67]✓89.7490.0965.66
Self-supervised
MCF [59]†×88.1786.8660.98
Bulat et al. [3, 8]×--60.20
BYOL [20]×89.2589.5365.65
LEWEL [25]×85.6181.8561.20
PCL [41]×85.8785.9260.77
FRA (LP)×78.1373.8957.38
FRA (FT)×89.7889.9566.16
FRA (EAC)×90.6290.7665.85
" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.612, + 0.472, + 0.695 + ], + "angle": 0, + "content": "the pre-trained weights to initialize the backbone of downstream tasks and then learn the backbone and task-specific head (attached to the backbone) jointly. Following [72], we report the performance with linear probe (denoted by \"LP\") and fine-tuning (denoted by \"FT\"). The details of the downstream tasks are described as follows:" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.698, + 0.472, + 0.81 + ], + "angle": 0, + "content": "Facial expression recognition is a multi-class classification task where the goal is to categorize the emotional expressions (e.g., anger, fear and surprise) for a given face image. Three widely-used datasets are adopted: FERPlus [2], RAF-DB [38] and AffectNet [45]. For RAF-DB, we use the basic emotion subset following [32, 41, 70]. For AffectNet, we report the results with 7 emotion classes (i.e., neutral, happy, sad, surprise, fear, anger, disgust) following [32, 70]." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.812, + 0.473, + 0.867 + ], + "angle": 0, + "content": "Facial attribute recognition is a multi-label classification task to predict various attributes (e.g., gender, age and race) of a given face image. We adopt the popular benchmark CelebA [42], which consists of more than 200K face" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.273, + 0.859, + 0.312 + ], + "angle": 0, + "content": "Table 3. Comparisons on CelebA [42] facial attribute recognition. We report the averaged accuracy over all attributes. †: our reproduction using the official codes. *: results cited from [72]." + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.321, + 0.784, + 0.551 + ], + "angle": 0, + "content": "
MethodAcc. (%)
Supervised
DMM [44]91.70
SlimCNN [51]91.24
AFFAIR [34]91.45
Self-supervised
SSPL [52]91.77
Bulat et al. [3, 8]*89.65
SimCLR [10]*91.08
BYOL [20]91.56
LEWEL [25]90.69
PCL [41]91.48
MCF [59]†91.33
FRA (LP)90.86
FRA (FT)92.02
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.573, + 0.857, + 0.601 + ], + "angle": 0, + "content": "images with 40 facial attributes per image. Following [72], we report the averaged accuracy over all attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.859, + 0.712 + ], + "angle": 0, + "content": "Face alignment is a regression task to predict 2D face landmark coordinates on a face image. We use two popular benchmarks: WFLW [60] and 300W [48-50]. Following the common practice [14, 26, 74], we report normalized mean error (NME), failure rate (FR) and AUC. For 300W, we report the results on full test set, common (554 images) and challenge (135 images) splits of the test set following [26, 74]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.721, + 0.858, + 0.75 + ], + "angle": 0, + "content": "4.3. Comparisons with weakly-supervised pretraining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.859, + 0.868 + ], + "angle": 0, + "content": "In Tab. 1, we compare our FRA with SOTA pre-trained Transformer FaRL [72], which is a weakly-supervised model pre-trained on 20M visual-linguistic data (face image and text) with image-text contrastive learning and mask image modeling. We fine-tune both the pre-trained feature backbone and the task-specific head on the corresponding downstream facial analysis task. Our self-supervised FRA with 24M parameter ResNet-50 achieves superior perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.515, + 0.902 + ], + "angle": 0, + "content": "2086" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.249, + 0.124, + 0.722, + 0.138 + ], + "angle": 0, + "content": "Table 4. Comparisons on face alignment. †: our reproduction using the official codes." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.148, + 0.87, + 0.382 + ], + "angle": 0, + "content": "
MethodVenueArch.WFLW300W (NME ↓)
NME ↓FR10% ↓AUC10% ↑FullComm.Chal.
Supervised
SLPT [61][CVPR'22]ResNet [21]4.203.040.5883.202.784.93
DTLD [33][CVPR'22]ResNet [21]4.082.76-2.962.594.50
RePFormer [37][IJCAI'22]ResNet [21]4.11--3.01--
ADNet [26][ICCV'21]Hourglass [63]4.142.720.6022.932.534.58
STAR [74][CVPR'23]Hourglass [63]4.022.320.6052.872.524.32
Self-supervised
MCF [59] (concurrent work)[ACMM'M'23]ViT [17]3.961.400.6092.982.604.51
Bulat et al. [3, 8][ECCV'22]ResNet [21]4.57--3.20--
BYOL [20][NeurIPS'20]ResNet [21]4.292.960.5793.032.664.55
LEWEL [25][CVPR'22]ResNet [21]4.524.500.5633.092.704.71
PCL [41]†[CVPR'23]ResNet [21]4.846.180.5353.352.775.12
FRAOursResNet [21]4.112.530.5912.912.604.46
" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.405, + 0.471, + 0.433 + ], + "angle": 0, + "content": "mance compared with weakly-supervised FaRL [72] with 86M parameter ViT-B/16 and text supervision on all tasks." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.454, + 0.27, + 0.469 + ], + "angle": 0, + "content": "4.4. Transfer learning" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.48, + 0.471, + 0.535 + ], + "angle": 0, + "content": "In this section, we compare our FRA with self-supervised pre-training approaches and SOTA methods in several downstream tasks. Please refer to the supplementary material for setup details." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.575, + 0.347, + 0.589 + ], + "angle": 0, + "content": "4.4.1 Facial expression recognition" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.604, + 0.474, + 0.867 + ], + "angle": 0, + "content": "The results on facial expression recognition are reported in Tab. 2. We observe: (1) With the setting of fine-tuning (FT), our FRA outperforms previous self-supervised pre-training approaches for visual images (e.g., BYOL and LEWEL) and pre-training approaches tailored for facial images (e.g., PCL, MCF). In particular, our FRA using a 24M parameter ResNet-50 surpasses the concurrent work MCF [59] with 86M parameter ViT-B/16 [17]. (2) By simply learning a linear classifier on top of the encoder backbone, our FRA outperforms SOTA facial expression recognition methods with sophisticated designs (e.g., EAC [70]) on AffectNet [45], the largest facial expression recognition dataset. (3) More importantly, by using our pre-trained model to initialize the backbone of SOTA facial expression recognition method EAC [70], our variant \"FRA (EAC)\" consistently improves EAC [70] on all datasets, which suggests \"FRA (EAC)\" outperforms SOTA FER methods and demonstrates the superiority of the proposed self-supervised pre-training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.724, + 0.419 + ], + "angle": 0, + "content": "4.4.2 Facial attribute recognition" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.428, + 0.859, + 0.511 + ], + "angle": 0, + "content": "As shown in Tab. 3, our FRA outperforms both self-supervised pre-training approaches for visual images and pre-training approaches tailored for facial images. The results on facial expression recognition and facial attribute recognition show that our FRA learns better facial representations for facial classification task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.532, + 0.647, + 0.546 + ], + "angle": 0, + "content": "4.4.3 Face alignment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.859, + 0.707 + ], + "angle": 0, + "content": "As shown in Tab. 4, despite SOTA face alignment methods (e.g., ADNet [26] and STAR [74]) commonly rely on hourglass network [63] for feature extraction, which is tailored for regression tasks like landmark detection, our method based on ResNet backbone achieves comparable performance with these SOTA methods (e.g., 2.91 vs. 2.87 on 300W). The results on classification (e.g., facial expression recognition) and regression tasks (e.g., face alignment) show that our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.718, + 0.647, + 0.731 + ], + "angle": 0, + "content": "4.5. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.74, + 0.859, + 0.781 + ], + "angle": 0, + "content": "We pre-train the model on VGGFace2 [6] and then evaluate it on facial expression recognition (RAF-DB) and facial attribute recognition (CelebA), as described in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.72, + 0.815 + ], + "angle": 0, + "content": "4.5.1 Effect of different modules" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.825, + 0.859, + 0.867 + ], + "angle": 0, + "content": "In Tab. 5, we investigate the contributions of the proposed semantic consistency loss (i.e., global consistency of whole face and local consistency of facial regions) and semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2087" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.124, + 0.472, + 0.176 + ], + "angle": 0, + "content": "Table 5. Effect of different modules. GC denotes the global consistency for aligning images, LC denotes the local consistency for aligning facial regions with heatmaps and SR represents semantic relation for aligning pixels and heatmaps." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.185, + 0.462, + 0.284 + ], + "angle": 0, + "content": "
GCLCSRRAF-DB ↑CelebA ↑300W ↓
✓--87.8290.633.56
--✓85.3490.663.25
-✓-87.4690.783.19
✓✓-88.0590.893.26
✓✓✓88.7291.183.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.301, + 0.416, + 0.314 + ], + "angle": 0, + "content": "Table 6. Effect of the number of heatmaps \\( N \\) ." + }, + { + "type": "table", + "bbox": [ + 0.178, + 0.324, + 0.408, + 0.382 + ], + "angle": 0, + "content": "
N83264
RAF-DB88.7288.3688.18
CelebA91.1891.0190.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.398, + 0.471, + 0.424 + ], + "angle": 0, + "content": "Table 7. Effect of the loss weights. Please refer to Tab. 5 for GC, LC and SR." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.435, + 0.468, + 0.533 + ], + "angle": 0, + "content": "
SettingsλcλrRAF-DBCelebA
GC1.0087.8290.63
GC + LC0.5088.0590.89
GC + LC + SR (FRA)0.50.188.7291.18
GC + LC + SR (FRA)0.50.588.4591.04
GC + LC + SR (FRA)0.51.088.0890.46
" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.561, + 0.472, + 0.685 + ], + "angle": 0, + "content": "relation loss to our approach. Note that the global consistency (first row) is BYOL [20]. We have the following observations: (1) The variant using all losses achieves the best results. (2) GC is essential to avoid degeneration on classification task. (3) LC or SR alone benefits regression task (landmark). Altogether, LC and SR improve BYOL [20] (GC) on both classification and regression by capturing spatial/local information, which validates our facial region awareness." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.716, + 0.376, + 0.73 + ], + "angle": 0, + "content": "4.5.2 Effect of the number of heatmaps" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.742, + 0.472, + 0.866 + ], + "angle": 0, + "content": "In Tab. 6, we study the effect of the number of heatmaps. We observe that the best setting is \\( N = 8 \\), which is close to the facial landmarks number 5. This suggests that given enough face images for training, a suitable \\( N \\) can encourage the model to learn face-specific patterns, which helps the transfer learning performance on various facial analysis tasks. Further increasing the number of heatmaps might force the model to look into fine-grained patterns that may not be suitable for facial tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.127, + 0.683, + 0.14 + ], + "angle": 0, + "content": "4.5.3 Effect of loss weights" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.148, + 0.858, + 0.314 + ], + "angle": 0, + "content": "In Tab. 7, we ablate the weights for the semantic consistency loss and semantic relation loss. We find that setting \\(\\lambda_{\\mathrm{c}} = 0.5\\) and \\(\\lambda_{\\mathrm{r}} = 0.1\\) works best. When \\(\\lambda_{\\mathrm{c}} = 1.0\\) and \\(\\lambda_{\\mathrm{r}} = 0\\), only the consistency of global representations is applied, and the model performs relatively worse, which suggests the importance of the consistency of local representations and the semantic relation loss. By using the semantic relation objective, the performance is significantly improved. However, when \\(\\lambda_{\\mathrm{r}}\\) is too high, the performance degrades as the pixel-level consistency between the online and momentum network might affect the capture of image/object-level information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.784, + 0.348 + ], + "angle": 0, + "content": "4.5.4 Effect of Transformer decoder layers" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.368, + 0.859, + 0.406 + ], + "angle": 0, + "content": "Table 8. Effect of Transformer decoder layers. 0 decoder layer represents BYOL [20] where only the consistency of global representation is enforced." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.417, + 0.84, + 0.474 + ], + "angle": 0, + "content": "
# decoder layer0123
RAF-DB87.8288.7289.0189.06
CelebA90.6391.1891.3091.35
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.859, + 0.585 + ], + "angle": 0, + "content": "In Tab. 8, we study the effect of the number of decoder layers used for heatmap prediction. We observe that a single decoder layer is able to produce decent results, showing that a 1-layer decoder is large enough to capture the facial region (landmarks) relations in face images. The performance gain diminishes as the decoder depth increases. By default, we only use 1 decoder layer for fast training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.608, + 0.613 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.859, + 0.788 + ], + "angle": 0, + "content": "In this work, we propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). We learn a set of heatmaps indicating facial regions from learnable positional embeddings, which leverages the attention mechanism to look up facial image globally for facial regions. We show that our FRA outperforms previous pre-trained models on several facial classification and regression tasks. More importantly, using ResNet as the unified backbone, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.645, + 0.817 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.858, + 0.867 + ], + "angle": 0, + "content": "This work was supported by the EU H2020 AI4Media No.951911 project. We thank Zengqun Zhao for his helpful comments on facial expression recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2088" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.125, + 0.201, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.148, + 0.473, + 0.212 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.213, + 0.473, + 0.29 + ], + "angle": 0, + "content": "[2] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM International Conference on Multimodal Interaction, page 279-283, New York, NY, USA, 2016. Association for Computing Machinery. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.292, + 0.473, + 0.355 + ], + "angle": 0, + "content": "[3] Adrian Bulat, Shiyang Cheng, Jing Yang, Andrew Garbett, Enrique Sanchez, and Georgios Tzimiropoulos. Pre-training strategies and datasets for facial representation learning. In Computer Vision - ECCV 2022, pages 107-125, Cham, 2022. Springer Nature Switzerland. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.357, + 0.473, + 0.433 + ], + "angle": 0, + "content": "[4] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. Marlin: Masked autoencoder for facial video representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1493-1504, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.435, + 0.473, + 0.499 + ], + "angle": 0, + "content": "[5] Jiajiong Cao, Yingming Li, and Zhongfei Zhang. Partially shared multi-task convolutional neural network with local constraint for face attribute learning. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4290-4299, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.501, + 0.473, + 0.564 + ], + "angle": 0, + "content": "[6] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.567, + 0.473, + 0.63 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020, pages 213-229, Cham, 2020. Springer International Publishing. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.632, + 0.473, + 0.695 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.698, + 0.473, + 0.749 + ], + "angle": 0, + "content": "[9] Jia-Ren Chang, Yong-Sheng Chen, and Wei-Chen Chiu. Learning facial representations from the cycle-consistency of face. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9660-9669, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.751, + 0.473, + 0.813 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.816, + 0.473, + 0.866 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.148, + 0.473, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.166 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.168, + 0.857, + 0.218 + ], + "angle": 0, + "content": "[13] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875. Curran Associates, Inc., 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.857, + 0.271 + ], + "angle": 0, + "content": "[14] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9877-9886, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.272, + 0.857, + 0.386 + ], + "angle": 0, + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.857, + 0.452 + ], + "angle": 0, + "content": "[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In Computer Vision - ECCV 2022, pages 247-264, Cham, 2022. Springer Nature Switzerland. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.454, + 0.857, + 0.542 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.544, + 0.857, + 0.607 + ], + "angle": 0, + "content": "[18] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.857, + 0.711 + ], + "angle": 0, + "content": "[19] Ian J Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, et al. Challenges in representation learning: A report on three machine learning contests. In Neural Information Processing: 20th International Conference, ICONIP 2013, Daegu, Korea, November 3-7, 2013. Proceedings, Part III 20, pages 117-124. Springer, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.713, + 0.857, + 0.813 + ], + "angle": 0, + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.857, + 0.866 + ], + "angle": 0, + "content": "[21] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2089" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.178 + ], + "angle": 0, + "content": "[22] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.18, + 0.473, + 0.242 + ], + "angle": 0, + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.245, + 0.473, + 0.307 + ], + "angle": 0, + "content": "[24] Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Philip Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR 2019. ICLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.309, + 0.473, + 0.373 + ], + "angle": 0, + "content": "[25] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14431-14440, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.375, + 0.473, + 0.437 + ], + "angle": 0, + "content": "[26] Yangyu Huang, Hao Yang, Chong Li, Jongyoo Kim, and Fangyun Wei. Adnet: Leveraging error-bias towards normal direction in face alignment. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3060-3070, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.439, + 0.473, + 0.502 + ], + "angle": 0, + "content": "[27] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. In Advances in Neural Information Processing Systems, pages 4016-4027. Curran Associates, Inc., 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.504, + 0.473, + 0.555 + ], + "angle": 0, + "content": "[28] X. Ji, A. Vedaldi, and J. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9864-9873, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.557, + 0.473, + 0.631 + ], + "angle": 0, + "content": "[29] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.633, + 0.473, + 0.697 + ], + "angle": 0, + "content": "[30] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Advances in Neural Information Processing Systems, pages 18661-18673. Curran Associates, Inc., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.699, + 0.473, + 0.774 + ], + "angle": 0, + "content": "[31] Abhinav Kumar, Tim K. Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8233-8243, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.776, + 0.473, + 0.827 + ], + "angle": 0, + "content": "[32] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via c-f labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.828, + 0.473, + 0.867 + ], + "angle": 0, + "content": "[33] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In 2022 IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.128, + 0.857, + 0.152 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 4166-4175, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.154, + 0.857, + 0.203 + ], + "angle": 0, + "content": "[34] Jianshu Li, Fang Zhao, Jiashi Feng, Sujoy Roy, Shuicheng Yan, and Terence Sim. Landmark free face attribute prediction. IEEE Transactions on Image Processing, 27(9):4651-4662, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.205, + 0.857, + 0.279 + ], + "angle": 0, + "content": "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.281, + 0.857, + 0.33 + ], + "angle": 0, + "content": "[36] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.332, + 0.857, + 0.419 + ], + "angle": 0, + "content": "[37] Jinpeng Li, Haibo Jin, Shengcai Liao, Ling Shao, and Pheng-Ann Heng. Repformer: Refinement pyramid transformer for robust facial landmark detection. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 1088–1094. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.421, + 0.857, + 0.483 + ], + "angle": 0, + "content": "[38] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2584-2593, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.485, + 0.857, + 0.547 + ], + "angle": 0, + "content": "[39] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Self-supervised representation learning from videos for facial action unit detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10916-10925, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.548, + 0.857, + 0.598 + ], + "angle": 0, + "content": "[40] Yong Li, Jiabei Zeng, and Shiguang Shan. Learning representations for facial actions from unlabeled videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):302-317, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.6, + 0.857, + 0.674 + ], + "angle": 0, + "content": "[41] Yuanyuan Liu, Wenbin Wang, Yibing Zhan, Shaoze Feng, Kejun Liu, and Zhe Chen. Pose-disentangled contrastive learning for self-supervised facial representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9717-9728, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.676, + 0.857, + 0.725 + ], + "angle": 0, + "content": "[42] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, 2015. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.727, + 0.857, + 0.79 + ], + "angle": 0, + "content": "[43] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1557-1567, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.791, + 0.857, + 0.84 + ], + "angle": 0, + "content": "[44] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.842, + 0.857, + 0.867 + ], + "angle": 0, + "content": "[45] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Ma-hoor. Affectnet: A database for facial expression, valence," + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.128, + 0.857, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2090" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.144, + 0.127, + 0.472, + 0.154 + ], + "angle": 0, + "content": "and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.155, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[46] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-bert: Bert-based facial micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1492, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.219, + 0.472, + 0.308 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.31, + 0.472, + 0.374 + ], + "angle": 0, + "content": "[48] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.375, + 0.472, + 0.437 + ], + "angle": 0, + "content": "[49] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. A semi-automatic methodology for facial landmark annotation. In 2013 IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 896-903, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.439, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[50] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: Database and results. Image and vision computing, 47:3-18, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.491, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[51] Ankit Kumar Sharma and Hassan Foroosh. Slim-cnn: A light-weight cnn for face attribute prediction. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 329-335, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.544, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[52] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11911-11920, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.608, + 0.472, + 0.672 + ], + "angle": 0, + "content": "[53] Yuxuan Shu, Xiao Gu, Guang-Zhong Yang, and Benny P L Lo. Revisiting self-supervised contrastive learning for facial expression recognition. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.673, + 0.472, + 0.748 + ], + "angle": 0, + "content": "[54] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2132-2141, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.75, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[55] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.803, + 0.472, + 0.866 + ], + "angle": 0, + "content": "[56] James Thewlis, Samuel Albanie, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of landmarks by descriptor vector exchange. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6360-6370, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.127, + 0.472, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.19 + ], + "angle": 0, + "content": "[57] Hao Wang, Min Li, Yangyang Song, Youjian Zhang, and Liying Chi. Ucol: Unsupervised learning of discriminative facial representations via uncertainty-aware contrast. Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):2510-2518, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.857, + 0.255 + ], + "angle": 0, + "content": "[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.257, + 0.857, + 0.333 + ], + "angle": 0, + "content": "[59] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Liang Liu, Yabiao Wang, and Chengjie Wang. Toward high quality facial representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, page 5048-5058, New York, NY, USA, 2023. Association for Computing Machinery. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.857, + 0.397 + ], + "angle": 0, + "content": "[60] Wenyan Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2138, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.857, + 0.464 + ], + "angle": 0, + "content": "[61] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4042-4051, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.465, + 0.857, + 0.539 + ], + "angle": 0, + "content": "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.542, + 0.857, + 0.605 + ], + "angle": 0, + "content": "[63] Jing Yang, Qingshan Liu, and Kaihua Zhang. Stacked hourglass network for robust facial landmark localisation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2025-2033, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.607, + 0.857, + 0.657 + ], + "angle": 0, + "content": "[64] Sejong Yang, Subin Jeon, Seonghyeon Nam, and Seon Joo Kim. Dense interspecies face embedding. In Advances in Neural Information Processing Systems, pages 33275-33288. Curran Associates, Inc., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.857, + 0.71 + ], + "angle": 0, + "content": "[65] M. Ye, X. Zhang, P. C. Yuen, and S. Chang. Unsupervised embedding learning via invariant and spreading instance feature. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6203-6212, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.711, + 0.857, + 0.774 + ], + "angle": 0, + "content": "[66] Lijun Yin, Xiaozhou Wei, Yi Sun, Jun Wang, and Matthew J Rosato. A 3d facial expression database for facial behavior research. In 7th international conference on automatic face and gesture recognition (FGR06), pages 211-216. IEEE, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.857, + 0.84 + ], + "angle": 0, + "content": "[67] Xiang Zhang, Taoyue Wang, Xiaotian Li, Huiyuan Yang, and Lijun Yin. Weakly-supervised text-driven contrastive learning for facial behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20751-20762, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.857, + 0.867 + ], + "angle": 0, + "content": "[68] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object land" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.513, + 0.902 + ], + "angle": 0, + "content": "2091" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.141, + 0.127, + 0.472, + 0.165 + ], + "angle": 0, + "content": "marks as structural representations. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.167, + 0.473, + 0.217 + ], + "angle": 0, + "content": "[69] Yuhang Zhang, Chengrui Wang, and Weihong Deng. Relative uncertainty learning for facial expression recognition. In Advances in Neural Information Processing Systems, pages 17616-17627. Curran Associates, Inc., 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.218, + 0.473, + 0.281 + ], + "angle": 0, + "content": "[70] Yuhang Zhang, Chengrui Wang, Xu Ling, and Weihong Deng. Learn from all: Erasing attention consistency for noisy label facial expression recognition. In Computer Vision - ECCV 2022, pages 418-434, Cham, 2022. Springer Nature Switzerland. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.283, + 0.473, + 0.333 + ], + "angle": 0, + "content": "[71] Kaili Zhao, Wen-Sheng Chu, and Honggang Zhang. Deep region and multi-label learning for facial action unit detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3399, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.334, + 0.473, + 0.41 + ], + "angle": 0, + "content": "[72] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18676-18688, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.411, + 0.473, + 0.463 + ], + "angle": 0, + "content": "[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.464, + 0.473, + 0.527 + ], + "angle": 0, + "content": "[74] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15475-15484, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.528, + 0.473, + 0.579 + ], + "angle": 0, + "content": "[75] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 146-155, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.515, + 0.902 + ], + "angle": 0, + "content": "2092" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_origin.pdf b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..525f4a1dcc4cbe562f57e66cfbd60baa10aa0ac8 --- /dev/null +++ b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/bb83bfc2-6351-436b-90f5-921020f4fe95_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2987617cbb93a3f44e2ea8b9fa8edb5719e201028e883a23d5c888ac8b7c3d2 +size 349603 diff --git a/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/full.md b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3c691c2b23ef03dd676f5978f14b2caaf0559527 --- /dev/null +++ b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/full.md @@ -0,0 +1,335 @@ +# Self-Supervised Facial Representation Learning with Facial Region Awareness + +Zheng Gao Ioannis Patras + +Queen Mary University of London, Mile End Road, London, E1 4NS + +{z.gao, i.patras}@qmul.ac.uk + +# Abstract + +Self-supervised pre-training has been proved to be effective in learning transferable representations that benefit various visual tasks. This paper asks this question: can self-supervised pre-training learn general facial representations for various facial analysis tasks? Recent efforts toward this goal are limited to treating each face image as a whole, i.e., learning consistent facial representations at the image-level, which overlooks the "consistency of local facial representations" (i.e., facial regions like eyes, nose, etc). In this work, we make a first attempt to propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). Specifically, we explicitly enforce the consistency of facial regions by matching the local facial representations across views, which are extracted with learned heatmaps highlighting the facial regions. Inspired by the mask prediction in supervised semantic segmentation, we obtain the heatmaps via cosine similarity between the per-pixel projection of feature maps and "facial mask embeddings" computed from learnable positional embeddings, which leverage the attention mechanism to globally look up the facial image for facial regions. To learn such heatmaps, we formulate the learning of facial mask embeddings as a deep clustering problem by assigning the pixel features from the feature maps to them. The transfer learning results on facial classification and regression tasks show that our FRA outperforms previous pre-trained models and more importantly, using ResNet as the unified backbone for various tasks, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks. + +# 1. Introduction + +Human face understanding is an important and challenging topic in computer vision [41, 72] and supervised learning algorithms have shown promising results on a wide range of facial analysis tasks recently [5, 31, 55, 71]. Despite + +the impressive progress, these methods require large-scale well-annotated training data, which is expensive to collect. + +Recent works in self-supervised representation learning for visual images have shown that self-supervised pretraining is effective in improving the performance on various downstream tasks such as image classification, object detection and segmentation as it can learn general representations from unlabeled data that could be transferred to downstream visual tasks, especially tasks with limited labeled data [18, 24, 28, 30, 58, 62, 65]. Among them, instance discrimination (including contrastive learning [10, 22, 36] and non-contrastive learning [11, 20] paradigms) has been shown to be effective in learning generalizable self-supervised features. Instance discrimination aims to learn view-invariant representations by matching the global representations between the augmented views generated by aggressive image augmentations, i.e., the image-level representations of the augmented views should be similar [10-12, 20, 22, 36]. Another self-supervised learning paradigm, masked image modeling (MIM) [23, 54] learns visual representations by reconstructing image content from a masked image, achieving excellent performance in full model fine-tuning. This leads to the question: can self-supervised pre-training learn general facial representations which benefit downstream facial analysis tasks? + +Several attempts have been made to learn general facial representations for facial analysis tasks [3, 41, 72]. For example, Bulat et al. [3] directly applies the contrastive objective to facial features. FaRL [72] and MCF [59] combine contrastive learning and mask image modeling [23]. PCL [41] proposes to disentangle the pose-related and pose-unrelated features, achieving strong performance on both pose-related (regression) and pose-unrelated (classification) tasks. However, it runs the model forward and backward three times for each training step, which is time-consuming. Despite different techniques, these methods commonly treats each face image as a whole to learn consistent global representations at the image-level and overlook the "spatial consistency of local representations", i.e., local facial regions (e.g., eyes, nose and mouth) should also be + +similar across the augmented views, thus limiting the generalization to downstream tasks. This brings us to the focus: learning consistent global and local representations for facial representation learning. + +We argue that in order to learn consistent local representations, the model needs to look into facial regions. Towards that goal, we predict a set of heatmaps highlighting different facial regions by leveraging learnable positional embeddings as facial queries (the feature maps as keys and values) to look up the facial image globally for facial regions, which is inspired by the mask prediction in supervised segmentation [13]. For visual images, the attention mechanism of Transformer allows the learnable positional embeddings to serve as object queries for visual pattern look-up [7, 13]. In our case (facial images), the learnable positional embeddings can be used as facial queries for facial regions (see the visualization in the supplementary material). + +In this work, taking the consistency of facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns general facial representations by enforcing consistent global and local facial representations, based on a popular instance discrimination baseline BYOL [20] for its simplicity. Specifically, we learn consistent local facial representations by match them across augmented views, which are extracted by aggregating the feature maps using learned heatmaps highlighting the facial regions as weights. Inspired by the mask prediction in MaskFormer [13], we produce the heatmaps from a set of learnable positional embeddings, which are used as facial queries to look up the facial image for facial regions. A Transformer decoder takes as input the feature maps from the encoder and the learnable positional embeddings to output a set of "facial mask embeddings", each associated with a facial region. The facial mask embeddings are used to compute cosine similarity with the per-pixel projection of feature maps to produce the heatmaps. In addition, we enforce the consistency of global representations across views simultaneously so that the image-level information is preserved. In order to learn the heatmaps (facial mask embeddings), inspired by deep clustering [8] that learns to assign samples to clusters, we treat the facial mask embeddings as clusters and learn to assign pixel features from the feature maps to them. Specifically, we align the per-pixel cluster assignments of each pixel feature over the facial region clusters between the online and momentum network for the same augmented view (i.e., each pixel feature should have similar similarity distribution over the facial mask embeddings between the momentum teacher and online student). In contrast to supervised segmentation that directly uses ground-truth masks to supervise the learning of the masks (heatmaps) with a per-pixel binary mask loss, we formulate the learning of heatmaps as a deep cluster + +ing [8] problem that learns to assign pixel features to clusters (facial mask embeddings) in a self-supervised manner. + +Our contributions can be summarized as follows: + +- Taking into the consistency of local facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns consistent global and local facial representations. +- We show that the learned heatmaps can roughly discover facial regions in the supplementary material. +- In previous works, different backbones are adopted for different facial analysis tasks (e.g., in face alignment the common backbone is hourglass network [63] while in facial expression recognition ResNet [21] is the popular backbone). In this work, our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks. +- Our FRA outperforms existing self-supervised pretraining approaches (e.g., BYOL [20] and PCL [41]) on facial classification (i.e., facial expression recognition [19, 38] and facial attribute recognition [42]) and regression (i.e., face alignment [48-50, 60]) tasks. More importantly, our FRA achieves comparable (e.g., face alignment) or even better performance (e.g., facial expression recognition) compared with SOTA methods in the corresponding facial analysis tasks. + +# 2. Related work + +# 2.1. Visual representation learning + +As one of the main paradigms for self-supervised pretraining, instance discrimination learns representations by treating an image as a whole and enforcing the consistency of global representations at the image-level across augmented views. Generally, instance discrimination includes two paradigms: contrastive learning [10, 22, 36] and non-contrastive learning [11, 20]. Contrastive learning considers each image and its transformations as a distinct class, i.e., "positive" samples are pulled together while "negative" samples are pushed apart in the latent space. Unlike contrastive learning that relies on negative samples to avoid collapse, non-contrastive learning directly maximizes the similarity of the global representations between the augmented views without involving negative samples based on techniques like stop-gradient [11] and predictor [20]. Further works perform visual-language pre-training by applying contrastive objective to image-text pairs [29, 35, 47]. + +Another line of work, masked image modeling (MIM) learns visual representations by reconstructing image content from a masked image [1, 23, 54], which is inspired by the masked language modeling in NLP [15]. In contrast to instance discrimination, MIM achieves strong full model fine-tuning performance with Vision Transformers + +pre-trained for enough epochs. However, these works suffer from poor linear separability and are less data-efficient than instance discrimination in few-shot scenarios [1]. + +# 2.2. Facial representation learning + +Recent works on facial analysis explore self-supervised learning for several face-related tasks, such as facial expression recognition [9, 53], face recognition [9, 57], facial micro-expression recognition [46], AU detection [39, 40], face alignment (facial landmark detection) [14, 64], etc. However, these methods are task-specific, i.e., tailored for a specific facial task and thus lack the ability to generalize to various facial analysis tasks [41]. Further efforts [3, 41, 72] focus on learning general facial representations with contrastive learning and mask image modeling [23, 54]. Bulat et al. [3] directly apply the contrastive objective to augmented views of the same face image, showing that general facial representation learned from pre-training can benefit various facial analysis tasks. FaRL [72] performs pre-training in a visual-linguistic manner by employing image-text contrastive learning and masked image modeling. MCF [59] leverages image-level contrastive learning and masked image modeling, along with the knowledge distilled from external ImageNet pre-trained model for facial representation learning. PCL [41] argues that directly applying the contrastive objective to face images overlooks the variances of facial poses and thus leads to pose-invariant representations, limiting the performance on pose-related tasks [66, 75]. Therefore, PCL [41] disentangles the pose-related and pose-unrelated features and then performs contrastive learning on these features, achieving strong performance on both pose-related and pose-unrelated facial analysis tasks. Despite the success, it performs forward and backward three times for each input image, which brings significant increase on computational cost. These works are commonly limited by instance discrimination and overlook the consistency of local facial regions. In contrast, inspired by supervised semantic segmentation, we learn consistent global and local facial representations by learning a set of heatmaps indicating facial regions from learnable positional embeddings, which leverage the attention mechanism to look up facial image globally for facial regions. + +# 2.3. Facial region discovery + +There are some approaches leveraging facial region (landmark) discovery for facial analysis [27, 43, 61]. Some focus on landmark detection by either learning a heatmap for each landmark via image reconstruction [27, 68], or performing pixel-level matching with an equivariance loss [56, 68]. Despite different techniques, these methods are task-specific, i.e., landmark detection with discovery of local information, while our method is task-agnostic, i.e., learn general facial representations for various tasks by preserving + +global and local information in a image, region and pixel-level contrastive manner. MARLIN [4] applies masked image modeling to learn general representations for facial videos by utilizing an external face parsing algorithm to discover the facial regions (e.g., eyes, nose and mouth), which are used to guide the masking for the masked autoencoder. A closely related work SLPT [61] leverages the attention mechanism to estimate facial landmarks from initial facial landmarks estimates of the mean face through supervised learning. These works commonly rely on external supervisory signal, whether it is from ground-truths or additional algorithms. In contrast, we learn to discover the facial regions in an end-to-end self-supervised manner for facial image representation learning. + +# 3. Methodology + +# 3.1. Overview + +The overview of the proposed FRA is shown in Fig. 1. The goal is to learn consistent global and local facial representations. Toward this goal, we propose two objectives: pixel-level semantic relation and image/region-level semantic consistency. Semantic relation aligns the per-pixel cluster assignments of each pixel feature over the facial mask embeddings between the online and momentum network to learn heatmaps for facial regions (Sec. 3.2) while semantic consistency directly matches the global and local facial representations across augmented views (Sec. 3.3) with the learned heatmaps. + +# 3.2. Semantic relation + +As shown in Fig. 1, our method adopts the Siamese structure of BYOL [20], a popular self-supervised pretraining baseline based on instance discrimination. Following BYOL [20], we employ two branches: the online network parameterized by $\theta$ and the momentum network parameterized by $\xi$ . The online network $\theta$ consists of an encoder $E_{\theta}$ , a global projector $H_{\theta}^{g}$ and a local projector $H_{\theta}^{l}$ . The momentum network $\xi$ has the same architecture with the online network, except $\xi$ is updated with an exponential moving average of $\theta$ . As in BYOL [20], we also adopt additional predictors $G_{\theta}^{g}$ and $G_{\theta}^{l}$ on top of the projectors in the online network. Note that this is omitted for brevity in Fig. 1. + +Given an input image $\mathbf{x}$ , two random augmentations are applied to generate two augmented views $\mathbf{x}_1 = \mathcal{T}_1(\mathbf{x})$ and $\mathbf{x}_2 = \mathcal{T}_2(\mathbf{x})$ , following BYOL [20]. Each augmented view $\mathbf{x}_i \in \{\mathbf{x}_1, \mathbf{x}_2\}$ is fed into an encoder $E$ to obtain a feature map $\mathbf{F}_i \in \mathbb{R}^{C \times H \times W}$ (before global average pooling), where $C, H, W$ are the number of channels, height and width of $\mathbf{F}_i$ and a latent representation $\mathbf{h}_i \in \{\mathbf{h}_1, \mathbf{h}_2\}$ (after global average pooling), i.e., $\mathbf{h}_1 = E_{\theta}(\mathbf{x}_1)$ and $\mathbf{h}_2 = E_{\xi}(\mathbf{x}_2)$ . Then each latent representation $\mathbf{h}_i$ is transformed + +![](images/4db6d3a6998f58ae0a8d163e71b2d81b1929bebbd4b443f85e50ee48549d111b.jpg) +Figure 1. Overview of the proposed FRA framework. $\odot$ denotes cosine similarity. For each input image $\mathbf{x}$ , its augmented views $\mathbf{x}_1$ and $\mathbf{x}_2$ are passed into two network branches to produce the global embeddings $\mathbf{z}_1$ and $\mathbf{z}_2$ . In addition, we produce a set of heatmaps $\mathbf{M}_1$ and $\mathbf{M}_2$ indicating the local facial regions, via the correlation between the pixel features and "facial mask embeddings" computed from a set of learnable positional embeddings. Then we aggregate the feature map to obtain the local facial embeddings $\{\mathbf{z}_1^m\}$ and $\{\mathbf{z}_2^m\}$ . The semantic consistency loss is applied to global embeddings and facial embeddings to maximize the similarity across augmented views. To learn such heatmaps, i.e., facial mask embeddings, we treat the facial mask embeddings as facial region clusters and propose a semantic relation loss to align the cluster assignments of each pixel feature over the facial region clusters between the online and momentum network. + +![](images/12e98237741f2f7ca67d9f53d279a2aae367a7bf7b91549941c016c0bf2bdb16.jpg) +Figure 2. Generation of heatmaps using learnable positional embeddings as facial queries and the feature maps as keys and values. + +by a global projector $H^g$ to produce a global embedding $\mathbf{z}_i\in \{\mathbf{z}_1,\mathbf{z}_2\}$ of dimension $\mathbf{z}_i\in \mathbb{R}^D$ , i.e., $\mathbf{z}_1 = H_\theta^g (\mathbf{h}_1)$ and $\mathbf{z}_2 = H_{\xi}^{g}(\mathbf{h}_{2})$ + +Then we obtain a set of heatmaps $\mathbf{M}_i\in \{\mathbf{M}_1,\mathbf{M}_2\}$ highlighting the facial regions from the feature map $\mathbf{F}_i$ for each view, which is inspired by mask classification-based supervised segmentation [7, 13] that leverages attention mechanism to look up visual patterns globally. First the local projector (e.g., $H_{\theta}^{l}$ ) is applied to project the pixel features of $\mathbf{F}_i$ in a pixel-wise manner, mapping it to $D$ dimensions to get the dense feature map $\mathbf{F}_i^{\mathrm{dense}}\in \mathbb{R}^{D\times H\times W}$ . Take view $\mathbf{x}_1$ as an example, the projected feature map can be expressed as: + +$$ +\mathbf {F} _ {1} ^ {\text {d e n s e}} [ *, u, v ] = H _ {\theta} ^ {l} \left(\mathbf {F} _ {1} [ *, u, v ]\right), \tag {1} +$$ + +where $\mathbf{F}_1[*,u,v]\in \mathbb{R}^C$ is the pixel feature at the $(u,v)$ -th grid of $\mathbf{F}_1$ . Then as shown in Fig. 2, inspired by supervised segmentation [13], we use a Transformer decoder followed by a MLP, which takes as input the feature map $\mathbf{F}_i$ and $N$ learnable positional embeddings (i.e., facial queries for looking up the facial image globally for facial regions) to predict $N$ "facial mask embeddings" $\mathbf{Q}_i\in \mathbb{R}^{N\times D}$ of dimension $D$ , where each row associated with a facial region. Next, we compute the cosine similarity between facial mask embeddings $\mathbf{Q}_i$ and dense feature map $\mathbf{F}_i^{\mathrm{dense}}$ along the channel dimension, yielding per-pixel cluster assignments $\mathbf{S}_i\in \mathbb{R}^{N\times H\times W}$ , where $\mathbf{S}_i[*,u,v]$ denotes the relation between the dense pixel feature $\mathbf{F}_1^{\mathrm{dense}}[*,u,v]$ and facial mask embeddings $\mathbf{Q}_i$ . Finally, we normalize $\mathbf{S}_i$ along the channel dimension with a softmax operation to encourage each channel to capture a different pattern, obtaining $N$ heatmaps $\mathbf{M}_i\in \mathbb{R}^{N\times H\times W}$ where each vector at location $(u,v)$ is a probabilistic similarity distribution (i.e., normalized per-pixel cluster assignments) $\mathbf{s}_1^{u,v}$ between $\mathbf{F}_i^{\mathrm{dense}}[*,u,v]$ and $\mathbf{Q}_i$ . Note that $\mathbf{M}_i$ is a set of heatmaps where each channel of $\mathbf{M}_i$ represents a 2D heatmap $\mathbf{M}_i^{(m)}$ . + +To learn such heatmaps, i.e., facial mask embeddings, inspired by deep clustering [8], we treat the facial mask embeddings as facial region clusters and align the per-pixel cluster assignments of each pixel feature over these clusters between the online and momentum network for the same augmented view, using the momentum network as momentum teacher [16, 73] to provide reliable target. + +Following BYOL [20], we pass both augmented views + +$\mathbf{x}_1$ and $\mathbf{x}_2$ through the online and momentum network. Take $\mathbf{x}_1$ as an example, the online network $\theta$ outputs the normalized per-pixel cluster assignments $\mathbf{s}_1^{u,v}$ and the momentum network outputs normalized assignments $\widehat{\mathbf{s}}_1^{u,v}$ for view $\mathbf{x}_1$ . Then we learn $\mathbf{s}_1^{u,v}$ using $\widehat{\mathbf{s}}_1^{u,v}$ as guidance based on the following cross-entropy loss: + +$$ +C E \left(\mathbf {s} _ {1} ^ {u, v}, \widehat {\mathbf {s}} _ {1} ^ {u, v}\right) = - \sum_ {m = 1} ^ {N} \widehat {\mathbf {s}} _ {1} ^ {u, v} [ m ] \log \mathbf {s} _ {1} ^ {u, v} [ m ]. \tag {2} +$$ + +For both augmented views, we define the symmetrized semantic relation objective as: + +$$ +\mathcal {L} _ {\mathrm {r}} = \frac {1}{H W} \sum_ {u, v} \left(C E \left(\mathbf {s} _ {1} ^ {u, v}, \widehat {\mathbf {s}} _ {1} ^ {u, v}\right) + C E \left(\mathbf {s} _ {2} ^ {u, v}, \widehat {\mathbf {s}} _ {2} ^ {u, v}\right)\right), \tag {3} +$$ + +where $CE(\mathbf{s}_2^{u,v},\widehat{\mathbf{s}}_2^{u,v})$ is the cross-entropy loss for view $\mathbf{x}_2$ . We apply the Sinkhorn-Knopp normalization to the target assignments from the momentum network following [8] to avoid collapse and the mean entropy maximization (ME-MAX) regularizer [1] to maximize the entropy of the prediction to encourage full use of the clusters. + +# 3.3. Semantic consistency + +In this section, we enforce the consistency of global embeddings and local facial embeddings. With the learned heatmaps $\mathbf{M}_i$ , we generate the latent representations for the local facial regions through weighted average pooling: + +$$ +\begin{array}{l} \mathbf {h} _ {i} ^ {m} = \mathbf {M} _ {i} ^ {(m)} \otimes \mathbf {F} _ {i} \\ = \frac {1}{\sum_ {u , v} \mathbf {M} _ {i} [ m , u , v ]} \sum_ {u, v} \mathbf {M} _ {i} [ m, u, v ] \mathbf {F} _ {i} [ *, u, v ], \tag {4} \\ \end{array} +$$ + +where $\otimes$ denotes channel-wise weighted average pooling, $\mathbf{M}_i^{(m)}$ is the $m$ -th channel (heatmap) of $\mathbf{M}_i$ and $\mathbf{h}_i^m \in \mathbb{R}^C$ is the corresponding latent representation produced with $\mathbf{M}_i^{(m)}$ . The facial embeddings $\{\mathbf{z}_1^m : \mathbf{z}_1^m \in \mathbb{R}^D\}_{m=1}^N$ and $\{\mathbf{z}_2^m : \mathbf{z}_2^m \in \mathbb{R}^D\}_{m=1}^N$ are obtained accordingly with the local projector $H_\theta^l$ and $H_\xi^l$ : + +$$ +\begin{array}{l} \mathbf {z} _ {1} ^ {m} = H _ {\theta} ^ {l} \left(\mathbf {h} _ {1} ^ {m}\right), \\ \mathbf {z} _ {2} ^ {m} = H _ {\xi} ^ {l} \left(\mathbf {h} _ {2} ^ {m}\right). \end{array} \tag {5} +$$ + +We then match the global embeddings and local facial embeddings across views using the negative cosine similarity in BYOL [20]: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s i m}} \left(\mathbf {z} _ {1}, \mathbf {z} _ {2}\right) = - \left(\lambda_ {\mathrm {c}} \times f _ {\mathrm {s}} \left(G _ {\theta} ^ {g} \left(\mathbf {z} _ {1}\right), \mathbf {z} _ {2}\right) + \right. \\ + \left(1 - \lambda_ {\mathrm {c}}\right) \times \frac {1}{N} \sum_ {m = 1} ^ {N} f _ {\mathrm {s}} \left(G _ {\theta} ^ {l} \left(\mathbf {z} _ {1} ^ {m}\right), \mathbf {z} _ {2} ^ {m}\right)), \tag {6} \\ \end{array} +$$ + +where $f_{\mathrm{s}}(\mathbf{u},\mathbf{v}) = \frac{\mathbf{u}^{\top}\mathbf{v}}{\|\mathbf{u}\|_{2}\|\mathbf{v}\|_{2}}$ denotes the cosine similarity between the vectors $\mathbf{u}$ and $\mathbf{v}$ , $\lambda_{\mathrm{c}}$ is the loss weight, $G_{\theta}^{g}$ and $G_{\theta}^{l}$ are the predictors on top of the projectors $H_{\theta}^{g}$ and $H_{\theta}^{l}$ , respectively. Following BYOL [20], we symmetrize the loss $\mathcal{L}_{\mathrm{sim}}(\mathbf{z}_1,\mathbf{z}_2)$ defined in Eq. (6) by passing $\mathbf{x}_1$ through the momentum network $\xi$ and $\mathbf{x}_2$ through the online network $\theta$ to compute $\mathcal{L}_{\mathrm{sim}}(\mathbf{z}_2,\mathbf{z}_1)$ . The semantic consistency objective can be expressed as follows: + +$$ +\mathcal {L} _ {\mathrm {c}} = \mathcal {L} _ {\text {s i m}} \left(\mathbf {z} _ {1}, \mathbf {z} _ {2}\right) + \mathcal {L} _ {\text {s i m}} \left(\mathbf {z} _ {2}, \mathbf {z} _ {1}\right). \tag {7} +$$ + +# 3.4. Overall objective + +We jointly optimize the semantic relation objective Eq. (3) and the semantic consistency objective Eq. (7), leading to the following overall objective: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {c}} + \lambda_ {\mathrm {r}} \mathcal {L} _ {\mathrm {r}}, \tag {8} +$$ + +where $\lambda_{\mathrm{r}}$ is the loss weight for balancing $\mathcal{L}_{\mathrm{c}}$ and $\mathcal{L}_{\mathrm{r}}$ + +# 4. Experiments + +# 4.1. Experimental setups + +# 4.1.1 Implementation details + +We use the same augmentation strategy as in [20, 25]. The number of heatmaps $N$ is set to 8 empirically. The loss weight $\lambda_{\mathrm{c}}$ and $\lambda_{\mathrm{r}}$ are set to $0.5 / 0.1$ , respectively. For fair comparisons, the other hyper-parameters are kept the same as BYOL [20] in all experiments. The architecture and pretraining details are provided in the supplementary material. + +# 4.1.2 Baselines + +Our baselines are self-supervised pre-training approaches for visual images (e.g., BYOL [20] and LEWEL [25]), and pre-training approaches for facial images (e.g., Bulat et al. [3] and PCL [41]). Note that SwAV [8] is equivalent to Bulat et al. [3]. As we adopt BYOL [20] as the pretraining backbone, we compare our FRA with BYOL [20] in all experiments. We also compare our FRA with another pre-training method LEWEL [25], which learns consistent local representations for visual images. Moreover, we perform comparisons with SOTA methods in the corresponding downstream tasks. + +# 4.2. Evaluation protocols + +Following the common practice in previous works [41, 72], we evaluate the transfer performance of the self-supervised pre-trained facial representations on several popular downstream facial analysis tasks: facial expression recognition (FER) [2, 38], facial attribute recognition (FAR) [42] and face alignment (FA) [48-50, 60]. Specifically, we use + +Table 1. Comparisons with weakly-supervised pre-trained vision transformer on several downstream facial analysis tasks, including facial expression recognition (AffectNet), facial attribute recognition (CelebA) and face alignment (300W). + +
MethodArch.Params.Pre-training settingsDownstream performances
DatasetScaleSupervisionAffectNet Acc. ↑CelebA Acc. ↑300W NME ↓
FaRL [72]ViT-B/16 [17]86MLAION-FACE [72]20Mface image + text64.8591.883.08
FRAR50 [21]24MVGGFace2 [6]3.3Mface image66.1692.022.91
+ +Table 2. Comparisons on facial expression recognition. We report the Top-1 accuracy on test set. Text denotes text supervision.†: our reproduction using the official codes. + +
MethodTextFERPlusRAF-DBAffectNet
Supervised
KTN [32]×90.4988.0763.97
RUL [69]×88.7588.9861.43
EAC [70]×90.0590.3565.32
Weakly-Supervised
FaRL [72]†✓88.6288.3164.85
CLEF [67]✓89.7490.0965.66
Self-supervised
MCF [59]†×88.1786.8660.98
Bulat et al. [3, 8]×--60.20
BYOL [20]×89.2589.5365.65
LEWEL [25]×85.6181.8561.20
PCL [41]×85.8785.9260.77
FRA (LP)×78.1373.8957.38
FRA (FT)×89.7889.9566.16
FRA (EAC)×90.6290.7665.85
+ +the pre-trained weights to initialize the backbone of downstream tasks and then learn the backbone and task-specific head (attached to the backbone) jointly. Following [72], we report the performance with linear probe (denoted by "LP") and fine-tuning (denoted by "FT"). The details of the downstream tasks are described as follows: + +Facial expression recognition is a multi-class classification task where the goal is to categorize the emotional expressions (e.g., anger, fear and surprise) for a given face image. Three widely-used datasets are adopted: FERPlus [2], RAF-DB [38] and AffectNet [45]. For RAF-DB, we use the basic emotion subset following [32, 41, 70]. For AffectNet, we report the results with 7 emotion classes (i.e., neutral, happy, sad, surprise, fear, anger, disgust) following [32, 70]. + +Facial attribute recognition is a multi-label classification task to predict various attributes (e.g., gender, age and race) of a given face image. We adopt the popular benchmark CelebA [42], which consists of more than 200K face + +Table 3. Comparisons on CelebA [42] facial attribute recognition. We report the averaged accuracy over all attributes. †: our reproduction using the official codes. *: results cited from [72]. + +
MethodAcc. (%)
Supervised
DMM [44]91.70
SlimCNN [51]91.24
AFFAIR [34]91.45
Self-supervised
SSPL [52]91.77
Bulat et al. [3, 8]*89.65
SimCLR [10]*91.08
BYOL [20]91.56
LEWEL [25]90.69
PCL [41]91.48
MCF [59]†91.33
FRA (LP)90.86
FRA (FT)92.02
+ +images with 40 facial attributes per image. Following [72], we report the averaged accuracy over all attributes. + +Face alignment is a regression task to predict 2D face landmark coordinates on a face image. We use two popular benchmarks: WFLW [60] and 300W [48-50]. Following the common practice [14, 26, 74], we report normalized mean error (NME), failure rate (FR) and AUC. For 300W, we report the results on full test set, common (554 images) and challenge (135 images) splits of the test set following [26, 74]. + +# 4.3. Comparisons with weakly-supervised pretraining + +In Tab. 1, we compare our FRA with SOTA pre-trained Transformer FaRL [72], which is a weakly-supervised model pre-trained on 20M visual-linguistic data (face image and text) with image-text contrastive learning and mask image modeling. We fine-tune both the pre-trained feature backbone and the task-specific head on the corresponding downstream facial analysis task. Our self-supervised FRA with 24M parameter ResNet-50 achieves superior perfor + +Table 4. Comparisons on face alignment. †: our reproduction using the official codes. + +
MethodVenueArch.WFLW300W (NME ↓)
NME ↓FR10% ↓AUC10% ↑FullComm.Chal.
Supervised
SLPT [61][CVPR'22]ResNet [21]4.203.040.5883.202.784.93
DTLD [33][CVPR'22]ResNet [21]4.082.76-2.962.594.50
RePFormer [37][IJCAI'22]ResNet [21]4.11--3.01--
ADNet [26][ICCV'21]Hourglass [63]4.142.720.6022.932.534.58
STAR [74][CVPR'23]Hourglass [63]4.022.320.6052.872.524.32
Self-supervised
MCF [59] (concurrent work)[ACMM'M'23]ViT [17]3.961.400.6092.982.604.51
Bulat et al. [3, 8][ECCV'22]ResNet [21]4.57--3.20--
BYOL [20][NeurIPS'20]ResNet [21]4.292.960.5793.032.664.55
LEWEL [25][CVPR'22]ResNet [21]4.524.500.5633.092.704.71
PCL [41]†[CVPR'23]ResNet [21]4.846.180.5353.352.775.12
FRAOursResNet [21]4.112.530.5912.912.604.46
+ +mance compared with weakly-supervised FaRL [72] with 86M parameter ViT-B/16 and text supervision on all tasks. + +# 4.4. Transfer learning + +In this section, we compare our FRA with self-supervised pre-training approaches and SOTA methods in several downstream tasks. Please refer to the supplementary material for setup details. + +# 4.4.1 Facial expression recognition + +The results on facial expression recognition are reported in Tab. 2. We observe: (1) With the setting of fine-tuning (FT), our FRA outperforms previous self-supervised pre-training approaches for visual images (e.g., BYOL and LEWEL) and pre-training approaches tailored for facial images (e.g., PCL, MCF). In particular, our FRA using a 24M parameter ResNet-50 surpasses the concurrent work MCF [59] with 86M parameter ViT-B/16 [17]. (2) By simply learning a linear classifier on top of the encoder backbone, our FRA outperforms SOTA facial expression recognition methods with sophisticated designs (e.g., EAC [70]) on AffectNet [45], the largest facial expression recognition dataset. (3) More importantly, by using our pre-trained model to initialize the backbone of SOTA facial expression recognition method EAC [70], our variant "FRA (EAC)" consistently improves EAC [70] on all datasets, which suggests "FRA (EAC)" outperforms SOTA FER methods and demonstrates the superiority of the proposed self-supervised pre-training. + +# 4.4.2 Facial attribute recognition + +As shown in Tab. 3, our FRA outperforms both self-supervised pre-training approaches for visual images and pre-training approaches tailored for facial images. The results on facial expression recognition and facial attribute recognition show that our FRA learns better facial representations for facial classification task. + +# 4.4.3 Face alignment + +As shown in Tab. 4, despite SOTA face alignment methods (e.g., ADNet [26] and STAR [74]) commonly rely on hourglass network [63] for feature extraction, which is tailored for regression tasks like landmark detection, our method based on ResNet backbone achieves comparable performance with these SOTA methods (e.g., 2.91 vs. 2.87 on 300W). The results on classification (e.g., facial expression recognition) and regression tasks (e.g., face alignment) show that our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks. + +# 4.5. Ablation studies + +We pre-train the model on VGGFace2 [6] and then evaluate it on facial expression recognition (RAF-DB) and facial attribute recognition (CelebA), as described in Sec. 4.2. + +# 4.5.1 Effect of different modules + +In Tab. 5, we investigate the contributions of the proposed semantic consistency loss (i.e., global consistency of whole face and local consistency of facial regions) and semantic + +Table 5. Effect of different modules. GC denotes the global consistency for aligning images, LC denotes the local consistency for aligning facial regions with heatmaps and SR represents semantic relation for aligning pixels and heatmaps. + +
GCLCSRRAF-DB ↑CelebA ↑300W ↓
✓--87.8290.633.56
--✓85.3490.663.25
-✓-87.4690.783.19
✓✓-88.0590.893.26
✓✓✓88.7291.183.14
+ +Table 6. Effect of the number of heatmaps $N$ . + +
N83264
RAF-DB88.7288.3688.18
CelebA91.1891.0190.98
+ +Table 7. Effect of the loss weights. Please refer to Tab. 5 for GC, LC and SR. + +
SettingsλcλrRAF-DBCelebA
GC1.0087.8290.63
GC + LC0.5088.0590.89
GC + LC + SR (FRA)0.50.188.7291.18
GC + LC + SR (FRA)0.50.588.4591.04
GC + LC + SR (FRA)0.51.088.0890.46
+ +relation loss to our approach. Note that the global consistency (first row) is BYOL [20]. We have the following observations: (1) The variant using all losses achieves the best results. (2) GC is essential to avoid degeneration on classification task. (3) LC or SR alone benefits regression task (landmark). Altogether, LC and SR improve BYOL [20] (GC) on both classification and regression by capturing spatial/local information, which validates our facial region awareness. + +# 4.5.2 Effect of the number of heatmaps + +In Tab. 6, we study the effect of the number of heatmaps. We observe that the best setting is $N = 8$ , which is close to the facial landmarks number 5. This suggests that given enough face images for training, a suitable $N$ can encourage the model to learn face-specific patterns, which helps the transfer learning performance on various facial analysis tasks. Further increasing the number of heatmaps might force the model to look into fine-grained patterns that may not be suitable for facial tasks. + +# 4.5.3 Effect of loss weights + +In Tab. 7, we ablate the weights for the semantic consistency loss and semantic relation loss. We find that setting $\lambda_{\mathrm{c}} = 0.5$ and $\lambda_{\mathrm{r}} = 0.1$ works best. When $\lambda_{\mathrm{c}} = 1.0$ and $\lambda_{\mathrm{r}} = 0$ , only the consistency of global representations is applied, and the model performs relatively worse, which suggests the importance of the consistency of local representations and the semantic relation loss. By using the semantic relation objective, the performance is significantly improved. However, when $\lambda_{\mathrm{r}}$ is too high, the performance degrades as the pixel-level consistency between the online and momentum network might affect the capture of image/object-level information. + +# 4.5.4 Effect of Transformer decoder layers + +Table 8. Effect of Transformer decoder layers. 0 decoder layer represents BYOL [20] where only the consistency of global representation is enforced. + +
# decoder layer0123
RAF-DB87.8288.7289.0189.06
CelebA90.6391.1891.3091.35
+ +In Tab. 8, we study the effect of the number of decoder layers used for heatmap prediction. We observe that a single decoder layer is able to produce decent results, showing that a 1-layer decoder is large enough to capture the facial region (landmarks) relations in face images. The performance gain diminishes as the decoder depth increases. By default, we only use 1 decoder layer for fast training. + +# 5. Conclusion + +In this work, we propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). We learn a set of heatmaps indicating facial regions from learnable positional embeddings, which leverages the attention mechanism to look up facial image globally for facial regions. We show that our FRA outperforms previous pre-trained models on several facial classification and regression tasks. More importantly, using ResNet as the unified backbone, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks. + +# Acknowledgement + +This work was supported by the EU H2020 AI4Media No.951911 project. We thank Zengqun Zhao for his helpful comments on facial expression recognition. + +# References + +[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 2, 3, 5 +[2] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM International Conference on Multimodal Interaction, page 279-283, New York, NY, USA, 2016. Association for Computing Machinery. 5, 6 +[3] Adrian Bulat, Shiyang Cheng, Jing Yang, Andrew Garbett, Enrique Sanchez, and Georgios Tzimiropoulos. Pre-training strategies and datasets for facial representation learning. In Computer Vision - ECCV 2022, pages 107-125, Cham, 2022. Springer Nature Switzerland. 1, 3, 5, 6, 7 +[4] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. Marlin: Masked autoencoder for facial video representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1493-1504, 2023. 3 +[5] Jiajiong Cao, Yingming Li, and Zhongfei Zhang. Partially shared multi-task convolutional neural network with local constraint for face attribute learning. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4290-4299, 2018. 1 +[6] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6, 7 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020, pages 213-229, Cham, 2020. Springer International Publishing. 2, 4 +[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 5, 6, 7 +[9] Jia-Ren Chang, Yong-Sheng Chen, and Wei-Chen Chiu. Learning facial representations from the cycle-consistency of face. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9660-9669, 2021. 3 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2, 6 +[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 1, 2 + +[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1 +[13] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875. Curran Associates, Inc., 2021. 2, 4 +[14] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9877-9886, 2021. 3, 6 +[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2 +[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In Computer Vision - ECCV 2022, pages 247-264, Cham, 2022. Springer Nature Switzerland. 4 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7 +[18] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 1 +[19] Ian J Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, et al. Challenges in representation learning: A report on three machine learning contests. In Neural Information Processing: 20th International Conference, ICONIP 2013, Daegu, Korea, November 3-7, 2013. Proceedings, Part III 20, pages 117-124. Springer, 2013. 2 +[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2, 3, 4, 5, 6, 7, 8 +[21] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 6, 7 + +[22] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 1, 2 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 1, 2, 3 +[24] Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Philip Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR 2019. ICLR, 2019. 1 +[25] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14431-14440, 2022. 5, 6, 7 +[26] Yangyu Huang, Hao Yang, Chong Li, Jongyoo Kim, and Fangyun Wei. Adnet: Leveraging error-bias towards normal direction in face alignment. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3060-3070, 2021. 6, 7 +[27] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. In Advances in Neural Information Processing Systems, pages 4016-4027. Curran Associates, Inc., 2018. 3 +[28] X. Ji, A. Vedaldi, and J. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9864-9873, 2019. 1 +[29] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[30] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Advances in Neural Information Processing Systems, pages 18661-18673. Curran Associates, Inc., 2020. 1 +[31] Abhinav Kumar, Tim K. Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8233-8243, 2020. 1 +[32] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via c-f labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 6 +[33] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In 2022 IEEE/CVF Conference + +on Computer Vision and Pattern Recognition (CVPR), pages 4166-4175, 2022. 7 +[34] Jianshu Li, Fang Zhao, Jiashi Feng, Sujoy Roy, Shuicheng Yan, and Terence Sim. Landmark free face attribute prediction. IEEE Transactions on Image Processing, 27(9):4651-4662, 2018. 6 +[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2 +[36] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 1, 2 +[37] Jinpeng Li, Haibo Jin, Shengcai Liao, Ling Shao, and Pheng-Ann Heng. Repformer: Refinement pyramid transformer for robust facial landmark detection. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 1088–1094. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 7 +[38] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2584-2593, 2017. 2, 5, 6 +[39] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Self-supervised representation learning from videos for facial action unit detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10916-10925, 2019. 3 +[40] Yong Li, Jiabei Zeng, and Shiguang Shan. Learning representations for facial actions from unlabeled videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):302-317, 2022. 3 +[41] Yuanyuan Liu, Wenbin Wang, Yibing Zhan, Shaoze Feng, Kejun Liu, and Zhe Chen. Pose-disentangled contrastive learning for self-supervised facial representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9717-9728, 2023. 1, 2, 3, 5, 6, 7 +[42] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, 2015. 2, 5, 6 +[43] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1557-1567, 2023. 3 +[44] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2022. 6 +[45] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Ma-hoor. Affectnet: A database for facial expression, valence, + +and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 6, 7 +[46] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-bert: Bert-based facial micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1492, 2023. 3 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[48] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 2, 5, 6 +[49] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. A semi-automatic methodology for facial landmark annotation. In 2013 IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 896-903, 2013. +[50] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: Database and results. Image and vision computing, 47:3-18, 2016. 2, 5, 6 +[51] Ankit Kumar Sharma and Hassan Foroosh. Slim-cnn: A light-weight cnn for face attribute prediction. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 329-335, 2020. 6 +[52] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11911-11920, 2021. 6 +[53] Yuxuan Shu, Xiao Gu, Guang-Zhong Yang, and Benny P L Lo. Revisiting self-supervised contrastive learning for facial expression recognition. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 3 +[54] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2132-2141, 2023. 1, 2, 3 +[55] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 1 +[56] James Thewlis, Samuel Albanie, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of landmarks by descriptor vector exchange. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6360-6370, 2019. 3 + +[57] Hao Wang, Min Li, Yangyang Song, Youjian Zhang, and Liying Chi. Ucol: Unsupervised learning of discriminative facial representations via uncertainty-aware contrast. Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):2510-2518, 2023. 3 +[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 1 +[59] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Liang Liu, Yabiao Wang, and Chengjie Wang. Toward high quality facial representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, page 5048-5058, New York, NY, USA, 2023. Association for Computing Machinery. 1, 3, 6, 7 +[60] Wenyan Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2138, 2018. 2, 5, 6 +[61] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4042-4051, 2022. 3, 7 +[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1 +[63] Jing Yang, Qingshan Liu, and Kaihua Zhang. Stacked hourglass network for robust facial landmark localisation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2025-2033, 2017. 2, 7 +[64] Sejong Yang, Subin Jeon, Seonghyeon Nam, and Seon Joo Kim. Dense interspecies face embedding. In Advances in Neural Information Processing Systems, pages 33275-33288. Curran Associates, Inc., 2022. 3 +[65] M. Ye, X. Zhang, P. C. Yuen, and S. Chang. Unsupervised embedding learning via invariant and spreading instance feature. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6203-6212, 2019. 1 +[66] Lijun Yin, Xiaozhou Wei, Yi Sun, Jun Wang, and Matthew J Rosato. A 3d facial expression database for facial behavior research. In 7th international conference on automatic face and gesture recognition (FGR06), pages 211-216. IEEE, 2006. 3 +[67] Xiang Zhang, Taoyue Wang, Xiaotian Li, Huiyuan Yang, and Lijun Yin. Weakly-supervised text-driven contrastive learning for facial behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20751-20762, 2023. 6 +[68] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object land + +marks as structural representations. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3 +[69] Yuhang Zhang, Chengrui Wang, and Weihong Deng. Relative uncertainty learning for facial expression recognition. In Advances in Neural Information Processing Systems, pages 17616-17627. Curran Associates, Inc., 2021. 6 +[70] Yuhang Zhang, Chengrui Wang, Xu Ling, and Weihong Deng. Learn from all: Erasing attention consistency for noisy label facial expression recognition. In Computer Vision - ECCV 2022, pages 418-434, Cham, 2022. Springer Nature Switzerland. 6, 7 +[71] Kaili Zhao, Wen-Sheng Chu, and Honggang Zhang. Deep region and multi-label learning for facial action unit detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3399, 2016. 1 +[72] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18676-18688, 2022. 1, 3, 5, 6, 7 +[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 4 +[74] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15475-15484, 2023. 6, 7 +[75] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 146-155, 2016. 3 \ No newline at end of file diff --git a/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/images.zip b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44eb31bd10521dd65bd71819691c647b15b51ca9 --- /dev/null +++ b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c75cc53b39f8441896693bb8fd0371770ede2fcf80d20fb5bfefa9c1106237b0 +size 378358 diff --git a/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/layout.json b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de930d0b552c012fe00a0a582636b84830d941f8 --- /dev/null +++ b/2024/Self-Supervised Facial Representation Learning with Facial Region Awareness/layout.json @@ -0,0 +1,9752 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 128, + 518, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 128, + 518, + 144 + ], + "spans": [ + { + "bbox": [ + 77, + 128, + 518, + 144 + ], + "type": "text", + "content": "Self-Supervised Facial Representation Learning with Facial Region Awareness" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 165, + 376, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 165, + 376, + 177 + ], + "spans": [ + { + "bbox": [ + 217, + 165, + 376, + 177 + ], + "type": "text", + "content": "Zheng Gao Ioannis Patras" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 190, + 451, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 190, + 451, + 203 + ], + "spans": [ + { + "bbox": [ + 144, + 190, + 451, + 203 + ], + "type": "text", + "content": "Queen Mary University of London, Mile End Road, London, E1 4NS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 227, + 205, + 365, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 205, + 365, + 216 + ], + "spans": [ + { + "bbox": [ + 227, + 205, + 365, + 216 + ], + "type": "text", + "content": "{z.gao, i.patras}@qmul.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 157, + 241, + 201, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 241, + 201, + 253 + ], + "spans": [ + { + "bbox": [ + 157, + 241, + 201, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 264, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 264, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 67, + 264, + 290, + 605 + ], + "type": "text", + "content": "Self-supervised pre-training has been proved to be effective in learning transferable representations that benefit various visual tasks. This paper asks this question: can self-supervised pre-training learn general facial representations for various facial analysis tasks? Recent efforts toward this goal are limited to treating each face image as a whole, i.e., learning consistent facial representations at the image-level, which overlooks the \"consistency of local facial representations\" (i.e., facial regions like eyes, nose, etc). In this work, we make a first attempt to propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). Specifically, we explicitly enforce the consistency of facial regions by matching the local facial representations across views, which are extracted with learned heatmaps highlighting the facial regions. Inspired by the mask prediction in supervised semantic segmentation, we obtain the heatmaps via cosine similarity between the per-pixel projection of feature maps and \"facial mask embeddings\" computed from learnable positional embeddings, which leverage the attention mechanism to globally look up the facial image for facial regions. To learn such heatmaps, we formulate the learning of facial mask embeddings as a deep clustering problem by assigning the pixel features from the feature maps to them. The transfer learning results on facial classification and regression tasks show that our FRA outperforms previous pre-trained models and more importantly, using ResNet as the unified backbone for various tasks, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 624, + 143, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 624, + 143, + 635 + ], + "spans": [ + { + "bbox": [ + 69, + 624, + 143, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 642, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 642, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 642, + 288, + 687 + ], + "type": "text", + "content": "Human face understanding is an important and challenging topic in computer vision [41, 72] and supervised learning algorithms have shown promising results on a wide range of facial analysis tasks recently [5, 31, 55, 71]. Despite" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 243, + 525, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 525, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 525, + 264 + ], + "type": "text", + "content": "the impressive progress, these methods require large-scale well-annotated training data, which is expensive to collect." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 267, + 526, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 526, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 526, + 519 + ], + "type": "text", + "content": "Recent works in self-supervised representation learning for visual images have shown that self-supervised pretraining is effective in improving the performance on various downstream tasks such as image classification, object detection and segmentation as it can learn general representations from unlabeled data that could be transferred to downstream visual tasks, especially tasks with limited labeled data [18, 24, 28, 30, 58, 62, 65]. Among them, instance discrimination (including contrastive learning [10, 22, 36] and non-contrastive learning [11, 20] paradigms) has been shown to be effective in learning generalizable self-supervised features. Instance discrimination aims to learn view-invariant representations by matching the global representations between the augmented views generated by aggressive image augmentations, i.e., the image-level representations of the augmented views should be similar [10-12, 20, 22, 36]. Another self-supervised learning paradigm, masked image modeling (MIM) [23, 54] learns visual representations by reconstructing image content from a masked image, achieving excellent performance in full model fine-tuning. This leads to the question: can self-supervised pre-training learn general facial representations which benefit downstream facial analysis tasks?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 526, + 687 + ], + "type": "text", + "content": "Several attempts have been made to learn general facial representations for facial analysis tasks [3, 41, 72]. For example, Bulat et al. [3] directly applies the contrastive objective to facial features. FaRL [72] and MCF [59] combine contrastive learning and mask image modeling [23]. PCL [41] proposes to disentangle the pose-related and pose-unrelated features, achieving strong performance on both pose-related (regression) and pose-unrelated (classification) tasks. However, it runs the model forward and backward three times for each training step, which is time-consuming. Despite different techniques, these methods commonly treats each face image as a whole to learn consistent global representations at the image-level and overlook the \"spatial consistency of local representations\", i.e., local facial regions (e.g., eyes, nose and mouth) should also be" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 34, + 479, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 34, + 479, + 67 + ], + "spans": [ + { + "bbox": [ + 159, + 34, + 479, + 67 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 705, + 313, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 313, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 313, + 714 + ], + "type": "text", + "content": "2081" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 100, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 100, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 68, + 100, + 288, + 144 + ], + "type": "text", + "content": "similar across the augmented views, thus limiting the generalization to downstream tasks. This brings us to the focus: learning consistent global and local representations for facial representation learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 147, + 290, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 147, + 290, + 289 + ], + "spans": [ + { + "bbox": [ + 68, + 147, + 290, + 289 + ], + "type": "text", + "content": "We argue that in order to learn consistent local representations, the model needs to look into facial regions. Towards that goal, we predict a set of heatmaps highlighting different facial regions by leveraging learnable positional embeddings as facial queries (the feature maps as keys and values) to look up the facial image globally for facial regions, which is inspired by the mask prediction in supervised segmentation [13]. For visual images, the attention mechanism of Transformer allows the learnable positional embeddings to serve as object queries for visual pattern look-up [7, 13]. In our case (facial images), the learnable positional embeddings can be used as facial queries for facial regions (see the visualization in the supplementary material)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 293, + 290, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 293, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 293, + 290, + 687 + ], + "type": "text", + "content": "In this work, taking the consistency of facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns general facial representations by enforcing consistent global and local facial representations, based on a popular instance discrimination baseline BYOL [20] for its simplicity. Specifically, we learn consistent local facial representations by match them across augmented views, which are extracted by aggregating the feature maps using learned heatmaps highlighting the facial regions as weights. Inspired by the mask prediction in MaskFormer [13], we produce the heatmaps from a set of learnable positional embeddings, which are used as facial queries to look up the facial image for facial regions. A Transformer decoder takes as input the feature maps from the encoder and the learnable positional embeddings to output a set of \"facial mask embeddings\", each associated with a facial region. The facial mask embeddings are used to compute cosine similarity with the per-pixel projection of feature maps to produce the heatmaps. In addition, we enforce the consistency of global representations across views simultaneously so that the image-level information is preserved. In order to learn the heatmaps (facial mask embeddings), inspired by deep clustering [8] that learns to assign samples to clusters, we treat the facial mask embeddings as clusters and learn to assign pixel features from the feature maps to them. Specifically, we align the per-pixel cluster assignments of each pixel feature over the facial region clusters between the online and momentum network for the same augmented view (i.e., each pixel feature should have similar similarity distribution over the facial mask embeddings between the momentum teacher and online student). In contrast to supervised segmentation that directly uses ground-truth masks to supervise the learning of the masks (heatmaps) with a per-pixel binary mask loss, we formulate the learning of heatmaps as a deep cluster" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 100, + 524, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 100, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 100, + 524, + 121 + ], + "type": "text", + "content": "ing [8] problem that learns to assign pixel features to clusters (facial mask embeddings) in a self-supervised manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 122, + 499, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 122, + 499, + 133 + ], + "spans": [ + { + "bbox": [ + 316, + 122, + 499, + 133 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 134, + 525, + 385 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 134, + 525, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 134, + 525, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 134, + 525, + 189 + ], + "type": "text", + "content": "- Taking into the consistency of local facial regions into account, we make a first attempt to propose a novel self-supervised facial representation learning framework, Facial Region Awareness (FRA) that learns consistent global and local facial representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 189, + 524, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 524, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 524, + 210 + ], + "type": "text", + "content": "- We show that the learned heatmaps can roughly discover facial regions in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 211, + 525, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 211, + 525, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 525, + 286 + ], + "type": "text", + "content": "- In previous works, different backbones are adopted for different facial analysis tasks (e.g., in face alignment the common backbone is hourglass network [63] while in facial expression recognition ResNet [21] is the popular backbone). In this work, our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 287, + 525, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 287, + 525, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 525, + 385 + ], + "type": "text", + "content": "- Our FRA outperforms existing self-supervised pretraining approaches (e.g., BYOL [20] and PCL [41]) on facial classification (i.e., facial expression recognition [19, 38] and facial attribute recognition [42]) and regression (i.e., face alignment [48-50, 60]) tasks. More importantly, our FRA achieves comparable (e.g., face alignment) or even better performance (e.g., facial expression recognition) compared with SOTA methods in the corresponding facial analysis tasks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 397, + 382, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 382, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 382, + 408 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 416, + 456, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 416, + 456, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 456, + 428 + ], + "type": "text", + "content": "2.1. Visual representation learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 434, + 525, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 525, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 525, + 620 + ], + "type": "text", + "content": "As one of the main paradigms for self-supervised pretraining, instance discrimination learns representations by treating an image as a whole and enforcing the consistency of global representations at the image-level across augmented views. Generally, instance discrimination includes two paradigms: contrastive learning [10, 22, 36] and non-contrastive learning [11, 20]. Contrastive learning considers each image and its transformations as a distinct class, i.e., \"positive\" samples are pulled together while \"negative\" samples are pushed apart in the latent space. Unlike contrastive learning that relies on negative samples to avoid collapse, non-contrastive learning directly maximizes the similarity of the global representations between the augmented views without involving negative samples based on techniques like stop-gradient [11] and predictor [20]. Further works perform visual-language pre-training by applying contrastive objective to image-text pairs [29, 35, 47]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 620, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 525, + 686 + ], + "type": "text", + "content": "Another line of work, masked image modeling (MIM) learns visual representations by reconstructing image content from a masked image [1, 23, 54], which is inspired by the masked language modeling in NLP [15]. In contrast to instance discrimination, MIM achieves strong full model fine-tuning performance with Vision Transformers" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "text", + "content": "2082" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 100, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 100, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 288, + 133 + ], + "type": "text", + "content": "pre-trained for enough epochs. However, these works suffer from poor linear separability and are less data-efficient than instance discrimination in few-shot scenarios [1]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 140, + 218, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 140, + 218, + 152 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 218, + 152 + ], + "type": "text", + "content": "2.2. Facial representation learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 157, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 157, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 69, + 157, + 290, + 563 + ], + "type": "text", + "content": "Recent works on facial analysis explore self-supervised learning for several face-related tasks, such as facial expression recognition [9, 53], face recognition [9, 57], facial micro-expression recognition [46], AU detection [39, 40], face alignment (facial landmark detection) [14, 64], etc. However, these methods are task-specific, i.e., tailored for a specific facial task and thus lack the ability to generalize to various facial analysis tasks [41]. Further efforts [3, 41, 72] focus on learning general facial representations with contrastive learning and mask image modeling [23, 54]. Bulat et al. [3] directly apply the contrastive objective to augmented views of the same face image, showing that general facial representation learned from pre-training can benefit various facial analysis tasks. FaRL [72] performs pre-training in a visual-linguistic manner by employing image-text contrastive learning and masked image modeling. MCF [59] leverages image-level contrastive learning and masked image modeling, along with the knowledge distilled from external ImageNet pre-trained model for facial representation learning. PCL [41] argues that directly applying the contrastive objective to face images overlooks the variances of facial poses and thus leads to pose-invariant representations, limiting the performance on pose-related tasks [66, 75]. Therefore, PCL [41] disentangles the pose-related and pose-unrelated features and then performs contrastive learning on these features, achieving strong performance on both pose-related and pose-unrelated facial analysis tasks. Despite the success, it performs forward and backward three times for each input image, which brings significant increase on computational cost. These works are commonly limited by instance discrimination and overlook the consistency of local facial regions. In contrast, inspired by supervised semantic segmentation, we learn consistent global and local facial representations by learning a set of heatmaps indicating facial regions from learnable positional embeddings, which leverage the attention mechanism to look up facial image globally for facial regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 570, + 188, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 570, + 188, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 188, + 582 + ], + "type": "text", + "content": "2.3. Facial region discovery" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 587, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 587, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 587, + 288, + 687 + ], + "type": "text", + "content": "There are some approaches leveraging facial region (landmark) discovery for facial analysis [27, 43, 61]. Some focus on landmark detection by either learning a heatmap for each landmark via image reconstruction [27, 68], or performing pixel-level matching with an equivariance loss [56, 68]. Despite different techniques, these methods are task-specific, i.e., landmark detection with discovery of local information, while our method is task-agnostic, i.e., learn general facial representations for various tasks by preserving" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 100, + 525, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 525, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 525, + 253 + ], + "type": "text", + "content": "global and local information in a image, region and pixel-level contrastive manner. MARLIN [4] applies masked image modeling to learn general representations for facial videos by utilizing an external face parsing algorithm to discover the facial regions (e.g., eyes, nose and mouth), which are used to guide the masking for the masked autoencoder. A closely related work SLPT [61] leverages the attention mechanism to estimate facial landmarks from initial facial landmarks estimates of the mean face through supervised learning. These works commonly rely on external supervisory signal, whether it is from ground-truths or additional algorithms. In contrast, we learn to discover the facial regions in an end-to-end self-supervised manner for facial image representation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 264, + 381, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 381, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 381, + 277 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 282, + 367, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 367, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 367, + 293 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 299, + 525, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 525, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 525, + 420 + ], + "type": "text", + "content": "The overview of the proposed FRA is shown in Fig. 1. The goal is to learn consistent global and local facial representations. Toward this goal, we propose two objectives: pixel-level semantic relation and image/region-level semantic consistency. Semantic relation aligns the per-pixel cluster assignments of each pixel feature over the facial mask embeddings between the online and momentum network to learn heatmaps for facial regions (Sec. 3.2) while semantic consistency directly matches the global and local facial representations across augmented views (Sec. 3.3) with the learned heatmaps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 427, + 402, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 402, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 402, + 438 + ], + "type": "text", + "content": "3.2. Semantic relation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": "As shown in Fig. 1, our method adopts the Siamese structure of BYOL [20], a popular self-supervised pretraining baseline based on instance discrimination. Following BYOL [20], we employ two branches: the online network parameterized by " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": " and the momentum network parameterized by " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": ". The online network " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": " consists of an encoder " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": ", a global projector " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "H_{\\theta}^{g}" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": " and a local projector " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "H_{\\theta}^{l}" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": ". The momentum network " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": " has the same architecture with the online network, except " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": " is updated with an exponential moving average of " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": ". As in BYOL [20], we also adopt additional predictors " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "G_{\\theta}^{g}" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "inline_equation", + "content": "G_{\\theta}^{l}" + }, + { + "bbox": [ + 304, + 445, + 525, + 587 + ], + "type": "text", + "content": " on top of the projectors in the online network. Note that this is omitted for brevity in Fig. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": ", two random augmentations are applied to generate two augmented views " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1 = \\mathcal{T}_1(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2 = \\mathcal{T}_2(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": ", following BYOL [20]. Each augmented view " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\{\\mathbf{x}_1, \\mathbf{x}_2\\}" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " is fed into an encoder " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " to obtain a feature map " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " (before global average pooling), where " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "C, H, W" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " are the number of channels, height and width of " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " and a latent representation " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i \\in \\{\\mathbf{h}_1, \\mathbf{h}_2\\}" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " (after global average pooling), i.e., " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_1 = E_{\\theta}(\\mathbf{x}_1)" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_2 = E_{\\xi}(\\mathbf{x}_2)" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": ". Then each latent representation " + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i" + }, + { + "bbox": [ + 304, + 588, + 525, + 687 + ], + "type": "text", + "content": " is transformed" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "text", + "content": "2083" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 98, + 501, + 268 + ], + "blocks": [ + { + "bbox": [ + 95, + 98, + 501, + 268 + ], + "lines": [ + { + "bbox": [ + 95, + 98, + 501, + 268 + ], + "spans": [ + { + "bbox": [ + 95, + 98, + 501, + 268 + ], + "type": "image", + "image_path": "4db6d3a6998f58ae0a8d163e71b2d81b1929bebbd4b443f85e50ee48549d111b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "lines": [ + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "spans": [ + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed FRA framework. " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": " denotes cosine similarity. For each input image " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": ", its augmented views " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": " are passed into two network branches to produce the global embeddings " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_2" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": ". In addition, we produce a set of heatmaps " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_1" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_2" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": " indicating the local facial regions, via the correlation between the pixel features and \"facial mask embeddings\" computed from a set of learnable positional embeddings. Then we aggregate the feature map to obtain the local facial embeddings " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_1^m\\}" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_2^m\\}" + }, + { + "bbox": [ + 67, + 276, + 527, + 348 + ], + "type": "text", + "content": ". The semantic consistency loss is applied to global embeddings and facial embeddings to maximize the similarity across augmented views. To learn such heatmaps, i.e., facial mask embeddings, we treat the facial mask embeddings as facial region clusters and propose a semantic relation loss to align the cluster assignments of each pixel feature over the facial region clusters between the online and momentum network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 365, + 289, + 478 + ], + "blocks": [ + { + "bbox": [ + 72, + 365, + 289, + 478 + ], + "lines": [ + { + "bbox": [ + 72, + 365, + 289, + 478 + ], + "spans": [ + { + "bbox": [ + 72, + 365, + 289, + 478 + ], + "type": "image", + "image_path": "12e98237741f2f7ca67d9f53d279a2aae367a7bf7b91549941c016c0bf2bdb16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 483, + 289, + 506 + ], + "lines": [ + { + "bbox": [ + 68, + 483, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 68, + 483, + 289, + 506 + ], + "type": "text", + "content": "Figure 2. Generation of heatmaps using learnable positional embeddings as facial queries and the feature maps as keys and values." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "text", + "content": "by a global projector " + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "inline_equation", + "content": "H^g" + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "text", + "content": " to produce a global embedding " + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\in \\{\\mathbf{z}_1,\\mathbf{z}_2\\}" + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "text", + "content": " , i.e., " + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1 = H_\\theta^g (\\mathbf{h}_1)" + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 524, + 288, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_2 = H_{\\xi}^{g}(\\mathbf{h}_{2})" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "spans": [ + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": "Then we obtain a set of heatmaps " + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i\\in \\{\\mathbf{M}_1,\\mathbf{M}_2\\}" + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": " highlighting the facial regions from the feature map " + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": " for each view, which is inspired by mask classification-based supervised segmentation [7, 13] that leverages attention mechanism to look up visual patterns globally. First the local projector (e.g., " + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "inline_equation", + "content": "H_{\\theta}^{l}" + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": ") is applied to project the pixel features of " + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": " in a pixel-wise manner, mapping it to " + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": " dimensions to get the dense feature map " + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{dense}}\\in \\mathbb{R}^{D\\times H\\times W}" + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": ". Take view " + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 68, + 556, + 290, + 666 + ], + "type": "text", + "content": " as an example, the projected feature map can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 673, + 289, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 673, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 116, + 673, + 289, + 688 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {1} ^ {\\text {d e n s e}} [ *, u, v ] = H _ {\\theta} ^ {l} \\left(\\mathbf {F} _ {1} [ *, u, v ]\\right), \\tag {1}", + "image_path": "e87ffc77535a93966268ff531e92548d108b9eca2c36ebc1494062f4fd30b12b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1[*,u,v]\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " is the pixel feature at the " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " -th grid of " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": ". Then as shown in Fig. 2, inspired by supervised segmentation [13], we use a Transformer decoder followed by a MLP, which takes as input the feature map " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " learnable positional embeddings (i.e., facial queries for looking up the facial image globally for facial regions) to predict " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " \"facial mask embeddings\" " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": ", where each row associated with a facial region. Next, we compute the cosine similarity between facial mask embeddings " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " and dense feature map " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{dense}}" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " along the channel dimension, yielding per-pixel cluster assignments " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i\\in \\mathbb{R}^{N\\times H\\times W}" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i[*,u,v]" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " denotes the relation between the dense pixel feature " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1^{\\mathrm{dense}}[*,u,v]" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " and facial mask embeddings " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": ". Finally, we normalize " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " along the channel dimension with a softmax operation to encourage each channel to capture a different pattern, obtaining " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " heatmaps " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i\\in \\mathbb{R}^{N\\times H\\times W}" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " where each vector at location " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " is a probabilistic similarity distribution (i.e., normalized per-pixel cluster assignments) " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_1^{u,v}" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{dense}}[*,u,v]" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " is a set of heatmaps where each channel of " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": " represents a 2D heatmap " + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^{(m)}" + }, + { + "bbox": [ + 304, + 365, + 526, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 598, + 525, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 525, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 525, + 674 + ], + "type": "text", + "content": "To learn such heatmaps, i.e., facial mask embeddings, inspired by deep clustering [8], we treat the facial mask embeddings as facial region clusters and align the per-pixel cluster assignments of each pixel feature over these clusters between the online and momentum network for the same augmented view, using the momentum network as momentum teacher [16, 73] to provide reliable target." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 675, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 675, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 675, + 525, + 686 + ], + "type": "text", + "content": "Following BYOL [20], we pass both augmented views" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "text", + "content": "2084" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "spans": [ + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " through the online and momentum network. Take " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " as an example, the online network " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " outputs the normalized per-pixel cluster assignments " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_1^{u,v}" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " and the momentum network outputs normalized assignments " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{s}}_1^{u,v}" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " for view " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": ". Then we learn " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_1^{u,v}" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{s}}_1^{u,v}" + }, + { + "bbox": [ + 68, + 100, + 289, + 166 + ], + "type": "text", + "content": " as guidance based on the following cross-entropy loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 174, + 289, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 174, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 93, + 174, + 289, + 205 + ], + "type": "interline_equation", + "content": "C E \\left(\\mathbf {s} _ {1} ^ {u, v}, \\widehat {\\mathbf {s}} _ {1} ^ {u, v}\\right) = - \\sum_ {m = 1} ^ {N} \\widehat {\\mathbf {s}} _ {1} ^ {u, v} [ m ] \\log \\mathbf {s} _ {1} ^ {u, v} [ m ]. \\tag {2}", + "image_path": "a76e1dc3d228664b7013e268e0cf97cbaa20b65ff6a66cd5962c22bdff78ee42.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 213, + 289, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 213, + 289, + 235 + ], + "spans": [ + { + "bbox": [ + 69, + 213, + 289, + 235 + ], + "type": "text", + "content": "For both augmented views, we define the symmetrized semantic relation objective as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 243, + 289, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 243, + 289, + 270 + ], + "spans": [ + { + "bbox": [ + 77, + 243, + 289, + 270 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r}} = \\frac {1}{H W} \\sum_ {u, v} \\left(C E \\left(\\mathbf {s} _ {1} ^ {u, v}, \\widehat {\\mathbf {s}} _ {1} ^ {u, v}\\right) + C E \\left(\\mathbf {s} _ {2} ^ {u, v}, \\widehat {\\mathbf {s}} _ {2} ^ {u, v}\\right)\\right), \\tag {3}", + "image_path": "37af934611cee5b5d46d7edd42b23eaaee57b5624a8a7b3da31eda53c09398fa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 278, + 289, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 278, + 289, + 345 + ], + "spans": [ + { + "bbox": [ + 68, + 278, + 289, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 278, + 289, + 345 + ], + "type": "inline_equation", + "content": "CE(\\mathbf{s}_2^{u,v},\\widehat{\\mathbf{s}}_2^{u,v})" + }, + { + "bbox": [ + 68, + 278, + 289, + 345 + ], + "type": "text", + "content": " is the cross-entropy loss for view " + }, + { + "bbox": [ + 68, + 278, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 68, + 278, + 289, + 345 + ], + "type": "text", + "content": ". We apply the Sinkhorn-Knopp normalization to the target assignments from the momentum network following [8] to avoid collapse and the mean entropy maximization (ME-MAX) regularizer [1] to maximize the entropy of the prediction to encourage full use of the clusters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 353, + 181, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 353, + 181, + 365 + ], + "spans": [ + { + "bbox": [ + 69, + 353, + 181, + 365 + ], + "type": "text", + "content": "3.3. Semantic consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 369, + 289, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 369, + 289, + 415 + ], + "spans": [ + { + "bbox": [ + 68, + 369, + 289, + 415 + ], + "type": "text", + "content": "In this section, we enforce the consistency of global embeddings and local facial embeddings. With the learned heatmaps " + }, + { + "bbox": [ + 68, + 369, + 289, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 68, + 369, + 289, + 415 + ], + "type": "text", + "content": ", we generate the latent representations for the local facial regions through weighted average pooling:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 423, + 288, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 423, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 76, + 423, + 288, + 464 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {h} _ {i} ^ {m} = \\mathbf {M} _ {i} ^ {(m)} \\otimes \\mathbf {F} _ {i} \\\\ = \\frac {1}{\\sum_ {u , v} \\mathbf {M} _ {i} [ m , u , v ]} \\sum_ {u, v} \\mathbf {M} _ {i} [ m, u, v ] \\mathbf {F} _ {i} [ *, u, v ], \\tag {4} \\\\ \\end{array}", + "image_path": "496b04dce07aa0d069970ec451eae117f4254a798764a4bd36b9fd9185d1aad0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": " denotes channel-wise weighted average pooling, " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^{(m)}" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": "-th channel (heatmap) of " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^m \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": " is the corresponding latent representation produced with " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^{(m)}" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": ". The facial embeddings " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_1^m : \\mathbf{z}_1^m \\in \\mathbb{R}^D\\}_{m=1}^N" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_2^m : \\mathbf{z}_2^m \\in \\mathbb{R}^D\\}_{m=1}^N" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": " are obtained accordingly with the local projector " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "H_\\theta^l" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "inline_equation", + "content": "H_\\xi^l" + }, + { + "bbox": [ + 69, + 473, + 289, + 544 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 148, + 552, + 288, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 552, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 148, + 552, + 288, + 577 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {z} _ {1} ^ {m} = H _ {\\theta} ^ {l} \\left(\\mathbf {h} _ {1} ^ {m}\\right), \\\\ \\mathbf {z} _ {2} ^ {m} = H _ {\\xi} ^ {l} \\left(\\mathbf {h} _ {2} ^ {m}\\right). \\end{array} \\tag {5}", + "image_path": "3e096a5fb8985544ad921e14a795ae0435ea9b2f1982fafc36f5103cfb73a4e7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 585, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 585, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 68, + 585, + 288, + 619 + ], + "type": "text", + "content": "We then match the global embeddings and local facial embeddings across views using the negative cosine similarity in BYOL [20]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 628, + 288, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 628, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 76, + 628, + 288, + 683 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}\\right) = - \\left(\\lambda_ {\\mathrm {c}} \\times f _ {\\mathrm {s}} \\left(G _ {\\theta} ^ {g} \\left(\\mathbf {z} _ {1}\\right), \\mathbf {z} _ {2}\\right) + \\right. \\\\ + \\left(1 - \\lambda_ {\\mathrm {c}}\\right) \\times \\frac {1}{N} \\sum_ {m = 1} ^ {N} f _ {\\mathrm {s}} \\left(G _ {\\theta} ^ {l} \\left(\\mathbf {z} _ {1} ^ {m}\\right), \\mathbf {z} _ {2} ^ {m}\\right)), \\tag {6} \\\\ \\end{array}", + "image_path": "3070760f20bcb73840680f31968b472646e84a685ed62c53a98b232ff2a60d32.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{s}}(\\mathbf{u},\\mathbf{v}) = \\frac{\\mathbf{u}^{\\top}\\mathbf{v}}{\\|\\mathbf{u}\\|_{2}\\|\\mathbf{v}\\|_{2}}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " denotes the cosine similarity between the vectors " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " is the loss weight, " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "G_{\\theta}^{g}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "G_{\\theta}^{l}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " are the predictors on top of the projectors " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "H_{\\theta}^{g}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "H_{\\theta}^{l}" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": ", respectively. Following BYOL [20], we symmetrize the loss " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sim}}(\\mathbf{z}_1,\\mathbf{z}_2)" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " defined in Eq. (6) by passing " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " through the momentum network " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " through the online network " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sim}}(\\mathbf{z}_2,\\mathbf{z}_1)" + }, + { + "bbox": [ + 304, + 99, + 526, + 190 + ], + "type": "text", + "content": ". The semantic consistency objective can be expressed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 199, + 525, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 199, + 525, + 212 + ], + "spans": [ + { + "bbox": [ + 351, + 199, + 525, + 212 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}\\right) + \\mathcal {L} _ {\\text {s i m}} \\left(\\mathbf {z} _ {2}, \\mathbf {z} _ {1}\\right). \\tag {7}", + "image_path": "00d6b6ad0bf1adfdbff6db58bc45ac55961d68d29a5b619c84f6c473cc1d7b3b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 220, + 400, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 400, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 400, + 232 + ], + "type": "text", + "content": "3.4. Overall objective" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 238, + 525, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 525, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 525, + 271 + ], + "type": "text", + "content": "We jointly optimize the semantic relation objective Eq. (3) and the semantic consistency objective Eq. (7), leading to the following overall objective:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 384, + 280, + 525, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 280, + 525, + 292 + ], + "spans": [ + { + "bbox": [ + 384, + 280, + 525, + 292 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}}, \\tag {8}", + "image_path": "637a63c8eaa49226c28d491037c19cf8dfac119e43be49106435d465153a9f49.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{r}}" + }, + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "type": "text", + "content": " is the loss weight for balancing " + }, + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 300, + 498, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{r}}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 323, + 379, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 379, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 379, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 342, + 414, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 414, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 414, + 354 + ], + "type": "text", + "content": "4.1. Experimental setups" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 358, + 426, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 426, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 426, + 369 + ], + "type": "text", + "content": "4.1.1 Implementation details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "text", + "content": "We use the same augmentation strategy as in [20, 25]. The number of heatmaps " + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "text", + "content": " is set to 8 empirically. The loss weight " + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{r}}" + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "inline_equation", + "content": "0.5 / 0.1" + }, + { + "bbox": [ + 304, + 376, + 525, + 443 + ], + "type": "text", + "content": ", respectively. For fair comparisons, the other hyper-parameters are kept the same as BYOL [20] in all experiments. The architecture and pretraining details are provided in the supplementary material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 457, + 372, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 372, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 372, + 467 + ], + "type": "text", + "content": "4.1.2 Baselines" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 475, + 525, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 525, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 525, + 594 + ], + "type": "text", + "content": "Our baselines are self-supervised pre-training approaches for visual images (e.g., BYOL [20] and LEWEL [25]), and pre-training approaches for facial images (e.g., Bulat et al. [3] and PCL [41]). Note that SwAV [8] is equivalent to Bulat et al. [3]. As we adopt BYOL [20] as the pretraining backbone, we compare our FRA with BYOL [20] in all experiments. We also compare our FRA with another pre-training method LEWEL [25], which learns consistent local representations for visual images. Moreover, we perform comparisons with SOTA methods in the corresponding downstream tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 603, + 415, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 415, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 415, + 615 + ], + "type": "text", + "content": "4.2. Evaluation protocols" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 620, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 525, + 687 + ], + "type": "text", + "content": "Following the common practice in previous works [41, 72], we evaluate the transfer performance of the self-supervised pre-trained facial representations on several popular downstream facial analysis tasks: facial expression recognition (FER) [2, 38], facial attribute recognition (FAR) [42] and face alignment (FA) [48-50, 60]. Specifically, we use" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "text", + "content": "2085" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 126, + 513, + 200 + ], + "blocks": [ + { + "bbox": [ + 69, + 98, + 525, + 119 + ], + "lines": [ + { + "bbox": [ + 69, + 98, + 525, + 119 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 525, + 119 + ], + "type": "text", + "content": "Table 1. Comparisons with weakly-supervised pre-trained vision transformer on several downstream facial analysis tasks, including facial expression recognition (AffectNet), facial attribute recognition (CelebA) and face alignment (300W)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 126, + 513, + 200 + ], + "lines": [ + { + "bbox": [ + 82, + 126, + 513, + 200 + ], + "spans": [ + { + "bbox": [ + 82, + 126, + 513, + 200 + ], + "type": "table", + "html": "
MethodArch.Params.Pre-training settingsDownstream performances
DatasetScaleSupervisionAffectNet Acc. ↑CelebA Acc. ↑300W NME ↓
FaRL [72]ViT-B/16 [17]86MLAION-FACE [72]20Mface image + text64.8591.883.08
FRAR50 [21]24MVGGFace2 [6]3.3Mface image66.1692.022.91
", + "image_path": "fc3864c927a80c34fc398a856e0110025bd0ec861ba9b11b5fb3f691daffba3a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 254, + 290, + 462 + ], + "blocks": [ + { + "bbox": [ + 69, + 216, + 288, + 246 + ], + "lines": [ + { + "bbox": [ + 69, + 216, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 69, + 216, + 288, + 246 + ], + "type": "text", + "content": "Table 2. Comparisons on facial expression recognition. We report the Top-1 accuracy on test set. Text denotes text supervision.†: our reproduction using the official codes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 254, + 290, + 462 + ], + "lines": [ + { + "bbox": [ + 70, + 254, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 70, + 254, + 290, + 462 + ], + "type": "table", + "html": "
MethodTextFERPlusRAF-DBAffectNet
Supervised
KTN [32]×90.4988.0763.97
RUL [69]×88.7588.9861.43
EAC [70]×90.0590.3565.32
Weakly-Supervised
FaRL [72]†✓88.6288.3164.85
CLEF [67]✓89.7490.0965.66
Self-supervised
MCF [59]†×88.1786.8660.98
Bulat et al. [3, 8]×--60.20
BYOL [20]×89.2589.5365.65
LEWEL [25]×85.6181.8561.20
PCL [41]×85.8785.9260.77
FRA (LP)×78.1373.8957.38
FRA (FT)×89.7889.9566.16
FRA (EAC)×90.6290.7665.85
", + "image_path": "c4387200e8da7f4b2db8832951e592bb26971a8991dd23819c6892669337606f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 484, + 288, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 484, + 288, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 484, + 288, + 550 + ], + "type": "text", + "content": "the pre-trained weights to initialize the backbone of downstream tasks and then learn the backbone and task-specific head (attached to the backbone) jointly. Following [72], we report the performance with linear probe (denoted by \"LP\") and fine-tuning (denoted by \"FT\"). The details of the downstream tasks are described as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 552, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 552, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 67, + 552, + 288, + 641 + ], + "type": "text", + "content": "Facial expression recognition is a multi-class classification task where the goal is to categorize the emotional expressions (e.g., anger, fear and surprise) for a given face image. Three widely-used datasets are adopted: FERPlus [2], RAF-DB [38] and AffectNet [45]. For RAF-DB, we use the basic emotion subset following [32, 41, 70]. For AffectNet, we report the results with 7 emotion classes (i.e., neutral, happy, sad, surprise, fear, anger, disgust) following [32, 70]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 643, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 643, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 67, + 643, + 289, + 686 + ], + "type": "text", + "content": "Facial attribute recognition is a multi-label classification task to predict various attributes (e.g., gender, age and race) of a given face image. We adopt the popular benchmark CelebA [42], which consists of more than 200K face" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 353, + 254, + 479, + 436 + ], + "blocks": [ + { + "bbox": [ + 305, + 216, + 525, + 247 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 525, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 525, + 247 + ], + "type": "text", + "content": "Table 3. Comparisons on CelebA [42] facial attribute recognition. We report the averaged accuracy over all attributes. †: our reproduction using the official codes. *: results cited from [72]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 254, + 479, + 436 + ], + "lines": [ + { + "bbox": [ + 353, + 254, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 353, + 254, + 479, + 436 + ], + "type": "table", + "html": "
MethodAcc. (%)
Supervised
DMM [44]91.70
SlimCNN [51]91.24
AFFAIR [34]91.45
Self-supervised
SSPL [52]91.77
Bulat et al. [3, 8]*89.65
SimCLR [10]*91.08
BYOL [20]91.56
LEWEL [25]90.69
PCL [41]91.48
MCF [59]†91.33
FRA (LP)90.86
FRA (FT)92.02
", + "image_path": "1f87af4b59b60e2e6694e61b4e84346742a27bfd36b0cb63a9eeb51a98d8c3dd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 453, + 524, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 524, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 524, + 475 + ], + "type": "text", + "content": "images with 40 facial attributes per image. Following [72], we report the averaged accuracy over all attributes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 476, + 525, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 525, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 525, + 563 + ], + "type": "text", + "content": "Face alignment is a regression task to predict 2D face landmark coordinates on a face image. We use two popular benchmarks: WFLW [60] and 300W [48-50]. Following the common practice [14, 26, 74], we report normalized mean error (NME), failure rate (FR) and AUC. For 300W, we report the results on full test set, common (554 images) and challenge (135 images) splits of the test set following [26, 74]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 571, + 525, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 571, + 525, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 525, + 594 + ], + "type": "text", + "content": "4.3. Comparisons with weakly-supervised pretraining" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 598, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 525, + 687 + ], + "type": "text", + "content": "In Tab. 1, we compare our FRA with SOTA pre-trained Transformer FaRL [72], which is a weakly-supervised model pre-trained on 20M visual-linguistic data (face image and text) with image-text contrastive learning and mask image modeling. We fine-tune both the pre-trained feature backbone and the task-specific head on the corresponding downstream facial analysis task. Our self-supervised FRA with 24M parameter ResNet-50 achieves superior perfor" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "text", + "content": "2086" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 117, + 532, + 302 + ], + "blocks": [ + { + "bbox": [ + 152, + 98, + 441, + 109 + ], + "lines": [ + { + "bbox": [ + 152, + 98, + 441, + 109 + ], + "spans": [ + { + "bbox": [ + 152, + 98, + 441, + 109 + ], + "type": "text", + "content": "Table 4. Comparisons on face alignment. †: our reproduction using the official codes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 117, + 532, + 302 + ], + "lines": [ + { + "bbox": [ + 70, + 117, + 532, + 302 + ], + "spans": [ + { + "bbox": [ + 70, + 117, + 532, + 302 + ], + "type": "table", + "html": "
MethodVenueArch.WFLW300W (NME ↓)
NME ↓FR10% ↓AUC10% ↑FullComm.Chal.
Supervised
SLPT [61][CVPR'22]ResNet [21]4.203.040.5883.202.784.93
DTLD [33][CVPR'22]ResNet [21]4.082.76-2.962.594.50
RePFormer [37][IJCAI'22]ResNet [21]4.11--3.01--
ADNet [26][ICCV'21]Hourglass [63]4.142.720.6022.932.534.58
STAR [74][CVPR'23]Hourglass [63]4.022.320.6052.872.524.32
Self-supervised
MCF [59] (concurrent work)[ACMM'M'23]ViT [17]3.961.400.6092.982.604.51
Bulat et al. [3, 8][ECCV'22]ResNet [21]4.57--3.20--
BYOL [20][NeurIPS'20]ResNet [21]4.292.960.5793.032.664.55
LEWEL [25][CVPR'22]ResNet [21]4.524.500.5633.092.704.71
PCL [41]†[CVPR'23]ResNet [21]4.846.180.5353.352.775.12
FRAOursResNet [21]4.112.530.5912.912.604.46
", + "image_path": "8eeb4f22084a5ced5d065f40354f57d244db8603a83d1ad70b53c94f462f46fb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 320, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 320, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 69, + 320, + 288, + 342 + ], + "type": "text", + "content": "mance compared with weakly-supervised FaRL [72] with 86M parameter ViT-B/16 and text supervision on all tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 359, + 165, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 359, + 165, + 371 + ], + "spans": [ + { + "bbox": [ + 69, + 359, + 165, + 371 + ], + "type": "text", + "content": "4.4. Transfer learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 380, + 288, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 68, + 380, + 288, + 423 + ], + "type": "text", + "content": "In this section, we compare our FRA with self-supervised pre-training approaches and SOTA methods in several downstream tasks. Please refer to the supplementary material for setup details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 455, + 212, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 455, + 212, + 466 + ], + "spans": [ + { + "bbox": [ + 69, + 455, + 212, + 466 + ], + "type": "text", + "content": "4.4.1 Facial expression recognition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 478, + 290, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 478, + 290, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 478, + 290, + 686 + ], + "type": "text", + "content": "The results on facial expression recognition are reported in Tab. 2. We observe: (1) With the setting of fine-tuning (FT), our FRA outperforms previous self-supervised pre-training approaches for visual images (e.g., BYOL and LEWEL) and pre-training approaches tailored for facial images (e.g., PCL, MCF). In particular, our FRA using a 24M parameter ResNet-50 surpasses the concurrent work MCF [59] with 86M parameter ViT-B/16 [17]. (2) By simply learning a linear classifier on top of the encoder backbone, our FRA outperforms SOTA facial expression recognition methods with sophisticated designs (e.g., EAC [70]) on AffectNet [45], the largest facial expression recognition dataset. (3) More importantly, by using our pre-trained model to initialize the backbone of SOTA facial expression recognition method EAC [70], our variant \"FRA (EAC)\" consistently improves EAC [70] on all datasets, which suggests \"FRA (EAC)\" outperforms SOTA FER methods and demonstrates the superiority of the proposed self-supervised pre-training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 320, + 443, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 443, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 443, + 331 + ], + "type": "text", + "content": "4.4.2 Facial attribute recognition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 338, + 525, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 525, + 404 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 525, + 404 + ], + "type": "text", + "content": "As shown in Tab. 3, our FRA outperforms both self-supervised pre-training approaches for visual images and pre-training approaches tailored for facial images. The results on facial expression recognition and facial attribute recognition show that our FRA learns better facial representations for facial classification task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 421, + 395, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 395, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 395, + 432 + ], + "type": "text", + "content": "4.4.3 Face alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 439, + 525, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 525, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 525, + 559 + ], + "type": "text", + "content": "As shown in Tab. 4, despite SOTA face alignment methods (e.g., ADNet [26] and STAR [74]) commonly rely on hourglass network [63] for feature extraction, which is tailored for regression tasks like landmark detection, our method based on ResNet backbone achieves comparable performance with these SOTA methods (e.g., 2.91 vs. 2.87 on 300W). The results on classification (e.g., facial expression recognition) and regression tasks (e.g., face alignment) show that our FRA achieves SOTA results using vanilla ResNet [21] as the unified backbone for various facial analysis tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 568, + 395, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 568, + 395, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 568, + 395, + 578 + ], + "type": "text", + "content": "4.5. Ablation studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 586, + 525, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 525, + 618 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 525, + 618 + ], + "type": "text", + "content": "We pre-train the model on VGGFace2 [6] and then evaluate it on facial expression recognition (RAF-DB) and facial attribute recognition (CelebA), as described in Sec. 4.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 635, + 440, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 440, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 440, + 645 + ], + "type": "text", + "content": "4.5.1 Effect of different modules" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 653, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 525, + 686 + ], + "type": "text", + "content": "In Tab. 5, we investigate the contributions of the proposed semantic consistency loss (i.e., global consistency of whole face and local consistency of facial regions) and semantic" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "text", + "content": "2087" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 146, + 282, + 224 + ], + "blocks": [ + { + "bbox": [ + 69, + 98, + 288, + 139 + ], + "lines": [ + { + "bbox": [ + 69, + 98, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 288, + 139 + ], + "type": "text", + "content": "Table 5. Effect of different modules. GC denotes the global consistency for aligning images, LC denotes the local consistency for aligning facial regions with heatmaps and SR represents semantic relation for aligning pixels and heatmaps." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 146, + 282, + 224 + ], + "lines": [ + { + "bbox": [ + 75, + 146, + 282, + 224 + ], + "spans": [ + { + "bbox": [ + 75, + 146, + 282, + 224 + ], + "type": "table", + "html": "
GCLCSRRAF-DB ↑CelebA ↑300W ↓
✓--87.8290.633.56
--✓85.3490.663.25
-✓-87.4690.783.19
✓✓-88.0590.893.26
✓✓✓88.7291.183.14
", + "image_path": "cb7a6e19e300c60c00cfe8c8ab9f35cfaec370acda28c91eb8bac7d839a33e90.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 108, + 256, + 249, + 302 + ], + "blocks": [ + { + "bbox": [ + 102, + 238, + 254, + 248 + ], + "lines": [ + { + "bbox": [ + 102, + 238, + 254, + 248 + ], + "spans": [ + { + "bbox": [ + 102, + 238, + 254, + 248 + ], + "type": "text", + "content": "Table 6. Effect of the number of heatmaps " + }, + { + "bbox": [ + 102, + 238, + 254, + 248 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 102, + 238, + 254, + 248 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 256, + 249, + 302 + ], + "lines": [ + { + "bbox": [ + 108, + 256, + 249, + 302 + ], + "spans": [ + { + "bbox": [ + 108, + 256, + 249, + 302 + ], + "type": "table", + "html": "
N83264
RAF-DB88.7288.3688.18
CelebA91.1891.0190.98
", + "image_path": "0a098eb2653778887895f4b4fb92d31aa0714062d0306cdfc1093bfadc4190c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 73, + 344, + 286, + 422 + ], + "blocks": [ + { + "bbox": [ + 69, + 315, + 288, + 335 + ], + "lines": [ + { + "bbox": [ + 69, + 315, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 69, + 315, + 288, + 335 + ], + "type": "text", + "content": "Table 7. Effect of the loss weights. Please refer to Tab. 5 for GC, LC and SR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 344, + 286, + 422 + ], + "lines": [ + { + "bbox": [ + 73, + 344, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 73, + 344, + 286, + 422 + ], + "type": "table", + "html": "
SettingsλcλrRAF-DBCelebA
GC1.0087.8290.63
GC + LC0.5088.0590.89
GC + LC + SR (FRA)0.50.188.7291.18
GC + LC + SR (FRA)0.50.588.4591.04
GC + LC + SR (FRA)0.51.088.0890.46
", + "image_path": "13930d0d44b24126d33ca7d88c729f4a20c51cf4c94345d58fd7db899616e067.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 444, + 288, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 444, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 68, + 444, + 288, + 542 + ], + "type": "text", + "content": "relation loss to our approach. Note that the global consistency (first row) is BYOL [20]. We have the following observations: (1) The variant using all losses achieves the best results. (2) GC is essential to avoid degeneration on classification task. (3) LC or SR alone benefits regression task (landmark). Altogether, LC and SR improve BYOL [20] (GC) on both classification and regression by capturing spatial/local information, which validates our facial region awareness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 567, + 230, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 567, + 230, + 578 + ], + "spans": [ + { + "bbox": [ + 69, + 567, + 230, + 578 + ], + "type": "text", + "content": "4.5.2 Effect of the number of heatmaps" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 587, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 587, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 68, + 587, + 288, + 685 + ], + "type": "text", + "content": "In Tab. 6, we study the effect of the number of heatmaps. We observe that the best setting is " + }, + { + "bbox": [ + 68, + 587, + 288, + 685 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 68, + 587, + 288, + 685 + ], + "type": "text", + "content": ", which is close to the facial landmarks number 5. This suggests that given enough face images for training, a suitable " + }, + { + "bbox": [ + 68, + 587, + 288, + 685 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 68, + 587, + 288, + 685 + ], + "type": "text", + "content": " can encourage the model to learn face-specific patterns, which helps the transfer learning performance on various facial analysis tasks. Further increasing the number of heatmaps might force the model to look into fine-grained patterns that may not be suitable for facial tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 100, + 417, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 100, + 417, + 110 + ], + "spans": [ + { + "bbox": [ + 306, + 100, + 417, + 110 + ], + "type": "text", + "content": "4.5.3 Effect of loss weights" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "text", + "content": "In Tab. 7, we ablate the weights for the semantic consistency loss and semantic relation loss. We find that setting " + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{c}} = 0.5" + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{r}} = 0.1" + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "text", + "content": " works best. When " + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{c}} = 1.0" + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{r}} = 0" + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "text", + "content": ", only the consistency of global representations is applied, and the model performs relatively worse, which suggests the importance of the consistency of local representations and the semantic relation loss. By using the semantic relation objective, the performance is significantly improved. However, when " + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{r}}" + }, + { + "bbox": [ + 304, + 117, + 525, + 248 + ], + "type": "text", + "content": " is too high, the performance degrades as the pixel-level consistency between the online and momentum network might affect the capture of image/object-level information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 264, + 479, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 479, + 275 + ], + "type": "text", + "content": "4.5.4 Effect of Transformer decoder layers" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 318, + 330, + 514, + 375 + ], + "blocks": [ + { + "bbox": [ + 304, + 291, + 525, + 321 + ], + "lines": [ + { + "bbox": [ + 304, + 291, + 525, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 525, + 321 + ], + "type": "text", + "content": "Table 8. Effect of Transformer decoder layers. 0 decoder layer represents BYOL [20] where only the consistency of global representation is enforced." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 330, + 514, + 375 + ], + "lines": [ + { + "bbox": [ + 318, + 330, + 514, + 375 + ], + "spans": [ + { + "bbox": [ + 318, + 330, + 514, + 375 + ], + "type": "table", + "html": "
# decoder layer0123
RAF-DB87.8288.7289.0189.06
CelebA90.6391.1891.3091.35
", + "image_path": "5cfe6ac8ce5db9e46595577460fba60480a1351aa29350609cee8b36ae22dc8d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 386, + 525, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 525, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 525, + 463 + ], + "type": "text", + "content": "In Tab. 8, we study the effect of the number of decoder layers used for heatmap prediction. We observe that a single decoder layer is able to produce decent results, showing that a 1-layer decoder is large enough to capture the facial region (landmarks) relations in face images. The performance gain diminishes as the decoder depth increases. By default, we only use 1 decoder layer for fast training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 474, + 372, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 372, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 372, + 485 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 492, + 525, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 525, + 624 + ], + "type": "text", + "content": "In this work, we propose a novel self-supervised facial representation learning framework to learn consistent global and local facial representations, Facial Region Awareness (FRA). We learn a set of heatmaps indicating facial regions from learnable positional embeddings, which leverages the attention mechanism to look up facial image globally for facial regions. We show that our FRA outperforms previous pre-trained models on several facial classification and regression tasks. More importantly, using ResNet as the unified backbone, our FRA achieves comparable or even better performance compared with SOTA methods in facial analysis tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 635, + 394, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 394, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 394, + 647 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 525, + 686 + ], + "type": "text", + "content": "This work was supported by the EU H2020 AI4Media No.951911 project. We thank Zengqun Zhao for his helpful comments on facial expression recognition." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "text", + "content": "2088" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 117, + 289, + 685 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 73, + 117, + 289, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 117, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 73, + 117, + 289, + 167 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 168, + 289, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 168, + 289, + 229 + ], + "spans": [ + { + "bbox": [ + 73, + 168, + 289, + 229 + ], + "type": "text", + "content": "[2] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM International Conference on Multimodal Interaction, page 279-283, New York, NY, USA, 2016. Association for Computing Machinery. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 231, + 289, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 231, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 73, + 231, + 289, + 281 + ], + "type": "text", + "content": "[3] Adrian Bulat, Shiyang Cheng, Jing Yang, Andrew Garbett, Enrique Sanchez, and Georgios Tzimiropoulos. Pre-training strategies and datasets for facial representation learning. In Computer Vision - ECCV 2022, pages 107-125, Cham, 2022. Springer Nature Switzerland. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 282, + 289, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 282, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 73, + 282, + 289, + 342 + ], + "type": "text", + "content": "[4] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. Marlin: Masked autoencoder for facial video representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1493-1504, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 344, + 289, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 344, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 73, + 344, + 289, + 395 + ], + "type": "text", + "content": "[5] Jiajiong Cao, Yingming Li, and Zhongfei Zhang. Partially shared multi-task convolutional neural network with local constraint for face attribute learning. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4290-4299, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 396, + 289, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 396, + 289, + 446 + ], + "spans": [ + { + "bbox": [ + 73, + 396, + 289, + 446 + ], + "type": "text", + "content": "[6] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 449, + 289, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 449, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 73, + 449, + 289, + 498 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020, pages 213-229, Cham, 2020. Springer International Publishing. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 500, + 289, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 500, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 73, + 500, + 289, + 550 + ], + "type": "text", + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 73, + 552, + 289, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 552, + 289, + 593 + ], + "spans": [ + { + "bbox": [ + 73, + 552, + 289, + 593 + ], + "type": "text", + "content": "[9] Jia-Ren Chang, Yong-Sheng Chen, and Wei-Chen Chiu. Learning facial representations from the cycle-consistency of face. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9660-9669, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 594, + 289, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 594, + 289, + 643 + ], + "spans": [ + { + "bbox": [ + 70, + 594, + 289, + 643 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 646, + 289, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 646, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 70, + 646, + 289, + 685 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 524, + 685 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 101, + 524, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 101, + 524, + 131 + ], + "spans": [ + { + "bbox": [ + 307, + 101, + 524, + 131 + ], + "type": "text", + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 133, + 524, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 133, + 524, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 133, + 524, + 172 + ], + "type": "text", + "content": "[13] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems, pages 17864-17875. Curran Associates, Inc., 2021. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 524, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 524, + 214 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 524, + 214 + ], + "type": "text", + "content": "[14] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9877-9886, 2021. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 215, + 524, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 215, + 524, + 305 + ], + "spans": [ + { + "bbox": [ + 307, + 215, + 524, + 305 + ], + "type": "text", + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 307, + 524, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 307, + 524, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 307, + 524, + 357 + ], + "type": "text", + "content": "[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In Computer Vision - ECCV 2022, pages 247-264, Cham, 2022. Springer Nature Switzerland. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 359, + 524, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 359, + 524, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 359, + 524, + 429 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 430, + 524, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 524, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 524, + 480 + ], + "type": "text", + "content": "[18] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 482, + 524, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 524, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 524, + 563 + ], + "type": "text", + "content": "[19] Ian J Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, et al. Challenges in representation learning: A report on three machine learning contests. In Neural Information Processing: 20th International Conference, ICONIP 2013, Daegu, Korea, November 3-7, 2013. Proceedings, Part III 20, pages 117-124. Springer, 2013. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 564, + 524, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 524, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 524, + 643 + ], + "type": "text", + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "type": "text", + "content": "[21] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 6, 7" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2089" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 686 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 100, + 289, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 100, + 289, + 140 + ], + "type": "text", + "content": "[22] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 142, + 289, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 289, + 191 + ], + "type": "text", + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 194, + 289, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 194, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 70, + 194, + 289, + 243 + ], + "type": "text", + "content": "[24] Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Philip Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR 2019. ICLR, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 244, + 289, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 244, + 289, + 295 + ], + "spans": [ + { + "bbox": [ + 70, + 244, + 289, + 295 + ], + "type": "text", + "content": "[25] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14431-14440, 2022. 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 297, + 289, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 297, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 70, + 297, + 289, + 346 + ], + "type": "text", + "content": "[26] Yangyu Huang, Hao Yang, Chong Li, Jongyoo Kim, and Fangyun Wei. Adnet: Leveraging error-bias towards normal direction in face alignment. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3060-3070, 2021. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 347, + 289, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 347, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 70, + 347, + 289, + 397 + ], + "type": "text", + "content": "[27] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. In Advances in Neural Information Processing Systems, pages 4016-4027. Curran Associates, Inc., 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 399, + 289, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 399, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 70, + 399, + 289, + 439 + ], + "type": "text", + "content": "[28] X. Ji, A. Vedaldi, and J. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9864-9873, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 441, + 289, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 441, + 289, + 499 + ], + "spans": [ + { + "bbox": [ + 70, + 441, + 289, + 499 + ], + "type": "text", + "content": "[29] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 501, + 289, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 501, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 70, + 501, + 289, + 552 + ], + "type": "text", + "content": "[30] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Advances in Neural Information Processing Systems, pages 18661-18673. Curran Associates, Inc., 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 553, + 289, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 553, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 70, + 553, + 289, + 613 + ], + "type": "text", + "content": "[31] Abhinav Kumar, Tim K. Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8233-8243, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 614, + 289, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 614, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 70, + 614, + 289, + 654 + ], + "type": "text", + "content": "[32] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via c-f labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 655, + 289, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 655, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 70, + 655, + 289, + 686 + ], + "type": "text", + "content": "[33] Hui Li, Zidong Guo, Seon-Min Rhee, Seungju Han, and Jae-Joon Han. Towards accurate facial landmark detection via cascaded transformers. In 2022 IEEE/CVF Conference" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 524, + 686 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 325, + 101, + 524, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 101, + 524, + 120 + ], + "spans": [ + { + "bbox": [ + 325, + 101, + 524, + 120 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 4166-4175, 2022. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 121, + 524, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 121, + 524, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 121, + 524, + 160 + ], + "type": "text", + "content": "[34] Jianshu Li, Fang Zhao, Jiashi Feng, Sujoy Roy, Shuicheng Yan, and Terence Sim. Landmark free face attribute prediction. IEEE Transactions on Image Processing, 27(9):4651-4662, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 162, + 524, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 524, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 524, + 220 + ], + "type": "text", + "content": "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 222, + 524, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 524, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 524, + 261 + ], + "type": "text", + "content": "[36] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 262, + 524, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 524, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 524, + 331 + ], + "type": "text", + "content": "[37] Jinpeng Li, Haibo Jin, Shengcai Liao, Ling Shao, and Pheng-Ann Heng. Repformer: Refinement pyramid transformer for robust facial landmark detection. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 1088–1094. International Joint Conferences on Artificial Intelligence Organization, 2022. Main Track. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 333, + 524, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 524, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 524, + 382 + ], + "type": "text", + "content": "[38] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2584-2593, 2017. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 384, + 524, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 524, + 433 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 524, + 433 + ], + "type": "text", + "content": "[39] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Self-supervised representation learning from videos for facial action unit detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10916-10925, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 434, + 524, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 524, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 524, + 473 + ], + "type": "text", + "content": "[40] Yong Li, Jiabei Zeng, and Shiguang Shan. Learning representations for facial actions from unlabeled videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):302-317, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 475, + 524, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 524, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 524, + 533 + ], + "type": "text", + "content": "[41] Yuanyuan Liu, Wenbin Wang, Yibing Zhan, Shaoze Feng, Kejun Liu, and Zhe Chen. Pose-disentangled contrastive learning for self-supervised facial representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9717-9728, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 535, + 524, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 524, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 524, + 574 + ], + "type": "text", + "content": "[42] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, 2015. 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 575, + 524, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 575, + 524, + 625 + ], + "spans": [ + { + "bbox": [ + 307, + 575, + 524, + 625 + ], + "type": "text", + "content": "[43] Yunfan Lu, Zipeng Wang, Minjie Liu, Hongjian Wang, and Lin Wang. Learning spatial-temporal implicit neural representations for event-guided video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1557-1567, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 626, + 524, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 626, + 524, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 626, + 524, + 665 + ], + "type": "text", + "content": "[44] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "type": "text", + "content": "[45] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Ma-hoor. Affectnet: A database for facial expression, valence," + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2090" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 288, + 685 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 88, + 100, + 288, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 100, + 288, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 100, + 288, + 121 + ], + "type": "text", + "content": "and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 122, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 122, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 71, + 122, + 288, + 172 + ], + "type": "text", + "content": "[46] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-bert: Bert-based facial micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1492, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 173, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 173, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 70, + 173, + 288, + 243 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 245, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 245, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 70, + 245, + 288, + 296 + ], + "type": "text", + "content": "[48] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In 2013 IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 297, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 297, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 70, + 297, + 288, + 346 + ], + "type": "text", + "content": "[49] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. A semi-automatic methodology for facial landmark annotation. In 2013 IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 896-903, 2013." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 347, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 347, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 70, + 347, + 288, + 388 + ], + "type": "text", + "content": "[50] Christos Sagonas, Epameinondas Antonakos, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: Database and results. Image and vision computing, 47:3-18, 2016. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 388, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 388, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 70, + 388, + 288, + 430 + ], + "type": "text", + "content": "[51] Ankit Kumar Sharma and Hassan Foroosh. Slim-cnn: A light-weight cnn for face attribute prediction. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 329-335, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 430, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 430, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 70, + 430, + 288, + 480 + ], + "type": "text", + "content": "[52] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11911-11920, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 481, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 481, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 70, + 481, + 288, + 532 + ], + "type": "text", + "content": "[53] Yuxuan Shu, Xiao Gu, Guang-Zhong Yang, and Benny P L Lo. Revisiting self-supervised contrastive learning for facial expression recognition. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 533, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 533, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 70, + 533, + 288, + 592 + ], + "type": "text", + "content": "[54] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2132-2141, 2023. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 594, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 594, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 70, + 594, + 288, + 635 + ], + "type": "text", + "content": "[55] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 635, + 288, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 635, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 70, + 635, + 288, + 685 + ], + "type": "text", + "content": "[56] James Thewlis, Samuel Albanie, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of landmarks by descriptor vector exchange. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6360-6370, 2019. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 524, + 686 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 101, + 524, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 101, + 524, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 101, + 524, + 150 + ], + "type": "text", + "content": "[57] Hao Wang, Min Li, Yangyang Song, Youjian Zhang, and Liying Chi. Ucol: Unsupervised learning of discriminative facial representations via uncertainty-aware contrast. Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):2510-2518, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 152, + 524, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 524, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 524, + 201 + ], + "type": "text", + "content": "[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 203, + 524, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 203, + 524, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 203, + 524, + 263 + ], + "type": "text", + "content": "[59] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Liang Liu, Yabiao Wang, and Chengjie Wang. Toward high quality facial representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, page 5048-5058, New York, NY, USA, 2023. Association for Computing Machinery. 1, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 264, + 524, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 524, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 524, + 314 + ], + "type": "text", + "content": "[60] Wenyan Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2138, 2018. 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 316, + 524, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 524, + 367 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 524, + 367 + ], + "type": "text", + "content": "[61] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4042-4051, 2022. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 368, + 524, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 368, + 524, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 524, + 426 + ], + "type": "text", + "content": "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 429, + 524, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 524, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 524, + 479 + ], + "type": "text", + "content": "[63] Jing Yang, Qingshan Liu, and Kaihua Zhang. Stacked hourglass network for robust facial landmark localisation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2025-2033, 2017. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 480, + 524, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 480, + 524, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 480, + 524, + 520 + ], + "type": "text", + "content": "[64] Sejong Yang, Subin Jeon, Seonghyeon Nam, and Seon Joo Kim. Dense interspecies face embedding. In Advances in Neural Information Processing Systems, pages 33275-33288. Curran Associates, Inc., 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 521, + 524, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 524, + 562 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 524, + 562 + ], + "type": "text", + "content": "[65] M. Ye, X. Zhang, P. C. Yuen, and S. Chang. Unsupervised embedding learning via invariant and spreading instance feature. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6203-6212, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 563, + 524, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 524, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 524, + 613 + ], + "type": "text", + "content": "[66] Lijun Yin, Xiaozhou Wei, Yi Sun, Jun Wang, and Matthew J Rosato. A 3d facial expression database for facial behavior research. In 7th international conference on automatic face and gesture recognition (FGR06), pages 211-216. IEEE, 2006. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 614, + 524, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 524, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 524, + 665 + ], + "type": "text", + "content": "[67] Xiang Zhang, Taoyue Wang, Xiaotian Li, Huiyuan Yang, and Lijun Yin. Weakly-supervised text-driven contrastive learning for facial behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20751-20762, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 524, + 686 + ], + "type": "text", + "content": "[68] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object land" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 313, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 313, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 313, + 714 + ], + "type": "text", + "content": "2091" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 458 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 86, + 100, + 288, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 100, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 86, + 100, + 288, + 130 + ], + "type": "text", + "content": "marks as structural representations. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 132, + 289, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 132, + 289, + 171 + ], + "spans": [ + { + "bbox": [ + 70, + 132, + 289, + 171 + ], + "type": "text", + "content": "[69] Yuhang Zhang, Chengrui Wang, and Weihong Deng. Relative uncertainty learning for facial expression recognition. In Advances in Neural Information Processing Systems, pages 17616-17627. Curran Associates, Inc., 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 172, + 289, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 172, + 289, + 222 + ], + "spans": [ + { + "bbox": [ + 70, + 172, + 289, + 222 + ], + "type": "text", + "content": "[70] Yuhang Zhang, Chengrui Wang, Xu Ling, and Weihong Deng. Learn from all: Erasing attention consistency for noisy label facial expression recognition. In Computer Vision - ECCV 2022, pages 418-434, Cham, 2022. Springer Nature Switzerland. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 224, + 289, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 224, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 70, + 224, + 289, + 263 + ], + "type": "text", + "content": "[71] Kaili Zhao, Wen-Sheng Chu, and Honggang Zhang. Deep region and multi-label learning for facial action unit detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3399, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 264, + 289, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 264, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 70, + 264, + 289, + 324 + ], + "type": "text", + "content": "[72] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18676-18688, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 325, + 289, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 325, + 289, + 366 + ], + "spans": [ + { + "bbox": [ + 70, + 325, + 289, + 366 + ], + "type": "text", + "content": "[73] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 367, + 289, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 367, + 289, + 417 + ], + "spans": [ + { + "bbox": [ + 70, + 367, + 289, + 417 + ], + "type": "text", + "content": "[74] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15475-15484, 2023. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 418, + 289, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 418, + 289, + 458 + ], + "spans": [ + { + "bbox": [ + 70, + 418, + 289, + 458 + ], + "type": "text", + "content": "[75] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 146-155, 2016. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "text", + "content": "2092" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_content_list.json b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9186205097cd483e98740680e24a441045e047c4 --- /dev/null +++ b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_content_list.json @@ -0,0 +1,1601 @@ +[ + { + "type": "text", + "text": "Self-Supervised Multi-Object Tracking with Path Consistency", + "text_level": 1, + "bbox": [ + 171, + 130, + 797, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijia Lu*, Bing Shuai†, Yanbei Chen, Zhenlin Xu, Davide Modolo \nAWS AI Labs", + "bbox": [ + 207, + 180, + 759, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lu.zij@northeastern.edu, {bshuai,yanbec,xzhenlin,dmodolo}@amazon.com", + "bbox": [ + 178, + 219, + 792, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel concept of path consistency to learn robust object matching without using manual object identity supervision. Our key idea is that, to track a object through frames, we can obtain multiple different association results from a model by varying the frames it can observe, i.e., skipping frames in observation. As the differences in observations do not alter the identities of objects, the obtained association results should be consistent. Based on this rationale, we generate multiple observation paths, each specifying a different set of frames to be skipped, and formulate the Path Consistency Loss that enforces the association results are consistent across different observation paths. We use the proposed loss to train our object matching model with only self-supervision. By extensive experiments on three tracking datasets (MOT17, PersonPath22, KITTI), we demonstrate that our method outperforms existing unsupervised methods with consistent margins on various evaluation metrics, and even achieves performance close to supervised methods.", + "bbox": [ + 75, + 300, + 473, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 599, + 209, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-Object Tracking (MOT) is the task of identifying and tracking all object instances present in a video. It benefits a wide spectrum of applications, such as monitoring vehicle traffic, wildlife activity, pedestrians, etc. While there have been several advanced supervised methods developed for MOT [5, 8, 29, 33, 34, 46, 52, 55, 58], the overall advancement in MOT is slower compared to other vision tasks, mostly due to the prohibitively expensive cost of acquiring high-quality object identities annotations for MOT training. As of today, MOT17 [31], a dataset of 14 videos, remains one of the most widely used datasets in MOT.", + "bbox": [ + 75, + 625, + 468, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome this problem, some recent works [1, 16, 21, 22, 24, 30] started shifting their interest on unsupervised learning for MOT. These approaches leverage off-the-shelf object detectors to localize all the object instances in each", + "bbox": [ + 75, + 791, + 470, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f7fc16452a84e4b19f61616876f0c38b443b62bd3769352b8da47c7aad0f882.jpg", + "image_caption": [ + "Figure 1. We propose a novel concept of path consistency for self-supervised MOT. We define a observation path for an object as a temporal list of observed frames from the start to end frame. As such, for the same object, we can generate multiple paths by skipping intermediate frames. As different observations of the same object does not alter its identity, the association results should be consistent across different paths." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 890, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frame of a video and then learn the temporal object association across frames in an unsupervised manner. Multiple works [16, 21, 22] propose to learn the object associations by creating pseudo-tracklets to estimate object IDs. However, they are vulnerable to the noises in tracklets, especially when the groundtruth tracklet of an object is incorrectly broken into several pseudo-tracklets, each being given a different object ID from the others. To avoid relying on pseudo-labels, other works [24, 30] explore self-supervision with temporal consistency objective, by tracking objects from the current frame to a future frame then back to the current frame while requiring each object is associated back to itself. However, these works only consider future frames in very short temporal distances (e.g., 1-2 frames) where objects lack substantial appearance difference, thus limiting the quality of their object embeddings. Meanwhile, UNS [1] leverages a motion model and an appearance model to create cross-supervision between their tracking outputs, yet it also only considers matching objects among adjacent frames.", + "bbox": [ + 496, + 579, + 892, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In fact, the ability of robust object matching over long", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Currently at Northeastern University. Work conducted during an internship with AWS AI Labs.", + "bbox": [ + 75, + 862, + 468, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19016", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "temporal distances is critical for MOT models as objects often encounter occlusion during tracking, where their bounding boxes are lost for a potentially long period of time, thus cannot be tracked by mere frame-by-frame matching. It requires long-distance matching to re-associate their bounding boxes after occlusion. However, very few works have managed to propose reliable solutions that tackle occlusion without using groundtruth supervision.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce the novel concept of path consistency, which can be used as a reliable self-supervised signal to learn a robust object matching model. To elaborate, when tracking an object, we define an observation path of it as a temporal list of observed frames between a start frame and a end frame. As shown in Figure 1, we are able to generate multiple different observation paths for the same object by randomly skipping intermediate frames. Based on an observation path, we can compute the temporal association between the object and all other objects from the start frame to the end frame. As such, path consistency dictates that the temporal associations should be consistent across paths, since the identities of objects do not change in different paths. Based on this rationale, we propose the Path Consistency Loss (PCL) that estimates the associations of objects along different paths and trains the object matching model by minimizing the discrepancy among association probability derived from different paths.", + "bbox": [ + 75, + 214, + 470, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Distinct from prior self-supervised methods [1, 24, 30], our formulation learns object matching over both short and long distances. As paths contain frame skipping of varied lengths, the model trained with PCL is able to learn not only short-distance matching (from paths with minimal frame skipping), but also long-distance matching (from paths with consecutive frame skipping) which is essential to handle occlusion during inference. In addition, we sample challenging training examples where paths span long temporal horizons, hence enabling our path consistency formulation to learn robust object matching models and handle occlusion.", + "bbox": [ + 75, + 488, + 468, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on three benchmark tracking datasets, MOT17 [31], PersonPath22 [39] and KITTI [13], and show it successfully achieves new state-of-the-art results over prior unsupervised methods and performs even close to recent supervised methods. We also provide extensive ablation studies to validate the design of our framework and improvement in handling occlusion. Our contributions are summarized as follows:", + "bbox": [ + 75, + 657, + 468, + 777 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the novel concept of path consistency that can be used as a reliable self-supervised signal to train a robust object matching model. Based on the concept, we formulate the Path Consistency Loss (PCL) that is capable of learning robust long-distance matching, which is essential in handling occlusion and large appearance changes in tracking.", + "- We establish new state-of-the-art performance over prior" + ], + "bbox": [ + 76, + 780, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unsupervised methods on three challenging tracking benchmarks. Extensive ablation studies also validate our improvement in long-distance matching and tracking over occlusion.", + "bbox": [ + 511, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 171, + 640, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised multi-object tracking. Most supervised MOT methods [2, 5, 15, 17, 19, 25-29, 36-38, 43, 44, 44, 50, 51, 53, 55, 58, 60] follow the tracking-by-detection paradigm where all object instances are first detected in each frame, then they are linked based on their similarities across frames to form tracklets. This paradigm benefits from the advances in object detection [6, 14, 35, 45, 57, 59] and has two threads of works: offline and online MOT. Offline MOT methods [2, 15, 17, 36, 37, 44, 51, 53] focus on finding global optimal associations of objects and are computationally heavy in general. Online MOT methods [3, 5, 29, 32, 38, 48, 52, 55, 56, 58, 60] focus on optimizing local optimum on associating detected objects and have real-time speed, e.g., Tracktor [3] uses bounding box regression to estimate new positions of objects. Meanwhile, the joint learning of the detection and object matching has made rapid progress in performance [48, 55], e.g., [5, 29, 52] introduce transformers and object tokens to associate objects with existing tracklets or to identify new ones. Despite the efficacy of supervised methods, they require costly human-annotated labels [10, 13, 31, 39]. In this work, we learn an online object matching model without labels, which we believe is an important step towards mitigating the limited size of tracking datasets.", + "bbox": [ + 496, + 200, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised multi-object tracking. This task aims to alleviate the high annotation cost of tracking labels. While existing motion trackers [4, 7, 56] do not require labels for training, they cannot leverage the appearance information of objects in tracking. Thus often lose objects in the event of movement change or long occlusion. To leverage appearance information, recent unsupervised methods focused on developing new object matching models [1, 22, 24, 30]. [20] proposes a setting of open-vocabulary MOT.", + "bbox": [ + 496, + 566, + 890, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among recent works, pseudo-label methods [16, 22] generate pseudo-tracklets to train their object matching models. SimepleReID [16] and UEANet [21] obtain pseudo-tracklets from motion trackers without detecting tracklet errors, thus those errors propagate to their models. U2MOT [22] improves the consistency of object IDs within tracklets, yet is still vulnerable to errors in the early stage of training and, more importantly, errors between tracklets. Specifically, the groundtruth tracklet of an object is often broken into several pseudo-tracklets if it experiences occlusions, where each pseudo-tracklet is then given a different object ID, treated as negative samples of each other. On the other hand, self-supervised methods [1, 24, 30] derive new", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19017", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unsupervised training objective to replace pseudo-labels. OUTrack [24] and UCSL [30] utilize temporal consistency, i.e., an object tracked from current frame to a future frame then back to current frame should associate back to itself. However, this objective does not hold if objects disappear among the frames, which limits those methods to only allow future frames at near distances (1-2 frames) to ensure objects likely remain present. Unfortunately, objects lack substantial appearance changes in such short distances, causing their failure to handle occlusion by matching objects over long temporal distances. We show that object matching learned with short distances cannot generalize to long distances (Table 5). UNS [1] creates cross-supervision between motion and appearance trackers yet also learn short-distance matching. It discusses occlusion-based supervision yet has failed and discarded it, likely due to its focus on short-distance matching. Meanwhile, PointId [41] learns a person embedding model by data augmentation, but it requires pretraining on ReID datasets with groundtruth labels.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike existing methods, we propose a novel self-supervised path consistency loss that simultaneously learns object matching over short to long distances without requiring pseudo or groundtruth labels. Thus we are able to consistently track objects in the event of long occlusion.", + "bbox": [ + 75, + 378, + 470, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 469, + 210, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Overview", + "text_level": 1, + "bbox": [ + 76, + 494, + 281, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We tackle the unsupervised MOT task with the concept of Path Consistency, which trains a robust object matching model capable of associating objects over long temporal distances, enabling consistent tracking in the event of occlusion. Following existing unsupervised works [1, 37, 44], we assume the bounding boxes of objects in each video frame are provided by an off-the-shelf detector. Our object matching model takes the detected objects and estimate object embeddings, which encode the appearance and spatial location of each object. Specifically, let $o_i^t$ denote the $i$ -th object in frame $t$ and $h_i^t$ its embedding. We define the probability that $o_i^t$ matches to an object $o_j^r$ in another frame $r$ as", + "bbox": [ + 75, + 518, + 472, + 703 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right) = \\frac {\\exp \\left(h _ {i} ^ {t} \\cdot h _ {j} ^ {r}\\right)}{\\sum_ {u = 1} ^ {N _ {r}} \\exp \\left(h _ {i} ^ {t} \\cdot h _ {u} ^ {r}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 714, + 468, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_r$ is the total number of objects in frame $r$ . Similarly, we also compute $p(o_j^r \\to o_i^t)$ , the matching probability from $o_j^r$ to $o_i^t$ . Importantly, to deal with occlusions and object exit in object association, we add a special null object, $\\phi$ , to the list of localized objects in each frame. If a real object $o_i^t$ is no longer visible in the following frame $r$ , it gets matched to $\\phi$ .", + "bbox": [ + 75, + 763, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference, our method uses the learned probability $p$ to associate objects to existing tracklets. Given a", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tracklet $\\tau = \\{o_m^{t_m}\\}_{m = 1}^M$ , consisting of the latest $M$ objects, it estimates the similarity between $\\tau$ and a new object $o_i^t$ as", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{M} \\sum_ {m} \\operatorname {g m} \\left[ p \\left(o _ {m} ^ {t _ {m}} \\rightarrow o _ {i} ^ {t}\\right), p \\left(o _ {i} ^ {t} \\rightarrow o _ {m} ^ {t _ {m}}\\right)\\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 131, + 890, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in which, gm denotes the geometric mean operator. Similar to all other works [22, 24, 30, 55, 56], we perform bipartite matching based on the tracklet-object similarities to either extend / terminate existing tracklets, or initiate new ones.", + "bbox": [ + 496, + 176, + 893, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the absence of tracking annotations, we propose a novel path consistency loss to learn the object matching model over varied temporal distances with only self-supervision (Section 3.2) and two regularization losses to improve $\\beta$ convergence (Section 3.3). As our method is agnostic to model structure and inference procedure, we provide their details in the supplementary materials.", + "bbox": [ + 496, + 237, + 893, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Self Supervision with Path Consistency", + "text_level": 1, + "bbox": [ + 498, + 349, + 834, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate the concept of path consistency as a reliable self-supervision signal to train a robust object matching model. Technically, suppose that we have a query object $o_{i}^{t_{s}}$ in the start frame $t_{s}$ , we aim to associating it to its correspondent object, $o_{j}^{t_{e}}$ , in the end frame $t_{e}$ ( $t_{e} > t_{s}$ ). Note that the corresponding object $o_{j}^{t_{e}}$ is the null object in the case that the query object is not visible in frame $t_{e}$ (due to being occluded or exiting the scene).", + "bbox": [ + 496, + 375, + 890, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define an observation path as the list of frames a model can observe to compute the association probability distribution between $o_i^{t_s}$ and all visible objects in frame $t_e$ . We are able to construct multiple paths by randomly skipping frames, thus creating a different visual observation for each path. Take the example in Figure 2 (right) as illustration, all frames are observed in Path1 whereas only a subset of frames are observed in Path2 and Path3. The core principle of the path consistency is, despite the different observations in the paths, the identities of objects remain unchanged, thus the association probability between $o_i^{t_s}$ and $o_j^{t_e}$ should always be higher than all other alternatives. That is, the association probability distribution between $o_i^{t_s}$ and all visible objects in frame $t_e$ should be consistent across paths. Based on this principle, we introduce our path consistency loss in Section 3.2.1, and then present our method to sample frames $(t_s, t_e)$ in Section 3.2.2.", + "bbox": [ + 496, + 497, + 892, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Path Consistency Loss", + "text_level": 1, + "bbox": [ + 500, + 767, + 710, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Path Consistency Loss (PCL) computes the association probability distributions from all possible paths, $\\{q_{\\pi}|\\pi \\sim \\Pi \\}$ , and enforces consistency among them1, where $\\Pi$ is the set of all paths and $q_{\\pi}$ is the association probability distribution obtained using a specific path", + "bbox": [ + 496, + 790, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1When the number of paths is large, we conduct Monte Carlo estimation by computing the loss with a smaller set of sampled paths.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19018", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7f8c75a50e14eafe345dc1d2036338f6e8efa1aa2cc1507d45712794ffe0bb9a.jpg", + "image_caption": [ + "Figure 2. Overview of Path Consistency Loss (PCL). Our method takes a video clip as input, where objects are localized by an off-the-shelf detector, and uses a selection strategy to choose suitable query objects and their correspondent end frames, then computes PCL to learn association between query objects and objects in end frames. Association probabilities obtained from different paths provide cross-supervision among them and enables self-supervised model learning." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\pi = \\{t_s, t_1, \\dots, t_K, t_e\\}$ . $t_k$ is an intermediate frame in the path. We use $q_{\\pi}^{t_k}(j)$ to denote the probability that query object $o_i^{t_s}$ matches with the $j$ -th object at frame $t_k$ .", + "bbox": [ + 75, + 380, + 470, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computing Association. To obtain association probability $q_{\\pi}$ , we aggregate the matching probability $p$ defined in Eq.1 according to the path $\\pi$ . We have", + "bbox": [ + 75, + 445, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\pi} ^ {t _ {k}} (j) = \\left\\{\\begin{array}{l l}p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {j} ^ {t _ {1}}\\right),&\\text {i f} t _ {k} = t _ {1}\\\\\\sum_ {u} q _ {\\pi} ^ {t _ {k - 1}} (u) \\cdot p \\left(o _ {u} ^ {t _ {k - 1}} \\rightarrow o _ {j} ^ {t _ {k}}\\right),&\\text {o t h e r w i s e}\\end{array}\\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 508, + 468, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the first case defines the association from query object to objects in the first intermediate frame $t_1$ , which is the matching probability, $p$ , between the objects. The second case extends the association from a frame $t_{k-1}$ to $t_k$ by jointly considering the matching probability from all objects in $t_{k-1}$ to $o_j^{t_k}$ . In contrast, pseudo-label methods [16, 21, 22] only consider the most likely matching, leading to error accumulation. Overall, we have the association probability between query object and objects in end frame as", + "bbox": [ + 75, + 571, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\pi} ^ {t _ {e}} (j) = \\sum_ {u} p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {u} ^ {t _ {1}}\\right) \\dots \\sum_ {v} p \\left(o _ {v} ^ {t _ {K}} \\rightarrow o _ {j} ^ {t _ {e}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 738, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which cumulatively multiplies the matching probabilities according to the frames in path.", + "bbox": [ + 75, + 787, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reduce the space of possible associations and facilitate learning, we constrain that only two spatially close objects between two neighboring frames can be matched. We define a binary spatial constraint mask $C_{t_k,v}^{t_{k-1},u}$ , which equals to zero if $o_v^{t_k}$ is one of the $S$ furthest objects to $o_u^{t_{k-1}}$ .", + "bbox": [ + 75, + 821, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, we update $q_{\\pi}^{t_e}$ as", + "bbox": [ + 500, + 380, + 653, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{Z} \\sum_ {u} \\left(C _ {t _ {1}, u} ^ {t _ {s}, i}\\right) p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {u} ^ {t _ {1}}\\right) \\dots \\sum_ {v} \\left(C _ {t _ {e}, j} ^ {t _ {K}, v}\\right) p \\left(o _ {v} ^ {t _ {K}} \\rightarrow o _ {j} ^ {t _ {e}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 402, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Z$ is a normalization factor to ensure the probability $q_{\\pi}^{t_e}$ sums to 1.", + "bbox": [ + 498, + 459, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computing Loss. After computing the association probabilities for all paths $\\{q_{\\pi}^{t_e}|\\pi \\sim \\Pi \\}$ , We define the path consistency loss to enforce consistency among them to learn our model with self-supervision,", + "bbox": [ + 498, + 506, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P C}} \\left(o _ {i} ^ {t _ {s}}, t _ {e}\\right) = \\frac {1}{| \\Pi |} \\sum_ {\\pi} K L \\left(q _ {\\pi} ^ {t _ {e}} \\mid \\mid \\hat {q}\\right) + H \\left(q _ {\\pi} ^ {t _ {e}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 571, + 890, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which includes two terms. The first term is the KL Divergence between each association probability $q_{\\pi}^{t_e}$ and the average of all probabilities, $\\hat{q} = \\sum_{\\pi} q_{\\pi}^{t_e} / |\\Pi|$ . Thus, it is minimized only when all probabilities follow the same distribution. The second term minimizes the entropy, $H(\\cdot)$ , of each probability to require only one object in frame $t_e$ can be associated to the query object, thus avoids the trivial solution where all $q_{\\pi}^{t_e}$ are consistent yet being uniform distributions.", + "bbox": [ + 496, + 613, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our path consistency loss has the key advantages of $(\\pmb{i})$ linear complexity w.r.t the number of paths as we compare each $q_{\\pi}^{t_e}$ to the averaged probability $\\hat{q}$ , hence avoid computing pairwise discrepancy between all $q_{\\pi}^{t_e}$ . $(ii)$ $\\hat{q}$ reflects association consistency and modulates the loss. When the consistency is low, multiple objects can have high probabilities in $\\hat{q}$ , indicating different association results. Unlike pseudo-label methods [16, 21, 22] that will enforce a specific association, PCL allows exploring all potential associations by using $\\hat{q}$ in $KL(\\cdot)$ , while $\\hat{q}$ also gives higher weights to the more likely ones.", + "bbox": [ + 496, + 734, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19019", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Selecting Frame Pairs to compute PCL", + "text_level": 1, + "bbox": [ + 76, + 90, + 403, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In principle, based on our PCL formulation we can sample any two random frames $(t_s, t_e)$ from a video clip, and then use all objects in $t_s$ as query objects. However, we observe such unconstrained frame sampling leads to most query objects not visible in the end frames. The unbalance between visible and disappeared query objects causes the model to always matching objects to the null objects and makes the training hard to converge.", + "bbox": [ + 75, + 114, + 468, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To constrain the frame sampling, we pick frame pairs $(t_s, t_e)$ under the condition that one object in $t_s$ should remain visible in intermediate and end frames to serve as the query object while $t_s$ and $t_e$ should be temporally disclose to support learning long-distance matching. As such, we develop a simple heuristics that uses the temporal chain of the IOU similarity to reject most frame pairs. With the existing solution, we observe that the query objects are present in end frames $98\\%$ of all sampled pairs. We include the details of our approach in the supplementary details. Notice that we can also employ more advanced approaches to include more challenging training examples and reject more noisy frame pairs, which we will leave for future works.", + "bbox": [ + 75, + 237, + 470, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Regularization Losses", + "text_level": 1, + "bbox": [ + 76, + 443, + 282, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further introduce the following two losses to regularize the model training together with our path consistency loss.", + "bbox": [ + 75, + 467, + 468, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One-to-one matching loss. To avoid that the path for two different query object cross each other, we constrain that a real, non-null object in frame $r$ , $o_j^r$ , can not be matched by more than one real objects in frame $t$ . As elaborated earlier, the probability $p(o_i^t \\to o_j^r)$ will be close to 1 between two matched real objects $o_i^t$ and $o_j^r$ . As such, we compute $\\sum_{i} p(o_i^t \\to o_j^r)$ to approximate the number of real objects in frame $t$ that are matched to object $o_j^r$ and penalize the value if it exceeds 1. Mathematically, our loss is defined as:", + "bbox": [ + 75, + 500, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {O M}} = \\frac {1}{T ^ {2}} \\sum_ {t, r} \\frac {1}{N ^ {r}} \\sum_ {j} \\max \\left(1, \\sum_ {i} p \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 646, + 468, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which penalizes any many-to-one real object matching between every two frames $(t,r)$ in the input video clip of length $T$ . $N^r$ is the number of real objects in frame $r$ .", + "bbox": [ + 75, + 700, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bidirectional consistency loss. In our model, we compute matching probability from frame $t$ to a later frame $r$ (forward in time) but also probability from frame $r$ to frame $t$ (backward in time). To ensure the matching probabilities are invariant to the temporal directions, we design a bidirectional consistency loss as:", + "bbox": [ + 75, + 748, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B C}} = \\frac {1}{T ^ {2} N _ {t} N _ {r}} \\sum_ {t, r} \\sum_ {i, j} | | p \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right) - p \\left(o _ {j} ^ {r} \\rightarrow o _ {i} ^ {t}\\right) | | ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 849, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given Eq. 7 and Eq. 8, our overall model objective is the sum of path consistency loss and two regularization losses: $\\mathcal{L} = \\mathcal{L}_{\\mathrm{PC}} + \\mathcal{L}_{\\mathrm{OM}} + \\mathcal{L}_{\\mathrm{BC}}$ , where we combine all loss terms with a balance weight of 1.0.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 167, + 633, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 500, + 194, + 687, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on three tracking datasets: MOT17 [31] is a widely adopted person tracking dataset containing 14 video sequences, featuring crowded scenes in indoor shopping malls or outdoor streets. The videos are split into 7 training sequences and 7 test sequences. **PersonPath22** [39] is a large-scale pedestrian dataset with 236 videos, which are collected from varied environment settings that contain longer occlusions and more crowded scenes. It is split into 138 training videos and 98 test videos. **KITTI** [13] is person and car tracking dataset, featuring videos with lower frame rates. We follow UNS [1] to report performance on car tracking.", + "bbox": [ + 496, + 219, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Following the evaluation protocols of the datasets, we report our results with several metrics: HOTA (DetA, AssA, LocA), MOTA (IDsw, FP and FN) and IDF1. We emphasize on HOTA, AssA, IDsw and IDF1 as they measure association accuracy. Other metrics are more influenced by the detection accuracy of detectors.", + "bbox": [ + 496, + 404, + 890, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our model input is a video clip of $T = 48$ frames, which we found balances between performance and efficiency. For PCL, we sample at most $G = 25$ different paths for each query object and end frame and set $S = \\sqrt{N}$ in spatial mask constraint, where $\\bar{N}$ is average number of objects per frame. We implement our model with Pytorch and trained with Adam Optimizer on Tesla V100 and a learning rate of 0.0001. In inference, we use the latest $M = 4$ bounding boxes of each tracklet to estimate its association with new objects and maintain unmatched tracklets for 30 frames. We provide more implementation details in the supplementary materials.", + "bbox": [ + 496, + 497, + 892, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison to the state-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 691, + 799, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with recent unsupervised object matching methods, where SimpleReID [16], U2MOT [22] are pseudo-label methods and OUTrack [24], UNS [1], UCSL [30] are self-supervised methods. For complete comparison, we include the results of motion-based methods [4, 7, 56], as they do not require object ID annotation, and also results of recent supervised methods.", + "bbox": [ + 496, + 715, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison on MOT17. In Table 1, we report the results on MOT17 dataset with both private and public detections. For fair comparison, we use YoloX detections from [56] for private detection and follow the detection pre-processing of UNS [1] for public detection. Our method outperforms", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19020", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/edcdbebe557525e13c8226946a9b16f4900645f0327651fee997ee95c5f94893.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sup.HOTA ↑IDF1↑MOTA↑IDsw↓
Private Detection
TransTrack [42]✓54.163.974.53663
TrackFormer [29]✓57.368.074.12829
MeMOT [5]✓56.969.072.52724
MOTR [52]✓57.868.673.42439
MeMOTR [12]✓58.871.572.81902
FairMOT [55]✓59.372.373.71074
FineTrack [34]✓64.379.580.01270
UTM [49]✓64.078.781.81431
MotionTrack [33]✓65.180.181.11140
SimpleReID [16]×50.460.769.0-
OUTrack [24]×-70.273.54110
PointID [40]×-72.474.22748
ByteTrack [56] †×63.177.380.32196
OC-SORT [7] †×63.277.578.01950
UCSL [30]×58.470.473.0-
U2MOT [22] †×64.278.279.91506
Ours †×65.079.680.91749
Public Detection
Tracktor++ [3] ‡✓42.152.353.52072
MHT-BLSTM [18]✓41.051.947.52069
FAMNet [9]✓-48.752.03072
LSST [11]✓47.162.354.71243
GSM [23]✓45.757.856.41485
CenterTrack [58]✓-60.555.72540
TrackFormer [29]✓-62.357.64018
SORT [4] ‡×34.039.843.14852
UNS [1] ‡×46.458.356.81320
Ours ‡×49.061.258.81219
", + "bbox": [ + 86, + 88, + 459, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "all unsupervised works on both private and public detections, and even exceeds recent fully-supervised methods, e.g., UTM [49], FineTrack [34].", + "bbox": [ + 75, + 564, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing with pseudo-label methods, we surpass the best competitor, U2MOT [22], by $0.8\\%$ in HOTA, $1.4\\%$ in IDF1 and $1\\%$ in MOTA with the same YoloX detection, as it is vulnerable to errors in pseudo labels. In contrast, our PCL naturally measures model confidence with $\\hat{q}$ and avoids enforcing incorrect, unconfident associations. Meanwhile, we also exceed self-supervised methods [1, 24, 30] as they only learn short-distance matching among adjacent frames. Our method learns object matching over both short and long distances thus is robust to appearance change and occlusion. Using the same public detection as UNS [1], we surpass it by $2.6\\%$ in HOTA, $2.9\\%$ in IDF1 and $2\\%$ in MOTA. Some works [16, 24] use better detectors to improve public detection, hence their performance are not comparable to us.", + "bbox": [ + 75, + 611, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on PersonPath22. In Table 2, we report results on PersonPath22 dataset, which is more challenging than MOT17 as its videos contain more crowded scenes with long occlusions. We compare with both public and private detections and successfully outperform ByteTrack", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1252ae92967acf129b4a7179805b2582840394de4d318d297c569bb0e5f5c23f.jpg", + "table_caption": [ + "Table 1. Comparison on MOT17 [31] dataset. $\\uparrow / \\downarrow$ : higher/lower is better. \"Sup.\" means requiring supervision. \" $\\dagger$ \" denotes using YoloX detection. \" $\\ddagger$ \" denotes using Tracktor++ detection preprocessing. The best results of unsupervised methods are in bold." + ], + "table_footnote": [], + "table_body": "
Sup.IDF1↑MOTA↑IDsw ↓FP↓FN↓
CenterTrack [58]✓46.459.31031924K72K
SiamMOT [38]✓53.767.5894213K63K
FairMOT [55]✓61.161.8509515K80K
TrackFormer [29]✓57.169.7863323K47K
SORT [4]×39.653.499272K110K
UNS [1]×41.855.51387717K86K
PointID [40]×63.168.661489K66K
ByteTrack [56] *×57.264.655309k71k
ByteTrack [56] †×66.875.4593117k40k
OC-SORT [7] †×68.274.0493117k41k
Ours*×60.564.850087k73k
Ours †×69.475.8466315k38k
", + "bbox": [ + 500, + 88, + 890, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d69c027b6844436791a6661b987a3538243ab5e49a5a24cff1fb99900f5c39d.jpg", + "table_caption": [ + "Table 2. Comparison on PersonPath22 [39] dataset. ↑/↓: higher/lower is better. The best results of unsupervised methods are in bold. “*” denotes using the public detection of Personpath22 provided by the authors. “†” denotes using YoloX detections." + ], + "table_footnote": [], + "table_body": "
Sup.HOTA↑AssA↑DetA↑MOTA↑
FAMNet [9]✓52.645.561.075.9
mmMOT [54]✓62.154.072.383.2
CenterTrack [58]✓73.071.275.688.8
LGM [47]✓73.172.374.687.6
TuSimple [8]✓71.671.172.686.3
PermaTr [46]✓78.078.478.391.3
SORT [4]×42.541.344.053.2
UNS [1]×62.565.361.1-
OC-SORT [7]×76.576.477.290.8
Ours×78.880.377.991.0
", + "bbox": [ + 511, + 359, + 879, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison on KITTI [13] dataset (Tracking Car). $\\uparrow/\\downarrow$ : higher/lower is better. The best results of unsupervised methods are in bold. We use the same detections as OC-SORT.", + "bbox": [ + 498, + 539, + 890, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by $3.3\\%$ and $2.6\\%$ in IDF1 respectively, also reducing IDsw by large margins. The larger improvement in IDF1 demonstrates the better ability of our model in matching over long temporal distances and re-associate objects over occlusion to deliver more consistent tracking results. With even public detection, our model has achieved better performance than CenterTrack [58], TrackFormer [29] and surpasses FairMOT [55] in MOTA, despite being unsupervised. These results show that our model effectively benefits from more training data, exposing the possibility of large-scale self-supervised learning with enormous tracking videos available online.", + "bbox": [ + 496, + 611, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More importantly, in Table 4, we study model performance under challenging scenarios, including different occlusion length (measured by frames), object moving speeds and people density. Our model consistently outperform the best competitor, ByteTrack, by large margins, thanks to our PCL learning more robust object matching to occlusion and appearance changes.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19021", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/026c59291304282b1b96212ab2ab5791309a19db724a736796507c0caa4c91a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Occlusion Length (frames)
01-2020-4040-6060+
UNS [1]49.236.832.827.026.1
OC-SORT [7]77.961.950.139.339.2
Ours79.264.350.740.739.8
", + "bbox": [ + 119, + 88, + 428, + 162 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7372db1fe67ff5d040cf4bcd6ef5b574766cc3f1f2a044ea6bf61bd7a354fbfb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Movement SpeedPeople Density
slowmediumfastlowhigh
UNS [1]42.740.338.344.742.4
OC-SORT [7]65.963.661.769.764.8
Ours68.365.162.671.466.3
", + "bbox": [ + 101, + 172, + 444, + 251 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/79f3e38041805648ee77e2364933560ddfc21ad91e76895bf4e2cc0339729462.jpg", + "table_caption": [ + "Table 4. Model Performance (IDF1) under different challenging scenarios on PersonPath22." + ], + "table_footnote": [], + "table_body": "
|t-r|1-45-89-1617-3232-48
PCL (s=T)99.599.198.196.394.5
s=3299.599.298.296.393.0
s=2499.599.198.094.888.0
s=1699.699.198.091.481.7
s=899.599.195.784.671.3
s=499.598.091.878.364.1
", + "bbox": [ + 86, + 303, + 455, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on KITTI. In Table 3, we further report our performance on KITTI dataset for tracking car, following the evaluation protocol of UNS [1]. KITTI features varied moving velocity of objects and lower frame rate on average, thus objects have large appearance and location changes in adjacent frames [7]. We use the shared detections from OC-SORT [7].", + "bbox": [ + 75, + 488, + 467, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model exceeds the best competitor, OC-SORT [7], by $2.3\\%$ in HOTA, $3.9\\%$ in AssA, $0.7\\%$ in DetA and $0.2\\%$ in MOTA. This is because cars have changes in their moving speeds and bounding box aspect ratios during driving (e.g., when making turn), which leads to low overlap between bounding boxes of the same object in consecutive frames and failure of pure motion trackers [4, 7]. On the other hand, our method learns object embeddings to encode object appearance thus successfully links cars in various driving scenarios, further validating our learned object matching is robust to appearance and location changes.", + "bbox": [ + 75, + 594, + 467, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 768, + 230, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation study to examine our key design choices. We train our unsupervised models on MOT17 test set and evaluate on MOT17 train set, since the annotations of the former set are not publicly available. Finally, we use the publicly released SDP detections.", + "bbox": [ + 75, + 792, + 467, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Has the model learned long-distance object matching?", + "bbox": [ + 76, + 869, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We exam the accuracy of our learned matching probability,", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5857d04041accdb86096cfdc21193ef1cc227347d2825f51fa5ae59039370489.jpg", + "table_caption": [ + "Table 5. Accuracy of the matching probability $p(o_i^t \\rightarrow o_j^r)$ . Our method has learned accurate object matching over short to long temporal distances." + ], + "table_footnote": [], + "table_body": "
L=0L=10L=20L=30L=40L=50L=60
UNS64.062.558.055.053.552.050.0
OC-SORT67.766.965.865.163.963.562.5
Ours68.968.367.466.966.165.965.3
Gain+1.2+1.4+1.6+1.9+2.2+2.4+2.8
", + "bbox": [ + 511, + 89, + 879, + 166 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Comparison of IDF1 for tracking over occlusion. When increasing the minimal occlusion length (number of frames), our model has more robust performance and exceeds the best competitor OC-SORT by large margins (as shown in last row).", + "bbox": [ + 498, + 176, + 890, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$p(o_{i}^{t}\\to o_{j}^{r})$ , where $p$ is correct if $o_t^r$ is the correct matching of $o_i^t$ and has the largest matching probability. In the first two rows of Table 5, we show the accuracy and break it down by the temporal distances between objects $|t - r|$ . It shows our matching probability is robust to temporal distances. As $|t - r|$ increases, the accuracy drops only by small margins and remains $94.5\\%$ even at the distance of 32-48 frames, validating our model has learned object matching from short to long distances.", + "bbox": [ + 496, + 258, + 890, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can the model only learn short-distance matching? Next, we further exam if learning long-distance matching is necessary and if a model only learned short-distance matching can generalize to long-distance matching? In rows 2-6 of Table 5, we limit the maximal number of continuous frames that can be skipped in PCL, denoted by $s$ . A smaller $s$ means the model is only trained to match objects over short temporal distances.", + "bbox": [ + 496, + 398, + 890, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As $s$ decreases from 32 to 4, while accuracy for short distances (1-8) remains high, it deteriorates quickly for longer distances. Comparing to the first row where we do not limit $s$ , accuracy drops by $30.4\\%$ on the distance of 32-48. It shows that model trained for short-distance matching cannot generalize to longer distances, which is the drawback of previous self-supervised methods [1, 24, 30]. Our PCL has successfully learned matching probability over varied temporal distances while not requiring groundtruth annotations. We also observe that $50\\%$ of frame skipping in paths are longer than 8 frames, which boosts our model robustness for long distance matching and the appearance changes caused by it.", + "bbox": [ + 496, + 520, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can the model handle long occlusion? After validating the accuracy of probability matching, we exam our model ability for handling occlusion at inference. We notice that most occlusions in MOT17 videos are of short lengths, e.g., less than 5 frames, which cannot reflect a model's ability to handle long-distance matching. Hence, we increase the occlusion length in a video by keep dropping the bounding boxes of objects after occlusion, until the occlusion lengths reach $L$ frames. If the original length of an occlusion is longer than $L$ , we do not increase it. Thus, $L = 0$ means no occlusion is extended while $L = 60$ means all occlusions last for at least 60 frames.", + "bbox": [ + 496, + 719, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19022", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8fd927a3f4ab848bcc7d5ac751795898bb1bccbd4b668c497d1166031cd2340d.jpg", + "image_caption": [ + "Frame 1" + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 299, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/859c6361e0aacaa21e860d6d8f00f4bac4d3da18bf9d5c8f328dfc22e2bbbbf9.jpg", + "image_caption": [ + "Frame 20", + "Figure 3. Qualitative comparison between our model and UNS [1]. We visualize the tracking on three frames. UNS cannot track the person in green bounding once he is occluded and assigns him a new ID (purple) on frame 80. It also fails to track the person in pink bounding box when she is only partially visible in frame 80. We can consistently track with both people with the same IDs (green, pink)." + ], + "image_footnote": [], + "bbox": [ + 300, + 90, + 480, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8844fe0e04d3f73a936e2e6b4b7cb9b04a31c72b8b7b292baacef8f0ce386c42.jpg", + "image_caption": [ + "Frame 80 (Ours)" + ], + "image_footnote": [], + "bbox": [ + 480, + 90, + 663, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bebda331f59d9621356899e1c2591344d029dbeb2464d6635a289fabbb8bf16f.jpg", + "image_caption": [ + "Frame 80 (UNS)" + ], + "image_footnote": [], + "bbox": [ + 666, + 90, + 844, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/556d5142706caea6fcfc0ad99e3207a0b78c4eded19cd29b2f42b50e43929886.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\text{PC}} \\)\\( \\mathcal{L}_{\\text{OM}} \\)\\( \\mathcal{L}_{\\text{BC}} \\)IDF1HOTAMOTAIDsw
1×✓✓3.48.31.470k
2(a)✓✓26.926.837.829k
3(b)✓✓66.659.463.9329
4✓×✓62.757.163.4591
5✓✓×66.359.863.6466
6✓✓✓68.960.963.7257
", + "bbox": [ + 78, + 251, + 467, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on proposed losses. In row 1-3, we demonstrate the effect of our path consistency loss. In row 4-5, we show the effect of our regularization losses. \"(a)\" denotes randomly sampling start and end frames for paths and \"(b)\" denotes disabling spatial constraint mask in Eq 5.", + "bbox": [ + 75, + 391, + 468, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 6, we report IDF1 with different occlusion length $L$ and compare with UNS and one of the state-of-the-art trackers, OC-SORT, as a strong baseline. Since other related methods [16, 22, 24, 30] do not release their codes or models, we are unable to compare with them. As can be seen, our model consistently provides higher IDF1 for all ranges of occlusion lengths. More importantly, when increasing occlusion lengths, the performance of UNS and OC-SORT decreases quickly, while our performance is more robust and decreases slowly, which echos our results in Table 4. Our IDF1 at $L = 50$ is still higher than the IDF1 of OC-SORT at $L = 20$ , showing our model can learn to match the same objects over varied temporal distances and is more robust to occlusion.", + "bbox": [ + 75, + 470, + 468, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What is the effect of path consistency loss? In Table 7 (row 1-3), we exam the results of removing PCL and also two modifications of it: (a) randomly sampling the starting and end frames for paths; (b) removing the spatial constraint mask in Eq.5; Firstly, without PCL (row 1), the model cannot learn meaningful object matching using only the regularization losses thus the tracking accuracy is close to zeros. When randomly sampling the starting and end frames (row 2), it causes most query objects are not visible in the end frames, causing imbalance between visible and disappeared objects. Thus, model matches most objects to null objects, also leading to low IDF1. Next, when removing the spatial constraint mask (row 3), the model may incorrectly match two spatial disclose objects, decreasing IDF1 by $2.3\\%$ .", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What is the effect of regularization losses? In Table 7 (row 4-5), we ablate the effect of the regularization losses in Section 3.3. As can be seen, while PCL contributes most to the model performance, the two regularization losses $\\mathcal{L}_{\\mathrm{OM}}$ , $\\mathcal{L}_{\\mathrm{BC}}$ are also helpful for achieving better performance. Since $\\mathcal{L}_{\\mathrm{OM}}$ prevents multiple objects matching to the same objects and $\\mathcal{L}_{\\mathrm{BC}}$ ensures the consistency between forward and backward matching, removing either of these loss terms leads to around $6.2\\%$ and $2.6\\%$ drop in IDF1, respectively. These results show the collective benefits of using different losses to optimize our unsupervised object tracker.", + "bbox": [ + 496, + 253, + 893, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 500, + 430, + 676, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 3, we visualize the tracking results for one video from MOT17 using our model and UNS [1], where the color indicates their IDs. Both our model and UNS can track two non-occluded front people from frame 1 to frame 80. However, UNS lose the man in green bounding box after he is occluded and assign a new ID (purple) to him on frame 80. It also cannot associate the woman in the pink bounding box when she is only partially visible in frame 80. Our model correctly associate two people with the same IDs (green and pink). This qualitative comparison suggests that our model can better track the object over occlusion.", + "bbox": [ + 496, + 452, + 893, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 632, + 617, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To tackle unsupervised MOT task, we proposed a novel concept of path consistency. We defined a path to track an object as a temporal list of observed frames from the start to the end frame. As such, to track the same object, we can generate multiple paths by skipping intermediate frames while the association results should be consistent across different paths. We leveraged this consistency in our self-supervised Path Consistency Loss to learn object matching model without groundtruth ID supervision. Our PCL especially learns robust long distance matching to tackle occlusion at inference time. With experiments on three popular datasets, our approach has exceeded existing unsupervised methods and performed on par with recent state-of-the-art supervised methods. We provided a comprehensive ablation study to validate the model has learned long-distance matching and is robust under challenging scenarios.", + "bbox": [ + 496, + 657, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19023", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In Advances in Neural Information Processing Systems, pages 13695-13706. Curran Associates, Inc., 2021. 1, 2, 3, 5, 6, 7, 8", + "[2] Jerome Berclaz, Francois Fleuret, Engin Turetken, and Pascal Fua. Multiple object tracking using k-shortest paths optimization. TPAMI, 33(9):1806-1819, 2011. 2", + "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6", + "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE International Conference on Image Processing (ICIP), pages 3464-3468, 2016. 2, 5, 6, 7", + "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR 2022, 2022. 1, 2, 6", + "[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 2", + "[7] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 2, 5, 6, 7", + "[8] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015. 1, 6", + "[9] Peng Chu and Haibin Ling. Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6172-6181, 2019. 6", + "[10] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes, 2020. 2", + "[11] Weitao Feng, Zhihao Hu, Wei Wu, Junjie Yan, and Wanli Ouyang. Multi-object tracking with multiple cues and switcher-aware classification, 2019. 6", + "[12] Ruopeng Gao and Limin Wang. MeMOTR: Long-term memory-augmented transformer for multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9901-9910, 2023. 6", + "[13] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 2, 5, 6", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2", + "[15] Roberto Henschel, Laura Leal-Taixe, Daniel Cremers, and Bodo Rosenhahn. Improvements to frank-wolfe optimization for multi-detector multi-object tracking. arXiv preprint arXiv:1705.08314, 2017. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Shyamgopal Karthik, Ameya Prabhu, and Vineet Gandhi. Simple unsupervised multi-object tracking. CoRR, abs/2006.02609, 2020. 1, 2, 4, 5, 6, 8", + "[17] Chanho Kim, Fuxin Li, Arridhana Ciptadi, and James M Rehg. Multiple hypothesis tracking revisited. In ICCV, 2015. 2", + "[18] Chanho Kim, Fuxin Li, and James M. Rehg. Multi-object tracking with neural gating using bilinear lstm. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6", + "[19] Shuai Li, Yu Kong, and Hamid Rezatofighi. Learning of global objective for network flow in multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8855-8865, 2022. 2", + "[20] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 2", + "[21] Yu-Lei Li. Unsupervised embedding and association network for multi-object tracking. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. International Joint Conferences on Artificial Intelligence Organization, 2022. 1, 2, 4", + "[22] Kai Liu, Sheng Jin, Zhihang Fu, Ze Chen, Rongxin Jiang, and Jieping Ye. Uncertainty-aware unsupervised multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9996-10005, 2023. 1, 2, 3, 4, 5, 6, 8", + "[23] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 530–536. International Joint Conferences on Artificial Intelligence Organization, 2020. Main track. 6", + "[24] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333-347, 2022. 1, 2, 3, 5, 6, 7, 8", + "[25] Shuai Liu, Xin Li, Huchuan Lu, and You He. Multi-object tracking meets moving UAV. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8876-8885, 2022. 2", + "[26] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19045-19055, 2022.", + "[27] Fan Ma, Mike Zheng Shou, Linchao Zhu, Haoqi Fan, Yilei Xu, Yi Yang, and Zhicheng Yan. Unified transformer tracker for object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8781-8790, 2022.", + "[28] Christoph Mayer, Martin Danelljan, Goutam Bhat, Matthieu Paul, Danda Pani Paudel, Fisher Yu, and Luc Van Gool." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19024", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Transforming model prediction for tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8731-8740, 2022.", + "[29] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6", + "[30] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16264-16273, 2023. 1, 2, 3, 5, 6, 7, 8", + "[31] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. MOT16: A benchmark for multi-object tracking. CoRR, abs/1603.00831, 2016. 1, 2, 5, 6", + "[32] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2", + "[33] Z. Qin, S. Zhou, L. Wang, J. Duan, G. Hua, and W. Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6", + "[34] Hao Ren, Shoudong Han, Huilin Ding, Ziwen Zhang, Hongwei Wang, and Faquan Wang. Focus on details: Online multi-object tracking with diverse fine-grained representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2", + "[36] Ergys Ristani and Carlo Tomasi. Features for multi-target multi-camera tracking and re-identification. In CVPR, 2018. 2", + "[37] Hao Sheng, Yang Zhang, Jiahui Chen, Zhang Xiong, and Jun Zhang. Heterogeneous association graph fusion for target association in multiple object tracking. TCSVT, 29(11): 3269-3280, 2018. 2, 3", + "[38] Bing Shuai, Andrew Berneshawi, Xinyu Li, Davide Modolo, and Joseph Tighe. Siammot: Siamese multi-object tracking. In CVPR, 2021. 2, 6", + "[39] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joe Tighe. Large scale real-world multi person tracking. In European Conference on Computer Vision. Springer, 2022. 2, 5, 6", + "[40] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14669-14679, 2022. 6", + "[41] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14689-14699, 2022. 3", + "[42] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack:" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460, 2020. 6", + "[43] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20993-21002, 2022. 2", + "[44] Siyu Tang, Mykhaylo Andriluka, Bjoern Andres, and Bernt Schiele. Multiple people tracking by lifted multicut and person re-identification. In CVPR, 2017. 2, 3", + "[45] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2", + "[46] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10840-10849, 2021. 1, 6", + "[47] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021. 6", + "[48] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 107-122. Springer, 2020. 2", + "[49] S. You, H. Yao, B. Bao, and C. Xu. Utm: A unified multiple object tracking model with identity-aware feature enhancement. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[50] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8834-8843, 2022. 2", + "[51] Amir Roshan Zamir, Afshin Dehghan, and Mubarak Shah. Gmcp-tracker: Global multi-object tracking using generalized minimum clique graphs. In ECCV, 2012. 2", + "[52] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6", + "[53] Li Zhang, Yuan Li, and Ramakant Nevatia. Global data association for multi-object tracking using network flows. In CVPR, 2008. 2", + "[54] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2365-2374, 2019. 6", + "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1, 2, 3, 6" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19025", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6", + "[57] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2", + "[58] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. ECCV, 2020. 1, 2, 6", + "[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 2", + "[60] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 301 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19026", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_model.json b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4bce61e574319f0ba1f59432a82a9ddfee9dc38 --- /dev/null +++ b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.173, + 0.131, + 0.798, + 0.154 + ], + "angle": 0, + "content": "Self-Supervised Multi-Object Tracking with Path Consistency" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.181, + 0.761, + 0.216 + ], + "angle": 0, + "content": "Zijia Lu*, Bing Shuai†, Yanbei Chen, Zhenlin Xu, Davide Modolo \nAWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.22, + 0.793, + 0.234 + ], + "angle": 0, + "content": "lu.zij@northeastern.edu, {bshuai,yanbec,xzhenlin,dmodolo}@amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.59 + ], + "angle": 0, + "content": "In this paper, we propose a novel concept of path consistency to learn robust object matching without using manual object identity supervision. Our key idea is that, to track a object through frames, we can obtain multiple different association results from a model by varying the frames it can observe, i.e., skipping frames in observation. As the differences in observations do not alter the identities of objects, the obtained association results should be consistent. Based on this rationale, we generate multiple observation paths, each specifying a different set of frames to be skipped, and formulate the Path Consistency Loss that enforces the association results are consistent across different observation paths. We use the proposed loss to train our object matching model with only self-supervision. By extensive experiments on three tracking datasets (MOT17, PersonPath22, KITTI), we demonstrate that our method outperforms existing unsupervised methods with consistent margins on various evaluation metrics, and even achieves performance close to supervised methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.601, + 0.21, + 0.617 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.791 + ], + "angle": 0, + "content": "Multi-Object Tracking (MOT) is the task of identifying and tracking all object instances present in a video. It benefits a wide spectrum of applications, such as monitoring vehicle traffic, wildlife activity, pedestrians, etc. While there have been several advanced supervised methods developed for MOT [5, 8, 29, 33, 34, 46, 52, 55, 58], the overall advancement in MOT is slower compared to other vision tasks, mostly due to the prohibitively expensive cost of acquiring high-quality object identities annotations for MOT training. As of today, MOT17 [31], a dataset of 14 videos, remains one of the most widely used datasets in MOT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.471, + 0.854 + ], + "angle": 0, + "content": "To overcome this problem, some recent works [1, 16, 21, 22, 24, 30] started shifting their interest on unsupervised learning for MOT. These approaches leverage off-the-shelf object detectors to localize all the object instances in each" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.27, + 0.891, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.449, + 0.894, + 0.547 + ], + "angle": 0, + "content": "Figure 1. We propose a novel concept of path consistency for self-supervised MOT. We define a observation path for an object as a temporal list of observed frames from the start to end frame. As such, for the same object, we can generate multiple paths by skipping intermediate frames. As different observations of the same object does not alter its identity, the association results should be consistent across different paths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.882 + ], + "angle": 0, + "content": "frame of a video and then learn the temporal object association across frames in an unsupervised manner. Multiple works [16, 21, 22] propose to learn the object associations by creating pseudo-tracklets to estimate object IDs. However, they are vulnerable to the noises in tracklets, especially when the groundtruth tracklet of an object is incorrectly broken into several pseudo-tracklets, each being given a different object ID from the others. To avoid relying on pseudo-labels, other works [24, 30] explore self-supervision with temporal consistency objective, by tracking objects from the current frame to a future frame then back to the current frame while requiring each object is associated back to itself. However, these works only consider future frames in very short temporal distances (e.g., 1-2 frames) where objects lack substantial appearance difference, thus limiting the quality of their object embeddings. Meanwhile, UNS [1] leverages a motion model and an appearance model to create cross-supervision between their tracking outputs, yet it also only considers matching objects among adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In fact, the ability of robust object matching over long" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.887 + ], + "angle": 0, + "content": "*Currently at Northeastern University. Work conducted during an internship with AWS AI Labs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19016" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "temporal distances is critical for MOT models as objects often encounter occlusion during tracking, where their bounding boxes are lost for a potentially long period of time, thus cannot be tracked by mere frame-by-frame matching. It requires long-distance matching to re-associate their bounding boxes after occlusion. However, very few works have managed to propose reliable solutions that tackle occlusion without using groundtruth supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.486 + ], + "angle": 0, + "content": "In this paper, we introduce the novel concept of path consistency, which can be used as a reliable self-supervised signal to learn a robust object matching model. To elaborate, when tracking an object, we define an observation path of it as a temporal list of observed frames between a start frame and a end frame. As shown in Figure 1, we are able to generate multiple different observation paths for the same object by randomly skipping intermediate frames. Based on an observation path, we can compute the temporal association between the object and all other objects from the start frame to the end frame. As such, path consistency dictates that the temporal associations should be consistent across paths, since the identities of objects do not change in different paths. Based on this rationale, we propose the Path Consistency Loss (PCL) that estimates the associations of objects along different paths and trains the object matching model by minimizing the discrepancy among association probability derived from different paths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Distinct from prior self-supervised methods [1, 24, 30], our formulation learns object matching over both short and long distances. As paths contain frame skipping of varied lengths, the model trained with PCL is able to learn not only short-distance matching (from paths with minimal frame skipping), but also long-distance matching (from paths with consecutive frame skipping) which is essential to handle occlusion during inference. In addition, we sample challenging training examples where paths span long temporal horizons, hence enabling our path consistency formulation to learn robust object matching models and handle occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.779 + ], + "angle": 0, + "content": "We evaluate our method on three benchmark tracking datasets, MOT17 [31], PersonPath22 [39] and KITTI [13], and show it successfully achieves new state-of-the-art results over prior unsupervised methods and performs even close to recent supervised methods. We also provide extensive ablation studies to validate the design of our framework and improvement in handling occlusion. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.781, + 0.47, + 0.886 + ], + "angle": 0, + "content": "- We propose the novel concept of path consistency that can be used as a reliable self-supervised signal to train a robust object matching model. Based on the concept, we formulate the Path Consistency Loss (PCL) that is capable of learning robust long-distance matching, which is essential in handling occlusion and large appearance changes in tracking." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We establish new state-of-the-art performance over prior" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "unsupervised methods on three challenging tracking benchmarks. Extensive ablation studies also validate our improvement in long-distance matching and tracking over occlusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.172, + 0.642, + 0.187 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Supervised multi-object tracking. Most supervised MOT methods [2, 5, 15, 17, 19, 25-29, 36-38, 43, 44, 44, 50, 51, 53, 55, 58, 60] follow the tracking-by-detection paradigm where all object instances are first detected in each frame, then they are linked based on their similarities across frames to form tracklets. This paradigm benefits from the advances in object detection [6, 14, 35, 45, 57, 59] and has two threads of works: offline and online MOT. Offline MOT methods [2, 15, 17, 36, 37, 44, 51, 53] focus on finding global optimal associations of objects and are computationally heavy in general. Online MOT methods [3, 5, 29, 32, 38, 48, 52, 55, 56, 58, 60] focus on optimizing local optimum on associating detected objects and have real-time speed, e.g., Tracktor [3] uses bounding box regression to estimate new positions of objects. Meanwhile, the joint learning of the detection and object matching has made rapid progress in performance [48, 55], e.g., [5, 29, 52] introduce transformers and object tokens to associate objects with existing tracklets or to identify new ones. Despite the efficacy of supervised methods, they require costly human-annotated labels [10, 13, 31, 39]. In this work, we learn an online object matching model without labels, which we believe is an important step towards mitigating the limited size of tracking datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Unsupervised multi-object tracking. This task aims to alleviate the high annotation cost of tracking labels. While existing motion trackers [4, 7, 56] do not require labels for training, they cannot leverage the appearance information of objects in tracking. Thus often lose objects in the event of movement change or long occlusion. To leverage appearance information, recent unsupervised methods focused on developing new object matching models [1, 22, 24, 30]. [20] proposes a setting of open-vocabulary MOT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Among recent works, pseudo-label methods [16, 22] generate pseudo-tracklets to train their object matching models. SimepleReID [16] and UEANet [21] obtain pseudo-tracklets from motion trackers without detecting tracklet errors, thus those errors propagate to their models. U2MOT [22] improves the consistency of object IDs within tracklets, yet is still vulnerable to errors in the early stage of training and, more importantly, errors between tracklets. Specifically, the groundtruth tracklet of an object is often broken into several pseudo-tracklets if it experiences occlusions, where each pseudo-tracklet is then given a different object ID, treated as negative samples of each other. On the other hand, self-supervised methods [1, 24, 30] derive new" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19017" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.379 + ], + "angle": 0, + "content": "unsupervised training objective to replace pseudo-labels. OUTrack [24] and UCSL [30] utilize temporal consistency, i.e., an object tracked from current frame to a future frame then back to current frame should associate back to itself. However, this objective does not hold if objects disappear among the frames, which limits those methods to only allow future frames at near distances (1-2 frames) to ensure objects likely remain present. Unfortunately, objects lack substantial appearance changes in such short distances, causing their failure to handle occlusion by matching objects over long temporal distances. We show that object matching learned with short distances cannot generalize to long distances (Table 5). UNS [1] creates cross-supervision between motion and appearance trackers yet also learn short-distance matching. It discusses occlusion-based supervision yet has failed and discarded it, likely due to its focus on short-distance matching. Meanwhile, PointId [41] learns a person embedding model by data augmentation, but it requires pretraining on ReID datasets with groundtruth labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.472, + 0.456 + ], + "angle": 0, + "content": "Unlike existing methods, we propose a novel self-supervised path consistency loss that simultaneously learns object matching over short to long distances without requiring pseudo or groundtruth labels. Thus we are able to consistently track objects in the event of long occlusion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.212, + 0.488 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.282, + 0.51 + ], + "angle": 0, + "content": "3.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.704 + ], + "angle": 0, + "content": "We tackle the unsupervised MOT task with the concept of Path Consistency, which trains a robust object matching model capable of associating objects over long temporal distances, enabling consistent tracking in the event of occlusion. Following existing unsupervised works [1, 37, 44], we assume the bounding boxes of objects in each video frame are provided by an off-the-shelf detector. Our object matching model takes the detected objects and estimate object embeddings, which encode the appearance and spatial location of each object. Specifically, let \\( o_i^t \\) denote the \\( i \\)-th object in frame \\( t \\) and \\( h_i^t \\) its embedding. We define the probability that \\( o_i^t \\) matches to an object \\( o_j^r \\) in another frame \\( r \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.715, + 0.47, + 0.755 + ], + "angle": 0, + "content": "\\[\np \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right) = \\frac {\\exp \\left(h _ {i} ^ {t} \\cdot h _ {j} ^ {r}\\right)}{\\sum_ {u = 1} ^ {N _ {r}} \\exp \\left(h _ {i} ^ {t} \\cdot h _ {u} ^ {r}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.869 + ], + "angle": 0, + "content": "where \\( N_r \\) is the total number of objects in frame \\( r \\). Similarly, we also compute \\( p(o_j^r \\to o_i^t) \\), the matching probability from \\( o_j^r \\) to \\( o_i^t \\). Importantly, to deal with occlusions and object exit in object association, we add a special null object, \\( \\phi \\), to the list of localized objects in each frame. If a real object \\( o_i^t \\) is no longer visible in the following frame \\( r \\), it gets matched to \\( \\phi \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "During inference, our method uses the learned probability \\( p \\) to associate objects to existing tracklets. Given a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.123 + ], + "angle": 0, + "content": "tracklet \\(\\tau = \\{o_m^{t_m}\\}_{m = 1}^M\\), consisting of the latest \\(M\\) objects, it estimates the similarity between \\(\\tau\\) and a new object \\(o_i^t\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.132, + 0.892, + 0.167 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{M} \\sum_ {m} \\operatorname {g m} \\left[ p \\left(o _ {m} ^ {t _ {m}} \\rightarrow o _ {i} ^ {t}\\right), p \\left(o _ {i} ^ {t} \\rightarrow o _ {m} ^ {t _ {m}}\\right)\\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.177, + 0.894, + 0.236 + ], + "angle": 0, + "content": "in which, gm denotes the geometric mean operator. Similar to all other works [22, 24, 30, 55, 56], we perform bipartite matching based on the tracklet-object similarities to either extend / terminate existing tracklets, or initiate new ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.895, + 0.343 + ], + "angle": 0, + "content": "In the absence of tracking annotations, we propose a novel path consistency loss to learn the object matching model over varied temporal distances with only self-supervision (Section 3.2) and two regularization losses to improve \\(\\beta\\) convergence (Section 3.3). As our method is agnostic to model structure and inference procedure, we provide their details in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.351, + 0.836, + 0.368 + ], + "angle": 0, + "content": "3.2. Self Supervision with Path Consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.497 + ], + "angle": 0, + "content": "In this section, we elaborate the concept of path consistency as a reliable self-supervision signal to train a robust object matching model. Technically, suppose that we have a query object \\( o_{i}^{t_{s}} \\) in the start frame \\( t_{s} \\), we aim to associating it to its correspondent object, \\( o_{j}^{t_{e}} \\), in the end frame \\( t_{e} \\) (\\( t_{e} > t_{s} \\)). Note that the corresponding object \\( o_{j}^{t_{e}} \\) is the null object in the case that the query object is not visible in frame \\( t_{e} \\) (due to being occluded or exiting the scene)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.893, + 0.758 + ], + "angle": 0, + "content": "We define an observation path as the list of frames a model can observe to compute the association probability distribution between \\( o_i^{t_s} \\) and all visible objects in frame \\( t_e \\). We are able to construct multiple paths by randomly skipping frames, thus creating a different visual observation for each path. Take the example in Figure 2 (right) as illustration, all frames are observed in Path1 whereas only a subset of frames are observed in Path2 and Path3. The core principle of the path consistency is, despite the different observations in the paths, the identities of objects remain unchanged, thus the association probability between \\( o_i^{t_s} \\) and \\( o_j^{t_e} \\) should always be higher than all other alternatives. That is, the association probability distribution between \\( o_i^{t_s} \\) and all visible objects in frame \\( t_e \\) should be consistent across paths. Based on this principle, we introduce our path consistency loss in Section 3.2.1, and then present our method to sample frames \\( (t_s, t_e) \\) in Section 3.2.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.712, + 0.783 + ], + "angle": 0, + "content": "3.2.1 Path Consistency Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Path Consistency Loss (PCL) computes the association probability distributions from all possible paths, \\(\\{q_{\\pi}|\\pi \\sim \\Pi \\}\\), and enforces consistency among them1, where \\(\\Pi\\) is the set of all paths and \\(q_{\\pi}\\) is the association probability distribution obtained using a specific path" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1When the number of paths is large, we conduct Monte Carlo estimation by computing the loss with a smaller set of sampled paths." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19018" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.891, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.893, + 0.37 + ], + "angle": 0, + "content": "Figure 2. Overview of Path Consistency Loss (PCL). Our method takes a video clip as input, where objects are localized by an off-the-shelf detector, and uses a selection strategy to choose suitable query objects and their correspondent end frames, then computes PCL to learn association between query objects and objects in end frames. Association probabilities obtained from different paths provide cross-supervision among them and enables self-supervised model learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.471, + 0.429 + ], + "angle": 0, + "content": "\\(\\pi = \\{t_s, t_1, \\dots, t_K, t_e\\}\\). \\(t_k\\) is an intermediate frame in the path. We use \\(q_{\\pi}^{t_k}(j)\\) to denote the probability that query object \\(o_i^{t_s}\\) matches with the \\(j\\)-th object at frame \\(t_k\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.469, + 0.492 + ], + "angle": 0, + "content": "Computing Association. To obtain association probability \\( q_{\\pi} \\), we aggregate the matching probability \\( p \\) defined in Eq.1 according to the path \\( \\pi \\). We have" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.51, + 0.47, + 0.555 + ], + "angle": 0, + "content": "\\[\nq _ {\\pi} ^ {t _ {k}} (j) = \\left\\{\\begin{array}{l l}p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {j} ^ {t _ {1}}\\right),&\\text {i f} t _ {k} = t _ {1}\\\\\\sum_ {u} q _ {\\pi} ^ {t _ {k - 1}} (u) \\cdot p \\left(o _ {u} ^ {t _ {k - 1}} \\rightarrow o _ {j} ^ {t _ {k}}\\right),&\\text {o t h e r w i s e}\\end{array}\\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.724 + ], + "angle": 0, + "content": "where the first case defines the association from query object to objects in the first intermediate frame \\( t_1 \\), which is the matching probability, \\( p \\), between the objects. The second case extends the association from a frame \\( t_{k-1} \\) to \\( t_k \\) by jointly considering the matching probability from all objects in \\( t_{k-1} \\) to \\( o_j^{t_k} \\). In contrast, pseudo-label methods [16, 21, 22] only consider the most likely matching, leading to error accumulation. Overall, we have the association probability between query object and objects in end frame as" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.739, + 0.47, + 0.772 + ], + "angle": 0, + "content": "\\[\nq _ {\\pi} ^ {t _ {e}} (j) = \\sum_ {u} p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {u} ^ {t _ {1}}\\right) \\dots \\sum_ {v} p \\left(o _ {v} ^ {t _ {K}} \\rightarrow o _ {j} ^ {t _ {e}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.469, + 0.82 + ], + "angle": 0, + "content": "which cumulatively multiplies the matching probabilities according to the frames in path." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To reduce the space of possible associations and facilitate learning, we constrain that only two spatially close objects between two neighboring frames can be matched. We define a binary spatial constraint mask \\( C_{t_k,v}^{t_{k-1},u} \\), which equals to zero if \\( o_v^{t_k} \\) is one of the \\( S \\) furthest objects to \\( o_u^{t_{k-1}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.381, + 0.655, + 0.399 + ], + "angle": 0, + "content": "Thus, we update \\(q_{\\pi}^{t_e}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.403, + 0.892, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{Z} \\sum_ {u} \\left(C _ {t _ {1}, u} ^ {t _ {s}, i}\\right) p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {u} ^ {t _ {1}}\\right) \\dots \\sum_ {v} \\left(C _ {t _ {e}, j} ^ {t _ {K}, v}\\right) p \\left(o _ {v} ^ {t _ {K}} \\rightarrow o _ {j} ^ {t _ {e}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.892, + 0.492 + ], + "angle": 0, + "content": "where \\( Z \\) is a normalization factor to ensure the probability \\( q_{\\pi}^{t_e} \\) sums to 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.507, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Computing Loss. After computing the association probabilities for all paths \\(\\{q_{\\pi}^{t_e}|\\pi \\sim \\Pi \\}\\), We define the path consistency loss to enforce consistency among them to learn our model with self-supervision," + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.573, + 0.892, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P C}} \\left(o _ {i} ^ {t _ {s}}, t _ {e}\\right) = \\frac {1}{| \\Pi |} \\sum_ {\\pi} K L \\left(q _ {\\pi} ^ {t _ {e}} \\mid \\mid \\hat {q}\\right) + H \\left(q _ {\\pi} ^ {t _ {e}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.735 + ], + "angle": 0, + "content": "which includes two terms. The first term is the KL Divergence between each association probability \\( q_{\\pi}^{t_e} \\) and the average of all probabilities, \\( \\hat{q} = \\sum_{\\pi} q_{\\pi}^{t_e} / |\\Pi| \\). Thus, it is minimized only when all probabilities follow the same distribution. The second term minimizes the entropy, \\( H(\\cdot) \\), of each probability to require only one object in frame \\( t_e \\) can be associated to the query object, thus avoids the trivial solution where all \\( q_{\\pi}^{t_e} \\) are consistent yet being uniform distributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Our path consistency loss has the key advantages of \\((\\pmb{i})\\) linear complexity w.r.t the number of paths as we compare each \\(q_{\\pi}^{t_e}\\) to the averaged probability \\(\\hat{q}\\), hence avoid computing pairwise discrepancy between all \\(q_{\\pi}^{t_e}\\). \\((ii)\\) \\(\\hat{q}\\) reflects association consistency and modulates the loss. When the consistency is low, multiple objects can have high probabilities in \\(\\hat{q}\\), indicating different association results. Unlike pseudo-label methods [16, 21, 22] that will enforce a specific association, PCL allows exploring all potential associations by using \\(\\hat{q}\\) in \\(KL(\\cdot)\\), while \\(\\hat{q}\\) also gives higher weights to the more likely ones." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19019" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.405, + 0.108 + ], + "angle": 0, + "content": "3.2.2 Selecting Frame Pairs to compute PCL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.237 + ], + "angle": 0, + "content": "In principle, based on our PCL formulation we can sample any two random frames \\((t_s, t_e)\\) from a video clip, and then use all objects in \\(t_s\\) as query objects. However, we observe such unconstrained frame sampling leads to most query objects not visible in the end frames. The unbalance between visible and disappeared query objects causes the model to always matching objects to the null objects and makes the training hard to converge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.471, + 0.434 + ], + "angle": 0, + "content": "To constrain the frame sampling, we pick frame pairs \\((t_s, t_e)\\) under the condition that one object in \\(t_s\\) should remain visible in intermediate and end frames to serve as the query object while \\(t_s\\) and \\(t_e\\) should be temporally disclose to support learning long-distance matching. As such, we develop a simple heuristics that uses the temporal chain of the IOU similarity to reject most frame pairs. With the existing solution, we observe that the query objects are present in end frames \\(98\\%\\) of all sampled pairs. We include the details of our approach in the supplementary details. Notice that we can also employ more advanced approaches to include more challenging training examples and reject more noisy frame pairs, which we will leave for future works." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.444, + 0.283, + 0.459 + ], + "angle": 0, + "content": "3.3. Regularization Losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.469, + 0.498 + ], + "angle": 0, + "content": "We further introduce the following two losses to regularize the model training together with our path consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.469, + 0.636 + ], + "angle": 0, + "content": "One-to-one matching loss. To avoid that the path for two different query object cross each other, we constrain that a real, non-null object in frame \\( r \\), \\( o_j^r \\), can not be matched by more than one real objects in frame \\( t \\). As elaborated earlier, the probability \\( p(o_i^t \\to o_j^r) \\) will be close to 1 between two matched real objects \\( o_i^t \\) and \\( o_j^r \\). As such, we compute \\( \\sum_{i} p(o_i^t \\to o_j^r) \\) to approximate the number of real objects in frame \\( t \\) that are matched to object \\( o_j^r \\) and penalize the value if it exceeds 1. Mathematically, our loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.647, + 0.47, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {O M}} = \\frac {1}{T ^ {2}} \\sum_ {t, r} \\frac {1}{N ^ {r}} \\sum_ {j} \\max \\left(1, \\sum_ {i} p \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.746 + ], + "angle": 0, + "content": "which penalizes any many-to-one real object matching between every two frames \\((t,r)\\) in the input video clip of length \\(T\\). \\(N^r\\) is the number of real objects in frame \\(r\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Bidirectional consistency loss. In our model, we compute matching probability from frame \\( t \\) to a later frame \\( r \\) (forward in time) but also probability from frame \\( r \\) to frame \\( t \\) (backward in time). To ensure the matching probabilities are invariant to the temporal directions, we design a bidirectional consistency loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.85, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B C}} = \\frac {1}{T ^ {2} N _ {t} N _ {r}} \\sum_ {t, r} \\sum_ {i, j} | | p \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right) - p \\left(o _ {j} ^ {r} \\rightarrow o _ {i} ^ {t}\\right) | | ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "Given Eq. 7 and Eq. 8, our overall model objective is the sum of path consistency loss and two regularization losses: \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{PC}} + \\mathcal{L}_{\\mathrm{OM}} + \\mathcal{L}_{\\mathrm{BC}}\\), where we combine all loss terms with a balance weight of 1.0." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.169, + 0.634, + 0.186 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.195, + 0.688, + 0.211 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.401 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on three tracking datasets: MOT17 [31] is a widely adopted person tracking dataset containing 14 video sequences, featuring crowded scenes in indoor shopping malls or outdoor streets. The videos are split into 7 training sequences and 7 test sequences. **PersonPath22** [39] is a large-scale pedestrian dataset with 236 videos, which are collected from varied environment settings that contain longer occlusions and more crowded scenes. It is split into 138 training videos and 98 test videos. **KITTI** [13] is person and car tracking dataset, featuring videos with lower frame rates. We follow UNS [1] to report performance on car tracking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.495 + ], + "angle": 0, + "content": "Evaluation metrics. Following the evaluation protocols of the datasets, we report our results with several metrics: HOTA (DetA, AssA, LocA), MOTA (IDsw, FP and FN) and IDF1. We emphasize on HOTA, AssA, IDsw and IDF1 as they measure association accuracy. Other metrics are more influenced by the detection accuracy of detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.893, + 0.68 + ], + "angle": 0, + "content": "Implementation details. Our model input is a video clip of \\( T = 48 \\) frames, which we found balances between performance and efficiency. For PCL, we sample at most \\( G = 25 \\) different paths for each query object and end frame and set \\( S = \\sqrt{N} \\) in spatial mask constraint, where \\( \\bar{N} \\) is average number of objects per frame. We implement our model with Pytorch and trained with Adam Optimizer on Tesla V100 and a learning rate of 0.0001. In inference, we use the latest \\( M = 4 \\) bounding boxes of each tracklet to estimate its association with new objects and maintain unmatched tracklets for 30 frames. We provide more implementation details in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.693, + 0.8, + 0.708 + ], + "angle": 0, + "content": "4.2. Comparison to the state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.822 + ], + "angle": 0, + "content": "We compare our method with recent unsupervised object matching methods, where SimpleReID [16], U2MOT [22] are pseudo-label methods and OUTrack [24], UNS [1], UCSL [30] are self-supervised methods. For complete comparison, we include the results of motion-based methods [4, 7, 56], as they do not require object ID annotation, and also results of recent supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison on MOT17. In Table 1, we report the results on MOT17 dataset with both private and public detections. For fair comparison, we use YoloX detections from [56] for private detection and follow the detection pre-processing of UNS [1] for public detection. Our method outperforms" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19020" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.46, + 0.472 + ], + "angle": 0, + "content": "
Sup.HOTA ↑IDF1↑MOTA↑IDsw↓
Private Detection
TransTrack [42]✓54.163.974.53663
TrackFormer [29]✓57.368.074.12829
MeMOT [5]✓56.969.072.52724
MOTR [52]✓57.868.673.42439
MeMOTR [12]✓58.871.572.81902
FairMOT [55]✓59.372.373.71074
FineTrack [34]✓64.379.580.01270
UTM [49]✓64.078.781.81431
MotionTrack [33]✓65.180.181.11140
SimpleReID [16]×50.460.769.0-
OUTrack [24]×-70.273.54110
PointID [40]×-72.474.22748
ByteTrack [56] †×63.177.380.32196
OC-SORT [7] †×63.277.578.01950
UCSL [30]×58.470.473.0-
U2MOT [22] †×64.278.279.91506
Ours †×65.079.680.91749
Public Detection
Tracktor++ [3] ‡✓42.152.353.52072
MHT-BLSTM [18]✓41.051.947.52069
FAMNet [9]✓-48.752.03072
LSST [11]✓47.162.354.71243
GSM [23]✓45.757.856.41485
CenterTrack [58]✓-60.555.72540
TrackFormer [29]✓-62.357.64018
SORT [4] ‡×34.039.843.14852
UNS [1] ‡×46.458.356.81320
Ours ‡×49.061.258.81219
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.48, + 0.472, + 0.539 + ], + "angle": 0, + "content": "Table 1. Comparison on MOT17 [31] dataset. \\(\\uparrow / \\downarrow\\): higher/lower is better. \"Sup.\" means requiring supervision. \"\\(\\dagger\\)\" denotes using YoloX detection. \"\\(\\ddagger\\)\" denotes using Tracktor++ detection preprocessing. The best results of unsupervised methods are in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.611 + ], + "angle": 0, + "content": "all unsupervised works on both private and public detections, and even exceeds recent fully-supervised methods, e.g., UTM [49], FineTrack [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Comparing with pseudo-label methods, we surpass the best competitor, U2MOT [22], by \\(0.8\\%\\) in HOTA, \\(1.4\\%\\) in IDF1 and \\(1\\%\\) in MOTA with the same YoloX detection, as it is vulnerable to errors in pseudo labels. In contrast, our PCL naturally measures model confidence with \\(\\hat{q}\\) and avoids enforcing incorrect, unconfident associations. Meanwhile, we also exceed self-supervised methods [1, 24, 30] as they only learn short-distance matching among adjacent frames. Our method learns object matching over both short and long distances thus is robust to appearance change and occlusion. Using the same public detection as UNS [1], we surpass it by \\(2.6\\%\\) in HOTA, \\(2.9\\%\\) in IDF1 and \\(2\\%\\) in MOTA. Some works [16, 24] use better detectors to improve public detection, hence their performance are not comparable to us." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Comparison on PersonPath22. In Table 2, we report results on PersonPath22 dataset, which is more challenging than MOT17 as its videos contain more crowded scenes with long occlusions. We compare with both public and private detections and successfully outperform ByteTrack" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.891, + 0.279 + ], + "angle": 0, + "content": "
Sup.IDF1↑MOTA↑IDsw ↓FP↓FN↓
CenterTrack [58]✓46.459.31031924K72K
SiamMOT [38]✓53.767.5894213K63K
FairMOT [55]✓61.161.8509515K80K
TrackFormer [29]✓57.169.7863323K47K
SORT [4]×39.653.499272K110K
UNS [1]×41.855.51387717K86K
PointID [40]×63.168.661489K66K
ByteTrack [56] *×57.264.655309k71k
ByteTrack [56] †×66.875.4593117k40k
OC-SORT [7] †×68.274.0493117k41k
Ours*×60.564.850087k73k
Ours †×69.475.8466315k38k
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.288, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Table 2. Comparison on PersonPath22 [39] dataset. ↑/↓: higher/lower is better. The best results of unsupervised methods are in bold. “*” denotes using the public detection of Personpath22 provided by the authors. “†” denotes using YoloX detections." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.361, + 0.88, + 0.53 + ], + "angle": 0, + "content": "
Sup.HOTA↑AssA↑DetA↑MOTA↑
FAMNet [9]✓52.645.561.075.9
mmMOT [54]✓62.154.072.383.2
CenterTrack [58]✓73.071.275.688.8
LGM [47]✓73.172.374.687.6
TuSimple [8]✓71.671.172.686.3
PermaTr [46]✓78.078.478.391.3
SORT [4]×42.541.344.053.2
UNS [1]×62.565.361.1-
OC-SORT [7]×76.576.477.290.8
Ours×78.880.377.991.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.54, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Table 3. Comparison on KITTI [13] dataset (Tracking Car). \\(\\uparrow/\\downarrow\\): higher/lower is better. The best results of unsupervised methods are in bold. We use the same detections as OC-SORT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.793 + ], + "angle": 0, + "content": "by \\(3.3\\%\\) and \\(2.6\\%\\) in IDF1 respectively, also reducing IDsw by large margins. The larger improvement in IDF1 demonstrates the better ability of our model in matching over long temporal distances and re-associate objects over occlusion to deliver more consistent tracking results. With even public detection, our model has achieved better performance than CenterTrack [58], TrackFormer [29] and surpasses FairMOT [55] in MOTA, despite being unsupervised. These results show that our model effectively benefits from more training data, exposing the possibility of large-scale self-supervised learning with enormous tracking videos available online." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "More importantly, in Table 4, we study model performance under challenging scenarios, including different occlusion length (measured by frames), object moving speeds and people density. Our model consistently outperform the best competitor, ByteTrack, by large margins, thanks to our PCL learning more robust object matching to occlusion and appearance changes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19021" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.429, + 0.164 + ], + "angle": 0, + "content": "
Occlusion Length (frames)
01-2020-4040-6060+
UNS [1]49.236.832.827.026.1
OC-SORT [7]77.961.950.139.339.2
Ours79.264.350.740.739.8
" + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.173, + 0.446, + 0.252 + ], + "angle": 0, + "content": "
Movement SpeedPeople Density
slowmediumfastlowhigh
UNS [1]42.740.338.344.742.4
OC-SORT [7]65.963.661.769.764.8
Ours68.365.162.671.466.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.262, + 0.468, + 0.292 + ], + "angle": 0, + "content": "Table 4. Model Performance (IDF1) under different challenging scenarios on PersonPath22." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.304, + 0.456, + 0.418 + ], + "angle": 0, + "content": "
|t-r|1-45-89-1617-3232-48
PCL (s=T)99.599.198.196.394.5
s=3299.599.298.296.393.0
s=2499.599.198.094.888.0
s=1699.699.198.091.481.7
s=899.599.195.784.671.3
s=499.598.091.878.364.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.421, + 0.47, + 0.464 + ], + "angle": 0, + "content": "Table 5. Accuracy of the matching probability \\( p(o_i^t \\rightarrow o_j^r) \\). Our method has learned accurate object matching over short to long temporal distances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.468, + 0.594 + ], + "angle": 0, + "content": "Comparison on KITTI. In Table 3, we further report our performance on KITTI dataset for tracking car, following the evaluation protocol of UNS [1]. KITTI features varied moving velocity of objects and lower frame rate on average, thus objects have large appearance and location changes in adjacent frames [7]. We use the shared detections from OC-SORT [7]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.468, + 0.762 + ], + "angle": 0, + "content": "Our model exceeds the best competitor, OC-SORT [7], by \\(2.3\\%\\) in HOTA, \\(3.9\\%\\) in AssA, \\(0.7\\%\\) in DetA and \\(0.2\\%\\) in MOTA. This is because cars have changes in their moving speeds and bounding box aspect ratios during driving (e.g., when making turn), which leads to low overlap between bounding boxes of the same object in consecutive frames and failure of pure motion trackers [4, 7]. On the other hand, our method learns object embeddings to encode object appearance thus successfully links cars in various driving scenarios, further validating our learned object matching is robust to appearance and location changes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.231, + 0.787 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.468, + 0.868 + ], + "angle": 0, + "content": "We conduct ablation study to examine our key design choices. We train our unsupervised models on MOT17 test set and evaluate on MOT17 train set, since the annotations of the former set are not publicly available. Finally, we use the publicly released SDP detections." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.885 + ], + "angle": 0, + "content": "Has the model learned long-distance object matching?" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We exam the accuracy of our learned matching probability," + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.09, + 0.88, + 0.167 + ], + "angle": 0, + "content": "
L=0L=10L=20L=30L=40L=50L=60
UNS64.062.558.055.053.552.050.0
OC-SORT67.766.965.865.163.963.562.5
Ours68.968.367.466.966.165.965.3
Gain+1.2+1.4+1.6+1.9+2.2+2.4+2.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.177, + 0.892, + 0.233 + ], + "angle": 0, + "content": "Table 6. Comparison of IDF1 for tracking over occlusion. When increasing the minimal occlusion length (number of frames), our model has more robust performance and exceeds the best competitor OC-SORT by large margins (as shown in last row)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.397 + ], + "angle": 0, + "content": "\\(p(o_{i}^{t}\\to o_{j}^{r})\\), where \\(p\\) is correct if \\(o_t^r\\) is the correct matching of \\(o_i^t\\) and has the largest matching probability. In the first two rows of Table 5, we show the accuracy and break it down by the temporal distances between objects \\(|t - r|\\). It shows our matching probability is robust to temporal distances. As \\(|t - r|\\) increases, the accuracy drops only by small margins and remains \\(94.5\\%\\) even at the distance of 32-48 frames, validating our model has learned object matching from short to long distances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Can the model only learn short-distance matching? Next, we further exam if learning long-distance matching is necessary and if a model only learned short-distance matching can generalize to long-distance matching? In rows 2-6 of Table 5, we limit the maximal number of continuous frames that can be skipped in PCL, denoted by \\( s \\). A smaller \\( s \\) means the model is only trained to match objects over short temporal distances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.718 + ], + "angle": 0, + "content": "As \\( s \\) decreases from 32 to 4, while accuracy for short distances (1-8) remains high, it deteriorates quickly for longer distances. Comparing to the first row where we do not limit \\( s \\), accuracy drops by \\( 30.4\\% \\) on the distance of 32-48. It shows that model trained for short-distance matching cannot generalize to longer distances, which is the drawback of previous self-supervised methods [1, 24, 30]. Our PCL has successfully learned matching probability over varied temporal distances while not requiring groundtruth annotations. We also observe that \\( 50\\% \\) of frame skipping in paths are longer than 8 frames, which boosts our model robustness for long distance matching and the appearance changes caused by it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Can the model handle long occlusion? After validating the accuracy of probability matching, we exam our model ability for handling occlusion at inference. We notice that most occlusions in MOT17 videos are of short lengths, e.g., less than 5 frames, which cannot reflect a model's ability to handle long-distance matching. Hence, we increase the occlusion length in a video by keep dropping the bounding boxes of objects after occlusion, until the occlusion lengths reach \\( L \\) frames. If the original length of an occlusion is longer than \\( L \\), we do not increase it. Thus, \\( L = 0 \\) means no occlusion is extended while \\( L = 60 \\) means all occlusions last for at least 60 frames." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19022" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.091, + 0.3, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.19, + 0.233, + 0.2 + ], + "angle": 0, + "content": "Frame 1" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.091, + 0.482, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.19, + 0.421, + 0.2 + ], + "angle": 0, + "content": "Frame 20" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.091, + 0.665, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.19, + 0.609, + 0.2 + ], + "angle": 0, + "content": "Frame 80 (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.091, + 0.846, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.19, + 0.793, + 0.199 + ], + "angle": 0, + "content": "Frame 80 (UNS)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.204, + 0.894, + 0.247 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison between our model and UNS [1]. We visualize the tracking on three frames. UNS cannot track the person in green bounding once he is occluded and assigns him a new ID (purple) on frame 80. It also fails to track the person in pink bounding box when she is only partially visible in frame 80. We can consistently track with both people with the same IDs (green, pink)." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.252, + 0.468, + 0.382 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\text{PC}} \\)\\( \\mathcal{L}_{\\text{OM}} \\)\\( \\mathcal{L}_{\\text{BC}} \\)IDF1HOTAMOTAIDsw
1×✓✓3.48.31.470k
2(a)✓✓26.926.837.829k
3(b)✓✓66.659.463.9329
4✓×✓62.757.163.4591
5✓✓×66.359.863.6466
6✓✓✓68.960.963.7257
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Table 7. Ablation study on proposed losses. In row 1-3, we demonstrate the effect of our path consistency loss. In row 4-5, we show the effect of our regularization losses. \"(a)\" denotes randomly sampling start and end frames for paths and \"(b)\" denotes disabling spatial constraint mask in Eq 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.683 + ], + "angle": 0, + "content": "In Table 6, we report IDF1 with different occlusion length \\( L \\) and compare with UNS and one of the state-of-the-art trackers, OC-SORT, as a strong baseline. Since other related methods [16, 22, 24, 30] do not release their codes or models, we are unable to compare with them. As can be seen, our model consistently provides higher IDF1 for all ranges of occlusion lengths. More importantly, when increasing occlusion lengths, the performance of UNS and OC-SORT decreases quickly, while our performance is more robust and decreases slowly, which echos our results in Table 4. Our IDF1 at \\( L = 50 \\) is still higher than the IDF1 of OC-SORT at \\( L = 20 \\), showing our model can learn to match the same objects over varied temporal distances and is more robust to occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "What is the effect of path consistency loss? In Table 7 (row 1-3), we exam the results of removing PCL and also two modifications of it: (a) randomly sampling the starting and end frames for paths; (b) removing the spatial constraint mask in Eq.5; Firstly, without PCL (row 1), the model cannot learn meaningful object matching using only the regularization losses thus the tracking accuracy is close to zeros. When randomly sampling the starting and end frames (row 2), it causes most query objects are not visible in the end frames, causing imbalance between visible and disappeared objects. Thus, model matches most objects to null objects, also leading to low IDF1. Next, when removing the spatial constraint mask (row 3), the model may incorrectly match two spatial disclose objects, decreasing IDF1 by \\(2.3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.895, + 0.422 + ], + "angle": 0, + "content": "What is the effect of regularization losses? In Table 7 (row 4-5), we ablate the effect of the regularization losses in Section 3.3. As can be seen, while PCL contributes most to the model performance, the two regularization losses \\(\\mathcal{L}_{\\mathrm{OM}}\\), \\(\\mathcal{L}_{\\mathrm{BC}}\\) are also helpful for achieving better performance. Since \\(\\mathcal{L}_{\\mathrm{OM}}\\) prevents multiple objects matching to the same objects and \\(\\mathcal{L}_{\\mathrm{BC}}\\) ensures the consistency between forward and backward matching, removing either of these loss terms leads to around \\(6.2\\%\\) and \\(2.6\\%\\) drop in IDF1, respectively. These results show the collective benefits of using different losses to optimize our unsupervised object tracker." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.677, + 0.446 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.894, + 0.62 + ], + "angle": 0, + "content": "In Figure 3, we visualize the tracking results for one video from MOT17 using our model and UNS [1], where the color indicates their IDs. Both our model and UNS can track two non-occluded front people from frame 1 to frame 80. However, UNS lose the man in green bounding box after he is occluded and assign a new ID (purple) to him on frame 80. It also cannot associate the woman in the pink bounding box when she is only partially visible in frame 80. Our model correctly associate two people with the same IDs (green and pink). This qualitative comparison suggests that our model can better track the object over occlusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.619, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To tackle unsupervised MOT task, we proposed a novel concept of path consistency. We defined a path to track an object as a temporal list of observed frames from the start to the end frame. As such, to track the same object, we can generate multiple paths by skipping intermediate frames while the association results should be consistent across different paths. We leveraged this consistency in our self-supervised Path Consistency Loss to learn object matching model without groundtruth ID supervision. Our PCL especially learns robust long distance matching to tackle occlusion at inference time. With experiments on three popular datasets, our approach has exceeded existing unsupervised methods and performed on par with recent state-of-the-art supervised methods. We provided a comprehensive ablation study to validate the model has learned long-distance matching and is robust under challenging scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19023" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In Advances in Neural Information Processing Systems, pages 13695-13706. Curran Associates, Inc., 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Jerome Berclaz, Francois Fleuret, Engin Turetken, and Pascal Fua. Multiple object tracking using k-shortest paths optimization. TPAMI, 33(9):1806-1819, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.276, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE International Conference on Image Processing (ICIP), pages 3464-3468, 2016. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.47, + 0.376 + ], + "angle": 0, + "content": "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR 2022, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.378, + 0.47, + 0.406 + ], + "angle": 0, + "content": "[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.472, + 0.478 + ], + "angle": 0, + "content": "[7] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[8] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[9] Peng Chu and Haibin Ling. Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6172-6181, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.611, + 0.47, + 0.666 + ], + "angle": 0, + "content": "[10] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.668, + 0.47, + 0.71 + ], + "angle": 0, + "content": "[11] Weitao Feng, Zhihao Hu, Wei Wu, Junjie Yan, and Wanli Ouyang. Multi-object tracking with multiple cues and switcher-aware classification, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.47, + 0.768 + ], + "angle": 0, + "content": "[12] Ruopeng Gao and Limin Wang. MeMOTR: Long-term memory-augmented transformer for multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9901-9910, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[13] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.842 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[15] Roberto Henschel, Laura Leal-Taixe, Daniel Cremers, and Bodo Rosenhahn. Improvements to frank-wolfe optimization for multi-detector multi-object tracking. arXiv preprint arXiv:1705.08314, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Shyamgopal Karthik, Ameya Prabhu, and Vineet Gandhi. Simple unsupervised multi-object tracking. CoRR, abs/2006.02609, 2020. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[17] Chanho Kim, Fuxin Li, Arridhana Ciptadi, and James M Rehg. Multiple hypothesis tracking revisited. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[18] Chanho Kim, Fuxin Li, and James M. Rehg. Multi-object tracking with neural gating using bilinear lstm. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[19] Shuai Li, Yu Kong, and Hamid Rezatofighi. Learning of global objective for network flow in multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8855-8865, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[20] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[21] Yu-Lei Li. Unsupervised embedding and association network for multi-object tracking. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. International Joint Conferences on Artificial Intelligence Organization, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[22] Kai Liu, Sheng Jin, Zhihang Fu, Ze Chen, Rongxin Jiang, and Jieping Ye. Uncertainty-aware unsupervised multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9996-10005, 2023. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[23] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 530–536. International Joint Conferences on Artificial Intelligence Organization, 2020. Main track. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[24] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333-347, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[25] Shuai Liu, Xin Li, Huchuan Lu, and You He. Multi-object tracking meets moving UAV. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8876-8885, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19045-19055, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[27] Fan Ma, Mike Zheng Shou, Linchao Zhu, Haoqi Fan, Yilei Xu, Yi Yang, and Zhicheng Yan. Unified transformer tracker for object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8781-8790, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Christoph Mayer, Martin Danelljan, Goutam Bhat, Matthieu Paul, Danda Pani Paudel, Fisher Yu, and Luc Van Gool." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19024" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Transforming model prediction for tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8731-8740, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[29] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[30] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16264-16273, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[31] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. MOT16: A benchmark for multi-object tracking. CoRR, abs/1603.00831, 2016. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[32] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.445 + ], + "angle": 0, + "content": "[33] Z. Qin, S. Zhou, L. Wang, J. Duan, G. Hua, and W. Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.471, + 0.516 + ], + "angle": 0, + "content": "[34] Hao Ren, Shoudong Han, Huilin Ding, Ziwen Zhang, Hongwei Wang, and Faquan Wang. Focus on details: Online multi-object tracking with diverse fine-grained representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.6 + ], + "angle": 0, + "content": "[36] Ergys Ristani and Carlo Tomasi. Features for multi-target multi-camera tracking and re-identification. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[37] Hao Sheng, Yang Zhang, Jiahui Chen, Zhang Xiong, and Jun Zhang. Heterogeneous association graph fusion for target association in multiple object tracking. TCSVT, 29(11): 3269-3280, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.701 + ], + "angle": 0, + "content": "[38] Bing Shuai, Andrew Berneshawi, Xinyu Li, Davide Modolo, and Joseph Tighe. Siammot: Siamese multi-object tracking. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[39] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joe Tighe. Large scale real-world multi person tracking. In European Conference on Computer Vision. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[40] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14669-14679, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[41] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14689-14699, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[42] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[43] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20993-21002, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[44] Siyu Tang, Mykhaylo Andriluka, Bjoern Andres, and Bernt Schiele. Multiple people tracking by lifted multicut and person re-identification. In CVPR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[45] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[46] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10840-10849, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[47] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[48] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 107-122. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[49] S. You, H. Yao, B. Bao, and C. Xu. Utm: A unified multiple object tracking model with identity-aware feature enhancement. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[50] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8834-8843, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[51] Amir Roshan Zamir, Afshin Dehghan, and Mubarak Shah. Gmcp-tracker: Global multi-object tracking using generalized minimum clique graphs. In ECCV, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[52] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[53] Li Zhang, Yuan Li, and Ramakant Nevatia. Global data association for multi-object tracking using network flows. In CVPR, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[54] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2365-2374, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19025" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[57] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.22 + ], + "angle": 0, + "content": "[58] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. ECCV, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[60] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19026" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_origin.pdf b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..705ed170d03c4294b4d35963cc071b51fe6830f6 --- /dev/null +++ b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/126e321e-aac6-4315-8a05-8ce49f269b09_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af84bb5184aafd57f3fbe65ec561f66990d6e691b1728d36d68432f1969ef391 +size 1337183 diff --git a/2024/Self-Supervised Multi-Object Tracking with Path Consistency/full.md b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7cdd3b0310fdcaf9c261b7f67bdf6cf0ec541a0 --- /dev/null +++ b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/full.md @@ -0,0 +1,319 @@ +# Self-Supervised Multi-Object Tracking with Path Consistency + +Zijia Lu*, Bing Shuai†, Yanbei Chen, Zhenlin Xu, Davide Modolo +AWS AI Labs + +lu.zij@northeastern.edu, {bshuai,yanbec,xzhenlin,dmodolo}@amazon.com + +# Abstract + +In this paper, we propose a novel concept of path consistency to learn robust object matching without using manual object identity supervision. Our key idea is that, to track a object through frames, we can obtain multiple different association results from a model by varying the frames it can observe, i.e., skipping frames in observation. As the differences in observations do not alter the identities of objects, the obtained association results should be consistent. Based on this rationale, we generate multiple observation paths, each specifying a different set of frames to be skipped, and formulate the Path Consistency Loss that enforces the association results are consistent across different observation paths. We use the proposed loss to train our object matching model with only self-supervision. By extensive experiments on three tracking datasets (MOT17, PersonPath22, KITTI), we demonstrate that our method outperforms existing unsupervised methods with consistent margins on various evaluation metrics, and even achieves performance close to supervised methods. + +# 1. Introduction + +Multi-Object Tracking (MOT) is the task of identifying and tracking all object instances present in a video. It benefits a wide spectrum of applications, such as monitoring vehicle traffic, wildlife activity, pedestrians, etc. While there have been several advanced supervised methods developed for MOT [5, 8, 29, 33, 34, 46, 52, 55, 58], the overall advancement in MOT is slower compared to other vision tasks, mostly due to the prohibitively expensive cost of acquiring high-quality object identities annotations for MOT training. As of today, MOT17 [31], a dataset of 14 videos, remains one of the most widely used datasets in MOT. + +To overcome this problem, some recent works [1, 16, 21, 22, 24, 30] started shifting their interest on unsupervised learning for MOT. These approaches leverage off-the-shelf object detectors to localize all the object instances in each + +![](images/4f7fc16452a84e4b19f61616876f0c38b443b62bd3769352b8da47c7aad0f882.jpg) +Figure 1. We propose a novel concept of path consistency for self-supervised MOT. We define a observation path for an object as a temporal list of observed frames from the start to end frame. As such, for the same object, we can generate multiple paths by skipping intermediate frames. As different observations of the same object does not alter its identity, the association results should be consistent across different paths. + +frame of a video and then learn the temporal object association across frames in an unsupervised manner. Multiple works [16, 21, 22] propose to learn the object associations by creating pseudo-tracklets to estimate object IDs. However, they are vulnerable to the noises in tracklets, especially when the groundtruth tracklet of an object is incorrectly broken into several pseudo-tracklets, each being given a different object ID from the others. To avoid relying on pseudo-labels, other works [24, 30] explore self-supervision with temporal consistency objective, by tracking objects from the current frame to a future frame then back to the current frame while requiring each object is associated back to itself. However, these works only consider future frames in very short temporal distances (e.g., 1-2 frames) where objects lack substantial appearance difference, thus limiting the quality of their object embeddings. Meanwhile, UNS [1] leverages a motion model and an appearance model to create cross-supervision between their tracking outputs, yet it also only considers matching objects among adjacent frames. + +In fact, the ability of robust object matching over long + +temporal distances is critical for MOT models as objects often encounter occlusion during tracking, where their bounding boxes are lost for a potentially long period of time, thus cannot be tracked by mere frame-by-frame matching. It requires long-distance matching to re-associate their bounding boxes after occlusion. However, very few works have managed to propose reliable solutions that tackle occlusion without using groundtruth supervision. + +In this paper, we introduce the novel concept of path consistency, which can be used as a reliable self-supervised signal to learn a robust object matching model. To elaborate, when tracking an object, we define an observation path of it as a temporal list of observed frames between a start frame and a end frame. As shown in Figure 1, we are able to generate multiple different observation paths for the same object by randomly skipping intermediate frames. Based on an observation path, we can compute the temporal association between the object and all other objects from the start frame to the end frame. As such, path consistency dictates that the temporal associations should be consistent across paths, since the identities of objects do not change in different paths. Based on this rationale, we propose the Path Consistency Loss (PCL) that estimates the associations of objects along different paths and trains the object matching model by minimizing the discrepancy among association probability derived from different paths. + +Distinct from prior self-supervised methods [1, 24, 30], our formulation learns object matching over both short and long distances. As paths contain frame skipping of varied lengths, the model trained with PCL is able to learn not only short-distance matching (from paths with minimal frame skipping), but also long-distance matching (from paths with consecutive frame skipping) which is essential to handle occlusion during inference. In addition, we sample challenging training examples where paths span long temporal horizons, hence enabling our path consistency formulation to learn robust object matching models and handle occlusion. + +We evaluate our method on three benchmark tracking datasets, MOT17 [31], PersonPath22 [39] and KITTI [13], and show it successfully achieves new state-of-the-art results over prior unsupervised methods and performs even close to recent supervised methods. We also provide extensive ablation studies to validate the design of our framework and improvement in handling occlusion. Our contributions are summarized as follows: + +- We propose the novel concept of path consistency that can be used as a reliable self-supervised signal to train a robust object matching model. Based on the concept, we formulate the Path Consistency Loss (PCL) that is capable of learning robust long-distance matching, which is essential in handling occlusion and large appearance changes in tracking. +- We establish new state-of-the-art performance over prior + +unsupervised methods on three challenging tracking benchmarks. Extensive ablation studies also validate our improvement in long-distance matching and tracking over occlusion. + +# 2. Related Work + +Supervised multi-object tracking. Most supervised MOT methods [2, 5, 15, 17, 19, 25-29, 36-38, 43, 44, 44, 50, 51, 53, 55, 58, 60] follow the tracking-by-detection paradigm where all object instances are first detected in each frame, then they are linked based on their similarities across frames to form tracklets. This paradigm benefits from the advances in object detection [6, 14, 35, 45, 57, 59] and has two threads of works: offline and online MOT. Offline MOT methods [2, 15, 17, 36, 37, 44, 51, 53] focus on finding global optimal associations of objects and are computationally heavy in general. Online MOT methods [3, 5, 29, 32, 38, 48, 52, 55, 56, 58, 60] focus on optimizing local optimum on associating detected objects and have real-time speed, e.g., Tracktor [3] uses bounding box regression to estimate new positions of objects. Meanwhile, the joint learning of the detection and object matching has made rapid progress in performance [48, 55], e.g., [5, 29, 52] introduce transformers and object tokens to associate objects with existing tracklets or to identify new ones. Despite the efficacy of supervised methods, they require costly human-annotated labels [10, 13, 31, 39]. In this work, we learn an online object matching model without labels, which we believe is an important step towards mitigating the limited size of tracking datasets. + +Unsupervised multi-object tracking. This task aims to alleviate the high annotation cost of tracking labels. While existing motion trackers [4, 7, 56] do not require labels for training, they cannot leverage the appearance information of objects in tracking. Thus often lose objects in the event of movement change or long occlusion. To leverage appearance information, recent unsupervised methods focused on developing new object matching models [1, 22, 24, 30]. [20] proposes a setting of open-vocabulary MOT. + +Among recent works, pseudo-label methods [16, 22] generate pseudo-tracklets to train their object matching models. SimepleReID [16] and UEANet [21] obtain pseudo-tracklets from motion trackers without detecting tracklet errors, thus those errors propagate to their models. U2MOT [22] improves the consistency of object IDs within tracklets, yet is still vulnerable to errors in the early stage of training and, more importantly, errors between tracklets. Specifically, the groundtruth tracklet of an object is often broken into several pseudo-tracklets if it experiences occlusions, where each pseudo-tracklet is then given a different object ID, treated as negative samples of each other. On the other hand, self-supervised methods [1, 24, 30] derive new + +unsupervised training objective to replace pseudo-labels. OUTrack [24] and UCSL [30] utilize temporal consistency, i.e., an object tracked from current frame to a future frame then back to current frame should associate back to itself. However, this objective does not hold if objects disappear among the frames, which limits those methods to only allow future frames at near distances (1-2 frames) to ensure objects likely remain present. Unfortunately, objects lack substantial appearance changes in such short distances, causing their failure to handle occlusion by matching objects over long temporal distances. We show that object matching learned with short distances cannot generalize to long distances (Table 5). UNS [1] creates cross-supervision between motion and appearance trackers yet also learn short-distance matching. It discusses occlusion-based supervision yet has failed and discarded it, likely due to its focus on short-distance matching. Meanwhile, PointId [41] learns a person embedding model by data augmentation, but it requires pretraining on ReID datasets with groundtruth labels. + +Unlike existing methods, we propose a novel self-supervised path consistency loss that simultaneously learns object matching over short to long distances without requiring pseudo or groundtruth labels. Thus we are able to consistently track objects in the event of long occlusion. + +# 3. Methodology + +# 3.1. Framework Overview + +We tackle the unsupervised MOT task with the concept of Path Consistency, which trains a robust object matching model capable of associating objects over long temporal distances, enabling consistent tracking in the event of occlusion. Following existing unsupervised works [1, 37, 44], we assume the bounding boxes of objects in each video frame are provided by an off-the-shelf detector. Our object matching model takes the detected objects and estimate object embeddings, which encode the appearance and spatial location of each object. Specifically, let $o_i^t$ denote the $i$ -th object in frame $t$ and $h_i^t$ its embedding. We define the probability that $o_i^t$ matches to an object $o_j^r$ in another frame $r$ as + +$$ +p \left(o _ {i} ^ {t} \rightarrow o _ {j} ^ {r}\right) = \frac {\exp \left(h _ {i} ^ {t} \cdot h _ {j} ^ {r}\right)}{\sum_ {u = 1} ^ {N _ {r}} \exp \left(h _ {i} ^ {t} \cdot h _ {u} ^ {r}\right)}, \tag {1} +$$ + +where $N_r$ is the total number of objects in frame $r$ . Similarly, we also compute $p(o_j^r \to o_i^t)$ , the matching probability from $o_j^r$ to $o_i^t$ . Importantly, to deal with occlusions and object exit in object association, we add a special null object, $\phi$ , to the list of localized objects in each frame. If a real object $o_i^t$ is no longer visible in the following frame $r$ , it gets matched to $\phi$ . + +During inference, our method uses the learned probability $p$ to associate objects to existing tracklets. Given a + +tracklet $\tau = \{o_m^{t_m}\}_{m = 1}^M$ , consisting of the latest $M$ objects, it estimates the similarity between $\tau$ and a new object $o_i^t$ as + +$$ +\frac {1}{M} \sum_ {m} \operatorname {g m} \left[ p \left(o _ {m} ^ {t _ {m}} \rightarrow o _ {i} ^ {t}\right), p \left(o _ {i} ^ {t} \rightarrow o _ {m} ^ {t _ {m}}\right)\right], \tag {2} +$$ + +in which, gm denotes the geometric mean operator. Similar to all other works [22, 24, 30, 55, 56], we perform bipartite matching based on the tracklet-object similarities to either extend / terminate existing tracklets, or initiate new ones. + +In the absence of tracking annotations, we propose a novel path consistency loss to learn the object matching model over varied temporal distances with only self-supervision (Section 3.2) and two regularization losses to improve $\beta$ convergence (Section 3.3). As our method is agnostic to model structure and inference procedure, we provide their details in the supplementary materials. + +# 3.2. Self Supervision with Path Consistency + +In this section, we elaborate the concept of path consistency as a reliable self-supervision signal to train a robust object matching model. Technically, suppose that we have a query object $o_{i}^{t_{s}}$ in the start frame $t_{s}$ , we aim to associating it to its correspondent object, $o_{j}^{t_{e}}$ , in the end frame $t_{e}$ ( $t_{e} > t_{s}$ ). Note that the corresponding object $o_{j}^{t_{e}}$ is the null object in the case that the query object is not visible in frame $t_{e}$ (due to being occluded or exiting the scene). + +We define an observation path as the list of frames a model can observe to compute the association probability distribution between $o_i^{t_s}$ and all visible objects in frame $t_e$ . We are able to construct multiple paths by randomly skipping frames, thus creating a different visual observation for each path. Take the example in Figure 2 (right) as illustration, all frames are observed in Path1 whereas only a subset of frames are observed in Path2 and Path3. The core principle of the path consistency is, despite the different observations in the paths, the identities of objects remain unchanged, thus the association probability between $o_i^{t_s}$ and $o_j^{t_e}$ should always be higher than all other alternatives. That is, the association probability distribution between $o_i^{t_s}$ and all visible objects in frame $t_e$ should be consistent across paths. Based on this principle, we introduce our path consistency loss in Section 3.2.1, and then present our method to sample frames $(t_s, t_e)$ in Section 3.2.2. + +# 3.2.1 Path Consistency Loss + +Path Consistency Loss (PCL) computes the association probability distributions from all possible paths, $\{q_{\pi}|\pi \sim \Pi \}$ , and enforces consistency among them1, where $\Pi$ is the set of all paths and $q_{\pi}$ is the association probability distribution obtained using a specific path + +![](images/7f8c75a50e14eafe345dc1d2036338f6e8efa1aa2cc1507d45712794ffe0bb9a.jpg) +Figure 2. Overview of Path Consistency Loss (PCL). Our method takes a video clip as input, where objects are localized by an off-the-shelf detector, and uses a selection strategy to choose suitable query objects and their correspondent end frames, then computes PCL to learn association between query objects and objects in end frames. Association probabilities obtained from different paths provide cross-supervision among them and enables self-supervised model learning. + +$\pi = \{t_s, t_1, \dots, t_K, t_e\}$ . $t_k$ is an intermediate frame in the path. We use $q_{\pi}^{t_k}(j)$ to denote the probability that query object $o_i^{t_s}$ matches with the $j$ -th object at frame $t_k$ . + +Computing Association. To obtain association probability $q_{\pi}$ , we aggregate the matching probability $p$ defined in Eq.1 according to the path $\pi$ . We have + +$$ +q _ {\pi} ^ {t _ {k}} (j) = \left\{\begin{array}{l l}p \left(o _ {i} ^ {t _ {s}} \rightarrow o _ {j} ^ {t _ {1}}\right),&\text {i f} t _ {k} = t _ {1}\\\sum_ {u} q _ {\pi} ^ {t _ {k - 1}} (u) \cdot p \left(o _ {u} ^ {t _ {k - 1}} \rightarrow o _ {j} ^ {t _ {k}}\right),&\text {o t h e r w i s e}\end{array}\right. \tag {3} +$$ + +where the first case defines the association from query object to objects in the first intermediate frame $t_1$ , which is the matching probability, $p$ , between the objects. The second case extends the association from a frame $t_{k-1}$ to $t_k$ by jointly considering the matching probability from all objects in $t_{k-1}$ to $o_j^{t_k}$ . In contrast, pseudo-label methods [16, 21, 22] only consider the most likely matching, leading to error accumulation. Overall, we have the association probability between query object and objects in end frame as + +$$ +q _ {\pi} ^ {t _ {e}} (j) = \sum_ {u} p \left(o _ {i} ^ {t _ {s}} \rightarrow o _ {u} ^ {t _ {1}}\right) \dots \sum_ {v} p \left(o _ {v} ^ {t _ {K}} \rightarrow o _ {j} ^ {t _ {e}}\right). \tag {4} +$$ + +which cumulatively multiplies the matching probabilities according to the frames in path. + +To reduce the space of possible associations and facilitate learning, we constrain that only two spatially close objects between two neighboring frames can be matched. We define a binary spatial constraint mask $C_{t_k,v}^{t_{k-1},u}$ , which equals to zero if $o_v^{t_k}$ is one of the $S$ furthest objects to $o_u^{t_{k-1}}$ . + +Thus, we update $q_{\pi}^{t_e}$ as + +$$ +\frac {1}{Z} \sum_ {u} \left(C _ {t _ {1}, u} ^ {t _ {s}, i}\right) p \left(o _ {i} ^ {t _ {s}} \rightarrow o _ {u} ^ {t _ {1}}\right) \dots \sum_ {v} \left(C _ {t _ {e}, j} ^ {t _ {K}, v}\right) p \left(o _ {v} ^ {t _ {K}} \rightarrow o _ {j} ^ {t _ {e}}\right), \tag {5} +$$ + +where $Z$ is a normalization factor to ensure the probability $q_{\pi}^{t_e}$ sums to 1. + +Computing Loss. After computing the association probabilities for all paths $\{q_{\pi}^{t_e}|\pi \sim \Pi \}$ , We define the path consistency loss to enforce consistency among them to learn our model with self-supervision, + +$$ +\mathcal {L} _ {\mathrm {P C}} \left(o _ {i} ^ {t _ {s}}, t _ {e}\right) = \frac {1}{| \Pi |} \sum_ {\pi} K L \left(q _ {\pi} ^ {t _ {e}} \mid \mid \hat {q}\right) + H \left(q _ {\pi} ^ {t _ {e}}\right), \tag {6} +$$ + +which includes two terms. The first term is the KL Divergence between each association probability $q_{\pi}^{t_e}$ and the average of all probabilities, $\hat{q} = \sum_{\pi} q_{\pi}^{t_e} / |\Pi|$ . Thus, it is minimized only when all probabilities follow the same distribution. The second term minimizes the entropy, $H(\cdot)$ , of each probability to require only one object in frame $t_e$ can be associated to the query object, thus avoids the trivial solution where all $q_{\pi}^{t_e}$ are consistent yet being uniform distributions. + +Our path consistency loss has the key advantages of $(\pmb{i})$ linear complexity w.r.t the number of paths as we compare each $q_{\pi}^{t_e}$ to the averaged probability $\hat{q}$ , hence avoid computing pairwise discrepancy between all $q_{\pi}^{t_e}$ . $(ii)$ $\hat{q}$ reflects association consistency and modulates the loss. When the consistency is low, multiple objects can have high probabilities in $\hat{q}$ , indicating different association results. Unlike pseudo-label methods [16, 21, 22] that will enforce a specific association, PCL allows exploring all potential associations by using $\hat{q}$ in $KL(\cdot)$ , while $\hat{q}$ also gives higher weights to the more likely ones. + +# 3.2.2 Selecting Frame Pairs to compute PCL + +In principle, based on our PCL formulation we can sample any two random frames $(t_s, t_e)$ from a video clip, and then use all objects in $t_s$ as query objects. However, we observe such unconstrained frame sampling leads to most query objects not visible in the end frames. The unbalance between visible and disappeared query objects causes the model to always matching objects to the null objects and makes the training hard to converge. + +To constrain the frame sampling, we pick frame pairs $(t_s, t_e)$ under the condition that one object in $t_s$ should remain visible in intermediate and end frames to serve as the query object while $t_s$ and $t_e$ should be temporally disclose to support learning long-distance matching. As such, we develop a simple heuristics that uses the temporal chain of the IOU similarity to reject most frame pairs. With the existing solution, we observe that the query objects are present in end frames $98\%$ of all sampled pairs. We include the details of our approach in the supplementary details. Notice that we can also employ more advanced approaches to include more challenging training examples and reject more noisy frame pairs, which we will leave for future works. + +# 3.3. Regularization Losses + +We further introduce the following two losses to regularize the model training together with our path consistency loss. + +One-to-one matching loss. To avoid that the path for two different query object cross each other, we constrain that a real, non-null object in frame $r$ , $o_j^r$ , can not be matched by more than one real objects in frame $t$ . As elaborated earlier, the probability $p(o_i^t \to o_j^r)$ will be close to 1 between two matched real objects $o_i^t$ and $o_j^r$ . As such, we compute $\sum_{i} p(o_i^t \to o_j^r)$ to approximate the number of real objects in frame $t$ that are matched to object $o_j^r$ and penalize the value if it exceeds 1. Mathematically, our loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {O M}} = \frac {1}{T ^ {2}} \sum_ {t, r} \frac {1}{N ^ {r}} \sum_ {j} \max \left(1, \sum_ {i} p \left(o _ {i} ^ {t} \rightarrow o _ {j} ^ {r}\right)\right), \tag {7} +$$ + +which penalizes any many-to-one real object matching between every two frames $(t,r)$ in the input video clip of length $T$ . $N^r$ is the number of real objects in frame $r$ . + +Bidirectional consistency loss. In our model, we compute matching probability from frame $t$ to a later frame $r$ (forward in time) but also probability from frame $r$ to frame $t$ (backward in time). To ensure the matching probabilities are invariant to the temporal directions, we design a bidirectional consistency loss as: + +$$ +\mathcal {L} _ {\mathrm {B C}} = \frac {1}{T ^ {2} N _ {t} N _ {r}} \sum_ {t, r} \sum_ {i, j} | | p \left(o _ {i} ^ {t} \rightarrow o _ {j} ^ {r}\right) - p \left(o _ {j} ^ {r} \rightarrow o _ {i} ^ {t}\right) | | ^ {2}. \tag {8} +$$ + +Given Eq. 7 and Eq. 8, our overall model objective is the sum of path consistency loss and two regularization losses: $\mathcal{L} = \mathcal{L}_{\mathrm{PC}} + \mathcal{L}_{\mathrm{OM}} + \mathcal{L}_{\mathrm{BC}}$ , where we combine all loss terms with a balance weight of 1.0. + +# 4. Experiments + +# 4.1. Experimental setup + +Datasets. We conduct experiments on three tracking datasets: MOT17 [31] is a widely adopted person tracking dataset containing 14 video sequences, featuring crowded scenes in indoor shopping malls or outdoor streets. The videos are split into 7 training sequences and 7 test sequences. **PersonPath22** [39] is a large-scale pedestrian dataset with 236 videos, which are collected from varied environment settings that contain longer occlusions and more crowded scenes. It is split into 138 training videos and 98 test videos. **KITTI** [13] is person and car tracking dataset, featuring videos with lower frame rates. We follow UNS [1] to report performance on car tracking. + +Evaluation metrics. Following the evaluation protocols of the datasets, we report our results with several metrics: HOTA (DetA, AssA, LocA), MOTA (IDsw, FP and FN) and IDF1. We emphasize on HOTA, AssA, IDsw and IDF1 as they measure association accuracy. Other metrics are more influenced by the detection accuracy of detectors. + +Implementation details. Our model input is a video clip of $T = 48$ frames, which we found balances between performance and efficiency. For PCL, we sample at most $G = 25$ different paths for each query object and end frame and set $S = \sqrt{N}$ in spatial mask constraint, where $\bar{N}$ is average number of objects per frame. We implement our model with Pytorch and trained with Adam Optimizer on Tesla V100 and a learning rate of 0.0001. In inference, we use the latest $M = 4$ bounding boxes of each tracklet to estimate its association with new objects and maintain unmatched tracklets for 30 frames. We provide more implementation details in the supplementary materials. + +# 4.2. Comparison to the state-of-the-art + +We compare our method with recent unsupervised object matching methods, where SimpleReID [16], U2MOT [22] are pseudo-label methods and OUTrack [24], UNS [1], UCSL [30] are self-supervised methods. For complete comparison, we include the results of motion-based methods [4, 7, 56], as they do not require object ID annotation, and also results of recent supervised methods. + +Comparison on MOT17. In Table 1, we report the results on MOT17 dataset with both private and public detections. For fair comparison, we use YoloX detections from [56] for private detection and follow the detection pre-processing of UNS [1] for public detection. Our method outperforms + +
Sup.HOTA ↑IDF1↑MOTA↑IDsw↓
Private Detection
TransTrack [42]✓54.163.974.53663
TrackFormer [29]✓57.368.074.12829
MeMOT [5]✓56.969.072.52724
MOTR [52]✓57.868.673.42439
MeMOTR [12]✓58.871.572.81902
FairMOT [55]✓59.372.373.71074
FineTrack [34]✓64.379.580.01270
UTM [49]✓64.078.781.81431
MotionTrack [33]✓65.180.181.11140
SimpleReID [16]×50.460.769.0-
OUTrack [24]×-70.273.54110
PointID [40]×-72.474.22748
ByteTrack [56] †×63.177.380.32196
OC-SORT [7] †×63.277.578.01950
UCSL [30]×58.470.473.0-
U2MOT [22] †×64.278.279.91506
Ours †×65.079.680.91749
Public Detection
Tracktor++ [3] ‡✓42.152.353.52072
MHT-BLSTM [18]✓41.051.947.52069
FAMNet [9]✓-48.752.03072
LSST [11]✓47.162.354.71243
GSM [23]✓45.757.856.41485
CenterTrack [58]✓-60.555.72540
TrackFormer [29]✓-62.357.64018
SORT [4] ‡×34.039.843.14852
UNS [1] ‡×46.458.356.81320
Ours ‡×49.061.258.81219
+ +all unsupervised works on both private and public detections, and even exceeds recent fully-supervised methods, e.g., UTM [49], FineTrack [34]. + +Comparing with pseudo-label methods, we surpass the best competitor, U2MOT [22], by $0.8\%$ in HOTA, $1.4\%$ in IDF1 and $1\%$ in MOTA with the same YoloX detection, as it is vulnerable to errors in pseudo labels. In contrast, our PCL naturally measures model confidence with $\hat{q}$ and avoids enforcing incorrect, unconfident associations. Meanwhile, we also exceed self-supervised methods [1, 24, 30] as they only learn short-distance matching among adjacent frames. Our method learns object matching over both short and long distances thus is robust to appearance change and occlusion. Using the same public detection as UNS [1], we surpass it by $2.6\%$ in HOTA, $2.9\%$ in IDF1 and $2\%$ in MOTA. Some works [16, 24] use better detectors to improve public detection, hence their performance are not comparable to us. + +Comparison on PersonPath22. In Table 2, we report results on PersonPath22 dataset, which is more challenging than MOT17 as its videos contain more crowded scenes with long occlusions. We compare with both public and private detections and successfully outperform ByteTrack + +Table 1. Comparison on MOT17 [31] dataset. $\uparrow / \downarrow$ : higher/lower is better. "Sup." means requiring supervision. " $\dagger$ " denotes using YoloX detection. " $\ddagger$ " denotes using Tracktor++ detection preprocessing. The best results of unsupervised methods are in bold. + +
Sup.IDF1↑MOTA↑IDsw ↓FP↓FN↓
CenterTrack [58]✓46.459.31031924K72K
SiamMOT [38]✓53.767.5894213K63K
FairMOT [55]✓61.161.8509515K80K
TrackFormer [29]✓57.169.7863323K47K
SORT [4]×39.653.499272K110K
UNS [1]×41.855.51387717K86K
PointID [40]×63.168.661489K66K
ByteTrack [56] *×57.264.655309k71k
ByteTrack [56] †×66.875.4593117k40k
OC-SORT [7] †×68.274.0493117k41k
Ours*×60.564.850087k73k
Ours †×69.475.8466315k38k
+ +Table 2. Comparison on PersonPath22 [39] dataset. ↑/↓: higher/lower is better. The best results of unsupervised methods are in bold. “*” denotes using the public detection of Personpath22 provided by the authors. “†” denotes using YoloX detections. + +
Sup.HOTA↑AssA↑DetA↑MOTA↑
FAMNet [9]✓52.645.561.075.9
mmMOT [54]✓62.154.072.383.2
CenterTrack [58]✓73.071.275.688.8
LGM [47]✓73.172.374.687.6
TuSimple [8]✓71.671.172.686.3
PermaTr [46]✓78.078.478.391.3
SORT [4]×42.541.344.053.2
UNS [1]×62.565.361.1-
OC-SORT [7]×76.576.477.290.8
Ours×78.880.377.991.0
+ +Table 3. Comparison on KITTI [13] dataset (Tracking Car). $\uparrow/\downarrow$ : higher/lower is better. The best results of unsupervised methods are in bold. We use the same detections as OC-SORT. + +by $3.3\%$ and $2.6\%$ in IDF1 respectively, also reducing IDsw by large margins. The larger improvement in IDF1 demonstrates the better ability of our model in matching over long temporal distances and re-associate objects over occlusion to deliver more consistent tracking results. With even public detection, our model has achieved better performance than CenterTrack [58], TrackFormer [29] and surpasses FairMOT [55] in MOTA, despite being unsupervised. These results show that our model effectively benefits from more training data, exposing the possibility of large-scale self-supervised learning with enormous tracking videos available online. + +More importantly, in Table 4, we study model performance under challenging scenarios, including different occlusion length (measured by frames), object moving speeds and people density. Our model consistently outperform the best competitor, ByteTrack, by large margins, thanks to our PCL learning more robust object matching to occlusion and appearance changes. + +
Occlusion Length (frames)
01-2020-4040-6060+
UNS [1]49.236.832.827.026.1
OC-SORT [7]77.961.950.139.339.2
Ours79.264.350.740.739.8
+ +
Movement SpeedPeople Density
slowmediumfastlowhigh
UNS [1]42.740.338.344.742.4
OC-SORT [7]65.963.661.769.764.8
Ours68.365.162.671.466.3
+ +Table 4. Model Performance (IDF1) under different challenging scenarios on PersonPath22. + +
|t-r|1-45-89-1617-3232-48
PCL (s=T)99.599.198.196.394.5
s=3299.599.298.296.393.0
s=2499.599.198.094.888.0
s=1699.699.198.091.481.7
s=899.599.195.784.671.3
s=499.598.091.878.364.1
+ +Comparison on KITTI. In Table 3, we further report our performance on KITTI dataset for tracking car, following the evaluation protocol of UNS [1]. KITTI features varied moving velocity of objects and lower frame rate on average, thus objects have large appearance and location changes in adjacent frames [7]. We use the shared detections from OC-SORT [7]. + +Our model exceeds the best competitor, OC-SORT [7], by $2.3\%$ in HOTA, $3.9\%$ in AssA, $0.7\%$ in DetA and $0.2\%$ in MOTA. This is because cars have changes in their moving speeds and bounding box aspect ratios during driving (e.g., when making turn), which leads to low overlap between bounding boxes of the same object in consecutive frames and failure of pure motion trackers [4, 7]. On the other hand, our method learns object embeddings to encode object appearance thus successfully links cars in various driving scenarios, further validating our learned object matching is robust to appearance and location changes. + +# 4.3. Ablation Study + +We conduct ablation study to examine our key design choices. We train our unsupervised models on MOT17 test set and evaluate on MOT17 train set, since the annotations of the former set are not publicly available. Finally, we use the publicly released SDP detections. + +Has the model learned long-distance object matching? + +We exam the accuracy of our learned matching probability, + +Table 5. Accuracy of the matching probability $p(o_i^t \rightarrow o_j^r)$ . Our method has learned accurate object matching over short to long temporal distances. + +
L=0L=10L=20L=30L=40L=50L=60
UNS64.062.558.055.053.552.050.0
OC-SORT67.766.965.865.163.963.562.5
Ours68.968.367.466.966.165.965.3
Gain+1.2+1.4+1.6+1.9+2.2+2.4+2.8
+ +Table 6. Comparison of IDF1 for tracking over occlusion. When increasing the minimal occlusion length (number of frames), our model has more robust performance and exceeds the best competitor OC-SORT by large margins (as shown in last row). + +$p(o_{i}^{t}\to o_{j}^{r})$ , where $p$ is correct if $o_t^r$ is the correct matching of $o_i^t$ and has the largest matching probability. In the first two rows of Table 5, we show the accuracy and break it down by the temporal distances between objects $|t - r|$ . It shows our matching probability is robust to temporal distances. As $|t - r|$ increases, the accuracy drops only by small margins and remains $94.5\%$ even at the distance of 32-48 frames, validating our model has learned object matching from short to long distances. + +Can the model only learn short-distance matching? Next, we further exam if learning long-distance matching is necessary and if a model only learned short-distance matching can generalize to long-distance matching? In rows 2-6 of Table 5, we limit the maximal number of continuous frames that can be skipped in PCL, denoted by $s$ . A smaller $s$ means the model is only trained to match objects over short temporal distances. + +As $s$ decreases from 32 to 4, while accuracy for short distances (1-8) remains high, it deteriorates quickly for longer distances. Comparing to the first row where we do not limit $s$ , accuracy drops by $30.4\%$ on the distance of 32-48. It shows that model trained for short-distance matching cannot generalize to longer distances, which is the drawback of previous self-supervised methods [1, 24, 30]. Our PCL has successfully learned matching probability over varied temporal distances while not requiring groundtruth annotations. We also observe that $50\%$ of frame skipping in paths are longer than 8 frames, which boosts our model robustness for long distance matching and the appearance changes caused by it. + +Can the model handle long occlusion? After validating the accuracy of probability matching, we exam our model ability for handling occlusion at inference. We notice that most occlusions in MOT17 videos are of short lengths, e.g., less than 5 frames, which cannot reflect a model's ability to handle long-distance matching. Hence, we increase the occlusion length in a video by keep dropping the bounding boxes of objects after occlusion, until the occlusion lengths reach $L$ frames. If the original length of an occlusion is longer than $L$ , we do not increase it. Thus, $L = 0$ means no occlusion is extended while $L = 60$ means all occlusions last for at least 60 frames. + +![](images/8fd927a3f4ab848bcc7d5ac751795898bb1bccbd4b668c497d1166031cd2340d.jpg) +Frame 1 + +![](images/859c6361e0aacaa21e860d6d8f00f4bac4d3da18bf9d5c8f328dfc22e2bbbbf9.jpg) +Frame 20 +Figure 3. Qualitative comparison between our model and UNS [1]. We visualize the tracking on three frames. UNS cannot track the person in green bounding once he is occluded and assigns him a new ID (purple) on frame 80. It also fails to track the person in pink bounding box when she is only partially visible in frame 80. We can consistently track with both people with the same IDs (green, pink). + +![](images/8844fe0e04d3f73a936e2e6b4b7cb9b04a31c72b8b7b292baacef8f0ce386c42.jpg) +Frame 80 (Ours) + +![](images/bebda331f59d9621356899e1c2591344d029dbeb2464d6635a289fabbb8bf16f.jpg) +Frame 80 (UNS) + +
\( \mathcal{L}_{\text{PC}} \)\( \mathcal{L}_{\text{OM}} \)\( \mathcal{L}_{\text{BC}} \)IDF1HOTAMOTAIDsw
1×✓✓3.48.31.470k
2(a)✓✓26.926.837.829k
3(b)✓✓66.659.463.9329
4✓×✓62.757.163.4591
5✓✓×66.359.863.6466
6✓✓✓68.960.963.7257
+ +Table 7. Ablation study on proposed losses. In row 1-3, we demonstrate the effect of our path consistency loss. In row 4-5, we show the effect of our regularization losses. "(a)" denotes randomly sampling start and end frames for paths and "(b)" denotes disabling spatial constraint mask in Eq 5. + +In Table 6, we report IDF1 with different occlusion length $L$ and compare with UNS and one of the state-of-the-art trackers, OC-SORT, as a strong baseline. Since other related methods [16, 22, 24, 30] do not release their codes or models, we are unable to compare with them. As can be seen, our model consistently provides higher IDF1 for all ranges of occlusion lengths. More importantly, when increasing occlusion lengths, the performance of UNS and OC-SORT decreases quickly, while our performance is more robust and decreases slowly, which echos our results in Table 4. Our IDF1 at $L = 50$ is still higher than the IDF1 of OC-SORT at $L = 20$ , showing our model can learn to match the same objects over varied temporal distances and is more robust to occlusion. + +What is the effect of path consistency loss? In Table 7 (row 1-3), we exam the results of removing PCL and also two modifications of it: (a) randomly sampling the starting and end frames for paths; (b) removing the spatial constraint mask in Eq.5; Firstly, without PCL (row 1), the model cannot learn meaningful object matching using only the regularization losses thus the tracking accuracy is close to zeros. When randomly sampling the starting and end frames (row 2), it causes most query objects are not visible in the end frames, causing imbalance between visible and disappeared objects. Thus, model matches most objects to null objects, also leading to low IDF1. Next, when removing the spatial constraint mask (row 3), the model may incorrectly match two spatial disclose objects, decreasing IDF1 by $2.3\%$ . + +What is the effect of regularization losses? In Table 7 (row 4-5), we ablate the effect of the regularization losses in Section 3.3. As can be seen, while PCL contributes most to the model performance, the two regularization losses $\mathcal{L}_{\mathrm{OM}}$ , $\mathcal{L}_{\mathrm{BC}}$ are also helpful for achieving better performance. Since $\mathcal{L}_{\mathrm{OM}}$ prevents multiple objects matching to the same objects and $\mathcal{L}_{\mathrm{BC}}$ ensures the consistency between forward and backward matching, removing either of these loss terms leads to around $6.2\%$ and $2.6\%$ drop in IDF1, respectively. These results show the collective benefits of using different losses to optimize our unsupervised object tracker. + +# 4.4. Qualitative results + +In Figure 3, we visualize the tracking results for one video from MOT17 using our model and UNS [1], where the color indicates their IDs. Both our model and UNS can track two non-occluded front people from frame 1 to frame 80. However, UNS lose the man in green bounding box after he is occluded and assign a new ID (purple) to him on frame 80. It also cannot associate the woman in the pink bounding box when she is only partially visible in frame 80. Our model correctly associate two people with the same IDs (green and pink). This qualitative comparison suggests that our model can better track the object over occlusion. + +# 5. Conclusion + +To tackle unsupervised MOT task, we proposed a novel concept of path consistency. We defined a path to track an object as a temporal list of observed frames from the start to the end frame. As such, to track the same object, we can generate multiple paths by skipping intermediate frames while the association results should be consistent across different paths. We leveraged this consistency in our self-supervised Path Consistency Loss to learn object matching model without groundtruth ID supervision. Our PCL especially learns robust long distance matching to tackle occlusion at inference time. With experiments on three popular datasets, our approach has exceeded existing unsupervised methods and performed on par with recent state-of-the-art supervised methods. We provided a comprehensive ablation study to validate the model has learned long-distance matching and is robust under challenging scenarios. + +# References + +[1] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In Advances in Neural Information Processing Systems, pages 13695-13706. Curran Associates, Inc., 2021. 1, 2, 3, 5, 6, 7, 8 +[2] Jerome Berclaz, Francois Fleuret, Engin Turetken, and Pascal Fua. Multiple object tracking using k-shortest paths optimization. TPAMI, 33(9):1806-1819, 2011. 2 +[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6 +[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE International Conference on Image Processing (ICIP), pages 3464-3468, 2016. 2, 5, 6, 7 +[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR 2022, 2022. 1, 2, 6 +[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 2 +[7] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 2, 5, 6, 7 +[8] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015. 1, 6 +[9] Peng Chu and Haibin Ling. Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6172-6181, 2019. 6 +[10] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes, 2020. 2 +[11] Weitao Feng, Zhihao Hu, Wei Wu, Junjie Yan, and Wanli Ouyang. Multi-object tracking with multiple cues and switcher-aware classification, 2019. 6 +[12] Ruopeng Gao and Limin Wang. MeMOTR: Long-term memory-augmented transformer for multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9901-9910, 2023. 6 +[13] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 2, 5, 6 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2 +[15] Roberto Henschel, Laura Leal-Taixe, Daniel Cremers, and Bodo Rosenhahn. Improvements to frank-wolfe optimization for multi-detector multi-object tracking. arXiv preprint arXiv:1705.08314, 2017. 2 + +[16] Shyamgopal Karthik, Ameya Prabhu, and Vineet Gandhi. Simple unsupervised multi-object tracking. CoRR, abs/2006.02609, 2020. 1, 2, 4, 5, 6, 8 +[17] Chanho Kim, Fuxin Li, Arridhana Ciptadi, and James M Rehg. Multiple hypothesis tracking revisited. In ICCV, 2015. 2 +[18] Chanho Kim, Fuxin Li, and James M. Rehg. Multi-object tracking with neural gating using bilinear lstm. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6 +[19] Shuai Li, Yu Kong, and Hamid Rezatofighi. Learning of global objective for network flow in multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8855-8865, 2022. 2 +[20] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 2 +[21] Yu-Lei Li. Unsupervised embedding and association network for multi-object tracking. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. International Joint Conferences on Artificial Intelligence Organization, 2022. 1, 2, 4 +[22] Kai Liu, Sheng Jin, Zhihang Fu, Ze Chen, Rongxin Jiang, and Jieping Ye. Uncertainty-aware unsupervised multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9996-10005, 2023. 1, 2, 3, 4, 5, 6, 8 +[23] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 530–536. International Joint Conferences on Artificial Intelligence Organization, 2020. Main track. 6 +[24] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333-347, 2022. 1, 2, 3, 5, 6, 7, 8 +[25] Shuai Liu, Xin Li, Huchuan Lu, and You He. Multi-object tracking meets moving UAV. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8876-8885, 2022. 2 +[26] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19045-19055, 2022. +[27] Fan Ma, Mike Zheng Shou, Linchao Zhu, Haoqi Fan, Yilei Xu, Yi Yang, and Zhicheng Yan. Unified transformer tracker for object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8781-8790, 2022. +[28] Christoph Mayer, Martin Danelljan, Goutam Bhat, Matthieu Paul, Danda Pani Paudel, Fisher Yu, and Luc Van Gool. + +Transforming model prediction for tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8731-8740, 2022. +[29] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6 +[30] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16264-16273, 2023. 1, 2, 3, 5, 6, 7, 8 +[31] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. MOT16: A benchmark for multi-object tracking. CoRR, abs/1603.00831, 2016. 1, 2, 5, 6 +[32] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2 +[33] Z. Qin, S. Zhou, L. Wang, J. Duan, G. Hua, and W. Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6 +[34] Hao Ren, Shoudong Han, Huilin Ding, Ziwen Zhang, Hongwei Wang, and Faquan Wang. Focus on details: Online multi-object tracking with diverse fine-grained representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2 +[36] Ergys Ristani and Carlo Tomasi. Features for multi-target multi-camera tracking and re-identification. In CVPR, 2018. 2 +[37] Hao Sheng, Yang Zhang, Jiahui Chen, Zhang Xiong, and Jun Zhang. Heterogeneous association graph fusion for target association in multiple object tracking. TCSVT, 29(11): 3269-3280, 2018. 2, 3 +[38] Bing Shuai, Andrew Berneshawi, Xinyu Li, Davide Modolo, and Joseph Tighe. Siammot: Siamese multi-object tracking. In CVPR, 2021. 2, 6 +[39] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joe Tighe. Large scale real-world multi person tracking. In European Conference on Computer Vision. Springer, 2022. 2, 5, 6 +[40] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14669-14679, 2022. 6 +[41] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14689-14699, 2022. 3 +[42] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: + +Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460, 2020. 6 +[43] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20993-21002, 2022. 2 +[44] Siyu Tang, Mykhaylo Andriluka, Bjoern Andres, and Bernt Schiele. Multiple people tracking by lifted multicut and person re-identification. In CVPR, 2017. 2, 3 +[45] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2 +[46] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10840-10849, 2021. 1, 6 +[47] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021. 6 +[48] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 107-122. Springer, 2020. 2 +[49] S. You, H. Yao, B. Bao, and C. Xu. Utm: A unified multiple object tracking model with identity-aware feature enhancement. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[50] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8834-8843, 2022. 2 +[51] Amir Roshan Zamir, Afshin Dehghan, and Mubarak Shah. Gmcp-tracker: Global multi-object tracking using generalized minimum clique graphs. In ECCV, 2012. 2 +[52] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6 +[53] Li Zhang, Yuan Li, and Ramakant Nevatia. Global data association for multi-object tracking using network flows. In CVPR, 2008. 2 +[54] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2365-2374, 2019. 6 +[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1, 2, 3, 6 + +[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6 +[57] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2 +[58] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. ECCV, 2020. 1, 2, 6 +[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 2 +[60] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022. 2 \ No newline at end of file diff --git a/2024/Self-Supervised Multi-Object Tracking with Path Consistency/images.zip b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c46fa155f5234de9cbde1e87bfd0107bea3406f9 --- /dev/null +++ b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f888b84365c032f5547fc78470066bd59dd94142bc877db0b60d161b2ebd03a +size 501185 diff --git a/2024/Self-Supervised Multi-Object Tracking with Path Consistency/layout.json b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c7f502d400d9d1bd20440393b0e40e7e440591b5 --- /dev/null +++ b/2024/Self-Supervised Multi-Object Tracking with Path Consistency/layout.json @@ -0,0 +1,10005 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 488, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 488, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 488, + 121 + ], + "type": "text", + "content": "Self-Supervised Multi-Object Tracking with Path Consistency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 143, + 465, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 143, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 127, + 143, + 465, + 171 + ], + "type": "text", + "content": "Zijia Lu*, Bing Shuai†, Yanbei Chen, Zhenlin Xu, Davide Modolo \nAWS AI Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 174, + 485, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 174, + 485, + 185 + ], + "spans": [ + { + "bbox": [ + 109, + 174, + 485, + 185 + ], + "type": "text", + "content": "lu.zij@northeastern.edu, {bshuai,yanbec,xzhenlin,dmodolo}@amazon.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 467 + ], + "type": "text", + "content": "In this paper, we propose a novel concept of path consistency to learn robust object matching without using manual object identity supervision. Our key idea is that, to track a object through frames, we can obtain multiple different association results from a model by varying the frames it can observe, i.e., skipping frames in observation. As the differences in observations do not alter the identities of objects, the obtained association results should be consistent. Based on this rationale, we generate multiple observation paths, each specifying a different set of frames to be skipped, and formulate the Path Consistency Loss that enforces the association results are consistent across different observation paths. We use the proposed loss to train our object matching model with only self-supervision. By extensive experiments on three tracking datasets (MOT17, PersonPath22, KITTI), we demonstrate that our method outperforms existing unsupervised methods with consistent margins on various evaluation metrics, and even achieves performance close to supervised methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 475, + 128, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 128, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 128, + 488 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 495, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 626 + ], + "type": "text", + "content": "Multi-Object Tracking (MOT) is the task of identifying and tracking all object instances present in a video. It benefits a wide spectrum of applications, such as monitoring vehicle traffic, wildlife activity, pedestrians, etc. While there have been several advanced supervised methods developed for MOT [5, 8, 29, 33, 34, 46, 52, 55, 58], the overall advancement in MOT is slower compared to other vision tasks, mostly due to the prohibitively expensive cost of acquiring high-quality object identities annotations for MOT training. As of today, MOT17 [31], a dataset of 14 videos, remains one of the most widely used datasets in MOT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 627, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 288, + 676 + ], + "type": "text", + "content": "To overcome this problem, some recent works [1, 16, 21, 22, 24, 30] started shifting their interest on unsupervised learning for MOT. These approaches leverage off-the-shelf object detectors to localize all the object instances in each" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 545, + 345 + ], + "type": "image", + "image_path": "4f7fc16452a84e4b19f61616876f0c38b443b62bd3769352b8da47c7aad0f882.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 355, + 547, + 433 + ], + "lines": [ + { + "bbox": [ + 305, + 355, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 547, + 433 + ], + "type": "text", + "content": "Figure 1. We propose a novel concept of path consistency for self-supervised MOT. We define a observation path for an object as a temporal list of observed frames from the start to end frame. As such, for the same object, we can generate multiple paths by skipping intermediate frames. As different observations of the same object does not alter its identity, the association results should be consistent across different paths." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "frame of a video and then learn the temporal object association across frames in an unsupervised manner. Multiple works [16, 21, 22] propose to learn the object associations by creating pseudo-tracklets to estimate object IDs. However, they are vulnerable to the noises in tracklets, especially when the groundtruth tracklet of an object is incorrectly broken into several pseudo-tracklets, each being given a different object ID from the others. To avoid relying on pseudo-labels, other works [24, 30] explore self-supervision with temporal consistency objective, by tracking objects from the current frame to a future frame then back to the current frame while requiring each object is associated back to itself. However, these works only consider future frames in very short temporal distances (e.g., 1-2 frames) where objects lack substantial appearance difference, thus limiting the quality of their object embeddings. Meanwhile, UNS [1] leverages a motion model and an appearance model to create cross-supervision between their tracking outputs, yet it also only considers matching objects among adjacent frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In fact, the ability of robust object matching over long" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 702 + ], + "type": "text", + "content": "*Currently at Northeastern University. Work conducted during an internship with AWS AI Labs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19016" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "temporal distances is critical for MOT models as objects often encounter occlusion during tracking, where their bounding boxes are lost for a potentially long period of time, thus cannot be tracked by mere frame-by-frame matching. It requires long-distance matching to re-associate their bounding boxes after occlusion. However, very few works have managed to propose reliable solutions that tackle occlusion without using groundtruth supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 384 + ], + "type": "text", + "content": "In this paper, we introduce the novel concept of path consistency, which can be used as a reliable self-supervised signal to learn a robust object matching model. To elaborate, when tracking an object, we define an observation path of it as a temporal list of observed frames between a start frame and a end frame. As shown in Figure 1, we are able to generate multiple different observation paths for the same object by randomly skipping intermediate frames. Based on an observation path, we can compute the temporal association between the object and all other objects from the start frame to the end frame. As such, path consistency dictates that the temporal associations should be consistent across paths, since the identities of objects do not change in different paths. Based on this rationale, we propose the Path Consistency Loss (PCL) that estimates the associations of objects along different paths and trains the object matching model by minimizing the discrepancy among association probability derived from different paths." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 387, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 518 + ], + "type": "text", + "content": "Distinct from prior self-supervised methods [1, 24, 30], our formulation learns object matching over both short and long distances. As paths contain frame skipping of varied lengths, the model trained with PCL is able to learn not only short-distance matching (from paths with minimal frame skipping), but also long-distance matching (from paths with consecutive frame skipping) which is essential to handle occlusion during inference. In addition, we sample challenging training examples where paths span long temporal horizons, hence enabling our path consistency formulation to learn robust object matching models and handle occlusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "content": "We evaluate our method on three benchmark tracking datasets, MOT17 [31], PersonPath22 [39] and KITTI [13], and show it successfully achieves new state-of-the-art results over prior unsupervised methods and performs even close to recent supervised methods. We also provide extensive ablation studies to validate the design of our framework and improvement in handling occlusion. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 618, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 618, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 287, + 701 + ], + "type": "text", + "content": "- We propose the novel concept of path consistency that can be used as a reliable self-supervised signal to train a robust object matching model. Based on the concept, we formulate the Path Consistency Loss (PCL) that is capable of learning robust long-distance matching, which is essential in handling occlusion and large appearance changes in tracking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "- We establish new state-of-the-art performance over prior" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 119 + ], + "type": "text", + "content": "unsupervised methods on three challenging tracking benchmarks. Extensive ablation studies also validate our improvement in long-distance matching and tracking over occlusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 136, + 392, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 136, + 392, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 136, + 392, + 148 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 159, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 445 + ], + "type": "text", + "content": "Supervised multi-object tracking. Most supervised MOT methods [2, 5, 15, 17, 19, 25-29, 36-38, 43, 44, 44, 50, 51, 53, 55, 58, 60] follow the tracking-by-detection paradigm where all object instances are first detected in each frame, then they are linked based on their similarities across frames to form tracklets. This paradigm benefits from the advances in object detection [6, 14, 35, 45, 57, 59] and has two threads of works: offline and online MOT. Offline MOT methods [2, 15, 17, 36, 37, 44, 51, 53] focus on finding global optimal associations of objects and are computationally heavy in general. Online MOT methods [3, 5, 29, 32, 38, 48, 52, 55, 56, 58, 60] focus on optimizing local optimum on associating detected objects and have real-time speed, e.g., Tracktor [3] uses bounding box regression to estimate new positions of objects. Meanwhile, the joint learning of the detection and object matching has made rapid progress in performance [48, 55], e.g., [5, 29, 52] introduce transformers and object tokens to associate objects with existing tracklets or to identify new ones. Despite the efficacy of supervised methods, they require costly human-annotated labels [10, 13, 31, 39]. In this work, we learn an online object matching model without labels, which we believe is an important step towards mitigating the limited size of tracking datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 556 + ], + "type": "text", + "content": "Unsupervised multi-object tracking. This task aims to alleviate the high annotation cost of tracking labels. While existing motion trackers [4, 7, 56] do not require labels for training, they cannot leverage the appearance information of objects in tracking. Thus often lose objects in the event of movement change or long occlusion. To leverage appearance information, recent unsupervised methods focused on developing new object matching models [1, 22, 24, 30]. [20] proposes a setting of open-vocabulary MOT." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Among recent works, pseudo-label methods [16, 22] generate pseudo-tracklets to train their object matching models. SimepleReID [16] and UEANet [21] obtain pseudo-tracklets from motion trackers without detecting tracklet errors, thus those errors propagate to their models. U2MOT [22] improves the consistency of object IDs within tracklets, yet is still vulnerable to errors in the early stage of training and, more importantly, errors between tracklets. Specifically, the groundtruth tracklet of an object is often broken into several pseudo-tracklets if it experiences occlusions, where each pseudo-tracklet is then given a different object ID, treated as negative samples of each other. On the other hand, self-supervised methods [1, 24, 30] derive new" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19017" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "unsupervised training objective to replace pseudo-labels. OUTrack [24] and UCSL [30] utilize temporal consistency, i.e., an object tracked from current frame to a future frame then back to current frame should associate back to itself. However, this objective does not hold if objects disappear among the frames, which limits those methods to only allow future frames at near distances (1-2 frames) to ensure objects likely remain present. Unfortunately, objects lack substantial appearance changes in such short distances, causing their failure to handle occlusion by matching objects over long temporal distances. We show that object matching learned with short distances cannot generalize to long distances (Table 5). UNS [1] creates cross-supervision between motion and appearance trackers yet also learn short-distance matching. It discusses occlusion-based supervision yet has failed and discarded it, likely due to its focus on short-distance matching. Meanwhile, PointId [41] learns a person embedding model by data augmentation, but it requires pretraining on ReID datasets with groundtruth labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 361 + ], + "type": "text", + "content": "Unlike existing methods, we propose a novel self-supervised path consistency loss that simultaneously learns object matching over short to long distances without requiring pseudo or groundtruth labels. Thus we are able to consistently track objects in the event of long occlusion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 372, + 129, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 129, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 129, + 386 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 172, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 172, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 172, + 403 + ], + "type": "text", + "content": "3.1. Framework Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": "We tackle the unsupervised MOT task with the concept of Path Consistency, which trains a robust object matching model capable of associating objects over long temporal distances, enabling consistent tracking in the event of occlusion. Following existing unsupervised works [1, 37, 44], we assume the bounding boxes of objects in each video frame are provided by an off-the-shelf detector. Our object matching model takes the detected objects and estimate object embeddings, which encode the appearance and spatial location of each object. Specifically, let " + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "inline_equation", + "content": "o_i^t" + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": "-th object in frame " + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "inline_equation", + "content": "h_i^t" + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": " its embedding. We define the probability that " + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "inline_equation", + "content": "o_i^t" + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": " matches to an object " + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "inline_equation", + "content": "o_j^r" + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": " in another frame " + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 411, + 289, + 557 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 566, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 566, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 94, + 566, + 287, + 597 + ], + "type": "interline_equation", + "content": "p \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right) = \\frac {\\exp \\left(h _ {i} ^ {t} \\cdot h _ {j} ^ {r}\\right)}{\\sum_ {u = 1} ^ {N _ {r}} \\exp \\left(h _ {i} ^ {t} \\cdot h _ {u} ^ {r}\\right)}, \\tag {1}", + "image_path": "d1e2d802c0a5bc271bc4cc2738ec5dd0d6b4d445debc65a6ae270f23345992bc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": " is the total number of objects in frame " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": ". Similarly, we also compute " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "p(o_j^r \\to o_i^t)" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": ", the matching probability from " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "o_j^r" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "o_i^t" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": ". Importantly, to deal with occlusions and object exit in object association, we add a special null object, " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": ", to the list of localized objects in each frame. If a real object " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "o_i^t" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": " is no longer visible in the following frame " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": ", it gets matched to " + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "During inference, our method uses the learned probability " + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": " to associate objects to existing tracklets. Given a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "tracklet " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\tau = \\{o_m^{t_m}\\}_{m = 1}^M" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": ", consisting of the latest " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " objects, it estimates the similarity between " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " and a new object " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "o_i^t" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 104, + 545, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 104, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 344, + 104, + 545, + 132 + ], + "type": "interline_equation", + "content": "\\frac {1}{M} \\sum_ {m} \\operatorname {g m} \\left[ p \\left(o _ {m} ^ {t _ {m}} \\rightarrow o _ {i} ^ {t}\\right), p \\left(o _ {i} ^ {t} \\rightarrow o _ {m} ^ {t _ {m}}\\right)\\right], \\tag {2}", + "image_path": "816cc4b89c8b74200b3036ff131944ae1b3a6d1fac3a904ec2f138dc68fa40ef.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 140, + 547, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 547, + 186 + ], + "type": "text", + "content": "in which, gm denotes the geometric mean operator. Similar to all other works [22, 24, 30, 55, 56], we perform bipartite matching based on the tracklet-object similarities to either extend / terminate existing tracklets, or initiate new ones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 188, + 547, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 547, + 271 + ], + "type": "text", + "content": "In the absence of tracking annotations, we propose a novel path consistency loss to learn the object matching model over varied temporal distances with only self-supervision (Section 3.2) and two regularization losses to improve " + }, + { + "bbox": [ + 304, + 188, + 547, + 271 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 188, + 547, + 271 + ], + "type": "text", + "content": " convergence (Section 3.3). As our method is agnostic to model structure and inference procedure, we provide their details in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 277, + 511, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 511, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 511, + 291 + ], + "type": "text", + "content": "3.2. Self Supervision with Path Consistency" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": "In this section, we elaborate the concept of path consistency as a reliable self-supervision signal to train a robust object matching model. Technically, suppose that we have a query object " + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "inline_equation", + "content": "o_{i}^{t_{s}}" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": " in the start frame " + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "inline_equation", + "content": "t_{s}" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": ", we aim to associating it to its correspondent object, " + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "inline_equation", + "content": "o_{j}^{t_{e}}" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": ", in the end frame " + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "inline_equation", + "content": "t_{e}" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "inline_equation", + "content": "t_{e} > t_{s}" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": "). Note that the corresponding object " + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "inline_equation", + "content": "o_{j}^{t_{e}}" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": " is the null object in the case that the query object is not visible in frame " + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "inline_equation", + "content": "t_{e}" + }, + { + "bbox": [ + 304, + 297, + 545, + 393 + ], + "type": "text", + "content": " (due to being occluded or exiting the scene)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": "We define an observation path as the list of frames a model can observe to compute the association probability distribution between " + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "inline_equation", + "content": "o_i^{t_s}" + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": " and all visible objects in frame " + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": ". We are able to construct multiple paths by randomly skipping frames, thus creating a different visual observation for each path. Take the example in Figure 2 (right) as illustration, all frames are observed in Path1 whereas only a subset of frames are observed in Path2 and Path3. The core principle of the path consistency is, despite the different observations in the paths, the identities of objects remain unchanged, thus the association probability between " + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "inline_equation", + "content": "o_i^{t_s}" + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "inline_equation", + "content": "o_j^{t_e}" + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": " should always be higher than all other alternatives. That is, the association probability distribution between " + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "inline_equation", + "content": "o_i^{t_s}" + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": " and all visible objects in frame " + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": " should be consistent across paths. Based on this principle, we introduce our path consistency loss in Section 3.2.1, and then present our method to sample frames " + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "inline_equation", + "content": "(t_s, t_e)" + }, + { + "bbox": [ + 304, + 394, + 546, + 600 + ], + "type": "text", + "content": " in Section 3.2.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 608, + 435, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 435, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 435, + 620 + ], + "type": "text", + "content": "3.2.1 Path Consistency Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "content": "Path Consistency Loss (PCL) computes the association probability distributions from all possible paths, " + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\{q_{\\pi}|\\pi \\sim \\Pi \\}" + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "content": ", and enforces consistency among them1, where " + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "content": " is the set of all paths and " + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "inline_equation", + "content": "q_{\\pi}" + }, + { + "bbox": [ + 304, + 626, + 545, + 687 + ], + "type": "text", + "content": " is the association probability distribution obtained using a specific path" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1When the number of paths is large, we conduct Monte Carlo estimation by computing the loss with a smaller set of sampled paths." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19018" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 235 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 235 + ], + "type": "image", + "image_path": "7f8c75a50e14eafe345dc1d2036338f6e8efa1aa2cc1507d45712794ffe0bb9a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 546, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 546, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of Path Consistency Loss (PCL). Our method takes a video clip as input, where objects are localized by an off-the-shelf detector, and uses a selection strategy to choose suitable query objects and their correspondent end frames, then computes PCL to learn association between query objects and objects in end frames. Association probabilities obtained from different paths provide cross-supervision among them and enables self-supervised model learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\pi = \\{t_s, t_1, \\dots, t_K, t_e\\}" + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "text", + "content": " is an intermediate frame in the path. We use " + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "inline_equation", + "content": "q_{\\pi}^{t_k}(j)" + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "text", + "content": " to denote the probability that query object " + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "inline_equation", + "content": "o_i^{t_s}" + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "text", + "content": " matches with the " + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "text", + "content": "-th object at frame " + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 46, + 301, + 288, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": "Computing Association. To obtain association probability " + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "q_{\\pi}" + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": ", we aggregate the matching probability " + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": " defined in Eq.1 according to the path " + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": ". We have" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 287, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 287, + 439 + ], + "type": "interline_equation", + "content": "q _ {\\pi} ^ {t _ {k}} (j) = \\left\\{\\begin{array}{l l}p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {j} ^ {t _ {1}}\\right),&\\text {i f} t _ {k} = t _ {1}\\\\\\sum_ {u} q _ {\\pi} ^ {t _ {k - 1}} (u) \\cdot p \\left(o _ {u} ^ {t _ {k - 1}} \\rightarrow o _ {j} ^ {t _ {k}}\\right),&\\text {o t h e r w i s e}\\end{array}\\right. \\tag {3}", + "image_path": "c820c754d387ec8a30a7f864078fef38bf47c6aeb41e3f3282aaa79c676d9bd1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "content": "where the first case defines the association from query object to objects in the first intermediate frame " + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "content": ", which is the matching probability, " + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "content": ", between the objects. The second case extends the association from a frame " + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "inline_equation", + "content": "t_{k-1}" + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "content": " by jointly considering the matching probability from all objects in " + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "inline_equation", + "content": "t_{k-1}" + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "inline_equation", + "content": "o_j^{t_k}" + }, + { + "bbox": [ + 46, + 453, + 287, + 573 + ], + "type": "text", + "content": ". In contrast, pseudo-label methods [16, 21, 22] only consider the most likely matching, leading to error accumulation. Overall, we have the association probability between query object and objects in end frame as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 585, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 585, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 63, + 585, + 287, + 611 + ], + "type": "interline_equation", + "content": "q _ {\\pi} ^ {t _ {e}} (j) = \\sum_ {u} p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {u} ^ {t _ {1}}\\right) \\dots \\sum_ {v} p \\left(o _ {v} ^ {t _ {K}} \\rightarrow o _ {j} ^ {t _ {e}}\\right). \\tag {4}", + "image_path": "1c0da99fb4f21b8c407ab0dddd1efdd20511a083d031dab953f063201e3427b2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 624, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 649 + ], + "type": "text", + "content": "which cumulatively multiplies the matching probabilities according to the frames in path." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "text", + "content": "To reduce the space of possible associations and facilitate learning, we constrain that only two spatially close objects between two neighboring frames can be matched. We define a binary spatial constraint mask " + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "C_{t_k,v}^{t_{k-1},u}" + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "text", + "content": ", which equals to zero if " + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "o_v^{t_k}" + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "text", + "content": " is one of the " + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "text", + "content": " furthest objects to " + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "inline_equation", + "content": "o_u^{t_{k-1}}" + }, + { + "bbox": [ + 46, + 651, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 301, + 400, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 400, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 400, + 316 + ], + "type": "text", + "content": "Thus, we update " + }, + { + "bbox": [ + 306, + 301, + 400, + 316 + ], + "type": "inline_equation", + "content": "q_{\\pi}^{t_e}" + }, + { + "bbox": [ + 306, + 301, + 400, + 316 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 319, + 545, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 319, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 319, + 545, + 358 + ], + "type": "interline_equation", + "content": "\\frac {1}{Z} \\sum_ {u} \\left(C _ {t _ {1}, u} ^ {t _ {s}, i}\\right) p \\left(o _ {i} ^ {t _ {s}} \\rightarrow o _ {u} ^ {t _ {1}}\\right) \\dots \\sum_ {v} \\left(C _ {t _ {e}, j} ^ {t _ {K}, v}\\right) p \\left(o _ {v} ^ {t _ {K}} \\rightarrow o _ {j} ^ {t _ {e}}\\right), \\tag {5}", + "image_path": "124d32103593e3a996322f62f486a20bdf54948e64973a7d5deb9ff3a9b10f2f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 364, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 545, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 364, + 545, + 389 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 305, + 364, + 545, + 389 + ], + "type": "text", + "content": " is a normalization factor to ensure the probability " + }, + { + "bbox": [ + 305, + 364, + 545, + 389 + ], + "type": "inline_equation", + "content": "q_{\\pi}^{t_e}" + }, + { + "bbox": [ + 305, + 364, + 545, + 389 + ], + "type": "text", + "content": " sums to 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "text", + "content": "Computing Loss. After computing the association probabilities for all paths " + }, + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\{q_{\\pi}^{t_e}|\\pi \\sim \\Pi \\}" + }, + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "text", + "content": ", We define the path consistency loss to enforce consistency among them to learn our model with self-supervision," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 453, + 545, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 453, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 330, + 453, + 545, + 482 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P C}} \\left(o _ {i} ^ {t _ {s}}, t _ {e}\\right) = \\frac {1}{| \\Pi |} \\sum_ {\\pi} K L \\left(q _ {\\pi} ^ {t _ {e}} \\mid \\mid \\hat {q}\\right) + H \\left(q _ {\\pi} ^ {t _ {e}}\\right), \\tag {6}", + "image_path": "7c7b1e5f7ce74d645f67a7b71a3a21d4ebe4b62bcc4b40253121f1b1fd65be8b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "text", + "content": "which includes two terms. The first term is the KL Divergence between each association probability " + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "inline_equation", + "content": "q_{\\pi}^{t_e}" + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "text", + "content": " and the average of all probabilities, " + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\hat{q} = \\sum_{\\pi} q_{\\pi}^{t_e} / |\\Pi|" + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "text", + "content": ". Thus, it is minimized only when all probabilities follow the same distribution. The second term minimizes the entropy, " + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "text", + "content": ", of each probability to require only one object in frame " + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "text", + "content": " can be associated to the query object, thus avoids the trivial solution where all " + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "inline_equation", + "content": "q_{\\pi}^{t_e}" + }, + { + "bbox": [ + 304, + 486, + 545, + 582 + ], + "type": "text", + "content": " are consistent yet being uniform distributions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": "Our path consistency loss has the key advantages of " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "(\\pmb{i})" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " linear complexity w.r.t the number of paths as we compare each " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "q_{\\pi}^{t_e}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " to the averaged probability " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ", hence avoid computing pairwise discrepancy between all " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "q_{\\pi}^{t_e}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " reflects association consistency and modulates the loss. When the consistency is low, multiple objects can have high probabilities in " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ", indicating different association results. Unlike pseudo-label methods [16, 21, 22] that will enforce a specific association, PCL allows exploring all potential associations by using " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "KL(\\cdot)" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " also gives higher weights to the more likely ones." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19019" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 247, + 85 + ], + "type": "text", + "content": "3.2.2 Selecting Frame Pairs to compute PCL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "text", + "content": "In principle, based on our PCL formulation we can sample any two random frames " + }, + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "inline_equation", + "content": "(t_s, t_e)" + }, + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "text", + "content": " from a video clip, and then use all objects in " + }, + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "text", + "content": " as query objects. However, we observe such unconstrained frame sampling leads to most query objects not visible in the end frames. The unbalance between visible and disappeared query objects causes the model to always matching objects to the null objects and makes the training hard to converge." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "text", + "content": "To constrain the frame sampling, we pick frame pairs " + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "inline_equation", + "content": "(t_s, t_e)" + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "text", + "content": " under the condition that one object in " + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "text", + "content": " should remain visible in intermediate and end frames to serve as the query object while " + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "text", + "content": " should be temporally disclose to support learning long-distance matching. As such, we develop a simple heuristics that uses the temporal chain of the IOU similarity to reject most frame pairs. With the existing solution, we observe that the query objects are present in end frames " + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 46, + 188, + 288, + 343 + ], + "type": "text", + "content": " of all sampled pairs. We include the details of our approach in the supplementary details. Notice that we can also employ more advanced approaches to include more challenging training examples and reject more noisy frame pairs, which we will leave for future works." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 351, + 173, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 173, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 173, + 363 + ], + "type": "text", + "content": "3.3. Regularization Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 370, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 394 + ], + "type": "text", + "content": "We further introduce the following two losses to regularize the model training together with our path consistency loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": "One-to-one matching loss. To avoid that the path for two different query object cross each other, we constrain that a real, non-null object in frame " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "o_j^r" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": ", can not be matched by more than one real objects in frame " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": ". As elaborated earlier, the probability " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "p(o_i^t \\to o_j^r)" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": " will be close to 1 between two matched real objects " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "o_i^t" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "o_j^r" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": ". As such, we compute " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\sum_{i} p(o_i^t \\to o_j^r)" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": " to approximate the number of real objects in frame " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": " that are matched to object " + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "inline_equation", + "content": "o_j^r" + }, + { + "bbox": [ + 46, + 396, + 287, + 503 + ], + "type": "text", + "content": " and penalize the value if it exceeds 1. Mathematically, our loss is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 512, + 287, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 287, + 546 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {O M}} = \\frac {1}{T ^ {2}} \\sum_ {t, r} \\frac {1}{N ^ {r}} \\sum_ {j} \\max \\left(1, \\sum_ {i} p \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right)\\right), \\tag {7}", + "image_path": "217d6443d05b719d20b64b332d402be843bca0715b6a9b4b9b4d826163fd037c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": "which penalizes any many-to-one real object matching between every two frames " + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "inline_equation", + "content": "(t,r)" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": " in the input video clip of length " + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "inline_equation", + "content": "N^r" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": " is the number of real objects in frame " + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "text", + "content": "Bidirectional consistency loss. In our model, we compute matching probability from frame " + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "text", + "content": " to a later frame " + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "text", + "content": " (forward in time) but also probability from frame " + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 593, + 287, + 664 + ], + "type": "text", + "content": " (backward in time). To ensure the matching probabilities are invariant to the temporal directions, we design a bidirectional consistency loss as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 673, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B C}} = \\frac {1}{T ^ {2} N _ {t} N _ {r}} \\sum_ {t, r} \\sum_ {i, j} | | p \\left(o _ {i} ^ {t} \\rightarrow o _ {j} ^ {r}\\right) - p \\left(o _ {j} ^ {r} \\rightarrow o _ {i} ^ {t}\\right) | | ^ {2}. \\tag {8}", + "image_path": "55f41fb34d0063aecd8d7cf2ec9b72d887891d8a6921f681e2487b906282d9fd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "Given Eq. 7 and Eq. 8, our overall model objective is the sum of path consistency loss and two regularization losses: " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{PC}} + \\mathcal{L}_{\\mathrm{OM}} + \\mathcal{L}_{\\mathrm{BC}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": ", where we combine all loss terms with a balance weight of 1.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 133, + 388, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 388, + 147 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 388, + 147 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 154, + 421, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 154, + 421, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 154, + 421, + 167 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 174, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 317 + ], + "type": "text", + "content": "Datasets. We conduct experiments on three tracking datasets: MOT17 [31] is a widely adopted person tracking dataset containing 14 video sequences, featuring crowded scenes in indoor shopping malls or outdoor streets. The videos are split into 7 training sequences and 7 test sequences. **PersonPath22** [39] is a large-scale pedestrian dataset with 236 videos, which are collected from varied environment settings that contain longer occlusions and more crowded scenes. It is split into 138 training videos and 98 test videos. **KITTI** [13] is person and car tracking dataset, featuring videos with lower frame rates. We follow UNS [1] to report performance on car tracking." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 320, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 392 + ], + "type": "text", + "content": "Evaluation metrics. Following the evaluation protocols of the datasets, we report our results with several metrics: HOTA (DetA, AssA, LocA), MOTA (IDsw, FP and FN) and IDF1. We emphasize on HOTA, AssA, IDsw and IDF1 as they measure association accuracy. Other metrics are more influenced by the detection accuracy of detectors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "text", + "content": "Implementation details. Our model input is a video clip of " + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "inline_equation", + "content": "T = 48" + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "text", + "content": " frames, which we found balances between performance and efficiency. For PCL, we sample at most " + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "inline_equation", + "content": "G = 25" + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "text", + "content": " different paths for each query object and end frame and set " + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "inline_equation", + "content": "S = \\sqrt{N}" + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "text", + "content": " in spatial mask constraint, where " + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "inline_equation", + "content": "\\bar{N}" + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "text", + "content": " is average number of objects per frame. We implement our model with Pytorch and trained with Adam Optimizer on Tesla V100 and a learning rate of 0.0001. In inference, we use the latest " + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 304, + 394, + 546, + 538 + ], + "type": "text", + "content": " bounding boxes of each tracklet to estimate its association with new objects and maintain unmatched tracklets for 30 frames. We provide more implementation details in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 548, + 489, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 489, + 560 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 489, + 560 + ], + "type": "text", + "content": "4.2. Comparison to the state-of-the-art" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": "We compare our method with recent unsupervised object matching methods, where SimpleReID [16], U2MOT [22] are pseudo-label methods and OUTrack [24], UNS [1], UCSL [30] are self-supervised methods. For complete comparison, we include the results of motion-based methods [4, 7, 56], as they do not require object ID annotation, and also results of recent supervised methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Comparison on MOT17. In Table 1, we report the results on MOT17 dataset with both private and public detections. For fair comparison, we use YoloX detections from [56] for private detection and follow the detection pre-processing of UNS [1] for public detection. Our method outperforms" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19020" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 281, + 373 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 281, + 373 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 281, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 281, + 373 + ], + "type": "table", + "html": "
Sup.HOTA ↑IDF1↑MOTA↑IDsw↓
Private Detection
TransTrack [42]✓54.163.974.53663
TrackFormer [29]✓57.368.074.12829
MeMOT [5]✓56.969.072.52724
MOTR [52]✓57.868.673.42439
MeMOTR [12]✓58.871.572.81902
FairMOT [55]✓59.372.373.71074
FineTrack [34]✓64.379.580.01270
UTM [49]✓64.078.781.81431
MotionTrack [33]✓65.180.181.11140
SimpleReID [16]×50.460.769.0-
OUTrack [24]×-70.273.54110
PointID [40]×-72.474.22748
ByteTrack [56] †×63.177.380.32196
OC-SORT [7] †×63.277.578.01950
UCSL [30]×58.470.473.0-
U2MOT [22] †×64.278.279.91506
Ours †×65.079.680.91749
Public Detection
Tracktor++ [3] ‡✓42.152.353.52072
MHT-BLSTM [18]✓41.051.947.52069
FAMNet [9]✓-48.752.03072
LSST [11]✓47.162.354.71243
GSM [23]✓45.757.856.41485
CenterTrack [58]✓-60.555.72540
TrackFormer [29]✓-62.357.64018
SORT [4] ‡×34.039.843.14852
UNS [1] ‡×46.458.356.81320
Ours ‡×49.061.258.81219
", + "image_path": "edcdbebe557525e13c8226946a9b16f4900645f0327651fee997ee95c5f94893.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 447, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 483 + ], + "type": "text", + "content": "all unsupervised works on both private and public detections, and even exceeds recent fully-supervised methods, e.g., UTM [49], FineTrack [34]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": "Comparing with pseudo-label methods, we surpass the best competitor, U2MOT [22], by " + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": " in HOTA, " + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": " in IDF1 and " + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": " in MOTA with the same YoloX detection, as it is vulnerable to errors in pseudo labels. In contrast, our PCL naturally measures model confidence with " + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": " and avoids enforcing incorrect, unconfident associations. Meanwhile, we also exceed self-supervised methods [1, 24, 30] as they only learn short-distance matching among adjacent frames. Our method learns object matching over both short and long distances thus is robust to appearance change and occlusion. Using the same public detection as UNS [1], we surpass it by " + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": " in HOTA, " + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": " in IDF1 and " + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": " in MOTA. Some works [16, 24] use better detectors to improve public detection, hence their performance are not comparable to us." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Comparison on PersonPath22. In Table 2, we report results on PersonPath22 dataset, which is more challenging than MOT17 as its videos contain more crowded scenes with long occlusions. We compare with both public and private detections and successfully outperform ByteTrack" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "type": "text", + "content": "Table 1. Comparison on MOT17 [31] dataset. " + }, + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\uparrow / \\downarrow" + }, + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "type": "text", + "content": ": higher/lower is better. \"Sup.\" means requiring supervision. \"" + }, + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "type": "text", + "content": "\" denotes using YoloX detection. \"" + }, + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 380, + 288, + 426 + ], + "type": "text", + "content": "\" denotes using Tracktor++ detection preprocessing. The best results of unsupervised methods are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 220 + ], + "type": "table", + "html": "
Sup.IDF1↑MOTA↑IDsw ↓FP↓FN↓
CenterTrack [58]✓46.459.31031924K72K
SiamMOT [38]✓53.767.5894213K63K
FairMOT [55]✓61.161.8509515K80K
TrackFormer [29]✓57.169.7863323K47K
SORT [4]×39.653.499272K110K
UNS [1]×41.855.51387717K86K
PointID [40]×63.168.661489K66K
ByteTrack [56] *×57.264.655309k71k
ByteTrack [56] †×66.875.4593117k40k
OC-SORT [7] †×68.274.0493117k41k
Ours*×60.564.850087k73k
Ours †×69.475.8466315k38k
", + "image_path": "1252ae92967acf129b4a7179805b2582840394de4d318d297c569bb0e5f5c23f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 285, + 538, + 419 + ], + "blocks": [ + { + "bbox": [ + 305, + 228, + 546, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 546, + 275 + ], + "type": "text", + "content": "Table 2. Comparison on PersonPath22 [39] dataset. ↑/↓: higher/lower is better. The best results of unsupervised methods are in bold. “*” denotes using the public detection of Personpath22 provided by the authors. “†” denotes using YoloX detections." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 285, + 538, + 419 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 538, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 538, + 419 + ], + "type": "table", + "html": "
Sup.HOTA↑AssA↑DetA↑MOTA↑
FAMNet [9]✓52.645.561.075.9
mmMOT [54]✓62.154.072.383.2
CenterTrack [58]✓73.071.275.688.8
LGM [47]✓73.172.374.687.6
TuSimple [8]✓71.671.172.686.3
PermaTr [46]✓78.078.478.391.3
SORT [4]×42.541.344.053.2
UNS [1]×62.565.361.1-
OC-SORT [7]×76.576.477.290.8
Ours×78.880.377.991.0
", + "image_path": "9d69c027b6844436791a6661b987a3538243ab5e49a5a24cff1fb99900f5c39d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 427, + 545, + 461 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 545, + 461 + ], + "type": "text", + "content": "Table 3. Comparison on KITTI [13] dataset (Tracking Car). " + }, + { + "bbox": [ + 305, + 427, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\uparrow/\\downarrow" + }, + { + "bbox": [ + 305, + 427, + 545, + 461 + ], + "type": "text", + "content": ": higher/lower is better. The best results of unsupervised methods are in bold. We use the same detections as OC-SORT." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " in IDF1 respectively, also reducing IDsw by large margins. The larger improvement in IDF1 demonstrates the better ability of our model in matching over long temporal distances and re-associate objects over occlusion to deliver more consistent tracking results. With even public detection, our model has achieved better performance than CenterTrack [58], TrackFormer [29] and surpasses FairMOT [55] in MOTA, despite being unsupervised. These results show that our model effectively benefits from more training data, exposing the possibility of large-scale self-supervised learning with enormous tracking videos available online." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "More importantly, in Table 4, we study model performance under challenging scenarios, including different occlusion length (measured by frames), object moving speeds and people density. Our model consistently outperform the best competitor, ByteTrack, by large margins, thanks to our PCL learning more robust object matching to occlusion and appearance changes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19021" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 262, + 129 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 262, + 129 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 262, + 129 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 262, + 129 + ], + "type": "table", + "html": "
Occlusion Length (frames)
01-2020-4040-6060+
UNS [1]49.236.832.827.026.1
OC-SORT [7]77.961.950.139.339.2
Ours79.264.350.740.739.8
", + "image_path": "026c59291304282b1b96212ab2ab5791309a19db724a736796507c0caa4c91a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 137, + 272, + 199 + ], + "blocks": [ + { + "bbox": [ + 62, + 137, + 272, + 199 + ], + "lines": [ + { + "bbox": [ + 62, + 137, + 272, + 199 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 272, + 199 + ], + "type": "table", + "html": "
Movement SpeedPeople Density
slowmediumfastlowhigh
UNS [1]42.740.338.344.742.4
OC-SORT [7]65.963.661.769.764.8
Ours68.365.162.671.466.3
", + "image_path": "7372db1fe67ff5d040cf4bcd6ef5b574766cc3f1f2a044ea6bf61bd7a354fbfb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 240, + 279, + 331 + ], + "blocks": [ + { + "bbox": [ + 47, + 207, + 286, + 231 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 286, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 286, + 231 + ], + "type": "text", + "content": "Table 4. Model Performance (IDF1) under different challenging scenarios on PersonPath22." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 240, + 279, + 331 + ], + "lines": [ + { + "bbox": [ + 53, + 240, + 279, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 279, + 331 + ], + "type": "table", + "html": "
|t-r|1-45-89-1617-3232-48
PCL (s=T)99.599.198.196.394.5
s=3299.599.298.296.393.0
s=2499.599.198.094.888.0
s=1699.699.198.091.481.7
s=899.599.195.784.671.3
s=499.598.091.878.364.1
", + "image_path": "79f3e38041805648ee77e2364933560ddfc21ad91e76895bf4e2cc0339729462.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 387, + 286, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 286, + 470 + ], + "type": "text", + "content": "Comparison on KITTI. In Table 3, we further report our performance on KITTI dataset for tracking car, following the evaluation protocol of UNS [1]. KITTI features varied moving velocity of objects and lower frame rate on average, thus objects have large appearance and location changes in adjacent frames [7]. We use the shared detections from OC-SORT [7]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "text", + "content": "Our model exceeds the best competitor, OC-SORT [7], by " + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "text", + "content": " in HOTA, " + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "text", + "content": " in AssA, " + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "text", + "content": " in DetA and " + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 603 + ], + "type": "text", + "content": " in MOTA. This is because cars have changes in their moving speeds and bounding box aspect ratios during driving (e.g., when making turn), which leads to low overlap between bounding boxes of the same object in consecutive frames and failure of pure motion trackers [4, 7]. On the other hand, our method learns object embeddings to encode object appearance thus successfully links cars in various driving scenarios, further validating our learned object matching is robust to appearance and location changes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 141, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 141, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 141, + 623 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 628, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 687 + ], + "type": "text", + "content": "We conduct ablation study to examine our key design choices. We train our unsupervised models on MOT17 test set and evaluate on MOT17 train set, since the annotations of the former set are not publicly available. Finally, we use the publicly released SDP detections." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 700 + ], + "type": "text", + "content": "Has the model learned long-distance object matching?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "We exam the accuracy of our learned matching probability," + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 313, + 71, + 538, + 132 + ], + "blocks": [ + { + "bbox": [ + 47, + 333, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 47, + 333, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 287, + 367 + ], + "type": "text", + "content": "Table 5. Accuracy of the matching probability " + }, + { + "bbox": [ + 47, + 333, + 287, + 367 + ], + "type": "inline_equation", + "content": "p(o_i^t \\rightarrow o_j^r)" + }, + { + "bbox": [ + 47, + 333, + 287, + 367 + ], + "type": "text", + "content": ". Our method has learned accurate object matching over short to long temporal distances." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 71, + 538, + 132 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 538, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 538, + 132 + ], + "type": "table", + "html": "
L=0L=10L=20L=30L=40L=50L=60
UNS64.062.558.055.053.552.050.0
OC-SORT67.766.965.865.163.963.562.5
Ours68.968.367.466.966.165.965.3
Gain+1.2+1.4+1.6+1.9+2.2+2.4+2.8
", + "image_path": "5857d04041accdb86096cfdc21193ef1cc227347d2825f51fa5ae59039370489.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 140, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 184 + ], + "type": "text", + "content": "Table 6. Comparison of IDF1 for tracking over occlusion. When increasing the minimal occlusion length (number of frames), our model has more robust performance and exceeds the best competitor OC-SORT by large margins (as shown in last row)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "inline_equation", + "content": "p(o_{i}^{t}\\to o_{j}^{r})" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "content": " is correct if " + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "inline_equation", + "content": "o_t^r" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "content": " is the correct matching of " + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "inline_equation", + "content": "o_i^t" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "content": " and has the largest matching probability. In the first two rows of Table 5, we show the accuracy and break it down by the temporal distances between objects " + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "inline_equation", + "content": "|t - r|" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "content": ". It shows our matching probability is robust to temporal distances. As " + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "inline_equation", + "content": "|t - r|" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "content": " increases, the accuracy drops only by small margins and remains " + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "inline_equation", + "content": "94.5\\%" + }, + { + "bbox": [ + 304, + 205, + 545, + 314 + ], + "type": "text", + "content": " even at the distance of 32-48 frames, validating our model has learned object matching from short to long distances." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "text", + "content": "Can the model only learn short-distance matching? Next, we further exam if learning long-distance matching is necessary and if a model only learned short-distance matching can generalize to long-distance matching? In rows 2-6 of Table 5, we limit the maximal number of continuous frames that can be skipped in PCL, denoted by " + }, + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "text", + "content": ". A smaller " + }, + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 316, + 545, + 411 + ], + "type": "text", + "content": " means the model is only trained to match objects over short temporal distances." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "text", + "content": " decreases from 32 to 4, while accuracy for short distances (1-8) remains high, it deteriorates quickly for longer distances. Comparing to the first row where we do not limit " + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "text", + "content": ", accuracy drops by " + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "inline_equation", + "content": "30.4\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "text", + "content": " on the distance of 32-48. It shows that model trained for short-distance matching cannot generalize to longer distances, which is the drawback of previous self-supervised methods [1, 24, 30]. Our PCL has successfully learned matching probability over varied temporal distances while not requiring groundtruth annotations. We also observe that " + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 568 + ], + "type": "text", + "content": " of frame skipping in paths are longer than 8 frames, which boosts our model robustness for long distance matching and the appearance changes caused by it." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": "Can the model handle long occlusion? After validating the accuracy of probability matching, we exam our model ability for handling occlusion at inference. We notice that most occlusions in MOT17 videos are of short lengths, e.g., less than 5 frames, which cannot reflect a model's ability to handle long-distance matching. Hence, we increase the occlusion length in a video by keep dropping the bounding boxes of objects after occlusion, until the occlusion lengths reach " + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": " frames. If the original length of an occlusion is longer than " + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": ", we do not increase it. Thus, " + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "inline_equation", + "content": "L = 0" + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": " means no occlusion is extended while " + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "inline_equation", + "content": "L = 60" + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": " means all occlusions last for at least 60 frames." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19022" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 183, + 148 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 183, + 148 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 183, + 148 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 183, + 148 + ], + "type": "image", + "image_path": "8fd927a3f4ab848bcc7d5ac751795898bb1bccbd4b668c497d1166031cd2340d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 150, + 142, + 158 + ], + "lines": [ + { + "bbox": [ + 117, + 150, + 142, + 158 + ], + "spans": [ + { + "bbox": [ + 117, + 150, + 142, + 158 + ], + "type": "text", + "content": "Frame 1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 72, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 184, + 72, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 184, + 72, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 184, + 72, + 294, + 148 + ], + "type": "image", + "image_path": "859c6361e0aacaa21e860d6d8f00f4bac4d3da18bf9d5c8f328dfc22e2bbbbf9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 150, + 257, + 158 + ], + "lines": [ + { + "bbox": [ + 229, + 150, + 257, + 158 + ], + "spans": [ + { + "bbox": [ + 229, + 150, + 257, + 158 + ], + "type": "text", + "content": "Frame 20" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 161, + 547, + 195 + ], + "lines": [ + { + "bbox": [ + 46, + 161, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 547, + 195 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison between our model and UNS [1]. We visualize the tracking on three frames. UNS cannot track the person in green bounding once he is occluded and assigns him a new ID (purple) on frame 80. It also fails to track the person in pink bounding box when she is only partially visible in frame 80. We can consistently track with both people with the same IDs (green, pink)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 72, + 406, + 148 + ], + "blocks": [ + { + "bbox": [ + 294, + 72, + 406, + 148 + ], + "lines": [ + { + "bbox": [ + 294, + 72, + 406, + 148 + ], + "spans": [ + { + "bbox": [ + 294, + 72, + 406, + 148 + ], + "type": "image", + "image_path": "8844fe0e04d3f73a936e2e6b4b7cb9b04a31c72b8b7b292baacef8f0ce386c42.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 150, + 372, + 158 + ], + "lines": [ + { + "bbox": [ + 326, + 150, + 372, + 158 + ], + "spans": [ + { + "bbox": [ + 326, + 150, + 372, + 158 + ], + "type": "text", + "content": "Frame 80 (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 517, + 148 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 517, + 148 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 517, + 148 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 517, + 148 + ], + "type": "image", + "image_path": "bebda331f59d9621356899e1c2591344d029dbeb2464d6635a289fabbb8bf16f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 150, + 485, + 157 + ], + "lines": [ + { + "bbox": [ + 440, + 150, + 485, + 157 + ], + "spans": [ + { + "bbox": [ + 440, + 150, + 485, + 157 + ], + "type": "text", + "content": "Frame 80 (UNS)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 199, + 286, + 302 + ], + "blocks": [ + { + "bbox": [ + 48, + 199, + 286, + 302 + ], + "lines": [ + { + "bbox": [ + 48, + 199, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 286, + 302 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\text{PC}} \\)\\( \\mathcal{L}_{\\text{OM}} \\)\\( \\mathcal{L}_{\\text{BC}} \\)IDF1HOTAMOTAIDsw
1×✓✓3.48.31.470k
2(a)✓✓26.926.837.829k
3(b)✓✓66.659.463.9329
4✓×✓62.757.163.4591
5✓✓×66.359.863.6466
6✓✓✓68.960.963.7257
", + "image_path": "556d5142706caea6fcfc0ad99e3207a0b78c4eded19cd29b2f42b50e43929886.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 310, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 365 + ], + "type": "text", + "content": "Table 7. Ablation study on proposed losses. In row 1-3, we demonstrate the effect of our path consistency loss. In row 4-5, we show the effect of our regularization losses. \"(a)\" denotes randomly sampling start and end frames for paths and \"(b)\" denotes disabling spatial constraint mask in Eq 5." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "text", + "content": "In Table 6, we report IDF1 with different occlusion length " + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "text", + "content": " and compare with UNS and one of the state-of-the-art trackers, OC-SORT, as a strong baseline. Since other related methods [16, 22, 24, 30] do not release their codes or models, we are unable to compare with them. As can be seen, our model consistently provides higher IDF1 for all ranges of occlusion lengths. More importantly, when increasing occlusion lengths, the performance of UNS and OC-SORT decreases quickly, while our performance is more robust and decreases slowly, which echos our results in Table 4. Our IDF1 at " + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "inline_equation", + "content": "L = 50" + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "text", + "content": " is still higher than the IDF1 of OC-SORT at " + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "inline_equation", + "content": "L = 20" + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "text", + "content": ", showing our model can learn to match the same objects over varied temporal distances and is more robust to occlusion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "What is the effect of path consistency loss? In Table 7 (row 1-3), we exam the results of removing PCL and also two modifications of it: (a) randomly sampling the starting and end frames for paths; (b) removing the spatial constraint mask in Eq.5; Firstly, without PCL (row 1), the model cannot learn meaningful object matching using only the regularization losses thus the tracking accuracy is close to zeros. When randomly sampling the starting and end frames (row 2), it causes most query objects are not visible in the end frames, causing imbalance between visible and disappeared objects. Thus, model matches most objects to null objects, also leading to low IDF1. Next, when removing the spatial constraint mask (row 3), the model may incorrectly match two spatial disclose objects, decreasing IDF1 by " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "content": "What is the effect of regularization losses? In Table 7 (row 4-5), we ablate the effect of the regularization losses in Section 3.3. As can be seen, while PCL contributes most to the model performance, the two regularization losses " + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{OM}}" + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BC}}" + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "content": " are also helpful for achieving better performance. Since " + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{OM}}" + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "content": " prevents multiple objects matching to the same objects and " + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BC}}" + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "content": " ensures the consistency between forward and backward matching, removing either of these loss terms leads to around " + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 201, + 547, + 334 + ], + "type": "text", + "content": " drop in IDF1, respectively. These results show the collective benefits of using different losses to optimize our unsupervised object tracker." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 341, + 414, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 414, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 414, + 353 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 358, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 547, + 491 + ], + "type": "text", + "content": "In Figure 3, we visualize the tracking results for one video from MOT17 using our model and UNS [1], where the color indicates their IDs. Both our model and UNS can track two non-occluded front people from frame 1 to frame 80. However, UNS lose the man in green bounding box after he is occluded and assign a new ID (purple) to him on frame 80. It also cannot associate the woman in the pink bounding box when she is only partially visible in frame 80. Our model correctly associate two people with the same IDs (green and pink). This qualitative comparison suggests that our model can better track the object over occlusion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "type": "text", + "content": "To tackle unsupervised MOT task, we proposed a novel concept of path consistency. We defined a path to track an object as a temporal list of observed frames from the start to the end frame. As such, to track the same object, we can generate multiple paths by skipping intermediate frames while the association results should be consistent across different paths. We leveraged this consistency in our self-supervised Path Consistency Loss to learn object matching model without groundtruth ID supervision. Our PCL especially learns robust long distance matching to tackle occlusion at inference time. With experiments on three popular datasets, our approach has exceeded existing unsupervised methods and performed on par with recent state-of-the-art supervised methods. We provided a comprehensive ablation study to validate the model has learned long-distance matching and is robust under challenging scenarios." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19023" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In Advances in Neural Information Processing Systems, pages 13695-13706. Curran Associates, Inc., 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "text", + "content": "[2] Jerome Berclaz, Francois Fleuret, Engin Turetken, and Pascal Fua. Multiple object tracking using k-shortest paths optimization. TPAMI, 33(9):1806-1819, 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 183, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 287, + 216 + ], + "type": "text", + "content": "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 218, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 218, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 287, + 262 + ], + "type": "text", + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE International Conference on Image Processing (ICIP), pages 3464-3468, 2016. 2, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 297 + ], + "type": "text", + "content": "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR 2022, 2022. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 299, + 287, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 287, + 321 + ], + "type": "text", + "content": "[6] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 323, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 288, + 378 + ], + "type": "text", + "content": "[7] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "text", + "content": "[8] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 287, + 480 + ], + "type": "text", + "content": "[9] Peng Chu and Haibin Ling. Famnet: Joint learning of feature, affinity and multi-dimensional assignment for online multiple object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6172-6181, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 483, + 287, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 287, + 527 + ], + "type": "text", + "content": "[10] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 529, + 287, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 287, + 562 + ], + "type": "text", + "content": "[11] Weitao Feng, Zhihao Hu, Wei Wu, Junjie Yan, and Wanli Ouyang. Multi-object tracking with multiple cues and switcher-aware classification, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 564, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 287, + 608 + ], + "type": "text", + "content": "[12] Ruopeng Gao and Limin Wang. MeMOTR: Long-term memory-augmented transformer for multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9901-9910, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 287, + 643 + ], + "type": "text", + "content": "[13] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 666 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[15] Roberto Henschel, Laura Leal-Taixe, Daniel Cremers, and Bodo Rosenhahn. Improvements to frank-wolfe optimization for multi-detector multi-object tracking. arXiv preprint arXiv:1705.08314, 2017. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Shyamgopal Karthik, Ameya Prabhu, and Vineet Gandhi. Simple unsupervised multi-object tracking. CoRR, abs/2006.02609, 2020. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "text", + "content": "[17] Chanho Kim, Fuxin Li, Arridhana Ciptadi, and James M Rehg. Multiple hypothesis tracking revisited. In ICCV, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[18] Chanho Kim, Fuxin Li, and James M. Rehg. Multi-object tracking with neural gating using bilinear lstm. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 240 + ], + "type": "text", + "content": "[19] Shuai Li, Yu Kong, and Hamid Rezatofighi. Learning of global objective for network flow in multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8855-8865, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "text", + "content": "[20] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 299, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 354 + ], + "type": "text", + "content": "[21] Yu-Lei Li. Unsupervised embedding and association network for multi-object tracking. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. International Joint Conferences on Artificial Intelligence Organization, 2022. 1, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 355, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 409 + ], + "type": "text", + "content": "[22] Kai Liu, Sheng Jin, Zhihang Fu, Ze Chen, Rongxin Jiang, and Jieping Ye. Uncertainty-aware unsupervised multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9996-10005, 2023. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 411, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 475 + ], + "type": "text", + "content": "[23] Qiankun Liu, Qi Chu, Bin Liu, and Nenghai Yu. Gsm: Graph similarity model for multi-object tracking. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 530–536. International Joint Conferences on Artificial Intelligence Organization, 2020. Main track. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "text", + "content": "[24] Qiankun Liu, Dongdong Chen, Qi Chu, Lu Yuan, Bin Liu, Lei Zhang, and Nenghai Yu. Online multi-object tracking with unsupervised re-identification learning and occlusion estimation. Neurocomputing, 483:333-347, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[25] Shuai Liu, Xin Li, Huchuan Lu, and You He. Multi-object tracking meets moving UAV. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8876-8885, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 580, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 634 + ], + "type": "text", + "content": "[26] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19045-19055, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "text", + "content": "[27] Fan Ma, Mike Zheng Shou, Linchao Zhu, Haoqi Fan, Yilei Xu, Yi Yang, and Zhicheng Yan. Unified transformer tracker for object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8781-8790, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[28] Christoph Mayer, Martin Danelljan, Goutam Bhat, Matthieu Paul, Danda Pani Paudel, Fisher Yu, and Luc Van Gool." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19024" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Transforming model prediction for tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8731-8740, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "text", + "content": "[29] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[30] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16264-16273, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "type": "text", + "content": "[31] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. MOT16: A benchmark for multi-object tracking. CoRR, abs/1603.00831, 2016. 1, 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "text", + "content": "[32] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 352 + ], + "type": "text", + "content": "[33] Z. Qin, S. Zhou, L. Wang, J. Duan, G. Hua, and W. Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 408 + ], + "type": "text", + "content": "[34] Hao Ren, Shoudong Han, Huilin Ding, Ziwen Zhang, Hongwei Wang, and Faquan Wang. Focus on details: Online multi-object tracking with diverse fine-grained representation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 442 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 475 + ], + "type": "text", + "content": "[36] Ergys Ristani and Carlo Tomasi. Features for multi-target multi-camera tracking and re-identification. In CVPR, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 477, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 288, + 520 + ], + "type": "text", + "content": "[37] Hao Sheng, Yang Zhang, Jiahui Chen, Zhang Xiong, and Jun Zhang. Heterogeneous association graph fusion for target association in multiple object tracking. TCSVT, 29(11): 3269-3280, 2018. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 555 + ], + "type": "text", + "content": "[38] Bing Shuai, Andrew Berneshawi, Xinyu Li, Davide Modolo, and Joseph Tighe. Siammot: Siamese multi-object tracking. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "type": "text", + "content": "[39] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joe Tighe. Large scale real-world multi person tracking. In European Conference on Computer Vision. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "type": "text", + "content": "[40] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14669-14679, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[41] Bing Shuai, Xinyu Li, Kaustav Kundu, and Joseph Tighe. Id-free person similarity learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14689-14699, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[42] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[43] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20993-21002, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[44] Siyu Tang, Mykhaylo Andriluka, Bjoern Andres, and Bernt Schiele. Multiple people tracking by lifted multicut and person re-identification. In CVPR, 2017. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[45] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "text", + "content": "[46] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10840-10849, 2021. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 265, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 331 + ], + "type": "text", + "content": "[47] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "text", + "content": "[48] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 107-122. Springer, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[49] S. You, H. Yao, B. Bao, and C. Xu. Utm: A unified multiple object tracking model with identity-aware feature enhancement. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 434, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 547, + 487 + ], + "type": "text", + "content": "[50] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8834-8843, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[51] Amir Roshan Zamir, Afshin Dehghan, and Mubarak Shah. Gmcp-tracker: Global multi-object tracking using generalized minimum clique graphs. In ECCV, 2012. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[52] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "text", + "content": "[53] Li Zhang, Yuan Li, and Ramakant Nevatia. Global data association for multi-object tracking using network flows. In CVPR, 2008. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "text", + "content": "[54] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2365-2374, 2019. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19025" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 239 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[56] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 150 + ], + "type": "text", + "content": "[57] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 174 + ], + "type": "text", + "content": "[58] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. ECCV, 2020. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 206 + ], + "type": "text", + "content": "[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 239 + ], + "type": "text", + "content": "[60] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19026" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_content_list.json b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b892517e0e15d89c1cd216c4dbb914e093872ebe --- /dev/null +++ b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_content_list.json @@ -0,0 +1,1634 @@ +[ + { + "type": "text", + "text": "Self-Supervised Representation Learning from Arbitrary Scenarios", + "text_level": 1, + "bbox": [ + 143, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaowen Li $^{1,2}$ Yousong Zhu $^{1\\boxtimes}$ Zhiyang Chen $^{1,2}$ Zongxin Gao $^{7}$ Rui Zhao $^{6}$ Chaoyang Zhao $^{1,5}$ Ming Tang $^{1}$ Jinqiao Wang $^{1,2,3,4,5\\boxtimes}$ \nFoundation Model Research Center, Institute of Automation, Chinese Academy of Science $^{1}$ School of Artificial Intelligence, University of Chinese Academy of Sciences $^{2}$ Peng Cheng Laboratory $^{3}$ Wuhan AI Research $^{4}$ Objectye Inc. $^{5}$ Qing Yuan Research Institute, Shanghai Jiao Tong University $^{6}$ Independent Researcher $^{7}$ {zhaowen.li, yousong.zhu, jqwang}@nlpr.ia.ac.cn", + "bbox": [ + 119, + 178, + 843, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current self-supervised methods can primarily be categorized into contrastive learning and masked image modeling. Extensive studies have demonstrated that combining these two approaches can achieve state-of-the-art performance. However, these methods essentially reinforce the global consistency of contrastive learning without taking into account the conflicts between these two approaches, which hinders their generalizability to arbitrary scenarios. In this paper, we theoretically prove that MAE serves as a patch-level contrastive learning, where each patch within an image is considered as a distinct category. This presents a significant conflict with global-level contrastive learning, which treats all patches in an image as an identical category. To address this conflict, this work abandons the nongenerizable global-level constraints and proposes explicit patch-level contrastive learning as a solution. Specifically, this work employs the encoder of MAE to generate dual-branch features, which then perform patch-level learning through a decoder. In contrast to global-level data augmentation in contrastive learning, our approach leverages patch-level feature augmentation to mitigate interference from global-level learning. Consequently, our approach can learn heterogeneous representations from a single image while avoiding the conflicts encountered by previous methods. Massive experiments affirm the potential of our method for learning from arbitrary scenarios.", + "bbox": [ + 75, + 369, + 472, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 791, + 207, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, self-supervised learning (SSL) in computer vision [1, 2, 4-7, 18, 21, 22, 35, 59] is capable of adopting self-defined pseudo labels as supervision [28, 29] and holds the promise in leveraging large amounts of unlabeled data", + "bbox": [ + 75, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bafcc348137408f74f60e64216fc5772dc8bdf0e5756deb50592b67a4ef031df.jpg", + "image_caption": [ + "Figure 1. Hybrid methods further enhance the global consistency of CL. We measures representational similarity among tokens for some popular self-supervised pre-trained models. Among them, ccMIM [67] and SIM [54] employ MAE as the MIM baseline method while iBOT [69] adopts DINO [3] as the CL baseline method for hybrid approach. It indicates that hybrid methods, while increasing linear evaluation scores, also enhance the representational similarity among tokens. This suggests that hybrid methods fundamentally further strengthen the global semantic consistency of CL. All of the models adopt ViT-B." + ], + "image_footnote": [], + "bbox": [ + 516, + 362, + 857, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to build the foundational models. SSL methods focus on designing different pretext tasks, and there are two mainstream approaches for learning visual features in the community. One of the most promising directions among them is contrastive learning (CL) [2, 5-7, 18, 21, 25, 48]. It assumes that different views of a single image are of the same instance. Its objective is to learn a global-level feature representation that discriminates among images. CL leverages trainable networks to generate semantic pseudo labels [55], thereby enhancing the model's capability to extract semantic information. On the other hand, the masked image mod", + "bbox": [ + 496, + 733, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author: Yousong Zhu and Jinqiao Wang.", + "bbox": [ + 99, + 886, + 397, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22967", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/415dcbcc76660511c360d1e83dafd5030e5f27229ed021eaa846308f6f969e66.jpg", + "image_caption": [ + "Figure 2. Visualization of patches in an image from COCO dataset. The image is partitioned into multiple patches, and different patches from the image may represent different semantics." + ], + "image_footnote": [], + "bbox": [ + 88, + 93, + 287, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8b43cbeb1fc5d970d9510bcafa8b43228a83d08e40799b287b80fd8b8d49e885.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 97, + 464, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eling (MIM) [1, 4, 15, 22, 35, 63, 69] has become the focus in the community. MIM is a pretext task that involves masking some patches on the input image and predicting the original information of the masked patches based on their context. In these methods, BEiT [1] predicts discrete tokens created by the pre-trained model of VQVAE [56]. In MAE [22], the pseudo labels are the given pixels. Among these MIM methods, MAE shows impressive transfer performance on various dense prediction tasks.", + "bbox": [ + 75, + 323, + 467, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The learning methodologies of CL and MIM are so distinct that a fundamental question arises: can their concurrent learning mutually enhance each other? In fact, extensive work [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] shows hybrid methods of CL and MIM achieve state-of-the-art (SOTA) performance on benchmark ImageNet [13] dataset. In our study, we choose some pre-trained models trained using popular self-supervised methods and measure their representational similarity for tokens as well as linear probing results, as depicted in Figure 1. Specifically, ccMIM [67] and SIM [54] employ MAE as the MIM baseline, while iBOT [69] adopts DINO [3] as the CL baseline for the hybrid approach. The figure illustrates that these hybrid methods have achieved better linear probing results compared to their respective baseline methods. However, they have further reinforced representational similarity for tokens, resulting in homogeneous token representations. This also implies that the hybrid methods essentially enhance the global consistency of CL, and the performance of SOTA methods heavily relies on the gains achieved through CL. Consequently, the success of these methods heavily relies on the single-centric-object data [37, 41] such as those in ImageNet dataset, and they cannot be generalized to non-iconic scenarios, making them unsuitable for training in arbitrary scenarios - regardless of whether they come from single-centric-object scenarios or non-iconic ones. Also, a recent study [34] indicates that the hybrid approach of MAE and CL exhibits inferior performance compared to the baseline in large-scale arbitrary scenarios. This goes against the pri", + "bbox": [ + 75, + 463, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mary promise of self-supervised learning, which aimed to build foundation models using a large amount of data for broader applicability.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Actually, the inability of SOTA methods to train in arbitrary scenarios can be attributed to the conflict between CL and MIM. In this paper, we first theoretically prove that MAE is a patch-level CL, treating each patch within an image as a separate category. This conflicts with global-level CL, where all patches from the image are treated as an identical category. As illustrated in Figure 2, a non-iconic image from the COCO [40] dataset is partitioned into multiple non-overlapping patches according to ViT [15], and each patch may represent a different semantics. MAE treats each object/patch in these patches as a distinct category, whereas CL considers all patches as an identical category. Therefore, when these two methods are combined for training on non-iconic data, significant conflicts and confusion arise, rendering the hybrid approach incapable of learning effective representations for non-iconic data and arbitrary scenarios.", + "bbox": [ + 496, + 138, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the problems, in this paper, we propose a self-supervised framework, called Arbitrary Self-supervised Learning (ASL), to learn representation from arbitrary scenarios. The framework leverages the principles of CL to enhance the patch-level semantic consistency of the model while avoiding the conflict between CL and MAE. Specifically, our approach involves designing a patch-level feature augmentation to effectively circumvent the limitation of global-level data augmentation in CL. According to the patch-level augmentation, ASL employs the encoder of MAE to generate dual-branch features, which then perform patch-level contrastive learning through a decoder module. We also retain the original MIM pretext task to ensure the rationality of the model. It is noteworthy that our approach is proposed for arbitrary scenarios, primarily experimenting on the non-iconic COCO dataset that closely represents natural scenes. Simultaneously, we conduct the pre-training experiments on both ImageNet and COCO. The results indicate that ASL is the sole method of attaining superior performance in arbitrary pre-training, further underscoring the generalizability and robustness of our approach.", + "bbox": [ + 496, + 381, + 892, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, we make the following contributions:", + "text_level": 1, + "bbox": [ + 517, + 700, + 823, + 714 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We theoretically prove that MAE is a patch-level contrastive learning, which is in conflict with global contrastive learning, thereby rendering current hybrid methods incapable of achieving generality in arbitrary scenarios pre-training.", + "- We propose a self-supervised framework, named Arbitrary Self-supervised Learning (ASL), to mitigate conflicts and allow the model to train in arbitrary scenarios.", + "- Extensive experiments demonstrate the effectiveness and transfer ability of our framework. Specifically, the models pre-trained with ASL achieve SOTA performance in arbitrary scenarios." + ], + "bbox": [ + 500, + 715, + 890, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22968", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is known that the promise of self-supervised learning [1, 2, 4-7, 15, 18, 21, 22, 25, 30, 35, 37, 47, 48, 50, 51, 59, 63, 66, 69] in computer vision is to establish a visual foundation model by leveraging large dataset, just like ChatGPT [49] in natural language processing. These representations of the foundation model can be transferred to various downstream tasks [10, 12, 13, 19, 20, 33, 36, 38, 39, 53, 60].", + "bbox": [ + 75, + 114, + 468, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Instance discrimination", + "text_level": 1, + "bbox": [ + 76, + 227, + 294, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The basic principle of contrastive learning/instance discrimination is that different views of an image are still the same category. As a significant representative of these approaches, MoCo [21] improves the training of contrastive learning methods by storing representations from a momentum encoder instead of the trained network. Then, SimCLR [5] shows that the memory bank can be entirely replaced with the elements from the same batch when the batch size is set large enough. Furthermore, BYOL [18] proposes an asymmetric structure and directly bootstraps the representations by attracting different features from the same instance and shows that contrastive learning without negative samples can also learn excellent visual representations. Also, the asymmetric structure is often directly adopted by subsequent work. Some approaches [23, 24, 58, 61] attempt to transfer global-level prior to pixel-level or region-level learning and acquire visual representations for dense tasks. Also, UniVIP [37] proposes a three-level pre-training task and learns versatile representations from non-iconic images. Recently, MoCo v3 [8] and DINO [3] replace CNNs with ViT [15] and achieve superior performance.", + "bbox": [ + 75, + 250, + 472, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Masked image modeling", + "text_level": 1, + "bbox": [ + 76, + 575, + 302, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by MLM of BERT [14] in NLP, the pioneering work iGPT [4] operates on sequences of pixels and predicts unknown pixels. Benefiting from the proposal of Vision Transformer [11, 15, 42], MIM achieves performance comparable to CL. MST [35] is the first to introduce MIM into the siamese structure and proposes the attention-guided mask strategy. Then, iBOT [69] also adopts the siamese structure and obtains impressive performance. However, they have the prior/assumption of image semantic consistency. Besides, BEiT [1] proposes to predict discrete tokens for masked image modeling, yet its transfer performance mainly depends on the quality of the pre-trained model VQVAE [56]. Following the MIM design of ViT, SimMIM [63] predicts pixels of masked patches to perform the MIM task. Different from the above methods of performing MIM task in the encoder, MAE [22] designs a special encoder-decoder structure to make the decoder perform the MIM task and abandon the MIM-performing module in downstream visual tasks. Among these methods, MAE shows excellent performance in various downstream tasks.", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Hybrid methods", + "text_level": 1, + "bbox": [ + 500, + 90, + 663, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An intuitive idea is to combine the above approaches [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] to achieve better performance. MST [35] firstly adopts MIM to a contrastive learning framework [3] and restores the pixels of patches, then iBOT [69] predicts the semantic tokens. Moreover, DINO v2 [50] achieves SOTA performance based on the iBOT through several engineering optimizations, with the most significant being data curation. It reduces the original dataset of 1.3 billion images to 142 million images, primarily using the ImageNet dataset as the main set of query images. However, it is not suitable for training in arbitrary scenarios. More methods [27, 32, 44, 52, 54, 64, 65, 67] are based on the MAE framework rather than the MST/iBOT architecture, as it conserves approximately 5 times the computational resources when compared to the MST/iBOT framework under the ViT-B setting. Additionally, as the model scales up, the proportion of saved computational resources increases further. However, current SOTA methods primarily benefit from contrastive learning, thus demonstrating SOTA performance on the benchmark single-centric-object ImageNet dataset. When extended to arbitrary or natural scenes, a pronounced conflict between CL and MAE arises.", + "bbox": [ + 496, + 113, + 893, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 472, + 609, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 497, + 629, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Masked autoencoders", + "text_level": 1, + "bbox": [ + 500, + 521, + 692, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As noted in the prior work [22], for a dataset $X$ without manual annotations, an image $x$ can be divided into $n$ regular non-overlapping patches. MAE first samples a subset of patches and masks the remaining ones, and then acquires the masked patches $x_{m_i}$ , where $i = 1,\\dots,\\eta$ and $\\eta$ is the set of possible masks of length, and visible patches $x_{v_j}$ , where $j = 1,\\dots,\\kappa$ and $\\kappa = n - \\eta$ . The mask process of these patches follows a uniform distribution, named the random mask strategy. Moreover, MAE feds these visible patches into encoder (ViT) $f(\\cdot)$ , parameterized by $\\theta$ , and obtains the encoded visible tokens, then puts these encoded visible tokens with mask tokens into decoder $g(\\cdot)$ to perform the MIM task, parameterized by $\\xi$ , and obtains the predicted patches. Finally, MAE minimizes the mean squared error (MSE) between the reconstructed and original images in the pixel space. The loss of a single patch prediction is shown as Eq(1).", + "bbox": [ + 496, + 536, + 893, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i n g l e}} (i) = \\left\\| x _ {m _ {i}} - g \\left(\\xi ; f \\left(\\theta ; x _ {v _ {1}}, \\dots , x _ {v _ {\\kappa}}, \\mathbf {m a s k}\\right)\\right) \\right\\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 800, + 890, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For MAE, the total loss function that trains the complete dataset is defined as Eq (2).", + "bbox": [ + 498, + 840, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M A E} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\underset {i \\sim \\eta} {\\mathbb {E}} \\mathcal {L} _ {\\text {s i n g l e}} (i), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 878, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "22969", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Contrastive learning", + "text_level": 1, + "bbox": [ + 76, + 90, + 263, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, for an image $x$ in the dataset $X$ , contrastive learning first generates two views under random data augmentation [5, 18], which are then fed into the network separately. Commonly, the network consists of the encoder and momentum encoder. Then, the two global features $z_{s}$ and $z_{t}$ can be obtained. Let $\\mathrm{sim}(z_s,z_t) = z_s^\\top z_t / \\| z_s\\| \\| z_t\\|$ denotes the dot product between $\\ell_2$ normalized $z_{s}$ and $z_{t}$ (i.e. cosine similarity). The InfoNCE loss function [5, 21, 48, 59] for a positive pair of examples is defined as Eq (3), where $\\tau$ is a temperature coefficient. It pulls one prediction closer to the self-defined pseudo label while pushing other predictions in a mini-batch apart.", + "bbox": [ + 75, + 107, + 468, + 289 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {I n f o N C E}} (x) = - \\log \\frac {\\exp \\left(\\operatorname* {s i m} \\left(\\boldsymbol {z} _ {s} , \\boldsymbol {z} _ {t}\\right) / \\tau\\right)}{\\sum_ {k \\neq s} ^ {2 N} \\exp \\left(\\operatorname* {s i m} \\left(\\boldsymbol {z} _ {s} , \\boldsymbol {z} _ {k}\\right) / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 301, + 468, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, the total loss function that trains the complete dataset is defined as Eq(4).", + "bbox": [ + 76, + 358, + 468, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a s}} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\mathcal {L} _ {\\text {I n f o N C E}} (x), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 402, + 468, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is noted that the advantage of CL is that semantic learning and its self-defined pseudo label are acquired by the updatable and learnable network.", + "bbox": [ + 76, + 439, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. The analysis of MAE", + "text_level": 1, + "bbox": [ + 76, + 497, + 274, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to [46], it is well known that minimizing MSE can be equivalent to maximum likelihood estimation. Therefore, a single patch prediction of MAE can be considered as the mean of a noisy prediction distribution, and the distribution can be modeled as a Gaussian distribution in the classic probabilistic interpretation [43] in Eq (5).", + "bbox": [ + 75, + 521, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(x _ {m _ {i}} \\mid x _ {\\text {i n p u t s}}; \\theta , \\xi\\right) = \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 626, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$x_{input}$ is the input tokens and equal to $\\{x_{v_1},\\dots,x_{v_\\kappa},\\mathbf{mask}\\}$ , $x_{p_i}$ is the single predicted patch, $x_{m_i}$ is the pseudo label corresponding to this patch, and $\\sigma_{noise}$ is the scale of an independent and identically distributed error term $\\epsilon \\sim \\mathcal{N}(0,\\sigma_{noise}^2\\mathbf{I})$ . Meanwhile, MSE equals to the negative log likelihood (NLL) loss of the prediction distribution $p(x_{m_i}|x_{input};\\theta ,\\xi)$ [46]. The network trained by MSE predicts a single patch in fact learns to model $p(x_{m_i}|x_{input})$", + "bbox": [ + 76, + 657, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MAE shows impressive performance on various downstream tasks [22], it reconstructs the masked patches through learning from a large amount of data, and the pseudo labels are the masked image and have no interference from manual annotations. Hence, it is reasonable to consider the $p(x_{m_i})$ is uniform, and the MSE is equivalent to Eq(6), where $C$ is a constant and equal to $p(x_{m_i})$ , and", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the detailed derivation of this section in the Appendix.", + "bbox": [ + 500, + 90, + 856, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i n g l e}} (i) \\cong - \\log \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right) \\\\ + \\log \\int \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right) \\cdot p \\left(x _ {m _ {i}}\\right) d x _ {m _ {i}}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 113, + 888, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the second part of Eq(6), we utilize monte carlo method to solve $p(x_{m_i})$ and can obtain the Eq(7).", + "bbox": [ + 500, + 188, + 890, + 220 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\int \\mathcal {N} (x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {n o i s e} ^ {2} \\mathbf {I}) \\cdot p (x _ {m _ {i}}) d x _ {m _ {i}} \\\\ \\approx \\frac {1}{N} \\sum_ {b = 1} ^ {N} \\mathcal {N} \\left(x _ {m _ {i _ {(b)}}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I})\\right), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 228, + 888, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The monte carlo method treats all pseudo labels in a training batch as random samples from $p(x_{m_i})$ . Hence, for pseudo labels in a training batch $B = \\{x_{m_{i(1)}}, x_{m_{i(2)}}, \\ldots x_{m_{i(N)}}\\}$ , the loss is defined as Eq(8), where $\\lambda = 2\\sigma_{noise}^2$ is a temperature coefficient.", + "bbox": [ + 500, + 310, + 890, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i n g l e}} (i) \\cong - \\log \\frac {\\exp \\left(- \\left| \\left| x _ {p _ {i}} - x _ {m _ {i}} \\right| \\right| ^ {2} / \\lambda\\right)}{\\sum_ {x _ {m _ {i}} ^ {\\prime} \\in B} \\exp \\left(- \\left| \\left| x _ {p _ {i}} - x _ {m _ {i}} ^ {\\prime} \\right| \\right| ^ {2} / \\lambda\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 398, + 888, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Eq (3) and Eq (8), it can be observed that the form of the two loss functions is similar. Both of the two loss functions pull together the prediction and self-defined pseudo label in the representation space, and push other predictions in a mini-batch apart. The MAE adopts the Euclidean distance while the contrastive learning measures the cosine similarity in representation space. Actually, MAE regards each masked patch as a category and performs patch-level contrastive learning with unlearnable pixels. It explicitly aligns the prediction with its corresponding masked patch and implicitly distinguishes other patches.", + "bbox": [ + 496, + 460, + 890, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the above description, when confronted with arbitrary scenes, MAE and CL will face a severe conflict issue.", + "bbox": [ + 496, + 627, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Arbitrary self-supervised learning", + "text_level": 1, + "bbox": [ + 500, + 665, + 797, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the previous analysis, it can be concluded that a global prior of CL is both unreasonable and conflicts with MIM when facing arbitrary scenarios. To enable learning from arbitrary scenarios, can there be a self-supervised method that harnesses the advantages of CL in semantic learning (learnable self-pseudo labels) and the heterogeneous representation learning of MIM, while avoiding global constraints and conflicts? The answer is affirmative. In this paper, we propose a new framework, named Arbitrary Self-supervised Learning (ASL), to learn from arbitrary scenarios while retaining their advantages.", + "bbox": [ + 496, + 688, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we adopt the popular MAE method as the basic MIM to ensure the initial heterogeneous representations. Then, inspired by CL using a dual-branch structure", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22970", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4e811026b0a1b77f5ec139adc04c001cc27115c67064b85d413a5a1cea74bcf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 103, + 867, + 308 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2669a10c2431ffdfffe1b26f79450cf7d443afdbd44e71c059049ff749faac4e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 117, + 311, + 416, + 416 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/402b4a158a45160a2e1a7ff8820274cb19785fe3f003e35bfc4b01967b8e3cca.jpg", + "image_caption": [ + "(b)", + "Figure 3. The pipeline of our ASL. The visible patches are fed into the network and the dual-branch features can be obtained by our designs. Then the encoded features with mask tokens are fed into the decoder and output the predictions. Finally, we pull together the predictions using L1 loss by one-to-one correspondence. We propose two designs in the network to obtain the dual-branch features: (a) Dropout in the linear embedding. (b) Asymmetric structure at the end of the encoder. In the figure, we adopt the network to represent (a) or (b) design. The olive patch represents the mask token." + ], + "image_footnote": [], + "bbox": [ + 436, + 318, + 854, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to generate learnable self-pseudo labels, we propose two patch-level designs to generate dual-branch features based on the encoder of MAE while avoiding the global constraints. Subsequently, these dual-branch features are further processed through a decoder module, and CL is performed between the outputs patch by patch to for semantic learning. In contrast to traditional CL, our approach does not entail global data augmentation but instead focuses on patch-level feature augmentation. The specifics of our designs are outlined as follows:", + "bbox": [ + 75, + 510, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dropout in the linear embedding (DLE). According to previous work in NLP, SimCSE [17] has explored the use of dropout in word embedding for feature augmentation. In line with this, we also adopt dropout in the linear embedding of the encoder to perform patch-level augmentation and acquire two encoded features at the end of the encoder. The details of our method are described in Figure 3 (a).", + "bbox": [ + 75, + 667, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Asymmetric structure at the end of the encoder (AEE). Some prior work [8, 18, 62] have demonstrated that introducing an asymmetric structure, which essentially is a nonlinear function, enhances the semantic features. With this in mind, we incorporate a Transformer layer [57] after the encoder in our approach to build an asymmetric structure for patch-level feature augmentation. This step produces dual-branch features: one is the feature output by the", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encoder, and the other is the feature that has passed through the Transformer layer. Figure 3 (b) illustrates the details of our method. The design is our default experimental setup.", + "bbox": [ + 496, + 510, + 890, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The methodology described above pertains to the construction of feature augmentation. Our rationale for utilizing feature augmentation to construct dual-branch features is primarily based on two considerations: 1) Data augmentation in CL is global and not suitable for patch-level learning. 2) The asymmetric structure of CL inherently constitutes an implicit form of feature augmentation, and it is worth exploring whether its success in CL can be extrapolated to the broader domain of self-supervised learning.", + "bbox": [ + 496, + 555, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After feature augmentation, we obtain the dual-branch features, $fea_{ori}$ and $fea_{aug}$ , then we feed them into the decoder module to obtain the corresponding predictions, $pred_{ori}$ and $pred_{aug}$ , respectively. The patch-level CL between the outputs does not employ contrastive loss or cosine similarity loss but rather L1 loss.", + "bbox": [ + 496, + 691, + 890, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The alignment between any pair is defined as Eq(9), where $\\mathsf{sg}[\\cdot]$ stands for stop gradient. It is noted that our alignment method does not include the alignment of [cls] token according to Section 3.2, which unleashes the potential of the method for training in arbitrary scenarios.", + "bbox": [ + 496, + 782, + 890, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} (i) = \\left| \\left| \\operatorname {p r e d} _ {\\text {o r i} _ {i}} - s g \\left[ \\operatorname {p r e d} _ {\\text {a u g} _ {i}} \\right] \\right| \\right| + \\left| \\left| s g \\left[ \\operatorname {p r e d} _ {\\text {o r i} _ {i}} \\right] - \\operatorname {p r e d} _ {\\text {a u g} _ {i}} \\right| \\right|, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22971", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function of the complete dataset is defined as Eq(10).", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S E M} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\underset {i \\sim \\eta} {\\mathbb {E}} \\mathcal {L} _ {s} (i), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 121, + 468, + 143 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Objective function", + "text_level": 1, + "bbox": [ + 76, + 152, + 256, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The loss function of our ASL consists of the MAE loss and the semantic loss. Hence, the total loss of our ASL is formulated as Eq (11), and each loss coefficient is set to be 1 for equally weighted.", + "bbox": [ + 76, + 176, + 468, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {M A E}} + \\mathcal {L} _ {\\text {S E M}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 250, + 468, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 277, + 209, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and evaluation metrics", + "text_level": 1, + "bbox": [ + 76, + 303, + 356, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training datasets. In the pre-training stage, we select the four popular visual pre-training datasets, MS COCO [40], ImageNet [13], CIFAR10 [31], CIFAR100 [31], to investigate the versatility and transfer ability of the proposed method. For the COCO dataset, its train2017 set contains $\\sim 118\\mathrm{k}$ images. COCO contains more natural and diverse scenes and is a non-iconic image dataset. To further demonstrate the versatility of our algorithm for training in arbitrary scenarios, we conduct pre-training on both ImageNet and COCO datasets. Given the substantial difference in data volume between ImageNet and COCO, we select a subset of ImageNet, referred to as ImageNet-100, to maintain a balanced dataset size. Moreover, we pre-train on ImageNet-1K to demonstrate the effectiveness of ASL on curated data. Finally, CIFAR-10 and CIFAR-100 datasets are small-scale image datasets, and they all contain 60000 images with $32\\times 32$ size that belong to 10 and 100 categories, respectively. We pre-train models on the CIFAR-10 and CIFAR-100 datasets to demonstrate the effectiveness of ASL in the small-scale dataset.", + "bbox": [ + 75, + 325, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation datasets. In this paper, ASL is evaluated by the linear probing and fine-tuning classification task on ImageNet, CIFAR-10, and CIFAR-100 datasets, which are popular image classification datasets. Additionally, for computer vision, object detection, and semantic segmentation are dense prediction tasks since training images of these tasks contain massive objects inside each image. The validation of dense prediction tasks can better reflect the transfer ability of self-supervised models. Therefore, we conduct extensive experiments on COCO [40] and ADE20k [68] datasets to verify the generalization and transfer ability of ASL. COCO is a popular object detection and instance segmentation dataset, which train2017 contains about $118k$ images and test on the val2017 contains $5k$ images. The challenging dataset contains human annotations for 80 classes. Moreover, ADE20K is also a challenging semantic segmentation dataset, and it contains about $25k$ training images and $2k$ validation images with 150 categories.", + "bbox": [ + 75, + 628, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f18ade2db5087a21781a1bcd9f7a9963bdeca48293dda8f9f6e980f969aa04e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArch.FLOPsEpochsLPAPb
Pre-training on COCO dataset:
MAE [22]ViT-B1×80046.8%49.3%
CMAE [27]ViT-B~ 3×80046.6%49.0%
ccMIM [67]ViT-B~ 3.3×80046.7%49.0%
SIM [54]ViT-B~ 2.8×80046.0%48.7%
iBOT [69]ViT-B~ 6×80045.2%48.6%
ASLViT-B~ 1.5×80048.6%50.3%
", + "bbox": [ + 504, + 89, + 898, + 219 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bdfdd4599018376d37a7cf1100ca819223fd395b81d8506fc8ed0090e9ce79c2.jpg", + "table_caption": [ + "Table 1. Evaluation of pre-trained models on the COCO dataset. We report Linear Probing (LP) top-1 accuracy on the ImageNet-1K val set and bbox mAP on COCO dataset." + ], + "table_footnote": [], + "table_body": "
MethodArch.Param.EpochsLPFT
Pre-training on ImageNet-100 + COCO:
MAE [22]ViT-B8680074.5%90.9%
iBOT[69]ViT-B8680071.8%88.5%
ccMIM [67]ViT-B8680072.3%88.7%
ASLViT-B8680079.6%92.4%
ASLViT-B86400085.9%94.2%
", + "bbox": [ + 503, + 266, + 893, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. ImageNet-100 Top-1 accuracy of different methods under linear probing (LP) and fine-tuning (FT) setting to evaluate the performance of the model pre-trained on both ImageNet-100 and COCO. We report top-1 accuracy on the ImageNet-100 val set.", + "bbox": [ + 498, + 385, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The valuation metrics, pre-training settings, and the experimental settings of downstream tasks in the Appendix.", + "bbox": [ + 498, + 460, + 890, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Pre-training on COCO dataset", + "text_level": 1, + "bbox": [ + 500, + 500, + 772, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to validate the adaptability of our approach for training in arbitrary scenarios, we initially conduct pretraining on the COCO dataset and subsequently conduct a fair comparison with other methods under the same experimental settings. In this comparison, we evaluate the ImageNet linear evaluation and the COCO detection results. Linear evaluation, which involves using frozen features, provides a more comprehensive assessment of a model's representation. Given that the pre-training experiments are conducted on the COCO dataset, using human-labeled data from COCO to test the learned representations is reasonable. As shown in Table 1, we observe that methods combining MAE with contrastive learning, such as CMAE, ccMIM, and SIM, exhibit weaker performance when pretrained on the COCO dataset, even though they achieved SOTA performance on ImageNet. This observation suggests that existing SOTA methods are not well-suited for training in arbitrary scenarios. Furthermore, our approach demonstrates superior performance after pre-training on the COCO dataset, surpassing the performance of MAE. Generally, introducing a dual-branch structure similar to contrastive learning typically results in a significant increase in computational overhead, often exceeding threefold. Diverging from the emphasis on the encoder in hybrid methods, our AEE design prioritizes the decoder, significantly reducing", + "bbox": [ + 496, + 523, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22972", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ebde70f32203de049d49384a3e3fe90999a63ef2810fba7db7a3172e7b131f26.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-train IterationsDatasetObject detectionInstance segmentation
\\(AP^b\\)\\(AP_{50}^b\\)\\(AP_{75}^b\\)\\(AP^m\\)\\(AP_{50}^m\\)\\(AP_{75}^m\\)
Supervised [9]~ 94kImageNet-1K47.9%--42.9%--
MoCo v3 [8]~ 187kImageNet-1K47.9%--42.7%--
BEiT [1]~ 249kImageNet-1K + DALLE49.8%--44.4%--
MAE [22]~ 499kImageNet-1K50.4%70.8%55.7%44.9%68.3%48.9%
CAE [9]~ 499kImageNet-1K50.0%70.9%54.8%44.0%67.9%47.6%
R-MAE [45]~ 115kCOCO50.6%--45.0%--
R-MAE [45]~ 250kCOCO50.8%--45.2%--
ccMIM [67]~ 249kImageNet-1K50.3%71.2%55.0%44.5%68.4%47.9%
ASL~ 47kImageNet-100 + COCO50.7%71.1%55.6%45.0%68.3%49.0%
ASL~ 236kImageNet-100 + COCO51.5%71.9%57.2%45.7%69.2%50.0%
", + "bbox": [ + 83, + 88, + 885, + 295 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24227bee0d7f3652b311959aa42e78f860b84c0c18397c39a2253ba55eb00e35.jpg", + "table_caption": [ + "Table 3. Results of object detection and instance segmentation on COCO. The architecture of various methods is ViT-B, and we report the bounding box AP and mask AP on COCO val2017." + ], + "table_footnote": [], + "table_body": "
MethodDatasetPre-train IterationsADE mIoU
Supervised [22]ImageNet-1K~ 93k47.4%
SplitMask [16]ADE20K~ 128k45.7%
MoCo v3 [8]ImageNet-1K~ 187k47.3%
BEiT [22]ImageNet-1K+DALLE~ 249k47.1%
R-MAE [45]COCO~ 115k46.8%
R-MAE [45]COCO~ 250k47.0%
ccMIM [67]ImageNet-1K~ 249k47.7%
MAE [22]ImageNet-1K~ 499k48.1%
ASLImageNet-100 + COCO~ 236k48.4%
", + "bbox": [ + 81, + 327, + 475, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results of semantic segmentation on ADE20K. We report results measured by mIoU.", + "bbox": [ + 76, + 508, + 470, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing computational costs in comparison. Meanwhile, due to the fixed-size and lightweight design of the decoder, our design incurs decreasing computational overhead as the model scale increases. This highlights the potential of our method for pre-training in non-iconic scenarios.", + "bbox": [ + 75, + 544, + 470, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Pre-training on both COCO and ImageNet", + "text_level": 1, + "bbox": [ + 76, + 631, + 442, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further demonstrate the adaptability of ASL across diverse scenarios, we conduct the pre-training experiments on both ImageNet-100 and COCO datasets. We assess the performance of the model on benchmark tasks: image classification, object detection, and semantic segmentation. The ASL models are primarily trained with 800 epochs and 4000 epochs, with the number of training iterations for the 4000-epoch training similar to that of the ImageNet-1K iterations for the 800-epoch training.", + "bbox": [ + 75, + 655, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Image classification", + "text_level": 1, + "bbox": [ + 76, + 815, + 267, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our training data comprises the ImageNet-100 dataset, we conducted testing for classification performance through both linear probing and fine-tuning on ImageNet-100. The experimental results are presented in Table 2. The", + "bbox": [ + 76, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results of pre-training observed from the table align with those of pre-training solely on COCO: ccMIM exhibits relatively weak performance in the context of the pre-training. This can be attributed to conflicts between CL and MAE when dealing with arbitrary scenarios. In contrast, our method successfully mitigates such conflicts and produces more robust representations in arbitrary scenarios.", + "bbox": [ + 496, + 330, + 893, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Object detection and instance segmentation", + "text_level": 1, + "bbox": [ + 498, + 455, + 857, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the transfer ability of ASL, we test it on object detection and instance segmentation with COCO [40] dataset. Here we report Box AP and mask AP on the validation set. According to the benchmark validation [22, 33], we choose the Mask R-CNN [20] as the test framework. Due to variations in the sizes of the pre-training datasets used, we opt to avoid epoch-based comparisons in this scenario. We convert epochs to a fixed batch size of 4096 iterations for a fair comparison. Table 3 shows the results of the learned representation by supervised method and different self-supervised methods. It shows our ASL, with approximately 236k iterations (4000 epochs), achieves impressive performance and outperforms 1600-epoch MAE, 1600-epoch CAE and 800-epoch ccMIM pre-trained on ImageNet. ASL surpasses various supervised and self-supervised models pre-trained on ImageNet and COCO, indicating its success in arbitrary scenarios.", + "bbox": [ + 496, + 479, + 890, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3 Semantic segmentation", + "text_level": 1, + "bbox": [ + 498, + 756, + 714, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate the generalization of our method, we conduct the semantic segmentation experiments on ADE20K dataset [68]. As shown in Table 4, ASL achieves the $48.4\\%$ mIoU. It also outperforms various supervised and self-supervised models pre-trained on ImageNet.", + "bbox": [ + 496, + 779, + 890, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, our ASL shows impressive performance on various mainstream visual benchmark tasks when pre-training on both COCO and ImageNet-100, which demonstrates", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22973", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a44ec4862b99658d27accfaa4740e8ef98f34f475c6b1bab8ad57628d030cbc9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLPFTAPbAPmmIoU
MAE67.8%83.6%50.4%44.9%48.1%
ASL69.2%84.2%51.3%45.7%49.2%
", + "bbox": [ + 84, + 89, + 459, + 155 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a9fd7a7480ef63b82de45017db9dcf6780c1776130086e9d3410048b7fdbbac.jpg", + "table_caption": [ + "Table 5. The ImageNet-1K pre-training results. We report ImageNet classification, COCO detection, and instance segmentation, as well as ADE20k semantic segmentation results." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10CIFAR-100
Linear Prob.FinetuningLinear Prob.Finetuning
Supervised [26]-91.3%-64.1%
DINO [3]89.0%94.4%65.8%76.3%
MAE [22]87.3%95.9%54.0%81.1%
ASL91.4%96.5%60.3%82.1%
", + "bbox": [ + 81, + 203, + 467, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8fcefa8aa83418238c6a51abe6594d822783edb401177fd989ce646c1f645515.jpg", + "table_caption": [ + "Table 6. Linear probing and fine-tuning classification results on small-scale datasets. We adopt ViT-S as the base model." + ], + "table_footnote": [], + "table_body": "
IDDLEAEEalignmentFLOPsTop-1APb
(a)×××1×46.8%49.3%
(b)✓××~ 2×46.5%49.1%
(c)✓×✓~ 2×48.5%50.4%
(d)×✓×~ 1.5×46.6%49.2%
(e)×✓✓~ 1.5×48.6%50.3%
", + "bbox": [ + 81, + 309, + 467, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ASL has the ability to learn general visual representations from arbitrary scenarios.", + "bbox": [ + 76, + 452, + 467, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Pre-training on ImageNet-1K", + "text_level": 1, + "bbox": [ + 76, + 492, + 338, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method is designed for learning visual representations from uncurated arbitrary scenarios, and it can also yield stronger representations when applied to meticulously curated datasets. To demonstrate this point, we conduct experiments on the curated ImageNet dataset. As depicted in the experimental results shown in Table 5, ASL exhibits superior performance across various downstream tasks compared to MAE, highlighting the versatility of our approach.", + "bbox": [ + 75, + 516, + 467, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Small-scale data pre-training", + "text_level": 1, + "bbox": [ + 76, + 647, + 336, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 6, the results of our method on these small datasets all show impressive performance. With the same training settings for all experiments, ASL consistently improves MAE by at least $4.1\\%$ for linear probing. However, MIM still lags behind CL in linear evaluation on CIFAR100 and its advantage lies in the fine-tuning superior performance. Our fine-tuning results also reveal it. Totally, these results suggest that ASL is effective and can learn more semantic features than MAE in small-scale datasets.", + "bbox": [ + 75, + 671, + 467, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 816, + 225, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of feature augmentation. We propose two designs to obtain a dual-branch structure by performing feature augmentation. As shown in Table 7, (a) is the results of MAE as the baseline when pre-trained on COCO. Table", + "bbox": [ + 75, + 839, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7 (b) indicates the results of the DLE design are slightly inferior to the baseline. However, the results outperform the baseline MAE when aligning the predictions according to (c). Similarly, the results for (d) using AEE are not as favorable as MAE, while the results for (e) incorporating alignment operations outperform MAE. These experimental findings collectively indicate that creating learnable pseudo labels based on feature augmentation during the pretraining phase can improve the representation of pre-trained models. Due to the relatively comparable performance of (e) and (c), with (e) incurring significantly lower computational costs, we opt for (e) as the final experimental design.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of alignment loss. We compare two loss functions: one commonly used in contrastive learning, which is cosine similarity loss, and the other is L1 loss. The following is a comparison between cosine similarity loss and L1 loss:", + "bbox": [ + 496, + 273, + 890, + 347 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69ba308ba0f4e19f4d05085099d1e36f59f96cdf0c1a2becee02dcea9f068d79.jpg", + "table_caption": [ + "Table 7. Ablations for ASL: Effect of feature augmentation. We conduct pre-training experiments on COCO dataset and report linear evaluation on ImageNet-1K and detection results on COCO." + ], + "table_footnote": [], + "table_body": "
MAEcosine similarity lossL1 loss[cls] align
46.8%47.6%48.6%45.6%
", + "bbox": [ + 501, + 357, + 888, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the above, experimental results indicate that the L1 loss facilitates the model in learning better representations compared to cosine similarity loss. Besides, we conduct an additional experiment involving alignment with the [cls] token on top of the L1 loss. The results indicate that introducing global contrastive learning during pre-training on the COCO dataset has a detrimental effect on performance.", + "bbox": [ + 496, + 393, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 517, + 625, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper points out the inability of current hybrid SOTA methods to extend training to arbitrary scenarios due to the conflict between MAE and CL. It then theoretically identifies the root of the issue as MAE being patch-level CL, conflicting with global CL. To address this conflict, we introduce a novel self-supervised learning method ASL to learn visual representations from arbitrary scenarios. Specifically, this paper abandons the global-level consistency of CL and proposes explicit patch-level learning. ASL adopts patch-level feature augmentation to generate dual-branch features by the encoder of MAE, then these features perform patch-level learning through a decoder module. Finally, ASL can learn heterogeneous representations while avoiding the conflicts. Our method demonstrates robustness and versatility in multiple pre-training datasets and downstream visual tasks. We expect that our study will draw the attention of the community to large-scale pre-train from arbitrary scenarios and contribute to the development of visual foundation models.", + "bbox": [ + 496, + 542, + 890, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by National Key R&D Program of China under Grant No. 2021YFE0205700 and National Natural Science Foundation of China (No.62276260, No.62176254, No.62076235).", + "bbox": [ + 496, + 829, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22974", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 2, 3, 7", + "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, pages 9912-9924, 2020. 1, 3", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv: Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 8", + "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey K Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703, 2020. 1, 2, 3", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 1, 3, 4", + "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E. Hinton. Big self-supervised models are strong semi-supervised learners. In Advances in Neural Information Processing Systems, pages 22243-22255, 2020.", + "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3", + "[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 3, 5, 7", + "[9] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.7", + "[10] Zhiyang Chen, Yousong Zhu, Zhaowen Li, Fan Yang, Wei Li, Haixin Wang, Chaoyang Zhao, Liwei Wu, Rui Zhao, Jinqiao Wang, et al. Obj2seq: Formatting objects as sequences with class prompt for visual tasks. Advances in Neural Information Processing Systems, 35:2494-2506, 2022. 3", + "[11] Zhiyang Chen, Yousong Zhu, Fan Yang, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. The devil is in details: Delving into lite ffn design for vision transformers. Available at SSRN 4299967, 2023. 3", + "[12] Zhiyang Chen, Yousong Zhu, Yufei Zhan, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. Mitigating hallucination in visual language models with visual supervision. arXiv preprint arXiv:2311.16479, 2023. 3", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 6", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, 2018. 3", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3", + "[16] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 7", + "[17] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 5", + "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 1, 3, 4, 5", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In The IEEE International Conference on Computer Vision, 2017. 3, 7", + "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1, 3, 4", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 2, 3, 4, 6, 7, 8", + "[23] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and Joao Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 3", + "[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In European Conference on Computer Vision, pages 123-143. Springer, 2022. 3", + "[25] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. International Conference on Learning Representations, 2019. 1, 3", + "[26] Tianyu Hua, Yonglong Tian, Sucheng Ren, Hang Zhao, and Leonid Signal. Self-supervision through random seg" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22975", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ments with autoregressive coding (rand sac). arXiv preprint arXiv:2203.12054, 2022.8", + "[27] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. arXiv preprint arXiv:2207.13532, 2022. 2, 3, 6", + "[28] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 1", + "[29] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1", + "[30] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3", + "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[32] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. arXiv preprint arXiv:2210.02077, 2022. 2, 3", + "[33] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3, 7", + "[34] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 23390-23400, 2023. 2", + "[35] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Mst: Masked self-supervised transformer for visual representation. In Advances in Neural Information Processing Systems, 2021. 1, 2, 3", + "[36] Zhaowen Li, Xu Zhao, Chaoyang Zhao, Ming Tang, and Jinqiao Wang. Transfering low-frequency features for domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 3", + "[37] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 14627-14636, 2022. 2, 3", + "[38] Zhaowen Li, Xu Zhao, Peigeng Ding, Zongxing Gao, Yuting Yang, Ming Tang, and Jinqiao Wang. Freconv: Frequency branch-and-integration convolutional networks. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 258-263. IEEE, 2023. 3", + "[39] Zhaowen Li, Yousong Zhu, Zhiyang Chen, Wei Li, Chaoyang Zhao, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Efficient masked autoencoders with self-consistency. arXiv preprint arXiv:2302.14431, 2023. 3", + "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 6, 7", + "[41] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020. 2", + "[42] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3", + "[43] Peter McCullagh and John A Nelder. Generalized linear models. Routledge, 2019. 4", + "[44] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations. arXiv preprint arXiv:2210.16870, 2022. 2, 3", + "[45] Duy-Kien Nguyen, Vaibhav Aggarwal, Yanghao Li, Martin R Oswald, Alexander Kirillov, Cees GM Snoek, and Xinlei Chen. R-mae: Regions meet masked autoencoders. arXiv preprint arXiv:2306.05411, 2023. 7", + "[46] David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 IEEE international conference on neural networks (ICNN'94), pages 55-60. IEEE, 1994. 4", + "[47] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proceedings of the European Conference on Computer Vision, 2016. 3", + "[48] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3, 4", + "[49] OpenAI. Introducing chatgpt, 2022. 3", + "[50] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3", + "[51] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3", + "[52] Hao Quan, Xingyu Li, Weixing Chen, Mingchen Zou, Ruijie Yang, Tingting Zheng, Ruiqun Qi, Xinghua Gao, and Xiaoyu Cui. Global contrast masked autoencoders are powerful pathological representation learners. arXiv preprint arXiv:2205.09048, 2022. 2, 3", + "[53] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, 2015. 3", + "[54] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. Proceedings of the IEEE conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22976", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1", + "[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3", + "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 5998-6008, 2017. 5", + "[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. arXiv preprint arXiv:2011.09157, 2020.3", + "[59] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 3, 4", + "[60] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision, pages 418-434, 2018. 3", + "[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. arXiv preprint arXiv:2011.10043, 2020. 3", + "[62] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 5", + "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. 2, 3", + "[64] Kun Yi, Yixiao Ge, Xiaotong Li, Shusheng Yang, Dian Li, Jianping Wu, Ying Shan, and Xiaohu Qie. Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616, 2022. 2, 3", + "[65] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344, 2022. 2, 3", + "[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Proceedings of the European Conference on Computer Vision, 2016. 3", + "[67] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In The Eleventh International Conference on Learning Representations. 1, 2, 3, 6, 7", + "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6, 7", + "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 6", + "bbox": [ + 535, + 92, + 890, + 118 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22977", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_model.json b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ee96f9423471320b8aed4c635418f2da00b628b --- /dev/null +++ b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_model.json @@ -0,0 +1,2411 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.825, + 0.153 + ], + "angle": 0, + "content": "Self-Supervised Representation Learning from Arbitrary Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.179, + 0.844, + 0.306 + ], + "angle": 0, + "content": "Zhaowen Li \\(^{1,2}\\) Yousong Zhu \\(^{1\\boxtimes}\\) Zhiyang Chen \\(^{1,2}\\) Zongxin Gao \\(^{7}\\) Rui Zhao \\(^{6}\\) Chaoyang Zhao \\(^{1,5}\\) Ming Tang \\(^{1}\\) Jinqiao Wang \\(^{1,2,3,4,5\\boxtimes}\\) \nFoundation Model Research Center, Institute of Automation, Chinese Academy of Science \\(^{1}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences \\(^{2}\\) Peng Cheng Laboratory \\(^{3}\\) Wuhan AI Research \\(^{4}\\) Objectye Inc. \\(^{5}\\) Qing Yuan Research Institute, Shanghai Jiao Tong University \\(^{6}\\) Independent Researcher \\(^{7}\\) {zhaowen.li, yousong.zhu, jqwang}@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.473, + 0.766 + ], + "angle": 0, + "content": "Current self-supervised methods can primarily be categorized into contrastive learning and masked image modeling. Extensive studies have demonstrated that combining these two approaches can achieve state-of-the-art performance. However, these methods essentially reinforce the global consistency of contrastive learning without taking into account the conflicts between these two approaches, which hinders their generalizability to arbitrary scenarios. In this paper, we theoretically prove that MAE serves as a patch-level contrastive learning, where each patch within an image is considered as a distinct category. This presents a significant conflict with global-level contrastive learning, which treats all patches in an image as an identical category. To address this conflict, this work abandons the nongenerizable global-level constraints and proposes explicit patch-level contrastive learning as a solution. Specifically, this work employs the encoder of MAE to generate dual-branch features, which then perform patch-level learning through a decoder. In contrast to global-level data augmentation in contrastive learning, our approach leverages patch-level feature augmentation to mitigate interference from global-level learning. Consequently, our approach can learn heterogeneous representations from a single image while avoiding the conflicts encountered by previous methods. Massive experiments affirm the potential of our method for learning from arbitrary scenarios." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.792, + 0.209, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Currently, self-supervised learning (SSL) in computer vision [1, 2, 4-7, 18, 21, 22, 35, 59] is capable of adopting self-defined pseudo labels as supervision [28, 29] and holds the promise in leveraging large amounts of unlabeled data" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.363, + 0.858, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Figure 1. Hybrid methods further enhance the global consistency of CL. We measures representational similarity among tokens for some popular self-supervised pre-trained models. Among them, ccMIM [67] and SIM [54] employ MAE as the MIM baseline method while iBOT [69] adopts DINO [3] as the CL baseline method for hybrid approach. It indicates that hybrid methods, while increasing linear evaluation scores, also enhance the representational similarity among tokens. This suggests that hybrid methods fundamentally further strengthen the global semantic consistency of CL. All of the models adopt ViT-B." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.903 + ], + "angle": 0, + "content": "to build the foundational models. SSL methods focus on designing different pretext tasks, and there are two mainstream approaches for learning visual features in the community. One of the most promising directions among them is contrastive learning (CL) [2, 5-7, 18, 21, 25, 48]. It assumes that different views of a single image are of the same instance. Its objective is to learn a global-level feature representation that discriminates among images. CL leverages trainable networks to generate semantic pseudo labels [55], thereby enhancing the model's capability to extract semantic information. On the other hand, the masked image mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.398, + 0.902 + ], + "angle": 0, + "content": "Corresponding author: Yousong Zhu and Jinqiao Wang." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22967" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.094, + 0.289, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.098, + 0.465, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Visualization of patches in an image from COCO dataset. The image is partitioned into multiple patches, and different patches from the image may represent different semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.468, + 0.461 + ], + "angle": 0, + "content": "eling (MIM) [1, 4, 15, 22, 35, 63, 69] has become the focus in the community. MIM is a pretext task that involves masking some patches on the input image and predicting the original information of the masked patches based on their context. In these methods, BEiT [1] predicts discrete tokens created by the pre-trained model of VQVAE [56]. In MAE [22], the pseudo labels are the given pixels. Among these MIM methods, MAE shows impressive transfer performance on various dense prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The learning methodologies of CL and MIM are so distinct that a fundamental question arises: can their concurrent learning mutually enhance each other? In fact, extensive work [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] shows hybrid methods of CL and MIM achieve state-of-the-art (SOTA) performance on benchmark ImageNet [13] dataset. In our study, we choose some pre-trained models trained using popular self-supervised methods and measure their representational similarity for tokens as well as linear probing results, as depicted in Figure 1. Specifically, ccMIM [67] and SIM [54] employ MAE as the MIM baseline, while iBOT [69] adopts DINO [3] as the CL baseline for the hybrid approach. The figure illustrates that these hybrid methods have achieved better linear probing results compared to their respective baseline methods. However, they have further reinforced representational similarity for tokens, resulting in homogeneous token representations. This also implies that the hybrid methods essentially enhance the global consistency of CL, and the performance of SOTA methods heavily relies on the gains achieved through CL. Consequently, the success of these methods heavily relies on the single-centric-object data [37, 41] such as those in ImageNet dataset, and they cannot be generalized to non-iconic scenarios, making them unsuitable for training in arbitrary scenarios - regardless of whether they come from single-centric-object scenarios or non-iconic ones. Also, a recent study [34] indicates that the hybrid approach of MAE and CL exhibits inferior performance compared to the baseline in large-scale arbitrary scenarios. This goes against the pri" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "mary promise of self-supervised learning, which aimed to build foundation models using a large amount of data for broader applicability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Actually, the inability of SOTA methods to train in arbitrary scenarios can be attributed to the conflict between CL and MIM. In this paper, we first theoretically prove that MAE is a patch-level CL, treating each patch within an image as a separate category. This conflicts with global-level CL, where all patches from the image are treated as an identical category. As illustrated in Figure 2, a non-iconic image from the COCO [40] dataset is partitioned into multiple non-overlapping patches according to ViT [15], and each patch may represent a different semantics. MAE treats each object/patch in these patches as a distinct category, whereas CL considers all patches as an identical category. Therefore, when these two methods are combined for training on non-iconic data, significant conflicts and confusion arise, rendering the hybrid approach incapable of learning effective representations for non-iconic data and arbitrary scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.893, + 0.699 + ], + "angle": 0, + "content": "To alleviate the problems, in this paper, we propose a self-supervised framework, called Arbitrary Self-supervised Learning (ASL), to learn representation from arbitrary scenarios. The framework leverages the principles of CL to enhance the patch-level semantic consistency of the model while avoiding the conflict between CL and MAE. Specifically, our approach involves designing a patch-level feature augmentation to effectively circumvent the limitation of global-level data augmentation in CL. According to the patch-level augmentation, ASL employs the encoder of MAE to generate dual-branch features, which then perform patch-level contrastive learning through a decoder module. We also retain the original MIM pretext task to ensure the rationality of the model. It is noteworthy that our approach is proposed for arbitrary scenarios, primarily experimenting on the non-iconic COCO dataset that closely represents natural scenes. Simultaneously, we conduct the pre-training experiments on both ImageNet and COCO. The results indicate that ASL is the sole method of attaining superior performance in arbitrary pre-training, further underscoring the generalizability and robustness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.701, + 0.825, + 0.715 + ], + "angle": 0, + "content": "Overall, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.717, + 0.891, + 0.791 + ], + "angle": 0, + "content": "- We theoretically prove that MAE is a patch-level contrastive learning, which is in conflict with global contrastive learning, thereby rendering current hybrid methods incapable of achieving generality in arbitrary scenarios pre-training." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.793, + 0.89, + 0.837 + ], + "angle": 0, + "content": "- We propose a self-supervised framework, named Arbitrary Self-supervised Learning (ASL), to mitigate conflicts and allow the model to train in arbitrary scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.838, + 0.891, + 0.898 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness and transfer ability of our framework. Specifically, the models pre-trained with ASL achieve SOTA performance in arbitrary scenarios." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.717, + 0.891, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "22968" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.222 + ], + "angle": 0, + "content": "It is known that the promise of self-supervised learning [1, 2, 4-7, 15, 18, 21, 22, 25, 30, 35, 37, 47, 48, 50, 51, 59, 63, 66, 69] in computer vision is to establish a visual foundation model by leveraging large dataset, just like ChatGPT [49] in natural language processing. These representations of the foundation model can be transferred to various downstream tasks [10, 12, 13, 19, 20, 33, 36, 38, 39, 53, 60]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.228, + 0.295, + 0.243 + ], + "angle": 0, + "content": "2.1. Instance discrimination" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.473, + 0.569 + ], + "angle": 0, + "content": "The basic principle of contrastive learning/instance discrimination is that different views of an image are still the same category. As a significant representative of these approaches, MoCo [21] improves the training of contrastive learning methods by storing representations from a momentum encoder instead of the trained network. Then, SimCLR [5] shows that the memory bank can be entirely replaced with the elements from the same batch when the batch size is set large enough. Furthermore, BYOL [18] proposes an asymmetric structure and directly bootstraps the representations by attracting different features from the same instance and shows that contrastive learning without negative samples can also learn excellent visual representations. Also, the asymmetric structure is often directly adopted by subsequent work. Some approaches [23, 24, 58, 61] attempt to transfer global-level prior to pixel-level or region-level learning and acquire visual representations for dense tasks. Also, UniVIP [37] proposes a three-level pre-training task and learns versatile representations from non-iconic images. Recently, MoCo v3 [8] and DINO [3] replace CNNs with ViT [15] and achieve superior performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.303, + 0.592 + ], + "angle": 0, + "content": "2.2. Masked image modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Motivated by MLM of BERT [14] in NLP, the pioneering work iGPT [4] operates on sequences of pixels and predicts unknown pixels. Benefiting from the proposal of Vision Transformer [11, 15, 42], MIM achieves performance comparable to CL. MST [35] is the first to introduce MIM into the siamese structure and proposes the attention-guided mask strategy. Then, iBOT [69] also adopts the siamese structure and obtains impressive performance. However, they have the prior/assumption of image semantic consistency. Besides, BEiT [1] proposes to predict discrete tokens for masked image modeling, yet its transfer performance mainly depends on the quality of the pre-trained model VQVAE [56]. Following the MIM design of ViT, SimMIM [63] predicts pixels of masked patches to perform the MIM task. Different from the above methods of performing MIM task in the encoder, MAE [22] designs a special encoder-decoder structure to make the decoder perform the MIM task and abandon the MIM-performing module in downstream visual tasks. Among these methods, MAE shows excellent performance in various downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.664, + 0.107 + ], + "angle": 0, + "content": "2.3. Hybrid methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.462 + ], + "angle": 0, + "content": "An intuitive idea is to combine the above approaches [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] to achieve better performance. MST [35] firstly adopts MIM to a contrastive learning framework [3] and restores the pixels of patches, then iBOT [69] predicts the semantic tokens. Moreover, DINO v2 [50] achieves SOTA performance based on the iBOT through several engineering optimizations, with the most significant being data curation. It reduces the original dataset of 1.3 billion images to 142 million images, primarily using the ImageNet dataset as the main set of query images. However, it is not suitable for training in arbitrary scenarios. More methods [27, 32, 44, 52, 54, 64, 65, 67] are based on the MAE framework rather than the MST/iBOT architecture, as it conserves approximately 5 times the computational resources when compared to the MST/iBOT framework under the ViT-B setting. Additionally, as the model scales up, the proportion of saved computational resources increases further. However, current SOTA methods primarily benefit from contrastive learning, thus demonstrating SOTA performance on the benchmark single-centric-object ImageNet dataset. When extended to arbitrary or natural scenes, a pronounced conflict between CL and MAE arises." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.473, + 0.61, + 0.491 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.63, + 0.515 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.694, + 0.535 + ], + "angle": 0, + "content": "3.1.1 Masked autoencoders" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.894, + 0.793 + ], + "angle": 0, + "content": "As noted in the prior work [22], for a dataset \\(X\\) without manual annotations, an image \\(x\\) can be divided into \\(n\\) regular non-overlapping patches. MAE first samples a subset of patches and masks the remaining ones, and then acquires the masked patches \\(x_{m_i}\\), where \\(i = 1,\\dots,\\eta\\) and \\(\\eta\\) is the set of possible masks of length, and visible patches \\(x_{v_j}\\), where \\(j = 1,\\dots,\\kappa\\) and \\(\\kappa = n - \\eta\\). The mask process of these patches follows a uniform distribution, named the random mask strategy. Moreover, MAE feds these visible patches into encoder (ViT) \\(f(\\cdot)\\), parameterized by \\(\\theta\\), and obtains the encoded visible tokens, then puts these encoded visible tokens with mask tokens into decoder \\(g(\\cdot)\\) to perform the MIM task, parameterized by \\(\\xi\\), and obtains the predicted patches. Finally, MAE minimizes the mean squared error (MSE) between the reconstructed and original images in the pixel space. The loss of a single patch prediction is shown as Eq(1)." + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.801, + 0.892, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i n g l e}} (i) = \\left\\| x _ {m _ {i}} - g \\left(\\xi ; f \\left(\\theta ; x _ {v _ {1}}, \\dots , x _ {v _ {\\kappa}}, \\mathbf {m a s k}\\right)\\right) \\right\\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.871 + ], + "angle": 0, + "content": "For MAE, the total loss function that trains the complete dataset is defined as Eq (2)." + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.88, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M A E} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\underset {i \\sim \\eta} {\\mathbb {E}} \\mathcal {L} _ {\\text {s i n g l e}} (i), \\tag {2}\n\\]" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22969" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.264, + 0.107 + ], + "angle": 0, + "content": "3.1.2 Contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Similarly, for an image \\( x \\) in the dataset \\( X \\), contrastive learning first generates two views under random data augmentation [5, 18], which are then fed into the network separately. Commonly, the network consists of the encoder and momentum encoder. Then, the two global features \\( z_{s} \\) and \\( z_{t} \\) can be obtained. Let \\( \\mathrm{sim}(z_s,z_t) = z_s^\\top z_t / \\| z_s\\| \\| z_t\\| \\) denotes the dot product between \\( \\ell_2 \\) normalized \\( z_{s} \\) and \\( z_{t} \\) (i.e. cosine similarity). The InfoNCE loss function [5, 21, 48, 59] for a positive pair of examples is defined as Eq (3), where \\( \\tau \\) is a temperature coefficient. It pulls one prediction closer to the self-defined pseudo label while pushing other predictions in a mini-batch apart." + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.302, + 0.469, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {I n f o N C E}} (x) = - \\log \\frac {\\exp \\left(\\operatorname* {s i m} \\left(\\boldsymbol {z} _ {s} , \\boldsymbol {z} _ {t}\\right) / \\tau\\right)}{\\sum_ {k \\neq s} ^ {2 N} \\exp \\left(\\operatorname* {s i m} \\left(\\boldsymbol {z} _ {s} , \\boldsymbol {z} _ {k}\\right) / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.469, + 0.39 + ], + "angle": 0, + "content": "Hence, the total loss function that trains the complete dataset is defined as Eq(4)." + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.403, + 0.469, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t r a s}} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\mathcal {L} _ {\\text {I n f o N C E}} (x), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.469, + 0.485 + ], + "angle": 0, + "content": "It is noted that the advantage of CL is that semantic learning and its self-defined pseudo label are acquired by the updatable and learnable network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.276, + 0.513 + ], + "angle": 0, + "content": "3.2. The analysis of MAE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.469, + 0.613 + ], + "angle": 0, + "content": "According to [46], it is well known that minimizing MSE can be equivalent to maximum likelihood estimation. Therefore, a single patch prediction of MAE can be considered as the mean of a noisy prediction distribution, and the distribution can be modeled as a Gaussian distribution in the classic probabilistic interpretation [43] in Eq (5)." + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.627, + 0.469, + 0.645 + ], + "angle": 0, + "content": "\\[\np \\left(x _ {m _ {i}} \\mid x _ {\\text {i n p u t s}}; \\theta , \\xi\\right) = \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.469, + 0.795 + ], + "angle": 0, + "content": "\\(x_{input}\\) is the input tokens and equal to \\(\\{x_{v_1},\\dots,x_{v_\\kappa},\\mathbf{mask}\\}\\), \\(x_{p_i}\\) is the single predicted patch, \\(x_{m_i}\\) is the pseudo label corresponding to this patch, and \\(\\sigma_{noise}\\) is the scale of an independent and identically distributed error term \\(\\epsilon \\sim \\mathcal{N}(0,\\sigma_{noise}^2\\mathbf{I})\\). Meanwhile, MSE equals to the negative log likelihood (NLL) loss of the prediction distribution \\(p(x_{m_i}|x_{input};\\theta ,\\xi)\\) [46]. The network trained by MSE predicts a single patch in fact learns to model \\(p(x_{m_i}|x_{input})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "MAE shows impressive performance on various downstream tasks [22], it reconstructs the masked patches through learning from a large amount of data, and the pseudo labels are the masked image and have no interference from manual annotations. Hence, it is reasonable to consider the \\( p(x_{m_i}) \\) is uniform, and the MSE is equivalent to Eq(6), where \\( C \\) is a constant and equal to \\( p(x_{m_i}) \\), and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.857, + 0.107 + ], + "angle": 0, + "content": "the detailed derivation of this section in the Appendix." + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.114, + 0.89, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i n g l e}} (i) \\cong - \\log \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right) \\\\ + \\log \\int \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right) \\cdot p \\left(x _ {m _ {i}}\\right) d x _ {m _ {i}}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.189, + 0.892, + 0.221 + ], + "angle": 0, + "content": "For the second part of Eq(6), we utilize monte carlo method to solve \\( p(x_{m_i}) \\) and can obtain the Eq(7)." + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.229, + 0.89, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\int \\mathcal {N} (x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {n o i s e} ^ {2} \\mathbf {I}) \\cdot p (x _ {m _ {i}}) d x _ {m _ {i}} \\\\ \\approx \\frac {1}{N} \\sum_ {b = 1} ^ {N} \\mathcal {N} \\left(x _ {m _ {i _ {(b)}}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I})\\right), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.311, + 0.892, + 0.391 + ], + "angle": 0, + "content": "The monte carlo method treats all pseudo labels in a training batch as random samples from \\( p(x_{m_i}) \\). Hence, for pseudo labels in a training batch \\( B = \\{x_{m_{i(1)}}, x_{m_{i(2)}}, \\ldots x_{m_{i(N)}}\\} \\), the loss is defined as Eq(8), where \\( \\lambda = 2\\sigma_{noise}^2 \\) is a temperature coefficient." + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.399, + 0.89, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i n g l e}} (i) \\cong - \\log \\frac {\\exp \\left(- \\left| \\left| x _ {p _ {i}} - x _ {m _ {i}} \\right| \\right| ^ {2} / \\lambda\\right)}{\\sum_ {x _ {m _ {i}} ^ {\\prime} \\in B} \\exp \\left(- \\left| \\left| x _ {p _ {i}} - x _ {m _ {i}} ^ {\\prime} \\right| \\right| ^ {2} / \\lambda\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.627 + ], + "angle": 0, + "content": "According to Eq (3) and Eq (8), it can be observed that the form of the two loss functions is similar. Both of the two loss functions pull together the prediction and self-defined pseudo label in the representation space, and push other predictions in a mini-batch apart. The MAE adopts the Euclidean distance while the contrastive learning measures the cosine similarity in representation space. Actually, MAE regards each masked patch as a category and performs patch-level contrastive learning with unlearnable pixels. It explicitly aligns the prediction with its corresponding masked patch and implicitly distinguishes other patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.658 + ], + "angle": 0, + "content": "From the above description, when confronted with arbitrary scenes, MAE and CL will face a severe conflict issue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.799, + 0.683 + ], + "angle": 0, + "content": "3.3. Arbitrary self-supervised learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Based on the previous analysis, it can be concluded that a global prior of CL is both unreasonable and conflicts with MIM when facing arbitrary scenarios. To enable learning from arbitrary scenarios, can there be a self-supervised method that harnesses the advantages of CL in semantic learning (learnable self-pseudo labels) and the heterogeneous representation learning of MIM, while avoiding global constraints and conflicts? The answer is affirmative. In this paper, we propose a new framework, named Arbitrary Self-supervised Learning (ASL), to learn from arbitrary scenarios while retaining their advantages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Specifically, we adopt the popular MAE method as the basic MIM to ensure the initial heterogeneous representations. Then, inspired by CL using a dual-branch structure" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22970" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.104, + 0.868, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.312, + 0.418, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.42, + 0.264, + 0.431 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.319, + 0.856, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.419, + 0.581, + 0.431 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.432, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Figure 3. The pipeline of our ASL. The visible patches are fed into the network and the dual-branch features can be obtained by our designs. Then the encoded features with mask tokens are fed into the decoder and output the predictions. Finally, we pull together the predictions using L1 loss by one-to-one correspondence. We propose two designs in the network to obtain the dual-branch features: (a) Dropout in the linear embedding. (b) Asymmetric structure at the end of the encoder. In the figure, we adopt the network to represent (a) or (b) design. The olive patch represents the mask token." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.662 + ], + "angle": 0, + "content": "to generate learnable self-pseudo labels, we propose two patch-level designs to generate dual-branch features based on the encoder of MAE while avoiding the global constraints. Subsequently, these dual-branch features are further processed through a decoder module, and CL is performed between the outputs patch by patch to for semantic learning. In contrast to traditional CL, our approach does not entail global data augmentation but instead focuses on patch-level feature augmentation. The specifics of our designs are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Dropout in the linear embedding (DLE). According to previous work in NLP, SimCSE [17] has explored the use of dropout in word embedding for feature augmentation. In line with this, we also adopt dropout in the linear embedding of the encoder to perform patch-level augmentation and acquire two encoded features at the end of the encoder. The details of our method are described in Figure 3 (a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Asymmetric structure at the end of the encoder (AEE). Some prior work [8, 18, 62] have demonstrated that introducing an asymmetric structure, which essentially is a nonlinear function, enhances the semantic features. With this in mind, we incorporate a Transformer layer [57] after the encoder in our approach to build an asymmetric structure for patch-level feature augmentation. This step produces dual-branch features: one is the feature output by the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.556 + ], + "angle": 0, + "content": "encoder, and the other is the feature that has passed through the Transformer layer. Figure 3 (b) illustrates the details of our method. The design is our default experimental setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.692 + ], + "angle": 0, + "content": "The methodology described above pertains to the construction of feature augmentation. Our rationale for utilizing feature augmentation to construct dual-branch features is primarily based on two considerations: 1) Data augmentation in CL is global and not suitable for patch-level learning. 2) The asymmetric structure of CL inherently constitutes an implicit form of feature augmentation, and it is worth exploring whether its success in CL can be extrapolated to the broader domain of self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.782 + ], + "angle": 0, + "content": "After feature augmentation, we obtain the dual-branch features, \\( fea_{ori} \\) and \\( fea_{aug} \\), then we feed them into the decoder module to obtain the corresponding predictions, \\( pred_{ori} \\) and \\( pred_{aug} \\), respectively. The patch-level CL between the outputs does not employ contrastive loss or cosine similarity loss but rather L1 loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.783, + 0.892, + 0.859 + ], + "angle": 0, + "content": "The alignment between any pair is defined as Eq(9), where \\( \\mathsf{sg}[\\cdot] \\) stands for stop gradient. It is noted that our alignment method does not include the alignment of [cls] token according to Section 3.2, which unleashes the potential of the method for training in arbitrary scenarios." + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} (i) = \\left| \\left| \\operatorname {p r e d} _ {\\text {o r i} _ {i}} - s g \\left[ \\operatorname {p r e d} _ {\\text {a u g} _ {i}} \\right] \\right| \\right| + \\left| \\left| s g \\left[ \\operatorname {p r e d} _ {\\text {o r i} _ {i}} \\right] - \\operatorname {p r e d} _ {\\text {a u g} _ {i}} \\right| \\right|, \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22971" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "The loss function of the complete dataset is defined as Eq(10)." + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.122, + 0.469, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S E M} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\underset {i \\sim \\eta} {\\mathbb {E}} \\mathcal {L} _ {s} (i), \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.154, + 0.258, + 0.17 + ], + "angle": 0, + "content": "3.4. Objective function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.177, + 0.469, + 0.239 + ], + "angle": 0, + "content": "The loss function of our ASL consists of the MAE loss and the semantic loss. Hence, the total loss of our ASL is formulated as Eq (11), and each loss coefficient is set to be 1 for equally weighted." + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.251, + 0.469, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {M A E}} + \\mathcal {L} _ {\\text {S E M}}, \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.279, + 0.21, + 0.296 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.304, + 0.357, + 0.319 + ], + "angle": 0, + "content": "4.1. Datasets and evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Pre-training datasets. In the pre-training stage, we select the four popular visual pre-training datasets, MS COCO [40], ImageNet [13], CIFAR10 [31], CIFAR100 [31], to investigate the versatility and transfer ability of the proposed method. For the COCO dataset, its train2017 set contains \\(\\sim 118\\mathrm{k}\\) images. COCO contains more natural and diverse scenes and is a non-iconic image dataset. To further demonstrate the versatility of our algorithm for training in arbitrary scenarios, we conduct pre-training on both ImageNet and COCO datasets. Given the substantial difference in data volume between ImageNet and COCO, we select a subset of ImageNet, referred to as ImageNet-100, to maintain a balanced dataset size. Moreover, we pre-train on ImageNet-1K to demonstrate the effectiveness of ASL on curated data. Finally, CIFAR-10 and CIFAR-100 datasets are small-scale image datasets, and they all contain 60000 images with \\(32\\times 32\\) size that belong to 10 and 100 categories, respectively. We pre-train models on the CIFAR-10 and CIFAR-100 datasets to demonstrate the effectiveness of ASL in the small-scale dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation datasets. In this paper, ASL is evaluated by the linear probing and fine-tuning classification task on ImageNet, CIFAR-10, and CIFAR-100 datasets, which are popular image classification datasets. Additionally, for computer vision, object detection, and semantic segmentation are dense prediction tasks since training images of these tasks contain massive objects inside each image. The validation of dense prediction tasks can better reflect the transfer ability of self-supervised models. Therefore, we conduct extensive experiments on COCO [40] and ADE20k [68] datasets to verify the generalization and transfer ability of ASL. COCO is a popular object detection and instance segmentation dataset, which train2017 contains about \\(118k\\) images and test on the val2017 contains \\(5k\\) images. The challenging dataset contains human annotations for 80 classes. Moreover, ADE20K is also a challenging semantic segmentation dataset, and it contains about \\(25k\\) training images and \\(2k\\) validation images with 150 categories." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.9, + 0.22 + ], + "angle": 0, + "content": "
MethodArch.FLOPsEpochsLPAPb
Pre-training on COCO dataset:
MAE [22]ViT-B1×80046.8%49.3%
CMAE [27]ViT-B~ 3×80046.6%49.0%
ccMIM [67]ViT-B~ 3.3×80046.7%49.0%
SIM [54]ViT-B~ 2.8×80046.0%48.7%
iBOT [69]ViT-B~ 6×80045.2%48.6%
ASLViT-B~ 1.5×80048.6%50.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.222, + 0.892, + 0.264 + ], + "angle": 0, + "content": "Table 1. Evaluation of pre-trained models on the COCO dataset. We report Linear Probing (LP) top-1 accuracy on the ImageNet-1K val set and bbox mAP on COCO dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.267, + 0.895, + 0.384 + ], + "angle": 0, + "content": "
MethodArch.Param.EpochsLPFT
Pre-training on ImageNet-100 + COCO:
MAE [22]ViT-B8680074.5%90.9%
iBOT[69]ViT-B8680071.8%88.5%
ccMIM [67]ViT-B8680072.3%88.7%
ASLViT-B8680079.6%92.4%
ASLViT-B86400085.9%94.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.386, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Table 2. ImageNet-100 Top-1 accuracy of different methods under linear probing (LP) and fine-tuning (FT) setting to evaluate the performance of the model pre-trained on both ImageNet-100 and COCO. We report top-1 accuracy on the ImageNet-100 val set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.492 + ], + "angle": 0, + "content": "The valuation metrics, pre-training settings, and the experimental settings of downstream tasks in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.773, + 0.516 + ], + "angle": 0, + "content": "4.2. Pre-training on COCO dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In order to validate the adaptability of our approach for training in arbitrary scenarios, we initially conduct pretraining on the COCO dataset and subsequently conduct a fair comparison with other methods under the same experimental settings. In this comparison, we evaluate the ImageNet linear evaluation and the COCO detection results. Linear evaluation, which involves using frozen features, provides a more comprehensive assessment of a model's representation. Given that the pre-training experiments are conducted on the COCO dataset, using human-labeled data from COCO to test the learned representations is reasonable. As shown in Table 1, we observe that methods combining MAE with contrastive learning, such as CMAE, ccMIM, and SIM, exhibit weaker performance when pretrained on the COCO dataset, even though they achieved SOTA performance on ImageNet. This observation suggests that existing SOTA methods are not well-suited for training in arbitrary scenarios. Furthermore, our approach demonstrates superior performance after pre-training on the COCO dataset, surpassing the performance of MAE. Generally, introducing a dual-branch structure similar to contrastive learning typically results in a significant increase in computational overhead, often exceeding threefold. Diverging from the emphasis on the encoder in hybrid methods, our AEE design prioritizes the decoder, significantly reducing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22972" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.887, + 0.296 + ], + "angle": 0, + "content": "
MethodPre-train IterationsDatasetObject detectionInstance segmentation
\\(AP^b\\)\\(AP_{50}^b\\)\\(AP_{75}^b\\)\\(AP^m\\)\\(AP_{50}^m\\)\\(AP_{75}^m\\)
Supervised [9]~ 94kImageNet-1K47.9%--42.9%--
MoCo v3 [8]~ 187kImageNet-1K47.9%--42.7%--
BEiT [1]~ 249kImageNet-1K + DALLE49.8%--44.4%--
MAE [22]~ 499kImageNet-1K50.4%70.8%55.7%44.9%68.3%48.9%
CAE [9]~ 499kImageNet-1K50.0%70.9%54.8%44.0%67.9%47.6%
R-MAE [45]~ 115kCOCO50.6%--45.0%--
R-MAE [45]~ 250kCOCO50.8%--45.2%--
ccMIM [67]~ 249kImageNet-1K50.3%71.2%55.0%44.5%68.4%47.9%
ASL~ 47kImageNet-100 + COCO50.7%71.1%55.6%45.0%68.3%49.0%
ASL~ 236kImageNet-100 + COCO51.5%71.9%57.2%45.7%69.2%50.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.298, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Table 3. Results of object detection and instance segmentation on COCO. The architecture of various methods is ViT-B, and we report the bounding box AP and mask AP on COCO val2017." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.328, + 0.476, + 0.508 + ], + "angle": 0, + "content": "
MethodDatasetPre-train IterationsADE mIoU
Supervised [22]ImageNet-1K~ 93k47.4%
SplitMask [16]ADE20K~ 128k45.7%
MoCo v3 [8]ImageNet-1K~ 187k47.3%
BEiT [22]ImageNet-1K+DALLE~ 249k47.1%
R-MAE [45]COCO~ 115k46.8%
R-MAE [45]COCO~ 250k47.0%
ccMIM [67]ImageNet-1K~ 249k47.7%
MAE [22]ImageNet-1K~ 499k48.1%
ASLImageNet-100 + COCO~ 236k48.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.509, + 0.471, + 0.539 + ], + "angle": 0, + "content": "Table 4. Results of semantic segmentation on ADE20K. We report results measured by mIoU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.471, + 0.621 + ], + "angle": 0, + "content": "ing computational costs in comparison. Meanwhile, due to the fixed-size and lightweight design of the decoder, our design incurs decreasing computational overhead as the model scale increases. This highlights the potential of our method for pre-training in non-iconic scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.443, + 0.649 + ], + "angle": 0, + "content": "4.3. Pre-training on both COCO and ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.793 + ], + "angle": 0, + "content": "To further demonstrate the adaptability of ASL across diverse scenarios, we conduct the pre-training experiments on both ImageNet-100 and COCO datasets. We assess the performance of the model on benchmark tasks: image classification, object detection, and semantic segmentation. The ASL models are primarily trained with 800 epochs and 4000 epochs, with the number of training iterations for the 4000-epoch training similar to that of the ImageNet-1K iterations for the 800-epoch training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.268, + 0.831 + ], + "angle": 0, + "content": "4.3.1 Image classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Since our training data comprises the ImageNet-100 dataset, we conducted testing for classification performance through both linear probing and fine-tuning on ImageNet-100. The experimental results are presented in Table 2. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.895, + 0.439 + ], + "angle": 0, + "content": "results of pre-training observed from the table align with those of pre-training solely on COCO: ccMIM exhibits relatively weak performance in the context of the pre-training. This can be attributed to conflicts between CL and MAE when dealing with arbitrary scenarios. In contrast, our method successfully mitigates such conflicts and produces more robust representations in arbitrary scenarios." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.456, + 0.859, + 0.472 + ], + "angle": 0, + "content": "4.3.2 Object detection and instance segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.737 + ], + "angle": 0, + "content": "To verify the transfer ability of ASL, we test it on object detection and instance segmentation with COCO [40] dataset. Here we report Box AP and mask AP on the validation set. According to the benchmark validation [22, 33], we choose the Mask R-CNN [20] as the test framework. Due to variations in the sizes of the pre-training datasets used, we opt to avoid epoch-based comparisons in this scenario. We convert epochs to a fixed batch size of 4096 iterations for a fair comparison. Table 3 shows the results of the learned representation by supervised method and different self-supervised methods. It shows our ASL, with approximately 236k iterations (4000 epochs), achieves impressive performance and outperforms 1600-epoch MAE, 1600-epoch CAE and 800-epoch ccMIM pre-trained on ImageNet. ASL surpasses various supervised and self-supervised models pre-trained on ImageNet and COCO, indicating its success in arbitrary scenarios." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.715, + 0.772 + ], + "angle": 0, + "content": "4.3.3 Semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.856 + ], + "angle": 0, + "content": "To further evaluate the generalization of our method, we conduct the semantic segmentation experiments on ADE20K dataset [68]. As shown in Table 4, ASL achieves the \\(48.4\\%\\) mIoU. It also outperforms various supervised and self-supervised models pre-trained on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overall, our ASL shows impressive performance on various mainstream visual benchmark tasks when pre-training on both COCO and ImageNet-100, which demonstrates" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22973" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.09, + 0.46, + 0.156 + ], + "angle": 0, + "content": "
MethodLPFTAPbAPmmIoU
MAE67.8%83.6%50.4%44.9%48.1%
ASL69.2%84.2%51.3%45.7%49.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.158, + 0.468, + 0.199 + ], + "angle": 0, + "content": "Table 5. The ImageNet-1K pre-training results. We report ImageNet classification, COCO detection, and instance segmentation, as well as ADE20k semantic segmentation results." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.204, + 0.468, + 0.278 + ], + "angle": 0, + "content": "
MethodCIFAR-10CIFAR-100
Linear Prob.FinetuningLinear Prob.Finetuning
Supervised [26]-91.3%-64.1%
DINO [3]89.0%94.4%65.8%76.3%
MAE [22]87.3%95.9%54.0%81.1%
ASL91.4%96.5%60.3%82.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.279, + 0.468, + 0.306 + ], + "angle": 0, + "content": "Table 6. Linear probing and fine-tuning classification results on small-scale datasets. We adopt ViT-S as the base model." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.31, + 0.468, + 0.396 + ], + "angle": 0, + "content": "
IDDLEAEEalignmentFLOPsTop-1APb
(a)×××1×46.8%49.3%
(b)✓××~ 2×46.5%49.1%
(c)✓×✓~ 2×48.5%50.4%
(d)×✓×~ 1.5×46.6%49.2%
(e)×✓✓~ 1.5×48.6%50.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.399, + 0.468, + 0.44 + ], + "angle": 0, + "content": "Table 7. Ablations for ASL: Effect of feature augmentation. We conduct pre-training experiments on COCO dataset and report linear evaluation on ImageNet-1K and detection results on COCO." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.453, + 0.468, + 0.483 + ], + "angle": 0, + "content": "ASL has the ability to learn general visual representations from arbitrary scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.493, + 0.339, + 0.509 + ], + "angle": 0, + "content": "4.4. Pre-training on ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.468, + 0.638 + ], + "angle": 0, + "content": "Our method is designed for learning visual representations from uncurated arbitrary scenarios, and it can also yield stronger representations when applied to meticulously curated datasets. To demonstrate this point, we conduct experiments on the curated ImageNet dataset. As depicted in the experimental results shown in Table 5, ASL exhibits superior performance across various downstream tasks compared to MAE, highlighting the versatility of our approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.338, + 0.664 + ], + "angle": 0, + "content": "4.5. Small-scale data pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.468, + 0.806 + ], + "angle": 0, + "content": "As shown in Table 6, the results of our method on these small datasets all show impressive performance. With the same training settings for all experiments, ASL consistently improves MAE by at least \\(4.1\\%\\) for linear probing. However, MIM still lags behind CL in linear evaluation on CIFAR100 and its advantage lies in the fine-tuning superior performance. Our fine-tuning results also reveal it. Totally, these results suggest that ASL is effective and can learn more semantic features than MAE in small-scale datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.227, + 0.833 + ], + "angle": 0, + "content": "4.6. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The effect of feature augmentation. We propose two designs to obtain a dual-branch structure by performing feature augmentation. As shown in Table 7, (a) is the results of MAE as the baseline when pre-trained on COCO. Table" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "7 (b) indicates the results of the DLE design are slightly inferior to the baseline. However, the results outperform the baseline MAE when aligning the predictions according to (c). Similarly, the results for (d) using AEE are not as favorable as MAE, while the results for (e) incorporating alignment operations outperform MAE. These experimental findings collectively indicate that creating learnable pseudo labels based on feature augmentation during the pretraining phase can improve the representation of pre-trained models. Due to the relatively comparable performance of (e) and (c), with (e) incurring significantly lower computational costs, we opt for (e) as the final experimental design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.348 + ], + "angle": 0, + "content": "The effect of alignment loss. We compare two loss functions: one commonly used in contrastive learning, which is cosine similarity loss, and the other is L1 loss. The following is a comparison between cosine similarity loss and L1 loss:" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.358, + 0.89, + 0.389 + ], + "angle": 0, + "content": "
MAEcosine similarity lossL1 loss[cls] align
46.8%47.6%48.6%45.6%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.5 + ], + "angle": 0, + "content": "From the above, experimental results indicate that the L1 loss facilitates the model in learning better representations compared to cosine similarity loss. Besides, we conduct an additional experiment involving alignment with the [cls] token on top of the L1 loss. The results indicate that introducing global contrastive learning during pre-training on the COCO dataset has a detrimental effect on performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.627, + 0.533 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.829 + ], + "angle": 0, + "content": "This paper points out the inability of current hybrid SOTA methods to extend training to arbitrary scenarios due to the conflict between MAE and CL. It then theoretically identifies the root of the issue as MAE being patch-level CL, conflicting with global CL. To address this conflict, we introduce a novel self-supervised learning method ASL to learn visual representations from arbitrary scenarios. Specifically, this paper abandons the global-level consistency of CL and proposes explicit patch-level learning. ASL adopts patch-level feature augmentation to generate dual-branch features by the encoder of MAE, then these features perform patch-level learning through a decoder module. Finally, ASL can learn heterogeneous representations while avoiding the conflicts. Our method demonstrates robustness and versatility in multiple pre-training datasets and downstream visual tasks. We expect that our study will draw the attention of the community to large-scale pre-train from arbitrary scenarios and contribute to the development of visual foundation models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by National Key R&D Program of China under Grant No. 2021YFE0205700 and National Natural Science Foundation of China (No.62276260, No.62176254, No.62076235)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22974" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.225 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, pages 9912-9924, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv: Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey K Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E. Hinton. Big self-supervised models are strong semi-supervised learners. In Advances in Neural Information Processing Systems, pages 22243-22255, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[9] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[10] Zhiyang Chen, Yousong Zhu, Zhaowen Li, Fan Yang, Wei Li, Haixin Wang, Chaoyang Zhao, Liwei Wu, Rui Zhao, Jinqiao Wang, et al. Obj2seq: Formatting objects as sequences with class prompt for visual tasks. Advances in Neural Information Processing Systems, 35:2494-2506, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[11] Zhiyang Chen, Yousong Zhu, Fan Yang, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. The devil is in details: Delving into lite ffn design for vision transformers. Available at SSRN 4299967, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Zhiyang Chen, Yousong Zhu, Yufei Zhan, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. Mitigating hallucination in visual language models with visual supervision. arXiv preprint arXiv:2311.16479, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[16] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In The IEEE International Conference on Computer Vision, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[23] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and Joao Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In European Conference on Computer Vision, pages 123-143. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[25] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. International Conference on Learning Representations, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Tianyu Hua, Yonglong Tian, Sucheng Ren, Hang Zhao, and Leonid Signal. Self-supervision through random seg" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22975" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.47, + 0.119 + ], + "angle": 0, + "content": "ments with autoregressive coding (rand sac). arXiv preprint arXiv:2203.12054, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. arXiv preprint arXiv:2207.13532, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[28] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[29] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[30] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[32] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. arXiv preprint arXiv:2210.02077, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[33] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[34] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 23390-23400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[35] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Mst: Masked self-supervised transformer for visual representation. In Advances in Neural Information Processing Systems, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[36] Zhaowen Li, Xu Zhao, Chaoyang Zhao, Ming Tang, and Jinqiao Wang. Transfering low-frequency features for domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[37] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 14627-14636, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[38] Zhaowen Li, Xu Zhao, Peigeng Ding, Zongxing Gao, Yuting Yang, Ming Tang, and Jinqiao Wang. Freconv: Frequency branch-and-integration convolutional networks. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 258-263. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[39] Zhaowen Li, Yousong Zhu, Zhiyang Chen, Wei Li, Chaoyang Zhao, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Efficient masked autoencoders with self-consistency. arXiv preprint arXiv:2302.14431, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[41] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[42] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.266 + ], + "angle": 0, + "content": "[43] Peter McCullagh and John A Nelder. Generalized linear models. Routledge, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[44] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations. arXiv preprint arXiv:2210.16870, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[45] Duy-Kien Nguyen, Vaibhav Aggarwal, Yanghao Li, Martin R Oswald, Alexander Kirillov, Cees GM Snoek, and Xinlei Chen. R-mae: Regions meet masked autoencoders. arXiv preprint arXiv:2306.05411, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[46] David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 IEEE international conference on neural networks (ICNN'94), pages 55-60. IEEE, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[47] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proceedings of the European Conference on Computer Vision, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[48] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.763, + 0.572 + ], + "angle": 0, + "content": "[49] OpenAI. Introducing chatgpt, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.573, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[50] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[51] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[52] Hao Quan, Xingyu Li, Weixing Chen, Mingchen Zou, Ruijie Yang, Tingting Zheng, Ruiqun Qi, Xinghua Gao, and Xiaoyu Cui. Global contrast masked autoencoders are powerful pathological representation learners. arXiv preprint arXiv:2205.09048, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[53] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[54] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. Proceedings of the IEEE conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22976" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 5998-6008, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. arXiv preprint arXiv:2011.09157, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[59] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[60] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision, pages 418-434, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. arXiv preprint arXiv:2011.10043, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[62] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[64] Kun Yi, Yixiao Ge, Xiaotong Li, Shusheng Yang, Dian Li, Jianping Wu, Ying Shan, and Xiaohu Qie. Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[65] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Proceedings of the European Conference on Computer Vision, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.747, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[67] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In The Eleventh International Conference on Learning Representations. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 6" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22977" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_origin.pdf b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06241e2f21e85c7ce11879d5900bcfc80c1e30f0 --- /dev/null +++ b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/90fe2be8-b444-4296-b490-1791873650b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f3b456438d545ef654a1f192ccb633031ff61c5d0c02fc87844db44db278ffc +size 7687744 diff --git a/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/full.md b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d2163816cb8ff933e3dbd3b5d18c839c5132e5fe --- /dev/null +++ b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/full.md @@ -0,0 +1,342 @@ +# Self-Supervised Representation Learning from Arbitrary Scenarios + +Zhaowen Li $^{1,2}$ Yousong Zhu $^{1\boxtimes}$ Zhiyang Chen $^{1,2}$ Zongxin Gao $^{7}$ Rui Zhao $^{6}$ Chaoyang Zhao $^{1,5}$ Ming Tang $^{1}$ Jinqiao Wang $^{1,2,3,4,5\boxtimes}$ +Foundation Model Research Center, Institute of Automation, Chinese Academy of Science $^{1}$ School of Artificial Intelligence, University of Chinese Academy of Sciences $^{2}$ Peng Cheng Laboratory $^{3}$ Wuhan AI Research $^{4}$ Objectye Inc. $^{5}$ Qing Yuan Research Institute, Shanghai Jiao Tong University $^{6}$ Independent Researcher $^{7}$ {zhaowen.li, yousong.zhu, jqwang}@nlpr.ia.ac.cn + +# Abstract + +Current self-supervised methods can primarily be categorized into contrastive learning and masked image modeling. Extensive studies have demonstrated that combining these two approaches can achieve state-of-the-art performance. However, these methods essentially reinforce the global consistency of contrastive learning without taking into account the conflicts between these two approaches, which hinders their generalizability to arbitrary scenarios. In this paper, we theoretically prove that MAE serves as a patch-level contrastive learning, where each patch within an image is considered as a distinct category. This presents a significant conflict with global-level contrastive learning, which treats all patches in an image as an identical category. To address this conflict, this work abandons the nongenerizable global-level constraints and proposes explicit patch-level contrastive learning as a solution. Specifically, this work employs the encoder of MAE to generate dual-branch features, which then perform patch-level learning through a decoder. In contrast to global-level data augmentation in contrastive learning, our approach leverages patch-level feature augmentation to mitigate interference from global-level learning. Consequently, our approach can learn heterogeneous representations from a single image while avoiding the conflicts encountered by previous methods. Massive experiments affirm the potential of our method for learning from arbitrary scenarios. + +# 1. Introduction + +Currently, self-supervised learning (SSL) in computer vision [1, 2, 4-7, 18, 21, 22, 35, 59] is capable of adopting self-defined pseudo labels as supervision [28, 29] and holds the promise in leveraging large amounts of unlabeled data + +![](images/bafcc348137408f74f60e64216fc5772dc8bdf0e5756deb50592b67a4ef031df.jpg) +Figure 1. Hybrid methods further enhance the global consistency of CL. We measures representational similarity among tokens for some popular self-supervised pre-trained models. Among them, ccMIM [67] and SIM [54] employ MAE as the MIM baseline method while iBOT [69] adopts DINO [3] as the CL baseline method for hybrid approach. It indicates that hybrid methods, while increasing linear evaluation scores, also enhance the representational similarity among tokens. This suggests that hybrid methods fundamentally further strengthen the global semantic consistency of CL. All of the models adopt ViT-B. + +to build the foundational models. SSL methods focus on designing different pretext tasks, and there are two mainstream approaches for learning visual features in the community. One of the most promising directions among them is contrastive learning (CL) [2, 5-7, 18, 21, 25, 48]. It assumes that different views of a single image are of the same instance. Its objective is to learn a global-level feature representation that discriminates among images. CL leverages trainable networks to generate semantic pseudo labels [55], thereby enhancing the model's capability to extract semantic information. On the other hand, the masked image mod + +![](images/415dcbcc76660511c360d1e83dafd5030e5f27229ed021eaa846308f6f969e66.jpg) +Figure 2. Visualization of patches in an image from COCO dataset. The image is partitioned into multiple patches, and different patches from the image may represent different semantics. + +![](images/8b43cbeb1fc5d970d9510bcafa8b43228a83d08e40799b287b80fd8b8d49e885.jpg) + +eling (MIM) [1, 4, 15, 22, 35, 63, 69] has become the focus in the community. MIM is a pretext task that involves masking some patches on the input image and predicting the original information of the masked patches based on their context. In these methods, BEiT [1] predicts discrete tokens created by the pre-trained model of VQVAE [56]. In MAE [22], the pseudo labels are the given pixels. Among these MIM methods, MAE shows impressive transfer performance on various dense prediction tasks. + +The learning methodologies of CL and MIM are so distinct that a fundamental question arises: can their concurrent learning mutually enhance each other? In fact, extensive work [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] shows hybrid methods of CL and MIM achieve state-of-the-art (SOTA) performance on benchmark ImageNet [13] dataset. In our study, we choose some pre-trained models trained using popular self-supervised methods and measure their representational similarity for tokens as well as linear probing results, as depicted in Figure 1. Specifically, ccMIM [67] and SIM [54] employ MAE as the MIM baseline, while iBOT [69] adopts DINO [3] as the CL baseline for the hybrid approach. The figure illustrates that these hybrid methods have achieved better linear probing results compared to their respective baseline methods. However, they have further reinforced representational similarity for tokens, resulting in homogeneous token representations. This also implies that the hybrid methods essentially enhance the global consistency of CL, and the performance of SOTA methods heavily relies on the gains achieved through CL. Consequently, the success of these methods heavily relies on the single-centric-object data [37, 41] such as those in ImageNet dataset, and they cannot be generalized to non-iconic scenarios, making them unsuitable for training in arbitrary scenarios - regardless of whether they come from single-centric-object scenarios or non-iconic ones. Also, a recent study [34] indicates that the hybrid approach of MAE and CL exhibits inferior performance compared to the baseline in large-scale arbitrary scenarios. This goes against the pri + +mary promise of self-supervised learning, which aimed to build foundation models using a large amount of data for broader applicability. + +Actually, the inability of SOTA methods to train in arbitrary scenarios can be attributed to the conflict between CL and MIM. In this paper, we first theoretically prove that MAE is a patch-level CL, treating each patch within an image as a separate category. This conflicts with global-level CL, where all patches from the image are treated as an identical category. As illustrated in Figure 2, a non-iconic image from the COCO [40] dataset is partitioned into multiple non-overlapping patches according to ViT [15], and each patch may represent a different semantics. MAE treats each object/patch in these patches as a distinct category, whereas CL considers all patches as an identical category. Therefore, when these two methods are combined for training on non-iconic data, significant conflicts and confusion arise, rendering the hybrid approach incapable of learning effective representations for non-iconic data and arbitrary scenarios. + +To alleviate the problems, in this paper, we propose a self-supervised framework, called Arbitrary Self-supervised Learning (ASL), to learn representation from arbitrary scenarios. The framework leverages the principles of CL to enhance the patch-level semantic consistency of the model while avoiding the conflict between CL and MAE. Specifically, our approach involves designing a patch-level feature augmentation to effectively circumvent the limitation of global-level data augmentation in CL. According to the patch-level augmentation, ASL employs the encoder of MAE to generate dual-branch features, which then perform patch-level contrastive learning through a decoder module. We also retain the original MIM pretext task to ensure the rationality of the model. It is noteworthy that our approach is proposed for arbitrary scenarios, primarily experimenting on the non-iconic COCO dataset that closely represents natural scenes. Simultaneously, we conduct the pre-training experiments on both ImageNet and COCO. The results indicate that ASL is the sole method of attaining superior performance in arbitrary pre-training, further underscoring the generalizability and robustness of our approach. + +# Overall, we make the following contributions: + +- We theoretically prove that MAE is a patch-level contrastive learning, which is in conflict with global contrastive learning, thereby rendering current hybrid methods incapable of achieving generality in arbitrary scenarios pre-training. +- We propose a self-supervised framework, named Arbitrary Self-supervised Learning (ASL), to mitigate conflicts and allow the model to train in arbitrary scenarios. +- Extensive experiments demonstrate the effectiveness and transfer ability of our framework. Specifically, the models pre-trained with ASL achieve SOTA performance in arbitrary scenarios. + +# 2. Related Work + +It is known that the promise of self-supervised learning [1, 2, 4-7, 15, 18, 21, 22, 25, 30, 35, 37, 47, 48, 50, 51, 59, 63, 66, 69] in computer vision is to establish a visual foundation model by leveraging large dataset, just like ChatGPT [49] in natural language processing. These representations of the foundation model can be transferred to various downstream tasks [10, 12, 13, 19, 20, 33, 36, 38, 39, 53, 60]. + +# 2.1. Instance discrimination + +The basic principle of contrastive learning/instance discrimination is that different views of an image are still the same category. As a significant representative of these approaches, MoCo [21] improves the training of contrastive learning methods by storing representations from a momentum encoder instead of the trained network. Then, SimCLR [5] shows that the memory bank can be entirely replaced with the elements from the same batch when the batch size is set large enough. Furthermore, BYOL [18] proposes an asymmetric structure and directly bootstraps the representations by attracting different features from the same instance and shows that contrastive learning without negative samples can also learn excellent visual representations. Also, the asymmetric structure is often directly adopted by subsequent work. Some approaches [23, 24, 58, 61] attempt to transfer global-level prior to pixel-level or region-level learning and acquire visual representations for dense tasks. Also, UniVIP [37] proposes a three-level pre-training task and learns versatile representations from non-iconic images. Recently, MoCo v3 [8] and DINO [3] replace CNNs with ViT [15] and achieve superior performance. + +# 2.2. Masked image modeling + +Motivated by MLM of BERT [14] in NLP, the pioneering work iGPT [4] operates on sequences of pixels and predicts unknown pixels. Benefiting from the proposal of Vision Transformer [11, 15, 42], MIM achieves performance comparable to CL. MST [35] is the first to introduce MIM into the siamese structure and proposes the attention-guided mask strategy. Then, iBOT [69] also adopts the siamese structure and obtains impressive performance. However, they have the prior/assumption of image semantic consistency. Besides, BEiT [1] proposes to predict discrete tokens for masked image modeling, yet its transfer performance mainly depends on the quality of the pre-trained model VQVAE [56]. Following the MIM design of ViT, SimMIM [63] predicts pixels of masked patches to perform the MIM task. Different from the above methods of performing MIM task in the encoder, MAE [22] designs a special encoder-decoder structure to make the decoder perform the MIM task and abandon the MIM-performing module in downstream visual tasks. Among these methods, MAE shows excellent performance in various downstream tasks. + +# 2.3. Hybrid methods + +An intuitive idea is to combine the above approaches [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] to achieve better performance. MST [35] firstly adopts MIM to a contrastive learning framework [3] and restores the pixels of patches, then iBOT [69] predicts the semantic tokens. Moreover, DINO v2 [50] achieves SOTA performance based on the iBOT through several engineering optimizations, with the most significant being data curation. It reduces the original dataset of 1.3 billion images to 142 million images, primarily using the ImageNet dataset as the main set of query images. However, it is not suitable for training in arbitrary scenarios. More methods [27, 32, 44, 52, 54, 64, 65, 67] are based on the MAE framework rather than the MST/iBOT architecture, as it conserves approximately 5 times the computational resources when compared to the MST/iBOT framework under the ViT-B setting. Additionally, as the model scales up, the proportion of saved computational resources increases further. However, current SOTA methods primarily benefit from contrastive learning, thus demonstrating SOTA performance on the benchmark single-centric-object ImageNet dataset. When extended to arbitrary or natural scenes, a pronounced conflict between CL and MAE arises. + +# 3. Approach + +# 3.1. Preliminary + +# 3.1.1 Masked autoencoders + +As noted in the prior work [22], for a dataset $X$ without manual annotations, an image $x$ can be divided into $n$ regular non-overlapping patches. MAE first samples a subset of patches and masks the remaining ones, and then acquires the masked patches $x_{m_i}$ , where $i = 1,\dots,\eta$ and $\eta$ is the set of possible masks of length, and visible patches $x_{v_j}$ , where $j = 1,\dots,\kappa$ and $\kappa = n - \eta$ . The mask process of these patches follows a uniform distribution, named the random mask strategy. Moreover, MAE feds these visible patches into encoder (ViT) $f(\cdot)$ , parameterized by $\theta$ , and obtains the encoded visible tokens, then puts these encoded visible tokens with mask tokens into decoder $g(\cdot)$ to perform the MIM task, parameterized by $\xi$ , and obtains the predicted patches. Finally, MAE minimizes the mean squared error (MSE) between the reconstructed and original images in the pixel space. The loss of a single patch prediction is shown as Eq(1). + +$$ +\mathcal {L} _ {\text {s i n g l e}} (i) = \left\| x _ {m _ {i}} - g \left(\xi ; f \left(\theta ; x _ {v _ {1}}, \dots , x _ {v _ {\kappa}}, \mathbf {m a s k}\right)\right) \right\| ^ {2}, \tag {1} +$$ + +For MAE, the total loss function that trains the complete dataset is defined as Eq (2). + +$$ +\mathcal {L} _ {M A E} = \underset {\mathbf {x} \sim X} {\mathbb {E}} \underset {i \sim \eta} {\mathbb {E}} \mathcal {L} _ {\text {s i n g l e}} (i), \tag {2} +$$ + +# 3.1.2 Contrastive learning + +Similarly, for an image $x$ in the dataset $X$ , contrastive learning first generates two views under random data augmentation [5, 18], which are then fed into the network separately. Commonly, the network consists of the encoder and momentum encoder. Then, the two global features $z_{s}$ and $z_{t}$ can be obtained. Let $\mathrm{sim}(z_s,z_t) = z_s^\top z_t / \| z_s\| \| z_t\|$ denotes the dot product between $\ell_2$ normalized $z_{s}$ and $z_{t}$ (i.e. cosine similarity). The InfoNCE loss function [5, 21, 48, 59] for a positive pair of examples is defined as Eq (3), where $\tau$ is a temperature coefficient. It pulls one prediction closer to the self-defined pseudo label while pushing other predictions in a mini-batch apart. + +$$ +\mathcal {L} _ {\text {I n f o N C E}} (x) = - \log \frac {\exp \left(\operatorname* {s i m} \left(\boldsymbol {z} _ {s} , \boldsymbol {z} _ {t}\right) / \tau\right)}{\sum_ {k \neq s} ^ {2 N} \exp \left(\operatorname* {s i m} \left(\boldsymbol {z} _ {s} , \boldsymbol {z} _ {k}\right) / \tau\right)}, \tag {3} +$$ + +Hence, the total loss function that trains the complete dataset is defined as Eq(4). + +$$ +\mathcal {L} _ {\text {c o n t r a s}} = \underset {\mathbf {x} \sim X} {\mathbb {E}} \mathcal {L} _ {\text {I n f o N C E}} (x), \tag {4} +$$ + +It is noted that the advantage of CL is that semantic learning and its self-defined pseudo label are acquired by the updatable and learnable network. + +# 3.2. The analysis of MAE + +According to [46], it is well known that minimizing MSE can be equivalent to maximum likelihood estimation. Therefore, a single patch prediction of MAE can be considered as the mean of a noisy prediction distribution, and the distribution can be modeled as a Gaussian distribution in the classic probabilistic interpretation [43] in Eq (5). + +$$ +p \left(x _ {m _ {i}} \mid x _ {\text {i n p u t s}}; \theta , \xi\right) = \mathcal {N} \left(x _ {m _ {i}}; x _ {p _ {i}}, \sigma_ {\text {n o i s e}} ^ {2} \mathbf {I}\right), \tag {5} +$$ + +$x_{input}$ is the input tokens and equal to $\{x_{v_1},\dots,x_{v_\kappa},\mathbf{mask}\}$ , $x_{p_i}$ is the single predicted patch, $x_{m_i}$ is the pseudo label corresponding to this patch, and $\sigma_{noise}$ is the scale of an independent and identically distributed error term $\epsilon \sim \mathcal{N}(0,\sigma_{noise}^2\mathbf{I})$ . Meanwhile, MSE equals to the negative log likelihood (NLL) loss of the prediction distribution $p(x_{m_i}|x_{input};\theta ,\xi)$ [46]. The network trained by MSE predicts a single patch in fact learns to model $p(x_{m_i}|x_{input})$ + +MAE shows impressive performance on various downstream tasks [22], it reconstructs the masked patches through learning from a large amount of data, and the pseudo labels are the masked image and have no interference from manual annotations. Hence, it is reasonable to consider the $p(x_{m_i})$ is uniform, and the MSE is equivalent to Eq(6), where $C$ is a constant and equal to $p(x_{m_i})$ , and + +the detailed derivation of this section in the Appendix. + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s i n g l e}} (i) \cong - \log \mathcal {N} \left(x _ {m _ {i}}; x _ {p _ {i}}, \sigma_ {\text {n o i s e}} ^ {2} \mathbf {I}\right) \\ + \log \int \mathcal {N} \left(x _ {m _ {i}}; x _ {p _ {i}}, \sigma_ {\text {n o i s e}} ^ {2} \mathbf {I}\right) \cdot p \left(x _ {m _ {i}}\right) d x _ {m _ {i}}, \tag {6} \\ \end{array} +$$ + +For the second part of Eq(6), we utilize monte carlo method to solve $p(x_{m_i})$ and can obtain the Eq(7). + +$$ +\begin{array}{l} \int \mathcal {N} (x _ {m _ {i}}; x _ {p _ {i}}, \sigma_ {n o i s e} ^ {2} \mathbf {I}) \cdot p (x _ {m _ {i}}) d x _ {m _ {i}} \\ \approx \frac {1}{N} \sum_ {b = 1} ^ {N} \mathcal {N} \left(x _ {m _ {i _ {(b)}}}; x _ {p _ {i}}, \sigma_ {\text {n o i s e}} ^ {2} \mathbf {I})\right), \tag {7} \\ \end{array} +$$ + +The monte carlo method treats all pseudo labels in a training batch as random samples from $p(x_{m_i})$ . Hence, for pseudo labels in a training batch $B = \{x_{m_{i(1)}}, x_{m_{i(2)}}, \ldots x_{m_{i(N)}}\}$ , the loss is defined as Eq(8), where $\lambda = 2\sigma_{noise}^2$ is a temperature coefficient. + +$$ +\mathcal {L} _ {\text {s i n g l e}} (i) \cong - \log \frac {\exp \left(- \left| \left| x _ {p _ {i}} - x _ {m _ {i}} \right| \right| ^ {2} / \lambda\right)}{\sum_ {x _ {m _ {i}} ^ {\prime} \in B} \exp \left(- \left| \left| x _ {p _ {i}} - x _ {m _ {i}} ^ {\prime} \right| \right| ^ {2} / \lambda\right)}, \tag {8} +$$ + +According to Eq (3) and Eq (8), it can be observed that the form of the two loss functions is similar. Both of the two loss functions pull together the prediction and self-defined pseudo label in the representation space, and push other predictions in a mini-batch apart. The MAE adopts the Euclidean distance while the contrastive learning measures the cosine similarity in representation space. Actually, MAE regards each masked patch as a category and performs patch-level contrastive learning with unlearnable pixels. It explicitly aligns the prediction with its corresponding masked patch and implicitly distinguishes other patches. + +From the above description, when confronted with arbitrary scenes, MAE and CL will face a severe conflict issue. + +# 3.3. Arbitrary self-supervised learning + +Based on the previous analysis, it can be concluded that a global prior of CL is both unreasonable and conflicts with MIM when facing arbitrary scenarios. To enable learning from arbitrary scenarios, can there be a self-supervised method that harnesses the advantages of CL in semantic learning (learnable self-pseudo labels) and the heterogeneous representation learning of MIM, while avoiding global constraints and conflicts? The answer is affirmative. In this paper, we propose a new framework, named Arbitrary Self-supervised Learning (ASL), to learn from arbitrary scenarios while retaining their advantages. + +Specifically, we adopt the popular MAE method as the basic MIM to ensure the initial heterogeneous representations. Then, inspired by CL using a dual-branch structure + +![](images/4e811026b0a1b77f5ec139adc04c001cc27115c67064b85d413a5a1cea74bcf4.jpg) + +![](images/2669a10c2431ffdfffe1b26f79450cf7d443afdbd44e71c059049ff749faac4e.jpg) +(a) + +![](images/402b4a158a45160a2e1a7ff8820274cb19785fe3f003e35bfc4b01967b8e3cca.jpg) +(b) +Figure 3. The pipeline of our ASL. The visible patches are fed into the network and the dual-branch features can be obtained by our designs. Then the encoded features with mask tokens are fed into the decoder and output the predictions. Finally, we pull together the predictions using L1 loss by one-to-one correspondence. We propose two designs in the network to obtain the dual-branch features: (a) Dropout in the linear embedding. (b) Asymmetric structure at the end of the encoder. In the figure, we adopt the network to represent (a) or (b) design. The olive patch represents the mask token. + +to generate learnable self-pseudo labels, we propose two patch-level designs to generate dual-branch features based on the encoder of MAE while avoiding the global constraints. Subsequently, these dual-branch features are further processed through a decoder module, and CL is performed between the outputs patch by patch to for semantic learning. In contrast to traditional CL, our approach does not entail global data augmentation but instead focuses on patch-level feature augmentation. The specifics of our designs are outlined as follows: + +Dropout in the linear embedding (DLE). According to previous work in NLP, SimCSE [17] has explored the use of dropout in word embedding for feature augmentation. In line with this, we also adopt dropout in the linear embedding of the encoder to perform patch-level augmentation and acquire two encoded features at the end of the encoder. The details of our method are described in Figure 3 (a). + +Asymmetric structure at the end of the encoder (AEE). Some prior work [8, 18, 62] have demonstrated that introducing an asymmetric structure, which essentially is a nonlinear function, enhances the semantic features. With this in mind, we incorporate a Transformer layer [57] after the encoder in our approach to build an asymmetric structure for patch-level feature augmentation. This step produces dual-branch features: one is the feature output by the + +encoder, and the other is the feature that has passed through the Transformer layer. Figure 3 (b) illustrates the details of our method. The design is our default experimental setup. + +The methodology described above pertains to the construction of feature augmentation. Our rationale for utilizing feature augmentation to construct dual-branch features is primarily based on two considerations: 1) Data augmentation in CL is global and not suitable for patch-level learning. 2) The asymmetric structure of CL inherently constitutes an implicit form of feature augmentation, and it is worth exploring whether its success in CL can be extrapolated to the broader domain of self-supervised learning. + +After feature augmentation, we obtain the dual-branch features, $fea_{ori}$ and $fea_{aug}$ , then we feed them into the decoder module to obtain the corresponding predictions, $pred_{ori}$ and $pred_{aug}$ , respectively. The patch-level CL between the outputs does not employ contrastive loss or cosine similarity loss but rather L1 loss. + +The alignment between any pair is defined as Eq(9), where $\mathsf{sg}[\cdot]$ stands for stop gradient. It is noted that our alignment method does not include the alignment of [cls] token according to Section 3.2, which unleashes the potential of the method for training in arbitrary scenarios. + +$$ +\mathcal {L} _ {s} (i) = \left| \left| \operatorname {p r e d} _ {\text {o r i} _ {i}} - s g \left[ \operatorname {p r e d} _ {\text {a u g} _ {i}} \right] \right| \right| + \left| \left| s g \left[ \operatorname {p r e d} _ {\text {o r i} _ {i}} \right] - \operatorname {p r e d} _ {\text {a u g} _ {i}} \right| \right|, \tag {9} +$$ + +The loss function of the complete dataset is defined as Eq(10). + +$$ +\mathcal {L} _ {S E M} = \underset {\mathbf {x} \sim X} {\mathbb {E}} \underset {i \sim \eta} {\mathbb {E}} \mathcal {L} _ {s} (i), \tag {10} +$$ + +# 3.4. Objective function + +The loss function of our ASL consists of the MAE loss and the semantic loss. Hence, the total loss of our ASL is formulated as Eq (11), and each loss coefficient is set to be 1 for equally weighted. + +$$ +\mathcal {L} _ {\text {f i n a l}} = \mathcal {L} _ {\text {M A E}} + \mathcal {L} _ {\text {S E M}}, \tag {11} +$$ + +# 4. Experiments + +# 4.1. Datasets and evaluation metrics + +Pre-training datasets. In the pre-training stage, we select the four popular visual pre-training datasets, MS COCO [40], ImageNet [13], CIFAR10 [31], CIFAR100 [31], to investigate the versatility and transfer ability of the proposed method. For the COCO dataset, its train2017 set contains $\sim 118\mathrm{k}$ images. COCO contains more natural and diverse scenes and is a non-iconic image dataset. To further demonstrate the versatility of our algorithm for training in arbitrary scenarios, we conduct pre-training on both ImageNet and COCO datasets. Given the substantial difference in data volume between ImageNet and COCO, we select a subset of ImageNet, referred to as ImageNet-100, to maintain a balanced dataset size. Moreover, we pre-train on ImageNet-1K to demonstrate the effectiveness of ASL on curated data. Finally, CIFAR-10 and CIFAR-100 datasets are small-scale image datasets, and they all contain 60000 images with $32\times 32$ size that belong to 10 and 100 categories, respectively. We pre-train models on the CIFAR-10 and CIFAR-100 datasets to demonstrate the effectiveness of ASL in the small-scale dataset. + +Evaluation datasets. In this paper, ASL is evaluated by the linear probing and fine-tuning classification task on ImageNet, CIFAR-10, and CIFAR-100 datasets, which are popular image classification datasets. Additionally, for computer vision, object detection, and semantic segmentation are dense prediction tasks since training images of these tasks contain massive objects inside each image. The validation of dense prediction tasks can better reflect the transfer ability of self-supervised models. Therefore, we conduct extensive experiments on COCO [40] and ADE20k [68] datasets to verify the generalization and transfer ability of ASL. COCO is a popular object detection and instance segmentation dataset, which train2017 contains about $118k$ images and test on the val2017 contains $5k$ images. The challenging dataset contains human annotations for 80 classes. Moreover, ADE20K is also a challenging semantic segmentation dataset, and it contains about $25k$ training images and $2k$ validation images with 150 categories. + +
MethodArch.FLOPsEpochsLPAPb
Pre-training on COCO dataset:
MAE [22]ViT-B1×80046.8%49.3%
CMAE [27]ViT-B~ 3×80046.6%49.0%
ccMIM [67]ViT-B~ 3.3×80046.7%49.0%
SIM [54]ViT-B~ 2.8×80046.0%48.7%
iBOT [69]ViT-B~ 6×80045.2%48.6%
ASLViT-B~ 1.5×80048.6%50.3%
+ +Table 1. Evaluation of pre-trained models on the COCO dataset. We report Linear Probing (LP) top-1 accuracy on the ImageNet-1K val set and bbox mAP on COCO dataset. + +
MethodArch.Param.EpochsLPFT
Pre-training on ImageNet-100 + COCO:
MAE [22]ViT-B8680074.5%90.9%
iBOT[69]ViT-B8680071.8%88.5%
ccMIM [67]ViT-B8680072.3%88.7%
ASLViT-B8680079.6%92.4%
ASLViT-B86400085.9%94.2%
+ +Table 2. ImageNet-100 Top-1 accuracy of different methods under linear probing (LP) and fine-tuning (FT) setting to evaluate the performance of the model pre-trained on both ImageNet-100 and COCO. We report top-1 accuracy on the ImageNet-100 val set. + +The valuation metrics, pre-training settings, and the experimental settings of downstream tasks in the Appendix. + +# 4.2. Pre-training on COCO dataset + +In order to validate the adaptability of our approach for training in arbitrary scenarios, we initially conduct pretraining on the COCO dataset and subsequently conduct a fair comparison with other methods under the same experimental settings. In this comparison, we evaluate the ImageNet linear evaluation and the COCO detection results. Linear evaluation, which involves using frozen features, provides a more comprehensive assessment of a model's representation. Given that the pre-training experiments are conducted on the COCO dataset, using human-labeled data from COCO to test the learned representations is reasonable. As shown in Table 1, we observe that methods combining MAE with contrastive learning, such as CMAE, ccMIM, and SIM, exhibit weaker performance when pretrained on the COCO dataset, even though they achieved SOTA performance on ImageNet. This observation suggests that existing SOTA methods are not well-suited for training in arbitrary scenarios. Furthermore, our approach demonstrates superior performance after pre-training on the COCO dataset, surpassing the performance of MAE. Generally, introducing a dual-branch structure similar to contrastive learning typically results in a significant increase in computational overhead, often exceeding threefold. Diverging from the emphasis on the encoder in hybrid methods, our AEE design prioritizes the decoder, significantly reducing + +
MethodPre-train IterationsDatasetObject detectionInstance segmentation
\(AP^b\)\(AP_{50}^b\)\(AP_{75}^b\)\(AP^m\)\(AP_{50}^m\)\(AP_{75}^m\)
Supervised [9]~ 94kImageNet-1K47.9%--42.9%--
MoCo v3 [8]~ 187kImageNet-1K47.9%--42.7%--
BEiT [1]~ 249kImageNet-1K + DALLE49.8%--44.4%--
MAE [22]~ 499kImageNet-1K50.4%70.8%55.7%44.9%68.3%48.9%
CAE [9]~ 499kImageNet-1K50.0%70.9%54.8%44.0%67.9%47.6%
R-MAE [45]~ 115kCOCO50.6%--45.0%--
R-MAE [45]~ 250kCOCO50.8%--45.2%--
ccMIM [67]~ 249kImageNet-1K50.3%71.2%55.0%44.5%68.4%47.9%
ASL~ 47kImageNet-100 + COCO50.7%71.1%55.6%45.0%68.3%49.0%
ASL~ 236kImageNet-100 + COCO51.5%71.9%57.2%45.7%69.2%50.0%
+ +Table 3. Results of object detection and instance segmentation on COCO. The architecture of various methods is ViT-B, and we report the bounding box AP and mask AP on COCO val2017. + +
MethodDatasetPre-train IterationsADE mIoU
Supervised [22]ImageNet-1K~ 93k47.4%
SplitMask [16]ADE20K~ 128k45.7%
MoCo v3 [8]ImageNet-1K~ 187k47.3%
BEiT [22]ImageNet-1K+DALLE~ 249k47.1%
R-MAE [45]COCO~ 115k46.8%
R-MAE [45]COCO~ 250k47.0%
ccMIM [67]ImageNet-1K~ 249k47.7%
MAE [22]ImageNet-1K~ 499k48.1%
ASLImageNet-100 + COCO~ 236k48.4%
+ +Table 4. Results of semantic segmentation on ADE20K. We report results measured by mIoU. + +ing computational costs in comparison. Meanwhile, due to the fixed-size and lightweight design of the decoder, our design incurs decreasing computational overhead as the model scale increases. This highlights the potential of our method for pre-training in non-iconic scenarios. + +# 4.3. Pre-training on both COCO and ImageNet + +To further demonstrate the adaptability of ASL across diverse scenarios, we conduct the pre-training experiments on both ImageNet-100 and COCO datasets. We assess the performance of the model on benchmark tasks: image classification, object detection, and semantic segmentation. The ASL models are primarily trained with 800 epochs and 4000 epochs, with the number of training iterations for the 4000-epoch training similar to that of the ImageNet-1K iterations for the 800-epoch training. + +# 4.3.1 Image classification + +Since our training data comprises the ImageNet-100 dataset, we conducted testing for classification performance through both linear probing and fine-tuning on ImageNet-100. The experimental results are presented in Table 2. The + +results of pre-training observed from the table align with those of pre-training solely on COCO: ccMIM exhibits relatively weak performance in the context of the pre-training. This can be attributed to conflicts between CL and MAE when dealing with arbitrary scenarios. In contrast, our method successfully mitigates such conflicts and produces more robust representations in arbitrary scenarios. + +# 4.3.2 Object detection and instance segmentation + +To verify the transfer ability of ASL, we test it on object detection and instance segmentation with COCO [40] dataset. Here we report Box AP and mask AP on the validation set. According to the benchmark validation [22, 33], we choose the Mask R-CNN [20] as the test framework. Due to variations in the sizes of the pre-training datasets used, we opt to avoid epoch-based comparisons in this scenario. We convert epochs to a fixed batch size of 4096 iterations for a fair comparison. Table 3 shows the results of the learned representation by supervised method and different self-supervised methods. It shows our ASL, with approximately 236k iterations (4000 epochs), achieves impressive performance and outperforms 1600-epoch MAE, 1600-epoch CAE and 800-epoch ccMIM pre-trained on ImageNet. ASL surpasses various supervised and self-supervised models pre-trained on ImageNet and COCO, indicating its success in arbitrary scenarios. + +# 4.3.3 Semantic segmentation + +To further evaluate the generalization of our method, we conduct the semantic segmentation experiments on ADE20K dataset [68]. As shown in Table 4, ASL achieves the $48.4\%$ mIoU. It also outperforms various supervised and self-supervised models pre-trained on ImageNet. + +Overall, our ASL shows impressive performance on various mainstream visual benchmark tasks when pre-training on both COCO and ImageNet-100, which demonstrates + +
MethodLPFTAPbAPmmIoU
MAE67.8%83.6%50.4%44.9%48.1%
ASL69.2%84.2%51.3%45.7%49.2%
+ +Table 5. The ImageNet-1K pre-training results. We report ImageNet classification, COCO detection, and instance segmentation, as well as ADE20k semantic segmentation results. + +
MethodCIFAR-10CIFAR-100
Linear Prob.FinetuningLinear Prob.Finetuning
Supervised [26]-91.3%-64.1%
DINO [3]89.0%94.4%65.8%76.3%
MAE [22]87.3%95.9%54.0%81.1%
ASL91.4%96.5%60.3%82.1%
+ +Table 6. Linear probing and fine-tuning classification results on small-scale datasets. We adopt ViT-S as the base model. + +
IDDLEAEEalignmentFLOPsTop-1APb
(a)×××1×46.8%49.3%
(b)✓××~ 2×46.5%49.1%
(c)✓×✓~ 2×48.5%50.4%
(d)×✓×~ 1.5×46.6%49.2%
(e)×✓✓~ 1.5×48.6%50.3%
+ +ASL has the ability to learn general visual representations from arbitrary scenarios. + +# 4.4. Pre-training on ImageNet-1K + +Our method is designed for learning visual representations from uncurated arbitrary scenarios, and it can also yield stronger representations when applied to meticulously curated datasets. To demonstrate this point, we conduct experiments on the curated ImageNet dataset. As depicted in the experimental results shown in Table 5, ASL exhibits superior performance across various downstream tasks compared to MAE, highlighting the versatility of our approach. + +# 4.5. Small-scale data pre-training + +As shown in Table 6, the results of our method on these small datasets all show impressive performance. With the same training settings for all experiments, ASL consistently improves MAE by at least $4.1\%$ for linear probing. However, MIM still lags behind CL in linear evaluation on CIFAR100 and its advantage lies in the fine-tuning superior performance. Our fine-tuning results also reveal it. Totally, these results suggest that ASL is effective and can learn more semantic features than MAE in small-scale datasets. + +# 4.6. Ablation study + +The effect of feature augmentation. We propose two designs to obtain a dual-branch structure by performing feature augmentation. As shown in Table 7, (a) is the results of MAE as the baseline when pre-trained on COCO. Table + +7 (b) indicates the results of the DLE design are slightly inferior to the baseline. However, the results outperform the baseline MAE when aligning the predictions according to (c). Similarly, the results for (d) using AEE are not as favorable as MAE, while the results for (e) incorporating alignment operations outperform MAE. These experimental findings collectively indicate that creating learnable pseudo labels based on feature augmentation during the pretraining phase can improve the representation of pre-trained models. Due to the relatively comparable performance of (e) and (c), with (e) incurring significantly lower computational costs, we opt for (e) as the final experimental design. + +The effect of alignment loss. We compare two loss functions: one commonly used in contrastive learning, which is cosine similarity loss, and the other is L1 loss. The following is a comparison between cosine similarity loss and L1 loss: + +Table 7. Ablations for ASL: Effect of feature augmentation. We conduct pre-training experiments on COCO dataset and report linear evaluation on ImageNet-1K and detection results on COCO. + +
MAEcosine similarity lossL1 loss[cls] align
46.8%47.6%48.6%45.6%
+ +From the above, experimental results indicate that the L1 loss facilitates the model in learning better representations compared to cosine similarity loss. Besides, we conduct an additional experiment involving alignment with the [cls] token on top of the L1 loss. The results indicate that introducing global contrastive learning during pre-training on the COCO dataset has a detrimental effect on performance. + +# 5. Conclusions + +This paper points out the inability of current hybrid SOTA methods to extend training to arbitrary scenarios due to the conflict between MAE and CL. It then theoretically identifies the root of the issue as MAE being patch-level CL, conflicting with global CL. To address this conflict, we introduce a novel self-supervised learning method ASL to learn visual representations from arbitrary scenarios. Specifically, this paper abandons the global-level consistency of CL and proposes explicit patch-level learning. ASL adopts patch-level feature augmentation to generate dual-branch features by the encoder of MAE, then these features perform patch-level learning through a decoder module. Finally, ASL can learn heterogeneous representations while avoiding the conflicts. Our method demonstrates robustness and versatility in multiple pre-training datasets and downstream visual tasks. We expect that our study will draw the attention of the community to large-scale pre-train from arbitrary scenarios and contribute to the development of visual foundation models. + +Acknowledgement. This work was supported by National Key R&D Program of China under Grant No. 2021YFE0205700 and National Natural Science Foundation of China (No.62276260, No.62176254, No.62076235). + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 2, 3, 7 +[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, pages 9912-9924, 2020. 1, 3 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv: Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 8 +[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey K Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703, 2020. 1, 2, 3 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 1, 3, 4 +[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E. Hinton. Big self-supervised models are strong semi-supervised learners. In Advances in Neural Information Processing Systems, pages 22243-22255, 2020. +[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3 +[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 3, 5, 7 +[9] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.7 +[10] Zhiyang Chen, Yousong Zhu, Zhaowen Li, Fan Yang, Wei Li, Haixin Wang, Chaoyang Zhao, Liwei Wu, Rui Zhao, Jinqiao Wang, et al. Obj2seq: Formatting objects as sequences with class prompt for visual tasks. Advances in Neural Information Processing Systems, 35:2494-2506, 2022. 3 +[11] Zhiyang Chen, Yousong Zhu, Fan Yang, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. The devil is in details: Delving into lite ffn design for vision transformers. Available at SSRN 4299967, 2023. 3 +[12] Zhiyang Chen, Yousong Zhu, Yufei Zhan, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. Mitigating hallucination in visual language models with visual supervision. arXiv preprint arXiv:2311.16479, 2023. 3 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 6 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the + +2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, 2018. 3 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3 +[16] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 7 +[17] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 5 +[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 1, 3, 4, 5 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In The IEEE International Conference on Computer Vision, 2017. 3, 7 +[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1, 3, 4 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 2, 3, 4, 6, 7, 8 +[23] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and Joao Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 3 +[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In European Conference on Computer Vision, pages 123-143. Springer, 2022. 3 +[25] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. International Conference on Learning Representations, 2019. 1, 3 +[26] Tianyu Hua, Yonglong Tian, Sucheng Ren, Hang Zhao, and Leonid Signal. Self-supervision through random seg + +ments with autoregressive coding (rand sac). arXiv preprint arXiv:2203.12054, 2022.8 +[27] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. arXiv preprint arXiv:2207.13532, 2022. 2, 3, 6 +[28] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 1 +[29] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1 +[30] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3 +[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[32] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. arXiv preprint arXiv:2210.02077, 2022. 2, 3 +[33] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3, 7 +[34] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 23390-23400, 2023. 2 +[35] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Mst: Masked self-supervised transformer for visual representation. In Advances in Neural Information Processing Systems, 2021. 1, 2, 3 +[36] Zhaowen Li, Xu Zhao, Chaoyang Zhao, Ming Tang, and Jinqiao Wang. Transfering low-frequency features for domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 3 +[37] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 14627-14636, 2022. 2, 3 +[38] Zhaowen Li, Xu Zhao, Peigeng Ding, Zongxing Gao, Yuting Yang, Ming Tang, and Jinqiao Wang. Freconv: Frequency branch-and-integration convolutional networks. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 258-263. IEEE, 2023. 3 +[39] Zhaowen Li, Yousong Zhu, Zhiyang Chen, Wei Li, Chaoyang Zhao, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Efficient masked autoencoders with self-consistency. arXiv preprint arXiv:2302.14431, 2023. 3 +[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence + +Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 6, 7 +[41] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020. 2 +[42] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3 +[43] Peter McCullagh and John A Nelder. Generalized linear models. Routledge, 2019. 4 +[44] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations. arXiv preprint arXiv:2210.16870, 2022. 2, 3 +[45] Duy-Kien Nguyen, Vaibhav Aggarwal, Yanghao Li, Martin R Oswald, Alexander Kirillov, Cees GM Snoek, and Xinlei Chen. R-mae: Regions meet masked autoencoders. arXiv preprint arXiv:2306.05411, 2023. 7 +[46] David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 IEEE international conference on neural networks (ICNN'94), pages 55-60. IEEE, 1994. 4 +[47] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proceedings of the European Conference on Computer Vision, 2016. 3 +[48] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3, 4 +[49] OpenAI. Introducing chatgpt, 2022. 3 +[50] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3 +[51] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3 +[52] Hao Quan, Xingyu Li, Weixing Chen, Mingchen Zou, Ruijie Yang, Tingting Zheng, Ruiqun Qi, Xinghua Gao, and Xiaoyu Cui. Global contrast masked autoencoders are powerful pathological representation learners. arXiv preprint arXiv:2205.09048, 2022. 2, 3 +[53] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, 2015. 3 +[54] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. Proceedings of the IEEE conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6 + +[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1 +[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3 +[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 5998-6008, 2017. 5 +[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. arXiv preprint arXiv:2011.09157, 2020.3 +[59] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 3, 4 +[60] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision, pages 418-434, 2018. 3 +[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. arXiv preprint arXiv:2011.10043, 2020. 3 +[62] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 5 +[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. 2, 3 +[64] Kun Yi, Yixiao Ge, Xiaotong Li, Shusheng Yang, Dian Li, Jianping Wu, Ying Shan, and Xiaohu Qie. Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616, 2022. 2, 3 +[65] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344, 2022. 2, 3 +[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Proceedings of the European Conference on Computer Vision, 2016. 3 +[67] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In The Eleventh International Conference on Learning Representations. 1, 2, 3, 6, 7 +[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6, 7 +[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training + +with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 6 \ No newline at end of file diff --git a/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/images.zip b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c911b11d75c0e85b7f185ce0ab43a85a23156595 --- /dev/null +++ b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6991f0a5cd63df2b03d8a5c5c91b53881c059b8c0eecc899bf56ff5632e2ffdc +size 447619 diff --git a/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/layout.json b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7aa1c208a5fd766a31439ed1f8b19034ee0fc6f9 --- /dev/null +++ b/2024/Self-Supervised Representation Learning from Arbitrary Scenarios/layout.json @@ -0,0 +1,8641 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 504, + 121 + ], + "type": "text", + "content": "Self-Supervised Representation Learning from Arbitrary Scenarios" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "spans": [ + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": "Zhaowen Li " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Yousong Zhu " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Zhiyang Chen " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Zongxin Gao " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Rui Zhao " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Chaoyang Zhao " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Ming Tang " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Jinqiao Wang " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{1,2,3,4,5\\boxtimes}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " \nFoundation Model Research Center, Institute of Automation, Chinese Academy of Science " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Peng Cheng Laboratory " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Wuhan AI Research " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Objectye Inc. " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Qing Yuan Research Institute, Shanghai Jiao Tong University " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " Independent Researcher " + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 73, + 141, + 516, + 242 + ], + "type": "text", + "content": " {zhaowen.li, yousong.zhu, jqwang}@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 293, + 289, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 606 + ], + "type": "text", + "content": "Current self-supervised methods can primarily be categorized into contrastive learning and masked image modeling. Extensive studies have demonstrated that combining these two approaches can achieve state-of-the-art performance. However, these methods essentially reinforce the global consistency of contrastive learning without taking into account the conflicts between these two approaches, which hinders their generalizability to arbitrary scenarios. In this paper, we theoretically prove that MAE serves as a patch-level contrastive learning, where each patch within an image is considered as a distinct category. This presents a significant conflict with global-level contrastive learning, which treats all patches in an image as an identical category. To address this conflict, this work abandons the nongenerizable global-level constraints and proposes explicit patch-level contrastive learning as a solution. Specifically, this work employs the encoder of MAE to generate dual-branch features, which then perform patch-level learning through a decoder. In contrast to global-level data augmentation in contrastive learning, our approach leverages patch-level feature augmentation to mitigate interference from global-level learning. Consequently, our approach can learn heterogeneous representations from a single image while avoiding the conflicts encountered by previous methods. Massive experiments affirm the potential of our method for learning from arbitrary scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 127, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "content": "Currently, self-supervised learning (SSL) in computer vision [1, 2, 4-7, 18, 21, 22, 35, 59] is capable of adopting self-defined pseudo labels as supervision [28, 29] and holds the promise in leveraging large amounts of unlabeled data" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 287, + 525, + 446 + ], + "blocks": [ + { + "bbox": [ + 316, + 287, + 525, + 446 + ], + "lines": [ + { + "bbox": [ + 316, + 287, + 525, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 525, + 446 + ], + "type": "image", + "image_path": "bafcc348137408f74f60e64216fc5772dc8bdf0e5756deb50592b67a4ef031df.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 453, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 564 + ], + "type": "text", + "content": "Figure 1. Hybrid methods further enhance the global consistency of CL. We measures representational similarity among tokens for some popular self-supervised pre-trained models. Among them, ccMIM [67] and SIM [54] employ MAE as the MIM baseline method while iBOT [69] adopts DINO [3] as the CL baseline method for hybrid approach. It indicates that hybrid methods, while increasing linear evaluation scores, also enhance the representational similarity among tokens. This suggests that hybrid methods fundamentally further strengthen the global semantic consistency of CL. All of the models adopt ViT-B." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 581, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 715 + ], + "type": "text", + "content": "to build the foundational models. SSL methods focus on designing different pretext tasks, and there are two mainstream approaches for learning visual features in the community. One of the most promising directions among them is contrastive learning (CL) [2, 5-7, 18, 21, 25, 48]. It assumes that different views of a single image are of the same instance. Its objective is to learn a global-level feature representation that discriminates among images. CL leverages trainable networks to generate semantic pseudo labels [55], thereby enhancing the model's capability to extract semantic information. On the other hand, the masked image mod" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 243, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 243, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 243, + 714 + ], + "type": "text", + "content": "Corresponding author: Yousong Zhu and Jinqiao Wang." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22967" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 74, + 176, + 196 + ], + "blocks": [ + { + "bbox": [ + 54, + 74, + 176, + 196 + ], + "lines": [ + { + "bbox": [ + 54, + 74, + 176, + 196 + ], + "spans": [ + { + "bbox": [ + 54, + 74, + 176, + 196 + ], + "type": "image", + "image_path": "415dcbcc76660511c360d1e83dafd5030e5f27229ed021eaa846308f6f969e66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 237 + ], + "type": "text", + "content": "Figure 2. Visualization of patches in an image from COCO dataset. The image is partitioned into multiple patches, and different patches from the image may represent different semantics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 77, + 284, + 191 + ], + "blocks": [ + { + "bbox": [ + 184, + 77, + 284, + 191 + ], + "lines": [ + { + "bbox": [ + 184, + 77, + 284, + 191 + ], + "spans": [ + { + "bbox": [ + 184, + 77, + 284, + 191 + ], + "type": "image", + "image_path": "8b43cbeb1fc5d970d9510bcafa8b43228a83d08e40799b287b80fd8b8d49e885.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 256, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 286, + 365 + ], + "type": "text", + "content": "eling (MIM) [1, 4, 15, 22, 35, 63, 69] has become the focus in the community. MIM is a pretext task that involves masking some patches on the input image and predicting the original information of the masked patches based on their context. In these methods, BEiT [1] predicts discrete tokens created by the pre-trained model of VQVAE [56]. In MAE [22], the pseudo labels are the given pixels. Among these MIM methods, MAE shows impressive transfer performance on various dense prediction tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 367, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 713 + ], + "type": "text", + "content": "The learning methodologies of CL and MIM are so distinct that a fundamental question arises: can their concurrent learning mutually enhance each other? In fact, extensive work [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] shows hybrid methods of CL and MIM achieve state-of-the-art (SOTA) performance on benchmark ImageNet [13] dataset. In our study, we choose some pre-trained models trained using popular self-supervised methods and measure their representational similarity for tokens as well as linear probing results, as depicted in Figure 1. Specifically, ccMIM [67] and SIM [54] employ MAE as the MIM baseline, while iBOT [69] adopts DINO [3] as the CL baseline for the hybrid approach. The figure illustrates that these hybrid methods have achieved better linear probing results compared to their respective baseline methods. However, they have further reinforced representational similarity for tokens, resulting in homogeneous token representations. This also implies that the hybrid methods essentially enhance the global consistency of CL, and the performance of SOTA methods heavily relies on the gains achieved through CL. Consequently, the success of these methods heavily relies on the single-centric-object data [37, 41] such as those in ImageNet dataset, and they cannot be generalized to non-iconic scenarios, making them unsuitable for training in arbitrary scenarios - regardless of whether they come from single-centric-object scenarios or non-iconic ones. Also, a recent study [34] indicates that the hybrid approach of MAE and CL exhibits inferior performance compared to the baseline in large-scale arbitrary scenarios. This goes against the pri" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "mary promise of self-supervised learning, which aimed to build foundation models using a large amount of data for broader applicability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 110, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 300 + ], + "type": "text", + "content": "Actually, the inability of SOTA methods to train in arbitrary scenarios can be attributed to the conflict between CL and MIM. In this paper, we first theoretically prove that MAE is a patch-level CL, treating each patch within an image as a separate category. This conflicts with global-level CL, where all patches from the image are treated as an identical category. As illustrated in Figure 2, a non-iconic image from the COCO [40] dataset is partitioned into multiple non-overlapping patches according to ViT [15], and each patch may represent a different semantics. MAE treats each object/patch in these patches as a distinct category, whereas CL considers all patches as an identical category. Therefore, when these two methods are combined for training on non-iconic data, significant conflicts and confusion arise, rendering the hybrid approach incapable of learning effective representations for non-iconic data and arbitrary scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 302, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 546, + 553 + ], + "type": "text", + "content": "To alleviate the problems, in this paper, we propose a self-supervised framework, called Arbitrary Self-supervised Learning (ASL), to learn representation from arbitrary scenarios. The framework leverages the principles of CL to enhance the patch-level semantic consistency of the model while avoiding the conflict between CL and MAE. Specifically, our approach involves designing a patch-level feature augmentation to effectively circumvent the limitation of global-level data augmentation in CL. According to the patch-level augmentation, ASL employs the encoder of MAE to generate dual-branch features, which then perform patch-level contrastive learning through a decoder module. We also retain the original MIM pretext task to ensure the rationality of the model. It is noteworthy that our approach is proposed for arbitrary scenarios, primarily experimenting on the non-iconic COCO dataset that closely represents natural scenes. Simultaneously, we conduct the pre-training experiments on both ImageNet and COCO. The results indicate that ASL is the sole method of attaining superior performance in arbitrary pre-training, further underscoring the generalizability and robustness of our approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 555, + 504, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 504, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 504, + 566 + ], + "type": "text", + "content": "Overall, we make the following contributions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 567, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 567, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 545, + 626 + ], + "type": "text", + "content": "- We theoretically prove that MAE is a patch-level contrastive learning, which is in conflict with global contrastive learning, thereby rendering current hybrid methods incapable of achieving generality in arbitrary scenarios pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 628, + 544, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 544, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 544, + 662 + ], + "type": "text", + "content": "- We propose a self-supervised framework, named Arbitrary Self-supervised Learning (ASL), to mitigate conflicts and allow the model to train in arbitrary scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 663, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 545, + 711 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness and transfer ability of our framework. Specifically, the models pre-trained with ASL achieve SOTA performance in arbitrary scenarios." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22968" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": "It is known that the promise of self-supervised learning [1, 2, 4-7, 15, 18, 21, 22, 25, 30, 35, 37, 47, 48, 50, 51, 59, 63, 66, 69] in computer vision is to establish a visual foundation model by leveraging large dataset, just like ChatGPT [49] in natural language processing. These representations of the foundation model can be transferred to various downstream tasks [10, 12, 13, 19, 20, 33, 36, 38, 39, 53, 60]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 180, + 180, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 180, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 180, + 192 + ], + "type": "text", + "content": "2.1. Instance discrimination" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 198, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 289, + 450 + ], + "type": "text", + "content": "The basic principle of contrastive learning/instance discrimination is that different views of an image are still the same category. As a significant representative of these approaches, MoCo [21] improves the training of contrastive learning methods by storing representations from a momentum encoder instead of the trained network. Then, SimCLR [5] shows that the memory bank can be entirely replaced with the elements from the same batch when the batch size is set large enough. Furthermore, BYOL [18] proposes an asymmetric structure and directly bootstraps the representations by attracting different features from the same instance and shows that contrastive learning without negative samples can also learn excellent visual representations. Also, the asymmetric structure is often directly adopted by subsequent work. Some approaches [23, 24, 58, 61] attempt to transfer global-level prior to pixel-level or region-level learning and acquire visual representations for dense tasks. Also, UniVIP [37] proposes a three-level pre-training task and learns versatile representations from non-iconic images. Recently, MoCo v3 [8] and DINO [3] replace CNNs with ViT [15] and achieve superior performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 456, + 185, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 185, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 185, + 468 + ], + "type": "text", + "content": "2.2. Masked image modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "Motivated by MLM of BERT [14] in NLP, the pioneering work iGPT [4] operates on sequences of pixels and predicts unknown pixels. Benefiting from the proposal of Vision Transformer [11, 15, 42], MIM achieves performance comparable to CL. MST [35] is the first to introduce MIM into the siamese structure and proposes the attention-guided mask strategy. Then, iBOT [69] also adopts the siamese structure and obtains impressive performance. However, they have the prior/assumption of image semantic consistency. Besides, BEiT [1] proposes to predict discrete tokens for masked image modeling, yet its transfer performance mainly depends on the quality of the pre-trained model VQVAE [56]. Following the MIM design of ViT, SimMIM [63] predicts pixels of masked patches to perform the MIM task. Different from the above methods of performing MIM task in the encoder, MAE [22] designs a special encoder-decoder structure to make the decoder perform the MIM task and abandon the MIM-performing module in downstream visual tasks. Among these methods, MAE shows excellent performance in various downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 406, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 406, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 406, + 84 + ], + "type": "text", + "content": "2.3. Hybrid methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 90, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 365 + ], + "type": "text", + "content": "An intuitive idea is to combine the above approaches [27, 32, 35, 44, 50, 52, 54, 64, 65, 67, 69] to achieve better performance. MST [35] firstly adopts MIM to a contrastive learning framework [3] and restores the pixels of patches, then iBOT [69] predicts the semantic tokens. Moreover, DINO v2 [50] achieves SOTA performance based on the iBOT through several engineering optimizations, with the most significant being data curation. It reduces the original dataset of 1.3 billion images to 142 million images, primarily using the ImageNet dataset as the main set of query images. However, it is not suitable for training in arbitrary scenarios. More methods [27, 32, 44, 52, 54, 64, 65, 67] are based on the MAE framework rather than the MST/iBOT architecture, as it conserves approximately 5 times the computational resources when compared to the MST/iBOT framework under the ViT-B setting. Additionally, as the model scales up, the proportion of saved computational resources increases further. However, current SOTA methods primarily benefit from contrastive learning, thus demonstrating SOTA performance on the benchmark single-centric-object ImageNet dataset. When extended to arbitrary or natural scenes, a pronounced conflict between CL and MAE arises." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 374, + 373, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 373, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 373, + 388 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 394, + 385, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 385, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 385, + 407 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 413, + 424, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 424, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 424, + 423 + ], + "type": "text", + "content": "3.1.1 Masked autoencoders" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": "As noted in the prior work [22], for a dataset " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": " without manual annotations, an image " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": " can be divided into " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": " regular non-overlapping patches. MAE first samples a subset of patches and masks the remaining ones, and then acquires the masked patches " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "x_{m_i}" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "i = 1,\\dots,\\eta" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": " is the set of possible masks of length, and visible patches " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "x_{v_j}" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "j = 1,\\dots,\\kappa" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\kappa = n - \\eta" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": ". The mask process of these patches follows a uniform distribution, named the random mask strategy. Moreover, MAE feds these visible patches into encoder (ViT) " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": ", and obtains the encoded visible tokens, then puts these encoded visible tokens with mask tokens into decoder " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": " to perform the MIM task, parameterized by " + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": ", and obtains the predicted patches. Finally, MAE minimizes the mean squared error (MSE) between the reconstructed and original images in the pixel space. The loss of a single patch prediction is shown as Eq(1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 634, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 634, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 312, + 634, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i n g l e}} (i) = \\left\\| x _ {m _ {i}} - g \\left(\\xi ; f \\left(\\theta ; x _ {v _ {1}}, \\dots , x _ {v _ {\\kappa}}, \\mathbf {m a s k}\\right)\\right) \\right\\| ^ {2}, \\tag {1}", + "image_path": "062423b4b1e6509440383b2e1e2ed687f8a63df6281c05deea257666eef06cd1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 666, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 689 + ], + "type": "text", + "content": "For MAE, the total loss function that trains the complete dataset is defined as Eq (2)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 696, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 696, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 364, + 696, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M A E} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\underset {i \\sim \\eta} {\\mathbb {E}} \\mathcal {L} _ {\\text {s i n g l e}} (i), \\tag {2}", + "image_path": "01ae2e309fa8b65316e56e7d51e4b945ada651ad459c4c58719d0e1a56608132.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22969" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 161, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 161, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 161, + 84 + ], + "type": "text", + "content": "3.1.2 Contrastive learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": "Similarly, for an image " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " in the dataset " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": ", contrastive learning first generates two views under random data augmentation [5, 18], which are then fed into the network separately. Commonly, the network consists of the encoder and momentum encoder. Then, the two global features " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " can be obtained. Let " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(z_s,z_t) = z_s^\\top z_t / \\| z_s\\| \\| z_t\\|" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " denotes the dot product between " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " normalized " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " (i.e. cosine similarity). The InfoNCE loss function [5, 21, 48, 59] for a positive pair of examples is defined as Eq (3), where " + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 85, + 287, + 229 + ], + "type": "text", + "content": " is a temperature coefficient. It pulls one prediction closer to the self-defined pseudo label while pushing other predictions in a mini-batch apart." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 239, + 287, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 239, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 52, + 239, + 287, + 273 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {I n f o N C E}} (x) = - \\log \\frac {\\exp \\left(\\operatorname* {s i m} \\left(\\boldsymbol {z} _ {s} , \\boldsymbol {z} _ {t}\\right) / \\tau\\right)}{\\sum_ {k \\neq s} ^ {2 N} \\exp \\left(\\operatorname* {s i m} \\left(\\boldsymbol {z} _ {s} , \\boldsymbol {z} _ {k}\\right) / \\tau\\right)}, \\tag {3}", + "image_path": "0954c410242e0e7ec543c9b1290575779137237a5de1f3715143f9b32f304a8b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": "Hence, the total loss function that trains the complete dataset is defined as Eq(4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 319, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 319, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 100, + 319, + 287, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t r a s}} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\mathcal {L} _ {\\text {I n f o N C E}} (x), \\tag {4}", + "image_path": "f90b0e0a13afbe516ea69eb58e86505d36d7d1b39a39f1659f4c61ba7f9af42c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 348, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 384 + ], + "type": "text", + "content": "It is noted that the advantage of CL is that semantic learning and its self-defined pseudo label are acquired by the updatable and learnable network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 394, + 168, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 168, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 168, + 406 + ], + "type": "text", + "content": "3.2. The analysis of MAE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "type": "text", + "content": "According to [46], it is well known that minimizing MSE can be equivalent to maximum likelihood estimation. Therefore, a single patch prediction of MAE can be considered as the mean of a noisy prediction distribution, and the distribution can be modeled as a Gaussian distribution in the classic probabilistic interpretation [43] in Eq (5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 496, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 496, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 74, + 496, + 287, + 510 + ], + "type": "interline_equation", + "content": "p \\left(x _ {m _ {i}} \\mid x _ {\\text {i n p u t s}}; \\theta , \\xi\\right) = \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right), \\tag {5}", + "image_path": "e4a8ea06079d9ca531efd84dd82449edc3347ed7efaac0781bcc702842db6d5f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "x_{input}" + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "content": " is the input tokens and equal to " + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\{x_{v_1},\\dots,x_{v_\\kappa},\\mathbf{mask}\\}" + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "x_{p_i}" + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "content": " is the single predicted patch, " + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "x_{m_i}" + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "content": " is the pseudo label corresponding to this patch, and " + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_{noise}" + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "content": " is the scale of an independent and identically distributed error term " + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\sigma_{noise}^2\\mathbf{I})" + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "content": ". Meanwhile, MSE equals to the negative log likelihood (NLL) loss of the prediction distribution " + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "p(x_{m_i}|x_{input};\\theta ,\\xi)" + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "text", + "content": " [46]. The network trained by MSE predicts a single patch in fact learns to model " + }, + { + "bbox": [ + 47, + 521, + 287, + 629 + ], + "type": "inline_equation", + "content": "p(x_{m_i}|x_{input})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "MAE shows impressive performance on various downstream tasks [22], it reconstructs the masked patches through learning from a large amount of data, and the pseudo labels are the masked image and have no interference from manual annotations. Hence, it is reasonable to consider the " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "p(x_{m_i})" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " is uniform, and the MSE is equivalent to Eq(6), where " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " is a constant and equal to " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "p(x_{m_i})" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 524, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 524, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 524, + 84 + ], + "type": "text", + "content": "the detailed derivation of this section in the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 90, + 544, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 90, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 312, + 90, + 544, + 143 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s i n g l e}} (i) \\cong - \\log \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right) \\\\ + \\log \\int \\mathcal {N} \\left(x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I}\\right) \\cdot p \\left(x _ {m _ {i}}\\right) d x _ {m _ {i}}, \\tag {6} \\\\ \\end{array}", + "image_path": "fd977adc6206f52f0cadf157e1ac1b25d6a73aa8d5e1672c33084340529a7129.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 149, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 149, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 149, + 545, + 175 + ], + "type": "text", + "content": "For the second part of Eq(6), we utilize monte carlo method to solve " + }, + { + "bbox": [ + 306, + 149, + 545, + 175 + ], + "type": "inline_equation", + "content": "p(x_{m_i})" + }, + { + "bbox": [ + 306, + 149, + 545, + 175 + ], + "type": "text", + "content": " and can obtain the Eq(7)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 181, + 544, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 181, + 544, + 240 + ], + "spans": [ + { + "bbox": [ + 347, + 181, + 544, + 240 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\int \\mathcal {N} (x _ {m _ {i}}; x _ {p _ {i}}, \\sigma_ {n o i s e} ^ {2} \\mathbf {I}) \\cdot p (x _ {m _ {i}}) d x _ {m _ {i}} \\\\ \\approx \\frac {1}{N} \\sum_ {b = 1} ^ {N} \\mathcal {N} \\left(x _ {m _ {i _ {(b)}}}; x _ {p _ {i}}, \\sigma_ {\\text {n o i s e}} ^ {2} \\mathbf {I})\\right), \\tag {7} \\\\ \\end{array}", + "image_path": "9387ee3fb0bbac6f4d5e53698801da25c05f18bc234daf583b5d1cbb3c545c52.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "text", + "content": "The monte carlo method treats all pseudo labels in a training batch as random samples from " + }, + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "inline_equation", + "content": "p(x_{m_i})" + }, + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "text", + "content": ". Hence, for pseudo labels in a training batch " + }, + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "inline_equation", + "content": "B = \\{x_{m_{i(1)}}, x_{m_{i(2)}}, \\ldots x_{m_{i(N)}}\\}" + }, + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "text", + "content": ", the loss is defined as Eq(8), where " + }, + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\lambda = 2\\sigma_{noise}^2" + }, + { + "bbox": [ + 306, + 246, + 545, + 309 + ], + "type": "text", + "content": " is a temperature coefficient." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 316, + 544, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 544, + 358 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i n g l e}} (i) \\cong - \\log \\frac {\\exp \\left(- \\left| \\left| x _ {p _ {i}} - x _ {m _ {i}} \\right| \\right| ^ {2} / \\lambda\\right)}{\\sum_ {x _ {m _ {i}} ^ {\\prime} \\in B} \\exp \\left(- \\left| \\left| x _ {p _ {i}} - x _ {m _ {i}} ^ {\\prime} \\right| \\right| ^ {2} / \\lambda\\right)}, \\tag {8}", + "image_path": "16b7c3e7142b4a35ec3056d6a81e8c12def75366def601ad88acfda5dfcf5125.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": "According to Eq (3) and Eq (8), it can be observed that the form of the two loss functions is similar. Both of the two loss functions pull together the prediction and self-defined pseudo label in the representation space, and push other predictions in a mini-batch apart. The MAE adopts the Euclidean distance while the contrastive learning measures the cosine similarity in representation space. Actually, MAE regards each masked patch as a category and performs patch-level contrastive learning with unlearnable pixels. It explicitly aligns the prediction with its corresponding masked patch and implicitly distinguishes other patches." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 497, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 521 + ], + "type": "text", + "content": "From the above description, when confronted with arbitrary scenes, MAE and CL will face a severe conflict issue." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 527, + 488, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 488, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 488, + 540 + ], + "type": "text", + "content": "3.3. Arbitrary self-supervised learning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 677 + ], + "type": "text", + "content": "Based on the previous analysis, it can be concluded that a global prior of CL is both unreasonable and conflicts with MIM when facing arbitrary scenarios. To enable learning from arbitrary scenarios, can there be a self-supervised method that harnesses the advantages of CL in semantic learning (learnable self-pseudo labels) and the heterogeneous representation learning of MIM, while avoiding global constraints and conflicts? The answer is affirmative. In this paper, we propose a new framework, named Arbitrary Self-supervised Learning (ASL), to learn from arbitrary scenarios while retaining their advantages." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Specifically, we adopt the popular MAE method as the basic MIM to ensure the initial heterogeneous representations. Then, inspired by CL using a dual-branch structure" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22970" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 82, + 531, + 244 + ], + "blocks": [ + { + "bbox": [ + 65, + 82, + 531, + 244 + ], + "lines": [ + { + "bbox": [ + 65, + 82, + 531, + 244 + ], + "spans": [ + { + "bbox": [ + 65, + 82, + 531, + 244 + ], + "type": "image", + "image_path": "4e811026b0a1b77f5ec139adc04c001cc27115c67064b85d413a5a1cea74bcf4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 247, + 255, + 330 + ], + "blocks": [ + { + "bbox": [ + 72, + 247, + 255, + 330 + ], + "lines": [ + { + "bbox": [ + 72, + 247, + 255, + 330 + ], + "spans": [ + { + "bbox": [ + 72, + 247, + 255, + 330 + ], + "type": "image", + "image_path": "2669a10c2431ffdfffe1b26f79450cf7d443afdbd44e71c059049ff749faac4e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 332, + 161, + 341 + ], + "lines": [ + { + "bbox": [ + 150, + 332, + 161, + 341 + ], + "spans": [ + { + "bbox": [ + 150, + 332, + 161, + 341 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 267, + 252, + 523, + 331 + ], + "blocks": [ + { + "bbox": [ + 267, + 252, + 523, + 331 + ], + "lines": [ + { + "bbox": [ + 267, + 252, + 523, + 331 + ], + "spans": [ + { + "bbox": [ + 267, + 252, + 523, + 331 + ], + "type": "image", + "image_path": "402b4a158a45160a2e1a7ff8820274cb19785fe3f003e35bfc4b01967b8e3cca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 331, + 355, + 341 + ], + "lines": [ + { + "bbox": [ + 343, + 331, + 355, + 341 + ], + "spans": [ + { + "bbox": [ + 343, + 331, + 355, + 341 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 342, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 45, + 342, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 45, + 342, + 545, + 397 + ], + "type": "text", + "content": "Figure 3. The pipeline of our ASL. The visible patches are fed into the network and the dual-branch features can be obtained by our designs. Then the encoded features with mask tokens are fed into the decoder and output the predictions. Finally, we pull together the predictions using L1 loss by one-to-one correspondence. We propose two designs in the network to obtain the dual-branch features: (a) Dropout in the linear embedding. (b) Asymmetric structure at the end of the encoder. In the figure, we adopt the network to represent (a) or (b) design. The olive patch represents the mask token." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 404, + 287, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 524 + ], + "type": "text", + "content": "to generate learnable self-pseudo labels, we propose two patch-level designs to generate dual-branch features based on the encoder of MAE while avoiding the global constraints. Subsequently, these dual-branch features are further processed through a decoder module, and CL is performed between the outputs patch by patch to for semantic learning. In contrast to traditional CL, our approach does not entail global data augmentation but instead focuses on patch-level feature augmentation. The specifics of our designs are outlined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "type": "text", + "content": "Dropout in the linear embedding (DLE). According to previous work in NLP, SimCSE [17] has explored the use of dropout in word embedding for feature augmentation. In line with this, we also adopt dropout in the linear embedding of the encoder to perform patch-level augmentation and acquire two encoded features at the end of the encoder. The details of our method are described in Figure 3 (a)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Asymmetric structure at the end of the encoder (AEE). Some prior work [8, 18, 62] have demonstrated that introducing an asymmetric structure, which essentially is a nonlinear function, enhances the semantic features. With this in mind, we incorporate a Transformer layer [57] after the encoder in our approach to build an asymmetric structure for patch-level feature augmentation. This step produces dual-branch features: one is the feature output by the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 404, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 440 + ], + "type": "text", + "content": "encoder, and the other is the feature that has passed through the Transformer layer. Figure 3 (b) illustrates the details of our method. The design is our default experimental setup." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 440, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 548 + ], + "type": "text", + "content": "The methodology described above pertains to the construction of feature augmentation. Our rationale for utilizing feature augmentation to construct dual-branch features is primarily based on two considerations: 1) Data augmentation in CL is global and not suitable for patch-level learning. 2) The asymmetric structure of CL inherently constitutes an implicit form of feature augmentation, and it is worth exploring whether its success in CL can be extrapolated to the broader domain of self-supervised learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": "After feature augmentation, we obtain the dual-branch features, " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "fea_{ori}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "fea_{aug}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": ", then we feed them into the decoder module to obtain the corresponding predictions, " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "pred_{ori}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "pred_{aug}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": ", respectively. The patch-level CL between the outputs does not employ contrastive loss or cosine similarity loss but rather L1 loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 620, + 545, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 545, + 680 + ], + "type": "text", + "content": "The alignment between any pair is defined as Eq(9), where " + }, + { + "bbox": [ + 304, + 620, + 545, + 680 + ], + "type": "inline_equation", + "content": "\\mathsf{sg}[\\cdot]" + }, + { + "bbox": [ + 304, + 620, + 545, + 680 + ], + "type": "text", + "content": " stands for stop gradient. It is noted that our alignment method does not include the alignment of [cls] token according to Section 3.2, which unleashes the potential of the method for training in arbitrary scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} (i) = \\left| \\left| \\operatorname {p r e d} _ {\\text {o r i} _ {i}} - s g \\left[ \\operatorname {p r e d} _ {\\text {a u g} _ {i}} \\right] \\right| \\right| + \\left| \\left| s g \\left[ \\operatorname {p r e d} _ {\\text {o r i} _ {i}} \\right] - \\operatorname {p r e d} _ {\\text {a u g} _ {i}} \\right| \\right|, \\tag {9}", + "image_path": "c1455cbd4d43caca919c864b868254373bd9d2f8d29ca52a581b5735eee91df3.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22971" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "The loss function of the complete dataset is defined as Eq(10)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 96, + 287, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 96, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 115, + 96, + 287, + 114 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S E M} = \\underset {\\mathbf {x} \\sim X} {\\mathbb {E}} \\underset {i \\sim \\eta} {\\mathbb {E}} \\mathcal {L} _ {s} (i), \\tag {10}", + "image_path": "bd28d8c5e4b88994d4b61651a56c3038adcca6c8913ee7e21bef8cf5e9401254.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 121, + 157, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 157, + 134 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 157, + 134 + ], + "type": "text", + "content": "3.4. Objective function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 140, + 287, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 189 + ], + "type": "text", + "content": "The loss function of our ASL consists of the MAE loss and the semantic loss. Hence, the total loss of our ASL is formulated as Eq (11), and each loss coefficient is set to be 1 for equally weighted." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 198, + 287, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 198, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 110, + 198, + 287, + 211 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {M A E}} + \\mathcal {L} _ {\\text {S E M}}, \\tag {11}", + "image_path": "aa64ece7298dc1193c15c8e94359cdd351573f339b5aa533fa9d902d53e950b1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 220, + 128, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 128, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 128, + 234 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 240, + 218, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 218, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 218, + 252 + ], + "type": "text", + "content": "4.1. Datasets and evaluation metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 258, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 498 + ], + "type": "text", + "content": "Pre-training datasets. In the pre-training stage, we select the four popular visual pre-training datasets, MS COCO [40], ImageNet [13], CIFAR10 [31], CIFAR100 [31], to investigate the versatility and transfer ability of the proposed method. For the COCO dataset, its train2017 set contains " + }, + { + "bbox": [ + 46, + 258, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\sim 118\\mathrm{k}" + }, + { + "bbox": [ + 46, + 258, + 287, + 498 + ], + "type": "text", + "content": " images. COCO contains more natural and diverse scenes and is a non-iconic image dataset. To further demonstrate the versatility of our algorithm for training in arbitrary scenarios, we conduct pre-training on both ImageNet and COCO datasets. Given the substantial difference in data volume between ImageNet and COCO, we select a subset of ImageNet, referred to as ImageNet-100, to maintain a balanced dataset size. Moreover, we pre-train on ImageNet-1K to demonstrate the effectiveness of ASL on curated data. Finally, CIFAR-10 and CIFAR-100 datasets are small-scale image datasets, and they all contain 60000 images with " + }, + { + "bbox": [ + 46, + 258, + 287, + 498 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 46, + 258, + 287, + 498 + ], + "type": "text", + "content": " size that belong to 10 and 100 categories, respectively. We pre-train models on the CIFAR-10 and CIFAR-100 datasets to demonstrate the effectiveness of ASL in the small-scale dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": "Evaluation datasets. In this paper, ASL is evaluated by the linear probing and fine-tuning classification task on ImageNet, CIFAR-10, and CIFAR-100 datasets, which are popular image classification datasets. Additionally, for computer vision, object detection, and semantic segmentation are dense prediction tasks since training images of these tasks contain massive objects inside each image. The validation of dense prediction tasks can better reflect the transfer ability of self-supervised models. Therefore, we conduct extensive experiments on COCO [40] and ADE20k [68] datasets to verify the generalization and transfer ability of ASL. COCO is a popular object detection and instance segmentation dataset, which train2017 contains about " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "118k" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " images and test on the val2017 contains " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "5k" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " images. The challenging dataset contains human annotations for 80 classes. Moreover, ADE20K is also a challenging semantic segmentation dataset, and it contains about " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "25k" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " training images and " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "2k" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " validation images with 150 categories." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 550, + 174 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 550, + 174 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 550, + 174 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 550, + 174 + ], + "type": "table", + "html": "
MethodArch.FLOPsEpochsLPAPb
Pre-training on COCO dataset:
MAE [22]ViT-B1×80046.8%49.3%
CMAE [27]ViT-B~ 3×80046.6%49.0%
ccMIM [67]ViT-B~ 3.3×80046.7%49.0%
SIM [54]ViT-B~ 2.8×80046.0%48.7%
iBOT [69]ViT-B~ 6×80045.2%48.6%
ASLViT-B~ 1.5×80048.6%50.3%
", + "image_path": "f18ade2db5087a21781a1bcd9f7a9963bdeca48293dda8f9f6e980f969aa04e4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 211, + 547, + 304 + ], + "blocks": [ + { + "bbox": [ + 306, + 175, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 306, + 175, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 545, + 209 + ], + "type": "text", + "content": "Table 1. Evaluation of pre-trained models on the COCO dataset. We report Linear Probing (LP) top-1 accuracy on the ImageNet-1K val set and bbox mAP on COCO dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 211, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 308, + 211, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 547, + 304 + ], + "type": "table", + "html": "
MethodArch.Param.EpochsLPFT
Pre-training on ImageNet-100 + COCO:
MAE [22]ViT-B8680074.5%90.9%
iBOT[69]ViT-B8680071.8%88.5%
ccMIM [67]ViT-B8680072.3%88.7%
ASLViT-B8680079.6%92.4%
ASLViT-B86400085.9%94.2%
", + "image_path": "bdfdd4599018376d37a7cf1100ca819223fd395b81d8506fc8ed0090e9ce79c2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 305, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 305, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 545, + 361 + ], + "type": "text", + "content": "Table 2. ImageNet-100 Top-1 accuracy of different methods under linear probing (LP) and fine-tuning (FT) setting to evaluate the performance of the model pre-trained on both ImageNet-100 and COCO. We report top-1 accuracy on the ImageNet-100 val set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 365, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 389 + ], + "type": "text", + "content": "The valuation metrics, pre-training settings, and the experimental settings of downstream tasks in the Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 396, + 473, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 473, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 473, + 408 + ], + "type": "text", + "content": "4.2. Pre-training on COCO dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": "In order to validate the adaptability of our approach for training in arbitrary scenarios, we initially conduct pretraining on the COCO dataset and subsequently conduct a fair comparison with other methods under the same experimental settings. In this comparison, we evaluate the ImageNet linear evaluation and the COCO detection results. Linear evaluation, which involves using frozen features, provides a more comprehensive assessment of a model's representation. Given that the pre-training experiments are conducted on the COCO dataset, using human-labeled data from COCO to test the learned representations is reasonable. As shown in Table 1, we observe that methods combining MAE with contrastive learning, such as CMAE, ccMIM, and SIM, exhibit weaker performance when pretrained on the COCO dataset, even though they achieved SOTA performance on ImageNet. This observation suggests that existing SOTA methods are not well-suited for training in arbitrary scenarios. Furthermore, our approach demonstrates superior performance after pre-training on the COCO dataset, surpassing the performance of MAE. Generally, introducing a dual-branch structure similar to contrastive learning typically results in a significant increase in computational overhead, often exceeding threefold. Diverging from the emphasis on the encoder in hybrid methods, our AEE design prioritizes the decoder, significantly reducing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22972" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 542, + 234 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 542, + 234 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 542, + 234 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 542, + 234 + ], + "type": "table", + "html": "
MethodPre-train IterationsDatasetObject detectionInstance segmentation
\\(AP^b\\)\\(AP_{50}^b\\)\\(AP_{75}^b\\)\\(AP^m\\)\\(AP_{50}^m\\)\\(AP_{75}^m\\)
Supervised [9]~ 94kImageNet-1K47.9%--42.9%--
MoCo v3 [8]~ 187kImageNet-1K47.9%--42.7%--
BEiT [1]~ 249kImageNet-1K + DALLE49.8%--44.4%--
MAE [22]~ 499kImageNet-1K50.4%70.8%55.7%44.9%68.3%48.9%
CAE [9]~ 499kImageNet-1K50.0%70.9%54.8%44.0%67.9%47.6%
R-MAE [45]~ 115kCOCO50.6%--45.0%--
R-MAE [45]~ 250kCOCO50.8%--45.2%--
ccMIM [67]~ 249kImageNet-1K50.3%71.2%55.0%44.5%68.4%47.9%
ASL~ 47kImageNet-100 + COCO50.7%71.1%55.6%45.0%68.3%49.0%
ASL~ 236kImageNet-100 + COCO51.5%71.9%57.2%45.7%69.2%50.0%
", + "image_path": "ebde70f32203de049d49384a3e3fe90999a63ef2810fba7db7a3172e7b131f26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 259, + 291, + 402 + ], + "blocks": [ + { + "bbox": [ + 47, + 236, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 236, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 545, + 258 + ], + "type": "text", + "content": "Table 3. Results of object detection and instance segmentation on COCO. The architecture of various methods is ViT-B, and we report the bounding box AP and mask AP on COCO val2017." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 259, + 291, + 402 + ], + "lines": [ + { + "bbox": [ + 50, + 259, + 291, + 402 + ], + "spans": [ + { + "bbox": [ + 50, + 259, + 291, + 402 + ], + "type": "table", + "html": "
MethodDatasetPre-train IterationsADE mIoU
Supervised [22]ImageNet-1K~ 93k47.4%
SplitMask [16]ADE20K~ 128k45.7%
MoCo v3 [8]ImageNet-1K~ 187k47.3%
BEiT [22]ImageNet-1K+DALLE~ 249k47.1%
R-MAE [45]COCO~ 115k46.8%
R-MAE [45]COCO~ 250k47.0%
ccMIM [67]ImageNet-1K~ 249k47.7%
MAE [22]ImageNet-1K~ 499k48.1%
ASLImageNet-100 + COCO~ 236k48.4%
", + "image_path": "24227bee0d7f3652b311959aa42e78f860b84c0c18397c39a2253ba55eb00e35.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 288, + 426 + ], + "lines": [ + { + "bbox": [ + 47, + 403, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 288, + 426 + ], + "type": "text", + "content": "Table 4. Results of semantic segmentation on ADE20K. We report results measured by mIoU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 431, + 288, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 288, + 491 + ], + "type": "text", + "content": "ing computational costs in comparison. Meanwhile, due to the fixed-size and lightweight design of the decoder, our design incurs decreasing computational overhead as the model scale increases. This highlights the potential of our method for pre-training in non-iconic scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 500, + 271, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 271, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 271, + 514 + ], + "type": "text", + "content": "4.3. Pre-training on both COCO and ImageNet" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "text", + "content": "To further demonstrate the adaptability of ASL across diverse scenarios, we conduct the pre-training experiments on both ImageNet-100 and COCO datasets. We assess the performance of the model on benchmark tasks: image classification, object detection, and semantic segmentation. The ASL models are primarily trained with 800 epochs and 4000 epochs, with the number of training iterations for the 4000-epoch training similar to that of the ImageNet-1K iterations for the 800-epoch training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 646, + 164, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 164, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 164, + 658 + ], + "type": "text", + "content": "4.3.1 Image classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "content": "Since our training data comprises the ImageNet-100 dataset, we conducted testing for classification performance through both linear probing and fine-tuning on ImageNet-100. The experimental results are presented in Table 2. The" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 262, + 547, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 347 + ], + "type": "text", + "content": "results of pre-training observed from the table align with those of pre-training solely on COCO: ccMIM exhibits relatively weak performance in the context of the pre-training. This can be attributed to conflicts between CL and MAE when dealing with arbitrary scenarios. In contrast, our method successfully mitigates such conflicts and produces more robust representations in arbitrary scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 361, + 525, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 525, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 525, + 373 + ], + "type": "text", + "content": "4.3.2 Object detection and instance segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "type": "text", + "content": "To verify the transfer ability of ASL, we test it on object detection and instance segmentation with COCO [40] dataset. Here we report Box AP and mask AP on the validation set. According to the benchmark validation [22, 33], we choose the Mask R-CNN [20] as the test framework. Due to variations in the sizes of the pre-training datasets used, we opt to avoid epoch-based comparisons in this scenario. We convert epochs to a fixed batch size of 4096 iterations for a fair comparison. Table 3 shows the results of the learned representation by supervised method and different self-supervised methods. It shows our ASL, with approximately 236k iterations (4000 epochs), achieves impressive performance and outperforms 1600-epoch MAE, 1600-epoch CAE and 800-epoch ccMIM pre-trained on ImageNet. ASL surpasses various supervised and self-supervised models pre-trained on ImageNet and COCO, indicating its success in arbitrary scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 599, + 437, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 437, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 437, + 611 + ], + "type": "text", + "content": "4.3.3 Semantic segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": "To further evaluate the generalization of our method, we conduct the semantic segmentation experiments on ADE20K dataset [68]. As shown in Table 4, ASL achieves the " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "48.4\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " mIoU. It also outperforms various supervised and self-supervised models pre-trained on ImageNet." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Overall, our ASL shows impressive performance on various mainstream visual benchmark tasks when pre-training on both COCO and ImageNet-100, which demonstrates" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22973" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 71, + 281, + 123 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 281, + 123 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 281, + 123 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 281, + 123 + ], + "type": "table", + "html": "
MethodLPFTAPbAPmmIoU
MAE67.8%83.6%50.4%44.9%48.1%
ASL69.2%84.2%51.3%45.7%49.2%
", + "image_path": "a44ec4862b99658d27accfaa4740e8ef98f34f475c6b1bab8ad57628d030cbc9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 161, + 286, + 220 + ], + "blocks": [ + { + "bbox": [ + 47, + 125, + 286, + 157 + ], + "lines": [ + { + "bbox": [ + 47, + 125, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 286, + 157 + ], + "type": "text", + "content": "Table 5. The ImageNet-1K pre-training results. We report ImageNet classification, COCO detection, and instance segmentation, as well as ADE20k semantic segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 161, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 161, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 161, + 286, + 220 + ], + "type": "table", + "html": "
MethodCIFAR-10CIFAR-100
Linear Prob.FinetuningLinear Prob.Finetuning
Supervised [26]-91.3%-64.1%
DINO [3]89.0%94.4%65.8%76.3%
MAE [22]87.3%95.9%54.0%81.1%
ASL91.4%96.5%60.3%82.1%
", + "image_path": "7a9fd7a7480ef63b82de45017db9dcf6780c1776130086e9d3410048b7fdbbac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 245, + 286, + 313 + ], + "blocks": [ + { + "bbox": [ + 47, + 220, + 286, + 242 + ], + "lines": [ + { + "bbox": [ + 47, + 220, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 286, + 242 + ], + "type": "text", + "content": "Table 6. Linear probing and fine-tuning classification results on small-scale datasets. We adopt ViT-S as the base model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 245, + 286, + 313 + ], + "lines": [ + { + "bbox": [ + 50, + 245, + 286, + 313 + ], + "spans": [ + { + "bbox": [ + 50, + 245, + 286, + 313 + ], + "type": "table", + "html": "
IDDLEAEEalignmentFLOPsTop-1APb
(a)×××1×46.8%49.3%
(b)✓××~ 2×46.5%49.1%
(c)✓×✓~ 2×48.5%50.4%
(d)×✓×~ 1.5×46.6%49.2%
(e)×✓✓~ 1.5×48.6%50.3%
", + "image_path": "8fcefa8aa83418238c6a51abe6594d822783edb401177fd989ce646c1f645515.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 358, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 286, + 382 + ], + "type": "text", + "content": "ASL has the ability to learn general visual representations from arbitrary scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 390, + 207, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 207, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 207, + 403 + ], + "type": "text", + "content": "4.4. Pre-training on ImageNet-1K" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 409, + 286, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 286, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 286, + 505 + ], + "type": "text", + "content": "Our method is designed for learning visual representations from uncurated arbitrary scenarios, and it can also yield stronger representations when applied to meticulously curated datasets. To demonstrate this point, we conduct experiments on the curated ImageNet dataset. As depicted in the experimental results shown in Table 5, ASL exhibits superior performance across various downstream tasks compared to MAE, highlighting the versatility of our approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 513, + 206, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 206, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 206, + 525 + ], + "type": "text", + "content": "4.5. Small-scale data pre-training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 532, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 286, + 638 + ], + "type": "text", + "content": "As shown in Table 6, the results of our method on these small datasets all show impressive performance. With the same training settings for all experiments, ASL consistently improves MAE by at least " + }, + { + "bbox": [ + 46, + 532, + 286, + 638 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 46, + 532, + 286, + 638 + ], + "type": "text", + "content": " for linear probing. However, MIM still lags behind CL in linear evaluation on CIFAR100 and its advantage lies in the fine-tuning superior performance. Our fine-tuning results also reveal it. Totally, these results suggest that ASL is effective and can learn more semantic features than MAE in small-scale datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 138, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 138, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 138, + 659 + ], + "type": "text", + "content": "4.6. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": "The effect of feature augmentation. We propose two designs to obtain a dual-branch structure by performing feature augmentation. As shown in Table 7, (a) is the results of MAE as the baseline when pre-trained on COCO. Table" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "7 (b) indicates the results of the DLE design are slightly inferior to the baseline. However, the results outperform the baseline MAE when aligning the predictions according to (c). Similarly, the results for (d) using AEE are not as favorable as MAE, while the results for (e) incorporating alignment operations outperform MAE. These experimental findings collectively indicate that creating learnable pseudo labels based on feature augmentation during the pretraining phase can improve the representation of pre-trained models. Due to the relatively comparable performance of (e) and (c), with (e) incurring significantly lower computational costs, we opt for (e) as the final experimental design." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 217, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 275 + ], + "type": "text", + "content": "The effect of alignment loss. We compare two loss functions: one commonly used in contrastive learning, which is cosine similarity loss, and the other is L1 loss. The following is a comparison between cosine similarity loss and L1 loss:" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 307, + 283, + 544, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 316, + 286, + 348 + ], + "lines": [ + { + "bbox": [ + 47, + 316, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 286, + 348 + ], + "type": "text", + "content": "Table 7. Ablations for ASL: Effect of feature augmentation. We conduct pre-training experiments on COCO dataset and report linear evaluation on ImageNet-1K and detection results on COCO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 283, + 544, + 308 + ], + "lines": [ + { + "bbox": [ + 307, + 283, + 544, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 544, + 308 + ], + "type": "table", + "html": "
MAEcosine similarity lossL1 loss[cls] align
46.8%47.6%48.6%45.6%
", + "image_path": "69ba308ba0f4e19f4d05085099d1e36f59f96cdf0c1a2becee02dcea9f068d79.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "type": "text", + "content": "From the above, experimental results indicate that the L1 loss facilitates the model in learning better representations compared to cosine similarity loss. Besides, we conduct an additional experiment involving alignment with the [cls] token on top of the L1 loss. The results indicate that introducing global contrastive learning during pre-training on the COCO dataset has a detrimental effect on performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 410, + 383, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 383, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 383, + 422 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 430, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 656 + ], + "type": "text", + "content": "This paper points out the inability of current hybrid SOTA methods to extend training to arbitrary scenarios due to the conflict between MAE and CL. It then theoretically identifies the root of the issue as MAE being patch-level CL, conflicting with global CL. To address this conflict, we introduce a novel self-supervised learning method ASL to learn visual representations from arbitrary scenarios. Specifically, this paper abandons the global-level consistency of CL and proposes explicit patch-level learning. ASL adopts patch-level feature augmentation to generate dual-branch features by the encoder of MAE, then these features perform patch-level learning through a decoder module. Finally, ASL can learn heterogeneous representations while avoiding the conflicts. Our method demonstrates robustness and versatility in multiple pre-training datasets and downstream visual tasks. We expect that our study will draw the attention of the community to large-scale pre-train from arbitrary scenarios and contribute to the development of visual foundation models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 657, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 545, + 702 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by National Key R&D Program of China under Grant No. 2021YFE0205700 and National Natural Science Foundation of China (No.62276260, No.62176254, No.62076235)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22974" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 178 + ], + "type": "text", + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, pages 9912-9924, 2020. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv: Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "text", + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey K Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703, 2020. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 1, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 367 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E. Hinton. Big self-supervised models are strong semi-supervised learners. In Advances in Neural Information Processing Systems, pages 22243-22255, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "text", + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "text", + "content": "[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 3, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "text", + "content": "[9] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "text", + "content": "[10] Zhiyang Chen, Yousong Zhu, Zhaowen Li, Fan Yang, Wei Li, Haixin Wang, Chaoyang Zhao, Liwei Wu, Rui Zhao, Jinqiao Wang, et al. Obj2seq: Formatting objects as sequences with class prompt for visual tasks. Advances in Neural Information Processing Systems, 35:2494-2506, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "text", + "content": "[11] Zhiyang Chen, Yousong Zhu, Fan Yang, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. The devil is in details: Delving into lite ffn design for vision transformers. Available at SSRN 4299967, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[12] Zhiyang Chen, Yousong Zhu, Yufei Zhan, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang. Mitigating hallucination in visual language models with visual supervision. arXiv preprint arXiv:2311.16479, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 193 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 194, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 237 + ], + "type": "text", + "content": "[16] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[17] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 358 + ], + "type": "text", + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 1, 3, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "text", + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In The IEEE International Conference on Computer Vision, 2017. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 524 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "text", + "content": "[23] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and Joao Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[24] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In European Conference on Computer Vision, pages 123-143. Springer, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[25] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. International Conference on Learning Representations, 2019. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[26] Tianyu Hua, Yonglong Tian, Sucheng Ren, Hang Zhao, and Leonid Signal. Self-supervision through random seg" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22975" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 94 + ], + "type": "text", + "content": "ments with autoregressive coding (rand sac). arXiv preprint arXiv:2203.12054, 2022.8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[27] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. arXiv preprint arXiv:2207.13532, 2022. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[28] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "text", + "content": "[29] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 261 + ], + "type": "text", + "content": "[30] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 283 + ], + "type": "text", + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "text", + "content": "[32] Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teachers in self-supervised masked auto-encoders. arXiv preprint arXiv:2210.02077, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[33] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 372, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 425 + ], + "type": "text", + "content": "[34] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 23390-23400, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "type": "text", + "content": "[35] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Mst: Masked self-supervised transformer for visual representation. In Advances in Neural Information Processing Systems, 2021. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 483, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 287, + 525 + ], + "type": "text", + "content": "[36] Zhaowen Li, Xu Zhao, Chaoyang Zhao, Ming Tang, and Jinqiao Wang. Transfering low-frequency features for domain adaptation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01-06. IEEE, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 526, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 592 + ], + "type": "text", + "content": "[37] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 14627-14636, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 647 + ], + "type": "text", + "content": "[38] Zhaowen Li, Xu Zhao, Peigeng Ding, Zongxing Gao, Yuting Yang, Ming Tang, and Jinqiao Wang. Freconv: Frequency branch-and-integration convolutional networks. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 258-263. IEEE, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[39] Zhaowen Li, Yousong Zhu, Zhiyang Chen, Wei Li, Chaoyang Zhao, Liwei Wu, Rui Zhao, Ming Tang, and Jinqiao Wang. Efficient masked autoencoders with self-consistency. arXiv preprint arXiv:2302.14431, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[41] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "text", + "content": "[42] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 188, + 545, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 210 + ], + "type": "text", + "content": "[43] Peter McCullagh and John A Nelder. Generalized linear models. Routledge, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 212, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 266 + ], + "type": "text", + "content": "[44] Shlok Mishra, Joshua Robinson, Huiwen Chang, David Jacobs, Aaron Sarna, Aaron Maschinot, and Dilip Krishnan. A simple, efficient and scalable contrastive masked autoencoder for learning visual representations. arXiv preprint arXiv:2210.16870, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 268, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 312 + ], + "type": "text", + "content": "[45] Duy-Kien Nguyen, Vaibhav Aggarwal, Yanghao Li, Martin R Oswald, Alexander Kirillov, Cees GM Snoek, and Xinlei Chen. R-mae: Regions meet masked autoencoders. arXiv preprint arXiv:2306.05411, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 315, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 358 + ], + "type": "text", + "content": "[46] David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 IEEE international conference on neural networks (ICNN'94), pages 55-60. IEEE, 1994. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 403 + ], + "type": "text", + "content": "[47] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proceedings of the European Conference on Computer Vision, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 406, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 439 + ], + "type": "text", + "content": "[48] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 441, + 466, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 466, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 466, + 453 + ], + "type": "text", + "content": "[49] OpenAI. Introducing chatgpt, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 453, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 508 + ], + "type": "text", + "content": "[50] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 510, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 554 + ], + "type": "text", + "content": "[51] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 555, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 610 + ], + "type": "text", + "content": "[52] Hao Quan, Xingyu Li, Weixing Chen, Mingchen Zou, Ruijie Yang, Tingting Zheng, Ruiqun Qi, Xinghua Gao, and Xiaoyu Cui. Global contrast masked autoencoders are powerful pathological representation learners. arXiv preprint arXiv:2205.09048, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 612, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 657 + ], + "type": "text", + "content": "[53] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems, 2015. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[54] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. Proceedings of the IEEE conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22976" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "text", + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 5998-6008, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 250 + ], + "type": "text", + "content": "[58] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. arXiv preprint arXiv:2011.09157, 2020.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 287, + 297 + ], + "type": "text", + "content": "[59] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "type": "text", + "content": "[60] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision, pages 418-434, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 343, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 386 + ], + "type": "text", + "content": "[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. arXiv preprint arXiv:2011.10043, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "type": "text", + "content": "[62] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swim transformers. arXiv preprint arXiv:2105.04553, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 287, + 475 + ], + "type": "text", + "content": "[63] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 478, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 287, + 520 + ], + "type": "text", + "content": "[64] Kun Yi, Yixiao Ge, Xiaotong Li, Shusheng Yang, Dian Li, Jianping Wu, Ying Shan, and Xiaohu Qie. Masked image modeling with denoising contrast. arXiv preprint arXiv:2205.09616, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "text", + "content": "[65] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. arXiv preprint arXiv:2210.08344, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 557, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 589 + ], + "type": "text", + "content": "[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Proceedings of the European Conference on Computer Vision, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 591, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 591, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 591, + 287, + 644 + ], + "type": "text", + "content": "[67] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In The Eleventh International Conference on Learning Representations. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 328, + 73, + 545, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 545, + 94 + ], + "type": "text", + "content": "with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22977" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_content_list.json b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a8fb59b5cdb5d6554c4c179f97907879a7766d4 --- /dev/null +++ b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_content_list.json @@ -0,0 +1,1388 @@ +[ + { + "type": "text", + "text": "Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement", + "text_level": 1, + "bbox": [ + 89, + 130, + 879, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zaid Khan $^{1}$ Vijay Kumar BG $^{2}$ Samuel Schulter $^{2}$ Yun Fu $^{1}$ Manmohan Chandraker $^{2,3}$", + "bbox": [ + 112, + 202, + 852, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northeastern University $^{2}$ NEC Laboratories America $^{3}$ UC San Diego", + "bbox": [ + 259, + 222, + 709, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual program synthesis is a promising approach to exploit the reasoning abilities of large language models for compositional computer vision tasks. Previous work has used few-shot prompting with frozen LLMs to synthesize visual programs. Training an LLM to write better visual programs is an attractive prospect, but it is unclear how to accomplish this. No dataset of visual programs for training exists, and acquisition of a visual program dataset cannot be easily crowdsourced due to the need for expert annotators. To get around the lack of direct supervision, we explore improving the program synthesis abilities of an LLM using feedback from interactive experience. We propose a method where we exploit existing annotations for a vision-language task to improvise a coarse reward signal for that task, treat the LLM as a policy, and apply reinforced self-training to improve the visual program synthesis ability of the LLM for that task. We describe a series of experiments on object detection, compositional visual question answering, and image-text retrieval, and show that in each case, the self-trained LLM outperforms or performs on par with few-shot frozen LLMs that are an order of magnitude larger. Website:", + "bbox": [ + 73, + 305, + 473, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://zaidkhan.me/ViReP", + "bbox": [ + 76, + 623, + 326, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 662, + 207, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Complex visual queries can often be decomposed into simpler subtasks, many of which can be carried out by task-specific perception modules (e.g. object detection, captioning). For example, consider the problem of finding bounding boxes for the phrase \"white mug to the left of the sink\". This is a challenging query for single model such as an open vocabulary object detector. However, this query can be solved by writing a program that composes task-specific perception modules with logic: use an open vocabulary object detector to find a sink and white mugs in the scene, then compare the horizontal center of the sink and the mugs to find white mugs to the left of the sink. Program synthesis with large language models [1] is a promising approach to automate this process, and recent work has shown that proprietary large language", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66344b4b51bcfc966e6bbfbe756ff8b1eb4ab3f4251b96d71c5c74cf85faf09e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 273, + 890, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ef705e9a0f73bda7686e15c73263592e83053126f9652916b2145f1d7c832d9.jpg", + "image_caption": [ + "How can we train a LLM to write better visual programs for a task?", + "Figure 1. Visual program synthesis with LLMs has been treated as a $0 / \\mathrm{n}$ -shot task where the LLM is kept frozen. This limits opportunities for improvement. We ask whether it is possible to train a LLM to write more accurate programs. Given that there is no large scale dataset of accurate visual programs available, we propose improving the LLM using self-training." + ], + "image_footnote": [], + "bbox": [ + 508, + 435, + 864, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models can write programs for visual tasks [9, 27, 28]. Current approaches for visual program synthesis with LLMs use few-shot prompting and rely on the in-context learning abilities [32] of frozen, proprietary LLMs. (Fig. 1)", + "bbox": [ + 496, + 671, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot prompting with frozen LLMs for visual program synthesis as in ViperGPT [28], VisProg [9], or CodeVQA [27] has several limitations. The LLM needs to understand the competencies of the perception modules it is using. A open vocabulary object detector may be able to locate a common attribute-noun phrase such as \"white mug\" without problems, but struggle with a more abstract phrase such as \"microwaveable mug\" [24]. A VQA model might be able to answer \"is the car blue?\" without problems, but fail when logical modifiers are introduced, such as \"is the car not blue?\" [7]. In many cases, we do not precisely know", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14344", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/fde2adfd1f09c7887052d783c301e16cccc39fbb872ceba814607ee7989c09f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Task DomainSelf-TrainingSupervisionTool/API UseVisual Task DecompositionGrounded By FeedbackImproves LLM
VisReP (Ours)Visual Program SynthesisYesWeakYesYesYesYes
Haluptzok et al. [10]Programming PuzzlesYesWeakNoNoYesYes
ReST [8]Natural Language UnderstandingYes-NoNoYesYes
VisProg [9]Visual Program SynthesisNo-YesYesNoNo
ViperGPT [28]Visual Program SynthesisNoWeakYesYesNoNo
ToolLLM [20]Tool Usage by APINoStrongYesNoNoYes
GorillaLLM [19]Tool Usage by APINoStrongYesNoNoYes
ToolFormer [23]Natural Language UnderstandingYesWeakYesNoYesYes
", + "bbox": [ + 78, + 88, + 890, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Differences between our work and similar work. Strong supervision means that the training process requires examples of ground-truth programs to train the LLM. Weak supervision means that the training process does not require ground-truth programs. Tool / API use means that the LLM is required to use substantial functionality implemented by external modules (e.g. an object detector, a web search API) to solve tasks. Visual task decomposition means that the LLM can decompose a complex visual task into primitive subtasks. Grounded by feedback means that the LLM has been optimized not just for syntactic / semantic correctness (program does not hallucinate / cause errors), but for functional correctness (programs produce the correct answer). Improves LLM means that the work proposes a method to improve an LLM for a specific task, rather than using a frozen LLM.", + "bbox": [ + 75, + 199, + 892, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the weaknesses or competencies of a perception model [30]. Even if this were known, it is difficult to convey all of the competencies / weaknesses of a perception module through in-context examples. Second, it is the case in program synthesis that an LLM often can generate the correct solution to a problem, but the correct solution is not the solution the LLM places the highest probability on [4]. We would like to align the LLM to \"uncover\" the knowledge of the correct solution, but it not clear how to do this in a principled way with few-shot prompting alone.", + "bbox": [ + 75, + 318, + 470, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How can we train a large language model to write better visual programs for a specific task?", + "bbox": [ + 107, + 479, + 434, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to optimize the parameters of the language model so the accuracy of the synthesized programs is higher. Existing approaches that train LLMs to improve their ability to programatically use tools / APIs such as GorillaLLM [19], ToolLLM [20] do so by finetuning LLMs on examples of tool use or API use. This cannot be directly applied to visual program synthesis because there are no large scale datasets of visual programs, and collecting such a dataset would be extremely labor intensive. In the absence of a large scale dataset, how do we learn to write better programs for a visual task?", + "bbox": [ + 75, + 518, + 468, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We posit that grounding a language model with interactive feedback from a generic visual task will improve the general visual program synthesis abilities of the model.", + "bbox": [ + 107, + 695, + 434, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A natural way to learn from feedback is to use reinforcement learning. ReST [8] and RaFT [6] introduce a general framework for reinforced self-training in generative tasks and demonstrate success in machine translation and text-to-image generation. However, a crucial ingredient in their recipe is the availability of a fine-grained reward model. It is difficult to construct a fine-grained reward model for visual program synthesis, given both the absence of human preference datasets for visual programs, and the difficulty", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of devising a proxy metric. One alternative is to use unit tests to teach a neural reward model or give a coarse-grained reward. This technique has been used successfully in coding challenges by CodeRL [16] and Haluptzok et al. [10], but it is unclear how it can be applied to visual program synthesis. Our key idea is to use existing annotations for a vision-language as improvised unit tests to provide a coarse reward signal. Using the coarse reward signal, we can apply reinforced self-training by treating the language model as a policy and training it with a simple policy gradient algorithm. We alternate synthetic data generation steps in which we sample programs from the language model policy with optimization steps in which we improve the language model policy based on observations from executing the sampled programs. We name our proposed method VisReP, for Visually Reinforced Program Synthesis.", + "bbox": [ + 496, + 318, + 893, + 561 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose optimizing the parameters of a LLM so that the accuracy of the synthesized visual programs is higher, in contrast to previous works that use frozen LLMs.", + "- Since no dataset of accurate visual programs is available for finetuning, we hypothesize that we can instead use feedback from the execution environment to improve the visual program synthesis abilities of a language model.", + "- We propose VisReP, an offline, model agnostic recipe for reinforced self-training of large language models for visual program synthesis using existing vision-language annotations with a simple policy gradient algorithm.", + "- Our results show that it is possible to apply reinforced self-training for to improve large language models for visual program synthesis with only coarse rewards." + ], + "bbox": [ + 500, + 571, + 890, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of an CodeLlama-7B policy trained by VisReP on compositional visual question answering $(+9\\%)$ , complex object detection $(+5\\%)$ , and compositional image-text matching $(+15\\%)$ relative to the untrained policy. We show that the policy trained by VisReP exceeds the accuracy of a gpt-3.5-turbo policy on all three tasks.", + "bbox": [ + 496, + 794, + 893, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14345", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Self-Training", + "text_level": 1, + "bbox": [ + 76, + 114, + 217, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-training is an established paradigm which uses unlabeled data to improve performance. Self-training has been successfully applied in a number of fields. We restrict our coverage to usages with significant overlap.", + "bbox": [ + 75, + 137, + 470, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Program Synthesis Haluptzok et al. [10] showed that LLMs can improve their program synthesis abilities by generating programming puzzles and solving them. CodeRL [16] proposed an actor-critic framework to improve the program synthesis abilities of LLMs for programming problems accompanied by unit tests. CodeIT [3] and Rest-EM [26] also use a similar policy gradient approach for program synthesis. Our problem domain is different from these works, which focus on program synthesis for programming puzzles / problems. In addition, our work has an explicit focus on learning to use an API fluently.", + "bbox": [ + 75, + 198, + 470, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alignment ReST [8] and RaFT [6] introduced a generic framework for reinforced self-training and applied it to align machine translation outputs to human preferences and align foundation models on language understanding and image generation tasks respectively. These works share the same basic idea as our work, though they are in a substantially different task domain where human preferences are either known (conversational alignment) or can be estimated with an available neural model.", + "bbox": [ + 75, + 364, + 468, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language SelTDA [14] introduced a self-training approach for visual question answering. SelTDA proceeds by pseudolabeling unlabeled data, then finetuning a large VLM on the pseudolabeled data. In contrast to SelTDA, we improve a LLM for visual program synthesis.", + "bbox": [ + 75, + 500, + 468, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual Program Synthesis", + "text_level": 1, + "bbox": [ + 76, + 585, + 312, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual program synthesis with LLMs was proposed concurrently by ViperGPT [28], VisProg [9], and CodeVQA [27]. The common points between these three works is that (a) they use pretrained LLMs as code generators (b) they represent complex visual tasks as compositions of primitive visual subtasks (c) they use code to invoke task-specific models to perform the primitive subtasks. Our work is most similar to ViperGPT and CodeVQA as they produce code in a general purpose programming language rather than a DSL. All three works use a proprietary, frozen LLM. In contrast to all three, the focus of our work is on how we can improve the visual program synthesis abilities of an open LLM.", + "bbox": [ + 75, + 609, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Tool Use with LLMs", + "text_level": 1, + "bbox": [ + 76, + 801, + 269, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal tool-using LLMs were first introduced by Socratic Models [33]. However, their approach was to create fixed pipelines in which the output of a perception model such as CLIP [21] is fed to a LLM. Later approaches such as GorillaLLM [19] and ToolLLM [20] improved on this by", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "treating tool use as a program synthesis problem and creating LLMs that use a broad range of tools by learning to invoke APIs. However, one key limitation of these approaches in the context of visual program synthesis is that they do not learn to decompose problems into subproblems that can be solved by tools. Instead, they are trained to select the right tool for the problem and invoke it. Another limitation is that they are not optimized for functional correctness. They are trained for syntactic and semantic correctness, but they have not been provided feedback on whether their use of tools produces the desired answer. ToolFormer [23] is similar to our work in the sense that the LLM's usage of tools is grounded by feedback, but they focus on natural language understanding tasks rather than visual tasks.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 316, + 591, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Visual Program Synthesis with LLMs", + "text_level": 1, + "bbox": [ + 500, + 340, + 825, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Formulation Let $v$ be a visual input and $q$ be a textual query about $v$ . In visual program synthesis, we synthesize a program $p = \\pi_{\\theta}(q)$ with a program generator $\\pi_{\\theta}$ . The program $p$ and visual input $q$ are then fed into the execution engine $\\hat{y} = \\phi(v, p)$ to produce a result $\\hat{y}$ . The program generator is an auto-regressive large language model", + "bbox": [ + 496, + 364, + 893, + 457 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\theta} (\\boldsymbol {y} \\mid \\boldsymbol {x}) = \\prod_ {t = 1} ^ {T} \\pi_ {\\theta} \\left(p _ {t} \\mid \\mathbf {p} _ {1: t - 1}, \\boldsymbol {x}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 467, + 892, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{1:t}$ are the tokens of the program, and $x$ is the input to the large language model. The language model is kept frozen in previous work [28]. Our goal is to optimize the parameters $\\theta$ of the language model $\\pi$ so the accuracy of the synthesized programs is higher.", + "bbox": [ + 496, + 518, + 893, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implementation Following ViperGPT [28], we provide the specification of the ImagePatch API concatenated with the textual query $q$ as the prompt to the program generator. The synthesized program $p$ is a Python program that can invoke any Python builtins, control flow structures, and the ImagePatch API. Our implementation of the ImagePatch API is largely similar to ViperGPT. We remove some API methods that were not required for the tasks we evaluate on (such as l1m_query). We use BLIP [17] and GroundingDINO [18] as perception modules underlying find (object detection), simple-query (visual question answering), and verify_property (attribute verification).", + "bbox": [ + 496, + 595, + 893, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Reinforced Self-Training", + "text_level": 1, + "bbox": [ + 500, + 801, + 725, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rather than use a frozen large language model as the program generator $p_{\\theta}$ , we would like to optimize the parameters $\\theta$ of the language model so the accuracy of the synthesized programs is higher. It is not obvious how to do this. We can't backpropagate through the execution engine $\\phi(\\pi_{\\theta}(q), v)$ to", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14346", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/526901645863a3d4e02b783d757276e55698454498b792b7a9d9885b081a8b22.jpg", + "image_caption": [ + "Offline Policy Improvement" + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 457, + 164 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f3c47938ba1c105a42024486873d143a800ecb1bba29453f3bfd7cfef1cd069d.jpg", + "image_caption": [ + "Figure 2. VisReP can be applied to improve the visual synthesis abilities of an LLM for a vision-language task using existing annotations for a vision-language task (e.g. an object description + image + bounding boxes). A key idea is to construct a coarse reward by comparing the answer produced by a synthesized program to the ground-truth answer." + ], + "image_footnote": [], + "bbox": [ + 81, + 210, + 464, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "directly optimize $\\theta$ with respect to $q$ or $v$ . An alternative might be to use human labor to build a dataset of high-quality visual programs, and train the large language model $\\pi_{\\theta}$ on the manually-collected dataset. But collecting such a dataset is very labor intensive, and not scalable. Instead, we explore the idea of learning from experience by applying a simple policy gradient method, REINFORCE [31].", + "bbox": [ + 75, + 459, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose VisReP, which treats the program synthesis task as a growing batch RL problem [15], inspired by ReST [8]. We first define a coarse discrete reward function $R(\\cdot)$ from existing annotations for a vision-language task. We then alternate Grow steps, in which we sample trajectories (programs) from the policy (large language model), with Improve steps, in which we apply behavioral cloning with a reward-weighted negative-log likelihood loss to improve the policy. A diagram of our approach is depicted in Fig. 2.", + "bbox": [ + 75, + 566, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grow Step The grow step corresponds to the acting step in reinforcement learning, and can also be seen as synthetic data generation. Let $\\mathcal{D} = \\{(v_1, q_1, y_1), \\ldots, (v_n, q_n, y_n)\\}$ be a dataset for a vision-language task, where $v_i$ is an image, $q_i$ is a textual query, and $y_i$ is ground-truth for the $i$ -th triplet (e.g. a string for VQA, bounding boxes for object detection). We start with the frozen language model $\\pi_\\theta(p \\mid q)$ , where $p$ is a synthesized program and $q$ is a textual query. The language model $\\pi_\\theta$ represents our policy. We generate a dataset of trajectories $\\mathcal{D}_g$ by sampling many programs $p$ from the current policy $\\pi_\\theta: p \\sim \\pi_\\theta(p \\mid q)$ for $q \\sim \\mathcal{D}$ .", + "bbox": [ + 75, + 703, + 470, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Improve Step Our goal in this step is to use the dataset of synthetic programs $\\mathcal{D}_g$ to improve the policy $\\pi_{\\theta}$ . First, we", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "define a binary-valued reward function $R: p, v, y \\to \\{0, 1\\}$ on a given program, image, annotation triplet,", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR (v, p, y) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\phi (p, v) = y \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 131, + 893, + 171 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi(p, v)$ is the result of executing the program $p$ on an image $v$ . Note that $y$ is not a program but an existing annotation such as a string for VQA for a bounding box for object detection. To apply behavioral cloning, we then minimize the reward-weighted loss", + "bbox": [ + 498, + 181, + 893, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\theta) = \\mathbb {E} _ {(q, p) \\sim \\mathcal {D} _ {g}} [ R (v, p) \\mathcal {L} (p, q; \\theta) ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 268, + 892, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(p,q;\\theta)$ is the negative log-likelihood loss", + "bbox": [ + 498, + 296, + 834, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N L L}} (p, q; \\theta) = - \\mathbb {E} _ {(q, p) \\sim \\mathcal {D} _ {g}} \\left[ \\sum_ {t = 1} ^ {T} \\log \\pi_ {\\theta} \\left(p _ {t} \\mid p _ {1: t - 1}, q\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 321, + 892, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "over the pairs of textual queries $q$ and synthetic programs $p$ in $\\mathcal{D}_g$ .", + "bbox": [ + 498, + 377, + 892, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because the reward function only takes on binary values, we can simplify this and implement it by: First, generating a dataset of synthetic programs $\\mathcal{D}_g = \\{\\pi_\\theta (q):\\forall q\\in \\mathcal{D}\\}$ using the LLM $\\pi_{\\theta}$ on a dataset $\\mathcal{D}$ . Next, filtering $\\mathcal{D}_g$ to obtain $\\mathcal{D}_g^\\prime = \\{(q,v,p\\in \\mathcal{D}_g:R(q,v,p) > 0\\}$ , which corresponds to executing all synthetic programs and only keeping those that give correct answers. Finally, we finetune the language model $\\pi_{\\theta}$ on the filtered dataset $\\mathcal{D}_g^\\prime$ using the standard causal language modeling loss. We then iterate the process, initiating a new synthetic data generation step with the improved policy $\\pi_{\\theta}^{\\prime}$ .", + "bbox": [ + 496, + 407, + 892, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Iteration For the initial grow step, we use a frozen language model as the initial policy. For example, we use the pretrained codellama-7b-instruct-hf as the policy in the initial grow step. In subsequent steps, we use the policy trained in the previous improve step for the grow step.", + "bbox": [ + 496, + 573, + 893, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Understanding Self-Training", + "text_level": 1, + "bbox": [ + 500, + 661, + 766, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal in this section is to characterize the stability and sample efficiency of VisReP. We want to understand:", + "bbox": [ + 498, + 686, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. How does applying VisReP change the accuracy of synthesized programs?", + "2. What happens as VisReP is repeated?", + "3. How does data scarcity and diversity affect VisReP?" + ], + "bbox": [ + 500, + 717, + 893, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Implementation", + "text_level": 1, + "bbox": [ + 500, + 786, + 658, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start off with the GQA [13] dataset for visual question answering. We choose GQA because each question in GQA was constructed programatically and is thus a good candidate to be answered by program synthesis. GQA has over 2M questions, each belonging to one of $\\approx 100$ question types. We construct a training set by sampling 100 questions for", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14347", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7968f0e6cdf62333acfd91d8a7b4b75df1392b20dd62ccb7564916bd4920de7.jpg", + "image_caption": [ + "Figure 3. Self-training with VisReP produces qualitatively better programs. Here, we show programs written by the initial policy (on the left) and the policy after 10 iterations of self-training on GQA (on the right). In VQA example, the initial policy does not specifically check whether the empty basket is plastic. In the object detection example, the reasoning of the initial policy is correct, but it issues a confusingly worded query to the simple-query module, which returns the wrong answer. The learned policy uses simple-query more appropriately. In the image-text matching example, in the initial policy tries to use the object detector to search directly for \"meat in a box\" and \"donuts on a plate\", but this is too complicated for the object detector to localize. After self-training, the LLM policy no longer makes this mistake." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each question type, for a total of $\\approx 10\\mathrm{k}$ visual questions and answers. We construct a validation set following Gupta and Kembhavi [9]. We use the CodeLlama [22] family of models as our initial policy. We use LoRA [12] adapters during the Improve steps. We use the hyperparameters suggested by", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dettmers et al. [5]. Full implementation details are in the supplement.", + "bbox": [ + 498, + 824, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14348", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8cfa5e6436ce0085f6cd0d2bc693ad92fd2c51e33bdddf37c6bb8bbcb3abfb64.jpg", + "image_caption": [ + "$\\Delta$ Training Accuracy from Step 1 to Step 10 of Self-Training" + ], + "image_footnote": [], + "bbox": [ + 76, + 103, + 893, + 327 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fbd95f5ddc46e3bf5e646ac0c4a9db7effe1c6e91cf7b22792dd7e5b3aaf9425.jpg", + "image_caption": [ + "Figure 4. Iteratively applying VisReP allows a LLM to self-improve improve on almost all of GQA's $\\approx 100$ question types. The base of each bar is set to the accuracy of the initial policy (codellama-7b-instruct). A green bar indicates question types on which the policy at iteration 10 improved over the initial policy, and a red bar indicates question types on which the policy at iteration 10 was worse than the initial policy.", + "Figure 5. Supplying a small amount of human written corrections as in-context examples during training can increase the stability of the self-training process (green line). We show validation accuracy on GQA through multiple iterations of self-training with a policy instantiated from CodeLlama-7b. Without these corrections, proliferating errors cause performance to degrade in later iterations (red line). The translucent shading around each line indicates the standard deviation over 5 evaluations on the validation set." + ], + "image_footnote": [], + "bbox": [ + 76, + 401, + 472, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Persistent Errors Harm Iterated Self-Training", + "text_level": 1, + "bbox": [ + 75, + 724, + 468, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Applying the formulation of self-training in Sec. 3.2 results in a improvement, but iterating it further results in program synthesis quality degrading, rather than increasing (red line in Fig. 5). This is due to the self-training process inadvertently reinforcing incorrect reasoning. A program that uses flawed reasoning can occasionally produce a correct answer. The language model can thus be rewarded for a program that is right for the wrong reasons. If this goes uncorrected, the language model will learn incorrect reasoning patterns.", + "bbox": [ + 75, + 750, + 472, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We hypothesize that providing a small number of human-", + "bbox": [ + 96, + 885, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "written corrections for persistent reasoning errors can stabilize the self-training process. We use the question type annotations in GQA to identify question types for which training accuracy decreases over time. These are question types which the language model is not able to self-improve on. We denote them $\\mathcal{Q}_{\\text {hard }}$ . For each question type in $\\mathcal{Q}_{\\text {hard }}$ , we randomly sample one question $q$ for which the language model synthesized a program that produced the wrong answer. We examine the reasoning in that program, and if the reasoning is flawed, we correct it. We repeat this until we have a program with correct reasoning for each question type in $\\mathcal{Q}_{\\text {hard }}$ , and denote the bank of correct programs as $\\mathcal{P}_{\\text {gold }}$ .", + "bbox": [ + 496, + 406, + 893, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then retrieve from $\\mathcal{P}_{\\text{gold}}$ during self-training for use as in-context examples. If a question is annotated with a question type in $Q_{\\text{hard}}$ , we retrieve a correct human-written program from $\\mathcal{P}_{\\text{gold}}$ and use it as an in-context example. If a question is not annotated with a question type in $Q_{\\text{hard}}$ , we use a \"default\" in-context example which is the same for all question types not in $Q_{\\text{hard}}$ . We show in Fig. 5 (green line) that this stabilizes self-training and allows the language model to self-improve across all but a few question types (Fig. 4).", + "bbox": [ + 496, + 588, + 895, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Effect of Data Availability on Self-Training", + "text_level": 1, + "bbox": [ + 498, + 753, + 864, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training With Less Data We explore this in a controlled setting, by manipulating the number of samples per question type in GQA. Recall that we originally sample 100 questions per question type for self-training. This dataset had $\\approx 10\\mathrm{k}$ questions. We construct a training set with only 10 and 1 question per question type, for a total of $\\approx 1000$ and $\\approx 100$ questions respectively. Self-training improves upon the baseline (Fig. 6) even when there is an order of magni", + "bbox": [ + 496, + 779, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14349", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2111617b308482b9c5768a3544789a5eeaae4588bdf5132519f2602b6ac8d062.jpg", + "image_caption": [ + "Figure 6. VisReP works even when the amount of available data is reduced by an order of magnitude. We show validation accuracy on GQA. The notation $n \\times k$ indicates $n$ samples per question type, with $k$ passes at each sample. For example $10 \\times 10$ indicates 10 samples per question type, with 10 passes per sample. Although $10 \\times 10$ has 10x fewer unique samples than $100 \\times 1$ , there is a $< 2\\%$ accuracy difference between them, indicating that more passes per instance can partially mitigate data scarcity." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tude decrease in training data $(100\\to 10)$ . Only when the amount of available training data is reduced by two orders of magnitude $(100\\to 1)$ does self-training fail to produce an appreciable increase in performance.", + "bbox": [ + 75, + 411, + 468, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is it possible to mitigate data scarcity? We previously showed that the benefits of self-training reduce when available data is reduced significantly. We now test whether we can mitigate this data scarcity by allowing $\\pi_{\\theta}$ multiple attempts at a query $q$ during the Grow step. Concretely, we allow $\\pi_{\\theta}$ a total of 10 tries at each query under the setting in which we train with 1 and 10 samples per question type, for a total of 1k and 10k total samples respectively. We show in Fig. 6 that this mitigates the effect of reduced data. Although the data poor $1 \\times 10$ and $10 \\times 10$ have 10x fewer unique questions than $10 \\times 1$ and $100 \\times 1$ , their performance is within a standard deviation of their data rich counterparts.", + "bbox": [ + 75, + 472, + 470, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Quantifying Changes in Syntactic Structure", + "text_level": 1, + "bbox": [ + 76, + 664, + 449, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How do the programs synthesized by the policy change as self-training is iterated? We examine this by looking at how many unique abstract syntax trees are produced during the Grow step of each iteration. We parse the synthesized programs into abstract syntax trees, and then normalize the trees to remove irrelevant details such as variable names. In the left panel of Fig. 7, we show that the diversity of syntactic forms drops over time. At the beginning, the policy produces a large number of syntactic forms, but appears to \"hone in\" on a smaller number of forms as self-training continues, and the number of unique syntactic forms drops by almost half.", + "bbox": [ + 75, + 688, + 470, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A remarkably stable set of syntactic forms is conserved from step to step, roughly $\\approx 700$ (row above diagonal in right panel of Fig. 7). However, the syntactic forms produced", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/653672caefc87f0f4ac555f55469f762680e7f4462711d859412d2035775af08.jpg", + "image_caption": [ + "Figure 7. As self-training is iterated, the LLM policy \"hones in\" on a smaller set of syntactic forms, and gradually evolves away from syntactic forms produced by the initial policy. Left Panel: Number of unique normalized abstract syntax trees seen during each iteration of VisReP. Right Panel: Number of unique normalized abstract syntax trees in common between each training step. For example, the entry in row 1, column 6 corresponds to the number of unique abstract syntax trees produced by both the policy in iteration 1 (initial policy) and the policy in iteration 6." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 694, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14447310f7b76e4d930e540b76eb8d418796293d5a9cfbf024de25a3450da451.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 95, + 890, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by the policy are gradually evolving away from the syntactic forms the initial policy tries, which can be seen in the darkening of the first row in Fig. 7. Despite the coarse reward scheme, the LLM policy gradually explores and learns new syntactic forms.", + "bbox": [ + 496, + 402, + 893, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Evaluating Functional Correctness", + "text_level": 1, + "bbox": [ + 498, + 503, + 815, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We measure the functional correctness of the programs synthesized by the self-trained LLM policy $\\pi_{\\theta}^{\\prime}$ across three compositional tasks, with the aim of understanding whether:", + "bbox": [ + 496, + 532, + 893, + 577 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Are the programs produced after self-training more functionally correct than programs produced before self-training?", + "2. Is it possible to exceed or match the performance of a much larger proprietary LLM with self-training?" + ], + "bbox": [ + 500, + 580, + 893, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For compositional VQA, we use the GQA [13] dataset for the reasons outlined in Sec. 4.1. For complex object detection, we choose Omnilabel [24]. Omnilabel contains 28K free-form object descriptions over 25K images, and is a challenging task for existing open-vocabulary object detectors due to the complexity of the object descriptions. For compositional image-text matching, we choose WinoGround [29] and SugarCrepe [11]. State-of-art vision-language models have trouble reaching above chance accuracy on WinoGround, but SugarCrepe is substantially easier. However, both of these tasks pose significant problems for the ImagePatch API, because many of the relationships mentioned in the text are challenging to detect with the available perception modules. For all experiments, we use ViperGPT[28] as the backbone and adopt their prompts. Due to space limitations, many experimental details are in the supplement.", + "bbox": [ + 496, + 659, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14350", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a57b71079d1d969b86a044c99f65aee528999a49d92ee26451677be6270833bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VQAObject DetectionImage-Text Matching
MethodLLMGQAOmnilabelOmnilabel-HardWinogroundSugarCREPE
Frozen Proprietary LLMGPT-3.5-turbo53.9 ± 0.840.0 ± 1.226.0 ± 1.145.6 ± 1.648.9 ± 0.7
Frozen Open LLMCodeLlama-7B50.0 ± 1.737.3 ± 1.623.7 ± 1.441.3 ± 0.0343.5 ± 0.8
Open LLM + VisRePCodeLlama-7B59.2 ± 1.442.4 ± 1.028.1 ± 0.952.7 ± 0.658.7 ± 1.5
", + "bbox": [ + 81, + 88, + 890, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/449434d9b003b48ca66099610c2821c651a80366bf11760d1b50f91907a02f66.jpg", + "table_caption": [ + "Table 2. An open LLM policy self-trained with our method substantially outperforms the open policy without self-training, and even outperforms a gpt-3.5-turbo policy. All results use ViperGPT [28] as the backbone. $\\pm$ numbers are the standard deviation over 5 runs. On all datasets except Omnilabel, we report accuracy. On Omnilabel, we report Macro-F1. Higher is better." + ], + "table_footnote": [], + "table_body": "
Code GeneratorX-Dataset GeneralizationX-Task Generalization
VQAv2OK-VQAOmnilabelSugarCrepe
Frozen Few-Shot (7B)46.6 ± 1.112.6 ± 2.137.3 ± 1.643.5 ± 0.8
VisReP on GQA (7B)61.0 ± 1.033.7 ± 1.939.9 ± 0.751.0 ± 1.5
VisReP Advantage+14.4+21.1+2.6+7.5
", + "bbox": [ + 80, + 253, + 465, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. VisReP improves benchmark agnostic visual program synthesis. A policy self-trained on GQA with VisReP writes better programs for other VQA datasets and other task types.", + "bbox": [ + 78, + 339, + 468, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 80, + 402, + 264, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For each task, we apply VisReP as described in Sec. 3.2, and evaluate on a held-out subset. For a comparison with a large proprietary LLM, we use gpt-3.5-turbo. We evaluate on a subsampled version of each dataset to reduce token costs. Every LLM is provided the same prompts. Each prompt consists of the ImagePatch API specification used in ViperGPT [28], and 3 in-context examples for each task except for object detection, for which we provide 5 in-context examples.", + "bbox": [ + 78, + 426, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use GQA as described in Sec. 4.1. We prepare a compositional subset of Omnilabel [24] by filtering out all descriptions less than two words in length. We then sample a subset of 500 for evaluation, and a subset of 500 for training. To prepare Omnilabel-Hard, we use run a state of the art open-vocabulary object detector (GroundingDINO [18]) on the remaining OmniLabel samples, and select those which GroundingDINO completely fails on (no detections) to obtain a hard slice. We then sample a subset of 500 from the hard slice for evaluation. For SugarCrepe [11], we sample 100 positives and their associated negatives from each of the 6 categories, for a total of 600 balanced image-text pairs for validation. We sample 100 of the remaining instances from each category for training. We use all of WinoGround, as it is small enough that there is no need to subsample it. On WinoGround[29], we evaluate the policy trained on SugarCrepe rather than training on it. For VQAv2, we sample 10 questions for each of the top-50 most common answers from the compositional subset curated by [25].", + "bbox": [ + 78, + 550, + 468, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Examples of the inputs for each task are in Fig. 3. We use nucleus sampling with identical parameters for all local LLMs. We use the API default temperature for gpt-3.5-turbo. More details are in the supplement.", + "bbox": [ + 78, + 839, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Discussion", + "text_level": 1, + "bbox": [ + 503, + 266, + 614, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Across all three tasks, the policy trained by VisReP outperforms both the gpt-3.5-turbo policy, and the initial CodeLlama-7b policy (Tab. 2). On GQA, the self-trained policy achieves an absolute improvement of almost $9\\%$ over the initial policy, and $5\\%$ over the gpt-3.5-turbo policy. On Omnilabel, self-training produces a $5\\%$ improvement in Macro-F1 score with only 500 training samples. On Omnilabel-Hard, we demonstrate that the visual program synthesis paradigm can localize objects that state of the art open-vocabulary object detectors are unable to localize (Omnilabel-Hard was constructed by selecting instances GroundingDino[18]) cannot localize). Even on Omnilabel-Hard, the self-trained policy outperforms the others. WinoGround and SugarCrepe are difficult to solve by visual program synthesis because many of the relationships are hard to detect with the available perception modules. Despite the intrinsic difficulty of compositional image-text matching for the ImagePatch API, VisReP produces an increase of $+15\\%$ over the baseline policy. The policy trained on SugarCrepe transfers to WinoGround, outperforming the baseline policy by $+10\\%$ .", + "bbox": [ + 501, + 289, + 892, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion & Future Work", + "text_level": 1, + "bbox": [ + 503, + 604, + 751, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While few-shot prompting of LLMs for visual program synthesis has produced impressive results, it has limitations, because writing good visual programs requires experience with the visual world and the perception modules at ones disposal. We presented VisReP, which improves a LLM's program synthesis abilities using feedback from executing visual programs. We showed that VisReP produces strong increases over baseline across multiple tasks, and is competitive with gpt-3.5-turbo. Our work constructed a coarse-valued reward from existing vision-language annotations. Methods like RLAIF [2], ReST [8], and CodeRL [16] all rely on a neural reward model that can provide fine-grained rewards. Learning from fine-grained rewards is much easier than learning from coarse rewards. An interesting direction for future work would be to train a neural reward model for visual program synthesis. Such a reward model could provide fine-grained rewards, and open a broader range of reinforcement learning methods.", + "bbox": [ + 501, + 628, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14351", + "bbox": [ + 480, + 946, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V. Le, and Charles Sutton. Program synthesis with large language models. ArXiv, abs/2108.07732, 2021. 1", + "[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, E Perez, Jamie Kerr, Jared Mueller, Jeff Ladish, J Landau, Kamal Ndousse, Kamile Lukovs.iuté, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noem'i Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, T. J. Henighan, Tristan Hume, Sam Bowman, Zac Hatfield-Dodds, Benjamin Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom B. Brown, and Jared Kaplan. Constitutional ai: Harmlessness from ai feedback. ArXiv, abs/2212.08073, 2022. 8", + "[3] Natasha Butt, Blazej Manczak, Auke J. Wiggers, Corrado Rainone, David W. Zhang, Michael Defferrard, and Taco Cohen. Codeit: Self-improving language models with prioritized hindsight replay. ArXiv, abs/2402.04858, 2024. 3", + "[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 2", + "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. ArXiv, abs/2305.14314, 2023. 5", + "[6] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 2, 3", + "[7] Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. VQA-LOL: visual question answering under the lens of logic. In ECCV, 2020. 1", + "[8] Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alexa Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, A. Doucet, Orhan First, and Nando de Freitas. Re-" + ], + "bbox": [ + 86, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inforced self-training (rest) for language modeling. ArXiv, abs/2308.08998, 2023. 2, 3, 4, 8", + "[9] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. ArXiv, abs/2211.11559, 2022. 1, 2, 3, 5", + "[10] Patrick M. Haluptzok, Matthew Bowers, and Adam Tauman Kalai. Language models can teach themselves to program better. ArXiv, abs/2207.14502, 2022. 2, 3", + "[11] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 7, 8", + "[12] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 5", + "[13] Drew A. Hudson and Christopher D. Manning. GQA: A new dataset for real-world visual reasoning and compositional question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6700-6709. Computer Vision Foundation / IEEE, 2019. 4, 7", + "[14] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[15] Sascha Lange, Thomas Gabel, and Martin Riedmiller. Batch Reinforcement Learning, pages 45-73. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 4", + "[16] Hung Le, Yue Wang, Akhilesh Deepak Gotmare, Silvio Savarese, and Steven Hoi. CodeRL: Mastering code generation through pretrained models and deep reinforcement learning. In Advances in Neural Information Processing Systems, 2022. 2, 3, 8", + "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3", + "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 8", + "[19] Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 3", + "[20] Yujia Qin, Shi Liang, Yining Ye, Kunlun Zhu, Lan Yan, Ya Ting Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Marc H. Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis. ArXiv, abs/2307.16789, 2023. 2, 3", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14352", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 3", + "[22] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, I. Evtimov, Joanna Bitton, Manish P Bhatt, Cristian Cantón Ferrer, Aaron Grattafori, Wenhan Xiong, Alexandre D'efossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. Code llama: Open foundation models for code. ArXiv, abs/2308.12950, 2023. 5", + "[23] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761, 2023. 2, 3", + "[24] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 1, 7, 8", + "[25] Ramprasaath R. Selvaraju, Purva Tendulkar, Devi Parikh, Eric Horvitz, Marco Tulio Ribeiro, Besmira Nushi, and Ece Kamar. Squinting at vqa models: Introspecting vqa models with subquestions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8", + "[26] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Ke- nely, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell L. Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem- solving with language models. ArXiv, abs/2312.06585, 2023. 3", + "[27] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 747-761, Toronto, Canada, 2023. Association for Computational Linguistics. 1, 3", + "[28] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 7, 8", + "[29] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 7, 8", + "[30] Spencer Whitehead, Suzanne Petryk, Vedaad Shakib, Joseph Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rohrbach. Reliable visual question answering: Abstain rather than answer incorrectly. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[31] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229-256, 1992. 4", + "[32] Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations, 2022. 1", + "[33] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 3" + ], + "bbox": [ + 501, + 92, + 893, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14353", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_model.json b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8ef64e3db3168ca6ba5f89a083f023cc1fe7126 --- /dev/null +++ b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_model.json @@ -0,0 +1,1892 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.131, + 0.88, + 0.175 + ], + "angle": 0, + "content": "Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.203, + 0.854, + 0.221 + ], + "angle": 0, + "content": "Zaid Khan\\(^{1}\\) Vijay Kumar BG\\(^{2}\\) Samuel Schulter\\(^{2}\\) Yun Fu\\(^{1}\\) Manmohan Chandraker\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.223, + 0.71, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Northeastern University \\(^{2}\\)NEC Laboratories America \\(^{3}\\)UC San Diego" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.306, + 0.474, + 0.622 + ], + "angle": 0, + "content": "Visual program synthesis is a promising approach to exploit the reasoning abilities of large language models for compositional computer vision tasks. Previous work has used few-shot prompting with frozen LLMs to synthesize visual programs. Training an LLM to write better visual programs is an attractive prospect, but it is unclear how to accomplish this. No dataset of visual programs for training exists, and acquisition of a visual program dataset cannot be easily crowdsourced due to the need for expert annotators. To get around the lack of direct supervision, we explore improving the program synthesis abilities of an LLM using feedback from interactive experience. We propose a method where we exploit existing annotations for a vision-language task to improvise a coarse reward signal for that task, treat the LLM as a policy, and apply reinforced self-training to improve the visual program synthesis ability of the LLM for that task. We describe a series of experiments on object detection, compositional visual question answering, and image-text retrieval, and show that in each case, the self-trained LLM outperforms or performs on par with few-shot frozen LLMs that are an order of magnitude larger. Website:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.624, + 0.328, + 0.637 + ], + "angle": 0, + "content": "https://zaidkhan.me/ViReP" + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.664, + 0.209, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Complex visual queries can often be decomposed into simpler subtasks, many of which can be carried out by task-specific perception modules (e.g. object detection, captioning). For example, consider the problem of finding bounding boxes for the phrase \"white mug to the left of the sink\". This is a challenging query for single model such as an open vocabulary object detector. However, this query can be solved by writing a program that composes task-specific perception modules with logic: use an open vocabulary object detector to find a sink and white mugs in the scene, then compare the horizontal center of the sink and the mugs to find white mugs to the left of the sink. Program synthesis with large language models [1] is a promising approach to automate this process, and recent work has shown that proprietary large language" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.274, + 0.891, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.424, + 0.885, + 0.436 + ], + "angle": 0, + "content": "How can we train a LLM to write better visual programs for a task?" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.436, + 0.865, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.556, + 0.895, + 0.641 + ], + "angle": 0, + "content": "Figure 1. Visual program synthesis with LLMs has been treated as a \\(0 / \\mathrm{n}\\)-shot task where the LLM is kept frozen. This limits opportunities for improvement. We ask whether it is possible to train a LLM to write more accurate programs. Given that there is no large scale dataset of accurate visual programs available, we propose improving the LLM using self-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.895, + 0.734 + ], + "angle": 0, + "content": "models can write programs for visual tasks [9, 27, 28]. Current approaches for visual program synthesis with LLMs use few-shot prompting and rely on the in-context learning abilities [32] of frozen, proprietary LLMs. (Fig. 1)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Few-shot prompting with frozen LLMs for visual program synthesis as in ViperGPT [28], VisProg [9], or CodeVQA [27] has several limitations. The LLM needs to understand the competencies of the perception modules it is using. A open vocabulary object detector may be able to locate a common attribute-noun phrase such as \"white mug\" without problems, but struggle with a more abstract phrase such as \"microwaveable mug\" [24]. A VQA model might be able to answer \"is the car blue?\" without problems, but fail when logical modifiers are introduced, such as \"is the car not blue?\" [7]. In many cases, we do not precisely know" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14344" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.189 + ], + "angle": 0, + "content": "
Task DomainSelf-TrainingSupervisionTool/API UseVisual Task DecompositionGrounded By FeedbackImproves LLM
VisReP (Ours)Visual Program SynthesisYesWeakYesYesYesYes
Haluptzok et al. [10]Programming PuzzlesYesWeakNoNoYesYes
ReST [8]Natural Language UnderstandingYes-NoNoYesYes
VisProg [9]Visual Program SynthesisNo-YesYesNoNo
ViperGPT [28]Visual Program SynthesisNoWeakYesYesNoNo
ToolLLM [20]Tool Usage by APINoStrongYesNoNoYes
GorillaLLM [19]Tool Usage by APINoStrongYesNoNoYes
ToolFormer [23]Natural Language UnderstandingYesWeakYesNoYesYes
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.2, + 0.893, + 0.298 + ], + "angle": 0, + "content": "Table 1. Differences between our work and similar work. Strong supervision means that the training process requires examples of ground-truth programs to train the LLM. Weak supervision means that the training process does not require ground-truth programs. Tool / API use means that the LLM is required to use substantial functionality implemented by external modules (e.g. an object detector, a web search API) to solve tasks. Visual task decomposition means that the LLM can decompose a complex visual task into primitive subtasks. Grounded by feedback means that the LLM has been optimized not just for syntactic / semantic correctness (program does not hallucinate / cause errors), but for functional correctness (programs produce the correct answer). Improves LLM means that the work proposes a method to improve an LLM for a specific task, rather than using a frozen LLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.471, + 0.471 + ], + "angle": 0, + "content": "the weaknesses or competencies of a perception model [30]. Even if this were known, it is difficult to convey all of the competencies / weaknesses of a perception module through in-context examples. Second, it is the case in program synthesis that an LLM often can generate the correct solution to a problem, but the correct solution is not the solution the LLM places the highest probability on [4]. We would like to align the LLM to \"uncover\" the knowledge of the correct solution, but it not clear how to do this in a principled way with few-shot prompting alone." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.48, + 0.436, + 0.51 + ], + "angle": 0, + "content": "How can we train a large language model to write better visual programs for a specific task?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Our goal is to optimize the parameters of the language model so the accuracy of the synthesized programs is higher. Existing approaches that train LLMs to improve their ability to programatically use tools / APIs such as GorillaLLM [19], ToolLLM [20] do so by finetuning LLMs on examples of tool use or API use. This cannot be directly applied to visual program synthesis because there are no large scale datasets of visual programs, and collecting such a dataset would be extremely labor intensive. In the absence of a large scale dataset, how do we learn to write better programs for a visual task?" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.696, + 0.436, + 0.755 + ], + "angle": 0, + "content": "We posit that grounding a language model with interactive feedback from a generic visual task will improve the general visual program synthesis abilities of the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "A natural way to learn from feedback is to use reinforcement learning. ReST [8] and RaFT [6] introduce a general framework for reinforced self-training in generative tasks and demonstrate success in machine translation and text-to-image generation. However, a crucial ingredient in their recipe is the availability of a fine-grained reward model. It is difficult to construct a fine-grained reward model for visual program synthesis, given both the absence of human preference datasets for visual programs, and the difficulty" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.562 + ], + "angle": 0, + "content": "of devising a proxy metric. One alternative is to use unit tests to teach a neural reward model or give a coarse-grained reward. This technique has been used successfully in coding challenges by CodeRL [16] and Haluptzok et al. [10], but it is unclear how it can be applied to visual program synthesis. Our key idea is to use existing annotations for a vision-language as improvised unit tests to provide a coarse reward signal. Using the coarse reward signal, we can apply reinforced self-training by treating the language model as a policy and training it with a simple policy gradient algorithm. We alternate synthetic data generation steps in which we sample programs from the language model policy with optimization steps in which we improve the language model policy based on observations from executing the sampled programs. We name our proposed method VisReP, for Visually Reinforced Program Synthesis." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.573, + 0.892, + 0.617 + ], + "angle": 0, + "content": "- We propose optimizing the parameters of a LLM so that the accuracy of the synthesized visual programs is higher, in contrast to previous works that use frozen LLMs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.618, + 0.892, + 0.678 + ], + "angle": 0, + "content": "- Since no dataset of accurate visual programs is available for finetuning, we hypothesize that we can instead use feedback from the execution environment to improve the visual program synthesis abilities of a language model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.679, + 0.892, + 0.738 + ], + "angle": 0, + "content": "- We propose VisReP, an offline, model agnostic recipe for reinforced self-training of large language models for visual program synthesis using existing vision-language annotations with a simple policy gradient algorithm." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.739, + 0.892, + 0.784 + ], + "angle": 0, + "content": "- Our results show that it is possible to apply reinforced self-training for to improve large language models for visual program synthesis with only coarse rewards." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.573, + 0.892, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.9 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of an CodeLlama-7B policy trained by VisReP on compositional visual question answering \\((+9\\%)\\), complex object detection \\((+5\\%)\\), and compositional image-text matching \\((+15\\%)\\) relative to the untrained policy. We show that the policy trained by VisReP exceeds the accuracy of a gpt-3.5-turbo policy on all three tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14345" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.218, + 0.132 + ], + "angle": 0, + "content": "2.1. Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.199 + ], + "angle": 0, + "content": "Self-training is an established paradigm which uses unlabeled data to improve performance. Self-training has been successfully applied in a number of fields. We restrict our coverage to usages with significant overlap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.472, + 0.365 + ], + "angle": 0, + "content": "Program Synthesis Haluptzok et al. [10] showed that LLMs can improve their program synthesis abilities by generating programming puzzles and solving them. CodeRL [16] proposed an actor-critic framework to improve the program synthesis abilities of LLMs for programming problems accompanied by unit tests. CodeIT [3] and Rest-EM [26] also use a similar policy gradient approach for program synthesis. Our problem domain is different from these works, which focus on program synthesis for programming puzzles / problems. In addition, our work has an explicit focus on learning to use an API fluently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.501 + ], + "angle": 0, + "content": "Alignment ReST [8] and RaFT [6] introduced a generic framework for reinforced self-training and applied it to align machine translation outputs to human preferences and align foundation models on language understanding and image generation tasks respectively. These works share the same basic idea as our work, though they are in a substantially different task domain where human preferences are either known (conversational alignment) or can be estimated with an available neural model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.579 + ], + "angle": 0, + "content": "Vision-Language SelTDA [14] introduced a self-training approach for visual question answering. SelTDA proceeds by pseudolabeling unlabeled data, then finetuning a large VLM on the pseudolabeled data. In contrast to SelTDA, we improve a LLM for visual program synthesis." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.587, + 0.313, + 0.603 + ], + "angle": 0, + "content": "2.2. Visual Program Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.472, + 0.793 + ], + "angle": 0, + "content": "Visual program synthesis with LLMs was proposed concurrently by ViperGPT [28], VisProg [9], and CodeVQA [27]. The common points between these three works is that (a) they use pretrained LLMs as code generators (b) they represent complex visual tasks as compositions of primitive visual subtasks (c) they use code to invoke task-specific models to perform the primitive subtasks. Our work is most similar to ViperGPT and CodeVQA as they produce code in a general purpose programming language rather than a DSL. All three works use a proprietary, frozen LLM. In contrast to all three, the focus of our work is on how we can improve the visual program synthesis abilities of an open LLM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.271, + 0.817 + ], + "angle": 0, + "content": "2.3. Tool Use with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Multimodal tool-using LLMs were first introduced by Socratic Models [33]. However, their approach was to create fixed pipelines in which the output of a perception model such as CLIP [21] is fed to a LLM. Later approaches such as GorillaLLM [19] and ToolLLM [20] improved on this by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "treating tool use as a program synthesis problem and creating LLMs that use a broad range of tools by learning to invoke APIs. However, one key limitation of these approaches in the context of visual program synthesis is that they do not learn to decompose problems into subproblems that can be solved by tools. Instead, they are trained to select the right tool for the problem and invoke it. Another limitation is that they are not optimized for functional correctness. They are trained for syntactic and semantic correctness, but they have not been provided feedback on whether their use of tools produces the desired answer. ToolFormer [23] is similar to our work in the sense that the LLM's usage of tools is grounded by feedback, but they focus on natural language understanding tasks rather than visual tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.593, + 0.333 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.342, + 0.826, + 0.359 + ], + "angle": 0, + "content": "3.1. Visual Program Synthesis with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.894, + 0.458 + ], + "angle": 0, + "content": "Task Formulation Let \\( v \\) be a visual input and \\( q \\) be a textual query about \\( v \\). In visual program synthesis, we synthesize a program \\( p = \\pi_{\\theta}(q) \\) with a program generator \\( \\pi_{\\theta} \\). The program \\( p \\) and visual input \\( q \\) are then fed into the execution engine \\( \\hat{y} = \\phi(v, p) \\) to produce a result \\( \\hat{y} \\). The program generator is an auto-regressive large language model" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.468, + 0.893, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\theta} (\\boldsymbol {y} \\mid \\boldsymbol {x}) = \\prod_ {t = 1} ^ {T} \\pi_ {\\theta} \\left(p _ {t} \\mid \\mathbf {p} _ {1: t - 1}, \\boldsymbol {x}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.894, + 0.595 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{1:t}\\) are the tokens of the program, and \\(x\\) is the input to the large language model. The language model is kept frozen in previous work [28]. Our goal is to optimize the parameters \\(\\theta\\) of the language model \\(\\pi\\) so the accuracy of the synthesized programs is higher." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.792 + ], + "angle": 0, + "content": "Implementation Following ViperGPT [28], we provide the specification of the ImagePatch API concatenated with the textual query \\( q \\) as the prompt to the program generator. The synthesized program \\( p \\) is a Python program that can invoke any Python builtins, control flow structures, and the ImagePatch API. Our implementation of the ImagePatch API is largely similar to ViperGPT. We remove some API methods that were not required for the tasks we evaluate on (such as l1m_query). We use BLIP [17] and GroundingDINO [18] as perception modules underlying find (object detection), simple-query (visual question answering), and verify_property (attribute verification)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.727, + 0.818 + ], + "angle": 0, + "content": "3.2. Reinforced Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Rather than use a frozen large language model as the program generator \\( p_{\\theta} \\), we would like to optimize the parameters \\( \\theta \\) of the language model so the accuracy of the synthesized programs is higher. It is not obvious how to do this. We can't backpropagate through the execution engine \\( \\phi(\\pi_{\\theta}(q), v) \\) to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14346" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.458, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.186, + 0.362, + 0.201 + ], + "angle": 0, + "content": "Offline Policy Improvement" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.212, + 0.465, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.432 + ], + "angle": 0, + "content": "Figure 2. VisReP can be applied to improve the visual synthesis abilities of an LLM for a vision-language task using existing annotations for a vision-language task (e.g. an object description + image + bounding boxes). A key idea is to construct a coarse reward by comparing the answer produced by a synthesized program to the ground-truth answer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.567 + ], + "angle": 0, + "content": "directly optimize \\(\\theta\\) with respect to \\(q\\) or \\(v\\). An alternative might be to use human labor to build a dataset of high-quality visual programs, and train the large language model \\(\\pi_{\\theta}\\) on the manually-collected dataset. But collecting such a dataset is very labor intensive, and not scalable. Instead, we explore the idea of learning from experience by applying a simple policy gradient method, REINFORCE [31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.704 + ], + "angle": 0, + "content": "We propose VisReP, which treats the program synthesis task as a growing batch RL problem [15], inspired by ReST [8]. We first define a coarse discrete reward function \\( R(\\cdot) \\) from existing annotations for a vision-language task. We then alternate Grow steps, in which we sample trajectories (programs) from the policy (large language model), with Improve steps, in which we apply behavioral cloning with a reward-weighted negative-log likelihood loss to improve the policy. A diagram of our approach is depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.87 + ], + "angle": 0, + "content": "Grow Step The grow step corresponds to the acting step in reinforcement learning, and can also be seen as synthetic data generation. Let \\(\\mathcal{D} = \\{(v_1, q_1, y_1), \\ldots, (v_n, q_n, y_n)\\}\\) be a dataset for a vision-language task, where \\(v_i\\) is an image, \\(q_i\\) is a textual query, and \\(y_i\\) is ground-truth for the \\(i\\)-th triplet (e.g. a string for VQA, bounding boxes for object detection). We start with the frozen language model \\(\\pi_\\theta(p \\mid q)\\), where \\(p\\) is a synthesized program and \\(q\\) is a textual query. The language model \\(\\pi_\\theta\\) represents our policy. We generate a dataset of trajectories \\(\\mathcal{D}_g\\) by sampling many programs \\(p\\) from the current policy \\(\\pi_\\theta: p \\sim \\pi_\\theta(p \\mid q)\\) for \\(q \\sim \\mathcal{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Improve Step Our goal in this step is to use the dataset of synthetic programs \\(\\mathcal{D}_g\\) to improve the policy \\(\\pi_{\\theta}\\). First, we" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "define a binary-valued reward function \\( R: p, v, y \\to \\{0, 1\\} \\) on a given program, image, annotation triplet," + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.132, + 0.894, + 0.172 + ], + "angle": 0, + "content": "\\[\nR (v, p, y) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\phi (p, v) = y \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.183, + 0.894, + 0.258 + ], + "angle": 0, + "content": "where \\(\\phi(p, v)\\) is the result of executing the program \\(p\\) on an image \\(v\\). Note that \\(y\\) is not a program but an existing annotation such as a string for VQA for a bounding box for object detection. To apply behavioral cloning, we then minimize the reward-weighted loss" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.269, + 0.893, + 0.288 + ], + "angle": 0, + "content": "\\[\nJ (\\theta) = \\mathbb {E} _ {(q, p) \\sim \\mathcal {D} _ {g}} [ R (v, p) \\mathcal {L} (p, q; \\theta) ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.297, + 0.836, + 0.313 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(p,q;\\theta)\\) is the negative log-likelihood loss" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.323, + 0.893, + 0.376 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {N L L}} (p, q; \\theta) = - \\mathbb {E} _ {(q, p) \\sim \\mathcal {D} _ {g}} \\left[ \\sum_ {t = 1} ^ {T} \\log \\pi_ {\\theta} \\left(p _ {t} \\mid p _ {1: t - 1}, q\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.378, + 0.893, + 0.408 + ], + "angle": 0, + "content": "over the pairs of textual queries \\( q \\) and synthetic programs \\( p \\) in \\( \\mathcal{D}_g \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.893, + 0.574 + ], + "angle": 0, + "content": "Because the reward function only takes on binary values, we can simplify this and implement it by: First, generating a dataset of synthetic programs \\(\\mathcal{D}_g = \\{\\pi_\\theta (q):\\forall q\\in \\mathcal{D}\\}\\) using the LLM \\(\\pi_{\\theta}\\) on a dataset \\(\\mathcal{D}\\). Next, filtering \\(\\mathcal{D}_g\\) to obtain \\(\\mathcal{D}_g^\\prime = \\{(q,v,p\\in \\mathcal{D}_g:R(q,v,p) > 0\\}\\), which corresponds to executing all synthetic programs and only keeping those that give correct answers. Finally, we finetune the language model \\(\\pi_{\\theta}\\) on the filtered dataset \\(\\mathcal{D}_g^\\prime\\) using the standard causal language modeling loss. We then iterate the process, initiating a new synthetic data generation step with the improved policy \\(\\pi_{\\theta}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.894, + 0.649 + ], + "angle": 0, + "content": "Iteration For the initial grow step, we use a frozen language model as the initial policy. For example, we use the pretrained codellama-7b-instruct-hf as the policy in the initial grow step. In subsequent steps, we use the policy trained in the previous improve step for the grow step." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.662, + 0.767, + 0.68 + ], + "angle": 0, + "content": "4. Understanding Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.687, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Our goal in this section is to characterize the stability and sample efficiency of VisReP. We want to understand:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.718, + 0.894, + 0.748 + ], + "angle": 0, + "content": "1. How does applying VisReP change the accuracy of synthesized programs?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.748, + 0.773, + 0.763 + ], + "angle": 0, + "content": "2. What happens as VisReP is repeated?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.763, + 0.868, + 0.778 + ], + "angle": 0, + "content": "3. How does data scarcity and diversity affect VisReP?" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.718, + 0.894, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.66, + 0.803 + ], + "angle": 0, + "content": "4.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We start off with the GQA [13] dataset for visual question answering. We choose GQA because each question in GQA was constructed programatically and is thus a good candidate to be answered by program synthesis. GQA has over 2M questions, each belonging to one of \\(\\approx 100\\) question types. We construct a training set by sampling 100 questions for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14347" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.891, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.701, + 0.897, + 0.798 + ], + "angle": 0, + "content": "Figure 3. Self-training with VisReP produces qualitatively better programs. Here, we show programs written by the initial policy (on the left) and the policy after 10 iterations of self-training on GQA (on the right). In VQA example, the initial policy does not specifically check whether the empty basket is plastic. In the object detection example, the reasoning of the initial policy is correct, but it issues a confusingly worded query to the simple-query module, which returns the wrong answer. The learned policy uses simple-query more appropriately. In the image-text matching example, in the initial policy tries to use the object detector to search directly for \"meat in a box\" and \"donuts on a plate\", but this is too complicated for the object detector to localize. After self-training, the LLM policy no longer makes this mistake." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "each question type, for a total of \\(\\approx 10\\mathrm{k}\\) visual questions and answers. We construct a validation set following Gupta and Kembhavi [9]. We use the CodeLlama [22] family of models as our initial policy. We use LoRA [12] adapters during the Improve steps. We use the hyperparameters suggested by" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Dettmers et al. [5]. Full implementation details are in the supplement." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14348" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.088, + 0.71, + 0.103 + ], + "angle": 0, + "content": "\\(\\Delta\\) Training Accuracy from Step 1 to Step 10 of Self-Training" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.104, + 0.895, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.337, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Iteratively applying VisReP allows a LLM to self-improve improve on almost all of GQA's \\(\\approx 100\\) question types. The base of each bar is set to the accuracy of the initial policy (codellama-7b-instruct). A green bar indicates question types on which the policy at iteration 10 improved over the initial policy, and a red bar indicates question types on which the policy at iteration 10 was worse than the initial policy." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.402, + 0.473, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.588, + 0.472, + 0.7 + ], + "angle": 0, + "content": "Figure 5. Supplying a small amount of human written corrections as in-context examples during training can increase the stability of the self-training process (green line). We show validation accuracy on GQA through multiple iterations of self-training with a policy instantiated from CodeLlama-7b. Without these corrections, proliferating errors cause performance to degrade in later iterations (red line). The translucent shading around each line indicates the standard deviation over 5 evaluations on the validation set." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.743 + ], + "angle": 0, + "content": "4.2. Persistent Errors Harm Iterated Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.886 + ], + "angle": 0, + "content": "Applying the formulation of self-training in Sec. 3.2 results in a improvement, but iterating it further results in program synthesis quality degrading, rather than increasing (red line in Fig. 5). This is due to the self-training process inadvertently reinforcing incorrect reasoning. A program that uses flawed reasoning can occasionally produce a correct answer. The language model can thus be rewarded for a program that is right for the wrong reasons. If this goes uncorrected, the language model will learn incorrect reasoning patterns." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We hypothesize that providing a small number of human-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.895, + 0.588 + ], + "angle": 0, + "content": "written corrections for persistent reasoning errors can stabilize the self-training process. We use the question type annotations in GQA to identify question types for which training accuracy decreases over time. These are question types which the language model is not able to self-improve on. We denote them \\(\\mathcal{Q}_{\\text {hard }}\\). For each question type in \\(\\mathcal{Q}_{\\text {hard }}\\), we randomly sample one question \\(q\\) for which the language model synthesized a program that produced the wrong answer. We examine the reasoning in that program, and if the reasoning is flawed, we correct it. We repeat this until we have a program with correct reasoning for each question type in \\(\\mathcal{Q}_{\\text {hard }}\\), and denote the bank of correct programs as \\(\\mathcal{P}_{\\text {gold }}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.896, + 0.741 + ], + "angle": 0, + "content": "We then retrieve from \\(\\mathcal{P}_{\\text{gold}}\\) during self-training for use as in-context examples. If a question is annotated with a question type in \\(Q_{\\text{hard}}\\), we retrieve a correct human-written program from \\(\\mathcal{P}_{\\text{gold}}\\) and use it as an in-context example. If a question is not annotated with a question type in \\(Q_{\\text{hard}}\\), we use a \"default\" in-context example which is the same for all question types not in \\(Q_{\\text{hard}}\\). We show in Fig. 5 (green line) that this stabilizes self-training and allows the language model to self-improve across all but a few question types (Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.755, + 0.866, + 0.772 + ], + "angle": 0, + "content": "4.3. Effect of Data Availability on Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Training With Less Data We explore this in a controlled setting, by manipulating the number of samples per question type in GQA. Recall that we originally sample 100 questions per question type for self-training. This dataset had \\(\\approx 10\\mathrm{k}\\) questions. We construct a training set with only 10 and 1 question per question type, for a total of \\(\\approx 1000\\) and \\(\\approx 100\\) questions respectively. Self-training improves upon the baseline (Fig. 6) even when there is an order of magni" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14349" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.088, + 0.473, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.473, + 0.386 + ], + "angle": 0, + "content": "Figure 6. VisReP works even when the amount of available data is reduced by an order of magnitude. We show validation accuracy on GQA. The notation \\( n \\times k \\) indicates \\( n \\) samples per question type, with \\( k \\) passes at each sample. For example \\( 10 \\times 10 \\) indicates 10 samples per question type, with 10 passes per sample. Although \\( 10 \\times 10 \\) has 10x fewer unique samples than \\( 100 \\times 1 \\), there is a \\( < 2\\% \\) accuracy difference between them, indicating that more passes per instance can partially mitigate data scarcity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.472 + ], + "angle": 0, + "content": "tude decrease in training data \\((100\\to 10)\\). Only when the amount of available training data is reduced by two orders of magnitude \\((100\\to 1)\\) does self-training fail to produce an appreciable increase in performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.472, + 0.655 + ], + "angle": 0, + "content": "Is it possible to mitigate data scarcity? We previously showed that the benefits of self-training reduce when available data is reduced significantly. We now test whether we can mitigate this data scarcity by allowing \\(\\pi_{\\theta}\\) multiple attempts at a query \\(q\\) during the Grow step. Concretely, we allow \\(\\pi_{\\theta}\\) a total of 10 tries at each query under the setting in which we train with 1 and 10 samples per question type, for a total of 1k and 10k total samples respectively. We show in Fig. 6 that this mitigates the effect of reduced data. Although the data poor \\(1 \\times 10\\) and \\(10 \\times 10\\) have 10x fewer unique questions than \\(10 \\times 1\\) and \\(100 \\times 1\\), their performance is within a standard deviation of their data rich counterparts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.45, + 0.683 + ], + "angle": 0, + "content": "4.4. Quantifying Changes in Syntactic Structure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.856 + ], + "angle": 0, + "content": "How do the programs synthesized by the policy change as self-training is iterated? We examine this by looking at how many unique abstract syntax trees are produced during the Grow step of each iteration. We parse the synthesized programs into abstract syntax trees, and then normalize the trees to remove irrelevant details such as variable names. In the left panel of Fig. 7, we show that the diversity of syntactic forms drops over time. At the beginning, the policy produces a large number of syntactic forms, but appears to \"hone in\" on a smaller number of forms as self-training continues, and the number of unique syntactic forms drops by almost half." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "A remarkably stable set of syntactic forms is conserved from step to step, roughly \\(\\approx 700\\) (row above diagonal in right panel of Fig. 7). However, the syntactic forms produced" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.696, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.097, + 0.892, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.246, + 0.895, + 0.372 + ], + "angle": 0, + "content": "Figure 7. As self-training is iterated, the LLM policy \"hones in\" on a smaller set of syntactic forms, and gradually evolves away from syntactic forms produced by the initial policy. Left Panel: Number of unique normalized abstract syntax trees seen during each iteration of VisReP. Right Panel: Number of unique normalized abstract syntax trees in common between each training step. For example, the entry in row 1, column 6 corresponds to the number of unique abstract syntax trees produced by both the policy in iteration 1 (initial policy) and the policy in iteration 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.895, + 0.48 + ], + "angle": 0, + "content": "by the policy are gradually evolving away from the syntactic forms the initial policy tries, which can be seen in the darkening of the first row in Fig. 7. Despite the coarse reward scheme, the LLM policy gradually explores and learns new syntactic forms." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.504, + 0.816, + 0.521 + ], + "angle": 0, + "content": "5. Evaluating Functional Correctness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.578 + ], + "angle": 0, + "content": "We measure the functional correctness of the programs synthesized by the self-trained LLM policy \\(\\pi_{\\theta}^{\\prime}\\) across three compositional tasks, with the aim of understanding whether:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.581, + 0.895, + 0.626 + ], + "angle": 0, + "content": "1. Are the programs produced after self-training more functionally correct than programs produced before self-training?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.626, + 0.894, + 0.657 + ], + "angle": 0, + "content": "2. Is it possible to exceed or match the performance of a much larger proprietary LLM with self-training?" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.581, + 0.895, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.902 + ], + "angle": 0, + "content": "For compositional VQA, we use the GQA [13] dataset for the reasons outlined in Sec. 4.1. For complex object detection, we choose Omnilabel [24]. Omnilabel contains 28K free-form object descriptions over 25K images, and is a challenging task for existing open-vocabulary object detectors due to the complexity of the object descriptions. For compositional image-text matching, we choose WinoGround [29] and SugarCrepe [11]. State-of-art vision-language models have trouble reaching above chance accuracy on WinoGround, but SugarCrepe is substantially easier. However, both of these tasks pose significant problems for the ImagePatch API, because many of the relationships mentioned in the text are challenging to detect with the available perception modules. For all experiments, we use ViperGPT[28] as the backbone and adopt their prompts. Due to space limitations, many experimental details are in the supplement." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14350" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.187 + ], + "angle": 0, + "content": "
VQAObject DetectionImage-Text Matching
MethodLLMGQAOmnilabelOmnilabel-HardWinogroundSugarCREPE
Frozen Proprietary LLMGPT-3.5-turbo53.9 ± 0.840.0 ± 1.226.0 ± 1.145.6 ± 1.648.9 ± 0.7
Frozen Open LLMCodeLlama-7B50.0 ± 1.737.3 ± 1.623.7 ± 1.441.3 ± 0.0343.5 ± 0.8
Open LLM + VisRePCodeLlama-7B59.2 ± 1.442.4 ± 1.028.1 ± 0.952.7 ± 0.658.7 ± 1.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.199, + 0.893, + 0.24 + ], + "angle": 0, + "content": "Table 2. An open LLM policy self-trained with our method substantially outperforms the open policy without self-training, and even outperforms a gpt-3.5-turbo policy. All results use ViperGPT [28] as the backbone. \\(\\pm\\) numbers are the standard deviation over 5 runs. On all datasets except Omnilabel, we report accuracy. On Omnilabel, we report Macro-F1. Higher is better." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.254, + 0.466, + 0.332 + ], + "angle": 0, + "content": "
Code GeneratorX-Dataset GeneralizationX-Task Generalization
VQAv2OK-VQAOmnilabelSugarCrepe
Frozen Few-Shot (7B)46.6 ± 1.112.6 ± 2.137.3 ± 1.643.5 ± 0.8
VisReP on GQA (7B)61.0 ± 1.033.7 ± 1.939.9 ± 0.751.0 ± 1.5
VisReP Advantage+14.4+21.1+2.6+7.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.34, + 0.469, + 0.381 + ], + "angle": 0, + "content": "Table 3. VisReP improves benchmark agnostic visual program synthesis. A policy self-trained on GQA with VisReP writes better programs for other VQA datasets and other task types." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.403, + 0.265, + 0.418 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.428, + 0.47, + 0.548 + ], + "angle": 0, + "content": "For each task, we apply VisReP as described in Sec. 3.2, and evaluate on a held-out subset. For a comparison with a large proprietary LLM, we use gpt-3.5-turbo. We evaluate on a subsampled version of each dataset to reduce token costs. Every LLM is provided the same prompts. Each prompt consists of the ImagePatch API specification used in ViperGPT [28], and 3 in-context examples for each task except for object detection, for which we provide 5 in-context examples." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.551, + 0.47, + 0.838 + ], + "angle": 0, + "content": "We use GQA as described in Sec. 4.1. We prepare a compositional subset of Omnilabel [24] by filtering out all descriptions less than two words in length. We then sample a subset of 500 for evaluation, and a subset of 500 for training. To prepare Omnilabel-Hard, we use run a state of the art open-vocabulary object detector (GroundingDINO [18]) on the remaining OmniLabel samples, and select those which GroundingDINO completely fails on (no detections) to obtain a hard slice. We then sample a subset of 500 from the hard slice for evaluation. For SugarCrepe [11], we sample 100 positives and their associated negatives from each of the 6 categories, for a total of 600 balanced image-text pairs for validation. We sample 100 of the remaining instances from each category for training. We use all of WinoGround, as it is small enough that there is no need to subsample it. On WinoGround[29], we evaluate the policy trained on SugarCrepe rather than training on it. For VQAv2, we sample 10 questions for each of the top-50 most common answers from the compositional subset curated by [25]." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Examples of the inputs for each task are in Fig. 3. We use nucleus sampling with identical parameters for all local LLMs. We use the API default temperature for gpt-3.5-turbo. More details are in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.267, + 0.616, + 0.281 + ], + "angle": 0, + "content": "5.2. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.29, + 0.893, + 0.591 + ], + "angle": 0, + "content": "Across all three tasks, the policy trained by VisReP outperforms both the gpt-3.5-turbo policy, and the initial CodeLlama-7b policy (Tab. 2). On GQA, the self-trained policy achieves an absolute improvement of almost \\(9\\%\\) over the initial policy, and \\(5\\%\\) over the gpt-3.5-turbo policy. On Omnilabel, self-training produces a \\(5\\%\\) improvement in Macro-F1 score with only 500 training samples. On Omnilabel-Hard, we demonstrate that the visual program synthesis paradigm can localize objects that state of the art open-vocabulary object detectors are unable to localize (Omnilabel-Hard was constructed by selecting instances GroundingDino[18]) cannot localize). Even on Omnilabel-Hard, the self-trained policy outperforms the others. WinoGround and SugarCrepe are difficult to solve by visual program synthesis because many of the relationships are hard to detect with the available perception modules. Despite the intrinsic difficulty of compositional image-text matching for the ImagePatch API, VisReP produces an increase of \\(+15\\%\\) over the baseline policy. The policy trained on SugarCrepe transfers to WinoGround, outperforming the baseline policy by \\(+10\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.605, + 0.753, + 0.619 + ], + "angle": 0, + "content": "6. Conclusion & Future Work" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.629, + 0.893, + 0.9 + ], + "angle": 0, + "content": "While few-shot prompting of LLMs for visual program synthesis has produced impressive results, it has limitations, because writing good visual programs requires experience with the visual world and the perception modules at ones disposal. We presented VisReP, which improves a LLM's program synthesis abilities using feedback from executing visual programs. We showed that VisReP produces strong increases over baseline across multiple tasks, and is competitive with gpt-3.5-turbo. Our work constructed a coarse-valued reward from existing vision-language annotations. Methods like RLAIF [2], ReST [8], and CodeRL [16] all rely on a neural reward model that can provide fine-grained rewards. Learning from fine-grained rewards is much easier than learning from coarse rewards. An interesting direction for future work would be to train a neural reward model for visual program synthesis. Such a reward model could provide fine-grained rewards, and open a broader range of reinforcement learning methods." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.517, + 0.957 + ], + "angle": 0, + "content": "14351" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V. Le, and Charles Sutton. Program synthesis with large language models. ArXiv, abs/2108.07732, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.473, + 0.395 + ], + "angle": 0, + "content": "[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, E Perez, Jamie Kerr, Jared Mueller, Jeff Ladish, J Landau, Kamal Ndousse, Kamile Lukovs.iuté, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noem'i Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, T. J. Henighan, Tristan Hume, Sam Bowman, Zac Hatfield-Dodds, Benjamin Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom B. Brown, and Jared Kaplan. Constitutional ai: Harmlessness from ai feedback. ArXiv, abs/2212.08073, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.399, + 0.473, + 0.454 + ], + "angle": 0, + "content": "[3] Natasha Butt, Blazej Manczak, Auke J. Wiggers, Corrado Rainone, David W. Zhang, Michael Defferrard, and Taco Cohen. Codeit: Self-improving language models with prioritized hindsight replay. ArXiv, abs/2402.04858, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.473, + 0.693 + ], + "angle": 0, + "content": "[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.697, + 0.473, + 0.737 + ], + "angle": 0, + "content": "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. ArXiv, abs/2305.14314, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.742, + 0.473, + 0.797 + ], + "angle": 0, + "content": "[6] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.801, + 0.473, + 0.842 + ], + "angle": 0, + "content": "[7] Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. VQA-LOL: visual question answering under the lens of logic. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.473, + 0.9 + ], + "angle": 0, + "content": "[8] Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alexa Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, A. Doucet, Orhan First, and Nando de Freitas. Re-" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.12 + ], + "angle": 0, + "content": "inforced self-training (rest) for language modeling. ArXiv, abs/2308.08998, 2023. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.894, + 0.163 + ], + "angle": 0, + "content": "[9] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. ArXiv, abs/2211.11559, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[10] Patrick M. Haluptzok, Matthew Bowers, and Adam Tauman Kalai. Language models can teach themselves to program better. ArXiv, abs/2207.14502, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.895, + 0.276 + ], + "angle": 0, + "content": "[11] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.895, + 0.332 + ], + "angle": 0, + "content": "[12] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.336, + 0.895, + 0.417 + ], + "angle": 0, + "content": "[13] Drew A. Hudson and Christopher D. Manning. GQA: A new dataset for real-world visual reasoning and compositional question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6700-6709. Computer Vision Foundation / IEEE, 2019. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.895, + 0.501 + ], + "angle": 0, + "content": "[14] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.895, + 0.546 + ], + "angle": 0, + "content": "[15] Sascha Lange, Thomas Gabel, and Martin Riedmiller. Batch Reinforcement Learning, pages 45-73. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.895, + 0.616 + ], + "angle": 0, + "content": "[16] Hung Le, Yue Wang, Akhilesh Deepak Gotmare, Silvio Savarese, and Steven Hoi. CodeRL: Mastering code generation through pretrained models and deep reinforcement learning. In Advances in Neural Information Processing Systems, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.895, + 0.659 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.895, + 0.773 + ], + "angle": 0, + "content": "[19] Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.895, + 0.857 + ], + "angle": 0, + "content": "[20] Yujia Qin, Shi Liang, Yining Ye, Kunlun Zhu, Lan Yan, Ya Ting Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Marc H. Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis. ArXiv, abs/2307.16789, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.895, + 0.9 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14352" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.471, + 0.161 + ], + "angle": 0, + "content": "Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[22] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, I. Evtimov, Joanna Bitton, Manish P Bhatt, Cristian Cantón Ferrer, Aaron Grattafori, Wenhan Xiong, Alexandre D'efossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. Code llama: Open foundation models for code. ArXiv, abs/2308.12950, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[23] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[24] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[25] Ramprasaath R. Selvaraju, Purva Tendulkar, Devi Parikh, Eric Horvitz, Marco Tulio Ribeiro, Besmira Nushi, and Ece Kamar. Squinting at vqa models: Introspecting vqa models with subquestions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[26] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Ke- nely, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell L. Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem- solving with language models. ArXiv, abs/2312.06585, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[27] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 747-761, Toronto, Canada, 2023. Association for Computational Linguistics. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[28] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[29] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[30] Spencer Whitehead, Suzanne Petryk, Vedaad Shakib, Joseph Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Rohrbach. Reliable visual question answering: Abstain rather than answer incorrectly. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[31] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229-256, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[32] Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[33] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14353" + } + ] +] \ No newline at end of file diff --git a/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_origin.pdf b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..98f6dab52cdeffb84bb4d3ac940e43d3a5db4fa4 --- /dev/null +++ b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/011d32ff-cfb4-4400-94b5-e42371e5459f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3e3f925f0b4dfd714a08e0c8da75ff9f4bf411a649e545a1a68720d76e782df +size 930836 diff --git a/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/full.md b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..009e60cc01baa45f512d7a08c1a0a8d38e95ddd0 --- /dev/null +++ b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/full.md @@ -0,0 +1,261 @@ +# Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement + +Zaid Khan $^{1}$ Vijay Kumar BG $^{2}$ Samuel Schulter $^{2}$ Yun Fu $^{1}$ Manmohan Chandraker $^{2,3}$ + +$^{1}$ Northeastern University $^{2}$ NEC Laboratories America $^{3}$ UC San Diego + +# Abstract + +Visual program synthesis is a promising approach to exploit the reasoning abilities of large language models for compositional computer vision tasks. Previous work has used few-shot prompting with frozen LLMs to synthesize visual programs. Training an LLM to write better visual programs is an attractive prospect, but it is unclear how to accomplish this. No dataset of visual programs for training exists, and acquisition of a visual program dataset cannot be easily crowdsourced due to the need for expert annotators. To get around the lack of direct supervision, we explore improving the program synthesis abilities of an LLM using feedback from interactive experience. We propose a method where we exploit existing annotations for a vision-language task to improvise a coarse reward signal for that task, treat the LLM as a policy, and apply reinforced self-training to improve the visual program synthesis ability of the LLM for that task. We describe a series of experiments on object detection, compositional visual question answering, and image-text retrieval, and show that in each case, the self-trained LLM outperforms or performs on par with few-shot frozen LLMs that are an order of magnitude larger. Website: + +https://zaidkhan.me/ViReP + +# 1. Introduction + +Complex visual queries can often be decomposed into simpler subtasks, many of which can be carried out by task-specific perception modules (e.g. object detection, captioning). For example, consider the problem of finding bounding boxes for the phrase "white mug to the left of the sink". This is a challenging query for single model such as an open vocabulary object detector. However, this query can be solved by writing a program that composes task-specific perception modules with logic: use an open vocabulary object detector to find a sink and white mugs in the scene, then compare the horizontal center of the sink and the mugs to find white mugs to the left of the sink. Program synthesis with large language models [1] is a promising approach to automate this process, and recent work has shown that proprietary large language + +![](images/66344b4b51bcfc966e6bbfbe756ff8b1eb4ab3f4251b96d71c5c74cf85faf09e.jpg) + +![](images/0ef705e9a0f73bda7686e15c73263592e83053126f9652916b2145f1d7c832d9.jpg) +How can we train a LLM to write better visual programs for a task? +Figure 1. Visual program synthesis with LLMs has been treated as a $0 / \mathrm{n}$ -shot task where the LLM is kept frozen. This limits opportunities for improvement. We ask whether it is possible to train a LLM to write more accurate programs. Given that there is no large scale dataset of accurate visual programs available, we propose improving the LLM using self-training. + +models can write programs for visual tasks [9, 27, 28]. Current approaches for visual program synthesis with LLMs use few-shot prompting and rely on the in-context learning abilities [32] of frozen, proprietary LLMs. (Fig. 1) + +Few-shot prompting with frozen LLMs for visual program synthesis as in ViperGPT [28], VisProg [9], or CodeVQA [27] has several limitations. The LLM needs to understand the competencies of the perception modules it is using. A open vocabulary object detector may be able to locate a common attribute-noun phrase such as "white mug" without problems, but struggle with a more abstract phrase such as "microwaveable mug" [24]. A VQA model might be able to answer "is the car blue?" without problems, but fail when logical modifiers are introduced, such as "is the car not blue?" [7]. In many cases, we do not precisely know + +
Task DomainSelf-TrainingSupervisionTool/API UseVisual Task DecompositionGrounded By FeedbackImproves LLM
VisReP (Ours)Visual Program SynthesisYesWeakYesYesYesYes
Haluptzok et al. [10]Programming PuzzlesYesWeakNoNoYesYes
ReST [8]Natural Language UnderstandingYes-NoNoYesYes
VisProg [9]Visual Program SynthesisNo-YesYesNoNo
ViperGPT [28]Visual Program SynthesisNoWeakYesYesNoNo
ToolLLM [20]Tool Usage by APINoStrongYesNoNoYes
GorillaLLM [19]Tool Usage by APINoStrongYesNoNoYes
ToolFormer [23]Natural Language UnderstandingYesWeakYesNoYesYes
+ +Table 1. Differences between our work and similar work. Strong supervision means that the training process requires examples of ground-truth programs to train the LLM. Weak supervision means that the training process does not require ground-truth programs. Tool / API use means that the LLM is required to use substantial functionality implemented by external modules (e.g. an object detector, a web search API) to solve tasks. Visual task decomposition means that the LLM can decompose a complex visual task into primitive subtasks. Grounded by feedback means that the LLM has been optimized not just for syntactic / semantic correctness (program does not hallucinate / cause errors), but for functional correctness (programs produce the correct answer). Improves LLM means that the work proposes a method to improve an LLM for a specific task, rather than using a frozen LLM. + +the weaknesses or competencies of a perception model [30]. Even if this were known, it is difficult to convey all of the competencies / weaknesses of a perception module through in-context examples. Second, it is the case in program synthesis that an LLM often can generate the correct solution to a problem, but the correct solution is not the solution the LLM places the highest probability on [4]. We would like to align the LLM to "uncover" the knowledge of the correct solution, but it not clear how to do this in a principled way with few-shot prompting alone. + +How can we train a large language model to write better visual programs for a specific task? + +Our goal is to optimize the parameters of the language model so the accuracy of the synthesized programs is higher. Existing approaches that train LLMs to improve their ability to programatically use tools / APIs such as GorillaLLM [19], ToolLLM [20] do so by finetuning LLMs on examples of tool use or API use. This cannot be directly applied to visual program synthesis because there are no large scale datasets of visual programs, and collecting such a dataset would be extremely labor intensive. In the absence of a large scale dataset, how do we learn to write better programs for a visual task? + +We posit that grounding a language model with interactive feedback from a generic visual task will improve the general visual program synthesis abilities of the model. + +A natural way to learn from feedback is to use reinforcement learning. ReST [8] and RaFT [6] introduce a general framework for reinforced self-training in generative tasks and demonstrate success in machine translation and text-to-image generation. However, a crucial ingredient in their recipe is the availability of a fine-grained reward model. It is difficult to construct a fine-grained reward model for visual program synthesis, given both the absence of human preference datasets for visual programs, and the difficulty + +of devising a proxy metric. One alternative is to use unit tests to teach a neural reward model or give a coarse-grained reward. This technique has been used successfully in coding challenges by CodeRL [16] and Haluptzok et al. [10], but it is unclear how it can be applied to visual program synthesis. Our key idea is to use existing annotations for a vision-language as improvised unit tests to provide a coarse reward signal. Using the coarse reward signal, we can apply reinforced self-training by treating the language model as a policy and training it with a simple policy gradient algorithm. We alternate synthetic data generation steps in which we sample programs from the language model policy with optimization steps in which we improve the language model policy based on observations from executing the sampled programs. We name our proposed method VisReP, for Visually Reinforced Program Synthesis. + +- We propose optimizing the parameters of a LLM so that the accuracy of the synthesized visual programs is higher, in contrast to previous works that use frozen LLMs. +- Since no dataset of accurate visual programs is available for finetuning, we hypothesize that we can instead use feedback from the execution environment to improve the visual program synthesis abilities of a language model. +- We propose VisReP, an offline, model agnostic recipe for reinforced self-training of large language models for visual program synthesis using existing vision-language annotations with a simple policy gradient algorithm. +- Our results show that it is possible to apply reinforced self-training for to improve large language models for visual program synthesis with only coarse rewards. + +We demonstrate the effectiveness of an CodeLlama-7B policy trained by VisReP on compositional visual question answering $(+9\%)$ , complex object detection $(+5\%)$ , and compositional image-text matching $(+15\%)$ relative to the untrained policy. We show that the policy trained by VisReP exceeds the accuracy of a gpt-3.5-turbo policy on all three tasks. + +# 2. Related Work + +# 2.1. Self-Training + +Self-training is an established paradigm which uses unlabeled data to improve performance. Self-training has been successfully applied in a number of fields. We restrict our coverage to usages with significant overlap. + +Program Synthesis Haluptzok et al. [10] showed that LLMs can improve their program synthesis abilities by generating programming puzzles and solving them. CodeRL [16] proposed an actor-critic framework to improve the program synthesis abilities of LLMs for programming problems accompanied by unit tests. CodeIT [3] and Rest-EM [26] also use a similar policy gradient approach for program synthesis. Our problem domain is different from these works, which focus on program synthesis for programming puzzles / problems. In addition, our work has an explicit focus on learning to use an API fluently. + +Alignment ReST [8] and RaFT [6] introduced a generic framework for reinforced self-training and applied it to align machine translation outputs to human preferences and align foundation models on language understanding and image generation tasks respectively. These works share the same basic idea as our work, though they are in a substantially different task domain where human preferences are either known (conversational alignment) or can be estimated with an available neural model. + +Vision-Language SelTDA [14] introduced a self-training approach for visual question answering. SelTDA proceeds by pseudolabeling unlabeled data, then finetuning a large VLM on the pseudolabeled data. In contrast to SelTDA, we improve a LLM for visual program synthesis. + +# 2.2. Visual Program Synthesis + +Visual program synthesis with LLMs was proposed concurrently by ViperGPT [28], VisProg [9], and CodeVQA [27]. The common points between these three works is that (a) they use pretrained LLMs as code generators (b) they represent complex visual tasks as compositions of primitive visual subtasks (c) they use code to invoke task-specific models to perform the primitive subtasks. Our work is most similar to ViperGPT and CodeVQA as they produce code in a general purpose programming language rather than a DSL. All three works use a proprietary, frozen LLM. In contrast to all three, the focus of our work is on how we can improve the visual program synthesis abilities of an open LLM. + +# 2.3. Tool Use with LLMs + +Multimodal tool-using LLMs were first introduced by Socratic Models [33]. However, their approach was to create fixed pipelines in which the output of a perception model such as CLIP [21] is fed to a LLM. Later approaches such as GorillaLLM [19] and ToolLLM [20] improved on this by + +treating tool use as a program synthesis problem and creating LLMs that use a broad range of tools by learning to invoke APIs. However, one key limitation of these approaches in the context of visual program synthesis is that they do not learn to decompose problems into subproblems that can be solved by tools. Instead, they are trained to select the right tool for the problem and invoke it. Another limitation is that they are not optimized for functional correctness. They are trained for syntactic and semantic correctness, but they have not been provided feedback on whether their use of tools produces the desired answer. ToolFormer [23] is similar to our work in the sense that the LLM's usage of tools is grounded by feedback, but they focus on natural language understanding tasks rather than visual tasks. + +# 3. Method + +# 3.1. Visual Program Synthesis with LLMs + +Task Formulation Let $v$ be a visual input and $q$ be a textual query about $v$ . In visual program synthesis, we synthesize a program $p = \pi_{\theta}(q)$ with a program generator $\pi_{\theta}$ . The program $p$ and visual input $q$ are then fed into the execution engine $\hat{y} = \phi(v, p)$ to produce a result $\hat{y}$ . The program generator is an auto-regressive large language model + +$$ +\pi_ {\theta} (\boldsymbol {y} \mid \boldsymbol {x}) = \prod_ {t = 1} ^ {T} \pi_ {\theta} \left(p _ {t} \mid \mathbf {p} _ {1: t - 1}, \boldsymbol {x}\right), \tag {1} +$$ + +where $\mathbf{p}_{1:t}$ are the tokens of the program, and $x$ is the input to the large language model. The language model is kept frozen in previous work [28]. Our goal is to optimize the parameters $\theta$ of the language model $\pi$ so the accuracy of the synthesized programs is higher. + +Implementation Following ViperGPT [28], we provide the specification of the ImagePatch API concatenated with the textual query $q$ as the prompt to the program generator. The synthesized program $p$ is a Python program that can invoke any Python builtins, control flow structures, and the ImagePatch API. Our implementation of the ImagePatch API is largely similar to ViperGPT. We remove some API methods that were not required for the tasks we evaluate on (such as l1m_query). We use BLIP [17] and GroundingDINO [18] as perception modules underlying find (object detection), simple-query (visual question answering), and verify_property (attribute verification). + +# 3.2. Reinforced Self-Training + +Rather than use a frozen large language model as the program generator $p_{\theta}$ , we would like to optimize the parameters $\theta$ of the language model so the accuracy of the synthesized programs is higher. It is not obvious how to do this. We can't backpropagate through the execution engine $\phi(\pi_{\theta}(q), v)$ to + +![](images/526901645863a3d4e02b783d757276e55698454498b792b7a9d9885b081a8b22.jpg) +Offline Policy Improvement + +![](images/f3c47938ba1c105a42024486873d143a800ecb1bba29453f3bfd7cfef1cd069d.jpg) +Figure 2. VisReP can be applied to improve the visual synthesis abilities of an LLM for a vision-language task using existing annotations for a vision-language task (e.g. an object description + image + bounding boxes). A key idea is to construct a coarse reward by comparing the answer produced by a synthesized program to the ground-truth answer. + +directly optimize $\theta$ with respect to $q$ or $v$ . An alternative might be to use human labor to build a dataset of high-quality visual programs, and train the large language model $\pi_{\theta}$ on the manually-collected dataset. But collecting such a dataset is very labor intensive, and not scalable. Instead, we explore the idea of learning from experience by applying a simple policy gradient method, REINFORCE [31]. + +We propose VisReP, which treats the program synthesis task as a growing batch RL problem [15], inspired by ReST [8]. We first define a coarse discrete reward function $R(\cdot)$ from existing annotations for a vision-language task. We then alternate Grow steps, in which we sample trajectories (programs) from the policy (large language model), with Improve steps, in which we apply behavioral cloning with a reward-weighted negative-log likelihood loss to improve the policy. A diagram of our approach is depicted in Fig. 2. + +Grow Step The grow step corresponds to the acting step in reinforcement learning, and can also be seen as synthetic data generation. Let $\mathcal{D} = \{(v_1, q_1, y_1), \ldots, (v_n, q_n, y_n)\}$ be a dataset for a vision-language task, where $v_i$ is an image, $q_i$ is a textual query, and $y_i$ is ground-truth for the $i$ -th triplet (e.g. a string for VQA, bounding boxes for object detection). We start with the frozen language model $\pi_\theta(p \mid q)$ , where $p$ is a synthesized program and $q$ is a textual query. The language model $\pi_\theta$ represents our policy. We generate a dataset of trajectories $\mathcal{D}_g$ by sampling many programs $p$ from the current policy $\pi_\theta: p \sim \pi_\theta(p \mid q)$ for $q \sim \mathcal{D}$ . + +Improve Step Our goal in this step is to use the dataset of synthetic programs $\mathcal{D}_g$ to improve the policy $\pi_{\theta}$ . First, we + +define a binary-valued reward function $R: p, v, y \to \{0, 1\}$ on a given program, image, annotation triplet, + +$$ +R (v, p, y) = \left\{ \begin{array}{l l} 1, & \text {i f} \phi (p, v) = y \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +where $\phi(p, v)$ is the result of executing the program $p$ on an image $v$ . Note that $y$ is not a program but an existing annotation such as a string for VQA for a bounding box for object detection. To apply behavioral cloning, we then minimize the reward-weighted loss + +$$ +J (\theta) = \mathbb {E} _ {(q, p) \sim \mathcal {D} _ {g}} [ R (v, p) \mathcal {L} (p, q; \theta) ] \tag {3} +$$ + +where $\mathcal{L}(p,q;\theta)$ is the negative log-likelihood loss + +$$ +\mathcal {L} _ {\mathrm {N L L}} (p, q; \theta) = - \mathbb {E} _ {(q, p) \sim \mathcal {D} _ {g}} \left[ \sum_ {t = 1} ^ {T} \log \pi_ {\theta} \left(p _ {t} \mid p _ {1: t - 1}, q\right) \right] \tag {4} +$$ + +over the pairs of textual queries $q$ and synthetic programs $p$ in $\mathcal{D}_g$ . + +Because the reward function only takes on binary values, we can simplify this and implement it by: First, generating a dataset of synthetic programs $\mathcal{D}_g = \{\pi_\theta (q):\forall q\in \mathcal{D}\}$ using the LLM $\pi_{\theta}$ on a dataset $\mathcal{D}$ . Next, filtering $\mathcal{D}_g$ to obtain $\mathcal{D}_g^\prime = \{(q,v,p\in \mathcal{D}_g:R(q,v,p) > 0\}$ , which corresponds to executing all synthetic programs and only keeping those that give correct answers. Finally, we finetune the language model $\pi_{\theta}$ on the filtered dataset $\mathcal{D}_g^\prime$ using the standard causal language modeling loss. We then iterate the process, initiating a new synthetic data generation step with the improved policy $\pi_{\theta}^{\prime}$ . + +Iteration For the initial grow step, we use a frozen language model as the initial policy. For example, we use the pretrained codellama-7b-instruct-hf as the policy in the initial grow step. In subsequent steps, we use the policy trained in the previous improve step for the grow step. + +# 4. Understanding Self-Training + +Our goal in this section is to characterize the stability and sample efficiency of VisReP. We want to understand: + +1. How does applying VisReP change the accuracy of synthesized programs? +2. What happens as VisReP is repeated? +3. How does data scarcity and diversity affect VisReP? + +# 4.1. Implementation + +We start off with the GQA [13] dataset for visual question answering. We choose GQA because each question in GQA was constructed programatically and is thus a good candidate to be answered by program synthesis. GQA has over 2M questions, each belonging to one of $\approx 100$ question types. We construct a training set by sampling 100 questions for + +![](images/f7968f0e6cdf62333acfd91d8a7b4b75df1392b20dd62ccb7564916bd4920de7.jpg) +Figure 3. Self-training with VisReP produces qualitatively better programs. Here, we show programs written by the initial policy (on the left) and the policy after 10 iterations of self-training on GQA (on the right). In VQA example, the initial policy does not specifically check whether the empty basket is plastic. In the object detection example, the reasoning of the initial policy is correct, but it issues a confusingly worded query to the simple-query module, which returns the wrong answer. The learned policy uses simple-query more appropriately. In the image-text matching example, in the initial policy tries to use the object detector to search directly for "meat in a box" and "donuts on a plate", but this is too complicated for the object detector to localize. After self-training, the LLM policy no longer makes this mistake. + +each question type, for a total of $\approx 10\mathrm{k}$ visual questions and answers. We construct a validation set following Gupta and Kembhavi [9]. We use the CodeLlama [22] family of models as our initial policy. We use LoRA [12] adapters during the Improve steps. We use the hyperparameters suggested by + +Dettmers et al. [5]. Full implementation details are in the supplement. + +![](images/8cfa5e6436ce0085f6cd0d2bc693ad92fd2c51e33bdddf37c6bb8bbcb3abfb64.jpg) +$\Delta$ Training Accuracy from Step 1 to Step 10 of Self-Training + +![](images/fbd95f5ddc46e3bf5e646ac0c4a9db7effe1c6e91cf7b22792dd7e5b3aaf9425.jpg) +Figure 4. Iteratively applying VisReP allows a LLM to self-improve improve on almost all of GQA's $\approx 100$ question types. The base of each bar is set to the accuracy of the initial policy (codellama-7b-instruct). A green bar indicates question types on which the policy at iteration 10 improved over the initial policy, and a red bar indicates question types on which the policy at iteration 10 was worse than the initial policy. +Figure 5. Supplying a small amount of human written corrections as in-context examples during training can increase the stability of the self-training process (green line). We show validation accuracy on GQA through multiple iterations of self-training with a policy instantiated from CodeLlama-7b. Without these corrections, proliferating errors cause performance to degrade in later iterations (red line). The translucent shading around each line indicates the standard deviation over 5 evaluations on the validation set. + +# 4.2. Persistent Errors Harm Iterated Self-Training + +Applying the formulation of self-training in Sec. 3.2 results in a improvement, but iterating it further results in program synthesis quality degrading, rather than increasing (red line in Fig. 5). This is due to the self-training process inadvertently reinforcing incorrect reasoning. A program that uses flawed reasoning can occasionally produce a correct answer. The language model can thus be rewarded for a program that is right for the wrong reasons. If this goes uncorrected, the language model will learn incorrect reasoning patterns. + +We hypothesize that providing a small number of human- + +written corrections for persistent reasoning errors can stabilize the self-training process. We use the question type annotations in GQA to identify question types for which training accuracy decreases over time. These are question types which the language model is not able to self-improve on. We denote them $\mathcal{Q}_{\text {hard }}$ . For each question type in $\mathcal{Q}_{\text {hard }}$ , we randomly sample one question $q$ for which the language model synthesized a program that produced the wrong answer. We examine the reasoning in that program, and if the reasoning is flawed, we correct it. We repeat this until we have a program with correct reasoning for each question type in $\mathcal{Q}_{\text {hard }}$ , and denote the bank of correct programs as $\mathcal{P}_{\text {gold }}$ . + +We then retrieve from $\mathcal{P}_{\text{gold}}$ during self-training for use as in-context examples. If a question is annotated with a question type in $Q_{\text{hard}}$ , we retrieve a correct human-written program from $\mathcal{P}_{\text{gold}}$ and use it as an in-context example. If a question is not annotated with a question type in $Q_{\text{hard}}$ , we use a "default" in-context example which is the same for all question types not in $Q_{\text{hard}}$ . We show in Fig. 5 (green line) that this stabilizes self-training and allows the language model to self-improve across all but a few question types (Fig. 4). + +# 4.3. Effect of Data Availability on Self-Training + +Training With Less Data We explore this in a controlled setting, by manipulating the number of samples per question type in GQA. Recall that we originally sample 100 questions per question type for self-training. This dataset had $\approx 10\mathrm{k}$ questions. We construct a training set with only 10 and 1 question per question type, for a total of $\approx 1000$ and $\approx 100$ questions respectively. Self-training improves upon the baseline (Fig. 6) even when there is an order of magni + +![](images/2111617b308482b9c5768a3544789a5eeaae4588bdf5132519f2602b6ac8d062.jpg) +Figure 6. VisReP works even when the amount of available data is reduced by an order of magnitude. We show validation accuracy on GQA. The notation $n \times k$ indicates $n$ samples per question type, with $k$ passes at each sample. For example $10 \times 10$ indicates 10 samples per question type, with 10 passes per sample. Although $10 \times 10$ has 10x fewer unique samples than $100 \times 1$ , there is a $< 2\%$ accuracy difference between them, indicating that more passes per instance can partially mitigate data scarcity. + +tude decrease in training data $(100\to 10)$ . Only when the amount of available training data is reduced by two orders of magnitude $(100\to 1)$ does self-training fail to produce an appreciable increase in performance. + +Is it possible to mitigate data scarcity? We previously showed that the benefits of self-training reduce when available data is reduced significantly. We now test whether we can mitigate this data scarcity by allowing $\pi_{\theta}$ multiple attempts at a query $q$ during the Grow step. Concretely, we allow $\pi_{\theta}$ a total of 10 tries at each query under the setting in which we train with 1 and 10 samples per question type, for a total of 1k and 10k total samples respectively. We show in Fig. 6 that this mitigates the effect of reduced data. Although the data poor $1 \times 10$ and $10 \times 10$ have 10x fewer unique questions than $10 \times 1$ and $100 \times 1$ , their performance is within a standard deviation of their data rich counterparts. + +# 4.4. Quantifying Changes in Syntactic Structure + +How do the programs synthesized by the policy change as self-training is iterated? We examine this by looking at how many unique abstract syntax trees are produced during the Grow step of each iteration. We parse the synthesized programs into abstract syntax trees, and then normalize the trees to remove irrelevant details such as variable names. In the left panel of Fig. 7, we show that the diversity of syntactic forms drops over time. At the beginning, the policy produces a large number of syntactic forms, but appears to "hone in" on a smaller number of forms as self-training continues, and the number of unique syntactic forms drops by almost half. + +A remarkably stable set of syntactic forms is conserved from step to step, roughly $\approx 700$ (row above diagonal in right panel of Fig. 7). However, the syntactic forms produced + +![](images/653672caefc87f0f4ac555f55469f762680e7f4462711d859412d2035775af08.jpg) +Figure 7. As self-training is iterated, the LLM policy "hones in" on a smaller set of syntactic forms, and gradually evolves away from syntactic forms produced by the initial policy. Left Panel: Number of unique normalized abstract syntax trees seen during each iteration of VisReP. Right Panel: Number of unique normalized abstract syntax trees in common between each training step. For example, the entry in row 1, column 6 corresponds to the number of unique abstract syntax trees produced by both the policy in iteration 1 (initial policy) and the policy in iteration 6. + +![](images/14447310f7b76e4d930e540b76eb8d418796293d5a9cfbf024de25a3450da451.jpg) + +by the policy are gradually evolving away from the syntactic forms the initial policy tries, which can be seen in the darkening of the first row in Fig. 7. Despite the coarse reward scheme, the LLM policy gradually explores and learns new syntactic forms. + +# 5. Evaluating Functional Correctness + +We measure the functional correctness of the programs synthesized by the self-trained LLM policy $\pi_{\theta}^{\prime}$ across three compositional tasks, with the aim of understanding whether: + +1. Are the programs produced after self-training more functionally correct than programs produced before self-training? +2. Is it possible to exceed or match the performance of a much larger proprietary LLM with self-training? + +For compositional VQA, we use the GQA [13] dataset for the reasons outlined in Sec. 4.1. For complex object detection, we choose Omnilabel [24]. Omnilabel contains 28K free-form object descriptions over 25K images, and is a challenging task for existing open-vocabulary object detectors due to the complexity of the object descriptions. For compositional image-text matching, we choose WinoGround [29] and SugarCrepe [11]. State-of-art vision-language models have trouble reaching above chance accuracy on WinoGround, but SugarCrepe is substantially easier. However, both of these tasks pose significant problems for the ImagePatch API, because many of the relationships mentioned in the text are challenging to detect with the available perception modules. For all experiments, we use ViperGPT[28] as the backbone and adopt their prompts. Due to space limitations, many experimental details are in the supplement. + +
VQAObject DetectionImage-Text Matching
MethodLLMGQAOmnilabelOmnilabel-HardWinogroundSugarCREPE
Frozen Proprietary LLMGPT-3.5-turbo53.9 ± 0.840.0 ± 1.226.0 ± 1.145.6 ± 1.648.9 ± 0.7
Frozen Open LLMCodeLlama-7B50.0 ± 1.737.3 ± 1.623.7 ± 1.441.3 ± 0.0343.5 ± 0.8
Open LLM + VisRePCodeLlama-7B59.2 ± 1.442.4 ± 1.028.1 ± 0.952.7 ± 0.658.7 ± 1.5
+ +Table 2. An open LLM policy self-trained with our method substantially outperforms the open policy without self-training, and even outperforms a gpt-3.5-turbo policy. All results use ViperGPT [28] as the backbone. $\pm$ numbers are the standard deviation over 5 runs. On all datasets except Omnilabel, we report accuracy. On Omnilabel, we report Macro-F1. Higher is better. + +
Code GeneratorX-Dataset GeneralizationX-Task Generalization
VQAv2OK-VQAOmnilabelSugarCrepe
Frozen Few-Shot (7B)46.6 ± 1.112.6 ± 2.137.3 ± 1.643.5 ± 0.8
VisReP on GQA (7B)61.0 ± 1.033.7 ± 1.939.9 ± 0.751.0 ± 1.5
VisReP Advantage+14.4+21.1+2.6+7.5
+ +Table 3. VisReP improves benchmark agnostic visual program synthesis. A policy self-trained on GQA with VisReP writes better programs for other VQA datasets and other task types. + +# 5.1. Experimental Setup + +For each task, we apply VisReP as described in Sec. 3.2, and evaluate on a held-out subset. For a comparison with a large proprietary LLM, we use gpt-3.5-turbo. We evaluate on a subsampled version of each dataset to reduce token costs. Every LLM is provided the same prompts. Each prompt consists of the ImagePatch API specification used in ViperGPT [28], and 3 in-context examples for each task except for object detection, for which we provide 5 in-context examples. + +We use GQA as described in Sec. 4.1. We prepare a compositional subset of Omnilabel [24] by filtering out all descriptions less than two words in length. We then sample a subset of 500 for evaluation, and a subset of 500 for training. To prepare Omnilabel-Hard, we use run a state of the art open-vocabulary object detector (GroundingDINO [18]) on the remaining OmniLabel samples, and select those which GroundingDINO completely fails on (no detections) to obtain a hard slice. We then sample a subset of 500 from the hard slice for evaluation. For SugarCrepe [11], we sample 100 positives and their associated negatives from each of the 6 categories, for a total of 600 balanced image-text pairs for validation. We sample 100 of the remaining instances from each category for training. We use all of WinoGround, as it is small enough that there is no need to subsample it. On WinoGround[29], we evaluate the policy trained on SugarCrepe rather than training on it. For VQAv2, we sample 10 questions for each of the top-50 most common answers from the compositional subset curated by [25]. + +Examples of the inputs for each task are in Fig. 3. We use nucleus sampling with identical parameters for all local LLMs. We use the API default temperature for gpt-3.5-turbo. More details are in the supplement. + +# 5.2. Discussion + +Across all three tasks, the policy trained by VisReP outperforms both the gpt-3.5-turbo policy, and the initial CodeLlama-7b policy (Tab. 2). On GQA, the self-trained policy achieves an absolute improvement of almost $9\%$ over the initial policy, and $5\%$ over the gpt-3.5-turbo policy. On Omnilabel, self-training produces a $5\%$ improvement in Macro-F1 score with only 500 training samples. On Omnilabel-Hard, we demonstrate that the visual program synthesis paradigm can localize objects that state of the art open-vocabulary object detectors are unable to localize (Omnilabel-Hard was constructed by selecting instances GroundingDino[18]) cannot localize). Even on Omnilabel-Hard, the self-trained policy outperforms the others. WinoGround and SugarCrepe are difficult to solve by visual program synthesis because many of the relationships are hard to detect with the available perception modules. Despite the intrinsic difficulty of compositional image-text matching for the ImagePatch API, VisReP produces an increase of $+15\%$ over the baseline policy. The policy trained on SugarCrepe transfers to WinoGround, outperforming the baseline policy by $+10\%$ . + +# 6. Conclusion & Future Work + +While few-shot prompting of LLMs for visual program synthesis has produced impressive results, it has limitations, because writing good visual programs requires experience with the visual world and the perception modules at ones disposal. We presented VisReP, which improves a LLM's program synthesis abilities using feedback from executing visual programs. We showed that VisReP produces strong increases over baseline across multiple tasks, and is competitive with gpt-3.5-turbo. Our work constructed a coarse-valued reward from existing vision-language annotations. Methods like RLAIF [2], ReST [8], and CodeRL [16] all rely on a neural reward model that can provide fine-grained rewards. Learning from fine-grained rewards is much easier than learning from coarse rewards. An interesting direction for future work would be to train a neural reward model for visual program synthesis. Such a reward model could provide fine-grained rewards, and open a broader range of reinforcement learning methods. + +# References + +[1] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V. Le, and Charles Sutton. Program synthesis with large language models. ArXiv, abs/2108.07732, 2021. 1 +[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, E Perez, Jamie Kerr, Jared Mueller, Jeff Ladish, J Landau, Kamal Ndousse, Kamile Lukovs.iuté, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noem'i Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, T. J. Henighan, Tristan Hume, Sam Bowman, Zac Hatfield-Dodds, Benjamin Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom B. Brown, and Jared Kaplan. Constitutional ai: Harmlessness from ai feedback. ArXiv, abs/2212.08073, 2022. 8 +[3] Natasha Butt, Blazej Manczak, Auke J. Wiggers, Corrado Rainone, David W. Zhang, Michael Defferrard, and Taco Cohen. Codeit: Self-improving language models with prioritized hindsight replay. ArXiv, abs/2402.04858, 2024. 3 +[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 2 +[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. ArXiv, abs/2305.14314, 2023. 5 +[6] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 2, 3 +[7] Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. VQA-LOL: visual question answering under the lens of logic. In ECCV, 2020. 1 +[8] Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alexa Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, A. Doucet, Orhan First, and Nando de Freitas. Re- + +inforced self-training (rest) for language modeling. ArXiv, abs/2308.08998, 2023. 2, 3, 4, 8 +[9] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. ArXiv, abs/2211.11559, 2022. 1, 2, 3, 5 +[10] Patrick M. Haluptzok, Matthew Bowers, and Adam Tauman Kalai. Language models can teach themselves to program better. ArXiv, abs/2207.14502, 2022. 2, 3 +[11] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 7, 8 +[12] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 5 +[13] Drew A. Hudson and Christopher D. Manning. GQA: A new dataset for real-world visual reasoning and compositional question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6700-6709. Computer Vision Foundation / IEEE, 2019. 4, 7 +[14] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[15] Sascha Lange, Thomas Gabel, and Martin Riedmiller. Batch Reinforcement Learning, pages 45-73. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 4 +[16] Hung Le, Yue Wang, Akhilesh Deepak Gotmare, Silvio Savarese, and Steven Hoi. CodeRL: Mastering code generation through pretrained models and deep reinforcement learning. In Advances in Neural Information Processing Systems, 2022. 2, 3, 8 +[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3 +[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 8 +[19] Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 3 +[20] Yujia Qin, Shi Liang, Yining Ye, Kunlun Zhu, Lan Yan, Ya Ting Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Marc H. Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis. ArXiv, abs/2307.16789, 2023. 2, 3 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen + +Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 3 +[22] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, I. Evtimov, Joanna Bitton, Manish P Bhatt, Cristian Cantón Ferrer, Aaron Grattafori, Wenhan Xiong, Alexandre D'efossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. Code llama: Open foundation models for code. ArXiv, abs/2308.12950, 2023. 5 +[23] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761, 2023. 2, 3 +[24] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 1, 7, 8 +[25] Ramprasaath R. Selvaraju, Purva Tendulkar, Devi Parikh, Eric Horvitz, Marco Tulio Ribeiro, Besmira Nushi, and Ece Kamar. Squinting at vqa models: Introspecting vqa models with subquestions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8 +[26] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Ke- nely, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell L. Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem- solving with language models. ArXiv, abs/2312.06585, 2023. 3 +[27] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 747-761, Toronto, Canada, 2023. Association for Computational Linguistics. 1, 3 +[28] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 7, 8 +[29] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 7, 8 +[30] Spencer Whitehead, Suzanne Petryk, Vedaad Shakib, Joseph Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus + +Rohrbach. Reliable visual question answering: Abstain rather than answer incorrectly. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[31] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229-256, 1992. 4 +[32] Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations, 2022. 1 +[33] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 3 \ No newline at end of file diff --git a/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/images.zip b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d24f6498b3adf654b7bb57bf789d9bed73cf74a --- /dev/null +++ b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b14454ea422cb7c7e3483bbe91328c3da61941d03ce2163a9e0949ca9182c49 +size 633321 diff --git a/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/layout.json b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..39371aad149ba7ccc723c9329cae975d6e20af9b --- /dev/null +++ b/2024/Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement/layout.json @@ -0,0 +1,7886 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 538, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 538, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 538, + 138 + ], + "type": "text", + "content": "Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "text", + "content": "Zaid Khan" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "text", + "content": " Vijay Kumar BG" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "text", + "content": " Samuel Schulter" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "text", + "content": " Yun Fu" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "text", + "content": " Manmohan Chandraker" + }, + { + "bbox": [ + 69, + 160, + 522, + 175 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "spans": [ + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "type": "text", + "content": "Northeastern University " + }, + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "type": "text", + "content": "NEC Laboratories America " + }, + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 159, + 176, + 434, + 190 + ], + "type": "text", + "content": "UC San Diego" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 242, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 242, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 45, + 242, + 290, + 492 + ], + "type": "text", + "content": "Visual program synthesis is a promising approach to exploit the reasoning abilities of large language models for compositional computer vision tasks. Previous work has used few-shot prompting with frozen LLMs to synthesize visual programs. Training an LLM to write better visual programs is an attractive prospect, but it is unclear how to accomplish this. No dataset of visual programs for training exists, and acquisition of a visual program dataset cannot be easily crowdsourced due to the need for expert annotators. To get around the lack of direct supervision, we explore improving the program synthesis abilities of an LLM using feedback from interactive experience. We propose a method where we exploit existing annotations for a vision-language task to improvise a coarse reward signal for that task, treat the LLM as a policy, and apply reinforced self-training to improve the visual program synthesis ability of the LLM for that task. We describe a series of experiments on object detection, compositional visual question answering, and image-text retrieval, and show that in each case, the self-trained LLM outperforms or performs on par with few-shot frozen LLMs that are an order of magnitude larger. Website:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 494, + 200, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 200, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 200, + 504 + ], + "type": "text", + "content": "https://zaidkhan.me/ViReP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 525, + 127, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 127, + 538 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 127, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Complex visual queries can often be decomposed into simpler subtasks, many of which can be carried out by task-specific perception modules (e.g. object detection, captioning). For example, consider the problem of finding bounding boxes for the phrase \"white mug to the left of the sink\". This is a challenging query for single model such as an open vocabulary object detector. However, this query can be solved by writing a program that composes task-specific perception modules with logic: use an open vocabulary object detector to find a sink and white mugs in the scene, then compare the horizontal center of the sink and the mugs to find white mugs to the left of the sink. Program synthesis with large language models [1] is a promising approach to automate this process, and recent work has shown that proprietary large language" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 217, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 310, + 217, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 310, + 217, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 310, + 217, + 545, + 334 + ], + "type": "image", + "image_path": "66344b4b51bcfc966e6bbfbe756ff8b1eb4ab3f4251b96d71c5c74cf85faf09e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 345, + 529, + 425 + ], + "blocks": [ + { + "bbox": [ + 310, + 335, + 541, + 345 + ], + "lines": [ + { + "bbox": [ + 310, + 335, + 541, + 345 + ], + "spans": [ + { + "bbox": [ + 310, + 335, + 541, + 345 + ], + "type": "text", + "content": "How can we train a LLM to write better visual programs for a task?" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 345, + 529, + 425 + ], + "lines": [ + { + "bbox": [ + 311, + 345, + 529, + 425 + ], + "spans": [ + { + "bbox": [ + 311, + 345, + 529, + 425 + ], + "type": "image", + "image_path": "0ef705e9a0f73bda7686e15c73263592e83053126f9652916b2145f1d7c832d9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 440, + 547, + 507 + ], + "lines": [ + { + "bbox": [ + 304, + 440, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 547, + 507 + ], + "type": "text", + "content": "Figure 1. Visual program synthesis with LLMs has been treated as a " + }, + { + "bbox": [ + 304, + 440, + 547, + 507 + ], + "type": "inline_equation", + "content": "0 / \\mathrm{n}" + }, + { + "bbox": [ + 304, + 440, + 547, + 507 + ], + "type": "text", + "content": "-shot task where the LLM is kept frozen. This limits opportunities for improvement. We ask whether it is possible to train a LLM to write more accurate programs. Given that there is no large scale dataset of accurate visual programs available, we propose improving the LLM using self-training." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 532, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 581 + ], + "type": "text", + "content": "models can write programs for visual tasks [9, 27, 28]. Current approaches for visual program synthesis with LLMs use few-shot prompting and rely on the in-context learning abilities [32] of frozen, proprietary LLMs. (Fig. 1)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Few-shot prompting with frozen LLMs for visual program synthesis as in ViperGPT [28], VisProg [9], or CodeVQA [27] has several limitations. The LLM needs to understand the competencies of the perception modules it is using. A open vocabulary object detector may be able to locate a common attribute-noun phrase such as \"white mug\" without problems, but struggle with a more abstract phrase such as \"microwaveable mug\" [24]. A VQA model might be able to answer \"is the car blue?\" without problems, but fail when logical modifiers are introduced, such as \"is the car not blue?\" [7]. In many cases, we do not precisely know" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14344" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 545, + 149 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 149 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 149 + ], + "type": "table", + "html": "
Task DomainSelf-TrainingSupervisionTool/API UseVisual Task DecompositionGrounded By FeedbackImproves LLM
VisReP (Ours)Visual Program SynthesisYesWeakYesYesYesYes
Haluptzok et al. [10]Programming PuzzlesYesWeakNoNoYesYes
ReST [8]Natural Language UnderstandingYes-NoNoYesYes
VisProg [9]Visual Program SynthesisNo-YesYesNoNo
ViperGPT [28]Visual Program SynthesisNoWeakYesYesNoNo
ToolLLM [20]Tool Usage by APINoStrongYesNoNoYes
GorillaLLM [19]Tool Usage by APINoStrongYesNoNoYes
ToolFormer [23]Natural Language UnderstandingYesWeakYesNoYesYes
", + "image_path": "fde2adfd1f09c7887052d783c301e16cccc39fbb872ceba814607ee7989c09f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 546, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 546, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 546, + 236 + ], + "type": "text", + "content": "Table 1. Differences between our work and similar work. Strong supervision means that the training process requires examples of ground-truth programs to train the LLM. Weak supervision means that the training process does not require ground-truth programs. Tool / API use means that the LLM is required to use substantial functionality implemented by external modules (e.g. an object detector, a web search API) to solve tasks. Visual task decomposition means that the LLM can decompose a complex visual task into primitive subtasks. Grounded by feedback means that the LLM has been optimized not just for syntactic / semantic correctness (program does not hallucinate / cause errors), but for functional correctness (programs produce the correct answer). Improves LLM means that the work proposes a method to improve an LLM for a specific task, rather than using a frozen LLM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 252, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 288, + 373 + ], + "type": "text", + "content": "the weaknesses or competencies of a perception model [30]. Even if this were known, it is difficult to convey all of the competencies / weaknesses of a perception module through in-context examples. Second, it is the case in program synthesis that an LLM often can generate the correct solution to a problem, but the correct solution is not the solution the LLM places the highest probability on [4]. We would like to align the LLM to \"uncover\" the knowledge of the correct solution, but it not clear how to do this in a principled way with few-shot prompting alone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 380, + 266, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 380, + 266, + 403 + ], + "spans": [ + { + "bbox": [ + 66, + 380, + 266, + 403 + ], + "type": "text", + "content": "How can we train a large language model to write better visual programs for a specific task?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 543 + ], + "type": "text", + "content": "Our goal is to optimize the parameters of the language model so the accuracy of the synthesized programs is higher. Existing approaches that train LLMs to improve their ability to programatically use tools / APIs such as GorillaLLM [19], ToolLLM [20] do so by finetuning LLMs on examples of tool use or API use. This cannot be directly applied to visual program synthesis because there are no large scale datasets of visual programs, and collecting such a dataset would be extremely labor intensive. In the absence of a large scale dataset, how do we learn to write better programs for a visual task?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 551, + 266, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 551, + 266, + 597 + ], + "spans": [ + { + "bbox": [ + 66, + 551, + 266, + 597 + ], + "type": "text", + "content": "We posit that grounding a language model with interactive feedback from a generic visual task will improve the general visual program synthesis abilities of the model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "A natural way to learn from feedback is to use reinforcement learning. ReST [8] and RaFT [6] introduce a general framework for reinforced self-training in generative tasks and demonstrate success in machine translation and text-to-image generation. However, a crucial ingredient in their recipe is the availability of a fine-grained reward model. It is difficult to construct a fine-grained reward model for visual program synthesis, given both the absence of human preference datasets for visual programs, and the difficulty" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 252, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 445 + ], + "type": "text", + "content": "of devising a proxy metric. One alternative is to use unit tests to teach a neural reward model or give a coarse-grained reward. This technique has been used successfully in coding challenges by CodeRL [16] and Haluptzok et al. [10], but it is unclear how it can be applied to visual program synthesis. Our key idea is to use existing annotations for a vision-language as improvised unit tests to provide a coarse reward signal. Using the coarse reward signal, we can apply reinforced self-training by treating the language model as a policy and training it with a simple policy gradient algorithm. We alternate synthetic data generation steps in which we sample programs from the language model policy with optimization steps in which we improve the language model policy based on observations from executing the sampled programs. We name our proposed method VisReP, for Visually Reinforced Program Synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 453, + 545, + 620 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 453, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 545, + 488 + ], + "type": "text", + "content": "- We propose optimizing the parameters of a LLM so that the accuracy of the synthesized visual programs is higher, in contrast to previous works that use frozen LLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 489, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 545, + 536 + ], + "type": "text", + "content": "- Since no dataset of accurate visual programs is available for finetuning, we hypothesize that we can instead use feedback from the execution environment to improve the visual program synthesis abilities of a language model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 537, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 545, + 584 + ], + "type": "text", + "content": "- We propose VisReP, an offline, model agnostic recipe for reinforced self-training of large language models for visual program synthesis using existing vision-language annotations with a simple policy gradient algorithm." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 585, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 545, + 620 + ], + "type": "text", + "content": "- Our results show that it is possible to apply reinforced self-training for to improve large language models for visual program synthesis with only coarse rewards." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": "We demonstrate the effectiveness of an CodeLlama-7B policy trained by VisReP on compositional visual question answering " + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "inline_equation", + "content": "(+9\\%)" + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": ", complex object detection " + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "inline_equation", + "content": "(+5\\%)" + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": ", and compositional image-text matching " + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "inline_equation", + "content": "(+15\\%)" + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": " relative to the untrained policy. We show that the policy trained by VisReP exceeds the accuracy of a gpt-3.5-turbo policy on all three tasks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14345" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 133, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 133, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 133, + 104 + ], + "type": "text", + "content": "2.1. Self-Training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 109, + 288, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 157 + ], + "type": "text", + "content": "Self-training is an established paradigm which uses unlabeled data to improve performance. Self-training has been successfully applied in a number of fields. We restrict our coverage to usages with significant overlap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 157, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 289 + ], + "type": "text", + "content": "Program Synthesis Haluptzok et al. [10] showed that LLMs can improve their program synthesis abilities by generating programming puzzles and solving them. CodeRL [16] proposed an actor-critic framework to improve the program synthesis abilities of LLMs for programming problems accompanied by unit tests. CodeIT [3] and Rest-EM [26] also use a similar policy gradient approach for program synthesis. Our problem domain is different from these works, which focus on program synthesis for programming puzzles / problems. In addition, our work has an explicit focus on learning to use an API fluently." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 289, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 396 + ], + "type": "text", + "content": "Alignment ReST [8] and RaFT [6] introduced a generic framework for reinforced self-training and applied it to align machine translation outputs to human preferences and align foundation models on language understanding and image generation tasks respectively. These works share the same basic idea as our work, though they are in a substantially different task domain where human preferences are either known (conversational alignment) or can be estimated with an available neural model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 396, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 458 + ], + "type": "text", + "content": "Vision-Language SelTDA [14] introduced a self-training approach for visual question answering. SelTDA proceeds by pseudolabeling unlabeled data, then finetuning a large VLM on the pseudolabeled data. In contrast to SelTDA, we improve a LLM for visual program synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 464, + 191, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 191, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 191, + 477 + ], + "type": "text", + "content": "2.2. Visual Program Synthesis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 628 + ], + "type": "text", + "content": "Visual program synthesis with LLMs was proposed concurrently by ViperGPT [28], VisProg [9], and CodeVQA [27]. The common points between these three works is that (a) they use pretrained LLMs as code generators (b) they represent complex visual tasks as compositions of primitive visual subtasks (c) they use code to invoke task-specific models to perform the primitive subtasks. Our work is most similar to ViperGPT and CodeVQA as they produce code in a general purpose programming language rather than a DSL. All three works use a proprietary, frozen LLM. In contrast to all three, the focus of our work is on how we can improve the visual program synthesis abilities of an open LLM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 165, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 165, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 165, + 647 + ], + "type": "text", + "content": "2.3. Tool Use with LLMs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Multimodal tool-using LLMs were first introduced by Socratic Models [33]. However, their approach was to create fixed pipelines in which the output of a perception model such as CLIP [21] is fed to a LLM. Later approaches such as GorillaLLM [19] and ToolLLM [20] improved on this by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "treating tool use as a program synthesis problem and creating LLMs that use a broad range of tools by learning to invoke APIs. However, one key limitation of these approaches in the context of visual program synthesis is that they do not learn to decompose problems into subproblems that can be solved by tools. Instead, they are trained to select the right tool for the problem and invoke it. Another limitation is that they are not optimized for functional correctness. They are trained for syntactic and semantic correctness, but they have not been provided feedback on whether their use of tools produces the desired answer. ToolFormer [23] is similar to our work in the sense that the LLM's usage of tools is grounded by feedback, but they focus on natural language understanding tasks rather than visual tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 251, + 362, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 362, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 362, + 263 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 270, + 505, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 505, + 284 + ], + "type": "text", + "content": "3.1. Visual Program Synthesis with LLMs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": "Task Formulation Let " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": " be a visual input and " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": " be a textual query about " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": ". In visual program synthesis, we synthesize a program " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "p = \\pi_{\\theta}(q)" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": " with a program generator " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": ". The program " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": " and visual input " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": " are then fed into the execution engine " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\phi(v, p)" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": " to produce a result " + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 289, + 547, + 362 + ], + "type": "text", + "content": ". The program generator is an auto-regressive large language model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 370, + 546, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 370, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 351, + 370, + 546, + 403 + ], + "type": "interline_equation", + "content": "\\pi_ {\\theta} (\\boldsymbol {y} \\mid \\boldsymbol {x}) = \\prod_ {t = 1} ^ {T} \\pi_ {\\theta} \\left(p _ {t} \\mid \\mathbf {p} _ {1: t - 1}, \\boldsymbol {x}\\right), \\tag {1}", + "image_path": "683fa9cffaa3bc1011e40538fd4ada82bbb1b7b3ac802a10eb906a61dbecb3ee.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{1:t}" + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "text", + "content": " are the tokens of the program, and " + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "text", + "content": " is the input to the large language model. The language model is kept frozen in previous work [28]. Our goal is to optimize the parameters " + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "text", + "content": " of the language model " + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 411, + 547, + 471 + ], + "type": "text", + "content": " so the accuracy of the synthesized programs is higher." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 472, + 547, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 627 + ], + "type": "text", + "content": "Implementation Following ViperGPT [28], we provide the specification of the ImagePatch API concatenated with the textual query " + }, + { + "bbox": [ + 304, + 472, + 547, + 627 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 472, + 547, + 627 + ], + "type": "text", + "content": " as the prompt to the program generator. The synthesized program " + }, + { + "bbox": [ + 304, + 472, + 547, + 627 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 472, + 547, + 627 + ], + "type": "text", + "content": " is a Python program that can invoke any Python builtins, control flow structures, and the ImagePatch API. Our implementation of the ImagePatch API is largely similar to ViperGPT. We remove some API methods that were not required for the tasks we evaluate on (such as l1m_query). We use BLIP [17] and GroundingDINO [18] as perception modules underlying find (object detection), simple-query (visual question answering), and verify_property (attribute verification)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 635, + 444, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 444, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 444, + 647 + ], + "type": "text", + "content": "3.2. Reinforced Self-Training" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Rather than use a frozen large language model as the program generator " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", we would like to optimize the parameters " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " of the language model so the accuracy of the synthesized programs is higher. It is not obvious how to do this. We can't backpropagate through the execution engine " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi(\\pi_{\\theta}(q), v)" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14346" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 280, + 130 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 280, + 130 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 280, + 130 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 280, + 130 + ], + "type": "image", + "image_path": "526901645863a3d4e02b783d757276e55698454498b792b7a9d9885b081a8b22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 147, + 221, + 159 + ], + "lines": [ + { + "bbox": [ + 110, + 147, + 221, + 159 + ], + "spans": [ + { + "bbox": [ + 110, + 147, + 221, + 159 + ], + "type": "text", + "content": "Offline Policy Improvement" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 167, + 284, + 266 + ], + "blocks": [ + { + "bbox": [ + 50, + 167, + 284, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 167, + 284, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 167, + 284, + 266 + ], + "type": "image", + "image_path": "f3c47938ba1c105a42024486873d143a800ecb1bba29453f3bfd7cfef1cd069d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 289, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 342 + ], + "type": "text", + "content": "Figure 2. VisReP can be applied to improve the visual synthesis abilities of an LLM for a vision-language task using existing annotations for a vision-language task (e.g. an object description + image + bounding boxes). A key idea is to construct a coarse reward by comparing the answer produced by a synthesized program to the ground-truth answer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "text", + "content": "directly optimize " + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "text", + "content": ". An alternative might be to use human labor to build a dataset of high-quality visual programs, and train the large language model " + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 46, + 364, + 287, + 449 + ], + "type": "text", + "content": " on the manually-collected dataset. But collecting such a dataset is very labor intensive, and not scalable. Instead, we explore the idea of learning from experience by applying a simple policy gradient method, REINFORCE [31]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": "We propose VisReP, which treats the program synthesis task as a growing batch RL problem [15], inspired by ReST [8]. We first define a coarse discrete reward function " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " from existing annotations for a vision-language task. We then alternate Grow steps, in which we sample trajectories (programs) from the policy (large language model), with Improve steps, in which we apply behavioral cloning with a reward-weighted negative-log likelihood loss to improve the policy. A diagram of our approach is depicted in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": "Grow Step The grow step corresponds to the acting step in reinforcement learning, and can also be seen as synthetic data generation. Let " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(v_1, q_1, y_1), \\ldots, (v_n, q_n, y_n)\\}" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " be a dataset for a vision-language task, where " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " is an image, " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " is a textual query, and " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " is ground-truth for the " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": "-th triplet (e.g. a string for VQA, bounding boxes for object detection). We start with the frozen language model " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\pi_\\theta(p \\mid q)" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " is a synthesized program and " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " is a textual query. The language model " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\pi_\\theta" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " represents our policy. We generate a dataset of trajectories " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " by sampling many programs " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " from the current policy " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\pi_\\theta: p \\sim \\pi_\\theta(p \\mid q)" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "inline_equation", + "content": "q \\sim \\mathcal{D}" + }, + { + "bbox": [ + 46, + 557, + 288, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Improve Step Our goal in this step is to use the dataset of synthetic programs " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " to improve the policy " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": ". First, we" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "define a binary-valued reward function " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "R: p, v, y \\to \\{0, 1\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " on a given program, image, annotation triplet," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 104, + 547, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 104, + 547, + 136 + ], + "spans": [ + { + "bbox": [ + 356, + 104, + 547, + 136 + ], + "type": "interline_equation", + "content": "R (v, p, y) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\phi (p, v) = y \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "74286e7638dea4e27a557a589c740b9a234359673d64b32e89bfffc39919b56a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\phi(p, v)" + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "text", + "content": " is the result of executing the program " + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "text", + "content": " on an image " + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 144, + 547, + 204 + ], + "type": "text", + "content": " is not a program but an existing annotation such as a string for VQA for a bounding box for object detection. To apply behavioral cloning, we then minimize the reward-weighted loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 213, + 546, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 213, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 351, + 213, + 546, + 228 + ], + "type": "interline_equation", + "content": "J (\\theta) = \\mathbb {E} _ {(q, p) \\sim \\mathcal {D} _ {g}} [ R (v, p) \\mathcal {L} (p, q; \\theta) ] \\tag {3}", + "image_path": "97e8d9948d8a1e8c54849ec66ef2081de16d0dd2c93d4a27677efada783dd7a7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 235, + 511, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 511, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 511, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 235, + 511, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(p,q;\\theta)" + }, + { + "bbox": [ + 305, + 235, + 511, + 247 + ], + "type": "text", + "content": " is the negative log-likelihood loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 255, + 546, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 546, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {N L L}} (p, q; \\theta) = - \\mathbb {E} _ {(q, p) \\sim \\mathcal {D} _ {g}} \\left[ \\sum_ {t = 1} ^ {T} \\log \\pi_ {\\theta} \\left(p _ {t} \\mid p _ {1: t - 1}, q\\right) \\right] \\tag {4}", + "image_path": "7daef9c01ff496c6f1ce7321f1df3c9d808803f85335a238a5c4265b61c1cf16.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "text", + "content": "over the pairs of textual queries " + }, + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "text", + "content": " and synthetic programs " + }, + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 305, + 299, + 546, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": "Because the reward function only takes on binary values, we can simplify this and implement it by: First, generating a dataset of synthetic programs " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g = \\{\\pi_\\theta (q):\\forall q\\in \\mathcal{D}\\}" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": " using the LLM " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": " on a dataset " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": ". Next, filtering " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g^\\prime = \\{(q,v,p\\in \\mathcal{D}_g:R(q,v,p) > 0\\}" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": ", which corresponds to executing all synthetic programs and only keeping those that give correct answers. Finally, we finetune the language model " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": " on the filtered dataset " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g^\\prime" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": " using the standard causal language modeling loss. We then iterate the process, initiating a new synthetic data generation step with the improved policy " + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}^{\\prime}" + }, + { + "bbox": [ + 304, + 323, + 546, + 454 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 454, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 547, + 514 + ], + "type": "text", + "content": "Iteration For the initial grow step, we use a frozen language model as the initial policy. For example, we use the pretrained codellama-7b-instruct-hf as the policy in the initial grow step. In subsequent steps, we use the policy trained in the previous improve step for the grow step." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 524, + 469, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 469, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 469, + 538 + ], + "type": "text", + "content": "4. Understanding Self-Training" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 567 + ], + "type": "text", + "content": "Our goal in this section is to characterize the stability and sample efficiency of VisReP. We want to understand:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 568, + 547, + 616 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 306, + 568, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 568, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 306, + 568, + 547, + 592 + ], + "type": "text", + "content": "1. How does applying VisReP change the accuracy of synthesized programs?" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 592, + 473, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 473, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 473, + 604 + ], + "type": "text", + "content": "2. What happens as VisReP is repeated?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 604, + 531, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 531, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 531, + 616 + ], + "type": "text", + "content": "3. How does data scarcity and diversity affect VisReP?" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 623, + 403, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 403, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 403, + 635 + ], + "type": "text", + "content": "4.1. Implementation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "We start off with the GQA [13] dataset for visual question answering. We choose GQA because each question in GQA was constructed programatically and is thus a good candidate to be answered by program synthesis. GQA has over 2M questions, each belonging to one of " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\approx 100" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " question types. We construct a training set by sampling 100 questions for" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14347" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 542 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 542 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 542 + ], + "type": "image", + "image_path": "f7968f0e6cdf62333acfd91d8a7b4b75df1392b20dd62ccb7564916bd4920de7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 555, + 548, + 632 + ], + "lines": [ + { + "bbox": [ + 45, + 555, + 548, + 632 + ], + "spans": [ + { + "bbox": [ + 45, + 555, + 548, + 632 + ], + "type": "text", + "content": "Figure 3. Self-training with VisReP produces qualitatively better programs. Here, we show programs written by the initial policy (on the left) and the policy after 10 iterations of self-training on GQA (on the right). In VQA example, the initial policy does not specifically check whether the empty basket is plastic. In the object detection example, the reasoning of the initial policy is correct, but it issues a confusingly worded query to the simple-query module, which returns the wrong answer. The learned policy uses simple-query more appropriately. In the image-text matching example, in the initial policy tries to use the object detector to search directly for \"meat in a box\" and \"donuts on a plate\", but this is too complicated for the object detector to localize. After self-training, the LLM policy no longer makes this mistake." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "each question type, for a total of " + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\approx 10\\mathrm{k}" + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": " visual questions and answers. We construct a validation set following Gupta and Kembhavi [9]. We use the CodeLlama [22] family of models as our initial policy. We use LoRA [12] adapters during the Improve steps. We use the hyperparameters suggested by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 653, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 677 + ], + "type": "text", + "content": "Dettmers et al. [5]. Full implementation details are in the supplement." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14348" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 82, + 547, + 259 + ], + "blocks": [ + { + "bbox": [ + 184, + 69, + 434, + 81 + ], + "lines": [ + { + "bbox": [ + 184, + 69, + 434, + 81 + ], + "spans": [ + { + "bbox": [ + 184, + 69, + 434, + 81 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 184, + 69, + 434, + 81 + ], + "type": "text", + "content": " Training Accuracy from Step 1 to Step 10 of Self-Training" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 82, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 47, + 82, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 82, + 547, + 259 + ], + "type": "image", + "image_path": "8cfa5e6436ce0085f6cd0d2bc693ad92fd2c51e33bdddf37c6bb8bbcb3abfb64.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 318, + 289, + 459 + ], + "blocks": [ + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "type": "text", + "content": "Figure 4. Iteratively applying VisReP allows a LLM to self-improve improve on almost all of GQA's " + }, + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\approx 100" + }, + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "type": "text", + "content": " question types. The base of each bar is set to the accuracy of the initial policy (codellama-7b-instruct). A green bar indicates question types on which the policy at iteration 10 improved over the initial policy, and a red bar indicates question types on which the policy at iteration 10 was worse than the initial policy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 318, + 289, + 459 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 289, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 289, + 459 + ], + "type": "image", + "image_path": "fbd95f5ddc46e3bf5e646ac0c4a9db7effe1c6e91cf7b22792dd7e5b3aaf9425.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 465, + 288, + 554 + ], + "lines": [ + { + "bbox": [ + 46, + 465, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 288, + 554 + ], + "type": "text", + "content": "Figure 5. Supplying a small amount of human written corrections as in-context examples during training can increase the stability of the self-training process (green line). We show validation accuracy on GQA through multiple iterations of self-training with a policy instantiated from CodeLlama-7b. Without these corrections, proliferating errors cause performance to degrade in later iterations (red line). The translucent shading around each line indicates the standard deviation over 5 evaluations on the validation set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 574, + 287, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 588 + ], + "type": "text", + "content": "4.2. Persistent Errors Harm Iterated Self-Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 289, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 701 + ], + "type": "text", + "content": "Applying the formulation of self-training in Sec. 3.2 results in a improvement, but iterating it further results in program synthesis quality degrading, rather than increasing (red line in Fig. 5). This is due to the self-training process inadvertently reinforcing incorrect reasoning. A program that uses flawed reasoning can occasionally produce a correct answer. The language model can thus be rewarded for a program that is right for the wrong reasons. If this goes uncorrected, the language model will learn incorrect reasoning patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 289, + 714 + ], + "type": "text", + "content": "We hypothesize that providing a small number of human-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "text", + "content": "written corrections for persistent reasoning errors can stabilize the self-training process. We use the question type annotations in GQA to identify question types for which training accuracy decreases over time. These are question types which the language model is not able to self-improve on. We denote them " + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\text {hard }}" + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "text", + "content": ". For each question type in " + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\text {hard }}" + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "text", + "content": ", we randomly sample one question " + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "text", + "content": " for which the language model synthesized a program that produced the wrong answer. We examine the reasoning in that program, and if the reasoning is flawed, we correct it. We repeat this until we have a program with correct reasoning for each question type in " + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\text {hard }}" + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "text", + "content": ", and denote the bank of correct programs as " + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text {gold }}" + }, + { + "bbox": [ + 304, + 322, + 547, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "text", + "content": "We then retrieve from " + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text{gold}}" + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "text", + "content": " during self-training for use as in-context examples. If a question is annotated with a question type in " + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "inline_equation", + "content": "Q_{\\text{hard}}" + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "text", + "content": ", we retrieve a correct human-written program from " + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text{gold}}" + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "text", + "content": " and use it as an in-context example. If a question is not annotated with a question type in " + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "inline_equation", + "content": "Q_{\\text{hard}}" + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "text", + "content": ", we use a \"default\" in-context example which is the same for all question types not in " + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "inline_equation", + "content": "Q_{\\text{hard}}" + }, + { + "bbox": [ + 304, + 466, + 548, + 586 + ], + "type": "text", + "content": ". We show in Fig. 5 (green line) that this stabilizes self-training and allows the language model to self-improve across all but a few question types (Fig. 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 597, + 529, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 529, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 529, + 611 + ], + "type": "text", + "content": "4.3. Effect of Data Availability on Self-Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": "Training With Less Data We explore this in a controlled setting, by manipulating the number of samples per question type in GQA. Recall that we originally sample 100 questions per question type for self-training. This dataset had " + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\approx 10\\mathrm{k}" + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": " questions. We construct a training set with only 10 and 1 question per question type, for a total of " + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\approx 1000" + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\approx 100" + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": " questions respectively. Self-training improves upon the baseline (Fig. 6) even when there is an order of magni" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14349" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 209 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 209 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 209 + ], + "type": "image", + "image_path": "2111617b308482b9c5768a3544789a5eeaae4588bdf5132519f2602b6ac8d062.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": "Figure 6. VisReP works even when the amount of available data is reduced by an order of magnitude. We show validation accuracy on GQA. The notation " + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "inline_equation", + "content": "n \\times k" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": " samples per question type, with " + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": " passes at each sample. For example " + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "inline_equation", + "content": "10 \\times 10" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": " indicates 10 samples per question type, with 10 passes per sample. Although " + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "inline_equation", + "content": "10 \\times 10" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": " has 10x fewer unique samples than " + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "inline_equation", + "content": "100 \\times 1" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": ", there is a " + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "inline_equation", + "content": "< 2\\%" + }, + { + "bbox": [ + 46, + 215, + 289, + 305 + ], + "type": "text", + "content": " accuracy difference between them, indicating that more passes per instance can partially mitigate data scarcity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 373 + ], + "type": "text", + "content": "tude decrease in training data " + }, + { + "bbox": [ + 46, + 326, + 287, + 373 + ], + "type": "inline_equation", + "content": "(100\\to 10)" + }, + { + "bbox": [ + 46, + 326, + 287, + 373 + ], + "type": "text", + "content": ". Only when the amount of available training data is reduced by two orders of magnitude " + }, + { + "bbox": [ + 46, + 326, + 287, + 373 + ], + "type": "inline_equation", + "content": "(100\\to 1)" + }, + { + "bbox": [ + 46, + 326, + 287, + 373 + ], + "type": "text", + "content": " does self-training fail to produce an appreciable increase in performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": "Is it possible to mitigate data scarcity? We previously showed that the benefits of self-training reduce when available data is reduced significantly. We now test whether we can mitigate this data scarcity by allowing " + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": " multiple attempts at a query " + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": " during the Grow step. Concretely, we allow " + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": " a total of 10 tries at each query under the setting in which we train with 1 and 10 samples per question type, for a total of 1k and 10k total samples respectively. We show in Fig. 6 that this mitigates the effect of reduced data. Although the data poor " + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "inline_equation", + "content": "1 \\times 10" + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "inline_equation", + "content": "10 \\times 10" + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": " have 10x fewer unique questions than " + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "inline_equation", + "content": "10 \\times 1" + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "inline_equation", + "content": "100 \\times 1" + }, + { + "bbox": [ + 46, + 374, + 288, + 518 + ], + "type": "text", + "content": ", their performance is within a standard deviation of their data rich counterparts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 526, + 275, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 275, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 275, + 540 + ], + "type": "text", + "content": "4.4. Quantifying Changes in Syntactic Structure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 677 + ], + "type": "text", + "content": "How do the programs synthesized by the policy change as self-training is iterated? We examine this by looking at how many unique abstract syntax trees are produced during the Grow step of each iteration. We parse the synthesized programs into abstract syntax trees, and then normalize the trees to remove irrelevant details such as variable names. In the left panel of Fig. 7, we show that the diversity of syntactic forms drops over time. At the beginning, the policy produces a large number of syntactic forms, but appears to \"hone in\" on a smaller number of forms as self-training continues, and the number of unique syntactic forms drops by almost half." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "A remarkably stable set of syntactic forms is conserved from step to step, roughly " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\approx 700" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " (row above diagonal in right panel of Fig. 7). However, the syntactic forms produced" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 425, + 188 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 425, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 425, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 425, + 188 + ], + "type": "image", + "image_path": "653672caefc87f0f4ac555f55469f762680e7f4462711d859412d2035775af08.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 194, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 294 + ], + "type": "text", + "content": "Figure 7. As self-training is iterated, the LLM policy \"hones in\" on a smaller set of syntactic forms, and gradually evolves away from syntactic forms produced by the initial policy. Left Panel: Number of unique normalized abstract syntax trees seen during each iteration of VisReP. Right Panel: Number of unique normalized abstract syntax trees in common between each training step. For example, the entry in row 1, column 6 corresponds to the number of unique abstract syntax trees produced by both the policy in iteration 1 (initial policy) and the policy in iteration 6." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 76, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 426, + 76, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 426, + 76, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 426, + 76, + 545, + 178 + ], + "type": "image", + "image_path": "14447310f7b76e4d930e540b76eb8d418796293d5a9cfbf024de25a3450da451.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "text", + "content": "by the policy are gradually evolving away from the syntactic forms the initial policy tries, which can be seen in the darkening of the first row in Fig. 7. Despite the coarse reward scheme, the LLM policy gradually explores and learns new syntactic forms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 399, + 499, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 499, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 499, + 412 + ], + "type": "text", + "content": "5. Evaluating Functional Correctness" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 547, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 457 + ], + "type": "text", + "content": "We measure the functional correctness of the programs synthesized by the self-trained LLM policy " + }, + { + "bbox": [ + 304, + 422, + 547, + 457 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}^{\\prime}" + }, + { + "bbox": [ + 304, + 422, + 547, + 457 + ], + "type": "text", + "content": " across three compositional tasks, with the aim of understanding whether:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 460, + 547, + 520 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 460, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 547, + 495 + ], + "type": "text", + "content": "1. Are the programs produced after self-training more functionally correct than programs produced before self-training?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 495, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 547, + 520 + ], + "type": "text", + "content": "2. Is it possible to exceed or match the performance of a much larger proprietary LLM with self-training?" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "For compositional VQA, we use the GQA [13] dataset for the reasons outlined in Sec. 4.1. For complex object detection, we choose Omnilabel [24]. Omnilabel contains 28K free-form object descriptions over 25K images, and is a challenging task for existing open-vocabulary object detectors due to the complexity of the object descriptions. For compositional image-text matching, we choose WinoGround [29] and SugarCrepe [11]. State-of-art vision-language models have trouble reaching above chance accuracy on WinoGround, but SugarCrepe is substantially easier. However, both of these tasks pose significant problems for the ImagePatch API, because many of the relationships mentioned in the text are challenging to detect with the available perception modules. For all experiments, we use ViperGPT[28] as the backbone and adopt their prompts. Due to space limitations, many experimental details are in the supplement." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14350" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 148 + ], + "type": "table", + "html": "
VQAObject DetectionImage-Text Matching
MethodLLMGQAOmnilabelOmnilabel-HardWinogroundSugarCREPE
Frozen Proprietary LLMGPT-3.5-turbo53.9 ± 0.840.0 ± 1.226.0 ± 1.145.6 ± 1.648.9 ± 0.7
Frozen Open LLMCodeLlama-7B50.0 ± 1.737.3 ± 1.623.7 ± 1.441.3 ± 0.0343.5 ± 0.8
Open LLM + VisRePCodeLlama-7B59.2 ± 1.442.4 ± 1.028.1 ± 0.952.7 ± 0.658.7 ± 1.5
", + "image_path": "a57b71079d1d969b86a044c99f65aee528999a49d92ee26451677be6270833bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 201, + 285, + 262 + ], + "blocks": [ + { + "bbox": [ + 48, + 157, + 546, + 190 + ], + "lines": [ + { + "bbox": [ + 48, + 157, + 546, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 157, + 546, + 190 + ], + "type": "text", + "content": "Table 2. An open LLM policy self-trained with our method substantially outperforms the open policy without self-training, and even outperforms a gpt-3.5-turbo policy. All results use ViperGPT [28] as the backbone. " + }, + { + "bbox": [ + 48, + 157, + 546, + 190 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 48, + 157, + 546, + 190 + ], + "type": "text", + "content": " numbers are the standard deviation over 5 runs. On all datasets except Omnilabel, we report accuracy. On Omnilabel, we report Macro-F1. Higher is better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 201, + 285, + 262 + ], + "lines": [ + { + "bbox": [ + 49, + 201, + 285, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 285, + 262 + ], + "type": "table", + "html": "
Code GeneratorX-Dataset GeneralizationX-Task Generalization
VQAv2OK-VQAOmnilabelSugarCrepe
Frozen Few-Shot (7B)46.6 ± 1.112.6 ± 2.137.3 ± 1.643.5 ± 0.8
VisReP on GQA (7B)61.0 ± 1.033.7 ± 1.939.9 ± 0.751.0 ± 1.5
VisReP Advantage+14.4+21.1+2.6+7.5
", + "image_path": "449434d9b003b48ca66099610c2821c651a80366bf11760d1b50f91907a02f66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "type": "text", + "content": "Table 3. VisReP improves benchmark agnostic visual program synthesis. A policy self-trained on GQA with VisReP writes better programs for other VQA datasets and other task types." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 49, + 319, + 162, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 162, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 162, + 331 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 434 + ], + "type": "text", + "content": "For each task, we apply VisReP as described in Sec. 3.2, and evaluate on a held-out subset. For a comparison with a large proprietary LLM, we use gpt-3.5-turbo. We evaluate on a subsampled version of each dataset to reduce token costs. Every LLM is provided the same prompts. Each prompt consists of the ImagePatch API specification used in ViperGPT [28], and 3 in-context examples for each task except for object detection, for which we provide 5 in-context examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 436, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 663 + ], + "type": "text", + "content": "We use GQA as described in Sec. 4.1. We prepare a compositional subset of Omnilabel [24] by filtering out all descriptions less than two words in length. We then sample a subset of 500 for evaluation, and a subset of 500 for training. To prepare Omnilabel-Hard, we use run a state of the art open-vocabulary object detector (GroundingDINO [18]) on the remaining OmniLabel samples, and select those which GroundingDINO completely fails on (no detections) to obtain a hard slice. We then sample a subset of 500 from the hard slice for evaluation. For SugarCrepe [11], we sample 100 positives and their associated negatives from each of the 6 categories, for a total of 600 balanced image-text pairs for validation. We sample 100 of the remaining instances from each category for training. We use all of WinoGround, as it is small enough that there is no need to subsample it. On WinoGround[29], we evaluate the policy trained on SugarCrepe rather than training on it. For VQAv2, we sample 10 questions for each of the top-50 most common answers from the compositional subset curated by [25]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 287, + 712 + ], + "type": "text", + "content": "Examples of the inputs for each task are in Fig. 3. We use nucleus sampling with identical parameters for all local LLMs. We use the API default temperature for gpt-3.5-turbo. More details are in the supplement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 211, + 376, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 211, + 376, + 222 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 376, + 222 + ], + "type": "text", + "content": "5.2. Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "text", + "content": "Across all three tasks, the policy trained by VisReP outperforms both the gpt-3.5-turbo policy, and the initial CodeLlama-7b policy (Tab. 2). On GQA, the self-trained policy achieves an absolute improvement of almost " + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "text", + "content": " over the initial policy, and " + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "text", + "content": " over the gpt-3.5-turbo policy. On Omnilabel, self-training produces a " + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "text", + "content": " improvement in Macro-F1 score with only 500 training samples. On Omnilabel-Hard, we demonstrate that the visual program synthesis paradigm can localize objects that state of the art open-vocabulary object detectors are unable to localize (Omnilabel-Hard was constructed by selecting instances GroundingDino[18]) cannot localize). Even on Omnilabel-Hard, the self-trained policy outperforms the others. WinoGround and SugarCrepe are difficult to solve by visual program synthesis because many of the relationships are hard to detect with the available perception modules. Despite the intrinsic difficulty of compositional image-text matching for the ImagePatch API, VisReP produces an increase of " + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "inline_equation", + "content": "+15\\%" + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "text", + "content": " over the baseline policy. The policy trained on SugarCrepe transfers to WinoGround, outperforming the baseline policy by " + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "inline_equation", + "content": "+10\\%" + }, + { + "bbox": [ + 307, + 229, + 546, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 479, + 460, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 460, + 490 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 460, + 490 + ], + "type": "text", + "content": "6. Conclusion & Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 498, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 546, + 712 + ], + "type": "text", + "content": "While few-shot prompting of LLMs for visual program synthesis has produced impressive results, it has limitations, because writing good visual programs requires experience with the visual world and the perception modules at ones disposal. We presented VisReP, which improves a LLM's program synthesis abilities using feedback from executing visual programs. We showed that VisReP produces strong increases over baseline across multiple tasks, and is competitive with gpt-3.5-turbo. Our work constructed a coarse-valued reward from existing vision-language annotations. Methods like RLAIF [2], ReST [8], and CodeRL [16] all rely on a neural reward model that can provide fine-grained rewards. Learning from fine-grained rewards is much easier than learning from coarse rewards. An interesting direction for future work would be to train a neural reward model for visual program synthesis. Such a reward model could provide fine-grained rewards, and open a broader range of reinforcement learning methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 316, + 757 + ], + "type": "text", + "content": "14351" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V. Le, and Charles Sutton. Program synthesis with large language models. ArXiv, abs/2108.07732, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 289, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 289, + 312 + ], + "type": "text", + "content": "[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, E Perez, Jamie Kerr, Jared Mueller, Jeff Ladish, J Landau, Kamal Ndousse, Kamile Lukovs.iuté, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noem'i Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, T. J. Henighan, Tristan Hume, Sam Bowman, Zac Hatfield-Dodds, Benjamin Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom B. Brown, and Jared Kaplan. Constitutional ai: Harmlessness from ai feedback. ArXiv, abs/2212.08073, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 316, + 289, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 289, + 359 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 289, + 359 + ], + "type": "text", + "content": "[3] Natasha Butt, Blazej Manczak, Auke J. Wiggers, Corrado Rainone, David W. Zhang, Michael Defferrard, and Taco Cohen. Codeit: Self-improving language models with prioritized hindsight replay. ArXiv, abs/2402.04858, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 361, + 289, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 289, + 548 + ], + "type": "text", + "content": "[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 552, + 289, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 552, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 53, + 552, + 289, + 583 + ], + "type": "text", + "content": "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. ArXiv, abs/2305.14314, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 587, + 289, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 587, + 289, + 631 + ], + "spans": [ + { + "bbox": [ + 53, + 587, + 289, + 631 + ], + "type": "text", + "content": "[6] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 634, + 289, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 634, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 634, + 289, + 666 + ], + "type": "text", + "content": "[7] Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. VQA-LOL: visual question answering under the lens of logic. In ECCV, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 670, + 289, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 289, + 712 + ], + "type": "text", + "content": "[8] Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alexa Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, A. Doucet, Orhan First, and Nando de Freitas. Re-" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "text", + "content": "inforced self-training (rest) for language modeling. ArXiv, abs/2308.08998, 2023. 2, 3, 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 96, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 129 + ], + "type": "text", + "content": "[9] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. ArXiv, abs/2211.11559, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[10] Patrick M. Haluptzok, Matthew Bowers, and Adam Tauman Kalai. Language models can teach themselves to program better. ArXiv, abs/2207.14502, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 165, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 547, + 218 + ], + "type": "text", + "content": "[11] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 220, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 262 + ], + "type": "text", + "content": "[12] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 266, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 547, + 330 + ], + "type": "text", + "content": "[13] Drew A. Hudson and Christopher D. Manning. GQA: A new dataset for real-world visual reasoning and compositional question answering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 6700-6709. Computer Vision Foundation / IEEE, 2019. 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 332, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 396 + ], + "type": "text", + "content": "[14] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 399, + 547, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 547, + 432 + ], + "type": "text", + "content": "[15] Sascha Lange, Thomas Gabel, and Martin Riedmiller. Batch Reinforcement Learning, pages 45-73. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 434, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 547, + 487 + ], + "type": "text", + "content": "[16] Hung Le, Yue Wang, Akhilesh Deepak Gotmare, Silvio Savarese, and Steven Hoi. CodeRL: Mastering code generation through pretrained models and deep reinforcement learning. In Advances in Neural Information Processing Systems, 2022. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 490, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 547, + 521 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 523, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 547, + 578 + ], + "type": "text", + "content": "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 580, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 547, + 612 + ], + "type": "text", + "content": "[19] Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 614, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 547, + 678 + ], + "type": "text", + "content": "[20] Yujia Qin, Shi Liang, Yining Ye, Kunlun Zhu, Lan Yan, Ya Ting Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Marc H. Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis. ArXiv, abs/2307.16789, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 681, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 712 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14352" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 288, + 127 + ], + "type": "text", + "content": "Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 215 + ], + "type": "text", + "content": "[22] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, I. Evtimov, Joanna Bitton, Manish P Bhatt, Cristian Cantón Ferrer, Aaron Grattafori, Wenhan Xiong, Alexandre D'efossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. Code llama: Open foundation models for code. ArXiv, abs/2308.12950, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "text", + "content": "[23] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[24] Samuel Schulter, Vijay Kumar B G, Yumin Suh, Konstantinos M. Dafnis, Zhixing Zhang, Shiyu Zhao, and Dimitris Metaxas. Omnilabel: A challenging benchmark for language-based object detection. In ICCV, 2023. 1, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "type": "text", + "content": "[25] Ramprasaath R. Selvaraju, Purva Tendulkar, Devi Parikh, Eric Horvitz, Marco Tulio Ribeiro, Besmira Nushi, and Ece Kamar. Squinting at vqa models: Introspecting vqa models with subquestions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 372, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 288, + 513 + ], + "type": "text", + "content": "[26] Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Ke- nely, Kevin Swersky, Kshitteej Mahajan, Laura Culp, Lechao Xiao, Maxwell L. Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Brian Warkentin, Yundi Qian, Ethan Dyer, Behnam Neyshabur, Jascha Narain Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem- solving with language models. ArXiv, abs/2312.06585, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 514, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 602 + ], + "type": "text", + "content": "[27] Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. Modular visual question answering via code generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 747-761, Toronto, Canada, 2023. Association for Computational Linguistics. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 603, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 646 + ], + "type": "text", + "content": "[28] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "text", + "content": "[29] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[30] Spencer Whitehead, Suzanne Petryk, Vedaad Shakib, Joseph Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 251 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Rohrbach. Reliable visual question answering: Abstain rather than answer incorrectly. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[31] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229-256, 1992. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[32] Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 186, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 251 + ], + "type": "text", + "content": "[33] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14353" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_content_list.json b/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a6210898821bd6cb63619a6f71bce989ffcfb930 --- /dev/null +++ b/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "Self-correcting LLM-controlled Diffusion Models", + "text_level": 1, + "bbox": [ + 235, + 130, + 733, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsung-Han Wu*", + "bbox": [ + 130, + 180, + 261, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long Lian*", + "bbox": [ + 295, + 181, + 387, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joseph E. Gonzalez \nUC Berkeley", + "bbox": [ + 419, + 181, + 578, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boyi Li†", + "bbox": [ + 619, + 180, + 689, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trevor Darrell†", + "bbox": [ + 723, + 181, + 844, + 196 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5134225e7f860ac442562a0b5ef10c0cc620556e3b7975b8e52780e41086f53.jpg", + "image_caption": [ + "DALL-E3", + "A realistic cartoon-style image with a princess and four dwarfs", + "Figure 1. Existing diffusion-based text-to-image generators (e.g., DALL-E 3 [16]) generally struggle to precisely generate images that correctly align with complex input prompts, especially for the ones that require numeracy and spatial relationships. Our Self-correcting LLM-controlled Diffusion (SLD) framework empowers these diffusion models to automatically and iteratively rectify inaccuracies through applying a set of latent space operations (addition, deletion, repositioning, etc.), resulting in enhanced text-to-image alignment." + ], + "image_footnote": [], + "bbox": [ + 81, + 257, + 267, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df5fa1e7fbafa99f44cf65d8f65d7f9505a6dfdd01c0b90bfcca41c1611c22c9.jpg", + "image_caption": [ + "Numeracy Correction", + "DALL-E 3 + SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 289, + 257, + 475, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1efcfb89b8bb173e5258d71b5e811377ddeb59708f3d3905bde442218b50d165.jpg", + "image_caption": [ + "Spatial Correction", + "DALL-E3", + "A vivid photo with a woman on the right and a clown on the left walking in a dirty alley" + ], + "image_footnote": [], + "bbox": [ + 493, + 257, + 679, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7bd7f0277e99d0cf48cb67ad61b9ce1a46e68da8113c0b21a7c0f38c56347ca.jpg", + "image_caption": [ + "DALL-E 3 + SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 702, + 257, + 887, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 267, + 513, + 346, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image generation has witnessed significant progress with the advent of diffusion models. Despite the ability to generate photorealistic images, current text-to-image diffusion models still often struggle to accurately interpret and follow complex input text prompts. In contrast to existing models that aim to generate images only with their best effort, we introduce Self-correcting LLM-controlled Diffusion (SLD). SLD is a framework that generates an image from the input prompt, assesses its alignment with the prompt, and performs self-corrections on the inaccuracies in the generated image. Steered by an LLM controller, SLD turns text-to-image generation into an iterative closed-loop process, ensuring correctness in the resulting image. SLD is not only training-free but can also be seamlessly integrated with diffusion models behind API access, such as DALL-E 3, to further boost the performance of state-of-the-art diffusion models. Experimental results show that our approach can rectify a majority of incorrect generations, particularly in generative numeracy, attribute binding, and spatial relationships. Furthermore, by simply adjusting the instructions to the LLM, SLD can perform image editing tasks, bridging the gap between text-to-image generation", + "bbox": [ + 75, + 546, + 472, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and image editing pipelines. Our code is available at: https://self-correcting-llm-diffusion.github.io.", + "bbox": [ + 500, + 516, + 890, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 554, + 632, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image generation has made remarkable advancements, especially with the advent of diffusion models. However, these models often struggle with interpreting complex input text prompts, particularly those that require skills such as understanding the concept of numeracy, spatial relationships, and attribute binding with multiple objects. Despite the astonishing scaling of model sizes and training data, these challenges, as illustrated in Fig. 1, are still present in state-of-the-art open-source and proprietary diffusion models.", + "bbox": [ + 496, + 580, + 890, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several research and engineering efforts aim to overcome these limitations. For instance, methods such as DALL-E 3 [16] focus on the diffusion training process and incorporate high-quality captions into the training data at a massive scale. However, this approach not only incurs substantial costs but also frequently falls short in generating accurate images from complicated user prompts, as shown in Fig. 1. Other work harnesses the power of external models for a better understanding of the prompt in the inference process before the actual image generation. For example, [6, 11] leverages Large Language Models (LLMs) to pre", + "bbox": [ + 496, + 734, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Equal advising.", + "bbox": [ + 83, + 886, + 307, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6327", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "process textual prompts into structured image layouts and thus ensures the preliminary design aligns with the user's directives. However, such integration does not resolve the inaccuracies produced by the downstream diffusion models, particularly in images with complex scenarios like multiple objects, cluttered arrangements, or detailed attributes.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Drawing inspiration from the process of a human painting and a diffusion model in generating images, we observe a key distinction in their approach to creation. Consider a human artist tasked with painting a scene featuring two cats. Throughout the painting process, the artist remains cognizant of this requirement, ensuring that two cats are indeed present before considering the work complete. Should the artist find only one cat depicted, an additional one would be added to meet the prompt's criteria. This contrasts sharply with current text-to-image diffusion models, which operate on an open-loop basis. These models generate images through a predetermined number of diffusion steps and present the output to the user, regardless of its alignment with the initial user prompt. Such a process, irrespective of scaling training data or LLM pre-generation conditioning, lacks a robust mechanism to ensure the final image aligns with the user's expectations.", + "bbox": [ + 75, + 183, + 468, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this, we propose our method Self-correcting LLM-controlled Diffusion (SLD) that performs self-checks to confidently offer users guarantees of the alignment between the prompt and the generated images. Departing from conventional single-round generation methods, SLD is a novel closed-loop approach that equips diffusion models with the ability to iteratively identify and rectify errors. Our SLD framework, illustrated in Fig. 2, contains two main components: LLM-driven object detection as well as LLM-controlled assessment and correction.", + "bbox": [ + 75, + 443, + 468, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The SLD pipeline follows a standard text-to-image generation setting. Given a textual prompt that outlines the desired image, SLD begins with calling an image generation module (e.g., the aforementioned open-loop text-to-image diffusion models) to generate an image in a best-effort fashion. Given that these open-loop generators do not guarantee an output that aligns perfectly with the prompt, SLD then conducts a thorough evaluation of the produced image against the prompt, with an LLM parsing key phrases for an open-vocabulary detector to check. Subsequently, an LLM controller takes the detected bounding boxes and the initial prompt as input and checks for potential mismatches between the detection results and the prompt requirements, suggesting appropriate self-correction operations, such as adding, moving, and removing objects. Finally, utilizing a chosen base diffusion model (e.g., Stable Diffusion [20]), SLD employs latent space composition to implement these adjustments, thereby ensuring that the final image accurately reflects the user's initial text prompt.", + "bbox": [ + 75, + 595, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notably, SLD is agnostic to the initial generation pro", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cess, which allows us to use DALL-E 3 APIs as a subroutine to generate initial images and then employ the open-source Stable Diffusion model family to conduct latent operations to attain self-correction. Furthermore, none of these operations require additional training on our base diffusion model [20], which easily allows our method to be applied to various diffusion models without external human annotation or training.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate that our SLD framework can achieve significant improvement over current diffusion-based methods on complex prompts with the LMD benchmark [11]. The results show that our method can surpass $\\mathrm{LMD + }$ , which is a strong baseline that already leverages LLM in the image process generation, by $9.0\\%$ . More importantly, with DALL-E 3 for initial generation, the generated images from our method achieve $26.5\\%$ performance gains compared to ones before self-correction.", + "bbox": [ + 496, + 212, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, since the SLD pipeline is agnostic to the initially generated image, it can easily be transformed into an image editing pipeline by simply changing the prompts to the LLM. While text-to-image generation and image editing are often treated as distinct tasks by the generative modeling community, our SLD can perform these two tasks with a unified pipeline. We list our key contributions below:", + "bbox": [ + 496, + 347, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. SLD is the first to integrate a detector and an LLM to self-correct generative models, ensuring accurate generation without extra training or external data.", + "2. SLD offers a unified solution for both image generation and editing, enabling enhanced text-to-image alignment for any image generator (e.g., DALL-E 3) and object-level editing on any images.", + "3. Experimental results show that SLD can correct a majority of incorrect generations, particularly in aspects of numeracy, attribute binding, and spatial relationships." + ], + "bbox": [ + 500, + 455, + 890, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 625, + 640, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 650, + 779, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based text-to-image generation has advanced significantly. Initial studies [19-21] showed diffusion models' ability to create high-quality images, but they struggle with complex prompts. Subsequent research [3, 10, 18, 24-26] has incorporated additional inputs such as keypoints and bounding boxes to control the diffusion generation process.", + "bbox": [ + 496, + 672, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements have incorporated LLMs to control the generation of diffusion models, bypassing the need for additional complementary information as inputs [6, 7, 11-13, 27]. In these approaches, LLMs play a central role in directly interpreting user textual prompts and managing the initial layout configuration. Despite some progress, these models often operate in an open-loop fashion, producing images in one iteration that cannot guarantee the generated images align with user prompts.", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6328", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c6ef2e8b092fa1cf62a5be3dfea74fe4f09fda2ad776123a781fe294cdb27ee7.jpg", + "image_caption": [ + "Figure 2. Our proposed Self-correcting LLM-controlled Diffusion (SLD) enhances text-to-image alignment through an iterative self-correction process. It begins with LLM-driven object detection (Sec. 3.1), and subsequently performs LLM-controlled analysis and correction (Sec. 3.2). The entire pipeline is outlined in Algorithm 1." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 841, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike prior work, SLD is the first closed-loop diffusion-based generation method. Integrating advanced object detectors and LLMs, SLD performs iterative self-checking and correction, significantly enhancing text-to-image alignment. This improvement spans numeracy, attribute binding to multiple objects, and spatial reasoning, applicable to various models, including those like DALL-E 3. Also, SLD extends beyond existing NLP self-correction techniques [17] by incorporating bounding box spatial representation and novel latent editing mechanisms tailored for text-to-image tasks. These designs are essential for ensuring high-quality, precise self-correction in image generation tasks.", + "bbox": [ + 75, + 505, + 472, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion-based Image Editing", + "text_level": 1, + "bbox": [ + 76, + 696, + 349, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements in text-to-image diffusion models have significantly expanded their applications in image editing, encompassing both global and local editing tasks. Techniques like Prompt-2-prompt [8] and InstructPix2Pix [2] specialize in global edits, such as style transformations. Conversely, methods like SDEdit [15], DiffEdit [4], and Plug-and-Play [22] focus on local edits, targeting specific areas within images. Despite their progress, these methods often struggle with precise object-level manipulation and tasks that require spatial reasoning, such as resizing or repositioning objects. While recent approaches like Self-Guidance [5] offer fine-grained operations, they still necessities to be able to use the same techniques.", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sitate user inputs for specific coordinates when moving or repositioning objects.", + "bbox": [ + 498, + 505, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike these methods only focus on diffusion models, SLD introduces the combination of detectors and LLMs in the loop of editing, enabling fine-grained editing with only user prompts. Also, SLD excels in various object-level editing tasks, including adding, replacing, moving, and modifying attributes, swapping, and so on, demonstrating a notable improvement in both ease of use and editing capabilities.", + "bbox": [ + 496, + 536, + 892, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Self-correcting LLM-controlled Diffusion", + "text_level": 1, + "bbox": [ + 498, + 657, + 870, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our Self-correcting LLM-controlled Diffusion (SLD) framework. SLD consists of two main components: LLM-driven object detection (Sec. 3.1) as well as LLM-controlled assessment and correction (Sec. 3.2). Moreover, with a simple change of the LLM instructions, we show that SLD is applicable to image editing, unifying text-to-image generation, and editing as discussed in Sec. 3.3. The complete pipeline is shown in Algorithm 1.", + "bbox": [ + 496, + 683, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. LLM-driven Object Detection", + "text_level": 1, + "bbox": [ + 498, + 830, + 767, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our SLD framework starts with LLM-driven object detection, which extracts information necessary for downstream assessment and correction. As shown with green arrows in", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6329", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ef7b66afd1c8e32ee938f07acbf4cec29c40bfd4b0aa5f235184b10e95678a3.jpg", + "image_caption": [ + "Figure 3. Our self-correction pipeline is driven by two distinct LLMs: (a) The LLM parser analyzes user prompts $P$ to extract a list of key object information $S$ , which is then passed to the open-vocabulary detector. (b) The LLM controller, taking both the user prompt $P$ and currently detected bounding boxes $B_{curr}$ as input, outputs suggested new bounding boxes $B_{next}$ . These are subsequently transformed into a set of latent space operations $Ops$ for image manipulation." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 377, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19ecd4193069023ec26d3a6a51f7c80079e6560e6bbd68d5a24664a8d63fab59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 88, + 717, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4fdbe53657b21fc76c661e3adc40c354300523c8f2d83e0041b4876177d3c0b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 88, + 890, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2, the LLM-driven object detection includes two steps: 1) We leverage an LLM as a parser that parses the user prompt and outputs key phrases that are potentially relevant to image assessment. 2) The phrases are then passed into an open-vocabulary object detector. The detected boxes are supposed to contain information that supports the assessment of whether the image is aligned with the specifications in the user prompt.", + "bbox": [ + 75, + 385, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the initial step, an LLM parser is directed to extract a list of key object details, denoted as $S$ , from the user-provided text prompt $P$ . This parser, aided by text instructions and in-context examples, can easily accomplish this as shown in Fig. 3 (a). For a user prompt that includes phrases like \"a green motorcycle\" and \"a blue motorcycle,\" the LLM is expected to identify and output \"green\" and \"blue\" as attributes associated with \"motorcycle.\" When the prompt references objects without specific quantities or attributes, such as \"a monkey\" and \"a raccoon,\" these descriptors are appropriately left blank. Importantly, the LLM's role is not limited to merely identifying object nouns; it also entails identifying any associated quantities or attributes.", + "bbox": [ + 75, + 507, + 470, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second step, an open-vocabulary detector processes the list of key object information, $S$ , parsed in the first step, to detect and localize objects within the image. We prompt the open-vocabulary object detector with queries formatted as image of a/an [attribute] [object name], where the \"attribute\" and \"object name\" are sourced from the parser's output. The resulting bounding boxes, $B_{curr}$ , are then organized into a list format like [\"[attribute] [object name] [#object ID]\", [x, y, w, h])] for further processing. A special case is when the prompt poses constraints on the object quantity. For cases where attributed objects (e.g., \"blue dog\") fall short compared to the required", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "quantities, a supplementary count of non-attributed objects (e.g., \"dog\") is provided to provide context for the subsequent LLM controller deciding whether to add more \"blue dogs\" or simply alter the color of existing dogs to blue. We will explain these operations, including object addition and attribute modification, in greater detail in Sec. 3.2.1.", + "bbox": [ + 496, + 385, + 893, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. LLM-controlled Analysis and Correction", + "text_level": 1, + "bbox": [ + 498, + 494, + 852, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use an LLM controller for image analysis and the subsequent correction. The controller, given the user prompt $P$ and detected boxes $B_{curr}$ , is asked to analyze whether the image, represented by objects bounding boxes, aligns with the description of the user prompt and offers a list of corrected bounding boxes $B_{next}$ , as shown in Fig. 3 (b).", + "bbox": [ + 496, + 520, + 893, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SLD then programmatically analyzes the inconsistencies between the refined and original bounding boxes to output a set of editing operations $Ops$ , which includes addition, deletion, repositioning, and attribute modification. However, a simple set-of-boxes representation does not carry correspondence information, which does not allow an easy way to compare the input and the output layout of the LLM controller when multiple boxes share the same object name. For example, when there are two cat boxes in both the model input and the model output, whether one cat box corresponds to which cat box in the output layout is unclear. Rather than introducing another algorithm to guess the correspondence, we propose to let the LLM output correspondence with a very simple edit: we give an object ID to each bounding box, with the number increasing within each object type, as a suffix added after the object name. In the in-context examples, we demonstrate to the LLM that the object should have the same name and object ID before and after the proposed correction.", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6330", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Self-correction for Image Generation.", + "Input: User prompt $P$ , Initial generated image $I$ , Maximum number of self-correction round $K$ ." + ], + "code_body": "1: for $k\\gets 1$ to $K$ do \n2: $S\\gets \\mathsf{LLM - Parse}(P)$ \n3: $B_{curr}\\gets \\mathsf{Detector}(S)$ \n4: $B_{next}\\gets \\mathsf{LLM - Analysis}(P,B_{curr})$ \n5: $Ops\\gets \\mathsf{Diff}(B_{curr},B_{next})$ \n6: if $Ops\\neq \\emptyset$ then (i.e., $B_{next}\\neq B_{curr}$ ) \n7: $I = \\mathsf{Correction}(Ops,B_{next},B_{curr})$ \n8: else \n9: break \n10: end if \n11: end for \nOutput: Image $I$", + "bbox": [ + 86, + 141, + 421, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Latent Operations for Training-Free Image Correction", + "text_level": 1, + "bbox": [ + 76, + 349, + 467, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The LLM controller outputs a list of correction operations to apply. For each operation, we first transform the original image into latent features. Our approach then executes a series of operations $Ops$ , such as addition, deletion, repositioning, and attribute modification, applied to these latent layers. It is worth noting that although there are only four operations, they are sufficient to handle a majority of misalignments naturally. We explain how each operation is performed below.", + "bbox": [ + 75, + 388, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d8ea5419e73e8ec93f27cba35120f3e932f85bab62ba6e1918b4065f17a72568.jpg", + "image_caption": [ + "Figure 4. Our latent operations can be summarized into two key concepts: (1) latent in removed regions are re-initialized to Gaussian noise, and latent of newly added or modified objects are composed onto the canvas. (2) Latent composition is confined to the initial steps, followed by \"unfrozen\" steps for a standard forward diffusion process, enhancing visual quality and avoiding artificial copy-and-paste effects." + ], + "image_footnote": [], + "bbox": [ + 81, + 537, + 464, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Addition. Inspired by [11], the addition process entails two", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "phases: pre-generating an object and integrating its latent representation into the original image's latent space. Initially, we use a diffusion model to create an object within a designated bounding box, followed by precise segmentation using models (e.g., SAM [9]). This object is then processed through a backward diffusion sequence with our base diffusion model, yielding masked latent layers corresponding to the object, which are later merged with the original canvas.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deletion operation begins with SAM refining the boundary of the object within its bounding box. The latent layers associated with these specified regions are then removed and reset with Gaussian noise. This necessitates a complete regeneration of these areas in the final denoising process.", + "bbox": [ + 496, + 213, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Repositioning involves modifying the original image to align objects with new bounding boxes, taking care to preserve their original aspect ratios. The initial steps include shifting and resizing the bounding box in the image space. Following this, SAM refines the object boundary, succeeded by a backward diffusion process to generate its relevant latent layers, similar to the approach in the addition operation. Latent layers corresponding to the excised parts are replaced with Gaussian noise, while the newly added sections are integrated into the final image composition. An important consideration in repositioning is conducting object resizing in the image space rather than the latent space to maintain high-quality results.", + "bbox": [ + 496, + 289, + 892, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attribute modification starts with SAM refining the object boundary within the bounding box, followed by applying attribute modifications such as DiffEdit [4]. The base diffusion model then reverses the image, producing a series of masked latent layers ready for final composition.", + "bbox": [ + 496, + 487, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After editing operations on each object, we proceed to the recomposition phase as shown in Fig. 4. In this phase, while latents for removed or repositioned regions are reinitialized with Gaussian noise, the latents for added or modified latents are updated accordingly. For regions with multiple overlapping objects, we place the larger masks first to ensure the visibility of the smaller objects.", + "bbox": [ + 496, + 563, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The stack of modified latent then undergoes a final forward diffusion process, which begins with steps in which regions not reinitialized with Gaussian noise are frozen (i.e., forced to align with the unmodified latent at the same step). This is crucial for the accurate formation of updated objects while maintaining background consistency at the same time. The procedure finishes with several steps where everything is allowed to change, resulting in a visually coherent and correct image.", + "bbox": [ + 496, + 670, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Termination of the Self-Correction Process", + "text_level": 1, + "bbox": [ + 500, + 829, + 851, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While we observe that one round of generation is often enough for a majority of the cases, subsequent rounds could still benefit the performance in terms of correctness further,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6331", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f1dff01ee196d8237d17bf39ae8eae6d6103e1476fdf9c490a71a1f8b24fb58a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
MultiDiffusion [1]29%28%26%39%30.5%
Backward Guidance [3]22%37%26%67%38.0%
BoxDiff [23]22%30%37%71%40.0%
LayoutGPT + GLIGEN [6, 10]36%65%26%78%51.3%
DALL-E 3 [16]25%38%74%71%52.0%
+ 1-round SLD (Ours)90%61%80%83%78.5% (+26.5)
LMD+ [11]100%82%49%86%79.3%
+ 1-round SLD (Ours)100%98%63%92%88.3% (+9.0)
", + "bbox": [ + 163, + 88, + 803, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Our method can be applied to various image generation methods and improves the generation accuracy by a large margin.", + "bbox": [ + 99, + 285, + 867, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "making our self-correction an iterative process. Thus, determining the optimal number of self-correction rounds is critical for balancing efficiency and accuracy. As outlined in Algorithm 1, our method sets a maximum number of attempts on the correction rounds to ensure the process finishes within a reasonable amount of time.", + "bbox": [ + 75, + 318, + 467, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The process completes when the LLM outputs the same layout as the input (i.e., if the bounding boxes suggested by the LLM controller $(B_{next})$ align with the current detected bounding boxes $(B_{curr}))$ , or when the maximum rounds of generation are reached, indicating that the method is unable to make a correct generation for the prompt. This iterative process provides guarantees on the correctness of the image, up to the accuracy of the detector and the LLM controller, ensuring it aligns closely with the initial text prompt. We explore the efficacy of multi-round corrections in Sec. 4.3.", + "bbox": [ + 75, + 410, + 467, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Unified text-to-image generation and editing", + "text_level": 1, + "bbox": [ + 76, + 571, + 452, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to self-correcting image generation models, our SLD framework is readily adaptable for image editing applications, requiring only minimal modifications. A key distinction is in the format of user input prompts. Unlike image generation, where users provide scene descriptions, image editing requires users to detail both the original image and the desired changes. For instance, to edit an image with two apples and a banana by replacing the banana with an orange, the prompt could be: \"Replace the banana with an orange, while keeping the two apples unchanged.\"", + "bbox": [ + 75, + 597, + 467, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The editing process is similar to our self-correction mechanism. The LLM parser extracts key objects from the user's prompt. These objects are then identified by the open-vocabulary detector, establishing a list of current bounding boxes. The editing-focused LLM controller, equipped with specific task objectives, guidelines, and in-context examples, analyzes these inputs. It proposes updated bounding boxes and corresponding latent space operations for precise image manipulation.", + "bbox": [ + 75, + 748, + 467, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SLD's ability to perform detailed, object-level editing", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "distinguishes it from existing diffusion-based methods like InstructPix2Pix [2] and prompt2prompt [8], which mainly address global image style changes. Also, SLD outperforms tools like DiffEdit [4] and SDEdit [15], which are restricted to object replacement or attribute adjustment, by enabling comprehensive object repositioning, addition, and deletion with exact control. Our comparative analysis in Sec. 4.2 will further highlight SLD's superior editing capabilities over existing methods.", + "bbox": [ + 496, + 318, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 470, + 632, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with Image Generation Methods", + "text_level": 1, + "bbox": [ + 500, + 497, + 883, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. We evaluate the performance of the SLD framework with the LMD benchmark [11], which is specifically designed to evaluate generation methods on complex tasks such as handling negation, numeracy, accurate attribute binding to multiple objects, and spatial reasoning. For each task, 100 programmatically generated prompts are fed into various text-to-image generation methods to produce corresponding images. We evaluate the images generated by our method and the baselines with open-vocabulary detector OWL-ViT v2 [14] for a robust quantitative evaluation of the alignment between the input prompts and the generated images. We compared SLD with several leading text-to-image diffusion methods, such as Multidiffusion [1], BoxDiff [23], LayoutGPT [6], LMD+ [12], and DALL-E 3 [16]. To ensure fair comparisons, all models incorporating LLMs used the same GPT-4 model. In our SLD implementation, we utilized LMD+ as the base model for latent space operations and OWL-ViT v2 for the open-vocabulary object detector.", + "bbox": [ + 496, + 521, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 1, applying the SLD method to both open-source $(\\mathrm{LMD}+)$ and proprietary models (DALL-E 3) significantly enhances their performance in terms of generation correctness. For negation tasks, as $\\mathrm{LMD}+$ converts user prompts containing \"without\" information into negative prompts, which already achieves a remarkable $100\\%$ accuracy without SLD integration. In contrast, even", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6332", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a307dbf42a70969e49a48b6001524db16204d2c41dec9c66fbb9ad461604a612.jpg", + "image_caption": [ + "SDXL" + ], + "image_footnote": [], + "bbox": [ + 80, + 116, + 209, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00fc34a4c7af14058bf1a7722fd49a17c4db0b582deb310bcd4ce9fb88899df8.jpg", + "image_caption": [ + "SDXL w/ SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 215, + 116, + 344, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bb12932fe4e277153d47699dd97a98608cf6416fadd3164c73238cea0f8f6d9.jpg", + "image_caption": [ + "LMD+" + ], + "image_footnote": [], + "bbox": [ + 351, + 116, + 480, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49f6ff7fa7e479b56d53bc2e478e1c0fc720e4a34b0afa88a9dae0c0d1e60014.jpg", + "image_caption": [ + "LMD+ w/ SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 488, + 116, + 619, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ca07c4e9fbae07c9d83e64a0d2f1296b633b59fa8b67998138a99d2e7f5d3e6.jpg", + "image_caption": [ + "DALL-E3" + ], + "image_footnote": [], + "bbox": [ + 625, + 116, + 754, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/abfd065bb7574402eba14755d9d2dc6ae24e047486aabbc63a219ad46b10092c.jpg", + "image_caption": [ + "DALL-E 3 w/ SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 761, + 116, + 890, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A realistic photo with a monkey sitting above a green motorcycle on the left and another raccoon sitting above a blue motorcycle on the right", + "bbox": [ + 80, + 239, + 888, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc008f569b3f5b259f3f45264c6c20c4623ab6f0f55e0b5472bd50007d10e76a.jpg", + "image_caption": [ + "SDXL" + ], + "image_footnote": [], + "bbox": [ + 80, + 256, + 209, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/848e9760730f9b5c157b58e581e601b2d71e9aa89248623e39338c51c9f4a1aa.jpg", + "image_caption": [ + "SDXL w/ SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 215, + 256, + 344, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd39125a3f2b1c74e7a4d54713b29da87759770cd3bcb522e197a217fa1d8935.jpg", + "image_caption": [ + "LMD+" + ], + "image_footnote": [], + "bbox": [ + 352, + 256, + 480, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de80f2785db0c6313e30f13160fa251ea7158a19e0517a14096f9a2f073c8084.jpg", + "image_caption": [ + "LMD+ w/ SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 488, + 256, + 617, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9bf35ba779a91127c0b862ddfc7217808eae8325c3ea3226d126e65b4836ae48.jpg", + "image_caption": [ + "DALL-E3" + ], + "image_footnote": [], + "bbox": [ + 625, + 256, + 754, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59ae716dc097a9b7eceff7bca8ae16282faf8daa5fb4fc027c91f7b91d48c63a.jpg", + "image_caption": [ + "DALL-E 3 w/ SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 761, + 256, + 890, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "though DALL-E 3 also uses an LLM to rewrite the prompt, it still fails for some negation cases, likely because the LLM simply puts the negation keyword (e.g., \"without\") into the rewritten prompt. In this case, our SLD method can automatically rectify most of these errors. For numeracy tasks, integrating SLD with $\\mathrm{LMD + }$ results in a significant improvement, with up to $98\\%$ accuracy. We noted that DALL-E 3 often struggles to generate an image with the correct number of objects. However, this issue is substantially mitigated by SLD, which enhances performance by over $20\\%$ . For attribute binding tasks, SLD improves the performance of both DALL-E 3 and $\\mathrm{LMD + }$ by $6\\%$ and $14\\%$ , respectively. Notably, DALL-E 3 initially outperforms $\\mathrm{LMD + }$ in this task, likely due to its training on high-quality image caption datasets. Finally, for spatial reasoning tasks, the integration of SLD with both $\\mathrm{LMD + }$ and DALL-E 3 demonstrates enhanced performance by $12\\%$ and $6\\%$ , respectively.", + "bbox": [ + 75, + 460, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Application to Image Editing", + "text_level": 1, + "bbox": [ + 76, + 741, + 334, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.3, SLD excels in fine-grained image editing over existing methods. As demonstrated in Fig. 6, our integration of an open-language detector with LLMs enables precise modifications within localized latent space regions. SLD adeptly performs specific edits, like seamlessly replacing an apple with a pumpkin, while preserving the integrity of surrounding objects. In contrast, methods like InstructPix2Pix [2] are confined to global transformations, and DiffEdit [4] often fails to accurately locate objects for", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e132a16ef20d871051607858e51ed72b624f0a8495d5598edddc3d7172a56016.jpg", + "image_caption": [ + "Figure 5. SLD enhances text-to-image alignment across diverse diffusion-based generative models such as SDXL, LMD+, and DALL-E 3. Notably, as highlighted by the red boxes in the first row, SLD precisely positions a blue bicycle between a bench and a palm tree, while maintaining the accurate count of palm trees and seagulls. The second row further demonstrates SLD's robustness in complex, cluttered scenes, effectively managing object collision through our training-free latent operations." + ], + "image_footnote": [], + "bbox": [ + 756, + 449, + 870, + 535 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eac1e7e74054f9eb76ff33b109371fda001ae8dd74ef83d370f49350c3af9c94.jpg", + "image_caption": [ + "Replace the left apple with a pumpkin and keep the right apple and all bananas in the scene unchanged." + ], + "image_footnote": [], + "bbox": [ + 519, + 513, + 544, + 532 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6af6512aaefc4900907eca27f063f9bb02b2808b2c67c1030e9fb9efe93c4a0b.jpg", + "image_caption": [ + "Editing Instructions for SLD", + "InstructPix2Pix", + "Figure 6. When instructed to perform object-level image editing. InstructPix2Pix [2] completely fails to accomplish the task, and DiffEdit [4] falls short, as highlighted in the green box of the image. Conversely, our method demonstrates a significantly better performance in executing these object-level edits." + ], + "image_footnote": [], + "bbox": [ + 519, + 541, + 630, + 627 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba6a05105eb50bd4d5449e5beca659645801228aa972b74b87e24b685ffa18c5.jpg", + "image_caption": [ + "DiffEdit" + ], + "image_footnote": [], + "bbox": [ + 638, + 541, + 750, + 627 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcb7998c6696b4de99e8a1726bb07bfc50058240280ea802408ff1d6ddd1c9ad.jpg", + "image_caption": [ + "SLD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 756, + 541, + 869, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "modification, leading to undesired results.", + "bbox": [ + 500, + 734, + 776, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, as exemplified in Fig. 7, SLD supports a wide array of editing instructions, including counting control (such as adding, deleting, or replacing objects), attribute modification (like altering colors or materials), and intricate location control (encompassing object swapping, resizing, and moving). A standout example is featured in the \"Object Resize\" column of the first row, where SLD precisely enlarges the cup on the table by an exact factor of $1.25 \\times$ . We encourage readers to verify this with a ruler for a clear demonstration of our method's precision. This level of pre", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "An oil painting at the beach of a blue bicycle to the left of a bench and to the right of a palm tree with three seagulls in the sky", + "bbox": [ + 122, + 89, + 852, + 109 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6333", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/325b8eecc53f1075be21ccb45648ead52c39545d53058ec27de89160a00f7620.jpg", + "image_caption": [ + "Replace the bowl with two round apples and keep the cup unchanged" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 209, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cec1937d48bdd41ee6ad4a7a21e590d32cf2d06ad881e976dfa18af88f7f7a74.jpg", + "image_caption": [ + "Turn the cup into green and the bowl into golden" + ], + "image_footnote": [], + "bbox": [ + 215, + 90, + 344, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a54fbf66c7e2da6bbd5cb856e1b457297dbaab09b21650eaa3c8ce62f8b0de9.jpg", + "image_caption": [ + "Swap the position of the bowl and the cup" + ], + "image_footnote": [], + "bbox": [ + 352, + 90, + 480, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3f3fefc5dcd868bd91cee3be94e35b7eba1d1290f92a98df0f9ba781d361a90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 90, + 617, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/405001983c82bc96a9d7adce58e4359fe71dd446cb79c7eaef1780bd935bad7a.jpg", + "image_caption": [ + "Make cup 1.25 times its larger but keep the bowl's size" + ], + "image_footnote": [], + "bbox": [ + 624, + 90, + 754, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d45f0f869a2f801327f228e3c82f9a561e13d812f99d7eb23c8092b7d92c763.jpg", + "image_caption": [ + "Move the cup towards the bowl until they slightly overlap" + ], + "image_footnote": [], + "bbox": [ + 761, + 90, + 890, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/600d91a9dcd434d0c6a3a2844d7d41ab27582de36bc1432d15cb9ac508ba02d1.jpg", + "image_caption": [ + "User Instructions", + "User Instructions" + ], + "image_footnote": [], + "bbox": [ + 80, + 247, + 209, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03f715b515708b5e9eb1b474dac5a58de7f9cc050b10a78521d40c07be8187aa.jpg", + "image_caption": [ + "Replace the dog with a houseplant and keep the cat unchanged" + ], + "image_footnote": [], + "bbox": [ + 215, + 247, + 344, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0b5b08eaae92c485f84cb968e8e138c14c5e730817fa18b8b31936e85e5400f.jpg", + "image_caption": [ + "Transform the dog into a robotic dog and the cat into a sculptural cat" + ], + "image_footnote": [], + "bbox": [ + 352, + 247, + 480, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7cebd7be58e3ddce0898d7473c90d28aa1ee15d8557e3bc24b2dc1f4c4fa393f.jpg", + "image_caption": [ + "Swap the position of the black dog and the yellow cat" + ], + "image_footnote": [], + "bbox": [ + 488, + 247, + 617, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1df495a9187b4cb0d960b07880a18be359c114da945d3719857c473d24290162.jpg", + "image_caption": [ + "Increase the cat's size and reduce the dog's size" + ], + "image_footnote": [], + "bbox": [ + 624, + 247, + 754, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce4dec8db7f3f3b88574e4c426f5d9c56dc8dbdf29b61e725592283454c303a9.jpg", + "image_caption": [ + "Shift the dog slightly upward and move the cat slightly towards the dog" + ], + "image_footnote": [], + "bbox": [ + 761, + 247, + 890, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/537f67b74f4d53911293856a266d53587254fb5c76c38b2ff1901a04d2182a48.jpg", + "image_caption": [ + "Figure 7. SLD can handle a diverse array of image editing tasks guided by natural, human-like instructions. Its capabilities span from adjusting object counts to altering object attributes, positions, and sizes.", + "Remove hair of the rightmost person sitting at a cafe in this image", + "Original Image", + "Figure 8. SLD struggles with objects of complex shapes, as the SAM module may unintentionally segment adjacent parts during the process." + ], + "image_footnote": [], + "bbox": [ + 80, + 468, + 200, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/925c6863876afdcf3efe077f0345a18e343b4309fb0bc08fb8315ea1fb5dd0d2.jpg", + "image_caption": [ + "Removed Region" + ], + "image_footnote": [], + "bbox": [ + 212, + 468, + 334, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9a18118c8952e83e3bd93cfecd77665a6234d23a75eb1bbb6a3894191ff52dd.jpg", + "image_caption": [ + "Failed Result" + ], + "image_footnote": [], + "bbox": [ + 344, + 468, + 467, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cision stems from the detector's exact object localization coupled with the LLMs' ability in reasoning and suggestions for new placements. Such detailed control over spatial adjustments is unmatched by any previous method, highlighting SLD's contributions to fine-grained image editing.", + "bbox": [ + 75, + 642, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Discussion", + "text_level": 1, + "bbox": [ + 76, + 726, + 194, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-round self-correction. Our analysis in Tab. 2 highlights the benefits of multi-round self-correction and the fact that the first round correction is always the most effective one and has a marginal effect. The first round of corrections substantially mitigates issues inherent in Stable Diffusion [20]. Then, a second round of correction still yields significant improvements across all four tasks.", + "bbox": [ + 75, + 750, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. A limitation of our method is illustrated in Fig. 8, where SLD fails to accurately remove a person's hair. In this instance, despite the successful iden-", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4b826b6cddd9a1692b96f074149a0723df9095c8425cf6eb8ff2d594bc0951d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
SD [20]19%38%24%33%28.5%
+ 1-round SLD69%55%25%69%54.5% (+26.0)
+ 2-round SLD73%61%31%75%60.0% (+31.5)
", + "bbox": [ + 503, + 452, + 888, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. While the majority of errors are typically rectified in the first round, multi-round correction consistently outperforms a single-round approach.", + "bbox": [ + 498, + 546, + 892, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tification and localization of the hair, the complex nature of its shape poses a challenge to the SAM module used for region selection, resulting in the unintended removal of the person's face in addition to the hair. However, since the person's cloth is not removed, the base diffusion model fails to generate a natural composition. This suggests that a better region selection method is needed for further improvements in the generation and editing quality.", + "bbox": [ + 496, + 609, + 890, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 750, + 617, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce the Self-correcting Language-Driven (SLD) framework, a pioneering self-correction system using detectors and LLMs to significantly enhance text-to-image alignment. This method not only sets a new SOTA in the image generation benchmark but is also compatible with various generative models, including DALL-E 3. Also, SLD extends its utility to image editing applications, offering fine-grained object-level manipulation that surpasses existing methods.", + "bbox": [ + 496, + 773, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6334", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 6", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 6, 7", + "[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373, 2023. 2, 6", + "[4] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2023. 3, 5, 6, 7", + "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 3", + "[6] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 1, 2, 6", + "[7] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LLM blueprint: Enabling text-to-image generation with complex and detailed prompts. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 6", + "[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5", + "[10] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 6", + "[11] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 1, 2, 5, 6", + "[12] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 6", + "[13] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2", + "[14] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3, 6", + "[16] OpenAI. Dall-e 3 system card, 2023. 1, 6", + "[17] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 3", + "[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[19] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 8", + "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[22] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3", + "[23] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 6", + "[24] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2", + "[25] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023.", + "[26] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6335", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[27] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 2", + "bbox": [ + 78, + 90, + 470, + 135 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6336", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_model.json b/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e139117cba924a8b6e836454a4ac59ed13683b90 --- /dev/null +++ b/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.131, + 0.735, + 0.152 + ], + "angle": 0, + "content": "Self-correcting LLM-controlled Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.181, + 0.262, + 0.2 + ], + "angle": 0, + "content": "Tsung-Han Wu*" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.182, + 0.388, + 0.199 + ], + "angle": 0, + "content": "Long Lian*" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.182, + 0.58, + 0.218 + ], + "angle": 0, + "content": "Joseph E. Gonzalez \nUC Berkeley" + }, + { + "type": "text", + "bbox": [ + 0.62, + 0.181, + 0.69, + 0.199 + ], + "angle": 0, + "content": "Boyi Li†" + }, + { + "type": "text", + "bbox": [ + 0.724, + 0.182, + 0.845, + 0.198 + ], + "angle": 0, + "content": "Trevor Darrell†" + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.228, + 0.347, + 0.242 + ], + "angle": 0, + "content": "Numeracy Correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.246, + 0.205, + 0.257 + ], + "angle": 0, + "content": "DALL-E3" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.258, + 0.268, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.246, + 0.454, + 0.258 + ], + "angle": 0, + "content": "DALL-E 3 + SLD (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.258, + 0.476, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.406, + 0.458, + 0.419 + ], + "angle": 0, + "content": "A realistic cartoon-style image with a princess and four dwarfs" + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.228, + 0.753, + 0.242 + ], + "angle": 0, + "content": "Spatial Correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.246, + 0.617, + 0.257 + ], + "angle": 0, + "content": "DALL-E3" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.258, + 0.681, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.246, + 0.867, + 0.258 + ], + "angle": 0, + "content": "DALL-E 3 + SLD (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.258, + 0.888, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.407, + 0.88, + 0.433 + ], + "angle": 0, + "content": "A vivid photo with a woman on the right and a clown on the left walking in a dirty alley" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.447, + 0.894, + 0.503 + ], + "angle": 0, + "content": "Figure 1. Existing diffusion-based text-to-image generators (e.g., DALL-E 3 [16]) generally struggle to precisely generate images that correctly align with complex input prompts, especially for the ones that require numeracy and spatial relationships. Our Self-correcting LLM-controlled Diffusion (SLD) framework empowers these diffusion models to automatically and iteratively rectify inaccuracies through applying a set of latent space operations (addition, deletion, repositioning, etc.), resulting in enhanced text-to-image alignment." + }, + { + "type": "title", + "bbox": [ + 0.269, + 0.515, + 0.347, + 0.53 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.473, + 0.879 + ], + "angle": 0, + "content": "Text-to-image generation has witnessed significant progress with the advent of diffusion models. Despite the ability to generate photorealistic images, current text-to-image diffusion models still often struggle to accurately interpret and follow complex input text prompts. In contrast to existing models that aim to generate images only with their best effort, we introduce Self-correcting LLM-controlled Diffusion (SLD). SLD is a framework that generates an image from the input prompt, assesses its alignment with the prompt, and performs self-corrections on the inaccuracies in the generated image. Steered by an LLM controller, SLD turns text-to-image generation into an iterative closed-loop process, ensuring correctness in the resulting image. SLD is not only training-free but can also be seamlessly integrated with diffusion models behind API access, such as DALL-E 3, to further boost the performance of state-of-the-art diffusion models. Experimental results show that our approach can rectify a majority of incorrect generations, particularly in generative numeracy, attribute binding, and spatial relationships. Furthermore, by simply adjusting the instructions to the LLM, SLD can perform image editing tasks, bridging the gap between text-to-image generation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.517, + 0.892, + 0.547 + ], + "angle": 0, + "content": "and image editing pipelines. Our code is available at: https://self-correcting-llm-diffusion.github.io." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.555, + 0.633, + 0.57 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Text-to-image generation has made remarkable advancements, especially with the advent of diffusion models. However, these models often struggle with interpreting complex input text prompts, particularly those that require skills such as understanding the concept of numeracy, spatial relationships, and attribute binding with multiple objects. Despite the astonishing scaling of model sizes and training data, these challenges, as illustrated in Fig. 1, are still present in state-of-the-art open-source and proprietary diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Several research and engineering efforts aim to overcome these limitations. For instance, methods such as DALL-E 3 [16] focus on the diffusion training process and incorporate high-quality captions into the training data at a massive scale. However, this approach not only incurs substantial costs but also frequently falls short in generating accurate images from complicated user prompts, as shown in Fig. 1. Other work harnesses the power of external models for a better understanding of the prompt in the inference process before the actual image generation. For example, [6, 11] leverages Large Language Models (LLMs) to pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.084, + 0.887, + 0.308, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Equal advising." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6327" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "process textual prompts into structured image layouts and thus ensures the preliminary design aligns with the user's directives. However, such integration does not resolve the inaccuracies produced by the downstream diffusion models, particularly in images with complex scenarios like multiple objects, cluttered arrangements, or detailed attributes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Drawing inspiration from the process of a human painting and a diffusion model in generating images, we observe a key distinction in their approach to creation. Consider a human artist tasked with painting a scene featuring two cats. Throughout the painting process, the artist remains cognizant of this requirement, ensuring that two cats are indeed present before considering the work complete. Should the artist find only one cat depicted, an additional one would be added to meet the prompt's criteria. This contrasts sharply with current text-to-image diffusion models, which operate on an open-loop basis. These models generate images through a predetermined number of diffusion steps and present the output to the user, regardless of its alignment with the initial user prompt. Such a process, irrespective of scaling training data or LLM pre-generation conditioning, lacks a robust mechanism to ensure the final image aligns with the user's expectations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.594 + ], + "angle": 0, + "content": "In light of this, we propose our method Self-correcting LLM-controlled Diffusion (SLD) that performs self-checks to confidently offer users guarantees of the alignment between the prompt and the generated images. Departing from conventional single-round generation methods, SLD is a novel closed-loop approach that equips diffusion models with the ability to iteratively identify and rectify errors. Our SLD framework, illustrated in Fig. 2, contains two main components: LLM-driven object detection as well as LLM-controlled assessment and correction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.884 + ], + "angle": 0, + "content": "The SLD pipeline follows a standard text-to-image generation setting. Given a textual prompt that outlines the desired image, SLD begins with calling an image generation module (e.g., the aforementioned open-loop text-to-image diffusion models) to generate an image in a best-effort fashion. Given that these open-loop generators do not guarantee an output that aligns perfectly with the prompt, SLD then conducts a thorough evaluation of the produced image against the prompt, with an LLM parsing key phrases for an open-vocabulary detector to check. Subsequently, an LLM controller takes the detected bounding boxes and the initial prompt as input and checks for potential mismatches between the detection results and the prompt requirements, suggesting appropriate self-correction operations, such as adding, moving, and removing objects. Finally, utilizing a chosen base diffusion model (e.g., Stable Diffusion [20]), SLD employs latent space composition to implement these adjustments, thereby ensuring that the final image accurately reflects the user's initial text prompt." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Notably, SLD is agnostic to the initial generation pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "cess, which allows us to use DALL-E 3 APIs as a subroutine to generate initial images and then employ the open-source Stable Diffusion model family to conduct latent operations to attain self-correction. Furthermore, none of these operations require additional training on our base diffusion model [20], which easily allows our method to be applied to various diffusion models without external human annotation or training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.348 + ], + "angle": 0, + "content": "We demonstrate that our SLD framework can achieve significant improvement over current diffusion-based methods on complex prompts with the LMD benchmark [11]. The results show that our method can surpass \\(\\mathrm{LMD + }\\), which is a strong baseline that already leverages LLM in the image process generation, by \\(9.0\\%\\). More importantly, with DALL-E 3 for initial generation, the generated images from our method achieve \\(26.5\\%\\) performance gains compared to ones before self-correction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Finally, since the SLD pipeline is agnostic to the initially generated image, it can easily be transformed into an image editing pipeline by simply changing the prompts to the LLM. While text-to-image generation and image editing are often treated as distinct tasks by the generative modeling community, our SLD can perform these two tasks with a unified pipeline. We list our key contributions below:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.457, + 0.892, + 0.502 + ], + "angle": 0, + "content": "1. SLD is the first to integrate a detector and an LLM to self-correct generative models, ensuring accurate generation without extra training or external data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.502, + 0.892, + 0.563 + ], + "angle": 0, + "content": "2. SLD offers a unified solution for both image generation and editing, enabling enhanced text-to-image alignment for any image generator (e.g., DALL-E 3) and object-level editing on any images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.563, + 0.892, + 0.609 + ], + "angle": 0, + "content": "3. Experimental results show that SLD can correct a majority of incorrect generations, particularly in aspects of numeracy, attribute binding, and spatial relationships." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.457, + 0.892, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.641, + 0.641 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.78, + 0.667 + ], + "angle": 0, + "content": "2.1. Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Diffusion-based text-to-image generation has advanced significantly. Initial studies [19-21] showed diffusion models' ability to create high-quality images, but they struggle with complex prompts. Subsequent research [3, 10, 18, 24-26] has incorporated additional inputs such as keypoints and bounding boxes to control the diffusion generation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recent advancements have incorporated LLMs to control the generation of diffusion models, bypassing the need for additional complementary information as inputs [6, 7, 11-13, 27]. In these approaches, LLMs play a central role in directly interpreting user textual prompts and managing the initial layout configuration. Despite some progress, these models often operate in an open-loop fashion, producing images in one iteration that cannot guarantee the generated images align with user prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6328" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.086, + 0.843, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.447, + 0.895, + 0.492 + ], + "angle": 0, + "content": "Figure 2. Our proposed Self-correcting LLM-controlled Diffusion (SLD) enhances text-to-image alignment through an iterative self-correction process. It begins with LLM-driven object detection (Sec. 3.1), and subsequently performs LLM-controlled analysis and correction (Sec. 3.2). The entire pipeline is outlined in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.689 + ], + "angle": 0, + "content": "Unlike prior work, SLD is the first closed-loop diffusion-based generation method. Integrating advanced object detectors and LLMs, SLD performs iterative self-checking and correction, significantly enhancing text-to-image alignment. This improvement spans numeracy, attribute binding to multiple objects, and spatial reasoning, applicable to various models, including those like DALL-E 3. Also, SLD extends beyond existing NLP self-correction techniques [17] by incorporating bounding box spatial representation and novel latent editing mechanisms tailored for text-to-image tasks. These designs are essential for ensuring high-quality, precise self-correction in image generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.35, + 0.714 + ], + "angle": 0, + "content": "2.2. Diffusion-based Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Recent advancements in text-to-image diffusion models have significantly expanded their applications in image editing, encompassing both global and local editing tasks. Techniques like Prompt-2-prompt [8] and InstructPix2Pix [2] specialize in global edits, such as style transformations. Conversely, methods like SDEdit [15], DiffEdit [4], and Plug-and-Play [22] focus on local edits, targeting specific areas within images. Despite their progress, these methods often struggle with precise object-level manipulation and tasks that require spatial reasoning, such as resizing or repositioning objects. While recent approaches like Self-Guidance [5] offer fine-grained operations, they still necessities to be able to use the same techniques." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.506, + 0.892, + 0.537 + ], + "angle": 0, + "content": "sitate user inputs for specific coordinates when moving or repositioning objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.893, + 0.645 + ], + "angle": 0, + "content": "Unlike these methods only focus on diffusion models, SLD introduces the combination of detectors and LLMs in the loop of editing, enabling fine-grained editing with only user prompts. Also, SLD excels in various object-level editing tasks, including adding, replacing, moving, and modifying attributes, swapping, and so on, demonstrating a notable improvement in both ease of use and editing capabilities." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.658, + 0.871, + 0.675 + ], + "angle": 0, + "content": "3. Self-correcting LLM-controlled Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.821 + ], + "angle": 0, + "content": "In this section, we introduce our Self-correcting LLM-controlled Diffusion (SLD) framework. SLD consists of two main components: LLM-driven object detection (Sec. 3.1) as well as LLM-controlled assessment and correction (Sec. 3.2). Moreover, with a simple change of the LLM instructions, we show that SLD is applicable to image editing, unifying text-to-image generation, and editing as discussed in Sec. 3.3. The complete pipeline is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.832, + 0.768, + 0.848 + ], + "angle": 0, + "content": "3.1. LLM-driven Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Our SLD framework starts with LLM-driven object detection, which extracts information necessary for downstream assessment and correction. As shown with green arrows in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6329" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.378, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.089, + 0.718, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.089, + 0.892, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Our self-correction pipeline is driven by two distinct LLMs: (a) The LLM parser analyzes user prompts \\( P \\) to extract a list of key object information \\( S \\), which is then passed to the open-vocabulary detector. (b) The LLM controller, taking both the user prompt \\( P \\) and currently detected bounding boxes \\( B_{curr} \\) as input, outputs suggested new bounding boxes \\( B_{next} \\). These are subsequently transformed into a set of latent space operations \\( Ops \\) for image manipulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.508 + ], + "angle": 0, + "content": "Fig. 2, the LLM-driven object detection includes two steps: 1) We leverage an LLM as a parser that parses the user prompt and outputs key phrases that are potentially relevant to image assessment. 2) The phrases are then passed into an open-vocabulary object detector. The detected boxes are supposed to contain information that supports the assessment of whether the image is aligned with the specifications in the user prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.704 + ], + "angle": 0, + "content": "In the initial step, an LLM parser is directed to extract a list of key object details, denoted as \\( S \\), from the user-provided text prompt \\( P \\). This parser, aided by text instructions and in-context examples, can easily accomplish this as shown in Fig. 3 (a). For a user prompt that includes phrases like \"a green motorcycle\" and \"a blue motorcycle,\" the LLM is expected to identify and output \"green\" and \"blue\" as attributes associated with \"motorcycle.\" When the prompt references objects without specific quantities or attributes, such as \"a monkey\" and \"a raccoon,\" these descriptors are appropriately left blank. Importantly, the LLM's role is not limited to merely identifying object nouns; it also entails identifying any associated quantities or attributes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In the second step, an open-vocabulary detector processes the list of key object information, \\( S \\), parsed in the first step, to detect and localize objects within the image. We prompt the open-vocabulary object detector with queries formatted as image of a/an [attribute] [object name], where the \"attribute\" and \"object name\" are sourced from the parser's output. The resulting bounding boxes, \\( B_{curr} \\), are then organized into a list format like [\"[attribute] [object name] [#object ID]\", [x, y, w, h])] for further processing. A special case is when the prompt poses constraints on the object quantity. For cases where attributed objects (e.g., \"blue dog\") fall short compared to the required" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.895, + 0.477 + ], + "angle": 0, + "content": "quantities, a supplementary count of non-attributed objects (e.g., \"dog\") is provided to provide context for the subsequent LLM controller deciding whether to add more \"blue dogs\" or simply alter the color of existing dogs to blue. We will explain these operations, including object addition and attribute modification, in greater detail in Sec. 3.2.1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.495, + 0.854, + 0.511 + ], + "angle": 0, + "content": "3.2. LLM-controlled Analysis and Correction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.894, + 0.612 + ], + "angle": 0, + "content": "We use an LLM controller for image analysis and the subsequent correction. The controller, given the user prompt \\( P \\) and detected boxes \\( B_{curr} \\), is asked to analyze whether the image, represented by objects bounding boxes, aligns with the description of the user prompt and offers a list of corrected bounding boxes \\( B_{next} \\), as shown in Fig. 3 (b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "SLD then programmatically analyzes the inconsistencies between the refined and original bounding boxes to output a set of editing operations \\( Ops \\), which includes addition, deletion, repositioning, and attribute modification. However, a simple set-of-boxes representation does not carry correspondence information, which does not allow an easy way to compare the input and the output layout of the LLM controller when multiple boxes share the same object name. For example, when there are two cat boxes in both the model input and the model output, whether one cat box corresponds to which cat box in the output layout is unclear. Rather than introducing another algorithm to guess the correspondence, we propose to let the LLM output correspondence with a very simple edit: we give an object ID to each bounding box, with the number increasing within each object type, as a suffix added after the object name. In the in-context examples, we demonstrate to the LLM that the object should have the same name and object ID before and after the proposed correction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6330" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.081, + 0.092, + 0.416, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Self-correction for Image Generation." + }, + { + "type": "code_caption", + "bbox": [ + 0.081, + 0.111, + 0.468, + 0.14 + ], + "angle": 0, + "content": "Input: User prompt \\(P\\), Initial generated image \\(I\\), Maximum number of self-correction round \\(K\\)." + }, + { + "type": "algorithm", + "bbox": [ + 0.088, + 0.142, + 0.422, + 0.322 + ], + "angle": 0, + "content": "1: for \\(k\\gets 1\\) to \\(K\\) do \n2: \\(S\\gets \\mathsf{LLM - Parse}(P)\\) \n3: \\(B_{curr}\\gets \\mathsf{Detector}(S)\\) \n4: \\(B_{next}\\gets \\mathsf{LLM - Analysis}(P,B_{curr})\\) \n5: \\(Ops\\gets \\mathsf{Diff}(B_{curr},B_{next})\\) \n6: if \\(Ops\\neq \\emptyset\\) then (i.e., \\(B_{next}\\neq B_{curr}\\)) \n7: \\(I = \\mathsf{Correction}(Ops,B_{next},B_{curr})\\) \n8: else \n9: break \n10: end if \n11: end for \nOutput: Image \\(I\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.351, + 0.468, + 0.38 + ], + "angle": 0, + "content": "3.2.1 Latent Operations for Training-Free Image Correction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.469, + 0.525 + ], + "angle": 0, + "content": "The LLM controller outputs a list of correction operations to apply. For each operation, we first transform the original image into latent features. Our approach then executes a series of operations \\( Ops \\), such as addition, deletion, repositioning, and attribute modification, applied to these latent layers. It is worth noting that although there are only four operations, they are sufficient to handle a majority of misalignments naturally. We explain how each operation is performed below." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.538, + 0.465, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.778, + 0.469, + 0.876 + ], + "angle": 0, + "content": "Figure 4. Our latent operations can be summarized into two key concepts: (1) latent in removed regions are re-initialized to Gaussian noise, and latent of newly added or modified objects are composed onto the canvas. (2) Latent composition is confined to the initial steps, followed by \"unfrozen\" steps for a standard forward diffusion process, enhancing visual quality and avoiding artificial copy-and-paste effects." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Addition. Inspired by [11], the addition process entails two" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "phases: pre-generating an object and integrating its latent representation into the original image's latent space. Initially, we use a diffusion model to create an object within a designated bounding box, followed by precise segmentation using models (e.g., SAM [9]). This object is then processed through a backward diffusion sequence with our base diffusion model, yielding masked latent layers corresponding to the object, which are later merged with the original canvas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Deletion operation begins with SAM refining the boundary of the object within its bounding box. The latent layers associated with these specified regions are then removed and reset with Gaussian noise. This necessitates a complete regeneration of these areas in the final denoising process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Repositioning involves modifying the original image to align objects with new bounding boxes, taking care to preserve their original aspect ratios. The initial steps include shifting and resizing the bounding box in the image space. Following this, SAM refines the object boundary, succeeded by a backward diffusion process to generate its relevant latent layers, similar to the approach in the addition operation. Latent layers corresponding to the excised parts are replaced with Gaussian noise, while the newly added sections are integrated into the final image composition. An important consideration in repositioning is conducting object resizing in the image space rather than the latent space to maintain high-quality results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Attribute modification starts with SAM refining the object boundary within the bounding box, followed by applying attribute modifications such as DiffEdit [4]. The base diffusion model then reverses the image, producing a series of masked latent layers ready for final composition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.669 + ], + "angle": 0, + "content": "After editing operations on each object, we proceed to the recomposition phase as shown in Fig. 4. In this phase, while latents for removed or repositioned regions are reinitialized with Gaussian noise, the latents for added or modified latents are updated accordingly. For regions with multiple overlapping objects, we place the larger masks first to ensure the visibility of the smaller objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.806 + ], + "angle": 0, + "content": "The stack of modified latent then undergoes a final forward diffusion process, which begins with steps in which regions not reinitialized with Gaussian noise are frozen (i.e., forced to align with the unmodified latent at the same step). This is crucial for the accurate formation of updated objects while maintaining background consistency at the same time. The procedure finishes with several steps where everything is allowed to change, resulting in a visually coherent and correct image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.852, + 0.844 + ], + "angle": 0, + "content": "3.2.2 Termination of the Self-Correction Process" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While we observe that one round of generation is often enough for a majority of the cases, subsequent rounds could still benefit the performance in terms of correctness further," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6331" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.165, + 0.089, + 0.805, + 0.276 + ], + "angle": 0, + "content": "
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
MultiDiffusion [1]29%28%26%39%30.5%
Backward Guidance [3]22%37%26%67%38.0%
BoxDiff [23]22%30%37%71%40.0%
LayoutGPT + GLIGEN [6, 10]36%65%26%78%51.3%
DALL-E 3 [16]25%38%74%71%52.0%
+ 1-round SLD (Ours)90%61%80%83%78.5% (+26.5)
LMD+ [11]100%82%49%86%79.3%
+ 1-round SLD (Ours)100%98%63%92%88.3% (+9.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.286, + 0.868, + 0.3 + ], + "angle": 0, + "content": "Table 1. Our method can be applied to various image generation methods and improves the generation accuracy by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.468, + 0.409 + ], + "angle": 0, + "content": "making our self-correction an iterative process. Thus, determining the optimal number of self-correction rounds is critical for balancing efficiency and accuracy. As outlined in Algorithm 1, our method sets a maximum number of attempts on the correction rounds to ensure the process finishes within a reasonable amount of time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.468, + 0.562 + ], + "angle": 0, + "content": "The process completes when the LLM outputs the same layout as the input (i.e., if the bounding boxes suggested by the LLM controller \\((B_{next})\\) align with the current detected bounding boxes \\((B_{curr}))\\), or when the maximum rounds of generation are reached, indicating that the method is unable to make a correct generation for the prompt. This iterative process provides guarantees on the correctness of the image, up to the accuracy of the detector and the LLM controller, ensuring it aligns closely with the initial text prompt. We explore the efficacy of multi-round corrections in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.453, + 0.59 + ], + "angle": 0, + "content": "3.3. Unified text-to-image generation and editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.468, + 0.748 + ], + "angle": 0, + "content": "In addition to self-correcting image generation models, our SLD framework is readily adaptable for image editing applications, requiring only minimal modifications. A key distinction is in the format of user input prompts. Unlike image generation, where users provide scene descriptions, image editing requires users to detail both the original image and the desired changes. For instance, to edit an image with two apples and a banana by replacing the banana with an orange, the prompt could be: \"Replace the banana with an orange, while keeping the two apples unchanged.\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.468, + 0.884 + ], + "angle": 0, + "content": "The editing process is similar to our self-correction mechanism. The LLM parser extracts key objects from the user's prompt. These objects are then identified by the open-vocabulary detector, establishing a list of current bounding boxes. The editing-focused LLM controller, equipped with specific task objectives, guidelines, and in-context examples, analyzes these inputs. It proposes updated bounding boxes and corresponding latent space operations for precise image manipulation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "SLD's ability to perform detailed, object-level editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.456 + ], + "angle": 0, + "content": "distinguishes it from existing diffusion-based methods like InstructPix2Pix [2] and prompt2prompt [8], which mainly address global image style changes. Also, SLD outperforms tools like DiffEdit [4] and SDEdit [15], which are restricted to object replacement or attribute adjustment, by enabling comprehensive object repositioning, addition, and deletion with exact control. Our comparative analysis in Sec. 4.2 will further highlight SLD's superior editing capabilities over existing methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.472, + 0.633, + 0.489 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.884, + 0.514 + ], + "angle": 0, + "content": "4.1. Comparison with Image Generation Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Setup. We evaluate the performance of the SLD framework with the LMD benchmark [11], which is specifically designed to evaluate generation methods on complex tasks such as handling negation, numeracy, accurate attribute binding to multiple objects, and spatial reasoning. For each task, 100 programmatically generated prompts are fed into various text-to-image generation methods to produce corresponding images. We evaluate the images generated by our method and the baselines with open-vocabulary detector OWL-ViT v2 [14] for a robust quantitative evaluation of the alignment between the input prompts and the generated images. We compared SLD with several leading text-to-image diffusion methods, such as Multidiffusion [1], BoxDiff [23], LayoutGPT [6], LMD+ [12], and DALL-E 3 [16]. To ensure fair comparisons, all models incorporating LLMs used the same GPT-4 model. In our SLD implementation, we utilized LMD+ as the base model for latent space operations and OWL-ViT v2 for the open-vocabulary object detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results. As shown in Tab. 1, applying the SLD method to both open-source \\((\\mathrm{LMD}+)\\) and proprietary models (DALL-E 3) significantly enhances their performance in terms of generation correctness. For negation tasks, as \\(\\mathrm{LMD}+\\) converts user prompts containing \"without\" information into negative prompts, which already achieves a remarkable \\(100\\%\\) accuracy without SLD integration. In contrast, even" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6332" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.123, + 0.09, + 0.853, + 0.111 + ], + "angle": 0, + "content": "An oil painting at the beach of a blue bicycle to the left of a bench and to the right of a palm tree with three seagulls in the sky" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.117, + 0.21, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.218, + 0.16, + 0.228 + ], + "angle": 0, + "content": "SDXL" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.117, + 0.346, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.218, + 0.331, + 0.229 + ], + "angle": 0, + "content": "SDXL w/ SLD (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.117, + 0.482, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.218, + 0.435, + 0.228 + ], + "angle": 0, + "content": "LMD+" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.117, + 0.62, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.218, + 0.608, + 0.229 + ], + "angle": 0, + "content": "LMD+ w/ SLD (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.117, + 0.755, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.218, + 0.717, + 0.228 + ], + "angle": 0, + "content": "DALL-E3" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.117, + 0.892, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.218, + 0.888, + 0.229 + ], + "angle": 0, + "content": "DALL-E 3 w/ SLD (Ours)" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.241, + 0.89, + 0.254 + ], + "angle": 0, + "content": "A realistic photo with a monkey sitting above a green motorcycle on the left and another raccoon sitting above a blue motorcycle on the right" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.257, + 0.21, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.36, + 0.162, + 0.37 + ], + "angle": 0, + "content": "SDXL" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.257, + 0.346, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.36, + 0.331, + 0.371 + ], + "angle": 0, + "content": "SDXL w/ SLD (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.257, + 0.482, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.36, + 0.435, + 0.37 + ], + "angle": 0, + "content": "LMD+" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.257, + 0.619, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.36, + 0.609, + 0.371 + ], + "angle": 0, + "content": "LMD+ w/ SLD (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.257, + 0.756, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.36, + 0.716, + 0.37 + ], + "angle": 0, + "content": "DALL-E3" + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.257, + 0.892, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.36, + 0.888, + 0.371 + ], + "angle": 0, + "content": "DALL-E 3 w/ SLD (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Figure 5. SLD enhances text-to-image alignment across diverse diffusion-based generative models such as SDXL, LMD+, and DALL-E 3. Notably, as highlighted by the red boxes in the first row, SLD precisely positions a blue bicycle between a bench and a palm tree, while maintaining the accurate count of palm trees and seagulls. The second row further demonstrates SLD's robustness in complex, cluttered scenes, effectively managing object collision through our training-free latent operations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.734 + ], + "angle": 0, + "content": "though DALL-E 3 also uses an LLM to rewrite the prompt, it still fails for some negation cases, likely because the LLM simply puts the negation keyword (e.g., \"without\") into the rewritten prompt. In this case, our SLD method can automatically rectify most of these errors. For numeracy tasks, integrating SLD with \\(\\mathrm{LMD + }\\) results in a significant improvement, with up to \\(98\\%\\) accuracy. We noted that DALL-E 3 often struggles to generate an image with the correct number of objects. However, this issue is substantially mitigated by SLD, which enhances performance by over \\(20\\%\\). For attribute binding tasks, SLD improves the performance of both DALL-E 3 and \\(\\mathrm{LMD + }\\) by \\(6\\%\\) and \\(14\\%\\), respectively. Notably, DALL-E 3 initially outperforms \\(\\mathrm{LMD + }\\) in this task, likely due to its training on high-quality image caption datasets. Finally, for spatial reasoning tasks, the integration of SLD with both \\(\\mathrm{LMD + }\\) and DALL-E 3 demonstrates enhanced performance by \\(12\\%\\) and \\(6\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.336, + 0.759 + ], + "angle": 0, + "content": "4.2. Application to Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As discussed in Sec. 3.3, SLD excels in fine-grained image editing over existing methods. As demonstrated in Fig. 6, our integration of an open-language detector with LLMs enables precise modifications within localized latent space regions. SLD adeptly performs specific edits, like seamlessly replacing an apple with a pumpkin, while preserving the integrity of surrounding objects. In contrast, methods like InstructPix2Pix [2] are confined to global transformations, and DiffEdit [4] often fails to accurately locate objects for" + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.453, + 0.71, + 0.464 + ], + "angle": 0, + "content": "Editing Instructions for SLD" + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.475, + 0.747, + 0.507 + ], + "angle": 0, + "content": "Replace the left apple with a pumpkin and keep the right apple and all bananas in the scene unchanged." + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.45, + 0.871, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.514, + 0.545, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.542, + 0.631, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.629, + 0.617, + 0.638 + ], + "angle": 0, + "content": "InstructPix2Pix" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.542, + 0.75, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.629, + 0.714, + 0.638 + ], + "angle": 0, + "content": "DiffEdit" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.542, + 0.87, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.629, + 0.844, + 0.638 + ], + "angle": 0, + "content": "SLD (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.645, + 0.892, + 0.714 + ], + "angle": 0, + "content": "Figure 6. When instructed to perform object-level image editing. InstructPix2Pix [2] completely fails to accomplish the task, and DiffEdit [4] falls short, as highlighted in the green box of the image. Conversely, our method demonstrates a significantly better performance in executing these object-level edits." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.735, + 0.777, + 0.749 + ], + "angle": 0, + "content": "modification, leading to undesired results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Furthermore, as exemplified in Fig. 7, SLD supports a wide array of editing instructions, including counting control (such as adding, deleting, or replacing objects), attribute modification (like altering colors or materials), and intricate location control (encompassing object swapping, resizing, and moving). A standout example is featured in the \"Object Resize\" column of the first row, where SLD precisely enlarges the cup on the table by an exact factor of \\(1.25 \\times\\). We encourage readers to verify this with a ruler for a clear demonstration of our method's precision. This level of pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6333" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.21, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.216, + 0.192, + 0.226 + ], + "angle": 0, + "content": "User Instructions" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.091, + 0.346, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.205, + 0.343, + 0.237 + ], + "angle": 0, + "content": "Replace the bowl with two round apples and keep the cup unchanged" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.091, + 0.482, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.205, + 0.478, + 0.227 + ], + "angle": 0, + "content": "Turn the cup into green and the bowl into golden" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.619, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.205, + 0.612, + 0.226 + ], + "angle": 0, + "content": "Swap the position of the bowl and the cup" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.091, + 0.755, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.205, + 0.747, + 0.236 + ], + "angle": 0, + "content": "Make cup 1.25 times its larger but keep the bowl's size" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.091, + 0.892, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.205, + 0.876, + 0.237 + ], + "angle": 0, + "content": "Move the cup towards the bowl until they slightly overlap" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.248, + 0.21, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.371, + 0.194, + 0.381 + ], + "angle": 0, + "content": "User Instructions" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.248, + 0.346, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.361, + 0.341, + 0.393 + ], + "angle": 0, + "content": "Replace the dog with a houseplant and keep the cat unchanged" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.248, + 0.482, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.361, + 0.476, + 0.393 + ], + "angle": 0, + "content": "Transform the dog into a robotic dog and the cat into a sculptural cat" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.248, + 0.619, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.361, + 0.614, + 0.392 + ], + "angle": 0, + "content": "Swap the position of the black dog and the yellow cat" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.248, + 0.755, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.361, + 0.756, + 0.382 + ], + "angle": 0, + "content": "Increase the cat's size and reduce the dog's size" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.248, + 0.892, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.361, + 0.892, + 0.393 + ], + "angle": 0, + "content": "Shift the dog slightly upward and move the cat slightly towards the dog" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.41, + 0.893, + 0.438 + ], + "angle": 0, + "content": "Figure 7. SLD can handle a diverse array of image editing tasks guided by natural, human-like instructions. Its capabilities span from adjusting object counts to altering object attributes, positions, and sizes." + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.457, + 0.422, + 0.467 + ], + "angle": 0, + "content": "Remove hair of the rightmost person sitting at a cafe in this image" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.469, + 0.202, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.564, + 0.174, + 0.572 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.469, + 0.335, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.564, + 0.307, + 0.572 + ], + "angle": 0, + "content": "Removed Region" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.469, + 0.468, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.564, + 0.436, + 0.572 + ], + "angle": 0, + "content": "Failed Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Figure 8. SLD struggles with objects of complex shapes, as the SAM module may unintentionally segment adjacent parts during the process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.469, + 0.718 + ], + "angle": 0, + "content": "cision stems from the detector's exact object localization coupled with the LLMs' ability in reasoning and suggestions for new placements. Such detailed control over spatial adjustments is unmatched by any previous method, highlighting SLD's contributions to fine-grained image editing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.195, + 0.74 + ], + "angle": 0, + "content": "4.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Multi-round self-correction. Our analysis in Tab. 2 highlights the benefits of multi-round self-correction and the fact that the first round correction is always the most effective one and has a marginal effect. The first round of corrections substantially mitigates issues inherent in Stable Diffusion [20]. Then, a second round of correction still yields significant improvements across all four tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Limitations and future work. A limitation of our method is illustrated in Fig. 8, where SLD fails to accurately remove a person's hair. In this instance, despite the successful iden-" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.453, + 0.89, + 0.536 + ], + "angle": 0, + "content": "
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
SD [20]19%38%24%33%28.5%
+ 1-round SLD69%55%25%69%54.5% (+26.0)
+ 2-round SLD73%61%31%75%60.0% (+31.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.547, + 0.893, + 0.589 + ], + "angle": 0, + "content": "Table 2. While the majority of errors are typically rectified in the first round, multi-round correction consistently outperforms a single-round approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.732 + ], + "angle": 0, + "content": "tification and localization of the hair, the complex nature of its shape poses a challenge to the SAM module used for region selection, resulting in the unintended removal of the person's face in addition to the hair. However, since the person's cloth is not removed, the base diffusion model fails to generate a natural composition. This suggests that a better region selection method is needed for further improvements in the generation and editing quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.75, + 0.619, + 0.765 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce the Self-correcting Language-Driven (SLD) framework, a pioneering self-correction system using detectors and LLMs to significantly enhance text-to-image alignment. This method not only sets a new SOTA in the image generation benchmark but is also compatible with various generative models, including DALL-E 3. Also, SLD extends its utility to image editing applications, offering fine-grained object-level manipulation that surpasses existing methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6334" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[4] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[6] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.471, + 0.528 + ], + "angle": 0, + "content": "[7] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LLM blueprint: Enabling text-to-image generation with complex and detailed prompts. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.471, + 0.585 + ], + "angle": 0, + "content": "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[10] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[11] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[13] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.783, + 0.178 + ], + "angle": 0, + "content": "[16] OpenAI. Dall-e 3 system card, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.18, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[17] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.893, + 0.319 + ], + "angle": 0, + "content": "[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.893, + 0.376 + ], + "angle": 0, + "content": "[19] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.894, + 0.448 + ], + "angle": 0, + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.893, + 0.532 + ], + "angle": 0, + "content": "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[22] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.606, + 0.893, + 0.674 + ], + "angle": 0, + "content": "[23] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.893, + 0.758 + ], + "angle": 0, + "content": "[24] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.893, + 0.842 + ], + "angle": 0, + "content": "[25] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[26] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6335" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.136 + ], + "angle": 0, + "content": "[27] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6336" + } + ] +] \ No newline at end of file diff --git a/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_origin.pdf b/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a741b4b0babd2b8a63f56a458ab148ba7fcebf4 --- /dev/null +++ b/2024/Self-correcting LLM-controlled Diffusion Models/1d07a470-ae6d-42d9-bf7a-62b3e923ba50_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a16d2563959482148ab728924a2d7489c10e8515d250cb16a2e0d2097173f6e +size 8733883 diff --git a/2024/Self-correcting LLM-controlled Diffusion Models/full.md b/2024/Self-correcting LLM-controlled Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d2cc6f84b8f1f2d97f35b8499d8cfa080f682e1 --- /dev/null +++ b/2024/Self-correcting LLM-controlled Diffusion Models/full.md @@ -0,0 +1,340 @@ +# Self-correcting LLM-controlled Diffusion Models + +Tsung-Han Wu* + +Long Lian* + +Joseph E. Gonzalez +UC Berkeley + +Boyi Li† + +Trevor Darrell† + +![](images/b5134225e7f860ac442562a0b5ef10c0cc620556e3b7975b8e52780e41086f53.jpg) +DALL-E3 +A realistic cartoon-style image with a princess and four dwarfs +Figure 1. Existing diffusion-based text-to-image generators (e.g., DALL-E 3 [16]) generally struggle to precisely generate images that correctly align with complex input prompts, especially for the ones that require numeracy and spatial relationships. Our Self-correcting LLM-controlled Diffusion (SLD) framework empowers these diffusion models to automatically and iteratively rectify inaccuracies through applying a set of latent space operations (addition, deletion, repositioning, etc.), resulting in enhanced text-to-image alignment. + +![](images/df5fa1e7fbafa99f44cf65d8f65d7f9505a6dfdd01c0b90bfcca41c1611c22c9.jpg) +Numeracy Correction +DALL-E 3 + SLD (Ours) + +![](images/1efcfb89b8bb173e5258d71b5e811377ddeb59708f3d3905bde442218b50d165.jpg) +Spatial Correction +DALL-E3 +A vivid photo with a woman on the right and a clown on the left walking in a dirty alley + +![](images/e7bd7f0277e99d0cf48cb67ad61b9ce1a46e68da8113c0b21a7c0f38c56347ca.jpg) +DALL-E 3 + SLD (Ours) + +# Abstract + +Text-to-image generation has witnessed significant progress with the advent of diffusion models. Despite the ability to generate photorealistic images, current text-to-image diffusion models still often struggle to accurately interpret and follow complex input text prompts. In contrast to existing models that aim to generate images only with their best effort, we introduce Self-correcting LLM-controlled Diffusion (SLD). SLD is a framework that generates an image from the input prompt, assesses its alignment with the prompt, and performs self-corrections on the inaccuracies in the generated image. Steered by an LLM controller, SLD turns text-to-image generation into an iterative closed-loop process, ensuring correctness in the resulting image. SLD is not only training-free but can also be seamlessly integrated with diffusion models behind API access, such as DALL-E 3, to further boost the performance of state-of-the-art diffusion models. Experimental results show that our approach can rectify a majority of incorrect generations, particularly in generative numeracy, attribute binding, and spatial relationships. Furthermore, by simply adjusting the instructions to the LLM, SLD can perform image editing tasks, bridging the gap between text-to-image generation + +and image editing pipelines. Our code is available at: https://self-correcting-llm-diffusion.github.io. + +# 1. Introduction + +Text-to-image generation has made remarkable advancements, especially with the advent of diffusion models. However, these models often struggle with interpreting complex input text prompts, particularly those that require skills such as understanding the concept of numeracy, spatial relationships, and attribute binding with multiple objects. Despite the astonishing scaling of model sizes and training data, these challenges, as illustrated in Fig. 1, are still present in state-of-the-art open-source and proprietary diffusion models. + +Several research and engineering efforts aim to overcome these limitations. For instance, methods such as DALL-E 3 [16] focus on the diffusion training process and incorporate high-quality captions into the training data at a massive scale. However, this approach not only incurs substantial costs but also frequently falls short in generating accurate images from complicated user prompts, as shown in Fig. 1. Other work harnesses the power of external models for a better understanding of the prompt in the inference process before the actual image generation. For example, [6, 11] leverages Large Language Models (LLMs) to pre + +process textual prompts into structured image layouts and thus ensures the preliminary design aligns with the user's directives. However, such integration does not resolve the inaccuracies produced by the downstream diffusion models, particularly in images with complex scenarios like multiple objects, cluttered arrangements, or detailed attributes. + +Drawing inspiration from the process of a human painting and a diffusion model in generating images, we observe a key distinction in their approach to creation. Consider a human artist tasked with painting a scene featuring two cats. Throughout the painting process, the artist remains cognizant of this requirement, ensuring that two cats are indeed present before considering the work complete. Should the artist find only one cat depicted, an additional one would be added to meet the prompt's criteria. This contrasts sharply with current text-to-image diffusion models, which operate on an open-loop basis. These models generate images through a predetermined number of diffusion steps and present the output to the user, regardless of its alignment with the initial user prompt. Such a process, irrespective of scaling training data or LLM pre-generation conditioning, lacks a robust mechanism to ensure the final image aligns with the user's expectations. + +In light of this, we propose our method Self-correcting LLM-controlled Diffusion (SLD) that performs self-checks to confidently offer users guarantees of the alignment between the prompt and the generated images. Departing from conventional single-round generation methods, SLD is a novel closed-loop approach that equips diffusion models with the ability to iteratively identify and rectify errors. Our SLD framework, illustrated in Fig. 2, contains two main components: LLM-driven object detection as well as LLM-controlled assessment and correction. + +The SLD pipeline follows a standard text-to-image generation setting. Given a textual prompt that outlines the desired image, SLD begins with calling an image generation module (e.g., the aforementioned open-loop text-to-image diffusion models) to generate an image in a best-effort fashion. Given that these open-loop generators do not guarantee an output that aligns perfectly with the prompt, SLD then conducts a thorough evaluation of the produced image against the prompt, with an LLM parsing key phrases for an open-vocabulary detector to check. Subsequently, an LLM controller takes the detected bounding boxes and the initial prompt as input and checks for potential mismatches between the detection results and the prompt requirements, suggesting appropriate self-correction operations, such as adding, moving, and removing objects. Finally, utilizing a chosen base diffusion model (e.g., Stable Diffusion [20]), SLD employs latent space composition to implement these adjustments, thereby ensuring that the final image accurately reflects the user's initial text prompt. + +Notably, SLD is agnostic to the initial generation pro + +cess, which allows us to use DALL-E 3 APIs as a subroutine to generate initial images and then employ the open-source Stable Diffusion model family to conduct latent operations to attain self-correction. Furthermore, none of these operations require additional training on our base diffusion model [20], which easily allows our method to be applied to various diffusion models without external human annotation or training. + +We demonstrate that our SLD framework can achieve significant improvement over current diffusion-based methods on complex prompts with the LMD benchmark [11]. The results show that our method can surpass $\mathrm{LMD + }$ , which is a strong baseline that already leverages LLM in the image process generation, by $9.0\%$ . More importantly, with DALL-E 3 for initial generation, the generated images from our method achieve $26.5\%$ performance gains compared to ones before self-correction. + +Finally, since the SLD pipeline is agnostic to the initially generated image, it can easily be transformed into an image editing pipeline by simply changing the prompts to the LLM. While text-to-image generation and image editing are often treated as distinct tasks by the generative modeling community, our SLD can perform these two tasks with a unified pipeline. We list our key contributions below: + +1. SLD is the first to integrate a detector and an LLM to self-correct generative models, ensuring accurate generation without extra training or external data. +2. SLD offers a unified solution for both image generation and editing, enabling enhanced text-to-image alignment for any image generator (e.g., DALL-E 3) and object-level editing on any images. +3. Experimental results show that SLD can correct a majority of incorrect generations, particularly in aspects of numeracy, attribute binding, and spatial relationships. + +# 2. Related Work + +# 2.1. Text-to-Image Diffusion Models + +Diffusion-based text-to-image generation has advanced significantly. Initial studies [19-21] showed diffusion models' ability to create high-quality images, but they struggle with complex prompts. Subsequent research [3, 10, 18, 24-26] has incorporated additional inputs such as keypoints and bounding boxes to control the diffusion generation process. + +Recent advancements have incorporated LLMs to control the generation of diffusion models, bypassing the need for additional complementary information as inputs [6, 7, 11-13, 27]. In these approaches, LLMs play a central role in directly interpreting user textual prompts and managing the initial layout configuration. Despite some progress, these models often operate in an open-loop fashion, producing images in one iteration that cannot guarantee the generated images align with user prompts. + +![](images/c6ef2e8b092fa1cf62a5be3dfea74fe4f09fda2ad776123a781fe294cdb27ee7.jpg) +Figure 2. Our proposed Self-correcting LLM-controlled Diffusion (SLD) enhances text-to-image alignment through an iterative self-correction process. It begins with LLM-driven object detection (Sec. 3.1), and subsequently performs LLM-controlled analysis and correction (Sec. 3.2). The entire pipeline is outlined in Algorithm 1. + +Unlike prior work, SLD is the first closed-loop diffusion-based generation method. Integrating advanced object detectors and LLMs, SLD performs iterative self-checking and correction, significantly enhancing text-to-image alignment. This improvement spans numeracy, attribute binding to multiple objects, and spatial reasoning, applicable to various models, including those like DALL-E 3. Also, SLD extends beyond existing NLP self-correction techniques [17] by incorporating bounding box spatial representation and novel latent editing mechanisms tailored for text-to-image tasks. These designs are essential for ensuring high-quality, precise self-correction in image generation tasks. + +# 2.2. Diffusion-based Image Editing + +Recent advancements in text-to-image diffusion models have significantly expanded their applications in image editing, encompassing both global and local editing tasks. Techniques like Prompt-2-prompt [8] and InstructPix2Pix [2] specialize in global edits, such as style transformations. Conversely, methods like SDEdit [15], DiffEdit [4], and Plug-and-Play [22] focus on local edits, targeting specific areas within images. Despite their progress, these methods often struggle with precise object-level manipulation and tasks that require spatial reasoning, such as resizing or repositioning objects. While recent approaches like Self-Guidance [5] offer fine-grained operations, they still necessities to be able to use the same techniques. + +sitate user inputs for specific coordinates when moving or repositioning objects. + +Unlike these methods only focus on diffusion models, SLD introduces the combination of detectors and LLMs in the loop of editing, enabling fine-grained editing with only user prompts. Also, SLD excels in various object-level editing tasks, including adding, replacing, moving, and modifying attributes, swapping, and so on, demonstrating a notable improvement in both ease of use and editing capabilities. + +# 3. Self-correcting LLM-controlled Diffusion + +In this section, we introduce our Self-correcting LLM-controlled Diffusion (SLD) framework. SLD consists of two main components: LLM-driven object detection (Sec. 3.1) as well as LLM-controlled assessment and correction (Sec. 3.2). Moreover, with a simple change of the LLM instructions, we show that SLD is applicable to image editing, unifying text-to-image generation, and editing as discussed in Sec. 3.3. The complete pipeline is shown in Algorithm 1. + +# 3.1. LLM-driven Object Detection + +Our SLD framework starts with LLM-driven object detection, which extracts information necessary for downstream assessment and correction. As shown with green arrows in + +![](images/4ef7b66afd1c8e32ee938f07acbf4cec29c40bfd4b0aa5f235184b10e95678a3.jpg) +Figure 3. Our self-correction pipeline is driven by two distinct LLMs: (a) The LLM parser analyzes user prompts $P$ to extract a list of key object information $S$ , which is then passed to the open-vocabulary detector. (b) The LLM controller, taking both the user prompt $P$ and currently detected bounding boxes $B_{curr}$ as input, outputs suggested new bounding boxes $B_{next}$ . These are subsequently transformed into a set of latent space operations $Ops$ for image manipulation. + +![](images/19ecd4193069023ec26d3a6a51f7c80079e6560e6bbd68d5a24664a8d63fab59.jpg) + +![](images/4fdbe53657b21fc76c661e3adc40c354300523c8f2d83e0041b4876177d3c0b8.jpg) + +Fig. 2, the LLM-driven object detection includes two steps: 1) We leverage an LLM as a parser that parses the user prompt and outputs key phrases that are potentially relevant to image assessment. 2) The phrases are then passed into an open-vocabulary object detector. The detected boxes are supposed to contain information that supports the assessment of whether the image is aligned with the specifications in the user prompt. + +In the initial step, an LLM parser is directed to extract a list of key object details, denoted as $S$ , from the user-provided text prompt $P$ . This parser, aided by text instructions and in-context examples, can easily accomplish this as shown in Fig. 3 (a). For a user prompt that includes phrases like "a green motorcycle" and "a blue motorcycle," the LLM is expected to identify and output "green" and "blue" as attributes associated with "motorcycle." When the prompt references objects without specific quantities or attributes, such as "a monkey" and "a raccoon," these descriptors are appropriately left blank. Importantly, the LLM's role is not limited to merely identifying object nouns; it also entails identifying any associated quantities or attributes. + +In the second step, an open-vocabulary detector processes the list of key object information, $S$ , parsed in the first step, to detect and localize objects within the image. We prompt the open-vocabulary object detector with queries formatted as image of a/an [attribute] [object name], where the "attribute" and "object name" are sourced from the parser's output. The resulting bounding boxes, $B_{curr}$ , are then organized into a list format like ["[attribute] [object name] [#object ID]", [x, y, w, h])] for further processing. A special case is when the prompt poses constraints on the object quantity. For cases where attributed objects (e.g., "blue dog") fall short compared to the required + +quantities, a supplementary count of non-attributed objects (e.g., "dog") is provided to provide context for the subsequent LLM controller deciding whether to add more "blue dogs" or simply alter the color of existing dogs to blue. We will explain these operations, including object addition and attribute modification, in greater detail in Sec. 3.2.1. + +# 3.2. LLM-controlled Analysis and Correction + +We use an LLM controller for image analysis and the subsequent correction. The controller, given the user prompt $P$ and detected boxes $B_{curr}$ , is asked to analyze whether the image, represented by objects bounding boxes, aligns with the description of the user prompt and offers a list of corrected bounding boxes $B_{next}$ , as shown in Fig. 3 (b). + +SLD then programmatically analyzes the inconsistencies between the refined and original bounding boxes to output a set of editing operations $Ops$ , which includes addition, deletion, repositioning, and attribute modification. However, a simple set-of-boxes representation does not carry correspondence information, which does not allow an easy way to compare the input and the output layout of the LLM controller when multiple boxes share the same object name. For example, when there are two cat boxes in both the model input and the model output, whether one cat box corresponds to which cat box in the output layout is unclear. Rather than introducing another algorithm to guess the correspondence, we propose to let the LLM output correspondence with a very simple edit: we give an object ID to each bounding box, with the number increasing within each object type, as a suffix added after the object name. In the in-context examples, we demonstrate to the LLM that the object should have the same name and object ID before and after the proposed correction. + +Algorithm 1 Self-correction for Image Generation. +Input: User prompt $P$ , Initial generated image $I$ , Maximum number of self-correction round $K$ . +1: for $k\gets 1$ to $K$ do +2: $S\gets \mathsf{LLM - Parse}(P)$ +3: $B_{curr}\gets \mathsf{Detector}(S)$ +4: $B_{next}\gets \mathsf{LLM - Analysis}(P,B_{curr})$ +5: $Ops\gets \mathsf{Diff}(B_{curr},B_{next})$ +6: if $Ops\neq \emptyset$ then (i.e., $B_{next}\neq B_{curr}$ ) +7: $I = \mathsf{Correction}(Ops,B_{next},B_{curr})$ +8: else +9: break +10: end if +11: end for +Output: Image $I$ + +# 3.2.1 Latent Operations for Training-Free Image Correction + +The LLM controller outputs a list of correction operations to apply. For each operation, we first transform the original image into latent features. Our approach then executes a series of operations $Ops$ , such as addition, deletion, repositioning, and attribute modification, applied to these latent layers. It is worth noting that although there are only four operations, they are sufficient to handle a majority of misalignments naturally. We explain how each operation is performed below. + +![](images/d8ea5419e73e8ec93f27cba35120f3e932f85bab62ba6e1918b4065f17a72568.jpg) +Figure 4. Our latent operations can be summarized into two key concepts: (1) latent in removed regions are re-initialized to Gaussian noise, and latent of newly added or modified objects are composed onto the canvas. (2) Latent composition is confined to the initial steps, followed by "unfrozen" steps for a standard forward diffusion process, enhancing visual quality and avoiding artificial copy-and-paste effects. + +Addition. Inspired by [11], the addition process entails two + +phases: pre-generating an object and integrating its latent representation into the original image's latent space. Initially, we use a diffusion model to create an object within a designated bounding box, followed by precise segmentation using models (e.g., SAM [9]). This object is then processed through a backward diffusion sequence with our base diffusion model, yielding masked latent layers corresponding to the object, which are later merged with the original canvas. + +Deletion operation begins with SAM refining the boundary of the object within its bounding box. The latent layers associated with these specified regions are then removed and reset with Gaussian noise. This necessitates a complete regeneration of these areas in the final denoising process. + +Repositioning involves modifying the original image to align objects with new bounding boxes, taking care to preserve their original aspect ratios. The initial steps include shifting and resizing the bounding box in the image space. Following this, SAM refines the object boundary, succeeded by a backward diffusion process to generate its relevant latent layers, similar to the approach in the addition operation. Latent layers corresponding to the excised parts are replaced with Gaussian noise, while the newly added sections are integrated into the final image composition. An important consideration in repositioning is conducting object resizing in the image space rather than the latent space to maintain high-quality results. + +Attribute modification starts with SAM refining the object boundary within the bounding box, followed by applying attribute modifications such as DiffEdit [4]. The base diffusion model then reverses the image, producing a series of masked latent layers ready for final composition. + +After editing operations on each object, we proceed to the recomposition phase as shown in Fig. 4. In this phase, while latents for removed or repositioned regions are reinitialized with Gaussian noise, the latents for added or modified latents are updated accordingly. For regions with multiple overlapping objects, we place the larger masks first to ensure the visibility of the smaller objects. + +The stack of modified latent then undergoes a final forward diffusion process, which begins with steps in which regions not reinitialized with Gaussian noise are frozen (i.e., forced to align with the unmodified latent at the same step). This is crucial for the accurate formation of updated objects while maintaining background consistency at the same time. The procedure finishes with several steps where everything is allowed to change, resulting in a visually coherent and correct image. + +# 3.2.2 Termination of the Self-Correction Process + +While we observe that one round of generation is often enough for a majority of the cases, subsequent rounds could still benefit the performance in terms of correctness further, + +
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
MultiDiffusion [1]29%28%26%39%30.5%
Backward Guidance [3]22%37%26%67%38.0%
BoxDiff [23]22%30%37%71%40.0%
LayoutGPT + GLIGEN [6, 10]36%65%26%78%51.3%
DALL-E 3 [16]25%38%74%71%52.0%
+ 1-round SLD (Ours)90%61%80%83%78.5% (+26.5)
LMD+ [11]100%82%49%86%79.3%
+ 1-round SLD (Ours)100%98%63%92%88.3% (+9.0)
+ +Table 1. Our method can be applied to various image generation methods and improves the generation accuracy by a large margin. + +making our self-correction an iterative process. Thus, determining the optimal number of self-correction rounds is critical for balancing efficiency and accuracy. As outlined in Algorithm 1, our method sets a maximum number of attempts on the correction rounds to ensure the process finishes within a reasonable amount of time. + +The process completes when the LLM outputs the same layout as the input (i.e., if the bounding boxes suggested by the LLM controller $(B_{next})$ align with the current detected bounding boxes $(B_{curr}))$ , or when the maximum rounds of generation are reached, indicating that the method is unable to make a correct generation for the prompt. This iterative process provides guarantees on the correctness of the image, up to the accuracy of the detector and the LLM controller, ensuring it aligns closely with the initial text prompt. We explore the efficacy of multi-round corrections in Sec. 4.3. + +# 3.3. Unified text-to-image generation and editing + +In addition to self-correcting image generation models, our SLD framework is readily adaptable for image editing applications, requiring only minimal modifications. A key distinction is in the format of user input prompts. Unlike image generation, where users provide scene descriptions, image editing requires users to detail both the original image and the desired changes. For instance, to edit an image with two apples and a banana by replacing the banana with an orange, the prompt could be: "Replace the banana with an orange, while keeping the two apples unchanged." + +The editing process is similar to our self-correction mechanism. The LLM parser extracts key objects from the user's prompt. These objects are then identified by the open-vocabulary detector, establishing a list of current bounding boxes. The editing-focused LLM controller, equipped with specific task objectives, guidelines, and in-context examples, analyzes these inputs. It proposes updated bounding boxes and corresponding latent space operations for precise image manipulation. + +SLD's ability to perform detailed, object-level editing + +distinguishes it from existing diffusion-based methods like InstructPix2Pix [2] and prompt2prompt [8], which mainly address global image style changes. Also, SLD outperforms tools like DiffEdit [4] and SDEdit [15], which are restricted to object replacement or attribute adjustment, by enabling comprehensive object repositioning, addition, and deletion with exact control. Our comparative analysis in Sec. 4.2 will further highlight SLD's superior editing capabilities over existing methods. + +# 4. Experiments + +# 4.1. Comparison with Image Generation Methods + +Setup. We evaluate the performance of the SLD framework with the LMD benchmark [11], which is specifically designed to evaluate generation methods on complex tasks such as handling negation, numeracy, accurate attribute binding to multiple objects, and spatial reasoning. For each task, 100 programmatically generated prompts are fed into various text-to-image generation methods to produce corresponding images. We evaluate the images generated by our method and the baselines with open-vocabulary detector OWL-ViT v2 [14] for a robust quantitative evaluation of the alignment between the input prompts and the generated images. We compared SLD with several leading text-to-image diffusion methods, such as Multidiffusion [1], BoxDiff [23], LayoutGPT [6], LMD+ [12], and DALL-E 3 [16]. To ensure fair comparisons, all models incorporating LLMs used the same GPT-4 model. In our SLD implementation, we utilized LMD+ as the base model for latent space operations and OWL-ViT v2 for the open-vocabulary object detector. + +Results. As shown in Tab. 1, applying the SLD method to both open-source $(\mathrm{LMD}+)$ and proprietary models (DALL-E 3) significantly enhances their performance in terms of generation correctness. For negation tasks, as $\mathrm{LMD}+$ converts user prompts containing "without" information into negative prompts, which already achieves a remarkable $100\%$ accuracy without SLD integration. In contrast, even + +![](images/a307dbf42a70969e49a48b6001524db16204d2c41dec9c66fbb9ad461604a612.jpg) +SDXL + +![](images/00fc34a4c7af14058bf1a7722fd49a17c4db0b582deb310bcd4ce9fb88899df8.jpg) +SDXL w/ SLD (Ours) + +![](images/5bb12932fe4e277153d47699dd97a98608cf6416fadd3164c73238cea0f8f6d9.jpg) +LMD+ + +![](images/49f6ff7fa7e479b56d53bc2e478e1c0fc720e4a34b0afa88a9dae0c0d1e60014.jpg) +LMD+ w/ SLD (Ours) + +![](images/7ca07c4e9fbae07c9d83e64a0d2f1296b633b59fa8b67998138a99d2e7f5d3e6.jpg) +DALL-E3 + +![](images/abfd065bb7574402eba14755d9d2dc6ae24e047486aabbc63a219ad46b10092c.jpg) +DALL-E 3 w/ SLD (Ours) + +A realistic photo with a monkey sitting above a green motorcycle on the left and another raccoon sitting above a blue motorcycle on the right + +![](images/fc008f569b3f5b259f3f45264c6c20c4623ab6f0f55e0b5472bd50007d10e76a.jpg) +SDXL + +![](images/848e9760730f9b5c157b58e581e601b2d71e9aa89248623e39338c51c9f4a1aa.jpg) +SDXL w/ SLD (Ours) + +![](images/bd39125a3f2b1c74e7a4d54713b29da87759770cd3bcb522e197a217fa1d8935.jpg) +LMD+ + +![](images/de80f2785db0c6313e30f13160fa251ea7158a19e0517a14096f9a2f073c8084.jpg) +LMD+ w/ SLD (Ours) + +![](images/9bf35ba779a91127c0b862ddfc7217808eae8325c3ea3226d126e65b4836ae48.jpg) +DALL-E3 + +![](images/59ae716dc097a9b7eceff7bca8ae16282faf8daa5fb4fc027c91f7b91d48c63a.jpg) +DALL-E 3 w/ SLD (Ours) + +though DALL-E 3 also uses an LLM to rewrite the prompt, it still fails for some negation cases, likely because the LLM simply puts the negation keyword (e.g., "without") into the rewritten prompt. In this case, our SLD method can automatically rectify most of these errors. For numeracy tasks, integrating SLD with $\mathrm{LMD + }$ results in a significant improvement, with up to $98\%$ accuracy. We noted that DALL-E 3 often struggles to generate an image with the correct number of objects. However, this issue is substantially mitigated by SLD, which enhances performance by over $20\%$ . For attribute binding tasks, SLD improves the performance of both DALL-E 3 and $\mathrm{LMD + }$ by $6\%$ and $14\%$ , respectively. Notably, DALL-E 3 initially outperforms $\mathrm{LMD + }$ in this task, likely due to its training on high-quality image caption datasets. Finally, for spatial reasoning tasks, the integration of SLD with both $\mathrm{LMD + }$ and DALL-E 3 demonstrates enhanced performance by $12\%$ and $6\%$ , respectively. + +# 4.2. Application to Image Editing + +As discussed in Sec. 3.3, SLD excels in fine-grained image editing over existing methods. As demonstrated in Fig. 6, our integration of an open-language detector with LLMs enables precise modifications within localized latent space regions. SLD adeptly performs specific edits, like seamlessly replacing an apple with a pumpkin, while preserving the integrity of surrounding objects. In contrast, methods like InstructPix2Pix [2] are confined to global transformations, and DiffEdit [4] often fails to accurately locate objects for + +![](images/e132a16ef20d871051607858e51ed72b624f0a8495d5598edddc3d7172a56016.jpg) +Figure 5. SLD enhances text-to-image alignment across diverse diffusion-based generative models such as SDXL, LMD+, and DALL-E 3. Notably, as highlighted by the red boxes in the first row, SLD precisely positions a blue bicycle between a bench and a palm tree, while maintaining the accurate count of palm trees and seagulls. The second row further demonstrates SLD's robustness in complex, cluttered scenes, effectively managing object collision through our training-free latent operations. + +![](images/eac1e7e74054f9eb76ff33b109371fda001ae8dd74ef83d370f49350c3af9c94.jpg) +Replace the left apple with a pumpkin and keep the right apple and all bananas in the scene unchanged. + +![](images/6af6512aaefc4900907eca27f063f9bb02b2808b2c67c1030e9fb9efe93c4a0b.jpg) +Editing Instructions for SLD +InstructPix2Pix +Figure 6. When instructed to perform object-level image editing. InstructPix2Pix [2] completely fails to accomplish the task, and DiffEdit [4] falls short, as highlighted in the green box of the image. Conversely, our method demonstrates a significantly better performance in executing these object-level edits. + +![](images/ba6a05105eb50bd4d5449e5beca659645801228aa972b74b87e24b685ffa18c5.jpg) +DiffEdit + +![](images/dcb7998c6696b4de99e8a1726bb07bfc50058240280ea802408ff1d6ddd1c9ad.jpg) +SLD (Ours) + +modification, leading to undesired results. + +Furthermore, as exemplified in Fig. 7, SLD supports a wide array of editing instructions, including counting control (such as adding, deleting, or replacing objects), attribute modification (like altering colors or materials), and intricate location control (encompassing object swapping, resizing, and moving). A standout example is featured in the "Object Resize" column of the first row, where SLD precisely enlarges the cup on the table by an exact factor of $1.25 \times$ . We encourage readers to verify this with a ruler for a clear demonstration of our method's precision. This level of pre + +![](images/325b8eecc53f1075be21ccb45648ead52c39545d53058ec27de89160a00f7620.jpg) +Replace the bowl with two round apples and keep the cup unchanged + +![](images/cec1937d48bdd41ee6ad4a7a21e590d32cf2d06ad881e976dfa18af88f7f7a74.jpg) +Turn the cup into green and the bowl into golden + +![](images/2a54fbf66c7e2da6bbd5cb856e1b457297dbaab09b21650eaa3c8ce62f8b0de9.jpg) +Swap the position of the bowl and the cup + +![](images/f3f3fefc5dcd868bd91cee3be94e35b7eba1d1290f92a98df0f9ba781d361a90.jpg) + +![](images/405001983c82bc96a9d7adce58e4359fe71dd446cb79c7eaef1780bd935bad7a.jpg) +Make cup 1.25 times its larger but keep the bowl's size + +![](images/2d45f0f869a2f801327f228e3c82f9a561e13d812f99d7eb23c8092b7d92c763.jpg) +Move the cup towards the bowl until they slightly overlap + +![](images/600d91a9dcd434d0c6a3a2844d7d41ab27582de36bc1432d15cb9ac508ba02d1.jpg) +User Instructions +User Instructions + +![](images/03f715b515708b5e9eb1b474dac5a58de7f9cc050b10a78521d40c07be8187aa.jpg) +Replace the dog with a houseplant and keep the cat unchanged + +![](images/c0b5b08eaae92c485f84cb968e8e138c14c5e730817fa18b8b31936e85e5400f.jpg) +Transform the dog into a robotic dog and the cat into a sculptural cat + +![](images/7cebd7be58e3ddce0898d7473c90d28aa1ee15d8557e3bc24b2dc1f4c4fa393f.jpg) +Swap the position of the black dog and the yellow cat + +![](images/1df495a9187b4cb0d960b07880a18be359c114da945d3719857c473d24290162.jpg) +Increase the cat's size and reduce the dog's size + +![](images/ce4dec8db7f3f3b88574e4c426f5d9c56dc8dbdf29b61e725592283454c303a9.jpg) +Shift the dog slightly upward and move the cat slightly towards the dog + +![](images/537f67b74f4d53911293856a266d53587254fb5c76c38b2ff1901a04d2182a48.jpg) +Figure 7. SLD can handle a diverse array of image editing tasks guided by natural, human-like instructions. Its capabilities span from adjusting object counts to altering object attributes, positions, and sizes. +Remove hair of the rightmost person sitting at a cafe in this image +Original Image +Figure 8. SLD struggles with objects of complex shapes, as the SAM module may unintentionally segment adjacent parts during the process. + +![](images/925c6863876afdcf3efe077f0345a18e343b4309fb0bc08fb8315ea1fb5dd0d2.jpg) +Removed Region + +![](images/a9a18118c8952e83e3bd93cfecd77665a6234d23a75eb1bbb6a3894191ff52dd.jpg) +Failed Result + +cision stems from the detector's exact object localization coupled with the LLMs' ability in reasoning and suggestions for new placements. Such detailed control over spatial adjustments is unmatched by any previous method, highlighting SLD's contributions to fine-grained image editing. + +# 4.3. Discussion + +Multi-round self-correction. Our analysis in Tab. 2 highlights the benefits of multi-round self-correction and the fact that the first round correction is always the most effective one and has a marginal effect. The first round of corrections substantially mitigates issues inherent in Stable Diffusion [20]. Then, a second round of correction still yields significant improvements across all four tasks. + +Limitations and future work. A limitation of our method is illustrated in Fig. 8, where SLD fails to accurately remove a person's hair. In this instance, despite the successful iden- + +
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
SD [20]19%38%24%33%28.5%
+ 1-round SLD69%55%25%69%54.5% (+26.0)
+ 2-round SLD73%61%31%75%60.0% (+31.5)
+ +Table 2. While the majority of errors are typically rectified in the first round, multi-round correction consistently outperforms a single-round approach. + +tification and localization of the hair, the complex nature of its shape poses a challenge to the SAM module used for region selection, resulting in the unintended removal of the person's face in addition to the hair. However, since the person's cloth is not removed, the base diffusion model fails to generate a natural composition. This suggests that a better region selection method is needed for further improvements in the generation and editing quality. + +# 5. Conclusion + +We introduce the Self-correcting Language-Driven (SLD) framework, a pioneering self-correction system using detectors and LLMs to significantly enhance text-to-image alignment. This method not only sets a new SOTA in the image generation benchmark but is also compatible with various generative models, including DALL-E 3. Also, SLD extends its utility to image editing applications, offering fine-grained object-level manipulation that surpasses existing methods. + +# References + +[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 6 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 6, 7 +[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373, 2023. 2, 6 +[4] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2023. 3, 5, 6, 7 +[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 3 +[6] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 1, 2, 6 +[7] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LLM blueprint: Enabling text-to-image generation with complex and detailed prompts. In The Twelfth International Conference on Learning Representations, 2024. 2 +[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 6 +[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5 +[10] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 6 +[11] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 1, 2, 5, 6 +[12] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 6 +[13] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2 +[14] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 6 + +[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3, 6 +[16] OpenAI. Dall-e 3 system card, 2023. 1, 6 +[17] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 3 +[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[19] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 8 +[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[22] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3 +[23] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 6 +[24] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2 +[25] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. +[26] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 + +[27] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 2 \ No newline at end of file diff --git a/2024/Self-correcting LLM-controlled Diffusion Models/images.zip b/2024/Self-correcting LLM-controlled Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b7a6cf0d92db39013fd4a8c3cee8072a3967bf8 --- /dev/null +++ b/2024/Self-correcting LLM-controlled Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c41606f0421f9709d42530a731dbd60c948da68277714b8ee7e73a52aefd4800 +size 808846 diff --git a/2024/Self-correcting LLM-controlled Diffusion Models/layout.json b/2024/Self-correcting LLM-controlled Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85ff33413721270a94fe4edab3de894dcfbafb04 --- /dev/null +++ b/2024/Self-correcting LLM-controlled Diffusion Models/layout.json @@ -0,0 +1,8780 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 103, + 449, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 103, + 449, + 120 + ], + "spans": [ + { + "bbox": [ + 144, + 103, + 449, + 120 + ], + "type": "text", + "content": "Self-correcting LLM-controlled Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 143, + 160, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 143, + 160, + 158 + ], + "spans": [ + { + "bbox": [ + 80, + 143, + 160, + 158 + ], + "type": "text", + "content": "Tsung-Han Wu*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 144, + 237, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 144, + 237, + 157 + ], + "spans": [ + { + "bbox": [ + 181, + 144, + 237, + 157 + ], + "type": "text", + "content": "Long Lian*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 144, + 354, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 144, + 354, + 172 + ], + "spans": [ + { + "bbox": [ + 257, + 144, + 354, + 172 + ], + "type": "text", + "content": "Joseph E. Gonzalez \nUC Berkeley" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 379, + 143, + 422, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 143, + 422, + 157 + ], + "spans": [ + { + "bbox": [ + 379, + 143, + 422, + 157 + ], + "type": "text", + "content": "Boyi Li†" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 443, + 144, + 517, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 144, + 517, + 156 + ], + "spans": [ + { + "bbox": [ + 443, + 144, + 517, + 156 + ], + "type": "text", + "content": "Trevor Darrell†" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 204, + 164, + 317 + ], + "blocks": [ + { + "bbox": [ + 88, + 194, + 125, + 203 + ], + "lines": [ + { + "bbox": [ + 88, + 194, + 125, + 203 + ], + "spans": [ + { + "bbox": [ + 88, + 194, + 125, + 203 + ], + "type": "text", + "content": "DALL-E3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 204, + 164, + 317 + ], + "lines": [ + { + "bbox": [ + 50, + 204, + 164, + 317 + ], + "spans": [ + { + "bbox": [ + 50, + 204, + 164, + 317 + ], + "type": "image", + "image_path": "b5134225e7f860ac442562a0b5ef10c0cc620556e3b7975b8e52780e41086f53.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 321, + 280, + 331 + ], + "lines": [ + { + "bbox": [ + 52, + 321, + 280, + 331 + ], + "spans": [ + { + "bbox": [ + 52, + 321, + 280, + 331 + ], + "type": "text", + "content": "A realistic cartoon-style image with a princess and four dwarfs" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 354, + 547, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 547, + 398 + ], + "type": "text", + "content": "Figure 1. Existing diffusion-based text-to-image generators (e.g., DALL-E 3 [16]) generally struggle to precisely generate images that correctly align with complex input prompts, especially for the ones that require numeracy and spatial relationships. Our Self-correcting LLM-controlled Diffusion (SLD) framework empowers these diffusion models to automatically and iteratively rectify inaccuracies through applying a set of latent space operations (addition, deletion, repositioning, etc.), resulting in enhanced text-to-image alignment." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 177, + 204, + 291, + 317 + ], + "blocks": [ + { + "bbox": [ + 118, + 180, + 212, + 191 + ], + "lines": [ + { + "bbox": [ + 118, + 180, + 212, + 191 + ], + "spans": [ + { + "bbox": [ + 118, + 180, + 212, + 191 + ], + "type": "text", + "content": "Numeracy Correction" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 194, + 277, + 204 + ], + "lines": [ + { + "bbox": [ + 175, + 194, + 277, + 204 + ], + "spans": [ + { + "bbox": [ + 175, + 194, + 277, + 204 + ], + "type": "text", + "content": "DALL-E 3 + SLD (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 204, + 291, + 317 + ], + "lines": [ + { + "bbox": [ + 177, + 204, + 291, + 317 + ], + "spans": [ + { + "bbox": [ + 177, + 204, + 291, + 317 + ], + "type": "image", + "image_path": "df5fa1e7fbafa99f44cf65d8f65d7f9505a6dfdd01c0b90bfcca41c1611c22c9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 302, + 204, + 416, + 317 + ], + "blocks": [ + { + "bbox": [ + 381, + 180, + 460, + 191 + ], + "lines": [ + { + "bbox": [ + 381, + 180, + 460, + 191 + ], + "spans": [ + { + "bbox": [ + 381, + 180, + 460, + 191 + ], + "type": "text", + "content": "Spatial Correction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 194, + 377, + 203 + ], + "lines": [ + { + "bbox": [ + 341, + 194, + 377, + 203 + ], + "spans": [ + { + "bbox": [ + 341, + 194, + 377, + 203 + ], + "type": "text", + "content": "DALL-E3" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 204, + 416, + 317 + ], + "lines": [ + { + "bbox": [ + 302, + 204, + 416, + 317 + ], + "spans": [ + { + "bbox": [ + 302, + 204, + 416, + 317 + ], + "type": "image", + "image_path": "1efcfb89b8bb173e5258d71b5e811377ddeb59708f3d3905bde442218b50d165.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 322, + 538, + 342 + ], + "lines": [ + { + "bbox": [ + 304, + 322, + 538, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 538, + 342 + ], + "type": "text", + "content": "A vivid photo with a woman on the right and a clown on the left walking in a dirty alley" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 430, + 204, + 543, + 317 + ], + "blocks": [ + { + "bbox": [ + 443, + 194, + 530, + 204 + ], + "lines": [ + { + "bbox": [ + 443, + 194, + 530, + 204 + ], + "spans": [ + { + "bbox": [ + 443, + 194, + 530, + 204 + ], + "type": "text", + "content": "DALL-E 3 + SLD (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 204, + 543, + 317 + ], + "lines": [ + { + "bbox": [ + 430, + 204, + 543, + 317 + ], + "spans": [ + { + "bbox": [ + 430, + 204, + 543, + 317 + ], + "type": "image", + "image_path": "e7bd7f0277e99d0cf48cb67ad61b9ce1a46e68da8113c0b21a7c0f38c56347ca.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 164, + 407, + 212, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 407, + 212, + 419 + ], + "spans": [ + { + "bbox": [ + 164, + 407, + 212, + 419 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 433, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 289, + 696 + ], + "type": "text", + "content": "Text-to-image generation has witnessed significant progress with the advent of diffusion models. Despite the ability to generate photorealistic images, current text-to-image diffusion models still often struggle to accurately interpret and follow complex input text prompts. In contrast to existing models that aim to generate images only with their best effort, we introduce Self-correcting LLM-controlled Diffusion (SLD). SLD is a framework that generates an image from the input prompt, assesses its alignment with the prompt, and performs self-corrections on the inaccuracies in the generated image. Steered by an LLM controller, SLD turns text-to-image generation into an iterative closed-loop process, ensuring correctness in the resulting image. SLD is not only training-free but can also be seamlessly integrated with diffusion models behind API access, such as DALL-E 3, to further boost the performance of state-of-the-art diffusion models. Experimental results show that our approach can rectify a majority of incorrect generations, particularly in generative numeracy, attribute binding, and spatial relationships. Furthermore, by simply adjusting the instructions to the LLM, SLD can perform image editing tasks, bridging the gap between text-to-image generation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 409, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 545, + 433 + ], + "type": "text", + "content": "and image editing pipelines. Our code is available at: https://self-correcting-llm-diffusion.github.io." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 439, + 387, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 387, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 387, + 451 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "content": "Text-to-image generation has made remarkable advancements, especially with the advent of diffusion models. However, these models often struggle with interpreting complex input text prompts, particularly those that require skills such as understanding the concept of numeracy, spatial relationships, and attribute binding with multiple objects. Despite the astonishing scaling of model sizes and training data, these challenges, as illustrated in Fig. 1, are still present in state-of-the-art open-source and proprietary diffusion models." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "content": "Several research and engineering efforts aim to overcome these limitations. For instance, methods such as DALL-E 3 [16] focus on the diffusion training process and incorporate high-quality captions into the training data at a massive scale. However, this approach not only incurs substantial costs but also frequently falls short in generating accurate images from complicated user prompts, as shown in Fig. 1. Other work harnesses the power of external models for a better understanding of the prompt in the inference process before the actual image generation. For example, [6, 11] leverages Large Language Models (LLMs) to pre" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 702, + 188, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 702, + 188, + 713 + ], + "spans": [ + { + "bbox": [ + 51, + 702, + 188, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Equal advising." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6327" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "process textual prompts into structured image layouts and thus ensures the preliminary design aligns with the user's directives. However, such integration does not resolve the inaccuracies produced by the downstream diffusion models, particularly in images with complex scenarios like multiple objects, cluttered arrangements, or detailed attributes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 287, + 350 + ], + "type": "text", + "content": "Drawing inspiration from the process of a human painting and a diffusion model in generating images, we observe a key distinction in their approach to creation. Consider a human artist tasked with painting a scene featuring two cats. Throughout the painting process, the artist remains cognizant of this requirement, ensuring that two cats are indeed present before considering the work complete. Should the artist find only one cat depicted, an additional one would be added to meet the prompt's criteria. This contrasts sharply with current text-to-image diffusion models, which operate on an open-loop basis. These models generate images through a predetermined number of diffusion steps and present the output to the user, regardless of its alignment with the initial user prompt. Such a process, irrespective of scaling training data or LLM pre-generation conditioning, lacks a robust mechanism to ensure the final image aligns with the user's expectations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 351, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 470 + ], + "type": "text", + "content": "In light of this, we propose our method Self-correcting LLM-controlled Diffusion (SLD) that performs self-checks to confidently offer users guarantees of the alignment between the prompt and the generated images. Departing from conventional single-round generation methods, SLD is a novel closed-loop approach that equips diffusion models with the ability to iteratively identify and rectify errors. Our SLD framework, illustrated in Fig. 2, contains two main components: LLM-driven object detection as well as LLM-controlled assessment and correction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 700 + ], + "type": "text", + "content": "The SLD pipeline follows a standard text-to-image generation setting. Given a textual prompt that outlines the desired image, SLD begins with calling an image generation module (e.g., the aforementioned open-loop text-to-image diffusion models) to generate an image in a best-effort fashion. Given that these open-loop generators do not guarantee an output that aligns perfectly with the prompt, SLD then conducts a thorough evaluation of the produced image against the prompt, with an LLM parsing key phrases for an open-vocabulary detector to check. Subsequently, an LLM controller takes the detected bounding boxes and the initial prompt as input and checks for potential mismatches between the detection results and the prompt requirements, suggesting appropriate self-correction operations, such as adding, moving, and removing objects. Finally, utilizing a chosen base diffusion model (e.g., Stable Diffusion [20]), SLD employs latent space composition to implement these adjustments, thereby ensuring that the final image accurately reflects the user's initial text prompt." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Notably, SLD is agnostic to the initial generation pro" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "cess, which allows us to use DALL-E 3 APIs as a subroutine to generate initial images and then employ the open-source Stable Diffusion model family to conduct latent operations to attain self-correction. Furthermore, none of these operations require additional training on our base diffusion model [20], which easily allows our method to be applied to various diffusion models without external human annotation or training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "text", + "content": "We demonstrate that our SLD framework can achieve significant improvement over current diffusion-based methods on complex prompts with the LMD benchmark [11]. The results show that our method can surpass " + }, + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{LMD + }" + }, + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "text", + "content": ", which is a strong baseline that already leverages LLM in the image process generation, by " + }, + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "inline_equation", + "content": "9.0\\%" + }, + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "text", + "content": ". More importantly, with DALL-E 3 for initial generation, the generated images from our method achieve " + }, + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "inline_equation", + "content": "26.5\\%" + }, + { + "bbox": [ + 304, + 168, + 545, + 275 + ], + "type": "text", + "content": " performance gains compared to ones before self-correction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 275, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 361 + ], + "type": "text", + "content": "Finally, since the SLD pipeline is agnostic to the initially generated image, it can easily be transformed into an image editing pipeline by simply changing the prompts to the LLM. While text-to-image generation and image editing are often treated as distinct tasks by the generative modeling community, our SLD can perform these two tasks with a unified pipeline. We list our key contributions below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 361, + 545, + 482 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 361, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 397 + ], + "type": "text", + "content": "1. SLD is the first to integrate a detector and an LLM to self-correct generative models, ensuring accurate generation without extra training or external data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 397, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 545, + 445 + ], + "type": "text", + "content": "2. SLD offers a unified solution for both image generation and editing, enabling enhanced text-to-image alignment for any image generator (e.g., DALL-E 3) and object-level editing on any images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 445, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 545, + 482 + ], + "type": "text", + "content": "3. Experimental results show that SLD can correct a majority of incorrect generations, particularly in aspects of numeracy, attribute binding, and spatial relationships." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 495, + 392, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 392, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 392, + 507 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 515, + 477, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 477, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 477, + 528 + ], + "type": "text", + "content": "2.1. Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "content": "Diffusion-based text-to-image generation has advanced significantly. Initial studies [19-21] showed diffusion models' ability to create high-quality images, but they struggle with complex prompts. Subsequent research [3, 10, 18, 24-26] has incorporated additional inputs such as keypoints and bounding boxes to control the diffusion generation process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "Recent advancements have incorporated LLMs to control the generation of diffusion models, bypassing the need for additional complementary information as inputs [6, 7, 11-13, 27]. In these approaches, LLMs play a central role in directly interpreting user textual prompts and managing the initial layout configuration. Despite some progress, these models often operate in an open-loop fashion, producing images in one iteration that cannot guarantee the generated images align with user prompts." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6328" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 515, + 350 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 515, + 350 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 515, + 350 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 515, + 350 + ], + "type": "image", + "image_path": "c6ef2e8b092fa1cf62a5be3dfea74fe4f09fda2ad776123a781fe294cdb27ee7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 547, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 547, + 389 + ], + "type": "text", + "content": "Figure 2. Our proposed Self-correcting LLM-controlled Diffusion (SLD) enhances text-to-image alignment through an iterative self-correction process. It begins with LLM-driven object detection (Sec. 3.1), and subsequently performs LLM-controlled analysis and correction (Sec. 3.2). The entire pipeline is outlined in Algorithm 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 400, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 545 + ], + "type": "text", + "content": "Unlike prior work, SLD is the first closed-loop diffusion-based generation method. Integrating advanced object detectors and LLMs, SLD performs iterative self-checking and correction, significantly enhancing text-to-image alignment. This improvement spans numeracy, attribute binding to multiple objects, and spatial reasoning, applicable to various models, including those like DALL-E 3. Also, SLD extends beyond existing NLP self-correction techniques [17] by incorporating bounding box spatial representation and novel latent editing mechanisms tailored for text-to-image tasks. These designs are essential for ensuring high-quality, precise self-correction in image generation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 214, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 214, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 214, + 565 + ], + "type": "text", + "content": "2.2. Diffusion-based Image Editing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Recent advancements in text-to-image diffusion models have significantly expanded their applications in image editing, encompassing both global and local editing tasks. Techniques like Prompt-2-prompt [8] and InstructPix2Pix [2] specialize in global edits, such as style transformations. Conversely, methods like SDEdit [15], DiffEdit [4], and Plug-and-Play [22] focus on local edits, targeting specific areas within images. Despite their progress, these methods often struggle with precise object-level manipulation and tasks that require spatial reasoning, such as resizing or repositioning objects. While recent approaches like Self-Guidance [5] offer fine-grained operations, they still necessities to be able to use the same techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 400, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 545, + 425 + ], + "type": "text", + "content": "sitate user inputs for specific coordinates when moving or repositioning objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 425, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 546, + 510 + ], + "type": "text", + "content": "Unlike these methods only focus on diffusion models, SLD introduces the combination of detectors and LLMs in the loop of editing, enabling fine-grained editing with only user prompts. Also, SLD excels in various object-level editing tasks, including adding, replacing, moving, and modifying attributes, swapping, and so on, demonstrating a notable improvement in both ease of use and editing capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 521, + 533, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 533, + 534 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 533, + 534 + ], + "type": "text", + "content": "3. Self-correcting LLM-controlled Diffusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 541, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 650 + ], + "type": "text", + "content": "In this section, we introduce our Self-correcting LLM-controlled Diffusion (SLD) framework. SLD consists of two main components: LLM-driven object detection (Sec. 3.1) as well as LLM-controlled assessment and correction (Sec. 3.2). Moreover, with a simple change of the LLM instructions, we show that SLD is applicable to image editing, unifying text-to-image generation, and editing as discussed in Sec. 3.3. The complete pipeline is shown in Algorithm 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 658, + 470, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 470, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 470, + 671 + ], + "type": "text", + "content": "3.1. LLM-driven Object Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "Our SLD framework starts with LLM-driven object detection, which extracts information necessary for downstream assessment and correction. As shown with green arrows in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6329" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 231, + 241 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 231, + 241 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 231, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 231, + 241 + ], + "type": "image", + "image_path": "4ef7b66afd1c8e32ee938f07acbf4cec29c40bfd4b0aa5f235184b10e95678a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "text", + "content": "Figure 3. Our self-correction pipeline is driven by two distinct LLMs: (a) The LLM parser analyzes user prompts " + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "text", + "content": " to extract a list of key object information " + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "text", + "content": ", which is then passed to the open-vocabulary detector. (b) The LLM controller, taking both the user prompt " + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "text", + "content": " and currently detected bounding boxes " + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "inline_equation", + "content": "B_{curr}" + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "text", + "content": " as input, outputs suggested new bounding boxes " + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "inline_equation", + "content": "B_{next}" + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "text", + "content": ". These are subsequently transformed into a set of latent space operations " + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "inline_equation", + "content": "Ops" + }, + { + "bbox": [ + 46, + 246, + 547, + 291 + ], + "type": "text", + "content": " for image manipulation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 70, + 439, + 241 + ], + "blocks": [ + { + "bbox": [ + 238, + 70, + 439, + 241 + ], + "lines": [ + { + "bbox": [ + 238, + 70, + 439, + 241 + ], + "spans": [ + { + "bbox": [ + 238, + 70, + 439, + 241 + ], + "type": "image", + "image_path": "19ecd4193069023ec26d3a6a51f7c80079e6560e6bbd68d5a24664a8d63fab59.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 441, + 70, + 545, + 241 + ], + "blocks": [ + { + "bbox": [ + 441, + 70, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 441, + 70, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 441, + 70, + 545, + 241 + ], + "type": "image", + "image_path": "4fdbe53657b21fc76c661e3adc40c354300523c8f2d83e0041b4876177d3c0b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "content": "Fig. 2, the LLM-driven object detection includes two steps: 1) We leverage an LLM as a parser that parses the user prompt and outputs key phrases that are potentially relevant to image assessment. 2) The phrases are then passed into an open-vocabulary object detector. The detected boxes are supposed to contain information that supports the assessment of whether the image is aligned with the specifications in the user prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 288, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 557 + ], + "type": "text", + "content": "In the initial step, an LLM parser is directed to extract a list of key object details, denoted as " + }, + { + "bbox": [ + 46, + 402, + 288, + 557 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 402, + 288, + 557 + ], + "type": "text", + "content": ", from the user-provided text prompt " + }, + { + "bbox": [ + 46, + 402, + 288, + 557 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 402, + 288, + 557 + ], + "type": "text", + "content": ". This parser, aided by text instructions and in-context examples, can easily accomplish this as shown in Fig. 3 (a). For a user prompt that includes phrases like \"a green motorcycle\" and \"a blue motorcycle,\" the LLM is expected to identify and output \"green\" and \"blue\" as attributes associated with \"motorcycle.\" When the prompt references objects without specific quantities or attributes, such as \"a monkey\" and \"a raccoon,\" these descriptors are appropriately left blank. Importantly, the LLM's role is not limited to merely identifying object nouns; it also entails identifying any associated quantities or attributes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "In the second step, an open-vocabulary detector processes the list of key object information, " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", parsed in the first step, to detect and localize objects within the image. We prompt the open-vocabulary object detector with queries formatted as image of a/an [attribute] [object name], where the \"attribute\" and \"object name\" are sourced from the parser's output. The resulting bounding boxes, " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "B_{curr}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", are then organized into a list format like [\"[attribute] [object name] [#object ID]\", [x, y, w, h])] for further processing. A special case is when the prompt poses constraints on the object quantity. For cases where attributed objects (e.g., \"blue dog\") fall short compared to the required" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 305, + 547, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 547, + 377 + ], + "type": "text", + "content": "quantities, a supplementary count of non-attributed objects (e.g., \"dog\") is provided to provide context for the subsequent LLM controller deciding whether to add more \"blue dogs\" or simply alter the color of existing dogs to blue. We will explain these operations, including object addition and attribute modification, in greater detail in Sec. 3.2.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 392, + 522, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 522, + 404 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 522, + 404 + ], + "type": "text", + "content": "3.2. LLM-controlled Analysis and Correction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "text", + "content": "We use an LLM controller for image analysis and the subsequent correction. The controller, given the user prompt " + }, + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "text", + "content": " and detected boxes " + }, + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "B_{curr}" + }, + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "text", + "content": ", is asked to analyze whether the image, represented by objects bounding boxes, aligns with the description of the user prompt and offers a list of corrected bounding boxes " + }, + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "B_{next}" + }, + { + "bbox": [ + 304, + 412, + 547, + 484 + ], + "type": "text", + "content": ", as shown in Fig. 3 (b)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "SLD then programmatically analyzes the inconsistencies between the refined and original bounding boxes to output a set of editing operations " + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "inline_equation", + "content": "Ops" + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": ", which includes addition, deletion, repositioning, and attribute modification. However, a simple set-of-boxes representation does not carry correspondence information, which does not allow an easy way to compare the input and the output layout of the LLM controller when multiple boxes share the same object name. For example, when there are two cat boxes in both the model input and the model output, whether one cat box corresponds to which cat box in the output layout is unclear. Rather than introducing another algorithm to guess the correspondence, we propose to let the LLM output correspondence with a very simple edit: we give an object ID to each bounding box, with the number increasing within each object type, as a suffix added after the object name. In the in-context examples, we demonstrate to the LLM that the object should have the same name and object ID before and after the proposed correction." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6330" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 53, + 112, + 258, + 255 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 254, + 83 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 254, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 254, + 83 + ], + "type": "text", + "content": "Algorithm 1 Self-correction for Image Generation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "lines": [ + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "spans": [ + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "type": "text", + "content": "Input: User prompt " + }, + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "type": "text", + "content": ", Initial generated image " + }, + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "type": "text", + "content": ", Maximum number of self-correction round " + }, + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 49, + 87, + 286, + 110 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "lines": [ + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "k\\gets 1" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "S\\gets \\mathsf{LLM - Parse}(P)" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "B_{curr}\\gets \\mathsf{Detector}(S)" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "B_{next}\\gets \\mathsf{LLM - Analysis}(P,B_{curr})" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "Ops\\gets \\mathsf{Diff}(B_{curr},B_{next})" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " \n6: if " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "Ops\\neq \\emptyset" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " then (i.e., " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "B_{next}\\neq B_{curr}" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": ") \n7: " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "I = \\mathsf{Correction}(Ops,B_{next},B_{curr})" + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "text", + "content": " \n8: else \n9: break \n10: end if \n11: end for \nOutput: Image " + }, + { + "bbox": [ + 53, + 112, + 258, + 255 + ], + "type": "inline_equation", + "content": "I" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 277, + 286, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 286, + 300 + ], + "type": "text", + "content": "3.2.1 Latent Operations for Training-Free Image Correction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 415 + ], + "type": "text", + "content": "The LLM controller outputs a list of correction operations to apply. For each operation, we first transform the original image into latent features. Our approach then executes a series of operations " + }, + { + "bbox": [ + 46, + 308, + 287, + 415 + ], + "type": "inline_equation", + "content": "Ops" + }, + { + "bbox": [ + 46, + 308, + 287, + 415 + ], + "type": "text", + "content": ", such as addition, deletion, repositioning, and attribute modification, applied to these latent layers. It is worth noting that although there are only four operations, they are sufficient to handle a majority of misalignments naturally. We explain how each operation is performed below." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 426, + 284, + 610 + ], + "blocks": [ + { + "bbox": [ + 50, + 426, + 284, + 610 + ], + "lines": [ + { + "bbox": [ + 50, + 426, + 284, + 610 + ], + "spans": [ + { + "bbox": [ + 50, + 426, + 284, + 610 + ], + "type": "image", + "image_path": "d8ea5419e73e8ec93f27cba35120f3e932f85bab62ba6e1918b4065f17a72568.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 616, + 287, + 693 + ], + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 693 + ], + "type": "text", + "content": "Figure 4. Our latent operations can be summarized into two key concepts: (1) latent in removed regions are re-initialized to Gaussian noise, and latent of newly added or modified objects are composed onto the canvas. (2) Latent composition is confined to the initial steps, followed by \"unfrozen\" steps for a standard forward diffusion process, enhancing visual quality and avoiding artificial copy-and-paste effects." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Addition. Inspired by [11], the addition process entails two" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "phases: pre-generating an object and integrating its latent representation into the original image's latent space. Initially, we use a diffusion model to create an object within a designated bounding box, followed by precise segmentation using models (e.g., SAM [9]). This object is then processed through a backward diffusion sequence with our base diffusion model, yielding masked latent layers corresponding to the object, which are later merged with the original canvas." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 169, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 228 + ], + "type": "text", + "content": "Deletion operation begins with SAM refining the boundary of the object within its bounding box. The latent layers associated with these specified regions are then removed and reset with Gaussian noise. This necessitates a complete regeneration of these areas in the final denoising process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 229, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 546, + 384 + ], + "type": "text", + "content": "Repositioning involves modifying the original image to align objects with new bounding boxes, taking care to preserve their original aspect ratios. The initial steps include shifting and resizing the bounding box in the image space. Following this, SAM refines the object boundary, succeeded by a backward diffusion process to generate its relevant latent layers, similar to the approach in the addition operation. Latent layers corresponding to the excised parts are replaced with Gaussian noise, while the newly added sections are integrated into the final image composition. An important consideration in repositioning is conducting object resizing in the image space rather than the latent space to maintain high-quality results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 386, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 445 + ], + "type": "text", + "content": "Attribute modification starts with SAM refining the object boundary within the bounding box, followed by applying attribute modifications such as DiffEdit [4]. The base diffusion model then reverses the image, producing a series of masked latent layers ready for final composition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 446, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 529 + ], + "type": "text", + "content": "After editing operations on each object, we proceed to the recomposition phase as shown in Fig. 4. In this phase, while latents for removed or repositioned regions are reinitialized with Gaussian noise, the latents for added or modified latents are updated accordingly. For regions with multiple overlapping objects, we place the larger masks first to ensure the visibility of the smaller objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 531, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 638 + ], + "type": "text", + "content": "The stack of modified latent then undergoes a final forward diffusion process, which begins with steps in which regions not reinitialized with Gaussian noise are frozen (i.e., forced to align with the unmodified latent at the same step). This is crucial for the accurate formation of updated objects while maintaining background consistency at the same time. The procedure finishes with several steps where everything is allowed to change, resulting in a visually coherent and correct image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 657, + 521, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 521, + 668 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 521, + 668 + ], + "type": "text", + "content": "3.2.2 Termination of the Self-Correction Process" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "While we observe that one round of generation is often enough for a majority of the cases, subsequent rounds could still benefit the performance in terms of correctness further," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6331" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 70, + 492, + 218 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 492, + 218 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 492, + 218 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 492, + 218 + ], + "type": "table", + "html": "
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
MultiDiffusion [1]29%28%26%39%30.5%
Backward Guidance [3]22%37%26%67%38.0%
BoxDiff [23]22%30%37%71%40.0%
LayoutGPT + GLIGEN [6, 10]36%65%26%78%51.3%
DALL-E 3 [16]25%38%74%71%52.0%
+ 1-round SLD (Ours)90%61%80%83%78.5% (+26.5)
LMD+ [11]100%82%49%86%79.3%
+ 1-round SLD (Ours)100%98%63%92%88.3% (+9.0)
", + "image_path": "f1dff01ee196d8237d17bf39ae8eae6d6103e1476fdf9c490a71a1f8b24fb58a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 226, + 531, + 237 + ], + "lines": [ + { + "bbox": [ + 61, + 226, + 531, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 531, + 237 + ], + "type": "text", + "content": "Table 1. Our method can be applied to various image generation methods and improves the generation accuracy by a large margin." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 252, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 286, + 323 + ], + "type": "text", + "content": "making our self-correction an iterative process. Thus, determining the optimal number of self-correction rounds is critical for balancing efficiency and accuracy. As outlined in Algorithm 1, our method sets a maximum number of attempts on the correction rounds to ensure the process finishes within a reasonable amount of time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 286, + 445 + ], + "type": "text", + "content": "The process completes when the LLM outputs the same layout as the input (i.e., if the bounding boxes suggested by the LLM controller " + }, + { + "bbox": [ + 46, + 325, + 286, + 445 + ], + "type": "inline_equation", + "content": "(B_{next})" + }, + { + "bbox": [ + 46, + 325, + 286, + 445 + ], + "type": "text", + "content": " align with the current detected bounding boxes " + }, + { + "bbox": [ + 46, + 325, + 286, + 445 + ], + "type": "inline_equation", + "content": "(B_{curr}))" + }, + { + "bbox": [ + 46, + 325, + 286, + 445 + ], + "type": "text", + "content": ", or when the maximum rounds of generation are reached, indicating that the method is unable to make a correct generation for the prompt. This iterative process provides guarantees on the correctness of the image, up to the accuracy of the detector and the LLM controller, ensuring it aligns closely with the initial text prompt. We explore the efficacy of multi-round corrections in Sec. 4.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 277, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 277, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 277, + 467 + ], + "type": "text", + "content": "3.3. Unified text-to-image generation and editing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 286, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 286, + 592 + ], + "type": "text", + "content": "In addition to self-correcting image generation models, our SLD framework is readily adaptable for image editing applications, requiring only minimal modifications. A key distinction is in the format of user input prompts. Unlike image generation, where users provide scene descriptions, image editing requires users to detail both the original image and the desired changes. For instance, to edit an image with two apples and a banana by replacing the banana with an orange, the prompt could be: \"Replace the banana with an orange, while keeping the two apples unchanged.\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 286, + 700 + ], + "type": "text", + "content": "The editing process is similar to our self-correction mechanism. The LLM parser extracts key objects from the user's prompt. These objects are then identified by the open-vocabulary detector, establishing a list of current bounding boxes. The editing-focused LLM controller, equipped with specific task objectives, guidelines, and in-context examples, analyzes these inputs. It proposes updated bounding boxes and corresponding latent space operations for precise image manipulation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "SLD's ability to perform detailed, object-level editing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 252, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 361 + ], + "type": "text", + "content": "distinguishes it from existing diffusion-based methods like InstructPix2Pix [2] and prompt2prompt [8], which mainly address global image style changes. Also, SLD outperforms tools like DiffEdit [4] and SDEdit [15], which are restricted to object replacement or attribute adjustment, by enabling comprehensive object repositioning, addition, and deletion with exact control. Our comparative analysis in Sec. 4.2 will further highlight SLD's superior editing capabilities over existing methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 373, + 387, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 387, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 387, + 387 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 394, + 541, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 541, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 541, + 407 + ], + "type": "text", + "content": "4.1. Comparison with Image Generation Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 628 + ], + "type": "text", + "content": "Setup. We evaluate the performance of the SLD framework with the LMD benchmark [11], which is specifically designed to evaluate generation methods on complex tasks such as handling negation, numeracy, accurate attribute binding to multiple objects, and spatial reasoning. For each task, 100 programmatically generated prompts are fed into various text-to-image generation methods to produce corresponding images. We evaluate the images generated by our method and the baselines with open-vocabulary detector OWL-ViT v2 [14] for a robust quantitative evaluation of the alignment between the input prompts and the generated images. We compared SLD with several leading text-to-image diffusion methods, such as Multidiffusion [1], BoxDiff [23], LayoutGPT [6], LMD+ [12], and DALL-E 3 [16]. To ensure fair comparisons, all models incorporating LLMs used the same GPT-4 model. In our SLD implementation, we utilized LMD+ as the base model for latent space operations and OWL-ViT v2 for the open-vocabulary object detector." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Results. As shown in Tab. 1, applying the SLD method to both open-source " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{LMD}+)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and proprietary models (DALL-E 3) significantly enhances their performance in terms of generation correctness. For negation tasks, as " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LMD}+" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " converts user prompts containing \"without\" information into negative prompts, which already achieves a remarkable " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " accuracy without SLD integration. In contrast, even" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6332" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 92, + 128, + 171 + ], + "blocks": [ + { + "bbox": [ + 49, + 92, + 128, + 171 + ], + "lines": [ + { + "bbox": [ + 49, + 92, + 128, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 92, + 128, + 171 + ], + "type": "image", + "image_path": "a307dbf42a70969e49a48b6001524db16204d2c41dec9c66fbb9ad461604a612.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 172, + 97, + 180 + ], + "lines": [ + { + "bbox": [ + 79, + 172, + 97, + 180 + ], + "spans": [ + { + "bbox": [ + 79, + 172, + 97, + 180 + ], + "type": "text", + "content": "SDXL" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 132, + 92, + 211, + 171 + ], + "blocks": [ + { + "bbox": [ + 132, + 92, + 211, + 171 + ], + "lines": [ + { + "bbox": [ + 132, + 92, + 211, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 92, + 211, + 171 + ], + "type": "image", + "image_path": "00fc34a4c7af14058bf1a7722fd49a17c4db0b582deb310bcd4ce9fb88899df8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 172, + 202, + 181 + ], + "lines": [ + { + "bbox": [ + 137, + 172, + 202, + 181 + ], + "spans": [ + { + "bbox": [ + 137, + 172, + 202, + 181 + ], + "type": "text", + "content": "SDXL w/ SLD (Ours)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 215, + 92, + 294, + 171 + ], + "blocks": [ + { + "bbox": [ + 215, + 92, + 294, + 171 + ], + "lines": [ + { + "bbox": [ + 215, + 92, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 215, + 92, + 294, + 171 + ], + "type": "image", + "image_path": "5bb12932fe4e277153d47699dd97a98608cf6416fadd3164c73238cea0f8f6d9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 172, + 266, + 180 + ], + "lines": [ + { + "bbox": [ + 246, + 172, + 266, + 180 + ], + "spans": [ + { + "bbox": [ + 246, + 172, + 266, + 180 + ], + "type": "text", + "content": "LMD+" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 299, + 92, + 379, + 171 + ], + "blocks": [ + { + "bbox": [ + 299, + 92, + 379, + 171 + ], + "lines": [ + { + "bbox": [ + 299, + 92, + 379, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 92, + 379, + 171 + ], + "type": "image", + "image_path": "49f6ff7fa7e479b56d53bc2e478e1c0fc720e4a34b0afa88a9dae0c0d1e60014.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 172, + 372, + 181 + ], + "lines": [ + { + "bbox": [ + 306, + 172, + 372, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 172, + 372, + 181 + ], + "type": "text", + "content": "LMD+ w/ SLD (Ours)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 383, + 92, + 462, + 171 + ], + "blocks": [ + { + "bbox": [ + 383, + 92, + 462, + 171 + ], + "lines": [ + { + "bbox": [ + 383, + 92, + 462, + 171 + ], + "spans": [ + { + "bbox": [ + 383, + 92, + 462, + 171 + ], + "type": "image", + "image_path": "7ca07c4e9fbae07c9d83e64a0d2f1296b633b59fa8b67998138a99d2e7f5d3e6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 172, + 438, + 180 + ], + "lines": [ + { + "bbox": [ + 407, + 172, + 438, + 180 + ], + "spans": [ + { + "bbox": [ + 407, + 172, + 438, + 180 + ], + "type": "text", + "content": "DALL-E3" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 466, + 92, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 466, + 92, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 466, + 92, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 466, + 92, + 545, + 171 + ], + "type": "image", + "image_path": "abfd065bb7574402eba14755d9d2dc6ae24e047486aabbc63a219ad46b10092c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 172, + 543, + 181 + ], + "lines": [ + { + "bbox": [ + 466, + 172, + 543, + 181 + ], + "spans": [ + { + "bbox": [ + 466, + 172, + 543, + 181 + ], + "type": "text", + "content": "DALL-E 3 w/ SLD (Ours)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 190, + 544, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 190, + 544, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 190, + 544, + 201 + ], + "type": "text", + "content": "A realistic photo with a monkey sitting above a green motorcycle on the left and another raccoon sitting above a blue motorcycle on the right" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 49, + 203, + 128, + 282 + ], + "blocks": [ + { + "bbox": [ + 49, + 203, + 128, + 282 + ], + "lines": [ + { + "bbox": [ + 49, + 203, + 128, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 203, + 128, + 282 + ], + "type": "image", + "image_path": "fc008f569b3f5b259f3f45264c6c20c4623ab6f0f55e0b5472bd50007d10e76a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 285, + 99, + 293 + ], + "lines": [ + { + "bbox": [ + 80, + 285, + 99, + 293 + ], + "spans": [ + { + "bbox": [ + 80, + 285, + 99, + 293 + ], + "type": "text", + "content": "SDXL" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 132, + 203, + 211, + 282 + ], + "blocks": [ + { + "bbox": [ + 132, + 203, + 211, + 282 + ], + "lines": [ + { + "bbox": [ + 132, + 203, + 211, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 211, + 282 + ], + "type": "image", + "image_path": "848e9760730f9b5c157b58e581e601b2d71e9aa89248623e39338c51c9f4a1aa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 285, + 202, + 293 + ], + "lines": [ + { + "bbox": [ + 137, + 285, + 202, + 293 + ], + "spans": [ + { + "bbox": [ + 137, + 285, + 202, + 293 + ], + "type": "text", + "content": "SDXL w/ SLD (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 216, + 203, + 294, + 282 + ], + "blocks": [ + { + "bbox": [ + 216, + 203, + 294, + 282 + ], + "lines": [ + { + "bbox": [ + 216, + 203, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 216, + 203, + 294, + 282 + ], + "type": "image", + "image_path": "bd39125a3f2b1c74e7a4d54713b29da87759770cd3bcb522e197a217fa1d8935.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 285, + 266, + 293 + ], + "lines": [ + { + "bbox": [ + 246, + 285, + 266, + 293 + ], + "spans": [ + { + "bbox": [ + 246, + 285, + 266, + 293 + ], + "type": "text", + "content": "LMD+" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 299, + 203, + 378, + 282 + ], + "blocks": [ + { + "bbox": [ + 299, + 203, + 378, + 282 + ], + "lines": [ + { + "bbox": [ + 299, + 203, + 378, + 282 + ], + "spans": [ + { + "bbox": [ + 299, + 203, + 378, + 282 + ], + "type": "image", + "image_path": "de80f2785db0c6313e30f13160fa251ea7158a19e0517a14096f9a2f073c8084.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 285, + 372, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 285, + 372, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 372, + 293 + ], + "type": "text", + "content": "LMD+ w/ SLD (Ours)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 383, + 203, + 462, + 282 + ], + "blocks": [ + { + "bbox": [ + 383, + 203, + 462, + 282 + ], + "lines": [ + { + "bbox": [ + 383, + 203, + 462, + 282 + ], + "spans": [ + { + "bbox": [ + 383, + 203, + 462, + 282 + ], + "type": "image", + "image_path": "9bf35ba779a91127c0b862ddfc7217808eae8325c3ea3226d126e65b4836ae48.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 285, + 438, + 293 + ], + "lines": [ + { + "bbox": [ + 407, + 285, + 438, + 293 + ], + "spans": [ + { + "bbox": [ + 407, + 285, + 438, + 293 + ], + "type": "text", + "content": "DALL-E3" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 466, + 203, + 545, + 282 + ], + "blocks": [ + { + "bbox": [ + 466, + 203, + 545, + 282 + ], + "lines": [ + { + "bbox": [ + 466, + 203, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 466, + 203, + 545, + 282 + ], + "type": "image", + "image_path": "59ae716dc097a9b7eceff7bca8ae16282faf8daa5fb4fc027c91f7b91d48c63a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 285, + 543, + 293 + ], + "lines": [ + { + "bbox": [ + 466, + 285, + 543, + 293 + ], + "spans": [ + { + "bbox": [ + 466, + 285, + 543, + 293 + ], + "type": "text", + "content": "DALL-E 3 w/ SLD (Ours)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": "though DALL-E 3 also uses an LLM to rewrite the prompt, it still fails for some negation cases, likely because the LLM simply puts the negation keyword (e.g., \"without\") into the rewritten prompt. In this case, our SLD method can automatically rectify most of these errors. For numeracy tasks, integrating SLD with " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{LMD + }" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": " results in a significant improvement, with up to " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": " accuracy. We noted that DALL-E 3 often struggles to generate an image with the correct number of objects. However, this issue is substantially mitigated by SLD, which enhances performance by over " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": ". For attribute binding tasks, SLD improves the performance of both DALL-E 3 and " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{LMD + }" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": ", respectively. Notably, DALL-E 3 initially outperforms " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{LMD + }" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": " in this task, likely due to its training on high-quality image caption datasets. Finally, for spatial reasoning tasks, the integration of SLD with both " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{LMD + }" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": " and DALL-E 3 demonstrates enhanced performance by " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 581 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 587, + 205, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 205, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 205, + 601 + ], + "type": "text", + "content": "4.2. Application to Image Editing" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "As discussed in Sec. 3.3, SLD excels in fine-grained image editing over existing methods. As demonstrated in Fig. 6, our integration of an open-language detector with LLMs enables precise modifications within localized latent space regions. SLD adeptly performs specific edits, like seamlessly replacing an apple with a pumpkin, while preserving the integrity of surrounding objects. In contrast, methods like InstructPix2Pix [2] are confined to global transformations, and DiffEdit [4] often fails to accurately locate objects for" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 463, + 356, + 533, + 424 + ], + "blocks": [ + { + "bbox": [ + 46, + 300, + 546, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 546, + 344 + ], + "type": "text", + "content": "Figure 5. SLD enhances text-to-image alignment across diverse diffusion-based generative models such as SDXL, LMD+, and DALL-E 3. Notably, as highlighted by the red boxes in the first row, SLD precisely positions a blue bicycle between a bench and a palm tree, while maintaining the accurate count of palm trees and seagulls. The second row further demonstrates SLD's robustness in complex, cluttered scenes, effectively managing object collision through our training-free latent operations." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 356, + 533, + 424 + ], + "lines": [ + { + "bbox": [ + 463, + 356, + 533, + 424 + ], + "spans": [ + { + "bbox": [ + 463, + 356, + 533, + 424 + ], + "type": "image", + "image_path": "e132a16ef20d871051607858e51ed72b624f0a8495d5598edddc3d7172a56016.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 318, + 407, + 333, + 422 + ], + "blocks": [ + { + "bbox": [ + 338, + 376, + 457, + 401 + ], + "lines": [ + { + "bbox": [ + 338, + 376, + 457, + 401 + ], + "spans": [ + { + "bbox": [ + 338, + 376, + 457, + 401 + ], + "type": "text", + "content": "Replace the left apple with a pumpkin and keep the right apple and all bananas in the scene unchanged." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 407, + 333, + 422 + ], + "lines": [ + { + "bbox": [ + 318, + 407, + 333, + 422 + ], + "spans": [ + { + "bbox": [ + 318, + 407, + 333, + 422 + ], + "type": "image", + "image_path": "eac1e7e74054f9eb76ff33b109371fda001ae8dd74ef83d370f49350c3af9c94.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 318, + 429, + 386, + 497 + ], + "blocks": [ + { + "bbox": [ + 357, + 358, + 434, + 367 + ], + "lines": [ + { + "bbox": [ + 357, + 358, + 434, + 367 + ], + "spans": [ + { + "bbox": [ + 357, + 358, + 434, + 367 + ], + "type": "text", + "content": "Editing Instructions for SLD" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 429, + 386, + 497 + ], + "lines": [ + { + "bbox": [ + 318, + 429, + 386, + 497 + ], + "spans": [ + { + "bbox": [ + 318, + 429, + 386, + 497 + ], + "type": "image", + "image_path": "6af6512aaefc4900907eca27f063f9bb02b2808b2c67c1030e9fb9efe93c4a0b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 498, + 377, + 505 + ], + "lines": [ + { + "bbox": [ + 329, + 498, + 377, + 505 + ], + "spans": [ + { + "bbox": [ + 329, + 498, + 377, + 505 + ], + "type": "text", + "content": "InstructPix2Pix" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 510, + 545, + 565 + ], + "lines": [ + { + "bbox": [ + 306, + 510, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 545, + 565 + ], + "type": "text", + "content": "Figure 6. When instructed to perform object-level image editing. InstructPix2Pix [2] completely fails to accomplish the task, and DiffEdit [4] falls short, as highlighted in the green box of the image. Conversely, our method demonstrates a significantly better performance in executing these object-level edits." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 391, + 429, + 459, + 497 + ], + "blocks": [ + { + "bbox": [ + 391, + 429, + 459, + 497 + ], + "lines": [ + { + "bbox": [ + 391, + 429, + 459, + 497 + ], + "spans": [ + { + "bbox": [ + 391, + 429, + 459, + 497 + ], + "type": "image", + "image_path": "ba6a05105eb50bd4d5449e5beca659645801228aa972b74b87e24b685ffa18c5.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 498, + 436, + 505 + ], + "lines": [ + { + "bbox": [ + 413, + 498, + 436, + 505 + ], + "spans": [ + { + "bbox": [ + 413, + 498, + 436, + 505 + ], + "type": "text", + "content": "DiffEdit" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 463, + 429, + 532, + 497 + ], + "blocks": [ + { + "bbox": [ + 463, + 429, + 532, + 497 + ], + "lines": [ + { + "bbox": [ + 463, + 429, + 532, + 497 + ], + "spans": [ + { + "bbox": [ + 463, + 429, + 532, + 497 + ], + "type": "image", + "image_path": "dcb7998c6696b4de99e8a1726bb07bfc50058240280ea802408ff1d6ddd1c9ad.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 498, + 516, + 505 + ], + "lines": [ + { + "bbox": [ + 479, + 498, + 516, + 505 + ], + "spans": [ + { + "bbox": [ + 479, + 498, + 516, + 505 + ], + "type": "text", + "content": "SLD (Ours)" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 306, + 582, + 475, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 475, + 593 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 475, + 593 + ], + "type": "text", + "content": "modification, leading to undesired results." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Furthermore, as exemplified in Fig. 7, SLD supports a wide array of editing instructions, including counting control (such as adding, deleting, or replacing objects), attribute modification (like altering colors or materials), and intricate location control (encompassing object swapping, resizing, and moving). A standout example is featured in the \"Object Resize\" column of the first row, where SLD precisely enlarges the cup on the table by an exact factor of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.25 \\times" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". We encourage readers to verify this with a ruler for a clear demonstration of our method's precision. This level of pre" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 75, + 71, + 522, + 87 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 71, + 522, + 87 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 522, + 87 + ], + "type": "text", + "content": "An oil painting at the beach of a blue bicycle to the left of a bench and to the right of a palm tree with three seagulls in the sky" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6333" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 128, + 159 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 128, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 128, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 128, + 159 + ], + "type": "image", + "image_path": "325b8eecc53f1075be21ccb45648ead52c39545d53058ec27de89160a00f7620.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 162, + 209, + 187 + ], + "lines": [ + { + "bbox": [ + 132, + 162, + 209, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 209, + 187 + ], + "type": "text", + "content": "Replace the bowl with two round apples and keep the cup unchanged" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 132, + 72, + 211, + 159 + ], + "blocks": [ + { + "bbox": [ + 132, + 72, + 211, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 72, + 211, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 211, + 159 + ], + "type": "image", + "image_path": "cec1937d48bdd41ee6ad4a7a21e590d32cf2d06ad881e976dfa18af88f7f7a74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 162, + 292, + 179 + ], + "lines": [ + { + "bbox": [ + 215, + 162, + 292, + 179 + ], + "spans": [ + { + "bbox": [ + 215, + 162, + 292, + 179 + ], + "type": "text", + "content": "Turn the cup into green and the bowl into golden" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 72, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 216, + 72, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 216, + 72, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 216, + 72, + 294, + 159 + ], + "type": "image", + "image_path": "2a54fbf66c7e2da6bbd5cb856e1b457297dbaab09b21650eaa3c8ce62f8b0de9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 162, + 374, + 178 + ], + "lines": [ + { + "bbox": [ + 299, + 162, + 374, + 178 + ], + "spans": [ + { + "bbox": [ + 299, + 162, + 374, + 178 + ], + "type": "text", + "content": "Swap the position of the bowl and the cup" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 378, + 158 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 378, + 158 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 378, + 158 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 378, + 158 + ], + "type": "image", + "image_path": "f3f3fefc5dcd868bd91cee3be94e35b7eba1d1290f92a98df0f9ba781d361a90.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 382, + 72, + 462, + 158 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 462, + 158 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 462, + 158 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 462, + 158 + ], + "type": "image", + "image_path": "405001983c82bc96a9d7adce58e4359fe71dd446cb79c7eaef1780bd935bad7a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 162, + 457, + 186 + ], + "lines": [ + { + "bbox": [ + 382, + 162, + 457, + 186 + ], + "spans": [ + { + "bbox": [ + 382, + 162, + 457, + 186 + ], + "type": "text", + "content": "Make cup 1.25 times its larger but keep the bowl's size" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 466, + 72, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 466, + 72, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 466, + 72, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 466, + 72, + 545, + 159 + ], + "type": "image", + "image_path": "2d45f0f869a2f801327f228e3c82f9a561e13d812f99d7eb23c8092b7d92c763.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 162, + 536, + 187 + ], + "lines": [ + { + "bbox": [ + 466, + 162, + 536, + 187 + ], + "spans": [ + { + "bbox": [ + 466, + 162, + 536, + 187 + ], + "type": "text", + "content": "Move the cup towards the bowl until they slightly overlap" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 49, + 196, + 128, + 282 + ], + "blocks": [ + { + "bbox": [ + 59, + 171, + 117, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 171, + 117, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 171, + 117, + 178 + ], + "type": "text", + "content": "User Instructions" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 196, + 128, + 282 + ], + "lines": [ + { + "bbox": [ + 49, + 196, + 128, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 196, + 128, + 282 + ], + "type": "image", + "image_path": "600d91a9dcd434d0c6a3a2844d7d41ab27582de36bc1432d15cb9ac508ba02d1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 293, + 118, + 301 + ], + "lines": [ + { + "bbox": [ + 61, + 293, + 118, + 301 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 118, + 301 + ], + "type": "text", + "content": "User Instructions" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 132, + 196, + 211, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 196, + 211, + 281 + ], + "lines": [ + { + "bbox": [ + 132, + 196, + 211, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 211, + 281 + ], + "type": "image", + "image_path": "03f715b515708b5e9eb1b474dac5a58de7f9cc050b10a78521d40c07be8187aa.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 285, + 208, + 311 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 208, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 208, + 311 + ], + "type": "text", + "content": "Replace the dog with a houseplant and keep the cat unchanged" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 216, + 196, + 294, + 281 + ], + "blocks": [ + { + "bbox": [ + 216, + 196, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 216, + 196, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 216, + 196, + 294, + 281 + ], + "type": "image", + "image_path": "c0b5b08eaae92c485f84cb968e8e138c14c5e730817fa18b8b31936e85e5400f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 285, + 291, + 311 + ], + "lines": [ + { + "bbox": [ + 215, + 285, + 291, + 311 + ], + "spans": [ + { + "bbox": [ + 215, + 285, + 291, + 311 + ], + "type": "text", + "content": "Transform the dog into a robotic dog and the cat into a sculptural cat" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 299, + 196, + 378, + 281 + ], + "blocks": [ + { + "bbox": [ + 299, + 196, + 378, + 281 + ], + "lines": [ + { + "bbox": [ + 299, + 196, + 378, + 281 + ], + "spans": [ + { + "bbox": [ + 299, + 196, + 378, + 281 + ], + "type": "image", + "image_path": "7cebd7be58e3ddce0898d7473c90d28aa1ee15d8557e3bc24b2dc1f4c4fa393f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 285, + 375, + 310 + ], + "lines": [ + { + "bbox": [ + 299, + 285, + 375, + 310 + ], + "spans": [ + { + "bbox": [ + 299, + 285, + 375, + 310 + ], + "type": "text", + "content": "Swap the position of the black dog and the yellow cat" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 382, + 196, + 462, + 281 + ], + "blocks": [ + { + "bbox": [ + 382, + 196, + 462, + 281 + ], + "lines": [ + { + "bbox": [ + 382, + 196, + 462, + 281 + ], + "spans": [ + { + "bbox": [ + 382, + 196, + 462, + 281 + ], + "type": "image", + "image_path": "1df495a9187b4cb0d960b07880a18be359c114da945d3719857c473d24290162.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 285, + 462, + 302 + ], + "lines": [ + { + "bbox": [ + 383, + 285, + 462, + 302 + ], + "spans": [ + { + "bbox": [ + 383, + 285, + 462, + 302 + ], + "type": "text", + "content": "Increase the cat's size and reduce the dog's size" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 196, + 545, + 282 + ], + "blocks": [ + { + "bbox": [ + 466, + 196, + 545, + 282 + ], + "lines": [ + { + "bbox": [ + 466, + 196, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 466, + 196, + 545, + 282 + ], + "type": "image", + "image_path": "ce4dec8db7f3f3b88574e4c426f5d9c56dc8dbdf29b61e725592283454c303a9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 285, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 465, + 285, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 465, + 285, + 545, + 311 + ], + "type": "text", + "content": "Shift the dog slightly upward and move the cat slightly towards the dog" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 49, + 371, + 123, + 445 + ], + "blocks": [ + { + "bbox": [ + 46, + 324, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 324, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 546, + 346 + ], + "type": "text", + "content": "Figure 7. SLD can handle a diverse array of image editing tasks guided by natural, human-like instructions. Its capabilities span from adjusting object counts to altering object attributes, positions, and sizes." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 361, + 258, + 369 + ], + "lines": [ + { + "bbox": [ + 76, + 361, + 258, + 369 + ], + "spans": [ + { + "bbox": [ + 76, + 361, + 258, + 369 + ], + "type": "text", + "content": "Remove hair of the rightmost person sitting at a cafe in this image" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 371, + 123, + 445 + ], + "lines": [ + { + "bbox": [ + 49, + 371, + 123, + 445 + ], + "spans": [ + { + "bbox": [ + 49, + 371, + 123, + 445 + ], + "type": "image", + "image_path": "537f67b74f4d53911293856a266d53587254fb5c76c38b2ff1901a04d2182a48.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 446, + 106, + 453 + ], + "lines": [ + { + "bbox": [ + 67, + 446, + 106, + 453 + ], + "spans": [ + { + "bbox": [ + 67, + 446, + 106, + 453 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 461, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 495 + ], + "type": "text", + "content": "Figure 8. SLD struggles with objects of complex shapes, as the SAM module may unintentionally segment adjacent parts during the process." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 130, + 371, + 205, + 445 + ], + "blocks": [ + { + "bbox": [ + 130, + 371, + 205, + 445 + ], + "lines": [ + { + "bbox": [ + 130, + 371, + 205, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 205, + 445 + ], + "type": "image", + "image_path": "925c6863876afdcf3efe077f0345a18e343b4309fb0bc08fb8315ea1fb5dd0d2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 446, + 187, + 453 + ], + "lines": [ + { + "bbox": [ + 143, + 446, + 187, + 453 + ], + "spans": [ + { + "bbox": [ + 143, + 446, + 187, + 453 + ], + "type": "text", + "content": "Removed Region" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 211, + 371, + 286, + 445 + ], + "blocks": [ + { + "bbox": [ + 211, + 371, + 286, + 445 + ], + "lines": [ + { + "bbox": [ + 211, + 371, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 211, + 371, + 286, + 445 + ], + "type": "image", + "image_path": "a9a18118c8952e83e3bd93cfecd77665a6234d23a75eb1bbb6a3894191ff52dd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 446, + 266, + 453 + ], + "lines": [ + { + "bbox": [ + 231, + 446, + 266, + 453 + ], + "spans": [ + { + "bbox": [ + 231, + 446, + 266, + 453 + ], + "type": "text", + "content": "Failed Result" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "content": "cision stems from the detector's exact object localization coupled with the LLMs' ability in reasoning and suggestions for new placements. Such detailed control over spatial adjustments is unmatched by any previous method, highlighting SLD's contributions to fine-grained image editing." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 575, + 119, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 119, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 119, + 586 + ], + "type": "text", + "content": "4.3. Discussion" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "content": "Multi-round self-correction. Our analysis in Tab. 2 highlights the benefits of multi-round self-correction and the fact that the first round correction is always the most effective one and has a marginal effect. The first round of corrections substantially mitigates issues inherent in Stable Diffusion [20]. Then, a second round of correction still yields significant improvements across all four tasks." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Limitations and future work. A limitation of our method is illustrated in Fig. 8, where SLD fails to accurately remove a person's hair. In this instance, despite the successful iden-" + } + ] + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 308, + 358, + 544, + 424 + ], + "blocks": [ + { + "bbox": [ + 308, + 358, + 544, + 424 + ], + "lines": [ + { + "bbox": [ + 308, + 358, + 544, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 358, + 544, + 424 + ], + "type": "table", + "html": "
MethodAccuracy
NegationNumeracyAttributeSpatialAverage
SD [20]19%38%24%33%28.5%
+ 1-round SLD69%55%25%69%54.5% (+26.0)
+ 2-round SLD73%61%31%75%60.0% (+31.5)
", + "image_path": "4b826b6cddd9a1692b96f074149a0723df9095c8425cf6eb8ff2d594bc0951d9.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 433, + 546, + 466 + ], + "lines": [ + { + "bbox": [ + 305, + 433, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 433, + 546, + 466 + ], + "type": "text", + "content": "Table 2. While the majority of errors are typically rectified in the first round, multi-round correction consistently outperforms a single-round approach." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "content": "tification and localization of the hair, the complex nature of its shape poses a challenge to the SAM module used for region selection, resulting in the unintended removal of the person's face in addition to the hair. However, since the person's cloth is not removed, the base diffusion model fails to generate a natural composition. This suggests that a better region selection method is needed for further improvements in the generation and editing quality." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 594, + 378, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 378, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 378, + 605 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 613, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 713 + ], + "type": "text", + "content": "We introduce the Self-correcting Language-Driven (SLD) framework, a pioneering self-correction system using detectors and LLMs to significantly enhance text-to-image alignment. This method not only sets a new SOTA in the image generation benchmark but is also compatible with various generative models, including DALL-E 3. Also, SLD extends its utility to image editing applications, offering fine-grained object-level manipulation that surpasses existing methods." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6334" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 214 + ], + "type": "text", + "content": "[3] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 259 + ], + "type": "text", + "content": "[4] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 304 + ], + "type": "text", + "content": "[5] Dave Epstein, Allan Jabri, Ben Poole, Alexei A. Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "text", + "content": "[6] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "type": "text", + "content": "[7] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LLM blueprint: Enabling text-to-image generation with complex and detailed prompts. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "type": "text", + "content": "[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 464, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 508 + ], + "type": "text", + "content": "[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "text", + "content": "[10] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[11] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "text", + "content": "[12] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[15] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 479, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 479, + 140 + ], + "type": "text", + "content": "[16] OpenAI. Dall-e 3 system card, 2023. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 142, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 546, + 196 + ], + "type": "text", + "content": "[17] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 546, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 546, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 546, + 252 + ], + "type": "text", + "content": "[18] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 546, + 297 + ], + "type": "text", + "content": "[19] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 300, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 547, + 354 + ], + "type": "text", + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 355, + 546, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 546, + 421 + ], + "type": "text", + "content": "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 546, + 477 + ], + "type": "text", + "content": "[22] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 479, + 546, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 546, + 533 + ], + "type": "text", + "content": "[23] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 535, + 546, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 546, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 546, + 600 + ], + "type": "text", + "content": "[24] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 602, + 546, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 546, + 666 + ], + "type": "text", + "content": "[25] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "type": "text", + "content": "[26] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6335" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "type": "text", + "content": "[27] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6336" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_content_list.json b/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd8e9da05ffa01b8fa0cf3175aa65000cfce8ace --- /dev/null +++ b/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_content_list.json @@ -0,0 +1,1557 @@ +[ + { + "type": "text", + "text": "Self-supervised debiasing using low rank regularization", + "text_level": 1, + "bbox": [ + 204, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geon Yeong Park1, Chanyong Jung1, Sangmin Lee2, Jong Chul Ye1,2,3* Sang Wan Lee1,4*", + "bbox": [ + 130, + 179, + 836, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Bio and Brain Engineering, $^{2}$ Mathematical Sciences,", + "bbox": [ + 272, + 196, + 696, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Kim Jaechul Graduate School of AI, 4Brain and Cognitive Sciences", + "bbox": [ + 214, + 215, + 754, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea", + "bbox": [ + 171, + 234, + 797, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pky3436, jcy132, leeeeesang,jong.ye,sangwan}@kaist.ac.kr", + "bbox": [ + 220, + 253, + 743, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spurious correlations can cause strong biases in deep neural networks, impairing generalization ability. While most existing debiasing methods require full supervision on either spurious attributes or target labels, training a debiased model from a limited amount of both annotations is still an open question. To address this issue, we investigate an interesting phenomenon using the spectral analysis of latent representations: spuriously correlated attributes make neural networks inductively biased towards encoding lower effective rank representations. We also show that a rank regularization can amplify this bias in a way that encourages highly correlated features. Leveraging these findings, we propose a self-supervised debiasing framework potentially compatible with unlabeled samples. Specifically, we first pretrain a biased encoder in a self-supervised manner with the rank regularization, serving as a semantic bottleneck to enforce the encoder to learn the spuriously correlated attributes. This biased encoder is then used to discover and upweight bias-conflicting samples in a downstream task, serving as a boosting to effectively debias the main model. Remarkably, the proposed debiasing framework significantly improves the generalization performance of self-supervised learning baselines and, in some cases, even outperforms state-of-the-art supervised debiasing approaches.", + "bbox": [ + 76, + 335, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While modern deep learning solves several challenging tasks successfully, a series of recent works [16, 18, 20] have reported that the high accuracy of deep networks on indistribution samples does not always guarantee low test error on out-of-distribution (OOD) samples, especially in the context of spurious correlations. Existing studies [3, 38, 50] suggest that the deep networks can be potentially biased to the spuriously correlated attributes, or dataset bias, which are misleading statistical heuristics that are closely correlated but not causally related to the target label.", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These catastrophic pitfalls of dataset bias have facilitated the development of debiasing methods, which can be roughly categorized into approaches: (1) leveraging annotations of spurious attributes, i.e., bias label [27, 46, 49, 56]; (2) presuming specific type of bias, e.g., color and texture [5, 17, 53]; or (3) without using explicit kinds of supervisions on dataset bias [30, 31, 35, 39, 60].", + "bbox": [ + 496, + 305, + 893, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While substantial advances have been made in this regard, these approaches still fail to address the problem: how to train a debiased classifier by fully exploiting unlabeled samples lacking both bias and target label. More specifically, while the large-scale unlabeled dataset can be potentially biased towards spuriously correlated sensitive attributes, e.g., ethnicity, gender, or age [1, 2], current existing debiasing frameworks are not designed to deal with this real-world unsupervised settings. Here we also confirm that most supervised debiasing frameworks suffer from performance degradation in the low-labeled data setting. Moreover, recent works have suggested that self-supervised learning might not be sufficient to deal with OOD generalization [10, 19, 44] when dataset bias remains after data augmentation.", + "bbox": [ + 496, + 411, + 895, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this issue, we first make a series of empirical observations that allow us to examine the fundamental difference between biased and unbiased representations. Interestingly, we found that spurious correlations suppress the effective rank [45] of latent representations, which severely deteriorates the semantic diversity of representations and leads to the degradation of feature discriminability. Another notable aspect of our findings is that the intentional increase of feature redundancy amplifies \"prejudice\" in neural networks. To be specific, as we enforce the correlation among latent features to regularize the effective rank of representations (i.e., rank regularization), the accuracy on bias-conflicting samples quickly declines while the model still performs reasonably well on the bias-aligned $^1$ samples.", + "bbox": [ + 496, + 625, + 895, + 838 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The bias-aligned samples refer to data with a strong correlation between (potentially latent) spurious features and target labels. The bias-conflicting samples refer to the opposite cases where spurious correlations do not exist.", + "bbox": [ + 498, + 849, + 893, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32adb27b77ae1b31508f6744fe78fa506eb842d8066fc972fdffc8511748ad58.jpg", + "image_caption": [ + "Figure 1. Concept. Based on the bias-rank relationship (Section 2), we introduce a novel debiasing framework centered on rank regularization, which intentionally amplifies spurious correlation by enforcing feature components to be entangled with both spurious and invariant attributes." + ], + "image_footnote": [], + "bbox": [ + 117, + 92, + 836, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these observations, we propose a novel debiasing framework that can utilize both labeled and unlabeled biased samples with rank regularization. The proposed method is fully compatible with both supervised and self-supervised scenarios, where such compatibility arises from the rank regularization that does not rely on any labels. Specifically, for a supervised (self-supervised) setting, we train 1) a biased classifier (encoder) with rank regularization, which serves as a semantic bottleneck limiting the semantic diversity of feature components, and 2) the main classifier (encoder) with standard (self-)supervised learning approaches. The biased model affords us the leverage to uncover spurious correlations and identify bias-conflicting samples in a downstream task.", + "bbox": [ + 75, + 358, + 473, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work is the first to unveil the bias-rank relationships and introduce an effective debiasing strategy to exploit potentially unlabeled data samples. We demonstrate the effectiveness of the proposed debiasing framework with various challenging real-world biased datasets, including MultiCM-NIST [34], biased Chest X-ray databases, UTKFace, CelebA, etc., in both a supervised and self-supervised scenario. These experiments show that our method significantly outperforms other self-supervised baselines, and even state-of-the-art supervised debiasing methods in some cases.", + "bbox": [ + 75, + 569, + 470, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Low-rank bias of biased representations", + "text_level": 1, + "bbox": [ + 76, + 731, + 437, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 756, + 217, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Throughout the paper, we denote $x \\in \\mathbb{R}^m$ and $y \\in \\mathcal{V}$ as $m$ -dimensional input sample and its corresponding predicting label, respectively. Then we denote $X = \\{x_k\\}_{k=1}^n$ as a batch of $n$ samples from a dataset which is fed to an encoder $f_\\theta : \\mathbb{R}^m \\to \\mathbb{R}^d$ , parameterized by $\\theta$ . Then we construct a matrix $Z \\in \\mathbb{R}^{n \\times d}$ where each $i$ th row is the output representations of the encoder $f_\\theta(x_i)^T$ for $x_i \\in X$ . For every analysis in this section, we use $Z$ as our latent representation.", + "bbox": [ + 76, + 779, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions, where the neural backbone of the encoder may vary as simple convolutional networks, ResNet-18, or ViT [15] (Experimental details provided in Appendix C.1 and D).", + "bbox": [ + 496, + 358, + 890, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the semantic diversity of given representation matrix, we introduce effective rank [45] which is a widely used metric to measure the effective dimensionality of matrix and analyze the spectral properties of features in neural networks [4, 6, 23, 42]:", + "bbox": [ + 496, + 402, + 892, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1. Given the matrix $X \\in \\mathbb{R}^{m \\times n}$ and its singular values $\\{\\sigma_i\\}_{i=1}^{\\min(m,n)}$ , the effective rank $\\rho$ of $X$ is defined as the shannon entropy of normalized singular values:", + "bbox": [ + 496, + 484, + 893, + 532 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (X) = - \\sum_ {i = 1} ^ {\\min (m, n)} \\bar {\\sigma} _ {i} \\log \\bar {\\sigma} _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 541, + 892, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\bar{\\sigma}_i = \\sigma_i / \\sum_k\\sigma_k$ is $i$ -th normalized singular value. Without loss of generality, we omit the exponentiation of $\\rho (X)$ as done in [45].", + "bbox": [ + 496, + 590, + 893, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Effective rank is also referred to as spectral entropy where its value is maximized when the singular values are all equal and minimized when a top singular value dominates relative to all others. Recent works [12, 13] have revealed that the discriminability of representations resides on wide range of eigenvectors since the rich discriminative information for the classification task cannot be transmitted by only few eigenvectors with top singular values. Thus from a spectral analysis perspective, effective rank quantifies how diverse the semantic information encoded by each eigenfeature is, which is closely related to the feature discriminability across target label categories. In the rest of paper, we interchangeably use effective rank and rank by following prior works.", + "bbox": [ + 496, + 643, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Spectral analysis of the bias-rank relationships", + "text_level": 1, + "bbox": [ + 500, + 847, + 890, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We now present experiments showing that the deep networks may tend to encode lower-rank representations in the pres", + "bbox": [ + 496, + 869, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12396", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b089feec5af2202506781bd6edff6d8f264af3d582e1102e8036b1d006ef70ea.jpg", + "image_caption": [ + "(a) Unbiased corr." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 251, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/efb658972a9fd5033c30fe74a24be47c6a2bc58e0fb7b165aebe7ad17766af7e.jpg", + "image_caption": [ + "(b) Biased corr." + ], + "image_footnote": [], + "bbox": [ + 303, + 90, + 467, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7a8bda690c1ff40e5b224fcc30a1e4e81071c04c00ba971fe6395bfcbf7edae6.jpg", + "image_caption": [ + "(c) Effective rank" + ], + "image_footnote": [], + "bbox": [ + 78, + 212, + 251, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d5b645d1e5d623cd3a9dc01051a7689d30b99996e49be5e08c1806acb6ed8467.jpg", + "image_caption": [ + "(d) SVD analysis", + "Figure 2. Empirical analysis on rank reduction phenomenon. For every analysis, we used the output $Z$ of the encoder (Sec. 2.1). (a, b): Hierarchically clustered auto-correlation matrix of unbiased and biased representations (Bias ratio=99%). (c): Effective rank with color bias. 'Unbiased' represents the case training model with perfectly unbiased dataset, i.e. random color for each sample. (d): SVD analysis with max-normalized singular values. Top 100 values are shown in the figure (total: 256)." + ], + "image_footnote": [], + "bbox": [ + 263, + 200, + 467, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ence of stronger spurious correlations. To arbitrarily control the degree of spurious correlations, we introduce synthetic biased datasets, Color-MNIST (CMNIST) and Corrupted CIFAR-10 (CIFAR-10C, [21]), with color and corruption bias types, respectively. We define the degree of spurious correlations as the ratio of bias-aligned samples included in the training set, or bias ratio, where most of the samples are bias-aligned in the context of strong spurious correlations.", + "bbox": [ + 75, + 477, + 468, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2c shows that the rank of latent representations from a penultimate layer of the simple convolutional classifier decreases as the bias ratio increases in CMNIST. We provide similar rank reduction results of CIFAR-10C with ResNet-18 and ViT in the Appendix C.1. We further compare the correlation matrix of biased and unbiased latent representations in the penultimate layer of biased and unbiased classifiers, respectively. In Figure 2a and 2b, we observe that the block structure in the correlation matrix is more evident in the biased representations after the hierarchical clustering, indicating that the features become highly correlated which may limit the semantic diversity of networks. To investigate the rank reduction phenomenon in-depth, we compare the normalized singular values of biased and unbiased representations. We conduct singular value decomposition (SVD) on the feature matrices of both biased and unbiased classifiers and plot the singular values normalized by the spectral norm of the corresponding matrix. Figure 2d shows that the top few normalized singular values of biased representations are similar to or even greater than those of unbiased rep", + "bbox": [ + 75, + 599, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resentations. However, the remaining majority of singular values decay significantly faster in biased representations, greatly weakening the informative signals of eigenvectors with smaller singular values and deteriorating feature discriminability [12, 13].", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Rank regularization", + "text_level": 1, + "bbox": [ + 500, + 175, + 691, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated from the aforementioned rank reduction phenomenon, we ask an opposite-directional question: \"Can we intentionally amplify the prejudice of deep networks by maximizing the redundancy between the components of latent representations?\" If the feature components are extremely correlated, the corresponding representations may exhibit most of its spectral energy along the direction of one singular vector. For this case, effective rank may converge to 0. In other words, our goal is to design a semantic bottleneck of representations that restricts the semantic diversity of feature vectors. To implement the bottleneck in practice, motivated from Figure 2b, we compute the auto-correlation matrix of the output of encoder.", + "bbox": [ + 496, + 199, + 893, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\bar{Z}$ denote the mean-centered representations $Z$ along the batch dimension. The normalized auto-correlation matrix $C\\in \\mathbb{R}^{d\\times d}$ of $\\bar{Z}$ is defined as follow:", + "bbox": [ + 498, + 395, + 890, + 439 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC _ {i, j} = \\frac {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , i} \\bar {Z} _ {b , j}}{\\sqrt {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , i} ^ {2}} \\sqrt {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , j} ^ {2}}} \\quad 1 \\leq \\forall i, j \\leq d, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 446, + 890, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $b$ is an index of sample and $i, j$ are index of each vector dimension. Then we define our regularization term as the negative of a sum of squared off-diagonal terms in $C$ :", + "bbox": [ + 496, + 501, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {r e g} (X; \\theta) = - \\sum_ {i} \\sum_ {j \\neq i} C _ {i, j} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 570, + 890, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we refer to it as the rank loss. Note that the target labels on $X$ is not used at all.", + "bbox": [ + 496, + 613, + 890, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis of rank-regularized networks. To investigate the impacts of rank regularization in deep neural networks, we construct the classification model by combining the linear classifier $f_{W}:\\mathbb{R}^{d}\\to \\mathbb{R}^{c}$ parameterized by $W\\in \\mathcal{W}$ on top of the encoder $f_{\\theta}$ , where $c = |\\mathcal{V}|$ is the number of classes. Then we trained models by cross entropy loss $\\ell_{CE}$ combined with $\\lambda_{reg}\\ell_{reg}$ , where $\\lambda_{reg} > 0$ is a Lagrangian multiplier. We use CMNIST, CIFAR-10C, and Waterbirds dataset [51], and evaluate the trained models on an unbiased test set following [30, 39]. After training models with varying the hyperparameter $\\lambda_{reg}$ , we compare bias-aligned and bias-conflict accuracy, which are the average accuracy on bias-aligned and bias-conflicting samples in the unbiased test set, respectively, for CMNIST and CIFAR-10C. Test accuracy on every individual data group is reported for Waterbirds. Figure 3 shows that models suffer more from poor OOD generalization as trained with larger $\\lambda_{reg}$ . The average", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16631f1a8a4c15d5116077b9006a0e6aeffc1ee152c2a80ceb95d868aa3b9ade.jpg", + "image_caption": [ + "(a) CMNIST" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 321, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/93667851c321fa8d62815232046c5a40e097a187b7cc04f600783039e4d37f3b.jpg", + "image_caption": [ + "(b) CIFAR-10C", + "Figure 3. (a, b): Bias-conflict and Bias-aligned accuracy on CMNIST and CIFAR-10C (Bias ratio=95%). (c): Group accuracy on Waterbirds. Detailed simulation settings are in the Appendix D." + ], + "image_footnote": [], + "bbox": [ + 359, + 90, + 602, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0bd0291c9fd0a4876d2fde24aa2ef9c51600adb1ed90d85f86dd88ebffac7a8b.jpg", + "image_caption": [ + "(c) Waterbirds" + ], + "image_footnote": [], + "bbox": [ + 643, + 90, + 883, + 218 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/6b62aadfe3f36b3453fca9621e001330e1f2f0c9897047307f39c1b4040dbc8e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CMNISTCIFAR-10C
P (%)R (%)P (%)R (%)
ERM85.5919.7652.030.06
+ Rank reg98.8395.9171.3951.43
", + "bbox": [ + 120, + 301, + 426, + 385 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bfb912d082d63259dbd93bc12a5282da353c75d75be02bfd7ce13102ea152e89.jpg", + "table_caption": [ + "(a) CMNIST, CIFAR-10C" + ], + "table_footnote": [], + "table_body": "
MetricsERMJTTRank reg
Precision (%)37.8448.9554.77
Recall (%)11.6748.7555.01
", + "bbox": [ + 133, + 400, + 413, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(b) Waterbirds", + "bbox": [ + 233, + 465, + 312, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Precision (P) and Recall (R) of bias-conflicting samples. (a): Bias-conflicting samples are identified in the error set of ERM model trained with and without rank regularization (Bias ratio=95% for both datasets). (b): Bias-conflicting samples are similarly identified by ERM, JTT, and the proposed biased model in Waterbirds dataset.", + "bbox": [ + 75, + 489, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "accuracy on bias-conflicting groups is significantly degraded, while the accuracy on bias-aligned groups is maintained to some extent. It implies that rank regularization may force deep networks to focus on spurious attributes.", + "bbox": [ + 75, + 607, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Minority mining performance. Table 1a and 1b support that the biased models with strong regularization can effectively probe out the bias-conflicting samples in the training set. Specifically, we train a biased classifier with rank regularization and distill an error set $E$ of misclassified training samples as bias-conflicting samples proxies. As reported in Table 1a, we observe that our biased classifier is relatively robust to the unintended memorization of bias-conflicting samples [47] in contrast to the standard models trained by Empirical Risk Minimization (ERM). Moreover, Table 1b shows that the proposed rank regularization improves the precision and recall of identified bias-conflicting samples compared to JTT [35]. Detailed simulation settings are in the Appendix D.", + "bbox": [ + 75, + 670, + 470, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction of biased representations. To under", + "bbox": [ + 96, + 885, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stand the relationship between rank regularization and spurious correlations more deeply, we visualize the pretrained representations with varying degrees of bias. We first trained deep networks on: (a) unbiased CMNIST (random background color), (b) biased CMNIST (bias ratio $= 95\\%$ ) without rank regularization and (c) with rank regularization $(\\lambda_{reg} = 50)$ . Then, we train the auxiliary decoder, which reconstructs the bias-conflicting images from the freezed latent representations of each pretrained network. Results show that rank regularization may cause the representation to lose information on complex invariant features, resulting in a loss of feature discriminability and informative signals. While both digit and color are well reconstructed with biased representations (b), the decoder fails to reconstruct bias-conflicting images from the (c) biased representations pretrained with rank regularization. The foreground digit is blurred, and its class changes following the color-digit assignment in Figure 4d.", + "bbox": [ + 496, + 303, + 893, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These observations afford us some key insights into rank regularization: First, the rank-regularized representation may lose its information on complex invariant features (i.e., shape and style of the foreground digit), specifically undermining the feature discriminability and informative signals. Second, the limited semantic diversity makes it harder to identify the true underlying independent generative factors for multidimensional data; instead, it may encode feature components entangled with both spurious and invariant attributes as the digit class of the reconstructed image is erroneously determined by the background color in 4c.", + "bbox": [ + 496, + 578, + 893, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multiple bias attributes. To further investigate the generalizability of rank regularization, we evaluate the biased representations with Multi-Color MNIST (MultiCMNIST) dataset [34], which is similar to the CMNIST but have two bias attributes: left and right background colors. We set bias ratio $= 99\\%$ for the left color and bias ratio $= 95\\%$ for the right color, i.e., the left color is a more salient bias than the right color (Dataset details are provided in Appendix D).", + "bbox": [ + 496, + 746, + 893, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2 shows that the rank regularization successfully biases the model w.r.t both bias attributes, while LfF [39]", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12398", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/05f8c9739d4ab43b343e2d3095e534d34b9d18c0b42e7bb4b80657c07387e638.jpg", + "image_caption": [ + "(a) Unbiased" + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 310, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/96a7397b2d57f6c185f97c008c8f726b90fc3c1dec4301d34b34f4637c13f5fe.jpg", + "image_caption": [ + "(b) Biased", + "Figure 4. Randomly selected reconstructed images from representations with varying degrees of bias. First and third row correspond to the input bias-conflicting images. Second and fourth row correspond to the reconstructed images. Reconstructed from (a) unbiased representations, (b) biased representations, and (c) biased representations with rank regularization (bias ratio=95% in b, c). (d) Examples of bias-aligned CMNIST images." + ], + "image_footnote": [], + "bbox": [ + 310, + 88, + 540, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/735bf7143f5bc175e63c3cc20101e97dd51d39ca6dbbbf30cae9c4281510624c.jpg", + "image_caption": [ + "(c) Biased with rank reg." + ], + "image_footnote": [], + "bbox": [ + 540, + 88, + 769, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bda856b5e403e77bfd7c89046dc2a08be7baae0f4b5977c6ee0e510f79889506.jpg", + "image_caption": [ + "(d) Examples." + ], + "image_footnote": [], + "bbox": [ + 769, + 90, + 888, + 166 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2c6a520e00dd406bf6a374402b0b161e17a8451721825b3ec2a5ba84f7cc4ab4.jpg", + "table_caption": [ + "Table 2. (a) Test accuracy (\\%) on MultiCMNIST. Lower is better for this results. BC for bias-conflicting, and BA for bias-aligned. Bias ratio=99(\\%) for left color, and 95(\\%) for right color. $\\lambda_{reg} = 50$ is used for rank regularization. (b) Debiasing results. Higher is better for this results. Baseline results are from [34]. $\\lambda_{up} = 50$ is used for upweighting in the proposed framework ( $\\lambda_{up}$ : a manual rescaling weight given to each identified bias-conflicting samples in cross entropy loss). Pseudo-code and experimental details are provided in Appendix A and D, respectively." + ], + "table_footnote": [], + "table_body": "
IdxLeft colorRight color(a) Biased accuracy (%)(b) Debiased accuracy (%)
ERMLfF [39]Rank reg.LFFDebiAN [34]Ours
(1)BABC100.0100.0100.099.6100.0100.0
(2)BABC96.698.841.64.795.697.0
(3)BCBA29.33.28.798.676.579.1
(4)BCBC7.61.36.15.116.018.3
(1) ~ (4) average acc.58.3850.8339.152.072.073.6
", + "bbox": [ + 81, + 411, + 465, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "completely fails to amplify the right color bias, i.e. less salient bias, as shown in the second row (Biased accuracy part). This leads to the abnormal debiasing results of LfF as shown in Table 2 where it records unbalanced accuracy for the left- and right-color-bias-conflicting samples. In contrast, the proposed framework shows superior performance by simply weighting the misclassified bias-conflicting proxies, as done in [35].", + "bbox": [ + 75, + 525, + 468, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taken together, these results indicate that the rank regularization encourages the network to focus more on spurious correlations in a way that minimizes semantic diversity and entangles invariant and spurious features [41], which is a fundamentally different mechanism compared to the LfF [39] with its easy-to-learn assumption. More details on the upweighting strategy will be provided in Section 3 and pseudo-code in Appendix A.", + "bbox": [ + 75, + 647, + 470, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. DeFund: Debiasing framework with unlabeled data", + "text_level": 1, + "bbox": [ + 76, + 781, + 472, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivated by the observations in Section 2, we propose a self-supervised debiasing framework with unlabeled data, coined DeFund (Debiasing Framework with Unlabeled Data). A notable distinction from previous studies [5, 60] lies in the proposed framework's ability to effectively harness un", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "labeled data for learning biased representations. This is achieved through the application of self-supervised learning and rank regularization techniques.", + "bbox": [ + 496, + 277, + 892, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed framework is composed of two stages: We first train the biased encoder, which can be potentially adopted to detect the bias-conflicting samples in a downstream task, along with the main encoder by self-supervised learning, both without any labels. After pretraining, we identify the bias-conflicting samples in the downstream task using linear evaluation protocol [9, 40]. This set of samples serves as a boosting to debias the main model.", + "bbox": [ + 496, + 327, + 893, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notation. We denote $f_{\\theta}^{bias} : \\mathcal{X} \\to \\mathbb{R}^d$ and $f_{\\phi}^{main} : \\mathcal{X} \\to \\mathbb{R}^d$ as biased encoder and main encoder parameterized by $\\theta \\in \\Theta$ and $\\phi \\in \\Theta$ , respectively, where $d$ is the dimensionality of latent representations. Then we can compute the rank loss in (3) with introduced encoders and given batch $\\{x_k\\}_{k=1}^N$ with size $N$ . Let $f_{W_b}^{cls} : \\mathbb{R}^d \\to \\mathbb{R}^C$ be a single-layer classifier parameterized by $W_b \\in \\mathcal{W}$ which is placed on top of biased encoder $f_{\\theta}^{bias}$ , where $C = |\\mathcal{Y}|$ is the number of classes. We similarly define the linear classifier $f_{W_m}^{cls}$ for the main encoder. Then we refer to $f^{bias} : \\mathcal{X} \\to \\mathbb{R}^C$ as biased model, where $f^{bias}(x) = f_{W_b}^{cls}\\big(f_{\\theta}^{bias}(x)\\big), \\forall x \\in \\mathcal{X}$ . We similarly define the main model $f^{main}$ as $f^{main}(x) = f_{W_m}^{cls}\\big(f_{\\phi}^{main}(x)\\big), \\forall x \\in \\mathcal{X}$ . While the projection networks [9] are employed as well, we omit the notations because they are not engaged in classification.", + "bbox": [ + 496, + 452, + 893, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1. Training a biased encoder. To train the biased encoder $f_{\\theta}^{bias}$ , we revisit the proposed rank regularization term (3) in context of instance discrimination task. Building upon the observations in Section 2.3, we conjecture that rank regularization may amplify bias in self-supervised learning as well by entangling invariant and spurious features. Based on these intuitions, we apply rank regularization directly to the output of the base encoder, which encourages each feature component to be highly correlated. From these applications, several noteworthy observations have emerged: (a) The representation becomes more biased as it is trained with stronger regularization (Appendix C.1). (b) While the overall performance may be upper-bounded due to the constraint on effective dimensionality [25], the bias-conflict accuracy", + "bbox": [ + 496, + 689, + 895, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is primarily sacrificed compared to the bias-aligned accuracy (Section 4).", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 2. Debiasing downstream tasks. After training the biased encoder, our goal is to debias the main model, which was pretrained using standard self-supervised learning methods on the same dataset. Here, assume that we have an ideal pretrained main encoder of which each output component corresponds to the latent factor of data variation [63]. While this ideal encoder should seamlessly adapt to downstream classification tasks, if most downstream task samples are bias-aligned, they may misguide the model to upweight spuriously correlated latent factors, leading to a biased solution despite well-generalized representations. We refer to this problem as the biased downstream application (Related analysis in Appendix C.1).", + "bbox": [ + 75, + 121, + 468, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above contradiction elucidates the importance of bias-conflicting samples, which serve as counterexamples of spuriously correlated feature components, thereby preventing the alleged involvement of such components in prediction. Based on these intuitions, we introduce a novel debiasing protocol that probes and upweights bias-conflicting samples to find and fully exploit feature components independent of spurious correlations. We apply our framework on two scenarios: linear evaluation and semi-supervised learning.", + "bbox": [ + 75, + 316, + 470, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Linear evaluation. To validate our hypothesis on the biased downstream application, we conduct linear evaluation [40, 61] following the conventional protocol of self-supervised learning. Specifically, a linear classifier is trained on top of unsupervised pretrained representations by using target labels of training samples. After training a linear classifier $f_{W_b}^{cls}$ with pretrained biased encoder $f_{\\theta}^{bias}$ given the whole training set $D = \\{(x_k, y_k)\\}_{k=1}^N$ with size $N$ , an error set $E$ of misclassified samples and corresponding labels is regarded as bias-conflicting pairs. Then we train a linear classifier $f_{W_m}^{cls}$ on intentionally frozen representations of main encoder $f_{\\phi}^{main}$ by weighting the identified samples in $E$ with $\\lambda_{up} > 0$ . The loss function for debiased linear evaluation is defined as follows:", + "bbox": [ + 75, + 454, + 470, + 667 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\ell_ {d e b i a s} (D; W _ {m}) \\\\ = \\lambda_ {u p} \\sum_ {(x, y) \\in E} \\ell (x, y; W _ {m}) + \\sum_ {(x, y) \\in D \\setminus E} \\ell (x, y; W _ {m}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 675, + 449, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we use cross entropy loss for $\\ell : \\mathcal{X} \\times \\mathcal{Y} \\times \\mathcal{W} \\to \\mathbb{R}^+$ . Note that the target labels are only used in training linear classifiers after pretraining.", + "bbox": [ + 75, + 734, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the debiased linear evaluation is not meant to compete directly with other supervised baselines. Instead, it aims to: (a) examine the potential origin of the failure in OOD generalization, (b) provide a rough estimate of the potential improvement achievable with frozen latent representations, and (c) compare with standard self-supervised baselines and identify the optimal learning algorithms, e.g. SimCLR [9], for training the main encoder.", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semi-supervised learning. We further compare our method directly to other supervised debiasing methods in the context of semi-supervised learning. Here we assume that the training dataset includes only a small amount of labeled data combined with a large amount of unlabeled data. As in linear evaluation, we train a linear classifier on top of the biased encoder by using labeled samples. After obtaining an error set $E$ of misclassified samples, we finetuned the whole main model by weighting the identified samples in $E$ with $\\lambda_{up}$ . Note that supervised baselines are restricted to using only a small fraction of labeled samples, while the proposed approach benefits from the abundant unlabeled samples during pre-training of the biased encoder (Pseudo-code in the Appendix section A).", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 319, + 586, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Methods", + "text_level": 1, + "bbox": [ + 500, + 344, + 602, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate several supervised and self-supervised baselines on MIMIC-CXR + NIH [32], UTKFace [62] and CelebA [36] in which prior work reported poor generalization performance due to spurious correlations (Dataset details in Appendix).", + "bbox": [ + 496, + 369, + 893, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For MIMIC-CXR + NIH, we mixed the MIMIC-CXR [26] and NIH [55] following [32] where the target categories are no finding and pneumonia. Most pneumonia images are collected from MIMIC-CXR, while most no finding images are from NIH. In other words, the biases come from systematic differences in data sources, where the classifier may erroneously rely on spurious radiographic features tied to variations in data acquisition pipelines [14] instead of true pathological indicators (Example images in Figure 8).", + "bbox": [ + 496, + 446, + 893, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For UTKFace, we conduct binary classifications using (Gender, Age) and (Race, Gender) as (target, spurious) attribute pair, which we refer to UTKFace (age) and UTKFace (gender), respectively. For CelebA, we consider (HeavyMakeup, Male) and (Blonde Hair, Male) as (target, spurious) attribute pairs, which are referred to CelebA (makeup) and CelebA (blonde), respectively. The results of CelebA (blonde) are reported in Appendix C.4. Following [22, 39], we report bias-conflict accuracy together with unbiased accuracy, which is evaluated on the explicitly constructed validation set. We exclude the dataset in Figure 3 based on the observations that the SimCLR models are already invariant w.r.t spurious attributes.", + "bbox": [ + 496, + 598, + 893, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We mainly target baselines consisting of recent advanced self-supervised learning methods, SimCLR [9], VICReg [7], and SimSiam [11], which can be categorized into contrastive (SimCLR) and non-contrastive (VICReg, SimSiam) methods. We further report the performance of vanilla networks trained by ERM, and other supervised debiasing methods such as LNL [27], EnD [49], and", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12400", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b67e85b24028acc1d21174bc6f55d688c15546f6c1d73cf78199af1cadd1cc6d.jpg", + "table_caption": [ + "Table 3. (Supervised learning) Bias-conflict and unbiased accuracy (%) on MIMIC-CXR + NIH. Each $\\checkmark$ marker represents whether the model requires information on dataset bias. Bias ratio=10%." + ], + "table_footnote": [], + "table_body": "
AccuracyLNL √EnD √LfF XJTT XCVaR DRO XERM XSimCLR XOurs X
Conflict43.8±0.550.4±2.325.2±2.147.9±0.244.6±0.541.7±1.235.5±1.356.8±1.7
Unbiased68.1±1.071.8±1.460.8±0.268.9±1.065.8±1.267.8±1.062.0±1.469.8±0.2
", + "bbox": [ + 86, + 128, + 885, + 223 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/061b5c91c97d8e493ff9f7598c8af75b1a2a97ed1848ccdfb6d0c42653da1392.jpg", + "table_caption": [ + "Table 4. (Linear evaluation) Bias-conflict and unbiased test accuracy (%) evaluated on UTKFace and CelebA. Models requiring information on target class or dataset bias in the (pre)training stage are denoted with √in columns Y and B, respectively. Our results are marked in bold to highlight the improvements compared to the mainly interested self-supervised learning baselines (Gray rows)." + ], + "table_footnote": [], + "table_body": "
ModelYBUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
LNL✓✓45.8±0.672.6±0.373.1±1.684.9±0.855.9±2.176.0±0.6
EnD✓✓45.3±0.972.2±0.275.5±1.185.5±0.457.3±2.476.4±1.4
JTT✓✗63.8±0.969.4±1.371.2±0.377.6±0.462.4±1.274.7±0.8
CVaR DRO✓✗45.7±2.071.4±0.368.6±1.081.0±0.858.0±1.776.5±0.6
ERM✓✗45.4±2.171.0±1.265.7±1.479.5±0.654.2±0.274.1±1.4
SimSiam✗✗28.2±0.962.6±0.748.5±1.069.8±0.739.9±0.666.7±0.6
VICReg✗✗32.3±0.664.6±0.351.0±1.471.3±0.748.6±0.671.9±0.2
SimCLR✗✗36.4±1.566.3±0.656.3±0.274.2±0.246.9±1.069.8±0.4
DeFund✗✗59.5±0.870.6±0.863.7±2.074.9±0.958.4±0.673.1±1.0
", + "bbox": [ + 166, + 287, + 805, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "upweighting-based algorithms, JTT [35] and CVaR DRO [31], which can be categorized into methods that leverage annotations on dataset bias (LNL, EnD) or not (JTT, CVaR DRO).", + "bbox": [ + 75, + 513, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimization setting. Both bias and main encoder is pretrained with SimCLR [9] for 100 epochs on UTKFace, and 20 epochs on CelebA, respectively, using ResNet-18, Adam optimizer and cosine annealing learning rate scheduling [37]. We use a MLP with one hidden layer for projection networks as in SimCLR. All the other baseline results are reproduced by tuning the hyperparameters and optimization settings using the same backbone architecture. We report the results of the model with the highest bias-conflicting test accuracy over those with improved unbiased test accuracy compared to the corresponding baseline algorithms, i.e., SimCLR for ours (More experimental details in Appendix D).", + "bbox": [ + 75, + 575, + 470, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation results", + "text_level": 1, + "bbox": [ + 76, + 770, + 253, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Supervised learning. To quantify the effectiveness of the rank regularization in-depth, we first consider a standard supervised debiasing scenario as similarly done in Table 2. For a MIMIC-CXR + NIH dataset, we found that the proposed framework outperforms other supervised baselines with respect to bias-conflict accuracy. Table 14 in the Appendix shows that the rank-regularized networks effectively", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "discover the bias-conflicting samples which are consistent with Table 1a, 1b, and 2.", + "bbox": [ + 496, + 513, + 890, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Linear evaluation. We also found that DeFund outperforms every self-supervised baseline by a large margin in a linear evaluation protocol, including SimCLR, SimSiam and VICReg, with respect to both bias-conflict and unbiased accuracy (Table 4). Moreover, in some cases, DeFund even outperforms ERM models or supervised debiasing approaches regarding bias-conflict accuracy. Note that there is an inherent gap between ERM models and self-supervised baselines, roughly $8.7\\%$ on average. Moreover, we found that non-contrastive learning methods generally perform worse than the contrastive learning method. This warns us against training the main model using a non-contrastive learning approach, while it may be a viable option for the biased model. Results of the proposed framework with non-contrastive learning methods are provided in the Appendix section C.5.", + "bbox": [ + 496, + 547, + 893, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semi-supervised learning. To compare supervised and self-supervised methods in a more practical and fair scenario, we randomly sample $10\\%$ of the labeled CelebA training dataset. The remaining $90\\%$ samples are treated as unlabeled ones and engaged only in pretraining encoders for self-supervised baselines. Labeled samples are provided equally to both supervised and self-supervised methods.", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d317a7308fc22a422048a97686430f54493974e8aa99fc33ce356f6a7eb22d4.jpg", + "table_caption": [ + "Table 5. (Semi-supervised learning) Accuracy results (%) on CelebA. Label fraction $= {10}\\%$ ." + ], + "table_footnote": [], + "table_body": "
AccuracyCelebA (Makeup)CelebA (Blonde)
LNLEnDJTTCVaR DROERMSimCLRDeFundJTTDeFund
Conflict55.7±1.455.3±1.551.5±1.955.6±1.551.5±1.150.5±4.760.5±0.470.6±1.075.1±0.8
Unbiased75.6±0.576.2±0.871.4±1.375.7±1.073.1±0.371.6±1.975.6±0.278.8±1.785.8±0.3
", + "bbox": [ + 106, + 114, + 864, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remarkably, Table 5 and Table 16 in Appendix show that the proposed framework outperforms other state-of-the-art supervised debiasing methods. Existing upweighting protocols, such as JTT, fail to prevent deep networks from memorizing minority counterexamples. However, the proposed framework can fully utilize unlabeled samples with contrastive learning to prevent memorization. Existing bias-conflicting sample mining algorithms may be affected by the implicit bias of overparameterized networks, but this is unlikely to happen with the proposed framework since it only trains a simple linear classifier on top of a frozen biased encoder to identify such samples.", + "bbox": [ + 75, + 224, + 472, + 407 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e57ca90ffec5b11ee87731b251ded63efe8afa1d10d676c8e4a15572626c25c3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
SimCLR36.466.356.374.246.969.8
+ Rank reg26.661.350.970.343.968.3
+ Upweight53.064.658.374.550.170.4
DeFund59.570.663.774.958.473.1
", + "bbox": [ + 84, + 419, + 470, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0e9bb32559e62f0b2ac189d7c1df3c1407c2207f9744b9c469929701c6cf0bca.jpg", + "table_caption": [ + "(a) Ablation study" + ], + "table_footnote": [], + "table_body": "
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
PrecisionRecallPrecisionRecallPrecisionRecall
SimCLR68.3144.6333.3639.5952.2528.23
DeFund68.6775.9429.9850.9355.2932.46
", + "bbox": [ + 86, + 512, + 470, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Precision and recall", + "bbox": [ + 215, + 582, + 341, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. (a) Ablation study on introduced modules. (b) Precision and recall (\\%) of bias-conflicting samples identified by SimCLR and our biased model. Both case used linear evaluation.", + "bbox": [ + 75, + 606, + 468, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. To quantify the extent of performance improvement achieved by each introduced module, we compared the linear evaluation results of (a) vanilla SimCLR, (b) SimCLR with rank regularization, (c) SimCLR with upweighting error set $E$ of the main model, and (d) DeFund. Note that (c) does not use a biased model at all. Table 6a shows that every module plays an important role in OOD generalization. Considering that the main model is already biased to some extent, we found that bias-conflict accuracy can be improved even without a biased model, where the error set $E$ of the biased model further boosts the generalization performance. We also measures the precision and recall of identified bias-conflicting samples in E, finding that the biased model detects more diverse bias-conflicting samples than the baseline (Table 6b). The improvement of recall in CelebA may seem marginal, but it is significant given the", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "larger number of samples compared to UTKFace.", + "bbox": [ + 496, + 224, + 828, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational costs. Our framework is computationally affordable as it only trains the linaer classifier (linear eval.) or finetune networks with a few epochs, e.g., about 30 epochs for UTKFace in debiasing stage. Self-supervised pre-training and linear evaluation takes 19.3 and 4.5 minutes with a NVIDIA GeForce RTX 2080Ti, respectively.", + "bbox": [ + 496, + 241, + 893, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 339, + 619, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel solution to the challenging self-supervised debiasing, an important problem that has received little attention so far. We (a) unveil the inductive bias towards encoding low effective rank representations in the presence of spurious correlations. Based on this, we (b) design a rank regularization that amplifies the feature redundancy by reducing the spectral entropy of latent representations. Then we (c) design a debiasing framework empowered by the biased model pretrained with abundant unlabeled samples.", + "bbox": [ + 496, + 364, + 893, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 498, + 513, + 633, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by National Research foundation of Korea(NRF) (\\*\\*RS-2023-00262527**), Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137899, KMDF_PR_20200901_0015), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT, Ministry of Science and ICT) (No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation), (No.2019-0-00075, Artificial Intelligence Graduate School Program(KAIST)), (No. RS-2023-00233251, System3 reinforcement learning with high-level brain functions), Electronics and Telecommunications Research Institute(ETRI) grant funded by the Korean government (24ZS1100, Core Technology Research for Self-Improving Integrated Artificial Intelligence System), [N01230878, Development of Beyond X-verse Core Technology for Hyper-realistic interactions by Synchronizing the Real World and Virtual Space], a grant of the KAIST-KT joint research project through AI2XL Laboratory, Institute of convergence Technology, funded by KT (Project No.3, Project title: Genie Brain).", + "bbox": [ + 496, + 537, + 895, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12402", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abubakar Abid, Maheen Farooqi, and James Zou. Large language models associate muslims with violence. Nature Machine Intelligence, 3(6):461-463, 2021.", + "[2] Sandhini Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, and Miles Brundage. Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818, 2021.", + "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019.", + "[4] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. Advances in Neural Information Processing Systems, 32, 2019.", + "[5] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In International Conference on Machine Learning, pages 528-539. PMLR, 2020.", + "[6] Aristide Baratin, Thomas George, César Laurent, R Devon Hjelm, Guillaume Lajoie, Pascal Vincent, and Simon Lacoste-Julien. Implicit regularization via neural feature alignment. In International Conference on Artificial Intelligence and Statistics, pages 2269-2277. PMLR, 2021.", + "[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021.", + "[8] Tiffany Tianhui Cai, Jonathan Frankle, David J Schwab, and Ari S Morcos. Are all negatives created equal in contrastive instance discrimination? arXiv preprint arXiv:2010.06682, 2020.", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[10] Ting Chen, Calvin Luo, and Lala Li. Intriguing properties of contrastive losses. Advances in Neural Information Processing Systems, 34:11834-11845, 2021.", + "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021.", + "[12] Xinyang Chen, Sinan Wang, Bo Fu, Mingsheng Long, and Jianmin Wang. Catastrophic forgetting meets negative transfer: Batch spectral shrinkage for safe transfer learning. Advances in Neural Information Processing Systems, 32, 2019.", + "[13] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019.", + "[14] Alex J DeGrave, Joseph D Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021.", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[16] Michael Feldman, Sorelle A Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. Certifying and removing disparate impact. In proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, pages 259-268, 2015.", + "[17] Songwei Ge, Shlok Mishra, Chun-Liang Li, Haohan Wang, and David Jacobs. Robust contrastive learning using negative samples with diminished semantics. Advances in Neural Information Processing Systems, 34:27356-27368, 2021.", + "[18] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018.", + "[19] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models. arXiv preprint arXiv:2010.08377, 2020.", + "[20] Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324, 2018.", + "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019.", + "[22] Youngkyu Hong and Eunho Yang. Unbiased classification through bias-contrastive and bias-balanced learning. Advances in Neural Information Processing Systems, 34:26449-26461, 2021.", + "[23] Minyoung Huh, Hossein Mobahi, Richard Zhang, Brian Cheung, Pulkit Agrawal, and Phillip Isola. The low-rank simplicity bias in deep networks. arXiv preprint arXiv:2103.10427, 2021.", + "[24] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022.", + "[25] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348, 2021.", + "[26] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019.", + "[27] Byungju Kim, Hyunwoo Kim, Kyungsu Kim, Sungjin Kim, and Junmo Kim. Learning not to learn: Training deep neural networks with biased data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9012-9020, 2019.", + "[28] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022.", + "[29] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12403", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021.", + "[30] Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, and Jaegul Choo. Learning debiased representation via disentangled feature augmentation. Advances in Neural Information Processing Systems, 34:25123-25133, 2021.", + "[31] Daniel Levy, Yair Carmon, John C Duchi, and Aaron Sidford. Large-scale methods for distributionally robust optimization. Advances in Neural Information Processing Systems, 33:8847-8860, 2020.", + "[32] Jiaxuan Li, Duc Minh Vo, and Hideki Nakayama. Partition-and-debias: Agnostic biases mitigation via a mixture of biases-specific experts. arXiv preprint arXiv:2308.10005, 2023.", + "[33] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14970-14979, 2021.", + "[34] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In European Conference on Computer Vision, pages 270-288. Springer, 2022.", + "[35] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021.", + "[36] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015.", + "[37] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.", + "[38] Vaishnavh Nagarajan, Anders Andreassen, and Behnam Neyshabur. Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775, 2020.", + "[39] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020.", + "[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.", + "[41] Geon Yeong Park, Sangmin Lee, Sang Wan Lee, and Jong Chul Ye. Training debiased subnetworks with contrastive weight pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7929-7938, 2023.", + "[42] Noam Razin and Nadav Cohen. Implicit regularization in deep learning may not be explainable by norms. Advances in neural information processing systems, 33:21174-21187, 2020." + ], + "bbox": [ + 78, + 90, + 472, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020.", + "[44] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021.", + "[45] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007.", + "[46] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019.", + "[47] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020.", + "[48] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv preprint arXiv:2206.01197, 2022.", + "[49] Enzo Tartaglione, Carlo Alberto Barbano, and Marco Grangetto. End: Entangling and disentangling deep representations for bias correction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13508-13517, 2021.", + "[50] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018.", + "[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011.", + "[52] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021.", + "[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019.", + "[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020.", + "[55] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017.", + "[56] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8919-8928, 2020." + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12404", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Yadollah Yaghoobzadeh, Soroush Mehri, Remi Tachet, Timothy J Hazen, and Alessandro Sordoni. Increasing robustness to spurious correlations using forgettable examples. arXiv preprint arXiv:1911.03861, 2019.", + "[58] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021.", + "[59] Dinghuai Zhang, Kartik Ahuja, Yilun Xu, Yisen Wang, and Aaron Courville. Can subnetwork structure be the key to out-of-distribution generalization? In International Conference on Machine Learning, pages 12356-12367. PMLR, 2021.", + "[60] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022.", + "[61] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016.", + "[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017.", + "[63] Roland S Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive learning inverts the data generating process. In International Conference on Machine Learning, pages 12979-12990. PMLR, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 484 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12405", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_model.json b/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb0fad21dbbc2932880e084bff9531b62c62be52 --- /dev/null +++ b/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "Self-supervised debiasing using low rank regularization" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.18, + 0.838, + 0.2 + ], + "angle": 0, + "content": "Geon Yeong Park1, Chanyong Jung1, Sangmin Lee2, Jong Chul Ye1,2,3* Sang Wan Lee1,4*" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.198, + 0.697, + 0.216 + ], + "angle": 0, + "content": "\\(^{1}\\)Bio and Brain Engineering, \\(^{2}\\)Mathematical Sciences," + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.216, + 0.756, + 0.234 + ], + "angle": 0, + "content": "3Kim Jaechul Graduate School of AI, 4Brain and Cognitive Sciences" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.235, + 0.799, + 0.252 + ], + "angle": 0, + "content": "Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.255, + 0.744, + 0.27 + ], + "angle": 0, + "content": "{pky3436, jcy132, leeeeesang,jong.ye,sangwan}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Spurious correlations can cause strong biases in deep neural networks, impairing generalization ability. While most existing debiasing methods require full supervision on either spurious attributes or target labels, training a debiased model from a limited amount of both annotations is still an open question. To address this issue, we investigate an interesting phenomenon using the spectral analysis of latent representations: spuriously correlated attributes make neural networks inductively biased towards encoding lower effective rank representations. We also show that a rank regularization can amplify this bias in a way that encourages highly correlated features. Leveraging these findings, we propose a self-supervised debiasing framework potentially compatible with unlabeled samples. Specifically, we first pretrain a biased encoder in a self-supervised manner with the rank regularization, serving as a semantic bottleneck to enforce the encoder to learn the spuriously correlated attributes. This biased encoder is then used to discover and upweight bias-conflicting samples in a downstream task, serving as a boosting to effectively debias the main model. Remarkably, the proposed debiasing framework significantly improves the generalization performance of self-supervised learning baselines and, in some cases, even outperforms state-of-the-art supervised debiasing approaches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.472, + 0.902 + ], + "angle": 0, + "content": "While modern deep learning solves several challenging tasks successfully, a series of recent works [16, 18, 20] have reported that the high accuracy of deep networks on indistribution samples does not always guarantee low test error on out-of-distribution (OOD) samples, especially in the context of spurious correlations. Existing studies [3, 38, 50] suggest that the deep networks can be potentially biased to the spuriously correlated attributes, or dataset bias, which are misleading statistical heuristics that are closely correlated but not causally related to the target label." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.412 + ], + "angle": 0, + "content": "These catastrophic pitfalls of dataset bias have facilitated the development of debiasing methods, which can be roughly categorized into approaches: (1) leveraging annotations of spurious attributes, i.e., bias label [27, 46, 49, 56]; (2) presuming specific type of bias, e.g., color and texture [5, 17, 53]; or (3) without using explicit kinds of supervisions on dataset bias [30, 31, 35, 39, 60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.896, + 0.625 + ], + "angle": 0, + "content": "While substantial advances have been made in this regard, these approaches still fail to address the problem: how to train a debiased classifier by fully exploiting unlabeled samples lacking both bias and target label. More specifically, while the large-scale unlabeled dataset can be potentially biased towards spuriously correlated sensitive attributes, e.g., ethnicity, gender, or age [1, 2], current existing debiasing frameworks are not designed to deal with this real-world unsupervised settings. Here we also confirm that most supervised debiasing frameworks suffer from performance degradation in the low-labeled data setting. Moreover, recent works have suggested that self-supervised learning might not be sufficient to deal with OOD generalization [10, 19, 44] when dataset bias remains after data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.897, + 0.839 + ], + "angle": 0, + "content": "To tackle this issue, we first make a series of empirical observations that allow us to examine the fundamental difference between biased and unbiased representations. Interestingly, we found that spurious correlations suppress the effective rank [45] of latent representations, which severely deteriorates the semantic diversity of representations and leads to the degradation of feature discriminability. Another notable aspect of our findings is that the intentional increase of feature redundancy amplifies \"prejudice\" in neural networks. To be specific, as we enforce the correlation among latent features to regularize the effective rank of representations (i.e., rank regularization), the accuracy on bias-conflicting samples quickly declines while the model still performs reasonably well on the bias-aligned \\(^1\\) samples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.851, + 0.895, + 0.9 + ], + "angle": 0, + "content": "1The bias-aligned samples refer to data with a strong correlation between (potentially latent) spurious features and target labels. The bias-conflicting samples refer to the opposite cases where spurious correlations do not exist." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12395" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.093, + 0.837, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Figure 1. Concept. Based on the bias-rank relationship (Section 2), we introduce a novel debiasing framework centered on rank regularization, which intentionally amplifies spurious correlation by enforcing feature components to be entangled with both spurious and invariant attributes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.569 + ], + "angle": 0, + "content": "Based on these observations, we propose a novel debiasing framework that can utilize both labeled and unlabeled biased samples with rank regularization. The proposed method is fully compatible with both supervised and self-supervised scenarios, where such compatibility arises from the rank regularization that does not rely on any labels. Specifically, for a supervised (self-supervised) setting, we train 1) a biased classifier (encoder) with rank regularization, which serves as a semantic bottleneck limiting the semantic diversity of feature components, and 2) the main classifier (encoder) with standard (self-)supervised learning approaches. The biased model affords us the leverage to uncover spurious correlations and identify bias-conflicting samples in a downstream task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.472, + 0.722 + ], + "angle": 0, + "content": "Our work is the first to unveil the bias-rank relationships and introduce an effective debiasing strategy to exploit potentially unlabeled data samples. We demonstrate the effectiveness of the proposed debiasing framework with various challenging real-world biased datasets, including MultiCM-NIST [34], biased Chest X-ray databases, UTKFace, CelebA, etc., in both a supervised and self-supervised scenario. These experiments show that our method significantly outperforms other self-supervised baselines, and even state-of-the-art supervised debiasing methods in some cases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.439, + 0.75 + ], + "angle": 0, + "content": "2. Low-rank bias of biased representations" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.218, + 0.772 + ], + "angle": 0, + "content": "2.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Throughout the paper, we denote \\( x \\in \\mathbb{R}^m \\) and \\( y \\in \\mathcal{V} \\) as \\( m \\)-dimensional input sample and its corresponding predicting label, respectively. Then we denote \\( X = \\{x_k\\}_{k=1}^n \\) as a batch of \\( n \\) samples from a dataset which is fed to an encoder \\( f_\\theta : \\mathbb{R}^m \\to \\mathbb{R}^d \\), parameterized by \\( \\theta \\). Then we construct a matrix \\( Z \\in \\mathbb{R}^{n \\times d} \\) where each \\( i \\)th row is the output representations of the encoder \\( f_\\theta(x_i)^T \\) for \\( x_i \\in X \\). For every analysis in this section, we use \\( Z \\) as our latent representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.404 + ], + "angle": 0, + "content": "tions, where the neural backbone of the encoder may vary as simple convolutional networks, ResNet-18, or ViT [15] (Experimental details provided in Appendix C.1 and D)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.893, + 0.48 + ], + "angle": 0, + "content": "To evaluate the semantic diversity of given representation matrix, we introduce effective rank [45] which is a widely used metric to measure the effective dimensionality of matrix and analyze the spectral properties of features in neural networks [4, 6, 23, 42]:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.894, + 0.534 + ], + "angle": 0, + "content": "Definition 1. Given the matrix \\( X \\in \\mathbb{R}^{m \\times n} \\) and its singular values \\( \\{\\sigma_i\\}_{i=1}^{\\min(m,n)} \\), the effective rank \\( \\rho \\) of \\( X \\) is defined as the shannon entropy of normalized singular values:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.542, + 0.893, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\rho (X) = - \\sum_ {i = 1} ^ {\\min (m, n)} \\bar {\\sigma} _ {i} \\log \\bar {\\sigma} _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.895, + 0.638 + ], + "angle": 0, + "content": "where \\(\\bar{\\sigma}_i = \\sigma_i / \\sum_k\\sigma_k\\) is \\(i\\)-th normalized singular value. Without loss of generality, we omit the exponentiation of \\(\\rho (X)\\) as done in [45]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Effective rank is also referred to as spectral entropy where its value is maximized when the singular values are all equal and minimized when a top singular value dominates relative to all others. Recent works [12, 13] have revealed that the discriminability of representations resides on wide range of eigenvectors since the rich discriminative information for the classification task cannot be transmitted by only few eigenvectors with top singular values. Thus from a spectral analysis perspective, effective rank quantifies how diverse the semantic information encoded by each eigenfeature is, which is closely related to the feature discriminability across target label categories. In the rest of paper, we interchangeably use effective rank and rank by following prior works." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.892, + 0.865 + ], + "angle": 0, + "content": "2.2. Spectral analysis of the bias-rank relationships" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We now present experiments showing that the deep networks may tend to encode lower-rank representations in the pres" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.252, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.187, + 0.214, + 0.198 + ], + "angle": 0, + "content": "(a) Unbiased corr." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.092, + 0.468, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.185, + 0.428, + 0.196 + ], + "angle": 0, + "content": "(b) Biased corr." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.213, + 0.253, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.301, + 0.212, + 0.312 + ], + "angle": 0, + "content": "(c) Effective rank" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.201, + 0.468, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.314, + 0.412, + 0.326 + ], + "angle": 0, + "content": "(d) SVD analysis" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.473, + 0.45 + ], + "angle": 0, + "content": "Figure 2. Empirical analysis on rank reduction phenomenon. For every analysis, we used the output \\( Z \\) of the encoder (Sec. 2.1). (a, b): Hierarchically clustered auto-correlation matrix of unbiased and biased representations (Bias ratio=99%). (c): Effective rank with color bias. 'Unbiased' represents the case training model with perfectly unbiased dataset, i.e. random color for each sample. (d): SVD analysis with max-normalized singular values. Top 100 values are shown in the figure (total: 256)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.599 + ], + "angle": 0, + "content": "ence of stronger spurious correlations. To arbitrarily control the degree of spurious correlations, we introduce synthetic biased datasets, Color-MNIST (CMNIST) and Corrupted CIFAR-10 (CIFAR-10C, [21]), with color and corruption bias types, respectively. We define the degree of spurious correlations as the ratio of bias-aligned samples included in the training set, or bias ratio, where most of the samples are bias-aligned in the context of strong spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Figure 2c shows that the rank of latent representations from a penultimate layer of the simple convolutional classifier decreases as the bias ratio increases in CMNIST. We provide similar rank reduction results of CIFAR-10C with ResNet-18 and ViT in the Appendix C.1. We further compare the correlation matrix of biased and unbiased latent representations in the penultimate layer of biased and unbiased classifiers, respectively. In Figure 2a and 2b, we observe that the block structure in the correlation matrix is more evident in the biased representations after the hierarchical clustering, indicating that the features become highly correlated which may limit the semantic diversity of networks. To investigate the rank reduction phenomenon in-depth, we compare the normalized singular values of biased and unbiased representations. We conduct singular value decomposition (SVD) on the feature matrices of both biased and unbiased classifiers and plot the singular values normalized by the spectral norm of the corresponding matrix. Figure 2d shows that the top few normalized singular values of biased representations are similar to or even greater than those of unbiased rep" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "resentations. However, the remaining majority of singular values decay significantly faster in biased representations, greatly weakening the informative signals of eigenvectors with smaller singular values and deteriorating feature discriminability [12, 13]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.176, + 0.692, + 0.192 + ], + "angle": 0, + "content": "2.3. Rank regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Motivated from the aforementioned rank reduction phenomenon, we ask an opposite-directional question: \"Can we intentionally amplify the prejudice of deep networks by maximizing the redundancy between the components of latent representations?\" If the feature components are extremely correlated, the corresponding representations may exhibit most of its spectral energy along the direction of one singular vector. For this case, effective rank may converge to 0. In other words, our goal is to design a semantic bottleneck of representations that restricts the semantic diversity of feature vectors. To implement the bottleneck in practice, motivated from Figure 2b, we compute the auto-correlation matrix of the output of encoder." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.396, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Let \\(\\bar{Z}\\) denote the mean-centered representations \\(Z\\) along the batch dimension. The normalized auto-correlation matrix \\(C\\in \\mathbb{R}^{d\\times d}\\) of \\(\\bar{Z}\\) is defined as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.448, + 0.892, + 0.493 + ], + "angle": 0, + "content": "\\[\nC _ {i, j} = \\frac {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , i} \\bar {Z} _ {b , j}}{\\sqrt {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , i} ^ {2}} \\sqrt {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , j} ^ {2}}} \\quad 1 \\leq \\forall i, j \\leq d, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.548 + ], + "angle": 0, + "content": "where \\( b \\) is an index of sample and \\( i, j \\) are index of each vector dimension. Then we define our regularization term as the negative of a sum of squared off-diagonal terms in \\( C \\):" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.571, + 0.892, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\ell_ {r e g} (X; \\theta) = - \\sum_ {i} \\sum_ {j \\neq i} C _ {i, j} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.643 + ], + "angle": 0, + "content": "where we refer to it as the rank loss. Note that the target labels on \\( X \\) is not used at all." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Analysis of rank-regularized networks. To investigate the impacts of rank regularization in deep neural networks, we construct the classification model by combining the linear classifier \\( f_{W}:\\mathbb{R}^{d}\\to \\mathbb{R}^{c} \\) parameterized by \\( W\\in \\mathcal{W} \\) on top of the encoder \\( f_{\\theta} \\), where \\( c = |\\mathcal{V}| \\) is the number of classes. Then we trained models by cross entropy loss \\( \\ell_{CE} \\) combined with \\( \\lambda_{reg}\\ell_{reg} \\), where \\( \\lambda_{reg} > 0 \\) is a Lagrangian multiplier. We use CMNIST, CIFAR-10C, and Waterbirds dataset [51], and evaluate the trained models on an unbiased test set following [30, 39]. After training models with varying the hyperparameter \\( \\lambda_{reg} \\), we compare bias-aligned and bias-conflict accuracy, which are the average accuracy on bias-aligned and bias-conflicting samples in the unbiased test set, respectively, for CMNIST and CIFAR-10C. Test accuracy on every individual data group is reported for Waterbirds. Figure 3 shows that models suffer more from poor OOD generalization as trained with larger \\( \\lambda_{reg} \\). The average" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12397" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.322, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.225, + 0.236, + 0.237 + ], + "angle": 0, + "content": "(a) CMNIST" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.091, + 0.604, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.225, + 0.526, + 0.236 + ], + "angle": 0, + "content": "(b) CIFAR-10C" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.092, + 0.884, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.225, + 0.802, + 0.237 + ], + "angle": 0, + "content": "(c) Waterbirds" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.894, + 0.278 + ], + "angle": 0, + "content": "Figure 3. (a, b): Bias-conflict and Bias-aligned accuracy on CMNIST and CIFAR-10C (Bias ratio=95%). (c): Group accuracy on Waterbirds. Detailed simulation settings are in the Appendix D." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.302, + 0.428, + 0.386 + ], + "angle": 0, + "content": "
CMNISTCIFAR-10C
P (%)R (%)P (%)R (%)
ERM85.5919.7652.030.06
+ Rank reg98.8395.9171.3951.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.205, + 0.388, + 0.345, + 0.4 + ], + "angle": 0, + "content": "(a) CMNIST, CIFAR-10C" + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.401, + 0.415, + 0.464 + ], + "angle": 0, + "content": "
MetricsERMJTTRank reg
Precision (%)37.8448.9554.77
Recall (%)11.6748.7555.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.467, + 0.313, + 0.479 + ], + "angle": 0, + "content": "(b) Waterbirds" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Table 1. Precision (P) and Recall (R) of bias-conflicting samples. (a): Bias-conflicting samples are identified in the error set of ERM model trained with and without rank regularization (Bias ratio=95% for both datasets). (b): Bias-conflicting samples are similarly identified by ERM, JTT, and the proposed biased model in Waterbirds dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.669 + ], + "angle": 0, + "content": "accuracy on bias-conflicting groups is significantly degraded, while the accuracy on bias-aligned groups is maintained to some extent. It implies that rank regularization may force deep networks to focus on spurious attributes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.471, + 0.883 + ], + "angle": 0, + "content": "Minority mining performance. Table 1a and 1b support that the biased models with strong regularization can effectively probe out the bias-conflicting samples in the training set. Specifically, we train a biased classifier with rank regularization and distill an error set \\( E \\) of misclassified training samples as bias-conflicting samples proxies. As reported in Table 1a, we observe that our biased classifier is relatively robust to the unintended memorization of bias-conflicting samples [47] in contrast to the standard models trained by Empirical Risk Minimization (ERM). Moreover, Table 1b shows that the proposed rank regularization improves the precision and recall of identified bias-conflicting samples compared to JTT [35]. Detailed simulation settings are in the Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Reconstruction of biased representations. To under" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.576 + ], + "angle": 0, + "content": "stand the relationship between rank regularization and spurious correlations more deeply, we visualize the pretrained representations with varying degrees of bias. We first trained deep networks on: (a) unbiased CMNIST (random background color), (b) biased CMNIST (bias ratio \\(= 95\\%\\)) without rank regularization and (c) with rank regularization \\((\\lambda_{reg} = 50)\\). Then, we train the auxiliary decoder, which reconstructs the bias-conflicting images from the freezed latent representations of each pretrained network. Results show that rank regularization may cause the representation to lose information on complex invariant features, resulting in a loss of feature discriminability and informative signals. While both digit and color are well reconstructed with biased representations (b), the decoder fails to reconstruct bias-conflicting images from the (c) biased representations pretrained with rank regularization. The foreground digit is blurred, and its class changes following the color-digit assignment in Figure 4d." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.744 + ], + "angle": 0, + "content": "These observations afford us some key insights into rank regularization: First, the rank-regularized representation may lose its information on complex invariant features (i.e., shape and style of the foreground digit), specifically undermining the feature discriminability and informative signals. Second, the limited semantic diversity makes it harder to identify the true underlying independent generative factors for multidimensional data; instead, it may encode feature components entangled with both spurious and invariant attributes as the digit class of the reconstructed image is erroneously determined by the background color in 4c." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.894, + 0.869 + ], + "angle": 0, + "content": "Multiple bias attributes. To further investigate the generalizability of rank regularization, we evaluate the biased representations with Multi-Color MNIST (MultiCMNIST) dataset [34], which is similar to the CMNIST but have two bias attributes: left and right background colors. We set bias ratio \\(= 99\\%\\) for the left color and bias ratio \\(= 95\\%\\) for the right color, i.e., the left color is a more salient bias than the right color (Dataset details are provided in Appendix D)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Table 2 shows that the rank regularization successfully biases the model w.r.t both bias attributes, while LfF [39]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12398" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.311, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.171, + 0.231, + 0.183 + ], + "angle": 0, + "content": "(a) Unbiased" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.089, + 0.541, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.171, + 0.455, + 0.183 + ], + "angle": 0, + "content": "(b) Biased" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.089, + 0.771, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.171, + 0.721, + 0.183 + ], + "angle": 0, + "content": "(c) Biased with rank reg." + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.091, + 0.89, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.168, + 0.867, + 0.18 + ], + "angle": 0, + "content": "(d) Examples." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.195, + 0.894, + 0.252 + ], + "angle": 0, + "content": "Figure 4. Randomly selected reconstructed images from representations with varying degrees of bias. First and third row correspond to the input bias-conflicting images. Second and fourth row correspond to the reconstructed images. Reconstructed from (a) unbiased representations, (b) biased representations, and (c) biased representations with rank regularization (bias ratio=95% in b, c). (d) Examples of bias-aligned CMNIST images." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.402 + ], + "angle": 0, + "content": "Table 2. (a) Test accuracy (\\%) on MultiCMNIST. Lower is better for this results. BC for bias-conflicting, and BA for bias-aligned. Bias ratio=99(\\%) for left color, and 95(\\%) for right color. \\(\\lambda_{reg} = 50\\) is used for rank regularization. (b) Debiasing results. Higher is better for this results. Baseline results are from [34]. \\(\\lambda_{up} = 50\\) is used for upweighting in the proposed framework (\\(\\lambda_{up}\\): a manual rescaling weight given to each identified bias-conflicting samples in cross entropy loss). Pseudo-code and experimental details are provided in Appendix A and D, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.412, + 0.466, + 0.501 + ], + "angle": 0, + "content": "
IdxLeft colorRight color(a) Biased accuracy (%)(b) Debiased accuracy (%)
ERMLfF [39]Rank reg.LFFDebiAN [34]Ours
(1)BABC100.0100.0100.099.6100.0100.0
(2)BABC96.698.841.64.795.697.0
(3)BCBA29.33.28.798.676.579.1
(4)BCBC7.61.36.15.116.018.3
(1) ~ (4) average acc.58.3850.8339.152.072.073.6
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.646 + ], + "angle": 0, + "content": "completely fails to amplify the right color bias, i.e. less salient bias, as shown in the second row (Biased accuracy part). This leads to the abnormal debiasing results of LfF as shown in Table 2 where it records unbalanced accuracy for the left- and right-color-bias-conflicting samples. In contrast, the proposed framework shows superior performance by simply weighting the misclassified bias-conflicting proxies, as done in [35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.471, + 0.768 + ], + "angle": 0, + "content": "Taken together, these results indicate that the rank regularization encourages the network to focus more on spurious correlations in a way that minimizes semantic diversity and entangles invariant and spurious features [41], which is a fundamentally different mechanism compared to the LfF [39] with its easy-to-learn assumption. More details on the upweighting strategy will be provided in Section 3 and pseudo-code in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.782, + 0.473, + 0.816 + ], + "angle": 0, + "content": "3. DeFund: Debiasing framework with unlabeled data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Motivated by the observations in Section 2, we propose a self-supervised debiasing framework with unlabeled data, coined DeFund (Debiasing Framework with Unlabeled Data). A notable distinction from previous studies [5, 60] lies in the proposed framework's ability to effectively harness un" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.893, + 0.324 + ], + "angle": 0, + "content": "labeled data for learning biased representations. This is achieved through the application of self-supervised learning and rank regularization techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.895, + 0.448 + ], + "angle": 0, + "content": "The proposed framework is composed of two stages: We first train the biased encoder, which can be potentially adopted to detect the bias-conflicting samples in a downstream task, along with the main encoder by self-supervised learning, both without any labels. After pretraining, we identify the bias-conflicting samples in the downstream task using linear evaluation protocol [9, 40]. This set of samples serves as a boosting to debias the main model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.895, + 0.685 + ], + "angle": 0, + "content": "Notation. We denote \\( f_{\\theta}^{bias} : \\mathcal{X} \\to \\mathbb{R}^d \\) and \\( f_{\\phi}^{main} : \\mathcal{X} \\to \\mathbb{R}^d \\) as biased encoder and main encoder parameterized by \\( \\theta \\in \\Theta \\) and \\( \\phi \\in \\Theta \\), respectively, where \\( d \\) is the dimensionality of latent representations. Then we can compute the rank loss in (3) with introduced encoders and given batch \\( \\{x_k\\}_{k=1}^N \\) with size \\( N \\). Let \\( f_{W_b}^{cls} : \\mathbb{R}^d \\to \\mathbb{R}^C \\) be a single-layer classifier parameterized by \\( W_b \\in \\mathcal{W} \\) which is placed on top of biased encoder \\( f_{\\theta}^{bias} \\), where \\( C = |\\mathcal{Y}| \\) is the number of classes. We similarly define the linear classifier \\( f_{W_m}^{cls} \\) for the main encoder. Then we refer to \\( f^{bias} : \\mathcal{X} \\to \\mathbb{R}^C \\) as biased model, where \\( f^{bias}(x) = f_{W_b}^{cls}\\big(f_{\\theta}^{bias}(x)\\big), \\forall x \\in \\mathcal{X} \\). We similarly define the main model \\( f^{main} \\) as \\( f^{main}(x) = f_{W_m}^{cls}\\big(f_{\\phi}^{main}(x)\\big), \\forall x \\in \\mathcal{X} \\). While the projection networks [9] are employed as well, we omit the notations because they are not engaged in classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Stage 1. Training a biased encoder. To train the biased encoder \\( f_{\\theta}^{bias} \\), we revisit the proposed rank regularization term (3) in context of instance discrimination task. Building upon the observations in Section 2.3, we conjecture that rank regularization may amplify bias in self-supervised learning as well by entangling invariant and spurious features. Based on these intuitions, we apply rank regularization directly to the output of the base encoder, which encourages each feature component to be highly correlated. From these applications, several noteworthy observations have emerged: (a) The representation becomes more biased as it is trained with stronger regularization (Appendix C.1). (b) While the overall performance may be upper-bounded due to the constraint on effective dimensionality [25], the bias-conflict accuracy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12399" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "is primarily sacrificed compared to the bias-aligned accuracy (Section 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Stage 2. Debiasing downstream tasks. After training the biased encoder, our goal is to debias the main model, which was pretrained using standard self-supervised learning methods on the same dataset. Here, assume that we have an ideal pretrained main encoder of which each output component corresponds to the latent factor of data variation [63]. While this ideal encoder should seamlessly adapt to downstream classification tasks, if most downstream task samples are bias-aligned, they may misguide the model to upweight spuriously correlated latent factors, leading to a biased solution despite well-generalized representations. We refer to this problem as the biased downstream application (Related analysis in Appendix C.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.453 + ], + "angle": 0, + "content": "The above contradiction elucidates the importance of bias-conflicting samples, which serve as counterexamples of spuriously correlated feature components, thereby preventing the alleged involvement of such components in prediction. Based on these intuitions, we introduce a novel debiasing protocol that probes and upweights bias-conflicting samples to find and fully exploit feature components independent of spurious correlations. We apply our framework on two scenarios: linear evaluation and semi-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.471, + 0.668 + ], + "angle": 0, + "content": "Linear evaluation. To validate our hypothesis on the biased downstream application, we conduct linear evaluation [40, 61] following the conventional protocol of self-supervised learning. Specifically, a linear classifier is trained on top of unsupervised pretrained representations by using target labels of training samples. After training a linear classifier \\( f_{W_b}^{cls} \\) with pretrained biased encoder \\( f_{\\theta}^{bias} \\) given the whole training set \\( D = \\{(x_k, y_k)\\}_{k=1}^N \\) with size \\( N \\), an error set \\( E \\) of misclassified samples and corresponding labels is regarded as bias-conflicting pairs. Then we train a linear classifier \\( f_{W_m}^{cls} \\) on intentionally frozen representations of main encoder \\( f_{\\phi}^{main} \\) by weighting the identified samples in \\( E \\) with \\( \\lambda_{up} > 0 \\). The loss function for debiased linear evaluation is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.676, + 0.45, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\ell_ {d e b i a s} (D; W _ {m}) \\\\ = \\lambda_ {u p} \\sum_ {(x, y) \\in E} \\ell (x, y; W _ {m}) + \\sum_ {(x, y) \\in D \\setminus E} \\ell (x, y; W _ {m}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.78 + ], + "angle": 0, + "content": "where we use cross entropy loss for \\(\\ell : \\mathcal{X} \\times \\mathcal{Y} \\times \\mathcal{W} \\to \\mathbb{R}^+\\). Note that the target labels are only used in training linear classifiers after pretraining." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Note that the debiased linear evaluation is not meant to compete directly with other supervised baselines. Instead, it aims to: (a) examine the potential origin of the failure in OOD generalization, (b) provide a rough estimate of the potential improvement achievable with frozen latent representations, and (c) compare with standard self-supervised baselines and identify the optimal learning algorithms, e.g. SimCLR [9], for training the main encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Semi-supervised learning. We further compare our method directly to other supervised debiasing methods in the context of semi-supervised learning. Here we assume that the training dataset includes only a small amount of labeled data combined with a large amount of unlabeled data. As in linear evaluation, we train a linear classifier on top of the biased encoder by using labeled samples. After obtaining an error set \\( E \\) of misclassified samples, we finetuned the whole main model by weighting the identified samples in \\( E \\) with \\( \\lambda_{up} \\). Note that supervised baselines are restricted to using only a small fraction of labeled samples, while the proposed approach benefits from the abundant unlabeled samples during pre-training of the biased encoder (Pseudo-code in the Appendix section A)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.587, + 0.335 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.345, + 0.604, + 0.359 + ], + "angle": 0, + "content": "4.1. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.894, + 0.445 + ], + "angle": 0, + "content": "Dataset. We evaluate several supervised and self-supervised baselines on MIMIC-CXR + NIH [32], UTKFace [62] and CelebA [36] in which prior work reported poor generalization performance due to spurious correlations (Dataset details in Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.894, + 0.597 + ], + "angle": 0, + "content": "For MIMIC-CXR + NIH, we mixed the MIMIC-CXR [26] and NIH [55] following [32] where the target categories are no finding and pneumonia. Most pneumonia images are collected from MIMIC-CXR, while most no finding images are from NIH. In other words, the biases come from systematic differences in data sources, where the classifier may erroneously rely on spurious radiographic features tied to variations in data acquisition pipelines [14] instead of true pathological indicators (Example images in Figure 8)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.794 + ], + "angle": 0, + "content": "For UTKFace, we conduct binary classifications using (Gender, Age) and (Race, Gender) as (target, spurious) attribute pair, which we refer to UTKFace (age) and UTKFace (gender), respectively. For CelebA, we consider (HeavyMakeup, Male) and (Blonde Hair, Male) as (target, spurious) attribute pairs, which are referred to CelebA (makeup) and CelebA (blonde), respectively. The results of CelebA (blonde) are reported in Appendix C.4. Following [22, 39], we report bias-conflict accuracy together with unbiased accuracy, which is evaluated on the explicitly constructed validation set. We exclude the dataset in Figure 3 based on the observations that the SimCLR models are already invariant w.r.t spurious attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Baselines. We mainly target baselines consisting of recent advanced self-supervised learning methods, SimCLR [9], VICReg [7], and SimSiam [11], which can be categorized into contrastive (SimCLR) and non-contrastive (VICReg, SimSiam) methods. We further report the performance of vanilla networks trained by ERM, and other supervised debiasing methods such as LNL [27], EnD [49], and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12400" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 3. (Supervised learning) Bias-conflict and unbiased accuracy (%) on MIMIC-CXR + NIH. Each \\(\\checkmark\\) marker represents whether the model requires information on dataset bias. Bias ratio=10%." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.129, + 0.887, + 0.224 + ], + "angle": 0, + "content": "
AccuracyLNL √EnD √LfF XJTT XCVaR DRO XERM XSimCLR XOurs X
Conflict43.8±0.550.4±2.325.2±2.147.9±0.244.6±0.541.7±1.235.5±1.356.8±1.7
Unbiased68.1±1.071.8±1.460.8±0.268.9±1.065.8±1.267.8±1.062.0±1.469.8±0.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.235, + 0.893, + 0.279 + ], + "angle": 0, + "content": "Table 4. (Linear evaluation) Bias-conflict and unbiased test accuracy (%) evaluated on UTKFace and CelebA. Models requiring information on target class or dataset bias in the (pre)training stage are denoted with √in columns Y and B, respectively. Our results are marked in bold to highlight the improvements compared to the mainly interested self-supervised learning baselines (Gray rows)." + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.288, + 0.806, + 0.49 + ], + "angle": 0, + "content": "
ModelYBUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
LNL✓✓45.8±0.672.6±0.373.1±1.684.9±0.855.9±2.176.0±0.6
EnD✓✓45.3±0.972.2±0.275.5±1.185.5±0.457.3±2.476.4±1.4
JTT✓✗63.8±0.969.4±1.371.2±0.377.6±0.462.4±1.274.7±0.8
CVaR DRO✓✗45.7±2.071.4±0.368.6±1.081.0±0.858.0±1.776.5±0.6
ERM✓✗45.4±2.171.0±1.265.7±1.479.5±0.654.2±0.274.1±1.4
SimSiam✗✗28.2±0.962.6±0.748.5±1.069.8±0.739.9±0.666.7±0.6
VICReg✗✗32.3±0.664.6±0.351.0±1.471.3±0.748.6±0.671.9±0.2
SimCLR✗✗36.4±1.566.3±0.656.3±0.274.2±0.246.9±1.069.8±0.4
DeFund✗✗59.5±0.870.6±0.863.7±2.074.9±0.958.4±0.673.1±1.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.576 + ], + "angle": 0, + "content": "upweighting-based algorithms, JTT [35] and CVaR DRO [31], which can be categorized into methods that leverage annotations on dataset bias (LNL, EnD) or not (JTT, CVaR DRO)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.471, + 0.759 + ], + "angle": 0, + "content": "Optimization setting. Both bias and main encoder is pretrained with SimCLR [9] for 100 epochs on UTKFace, and 20 epochs on CelebA, respectively, using ResNet-18, Adam optimizer and cosine annealing learning rate scheduling [37]. We use a MLP with one hidden layer for projection networks as in SimCLR. All the other baseline results are reproduced by tuning the hyperparameters and optimization settings using the same backbone architecture. We report the results of the model with the highest bias-conflicting test accuracy over those with improved unbiased test accuracy compared to the corresponding baseline algorithms, i.e., SimCLR for ours (More experimental details in Appendix D)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.254, + 0.787 + ], + "angle": 0, + "content": "4.2. Evaluation results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Supervised learning. To quantify the effectiveness of the rank regularization in-depth, we first consider a standard supervised debiasing scenario as similarly done in Table 2. For a MIMIC-CXR + NIH dataset, we found that the proposed framework outperforms other supervised baselines with respect to bias-conflict accuracy. Table 14 in the Appendix shows that the rank-regularized networks effectively" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.545 + ], + "angle": 0, + "content": "discover the bias-conflicting samples which are consistent with Table 1a, 1b, and 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.791 + ], + "angle": 0, + "content": "Linear evaluation. We also found that DeFund outperforms every self-supervised baseline by a large margin in a linear evaluation protocol, including SimCLR, SimSiam and VICReg, with respect to both bias-conflict and unbiased accuracy (Table 4). Moreover, in some cases, DeFund even outperforms ERM models or supervised debiasing approaches regarding bias-conflict accuracy. Note that there is an inherent gap between ERM models and self-supervised baselines, roughly \\(8.7\\%\\) on average. Moreover, we found that non-contrastive learning methods generally perform worse than the contrastive learning method. This warns us against training the main model using a non-contrastive learning approach, while it may be a viable option for the biased model. Results of the proposed framework with non-contrastive learning methods are provided in the Appendix section C.5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Semi-supervised learning. To compare supervised and self-supervised methods in a more practical and fair scenario, we randomly sample \\(10\\%\\) of the labeled CelebA training dataset. The remaining \\(90\\%\\) samples are treated as unlabeled ones and engaged only in pretraining encoders for self-supervised baselines. Labeled samples are provided equally to both supervised and self-supervised methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12401" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.21, + 0.089, + 0.759, + 0.104 + ], + "angle": 0, + "content": "Table 5. (Semi-supervised learning) Accuracy results (%) on CelebA. Label fraction \\( = {10}\\% \\) ." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.115, + 0.865, + 0.201 + ], + "angle": 0, + "content": "
AccuracyCelebA (Makeup)CelebA (Blonde)
LNLEnDJTTCVaR DROERMSimCLRDeFundJTTDeFund
Conflict55.7±1.455.3±1.551.5±1.955.6±1.551.5±1.150.5±4.760.5±0.470.6±1.075.1±0.8
Unbiased75.6±0.576.2±0.871.4±1.375.7±1.073.1±0.371.6±1.975.6±0.278.8±1.785.8±0.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.226, + 0.473, + 0.409 + ], + "angle": 0, + "content": "Remarkably, Table 5 and Table 16 in Appendix show that the proposed framework outperforms other state-of-the-art supervised debiasing methods. Existing upweighting protocols, such as JTT, fail to prevent deep networks from memorizing minority counterexamples. However, the proposed framework can fully utilize unlabeled samples with contrastive learning to prevent memorization. Existing bias-conflicting sample mining algorithms may be affected by the implicit bias of overparameterized networks, but this is unlikely to happen with the proposed framework since it only trains a simple linear classifier on top of a frozen biased encoder to identify such samples." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.42, + 0.472, + 0.5 + ], + "angle": 0, + "content": "
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
SimCLR36.466.356.374.246.969.8
+ Rank reg26.661.350.970.343.968.3
+ Upweight53.064.658.374.550.170.4
DeFund59.570.663.774.958.473.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.501, + 0.329, + 0.513 + ], + "angle": 0, + "content": "(a) Ablation study" + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.513, + 0.472, + 0.581 + ], + "angle": 0, + "content": "
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
PrecisionRecallPrecisionRecallPrecisionRecall
SimCLR68.3144.6333.3639.5952.2528.23
DeFund68.6775.9429.9850.9355.2932.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.583, + 0.342, + 0.595 + ], + "angle": 0, + "content": "(b) Precision and recall" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.649 + ], + "angle": 0, + "content": "Table 6. (a) Ablation study on introduced modules. (b) Precision and recall (\\%) of bias-conflicting samples identified by SimCLR and our biased model. Both case used linear evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Ablation study. To quantify the extent of performance improvement achieved by each introduced module, we compared the linear evaluation results of (a) vanilla SimCLR, (b) SimCLR with rank regularization, (c) SimCLR with upweighting error set \\( E \\) of the main model, and (d) DeFund. Note that (c) does not use a biased model at all. Table 6a shows that every module plays an important role in OOD generalization. Considering that the main model is already biased to some extent, we found that bias-conflict accuracy can be improved even without a biased model, where the error set \\( E \\) of the biased model further boosts the generalization performance. We also measures the precision and recall of identified bias-conflicting samples in E, finding that the biased model detects more diverse bias-conflicting samples than the baseline (Table 6b). The improvement of recall in CelebA may seem marginal, but it is significant given the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.829, + 0.241 + ], + "angle": 0, + "content": "larger number of samples compared to UTKFace." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.895, + 0.333 + ], + "angle": 0, + "content": "Computational costs. Our framework is computationally affordable as it only trains the linaer classifier (linear eval.) or finetune networks with a few epochs, e.g., about 30 epochs for UTKFace in debiasing stage. Self-supervised pre-training and linear evaluation takes 19.3 and 4.5 minutes with a NVIDIA GeForce RTX 2080Ti, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.34, + 0.62, + 0.356 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.895, + 0.502 + ], + "angle": 0, + "content": "We present a novel solution to the challenging self-supervised debiasing, an important problem that has received little attention so far. We (a) unveil the inductive bias towards encoding low effective rank representations in the presence of spurious correlations. Based on this, we (b) design a rank regularization that amplifies the feature redundancy by reducing the spectral entropy of latent representations. Then we (c) design a debiasing framework empowered by the biased model pretrained with abundant unlabeled samples." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.634, + 0.53 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.897, + 0.903 + ], + "angle": 0, + "content": "This research was supported by National Research foundation of Korea(NRF) (\\*\\*RS-2023-00262527**), Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137899, KMDF_PR_20200901_0015), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT, Ministry of Science and ICT) (No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation), (No.2019-0-00075, Artificial Intelligence Graduate School Program(KAIST)), (No. RS-2023-00233251, System3 reinforcement learning with high-level brain functions), Electronics and Telecommunications Research Institute(ETRI) grant funded by the Korean government (24ZS1100, Core Technology Research for Self-Improving Integrated Artificial Intelligence System), [N01230878, Development of Beyond X-verse Core Technology for Hyper-realistic interactions by Synchronizing the Real World and Virtual Space], a grant of the KAIST-KT joint research project through AI2XL Laboratory, Institute of convergence Technology, funded by KT (Project No.3, Project title: Genie Brain)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12402" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Abubakar Abid, Maheen Farooqi, and James Zou. Large language models associate muslims with violence. Nature Machine Intelligence, 3(6):461-463, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Sandhini Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, and Miles Brundage. Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.255 + ], + "angle": 0, + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[4] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[5] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In International Conference on Machine Learning, pages 528-539. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.423 + ], + "angle": 0, + "content": "[6] Aristide Baratin, Thomas George, César Laurent, R Devon Hjelm, Guillaume Lajoie, Pascal Vincent, and Simon Lacoste-Julien. Implicit regularization via neural feature alignment. In International Conference on Artificial Intelligence and Statistics, pages 2269-2277. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[8] Tiffany Tianhui Cai, Jonathan Frankle, David J Schwab, and Ari S Morcos. Are all negatives created equal in contrastive instance discrimination? arXiv preprint arXiv:2010.06682, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.579, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[10] Ting Chen, Calvin Luo, and Lala Li. Intriguing properties of contrastive losses. Advances in Neural Information Processing Systems, 34:11834-11845, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[12] Xinyang Chen, Sinan Wang, Bo Fu, Mingsheng Long, and Jianmin Wang. Catastrophic forgetting meets negative transfer: Batch spectral shrinkage for safe transfer learning. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.734, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[13] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.802, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[14] Alex J DeGrave, Joseph D Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Michael Feldman, Sorelle A Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. Certifying and removing disparate impact. In proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, pages 259-268, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[17] Songwei Ge, Shlok Mishra, Chun-Liang Li, Haohan Wang, and David Jacobs. Robust contrastive learning using negative samples with diminished semantics. Advances in Neural Information Processing Systems, 34:27356-27368, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[18] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[19] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models. arXiv preprint arXiv:2010.08377, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[22] Youngkyu Hong and Eunho Yang. Unbiased classification through bias-contrastive and bias-balanced learning. Advances in Neural Information Processing Systems, 34:26449-26461, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[23] Minyoung Huh, Hossein Mobahi, Richard Zhang, Brian Cheung, Pulkit Agrawal, and Phillip Isola. The low-rank simplicity bias in deep networks. arXiv preprint arXiv:2103.10427, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[24] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[25] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[26] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[27] Byungju Kim, Hyunwoo Kim, Kyungsu Kim, Sungjin Kim, and Junmo Kim. Learning not to learn: Training deep neural networks with biased data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9012-9020, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[28] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12403" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[30] Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, and Jaegul Choo. Learning debiased representation via disentangled feature augmentation. Advances in Neural Information Processing Systems, 34:25123-25133, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.473, + 0.263 + ], + "angle": 0, + "content": "[31] Daniel Levy, Yair Carmon, John C Duchi, and Aaron Sidford. Large-scale methods for distributionally robust optimization. Advances in Neural Information Processing Systems, 33:8847-8860, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.472, + 0.32 + ], + "angle": 0, + "content": "[32] Jiaxuan Li, Duc Minh Vo, and Hideki Nakayama. Partition-and-debias: Agnostic biases mitigation via a mixture of biases-specific experts. arXiv preprint arXiv:2308.10005, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.472, + 0.38 + ], + "angle": 0, + "content": "[33] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14970-14979, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.472, + 0.437 + ], + "angle": 0, + "content": "[34] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In European Conference on Computer Vision, pages 270-288. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.472, + 0.509 + ], + "angle": 0, + "content": "[35] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[36] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.57, + 0.472, + 0.61 + ], + "angle": 0, + "content": "[37] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.472, + 0.668 + ], + "angle": 0, + "content": "[38] Vaishnavh Nagarajan, Anders Andreassen, and Behnam Neyshabur. Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.472, + 0.727 + ], + "angle": 0, + "content": "[39] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.472, + 0.771 + ], + "angle": 0, + "content": "[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[41] Geon Yeong Park, Sangmin Lee, Sang Wan Lee, and Jong Chul Ye. Training debiased subnetworks with contrastive weight pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7929-7938, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[42] Noam Razin and Nadav Cohen. Implicit regularization in deep learning may not be explainable by norms. Advances in neural information processing systems, 33:21174-21187, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.473, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[43] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[44] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[45] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.291 + ], + "angle": 0, + "content": "[46] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[47] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[48] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv preprint arXiv:2206.01197, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.894, + 0.474 + ], + "angle": 0, + "content": "[49] Enzo Tartaglione, Carlo Alberto Barbano, and Marco Grangetto. End: Entangling and disentangling deep representations for bias correction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13508-13517, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.894, + 0.518 + ], + "angle": 0, + "content": "[50] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[52] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.894, + 0.661 + ], + "angle": 0, + "content": "[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.663, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.894, + 0.814 + ], + "angle": 0, + "content": "[55] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[56] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8919-8928, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12404" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[57] Yadollah Yaghoobzadeh, Soroush Mehri, Remi Tachet, Timothy J Hazen, and Alessandro Sordoni. Increasing robustness to spurious correlations using forgettable examples. arXiv preprint arXiv:1911.03861, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[58] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[59] Dinghuai Zhang, Kartik Ahuja, Yilun Xu, Yisen Wang, and Aaron Courville. Can subnetwork structure be the key to out-of-distribution generalization? In International Conference on Machine Learning, pages 12356-12367. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.471, + 0.318 + ], + "angle": 0, + "content": "[60] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[61] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.419, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[63] Roland S Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive learning inverts the data generating process. In International Conference on Machine Learning, pages 12979-12990. PMLR, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12405" + } + ] +] \ No newline at end of file diff --git a/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_origin.pdf b/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9a5da7a1e001d943112e3ccd91350eabfde86d4d --- /dev/null +++ b/2024/Self-supervised Debiasing Using Low Rank Regularization/9a99109e-c17a-4511-ac59-3693469d8cef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:426188ddd1f5298ad5901b6e8fe3f62466fe6ab9eb825d0c92d23e4df31b7444 +size 1167770 diff --git a/2024/Self-supervised Debiasing Using Low Rank Regularization/full.md b/2024/Self-supervised Debiasing Using Low Rank Regularization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d4461bdda258c22d20679ae9e3478be70b59377e --- /dev/null +++ b/2024/Self-supervised Debiasing Using Low Rank Regularization/full.md @@ -0,0 +1,316 @@ +# Self-supervised debiasing using low rank regularization + +Geon Yeong Park1, Chanyong Jung1, Sangmin Lee2, Jong Chul Ye1,2,3* Sang Wan Lee1,4* + +$^{1}$ Bio and Brain Engineering, $^{2}$ Mathematical Sciences, + +3Kim Jaechul Graduate School of AI, 4Brain and Cognitive Sciences + +Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea + +{pky3436, jcy132, leeeeesang,jong.ye,sangwan}@kaist.ac.kr + +# Abstract + +Spurious correlations can cause strong biases in deep neural networks, impairing generalization ability. While most existing debiasing methods require full supervision on either spurious attributes or target labels, training a debiased model from a limited amount of both annotations is still an open question. To address this issue, we investigate an interesting phenomenon using the spectral analysis of latent representations: spuriously correlated attributes make neural networks inductively biased towards encoding lower effective rank representations. We also show that a rank regularization can amplify this bias in a way that encourages highly correlated features. Leveraging these findings, we propose a self-supervised debiasing framework potentially compatible with unlabeled samples. Specifically, we first pretrain a biased encoder in a self-supervised manner with the rank regularization, serving as a semantic bottleneck to enforce the encoder to learn the spuriously correlated attributes. This biased encoder is then used to discover and upweight bias-conflicting samples in a downstream task, serving as a boosting to effectively debias the main model. Remarkably, the proposed debiasing framework significantly improves the generalization performance of self-supervised learning baselines and, in some cases, even outperforms state-of-the-art supervised debiasing approaches. + +# 1. Introduction + +While modern deep learning solves several challenging tasks successfully, a series of recent works [16, 18, 20] have reported that the high accuracy of deep networks on indistribution samples does not always guarantee low test error on out-of-distribution (OOD) samples, especially in the context of spurious correlations. Existing studies [3, 38, 50] suggest that the deep networks can be potentially biased to the spuriously correlated attributes, or dataset bias, which are misleading statistical heuristics that are closely correlated but not causally related to the target label. + +These catastrophic pitfalls of dataset bias have facilitated the development of debiasing methods, which can be roughly categorized into approaches: (1) leveraging annotations of spurious attributes, i.e., bias label [27, 46, 49, 56]; (2) presuming specific type of bias, e.g., color and texture [5, 17, 53]; or (3) without using explicit kinds of supervisions on dataset bias [30, 31, 35, 39, 60]. + +While substantial advances have been made in this regard, these approaches still fail to address the problem: how to train a debiased classifier by fully exploiting unlabeled samples lacking both bias and target label. More specifically, while the large-scale unlabeled dataset can be potentially biased towards spuriously correlated sensitive attributes, e.g., ethnicity, gender, or age [1, 2], current existing debiasing frameworks are not designed to deal with this real-world unsupervised settings. Here we also confirm that most supervised debiasing frameworks suffer from performance degradation in the low-labeled data setting. Moreover, recent works have suggested that self-supervised learning might not be sufficient to deal with OOD generalization [10, 19, 44] when dataset bias remains after data augmentation. + +To tackle this issue, we first make a series of empirical observations that allow us to examine the fundamental difference between biased and unbiased representations. Interestingly, we found that spurious correlations suppress the effective rank [45] of latent representations, which severely deteriorates the semantic diversity of representations and leads to the degradation of feature discriminability. Another notable aspect of our findings is that the intentional increase of feature redundancy amplifies "prejudice" in neural networks. To be specific, as we enforce the correlation among latent features to regularize the effective rank of representations (i.e., rank regularization), the accuracy on bias-conflicting samples quickly declines while the model still performs reasonably well on the bias-aligned $^1$ samples. + +![](images/32adb27b77ae1b31508f6744fe78fa506eb842d8066fc972fdffc8511748ad58.jpg) +Figure 1. Concept. Based on the bias-rank relationship (Section 2), we introduce a novel debiasing framework centered on rank regularization, which intentionally amplifies spurious correlation by enforcing feature components to be entangled with both spurious and invariant attributes. + +Based on these observations, we propose a novel debiasing framework that can utilize both labeled and unlabeled biased samples with rank regularization. The proposed method is fully compatible with both supervised and self-supervised scenarios, where such compatibility arises from the rank regularization that does not rely on any labels. Specifically, for a supervised (self-supervised) setting, we train 1) a biased classifier (encoder) with rank regularization, which serves as a semantic bottleneck limiting the semantic diversity of feature components, and 2) the main classifier (encoder) with standard (self-)supervised learning approaches. The biased model affords us the leverage to uncover spurious correlations and identify bias-conflicting samples in a downstream task. + +Our work is the first to unveil the bias-rank relationships and introduce an effective debiasing strategy to exploit potentially unlabeled data samples. We demonstrate the effectiveness of the proposed debiasing framework with various challenging real-world biased datasets, including MultiCM-NIST [34], biased Chest X-ray databases, UTKFace, CelebA, etc., in both a supervised and self-supervised scenario. These experiments show that our method significantly outperforms other self-supervised baselines, and even state-of-the-art supervised debiasing methods in some cases. + +# 2. Low-rank bias of biased representations + +# 2.1. Preliminaries + +Throughout the paper, we denote $x \in \mathbb{R}^m$ and $y \in \mathcal{V}$ as $m$ -dimensional input sample and its corresponding predicting label, respectively. Then we denote $X = \{x_k\}_{k=1}^n$ as a batch of $n$ samples from a dataset which is fed to an encoder $f_\theta : \mathbb{R}^m \to \mathbb{R}^d$ , parameterized by $\theta$ . Then we construct a matrix $Z \in \mathbb{R}^{n \times d}$ where each $i$ th row is the output representations of the encoder $f_\theta(x_i)^T$ for $x_i \in X$ . For every analysis in this section, we use $Z$ as our latent representation. + +tions, where the neural backbone of the encoder may vary as simple convolutional networks, ResNet-18, or ViT [15] (Experimental details provided in Appendix C.1 and D). + +To evaluate the semantic diversity of given representation matrix, we introduce effective rank [45] which is a widely used metric to measure the effective dimensionality of matrix and analyze the spectral properties of features in neural networks [4, 6, 23, 42]: + +Definition 1. Given the matrix $X \in \mathbb{R}^{m \times n}$ and its singular values $\{\sigma_i\}_{i=1}^{\min(m,n)}$ , the effective rank $\rho$ of $X$ is defined as the shannon entropy of normalized singular values: + +$$ +\rho (X) = - \sum_ {i = 1} ^ {\min (m, n)} \bar {\sigma} _ {i} \log \bar {\sigma} _ {i}, \tag {1} +$$ + +where $\bar{\sigma}_i = \sigma_i / \sum_k\sigma_k$ is $i$ -th normalized singular value. Without loss of generality, we omit the exponentiation of $\rho (X)$ as done in [45]. + +Effective rank is also referred to as spectral entropy where its value is maximized when the singular values are all equal and minimized when a top singular value dominates relative to all others. Recent works [12, 13] have revealed that the discriminability of representations resides on wide range of eigenvectors since the rich discriminative information for the classification task cannot be transmitted by only few eigenvectors with top singular values. Thus from a spectral analysis perspective, effective rank quantifies how diverse the semantic information encoded by each eigenfeature is, which is closely related to the feature discriminability across target label categories. In the rest of paper, we interchangeably use effective rank and rank by following prior works. + +# 2.2. Spectral analysis of the bias-rank relationships + +We now present experiments showing that the deep networks may tend to encode lower-rank representations in the pres + +![](images/b089feec5af2202506781bd6edff6d8f264af3d582e1102e8036b1d006ef70ea.jpg) +(a) Unbiased corr. + +![](images/efb658972a9fd5033c30fe74a24be47c6a2bc58e0fb7b165aebe7ad17766af7e.jpg) +(b) Biased corr. + +![](images/7a8bda690c1ff40e5b224fcc30a1e4e81071c04c00ba971fe6395bfcbf7edae6.jpg) +(c) Effective rank + +![](images/d5b645d1e5d623cd3a9dc01051a7689d30b99996e49be5e08c1806acb6ed8467.jpg) +(d) SVD analysis +Figure 2. Empirical analysis on rank reduction phenomenon. For every analysis, we used the output $Z$ of the encoder (Sec. 2.1). (a, b): Hierarchically clustered auto-correlation matrix of unbiased and biased representations (Bias ratio=99%). (c): Effective rank with color bias. 'Unbiased' represents the case training model with perfectly unbiased dataset, i.e. random color for each sample. (d): SVD analysis with max-normalized singular values. Top 100 values are shown in the figure (total: 256). + +ence of stronger spurious correlations. To arbitrarily control the degree of spurious correlations, we introduce synthetic biased datasets, Color-MNIST (CMNIST) and Corrupted CIFAR-10 (CIFAR-10C, [21]), with color and corruption bias types, respectively. We define the degree of spurious correlations as the ratio of bias-aligned samples included in the training set, or bias ratio, where most of the samples are bias-aligned in the context of strong spurious correlations. + +Figure 2c shows that the rank of latent representations from a penultimate layer of the simple convolutional classifier decreases as the bias ratio increases in CMNIST. We provide similar rank reduction results of CIFAR-10C with ResNet-18 and ViT in the Appendix C.1. We further compare the correlation matrix of biased and unbiased latent representations in the penultimate layer of biased and unbiased classifiers, respectively. In Figure 2a and 2b, we observe that the block structure in the correlation matrix is more evident in the biased representations after the hierarchical clustering, indicating that the features become highly correlated which may limit the semantic diversity of networks. To investigate the rank reduction phenomenon in-depth, we compare the normalized singular values of biased and unbiased representations. We conduct singular value decomposition (SVD) on the feature matrices of both biased and unbiased classifiers and plot the singular values normalized by the spectral norm of the corresponding matrix. Figure 2d shows that the top few normalized singular values of biased representations are similar to or even greater than those of unbiased rep + +resentations. However, the remaining majority of singular values decay significantly faster in biased representations, greatly weakening the informative signals of eigenvectors with smaller singular values and deteriorating feature discriminability [12, 13]. + +# 2.3. Rank regularization + +Motivated from the aforementioned rank reduction phenomenon, we ask an opposite-directional question: "Can we intentionally amplify the prejudice of deep networks by maximizing the redundancy between the components of latent representations?" If the feature components are extremely correlated, the corresponding representations may exhibit most of its spectral energy along the direction of one singular vector. For this case, effective rank may converge to 0. In other words, our goal is to design a semantic bottleneck of representations that restricts the semantic diversity of feature vectors. To implement the bottleneck in practice, motivated from Figure 2b, we compute the auto-correlation matrix of the output of encoder. + +Let $\bar{Z}$ denote the mean-centered representations $Z$ along the batch dimension. The normalized auto-correlation matrix $C\in \mathbb{R}^{d\times d}$ of $\bar{Z}$ is defined as follow: + +$$ +C _ {i, j} = \frac {\sum_ {b = 1} ^ {n} \bar {Z} _ {b , i} \bar {Z} _ {b , j}}{\sqrt {\sum_ {b = 1} ^ {n} \bar {Z} _ {b , i} ^ {2}} \sqrt {\sum_ {b = 1} ^ {n} \bar {Z} _ {b , j} ^ {2}}} \quad 1 \leq \forall i, j \leq d, \tag {2} +$$ + +where $b$ is an index of sample and $i, j$ are index of each vector dimension. Then we define our regularization term as the negative of a sum of squared off-diagonal terms in $C$ : + +$$ +\ell_ {r e g} (X; \theta) = - \sum_ {i} \sum_ {j \neq i} C _ {i, j} ^ {2}, \tag {3} +$$ + +where we refer to it as the rank loss. Note that the target labels on $X$ is not used at all. + +Analysis of rank-regularized networks. To investigate the impacts of rank regularization in deep neural networks, we construct the classification model by combining the linear classifier $f_{W}:\mathbb{R}^{d}\to \mathbb{R}^{c}$ parameterized by $W\in \mathcal{W}$ on top of the encoder $f_{\theta}$ , where $c = |\mathcal{V}|$ is the number of classes. Then we trained models by cross entropy loss $\ell_{CE}$ combined with $\lambda_{reg}\ell_{reg}$ , where $\lambda_{reg} > 0$ is a Lagrangian multiplier. We use CMNIST, CIFAR-10C, and Waterbirds dataset [51], and evaluate the trained models on an unbiased test set following [30, 39]. After training models with varying the hyperparameter $\lambda_{reg}$ , we compare bias-aligned and bias-conflict accuracy, which are the average accuracy on bias-aligned and bias-conflicting samples in the unbiased test set, respectively, for CMNIST and CIFAR-10C. Test accuracy on every individual data group is reported for Waterbirds. Figure 3 shows that models suffer more from poor OOD generalization as trained with larger $\lambda_{reg}$ . The average + +![](images/16631f1a8a4c15d5116077b9006a0e6aeffc1ee152c2a80ceb95d868aa3b9ade.jpg) +(a) CMNIST + +![](images/93667851c321fa8d62815232046c5a40e097a187b7cc04f600783039e4d37f3b.jpg) +(b) CIFAR-10C +Figure 3. (a, b): Bias-conflict and Bias-aligned accuracy on CMNIST and CIFAR-10C (Bias ratio=95%). (c): Group accuracy on Waterbirds. Detailed simulation settings are in the Appendix D. + +![](images/0bd0291c9fd0a4876d2fde24aa2ef9c51600adb1ed90d85f86dd88ebffac7a8b.jpg) +(c) Waterbirds + +
CMNISTCIFAR-10C
P (%)R (%)P (%)R (%)
ERM85.5919.7652.030.06
+ Rank reg98.8395.9171.3951.43
+ +(a) CMNIST, CIFAR-10C + +
MetricsERMJTTRank reg
Precision (%)37.8448.9554.77
Recall (%)11.6748.7555.01
+ +(b) Waterbirds + +Table 1. Precision (P) and Recall (R) of bias-conflicting samples. (a): Bias-conflicting samples are identified in the error set of ERM model trained with and without rank regularization (Bias ratio=95% for both datasets). (b): Bias-conflicting samples are similarly identified by ERM, JTT, and the proposed biased model in Waterbirds dataset. + +accuracy on bias-conflicting groups is significantly degraded, while the accuracy on bias-aligned groups is maintained to some extent. It implies that rank regularization may force deep networks to focus on spurious attributes. + +Minority mining performance. Table 1a and 1b support that the biased models with strong regularization can effectively probe out the bias-conflicting samples in the training set. Specifically, we train a biased classifier with rank regularization and distill an error set $E$ of misclassified training samples as bias-conflicting samples proxies. As reported in Table 1a, we observe that our biased classifier is relatively robust to the unintended memorization of bias-conflicting samples [47] in contrast to the standard models trained by Empirical Risk Minimization (ERM). Moreover, Table 1b shows that the proposed rank regularization improves the precision and recall of identified bias-conflicting samples compared to JTT [35]. Detailed simulation settings are in the Appendix D. + +Reconstruction of biased representations. To under + +stand the relationship between rank regularization and spurious correlations more deeply, we visualize the pretrained representations with varying degrees of bias. We first trained deep networks on: (a) unbiased CMNIST (random background color), (b) biased CMNIST (bias ratio $= 95\%$ ) without rank regularization and (c) with rank regularization $(\lambda_{reg} = 50)$ . Then, we train the auxiliary decoder, which reconstructs the bias-conflicting images from the freezed latent representations of each pretrained network. Results show that rank regularization may cause the representation to lose information on complex invariant features, resulting in a loss of feature discriminability and informative signals. While both digit and color are well reconstructed with biased representations (b), the decoder fails to reconstruct bias-conflicting images from the (c) biased representations pretrained with rank regularization. The foreground digit is blurred, and its class changes following the color-digit assignment in Figure 4d. + +These observations afford us some key insights into rank regularization: First, the rank-regularized representation may lose its information on complex invariant features (i.e., shape and style of the foreground digit), specifically undermining the feature discriminability and informative signals. Second, the limited semantic diversity makes it harder to identify the true underlying independent generative factors for multidimensional data; instead, it may encode feature components entangled with both spurious and invariant attributes as the digit class of the reconstructed image is erroneously determined by the background color in 4c. + +Multiple bias attributes. To further investigate the generalizability of rank regularization, we evaluate the biased representations with Multi-Color MNIST (MultiCMNIST) dataset [34], which is similar to the CMNIST but have two bias attributes: left and right background colors. We set bias ratio $= 99\%$ for the left color and bias ratio $= 95\%$ for the right color, i.e., the left color is a more salient bias than the right color (Dataset details are provided in Appendix D). + +Table 2 shows that the rank regularization successfully biases the model w.r.t both bias attributes, while LfF [39] + +![](images/05f8c9739d4ab43b343e2d3095e534d34b9d18c0b42e7bb4b80657c07387e638.jpg) +(a) Unbiased + +![](images/96a7397b2d57f6c185f97c008c8f726b90fc3c1dec4301d34b34f4637c13f5fe.jpg) +(b) Biased +Figure 4. Randomly selected reconstructed images from representations with varying degrees of bias. First and third row correspond to the input bias-conflicting images. Second and fourth row correspond to the reconstructed images. Reconstructed from (a) unbiased representations, (b) biased representations, and (c) biased representations with rank regularization (bias ratio=95% in b, c). (d) Examples of bias-aligned CMNIST images. + +![](images/735bf7143f5bc175e63c3cc20101e97dd51d39ca6dbbbf30cae9c4281510624c.jpg) +(c) Biased with rank reg. + +![](images/bda856b5e403e77bfd7c89046dc2a08be7baae0f4b5977c6ee0e510f79889506.jpg) +(d) Examples. + +Table 2. (a) Test accuracy (\%) on MultiCMNIST. Lower is better for this results. BC for bias-conflicting, and BA for bias-aligned. Bias ratio=99(\%) for left color, and 95(\%) for right color. $\lambda_{reg} = 50$ is used for rank regularization. (b) Debiasing results. Higher is better for this results. Baseline results are from [34]. $\lambda_{up} = 50$ is used for upweighting in the proposed framework ( $\lambda_{up}$ : a manual rescaling weight given to each identified bias-conflicting samples in cross entropy loss). Pseudo-code and experimental details are provided in Appendix A and D, respectively. + +
IdxLeft colorRight color(a) Biased accuracy (%)(b) Debiased accuracy (%)
ERMLfF [39]Rank reg.LFFDebiAN [34]Ours
(1)BABC100.0100.0100.099.6100.0100.0
(2)BABC96.698.841.64.795.697.0
(3)BCBA29.33.28.798.676.579.1
(4)BCBC7.61.36.15.116.018.3
(1) ~ (4) average acc.58.3850.8339.152.072.073.6
+ +completely fails to amplify the right color bias, i.e. less salient bias, as shown in the second row (Biased accuracy part). This leads to the abnormal debiasing results of LfF as shown in Table 2 where it records unbalanced accuracy for the left- and right-color-bias-conflicting samples. In contrast, the proposed framework shows superior performance by simply weighting the misclassified bias-conflicting proxies, as done in [35]. + +Taken together, these results indicate that the rank regularization encourages the network to focus more on spurious correlations in a way that minimizes semantic diversity and entangles invariant and spurious features [41], which is a fundamentally different mechanism compared to the LfF [39] with its easy-to-learn assumption. More details on the upweighting strategy will be provided in Section 3 and pseudo-code in Appendix A. + +# 3. DeFund: Debiasing framework with unlabeled data + +Motivated by the observations in Section 2, we propose a self-supervised debiasing framework with unlabeled data, coined DeFund (Debiasing Framework with Unlabeled Data). A notable distinction from previous studies [5, 60] lies in the proposed framework's ability to effectively harness un + +labeled data for learning biased representations. This is achieved through the application of self-supervised learning and rank regularization techniques. + +The proposed framework is composed of two stages: We first train the biased encoder, which can be potentially adopted to detect the bias-conflicting samples in a downstream task, along with the main encoder by self-supervised learning, both without any labels. After pretraining, we identify the bias-conflicting samples in the downstream task using linear evaluation protocol [9, 40]. This set of samples serves as a boosting to debias the main model. + +Notation. We denote $f_{\theta}^{bias} : \mathcal{X} \to \mathbb{R}^d$ and $f_{\phi}^{main} : \mathcal{X} \to \mathbb{R}^d$ as biased encoder and main encoder parameterized by $\theta \in \Theta$ and $\phi \in \Theta$ , respectively, where $d$ is the dimensionality of latent representations. Then we can compute the rank loss in (3) with introduced encoders and given batch $\{x_k\}_{k=1}^N$ with size $N$ . Let $f_{W_b}^{cls} : \mathbb{R}^d \to \mathbb{R}^C$ be a single-layer classifier parameterized by $W_b \in \mathcal{W}$ which is placed on top of biased encoder $f_{\theta}^{bias}$ , where $C = |\mathcal{Y}|$ is the number of classes. We similarly define the linear classifier $f_{W_m}^{cls}$ for the main encoder. Then we refer to $f^{bias} : \mathcal{X} \to \mathbb{R}^C$ as biased model, where $f^{bias}(x) = f_{W_b}^{cls}\big(f_{\theta}^{bias}(x)\big), \forall x \in \mathcal{X}$ . We similarly define the main model $f^{main}$ as $f^{main}(x) = f_{W_m}^{cls}\big(f_{\phi}^{main}(x)\big), \forall x \in \mathcal{X}$ . While the projection networks [9] are employed as well, we omit the notations because they are not engaged in classification. + +Stage 1. Training a biased encoder. To train the biased encoder $f_{\theta}^{bias}$ , we revisit the proposed rank regularization term (3) in context of instance discrimination task. Building upon the observations in Section 2.3, we conjecture that rank regularization may amplify bias in self-supervised learning as well by entangling invariant and spurious features. Based on these intuitions, we apply rank regularization directly to the output of the base encoder, which encourages each feature component to be highly correlated. From these applications, several noteworthy observations have emerged: (a) The representation becomes more biased as it is trained with stronger regularization (Appendix C.1). (b) While the overall performance may be upper-bounded due to the constraint on effective dimensionality [25], the bias-conflict accuracy + +is primarily sacrificed compared to the bias-aligned accuracy (Section 4). + +Stage 2. Debiasing downstream tasks. After training the biased encoder, our goal is to debias the main model, which was pretrained using standard self-supervised learning methods on the same dataset. Here, assume that we have an ideal pretrained main encoder of which each output component corresponds to the latent factor of data variation [63]. While this ideal encoder should seamlessly adapt to downstream classification tasks, if most downstream task samples are bias-aligned, they may misguide the model to upweight spuriously correlated latent factors, leading to a biased solution despite well-generalized representations. We refer to this problem as the biased downstream application (Related analysis in Appendix C.1). + +The above contradiction elucidates the importance of bias-conflicting samples, which serve as counterexamples of spuriously correlated feature components, thereby preventing the alleged involvement of such components in prediction. Based on these intuitions, we introduce a novel debiasing protocol that probes and upweights bias-conflicting samples to find and fully exploit feature components independent of spurious correlations. We apply our framework on two scenarios: linear evaluation and semi-supervised learning. + +Linear evaluation. To validate our hypothesis on the biased downstream application, we conduct linear evaluation [40, 61] following the conventional protocol of self-supervised learning. Specifically, a linear classifier is trained on top of unsupervised pretrained representations by using target labels of training samples. After training a linear classifier $f_{W_b}^{cls}$ with pretrained biased encoder $f_{\theta}^{bias}$ given the whole training set $D = \{(x_k, y_k)\}_{k=1}^N$ with size $N$ , an error set $E$ of misclassified samples and corresponding labels is regarded as bias-conflicting pairs. Then we train a linear classifier $f_{W_m}^{cls}$ on intentionally frozen representations of main encoder $f_{\phi}^{main}$ by weighting the identified samples in $E$ with $\lambda_{up} > 0$ . The loss function for debiased linear evaluation is defined as follows: + +$$ +\begin{array}{l} \ell_ {d e b i a s} (D; W _ {m}) \\ = \lambda_ {u p} \sum_ {(x, y) \in E} \ell (x, y; W _ {m}) + \sum_ {(x, y) \in D \setminus E} \ell (x, y; W _ {m}), \\ \end{array} +$$ + +where we use cross entropy loss for $\ell : \mathcal{X} \times \mathcal{Y} \times \mathcal{W} \to \mathbb{R}^+$ . Note that the target labels are only used in training linear classifiers after pretraining. + +Note that the debiased linear evaluation is not meant to compete directly with other supervised baselines. Instead, it aims to: (a) examine the potential origin of the failure in OOD generalization, (b) provide a rough estimate of the potential improvement achievable with frozen latent representations, and (c) compare with standard self-supervised baselines and identify the optimal learning algorithms, e.g. SimCLR [9], for training the main encoder. + +Semi-supervised learning. We further compare our method directly to other supervised debiasing methods in the context of semi-supervised learning. Here we assume that the training dataset includes only a small amount of labeled data combined with a large amount of unlabeled data. As in linear evaluation, we train a linear classifier on top of the biased encoder by using labeled samples. After obtaining an error set $E$ of misclassified samples, we finetuned the whole main model by weighting the identified samples in $E$ with $\lambda_{up}$ . Note that supervised baselines are restricted to using only a small fraction of labeled samples, while the proposed approach benefits from the abundant unlabeled samples during pre-training of the biased encoder (Pseudo-code in the Appendix section A). + +# 4. Results + +# 4.1. Methods + +Dataset. We evaluate several supervised and self-supervised baselines on MIMIC-CXR + NIH [32], UTKFace [62] and CelebA [36] in which prior work reported poor generalization performance due to spurious correlations (Dataset details in Appendix). + +For MIMIC-CXR + NIH, we mixed the MIMIC-CXR [26] and NIH [55] following [32] where the target categories are no finding and pneumonia. Most pneumonia images are collected from MIMIC-CXR, while most no finding images are from NIH. In other words, the biases come from systematic differences in data sources, where the classifier may erroneously rely on spurious radiographic features tied to variations in data acquisition pipelines [14] instead of true pathological indicators (Example images in Figure 8). + +For UTKFace, we conduct binary classifications using (Gender, Age) and (Race, Gender) as (target, spurious) attribute pair, which we refer to UTKFace (age) and UTKFace (gender), respectively. For CelebA, we consider (HeavyMakeup, Male) and (Blonde Hair, Male) as (target, spurious) attribute pairs, which are referred to CelebA (makeup) and CelebA (blonde), respectively. The results of CelebA (blonde) are reported in Appendix C.4. Following [22, 39], we report bias-conflict accuracy together with unbiased accuracy, which is evaluated on the explicitly constructed validation set. We exclude the dataset in Figure 3 based on the observations that the SimCLR models are already invariant w.r.t spurious attributes. + +Baselines. We mainly target baselines consisting of recent advanced self-supervised learning methods, SimCLR [9], VICReg [7], and SimSiam [11], which can be categorized into contrastive (SimCLR) and non-contrastive (VICReg, SimSiam) methods. We further report the performance of vanilla networks trained by ERM, and other supervised debiasing methods such as LNL [27], EnD [49], and + +Table 3. (Supervised learning) Bias-conflict and unbiased accuracy (%) on MIMIC-CXR + NIH. Each $\checkmark$ marker represents whether the model requires information on dataset bias. Bias ratio=10%. + +
AccuracyLNL √EnD √LfF XJTT XCVaR DRO XERM XSimCLR XOurs X
Conflict43.8±0.550.4±2.325.2±2.147.9±0.244.6±0.541.7±1.235.5±1.356.8±1.7
Unbiased68.1±1.071.8±1.460.8±0.268.9±1.065.8±1.267.8±1.062.0±1.469.8±0.2
+ +Table 4. (Linear evaluation) Bias-conflict and unbiased test accuracy (%) evaluated on UTKFace and CelebA. Models requiring information on target class or dataset bias in the (pre)training stage are denoted with √in columns Y and B, respectively. Our results are marked in bold to highlight the improvements compared to the mainly interested self-supervised learning baselines (Gray rows). + +
ModelYBUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
LNL✓✓45.8±0.672.6±0.373.1±1.684.9±0.855.9±2.176.0±0.6
EnD✓✓45.3±0.972.2±0.275.5±1.185.5±0.457.3±2.476.4±1.4
JTT✓✗63.8±0.969.4±1.371.2±0.377.6±0.462.4±1.274.7±0.8
CVaR DRO✓✗45.7±2.071.4±0.368.6±1.081.0±0.858.0±1.776.5±0.6
ERM✓✗45.4±2.171.0±1.265.7±1.479.5±0.654.2±0.274.1±1.4
SimSiam✗✗28.2±0.962.6±0.748.5±1.069.8±0.739.9±0.666.7±0.6
VICReg✗✗32.3±0.664.6±0.351.0±1.471.3±0.748.6±0.671.9±0.2
SimCLR✗✗36.4±1.566.3±0.656.3±0.274.2±0.246.9±1.069.8±0.4
DeFund✗✗59.5±0.870.6±0.863.7±2.074.9±0.958.4±0.673.1±1.0
+ +upweighting-based algorithms, JTT [35] and CVaR DRO [31], which can be categorized into methods that leverage annotations on dataset bias (LNL, EnD) or not (JTT, CVaR DRO). + +Optimization setting. Both bias and main encoder is pretrained with SimCLR [9] for 100 epochs on UTKFace, and 20 epochs on CelebA, respectively, using ResNet-18, Adam optimizer and cosine annealing learning rate scheduling [37]. We use a MLP with one hidden layer for projection networks as in SimCLR. All the other baseline results are reproduced by tuning the hyperparameters and optimization settings using the same backbone architecture. We report the results of the model with the highest bias-conflicting test accuracy over those with improved unbiased test accuracy compared to the corresponding baseline algorithms, i.e., SimCLR for ours (More experimental details in Appendix D). + +# 4.2. Evaluation results + +Supervised learning. To quantify the effectiveness of the rank regularization in-depth, we first consider a standard supervised debiasing scenario as similarly done in Table 2. For a MIMIC-CXR + NIH dataset, we found that the proposed framework outperforms other supervised baselines with respect to bias-conflict accuracy. Table 14 in the Appendix shows that the rank-regularized networks effectively + +discover the bias-conflicting samples which are consistent with Table 1a, 1b, and 2. + +Linear evaluation. We also found that DeFund outperforms every self-supervised baseline by a large margin in a linear evaluation protocol, including SimCLR, SimSiam and VICReg, with respect to both bias-conflict and unbiased accuracy (Table 4). Moreover, in some cases, DeFund even outperforms ERM models or supervised debiasing approaches regarding bias-conflict accuracy. Note that there is an inherent gap between ERM models and self-supervised baselines, roughly $8.7\%$ on average. Moreover, we found that non-contrastive learning methods generally perform worse than the contrastive learning method. This warns us against training the main model using a non-contrastive learning approach, while it may be a viable option for the biased model. Results of the proposed framework with non-contrastive learning methods are provided in the Appendix section C.5. + +Semi-supervised learning. To compare supervised and self-supervised methods in a more practical and fair scenario, we randomly sample $10\%$ of the labeled CelebA training dataset. The remaining $90\%$ samples are treated as unlabeled ones and engaged only in pretraining encoders for self-supervised baselines. Labeled samples are provided equally to both supervised and self-supervised methods. + +Table 5. (Semi-supervised learning) Accuracy results (%) on CelebA. Label fraction $= {10}\%$ . + +
AccuracyCelebA (Makeup)CelebA (Blonde)
LNLEnDJTTCVaR DROERMSimCLRDeFundJTTDeFund
Conflict55.7±1.455.3±1.551.5±1.955.6±1.551.5±1.150.5±4.760.5±0.470.6±1.075.1±0.8
Unbiased75.6±0.576.2±0.871.4±1.375.7±1.073.1±0.371.6±1.975.6±0.278.8±1.785.8±0.3
+ +Remarkably, Table 5 and Table 16 in Appendix show that the proposed framework outperforms other state-of-the-art supervised debiasing methods. Existing upweighting protocols, such as JTT, fail to prevent deep networks from memorizing minority counterexamples. However, the proposed framework can fully utilize unlabeled samples with contrastive learning to prevent memorization. Existing bias-conflicting sample mining algorithms may be affected by the implicit bias of overparameterized networks, but this is unlikely to happen with the proposed framework since it only trains a simple linear classifier on top of a frozen biased encoder to identify such samples. + +
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
SimCLR36.466.356.374.246.969.8
+ Rank reg26.661.350.970.343.968.3
+ Upweight53.064.658.374.550.170.4
DeFund59.570.663.774.958.473.1
+ +(a) Ablation study + +
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
PrecisionRecallPrecisionRecallPrecisionRecall
SimCLR68.3144.6333.3639.5952.2528.23
DeFund68.6775.9429.9850.9355.2932.46
+ +(b) Precision and recall + +Table 6. (a) Ablation study on introduced modules. (b) Precision and recall (\%) of bias-conflicting samples identified by SimCLR and our biased model. Both case used linear evaluation. + +Ablation study. To quantify the extent of performance improvement achieved by each introduced module, we compared the linear evaluation results of (a) vanilla SimCLR, (b) SimCLR with rank regularization, (c) SimCLR with upweighting error set $E$ of the main model, and (d) DeFund. Note that (c) does not use a biased model at all. Table 6a shows that every module plays an important role in OOD generalization. Considering that the main model is already biased to some extent, we found that bias-conflict accuracy can be improved even without a biased model, where the error set $E$ of the biased model further boosts the generalization performance. We also measures the precision and recall of identified bias-conflicting samples in E, finding that the biased model detects more diverse bias-conflicting samples than the baseline (Table 6b). The improvement of recall in CelebA may seem marginal, but it is significant given the + +larger number of samples compared to UTKFace. + +Computational costs. Our framework is computationally affordable as it only trains the linaer classifier (linear eval.) or finetune networks with a few epochs, e.g., about 30 epochs for UTKFace in debiasing stage. Self-supervised pre-training and linear evaluation takes 19.3 and 4.5 minutes with a NVIDIA GeForce RTX 2080Ti, respectively. + +# 5. Conclusion + +We present a novel solution to the challenging self-supervised debiasing, an important problem that has received little attention so far. We (a) unveil the inductive bias towards encoding low effective rank representations in the presence of spurious correlations. Based on this, we (b) design a rank regularization that amplifies the feature redundancy by reducing the spectral entropy of latent representations. Then we (c) design a debiasing framework empowered by the biased model pretrained with abundant unlabeled samples. + +# Acknowledgments + +This research was supported by National Research foundation of Korea(NRF) (\*\*RS-2023-00262527**), Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137899, KMDF_PR_20200901_0015), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT, Ministry of Science and ICT) (No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation), (No.2019-0-00075, Artificial Intelligence Graduate School Program(KAIST)), (No. RS-2023-00233251, System3 reinforcement learning with high-level brain functions), Electronics and Telecommunications Research Institute(ETRI) grant funded by the Korean government (24ZS1100, Core Technology Research for Self-Improving Integrated Artificial Intelligence System), [N01230878, Development of Beyond X-verse Core Technology for Hyper-realistic interactions by Synchronizing the Real World and Virtual Space], a grant of the KAIST-KT joint research project through AI2XL Laboratory, Institute of convergence Technology, funded by KT (Project No.3, Project title: Genie Brain). + +# References + +[1] Abubakar Abid, Maheen Farooqi, and James Zou. Large language models associate muslims with violence. Nature Machine Intelligence, 3(6):461-463, 2021. +[2] Sandhini Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, and Miles Brundage. Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818, 2021. +[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. +[4] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. Advances in Neural Information Processing Systems, 32, 2019. +[5] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In International Conference on Machine Learning, pages 528-539. PMLR, 2020. +[6] Aristide Baratin, Thomas George, César Laurent, R Devon Hjelm, Guillaume Lajoie, Pascal Vincent, and Simon Lacoste-Julien. Implicit regularization via neural feature alignment. In International Conference on Artificial Intelligence and Statistics, pages 2269-2277. PMLR, 2021. +[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. +[8] Tiffany Tianhui Cai, Jonathan Frankle, David J Schwab, and Ari S Morcos. Are all negatives created equal in contrastive instance discrimination? arXiv preprint arXiv:2010.06682, 2020. +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[10] Ting Chen, Calvin Luo, and Lala Li. Intriguing properties of contrastive losses. Advances in Neural Information Processing Systems, 34:11834-11845, 2021. +[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. +[12] Xinyang Chen, Sinan Wang, Bo Fu, Mingsheng Long, and Jianmin Wang. Catastrophic forgetting meets negative transfer: Batch spectral shrinkage for safe transfer learning. Advances in Neural Information Processing Systems, 32, 2019. +[13] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. +[14] Alex J DeGrave, Joseph D Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021. +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans + +formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[16] Michael Feldman, Sorelle A Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. Certifying and removing disparate impact. In proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, pages 259-268, 2015. +[17] Songwei Ge, Shlok Mishra, Chun-Liang Li, Haohan Wang, and David Jacobs. Robust contrastive learning using negative samples with diminished semantics. Advances in Neural Information Processing Systems, 34:27356-27368, 2021. +[18] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. +[19] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models. arXiv preprint arXiv:2010.08377, 2020. +[20] Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324, 2018. +[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. +[22] Youngkyu Hong and Eunho Yang. Unbiased classification through bias-contrastive and bias-balanced learning. Advances in Neural Information Processing Systems, 34:26449-26461, 2021. +[23] Minyoung Huh, Hossein Mobahi, Richard Zhang, Brian Cheung, Pulkit Agrawal, and Phillip Isola. The low-rank simplicity bias in deep networks. arXiv preprint arXiv:2103.10427, 2021. +[24] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. +[25] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348, 2021. +[26] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. +[27] Byungju Kim, Hyunwoo Kim, Kyungsu Kim, Sungjin Kim, and Junmo Kim. Learning not to learn: Training deep neural networks with biased data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9012-9020, 2019. +[28] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. +[29] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip + +Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021. +[30] Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, and Jaegul Choo. Learning debiased representation via disentangled feature augmentation. Advances in Neural Information Processing Systems, 34:25123-25133, 2021. +[31] Daniel Levy, Yair Carmon, John C Duchi, and Aaron Sidford. Large-scale methods for distributionally robust optimization. Advances in Neural Information Processing Systems, 33:8847-8860, 2020. +[32] Jiaxuan Li, Duc Minh Vo, and Hideki Nakayama. Partition-and-debias: Agnostic biases mitigation via a mixture of biases-specific experts. arXiv preprint arXiv:2308.10005, 2023. +[33] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14970-14979, 2021. +[34] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In European Conference on Computer Vision, pages 270-288. Springer, 2022. +[35] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. +[36] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. +[37] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. +[38] Vaishnavh Nagarajan, Anders Andreassen, and Behnam Neyshabur. Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775, 2020. +[39] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020. +[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. +[41] Geon Yeong Park, Sangmin Lee, Sang Wan Lee, and Jong Chul Ye. Training debiased subnetworks with contrastive weight pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7929-7938, 2023. +[42] Noam Razin and Nadav Cohen. Implicit regularization in deep learning may not be explainable by norms. Advances in neural information processing systems, 33:21174-21187, 2020. + +[43] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020. +[44] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021. +[45] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007. +[46] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. +[47] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. +[48] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv preprint arXiv:2206.01197, 2022. +[49] Enzo Tartaglione, Carlo Alberto Barbano, and Marco Grangetto. End: Entangling and disentangling deep representations for bias correction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13508-13517, 2021. +[50] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. +[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. +[52] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. +[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019. +[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. +[55] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. +[56] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8919-8928, 2020. + +[57] Yadollah Yaghoobzadeh, Soroush Mehri, Remi Tachet, Timothy J Hazen, and Alessandro Sordoni. Increasing robustness to spurious correlations using forgettable examples. arXiv preprint arXiv:1911.03861, 2019. +[58] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021. +[59] Dinghuai Zhang, Kartik Ahuja, Yilun Xu, Yisen Wang, and Aaron Courville. Can subnetwork structure be the key to out-of-distribution generalization? In International Conference on Machine Learning, pages 12356-12367. PMLR, 2021. +[60] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022. +[61] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016. +[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. +[63] Roland S Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive learning inverts the data generating process. In International Conference on Machine Learning, pages 12979-12990. PMLR, 2021. \ No newline at end of file diff --git a/2024/Self-supervised Debiasing Using Low Rank Regularization/images.zip b/2024/Self-supervised Debiasing Using Low Rank Regularization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f8433eb2dd292345922f6bbb069b02e592c17ae9 --- /dev/null +++ b/2024/Self-supervised Debiasing Using Low Rank Regularization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6a381b6e45e4c6b2997fd8751a455fff43a8dfb015ccbee74292f674746b445 +size 478185 diff --git a/2024/Self-supervised Debiasing Using Low Rank Regularization/layout.json b/2024/Self-supervised Debiasing Using Low Rank Regularization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d60a3fa78531a703aa5054a9d11ce126be3993 --- /dev/null +++ b/2024/Self-supervised Debiasing Using Low Rank Regularization/layout.json @@ -0,0 +1,8937 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 470, + 121 + ], + "type": "text", + "content": "Self-supervised debiasing using low rank regularization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 142, + 512, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 142, + 512, + 158 + ], + "spans": [ + { + "bbox": [ + 80, + 142, + 512, + 158 + ], + "type": "text", + "content": "Geon Yeong Park1, Chanyong Jung1, Sangmin Lee2, Jong Chul Ye1,2,3* Sang Wan Lee1,4*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 156, + 426, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 156, + 426, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 156, + 426, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 167, + 156, + 426, + 171 + ], + "type": "text", + "content": "Bio and Brain Engineering, " + }, + { + "bbox": [ + 167, + 156, + 426, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 167, + 156, + 426, + 171 + ], + "type": "text", + "content": "Mathematical Sciences," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 171, + 462, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 462, + 185 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 462, + 185 + ], + "type": "text", + "content": "3Kim Jaechul Graduate School of AI, 4Brain and Cognitive Sciences" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 186, + 488, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 186, + 488, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 488, + 199 + ], + "type": "text", + "content": "Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 201, + 455, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 201, + 455, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 201, + 455, + 213 + ], + "type": "text", + "content": "{pky3436, jcy132, leeeeesang,jong.ye,sangwan}@kaist.ac.kr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 266, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 553 + ], + "type": "text", + "content": "Spurious correlations can cause strong biases in deep neural networks, impairing generalization ability. While most existing debiasing methods require full supervision on either spurious attributes or target labels, training a debiased model from a limited amount of both annotations is still an open question. To address this issue, we investigate an interesting phenomenon using the spectral analysis of latent representations: spuriously correlated attributes make neural networks inductively biased towards encoding lower effective rank representations. We also show that a rank regularization can amplify this bias in a way that encourages highly correlated features. Leveraging these findings, we propose a self-supervised debiasing framework potentially compatible with unlabeled samples. Specifically, we first pretrain a biased encoder in a self-supervised manner with the rank regularization, serving as a semantic bottleneck to enforce the encoder to learn the spuriously correlated attributes. This biased encoder is then used to discover and upweight bias-conflicting samples in a downstream task, serving as a boosting to effectively debias the main model. Remarkably, the proposed debiasing framework significantly improves the generalization performance of self-supervised learning baselines and, in some cases, even outperforms state-of-the-art supervised debiasing approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "While modern deep learning solves several challenging tasks successfully, a series of recent works [16, 18, 20] have reported that the high accuracy of deep networks on indistribution samples does not always guarantee low test error on out-of-distribution (OOD) samples, especially in the context of spurious correlations. Existing studies [3, 38, 50] suggest that the deep networks can be potentially biased to the spuriously correlated attributes, or dataset bias, which are misleading statistical heuristics that are closely correlated but not causally related to the target label." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "content": "These catastrophic pitfalls of dataset bias have facilitated the development of debiasing methods, which can be roughly categorized into approaches: (1) leveraging annotations of spurious attributes, i.e., bias label [27, 46, 49, 56]; (2) presuming specific type of bias, e.g., color and texture [5, 17, 53]; or (3) without using explicit kinds of supervisions on dataset bias [30, 31, 35, 39, 60]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 326, + 548, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 548, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 548, + 495 + ], + "type": "text", + "content": "While substantial advances have been made in this regard, these approaches still fail to address the problem: how to train a debiased classifier by fully exploiting unlabeled samples lacking both bias and target label. More specifically, while the large-scale unlabeled dataset can be potentially biased towards spuriously correlated sensitive attributes, e.g., ethnicity, gender, or age [1, 2], current existing debiasing frameworks are not designed to deal with this real-world unsupervised settings. Here we also confirm that most supervised debiasing frameworks suffer from performance degradation in the low-labeled data setting. Moreover, recent works have suggested that self-supervised learning might not be sufficient to deal with OOD generalization [10, 19, 44] when dataset bias remains after data augmentation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 495, + 548, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 548, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 548, + 664 + ], + "type": "text", + "content": "To tackle this issue, we first make a series of empirical observations that allow us to examine the fundamental difference between biased and unbiased representations. Interestingly, we found that spurious correlations suppress the effective rank [45] of latent representations, which severely deteriorates the semantic diversity of representations and leads to the degradation of feature discriminability. Another notable aspect of our findings is that the intentional increase of feature redundancy amplifies \"prejudice\" in neural networks. To be specific, as we enforce the correlation among latent features to regularize the effective rank of representations (i.e., rank regularization), the accuracy on bias-conflicting samples quickly declines while the model still performs reasonably well on the bias-aligned " + }, + { + "bbox": [ + 304, + 495, + 548, + 664 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 495, + 548, + 664 + ], + "type": "text", + "content": " samples." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 305, + 673, + 547, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 673, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 547, + 712 + ], + "type": "text", + "content": "1The bias-aligned samples refer to data with a strong correlation between (potentially latent) spurious features and target labels. The bias-conflicting samples refer to the opposite cases where spurious correlations do not exist." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12395" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 73, + 512, + 219 + ], + "blocks": [ + { + "bbox": [ + 72, + 73, + 512, + 219 + ], + "lines": [ + { + "bbox": [ + 72, + 73, + 512, + 219 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 512, + 219 + ], + "type": "image", + "image_path": "32adb27b77ae1b31508f6744fe78fa506eb842d8066fc972fdffc8511748ad58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "type": "text", + "content": "Figure 1. Concept. Based on the bias-rank relationship (Section 2), we introduce a novel debiasing framework centered on rank regularization, which intentionally amplifies spurious correlation by enforcing feature components to be entangled with both spurious and invariant attributes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 284, + 290, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 450 + ], + "type": "text", + "content": "Based on these observations, we propose a novel debiasing framework that can utilize both labeled and unlabeled biased samples with rank regularization. The proposed method is fully compatible with both supervised and self-supervised scenarios, where such compatibility arises from the rank regularization that does not rely on any labels. Specifically, for a supervised (self-supervised) setting, we train 1) a biased classifier (encoder) with rank regularization, which serves as a semantic bottleneck limiting the semantic diversity of feature components, and 2) the main classifier (encoder) with standard (self-)supervised learning approaches. The biased model affords us the leverage to uncover spurious correlations and identify bias-conflicting samples in a downstream task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 451, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 288, + 571 + ], + "type": "text", + "content": "Our work is the first to unveil the bias-rank relationships and introduce an effective debiasing strategy to exploit potentially unlabeled data samples. We demonstrate the effectiveness of the proposed debiasing framework with various challenging real-world biased datasets, including MultiCM-NIST [34], biased Chest X-ray databases, UTKFace, CelebA, etc., in both a supervised and self-supervised scenario. These experiments show that our method significantly outperforms other self-supervised baselines, and even state-of-the-art supervised debiasing methods in some cases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 579, + 268, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 268, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 268, + 594 + ], + "type": "text", + "content": "2. Low-rank bias of biased representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 133, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 133, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 133, + 611 + ], + "type": "text", + "content": "2.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": "Throughout the paper, we denote " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{V}" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": "-dimensional input sample and its corresponding predicting label, respectively. Then we denote " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "X = \\{x_k\\}_{k=1}^n" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": " as a batch of " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": " samples from a dataset which is fed to an encoder " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_\\theta : \\mathbb{R}^m \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": ". Then we construct a matrix " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": " where each " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": "th row is the output representations of the encoder " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "f_\\theta(x_i)^T" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "x_i \\in X" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": ". For every analysis in this section, we use " + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 617, + 289, + 714 + ], + "type": "text", + "content": " as our latent representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": "tions, where the neural backbone of the encoder may vary as simple convolutional networks, ResNet-18, or ViT [15] (Experimental details provided in Appendix C.1 and D)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 319, + 546, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 380 + ], + "type": "text", + "content": "To evaluate the semantic diversity of given representation matrix, we introduce effective rank [45] which is a widely used metric to measure the effective dimensionality of matrix and analyze the spectral properties of features in neural networks [4, 6, 23, 42]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "text", + "content": "Definition 1. Given the matrix " + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "text", + "content": " and its singular values " + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i\\}_{i=1}^{\\min(m,n)}" + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "text", + "content": ", the effective rank " + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 384, + 547, + 422 + ], + "type": "text", + "content": " is defined as the shannon entropy of normalized singular values:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 429, + 546, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 429, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 364, + 429, + 546, + 462 + ], + "type": "interline_equation", + "content": "\\rho (X) = - \\sum_ {i = 1} ^ {\\min (m, n)} \\bar {\\sigma} _ {i} \\log \\bar {\\sigma} _ {i}, \\tag {1}", + "image_path": "16eac6abb94c693458106a9c849c44a94a6b3b6e8e7cc0ec4965db6d4c1251da.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "inline_equation", + "content": "\\bar{\\sigma}_i = \\sigma_i / \\sum_k\\sigma_k" + }, + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "text", + "content": "-th normalized singular value. Without loss of generality, we omit the exponentiation of " + }, + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "inline_equation", + "content": "\\rho (X)" + }, + { + "bbox": [ + 304, + 468, + 547, + 505 + ], + "type": "text", + "content": " as done in [45]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 665 + ], + "type": "text", + "content": "Effective rank is also referred to as spectral entropy where its value is maximized when the singular values are all equal and minimized when a top singular value dominates relative to all others. Recent works [12, 13] have revealed that the discriminability of representations resides on wide range of eigenvectors since the rich discriminative information for the classification task cannot be transmitted by only few eigenvectors with top singular values. Thus from a spectral analysis perspective, effective rank quantifies how diverse the semantic information encoded by each eigenfeature is, which is closely related to the feature discriminability across target label categories. In the rest of paper, we interchangeably use effective rank and rank by following prior works." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 671, + 545, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 545, + 685 + ], + "type": "text", + "content": "2.2. Spectral analysis of the bias-rank relationships" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 713 + ], + "type": "text", + "content": "We now present experiments showing that the deep networks may tend to encode lower-rank representations in the pres" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12396" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 154, + 145 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 154, + 145 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 154, + 145 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 154, + 145 + ], + "type": "image", + "image_path": "b089feec5af2202506781bd6edff6d8f264af3d582e1102e8036b1d006ef70ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 148, + 130, + 156 + ], + "lines": [ + { + "bbox": [ + 71, + 148, + 130, + 156 + ], + "spans": [ + { + "bbox": [ + 71, + 148, + 130, + 156 + ], + "type": "text", + "content": "(a) Unbiased corr." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 286, + 143 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 286, + 143 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 286, + 143 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 286, + 143 + ], + "type": "image", + "image_path": "efb658972a9fd5033c30fe74a24be47c6a2bc58e0fb7b165aebe7ad17766af7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 146, + 261, + 155 + ], + "lines": [ + { + "bbox": [ + 211, + 146, + 261, + 155 + ], + "spans": [ + { + "bbox": [ + 211, + 146, + 261, + 155 + ], + "type": "text", + "content": "(b) Biased corr." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 168, + 154, + 235 + ], + "blocks": [ + { + "bbox": [ + 48, + 168, + 154, + 235 + ], + "lines": [ + { + "bbox": [ + 48, + 168, + 154, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 168, + 154, + 235 + ], + "type": "image", + "image_path": "7a8bda690c1ff40e5b224fcc30a1e4e81071c04c00ba971fe6395bfcbf7edae6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 238, + 129, + 247 + ], + "lines": [ + { + "bbox": [ + 72, + 238, + 129, + 247 + ], + "spans": [ + { + "bbox": [ + 72, + 238, + 129, + 247 + ], + "type": "text", + "content": "(c) Effective rank" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 161, + 159, + 286, + 245 + ], + "blocks": [ + { + "bbox": [ + 161, + 159, + 286, + 245 + ], + "lines": [ + { + "bbox": [ + 161, + 159, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 161, + 159, + 286, + 245 + ], + "type": "image", + "image_path": "d5b645d1e5d623cd3a9dc01051a7689d30b99996e49be5e08c1806acb6ed8467.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 248, + 252, + 258 + ], + "lines": [ + { + "bbox": [ + 195, + 248, + 252, + 258 + ], + "spans": [ + { + "bbox": [ + 195, + 248, + 252, + 258 + ], + "type": "text", + "content": "(d) SVD analysis" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 268, + 289, + 356 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 289, + 356 + ], + "type": "text", + "content": "Figure 2. Empirical analysis on rank reduction phenomenon. For every analysis, we used the output " + }, + { + "bbox": [ + 46, + 268, + 289, + 356 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 268, + 289, + 356 + ], + "type": "text", + "content": " of the encoder (Sec. 2.1). (a, b): Hierarchically clustered auto-correlation matrix of unbiased and biased representations (Bias ratio=99%). (c): Effective rank with color bias. 'Unbiased' represents the case training model with perfectly unbiased dataset, i.e. random color for each sample. (d): SVD analysis with max-normalized singular values. Top 100 values are shown in the figure (total: 256)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 378, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 474 + ], + "type": "text", + "content": "ence of stronger spurious correlations. To arbitrarily control the degree of spurious correlations, we introduce synthetic biased datasets, Color-MNIST (CMNIST) and Corrupted CIFAR-10 (CIFAR-10C, [21]), with color and corruption bias types, respectively. We define the degree of spurious correlations as the ratio of bias-aligned samples included in the training set, or bias ratio, where most of the samples are bias-aligned in the context of strong spurious correlations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 475, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 289, + 714 + ], + "type": "text", + "content": "Figure 2c shows that the rank of latent representations from a penultimate layer of the simple convolutional classifier decreases as the bias ratio increases in CMNIST. We provide similar rank reduction results of CIFAR-10C with ResNet-18 and ViT in the Appendix C.1. We further compare the correlation matrix of biased and unbiased latent representations in the penultimate layer of biased and unbiased classifiers, respectively. In Figure 2a and 2b, we observe that the block structure in the correlation matrix is more evident in the biased representations after the hierarchical clustering, indicating that the features become highly correlated which may limit the semantic diversity of networks. To investigate the rank reduction phenomenon in-depth, we compare the normalized singular values of biased and unbiased representations. We conduct singular value decomposition (SVD) on the feature matrices of both biased and unbiased classifiers and plot the singular values normalized by the spectral norm of the corresponding matrix. Figure 2d shows that the top few normalized singular values of biased representations are similar to or even greater than those of unbiased rep" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "resentations. However, the remaining majority of singular values decay significantly faster in biased representations, greatly weakening the informative signals of eigenvectors with smaller singular values and deteriorating feature discriminability [12, 13]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 139, + 423, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 139, + 423, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 139, + 423, + 152 + ], + "type": "text", + "content": "2.3. Rank regularization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 158, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 547, + 312 + ], + "type": "text", + "content": "Motivated from the aforementioned rank reduction phenomenon, we ask an opposite-directional question: \"Can we intentionally amplify the prejudice of deep networks by maximizing the redundancy between the components of latent representations?\" If the feature components are extremely correlated, the corresponding representations may exhibit most of its spectral energy along the direction of one singular vector. For this case, effective rank may converge to 0. In other words, our goal is to design a semantic bottleneck of representations that restricts the semantic diversity of feature vectors. To implement the bottleneck in practice, motivated from Figure 2b, we compute the auto-correlation matrix of the output of encoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\bar{Z}" + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "text", + "content": " denote the mean-centered representations " + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "text", + "content": " along the batch dimension. The normalized auto-correlation matrix " + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "inline_equation", + "content": "C\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\bar{Z}" + }, + { + "bbox": [ + 305, + 313, + 545, + 348 + ], + "type": "text", + "content": " is defined as follow:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 354, + 545, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 354, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 545, + 390 + ], + "type": "interline_equation", + "content": "C _ {i, j} = \\frac {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , i} \\bar {Z} _ {b , j}}{\\sqrt {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , i} ^ {2}} \\sqrt {\\sum_ {b = 1} ^ {n} \\bar {Z} _ {b , j} ^ {2}}} \\quad 1 \\leq \\forall i, j \\leq d, \\tag {2}", + "image_path": "57d4bad4d157f00b14d134368fa4257fce2dc57073e9768369f47e72543bcf2e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": " is an index of sample and " + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": " are index of each vector dimension. Then we define our regularization term as the negative of a sum of squared off-diagonal terms in " + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 365, + 452, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 452, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 365, + 452, + 545, + 478 + ], + "type": "interline_equation", + "content": "\\ell_ {r e g} (X; \\theta) = - \\sum_ {i} \\sum_ {j \\neq i} C _ {i, j} ^ {2}, \\tag {3}", + "image_path": "d3e90ee2551779152e612a92716fc46e2f26e1821aa3da16ea387dbafff8e60a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 486, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 509 + ], + "type": "text", + "content": "where we refer to it as the rank loss. Note that the target labels on " + }, + { + "bbox": [ + 304, + 486, + 545, + 509 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 486, + 545, + 509 + ], + "type": "text", + "content": " is not used at all." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "Analysis of rank-regularized networks. To investigate the impacts of rank regularization in deep neural networks, we construct the classification model by combining the linear classifier " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_{W}:\\mathbb{R}^{d}\\to \\mathbb{R}^{c}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "W\\in \\mathcal{W}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " on top of the encoder " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "c = |\\mathcal{V}|" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " is the number of classes. Then we trained models by cross entropy loss " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{CE}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " combined with " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}\\ell_{reg}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{reg} > 0" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " is a Lagrangian multiplier. We use CMNIST, CIFAR-10C, and Waterbirds dataset [51], and evaluate the trained models on an unbiased test set following [30, 39]. After training models with varying the hyperparameter " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": ", we compare bias-aligned and bias-conflict accuracy, which are the average accuracy on bias-aligned and bias-conflicting samples in the unbiased test set, respectively, for CMNIST and CIFAR-10C. Test accuracy on every individual data group is reported for Waterbirds. Figure 3 shows that models suffer more from poor OOD generalization as trained with larger " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": ". The average" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12397" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 197, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 197, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 197, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 197, + 173 + ], + "type": "image", + "image_path": "16631f1a8a4c15d5116077b9006a0e6aeffc1ee152c2a80ceb95d868aa3b9ade.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 178, + 144, + 187 + ], + "lines": [ + { + "bbox": [ + 102, + 178, + 144, + 187 + ], + "spans": [ + { + "bbox": [ + 102, + 178, + 144, + 187 + ], + "type": "text", + "content": "(a) CMNIST" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 72, + 369, + 173 + ], + "blocks": [ + { + "bbox": [ + 220, + 72, + 369, + 173 + ], + "lines": [ + { + "bbox": [ + 220, + 72, + 369, + 173 + ], + "spans": [ + { + "bbox": [ + 220, + 72, + 369, + 173 + ], + "type": "image", + "image_path": "93667851c321fa8d62815232046c5a40e097a187b7cc04f600783039e4d37f3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 178, + 321, + 186 + ], + "lines": [ + { + "bbox": [ + 270, + 178, + 321, + 186 + ], + "spans": [ + { + "bbox": [ + 270, + 178, + 321, + 186 + ], + "type": "text", + "content": "(b) CIFAR-10C" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "type": "text", + "content": "Figure 3. (a, b): Bias-conflict and Bias-aligned accuracy on CMNIST and CIFAR-10C (Bias ratio=95%). (c): Group accuracy on Waterbirds. Detailed simulation settings are in the Appendix D." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 541, + 173 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 541, + 173 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 541, + 173 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 541, + 173 + ], + "type": "image", + "image_path": "0bd0291c9fd0a4876d2fde24aa2ef9c51600adb1ed90d85f86dd88ebffac7a8b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 178, + 490, + 187 + ], + "lines": [ + { + "bbox": [ + 444, + 178, + 490, + 187 + ], + "spans": [ + { + "bbox": [ + 444, + 178, + 490, + 187 + ], + "type": "text", + "content": "(c) Waterbirds" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 74, + 239, + 261, + 305 + ], + "blocks": [ + { + "bbox": [ + 74, + 239, + 261, + 305 + ], + "lines": [ + { + "bbox": [ + 74, + 239, + 261, + 305 + ], + "spans": [ + { + "bbox": [ + 74, + 239, + 261, + 305 + ], + "type": "table", + "html": "
CMNISTCIFAR-10C
P (%)R (%)P (%)R (%)
ERM85.5919.7652.030.06
+ Rank reg98.8395.9171.3951.43
", + "image_path": "6b62aadfe3f36b3453fca9621e001330e1f2f0c9897047307f39c1b4040dbc8e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 82, + 317, + 253, + 367 + ], + "blocks": [ + { + "bbox": [ + 125, + 307, + 211, + 316 + ], + "lines": [ + { + "bbox": [ + 125, + 307, + 211, + 316 + ], + "spans": [ + { + "bbox": [ + 125, + 307, + 211, + 316 + ], + "type": "text", + "content": "(a) CMNIST, CIFAR-10C" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 317, + 253, + 367 + ], + "lines": [ + { + "bbox": [ + 82, + 317, + 253, + 367 + ], + "spans": [ + { + "bbox": [ + 82, + 317, + 253, + 367 + ], + "type": "table", + "html": "
MetricsERMJTTRank reg
Precision (%)37.8448.9554.77
Recall (%)11.6748.7555.01
", + "image_path": "bfb912d082d63259dbd93bc12a5282da353c75d75be02bfd7ce13102ea152e89.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 369, + 191, + 379 + ], + "lines": [ + { + "bbox": [ + 143, + 369, + 191, + 379 + ], + "spans": [ + { + "bbox": [ + 143, + 369, + 191, + 379 + ], + "type": "text", + "content": "(b) Waterbirds" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 388, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 453 + ], + "type": "text", + "content": "Table 1. Precision (P) and Recall (R) of bias-conflicting samples. (a): Bias-conflicting samples are identified in the error set of ERM model trained with and without rank regularization (Bias ratio=95% for both datasets). (b): Bias-conflicting samples are similarly identified by ERM, JTT, and the proposed biased model in Waterbirds dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 481, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 529 + ], + "type": "text", + "content": "accuracy on bias-conflicting groups is significantly degraded, while the accuracy on bias-aligned groups is maintained to some extent. It implies that rank regularization may force deep networks to focus on spurious attributes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 531, + 288, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 288, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 288, + 699 + ], + "type": "text", + "content": "Minority mining performance. Table 1a and 1b support that the biased models with strong regularization can effectively probe out the bias-conflicting samples in the training set. Specifically, we train a biased classifier with rank regularization and distill an error set " + }, + { + "bbox": [ + 46, + 531, + 288, + 699 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 531, + 288, + 699 + ], + "type": "text", + "content": " of misclassified training samples as bias-conflicting samples proxies. As reported in Table 1a, we observe that our biased classifier is relatively robust to the unintended memorization of bias-conflicting samples [47] in contrast to the standard models trained by Empirical Risk Minimization (ERM). Moreover, Table 1b shows that the proposed rank regularization improves the precision and recall of identified bias-conflicting samples compared to JTT [35]. Detailed simulation settings are in the Appendix D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "content": "Reconstruction of biased representations. To under" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 240, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 456 + ], + "type": "text", + "content": "stand the relationship between rank regularization and spurious correlations more deeply, we visualize the pretrained representations with varying degrees of bias. We first trained deep networks on: (a) unbiased CMNIST (random background color), (b) biased CMNIST (bias ratio " + }, + { + "bbox": [ + 304, + 240, + 547, + 456 + ], + "type": "inline_equation", + "content": "= 95\\%" + }, + { + "bbox": [ + 304, + 240, + 547, + 456 + ], + "type": "text", + "content": ") without rank regularization and (c) with rank regularization " + }, + { + "bbox": [ + 304, + 240, + 547, + 456 + ], + "type": "inline_equation", + "content": "(\\lambda_{reg} = 50)" + }, + { + "bbox": [ + 304, + 240, + 547, + 456 + ], + "type": "text", + "content": ". Then, we train the auxiliary decoder, which reconstructs the bias-conflicting images from the freezed latent representations of each pretrained network. Results show that rank regularization may cause the representation to lose information on complex invariant features, resulting in a loss of feature discriminability and informative signals. While both digit and color are well reconstructed with biased representations (b), the decoder fails to reconstruct bias-conflicting images from the (c) biased representations pretrained with rank regularization. The foreground digit is blurred, and its class changes following the color-digit assignment in Figure 4d." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 458, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 589 + ], + "type": "text", + "content": "These observations afford us some key insights into rank regularization: First, the rank-regularized representation may lose its information on complex invariant features (i.e., shape and style of the foreground digit), specifically undermining the feature discriminability and informative signals. Second, the limited semantic diversity makes it harder to identify the true underlying independent generative factors for multidimensional data; instead, it may encode feature components entangled with both spurious and invariant attributes as the digit class of the reconstructed image is erroneously determined by the background color in 4c." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": "Multiple bias attributes. To further investigate the generalizability of rank regularization, we evaluate the biased representations with Multi-Color MNIST (MultiCMNIST) dataset [34], which is similar to the CMNIST but have two bias attributes: left and right background colors. We set bias ratio " + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "inline_equation", + "content": "= 99\\%" + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": " for the left color and bias ratio " + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "inline_equation", + "content": "= 95\\%" + }, + { + "bbox": [ + 304, + 591, + 547, + 688 + ], + "type": "text", + "content": " for the right color, i.e., the left color is a more salient bias than the right color (Dataset details are provided in Appendix D)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Table 2 shows that the rank regularization successfully biases the model w.r.t both bias attributes, while LfF [39]" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12398" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 190, + 133 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 190, + 133 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 190, + 133 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 190, + 133 + ], + "type": "image", + "image_path": "05f8c9739d4ab43b343e2d3095e534d34b9d18c0b42e7bb4b80657c07387e638.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 135, + 141, + 144 + ], + "lines": [ + { + "bbox": [ + 99, + 135, + 141, + 144 + ], + "spans": [ + { + "bbox": [ + 99, + 135, + 141, + 144 + ], + "type": "text", + "content": "(a) Unbiased" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 70, + 331, + 133 + ], + "blocks": [ + { + "bbox": [ + 190, + 70, + 331, + 133 + ], + "lines": [ + { + "bbox": [ + 190, + 70, + 331, + 133 + ], + "spans": [ + { + "bbox": [ + 190, + 70, + 331, + 133 + ], + "type": "image", + "image_path": "96a7397b2d57f6c185f97c008c8f726b90fc3c1dec4301d34b34f4637c13f5fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 135, + 278, + 144 + ], + "lines": [ + { + "bbox": [ + 243, + 135, + 278, + 144 + ], + "spans": [ + { + "bbox": [ + 243, + 135, + 278, + 144 + ], + "type": "text", + "content": "(b) Biased" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 154, + 547, + 199 + ], + "lines": [ + { + "bbox": [ + 45, + 154, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 45, + 154, + 547, + 199 + ], + "type": "text", + "content": "Figure 4. Randomly selected reconstructed images from representations with varying degrees of bias. First and third row correspond to the input bias-conflicting images. Second and fourth row correspond to the reconstructed images. Reconstructed from (a) unbiased representations, (b) biased representations, and (c) biased representations with rank regularization (bias ratio=95% in b, c). (d) Examples of bias-aligned CMNIST images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 471, + 133 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 471, + 133 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 471, + 133 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 471, + 133 + ], + "type": "image", + "image_path": "735bf7143f5bc175e63c3cc20101e97dd51d39ca6dbbbf30cae9c4281510624c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 135, + 441, + 144 + ], + "lines": [ + { + "bbox": [ + 361, + 135, + 441, + 144 + ], + "spans": [ + { + "bbox": [ + 361, + 135, + 441, + 144 + ], + "type": "text", + "content": "(c) Biased with rank reg." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 544, + 132 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 544, + 132 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 544, + 132 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 544, + 132 + ], + "type": "image", + "image_path": "bda856b5e403e77bfd7c89046dc2a08be7baae0f4b5977c6ee0e510f79889506.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 133, + 530, + 142 + ], + "lines": [ + { + "bbox": [ + 484, + 133, + 530, + 142 + ], + "spans": [ + { + "bbox": [ + 484, + 133, + 530, + 142 + ], + "type": "text", + "content": "(d) Examples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 326, + 285, + 396 + ], + "blocks": [ + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "type": "text", + "content": "Table 2. (a) Test accuracy (\\%) on MultiCMNIST. Lower is better for this results. BC for bias-conflicting, and BA for bias-aligned. Bias ratio=99(\\%) for left color, and 95(\\%) for right color. " + }, + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "type": "inline_equation", + "content": "\\lambda_{reg} = 50" + }, + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "type": "text", + "content": " is used for rank regularization. (b) Debiasing results. Higher is better for this results. Baseline results are from [34]. " + }, + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "type": "inline_equation", + "content": "\\lambda_{up} = 50" + }, + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "type": "text", + "content": " is used for upweighting in the proposed framework (" + }, + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "type": "inline_equation", + "content": "\\lambda_{up}" + }, + { + "bbox": [ + 46, + 218, + 289, + 318 + ], + "type": "text", + "content": ": a manual rescaling weight given to each identified bias-conflicting samples in cross entropy loss). Pseudo-code and experimental details are provided in Appendix A and D, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 326, + 285, + 396 + ], + "lines": [ + { + "bbox": [ + 50, + 326, + 285, + 396 + ], + "spans": [ + { + "bbox": [ + 50, + 326, + 285, + 396 + ], + "type": "table", + "html": "
IdxLeft colorRight color(a) Biased accuracy (%)(b) Debiased accuracy (%)
ERMLfF [39]Rank reg.LFFDebiAN [34]Ours
(1)BABC100.0100.0100.099.6100.0100.0
(2)BABC96.698.841.64.795.697.0
(3)BCBA29.33.28.798.676.579.1
(4)BCBC7.61.36.15.116.018.3
(1) ~ (4) average acc.58.3850.8339.152.072.073.6
", + "image_path": "2c6a520e00dd406bf6a374402b0b161e17a8451721825b3ec2a5ba84f7cc4ab4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 416, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 511 + ], + "type": "text", + "content": "completely fails to amplify the right color bias, i.e. less salient bias, as shown in the second row (Biased accuracy part). This leads to the abnormal debiasing results of LfF as shown in Table 2 where it records unbalanced accuracy for the left- and right-color-bias-conflicting samples. In contrast, the proposed framework shows superior performance by simply weighting the misclassified bias-conflicting proxies, as done in [35]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 513, + 288, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 288, + 608 + ], + "type": "text", + "content": "Taken together, these results indicate that the rank regularization encourages the network to focus more on spurious correlations in a way that minimizes semantic diversity and entangles invariant and spurious features [41], which is a fundamentally different mechanism compared to the LfF [39] with its easy-to-learn assumption. More details on the upweighting strategy will be provided in Section 3 and pseudo-code in Appendix A." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 619, + 289, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 289, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 289, + 646 + ], + "type": "text", + "content": "3. DeFund: Debiasing framework with unlabeled data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Motivated by the observations in Section 2, we propose a self-supervised debiasing framework with unlabeled data, coined DeFund (Debiasing Framework with Unlabeled Data). A notable distinction from previous studies [5, 60] lies in the proposed framework's ability to effectively harness un" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 220, + 546, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 546, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 546, + 256 + ], + "type": "text", + "content": "labeled data for learning biased representations. This is achieved through the application of self-supervised learning and rank regularization techniques." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": "The proposed framework is composed of two stages: We first train the biased encoder, which can be potentially adopted to detect the bias-conflicting samples in a downstream task, along with the main encoder by self-supervised learning, both without any labels. After pretraining, we identify the bias-conflicting samples in the downstream task using linear evaluation protocol [9, 40]. This set of samples serves as a boosting to debias the main model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": "Notation. We denote " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{bias} : \\mathcal{X} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f_{\\phi}^{main} : \\mathcal{X} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " as biased encoder and main encoder parameterized by " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "\\theta \\in \\Theta" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "\\phi \\in \\Theta" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " is the dimensionality of latent representations. Then we can compute the rank loss in (3) with introduced encoders and given batch " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "\\{x_k\\}_{k=1}^N" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f_{W_b}^{cls} : \\mathbb{R}^d \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " be a single-layer classifier parameterized by " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "W_b \\in \\mathcal{W}" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " which is placed on top of biased encoder " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{bias}" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "C = |\\mathcal{Y}|" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " is the number of classes. We similarly define the linear classifier " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f_{W_m}^{cls}" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " for the main encoder. Then we refer to " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f^{bias} : \\mathcal{X} \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " as biased model, where " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f^{bias}(x) = f_{W_b}^{cls}\\big(f_{\\theta}^{bias}(x)\\big), \\forall x \\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": ". We similarly define the main model " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f^{main}" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "inline_equation", + "content": "f^{main}(x) = f_{W_m}^{cls}\\big(f_{\\phi}^{main}(x)\\big), \\forall x \\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 358, + 547, + 542 + ], + "type": "text", + "content": ". While the projection networks [9] are employed as well, we omit the notations because they are not engaged in classification." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": "Stage 1. Training a biased encoder. To train the biased encoder " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{bias}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ", we revisit the proposed rank regularization term (3) in context of instance discrimination task. Building upon the observations in Section 2.3, we conjecture that rank regularization may amplify bias in self-supervised learning as well by entangling invariant and spurious features. Based on these intuitions, we apply rank regularization directly to the output of the base encoder, which encourages each feature component to be highly correlated. From these applications, several noteworthy observations have emerged: (a) The representation becomes more biased as it is trained with stronger regularization (Appendix C.1). (b) While the overall performance may be upper-bounded due to the constraint on effective dimensionality [25], the bias-conflict accuracy" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12399" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "is primarily sacrificed compared to the bias-aligned accuracy (Section 4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 251 + ], + "type": "text", + "content": "Stage 2. Debiasing downstream tasks. After training the biased encoder, our goal is to debias the main model, which was pretrained using standard self-supervised learning methods on the same dataset. Here, assume that we have an ideal pretrained main encoder of which each output component corresponds to the latent factor of data variation [63]. While this ideal encoder should seamlessly adapt to downstream classification tasks, if most downstream task samples are bias-aligned, they may misguide the model to upweight spuriously correlated latent factors, leading to a biased solution despite well-generalized representations. We refer to this problem as the biased downstream application (Related analysis in Appendix C.1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": "The above contradiction elucidates the importance of bias-conflicting samples, which serve as counterexamples of spuriously correlated feature components, thereby preventing the alleged involvement of such components in prediction. Based on these intuitions, we introduce a novel debiasing protocol that probes and upweights bias-conflicting samples to find and fully exploit feature components independent of spurious correlations. We apply our framework on two scenarios: linear evaluation and semi-supervised learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": "Linear evaluation. To validate our hypothesis on the biased downstream application, we conduct linear evaluation [40, 61] following the conventional protocol of self-supervised learning. Specifically, a linear classifier is trained on top of unsupervised pretrained representations by using target labels of training samples. After training a linear classifier " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "f_{W_b}^{cls}" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": " with pretrained biased encoder " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{bias}" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": " given the whole training set " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "D = \\{(x_k, y_k)\\}_{k=1}^N" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": ", an error set " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": " of misclassified samples and corresponding labels is regarded as bias-conflicting pairs. Then we train a linear classifier " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "f_{W_m}^{cls}" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": " on intentionally frozen representations of main encoder " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "f_{\\phi}^{main}" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": " by weighting the identified samples in " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\lambda_{up} > 0" + }, + { + "bbox": [ + 46, + 360, + 288, + 529 + ], + "type": "text", + "content": ". The loss function for debiased linear evaluation is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 535, + 275, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 535, + 275, + 575 + ], + "spans": [ + { + "bbox": [ + 60, + 535, + 275, + 575 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\ell_ {d e b i a s} (D; W _ {m}) \\\\ = \\lambda_ {u p} \\sum_ {(x, y) \\in E} \\ell (x, y; W _ {m}) + \\sum_ {(x, y) \\in D \\setminus E} \\ell (x, y; W _ {m}), \\\\ \\end{array}", + "image_path": "eaeff5011b6387af79c67f492fed30b11d7ffe48122e9f3a0458c11c0d56ace4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "type": "text", + "content": "where we use cross entropy loss for " + }, + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\ell : \\mathcal{X} \\times \\mathcal{Y} \\times \\mathcal{W} \\to \\mathbb{R}^+" + }, + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "type": "text", + "content": ". Note that the target labels are only used in training linear classifiers after pretraining." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "Note that the debiased linear evaluation is not meant to compete directly with other supervised baselines. Instead, it aims to: (a) examine the potential origin of the failure in OOD generalization, (b) provide a rough estimate of the potential improvement achievable with frozen latent representations, and (c) compare with standard self-supervised baselines and identify the optimal learning algorithms, e.g. SimCLR [9], for training the main encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "Semi-supervised learning. We further compare our method directly to other supervised debiasing methods in the context of semi-supervised learning. Here we assume that the training dataset includes only a small amount of labeled data combined with a large amount of unlabeled data. As in linear evaluation, we train a linear classifier on top of the biased encoder by using labeled samples. After obtaining an error set " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " of misclassified samples, we finetuned the whole main model by weighting the identified samples in " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\lambda_{up}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": ". Note that supervised baselines are restricted to using only a small fraction of labeled samples, while the proposed approach benefits from the abundant unlabeled samples during pre-training of the biased encoder (Pseudo-code in the Appendix section A)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 253, + 359, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 359, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 359, + 265 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 273, + 369, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 273, + 369, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 369, + 284 + ], + "type": "text", + "content": "4.1. Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 293, + 547, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 352 + ], + "type": "text", + "content": "Dataset. We evaluate several supervised and self-supervised baselines on MIMIC-CXR + NIH [32], UTKFace [62] and CelebA [36] in which prior work reported poor generalization performance due to spurious correlations (Dataset details in Appendix)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 472 + ], + "type": "text", + "content": "For MIMIC-CXR + NIH, we mixed the MIMIC-CXR [26] and NIH [55] following [32] where the target categories are no finding and pneumonia. Most pneumonia images are collected from MIMIC-CXR, while most no finding images are from NIH. In other words, the biases come from systematic differences in data sources, where the classifier may erroneously rely on spurious radiographic features tied to variations in data acquisition pipelines [14] instead of true pathological indicators (Example images in Figure 8)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "type": "text", + "content": "For UTKFace, we conduct binary classifications using (Gender, Age) and (Race, Gender) as (target, spurious) attribute pair, which we refer to UTKFace (age) and UTKFace (gender), respectively. For CelebA, we consider (HeavyMakeup, Male) and (Blonde Hair, Male) as (target, spurious) attribute pairs, which are referred to CelebA (makeup) and CelebA (blonde), respectively. The results of CelebA (blonde) are reported in Appendix C.4. Following [22, 39], we report bias-conflict accuracy together with unbiased accuracy, which is evaluated on the explicitly constructed validation set. We exclude the dataset in Figure 3 based on the observations that the SimCLR models are already invariant w.r.t spurious attributes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Baselines. We mainly target baselines consisting of recent advanced self-supervised learning methods, SimCLR [9], VICReg [7], and SimSiam [11], which can be categorized into contrastive (SimCLR) and non-contrastive (VICReg, SimSiam) methods. We further report the performance of vanilla networks trained by ERM, and other supervised debiasing methods such as LNL [27], EnD [49], and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12400" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 102, + 542, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 3. (Supervised learning) Bias-conflict and unbiased accuracy (%) on MIMIC-CXR + NIH. Each " + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": " marker represents whether the model requires information on dataset bias. Bias ratio=10%." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 102, + 542, + 177 + ], + "lines": [ + { + "bbox": [ + 53, + 102, + 542, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 542, + 177 + ], + "type": "table", + "html": "
AccuracyLNL √EnD √LfF XJTT XCVaR DRO XERM XSimCLR XOurs X
Conflict43.8±0.550.4±2.325.2±2.147.9±0.244.6±0.541.7±1.235.5±1.356.8±1.7
Unbiased68.1±1.071.8±1.460.8±0.268.9±1.065.8±1.267.8±1.062.0±1.469.8±0.2
", + "image_path": "b67e85b24028acc1d21174bc6f55d688c15546f6c1d73cf78199af1cadd1cc6d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 102, + 228, + 493, + 388 + ], + "blocks": [ + { + "bbox": [ + 46, + 186, + 546, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 546, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 546, + 220 + ], + "type": "text", + "content": "Table 4. (Linear evaluation) Bias-conflict and unbiased test accuracy (%) evaluated on UTKFace and CelebA. Models requiring information on target class or dataset bias in the (pre)training stage are denoted with √in columns Y and B, respectively. Our results are marked in bold to highlight the improvements compared to the mainly interested self-supervised learning baselines (Gray rows)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 228, + 493, + 388 + ], + "lines": [ + { + "bbox": [ + 102, + 228, + 493, + 388 + ], + "spans": [ + { + "bbox": [ + 102, + 228, + 493, + 388 + ], + "type": "table", + "html": "
ModelYBUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
LNL✓✓45.8±0.672.6±0.373.1±1.684.9±0.855.9±2.176.0±0.6
EnD✓✓45.3±0.972.2±0.275.5±1.185.5±0.457.3±2.476.4±1.4
JTT✓✗63.8±0.969.4±1.371.2±0.377.6±0.462.4±1.274.7±0.8
CVaR DRO✓✗45.7±2.071.4±0.368.6±1.081.0±0.858.0±1.776.5±0.6
ERM✓✗45.4±2.171.0±1.265.7±1.479.5±0.654.2±0.274.1±1.4
SimSiam✗✗28.2±0.962.6±0.748.5±1.069.8±0.739.9±0.666.7±0.6
VICReg✗✗32.3±0.664.6±0.351.0±1.471.3±0.748.6±0.671.9±0.2
SimCLR✗✗36.4±1.566.3±0.656.3±0.274.2±0.246.9±1.069.8±0.4
DeFund✗✗59.5±0.870.6±0.863.7±2.074.9±0.958.4±0.673.1±1.0
", + "image_path": "061b5c91c97d8e493ff9f7598c8af75b1a2a97ed1848ccdfb6d0c42653da1392.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 456 + ], + "type": "text", + "content": "upweighting-based algorithms, JTT [35] and CVaR DRO [31], which can be categorized into methods that leverage annotations on dataset bias (LNL, EnD) or not (JTT, CVaR DRO)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 601 + ], + "type": "text", + "content": "Optimization setting. Both bias and main encoder is pretrained with SimCLR [9] for 100 epochs on UTKFace, and 20 epochs on CelebA, respectively, using ResNet-18, Adam optimizer and cosine annealing learning rate scheduling [37]. We use a MLP with one hidden layer for projection networks as in SimCLR. All the other baseline results are reproduced by tuning the hyperparameters and optimization settings using the same backbone architecture. We report the results of the model with the highest bias-conflicting test accuracy over those with improved unbiased test accuracy compared to the corresponding baseline algorithms, i.e., SimCLR for ours (More experimental details in Appendix D)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 610, + 155, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 155, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 155, + 623 + ], + "type": "text", + "content": "4.2. Evaluation results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Supervised learning. To quantify the effectiveness of the rank regularization in-depth, we first consider a standard supervised debiasing scenario as similarly done in Table 2. For a MIMIC-CXR + NIH dataset, we found that the proposed framework outperforms other supervised baselines with respect to bias-conflict accuracy. Table 14 in the Appendix shows that the rank-regularized networks effectively" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 431 + ], + "type": "text", + "content": "discover the bias-conflicting samples which are consistent with Table 1a, 1b, and 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 434, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 626 + ], + "type": "text", + "content": "Linear evaluation. We also found that DeFund outperforms every self-supervised baseline by a large margin in a linear evaluation protocol, including SimCLR, SimSiam and VICReg, with respect to both bias-conflict and unbiased accuracy (Table 4). Moreover, in some cases, DeFund even outperforms ERM models or supervised debiasing approaches regarding bias-conflict accuracy. Note that there is an inherent gap between ERM models and self-supervised baselines, roughly " + }, + { + "bbox": [ + 304, + 434, + 547, + 626 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 304, + 434, + 547, + 626 + ], + "type": "text", + "content": " on average. Moreover, we found that non-contrastive learning methods generally perform worse than the contrastive learning method. This warns us against training the main model using a non-contrastive learning approach, while it may be a viable option for the biased model. Results of the proposed framework with non-contrastive learning methods are provided in the Appendix section C.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Semi-supervised learning. To compare supervised and self-supervised methods in a more practical and fair scenario, we randomly sample " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": " of the labeled CelebA training dataset. The remaining " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": " samples are treated as unlabeled ones and engaged only in pretraining encoders for self-supervised baselines. Labeled samples are provided equally to both supervised and self-supervised methods." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12401" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 91, + 529, + 159 + ], + "blocks": [ + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "lines": [ + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "spans": [ + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "type": "text", + "content": "Table 5. (Semi-supervised learning) Accuracy results (%) on CelebA. Label fraction " + }, + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "type": "inline_equation", + "content": "= {10}\\%" + }, + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 91, + 529, + 159 + ], + "lines": [ + { + "bbox": [ + 65, + 91, + 529, + 159 + ], + "spans": [ + { + "bbox": [ + 65, + 91, + 529, + 159 + ], + "type": "table", + "html": "
AccuracyCelebA (Makeup)CelebA (Blonde)
LNLEnDJTTCVaR DROERMSimCLRDeFundJTTDeFund
Conflict55.7±1.455.3±1.551.5±1.955.6±1.551.5±1.150.5±4.760.5±0.470.6±1.075.1±0.8
Unbiased75.6±0.576.2±0.871.4±1.375.7±1.073.1±0.371.6±1.975.6±0.278.8±1.785.8±0.3
", + "image_path": "2d317a7308fc22a422048a97686430f54493974e8aa99fc33ce356f6a7eb22d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 178, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 178, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 289, + 323 + ], + "type": "text", + "content": "Remarkably, Table 5 and Table 16 in Appendix show that the proposed framework outperforms other state-of-the-art supervised debiasing methods. Existing upweighting protocols, such as JTT, fail to prevent deep networks from memorizing minority counterexamples. However, the proposed framework can fully utilize unlabeled samples with contrastive learning to prevent memorization. Existing bias-conflicting sample mining algorithms may be affected by the implicit bias of overparameterized networks, but this is unlikely to happen with the proposed framework since it only trains a simple linear classifier on top of a frozen biased encoder to identify such samples." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 332, + 288, + 396 + ], + "blocks": [ + { + "bbox": [ + 52, + 332, + 288, + 396 + ], + "lines": [ + { + "bbox": [ + 52, + 332, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 52, + 332, + 288, + 396 + ], + "type": "table", + "html": "
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
ConflictUnbiasedConflictUnbiasedConflictUnbiased
SimCLR36.466.356.374.246.969.8
+ Rank reg26.661.350.970.343.968.3
+ Upweight53.064.658.374.550.170.4
DeFund59.570.663.774.958.473.1
", + "image_path": "e57ca90ffec5b11ee87731b251ded63efe8afa1d10d676c8e4a15572626c25c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 53, + 406, + 288, + 460 + ], + "blocks": [ + { + "bbox": [ + 140, + 396, + 201, + 406 + ], + "lines": [ + { + "bbox": [ + 140, + 396, + 201, + 406 + ], + "spans": [ + { + "bbox": [ + 140, + 396, + 201, + 406 + ], + "type": "text", + "content": "(a) Ablation study" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 406, + 288, + 460 + ], + "lines": [ + { + "bbox": [ + 53, + 406, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 288, + 460 + ], + "type": "table", + "html": "
MethodUTKFace (age)UTKFace (gender)CelebA (makeup)
PrecisionRecallPrecisionRecallPrecisionRecall
SimCLR68.3144.6333.3639.5952.2528.23
DeFund68.6775.9429.9850.9355.2932.46
", + "image_path": "0e9bb32559e62f0b2ac189d7c1df3c1407c2207f9744b9c469929701c6cf0bca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 461, + 209, + 471 + ], + "lines": [ + { + "bbox": [ + 132, + 461, + 209, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 209, + 471 + ], + "type": "text", + "content": "(b) Precision and recall" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 480, + 287, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 514 + ], + "type": "text", + "content": "Table 6. (a) Ablation study on introduced modules. (b) Precision and recall (\\%) of bias-conflicting samples identified by SimCLR and our biased model. Both case used linear evaluation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "Ablation study. To quantify the extent of performance improvement achieved by each introduced module, we compared the linear evaluation results of (a) vanilla SimCLR, (b) SimCLR with rank regularization, (c) SimCLR with upweighting error set " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " of the main model, and (d) DeFund. Note that (c) does not use a biased model at all. Table 6a shows that every module plays an important role in OOD generalization. Considering that the main model is already biased to some extent, we found that bias-conflict accuracy can be improved even without a biased model, where the error set " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " of the biased model further boosts the generalization performance. We also measures the precision and recall of identified bias-conflicting samples in E, finding that the biased model detects more diverse bias-conflicting samples than the baseline (Table 6b). The improvement of recall in CelebA may seem marginal, but it is significant given the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 178, + 507, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 507, + 190 + ], + "type": "text", + "content": "larger number of samples compared to UTKFace." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 191, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 547, + 263 + ], + "type": "text", + "content": "Computational costs. Our framework is computationally affordable as it only trains the linaer classifier (linear eval.) or finetune networks with a few epochs, e.g., about 30 epochs for UTKFace in debiasing stage. Self-supervised pre-training and linear evaluation takes 19.3 and 4.5 minutes with a NVIDIA GeForce RTX 2080Ti, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 269, + 379, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 379, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 379, + 281 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 397 + ], + "type": "text", + "content": "We present a novel solution to the challenging self-supervised debiasing, an important problem that has received little attention so far. We (a) unveil the inductive bias towards encoding low effective rank representations in the presence of spurious correlations. Based on this, we (b) design a rank regularization that amplifies the feature redundancy by reducing the spectral entropy of latent representations. Then we (c) design a debiasing framework empowered by the biased model pretrained with abundant unlabeled samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 407, + 388, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 388, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 388, + 419 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 548, + 715 + ], + "type": "text", + "content": "This research was supported by National Research foundation of Korea(NRF) (\\*\\*RS-2023-00262527**), Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: 1711137899, KMDF_PR_20200901_0015), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT, Ministry of Science and ICT) (No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation), (No.2019-0-00075, Artificial Intelligence Graduate School Program(KAIST)), (No. RS-2023-00233251, System3 reinforcement learning with high-level brain functions), Electronics and Telecommunications Research Institute(ETRI) grant funded by the Korean government (24ZS1100, Core Technology Research for Self-Improving Integrated Artificial Intelligence System), [N01230878, Development of Beyond X-verse Core Technology for Hyper-realistic interactions by Synchronizing the Real World and Virtual Space], a grant of the KAIST-KT joint research project through AI2XL Laboratory, Institute of convergence Technology, funded by KT (Project No.3, Project title: Genie Brain)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12402" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Abubakar Abid, Maheen Farooqi, and James Zou. Large language models associate muslims with violence. Nature Machine Intelligence, 3(6):461-463, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Sandhini Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, and Miles Brundage. Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "text", + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "type": "text", + "content": "[4] Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[5] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In International Conference on Machine Learning, pages 528-539. PMLR, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 335 + ], + "type": "text", + "content": "[6] Aristide Baratin, Thomas George, César Laurent, R Devon Hjelm, Guillaume Lajoie, Pascal Vincent, and Simon Lacoste-Julien. Implicit regularization via neural feature alignment. In International Conference on Artificial Intelligence and Statistics, pages 2269-2277. PMLR, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 369 + ], + "type": "text", + "content": "[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "type": "text", + "content": "[8] Tiffany Tianhui Cai, Jonathan Frankle, David J Schwab, and Ari S Morcos. Are all negatives created equal in contrastive instance discrimination? arXiv preprint arXiv:2010.06682, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 457 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 288, + 491 + ], + "type": "text", + "content": "[10] Ting Chen, Calvin Luo, and Lala Li. Intriguing properties of contrastive losses. Advances in Neural Information Processing Systems, 34:11834-11845, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "type": "text", + "content": "[12] Xinyang Chen, Sinan Wang, Bo Fu, Mingsheng Long, and Jianmin Wang. Catastrophic forgetting meets negative transfer: Batch spectral shrinkage for safe transfer learning. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 634 + ], + "type": "text", + "content": "[13] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "text", + "content": "[14] Alex J DeGrave, Joseph D Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[16] Michael Feldman, Sorelle A Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. Certifying and removing disparate impact. In proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, pages 259-268, 2015." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[17] Songwei Ge, Shlok Mishra, Chun-Liang Li, Haohan Wang, and David Jacobs. Robust contrastive learning using negative samples with diminished semantics. Advances in Neural Information Processing Systems, 34:27356-27368, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 251 + ], + "type": "text", + "content": "[18] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 252, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 545, + 297 + ], + "type": "text", + "content": "[19] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models. arXiv preprint arXiv:2010.08377, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "text", + "content": "[20] Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 342, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 375 + ], + "type": "text", + "content": "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "text", + "content": "[22] Youngkyu Hong and Eunho Yang. Unbiased classification through bias-contrastive and bias-balanced learning. Advances in Neural Information Processing Systems, 34:26449-26461, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "text", + "content": "[23] Minyoung Huh, Hossein Mobahi, Richard Zhang, Brian Cheung, Pulkit Agrawal, and Phillip Isola. The low-rank simplicity bias in deep networks. arXiv preprint arXiv:2103.10427, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "text", + "content": "[24] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 511, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 544 + ], + "type": "text", + "content": "[25] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. arXiv preprint arXiv:2110.09348, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "text", + "content": "[26] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "type": "text", + "content": "[27] Byungju Kim, Hyunwoo Kim, Kyungsu Kim, Sungjin Kim, and Junmo Kim. Learning not to learn: Training deep neural networks with biased data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9012-9020, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "type": "text", + "content": "[28] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T Freeman, Phillip" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12403" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "type": "text", + "content": "Isola, Amir Globerson, Michal Irani, et al. Explaining in style: Training a gan to explain a classifier in stylespace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 693-702, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "text", + "content": "[30] Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, and Jaegul Choo. Learning debiased representation via disentangled feature augmentation. Advances in Neural Information Processing Systems, 34:25123-25133, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 289, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 289, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 289, + 208 + ], + "type": "text", + "content": "[31] Daniel Levy, Yair Carmon, John C Duchi, and Aaron Sidford. Large-scale methods for distributionally robust optimization. Advances in Neural Information Processing Systems, 33:8847-8860, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 288, + 253 + ], + "type": "text", + "content": "[32] Jiaxuan Li, Duc Minh Vo, and Hideki Nakayama. Partition-and-debias: Agnostic biases mitigation via a mixture of biases-specific experts. arXiv preprint arXiv:2308.10005, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 288, + 300 + ], + "type": "text", + "content": "[33] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14970-14979, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 302, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 288, + 346 + ], + "type": "text", + "content": "[34] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In European Conference on Computer Vision, pages 270-288. Springer, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 348, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 288, + 403 + ], + "type": "text", + "content": "[35] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 405, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 288, + 448 + ], + "type": "text", + "content": "[36] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 451, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 451, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 451, + 288, + 483 + ], + "type": "text", + "content": "[37] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 486, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 288, + 529 + ], + "type": "text", + "content": "[38] Vaishnavh Nagarajan, Anders Andreassen, and Behnam Neyshabur. Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 575 + ], + "type": "text", + "content": "[39] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 610 + ], + "type": "text", + "content": "[40] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[41] Geon Yeong Park, Sangmin Lee, Sang Wan Lee, and Jong Chul Ye. Training debiased subnetworks with contrastive weight pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7929-7938, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[42] Noam Razin and Nadav Cohen. Implicit regularization in deep learning may not be explainable by norms. Advances in neural information processing systems, 33:21174-21187, 2020." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[43] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 151 + ], + "type": "text", + "content": "[44] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 185 + ], + "type": "text", + "content": "[45] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "text", + "content": "[46] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 547, + 274 + ], + "type": "text", + "content": "[47] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 277, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 547, + 319 + ], + "type": "text", + "content": "[48] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv preprint arXiv:2206.01197, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 321, + 547, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 547, + 375 + ], + "type": "text", + "content": "[49] Enzo Tartaglione, Carlo Alberto Barbano, and Marco Grangetto. End: Entangling and disentangling deep representations for bias correction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13508-13517, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 377, + 547, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 547, + 410 + ], + "type": "text", + "content": "[50] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 547, + 443 + ], + "type": "text", + "content": "[51] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 547, + 488 + ], + "type": "text", + "content": "[52] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 491, + 547, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 547, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 547, + 523 + ], + "type": "text", + "content": "[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 525, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 547, + 568 + ], + "type": "text", + "content": "[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 570, + 547, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 547, + 644 + ], + "type": "text", + "content": "[55] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "type": "text", + "content": "[56] Zeyu Wang, Clint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8919-8928, 2020." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12404" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 384 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[57] Yadollah Yaghoobzadeh, Soroush Mehri, Remi Tachet, Timothy J Hazen, and Alessandro Sordoni. Increasing robustness to spurious correlations using forgettable examples. arXiv preprint arXiv:1911.03861, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[58] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "type": "text", + "content": "[59] Dinghuai Zhang, Kartik Ahuja, Yilun Xu, Yisen Wang, and Aaron Courville. Can subnetwork structure be the key to out-of-distribution generalization? In International Conference on Machine Learning, pages 12356-12367. PMLR, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 288, + 251 + ], + "type": "text", + "content": "[60] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 288, + 285 + ], + "type": "text", + "content": "[61] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 288, + 330 + ], + "type": "text", + "content": "[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 331, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 288, + 384 + ], + "type": "text", + "content": "[63] Roland S Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive learning inverts the data generating process. In International Conference on Machine Learning, pages 12979-12990. PMLR, 2021." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12405" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_content_list.json b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab0efe7b82053f9f0d87ea51d45d132629dac6ad --- /dev/null +++ b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_content_list.json @@ -0,0 +1,1862 @@ +[ + { + "type": "text", + "text": "SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 156, + 69, + 810, + 92 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanhui Huang* Wenzhao Zheng* Borui Zhang Jie Zhou Jiwen Lu† \nBeijing National Research Center for Information Science and Technology, China \nDepartment of Automation, Tsinghua University, China", + "bbox": [ + 186, + 119, + 833, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{huangyh22,zhang-br21}@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com; {jzhou,lujiwen}@tsinghua.edu.cn", + "bbox": [ + 88, + 172, + 929, + 209 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd6ae48a60c57315706bb11fa76122d06eda1f364d4ff290b4ded201403225f2.jpg", + "image_caption": [ + "Figure 1. Trained with only video sequences as supervision, our model can predict meaningful geometry for the scene given surround-camera RGB images, which can be further extended to semantic occupancy prediction if 2D segmentation maps are available e.g. from an off-the-shelf segmentor. This task is challenging because it completely depends on video sequences to reconstruct scenes without any 3D supervision. We observe that our model can produce dense and consistent occupancy prediction and even infer the back side of cars." + ], + "image_footnote": [], + "bbox": [ + 86, + 215, + 890, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 450, + 313, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D occupancy prediction is an important task for the robustness of vision-centric autonomous driving, which aims to predict whether each point is occupied in the surrounding 3D space. Existing methods usually require 3D occupancy labels to produce meaningful results. However, it is very laborious to annotate the occupancy status of each voxel. In this paper, we propose SelfOcc to explore a self-supervised way to learn 3D occupancy using only video sequences. We first transform the images into the 3D space (e.g., bird's eye view) to obtain 3D representation of the scene. We directly impose constraints on the 3D representations by treating them as signed distance fields. We can then render 2D images of previous and future frames as self-supervision signals to learn the 3D representations. We propose an MVS-embedded strategy to directly optimize the SDF-induced weights with multiple depth proposals. Our SelfOcc outperforms the previous best method SceneRF by $58.7\\%$ using a single frame as input on SemanticKITTI and is the first self-supervised work that produces reasonable 3D occupancy for surround cameras on nuScenes. SelfOcc produces high-quality depth and achieves state-of-the-art results on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI, KITTI-2015, and nuScenes, respectively. Code: https://github.com/huang-yh/SelfOcc.", + "bbox": [ + 73, + 483, + 473, + 862 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 450, + 630, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The vision-centric paradigm has demonstrated its potential for efficient autonomous driving [29, 31, 32, 40], where the key procedure is to obtain a perceptual 3D representation, e.g. the bird's eye view (BEV) [50] or tri-perspective view (TPV) [31] of the surrounding scene given only 2D images [40, 54]. Most existing methods achieve this using 3D annotations as supervision and achieve promising results on various tasks including semantic map construction [28, 50, 54], 3D object detection [30, 39, 40, 81], and 3D semantic occupancy prediction [6, 31, 63, 66].", + "bbox": [ + 496, + 476, + 892, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the strong performance of supervised 3D representation that facilitates accurate subsequent prediction and planning [29, 32, 64], it is very laborious to annotate 3D labels and thus difficult to scale to large-scale training data. As one workaround, data auto-labeling creates data cycles and train large networks to assist human annotation, yet it is still time-consuming and resource-hungry [11, 55, 74]. It is also noticed to be the bottleneck towards building and updating large models for autonomous driving [9]. Therefore, it is desired to learn meaningful 3D representations only from recorded video sequences without 3D labels. While existing works have explored learning 3D information (e.g., depth) from images [7, 23, 71], they are still based on 2D representations in each monocular camera space. It remains challenging to obtain reasonable and comprehensive 3D representations useful for further autonomous driving prediction and planning [29] in a self-supervised manner.", + "bbox": [ + 496, + 628, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the problem of learning self", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 94, + 886, + 333, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19946", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "supervised 3D representation (BEV and TPV) and adopt 3D occupancy prediction as the main task to evaluate the quality of the learned 3D representation, which aims to classify each point in the 3D space into occupied or empty. We first lift 2D image features into 3D space with deformable attention layers [31, 40] to enable feature interactions in the 3D space and avoid ambiguities from multiple cameras. Inspired by recent advances in neural implicit surface reconstruction [52, 65], we then transform the 3D representation into a signed distance function (SDF) field to allow for meaningful regularization and straightforward determination of occupancy boundaries. When exploiting the temporal consistency inherent in video sequences to optimize the SDF field, we propose an MVS-embedded strategy that directly optimizes the SDF-induced weight values with multiple depth proposals. Combined with the temporal supervision scheme and loss formulation tailored for 3D occupancy prediction, our method can reconstruct meaningful 3D occupancy in a self-supervised manner.", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on various datasets to demonstrate the effectiveness of the proposed SelfOcc. We apply SelfOcc to two types of 3D representations BEV and TPV. For surround-view 3D occupancy prediction, SelfOcc is the first self-supervised method that is able to produce reasonable occupancy results using only video supervision (with an IoU of 45.01 and mIoU of 9.30 on Occ3D [63]). For monocular 3D occupancy prediction, SelfOcc outperforms the previous best method SceneRF [7] by $58.7\\%$ with an IoU of 21.97 over 13.84 on SemanticKITTI [3]. To further evaluate the usefulness of the learned 3D representations, we render depth maps for each camera from the self-supervised BEV and TPV representation. Our SelfOcc outperforms the mainstream method SceneRF [7], MonoDepth2 [23], and SurroundDepth [69] on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI [3], KITTI-2015 [21], and nuScenes [4], respectively.", + "bbox": [ + 76, + 378, + 472, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 662, + 218, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Occupancy Prediction: While detection, tracking, prediction and planning have long been the focal tasks and composed a standard pipeline for autonomous driving, 3D occupancy prediction [31] or semantic scene completion [62] has recently gained attention as a more fundamental task encompassing the most fine-grained prediction of both occupancy and semantics of the environment. Pioneering works rely on 3D inputs such as depth [13, 20, 34-36, 38, 44], occupancy grids [20, 59, 72, 73], point cloud [57, 83], or truncated signed distance function (TSDF) [5, 10, 12, 13, 16, 17, 35, 62, 67, 78, 80]. Cao et al. [6] and Huang et al. [31] first attempt to predict semantic occupancy from image input only. Recent advancements related to 3D occupancy prediction include modal", + "bbox": [ + 76, + 688, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity fusion [66], multi-task learning [64], and end-to-end autonomous driving [29]. Despite the promising performance, all these methods require 3D ground truth from laborious annotation for supervision. Most related to our work is BTS [71] and SceneRF [7], which learn 3D occupancy in a self-supervised manner in monocular scenarios. However, it is nontrivial to adapt them to surround views, while our work can inherently handle monocular and surround cases.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields: One of the most predominant paradigms for self-supervised 3D reconstruction is neural radiance fields [47], which learn a multi-layer perceptron (MLP) per scene to map spatial locations to radiance and density values. To synthesize a novel view, volume rendering [46] is often leveraged which queries the MLP at multiple locations along rays and integrates radiance over density-induced weights. Training of NeRFs [1, 2, 8, 18, 47, 49] depends on a large number of images of the same scene with ground truth poses. To facilitate sparse view reconstruction and generalization for object-level data, several methods [42, 48, 75] propose to condition NeRFs on image features which establish the foundation for online NeRFs. Recently, this boundary has been further extended to scene-level online reconstruction by [7, 37, 71]. However, these works still use image features as conditions for NeRFs, ignoring 3D feature encoding crucial for 3D perception, while our method learns 3D representations explicitly.", + "bbox": [ + 496, + 210, + 892, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Depth Prediction: Training of NeRFs is time-consuming and prone to overfitting due to geometry-appearance coupling, which can be alleviated with additional depth supervision from separately trained depth prediction modules [15, 58, 68] or LiDAR data [56]. To integrate explicit depth optimization in self-supervised generalizable NeRFs, BTS [71] and SceneRF [7] leverage the reprojection loss from self-supervised depth estimation literature [22, 23, 76, 85], which minimizes the discrepancy between the target pixel and its warped counterpart. However, the extra degrees of freedom from volume rendering and the local receptive field from bilinear interpolation may still hinder depth optimization, while we propose a novel MVS-embedded approach to learn depth in NeRFs.", + "bbox": [ + 496, + 484, + 892, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 500, + 707, + 691, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we present a self-supervised vision-based 3D occupancy prediction method for autonomous driving.", + "bbox": [ + 496, + 732, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. From Image to Occupancy", + "text_level": 1, + "bbox": [ + 500, + 771, + 741, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To understand the complicated interactions among traffic participants and their surrounding environment, it is fundamental for autonomous driving systems to reconstruct the accurate 3D structure of the scene. 3D occupancy prediction is one prevalent proxy for scene reconstruction given its fine granularity and comprehensiveness, which aims at producing a voxelized prediction $\\mathbf{O} \\in \\mathcal{C}^{H \\times W \\times D}$", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8b853f5cd0d0d93dda2343e2340c2191a59949fdd685706de62f7d4cf4c46844.jpg", + "image_caption": [ + "Figure 2. Our MVS-embedded strategy effectively enlarges the receptive field of the depth optimization process across the whole epipolar line, which provides a good depth prior (Raw Argmin)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encoding per-voxel occupancy (and semantic) information. Here $H, W, D$ and $\\mathcal{C}$ denote the resolution of the occupancy grid and the set of predefined classes. First proposed by MonoScene [6], vision-based methods have become a promising approach to 3D occupancy prediction because of their low cost and effectiveness compared with 3D-based methods. Vision-based methods generally learn a mapping $\\mathcal{M} = \\mathcal{D} \\circ \\mathcal{F} \\circ \\mathcal{E}$ from RGB images $\\mathbf{I} = \\{\\mathbf{I}_n | n = 1, \\dots, N\\}$ captured by $N$ cameras to 3D occupancy $\\mathbf{O}$ :", + "bbox": [ + 75, + 273, + 467, + 409 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} = \\mathcal {M} (\\mathbf {I}) = \\mathcal {D} \\left(\\mathcal {F} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {1}\\right), \\dots , \\mathcal {E} \\left(\\mathbf {I} _ {N}\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 421, + 468, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where a 2D backbone $\\mathcal{E}$ first encodes $N$ input images into multi-scale image features $\\mathbf{F} = \\mathcal{E}(\\mathbf{I})$ , and a 3D encoder $\\mathcal{F}$ further lifts 2D features into 3D representation $\\mathbf{R} = \\mathcal{F}(\\mathbf{F})$ through attention mechanism [31, 40], lift-splat-shoot (LLS) [39] or identical transformation for monocular scenario [6], followed by a decoder network $\\mathcal{D}$ to generate the final occupancy prediction result $\\mathbf{O} = \\mathcal{D}(\\mathbf{R})$ .", + "bbox": [ + 75, + 450, + 468, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although recent supervised vision-based approaches have achieved encouraging performance on 3D occupancy prediction given only visual inputs at inference, they still rely on 3D (semantic) supervision during training, e.g. Li-DAR point cloud or dense occupancy annotation. On the contrary, self-supervised counterparts [7, 71] take only temporal correspondences inherent in video sequences as supervision to predict meaningful occupancy, and therefore can effortlessly exploit the vast amount of unlabeled driving images. However, existing self-supervised methods [7, 71] consider only one monocular camera and decode forward occupancy $\\mathbf{O}_{\\text{front}}$ directly from 2D image features by instantiating the 3D encoder $\\mathcal{F}$ with identical transformation:", + "bbox": [ + 75, + 556, + 468, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {f r o n t} = \\mathcal {D} \\left(\\mathcal {F} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {m o n o}\\right)\\right)\\right) = \\mathcal {D} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {m o n o}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 765, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite being straightforward, ignoring $\\mathcal{F}$ may cause multicamera inconsistency in the surround view scenario and inadequate geometry reasoning in the 3D space.", + "bbox": [ + 75, + 794, + 467, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose to transform the images into the BEV or TPV space to obtain a 3D representation of the scene in order to enable 3D feature interactions even in the monocular setting and also avoid ambiguities from multiple", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cameras. Similar to the supervised literature [31, 40], we leverage the deformable cross-attention (CA) to adaptively aggregate information from the image features $\\mathbf{F}$ , in which a set of learnable 3D tokens $\\mathbf{Q}$ serve as queries and the corresponding local image features serve as keys and values. Each of these learnable tokens represents a pillar area as in BEV or TPV representation. Moreover, we determine the correspondences between 3D tokens and image features with the projection matrices $\\mathbf{T} = \\{\\mathbf{T}_n|n = 1,\\dots,N\\}$ from the ego car to pixel coordinates. We interleave deformable self-attention (SA), deformable cross-attention (CA) and feed forward network (FFN) to build the 3D encoder $\\mathcal{F}$ .", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {l} = \\operatorname {F F N} _ {l} \\left(\\mathrm {C A} _ {l} \\left(\\mathrm {S A} _ {l} (\\mathbf {Q} _ {l}); \\mathbf {F}, \\mathbf {T}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 280, + 890, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the subscription $l$ represents the $l$ -th block.", + "bbox": [ + 500, + 305, + 826, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.From Occupancy to Image", + "text_level": 1, + "bbox": [ + 500, + 328, + 741, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Because 3D ground truth is unavailable, we project occupancy O from Eq. (1) back into 2D views and leverage multi-view consistency inherent in video sequences to optimize a self-supervised vision-based occupancy prediction system. Similar to previous work [7, 71], we employ differentiable volume rendering [47] to synthesize color and depth views, which enables the seamless integration of gradient information into the rendering pipeline and effective exploitation of supervision from multiple viewpoints.", + "bbox": [ + 496, + 351, + 890, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first transform the 3D representation $\\mathbf{R}$ into an SDF field $\\mathbf{S} \\in \\mathbb{R}^{H \\times W \\times D}$ , which indicates the distance of every voxel center to the nearest surface of an object, by implementing the decoder network $\\mathcal{D}$ with an MLP. We use different MLPs for different heights for BEV representation, while we construct a 3D feature volume from TPV representation with broadcasting and summation [31] before feeding the feature volume through a single MLP. For a continuous 3D coordinate $\\mathbf{p}$ , we employ bicubic interpolation (BI) to predict the SDF value $s_{\\mathbf{p}}$ and determine its occupancy status $o_{\\mathbf{p}}$ according to the sign of $s_{\\mathbf{p}}$ :", + "bbox": [ + 496, + 487, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no _ {\\mathbf {p}} = \\operatorname {s g n} \\left(s _ {\\mathbf {p}}\\right), \\quad s _ {\\mathbf {p}} = \\operatorname {B I} (\\mathbf {S}, \\mathbf {p}) = \\operatorname {B I} (\\mathcal {D} (\\mathbf {R}), \\mathbf {p}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 661, + 890, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The preference for SDF field over density field as in [7, 71] is based on two considerations: 1) An SDF field has more explicit physical meaning compared with a density field and inherently holds assumptions about the gradient magnitudes [24], thus allowing for easier regularization and optimization. 2) The signed nature of SDF facilitates straightforward determination of whether a point lies inside or outside a surface, thus enabling accurate differentiation of intricate geometries. Similar to SDF values, we can also decode color c from $\\mathbf{R}$ by using a separate MLP $\\mathcal{D}_c$ as the decoder:", + "bbox": [ + 496, + 686, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\mathbf {p}} = \\operatorname {B I} (\\mathbf {C}, \\mathbf {p}) = \\operatorname {B I} \\left(\\mathcal {D} _ {c} (\\mathbf {R}), \\mathbf {p}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 845, + 890, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Take the volume rendering process of one single ray as an example. We first determine the origin and direction of", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19948", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/15619fe9dc9b165b4de29b2ffa4604395f0c8b45b89360fe362cace141e97870.jpg", + "image_caption": [ + "Figure 3. Framework of the proposed SelfOcc for self-supervised vision-based 3D occupancy prediction. We employ an image backbone and a 3D encoder to generate 3D representations as in BEVFormer [40] or TPVFormer [31]. To render a novel view, we apply a lightweight MLP on the 3D features to predict the SDF values, color and semantic vectors. We then perform volume rendering to synthesize color, depth and semantic views. We use simple 0-thresholding to predict the occupancy volume." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 893, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the ray and uniformly sample $M$ points $\\mathbf{P} = \\{\\mathbf{p}_m|m = 1,\\dots,M\\}$ between the origin and the intersection of the ray with the border of the 3D representation. The SDF values of these points $\\mathbf{s} = \\{s_m|m = 1,\\dots,M\\}$ are then calculated with Eq. (4). According to NeuS [65], the discrete opacity value $\\alpha_{m}$ , which is the probability of the ray terminating between $\\mathbf{p}_m$ and $\\mathbf{p}_{m + 1}$ , can be derived with", + "bbox": [ + 76, + 301, + 473, + 409 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {m} = \\max \\left(\\frac {\\Phi_ {a} (s _ {m}) - \\Phi_ {a} (s _ {m + 1})}{\\Phi_ {a} (s _ {m})}, 0\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 421, + 468, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi_{a}(x) = (1 + e^{-ax})^{-1}$ is the sigmoid function with $a$ being a learnable parameter. Further, the probability of the ray passing through $\\mathbf{p}_m$ , i.e. the discrete accumulated transmittance, can be defined by $T_{m} = \\prod_{i = 1}^{m - 1}(1 - \\alpha_{i})$ . Finally, we can calculate the probability of the ray ending at $\\mathbf{p}_m$ after emission as $w_{m} = T_{m}\\alpha_{m}$ , and the rendering process is an integration of per-point attribute over the probability distribution $\\mathbf{w} = \\{w_m|m = 1,\\dots,M\\}$ . Eq. (7) calculates the rendered color $\\mathbf{c}_r$ and depth $d_r$ of the ray:", + "bbox": [ + 75, + 469, + 470, + 608 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {r} = \\sum_ {m = 1} ^ {M} w _ {m} \\mathbf {c} _ {m}, \\quad d _ {r} = \\sum_ {m = 1} ^ {M} w _ {m} d _ {m}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 621, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_m, d_m$ denote the color and depth of the $m$ -th point. Figure 3 shows the overall framework of our method.", + "bbox": [ + 76, + 676, + 467, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Occupancy-Oriented Supervision", + "text_level": 1, + "bbox": [ + 76, + 722, + 370, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we elaborate on the supervision formulation that facilitates self-supervised vision-based 3D occupancy prediction for autonomous driving.", + "bbox": [ + 75, + 747, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MVS-embedded Depth Learning. Depth supervision enhances the convergence of neural radiance fields [68], enabling faster and more effective learning of accurate geometry, especially with sparse views. Self-supervised depth estimation in NeRFs usually relies on the photometric reprojection loss $L_{rpj}$ which aims to maximize the similarity between the target image $\\mathbf{I}_t$ and sampled pixels on the", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "source image $\\mathbf{I}_s$ by refining the depth prediction $\\mathbf{D}(\\mathbf{I}_t,\\pmb {\\theta})$", + "bbox": [ + 500, + 301, + 883, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {r p j} (\\mathbf {x}, \\mathbf {I} _ {t}, \\mathbf {I} _ {s}; \\boldsymbol {\\theta}) = \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {s} (\\hat {\\mathbf {x}}) \\| \\tag {8} \\\\ = \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {s} \\left(\\pi (\\mathbf {x}, \\mathbf {D} (\\mathbf {I} _ {t}; \\boldsymbol {\\theta}) | _ {\\mathbf {x}}, \\Pi)\\right) \\|, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 325, + 890, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta, \\mathbf{x}, \\hat{\\mathbf{x}}, \\Pi$ denote the learnable parameters, a random pixel on the target image, the warped pixel on the source image and the projection matrix from the target to the source image coordinate, respectively. $\\mathbf{I}(\\mathbf{x})$ and $\\mathbf{D}|_{\\mathbf{x}}$ denote bilinear interpolation of corresponding 2D tensors, and $\\pi (\\mathbf{x},d_{\\mathbf{x}},\\Pi)$ warps a pixel $\\mathbf{x}$ to another image according to the depth $d_{\\mathbf{x}}$ and projection matrix $\\Pi$ . The inherent local receptive field of the bilinear interpolation $\\mathbf{I}_s(\\hat{\\mathbf{x}})$ has an adverse effect on the optimization property of $L_{rpj}$ since the differences between $\\mathbf{I}_t(\\mathbf{x})$ and the colors of the four adjacent grid pixels of $\\hat{\\mathbf{x}}$ are the only reference information for the optimization of $\\theta$ , which could be easily misled by poor initialization or low-texture regions.", + "bbox": [ + 496, + 371, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we propose an MVS-embedded strategy to extend the receptive field of the depth optimization for a ray across the whole epipolar line. Specifically, we first take the uniformly spaced depths $\\mathbf{d} = \\{d_m|m = 1,\\dots,M\\}$ associated with $\\mathbf{P} = \\{\\mathbf{p}_m|m = 1,\\dots,M\\}$ as depth proposals. Then we define $L_{mvs}$ as an average of the dissimilarity at each proposal $d_{m}$ with weight $w_{m}$ from volume rendering:", + "bbox": [ + 496, + 569, + 892, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {m v s} (\\mathbf {x}, \\mathbf {I} _ {t}, \\mathbf {I} _ {s}; \\boldsymbol {\\theta}) \\\\ = \\sum_ {m = 1} ^ {M} w _ {m} (\\boldsymbol {\\theta}) \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {x} (\\pi (\\mathbf {x}, d _ {m}, \\boldsymbol {\\Pi})) \\|. \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 683, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As Figure 2 shows, $L_{mvs}$ enlarges the receptive field of depth optimization by associating $\\theta$ with multiple depth proposals along the ray, thus stabilizing and facilitating self-supervised depth optimization in NeRFs. Similar to Mon-depth2 [23], we formulate the final depth loss by taking the minimum of two $L_{mvs}$ using the previous image $\\mathbf{I}_{t-1}$ or the next image $\\mathbf{I}_{t+1}$ as the source image and leverage the auto-masking strategy to filter out low texture regions:", + "bbox": [ + 496, + 753, + 893, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d e p} (\\mathbf {x}, \\mathbf {I} _ {t}) = \\operatorname {a u t o m a s k} \\left(\\min \\left(L _ {m v s} ^ {(t - 1)}, L _ {m v s} ^ {(t + 1)}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19949", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3c7133485af5a130d169a3c1f068c3dd6ce103a59610e03da5fdc12812af8073.jpg", + "table_caption": [ + "Table 1. 3D occupancy prediction performance on the Occ3D-nuScenes [63] dataset. Cons. veh. and drive. surf. represent construction vehicle and driveable surface, respectively. Our method learns meaningful geometry and reasonable semantics compared with 3D-supervised methods, and achieves even higher IoU than LiDAR-supervised TPVFormer [31]." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionIoUmIoUothersbarrierbicyclebuscarcons. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetation
MonoScene [6]3D-6.061.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.65
OccFormer [82]3D-21.935.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.97
BEVFormer [40]3D-26.885.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.028.0620.0417.69
CTF-Occ [63]3D-28.538.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.0
TPVFormer [31]3D-27.837.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.78
FB-OCC [41]3D-42.0614.3049.7130.0046.6251.5429.3029.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58
RenderOcc [53]L-23.935.6927.5614.3619.9120.5611.9612.4212.1414.3420.8118.9468.8533.3542.0143.9417.3622.61
TPVFormer [31]L17.2013.570.0014.809.3621.2716.8114.4513.7611.235.3216.0519.7310.759.439.5011.1616.5117.04
SelfOcc (BEV)C44.336.760.000.000.000.009.820.000.000.000.000.006.9747.030.0018.7516.5811.933.81
SelfOcc (TPV)C45.019.300.000.150.665.4612.540.000.802.100.000.008.2555.490.0026.3026.5414.225.60
", + "bbox": [ + 80, + 132, + 887, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we omit the learnable parameters $\\theta$ .", + "bbox": [ + 76, + 334, + 359, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervision from temporal frames. Unlike depth estimation which predicts only the visible surfaces, 3D occupancy prediction reconstructs the complete geometry of the scene which requires self-supervision signals from multiple viewpoints as in NeRFs. Considering the limited overlapping field of view for cameras, we take advantage of the temporal frames as supervision. Given a video sequence $\\mathbf{V} = \\{\\mathbf{V}_1,\\dots,\\mathbf{V}_T\\}$ with $T$ frames and the current timesamp $t$ , we use $\\mathbf{V}_t$ as the input to our model $\\mathcal{M}$ to predict the SDF and radiance fields $\\{\\mathbf{S}_t,\\mathbf{C}_t\\}$ . Then we decide whether to use a temporal frame as supervision with probability $p$ . If positive, we randomly sample one frame $t^\\prime$ for supervision satisfying that the distance between the ego cars at timestamps $t$ and $t^\\prime$ falls inside $[l_1,l_2]$ , which guarantees diversity and adheres to the range of the 3D representation. Otherwise, we use the current frame $t$ for supervision.", + "bbox": [ + 75, + 349, + 470, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SDF Field Regularization. Neural radiance fields with sparse views suffer from overfitting because of a large solution set [51], which is exacerbated in autonomous driving with outward-facing cameras. Therefore, it is crucial to regularize the SDF field in order to restrict the solution set. To learn a smooth SDF field, we minimize the magnitude of the second-order derivatives of the SDF field with a Hessian loss [79] $L_{H}(\\mathbf{p}) = \\| \\mathbf{H}(\\mathbf{p})\\|_{1}$ , where $\\| \\cdot \\|_1$ is the elementwise matrix 1-norm, and $\\mathbf{H}(\\mathbf{p})$ denotes the Hessian matrix of the SDF field with respect to the 3D coordinates at location $\\mathbf{p}$ . To impose sparsity of occupancy in the invisible area such as those blocked by a wall, we propose a simple yet effective regularization $L_{s}(\\mathbf{p}) = \\max(-s_{\\mathbf{p}}, 0)$ , which encourages negative SDF values to become larger. In addition, we include the Eikonal term [24] $L_{E}(\\mathbf{p}) = ||\\nabla s(\\mathbf{p})|| - 1|$ for compliance with physical meaning of SDF. For optional semantic prediction, we use 2D segmentation maps from off-the-shelf segmentors as supervision similar to colors.", + "bbox": [ + 75, + 590, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Til now, we define the training loss $L$ as follows:", + "bbox": [ + 96, + 863, + 419, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {d e p} + \\lambda_ {c} L _ {r g b} + \\lambda_ {e} L _ {E} + \\lambda_ {H} L _ {H} + \\lambda_ {s} L _ {s} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 885, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b333d38e19cdc69d9f66eb7d4b2209313ef8006644c88d7056ad9f9f8e058b01.jpg", + "table_caption": [ + "Table 2. 3D occupancy prediction performance on the SemanticKITTI [3] dataset. MonoScene* is supervised by depth predictions of Monodepth2 [23] trained with ground-truth poses." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionIoUPrec.Rec.
3DDepth Image
MonoScene [6]✓37.1449.9059.24
\\(LMSCNet^{rgb}\\) [59]✓12.0813.0063.16
\\(3DSketch^{rgb}\\) [10]✓12.0112.9562.31
\\(AICNet^{rgb}\\) [36]✓11.2811.8470.89
MonoScene [6]✓13.5316.9840.06
MonoScene* [6]11.1813.1540.22
SceneRF [7]✓13.8417.2840.96
SelfOcc (BEV)✓20.9537.2732.37
SelfOcc (TPV)✓21.9734.8337.31
", + "bbox": [ + 501, + 377, + 880, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the four $\\lambda$ s are hyperparameters and $L_{rgb}$ measures the dissimilarity between the rendered color $\\mathbf{c}_r$ of a ray and the ground truth color on the target image. For semantic prediction, we optionally add a semantic loss $L_{sem}$ similar to $L_{rgb}$ and an off-the-shelf 2D segmentor.", + "bbox": [ + 496, + 565, + 893, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 655, + 633, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Task Descriptions", + "text_level": 1, + "bbox": [ + 500, + 680, + 671, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The 3D occupancy prediction task aims to predict the dense occupancy (and semantic) states for a voxel grid, which is a fundamental task in autonomous driving due to its fine granularity and least abstraction of the real world. We conduct this task on the Occ3D-nuScenes [63] dataset, and use IoU and mIoU as geometric and semantic metrics, respectively. In addition, we evaluate our method on the SemanticKITTI [3] dataset for the geometry-only 3D occupancy prediction task and report IoU, Precision and Recall.", + "bbox": [ + 496, + 703, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The novel depth synthesis task synthesizes depth maps from novel viewpoints within a certain distance given the input frame [7]. This task is more reliable as a proxy for 3D reconstruction compared with novel view synthesis because", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19950", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a26d3e74758eb5d0bc52fe8f0de063b15395674371db9c6b1c114d1512afac8e.jpg", + "table_caption": [ + "Table 3. Novel depth synthesis on SemanticKITTI [3] and nuScenes [4]. We use the results reported in SceneRF [7] and all methods are trained using only images with ground truth poses. We adapt SceneRF [7] to the nuScenes dataset based on the official implementation." + ], + "table_footnote": [], + "table_body": "
MethodDatasetAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
MonoDepth2 [23]0.52597.11314.431.029210.4426.3241.43
SynSin [70]0.36113.4838.8240.429052.6174.5686.50
PixelNeRF [75]0.23642.0806.4490.335465.8185.4392.90
MINE [37]SemanticKITTI0.22481.7876.3430.328365.8785.5293.30
VisionNerf [42]0.20541.4905.8410.307369.1188.2894.37
SceneRF [7]0.16811.2915.7810.285175.0789.0994.50
SelfOcc (TPV)0.15621.1535.3490.253478.9291.1295.92
SceneRF [7]nuScenes0.763115.3411.500.624040.2664.8677.02
SelfOcc (TPV)0.40036.7238.4600.416564.5080.4088.17
", + "bbox": [ + 98, + 119, + 867, + 291 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d4600b0850b3c5de3ff6116110ffd850941f79429996d3375d827efab3e42fb7.jpg", + "table_caption": [ + "Table 4. Self-supervised depth estimation on the nuScenes [4] and KITTI-2015 [21] dataset. M: trained with monocular sequence. MS: trained with both monocular sequence and stereo images. SV: trained with surround-view images. TO: temporal offline refinement. We train SurroundDepth* with ground truth poses based on the official implementation." + ], + "table_footnote": [], + "table_body": "
MethodDatasetTypeScale MethodAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
Monodepth2 [23]nuScenesSVMedian0.2873.3497.1840.3450.6410.8450.925
FSM [26]SVMedian0.299------
SurroundDepth [69]SVSfM Pretrain0.2804.4017.4670.3640.6610.8440.917
SurroundDepth* [69]SVPose GT0.3427.9178.0190.3600.7160.8580.918
SimpleOcc [19]SVPose GT0.2243.3837.1650.3330.7530.8770.930
SelfOcc (TPV)SVPose GT0.2152.7436.7060.3160.7530.8750.932
R3D3 [60]nuScenesTOExtrinsics0.2534.7597.150-0.729--
Johnston [33]KITTI-2015MMedian0.1060.8614.6990.1850.8890.9620.982
FeatDepth(R50) [61]MMedian0.1040.7294.4810.1790.8930.9650.984
PackNet-SfM [25]MMedian0.1070.8024.5380.1860.8890.9620.981
R-MSFM6 [86]MMedian0.1080.7484.4700.1850.8890.9630.982
DevNet [84]MMedian0.1000.6994.4120.1740.8930.9660.985
SelfOcc (TPV)MPose GT0.1030.7924.6730.1870.8770.9590.982
MonoDepth2(R50) [23]KITTI-2015MSExtrinsics0.1060.8184.750.1960.8740.9570.979
FeatDepth(R50) [61]MSExtrinsics0.0990.6974.4270.1840.8890.9630.982
R-MSFM6 [86]MSExtrinsics0.1080.7534.4690.1850.8880.9630.982
DevNet [84]MSExtrinsics0.0950.6714.3650.1740.8950.9700.988
BTS [71]MSPose GT0.1020.7514.4070.1880.8820.9610.982
SelfOcc (TPV)MSPose GT0.0990.7114.5860.1860.8800.9600.982
", + "bbox": [ + 99, + 339, + 870, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the rendered color is an integration of predicted radiance at sampled locations, which introduces unnecessary degrees of freedom. We conduct experiments on nuScenes [4] and SemanticKITTI [3] and use Abs Rel, Sq Rel, RMSE, RMSE log and threshold accuracies $(\\delta 1, \\delta 2, \\delta 3)$ as metrics.", + "bbox": [ + 75, + 648, + 468, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The depth estimation task is a traditional task for 3D perception, which predicts the pixel-aligned depth map for a given frame. We conduct surround-view depth estimation on nuScenes [4] and monocular depth estimation on KITTI-2015 [21]. The same metrics are used as in novel depth synthesis. We detail the datasets in the supplementary.", + "bbox": [ + 75, + 726, + 468, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 830, + 294, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the identical network architecture for all three tasks, while the temporal supervision and the loss formulation may differ according to their respective objectives. We", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adopt ResNet50 [27] pretrained on ImageNet [14] as the 2D backbone and an feature pyramid network (FPN) [43] to generate multi-scale image features. The 3D encoder $\\mathcal{F}$ is the same as in BEVFormer [40] or TPVFormer [31] depending on the 3D representation we use. We further employ a two-layer MLP as the decoder network $\\mathcal{D}$ to generate the SDF field, color and semantic logits (optional).", + "bbox": [ + 496, + 648, + 890, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We activate temporal supervision for 3D occupancy prediction and novel view synthesis which both involve scene reconstruction from multiple viewpoints, while we use only the current frame as supervision for depth estimation since it predicts the geometry only for the input view. As for loss formulation, we use $L_{dep}$ , $L_{E}$ and an edge loss $L_{edg}$ [22] for depth estimation since RGB supervision from the input view is meaningless, while $L_{H}$ and $L_{s}$ focus on improving the overall geometry, thus helpless for depth es", + "bbox": [ + 496, + 763, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8398e98949e52a90bf7242a434ae1750223b5147e9f22f2969af8c960866e6fa.jpg", + "table_caption": [ + "Table 5. Architecture ablation on the nuScenes dataset. $L_{mvs}$ : the proposed $L_{mvs}(\\checkmark)$ or $L_{rpj}$ for depth optimization. SDF: SDF field ( $\\checkmark$ ) or density field for volume rendering." + ], + "table_footnote": [], + "table_body": "
LmvsSDFOcc.Novel DepthDepth
IoUmIoUAbs Rel↓RMSE↓Abs Rel↓RMSE↓
✓18.902.810.810112.4080.65513.242
✓36.686.100.40078.1490.3188.524
✓✓38.436.940.43768.7340.3127.544
", + "bbox": [ + 88, + 132, + 455, + 209 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be7a29dd85e42a4ffede6079ff723c499f18d5dbd638a19feca1dd49298aa76c.jpg", + "table_caption": [ + "Table 6. Temporal supervision ablation on nuScenes [4]. We do not use semantic supervision for occupancy prediction here." + ], + "table_footnote": [], + "table_body": "
SupervisionRatioOcc.NovelDepthDepth
Curr.Prev.NextIoUPrec.Rec.Abs RelRMSEAbs RelRMSE
10039.6766.9649.330.4718.8520.3127.544
21140.5067.5150.310.4468.8270.3658.046
11140.4763.9852.410.4388.7340.3637.975
12240.4363.3552.780.4428.7810.3667.998
01142.4870.0851.890.4628.7320.3928.075
", + "bbox": [ + 78, + 250, + 467, + 349 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/096927ef2c5b16393b7e0351936c1f6b6abf5bce7ca8eb0542da09f082f7988e.jpg", + "table_caption": [ + "Table 7. Loss and regularization ablation on the SemanticKITTI [3] (occupancy prediction and novel depth synthesis) and KITTI-2015 [21] (depth estimation) datasets." + ], + "table_footnote": [], + "table_body": "
LdepLrgbLHLsOcc.Novel DepthDepth
IoUPrec.Rec.Abs RelRMSEAbs RelRMSE
✓9.819.8596.180.385211.67--
✓23.5141.0135.510.15795.3480.1024.676
✓✓24.0737.7839.880.15735.3520.1004.773
✓✓✓12.2713.0268.140.17225.7480.1075.091
✓✓✓✓22.1633.5439.500.16945.6010.1135.078
", + "bbox": [ + 78, + 407, + 467, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2fd7c0769098dfb48e549522f7fd067b121e774e1fd08c819dfd8f03dfc59e3.jpg", + "image_caption": [ + "Figure 4. Visualization of the effect of different loss functions. The occupancy quality gradually improves with $L_{dep}$ , $L_{H}$ and $L_{s}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 515, + 467, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "timation. For novel depth synthesis, we leverage $L_{dep}$ , $L_{rgb}$ and $L_{E}$ . And for 3D occupancy prediction, we further activate $L_{H}$ and $L_{s}$ on the basis of novel depth synthesis to impose smoothness and sparsity constraints on the predicted occupancy grid. For semantic prediction, we use the off-the-shelf open-vocabulary segmentation network OpenSeeD [77] to generate pseudo segmentation labels for nuScenes [4]. We use the AdamW [45] optimizer with an initial learning rate of 1e-4 which decays to zero following the cosine schedule. We train our models for 12 epochs on the nuScenes [4] dataset and 24 epochs on the SemanticKITTI [3] and KITTI-2015 [21] datasets.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results and Analysis", + "text_level": 1, + "bbox": [ + 500, + 90, + 694, + 106 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Occupancy Prediction. We report the results for the 3D occupancy prediction task in Table 1 and 2. On the Occ3D-nuScenes [63] dataset, our methods achieve comparable IoU and reasonable mIoU without any form of 3D supervision compared with supervised approaches. The IoU metric $(45.01\\%)$ indicates that SelfOcc learns meaningful geometry given only video sequences. We notice that SelfOcc is better at predicting background classes than foreground classes, which might be related to the semantic misalignment of the open-vocabulary 2D segmentation network we use. Nonetheless, our method still outperforms MonoScene [6] supervised by ground truth in mIoU and LiDAR-supervised TPVFormer [31] in IoU.", + "bbox": [ + 496, + 121, + 890, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the SemanticKITTI [3] dataset, we use the results reported in SceneRF [7], where baselines are grouped into 3D, depth and image supervised categories, and the superscription ${}^{rgb}$ indicates the adaptation from 3D input to image input. From Table 2, SelfOcc sets the new state-of-the-art for vision-based depth-supervised and self-supervised 3D occupancy prediction tasks, which outperforms SceneRF [7] by ${58.7}\\%$ in IoU,with much higher Precision and slightly lower Recall. Visualizations for 3D occupancy predictions are shown in Figure 5.", + "bbox": [ + 496, + 325, + 890, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel Depth Synthesis. We report the results for novel depth synthesis in Table 3, which averages depth metrics over all frames no further than 10 meters away from the input frame. Following SceneRF [7], we compare against methods based on monocular depth estimation [23], generalizable NeRF [37, 42, 75] and 3D-aware GAN [70]. Our method outperforms previous state-of-the-art SceneRF [7] on all metrics on the SemanticKITTI [3] dataset. In addition, we modify SceneRF for the nuScenes [4] dataset by randomly selecting one camera every iteration during training, and directly report the depth metric for the input frame considering the complexity of volume rendering conditioned on image features from surround cameras. SelfOcc still achieves better results than SceneRF, which confirms its superiority in 3D reconstruction from surround views. We provide novel depth visualizations in the supplementary.", + "bbox": [ + 496, + 484, + 890, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth Estimation. We report the results for depth estimation in Table 4. Our method achieves state-of-the-art performance on nuScenes for self-supervised surround-view depth estimation. We also train SurroundDepth [69] using the official implementation with ground-truth poses to eliminate their influence, which indicates that ground-truth poses do not necessarily boost performance compared with median-scaling applied for scale-ambiguous methods. Additionally, SelfOcc performs on par with the state-of-the-art for self-supervised depth estimation on KITTI-2015 [21]. Visualizations are provided in the supplementary.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19952", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec6a5899ee9c761da65f0dfb932e3b00be50f5d89b50d7fd90da1fe1c50f61c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 87, + 316, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edccc952ae8cd9ba197a15d7b304e802302fa6e30118f12ba70338e710b88496.jpg", + "image_caption": [ + "Semantic Occupancy Prediction" + ], + "image_footnote": [], + "bbox": [ + 325, + 97, + 462, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2042a438150f9db3fc1ee4c8c58314ca60d17ca71b72a0412f52523149151f05.jpg", + "image_caption": [ + "Semantic Occupancy GT" + ], + "image_footnote": [], + "bbox": [ + 467, + 99, + 604, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d061b9992af1f3b3ba806096e8e503010b3ce011b7fef528dbd2b0daf9fc918.jpg", + "image_caption": [ + "Geometric Occupancy Prediction" + ], + "image_footnote": [], + "bbox": [ + 612, + 99, + 745, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9fb842a46eb7d5a21734dbbfd05f63a8a041607ae85e772ffda6b019f58b5689.jpg", + "image_caption": [ + "Geometric Occupancy GT" + ], + "image_footnote": [], + "bbox": [ + 754, + 97, + 890, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7588c220b99e22bcd38339a533f2b9282a817e78b5d4f4251fb88110bfb42a11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 170, + 316, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63cb76b35759bbab4a90d88da8dd0c6a0ff9c71ad4a99fdeb02471c71dd5afb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 171, + 460, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/202d1eeb53be0876bc39c50c154dc5c61c1ed1327fb74fa572cd8047fea29588.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 171, + 602, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a47d1ca153c83a1aa239bdba9bbb932277c0167efe018071bb485485a2e2bbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 172, + 745, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a69f4ddad804a9011b47b2c09079bf3a3e400b73d78d7998faabe670f82d6e04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 171, + 890, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/368f2eb8a629d3f1e4357d64b8a6955422164249c9bf3dfae2702b6ba41c7152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 244, + 316, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62d9aa8bfae0255fc815d5334757d8b976f23ee64bdc1ac0d87156dbaaede636.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 244, + 460, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bdaa6aed7018b969a106d843c98abd73c782c39a84704d2086798fd436341dda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 244, + 602, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72819640aed9cbc8d5b701b55233029e74201010987a22b6a976325140d84c5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 244, + 745, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f13c3971692f81024c4886725341eee39f18a07daf0e5fb73c986c861499b99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 244, + 890, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0de870ab8bcd2f24bbda702b515c0437caad3ea23e59e5f66372a90e6d61fedb.jpg", + "image_caption": [], + "image_footnote": [ + "BACK LEFT", + "BACK", + "BACK RIGHT" + ], + "bbox": [ + 80, + 318, + 316, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee1ce8b951504b5e64e871396105535f271762991502ecc9c49c227bcfab4b5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 320, + 460, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec3ba5f0200396d913c9d841988dede7372a9a15320383a00ea360141e403738.jpg", + "image_caption": [], + "image_footnote": [ + "other flat", + "pedc" + ], + "bbox": [ + 467, + 320, + 602, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e91500f337661ed575abab38f918db33867c79c2d21c78c228aa0e28bc21f0d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 320, + 745, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/058acdca8fe2904be7f1efe1ebdd7dbe55ab212b9d284f73ba1309a36bf31d09.jpg", + "image_caption": [], + "image_footnote": [ + "higher" + ], + "bbox": [ + 754, + 320, + 890, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b94dc920af79801cfdc67c72583b81a59122694651e28b8cb8d0bf08bcaae681.jpg", + "image_caption": [ + "Veabl s", + "Figure 5. Visualization results for 3D occupancy prediction on the Occ3D-nuScenes [63] dataset. Our method achieves comparable visualization quality with ground truth for both semantic and geometric occupancy prediction tasks." + ], + "image_footnote": [], + "bbox": [ + 81, + 404, + 109, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db925c8d306551a7c10073ee772ed908a67c61baec771db81994f65799ec1054.jpg", + "image_caption": [ + "${13}/{14}$" + ], + "image_footnote": [], + "bbox": [ + 155, + 405, + 173, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b52e703fd9a719beebfd36b7ba893aaf7edc9eaf1e1772bdd39899fe8971a94.jpg", + "image_caption": [ + "ous" + ], + "image_footnote": [], + "bbox": [ + 223, + 405, + 241, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dbb839b407618438fe650a706ec637f0739d8332e852bfb88cfdc81db63ec1d.jpg", + "image_caption": [ + "terrain" + ], + "image_footnote": [], + "bbox": [ + 294, + 405, + 331, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e493d18eeaf104a74429ce83be94d0226542b28b6537ae212e8ba2222d575e27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 405, + 424, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7411a4c4205fdbd1355c4dc73eecf4a66b78b744676b5ed7ccada8d40d7c74fc.jpg", + "image_caption": [ + "dewalk" + ], + "image_footnote": [], + "bbox": [ + 436, + 405, + 452, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fc63cd1b1ef2ace70724c2788b9096be492123d74c4488c756ce762255ce4f5.jpg", + "image_caption": [ + "lower" + ], + "image_footnote": [ + "1", + "${13}/{14}$", + "lower" + ], + "bbox": [ + 653, + 405, + 656, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 472, + 228, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies on architectural choices, temporal supervision, and loss formulations to validate the effectiveness of every component of our method. All experiments are trained for half of the original epochs.", + "bbox": [ + 75, + 498, + 468, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture. We investigate the effect of $L_{mvs}$ and the SDF field in Table 5. MVS-embedded depth optimization brings consistent improvement for all three tasks. While the SDF field outperforms the density field in occupancy prediction and depth estimation, it achieves worse results in novel depth synthesis. We think it is because the SDF field holds an inherent smoothness prior and is thus more difficult to optimize for novel depth synthesis.", + "bbox": [ + 75, + 561, + 468, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal Supervision. We report the results of different temporal supervision settings in Table 6. It is shown that increasing temporal supervision benefits occupancy prediction, but deteriorates depth estimation performance. In addition, the best score for novel depth synthesis is achieved with a moderate level of temporal supervision (1:1:1).", + "bbox": [ + 75, + 686, + 468, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Formulations. We ablate the effect of loss functions in Table 7. Note that we replace $L_{rbg}$ with $L_{edg}$ for depth estimation task. Models trained with depth loss $L_{dep}$ and color loss $L_{rgb}$ achieve best performance for depth-related tasks. As shown in Figure 4, we use the Hessian loss $L_{H}$ and sparsity loss $L_{s}$ to regularize the SDF field for better visualization quality, although models trained without them can also achieve high IoU during evaluation. In fact,", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the sparsity loss makes the prediction more similar to the ground truth, but it may not improve the IoU under the evaluation protocol of using camera masks as in Occ3D [63].", + "bbox": [ + 496, + 473, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. 3D Occupancy Visualization", + "text_level": 1, + "bbox": [ + 498, + 527, + 754, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide the visualization results for 3D occupancy prediction on nuScenes [4] in Figure 5. Despite being trained with video sequences only, our model produces meaningful geometric and reasonable semantic predictions.", + "bbox": [ + 496, + 551, + 890, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 625, + 619, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented a self-supervised vision-based 3D occupancy prediction method (SelfOcc) for autonomous driving, which predicts meaningful geometry and semantics from video sequences only. To facilitate feature extraction in the 3D space, we have proposed to condition NeRFs on 3D representations. Furthermore, we have designed an MVS-embedded strategy to boost depth optimization for NeRFs. SelfOcc has set new state-of-the-art for self-supervised vision-based occupancy prediction on SemanticKITTI and has predicted reasonable 3D occupancy for surrounding cameras on Occ3D for the first time.", + "bbox": [ + 496, + 650, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 830, + 660, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62321005, Grant 62336004, and Grant 62125603.", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19953", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 2", + "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 5, 6, 7", + "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 7, 8", + "[5] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2", + "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2, 3, 5, 7", + "[7] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 1, 2, 3, 5, 6, 7", + "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 2", + "[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1", + "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, 2020. 2, 5", + "[11] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In ECCV, pages 680-697. Springer, 2022. 1", + "[12] Yueh-Tung Chen, Martin Garbade, and Juergen Gall. 3d semantic scene completion from a single depth image using adversarial training. In ICIP, pages 1835-1839. IEEE, 2019. 2", + "[13] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161. PMLR, 2021. 2", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2", + "[16] Aloisio Dourado, Hansung Kim, Teofilo E de Campos, and Adrian Hilton. Semantic scene completion from a single 360-degree image and depth map. In VISIGRAPP (5: VIS-APP), pages 36-46, 2020. 2", + "[17] Aloisio Dourado, Teofilo E De Campos, Hansung Kim, and Adrian Hilton. Edgenet: Semantic scene completion from a single rgb-d image. In ICPR, pages 503-510. IEEE, 2021. 2", + "[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501-5510, 2022. 2", + "[19] Wanshui Gan, Ningkai Mo, Hongbin Xu, and Naoto Yokoya. A simple attempt for 3d occupancy estimation in autonomous driving. arXiv preprint arXiv:2303.10076, 2023. 6", + "[20] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPR Workshops, pages 0–0, 2019. 2", + "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 2, 6, 7", + "[22] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 2, 6", + "[23] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 1, 2, 4, 5, 6, 7", + "[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 3, 5", + "[25] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, pages 2485-2494, 2020. 6", + "[26] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. arXiv preprint arXiv:2104.00152, 2021. 6", + "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 6", + "[28] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 1", + "[29] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19954", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1", + "[31] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2, 3, 4, 5, 6, 7", + "[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1", + "[33] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 6", + "[34] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2", + "[35] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters, 5(1):219-226, 2019. 2", + "[36] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, 2020. 2, 5", + "[37] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 6, 7", + "[38] Siqi Li, Changqing Zou, Yipeng Li, Xibin Zhao, and Yue Gao. Attention-based multi-modal fusion network for semantic scene completion. In AAAI, pages 11402-11409, 2020. 2", + "[39] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 3", + "[40] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 1, 2, 3, 4, 5, 6", + "[41] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 5", + "[42] Kai-En Lin, Yen-Chen Lin, Wei-Sheng Lai, Tsung-Yi Lin, Yichang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2, 6, 7", + "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6", + "[44] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. NIPS, 31, 2018. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7", + "[46] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2", + "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3", + "[48] Norman Müller, Andrea Simonelli, Lorenzo Porzi, Samuel Rota Bulo, Matthias Nießner, and Peter Kontschieder. Autorf: Learning 3d object radiance fields from single view observations. In CVPR, pages 3971-3980, 2022. 2", + "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ToG, 41(4):1-15, 2022. 2", + "[50] Mong H Ng, Kaahan Radia, Jianfei Chen, Dequan Wang, Ionel Gog, and Joseph E Gonzalez. Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. arXiv preprint arXiv:2006.11436, 2020. 1", + "[51] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 5", + "[52] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 2", + "[53] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 5", + "[54] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 1", + "[55] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, pages 6134-6144, 2021. 1", + "[56] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2", + "[57] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2", + "[58] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2", + "[59] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In ThreeDV, 2020. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, pages 3216-3226, 2023. 6", + "[61] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 6", + "[62] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2", + "[63] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 5, 7, 8", + "[64] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 1, 2", + "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4", + "[66] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 2", + "[67] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In ICCV, pages 8608–8617, 2019. 2", + "[68] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 2, 4", + "[69] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, pages 539-549. PMLR, 2023. 2, 6, 7", + "[70] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 6, 7", + "[71] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9076-9086, 2023. 1, 2, 3, 6", + "[72] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2", + "[73] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1", + "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2, 6, 7", + "[76] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In CVPR, 2018. 2", + "[77] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In ICCV, pages 1020-1031, 2023. 7", + "[78] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, pages 733-749, 2018. 2", + "[79] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 5", + "[80] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In ICCV, pages 7801-7810, 2019. 2", + "[81] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1", + "[82] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 5", + "[83] Min Zhong and Gang Zeng. Semantic point completion network for 3d semantic scene completion. In ECAI 2020, pages 2824-2831. IOS Press, 2020. 2", + "[84] Kaichen Zhou, Lanqing Hong, Changhao Chen, Hang Xu, Chaoqiang Ye, Qingyong Hu, and Zhenguo Li. Devnet: Self-supervised monocular depth learning via density volume construction. In ECCV, pages 125-142. Springer, 2022. 6", + "[85] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017. 2", + "[86] Zhongkai Zhou, Xinnan Fan, Pengfei Shi, and Yuanxue Xin. R-MSFM: Recurrent multi-scale feature modulation for monocular depth estimating. In ICCV, 2021. 6" + ], + "bbox": [ + 501, + 92, + 890, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19956", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_model.json b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35c8f2944fc6cdff8b16c9d40f086d074a3b6da4 --- /dev/null +++ b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_model.json @@ -0,0 +1,2939 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.07, + 0.812, + 0.093 + ], + "angle": 0, + "content": "SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.12, + 0.834, + 0.174 + ], + "angle": 0, + "content": "Yuanhui Huang* Wenzhao Zheng* Borui Zhang Jie Zhou Jiwen Lu† \nBeijing National Research Center for Information Science and Technology, China \nDepartment of Automation, Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.174, + 0.93, + 0.21 + ], + "angle": 0, + "content": "{huangyh22,zhang-br21}@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com; {jzhou,lujiwen}@tsinghua.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.217, + 0.891, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.894, + 0.44 + ], + "angle": 0, + "content": "Figure 1. Trained with only video sequences as supervision, our model can predict meaningful geometry for the scene given surround-camera RGB images, which can be further extended to semantic occupancy prediction if 2D segmentation maps are available e.g. from an off-the-shelf segmentor. This task is challenging because it completely depends on video sequences to reconstruct scenes without any 3D supervision. We observe that our model can produce dense and consistent occupancy prediction and even infer the back side of cars." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.451, + 0.314, + 0.466 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.484, + 0.474, + 0.863 + ], + "angle": 0, + "content": "3D occupancy prediction is an important task for the robustness of vision-centric autonomous driving, which aims to predict whether each point is occupied in the surrounding 3D space. Existing methods usually require 3D occupancy labels to produce meaningful results. However, it is very laborious to annotate the occupancy status of each voxel. In this paper, we propose SelfOcc to explore a self-supervised way to learn 3D occupancy using only video sequences. We first transform the images into the 3D space (e.g., bird's eye view) to obtain 3D representation of the scene. We directly impose constraints on the 3D representations by treating them as signed distance fields. We can then render 2D images of previous and future frames as self-supervision signals to learn the 3D representations. We propose an MVS-embedded strategy to directly optimize the SDF-induced weights with multiple depth proposals. Our SelfOcc outperforms the previous best method SceneRF by \\(58.7\\%\\) using a single frame as input on SemanticKITTI and is the first self-supervised work that produces reasonable 3D occupancy for surround cameras on nuScenes. SelfOcc produces high-quality depth and achieves state-of-the-art results on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI, KITTI-2015, and nuScenes, respectively. Code: https://github.com/huang-yh/SelfOcc." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.451, + 0.632, + 0.466 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.893, + 0.628 + ], + "angle": 0, + "content": "The vision-centric paradigm has demonstrated its potential for efficient autonomous driving [29, 31, 32, 40], where the key procedure is to obtain a perceptual 3D representation, e.g. the bird's eye view (BEV) [50] or tri-perspective view (TPV) [31] of the surrounding scene given only 2D images [40, 54]. Most existing methods achieve this using 3D annotations as supervision and achieve promising results on various tasks including semantic map construction [28, 50, 54], 3D object detection [30, 39, 40, 81], and 3D semantic occupancy prediction [6, 31, 63, 66]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Despite the strong performance of supervised 3D representation that facilitates accurate subsequent prediction and planning [29, 32, 64], it is very laborious to annotate 3D labels and thus difficult to scale to large-scale training data. As one workaround, data auto-labeling creates data cycles and train large networks to assist human annotation, yet it is still time-consuming and resource-hungry [11, 55, 74]. It is also noticed to be the bottleneck towards building and updating large models for autonomous driving [9]. Therefore, it is desired to learn meaningful 3D representations only from recorded video sequences without 3D labels. While existing works have explored learning 3D information (e.g., depth) from images [7, 23, 71], they are still based on 2D representations in each monocular camera space. It remains challenging to obtain reasonable and comprehensive 3D representations useful for further autonomous driving prediction and planning [29] in a self-supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this paper, we explore the problem of learning self" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.334, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.379 + ], + "angle": 0, + "content": "supervised 3D representation (BEV and TPV) and adopt 3D occupancy prediction as the main task to evaluate the quality of the learned 3D representation, which aims to classify each point in the 3D space into occupied or empty. We first lift 2D image features into 3D space with deformable attention layers [31, 40] to enable feature interactions in the 3D space and avoid ambiguities from multiple cameras. Inspired by recent advances in neural implicit surface reconstruction [52, 65], we then transform the 3D representation into a signed distance function (SDF) field to allow for meaningful regularization and straightforward determination of occupancy boundaries. When exploiting the temporal consistency inherent in video sequences to optimize the SDF field, we propose an MVS-embedded strategy that directly optimizes the SDF-induced weight values with multiple depth proposals. Combined with the temporal supervision scheme and loss formulation tailored for 3D occupancy prediction, our method can reconstruct meaningful 3D occupancy in a self-supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.473, + 0.651 + ], + "angle": 0, + "content": "We conduct extensive experiments on various datasets to demonstrate the effectiveness of the proposed SelfOcc. We apply SelfOcc to two types of 3D representations BEV and TPV. For surround-view 3D occupancy prediction, SelfOcc is the first self-supervised method that is able to produce reasonable occupancy results using only video supervision (with an IoU of 45.01 and mIoU of 9.30 on Occ3D [63]). For monocular 3D occupancy prediction, SelfOcc outperforms the previous best method SceneRF [7] by \\(58.7\\%\\) with an IoU of 21.97 over 13.84 on SemanticKITTI [3]. To further evaluate the usefulness of the learned 3D representations, we render depth maps for each camera from the self-supervised BEV and TPV representation. Our SelfOcc outperforms the mainstream method SceneRF [7], MonoDepth2 [23], and SurroundDepth [69] on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI [3], KITTI-2015 [21], and nuScenes [4], respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.22, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.473, + 0.902 + ], + "angle": 0, + "content": "3D Occupancy Prediction: While detection, tracking, prediction and planning have long been the focal tasks and composed a standard pipeline for autonomous driving, 3D occupancy prediction [31] or semantic scene completion [62] has recently gained attention as a more fundamental task encompassing the most fine-grained prediction of both occupancy and semantics of the environment. Pioneering works rely on 3D inputs such as depth [13, 20, 34-36, 38, 44], occupancy grids [20, 59, 72, 73], point cloud [57, 83], or truncated signed distance function (TSDF) [5, 10, 12, 13, 16, 17, 35, 62, 67, 78, 80]. Cao et al. [6] and Huang et al. [31] first attempt to predict semantic occupancy from image input only. Recent advancements related to 3D occupancy prediction include modal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "ity fusion [66], multi-task learning [64], and end-to-end autonomous driving [29]. Despite the promising performance, all these methods require 3D ground truth from laborious annotation for supervision. Most related to our work is BTS [71] and SceneRF [7], which learn 3D occupancy in a self-supervised manner in monocular scenarios. However, it is nontrivial to adapt them to surround views, while our work can inherently handle monocular and surround cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Neural Radiance Fields: One of the most predominant paradigms for self-supervised 3D reconstruction is neural radiance fields [47], which learn a multi-layer perceptron (MLP) per scene to map spatial locations to radiance and density values. To synthesize a novel view, volume rendering [46] is often leveraged which queries the MLP at multiple locations along rays and integrates radiance over density-induced weights. Training of NeRFs [1, 2, 8, 18, 47, 49] depends on a large number of images of the same scene with ground truth poses. To facilitate sparse view reconstruction and generalization for object-level data, several methods [42, 48, 75] propose to condition NeRFs on image features which establish the foundation for online NeRFs. Recently, this boundary has been further extended to scene-level online reconstruction by [7, 37, 71]. However, these works still use image features as conditions for NeRFs, ignoring 3D feature encoding crucial for 3D perception, while our method learns 3D representations explicitly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.893, + 0.696 + ], + "angle": 0, + "content": "Self-supervised Depth Prediction: Training of NeRFs is time-consuming and prone to overfitting due to geometry-appearance coupling, which can be alleviated with additional depth supervision from separately trained depth prediction modules [15, 58, 68] or LiDAR data [56]. To integrate explicit depth optimization in self-supervised generalizable NeRFs, BTS [71] and SceneRF [7] leverage the reprojection loss from self-supervised depth estimation literature [22, 23, 76, 85], which minimizes the discrepancy between the target pixel and its warped counterpart. However, the extra degrees of freedom from volume rendering and the local receptive field from bilinear interpolation may still hinder depth optimization, while we propose a novel MVS-embedded approach to learn depth in NeRFs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.692, + 0.726 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In this section, we present a self-supervised vision-based 3D occupancy prediction method for autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.742, + 0.789 + ], + "angle": 0, + "content": "3.1. From Image to Occupancy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To understand the complicated interactions among traffic participants and their surrounding environment, it is fundamental for autonomous driving systems to reconstruct the accurate 3D structure of the scene. 3D occupancy prediction is one prevalent proxy for scene reconstruction given its fine granularity and comprehensiveness, which aims at producing a voxelized prediction \\(\\mathbf{O} \\in \\mathcal{C}^{H \\times W \\times D}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.264 + ], + "angle": 0, + "content": "Figure 2. Our MVS-embedded strategy effectively enlarges the receptive field of the depth optimization process across the whole epipolar line, which provides a good depth prior (Raw Argmin)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.468, + 0.41 + ], + "angle": 0, + "content": "encoding per-voxel occupancy (and semantic) information. Here \\(H, W, D\\) and \\(\\mathcal{C}\\) denote the resolution of the occupancy grid and the set of predefined classes. First proposed by MonoScene [6], vision-based methods have become a promising approach to 3D occupancy prediction because of their low cost and effectiveness compared with 3D-based methods. Vision-based methods generally learn a mapping \\(\\mathcal{M} = \\mathcal{D} \\circ \\mathcal{F} \\circ \\mathcal{E}\\) from RGB images \\(\\mathbf{I} = \\{\\mathbf{I}_n | n = 1, \\dots, N\\}\\) captured by \\(N\\) cameras to 3D occupancy \\(\\mathbf{O}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.422, + 0.469, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} = \\mathcal {M} (\\mathbf {I}) = \\mathcal {D} \\left(\\mathcal {F} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {1}\\right), \\dots , \\mathcal {E} \\left(\\mathbf {I} _ {N}\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.469, + 0.556 + ], + "angle": 0, + "content": "where a 2D backbone \\(\\mathcal{E}\\) first encodes \\(N\\) input images into multi-scale image features \\(\\mathbf{F} = \\mathcal{E}(\\mathbf{I})\\), and a 3D encoder \\(\\mathcal{F}\\) further lifts 2D features into 3D representation \\(\\mathbf{R} = \\mathcal{F}(\\mathbf{F})\\) through attention mechanism [31, 40], lift-splat-shoot (LLS) [39] or identical transformation for monocular scenario [6], followed by a decoder network \\(\\mathcal{D}\\) to generate the final occupancy prediction result \\(\\mathbf{O} = \\mathcal{D}(\\mathbf{R})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.469, + 0.753 + ], + "angle": 0, + "content": "Although recent supervised vision-based approaches have achieved encouraging performance on 3D occupancy prediction given only visual inputs at inference, they still rely on 3D (semantic) supervision during training, e.g. Li-DAR point cloud or dense occupancy annotation. On the contrary, self-supervised counterparts [7, 71] take only temporal correspondences inherent in video sequences as supervision to predict meaningful occupancy, and therefore can effortlessly exploit the vast amount of unlabeled driving images. However, existing self-supervised methods [7, 71] consider only one monocular camera and decode forward occupancy \\(\\mathbf{O}_{\\text{front}}\\) directly from 2D image features by instantiating the 3D encoder \\(\\mathcal{F}\\) with identical transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.766, + 0.469, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {f r o n t} = \\mathcal {D} \\left(\\mathcal {F} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {m o n o}\\right)\\right)\\right) = \\mathcal {D} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {m o n o}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Despite being straightforward, ignoring \\(\\mathcal{F}\\) may cause multicamera inconsistency in the surround view scenario and inadequate geometry reasoning in the 3D space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To this end, we propose to transform the images into the BEV or TPV space to obtain a 3D representation of the scene in order to enable 3D feature interactions even in the monocular setting and also avoid ambiguities from multiple" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "cameras. Similar to the supervised literature [31, 40], we leverage the deformable cross-attention (CA) to adaptively aggregate information from the image features \\(\\mathbf{F}\\), in which a set of learnable 3D tokens \\(\\mathbf{Q}\\) serve as queries and the corresponding local image features serve as keys and values. Each of these learnable tokens represents a pillar area as in BEV or TPV representation. Moreover, we determine the correspondences between 3D tokens and image features with the projection matrices \\(\\mathbf{T} = \\{\\mathbf{T}_n|n = 1,\\dots,N\\}\\) from the ego car to pixel coordinates. We interleave deformable self-attention (SA), deformable cross-attention (CA) and feed forward network (FFN) to build the 3D encoder \\(\\mathcal{F}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.281, + 0.892, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {l} = \\operatorname {F F N} _ {l} \\left(\\mathrm {C A} _ {l} \\left(\\mathrm {S A} _ {l} (\\mathbf {Q} _ {l}); \\mathbf {F}, \\mathbf {T}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.827, + 0.321 + ], + "angle": 0, + "content": "where the subscription \\(l\\) represents the \\(l\\)-th block." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.329, + 0.742, + 0.345 + ], + "angle": 0, + "content": "3.2.From Occupancy to Image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Because 3D ground truth is unavailable, we project occupancy O from Eq. (1) back into 2D views and leverage multi-view consistency inherent in video sequences to optimize a self-supervised vision-based occupancy prediction system. Similar to previous work [7, 71], we employ differentiable volume rendering [47] to synthesize color and depth views, which enables the seamless integration of gradient information into the rendering pipeline and effective exploitation of supervision from multiple viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.655 + ], + "angle": 0, + "content": "We first transform the 3D representation \\(\\mathbf{R}\\) into an SDF field \\(\\mathbf{S} \\in \\mathbb{R}^{H \\times W \\times D}\\), which indicates the distance of every voxel center to the nearest surface of an object, by implementing the decoder network \\(\\mathcal{D}\\) with an MLP. We use different MLPs for different heights for BEV representation, while we construct a 3D feature volume from TPV representation with broadcasting and summation [31] before feeding the feature volume through a single MLP. For a continuous 3D coordinate \\(\\mathbf{p}\\), we employ bicubic interpolation (BI) to predict the SDF value \\(s_{\\mathbf{p}}\\) and determine its occupancy status \\(o_{\\mathbf{p}}\\) according to the sign of \\(s_{\\mathbf{p}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.662, + 0.891, + 0.68 + ], + "angle": 0, + "content": "\\[\no _ {\\mathbf {p}} = \\operatorname {s g n} \\left(s _ {\\mathbf {p}}\\right), \\quad s _ {\\mathbf {p}} = \\operatorname {B I} (\\mathbf {S}, \\mathbf {p}) = \\operatorname {B I} (\\mathcal {D} (\\mathbf {R}), \\mathbf {p}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.838 + ], + "angle": 0, + "content": "The preference for SDF field over density field as in [7, 71] is based on two considerations: 1) An SDF field has more explicit physical meaning compared with a density field and inherently holds assumptions about the gradient magnitudes [24], thus allowing for easier regularization and optimization. 2) The signed nature of SDF facilitates straightforward determination of whether a point lies inside or outside a surface, thus enabling accurate differentiation of intricate geometries. Similar to SDF values, we can also decode color c from \\(\\mathbf{R}\\) by using a separate MLP \\(\\mathcal{D}_c\\) as the decoder:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.846, + 0.891, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\mathbf {p}} = \\operatorname {B I} (\\mathbf {C}, \\mathbf {p}) = \\operatorname {B I} \\left(\\mathcal {D} _ {c} (\\mathbf {R}), \\mathbf {p}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Take the volume rendering process of one single ray as an example. We first determine the origin and direction of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.087, + 0.895, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 3. Framework of the proposed SelfOcc for self-supervised vision-based 3D occupancy prediction. We employ an image backbone and a 3D encoder to generate 3D representations as in BEVFormer [40] or TPVFormer [31]. To render a novel view, we apply a lightweight MLP on the 3D features to predict the SDF values, color and semantic vectors. We then perform volume rendering to synthesize color, depth and semantic views. We use simple 0-thresholding to predict the occupancy volume." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.41 + ], + "angle": 0, + "content": "the ray and uniformly sample \\(M\\) points \\(\\mathbf{P} = \\{\\mathbf{p}_m|m = 1,\\dots,M\\}\\) between the origin and the intersection of the ray with the border of the 3D representation. The SDF values of these points \\(\\mathbf{s} = \\{s_m|m = 1,\\dots,M\\}\\) are then calculated with Eq. (4). According to NeuS [65], the discrete opacity value \\(\\alpha_{m}\\), which is the probability of the ray terminating between \\(\\mathbf{p}_m\\) and \\(\\mathbf{p}_{m + 1}\\), can be derived with" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.422, + 0.47, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {m} = \\max \\left(\\frac {\\Phi_ {a} (s _ {m}) - \\Phi_ {a} (s _ {m + 1})}{\\Phi_ {a} (s _ {m})}, 0\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.471, + 0.609 + ], + "angle": 0, + "content": "where \\(\\Phi_{a}(x) = (1 + e^{-ax})^{-1}\\) is the sigmoid function with \\(a\\) being a learnable parameter. Further, the probability of the ray passing through \\(\\mathbf{p}_m\\), i.e. the discrete accumulated transmittance, can be defined by \\(T_{m} = \\prod_{i = 1}^{m - 1}(1 - \\alpha_{i})\\). Finally, we can calculate the probability of the ray ending at \\(\\mathbf{p}_m\\) after emission as \\(w_{m} = T_{m}\\alpha_{m}\\), and the rendering process is an integration of per-point attribute over the probability distribution \\(\\mathbf{w} = \\{w_m|m = 1,\\dots,M\\}\\). Eq. (7) calculates the rendered color \\(\\mathbf{c}_r\\) and depth \\(d_r\\) of the ray:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.622, + 0.47, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {r} = \\sum_ {m = 1} ^ {M} w _ {m} \\mathbf {c} _ {m}, \\quad d _ {r} = \\sum_ {m = 1} ^ {M} w _ {m} d _ {m}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.678, + 0.468, + 0.708 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_m, d_m\\) denote the color and depth of the \\(m\\)-th point. Figure 3 shows the overall framework of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.371, + 0.74 + ], + "angle": 0, + "content": "3.3. Occupancy-Oriented Supervision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.794 + ], + "angle": 0, + "content": "In this section, we elaborate on the supervision formulation that facilitates self-supervised vision-based 3D occupancy prediction for autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "MVS-embedded Depth Learning. Depth supervision enhances the convergence of neural radiance fields [68], enabling faster and more effective learning of accurate geometry, especially with sparse views. Self-supervised depth estimation in NeRFs usually relies on the photometric reprojection loss \\( L_{rpj} \\) which aims to maximize the similarity between the target image \\( \\mathbf{I}_t \\) and sampled pixels on the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.302, + 0.885, + 0.318 + ], + "angle": 0, + "content": "source image \\(\\mathbf{I}_s\\) by refining the depth prediction \\(\\mathbf{D}(\\mathbf{I}_t,\\pmb {\\theta})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.327, + 0.892, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {r p j} (\\mathbf {x}, \\mathbf {I} _ {t}, \\mathbf {I} _ {s}; \\boldsymbol {\\theta}) = \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {s} (\\hat {\\mathbf {x}}) \\| \\tag {8} \\\\ = \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {s} \\left(\\pi (\\mathbf {x}, \\mathbf {D} (\\mathbf {I} _ {t}; \\boldsymbol {\\theta}) | _ {\\mathbf {x}}, \\Pi)\\right) \\|, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.569 + ], + "angle": 0, + "content": "where \\(\\theta, \\mathbf{x}, \\hat{\\mathbf{x}}, \\Pi\\) denote the learnable parameters, a random pixel on the target image, the warped pixel on the source image and the projection matrix from the target to the source image coordinate, respectively. \\(\\mathbf{I}(\\mathbf{x})\\) and \\(\\mathbf{D}|_{\\mathbf{x}}\\) denote bilinear interpolation of corresponding 2D tensors, and \\(\\pi (\\mathbf{x},d_{\\mathbf{x}},\\Pi)\\) warps a pixel \\(\\mathbf{x}\\) to another image according to the depth \\(d_{\\mathbf{x}}\\) and projection matrix \\(\\Pi\\). The inherent local receptive field of the bilinear interpolation \\(\\mathbf{I}_s(\\hat{\\mathbf{x}})\\) has an adverse effect on the optimization property of \\(L_{rpj}\\) since the differences between \\(\\mathbf{I}_t(\\mathbf{x})\\) and the colors of the four adjacent grid pixels of \\(\\hat{\\mathbf{x}}\\) are the only reference information for the optimization of \\(\\theta\\), which could be easily misled by poor initialization or low-texture regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.893, + 0.676 + ], + "angle": 0, + "content": "Therefore, we propose an MVS-embedded strategy to extend the receptive field of the depth optimization for a ray across the whole epipolar line. Specifically, we first take the uniformly spaced depths \\(\\mathbf{d} = \\{d_m|m = 1,\\dots,M\\}\\) associated with \\(\\mathbf{P} = \\{\\mathbf{p}_m|m = 1,\\dots,M\\}\\) as depth proposals. Then we define \\(L_{mvs}\\) as an average of the dissimilarity at each proposal \\(d_{m}\\) with weight \\(w_{m}\\) from volume rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.684, + 0.892, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {m v s} (\\mathbf {x}, \\mathbf {I} _ {t}, \\mathbf {I} _ {s}; \\boldsymbol {\\theta}) \\\\ = \\sum_ {m = 1} ^ {M} w _ {m} (\\boldsymbol {\\theta}) \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {x} (\\pi (\\mathbf {x}, d _ {m}, \\boldsymbol {\\Pi})) \\|. \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.894, + 0.876 + ], + "angle": 0, + "content": "As Figure 2 shows, \\( L_{mvs} \\) enlarges the receptive field of depth optimization by associating \\( \\theta \\) with multiple depth proposals along the ray, thus stabilizing and facilitating self-supervised depth optimization in NeRFs. Similar to Mon-depth2 [23], we formulate the final depth loss by taking the minimum of two \\( L_{mvs} \\) using the previous image \\( \\mathbf{I}_{t-1} \\) or the next image \\( \\mathbf{I}_{t+1} \\) as the source image and leverage the auto-masking strategy to filter out low texture regions:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.884, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\nL _ {d e p} (\\mathbf {x}, \\mathbf {I} _ {t}) = \\operatorname {a u t o m a s k} \\left(\\min \\left(L _ {m v s} ^ {(t - 1)}, L _ {m v s} ^ {(t + 1)}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19949" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. 3D occupancy prediction performance on the Occ3D-nuScenes [63] dataset. Cons. veh. and drive. surf. represent construction vehicle and driveable surface, respectively. Our method learns meaningful geometry and reasonable semantics compared with 3D-supervised methods, and achieves even higher IoU than LiDAR-supervised TPVFormer [31]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.133, + 0.888, + 0.328 + ], + "angle": 0, + "content": "
MethodSupervisionIoUmIoUothersbarrierbicyclebuscarcons. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetation
MonoScene [6]3D-6.061.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.65
OccFormer [82]3D-21.935.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.97
BEVFormer [40]3D-26.885.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.028.0620.0417.69
CTF-Occ [63]3D-28.538.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.0
TPVFormer [31]3D-27.837.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.78
FB-OCC [41]3D-42.0614.3049.7130.0046.6251.5429.3029.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58
RenderOcc [53]L-23.935.6927.5614.3619.9120.5611.9612.4212.1414.3420.8118.9468.8533.3542.0143.9417.3622.61
TPVFormer [31]L17.2013.570.0014.809.3621.2716.8114.4513.7611.235.3216.0519.7310.759.439.5011.1616.5117.04
SelfOcc (BEV)C44.336.760.000.000.000.009.820.000.000.000.000.006.9747.030.0018.7516.5811.933.81
SelfOcc (TPV)C45.019.300.000.150.665.4612.540.000.802.100.000.008.2555.490.0026.3026.5414.225.60
" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.335, + 0.361, + 0.35 + ], + "angle": 0, + "content": "where we omit the learnable parameters \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.471, + 0.591 + ], + "angle": 0, + "content": "Supervision from temporal frames. Unlike depth estimation which predicts only the visible surfaces, 3D occupancy prediction reconstructs the complete geometry of the scene which requires self-supervision signals from multiple viewpoints as in NeRFs. Considering the limited overlapping field of view for cameras, we take advantage of the temporal frames as supervision. Given a video sequence \\(\\mathbf{V} = \\{\\mathbf{V}_1,\\dots,\\mathbf{V}_T\\}\\) with \\(T\\) frames and the current timesamp \\(t\\), we use \\(\\mathbf{V}_t\\) as the input to our model \\(\\mathcal{M}\\) to predict the SDF and radiance fields \\(\\{\\mathbf{S}_t,\\mathbf{C}_t\\}\\). Then we decide whether to use a temporal frame as supervision with probability \\(p\\). If positive, we randomly sample one frame \\(t^\\prime\\) for supervision satisfying that the distance between the ego cars at timestamps \\(t\\) and \\(t^\\prime\\) falls inside \\([l_1,l_2]\\), which guarantees diversity and adheres to the range of the 3D representation. Otherwise, we use the current frame \\(t\\) for supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.863 + ], + "angle": 0, + "content": "SDF Field Regularization. Neural radiance fields with sparse views suffer from overfitting because of a large solution set [51], which is exacerbated in autonomous driving with outward-facing cameras. Therefore, it is crucial to regularize the SDF field in order to restrict the solution set. To learn a smooth SDF field, we minimize the magnitude of the second-order derivatives of the SDF field with a Hessian loss [79] \\( L_{H}(\\mathbf{p}) = \\| \\mathbf{H}(\\mathbf{p})\\|_{1} \\), where \\( \\| \\cdot \\|_1 \\) is the elementwise matrix 1-norm, and \\( \\mathbf{H}(\\mathbf{p}) \\) denotes the Hessian matrix of the SDF field with respect to the 3D coordinates at location \\( \\mathbf{p} \\). To impose sparsity of occupancy in the invisible area such as those blocked by a wall, we propose a simple yet effective regularization \\( L_{s}(\\mathbf{p}) = \\max(-s_{\\mathbf{p}}, 0) \\), which encourages negative SDF values to become larger. In addition, we include the Eikonal term [24] \\( L_{E}(\\mathbf{p}) = ||\\nabla s(\\mathbf{p})|| - 1| \\) for compliance with physical meaning of SDF. For optional semantic prediction, we use 2D segmentation maps from off-the-shelf segmentors as supervision similar to colors." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.864, + 0.421, + 0.878 + ], + "angle": 0, + "content": "Til now, we define the training loss \\(L\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.886, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nL = L _ {d e p} + \\lambda_ {c} L _ {r g b} + \\lambda_ {e} L _ {E} + \\lambda_ {H} L _ {H} + \\lambda_ {s} L _ {s} \\tag {11}\n\\]" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.333, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Table 2. 3D occupancy prediction performance on the SemanticKITTI [3] dataset. MonoScene* is supervised by depth predictions of Monodepth2 [23] trained with ground-truth poses." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.378, + 0.882, + 0.558 + ], + "angle": 0, + "content": "
MethodSupervisionIoUPrec.Rec.
3DDepth Image
MonoScene [6]✓37.1449.9059.24
\\(LMSCNet^{rgb}\\) [59]✓12.0813.0063.16
\\(3DSketch^{rgb}\\) [10]✓12.0112.9562.31
\\(AICNet^{rgb}\\) [36]✓11.2811.8470.89
MonoScene [6]✓13.5316.9840.06
MonoScene* [6]11.1813.1540.22
SceneRF [7]✓13.8417.2840.96
SelfOcc (BEV)✓20.9537.2732.37
SelfOcc (TPV)✓21.9734.8337.31
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.894, + 0.643 + ], + "angle": 0, + "content": "where the four \\(\\lambda\\)s are hyperparameters and \\(L_{rgb}\\) measures the dissimilarity between the rendered color \\(\\mathbf{c}_r\\) of a ray and the ground truth color on the target image. For semantic prediction, we optionally add a semantic loss \\(L_{sem}\\) similar to \\(L_{rgb}\\) and an off-the-shelf 2D segmentor." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.656, + 0.634, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.673, + 0.697 + ], + "angle": 0, + "content": "4.1. Task Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The 3D occupancy prediction task aims to predict the dense occupancy (and semantic) states for a voxel grid, which is a fundamental task in autonomous driving due to its fine granularity and least abstraction of the real world. We conduct this task on the Occ3D-nuScenes [63] dataset, and use IoU and mIoU as geometric and semantic metrics, respectively. In addition, we evaluate our method on the SemanticKITTI [3] dataset for the geometry-only 3D occupancy prediction task and report IoU, Precision and Recall." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The novel depth synthesis task synthesizes depth maps from novel viewpoints within a certain distance given the input frame [7]. This task is more reliable as a proxy for 3D reconstruction compared with novel view synthesis because" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19950" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 3. Novel depth synthesis on SemanticKITTI [3] and nuScenes [4]. We use the results reported in SceneRF [7] and all methods are trained using only images with ground truth poses. We adapt SceneRF [7] to the nuScenes dataset based on the official implementation." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.12, + 0.868, + 0.292 + ], + "angle": 0, + "content": "
MethodDatasetAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
MonoDepth2 [23]0.52597.11314.431.029210.4426.3241.43
SynSin [70]0.36113.4838.8240.429052.6174.5686.50
PixelNeRF [75]0.23642.0806.4490.335465.8185.4392.90
MINE [37]SemanticKITTI0.22481.7876.3430.328365.8785.5293.30
VisionNerf [42]0.20541.4905.8410.307369.1188.2894.37
SceneRF [7]0.16811.2915.7810.285175.0789.0994.50
SelfOcc (TPV)0.15621.1535.3490.253478.9291.1295.92
SceneRF [7]nuScenes0.763115.3411.500.624040.2664.8677.02
SelfOcc (TPV)0.40036.7238.4600.416564.5080.4088.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.298, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Table 4. Self-supervised depth estimation on the nuScenes [4] and KITTI-2015 [21] dataset. M: trained with monocular sequence. MS: trained with both monocular sequence and stereo images. SV: trained with surround-view images. TO: temporal offline refinement. We train SurroundDepth* with ground truth poses based on the official implementation." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.34, + 0.872, + 0.643 + ], + "angle": 0, + "content": "
MethodDatasetTypeScale MethodAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
Monodepth2 [23]nuScenesSVMedian0.2873.3497.1840.3450.6410.8450.925
FSM [26]SVMedian0.299------
SurroundDepth [69]SVSfM Pretrain0.2804.4017.4670.3640.6610.8440.917
SurroundDepth* [69]SVPose GT0.3427.9178.0190.3600.7160.8580.918
SimpleOcc [19]SVPose GT0.2243.3837.1650.3330.7530.8770.930
SelfOcc (TPV)SVPose GT0.2152.7436.7060.3160.7530.8750.932
R3D3 [60]nuScenesTOExtrinsics0.2534.7597.150-0.729--
Johnston [33]KITTI-2015MMedian0.1060.8614.6990.1850.8890.9620.982
FeatDepth(R50) [61]MMedian0.1040.7294.4810.1790.8930.9650.984
PackNet-SfM [25]MMedian0.1070.8024.5380.1860.8890.9620.981
R-MSFM6 [86]MMedian0.1080.7484.4700.1850.8890.9630.982
DevNet [84]MMedian0.1000.6994.4120.1740.8930.9660.985
SelfOcc (TPV)MPose GT0.1030.7924.6730.1870.8770.9590.982
MonoDepth2(R50) [23]KITTI-2015MSExtrinsics0.1060.8184.750.1960.8740.9570.979
FeatDepth(R50) [61]MSExtrinsics0.0990.6974.4270.1840.8890.9630.982
R-MSFM6 [86]MSExtrinsics0.1080.7534.4690.1850.8880.9630.982
DevNet [84]MSExtrinsics0.0950.6714.3650.1740.8950.9700.988
BTS [71]MSPose GT0.1020.7514.4070.1880.8820.9610.982
SelfOcc (TPV)MSPose GT0.0990.7114.5860.1860.8800.9600.982
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.725 + ], + "angle": 0, + "content": "the rendered color is an integration of predicted radiance at sampled locations, which introduces unnecessary degrees of freedom. We conduct experiments on nuScenes [4] and SemanticKITTI [3] and use Abs Rel, Sq Rel, RMSE, RMSE log and threshold accuracies \\((\\delta 1, \\delta 2, \\delta 3)\\) as metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.819 + ], + "angle": 0, + "content": "The depth estimation task is a traditional task for 3D perception, which predicts the pixel-aligned depth map for a given frame. We conduct surround-view depth estimation on nuScenes [4] and monocular depth estimation on KITTI-2015 [21]. The same metrics are used as in novel depth synthesis. We detail the datasets in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.295, + 0.848 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We use the identical network architecture for all three tasks, while the temporal supervision and the loss formulation may differ according to their respective objectives. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.757 + ], + "angle": 0, + "content": "adopt ResNet50 [27] pretrained on ImageNet [14] as the 2D backbone and an feature pyramid network (FPN) [43] to generate multi-scale image features. The 3D encoder \\(\\mathcal{F}\\) is the same as in BEVFormer [40] or TPVFormer [31] depending on the 3D representation we use. We further employ a two-layer MLP as the decoder network \\(\\mathcal{D}\\) to generate the SDF field, color and semantic logits (optional)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.903 + ], + "angle": 0, + "content": "We activate temporal supervision for 3D occupancy prediction and novel view synthesis which both involve scene reconstruction from multiple viewpoints, while we use only the current frame as supervision for depth estimation since it predicts the geometry only for the input view. As for loss formulation, we use \\( L_{dep} \\), \\( L_{E} \\) and an edge loss \\( L_{edg} \\) [22] for depth estimation since RGB supervision from the input view is meaningless, while \\( L_{H} \\) and \\( L_{s} \\) focus on improving the overall geometry, thus helpless for depth es" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19951" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.131 + ], + "angle": 0, + "content": "Table 5. Architecture ablation on the nuScenes dataset. \\( L_{mvs} \\): the proposed \\( L_{mvs}(\\checkmark) \\) or \\( L_{rpj} \\) for depth optimization. SDF: SDF field (\\( \\checkmark \\)) or density field for volume rendering." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.133, + 0.457, + 0.21 + ], + "angle": 0, + "content": "
LmvsSDFOcc.Novel DepthDepth
IoUmIoUAbs Rel↓RMSE↓Abs Rel↓RMSE↓
✓18.902.810.810112.4080.65513.242
✓36.686.100.40078.1490.3188.524
✓✓38.436.940.43768.7340.3127.544
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.223, + 0.468, + 0.249 + ], + "angle": 0, + "content": "Table 6. Temporal supervision ablation on nuScenes [4]. We do not use semantic supervision for occupancy prediction here." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.251, + 0.468, + 0.351 + ], + "angle": 0, + "content": "
SupervisionRatioOcc.NovelDepthDepth
Curr.Prev.NextIoUPrec.Rec.Abs RelRMSEAbs RelRMSE
10039.6766.9649.330.4718.8520.3127.544
21140.5067.5150.310.4468.8270.3658.046
11140.4763.9852.410.4388.7340.3637.975
12240.4363.3552.780.4428.7810.3667.998
01142.4870.0851.890.4628.7320.3928.075
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.367, + 0.468, + 0.407 + ], + "angle": 0, + "content": "Table 7. Loss and regularization ablation on the SemanticKITTI [3] (occupancy prediction and novel depth synthesis) and KITTI-2015 [21] (depth estimation) datasets." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.409, + 0.468, + 0.509 + ], + "angle": 0, + "content": "
LdepLrgbLHLsOcc.Novel DepthDepth
IoUPrec.Rec.Abs RelRMSEAbs RelRMSE
✓9.819.8596.180.385211.67--
✓23.5141.0135.510.15795.3480.1024.676
✓✓24.0737.7839.880.15735.3520.1004.773
✓✓✓12.2713.0268.140.17225.7480.1075.091
✓✓✓✓22.1633.5439.500.16945.6010.1135.078
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.516, + 0.468, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.672, + 0.468, + 0.701 + ], + "angle": 0, + "content": "Figure 4. Visualization of the effect of different loss functions. The occupancy quality gradually improves with \\( L_{dep} \\), \\( L_{H} \\) and \\( L_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "timation. For novel depth synthesis, we leverage \\(L_{dep}\\), \\(L_{rgb}\\) and \\(L_{E}\\). And for 3D occupancy prediction, we further activate \\(L_{H}\\) and \\(L_{s}\\) on the basis of novel depth synthesis to impose smoothness and sparsity constraints on the predicted occupancy grid. For semantic prediction, we use the off-the-shelf open-vocabulary segmentation network OpenSeeD [77] to generate pseudo segmentation labels for nuScenes [4]. We use the AdamW [45] optimizer with an initial learning rate of 1e-4 which decays to zero following the cosine schedule. We train our models for 12 epochs on the nuScenes [4] dataset and 24 epochs on the SemanticKITTI [3] and KITTI-2015 [21] datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.695, + 0.107 + ], + "angle": 0, + "content": "4.3. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.319 + ], + "angle": 0, + "content": "3D Occupancy Prediction. We report the results for the 3D occupancy prediction task in Table 1 and 2. On the Occ3D-nuScenes [63] dataset, our methods achieve comparable IoU and reasonable mIoU without any form of 3D supervision compared with supervised approaches. The IoU metric \\((45.01\\%)\\) indicates that SelfOcc learns meaningful geometry given only video sequences. We notice that SelfOcc is better at predicting background classes than foreground classes, which might be related to the semantic misalignment of the open-vocabulary 2D segmentation network we use. Nonetheless, our method still outperforms MonoScene [6] supervised by ground truth in mIoU and LiDAR-supervised TPVFormer [31] in IoU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.478 + ], + "angle": 0, + "content": "On the SemanticKITTI [3] dataset, we use the results reported in SceneRF [7], where baselines are grouped into 3D, depth and image supervised categories, and the superscription \\( {}^{rgb} \\) indicates the adaptation from 3D input to image input. From Table 2, SelfOcc sets the new state-of-the-art for vision-based depth-supervised and self-supervised 3D occupancy prediction tasks, which outperforms SceneRF [7] by \\( {58.7}\\% \\) in IoU,with much higher Precision and slightly lower Recall. Visualizations for 3D occupancy predictions are shown in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Novel Depth Synthesis. We report the results for novel depth synthesis in Table 3, which averages depth metrics over all frames no further than 10 meters away from the input frame. Following SceneRF [7], we compare against methods based on monocular depth estimation [23], generalizable NeRF [37, 42, 75] and 3D-aware GAN [70]. Our method outperforms previous state-of-the-art SceneRF [7] on all metrics on the SemanticKITTI [3] dataset. In addition, we modify SceneRF for the nuScenes [4] dataset by randomly selecting one camera every iteration during training, and directly report the depth metric for the input frame considering the complexity of volume rendering conditioned on image features from surround cameras. SelfOcc still achieves better results than SceneRF, which confirms its superiority in 3D reconstruction from surround views. We provide novel depth visualizations in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Depth Estimation. We report the results for depth estimation in Table 4. Our method achieves state-of-the-art performance on nuScenes for self-supervised surround-view depth estimation. We also train SurroundDepth [69] using the official implementation with ground-truth poses to eliminate their influence, which indicates that ground-truth poses do not necessarily boost performance compared with median-scaling applied for scale-ambiguous methods. Additionally, SelfOcc performs on par with the state-of-the-art for self-supervised depth estimation on KITTI-2015 [21]. Visualizations are provided in the supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19952" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.317, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.089, + 0.457, + 0.097 + ], + "angle": 0, + "content": "Semantic Occupancy Prediction" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.098, + 0.463, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.089, + 0.585, + 0.097 + ], + "angle": 0, + "content": "Semantic Occupancy GT" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.1, + 0.605, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.089, + 0.747, + 0.098 + ], + "angle": 0, + "content": "Geometric Occupancy Prediction" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.1, + 0.746, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.089, + 0.873, + 0.097 + ], + "angle": 0, + "content": "Geometric Occupancy GT" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.098, + 0.891, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.171, + 0.317, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.172, + 0.461, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.172, + 0.604, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.173, + 0.746, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.172, + 0.891, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.245, + 0.317, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.245, + 0.461, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.245, + 0.604, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.246, + 0.746, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.245, + 0.891, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.319, + 0.317, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.321, + 0.461, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.321, + 0.604, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.321, + 0.746, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.321, + 0.891, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.391, + 0.15, + 0.4 + ], + "angle": 0, + "content": "BACK LEFT" + }, + { + "type": "image_footnote", + "bbox": [ + 0.184, + 0.391, + 0.214, + 0.4 + ], + "angle": 0, + "content": "BACK" + }, + { + "type": "image_footnote", + "bbox": [ + 0.245, + 0.391, + 0.308, + 0.4 + ], + "angle": 0, + "content": "BACK RIGHT" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.405, + 0.111, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.406, + 0.141, + 0.415 + ], + "angle": 0, + "content": "Veabl s" + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.406, + 0.174, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.406, + 0.192, + 0.415 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.406, + 0.242, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.406, + 0.261, + 0.415 + ], + "angle": 0, + "content": "ous" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.406, + 0.332, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.406, + 0.359, + 0.415 + ], + "angle": 0, + "content": "terrain" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.406, + 0.426, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.406, + 0.454, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.464, + 0.406, + 0.491, + 0.415 + ], + "angle": 0, + "content": "dewalk" + }, + { + "type": "image_footnote", + "bbox": [ + 0.507, + 0.406, + 0.558, + 0.415 + ], + "angle": 0, + "content": "other flat" + }, + { + "type": "image_footnote", + "bbox": [ + 0.571, + 0.406, + 0.604, + 0.415 + ], + "angle": 0, + "content": "pedc" + }, + { + "type": "list", + "bbox": [ + 0.507, + 0.406, + 0.604, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.627, + 0.406, + 0.656, + 0.415 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_footnote", + "bbox": [ + 0.655, + 0.406, + 0.682, + 0.415 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "list", + "bbox": [ + 0.627, + 0.406, + 0.682, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.655, + 0.421, + 0.68, + 0.43 + ], + "angle": 0, + "content": "lower" + }, + { + "type": "list", + "bbox": [ + 0.655, + 0.406, + 0.68, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.406, + 0.658, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.421, + 0.68, + 0.43 + ], + "angle": 0, + "content": "lower" + }, + { + "type": "image_footnote", + "bbox": [ + 0.853, + 0.421, + 0.882, + 0.431 + ], + "angle": 0, + "content": "higher" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.437, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Figure 5. Visualization results for 3D occupancy prediction on the Occ3D-nuScenes [63] dataset. Our method achieves comparable visualization quality with ground truth for both semantic and geometric occupancy prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.23, + 0.489 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.469, + 0.56 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies on architectural choices, temporal supervision, and loss formulations to validate the effectiveness of every component of our method. All experiments are trained for half of the original epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.684 + ], + "angle": 0, + "content": "Architecture. We investigate the effect of \\( L_{mvs} \\) and the SDF field in Table 5. MVS-embedded depth optimization brings consistent improvement for all three tasks. While the SDF field outperforms the density field in occupancy prediction and depth estimation, it achieves worse results in novel depth synthesis. We think it is because the SDF field holds an inherent smoothness prior and is thus more difficult to optimize for novel depth synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.777 + ], + "angle": 0, + "content": "Temporal Supervision. We report the results of different temporal supervision settings in Table 6. It is shown that increasing temporal supervision benefits occupancy prediction, but deteriorates depth estimation performance. In addition, the best score for novel depth synthesis is achieved with a moderate level of temporal supervision (1:1:1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Loss Formulations. We ablate the effect of loss functions in Table 7. Note that we replace \\( L_{rbg} \\) with \\( L_{edg} \\) for depth estimation task. Models trained with depth loss \\( L_{dep} \\) and color loss \\( L_{rgb} \\) achieve best performance for depth-related tasks. As shown in Figure 4, we use the Hessian loss \\( L_{H} \\) and sparsity loss \\( L_{s} \\) to regularize the SDF field for better visualization quality, although models trained without them can also achieve high IoU during evaluation. In fact," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.52 + ], + "angle": 0, + "content": "the sparsity loss makes the prediction more similar to the ground truth, but it may not improve the IoU under the evaluation protocol of using camera masks as in Occ3D [63]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.755, + 0.544 + ], + "angle": 0, + "content": "4.5. 3D Occupancy Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.613 + ], + "angle": 0, + "content": "We provide the visualization results for 3D occupancy prediction on nuScenes [4] in Figure 5. Despite being trained with video sequences only, our model produces meaningful geometric and reasonable semantic predictions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.626, + 0.62, + 0.641 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.817 + ], + "angle": 0, + "content": "In this paper, we have presented a self-supervised vision-based 3D occupancy prediction method (SelfOcc) for autonomous driving, which predicts meaningful geometry and semantics from video sequences only. To facilitate feature extraction in the 3D space, we have proposed to condition NeRFs on 3D representations. Furthermore, we have designed an MVS-embedded strategy to boost depth optimization for NeRFs. SelfOcc has set new state-of-the-art for self-supervised vision-based occupancy prediction on SemanticKITTI and has predicted reasonable 3D occupancy for surrounding cameras on Occ3D for the first time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.661, + 0.847 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62321005, Grant 62336004, and Grant 62125603." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19953" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.355 + ], + "angle": 0, + "content": "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.541 + ], + "angle": 0, + "content": "[7] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.47, + 0.641 + ], + "angle": 0, + "content": "[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[11] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In ECCV, pages 680-697. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[12] Yueh-Tung Chen, Martin Garbade, and Juergen Gall. 3d semantic scene completion from a single depth image using adversarial training. In ICIP, pages 1835-1839. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[13] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[16] Aloisio Dourado, Hansung Kim, Teofilo E de Campos, and Adrian Hilton. Semantic scene completion from a single 360-degree image and depth map. In VISIGRAPP (5: VIS-APP), pages 36-46, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[17] Aloisio Dourado, Teofilo E De Campos, Hansung Kim, and Adrian Hilton. Edgenet: Semantic scene completion from a single rgb-d image. In ICPR, pages 503-510. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501-5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[19] Wanshui Gan, Ningkai Mo, Hongbin Xu, and Naoto Yokoya. A simple attempt for 3d occupancy estimation in autonomous driving. arXiv preprint arXiv:2303.10076, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.396 + ], + "angle": 0, + "content": "[20] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPR Workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[22] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[23] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[25] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, pages 2485-2494, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[26] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. arXiv preprint arXiv:2104.00152, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[28] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19954" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[30] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[31] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[33] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[34] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[35] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters, 5(1):219-226, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[36] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[37] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[38] Siqi Li, Changqing Zou, Yipeng Li, Xibin Zhao, and Yue Gao. Attention-based multi-modal fusion network for semantic scene completion. In AAAI, pages 11402-11409, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[39] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[40] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[41] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[42] Kai-En Lin, Yen-Chen Lin, Wei-Sheng Lai, Tsung-Yi Lin, Yichang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[44] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. NIPS, 31, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[46] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[48] Norman Müller, Andrea Simonelli, Lorenzo Porzi, Samuel Rota Bulo, Matthias Nießner, and Peter Kontschieder. Autorf: Learning 3d object radiance fields from single view observations. In CVPR, pages 3971-3980, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ToG, 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[50] Mong H Ng, Kaahan Radia, Jianfei Chen, Dequan Wang, Ionel Gog, and Joseph E Gonzalez. Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. arXiv preprint arXiv:2006.11436, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[51] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[52] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[53] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[54] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[55] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, pages 6134-6144, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[56] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[57] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[58] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[59] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In ThreeDV, 2020. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19955" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[60] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, pages 3216-3226, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[61] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.244 + ], + "angle": 0, + "content": "[62] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.246, + 0.469, + 0.301 + ], + "angle": 0, + "content": "[63] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.303, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[64] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.414, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[66] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.484, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[67] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In ICCV, pages 8608–8617, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[68] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.609, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[69] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, pages 539-549. PMLR, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[70] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.721, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[71] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9076-9086, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.79, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[72] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[73] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[74] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[76] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[77] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In ICCV, pages 1020-1031, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[78] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, pages 733-749, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[79] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[80] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In ICCV, pages 7801-7810, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[81] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[82] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[83] Min Zhong and Gang Zeng. Semantic point completion network for 3d semantic scene completion. In ECAI 2020, pages 2824-2831. IOS Press, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[84] Kaichen Zhou, Lanqing Hong, Changhao Chen, Hang Xu, Chaoqiang Ye, Qingyong Hu, and Zhenguo Li. Devnet: Self-supervised monocular depth learning via density volume construction. In ECCV, pages 125-142. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[85] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[86] Zhongkai Zhou, Xinnan Fan, Pengfei Shi, and Yuanxue Xin. R-MSFM: Recurrent multi-scale feature modulation for monocular depth estimating. In ICCV, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19956" + } + ] +] \ No newline at end of file diff --git a/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_origin.pdf b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8949527a15afede1fc4243069edd785ecc2d4c41 --- /dev/null +++ b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/322bb044-d09b-4af1-b281-2137f8b01f03_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78fd2f1a9444b615f2ce95c46b8cb5bc04d3309dee64a7e5eebe43bee3fcf2e3 +size 5885600 diff --git a/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/full.md b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eef2f682cca40c570c888859ff2e6385ddfb8d33 --- /dev/null +++ b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/full.md @@ -0,0 +1,403 @@ +# SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction + +Yuanhui Huang* Wenzhao Zheng* Borui Zhang Jie Zhou Jiwen Lu† +Beijing National Research Center for Information Science and Technology, China +Department of Automation, Tsinghua University, China + +{huangyh22,zhang-br21}@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com; {jzhou,lujiwen}@tsinghua.edu.cn + +![](images/dd6ae48a60c57315706bb11fa76122d06eda1f364d4ff290b4ded201403225f2.jpg) +Figure 1. Trained with only video sequences as supervision, our model can predict meaningful geometry for the scene given surround-camera RGB images, which can be further extended to semantic occupancy prediction if 2D segmentation maps are available e.g. from an off-the-shelf segmentor. This task is challenging because it completely depends on video sequences to reconstruct scenes without any 3D supervision. We observe that our model can produce dense and consistent occupancy prediction and even infer the back side of cars. + +# Abstract + +3D occupancy prediction is an important task for the robustness of vision-centric autonomous driving, which aims to predict whether each point is occupied in the surrounding 3D space. Existing methods usually require 3D occupancy labels to produce meaningful results. However, it is very laborious to annotate the occupancy status of each voxel. In this paper, we propose SelfOcc to explore a self-supervised way to learn 3D occupancy using only video sequences. We first transform the images into the 3D space (e.g., bird's eye view) to obtain 3D representation of the scene. We directly impose constraints on the 3D representations by treating them as signed distance fields. We can then render 2D images of previous and future frames as self-supervision signals to learn the 3D representations. We propose an MVS-embedded strategy to directly optimize the SDF-induced weights with multiple depth proposals. Our SelfOcc outperforms the previous best method SceneRF by $58.7\%$ using a single frame as input on SemanticKITTI and is the first self-supervised work that produces reasonable 3D occupancy for surround cameras on nuScenes. SelfOcc produces high-quality depth and achieves state-of-the-art results on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI, KITTI-2015, and nuScenes, respectively. Code: https://github.com/huang-yh/SelfOcc. + +# 1. Introduction + +The vision-centric paradigm has demonstrated its potential for efficient autonomous driving [29, 31, 32, 40], where the key procedure is to obtain a perceptual 3D representation, e.g. the bird's eye view (BEV) [50] or tri-perspective view (TPV) [31] of the surrounding scene given only 2D images [40, 54]. Most existing methods achieve this using 3D annotations as supervision and achieve promising results on various tasks including semantic map construction [28, 50, 54], 3D object detection [30, 39, 40, 81], and 3D semantic occupancy prediction [6, 31, 63, 66]. + +Despite the strong performance of supervised 3D representation that facilitates accurate subsequent prediction and planning [29, 32, 64], it is very laborious to annotate 3D labels and thus difficult to scale to large-scale training data. As one workaround, data auto-labeling creates data cycles and train large networks to assist human annotation, yet it is still time-consuming and resource-hungry [11, 55, 74]. It is also noticed to be the bottleneck towards building and updating large models for autonomous driving [9]. Therefore, it is desired to learn meaningful 3D representations only from recorded video sequences without 3D labels. While existing works have explored learning 3D information (e.g., depth) from images [7, 23, 71], they are still based on 2D representations in each monocular camera space. It remains challenging to obtain reasonable and comprehensive 3D representations useful for further autonomous driving prediction and planning [29] in a self-supervised manner. + +In this paper, we explore the problem of learning self + +supervised 3D representation (BEV and TPV) and adopt 3D occupancy prediction as the main task to evaluate the quality of the learned 3D representation, which aims to classify each point in the 3D space into occupied or empty. We first lift 2D image features into 3D space with deformable attention layers [31, 40] to enable feature interactions in the 3D space and avoid ambiguities from multiple cameras. Inspired by recent advances in neural implicit surface reconstruction [52, 65], we then transform the 3D representation into a signed distance function (SDF) field to allow for meaningful regularization and straightforward determination of occupancy boundaries. When exploiting the temporal consistency inherent in video sequences to optimize the SDF field, we propose an MVS-embedded strategy that directly optimizes the SDF-induced weight values with multiple depth proposals. Combined with the temporal supervision scheme and loss formulation tailored for 3D occupancy prediction, our method can reconstruct meaningful 3D occupancy in a self-supervised manner. + +We conduct extensive experiments on various datasets to demonstrate the effectiveness of the proposed SelfOcc. We apply SelfOcc to two types of 3D representations BEV and TPV. For surround-view 3D occupancy prediction, SelfOcc is the first self-supervised method that is able to produce reasonable occupancy results using only video supervision (with an IoU of 45.01 and mIoU of 9.30 on Occ3D [63]). For monocular 3D occupancy prediction, SelfOcc outperforms the previous best method SceneRF [7] by $58.7\%$ with an IoU of 21.97 over 13.84 on SemanticKITTI [3]. To further evaluate the usefulness of the learned 3D representations, we render depth maps for each camera from the self-supervised BEV and TPV representation. Our SelfOcc outperforms the mainstream method SceneRF [7], MonoDepth2 [23], and SurroundDepth [69] on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI [3], KITTI-2015 [21], and nuScenes [4], respectively. + +# 2. Related Work + +3D Occupancy Prediction: While detection, tracking, prediction and planning have long been the focal tasks and composed a standard pipeline for autonomous driving, 3D occupancy prediction [31] or semantic scene completion [62] has recently gained attention as a more fundamental task encompassing the most fine-grained prediction of both occupancy and semantics of the environment. Pioneering works rely on 3D inputs such as depth [13, 20, 34-36, 38, 44], occupancy grids [20, 59, 72, 73], point cloud [57, 83], or truncated signed distance function (TSDF) [5, 10, 12, 13, 16, 17, 35, 62, 67, 78, 80]. Cao et al. [6] and Huang et al. [31] first attempt to predict semantic occupancy from image input only. Recent advancements related to 3D occupancy prediction include modal + +ity fusion [66], multi-task learning [64], and end-to-end autonomous driving [29]. Despite the promising performance, all these methods require 3D ground truth from laborious annotation for supervision. Most related to our work is BTS [71] and SceneRF [7], which learn 3D occupancy in a self-supervised manner in monocular scenarios. However, it is nontrivial to adapt them to surround views, while our work can inherently handle monocular and surround cases. + +Neural Radiance Fields: One of the most predominant paradigms for self-supervised 3D reconstruction is neural radiance fields [47], which learn a multi-layer perceptron (MLP) per scene to map spatial locations to radiance and density values. To synthesize a novel view, volume rendering [46] is often leveraged which queries the MLP at multiple locations along rays and integrates radiance over density-induced weights. Training of NeRFs [1, 2, 8, 18, 47, 49] depends on a large number of images of the same scene with ground truth poses. To facilitate sparse view reconstruction and generalization for object-level data, several methods [42, 48, 75] propose to condition NeRFs on image features which establish the foundation for online NeRFs. Recently, this boundary has been further extended to scene-level online reconstruction by [7, 37, 71]. However, these works still use image features as conditions for NeRFs, ignoring 3D feature encoding crucial for 3D perception, while our method learns 3D representations explicitly. + +Self-supervised Depth Prediction: Training of NeRFs is time-consuming and prone to overfitting due to geometry-appearance coupling, which can be alleviated with additional depth supervision from separately trained depth prediction modules [15, 58, 68] or LiDAR data [56]. To integrate explicit depth optimization in self-supervised generalizable NeRFs, BTS [71] and SceneRF [7] leverage the reprojection loss from self-supervised depth estimation literature [22, 23, 76, 85], which minimizes the discrepancy between the target pixel and its warped counterpart. However, the extra degrees of freedom from volume rendering and the local receptive field from bilinear interpolation may still hinder depth optimization, while we propose a novel MVS-embedded approach to learn depth in NeRFs. + +# 3. Proposed Approach + +In this section, we present a self-supervised vision-based 3D occupancy prediction method for autonomous driving. + +# 3.1. From Image to Occupancy + +To understand the complicated interactions among traffic participants and their surrounding environment, it is fundamental for autonomous driving systems to reconstruct the accurate 3D structure of the scene. 3D occupancy prediction is one prevalent proxy for scene reconstruction given its fine granularity and comprehensiveness, which aims at producing a voxelized prediction $\mathbf{O} \in \mathcal{C}^{H \times W \times D}$ + +![](images/8b853f5cd0d0d93dda2343e2340c2191a59949fdd685706de62f7d4cf4c46844.jpg) +Figure 2. Our MVS-embedded strategy effectively enlarges the receptive field of the depth optimization process across the whole epipolar line, which provides a good depth prior (Raw Argmin). + +encoding per-voxel occupancy (and semantic) information. Here $H, W, D$ and $\mathcal{C}$ denote the resolution of the occupancy grid and the set of predefined classes. First proposed by MonoScene [6], vision-based methods have become a promising approach to 3D occupancy prediction because of their low cost and effectiveness compared with 3D-based methods. Vision-based methods generally learn a mapping $\mathcal{M} = \mathcal{D} \circ \mathcal{F} \circ \mathcal{E}$ from RGB images $\mathbf{I} = \{\mathbf{I}_n | n = 1, \dots, N\}$ captured by $N$ cameras to 3D occupancy $\mathbf{O}$ : + +$$ +\mathbf {O} = \mathcal {M} (\mathbf {I}) = \mathcal {D} \left(\mathcal {F} \left(\mathcal {E} \left(\mathbf {I} _ {1}\right), \dots , \mathcal {E} \left(\mathbf {I} _ {N}\right)\right)\right), \tag {1} +$$ + +where a 2D backbone $\mathcal{E}$ first encodes $N$ input images into multi-scale image features $\mathbf{F} = \mathcal{E}(\mathbf{I})$ , and a 3D encoder $\mathcal{F}$ further lifts 2D features into 3D representation $\mathbf{R} = \mathcal{F}(\mathbf{F})$ through attention mechanism [31, 40], lift-splat-shoot (LLS) [39] or identical transformation for monocular scenario [6], followed by a decoder network $\mathcal{D}$ to generate the final occupancy prediction result $\mathbf{O} = \mathcal{D}(\mathbf{R})$ . + +Although recent supervised vision-based approaches have achieved encouraging performance on 3D occupancy prediction given only visual inputs at inference, they still rely on 3D (semantic) supervision during training, e.g. Li-DAR point cloud or dense occupancy annotation. On the contrary, self-supervised counterparts [7, 71] take only temporal correspondences inherent in video sequences as supervision to predict meaningful occupancy, and therefore can effortlessly exploit the vast amount of unlabeled driving images. However, existing self-supervised methods [7, 71] consider only one monocular camera and decode forward occupancy $\mathbf{O}_{\text{front}}$ directly from 2D image features by instantiating the 3D encoder $\mathcal{F}$ with identical transformation: + +$$ +\mathbf {O} _ {f r o n t} = \mathcal {D} \left(\mathcal {F} \left(\mathcal {E} \left(\mathbf {I} _ {m o n o}\right)\right)\right) = \mathcal {D} \left(\mathcal {E} \left(\mathbf {I} _ {m o n o}\right)\right). \tag {2} +$$ + +Despite being straightforward, ignoring $\mathcal{F}$ may cause multicamera inconsistency in the surround view scenario and inadequate geometry reasoning in the 3D space. + +To this end, we propose to transform the images into the BEV or TPV space to obtain a 3D representation of the scene in order to enable 3D feature interactions even in the monocular setting and also avoid ambiguities from multiple + +cameras. Similar to the supervised literature [31, 40], we leverage the deformable cross-attention (CA) to adaptively aggregate information from the image features $\mathbf{F}$ , in which a set of learnable 3D tokens $\mathbf{Q}$ serve as queries and the corresponding local image features serve as keys and values. Each of these learnable tokens represents a pillar area as in BEV or TPV representation. Moreover, we determine the correspondences between 3D tokens and image features with the projection matrices $\mathbf{T} = \{\mathbf{T}_n|n = 1,\dots,N\}$ from the ego car to pixel coordinates. We interleave deformable self-attention (SA), deformable cross-attention (CA) and feed forward network (FFN) to build the 3D encoder $\mathcal{F}$ . + +$$ +\mathcal {F} _ {l} = \operatorname {F F N} _ {l} \left(\mathrm {C A} _ {l} \left(\mathrm {S A} _ {l} (\mathbf {Q} _ {l}); \mathbf {F}, \mathbf {T}\right)\right), \tag {3} +$$ + +where the subscription $l$ represents the $l$ -th block. + +# 3.2.From Occupancy to Image + +Because 3D ground truth is unavailable, we project occupancy O from Eq. (1) back into 2D views and leverage multi-view consistency inherent in video sequences to optimize a self-supervised vision-based occupancy prediction system. Similar to previous work [7, 71], we employ differentiable volume rendering [47] to synthesize color and depth views, which enables the seamless integration of gradient information into the rendering pipeline and effective exploitation of supervision from multiple viewpoints. + +We first transform the 3D representation $\mathbf{R}$ into an SDF field $\mathbf{S} \in \mathbb{R}^{H \times W \times D}$ , which indicates the distance of every voxel center to the nearest surface of an object, by implementing the decoder network $\mathcal{D}$ with an MLP. We use different MLPs for different heights for BEV representation, while we construct a 3D feature volume from TPV representation with broadcasting and summation [31] before feeding the feature volume through a single MLP. For a continuous 3D coordinate $\mathbf{p}$ , we employ bicubic interpolation (BI) to predict the SDF value $s_{\mathbf{p}}$ and determine its occupancy status $o_{\mathbf{p}}$ according to the sign of $s_{\mathbf{p}}$ : + +$$ +o _ {\mathbf {p}} = \operatorname {s g n} \left(s _ {\mathbf {p}}\right), \quad s _ {\mathbf {p}} = \operatorname {B I} (\mathbf {S}, \mathbf {p}) = \operatorname {B I} (\mathcal {D} (\mathbf {R}), \mathbf {p}). \tag {4} +$$ + +The preference for SDF field over density field as in [7, 71] is based on two considerations: 1) An SDF field has more explicit physical meaning compared with a density field and inherently holds assumptions about the gradient magnitudes [24], thus allowing for easier regularization and optimization. 2) The signed nature of SDF facilitates straightforward determination of whether a point lies inside or outside a surface, thus enabling accurate differentiation of intricate geometries. Similar to SDF values, we can also decode color c from $\mathbf{R}$ by using a separate MLP $\mathcal{D}_c$ as the decoder: + +$$ +\mathbf {c} _ {\mathbf {p}} = \operatorname {B I} (\mathbf {C}, \mathbf {p}) = \operatorname {B I} \left(\mathcal {D} _ {c} (\mathbf {R}), \mathbf {p}\right). \tag {5} +$$ + +Take the volume rendering process of one single ray as an example. We first determine the origin and direction of + +![](images/15619fe9dc9b165b4de29b2ffa4604395f0c8b45b89360fe362cace141e97870.jpg) +Figure 3. Framework of the proposed SelfOcc for self-supervised vision-based 3D occupancy prediction. We employ an image backbone and a 3D encoder to generate 3D representations as in BEVFormer [40] or TPVFormer [31]. To render a novel view, we apply a lightweight MLP on the 3D features to predict the SDF values, color and semantic vectors. We then perform volume rendering to synthesize color, depth and semantic views. We use simple 0-thresholding to predict the occupancy volume. + +the ray and uniformly sample $M$ points $\mathbf{P} = \{\mathbf{p}_m|m = 1,\dots,M\}$ between the origin and the intersection of the ray with the border of the 3D representation. The SDF values of these points $\mathbf{s} = \{s_m|m = 1,\dots,M\}$ are then calculated with Eq. (4). According to NeuS [65], the discrete opacity value $\alpha_{m}$ , which is the probability of the ray terminating between $\mathbf{p}_m$ and $\mathbf{p}_{m + 1}$ , can be derived with + +$$ +\alpha_ {m} = \max \left(\frac {\Phi_ {a} (s _ {m}) - \Phi_ {a} (s _ {m + 1})}{\Phi_ {a} (s _ {m})}, 0\right), \tag {6} +$$ + +where $\Phi_{a}(x) = (1 + e^{-ax})^{-1}$ is the sigmoid function with $a$ being a learnable parameter. Further, the probability of the ray passing through $\mathbf{p}_m$ , i.e. the discrete accumulated transmittance, can be defined by $T_{m} = \prod_{i = 1}^{m - 1}(1 - \alpha_{i})$ . Finally, we can calculate the probability of the ray ending at $\mathbf{p}_m$ after emission as $w_{m} = T_{m}\alpha_{m}$ , and the rendering process is an integration of per-point attribute over the probability distribution $\mathbf{w} = \{w_m|m = 1,\dots,M\}$ . Eq. (7) calculates the rendered color $\mathbf{c}_r$ and depth $d_r$ of the ray: + +$$ +\mathbf {c} _ {r} = \sum_ {m = 1} ^ {M} w _ {m} \mathbf {c} _ {m}, \quad d _ {r} = \sum_ {m = 1} ^ {M} w _ {m} d _ {m}, \tag {7} +$$ + +where $\mathbf{c}_m, d_m$ denote the color and depth of the $m$ -th point. Figure 3 shows the overall framework of our method. + +# 3.3. Occupancy-Oriented Supervision + +In this section, we elaborate on the supervision formulation that facilitates self-supervised vision-based 3D occupancy prediction for autonomous driving. + +MVS-embedded Depth Learning. Depth supervision enhances the convergence of neural radiance fields [68], enabling faster and more effective learning of accurate geometry, especially with sparse views. Self-supervised depth estimation in NeRFs usually relies on the photometric reprojection loss $L_{rpj}$ which aims to maximize the similarity between the target image $\mathbf{I}_t$ and sampled pixels on the + +source image $\mathbf{I}_s$ by refining the depth prediction $\mathbf{D}(\mathbf{I}_t,\pmb {\theta})$ + +$$ +\begin{array}{l} L _ {r p j} (\mathbf {x}, \mathbf {I} _ {t}, \mathbf {I} _ {s}; \boldsymbol {\theta}) = \| \mathbf {I} _ {t} (\mathbf {x}) - \mathbf {I} _ {s} (\hat {\mathbf {x}}) \| \tag {8} \\ = \| \mathbf {I} _ {t} (\mathbf {x}) - \mathbf {I} _ {s} \left(\pi (\mathbf {x}, \mathbf {D} (\mathbf {I} _ {t}; \boldsymbol {\theta}) | _ {\mathbf {x}}, \Pi)\right) \|, \\ \end{array} +$$ + +where $\theta, \mathbf{x}, \hat{\mathbf{x}}, \Pi$ denote the learnable parameters, a random pixel on the target image, the warped pixel on the source image and the projection matrix from the target to the source image coordinate, respectively. $\mathbf{I}(\mathbf{x})$ and $\mathbf{D}|_{\mathbf{x}}$ denote bilinear interpolation of corresponding 2D tensors, and $\pi (\mathbf{x},d_{\mathbf{x}},\Pi)$ warps a pixel $\mathbf{x}$ to another image according to the depth $d_{\mathbf{x}}$ and projection matrix $\Pi$ . The inherent local receptive field of the bilinear interpolation $\mathbf{I}_s(\hat{\mathbf{x}})$ has an adverse effect on the optimization property of $L_{rpj}$ since the differences between $\mathbf{I}_t(\mathbf{x})$ and the colors of the four adjacent grid pixels of $\hat{\mathbf{x}}$ are the only reference information for the optimization of $\theta$ , which could be easily misled by poor initialization or low-texture regions. + +Therefore, we propose an MVS-embedded strategy to extend the receptive field of the depth optimization for a ray across the whole epipolar line. Specifically, we first take the uniformly spaced depths $\mathbf{d} = \{d_m|m = 1,\dots,M\}$ associated with $\mathbf{P} = \{\mathbf{p}_m|m = 1,\dots,M\}$ as depth proposals. Then we define $L_{mvs}$ as an average of the dissimilarity at each proposal $d_{m}$ with weight $w_{m}$ from volume rendering: + +$$ +\begin{array}{l} L _ {m v s} (\mathbf {x}, \mathbf {I} _ {t}, \mathbf {I} _ {s}; \boldsymbol {\theta}) \\ = \sum_ {m = 1} ^ {M} w _ {m} (\boldsymbol {\theta}) \| \mathbf {I} _ {t} (\mathbf {x}) - \mathbf {I} _ {x} (\pi (\mathbf {x}, d _ {m}, \boldsymbol {\Pi})) \|. \tag {9} \\ \end{array} +$$ + +As Figure 2 shows, $L_{mvs}$ enlarges the receptive field of depth optimization by associating $\theta$ with multiple depth proposals along the ray, thus stabilizing and facilitating self-supervised depth optimization in NeRFs. Similar to Mon-depth2 [23], we formulate the final depth loss by taking the minimum of two $L_{mvs}$ using the previous image $\mathbf{I}_{t-1}$ or the next image $\mathbf{I}_{t+1}$ as the source image and leverage the auto-masking strategy to filter out low texture regions: + +$$ +L _ {d e p} (\mathbf {x}, \mathbf {I} _ {t}) = \operatorname {a u t o m a s k} \left(\min \left(L _ {m v s} ^ {(t - 1)}, L _ {m v s} ^ {(t + 1)}\right)\right), \tag {10} +$$ + +Table 1. 3D occupancy prediction performance on the Occ3D-nuScenes [63] dataset. Cons. veh. and drive. surf. represent construction vehicle and driveable surface, respectively. Our method learns meaningful geometry and reasonable semantics compared with 3D-supervised methods, and achieves even higher IoU than LiDAR-supervised TPVFormer [31]. + +
MethodSupervisionIoUmIoUothersbarrierbicyclebuscarcons. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetation
MonoScene [6]3D-6.061.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.65
OccFormer [82]3D-21.935.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.97
BEVFormer [40]3D-26.885.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.028.0620.0417.69
CTF-Occ [63]3D-28.538.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.0
TPVFormer [31]3D-27.837.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.78
FB-OCC [41]3D-42.0614.3049.7130.0046.6251.5429.3029.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58
RenderOcc [53]L-23.935.6927.5614.3619.9120.5611.9612.4212.1414.3420.8118.9468.8533.3542.0143.9417.3622.61
TPVFormer [31]L17.2013.570.0014.809.3621.2716.8114.4513.7611.235.3216.0519.7310.759.439.5011.1616.5117.04
SelfOcc (BEV)C44.336.760.000.000.000.009.820.000.000.000.000.006.9747.030.0018.7516.5811.933.81
SelfOcc (TPV)C45.019.300.000.150.665.4612.540.000.802.100.000.008.2555.490.0026.3026.5414.225.60
+ +where we omit the learnable parameters $\theta$ . + +Supervision from temporal frames. Unlike depth estimation which predicts only the visible surfaces, 3D occupancy prediction reconstructs the complete geometry of the scene which requires self-supervision signals from multiple viewpoints as in NeRFs. Considering the limited overlapping field of view for cameras, we take advantage of the temporal frames as supervision. Given a video sequence $\mathbf{V} = \{\mathbf{V}_1,\dots,\mathbf{V}_T\}$ with $T$ frames and the current timesamp $t$ , we use $\mathbf{V}_t$ as the input to our model $\mathcal{M}$ to predict the SDF and radiance fields $\{\mathbf{S}_t,\mathbf{C}_t\}$ . Then we decide whether to use a temporal frame as supervision with probability $p$ . If positive, we randomly sample one frame $t^\prime$ for supervision satisfying that the distance between the ego cars at timestamps $t$ and $t^\prime$ falls inside $[l_1,l_2]$ , which guarantees diversity and adheres to the range of the 3D representation. Otherwise, we use the current frame $t$ for supervision. + +SDF Field Regularization. Neural radiance fields with sparse views suffer from overfitting because of a large solution set [51], which is exacerbated in autonomous driving with outward-facing cameras. Therefore, it is crucial to regularize the SDF field in order to restrict the solution set. To learn a smooth SDF field, we minimize the magnitude of the second-order derivatives of the SDF field with a Hessian loss [79] $L_{H}(\mathbf{p}) = \| \mathbf{H}(\mathbf{p})\|_{1}$ , where $\| \cdot \|_1$ is the elementwise matrix 1-norm, and $\mathbf{H}(\mathbf{p})$ denotes the Hessian matrix of the SDF field with respect to the 3D coordinates at location $\mathbf{p}$ . To impose sparsity of occupancy in the invisible area such as those blocked by a wall, we propose a simple yet effective regularization $L_{s}(\mathbf{p}) = \max(-s_{\mathbf{p}}, 0)$ , which encourages negative SDF values to become larger. In addition, we include the Eikonal term [24] $L_{E}(\mathbf{p}) = ||\nabla s(\mathbf{p})|| - 1|$ for compliance with physical meaning of SDF. For optional semantic prediction, we use 2D segmentation maps from off-the-shelf segmentors as supervision similar to colors. + +Til now, we define the training loss $L$ as follows: + +$$ +L = L _ {d e p} + \lambda_ {c} L _ {r g b} + \lambda_ {e} L _ {E} + \lambda_ {H} L _ {H} + \lambda_ {s} L _ {s} \tag {11} +$$ + +Table 2. 3D occupancy prediction performance on the SemanticKITTI [3] dataset. MonoScene* is supervised by depth predictions of Monodepth2 [23] trained with ground-truth poses. + +
MethodSupervisionIoUPrec.Rec.
3DDepth Image
MonoScene [6]✓37.1449.9059.24
\(LMSCNet^{rgb}\) [59]✓12.0813.0063.16
\(3DSketch^{rgb}\) [10]✓12.0112.9562.31
\(AICNet^{rgb}\) [36]✓11.2811.8470.89
MonoScene [6]✓13.5316.9840.06
MonoScene* [6]11.1813.1540.22
SceneRF [7]✓13.8417.2840.96
SelfOcc (BEV)✓20.9537.2732.37
SelfOcc (TPV)✓21.9734.8337.31
+ +where the four $\lambda$ s are hyperparameters and $L_{rgb}$ measures the dissimilarity between the rendered color $\mathbf{c}_r$ of a ray and the ground truth color on the target image. For semantic prediction, we optionally add a semantic loss $L_{sem}$ similar to $L_{rgb}$ and an off-the-shelf 2D segmentor. + +# 4. Experiments + +# 4.1. Task Descriptions + +The 3D occupancy prediction task aims to predict the dense occupancy (and semantic) states for a voxel grid, which is a fundamental task in autonomous driving due to its fine granularity and least abstraction of the real world. We conduct this task on the Occ3D-nuScenes [63] dataset, and use IoU and mIoU as geometric and semantic metrics, respectively. In addition, we evaluate our method on the SemanticKITTI [3] dataset for the geometry-only 3D occupancy prediction task and report IoU, Precision and Recall. + +The novel depth synthesis task synthesizes depth maps from novel viewpoints within a certain distance given the input frame [7]. This task is more reliable as a proxy for 3D reconstruction compared with novel view synthesis because + +Table 3. Novel depth synthesis on SemanticKITTI [3] and nuScenes [4]. We use the results reported in SceneRF [7] and all methods are trained using only images with ground truth poses. We adapt SceneRF [7] to the nuScenes dataset based on the official implementation. + +
MethodDatasetAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
MonoDepth2 [23]0.52597.11314.431.029210.4426.3241.43
SynSin [70]0.36113.4838.8240.429052.6174.5686.50
PixelNeRF [75]0.23642.0806.4490.335465.8185.4392.90
MINE [37]SemanticKITTI0.22481.7876.3430.328365.8785.5293.30
VisionNerf [42]0.20541.4905.8410.307369.1188.2894.37
SceneRF [7]0.16811.2915.7810.285175.0789.0994.50
SelfOcc (TPV)0.15621.1535.3490.253478.9291.1295.92
SceneRF [7]nuScenes0.763115.3411.500.624040.2664.8677.02
SelfOcc (TPV)0.40036.7238.4600.416564.5080.4088.17
+ +Table 4. Self-supervised depth estimation on the nuScenes [4] and KITTI-2015 [21] dataset. M: trained with monocular sequence. MS: trained with both monocular sequence and stereo images. SV: trained with surround-view images. TO: temporal offline refinement. We train SurroundDepth* with ground truth poses based on the official implementation. + +
MethodDatasetTypeScale MethodAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
Monodepth2 [23]nuScenesSVMedian0.2873.3497.1840.3450.6410.8450.925
FSM [26]SVMedian0.299------
SurroundDepth [69]SVSfM Pretrain0.2804.4017.4670.3640.6610.8440.917
SurroundDepth* [69]SVPose GT0.3427.9178.0190.3600.7160.8580.918
SimpleOcc [19]SVPose GT0.2243.3837.1650.3330.7530.8770.930
SelfOcc (TPV)SVPose GT0.2152.7436.7060.3160.7530.8750.932
R3D3 [60]nuScenesTOExtrinsics0.2534.7597.150-0.729--
Johnston [33]KITTI-2015MMedian0.1060.8614.6990.1850.8890.9620.982
FeatDepth(R50) [61]MMedian0.1040.7294.4810.1790.8930.9650.984
PackNet-SfM [25]MMedian0.1070.8024.5380.1860.8890.9620.981
R-MSFM6 [86]MMedian0.1080.7484.4700.1850.8890.9630.982
DevNet [84]MMedian0.1000.6994.4120.1740.8930.9660.985
SelfOcc (TPV)MPose GT0.1030.7924.6730.1870.8770.9590.982
MonoDepth2(R50) [23]KITTI-2015MSExtrinsics0.1060.8184.750.1960.8740.9570.979
FeatDepth(R50) [61]MSExtrinsics0.0990.6974.4270.1840.8890.9630.982
R-MSFM6 [86]MSExtrinsics0.1080.7534.4690.1850.8880.9630.982
DevNet [84]MSExtrinsics0.0950.6714.3650.1740.8950.9700.988
BTS [71]MSPose GT0.1020.7514.4070.1880.8820.9610.982
SelfOcc (TPV)MSPose GT0.0990.7114.5860.1860.8800.9600.982
+ +the rendered color is an integration of predicted radiance at sampled locations, which introduces unnecessary degrees of freedom. We conduct experiments on nuScenes [4] and SemanticKITTI [3] and use Abs Rel, Sq Rel, RMSE, RMSE log and threshold accuracies $(\delta 1, \delta 2, \delta 3)$ as metrics. + +The depth estimation task is a traditional task for 3D perception, which predicts the pixel-aligned depth map for a given frame. We conduct surround-view depth estimation on nuScenes [4] and monocular depth estimation on KITTI-2015 [21]. The same metrics are used as in novel depth synthesis. We detail the datasets in the supplementary. + +# 4.2. Implementation Details + +We use the identical network architecture for all three tasks, while the temporal supervision and the loss formulation may differ according to their respective objectives. We + +adopt ResNet50 [27] pretrained on ImageNet [14] as the 2D backbone and an feature pyramid network (FPN) [43] to generate multi-scale image features. The 3D encoder $\mathcal{F}$ is the same as in BEVFormer [40] or TPVFormer [31] depending on the 3D representation we use. We further employ a two-layer MLP as the decoder network $\mathcal{D}$ to generate the SDF field, color and semantic logits (optional). + +We activate temporal supervision for 3D occupancy prediction and novel view synthesis which both involve scene reconstruction from multiple viewpoints, while we use only the current frame as supervision for depth estimation since it predicts the geometry only for the input view. As for loss formulation, we use $L_{dep}$ , $L_{E}$ and an edge loss $L_{edg}$ [22] for depth estimation since RGB supervision from the input view is meaningless, while $L_{H}$ and $L_{s}$ focus on improving the overall geometry, thus helpless for depth es + +Table 5. Architecture ablation on the nuScenes dataset. $L_{mvs}$ : the proposed $L_{mvs}(\checkmark)$ or $L_{rpj}$ for depth optimization. SDF: SDF field ( $\checkmark$ ) or density field for volume rendering. + +
LmvsSDFOcc.Novel DepthDepth
IoUmIoUAbs Rel↓RMSE↓Abs Rel↓RMSE↓
✓18.902.810.810112.4080.65513.242
✓36.686.100.40078.1490.3188.524
✓✓38.436.940.43768.7340.3127.544
+ +Table 6. Temporal supervision ablation on nuScenes [4]. We do not use semantic supervision for occupancy prediction here. + +
SupervisionRatioOcc.NovelDepthDepth
Curr.Prev.NextIoUPrec.Rec.Abs RelRMSEAbs RelRMSE
10039.6766.9649.330.4718.8520.3127.544
21140.5067.5150.310.4468.8270.3658.046
11140.4763.9852.410.4388.7340.3637.975
12240.4363.3552.780.4428.7810.3667.998
01142.4870.0851.890.4628.7320.3928.075
+ +Table 7. Loss and regularization ablation on the SemanticKITTI [3] (occupancy prediction and novel depth synthesis) and KITTI-2015 [21] (depth estimation) datasets. + +
LdepLrgbLHLsOcc.Novel DepthDepth
IoUPrec.Rec.Abs RelRMSEAbs RelRMSE
✓9.819.8596.180.385211.67--
✓23.5141.0135.510.15795.3480.1024.676
✓✓24.0737.7839.880.15735.3520.1004.773
✓✓✓12.2713.0268.140.17225.7480.1075.091
✓✓✓✓22.1633.5439.500.16945.6010.1135.078
+ +![](images/c2fd7c0769098dfb48e549522f7fd067b121e774e1fd08c819dfd8f03dfc59e3.jpg) +Figure 4. Visualization of the effect of different loss functions. The occupancy quality gradually improves with $L_{dep}$ , $L_{H}$ and $L_{s}$ . + +timation. For novel depth synthesis, we leverage $L_{dep}$ , $L_{rgb}$ and $L_{E}$ . And for 3D occupancy prediction, we further activate $L_{H}$ and $L_{s}$ on the basis of novel depth synthesis to impose smoothness and sparsity constraints on the predicted occupancy grid. For semantic prediction, we use the off-the-shelf open-vocabulary segmentation network OpenSeeD [77] to generate pseudo segmentation labels for nuScenes [4]. We use the AdamW [45] optimizer with an initial learning rate of 1e-4 which decays to zero following the cosine schedule. We train our models for 12 epochs on the nuScenes [4] dataset and 24 epochs on the SemanticKITTI [3] and KITTI-2015 [21] datasets. + +# 4.3. Results and Analysis + +3D Occupancy Prediction. We report the results for the 3D occupancy prediction task in Table 1 and 2. On the Occ3D-nuScenes [63] dataset, our methods achieve comparable IoU and reasonable mIoU without any form of 3D supervision compared with supervised approaches. The IoU metric $(45.01\%)$ indicates that SelfOcc learns meaningful geometry given only video sequences. We notice that SelfOcc is better at predicting background classes than foreground classes, which might be related to the semantic misalignment of the open-vocabulary 2D segmentation network we use. Nonetheless, our method still outperforms MonoScene [6] supervised by ground truth in mIoU and LiDAR-supervised TPVFormer [31] in IoU. + +On the SemanticKITTI [3] dataset, we use the results reported in SceneRF [7], where baselines are grouped into 3D, depth and image supervised categories, and the superscription ${}^{rgb}$ indicates the adaptation from 3D input to image input. From Table 2, SelfOcc sets the new state-of-the-art for vision-based depth-supervised and self-supervised 3D occupancy prediction tasks, which outperforms SceneRF [7] by ${58.7}\%$ in IoU,with much higher Precision and slightly lower Recall. Visualizations for 3D occupancy predictions are shown in Figure 5. + +Novel Depth Synthesis. We report the results for novel depth synthesis in Table 3, which averages depth metrics over all frames no further than 10 meters away from the input frame. Following SceneRF [7], we compare against methods based on monocular depth estimation [23], generalizable NeRF [37, 42, 75] and 3D-aware GAN [70]. Our method outperforms previous state-of-the-art SceneRF [7] on all metrics on the SemanticKITTI [3] dataset. In addition, we modify SceneRF for the nuScenes [4] dataset by randomly selecting one camera every iteration during training, and directly report the depth metric for the input frame considering the complexity of volume rendering conditioned on image features from surround cameras. SelfOcc still achieves better results than SceneRF, which confirms its superiority in 3D reconstruction from surround views. We provide novel depth visualizations in the supplementary. + +Depth Estimation. We report the results for depth estimation in Table 4. Our method achieves state-of-the-art performance on nuScenes for self-supervised surround-view depth estimation. We also train SurroundDepth [69] using the official implementation with ground-truth poses to eliminate their influence, which indicates that ground-truth poses do not necessarily boost performance compared with median-scaling applied for scale-ambiguous methods. Additionally, SelfOcc performs on par with the state-of-the-art for self-supervised depth estimation on KITTI-2015 [21]. Visualizations are provided in the supplementary. + +![](images/ec6a5899ee9c761da65f0dfb932e3b00be50f5d89b50d7fd90da1fe1c50f61c4.jpg) + +![](images/edccc952ae8cd9ba197a15d7b304e802302fa6e30118f12ba70338e710b88496.jpg) +Semantic Occupancy Prediction + +![](images/2042a438150f9db3fc1ee4c8c58314ca60d17ca71b72a0412f52523149151f05.jpg) +Semantic Occupancy GT + +![](images/1d061b9992af1f3b3ba806096e8e503010b3ce011b7fef528dbd2b0daf9fc918.jpg) +Geometric Occupancy Prediction + +![](images/9fb842a46eb7d5a21734dbbfd05f63a8a041607ae85e772ffda6b019f58b5689.jpg) +Geometric Occupancy GT + +![](images/7588c220b99e22bcd38339a533f2b9282a817e78b5d4f4251fb88110bfb42a11.jpg) + +![](images/63cb76b35759bbab4a90d88da8dd0c6a0ff9c71ad4a99fdeb02471c71dd5afb4.jpg) + +![](images/202d1eeb53be0876bc39c50c154dc5c61c1ed1327fb74fa572cd8047fea29588.jpg) + +![](images/8a47d1ca153c83a1aa239bdba9bbb932277c0167efe018071bb485485a2e2bbf.jpg) + +![](images/a69f4ddad804a9011b47b2c09079bf3a3e400b73d78d7998faabe670f82d6e04.jpg) + +![](images/368f2eb8a629d3f1e4357d64b8a6955422164249c9bf3dfae2702b6ba41c7152.jpg) + +![](images/62d9aa8bfae0255fc815d5334757d8b976f23ee64bdc1ac0d87156dbaaede636.jpg) + +![](images/bdaa6aed7018b969a106d843c98abd73c782c39a84704d2086798fd436341dda.jpg) + +![](images/72819640aed9cbc8d5b701b55233029e74201010987a22b6a976325140d84c5b.jpg) + +![](images/8f13c3971692f81024c4886725341eee39f18a07daf0e5fb73c986c861499b99.jpg) + +![](images/0de870ab8bcd2f24bbda702b515c0437caad3ea23e59e5f66372a90e6d61fedb.jpg) +BACK LEFT +BACK +BACK RIGHT + +![](images/ee1ce8b951504b5e64e871396105535f271762991502ecc9c49c227bcfab4b5c.jpg) + +![](images/ec3ba5f0200396d913c9d841988dede7372a9a15320383a00ea360141e403738.jpg) +other flat +pedc + +![](images/e91500f337661ed575abab38f918db33867c79c2d21c78c228aa0e28bc21f0d5.jpg) + +![](images/058acdca8fe2904be7f1efe1ebdd7dbe55ab212b9d284f73ba1309a36bf31d09.jpg) +higher + +![](images/b94dc920af79801cfdc67c72583b81a59122694651e28b8cb8d0bf08bcaae681.jpg) +Veabl s +Figure 5. Visualization results for 3D occupancy prediction on the Occ3D-nuScenes [63] dataset. Our method achieves comparable visualization quality with ground truth for both semantic and geometric occupancy prediction tasks. + +![](images/db925c8d306551a7c10073ee772ed908a67c61baec771db81994f65799ec1054.jpg) +${13}/{14}$ + +![](images/4b52e703fd9a719beebfd36b7ba893aaf7edc9eaf1e1772bdd39899fe8971a94.jpg) +ous + +![](images/7dbb839b407618438fe650a706ec637f0739d8332e852bfb88cfdc81db63ec1d.jpg) +terrain + +![](images/e493d18eeaf104a74429ce83be94d0226542b28b6537ae212e8ba2222d575e27.jpg) + +![](images/7411a4c4205fdbd1355c4dc73eecf4a66b78b744676b5ed7ccada8d40d7c74fc.jpg) +dewalk + +lower +![](images/3fc63cd1b1ef2ace70724c2788b9096be492123d74c4488c756ce762255ce4f5.jpg) +1 +${13}/{14}$ +lower + +# 4.4. Ablation Study + +We conduct comprehensive ablation studies on architectural choices, temporal supervision, and loss formulations to validate the effectiveness of every component of our method. All experiments are trained for half of the original epochs. + +Architecture. We investigate the effect of $L_{mvs}$ and the SDF field in Table 5. MVS-embedded depth optimization brings consistent improvement for all three tasks. While the SDF field outperforms the density field in occupancy prediction and depth estimation, it achieves worse results in novel depth synthesis. We think it is because the SDF field holds an inherent smoothness prior and is thus more difficult to optimize for novel depth synthesis. + +Temporal Supervision. We report the results of different temporal supervision settings in Table 6. It is shown that increasing temporal supervision benefits occupancy prediction, but deteriorates depth estimation performance. In addition, the best score for novel depth synthesis is achieved with a moderate level of temporal supervision (1:1:1). + +Loss Formulations. We ablate the effect of loss functions in Table 7. Note that we replace $L_{rbg}$ with $L_{edg}$ for depth estimation task. Models trained with depth loss $L_{dep}$ and color loss $L_{rgb}$ achieve best performance for depth-related tasks. As shown in Figure 4, we use the Hessian loss $L_{H}$ and sparsity loss $L_{s}$ to regularize the SDF field for better visualization quality, although models trained without them can also achieve high IoU during evaluation. In fact, + +the sparsity loss makes the prediction more similar to the ground truth, but it may not improve the IoU under the evaluation protocol of using camera masks as in Occ3D [63]. + +# 4.5. 3D Occupancy Visualization + +We provide the visualization results for 3D occupancy prediction on nuScenes [4] in Figure 5. Despite being trained with video sequences only, our model produces meaningful geometric and reasonable semantic predictions. + +# 5. Conclusion + +In this paper, we have presented a self-supervised vision-based 3D occupancy prediction method (SelfOcc) for autonomous driving, which predicts meaningful geometry and semantics from video sequences only. To facilitate feature extraction in the 3D space, we have proposed to condition NeRFs on 3D representations. Furthermore, we have designed an MVS-embedded strategy to boost depth optimization for NeRFs. SelfOcc has set new state-of-the-art for self-supervised vision-based occupancy prediction on SemanticKITTI and has predicted reasonable 3D occupancy for surrounding cameras on Occ3D for the first time. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant 62321005, Grant 62336004, and Grant 62125603. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 2 +[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 5, 6, 7 +[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 7, 8 +[5] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2 +[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2, 3, 5, 7 +[7] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 1, 2, 3, 5, 6, 7 +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 2 +[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1 +[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, 2020. 2, 5 +[11] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In ECCV, pages 680-697. Springer, 2022. 1 +[12] Yueh-Tung Chen, Martin Garbade, and Juergen Gall. 3d semantic scene completion from a single depth image using adversarial training. In ICIP, pages 1835-1839. IEEE, 2019. 2 +[13] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161. PMLR, 2021. 2 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6 + +[15] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2 +[16] Aloisio Dourado, Hansung Kim, Teofilo E de Campos, and Adrian Hilton. Semantic scene completion from a single 360-degree image and depth map. In VISIGRAPP (5: VIS-APP), pages 36-46, 2020. 2 +[17] Aloisio Dourado, Teofilo E De Campos, Hansung Kim, and Adrian Hilton. Edgenet: Semantic scene completion from a single rgb-d image. In ICPR, pages 503-510. IEEE, 2021. 2 +[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501-5510, 2022. 2 +[19] Wanshui Gan, Ningkai Mo, Hongbin Xu, and Naoto Yokoya. A simple attempt for 3d occupancy estimation in autonomous driving. arXiv preprint arXiv:2303.10076, 2023. 6 +[20] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPR Workshops, pages 0–0, 2019. 2 +[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 2, 6, 7 +[22] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 2, 6 +[23] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 1, 2, 4, 5, 6, 7 +[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 3, 5 +[25] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, pages 2485-2494, 2020. 6 +[26] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. arXiv preprint arXiv:2104.00152, 2021. 6 +[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 6 +[28] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 1 +[29] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1, 2 + +[30] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1 +[31] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2, 3, 4, 5, 6, 7 +[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1 +[33] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 6 +[34] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2 +[35] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters, 5(1):219-226, 2019. 2 +[36] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, 2020. 2, 5 +[37] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 6, 7 +[38] Siqi Li, Changqing Zou, Yipeng Li, Xibin Zhao, and Yue Gao. Attention-based multi-modal fusion network for semantic scene completion. In AAAI, pages 11402-11409, 2020. 2 +[39] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 3 +[40] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 1, 2, 3, 4, 5, 6 +[41] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 5 +[42] Kai-En Lin, Yen-Chen Lin, Wei-Sheng Lai, Tsung-Yi Lin, Yichang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2, 6, 7 +[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6 +[44] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. NIPS, 31, 2018. 2 + +[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[46] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2 +[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3 +[48] Norman Müller, Andrea Simonelli, Lorenzo Porzi, Samuel Rota Bulo, Matthias Nießner, and Peter Kontschieder. Autorf: Learning 3d object radiance fields from single view observations. In CVPR, pages 3971-3980, 2022. 2 +[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ToG, 41(4):1-15, 2022. 2 +[50] Mong H Ng, Kaahan Radia, Jianfei Chen, Dequan Wang, Ionel Gog, and Joseph E Gonzalez. Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. arXiv preprint arXiv:2006.11436, 2020. 1 +[51] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 5 +[52] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 2 +[53] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 5 +[54] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 1 +[55] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, pages 6134-6144, 2021. 1 +[56] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2 +[57] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2 +[58] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2 +[59] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In ThreeDV, 2020. 2, 5 + +[60] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, pages 3216-3226, 2023. 6 +[61] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 6 +[62] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2 +[63] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 5, 7, 8 +[64] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 1, 2 +[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4 +[66] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 2 +[67] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In ICCV, pages 8608–8617, 2019. 2 +[68] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 2, 4 +[69] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, pages 539-549. PMLR, 2023. 2, 6, 7 +[70] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 6, 7 +[71] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9076-9086, 2023. 1, 2, 3, 6 +[72] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2 +[73] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2 + +[74] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1 +[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2, 6, 7 +[76] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In CVPR, 2018. 2 +[77] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In ICCV, pages 1020-1031, 2023. 7 +[78] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, pages 733-749, 2018. 2 +[79] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 5 +[80] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In ICCV, pages 7801-7810, 2019. 2 +[81] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1 +[82] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 5 +[83] Min Zhong and Gang Zeng. Semantic point completion network for 3d semantic scene completion. In ECAI 2020, pages 2824-2831. IOS Press, 2020. 2 +[84] Kaichen Zhou, Lanqing Hong, Changhao Chen, Hang Xu, Chaoqiang Ye, Qingyong Hu, and Zhenguo Li. Devnet: Self-supervised monocular depth learning via density volume construction. In ECCV, pages 125-142. Springer, 2022. 6 +[85] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017. 2 +[86] Zhongkai Zhou, Xinnan Fan, Pengfei Shi, and Yuanxue Xin. R-MSFM: Recurrent multi-scale feature modulation for monocular depth estimating. In ICCV, 2021. 6 \ No newline at end of file diff --git a/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/images.zip b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..847b803f91d4aa37e2aa87ad1def27a1f3d90761 --- /dev/null +++ b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c211cbf8e8bdfede699d2cbb6c468700bd41f6a96f10c30e276ed7a4c51c7c12 +size 992280 diff --git a/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/layout.json b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e5069e8d5bd613896d774caa2038caf678ac0fd9 --- /dev/null +++ b/2024/SelfOcc_ Self-Supervised Vision-Based 3D Occupancy Prediction/layout.json @@ -0,0 +1,11753 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 55, + 496, + 73 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 55, + 496, + 73 + ], + "spans": [ + { + "bbox": [ + 96, + 55, + 496, + 73 + ], + "type": "text", + "content": "SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 95, + 510, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 95, + 510, + 137 + ], + "spans": [ + { + "bbox": [ + 114, + 95, + 510, + 137 + ], + "type": "text", + "content": "Yuanhui Huang* Wenzhao Zheng* Borui Zhang Jie Zhou Jiwen Lu† \nBeijing National Research Center for Information Science and Technology, China \nDepartment of Automation, Tsinghua University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 137, + 569, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 137, + 569, + 166 + ], + "spans": [ + { + "bbox": [ + 54, + 137, + 569, + 166 + ], + "type": "text", + "content": "{huangyh22,zhang-br21}@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com; {jzhou,lujiwen}@tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 171, + 545, + 303 + ], + "blocks": [ + { + "bbox": [ + 53, + 171, + 545, + 303 + ], + "lines": [ + { + "bbox": [ + 53, + 171, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 545, + 303 + ], + "type": "image", + "image_path": "dd6ae48a60c57315706bb11fa76122d06eda1f364d4ff290b4ded201403225f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 547, + 348 + ], + "type": "text", + "content": "Figure 1. Trained with only video sequences as supervision, our model can predict meaningful geometry for the scene given surround-camera RGB images, which can be further extended to semantic occupancy prediction if 2D segmentation maps are available e.g. from an off-the-shelf segmentor. This task is challenging because it completely depends on video sequences to reconstruct scenes without any 3D supervision. We observe that our model can produce dense and consistent occupancy prediction and even infer the back side of cars." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 357, + 192, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 357, + 192, + 369 + ], + "spans": [ + { + "bbox": [ + 143, + 357, + 192, + 369 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 383, + 290, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 383, + 290, + 683 + ], + "spans": [ + { + "bbox": [ + 45, + 383, + 290, + 683 + ], + "type": "text", + "content": "3D occupancy prediction is an important task for the robustness of vision-centric autonomous driving, which aims to predict whether each point is occupied in the surrounding 3D space. Existing methods usually require 3D occupancy labels to produce meaningful results. However, it is very laborious to annotate the occupancy status of each voxel. In this paper, we propose SelfOcc to explore a self-supervised way to learn 3D occupancy using only video sequences. We first transform the images into the 3D space (e.g., bird's eye view) to obtain 3D representation of the scene. We directly impose constraints on the 3D representations by treating them as signed distance fields. We can then render 2D images of previous and future frames as self-supervision signals to learn the 3D representations. We propose an MVS-embedded strategy to directly optimize the SDF-induced weights with multiple depth proposals. Our SelfOcc outperforms the previous best method SceneRF by " + }, + { + "bbox": [ + 45, + 383, + 290, + 683 + ], + "type": "inline_equation", + "content": "58.7\\%" + }, + { + "bbox": [ + 45, + 383, + 290, + 683 + ], + "type": "text", + "content": " using a single frame as input on SemanticKITTI and is the first self-supervised work that produces reasonable 3D occupancy for surround cameras on nuScenes. SelfOcc produces high-quality depth and achieves state-of-the-art results on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI, KITTI-2015, and nuScenes, respectively. Code: https://github.com/huang-yh/SelfOcc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 357, + 386, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 357, + 386, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 357, + 386, + 369 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "type": "text", + "content": "The vision-centric paradigm has demonstrated its potential for efficient autonomous driving [29, 31, 32, 40], where the key procedure is to obtain a perceptual 3D representation, e.g. the bird's eye view (BEV) [50] or tri-perspective view (TPV) [31] of the surrounding scene given only 2D images [40, 54]. Most existing methods achieve this using 3D annotations as supervision and achieve promising results on various tasks including semantic map construction [28, 50, 54], 3D object detection [30, 39, 40, 81], and 3D semantic occupancy prediction [6, 31, 63, 66]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 700 + ], + "type": "text", + "content": "Despite the strong performance of supervised 3D representation that facilitates accurate subsequent prediction and planning [29, 32, 64], it is very laborious to annotate 3D labels and thus difficult to scale to large-scale training data. As one workaround, data auto-labeling creates data cycles and train large networks to assist human annotation, yet it is still time-consuming and resource-hungry [11, 55, 74]. It is also noticed to be the bottleneck towards building and updating large models for autonomous driving [9]. Therefore, it is desired to learn meaningful 3D representations only from recorded video sequences without 3D labels. While existing works have explored learning 3D information (e.g., depth) from images [7, 23, 71], they are still based on 2D representations in each monocular camera space. It remains challenging to obtain reasonable and comprehensive 3D representations useful for further autonomous driving prediction and planning [29] in a self-supervised manner." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In this paper, we explore the problem of learning self" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 204, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 204, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 204, + 712 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19946" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "supervised 3D representation (BEV and TPV) and adopt 3D occupancy prediction as the main task to evaluate the quality of the learned 3D representation, which aims to classify each point in the 3D space into occupied or empty. We first lift 2D image features into 3D space with deformable attention layers [31, 40] to enable feature interactions in the 3D space and avoid ambiguities from multiple cameras. Inspired by recent advances in neural implicit surface reconstruction [52, 65], we then transform the 3D representation into a signed distance function (SDF) field to allow for meaningful regularization and straightforward determination of occupancy boundaries. When exploiting the temporal consistency inherent in video sequences to optimize the SDF field, we propose an MVS-embedded strategy that directly optimizes the SDF-induced weight values with multiple depth proposals. Combined with the temporal supervision scheme and loss formulation tailored for 3D occupancy prediction, our method can reconstruct meaningful 3D occupancy in a self-supervised manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 300, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 289, + 515 + ], + "type": "text", + "content": "We conduct extensive experiments on various datasets to demonstrate the effectiveness of the proposed SelfOcc. We apply SelfOcc to two types of 3D representations BEV and TPV. For surround-view 3D occupancy prediction, SelfOcc is the first self-supervised method that is able to produce reasonable occupancy results using only video supervision (with an IoU of 45.01 and mIoU of 9.30 on Occ3D [63]). For monocular 3D occupancy prediction, SelfOcc outperforms the previous best method SceneRF [7] by " + }, + { + "bbox": [ + 47, + 300, + 289, + 515 + ], + "type": "inline_equation", + "content": "58.7\\%" + }, + { + "bbox": [ + 47, + 300, + 289, + 515 + ], + "type": "text", + "content": " with an IoU of 21.97 over 13.84 on SemanticKITTI [3]. To further evaluate the usefulness of the learned 3D representations, we render depth maps for each camera from the self-supervised BEV and TPV representation. Our SelfOcc outperforms the mainstream method SceneRF [7], MonoDepth2 [23], and SurroundDepth [69] on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI [3], KITTI-2015 [21], and nuScenes [4], respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 289, + 714 + ], + "type": "text", + "content": "3D Occupancy Prediction: While detection, tracking, prediction and planning have long been the focal tasks and composed a standard pipeline for autonomous driving, 3D occupancy prediction [31] or semantic scene completion [62] has recently gained attention as a more fundamental task encompassing the most fine-grained prediction of both occupancy and semantics of the environment. Pioneering works rely on 3D inputs such as depth [13, 20, 34-36, 38, 44], occupancy grids [20, 59, 72, 73], point cloud [57, 83], or truncated signed distance function (TSDF) [5, 10, 12, 13, 16, 17, 35, 62, 67, 78, 80]. Cao et al. [6] and Huang et al. [31] first attempt to predict semantic occupancy from image input only. Recent advancements related to 3D occupancy prediction include modal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "ity fusion [66], multi-task learning [64], and end-to-end autonomous driving [29]. Despite the promising performance, all these methods require 3D ground truth from laborious annotation for supervision. Most related to our work is BTS [71] and SceneRF [7], which learn 3D occupancy in a self-supervised manner in monocular scenarios. However, it is nontrivial to adapt them to surround views, while our work can inherently handle monocular and surround cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 167, + 546, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 546, + 383 + ], + "type": "text", + "content": "Neural Radiance Fields: One of the most predominant paradigms for self-supervised 3D reconstruction is neural radiance fields [47], which learn a multi-layer perceptron (MLP) per scene to map spatial locations to radiance and density values. To synthesize a novel view, volume rendering [46] is often leveraged which queries the MLP at multiple locations along rays and integrates radiance over density-induced weights. Training of NeRFs [1, 2, 8, 18, 47, 49] depends on a large number of images of the same scene with ground truth poses. To facilitate sparse view reconstruction and generalization for object-level data, several methods [42, 48, 75] propose to condition NeRFs on image features which establish the foundation for online NeRFs. Recently, this boundary has been further extended to scene-level online reconstruction by [7, 37, 71]. However, these works still use image features as conditions for NeRFs, ignoring 3D feature encoding crucial for 3D perception, while our method learns 3D representations explicitly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 384, + 546, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 551 + ], + "type": "text", + "content": "Self-supervised Depth Prediction: Training of NeRFs is time-consuming and prone to overfitting due to geometry-appearance coupling, which can be alleviated with additional depth supervision from separately trained depth prediction modules [15, 58, 68] or LiDAR data [56]. To integrate explicit depth optimization in self-supervised generalizable NeRFs, BTS [71] and SceneRF [7] leverage the reprojection loss from self-supervised depth estimation literature [22, 23, 76, 85], which minimizes the discrepancy between the target pixel and its warped counterpart. However, the extra degrees of freedom from volume rendering and the local receptive field from bilinear interpolation may still hinder depth optimization, while we propose a novel MVS-embedded approach to learn depth in NeRFs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 560, + 423, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 423, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 423, + 574 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "type": "text", + "content": "In this section, we present a self-supervised vision-based 3D occupancy prediction method for autonomous driving." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 611, + 454, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 454, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 454, + 624 + ], + "type": "text", + "content": "3.1. From Image to Occupancy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "To understand the complicated interactions among traffic participants and their surrounding environment, it is fundamental for autonomous driving systems to reconstruct the accurate 3D structure of the scene. 3D occupancy prediction is one prevalent proxy for scene reconstruction given its fine granularity and comprehensiveness, which aims at producing a voxelized prediction " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{O} \\in \\mathcal{C}^{H \\times W \\times D}" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19947" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 175 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 175 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 175 + ], + "type": "image", + "image_path": "8b853f5cd0d0d93dda2343e2340c2191a59949fdd685706de62f7d4cf4c46844.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 209 + ], + "type": "text", + "content": "Figure 2. Our MVS-embedded strategy effectively enlarges the receptive field of the depth optimization process across the whole epipolar line, which provides a good depth prior (Raw Argmin)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "content": "encoding per-voxel occupancy (and semantic) information. Here " + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "inline_equation", + "content": "H, W, D" + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "content": " denote the resolution of the occupancy grid and the set of predefined classes. First proposed by MonoScene [6], vision-based methods have become a promising approach to 3D occupancy prediction because of their low cost and effectiveness compared with 3D-based methods. Vision-based methods generally learn a mapping " + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\mathcal{D} \\circ \\mathcal{F} \\circ \\mathcal{E}" + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "content": " from RGB images " + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{\\mathbf{I}_n | n = 1, \\dots, N\\}" + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "content": " captured by " + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "content": " cameras to 3D occupancy " + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 46, + 217, + 286, + 324 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 334, + 287, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 334, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 85, + 334, + 287, + 347 + ], + "type": "interline_equation", + "content": "\\mathbf {O} = \\mathcal {M} (\\mathbf {I}) = \\mathcal {D} \\left(\\mathcal {F} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {1}\\right), \\dots , \\mathcal {E} \\left(\\mathbf {I} _ {N}\\right)\\right)\\right), \\tag {1}", + "image_path": "9b209bf59983e52b48b9b01c54316beb9573ba78710a39fbdae9360773efab65.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": "where a 2D backbone " + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": " first encodes " + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": " input images into multi-scale image features " + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\mathcal{E}(\\mathbf{I})" + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": ", and a 3D encoder " + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": " further lifts 2D features into 3D representation " + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = \\mathcal{F}(\\mathbf{F})" + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": " through attention mechanism [31, 40], lift-splat-shoot (LLS) [39] or identical transformation for monocular scenario [6], followed by a decoder network " + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": " to generate the final occupancy prediction result " + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = \\mathcal{D}(\\mathbf{R})" + }, + { + "bbox": [ + 46, + 357, + 287, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 596 + ], + "type": "text", + "content": "Although recent supervised vision-based approaches have achieved encouraging performance on 3D occupancy prediction given only visual inputs at inference, they still rely on 3D (semantic) supervision during training, e.g. Li-DAR point cloud or dense occupancy annotation. On the contrary, self-supervised counterparts [7, 71] take only temporal correspondences inherent in video sequences as supervision to predict meaningful occupancy, and therefore can effortlessly exploit the vast amount of unlabeled driving images. However, existing self-supervised methods [7, 71] consider only one monocular camera and decode forward occupancy " + }, + { + "bbox": [ + 46, + 441, + 287, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\text{front}}" + }, + { + "bbox": [ + 46, + 441, + 287, + 596 + ], + "type": "text", + "content": " directly from 2D image features by instantiating the 3D encoder " + }, + { + "bbox": [ + 46, + 441, + 287, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 441, + 287, + 596 + ], + "type": "text", + "content": " with identical transformation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 606, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 606, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 73, + 606, + 287, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {f r o n t} = \\mathcal {D} \\left(\\mathcal {F} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {m o n o}\\right)\\right)\\right) = \\mathcal {D} \\left(\\mathcal {E} \\left(\\mathbf {I} _ {m o n o}\\right)\\right). \\tag {2}", + "image_path": "8a2a242ae855bd017556fa2e80dc3cc0e15121307c32fd7a4d468e3aee89d8d6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 665 + ], + "type": "text", + "content": "Despite being straightforward, ignoring " + }, + { + "bbox": [ + 46, + 629, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 629, + 286, + 665 + ], + "type": "text", + "content": " may cause multicamera inconsistency in the surround view scenario and inadequate geometry reasoning in the 3D space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "To this end, we propose to transform the images into the BEV or TPV space to obtain a 3D representation of the scene in order to enable 3D feature interactions even in the monocular setting and also avoid ambiguities from multiple" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "cameras. Similar to the supervised literature [31, 40], we leverage the deformable cross-attention (CA) to adaptively aggregate information from the image features " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ", in which a set of learnable 3D tokens " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " serve as queries and the corresponding local image features serve as keys and values. Each of these learnable tokens represents a pillar area as in BEV or TPV representation. Moreover, we determine the correspondences between 3D tokens and image features with the projection matrices " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = \\{\\mathbf{T}_n|n = 1,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " from the ego car to pixel coordinates. We interleave deformable self-attention (SA), deformable cross-attention (CA) and feed forward network (FFN) to build the 3D encoder " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 352, + 222, + 545, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 222, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 352, + 222, + 545, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {l} = \\operatorname {F F N} _ {l} \\left(\\mathrm {C A} _ {l} \\left(\\mathrm {S A} _ {l} (\\mathbf {Q} _ {l}); \\mathbf {F}, \\mathbf {T}\\right)\\right), \\tag {3}", + "image_path": "c485d31c6bef760e6811b825e15994b7e8ca8eea27ca38f086688fcb36c9b1dd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 242, + 506, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 506, + 254 + ], + "type": "text", + "content": "where the subscription " + }, + { + "bbox": [ + 306, + 242, + 506, + 254 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 306, + 242, + 506, + 254 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 306, + 242, + 506, + 254 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 306, + 242, + 506, + 254 + ], + "type": "text", + "content": "-th block." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 260, + 454, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 454, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 454, + 273 + ], + "type": "text", + "content": "3.2.From Occupancy to Image" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "content": "Because 3D ground truth is unavailable, we project occupancy O from Eq. (1) back into 2D views and leverage multi-view consistency inherent in video sequences to optimize a self-supervised vision-based occupancy prediction system. Similar to previous work [7, 71], we employ differentiable volume rendering [47] to synthesize color and depth views, which enables the seamless integration of gradient information into the rendering pipeline and effective exploitation of supervision from multiple viewpoints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": "We first transform the 3D representation " + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": " into an SDF field " + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": ", which indicates the distance of every voxel center to the nearest surface of an object, by implementing the decoder network " + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": " with an MLP. We use different MLPs for different heights for BEV representation, while we construct a 3D feature volume from TPV representation with broadcasting and summation [31] before feeding the feature volume through a single MLP. For a continuous 3D coordinate " + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": ", we employ bicubic interpolation (BI) to predict the SDF value " + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "inline_equation", + "content": "s_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": " and determine its occupancy status " + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "inline_equation", + "content": "o_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": " according to the sign of " + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "inline_equation", + "content": "s_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 386, + 545, + 518 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 524, + 545, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 524, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 318, + 524, + 545, + 538 + ], + "type": "interline_equation", + "content": "o _ {\\mathbf {p}} = \\operatorname {s g n} \\left(s _ {\\mathbf {p}}\\right), \\quad s _ {\\mathbf {p}} = \\operatorname {B I} (\\mathbf {S}, \\mathbf {p}) = \\operatorname {B I} (\\mathcal {D} (\\mathbf {R}), \\mathbf {p}). \\tag {4}", + "image_path": "27544935e1c34b4287eb44acfdd346d124347633be635cab555ee97f34a652c1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 544, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 663 + ], + "type": "text", + "content": "The preference for SDF field over density field as in [7, 71] is based on two considerations: 1) An SDF field has more explicit physical meaning compared with a density field and inherently holds assumptions about the gradient magnitudes [24], thus allowing for easier regularization and optimization. 2) The signed nature of SDF facilitates straightforward determination of whether a point lies inside or outside a surface, thus enabling accurate differentiation of intricate geometries. Similar to SDF values, we can also decode color c from " + }, + { + "bbox": [ + 304, + 544, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 544, + 545, + 663 + ], + "type": "text", + "content": " by using a separate MLP " + }, + { + "bbox": [ + 304, + 544, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 304, + 544, + 545, + 663 + ], + "type": "text", + "content": " as the decoder:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 670, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 670, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 356, + 670, + 545, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\mathbf {p}} = \\operatorname {B I} (\\mathbf {C}, \\mathbf {p}) = \\operatorname {B I} \\left(\\mathcal {D} _ {c} (\\mathbf {R}), \\mathbf {p}\\right). \\tag {5}", + "image_path": "222dd39ef19ce13d7764b1b733ed9ff4476ccb47abc945df6965d6233b76087b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Take the volume rendering process of one single ray as an example. We first determine the origin and direction of" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19948" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 547, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 547, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 547, + 186 + ], + "type": "image", + "image_path": "15619fe9dc9b165b4de29b2ffa4604395f0c8b45b89360fe362cace141e97870.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 234 + ], + "type": "text", + "content": "Figure 3. Framework of the proposed SelfOcc for self-supervised vision-based 3D occupancy prediction. We employ an image backbone and a 3D encoder to generate 3D representations as in BEVFormer [40] or TPVFormer [31]. To render a novel view, we apply a lightweight MLP on the 3D features to predict the SDF values, color and semantic vectors. We then perform volume rendering to synthesize color, depth and semantic views. We use simple 0-thresholding to predict the occupancy volume." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "content": "the ray and uniformly sample " + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{p}_m|m = 1,\\dots,M\\}" + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "content": " between the origin and the intersection of the ray with the border of the 3D representation. The SDF values of these points " + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = \\{s_m|m = 1,\\dots,M\\}" + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "content": " are then calculated with Eq. (4). According to NeuS [65], the discrete opacity value " + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "inline_equation", + "content": "\\alpha_{m}" + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "content": ", which is the probability of the ray terminating between " + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_m" + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{m + 1}" + }, + { + "bbox": [ + 47, + 239, + 290, + 324 + ], + "type": "text", + "content": ", can be derived with" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 334, + 287, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 334, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 87, + 334, + 287, + 360 + ], + "type": "interline_equation", + "content": "\\alpha_ {m} = \\max \\left(\\frac {\\Phi_ {a} (s _ {m}) - \\Phi_ {a} (s _ {m + 1})}{\\Phi_ {a} (s _ {m})}, 0\\right), \\tag {6}", + "image_path": "98965f98c6ee892439c05786f0e41273940302183551027dd7286d539cc5b291.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\Phi_{a}(x) = (1 + e^{-ax})^{-1}" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": " is the sigmoid function with " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": " being a learnable parameter. Further, the probability of the ray passing through " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_m" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": ", i.e. the discrete accumulated transmittance, can be defined by " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "T_{m} = \\prod_{i = 1}^{m - 1}(1 - \\alpha_{i})" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": ". Finally, we can calculate the probability of the ray ending at " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_m" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": " after emission as " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "w_{m} = T_{m}\\alpha_{m}" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": ", and the rendering process is an integration of per-point attribute over the probability distribution " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\{w_m|m = 1,\\dots,M\\}" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": ". Eq. (7) calculates the rendered color " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_r" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "inline_equation", + "content": "d_r" + }, + { + "bbox": [ + 46, + 372, + 288, + 482 + ], + "type": "text", + "content": " of the ray:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 492, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 492, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 87, + 492, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {r} = \\sum_ {m = 1} ^ {M} w _ {m} \\mathbf {c} _ {m}, \\quad d _ {r} = \\sum_ {m = 1} ^ {M} w _ {m} d _ {m}, \\tag {7}", + "image_path": "b3b69546524f387e91f2f8cb769954b9cb74b2033c591f4c50ce8906908f126f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 286, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 286, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 286, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 536, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_m, d_m" + }, + { + "bbox": [ + 47, + 536, + 286, + 560 + ], + "type": "text", + "content": " denote the color and depth of the " + }, + { + "bbox": [ + 47, + 536, + 286, + 560 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 536, + 286, + 560 + ], + "type": "text", + "content": "-th point. Figure 3 shows the overall framework of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 572, + 227, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 227, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 227, + 586 + ], + "type": "text", + "content": "3.3. Occupancy-Oriented Supervision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "content": "In this section, we elaborate on the supervision formulation that facilitates self-supervised vision-based 3D occupancy prediction for autonomous driving." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "MVS-embedded Depth Learning. Depth supervision enhances the convergence of neural radiance fields [68], enabling faster and more effective learning of accurate geometry, especially with sparse views. Self-supervised depth estimation in NeRFs usually relies on the photometric reprojection loss " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "L_{rpj}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " which aims to maximize the similarity between the target image " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " and sampled pixels on the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 239, + 541, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 541, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 541, + 251 + ], + "type": "text", + "content": "source image " + }, + { + "bbox": [ + 306, + 239, + 541, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s" + }, + { + "bbox": [ + 306, + 239, + 541, + 251 + ], + "type": "text", + "content": " by refining the depth prediction " + }, + { + "bbox": [ + 306, + 239, + 541, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\mathbf{I}_t,\\pmb {\\theta})" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 258, + 545, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 258, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 311, + 258, + 545, + 288 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {r p j} (\\mathbf {x}, \\mathbf {I} _ {t}, \\mathbf {I} _ {s}; \\boldsymbol {\\theta}) = \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {s} (\\hat {\\mathbf {x}}) \\| \\tag {8} \\\\ = \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {s} \\left(\\pi (\\mathbf {x}, \\mathbf {D} (\\mathbf {I} _ {t}; \\boldsymbol {\\theta}) | _ {\\mathbf {x}}, \\Pi)\\right) \\|, \\\\ \\end{array}", + "image_path": "bfa64c1956eb29b54f5cf1c3c749be46f915c0a0d23e3275dfdb50501c5fcba0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\theta, \\mathbf{x}, \\hat{\\mathbf{x}}, \\Pi" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " denote the learnable parameters, a random pixel on the target image, the warped pixel on the source image and the projection matrix from the target to the source image coordinate, respectively. " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{D}|_{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " denote bilinear interpolation of corresponding 2D tensors, and " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\pi (\\mathbf{x},d_{\\mathbf{x}},\\Pi)" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " warps a pixel " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " to another image according to the depth " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " and projection matrix " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": ". The inherent local receptive field of the bilinear interpolation " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s(\\hat{\\mathbf{x}})" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " has an adverse effect on the optimization property of " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "L_{rpj}" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " since the differences between " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " and the colors of the four adjacent grid pixels of " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": " are the only reference information for the optimization of " + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 294, + 545, + 450 + ], + "type": "text", + "content": ", which could be easily misled by poor initialization or low-texture regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "text", + "content": "Therefore, we propose an MVS-embedded strategy to extend the receptive field of the depth optimization for a ray across the whole epipolar line. Specifically, we first take the uniformly spaced depths " + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = \\{d_m|m = 1,\\dots,M\\}" + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "text", + "content": " associated with " + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{p}_m|m = 1,\\dots,M\\}" + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "text", + "content": " as depth proposals. Then we define " + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "inline_equation", + "content": "L_{mvs}" + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "text", + "content": " as an average of the dissimilarity at each proposal " + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "text", + "content": " with weight " + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "inline_equation", + "content": "w_{m}" + }, + { + "bbox": [ + 304, + 451, + 546, + 535 + ], + "type": "text", + "content": " from volume rendering:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 541, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 541, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 319, + 541, + 545, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {m v s} (\\mathbf {x}, \\mathbf {I} _ {t}, \\mathbf {I} _ {s}; \\boldsymbol {\\theta}) \\\\ = \\sum_ {m = 1} ^ {M} w _ {m} (\\boldsymbol {\\theta}) \\| \\mathbf {I} _ {t} (\\mathbf {x}) - \\mathbf {I} _ {x} (\\pi (\\mathbf {x}, d _ {m}, \\boldsymbol {\\Pi})) \\|. \\tag {9} \\\\ \\end{array}", + "image_path": "f931ff18bfd1183f866b22c1357600bf397c07a24e0d169a973aef839c69dd29.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "text", + "content": "As Figure 2 shows, " + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "inline_equation", + "content": "L_{mvs}" + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "text", + "content": " enlarges the receptive field of depth optimization by associating " + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "text", + "content": " with multiple depth proposals along the ray, thus stabilizing and facilitating self-supervised depth optimization in NeRFs. Similar to Mon-depth2 [23], we formulate the final depth loss by taking the minimum of two " + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "inline_equation", + "content": "L_{mvs}" + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "text", + "content": " using the previous image " + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{t-1}" + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "text", + "content": " or the next image " + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{t+1}" + }, + { + "bbox": [ + 304, + 597, + 547, + 693 + ], + "type": "text", + "content": " as the source image and leverage the auto-masking strategy to filter out low texture regions:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 319, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "L _ {d e p} (\\mathbf {x}, \\mathbf {I} _ {t}) = \\operatorname {a u t o m a s k} \\left(\\min \\left(L _ {m v s} ^ {(t - 1)}, L _ {m v s} ^ {(t + 1)}\\right)\\right), \\tag {10}", + "image_path": "d3598ae749fdcd01411f7320d34a201feda78d9c218f4b1353d800be08655fe0.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19949" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 105, + 543, + 259 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. 3D occupancy prediction performance on the Occ3D-nuScenes [63] dataset. Cons. veh. and drive. surf. represent construction vehicle and driveable surface, respectively. Our method learns meaningful geometry and reasonable semantics compared with 3D-supervised methods, and achieves even higher IoU than LiDAR-supervised TPVFormer [31]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 105, + 543, + 259 + ], + "lines": [ + { + "bbox": [ + 49, + 105, + 543, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 105, + 543, + 259 + ], + "type": "table", + "html": "
MethodSupervisionIoUmIoUothersbarrierbicyclebuscarcons. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetation
MonoScene [6]3D-6.061.757.234.264.939.385.673.983.015.904.457.1714.916.327.927.431.017.65
OccFormer [82]3D-21.935.9430.2912.3234.4039.1714.4416.4517.229.2713.9026.3650.9930.9634.6622.736.766.97
BEVFormer [40]3D-26.885.8537.8317.8740.4442.437.3623.8821.8120.9822.3830.7055.3528.3636.028.0620.0417.69
CTF-Occ [63]3D-28.538.0939.3320.5638.2942.2416.9324.5222.7221.0522.9831.1153.3333.8437.9833.2320.7918.0
TPVFormer [31]3D-27.837.2238.9013.6740.7845.9017.2319.9918.8514.3026.6934.1755.6535.4737.5530.7019.4016.78
FB-OCC [41]3D-42.0614.3049.7130.0046.6251.5429.3029.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58
RenderOcc [53]L-23.935.6927.5614.3619.9120.5611.9612.4212.1414.3420.8118.9468.8533.3542.0143.9417.3622.61
TPVFormer [31]L17.2013.570.0014.809.3621.2716.8114.4513.7611.235.3216.0519.7310.759.439.5011.1616.5117.04
SelfOcc (BEV)C44.336.760.000.000.000.009.820.000.000.000.000.006.9747.030.0018.7516.5811.933.81
SelfOcc (TPV)C45.019.300.000.150.665.4612.540.000.802.100.000.008.2555.490.0026.3026.5414.225.60
", + "image_path": "3c7133485af5a130d169a3c1f068c3dd6ce103a59610e03da5fdc12812af8073.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 265, + 220, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 220, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 220, + 277 + ], + "type": "text", + "content": "where we omit the learnable parameters " + }, + { + "bbox": [ + 47, + 265, + 220, + 277 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 265, + 220, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": "Supervision from temporal frames. Unlike depth estimation which predicts only the visible surfaces, 3D occupancy prediction reconstructs the complete geometry of the scene which requires self-supervision signals from multiple viewpoints as in NeRFs. Considering the limited overlapping field of view for cameras, we take advantage of the temporal frames as supervision. Given a video sequence " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\mathbf{V}_1,\\dots,\\mathbf{V}_T\\}" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " frames and the current timesamp " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " as the input to our model " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " to predict the SDF and radiance fields " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_t,\\mathbf{C}_t\\}" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": ". Then we decide whether to use a temporal frame as supervision with probability " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": ". If positive, we randomly sample one frame " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "t^\\prime" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " for supervision satisfying that the distance between the ego cars at timestamps " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "t^\\prime" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " falls inside " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "[l_1,l_2]" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": ", which guarantees diversity and adheres to the range of the 3D representation. Otherwise, we use the current frame " + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 277, + 288, + 468 + ], + "type": "text", + "content": " for supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "content": "SDF Field Regularization. Neural radiance fields with sparse views suffer from overfitting because of a large solution set [51], which is exacerbated in autonomous driving with outward-facing cameras. Therefore, it is crucial to regularize the SDF field in order to restrict the solution set. To learn a smooth SDF field, we minimize the magnitude of the second-order derivatives of the SDF field with a Hessian loss [79] " + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "inline_equation", + "content": "L_{H}(\\mathbf{p}) = \\| \\mathbf{H}(\\mathbf{p})\\|_{1}" + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_1" + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "content": " is the elementwise matrix 1-norm, and " + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{H}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "content": " denotes the Hessian matrix of the SDF field with respect to the 3D coordinates at location " + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "content": ". To impose sparsity of occupancy in the invisible area such as those blocked by a wall, we propose a simple yet effective regularization " + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "inline_equation", + "content": "L_{s}(\\mathbf{p}) = \\max(-s_{\\mathbf{p}}, 0)" + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "content": ", which encourages negative SDF values to become larger. In addition, we include the Eikonal term [24] " + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "inline_equation", + "content": "L_{E}(\\mathbf{p}) = ||\\nabla s(\\mathbf{p})|| - 1|" + }, + { + "bbox": [ + 46, + 468, + 287, + 683 + ], + "type": "text", + "content": " for compliance with physical meaning of SDF. For optional semantic prediction, we use 2D segmentation maps from off-the-shelf segmentors as supervision similar to colors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 684, + 257, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 684, + 257, + 695 + ], + "spans": [ + { + "bbox": [ + 59, + 684, + 257, + 695 + ], + "type": "text", + "content": "Til now, we define the training loss " + }, + { + "bbox": [ + 59, + 684, + 257, + 695 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 59, + 684, + 257, + 695 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 701, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 701, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 62, + 701, + 287, + 715 + ], + "type": "interline_equation", + "content": "L = L _ {d e p} + \\lambda_ {c} L _ {r g b} + \\lambda_ {e} L _ {E} + \\lambda_ {H} L _ {H} + \\lambda_ {s} L _ {s} \\tag {11}", + "image_path": "acdd7a88814bb49011ac790bbdfa054d7e44c341da17b3ea5ed482ba2396724b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 299, + 539, + 441 + ], + "blocks": [ + { + "bbox": [ + 305, + 263, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 263, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 545, + 297 + ], + "type": "text", + "content": "Table 2. 3D occupancy prediction performance on the SemanticKITTI [3] dataset. MonoScene* is supervised by depth predictions of Monodepth2 [23] trained with ground-truth poses." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 299, + 539, + 441 + ], + "lines": [ + { + "bbox": [ + 307, + 299, + 539, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 539, + 441 + ], + "type": "table", + "html": "
MethodSupervisionIoUPrec.Rec.
3DDepth Image
MonoScene [6]✓37.1449.9059.24
\\(LMSCNet^{rgb}\\) [59]✓12.0813.0063.16
\\(3DSketch^{rgb}\\) [10]✓12.0112.9562.31
\\(AICNet^{rgb}\\) [36]✓11.2811.8470.89
MonoScene [6]✓13.5316.9840.06
MonoScene* [6]11.1813.1540.22
SceneRF [7]✓13.8417.2840.96
SelfOcc (BEV)✓20.9537.2732.37
SelfOcc (TPV)✓21.9734.8337.31
", + "image_path": "b333d38e19cdc69d9f66eb7d4b2209313ef8006644c88d7056ad9f9f8e058b01.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "text", + "content": "where the four " + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "text", + "content": "s are hyperparameters and " + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "inline_equation", + "content": "L_{rgb}" + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "text", + "content": " measures the dissimilarity between the rendered color " + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_r" + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "text", + "content": " of a ray and the ground truth color on the target image. For semantic prediction, we optionally add a semantic loss " + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "inline_equation", + "content": "L_{sem}" + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "text", + "content": " similar to " + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "inline_equation", + "content": "L_{rgb}" + }, + { + "bbox": [ + 304, + 448, + 547, + 509 + ], + "type": "text", + "content": " and an off-the-shelf 2D segmentor." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 519, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 519, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 519, + 388, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 539, + 411, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 411, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 411, + 552 + ], + "type": "text", + "content": "4.1. Task Descriptions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 665 + ], + "type": "text", + "content": "The 3D occupancy prediction task aims to predict the dense occupancy (and semantic) states for a voxel grid, which is a fundamental task in autonomous driving due to its fine granularity and least abstraction of the real world. We conduct this task on the Occ3D-nuScenes [63] dataset, and use IoU and mIoU as geometric and semantic metrics, respectively. In addition, we evaluate our method on the SemanticKITTI [3] dataset for the geometry-only 3D occupancy prediction task and report IoU, Precision and Recall." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "The novel depth synthesis task synthesizes depth maps from novel viewpoints within a certain distance given the input frame [7]. This task is more reliable as a proxy for 3D reconstruction compared with novel view synthesis because" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19950" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 95, + 531, + 231 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 3. Novel depth synthesis on SemanticKITTI [3] and nuScenes [4]. We use the results reported in SceneRF [7] and all methods are trained using only images with ground truth poses. We adapt SceneRF [7] to the nuScenes dataset based on the official implementation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 95, + 531, + 231 + ], + "lines": [ + { + "bbox": [ + 60, + 95, + 531, + 231 + ], + "spans": [ + { + "bbox": [ + 60, + 95, + 531, + 231 + ], + "type": "table", + "html": "
MethodDatasetAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
MonoDepth2 [23]0.52597.11314.431.029210.4426.3241.43
SynSin [70]0.36113.4838.8240.429052.6174.5686.50
PixelNeRF [75]0.23642.0806.4490.335465.8185.4392.90
MINE [37]SemanticKITTI0.22481.7876.3430.328365.8785.5293.30
VisionNerf [42]0.20541.4905.8410.307369.1188.2894.37
SceneRF [7]0.16811.2915.7810.285175.0789.0994.50
SelfOcc (TPV)0.15621.1535.3490.253478.9291.1295.92
SceneRF [7]nuScenes0.763115.3411.500.624040.2664.8677.02
SelfOcc (TPV)0.40036.7238.4600.416564.5080.4088.17
", + "image_path": "a26d3e74758eb5d0bc52fe8f0de063b15395674371db9c6b1c114d1512afac8e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 269, + 533, + 509 + ], + "blocks": [ + { + "bbox": [ + 47, + 236, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 236, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 545, + 269 + ], + "type": "text", + "content": "Table 4. Self-supervised depth estimation on the nuScenes [4] and KITTI-2015 [21] dataset. M: trained with monocular sequence. MS: trained with both monocular sequence and stereo images. SV: trained with surround-view images. TO: temporal offline refinement. We train SurroundDepth* with ground truth poses based on the official implementation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 269, + 533, + 509 + ], + "lines": [ + { + "bbox": [ + 61, + 269, + 533, + 509 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 533, + 509 + ], + "type": "table", + "html": "
MethodDatasetTypeScale MethodAbs Rel↓Sq Rel↓RMSE↓RMSE log↓δ1↑δ2↑δ3↑
Monodepth2 [23]nuScenesSVMedian0.2873.3497.1840.3450.6410.8450.925
FSM [26]SVMedian0.299------
SurroundDepth [69]SVSfM Pretrain0.2804.4017.4670.3640.6610.8440.917
SurroundDepth* [69]SVPose GT0.3427.9178.0190.3600.7160.8580.918
SimpleOcc [19]SVPose GT0.2243.3837.1650.3330.7530.8770.930
SelfOcc (TPV)SVPose GT0.2152.7436.7060.3160.7530.8750.932
R3D3 [60]nuScenesTOExtrinsics0.2534.7597.150-0.729--
Johnston [33]KITTI-2015MMedian0.1060.8614.6990.1850.8890.9620.982
FeatDepth(R50) [61]MMedian0.1040.7294.4810.1790.8930.9650.984
PackNet-SfM [25]MMedian0.1070.8024.5380.1860.8890.9620.981
R-MSFM6 [86]MMedian0.1080.7484.4700.1850.8890.9630.982
DevNet [84]MMedian0.1000.6994.4120.1740.8930.9660.985
SelfOcc (TPV)MPose GT0.1030.7924.6730.1870.8770.9590.982
MonoDepth2(R50) [23]KITTI-2015MSExtrinsics0.1060.8184.750.1960.8740.9570.979
FeatDepth(R50) [61]MSExtrinsics0.0990.6974.4270.1840.8890.9630.982
R-MSFM6 [86]MSExtrinsics0.1080.7534.4690.1850.8880.9630.982
DevNet [84]MSExtrinsics0.0950.6714.3650.1740.8950.9700.988
BTS [71]MSPose GT0.1020.7514.4070.1880.8820.9610.982
SelfOcc (TPV)MSPose GT0.0990.7114.5860.1860.8800.9600.982
", + "image_path": "d4600b0850b3c5de3ff6116110ffd850941f79429996d3375d827efab3e42fb7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "text", + "content": "the rendered color is an integration of predicted radiance at sampled locations, which introduces unnecessary degrees of freedom. We conduct experiments on nuScenes [4] and SemanticKITTI [3] and use Abs Rel, Sq Rel, RMSE, RMSE log and threshold accuracies " + }, + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "inline_equation", + "content": "(\\delta 1, \\delta 2, \\delta 3)" + }, + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "text", + "content": " as metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "content": "The depth estimation task is a traditional task for 3D perception, which predicts the pixel-aligned depth map for a given frame. We conduct surround-view depth estimation on nuScenes [4] and monocular depth estimation on KITTI-2015 [21]. The same metrics are used as in novel depth synthesis. We detail the datasets in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 180, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 180, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 180, + 671 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "We use the identical network architecture for all three tasks, while the temporal supervision and the loss formulation may differ according to their respective objectives. We" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 514, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 599 + ], + "type": "text", + "content": "adopt ResNet50 [27] pretrained on ImageNet [14] as the 2D backbone and an feature pyramid network (FPN) [43] to generate multi-scale image features. The 3D encoder " + }, + { + "bbox": [ + 304, + 514, + 545, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 514, + 545, + 599 + ], + "type": "text", + "content": " is the same as in BEVFormer [40] or TPVFormer [31] depending on the 3D representation we use. We further employ a two-layer MLP as the decoder network " + }, + { + "bbox": [ + 304, + 514, + 545, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 514, + 545, + 599 + ], + "type": "text", + "content": " to generate the SDF field, color and semantic logits (optional)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": "We activate temporal supervision for 3D occupancy prediction and novel view synthesis which both involve scene reconstruction from multiple viewpoints, while we use only the current frame as supervision for depth estimation since it predicts the geometry only for the input view. As for loss formulation, we use " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "L_{dep}" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "L_{E}" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": " and an edge loss " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "L_{edg}" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": " [22] for depth estimation since RGB supervision from the input view is meaningless, while " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "L_{H}" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": " focus on improving the overall geometry, thus helpless for depth es" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19951" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 105, + 279, + 166 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": "Table 5. Architecture ablation on the nuScenes dataset. " + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "inline_equation", + "content": "L_{mvs}" + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": ": the proposed " + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "inline_equation", + "content": "L_{mvs}(\\checkmark)" + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "inline_equation", + "content": "L_{rpj}" + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": " for depth optimization. SDF: SDF field (" + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": ") or density field for volume rendering." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 105, + 279, + 166 + ], + "lines": [ + { + "bbox": [ + 54, + 105, + 279, + 166 + ], + "spans": [ + { + "bbox": [ + 54, + 105, + 279, + 166 + ], + "type": "table", + "html": "
LmvsSDFOcc.Novel DepthDepth
IoUmIoUAbs Rel↓RMSE↓Abs Rel↓RMSE↓
✓18.902.810.810112.4080.65513.242
✓36.686.100.40078.1490.3188.524
✓✓38.436.940.43768.7340.3127.544
", + "image_path": "8398e98949e52a90bf7242a434ae1750223b5147e9f22f2969af8c960866e6fa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 198, + 286, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 176, + 286, + 197 + ], + "lines": [ + { + "bbox": [ + 47, + 176, + 286, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 286, + 197 + ], + "type": "text", + "content": "Table 6. Temporal supervision ablation on nuScenes [4]. We do not use semantic supervision for occupancy prediction here." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 198, + 286, + 277 + ], + "lines": [ + { + "bbox": [ + 48, + 198, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 286, + 277 + ], + "type": "table", + "html": "
SupervisionRatioOcc.NovelDepthDepth
Curr.Prev.NextIoUPrec.Rec.Abs RelRMSEAbs RelRMSE
10039.6766.9649.330.4718.8520.3127.544
21140.5067.5150.310.4468.8270.3658.046
11140.4763.9852.410.4388.7340.3637.975
12240.4363.3552.780.4428.7810.3667.998
01142.4870.0851.890.4628.7320.3928.075
", + "image_path": "be7a29dd85e42a4ffede6079ff723c499f18d5dbd638a19feca1dd49298aa76c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 323, + 286, + 403 + ], + "blocks": [ + { + "bbox": [ + 47, + 290, + 286, + 322 + ], + "lines": [ + { + "bbox": [ + 47, + 290, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 286, + 322 + ], + "type": "text", + "content": "Table 7. Loss and regularization ablation on the SemanticKITTI [3] (occupancy prediction and novel depth synthesis) and KITTI-2015 [21] (depth estimation) datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 323, + 286, + 403 + ], + "lines": [ + { + "bbox": [ + 48, + 323, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 286, + 403 + ], + "type": "table", + "html": "
LdepLrgbLHLsOcc.Novel DepthDepth
IoUPrec.Rec.Abs RelRMSEAbs RelRMSE
✓9.819.8596.180.385211.67--
✓23.5141.0135.510.15795.3480.1024.676
✓✓24.0737.7839.880.15735.3520.1004.773
✓✓✓12.2713.0268.140.17225.7480.1075.091
✓✓✓✓22.1633.5439.500.16945.6010.1135.078
", + "image_path": "096927ef2c5b16393b7e0351936c1f6b6abf5bce7ca8eb0542da09f082f7988e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 408, + 286, + 528 + ], + "blocks": [ + { + "bbox": [ + 50, + 408, + 286, + 528 + ], + "lines": [ + { + "bbox": [ + 50, + 408, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 50, + 408, + 286, + 528 + ], + "type": "image", + "image_path": "c2fd7c0769098dfb48e549522f7fd067b121e774e1fd08c819dfd8f03dfc59e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "lines": [ + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "text", + "content": "Figure 4. Visualization of the effect of different loss functions. The occupancy quality gradually improves with " + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "inline_equation", + "content": "L_{dep}" + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "inline_equation", + "content": "L_{H}" + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 47, + 532, + 286, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "timation. For novel depth synthesis, we leverage " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "L_{dep}" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "L_{rgb}" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "L_{E}" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": ". And for 3D occupancy prediction, we further activate " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "L_{H}" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " on the basis of novel depth synthesis to impose smoothness and sparsity constraints on the predicted occupancy grid. For semantic prediction, we use the off-the-shelf open-vocabulary segmentation network OpenSeeD [77] to generate pseudo segmentation labels for nuScenes [4]. We use the AdamW [45] optimizer with an initial learning rate of 1e-4 which decays to zero following the cosine schedule. We train our models for 12 epochs on the nuScenes [4] dataset and 24 epochs on the SemanticKITTI [3] and KITTI-2015 [21] datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 425, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 425, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 425, + 84 + ], + "type": "text", + "content": "4.3. Results and Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "text", + "content": "3D Occupancy Prediction. We report the results for the 3D occupancy prediction task in Table 1 and 2. On the Occ3D-nuScenes [63] dataset, our methods achieve comparable IoU and reasonable mIoU without any form of 3D supervision compared with supervised approaches. The IoU metric " + }, + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "inline_equation", + "content": "(45.01\\%)" + }, + { + "bbox": [ + 304, + 96, + 545, + 252 + ], + "type": "text", + "content": " indicates that SelfOcc learns meaningful geometry given only video sequences. We notice that SelfOcc is better at predicting background classes than foreground classes, which might be related to the semantic misalignment of the open-vocabulary 2D segmentation network we use. Nonetheless, our method still outperforms MonoScene [6] supervised by ground truth in mIoU and LiDAR-supervised TPVFormer [31] in IoU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": "On the SemanticKITTI [3] dataset, we use the results reported in SceneRF [7], where baselines are grouped into 3D, depth and image supervised categories, and the superscription " + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "inline_equation", + "content": "{}^{rgb}" + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": " indicates the adaptation from 3D input to image input. From Table 2, SelfOcc sets the new state-of-the-art for vision-based depth-supervised and self-supervised 3D occupancy prediction tasks, which outperforms SceneRF [7] by " + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "inline_equation", + "content": "{58.7}\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": " in IoU,with much higher Precision and slightly lower Recall. Visualizations for 3D occupancy predictions are shown in Figure 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 384, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 576 + ], + "type": "text", + "content": "Novel Depth Synthesis. We report the results for novel depth synthesis in Table 3, which averages depth metrics over all frames no further than 10 meters away from the input frame. Following SceneRF [7], we compare against methods based on monocular depth estimation [23], generalizable NeRF [37, 42, 75] and 3D-aware GAN [70]. Our method outperforms previous state-of-the-art SceneRF [7] on all metrics on the SemanticKITTI [3] dataset. In addition, we modify SceneRF for the nuScenes [4] dataset by randomly selecting one camera every iteration during training, and directly report the depth metric for the input frame considering the complexity of volume rendering conditioned on image features from surround cameras. SelfOcc still achieves better results than SceneRF, which confirms its superiority in 3D reconstruction from surround views. We provide novel depth visualizations in the supplementary." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Depth Estimation. We report the results for depth estimation in Table 4. Our method achieves state-of-the-art performance on nuScenes for self-supervised surround-view depth estimation. We also train SurroundDepth [69] using the official implementation with ground-truth poses to eliminate their influence, which indicates that ground-truth poses do not necessarily boost performance compared with median-scaling applied for scale-ambiguous methods. Additionally, SelfOcc performs on par with the state-of-the-art for self-supervised depth estimation on KITTI-2015 [21]. Visualizations are provided in the supplementary." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19952" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 194, + 132 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 194, + 132 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 194, + 132 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 194, + 132 + ], + "type": "image", + "image_path": "ec6a5899ee9c761da65f0dfb932e3b00be50f5d89b50d7fd90da1fe1c50f61c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 199, + 77, + 283, + 133 + ], + "blocks": [ + { + "bbox": [ + 199, + 70, + 279, + 76 + ], + "lines": [ + { + "bbox": [ + 199, + 70, + 279, + 76 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 279, + 76 + ], + "type": "text", + "content": "Semantic Occupancy Prediction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 199, + 77, + 283, + 133 + ], + "lines": [ + { + "bbox": [ + 199, + 77, + 283, + 133 + ], + "spans": [ + { + "bbox": [ + 199, + 77, + 283, + 133 + ], + "type": "image", + "image_path": "edccc952ae8cd9ba197a15d7b304e802302fa6e30118f12ba70338e710b88496.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 79, + 370, + 133 + ], + "blocks": [ + { + "bbox": [ + 294, + 70, + 358, + 76 + ], + "lines": [ + { + "bbox": [ + 294, + 70, + 358, + 76 + ], + "spans": [ + { + "bbox": [ + 294, + 70, + 358, + 76 + ], + "type": "text", + "content": "Semantic Occupancy GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 79, + 370, + 133 + ], + "lines": [ + { + "bbox": [ + 286, + 79, + 370, + 133 + ], + "spans": [ + { + "bbox": [ + 286, + 79, + 370, + 133 + ], + "type": "image", + "image_path": "2042a438150f9db3fc1ee4c8c58314ca60d17ca71b72a0412f52523149151f05.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 375, + 79, + 456, + 132 + ], + "blocks": [ + { + "bbox": [ + 375, + 70, + 457, + 77 + ], + "lines": [ + { + "bbox": [ + 375, + 70, + 457, + 77 + ], + "spans": [ + { + "bbox": [ + 375, + 70, + 457, + 77 + ], + "type": "text", + "content": "Geometric Occupancy Prediction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 79, + 456, + 132 + ], + "lines": [ + { + "bbox": [ + 375, + 79, + 456, + 132 + ], + "spans": [ + { + "bbox": [ + 375, + 79, + 456, + 132 + ], + "type": "image", + "image_path": "1d061b9992af1f3b3ba806096e8e503010b3ce011b7fef528dbd2b0daf9fc918.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 462, + 77, + 545, + 132 + ], + "blocks": [ + { + "bbox": [ + 468, + 70, + 534, + 76 + ], + "lines": [ + { + "bbox": [ + 468, + 70, + 534, + 76 + ], + "spans": [ + { + "bbox": [ + 468, + 70, + 534, + 76 + ], + "type": "text", + "content": "Geometric Occupancy GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 77, + 545, + 132 + ], + "lines": [ + { + "bbox": [ + 462, + 77, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 462, + 77, + 545, + 132 + ], + "type": "image", + "image_path": "9fb842a46eb7d5a21734dbbfd05f63a8a041607ae85e772ffda6b019f58b5689.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 135, + 194, + 191 + ], + "blocks": [ + { + "bbox": [ + 49, + 135, + 194, + 191 + ], + "lines": [ + { + "bbox": [ + 49, + 135, + 194, + 191 + ], + "spans": [ + { + "bbox": [ + 49, + 135, + 194, + 191 + ], + "type": "image", + "image_path": "7588c220b99e22bcd38339a533f2b9282a817e78b5d4f4251fb88110bfb42a11.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 199, + 136, + 282, + 191 + ], + "blocks": [ + { + "bbox": [ + 199, + 136, + 282, + 191 + ], + "lines": [ + { + "bbox": [ + 199, + 136, + 282, + 191 + ], + "spans": [ + { + "bbox": [ + 199, + 136, + 282, + 191 + ], + "type": "image", + "image_path": "63cb76b35759bbab4a90d88da8dd0c6a0ff9c71ad4a99fdeb02471c71dd5afb4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 286, + 136, + 369, + 191 + ], + "blocks": [ + { + "bbox": [ + 286, + 136, + 369, + 191 + ], + "lines": [ + { + "bbox": [ + 286, + 136, + 369, + 191 + ], + "spans": [ + { + "bbox": [ + 286, + 136, + 369, + 191 + ], + "type": "image", + "image_path": "202d1eeb53be0876bc39c50c154dc5c61c1ed1327fb74fa572cd8047fea29588.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 375, + 137, + 456, + 191 + ], + "blocks": [ + { + "bbox": [ + 375, + 137, + 456, + 191 + ], + "lines": [ + { + "bbox": [ + 375, + 137, + 456, + 191 + ], + "spans": [ + { + "bbox": [ + 375, + 137, + 456, + 191 + ], + "type": "image", + "image_path": "8a47d1ca153c83a1aa239bdba9bbb932277c0167efe018071bb485485a2e2bbf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 462, + 136, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 462, + 136, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 462, + 136, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 462, + 136, + 545, + 191 + ], + "type": "image", + "image_path": "a69f4ddad804a9011b47b2c09079bf3a3e400b73d78d7998faabe670f82d6e04.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 49, + 194, + 194, + 250 + ], + "blocks": [ + { + "bbox": [ + 49, + 194, + 194, + 250 + ], + "lines": [ + { + "bbox": [ + 49, + 194, + 194, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 194, + 250 + ], + "type": "image", + "image_path": "368f2eb8a629d3f1e4357d64b8a6955422164249c9bf3dfae2702b6ba41c7152.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 199, + 194, + 282, + 250 + ], + "blocks": [ + { + "bbox": [ + 199, + 194, + 282, + 250 + ], + "lines": [ + { + "bbox": [ + 199, + 194, + 282, + 250 + ], + "spans": [ + { + "bbox": [ + 199, + 194, + 282, + 250 + ], + "type": "image", + "image_path": "62d9aa8bfae0255fc815d5334757d8b976f23ee64bdc1ac0d87156dbaaede636.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 286, + 194, + 369, + 250 + ], + "blocks": [ + { + "bbox": [ + 286, + 194, + 369, + 250 + ], + "lines": [ + { + "bbox": [ + 286, + 194, + 369, + 250 + ], + "spans": [ + { + "bbox": [ + 286, + 194, + 369, + 250 + ], + "type": "image", + "image_path": "bdaa6aed7018b969a106d843c98abd73c782c39a84704d2086798fd436341dda.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 375, + 194, + 456, + 250 + ], + "blocks": [ + { + "bbox": [ + 375, + 194, + 456, + 250 + ], + "lines": [ + { + "bbox": [ + 375, + 194, + 456, + 250 + ], + "spans": [ + { + "bbox": [ + 375, + 194, + 456, + 250 + ], + "type": "image", + "image_path": "72819640aed9cbc8d5b701b55233029e74201010987a22b6a976325140d84c5b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 462, + 194, + 545, + 250 + ], + "blocks": [ + { + "bbox": [ + 462, + 194, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 462, + 194, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 462, + 194, + 545, + 250 + ], + "type": "image", + "image_path": "8f13c3971692f81024c4886725341eee39f18a07daf0e5fb73c986c861499b99.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 49, + 252, + 194, + 308 + ], + "blocks": [ + { + "bbox": [ + 49, + 252, + 194, + 308 + ], + "lines": [ + { + "bbox": [ + 49, + 252, + 194, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 194, + 308 + ], + "type": "image", + "image_path": "0de870ab8bcd2f24bbda702b515c0437caad3ea23e59e5f66372a90e6d61fedb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 309, + 91, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 309, + 91, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 91, + 316 + ], + "type": "text", + "content": "BACK LEFT" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 112, + 309, + 130, + 316 + ], + "lines": [ + { + "bbox": [ + 112, + 309, + 130, + 316 + ], + "spans": [ + { + "bbox": [ + 112, + 309, + 130, + 316 + ], + "type": "text", + "content": "BACK" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 149, + 309, + 188, + 316 + ], + "lines": [ + { + "bbox": [ + 149, + 309, + 188, + 316 + ], + "spans": [ + { + "bbox": [ + 149, + 309, + 188, + 316 + ], + "type": "text", + "content": "BACK RIGHT" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 199, + 254, + 282, + 308 + ], + "blocks": [ + { + "bbox": [ + 199, + 254, + 282, + 308 + ], + "lines": [ + { + "bbox": [ + 199, + 254, + 282, + 308 + ], + "spans": [ + { + "bbox": [ + 199, + 254, + 282, + 308 + ], + "type": "image", + "image_path": "ee1ce8b951504b5e64e871396105535f271762991502ecc9c49c227bcfab4b5c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 286, + 254, + 369, + 308 + ], + "blocks": [ + { + "bbox": [ + 286, + 254, + 369, + 308 + ], + "lines": [ + { + "bbox": [ + 286, + 254, + 369, + 308 + ], + "spans": [ + { + "bbox": [ + 286, + 254, + 369, + 308 + ], + "type": "image", + "image_path": "ec3ba5f0200396d913c9d841988dede7372a9a15320383a00ea360141e403738.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 321, + 341, + 328 + ], + "lines": [ + { + "bbox": [ + 310, + 321, + 341, + 328 + ], + "spans": [ + { + "bbox": [ + 310, + 321, + 341, + 328 + ], + "type": "text", + "content": "other flat" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 349, + 321, + 369, + 328 + ], + "lines": [ + { + "bbox": [ + 349, + 321, + 369, + 328 + ], + "spans": [ + { + "bbox": [ + 349, + 321, + 369, + 328 + ], + "type": "text", + "content": "pedc" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 375, + 254, + 456, + 308 + ], + "blocks": [ + { + "bbox": [ + 375, + 254, + 456, + 308 + ], + "lines": [ + { + "bbox": [ + 375, + 254, + 456, + 308 + ], + "spans": [ + { + "bbox": [ + 375, + 254, + 456, + 308 + ], + "type": "image", + "image_path": "e91500f337661ed575abab38f918db33867c79c2d21c78c228aa0e28bc21f0d5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 462, + 254, + 545, + 308 + ], + "blocks": [ + { + "bbox": [ + 462, + 254, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 462, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 462, + 254, + 545, + 308 + ], + "type": "image", + "image_path": "058acdca8fe2904be7f1efe1ebdd7dbe55ab212b9d284f73ba1309a36bf31d09.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 522, + 333, + 539, + 341 + ], + "lines": [ + { + "bbox": [ + 522, + 333, + 539, + 341 + ], + "spans": [ + { + "bbox": [ + 522, + 333, + 539, + 341 + ], + "type": "text", + "content": "higher" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 50, + 320, + 67, + 328 + ], + "blocks": [ + { + "bbox": [ + 50, + 320, + 67, + 328 + ], + "lines": [ + { + "bbox": [ + 50, + 320, + 67, + 328 + ], + "spans": [ + { + "bbox": [ + 50, + 320, + 67, + 328 + ], + "type": "image", + "image_path": "b94dc920af79801cfdc67c72583b81a59122694651e28b8cb8d0bf08bcaae681.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 321, + 86, + 328 + ], + "lines": [ + { + "bbox": [ + 67, + 321, + 86, + 328 + ], + "spans": [ + { + "bbox": [ + 67, + 321, + 86, + 328 + ], + "type": "text", + "content": "Veabl s" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 346, + 545, + 368 + ], + "lines": [ + { + "bbox": [ + 47, + 346, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 545, + 368 + ], + "type": "text", + "content": "Figure 5. Visualization results for 3D occupancy prediction on the Occ3D-nuScenes [63] dataset. Our method achieves comparable visualization quality with ground truth for both semantic and geometric occupancy prediction tasks." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 95, + 321, + 106, + 328 + ], + "blocks": [ + { + "bbox": [ + 95, + 321, + 106, + 328 + ], + "lines": [ + { + "bbox": [ + 95, + 321, + 106, + 328 + ], + "spans": [ + { + "bbox": [ + 95, + 321, + 106, + 328 + ], + "type": "image", + "image_path": "db925c8d306551a7c10073ee772ed908a67c61baec771db81994f65799ec1054.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 321, + 117, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 321, + 117, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 117, + 328 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 137, + 321, + 148, + 328 + ], + "blocks": [ + { + "bbox": [ + 137, + 321, + 148, + 328 + ], + "lines": [ + { + "bbox": [ + 137, + 321, + 148, + 328 + ], + "spans": [ + { + "bbox": [ + 137, + 321, + 148, + 328 + ], + "type": "image", + "image_path": "4b52e703fd9a719beebfd36b7ba893aaf7edc9eaf1e1772bdd39899fe8971a94.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 321, + 159, + 328 + ], + "lines": [ + { + "bbox": [ + 148, + 321, + 159, + 328 + ], + "spans": [ + { + "bbox": [ + 148, + 321, + 159, + 328 + ], + "type": "text", + "content": "ous" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 180, + 321, + 203, + 328 + ], + "blocks": [ + { + "bbox": [ + 180, + 321, + 203, + 328 + ], + "lines": [ + { + "bbox": [ + 180, + 321, + 203, + 328 + ], + "spans": [ + { + "bbox": [ + 180, + 321, + 203, + 328 + ], + "type": "image", + "image_path": "7dbb839b407618438fe650a706ec637f0739d8332e852bfb88cfdc81db63ec1d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 321, + 219, + 328 + ], + "lines": [ + { + "bbox": [ + 203, + 321, + 219, + 328 + ], + "spans": [ + { + "bbox": [ + 203, + 321, + 219, + 328 + ], + "type": "text", + "content": "terrain" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 235, + 321, + 260, + 328 + ], + "blocks": [ + { + "bbox": [ + 235, + 321, + 260, + 328 + ], + "lines": [ + { + "bbox": [ + 235, + 321, + 260, + 328 + ], + "spans": [ + { + "bbox": [ + 235, + 321, + 260, + 328 + ], + "type": "image", + "image_path": "e493d18eeaf104a74429ce83be94d0226542b28b6537ae212e8ba2222d575e27.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 267, + 321, + 277, + 328 + ], + "blocks": [ + { + "bbox": [ + 267, + 321, + 277, + 328 + ], + "lines": [ + { + "bbox": [ + 267, + 321, + 277, + 328 + ], + "spans": [ + { + "bbox": [ + 267, + 321, + 277, + 328 + ], + "type": "image", + "image_path": "7411a4c4205fdbd1355c4dc73eecf4a66b78b744676b5ed7ccada8d40d7c74fc.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 321, + 300, + 328 + ], + "lines": [ + { + "bbox": [ + 283, + 321, + 300, + 328 + ], + "spans": [ + { + "bbox": [ + 283, + 321, + 300, + 328 + ], + "type": "text", + "content": "dewalk" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 400, + 321, + 402, + 328 + ], + "blocks": [ + { + "bbox": [ + 383, + 321, + 401, + 328 + ], + "lines": [ + { + "bbox": [ + 383, + 321, + 401, + 328 + ], + "spans": [ + { + "bbox": [ + 383, + 321, + 401, + 328 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 400, + 321, + 417, + 328 + ], + "lines": [ + { + "bbox": [ + 400, + 321, + 417, + 328 + ], + "spans": [ + { + "bbox": [ + 400, + 321, + 417, + 328 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 400, + 333, + 416, + 340 + ], + "lines": [ + { + "bbox": [ + 400, + 333, + 416, + 340 + ], + "spans": [ + { + "bbox": [ + 400, + 333, + 416, + 340 + ], + "type": "text", + "content": "lower" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 400, + 321, + 402, + 328 + ], + "lines": [ + { + "bbox": [ + 400, + 321, + 402, + 328 + ], + "spans": [ + { + "bbox": [ + 400, + 321, + 402, + 328 + ], + "type": "image", + "image_path": "3fc63cd1b1ef2ace70724c2788b9096be492123d74c4488c756ce762255ce4f5.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 333, + 416, + 340 + ], + "lines": [ + { + "bbox": [ + 400, + 333, + 416, + 340 + ], + "spans": [ + { + "bbox": [ + 400, + 333, + 416, + 340 + ], + "type": "text", + "content": "lower" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 47, + 374, + 140, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 140, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 140, + 387 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 46, + 395, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 443 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies on architectural choices, temporal supervision, and loss formulations to validate the effectiveness of every component of our method. All experiments are trained for half of the original epochs." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "text", + "content": "Architecture. We investigate the effect of " + }, + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "inline_equation", + "content": "L_{mvs}" + }, + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "text", + "content": " and the SDF field in Table 5. MVS-embedded depth optimization brings consistent improvement for all three tasks. While the SDF field outperforms the density field in occupancy prediction and depth estimation, it achieves worse results in novel depth synthesis. We think it is because the SDF field holds an inherent smoothness prior and is thus more difficult to optimize for novel depth synthesis." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 46, + 544, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 615 + ], + "type": "text", + "content": "Temporal Supervision. We report the results of different temporal supervision settings in Table 6. It is shown that increasing temporal supervision benefits occupancy prediction, but deteriorates depth estimation performance. In addition, the best score for novel depth synthesis is achieved with a moderate level of temporal supervision (1:1:1)." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Loss Formulations. We ablate the effect of loss functions in Table 7. Note that we replace " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{rbg}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{edg}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " for depth estimation task. Models trained with depth loss " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{dep}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and color loss " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{rgb}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " achieve best performance for depth-related tasks. As shown in Figure 4, we use the Hessian loss " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{H}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and sparsity loss " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " to regularize the SDF field for better visualization quality, although models trained without them can also achieve high IoU during evaluation. In fact," + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 375, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 411 + ], + "type": "text", + "content": "the sparsity loss makes the prediction more similar to the ground truth, but it may not improve the IoU under the evaluation protocol of using camera masks as in Occ3D [63]." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 305, + 418, + 462, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 462, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 462, + 430 + ], + "type": "text", + "content": "4.5. 3D Occupancy Visualization" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 437, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 485 + ], + "type": "text", + "content": "We provide the visualization results for 3D occupancy prediction on nuScenes [4] in Figure 5. Despite being trained with video sequences only, our model produces meaningful geometric and reasonable semantic predictions." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 305, + 495, + 379, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 379, + 507 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 515, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 647 + ], + "type": "text", + "content": "In this paper, we have presented a self-supervised vision-based 3D occupancy prediction method (SelfOcc) for autonomous driving, which predicts meaningful geometry and semantics from video sequences only. To facilitate feature extraction in the 3D space, we have proposed to condition NeRFs on 3D representations. Furthermore, we have designed an MVS-embedded strategy to boost depth optimization for NeRFs. SelfOcc has set new state-of-the-art for self-supervised vision-based occupancy prediction on SemanticKITTI and has predicted reasonable 3D occupancy for surrounding cameras on Occ3D for the first time." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 306, + 658, + 404, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 404, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 404, + 670 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62321005, Grant 62336004, and Grant 62125603." + } + ] + } + ], + "index": 61 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19953" + } + ] + } + ], + "index": 62 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "text", + "content": "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "type": "text", + "content": "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 287, + 327 + ], + "type": "text", + "content": "[5] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 372 + ], + "type": "text", + "content": "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 428 + ], + "type": "text", + "content": "[7] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 430, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 463 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 464, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 507 + ], + "type": "text", + "content": "[9] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "text", + "content": "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, 2020. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[11] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In ECCV, pages 680-697. Springer, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "text", + "content": "[12] Yueh-Tung Chen, Martin Garbade, and Juergen Gall. 3d semantic scene completion from a single depth image using adversarial training. In ICIP, pages 1835-1839. IEEE, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[13] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161. PMLR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "text", + "content": "[16] Aloisio Dourado, Hansung Kim, Teofilo E de Campos, and Adrian Hilton. Semantic scene completion from a single 360-degree image and depth map. In VISIGRAPP (5: VIS-APP), pages 36-46, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "text", + "content": "[17] Aloisio Dourado, Teofilo E De Campos, Hansung Kim, and Adrian Hilton. Edgenet: Semantic scene completion from a single rgb-d image. In ICPR, pages 503-510. IEEE, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 189, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 545, + 231 + ], + "type": "text", + "content": "[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501-5510, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 234, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 545, + 276 + ], + "type": "text", + "content": "[19] Wanshui Gan, Ningkai Mo, Hongbin Xu, and Naoto Yokoya. A simple attempt for 3d occupancy estimation in autonomous driving. arXiv preprint arXiv:2303.10076, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 280, + 545, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 313 + ], + "type": "text", + "content": "[20] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPR Workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 315, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 347 + ], + "type": "text", + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "text", + "content": "[22] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 384, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 437 + ], + "type": "text", + "content": "[23] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 441, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 474 + ], + "type": "text", + "content": "[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "text", + "content": "[25] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, pages 2485-2494, 2020. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 564 + ], + "type": "text", + "content": "[26] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. arXiv preprint arXiv:2104.00152, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "text", + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 666 + ], + "type": "text", + "content": "[28] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[29] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 1, 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19954" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "text", + "content": "[31] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 216 + ], + "type": "text", + "content": "[32] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "text", + "content": "[33] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "text", + "content": "[34] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "type": "text", + "content": "[35] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters, 5(1):219-226, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[36] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, 2020. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 413 + ], + "type": "text", + "content": "[37] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "text", + "content": "[38] Siqi Li, Changqing Zou, Yipeng Li, Xibin Zhao, and Yue Gao. Attention-based multi-modal fusion network for semantic scene completion. In AAAI, pages 11402-11409, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 502 + ], + "type": "text", + "content": "[39] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 557 + ], + "type": "text", + "content": "[40] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[41] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "text", + "content": "[42] Kai-En Lin, Yen-Chen Lin, Wei-Sheng Lai, Tsung-Yi Lin, Yichang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[44] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. NIPS, 31, 2018. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 127 + ], + "type": "text", + "content": "[46] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "text", + "content": "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 183, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 237 + ], + "type": "text", + "content": "[48] Norman Müller, Andrea Simonelli, Lorenzo Porzi, Samuel Rota Bulo, Matthias Nießner, and Peter Kontschieder. Autorf: Learning 3d object radiance fields from single view observations. In CVPR, pages 3971-3980, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "text", + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ToG, 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[50] Mong H Ng, Kaahan Radia, Jianfei Chen, Dequan Wang, Ionel Gog, and Joseph E Gonzalez. Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. arXiv preprint arXiv:2006.11436, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "type": "text", + "content": "[51] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "text", + "content": "[52] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "text", + "content": "[53] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 471, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 502 + ], + "type": "text", + "content": "[54] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "text", + "content": "[55] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, pages 6134-6144, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "text", + "content": "[56] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "text", + "content": "[57] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[58] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[59] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In ThreeDV, 2020. 2, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19955" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[60] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In ICCV, pages 3216-3226, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "text", + "content": "[61] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In ECCV, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "type": "text", + "content": "[62] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 194, + 287, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 287, + 238 + ], + "type": "text", + "content": "[63] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 281 + ], + "type": "text", + "content": "[64] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 284, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 326 + ], + "type": "text", + "content": "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 327, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 287, + 380 + ], + "type": "text", + "content": "[66] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 383, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 383, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 383, + 287, + 425 + ], + "type": "text", + "content": "[67] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In ICCV, pages 8608–8617, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 480 + ], + "type": "text", + "content": "[68] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 482, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 482, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 49, + 482, + 287, + 535 + ], + "type": "text", + "content": "[69] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, pages 539-549. PMLR, 2023. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 537, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 287, + 569 + ], + "type": "text", + "content": "[70] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 49, + 571, + 287, + 624 + ], + "type": "text", + "content": "[71] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9076-9086, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 625, + 287, + 667 + ], + "type": "text", + "content": "[72] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[73] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 643 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[74] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "text", + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[76] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In CVPR, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "text", + "content": "[77] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In ICCV, pages 1020-1031, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "type": "text", + "content": "[78] Jiahui Zhang, Hao Zhao, Anbang Yao, Yurong Chen, Li Zhang, and Hongen Liao. Efficient semantic scene completion network with spatial group convolution. In ECCV, pages 733-749, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[79] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "type": "text", + "content": "[80] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In ICCV, pages 7801-7810, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "text", + "content": "[81] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "type": "text", + "content": "[82] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 487, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 519 + ], + "type": "text", + "content": "[83] Min Zhong and Gang Zeng. Semantic point completion network for 3d semantic scene completion. In ECAI 2020, pages 2824-2831. IOS Press, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 521, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 574 + ], + "type": "text", + "content": "[84] Kaichen Zhou, Lanqing Hong, Changhao Chen, Hang Xu, Chaoqiang Ye, Qingyong Hu, and Zhenguo Li. Devnet: Self-supervised monocular depth learning via density volume construction. In ECCV, pages 125-142. Springer, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 610 + ], + "type": "text", + "content": "[85] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 611, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 643 + ], + "type": "text", + "content": "[86] Zhongkai Zhou, Xinnan Fan, Pengfei Shi, and Yuanxue Xin. R-MSFM: Recurrent multi-scale feature modulation for monocular depth estimating. In ICCV, 2021. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19956" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_content_list.json b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40592e9498f6cf75285d055e69bbc00839a22d56 --- /dev/null +++ b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_content_list.json @@ -0,0 +1,1560 @@ +[ + { + "type": "text", + "text": "SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vinkle Srivastav $^{1,2*}$ Keqi Chen $^{1*}$ Nicolas Padoy $^{1,2}$", + "bbox": [ + 261, + 179, + 702, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France \n2IHU Strasbourg, France", + "bbox": [ + 209, + 203, + 759, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "srivastav@unistra.fr keqi.chen@unistra.fr npadoy@unistra.fr", + "bbox": [ + 205, + 241, + 759, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 290, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a new self-supervised approach, SelfPose3d, for estimating 3d poses of multiple persons from multiple camera views. Unlike current state-of-the-art fully-supervised methods, our approach does not require any 2d or 3d ground-truth poses and uses only the multiview input images from a calibrated camera setup and 2d pseudo poses generated from an off-the-shelf 2d human pose estimator. We propose two self-supervised learning objectives: self-supervised person localization in 3d space and self-supervised 3d pose estimation. We achieve self-supervised 3d person localization by training the model on synthetically generated 3d points, serving as 3d person root positions, and on the projected root-heatmaps in all the views. We then model the 3d poses of all the localized persons with a bottleneck representation, map them onto all views obtaining 2d joints, and render them using 2d Gaussian heatmaps in an end-to-end differentiable manner. Afterwards, we use the corresponding 2d joints and heatmaps from the pseudo 2d poses for learning. To alleviate the intrinsic inaccuracy of the pseudo labels, we propose an adaptive supervision attention mechanism to guide the self-supervision. Our experiments and analysis on three public benchmark datasets, including Panoptic, Shelf, and Campus, show the effectiveness of our approach, which is comparable to fully-supervised methods. Code is available at https://github.com/CAMMA-public/SelfPose3D.", + "bbox": [ + 75, + 321, + 472, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 760, + 209, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of estimating 3d poses for multiple persons using a few calibrated cameras is a challenging computer vision problem [10, 18, 32, 52, 57]. A significant part of this challenge lies in identifying and matching the same person across different camera views. The solutions developed so far generally use one of the two paradigms: learning-based", + "bbox": [ + 75, + 785, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods and optimization-based methods. The learning-based methods develop novel deep-learning models and use 3d ground-truth poses for both training the models and establishing person correspondences across different views [39, 52, 57, 59, 63]. The accurate 3d ground-truth poses are typically generated using a dense camera system [32]. In contrast, the optimization-based methods formulate the 3d pose reconstruction as a mathematical optimization task, primarily focusing on aligning and matching the 2d poses across different camera views to infer 3d poses using triangulation within the framework of multi-view geometry [10, 18, 31-33, 49]. The 2d poses are estimated using off-the-shelf 2d human pose detectors [5, 55]. These methods apply geometric and spatial constraints in the optimization loop to ensure the anatomical plausibility and consistency of the inferred 3d poses. Although these methods do not require 3d ground-truth poses, their effectiveness is somewhat limited compared to the fully-supervised learning-based methods, see Table 1.", + "bbox": [ + 500, + 291, + 893, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the possibility of combining the strengths of both paradigms. Specifically, we investigate whether it's feasible to utilize a learning-based model for multi-view, multi-person 3D pose estimation and simultaneously eliminate its dependence on 3D ground-truth poses by incorporating geometric and appearance constraints, drawing inspiration from optimization-based methods.", + "bbox": [ + 496, + 580, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose, SelfPose3d, a self-supervised learning-based approach to estimate the 3d poses of multiple persons from a few calibrated cameras without using any 2d or 3d ground-truth poses. Our approach requires only 2d pseudo poses obtained using an off-the-shelf 2d pose detector [55]. Learning 3d poses without 3D ground-truth poses would require suitable supervisory signals to train a learning-based model. We follow the learning-by-projection paradigm, where the main idea is to learn the 3d output by comparing the projected bottleneck 3d output against the 2d input features [12].", + "bbox": [ + 496, + 686, + 893, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider VoxelPose [57] as a learning-based method and use its output 3d poses as a bottleneck representation. To recover the accurate underlying 3d poses, we propose us-", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* co-first authors with equal contributions.", + "bbox": [ + 94, + 886, + 316, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2502", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/121056602f4f7bb080107f9f99ed3f302a87fe053607a093d29e93a77534b294.jpg", + "image_caption": [ + "Figure 1. Our self-supervised approach, called SelfPose3d, estimates multi-person 3d poses from multi-view images and pseudo 2d poses generated using an off-the-shelf 2d human pose estimator. We propose a self-supervised learning objective that generates differentiable and geometrically constrained 2d joints and heatmaps across multiple views from bottleneck 3d poses. On the right, we show 3d pose outputs from our approach along with estimated body meshes (using SMPL body mesh fitting on 3d poses [4, 41]) and the projected 2d poses." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 423, + 339 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d23c63683cc97010ffee781c25bd2a8ac5e200461236303f50bb887f7d097456.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 89, + 870, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing differentiable multi-view 2d representations and cross-affine-view consistency. In particular, given a multi-view input image, we apply two random affine augmentations and pass them to the VoxelPose. It generates the bottleneck 3d poses corresponding to each affine augmented multi-view image. To enforce the model to learn and reason in the spatial dimension, we project the bottleneck 3d poses onto each view, obtaining 2d joints, and rendering them into spatial 2d heatmap representations in an end-to-end differentiable way. We further put tight geometric constraints by cross-affine-view operation, i.e. the bottleneck 3d poses from the 1st affine augmented multi-view image is mapped and rendered in the 2nd affine augmented multi-view image space and vice versa. Finally, we use the affine transformed 2d joints and heatmaps from the 2d pseudo poses to enable the geometrically constrained learning, with $L_{1}$ and $L_{2}$ losses respectively.", + "bbox": [ + 75, + 422, + 472, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the 2d pseudo poses contain non-negligible noises (mostly due to occlusions, see Figure 3), we propose adaptive supervision attention to guide our model to focus on more reliable regions. We apply two strategies towards $L_{1}$ joint loss and $L_{2}$ heatmap loss; for $L_{1}$ joint loss supervision, we employ hard attention, where we ignore the one view with the largest absolute error for each multi-view image set; for $L_{2}$ heatmap loss supervision, we employ soft attention using a lighter backbone to process each view, obtaining same-size attention heatmaps. During $L_{2}$ loss computation, we compute the element-wise product of the attention heatmaps and the square error before averaging. To avoid obtaining zero attention, which the model tends to do, we add a regularization term, where we create tensors of all", + "bbox": [ + 75, + 689, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ones as the attention heatmap labels and use $L_{2}$ loss as the attention loss.", + "bbox": [ + 498, + 422, + 890, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, specific to our choice of learning-based method, i.e., VoxelPose, which uses a voxel-based 3d root localization model to localize the persons in space using ground-truth 3d root joints (mid-hip joint), we use a simple but effective strategy to localize persons in space. Specifically, we randomly place 3d points in 3d world-space and project them to each view using the given camera parameters, subsequently rendering the projected 2d points as heatmap representations. This generates a synthetic dataset containing 3d points (roots) and their corresponding rendered multiview root-heatmaps. We then use this dataset to train a 3d root localization model that takes multi-view root-heatmaps as input and predicts the 3d roots as output. We further regularize the model by enforcing invariant constraints between pairs of affine augmented root-heatmaps coming from the real multi-view input.", + "bbox": [ + 496, + 455, + 892, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation on three 3d pose benchmarks datasets, Panoptic [32], Shelf [1] and Campus [1], along with extensive ablation studies on the Panoptic [32] dataset, show the effectiveness of our approach. Our approach reaches a performance comparable to learning-based fully-supervised approaches and performs significantly better than optimization-based approaches. Moreover, SMPL body mesh fitting [4, 41] on our estimated 3d poses generates geometrically plausible body shapes (see Figure 1 and Figure 4).", + "bbox": [ + 496, + 700, + 893, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows: 1) We address the challenging multi-person multi-view 3d person pose estimation problem using a self-supervised approach", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2503", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "without any 2d or 3d ground truth. 2) We propose self-supervised 3d pose estimation by using a new method to recover geometrically constrained 2d joints and heatmap representations from the bottleneck 3d poses. 3) We propose adaptive supervision attention to address the misinformation caused by the inaccurate pseudo labels. 4) We propose self-supervised 3d root localization to estimate the 3d root location utilizing synthetic 3d roots and the corresponding rendered multi-view root heatmaps.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 241, + 215, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we briefly review current works related to fully-supervised learning-based methods for 3d pose estimation, optimization-based methods for 3D pose estimation, and self-supervised learning.", + "bbox": [ + 75, + 266, + 468, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fully-supervised methods: Monocular 3d pose estimation [22, 43, 45, 47, 50, 56, 62, 65] is an ill-posed problem due to depth ambiguities as multiple 3d poses can produce same 2d pose projection. Having access to multi-view cameras can alleviate such depth ambiguities achieving the state-of-art results on benchmark datasets [23, 27, 28, 39, 51, 52, 57, 59, 63]. For single-person scenes, these approaches exploit multi-view geometry [24] to either fuse the visual features [27, 51], perform triangulation on heatmaps [28, 53], or use pictorial structural models for 3d reconstruction [48, 51]. The multi-person scene offers extra complexity due to the variability in the number of person in each view and the unknown cross-view correspondence. Existing multi-person multi-view approaches are based on volumetric paradigm [52, 57, 64], or direct regression [63] based on transformers [6, 58, 66]. Despite their good performance, these approaches rely on ground-truth 3d poses, which are generated using dense camera systems [32].", + "bbox": [ + 75, + 327, + 472, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization-based 3d pose estimation: For the multi-person and multi-view scenario, optimization-based approaches use an off-the-shelf person-id detector across all the views to solve the correspondence and triangulation problem and temporal refinement along with training a reinforcement learning agent to find the best camera locations for 3d pose reconstruction [49]. More recent approaches utilize multi-view 3d reconstruction in the optimization loop inferring 3D poses that are geometrically and spatially coherent [10, 18, 33].", + "bbox": [ + 75, + 598, + 472, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised learning: Self-supervised learning can be broadly classified into self-supervised representation learning and self-supervised task learning. Self-supervised representation learning aims to use large-scale unlabeled data to learn generic feature representations. The recent promising results from these approaches have started to surpass the fully-supervised baselines for various downstream tasks [7, 11, 26, 54]. Self-supervised task learning aims to learn a particular downstream task without using ground truth labels and has been applied to 2d pose estimation [29, 30],", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "single-person 3d pose estimation [9, 19, 34, 36, 38], and surface correspondences estimation [3]. Self-supervised approaches for 3d pose estimation have primarily been developed for single-person scenarios. Given 2d poses, estimated by utilizing advances in the 2d pose estimation methods [5, 13, 14, 21, 35, 42, 44, 46, 55, 60], these approaches use the supervisory signals generated from multi-view geometry [34], video constraints [38], or adversarial learning [9, 19, 36].", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work proposes a learning-based approach to model the 3d poses as bottleneck representations and recover geometrically constrained and spatially accurate 2d joints and heatmap representations in an end-to-end differentiable manner.", + "bbox": [ + 496, + 229, + 893, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 327, + 635, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem overview", + "text_level": 1, + "bbox": [ + 498, + 354, + 674, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a training dataset of multi-view images $\\mathcal{D} = \\{\\mathbf{x}|\\mathbf{y}^*\\}$ where $\\mathbf{x} \\in \\mathcal{R}^{C \\times 3 \\times H \\times W}$ is a multi-view image set from $C$ cameras with height $H$ and width $W$ , and $\\mathbf{y}^* \\in \\mathcal{R}^{C \\times P \\times J \\times 2}$ represents the 2d pseudo poses for $P$ persons with $J$ joints, the goal is to learn a deep learning model that estimates the 3d poses $\\mathcal{V} \\in \\mathcal{R}^{P \\times J \\times 3}$ of all the $P$ persons from the multi-view input image $\\mathbf{x}$ . It is to be noted that $P$ can vary in each camera view due to occlusion and noisy pseudo 2d pose estimation. For simplicity in the notation, we keep the same variable $\\mathbb{P}$ .", + "bbox": [ + 496, + 381, + 893, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fully-supervised approaches rely on 3d ground-truth poses, while we only have 2d pseudo poses $\\mathbf{y}^*$ . Therefore, after obtaining 3d poses $\\mathcal{V} \\in \\mathcal{R}^{P \\times J \\times 3}$ following traditional approach, we propose to project the poses to each view obtaining 2d poses $\\mathbf{y} \\in \\mathcal{R}^{C \\times P \\times J \\times 2}$ , and train the model from 2d pseudo poses $\\mathbf{y}^* \\in \\mathcal{R}^{C \\times P \\times J \\times 2}$ .", + "bbox": [ + 496, + 536, + 893, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we present our self-supervised approach based on fully-supervised VoxelPose [57]. We first generate pseudo 2d poses, and then propose self-supervised 3d root localization, self-supervised 3d pose estimation, and an adaptive supervision attention to learn 3d poses in a self-supervised manner, without modifying the original Voxel-Pose structure.", + "bbox": [ + 496, + 628, + 893, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Generating pseudo 2d poses", + "text_level": 1, + "bbox": [ + 498, + 753, + 751, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To circumvent the dependence on the ground-truth 2d poses, we generate the 2d pseudo poses on the training dataset using Mask R-CNN [25] to generate person bounding boxes followed by using HRNet [55] to generate 2d poses of each detected person bounding box. The two-stage approach is chosen based on its state-of-art performance on the COCO dataset [40]. We first pre-train the 2d CNN backbone heatmap_net $_{2d}$ with pseudo 2d poses.", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2504", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/15683e8d594f83f5225fd1681555b2ce1e29c3713f4977fe874c63317b726872.jpg", + "image_caption": [ + "SelfPose3d Framework", + "Figure 2. Illustrating our self-supervised SelfPose3d approaches for multi-view multi-person 3d pose estimation. Instead of using ground-truth 3d poses for learning, we propose self-supervised learning objectives to localize 3d roots (mid-hip location of the person) and estimate their 3d poses. We utilize a synthetic 3d roots dataset, two different affine transformations on the multi-view input images $(t_{r,s}^{1}, t_{r,s}^{2})$ parametrized by rotation $r$ and scale $s$ , a differentiable cross-affine-view 2d joints and heatmaps rendering from the bottleneck 3d poses, and an adaptive supervision attention mechanism to automatically learn the 3d poses in world-space." + ], + "image_footnote": [], + "bbox": [ + 86, + 101, + 885, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Self-supervised 3d root localization", + "text_level": 1, + "bbox": [ + 76, + 462, + 382, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given 2d multi-view heatmaps from all the views and all the joints $\\mathcal{H} \\in \\mathcal{R}^{C \\times J \\times \\frac{H}{4} \\times \\frac{W}{4}}$ estimated using a 2d backbone model heatmap。 $\\mathrm{net}_{2\\mathrm{d}}$ , we use [28] to construct a discretized 3d feature volume $\\mathcal{F} \\in \\mathcal{R}^{J \\times X \\times Y \\times Z}$ for each joint by un-projecting the 2d multi-view heatmaps to 3d space:", + "bbox": [ + 75, + 486, + 470, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {\\text {u n p r o j}} (\\operatorname {c a m}, \\text {c e n t e r}, t _ {r, s}): \\mathcal {H} \\longrightarrow \\mathcal {F}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 575, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To localize persons' root (mid-hip) joint in 3d space without using 3d ground truth, we hypothesize that 2d multi-view heatmaps of the root location $\\mathcal{H}_{\\mathrm{root}} \\in \\mathcal{R}^{C \\times \\frac{H}{4} \\times \\frac{W}{4}}$ are sufficient for 3d root localization (see Sec. 7.6 for verification). Then we generate the 3d feature volume for the root location $\\mathcal{F}_{\\mathrm{root}} \\in \\mathcal{R}^{X \\times Y \\times Z}$ using $\\mathcal{H}_{\\mathrm{root}}$ using Eq. (1), which has the same dimensions as predicted root-volumes $\\mathcal{G}$ . This allows us to establish a one-to-one relationship between 3d root-volumes $\\mathcal{G}$ and 2d multi-view root-heatmaps $\\mathcal{H}_{\\mathrm{root}}$ . We generate a synthetic root dataset $\\mathcal{D}_{\\mathrm{root}} = \\{\\mathcal{G}_i^{\\mathrm{syn*}} | \\mathcal{H}_{\\mathrm{root\\_i}}^{\\mathrm{syn}}\\}_{i=1}^N$ where $\\mathcal{G}_i^{\\mathrm{syn*}} \\in \\mathcal{R}^{X \\times Y \\times Z}$ contains the root-volumes of randomly placed 3d points, and $\\mathcal{H}_{\\mathrm{root\\_i}}^{\\mathrm{syn}}$ is the corresponding 2d multi-view heatmaps generated by projecting the random 3d points to each view using camera parameters cam. After unprojecting $\\mathcal{H}_{\\mathrm{root}}^{\\mathrm{syn}}$ to $\\mathcal{F}_{\\mathrm{root}}^{\\mathrm{syn}}$ , and passing it through root\\_net obtaining $\\mathcal{G}_{\\mathrm{root}}^{\\mathrm{syn}}$ , we compute the $L_2$ loss error as the synthetic root loss $l_{\\mathrm{root\\_syn}}$ :", + "bbox": [ + 75, + 607, + 472, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {r o o t - s y n}} = \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {\\text {s y n}}, \\mathcal {G} ^ {\\text {s y n *}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further regularize root_net on the real-world 2d multi-view root-heatmaps, we propose the root consistency loss. Given a multi-view training image set $x^0$ , we apply two affine transformations $(t_{r,s}^{1}, t_{r,s}^{2})$ with random rotation and scaling $(r, s)$ to generate two affine transformed multiview images $(x^1, x^2)$ . We pass $x^0$ , $x^1$ and $x^2$ through heatmap_net2d, construct the root feature volumes using Eq. (1) with corresponding affine transformation parameters, and finally obtain $\\mathcal{G}^0$ , $\\mathcal{G}^1$ and $\\mathcal{G}^2$ through root_net. Since $\\mathcal{G}^1$ and $\\mathcal{G}^2$ are invariant to the applied affine transformations $t_{r,s}^1$ and $t_{r,s}^2$ , we use $\\mathcal{G}^0$ as the baseline to compute the $L_2$ loss error between $\\mathcal{G}^0$ , $\\mathcal{G}^1$ and $\\mathcal{G}^2$ as the root consistency loss $l_{\\mathrm{root\\_C}}$ :", + "bbox": [ + 496, + 463, + 890, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {r o o t} C} = \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {0}, \\mathcal {G} ^ {1}\\right) + \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {0}, \\mathcal {G} ^ {2}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 672, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train root_net by minimizing Eq. (2) and Eq. (3). We generate the person proposals $\\{\\mathrm{root}_i\\}_{i=1}^K$ by applying nonmaximum suppression (NMS) and thresholding on $\\mathcal{G}^2$ ( $\\mathcal{G}^1$ would also work).", + "bbox": [ + 496, + 694, + 890, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Self-supervised 3d pose estimation", + "text_level": 1, + "bbox": [ + 500, + 762, + 799, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given pseudo 2d poses $\\mathbf{y}_{2\\mathrm{d}}^{*}$ , person proposals $\\{\\mathrm{root}_i\\}_{i = 1}^P$ and 2d multi-view heatmaps $\\mathcal{H}^1,\\mathcal{H}^2$ predicted using heatmap_net2d, we describe our self-supervised 3d pose estimation approach.", + "bbox": [ + 496, + 784, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The person proposals $\\{\\mathrm{root}_i\\}_{i = 1}^P$ are used to generate the 3d feature volumes i.e $\\left\\{\\mathcal{F}_i^1\\right\\}_{i = 1}^P = \\left\\{\\mathcal{P}_{\\mathrm{unproj}}(\\mathrm{cam},r_i,t_{r,s}^1)\\right\\}_{i = 1}^P$ and $\\left\\{\\mathcal{F}_i^2\\right\\}_{i = 1}^P =$", + "bbox": [ + 496, + 845, + 893, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2505", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\left\\{\\mathcal{P}_{\\mathrm{unproj}}(\\mathrm{cam},r_i,t_{r,s}^2)\\right\\}_{i = 1}^P$ corresponding to the person feature volumes for each affine augmented multi-view input image $x^{1}$ and $x^{2}$ , respectively. $\\left\\{\\mathcal{F}_i^1\\right\\}_{i = 1}^P$ and $\\left\\{\\mathcal{F}_i^2\\right\\}_{i = 1}^P$ are passed through pose_net3d and soft-argmax [8] to estimate the 3d poses $\\mathcal{V}^1,\\mathcal{V}^2\\in \\mathcal{R}^{P\\times J\\times 3}$ . These 3d poses serve as a bottleneck representation.", + "bbox": [ + 76, + 88, + 467, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the camera parameters cam and the affine transformation parameters $t_{r,s}^{1}, t_{r,s}^{2}$ , we project the bottleneck 3d poses in cross-affine-view i.e. $\\mathcal{V}^1$ are projected to $x^2$ image space using $t_{r,s}^{2}$ to generate multi-view 2d poses $\\mathbf{y}^2 \\in \\mathcal{R}^{C \\times P \\times J \\times 2}$ , and $\\mathcal{V}^2$ are projected to the $x^1$ image space using $t_{r,s}^{1}$ to generate multi-view 2d poses $\\mathbf{y}^1 \\in \\mathcal{R}^{C \\times P \\times J \\times 2}$ .", + "bbox": [ + 76, + 191, + 467, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose to render $\\mathbf{y}^1$ and $\\mathbf{y}^2$ 2d poses in the 2d heatmap representation. The heatmap representation encodes the per-pixel likelihood of a body joint and has been a vital component to enable the state-of-the-art 2d pose estimation approaches [55]. Generating heatmap representation has essentially been a pre-processing step where state-of-the-art approaches quantize the 2d joints before generating the heatmap [55]. However, this quantization step is non-differentiable and could cut the backward gradient flow. Zhang et al. [61] show that encoding floating point 2d joints into heatmap representation in their pre-processing step leads to improved performance. We propose to use the same differentiable approach in an online way to render the projected 2d joints into the heatmap representation.", + "bbox": [ + 75, + 284, + 467, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We render $\\mathbf{y}^1$ and $\\mathbf{y}^2$ into heatmap representation to generate 2d multi-view heatmaps $\\mathcal{H}^1$ , and $\\mathcal{H}^2$ , respectively. We apply the affine transformations $t_{r,s}^1, t_{r,s}^2$ on the pseudo 2d poses $\\mathbf{y}_{2\\mathrm{d}}^{*}$ to generate pseudo 2d multi-view joints $\\mathbf{y}_{2\\mathrm{d}}^{1*}, \\mathbf{y}_{2\\mathrm{d}}^{2*}$ and heatmaps $\\mathcal{H}^{1*}, \\mathcal{H}^{2*}$ . Then, we compute the $L_2$ loss between heatmaps as the pose heatmap loss $l_{\\mathrm{pose\\_H}}$ :", + "bbox": [ + 75, + 494, + 467, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {p o s e} \\mathrm {H}} = \\mathcal {L} _ {2} \\left(\\mathcal {H} ^ {1}, \\mathcal {H} ^ {1 *}\\right) + \\mathcal {L} _ {2} \\left(\\mathcal {H} ^ {2}, \\mathcal {H} ^ {2 *}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 599, + 467, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training with $l_{\\mathrm{pose\\_H}}$ preliminarily, we propose to add the $L_{1}$ loss between 2d joints to further fine-tune the model. For each view, we employ the Hungarian algorithm [37] to obtain the optimal assignment between $\\mathbf{y}$ and $\\mathbf{y}_{2\\mathrm{d}}^{*}$ , where the matching cost is the mean absolute error. Based on the optimal assignment, we obtain the $L_{1}$ loss as $l_{\\mathrm{pose\\_J}}$ . Then we use $l_{\\mathrm{pose\\_H}}$ and $l_{\\mathrm{pose\\_J}}$ together to train the whole network, where $\\lambda$ is a manually defined weight:", + "bbox": [ + 75, + 626, + 467, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {p o s e - J}} = \\mathcal {L} _ {1} \\left(\\mathbf {y} ^ {1}, \\mathbf {y} _ {2 \\mathrm {d}} ^ {1 *}\\right) + \\mathcal {L} _ {1} \\left(\\mathbf {y} ^ {2}, \\mathbf {y} _ {2 \\mathrm {d}} ^ {2 *}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 760, + 467, + 779 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {p o s e - 3 d}} = l _ {\\text {p o s e - H}} + \\lambda l _ {\\text {p o s e - J}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 801, + 467, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the network needs to reason about the 2d joint locations in spatial dimension, it implicitly solves the person correspondence problem. Training pose_net3d with 3d pose loss $l_{\\mathrm{pose\\_3d}}$ performs decently, but to achieve even better results, we introduce the adaptive supervision attention.", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27a3d8b7034de1a41fcf05a0ea8360107c945bc08060162bc16a013846740d43.jpg", + "image_caption": [ + "ground truth 2d poses", + "Figure 3. Comparing ground-truth 2d poses generated by projecting the ground-truth 3d poses to each multi-view image and our pseudo 2d poses generated by running HRNet human pose estimation model [55] on the training dataset. Pseudo 2d poses contain localization errors due to occlusion (see the red arrows), and ground-truth 2d poses exist for partially or even entirely occluded persons (see the blue dotted arrows)." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 694, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec0de6ec05ddac0d1a71ef0f200a5a00a5df47cf062f8753ec123a2a6907381b.jpg", + "image_caption": [ + "pseudo 2d poses" + ], + "image_footnote": [], + "bbox": [ + 694, + 89, + 887, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Adaptive supervision attention", + "text_level": 1, + "bbox": [ + 500, + 393, + 769, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional $L_{1}$ and $L_{2}$ losses treat each label equally, which is sub-optimal in two aspects: (1) the 2d human pose detector may generate inaccurate labels due to occlusions (see the red arrows in Figure 3); (2) the 3d-2d projection will output 2d joints in certain views even when the person is entirely occluded (see the blue dotted arrows in Figure 3). Therefore, we propose to employ attentions to adaptively guide the supervision process.", + "bbox": [ + 496, + 415, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $L_{2}$ loss supervision, we use the soft attention. Specifically, we use ResNet-18 to extract the visual features of the views, followed by deconvolutional layers to obtain the attention heatmaps $\\mathcal{A}$ (see attn\\_net $_{2d}$ in Figure 2). Then we compute the element-wise product of $\\mathcal{A}$ and the square error before averaging, as the new loss $l_{\\mathrm{pose\\_H}}^{\\mathrm{attn}}$ :", + "bbox": [ + 496, + 537, + 890, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {p o s e - H}} ^ {\\mathrm {a t t n}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {A} _ {i} \\otimes \\left(\\mathcal {H} _ {i} - \\mathcal {H} _ {i} ^ {*}\\right) ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 638, + 890, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To avoid that $\\mathcal{A}$ becomes zero, we add a regularization term. We create tensors of all ones $\\mathbb{1}$ as the attention heatmap labels, and compute $L_{2}$ error as the attention loss $l_{\\mathrm{attn}}$ . If $l_{\\mathrm{attn}}$ becomes zero, $l_{\\mathrm{pose\\_H}}^{\\mathrm{attn}}$ degrades to non-attentive version.", + "bbox": [ + 496, + 683, + 890, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\mathrm {a t t n}} = \\mathcal {L} _ {2} (\\mathcal {A}, \\mathbb {1}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 773, + 890, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $L_{1}$ loss supervision, we use the hard attention. For each input with $K$ views, we compute the $L_{1}$ loss of each view, find the view with the largest loss, and ignore it when averaging the final loss $l_{\\mathrm{pose\\_J}}^{\\mathrm{att}}$ :", + "bbox": [ + 496, + 794, + 890, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nj = \\underset {i} {\\arg \\max } \\mathcal {L} _ {1} \\left(\\mathbf {y} _ {i}, \\mathbf {y} _ {i, 2 \\mathrm {d}} ^ {*}\\right) \\quad (i = 1, 2, \\dots , K) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 878, + 890, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2506", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {p o s e - J}} ^ {\\mathrm {a t t n}} = \\frac {1}{K - 1} \\sum_ {i = 1, i \\neq j} ^ {K} \\mathcal {L} _ {1} \\left(\\mathbf {y} _ {i}, \\mathbf {y} _ {i, 2 \\mathrm {d}} ^ {*}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 102, + 468, + 145 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In general, the final 3d pose loss $l_{\\mathrm{pose\\_3d}}^{\\mathrm{attn}}$ is as follows, where $\\lambda$ and $\\sigma$ are manually defined weights:", + "bbox": [ + 76, + 154, + 468, + 185 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {p o s e - 3 d}} ^ {\\text {a t t n}} = l _ {\\text {p o s e - H}} ^ {\\text {a t t n}} + \\lambda l _ {\\text {p o s e - J}} ^ {\\text {a t t n}} + \\sigma l _ {\\text {a t t n}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 199, + 468, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train pose_net3d by minimizing $l_{\\mathrm{pose\\_3d}}^{\\mathrm{attn}}$ . Our self-supervised approach is visually described in Figure 2.", + "bbox": [ + 76, + 224, + 468, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 265, + 290, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training strategies For the Panoptic dataset, similar to VoxelPose [57], we first train heatmap_net2d for 20 epochs with pseudo 2d poses. We use the Adam optimizer with an initial learning rate of 1e-4, which decreases to 1e-5 and 1e-6 at the 10th and 15th epochs, respectively. Then, we train the root_net for 1 epoch, followed by end-to-end joint training of the whole network for 5 epochs using only the $L_{2}$ loss, with a learning rate of 1e-4. Afterwards, we add $L_{1}$ loss to train the whole network for another 5 epochs with a learning rate of 5e-5. $\\lambda$ and $\\sigma$ in Eq. (11) are set to 0.01 and 0.1 respectively.", + "bbox": [ + 76, + 287, + 468, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the random rotation between $-45^{\\circ}$ to $45^{\\circ}$ and random scale between $-0.35$ to 0.35. We also apply spatial augmentations using rand-augment [16] and rand-cutout [17] using python image library1. The rand-augment consists of \"contrast-jittering\", \"auto-contrast\", \"equalize\", \"color-jittering\", \"sharpness-jittering\", and \"brightness-jittering\", and the rand-cutout places random square boxes of sizes between 20 to 40 pixels at random locations in the image. We use the SMPL model and optimization-based body fitting approach2[4, 41] to estimate body mesh parameters.", + "bbox": [ + 75, + 455, + 470, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference pipeline During inference, we input the multiview RGB images, and obtain the estimated 3d poses in an end-to-end pipeline. For each view, the backbone generates corresponding 2d heatmaps for cuboid construction. Then, given constructed cuboid of the whole space, the root_net predicts root joint locations of all persons. Finally, the pose_net outputs the regressed 3d locations of each joint for every cuboid proposal of the root joints.", + "bbox": [ + 75, + 626, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and evaluation metrics", + "text_level": 1, + "bbox": [ + 76, + 786, + 356, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on three benchmark datasets: Panoptic [32], Campus [1], and Shelf [1].", + "bbox": [ + 76, + 810, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/01fa1677e36f8bbff62ce3dc00577a0bdcf2b01ed13fc81549f144a2b08d0953.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAP25AP50AP100AP150Recall@500MPJPE[mm]
FSVoxelPose [57]83.698.399.899.998.817.7
Lin et al. [39]92.199.099.899.8-16.8
MvP [63]92.396.697.597.798.215.8
Wu et al. [59]----98.715.8
TEMPO [15]89.099.199.899.9-14.7
OBACTOR [49]-----168.4
MvPose [18]0.02.9759.9381.5398.2384.2
SSSelfPose3d (ours)55.196.498.599.099.624.5
", + "bbox": [ + 506, + 89, + 885, + 190 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1d675c3ab67e79a6b3e785ef9e07b79cb86e34ac279e618b15a4decb07f7e479.jpg", + "table_caption": [ + "Table 1. Result on the Panoptic dataset (FS = fully-supervised, OB = optimization-based, SS = self-supervised)." + ], + "table_footnote": [], + "table_body": "
MethodsShelfCampus
Actor 1Actor 2Actor 3AverageActor 1Actor 2Actor 3Average
FSErshadi et al. [20]93.375.994.888.094.292.984.690.6
Wu et al. [59]99.396.597.397.7----
MvP [63]99.395.197.897.498.294.197.496.6
VoxelPose [57]99.394.197.697.097.693.898.896.7
VoxelPose* [57]99.593.597.896.993.186.593.290.9
OB3DPS [2]75.369.787.677.593.575.784.484.5
MvPose [18]98.894.197.896.997.693.398.096.3
SSSelfPose3d97.290.397.995.192.582.289.287.9
", + "bbox": [ + 504, + 244, + 888, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results (in PCP) on Shelf and Campus datasets (FS = fully-supervised, OB = optimization-based, SS = self-supervised, * = reproduced results). SelfPose3d is trained from the pseudo 3d poses from the Panoptic training set.", + "bbox": [ + 498, + 352, + 892, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Panoptic dataset is a large-scale dataset captured inside a dome environment containing multiple persons performing daily social activities. We conduct extensive experiments on this dataset to evaluate and assess various components of our approach. We use the same data sequences for training and testing as VoxelPose [57] except that our training set doesn't include '160906_band3'. In other words, we are only using 9 multi-view videos for training (the '160906_band3' video is not available due to the broken images on the source website). We use the five HD camera images (3, 6, 12, 13, 23) to train and report the performance in our experiments. We use Average Precision (AP), Recall, and Mean Per Joint Position Error (MPJPE) in millimeters (mm) as evaluation metrics (higher AP and lower MPJPE are better) [57].", + "bbox": [ + 498, + 436, + 890, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Shelf and Campus are two multi-person datasets capturing activities in the indoor and outdoor environments, respectively [1]. We use the same training and test split as [1, 57]. As used in [1, 57], we use the Percentage of Correct Parts (PCP) as evaluation metrics for these two datasets.", + "bbox": [ + 498, + 665, + 890, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Panoptic Table 1 shows the results on the challenging Panoptic dataset. All the fully-supervised approaches utilizing 2d and 3d ground-truth 3d poses reach nearly the same performance. SelfPose3d, without using any 3d or 2d ground-truth poses, achieves comparable results to fully-supervised approaches. Nevertheless, there still exists a gap compared to the fully-supervised VoxelPose model (96.4 v.s 98.3 AP50 and 24.5 v.s 17.7 MPJPE). However, unlike VoxelPose, which relies on heat maps from all the joints to es", + "bbox": [ + 496, + 763, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https : / / github . com / jizongFox / pytorch - randaugment 2https : // github . com / JiangWenPL / multiperson/ tree/master/misc/smplify-x", + "bbox": [ + 76, + 849, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2507", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/284dc7761ab5c730b5f21a0f5c57b543155f46c8bf22803f2fd71a6d2ac22fa4.jpg", + "image_caption": [ + "Figure 4. Qualitative results for the 3d pose estimations, 2d projections on the multi-view images, and estimated SMPL body shapes on some example images from the Panoptic dataset" + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 851, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "timate 3d roots, we only use root-heatmaps to do the same. This results in the reduction of the input channel from 15 (number of keypoints) to 1 for the root_net, making our approach computationally faster.", + "bbox": [ + 75, + 491, + 470, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare our approach with optimization-based baselines from Pirinen et al. [49] and Dong et al. [18]. These non-learning-based approaches fail to capture the multi-person interaction in a complex scene from a few sparse multi-view cameras. Our learning-based self-supervised approach achieves much better performance. It is to be noted that Pirinen et al. evaluate their approach on two multi-person sequences, whereas we evaluate on four multi-person sequences.", + "bbox": [ + 75, + 556, + 470, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shelf and Campus We compare our approach with the state-of-the-art methods on the Shelf and Campus dataset. VoxelPose uses the 3d ground-truth from the Panoptic dataset to train their approach to these datasets due to noisy and incomplete 3d ground-truth poses. For a fair comparison with VoxelPose, we use the pseudo 3d poses (by running SelfPose3d on the Panoptic training set) and train on these two datasets in a fully supervised manner. As shown in Table 2, our approach using pseudo 3d poses from the Panoptic dataset also reaches the same performance as the fully-supervised approaches.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative visualizations Figure 4 shows 3d pose estimation results from our SelfPose3d approach on the challenging Panoptic dataset. Without using any 3d ground-truth, we can see that SelfPose3d is robust to occlusions and multiple persons while correctly identifying the person identities across all the views (see the corresponding 2d projections in Figure 4). We also show the qualitative results for the SMPL body mesh fitting [4, 41] on the estimated 3d poses. All these results demonstrate both the effectiveness and extendability of SelfPose3d. Please see the supplementary for more results.", + "bbox": [ + 496, + 491, + 893, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 498, + 666, + 661, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ground-truth 2d poses v.s pseudo 2d poses As shown in Table 3, when we use the ground-truth 2d poses in our self-supervised framework, 3d reconstruction error significantly reduces. To inspect the better performance when using the ground-truth 2d poses, we qualitatively compare the ground-truth 2d poses with the pseudo 2d poses on some example training images. Pseudo 2d poses contain localization errors due to occlusion, whereas ground-truth 2d poses exist for partially or even entirely occluded persons as shown in Figure 3. As the ground-truth 2d poses are generated by projecting the ground-truth 3d poses to each multi-view image, they serve as a suitable proxy for the 3d poses, thereby reaching a performance close to the fully-supervised approaches. However, obtaining the ground 2d", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2508", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fa796cd60a174944608efff4f7e1758eaf5d265b6f63f162c69261dfaeba2d91.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
2d posesAP50AP100MPJPE
ground-truth98.899.619.9
pseudo96.498.524.5
", + "bbox": [ + 171, + 88, + 375, + 132 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6bb6a7b22feb9503eaecc463e56f2304baa7a0560d17cd018a4c9e6db0da7725.jpg", + "table_caption": [ + "Table 3. The ground-truth 2d poses in our self-supervised framework decrease the 3d reconstruction error and reach the performance close to the fully-supervised approaches." + ], + "table_footnote": [], + "table_body": "
cross-affine-view consistencyaffine augsAP50AP100MPJPE
86.096.234.7
✓83.397.533.3
✓✓93.898.129.3
", + "bbox": [ + 101, + 199, + 444, + 253 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6fba851443d10ec557b624aef444c76522e244a68dc5d74336d4ca920555ee7b.jpg", + "table_caption": [ + "Table 4. Affine augmentations and cross-affine-view consistency significantly improves the 3d pose reconstruction accuracy. All three models are trained for two epochs with frozen backbone and frozen root_net and no attention." + ], + "table_footnote": [], + "table_body": "
L1 lossL2 lossAP25AP50AP100MPJPE
✓----
✓43.895.898.225.7
✓✓55.196.498.524.5
", + "bbox": [ + 140, + 335, + 406, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "poses in this way would be as challenging as acquiring the ground-truth 3d poses.", + "bbox": [ + 76, + 455, + 467, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of cross-affine-view consistency and affine augmentations We also examine the effect of affine augmentations on the multi-view images and cross-affine-view consistency when generating differentiable 2d representations from the bottleneck 3d poses. As shown in Table 4, the affine augmentations and cross-affine-view consistency significantly improve the 3d pose reconstruction as they provide necessary geometric constraints during training.", + "bbox": [ + 75, + 507, + 467, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of $L_{2}$ and $L_{1}$ pose losses We conduct experiments to analyze the use of $L_{1}$ and $L_{2}$ pose losses. As shown in Table 5, using $L_{2}$ and $L_{1}$ losses together can obtain better results than using $L_{2}$ loss solely. Also, using $L_{1}$ loss solely doesn't converge due to the label noises.", + "bbox": [ + 76, + 648, + 467, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of adaptive supervision attention We also examine the necessity of adaptive supervision attention. Table 6 shows that supervision attention for both $L_{1}$ and $L_{2}$ losses are necessary for training.", + "bbox": [ + 76, + 744, + 467, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of different 2d human pose estimation models Finally, we show how pseudo 2d poses generated from different 2d human pose estimation models affect the performance. As shown in Table 7, models that perform well on the COCO dataset [40] also generate better pseudo 2d poses", + "bbox": [ + 76, + 824, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/231b8c119ee1243b5556c3da3363028f9930fd62b854130db2b59d8957441b7e.jpg", + "table_caption": [ + "Table 5. Training using $L_{1}$ and $L_{2}$ pose losses together achieves the best performance." + ], + "table_footnote": [], + "table_body": "
L1 loss attentionL2 loss attentionAP25AP50AP100MPJPE
32.594.197.828.5
✓37.995.898.026.3
✓47.496.698.225.0
✓✓55.196.498.524.5
", + "bbox": [ + 509, + 89, + 883, + 157 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ac36886afdb7d5f0aa02745cb0acdec838b7cf834f55f02c6970c1a0490658c5.jpg", + "table_caption": [ + "Table 6. Training using $L_{1}$ and $L_{2}$ loss supervisions together achieves the best performance." + ], + "table_footnote": [], + "table_body": "
Method for 2d pseudo pose generationAP50AP100MPJPEKeypoint AP on COCO-val[40]
Keypoint R-CNN (R-101) [25]89.297.631.966.1
HRNet (w48 384x288) [55]93.898.129.376.3
", + "bbox": [ + 521, + 210, + 870, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Comparing different models for generating pseudo 2d poses. Models that perform well on the COCO dataset [40] also generate better pseudo 2d poses for the Panoptic dataset, helping SelfPose3d to achieve better performance.", + "bbox": [ + 498, + 272, + 890, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for the Panoptic dataset, helping SelfPose3d to achieve better performance.", + "bbox": [ + 500, + 354, + 888, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 398, + 617, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a self-supervised approach, called SelfPose3d, to address the challenging problem of multi-view multiperson 3d human pose estimation. Unlike current state-of-the-art methods that use difficult-to-acquire 3d ground-truth poses to train the model, SelfPose3d requires only multiview input images and an off-the-shelf 2d human pose detector. We propose a novel self-supervised learning objective that aims to recover 2d joints and heatmaps under different affine transformations from the bottleneck 3d poses. We further improve the performance of our approach by integrating adaptive supervision attention to address the misinformation caused by the inaccurate 2d pseudo labels from the off-the-shelf 2d human pose detector. We conduct extensive experiments on large-scale benchmark datasets, assess various components of our approach, and show that SelfPose3d reaches a performance on par with the well-established fully-supervised baselines. We visualize the 3d pose reconstruction in the complex multiple-person scenes and show that body shape meshes fitted on the estimated 3d poses look geometrically plausible under different viewpoints.", + "bbox": [ + 496, + 422, + 890, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 753, + 686, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by French state funds managed by the ANR under references ANR-20-CHIA0029-01 (National AI Chair AI4ORSafety), ANR-10IAHU-02 (IHU Strasbourg), ANR-18-CE45-0011-03 (OptimiX), and by BPI France (project 5G-OR). This work was also granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3].", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2509", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures for multiple human pose estimation. In CVPR, pages 1669-1676, 2014. 2, 6", + "[2] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. TPAMI, 38(10):1929-1942, 2015. 6", + "[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 3", + "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578. Springer, 2016. 2, 6, 7", + "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, pages 7291-7299, 2017. 1, 3", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3", + "[8] Olivier Chapelle and Mingrui Wu. Gradient descent optimization of smoothed information retrieval metrics. Information retrieval, 13(3):216-235, 2010. 5", + "[9] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, M. V. Rohith, Stefan Stojanov, and James M. Rehg. Unsupervised 3d pose estimation with geometric self-supervision. CoRR, abs/1904.04812, 2019. 3", + "[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. In European Conference on Computer Vision, pages 541-557. Springer, 2020. 1, 3", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3", + "[12] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 1", + "[13] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020. 3", + "[15] Rohan Choudhury, Kris M Kitani, and László A Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14750-14760, 2023. 6", + "[16] Ekin D Cubuk, Barret Zoph, Jonathon Shiens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6", + "[17] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6", + "[18] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In CVPR, pages 7792-7801, 2019. 1, 3, 6, 7", + "[19] Dylan Drover, Ching-Hang Chen, Amit Agrawal, Ambrish Tyagi, and Cong Phuoc Huynh. Can 3d pose be learned from 2d projections alone? In Proceedings of the European Conference on Computer Vision (ECCV), pages 0–0, 2018. 3", + "[20] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77(12): 15573-15601, 2018. 6", + "[21] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2334-2343, 2017. 3", + "[22] Kehong Gong, Jianfeng Zhang, and Jiashi Feng. Poseaug: A differentiable pose augmentation framework for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8575-8584, 2021. 3", + "[23] Brian Gordon, Sigal Raab, Guy Azov, Raja Giryes, and Daniel Cohen-Or. Flex: Parameter-free multiview 3d human motion reconstruction. arXiv preprint arXiv:2105.01937, 2021. 3", + "[24] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3", + "[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 8", + "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3", + "[27] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7779-7788, 2020. 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2510", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. arXiv preprint arXiv:1905.05754, 2019. 3, 4", + "[29] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3", + "[30] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3", + "[31] Hanbyul Joo, Hyun Soo Park, and Yaser Sheikh. Map visibility estimation for large-scale dynamic 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1122-1129, 2014. 1", + "[32] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In Proceedings of the IEEE International Conference on Computer Vision, pages 3334-3342, 2015. 1, 2, 3, 6", + "[33] Abdelrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):1-14, 2021. 1, 3", + "[34] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1077-1086, 2019. 3", + "[35] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR, pages 11977-11986, 2019. 3", + "[36] Yasunori Kudo, Keisuke Ogaki, Yusuke Matsui, and Yuri Odagiri. Unsupervised adversarial learning of 3d human pose from 2d joint locations. arXiv preprint arXiv:1803.08244, 2018. 3", + "[37] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5", + "[38] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6152-6162, 2020. 3", + "[39] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11886-11895, 2021. 1, 3, 6", + "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 3, 8", + "[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "person linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 6, 7", + "[42] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 3", + "[43] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2640-2649, 2017. 3", + "[44] William McNally, Kanav Vats, Alexander Wong, and John McPhee. Evopose2d: Pushing the boundaries of 2d human pose estimation using neuroevolution. arXiv preprint arXiv:2011.08446, 2020. 3", + "[45] Dushyant Mehta, Srinath Sridhar, Oleksandr Sotnychenko, Helge Rhodin, Mohammad Shafiei, Hans-Peter Seidel, Weipeng Xu, Dan Casas, and Christian Theobalt. Vnect: Real-time 3d human pose estimation with a single rgb camera. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 3", + "[46] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. In NIPS, pages 2277-2287, 2017. 3", + "[47] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 3", + "[48] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In CVPR, pages 1253-1262, 2017. 3", + "[49] Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Domes to drones: Self-supervised active triangulation for 3d human pose reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 6, 7", + "[50] Alin-Ionut Popa, Mihai Zanfir, and Cristian Sminchisescu. Deep multitask architecture for integrated 2d and 3d human sensing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6289-6298, 2017. 3", + "[51] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In ICCV, pages 4342-4351, 2019. 3", + "[52] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15190-15200, 2021. 1, 3", + "[53] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In CVPR, pages 6040-6049, 2020. 3", + "[54] Pierre H Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, et al." + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2511", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Byol works even without batch statistics. arXiv preprint arXiv:2010.10241, 2020.3", + "[55] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, pages 5693-5703, 2019. 1, 3, 5, 8", + "[56] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 529-545, 2018. 3", + "[57] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 1, 3, 6, 2", + "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[59] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11148-11157, 2021. 1, 3, 6", + "[60] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, pages 466-481, 2018. 3", + "[61] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7093-7102, 2020. 5", + "[62] Jianfeng Zhang, Xuecheng Nie, and Jiashi Feng. Inference stage optimization for cross-scenario 3d human pose estimation. Advances in Neural Information Processing Systems, 33:2408-2419, 2020. 3", + "[63] Jianfeng Zhang, Yujun Cai, Shuicheng Yan, Jiashi Feng, et al. Direct multi-view multi-person 3d pose estimation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6", + "[64] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. arXiv preprint arXiv:2108.02452, 2021. 3", + "[65] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In ICCV, 2017. 3", + "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + ], + "bbox": [ + 78, + 90, + 468, + 771 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2512", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_model.json b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c964bf9cd94926a9a336ced10cc2a03d10725df7 --- /dev/null +++ b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.856, + 0.152 + ], + "angle": 0, + "content": "SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.18, + 0.703, + 0.2 + ], + "angle": 0, + "content": "Vinkle Srivastav\\(^{1,2*}\\) Keqi Chen\\(^{1*}\\) Nicolas Padoy\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.204, + 0.761, + 0.238 + ], + "angle": 0, + "content": "1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France \n2IHU Strasbourg, France" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.242, + 0.76, + 0.256 + ], + "angle": 0, + "content": "srivastav@unistra.fr keqi.chen@unistra.fr npadoy@unistra.fr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.291, + 0.314, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.732 + ], + "angle": 0, + "content": "We present a new self-supervised approach, SelfPose3d, for estimating 3d poses of multiple persons from multiple camera views. Unlike current state-of-the-art fully-supervised methods, our approach does not require any 2d or 3d ground-truth poses and uses only the multiview input images from a calibrated camera setup and 2d pseudo poses generated from an off-the-shelf 2d human pose estimator. We propose two self-supervised learning objectives: self-supervised person localization in 3d space and self-supervised 3d pose estimation. We achieve self-supervised 3d person localization by training the model on synthetically generated 3d points, serving as 3d person root positions, and on the projected root-heatmaps in all the views. We then model the 3d poses of all the localized persons with a bottleneck representation, map them onto all views obtaining 2d joints, and render them using 2d Gaussian heatmaps in an end-to-end differentiable manner. Afterwards, we use the corresponding 2d joints and heatmaps from the pseudo 2d poses for learning. To alleviate the intrinsic inaccuracy of the pseudo labels, we propose an adaptive supervision attention mechanism to guide the self-supervision. Our experiments and analysis on three public benchmark datasets, including Panoptic, Shelf, and Campus, show the effectiveness of our approach, which is comparable to fully-supervised methods. Code is available at https://github.com/CAMMA-public/SelfPose3D." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.761, + 0.21, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.786, + 0.471, + 0.879 + ], + "angle": 0, + "content": "The task of estimating 3d poses for multiple persons using a few calibrated cameras is a challenging computer vision problem [10, 18, 32, 52, 57]. A significant part of this challenge lies in identifying and matching the same person across different camera views. The solutions developed so far generally use one of the two paradigms: learning-based" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.292, + 0.895, + 0.58 + ], + "angle": 0, + "content": "methods and optimization-based methods. The learning-based methods develop novel deep-learning models and use 3d ground-truth poses for both training the models and establishing person correspondences across different views [39, 52, 57, 59, 63]. The accurate 3d ground-truth poses are typically generated using a dense camera system [32]. In contrast, the optimization-based methods formulate the 3d pose reconstruction as a mathematical optimization task, primarily focusing on aligning and matching the 2d poses across different camera views to infer 3d poses using triangulation within the framework of multi-view geometry [10, 18, 31-33, 49]. The 2d poses are estimated using off-the-shelf 2d human pose detectors [5, 55]. These methods apply geometric and spatial constraints in the optimization loop to ensure the anatomical plausibility and consistency of the inferred 3d poses. Although these methods do not require 3d ground-truth poses, their effectiveness is somewhat limited compared to the fully-supervised learning-based methods, see Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.894, + 0.687 + ], + "angle": 0, + "content": "In this paper, we explore the possibility of combining the strengths of both paradigms. Specifically, we investigate whether it's feasible to utilize a learning-based model for multi-view, multi-person 3D pose estimation and simultaneously eliminate its dependence on 3D ground-truth poses by incorporating geometric and appearance constraints, drawing inspiration from optimization-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.853 + ], + "angle": 0, + "content": "We propose, SelfPose3d, a self-supervised learning-based approach to estimate the 3d poses of multiple persons from a few calibrated cameras without using any 2d or 3d ground-truth poses. Our approach requires only 2d pseudo poses obtained using an off-the-shelf 2d pose detector [55]. Learning 3d poses without 3D ground-truth poses would require suitable supervisory signals to train a learning-based model. We follow the learning-by-projection paradigm, where the main idea is to learn the 3d output by comparing the projected bottleneck 3d output against the 2d input features [12]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We consider VoxelPose [57] as a learning-based method and use its output 3d poses as a bottleneck representation. To recover the accurate underlying 3d poses, we propose us-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.318, + 0.9 + ], + "angle": 0, + "content": "* co-first authors with equal contributions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "2502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.424, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.09, + 0.871, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.35, + 0.895, + 0.409 + ], + "angle": 0, + "content": "Figure 1. Our self-supervised approach, called SelfPose3d, estimates multi-person 3d poses from multi-view images and pseudo 2d poses generated using an off-the-shelf 2d human pose estimator. We propose a self-supervised learning objective that generates differentiable and geometrically constrained 2d joints and heatmaps across multiple views from bottleneck 3d poses. On the right, we show 3d pose outputs from our approach along with estimated body meshes (using SMPL body mesh fitting on 3d poses [4, 41]) and the projected 2d poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.473, + 0.681 + ], + "angle": 0, + "content": "ing differentiable multi-view 2d representations and cross-affine-view consistency. In particular, given a multi-view input image, we apply two random affine augmentations and pass them to the VoxelPose. It generates the bottleneck 3d poses corresponding to each affine augmented multi-view image. To enforce the model to learn and reason in the spatial dimension, we project the bottleneck 3d poses onto each view, obtaining 2d joints, and rendering them into spatial 2d heatmap representations in an end-to-end differentiable way. We further put tight geometric constraints by cross-affine-view operation, i.e. the bottleneck 3d poses from the 1st affine augmented multi-view image is mapped and rendered in the 2nd affine augmented multi-view image space and vice versa. Finally, we use the affine transformed 2d joints and heatmaps from the 2d pseudo poses to enable the geometrically constrained learning, with \\( L_{1} \\) and \\( L_{2} \\) losses respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.901 + ], + "angle": 0, + "content": "As the 2d pseudo poses contain non-negligible noises (mostly due to occlusions, see Figure 3), we propose adaptive supervision attention to guide our model to focus on more reliable regions. We apply two strategies towards \\( L_{1} \\) joint loss and \\( L_{2} \\) heatmap loss; for \\( L_{1} \\) joint loss supervision, we employ hard attention, where we ignore the one view with the largest absolute error for each multi-view image set; for \\( L_{2} \\) heatmap loss supervision, we employ soft attention using a lighter backbone to process each view, obtaining same-size attention heatmaps. During \\( L_{2} \\) loss computation, we compute the element-wise product of the attention heatmaps and the square error before averaging. To avoid obtaining zero attention, which the model tends to do, we add a regularization term, where we create tensors of all" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.423, + 0.892, + 0.452 + ], + "angle": 0, + "content": "ones as the attention heatmap labels and use \\( L_{2} \\) loss as the attention loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.893, + 0.699 + ], + "angle": 0, + "content": "Finally, specific to our choice of learning-based method, i.e., VoxelPose, which uses a voxel-based 3d root localization model to localize the persons in space using ground-truth 3d root joints (mid-hip joint), we use a simple but effective strategy to localize persons in space. Specifically, we randomly place 3d points in 3d world-space and project them to each view using the given camera parameters, subsequently rendering the projected 2d points as heatmap representations. This generates a synthetic dataset containing 3d points (roots) and their corresponding rendered multiview root-heatmaps. We then use this dataset to train a 3d root localization model that takes multi-view root-heatmaps as input and predicts the 3d roots as output. We further regularize the model by enforcing invariant constraints between pairs of affine augmented root-heatmaps coming from the real multi-view input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.895, + 0.853 + ], + "angle": 0, + "content": "Evaluation on three 3d pose benchmarks datasets, Panoptic [32], Shelf [1] and Campus [1], along with extensive ablation studies on the Panoptic [32] dataset, show the effectiveness of our approach. Our approach reaches a performance comparable to learning-based fully-supervised approaches and performs significantly better than optimization-based approaches. Moreover, SMPL body mesh fitting [4, 41] on our estimated 3d poses generates geometrically plausible body shapes (see Figure 1 and Figure 4)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We summarize our contributions as follows: 1) We address the challenging multi-person multi-view 3d person pose estimation problem using a self-supervised approach" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2503" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "without any 2d or 3d ground truth. 2) We propose self-supervised 3d pose estimation by using a new method to recover geometrically constrained 2d joints and heatmap representations from the bottleneck 3d poses. 3) We propose adaptive supervision attention to address the misinformation caused by the inaccurate pseudo labels. 4) We propose self-supervised 3d root localization to estimate the 3d root location utilizing synthetic 3d roots and the corresponding rendered multi-view root heatmaps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.242, + 0.216, + 0.258 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.47, + 0.327 + ], + "angle": 0, + "content": "In this section, we briefly review current works related to fully-supervised learning-based methods for 3d pose estimation, optimization-based methods for 3D pose estimation, and self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.473, + 0.599 + ], + "angle": 0, + "content": "Fully-supervised methods: Monocular 3d pose estimation [22, 43, 45, 47, 50, 56, 62, 65] is an ill-posed problem due to depth ambiguities as multiple 3d poses can produce same 2d pose projection. Having access to multi-view cameras can alleviate such depth ambiguities achieving the state-of-art results on benchmark datasets [23, 27, 28, 39, 51, 52, 57, 59, 63]. For single-person scenes, these approaches exploit multi-view geometry [24] to either fuse the visual features [27, 51], perform triangulation on heatmaps [28, 53], or use pictorial structural models for 3d reconstruction [48, 51]. The multi-person scene offers extra complexity due to the variability in the number of person in each view and the unknown cross-view correspondence. Existing multi-person multi-view approaches are based on volumetric paradigm [52, 57, 64], or direct regression [63] based on transformers [6, 58, 66]. Despite their good performance, these approaches rely on ground-truth 3d poses, which are generated using dense camera systems [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.75 + ], + "angle": 0, + "content": "Optimization-based 3d pose estimation: For the multi-person and multi-view scenario, optimization-based approaches use an off-the-shelf person-id detector across all the views to solve the correspondence and triangulation problem and temporal refinement along with training a reinforcement learning agent to find the best camera locations for 3d pose reconstruction [49]. More recent approaches utilize multi-view 3d reconstruction in the optimization loop inferring 3D poses that are geometrically and spatially coherent [10, 18, 33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Self-supervised learning: Self-supervised learning can be broadly classified into self-supervised representation learning and self-supervised task learning. Self-supervised representation learning aims to use large-scale unlabeled data to learn generic feature representations. The recent promising results from these approaches have started to surpass the fully-supervised baselines for various downstream tasks [7, 11, 26, 54]. Self-supervised task learning aims to learn a particular downstream task without using ground truth labels and has been applied to 2d pose estimation [29, 30]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "single-person 3d pose estimation [9, 19, 34, 36, 38], and surface correspondences estimation [3]. Self-supervised approaches for 3d pose estimation have primarily been developed for single-person scenarios. Given 2d poses, estimated by utilizing advances in the 2d pose estimation methods [5, 13, 14, 21, 35, 42, 44, 46, 55, 60], these approaches use the supervisory signals generated from multi-view geometry [34], video constraints [38], or adversarial learning [9, 19, 36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.895, + 0.306 + ], + "angle": 0, + "content": "Our work proposes a learning-based approach to model the 3d poses as bottleneck representations and recover geometrically constrained and spatially accurate 2d joints and heatmap representations in an end-to-end differentiable manner." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.328, + 0.637, + 0.347 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.356, + 0.676, + 0.372 + ], + "angle": 0, + "content": "3.1. Problem overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.895, + 0.534 + ], + "angle": 0, + "content": "Given a training dataset of multi-view images \\(\\mathcal{D} = \\{\\mathbf{x}|\\mathbf{y}^*\\}\\) where \\(\\mathbf{x} \\in \\mathcal{R}^{C \\times 3 \\times H \\times W}\\) is a multi-view image set from \\(C\\) cameras with height \\(H\\) and width \\(W\\), and \\(\\mathbf{y}^* \\in \\mathcal{R}^{C \\times P \\times J \\times 2}\\) represents the 2d pseudo poses for \\(P\\) persons with \\(J\\) joints, the goal is to learn a deep learning model that estimates the 3d poses \\(\\mathcal{V} \\in \\mathcal{R}^{P \\times J \\times 3}\\) of all the \\(P\\) persons from the multi-view input image \\(\\mathbf{x}\\). It is to be noted that \\(P\\) can vary in each camera view due to occlusion and noisy pseudo 2d pose estimation. For simplicity in the notation, we keep the same variable \\(\\mathbb{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.628 + ], + "angle": 0, + "content": "Fully-supervised approaches rely on 3d ground-truth poses, while we only have 2d pseudo poses \\(\\mathbf{y}^*\\). Therefore, after obtaining 3d poses \\(\\mathcal{V} \\in \\mathcal{R}^{P \\times J \\times 3}\\) following traditional approach, we propose to project the poses to each view obtaining 2d poses \\(\\mathbf{y} \\in \\mathcal{R}^{C \\times P \\times J \\times 2}\\), and train the model from 2d pseudo poses \\(\\mathbf{y}^* \\in \\mathcal{R}^{C \\times P \\times J \\times 2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.735 + ], + "angle": 0, + "content": "In the following, we present our self-supervised approach based on fully-supervised VoxelPose [57]. We first generate pseudo 2d poses, and then propose self-supervised 3d root localization, self-supervised 3d pose estimation, and an adaptive supervision attention to learn 3d poses in a self-supervised manner, without modifying the original Voxel-Pose structure." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.754, + 0.753, + 0.771 + ], + "angle": 0, + "content": "3.2. Generating pseudo 2d poses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To circumvent the dependence on the ground-truth 2d poses, we generate the 2d pseudo poses on the training dataset using Mask R-CNN [25] to generate person bounding boxes followed by using HRNet [55] to generate 2d poses of each detected person bounding box. The two-stage approach is chosen based on its state-of-art performance on the COCO dataset [40]. We first pre-train the 2d CNN backbone heatmap_net\\(_{2d}\\) with pseudo 2d poses." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2504" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.091, + 0.543, + 0.101 + ], + "angle": 0, + "content": "SelfPose3d Framework" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.102, + 0.887, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Figure 2. Illustrating our self-supervised SelfPose3d approaches for multi-view multi-person 3d pose estimation. Instead of using ground-truth 3d poses for learning, we propose self-supervised learning objectives to localize 3d roots (mid-hip location of the person) and estimate their 3d poses. We utilize a synthetic 3d roots dataset, two different affine transformations on the multi-view input images \\((t_{r,s}^{1}, t_{r,s}^{2})\\) parametrized by rotation \\(r\\) and scale \\(s\\), a differentiable cross-affine-view 2d joints and heatmaps rendering from the bottleneck 3d poses, and an adaptive supervision attention mechanism to automatically learn the 3d poses in world-space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.463, + 0.383, + 0.48 + ], + "angle": 0, + "content": "3.3. Self-supervised 3d root localization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.565 + ], + "angle": 0, + "content": "Given 2d multi-view heatmaps from all the views and all the joints \\(\\mathcal{H} \\in \\mathcal{R}^{C \\times J \\times \\frac{H}{4} \\times \\frac{W}{4}}\\) estimated using a 2d backbone model heatmap。\\(\\mathrm{net}_{2\\mathrm{d}}\\), we use [28] to construct a discretized 3d feature volume \\(\\mathcal{F} \\in \\mathcal{R}^{J \\times X \\times Y \\times Z}\\) for each joint by un-projecting the 2d multi-view heatmaps to 3d space:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.577, + 0.47, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {\\text {u n p r o j}} (\\operatorname {c a m}, \\text {c e n t e r}, t _ {r, s}): \\mathcal {H} \\longrightarrow \\mathcal {F}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.473, + 0.868 + ], + "angle": 0, + "content": "To localize persons' root (mid-hip) joint in 3d space without using 3d ground truth, we hypothesize that 2d multi-view heatmaps of the root location \\(\\mathcal{H}_{\\mathrm{root}} \\in \\mathcal{R}^{C \\times \\frac{H}{4} \\times \\frac{W}{4}}\\) are sufficient for 3d root localization (see Sec. 7.6 for verification). Then we generate the 3d feature volume for the root location \\(\\mathcal{F}_{\\mathrm{root}} \\in \\mathcal{R}^{X \\times Y \\times Z}\\) using \\(\\mathcal{H}_{\\mathrm{root}}\\) using Eq. (1), which has the same dimensions as predicted root-volumes \\(\\mathcal{G}\\). This allows us to establish a one-to-one relationship between 3d root-volumes \\(\\mathcal{G}\\) and 2d multi-view root-heatmaps \\(\\mathcal{H}_{\\mathrm{root}}\\). We generate a synthetic root dataset \\(\\mathcal{D}_{\\mathrm{root}} = \\{\\mathcal{G}_i^{\\mathrm{syn*}} | \\mathcal{H}_{\\mathrm{root\\_i}}^{\\mathrm{syn}}\\}_{i=1}^N\\) where \\(\\mathcal{G}_i^{\\mathrm{syn*}} \\in \\mathcal{R}^{X \\times Y \\times Z}\\) contains the root-volumes of randomly placed 3d points, and \\(\\mathcal{H}_{\\mathrm{root\\_i}}^{\\mathrm{syn}}\\) is the corresponding 2d multi-view heatmaps generated by projecting the random 3d points to each view using camera parameters cam. After unprojecting \\(\\mathcal{H}_{\\mathrm{root}}^{\\mathrm{syn}}\\) to \\(\\mathcal{F}_{\\mathrm{root}}^{\\mathrm{syn}}\\), and passing it through root\\_net obtaining \\(\\mathcal{G}_{\\mathrm{root}}^{\\mathrm{syn}}\\), we compute the \\(L_2\\) loss error as the synthetic root loss \\(l_{\\mathrm{root\\_syn}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.885, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {r o o t - s y n}} = \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {\\text {s y n}}, \\mathcal {G} ^ {\\text {s y n *}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.66 + ], + "angle": 0, + "content": "To further regularize root_net on the real-world 2d multi-view root-heatmaps, we propose the root consistency loss. Given a multi-view training image set \\( x^0 \\), we apply two affine transformations \\( (t_{r,s}^{1}, t_{r,s}^{2}) \\) with random rotation and scaling \\( (r, s) \\) to generate two affine transformed multiview images \\( (x^1, x^2) \\). We pass \\( x^0 \\), \\( x^1 \\) and \\( x^2 \\) through heatmap_net2d, construct the root feature volumes using Eq. (1) with corresponding affine transformation parameters, and finally obtain \\( \\mathcal{G}^0 \\), \\( \\mathcal{G}^1 \\) and \\( \\mathcal{G}^2 \\) through root_net. Since \\( \\mathcal{G}^1 \\) and \\( \\mathcal{G}^2 \\) are invariant to the applied affine transformations \\( t_{r,s}^1 \\) and \\( t_{r,s}^2 \\), we use \\( \\mathcal{G}^0 \\) as the baseline to compute the \\( L_2 \\) loss error between \\( \\mathcal{G}^0 \\), \\( \\mathcal{G}^1 \\) and \\( \\mathcal{G}^2 \\) as the root consistency loss \\( l_{\\mathrm{root\\_C}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.673, + 0.891, + 0.693 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {r o o t} C} = \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {0}, \\mathcal {G} ^ {1}\\right) + \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {0}, \\mathcal {G} ^ {2}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.756 + ], + "angle": 0, + "content": "We train root_net by minimizing Eq. (2) and Eq. (3). We generate the person proposals \\(\\{\\mathrm{root}_i\\}_{i=1}^K\\) by applying nonmaximum suppression (NMS) and thresholding on \\(\\mathcal{G}^2\\) (\\(\\mathcal{G}^1\\) would also work)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.8, + 0.78 + ], + "angle": 0, + "content": "3.4. Self-supervised 3d pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.847 + ], + "angle": 0, + "content": "Given pseudo 2d poses \\(\\mathbf{y}_{2\\mathrm{d}}^{*}\\), person proposals \\(\\{\\mathrm{root}_i\\}_{i = 1}^P\\) and 2d multi-view heatmaps \\(\\mathcal{H}^1,\\mathcal{H}^2\\) predicted using heatmap_net2d, we describe our self-supervised 3d pose estimation approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.846, + 0.894, + 0.905 + ], + "angle": 0, + "content": "The person proposals \\(\\{\\mathrm{root}_i\\}_{i = 1}^P\\) are used to generate the 3d feature volumes i.e \\(\\left\\{\\mathcal{F}_i^1\\right\\}_{i = 1}^P = \\left\\{\\mathcal{P}_{\\mathrm{unproj}}(\\mathrm{cam},r_i,t_{r,s}^1)\\right\\}_{i = 1}^P\\) and \\(\\left\\{\\mathcal{F}_i^2\\right\\}_{i = 1}^P =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2505" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.089, + 0.468, + 0.192 + ], + "angle": 0, + "content": "\\(\\left\\{\\mathcal{P}_{\\mathrm{unproj}}(\\mathrm{cam},r_i,t_{r,s}^2)\\right\\}_{i = 1}^P\\) corresponding to the person feature volumes for each affine augmented multi-view input image \\(x^{1}\\) and \\(x^{2}\\), respectively. \\(\\left\\{\\mathcal{F}_i^1\\right\\}_{i = 1}^P\\) and \\(\\left\\{\\mathcal{F}_i^2\\right\\}_{i = 1}^P\\) are passed through pose_net3d and soft-argmax [8] to estimate the 3d poses \\(\\mathcal{V}^1,\\mathcal{V}^2\\in \\mathcal{R}^{P\\times J\\times 3}\\). These 3d poses serve as a bottleneck representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.193, + 0.468, + 0.285 + ], + "angle": 0, + "content": "Given the camera parameters cam and the affine transformation parameters \\(t_{r,s}^{1}, t_{r,s}^{2}\\), we project the bottleneck 3d poses in cross-affine-view i.e. \\(\\mathcal{V}^1\\) are projected to \\(x^2\\) image space using \\(t_{r,s}^{2}\\) to generate multi-view 2d poses \\(\\mathbf{y}^2 \\in \\mathcal{R}^{C \\times P \\times J \\times 2}\\), and \\(\\mathcal{V}^2\\) are projected to the \\(x^1\\) image space using \\(t_{r,s}^{1}\\) to generate multi-view 2d poses \\(\\mathbf{y}^1 \\in \\mathcal{R}^{C \\times P \\times J \\times 2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.468, + 0.495 + ], + "angle": 0, + "content": "We propose to render \\(\\mathbf{y}^1\\) and \\(\\mathbf{y}^2\\) 2d poses in the 2d heatmap representation. The heatmap representation encodes the per-pixel likelihood of a body joint and has been a vital component to enable the state-of-the-art 2d pose estimation approaches [55]. Generating heatmap representation has essentially been a pre-processing step where state-of-the-art approaches quantize the 2d joints before generating the heatmap [55]. However, this quantization step is non-differentiable and could cut the backward gradient flow. Zhang et al. [61] show that encoding floating point 2d joints into heatmap representation in their pre-processing step leads to improved performance. We propose to use the same differentiable approach in an online way to render the projected 2d joints into the heatmap representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.468, + 0.588 + ], + "angle": 0, + "content": "We render \\(\\mathbf{y}^1\\) and \\(\\mathbf{y}^2\\) into heatmap representation to generate 2d multi-view heatmaps \\(\\mathcal{H}^1\\), and \\(\\mathcal{H}^2\\), respectively. We apply the affine transformations \\(t_{r,s}^1, t_{r,s}^2\\) on the pseudo 2d poses \\(\\mathbf{y}_{2\\mathrm{d}}^{*}\\) to generate pseudo 2d multi-view joints \\(\\mathbf{y}_{2\\mathrm{d}}^{1*}, \\mathbf{y}_{2\\mathrm{d}}^{2*}\\) and heatmaps \\(\\mathcal{H}^{1*}, \\mathcal{H}^{2*}\\). Then, we compute the \\(L_2\\) loss between heatmaps as the pose heatmap loss \\(l_{\\mathrm{pose\\_H}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.601, + 0.468, + 0.62 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {p o s e} \\mathrm {H}} = \\mathcal {L} _ {2} \\left(\\mathcal {H} ^ {1}, \\mathcal {H} ^ {1 *}\\right) + \\mathcal {L} _ {2} \\left(\\mathcal {H} ^ {2}, \\mathcal {H} ^ {2 *}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.468, + 0.747 + ], + "angle": 0, + "content": "After training with \\( l_{\\mathrm{pose\\_H}} \\) preliminarily, we propose to add the \\( L_{1} \\) loss between 2d joints to further fine-tune the model. For each view, we employ the Hungarian algorithm [37] to obtain the optimal assignment between \\( \\mathbf{y} \\) and \\( \\mathbf{y}_{2\\mathrm{d}}^{*} \\), where the matching cost is the mean absolute error. Based on the optimal assignment, we obtain the \\( L_{1} \\) loss as \\( l_{\\mathrm{pose\\_J}} \\). Then we use \\( l_{\\mathrm{pose\\_H}} \\) and \\( l_{\\mathrm{pose\\_J}} \\) together to train the whole network, where \\( \\lambda \\) is a manually defined weight:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.761, + 0.468, + 0.78 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {p o s e - J}} = \\mathcal {L} _ {1} \\left(\\mathbf {y} ^ {1}, \\mathbf {y} _ {2 \\mathrm {d}} ^ {1 *}\\right) + \\mathcal {L} _ {1} \\left(\\mathbf {y} ^ {2}, \\mathbf {y} _ {2 \\mathrm {d}} ^ {2 *}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.802, + 0.468, + 0.819 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {p o s e - 3 d}} = l _ {\\text {p o s e - H}} + \\lambda l _ {\\text {p o s e - J}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As the network needs to reason about the 2d joint locations in spatial dimension, it implicitly solves the person correspondence problem. Training pose_net3d with 3d pose loss \\( l_{\\mathrm{pose\\_3d}} \\) performs decently, but to achieve even better results, we introduce the adaptive supervision attention." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.696, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.254, + 0.655, + 0.262 + ], + "angle": 0, + "content": "ground truth 2d poses" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.09, + 0.888, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.254, + 0.835, + 0.261 + ], + "angle": 0, + "content": "pseudo 2d poses" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Figure 3. Comparing ground-truth 2d poses generated by projecting the ground-truth 3d poses to each multi-view image and our pseudo 2d poses generated by running HRNet human pose estimation model [55] on the training dataset. Pseudo 2d poses contain localization errors due to occlusion (see the red arrows), and ground-truth 2d poses exist for partially or even entirely occluded persons (see the blue dotted arrows)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.394, + 0.771, + 0.409 + ], + "angle": 0, + "content": "3.5. Adaptive supervision attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Traditional \\( L_{1} \\) and \\( L_{2} \\) losses treat each label equally, which is sub-optimal in two aspects: (1) the 2d human pose detector may generate inaccurate labels due to occlusions (see the red arrows in Figure 3); (2) the 3d-2d projection will output 2d joints in certain views even when the person is entirely occluded (see the blue dotted arrows in Figure 3). Therefore, we propose to employ attentions to adaptively guide the supervision process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.632 + ], + "angle": 0, + "content": "For \\( L_{2} \\) loss supervision, we use the soft attention. Specifically, we use ResNet-18 to extract the visual features of the views, followed by deconvolutional layers to obtain the attention heatmaps \\( \\mathcal{A} \\) (see attn\\_net\\(_{2d}\\) in Figure 2). Then we compute the element-wise product of \\( \\mathcal{A} \\) and the square error before averaging, as the new loss \\( l_{\\mathrm{pose\\_H}}^{\\mathrm{attn}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.64, + 0.891, + 0.68 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {p o s e - H}} ^ {\\mathrm {a t t n}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {A} _ {i} \\otimes \\left(\\mathcal {H} _ {i} - \\mathcal {H} _ {i} ^ {*}\\right) ^ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.759 + ], + "angle": 0, + "content": "To avoid that \\(\\mathcal{A}\\) becomes zero, we add a regularization term. We create tensors of all ones \\(\\mathbb{1}\\) as the attention heatmap labels, and compute \\(L_{2}\\) error as the attention loss \\(l_{\\mathrm{attn}}\\). If \\(l_{\\mathrm{attn}}\\) becomes zero, \\(l_{\\mathrm{pose\\_H}}^{\\mathrm{attn}}\\) degrades to non-attentive version." + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.774, + 0.891, + 0.79 + ], + "angle": 0, + "content": "\\[\nl _ {\\mathrm {a t t n}} = \\mathcal {L} _ {2} (\\mathcal {A}, \\mathbb {1}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.858 + ], + "angle": 0, + "content": "For \\( L_{1} \\) loss supervision, we use the hard attention. For each input with \\( K \\) views, we compute the \\( L_{1} \\) loss of each view, find the view with the largest loss, and ignore it when averaging the final loss \\( l_{\\mathrm{pose\\_J}}^{\\mathrm{att}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.88, + 0.891, + 0.905 + ], + "angle": 0, + "content": "\\[\nj = \\underset {i} {\\arg \\max } \\mathcal {L} _ {1} \\left(\\mathbf {y} _ {i}, \\mathbf {y} _ {i, 2 \\mathrm {d}} ^ {*}\\right) \\quad (i = 1, 2, \\dots , K) \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2506" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.146, + 0.103, + 0.47, + 0.146 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {p o s e - J}} ^ {\\mathrm {a t t n}} = \\frac {1}{K - 1} \\sum_ {i = 1, i \\neq j} ^ {K} \\mathcal {L} _ {1} \\left(\\mathbf {y} _ {i}, \\mathbf {y} _ {i, 2 \\mathrm {d}} ^ {*}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.155, + 0.47, + 0.186 + ], + "angle": 0, + "content": "In general, the final 3d pose loss \\( l_{\\mathrm{pose\\_3d}}^{\\mathrm{attn}} \\) is as follows, where \\( \\lambda \\) and \\( \\sigma \\) are manually defined weights:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.2, + 0.469, + 0.22 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {p o s e - 3 d}} ^ {\\text {a t t n}} = l _ {\\text {p o s e - H}} ^ {\\text {a t t n}} + \\lambda l _ {\\text {p o s e - J}} ^ {\\text {a t t n}} + \\sigma l _ {\\text {a t t n}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.225, + 0.469, + 0.256 + ], + "angle": 0, + "content": "We train pose_net3d by minimizing \\( l_{\\mathrm{pose\\_3d}}^{\\mathrm{attn}} \\). Our self-supervised approach is visually described in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.266, + 0.292, + 0.282 + ], + "angle": 0, + "content": "3.6. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.289, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Training strategies For the Panoptic dataset, similar to VoxelPose [57], we first train heatmap_net2d for 20 epochs with pseudo 2d poses. We use the Adam optimizer with an initial learning rate of 1e-4, which decreases to 1e-5 and 1e-6 at the 10th and 15th epochs, respectively. Then, we train the root_net for 1 epoch, followed by end-to-end joint training of the whole network for 5 epochs using only the \\( L_{2} \\) loss, with a learning rate of 1e-4. Afterwards, we add \\( L_{1} \\) loss to train the whole network for another 5 epochs with a learning rate of 5e-5. \\( \\lambda \\) and \\( \\sigma \\) in Eq. (11) are set to 0.01 and 0.1 respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.471, + 0.609 + ], + "angle": 0, + "content": "We use the random rotation between \\(-45^{\\circ}\\) to \\(45^{\\circ}\\) and random scale between \\(-0.35\\) to 0.35. We also apply spatial augmentations using rand-augment [16] and rand-cutout [17] using python image library1. The rand-augment consists of \"contrast-jittering\", \"auto-contrast\", \"equalize\", \"color-jittering\", \"sharpness-jittering\", and \"brightness-jittering\", and the rand-cutout places random square boxes of sizes between 20 to 40 pixels at random locations in the image. We use the SMPL model and optimization-based body fitting approach2[4, 41] to estimate body mesh parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Inference pipeline During inference, we input the multiview RGB images, and obtain the estimated 3d poses in an end-to-end pipeline. For each view, the backbone generates corresponding 2d heatmaps for cuboid construction. Then, given constructed cuboid of the whole space, the root_net predicts root joint locations of all persons. Finally, the pose_net outputs the regressed 3d locations of each joint for every cuboid proposal of the root joints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.21, + 0.779 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.357, + 0.802 + ], + "angle": 0, + "content": "4.1. Datasets and evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.842 + ], + "angle": 0, + "content": "We conduct experiments on three benchmark datasets: Panoptic [32], Campus [1], and Shelf [1]." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.09, + 0.886, + 0.191 + ], + "angle": 0, + "content": "
MethodsAP25AP50AP100AP150Recall@500MPJPE[mm]
FSVoxelPose [57]83.698.399.899.998.817.7
Lin et al. [39]92.199.099.899.8-16.8
MvP [63]92.396.697.597.798.215.8
Wu et al. [59]----98.715.8
TEMPO [15]89.099.199.899.9-14.7
OBACTOR [49]-----168.4
MvPose [18]0.02.9759.9381.5398.2384.2
SSSelfPose3d (ours)55.196.498.599.099.624.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.202, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Table 1. Result on the Panoptic dataset (FS = fully-supervised, OB = optimization-based, SS = self-supervised)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.246, + 0.89, + 0.343 + ], + "angle": 0, + "content": "
MethodsShelfCampus
Actor 1Actor 2Actor 3AverageActor 1Actor 2Actor 3Average
FSErshadi et al. [20]93.375.994.888.094.292.984.690.6
Wu et al. [59]99.396.597.397.7----
MvP [63]99.395.197.897.498.294.197.496.6
VoxelPose [57]99.394.197.697.097.693.898.896.7
VoxelPose* [57]99.593.597.896.993.186.593.290.9
OB3DPS [2]75.369.787.677.593.575.784.484.5
MvPose [18]98.894.197.896.997.693.398.096.3
SSSelfPose3d97.290.397.995.192.582.289.287.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.353, + 0.893, + 0.41 + ], + "angle": 0, + "content": "Table 2. Results (in PCP) on Shelf and Campus datasets (FS = fully-supervised, OB = optimization-based, SS = self-supervised, * = reproduced results). SelfPose3d is trained from the pseudo 3d poses from the Panoptic training set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.665 + ], + "angle": 0, + "content": "The Panoptic dataset is a large-scale dataset captured inside a dome environment containing multiple persons performing daily social activities. We conduct extensive experiments on this dataset to evaluate and assess various components of our approach. We use the same data sequences for training and testing as VoxelPose [57] except that our training set doesn't include '160906_band3'. In other words, we are only using 9 multi-view videos for training (the '160906_band3' video is not available due to the broken images on the source website). We use the five HD camera images (3, 6, 12, 13, 23) to train and report the performance in our experiments. We use Average Precision (AP), Recall, and Mean Per Joint Position Error (MPJPE) in millimeters (mm) as evaluation metrics (higher AP and lower MPJPE are better) [57]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.666, + 0.892, + 0.741 + ], + "angle": 0, + "content": "The Shelf and Campus are two multi-person datasets capturing activities in the indoor and outdoor environments, respectively [1]. We use the same training and test split as [1, 57]. As used in [1, 57], we use the Percentage of Correct Parts (PCP) as evaluation metrics for these two datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Panoptic Table 1 shows the results on the challenging Panoptic dataset. All the fully-supervised approaches utilizing 2d and 3d ground-truth 3d poses reach nearly the same performance. SelfPose3d, without using any 3d or 2d ground-truth poses, achieves comparable results to fully-supervised approaches. Nevertheless, there still exists a gap compared to the fully-supervised VoxelPose model (96.4 v.s 98.3 AP50 and 24.5 v.s 17.7 MPJPE). However, unlike VoxelPose, which relies on heat maps from all the joints to es" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.85, + 0.468, + 0.901 + ], + "angle": 0, + "content": "1https : / / github . com / jizongFox / pytorch - randaugment 2https : // github . com / JiangWenPL / multiperson/ tree/master/misc/smplify-x" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2507" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.09, + 0.852, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.895, + 0.457 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for the 3d pose estimations, 2d projections on the multi-view images, and estimated SMPL body shapes on some example images from the Panoptic dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.553 + ], + "angle": 0, + "content": "timate 3d roots, we only use root-heatmaps to do the same. This results in the reduction of the input channel from 15 (number of keypoints) to 1 for the root_net, making our approach computationally faster." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.471, + 0.693 + ], + "angle": 0, + "content": "We also compare our approach with optimization-based baselines from Pirinen et al. [49] and Dong et al. [18]. These non-learning-based approaches fail to capture the multi-person interaction in a complex scene from a few sparse multi-view cameras. Our learning-based self-supervised approach achieves much better performance. It is to be noted that Pirinen et al. evaluate their approach on two multi-person sequences, whereas we evaluate on four multi-person sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Shelf and Campus We compare our approach with the state-of-the-art methods on the Shelf and Campus dataset. VoxelPose uses the 3d ground-truth from the Panoptic dataset to train their approach to these datasets due to noisy and incomplete 3d ground-truth poses. For a fair comparison with VoxelPose, we use the pseudo 3d poses (by running SelfPose3d on the Panoptic training set) and train on these two datasets in a fully supervised manner. As shown in Table 2, our approach using pseudo 3d poses from the Panoptic dataset also reaches the same performance as the fully-supervised approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.895, + 0.658 + ], + "angle": 0, + "content": "Qualitative visualizations Figure 4 shows 3d pose estimation results from our SelfPose3d approach on the challenging Panoptic dataset. Without using any 3d ground-truth, we can see that SelfPose3d is robust to occlusions and multiple persons while correctly identifying the person identities across all the views (see the corresponding 2d projections in Figure 4). We also show the qualitative results for the SMPL body mesh fitting [4, 41] on the estimated 3d poses. All these results demonstrate both the effectiveness and extendability of SelfPose3d. Please see the supplementary for more results." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.667, + 0.663, + 0.682 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Ground-truth 2d poses v.s pseudo 2d poses As shown in Table 3, when we use the ground-truth 2d poses in our self-supervised framework, 3d reconstruction error significantly reduces. To inspect the better performance when using the ground-truth 2d poses, we qualitatively compare the ground-truth 2d poses with the pseudo 2d poses on some example training images. Pseudo 2d poses contain localization errors due to occlusion, whereas ground-truth 2d poses exist for partially or even entirely occluded persons as shown in Figure 3. As the ground-truth 2d poses are generated by projecting the ground-truth 3d poses to each multi-view image, they serve as a suitable proxy for the 3d poses, thereby reaching a performance close to the fully-supervised approaches. However, obtaining the ground 2d" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2508" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.172, + 0.089, + 0.376, + 0.133 + ], + "angle": 0, + "content": "
2d posesAP50AP100MPJPE
ground-truth98.899.619.9
pseudo96.498.524.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.144, + 0.47, + 0.187 + ], + "angle": 0, + "content": "Table 3. The ground-truth 2d poses in our self-supervised framework decrease the 3d reconstruction error and reach the performance close to the fully-supervised approaches." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.2, + 0.446, + 0.255 + ], + "angle": 0, + "content": "
cross-affine-view consistencyaffine augsAP50AP100MPJPE
86.096.234.7
✓83.397.533.3
✓✓93.898.129.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.266, + 0.468, + 0.32 + ], + "angle": 0, + "content": "Table 4. Affine augmentations and cross-affine-view consistency significantly improves the 3d pose reconstruction accuracy. All three models are trained for two epochs with frozen backbone and frozen root_net and no attention." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.336, + 0.408, + 0.391 + ], + "angle": 0, + "content": "
L1 lossL2 lossAP25AP50AP100MPJPE
✓----
✓43.895.898.225.7
✓✓55.196.498.524.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.402, + 0.468, + 0.43 + ], + "angle": 0, + "content": "Table 5. Training using \\( L_{1} \\) and \\( L_{2} \\) pose losses together achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.457, + 0.468, + 0.487 + ], + "angle": 0, + "content": "poses in this way would be as challenging as acquiring the ground-truth 3d poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.468, + 0.629 + ], + "angle": 0, + "content": "Importance of cross-affine-view consistency and affine augmentations We also examine the effect of affine augmentations on the multi-view images and cross-affine-view consistency when generating differentiable 2d representations from the bottleneck 3d poses. As shown in Table 4, the affine augmentations and cross-affine-view consistency significantly improve the 3d pose reconstruction as they provide necessary geometric constraints during training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.468, + 0.724 + ], + "angle": 0, + "content": "Analysis of \\( L_{2} \\) and \\( L_{1} \\) pose losses We conduct experiments to analyze the use of \\( L_{1} \\) and \\( L_{2} \\) pose losses. As shown in Table 5, using \\( L_{2} \\) and \\( L_{1} \\) losses together can obtain better results than using \\( L_{2} \\) loss solely. Also, using \\( L_{1} \\) loss solely doesn't converge due to the label noises." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.745, + 0.468, + 0.805 + ], + "angle": 0, + "content": "Importance of adaptive supervision attention We also examine the necessity of adaptive supervision attention. Table 6 shows that supervision attention for both \\( L_{1} \\) and \\( L_{2} \\) losses are necessary for training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Influence of different 2d human pose estimation models Finally, we show how pseudo 2d poses generated from different 2d human pose estimation models affect the performance. As shown in Table 7, models that perform well on the COCO dataset [40] also generate better pseudo 2d poses" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.09, + 0.885, + 0.158 + ], + "angle": 0, + "content": "
L1 loss attentionL2 loss attentionAP25AP50AP100MPJPE
32.594.197.828.5
✓37.995.898.026.3
✓47.496.698.225.0
✓✓55.196.498.524.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.17, + 0.892, + 0.197 + ], + "angle": 0, + "content": "Table 6. Training using \\( L_{1} \\) and \\( L_{2} \\) loss supervisions together achieves the best performance." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.212, + 0.871, + 0.263 + ], + "angle": 0, + "content": "
Method for 2d pseudo pose generationAP50AP100MPJPEKeypoint AP on COCO-val[40]
Keypoint R-CNN (R-101) [25]89.297.631.966.1
HRNet (w48 384x288) [55]93.898.129.376.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.329 + ], + "angle": 0, + "content": "Table 7. Comparing different models for generating pseudo 2d poses. Models that perform well on the COCO dataset [40] also generate better pseudo 2d poses for the Panoptic dataset, helping SelfPose3d to achieve better performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.355, + 0.89, + 0.386 + ], + "angle": 0, + "content": "for the Panoptic dataset, helping SelfPose3d to achieve better performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.399, + 0.618, + 0.415 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.742 + ], + "angle": 0, + "content": "We present a self-supervised approach, called SelfPose3d, to address the challenging problem of multi-view multiperson 3d human pose estimation. Unlike current state-of-the-art methods that use difficult-to-acquire 3d ground-truth poses to train the model, SelfPose3d requires only multiview input images and an off-the-shelf 2d human pose detector. We propose a novel self-supervised learning objective that aims to recover 2d joints and heatmaps under different affine transformations from the bottleneck 3d poses. We further improve the performance of our approach by integrating adaptive supervision attention to address the misinformation caused by the inaccurate 2d pseudo labels from the off-the-shelf 2d human pose detector. We conduct extensive experiments on large-scale benchmark datasets, assess various components of our approach, and show that SelfPose3d reaches a performance on par with the well-established fully-supervised baselines. We visualize the 3d pose reconstruction in the complex multiple-person scenes and show that body shape meshes fitted on the estimated 3d poses look geometrically plausible under different viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.687, + 0.772 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was partially supported by French state funds managed by the ANR under references ANR-20-CHIA0029-01 (National AI Chair AI4ORSafety), ANR-10IAHU-02 (IHU Strasbourg), ANR-18-CE45-0011-03 (OptimiX), and by BPI France (project 5G-OR). This work was also granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2509" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures for multiple human pose estimation. In CVPR, pages 1669-1676, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. TPAMI, 38(10):1929-1942, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578. Springer, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, pages 7291-7299, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[8] Olivier Chapelle and Mingrui Wu. Gradient descent optimization of smoothed information retrieval metrics. Information retrieval, 13(3):216-235, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[9] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, M. V. Rohith, Stefan Stojanov, and James M. Rehg. Unsupervised 3d pose estimation with geometric self-supervision. CoRR, abs/1904.04812, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. In European Conference on Computer Vision, pages 541-557. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[12] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.216 + ], + "angle": 0, + "content": "[15] Rohan Choudhury, Kris M Kitani, and László A Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14750-14760, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[16] Ekin D Cubuk, Barret Zoph, Jonathon Shiens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[17] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In CVPR, pages 7792-7801, 2019. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[19] Dylan Drover, Ching-Hang Chen, Amit Agrawal, Ambrish Tyagi, and Cong Phuoc Huynh. Can 3d pose be learned from 2d projections alone? In Proceedings of the European Conference on Computer Vision (ECCV), pages 0–0, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77(12): 15573-15601, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2334-2343, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[22] Kehong Gong, Jianfeng Zhang, and Jiashi Feng. Poseaug: A differentiable pose augmentation framework for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8575-8584, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[23] Brian Gordon, Sigal Raab, Guy Azov, Raja Giryes, and Daniel Cohen-Or. Flex: Parameter-free multiview 3d human motion reconstruction. arXiv preprint arXiv:2105.01937, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[24] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7779-7788, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2510" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[28] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. arXiv preprint arXiv:1905.05754, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[29] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[30] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[31] Hanbyul Joo, Hyun Soo Park, and Yaser Sheikh. Map visibility estimation for large-scale dynamic 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1122-1129, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[32] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In Proceedings of the IEEE International Conference on Computer Vision, pages 3334-3342, 2015. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[33] Abdelrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):1-14, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[34] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1077-1086, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.47, + 0.573 + ], + "angle": 0, + "content": "[35] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR, pages 11977-11986, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[36] Yasunori Kudo, Keisuke Ogaki, Yusuke Matsui, and Yuri Odagiri. Unsupervised adversarial learning of 3d human pose from 2d joint locations. arXiv preprint arXiv:1803.08244, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[37] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[38] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6152-6162, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[39] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11886-11895, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "person linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[42] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[43] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2640-2649, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[44] William McNally, Kanav Vats, Alexander Wong, and John McPhee. Evopose2d: Pushing the boundaries of 2d human pose estimation using neuroevolution. arXiv preprint arXiv:2011.08446, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[45] Dushyant Mehta, Srinath Sridhar, Oleksandr Sotnychenko, Helge Rhodin, Mohammad Shafiei, Hans-Peter Seidel, Weipeng Xu, Dan Casas, and Christian Theobalt. Vnect: Real-time 3d human pose estimation with a single rgb camera. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[46] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. In NIPS, pages 2277-2287, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[47] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[48] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In CVPR, pages 1253-1262, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[49] Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Domes to drones: Self-supervised active triangulation for 3d human pose reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[50] Alin-Ionut Popa, Mihai Zanfir, and Cristian Sminchisescu. Deep multitask architecture for integrated 2d and 3d human sensing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6289-6298, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[51] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In ICCV, pages 4342-4351, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[52] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15190-15200, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[53] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In CVPR, pages 6040-6049, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[54] Pierre H Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, et al." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2511" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Byol works even without batch statistics. arXiv preprint arXiv:2010.10241, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[55] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, pages 5693-5703, 2019. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[56] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 529-545, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.47, + 0.263 + ], + "angle": 0, + "content": "[57] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 1, 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[59] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11148-11157, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[60] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, pages 466-481, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[61] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7093-7102, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[62] Jianfeng Zhang, Xuecheng Nie, and Jiashi Feng. Inference stage optimization for cross-scenario 3d human pose estimation. Advances in Neural Information Processing Systems, 33:2408-2419, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[63] Jianfeng Zhang, Yujun Cai, Shuicheng Yan, Jiashi Feng, et al. Direct multi-view multi-person 3d pose estimation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.672 + ], + "angle": 0, + "content": "[64] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. arXiv preprint arXiv:2108.02452, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[65] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2512" + } + ] +] \ No newline at end of file diff --git a/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_origin.pdf b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d2e8507cbe9384c6faaebdad7c1cc8d34c4ca8c --- /dev/null +++ b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/4659f454-3934-4652-8e58-699358f95d03_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e091ccbbfed3e7ae3adac14b525daa6054d6095b77ae88ed358bc6436a3284c +size 4160135 diff --git a/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/full.md b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1abd521beb03d0de7460dfe6f32001d7477b438 --- /dev/null +++ b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/full.md @@ -0,0 +1,326 @@ +# SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation + +Vinkle Srivastav $^{1,2*}$ Keqi Chen $^{1*}$ Nicolas Padoy $^{1,2}$ + +1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France +2IHU Strasbourg, France + +srivastav@unistra.fr keqi.chen@unistra.fr npadoy@unistra.fr + +# Abstract + +We present a new self-supervised approach, SelfPose3d, for estimating 3d poses of multiple persons from multiple camera views. Unlike current state-of-the-art fully-supervised methods, our approach does not require any 2d or 3d ground-truth poses and uses only the multiview input images from a calibrated camera setup and 2d pseudo poses generated from an off-the-shelf 2d human pose estimator. We propose two self-supervised learning objectives: self-supervised person localization in 3d space and self-supervised 3d pose estimation. We achieve self-supervised 3d person localization by training the model on synthetically generated 3d points, serving as 3d person root positions, and on the projected root-heatmaps in all the views. We then model the 3d poses of all the localized persons with a bottleneck representation, map them onto all views obtaining 2d joints, and render them using 2d Gaussian heatmaps in an end-to-end differentiable manner. Afterwards, we use the corresponding 2d joints and heatmaps from the pseudo 2d poses for learning. To alleviate the intrinsic inaccuracy of the pseudo labels, we propose an adaptive supervision attention mechanism to guide the self-supervision. Our experiments and analysis on three public benchmark datasets, including Panoptic, Shelf, and Campus, show the effectiveness of our approach, which is comparable to fully-supervised methods. Code is available at https://github.com/CAMMA-public/SelfPose3D. + +# 1. Introduction + +The task of estimating 3d poses for multiple persons using a few calibrated cameras is a challenging computer vision problem [10, 18, 32, 52, 57]. A significant part of this challenge lies in identifying and matching the same person across different camera views. The solutions developed so far generally use one of the two paradigms: learning-based + +methods and optimization-based methods. The learning-based methods develop novel deep-learning models and use 3d ground-truth poses for both training the models and establishing person correspondences across different views [39, 52, 57, 59, 63]. The accurate 3d ground-truth poses are typically generated using a dense camera system [32]. In contrast, the optimization-based methods formulate the 3d pose reconstruction as a mathematical optimization task, primarily focusing on aligning and matching the 2d poses across different camera views to infer 3d poses using triangulation within the framework of multi-view geometry [10, 18, 31-33, 49]. The 2d poses are estimated using off-the-shelf 2d human pose detectors [5, 55]. These methods apply geometric and spatial constraints in the optimization loop to ensure the anatomical plausibility and consistency of the inferred 3d poses. Although these methods do not require 3d ground-truth poses, their effectiveness is somewhat limited compared to the fully-supervised learning-based methods, see Table 1. + +In this paper, we explore the possibility of combining the strengths of both paradigms. Specifically, we investigate whether it's feasible to utilize a learning-based model for multi-view, multi-person 3D pose estimation and simultaneously eliminate its dependence on 3D ground-truth poses by incorporating geometric and appearance constraints, drawing inspiration from optimization-based methods. + +We propose, SelfPose3d, a self-supervised learning-based approach to estimate the 3d poses of multiple persons from a few calibrated cameras without using any 2d or 3d ground-truth poses. Our approach requires only 2d pseudo poses obtained using an off-the-shelf 2d pose detector [55]. Learning 3d poses without 3D ground-truth poses would require suitable supervisory signals to train a learning-based model. We follow the learning-by-projection paradigm, where the main idea is to learn the 3d output by comparing the projected bottleneck 3d output against the 2d input features [12]. + +We consider VoxelPose [57] as a learning-based method and use its output 3d poses as a bottleneck representation. To recover the accurate underlying 3d poses, we propose us- + +![](images/121056602f4f7bb080107f9f99ed3f302a87fe053607a093d29e93a77534b294.jpg) +Figure 1. Our self-supervised approach, called SelfPose3d, estimates multi-person 3d poses from multi-view images and pseudo 2d poses generated using an off-the-shelf 2d human pose estimator. We propose a self-supervised learning objective that generates differentiable and geometrically constrained 2d joints and heatmaps across multiple views from bottleneck 3d poses. On the right, we show 3d pose outputs from our approach along with estimated body meshes (using SMPL body mesh fitting on 3d poses [4, 41]) and the projected 2d poses. + +![](images/d23c63683cc97010ffee781c25bd2a8ac5e200461236303f50bb887f7d097456.jpg) + +ing differentiable multi-view 2d representations and cross-affine-view consistency. In particular, given a multi-view input image, we apply two random affine augmentations and pass them to the VoxelPose. It generates the bottleneck 3d poses corresponding to each affine augmented multi-view image. To enforce the model to learn and reason in the spatial dimension, we project the bottleneck 3d poses onto each view, obtaining 2d joints, and rendering them into spatial 2d heatmap representations in an end-to-end differentiable way. We further put tight geometric constraints by cross-affine-view operation, i.e. the bottleneck 3d poses from the 1st affine augmented multi-view image is mapped and rendered in the 2nd affine augmented multi-view image space and vice versa. Finally, we use the affine transformed 2d joints and heatmaps from the 2d pseudo poses to enable the geometrically constrained learning, with $L_{1}$ and $L_{2}$ losses respectively. + +As the 2d pseudo poses contain non-negligible noises (mostly due to occlusions, see Figure 3), we propose adaptive supervision attention to guide our model to focus on more reliable regions. We apply two strategies towards $L_{1}$ joint loss and $L_{2}$ heatmap loss; for $L_{1}$ joint loss supervision, we employ hard attention, where we ignore the one view with the largest absolute error for each multi-view image set; for $L_{2}$ heatmap loss supervision, we employ soft attention using a lighter backbone to process each view, obtaining same-size attention heatmaps. During $L_{2}$ loss computation, we compute the element-wise product of the attention heatmaps and the square error before averaging. To avoid obtaining zero attention, which the model tends to do, we add a regularization term, where we create tensors of all + +ones as the attention heatmap labels and use $L_{2}$ loss as the attention loss. + +Finally, specific to our choice of learning-based method, i.e., VoxelPose, which uses a voxel-based 3d root localization model to localize the persons in space using ground-truth 3d root joints (mid-hip joint), we use a simple but effective strategy to localize persons in space. Specifically, we randomly place 3d points in 3d world-space and project them to each view using the given camera parameters, subsequently rendering the projected 2d points as heatmap representations. This generates a synthetic dataset containing 3d points (roots) and their corresponding rendered multiview root-heatmaps. We then use this dataset to train a 3d root localization model that takes multi-view root-heatmaps as input and predicts the 3d roots as output. We further regularize the model by enforcing invariant constraints between pairs of affine augmented root-heatmaps coming from the real multi-view input. + +Evaluation on three 3d pose benchmarks datasets, Panoptic [32], Shelf [1] and Campus [1], along with extensive ablation studies on the Panoptic [32] dataset, show the effectiveness of our approach. Our approach reaches a performance comparable to learning-based fully-supervised approaches and performs significantly better than optimization-based approaches. Moreover, SMPL body mesh fitting [4, 41] on our estimated 3d poses generates geometrically plausible body shapes (see Figure 1 and Figure 4). + +We summarize our contributions as follows: 1) We address the challenging multi-person multi-view 3d person pose estimation problem using a self-supervised approach + +without any 2d or 3d ground truth. 2) We propose self-supervised 3d pose estimation by using a new method to recover geometrically constrained 2d joints and heatmap representations from the bottleneck 3d poses. 3) We propose adaptive supervision attention to address the misinformation caused by the inaccurate pseudo labels. 4) We propose self-supervised 3d root localization to estimate the 3d root location utilizing synthetic 3d roots and the corresponding rendered multi-view root heatmaps. + +# 2. Related work + +In this section, we briefly review current works related to fully-supervised learning-based methods for 3d pose estimation, optimization-based methods for 3D pose estimation, and self-supervised learning. + +Fully-supervised methods: Monocular 3d pose estimation [22, 43, 45, 47, 50, 56, 62, 65] is an ill-posed problem due to depth ambiguities as multiple 3d poses can produce same 2d pose projection. Having access to multi-view cameras can alleviate such depth ambiguities achieving the state-of-art results on benchmark datasets [23, 27, 28, 39, 51, 52, 57, 59, 63]. For single-person scenes, these approaches exploit multi-view geometry [24] to either fuse the visual features [27, 51], perform triangulation on heatmaps [28, 53], or use pictorial structural models for 3d reconstruction [48, 51]. The multi-person scene offers extra complexity due to the variability in the number of person in each view and the unknown cross-view correspondence. Existing multi-person multi-view approaches are based on volumetric paradigm [52, 57, 64], or direct regression [63] based on transformers [6, 58, 66]. Despite their good performance, these approaches rely on ground-truth 3d poses, which are generated using dense camera systems [32]. + +Optimization-based 3d pose estimation: For the multi-person and multi-view scenario, optimization-based approaches use an off-the-shelf person-id detector across all the views to solve the correspondence and triangulation problem and temporal refinement along with training a reinforcement learning agent to find the best camera locations for 3d pose reconstruction [49]. More recent approaches utilize multi-view 3d reconstruction in the optimization loop inferring 3D poses that are geometrically and spatially coherent [10, 18, 33]. + +Self-supervised learning: Self-supervised learning can be broadly classified into self-supervised representation learning and self-supervised task learning. Self-supervised representation learning aims to use large-scale unlabeled data to learn generic feature representations. The recent promising results from these approaches have started to surpass the fully-supervised baselines for various downstream tasks [7, 11, 26, 54]. Self-supervised task learning aims to learn a particular downstream task without using ground truth labels and has been applied to 2d pose estimation [29, 30], + +single-person 3d pose estimation [9, 19, 34, 36, 38], and surface correspondences estimation [3]. Self-supervised approaches for 3d pose estimation have primarily been developed for single-person scenarios. Given 2d poses, estimated by utilizing advances in the 2d pose estimation methods [5, 13, 14, 21, 35, 42, 44, 46, 55, 60], these approaches use the supervisory signals generated from multi-view geometry [34], video constraints [38], or adversarial learning [9, 19, 36]. + +Our work proposes a learning-based approach to model the 3d poses as bottleneck representations and recover geometrically constrained and spatially accurate 2d joints and heatmap representations in an end-to-end differentiable manner. + +# 3. Methodology + +# 3.1. Problem overview + +Given a training dataset of multi-view images $\mathcal{D} = \{\mathbf{x}|\mathbf{y}^*\}$ where $\mathbf{x} \in \mathcal{R}^{C \times 3 \times H \times W}$ is a multi-view image set from $C$ cameras with height $H$ and width $W$ , and $\mathbf{y}^* \in \mathcal{R}^{C \times P \times J \times 2}$ represents the 2d pseudo poses for $P$ persons with $J$ joints, the goal is to learn a deep learning model that estimates the 3d poses $\mathcal{V} \in \mathcal{R}^{P \times J \times 3}$ of all the $P$ persons from the multi-view input image $\mathbf{x}$ . It is to be noted that $P$ can vary in each camera view due to occlusion and noisy pseudo 2d pose estimation. For simplicity in the notation, we keep the same variable $\mathbb{P}$ . + +Fully-supervised approaches rely on 3d ground-truth poses, while we only have 2d pseudo poses $\mathbf{y}^*$ . Therefore, after obtaining 3d poses $\mathcal{V} \in \mathcal{R}^{P \times J \times 3}$ following traditional approach, we propose to project the poses to each view obtaining 2d poses $\mathbf{y} \in \mathcal{R}^{C \times P \times J \times 2}$ , and train the model from 2d pseudo poses $\mathbf{y}^* \in \mathcal{R}^{C \times P \times J \times 2}$ . + +In the following, we present our self-supervised approach based on fully-supervised VoxelPose [57]. We first generate pseudo 2d poses, and then propose self-supervised 3d root localization, self-supervised 3d pose estimation, and an adaptive supervision attention to learn 3d poses in a self-supervised manner, without modifying the original Voxel-Pose structure. + +# 3.2. Generating pseudo 2d poses + +To circumvent the dependence on the ground-truth 2d poses, we generate the 2d pseudo poses on the training dataset using Mask R-CNN [25] to generate person bounding boxes followed by using HRNet [55] to generate 2d poses of each detected person bounding box. The two-stage approach is chosen based on its state-of-art performance on the COCO dataset [40]. We first pre-train the 2d CNN backbone heatmap_net $_{2d}$ with pseudo 2d poses. + +![](images/15683e8d594f83f5225fd1681555b2ce1e29c3713f4977fe874c63317b726872.jpg) +SelfPose3d Framework +Figure 2. Illustrating our self-supervised SelfPose3d approaches for multi-view multi-person 3d pose estimation. Instead of using ground-truth 3d poses for learning, we propose self-supervised learning objectives to localize 3d roots (mid-hip location of the person) and estimate their 3d poses. We utilize a synthetic 3d roots dataset, two different affine transformations on the multi-view input images $(t_{r,s}^{1}, t_{r,s}^{2})$ parametrized by rotation $r$ and scale $s$ , a differentiable cross-affine-view 2d joints and heatmaps rendering from the bottleneck 3d poses, and an adaptive supervision attention mechanism to automatically learn the 3d poses in world-space. + +# 3.3. Self-supervised 3d root localization + +Given 2d multi-view heatmaps from all the views and all the joints $\mathcal{H} \in \mathcal{R}^{C \times J \times \frac{H}{4} \times \frac{W}{4}}$ estimated using a 2d backbone model heatmap。 $\mathrm{net}_{2\mathrm{d}}$ , we use [28] to construct a discretized 3d feature volume $\mathcal{F} \in \mathcal{R}^{J \times X \times Y \times Z}$ for each joint by un-projecting the 2d multi-view heatmaps to 3d space: + +$$ +\mathcal {P} _ {\text {u n p r o j}} (\operatorname {c a m}, \text {c e n t e r}, t _ {r, s}): \mathcal {H} \longrightarrow \mathcal {F}, \tag {1} +$$ + +To localize persons' root (mid-hip) joint in 3d space without using 3d ground truth, we hypothesize that 2d multi-view heatmaps of the root location $\mathcal{H}_{\mathrm{root}} \in \mathcal{R}^{C \times \frac{H}{4} \times \frac{W}{4}}$ are sufficient for 3d root localization (see Sec. 7.6 for verification). Then we generate the 3d feature volume for the root location $\mathcal{F}_{\mathrm{root}} \in \mathcal{R}^{X \times Y \times Z}$ using $\mathcal{H}_{\mathrm{root}}$ using Eq. (1), which has the same dimensions as predicted root-volumes $\mathcal{G}$ . This allows us to establish a one-to-one relationship between 3d root-volumes $\mathcal{G}$ and 2d multi-view root-heatmaps $\mathcal{H}_{\mathrm{root}}$ . We generate a synthetic root dataset $\mathcal{D}_{\mathrm{root}} = \{\mathcal{G}_i^{\mathrm{syn*}} | \mathcal{H}_{\mathrm{root\_i}}^{\mathrm{syn}}\}_{i=1}^N$ where $\mathcal{G}_i^{\mathrm{syn*}} \in \mathcal{R}^{X \times Y \times Z}$ contains the root-volumes of randomly placed 3d points, and $\mathcal{H}_{\mathrm{root\_i}}^{\mathrm{syn}}$ is the corresponding 2d multi-view heatmaps generated by projecting the random 3d points to each view using camera parameters cam. After unprojecting $\mathcal{H}_{\mathrm{root}}^{\mathrm{syn}}$ to $\mathcal{F}_{\mathrm{root}}^{\mathrm{syn}}$ , and passing it through root\_net obtaining $\mathcal{G}_{\mathrm{root}}^{\mathrm{syn}}$ , we compute the $L_2$ loss error as the synthetic root loss $l_{\mathrm{root\_syn}}$ : + +$$ +l _ {\text {r o o t - s y n}} = \mathcal {L} _ {2} \left(\mathcal {G} ^ {\text {s y n}}, \mathcal {G} ^ {\text {s y n *}}\right) \tag {2} +$$ + +To further regularize root_net on the real-world 2d multi-view root-heatmaps, we propose the root consistency loss. Given a multi-view training image set $x^0$ , we apply two affine transformations $(t_{r,s}^{1}, t_{r,s}^{2})$ with random rotation and scaling $(r, s)$ to generate two affine transformed multiview images $(x^1, x^2)$ . We pass $x^0$ , $x^1$ and $x^2$ through heatmap_net2d, construct the root feature volumes using Eq. (1) with corresponding affine transformation parameters, and finally obtain $\mathcal{G}^0$ , $\mathcal{G}^1$ and $\mathcal{G}^2$ through root_net. Since $\mathcal{G}^1$ and $\mathcal{G}^2$ are invariant to the applied affine transformations $t_{r,s}^1$ and $t_{r,s}^2$ , we use $\mathcal{G}^0$ as the baseline to compute the $L_2$ loss error between $\mathcal{G}^0$ , $\mathcal{G}^1$ and $\mathcal{G}^2$ as the root consistency loss $l_{\mathrm{root\_C}}$ : + +$$ +l _ {\text {r o o t} C} = \mathcal {L} _ {2} \left(\mathcal {G} ^ {0}, \mathcal {G} ^ {1}\right) + \mathcal {L} _ {2} \left(\mathcal {G} ^ {0}, \mathcal {G} ^ {2}\right) \tag {3} +$$ + +We train root_net by minimizing Eq. (2) and Eq. (3). We generate the person proposals $\{\mathrm{root}_i\}_{i=1}^K$ by applying nonmaximum suppression (NMS) and thresholding on $\mathcal{G}^2$ ( $\mathcal{G}^1$ would also work). + +# 3.4. Self-supervised 3d pose estimation + +Given pseudo 2d poses $\mathbf{y}_{2\mathrm{d}}^{*}$ , person proposals $\{\mathrm{root}_i\}_{i = 1}^P$ and 2d multi-view heatmaps $\mathcal{H}^1,\mathcal{H}^2$ predicted using heatmap_net2d, we describe our self-supervised 3d pose estimation approach. + +The person proposals $\{\mathrm{root}_i\}_{i = 1}^P$ are used to generate the 3d feature volumes i.e $\left\{\mathcal{F}_i^1\right\}_{i = 1}^P = \left\{\mathcal{P}_{\mathrm{unproj}}(\mathrm{cam},r_i,t_{r,s}^1)\right\}_{i = 1}^P$ and $\left\{\mathcal{F}_i^2\right\}_{i = 1}^P =$ + +$\left\{\mathcal{P}_{\mathrm{unproj}}(\mathrm{cam},r_i,t_{r,s}^2)\right\}_{i = 1}^P$ corresponding to the person feature volumes for each affine augmented multi-view input image $x^{1}$ and $x^{2}$ , respectively. $\left\{\mathcal{F}_i^1\right\}_{i = 1}^P$ and $\left\{\mathcal{F}_i^2\right\}_{i = 1}^P$ are passed through pose_net3d and soft-argmax [8] to estimate the 3d poses $\mathcal{V}^1,\mathcal{V}^2\in \mathcal{R}^{P\times J\times 3}$ . These 3d poses serve as a bottleneck representation. + +Given the camera parameters cam and the affine transformation parameters $t_{r,s}^{1}, t_{r,s}^{2}$ , we project the bottleneck 3d poses in cross-affine-view i.e. $\mathcal{V}^1$ are projected to $x^2$ image space using $t_{r,s}^{2}$ to generate multi-view 2d poses $\mathbf{y}^2 \in \mathcal{R}^{C \times P \times J \times 2}$ , and $\mathcal{V}^2$ are projected to the $x^1$ image space using $t_{r,s}^{1}$ to generate multi-view 2d poses $\mathbf{y}^1 \in \mathcal{R}^{C \times P \times J \times 2}$ . + +We propose to render $\mathbf{y}^1$ and $\mathbf{y}^2$ 2d poses in the 2d heatmap representation. The heatmap representation encodes the per-pixel likelihood of a body joint and has been a vital component to enable the state-of-the-art 2d pose estimation approaches [55]. Generating heatmap representation has essentially been a pre-processing step where state-of-the-art approaches quantize the 2d joints before generating the heatmap [55]. However, this quantization step is non-differentiable and could cut the backward gradient flow. Zhang et al. [61] show that encoding floating point 2d joints into heatmap representation in their pre-processing step leads to improved performance. We propose to use the same differentiable approach in an online way to render the projected 2d joints into the heatmap representation. + +We render $\mathbf{y}^1$ and $\mathbf{y}^2$ into heatmap representation to generate 2d multi-view heatmaps $\mathcal{H}^1$ , and $\mathcal{H}^2$ , respectively. We apply the affine transformations $t_{r,s}^1, t_{r,s}^2$ on the pseudo 2d poses $\mathbf{y}_{2\mathrm{d}}^{*}$ to generate pseudo 2d multi-view joints $\mathbf{y}_{2\mathrm{d}}^{1*}, \mathbf{y}_{2\mathrm{d}}^{2*}$ and heatmaps $\mathcal{H}^{1*}, \mathcal{H}^{2*}$ . Then, we compute the $L_2$ loss between heatmaps as the pose heatmap loss $l_{\mathrm{pose\_H}}$ : + +$$ +l _ {\text {p o s e} \mathrm {H}} = \mathcal {L} _ {2} \left(\mathcal {H} ^ {1}, \mathcal {H} ^ {1 *}\right) + \mathcal {L} _ {2} \left(\mathcal {H} ^ {2}, \mathcal {H} ^ {2 *}\right) \tag {4} +$$ + +After training with $l_{\mathrm{pose\_H}}$ preliminarily, we propose to add the $L_{1}$ loss between 2d joints to further fine-tune the model. For each view, we employ the Hungarian algorithm [37] to obtain the optimal assignment between $\mathbf{y}$ and $\mathbf{y}_{2\mathrm{d}}^{*}$ , where the matching cost is the mean absolute error. Based on the optimal assignment, we obtain the $L_{1}$ loss as $l_{\mathrm{pose\_J}}$ . Then we use $l_{\mathrm{pose\_H}}$ and $l_{\mathrm{pose\_J}}$ together to train the whole network, where $\lambda$ is a manually defined weight: + +$$ +l _ {\text {p o s e - J}} = \mathcal {L} _ {1} \left(\mathbf {y} ^ {1}, \mathbf {y} _ {2 \mathrm {d}} ^ {1 *}\right) + \mathcal {L} _ {1} \left(\mathbf {y} ^ {2}, \mathbf {y} _ {2 \mathrm {d}} ^ {2 *}\right) \tag {5} +$$ + +$$ +l _ {\text {p o s e - 3 d}} = l _ {\text {p o s e - H}} + \lambda l _ {\text {p o s e - J}} \tag {6} +$$ + +As the network needs to reason about the 2d joint locations in spatial dimension, it implicitly solves the person correspondence problem. Training pose_net3d with 3d pose loss $l_{\mathrm{pose\_3d}}$ performs decently, but to achieve even better results, we introduce the adaptive supervision attention. + +![](images/27a3d8b7034de1a41fcf05a0ea8360107c945bc08060162bc16a013846740d43.jpg) +ground truth 2d poses +Figure 3. Comparing ground-truth 2d poses generated by projecting the ground-truth 3d poses to each multi-view image and our pseudo 2d poses generated by running HRNet human pose estimation model [55] on the training dataset. Pseudo 2d poses contain localization errors due to occlusion (see the red arrows), and ground-truth 2d poses exist for partially or even entirely occluded persons (see the blue dotted arrows). + +![](images/ec0de6ec05ddac0d1a71ef0f200a5a00a5df47cf062f8753ec123a2a6907381b.jpg) +pseudo 2d poses + +# 3.5. Adaptive supervision attention + +Traditional $L_{1}$ and $L_{2}$ losses treat each label equally, which is sub-optimal in two aspects: (1) the 2d human pose detector may generate inaccurate labels due to occlusions (see the red arrows in Figure 3); (2) the 3d-2d projection will output 2d joints in certain views even when the person is entirely occluded (see the blue dotted arrows in Figure 3). Therefore, we propose to employ attentions to adaptively guide the supervision process. + +For $L_{2}$ loss supervision, we use the soft attention. Specifically, we use ResNet-18 to extract the visual features of the views, followed by deconvolutional layers to obtain the attention heatmaps $\mathcal{A}$ (see attn\_net $_{2d}$ in Figure 2). Then we compute the element-wise product of $\mathcal{A}$ and the square error before averaging, as the new loss $l_{\mathrm{pose\_H}}^{\mathrm{attn}}$ : + +$$ +l _ {\text {p o s e - H}} ^ {\mathrm {a t t n}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {A} _ {i} \otimes \left(\mathcal {H} _ {i} - \mathcal {H} _ {i} ^ {*}\right) ^ {2} \tag {7} +$$ + +To avoid that $\mathcal{A}$ becomes zero, we add a regularization term. We create tensors of all ones $\mathbb{1}$ as the attention heatmap labels, and compute $L_{2}$ error as the attention loss $l_{\mathrm{attn}}$ . If $l_{\mathrm{attn}}$ becomes zero, $l_{\mathrm{pose\_H}}^{\mathrm{attn}}$ degrades to non-attentive version. + +$$ +l _ {\mathrm {a t t n}} = \mathcal {L} _ {2} (\mathcal {A}, \mathbb {1}) \tag {8} +$$ + +For $L_{1}$ loss supervision, we use the hard attention. For each input with $K$ views, we compute the $L_{1}$ loss of each view, find the view with the largest loss, and ignore it when averaging the final loss $l_{\mathrm{pose\_J}}^{\mathrm{att}}$ : + +$$ +j = \underset {i} {\arg \max } \mathcal {L} _ {1} \left(\mathbf {y} _ {i}, \mathbf {y} _ {i, 2 \mathrm {d}} ^ {*}\right) \quad (i = 1, 2, \dots , K) \tag {9} +$$ + +$$ +l _ {\text {p o s e - J}} ^ {\mathrm {a t t n}} = \frac {1}{K - 1} \sum_ {i = 1, i \neq j} ^ {K} \mathcal {L} _ {1} \left(\mathbf {y} _ {i}, \mathbf {y} _ {i, 2 \mathrm {d}} ^ {*}\right) \tag {10} +$$ + +In general, the final 3d pose loss $l_{\mathrm{pose\_3d}}^{\mathrm{attn}}$ is as follows, where $\lambda$ and $\sigma$ are manually defined weights: + +$$ +l _ {\text {p o s e - 3 d}} ^ {\text {a t t n}} = l _ {\text {p o s e - H}} ^ {\text {a t t n}} + \lambda l _ {\text {p o s e - J}} ^ {\text {a t t n}} + \sigma l _ {\text {a t t n}} \tag {11} +$$ + +We train pose_net3d by minimizing $l_{\mathrm{pose\_3d}}^{\mathrm{attn}}$ . Our self-supervised approach is visually described in Figure 2. + +# 3.6. Implementation details + +Training strategies For the Panoptic dataset, similar to VoxelPose [57], we first train heatmap_net2d for 20 epochs with pseudo 2d poses. We use the Adam optimizer with an initial learning rate of 1e-4, which decreases to 1e-5 and 1e-6 at the 10th and 15th epochs, respectively. Then, we train the root_net for 1 epoch, followed by end-to-end joint training of the whole network for 5 epochs using only the $L_{2}$ loss, with a learning rate of 1e-4. Afterwards, we add $L_{1}$ loss to train the whole network for another 5 epochs with a learning rate of 5e-5. $\lambda$ and $\sigma$ in Eq. (11) are set to 0.01 and 0.1 respectively. + +We use the random rotation between $-45^{\circ}$ to $45^{\circ}$ and random scale between $-0.35$ to 0.35. We also apply spatial augmentations using rand-augment [16] and rand-cutout [17] using python image library1. The rand-augment consists of "contrast-jittering", "auto-contrast", "equalize", "color-jittering", "sharpness-jittering", and "brightness-jittering", and the rand-cutout places random square boxes of sizes between 20 to 40 pixels at random locations in the image. We use the SMPL model and optimization-based body fitting approach2[4, 41] to estimate body mesh parameters. + +Inference pipeline During inference, we input the multiview RGB images, and obtain the estimated 3d poses in an end-to-end pipeline. For each view, the backbone generates corresponding 2d heatmaps for cuboid construction. Then, given constructed cuboid of the whole space, the root_net predicts root joint locations of all persons. Finally, the pose_net outputs the regressed 3d locations of each joint for every cuboid proposal of the root joints. + +# 4. Experiments + +# 4.1. Datasets and evaluation metrics + +We conduct experiments on three benchmark datasets: Panoptic [32], Campus [1], and Shelf [1]. + +
MethodsAP25AP50AP100AP150Recall@500MPJPE[mm]
FSVoxelPose [57]83.698.399.899.998.817.7
Lin et al. [39]92.199.099.899.8-16.8
MvP [63]92.396.697.597.798.215.8
Wu et al. [59]----98.715.8
TEMPO [15]89.099.199.899.9-14.7
OBACTOR [49]-----168.4
MvPose [18]0.02.9759.9381.5398.2384.2
SSSelfPose3d (ours)55.196.498.599.099.624.5
+ +Table 1. Result on the Panoptic dataset (FS = fully-supervised, OB = optimization-based, SS = self-supervised). + +
MethodsShelfCampus
Actor 1Actor 2Actor 3AverageActor 1Actor 2Actor 3Average
FSErshadi et al. [20]93.375.994.888.094.292.984.690.6
Wu et al. [59]99.396.597.397.7----
MvP [63]99.395.197.897.498.294.197.496.6
VoxelPose [57]99.394.197.697.097.693.898.896.7
VoxelPose* [57]99.593.597.896.993.186.593.290.9
OB3DPS [2]75.369.787.677.593.575.784.484.5
MvPose [18]98.894.197.896.997.693.398.096.3
SSSelfPose3d97.290.397.995.192.582.289.287.9
+ +Table 2. Results (in PCP) on Shelf and Campus datasets (FS = fully-supervised, OB = optimization-based, SS = self-supervised, * = reproduced results). SelfPose3d is trained from the pseudo 3d poses from the Panoptic training set. + +The Panoptic dataset is a large-scale dataset captured inside a dome environment containing multiple persons performing daily social activities. We conduct extensive experiments on this dataset to evaluate and assess various components of our approach. We use the same data sequences for training and testing as VoxelPose [57] except that our training set doesn't include '160906_band3'. In other words, we are only using 9 multi-view videos for training (the '160906_band3' video is not available due to the broken images on the source website). We use the five HD camera images (3, 6, 12, 13, 23) to train and report the performance in our experiments. We use Average Precision (AP), Recall, and Mean Per Joint Position Error (MPJPE) in millimeters (mm) as evaluation metrics (higher AP and lower MPJPE are better) [57]. + +The Shelf and Campus are two multi-person datasets capturing activities in the indoor and outdoor environments, respectively [1]. We use the same training and test split as [1, 57]. As used in [1, 57], we use the Percentage of Correct Parts (PCP) as evaluation metrics for these two datasets. + +Panoptic Table 1 shows the results on the challenging Panoptic dataset. All the fully-supervised approaches utilizing 2d and 3d ground-truth 3d poses reach nearly the same performance. SelfPose3d, without using any 3d or 2d ground-truth poses, achieves comparable results to fully-supervised approaches. Nevertheless, there still exists a gap compared to the fully-supervised VoxelPose model (96.4 v.s 98.3 AP50 and 24.5 v.s 17.7 MPJPE). However, unlike VoxelPose, which relies on heat maps from all the joints to es + +![](images/284dc7761ab5c730b5f21a0f5c57b543155f46c8bf22803f2fd71a6d2ac22fa4.jpg) +Figure 4. Qualitative results for the 3d pose estimations, 2d projections on the multi-view images, and estimated SMPL body shapes on some example images from the Panoptic dataset + +timate 3d roots, we only use root-heatmaps to do the same. This results in the reduction of the input channel from 15 (number of keypoints) to 1 for the root_net, making our approach computationally faster. + +We also compare our approach with optimization-based baselines from Pirinen et al. [49] and Dong et al. [18]. These non-learning-based approaches fail to capture the multi-person interaction in a complex scene from a few sparse multi-view cameras. Our learning-based self-supervised approach achieves much better performance. It is to be noted that Pirinen et al. evaluate their approach on two multi-person sequences, whereas we evaluate on four multi-person sequences. + +Shelf and Campus We compare our approach with the state-of-the-art methods on the Shelf and Campus dataset. VoxelPose uses the 3d ground-truth from the Panoptic dataset to train their approach to these datasets due to noisy and incomplete 3d ground-truth poses. For a fair comparison with VoxelPose, we use the pseudo 3d poses (by running SelfPose3d on the Panoptic training set) and train on these two datasets in a fully supervised manner. As shown in Table 2, our approach using pseudo 3d poses from the Panoptic dataset also reaches the same performance as the fully-supervised approaches. + +Qualitative visualizations Figure 4 shows 3d pose estimation results from our SelfPose3d approach on the challenging Panoptic dataset. Without using any 3d ground-truth, we can see that SelfPose3d is robust to occlusions and multiple persons while correctly identifying the person identities across all the views (see the corresponding 2d projections in Figure 4). We also show the qualitative results for the SMPL body mesh fitting [4, 41] on the estimated 3d poses. All these results demonstrate both the effectiveness and extendability of SelfPose3d. Please see the supplementary for more results. + +# 4.2. Ablation studies + +Ground-truth 2d poses v.s pseudo 2d poses As shown in Table 3, when we use the ground-truth 2d poses in our self-supervised framework, 3d reconstruction error significantly reduces. To inspect the better performance when using the ground-truth 2d poses, we qualitatively compare the ground-truth 2d poses with the pseudo 2d poses on some example training images. Pseudo 2d poses contain localization errors due to occlusion, whereas ground-truth 2d poses exist for partially or even entirely occluded persons as shown in Figure 3. As the ground-truth 2d poses are generated by projecting the ground-truth 3d poses to each multi-view image, they serve as a suitable proxy for the 3d poses, thereby reaching a performance close to the fully-supervised approaches. However, obtaining the ground 2d + +
2d posesAP50AP100MPJPE
ground-truth98.899.619.9
pseudo96.498.524.5
+ +Table 3. The ground-truth 2d poses in our self-supervised framework decrease the 3d reconstruction error and reach the performance close to the fully-supervised approaches. + +
cross-affine-view consistencyaffine augsAP50AP100MPJPE
86.096.234.7
✓83.397.533.3
✓✓93.898.129.3
+ +Table 4. Affine augmentations and cross-affine-view consistency significantly improves the 3d pose reconstruction accuracy. All three models are trained for two epochs with frozen backbone and frozen root_net and no attention. + +
L1 lossL2 lossAP25AP50AP100MPJPE
✓----
✓43.895.898.225.7
✓✓55.196.498.524.5
+ +poses in this way would be as challenging as acquiring the ground-truth 3d poses. + +Importance of cross-affine-view consistency and affine augmentations We also examine the effect of affine augmentations on the multi-view images and cross-affine-view consistency when generating differentiable 2d representations from the bottleneck 3d poses. As shown in Table 4, the affine augmentations and cross-affine-view consistency significantly improve the 3d pose reconstruction as they provide necessary geometric constraints during training. + +Analysis of $L_{2}$ and $L_{1}$ pose losses We conduct experiments to analyze the use of $L_{1}$ and $L_{2}$ pose losses. As shown in Table 5, using $L_{2}$ and $L_{1}$ losses together can obtain better results than using $L_{2}$ loss solely. Also, using $L_{1}$ loss solely doesn't converge due to the label noises. + +Importance of adaptive supervision attention We also examine the necessity of adaptive supervision attention. Table 6 shows that supervision attention for both $L_{1}$ and $L_{2}$ losses are necessary for training. + +Influence of different 2d human pose estimation models Finally, we show how pseudo 2d poses generated from different 2d human pose estimation models affect the performance. As shown in Table 7, models that perform well on the COCO dataset [40] also generate better pseudo 2d poses + +Table 5. Training using $L_{1}$ and $L_{2}$ pose losses together achieves the best performance. + +
L1 loss attentionL2 loss attentionAP25AP50AP100MPJPE
32.594.197.828.5
✓37.995.898.026.3
✓47.496.698.225.0
✓✓55.196.498.524.5
+ +Table 6. Training using $L_{1}$ and $L_{2}$ loss supervisions together achieves the best performance. + +
Method for 2d pseudo pose generationAP50AP100MPJPEKeypoint AP on COCO-val[40]
Keypoint R-CNN (R-101) [25]89.297.631.966.1
HRNet (w48 384x288) [55]93.898.129.376.3
+ +Table 7. Comparing different models for generating pseudo 2d poses. Models that perform well on the COCO dataset [40] also generate better pseudo 2d poses for the Panoptic dataset, helping SelfPose3d to achieve better performance. + +for the Panoptic dataset, helping SelfPose3d to achieve better performance. + +# 5. Conclusion + +We present a self-supervised approach, called SelfPose3d, to address the challenging problem of multi-view multiperson 3d human pose estimation. Unlike current state-of-the-art methods that use difficult-to-acquire 3d ground-truth poses to train the model, SelfPose3d requires only multiview input images and an off-the-shelf 2d human pose detector. We propose a novel self-supervised learning objective that aims to recover 2d joints and heatmaps under different affine transformations from the bottleneck 3d poses. We further improve the performance of our approach by integrating adaptive supervision attention to address the misinformation caused by the inaccurate 2d pseudo labels from the off-the-shelf 2d human pose detector. We conduct extensive experiments on large-scale benchmark datasets, assess various components of our approach, and show that SelfPose3d reaches a performance on par with the well-established fully-supervised baselines. We visualize the 3d pose reconstruction in the complex multiple-person scenes and show that body shape meshes fitted on the estimated 3d poses look geometrically plausible under different viewpoints. + +# 6. Acknowledgements + +This work was partially supported by French state funds managed by the ANR under references ANR-20-CHIA0029-01 (National AI Chair AI4ORSafety), ANR-10IAHU-02 (IHU Strasbourg), ANR-18-CE45-0011-03 (OptimiX), and by BPI France (project 5G-OR). This work was also granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3]. + +# References + +[1] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures for multiple human pose estimation. In CVPR, pages 1669-1676, 2014. 2, 6 +[2] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. TPAMI, 38(10):1929-1942, 2015. 6 +[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 3 +[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578. Springer, 2016. 2, 6, 7 +[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, pages 7291-7299, 2017. 1, 3 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3 +[8] Olivier Chapelle and Mingrui Wu. Gradient descent optimization of smoothed information retrieval metrics. Information retrieval, 13(3):216-235, 2010. 5 +[9] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, M. V. Rohith, Stefan Stojanov, and James M. Rehg. Unsupervised 3d pose estimation with geometric self-supervision. CoRR, abs/1904.04812, 2019. 3 +[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. In European Conference on Computer Vision, pages 541-557. Springer, 2020. 1, 3 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3 +[12] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 1 +[13] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 3 + +[14] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020. 3 +[15] Rohan Choudhury, Kris M Kitani, and László A Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14750-14760, 2023. 6 +[16] Ekin D Cubuk, Barret Zoph, Jonathon Shiens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6 +[17] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6 +[18] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In CVPR, pages 7792-7801, 2019. 1, 3, 6, 7 +[19] Dylan Drover, Ching-Hang Chen, Amit Agrawal, Ambrish Tyagi, and Cong Phuoc Huynh. Can 3d pose be learned from 2d projections alone? In Proceedings of the European Conference on Computer Vision (ECCV), pages 0–0, 2018. 3 +[20] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77(12): 15573-15601, 2018. 6 +[21] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2334-2343, 2017. 3 +[22] Kehong Gong, Jianfeng Zhang, and Jiashi Feng. Poseaug: A differentiable pose augmentation framework for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8575-8584, 2021. 3 +[23] Brian Gordon, Sigal Raab, Guy Azov, Raja Giryes, and Daniel Cohen-Or. Flex: Parameter-free multiview 3d human motion reconstruction. arXiv preprint arXiv:2105.01937, 2021. 3 +[24] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3 +[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 8 +[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3 +[27] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7779-7788, 2020. 3 + +[28] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. arXiv preprint arXiv:1905.05754, 2019. 3, 4 +[29] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3 +[30] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3 +[31] Hanbyul Joo, Hyun Soo Park, and Yaser Sheikh. Map visibility estimation for large-scale dynamic 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1122-1129, 2014. 1 +[32] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In Proceedings of the IEEE International Conference on Computer Vision, pages 3334-3342, 2015. 1, 2, 3, 6 +[33] Abdelrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):1-14, 2021. 1, 3 +[34] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1077-1086, 2019. 3 +[35] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR, pages 11977-11986, 2019. 3 +[36] Yasunori Kudo, Keisuke Ogaki, Yusuke Matsui, and Yuri Odagiri. Unsupervised adversarial learning of 3d human pose from 2d joint locations. arXiv preprint arXiv:1803.08244, 2018. 3 +[37] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5 +[38] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6152-6162, 2020. 3 +[39] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11886-11895, 2021. 1, 3, 6 +[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 3, 8 +[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi + +person linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 6, 7 +[42] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 3 +[43] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2640-2649, 2017. 3 +[44] William McNally, Kanav Vats, Alexander Wong, and John McPhee. Evopose2d: Pushing the boundaries of 2d human pose estimation using neuroevolution. arXiv preprint arXiv:2011.08446, 2020. 3 +[45] Dushyant Mehta, Srinath Sridhar, Oleksandr Sotnychenko, Helge Rhodin, Mohammad Shafiei, Hans-Peter Seidel, Weipeng Xu, Dan Casas, and Christian Theobalt. Vnect: Real-time 3d human pose estimation with a single rgb camera. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 3 +[46] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. In NIPS, pages 2277-2287, 2017. 3 +[47] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 3 +[48] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In CVPR, pages 1253-1262, 2017. 3 +[49] Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Domes to drones: Self-supervised active triangulation for 3d human pose reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 6, 7 +[50] Alin-Ionut Popa, Mihai Zanfir, and Cristian Sminchisescu. Deep multitask architecture for integrated 2d and 3d human sensing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6289-6298, 2017. 3 +[51] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In ICCV, pages 4342-4351, 2019. 3 +[52] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15190-15200, 2021. 1, 3 +[53] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In CVPR, pages 6040-6049, 2020. 3 +[54] Pierre H Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, et al. + +Byol works even without batch statistics. arXiv preprint arXiv:2010.10241, 2020.3 +[55] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, pages 5693-5703, 2019. 1, 3, 5, 8 +[56] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 529-545, 2018. 3 +[57] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 1, 3, 6, 2 +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[59] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11148-11157, 2021. 1, 3, 6 +[60] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, pages 466-481, 2018. 3 +[61] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7093-7102, 2020. 5 +[62] Jianfeng Zhang, Xuecheng Nie, and Jiashi Feng. Inference stage optimization for cross-scenario 3d human pose estimation. Advances in Neural Information Processing Systems, 33:2408-2419, 2020. 3 +[63] Jianfeng Zhang, Yujun Cai, Shuicheng Yan, Jiashi Feng, et al. Direct multi-view multi-person 3d pose estimation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6 +[64] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. arXiv preprint arXiv:2108.02452, 2021. 3 +[65] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In ICCV, 2017. 3 +[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3 \ No newline at end of file diff --git a/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/images.zip b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0d9f567ea89ba493e335f92473d12898175e3dea --- /dev/null +++ b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4835df995029bdf0f4c534a5f07325f18abf3b2de60600ae61690a461527d050 +size 525518 diff --git a/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/layout.json b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..30d1ed025c4bdedf57f8f9b57f5e49c13aa2c1d7 --- /dev/null +++ b/2024/SelfPose3d_ Self-Supervised Multi-Person Multi-View 3d Pose Estimation/layout.json @@ -0,0 +1,10147 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 120 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 120 + ], + "type": "text", + "content": "SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "spans": [ + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "type": "text", + "content": "Vinkle Srivastav" + }, + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "type": "text", + "content": " Keqi Chen" + }, + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "type": "text", + "content": " Nicolas Padoy" + }, + { + "bbox": [ + 160, + 142, + 430, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 161, + 465, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 161, + 465, + 188 + ], + "spans": [ + { + "bbox": [ + 128, + 161, + 465, + 188 + ], + "type": "text", + "content": "1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France \n2IHU Strasbourg, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 191, + 465, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 191, + 465, + 202 + ], + "spans": [ + { + "bbox": [ + 126, + 191, + 465, + 202 + ], + "type": "text", + "content": "srivastav@unistra.fr keqi.chen@unistra.fr npadoy@unistra.fr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 255, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 579 + ], + "type": "text", + "content": "We present a new self-supervised approach, SelfPose3d, for estimating 3d poses of multiple persons from multiple camera views. Unlike current state-of-the-art fully-supervised methods, our approach does not require any 2d or 3d ground-truth poses and uses only the multiview input images from a calibrated camera setup and 2d pseudo poses generated from an off-the-shelf 2d human pose estimator. We propose two self-supervised learning objectives: self-supervised person localization in 3d space and self-supervised 3d pose estimation. We achieve self-supervised 3d person localization by training the model on synthetically generated 3d points, serving as 3d person root positions, and on the projected root-heatmaps in all the views. We then model the 3d poses of all the localized persons with a bottleneck representation, map them onto all views obtaining 2d joints, and render them using 2d Gaussian heatmaps in an end-to-end differentiable manner. Afterwards, we use the corresponding 2d joints and heatmaps from the pseudo 2d poses for learning. To alleviate the intrinsic inaccuracy of the pseudo labels, we propose an adaptive supervision attention mechanism to guide the self-supervision. Our experiments and analysis on three public benchmark datasets, including Panoptic, Shelf, and Campus, show the effectiveness of our approach, which is comparable to fully-supervised methods. Code is available at https://github.com/CAMMA-public/SelfPose3D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 622, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 622, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 288, + 696 + ], + "type": "text", + "content": "The task of estimating 3d poses for multiple persons using a few calibrated cameras is a challenging computer vision problem [10, 18, 32, 52, 57]. A significant part of this challenge lies in identifying and matching the same person across different camera views. The solutions developed so far generally use one of the two paradigms: learning-based" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 231, + 547, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 547, + 459 + ], + "type": "text", + "content": "methods and optimization-based methods. The learning-based methods develop novel deep-learning models and use 3d ground-truth poses for both training the models and establishing person correspondences across different views [39, 52, 57, 59, 63]. The accurate 3d ground-truth poses are typically generated using a dense camera system [32]. In contrast, the optimization-based methods formulate the 3d pose reconstruction as a mathematical optimization task, primarily focusing on aligning and matching the 2d poses across different camera views to infer 3d poses using triangulation within the framework of multi-view geometry [10, 18, 31-33, 49]. The 2d poses are estimated using off-the-shelf 2d human pose detectors [5, 55]. These methods apply geometric and spatial constraints in the optimization loop to ensure the anatomical plausibility and consistency of the inferred 3d poses. Although these methods do not require 3d ground-truth poses, their effectiveness is somewhat limited compared to the fully-supervised learning-based methods, see Table 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 544 + ], + "type": "text", + "content": "In this paper, we explore the possibility of combining the strengths of both paradigms. Specifically, we investigate whether it's feasible to utilize a learning-based model for multi-view, multi-person 3D pose estimation and simultaneously eliminate its dependence on 3D ground-truth poses by incorporating geometric and appearance constraints, drawing inspiration from optimization-based methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 675 + ], + "type": "text", + "content": "We propose, SelfPose3d, a self-supervised learning-based approach to estimate the 3d poses of multiple persons from a few calibrated cameras without using any 2d or 3d ground-truth poses. Our approach requires only 2d pseudo poses obtained using an off-the-shelf 2d pose detector [55]. Learning 3d poses without 3D ground-truth poses would require suitable supervisory signals to train a learning-based model. We follow the learning-by-projection paradigm, where the main idea is to learn the 3d output by comparing the projected bottleneck 3d output against the 2d input features [12]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "We consider VoxelPose [57] as a learning-based method and use its output 3d poses as a bottleneck representation. To recover the accurate underlying 3d poses, we propose us-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 194, + 712 + ], + "type": "text", + "content": "* co-first authors with equal contributions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "2502" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 259, + 269 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 259, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 259, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 259, + 269 + ], + "type": "image", + "image_path": "121056602f4f7bb080107f9f99ed3f302a87fe053607a093d29e93a77534b294.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 547, + 323 + ], + "type": "text", + "content": "Figure 1. Our self-supervised approach, called SelfPose3d, estimates multi-person 3d poses from multi-view images and pseudo 2d poses generated using an off-the-shelf 2d human pose estimator. We propose a self-supervised learning objective that generates differentiable and geometrically constrained 2d joints and heatmaps across multiple views from bottleneck 3d poses. On the right, we show 3d pose outputs from our approach along with estimated body meshes (using SMPL body mesh fitting on 3d poses [4, 41]) and the projected 2d poses." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 71, + 533, + 268 + ], + "blocks": [ + { + "bbox": [ + 276, + 71, + 533, + 268 + ], + "lines": [ + { + "bbox": [ + 276, + 71, + 533, + 268 + ], + "spans": [ + { + "bbox": [ + 276, + 71, + 533, + 268 + ], + "type": "image", + "image_path": "d23c63683cc97010ffee781c25bd2a8ac5e200461236303f50bb887f7d097456.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 335, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 539 + ], + "type": "text", + "content": "ing differentiable multi-view 2d representations and cross-affine-view consistency. In particular, given a multi-view input image, we apply two random affine augmentations and pass them to the VoxelPose. It generates the bottleneck 3d poses corresponding to each affine augmented multi-view image. To enforce the model to learn and reason in the spatial dimension, we project the bottleneck 3d poses onto each view, obtaining 2d joints, and rendering them into spatial 2d heatmap representations in an end-to-end differentiable way. We further put tight geometric constraints by cross-affine-view operation, i.e. the bottleneck 3d poses from the 1st affine augmented multi-view image is mapped and rendered in the 2nd affine augmented multi-view image space and vice versa. Finally, we use the affine transformed 2d joints and heatmaps from the 2d pseudo poses to enable the geometrically constrained learning, with " + }, + { + "bbox": [ + 46, + 335, + 289, + 539 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 335, + 289, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 289, + 539 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 335, + 289, + 539 + ], + "type": "text", + "content": " losses respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": "As the 2d pseudo poses contain non-negligible noises (mostly due to occlusions, see Figure 3), we propose adaptive supervision attention to guide our model to focus on more reliable regions. We apply two strategies towards " + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": " joint loss and " + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": " heatmap loss; for " + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": " joint loss supervision, we employ hard attention, where we ignore the one view with the largest absolute error for each multi-view image set; for " + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": " heatmap loss supervision, we employ soft attention using a lighter backbone to process each view, obtaining same-size attention heatmaps. During " + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": " loss computation, we compute the element-wise product of the attention heatmaps and the square error before averaging. To avoid obtaining zero attention, which the model tends to do, we add a regularization term, where we create tensors of all" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 335, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 357 + ], + "type": "text", + "content": "ones as the attention heatmap labels and use " + }, + { + "bbox": [ + 305, + 335, + 545, + 357 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 335, + 545, + 357 + ], + "type": "text", + "content": " loss as the attention loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 361, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 553 + ], + "type": "text", + "content": "Finally, specific to our choice of learning-based method, i.e., VoxelPose, which uses a voxel-based 3d root localization model to localize the persons in space using ground-truth 3d root joints (mid-hip joint), we use a simple but effective strategy to localize persons in space. Specifically, we randomly place 3d points in 3d world-space and project them to each view using the given camera parameters, subsequently rendering the projected 2d points as heatmap representations. This generates a synthetic dataset containing 3d points (roots) and their corresponding rendered multiview root-heatmaps. We then use this dataset to train a 3d root localization model that takes multi-view root-heatmaps as input and predicts the 3d roots as output. We further regularize the model by enforcing invariant constraints between pairs of affine augmented root-heatmaps coming from the real multi-view input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 675 + ], + "type": "text", + "content": "Evaluation on three 3d pose benchmarks datasets, Panoptic [32], Shelf [1] and Campus [1], along with extensive ablation studies on the Panoptic [32] dataset, show the effectiveness of our approach. Our approach reaches a performance comparable to learning-based fully-supervised approaches and performs significantly better than optimization-based approaches. Moreover, SMPL body mesh fitting [4, 41] on our estimated 3d poses generates geometrically plausible body shapes (see Figure 1 and Figure 4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "We summarize our contributions as follows: 1) We address the challenging multi-person multi-view 3d person pose estimation problem using a self-supervised approach" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2503" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "without any 2d or 3d ground truth. 2) We propose self-supervised 3d pose estimation by using a new method to recover geometrically constrained 2d joints and heatmap representations from the bottleneck 3d poses. 3) We propose adaptive supervision attention to address the misinformation caused by the inaccurate pseudo labels. 4) We propose self-supervised 3d root localization to estimate the 3d root location utilizing synthetic 3d roots and the corresponding rendered multi-view root heatmaps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 191, + 132, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 132, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 132, + 204 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 211, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 287, + 258 + ], + "type": "text", + "content": "In this section, we briefly review current works related to fully-supervised learning-based methods for 3d pose estimation, optimization-based methods for 3D pose estimation, and self-supervised learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 259, + 289, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 289, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 289, + 474 + ], + "type": "text", + "content": "Fully-supervised methods: Monocular 3d pose estimation [22, 43, 45, 47, 50, 56, 62, 65] is an ill-posed problem due to depth ambiguities as multiple 3d poses can produce same 2d pose projection. Having access to multi-view cameras can alleviate such depth ambiguities achieving the state-of-art results on benchmark datasets [23, 27, 28, 39, 51, 52, 57, 59, 63]. For single-person scenes, these approaches exploit multi-view geometry [24] to either fuse the visual features [27, 51], perform triangulation on heatmaps [28, 53], or use pictorial structural models for 3d reconstruction [48, 51]. The multi-person scene offers extra complexity due to the variability in the number of person in each view and the unknown cross-view correspondence. Existing multi-person multi-view approaches are based on volumetric paradigm [52, 57, 64], or direct regression [63] based on transformers [6, 58, 66]. Despite their good performance, these approaches rely on ground-truth 3d poses, which are generated using dense camera systems [32]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 594 + ], + "type": "text", + "content": "Optimization-based 3d pose estimation: For the multi-person and multi-view scenario, optimization-based approaches use an off-the-shelf person-id detector across all the views to solve the correspondence and triangulation problem and temporal refinement along with training a reinforcement learning agent to find the best camera locations for 3d pose reconstruction [49]. More recent approaches utilize multi-view 3d reconstruction in the optimization loop inferring 3D poses that are geometrically and spatially coherent [10, 18, 33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Self-supervised learning: Self-supervised learning can be broadly classified into self-supervised representation learning and self-supervised task learning. Self-supervised representation learning aims to use large-scale unlabeled data to learn generic feature representations. The recent promising results from these approaches have started to surpass the fully-supervised baselines for various downstream tasks [7, 11, 26, 54]. Self-supervised task learning aims to learn a particular downstream task without using ground truth labels and has been applied to 2d pose estimation [29, 30]," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "single-person 3d pose estimation [9, 19, 34, 36, 38], and surface correspondences estimation [3]. Self-supervised approaches for 3d pose estimation have primarily been developed for single-person scenarios. Given 2d poses, estimated by utilizing advances in the 2d pose estimation methods [5, 13, 14, 21, 35, 42, 44, 46, 55, 60], these approaches use the supervisory signals generated from multi-view geometry [34], video constraints [38], or adversarial learning [9, 19, 36]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 182, + 547, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 242 + ], + "type": "text", + "content": "Our work proposes a learning-based approach to model the 3d poses as bottleneck representations and recover geometrically constrained and spatially accurate 2d joints and heatmap representations in an end-to-end differentiable manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 259, + 389, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 259, + 389, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 259, + 389, + 274 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 281, + 413, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 413, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 413, + 294 + ], + "type": "text", + "content": "3.1. Problem overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": "Given a training dataset of multi-view images " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{x}|\\mathbf{y}^*\\}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{R}^{C \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " is a multi-view image set from " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " cameras with height " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^* \\in \\mathcal{R}^{C \\times P \\times J \\times 2}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " represents the 2d pseudo poses for " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " persons with " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " joints, the goal is to learn a deep learning model that estimates the 3d poses " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathcal{R}^{P \\times J \\times 3}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " of all the " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " persons from the multi-view input image " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": ". It is to be noted that " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": " can vary in each camera view due to occlusion and noisy pseudo 2d pose estimation. For simplicity in the notation, we keep the same variable " + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 304, + 302, + 547, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "text", + "content": "Fully-supervised approaches rely on 3d ground-truth poses, while we only have 2d pseudo poses " + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^*" + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "text", + "content": ". Therefore, after obtaining 3d poses " + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathcal{R}^{P \\times J \\times 3}" + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "text", + "content": " following traditional approach, we propose to project the poses to each view obtaining 2d poses " + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathcal{R}^{C \\times P \\times J \\times 2}" + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "text", + "content": ", and train the model from 2d pseudo poses " + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^* \\in \\mathcal{R}^{C \\times P \\times J \\times 2}" + }, + { + "bbox": [ + 304, + 425, + 547, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 582 + ], + "type": "text", + "content": "In the following, we present our self-supervised approach based on fully-supervised VoxelPose [57]. We first generate pseudo 2d poses, and then propose self-supervised 3d root localization, self-supervised 3d pose estimation, and an adaptive supervision attention to learn 3d poses in a self-supervised manner, without modifying the original Voxel-Pose structure." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 597, + 460, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 460, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 460, + 610 + ], + "type": "text", + "content": "3.2. Generating pseudo 2d poses" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "To circumvent the dependence on the ground-truth 2d poses, we generate the 2d pseudo poses on the training dataset using Mask R-CNN [25] to generate person bounding boxes followed by using HRNet [55] to generate 2d poses of each detected person bounding box. The two-stage approach is chosen based on its state-of-art performance on the COCO dataset [40]. We first pre-train the 2d CNN backbone heatmap_net" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "_{2d}" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " with pseudo 2d poses." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2504" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 80, + 542, + 272 + ], + "blocks": [ + { + "bbox": [ + 256, + 72, + 332, + 79 + ], + "lines": [ + { + "bbox": [ + 256, + 72, + 332, + 79 + ], + "spans": [ + { + "bbox": [ + 256, + 72, + 332, + 79 + ], + "type": "text", + "content": "SelfPose3d Framework" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 80, + 542, + 272 + ], + "lines": [ + { + "bbox": [ + 53, + 80, + 542, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 80, + 542, + 272 + ], + "type": "image", + "image_path": "15683e8d594f83f5225fd1681555b2ce1e29c3713f4977fe874c63317b726872.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "type": "text", + "content": "Figure 2. Illustrating our self-supervised SelfPose3d approaches for multi-view multi-person 3d pose estimation. Instead of using ground-truth 3d poses for learning, we propose self-supervised learning objectives to localize 3d roots (mid-hip location of the person) and estimate their 3d poses. We utilize a synthetic 3d roots dataset, two different affine transformations on the multi-view input images " + }, + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "type": "inline_equation", + "content": "(t_{r,s}^{1}, t_{r,s}^{2})" + }, + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "type": "text", + "content": " parametrized by rotation " + }, + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 290, + 547, + 348 + ], + "type": "text", + "content": ", a differentiable cross-affine-view 2d joints and heatmaps rendering from the bottleneck 3d poses, and an adaptive supervision attention mechanism to automatically learn the 3d poses in world-space." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 366, + 234, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 234, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 234, + 380 + ], + "type": "text", + "content": "3.3. Self-supervised 3d root localization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "text", + "content": "Given 2d multi-view heatmaps from all the views and all the joints " + }, + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\in \\mathcal{R}^{C \\times J \\times \\frac{H}{4} \\times \\frac{W}{4}}" + }, + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "text", + "content": " estimated using a 2d backbone model heatmap。" + }, + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "inline_equation", + "content": "\\mathrm{net}_{2\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "text", + "content": ", we use [28] to construct a discretized 3d feature volume " + }, + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\in \\mathcal{R}^{J \\times X \\times Y \\times Z}" + }, + { + "bbox": [ + 46, + 385, + 288, + 447 + ], + "type": "text", + "content": " for each joint by un-projecting the 2d multi-view heatmaps to 3d space:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 456, + 287, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 456, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 88, + 456, + 287, + 471 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {\\text {u n p r o j}} (\\operatorname {c a m}, \\text {c e n t e r}, t _ {r, s}): \\mathcal {H} \\longrightarrow \\mathcal {F}, \\tag {1}", + "image_path": "47b9623c88e3f21a125c27db1d15b9b3dc07a94a571998715238218d51ba3953.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": "To localize persons' root (mid-hip) joint in 3d space without using 3d ground truth, we hypothesize that 2d multi-view heatmaps of the root location " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{root}} \\in \\mathcal{R}^{C \\times \\frac{H}{4} \\times \\frac{W}{4}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " are sufficient for 3d root localization (see Sec. 7.6 for verification). Then we generate the 3d feature volume for the root location " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{root}} \\in \\mathcal{R}^{X \\times Y \\times Z}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{root}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " using Eq. (1), which has the same dimensions as predicted root-volumes " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": ". This allows us to establish a one-to-one relationship between 3d root-volumes " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " and 2d multi-view root-heatmaps " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{root}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": ". We generate a synthetic root dataset " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{root}} = \\{\\mathcal{G}_i^{\\mathrm{syn*}} | \\mathcal{H}_{\\mathrm{root\\_i}}^{\\mathrm{syn}}\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i^{\\mathrm{syn*}} \\in \\mathcal{R}^{X \\times Y \\times Z}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " contains the root-volumes of randomly placed 3d points, and " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{root\\_i}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " is the corresponding 2d multi-view heatmaps generated by projecting the random 3d points to each view using camera parameters cam. After unprojecting " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{root}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{root}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": ", and passing it through root\\_net obtaining " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{root}}^{\\mathrm{syn}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": ", we compute the " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": " loss error as the synthetic root loss " + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{root\\_syn}}" + }, + { + "bbox": [ + 46, + 481, + 289, + 687 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 109, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "l _ {\\text {r o o t - s y n}} = \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {\\text {s y n}}, \\mathcal {G} ^ {\\text {s y n *}}\\right) \\tag {2}", + "image_path": "6c4bd0a2d7314c4b3c4fb0639ebc302a5eb62c246263a7a45c8832c61034849d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": "To further regularize root_net on the real-world 2d multi-view root-heatmaps, we propose the root consistency loss. Given a multi-view training image set " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": ", we apply two affine transformations " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "(t_{r,s}^{1}, t_{r,s}^{2})" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " with random rotation and scaling " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "(r, s)" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " to generate two affine transformed multiview images " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "(x^1, x^2)" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": ". We pass " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "x^1" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "x^2" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " through heatmap_net2d, construct the root feature volumes using Eq. (1) with corresponding affine transformation parameters, and finally obtain " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^0" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^1" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^2" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " through root_net. Since " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^1" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^2" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " are invariant to the applied affine transformations " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "t_{r,s}^1" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "t_{r,s}^2" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^0" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " as the baseline to compute the " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " loss error between " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^0" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^1" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^2" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": " as the root consistency loss " + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{root\\_C}}" + }, + { + "bbox": [ + 304, + 367, + 545, + 522 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 533, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 533, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 351, + 533, + 545, + 548 + ], + "type": "interline_equation", + "content": "l _ {\\text {r o o t} C} = \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {0}, \\mathcal {G} ^ {1}\\right) + \\mathcal {L} _ {2} \\left(\\mathcal {G} ^ {0}, \\mathcal {G} ^ {2}\\right) \\tag {3}", + "image_path": "f2153d509317b5267180255e81ffc38ef59bfd8c8ceb257f4c5db44f880e0690.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "text", + "content": "We train root_net by minimizing Eq. (2) and Eq. (3). We generate the person proposals " + }, + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{root}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "text", + "content": " by applying nonmaximum suppression (NMS) and thresholding on " + }, + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^2" + }, + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^1" + }, + { + "bbox": [ + 304, + 550, + 545, + 598 + ], + "type": "text", + "content": " would also work)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 604, + 489, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 489, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 489, + 617 + ], + "type": "text", + "content": "3.4. Self-supervised 3d pose estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "text", + "content": "Given pseudo 2d poses " + }, + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{2\\mathrm{d}}^{*}" + }, + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "text", + "content": ", person proposals " + }, + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{root}_i\\}_{i = 1}^P" + }, + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "text", + "content": " and 2d multi-view heatmaps " + }, + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^1,\\mathcal{H}^2" + }, + { + "bbox": [ + 304, + 621, + 545, + 670 + ], + "type": "text", + "content": " predicted using heatmap_net2d, we describe our self-supervised 3d pose estimation approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "type": "text", + "content": "The person proposals " + }, + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{root}_i\\}_{i = 1}^P" + }, + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "type": "text", + "content": " are used to generate the 3d feature volumes i.e " + }, + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{F}_i^1\\right\\}_{i = 1}^P = \\left\\{\\mathcal{P}_{\\mathrm{unproj}}(\\mathrm{cam},r_i,t_{r,s}^1)\\right\\}_{i = 1}^P" + }, + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 670, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{F}_i^2\\right\\}_{i = 1}^P =" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2505" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{P}_{\\mathrm{unproj}}(\\mathrm{cam},r_i,t_{r,s}^2)\\right\\}_{i = 1}^P" + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "text", + "content": " corresponding to the person feature volumes for each affine augmented multi-view input image " + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "inline_equation", + "content": "x^{1}" + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "inline_equation", + "content": "x^{2}" + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{F}_i^1\\right\\}_{i = 1}^P" + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{F}_i^2\\right\\}_{i = 1}^P" + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "text", + "content": " are passed through pose_net3d and soft-argmax [8] to estimate the 3d poses " + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^1,\\mathcal{V}^2\\in \\mathcal{R}^{P\\times J\\times 3}" + }, + { + "bbox": [ + 47, + 70, + 286, + 152 + ], + "type": "text", + "content": ". These 3d poses serve as a bottleneck representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": "Given the camera parameters cam and the affine transformation parameters " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "t_{r,s}^{1}, t_{r,s}^{2}" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": ", we project the bottleneck 3d poses in cross-affine-view i.e. " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^1" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": " are projected to " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "x^2" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": " image space using " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "t_{r,s}^{2}" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": " to generate multi-view 2d poses " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^2 \\in \\mathcal{R}^{C \\times P \\times J \\times 2}" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^2" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": " are projected to the " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "x^1" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": " image space using " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "t_{r,s}^{1}" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": " to generate multi-view 2d poses " + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^1 \\in \\mathcal{R}^{C \\times P \\times J \\times 2}" + }, + { + "bbox": [ + 47, + 152, + 286, + 225 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 225, + 286, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 286, + 392 + ], + "type": "text", + "content": "We propose to render " + }, + { + "bbox": [ + 46, + 225, + 286, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^1" + }, + { + "bbox": [ + 46, + 225, + 286, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 225, + 286, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^2" + }, + { + "bbox": [ + 46, + 225, + 286, + 392 + ], + "type": "text", + "content": " 2d poses in the 2d heatmap representation. The heatmap representation encodes the per-pixel likelihood of a body joint and has been a vital component to enable the state-of-the-art 2d pose estimation approaches [55]. Generating heatmap representation has essentially been a pre-processing step where state-of-the-art approaches quantize the 2d joints before generating the heatmap [55]. However, this quantization step is non-differentiable and could cut the backward gradient flow. Zhang et al. [61] show that encoding floating point 2d joints into heatmap representation in their pre-processing step leads to improved performance. We propose to use the same differentiable approach in an online way to render the projected 2d joints into the heatmap representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": "We render " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^1" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^2" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": " into heatmap representation to generate 2d multi-view heatmaps " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^1" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^2" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": ", respectively. We apply the affine transformations " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "t_{r,s}^1, t_{r,s}^2" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": " on the pseudo 2d poses " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{2\\mathrm{d}}^{*}" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": " to generate pseudo 2d multi-view joints " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{2\\mathrm{d}}^{1*}, \\mathbf{y}_{2\\mathrm{d}}^{2*}" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": " and heatmaps " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{1*}, \\mathcal{H}^{2*}" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": ". Then, we compute the " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": " loss between heatmaps as the pose heatmap loss " + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_H}}" + }, + { + "bbox": [ + 46, + 392, + 286, + 465 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 475, + 286, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 475, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 85, + 475, + 286, + 491 + ], + "type": "interline_equation", + "content": "l _ {\\text {p o s e} \\mathrm {H}} = \\mathcal {L} _ {2} \\left(\\mathcal {H} ^ {1}, \\mathcal {H} ^ {1 *}\\right) + \\mathcal {L} _ {2} \\left(\\mathcal {H} ^ {2}, \\mathcal {H} ^ {2 *}\\right) \\tag {4}", + "image_path": "324cfd05169967dfc33e21cc77b8e3e54a5c7374304a50dd8c15d345fd9f808a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": "After training with " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_H}}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": " preliminarily, we propose to add the " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": " loss between 2d joints to further fine-tune the model. For each view, we employ the Hungarian algorithm [37] to obtain the optimal assignment between " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{2\\mathrm{d}}^{*}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": ", where the matching cost is the mean absolute error. Based on the optimal assignment, we obtain the " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": " loss as " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_J}}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": ". Then we use " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_H}}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_J}}" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": " together to train the whole network, where " + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 496, + 286, + 591 + ], + "type": "text", + "content": " is a manually defined weight:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 602, + 286, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 602, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 90, + 602, + 286, + 617 + ], + "type": "interline_equation", + "content": "l _ {\\text {p o s e - J}} = \\mathcal {L} _ {1} \\left(\\mathbf {y} ^ {1}, \\mathbf {y} _ {2 \\mathrm {d}} ^ {1 *}\\right) + \\mathcal {L} _ {1} \\left(\\mathbf {y} ^ {2}, \\mathbf {y} _ {2 \\mathrm {d}} ^ {2 *}\\right) \\tag {5}", + "image_path": "1355cc47286a740b5cd9032ee92cc55be999cc4064b32fc7d232c86533977dcd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 635, + 286, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 635, + 286, + 648 + ], + "spans": [ + { + "bbox": [ + 108, + 635, + 286, + 648 + ], + "type": "interline_equation", + "content": "l _ {\\text {p o s e - 3 d}} = l _ {\\text {p o s e - H}} + \\lambda l _ {\\text {p o s e - J}} \\tag {6}", + "image_path": "ae01e2f26a9336d8d8237a3a106cbd06c32a12e7b8c3a30bd35a3a574b86da59.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "As the network needs to reason about the 2d joint locations in spatial dimension, it implicitly solves the person correspondence problem. Training pose_net3d with 3d pose loss " + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_3d}}" + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": " performs decently, but to achieve even better results, we introduce the adaptive supervision attention." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 425, + 201 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 425, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 425, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 425, + 201 + ], + "type": "image", + "image_path": "27a3d8b7034de1a41fcf05a0ea8360107c945bc08060162bc16a013846740d43.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 201, + 400, + 207 + ], + "lines": [ + { + "bbox": [ + 333, + 201, + 400, + 207 + ], + "spans": [ + { + "bbox": [ + 333, + 201, + 400, + 207 + ], + "type": "text", + "content": "ground truth 2d poses" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 216, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 292 + ], + "type": "text", + "content": "Figure 3. Comparing ground-truth 2d poses generated by projecting the ground-truth 3d poses to each multi-view image and our pseudo 2d poses generated by running HRNet human pose estimation model [55] on the training dataset. Pseudo 2d poses contain localization errors due to occlusion (see the red arrows), and ground-truth 2d poses exist for partially or even entirely occluded persons (see the blue dotted arrows)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 71, + 543, + 201 + ], + "blocks": [ + { + "bbox": [ + 425, + 71, + 543, + 201 + ], + "lines": [ + { + "bbox": [ + 425, + 71, + 543, + 201 + ], + "spans": [ + { + "bbox": [ + 425, + 71, + 543, + 201 + ], + "type": "image", + "image_path": "ec0de6ec05ddac0d1a71ef0f200a5a00a5df47cf062f8753ec123a2a6907381b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 201, + 511, + 206 + ], + "lines": [ + { + "bbox": [ + 462, + 201, + 511, + 206 + ], + "spans": [ + { + "bbox": [ + 462, + 201, + 511, + 206 + ], + "type": "text", + "content": "pseudo 2d poses" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 312, + 471, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 471, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 471, + 323 + ], + "type": "text", + "content": "3.5. Adaptive supervision attention" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 329, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 426 + ], + "type": "text", + "content": "Traditional " + }, + { + "bbox": [ + 304, + 329, + 545, + 426 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 329, + 545, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 329, + 545, + 426 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 329, + 545, + 426 + ], + "type": "text", + "content": " losses treat each label equally, which is sub-optimal in two aspects: (1) the 2d human pose detector may generate inaccurate labels due to occlusions (see the red arrows in Figure 3); (2) the 3d-2d projection will output 2d joints in certain views even when the person is entirely occluded (see the blue dotted arrows in Figure 3). Therefore, we propose to employ attentions to adaptively guide the supervision process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "text", + "content": " loss supervision, we use the soft attention. Specifically, we use ResNet-18 to extract the visual features of the views, followed by deconvolutional layers to obtain the attention heatmaps " + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "text", + "content": " (see attn\\_net" + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "inline_equation", + "content": "_{2d}" + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "text", + "content": " in Figure 2). Then we compute the element-wise product of " + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "text", + "content": " and the square error before averaging, as the new loss " + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_H}}^{\\mathrm{attn}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 500 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 352, + 506, + 545, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 506, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 352, + 506, + 545, + 538 + ], + "type": "interline_equation", + "content": "l _ {\\text {p o s e - H}} ^ {\\mathrm {a t t n}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {A} _ {i} \\otimes \\left(\\mathcal {H} _ {i} - \\mathcal {H} _ {i} ^ {*}\\right) ^ {2} \\tag {7}", + "image_path": "475626fb85ef01232616a3a7048ccb98e4a151ac335658dba4125295456327d4.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "content": "To avoid that " + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "content": " becomes zero, we add a regularization term. We create tensors of all ones " + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "content": " as the attention heatmap labels, and compute " + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "content": " error as the attention loss " + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{attn}}" + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{attn}}" + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "content": " becomes zero, " + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_H}}^{\\mathrm{attn}}" + }, + { + "bbox": [ + 304, + 541, + 545, + 601 + ], + "type": "text", + "content": " degrades to non-attentive version." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 389, + 613, + 545, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 613, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 389, + 613, + 545, + 625 + ], + "type": "interline_equation", + "content": "l _ {\\mathrm {a t t n}} = \\mathcal {L} _ {2} (\\mathcal {A}, \\mathbb {1}) \\tag {8}", + "image_path": "b2390fd5effcf437cb44a92160bae71d1ee87a7050798aec6962f0eb795e4906.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "text", + "content": " loss supervision, we use the hard attention. For each input with " + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "text", + "content": " views, we compute the " + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "text", + "content": " loss of each view, find the view with the largest loss, and ignore it when averaging the final loss " + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_J}}^{\\mathrm{att}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 679 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 332, + 696, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 696, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 332, + 696, + 545, + 716 + ], + "type": "interline_equation", + "content": "j = \\underset {i} {\\arg \\max } \\mathcal {L} _ {1} \\left(\\mathbf {y} _ {i}, \\mathbf {y} _ {i, 2 \\mathrm {d}} ^ {*}\\right) \\quad (i = 1, 2, \\dots , K) \\tag {9}", + "image_path": "c821dc6b84ad0879a197980a65d78b56da055c19e3c5513196155c4e656773a5.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2506" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 89, + 81, + 287, + 115 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 81, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 89, + 81, + 287, + 115 + ], + "type": "interline_equation", + "content": "l _ {\\text {p o s e - J}} ^ {\\mathrm {a t t n}} = \\frac {1}{K - 1} \\sum_ {i = 1, i \\neq j} ^ {K} \\mathcal {L} _ {1} \\left(\\mathbf {y} _ {i}, \\mathbf {y} _ {i, 2 \\mathrm {d}} ^ {*}\\right) \\tag {10}", + "image_path": "7eb4af776bd2f070a3b0e252b2c1276b294e20e5aceec72b25c46efd84b44fda.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "text", + "content": "In general, the final 3d pose loss " + }, + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_3d}}^{\\mathrm{attn}}" + }, + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "text", + "content": " is as follows, where " + }, + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 122, + 287, + 147 + ], + "type": "text", + "content": " are manually defined weights:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 90, + 158, + 287, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 158, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 90, + 158, + 287, + 174 + ], + "type": "interline_equation", + "content": "l _ {\\text {p o s e - 3 d}} ^ {\\text {a t t n}} = l _ {\\text {p o s e - H}} ^ {\\text {a t t n}} + \\lambda l _ {\\text {p o s e - J}} ^ {\\text {a t t n}} + \\sigma l _ {\\text {a t t n}} \\tag {11}", + "image_path": "86f254d0130792df07839a199cc7cf4f0fd1cc401d245fa43120803be98bc301.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 178, + 287, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 202 + ], + "type": "text", + "content": "We train pose_net3d by minimizing " + }, + { + "bbox": [ + 47, + 178, + 287, + 202 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pose\\_3d}}^{\\mathrm{attn}}" + }, + { + "bbox": [ + 47, + 178, + 287, + 202 + ], + "type": "text", + "content": ". Our self-supervised approach is visually described in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 210, + 178, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 178, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 178, + 223 + ], + "type": "text", + "content": "3.6. Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "text", + "content": "Training strategies For the Panoptic dataset, similar to VoxelPose [57], we first train heatmap_net2d for 20 epochs with pseudo 2d poses. We use the Adam optimizer with an initial learning rate of 1e-4, which decreases to 1e-5 and 1e-6 at the 10th and 15th epochs, respectively. Then, we train the root_net for 1 epoch, followed by end-to-end joint training of the whole network for 5 epochs using only the " + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "text", + "content": " loss, with a learning rate of 1e-4. Afterwards, we add " + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "text", + "content": " loss to train the whole network for another 5 epochs with a learning rate of 5e-5. " + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 228, + 287, + 361 + ], + "type": "text", + "content": " in Eq. (11) are set to 0.01 and 0.1 respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": "We use the random rotation between " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "-45^{\\circ}" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": " and random scale between " + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "inline_equation", + "content": "-0.35" + }, + { + "bbox": [ + 46, + 361, + 288, + 482 + ], + "type": "text", + "content": " to 0.35. We also apply spatial augmentations using rand-augment [16] and rand-cutout [17] using python image library1. The rand-augment consists of \"contrast-jittering\", \"auto-contrast\", \"equalize\", \"color-jittering\", \"sharpness-jittering\", and \"brightness-jittering\", and the rand-cutout places random square boxes of sizes between 20 to 40 pixels at random locations in the image. We use the SMPL model and optimization-based body fitting approach2[4, 41] to estimate body mesh parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": "Inference pipeline During inference, we input the multiview RGB images, and obtain the estimated 3d poses in an end-to-end pipeline. For each view, the backbone generates corresponding 2d heatmaps for cuboid construction. Then, given constructed cuboid of the whole space, the root_net predicts root joint locations of all persons. Finally, the pose_net outputs the regressed 3d locations of each joint for every cuboid proposal of the root joints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 218, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 218, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 218, + 635 + ], + "type": "text", + "content": "4.1. Datasets and evaluation metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 642, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 666 + ], + "type": "text", + "content": "We conduct experiments on three benchmark datasets: Panoptic [32], Campus [1], and Shelf [1]." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 542, + 151 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 151 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 151 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 151 + ], + "type": "table", + "html": "
MethodsAP25AP50AP100AP150Recall@500MPJPE[mm]
FSVoxelPose [57]83.698.399.899.998.817.7
Lin et al. [39]92.199.099.899.8-16.8
MvP [63]92.396.697.597.798.215.8
Wu et al. [59]----98.715.8
TEMPO [15]89.099.199.899.9-14.7
OBACTOR [49]-----168.4
MvPose [18]0.02.9759.9381.5398.2384.2
SSSelfPose3d (ours)55.196.498.599.099.624.5
", + "image_path": "01fa1677e36f8bbff62ce3dc00577a0bdcf2b01ed13fc81549f144a2b08d0953.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 194, + 544, + 271 + ], + "blocks": [ + { + "bbox": [ + 306, + 159, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 159, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 545, + 182 + ], + "type": "text", + "content": "Table 1. Result on the Panoptic dataset (FS = fully-supervised, OB = optimization-based, SS = self-supervised)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 194, + 544, + 271 + ], + "lines": [ + { + "bbox": [ + 309, + 194, + 544, + 271 + ], + "spans": [ + { + "bbox": [ + 309, + 194, + 544, + 271 + ], + "type": "table", + "html": "
MethodsShelfCampus
Actor 1Actor 2Actor 3AverageActor 1Actor 2Actor 3Average
FSErshadi et al. [20]93.375.994.888.094.292.984.690.6
Wu et al. [59]99.396.597.397.7----
MvP [63]99.395.197.897.498.294.197.496.6
VoxelPose [57]99.394.197.697.097.693.898.896.7
VoxelPose* [57]99.593.597.896.993.186.593.290.9
OB3DPS [2]75.369.787.677.593.575.784.484.5
MvPose [18]98.894.197.896.997.693.398.096.3
SSSelfPose3d97.290.397.995.192.582.289.287.9
", + "image_path": "1d675c3ab67e79a6b3e785ef9e07b79cb86e34ac279e618b15a4decb07f7e479.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 279, + 546, + 324 + ], + "lines": [ + { + "bbox": [ + 305, + 279, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 546, + 324 + ], + "type": "text", + "content": "Table 2. Results (in PCP) on Shelf and Campus datasets (FS = fully-supervised, OB = optimization-based, SS = self-supervised, * = reproduced results). SelfPose3d is trained from the pseudo 3d poses from the Panoptic training set." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 346, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 526 + ], + "type": "text", + "content": "The Panoptic dataset is a large-scale dataset captured inside a dome environment containing multiple persons performing daily social activities. We conduct extensive experiments on this dataset to evaluate and assess various components of our approach. We use the same data sequences for training and testing as VoxelPose [57] except that our training set doesn't include '160906_band3'. In other words, we are only using 9 multi-view videos for training (the '160906_band3' video is not available due to the broken images on the source website). We use the five HD camera images (3, 6, 12, 13, 23) to train and report the performance in our experiments. We use Average Precision (AP), Recall, and Mean Per Joint Position Error (MPJPE) in millimeters (mm) as evaluation metrics (higher AP and lower MPJPE are better) [57]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 527, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 545, + 586 + ], + "type": "text", + "content": "The Shelf and Campus are two multi-person datasets capturing activities in the indoor and outdoor environments, respectively [1]. We use the same training and test split as [1, 57]. As used in [1, 57], we use the Percentage of Correct Parts (PCP) as evaluation metrics for these two datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": "Panoptic Table 1 shows the results on the challenging Panoptic dataset. All the fully-supervised approaches utilizing 2d and 3d ground-truth 3d poses reach nearly the same performance. SelfPose3d, without using any 3d or 2d ground-truth poses, achieves comparable results to fully-supervised approaches. Nevertheless, there still exists a gap compared to the fully-supervised VoxelPose model (96.4 v.s 98.3 AP50 and 24.5 v.s 17.7 MPJPE). However, unlike VoxelPose, which relies on heat maps from all the joints to es" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 673, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 286, + 713 + ], + "type": "text", + "content": "1https : / / github . com / jizongFox / pytorch - randaugment 2https : // github . com / JiangWenPL / multiperson/ tree/master/misc/smplify-x" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2507" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 521, + 330 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 521, + 330 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 521, + 330 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 521, + 330 + ], + "type": "image", + "image_path": "284dc7761ab5c730b5f21a0f5c57b543155f46c8bf22803f2fd71a6d2ac22fa4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 337, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 547, + 361 + ], + "type": "text", + "content": "Figure 4. Qualitative results for the 3d pose estimations, 2d projections on the multi-view images, and estimated SMPL body shapes on some example images from the Panoptic dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 389, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 437 + ], + "type": "text", + "content": "timate 3d roots, we only use root-heatmaps to do the same. This results in the reduction of the input channel from 15 (number of keypoints) to 1 for the root_net, making our approach computationally faster." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 441, + 288, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 548 + ], + "type": "text", + "content": "We also compare our approach with optimization-based baselines from Pirinen et al. [49] and Dong et al. [18]. These non-learning-based approaches fail to capture the multi-person interaction in a complex scene from a few sparse multi-view cameras. Our learning-based self-supervised approach achieves much better performance. It is to be noted that Pirinen et al. evaluate their approach on two multi-person sequences, whereas we evaluate on four multi-person sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Shelf and Campus We compare our approach with the state-of-the-art methods on the Shelf and Campus dataset. VoxelPose uses the 3d ground-truth from the Panoptic dataset to train their approach to these datasets due to noisy and incomplete 3d ground-truth poses. For a fair comparison with VoxelPose, we use the pseudo 3d poses (by running SelfPose3d on the Panoptic training set) and train on these two datasets in a fully supervised manner. As shown in Table 2, our approach using pseudo 3d poses from the Panoptic dataset also reaches the same performance as the fully-supervised approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 389, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 547, + 521 + ], + "type": "text", + "content": "Qualitative visualizations Figure 4 shows 3d pose estimation results from our SelfPose3d approach on the challenging Panoptic dataset. Without using any 3d ground-truth, we can see that SelfPose3d is robust to occlusions and multiple persons while correctly identifying the person identities across all the views (see the corresponding 2d projections in Figure 4). We also show the qualitative results for the SMPL body mesh fitting [4, 41] on the estimated 3d poses. All these results demonstrate both the effectiveness and extendability of SelfPose3d. Please see the supplementary for more results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 528, + 405, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 528, + 405, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 405, + 540 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Ground-truth 2d poses v.s pseudo 2d poses As shown in Table 3, when we use the ground-truth 2d poses in our self-supervised framework, 3d reconstruction error significantly reduces. To inspect the better performance when using the ground-truth 2d poses, we qualitatively compare the ground-truth 2d poses with the pseudo 2d poses on some example training images. Pseudo 2d poses contain localization errors due to occlusion, whereas ground-truth 2d poses exist for partially or even entirely occluded persons as shown in Figure 3. As the ground-truth 2d poses are generated by projecting the ground-truth 3d poses to each multi-view image, they serve as a suitable proxy for the 3d poses, thereby reaching a performance close to the fully-supervised approaches. However, obtaining the ground 2d" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2508" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 70, + 230, + 105 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 230, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 230, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 230, + 105 + ], + "type": "table", + "html": "
2d posesAP50AP100MPJPE
ground-truth98.899.619.9
pseudo96.498.524.5
", + "image_path": "fa796cd60a174944608efff4f7e1758eaf5d265b6f63f162c69261dfaeba2d91.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 158, + 272, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "type": "text", + "content": "Table 3. The ground-truth 2d poses in our self-supervised framework decrease the 3d reconstruction error and reach the performance close to the fully-supervised approaches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 158, + 272, + 201 + ], + "lines": [ + { + "bbox": [ + 62, + 158, + 272, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 272, + 201 + ], + "type": "table", + "html": "
cross-affine-view consistencyaffine augsAP50AP100MPJPE
86.096.234.7
✓83.397.533.3
✓✓93.898.129.3
", + "image_path": "6bb6a7b22feb9503eaecc463e56f2304baa7a0560d17cd018a4c9e6db0da7725.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 86, + 266, + 249, + 309 + ], + "blocks": [ + { + "bbox": [ + 47, + 210, + 286, + 253 + ], + "lines": [ + { + "bbox": [ + 47, + 210, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 286, + 253 + ], + "type": "text", + "content": "Table 4. Affine augmentations and cross-affine-view consistency significantly improves the 3d pose reconstruction accuracy. All three models are trained for two epochs with frozen backbone and frozen root_net and no attention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 266, + 249, + 309 + ], + "lines": [ + { + "bbox": [ + 86, + 266, + 249, + 309 + ], + "spans": [ + { + "bbox": [ + 86, + 266, + 249, + 309 + ], + "type": "table", + "html": "
L1 lossL2 lossAP25AP50AP100MPJPE
✓----
✓43.895.898.225.7
✓✓55.196.498.524.5
", + "image_path": "6fba851443d10ec557b624aef444c76522e244a68dc5d74336d4ca920555ee7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 361, + 286, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 286, + 385 + ], + "type": "text", + "content": "poses in this way would be as challenging as acquiring the ground-truth 3d poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 402, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 286, + 498 + ], + "type": "text", + "content": "Importance of cross-affine-view consistency and affine augmentations We also examine the effect of affine augmentations on the multi-view images and cross-affine-view consistency when generating differentiable 2d representations from the bottleneck 3d poses. As shown in Table 4, the affine augmentations and cross-affine-view consistency significantly improve the 3d pose reconstruction as they provide necessary geometric constraints during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": "Analysis of " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " pose losses We conduct experiments to analyze the use of " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " pose losses. As shown in Table 5, using " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " losses together can obtain better results than using " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " loss solely. Also, using " + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 514, + 286, + 573 + ], + "type": "text", + "content": " loss solely doesn't converge due to the label noises." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 590, + 286, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 286, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 286, + 637 + ], + "type": "text", + "content": "Importance of adaptive supervision attention We also examine the necessity of adaptive supervision attention. Table 6 shows that supervision attention for both " + }, + { + "bbox": [ + 47, + 590, + 286, + 637 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 590, + 286, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 590, + 286, + 637 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 590, + 286, + 637 + ], + "type": "text", + "content": " losses are necessary for training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "type": "text", + "content": "Influence of different 2d human pose estimation models Finally, we show how pseudo 2d poses generated from different 2d human pose estimation models affect the performance. As shown in Table 7, models that perform well on the COCO dataset [40] also generate better pseudo 2d poses" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 312, + 71, + 541, + 125 + ], + "blocks": [ + { + "bbox": [ + 47, + 318, + 286, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 286, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 286, + 340 + ], + "type": "text", + "content": "Table 5. Training using " + }, + { + "bbox": [ + 47, + 318, + 286, + 340 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 318, + 286, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 318, + 286, + 340 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 318, + 286, + 340 + ], + "type": "text", + "content": " pose losses together achieves the best performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 71, + 541, + 125 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 541, + 125 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 541, + 125 + ], + "type": "table", + "html": "
L1 loss attentionL2 loss attentionAP25AP50AP100MPJPE
32.594.197.828.5
✓37.995.898.026.3
✓47.496.698.225.0
✓✓55.196.498.524.5
", + "image_path": "231b8c119ee1243b5556c3da3363028f9930fd62b854130db2b59d8957441b7e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 319, + 167, + 533, + 208 + ], + "blocks": [ + { + "bbox": [ + 306, + 134, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 306, + 134, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 134, + 545, + 156 + ], + "type": "text", + "content": "Table 6. Training using " + }, + { + "bbox": [ + 306, + 134, + 545, + 156 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 306, + 134, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 134, + 545, + 156 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 306, + 134, + 545, + 156 + ], + "type": "text", + "content": " loss supervisions together achieves the best performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 167, + 533, + 208 + ], + "lines": [ + { + "bbox": [ + 319, + 167, + 533, + 208 + ], + "spans": [ + { + "bbox": [ + 319, + 167, + 533, + 208 + ], + "type": "table", + "html": "
Method for 2d pseudo pose generationAP50AP100MPJPEKeypoint AP on COCO-val[40]
Keypoint R-CNN (R-101) [25]89.297.631.966.1
HRNet (w48 384x288) [55]93.898.129.376.3
", + "image_path": "ac36886afdb7d5f0aa02745cb0acdec838b7cf834f55f02c6970c1a0490658c5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 216, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 260 + ], + "type": "text", + "content": "Table 7. Comparing different models for generating pseudo 2d poses. Models that perform well on the COCO dataset [40] also generate better pseudo 2d poses for the Panoptic dataset, helping SelfPose3d to achieve better performance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 281, + 544, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 544, + 305 + ], + "type": "text", + "content": "for the Panoptic dataset, helping SelfPose3d to achieve better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 316, + 378, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 378, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 378, + 328 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 335, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 587 + ], + "type": "text", + "content": "We present a self-supervised approach, called SelfPose3d, to address the challenging problem of multi-view multiperson 3d human pose estimation. Unlike current state-of-the-art methods that use difficult-to-acquire 3d ground-truth poses to train the model, SelfPose3d requires only multiview input images and an off-the-shelf 2d human pose detector. We propose a novel self-supervised learning objective that aims to recover 2d joints and heatmaps under different affine transformations from the bottleneck 3d poses. We further improve the performance of our approach by integrating adaptive supervision attention to address the misinformation caused by the inaccurate 2d pseudo labels from the off-the-shelf 2d human pose detector. We conduct extensive experiments on large-scale benchmark datasets, assess various components of our approach, and show that SelfPose3d reaches a performance on par with the well-established fully-supervised baselines. We visualize the 3d pose reconstruction in the complex multiple-person scenes and show that body shape meshes fitted on the estimated 3d poses look geometrically plausible under different viewpoints." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 597, + 420, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 420, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 420, + 611 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "This work was partially supported by French state funds managed by the ANR under references ANR-20-CHIA0029-01 (National AI Chair AI4ORSafety), ANR-10IAHU-02 (IHU Strasbourg), ANR-18-CE45-0011-03 (OptimiX), and by BPI France (project 5G-OR). This work was also granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3]." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2509" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures for multiple human pose estimation. In CVPR, pages 1669-1676, 2014. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "text", + "content": "[2] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. TPAMI, 38(10):1929-1942, 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "text", + "content": "[3] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Loopreg: Self-supervised learning of implicit surface correspondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems, 33:12909-12922, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578. Springer, 2016. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "type": "text", + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, pages 7291-7299, 2017. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 424 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 457 + ], + "type": "text", + "content": "[8] Olivier Chapelle and Mingrui Wu. Gradient descent optimization of smoothed information retrieval metrics. Information retrieval, 13(3):216-235, 2010. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 459, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 502 + ], + "type": "text", + "content": "[9] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, M. V. Rohith, Stefan Stojanov, and James M. Rehg. Unsupervised 3d pose estimation with geometric self-supervision. CoRR, abs/1904.04812, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. In European Conference on Computer Vision, pages 541-557. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "text", + "content": "[12] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[13] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[15] Rohan Choudhury, Kris M Kitani, and László A Jeni. Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14750-14760, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "text", + "content": "[16] Ekin D Cubuk, Barret Zoph, Jonathon Shiens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": "[17] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In CVPR, pages 7792-7801, 2019. 1, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 350 + ], + "type": "text", + "content": "[19] Dylan Drover, Ching-Hang Chen, Amit Agrawal, Ambrish Tyagi, and Cong Phuoc Huynh. Can 3d pose be learned from 2d projections alone? In Proceedings of the European Conference on Computer Vision (ECCV), pages 0–0, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[20] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77(12): 15573-15601, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "text", + "content": "[21] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2334-2343, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "text", + "content": "[22] Kehong Gong, Jianfeng Zhang, and Jiashi Feng. Poseaug: A differentiable pose augmentation framework for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8575-8584, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 492, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 536 + ], + "type": "text", + "content": "[23] Brian Gordon, Sigal Raab, Guy Azov, Raja Giryes, and Daniel Cohen-Or. Flex: Parameter-free multiview 3d human motion reconstruction. arXiv preprint arXiv:2105.01937, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 568 + ], + "type": "text", + "content": "[24] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "text", + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 614, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 668 + ], + "type": "text", + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[27] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7779-7788, 2020. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2510" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[28] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. arXiv preprint arXiv:1905.05754, 2019. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[29] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[30] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabelled videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8787-8797, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[31] Hanbyul Joo, Hyun Soo Park, and Yaser Sheikh. Map visibility estimation for large-scale dynamic 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1122-1129, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "type": "text", + "content": "[32] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In Proceedings of the IEEE International Conference on Computer Vision, pages 3334-3342, 2015. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "text", + "content": "[33] Abdelrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):1-14, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 419 + ], + "type": "text", + "content": "[34] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1077-1086, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "text", + "content": "[35] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR, pages 11977-11986, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "text", + "content": "[36] Yasunori Kudo, Keisuke Ogaki, Yusuke Matsui, and Yuri Odagiri. Unsupervised adversarial learning of 3d human pose from 2d joint locations. arXiv preprint arXiv:1803.08244, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[37] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 600 + ], + "type": "text", + "content": "[38] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6152-6162, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[39] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11886-11895, 2021. 1, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "person linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "text", + "content": "[42] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[43] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 2640-2649, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[44] William McNally, Kanav Vats, Alexander Wong, and John McPhee. Evopose2d: Pushing the boundaries of 2d human pose estimation using neuroevolution. arXiv preprint arXiv:2011.08446, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 307 + ], + "type": "text", + "content": "[45] Dushyant Mehta, Srinath Sridhar, Oleksandr Sotnychenko, Helge Rhodin, Mohammad Shafiei, Hans-Peter Seidel, Weipeng Xu, Dan Casas, and Christian Theobalt. Vnect: Real-time 3d human pose estimation with a single rgb camera. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "text", + "content": "[46] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. In NIPS, pages 2277-2287, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "type": "text", + "content": "[47] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6951-6960, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[48] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In CVPR, pages 1253-1262, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[49] Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Domes to drones: Self-supervised active triangulation for 3d human pose reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "text", + "content": "[50] Alin-Ionut Popa, Mihai Zanfir, and Cristian Sminchisescu. Deep multitask architecture for integrated 2d and 3d human sensing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6289-6298, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "text", + "content": "[51] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In ICCV, pages 4342-4351, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 632 + ], + "type": "text", + "content": "[52] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G Narasimhan. Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15190-15200, 2021. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[53] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In CVPR, pages 6040-6049, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "type": "text", + "content": "[54] Pierre H Richemond, Jean-Bastien Grill, Florent Altché, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, et al." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2511" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 611 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Byol works even without batch statistics. arXiv preprint arXiv:2010.10241, 2020.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[55] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, pages 5693-5703, 2019. 1, 3, 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 162 + ], + "type": "text", + "content": "[56] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV, pages 529-545, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "text", + "content": "[57] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 1, 3, 6, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[59] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11148-11157, 2021. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 342 + ], + "type": "text", + "content": "[60] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, pages 466-481, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "type": "text", + "content": "[61] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7093-7102, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "text", + "content": "[62] Jianfeng Zhang, Xuecheng Nie, and Jiashi Feng. Inference stage optimization for cross-scenario 3d human pose estimation. Advances in Neural Information Processing Systems, 33:2408-2419, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[63] Jianfeng Zhang, Yujun Cai, Shuicheng Yan, Jiashi Feng, et al. Direct multi-view multi-person 3d pose estimation. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "type": "text", + "content": "[64] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. arXiv preprint arXiv:2108.02452, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "type": "text", + "content": "[65] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In ICCV, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[66] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2512" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_content_list.json b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2b575b170c9f5c581e0e8929957a1eb96b6d9d87 --- /dev/null +++ b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_content_list.json @@ -0,0 +1,1469 @@ +[ + { + "type": "text", + "text": "SemCity: Semantic Scene Generation with Triplane Diffusion", + "text_level": 1, + "bbox": [ + 173, + 130, + 795, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jumin Lee $^{1*}$ Sebin Lee $^{1*}$ Changho Jo $^{2}$ Woobin Im $^{1}$ Juhyeong Seon $^{1}$ Sung-Eui Yoon $^{1}$ $^{1}$ KAIST $^{2}$ Neosapience, Inc.", + "bbox": [ + 114, + 179, + 851, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41e3c143eef09caf41f959ed3acc2728a1a2eda481960466c73f6825302a925c.jpg", + "image_caption": [ + "(a) Scene generation" + ], + "image_footnote": [], + "bbox": [ + 81, + 250, + 282, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3564e47203d7a727abfbd1630678f15064257f8ea24a305c1b3853f8db2f7e88.jpg", + "image_caption": [ + "(c) Scene outpainting" + ], + "image_footnote": [], + "bbox": [ + 287, + 252, + 720, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db600fd2c966d0be4c3d13c0eb9c5a682c3eebb88015381d387babaae39f3f73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 256, + 880, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c599459d462a207f589abd6de1133221a987f1391f5b62a3e924909b8cd5add5.jpg", + "image_caption": [ + "(b) Semantic scene completion refinement" + ], + "image_footnote": [], + "bbox": [ + 287, + 386, + 679, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5adfb0a1a22d8c604a474870f7dfa47b5b861f77d72c48444a7cf65bc9f34b81.jpg", + "image_caption": [ + "(d) Scene inpainting", + "Figure 1. We introduce a diffusion framework, SemCity, designed for generating semantic scenes in real-world outdoor environments as shown in (a). We extend our diffusion model to various practical tasks: semantic scene completion refinement, scene outpainting, and scene inpainting. For instance, the comprehensive scenario is displayed in $(\\mathrm{b})\\rightarrow (\\mathrm{c})\\rightarrow (\\mathrm{d})$ : the refined scene (SSC refinement) (b) is outpainted to a broader scene (c); then, an object (in this case, a car) is seamlessly integrated into the scene via our inpainting process (d)." + ], + "image_footnote": [], + "bbox": [ + 692, + 400, + 880, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 570, + 313, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present \"SemCity,\" a 3D diffusion model for semantic scene generation in real-world outdoor environments. Most 3D diffusion models focus on generating a single object, synthetic indoor scenes, or synthetic outdoor scenes, while the generation of real-world outdoor scenes is rarely addressed. In this paper, we concentrate on generating a real-outdoor scene through learning a diffusion model on a real-world outdoor dataset. In contrast to synthetic data, real-outdoor datasets often contain more empty spaces due to sensor limitations, causing challenges in learning realoutdoor distributions. To address this issue, we exploit a triplane representation as a proxy form of scene distributions to be learned by our diffusion model. Furthermore, we propose a triplane manipulation that integrates seamlessly with our triplane diffusion model. The manipulation improves our diffusion model's applicability in a variety of downstream tasks related to outdoor scene generation such as scene inpainting, scene outpainting, and semantic", + "bbox": [ + 75, + 602, + 472, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene completion refinements. In experimental results, we demonstrate that our triplane diffusion model shows meaningful generation results compared with existing work in a real-outdoor dataset, SemanticKITTI. We also show our triplane manipulation facilitates seamlessly adding, removing, or modifying objects within a scene. Further, it also enables the expansion of scenes toward a city-level scale. Finally, we evaluate our method on semantic scene completion refinements where our diffusion model enhances predictions of semantic scene completion networks by learning scene distribution. Our code is available at https://github.com/zoomin-lee/SemCity.", + "bbox": [ + 500, + 571, + 890, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 784, + 632, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [16] have emerged as a promising generation tool owing to its state-of-the-art generation results in image domain [35, 36]. This advance has led to active exploration in extending diffusion models to 3D data generation across both academic and industrial groups. In the 3D domain, diffusion models have shown remarkable", + "bbox": [ + 496, + 809, + 893, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Both authors contributed equally to this work as co-first authors.", + "bbox": [ + 99, + 886, + 444, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28337", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "capabilities in generating diverse 3D forms (e.g., voxels, meshes) [24, 26]. While those 3D diffusion models primarily aim to craft a single object, generating scenes consisting of multiple objects remains a relatively unexplored area in the 3D diffusion domain.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene generative diffusion models focus on crafting both geometrically and semantically coherent environments. Compared with a single object generation, generating a scene with multiple objects requires an understanding of more complex geometric and semantic structure due to a larger spatial extent [41]. There are primarily two streams of scene generative diffusion models, each tailored to either indoor or outdoor settings. In particular, the outdoor environments have inherent challenges caused by a broader landscape than indoor ones.", + "bbox": [ + 75, + 167, + 468, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to leverage a triplane representation [10] for broader outdoor scenes, a method of factorizing 3D data onto three orthogonal 2D planes, as utilized in 3D object reconstruction and NeRF models [2, 31, 46]. We excavate its advantages in addressing the data sparsity problem typically found in outdoor datasets due to the sensor limitations (e.g., occlusions, range constraints) in capturing outdoor scenes. Triplane representation helps to reduce the inclusion of unnecessary empty information through the factorization of 3D data to 2D planes [10]. This efficiency in capturing relevant spatial detail makes it an effective tool for representing the many objects typically found in outdoor environments.", + "bbox": [ + 75, + 321, + 468, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we design our diffusion framework based on triplane representations. Our triplane autoencoder learns to compress a voxelized scene into a triplane representation by reconstructing semantic labels of the scene. Following this, the triplane diffusion model is trained and used to generate new scenes, as shown in Fig. 1(a), by creating novel triplanes based on the efficient representation. Further, we propose a triplane manipulation method, which extends our triplane diffusion model toward several practical tasks (i.e., scene inpainting, scene outpainting, and semantic scene completion refinements) as shown in Fig. 1(b-d). Our method can seamlessly add, remove, and modify objects in real-outdoor scenes while maintaining the semantic coherence of environments.", + "bbox": [ + 75, + 503, + 468, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 96, + 717, + 398, + 731 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We disclose the applicability of the triplane representation through generating semantic scenes for real-outdoor environments and extend its views in practical downstream tasks: scene inpainting, scene outpainting, and semantic scene completion refinement.", + "- We propose to manipulate triplane features during our diffusion process, facilitating seamlessly extending our method toward the downstream tasks.", + "- We demonstrate that the proposed method significantly enhances the quality of generated scenes in real-world outdoor environments." + ], + "bbox": [ + 76, + 734, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models [16] learn data distributions via iterative denoising processes based on score functions [40]. Its generated results have shown remarkably realistic appearances with high fidelity and diversity in a variety of 2D image synthesis such as outpainting [35, 57], inpainting [27, 35] and text-to-image generation [32, 36]. Built upon these achievements, diffusion models have also been extended into the 3D domain, generating impressive results in various 3D shapes, including voxel grids [24, 58], point clouds [28, 53, 54], meshes [26], and implicit functions [20, 37, 38, 48]. While these models can craft a single 3D object, our model focuses on generating a 3D scene composed of multiple objects using a categorical voxel data structure, which is a relatively under-explored area in the 3D diffusion domain.", + "bbox": [ + 496, + 117, + 890, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models for Scene Generation. In contrast to a single object generation, scene generation involves an understanding of the larger 3D space, causing more semantic and geometric complexities [41]. Diffusion models for scene generation have been studied in both indoor and outdoor environments. In indoor settings, diffusion models aim to learn distributions of relations among objects by representing them as scene graphs [19]. The scene graphs contain object attributes (e.g., location, orientation, and size), capturing the intricate inter-object relationships within bounded spaces [41, 55]. For outdoor scenes, the challenges are distinct, frequently including a lot of empty areas (e.g., sky, open areas) resulting from the broader landscapes. Traditional approach [22] has relied on discrete diffusion methods [17] on voxel space, necessitating a detailed representation of every air volume.", + "bbox": [ + 496, + 383, + 890, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we demonstrate that triplane diffusion is highly effective for generating real-outdoor scenes. By abstracting 3D spaces into three orthogonal 2D planes, triplane representation [10] effectively captures the vastness of outdoor environments, predominantly composed of air. Beyond its data efficiency, the triplane excels in focusing on other significant objects (e.g., vehicles, buildings) by allocating lesser attention to less informative elements like air. Our approach stands in stark contrast to prior work [22], which were constrained to synthetic datasets with considerably less empty space. In real datasets, inherent sensor limitations, such as a limited field-of-view, a limited ranges, and the inability to capture occluded areas like the rears of buildings, lead to a prevalence of empty space. Furthermore, we emphasize the versatility of our framework by demonstrating its extension to various downstream tasks, including scene inpainting, outpainting, and semantic scene completion refinement.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28338", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c0cb1b2c89dd38346bd4de7735e64d2632e7f6dcd0eeeb43c77c5675768555b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 697, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/826c63a724b7e726ff79b1c07c5ce0f74cd67193c227f600214d1796f3c40aae.jpg", + "image_caption": [ + "Figure 2. Overview of ours. (a) A 3D semantic map $\\mathbf{x}$ is encoded by a triplane encoder $f_{\\theta}$ and factorized to a triplane $\\mathbf{h}$ . The triplane coupled with a positional encoding $\\mathbb{P}\\mathbb{E}(\\mathbf{p})$ is decoded by an implicit decoder $g_{\\theta}$ , resulting in class probabilities for each coordinate $\\mathbf{p}$ . (b) Our triplane diffusion model $D_{\\phi}$ learns to generate a novel triplane for semantic scene generation via denoising diffusion process. (c) We further extend our triplane diffusion beyond a simple scene generation toward various practical scenarios by manipulating triplanes in (b)." + ], + "image_footnote": [], + "bbox": [ + 80, + 181, + 697, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f46fdb4aad04ff6e1b856abdf10c4d852743922fef0451ad2c6e129240378c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 890, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Inpainting and Outpainting. In 3D inpainting, the primary objective is to fill in missing portions or modify existing elements of 3D data while maintaining geometric consistency. Most existing works concentrate on single-object inpainting [3, 23, 44]; for instance, they seamlessly transit a 3D chair's leg count from three to four. Contrary to inpainting, 3D outpainting is to extrapolate a given scene over an unobserved space. The existing work [1] focuses on scene outpainting within bounded indoor environments such as rooms. While in 2D images, inpainting and outpainting are not restricted to a single object [12, 18, 52]. Likewise, in 3D space, we focus on scene-level inpainting, which seamlessly adds, removes, or modifies objects in a scene. Further, our scene-level outpainting is not constrained to the bounded scenes; we extrapolate the outdoor scene from sensor range (e.g., LiDAR) to city-scale.", + "bbox": [ + 75, + 378, + 472, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic Scene Completion. Semantic scene completion (SSC) [39] is pivotal for 3D scene understanding, where it jointly infers completion and semantic segmentation of the 3D scene from sensor observations such as RGB images [9, 25, 29] or point clouds [49-51]. In addition, SSC plays a crucial role in supporting comprehensive autonomous navigation systems, notably in essential downstream tasks like path planning [5, 11, 43] and map construction [13, 42]. Despite significant progress in the field, a persistent challenge is the semantic and geometric discrepancies between the SSC-estimated scenes and their real counterparts, as illustrated in Fig. 1(c). These discrepancies can undermine the performance of downstream tasks. Our triplane diffusion model can help to bridge this gap by exploiting 3D scene priors. This approach enhances the reliability and effectiveness of SSC, which is expected to improve its application in autonomous navigation systems.", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 377, + 589, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elucidate our triplane diffusion model and its extensions. Our triplane diffusion model aims to synthesize novel real-outdoor scenes by generating a triplane, a proxy representation, which effectively addresses the inherent challenges of real-outdoor scene synthesis. This triplane representation is learned by our triplane autoencoder, which abstracts the geometric and semantic intricacies of a scene into three orthogonal 2D feature planes, namely, the $xy$ , $xz$ , and $yz$ planes (Sec. 3.1). Then, our diffusion model learns triplane distributions of scenes, generating novel triplanes (Sec. 3.2). We extend our triplane diffusion model toward various practical scenarios: scene inpainting, scene outpainting, and semantic scene completion refinements (Sec. 3.3).", + "bbox": [ + 496, + 404, + 890, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Representing a Semantic Scene with Triplane", + "text_level": 1, + "bbox": [ + 498, + 631, + 883, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To represent a 3D scene as a triplane, our triplane autoencoder learns to compress a 3D scene into a triplane representation as shown in Fig. 2(a). The autoencoder consists of two modules: (1) an encoder $f_{\\theta}$ yielding a triplane, and (2) an implicit multi-layer perceptron (MLP) decoder $g_{\\theta}$ for reconstruction from the triplane.", + "bbox": [ + 496, + 656, + 890, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoder $f_{\\theta}$ takes a voxelized scene $\\mathbf{x} \\in \\mathbb{R}^{X \\times Y \\times Z}$ containing $N$ classes within a spatial grid of resolution $X \\times Y \\times Z$ . It then yields an axis-aligned triplane representation $\\mathbf{h} = [\\mathbf{h}^{xy}, \\mathbf{h}^{xz}, \\mathbf{h}^{yz}]$ . The triplane consists of three planes, each characterized by distinct dimensional properties: $\\mathbf{h}^{xy} \\in \\mathbb{R}^{C_h \\times X_h \\times Y_h}$ , $\\mathbf{h}^{xz} \\in \\mathbb{R}^{C_h \\times X_h \\times Z_h}$ , and $\\mathbf{h}^{yz} \\in \\mathbb{R}^{C_h \\times Y_h \\times Z_h}$ , where $C_h$ stands for a feature dimension, and $X_h, Y_h$ , and $Z_h$ denote spatial dimension of the triplane. During the encoding phase, a 3D feature volume is extracted by 3D convolutional layers from the scene $\\mathbf{x}$ , re", + "bbox": [ + 496, + 748, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28339", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sulting in the triplane via axis-wise average pooling. Given a 3D coordinate $\\mathbf{p} = (x,y,z)$ , the triplane is interpreted as a summation of vectors bilinearly interpolated from each plane: $\\mathbf{h}(\\mathbf{p}) = \\mathbf{h}^{xy}(x,y) + \\mathbf{h}^{xz}(x,z) + \\mathbf{h}^{yz}(y,z)$ .", + "bbox": [ + 75, + 90, + 467, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reconstruct the 3D scene $\\mathbf{x}$ , we decode the encoded triplane $\\mathbf{h}$ with an implicit MLP decoder $g_{\\theta}$ that predicts semantic class probabilities. The decoder takes the triplane vector $\\mathbf{h}(\\mathbf{p})$ with its sinusoidal positional embedding $\\mathrm{PE}(\\mathbf{p})$ [30], resulting in class probabilities $\\mathbf{c}(\\mathbf{p}) = g_{\\theta}(\\mathbf{h}(\\mathbf{p}),\\mathrm{PE}(\\mathbf{p}))\\in [0,1]^N$ . The positional embedding produces high-frequency features according to the coordinates $\\mathbf{p}$ , which helps the implicit decoder $g_{\\theta}$ represent high-frequency scene contents [45].", + "bbox": [ + 75, + 152, + 467, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The encoder $f_{\\theta}$ and the MLP decoder $g_{\\theta}$ are trained with the autoencoder loss $\\mathcal{L}_{\\mathrm{AE}}$ and scene label $\\mathbf{x}(\\mathbf{p})$ as:", + "bbox": [ + 76, + 287, + 467, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A E}} = \\mathbb {E} _ {\\mathbf {p} \\sim \\mathcal {P}} \\left[ \\ell_ {\\mathrm {C E}} (\\mathbf {c} (\\mathbf {p}), \\mathbf {x} (\\mathbf {p})) + \\lambda \\ell_ {\\mathrm {L Z}} (\\mathbf {c} (\\mathbf {p}), \\mathbf {x} (\\mathbf {p})) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 328, + 467, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a loss weight, and $\\mathcal{P}$ is the set of grid coordinates of the scene. We use the weighted cross-entropy loss $\\ell_{\\mathrm{CE}}$ [33] and the Lovász-softmax loss $\\ell_{\\mathrm{LZ}}$ [7] to learn imbalanced semantic distributions of the scene.", + "bbox": [ + 76, + 354, + 467, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Triplane Diffusion", + "text_level": 1, + "bbox": [ + 76, + 424, + 253, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the triplane representation of the 3D semantic scene, our triplane diffusion model $D_{\\phi}$ learns to generate a novel triplane through denoising diffusion probabilistic models [16] as shown in Fig. 2(b). This triplane generation leads to generation of 3D scene through decoding the generated triplane with the implicit MLP decoder $g_{\\theta}$ . Through the $x_0$ -parameterization [4], the diffusion model $D_{\\phi}$ is trained to reconstruct the triplane $\\mathbf{h}$ given its corrupted triplane $\\mathbf{h}_t$ sampled from a diffusion process $q(\\mathbf{h}_t|\\mathbf{h}) = \\mathcal{N}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{h}, (1 - \\bar{\\alpha}_t)\\mathbf{I})$ , where $\\mathcal{N}$ is the Gaussian distribution, $\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i$ , and $\\alpha_t = 1 - \\beta_t$ with a variance schedule $\\beta_t$ . The diffusion process $q(\\mathbf{h}_t|\\mathbf{h})$ is derived from the Markovian chain rule with a single step's diffusion process $q(\\mathbf{h}_t|\\mathbf{h}_{t-1}) = \\mathcal{N}(\\sqrt{1 - \\beta_t}\\mathbf{h}_{t-1}, \\beta_t\\mathbf{I})$ . Thus, the triplane diffusion loss is defined as:", + "bbox": [ + 75, + 446, + 467, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D}} = \\mathbb {E} _ {t \\sim \\mathcal {U} (1, T)} | | \\mathbf {h} - D _ {\\phi} (\\mathbf {h} _ {t}, t) | | _ {p}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 685, + 467, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the number of denoising steps, and $p$ represents the order of the norm. The timestep $t$ is sampled from the discrete uniform distribution $\\mathcal{U}$ .", + "bbox": [ + 75, + 710, + 467, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After training, the diffusion model $D_{\\phi}$ generates a novel triplane $\\mathbf{h}_0$ via the iterative DDPM generation process [16] starting from $\\mathbf{h}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ :", + "bbox": [ + 75, + 757, + 467, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {t - 1} \\sim \\mathcal {N} \\left(\\gamma_ {t} \\mathbf {h} _ {t} + \\delta_ {t} D _ {\\phi} \\left(\\mathbf {h} _ {t}, t\\right), \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 811, + 467, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\gamma_{t} \\coloneqq \\sqrt{\\alpha_{t}} (1 - \\bar{\\alpha}_{t-1}) / (1 - \\bar{\\alpha}_{t})$ and $\\delta_{t} \\coloneqq \\sqrt{\\bar{\\alpha}_{t-1}} \\beta_{t} / (1 - \\bar{\\alpha}_{t})$ . From the generated triplane $\\mathbf{h}_0$ , we generate a novel 3D semantic scene $\\mathbf{x}_0$ by querying coordinates $\\mathbf{p}$ to the implicit decoder, i.e., $g_{\\theta}(\\mathbf{h}_0(\\mathbf{p}), \\mathsf{PE}(\\mathbf{p}))$ .", + "bbox": [ + 76, + 839, + 467, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Applications with Triplane Manipulation", + "text_level": 1, + "bbox": [ + 500, + 90, + 849, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on the triplane diffusion process (Sec. 3.2), we propose a triplane manipulation that allows our model to facilitate a variety of practical downstream tasks with few modifications, as illustrated in Fig. 2(c).", + "bbox": [ + 498, + 114, + 890, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Inpainting. Our scene inpainting randomly edits a 3D scene, seamlessly adding, modifying, or removing objects while maintaining the consistency and realism of the scene. For instance, the inpainting includes scenarios where cars or sidewalks appear and then disappear, or vice versa, as shown in Fig. 1(d). Inspired by the RePaint sampling strategy [27], we propose a 3D-aware inpainting approach with semantic coherence. RePaint focuses on the image domain without explicitly considering the fidelity of the underlying 3D scene. To facilitate 3D-aware inpainting, we inpaint the triplane, serving as a compact proxy representation for the scene. We define a binary spatial trimask $\\mathbf{m} = [\\mathbf{m}^{xy},\\mathbf{m}^{xz},\\mathbf{m}^{yz}]$ covering inpainting regions on triplane space, allowing us to control the generation process on the masked region. The trimask $\\mathbf{m}$ is set to have ones for inpainting regions and zeros for others. We override the $t$ -th triplane $\\mathbf{h}_t = [\\mathbf{h}_t^{xy},\\mathbf{h}_t^{xz},\\mathbf{h}_t^{yz}]$ of the generation process (Eq. 3) as follows:", + "bbox": [ + 496, + 204, + 890, + 476 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {t} \\leftarrow \\mathbf {m} \\otimes \\mathbf {h} _ {t} + (\\mathbf {1} - \\mathbf {m}) \\otimes \\mathbf {h} _ {t} ^ {\\text {k n o w n}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 492, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ is the element-wise product, and the known triplane $\\mathbf{h}_t^{\\mathrm{known}}$ for intact regions is sampled from the diffusion process, i.e., $q(\\mathbf{h}_t^{\\mathrm{known}}|\\mathbf{h})\\coloneqq q(\\mathbf{h}_t|\\mathbf{h})$ , which adheres to a known Gaussian distribution.", + "bbox": [ + 498, + 523, + 890, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Outpainting. Our scene outpainting extends the boundaries of the 3D scene without additional training as with inpainting. To seamlessly outpaint a scene, the regions to be extended should be conditioned on the original scene. We propose to inject this intuition into triplane representation as shown in Fig. 2(c). For covering regions to be outpainted, our diffusion model generates a novel triplane that is partially overlapped with the original triplane. We implement this with the concept of the trimask m and the known triplane $\\mathbf{h}_t^{\\mathrm{known}}$ as in our scene inpainting. The trimask m covers regions to be outpainted, and the $\\mathbf{h}_t^{\\mathrm{known}}$ is obtained from the intersections of triplanes between original and outpainted regions. Based on the outpainting strategy, we extend a given scene toward cardinal and intercardinal directions, facilitating the creation of an unbounded scene. While our triplane diffusion model is trained using triplanes of a fixed size, it demonstrates the capability to outpaint scenes to be several times bigger than the original scene, as illustrated in Fig. 1(c).", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28340", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a5dc446c9fe8e2b8234df33be7aff550d99df5adda0b47342e66fffe12ccf457.jpg", + "image_caption": [ + "Figure 3. Scene generation results using both real and synthetic outdoor datasets - SemanticKITTI [6] and CarlaSC [47]. Our results showcase the effective generation of overall structures, including roads and buildings, along with detailed objects such as cars." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 893, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Scene Completion Refinement. SSC models complete and segment a 3D scene from sensor observations such as images or point clouds. We observe that the SSC results show a geometric and semantic discrepancy compared with data distributions, as shown in Fig. 1(b). We extend our triplane diffusion model to refine predictions of SSC models toward reducing the discrepancy. To effectively condition our triplane diffusion scheme to SSC model's prediction $\\mathbf{x}^{\\mathrm{ssc}}$ , we utilize its triplane representation $\\mathbf{h}^{\\mathrm{ssc}} = f_{\\theta}(\\mathbf{x}^{\\mathrm{ssc}})$ derived by our triplane encoder $f_{\\theta}$ . We extend our triplane diffusion scheme with a simple modification of the triplane $\\mathbf{h}_t$ in the diffusion loss (Eq. 2) and the generation process (Eq. 3) as follows:", + "bbox": [ + 75, + 500, + 472, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {t} = \\mathbf {h} _ {t} ^ {\\mathrm {s s c}} \\oplus \\mathbf {h} ^ {\\mathrm {s s c}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 707, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ is concatenation, and $\\mathbf{h}_t^{\\mathrm{ssc}}$ is a $t$ -th diffused triplane sampled by the DDPS diffusion process [21] with the SSC prediction's triplane $\\mathbf{h}^{\\mathrm{ssc}}$ .", + "bbox": [ + 75, + 733, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Details", + "text_level": 1, + "bbox": [ + 76, + 816, + 276, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Dataset. We validate our method on the SemanticKITTI [6] and CarlaSC [47] datasets. SemanticKITTI provides 3D semantic scenes of real-outdoor environments with labels for 20 semantic classes. Each", + "bbox": [ + 75, + 839, + 470, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scene is represented by a voxel grid of $256 \\times 256 \\times 32$ , covering an area of $51.2\\mathrm{m}$ in front of the car, extending $51.2\\mathrm{m}$ on each side, and reaching up to a height of $6.4\\mathrm{m}$ . The dataset retains object motion traces as a result of sensor frame integration, which is employed to establish a dense ground truth. In contrast, CarlaSC is a synthetic dataset that provides 3D semantic outdoor scenes without the trace of moving objects. The dataset contains annotated 11 semantic classes with a voxel grid of $128 \\times 128 \\times 8$ and covers a distance of $25.6\\mathrm{m}$ in front and behind the car, $25.6\\mathrm{m}$ laterally on each side, and $3\\mathrm{m}$ in height.", + "bbox": [ + 496, + 500, + 893, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our experiments are deployed on a single NVIDIA RTX 3090 GPU with a batch size of 4 for the triplane autoencoder and 18 for the triplane diffusion model. For the triplane autoencoder, the input scene is encoded to triplane with a spatial resolution $(X_h, Y_h, Z_h) = (128, 128, 32)$ , and the feature dimension $C_h$ is 16. The loss weight $\\lambda$ in Eq. 1 is set to 1.0. The order of the norm $p$ in Eq. 2 is set to 1 for SSC refinements and 2 for other cases. For the diffusion model, the learning rate is initialized to 1e-4 and then decreases linearly. During the diffusion process, we use the default settings [34] with 100 time steps $(T)$ . For our triplane inpainting and outpainting, we employ the RePaint sampling strategy [27] as a reference and perform a repaint with 5 resampling and a jump size of 20.", + "bbox": [ + 496, + 689, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fb82f8d7a57c34f36b32e7df2dd537a33b90bbfbc8709e4bd7525b6d1b51da3b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFID ↓KID ↓IS ↑Prec ↑Rec ↑
SemanticKITTI [6]
SSD [22]112.820.122.230.010.08
SemCity (Ours)56.550.043.250.390.32
CarlaSC [47]
SSD [22]87.390.092.440.140.07
SemCity (Ours)40.630.023.510.310.09
", + "bbox": [ + 78, + 88, + 467, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results of semantic scene generation. Each metric is computed between the rendered image of the generated scene and the ones of the actual scene at a resolution of $1440 \\times 2048$ .", + "bbox": [ + 75, + 200, + 468, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following the scene generation works [41, 55], we evaluate the performance of semantic scene generation by examining both the diversity and fidelity of 3D semantic scenes within the rendered images. We use recall to evaluate diversity, while precision and inception score (IS) are used to evaluate fidelity. The Fréchet Inception Distance (FID) [15] and Kernel Inception Distance (KID) [8] metrics are also utilized, as they reflect the combined effect of both diversity and fidelity on scene quality [14]. In terms of semantic scene completion (SSC) refinement performance, we follow the protocols defined in SSC works [9, 29, 49, 50]. The Intersection-over-Union (IoU) metric is used to quantify scene completeness, while the mean IoU (mIoU) provides a measure for the quality of semantic segmentation. These metrics together enable a comprehensive evaluation of how well the semantic scene completion methods perform in terms of accurately filling in and labeling the scene components.", + "bbox": [ + 75, + 263, + 470, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Semantic Scene Generation", + "text_level": 1, + "bbox": [ + 76, + 545, + 323, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 illustrates a qualitative comparison on the SemanticKITTI [6] and CarlaSC [47] datasets. SSD [22] shows impressive results on the CarlaSC dataset. However, its performance on the SemanticKITTI dataset, which is a real-world dataset, is notably constrained. This limitation primarily arises from the SSD's voxel-based representation, which struggles with the more prevalent empty spaces in real-outdoor datasets compared to synthetic ones. This issue is especially evident in the generation of buildings and roads, where SSD often fails to define boundaries accurately. Furthermore, the model's limitations extend to representing finer structures, such as trunks and leaves, as well as traffic light poles and signals. It also struggles with generating uniform shapes of vehicles, often resulting in irregular shapes. In contrast, our method demonstrates the ability to effectively synthesize detailed scenes even on the real dataset, as illustrated in Fig. 3. It performs better than SSD [22] in accurately capturing complex building shapes on the CarlaSC dataset. In addition, our method exhibits remarkable proficiency in generating the overall contours of roads and buildings, along with intricate details on the SemanticKITTI dataset. Tab. 1 provides a detailed compar", + "bbox": [ + 75, + 568, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f08e41b86a070b9dd13c86786f4d738f88654973ddfad42dc94cd245d130cf51.jpg", + "image_caption": [ + "Figure 4. Higher-resolution scene generation. Building upon our implicit decoder, higher-resolution scene $(1024\\times 1024\\times 128)$ can be generated compared with a resolution of training dataset $(256\\times 256\\times 32)$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 697, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20d615566abf8ce92c8dc0b6b718f12c308db3b0fa709aecf4dea9df4d99628c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 890, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31e88e24c962750778056f56b2a984e75b8cc69ec42fda94cd144eabbbc23ef0.jpg", + "image_caption": [ + "Figure 5. Scene inpainting of our method. The red boxes denote inpainting regions. (a) and (b) show our inpainting examples from reference images." + ], + "image_footnote": [], + "bbox": [ + 503, + 292, + 890, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ative evaluation using various metrics. Our model shows significant improvements in both the fidelity and diversity of the generated scenes. Moreover, our generated result is not tied to fixed resolution by means of the implicit neural representation, as depicted in Fig. 4. For additional results, please refer to the Supplementary Material.", + "bbox": [ + 496, + 702, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Applications of Triplane Diffusion", + "text_level": 1, + "bbox": [ + 498, + 801, + 797, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Inpainting. Fig. 5 presents the qualitative results of our inpainting, demonstrating its effectiveness in inpainting both small and large regions within a scene, while maintaining the coherence of 3D contexts. In detail, the second row of (a) and (b) illustrates the model's seamless removal", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28342", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc51c5c0950bad760add063b152dd4947ca9b0e8423036bcafff1a89aad4f9d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 99, + 890, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3f28b3c91df46837035313a79c64064d686caf70078b6f3ed4996796de6861a.jpg", + "image_caption": [ + "Figure 6. Our scene outpainting results. The red box at the center of the figure in the first row represents the given scene for outpainting. The zoomed views of the outpainted scene are depicted in the second row. The outpainted scene is expanded from the given size of $256 \\times 256 \\times 32$ to $1792 \\times 3328 \\times 32$ ." + ], + "image_footnote": [], + "bbox": [ + 80, + 218, + 320, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f1712c7bfa47ec1efce17dac5fd8a5a35134609050cec64c712e36dc70750e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 218, + 576, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fb3650c0b537da7b60f831f511178f19f79db80c939678746cdda2f621a8db3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 218, + 888, + 314 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f639df30dcff6434d27938b80c54456d811edb0b756585100adeca64863262f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SSC InputMethodIoU ↑mIoU ↑
RGBMonoScene [9]37.1211.50
MonoScene + Ours50.4417.08
OccDepth [29]41.6012.84
OccDepth + Ours50.2016.79
Point CloudSSA-SC [51]58.2524.54
SSA-SC + Ours60.7125.58
SCPNet [49]50.2437.55
SCPNet + Ours59.2538.19
", + "bbox": [ + 78, + 375, + 467, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of refining SSC on SemanticKITTI validation set [6]. The results are based on the weights released by the authors on GitHub.", + "bbox": [ + 76, + 507, + 468, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of vehicles, which harmonizes with the adjacent road. The third row demonstrates the insertion of new entities — a vehicle in (a) and a person in (b) — that are contextually congruent with the reference scene. The fourth row in (a) exemplifies the model's dual functionality in both modifying and adding vehicles within the scene. The fifth row in both columns underscores the model's proficiency in modifying scenes. Here, the model alters existing scene components, showcasing its ability to transform the overall ambiance of the scene. These results show our model's adeptness not just in object-level inpainting but also in scene-level inpainting.", + "bbox": [ + 75, + 573, + 468, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Outpainting. Fig. 6 illustrates a generated outpainting city-level scene, extending a $256 \\times 256 \\times 32$ scene to a substantial $1792 \\times 3328 \\times 32$ landscape. Although our model was not designed to generate cityscapes, it demonstrates the capability of maintaining coherence over a large area. The roads are connected in a meaningful and varied way, and various objects such as buildings, cars, and people have been created around them. Please check the details of", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/689f5b59eb6ae1552c4116a3092abcdb3fd6d14e3a5f27725332f7b37af2ae17.jpg", + "image_caption": [ + "Figure 7. Semantic scene completion refinement. The SSC metrics are reported in the parentheses as (IoU, mIoU)." + ], + "image_footnote": [], + "bbox": [ + 500, + 376, + 890, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the scene in Fig. 6.", + "bbox": [ + 500, + 650, + 627, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Scene Completion Refinement. In Fig. 7, there is a notable semantic and geometric discrepancy between scenes predicted by existing semantic scene completion (SSC) methods and their real scene counterparts. Our model helps to bridge this gap by employing a 3D scene prior, which is effectively modeled through our diffusion model. SCPNet [49] also attempts to learn geometric priors from a teacher model trained with merged sequential frames. Yet, its completion network is inclined toward conservative estimations, which frequently result in partially filled spaces. In the case of SSA-SC [51], which utilizes bird's-eye view features, it is sometimes inappropriate due to the label's inherent limitations of the bird's-eye perspective in capturing certain semantic details. This discrepancy", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28343", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e45ee7569fe3152caba6e7c7df31901e4786ef791f0699cf51da81d0e164a95e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 93, + 269, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e05e11e232fa582a276e9d5dd1f5b96f50dda6f84e1345341f43b9d55752a058.jpg", + "image_caption": [ + "(b)", + "Figure 8. Semantic scene to RGB. (a) Semantic maps are rendered into a driving view from our generated scenes. (b) We generate images through ControlNet [56], an easily accessible imaged-to-image model, with generated semantic maps without shadows." + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 464, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8933d535b3a98fad44528baec0129a4af278c05d1c38e0787d8bf46b2e25e0ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓KID ↓IS ↑Prec ↑Rec ↑
Ours56.550.043.250.390.32
w/o PE59.140.043.210.300.31
xy-plane (w/o triplane)75.560.063.080.280.27
3D-volume (w/o triplane)136.500.151.960.140.12
", + "bbox": [ + 78, + 268, + 467, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on scene generation. We ablate our model variants on SemanticKITTI [6].", + "bbox": [ + 75, + 354, + 467, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "issue is even more evident in RGB-based methods, such as MonoScene [9] and OccDepth [29], which often exhibit diminished sharpness when image features are projected into 3D space. While SSC models show variations from real-world data distributions, our model shows potential in aligning these more closely with reality, as shown in Fig. 7.", + "bbox": [ + 75, + 401, + 467, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As indicated in Tab. 2, our SSC refinement process appears to offer improvements to all state-of-the-art SSC models. These preliminary results suggest our model's effectiveness in providing not just more accurate semantic segmentation but also a more complete scene.", + "bbox": [ + 75, + 492, + 467, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic Scene to RGB Image. As illustrated in Fig. 8, we conducted an image-to-image generation experiment. The semantic maps are rendered into a driving view without shadows; then, they are utilized as inputs for ControlNet [56]. The generated RGB images are geometrically and semantically plausible but display a synthetic quality since the pretrained ControlNet was not trained on actual autonomous driving datasets.", + "bbox": [ + 75, + 585, + 467, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation", + "text_level": 1, + "bbox": [ + 76, + 715, + 181, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 3, we conducted ablation studies on our triplane diffusion model for scene generation, focusing on two key design elements:", + "bbox": [ + 75, + 739, + 467, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positional Embedding. Our variant excluded positional embedding PE, which generates high-frequency features critical for detailed scene reconstruction. Its absence resulted in lower performance across all metrics.", + "bbox": [ + 75, + 801, + 467, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Triplane Representation. We evaluated the effectiveness of the triplane representation for real-outdoor scene gen", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eration. The triplane and $xy$ -plane enhance the generation quality compared with 3D features, while the $xy$ -plane shows lower performance than triplanes. We suspect that the excessive factorization limits the representation capability of the $xy$ -plane compared to the triplane.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitation", + "text_level": 1, + "bbox": [ + 500, + 174, + 617, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our model demonstrates significant progress in the generation of 3D real-outdoor scenes, it inherently reflects the characteristics of its training data, as detailed in Sec. 4.1. This reliance introduces several limitations. One notable challenge is the model's difficulty in accurately depicting areas occluded from the sensor's viewpoint, such as the rear sides of buildings, often leading to their incomplete representation in our generated scenes. Moreover, since the dataset is captured from a driving view, there is an inherent shortfall in capturing the full height of buildings. This brings a partial representation of the vertical structure of buildings and other tall elements in the scenes. Another issue is the model tends to produce traces of moving objects stemming from dataset pre-processing that merges sequential frames. For future work, incorporating prior knowledge of the city into the model could yield better results, such as addressing occluded areas or building heights.", + "bbox": [ + 496, + 196, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 467, + 617, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a diffusion framework called SemCity for real-world outdoor scene generation. The seminal idea is to generate a scene by factorizing real-outdoor scenes into triplane representations. Our triplane representation outperforms traditional voxel-based approaches, producing scenes that are not only visually more appealing but also rich in semantic detail, effectively capturing the complexity of various objects within the scene. Ours is not constrained by fixed resolutions thanks to the incorporation of an implicit neural representation. We have further expanded the capabilities of our triplane diffusion model to several practical applications, including scene inpainting, scene outpainting, and semantic scene completion refinement. Specifically, by manipulating triplanes during the diffusion process, we achieve seamless inpainting and outpainting at both the object and scene levels. Ours is used to more closely align the scenes predicted by existing semantic scene completion methods with the actual data distribution using our learned 3D prior. We believe that our work provides a road map of real-outdoor scene generation to research communities.", + "bbox": [ + 496, + 492, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by IITP (Institute of Information & Communications Technology Planning & Evaluation) and ITRC (Information Technology Research Center), funded by Korea government (MSIT) (RS-2023-00237965(2024) and IITP-2024-2020-0-01460). Prof. Sung-Eui Yoon is a corresponding author.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28344", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Abbasi, Sinan Kalkan, and Yusuf Sahillioğlu. Deep 3d semantic scene extrapolation. The Visual Computer, 35:271-279, 2019. 3", + "[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12608-12618, 2023. 2", + "[3] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12608-12618, 2023. 3", + "[4] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 4", + "[5] Luca Bartolomei, Lucas Teixeira, and Margarita Chli. Perception-aware path planning for uavs using semantic segmentation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5808-5815, 2020. 3", + "[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 5, 6, 7, 8", + "[7] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 4", + "[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6", + "[9] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 6, 7, 8", + "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[11] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 3", + "[12] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 3", + "[13] Xu Cui, Chenggang Lu, and Jinxiang Wang. 3d semantic map construction using improved orb-slam2 for mobile robot in edge computing environment. IEEE Access, 8:67179-67191, 2020. 3", + "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 4", + "[17] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2", + "[18] Changho Jo, Woobin Im, and Sung-Eui Yoon. In-n-out: Towards good initialization for inpainting and outpainting. arXiv preprint arXiv:2106.13953, 2021. 3", + "[19] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 2", + "[20] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2", + "[21] Zeqiang Lai, Yuchen Duan, Jifeng Dai, Ziheng Li, Ying Fu, Hongsheng Li, Yu Qiao, and Wenhai Wang. Denoising diffusion semantic segmentation with mask prior modeling. arXiv preprint arXiv:2306.01721, 2023. 5", + "[22] Jumin Lee, Woobin Im, Sebin Lee, and Sung-Eui Yoon. Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527, 2023. 2, 6", + "[23] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. arXiv preprint arXiv:2212.05993, 2022. 3", + "[24] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2", + "[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 3", + "[26] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28345", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 4, 5", + "[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 2", + "[29] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 3, 6, 7, 8", + "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 4", + "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[33] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 4", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5", + "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1, 2", + "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2", + "[37] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2", + "[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2", + "[39] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene com" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pletion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 3", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 2", + "[41] Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner. Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207, 2023. 2, 6", + "[42] Chaoqun Wang, Wenzheng Chi, Yuxiang Sun, and Max Q-H Meng. Autonomous robotic exploration by incremental road map construction. IEEE Transactions on Automation Science and Engineering, 16(4):1720-1731, 2019. 3", + "[43] Chaoqun Wang, Delong Zhu, Teng Li, Max Q-H Meng, and Clarence W De Silva. Efficient autonomous robotic exploration with semantic road map in indoor environments. IEEE Robotics and Automation Letters, 4(3):2989-2996, 2019. 3", + "[44] Xinying Wang, Dikai Xu, and Fangming Gu. 3d model inpainting based on 3d deep convolutional generative adversarial network. IEEE Access, 8:170355-170363, 2020. 3", + "[45] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Pet-neus: Positional encoding tri-planes for neural surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12598–12607, 2023. 4", + "[46] Zhen Wang, Shijie Zhou, Jeong Joon Park, Despoina Paschalidou, Suya You, Gordon Wetzstein, Leonidas Guibas, and Achuta Kadambi. Alto: Alternating latent topologies for implicit 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 259-270, 2023. 2", + "[47] Joey Wilson, Jingyu Song, Yuewei Fu, Arthur Zhang, Andrew Capodieci, Paramsothy Jayakumar, Kira Barton, and Maani Ghaffari. *Motionsc: Data set and network for real-time semantic mapping in dynamic environments*. arXiv preprint arXiv:2203.07060, 2022. 5, 6", + "[48] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2", + "[49] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 3, 6, 7", + "[50] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6", + "[51] Xuemeng Yang, Hao Zou, Xin Kong, Tianxin Huang, Yong Liu, Wanlong Li, Feng Wen, and Hongbo Zhang. Semantic segmentation-assisted scene completion for lidar point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3555-3562. IEEE, 2021. 3, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28346", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3", + "[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4206-4216, 2023. 2", + "[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[55] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2, 6", + "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 8", + "[57] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10188-10198, 2023. 2", + "[58] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28347", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_model.json b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f5ed4114c8a57c4da124a6979848c36d1d6a475 --- /dev/null +++ b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.797, + 0.154 + ], + "angle": 0, + "content": "SemCity: Semantic Scene Generation with Triplane Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.18, + 0.852, + 0.234 + ], + "angle": 0, + "content": "Jumin Lee\\(^{1*}\\) Sebin Lee\\(^{1*}\\) Changho Jo\\(^{2}\\) Woobin Im\\(^{1}\\) Juhyeong Seon\\(^{1}\\) Sung-Eui Yoon\\(^{1}\\) \n\\(^{1}\\)KAIST \\(^{2}\\)Neosapience, Inc." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.251, + 0.283, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.465, + 0.258, + 0.479 + ], + "angle": 0, + "content": "(a) Scene generation" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.253, + 0.721, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.371, + 0.488, + 0.385 + ], + "angle": 0, + "content": "(c) Scene outpainting" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.257, + 0.882, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.387, + 0.681, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.463, + 0.642, + 0.477 + ], + "angle": 0, + "content": "(b) Semantic scene completion refinement" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.401, + 0.882, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.463, + 0.862, + 0.477 + ], + "angle": 0, + "content": "(d) Scene inpainting" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.491, + 0.894, + 0.547 + ], + "angle": 0, + "content": "Figure 1. We introduce a diffusion framework, SemCity, designed for generating semantic scenes in real-world outdoor environments as shown in (a). We extend our diffusion model to various practical tasks: semantic scene completion refinement, scene outpainting, and scene inpainting. For instance, the comprehensive scenario is displayed in \\((\\mathrm{b})\\rightarrow (\\mathrm{c})\\rightarrow (\\mathrm{d})\\) : the refined scene (SSC refinement) (b) is outpainted to a broader scene (c); then, an object (in this case, a car) is seamlessly integrated into the scene via our inpainting process (d)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.571, + 0.314, + 0.587 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.473, + 0.876 + ], + "angle": 0, + "content": "We present \"SemCity,\" a 3D diffusion model for semantic scene generation in real-world outdoor environments. Most 3D diffusion models focus on generating a single object, synthetic indoor scenes, or synthetic outdoor scenes, while the generation of real-world outdoor scenes is rarely addressed. In this paper, we concentrate on generating a real-outdoor scene through learning a diffusion model on a real-world outdoor dataset. In contrast to synthetic data, real-outdoor datasets often contain more empty spaces due to sensor limitations, causing challenges in learning realoutdoor distributions. To address this issue, we exploit a triplane representation as a proxy form of scene distributions to be learned by our diffusion model. Furthermore, we propose a triplane manipulation that integrates seamlessly with our triplane diffusion model. The manipulation improves our diffusion model's applicability in a variety of downstream tasks related to outdoor scene generation such as scene inpainting, scene outpainting, and semantic" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.573, + 0.892, + 0.753 + ], + "angle": 0, + "content": "scene completion refinements. In experimental results, we demonstrate that our triplane diffusion model shows meaningful generation results compared with existing work in a real-outdoor dataset, SemanticKITTI. We also show our triplane manipulation facilitates seamlessly adding, removing, or modifying objects within a scene. Further, it also enables the expansion of scenes toward a city-level scale. Finally, we evaluate our method on semantic scene completion refinements where our diffusion model enhances predictions of semantic scene completion networks by learning scene distribution. Our code is available at https://github.com/zoomin-lee/SemCity." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.785, + 0.633, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Diffusion models [16] have emerged as a promising generation tool owing to its state-of-the-art generation results in image domain [35, 36]. This advance has led to active exploration in extending diffusion models to 3D data generation across both academic and industrial groups. In the 3D domain, diffusion models have shown remarkable" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.446, + 0.9 + ], + "angle": 0, + "content": "*Both authors contributed equally to this work as co-first authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28337" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "capabilities in generating diverse 3D forms (e.g., voxels, meshes) [24, 26]. While those 3D diffusion models primarily aim to craft a single object, generating scenes consisting of multiple objects remains a relatively unexplored area in the 3D diffusion domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.47, + 0.319 + ], + "angle": 0, + "content": "Scene generative diffusion models focus on crafting both geometrically and semantically coherent environments. Compared with a single object generation, generating a scene with multiple objects requires an understanding of more complex geometric and semantic structure due to a larger spatial extent [41]. There are primarily two streams of scene generative diffusion models, each tailored to either indoor or outdoor settings. In particular, the outdoor environments have inherent challenges caused by a broader landscape than indoor ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.502 + ], + "angle": 0, + "content": "We propose to leverage a triplane representation [10] for broader outdoor scenes, a method of factorizing 3D data onto three orthogonal 2D planes, as utilized in 3D object reconstruction and NeRF models [2, 31, 46]. We excavate its advantages in addressing the data sparsity problem typically found in outdoor datasets due to the sensor limitations (e.g., occlusions, range constraints) in capturing outdoor scenes. Triplane representation helps to reduce the inclusion of unnecessary empty information through the factorization of 3D data to 2D planes [10]. This efficiency in capturing relevant spatial detail makes it an effective tool for representing the many objects typically found in outdoor environments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.716 + ], + "angle": 0, + "content": "In this paper, we design our diffusion framework based on triplane representations. Our triplane autoencoder learns to compress a voxelized scene into a triplane representation by reconstructing semantic labels of the scene. Following this, the triplane diffusion model is trained and used to generate new scenes, as shown in Fig. 1(a), by creating novel triplanes based on the efficient representation. Further, we propose a triplane manipulation method, which extends our triplane diffusion model toward several practical tasks (i.e., scene inpainting, scene outpainting, and semantic scene completion refinements) as shown in Fig. 1(b-d). Our method can seamlessly add, remove, and modify objects in real-outdoor scenes while maintaining the semantic coherence of environments." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.718, + 0.4, + 0.732 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.735, + 0.468, + 0.809 + ], + "angle": 0, + "content": "- We disclose the applicability of the triplane representation through generating semantic scenes for real-outdoor environments and extend its views in practical downstream tasks: scene inpainting, scene outpainting, and semantic scene completion refinement." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.853 + ], + "angle": 0, + "content": "- We propose to manipulate triplane features during our diffusion process, facilitating seamlessly extending our method toward the downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.468, + 0.899 + ], + "angle": 0, + "content": "- We demonstrate that the proposed method significantly enhances the quality of generated scenes in real-world outdoor environments." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.735, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.118, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models [16] learn data distributions via iterative denoising processes based on score functions [40]. Its generated results have shown remarkably realistic appearances with high fidelity and diversity in a variety of 2D image synthesis such as outpainting [35, 57], inpainting [27, 35] and text-to-image generation [32, 36]. Built upon these achievements, diffusion models have also been extended into the 3D domain, generating impressive results in various 3D shapes, including voxel grids [24, 58], point clouds [28, 53, 54], meshes [26], and implicit functions [20, 37, 38, 48]. While these models can craft a single 3D object, our model focuses on generating a 3D scene composed of multiple objects using a categorical voxel data structure, which is a relatively under-explored area in the 3D diffusion domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.626 + ], + "angle": 0, + "content": "Diffusion Models for Scene Generation. In contrast to a single object generation, scene generation involves an understanding of the larger 3D space, causing more semantic and geometric complexities [41]. Diffusion models for scene generation have been studied in both indoor and outdoor environments. In indoor settings, diffusion models aim to learn distributions of relations among objects by representing them as scene graphs [19]. The scene graphs contain object attributes (e.g., location, orientation, and size), capturing the intricate inter-object relationships within bounded spaces [41, 55]. For outdoor scenes, the challenges are distinct, frequently including a lot of empty areas (e.g., sky, open areas) resulting from the broader landscapes. Traditional approach [22] has relied on discrete diffusion methods [17] on voxel space, necessitating a detailed representation of every air volume." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we demonstrate that triplane diffusion is highly effective for generating real-outdoor scenes. By abstracting 3D spaces into three orthogonal 2D planes, triplane representation [10] effectively captures the vastness of outdoor environments, predominantly composed of air. Beyond its data efficiency, the triplane excels in focusing on other significant objects (e.g., vehicles, buildings) by allocating lesser attention to less informative elements like air. Our approach stands in stark contrast to prior work [22], which were constrained to synthetic datasets with considerably less empty space. In real datasets, inherent sensor limitations, such as a limited field-of-view, a limited ranges, and the inability to capture occluded areas like the rears of buildings, lead to a prevalence of empty space. Furthermore, we emphasize the versatility of our framework by demonstrating its extension to various downstream tasks, including scene inpainting, outpainting, and semantic scene completion refinement." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28338" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.698, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.182, + 0.698, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.891, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.291, + 0.893, + 0.349 + ], + "angle": 0, + "content": "Figure 2. Overview of ours. (a) A 3D semantic map \\(\\mathbf{x}\\) is encoded by a triplane encoder \\(f_{\\theta}\\) and factorized to a triplane \\(\\mathbf{h}\\). The triplane coupled with a positional encoding \\(\\mathbb{P}\\mathbb{E}(\\mathbf{p})\\) is decoded by an implicit decoder \\(g_{\\theta}\\), resulting in class probabilities for each coordinate \\(\\mathbf{p}\\). (b) Our triplane diffusion model \\(D_{\\phi}\\) learns to generate a novel triplane for semantic scene generation via denoising diffusion process. (c) We further extend our triplane diffusion beyond a simple scene generation toward various practical scenarios by manipulating triplanes in (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.622 + ], + "angle": 0, + "content": "3D Inpainting and Outpainting. In 3D inpainting, the primary objective is to fill in missing portions or modify existing elements of 3D data while maintaining geometric consistency. Most existing works concentrate on single-object inpainting [3, 23, 44]; for instance, they seamlessly transit a 3D chair's leg count from three to four. Contrary to inpainting, 3D outpainting is to extrapolate a given scene over an unobserved space. The existing work [1] focuses on scene outpainting within bounded indoor environments such as rooms. While in 2D images, inpainting and outpainting are not restricted to a single object [12, 18, 52]. Likewise, in 3D space, we focus on scene-level inpainting, which seamlessly adds, removes, or modifies objects in a scene. Further, our scene-level outpainting is not constrained to the bounded scenes; we extrapolate the outdoor scene from sensor range (e.g., LiDAR) to city-scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Semantic Scene Completion. Semantic scene completion (SSC) [39] is pivotal for 3D scene understanding, where it jointly infers completion and semantic segmentation of the 3D scene from sensor observations such as RGB images [9, 25, 29] or point clouds [49-51]. In addition, SSC plays a crucial role in supporting comprehensive autonomous navigation systems, notably in essential downstream tasks like path planning [5, 11, 43] and map construction [13, 42]. Despite significant progress in the field, a persistent challenge is the semantic and geometric discrepancies between the SSC-estimated scenes and their real counterparts, as illustrated in Fig. 1(c). These discrepancies can undermine the performance of downstream tasks. Our triplane diffusion model can help to bridge this gap by exploiting 3D scene priors. This approach enhances the reliability and effectiveness of SSC, which is expected to improve its application in autonomous navigation systems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.591, + 0.394 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.617 + ], + "angle": 0, + "content": "In this section, we elucidate our triplane diffusion model and its extensions. Our triplane diffusion model aims to synthesize novel real-outdoor scenes by generating a triplane, a proxy representation, which effectively addresses the inherent challenges of real-outdoor scene synthesis. This triplane representation is learned by our triplane autoencoder, which abstracts the geometric and semantic intricacies of a scene into three orthogonal 2D feature planes, namely, the \\(xy\\), \\(xz\\), and \\(yz\\) planes (Sec. 3.1). Then, our diffusion model learns triplane distributions of scenes, generating novel triplanes (Sec. 3.2). We extend our triplane diffusion model toward various practical scenarios: scene inpainting, scene outpainting, and semantic scene completion refinements (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.632, + 0.885, + 0.649 + ], + "angle": 0, + "content": "3.1. Representing a Semantic Scene with Triplane" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.748 + ], + "angle": 0, + "content": "To represent a 3D scene as a triplane, our triplane autoencoder learns to compress a 3D scene into a triplane representation as shown in Fig. 2(a). The autoencoder consists of two modules: (1) an encoder \\( f_{\\theta} \\) yielding a triplane, and (2) an implicit multi-layer perceptron (MLP) decoder \\( g_{\\theta} \\) for reconstruction from the triplane." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The encoder \\(f_{\\theta}\\) takes a voxelized scene \\(\\mathbf{x} \\in \\mathbb{R}^{X \\times Y \\times Z}\\) containing \\(N\\) classes within a spatial grid of resolution \\(X \\times Y \\times Z\\). It then yields an axis-aligned triplane representation \\(\\mathbf{h} = [\\mathbf{h}^{xy}, \\mathbf{h}^{xz}, \\mathbf{h}^{yz}]\\). The triplane consists of three planes, each characterized by distinct dimensional properties: \\(\\mathbf{h}^{xy} \\in \\mathbb{R}^{C_h \\times X_h \\times Y_h}\\), \\(\\mathbf{h}^{xz} \\in \\mathbb{R}^{C_h \\times X_h \\times Z_h}\\), and \\(\\mathbf{h}^{yz} \\in \\mathbb{R}^{C_h \\times Y_h \\times Z_h}\\), where \\(C_h\\) stands for a feature dimension, and \\(X_h, Y_h\\), and \\(Z_h\\) denote spatial dimension of the triplane. During the encoding phase, a 3D feature volume is extracted by 3D convolutional layers from the scene \\(\\mathbf{x}\\), re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28339" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "sulting in the triplane via axis-wise average pooling. Given a 3D coordinate \\(\\mathbf{p} = (x,y,z)\\), the triplane is interpreted as a summation of vectors bilinearly interpolated from each plane: \\(\\mathbf{h}(\\mathbf{p}) = \\mathbf{h}^{xy}(x,y) + \\mathbf{h}^{xz}(x,z) + \\mathbf{h}^{yz}(y,z)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.287 + ], + "angle": 0, + "content": "To reconstruct the 3D scene \\(\\mathbf{x}\\), we decode the encoded triplane \\(\\mathbf{h}\\) with an implicit MLP decoder \\(g_{\\theta}\\) that predicts semantic class probabilities. The decoder takes the triplane vector \\(\\mathbf{h}(\\mathbf{p})\\) with its sinusoidal positional embedding \\(\\mathrm{PE}(\\mathbf{p})\\) [30], resulting in class probabilities \\(\\mathbf{c}(\\mathbf{p}) = g_{\\theta}(\\mathbf{h}(\\mathbf{p}),\\mathrm{PE}(\\mathbf{p}))\\in [0,1]^N\\). The positional embedding produces high-frequency features according to the coordinates \\(\\mathbf{p}\\), which helps the implicit decoder \\(g_{\\theta}\\) represent high-frequency scene contents [45]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.288, + 0.468, + 0.318 + ], + "angle": 0, + "content": "The encoder \\( f_{\\theta} \\) and the MLP decoder \\( g_{\\theta} \\) are trained with the autoencoder loss \\( \\mathcal{L}_{\\mathrm{AE}} \\) and scene label \\( \\mathbf{x}(\\mathbf{p}) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.329, + 0.468, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A E}} = \\mathbb {E} _ {\\mathbf {p} \\sim \\mathcal {P}} \\left[ \\ell_ {\\mathrm {C E}} (\\mathbf {c} (\\mathbf {p}), \\mathbf {x} (\\mathbf {p})) + \\lambda \\ell_ {\\mathrm {L Z}} (\\mathbf {c} (\\mathbf {p}), \\mathbf {x} (\\mathbf {p})) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.468, + 0.415 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a loss weight, and \\(\\mathcal{P}\\) is the set of grid coordinates of the scene. We use the weighted cross-entropy loss \\(\\ell_{\\mathrm{CE}}\\) [33] and the Lovász-softmax loss \\(\\ell_{\\mathrm{LZ}}\\) [7] to learn imbalanced semantic distributions of the scene." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.254, + 0.44 + ], + "angle": 0, + "content": "3.2. Triplane Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.468, + 0.673 + ], + "angle": 0, + "content": "Based on the triplane representation of the 3D semantic scene, our triplane diffusion model \\( D_{\\phi} \\) learns to generate a novel triplane through denoising diffusion probabilistic models [16] as shown in Fig. 2(b). This triplane generation leads to generation of 3D scene through decoding the generated triplane with the implicit MLP decoder \\( g_{\\theta} \\). Through the \\( x_0 \\)-parameterization [4], the diffusion model \\( D_{\\phi} \\) is trained to reconstruct the triplane \\( \\mathbf{h} \\) given its corrupted triplane \\( \\mathbf{h}_t \\) sampled from a diffusion process \\( q(\\mathbf{h}_t|\\mathbf{h}) = \\mathcal{N}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{h}, (1 - \\bar{\\alpha}_t)\\mathbf{I}) \\), where \\( \\mathcal{N} \\) is the Gaussian distribution, \\( \\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i \\), and \\( \\alpha_t = 1 - \\beta_t \\) with a variance schedule \\( \\beta_t \\). The diffusion process \\( q(\\mathbf{h}_t|\\mathbf{h}) \\) is derived from the Markovian chain rule with a single step's diffusion process \\( q(\\mathbf{h}_t|\\mathbf{h}_{t-1}) = \\mathcal{N}(\\sqrt{1 - \\beta_t}\\mathbf{h}_{t-1}, \\beta_t\\mathbf{I}) \\). Thus, the triplane diffusion loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.686, + 0.468, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D}} = \\mathbb {E} _ {t \\sim \\mathcal {U} (1, T)} | | \\mathbf {h} - D _ {\\phi} (\\mathbf {h} _ {t}, t) | | _ {p}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.468, + 0.756 + ], + "angle": 0, + "content": "where \\(T\\) is the number of denoising steps, and \\(p\\) represents the order of the norm. The timestep \\(t\\) is sampled from the discrete uniform distribution \\(\\mathcal{U}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.468, + 0.803 + ], + "angle": 0, + "content": "After training, the diffusion model \\(D_{\\phi}\\) generates a novel triplane \\(\\mathbf{h}_0\\) via the iterative DDPM generation process [16] starting from \\(\\mathbf{h}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.813, + 0.468, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {t - 1} \\sim \\mathcal {N} \\left(\\gamma_ {t} \\mathbf {h} _ {t} + \\delta_ {t} D _ {\\phi} \\left(\\mathbf {h} _ {t}, t\\right), \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.902 + ], + "angle": 0, + "content": "with \\(\\gamma_{t} \\coloneqq \\sqrt{\\alpha_{t}} (1 - \\bar{\\alpha}_{t-1}) / (1 - \\bar{\\alpha}_{t})\\) and \\(\\delta_{t} \\coloneqq \\sqrt{\\bar{\\alpha}_{t-1}} \\beta_{t} / (1 - \\bar{\\alpha}_{t})\\). From the generated triplane \\(\\mathbf{h}_0\\), we generate a novel 3D semantic scene \\(\\mathbf{x}_0\\) by querying coordinates \\(\\mathbf{p}\\) to the implicit decoder, i.e., \\(g_{\\theta}(\\mathbf{h}_0(\\mathbf{p}), \\mathsf{PE}(\\mathbf{p}))\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.85, + 0.107 + ], + "angle": 0, + "content": "3.3. Applications with Triplane Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.892, + 0.176 + ], + "angle": 0, + "content": "Building on the triplane diffusion process (Sec. 3.2), we propose a triplane manipulation that allows our model to facilitate a variety of practical downstream tasks with few modifications, as illustrated in Fig. 2(c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.205, + 0.892, + 0.477 + ], + "angle": 0, + "content": "Scene Inpainting. Our scene inpainting randomly edits a 3D scene, seamlessly adding, modifying, or removing objects while maintaining the consistency and realism of the scene. For instance, the inpainting includes scenarios where cars or sidewalks appear and then disappear, or vice versa, as shown in Fig. 1(d). Inspired by the RePaint sampling strategy [27], we propose a 3D-aware inpainting approach with semantic coherence. RePaint focuses on the image domain without explicitly considering the fidelity of the underlying 3D scene. To facilitate 3D-aware inpainting, we inpaint the triplane, serving as a compact proxy representation for the scene. We define a binary spatial trimask \\(\\mathbf{m} = [\\mathbf{m}^{xy},\\mathbf{m}^{xz},\\mathbf{m}^{yz}]\\) covering inpainting regions on triplane space, allowing us to control the generation process on the masked region. The trimask \\(\\mathbf{m}\\) is set to have ones for inpainting regions and zeros for others. We override the \\(t\\)-th triplane \\(\\mathbf{h}_t = [\\mathbf{h}_t^{xy},\\mathbf{h}_t^{xz},\\mathbf{h}_t^{yz}]\\) of the generation process (Eq. 3) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.493, + 0.891, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {t} \\leftarrow \\mathbf {m} \\otimes \\mathbf {h} _ {t} + (\\mathbf {1} - \\mathbf {m}) \\otimes \\mathbf {h} _ {t} ^ {\\text {k n o w n}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.525, + 0.892, + 0.584 + ], + "angle": 0, + "content": "where \\(\\otimes\\) is the element-wise product, and the known triplane \\(\\mathbf{h}_t^{\\mathrm{known}}\\) for intact regions is sampled from the diffusion process, i.e., \\(q(\\mathbf{h}_t^{\\mathrm{known}}|\\mathbf{h})\\coloneqq q(\\mathbf{h}_t|\\mathbf{h})\\), which adheres to a known Gaussian distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Scene Outpainting. Our scene outpainting extends the boundaries of the 3D scene without additional training as with inpainting. To seamlessly outpaint a scene, the regions to be extended should be conditioned on the original scene. We propose to inject this intuition into triplane representation as shown in Fig. 2(c). For covering regions to be outpainted, our diffusion model generates a novel triplane that is partially overlapped with the original triplane. We implement this with the concept of the trimask m and the known triplane \\(\\mathbf{h}_t^{\\mathrm{known}}\\) as in our scene inpainting. The trimask m covers regions to be outpainted, and the \\(\\mathbf{h}_t^{\\mathrm{known}}\\) is obtained from the intersections of triplanes between original and outpainted regions. Based on the outpainting strategy, we extend a given scene toward cardinal and intercardinal directions, facilitating the creation of an unbounded scene. While our triplane diffusion model is trained using triplanes of a fixed size, it demonstrates the capability to outpaint scenes to be several times bigger than the original scene, as illustrated in Fig. 1(c)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.086, + 0.895, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.451, + 0.893, + 0.48 + ], + "angle": 0, + "content": "Figure 3. Scene generation results using both real and synthetic outdoor datasets - SemanticKITTI [6] and CarlaSC [47]. Our results showcase the effective generation of overall structures, including roads and buildings, along with detailed objects such as cars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.473, + 0.697 + ], + "angle": 0, + "content": "Semantic Scene Completion Refinement. SSC models complete and segment a 3D scene from sensor observations such as images or point clouds. We observe that the SSC results show a geometric and semantic discrepancy compared with data distributions, as shown in Fig. 1(b). We extend our triplane diffusion model to refine predictions of SSC models toward reducing the discrepancy. To effectively condition our triplane diffusion scheme to SSC model's prediction \\(\\mathbf{x}^{\\mathrm{ssc}}\\), we utilize its triplane representation \\(\\mathbf{h}^{\\mathrm{ssc}} = f_{\\theta}(\\mathbf{x}^{\\mathrm{ssc}})\\) derived by our triplane encoder \\(f_{\\theta}\\). We extend our triplane diffusion scheme with a simple modification of the triplane \\(\\mathbf{h}_t\\) in the diffusion loss (Eq. 2) and the generation process (Eq. 3) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.708, + 0.47, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {t} = \\mathbf {h} _ {t} ^ {\\mathrm {s s c}} \\oplus \\mathbf {h} ^ {\\mathrm {s s c}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.78 + ], + "angle": 0, + "content": "where \\(\\oplus\\) is concatenation, and \\(\\mathbf{h}_t^{\\mathrm{ssc}}\\) is a \\(t\\)-th diffused triplane sampled by the DDPS diffusion process [21] with the SSC prediction's triplane \\(\\mathbf{h}^{\\mathrm{ssc}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.793, + 0.21, + 0.81 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.277, + 0.833 + ], + "angle": 0, + "content": "4.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Training Dataset. We validate our method on the SemanticKITTI [6] and CarlaSC [47] datasets. SemanticKITTI provides 3D semantic scenes of real-outdoor environments with labels for 20 semantic classes. Each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.668 + ], + "angle": 0, + "content": "scene is represented by a voxel grid of \\(256 \\times 256 \\times 32\\), covering an area of \\(51.2\\mathrm{m}\\) in front of the car, extending \\(51.2\\mathrm{m}\\) on each side, and reaching up to a height of \\(6.4\\mathrm{m}\\). The dataset retains object motion traces as a result of sensor frame integration, which is employed to establish a dense ground truth. In contrast, CarlaSC is a synthetic dataset that provides 3D semantic outdoor scenes without the trace of moving objects. The dataset contains annotated 11 semantic classes with a voxel grid of \\(128 \\times 128 \\times 8\\) and covers a distance of \\(25.6\\mathrm{m}\\) in front and behind the car, \\(25.6\\mathrm{m}\\) laterally on each side, and \\(3\\mathrm{m}\\) in height." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. Our experiments are deployed on a single NVIDIA RTX 3090 GPU with a batch size of 4 for the triplane autoencoder and 18 for the triplane diffusion model. For the triplane autoencoder, the input scene is encoded to triplane with a spatial resolution \\((X_h, Y_h, Z_h) = (128, 128, 32)\\), and the feature dimension \\(C_h\\) is 16. The loss weight \\(\\lambda\\) in Eq. 1 is set to 1.0. The order of the norm \\(p\\) in Eq. 2 is set to 1 for SSC refinements and 2 for other cases. For the diffusion model, the learning rate is initialized to 1e-4 and then decreases linearly. During the diffusion process, we use the default settings [34] with 100 time steps \\((T)\\). For our triplane inpainting and outpainting, we employ the RePaint sampling strategy [27] as a reference and perform a repaint with 5 resampling and a jump size of 20." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28341" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.19 + ], + "angle": 0, + "content": "
ModelFID ↓KID ↓IS ↑Prec ↑Rec ↑
SemanticKITTI [6]
SSD [22]112.820.122.230.010.08
SemCity (Ours)56.550.043.250.390.32
CarlaSC [47]
SSD [22]87.390.092.440.140.07
SemCity (Ours)40.630.023.510.310.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.201, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Table 1. Results of semantic scene generation. Each metric is computed between the rendered image of the generated scene and the ones of the actual scene at a resolution of \\(1440 \\times 2048\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.471, + 0.537 + ], + "angle": 0, + "content": "Evaluation Metrics. Following the scene generation works [41, 55], we evaluate the performance of semantic scene generation by examining both the diversity and fidelity of 3D semantic scenes within the rendered images. We use recall to evaluate diversity, while precision and inception score (IS) are used to evaluate fidelity. The Fréchet Inception Distance (FID) [15] and Kernel Inception Distance (KID) [8] metrics are also utilized, as they reflect the combined effect of both diversity and fidelity on scene quality [14]. In terms of semantic scene completion (SSC) refinement performance, we follow the protocols defined in SSC works [9, 29, 49, 50]. The Intersection-over-Union (IoU) metric is used to quantify scene completeness, while the mean IoU (mIoU) provides a measure for the quality of semantic segmentation. These metrics together enable a comprehensive evaluation of how well the semantic scene completion methods perform in terms of accurately filling in and labeling the scene components." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.325, + 0.56 + ], + "angle": 0, + "content": "4.2. Semantic Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 illustrates a qualitative comparison on the SemanticKITTI [6] and CarlaSC [47] datasets. SSD [22] shows impressive results on the CarlaSC dataset. However, its performance on the SemanticKITTI dataset, which is a real-world dataset, is notably constrained. This limitation primarily arises from the SSD's voxel-based representation, which struggles with the more prevalent empty spaces in real-outdoor datasets compared to synthetic ones. This issue is especially evident in the generation of buildings and roads, where SSD often fails to define boundaries accurately. Furthermore, the model's limitations extend to representing finer structures, such as trunks and leaves, as well as traffic light poles and signals. It also struggles with generating uniform shapes of vehicles, often resulting in irregular shapes. In contrast, our method demonstrates the ability to effectively synthesize detailed scenes even on the real dataset, as illustrated in Fig. 3. It performs better than SSD [22] in accurately capturing complex building shapes on the CarlaSC dataset. In addition, our method exhibits remarkable proficiency in generating the overall contours of roads and buildings, along with intricate details on the SemanticKITTI dataset. Tab. 1 provides a detailed compar" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.699, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.892, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.225, + 0.894, + 0.28 + ], + "angle": 0, + "content": "Figure 4. Higher-resolution scene generation. Building upon our implicit decoder, higher-resolution scene \\((1024\\times 1024\\times 128)\\) can be generated compared with a resolution of training dataset \\((256\\times 256\\times 32)\\)." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.294, + 0.892, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.634, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Figure 5. Scene inpainting of our method. The red boxes denote inpainting regions. (a) and (b) show our inpainting examples from reference images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.793 + ], + "angle": 0, + "content": "ative evaluation using various metrics. Our model shows significant improvements in both the fidelity and diversity of the generated scenes. Moreover, our generated result is not tied to fixed resolution by means of the implicit neural representation, as depicted in Fig. 4. For additional results, please refer to the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.802, + 0.798, + 0.818 + ], + "angle": 0, + "content": "4.3. Applications of Triplane Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Scene Inpainting. Fig. 5 presents the qualitative results of our inpainting, demonstrating its effectiveness in inpainting both small and large regions within a scene, while maintaining the coherence of 3D contexts. In detail, the second row of (a) and (b) illustrates the model's seamless removal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28342" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.101, + 0.891, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.219, + 0.321, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.219, + 0.578, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.219, + 0.889, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.893, + 0.364 + ], + "angle": 0, + "content": "Figure 6. Our scene outpainting results. The red box at the center of the figure in the first row represents the given scene for outpainting. The zoomed views of the outpainted scene are depicted in the second row. The outpainted scene is expanded from the given size of \\(256 \\times 256 \\times 32\\) to \\(1792 \\times 3328 \\times 32\\)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.376, + 0.468, + 0.505 + ], + "angle": 0, + "content": "
SSC InputMethodIoU ↑mIoU ↑
RGBMonoScene [9]37.1211.50
MonoScene + Ours50.4417.08
OccDepth [29]41.6012.84
OccDepth + Ours50.2016.79
Point CloudSSA-SC [51]58.2524.54
SSA-SC + Ours60.7125.58
SCPNet [49]50.2437.55
SCPNet + Ours59.2538.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.508, + 0.47, + 0.551 + ], + "angle": 0, + "content": "Table 2. Quantitative results of refining SSC on SemanticKITTI validation set [6]. The results are based on the weights released by the authors on GitHub." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.742 + ], + "angle": 0, + "content": "of vehicles, which harmonizes with the adjacent road. The third row demonstrates the insertion of new entities — a vehicle in (a) and a person in (b) — that are contextually congruent with the reference scene. The fourth row in (a) exemplifies the model's dual functionality in both modifying and adding vehicles within the scene. The fifth row in both columns underscores the model's proficiency in modifying scenes. Here, the model alters existing scene components, showcasing its ability to transform the overall ambiance of the scene. These results show our model's adeptness not just in object-level inpainting but also in scene-level inpainting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Scene Outpainting. Fig. 6 illustrates a generated outpainting city-level scene, extending a \\(256 \\times 256 \\times 32\\) scene to a substantial \\(1792 \\times 3328 \\times 32\\) landscape. Although our model was not designed to generate cityscapes, it demonstrates the capability of maintaining coherence over a large area. The roads are connected in a meaningful and varied way, and various objects such as buildings, cars, and people have been created around them. Please check the details of" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.377, + 0.892, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.603, + 0.891, + 0.632 + ], + "angle": 0, + "content": "Figure 7. Semantic scene completion refinement. The SSC metrics are reported in the parentheses as (IoU, mIoU)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.651, + 0.628, + 0.665 + ], + "angle": 0, + "content": "the scene in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Semantic Scene Completion Refinement. In Fig. 7, there is a notable semantic and geometric discrepancy between scenes predicted by existing semantic scene completion (SSC) methods and their real scene counterparts. Our model helps to bridge this gap by employing a 3D scene prior, which is effectively modeled through our diffusion model. SCPNet [49] also attempts to learn geometric priors from a teacher model trained with merged sequential frames. Yet, its completion network is inclined toward conservative estimations, which frequently result in partially filled spaces. In the case of SSA-SC [51], which utilizes bird's-eye view features, it is sometimes inappropriate due to the label's inherent limitations of the bird's-eye perspective in capturing certain semantic details. This discrepancy" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28343" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.094, + 0.27, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.185, + 0.184, + 0.197 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.465, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.185, + 0.378, + 0.197 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.202, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Figure 8. Semantic scene to RGB. (a) Semantic maps are rendered into a driving view from our generated scenes. (b) We generate images through ControlNet [56], an easily accessible imaged-to-image model, with generated semantic maps without shadows." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.27, + 0.468, + 0.345 + ], + "angle": 0, + "content": "
MethodFID ↓KID ↓IS ↑Prec ↑Rec ↑
Ours56.550.043.250.390.32
w/o PE59.140.043.210.300.31
xy-plane (w/o triplane)75.560.063.080.280.27
3D-volume (w/o triplane)136.500.151.960.140.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.356, + 0.468, + 0.384 + ], + "angle": 0, + "content": "Table 3. Ablation studies on scene generation. We ablate our model variants on SemanticKITTI [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.468, + 0.493 + ], + "angle": 0, + "content": "issue is even more evident in RGB-based methods, such as MonoScene [9] and OccDepth [29], which often exhibit diminished sharpness when image features are projected into 3D space. While SSC models show variations from real-world data distributions, our model shows potential in aligning these more closely with reality, as shown in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.569 + ], + "angle": 0, + "content": "As indicated in Tab. 2, our SSC refinement process appears to offer improvements to all state-of-the-art SSC models. These preliminary results suggest our model's effectiveness in providing not just more accurate semantic segmentation but also a more complete scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.468, + 0.707 + ], + "angle": 0, + "content": "Semantic Scene to RGB Image. As illustrated in Fig. 8, we conducted an image-to-image generation experiment. The semantic maps are rendered into a driving view without shadows; then, they are utilized as inputs for ControlNet [56]. The generated RGB images are geometrically and semantically plausible but display a synthetic quality since the pretrained ControlNet was not trained on actual autonomous driving datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.182, + 0.731 + ], + "angle": 0, + "content": "4.4. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.468, + 0.785 + ], + "angle": 0, + "content": "As shown in Tab. 3, we conducted ablation studies on our triplane diffusion model for scene generation, focusing on two key design elements:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.468, + 0.863 + ], + "angle": 0, + "content": "Positional Embedding. Our variant excluded positional embedding PE, which generates high-frequency features critical for detailed scene reconstruction. Its absence resulted in lower performance across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Triplane Representation. We evaluated the effectiveness of the triplane representation for real-outdoor scene gen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "eration. The triplane and \\( xy \\)-plane enhance the generation quality compared with 3D features, while the \\( xy \\)-plane shows lower performance than triplanes. We suspect that the excessive factorization limits the representation capability of the \\( xy \\)-plane compared to the triplane." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.175, + 0.618, + 0.19 + ], + "angle": 0, + "content": "4.5. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.456 + ], + "angle": 0, + "content": "While our model demonstrates significant progress in the generation of 3D real-outdoor scenes, it inherently reflects the characteristics of its training data, as detailed in Sec. 4.1. This reliance introduces several limitations. One notable challenge is the model's difficulty in accurately depicting areas occluded from the sensor's viewpoint, such as the rear sides of buildings, often leading to their incomplete representation in our generated scenes. Moreover, since the dataset is captured from a driving view, there is an inherent shortfall in capturing the full height of buildings. This brings a partial representation of the vertical structure of buildings and other tall elements in the scenes. Another issue is the model tends to produce traces of moving objects stemming from dataset pre-processing that merges sequential frames. For future work, incorporating prior knowledge of the city into the model could yield better results, such as addressing occluded areas or building heights." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.618, + 0.483 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We have proposed a diffusion framework called SemCity for real-world outdoor scene generation. The seminal idea is to generate a scene by factorizing real-outdoor scenes into triplane representations. Our triplane representation outperforms traditional voxel-based approaches, producing scenes that are not only visually more appealing but also rich in semantic detail, effectively capturing the complexity of various objects within the scene. Ours is not constrained by fixed resolutions thanks to the incorporation of an implicit neural representation. We have further expanded the capabilities of our triplane diffusion model to several practical applications, including scene inpainting, scene outpainting, and semantic scene completion refinement. Specifically, by manipulating triplanes during the diffusion process, we achieve seamless inpainting and outpainting at both the object and scene levels. Ours is used to more closely align the scenes predicted by existing semantic scene completion methods with the actual data distribution using our learned 3D prior. We believe that our work provides a road map of real-outdoor scene generation to research communities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by IITP (Institute of Information & Communications Technology Planning & Evaluation) and ITRC (Information Technology Research Center), funded by Korea government (MSIT) (RS-2023-00237965(2024) and IITP-2024-2020-0-01460). Prof. Sung-Eui Yoon is a corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28344" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Ali Abbasi, Sinan Kalkan, and Yusuf Sahillioğlu. Deep 3d semantic scene extrapolation. The Visual Computer, 35:271-279, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12608-12618, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[3] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12608-12618, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[4] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[5] Luca Bartolomei, Lucas Teixeira, and Margarita Chli. Perception-aware path planning for uavs using semantic segmentation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5808-5815, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[7] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.62, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[9] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[11] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[12] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.895, + 0.135 + ], + "angle": 0, + "content": "image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[13] Xu Cui, Chenggang Lu, and Jinxiang Wang. 3d semantic map construction using improved orb-slam2 for mobile robot in edge computing environment. IEEE Access, 8:67179-67191, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[18] Changho Jo, Woobin Im, and Sung-Eui Yoon. In-n-out: Towards good initialization for inpainting and outpainting. arXiv preprint arXiv:2106.13953, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[21] Zeqiang Lai, Yuchen Duan, Jifeng Dai, Ziheng Li, Ying Fu, Hongsheng Li, Yu Qiao, and Wenhai Wang. Denoising diffusion semantic segmentation with mask prior modeling. arXiv preprint arXiv:2306.01721, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[22] Jumin Lee, Woobin Im, Sebin Lee, and Sung-Eui Yoon. Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[23] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. arXiv preprint arXiv:2212.05993, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.895, + 0.76 + ], + "angle": 0, + "content": "[24] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.895, + 0.844 + ], + "angle": 0, + "content": "[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[26] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28345" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[29] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.328, + 0.471, + 0.397 + ], + "angle": 0, + "content": "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.398, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[33] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[37] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[39] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene com" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "pletion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.203, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[41] Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner. Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.273, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[42] Chaoqun Wang, Wenzheng Chi, Yuxiang Sun, and Max Q-H Meng. Autonomous robotic exploration by incremental road map construction. IEEE Transactions on Automation Science and Engineering, 16(4):1720-1731, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.328, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[43] Chaoqun Wang, Delong Zhu, Teng Li, Max Q-H Meng, and Clarence W De Silva. Efficient autonomous robotic exploration with semantic road map in indoor environments. IEEE Robotics and Automation Letters, 4(3):2989-2996, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[44] Xinying Wang, Dikai Xu, and Fangming Gu. 3d model inpainting based on 3d deep convolutional generative adversarial network. IEEE Access, 8:170355-170363, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[45] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Pet-neus: Positional encoding tri-planes for neural surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12598–12607, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[46] Zhen Wang, Shijie Zhou, Jeong Joon Park, Despoina Paschalidou, Suya You, Gordon Wetzstein, Leonidas Guibas, and Achuta Kadambi. Alto: Alternating latent topologies for implicit 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 259-270, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[47] Joey Wilson, Jingyu Song, Yuewei Fu, Arthur Zhang, Andrew Capodieci, Paramsothy Jayakumar, Kira Barton, and Maani Ghaffari. *Motionsc: Data set and network for real-time semantic mapping in dynamic environments*. arXiv preprint arXiv:2203.07060, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[48] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.678, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[49] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[50] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Xuemeng Yang, Hao Zou, Xin Kong, Tianxin Huang, Yong Liu, Wanlong Li, Feng Wen, and Hongbo Zhang. Semantic segmentation-assisted scene completion for lidar point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3555-3562. IEEE, 2021. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28346" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[52] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4206-4216, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[55] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.472, + 0.431 + ], + "angle": 0, + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.472, + 0.5 + ], + "angle": 0, + "content": "[57] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10188-10198, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.502, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[58] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28347" + } + ] +] \ No newline at end of file diff --git a/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_origin.pdf b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..65ddbc63ec50c76c5531f21ec2732979e066b43e --- /dev/null +++ b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/36ac8a85-7845-4237-8f1d-f2203275c4a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef33409366aad96f17205e34be2f84488db6d84ddbbfb044f1131066550f1694 +size 9048717 diff --git a/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/full.md b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..07e8068202a704282465c5f1e7d25317d9cc0562 --- /dev/null +++ b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/full.md @@ -0,0 +1,288 @@ +# SemCity: Semantic Scene Generation with Triplane Diffusion + +Jumin Lee $^{1*}$ Sebin Lee $^{1*}$ Changho Jo $^{2}$ Woobin Im $^{1}$ Juhyeong Seon $^{1}$ Sung-Eui Yoon $^{1}$ $^{1}$ KAIST $^{2}$ Neosapience, Inc. + +![](images/41e3c143eef09caf41f959ed3acc2728a1a2eda481960466c73f6825302a925c.jpg) +(a) Scene generation + +![](images/3564e47203d7a727abfbd1630678f15064257f8ea24a305c1b3853f8db2f7e88.jpg) +(c) Scene outpainting + +![](images/db600fd2c966d0be4c3d13c0eb9c5a682c3eebb88015381d387babaae39f3f73.jpg) + +![](images/c599459d462a207f589abd6de1133221a987f1391f5b62a3e924909b8cd5add5.jpg) +(b) Semantic scene completion refinement + +![](images/5adfb0a1a22d8c604a474870f7dfa47b5b861f77d72c48444a7cf65bc9f34b81.jpg) +(d) Scene inpainting +Figure 1. We introduce a diffusion framework, SemCity, designed for generating semantic scenes in real-world outdoor environments as shown in (a). We extend our diffusion model to various practical tasks: semantic scene completion refinement, scene outpainting, and scene inpainting. For instance, the comprehensive scenario is displayed in $(\mathrm{b})\rightarrow (\mathrm{c})\rightarrow (\mathrm{d})$ : the refined scene (SSC refinement) (b) is outpainted to a broader scene (c); then, an object (in this case, a car) is seamlessly integrated into the scene via our inpainting process (d). + +# Abstract + +We present "SemCity," a 3D diffusion model for semantic scene generation in real-world outdoor environments. Most 3D diffusion models focus on generating a single object, synthetic indoor scenes, or synthetic outdoor scenes, while the generation of real-world outdoor scenes is rarely addressed. In this paper, we concentrate on generating a real-outdoor scene through learning a diffusion model on a real-world outdoor dataset. In contrast to synthetic data, real-outdoor datasets often contain more empty spaces due to sensor limitations, causing challenges in learning realoutdoor distributions. To address this issue, we exploit a triplane representation as a proxy form of scene distributions to be learned by our diffusion model. Furthermore, we propose a triplane manipulation that integrates seamlessly with our triplane diffusion model. The manipulation improves our diffusion model's applicability in a variety of downstream tasks related to outdoor scene generation such as scene inpainting, scene outpainting, and semantic + +scene completion refinements. In experimental results, we demonstrate that our triplane diffusion model shows meaningful generation results compared with existing work in a real-outdoor dataset, SemanticKITTI. We also show our triplane manipulation facilitates seamlessly adding, removing, or modifying objects within a scene. Further, it also enables the expansion of scenes toward a city-level scale. Finally, we evaluate our method on semantic scene completion refinements where our diffusion model enhances predictions of semantic scene completion networks by learning scene distribution. Our code is available at https://github.com/zoomin-lee/SemCity. + +# 1. Introduction + +Diffusion models [16] have emerged as a promising generation tool owing to its state-of-the-art generation results in image domain [35, 36]. This advance has led to active exploration in extending diffusion models to 3D data generation across both academic and industrial groups. In the 3D domain, diffusion models have shown remarkable + +capabilities in generating diverse 3D forms (e.g., voxels, meshes) [24, 26]. While those 3D diffusion models primarily aim to craft a single object, generating scenes consisting of multiple objects remains a relatively unexplored area in the 3D diffusion domain. + +Scene generative diffusion models focus on crafting both geometrically and semantically coherent environments. Compared with a single object generation, generating a scene with multiple objects requires an understanding of more complex geometric and semantic structure due to a larger spatial extent [41]. There are primarily two streams of scene generative diffusion models, each tailored to either indoor or outdoor settings. In particular, the outdoor environments have inherent challenges caused by a broader landscape than indoor ones. + +We propose to leverage a triplane representation [10] for broader outdoor scenes, a method of factorizing 3D data onto three orthogonal 2D planes, as utilized in 3D object reconstruction and NeRF models [2, 31, 46]. We excavate its advantages in addressing the data sparsity problem typically found in outdoor datasets due to the sensor limitations (e.g., occlusions, range constraints) in capturing outdoor scenes. Triplane representation helps to reduce the inclusion of unnecessary empty information through the factorization of 3D data to 2D planes [10]. This efficiency in capturing relevant spatial detail makes it an effective tool for representing the many objects typically found in outdoor environments. + +In this paper, we design our diffusion framework based on triplane representations. Our triplane autoencoder learns to compress a voxelized scene into a triplane representation by reconstructing semantic labels of the scene. Following this, the triplane diffusion model is trained and used to generate new scenes, as shown in Fig. 1(a), by creating novel triplanes based on the efficient representation. Further, we propose a triplane manipulation method, which extends our triplane diffusion model toward several practical tasks (i.e., scene inpainting, scene outpainting, and semantic scene completion refinements) as shown in Fig. 1(b-d). Our method can seamlessly add, remove, and modify objects in real-outdoor scenes while maintaining the semantic coherence of environments. + +Our contributions are summarized as follows: + +- We disclose the applicability of the triplane representation through generating semantic scenes for real-outdoor environments and extend its views in practical downstream tasks: scene inpainting, scene outpainting, and semantic scene completion refinement. +- We propose to manipulate triplane features during our diffusion process, facilitating seamlessly extending our method toward the downstream tasks. +- We demonstrate that the proposed method significantly enhances the quality of generated scenes in real-world outdoor environments. + +# 2. Related Work + +Diffusion Models. Diffusion models [16] learn data distributions via iterative denoising processes based on score functions [40]. Its generated results have shown remarkably realistic appearances with high fidelity and diversity in a variety of 2D image synthesis such as outpainting [35, 57], inpainting [27, 35] and text-to-image generation [32, 36]. Built upon these achievements, diffusion models have also been extended into the 3D domain, generating impressive results in various 3D shapes, including voxel grids [24, 58], point clouds [28, 53, 54], meshes [26], and implicit functions [20, 37, 38, 48]. While these models can craft a single 3D object, our model focuses on generating a 3D scene composed of multiple objects using a categorical voxel data structure, which is a relatively under-explored area in the 3D diffusion domain. + +Diffusion Models for Scene Generation. In contrast to a single object generation, scene generation involves an understanding of the larger 3D space, causing more semantic and geometric complexities [41]. Diffusion models for scene generation have been studied in both indoor and outdoor environments. In indoor settings, diffusion models aim to learn distributions of relations among objects by representing them as scene graphs [19]. The scene graphs contain object attributes (e.g., location, orientation, and size), capturing the intricate inter-object relationships within bounded spaces [41, 55]. For outdoor scenes, the challenges are distinct, frequently including a lot of empty areas (e.g., sky, open areas) resulting from the broader landscapes. Traditional approach [22] has relied on discrete diffusion methods [17] on voxel space, necessitating a detailed representation of every air volume. + +In this paper, we demonstrate that triplane diffusion is highly effective for generating real-outdoor scenes. By abstracting 3D spaces into three orthogonal 2D planes, triplane representation [10] effectively captures the vastness of outdoor environments, predominantly composed of air. Beyond its data efficiency, the triplane excels in focusing on other significant objects (e.g., vehicles, buildings) by allocating lesser attention to less informative elements like air. Our approach stands in stark contrast to prior work [22], which were constrained to synthetic datasets with considerably less empty space. In real datasets, inherent sensor limitations, such as a limited field-of-view, a limited ranges, and the inability to capture occluded areas like the rears of buildings, lead to a prevalence of empty space. Furthermore, we emphasize the versatility of our framework by demonstrating its extension to various downstream tasks, including scene inpainting, outpainting, and semantic scene completion refinement. + +![](images/c0cb1b2c89dd38346bd4de7735e64d2632e7f6dcd0eeeb43c77c5675768555b9.jpg) + +![](images/826c63a724b7e726ff79b1c07c5ce0f74cd67193c227f600214d1796f3c40aae.jpg) +Figure 2. Overview of ours. (a) A 3D semantic map $\mathbf{x}$ is encoded by a triplane encoder $f_{\theta}$ and factorized to a triplane $\mathbf{h}$ . The triplane coupled with a positional encoding $\mathbb{P}\mathbb{E}(\mathbf{p})$ is decoded by an implicit decoder $g_{\theta}$ , resulting in class probabilities for each coordinate $\mathbf{p}$ . (b) Our triplane diffusion model $D_{\phi}$ learns to generate a novel triplane for semantic scene generation via denoising diffusion process. (c) We further extend our triplane diffusion beyond a simple scene generation toward various practical scenarios by manipulating triplanes in (b). + +![](images/1f46fdb4aad04ff6e1b856abdf10c4d852743922fef0451ad2c6e129240378c4.jpg) + +3D Inpainting and Outpainting. In 3D inpainting, the primary objective is to fill in missing portions or modify existing elements of 3D data while maintaining geometric consistency. Most existing works concentrate on single-object inpainting [3, 23, 44]; for instance, they seamlessly transit a 3D chair's leg count from three to four. Contrary to inpainting, 3D outpainting is to extrapolate a given scene over an unobserved space. The existing work [1] focuses on scene outpainting within bounded indoor environments such as rooms. While in 2D images, inpainting and outpainting are not restricted to a single object [12, 18, 52]. Likewise, in 3D space, we focus on scene-level inpainting, which seamlessly adds, removes, or modifies objects in a scene. Further, our scene-level outpainting is not constrained to the bounded scenes; we extrapolate the outdoor scene from sensor range (e.g., LiDAR) to city-scale. + +Semantic Scene Completion. Semantic scene completion (SSC) [39] is pivotal for 3D scene understanding, where it jointly infers completion and semantic segmentation of the 3D scene from sensor observations such as RGB images [9, 25, 29] or point clouds [49-51]. In addition, SSC plays a crucial role in supporting comprehensive autonomous navigation systems, notably in essential downstream tasks like path planning [5, 11, 43] and map construction [13, 42]. Despite significant progress in the field, a persistent challenge is the semantic and geometric discrepancies between the SSC-estimated scenes and their real counterparts, as illustrated in Fig. 1(c). These discrepancies can undermine the performance of downstream tasks. Our triplane diffusion model can help to bridge this gap by exploiting 3D scene priors. This approach enhances the reliability and effectiveness of SSC, which is expected to improve its application in autonomous navigation systems. + +# 3. Method + +In this section, we elucidate our triplane diffusion model and its extensions. Our triplane diffusion model aims to synthesize novel real-outdoor scenes by generating a triplane, a proxy representation, which effectively addresses the inherent challenges of real-outdoor scene synthesis. This triplane representation is learned by our triplane autoencoder, which abstracts the geometric and semantic intricacies of a scene into three orthogonal 2D feature planes, namely, the $xy$ , $xz$ , and $yz$ planes (Sec. 3.1). Then, our diffusion model learns triplane distributions of scenes, generating novel triplanes (Sec. 3.2). We extend our triplane diffusion model toward various practical scenarios: scene inpainting, scene outpainting, and semantic scene completion refinements (Sec. 3.3). + +# 3.1. Representing a Semantic Scene with Triplane + +To represent a 3D scene as a triplane, our triplane autoencoder learns to compress a 3D scene into a triplane representation as shown in Fig. 2(a). The autoencoder consists of two modules: (1) an encoder $f_{\theta}$ yielding a triplane, and (2) an implicit multi-layer perceptron (MLP) decoder $g_{\theta}$ for reconstruction from the triplane. + +The encoder $f_{\theta}$ takes a voxelized scene $\mathbf{x} \in \mathbb{R}^{X \times Y \times Z}$ containing $N$ classes within a spatial grid of resolution $X \times Y \times Z$ . It then yields an axis-aligned triplane representation $\mathbf{h} = [\mathbf{h}^{xy}, \mathbf{h}^{xz}, \mathbf{h}^{yz}]$ . The triplane consists of three planes, each characterized by distinct dimensional properties: $\mathbf{h}^{xy} \in \mathbb{R}^{C_h \times X_h \times Y_h}$ , $\mathbf{h}^{xz} \in \mathbb{R}^{C_h \times X_h \times Z_h}$ , and $\mathbf{h}^{yz} \in \mathbb{R}^{C_h \times Y_h \times Z_h}$ , where $C_h$ stands for a feature dimension, and $X_h, Y_h$ , and $Z_h$ denote spatial dimension of the triplane. During the encoding phase, a 3D feature volume is extracted by 3D convolutional layers from the scene $\mathbf{x}$ , re + +sulting in the triplane via axis-wise average pooling. Given a 3D coordinate $\mathbf{p} = (x,y,z)$ , the triplane is interpreted as a summation of vectors bilinearly interpolated from each plane: $\mathbf{h}(\mathbf{p}) = \mathbf{h}^{xy}(x,y) + \mathbf{h}^{xz}(x,z) + \mathbf{h}^{yz}(y,z)$ . + +To reconstruct the 3D scene $\mathbf{x}$ , we decode the encoded triplane $\mathbf{h}$ with an implicit MLP decoder $g_{\theta}$ that predicts semantic class probabilities. The decoder takes the triplane vector $\mathbf{h}(\mathbf{p})$ with its sinusoidal positional embedding $\mathrm{PE}(\mathbf{p})$ [30], resulting in class probabilities $\mathbf{c}(\mathbf{p}) = g_{\theta}(\mathbf{h}(\mathbf{p}),\mathrm{PE}(\mathbf{p}))\in [0,1]^N$ . The positional embedding produces high-frequency features according to the coordinates $\mathbf{p}$ , which helps the implicit decoder $g_{\theta}$ represent high-frequency scene contents [45]. + +The encoder $f_{\theta}$ and the MLP decoder $g_{\theta}$ are trained with the autoencoder loss $\mathcal{L}_{\mathrm{AE}}$ and scene label $\mathbf{x}(\mathbf{p})$ as: + +$$ +\mathcal {L} _ {\mathrm {A E}} = \mathbb {E} _ {\mathbf {p} \sim \mathcal {P}} \left[ \ell_ {\mathrm {C E}} (\mathbf {c} (\mathbf {p}), \mathbf {x} (\mathbf {p})) + \lambda \ell_ {\mathrm {L Z}} (\mathbf {c} (\mathbf {p}), \mathbf {x} (\mathbf {p})) \right], \tag {1} +$$ + +where $\lambda$ is a loss weight, and $\mathcal{P}$ is the set of grid coordinates of the scene. We use the weighted cross-entropy loss $\ell_{\mathrm{CE}}$ [33] and the Lovász-softmax loss $\ell_{\mathrm{LZ}}$ [7] to learn imbalanced semantic distributions of the scene. + +# 3.2. Triplane Diffusion + +Based on the triplane representation of the 3D semantic scene, our triplane diffusion model $D_{\phi}$ learns to generate a novel triplane through denoising diffusion probabilistic models [16] as shown in Fig. 2(b). This triplane generation leads to generation of 3D scene through decoding the generated triplane with the implicit MLP decoder $g_{\theta}$ . Through the $x_0$ -parameterization [4], the diffusion model $D_{\phi}$ is trained to reconstruct the triplane $\mathbf{h}$ given its corrupted triplane $\mathbf{h}_t$ sampled from a diffusion process $q(\mathbf{h}_t|\mathbf{h}) = \mathcal{N}(\sqrt{\bar{\alpha}_t}\mathbf{h}, (1 - \bar{\alpha}_t)\mathbf{I})$ , where $\mathcal{N}$ is the Gaussian distribution, $\bar{\alpha}_t = \prod_{i=1}^t \alpha_i$ , and $\alpha_t = 1 - \beta_t$ with a variance schedule $\beta_t$ . The diffusion process $q(\mathbf{h}_t|\mathbf{h})$ is derived from the Markovian chain rule with a single step's diffusion process $q(\mathbf{h}_t|\mathbf{h}_{t-1}) = \mathcal{N}(\sqrt{1 - \beta_t}\mathbf{h}_{t-1}, \beta_t\mathbf{I})$ . Thus, the triplane diffusion loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {D}} = \mathbb {E} _ {t \sim \mathcal {U} (1, T)} | | \mathbf {h} - D _ {\phi} (\mathbf {h} _ {t}, t) | | _ {p}, \tag {2} +$$ + +where $T$ is the number of denoising steps, and $p$ represents the order of the norm. The timestep $t$ is sampled from the discrete uniform distribution $\mathcal{U}$ . + +After training, the diffusion model $D_{\phi}$ generates a novel triplane $\mathbf{h}_0$ via the iterative DDPM generation process [16] starting from $\mathbf{h}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ : + +$$ +\mathbf {h} _ {t - 1} \sim \mathcal {N} \left(\gamma_ {t} \mathbf {h} _ {t} + \delta_ {t} D _ {\phi} \left(\mathbf {h} _ {t}, t\right), \beta_ {t} ^ {2} \mathbf {I}\right), \tag {3} +$$ + +with $\gamma_{t} \coloneqq \sqrt{\alpha_{t}} (1 - \bar{\alpha}_{t-1}) / (1 - \bar{\alpha}_{t})$ and $\delta_{t} \coloneqq \sqrt{\bar{\alpha}_{t-1}} \beta_{t} / (1 - \bar{\alpha}_{t})$ . From the generated triplane $\mathbf{h}_0$ , we generate a novel 3D semantic scene $\mathbf{x}_0$ by querying coordinates $\mathbf{p}$ to the implicit decoder, i.e., $g_{\theta}(\mathbf{h}_0(\mathbf{p}), \mathsf{PE}(\mathbf{p}))$ . + +# 3.3. Applications with Triplane Manipulation + +Building on the triplane diffusion process (Sec. 3.2), we propose a triplane manipulation that allows our model to facilitate a variety of practical downstream tasks with few modifications, as illustrated in Fig. 2(c). + +Scene Inpainting. Our scene inpainting randomly edits a 3D scene, seamlessly adding, modifying, or removing objects while maintaining the consistency and realism of the scene. For instance, the inpainting includes scenarios where cars or sidewalks appear and then disappear, or vice versa, as shown in Fig. 1(d). Inspired by the RePaint sampling strategy [27], we propose a 3D-aware inpainting approach with semantic coherence. RePaint focuses on the image domain without explicitly considering the fidelity of the underlying 3D scene. To facilitate 3D-aware inpainting, we inpaint the triplane, serving as a compact proxy representation for the scene. We define a binary spatial trimask $\mathbf{m} = [\mathbf{m}^{xy},\mathbf{m}^{xz},\mathbf{m}^{yz}]$ covering inpainting regions on triplane space, allowing us to control the generation process on the masked region. The trimask $\mathbf{m}$ is set to have ones for inpainting regions and zeros for others. We override the $t$ -th triplane $\mathbf{h}_t = [\mathbf{h}_t^{xy},\mathbf{h}_t^{xz},\mathbf{h}_t^{yz}]$ of the generation process (Eq. 3) as follows: + +$$ +\mathbf {h} _ {t} \leftarrow \mathbf {m} \otimes \mathbf {h} _ {t} + (\mathbf {1} - \mathbf {m}) \otimes \mathbf {h} _ {t} ^ {\text {k n o w n}}, \tag {4} +$$ + +where $\otimes$ is the element-wise product, and the known triplane $\mathbf{h}_t^{\mathrm{known}}$ for intact regions is sampled from the diffusion process, i.e., $q(\mathbf{h}_t^{\mathrm{known}}|\mathbf{h})\coloneqq q(\mathbf{h}_t|\mathbf{h})$ , which adheres to a known Gaussian distribution. + +Scene Outpainting. Our scene outpainting extends the boundaries of the 3D scene without additional training as with inpainting. To seamlessly outpaint a scene, the regions to be extended should be conditioned on the original scene. We propose to inject this intuition into triplane representation as shown in Fig. 2(c). For covering regions to be outpainted, our diffusion model generates a novel triplane that is partially overlapped with the original triplane. We implement this with the concept of the trimask m and the known triplane $\mathbf{h}_t^{\mathrm{known}}$ as in our scene inpainting. The trimask m covers regions to be outpainted, and the $\mathbf{h}_t^{\mathrm{known}}$ is obtained from the intersections of triplanes between original and outpainted regions. Based on the outpainting strategy, we extend a given scene toward cardinal and intercardinal directions, facilitating the creation of an unbounded scene. While our triplane diffusion model is trained using triplanes of a fixed size, it demonstrates the capability to outpaint scenes to be several times bigger than the original scene, as illustrated in Fig. 1(c). + +![](images/a5dc446c9fe8e2b8234df33be7aff550d99df5adda0b47342e66fffe12ccf457.jpg) +Figure 3. Scene generation results using both real and synthetic outdoor datasets - SemanticKITTI [6] and CarlaSC [47]. Our results showcase the effective generation of overall structures, including roads and buildings, along with detailed objects such as cars. + +Semantic Scene Completion Refinement. SSC models complete and segment a 3D scene from sensor observations such as images or point clouds. We observe that the SSC results show a geometric and semantic discrepancy compared with data distributions, as shown in Fig. 1(b). We extend our triplane diffusion model to refine predictions of SSC models toward reducing the discrepancy. To effectively condition our triplane diffusion scheme to SSC model's prediction $\mathbf{x}^{\mathrm{ssc}}$ , we utilize its triplane representation $\mathbf{h}^{\mathrm{ssc}} = f_{\theta}(\mathbf{x}^{\mathrm{ssc}})$ derived by our triplane encoder $f_{\theta}$ . We extend our triplane diffusion scheme with a simple modification of the triplane $\mathbf{h}_t$ in the diffusion loss (Eq. 2) and the generation process (Eq. 3) as follows: + +$$ +\mathbf {h} _ {t} = \mathbf {h} _ {t} ^ {\mathrm {s s c}} \oplus \mathbf {h} ^ {\mathrm {s s c}}, \tag {5} +$$ + +where $\oplus$ is concatenation, and $\mathbf{h}_t^{\mathrm{ssc}}$ is a $t$ -th diffused triplane sampled by the DDPS diffusion process [21] with the SSC prediction's triplane $\mathbf{h}^{\mathrm{ssc}}$ . + +# 4. Experiments + +# 4.1. Experimental Details + +Training Dataset. We validate our method on the SemanticKITTI [6] and CarlaSC [47] datasets. SemanticKITTI provides 3D semantic scenes of real-outdoor environments with labels for 20 semantic classes. Each + +scene is represented by a voxel grid of $256 \times 256 \times 32$ , covering an area of $51.2\mathrm{m}$ in front of the car, extending $51.2\mathrm{m}$ on each side, and reaching up to a height of $6.4\mathrm{m}$ . The dataset retains object motion traces as a result of sensor frame integration, which is employed to establish a dense ground truth. In contrast, CarlaSC is a synthetic dataset that provides 3D semantic outdoor scenes without the trace of moving objects. The dataset contains annotated 11 semantic classes with a voxel grid of $128 \times 128 \times 8$ and covers a distance of $25.6\mathrm{m}$ in front and behind the car, $25.6\mathrm{m}$ laterally on each side, and $3\mathrm{m}$ in height. + +Implementation Details. Our experiments are deployed on a single NVIDIA RTX 3090 GPU with a batch size of 4 for the triplane autoencoder and 18 for the triplane diffusion model. For the triplane autoencoder, the input scene is encoded to triplane with a spatial resolution $(X_h, Y_h, Z_h) = (128, 128, 32)$ , and the feature dimension $C_h$ is 16. The loss weight $\lambda$ in Eq. 1 is set to 1.0. The order of the norm $p$ in Eq. 2 is set to 1 for SSC refinements and 2 for other cases. For the diffusion model, the learning rate is initialized to 1e-4 and then decreases linearly. During the diffusion process, we use the default settings [34] with 100 time steps $(T)$ . For our triplane inpainting and outpainting, we employ the RePaint sampling strategy [27] as a reference and perform a repaint with 5 resampling and a jump size of 20. + +
ModelFID ↓KID ↓IS ↑Prec ↑Rec ↑
SemanticKITTI [6]
SSD [22]112.820.122.230.010.08
SemCity (Ours)56.550.043.250.390.32
CarlaSC [47]
SSD [22]87.390.092.440.140.07
SemCity (Ours)40.630.023.510.310.09
+ +Table 1. Results of semantic scene generation. Each metric is computed between the rendered image of the generated scene and the ones of the actual scene at a resolution of $1440 \times 2048$ . + +Evaluation Metrics. Following the scene generation works [41, 55], we evaluate the performance of semantic scene generation by examining both the diversity and fidelity of 3D semantic scenes within the rendered images. We use recall to evaluate diversity, while precision and inception score (IS) are used to evaluate fidelity. The Fréchet Inception Distance (FID) [15] and Kernel Inception Distance (KID) [8] metrics are also utilized, as they reflect the combined effect of both diversity and fidelity on scene quality [14]. In terms of semantic scene completion (SSC) refinement performance, we follow the protocols defined in SSC works [9, 29, 49, 50]. The Intersection-over-Union (IoU) metric is used to quantify scene completeness, while the mean IoU (mIoU) provides a measure for the quality of semantic segmentation. These metrics together enable a comprehensive evaluation of how well the semantic scene completion methods perform in terms of accurately filling in and labeling the scene components. + +# 4.2. Semantic Scene Generation + +Fig. 3 illustrates a qualitative comparison on the SemanticKITTI [6] and CarlaSC [47] datasets. SSD [22] shows impressive results on the CarlaSC dataset. However, its performance on the SemanticKITTI dataset, which is a real-world dataset, is notably constrained. This limitation primarily arises from the SSD's voxel-based representation, which struggles with the more prevalent empty spaces in real-outdoor datasets compared to synthetic ones. This issue is especially evident in the generation of buildings and roads, where SSD often fails to define boundaries accurately. Furthermore, the model's limitations extend to representing finer structures, such as trunks and leaves, as well as traffic light poles and signals. It also struggles with generating uniform shapes of vehicles, often resulting in irregular shapes. In contrast, our method demonstrates the ability to effectively synthesize detailed scenes even on the real dataset, as illustrated in Fig. 3. It performs better than SSD [22] in accurately capturing complex building shapes on the CarlaSC dataset. In addition, our method exhibits remarkable proficiency in generating the overall contours of roads and buildings, along with intricate details on the SemanticKITTI dataset. Tab. 1 provides a detailed compar + +![](images/f08e41b86a070b9dd13c86786f4d738f88654973ddfad42dc94cd245d130cf51.jpg) +Figure 4. Higher-resolution scene generation. Building upon our implicit decoder, higher-resolution scene $(1024\times 1024\times 128)$ can be generated compared with a resolution of training dataset $(256\times 256\times 32)$ . + +![](images/20d615566abf8ce92c8dc0b6b718f12c308db3b0fa709aecf4dea9df4d99628c.jpg) + +![](images/31e88e24c962750778056f56b2a984e75b8cc69ec42fda94cd144eabbbc23ef0.jpg) +Figure 5. Scene inpainting of our method. The red boxes denote inpainting regions. (a) and (b) show our inpainting examples from reference images. + +ative evaluation using various metrics. Our model shows significant improvements in both the fidelity and diversity of the generated scenes. Moreover, our generated result is not tied to fixed resolution by means of the implicit neural representation, as depicted in Fig. 4. For additional results, please refer to the Supplementary Material. + +# 4.3. Applications of Triplane Diffusion + +Scene Inpainting. Fig. 5 presents the qualitative results of our inpainting, demonstrating its effectiveness in inpainting both small and large regions within a scene, while maintaining the coherence of 3D contexts. In detail, the second row of (a) and (b) illustrates the model's seamless removal + +![](images/fc51c5c0950bad760add063b152dd4947ca9b0e8423036bcafff1a89aad4f9d1.jpg) + +![](images/c3f28b3c91df46837035313a79c64064d686caf70078b6f3ed4996796de6861a.jpg) +Figure 6. Our scene outpainting results. The red box at the center of the figure in the first row represents the given scene for outpainting. The zoomed views of the outpainted scene are depicted in the second row. The outpainted scene is expanded from the given size of $256 \times 256 \times 32$ to $1792 \times 3328 \times 32$ . + +![](images/4f1712c7bfa47ec1efce17dac5fd8a5a35134609050cec64c712e36dc70750e7.jpg) + +![](images/5fb3650c0b537da7b60f831f511178f19f79db80c939678746cdda2f621a8db3.jpg) + +
SSC InputMethodIoU ↑mIoU ↑
RGBMonoScene [9]37.1211.50
MonoScene + Ours50.4417.08
OccDepth [29]41.6012.84
OccDepth + Ours50.2016.79
Point CloudSSA-SC [51]58.2524.54
SSA-SC + Ours60.7125.58
SCPNet [49]50.2437.55
SCPNet + Ours59.2538.19
+ +Table 2. Quantitative results of refining SSC on SemanticKITTI validation set [6]. The results are based on the weights released by the authors on GitHub. + +of vehicles, which harmonizes with the adjacent road. The third row demonstrates the insertion of new entities — a vehicle in (a) and a person in (b) — that are contextually congruent with the reference scene. The fourth row in (a) exemplifies the model's dual functionality in both modifying and adding vehicles within the scene. The fifth row in both columns underscores the model's proficiency in modifying scenes. Here, the model alters existing scene components, showcasing its ability to transform the overall ambiance of the scene. These results show our model's adeptness not just in object-level inpainting but also in scene-level inpainting. + +Scene Outpainting. Fig. 6 illustrates a generated outpainting city-level scene, extending a $256 \times 256 \times 32$ scene to a substantial $1792 \times 3328 \times 32$ landscape. Although our model was not designed to generate cityscapes, it demonstrates the capability of maintaining coherence over a large area. The roads are connected in a meaningful and varied way, and various objects such as buildings, cars, and people have been created around them. Please check the details of + +![](images/689f5b59eb6ae1552c4116a3092abcdb3fd6d14e3a5f27725332f7b37af2ae17.jpg) +Figure 7. Semantic scene completion refinement. The SSC metrics are reported in the parentheses as (IoU, mIoU). + +the scene in Fig. 6. + +Semantic Scene Completion Refinement. In Fig. 7, there is a notable semantic and geometric discrepancy between scenes predicted by existing semantic scene completion (SSC) methods and their real scene counterparts. Our model helps to bridge this gap by employing a 3D scene prior, which is effectively modeled through our diffusion model. SCPNet [49] also attempts to learn geometric priors from a teacher model trained with merged sequential frames. Yet, its completion network is inclined toward conservative estimations, which frequently result in partially filled spaces. In the case of SSA-SC [51], which utilizes bird's-eye view features, it is sometimes inappropriate due to the label's inherent limitations of the bird's-eye perspective in capturing certain semantic details. This discrepancy + +![](images/e45ee7569fe3152caba6e7c7df31901e4786ef791f0699cf51da81d0e164a95e.jpg) +(a) + +![](images/e05e11e232fa582a276e9d5dd1f5b96f50dda6f84e1345341f43b9d55752a058.jpg) +(b) +Figure 8. Semantic scene to RGB. (a) Semantic maps are rendered into a driving view from our generated scenes. (b) We generate images through ControlNet [56], an easily accessible imaged-to-image model, with generated semantic maps without shadows. + +
MethodFID ↓KID ↓IS ↑Prec ↑Rec ↑
Ours56.550.043.250.390.32
w/o PE59.140.043.210.300.31
xy-plane (w/o triplane)75.560.063.080.280.27
3D-volume (w/o triplane)136.500.151.960.140.12
+ +Table 3. Ablation studies on scene generation. We ablate our model variants on SemanticKITTI [6]. + +issue is even more evident in RGB-based methods, such as MonoScene [9] and OccDepth [29], which often exhibit diminished sharpness when image features are projected into 3D space. While SSC models show variations from real-world data distributions, our model shows potential in aligning these more closely with reality, as shown in Fig. 7. + +As indicated in Tab. 2, our SSC refinement process appears to offer improvements to all state-of-the-art SSC models. These preliminary results suggest our model's effectiveness in providing not just more accurate semantic segmentation but also a more complete scene. + +Semantic Scene to RGB Image. As illustrated in Fig. 8, we conducted an image-to-image generation experiment. The semantic maps are rendered into a driving view without shadows; then, they are utilized as inputs for ControlNet [56]. The generated RGB images are geometrically and semantically plausible but display a synthetic quality since the pretrained ControlNet was not trained on actual autonomous driving datasets. + +# 4.4. Ablation + +As shown in Tab. 3, we conducted ablation studies on our triplane diffusion model for scene generation, focusing on two key design elements: + +Positional Embedding. Our variant excluded positional embedding PE, which generates high-frequency features critical for detailed scene reconstruction. Its absence resulted in lower performance across all metrics. + +Triplane Representation. We evaluated the effectiveness of the triplane representation for real-outdoor scene gen + +eration. The triplane and $xy$ -plane enhance the generation quality compared with 3D features, while the $xy$ -plane shows lower performance than triplanes. We suspect that the excessive factorization limits the representation capability of the $xy$ -plane compared to the triplane. + +# 4.5. Limitation + +While our model demonstrates significant progress in the generation of 3D real-outdoor scenes, it inherently reflects the characteristics of its training data, as detailed in Sec. 4.1. This reliance introduces several limitations. One notable challenge is the model's difficulty in accurately depicting areas occluded from the sensor's viewpoint, such as the rear sides of buildings, often leading to their incomplete representation in our generated scenes. Moreover, since the dataset is captured from a driving view, there is an inherent shortfall in capturing the full height of buildings. This brings a partial representation of the vertical structure of buildings and other tall elements in the scenes. Another issue is the model tends to produce traces of moving objects stemming from dataset pre-processing that merges sequential frames. For future work, incorporating prior knowledge of the city into the model could yield better results, such as addressing occluded areas or building heights. + +# 5. Conclusion + +We have proposed a diffusion framework called SemCity for real-world outdoor scene generation. The seminal idea is to generate a scene by factorizing real-outdoor scenes into triplane representations. Our triplane representation outperforms traditional voxel-based approaches, producing scenes that are not only visually more appealing but also rich in semantic detail, effectively capturing the complexity of various objects within the scene. Ours is not constrained by fixed resolutions thanks to the incorporation of an implicit neural representation. We have further expanded the capabilities of our triplane diffusion model to several practical applications, including scene inpainting, scene outpainting, and semantic scene completion refinement. Specifically, by manipulating triplanes during the diffusion process, we achieve seamless inpainting and outpainting at both the object and scene levels. Ours is used to more closely align the scenes predicted by existing semantic scene completion methods with the actual data distribution using our learned 3D prior. We believe that our work provides a road map of real-outdoor scene generation to research communities. + +Acknowledgement. This work was supported by IITP (Institute of Information & Communications Technology Planning & Evaluation) and ITRC (Information Technology Research Center), funded by Korea government (MSIT) (RS-2023-00237965(2024) and IITP-2024-2020-0-01460). Prof. Sung-Eui Yoon is a corresponding author. + +# References + +[1] Ali Abbasi, Sinan Kalkan, and Yusuf Sahillioğlu. Deep 3d semantic scene extrapolation. The Visual Computer, 35:271-279, 2019. 3 +[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12608-12618, 2023. 2 +[3] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12608-12618, 2023. 3 +[4] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 4 +[5] Luca Bartolomei, Lucas Teixeira, and Margarita Chli. Perception-aware path planning for uavs using semantic segmentation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5808-5815, 2020. 3 +[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 5, 6, 7, 8 +[7] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 4 +[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6 +[9] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 6, 7, 8 +[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[11] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 3 +[12] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse + +image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 3 +[13] Xu Cui, Chenggang Lu, and Jinxiang Wang. 3d semantic map construction using improved orb-slam2 for mobile robot in edge computing environment. IEEE Access, 8:67179-67191, 2020. 3 +[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 4 +[17] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2 +[18] Changho Jo, Woobin Im, and Sung-Eui Yoon. In-n-out: Towards good initialization for inpainting and outpainting. arXiv preprint arXiv:2106.13953, 2021. 3 +[19] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 2 +[20] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2 +[21] Zeqiang Lai, Yuchen Duan, Jifeng Dai, Ziheng Li, Ying Fu, Hongsheng Li, Yu Qiao, and Wenhai Wang. Denoising diffusion semantic segmentation with mask prior modeling. arXiv preprint arXiv:2306.01721, 2023. 5 +[22] Jumin Lee, Woobin Im, Sebin Lee, and Sung-Eui Yoon. Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527, 2023. 2, 6 +[23] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. arXiv preprint arXiv:2212.05993, 2022. 3 +[24] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2 +[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 3 +[26] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2 + +[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 4, 5 +[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 2 +[29] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 3, 6, 7, 8 +[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 4 +[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[33] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 4 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5 +[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1, 2 +[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2 +[37] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2 +[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2 +[39] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene com + +pletion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 3 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 2 +[41] Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner. Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207, 2023. 2, 6 +[42] Chaoqun Wang, Wenzheng Chi, Yuxiang Sun, and Max Q-H Meng. Autonomous robotic exploration by incremental road map construction. IEEE Transactions on Automation Science and Engineering, 16(4):1720-1731, 2019. 3 +[43] Chaoqun Wang, Delong Zhu, Teng Li, Max Q-H Meng, and Clarence W De Silva. Efficient autonomous robotic exploration with semantic road map in indoor environments. IEEE Robotics and Automation Letters, 4(3):2989-2996, 2019. 3 +[44] Xinying Wang, Dikai Xu, and Fangming Gu. 3d model inpainting based on 3d deep convolutional generative adversarial network. IEEE Access, 8:170355-170363, 2020. 3 +[45] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Pet-neus: Positional encoding tri-planes for neural surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12598–12607, 2023. 4 +[46] Zhen Wang, Shijie Zhou, Jeong Joon Park, Despoina Paschalidou, Suya You, Gordon Wetzstein, Leonidas Guibas, and Achuta Kadambi. Alto: Alternating latent topologies for implicit 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 259-270, 2023. 2 +[47] Joey Wilson, Jingyu Song, Yuewei Fu, Arthur Zhang, Andrew Capodieci, Paramsothy Jayakumar, Kira Barton, and Maani Ghaffari. *Motionsc: Data set and network for real-time semantic mapping in dynamic environments*. arXiv preprint arXiv:2203.07060, 2022. 5, 6 +[48] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2 +[49] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 3, 6, 7 +[50] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6 +[51] Xuemeng Yang, Hao Zou, Xin Kong, Tianxin Huang, Yong Liu, Wanlong Li, Feng Wen, and Hongbo Zhang. Semantic segmentation-assisted scene completion for lidar point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3555-3562. IEEE, 2021. 3, 7 + +[52] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3 +[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4206-4216, 2023. 2 +[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[55] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2, 6 +[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 8 +[57] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10188-10198, 2023. 2 +[58] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2 \ No newline at end of file diff --git a/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/images.zip b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..355ba2dd275c4759d49a07a5db58b23d69575a2a --- /dev/null +++ b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c9ffc3c8637429f53785b5a7d6d28abdc924f642a91fb9a668195b241540619 +size 928490 diff --git a/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/layout.json b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bbd64da413e400e12f8c3dc99abf910937c54cfd --- /dev/null +++ b/2024/SemCity_ Semantic Scene Generation with Triplane Diffusion/layout.json @@ -0,0 +1,9050 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 487, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 487, + 121 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 487, + 121 + ], + "type": "text", + "content": "SemCity: Semantic Scene Generation with Triplane Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": "Jumin Lee" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": " Sebin Lee" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": " Changho Jo" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": " Woobin Im" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": " Juhyeong Seon" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": " Sung-Eui Yoon" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 142, + 521, + 185 + ], + "type": "text", + "content": "Neosapience, Inc." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 198, + 173, + 366 + ], + "blocks": [ + { + "bbox": [ + 50, + 198, + 173, + 366 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 173, + 366 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 173, + 366 + ], + "type": "image", + "image_path": "41e3c143eef09caf41f959ed3acc2728a1a2eda481960466c73f6825302a925c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 368, + 157, + 379 + ], + "lines": [ + { + "bbox": [ + 64, + 368, + 157, + 379 + ], + "spans": [ + { + "bbox": [ + 64, + 368, + 157, + 379 + ], + "type": "text", + "content": "(a) Scene generation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 200, + 441, + 291 + ], + "blocks": [ + { + "bbox": [ + 176, + 200, + 441, + 291 + ], + "lines": [ + { + "bbox": [ + 176, + 200, + 441, + 291 + ], + "spans": [ + { + "bbox": [ + 176, + 200, + 441, + 291 + ], + "type": "image", + "image_path": "3564e47203d7a727abfbd1630678f15064257f8ea24a305c1b3853f8db2f7e88.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 293, + 298, + 304 + ], + "lines": [ + { + "bbox": [ + 200, + 293, + 298, + 304 + ], + "spans": [ + { + "bbox": [ + 200, + 293, + 298, + 304 + ], + "type": "text", + "content": "(c) Scene outpainting" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 449, + 203, + 539, + 291 + ], + "blocks": [ + { + "bbox": [ + 449, + 203, + 539, + 291 + ], + "lines": [ + { + "bbox": [ + 449, + 203, + 539, + 291 + ], + "spans": [ + { + "bbox": [ + 449, + 203, + 539, + 291 + ], + "type": "image", + "image_path": "db600fd2c966d0be4c3d13c0eb9c5a682c3eebb88015381d387babaae39f3f73.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 176, + 306, + 416, + 364 + ], + "blocks": [ + { + "bbox": [ + 176, + 306, + 416, + 364 + ], + "lines": [ + { + "bbox": [ + 176, + 306, + 416, + 364 + ], + "spans": [ + { + "bbox": [ + 176, + 306, + 416, + 364 + ], + "type": "image", + "image_path": "c599459d462a207f589abd6de1133221a987f1391f5b62a3e924909b8cd5add5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 366, + 392, + 377 + ], + "lines": [ + { + "bbox": [ + 201, + 366, + 392, + 377 + ], + "spans": [ + { + "bbox": [ + 201, + 366, + 392, + 377 + ], + "type": "text", + "content": "(b) Semantic scene completion refinement" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 424, + 317, + 539, + 365 + ], + "blocks": [ + { + "bbox": [ + 424, + 317, + 539, + 365 + ], + "lines": [ + { + "bbox": [ + 424, + 317, + 539, + 365 + ], + "spans": [ + { + "bbox": [ + 424, + 317, + 539, + 365 + ], + "type": "image", + "image_path": "5adfb0a1a22d8c604a474870f7dfa47b5b861f77d72c48444a7cf65bc9f34b81.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 366, + 527, + 377 + ], + "lines": [ + { + "bbox": [ + 436, + 366, + 527, + 377 + ], + "spans": [ + { + "bbox": [ + 436, + 366, + 527, + 377 + ], + "type": "text", + "content": "(d) Scene inpainting" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 388, + 547, + 433 + ], + "lines": [ + { + "bbox": [ + 45, + 388, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 45, + 388, + 547, + 433 + ], + "type": "text", + "content": "Figure 1. We introduce a diffusion framework, SemCity, designed for generating semantic scenes in real-world outdoor environments as shown in (a). We extend our diffusion model to various practical tasks: semantic scene completion refinement, scene outpainting, and scene inpainting. For instance, the comprehensive scenario is displayed in " + }, + { + "bbox": [ + 45, + 388, + 547, + 433 + ], + "type": "inline_equation", + "content": "(\\mathrm{b})\\rightarrow (\\mathrm{c})\\rightarrow (\\mathrm{d})" + }, + { + "bbox": [ + 45, + 388, + 547, + 433 + ], + "type": "text", + "content": " : the refined scene (SSC refinement) (b) is outpainted to a broader scene (c); then, an object (in this case, a car) is seamlessly integrated into the scene via our inpainting process (d)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 452, + 192, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 452, + 192, + 464 + ], + "spans": [ + { + "bbox": [ + 143, + 452, + 192, + 464 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 477, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 289, + 693 + ], + "type": "text", + "content": "We present \"SemCity,\" a 3D diffusion model for semantic scene generation in real-world outdoor environments. Most 3D diffusion models focus on generating a single object, synthetic indoor scenes, or synthetic outdoor scenes, while the generation of real-world outdoor scenes is rarely addressed. In this paper, we concentrate on generating a real-outdoor scene through learning a diffusion model on a real-world outdoor dataset. In contrast to synthetic data, real-outdoor datasets often contain more empty spaces due to sensor limitations, causing challenges in learning realoutdoor distributions. To address this issue, we exploit a triplane representation as a proxy form of scene distributions to be learned by our diffusion model. Furthermore, we propose a triplane manipulation that integrates seamlessly with our triplane diffusion model. The manipulation improves our diffusion model's applicability in a variety of downstream tasks related to outdoor scene generation such as scene inpainting, scene outpainting, and semantic" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 453, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 545, + 596 + ], + "type": "text", + "content": "scene completion refinements. In experimental results, we demonstrate that our triplane diffusion model shows meaningful generation results compared with existing work in a real-outdoor dataset, SemanticKITTI. We also show our triplane manipulation facilitates seamlessly adding, removing, or modifying objects within a scene. Further, it also enables the expansion of scenes toward a city-level scale. Finally, we evaluate our method on semantic scene completion refinements where our diffusion model enhances predictions of semantic scene completion networks by learning scene distribution. Our code is available at https://github.com/zoomin-lee/SemCity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 621, + 387, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 387, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 387, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 712 + ], + "type": "text", + "content": "Diffusion models [16] have emerged as a promising generation tool owing to its state-of-the-art generation results in image domain [35, 36]. This advance has led to active exploration in extending diffusion models to 3D data generation across both academic and industrial groups. In the 3D domain, diffusion models have shown remarkable" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 272, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 272, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 272, + 712 + ], + "type": "text", + "content": "*Both authors contributed equally to this work as co-first authors." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28337" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "capabilities in generating diverse 3D forms (e.g., voxels, meshes) [24, 26]. While those 3D diffusion models primarily aim to craft a single object, generating scenes consisting of multiple objects remains a relatively unexplored area in the 3D diffusion domain." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 252 + ], + "type": "text", + "content": "Scene generative diffusion models focus on crafting both geometrically and semantically coherent environments. Compared with a single object generation, generating a scene with multiple objects requires an understanding of more complex geometric and semantic structure due to a larger spatial extent [41]. There are primarily two streams of scene generative diffusion models, each tailored to either indoor or outdoor settings. In particular, the outdoor environments have inherent challenges caused by a broader landscape than indoor ones." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 397 + ], + "type": "text", + "content": "We propose to leverage a triplane representation [10] for broader outdoor scenes, a method of factorizing 3D data onto three orthogonal 2D planes, as utilized in 3D object reconstruction and NeRF models [2, 31, 46]. We excavate its advantages in addressing the data sparsity problem typically found in outdoor datasets due to the sensor limitations (e.g., occlusions, range constraints) in capturing outdoor scenes. Triplane representation helps to reduce the inclusion of unnecessary empty information through the factorization of 3D data to 2D planes [10]. This efficiency in capturing relevant spatial detail makes it an effective tool for representing the many objects typically found in outdoor environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": "In this paper, we design our diffusion framework based on triplane representations. Our triplane autoencoder learns to compress a voxelized scene into a triplane representation by reconstructing semantic labels of the scene. Following this, the triplane diffusion model is trained and used to generate new scenes, as shown in Fig. 1(a), by creating novel triplanes based on the efficient representation. Further, we propose a triplane manipulation method, which extends our triplane diffusion model toward several practical tasks (i.e., scene inpainting, scene outpainting, and semantic scene completion refinements) as shown in Fig. 1(b-d). Our method can seamlessly add, remove, and modify objects in real-outdoor scenes while maintaining the semantic coherence of environments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 568, + 244, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 568, + 244, + 579 + ], + "spans": [ + { + "bbox": [ + 59, + 568, + 244, + 579 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 582, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 582, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 286, + 640 + ], + "type": "text", + "content": "- We disclose the applicability of the triplane representation through generating semantic scenes for real-outdoor environments and extend its views in practical downstream tasks: scene inpainting, scene outpainting, and semantic scene completion refinement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 642, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 675 + ], + "type": "text", + "content": "- We propose to manipulate triplane features during our diffusion process, facilitating seamlessly extending our method toward the downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "- We demonstrate that the proposed method significantly enhances the quality of generated scenes in real-world outdoor environments." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 93, + 545, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 93, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 304, + 93, + 545, + 273 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models [16] learn data distributions via iterative denoising processes based on score functions [40]. Its generated results have shown remarkably realistic appearances with high fidelity and diversity in a variety of 2D image synthesis such as outpainting [35, 57], inpainting [27, 35] and text-to-image generation [32, 36]. Built upon these achievements, diffusion models have also been extended into the 3D domain, generating impressive results in various 3D shapes, including voxel grids [24, 58], point clouds [28, 53, 54], meshes [26], and implicit functions [20, 37, 38, 48]. While these models can craft a single 3D object, our model focuses on generating a 3D scene composed of multiple objects using a categorical voxel data structure, which is a relatively under-explored area in the 3D diffusion domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 495 + ], + "type": "text", + "content": "Diffusion Models for Scene Generation. In contrast to a single object generation, scene generation involves an understanding of the larger 3D space, causing more semantic and geometric complexities [41]. Diffusion models for scene generation have been studied in both indoor and outdoor environments. In indoor settings, diffusion models aim to learn distributions of relations among objects by representing them as scene graphs [19]. The scene graphs contain object attributes (e.g., location, orientation, and size), capturing the intricate inter-object relationships within bounded spaces [41, 55]. For outdoor scenes, the challenges are distinct, frequently including a lot of empty areas (e.g., sky, open areas) resulting from the broader landscapes. Traditional approach [22] has relied on discrete diffusion methods [17] on voxel space, necessitating a detailed representation of every air volume." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we demonstrate that triplane diffusion is highly effective for generating real-outdoor scenes. By abstracting 3D spaces into three orthogonal 2D planes, triplane representation [10] effectively captures the vastness of outdoor environments, predominantly composed of air. Beyond its data efficiency, the triplane excels in focusing on other significant objects (e.g., vehicles, buildings) by allocating lesser attention to less informative elements like air. Our approach stands in stark contrast to prior work [22], which were constrained to synthetic datasets with considerably less empty space. In real datasets, inherent sensor limitations, such as a limited field-of-view, a limited ranges, and the inability to capture occluded areas like the rears of buildings, lead to a prevalence of empty space. Furthermore, we emphasize the versatility of our framework by demonstrating its extension to various downstream tasks, including scene inpainting, outpainting, and semantic scene completion refinement." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28338" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 427, + 141 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 427, + 141 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 427, + 141 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 427, + 141 + ], + "type": "image", + "image_path": "c0cb1b2c89dd38346bd4de7735e64d2632e7f6dcd0eeeb43c77c5675768555b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 144, + 427, + 220 + ], + "blocks": [ + { + "bbox": [ + 49, + 144, + 427, + 220 + ], + "lines": [ + { + "bbox": [ + 49, + 144, + 427, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 427, + 220 + ], + "type": "image", + "image_path": "826c63a724b7e726ff79b1c07c5ce0f74cd67193c227f600214d1796f3c40aae.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": "Figure 2. Overview of ours. (a) A 3D semantic map " + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": " is encoded by a triplane encoder " + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": " and factorized to a triplane " + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": ". The triplane coupled with a positional encoding " + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "inline_equation", + "content": "\\mathbb{P}\\mathbb{E}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": " is decoded by an implicit decoder " + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": ", resulting in class probabilities for each coordinate " + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": ". (b) Our triplane diffusion model " + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 46, + 230, + 546, + 276 + ], + "type": "text", + "content": " learns to generate a novel triplane for semantic scene generation via denoising diffusion process. (c) We further extend our triplane diffusion beyond a simple scene generation toward various practical scenarios by manipulating triplanes in (b)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 545, + 220 + ], + "type": "image", + "image_path": "1f46fdb4aad04ff6e1b856abdf10c4d852743922fef0451ad2c6e129240378c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 492 + ], + "type": "text", + "content": "3D Inpainting and Outpainting. In 3D inpainting, the primary objective is to fill in missing portions or modify existing elements of 3D data while maintaining geometric consistency. Most existing works concentrate on single-object inpainting [3, 23, 44]; for instance, they seamlessly transit a 3D chair's leg count from three to four. Contrary to inpainting, 3D outpainting is to extrapolate a given scene over an unobserved space. The existing work [1] focuses on scene outpainting within bounded indoor environments such as rooms. While in 2D images, inpainting and outpainting are not restricted to a single object [12, 18, 52]. Likewise, in 3D space, we focus on scene-level inpainting, which seamlessly adds, removes, or modifies objects in a scene. Further, our scene-level outpainting is not constrained to the bounded scenes; we extrapolate the outdoor scene from sensor range (e.g., LiDAR) to city-scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Semantic Scene Completion. Semantic scene completion (SSC) [39] is pivotal for 3D scene understanding, where it jointly infers completion and semantic segmentation of the 3D scene from sensor observations such as RGB images [9, 25, 29] or point clouds [49-51]. In addition, SSC plays a crucial role in supporting comprehensive autonomous navigation systems, notably in essential downstream tasks like path planning [5, 11, 43] and map construction [13, 42]. Despite significant progress in the field, a persistent challenge is the semantic and geometric discrepancies between the SSC-estimated scenes and their real counterparts, as illustrated in Fig. 1(c). These discrepancies can undermine the performance of downstream tasks. Our triplane diffusion model can help to bridge this gap by exploiting 3D scene priors. This approach enhances the reliability and effectiveness of SSC, which is expected to improve its application in autonomous navigation systems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 299, + 361, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 361, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 361, + 312 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "text", + "content": "In this section, we elucidate our triplane diffusion model and its extensions. Our triplane diffusion model aims to synthesize novel real-outdoor scenes by generating a triplane, a proxy representation, which effectively addresses the inherent challenges of real-outdoor scene synthesis. This triplane representation is learned by our triplane autoencoder, which abstracts the geometric and semantic intricacies of a scene into three orthogonal 2D feature planes, namely, the " + }, + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "inline_equation", + "content": "xz" + }, + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "inline_equation", + "content": "yz" + }, + { + "bbox": [ + 304, + 320, + 545, + 488 + ], + "type": "text", + "content": " planes (Sec. 3.1). Then, our diffusion model learns triplane distributions of scenes, generating novel triplanes (Sec. 3.2). We extend our triplane diffusion model toward various practical scenarios: scene inpainting, scene outpainting, and semantic scene completion refinements (Sec. 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 500, + 541, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 500, + 541, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 500, + 541, + 514 + ], + "type": "text", + "content": "3.1. Representing a Semantic Scene with Triplane" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "text", + "content": "To represent a 3D scene as a triplane, our triplane autoencoder learns to compress a 3D scene into a triplane representation as shown in Fig. 2(a). The autoencoder consists of two modules: (1) an encoder " + }, + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "text", + "content": " yielding a triplane, and (2) an implicit multi-layer perceptron (MLP) decoder " + }, + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "text", + "content": " for reconstruction from the triplane." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": " takes a voxelized scene " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{X \\times Y \\times Z}" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": " classes within a spatial grid of resolution " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "X \\times Y \\times Z" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": ". It then yields an axis-aligned triplane representation " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = [\\mathbf{h}^{xy}, \\mathbf{h}^{xz}, \\mathbf{h}^{yz}]" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": ". The triplane consists of three planes, each characterized by distinct dimensional properties: " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{xy} \\in \\mathbb{R}^{C_h \\times X_h \\times Y_h}" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{xz} \\in \\mathbb{R}^{C_h \\times X_h \\times Z_h}" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{yz} \\in \\mathbb{R}^{C_h \\times Y_h \\times Z_h}" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": " stands for a feature dimension, and " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "X_h, Y_h" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "Z_h" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": " denote spatial dimension of the triplane. During the encoding phase, a 3D feature volume is extracted by 3D convolutional layers from the scene " + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 593, + 546, + 713 + ], + "type": "text", + "content": ", re" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28339" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "sulting in the triplane via axis-wise average pooling. Given a 3D coordinate " + }, + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (x,y,z)" + }, + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": ", the triplane is interpreted as a summation of vectors bilinearly interpolated from each plane: " + }, + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(\\mathbf{p}) = \\mathbf{h}^{xy}(x,y) + \\mathbf{h}^{xz}(x,z) + \\mathbf{h}^{yz}(y,z)" + }, + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": "To reconstruct the 3D scene " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": ", we decode the encoded triplane " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": " with an implicit MLP decoder " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": " that predicts semantic class probabilities. The decoder takes the triplane vector " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": " with its sinusoidal positional embedding " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\mathrm{PE}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": " [30], resulting in class probabilities " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(\\mathbf{p}) = g_{\\theta}(\\mathbf{h}(\\mathbf{p}),\\mathrm{PE}(\\mathbf{p}))\\in [0,1]^N" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": ". The positional embedding produces high-frequency features according to the coordinates " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": ", which helps the implicit decoder " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": " represent high-frequency scene contents [45]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "text", + "content": " and the MLP decoder " + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "text", + "content": " are trained with the autoencoder loss " + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AE}}" + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "text", + "content": " and scene label " + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(\\mathbf{p})" + }, + { + "bbox": [ + 47, + 228, + 286, + 251 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 260, + 286, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 286, + 274 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A E}} = \\mathbb {E} _ {\\mathbf {p} \\sim \\mathcal {P}} \\left[ \\ell_ {\\mathrm {C E}} (\\mathbf {c} (\\mathbf {p}), \\mathbf {x} (\\mathbf {p})) + \\lambda \\ell_ {\\mathrm {L Z}} (\\mathbf {c} (\\mathbf {p}), \\mathbf {x} (\\mathbf {p})) \\right], \\tag {1}", + "image_path": "90c4d5ba8b3c46c3d185faa3a0ea676373db63fc20ec4ca8dd04b18c3c439379.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "text", + "content": " is a loss weight, and " + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "text", + "content": " is the set of grid coordinates of the scene. We use the weighted cross-entropy loss " + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{CE}}" + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "text", + "content": " [33] and the Lovász-softmax loss " + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{LZ}}" + }, + { + "bbox": [ + 47, + 281, + 286, + 328 + ], + "type": "text", + "content": " [7] to learn imbalanced semantic distributions of the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 336, + 155, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 155, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 155, + 348 + ], + "type": "text", + "content": "3.2. Triplane Diffusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": "Based on the triplane representation of the 3D semantic scene, our triplane diffusion model " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": " learns to generate a novel triplane through denoising diffusion probabilistic models [16] as shown in Fig. 2(b). This triplane generation leads to generation of 3D scene through decoding the generated triplane with the implicit MLP decoder " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": ". Through the " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": "-parameterization [4], the diffusion model " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": " is trained to reconstruct the triplane " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": " given its corrupted triplane " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": " sampled from a diffusion process " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "q(\\mathbf{h}_t|\\mathbf{h}) = \\mathcal{N}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{h}, (1 - \\bar{\\alpha}_t)\\mathbf{I})" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": " is the Gaussian distribution, " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": " with a variance schedule " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": ". The diffusion process " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "q(\\mathbf{h}_t|\\mathbf{h})" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": " is derived from the Markovian chain rule with a single step's diffusion process " + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "inline_equation", + "content": "q(\\mathbf{h}_t|\\mathbf{h}_{t-1}) = \\mathcal{N}(\\sqrt{1 - \\beta_t}\\mathbf{h}_{t-1}, \\beta_t\\mathbf{I})" + }, + { + "bbox": [ + 46, + 354, + 286, + 533 + ], + "type": "text", + "content": ". Thus, the triplane diffusion loss is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 543, + 286, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 543, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 94, + 543, + 286, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D}} = \\mathbb {E} _ {t \\sim \\mathcal {U} (1, T)} | | \\mathbf {h} - D _ {\\phi} (\\mathbf {h} _ {t}, t) | | _ {p}, \\tag {2}", + "image_path": "925ecd082f6417d341c49039f8c531decbcabf3bf9fc0483bd65adfb4bb7112e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "text", + "content": " is the number of denoising steps, and " + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "text", + "content": " represents the order of the norm. The timestep " + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "text", + "content": " is sampled from the discrete uniform distribution " + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 563, + 286, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "text", + "content": "After training, the diffusion model " + }, + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "text", + "content": " generates a novel triplane " + }, + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_0" + }, + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "text", + "content": " via the iterative DDPM generation process [16] starting from " + }, + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 600, + 286, + 635 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 643, + 286, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 643, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 88, + 643, + 286, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {t - 1} \\sim \\mathcal {N} \\left(\\gamma_ {t} \\mathbf {h} _ {t} + \\delta_ {t} D _ {\\phi} \\left(\\mathbf {h} _ {t}, t\\right), \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\tag {3}", + "image_path": "e4844c1ae8ed7671d689cc52cd12eda2f8d0915078b04ee5200093798ab1ee23.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{t} \\coloneqq \\sqrt{\\alpha_{t}} (1 - \\bar{\\alpha}_{t-1}) / (1 - \\bar{\\alpha}_{t})" + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{t} \\coloneqq \\sqrt{\\bar{\\alpha}_{t-1}} \\beta_{t} / (1 - \\bar{\\alpha}_{t})" + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "content": ". From the generated triplane " + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_0" + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "content": ", we generate a novel 3D semantic scene " + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "content": " by querying coordinates " + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "content": " to the implicit decoder, i.e., " + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "g_{\\theta}(\\mathbf{h}_0(\\mathbf{p}), \\mathsf{PE}(\\mathbf{p}))" + }, + { + "bbox": [ + 47, + 665, + 286, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "type": "text", + "content": "3.3. Applications with Triplane Manipulation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 91, + 545, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 139 + ], + "type": "text", + "content": "Building on the triplane diffusion process (Sec. 3.2), we propose a triplane manipulation that allows our model to facilitate a variety of practical downstream tasks with few modifications, as illustrated in Fig. 2(c)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "text", + "content": "Scene Inpainting. Our scene inpainting randomly edits a 3D scene, seamlessly adding, modifying, or removing objects while maintaining the consistency and realism of the scene. For instance, the inpainting includes scenarios where cars or sidewalks appear and then disappear, or vice versa, as shown in Fig. 1(d). Inspired by the RePaint sampling strategy [27], we propose a 3D-aware inpainting approach with semantic coherence. RePaint focuses on the image domain without explicitly considering the fidelity of the underlying 3D scene. To facilitate 3D-aware inpainting, we inpaint the triplane, serving as a compact proxy representation for the scene. We define a binary spatial trimask " + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{m} = [\\mathbf{m}^{xy},\\mathbf{m}^{xz},\\mathbf{m}^{yz}]" + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "text", + "content": " covering inpainting regions on triplane space, allowing us to control the generation process on the masked region. The trimask " + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "text", + "content": " is set to have ones for inpainting regions and zeros for others. We override the " + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "text", + "content": "-th triplane " + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t = [\\mathbf{h}_t^{xy},\\mathbf{h}_t^{xz},\\mathbf{h}_t^{yz}]" + }, + { + "bbox": [ + 304, + 162, + 545, + 377 + ], + "type": "text", + "content": " of the generation process (Eq. 3) as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 390, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 390, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 351, + 390, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {t} \\leftarrow \\mathbf {m} \\otimes \\mathbf {h} _ {t} + (\\mathbf {1} - \\mathbf {m}) \\otimes \\mathbf {h} _ {t} ^ {\\text {k n o w n}}, \\tag {4}", + "image_path": "83c1aa83d5b102c7fa9a551db70cdafa146ce6da8e7de45bd0c04f23b139cf19.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "text", + "content": " is the element-wise product, and the known triplane " + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^{\\mathrm{known}}" + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "text", + "content": " for intact regions is sampled from the diffusion process, i.e., " + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "inline_equation", + "content": "q(\\mathbf{h}_t^{\\mathrm{known}}|\\mathbf{h})\\coloneqq q(\\mathbf{h}_t|\\mathbf{h})" + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "text", + "content": ", which adheres to a known Gaussian distribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Scene Outpainting. Our scene outpainting extends the boundaries of the 3D scene without additional training as with inpainting. To seamlessly outpaint a scene, the regions to be extended should be conditioned on the original scene. We propose to inject this intuition into triplane representation as shown in Fig. 2(c). For covering regions to be outpainted, our diffusion model generates a novel triplane that is partially overlapped with the original triplane. We implement this with the concept of the trimask m and the known triplane " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^{\\mathrm{known}}" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " as in our scene inpainting. The trimask m covers regions to be outpainted, and the " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^{\\mathrm{known}}" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " is obtained from the intersections of triplanes between original and outpainted regions. Based on the outpainting strategy, we extend a given scene toward cardinal and intercardinal directions, facilitating the creation of an unbounded scene. While our triplane diffusion model is trained using triplanes of a fixed size, it demonstrates the capability to outpaint scenes to be several times bigger than the original scene, as illustrated in Fig. 1(c)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28340" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 547, + 350 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 547, + 350 + ], + "type": "image", + "image_path": "a5dc446c9fe8e2b8234df33be7aff550d99df5adda0b47342e66fffe12ccf457.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 546, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 546, + 380 + ], + "type": "text", + "content": "Figure 3. Scene generation results using both real and synthetic outdoor datasets - SemanticKITTI [6] and CarlaSC [47]. Our results showcase the effective generation of overall structures, including roads and buildings, along with detailed objects such as cars." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "content": "Semantic Scene Completion Refinement. SSC models complete and segment a 3D scene from sensor observations such as images or point clouds. We observe that the SSC results show a geometric and semantic discrepancy compared with data distributions, as shown in Fig. 1(b). We extend our triplane diffusion model to refine predictions of SSC models toward reducing the discrepancy. To effectively condition our triplane diffusion scheme to SSC model's prediction " + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{ssc}}" + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "content": ", we utilize its triplane representation " + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{\\mathrm{ssc}} = f_{\\theta}(\\mathbf{x}^{\\mathrm{ssc}})" + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "content": " derived by our triplane encoder " + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "content": ". We extend our triplane diffusion scheme with a simple modification of the triplane " + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t" + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "content": " in the diffusion loss (Eq. 2) and the generation process (Eq. 3) as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 560, + 287, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 287, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {t} = \\mathbf {h} _ {t} ^ {\\mathrm {s s c}} \\oplus \\mathbf {h} ^ {\\mathrm {s s c}}, \\tag {5}", + "image_path": "ec03c3b2d3a62934b625c835b75348b473b0c0f120b410cbc60c6d1ba6af54b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "text", + "content": " is concatenation, and " + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^{\\mathrm{ssc}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "text", + "content": "-th diffused triplane sampled by the DDPS diffusion process [21] with the SSC prediction's triplane " + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{\\mathrm{ssc}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 169, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 169, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 169, + 659 + ], + "type": "text", + "content": "4.1. Experimental Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": "Training Dataset. We validate our method on the SemanticKITTI [6] and CarlaSC [47] datasets. SemanticKITTI provides 3D semantic scenes of real-outdoor environments with labels for 20 semantic classes. Each" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": "scene is represented by a voxel grid of " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": ", covering an area of " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "51.2\\mathrm{m}" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": " in front of the car, extending " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "51.2\\mathrm{m}" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": " on each side, and reaching up to a height of " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "6.4\\mathrm{m}" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": ". The dataset retains object motion traces as a result of sensor frame integration, which is employed to establish a dense ground truth. In contrast, CarlaSC is a synthetic dataset that provides 3D semantic outdoor scenes without the trace of moving objects. The dataset contains annotated 11 semantic classes with a voxel grid of " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 8" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": " and covers a distance of " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "25.6\\mathrm{m}" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": " in front and behind the car, " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "25.6\\mathrm{m}" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": " laterally on each side, and " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "3\\mathrm{m}" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": " in height." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": "Implementation Details. Our experiments are deployed on a single NVIDIA RTX 3090 GPU with a batch size of 4 for the triplane autoencoder and 18 for the triplane diffusion model. For the triplane autoencoder, the input scene is encoded to triplane with a spatial resolution " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "(X_h, Y_h, Z_h) = (128, 128, 32)" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": ", and the feature dimension " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " is 16. The loss weight " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " in Eq. 1 is set to 1.0. The order of the norm " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " in Eq. 2 is set to 1 for SSC refinements and 2 for other cases. For the diffusion model, the learning rate is initialized to 1e-4 and then decreases linearly. During the diffusion process, we use the default settings [34] with 100 time steps " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": ". For our triplane inpainting and outpainting, we employ the RePaint sampling strategy [27] as a reference and perform a repaint with 5 resampling and a jump size of 20." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28341" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 150 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 150 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 150 + ], + "type": "table", + "html": "
ModelFID ↓KID ↓IS ↑Prec ↑Rec ↑
SemanticKITTI [6]
SSD [22]112.820.122.230.010.08
SemCity (Ours)56.550.043.250.390.32
CarlaSC [47]
SSD [22]87.390.092.440.140.07
SemCity (Ours)40.630.023.510.310.09
", + "image_path": "fb82f8d7a57c34f36b32e7df2dd537a33b90bbfbc8709e4bd7525b6d1b51da3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 191 + ], + "type": "text", + "content": "Table 1. Results of semantic scene generation. Each metric is computed between the rendered image of the generated scene and the ones of the actual scene at a resolution of " + }, + { + "bbox": [ + 46, + 159, + 287, + 191 + ], + "type": "inline_equation", + "content": "1440 \\times 2048" + }, + { + "bbox": [ + 46, + 159, + 287, + 191 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 209, + 288, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 425 + ], + "type": "text", + "content": "Evaluation Metrics. Following the scene generation works [41, 55], we evaluate the performance of semantic scene generation by examining both the diversity and fidelity of 3D semantic scenes within the rendered images. We use recall to evaluate diversity, while precision and inception score (IS) are used to evaluate fidelity. The Fréchet Inception Distance (FID) [15] and Kernel Inception Distance (KID) [8] metrics are also utilized, as they reflect the combined effect of both diversity and fidelity on scene quality [14]. In terms of semantic scene completion (SSC) refinement performance, we follow the protocols defined in SSC works [9, 29, 49, 50]. The Intersection-over-Union (IoU) metric is used to quantify scene completeness, while the mean IoU (mIoU) provides a measure for the quality of semantic segmentation. These metrics together enable a comprehensive evaluation of how well the semantic scene completion methods perform in terms of accurately filling in and labeling the scene components." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 432, + 198, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 198, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 198, + 443 + ], + "type": "text", + "content": "4.2. Semantic Scene Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": "Fig. 3 illustrates a qualitative comparison on the SemanticKITTI [6] and CarlaSC [47] datasets. SSD [22] shows impressive results on the CarlaSC dataset. However, its performance on the SemanticKITTI dataset, which is a real-world dataset, is notably constrained. This limitation primarily arises from the SSD's voxel-based representation, which struggles with the more prevalent empty spaces in real-outdoor datasets compared to synthetic ones. This issue is especially evident in the generation of buildings and roads, where SSD often fails to define boundaries accurately. Furthermore, the model's limitations extend to representing finer structures, such as trunks and leaves, as well as traffic light poles and signals. It also struggles with generating uniform shapes of vehicles, often resulting in irregular shapes. In contrast, our method demonstrates the ability to effectively synthesize detailed scenes even on the real dataset, as illustrated in Fig. 3. It performs better than SSD [22] in accurately capturing complex building shapes on the CarlaSC dataset. In addition, our method exhibits remarkable proficiency in generating the overall contours of roads and buildings, along with intricate details on the SemanticKITTI dataset. Tab. 1 provides a detailed compar" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 427, + 170 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 427, + 170 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 427, + 170 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 427, + 170 + ], + "type": "image", + "image_path": "f08e41b86a070b9dd13c86786f4d738f88654973ddfad42dc94cd245d130cf51.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 178, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 304, + 178, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 547, + 221 + ], + "type": "text", + "content": "Figure 4. Higher-resolution scene generation. Building upon our implicit decoder, higher-resolution scene " + }, + { + "bbox": [ + 304, + 178, + 547, + 221 + ], + "type": "inline_equation", + "content": "(1024\\times 1024\\times 128)" + }, + { + "bbox": [ + 304, + 178, + 547, + 221 + ], + "type": "text", + "content": " can be generated compared with a resolution of training dataset " + }, + { + "bbox": [ + 304, + 178, + 547, + 221 + ], + "type": "inline_equation", + "content": "(256\\times 256\\times 32)" + }, + { + "bbox": [ + 304, + 178, + 547, + 221 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 545, + 170 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 545, + 170 + ], + "type": "image", + "image_path": "20d615566abf8ce92c8dc0b6b718f12c308db3b0fa709aecf4dea9df4d99628c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 232, + 545, + 498 + ], + "blocks": [ + { + "bbox": [ + 308, + 232, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 498 + ], + "type": "image", + "image_path": "31e88e24c962750778056f56b2a984e75b8cc69ec42fda94cd144eabbbc23ef0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 502, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 535 + ], + "type": "text", + "content": "Figure 5. Scene inpainting of our method. The red boxes denote inpainting regions. (a) and (b) show our inpainting examples from reference images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "content": "ative evaluation using various metrics. Our model shows significant improvements in both the fidelity and diversity of the generated scenes. Moreover, our generated result is not tied to fixed resolution by means of the implicit neural representation, as depicted in Fig. 4. For additional results, please refer to the Supplementary Material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 635, + 488, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 488, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 488, + 647 + ], + "type": "text", + "content": "4.3. Applications of Triplane Diffusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Scene Inpainting. Fig. 5 presents the qualitative results of our inpainting, demonstrating its effectiveness in inpainting both small and large regions within a scene, while maintaining the coherence of 3D contexts. In detail, the second row of (a) and (b) illustrates the model's seamless removal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28342" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 79, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 49, + 79, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 49, + 79, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 79, + 545, + 171 + ], + "type": "image", + "image_path": "fc51c5c0950bad760add063b152dd4947ca9b0e8423036bcafff1a89aad4f9d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 173, + 196, + 249 + ], + "blocks": [ + { + "bbox": [ + 49, + 173, + 196, + 249 + ], + "lines": [ + { + "bbox": [ + 49, + 173, + 196, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 196, + 249 + ], + "type": "image", + "image_path": "c3f28b3c91df46837035313a79c64064d686caf70078b6f3ed4996796de6861a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 546, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 546, + 288 + ], + "type": "text", + "content": "Figure 6. Our scene outpainting results. The red box at the center of the figure in the first row represents the given scene for outpainting. The zoomed views of the outpainted scene are depicted in the second row. The outpainted scene is expanded from the given size of " + }, + { + "bbox": [ + 46, + 255, + 546, + 288 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 46, + 255, + 546, + 288 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 255, + 546, + 288 + ], + "type": "inline_equation", + "content": "1792 \\times 3328 \\times 32" + }, + { + "bbox": [ + 46, + 255, + 546, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 197, + 173, + 353, + 249 + ], + "blocks": [ + { + "bbox": [ + 197, + 173, + 353, + 249 + ], + "lines": [ + { + "bbox": [ + 197, + 173, + 353, + 249 + ], + "spans": [ + { + "bbox": [ + 197, + 173, + 353, + 249 + ], + "type": "image", + "image_path": "4f1712c7bfa47ec1efce17dac5fd8a5a35134609050cec64c712e36dc70750e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 173, + 544, + 249 + ], + "blocks": [ + { + "bbox": [ + 354, + 173, + 544, + 249 + ], + "lines": [ + { + "bbox": [ + 354, + 173, + 544, + 249 + ], + "spans": [ + { + "bbox": [ + 354, + 173, + 544, + 249 + ], + "type": "image", + "image_path": "5fb3650c0b537da7b60f831f511178f19f79db80c939678746cdda2f621a8db3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 297, + 286, + 399 + ], + "blocks": [ + { + "bbox": [ + 48, + 297, + 286, + 399 + ], + "lines": [ + { + "bbox": [ + 48, + 297, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 286, + 399 + ], + "type": "table", + "html": "
SSC InputMethodIoU ↑mIoU ↑
RGBMonoScene [9]37.1211.50
MonoScene + Ours50.4417.08
OccDepth [29]41.6012.84
OccDepth + Ours50.2016.79
Point CloudSSA-SC [51]58.2524.54
SSA-SC + Ours60.7125.58
SCPNet [49]50.2437.55
SCPNet + Ours59.2538.19
", + "image_path": "f639df30dcff6434d27938b80c54456d811edb0b756585100adeca64863262f0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 402, + 287, + 436 + ], + "lines": [ + { + "bbox": [ + 47, + 402, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 287, + 436 + ], + "type": "text", + "content": "Table 2. Quantitative results of refining SSC on SemanticKITTI validation set [6]. The results are based on the weights released by the authors on GitHub." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": "of vehicles, which harmonizes with the adjacent road. The third row demonstrates the insertion of new entities — a vehicle in (a) and a person in (b) — that are contextually congruent with the reference scene. The fourth row in (a) exemplifies the model's dual functionality in both modifying and adding vehicles within the scene. The fifth row in both columns underscores the model's proficiency in modifying scenes. Here, the model alters existing scene components, showcasing its ability to transform the overall ambiance of the scene. These results show our model's adeptness not just in object-level inpainting but also in scene-level inpainting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Scene Outpainting. Fig. 6 illustrates a generated outpainting city-level scene, extending a " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " scene to a substantial " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "1792 \\times 3328 \\times 32" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " landscape. Although our model was not designed to generate cityscapes, it demonstrates the capability of maintaining coherence over a large area. The roads are connected in a meaningful and varied way, and various objects such as buildings, cars, and people have been created around them. Please check the details of" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 298, + 545, + 471 + ], + "blocks": [ + { + "bbox": [ + 306, + 298, + 545, + 471 + ], + "lines": [ + { + "bbox": [ + 306, + 298, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 545, + 471 + ], + "type": "image", + "image_path": "689f5b59eb6ae1552c4116a3092abcdb3fd6d14e3a5f27725332f7b37af2ae17.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 477, + 545, + 500 + ], + "lines": [ + { + "bbox": [ + 306, + 477, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 545, + 500 + ], + "type": "text", + "content": "Figure 7. Semantic scene completion refinement. The SSC metrics are reported in the parentheses as (IoU, mIoU)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 515, + 384, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 384, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 384, + 526 + ], + "type": "text", + "content": "the scene in Fig. 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Semantic Scene Completion Refinement. In Fig. 7, there is a notable semantic and geometric discrepancy between scenes predicted by existing semantic scene completion (SSC) methods and their real scene counterparts. Our model helps to bridge this gap by employing a 3D scene prior, which is effectively modeled through our diffusion model. SCPNet [49] also attempts to learn geometric priors from a teacher model trained with merged sequential frames. Yet, its completion network is inclined toward conservative estimations, which frequently result in partially filled spaces. In the case of SSA-SC [51], which utilizes bird's-eye view features, it is sometimes inappropriate due to the label's inherent limitations of the bird's-eye perspective in capturing certain semantic details. This discrepancy" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28343" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 74, + 165, + 145 + ], + "blocks": [ + { + "bbox": [ + 49, + 74, + 165, + 145 + ], + "lines": [ + { + "bbox": [ + 49, + 74, + 165, + 145 + ], + "spans": [ + { + "bbox": [ + 49, + 74, + 165, + 145 + ], + "type": "image", + "image_path": "e45ee7569fe3152caba6e7c7df31901e4786ef791f0699cf51da81d0e164a95e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 146, + 112, + 156 + ], + "lines": [ + { + "bbox": [ + 101, + 146, + 112, + 156 + ], + "spans": [ + { + "bbox": [ + 101, + 146, + 112, + 156 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 284, + 145 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 284, + 145 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 284, + 145 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 284, + 145 + ], + "type": "image", + "image_path": "e05e11e232fa582a276e9d5dd1f5b96f50dda6f84e1345341f43b9d55752a058.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 146, + 231, + 156 + ], + "lines": [ + { + "bbox": [ + 220, + 146, + 231, + 156 + ], + "spans": [ + { + "bbox": [ + 220, + 146, + 231, + 156 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 159, + 286, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 286, + 204 + ], + "type": "text", + "content": "Figure 8. Semantic scene to RGB. (a) Semantic maps are rendered into a driving view from our generated scenes. (b) We generate images through ControlNet [56], an easily accessible imaged-to-image model, with generated semantic maps without shadows." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 213, + 286, + 273 + ], + "blocks": [ + { + "bbox": [ + 48, + 213, + 286, + 273 + ], + "lines": [ + { + "bbox": [ + 48, + 213, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 286, + 273 + ], + "type": "table", + "html": "
MethodFID ↓KID ↓IS ↑Prec ↑Rec ↑
Ours56.550.043.250.390.32
w/o PE59.140.043.210.300.31
xy-plane (w/o triplane)75.560.063.080.280.27
3D-volume (w/o triplane)136.500.151.960.140.12
", + "image_path": "8933d535b3a98fad44528baec0129a4af278c05d1c38e0787d8bf46b2e25e0ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 281, + 286, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 286, + 304 + ], + "type": "text", + "content": "Table 3. Ablation studies on scene generation. We ablate our model variants on SemanticKITTI [6]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 318, + 286, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 286, + 390 + ], + "type": "text", + "content": "issue is even more evident in RGB-based methods, such as MonoScene [9] and OccDepth [29], which often exhibit diminished sharpness when image features are projected into 3D space. While SSC models show variations from real-world data distributions, our model shows potential in aligning these more closely with reality, as shown in Fig. 7." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 390, + 286, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 450 + ], + "type": "text", + "content": "As indicated in Tab. 2, our SSC refinement process appears to offer improvements to all state-of-the-art SSC models. These preliminary results suggest our model's effectiveness in providing not just more accurate semantic segmentation but also a more complete scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 464, + 286, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 286, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 286, + 559 + ], + "type": "text", + "content": "Semantic Scene to RGB Image. As illustrated in Fig. 8, we conducted an image-to-image generation experiment. The semantic maps are rendered into a driving view without shadows; then, they are utilized as inputs for ControlNet [56]. The generated RGB images are geometrically and semantically plausible but display a synthetic quality since the pretrained ControlNet was not trained on actual autonomous driving datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 567, + 111, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 111, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 111, + 578 + ], + "type": "text", + "content": "4.4. Ablation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 586, + 286, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 286, + 621 + ], + "type": "text", + "content": "As shown in Tab. 3, we conducted ablation studies on our triplane diffusion model for scene generation, focusing on two key design elements:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 635, + 286, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 286, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 286, + 683 + ], + "type": "text", + "content": "Positional Embedding. Our variant excluded positional embedding PE, which generates high-frequency features critical for detailed scene reconstruction. Its absence resulted in lower performance across all metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Triplane Representation. We evaluated the effectiveness of the triplane representation for real-outdoor scene gen" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "eration. The triplane and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-plane enhance the generation quality compared with 3D features, while the " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-plane shows lower performance than triplanes. We suspect that the excessive factorization limits the representation capability of the " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-plane compared to the triplane." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 138, + 378, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 138, + 378, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 378, + 150 + ], + "type": "text", + "content": "4.5. Limitation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 156, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 361 + ], + "type": "text", + "content": "While our model demonstrates significant progress in the generation of 3D real-outdoor scenes, it inherently reflects the characteristics of its training data, as detailed in Sec. 4.1. This reliance introduces several limitations. One notable challenge is the model's difficulty in accurately depicting areas occluded from the sensor's viewpoint, such as the rear sides of buildings, often leading to their incomplete representation in our generated scenes. Moreover, since the dataset is captured from a driving view, there is an inherent shortfall in capturing the full height of buildings. This brings a partial representation of the vertical structure of buildings and other tall elements in the scenes. Another issue is the model tends to produce traces of moving objects stemming from dataset pre-processing that merges sequential frames. For future work, incorporating prior knowledge of the city into the model could yield better results, such as addressing occluded areas or building heights." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 370, + 378, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 378, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 378, + 382 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 390, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 629 + ], + "type": "text", + "content": "We have proposed a diffusion framework called SemCity for real-world outdoor scene generation. The seminal idea is to generate a scene by factorizing real-outdoor scenes into triplane representations. Our triplane representation outperforms traditional voxel-based approaches, producing scenes that are not only visually more appealing but also rich in semantic detail, effectively capturing the complexity of various objects within the scene. Ours is not constrained by fixed resolutions thanks to the incorporation of an implicit neural representation. We have further expanded the capabilities of our triplane diffusion model to several practical applications, including scene inpainting, scene outpainting, and semantic scene completion refinement. Specifically, by manipulating triplanes during the diffusion process, we achieve seamless inpainting and outpainting at both the object and scene levels. Ours is used to more closely align the scenes predicted by existing semantic scene completion methods with the actual data distribution using our learned 3D prior. We believe that our work provides a road map of real-outdoor scene generation to research communities." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by IITP (Institute of Information & Communications Technology Planning & Evaluation) and ITRC (Information Technology Research Center), funded by Korea government (MSIT) (RS-2023-00237965(2024) and IITP-2024-2020-0-01460). Prof. Sung-Eui Yoon is a corresponding author." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28344" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Ali Abbasi, Sinan Kalkan, and Yusuf Sahillioğlu. Deep 3d semantic scene extrapolation. The Visual Computer, 35:271-279, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "text", + "content": "[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12608-12618, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "text", + "content": "[3] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. Renderriffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12608-12618, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 302 + ], + "type": "text", + "content": "[4] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "type": "text", + "content": "[5] Luca Bartolomei, Lucas Teixeira, and Margarita Chli. Perception-aware path planning for uavs using semantic segmentation. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5808-5815, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "text", + "content": "[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 490 + ], + "type": "text", + "content": "[7] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 524 + ], + "type": "text", + "content": "[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "type": "text", + "content": "[9] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "text", + "content": "[10] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[11] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[12] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "type": "text", + "content": "[13] Xu Cui, Chenggang Lu, and Jinxiang Wang. 3d semantic map construction using improved orb-slam2 for mobile robot in edge computing environment. IEEE Access, 8:67179-67191, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[17] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "text", + "content": "[18] Changho Jo, Woobin Im, and Sung-Eui Yoon. In-n-out: Towards good initialization for inpainting and outpainting. arXiv preprint arXiv:2106.13953, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[19] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "text", + "content": "[20] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 488 + ], + "type": "text", + "content": "[21] Zeqiang Lai, Yuchen Duan, Jifeng Dai, Ziheng Li, Ying Fu, Hongsheng Li, Yu Qiao, and Wenhai Wang. Denoising diffusion semantic segmentation with mask prior modeling. arXiv preprint arXiv:2306.01721, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 522 + ], + "type": "text", + "content": "[22] Jumin Lee, Woobin Im, Sebin Lee, and Sung-Eui Yoon. Diffusion probabilistic models for scene-scale 3d categorical data. arXiv preprint arXiv:2301.00527, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "type": "text", + "content": "[23] Jiabao Lei, Jiapeng Tang, and Kui Jia. Generative scene synthesis via incremental view inpainting using rgbd diffusion models. arXiv preprint arXiv:2212.05993, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 557, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 547, + 601 + ], + "type": "text", + "content": "[24] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 547, + 668 + ], + "type": "text", + "content": "[25] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "type": "text", + "content": "[26] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28345" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "text", + "content": "[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[29] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 216, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 288, + 259 + ], + "type": "text", + "content": "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 259, + 288, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 259, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 288, + 314 + ], + "type": "text", + "content": "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 315, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 315, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 288, + 360 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "text", + "content": "[33] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 403, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 288, + 459 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 580 + ], + "type": "text", + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "text", + "content": "[37] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "text", + "content": "[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[39] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene com" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "pletion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 160, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 160, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 545, + 215 + ], + "type": "text", + "content": "[41] Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner. Diffuscene: Scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. arXiv preprint arXiv:2303.14207, 2023. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 216, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 216, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 545, + 259 + ], + "type": "text", + "content": "[42] Chaoqun Wang, Wenzheng Chi, Yuxiang Sun, and Max Q-H Meng. Autonomous robotic exploration by incremental road map construction. IEEE Transactions on Automation Science and Engineering, 16(4):1720-1731, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 259, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 259, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 545, + 304 + ], + "type": "text", + "content": "[43] Chaoqun Wang, Delong Zhu, Teng Li, Max Q-H Meng, and Clarence W De Silva. Efficient autonomous robotic exploration with semantic road map in indoor environments. IEEE Robotics and Automation Letters, 4(3):2989-2996, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 338 + ], + "type": "text", + "content": "[44] Xinying Wang, Dikai Xu, and Fangming Gu. 3d model inpainting based on 3d deep convolutional generative adversarial network. IEEE Access, 8:170355-170363, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "type": "text", + "content": "[45] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Pet-neus: Positional encoding tri-planes for neural surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12598–12607, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 449 + ], + "type": "text", + "content": "[46] Zhen Wang, Shijie Zhou, Jeong Joon Park, Despoina Paschalidou, Suya You, Gordon Wetzstein, Leonidas Guibas, and Achuta Kadambi. Alto: Alternating latent topologies for implicit 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 259-270, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 504 + ], + "type": "text", + "content": "[47] Joey Wilson, Jingyu Song, Yuewei Fu, Arthur Zhang, Andrew Capodieci, Paramsothy Jayakumar, Kira Barton, and Maani Ghaffari. *Motionsc: Data set and network for real-time semantic mapping in dynamic environments*. arXiv preprint arXiv:2203.07060, 2022. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "text", + "content": "[48] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 536, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 592 + ], + "type": "text", + "content": "[49] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 592, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 647 + ], + "type": "text", + "content": "[50] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[51] Xuemeng Yang, Hao Zou, Xin Kong, Tianxin Huang, Yong Liu, Wanlong Li, Feng Wen, and Hongbo Zhang. Semantic segmentation-assisted scene completion for lidar point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3555-3562. IEEE, 2021. 3, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28346" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 441 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[52] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[53] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4206-4216, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 295 + ], + "type": "text", + "content": "[55] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 341 + ], + "type": "text", + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 396 + ], + "type": "text", + "content": "[57] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10188-10198, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 288, + 441 + ], + "type": "text", + "content": "[58] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28347" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_content_list.json b/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..408c7a813ba31ae69d795f2ecde3ba5ce966a2e7 --- /dev/null +++ b/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_content_list.json @@ -0,0 +1,1607 @@ +[ + { + "type": "text", + "text": "Semantic Human Mesh Reconstruction with Textures", + "text_level": 1, + "bbox": [ + 214, + 130, + 756, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyu Zhan $^{1}$ , Jianxin Yang $^{1}$ , Yuanqi Li $^{1}$ , Jie Guo $^{1}$ , Yanwen Guo $^{1*}$ , and Wenping Wang $^{2}$ $^{1}$ Nanjing University $^{2}$ Texas A&M University", + "bbox": [ + 140, + 181, + 846, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhanxy, jianxin-yang, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn, wenping@tamu.edu https://zhanxy.xyz/projects/shert", + "bbox": [ + 89, + 220, + 898, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78a4fe5fe0d7ad43a6a35a99b377612be899fc9f4f38716f8cf85e7d3db20f2f.jpg", + "image_caption": [ + "Figure 1. Semantic Human mEsh Reconstruction with Textures (SHERT): (a) Given a target surface and the corresponding semantic guider, (b) SHERT reconstructs a detailed semantic model, which has stable UV unwrapping and skinning weights with high-quality triangle meshes. (c) It can either sample a texture map from the target surface or generate from the text prompts. (d) Based on our semantic representation, SHERT allows for high-precision facial reconstruction and animation of the body, face, and hands. (e) Moreover, SHERT is capable of inferring a fully textured avatar from a monocular image." + ], + "image_footnote": [], + "bbox": [ + 83, + 262, + 885, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 607, + 313, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of 3D detailed human mesh reconstruction has made significant progress in recent years. However, current methods still face challenges when used in industrial applications due to unstable results, low-quality meshes, and a lack of UV unwrapping and skinning weights. In this paper, we present SHERT, a novel pipeline that can reconstruct semantic human meshes with textures and high-precision details. SHERT applies semantic- and normal-based sampling between the detailed surface (e.g. mesh and SDF) and the corresponding SMPL-X model to obtain a partially sampled semantic mesh and then generates the complete semantic mesh by our specifically designed self-supervised completion and refinement networks. Using the complete semantic mesh as a basis, we employ a texture diffusion model to create human textures that are driven by both images and texts. Our reconstructed meshes have stable UV unwrapping, high-quality triangle meshes, and consistent semantic", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "information. The given SMPL-X model provides semantic information and shape priors, allowing SHERT to perform well even with incorrect and incomplete inputs. The semantic information also makes it easy to substitute and animate different body parts such as the face, body, and hands. Quantitative and qualitative experiments demonstrate that SHERT is capable of producing high-fidelity and robust semantic meshes that outperform state-of-the-art methods.", + "bbox": [ + 498, + 609, + 893, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 768, + 632, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recovering highly realistic details and textures of a human mesh from monocular images is crucial for various applications such as gaming, movies, cartoons, VR, virtual try-on, and digital avatars. Current approaches [1, 2, 4, 23, 24, 26, 39, 54, 55, 57, 61, 63, 64, 67, 70] primarily focus on the recovery of the geometric details that are associated with images, but their results are not yet practical for real-world ap", + "bbox": [ + 496, + 796, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "142", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plications. Recent advancements in parametric and explicit clothing models [12, 19, 29, 50, 62, 75] have shown promise in clothing reconstruction. However, they have limitations in accurately fitting different geometries and details. Implicit reconstruction approaches [23, 54, 55, 63, 64, 67, 70] excel at capturing clothing details but perform poorly in hands and face reconstruction. They may also generate incomplete and geometrically inseparable results.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, our goal is to reconstruct fully textured semantic human meshes through given detailed surfaces and corresponding semantic guiders. Our semantic human mesh is complete, animatable, and edit-friendly for both users and designers. It ensures that each vertex has deterministic semantic information and predefined skinning weights. This allows for easy substitution and animation of different body parts such as the face, body, and hands. The semantic information also guarantees stable and reasonable UV unwrapping, which is advantageous for editing and image-based texture generation.", + "bbox": [ + 75, + 213, + 472, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose SHERT, which generates a semantic human mesh from the detailed surface and its corresponding SMPL-X [46] model and optionally infers textures from images or colored surfaces. SHERT has four main processes. 1) Sampling, SHERT applies semantic and normal-based sampling to obtain a partially sampled semantic mesh based on the input detailed surface and SMPL-X model. We subdivide the original SMPL-X model to better capture the human details. 2) Completion, a self-supervised network is proposed to complete the partially sampled mesh. The network works in the 2D UV domain, which converts the 3D completion task into a 2D inpainting task. 3) Refinement, the origin image and front-back normal maps are used to enhance the geometry details. 4) Texture, our semantic human mesh is projected to the origin image in order to generate the partial texture map. Then we adapt the diffusion framework for text-driven partial texture inpainting and generation. The powerful generation ability of pre-trained diffusion model enables textures with rich clothing details and clear facial expressions.", + "bbox": [ + 75, + 381, + 472, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments have been conducted on both datasets and in-the-wild images. The quantitative and qualitative results show the robustness and superior performance of SHERT in reconstructing high-fidelity semantic human meshes and generating various high-resolution textures.", + "bbox": [ + 75, + 685, + 470, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this work include four-fold.", + "bbox": [ + 76, + 762, + 470, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SHERT, a novel pipeline to reconstruct high-quality semantic human meshes from the detailed 3D surfaces, represented either explicitly as meshes, or implicitly as signed distance field (SDF). SHERT is also capable of predicting robust and fully textured avatars with high-fidelity faces from monocular images.", + "- We propose a semantic- and normal-based sampling" + ], + "bbox": [ + 76, + 794, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method (SNS) and a self-supervised mesh completion network to achieve non-rigid 3D surface registration. The approach has the capability to process incomplete and inaccurate inputs by leveraging SMPL-X human priors.", + "bbox": [ + 509, + 90, + 893, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a self-supervised mesh refinement network working in the UV domain. It utilizes the images and front-back normal maps to improve the geometric mesh details.", + "- We use a diffusion model to infer high-resolution human textures from input images. The model can also accomplish text-driven texture inpainting and generation." + ], + "bbox": [ + 500, + 152, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 273, + 640, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Monocular 3D Human Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 300, + 831, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Clothed Human Reconstruction. Most of the current CNN-based monocular 3D clothed human shape estimation methods can be divided into two categories: explicit-based [1-4, 29, 73, 74] and implicit-based [5, 12, 23-26, 39, 54, 55, 57, 63, 64, 67, 70] approaches, based on their representations. Explicit-based approaches usually infer the 3D offsets on top of the parametric human model [41, 46, 49, 65]. However, these methods are difficult to apply to flexible human topologies and cannot capture details well. Implicit-based approaches have the advantage of representing arbitrary 3D clothed human shapes that are free from the limitations of parametric human models. There are also some works [13, 25, 39, 63, 64, 70] that mix the implicit and explicit representations to achieve the detailed and robust 3D clothed human reconstruction. But both implicit-based and mixed methods still cannot maintain the stability of the human body shape well, often resulting in problems such as blurring and missing local body parts in the predicted results. Some methods [12, 19, 29, 50, 62, 75] propose additional parametric or implicit clothing models to fit loose clothes. These clothing models often struggle to accurately capture the details present in the images. Nevertheless, they are still useful for generating rough approximations of the clothing. Despite considerable progress made in monocular human reconstruction capture, there are still certain constraints, especially in terms of reliability, availability, and user-friendliness.", + "bbox": [ + 496, + 324, + 893, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Texture Prediction. Previous works [5, 13, 24-26, 44, 54, 57, 59] have also focused on predicting image-based human textures. However, the current emphasis of implicit-based methods [5, 13, 24-26, 54] is still on predicting vertex colors, which cannot be easily converted into usable texture maps for industrial applications. DINAR [59] proposes neural textures for modeling human avatars and achieves texture inpainting by utilizing the diffusion framework. Nevertheless, the neural texture is highly coupled with human avatars, and the resolution is relatively low.", + "bbox": [ + 496, + 733, + 893, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face Reconstruction. In recent years, parameterized face", + "bbox": [ + 500, + 885, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "143", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dfcc98ab90beb3fbca00a919737c29c38021b7a01a830b65db5843c02659f70d.jpg", + "image_caption": [ + "Figure 2. Overview of SHERT for monocular image reconstruction. Given an RGB image $\\mathcal{I}$ , SHERT first infers the detailed mesh $\\mathcal{M}_t$ and corresponding sub-SMPLX model $\\mathcal{M}_x$ . It then applies SNS to obtain the partial semantic mesh $\\mathcal{M}_s$ (in Sec. 3.2). The Completion Net infers $\\mathcal{M}_c$ by filling the UV holes in $\\mathcal{M}_s$ (in Sec. 3.3). Image and normal maps are utilized to generate $\\mathcal{M}_r$ , which contains sharper geometry details (in Sec. 3.4). Finally, SHERT uses a diffusion model to achieve text-driven texture inpainting and generation (in Sec. 3.5)." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models [8, 9, 36, 47] have been widely used in monocular face reconstruction algorithms [14, 17, 18, 56, 76], contributing to the success of achieving realistic high-quality reconstruction results. In light of the incorporation of FLAME [36] into SMPL-X [46], we can utilize high-quality facial reconstruction results from previous works to enhance the accuracy and realism of the SMPL-X based human body reconstruction results.", + "bbox": [ + 75, + 436, + 472, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Non-rigid 3D Registration", + "text_level": 1, + "bbox": [ + 76, + 566, + 313, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Non-rigid 3D surface registration approaches [6, 10, 22, 30, 35, 45, 66, 68] usually compute a deformation that aligns a source surface with a target surface [15]. SHERT aims to transfer the semantic information of parametric human models to the corresponding detailed surfaces through registration. However, existing methods mainly focus on the precise registration of the source and target, while the target detailed human surfaces in real-world tasks often have incomplete and erroneous data.", + "bbox": [ + 75, + 590, + 468, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 739, + 168, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SHERT is capable of reconstructing a high-fidelity fully textured semantic human mesh based on a pair of detailed 3D surface and corresponding SMPL-X [46] model. Our results have high robustness and will not result in missing body parts in despite of the incomplete inputs. Furthermore, the semantic mesh ensures that each vertex has deterministic semantic information and predefined skinning weights, making it possible to replace and animate the human face, body, and hands. SHERT can also generate realistic hu", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "man textures from texts and images. These features make it easy for users and designers to use or further edit our results. To achieve this, SHERT first subdivides the SMPL-X model to improve the accuracy in capturing human details (in Sec. 3.1) and then applies the semantic- and normal-based sampling to obtain a partially sampled mesh (in Sec. 3.2). We complete the partial result using a self-supervised mesh completion network (in Sec. 3.3), and finally enhance the geometry details through refinement (in Sec. 3.4). In addition, SHERT uses a diffusion model for text-driven human texture inpainting and generation (in Sec. 3.5).", + "bbox": [ + 496, + 436, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Subdivided SMPL-X model", + "text_level": 1, + "bbox": [ + 498, + 617, + 746, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we use SMPL-X as the semantic guider. SMPL-X is a parameterized body model that combines SMPL [41] with the FLAME [37] head model and the MANO [52] hand model. It is parameterized with shape and pose, has 10, 475 vertices and 54 joints, including joints for the neck, jaw, eyeballs, and fingers. The SMPL-X model provides basic skinning weights and corresponding semantic information for each vertex.", + "bbox": [ + 496, + 642, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SHERT defines the sub-SMPLX based on SMPL-X since 10,475 vertices are not sufficient to accurately represent the details of the human body and clothing, such as facial details and clothing wrinkles. Taking into consideration both the expressiveness of the model and computational cost, we apply the mid-point subdivision algorithm twice on a standard SMPL-X model (with eyeballs removed) to obtain the sub-SMPLX with 149,921 vertices and 299,712 faces.", + "bbox": [ + 496, + 763, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "144", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/21d0bec775a267e2e481f6bb85b3278d5a9ca4a728769cd55e4f9bae5846c565.jpg", + "image_caption": [ + "Figure 3. Explicit and implicit sampling. a) SNS shoots a ray from the starting vertex along the vertex normal to locate the intersection point with the target surface. b) SNS takes samples at fixed intervals along the vertex normal ray and search for the point that is closest to the zero isosurface." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 457, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Semantic- and Normal-based Sampling (SNS)", + "text_level": 1, + "bbox": [ + 76, + 270, + 464, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our objective is to accurately predict the non-rigid deformation between a source surface and a target surface. Given a source surface $\\mathcal{M}_x$ , which is represented as a sub-SMPLX in our work, and the target surface $M_t$ , we learn a mapping function $D: \\mathcal{M}_x^{149921 \\times 3} \\to \\mathcal{M}_s^{149921 \\times 3}$ such that $\\mathcal{M}_s$ can be aligned with $\\mathcal{M}_t$ through the learned mapping in geometry, and ignoring the incorrect parts.", + "bbox": [ + 75, + 294, + 468, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we obtain a partially sampled mesh through a sampling scheme based on the vertex normals $N^{149921 \\times 3}$ . Starting from a point on $\\mathcal{M}_x$ , we search for the intersection point with the target surface $\\mathcal{M}_t$ along the vertex normal ray. We can extend the sampling scheme to implicit surfaces with constant step Ray Marching [28, 48], shown in Fig. 3 and Fig. 4.", + "bbox": [ + 75, + 401, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It should be noted that SNS does not always return satisfactory results since the ray may not intersect with the target surface. We need to locate and label the vertices that have failed to register. In addition, due to the large geometric differences and the possibility of incorrect alignment, the sampling scheme may retrieve incorrect results, which should also be detected. In order to remove incorrect points and triangle meshes from the sampled results, we calculate $\\theta$ (the angle between the normal vectors of the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), $s$ (the area ratio between the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), and $r$ (the edge ratio between the longest and shortest edges of the sampled triangle mesh) for each triangle mesh in the sampled result. We then perform mesh culling by removing low-quality triangle meshes that have at least one indicator exceeding the threshold among the three mentioned above. Finally, we perform a connectivity check on the processed sampled mesh and remove the sets with a number of connected triangle meshes less than $g$ . By default, we set $\\theta = 2$ , $s = 3$ , $r = 3$ , and $g = 500$ .", + "bbox": [ + 75, + 507, + 470, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We found that culling meshes only in the current pose would result in unreasonable deformations in human animation tasks. Therefore, SNS repeats the processing above in the canonical space. This ensures that we remove almost all unreasonable sampling results and obtain the partial se", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be20819802bf73bbd27d596bd85d23e3300a1b077f7db0c3d73a3faedc289789.jpg", + "image_caption": [ + "of SNS and the Marching Cubes algorithm [42] in the implicit field predicted by the multi-view PIFu [54]. The results show that our SNS and completion network correctly reconstruct the implicit field and obtain a smoother surface compared to Marching Cubes. mantic mesh $\\mathcal{M}_s$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 880, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Self-supervised Mesh Completion", + "text_level": 1, + "bbox": [ + 498, + 303, + 794, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Sec. 3.2, the vertices that failed to register have resulted in holes in $M_{s}$ . Currently, there are many mesh completion algorithms [11, 31-33, 40] that can fill in these mesh holes. However, these algorithms cannot be well adapted to our semantic reconstruction task due to the strong constraints on the number and relative positions of vertices in our representation in Sec. 3.1.", + "bbox": [ + 496, + 325, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we transform the partially sampled mesh into the UV domain [4, 16, 72] according to the semantic information. As a result, we convert the mesh completion task in 3D space into an inpainting task on a 2D image. Then, we design a self-supervised completion network that can fill in the holes of the partially sampled mesh. As shown in Fig. 5, by adding a random hole mask $\\mathcal{H}_r$ on the partial UV position map, we can obtain trainable pairs, which has a consistent distribution as the missing parts of the sampled mesh. The manually masked parts can provide the network with consistent supervision information as the input parts. This completion network is capable of generating robust meshes while maintaining both mesh quality and semantic consistency in the final results (refer to Fig. 6).", + "bbox": [ + 496, + 431, + 892, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further decrease the learning difficulties, we transform all incompletely sampled meshes to canonical space when facing the diverse range of poses and clothing styles. We believe that this approach has the potential to mitigate the challenges in problem-solving, as completion is no longer affected by pose and body shape. Meanwhile, since the sampled points are located on the normal ray of the sub-SMPLX vertices, we represent the deformation relative to sub-SMPLX as an offset based on the vertex normal. The transformed UV position map $\\overline{S}$ can be presented as:", + "bbox": [ + 496, + 643, + 893, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {d} = \\frac {S _ {\\text {s a m p l e}} - S _ {\\text {p o s e}}}{N _ {\\text {p o s e}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 804, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {S} = S _ {\\text {c a n o}} + N _ {\\text {c a n o}} \\cdot \\bar {d}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 847, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{\\text{sample}}$ denotes the partial UV position map of the sampled mesh. $S_{\\text{pose}}$ and $S_{\\text{cano}}$ refer to the UV position", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "145", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59c617ab77159f431c737eeada4b8c40254e1fd9ca69b9847bea1c9d24454f2b.jpg", + "image_caption": [ + "Figure 5. Completion network. The completion network transforms the result of SNS into canonical space and predicts the holes in the UV domain." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 450, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "maps of the original posed sub-SMPLX and the canonical sub-SMPLX respectively. $N_{\\text{pose}}$ and $N_{\\text{cano}}$ are the corresponding UV normal maps.", + "bbox": [ + 75, + 304, + 468, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complete UV position map $S$ and the input combined hole mask $\\mathcal{H}_{sup}$ are calculated by:", + "bbox": [ + 75, + 349, + 468, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {s u p} = \\mathcal {H} _ {r} \\cdot \\left(1 - \\mathcal {H} _ {o}\\right) + \\mathcal {H} _ {o}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 393, + 468, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS = S _ {c a n o} + N _ {c a n o} \\cdot \\left[ d \\cdot \\mathcal {H} _ {o} + \\bar {d} \\cdot \\left(1 - \\mathcal {H} _ {o}\\right) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 441, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_r$ and $\\mathcal{H}_o$ denote the randomly added hole mask and the hole mask generated by SNS respectively. $d\\in$ $\\mathbb{R}^{H\\times W\\times 3}$ is the estimated displacement UV map. The complete mesh $\\mathcal{M}_c$ is resampled from $S$ and then transformed to the original pose space.", + "bbox": [ + 75, + 470, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The primary loss used in our completion network is as follows:", + "bbox": [ + 75, + 546, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {p} = S _ {\\text {c a n o}} + N _ {\\text {c a n o}} \\cdot [ d \\cdot \\mathcal {H} _ {\\sup } + \\bar {d} \\cdot (1 - \\mathcal {H} _ {\\sup }) ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 602, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n p} = \\frac {\\left\\| \\left(S _ {p} - \\bar {S}\\right) \\cdot \\left(\\mathcal {H} _ {s u p} - \\mathcal {H} _ {o}\\right) \\right\\| _ {2} ^ {2}}{\\sum_ {i , j} \\left(\\mathcal {H} _ {s u p} - \\mathcal {H} _ {o}\\right) _ {i , j}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 641, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We find that completing the face, hands, and feet is much more difficult than other parts of the body, so we use the corresponding parts of sub-SMPLX for replacement. Furthermore, SHERT allows for the use of FLAME-based methods such as EMOCA [14] to replace the facial region in our completion results, enhancing the accuracy and realism of the facial details.", + "bbox": [ + 75, + 686, + 468, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Self-supervised Mesh Refinement", + "text_level": 1, + "bbox": [ + 76, + 801, + 367, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After completing the mesh, we have already reconstructed a semantic mesh with extremely high levels of detail. However, some geometric details may be lost during the SNS and completion process. Therefore, we design an additional self-supervised refinement network to further optimize the", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0182e7005cdd47679cad32888c6302e506948c9beb737252f3ee5d78528abb54.jpg", + "image_caption": [ + "Figure 6. The robustness of SNS and Completion Network. We present our complete meshes, which are reconstructed using the predictions of ECON [64] from in-the-wild images. Distinguishing itself from previous registration methods, SHERT has the capability to effectively process scenarios where the inputs are incomplete, inaccurate, or contain errors by leveraging SMPL-X human priors." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 883, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/25b4c87d12205b62882c53201b1fa3250df46a7927fdc10e17bed4fb3613be52.jpg", + "image_caption": [ + "Figure 7. Refinement network. The features extracted from the image and front-back normal maps are projected to the UV domain. These projected features are subsequently combined with the input UV position map to generate a refined mesh." + ], + "image_footnote": [], + "bbox": [ + 506, + 280, + 890, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mesh's details (e.g. cloth wrinkles, subtle deformations of body movements) using the image and normal maps. The input normal maps can be obtained either by rendering the scanned model or by using prediction networks. Our refinement network follows a U-Net [53] architecture and is capable of predicting a displacement UV map, denoted as $z \\in \\mathbb{R}^{H \\times W \\times 3}$ , in order to optimize the complete mesh $\\mathcal{M}_c$ . Since the input image and normal maps are not in the UV domain, we utilize a separate network to extract the image domain features and then project them to UV domain.", + "bbox": [ + 498, + 479, + 890, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to guide the network in learning the potential correspondence between the input features and mesh details, we apply Laplacian Smoothing [20] on the existing complete meshes, as shown in Fig. 7. The smoothed mesh $\\mathcal{M}_l$ is used as input to the network, and the original mesh $\\mathcal{M}_c$ is used for supervision.", + "bbox": [ + 498, + 630, + 890, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the image $\\mathcal{I}$ , front-view normal map $\\mathcal{N}_f$ , backview normal map $\\mathcal{N}_b$ , and the corresponding complete UV position map $S_l$ , the projected feature $F$ is typically represented as", + "bbox": [ + 498, + 720, + 890, + 780 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF = \\mathcal {P} \\left(\\mathcal {F} \\left(\\mathcal {I}, \\mathcal {N} _ {f}, \\mathcal {N} _ {b}\\right), S _ {l}, c\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 799, + 890, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c$ represents the camera parameters, $\\mathcal{F}$ denotes the image domain feature extraction network, and $\\mathcal{P}$ is the projection function that can project the image domain features to the UV domain according to the camera parameters and the 3D coordinates of points on $S_{l}$ .", + "bbox": [ + 498, + 824, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "146", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df8663b470ba65973de2791b24d60924a77afe1309f421bdb2c64c7267432914.jpg", + "image_caption": [ + "Figure 8. Texture diffusion. We use two strategies to finetune the diffusion model for text-driven texture inpainting and inpainting separately. During inference, instead of just using random noise, the encoded partial texture map and mask can be manually added to the latent features to preserve the original information of the input. Additionally, we use the super resolution method Real-esrgan [60], which can specifically optimize both the body and face, to further enhance the final outputs." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 519, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb73cb07c45c17ef49634b845d39cdc7973ff5c29d9d7afbb99b6d884e81328f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 88, + 883, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The UV position map $S_{r}$ of the refined mesh $\\mathcal{M}_r$ is given by", + "bbox": [ + 75, + 297, + 470, + 329 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS _ {r} = S _ {l} + N _ {l} \\cdot z, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 343, + 468, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z \\in \\mathbb{R}^{H \\times W \\times 3}$ represents the predicted displacement UV map of the refine network, $S_{l}$ and $N_{l}$ are the UV position map and normal UV map of $\\mathcal{M}_l$ respectively.", + "bbox": [ + 76, + 364, + 468, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The loss we used in refine network can be represented as", + "bbox": [ + 96, + 412, + 468, + 426 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s} = M S E \\left(z - \\frac {S _ {c} - S _ {l}}{N _ {l}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 436, + 468, + 467 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = M S E \\left(N _ {r} - N _ {c}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 482, + 468, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{c}$ is the UV position map of $\\mathcal{M}_c$ . $N_{r}$ and $N_{c}$ denote the normal UV maps of $\\mathcal{M}_r$ and $\\mathcal{M}_c$ respectively.", + "bbox": [ + 76, + 505, + 468, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, the refine network can be iteratively employed by users to enhance the mesh details according to their desired level (refer to Fig. 11).", + "bbox": [ + 75, + 536, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Text-driven Texture Inpainting and Generation", + "text_level": 1, + "bbox": [ + 76, + 590, + 468, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Sec. 3.4, all the reconstructed meshes share the same semantic information from SMPL-X. This means that when the models are unfolded to the UV domain, vertices with the same semantic information will be projected to a fixed UV position, resulting in a stable texture map that allows for painting and transferring textures. To perform texture inpainting on incomplete human body textures and generate high-quality human body textures driven by text, we fine-tune the Stable Diffusion model [51] with ControlNet [71], similar to the approach taken by Dinar [59].", + "bbox": [ + 75, + 613, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As our reconstructed semantic mesh closely matches the scanned model's geometry, we can use the ICP algorithm [7] to register the vertex colors from the scanned model onto our result. This enables us to convert the texture maps into the SMPL-X format. As shown in Fig. 8, we obtain a partial texture map from the input image based on the semantic mesh and camera parameters and hope to generate the invisible parts. ControlNet enable us to add conditional inputs (partial texture maps) to the generation of the diffusion", + "bbox": [ + 75, + 765, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "network at a low cost. We use two strategies to finetune the diffusion model, distinguished by whether to take the encoded partial texture map and mask as the additional inputs [27]. During the inference process, we can directly input partial texture maps into the ControlNet and generate various results by adding random noise. Alternatively, we can enhance the preservation of existing information in images by concatenating the encoded partial texture maps and masks with the latent features. Although Stable Diffusion has not worked with UV-parameterized images before, the well-designed UV parameterization keeps the shapes of the face, body, and limbs stable, ensuring a learnable space for the model. The texture map for the body and limbs focuses more on color and patterns rather than shapes, also resulting in excellent outcomes when cropped with a mask.", + "bbox": [ + 496, + 297, + 893, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 537, + 633, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Networks", + "text_level": 1, + "bbox": [ + 500, + 563, + 714, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training data. The completion and refinement networks are trained using the first 499 scans of THuman2.0 [69]. During the training of the completion network, we randomly choose one of the remaining 498 masks as a random hole mask. To enrich the inputs for the refinement network,", + "bbox": [ + 496, + 585, + 892, + 662 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/082195cd63daa45fda7ab5594c4630a86d2459bfa85ed6dfcc24835325fde955.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCAPE[43]THuman2.0[69]
P2S↓Chamfer↓Normal↓P2S↓Chamfer↓Normal↓
PIFu[54]*2.11371.65370.07552.54932.36400.1042
PIFuHD[55]3.78463.57870.10023.07723.18080.1207
PaMIR[70]*1.45201.22410.06101.54391.33110.1102
ICON[63]0.88550.86090.03471.03611.08740.0607
ECON[64]0.94030.93860.03741.13041.20810.0661
2K2K[23]†---2.53422.61650.1030
Ours (ICON-comp)0.8550↑0.8107↑0.03591.04591.0465↑0.0604↑
Ours (ICON-refine)0.86330.81120.03801.0442↑1.04680.0603↑
Ours (ECON-comp)0.8561↑0.8242↑0.03781.1255↑1.1420↑0.0672
Ours (ECON-refine)0.85810.8144↑0.03981.0630↑1.0430↑0.0649↑
", + "bbox": [ + 504, + 670, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation for monocular image reconstruction. We evaluate the performance of our completion results (comp) and refinement results (refine) by comparing them with state-of-the-art methods. * methods are re-implemented in [63] to ensure a fair comparison. † method has only been tested with human-facing-forward images. ↑ and ↑ indicate the improvement achieved through completion and refinement, respectively.", + "bbox": [ + 496, + 799, + 893, + 896 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "147", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/822937ec2201c677ed60b7c2329350140938f139eab8e7c52282e35ae6fec2a7.jpg", + "image_caption": [ + "Figure 9. Qualitative comparison for monocular image reconstruction on in-the-wild image. For each method, we present two views of the reconstructed results. SHERT demonstrates the ability to handle challenging poses while providing clear details of facial and hand geometry." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 893, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d63ebdbc0db070e4904e914a70449e7c70ff2515c57cedc403cf020ee56f5ae.jpg", + "image_caption": [ + "Figure 10. Qualitative comparison for registration on THuman2.0. We compare the registration quality of various methods including N-ICP [6], RPTS [35], SVR- $\\mathcal{L}_0$ [22], Fast-RNRR [68], DR [45], MDA [30] and ours SNS. The holes present in our result are the eliminated faces, as described in Sec. 3.2. The results indicate that SNS exhibits excellent performance in terms of model details, mesh quality, and registration robustness. The quantitative comparisons are shown in Tab. 2." + ], + "image_footnote": [], + "bbox": [ + 78, + 282, + 890, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we rotate the meshes every 60 degrees, resulting in a total of 2994 different orientations. We utilized the ICP [7] algorithm to transfer the color of the THuman2.0 scans to the vertices of our completed result, thereby generating 499 UV textures for training and obtaining 2994 visible UV masks from the rotated meshes. The complete UV textures and visible UV masks are randomly combined as inputs for our texture diffusion.", + "bbox": [ + 75, + 503, + 470, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Testing data. We conduct quantitative and qualitative evaluations on CAPE [43], THuman2.0, and in-the-wild images. We use CAPE-NFP [63] (100 samples with 3 viewpoints for each), and the last 27 subjects of THuman2.0 scans (6 viewpoints, each differing by 60 degrees).", + "bbox": [ + 75, + 628, + 468, + 705 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e996b6a68ea71bfcf26ae30de643caebc4139e97048c7f70e9d386bde7c553de.jpg", + "image_caption": [ + "Figure 11. The ablation results (with face substitution). We present the results after completion and refinement. With an increasing number of refinement iterations, the details of the mesh will be enhanced. Please zoom-in to see more details." + ], + "image_footnote": [], + "bbox": [ + 88, + 718, + 460, + 839 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee9d315e3ea598a21c809697dc6326846540710f3b8752a2bdada8ff81cf43f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGPUP2S↓Chamfer↓G-avg↑θ<30°↓Time
N-ICP [6]-0.2130.1630.50655.27m 23s
RPTS [35]-0.4880.3600.56548.31m 55s
SVR-L0 [22]-0.4040.2960.53153.51h 23m 32s
Fast-RNRR [68]-0.1150.0970.59731.11m 4s
DR [45]✓0.3390.3470.58147.716m 17s
MDA [30]✓0.6710.7310.58748.54m 48s
Ours (SNS)-0.1070.0780.72917.323s
Ours (Comp)✓0.1390.1670.66228.927s (23 + 4)
", + "bbox": [ + 503, + 502, + 893, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative evaluation for registration on THuman2.0. We test all the methods on the first subject of THuman2.0. Following the previous researches [21, 38], we adapt G-avg as a method for evaluating the mesh quality. We also report the metric $\\theta < 30^{\\circ}$ , which denotes the percentage of triangle meshes in the given mesh that have an angle less than 30 degrees. Additionally, we present the metrics for our complete mesh.", + "bbox": [ + 496, + 618, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Networks. The completion net and refinement net are both trained for 100 epochs with a learning rate of $1 \\times 10^{-6}$ . The resolutions of the input and output data, including the UV position maps, images, masks and front-back normal maps, are all $1024 \\times 1024 \\times 3$ . During inference, the ECON's predicted front-back normals $(512 \\times 512 \\times 3)$ are upsampled to $1024 \\times 1024 \\times 3$ using bilinear interpolation. The texture diffusion is trained for 1 epoch with a learning rate of $2 \\times 10^{-5}$ . The sampler of texture diffusion is DDIM [58]. We use 30 steps by default and infer the texture UV map with a resolution of $1024 \\times 1024 \\times 3$ . All the networks are trained on three NVIDIA RTX 3090 GPUs.", + "bbox": [ + 496, + 719, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "148", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75314ed9458309b8171b568893726d227123995c068b539f6f74501d22424320.jpg", + "image_caption": [ + "Figure 12. Qualitative comparison for texture prediction on in-the-wild image. We display the front and back view rendering results for each method. Since PIFu [54] and PaMIR [70] predict vertex colors, we only exhibit the texture maps of DINAR [59] and SHERT (ours). Please zoom-in to see more details." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 286, + 197, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative comparisons. We conduct quantitative comparisons with mainstream state-of-the-art monocular image reconstruction approaches in Tab. 1. As in previous work [23, 55, 63, 64], we report the point-to-surface Euclidean distance (P2S, cm), the Chamfer Distance (cm), and the Normals difference (L2). To ensure a fair comparison, $\\mathrm{PIFu}^*$ [54] and $\\mathrm{PaMIR}^*$ [70] are re-implemented and restrained on THuman2.0, using the same settings as ICON [63]. The ground-truth SMPL/SMPL-X models are provided for evaluation. However, PIFu [54], PIFuHD [55] and 2K2K [23] do not utilize the parametric body priors, which may result in subpar performance. In Tab. 1, \"ICON-comp\" refers to the completion result achieved by leveraging ICON's [63] prediction, while \"ECON-refine\" denotes the refinement mesh obtained using ICON's [64] result and the predicted front-back normal maps. Additionally, we evaluate the registration quality of SNS against state-of-the-art non-rigid registration methods, as presented in Tab. 2", + "bbox": [ + 75, + 310, + 468, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative comparisons. We demonstrate a comparison between SHERT and state-of-the-art methods using in-the-wild images, with a focus on monocular image reconstruction (refer to Fig. 9) and texture prediction (refer to Fig. 12). Additionally, we compare the registration quality of our SNS with state-of-the-art non-rigid registration approaches on Thuman2.0 (refer to Fig. 10).", + "bbox": [ + 75, + 583, + 468, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Limitations", + "text_level": 1, + "bbox": [ + 76, + 699, + 202, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the geometric limitations of SMPL-X, SHERT performs weaker in reconstructing loose clothing, shoes and hair compared to implicit-based reconstruction methods. It is also difficult to ensure consistent results of texture diffusion at the seams of UVs. See more in SupMat.", + "bbox": [ + 75, + 723, + 468, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Applications", + "text_level": 1, + "bbox": [ + 76, + 814, + 207, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SHERT uses the skinning weights from SMPL-X to enable animated poses, expressions, and gestures on the reconstructed mesh through LBS [34] (refer to Fig. 13). It allows for both global texture repainting (refer to Fig. 14)", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a45b655e8185b81bcb2405d876487bd2ece54e322bb0c43fc70b30c7a281ff49.jpg", + "image_caption": [ + "Figure 13. Animation results. Please zoom-in to see more details." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 875, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/835cacc056a247a8e631bbeef29df6b0baf5d38baa09813f984f648f37e3efc9.jpg", + "image_caption": [ + "Figure 14. Global texture repainting. SHERT can repaint the texture through text prompts. Please zoom-in to see more details." + ], + "image_footnote": [], + "bbox": [ + 506, + 196, + 890, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20372b62f2cc4547d835a8c642eba666764096966a22dfacc7bc14a12a9cef84.jpg", + "image_caption": [ + "Figure 15. Localized texture repainting. SHERT can repaint the masked area through text prompts. Please zoom-in to see more details." + ], + "image_footnote": [], + "bbox": [ + 514, + 378, + 879, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and the option for users to provide custom masks and text prompts for localized texture repainting (refer to Fig. 15).", + "bbox": [ + 498, + 566, + 890, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 609, + 617, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose SHERT, which reconstructs a fully textured semantic human avatar from a detailed surface or a monocular image. It takes advantage of the geometric details of the target surface, along with semantic information and prior knowledge of the semantic guider. The reconstructed results have high-fidelity clothing details, high-quality triangle meshes, clear facial features, and complete hands geometry. SHERT is also capable of generating high-resolution texture maps with stable UV unwrapping. This approach bridges existing monocular reconstruction work and downstream industrial applications, and we believe it can promote the development of human avatars.", + "bbox": [ + 496, + 635, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 830, + 658, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (62032011) and the Natural Science Foundation of Jiangsu Province (No. BK20211147).", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "149", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2018. 1, 2", + "[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In IEEE Conf. 3D Vis.(3DV), 2018. 1", + "[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single rgb camera. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019.", + "[4] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4", + "[5] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2", + "[6] Brian Amberg, Sami Romdhani, and Thomas Vetter. Optimal step nonrigid icp algorithms for surface registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2007. 3, 7", + "[7] P.J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 1992. 6, 7", + "[8] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), 1999. 3", + "[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Trans. Vis. Comput. Graph.(TVCG), 2014. 3", + "[10] Kunyao Chen, Fei Yin, Bang Du, Baichuan Wu, and Truong Q. Nguyen. Efficient registration for human surfaces via isometric regularization on embedded deformation. IEEE Trans. Vis. Comput. Graph.(TVCG), 2022. 3", + "[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 4", + "[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 2", + "[13] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2", + "[14] Radek Danecek, Michael J. Black, and Timo Bolkart. EMOCA: Emotion driven monocular face capture and animation. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 3, 5" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Bailin Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. Computer Graphics Forum (Eurographics 2022 State-of-the-Art Reports), 2022. 3", + "[16] Jiankang Deng, Shiyang Cheng, Niannan Xue, Yuxiang Zhou, and Stefanos Zafeiriou. Uv-gan: Adversarial facial uv map completion for pose-invariant face recognition. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2018. 4", + "[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3", + "[18] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. ACM Trans. Graph.(TOG), 2021. 3", + "[19] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Proceedings, 2022. 2", + "[20] David A Field. Laplacian smoothing and delaunay triangulations. Communications in applied numerical methods, 1988. 5", + "[21] Pascal J Frey and Houman Borouchaki. Surface mesh quality evaluation. International journal for numerical methods in engineering, 1999. 7", + "[22] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Int. Conf. Comput. Vis.(ICCV), 2015. 3, 7", + "[23] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 6, 8", + "[24] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. $\\mathrm{ARCH} + +$ : Animation-ready clothed human reconstruction revisited. In Int. Conf. Comput. Vis.(ICCV), 2021. 1, 2", + "[25] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2", + "[26] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable reconstruction of clothed humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2", + "[27] Huggingface. Stable-diffusion-inpainting, 2022. https://huggingface.co/runwayml/stable-diffusion-inpainting.6", + "[28] Henrik Wann Jensen. Realistic image synthesis using photon mapping. AK Peters/crc Press, 2001. 4", + "[29] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In Eur. Conf. Comput. Vis.(ECCV), 2020. 2", + "[30] Yucheol Jung, Hyomin Kim, Gyeongha Hwang, Seung-Hwan Baek, and Seungyong Lee. Mesh density adapta" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "150", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion for template-based shape reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 7", + "[31] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph.(TOG), 2013. 4", + "[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006.", + "[33] Vladislav Kraevoy and Alla Sheffer. Template-based mesh completion. In Symposium on Geometry Processing, 2005. 4", + "[34] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of SIGGRAPH 2000, 2000. 8", + "[35] Kun Li, Jingyu Yang, Yu-Kun Lai, and Daoliang Guo. Robust non-rigid registration with reweighted position and transformation sparsity. IEEE Trans. Vis. Comput. Graph.(TVCG), 2018. 3, 7", + "[36] Tianye Li, Timo Bolkart, Michael Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph.(TOG), 2017. 3", + "[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Trans. Graph.(TOG), 2017. 3", + "[38] Yuanqi Li, Jianwei Guo, Xinran Yang, Shun Liu, Jie Guo, Xiaopeng Zhang, and Yanwen Guo. Deep point cloud simplification for high-quality surface reconstruction. arXiv preprint arXiv:2203.09088, 2022. 7", + "[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2", + "[40] Peter Liepa. Filling holes in meshes. In Proceedings of the 2003 Eurographics/ACM SIGGRAPH symposium on Geometry processing, 2003. 4", + "[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graph.(TOG), 2015. 2, 3", + "[42] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. Seminal graphics: pioneering efforts that shaped the field, 1998. 4", + "[43] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 6, 7", + "[44] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2", + "[45] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph.(TOG), 2021. 3, 7", + "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2019. 2, 3", + "[47] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, 2009. 3", + "[48] Ken Perlin and Eric M Hoffert. Hypertexture. In Proceedings of the 16th annual conference on Computer graphics and interactive techniques, 1989. 4", + "[49] Leonid Pishchulin, Stefanie Wuhrer, Thomas Helten, Christian Theobalt, and Bernt Schiele. Building statistical shape spaces for 3d human modeling. Pattern Recognition, 2017. 2", + "[50] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2022. 6", + "[52] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph.(TOG), 2017. 3", + "[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention(MICCAI), 2015, 2015. 5", + "[54] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4, 6, 8", + "[55] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2, 6, 8", + "[56] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Ming-min Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In Eur. Conf. Comput. Vis.(ECCV), 2020. 3", + "[57] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X- avatar: Expressive human avatars. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2", + "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 7", + "[59] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Int. Conf. Comput. Vis.(ICCV), 2023. 2, 6, 8", + "[60] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Int. Conf. Comput. Vis.(ICCV), 2021. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "151", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In IEEE Conf. 3D Vis.(3DV), 2020. 1", + "[62] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph.(TOG), 2021. 2", + "[63] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 1, 2, 6, 7, 8", + "[64] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 5, 6, 8", + "[65] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3d human shape and articulated pose models. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2", + "[66] Jingyu Yang, Ke Li, Kun Li, and Yu-Kun Lai. Sparse non-rigid registration of 3d shapes. Computer Graphics Forum, 2015. 3", + "[67] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-if: Uncertainty-aware human digitization via implicit distribution field. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2", + "[68] Yuxin Yao, Bailin Deng, Weiwei Xu, and Juyong Zhang. Quasi-newton solver for robust non-rigid registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 3, 7", + "[69] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 6", + "[70] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 1, 2, 6, 8", + "[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis.(ICCV), 2023. 6", + "[72] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 4", + "[73] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019. 2", + "[74] Hao Zhu, Xinxin Zuo, Haotian Yang, Sen Wang, Xun Cao, and Ruigang Yang. Detailed avatar recovery from single image. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 2", + "[75] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity gar" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ment mesh reconstruction from single images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2", + "[76] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2016. 3" + ], + "bbox": [ + 503, + 92, + 890, + 162 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "152", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_model.json b/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cc40530c4f600572b4c2205d02e6d1f3f38922f2 --- /dev/null +++ b/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.131, + 0.757, + 0.151 + ], + "angle": 0, + "content": "Semantic Human Mesh Reconstruction with Textures" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.183, + 0.847, + 0.22 + ], + "angle": 0, + "content": "Xiaoyu Zhan\\(^{1}\\), Jianxin Yang\\(^{1}\\), Yuanqi Li\\(^{1}\\), Jie Guo\\(^{1}\\), Yanwen Guo\\(^{1*}\\), and Wenping Wang\\(^{2}\\) \n\\(^{1}\\) Nanjing University \\(^{2}\\) Texas A&M University" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.899, + 0.256 + ], + "angle": 0, + "content": "{zhanxy, jianxin-yang, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn, wenping@tamu.edu https://zhanxy.xyz/projects/shert" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.263, + 0.887, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.52, + 0.893, + 0.588 + ], + "angle": 0, + "content": "Figure 1. Semantic Human mEsh Reconstruction with Textures (SHERT): (a) Given a target surface and the corresponding semantic guider, (b) SHERT reconstructs a detailed semantic model, which has stable UV unwrapping and skinning weights with high-quality triangle meshes. (c) It can either sample a texture map from the target surface or generate from the text prompts. (d) Based on our semantic representation, SHERT allows for high-precision facial reconstruction and animation of the body, face, and hands. (e) Moreover, SHERT is capable of inferring a fully textured avatar from a monocular image." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.608, + 0.314, + 0.625 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.904 + ], + "angle": 0, + "content": "The field of 3D detailed human mesh reconstruction has made significant progress in recent years. However, current methods still face challenges when used in industrial applications due to unstable results, low-quality meshes, and a lack of UV unwrapping and skinning weights. In this paper, we present SHERT, a novel pipeline that can reconstruct semantic human meshes with textures and high-precision details. SHERT applies semantic- and normal-based sampling between the detailed surface (e.g. mesh and SDF) and the corresponding SMPL-X model to obtain a partially sampled semantic mesh and then generates the complete semantic mesh by our specifically designed self-supervised completion and refinement networks. Using the complete semantic mesh as a basis, we employ a texture diffusion model to create human textures that are driven by both images and texts. Our reconstructed meshes have stable UV unwrapping, high-quality triangle meshes, and consistent semantic" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.895, + 0.732 + ], + "angle": 0, + "content": "information. The given SMPL-X model provides semantic information and shape priors, allowing SHERT to perform well even with incorrect and incomplete inputs. The semantic information also makes it easy to substitute and animate different body parts such as the face, body, and hands. Quantitative and qualitative experiments demonstrate that SHERT is capable of producing high-fidelity and robust semantic meshes that outperform state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.769, + 0.633, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Recovering highly realistic details and textures of a human mesh from monocular images is crucial for various applications such as gaming, movies, cartoons, VR, virtual try-on, and digital avatars. Current approaches [1, 2, 4, 23, 24, 26, 39, 54, 55, 57, 61, 63, 64, 67, 70] primarily focus on the recovery of the geometric details that are associated with images, but their results are not yet practical for real-world ap" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "142" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "plications. Recent advancements in parametric and explicit clothing models [12, 19, 29, 50, 62, 75] have shown promise in clothing reconstruction. However, they have limitations in accurately fitting different geometries and details. Implicit reconstruction approaches [23, 54, 55, 63, 64, 67, 70] excel at capturing clothing details but perform poorly in hands and face reconstruction. They may also generate incomplete and geometrically inseparable results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.473, + 0.382 + ], + "angle": 0, + "content": "In this work, our goal is to reconstruct fully textured semantic human meshes through given detailed surfaces and corresponding semantic guiders. Our semantic human mesh is complete, animatable, and edit-friendly for both users and designers. It ensures that each vertex has deterministic semantic information and predefined skinning weights. This allows for easy substitution and animation of different body parts such as the face, body, and hands. The semantic information also guarantees stable and reasonable UV unwrapping, which is advantageous for editing and image-based texture generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.473, + 0.685 + ], + "angle": 0, + "content": "Specifically, we propose SHERT, which generates a semantic human mesh from the detailed surface and its corresponding SMPL-X [46] model and optionally infers textures from images or colored surfaces. SHERT has four main processes. 1) Sampling, SHERT applies semantic and normal-based sampling to obtain a partially sampled semantic mesh based on the input detailed surface and SMPL-X model. We subdivide the original SMPL-X model to better capture the human details. 2) Completion, a self-supervised network is proposed to complete the partially sampled mesh. The network works in the 2D UV domain, which converts the 3D completion task into a 2D inpainting task. 3) Refinement, the origin image and front-back normal maps are used to enhance the geometry details. 4) Texture, our semantic human mesh is projected to the origin image in order to generate the partial texture map. Then we adapt the diffusion framework for text-driven partial texture inpainting and generation. The powerful generation ability of pre-trained diffusion model enables textures with rich clothing details and clear facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.472, + 0.763 + ], + "angle": 0, + "content": "Extensive experiments have been conducted on both datasets and in-the-wild images. The quantitative and qualitative results show the robustness and superior performance of SHERT in reconstructing high-fidelity semantic human meshes and generating various high-resolution textures." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.763, + 0.472, + 0.793 + ], + "angle": 0, + "content": "In summary, the main contributions of this work include four-fold." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.795, + 0.472, + 0.886 + ], + "angle": 0, + "content": "- We introduce SHERT, a novel pipeline to reconstruct high-quality semantic human meshes from the detailed 3D surfaces, represented either explicitly as meshes, or implicitly as signed distance field (SDF). SHERT is also capable of predicting robust and fully textured avatars with high-fidelity faces from monocular images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- We propose a semantic- and normal-based sampling" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.795, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "method (SNS) and a self-supervised mesh completion network to achieve non-rigid 3D surface registration. The approach has the capability to process incomplete and inaccurate inputs by leveraging SMPL-X human priors." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.894, + 0.211 + ], + "angle": 0, + "content": "- We present a self-supervised mesh refinement network working in the UV domain. It utilizes the images and front-back normal maps to improve the geometric mesh details." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.212, + 0.894, + 0.258 + ], + "angle": 0, + "content": "- We use a diffusion model to infer high-resolution human textures from input images. The model can also accomplish text-driven texture inpainting and generation." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.153, + 0.894, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.275, + 0.642, + 0.29 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.833, + 0.316 + ], + "angle": 0, + "content": "2.1. Monocular 3D Human Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.733 + ], + "angle": 0, + "content": "Clothed Human Reconstruction. Most of the current CNN-based monocular 3D clothed human shape estimation methods can be divided into two categories: explicit-based [1-4, 29, 73, 74] and implicit-based [5, 12, 23-26, 39, 54, 55, 57, 63, 64, 67, 70] approaches, based on their representations. Explicit-based approaches usually infer the 3D offsets on top of the parametric human model [41, 46, 49, 65]. However, these methods are difficult to apply to flexible human topologies and cannot capture details well. Implicit-based approaches have the advantage of representing arbitrary 3D clothed human shapes that are free from the limitations of parametric human models. There are also some works [13, 25, 39, 63, 64, 70] that mix the implicit and explicit representations to achieve the detailed and robust 3D clothed human reconstruction. But both implicit-based and mixed methods still cannot maintain the stability of the human body shape well, often resulting in problems such as blurring and missing local body parts in the predicted results. Some methods [12, 19, 29, 50, 62, 75] propose additional parametric or implicit clothing models to fit loose clothes. These clothing models often struggle to accurately capture the details present in the images. Nevertheless, they are still useful for generating rough approximations of the clothing. Despite considerable progress made in monocular human reconstruction capture, there are still certain constraints, especially in terms of reliability, availability, and user-friendliness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Texture Prediction. Previous works [5, 13, 24-26, 44, 54, 57, 59] have also focused on predicting image-based human textures. However, the current emphasis of implicit-based methods [5, 13, 24-26, 54] is still on predicting vertex colors, which cannot be easily converted into usable texture maps for industrial applications. DINAR [59] proposes neural textures for modeling human avatars and achieves texture inpainting by utilizing the diffusion framework. Nevertheless, the neural texture is highly coupled with human avatars, and the resolution is relatively low." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Face Reconstruction. In recent years, parameterized face" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "143" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Overview of SHERT for monocular image reconstruction. Given an RGB image \\(\\mathcal{I}\\), SHERT first infers the detailed mesh \\(\\mathcal{M}_t\\) and corresponding sub-SMPLX model \\(\\mathcal{M}_x\\). It then applies SNS to obtain the partial semantic mesh \\(\\mathcal{M}_s\\) (in Sec. 3.2). The Completion Net infers \\(\\mathcal{M}_c\\) by filling the UV holes in \\(\\mathcal{M}_s\\) (in Sec. 3.3). Image and normal maps are utilized to generate \\(\\mathcal{M}_r\\), which contains sharper geometry details (in Sec. 3.4). Finally, SHERT uses a diffusion model to achieve text-driven texture inpainting and generation (in Sec. 3.5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.473, + 0.558 + ], + "angle": 0, + "content": "models [8, 9, 36, 47] have been widely used in monocular face reconstruction algorithms [14, 17, 18, 56, 76], contributing to the success of achieving realistic high-quality reconstruction results. In light of the incorporation of FLAME [36] into SMPL-X [46], we can utilize high-quality facial reconstruction results from previous works to enhance the accuracy and realism of the SMPL-X based human body reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.315, + 0.584 + ], + "angle": 0, + "content": "2.2. Non-rigid 3D Registration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Non-rigid 3D surface registration approaches [6, 10, 22, 30, 35, 45, 66, 68] usually compute a deformation that aligns a source surface with a target surface [15]. SHERT aims to transfer the semantic information of parametric human models to the corresponding detailed surfaces through registration. However, existing methods mainly focus on the precise registration of the source and target, while the target detailed human surfaces in real-world tasks often have incomplete and erroneous data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.169, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "SHERT is capable of reconstructing a high-fidelity fully textured semantic human mesh based on a pair of detailed 3D surface and corresponding SMPL-X [46] model. Our results have high robustness and will not result in missing body parts in despite of the incomplete inputs. Furthermore, the semantic mesh ensures that each vertex has deterministic semantic information and predefined skinning weights, making it possible to replace and animate the human face, body, and hands. SHERT can also generate realistic hu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.606 + ], + "angle": 0, + "content": "man textures from texts and images. These features make it easy for users and designers to use or further edit our results. To achieve this, SHERT first subdivides the SMPL-X model to improve the accuracy in capturing human details (in Sec. 3.1) and then applies the semantic- and normal-based sampling to obtain a partially sampled mesh (in Sec. 3.2). We complete the partial result using a self-supervised mesh completion network (in Sec. 3.3), and finally enhance the geometry details through refinement (in Sec. 3.4). In addition, SHERT uses a diffusion model for text-driven human texture inpainting and generation (in Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.618, + 0.748, + 0.633 + ], + "angle": 0, + "content": "3.1. Subdivided SMPL-X model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In this work, we use SMPL-X as the semantic guider. SMPL-X is a parameterized body model that combines SMPL [41] with the FLAME [37] head model and the MANO [52] hand model. It is parameterized with shape and pose, has 10, 475 vertices and 54 joints, including joints for the neck, jaw, eyeballs, and fingers. The SMPL-X model provides basic skinning weights and corresponding semantic information for each vertex." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "SHERT defines the sub-SMPLX based on SMPL-X since 10,475 vertices are not sufficient to accurately represent the details of the human body and clothing, such as facial details and clothing wrinkles. Taking into consideration both the expressiveness of the model and computational cost, we apply the mid-point subdivision algorithm twice on a standard SMPL-X model (with eyeballs removed) to obtain the sub-SMPLX with 149,921 vertices and 299,712 faces." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.458, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.195, + 0.473, + 0.264 + ], + "angle": 0, + "content": "Figure 3. Explicit and implicit sampling. a) SNS shoots a ray from the starting vertex along the vertex normal to locate the intersection point with the target surface. b) SNS takes samples at fixed intervals along the vertex normal ray and search for the point that is closest to the zero isosurface." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.271, + 0.465, + 0.287 + ], + "angle": 0, + "content": "3.2. Semantic- and Normal-based Sampling (SNS)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.401 + ], + "angle": 0, + "content": "Our objective is to accurately predict the non-rigid deformation between a source surface and a target surface. Given a source surface \\(\\mathcal{M}_x\\), which is represented as a sub-SMPLX in our work, and the target surface \\(M_t\\), we learn a mapping function \\(D: \\mathcal{M}_x^{149921 \\times 3} \\to \\mathcal{M}_s^{149921 \\times 3}\\) such that \\(\\mathcal{M}_s\\) can be aligned with \\(\\mathcal{M}_t\\) through the learned mapping in geometry, and ignoring the incorrect parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Specifically, we obtain a partially sampled mesh through a sampling scheme based on the vertex normals \\( N^{149921 \\times 3} \\). Starting from a point on \\( \\mathcal{M}_x \\), we search for the intersection point with the target surface \\( \\mathcal{M}_t \\) along the vertex normal ray. We can extend the sampling scheme to implicit surfaces with constant step Ray Marching [28, 48], shown in Fig. 3 and Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.825 + ], + "angle": 0, + "content": "It should be noted that SNS does not always return satisfactory results since the ray may not intersect with the target surface. We need to locate and label the vertices that have failed to register. In addition, due to the large geometric differences and the possibility of incorrect alignment, the sampling scheme may retrieve incorrect results, which should also be detected. In order to remove incorrect points and triangle meshes from the sampled results, we calculate \\(\\theta\\) (the angle between the normal vectors of the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), \\(s\\) (the area ratio between the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), and \\(r\\) (the edge ratio between the longest and shortest edges of the sampled triangle mesh) for each triangle mesh in the sampled result. We then perform mesh culling by removing low-quality triangle meshes that have at least one indicator exceeding the threshold among the three mentioned above. Finally, we perform a connectivity check on the processed sampled mesh and remove the sets with a number of connected triangle meshes less than \\(g\\). By default, we set \\(\\theta = 2\\), \\(s = 3\\), \\(r = 3\\), and \\(g = 500\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We found that culling meshes only in the current pose would result in unreasonable deformations in human animation tasks. Therefore, SNS repeats the processing above in the canonical space. This ensures that we remove almost all unreasonable sampling results and obtain the partial se" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.088, + 0.882, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.218, + 0.894, + 0.297 + ], + "angle": 0, + "content": "of SNS and the Marching Cubes algorithm [42] in the implicit field predicted by the multi-view PIFu [54]. The results show that our SNS and completion network correctly reconstruct the implicit field and obtain a smoother surface compared to Marching Cubes. mantic mesh \\(\\mathcal{M}_s\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.304, + 0.795, + 0.321 + ], + "angle": 0, + "content": "3.3. Self-supervised Mesh Completion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.432 + ], + "angle": 0, + "content": "In Sec. 3.2, the vertices that failed to register have resulted in holes in \\( M_{s} \\). Currently, there are many mesh completion algorithms [11, 31-33, 40] that can fill in these mesh holes. However, these algorithms cannot be well adapted to our semantic reconstruction task due to the strong constraints on the number and relative positions of vertices in our representation in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Therefore, we transform the partially sampled mesh into the UV domain [4, 16, 72] according to the semantic information. As a result, we convert the mesh completion task in 3D space into an inpainting task on a 2D image. Then, we design a self-supervised completion network that can fill in the holes of the partially sampled mesh. As shown in Fig. 5, by adding a random hole mask \\(\\mathcal{H}_r\\) on the partial UV position map, we can obtain trainable pairs, which has a consistent distribution as the missing parts of the sampled mesh. The manually masked parts can provide the network with consistent supervision information as the input parts. This completion network is capable of generating robust meshes while maintaining both mesh quality and semantic consistency in the final results (refer to Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.894, + 0.796 + ], + "angle": 0, + "content": "To further decrease the learning difficulties, we transform all incompletely sampled meshes to canonical space when facing the diverse range of poses and clothing styles. We believe that this approach has the potential to mitigate the challenges in problem-solving, as completion is no longer affected by pose and body shape. Meanwhile, since the sampled points are located on the normal ray of the sub-SMPLX vertices, we represent the deformation relative to sub-SMPLX as an offset based on the vertex normal. The transformed UV position map \\(\\overline{S}\\) can be presented as:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.805, + 0.891, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\bar {d} = \\frac {S _ {\\text {s a m p l e}} - S _ {\\text {p o s e}}}{N _ {\\text {p o s e}}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.848, + 0.891, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\bar {S} = S _ {\\text {c a n o}} + N _ {\\text {c a n o}} \\cdot \\bar {d}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( S_{\\text{sample}} \\) denotes the partial UV position map of the sampled mesh. \\( S_{\\text{pose}} \\) and \\( S_{\\text{cano}} \\) refer to the UV position" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "145" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.452, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.47, + 0.295 + ], + "angle": 0, + "content": "Figure 5. Completion network. The completion network transforms the result of SNS into canonical space and predicts the holes in the UV domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.469, + 0.349 + ], + "angle": 0, + "content": "maps of the original posed sub-SMPLX and the canonical sub-SMPLX respectively. \\(N_{\\text{pose}}\\) and \\(N_{\\text{cano}}\\) are the corresponding UV normal maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.381 + ], + "angle": 0, + "content": "The complete UV position map \\(S\\) and the input combined hole mask \\(\\mathcal{H}_{sup}\\) are calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.395, + 0.469, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {s u p} = \\mathcal {H} _ {r} \\cdot \\left(1 - \\mathcal {H} _ {o}\\right) + \\mathcal {H} _ {o}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.443, + 0.469, + 0.461 + ], + "angle": 0, + "content": "\\[\nS = S _ {c a n o} + N _ {c a n o} \\cdot \\left[ d \\cdot \\mathcal {H} _ {o} + \\bar {d} \\cdot \\left(1 - \\mathcal {H} _ {o}\\right) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.469, + 0.546 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_r\\) and \\(\\mathcal{H}_o\\) denote the randomly added hole mask and the hole mask generated by SNS respectively. \\(d\\in\\) \\(\\mathbb{R}^{H\\times W\\times 3}\\) is the estimated displacement UV map. The complete mesh \\(\\mathcal{M}_c\\) is resampled from \\(S\\) and then transformed to the original pose space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.469, + 0.576 + ], + "angle": 0, + "content": "The primary loss used in our completion network is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.603, + 0.469, + 0.621 + ], + "angle": 0, + "content": "\\[\nS _ {p} = S _ {\\text {c a n o}} + N _ {\\text {c a n o}} \\cdot [ d \\cdot \\mathcal {H} _ {\\sup } + \\bar {d} \\cdot (1 - \\mathcal {H} _ {\\sup }) ], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.642, + 0.469, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n p} = \\frac {\\left\\| \\left(S _ {p} - \\bar {S}\\right) \\cdot \\left(\\mathcal {H} _ {s u p} - \\mathcal {H} _ {o}\\right) \\right\\| _ {2} ^ {2}}{\\sum_ {i , j} \\left(\\mathcal {H} _ {s u p} - \\mathcal {H} _ {o}\\right) _ {i , j}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.791 + ], + "angle": 0, + "content": "We find that completing the face, hands, and feet is much more difficult than other parts of the body, so we use the corresponding parts of sub-SMPLX for replacement. Furthermore, SHERT allows for the use of FLAME-based methods such as EMOCA [14] to replace the facial region in our completion results, enhancing the accuracy and realism of the facial details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.369, + 0.817 + ], + "angle": 0, + "content": "3.4. Self-supervised Mesh Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "After completing the mesh, we have already reconstructed a semantic mesh with extremely high levels of detail. However, some geometric details may be lost during the SNS and completion process. Therefore, we design an additional self-supervised refinement network to further optimize the" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.885, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.178, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Figure 6. The robustness of SNS and Completion Network. We present our complete meshes, which are reconstructed using the predictions of ECON [64] from in-the-wild images. Distinguishing itself from previous registration methods, SHERT has the capability to effectively process scenarios where the inputs are incomplete, inaccurate, or contain errors by leveraging SMPL-X human priors." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.281, + 0.891, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Figure 7. Refinement network. The features extracted from the image and front-back normal maps are projected to the UV domain. These projected features are subsequently combined with the input UV position map to generate a refined mesh." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.48, + 0.892, + 0.63 + ], + "angle": 0, + "content": "mesh's details (e.g. cloth wrinkles, subtle deformations of body movements) using the image and normal maps. The input normal maps can be obtained either by rendering the scanned model or by using prediction networks. Our refinement network follows a U-Net [53] architecture and is capable of predicting a displacement UV map, denoted as \\( z \\in \\mathbb{R}^{H \\times W \\times 3} \\), in order to optimize the complete mesh \\( \\mathcal{M}_c \\). Since the input image and normal maps are not in the UV domain, we utilize a separate network to extract the image domain features and then project them to UV domain." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.631, + 0.892, + 0.72 + ], + "angle": 0, + "content": "In order to guide the network in learning the potential correspondence between the input features and mesh details, we apply Laplacian Smoothing [20] on the existing complete meshes, as shown in Fig. 7. The smoothed mesh \\(\\mathcal{M}_l\\) is used as input to the network, and the original mesh \\(\\mathcal{M}_c\\) is used for supervision." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.722, + 0.892, + 0.781 + ], + "angle": 0, + "content": "Given the image \\(\\mathcal{I}\\), front-view normal map \\(\\mathcal{N}_f\\), backview normal map \\(\\mathcal{N}_b\\), and the corresponding complete UV position map \\(S_l\\), the projected feature \\(F\\) is typically represented as" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.8, + 0.891, + 0.816 + ], + "angle": 0, + "content": "\\[\nF = \\mathcal {P} \\left(\\mathcal {F} \\left(\\mathcal {I}, \\mathcal {N} _ {f}, \\mathcal {N} _ {b}\\right), S _ {l}, c\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where \\(c\\) represents the camera parameters, \\(\\mathcal{F}\\) denotes the image domain feature extraction network, and \\(\\mathcal{P}\\) is the projection function that can project the image domain features to the UV domain according to the camera parameters and the 3D coordinates of points on \\(S_{l}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "146" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.52, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.089, + 0.885, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.295 + ], + "angle": 0, + "content": "Figure 8. Texture diffusion. We use two strategies to finetune the diffusion model for text-driven texture inpainting and inpainting separately. During inference, instead of just using random noise, the encoded partial texture map and mask can be manually added to the latent features to preserve the original information of the input. Additionally, we use the super resolution method Real-esrgan [60], which can specifically optimize both the body and face, to further enhance the final outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.33 + ], + "angle": 0, + "content": "The UV position map \\( S_{r} \\) of the refined mesh \\( \\mathcal{M}_r \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.344, + 0.47, + 0.36 + ], + "angle": 0, + "content": "\\[\nS _ {r} = S _ {l} + N _ {l} \\cdot z, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.47, + 0.412 + ], + "angle": 0, + "content": "where \\(z \\in \\mathbb{R}^{H \\times W \\times 3}\\) represents the predicted displacement UV map of the refine network, \\(S_{l}\\) and \\(N_{l}\\) are the UV position map and normal UV map of \\(\\mathcal{M}_l\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.413, + 0.47, + 0.428 + ], + "angle": 0, + "content": "The loss we used in refine network can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.437, + 0.469, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s} = M S E \\left(z - \\frac {S _ {c} - S _ {l}}{N _ {l}}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.483, + 0.469, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = M S E \\left(N _ {r} - N _ {c}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.506, + 0.469, + 0.536 + ], + "angle": 0, + "content": "where \\(S_{c}\\) is the UV position map of \\(\\mathcal{M}_c\\). \\(N_{r}\\) and \\(N_{c}\\) denote the normal UV maps of \\(\\mathcal{M}_r\\) and \\(\\mathcal{M}_c\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.582 + ], + "angle": 0, + "content": "Moreover, the refine network can be iteratively employed by users to enhance the mesh details according to their desired level (refer to Fig. 11)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.469, + 0.607 + ], + "angle": 0, + "content": "3.5. Text-driven Texture Inpainting and Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.765 + ], + "angle": 0, + "content": "In Sec. 3.4, all the reconstructed meshes share the same semantic information from SMPL-X. This means that when the models are unfolded to the UV domain, vertices with the same semantic information will be projected to a fixed UV position, resulting in a stable texture map that allows for painting and transferring textures. To perform texture inpainting on incomplete human body textures and generate high-quality human body textures driven by text, we fine-tune the Stable Diffusion model [51] with ControlNet [71], similar to the approach taken by Dinar [59]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As our reconstructed semantic mesh closely matches the scanned model's geometry, we can use the ICP algorithm [7] to register the vertex colors from the scanned model onto our result. This enables us to convert the texture maps into the SMPL-X format. As shown in Fig. 8, we obtain a partial texture map from the input image based on the semantic mesh and camera parameters and hope to generate the invisible parts. ControlNet enable us to add conditional inputs (partial texture maps) to the generation of the diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.895, + 0.526 + ], + "angle": 0, + "content": "network at a low cost. We use two strategies to finetune the diffusion model, distinguished by whether to take the encoded partial texture map and mask as the additional inputs [27]. During the inference process, we can directly input partial texture maps into the ControlNet and generate various results by adding random noise. Alternatively, we can enhance the preservation of existing information in images by concatenating the encoded partial texture maps and masks with the latent features. Although Stable Diffusion has not worked with UV-parameterized images before, the well-designed UV parameterization keeps the shapes of the face, body, and limbs stable, ensuring a learnable space for the model. The texture map for the body and limbs focuses more on color and patterns rather than shapes, also resulting in excellent outcomes when cropped with a mask." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.539, + 0.634, + 0.556 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.564, + 0.715, + 0.578 + ], + "angle": 0, + "content": "4.1. Datasets and Networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.893, + 0.663 + ], + "angle": 0, + "content": "Training data. The completion and refinement networks are trained using the first 499 scans of THuman2.0 [69]. During the training of the completion network, we randomly choose one of the remaining 498 masks as a random hole mask. To enrich the inputs for the refinement network," + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.671, + 0.892, + 0.794 + ], + "angle": 0, + "content": "
MethodCAPE[43]THuman2.0[69]
P2S↓Chamfer↓Normal↓P2S↓Chamfer↓Normal↓
PIFu[54]*2.11371.65370.07552.54932.36400.1042
PIFuHD[55]3.78463.57870.10023.07723.18080.1207
PaMIR[70]*1.45201.22410.06101.54391.33110.1102
ICON[63]0.88550.86090.03471.03611.08740.0607
ECON[64]0.94030.93860.03741.13041.20810.0661
2K2K[23]†---2.53422.61650.1030
Ours (ICON-comp)0.8550↑0.8107↑0.03591.04591.0465↑0.0604↑
Ours (ICON-refine)0.86330.81120.03801.0442↑1.04680.0603↑
Ours (ECON-comp)0.8561↑0.8242↑0.03781.1255↑1.1420↑0.0672
Ours (ECON-refine)0.85810.8144↑0.03981.0630↑1.0430↑0.0649↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.8, + 0.894, + 0.897 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation for monocular image reconstruction. We evaluate the performance of our completion results (comp) and refinement results (refine) by comparing them with state-of-the-art methods. * methods are re-implemented in [63] to ensure a fair comparison. † method has only been tested with human-facing-forward images. ↑ and ↑ indicate the improvement achieved through completion and refinement, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.087, + 0.895, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison for monocular image reconstruction on in-the-wild image. For each method, we present two views of the reconstructed results. SHERT demonstrates the ability to handle challenging poses while providing clear details of facial and hand geometry." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.284, + 0.891, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.437, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparison for registration on THuman2.0. We compare the registration quality of various methods including N-ICP [6], RPTS [35], SVR-\\(\\mathcal{L}_0\\) [22], Fast-RNRR [68], DR [45], MDA [30] and ours SNS. The holes present in our result are the eliminated faces, as described in Sec. 3.2. The results indicate that SNS exhibits excellent performance in terms of model details, mesh quality, and registration robustness. The quantitative comparisons are shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.625 + ], + "angle": 0, + "content": "we rotate the meshes every 60 degrees, resulting in a total of 2994 different orientations. We utilized the ICP [7] algorithm to transfer the color of the THuman2.0 scans to the vertices of our completed result, thereby generating 499 UV textures for training and obtaining 2994 visible UV masks from the rotated meshes. The complete UV textures and visible UV masks are randomly combined as inputs for our texture diffusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Testing data. We conduct quantitative and qualitative evaluations on CAPE [43], THuman2.0, and in-the-wild images. We use CAPE-NFP [63] (100 samples with 3 viewpoints for each), and the last 27 subjects of THuman2.0 scans (6 viewpoints, each differing by 60 degrees)." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.719, + 0.462, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.842, + 0.47, + 0.896 + ], + "angle": 0, + "content": "Figure 11. The ablation results (with face substitution). We present the results after completion and refinement. With an increasing number of refinement iterations, the details of the mesh will be enhanced. Please zoom-in to see more details." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.503, + 0.894, + 0.611 + ], + "angle": 0, + "content": "
MethodGPUP2S↓Chamfer↓G-avg↑θ<30°↓Time
N-ICP [6]-0.2130.1630.50655.27m 23s
RPTS [35]-0.4880.3600.56548.31m 55s
SVR-L0 [22]-0.4040.2960.53153.51h 23m 32s
Fast-RNRR [68]-0.1150.0970.59731.11m 4s
DR [45]✓0.3390.3470.58147.716m 17s
MDA [30]✓0.6710.7310.58748.54m 48s
Ours (SNS)-0.1070.0780.72917.323s
Ours (Comp)✓0.1390.1670.66228.927s (23 + 4)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation for registration on THuman2.0. We test all the methods on the first subject of THuman2.0. Following the previous researches [21, 38], we adapt G-avg as a method for evaluating the mesh quality. We also report the metric \\(\\theta < 30^{\\circ}\\), which denotes the percentage of triangle meshes in the given mesh that have an angle less than 30 degrees. Additionally, we present the metrics for our complete mesh." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Networks. The completion net and refinement net are both trained for 100 epochs with a learning rate of \\(1 \\times 10^{-6}\\). The resolutions of the input and output data, including the UV position maps, images, masks and front-back normal maps, are all \\(1024 \\times 1024 \\times 3\\). During inference, the ECON's predicted front-back normals \\((512 \\times 512 \\times 3)\\) are upsampled to \\(1024 \\times 1024 \\times 3\\) using bilinear interpolation. The texture diffusion is trained for 1 epoch with a learning rate of \\(2 \\times 10^{-5}\\). The sampler of texture diffusion is DDIM [58]. We use 30 steps by default and infer the texture UV map with a resolution of \\(1024 \\times 1024 \\times 3\\). All the networks are trained on three NVIDIA RTX 3090 GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.206, + 0.471, + 0.275 + ], + "angle": 0, + "content": "Figure 12. Qualitative comparison for texture prediction on in-the-wild image. We display the front and back view rendering results for each method. Since PIFu [54] and PaMIR [70] predict vertex colors, we only exhibit the texture maps of DINAR [59] and SHERT (ours). Please zoom-in to see more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.287, + 0.199, + 0.302 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.583 + ], + "angle": 0, + "content": "Quantitative comparisons. We conduct quantitative comparisons with mainstream state-of-the-art monocular image reconstruction approaches in Tab. 1. As in previous work [23, 55, 63, 64], we report the point-to-surface Euclidean distance (P2S, cm), the Chamfer Distance (cm), and the Normals difference (L2). To ensure a fair comparison, \\(\\mathrm{PIFu}^*\\) [54] and \\(\\mathrm{PaMIR}^*\\) [70] are re-implemented and restrained on THuman2.0, using the same settings as ICON [63]. The ground-truth SMPL/SMPL-X models are provided for evaluation. However, PIFu [54], PIFuHD [55] and 2K2K [23] do not utilize the parametric body priors, which may result in subpar performance. In Tab. 1, \"ICON-comp\" refers to the completion result achieved by leveraging ICON's [63] prediction, while \"ECON-refine\" denotes the refinement mesh obtained using ICON's [64] result and the predicted front-back normal maps. Additionally, we evaluate the registration quality of SNS against state-of-the-art non-rigid registration methods, as presented in Tab. 2" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Qualitative comparisons. We demonstrate a comparison between SHERT and state-of-the-art methods using in-the-wild images, with a focus on monocular image reconstruction (refer to Fig. 9) and texture prediction (refer to Fig. 12). Additionally, we compare the registration quality of our SNS with state-of-the-art non-rigid registration approaches on Thuman2.0 (refer to Fig. 10)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.204, + 0.715 + ], + "angle": 0, + "content": "4.3. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.47, + 0.8 + ], + "angle": 0, + "content": "Due to the geometric limitations of SMPL-X, SHERT performs weaker in reconstructing loose clothing, shoes and hair compared to implicit-based reconstruction methods. It is also difficult to ensure consistent results of texture diffusion at the seams of UVs. See more in SupMat." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.208, + 0.832 + ], + "angle": 0, + "content": "5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "SHERT uses the skinning weights from SMPL-X to enable animated poses, expressions, and gestures on the reconstructed mesh through LBS [34] (refer to Fig. 13). It allows for both global texture repainting (refer to Fig. 14)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.088, + 0.877, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.182, + 0.892, + 0.195 + ], + "angle": 0, + "content": "Figure 13. Animation results. Please zoom-in to see more details." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.197, + 0.891, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.353, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Figure 14. Global texture repainting. SHERT can repaint the texture through text prompts. Please zoom-in to see more details." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.38, + 0.88, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.525, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Figure 15. Localized texture repainting. SHERT can repaint the masked area through text prompts. Please zoom-in to see more details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.892, + 0.599 + ], + "angle": 0, + "content": "and the option for users to provide custom masks and text prompts for localized texture repainting (refer to Fig. 15)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.619, + 0.627 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.818 + ], + "angle": 0, + "content": "We propose SHERT, which reconstructs a fully textured semantic human avatar from a detailed surface or a monocular image. It takes advantage of the geometric details of the target surface, along with semantic information and prior knowledge of the semantic guider. The reconstructed results have high-fidelity clothing details, high-quality triangle meshes, clear facial features, and complete hands geometry. SHERT is also capable of generating high-resolution texture maps with stable UV unwrapping. This approach bridges existing monocular reconstruction work and downstream industrial applications, and we believe it can promote the development of human avatars." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.66, + 0.848 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (62032011) and the Natural Science Foundation of Jiangsu Province (No. BK20211147)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "149" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In IEEE Conf. 3D Vis.(3DV), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single rgb camera. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[4] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[5] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[6] Brian Amberg, Sami Romdhani, and Thomas Vetter. Optimal step nonrigid icp algorithms for surface registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2007. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[7] P.J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 1992. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[8] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Trans. Vis. Comput. Graph.(TVCG), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[10] Kunyao Chen, Fei Yin, Bang Du, Baichuan Wu, and Truong Q. Nguyen. Efficient registration for human surfaces via isometric regularization on embedded deformation. IEEE Trans. Vis. Comput. Graph.(TVCG), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[13] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[14] Radek Danecek, Michael J. Black, and Timo Bolkart. EMOCA: Emotion driven monocular face capture and animation. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Bailin Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. Computer Graphics Forum (Eurographics 2022 State-of-the-Art Reports), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Jiankang Deng, Shiyang Cheng, Niannan Xue, Yuxiang Zhou, and Stefanos Zafeiriou. Uv-gan: Adversarial facial uv map completion for pose-invariant face recognition. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[18] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. ACM Trans. Graph.(TOG), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[19] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[20] David A Field. Laplacian smoothing and delaunay triangulations. Communications in applied numerical methods, 1988. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[21] Pascal J Frey and Houman Borouchaki. Surface mesh quality evaluation. International journal for numerical methods in engineering, 1999. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[22] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Int. Conf. Comput. Vis.(ICCV), 2015. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[23] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[24] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. \\(\\mathrm{ARCH} + +\\): Animation-ready clothed human reconstruction revisited. In Int. Conf. Comput. Vis.(ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[25] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[26] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable reconstruction of clothed humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[27] Huggingface. Stable-diffusion-inpainting, 2022. https://huggingface.co/runwayml/stable-diffusion-inpainting.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[28] Henrik Wann Jensen. Realistic image synthesis using photon mapping. AK Peters/crc Press, 2001. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[29] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In Eur. Conf. Comput. Vis.(ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Yucheol Jung, Hyomin Kim, Gyeongha Hwang, Seung-Hwan Baek, and Seungyong Lee. Mesh density adapta" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "150" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "tion for template-based shape reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[31] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph.(TOG), 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[33] Vladislav Kraevoy and Alla Sheffer. Template-based mesh completion. In Symposium on Geometry Processing, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[34] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of SIGGRAPH 2000, 2000. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[35] Kun Li, Jingyu Yang, Yu-Kun Lai, and Daoliang Guo. Robust non-rigid registration with reweighted position and transformation sparsity. IEEE Trans. Vis. Comput. Graph.(TVCG), 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[36] Tianye Li, Timo Bolkart, Michael Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph.(TOG), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Trans. Graph.(TOG), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[38] Yuanqi Li, Jianwei Guo, Xinran Yang, Shun Liu, Jie Guo, Xiaopeng Zhang, and Yanwen Guo. Deep point cloud simplification for high-quality surface reconstruction. arXiv preprint arXiv:2203.09088, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[40] Peter Liepa. Filling holes in meshes. In Proceedings of the 2003 Eurographics/ACM SIGGRAPH symposium on Geometry processing, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graph.(TOG), 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[42] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. Seminal graphics: pioneering efforts that shaped the field, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[43] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[44] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[45] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph.(TOG), 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[47] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[48] Ken Perlin and Eric M Hoffert. Hypertexture. In Proceedings of the 16th annual conference on Computer graphics and interactive techniques, 1989. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[49] Leonid Pishchulin, Stefanie Wuhrer, Thomas Helten, Christian Theobalt, and Bernt Schiele. Building statistical shape spaces for 3d human modeling. Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[50] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[52] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph.(TOG), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention(MICCAI), 2015, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[54] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[55] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[56] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Ming-min Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In Eur. Conf. Comput. Vis.(ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[57] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X- avatar: Expressive human avatars. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[59] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Int. Conf. Comput. Vis.(ICCV), 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[60] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Int. Conf. Comput. Vis.(ICCV), 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[61] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In IEEE Conf. 3D Vis.(3DV), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[62] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph.(TOG), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[63] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[64] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[65] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3d human shape and articulated pose models. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[66] Jingyu Yang, Ke Li, Kun Li, and Yu-Kun Lai. Sparse non-rigid registration of 3d shapes. Computer Graphics Forum, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.468, + 0.488 + ], + "angle": 0, + "content": "[67] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-if: Uncertainty-aware human digitization via implicit distribution field. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[68] Yuxin Yao, Bailin Deng, Weiwei Xu, and Juyong Zhang. Quasi-newton solver for robust non-rigid registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[69] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[70] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis.(ICCV), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[72] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.761, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[73] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[74] Hao Zhu, Xinxin Zuo, Haotian Yang, Sen Wang, Xun Cao, and Ruigang Yang. Detailed avatar recovery from single image. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[75] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity gar" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ment mesh reconstruction from single images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[76] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "152" + } + ] +] \ No newline at end of file diff --git a/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_origin.pdf b/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8cf1444c2185a2578bf88e75ab69b7231ed5ccc7 --- /dev/null +++ b/2024/Semantic Human Mesh Reconstruction with Textures/08bc6bd7-0008-4e31-94a7-d23c37e8682d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bc84f5e01b05a5343d580668008b19284fdc77899ac8f9c54d97b4f3c2a0475 +size 9882032 diff --git a/2024/Semantic Human Mesh Reconstruction with Textures/full.md b/2024/Semantic Human Mesh Reconstruction with Textures/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d43bb0996ca511eaf3c7a41d469874fcd5372528 --- /dev/null +++ b/2024/Semantic Human Mesh Reconstruction with Textures/full.md @@ -0,0 +1,339 @@ +# Semantic Human Mesh Reconstruction with Textures + +Xiaoyu Zhan $^{1}$ , Jianxin Yang $^{1}$ , Yuanqi Li $^{1}$ , Jie Guo $^{1}$ , Yanwen Guo $^{1*}$ , and Wenping Wang $^{2}$ $^{1}$ Nanjing University $^{2}$ Texas A&M University + +{zhanxy, jianxin-yang, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn, wenping@tamu.edu https://zhanxy.xyz/projects/shert + +![](images/78a4fe5fe0d7ad43a6a35a99b377612be899fc9f4f38716f8cf85e7d3db20f2f.jpg) +Figure 1. Semantic Human mEsh Reconstruction with Textures (SHERT): (a) Given a target surface and the corresponding semantic guider, (b) SHERT reconstructs a detailed semantic model, which has stable UV unwrapping and skinning weights with high-quality triangle meshes. (c) It can either sample a texture map from the target surface or generate from the text prompts. (d) Based on our semantic representation, SHERT allows for high-precision facial reconstruction and animation of the body, face, and hands. (e) Moreover, SHERT is capable of inferring a fully textured avatar from a monocular image. + +# Abstract + +The field of 3D detailed human mesh reconstruction has made significant progress in recent years. However, current methods still face challenges when used in industrial applications due to unstable results, low-quality meshes, and a lack of UV unwrapping and skinning weights. In this paper, we present SHERT, a novel pipeline that can reconstruct semantic human meshes with textures and high-precision details. SHERT applies semantic- and normal-based sampling between the detailed surface (e.g. mesh and SDF) and the corresponding SMPL-X model to obtain a partially sampled semantic mesh and then generates the complete semantic mesh by our specifically designed self-supervised completion and refinement networks. Using the complete semantic mesh as a basis, we employ a texture diffusion model to create human textures that are driven by both images and texts. Our reconstructed meshes have stable UV unwrapping, high-quality triangle meshes, and consistent semantic + +information. The given SMPL-X model provides semantic information and shape priors, allowing SHERT to perform well even with incorrect and incomplete inputs. The semantic information also makes it easy to substitute and animate different body parts such as the face, body, and hands. Quantitative and qualitative experiments demonstrate that SHERT is capable of producing high-fidelity and robust semantic meshes that outperform state-of-the-art methods. + +# 1. Introduction + +Recovering highly realistic details and textures of a human mesh from monocular images is crucial for various applications such as gaming, movies, cartoons, VR, virtual try-on, and digital avatars. Current approaches [1, 2, 4, 23, 24, 26, 39, 54, 55, 57, 61, 63, 64, 67, 70] primarily focus on the recovery of the geometric details that are associated with images, but their results are not yet practical for real-world ap + +plications. Recent advancements in parametric and explicit clothing models [12, 19, 29, 50, 62, 75] have shown promise in clothing reconstruction. However, they have limitations in accurately fitting different geometries and details. Implicit reconstruction approaches [23, 54, 55, 63, 64, 67, 70] excel at capturing clothing details but perform poorly in hands and face reconstruction. They may also generate incomplete and geometrically inseparable results. + +In this work, our goal is to reconstruct fully textured semantic human meshes through given detailed surfaces and corresponding semantic guiders. Our semantic human mesh is complete, animatable, and edit-friendly for both users and designers. It ensures that each vertex has deterministic semantic information and predefined skinning weights. This allows for easy substitution and animation of different body parts such as the face, body, and hands. The semantic information also guarantees stable and reasonable UV unwrapping, which is advantageous for editing and image-based texture generation. + +Specifically, we propose SHERT, which generates a semantic human mesh from the detailed surface and its corresponding SMPL-X [46] model and optionally infers textures from images or colored surfaces. SHERT has four main processes. 1) Sampling, SHERT applies semantic and normal-based sampling to obtain a partially sampled semantic mesh based on the input detailed surface and SMPL-X model. We subdivide the original SMPL-X model to better capture the human details. 2) Completion, a self-supervised network is proposed to complete the partially sampled mesh. The network works in the 2D UV domain, which converts the 3D completion task into a 2D inpainting task. 3) Refinement, the origin image and front-back normal maps are used to enhance the geometry details. 4) Texture, our semantic human mesh is projected to the origin image in order to generate the partial texture map. Then we adapt the diffusion framework for text-driven partial texture inpainting and generation. The powerful generation ability of pre-trained diffusion model enables textures with rich clothing details and clear facial expressions. + +Extensive experiments have been conducted on both datasets and in-the-wild images. The quantitative and qualitative results show the robustness and superior performance of SHERT in reconstructing high-fidelity semantic human meshes and generating various high-resolution textures. + +In summary, the main contributions of this work include four-fold. + +- We introduce SHERT, a novel pipeline to reconstruct high-quality semantic human meshes from the detailed 3D surfaces, represented either explicitly as meshes, or implicitly as signed distance field (SDF). SHERT is also capable of predicting robust and fully textured avatars with high-fidelity faces from monocular images. +- We propose a semantic- and normal-based sampling + +method (SNS) and a self-supervised mesh completion network to achieve non-rigid 3D surface registration. The approach has the capability to process incomplete and inaccurate inputs by leveraging SMPL-X human priors. + +- We present a self-supervised mesh refinement network working in the UV domain. It utilizes the images and front-back normal maps to improve the geometric mesh details. +- We use a diffusion model to infer high-resolution human textures from input images. The model can also accomplish text-driven texture inpainting and generation. + +# 2. Related Work + +# 2.1. Monocular 3D Human Reconstruction + +Clothed Human Reconstruction. Most of the current CNN-based monocular 3D clothed human shape estimation methods can be divided into two categories: explicit-based [1-4, 29, 73, 74] and implicit-based [5, 12, 23-26, 39, 54, 55, 57, 63, 64, 67, 70] approaches, based on their representations. Explicit-based approaches usually infer the 3D offsets on top of the parametric human model [41, 46, 49, 65]. However, these methods are difficult to apply to flexible human topologies and cannot capture details well. Implicit-based approaches have the advantage of representing arbitrary 3D clothed human shapes that are free from the limitations of parametric human models. There are also some works [13, 25, 39, 63, 64, 70] that mix the implicit and explicit representations to achieve the detailed and robust 3D clothed human reconstruction. But both implicit-based and mixed methods still cannot maintain the stability of the human body shape well, often resulting in problems such as blurring and missing local body parts in the predicted results. Some methods [12, 19, 29, 50, 62, 75] propose additional parametric or implicit clothing models to fit loose clothes. These clothing models often struggle to accurately capture the details present in the images. Nevertheless, they are still useful for generating rough approximations of the clothing. Despite considerable progress made in monocular human reconstruction capture, there are still certain constraints, especially in terms of reliability, availability, and user-friendliness. + +Texture Prediction. Previous works [5, 13, 24-26, 44, 54, 57, 59] have also focused on predicting image-based human textures. However, the current emphasis of implicit-based methods [5, 13, 24-26, 54] is still on predicting vertex colors, which cannot be easily converted into usable texture maps for industrial applications. DINAR [59] proposes neural textures for modeling human avatars and achieves texture inpainting by utilizing the diffusion framework. Nevertheless, the neural texture is highly coupled with human avatars, and the resolution is relatively low. + +Face Reconstruction. In recent years, parameterized face + +![](images/dfcc98ab90beb3fbca00a919737c29c38021b7a01a830b65db5843c02659f70d.jpg) +Figure 2. Overview of SHERT for monocular image reconstruction. Given an RGB image $\mathcal{I}$ , SHERT first infers the detailed mesh $\mathcal{M}_t$ and corresponding sub-SMPLX model $\mathcal{M}_x$ . It then applies SNS to obtain the partial semantic mesh $\mathcal{M}_s$ (in Sec. 3.2). The Completion Net infers $\mathcal{M}_c$ by filling the UV holes in $\mathcal{M}_s$ (in Sec. 3.3). Image and normal maps are utilized to generate $\mathcal{M}_r$ , which contains sharper geometry details (in Sec. 3.4). Finally, SHERT uses a diffusion model to achieve text-driven texture inpainting and generation (in Sec. 3.5). + +models [8, 9, 36, 47] have been widely used in monocular face reconstruction algorithms [14, 17, 18, 56, 76], contributing to the success of achieving realistic high-quality reconstruction results. In light of the incorporation of FLAME [36] into SMPL-X [46], we can utilize high-quality facial reconstruction results from previous works to enhance the accuracy and realism of the SMPL-X based human body reconstruction results. + +# 2.2. Non-rigid 3D Registration + +Non-rigid 3D surface registration approaches [6, 10, 22, 30, 35, 45, 66, 68] usually compute a deformation that aligns a source surface with a target surface [15]. SHERT aims to transfer the semantic information of parametric human models to the corresponding detailed surfaces through registration. However, existing methods mainly focus on the precise registration of the source and target, while the target detailed human surfaces in real-world tasks often have incomplete and erroneous data. + +# 3. Method + +SHERT is capable of reconstructing a high-fidelity fully textured semantic human mesh based on a pair of detailed 3D surface and corresponding SMPL-X [46] model. Our results have high robustness and will not result in missing body parts in despite of the incomplete inputs. Furthermore, the semantic mesh ensures that each vertex has deterministic semantic information and predefined skinning weights, making it possible to replace and animate the human face, body, and hands. SHERT can also generate realistic hu + +man textures from texts and images. These features make it easy for users and designers to use or further edit our results. To achieve this, SHERT first subdivides the SMPL-X model to improve the accuracy in capturing human details (in Sec. 3.1) and then applies the semantic- and normal-based sampling to obtain a partially sampled mesh (in Sec. 3.2). We complete the partial result using a self-supervised mesh completion network (in Sec. 3.3), and finally enhance the geometry details through refinement (in Sec. 3.4). In addition, SHERT uses a diffusion model for text-driven human texture inpainting and generation (in Sec. 3.5). + +# 3.1. Subdivided SMPL-X model + +In this work, we use SMPL-X as the semantic guider. SMPL-X is a parameterized body model that combines SMPL [41] with the FLAME [37] head model and the MANO [52] hand model. It is parameterized with shape and pose, has 10, 475 vertices and 54 joints, including joints for the neck, jaw, eyeballs, and fingers. The SMPL-X model provides basic skinning weights and corresponding semantic information for each vertex. + +SHERT defines the sub-SMPLX based on SMPL-X since 10,475 vertices are not sufficient to accurately represent the details of the human body and clothing, such as facial details and clothing wrinkles. Taking into consideration both the expressiveness of the model and computational cost, we apply the mid-point subdivision algorithm twice on a standard SMPL-X model (with eyeballs removed) to obtain the sub-SMPLX with 149,921 vertices and 299,712 faces. + +![](images/21d0bec775a267e2e481f6bb85b3278d5a9ca4a728769cd55e4f9bae5846c565.jpg) +Figure 3. Explicit and implicit sampling. a) SNS shoots a ray from the starting vertex along the vertex normal to locate the intersection point with the target surface. b) SNS takes samples at fixed intervals along the vertex normal ray and search for the point that is closest to the zero isosurface. + +# 3.2. Semantic- and Normal-based Sampling (SNS) + +Our objective is to accurately predict the non-rigid deformation between a source surface and a target surface. Given a source surface $\mathcal{M}_x$ , which is represented as a sub-SMPLX in our work, and the target surface $M_t$ , we learn a mapping function $D: \mathcal{M}_x^{149921 \times 3} \to \mathcal{M}_s^{149921 \times 3}$ such that $\mathcal{M}_s$ can be aligned with $\mathcal{M}_t$ through the learned mapping in geometry, and ignoring the incorrect parts. + +Specifically, we obtain a partially sampled mesh through a sampling scheme based on the vertex normals $N^{149921 \times 3}$ . Starting from a point on $\mathcal{M}_x$ , we search for the intersection point with the target surface $\mathcal{M}_t$ along the vertex normal ray. We can extend the sampling scheme to implicit surfaces with constant step Ray Marching [28, 48], shown in Fig. 3 and Fig. 4. + +It should be noted that SNS does not always return satisfactory results since the ray may not intersect with the target surface. We need to locate and label the vertices that have failed to register. In addition, due to the large geometric differences and the possibility of incorrect alignment, the sampling scheme may retrieve incorrect results, which should also be detected. In order to remove incorrect points and triangle meshes from the sampled results, we calculate $\theta$ (the angle between the normal vectors of the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), $s$ (the area ratio between the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), and $r$ (the edge ratio between the longest and shortest edges of the sampled triangle mesh) for each triangle mesh in the sampled result. We then perform mesh culling by removing low-quality triangle meshes that have at least one indicator exceeding the threshold among the three mentioned above. Finally, we perform a connectivity check on the processed sampled mesh and remove the sets with a number of connected triangle meshes less than $g$ . By default, we set $\theta = 2$ , $s = 3$ , $r = 3$ , and $g = 500$ . + +We found that culling meshes only in the current pose would result in unreasonable deformations in human animation tasks. Therefore, SNS repeats the processing above in the canonical space. This ensures that we remove almost all unreasonable sampling results and obtain the partial se + +![](images/be20819802bf73bbd27d596bd85d23e3300a1b077f7db0c3d73a3faedc289789.jpg) +of SNS and the Marching Cubes algorithm [42] in the implicit field predicted by the multi-view PIFu [54]. The results show that our SNS and completion network correctly reconstruct the implicit field and obtain a smoother surface compared to Marching Cubes. mantic mesh $\mathcal{M}_s$ . + +# 3.3. Self-supervised Mesh Completion + +In Sec. 3.2, the vertices that failed to register have resulted in holes in $M_{s}$ . Currently, there are many mesh completion algorithms [11, 31-33, 40] that can fill in these mesh holes. However, these algorithms cannot be well adapted to our semantic reconstruction task due to the strong constraints on the number and relative positions of vertices in our representation in Sec. 3.1. + +Therefore, we transform the partially sampled mesh into the UV domain [4, 16, 72] according to the semantic information. As a result, we convert the mesh completion task in 3D space into an inpainting task on a 2D image. Then, we design a self-supervised completion network that can fill in the holes of the partially sampled mesh. As shown in Fig. 5, by adding a random hole mask $\mathcal{H}_r$ on the partial UV position map, we can obtain trainable pairs, which has a consistent distribution as the missing parts of the sampled mesh. The manually masked parts can provide the network with consistent supervision information as the input parts. This completion network is capable of generating robust meshes while maintaining both mesh quality and semantic consistency in the final results (refer to Fig. 6). + +To further decrease the learning difficulties, we transform all incompletely sampled meshes to canonical space when facing the diverse range of poses and clothing styles. We believe that this approach has the potential to mitigate the challenges in problem-solving, as completion is no longer affected by pose and body shape. Meanwhile, since the sampled points are located on the normal ray of the sub-SMPLX vertices, we represent the deformation relative to sub-SMPLX as an offset based on the vertex normal. The transformed UV position map $\overline{S}$ can be presented as: + +$$ +\bar {d} = \frac {S _ {\text {s a m p l e}} - S _ {\text {p o s e}}}{N _ {\text {p o s e}}}, \tag {1} +$$ + +$$ +\bar {S} = S _ {\text {c a n o}} + N _ {\text {c a n o}} \cdot \bar {d}, \tag {2} +$$ + +where $S_{\text{sample}}$ denotes the partial UV position map of the sampled mesh. $S_{\text{pose}}$ and $S_{\text{cano}}$ refer to the UV position + +![](images/59c617ab77159f431c737eeada4b8c40254e1fd9ca69b9847bea1c9d24454f2b.jpg) +Figure 5. Completion network. The completion network transforms the result of SNS into canonical space and predicts the holes in the UV domain. + +maps of the original posed sub-SMPLX and the canonical sub-SMPLX respectively. $N_{\text{pose}}$ and $N_{\text{cano}}$ are the corresponding UV normal maps. + +The complete UV position map $S$ and the input combined hole mask $\mathcal{H}_{sup}$ are calculated by: + +$$ +\mathcal {H} _ {s u p} = \mathcal {H} _ {r} \cdot \left(1 - \mathcal {H} _ {o}\right) + \mathcal {H} _ {o}, \tag {3} +$$ + +$$ +S = S _ {c a n o} + N _ {c a n o} \cdot \left[ d \cdot \mathcal {H} _ {o} + \bar {d} \cdot \left(1 - \mathcal {H} _ {o}\right) \right], \tag {4} +$$ + +where $\mathcal{H}_r$ and $\mathcal{H}_o$ denote the randomly added hole mask and the hole mask generated by SNS respectively. $d\in$ $\mathbb{R}^{H\times W\times 3}$ is the estimated displacement UV map. The complete mesh $\mathcal{M}_c$ is resampled from $S$ and then transformed to the original pose space. + +The primary loss used in our completion network is as follows: + +$$ +S _ {p} = S _ {\text {c a n o}} + N _ {\text {c a n o}} \cdot [ d \cdot \mathcal {H} _ {\sup } + \bar {d} \cdot (1 - \mathcal {H} _ {\sup }) ], \tag {5} +$$ + +$$ +\mathcal {L} _ {i n p} = \frac {\left\| \left(S _ {p} - \bar {S}\right) \cdot \left(\mathcal {H} _ {s u p} - \mathcal {H} _ {o}\right) \right\| _ {2} ^ {2}}{\sum_ {i , j} \left(\mathcal {H} _ {s u p} - \mathcal {H} _ {o}\right) _ {i , j}}. \tag {6} +$$ + +We find that completing the face, hands, and feet is much more difficult than other parts of the body, so we use the corresponding parts of sub-SMPLX for replacement. Furthermore, SHERT allows for the use of FLAME-based methods such as EMOCA [14] to replace the facial region in our completion results, enhancing the accuracy and realism of the facial details. + +# 3.4. Self-supervised Mesh Refinement + +After completing the mesh, we have already reconstructed a semantic mesh with extremely high levels of detail. However, some geometric details may be lost during the SNS and completion process. Therefore, we design an additional self-supervised refinement network to further optimize the + +![](images/0182e7005cdd47679cad32888c6302e506948c9beb737252f3ee5d78528abb54.jpg) +Figure 6. The robustness of SNS and Completion Network. We present our complete meshes, which are reconstructed using the predictions of ECON [64] from in-the-wild images. Distinguishing itself from previous registration methods, SHERT has the capability to effectively process scenarios where the inputs are incomplete, inaccurate, or contain errors by leveraging SMPL-X human priors. + +![](images/25b4c87d12205b62882c53201b1fa3250df46a7927fdc10e17bed4fb3613be52.jpg) +Figure 7. Refinement network. The features extracted from the image and front-back normal maps are projected to the UV domain. These projected features are subsequently combined with the input UV position map to generate a refined mesh. + +mesh's details (e.g. cloth wrinkles, subtle deformations of body movements) using the image and normal maps. The input normal maps can be obtained either by rendering the scanned model or by using prediction networks. Our refinement network follows a U-Net [53] architecture and is capable of predicting a displacement UV map, denoted as $z \in \mathbb{R}^{H \times W \times 3}$ , in order to optimize the complete mesh $\mathcal{M}_c$ . Since the input image and normal maps are not in the UV domain, we utilize a separate network to extract the image domain features and then project them to UV domain. + +In order to guide the network in learning the potential correspondence between the input features and mesh details, we apply Laplacian Smoothing [20] on the existing complete meshes, as shown in Fig. 7. The smoothed mesh $\mathcal{M}_l$ is used as input to the network, and the original mesh $\mathcal{M}_c$ is used for supervision. + +Given the image $\mathcal{I}$ , front-view normal map $\mathcal{N}_f$ , backview normal map $\mathcal{N}_b$ , and the corresponding complete UV position map $S_l$ , the projected feature $F$ is typically represented as + +$$ +F = \mathcal {P} \left(\mathcal {F} \left(\mathcal {I}, \mathcal {N} _ {f}, \mathcal {N} _ {b}\right), S _ {l}, c\right), \tag {7} +$$ + +where $c$ represents the camera parameters, $\mathcal{F}$ denotes the image domain feature extraction network, and $\mathcal{P}$ is the projection function that can project the image domain features to the UV domain according to the camera parameters and the 3D coordinates of points on $S_{l}$ . + +![](images/df8663b470ba65973de2791b24d60924a77afe1309f421bdb2c64c7267432914.jpg) +Figure 8. Texture diffusion. We use two strategies to finetune the diffusion model for text-driven texture inpainting and inpainting separately. During inference, instead of just using random noise, the encoded partial texture map and mask can be manually added to the latent features to preserve the original information of the input. Additionally, we use the super resolution method Real-esrgan [60], which can specifically optimize both the body and face, to further enhance the final outputs. + +![](images/fb73cb07c45c17ef49634b845d39cdc7973ff5c29d9d7afbb99b6d884e81328f.jpg) + +The UV position map $S_{r}$ of the refined mesh $\mathcal{M}_r$ is given by + +$$ +S _ {r} = S _ {l} + N _ {l} \cdot z, \tag {8} +$$ + +where $z \in \mathbb{R}^{H \times W \times 3}$ represents the predicted displacement UV map of the refine network, $S_{l}$ and $N_{l}$ are the UV position map and normal UV map of $\mathcal{M}_l$ respectively. + +The loss we used in refine network can be represented as + +$$ +\mathcal {L} _ {d i s} = M S E \left(z - \frac {S _ {c} - S _ {l}}{N _ {l}}\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {\text {n o r m a l}} = M S E \left(N _ {r} - N _ {c}\right), \tag {10} +$$ + +where $S_{c}$ is the UV position map of $\mathcal{M}_c$ . $N_{r}$ and $N_{c}$ denote the normal UV maps of $\mathcal{M}_r$ and $\mathcal{M}_c$ respectively. + +Moreover, the refine network can be iteratively employed by users to enhance the mesh details according to their desired level (refer to Fig. 11). + +# 3.5. Text-driven Texture Inpainting and Generation + +In Sec. 3.4, all the reconstructed meshes share the same semantic information from SMPL-X. This means that when the models are unfolded to the UV domain, vertices with the same semantic information will be projected to a fixed UV position, resulting in a stable texture map that allows for painting and transferring textures. To perform texture inpainting on incomplete human body textures and generate high-quality human body textures driven by text, we fine-tune the Stable Diffusion model [51] with ControlNet [71], similar to the approach taken by Dinar [59]. + +As our reconstructed semantic mesh closely matches the scanned model's geometry, we can use the ICP algorithm [7] to register the vertex colors from the scanned model onto our result. This enables us to convert the texture maps into the SMPL-X format. As shown in Fig. 8, we obtain a partial texture map from the input image based on the semantic mesh and camera parameters and hope to generate the invisible parts. ControlNet enable us to add conditional inputs (partial texture maps) to the generation of the diffusion + +network at a low cost. We use two strategies to finetune the diffusion model, distinguished by whether to take the encoded partial texture map and mask as the additional inputs [27]. During the inference process, we can directly input partial texture maps into the ControlNet and generate various results by adding random noise. Alternatively, we can enhance the preservation of existing information in images by concatenating the encoded partial texture maps and masks with the latent features. Although Stable Diffusion has not worked with UV-parameterized images before, the well-designed UV parameterization keeps the shapes of the face, body, and limbs stable, ensuring a learnable space for the model. The texture map for the body and limbs focuses more on color and patterns rather than shapes, also resulting in excellent outcomes when cropped with a mask. + +# 4. Experiments + +# 4.1. Datasets and Networks + +Training data. The completion and refinement networks are trained using the first 499 scans of THuman2.0 [69]. During the training of the completion network, we randomly choose one of the remaining 498 masks as a random hole mask. To enrich the inputs for the refinement network, + +
MethodCAPE[43]THuman2.0[69]
P2S↓Chamfer↓Normal↓P2S↓Chamfer↓Normal↓
PIFu[54]*2.11371.65370.07552.54932.36400.1042
PIFuHD[55]3.78463.57870.10023.07723.18080.1207
PaMIR[70]*1.45201.22410.06101.54391.33110.1102
ICON[63]0.88550.86090.03471.03611.08740.0607
ECON[64]0.94030.93860.03741.13041.20810.0661
2K2K[23]†---2.53422.61650.1030
Ours (ICON-comp)0.8550↑0.8107↑0.03591.04591.0465↑0.0604↑
Ours (ICON-refine)0.86330.81120.03801.0442↑1.04680.0603↑
Ours (ECON-comp)0.8561↑0.8242↑0.03781.1255↑1.1420↑0.0672
Ours (ECON-refine)0.85810.8144↑0.03981.0630↑1.0430↑0.0649↑
+ +Table 1. Quantitative evaluation for monocular image reconstruction. We evaluate the performance of our completion results (comp) and refinement results (refine) by comparing them with state-of-the-art methods. * methods are re-implemented in [63] to ensure a fair comparison. † method has only been tested with human-facing-forward images. ↑ and ↑ indicate the improvement achieved through completion and refinement, respectively. + +![](images/822937ec2201c677ed60b7c2329350140938f139eab8e7c52282e35ae6fec2a7.jpg) +Figure 9. Qualitative comparison for monocular image reconstruction on in-the-wild image. For each method, we present two views of the reconstructed results. SHERT demonstrates the ability to handle challenging poses while providing clear details of facial and hand geometry. + +![](images/4d63ebdbc0db070e4904e914a70449e7c70ff2515c57cedc403cf020ee56f5ae.jpg) +Figure 10. Qualitative comparison for registration on THuman2.0. We compare the registration quality of various methods including N-ICP [6], RPTS [35], SVR- $\mathcal{L}_0$ [22], Fast-RNRR [68], DR [45], MDA [30] and ours SNS. The holes present in our result are the eliminated faces, as described in Sec. 3.2. The results indicate that SNS exhibits excellent performance in terms of model details, mesh quality, and registration robustness. The quantitative comparisons are shown in Tab. 2. + +we rotate the meshes every 60 degrees, resulting in a total of 2994 different orientations. We utilized the ICP [7] algorithm to transfer the color of the THuman2.0 scans to the vertices of our completed result, thereby generating 499 UV textures for training and obtaining 2994 visible UV masks from the rotated meshes. The complete UV textures and visible UV masks are randomly combined as inputs for our texture diffusion. + +Testing data. We conduct quantitative and qualitative evaluations on CAPE [43], THuman2.0, and in-the-wild images. We use CAPE-NFP [63] (100 samples with 3 viewpoints for each), and the last 27 subjects of THuman2.0 scans (6 viewpoints, each differing by 60 degrees). + +![](images/e996b6a68ea71bfcf26ae30de643caebc4139e97048c7f70e9d386bde7c553de.jpg) +Figure 11. The ablation results (with face substitution). We present the results after completion and refinement. With an increasing number of refinement iterations, the details of the mesh will be enhanced. Please zoom-in to see more details. + +
MethodGPUP2S↓Chamfer↓G-avg↑θ<30°↓Time
N-ICP [6]-0.2130.1630.50655.27m 23s
RPTS [35]-0.4880.3600.56548.31m 55s
SVR-L0 [22]-0.4040.2960.53153.51h 23m 32s
Fast-RNRR [68]-0.1150.0970.59731.11m 4s
DR [45]✓0.3390.3470.58147.716m 17s
MDA [30]✓0.6710.7310.58748.54m 48s
Ours (SNS)-0.1070.0780.72917.323s
Ours (Comp)✓0.1390.1670.66228.927s (23 + 4)
+ +Table 2. Quantitative evaluation for registration on THuman2.0. We test all the methods on the first subject of THuman2.0. Following the previous researches [21, 38], we adapt G-avg as a method for evaluating the mesh quality. We also report the metric $\theta < 30^{\circ}$ , which denotes the percentage of triangle meshes in the given mesh that have an angle less than 30 degrees. Additionally, we present the metrics for our complete mesh. + +Networks. The completion net and refinement net are both trained for 100 epochs with a learning rate of $1 \times 10^{-6}$ . The resolutions of the input and output data, including the UV position maps, images, masks and front-back normal maps, are all $1024 \times 1024 \times 3$ . During inference, the ECON's predicted front-back normals $(512 \times 512 \times 3)$ are upsampled to $1024 \times 1024 \times 3$ using bilinear interpolation. The texture diffusion is trained for 1 epoch with a learning rate of $2 \times 10^{-5}$ . The sampler of texture diffusion is DDIM [58]. We use 30 steps by default and infer the texture UV map with a resolution of $1024 \times 1024 \times 3$ . All the networks are trained on three NVIDIA RTX 3090 GPUs. + +![](images/75314ed9458309b8171b568893726d227123995c068b539f6f74501d22424320.jpg) +Figure 12. Qualitative comparison for texture prediction on in-the-wild image. We display the front and back view rendering results for each method. Since PIFu [54] and PaMIR [70] predict vertex colors, we only exhibit the texture maps of DINAR [59] and SHERT (ours). Please zoom-in to see more details. + +# 4.2. Evaluation + +Quantitative comparisons. We conduct quantitative comparisons with mainstream state-of-the-art monocular image reconstruction approaches in Tab. 1. As in previous work [23, 55, 63, 64], we report the point-to-surface Euclidean distance (P2S, cm), the Chamfer Distance (cm), and the Normals difference (L2). To ensure a fair comparison, $\mathrm{PIFu}^*$ [54] and $\mathrm{PaMIR}^*$ [70] are re-implemented and restrained on THuman2.0, using the same settings as ICON [63]. The ground-truth SMPL/SMPL-X models are provided for evaluation. However, PIFu [54], PIFuHD [55] and 2K2K [23] do not utilize the parametric body priors, which may result in subpar performance. In Tab. 1, "ICON-comp" refers to the completion result achieved by leveraging ICON's [63] prediction, while "ECON-refine" denotes the refinement mesh obtained using ICON's [64] result and the predicted front-back normal maps. Additionally, we evaluate the registration quality of SNS against state-of-the-art non-rigid registration methods, as presented in Tab. 2 + +Qualitative comparisons. We demonstrate a comparison between SHERT and state-of-the-art methods using in-the-wild images, with a focus on monocular image reconstruction (refer to Fig. 9) and texture prediction (refer to Fig. 12). Additionally, we compare the registration quality of our SNS with state-of-the-art non-rigid registration approaches on Thuman2.0 (refer to Fig. 10). + +# 4.3. Limitations + +Due to the geometric limitations of SMPL-X, SHERT performs weaker in reconstructing loose clothing, shoes and hair compared to implicit-based reconstruction methods. It is also difficult to ensure consistent results of texture diffusion at the seams of UVs. See more in SupMat. + +# 5. Applications + +SHERT uses the skinning weights from SMPL-X to enable animated poses, expressions, and gestures on the reconstructed mesh through LBS [34] (refer to Fig. 13). It allows for both global texture repainting (refer to Fig. 14) + +![](images/a45b655e8185b81bcb2405d876487bd2ece54e322bb0c43fc70b30c7a281ff49.jpg) +Figure 13. Animation results. Please zoom-in to see more details. + +![](images/835cacc056a247a8e631bbeef29df6b0baf5d38baa09813f984f648f37e3efc9.jpg) +Figure 14. Global texture repainting. SHERT can repaint the texture through text prompts. Please zoom-in to see more details. + +![](images/20372b62f2cc4547d835a8c642eba666764096966a22dfacc7bc14a12a9cef84.jpg) +Figure 15. Localized texture repainting. SHERT can repaint the masked area through text prompts. Please zoom-in to see more details. + +and the option for users to provide custom masks and text prompts for localized texture repainting (refer to Fig. 15). + +# 6. Conclusion + +We propose SHERT, which reconstructs a fully textured semantic human avatar from a detailed surface or a monocular image. It takes advantage of the geometric details of the target surface, along with semantic information and prior knowledge of the semantic guider. The reconstructed results have high-fidelity clothing details, high-quality triangle meshes, clear facial features, and complete hands geometry. SHERT is also capable of generating high-resolution texture maps with stable UV unwrapping. This approach bridges existing monocular reconstruction work and downstream industrial applications, and we believe it can promote the development of human avatars. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (62032011) and the Natural Science Foundation of Jiangsu Province (No. BK20211147). + +# References + +[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2018. 1, 2 +[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In IEEE Conf. 3D Vis.(3DV), 2018. 1 +[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single rgb camera. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019. +[4] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4 +[5] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2 +[6] Brian Amberg, Sami Romdhani, and Thomas Vetter. Optimal step nonrigid icp algorithms for surface registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2007. 3, 7 +[7] P.J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 1992. 6, 7 +[8] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), 1999. 3 +[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Trans. Vis. Comput. Graph.(TVCG), 2014. 3 +[10] Kunyao Chen, Fei Yin, Bang Du, Baichuan Wu, and Truong Q. Nguyen. Efficient registration for human surfaces via isometric regularization on embedded deformation. IEEE Trans. Vis. Comput. Graph.(TVCG), 2022. 3 +[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 4 +[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 2 +[13] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2 +[14] Radek Danecek, Michael J. Black, and Timo Bolkart. EMOCA: Emotion driven monocular face capture and animation. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 3, 5 + +[15] Bailin Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. Computer Graphics Forum (Eurographics 2022 State-of-the-Art Reports), 2022. 3 +[16] Jiankang Deng, Shiyang Cheng, Niannan Xue, Yuxiang Zhou, and Stefanos Zafeiriou. Uv-gan: Adversarial facial uv map completion for pose-invariant face recognition. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2018. 4 +[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3 +[18] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. ACM Trans. Graph.(TOG), 2021. 3 +[19] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Proceedings, 2022. 2 +[20] David A Field. Laplacian smoothing and delaunay triangulations. Communications in applied numerical methods, 1988. 5 +[21] Pascal J Frey and Houman Borouchaki. Surface mesh quality evaluation. International journal for numerical methods in engineering, 1999. 7 +[22] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Int. Conf. Comput. Vis.(ICCV), 2015. 3, 7 +[23] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 6, 8 +[24] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. $\mathrm{ARCH} + +$ : Animation-ready clothed human reconstruction revisited. In Int. Conf. Comput. Vis.(ICCV), 2021. 1, 2 +[25] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2 +[26] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable reconstruction of clothed humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2 +[27] Huggingface. Stable-diffusion-inpainting, 2022. https://huggingface.co/runwayml/stable-diffusion-inpainting.6 +[28] Henrik Wann Jensen. Realistic image synthesis using photon mapping. AK Peters/crc Press, 2001. 4 +[29] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In Eur. Conf. Comput. Vis.(ECCV), 2020. 2 +[30] Yucheol Jung, Hyomin Kim, Gyeongha Hwang, Seung-Hwan Baek, and Seungyong Lee. Mesh density adapta + +tion for template-based shape reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 7 +[31] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph.(TOG), 2013. 4 +[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. +[33] Vladislav Kraevoy and Alla Sheffer. Template-based mesh completion. In Symposium on Geometry Processing, 2005. 4 +[34] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of SIGGRAPH 2000, 2000. 8 +[35] Kun Li, Jingyu Yang, Yu-Kun Lai, and Daoliang Guo. Robust non-rigid registration with reweighted position and transformation sparsity. IEEE Trans. Vis. Comput. Graph.(TVCG), 2018. 3, 7 +[36] Tianye Li, Timo Bolkart, Michael Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph.(TOG), 2017. 3 +[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Trans. Graph.(TOG), 2017. 3 +[38] Yuanqi Li, Jianwei Guo, Xinran Yang, Shun Liu, Jie Guo, Xiaopeng Zhang, and Yanwen Guo. Deep point cloud simplification for high-quality surface reconstruction. arXiv preprint arXiv:2203.09088, 2022. 7 +[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2 +[40] Peter Liepa. Filling holes in meshes. In Proceedings of the 2003 Eurographics/ACM SIGGRAPH symposium on Geometry processing, 2003. 4 +[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graph.(TOG), 2015. 2, 3 +[42] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. Seminal graphics: pioneering efforts that shaped the field, 1998. 4 +[43] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 6, 7 +[44] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2 +[45] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph.(TOG), 2021. 3, 7 +[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, + +and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2019. 2, 3 +[47] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, 2009. 3 +[48] Ken Perlin and Eric M Hoffert. Hypertexture. In Proceedings of the 16th annual conference on Computer graphics and interactive techniques, 1989. 4 +[49] Leonid Pishchulin, Stefanie Wuhrer, Thomas Helten, Christian Theobalt, and Bernt Schiele. Building statistical shape spaces for 3d human modeling. Pattern Recognition, 2017. 2 +[50] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2022. 6 +[52] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph.(TOG), 2017. 3 +[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention(MICCAI), 2015, 2015. 5 +[54] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4, 6, 8 +[55] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2, 6, 8 +[56] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Ming-min Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In Eur. Conf. Comput. Vis.(ECCV), 2020. 3 +[57] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X- avatar: Expressive human avatars. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 +[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 7 +[59] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Int. Conf. Comput. Vis.(ICCV), 2023. 2, 6, 8 +[60] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Int. Conf. Comput. Vis.(ICCV), 2021. 6 + +[61] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In IEEE Conf. 3D Vis.(3DV), 2020. 1 +[62] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph.(TOG), 2021. 2 +[63] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 1, 2, 6, 7, 8 +[64] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 5, 6, 8 +[65] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3d human shape and articulated pose models. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2 +[66] Jingyu Yang, Ke Li, Kun Li, and Yu-Kun Lai. Sparse non-rigid registration of 3d shapes. Computer Graphics Forum, 2015. 3 +[67] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-if: Uncertainty-aware human digitization via implicit distribution field. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2 +[68] Yuxin Yao, Bailin Deng, Weiwei Xu, and Juyong Zhang. Quasi-newton solver for robust non-rigid registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 3, 7 +[69] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 6 +[70] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 1, 2, 6, 8 +[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis.(ICCV), 2023. 6 +[72] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 4 +[73] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019. 2 +[74] Hao Zhu, Xinxin Zuo, Haotian Yang, Sen Wang, Xun Cao, and Ruigang Yang. Detailed avatar recovery from single image. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 2 +[75] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity gar + +ment mesh reconstruction from single images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2 +[76] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2016. 3 \ No newline at end of file diff --git a/2024/Semantic Human Mesh Reconstruction with Textures/images.zip b/2024/Semantic Human Mesh Reconstruction with Textures/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..29fd59e60ef29327d05f647257c3268d1f272da7 --- /dev/null +++ b/2024/Semantic Human Mesh Reconstruction with Textures/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2705197b3f2cd136557b57adca583a53f6b95ee00e609d09daa438046463145a +size 897225 diff --git a/2024/Semantic Human Mesh Reconstruction with Textures/layout.json b/2024/Semantic Human Mesh Reconstruction with Textures/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..49dfe40b360cb7d05918bdaf7a5694af6a4cb850 --- /dev/null +++ b/2024/Semantic Human Mesh Reconstruction with Textures/layout.json @@ -0,0 +1,9321 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 131, + 103, + 463, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 103, + 463, + 119 + ], + "spans": [ + { + "bbox": [ + 131, + 103, + 463, + 119 + ], + "type": "text", + "content": "Semantic Human Mesh Reconstruction with Textures" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "spans": [ + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": "Xiaoyu Zhan" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": ", Jianxin Yang" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": ", Yuanqi Li" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": ", Jie Guo" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": ", Yanwen Guo" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": ", and Wenping Wang" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": " Nanjing University " + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 144, + 518, + 174 + ], + "type": "text", + "content": " Texas A&M University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 175, + 550, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 550, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 550, + 202 + ], + "type": "text", + "content": "{zhanxy, jianxin-yang, yuanqili}@smail.nju.edu.cn, {guojie, ywguo}@nju.edu.cn, wenping@tamu.edu https://zhanxy.xyz/projects/shert" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 208, + 542, + 411 + ], + "blocks": [ + { + "bbox": [ + 51, + 208, + 542, + 411 + ], + "lines": [ + { + "bbox": [ + 51, + 208, + 542, + 411 + ], + "spans": [ + { + "bbox": [ + 51, + 208, + 542, + 411 + ], + "type": "image", + "image_path": "78a4fe5fe0d7ad43a6a35a99b377612be899fc9f4f38716f8cf85e7d3db20f2f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 411, + 546, + 465 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 546, + 465 + ], + "type": "text", + "content": "Figure 1. Semantic Human mEsh Reconstruction with Textures (SHERT): (a) Given a target surface and the corresponding semantic guider, (b) SHERT reconstructs a detailed semantic model, which has stable UV unwrapping and skinning weights with high-quality triangle meshes. (c) It can either sample a texture map from the target surface or generate from the text prompts. (d) Based on our semantic representation, SHERT allows for high-precision facial reconstruction and animation of the body, face, and hands. (e) Moreover, SHERT is capable of inferring a fully textured avatar from a monocular image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 481, + 192, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 481, + 192, + 495 + ], + "spans": [ + { + "bbox": [ + 143, + 481, + 192, + 495 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "The field of 3D detailed human mesh reconstruction has made significant progress in recent years. However, current methods still face challenges when used in industrial applications due to unstable results, low-quality meshes, and a lack of UV unwrapping and skinning weights. In this paper, we present SHERT, a novel pipeline that can reconstruct semantic human meshes with textures and high-precision details. SHERT applies semantic- and normal-based sampling between the detailed surface (e.g. mesh and SDF) and the corresponding SMPL-X model to obtain a partially sampled semantic mesh and then generates the complete semantic mesh by our specifically designed self-supervised completion and refinement networks. Using the complete semantic mesh as a basis, we employ a texture diffusion model to create human textures that are driven by both images and texts. Our reconstructed meshes have stable UV unwrapping, high-quality triangle meshes, and consistent semantic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 483, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 547, + 579 + ], + "type": "text", + "content": "information. The given SMPL-X model provides semantic information and shape priors, allowing SHERT to perform well even with incorrect and incomplete inputs. The semantic information also makes it easy to substitute and animate different body parts such as the face, body, and hands. Quantitative and qualitative experiments demonstrate that SHERT is capable of producing high-fidelity and robust semantic meshes that outperform state-of-the-art methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 609, + 387, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 387, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 387, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "content": "Recovering highly realistic details and textures of a human mesh from monocular images is crucial for various applications such as gaming, movies, cartoons, VR, virtual try-on, and digital avatars. Current approaches [1, 2, 4, 23, 24, 26, 39, 54, 55, 57, 61, 63, 64, 67, 70] primarily focus on the recovery of the geometric details that are associated with images, but their results are not yet practical for real-world ap" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "142" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "plications. Recent advancements in parametric and explicit clothing models [12, 19, 29, 50, 62, 75] have shown promise in clothing reconstruction. However, they have limitations in accurately fitting different geometries and details. Implicit reconstruction approaches [23, 54, 55, 63, 64, 67, 70] excel at capturing clothing details but perform poorly in hands and face reconstruction. They may also generate incomplete and geometrically inseparable results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 289, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 289, + 302 + ], + "type": "text", + "content": "In this work, our goal is to reconstruct fully textured semantic human meshes through given detailed surfaces and corresponding semantic guiders. Our semantic human mesh is complete, animatable, and edit-friendly for both users and designers. It ensures that each vertex has deterministic semantic information and predefined skinning weights. This allows for easy substitution and animation of different body parts such as the face, body, and hands. The semantic information also guarantees stable and reasonable UV unwrapping, which is advantageous for editing and image-based texture generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 302, + 289, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 289, + 542 + ], + "type": "text", + "content": "Specifically, we propose SHERT, which generates a semantic human mesh from the detailed surface and its corresponding SMPL-X [46] model and optionally infers textures from images or colored surfaces. SHERT has four main processes. 1) Sampling, SHERT applies semantic and normal-based sampling to obtain a partially sampled semantic mesh based on the input detailed surface and SMPL-X model. We subdivide the original SMPL-X model to better capture the human details. 2) Completion, a self-supervised network is proposed to complete the partially sampled mesh. The network works in the 2D UV domain, which converts the 3D completion task into a 2D inpainting task. 3) Refinement, the origin image and front-back normal maps are used to enhance the geometry details. 4) Texture, our semantic human mesh is projected to the origin image in order to generate the partial texture map. Then we adapt the diffusion framework for text-driven partial texture inpainting and generation. The powerful generation ability of pre-trained diffusion model enables textures with rich clothing details and clear facial expressions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 543, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 288, + 604 + ], + "type": "text", + "content": "Extensive experiments have been conducted on both datasets and in-the-wild images. The quantitative and qualitative results show the robustness and superior performance of SHERT in reconstructing high-fidelity semantic human meshes and generating various high-resolution textures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 604, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 288, + 628 + ], + "type": "text", + "content": "In summary, the main contributions of this work include four-fold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 629, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 629, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 701 + ], + "type": "text", + "content": "- We introduce SHERT, a novel pipeline to reconstruct high-quality semantic human meshes from the detailed 3D surfaces, represented either explicitly as meshes, or implicitly as signed distance field (SDF). SHERT is also capable of predicting robust and fully textured avatars with high-fidelity faces from monocular images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "- We propose a semantic- and normal-based sampling" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 547, + 120 + ], + "type": "text", + "content": "method (SNS) and a self-supervised mesh completion network to achieve non-rigid 3D surface registration. The approach has the capability to process incomplete and inaccurate inputs by leveraging SMPL-X human priors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 121, + 547, + 204 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 121, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 547, + 167 + ], + "type": "text", + "content": "- We present a self-supervised mesh refinement network working in the UV domain. It utilizes the images and front-back normal maps to improve the geometric mesh details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 167, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 547, + 204 + ], + "type": "text", + "content": "- We use a diffusion model to infer high-resolution human textures from input images. The model can also accomplish text-driven texture inpainting and generation." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 392, + 229 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 238, + 509, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 509, + 250 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 509, + 250 + ], + "type": "text", + "content": "2.1. Monocular 3D Human Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 257, + 547, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 580 + ], + "type": "text", + "content": "Clothed Human Reconstruction. Most of the current CNN-based monocular 3D clothed human shape estimation methods can be divided into two categories: explicit-based [1-4, 29, 73, 74] and implicit-based [5, 12, 23-26, 39, 54, 55, 57, 63, 64, 67, 70] approaches, based on their representations. Explicit-based approaches usually infer the 3D offsets on top of the parametric human model [41, 46, 49, 65]. However, these methods are difficult to apply to flexible human topologies and cannot capture details well. Implicit-based approaches have the advantage of representing arbitrary 3D clothed human shapes that are free from the limitations of parametric human models. There are also some works [13, 25, 39, 63, 64, 70] that mix the implicit and explicit representations to achieve the detailed and robust 3D clothed human reconstruction. But both implicit-based and mixed methods still cannot maintain the stability of the human body shape well, often resulting in problems such as blurring and missing local body parts in the predicted results. Some methods [12, 19, 29, 50, 62, 75] propose additional parametric or implicit clothing models to fit loose clothes. These clothing models often struggle to accurately capture the details present in the images. Nevertheless, they are still useful for generating rough approximations of the clothing. Despite considerable progress made in monocular human reconstruction capture, there are still certain constraints, especially in terms of reliability, availability, and user-friendliness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 581, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 700 + ], + "type": "text", + "content": "Texture Prediction. Previous works [5, 13, 24-26, 44, 54, 57, 59] have also focused on predicting image-based human textures. However, the current emphasis of implicit-based methods [5, 13, 24-26, 54] is still on predicting vertex colors, which cannot be easily converted into usable texture maps for industrial applications. DINAR [59] proposes neural textures for modeling human avatars and achieves texture inpainting by utilizing the diffusion framework. Nevertheless, the neural texture is highly coupled with human avatars, and the resolution is relatively low." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "type": "text", + "content": "Face Reconstruction. In recent years, parameterized face" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "143" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 281 + ], + "type": "image", + "image_path": "dfcc98ab90beb3fbca00a919737c29c38021b7a01a830b65db5843c02659f70d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": "Figure 2. Overview of SHERT for monocular image reconstruction. Given an RGB image " + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": ", SHERT first infers the detailed mesh " + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": " and corresponding sub-SMPLX model " + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_x" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": ". It then applies SNS to obtain the partial semantic mesh " + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": " (in Sec. 3.2). The Completion Net infers " + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": " by filling the UV holes in " + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": " (in Sec. 3.3). Image and normal maps are utilized to generate " + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r" + }, + { + "bbox": [ + 46, + 285, + 547, + 340 + ], + "type": "text", + "content": ", which contains sharper geometry details (in Sec. 3.4). Finally, SHERT uses a diffusion model to achieve text-driven texture inpainting and generation (in Sec. 3.5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 346, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 289, + 441 + ], + "type": "text", + "content": "models [8, 9, 36, 47] have been widely used in monocular face reconstruction algorithms [14, 17, 18, 56, 76], contributing to the success of achieving realistic high-quality reconstruction results. In light of the incorporation of FLAME [36] into SMPL-X [46], we can utilize high-quality facial reconstruction results from previous works to enhance the accuracy and realism of the SMPL-X based human body reconstruction results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 192, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 192, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 192, + 462 + ], + "type": "text", + "content": "2.2. Non-rigid 3D Registration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 575 + ], + "type": "text", + "content": "Non-rigid 3D surface registration approaches [6, 10, 22, 30, 35, 45, 66, 68] usually compute a deformation that aligns a source surface with a target surface [15]. SHERT aims to transfer the semantic information of parametric human models to the corresponding detailed surfaces through registration. However, existing methods mainly focus on the precise registration of the source and target, while the target detailed human surfaces in real-world tasks often have incomplete and erroneous data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 103, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 103, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 103, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "SHERT is capable of reconstructing a high-fidelity fully textured semantic human mesh based on a pair of detailed 3D surface and corresponding SMPL-X [46] model. Our results have high robustness and will not result in missing body parts in despite of the incomplete inputs. Furthermore, the semantic mesh ensures that each vertex has deterministic semantic information and predefined skinning weights, making it possible to replace and animate the human face, body, and hands. SHERT can also generate realistic hu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 346, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 479 + ], + "type": "text", + "content": "man textures from texts and images. These features make it easy for users and designers to use or further edit our results. To achieve this, SHERT first subdivides the SMPL-X model to improve the accuracy in capturing human details (in Sec. 3.1) and then applies the semantic- and normal-based sampling to obtain a partially sampled mesh (in Sec. 3.2). We complete the partial result using a self-supervised mesh completion network (in Sec. 3.3), and finally enhance the geometry details through refinement (in Sec. 3.4). In addition, SHERT uses a diffusion model for text-driven human texture inpainting and generation (in Sec. 3.5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 489, + 457, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 489, + 457, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 489, + 457, + 501 + ], + "type": "text", + "content": "3.1. Subdivided SMPL-X model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 509, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 604 + ], + "type": "text", + "content": "In this work, we use SMPL-X as the semantic guider. SMPL-X is a parameterized body model that combines SMPL [41] with the FLAME [37] head model and the MANO [52] hand model. It is parameterized with shape and pose, has 10, 475 vertices and 54 joints, including joints for the neck, jaw, eyeballs, and fingers. The SMPL-X model provides basic skinning weights and corresponding semantic information for each vertex." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": "SHERT defines the sub-SMPLX based on SMPL-X since 10,475 vertices are not sufficient to accurately represent the details of the human body and clothing, such as facial details and clothing wrinkles. Taking into consideration both the expressiveness of the model and computational cost, we apply the mid-point subdivision algorithm twice on a standard SMPL-X model (with eyeballs removed) to obtain the sub-SMPLX with 149,921 vertices and 299,712 faces." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "144" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 280, + 150 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 280, + 150 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 280, + 150 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 280, + 150 + ], + "type": "image", + "image_path": "21d0bec775a267e2e481f6bb85b3278d5a9ca4a728769cd55e4f9bae5846c565.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 154, + 289, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 289, + 209 + ], + "type": "text", + "content": "Figure 3. Explicit and implicit sampling. a) SNS shoots a ray from the starting vertex along the vertex normal to locate the intersection point with the target surface. b) SNS takes samples at fixed intervals along the vertex normal ray and search for the point that is closest to the zero isosurface." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 214, + 284, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 284, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 284, + 227 + ], + "type": "text", + "content": "3.2. Semantic- and Normal-based Sampling (SNS)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "text", + "content": "Our objective is to accurately predict the non-rigid deformation between a source surface and a target surface. Given a source surface " + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_x" + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "text", + "content": ", which is represented as a sub-SMPLX in our work, and the target surface " + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "text", + "content": ", we learn a mapping function " + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "inline_equation", + "content": "D: \\mathcal{M}_x^{149921 \\times 3} \\to \\mathcal{M}_s^{149921 \\times 3}" + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s" + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "text", + "content": " can be aligned with " + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 46, + 233, + 287, + 317 + ], + "type": "text", + "content": " through the learned mapping in geometry, and ignoring the incorrect parts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "content": "Specifically, we obtain a partially sampled mesh through a sampling scheme based on the vertex normals " + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "inline_equation", + "content": "N^{149921 \\times 3}" + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "content": ". Starting from a point on " + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_x" + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "content": ", we search for the intersection point with the target surface " + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "content": " along the vertex normal ray. We can extend the sampling scheme to implicit surfaces with constant step Ray Marching [28, 48], shown in Fig. 3 and Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": "It should be noted that SNS does not always return satisfactory results since the ray may not intersect with the target surface. We need to locate and label the vertices that have failed to register. In addition, due to the large geometric differences and the possibility of incorrect alignment, the sampling scheme may retrieve incorrect results, which should also be detected. In order to remove incorrect points and triangle meshes from the sampled results, we calculate " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": " (the angle between the normal vectors of the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": " (the area ratio between the sampled triangle mesh and the corresponding sub-SMPLX triangle mesh), and " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": " (the edge ratio between the longest and shortest edges of the sampled triangle mesh) for each triangle mesh in the sampled result. We then perform mesh culling by removing low-quality triangle meshes that have at least one indicator exceeding the threshold among the three mentioned above. Finally, we perform a connectivity check on the processed sampled mesh and remove the sets with a number of connected triangle meshes less than " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": ". By default, we set " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\theta = 2" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "s = 3" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "r = 3" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "inline_equation", + "content": "g = 500" + }, + { + "bbox": [ + 46, + 402, + 288, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "We found that culling meshes only in the current pose would result in unreasonable deformations in human animation tasks. Therefore, SNS repeats the processing above in the canonical space. This ensures that we remove almost all unreasonable sampling results and obtain the partial se" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 539, + 157 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 539, + 157 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 539, + 157 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 539, + 157 + ], + "type": "image", + "image_path": "be20819802bf73bbd27d596bd85d23e3300a1b077f7db0c3d73a3faedc289789.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 172, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 547, + 235 + ], + "type": "text", + "content": "of SNS and the Marching Cubes algorithm [42] in the implicit field predicted by the multi-view PIFu [54]. The results show that our SNS and completion network correctly reconstruct the implicit field and obtain a smoother surface compared to Marching Cubes. mantic mesh " + }, + { + "bbox": [ + 304, + 172, + 547, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s" + }, + { + "bbox": [ + 304, + 172, + 547, + 235 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 240, + 486, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 240, + 486, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 486, + 254 + ], + "type": "text", + "content": "3.3. Self-supervised Mesh Completion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "text", + "content": "In Sec. 3.2, the vertices that failed to register have resulted in holes in " + }, + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "text", + "content": ". Currently, there are many mesh completion algorithms [11, 31-33, 40] that can fill in these mesh holes. However, these algorithms cannot be well adapted to our semantic reconstruction task due to the strong constraints on the number and relative positions of vertices in our representation in Sec. 3.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 510 + ], + "type": "text", + "content": "Therefore, we transform the partially sampled mesh into the UV domain [4, 16, 72] according to the semantic information. As a result, we convert the mesh completion task in 3D space into an inpainting task on a 2D image. Then, we design a self-supervised completion network that can fill in the holes of the partially sampled mesh. As shown in Fig. 5, by adding a random hole mask " + }, + { + "bbox": [ + 304, + 342, + 546, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_r" + }, + { + "bbox": [ + 304, + 342, + 546, + 510 + ], + "type": "text", + "content": " on the partial UV position map, we can obtain trainable pairs, which has a consistent distribution as the missing parts of the sampled mesh. The manually masked parts can provide the network with consistent supervision information as the input parts. This completion network is capable of generating robust meshes while maintaining both mesh quality and semantic consistency in the final results (refer to Fig. 6)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 630 + ], + "type": "text", + "content": "To further decrease the learning difficulties, we transform all incompletely sampled meshes to canonical space when facing the diverse range of poses and clothing styles. We believe that this approach has the potential to mitigate the challenges in problem-solving, as completion is no longer affected by pose and body shape. Meanwhile, since the sampled points are located on the normal ray of the sub-SMPLX vertices, we represent the deformation relative to sub-SMPLX as an offset based on the vertex normal. The transformed UV position map " + }, + { + "bbox": [ + 304, + 510, + 547, + 630 + ], + "type": "inline_equation", + "content": "\\overline{S}" + }, + { + "bbox": [ + 304, + 510, + 547, + 630 + ], + "type": "text", + "content": " can be presented as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 637, + 545, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 637, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 378, + 637, + 545, + 665 + ], + "type": "interline_equation", + "content": "\\bar {d} = \\frac {S _ {\\text {s a m p l e}} - S _ {\\text {p o s e}}}{N _ {\\text {p o s e}}}, \\tag {1}", + "image_path": "98e4ea9e47756c4880e5fb2bda7f7c8e0c674d2ef6a92a47739027cffd4f1ab6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 671, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 671, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 376, + 671, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\bar {S} = S _ {\\text {c a n o}} + N _ {\\text {c a n o}} \\cdot \\bar {d}, \\tag {2}", + "image_path": "d2d8b819b10fae8179eb99603e7b13762d50331ebb5ffbe4e30ec94a976389ef.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{\\text{sample}}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " denotes the partial UV position map of the sampled mesh. " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{\\text{pose}}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "S_{\\text{cano}}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " refer to the UV position" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "145" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 276, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 276, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 276, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 276, + 198 + ], + "type": "image", + "image_path": "59c617ab77159f431c737eeada4b8c40254e1fd9ca69b9847bea1c9d24454f2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 287, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 233 + ], + "type": "text", + "content": "Figure 5. Completion network. The completion network transforms the result of SNS into canonical space and predicts the holes in the UV domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 276 + ], + "type": "text", + "content": "maps of the original posed sub-SMPLX and the canonical sub-SMPLX respectively. " + }, + { + "bbox": [ + 46, + 241, + 287, + 276 + ], + "type": "inline_equation", + "content": "N_{\\text{pose}}" + }, + { + "bbox": [ + 46, + 241, + 287, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 241, + 287, + 276 + ], + "type": "inline_equation", + "content": "N_{\\text{cano}}" + }, + { + "bbox": [ + 46, + 241, + 287, + 276 + ], + "type": "text", + "content": " are the corresponding UV normal maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": "The complete UV position map " + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": " and the input combined hole mask " + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{sup}" + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": " are calculated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 312, + 287, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 312, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 287, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {s u p} = \\mathcal {H} _ {r} \\cdot \\left(1 - \\mathcal {H} _ {o}\\right) + \\mathcal {H} _ {o}, \\tag {3}", + "image_path": "2a5e175a256b93f7095e2c4bbbd516ca57b24bd16464a889ef0504e5e3fd2ae0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 350, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 350, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 72, + 350, + 287, + 365 + ], + "type": "interline_equation", + "content": "S = S _ {c a n o} + N _ {c a n o} \\cdot \\left[ d \\cdot \\mathcal {H} _ {o} + \\bar {d} \\cdot \\left(1 - \\mathcal {H} _ {o}\\right) \\right], \\tag {4}", + "image_path": "55e54e53147f979bab5281c6af37b48678d5b4b1d97efaed5f33a633f87638bd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_r" + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_o" + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "text", + "content": " denote the randomly added hole mask and the hole mask generated by SNS respectively. " + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "inline_equation", + "content": "d\\in" + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "text", + "content": " is the estimated displacement UV map. The complete mesh " + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "text", + "content": " is resampled from " + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 373, + 287, + 432 + ], + "type": "text", + "content": " and then transformed to the original pose space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 456 + ], + "type": "text", + "content": "The primary loss used in our completion network is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 287, + 491 + ], + "type": "interline_equation", + "content": "S _ {p} = S _ {\\text {c a n o}} + N _ {\\text {c a n o}} \\cdot [ d \\cdot \\mathcal {H} _ {\\sup } + \\bar {d} \\cdot (1 - \\mathcal {H} _ {\\sup }) ], \\tag {5}", + "image_path": "a6da567ad630b4eece0310dd4d1f8cffbbe5b1924ed5561af78d4872e69196cb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 508, + 287, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 508, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 89, + 508, + 287, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n p} = \\frac {\\left\\| \\left(S _ {p} - \\bar {S}\\right) \\cdot \\left(\\mathcal {H} _ {s u p} - \\mathcal {H} _ {o}\\right) \\right\\| _ {2} ^ {2}}{\\sum_ {i , j} \\left(\\mathcal {H} _ {s u p} - \\mathcal {H} _ {o}\\right) _ {i , j}}. \\tag {6}", + "image_path": "1d6e30d05bb981ca93b4f1de54542215056d1c1b18c15195004fd21fd2197e14.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 544, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 626 + ], + "type": "text", + "content": "We find that completing the face, hands, and feet is much more difficult than other parts of the body, so we use the corresponding parts of sub-SMPLX for replacement. Furthermore, SHERT allows for the use of FLAME-based methods such as EMOCA [14] to replace the facial region in our completion results, enhancing the accuracy and realism of the facial details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 635, + 225, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 225, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 225, + 647 + ], + "type": "text", + "content": "3.4. Self-supervised Mesh Refinement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "After completing the mesh, we have already reconstructed a semantic mesh with extremely high levels of detail. However, some geometric details may be lost during the SNS and completion process. Therefore, we design an additional self-supervised refinement network to further optimize the" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 541, + 138 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 541, + 138 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 541, + 138 + ], + "type": "image", + "image_path": "0182e7005cdd47679cad32888c6302e506948c9beb737252f3ee5d78528abb54.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 140, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 217 + ], + "type": "text", + "content": "Figure 6. The robustness of SNS and Completion Network. We present our complete meshes, which are reconstructed using the predictions of ECON [64] from in-the-wild images. Distinguishing itself from previous registration methods, SHERT has the capability to effectively process scenarios where the inputs are incomplete, inaccurate, or contain errors by leveraging SMPL-X human priors." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 222, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 310, + 222, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 310, + 222, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 310, + 222, + 545, + 323 + ], + "type": "image", + "image_path": "25b4c87d12205b62882c53201b1fa3250df46a7927fdc10e17bed4fb3613be52.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 327, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 371 + ], + "type": "text", + "content": "Figure 7. Refinement network. The features extracted from the image and front-back normal maps are projected to the UV domain. These projected features are subsequently combined with the input UV position map to generate a refined mesh." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 380, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 498 + ], + "type": "text", + "content": "mesh's details (e.g. cloth wrinkles, subtle deformations of body movements) using the image and normal maps. The input normal maps can be obtained either by rendering the scanned model or by using prediction networks. Our refinement network follows a U-Net [53] architecture and is capable of predicting a displacement UV map, denoted as " + }, + { + "bbox": [ + 305, + 380, + 545, + 498 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 305, + 380, + 545, + 498 + ], + "type": "text", + "content": ", in order to optimize the complete mesh " + }, + { + "bbox": [ + 305, + 380, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 305, + 380, + 545, + 498 + ], + "type": "text", + "content": ". Since the input image and normal maps are not in the UV domain, we utilize a separate network to extract the image domain features and then project them to UV domain." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 499, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 499, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 305, + 499, + 545, + 570 + ], + "type": "text", + "content": "In order to guide the network in learning the potential correspondence between the input features and mesh details, we apply Laplacian Smoothing [20] on the existing complete meshes, as shown in Fig. 7. The smoothed mesh " + }, + { + "bbox": [ + 305, + 499, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_l" + }, + { + "bbox": [ + 305, + 499, + 545, + 570 + ], + "type": "text", + "content": " is used as input to the network, and the original mesh " + }, + { + "bbox": [ + 305, + 499, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 305, + 499, + 545, + 570 + ], + "type": "text", + "content": " is used for supervision." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "text", + "content": "Given the image " + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "text", + "content": ", front-view normal map " + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_f" + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "text", + "content": ", backview normal map " + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_b" + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "text", + "content": ", and the corresponding complete UV position map " + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "inline_equation", + "content": "S_l" + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "text", + "content": ", the projected feature " + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 571, + 545, + 618 + ], + "type": "text", + "content": " is typically represented as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 365, + 633, + 545, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 633, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 365, + 633, + 545, + 646 + ], + "type": "interline_equation", + "content": "F = \\mathcal {P} \\left(\\mathcal {F} \\left(\\mathcal {I}, \\mathcal {N} _ {f}, \\mathcal {N} _ {b}\\right), S _ {l}, c\\right), \\tag {7}", + "image_path": "c54e81d262efef4d5a812c92f1bc8507864b1c525e5cec01674b7d11e3edf68d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " represents the camera parameters, " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " denotes the image domain feature extraction network, and " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " is the projection function that can project the image domain features to the UV domain according to the camera parameters and the 3D coordinates of points on " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "146" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 318, + 188 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 318, + 188 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 318, + 188 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 318, + 188 + ], + "type": "image", + "image_path": "df8663b470ba65973de2791b24d60924a77afe1309f421bdb2c64c7267432914.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 547, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 233 + ], + "type": "text", + "content": "Figure 8. Texture diffusion. We use two strategies to finetune the diffusion model for text-driven texture inpainting and inpainting separately. During inference, instead of just using random noise, the encoded partial texture map and mask can be manually added to the latent features to preserve the original information of the input. Additionally, we use the super resolution method Real-esrgan [60], which can specifically optimize both the body and face, to further enhance the final outputs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 324, + 70, + 541, + 187 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 541, + 187 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 541, + 187 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 541, + 187 + ], + "type": "image", + "image_path": "fb73cb07c45c17ef49634b845d39cdc7973ff5c29d9d7afbb99b6d884e81328f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "text", + "content": "The UV position map " + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "text", + "content": " of the refined mesh " + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r" + }, + { + "bbox": [ + 46, + 236, + 288, + 261 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 272, + 287, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 272, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 129, + 272, + 287, + 285 + ], + "type": "interline_equation", + "content": "S _ {r} = S _ {l} + N _ {l} \\cdot z, \\tag {8}", + "image_path": "9d85433b88de8d71c56bd80fc0bad032833f3ca4106a75db433461f35c531be4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "text", + "content": " represents the predicted displacement UV map of the refine network, " + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "text", + "content": " are the UV position map and normal UV map of " + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_l" + }, + { + "bbox": [ + 47, + 289, + 287, + 326 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 327, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 327, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 327, + 287, + 338 + ], + "type": "text", + "content": "The loss we used in refine network can be represented as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 346, + 287, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 346, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 287, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s} = M S E \\left(z - \\frac {S _ {c} - S _ {l}}{N _ {l}}\\right), \\tag {9}", + "image_path": "6b87fdbf803265f1e5794e89738f620c669ce53c6ce41d51ba3efa32187363d1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 382, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 382, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = M S E \\left(N _ {r} - N _ {c}\\right), \\tag {10}", + "image_path": "db81781cddfcc38204ae8130963ed324a48a3da461e7bdacff7f0f900923e7cd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": " is the UV position map of " + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": " denote the normal UV maps of " + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r" + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 47, + 400, + 287, + 424 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "content": "Moreover, the refine network can be iteratively employed by users to enhance the mesh details according to their desired level (refer to Fig. 11)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 468, + 287, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 480 + ], + "type": "text", + "content": "3.5. Text-driven Texture Inpainting and Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "In Sec. 3.4, all the reconstructed meshes share the same semantic information from SMPL-X. This means that when the models are unfolded to the UV domain, vertices with the same semantic information will be projected to a fixed UV position, resulting in a stable texture map that allows for painting and transferring textures. To perform texture inpainting on incomplete human body textures and generate high-quality human body textures driven by text, we fine-tune the Stable Diffusion model [51] with ControlNet [71], similar to the approach taken by Dinar [59]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": "As our reconstructed semantic mesh closely matches the scanned model's geometry, we can use the ICP algorithm [7] to register the vertex colors from the scanned model onto our result. This enables us to convert the texture maps into the SMPL-X format. As shown in Fig. 8, we obtain a partial texture map from the input image based on the semantic mesh and camera parameters and hope to generate the invisible parts. ControlNet enable us to add conditional inputs (partial texture maps) to the generation of the diffusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 236, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 416 + ], + "type": "text", + "content": "network at a low cost. We use two strategies to finetune the diffusion model, distinguished by whether to take the encoded partial texture map and mask as the additional inputs [27]. During the inference process, we can directly input partial texture maps into the ControlNet and generate various results by adding random noise. Alternatively, we can enhance the preservation of existing information in images by concatenating the encoded partial texture maps and masks with the latent features. Although Stable Diffusion has not worked with UV-parameterized images before, the well-designed UV parameterization keeps the shapes of the face, body, and limbs stable, ensuring a learnable space for the model. The texture map for the body and limbs focuses more on color and patterns rather than shapes, also resulting in excellent outcomes when cropped with a mask." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 426, + 388, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 388, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 388, + 440 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 446, + 437, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 437, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 437, + 457 + ], + "type": "text", + "content": "4.1. Datasets and Networks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 464, + 546, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 525 + ], + "type": "text", + "content": "Training data. The completion and refinement networks are trained using the first 499 scans of THuman2.0 [69]. During the training of the completion network, we randomly choose one of the remaining 498 masks as a random hole mask. To enrich the inputs for the refinement network," + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 309, + 531, + 545, + 628 + ], + "blocks": [ + { + "bbox": [ + 309, + 531, + 545, + 628 + ], + "lines": [ + { + "bbox": [ + 309, + 531, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 309, + 531, + 545, + 628 + ], + "type": "table", + "html": "
MethodCAPE[43]THuman2.0[69]
P2S↓Chamfer↓Normal↓P2S↓Chamfer↓Normal↓
PIFu[54]*2.11371.65370.07552.54932.36400.1042
PIFuHD[55]3.78463.57870.10023.07723.18080.1207
PaMIR[70]*1.45201.22410.06101.54391.33110.1102
ICON[63]0.88550.86090.03471.03611.08740.0607
ECON[64]0.94030.93860.03741.13041.20810.0661
2K2K[23]†---2.53422.61650.1030
Ours (ICON-comp)0.8550↑0.8107↑0.03591.04591.0465↑0.0604↑
Ours (ICON-refine)0.86330.81120.03801.0442↑1.04680.0603↑
Ours (ECON-comp)0.8561↑0.8242↑0.03781.1255↑1.1420↑0.0672
Ours (ECON-refine)0.85810.8144↑0.03981.0630↑1.0430↑0.0649↑
", + "image_path": "082195cd63daa45fda7ab5594c4630a86d2459bfa85ed6dfcc24835325fde955.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 633, + 547, + 710 + ], + "lines": [ + { + "bbox": [ + 304, + 633, + 547, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 547, + 710 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation for monocular image reconstruction. We evaluate the performance of our completion results (comp) and refinement results (refine) by comparing them with state-of-the-art methods. * methods are re-implemented in [63] to ensure a fair comparison. † method has only been tested with human-facing-forward images. ↑ and ↑ indicate the improvement achieved through completion and refinement, respectively." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "147" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 547, + 187 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 547, + 187 + ], + "type": "image", + "image_path": "822937ec2201c677ed60b7c2329350140938f139eab8e7c52282e35ae6fec2a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 222 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison for monocular image reconstruction on in-the-wild image. For each method, we present two views of the reconstructed results. SHERT demonstrates the ability to handle challenging poses while providing clear details of facial and hand geometry." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 224, + 545, + 344 + ], + "blocks": [ + { + "bbox": [ + 48, + 224, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 48, + 224, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 545, + 344 + ], + "type": "image", + "image_path": "4d63ebdbc0db070e4904e914a70449e7c70ff2515c57cedc403cf020ee56f5ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 547, + 390 + ], + "type": "text", + "content": "Figure 10. Qualitative comparison for registration on THuman2.0. We compare the registration quality of various methods including N-ICP [6], RPTS [35], SVR-" + }, + { + "bbox": [ + 46, + 346, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_0" + }, + { + "bbox": [ + 46, + 346, + 547, + 390 + ], + "type": "text", + "content": " [22], Fast-RNRR [68], DR [45], MDA [30] and ours SNS. The holes present in our result are the eliminated faces, as described in Sec. 3.2. The results indicate that SNS exhibits excellent performance in terms of model details, mesh quality, and registration robustness. The quantitative comparisons are shown in Tab. 2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 495 + ], + "type": "text", + "content": "we rotate the meshes every 60 degrees, resulting in a total of 2994 different orientations. We utilized the ICP [7] algorithm to transfer the color of the THuman2.0 scans to the vertices of our completed result, thereby generating 499 UV textures for training and obtaining 2994 visible UV masks from the rotated meshes. The complete UV textures and visible UV masks are randomly combined as inputs for our texture diffusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 559 + ], + "type": "text", + "content": "Testing data. We conduct quantitative and qualitative evaluations on CAPE [43], THuman2.0, and in-the-wild images. We use CAPE-NFP [63] (100 samples with 3 viewpoints for each), and the last 27 subjects of THuman2.0 scans (6 viewpoints, each differing by 60 degrees)." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 54, + 569, + 282, + 665 + ], + "blocks": [ + { + "bbox": [ + 54, + 569, + 282, + 665 + ], + "lines": [ + { + "bbox": [ + 54, + 569, + 282, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 569, + 282, + 665 + ], + "type": "image", + "image_path": "e996b6a68ea71bfcf26ae30de643caebc4139e97048c7f70e9d386bde7c553de.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 666, + 287, + 709 + ], + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 709 + ], + "type": "text", + "content": "Figure 11. The ablation results (with face substitution). We present the results after completion and refinement. With an increasing number of refinement iterations, the details of the mesh will be enhanced. Please zoom-in to see more details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 398, + 547, + 483 + ], + "blocks": [ + { + "bbox": [ + 308, + 398, + 547, + 483 + ], + "lines": [ + { + "bbox": [ + 308, + 398, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 547, + 483 + ], + "type": "table", + "html": "
MethodGPUP2S↓Chamfer↓G-avg↑θ<30°↓Time
N-ICP [6]-0.2130.1630.50655.27m 23s
RPTS [35]-0.4880.3600.56548.31m 55s
SVR-L0 [22]-0.4040.2960.53153.51h 23m 32s
Fast-RNRR [68]-0.1150.0970.59731.11m 4s
DR [45]✓0.3390.3470.58147.716m 17s
MDA [30]✓0.6710.7310.58748.54m 48s
Ours (SNS)-0.1070.0780.72917.323s
Ours (Comp)✓0.1390.1670.66228.927s (23 + 4)
", + "image_path": "ee9d315e3ea598a21c809697dc6326846540710f3b8752a2bdada8ff81cf43f8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 490, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 545, + 568 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation for registration on THuman2.0. We test all the methods on the first subject of THuman2.0. Following the previous researches [21, 38], we adapt G-avg as a method for evaluating the mesh quality. We also report the metric " + }, + { + "bbox": [ + 304, + 490, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\theta < 30^{\\circ}" + }, + { + "bbox": [ + 304, + 490, + 545, + 568 + ], + "type": "text", + "content": ", which denotes the percentage of triangle meshes in the given mesh that have an angle less than 30 degrees. Additionally, we present the metrics for our complete mesh." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "content": "Networks. The completion net and refinement net are both trained for 100 epochs with a learning rate of " + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "content": ". The resolutions of the input and output data, including the UV position maps, images, masks and front-back normal maps, are all " + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "inline_equation", + "content": "1024 \\times 1024 \\times 3" + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "content": ". During inference, the ECON's predicted front-back normals " + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "inline_equation", + "content": "(512 \\times 512 \\times 3)" + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "content": " are upsampled to " + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "inline_equation", + "content": "1024 \\times 1024 \\times 3" + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "content": " using bilinear interpolation. The texture diffusion is trained for 1 epoch with a learning rate of " + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "content": ". The sampler of texture diffusion is DDIM [58]. We use 30 steps by default and infer the texture UV map with a resolution of " + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "inline_equation", + "content": "1024 \\times 1024 \\times 3" + }, + { + "bbox": [ + 304, + 570, + 546, + 712 + ], + "type": "text", + "content": ". All the networks are trained on three NVIDIA RTX 3090 GPUs." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "148" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 162 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 162 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 162 + ], + "type": "image", + "image_path": "75314ed9458309b8171b568893726d227123995c068b539f6f74501d22424320.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 163, + 288, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 288, + 217 + ], + "type": "text", + "content": "Figure 12. Qualitative comparison for texture prediction on in-the-wild image. We display the front and back view rendering results for each method. Since PIFu [54] and PaMIR [70] predict vertex colors, we only exhibit the texture maps of DINAR [59] and SHERT (ours). Please zoom-in to see more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 227, + 121, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 121, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 121, + 239 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 246, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 461 + ], + "type": "text", + "content": "Quantitative comparisons. We conduct quantitative comparisons with mainstream state-of-the-art monocular image reconstruction approaches in Tab. 1. As in previous work [23, 55, 63, 64], we report the point-to-surface Euclidean distance (P2S, cm), the Chamfer Distance (cm), and the Normals difference (L2). To ensure a fair comparison, " + }, + { + "bbox": [ + 46, + 246, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathrm{PIFu}^*" + }, + { + "bbox": [ + 46, + 246, + 287, + 461 + ], + "type": "text", + "content": " [54] and " + }, + { + "bbox": [ + 46, + 246, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}^*" + }, + { + "bbox": [ + 46, + 246, + 287, + 461 + ], + "type": "text", + "content": " [70] are re-implemented and restrained on THuman2.0, using the same settings as ICON [63]. The ground-truth SMPL/SMPL-X models are provided for evaluation. However, PIFu [54], PIFuHD [55] and 2K2K [23] do not utilize the parametric body priors, which may result in subpar performance. In Tab. 1, \"ICON-comp\" refers to the completion result achieved by leveraging ICON's [63] prediction, while \"ECON-refine\" denotes the refinement mesh obtained using ICON's [64] result and the predicted front-back normal maps. Additionally, we evaluate the registration quality of SNS against state-of-the-art non-rigid registration methods, as presented in Tab. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 546 + ], + "type": "text", + "content": "Qualitative comparisons. We demonstrate a comparison between SHERT and state-of-the-art methods using in-the-wild images, with a focus on monocular image reconstruction (refer to Fig. 9) and texture prediction (refer to Fig. 12). Additionally, we compare the registration quality of our SNS with state-of-the-art non-rigid registration approaches on Thuman2.0 (refer to Fig. 10)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 554, + 124, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 124, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 124, + 566 + ], + "type": "text", + "content": "4.3. Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 573, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 633 + ], + "type": "text", + "content": "Due to the geometric limitations of SMPL-X, SHERT performs weaker in reconstructing loose clothing, shoes and hair compared to implicit-based reconstruction methods. It is also difficult to ensure consistent results of texture diffusion at the seams of UVs. See more in SupMat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "type": "text", + "content": "5. Applications" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "SHERT uses the skinning weights from SMPL-X to enable animated poses, expressions, and gestures on the reconstructed mesh through LBS [34] (refer to Fig. 13). It allows for both global texture repainting (refer to Fig. 14)" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 536, + 142 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 536, + 142 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 536, + 142 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 536, + 142 + ], + "type": "image", + "image_path": "a45b655e8185b81bcb2405d876487bd2ece54e322bb0c43fc70b30c7a281ff49.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 144, + 545, + 154 + ], + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 154 + ], + "type": "text", + "content": "Figure 13. Animation results. Please zoom-in to see more details." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 156, + 545, + 279 + ], + "blocks": [ + { + "bbox": [ + 310, + 156, + 545, + 279 + ], + "lines": [ + { + "bbox": [ + 310, + 156, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 310, + 156, + 545, + 279 + ], + "type": "image", + "image_path": "835cacc056a247a8e631bbeef29df6b0baf5d38baa09813f984f648f37e3efc9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 279, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 306, + 279, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 279, + 545, + 300 + ], + "type": "text", + "content": "Figure 14. Global texture repainting. SHERT can repaint the texture through text prompts. Please zoom-in to see more details." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 300, + 538, + 416 + ], + "blocks": [ + { + "bbox": [ + 315, + 300, + 538, + 416 + ], + "lines": [ + { + "bbox": [ + 315, + 300, + 538, + 416 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 538, + 416 + ], + "type": "image", + "image_path": "20372b62f2cc4547d835a8c642eba666764096966a22dfacc7bc14a12a9cef84.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 415, + 545, + 447 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 447 + ], + "type": "text", + "content": "Figure 15. Localized texture repainting. SHERT can repaint the masked area through text prompts. Please zoom-in to see more details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 449, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 474 + ], + "type": "text", + "content": "and the option for users to provide custom masks and text prompts for localized texture repainting (refer to Fig. 15)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 483, + 378, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 378, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 378, + 496 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": "We propose SHERT, which reconstructs a fully textured semantic human avatar from a detailed surface or a monocular image. It takes advantage of the geometric details of the target surface, along with semantic information and prior knowledge of the semantic guider. The reconstructed results have high-fidelity clothing details, high-quality triangle meshes, clear facial features, and complete hands geometry. SHERT is also capable of generating high-resolution texture maps with stable UV unwrapping. This approach bridges existing monocular reconstruction work and downstream industrial applications, and we believe it can promote the development of human avatars." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 658, + 403, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 403, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 403, + 671 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (62032011) and the Natural Science Foundation of Jiangsu Province (No. BK20211147)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "149" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "text", + "content": "[2] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In IEEE Conf. 3D Vis.(3DV), 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 287, + 225 + ], + "type": "text", + "content": "[3] Thiemo Alldieck, Marcus Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single rgb camera. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "text", + "content": "[4] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "text", + "content": "[5] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "type": "text", + "content": "[6] Brian Amberg, Sami Romdhani, and Thomas Vetter. Optimal step nonrigid icp algorithms for surface registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2007. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "type": "text", + "content": "[7] P.J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 1992. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "text", + "content": "[8] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In 26th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 1999), 1999. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 432, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 474 + ], + "type": "text", + "content": "[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Trans. Vis. Comput. Graph.(TVCG), 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "type": "text", + "content": "[10] Kunyao Chen, Fei Yin, Bang Du, Baichuan Wu, and Truong Q. Nguyen. Efficient registration for human surfaces via isometric regularization on embedded deformation. IEEE Trans. Vis. Comput. Graph.(TVCG), 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[13] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[14] Radek Danecek, Michael J. Black, and Timo Bolkart. EMOCA: Emotion driven monocular face capture and animation. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Bailin Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. Computer Graphics Forum (Eurographics 2022 State-of-the-Art Reports), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[16] Jiankang Deng, Shiyang Cheng, Niannan Xue, Yuxiang Zhou, and Stefanos Zafeiriou. Uv-gan: Adversarial facial uv map completion for pose-invariant face recognition. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 260 + ], + "type": "text", + "content": "[18] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. ACM Trans. Graph.(TOG), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "text", + "content": "[19] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia 2022 Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 308, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 339 + ], + "type": "text", + "content": "[20] David A Field. Laplacian smoothing and delaunay triangulations. Communications in applied numerical methods, 1988. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "text", + "content": "[21] Pascal J Frey and Houman Borouchaki. Surface mesh quality evaluation. International journal for numerical methods in engineering, 1999. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[22] Kaiwen Guo, Feng Xu, Yangang Wang, Yebin Liu, and Qionghai Dai. Robust non-rigid motion tracking and surface reconstruction using 10 regularization. In Int. Conf. Comput. Vis.(ICCV), 2015. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "text", + "content": "[23] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "text", + "content": "[24] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. " + }, + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "inline_equation", + "content": "\\mathrm{ARCH} + +" + }, + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "text", + "content": ": Animation-ready clothed human reconstruction revisited. In Int. Conf. Comput. Vis.(ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 510, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 554 + ], + "type": "text", + "content": "[25] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "text", + "content": "[26] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable reconstruction of clothed humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 622 + ], + "type": "text", + "content": "[27] Huggingface. Stable-diffusion-inpainting, 2022. https://huggingface.co/runwayml/stable-diffusion-inpainting.6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 645 + ], + "type": "text", + "content": "[28] Henrik Wann Jensen. Realistic image synthesis using photon mapping. AK Peters/crc Press, 2001. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[29] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In Eur. Conf. Comput. Vis.(ECCV), 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] Yucheol Jung, Hyomin Kim, Gyeongha Hwang, Seung-Hwan Baek, and Seungyong Lee. Mesh density adapta" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "150" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "tion for template-based shape reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "text", + "content": "[31] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph.(TOG), 2013. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 119, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 152 + ], + "type": "text", + "content": "[32] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 153, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 184 + ], + "type": "text", + "content": "[33] Vladislav Kraevoy and Alla Sheffer. Template-based mesh completion. In Symposium on Geometry Processing, 2005. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "text", + "content": "[34] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of SIGGRAPH 2000, 2000. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "text", + "content": "[35] Kun Li, Jingyu Yang, Yu-Kun Lai, and Daoliang Guo. Robust non-rigid registration with reweighted position and transformation sparsity. IEEE Trans. Vis. Comput. Graph.(TVCG), 2018. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 277, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 308 + ], + "type": "text", + "content": "[36] Tianye Li, Timo Bolkart, Michael Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph.(TOG), 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 310, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 352 + ], + "type": "text", + "content": "[37] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Trans. Graph.(TOG), 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 354, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 399 + ], + "type": "text", + "content": "[38] Yuanqi Li, Jianwei Guo, Xinran Yang, Shun Liu, Jie Guo, Xiaopeng Zhang, and Yanwen Guo. Deep point cloud simplification for high-quality surface reconstruction. arXiv preprint arXiv:2203.09088, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 454 + ], + "type": "text", + "content": "[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-fidelity clothed avatar reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "text", + "content": "[40] Peter Liepa. Filling holes in meshes. In Proceedings of the 2003 Eurographics/ACM SIGGRAPH symposium on Geometry processing, 2003. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 532 + ], + "type": "text", + "content": "[41] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graph.(TOG), 2015. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "text", + "content": "[42] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. Seminal graphics: pioneering efforts that shaped the field, 1998. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[43] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[44] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[45] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Trans. Graph.(TOG), 2021. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2019. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "text", + "content": "[47] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, 2009. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "type": "text", + "content": "[48] Ken Perlin and Eric M Hoffert. Hypertexture. In Proceedings of the 16th annual conference on Computer graphics and interactive techniques, 1989. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 184, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 225 + ], + "type": "text", + "content": "[49] Leonid Pishchulin, Stefanie Wuhrer, Thomas Helten, Christian Theobalt, and Bernt Schiele. Building statistical shape spaces for 3d human modeling. Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[50] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "text", + "content": "[52] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Trans. Graph.(TOG), 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[53] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention(MICCAI), 2015, 2015. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 393, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 446 + ], + "type": "text", + "content": "[54] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Int. Conf. Comput. Vis.(ICCV), 2019. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "text", + "content": "[55] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "text", + "content": "[56] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Ming-min Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In Eur. Conf. Comput. Vis.(ECCV), 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "text", + "content": "[57] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X- avatar: Expressive human avatars. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "text", + "content": "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[59] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Int. Conf. Comput. Vis.(ICCV), 2023. 2, 6, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[60] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Int. Conf. Comput. Vis.(ICCV), 2021. 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "151" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[61] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In IEEE Conf. 3D Vis.(3DV), 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[62] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM Trans. Graph.(TOG), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "text", + "content": "[63] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 251 + ], + "type": "text", + "content": "[64] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 287, + 307 + ], + "type": "text", + "content": "[65] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3d human shape and articulated pose models. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 309, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 287, + 341 + ], + "type": "text", + "content": "[66] Jingyu Yang, Ke Li, Kun Li, and Yu-Kun Lai. Sparse non-rigid registration of 3d shapes. Computer Graphics Forum, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 343, + 286, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 286, + 386 + ], + "type": "text", + "content": "[67] Xueting Yang, Yihao Luo, Yuliang Xiu, Wei Wang, Hao Xu, and Zhaoxin Fan. D-if: Uncertainty-aware human digitization via implicit distribution field. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "type": "text", + "content": "[68] Yuxin Yao, Bailin Deng, Weiwei Xu, and Juyong Zhang. Quasi-newton solver for robust non-rigid registration. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2020. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 433, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 287, + 476 + ], + "type": "text", + "content": "[69] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 478, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 287, + 521 + ], + "type": "text", + "content": "[70] Zheng Zerong, Yu Tao, Liu Yebin, and Dai Qionghai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 555 + ], + "type": "text", + "content": "[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis.(ICCV), 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 600 + ], + "type": "text", + "content": "[72] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 602, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 602, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 602, + 287, + 644 + ], + "type": "text", + "content": "[73] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[74] Hao Zhu, Xinxin Zuo, Haotian Yang, Sen Wang, Xun Cao, and Ruigang Yang. Detailed avatar recovery from single image. IEEE Trans. Pattern Anal. Mach. Intell.(TPAMI), 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 692, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 714 + ], + "type": "text", + "content": "[75] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity gar" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 129 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ment mesh reconstruction from single images. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[76] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z. Li. Face alignment across large poses: A 3d solution. In IEEE Conf. Comput. Vis. Pattern Recog.(CVPR), 2016. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "152" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_content_list.json b/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..77e1e104619ed721a64527809abb101330cfdda6 --- /dev/null +++ b/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_content_list.json @@ -0,0 +1,1994 @@ +[ + { + "type": "text", + "text": "Semantic Line Combination Detector", + "text_level": 1, + "bbox": [ + 312, + 162, + 658, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinwon Ko", + "bbox": [ + 205, + 208, + 285, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea University", + "bbox": [ + 184, + 224, + 308, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jwko@mcl.korea.ac.kr", + "bbox": [ + 163, + 243, + 326, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongkwon Jin", + "bbox": [ + 418, + 208, + 531, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea University", + "bbox": [ + 416, + 224, + 540, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dongkwonjin@mcl.korea.ac.kr", + "bbox": [ + 367, + 243, + 586, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chang-Su Kim*", + "bbox": [ + 660, + 208, + 779, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea University", + "bbox": [ + 656, + 224, + 779, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "changsukim@korea.ac.kr", + "bbox": [ + 627, + 243, + 807, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 287, + 328, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A novel algorithm, called semantic line combination detector (SLCD), to find an optimal combination of semantic lines is proposed in this paper. It processes all lines in each line combination at once to assess the overall harmony of the lines. First, we generate various line combinations from reliable lines. Second, we estimate the score of each line combination and determine the best one. Experimental results demonstrate that the proposed SLCD outperforms existing semantic line detectors on various datasets. Moreover, it is shown that SLCD can be applied effectively to three vision tasks of vanishing point detection, symmetry axis detection, and composition-based image retrieval. Our codes are available at https://github.com/Jinwon-Ko/SLCD.", + "bbox": [ + 111, + 316, + 473, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 522, + 233, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A semantic line [15] is defined as a meaningful line separating distinct semantic regions in an image. Besides this unary definition, multiple semantic lines in an image are supposed to convey the global scene structure properly [13]. It is challenging to detect such semantic lines because they are often implied by complicated region boundaries. Moreover, they should represent the image composition optimally by dividing it into semantic regions harmoniously.", + "bbox": [ + 111, + 542, + 470, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic lines are essential elements in many vision applications. For example, a horizon line [14, 30, 34, 40], which is a specific type of semantic line, can be exploited to adjust the levelness of an image [15, 29]. A reflection symmetry axis [3, 4, 6, 7, 20], which is another type of semantic line, provides visual cues for object recognition and pattern analysis. Vanishing points, conveying depth impression in images, can be estimated by detecting dominant parallel semantic lines in the 3D world [1, 12, 39]. In autonomous driving systems [13, 26, 36], boundaries of road lanes can be also described by semantic lines.", + "bbox": [ + 111, + 652, + 470, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several attempts [9, 12, 13, 15, 35] have been made to detect semantic lines. These techniques perform line detection and refinement sequentially. At the detection", + "bbox": [ + 111, + 805, + 470, + 845 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0eb69331fb07e6bf0443e794367a4b0c94d7a7b8b7cf946f443de0f04bfb0785.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 504, + 287, + 573, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/976615b3f9f48847ab871ec52e7071aa26f8150d32784f5c324ea2d15d079bd5.jpg", + "image_caption": [ + "(a) NMS" + ], + "image_footnote": [], + "bbox": [ + 596, + 290, + 717, + 348 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1cfc8e91f1857e4e81e2238a373ea847ce93b930b4a03ce18bf749fdc5b0011.jpg", + "image_caption": [ + "(b) Pairwise correlation" + ], + "image_footnote": [], + "bbox": [ + 730, + 291, + 851, + 348 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9cb3c69de85b36a772ea0bb6226738a4c810e2f30914daf1c934880e5514946c.jpg", + "image_caption": [ + "Ground-truth", + "Figure 1. After selecting reliable line candidates, there are two existing approaches to semantic line detection. The first approach in (a) focuses on locating a line near a region boundary and eliminating overlapping lines. However, a redundant line still remains, since this approach does not consider how well a group of detected lines represents the layout of a scene. The second approach in (b) takes into account only the pairwise correlation between two lines, so it may fail to assess the overall harmony of more than two semantic lines. In contrast, in (c), the proposed SLCD generates a number of line combinations, analyzes all lines in each combination at once, and then finds the most harmonious combination that conveys the global scene composition optimally." + ], + "image_footnote": [], + "bbox": [ + 504, + 359, + 573, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e31170d5cdeacfff1e62c5bff86153b2e160a3b78829f5183884e485a47e05ee.jpg", + "image_caption": [ + "(c) Proposed SLCD" + ], + "image_footnote": [], + "bbox": [ + 596, + 359, + 740, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56e41afe1fb1bdb8517a6f2c9bfffeffab508d297e7df904cfe1ddef5aedad07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 359, + 846, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stage, they extract deep line features to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing redundant lines. Specifically, to refine line candidates, non-maximum suppression (NMS) is performed in [9, 12, 15], as illustrated in Figure 1(a). Lee et al. [15] iteratively select the reliable line near boundary pixels and remove overlapping lines with the selected one. Han et al. [9] simplify the NMS process by adopting a Hough line space. Jin et al. [12] process each candidate through comparative ranking and matching. These techniques, however, do not consider the overall harmony among detected lines. To cope with this issue, Jin et al. [13] estimate the relation score for every pair of detected lines and then decide harmonious semantic lines via graph optimization. But, they may yield sub-optimal results, since only the pairwise relationships between lines are exploited, as in Figure 1(b).", + "bbox": [ + 496, + 631, + 859, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 129, + 853, + 245, + 864 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28066", + "bbox": [ + 478, + 928, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf3196727fb63a2bdfe55217b71bf63d4b9a90129c67f9cd4d02768c30b054b6.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed SLCD algorithm." + ], + "image_footnote": [], + "bbox": [ + 117, + 126, + 854, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel algorithm, called semantic line combination detector (SLCD), to find an optimal group of semantic lines. It processes all lines in a line group (or combination) simultaneously, instead of analyzing each pair of lines, to estimate the overall harmony, as in Figure 1(c). Figure 2 shows an overview of the proposed SLCD. First, we select reliable lines from line candidates and then generate a number of line combinations. Second, we score all line combinations and determine the combination with the highest score as the optimal group of semantic lines. To this end, we design two novel modules for semantic feature grouping and compositional feature extraction. We also introduce a novel loss function to guide the feature grouping module. Experimental results demonstrate that SLCD can detect semantic lines reliably on existing datasets and a new dataset, called compositionally diverse lines (CDL). Moreover, it is shown that SLCD can be used effectively in various applications.", + "bbox": [ + 109, + 268, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work has the following major contributions:", + "bbox": [ + 130, + 518, + 428, + 532 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SLCD finds an optimal combination of semantic lines by processing all lines in a line combination at once.", + "- We construct the CDL dataset containing compositionally diverse images with implied lines. It will be made publicly available.1", + "- SLCD outperforms conventional detectors on most datasets. Also, its effectiveness is demonstrated in three applications: vanishing point detection, symmetry axis detection, and composition-based image retrieval." + ], + "bbox": [ + 112, + 532, + 472, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 112, + 667, + 243, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic lines, located near the boundaries of different semantic regions, outline the layout and composition of an image. They play an important role in various vision applications. Horizons [14, 30, 34, 40] are a specific type of semantic lines, which can be applied to adjust the levelness of images and improve their aesthetics [15, 29]. In [1, 12, 39], dominant parallel lines in the 3D world are detected to estimate vanishing points, which convey depth impression on 2D images. Also, in [3, 6, 7, 20], reflection symmetry axes are identified to analyze the shapes of objects or patterns. These types of lines can be regarded as", + "bbox": [ + 109, + 691, + 472, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "highly implied semantic lines. In [13, 26, 36], straight lanes are detected to aid in vehicle maneuvers in road environments. Furthermore, semantic lines are essential visual cues in photographic composition [16, 18]. They direct viewers' attention and help to compose a visually balanced image.", + "bbox": [ + 496, + 268, + 857, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several semantic line detectors [9, 12, 13, 15] have been proposed. They perform in two stages: line detection and refinement. At the detection stage, deep line features are extracted to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing irrelevant candidates, based on NMS [9, 12, 15] or graph optimization [13]. More specifically, Lee et al. [15] detect reliable lines with classification probabilities higher than a threshold. They then iteratively select semantic lines and remove overlapping lines with the selected one, by employing the edge detector in [31]. Han et al. [9] predict the probability of each candidate in a Hough parametric space. They simplify NMS by computing the centroids of connected components in the Hough space. Jin et al. [12] design two comparators to estimate the priority and similarity between two lines. Then, they determine the most reliable lines and eliminate redundant ones alternately through pairwise comparisons. However, these techniques [9, 12, 15] do not consider how well a group of detected lines represents the global scene structure. To address this issue, Jin et al. [13] analyze the pairwise harmony of detected lines. They first estimate a harmony score for a pair of detected lines. They then construct a complete graph and determine harmonious semantic lines by finding a maximal weight clique. Their method, however, may yield sub-optimal results because it considers only pairwise relationships between lines. On the contrary, the proposed SLCD finds an optimal combination of semantic lines by analyzing all lines in each combination simultaneously.", + "bbox": [ + 496, + 338, + 859, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Algorithm", + "text_level": 1, + "bbox": [ + 498, + 746, + 679, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a novel algorithm, called SLCD, to detect an optimal combination of semantic lines, an overview of which is in Figure 2. First, we select $K$ reliable lines from line candidates and then generate a number of line combinations. Second, we score all the line combinations and determine the combination with the highest score as the optimal group of semantic lines.", + "bbox": [ + 496, + 768, + 859, + 867 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{1}$ CDL is available at https://github.com/Jinwon-Ko/SLCD.", + "bbox": [ + 129, + 853, + 413, + 864 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28067", + "bbox": [ + 478, + 928, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9b3a72ca4e2fa6e9e65d80f0a5a528ff6a36907331dfac4b96e86c5b476770b3.jpg", + "image_caption": [ + "Figure 3. The architecture of SLCD. CA and FC denote cross-attention and fully connected layers, respectively. The blue, red, yellow, and green boxes indicate encoding, semantic feature grouping, compositional feature extraction, and score regression, respectively." + ], + "image_footnote": [], + "bbox": [ + 117, + 125, + 854, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generating Line Combinations", + "text_level": 1, + "bbox": [ + 112, + 334, + 364, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initializing line candidates: We generate line candidates, which are end-to-end straight lines in an image. Each line candidate is parameterized by polar coordinates in the Hough space [9]. By quantizing the coordinates uniformly, we obtain $N$ line candidates. The default $N$ is 1024.", + "bbox": [ + 111, + 356, + 470, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Filtering line candidates: Given the $N$ line candidates, there are $2^{N}$ possible line combinations in total. Since this number of combinations is unmanageable, we filter out irrelevant lines to maintain $K$ reliable ones only. To this end, we design a simple line detector by modifying S-Net in [13]. Given an image, the detector obtains a convolutional feature map and extracts line features by aggregating the features of pixels along each line candidate. Then, it computes the classification probability and regression offset of each candidate. We select the most reliable candidate with the highest probability and remove overlapping lines with the selected one. We iterate this process $K$ times to determine $K$ reliable lines. The default $K$ is 8. Figure 4 shows examples of selected reliable lines. Note that, in this filtering, false negatives occur rarely because the $K$ lines are selected via NMS even though their classification probabilities are not high. The architecture of the modified S-Net is described in detail in the supplemental document.", + "bbox": [ + 111, + 429, + 473, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating line combinations: From the $K$ reliable lines, we generate all $2^{K}$ line combinations.", + "bbox": [ + 111, + 681, + 470, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Assessing Line Combinations", + "text_level": 1, + "bbox": [ + 112, + 719, + 352, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When a line combination divides an image insufficiently or over-segments it into unnecessary parts, it does not describe the overall structure of the scene properly. On the contrary, an optimal combination of semantic lines should convey the image composition reliably and efficiently (i.e. with a small number of lines). To find the best line combination, we develop the semantic line combination detector (SLCD). Figure 3 shows the structure of SLCD, which performs encoding, semantic feature grouping, compositional feature ex", + "bbox": [ + 111, + 741, + 473, + 867 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/751fdd3902bfde92d1219577e0419c95b15c8569b7c4f69c0fd16a5f126b813b.jpg", + "image_caption": [ + "Figure 4. From a set of line candidates, the line detector selects $K$ reliable lines, depicted in orange. A high recall rate is achieved because a sufficient number of reliable lines are selected through NMS. Ground-truth semantic lines in green are in the insets." + ], + "image_footnote": [], + "bbox": [ + 501, + 333, + 857, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "traction, and score regression. The detailed architecture of SLCD is described in the supplemental document.", + "bbox": [ + 498, + 518, + 857, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoding: Given an image, we extract multi-scale feature maps of ResNet50 [11]. Then, we match their spatial resolutions via bilinear interpolation and concatenate them in the channel dimension. We then squeeze the channels using 2D convolutional layers to obtain an aggregated feature map $F \\in \\mathbb{R}^{HW \\times C}$ , where $H$ , $W$ , and $C$ are the feature height, the feature width, and the number of channels.", + "bbox": [ + 498, + 551, + 857, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic feature grouping: Semantic lines are located near the boundaries between distinct regions. For their reliable detection, it is desirable to separate different regional parts more clearly. We hence attempt to group pixel features into multiple regions through cross-attention [2, 22, 32, 33]. We employ three cross-attention modules.", + "bbox": [ + 498, + 654, + 857, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $R \\in \\mathbb{R}^{M \\times C}$ be a learnable matrix representing $M$ regions, called region query matrix. Then, we convert $R$ into queries and $F$ into keys and values by", + "bbox": [ + 498, + 738, + 857, + 780 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {q} = R U _ {q}, F _ {k} = (F + S) U _ {k}, F _ {v} = (F + S) U _ {v}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 794, + 857, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S\\in \\mathbb{R}^{HW\\times C}$ denotes the sinusoidal positional encoding [28], and $U_{q},U_{k},U_{v}\\in \\mathbb{R}^{C\\times C}$ are projection matrices for queries, keys, and values. We then obtain an updated", + "bbox": [ + 498, + 823, + 857, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28068", + "bbox": [ + 478, + 928, + 517, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c89af6ab40adfaf0bd9a77941c6e68a8ffdda2acec7bf6f221d88a87872fbed4.jpg", + "image_caption": [ + "Figure 5. Visualization of semantic feature grouping results. The top row shows input images with ground-truth lines. The bottom one presents the membership maps, representing the semantic region that each pixel belongs to." + ], + "image_footnote": [], + "bbox": [ + 116, + 123, + 468, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "region query matrix $\\hat{R}$ by", + "bbox": [ + 112, + 277, + 269, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {s o f t} _ {M} \\max \\left(R _ {q} F _ {k} ^ {\\mathrm {T}} / \\tau\\right), \\hat {R} = A F _ {v} + R, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 301, + 470, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A\\in \\mathbb{R}^{M\\times HW}$ is the attention matrix with a scaling factor $\\tau$ and the letter $M$ means the softmax operation is done in the column direction, as in [22, 32, 33]. Then, in the last cross-attention, we generate a semantic feature map $F_{s}\\in \\mathbb{R}^{HW\\times C}$ by", + "bbox": [ + 111, + 332, + 472, + 400 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s} = A ^ {\\mathrm {T}} \\hat {R}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 254, + 401, + 470, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we channel-wise concatenate $F$ and $F_{s}$ to obtain a combined feature map $X\\in \\mathbb{R}^{HW\\times 2C}$ .", + "bbox": [ + 111, + 422, + 472, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, in (2), the softmax function is applied along the query axis. Thus, in $A = [\\mathbf{p}_1,\\mathbf{p}_2,\\dots ,\\mathbf{p}_{HW}]$ each column $\\mathbf{p}_i\\in \\mathbb{R}^M$ is a probability vector. Specifically, the mth element $p_i^m$ in $\\mathbf{p}_i$ is the probability that pixel $i$ belongs to the mth region query. Figure 5 visualizes membership maps, computed by", + "bbox": [ + 111, + 450, + 472, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {1}\\right), \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {2}\\right), \\dots , \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {H W}\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 541, + 470, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each element is the index of the region query that the corresponding pixel most likely belongs to. We see that each image is partitioned into $M$ meaningful regions. Thus, $F_{s}$ in (3) represents the regional membership of each pixel and is used to make the feature map $X$ more discriminative.", + "bbox": [ + 111, + 566, + 472, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To produce the attention matrix $A$ reliably, we design a novel loss function in Section 3.3. The default $M$ is 8.", + "bbox": [ + 111, + 635, + 472, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compositional feature extraction: For each line combination, we generate three types of feature maps to extract a compositional feature map. First, we generate a binary mask $B \\in \\mathbb{R}^{HW}$ for each line combination. Each element in $B$ is 1 if the corresponding pixel belongs to any line in the combination, and 0 otherwise. Then, we decompose the combined feature into a line feature map $X_{l}$ and a region feature map $X_{r}$ by", + "bbox": [ + 111, + 666, + 472, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {l} = X \\otimes B, X _ {r} = X \\otimes (1 - B), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 786, + 470, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ is the element-wise multiplication. In other words, $X_{l}\\in \\mathbb{R}^{HW\\times 2C}$ contains contextual information for the line pixels only, whereas $X_{r}\\in \\mathbb{R}^{HW\\times 2C}$ does for the pixels strictly inside the regions divided by the lines.", + "bbox": [ + 112, + 810, + 472, + 866 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f64950abe3173d58c32d60b1fd0f88e3890ccb5499261c4034b08c36494c9e1e.jpg", + "image_caption": [ + "Figure 6. Illustration of the line collection map generation when $K = 3$ : (a) reliable lines, (b) line combinations, and (c) line collection maps." + ], + "image_footnote": [], + "bbox": [ + 501, + 125, + 857, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, we generate a line collection map $L = [L_1,L_2,\\dots ,L_K]\\in \\mathbb{R}^{HW\\times K}$ , where $L_{k}\\in \\mathbb{R}^{HW}$ is a ternary mask for the $k$ th reliable line, as illustrated in Figure 6. If the $k$ th reliable line does not belong to the line combination, all elements in $L_{k}$ are set to 0. Otherwise, each value in $L_{k}$ is set to either 1 or $-1$ to indicate where the corresponding pixel is located between the two parts divided by the $k$ th reliable line. In other words, $L_{k}$ informs how the $k$ th line splits the image into two regions. We then produce a positional feature map $P\\in \\mathbb{R}^{HW\\times 2C}$ by applying a series of fully connected layers to $L$ . Then, we yield a compositional feature map $Z\\in \\mathbb{R}^{HW\\times 6C}$ by", + "bbox": [ + 496, + 282, + 859, + 449 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\left[ X _ {l}, X _ {r}, P \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 459, + 857, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which contains information about how the lines in the combination separate the image into multiple parts.", + "bbox": [ + 496, + 486, + 857, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Score regression: Lastly, using a regressor, we estimate a composition score for each line combination. Then, we declare the line combination with the highest score as the optimal group of semantic lines. The regressor takes the compositional feature map $Z$ in (6) as input and predicts a composition score $s$ within [0, 1]. It is implemented using a bilinear interpolation layer and a series of 2D convolution layers and fully connected layers with the ReLU activation.", + "bbox": [ + 496, + 517, + 859, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Loss Functions", + "text_level": 1, + "bbox": [ + 500, + 637, + 638, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train SLCD, we design two loss functions.", + "bbox": [ + 498, + 659, + 774, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic region separation loss: When two pixels $i$ and $j$ are located in distinct regions, they should be assigned to different region queries. In other words, their probability vectors $\\mathbf{p}_i$ and $\\mathbf{p}_j$ in the attention matrix $A$ should be far from each other. Let $X$ and $Y$ be the two regions divided by a ground-truth line. Their probability vectors are defined as", + "bbox": [ + 496, + 676, + 857, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {X} = \\frac {1}{| X |} \\sum_ {i \\in X} \\mathbf {p} _ {i}, \\quad \\mathbf {p} _ {Y} = \\frac {1}{| Y |} \\sum_ {j \\in Y} \\mathbf {p} _ {j}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 770, + 857, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which should be also far from each other, for the ground-truth line divides the image into two semantic regions.", + "bbox": [ + 496, + 810, + 856, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To measure the distance between probability vectors, we adopt the Kullback-Leibler divergence (KLD) [5]. Then,", + "bbox": [ + 498, + 838, + 857, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28069", + "bbox": [ + 478, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9ab17d575ca7742c214e5f6b6731cb5b27e54609641d4e07740989a0c89fa3e6.jpg", + "image_caption": [ + "Horizontal Vertical Diagonal Triangle Symmetric Low Front", + "Figure 7. The proposed CDL dataset contains 9,100 images in seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. The ground-truth semantic lines are depicted in green." + ], + "image_footnote": [], + "bbox": [ + 114, + 136, + 468, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we define the semantic region segmentation (SRS) loss, by employing all ground-truth lines, as", + "bbox": [ + 111, + 330, + 472, + 358 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S R S}} = - \\sum_ {l = 1} ^ {T} \\left[ D \\left(\\mathbf {p} _ {X _ {l}} \\| \\mathbf {p} _ {Y _ {l}}\\right) + D \\left(\\mathbf {p} _ {Y _ {l}} \\| \\mathbf {p} _ {X _ {l}}\\right) \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 367, + 472, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D$ denotes the KLD, $\\mathbf{p}_{X_l}$ and $\\mathbf{p}_{Y_l}$ are the regions divided by the $l$ th ground-truth line, and $T$ is the number of ground-truth lines. Both $D(\\mathbf{p}_{X_l} \\parallel \\mathbf{p}_{Y_l})$ and $D(\\mathbf{p}_{Y_l} \\parallel \\mathbf{p}_{X_l})$ are computed because $D$ is not commutative. With $\\mathcal{L}_{\\mathrm{SRS}}$ , it is possible to roughly segment an image into meaningful parts, as illustrated in Figure 5, even though no ground-truth labels for semantic segmentation are used for training.", + "bbox": [ + 111, + 412, + 472, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regression loss: We also design a loss for regressing the composition score of each line combination. To this end, we employ the HIoU metric [13] that quantifies the structural layout of a line combination. Let $\\mathbf{c}$ and $\\mathbf{c}^{\\star}$ denote a line combination and the ground-truth combination, respectively. We compute the HIoU between $\\mathbf{c}$ and $\\mathbf{c}^{\\star}$ and use it as the ground-truth composition score $\\overline{s}$ of $\\mathbf{c}$ . Then, the regression loss is defined as", + "bbox": [ + 111, + 513, + 472, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = (s - \\bar {s}) ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 632, + 470, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ is the predicted score of $\\mathbf{c}$ . To reduce $\\mathcal{L}_{\\mathrm{reg}}$ effectively, a ranking loss $\\mathcal{L}_{\\mathrm{rank}}$ is also employed as in [17].", + "bbox": [ + 111, + 657, + 472, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 112, + 696, + 302, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 112, + 720, + 312, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt ResNet50 [11] as the encoder of the proposed SLCD. We use the AdamW optimizer [19] with a learning rate $10^{-4}$ , a weight decay of $10^{-4}$ , $\\gamma = 0.5$ , $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ . We use a batch size of two for 400,000 iterations. Training images are resized to $480 \\times 480$ and augmented by random horizontal flipping. We fix the number of line candidates, reliable lines, and region queries to $N = 1024$ , $K = 8$ , and $M = 8$ , respectively. We set $H = 60$ , $W = 60$ , and $C = 96$ .", + "bbox": [ + 111, + 741, + 473, + 866 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5f9576c44757900c226d49acc436ebe6d0bcd93e03b7ae4acc6dd6ed06099b38.jpg", + "table_caption": [ + "Table 1. Comparison of the HIoU scores $(\\%)$ on the SEL, SEL_Hard, NKL, and CDL datasets." + ], + "table_footnote": [], + "table_body": "
SELSEL_HardNKLCDL
SLNet [15]77.8759.7165.4957.78
DHT [9]79.6263.3969.0863.24
DRM [12]80.2368.8367.4263.96
HSLD [13]81.0365.9974.2964.98
Proposed84.0968.1576.2168.85
", + "bbox": [ + 503, + 154, + 857, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 500, + 251, + 591, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SEL [15]: It is the first dataset for semantic line detection. It contains 1,750 images, which are split into 1,575 training and 175 test images. Each semantic line is annotated by the coordinates of two endpoints on an image boundary. In SEL, most images are high-quality landscapes, in which semantic lines are often obvious.", + "bbox": [ + 496, + 272, + 857, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SEL_Hard [12]: It contains 300 test images only, selected from the ADE20K segmentation dataset [37]. It is challenging because of cluttered scenes or occluded objects.", + "bbox": [ + 496, + 358, + 857, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NKL [35]: It is a relatively large dataset of 5,200 training and 1,300 testing images. It includes both indoor and outdoor scenes.", + "bbox": [ + 496, + 402, + 857, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CDL: We construct CDL to contain 7,100 scenes with diverse contents and compositions. It is split into 6,390 training and 710 test images. As in Figure 7, the images are categorized into seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. In Low and Front, humans and animals are essential parts of the image composition. In the other classes, most images are outdoor ones, as in the other datasets. To construct CDL, semantic lines were manually annotated in about 200 man-hours. More examples and the annotation process are provided in the supplemental document.", + "bbox": [ + 496, + 446, + 859, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Metrics", + "text_level": 1, + "bbox": [ + 500, + 607, + 588, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned earlier, the harmony of detected lines is more important than individual lines in semantic line detection. Therefore, the main paper discusses only HIoU performances. The precision, recall, and f-measure results, which are metrics used in previous work, are compared in the supplemental document.", + "bbox": [ + 496, + 628, + 857, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HIoU: An optimal group of semantic lines conveys a harmonious impression about the composition of an image. To assess the overall harmony of detection results, we use the HIoU metric [13]. It measures the consistency between the division of an image by detected lines and that by the ground truth. Let $S = \\{s_1,s_2,\\dots ,s_N\\}$ and $T = \\{t_{1},t_{2},\\ldots ,t_{M}\\}$ be the regions divided by the detected lines and the ground-truth lines, respectively. Then, HIoU is computed as", + "bbox": [ + 496, + 714, + 857, + 839 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H I o U} = \\frac {\\sum_ {i = 1} ^ {N} \\max _ {k} \\operatorname {I o U} \\left(s _ {i} , t _ {k}\\right) + \\sum_ {j = 1} ^ {M} \\max _ {k} \\operatorname {I o U} \\left(t _ {j} , s _ {k}\\right)}{N + M}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 845, + 857, + 869 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28070", + "bbox": [ + 478, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f40557f8557bd0bac1cbdba040bbc06303e37f7b735fa096631ecd5a814228d.jpg", + "image_caption": [ + "Figure 8. Comparison of detected semantic lines. From the left, two images are selected from each of the SEL, SEL_Hard, NKL, and CDL datasets." + ], + "image_footnote": [], + "bbox": [ + 117, + 114, + 857, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, for each $s_i$ , the most overlapping region $t_k$ is chosen and the intersection-over-union (IoU) is calculated. This is performed similarly for each $t_j$ . Then, the HIoU score is given by the average of these bi-directionally matching IoUs.", + "bbox": [ + 111, + 481, + 473, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparative Assessment", + "text_level": 1, + "bbox": [ + 112, + 560, + 323, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares the HIoU scores of the proposed SLCD with those of the existing detectors [9, 12, 13, 15] on the SEL, SEL_Hard, NKL, and CDL datasets. Also, Figure 8 compares detection results qualitatively. The existing detectors miss correct lines or fail to remove redundant lines, yielding sub-optimal results. In contrast, SLCD detects semantic lines more precisely and represents the composition more reliably than the existing detectors do. More detection results are available in the supplemental document.", + "bbox": [ + 111, + 582, + 473, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on SEL: In Table 1, SLCD outperforms the existing detectors on SEL. Compared with the second-best HSLD, SLCD yields a wide HIoU margin of 3.06. This indicates that SLCD finds an optimal group of semantic lines more effectively by processing all lines in each combination simultaneously, instead of performing pairwise comparisons in HSLD.", + "bbox": [ + 111, + 710, + 470, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on SEL_Hard: As done in [12], we conduct experiments on SEL_Hard using the networks trained on the SEL dataset. In Table 1, SLCD ranks 2nd on SEL_Hard. DRM provides a better result than SLCD, but it demands", + "bbox": [ + 111, + 810, + 473, + 867 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4ec32f932ef9c5ddbb6406f86fb06d037bf16468c985adf4bcf5c791363fc34d.jpg", + "table_caption": [ + "Table 2. Comparison of the HIoU scores $(\\%)$ on the CDL dataset according to the number $M$ of region queries." + ], + "table_footnote": [], + "table_body": "
M481632
H IoU66.4368.8568.1167.07
", + "bbox": [ + 509, + 511, + 846, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5cad2bf36d941e55e506c89673e99df94b5231413db42ebfcacd7e0bb5a9c9ee.jpg", + "image_caption": [ + "Figure 9. Visualization of the membership maps according to $M$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 556, + 856, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "much higher complexity, as will be discussed in Section 4.5.", + "bbox": [ + 498, + 689, + 856, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on NKL: SLCD surpasses all existing detectors. For example, its HIoU score is 1.92 points higher than the second-best HSLD.", + "bbox": [ + 498, + 708, + 857, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on CDL: Table 1 also lists HIoU scores on the proposed CDL dataset. For a fair comparison, we train the existing detectors on CDL using their publicly available source codes. We see that SLCD outperforms all existing detectors on CDL as well. SLNet, DHT, and DRM yield poor results because they do not consider the overall harmony of detected lines. HSLD is better than these detectors but is inferior to the proposed SLCD. We compare and", + "bbox": [ + 496, + 755, + 859, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28071", + "bbox": [ + 478, + 928, + 517, + 940 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb623a46b76daa409fb2eecc2b996a1e061d719a13e7bd9b4f02b39056242901.jpg", + "table_caption": [ + "Table 3. Comparison of the HIoU scores $(\\%)$ on the CDL dataset according to the usage of the SRS loss $\\mathcal{L}_{\\mathrm{SRS}}$" + ], + "table_footnote": [], + "table_body": "
w/o LSRProposed
HIoU66.7168.85
", + "bbox": [ + 122, + 154, + 460, + 189 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dc6e890ec4d1aa86c06743ca92b72b5a0ed70275bd72f58b9d859f85c27635df.jpg", + "table_caption": [ + "Table 4. Ablation studies for the compositional feature extraction on the CDL dataset." + ], + "table_footnote": [], + "table_body": "
LineRegionPositionalHIoU
I✓65.12
II✓65.74
III✓✓66.37
IV✓✓✓68.85
", + "bbox": [ + 122, + 231, + 460, + 300 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "discuss the detection results according to the composition classes in the supplemental document.", + "bbox": [ + 111, + 323, + 470, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Analysis", + "text_level": 1, + "bbox": [ + 112, + 359, + 207, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct several ablation studies to analyze each component of the proposed SLCD.", + "bbox": [ + 111, + 382, + 470, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The number of $M$ : Table 2 lists the HIoU performances on the CDL dataset, according to the number $M$ of region queries. At the default $M = 8$ , SLCD achieves the best performance. When $M$ is smaller or bigger than 8, the performance degrades due to under- or over-segmentation of semantic parts, respectively, as shown in Figure 9.", + "bbox": [ + 111, + 412, + 472, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficacy of $\\mathcal{L}_{\\mathrm{SRS}}$ : In Table 3, if the proposed SRS loss $\\mathcal{L}_{\\mathrm{SRS}}$ in (8) is excluded from the training, the performance drops by 2.14 points. It means that the composition analysis based on the SRS loss is essential for finding optimal line combinations.", + "bbox": [ + 111, + 500, + 472, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficacy of compositional feature extraction: SLCD extracts the compositional feature map $Z$ for each line combination, by processing the line feature map $X_{l}$ , the region feature map $X_{r}$ , and the positional feature map $P$ in (6). In Table 4, Method I uses the line feature map only, while II uses the region feature map only. Method III utilizes both feature maps. Method IV, which is the proposed SLCD, uses all three maps.", + "bbox": [ + 111, + 571, + 470, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Method I yields the worst result, for it uses only the contextual information near line pixels. Method II is slightly better than Method I, by exploiting the regional information. Using both line and region feature maps, III provides a better result. Moreover, IV further improves the performance significantly by utilizing the positional feature map. This is because the overall harmony is estimated more effectively, by combining the line structures in the positional feature map with scene contexts.", + "bbox": [ + 111, + 683, + 470, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Runtime: Table 5 compares the runtimes of SLCD and the existing detectors in seconds per frame (spf). We use a PC with AMD Ryzen 9 3900X CPU and NVIDIA RTX 2080 GPU. The proposed SLCD takes 0.114spf, adding", + "bbox": [ + 111, + 810, + 472, + 867 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8713657656d0c36a64321c4bb0145247e2b7d6da2a8416bd1b091c353b867d0a.jpg", + "table_caption": [ + "Table 5. Runtime comparison of the proposed SLCD and the existing detectors. The processing times are reported in seconds per frame (spf)." + ], + "table_footnote": [], + "table_body": "
SLNet [15]DHT [9]DRM [12]HSLD [13]Proposed
0.1360.0330.9520.0460.114
", + "bbox": [ + 504, + 167, + 849, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df92b167f65fd904d2e83b60fa11a7750248f4645163c77d1b2915c6d82eefc3.jpg", + "table_caption": [ + "Table 6. Comparison of the AA scores $(\\%)$ of VP detection." + ], + "table_footnote": [], + "table_body": "
AA1°AA2°AA10°
Zhou et al. [39]18.533.060.0
Jin et al. [12]8.622.968.3
Proposed16.636.978.3
", + "bbox": [ + 506, + 233, + 849, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/959a4c3b38ec86f392b4e02c6b4a0a280d95ae345be6a7250cb0e7239250a6b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 308, + 573, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/383589e349806de2088fb487d72d2f1b3f5e24dc5e8293ef75eece014f9bfacf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 308, + 643, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1ce6b0a7510c7a901fe5856c739fb376930ce7d418d30d8a4d2810cc2264807.jpg", + "image_caption": [ + "Figure 10. Detected dominant parallel lines and their intersections are depicted by yellow lines and green triangles, respectively. The ground-truth VPs are depicted by red dots." + ], + "image_footnote": [], + "bbox": [ + 573, + 349, + 643, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b50d7c4b343f89aa571b578a144ba3922851fd495db9c62dc98175268075b18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 308, + 712, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eade586151d69ba3145ba6a22b7a10d581dc3537733da735cd6b6994a28a0203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 349, + 712, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f2b0c4b156000e45d831c8818791939272e4089ac14af15dab5ceffdeb9d111.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 308, + 784, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/130d977ad799b84a0060fcba31190943ee9501631bf296dee4a57b2a57bfa2ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 349, + 784, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78a0814479b3248024a17fe086e70206428606b9f9138c8acc8dc90eca298c9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 308, + 854, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4a840a90cecb107679c05a08857c0d7eade2faab03adbb0e052c821ab165c83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 349, + 854, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "up 0.030spf and 0.074spf for generating and assessing line combinations, respectively. DHT is the fastest detector, but it is inferior to SLCD on all datasets. On the other hand, even though DRM performs better on SEL_Hard, it is about 8.4 times slower than SLCD.", + "bbox": [ + 496, + 455, + 857, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Applications", + "text_level": 1, + "bbox": [ + 500, + 542, + 617, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply the proposed SLCD to three vision tasks: dominant vanishing point detection, reflection symmetry axis detection, and composition-based image retrieval. Note that the first two tasks were considered in [12]. Due to the page limit, more details and results are described in the supplemental document.", + "bbox": [ + 496, + 566, + 857, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Dominant Vanishing Point Detection", + "text_level": 1, + "bbox": [ + 498, + 662, + 790, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A vanishing point (VP) facilitates understanding of the 3D geometric structure. We apply the proposed SLCD to detect dominant VPs, by identifying vanishing lines. We use the AVA landscape dataset [39], in which two dominant parallel lines are annotated for each image. To detect a dominant VP, we generate line combinations containing two lines only. We then find the best combination, whose intersecting point is declared as a VP. Table 6 compares this VP detection scheme with the existing line-based VP detectors in [12, 39]. The angle accuracies (AAs) [38] are compared. We see that SLCD is better than the existing detectors, except that its $\\mathrm{AA1}^{\\circ}$ score is lower than that of Zhou et al. [39]. Figure 10 shows some VP detection results.", + "bbox": [ + 496, + 686, + 859, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28072", + "bbox": [ + 478, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/10af0201aa97ea5ea3e60f7ddb9c8d9a1523d32f4f70e1694200e791bc0a6af7.jpg", + "table_caption": [ + "Table 7. Comparison of the AUC_A scores $(\\%)$ of symmetry axis detection." + ], + "table_footnote": [], + "table_body": "
ICCVNYUSYM_Hard
Cicconet et al. [3]80.8082.8568.99
Elawady et al. [6]87.2483.8373.90
Cicconet et al. [4]87.3887.6481.04
Loy & Eklundh [20]89.7790.8581.99
Jin et al. [12]90.6092.7884.73
Proposed93.1593.5188.03
", + "bbox": [ + 119, + 152, + 460, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2c64ee54b58ceab391233324a1453dd9c9ffb0f86dd748a3ac2736c7eb081d4.jpg", + "image_caption": [ + "Figure 11. Symmetry axis detection results. In the top row, the ground-truth and predicted axes are depicted by dashed red and solid yellow lines, respectively. The membership maps are also visualized in the bottom row." + ], + "image_footnote": [], + "bbox": [ + 117, + 243, + 468, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Reflection Symmetry Axis Detection", + "text_level": 1, + "bbox": [ + 112, + 392, + 398, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reflection symmetry is a visual cue for analyzing object shapes and patterns. Its detection is challenging because the symmetry is only implicit in many cases. We test the proposed SLCD on three datasets: ICCV [7], NYU [4], and SYM_Hard [12]. In these datasets, each image contains a single reflection symmetry axis. Thus, in the proposed SLCD, each of the $K$ reliable lines is regarded as a line combination. Table 7 compares the AUC_A scores [15] of the proposed SLCD and the conventional techniques [3, 4, 6, 12, 20]. SLCD outperforms all these techniques on all datasets. Figure 11 shows some detection results together with the membership maps. We see that regional parts are symmetrically divided along implied axes, enabling SLCD to identify those axes effectively.", + "bbox": [ + 111, + 409, + 473, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Composition-Based Image Retrieval", + "text_level": 1, + "bbox": [ + 112, + 613, + 398, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Existing image retrieval techniques focus on the visual contents of a query image to find similar images in a database [8, 21, 24, 25]. In this work, however, we attempt to discover images with similar compositions to a query image, i.e. composition-based image retrieval.", + "bbox": [ + 111, + 630, + 470, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test the proposed SLCD on the Oxford 5k and Paris 6k retrieval dataset [23]. We first detect semantic lines in every image, while storing the positional feature map $P$ in (6). We filter out some images whose composition scores are lower than a threshold since a low score indicates a low-quality image with inharmonious composition in general. Then, for a randomly selected query image, we compute the $\\ell_2$ -distances between the positional feature maps $P$ of the query and the remaining images. We determine the images with the smallest distances as retrieval results.", + "bbox": [ + 111, + 699, + 470, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 12 shows the top-4 retrieval results for four query images. We see that SLCD returns structurally similar im", + "bbox": [ + 111, + 838, + 472, + 867 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/576dbbd33363f999f807e75a277b54742f5fe14c30814966efee0663584381cd.jpg", + "image_caption": [ + "Query image" + ], + "image_footnote": [], + "bbox": [ + 503, + 125, + 575, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc92edc03a8701496e64266227c326dfb56783c1f8cc4e0fff47ad3a22fc319a.jpg", + "image_caption": [ + "Top-4 retrieval results" + ], + "image_footnote": [], + "bbox": [ + 578, + 125, + 856, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae0a8aa009561a0ca6d71390ed852a25cbe9500eaf2042fd3f89e7d85b2315b2.jpg", + "image_caption": [ + "Figure 12. Composition-based retrieval results for query images.", + "Figure 13. t-SNE visualization [27] of the feature space for the Oxford 5k and Paris 6k dataset [23]." + ], + "image_footnote": [], + "bbox": [ + 503, + 277, + 854, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ages to the queries. This means that the positional feature maps $P$ , containing the structural information of semantic lines, can be used to compute the compositional differences. Based on this observation, we perform data clustering by employing k-means [10] on the feature space of $P$ . Figure 13 is t-SNE visualization [27] of the clustering results. The nine images nearest to each centroid are also shown. Note that images with similar composition are grouped into the same cluster, confirming the efficacy of SLCD in the compositional analysis of images.", + "bbox": [ + 496, + 458, + 857, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 604, + 614, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a novel semantic line detector, SLCD, which processes a combination of lines at once to estimate the overall harmony reliably. We first generated all possible line combinations from reliable lines. Then, we estimated the score of each line combination and determined the best combination. Experimental results demonstrated that the proposed SLCD can detect semantic lines reliably on existing datasets, as well as on the new dataset CDL. Furthermore, SLCD can be successfully used in vanishing point detection, symmetry axis detection, and image retrieval.", + "bbox": [ + 496, + 622, + 857, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 767, + 653, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was conducted by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD) and was supported by the National Research Foundation of Korea (NRF) grants funded by the Korea government (MSIT) (No. NRF-2021R1A4A1031864 and No. NRF-2022R1A2B5B03002310).", + "bbox": [ + 496, + 782, + 857, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28073", + "bbox": [ + 478, + 928, + 517, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 125, + 200, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Charles Bazin, Yongduek Seo, Cedric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in Manhattan world. In Proc. IEEE CVPR, 2012. 1, 2", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 3", + "[3] Marcelo Cicconet, Vighnesh Birodkar, Mads Lund, Michael Werman, and Davi Geiger. A convolutional approach to reflection symmetry. Pattern Recog. Lett., 95:44-50, 2017. 1, 2, 8", + "[4] Marcelo Cicconet, David GC Hildebrand, and Hunter Elliott. Finding mirror symmetry via registration and optimal symmetric pairwise assignment of curves: Algorithm and results. In Proc. IEEE ICCV Workshops, 2017. 1, 8", + "[5] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley, 2nd edition, 2006. 4", + "[6] Mohamed Elawady, Christophe Ducottet, Olivier Alata, Cécile Barat, and Philippe Colantoni. Wavelet-based reflection symmetry detection via textural and color histograms. In Proc. IEEE ICCV Workshops, 2017. 1, 2, 8", + "[7] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 ICCV Challenge: Detecting symmetry in the wild. In Proc. IEEE ICCV, 2017. 1, 2, 8", + "[8] Albert Gordo, Jon Almazan, Jerome Revaud, and Diane Larlus. End-to-end learning of deep visual representations for image retrieval. Int. J. Comput. Vis., 124:237-254, 2017. 8", + "[9] Qi Han, Kai Zhao, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. In Proc. ECCV, 2020. 1, 2, 3, 5, 6, 7", + "[10] J. A. Hartigan and M. A. Wong. A K-means clustering algorithm. Journal of the Royal Statistical Society. Series C (Applied Statistics), 28(1):100-108, 1979. 8", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, 2016. 3, 5", + "[12] Dongkwon Jin, Jun-Tae Lee, and Chang-Su Kim. Semantic line detection using mirror attention and comparative ranking and matching. In Proc. ECCV, 2020. 1, 2, 5, 6, 7, 8", + "[13] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, and Chang-Su Kim. Harmonious semantic line detection via maximal weight clique selection. In Proc. IEEE CVPR, 2021. 1, 2, 3, 5, 6, 7", + "[14] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Temporally consistent horizon lines. In Proc. IEEE ICRA, 2020. 1, 2", + "[15] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Semantic line detection and its applications. In Proc. IEEE ICCV, 2017. 1, 2, 5, 6, 7, 8", + "[16] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Photographic composition classification and dominant geometric element detection for outdoor scenes. J. Vis. Commun. Image Represent., 55:91-105, 2018. 2" + ], + "bbox": [ + 114, + 147, + 472, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Debang Li, Junge Zhang, Kaiqi Huang, and Ming-Hsuan Yang. Composing good shots by exploiting mutual relations. In Proc. IEEE CVPR, 2020. 5", + "[18] Ligang Liu, Renjie Chen, Lior Wolf, and Daniel Cohen-Or. Optimizing photo composition. Computer Graphics Forum, 29:469-478, 2010. 2", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[20] Gareth Loy and Jan-Olof Eklundh. Detecting symmetry and symmetric constellations of features. In Proc. ECCV, 2006. 1, 2, 8", + "[21] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proc. IEEE ICCV, 2017. 8", + "[22] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. iDisc: Internal discretization for monocular depth estimation. In Proc. IEEE CVPR, 2023. 3, 4", + "[23] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proc. IEEE CVPR, 2018. 8", + "[24] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell., 41:1655-1668, 2018. 8", + "[25] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. IEEE Trans. Pattern Anal. Mach. Intell., 22:1349-1380, 2000. 8", + "[26] Lucas Tabelini, Rodrigo Berriel, Thiago M. Paixao, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE CVPR, 2021. 1, 2", + "[27] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8", + "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 3", + "[29] Yaoting Wang, Yongzhen Ke, Kai Wang, Jing Guo, and Shuai Yang. Spatial-invariant convolutional neural network for photographic composition prediction and automatic correction. J. Vis. Commun. Image Represent., 90:103751, 2023. 1, 2", + "[30] Scott Workman, Menghua Zhai, and Nathan Jacobs. Horizon lines in the wild. In Proc. BMVC, 2016. 1, 2", + "[31] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proc. IEEE ICCV, 2015. 2", + "[32] Qihang Yu, Huiyu Wang, Dahun Kim, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. CMT-DeepLab: Clustering mask transformers for panoptic segmentation. In Proc. IEEE CVPR, 2022. 3, 4", + "[33] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. $k$ -means mask transformer. In Proc. ECCV, 2022. 3, 4" + ], + "bbox": [ + 501, + 127, + 856, + 864 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28074", + "bbox": [ + 478, + 928, + 517, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proc. IEEE CVPR, 2016. 1, 2", + "[35] Kai Zhao, Qi Han, Chang-Bin Zhang, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. IEEE Trans. Pattern Anal. Mach. Intell., 44:4793-4806, 2021. 1, 5", + "[36] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. CLRNet: Cross layer refinement network for lane detection. In Proc. IEEE CVPR, 2022. 1, 2", + "[37] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In Proc. IEEE CVPR, 2017. 5", + "[38] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. NeurVPS: Neural vanishing point scanning via conic convolution. In Proc. NeurIPS, 2019. 7", + "[39] Zihan Zhou, Farshid Farhat, and James Z Wang. Detecting dominant vanishing points in natural scenes with application to composition-sensitive image retrieval. IEEE Trans. Multimedia, 19:2651-2665, 2017. 1, 2, 7", + "[40] Rui Zhu, Xingyi Yang, Yannick Hold-Geoffroy, Federico Perazzi, Jonathan Eisenmann, Kalyan Sunkavalli, and Manmohan Chandraker. Single view metrology in the wild. In Proc. ECCV, 2020. 1, 2" + ], + "bbox": [ + 114, + 126, + 470, + 449 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28075", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_model.json b/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e20d790ab0529af3c681695b5bbe685d0613bd6 --- /dev/null +++ b/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_model.json @@ -0,0 +1,2585 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.313, + 0.163, + 0.66, + 0.181 + ], + "angle": 0, + "content": "Semantic Line Combination Detector" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.209, + 0.287, + 0.224 + ], + "angle": 0, + "content": "Jinwon Ko" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.226, + 0.31, + 0.241 + ], + "angle": 0, + "content": "Korea University" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.244, + 0.328, + 0.256 + ], + "angle": 0, + "content": "jwko@mcl.korea.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.209, + 0.532, + 0.224 + ], + "angle": 0, + "content": "Dongkwon Jin" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.226, + 0.541, + 0.241 + ], + "angle": 0, + "content": "Korea University" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.244, + 0.588, + 0.256 + ], + "angle": 0, + "content": "dongkwonjin@mcl.korea.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.209, + 0.78, + 0.225 + ], + "angle": 0, + "content": "Chang-Su Kim*" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.226, + 0.78, + 0.241 + ], + "angle": 0, + "content": "Korea University" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.244, + 0.808, + 0.256 + ], + "angle": 0, + "content": "changsukim@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.289, + 0.329, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.318, + 0.474, + 0.497 + ], + "angle": 0, + "content": "A novel algorithm, called semantic line combination detector (SLCD), to find an optimal combination of semantic lines is proposed in this paper. It processes all lines in each line combination at once to assess the overall harmony of the lines. First, we generate various line combinations from reliable lines. Second, we estimate the score of each line combination and determine the best one. Experimental results demonstrate that the proposed SLCD outperforms existing semantic line detectors on various datasets. Moreover, it is shown that SLCD can be applied effectively to three vision tasks of vanishing point detection, symmetry axis detection, and composition-based image retrieval. Our codes are available at https://github.com/Jinwon-Ko/SLCD." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.523, + 0.234, + 0.537 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.543, + 0.472, + 0.653 + ], + "angle": 0, + "content": "A semantic line [15] is defined as a meaningful line separating distinct semantic regions in an image. Besides this unary definition, multiple semantic lines in an image are supposed to convey the global scene structure properly [13]. It is challenging to detect such semantic lines because they are often implied by complicated region boundaries. Moreover, they should represent the image composition optimally by dividing it into semantic regions harmoniously." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.654, + 0.472, + 0.805 + ], + "angle": 0, + "content": "Semantic lines are essential elements in many vision applications. For example, a horizon line [14, 30, 34, 40], which is a specific type of semantic line, can be exploited to adjust the levelness of an image [15, 29]. A reflection symmetry axis [3, 4, 6, 7, 20], which is another type of semantic line, provides visual cues for object recognition and pattern analysis. Vanishing points, conveying depth impression in images, can be estimated by detecting dominant parallel semantic lines in the 3D world [1, 12, 39]. In autonomous driving systems [13, 26, 36], boundaries of road lanes can be also described by semantic lines." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.806, + 0.472, + 0.847 + ], + "angle": 0, + "content": "Recently, several attempts [9, 12, 13, 15, 35] have been made to detect semantic lines. These techniques perform line detection and refinement sequentially. At the detection" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.289, + 0.574, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.332, + 0.549, + 0.339 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.291, + 0.718, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.351, + 0.675, + 0.359 + ], + "angle": 0, + "content": "(a) NMS" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.292, + 0.852, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.351, + 0.832, + 0.359 + ], + "angle": 0, + "content": "(b) Pairwise correlation" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.36, + 0.574, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.432, + 0.564, + 0.439 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.361, + 0.741, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.431, + 0.76, + 0.44 + ], + "angle": 0, + "content": "(c) Proposed SLCD" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.361, + 0.848, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.45, + 0.86, + 0.603 + ], + "angle": 0, + "content": "Figure 1. After selecting reliable line candidates, there are two existing approaches to semantic line detection. The first approach in (a) focuses on locating a line near a region boundary and eliminating overlapping lines. However, a redundant line still remains, since this approach does not consider how well a group of detected lines represents the layout of a scene. The second approach in (b) takes into account only the pairwise correlation between two lines, so it may fail to assess the overall harmony of more than two semantic lines. In contrast, in (c), the proposed SLCD generates a number of line combinations, analyzes all lines in each combination at once, and then finds the most harmonious combination that conveys the global scene composition optimally." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.86, + 0.868 + ], + "angle": 0, + "content": "stage, they extract deep line features to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing redundant lines. Specifically, to refine line candidates, non-maximum suppression (NMS) is performed in [9, 12, 15], as illustrated in Figure 1(a). Lee et al. [15] iteratively select the reliable line near boundary pixels and remove overlapping lines with the selected one. Han et al. [9] simplify the NMS process by adopting a Hough line space. Jin et al. [12] process each candidate through comparative ranking and matching. These techniques, however, do not consider the overall harmony among detected lines. To cope with this issue, Jin et al. [13] estimate the relation score for every pair of detected lines and then decide harmonious semantic lines via graph optimization. But, they may yield sub-optimal results, since only the pairwise relationships between lines are exploited, as in Figure 1(b)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.13, + 0.854, + 0.246, + 0.866 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "28066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.127, + 0.855, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.235, + 0.632, + 0.248 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed SLCD algorithm." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.27, + 0.473, + 0.518 + ], + "angle": 0, + "content": "In this paper, we propose a novel algorithm, called semantic line combination detector (SLCD), to find an optimal group of semantic lines. It processes all lines in a line group (or combination) simultaneously, instead of analyzing each pair of lines, to estimate the overall harmony, as in Figure 1(c). Figure 2 shows an overview of the proposed SLCD. First, we select reliable lines from line candidates and then generate a number of line combinations. Second, we score all line combinations and determine the combination with the highest score as the optimal group of semantic lines. To this end, we design two novel modules for semantic feature grouping and compositional feature extraction. We also introduce a novel loss function to guide the feature grouping module. Experimental results demonstrate that SLCD can detect semantic lines reliably on existing datasets and a new dataset, called compositionally diverse lines (CDL). Moreover, it is shown that SLCD can be used effectively in various applications." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.519, + 0.429, + 0.533 + ], + "angle": 0, + "content": "This work has the following major contributions:" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.533, + 0.472, + 0.561 + ], + "angle": 0, + "content": "- SLCD finds an optimal combination of semantic lines by processing all lines in a line combination at once." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.561, + 0.473, + 0.601 + ], + "angle": 0, + "content": "- We construct the CDL dataset containing compositionally diverse images with implied lines. It will be made publicly available.1" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.602, + 0.473, + 0.658 + ], + "angle": 0, + "content": "- SLCD outperforms conventional detectors on most datasets. Also, its effectiveness is demonstrated in three applications: vanishing point detection, symmetry axis detection, and composition-based image retrieval." + }, + { + "type": "list", + "bbox": [ + 0.113, + 0.533, + 0.473, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.668, + 0.245, + 0.683 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.692, + 0.473, + 0.844 + ], + "angle": 0, + "content": "Semantic lines, located near the boundaries of different semantic regions, outline the layout and composition of an image. They play an important role in various vision applications. Horizons [14, 30, 34, 40] are a specific type of semantic lines, which can be applied to adjust the levelness of images and improve their aesthetics [15, 29]. In [1, 12, 39], dominant parallel lines in the 3D world are detected to estimate vanishing points, which convey depth impression on 2D images. Also, in [3, 6, 7, 20], reflection symmetry axes are identified to analyze the shapes of objects or patterns. These types of lines can be regarded as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.13, + 0.854, + 0.414, + 0.866 + ], + "angle": 0, + "content": "\\(^{1}\\)CDL is available at https://github.com/Jinwon-Ko/SLCD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.858, + 0.339 + ], + "angle": 0, + "content": "highly implied semantic lines. In [13, 26, 36], straight lanes are detected to aid in vehicle maneuvers in road environments. Furthermore, semantic lines are essential visual cues in photographic composition [16, 18]. They direct viewers' attention and help to compose a visually balanced image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.86, + 0.74 + ], + "angle": 0, + "content": "Several semantic line detectors [9, 12, 13, 15] have been proposed. They perform in two stages: line detection and refinement. At the detection stage, deep line features are extracted to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing irrelevant candidates, based on NMS [9, 12, 15] or graph optimization [13]. More specifically, Lee et al. [15] detect reliable lines with classification probabilities higher than a threshold. They then iteratively select semantic lines and remove overlapping lines with the selected one, by employing the edge detector in [31]. Han et al. [9] predict the probability of each candidate in a Hough parametric space. They simplify NMS by computing the centroids of connected components in the Hough space. Jin et al. [12] design two comparators to estimate the priority and similarity between two lines. Then, they determine the most reliable lines and eliminate redundant ones alternately through pairwise comparisons. However, these techniques [9, 12, 15] do not consider how well a group of detected lines represents the global scene structure. To address this issue, Jin et al. [13] analyze the pairwise harmony of detected lines. They first estimate a harmony score for a pair of detected lines. They then construct a complete graph and determine harmonious semantic lines by finding a maximal weight clique. Their method, however, may yield sub-optimal results because it considers only pairwise relationships between lines. On the contrary, the proposed SLCD finds an optimal combination of semantic lines by analyzing all lines in each combination simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.747, + 0.681, + 0.763 + ], + "angle": 0, + "content": "3. Proposed Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.86, + 0.868 + ], + "angle": 0, + "content": "We propose a novel algorithm, called SLCD, to detect an optimal combination of semantic lines, an overview of which is in Figure 2. First, we select \\( K \\) reliable lines from line candidates and then generate a number of line combinations. Second, we score all the line combinations and determine the combination with the highest score as the optimal group of semantic lines." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "28067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.125, + 0.855, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.285, + 0.858, + 0.312 + ], + "angle": 0, + "content": "Figure 3. The architecture of SLCD. CA and FC denote cross-attention and fully connected layers, respectively. The blue, red, yellow, and green boxes indicate encoding, semantic feature grouping, compositional feature extraction, and score regression, respectively." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.335, + 0.365, + 0.349 + ], + "angle": 0, + "content": "3.1. Generating Line Combinations" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.357, + 0.471, + 0.426 + ], + "angle": 0, + "content": "Initializing line candidates: We generate line candidates, which are end-to-end straight lines in an image. Each line candidate is parameterized by polar coordinates in the Hough space [9]. By quantizing the coordinates uniformly, we obtain \\(N\\) line candidates. The default \\(N\\) is 1024." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.43, + 0.474, + 0.679 + ], + "angle": 0, + "content": "Filtering line candidates: Given the \\(N\\) line candidates, there are \\(2^{N}\\) possible line combinations in total. Since this number of combinations is unmanageable, we filter out irrelevant lines to maintain \\(K\\) reliable ones only. To this end, we design a simple line detector by modifying S-Net in [13]. Given an image, the detector obtains a convolutional feature map and extracts line features by aggregating the features of pixels along each line candidate. Then, it computes the classification probability and regression offset of each candidate. We select the most reliable candidate with the highest probability and remove overlapping lines with the selected one. We iterate this process \\(K\\) times to determine \\(K\\) reliable lines. The default \\(K\\) is 8. Figure 4 shows examples of selected reliable lines. Note that, in this filtering, false negatives occur rarely because the \\(K\\) lines are selected via NMS even though their classification probabilities are not high. The architecture of the modified S-Net is described in detail in the supplemental document." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.682, + 0.472, + 0.71 + ], + "angle": 0, + "content": "Generating line combinations: From the \\(K\\) reliable lines, we generate all \\(2^{K}\\) line combinations." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.72, + 0.353, + 0.735 + ], + "angle": 0, + "content": "3.2. Assessing Line Combinations" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.742, + 0.474, + 0.868 + ], + "angle": 0, + "content": "When a line combination divides an image insufficiently or over-segments it into unnecessary parts, it does not describe the overall structure of the scene properly. On the contrary, an optimal combination of semantic lines should convey the image composition reliably and efficiently (i.e. with a small number of lines). To find the best line combination, we develop the semantic line combination detector (SLCD). Figure 3 shows the structure of SLCD, which performs encoding, semantic feature grouping, compositional feature ex" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.334, + 0.858, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.441, + 0.86, + 0.492 + ], + "angle": 0, + "content": "Figure 4. From a set of line candidates, the line detector selects \\( K \\) reliable lines, depicted in orange. A high recall rate is achieved because a sufficient number of reliable lines are selected through NMS. Ground-truth semantic lines in green are in the insets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.859, + 0.548 + ], + "angle": 0, + "content": "traction, and score regression. The detailed architecture of SLCD is described in the supplemental document." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.859, + 0.65 + ], + "angle": 0, + "content": "Encoding: Given an image, we extract multi-scale feature maps of ResNet50 [11]. Then, we match their spatial resolutions via bilinear interpolation and concatenate them in the channel dimension. We then squeeze the channels using 2D convolutional layers to obtain an aggregated feature map \\( F \\in \\mathbb{R}^{HW \\times C} \\), where \\( H \\), \\( W \\), and \\( C \\) are the feature height, the feature width, and the number of channels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.859, + 0.737 + ], + "angle": 0, + "content": "Semantic feature grouping: Semantic lines are located near the boundaries between distinct regions. For their reliable detection, it is desirable to separate different regional parts more clearly. We hence attempt to group pixel features into multiple regions through cross-attention [2, 22, 32, 33]. We employ three cross-attention modules." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.739, + 0.859, + 0.781 + ], + "angle": 0, + "content": "Let \\( R \\in \\mathbb{R}^{M \\times C} \\) be a learnable matrix representing \\( M \\) regions, called region query matrix. Then, we convert \\( R \\) into queries and \\( F \\) into keys and values by" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.795, + 0.858, + 0.812 + ], + "angle": 0, + "content": "\\[\nR _ {q} = R U _ {q}, F _ {k} = (F + S) U _ {k}, F _ {v} = (F + S) U _ {v}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.824, + 0.859, + 0.868 + ], + "angle": 0, + "content": "where \\(S\\in \\mathbb{R}^{HW\\times C}\\) denotes the sinusoidal positional encoding [28], and \\(U_{q},U_{k},U_{v}\\in \\mathbb{R}^{C\\times C}\\) are projection matrices for queries, keys, and values. We then obtain an updated" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "28068" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.124, + 0.47, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.216, + 0.473, + 0.268 + ], + "angle": 0, + "content": "Figure 5. Visualization of semantic feature grouping results. The top row shows input images with ground-truth lines. The bottom one presents the membership maps, representing the semantic region that each pixel belongs to." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.279, + 0.271, + 0.295 + ], + "angle": 0, + "content": "region query matrix \\(\\hat{R}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.302, + 0.472, + 0.325 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {s o f t} _ {M} \\max \\left(R _ {q} F _ {k} ^ {\\mathrm {T}} / \\tau\\right), \\hat {R} = A F _ {v} + R, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.333, + 0.473, + 0.401 + ], + "angle": 0, + "content": "where \\(A\\in \\mathbb{R}^{M\\times HW}\\) is the attention matrix with a scaling factor \\(\\tau\\) and the letter \\(M\\) means the softmax operation is done in the column direction, as in [22, 32, 33]. Then, in the last cross-attention, we generate a semantic feature map \\(F_{s}\\in \\mathbb{R}^{HW\\times C}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.255, + 0.402, + 0.472, + 0.417 + ], + "angle": 0, + "content": "\\[\nF _ {s} = A ^ {\\mathrm {T}} \\hat {R}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.423, + 0.473, + 0.45 + ], + "angle": 0, + "content": "Finally, we channel-wise concatenate \\(F\\) and \\(F_{s}\\) to obtain a combined feature map \\(X\\in \\mathbb{R}^{HW\\times 2C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.451, + 0.473, + 0.533 + ], + "angle": 0, + "content": "Note that, in (2), the softmax function is applied along the query axis. Thus, in \\(A = [\\mathbf{p}_1,\\mathbf{p}_2,\\dots ,\\mathbf{p}_{HW}]\\) each column \\(\\mathbf{p}_i\\in \\mathbb{R}^M\\) is a probability vector. Specifically, the mth element \\(p_i^m\\) in \\(\\mathbf{p}_i\\) is the probability that pixel \\(i\\) belongs to the mth region query. Figure 5 visualizes membership maps, computed by" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.542, + 0.472, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\left[ \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {1}\\right), \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {2}\\right), \\dots , \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {H W}\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.567, + 0.473, + 0.636 + ], + "angle": 0, + "content": "where each element is the index of the region query that the corresponding pixel most likely belongs to. We see that each image is partitioned into \\(M\\) meaningful regions. Thus, \\(F_{s}\\) in (3) represents the regional membership of each pixel and is used to make the feature map \\(X\\) more discriminative." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.636, + 0.473, + 0.663 + ], + "angle": 0, + "content": "To produce the attention matrix \\(A\\) reliably, we design a novel loss function in Section 3.3. The default \\(M\\) is 8." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.667, + 0.473, + 0.778 + ], + "angle": 0, + "content": "Compositional feature extraction: For each line combination, we generate three types of feature maps to extract a compositional feature map. First, we generate a binary mask \\(B \\in \\mathbb{R}^{HW}\\) for each line combination. Each element in \\(B\\) is 1 if the corresponding pixel belongs to any line in the combination, and 0 otherwise. Then, we decompose the combined feature into a line feature map \\(X_{l}\\) and a region feature map \\(X_{r}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.787, + 0.472, + 0.802 + ], + "angle": 0, + "content": "\\[\nX _ {l} = X \\otimes B, X _ {r} = X \\otimes (1 - B), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.811, + 0.473, + 0.867 + ], + "angle": 0, + "content": "where \\(\\otimes\\) is the element-wise multiplication. In other words, \\(X_{l}\\in \\mathbb{R}^{HW\\times 2C}\\) contains contextual information for the line pixels only, whereas \\(X_{r}\\in \\mathbb{R}^{HW\\times 2C}\\) does for the pixels strictly inside the regions divided by the lines." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.125, + 0.858, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.229, + 0.858, + 0.268 + ], + "angle": 0, + "content": "Figure 6. Illustration of the line collection map generation when \\( K = 3 \\): (a) reliable lines, (b) line combinations, and (c) line collection maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.86, + 0.45 + ], + "angle": 0, + "content": "Moreover, we generate a line collection map \\(L = [L_1,L_2,\\dots ,L_K]\\in \\mathbb{R}^{HW\\times K}\\), where \\(L_{k}\\in \\mathbb{R}^{HW}\\) is a ternary mask for the \\(k\\)th reliable line, as illustrated in Figure 6. If the \\(k\\)th reliable line does not belong to the line combination, all elements in \\(L_{k}\\) are set to 0. Otherwise, each value in \\(L_{k}\\) is set to either 1 or \\(-1\\) to indicate where the corresponding pixel is located between the two parts divided by the \\(k\\)th reliable line. In other words, \\(L_{k}\\) informs how the \\(k\\)th line splits the image into two regions. We then produce a positional feature map \\(P\\in \\mathbb{R}^{HW\\times 2C}\\) by applying a series of fully connected layers to \\(L\\). Then, we yield a compositional feature map \\(Z\\in \\mathbb{R}^{HW\\times 6C}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.46, + 0.858, + 0.476 + ], + "angle": 0, + "content": "\\[\nZ = \\left[ X _ {l}, X _ {r}, P \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.859, + 0.515 + ], + "angle": 0, + "content": "which contains information about how the lines in the combination separate the image into multiple parts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.86, + 0.629 + ], + "angle": 0, + "content": "Score regression: Lastly, using a regressor, we estimate a composition score for each line combination. Then, we declare the line combination with the highest score as the optimal group of semantic lines. The regressor takes the compositional feature map \\( Z \\) in (6) as input and predicts a composition score \\( s \\) within [0, 1]. It is implemented using a bilinear interpolation layer and a series of 2D convolution layers and fully connected layers with the ReLU activation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.638, + 0.64, + 0.652 + ], + "angle": 0, + "content": "3.3. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.66, + 0.776, + 0.674 + ], + "angle": 0, + "content": "To train SLCD, we design two loss functions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.859, + 0.771 + ], + "angle": 0, + "content": "Semantic region separation loss: When two pixels \\(i\\) and \\(j\\) are located in distinct regions, they should be assigned to different region queries. In other words, their probability vectors \\(\\mathbf{p}_i\\) and \\(\\mathbf{p}_j\\) in the attention matrix \\(A\\) should be far from each other. Let \\(X\\) and \\(Y\\) be the two regions divided by a ground-truth line. Their probability vectors are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.771, + 0.858, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {X} = \\frac {1}{| X |} \\sum_ {i \\in X} \\mathbf {p} _ {i}, \\quad \\mathbf {p} _ {Y} = \\frac {1}{| Y |} \\sum_ {j \\in Y} \\mathbf {p} _ {j}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.857, + 0.838 + ], + "angle": 0, + "content": "which should be also far from each other, for the ground-truth line divides the image into two semantic regions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.858, + 0.867 + ], + "angle": 0, + "content": "To measure the distance between probability vectors, we adopt the Kullback-Leibler divergence (KLD) [5]. Then," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "28069" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.127, + 0.46, + 0.137 + ], + "angle": 0, + "content": "Horizontal Vertical Diagonal Triangle Symmetric Low Front" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.137, + 0.47, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.263, + 0.473, + 0.315 + ], + "angle": 0, + "content": "Figure 7. The proposed CDL dataset contains 9,100 images in seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. The ground-truth semantic lines are depicted in green." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.331, + 0.473, + 0.359 + ], + "angle": 0, + "content": "we define the semantic region segmentation (SRS) loss, by employing all ground-truth lines, as" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.368, + 0.473, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S R S}} = - \\sum_ {l = 1} ^ {T} \\left[ D \\left(\\mathbf {p} _ {X _ {l}} \\| \\mathbf {p} _ {Y _ {l}}\\right) + D \\left(\\mathbf {p} _ {Y _ {l}} \\| \\mathbf {p} _ {X _ {l}}\\right) \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.414, + 0.473, + 0.512 + ], + "angle": 0, + "content": "where \\(D\\) denotes the KLD, \\(\\mathbf{p}_{X_l}\\) and \\(\\mathbf{p}_{Y_l}\\) are the regions divided by the \\(l\\)th ground-truth line, and \\(T\\) is the number of ground-truth lines. Both \\(D(\\mathbf{p}_{X_l} \\parallel \\mathbf{p}_{Y_l})\\) and \\(D(\\mathbf{p}_{Y_l} \\parallel \\mathbf{p}_{X_l})\\) are computed because \\(D\\) is not commutative. With \\(\\mathcal{L}_{\\mathrm{SRS}}\\), it is possible to roughly segment an image into meaningful parts, as illustrated in Figure 5, even though no ground-truth labels for semantic segmentation are used for training." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.514, + 0.473, + 0.625 + ], + "angle": 0, + "content": "Regression loss: We also design a loss for regressing the composition score of each line combination. To this end, we employ the HIoU metric [13] that quantifies the structural layout of a line combination. Let \\(\\mathbf{c}\\) and \\(\\mathbf{c}^{\\star}\\) denote a line combination and the ground-truth combination, respectively. We compute the HIoU between \\(\\mathbf{c}\\) and \\(\\mathbf{c}^{\\star}\\) and use it as the ground-truth composition score \\(\\overline{s}\\) of \\(\\mathbf{c}\\). Then, the regression loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.633, + 0.472, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = (s - \\bar {s}) ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.659, + 0.473, + 0.688 + ], + "angle": 0, + "content": "where \\(s\\) is the predicted score of \\(\\mathbf{c}\\). To reduce \\(\\mathcal{L}_{\\mathrm{reg}}\\) effectively, a ranking loss \\(\\mathcal{L}_{\\mathrm{rank}}\\) is also employed as in [17]." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.698, + 0.303, + 0.714 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.721, + 0.313, + 0.736 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.742, + 0.474, + 0.867 + ], + "angle": 0, + "content": "We adopt ResNet50 [11] as the encoder of the proposed SLCD. We use the AdamW optimizer [19] with a learning rate \\( 10^{-4} \\), a weight decay of \\( 10^{-4} \\), \\( \\gamma = 0.5 \\), \\( \\beta_{1} = 0.9 \\) and \\( \\beta_{2} = 0.999 \\). We use a batch size of two for 400,000 iterations. Training images are resized to \\( 480 \\times 480 \\) and augmented by random horizontal flipping. We fix the number of line candidates, reliable lines, and region queries to \\( N = 1024 \\), \\( K = 8 \\), and \\( M = 8 \\), respectively. We set \\( H = 60 \\), \\( W = 60 \\), and \\( C = 96 \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.125, + 0.858, + 0.15 + ], + "angle": 0, + "content": "Table 1. Comparison of the HIoU scores \\((\\%)\\) on the SEL, SEL_Hard, NKL, and CDL datasets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.155, + 0.858, + 0.23 + ], + "angle": 0, + "content": "
SELSEL_HardNKLCDL
SLNet [15]77.8759.7165.4957.78
DHT [9]79.6263.3969.0863.24
DRM [12]80.2368.8367.4263.96
HSLD [13]81.0365.9974.2964.98
Proposed84.0968.1576.2168.85
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.252, + 0.593, + 0.265 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.859, + 0.356 + ], + "angle": 0, + "content": "SEL [15]: It is the first dataset for semantic line detection. It contains 1,750 images, which are split into 1,575 training and 175 test images. Each semantic line is annotated by the coordinates of two endpoints on an image boundary. In SEL, most images are high-quality landscapes, in which semantic lines are often obvious." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.859, + 0.401 + ], + "angle": 0, + "content": "SEL_Hard [12]: It contains 300 test images only, selected from the ADE20K segmentation dataset [37]. It is challenging because of cluttered scenes or occluded objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.859, + 0.445 + ], + "angle": 0, + "content": "NKL [35]: It is a relatively large dataset of 5,200 training and 1,300 testing images. It includes both indoor and outdoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.86, + 0.601 + ], + "angle": 0, + "content": "CDL: We construct CDL to contain 7,100 scenes with diverse contents and compositions. It is split into 6,390 training and 710 test images. As in Figure 7, the images are categorized into seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. In Low and Front, humans and animals are essential parts of the image composition. In the other classes, most images are outdoor ones, as in the other datasets. To construct CDL, semantic lines were manually annotated in about 200 man-hours. More examples and the annotation process are provided in the supplemental document." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.589, + 0.621 + ], + "angle": 0, + "content": "4.3. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.859, + 0.713 + ], + "angle": 0, + "content": "As mentioned earlier, the harmony of detected lines is more important than individual lines in semantic line detection. Therefore, the main paper discusses only HIoU performances. The precision, recall, and f-measure results, which are metrics used in previous work, are compared in the supplemental document." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.859, + 0.84 + ], + "angle": 0, + "content": "HIoU: An optimal group of semantic lines conveys a harmonious impression about the composition of an image. To assess the overall harmony of detection results, we use the HIoU metric [13]. It measures the consistency between the division of an image by detected lines and that by the ground truth. Let \\( S = \\{s_1,s_2,\\dots ,s_N\\} \\) and \\( T = \\{t_{1},t_{2},\\ldots ,t_{M}\\} \\) be the regions divided by the detected lines and the ground-truth lines, respectively. Then, HIoU is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.846, + 0.858, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H I o U} = \\frac {\\sum_ {i = 1} ^ {N} \\max _ {k} \\operatorname {I o U} \\left(s _ {i} , t _ {k}\\right) + \\sum_ {j = 1} ^ {M} \\max _ {k} \\operatorname {I o U} \\left(t _ {j} , s _ {k}\\right)}{N + M}. \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "28070" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.116, + 0.859, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.439, + 0.86, + 0.464 + ], + "angle": 0, + "content": "Figure 8. Comparison of detected semantic lines. From the left, two images are selected from each of the SEL, SEL_Hard, NKL, and CDL datasets." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.482, + 0.475, + 0.553 + ], + "angle": 0, + "content": "Specifically, for each \\( s_i \\), the most overlapping region \\( t_k \\) is chosen and the intersection-over-union (IoU) is calculated. This is performed similarly for each \\( t_j \\). Then, the HIoU score is given by the average of these bi-directionally matching IoUs." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.561, + 0.324, + 0.576 + ], + "angle": 0, + "content": "4.4. Comparative Assessment" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.583, + 0.474, + 0.708 + ], + "angle": 0, + "content": "Table 1 compares the HIoU scores of the proposed SLCD with those of the existing detectors [9, 12, 13, 15] on the SEL, SEL_Hard, NKL, and CDL datasets. Also, Figure 8 compares detection results qualitatively. The existing detectors miss correct lines or fail to remove redundant lines, yielding sub-optimal results. In contrast, SLCD detects semantic lines more precisely and represents the composition more reliably than the existing detectors do. More detection results are available in the supplemental document." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.711, + 0.472, + 0.807 + ], + "angle": 0, + "content": "Comparison on SEL: In Table 1, SLCD outperforms the existing detectors on SEL. Compared with the second-best HSLD, SLCD yields a wide HIoU margin of 3.06. This indicates that SLCD finds an optimal group of semantic lines more effectively by processing all lines in each combination simultaneously, instead of performing pairwise comparisons in HSLD." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.811, + 0.474, + 0.868 + ], + "angle": 0, + "content": "Comparison on SEL_Hard: As done in [12], we conduct experiments on SEL_Hard using the networks trained on the SEL dataset. In Table 1, SLCD ranks 2nd on SEL_Hard. DRM provides a better result than SLCD, but it demands" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.481, + 0.859, + 0.507 + ], + "angle": 0, + "content": "Table 2. Comparison of the HIoU scores \\((\\%)\\) on the CDL dataset according to the number \\(M\\) of region queries." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.512, + 0.848, + 0.546 + ], + "angle": 0, + "content": "
M481632
H IoU66.4368.8568.1167.07
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.557, + 0.857, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.655, + 0.857, + 0.669 + ], + "angle": 0, + "content": "Figure 9. Visualization of the membership maps according to \\(M\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.857, + 0.704 + ], + "angle": 0, + "content": "much higher complexity, as will be discussed in Section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.709, + 0.859, + 0.75 + ], + "angle": 0, + "content": "Comparison on NKL: SLCD surpasses all existing detectors. For example, its HIoU score is 1.92 points higher than the second-best HSLD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.86, + 0.868 + ], + "angle": 0, + "content": "Comparison on CDL: Table 1 also lists HIoU scores on the proposed CDL dataset. For a fair comparison, we train the existing detectors on CDL using their publicly available source codes. We see that SLCD outperforms all existing detectors on CDL as well. SLNet, DHT, and DRM yield poor results because they do not consider the overall harmony of detected lines. HSLD is better than these detectors but is inferior to the proposed SLCD. We compare and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.518, + 0.941 + ], + "angle": 0, + "content": "28071" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.124, + 0.473, + 0.15 + ], + "angle": 0, + "content": "Table 3. Comparison of the HIoU scores \\((\\%)\\) on the CDL dataset according to the usage of the SRS loss \\(\\mathcal{L}_{\\mathrm{SRS}}\\)" + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.155, + 0.462, + 0.19 + ], + "angle": 0, + "content": "
w/o LSRProposed
HIoU66.7168.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.203, + 0.471, + 0.227 + ], + "angle": 0, + "content": "Table 4. Ablation studies for the compositional feature extraction on the CDL dataset." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.232, + 0.462, + 0.301 + ], + "angle": 0, + "content": "
LineRegionPositionalHIoU
I✓65.12
II✓65.74
III✓✓66.37
IV✓✓✓68.85
" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.324, + 0.471, + 0.352 + ], + "angle": 0, + "content": "discuss the detection results according to the composition classes in the supplemental document." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.361, + 0.208, + 0.376 + ], + "angle": 0, + "content": "4.5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.383, + 0.471, + 0.41 + ], + "angle": 0, + "content": "We conduct several ablation studies to analyze each component of the proposed SLCD." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.414, + 0.473, + 0.497 + ], + "angle": 0, + "content": "The number of \\(M\\): Table 2 lists the HIoU performances on the CDL dataset, according to the number \\(M\\) of region queries. At the default \\(M = 8\\), SLCD achieves the best performance. When \\(M\\) is smaller or bigger than 8, the performance degrades due to under- or over-segmentation of semantic parts, respectively, as shown in Figure 9." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.5, + 0.473, + 0.569 + ], + "angle": 0, + "content": "Efficacy of \\(\\mathcal{L}_{\\mathrm{SRS}}\\): In Table 3, if the proposed SRS loss \\(\\mathcal{L}_{\\mathrm{SRS}}\\) in (8) is excluded from the training, the performance drops by 2.14 points. It means that the composition analysis based on the SRS loss is essential for finding optimal line combinations." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.573, + 0.472, + 0.683 + ], + "angle": 0, + "content": "Efficacy of compositional feature extraction: SLCD extracts the compositional feature map \\( Z \\) for each line combination, by processing the line feature map \\( X_{l} \\), the region feature map \\( X_{r} \\), and the positional feature map \\( P \\) in (6). In Table 4, Method I uses the line feature map only, while II uses the region feature map only. Method III utilizes both feature maps. Method IV, which is the proposed SLCD, uses all three maps." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.684, + 0.472, + 0.808 + ], + "angle": 0, + "content": "Method I yields the worst result, for it uses only the contextual information near line pixels. Method II is slightly better than Method I, by exploiting the regional information. Using both line and region feature maps, III provides a better result. Moreover, IV further improves the performance significantly by utilizing the positional feature map. This is because the overall harmony is estimated more effectively, by combining the line structures in the positional feature map with scene contexts." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.811, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Runtime: Table 5 compares the runtimes of SLCD and the existing detectors in seconds per frame (spf). We use a PC with AMD Ryzen 9 3900X CPU and NVIDIA RTX 2080 GPU. The proposed SLCD takes 0.114spf, adding" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.124, + 0.859, + 0.163 + ], + "angle": 0, + "content": "Table 5. Runtime comparison of the proposed SLCD and the existing detectors. The processing times are reported in seconds per frame (spf)." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.168, + 0.851, + 0.203 + ], + "angle": 0, + "content": "
SLNet [15]DHT [9]DRM [12]HSLD [13]Proposed
0.1360.0330.9520.0460.114
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.217, + 0.842, + 0.23 + ], + "angle": 0, + "content": "Table 6. Comparison of the AA scores \\((\\%)\\) of VP detection." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.234, + 0.851, + 0.294 + ], + "angle": 0, + "content": "
AA1°AA2°AA10°
Zhou et al. [39]18.533.060.0
Jin et al. [12]8.622.968.3
Proposed16.636.978.3
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.309, + 0.574, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.309, + 0.644, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.351, + 0.644, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.309, + 0.714, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.351, + 0.714, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.309, + 0.785, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.351, + 0.785, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.309, + 0.855, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.351, + 0.855, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.401, + 0.859, + 0.44 + ], + "angle": 0, + "content": "Figure 10. Detected dominant parallel lines and their intersections are depicted by yellow lines and green triangles, respectively. The ground-truth VPs are depicted by red dots." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.859, + 0.526 + ], + "angle": 0, + "content": "up 0.030spf and 0.074spf for generating and assessing line combinations, respectively. DHT is the fastest detector, but it is inferior to SLCD on all datasets. On the other hand, even though DRM performs better on SEL_Hard, it is about 8.4 times slower than SLCD." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.619, + 0.559 + ], + "angle": 0, + "content": "5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.859, + 0.651 + ], + "angle": 0, + "content": "We apply the proposed SLCD to three vision tasks: dominant vanishing point detection, reflection symmetry axis detection, and composition-based image retrieval. Note that the first two tasks were considered in [12]. Due to the page limit, more details and results are described in the supplemental document." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.664, + 0.791, + 0.679 + ], + "angle": 0, + "content": "5.1. Dominant Vanishing Point Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.86, + 0.868 + ], + "angle": 0, + "content": "A vanishing point (VP) facilitates understanding of the 3D geometric structure. We apply the proposed SLCD to detect dominant VPs, by identifying vanishing lines. We use the AVA landscape dataset [39], in which two dominant parallel lines are annotated for each image. To detect a dominant VP, we generate line combinations containing two lines only. We then find the best combination, whose intersecting point is declared as a VP. Table 6 compares this VP detection scheme with the existing line-based VP detectors in [12, 39]. The angle accuracies (AAs) [38] are compared. We see that SLCD is better than the existing detectors, except that its \\(\\mathrm{AA1}^{\\circ}\\) score is lower than that of Zhou et al. [39]. Figure 10 shows some VP detection results." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "28072" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.124, + 0.471, + 0.149 + ], + "angle": 0, + "content": "Table 7. Comparison of the AUC_A scores \\((\\%)\\) of symmetry axis detection." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.154, + 0.462, + 0.242 + ], + "angle": 0, + "content": "
ICCVNYUSYM_Hard
Cicconet et al. [3]80.8082.8568.99
Elawady et al. [6]87.2483.8373.90
Cicconet et al. [4]87.3887.6481.04
Loy & Eklundh [20]89.7790.8581.99
Jin et al. [12]90.6092.7884.73
Proposed93.1593.5188.03
" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.244, + 0.47, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.331, + 0.473, + 0.382 + ], + "angle": 0, + "content": "Figure 11. Symmetry axis detection results. In the top row, the ground-truth and predicted axes are depicted by dashed red and solid yellow lines, respectively. The membership maps are also visualized in the bottom row." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.393, + 0.4, + 0.408 + ], + "angle": 0, + "content": "5.2. Reflection Symmetry Axis Detection" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.41, + 0.474, + 0.604 + ], + "angle": 0, + "content": "Reflection symmetry is a visual cue for analyzing object shapes and patterns. Its detection is challenging because the symmetry is only implicit in many cases. We test the proposed SLCD on three datasets: ICCV [7], NYU [4], and SYM_Hard [12]. In these datasets, each image contains a single reflection symmetry axis. Thus, in the proposed SLCD, each of the \\(K\\) reliable lines is regarded as a line combination. Table 7 compares the AUC_A scores [15] of the proposed SLCD and the conventional techniques [3, 4, 6, 12, 20]. SLCD outperforms all these techniques on all datasets. Figure 11 shows some detection results together with the membership maps. We see that regional parts are symmetrically divided along implied axes, enabling SLCD to identify those axes effectively." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.614, + 0.4, + 0.629 + ], + "angle": 0, + "content": "5.3. Composition-Based Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.631, + 0.472, + 0.7 + ], + "angle": 0, + "content": "Existing image retrieval techniques focus on the visual contents of a query image to find similar images in a database [8, 21, 24, 25]. In this work, however, we attempt to discover images with similar compositions to a query image, i.e. composition-based image retrieval." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.7, + 0.472, + 0.838 + ], + "angle": 0, + "content": "We test the proposed SLCD on the Oxford 5k and Paris 6k retrieval dataset [23]. We first detect semantic lines in every image, while storing the positional feature map \\(P\\) in (6). We filter out some images whose composition scores are lower than a threshold since a low score indicates a low-quality image with inharmonious composition in general. Then, for a randomly selected query image, we compute the \\(\\ell_2\\)-distances between the positional feature maps \\(P\\) of the query and the remaining images. We determine the images with the smallest distances as retrieval results." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.839, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Figure 12 shows the top-4 retrieval results for four query images. We see that SLCD returns structurally similar im" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.125, + 0.576, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.248, + 0.573, + 0.258 + ], + "angle": 0, + "content": "Query image" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.125, + 0.857, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.248, + 0.771, + 0.258 + ], + "angle": 0, + "content": "Top-4 retrieval results" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.26, + 0.854, + 0.273 + ], + "angle": 0, + "content": "Figure 12. Composition-based retrieval results for query images." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.279, + 0.856, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.423, + 0.858, + 0.448 + ], + "angle": 0, + "content": "Figure 13. t-SNE visualization [27] of the feature space for the Oxford 5k and Paris 6k dataset [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.859, + 0.598 + ], + "angle": 0, + "content": "ages to the queries. This means that the positional feature maps \\(P\\), containing the structural information of semantic lines, can be used to compute the compositional differences. Based on this observation, we perform data clustering by employing k-means [10] on the feature space of \\(P\\). Figure 13 is t-SNE visualization [27] of the clustering results. The nine images nearest to each centroid are also shown. Note that images with similar composition are grouped into the same cluster, confirming the efficacy of SLCD in the compositional analysis of images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.616, + 0.619 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.859, + 0.763 + ], + "angle": 0, + "content": "We proposed a novel semantic line detector, SLCD, which processes a combination of lines at once to estimate the overall harmony reliably. We first generated all possible line combinations from reliable lines. Then, we estimated the score of each line combination and determined the best combination. Experimental results demonstrated that the proposed SLCD can detect semantic lines reliably on existing datasets, as well as on the new dataset CDL. Furthermore, SLCD can be successfully used in vanishing point detection, symmetry axis detection, and image retrieval." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.655, + 0.782 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.859, + 0.867 + ], + "angle": 0, + "content": "This work was conducted by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD) and was supported by the National Research Foundation of Korea (NRF) grants funded by the Korea government (MSIT) (No. NRF-2021R1A4A1031864 and No. NRF-2022R1A2B5B03002310)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "28073" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.125, + 0.201, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.148, + 0.473, + 0.21 + ], + "angle": 0, + "content": "[1] Jean-Charles Bazin, Yongduek Seo, Cedric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in Manhattan world. In Proc. IEEE CVPR, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.212, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.264, + 0.472, + 0.313 + ], + "angle": 0, + "content": "[3] Marcelo Cicconet, Vighnesh Birodkar, Mads Lund, Michael Werman, and Davi Geiger. A convolutional approach to reflection symmetry. Pattern Recog. Lett., 95:44-50, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.315, + 0.472, + 0.365 + ], + "angle": 0, + "content": "[4] Marcelo Cicconet, David GC Hildebrand, and Hunter Elliott. Finding mirror symmetry via registration and optimal symmetric pairwise assignment of curves: Algorithm and results. In Proc. IEEE ICCV Workshops, 2017. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.367, + 0.472, + 0.391 + ], + "angle": 0, + "content": "[5] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley, 2nd edition, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.392, + 0.472, + 0.442 + ], + "angle": 0, + "content": "[6] Mohamed Elawady, Christophe Ducottet, Olivier Alata, Cécile Barat, and Philippe Colantoni. Wavelet-based reflection symmetry detection via textural and color histograms. In Proc. IEEE ICCV Workshops, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.444, + 0.472, + 0.493 + ], + "angle": 0, + "content": "[7] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 ICCV Challenge: Detecting symmetry in the wild. In Proc. IEEE ICCV, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.495, + 0.472, + 0.533 + ], + "angle": 0, + "content": "[8] Albert Gordo, Jon Almazan, Jerome Revaud, and Diane Larlus. End-to-end learning of deep visual representations for image retrieval. Int. J. Comput. Vis., 124:237-254, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.534, + 0.472, + 0.571 + ], + "angle": 0, + "content": "[9] Qi Han, Kai Zhao, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. In Proc. ECCV, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.572, + 0.472, + 0.609 + ], + "angle": 0, + "content": "[10] J. A. Hartigan and M. A. Wong. A K-means clustering algorithm. Journal of the Royal Statistical Society. Series C (Applied Statistics), 28(1):100-108, 1979. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.611, + 0.472, + 0.647 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.649, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[12] Dongkwon Jin, Jun-Tae Lee, and Chang-Su Kim. Semantic line detection using mirror attention and comparative ranking and matching. In Proc. ECCV, 2020. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.688, + 0.472, + 0.737 + ], + "angle": 0, + "content": "[13] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, and Chang-Su Kim. Harmonious semantic line detection via maximal weight clique selection. In Proc. IEEE CVPR, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.739, + 0.472, + 0.776 + ], + "angle": 0, + "content": "[14] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Temporally consistent horizon lines. In Proc. IEEE ICRA, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.778, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[15] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Semantic line detection and its applications. In Proc. IEEE ICCV, 2017. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.817, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[16] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Photographic composition classification and dominant geometric element detection for outdoor scenes. J. Vis. Commun. Image Represent., 55:91-105, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.148, + 0.473, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.165 + ], + "angle": 0, + "content": "[17] Debang Li, Junge Zhang, Kaiqi Huang, and Ming-Hsuan Yang. Composing good shots by exploiting mutual relations. In Proc. IEEE CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.857, + 0.204 + ], + "angle": 0, + "content": "[18] Ligang Liu, Renjie Chen, Lior Wolf, and Daniel Cohen-Or. Optimizing photo composition. Computer Graphics Forum, 29:469-478, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.857, + 0.231 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.857, + 0.269 + ], + "angle": 0, + "content": "[20] Gareth Loy and Jan-Olof Eklundh. Detecting symmetry and symmetric constellations of features. In Proc. ECCV, 2006. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.271, + 0.857, + 0.309 + ], + "angle": 0, + "content": "[21] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proc. IEEE ICCV, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.857, + 0.348 + ], + "angle": 0, + "content": "[22] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. iDisc: Internal discretization for monocular depth estimation. In Proc. IEEE CVPR, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.857, + 0.399 + ], + "angle": 0, + "content": "[23] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proc. IEEE CVPR, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.857, + 0.438 + ], + "angle": 0, + "content": "[24] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell., 41:1655-1668, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.857, + 0.49 + ], + "angle": 0, + "content": "[25] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. IEEE Trans. Pattern Anal. Mach. Intell., 22:1349-1380, 2000. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.857, + 0.542 + ], + "angle": 0, + "content": "[26] Lucas Tabelini, Rodrigo Berriel, Thiago M. Paixao, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.544, + 0.857, + 0.581 + ], + "angle": 0, + "content": "[27] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.583, + 0.857, + 0.632 + ], + "angle": 0, + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.857, + 0.696 + ], + "angle": 0, + "content": "[29] Yaoting Wang, Yongzhen Ke, Kai Wang, Jing Guo, and Shuai Yang. Spatial-invariant convolutional neural network for photographic composition prediction and automatic correction. J. Vis. Commun. Image Represent., 90:103751, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.857, + 0.723 + ], + "angle": 0, + "content": "[30] Scott Workman, Menghua Zhai, and Nathan Jacobs. Horizon lines in the wild. In Proc. BMVC, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.726, + 0.857, + 0.75 + ], + "angle": 0, + "content": "[31] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proc. IEEE ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.752, + 0.857, + 0.814 + ], + "angle": 0, + "content": "[32] Qihang Yu, Huiyu Wang, Dahun Kim, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. CMT-DeepLab: Clustering mask transformers for panoptic segmentation. In Proc. IEEE CVPR, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.857, + 0.865 + ], + "angle": 0, + "content": "[33] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. \\(k\\)-means mask transformer. In Proc. ECCV, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.857, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "28074" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.127, + 0.472, + 0.165 + ], + "angle": 0, + "content": "[34] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proc. IEEE CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.167, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[35] Kai Zhao, Qi Han, Chang-Bin Zhang, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. IEEE Trans. Pattern Anal. Mach. Intell., 44:4793-4806, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.218, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[36] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. CLRNet: Cross layer refinement network for lane detection. In Proc. IEEE CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.271, + 0.472, + 0.307 + ], + "angle": 0, + "content": "[37] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In Proc. IEEE CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.31, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[38] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. NeurVPS: Neural vanishing point scanning via conic convolution. In Proc. NeurIPS, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.349, + 0.472, + 0.398 + ], + "angle": 0, + "content": "[39] Zihan Zhou, Farshid Farhat, and James Z Wang. Detecting dominant vanishing points in natural scenes with application to composition-sensitive image retrieval. IEEE Trans. Multimedia, 19:2651-2665, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.401, + 0.472, + 0.45 + ], + "angle": 0, + "content": "[40] Rui Zhu, Xingyi Yang, Yannick Hold-Geoffroy, Federico Perazzi, Jonathan Eisenmann, Kalyan Sunkavalli, and Manmohan Chandraker. Single view metrology in the wild. In Proc. ECCV, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.127, + 0.472, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.94 + ], + "angle": 0, + "content": "28075" + } + ] +] \ No newline at end of file diff --git a/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_origin.pdf b/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..753dedc0ff220849bc5d068c13773ffc6630482d --- /dev/null +++ b/2024/Semantic Line Combination Detector/09fa8b25-63f5-44a7-8c3c-7ce99a11a2c1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83e28ddd595b3d4f952c0ff41d909811ec29687b85f19360ba45cafaa449c5f7 +size 4650583 diff --git a/2024/Semantic Line Combination Detector/full.md b/2024/Semantic Line Combination Detector/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a3cbfb81b4457e9b7ea5bd4a7dab61cd67d3631 --- /dev/null +++ b/2024/Semantic Line Combination Detector/full.md @@ -0,0 +1,389 @@ +# Semantic Line Combination Detector + +Jinwon Ko + +Korea University + +jwko@mcl.korea.ac.kr + +Dongkwon Jin + +Korea University + +dongkwonjin@mcl.korea.ac.kr + +Chang-Su Kim* + +Korea University + +changsukim@korea.ac.kr + +# Abstract + +A novel algorithm, called semantic line combination detector (SLCD), to find an optimal combination of semantic lines is proposed in this paper. It processes all lines in each line combination at once to assess the overall harmony of the lines. First, we generate various line combinations from reliable lines. Second, we estimate the score of each line combination and determine the best one. Experimental results demonstrate that the proposed SLCD outperforms existing semantic line detectors on various datasets. Moreover, it is shown that SLCD can be applied effectively to three vision tasks of vanishing point detection, symmetry axis detection, and composition-based image retrieval. Our codes are available at https://github.com/Jinwon-Ko/SLCD. + +# 1. Introduction + +A semantic line [15] is defined as a meaningful line separating distinct semantic regions in an image. Besides this unary definition, multiple semantic lines in an image are supposed to convey the global scene structure properly [13]. It is challenging to detect such semantic lines because they are often implied by complicated region boundaries. Moreover, they should represent the image composition optimally by dividing it into semantic regions harmoniously. + +Semantic lines are essential elements in many vision applications. For example, a horizon line [14, 30, 34, 40], which is a specific type of semantic line, can be exploited to adjust the levelness of an image [15, 29]. A reflection symmetry axis [3, 4, 6, 7, 20], which is another type of semantic line, provides visual cues for object recognition and pattern analysis. Vanishing points, conveying depth impression in images, can be estimated by detecting dominant parallel semantic lines in the 3D world [1, 12, 39]. In autonomous driving systems [13, 26, 36], boundaries of road lanes can be also described by semantic lines. + +Recently, several attempts [9, 12, 13, 15, 35] have been made to detect semantic lines. These techniques perform line detection and refinement sequentially. At the detection + +![](images/0eb69331fb07e6bf0443e794367a4b0c94d7a7b8b7cf946f443de0f04bfb0785.jpg) +Input + +![](images/976615b3f9f48847ab871ec52e7071aa26f8150d32784f5c324ea2d15d079bd5.jpg) +(a) NMS + +![](images/a1cfc8e91f1857e4e81e2238a373ea847ce93b930b4a03ce18bf749fdc5b0011.jpg) +(b) Pairwise correlation + +![](images/9cb3c69de85b36a772ea0bb6226738a4c810e2f30914daf1c934880e5514946c.jpg) +Ground-truth +Figure 1. After selecting reliable line candidates, there are two existing approaches to semantic line detection. The first approach in (a) focuses on locating a line near a region boundary and eliminating overlapping lines. However, a redundant line still remains, since this approach does not consider how well a group of detected lines represents the layout of a scene. The second approach in (b) takes into account only the pairwise correlation between two lines, so it may fail to assess the overall harmony of more than two semantic lines. In contrast, in (c), the proposed SLCD generates a number of line combinations, analyzes all lines in each combination at once, and then finds the most harmonious combination that conveys the global scene composition optimally. + +![](images/e31170d5cdeacfff1e62c5bff86153b2e160a3b78829f5183884e485a47e05ee.jpg) +(c) Proposed SLCD + +![](images/56e41afe1fb1bdb8517a6f2c9bfffeffab508d297e7df904cfe1ddef5aedad07.jpg) + +stage, they extract deep line features to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing redundant lines. Specifically, to refine line candidates, non-maximum suppression (NMS) is performed in [9, 12, 15], as illustrated in Figure 1(a). Lee et al. [15] iteratively select the reliable line near boundary pixels and remove overlapping lines with the selected one. Han et al. [9] simplify the NMS process by adopting a Hough line space. Jin et al. [12] process each candidate through comparative ranking and matching. These techniques, however, do not consider the overall harmony among detected lines. To cope with this issue, Jin et al. [13] estimate the relation score for every pair of detected lines and then decide harmonious semantic lines via graph optimization. But, they may yield sub-optimal results, since only the pairwise relationships between lines are exploited, as in Figure 1(b). + +![](images/cf3196727fb63a2bdfe55217b71bf63d4b9a90129c67f9cd4d02768c30b054b6.jpg) +Figure 2. Overview of the proposed SLCD algorithm. + +In this paper, we propose a novel algorithm, called semantic line combination detector (SLCD), to find an optimal group of semantic lines. It processes all lines in a line group (or combination) simultaneously, instead of analyzing each pair of lines, to estimate the overall harmony, as in Figure 1(c). Figure 2 shows an overview of the proposed SLCD. First, we select reliable lines from line candidates and then generate a number of line combinations. Second, we score all line combinations and determine the combination with the highest score as the optimal group of semantic lines. To this end, we design two novel modules for semantic feature grouping and compositional feature extraction. We also introduce a novel loss function to guide the feature grouping module. Experimental results demonstrate that SLCD can detect semantic lines reliably on existing datasets and a new dataset, called compositionally diverse lines (CDL). Moreover, it is shown that SLCD can be used effectively in various applications. + +This work has the following major contributions: + +- SLCD finds an optimal combination of semantic lines by processing all lines in a line combination at once. +- We construct the CDL dataset containing compositionally diverse images with implied lines. It will be made publicly available.1 +- SLCD outperforms conventional detectors on most datasets. Also, its effectiveness is demonstrated in three applications: vanishing point detection, symmetry axis detection, and composition-based image retrieval. + +# 2. Related Work + +Semantic lines, located near the boundaries of different semantic regions, outline the layout and composition of an image. They play an important role in various vision applications. Horizons [14, 30, 34, 40] are a specific type of semantic lines, which can be applied to adjust the levelness of images and improve their aesthetics [15, 29]. In [1, 12, 39], dominant parallel lines in the 3D world are detected to estimate vanishing points, which convey depth impression on 2D images. Also, in [3, 6, 7, 20], reflection symmetry axes are identified to analyze the shapes of objects or patterns. These types of lines can be regarded as + +highly implied semantic lines. In [13, 26, 36], straight lanes are detected to aid in vehicle maneuvers in road environments. Furthermore, semantic lines are essential visual cues in photographic composition [16, 18]. They direct viewers' attention and help to compose a visually balanced image. + +Several semantic line detectors [9, 12, 13, 15] have been proposed. They perform in two stages: line detection and refinement. At the detection stage, deep line features are extracted to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing irrelevant candidates, based on NMS [9, 12, 15] or graph optimization [13]. More specifically, Lee et al. [15] detect reliable lines with classification probabilities higher than a threshold. They then iteratively select semantic lines and remove overlapping lines with the selected one, by employing the edge detector in [31]. Han et al. [9] predict the probability of each candidate in a Hough parametric space. They simplify NMS by computing the centroids of connected components in the Hough space. Jin et al. [12] design two comparators to estimate the priority and similarity between two lines. Then, they determine the most reliable lines and eliminate redundant ones alternately through pairwise comparisons. However, these techniques [9, 12, 15] do not consider how well a group of detected lines represents the global scene structure. To address this issue, Jin et al. [13] analyze the pairwise harmony of detected lines. They first estimate a harmony score for a pair of detected lines. They then construct a complete graph and determine harmonious semantic lines by finding a maximal weight clique. Their method, however, may yield sub-optimal results because it considers only pairwise relationships between lines. On the contrary, the proposed SLCD finds an optimal combination of semantic lines by analyzing all lines in each combination simultaneously. + +# 3. Proposed Algorithm + +We propose a novel algorithm, called SLCD, to detect an optimal combination of semantic lines, an overview of which is in Figure 2. First, we select $K$ reliable lines from line candidates and then generate a number of line combinations. Second, we score all the line combinations and determine the combination with the highest score as the optimal group of semantic lines. + +![](images/9b3a72ca4e2fa6e9e65d80f0a5a528ff6a36907331dfac4b96e86c5b476770b3.jpg) +Figure 3. The architecture of SLCD. CA and FC denote cross-attention and fully connected layers, respectively. The blue, red, yellow, and green boxes indicate encoding, semantic feature grouping, compositional feature extraction, and score regression, respectively. + +# 3.1. Generating Line Combinations + +Initializing line candidates: We generate line candidates, which are end-to-end straight lines in an image. Each line candidate is parameterized by polar coordinates in the Hough space [9]. By quantizing the coordinates uniformly, we obtain $N$ line candidates. The default $N$ is 1024. + +Filtering line candidates: Given the $N$ line candidates, there are $2^{N}$ possible line combinations in total. Since this number of combinations is unmanageable, we filter out irrelevant lines to maintain $K$ reliable ones only. To this end, we design a simple line detector by modifying S-Net in [13]. Given an image, the detector obtains a convolutional feature map and extracts line features by aggregating the features of pixels along each line candidate. Then, it computes the classification probability and regression offset of each candidate. We select the most reliable candidate with the highest probability and remove overlapping lines with the selected one. We iterate this process $K$ times to determine $K$ reliable lines. The default $K$ is 8. Figure 4 shows examples of selected reliable lines. Note that, in this filtering, false negatives occur rarely because the $K$ lines are selected via NMS even though their classification probabilities are not high. The architecture of the modified S-Net is described in detail in the supplemental document. + +Generating line combinations: From the $K$ reliable lines, we generate all $2^{K}$ line combinations. + +# 3.2. Assessing Line Combinations + +When a line combination divides an image insufficiently or over-segments it into unnecessary parts, it does not describe the overall structure of the scene properly. On the contrary, an optimal combination of semantic lines should convey the image composition reliably and efficiently (i.e. with a small number of lines). To find the best line combination, we develop the semantic line combination detector (SLCD). Figure 3 shows the structure of SLCD, which performs encoding, semantic feature grouping, compositional feature ex + +![](images/751fdd3902bfde92d1219577e0419c95b15c8569b7c4f69c0fd16a5f126b813b.jpg) +Figure 4. From a set of line candidates, the line detector selects $K$ reliable lines, depicted in orange. A high recall rate is achieved because a sufficient number of reliable lines are selected through NMS. Ground-truth semantic lines in green are in the insets. + +traction, and score regression. The detailed architecture of SLCD is described in the supplemental document. + +Encoding: Given an image, we extract multi-scale feature maps of ResNet50 [11]. Then, we match their spatial resolutions via bilinear interpolation and concatenate them in the channel dimension. We then squeeze the channels using 2D convolutional layers to obtain an aggregated feature map $F \in \mathbb{R}^{HW \times C}$ , where $H$ , $W$ , and $C$ are the feature height, the feature width, and the number of channels. + +Semantic feature grouping: Semantic lines are located near the boundaries between distinct regions. For their reliable detection, it is desirable to separate different regional parts more clearly. We hence attempt to group pixel features into multiple regions through cross-attention [2, 22, 32, 33]. We employ three cross-attention modules. + +Let $R \in \mathbb{R}^{M \times C}$ be a learnable matrix representing $M$ regions, called region query matrix. Then, we convert $R$ into queries and $F$ into keys and values by + +$$ +R _ {q} = R U _ {q}, F _ {k} = (F + S) U _ {k}, F _ {v} = (F + S) U _ {v}. \tag {1} +$$ + +where $S\in \mathbb{R}^{HW\times C}$ denotes the sinusoidal positional encoding [28], and $U_{q},U_{k},U_{v}\in \mathbb{R}^{C\times C}$ are projection matrices for queries, keys, and values. We then obtain an updated + +![](images/c89af6ab40adfaf0bd9a77941c6e68a8ffdda2acec7bf6f221d88a87872fbed4.jpg) +Figure 5. Visualization of semantic feature grouping results. The top row shows input images with ground-truth lines. The bottom one presents the membership maps, representing the semantic region that each pixel belongs to. + +region query matrix $\hat{R}$ by + +$$ +A = \operatorname {s o f t} _ {M} \max \left(R _ {q} F _ {k} ^ {\mathrm {T}} / \tau\right), \hat {R} = A F _ {v} + R, \tag {2} +$$ + +where $A\in \mathbb{R}^{M\times HW}$ is the attention matrix with a scaling factor $\tau$ and the letter $M$ means the softmax operation is done in the column direction, as in [22, 32, 33]. Then, in the last cross-attention, we generate a semantic feature map $F_{s}\in \mathbb{R}^{HW\times C}$ by + +$$ +F _ {s} = A ^ {\mathrm {T}} \hat {R}. \tag {3} +$$ + +Finally, we channel-wise concatenate $F$ and $F_{s}$ to obtain a combined feature map $X\in \mathbb{R}^{HW\times 2C}$ . + +Note that, in (2), the softmax function is applied along the query axis. Thus, in $A = [\mathbf{p}_1,\mathbf{p}_2,\dots ,\mathbf{p}_{HW}]$ each column $\mathbf{p}_i\in \mathbb{R}^M$ is a probability vector. Specifically, the mth element $p_i^m$ in $\mathbf{p}_i$ is the probability that pixel $i$ belongs to the mth region query. Figure 5 visualizes membership maps, computed by + +$$ +\left[ \operatorname {a r g m a x} \left(\mathbf {p} _ {1}\right), \operatorname {a r g m a x} \left(\mathbf {p} _ {2}\right), \dots , \operatorname {a r g m a x} \left(\mathbf {p} _ {H W}\right) \right] \tag {4} +$$ + +where each element is the index of the region query that the corresponding pixel most likely belongs to. We see that each image is partitioned into $M$ meaningful regions. Thus, $F_{s}$ in (3) represents the regional membership of each pixel and is used to make the feature map $X$ more discriminative. + +To produce the attention matrix $A$ reliably, we design a novel loss function in Section 3.3. The default $M$ is 8. + +Compositional feature extraction: For each line combination, we generate three types of feature maps to extract a compositional feature map. First, we generate a binary mask $B \in \mathbb{R}^{HW}$ for each line combination. Each element in $B$ is 1 if the corresponding pixel belongs to any line in the combination, and 0 otherwise. Then, we decompose the combined feature into a line feature map $X_{l}$ and a region feature map $X_{r}$ by + +$$ +X _ {l} = X \otimes B, X _ {r} = X \otimes (1 - B), \tag {5} +$$ + +where $\otimes$ is the element-wise multiplication. In other words, $X_{l}\in \mathbb{R}^{HW\times 2C}$ contains contextual information for the line pixels only, whereas $X_{r}\in \mathbb{R}^{HW\times 2C}$ does for the pixels strictly inside the regions divided by the lines. + +![](images/f64950abe3173d58c32d60b1fd0f88e3890ccb5499261c4034b08c36494c9e1e.jpg) +Figure 6. Illustration of the line collection map generation when $K = 3$ : (a) reliable lines, (b) line combinations, and (c) line collection maps. + +Moreover, we generate a line collection map $L = [L_1,L_2,\dots ,L_K]\in \mathbb{R}^{HW\times K}$ , where $L_{k}\in \mathbb{R}^{HW}$ is a ternary mask for the $k$ th reliable line, as illustrated in Figure 6. If the $k$ th reliable line does not belong to the line combination, all elements in $L_{k}$ are set to 0. Otherwise, each value in $L_{k}$ is set to either 1 or $-1$ to indicate where the corresponding pixel is located between the two parts divided by the $k$ th reliable line. In other words, $L_{k}$ informs how the $k$ th line splits the image into two regions. We then produce a positional feature map $P\in \mathbb{R}^{HW\times 2C}$ by applying a series of fully connected layers to $L$ . Then, we yield a compositional feature map $Z\in \mathbb{R}^{HW\times 6C}$ by + +$$ +Z = \left[ X _ {l}, X _ {r}, P \right], \tag {6} +$$ + +which contains information about how the lines in the combination separate the image into multiple parts. + +Score regression: Lastly, using a regressor, we estimate a composition score for each line combination. Then, we declare the line combination with the highest score as the optimal group of semantic lines. The regressor takes the compositional feature map $Z$ in (6) as input and predicts a composition score $s$ within [0, 1]. It is implemented using a bilinear interpolation layer and a series of 2D convolution layers and fully connected layers with the ReLU activation. + +# 3.3. Loss Functions + +To train SLCD, we design two loss functions. + +Semantic region separation loss: When two pixels $i$ and $j$ are located in distinct regions, they should be assigned to different region queries. In other words, their probability vectors $\mathbf{p}_i$ and $\mathbf{p}_j$ in the attention matrix $A$ should be far from each other. Let $X$ and $Y$ be the two regions divided by a ground-truth line. Their probability vectors are defined as + +$$ +\mathbf {p} _ {X} = \frac {1}{| X |} \sum_ {i \in X} \mathbf {p} _ {i}, \quad \mathbf {p} _ {Y} = \frac {1}{| Y |} \sum_ {j \in Y} \mathbf {p} _ {j}, \tag {7} +$$ + +which should be also far from each other, for the ground-truth line divides the image into two semantic regions. + +To measure the distance between probability vectors, we adopt the Kullback-Leibler divergence (KLD) [5]. Then, + +![](images/9ab17d575ca7742c214e5f6b6731cb5b27e54609641d4e07740989a0c89fa3e6.jpg) +Horizontal Vertical Diagonal Triangle Symmetric Low Front +Figure 7. The proposed CDL dataset contains 9,100 images in seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. The ground-truth semantic lines are depicted in green. + +we define the semantic region segmentation (SRS) loss, by employing all ground-truth lines, as + +$$ +\mathcal {L} _ {\mathrm {S R S}} = - \sum_ {l = 1} ^ {T} \left[ D \left(\mathbf {p} _ {X _ {l}} \| \mathbf {p} _ {Y _ {l}}\right) + D \left(\mathbf {p} _ {Y _ {l}} \| \mathbf {p} _ {X _ {l}}\right) \right] \tag {8} +$$ + +where $D$ denotes the KLD, $\mathbf{p}_{X_l}$ and $\mathbf{p}_{Y_l}$ are the regions divided by the $l$ th ground-truth line, and $T$ is the number of ground-truth lines. Both $D(\mathbf{p}_{X_l} \parallel \mathbf{p}_{Y_l})$ and $D(\mathbf{p}_{Y_l} \parallel \mathbf{p}_{X_l})$ are computed because $D$ is not commutative. With $\mathcal{L}_{\mathrm{SRS}}$ , it is possible to roughly segment an image into meaningful parts, as illustrated in Figure 5, even though no ground-truth labels for semantic segmentation are used for training. + +Regression loss: We also design a loss for regressing the composition score of each line combination. To this end, we employ the HIoU metric [13] that quantifies the structural layout of a line combination. Let $\mathbf{c}$ and $\mathbf{c}^{\star}$ denote a line combination and the ground-truth combination, respectively. We compute the HIoU between $\mathbf{c}$ and $\mathbf{c}^{\star}$ and use it as the ground-truth composition score $\overline{s}$ of $\mathbf{c}$ . Then, the regression loss is defined as + +$$ +\mathcal {L} _ {\text {r e g}} = (s - \bar {s}) ^ {2} \tag {9} +$$ + +where $s$ is the predicted score of $\mathbf{c}$ . To reduce $\mathcal{L}_{\mathrm{reg}}$ effectively, a ranking loss $\mathcal{L}_{\mathrm{rank}}$ is also employed as in [17]. + +# 4. Experimental Results + +# 4.1. Implementation Details + +We adopt ResNet50 [11] as the encoder of the proposed SLCD. We use the AdamW optimizer [19] with a learning rate $10^{-4}$ , a weight decay of $10^{-4}$ , $\gamma = 0.5$ , $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ . We use a batch size of two for 400,000 iterations. Training images are resized to $480 \times 480$ and augmented by random horizontal flipping. We fix the number of line candidates, reliable lines, and region queries to $N = 1024$ , $K = 8$ , and $M = 8$ , respectively. We set $H = 60$ , $W = 60$ , and $C = 96$ . + +Table 1. Comparison of the HIoU scores $(\%)$ on the SEL, SEL_Hard, NKL, and CDL datasets. + +
SELSEL_HardNKLCDL
SLNet [15]77.8759.7165.4957.78
DHT [9]79.6263.3969.0863.24
DRM [12]80.2368.8367.4263.96
HSLD [13]81.0365.9974.2964.98
Proposed84.0968.1576.2168.85
+ +# 4.2. Datasets + +SEL [15]: It is the first dataset for semantic line detection. It contains 1,750 images, which are split into 1,575 training and 175 test images. Each semantic line is annotated by the coordinates of two endpoints on an image boundary. In SEL, most images are high-quality landscapes, in which semantic lines are often obvious. + +SEL_Hard [12]: It contains 300 test images only, selected from the ADE20K segmentation dataset [37]. It is challenging because of cluttered scenes or occluded objects. + +NKL [35]: It is a relatively large dataset of 5,200 training and 1,300 testing images. It includes both indoor and outdoor scenes. + +CDL: We construct CDL to contain 7,100 scenes with diverse contents and compositions. It is split into 6,390 training and 710 test images. As in Figure 7, the images are categorized into seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. In Low and Front, humans and animals are essential parts of the image composition. In the other classes, most images are outdoor ones, as in the other datasets. To construct CDL, semantic lines were manually annotated in about 200 man-hours. More examples and the annotation process are provided in the supplemental document. + +# 4.3. Metrics + +As mentioned earlier, the harmony of detected lines is more important than individual lines in semantic line detection. Therefore, the main paper discusses only HIoU performances. The precision, recall, and f-measure results, which are metrics used in previous work, are compared in the supplemental document. + +HIoU: An optimal group of semantic lines conveys a harmonious impression about the composition of an image. To assess the overall harmony of detection results, we use the HIoU metric [13]. It measures the consistency between the division of an image by detected lines and that by the ground truth. Let $S = \{s_1,s_2,\dots ,s_N\}$ and $T = \{t_{1},t_{2},\ldots ,t_{M}\}$ be the regions divided by the detected lines and the ground-truth lines, respectively. Then, HIoU is computed as + +$$ +\mathrm {H I o U} = \frac {\sum_ {i = 1} ^ {N} \max _ {k} \operatorname {I o U} \left(s _ {i} , t _ {k}\right) + \sum_ {j = 1} ^ {M} \max _ {k} \operatorname {I o U} \left(t _ {j} , s _ {k}\right)}{N + M}. \tag {10} +$$ + +![](images/7f40557f8557bd0bac1cbdba040bbc06303e37f7b735fa096631ecd5a814228d.jpg) +Figure 8. Comparison of detected semantic lines. From the left, two images are selected from each of the SEL, SEL_Hard, NKL, and CDL datasets. + +Specifically, for each $s_i$ , the most overlapping region $t_k$ is chosen and the intersection-over-union (IoU) is calculated. This is performed similarly for each $t_j$ . Then, the HIoU score is given by the average of these bi-directionally matching IoUs. + +# 4.4. Comparative Assessment + +Table 1 compares the HIoU scores of the proposed SLCD with those of the existing detectors [9, 12, 13, 15] on the SEL, SEL_Hard, NKL, and CDL datasets. Also, Figure 8 compares detection results qualitatively. The existing detectors miss correct lines or fail to remove redundant lines, yielding sub-optimal results. In contrast, SLCD detects semantic lines more precisely and represents the composition more reliably than the existing detectors do. More detection results are available in the supplemental document. + +Comparison on SEL: In Table 1, SLCD outperforms the existing detectors on SEL. Compared with the second-best HSLD, SLCD yields a wide HIoU margin of 3.06. This indicates that SLCD finds an optimal group of semantic lines more effectively by processing all lines in each combination simultaneously, instead of performing pairwise comparisons in HSLD. + +Comparison on SEL_Hard: As done in [12], we conduct experiments on SEL_Hard using the networks trained on the SEL dataset. In Table 1, SLCD ranks 2nd on SEL_Hard. DRM provides a better result than SLCD, but it demands + +Table 2. Comparison of the HIoU scores $(\%)$ on the CDL dataset according to the number $M$ of region queries. + +
M481632
H IoU66.4368.8568.1167.07
+ +![](images/5cad2bf36d941e55e506c89673e99df94b5231413db42ebfcacd7e0bb5a9c9ee.jpg) +Figure 9. Visualization of the membership maps according to $M$ . + +much higher complexity, as will be discussed in Section 4.5. + +Comparison on NKL: SLCD surpasses all existing detectors. For example, its HIoU score is 1.92 points higher than the second-best HSLD. + +Comparison on CDL: Table 1 also lists HIoU scores on the proposed CDL dataset. For a fair comparison, we train the existing detectors on CDL using their publicly available source codes. We see that SLCD outperforms all existing detectors on CDL as well. SLNet, DHT, and DRM yield poor results because they do not consider the overall harmony of detected lines. HSLD is better than these detectors but is inferior to the proposed SLCD. We compare and + +Table 3. Comparison of the HIoU scores $(\%)$ on the CDL dataset according to the usage of the SRS loss $\mathcal{L}_{\mathrm{SRS}}$ + +
w/o LSRProposed
HIoU66.7168.85
+ +Table 4. Ablation studies for the compositional feature extraction on the CDL dataset. + +
LineRegionPositionalHIoU
I✓65.12
II✓65.74
III✓✓66.37
IV✓✓✓68.85
+ +discuss the detection results according to the composition classes in the supplemental document. + +# 4.5. Analysis + +We conduct several ablation studies to analyze each component of the proposed SLCD. + +The number of $M$ : Table 2 lists the HIoU performances on the CDL dataset, according to the number $M$ of region queries. At the default $M = 8$ , SLCD achieves the best performance. When $M$ is smaller or bigger than 8, the performance degrades due to under- or over-segmentation of semantic parts, respectively, as shown in Figure 9. + +Efficacy of $\mathcal{L}_{\mathrm{SRS}}$ : In Table 3, if the proposed SRS loss $\mathcal{L}_{\mathrm{SRS}}$ in (8) is excluded from the training, the performance drops by 2.14 points. It means that the composition analysis based on the SRS loss is essential for finding optimal line combinations. + +Efficacy of compositional feature extraction: SLCD extracts the compositional feature map $Z$ for each line combination, by processing the line feature map $X_{l}$ , the region feature map $X_{r}$ , and the positional feature map $P$ in (6). In Table 4, Method I uses the line feature map only, while II uses the region feature map only. Method III utilizes both feature maps. Method IV, which is the proposed SLCD, uses all three maps. + +Method I yields the worst result, for it uses only the contextual information near line pixels. Method II is slightly better than Method I, by exploiting the regional information. Using both line and region feature maps, III provides a better result. Moreover, IV further improves the performance significantly by utilizing the positional feature map. This is because the overall harmony is estimated more effectively, by combining the line structures in the positional feature map with scene contexts. + +Runtime: Table 5 compares the runtimes of SLCD and the existing detectors in seconds per frame (spf). We use a PC with AMD Ryzen 9 3900X CPU and NVIDIA RTX 2080 GPU. The proposed SLCD takes 0.114spf, adding + +Table 5. Runtime comparison of the proposed SLCD and the existing detectors. The processing times are reported in seconds per frame (spf). + +
SLNet [15]DHT [9]DRM [12]HSLD [13]Proposed
0.1360.0330.9520.0460.114
+ +Table 6. Comparison of the AA scores $(\%)$ of VP detection. + +
AA1°AA2°AA10°
Zhou et al. [39]18.533.060.0
Jin et al. [12]8.622.968.3
Proposed16.636.978.3
+ +![](images/959a4c3b38ec86f392b4e02c6b4a0a280d95ae345be6a7250cb0e7239250a6b3.jpg) + +![](images/383589e349806de2088fb487d72d2f1b3f5e24dc5e8293ef75eece014f9bfacf.jpg) + +![](images/b1ce6b0a7510c7a901fe5856c739fb376930ce7d418d30d8a4d2810cc2264807.jpg) +Figure 10. Detected dominant parallel lines and their intersections are depicted by yellow lines and green triangles, respectively. The ground-truth VPs are depicted by red dots. + +![](images/4b50d7c4b343f89aa571b578a144ba3922851fd495db9c62dc98175268075b18.jpg) + +![](images/eade586151d69ba3145ba6a22b7a10d581dc3537733da735cd6b6994a28a0203.jpg) + +![](images/6f2b0c4b156000e45d831c8818791939272e4089ac14af15dab5ceffdeb9d111.jpg) + +![](images/130d977ad799b84a0060fcba31190943ee9501631bf296dee4a57b2a57bfa2ad.jpg) + +![](images/78a0814479b3248024a17fe086e70206428606b9f9138c8acc8dc90eca298c9a.jpg) + +![](images/b4a840a90cecb107679c05a08857c0d7eade2faab03adbb0e052c821ab165c83.jpg) + +up 0.030spf and 0.074spf for generating and assessing line combinations, respectively. DHT is the fastest detector, but it is inferior to SLCD on all datasets. On the other hand, even though DRM performs better on SEL_Hard, it is about 8.4 times slower than SLCD. + +# 5. Applications + +We apply the proposed SLCD to three vision tasks: dominant vanishing point detection, reflection symmetry axis detection, and composition-based image retrieval. Note that the first two tasks were considered in [12]. Due to the page limit, more details and results are described in the supplemental document. + +# 5.1. Dominant Vanishing Point Detection + +A vanishing point (VP) facilitates understanding of the 3D geometric structure. We apply the proposed SLCD to detect dominant VPs, by identifying vanishing lines. We use the AVA landscape dataset [39], in which two dominant parallel lines are annotated for each image. To detect a dominant VP, we generate line combinations containing two lines only. We then find the best combination, whose intersecting point is declared as a VP. Table 6 compares this VP detection scheme with the existing line-based VP detectors in [12, 39]. The angle accuracies (AAs) [38] are compared. We see that SLCD is better than the existing detectors, except that its $\mathrm{AA1}^{\circ}$ score is lower than that of Zhou et al. [39]. Figure 10 shows some VP detection results. + +Table 7. Comparison of the AUC_A scores $(\%)$ of symmetry axis detection. + +
ICCVNYUSYM_Hard
Cicconet et al. [3]80.8082.8568.99
Elawady et al. [6]87.2483.8373.90
Cicconet et al. [4]87.3887.6481.04
Loy & Eklundh [20]89.7790.8581.99
Jin et al. [12]90.6092.7884.73
Proposed93.1593.5188.03
+ +![](images/b2c64ee54b58ceab391233324a1453dd9c9ffb0f86dd748a3ac2736c7eb081d4.jpg) +Figure 11. Symmetry axis detection results. In the top row, the ground-truth and predicted axes are depicted by dashed red and solid yellow lines, respectively. The membership maps are also visualized in the bottom row. + +# 5.2. Reflection Symmetry Axis Detection + +Reflection symmetry is a visual cue for analyzing object shapes and patterns. Its detection is challenging because the symmetry is only implicit in many cases. We test the proposed SLCD on three datasets: ICCV [7], NYU [4], and SYM_Hard [12]. In these datasets, each image contains a single reflection symmetry axis. Thus, in the proposed SLCD, each of the $K$ reliable lines is regarded as a line combination. Table 7 compares the AUC_A scores [15] of the proposed SLCD and the conventional techniques [3, 4, 6, 12, 20]. SLCD outperforms all these techniques on all datasets. Figure 11 shows some detection results together with the membership maps. We see that regional parts are symmetrically divided along implied axes, enabling SLCD to identify those axes effectively. + +# 5.3. Composition-Based Image Retrieval + +Existing image retrieval techniques focus on the visual contents of a query image to find similar images in a database [8, 21, 24, 25]. In this work, however, we attempt to discover images with similar compositions to a query image, i.e. composition-based image retrieval. + +We test the proposed SLCD on the Oxford 5k and Paris 6k retrieval dataset [23]. We first detect semantic lines in every image, while storing the positional feature map $P$ in (6). We filter out some images whose composition scores are lower than a threshold since a low score indicates a low-quality image with inharmonious composition in general. Then, for a randomly selected query image, we compute the $\ell_2$ -distances between the positional feature maps $P$ of the query and the remaining images. We determine the images with the smallest distances as retrieval results. + +Figure 12 shows the top-4 retrieval results for four query images. We see that SLCD returns structurally similar im + +![](images/576dbbd33363f999f807e75a277b54742f5fe14c30814966efee0663584381cd.jpg) +Query image + +![](images/dc92edc03a8701496e64266227c326dfb56783c1f8cc4e0fff47ad3a22fc319a.jpg) +Top-4 retrieval results + +![](images/ae0a8aa009561a0ca6d71390ed852a25cbe9500eaf2042fd3f89e7d85b2315b2.jpg) +Figure 12. Composition-based retrieval results for query images. +Figure 13. t-SNE visualization [27] of the feature space for the Oxford 5k and Paris 6k dataset [23]. + +ages to the queries. This means that the positional feature maps $P$ , containing the structural information of semantic lines, can be used to compute the compositional differences. Based on this observation, we perform data clustering by employing k-means [10] on the feature space of $P$ . Figure 13 is t-SNE visualization [27] of the clustering results. The nine images nearest to each centroid are also shown. Note that images with similar composition are grouped into the same cluster, confirming the efficacy of SLCD in the compositional analysis of images. + +# 6. Conclusions + +We proposed a novel semantic line detector, SLCD, which processes a combination of lines at once to estimate the overall harmony reliably. We first generated all possible line combinations from reliable lines. Then, we estimated the score of each line combination and determined the best combination. Experimental results demonstrated that the proposed SLCD can detect semantic lines reliably on existing datasets, as well as on the new dataset CDL. Furthermore, SLCD can be successfully used in vanishing point detection, symmetry axis detection, and image retrieval. + +# Acknowledgements + +This work was conducted by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD) and was supported by the National Research Foundation of Korea (NRF) grants funded by the Korea government (MSIT) (No. NRF-2021R1A4A1031864 and No. NRF-2022R1A2B5B03002310). + +# References + +[1] Jean-Charles Bazin, Yongduek Seo, Cedric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in Manhattan world. In Proc. IEEE CVPR, 2012. 1, 2 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 3 +[3] Marcelo Cicconet, Vighnesh Birodkar, Mads Lund, Michael Werman, and Davi Geiger. A convolutional approach to reflection symmetry. Pattern Recog. Lett., 95:44-50, 2017. 1, 2, 8 +[4] Marcelo Cicconet, David GC Hildebrand, and Hunter Elliott. Finding mirror symmetry via registration and optimal symmetric pairwise assignment of curves: Algorithm and results. In Proc. IEEE ICCV Workshops, 2017. 1, 8 +[5] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley, 2nd edition, 2006. 4 +[6] Mohamed Elawady, Christophe Ducottet, Olivier Alata, Cécile Barat, and Philippe Colantoni. Wavelet-based reflection symmetry detection via textural and color histograms. In Proc. IEEE ICCV Workshops, 2017. 1, 2, 8 +[7] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 ICCV Challenge: Detecting symmetry in the wild. In Proc. IEEE ICCV, 2017. 1, 2, 8 +[8] Albert Gordo, Jon Almazan, Jerome Revaud, and Diane Larlus. End-to-end learning of deep visual representations for image retrieval. Int. J. Comput. Vis., 124:237-254, 2017. 8 +[9] Qi Han, Kai Zhao, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. In Proc. ECCV, 2020. 1, 2, 3, 5, 6, 7 +[10] J. A. Hartigan and M. A. Wong. A K-means clustering algorithm. Journal of the Royal Statistical Society. Series C (Applied Statistics), 28(1):100-108, 1979. 8 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, 2016. 3, 5 +[12] Dongkwon Jin, Jun-Tae Lee, and Chang-Su Kim. Semantic line detection using mirror attention and comparative ranking and matching. In Proc. ECCV, 2020. 1, 2, 5, 6, 7, 8 +[13] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, and Chang-Su Kim. Harmonious semantic line detection via maximal weight clique selection. In Proc. IEEE CVPR, 2021. 1, 2, 3, 5, 6, 7 +[14] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Temporally consistent horizon lines. In Proc. IEEE ICRA, 2020. 1, 2 +[15] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Semantic line detection and its applications. In Proc. IEEE ICCV, 2017. 1, 2, 5, 6, 7, 8 +[16] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Photographic composition classification and dominant geometric element detection for outdoor scenes. J. Vis. Commun. Image Represent., 55:91-105, 2018. 2 + +[17] Debang Li, Junge Zhang, Kaiqi Huang, and Ming-Hsuan Yang. Composing good shots by exploiting mutual relations. In Proc. IEEE CVPR, 2020. 5 +[18] Ligang Liu, Renjie Chen, Lior Wolf, and Daniel Cohen-Or. Optimizing photo composition. Computer Graphics Forum, 29:469-478, 2010. 2 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[20] Gareth Loy and Jan-Olof Eklundh. Detecting symmetry and symmetric constellations of features. In Proc. ECCV, 2006. 1, 2, 8 +[21] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proc. IEEE ICCV, 2017. 8 +[22] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. iDisc: Internal discretization for monocular depth estimation. In Proc. IEEE CVPR, 2023. 3, 4 +[23] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proc. IEEE CVPR, 2018. 8 +[24] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell., 41:1655-1668, 2018. 8 +[25] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. IEEE Trans. Pattern Anal. Mach. Intell., 22:1349-1380, 2000. 8 +[26] Lucas Tabelini, Rodrigo Berriel, Thiago M. Paixao, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE CVPR, 2021. 1, 2 +[27] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8 +[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 3 +[29] Yaoting Wang, Yongzhen Ke, Kai Wang, Jing Guo, and Shuai Yang. Spatial-invariant convolutional neural network for photographic composition prediction and automatic correction. J. Vis. Commun. Image Represent., 90:103751, 2023. 1, 2 +[30] Scott Workman, Menghua Zhai, and Nathan Jacobs. Horizon lines in the wild. In Proc. BMVC, 2016. 1, 2 +[31] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proc. IEEE ICCV, 2015. 2 +[32] Qihang Yu, Huiyu Wang, Dahun Kim, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. CMT-DeepLab: Clustering mask transformers for panoptic segmentation. In Proc. IEEE CVPR, 2022. 3, 4 +[33] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. $k$ -means mask transformer. In Proc. ECCV, 2022. 3, 4 + +[34] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proc. IEEE CVPR, 2016. 1, 2 +[35] Kai Zhao, Qi Han, Chang-Bin Zhang, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. IEEE Trans. Pattern Anal. Mach. Intell., 44:4793-4806, 2021. 1, 5 +[36] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. CLRNet: Cross layer refinement network for lane detection. In Proc. IEEE CVPR, 2022. 1, 2 +[37] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In Proc. IEEE CVPR, 2017. 5 +[38] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. NeurVPS: Neural vanishing point scanning via conic convolution. In Proc. NeurIPS, 2019. 7 +[39] Zihan Zhou, Farshid Farhat, and James Z Wang. Detecting dominant vanishing points in natural scenes with application to composition-sensitive image retrieval. IEEE Trans. Multimedia, 19:2651-2665, 2017. 1, 2, 7 +[40] Rui Zhu, Xingyi Yang, Yannick Hold-Geoffroy, Federico Perazzi, Jonathan Eisenmann, Kalyan Sunkavalli, and Manmohan Chandraker. Single view metrology in the wild. In Proc. ECCV, 2020. 1, 2 \ No newline at end of file diff --git a/2024/Semantic Line Combination Detector/images.zip b/2024/Semantic Line Combination Detector/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da3c037eadc7184005e19c35a5ac09079797097e --- /dev/null +++ b/2024/Semantic Line Combination Detector/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9fafded18f8e1cf2af8c08eeec8cd81206a4c872f5b8bbe58cf19dd11fc0883 +size 735317 diff --git a/2024/Semantic Line Combination Detector/layout.json b/2024/Semantic Line Combination Detector/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..49bb67f7952997a8b8b3303cdbb65f02170a7001 --- /dev/null +++ b/2024/Semantic Line Combination Detector/layout.json @@ -0,0 +1,10814 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 191, + 129, + 403, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 129, + 403, + 143 + ], + "spans": [ + { + "bbox": [ + 191, + 129, + 403, + 143 + ], + "type": "text", + "content": "Semantic Line Combination Detector" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 165, + 175, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 165, + 175, + 177 + ], + "spans": [ + { + "bbox": [ + 126, + 165, + 175, + 177 + ], + "type": "text", + "content": "Jinwon Ko" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 178, + 189, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 178, + 189, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 178, + 189, + 190 + ], + "type": "text", + "content": "Korea University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 193, + 200, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 193, + 200, + 202 + ], + "spans": [ + { + "bbox": [ + 100, + 193, + 200, + 202 + ], + "type": "text", + "content": "jwko@mcl.korea.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 256, + 165, + 325, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 165, + 325, + 177 + ], + "spans": [ + { + "bbox": [ + 256, + 165, + 325, + 177 + ], + "type": "text", + "content": "Dongkwon Jin" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 178, + 331, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 178, + 331, + 190 + ], + "spans": [ + { + "bbox": [ + 255, + 178, + 331, + 190 + ], + "type": "text", + "content": "Korea University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 193, + 359, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 193, + 359, + 202 + ], + "spans": [ + { + "bbox": [ + 225, + 193, + 359, + 202 + ], + "type": "text", + "content": "dongkwonjin@mcl.korea.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 404, + 165, + 477, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 165, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 404, + 165, + 477, + 178 + ], + "type": "text", + "content": "Chang-Su Kim*" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 402, + 178, + 477, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 178, + 477, + 190 + ], + "spans": [ + { + "bbox": [ + 402, + 178, + 477, + 190 + ], + "type": "text", + "content": "Korea University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 384, + 193, + 494, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 193, + 494, + 202 + ], + "spans": [ + { + "bbox": [ + 384, + 193, + 494, + 202 + ], + "type": "text", + "content": "changsukim@korea.ac.kr" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 68, + 251, + 290, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 251, + 290, + 393 + ], + "spans": [ + { + "bbox": [ + 68, + 251, + 290, + 393 + ], + "type": "text", + "content": "A novel algorithm, called semantic line combination detector (SLCD), to find an optimal combination of semantic lines is proposed in this paper. It processes all lines in each line combination at once to assess the overall harmony of the lines. First, we generate various line combinations from reliable lines. Second, we estimate the score of each line combination and determine the best one. Experimental results demonstrate that the proposed SLCD outperforms existing semantic line detectors on various datasets. Moreover, it is shown that SLCD can be applied effectively to three vision tasks of vanishing point detection, symmetry axis detection, and composition-based image retrieval. Our codes are available at https://github.com/Jinwon-Ko/SLCD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 414, + 143, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 414, + 143, + 425 + ], + "spans": [ + { + "bbox": [ + 69, + 414, + 143, + 425 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 68, + 430, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 430, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 68, + 430, + 288, + 517 + ], + "type": "text", + "content": "A semantic line [15] is defined as a meaningful line separating distinct semantic regions in an image. Besides this unary definition, multiple semantic lines in an image are supposed to convey the global scene structure properly [13]. It is challenging to detect such semantic lines because they are often implied by complicated region boundaries. Moreover, they should represent the image composition optimally by dividing it into semantic regions harmoniously." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 68, + 517, + 288, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 517, + 288, + 637 + ], + "spans": [ + { + "bbox": [ + 68, + 517, + 288, + 637 + ], + "type": "text", + "content": "Semantic lines are essential elements in many vision applications. For example, a horizon line [14, 30, 34, 40], which is a specific type of semantic line, can be exploited to adjust the levelness of an image [15, 29]. A reflection symmetry axis [3, 4, 6, 7, 20], which is another type of semantic line, provides visual cues for object recognition and pattern analysis. Vanishing points, conveying depth impression in images, can be estimated by detecting dominant parallel semantic lines in the 3D world [1, 12, 39]. In autonomous driving systems [13, 26, 36], boundaries of road lanes can be also described by semantic lines." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 68, + 638, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 638, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 68, + 638, + 288, + 670 + ], + "type": "text", + "content": "Recently, several attempts [9, 12, 13, 15, 35] have been made to detect semantic lines. These techniques perform line detection and refinement sequentially. At the detection" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 228, + 351, + 262 + ], + "blocks": [ + { + "bbox": [ + 309, + 228, + 351, + 262 + ], + "lines": [ + { + "bbox": [ + 309, + 228, + 351, + 262 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 351, + 262 + ], + "type": "image", + "image_path": "0eb69331fb07e6bf0443e794367a4b0c94d7a7b8b7cf946f443de0f04bfb0785.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 262, + 335, + 268 + ], + "lines": [ + { + "bbox": [ + 324, + 262, + 335, + 268 + ], + "spans": [ + { + "bbox": [ + 324, + 262, + 335, + 268 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 365, + 230, + 439, + 276 + ], + "blocks": [ + { + "bbox": [ + 365, + 230, + 439, + 276 + ], + "lines": [ + { + "bbox": [ + 365, + 230, + 439, + 276 + ], + "spans": [ + { + "bbox": [ + 365, + 230, + 439, + 276 + ], + "type": "image", + "image_path": "976615b3f9f48847ab871ec52e7071aa26f8150d32784f5c324ea2d15d079bd5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 277, + 413, + 284 + ], + "lines": [ + { + "bbox": [ + 392, + 277, + 413, + 284 + ], + "spans": [ + { + "bbox": [ + 392, + 277, + 413, + 284 + ], + "type": "text", + "content": "(a) NMS" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 447, + 231, + 521, + 276 + ], + "blocks": [ + { + "bbox": [ + 447, + 231, + 521, + 276 + ], + "lines": [ + { + "bbox": [ + 447, + 231, + 521, + 276 + ], + "spans": [ + { + "bbox": [ + 447, + 231, + 521, + 276 + ], + "type": "image", + "image_path": "a1cfc8e91f1857e4e81e2238a373ea847ce93b930b4a03ce18bf749fdc5b0011.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 277, + 509, + 284 + ], + "lines": [ + { + "bbox": [ + 459, + 277, + 509, + 284 + ], + "spans": [ + { + "bbox": [ + 459, + 277, + 509, + 284 + ], + "type": "text", + "content": "(b) Pairwise correlation" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 309, + 285, + 351, + 342 + ], + "blocks": [ + { + "bbox": [ + 309, + 285, + 351, + 342 + ], + "lines": [ + { + "bbox": [ + 309, + 285, + 351, + 342 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 351, + 342 + ], + "type": "image", + "image_path": "9cb3c69de85b36a772ea0bb6226738a4c810e2f30914daf1c934880e5514946c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 342, + 345, + 347 + ], + "lines": [ + { + "bbox": [ + 316, + 342, + 345, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 345, + 347 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 356, + 526, + 477 + ], + "lines": [ + { + "bbox": [ + 304, + 356, + 526, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 526, + 477 + ], + "type": "text", + "content": "Figure 1. After selecting reliable line candidates, there are two existing approaches to semantic line detection. The first approach in (a) focuses on locating a line near a region boundary and eliminating overlapping lines. However, a redundant line still remains, since this approach does not consider how well a group of detected lines represents the layout of a scene. The second approach in (b) takes into account only the pairwise correlation between two lines, so it may fail to assess the overall harmony of more than two semantic lines. In contrast, in (c), the proposed SLCD generates a number of line combinations, analyzes all lines in each combination at once, and then finds the most harmonious combination that conveys the global scene composition optimally." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 365, + 285, + 453, + 339 + ], + "blocks": [ + { + "bbox": [ + 365, + 285, + 453, + 339 + ], + "lines": [ + { + "bbox": [ + 365, + 285, + 453, + 339 + ], + "spans": [ + { + "bbox": [ + 365, + 285, + 453, + 339 + ], + "type": "image", + "image_path": "e31170d5cdeacfff1e62c5bff86153b2e160a3b78829f5183884e485a47e05ee.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 341, + 465, + 348 + ], + "lines": [ + { + "bbox": [ + 421, + 341, + 465, + 348 + ], + "spans": [ + { + "bbox": [ + 421, + 341, + 465, + 348 + ], + "type": "text", + "content": "(c) Proposed SLCD" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 455, + 285, + 518, + 339 + ], + "blocks": [ + { + "bbox": [ + 455, + 285, + 518, + 339 + ], + "lines": [ + { + "bbox": [ + 455, + 285, + 518, + 339 + ], + "spans": [ + { + "bbox": [ + 455, + 285, + 518, + 339 + ], + "type": "image", + "image_path": "56e41afe1fb1bdb8517a6f2c9bfffeffab508d297e7df904cfe1ddef5aedad07.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 500, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 526, + 687 + ], + "type": "text", + "content": "stage, they extract deep line features to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing redundant lines. Specifically, to refine line candidates, non-maximum suppression (NMS) is performed in [9, 12, 15], as illustrated in Figure 1(a). Lee et al. [15] iteratively select the reliable line near boundary pixels and remove overlapping lines with the selected one. Han et al. [9] simplify the NMS process by adopting a Hough line space. Jin et al. [12] process each candidate through comparative ranking and matching. These techniques, however, do not consider the overall harmony among detected lines. To cope with this issue, Jin et al. [13] estimate the relation score for every pair of detected lines and then decide harmonious semantic lines via graph optimization. But, they may yield sub-optimal results, since only the pairwise relationships between lines are exploited, as in Figure 1(b)." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 676, + 150, + 685 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 676, + 150, + 685 + ], + "spans": [ + { + "bbox": [ + 79, + 676, + 150, + 685 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "text", + "content": "28066" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 100, + 523, + 178 + ], + "blocks": [ + { + "bbox": [ + 72, + 100, + 523, + 178 + ], + "lines": [ + { + "bbox": [ + 72, + 100, + 523, + 178 + ], + "spans": [ + { + "bbox": [ + 72, + 100, + 523, + 178 + ], + "type": "image", + "image_path": "cf3196727fb63a2bdfe55217b71bf63d4b9a90129c67f9cd4d02768c30b054b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 186, + 386, + 196 + ], + "lines": [ + { + "bbox": [ + 206, + 186, + 386, + 196 + ], + "spans": [ + { + "bbox": [ + 206, + 186, + 386, + 196 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed SLCD algorithm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 213, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 213, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 289, + 410 + ], + "type": "text", + "content": "In this paper, we propose a novel algorithm, called semantic line combination detector (SLCD), to find an optimal group of semantic lines. It processes all lines in a line group (or combination) simultaneously, instead of analyzing each pair of lines, to estimate the overall harmony, as in Figure 1(c). Figure 2 shows an overview of the proposed SLCD. First, we select reliable lines from line candidates and then generate a number of line combinations. Second, we score all line combinations and determine the combination with the highest score as the optimal group of semantic lines. To this end, we design two novel modules for semantic feature grouping and compositional feature extraction. We also introduce a novel loss function to guide the feature grouping module. Experimental results demonstrate that SLCD can detect semantic lines reliably on existing datasets and a new dataset, called compositionally diverse lines (CDL). Moreover, it is shown that SLCD can be used effectively in various applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 411, + 262, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 411, + 262, + 422 + ], + "spans": [ + { + "bbox": [ + 80, + 411, + 262, + 422 + ], + "type": "text", + "content": "This work has the following major contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 422, + 289, + 521 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 69, + 422, + 288, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 422, + 288, + 444 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 288, + 444 + ], + "type": "text", + "content": "- SLCD finds an optimal combination of semantic lines by processing all lines in a line combination at once." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 444, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 444, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 69, + 444, + 289, + 475 + ], + "type": "text", + "content": "- We construct the CDL dataset containing compositionally diverse images with implied lines. It will be made publicly available.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 476, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 476, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 69, + 476, + 289, + 521 + ], + "type": "text", + "content": "- SLCD outperforms conventional detectors on most datasets. Also, its effectiveness is demonstrated in three applications: vanishing point detection, symmetry axis detection, and composition-based image retrieval." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 529, + 149, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 529, + 149, + 540 + ], + "spans": [ + { + "bbox": [ + 69, + 529, + 149, + 540 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 548, + 289, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 548, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 67, + 548, + 289, + 668 + ], + "type": "text", + "content": "Semantic lines, located near the boundaries of different semantic regions, outline the layout and composition of an image. They play an important role in various vision applications. Horizons [14, 30, 34, 40] are a specific type of semantic lines, which can be applied to adjust the levelness of images and improve their aesthetics [15, 29]. In [1, 12, 39], dominant parallel lines in the 3D world are detected to estimate vanishing points, which convey depth impression on 2D images. Also, in [3, 6, 7, 20], reflection symmetry axes are identified to analyze the shapes of objects or patterns. These types of lines can be regarded as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 525, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 525, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 525, + 268 + ], + "type": "text", + "content": "highly implied semantic lines. In [13, 26, 36], straight lanes are detected to aid in vehicle maneuvers in road environments. Furthermore, semantic lines are essential visual cues in photographic composition [16, 18]. They direct viewers' attention and help to compose a visually balanced image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 268, + 526, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 526, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 526, + 586 + ], + "type": "text", + "content": "Several semantic line detectors [9, 12, 13, 15] have been proposed. They perform in two stages: line detection and refinement. At the detection stage, deep line features are extracted to classify and regress each line candidate. At the refinement stage, reliable semantic lines are determined by removing irrelevant candidates, based on NMS [9, 12, 15] or graph optimization [13]. More specifically, Lee et al. [15] detect reliable lines with classification probabilities higher than a threshold. They then iteratively select semantic lines and remove overlapping lines with the selected one, by employing the edge detector in [31]. Han et al. [9] predict the probability of each candidate in a Hough parametric space. They simplify NMS by computing the centroids of connected components in the Hough space. Jin et al. [12] design two comparators to estimate the priority and similarity between two lines. Then, they determine the most reliable lines and eliminate redundant ones alternately through pairwise comparisons. However, these techniques [9, 12, 15] do not consider how well a group of detected lines represents the global scene structure. To address this issue, Jin et al. [13] analyze the pairwise harmony of detected lines. They first estimate a harmony score for a pair of detected lines. They then construct a complete graph and determine harmonious semantic lines by finding a maximal weight clique. Their method, however, may yield sub-optimal results because it considers only pairwise relationships between lines. On the contrary, the proposed SLCD finds an optimal combination of semantic lines by analyzing all lines in each combination simultaneously." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 591, + 416, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 591, + 416, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 591, + 416, + 604 + ], + "type": "text", + "content": "3. Proposed Algorithm" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 609, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 526, + 687 + ], + "type": "text", + "content": "We propose a novel algorithm, called SLCD, to detect an optimal combination of semantic lines, an overview of which is in Figure 2. First, we select " + }, + { + "bbox": [ + 304, + 609, + 526, + 687 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 609, + 526, + 687 + ], + "type": "text", + "content": " reliable lines from line candidates and then generate a number of line combinations. Second, we score all the line combinations and determine the combination with the highest score as the optimal group of semantic lines." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 676, + 253, + 685 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 676, + 253, + 685 + ], + "spans": [ + { + "bbox": [ + 79, + 676, + 253, + 685 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 676, + 253, + 685 + ], + "type": "text", + "content": "CDL is available at https://github.com/Jinwon-Ko/SLCD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "text", + "content": "28067" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 99, + 523, + 219 + ], + "blocks": [ + { + "bbox": [ + 72, + 99, + 523, + 219 + ], + "lines": [ + { + "bbox": [ + 72, + 99, + 523, + 219 + ], + "spans": [ + { + "bbox": [ + 72, + 99, + 523, + 219 + ], + "type": "image", + "image_path": "9b3a72ca4e2fa6e9e65d80f0a5a528ff6a36907331dfac4b96e86c5b476770b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 225, + 525, + 247 + ], + "lines": [ + { + "bbox": [ + 68, + 225, + 525, + 247 + ], + "spans": [ + { + "bbox": [ + 68, + 225, + 525, + 247 + ], + "type": "text", + "content": "Figure 3. The architecture of SLCD. CA and FC denote cross-attention and fully connected layers, respectively. The blue, red, yellow, and green boxes indicate encoding, semantic feature grouping, compositional feature extraction, and score regression, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 265, + 223, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 265, + 223, + 276 + ], + "spans": [ + { + "bbox": [ + 69, + 265, + 223, + 276 + ], + "type": "text", + "content": "3.1. Generating Line Combinations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 282, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 282, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 68, + 282, + 288, + 337 + ], + "type": "text", + "content": "Initializing line candidates: We generate line candidates, which are end-to-end straight lines in an image. Each line candidate is parameterized by polar coordinates in the Hough space [9]. By quantizing the coordinates uniformly, we obtain " + }, + { + "bbox": [ + 68, + 282, + 288, + 337 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 68, + 282, + 288, + 337 + ], + "type": "text", + "content": " line candidates. The default " + }, + { + "bbox": [ + 68, + 282, + 288, + 337 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 68, + 282, + 288, + 337 + ], + "type": "text", + "content": " is 1024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "spans": [ + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": "Filtering line candidates: Given the " + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": " line candidates, there are " + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "inline_equation", + "content": "2^{N}" + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": " possible line combinations in total. Since this number of combinations is unmanageable, we filter out irrelevant lines to maintain " + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": " reliable ones only. To this end, we design a simple line detector by modifying S-Net in [13]. Given an image, the detector obtains a convolutional feature map and extracts line features by aggregating the features of pixels along each line candidate. Then, it computes the classification probability and regression offset of each candidate. We select the most reliable candidate with the highest probability and remove overlapping lines with the selected one. We iterate this process " + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": " times to determine " + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": " reliable lines. The default " + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": " is 8. Figure 4 shows examples of selected reliable lines. Note that, in this filtering, false negatives occur rarely because the " + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 340, + 290, + 537 + ], + "type": "text", + "content": " lines are selected via NMS even though their classification probabilities are not high. The architecture of the modified S-Net is described in detail in the supplemental document." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 540, + 288, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 540, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 68, + 540, + 288, + 562 + ], + "type": "text", + "content": "Generating line combinations: From the " + }, + { + "bbox": [ + 68, + 540, + 288, + 562 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 540, + 288, + 562 + ], + "type": "text", + "content": " reliable lines, we generate all " + }, + { + "bbox": [ + 68, + 540, + 288, + 562 + ], + "type": "inline_equation", + "content": "2^{K}" + }, + { + "bbox": [ + 68, + 540, + 288, + 562 + ], + "type": "text", + "content": " line combinations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 570, + 216, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 570, + 216, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 216, + 582 + ], + "type": "text", + "content": "3.2. Assessing Line Combinations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 587, + 290, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 587, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 587, + 290, + 687 + ], + "type": "text", + "content": "When a line combination divides an image insufficiently or over-segments it into unnecessary parts, it does not describe the overall structure of the scene properly. On the contrary, an optimal combination of semantic lines should convey the image composition reliably and efficiently (i.e. with a small number of lines). To find the best line combination, we develop the semantic line combination detector (SLCD). Figure 3 shows the structure of SLCD, which performs encoding, semantic feature grouping, compositional feature ex" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 264, + 525, + 342 + ], + "blocks": [ + { + "bbox": [ + 307, + 264, + 525, + 342 + ], + "lines": [ + { + "bbox": [ + 307, + 264, + 525, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 525, + 342 + ], + "type": "image", + "image_path": "751fdd3902bfde92d1219577e0419c95b15c8569b7c4f69c0fd16a5f126b813b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 349, + 526, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 349, + 526, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 526, + 389 + ], + "type": "text", + "content": "Figure 4. From a set of line candidates, the line detector selects " + }, + { + "bbox": [ + 305, + 349, + 526, + 389 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 349, + 526, + 389 + ], + "type": "text", + "content": " reliable lines, depicted in orange. A high recall rate is achieved because a sufficient number of reliable lines are selected through NMS. Ground-truth semantic lines in green are in the insets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 411, + 525, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 525, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 525, + 434 + ], + "type": "text", + "content": "traction, and score regression. The detailed architecture of SLCD is described in the supplemental document." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "text", + "content": "Encoding: Given an image, we extract multi-scale feature maps of ResNet50 [11]. Then, we match their spatial resolutions via bilinear interpolation and concatenate them in the channel dimension. We then squeeze the channels using 2D convolutional layers to obtain an aggregated feature map " + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 437, + 525, + 514 + ], + "type": "text", + "content": " are the feature height, the feature width, and the number of channels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 518, + 525, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 525, + 583 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 525, + 583 + ], + "type": "text", + "content": "Semantic feature grouping: Semantic lines are located near the boundaries between distinct regions. For their reliable detection, it is desirable to separate different regional parts more clearly. We hence attempt to group pixel features into multiple regions through cross-attention [2, 22, 32, 33]. We employ three cross-attention modules." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "text", + "content": " be a learnable matrix representing " + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "text", + "content": " regions, called region query matrix. Then, we convert " + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "text", + "content": " into queries and " + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 585, + 525, + 618 + ], + "type": "text", + "content": " into keys and values by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 629, + 525, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 629, + 525, + 643 + ], + "spans": [ + { + "bbox": [ + 314, + 629, + 525, + 643 + ], + "type": "interline_equation", + "content": "R _ {q} = R U _ {q}, F _ {k} = (F + S) U _ {k}, F _ {v} = (F + S) U _ {v}. \\tag {1}", + "image_path": "1755173100994ff2e02c0e8207430f123464804dfa055246fc27fcae827f38ef.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 652, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 652, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 652, + 525, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 652, + 525, + 687 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{HW\\times C}" + }, + { + "bbox": [ + 305, + 652, + 525, + 687 + ], + "type": "text", + "content": " denotes the sinusoidal positional encoding [28], and " + }, + { + "bbox": [ + 305, + 652, + 525, + 687 + ], + "type": "inline_equation", + "content": "U_{q},U_{k},U_{v}\\in \\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 305, + 652, + 525, + 687 + ], + "type": "text", + "content": " are projection matrices for queries, keys, and values. We then obtain an updated" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "text", + "content": "28068" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 98, + 287, + 164 + ], + "blocks": [ + { + "bbox": [ + 71, + 98, + 287, + 164 + ], + "lines": [ + { + "bbox": [ + 71, + 98, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 71, + 98, + 287, + 164 + ], + "type": "image", + "image_path": "c89af6ab40adfaf0bd9a77941c6e68a8ffdda2acec7bf6f221d88a87872fbed4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 171, + 289, + 212 + ], + "lines": [ + { + "bbox": [ + 68, + 171, + 289, + 212 + ], + "spans": [ + { + "bbox": [ + 68, + 171, + 289, + 212 + ], + "type": "text", + "content": "Figure 5. Visualization of semantic feature grouping results. The top row shows input images with ground-truth lines. The bottom one presents the membership maps, representing the semantic region that each pixel belongs to." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 220, + 165, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 220, + 165, + 233 + ], + "spans": [ + { + "bbox": [ + 69, + 220, + 165, + 233 + ], + "type": "text", + "content": "region query matrix " + }, + { + "bbox": [ + 69, + 220, + 165, + 233 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 69, + 220, + 165, + 233 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 239, + 288, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 239, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 99, + 239, + 288, + 257 + ], + "type": "interline_equation", + "content": "A = \\operatorname {s o f t} _ {M} \\max \\left(R _ {q} F _ {k} ^ {\\mathrm {T}} / \\tau\\right), \\hat {R} = A F _ {v} + R, \\tag {2}", + "image_path": "0279afa2cfce76f15e2b1b0aad484c4ab6c05b2d5b736c91cea642456ba256cd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "spans": [ + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{M\\times HW}" + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "text", + "content": " is the attention matrix with a scaling factor " + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "text", + "content": " and the letter " + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "text", + "content": " means the softmax operation is done in the column direction, as in [22, 32, 33]. Then, in the last cross-attention, we generate a semantic feature map " + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "inline_equation", + "content": "F_{s}\\in \\mathbb{R}^{HW\\times C}" + }, + { + "bbox": [ + 68, + 263, + 289, + 317 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 318, + 288, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 318, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 156, + 318, + 288, + 330 + ], + "type": "interline_equation", + "content": "F _ {s} = A ^ {\\mathrm {T}} \\hat {R}. \\tag {3}", + "image_path": "3b7752d17f8577a611f4759da4fc17226f9d80f5a1be41e45e8e42e8f89688b7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "text", + "content": "Finally, we channel-wise concatenate " + }, + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "text", + "content": " to obtain a combined feature map " + }, + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "inline_equation", + "content": "X\\in \\mathbb{R}^{HW\\times 2C}" + }, + { + "bbox": [ + 68, + 335, + 289, + 356 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "text", + "content": "Note that, in (2), the softmax function is applied along the query axis. Thus, in " + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "inline_equation", + "content": "A = [\\mathbf{p}_1,\\mathbf{p}_2,\\dots ,\\mathbf{p}_{HW}]" + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "text", + "content": " each column " + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i\\in \\mathbb{R}^M" + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "text", + "content": " is a probability vector. Specifically, the mth element " + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "inline_equation", + "content": "p_i^m" + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "text", + "content": " is the probability that pixel " + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 357, + 289, + 422 + ], + "type": "text", + "content": " belongs to the mth region query. Figure 5 visualizes membership maps, computed by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 429, + 288, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 429, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 80, + 429, + 288, + 443 + ], + "type": "interline_equation", + "content": "\\left[ \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {1}\\right), \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {2}\\right), \\dots , \\operatorname {a r g m a x} \\left(\\mathbf {p} _ {H W}\\right) \\right] \\tag {4}", + "image_path": "2b04449d13cc0c322b6c86d4740b9e6681112d38c4b984474065007759218c97.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "text", + "content": "where each element is the index of the region query that the corresponding pixel most likely belongs to. We see that each image is partitioned into " + }, + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "text", + "content": " meaningful regions. Thus, " + }, + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "text", + "content": " in (3) represents the regional membership of each pixel and is used to make the feature map " + }, + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 68, + 449, + 289, + 503 + ], + "type": "text", + "content": " more discriminative." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 503, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 503, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 68, + 503, + 289, + 525 + ], + "type": "text", + "content": "To produce the attention matrix " + }, + { + "bbox": [ + 68, + 503, + 289, + 525 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 68, + 503, + 289, + 525 + ], + "type": "text", + "content": " reliably, we design a novel loss function in Section 3.3. The default " + }, + { + "bbox": [ + 68, + 503, + 289, + 525 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 503, + 289, + 525 + ], + "type": "text", + "content": " is 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "text", + "content": "Compositional feature extraction: For each line combination, we generate three types of feature maps to extract a compositional feature map. First, we generate a binary mask " + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{HW}" + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "text", + "content": " for each line combination. Each element in " + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "text", + "content": " is 1 if the corresponding pixel belongs to any line in the combination, and 0 otherwise. Then, we decompose the combined feature into a line feature map " + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "text", + "content": " and a region feature map " + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "inline_equation", + "content": "X_{r}" + }, + { + "bbox": [ + 68, + 528, + 289, + 616 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 623, + 288, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 623, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 110, + 623, + 288, + 635 + ], + "type": "interline_equation", + "content": "X _ {l} = X \\otimes B, X _ {r} = X \\otimes (1 - B), \\tag {5}", + "image_path": "35bfa4f94b461ecbc017f3d9f3c0d2e2436f019865b5f027891a60421a42ae7f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "text", + "content": " is the element-wise multiplication. In other words, " + }, + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "inline_equation", + "content": "X_{l}\\in \\mathbb{R}^{HW\\times 2C}" + }, + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "text", + "content": " contains contextual information for the line pixels only, whereas " + }, + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "inline_equation", + "content": "X_{r}\\in \\mathbb{R}^{HW\\times 2C}" + }, + { + "bbox": [ + 69, + 642, + 289, + 686 + ], + "type": "text", + "content": " does for the pixels strictly inside the regions divided by the lines." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 99, + 525, + 176 + ], + "blocks": [ + { + "bbox": [ + 307, + 99, + 525, + 176 + ], + "lines": [ + { + "bbox": [ + 307, + 99, + 525, + 176 + ], + "spans": [ + { + "bbox": [ + 307, + 99, + 525, + 176 + ], + "type": "image", + "image_path": "f64950abe3173d58c32d60b1fd0f88e3890ccb5499261c4034b08c36494c9e1e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 181, + 525, + 212 + ], + "lines": [ + { + "bbox": [ + 305, + 181, + 525, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 525, + 212 + ], + "type": "text", + "content": "Figure 6. Illustration of the line collection map generation when " + }, + { + "bbox": [ + 305, + 181, + 525, + 212 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 305, + 181, + 525, + 212 + ], + "type": "text", + "content": ": (a) reliable lines, (b) line combinations, and (c) line collection maps." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": "Moreover, we generate a line collection map " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "L = [L_1,L_2,\\dots ,L_K]\\in \\mathbb{R}^{HW\\times K}" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "L_{k}\\in \\mathbb{R}^{HW}" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": " is a ternary mask for the " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": "th reliable line, as illustrated in Figure 6. If the " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": "th reliable line does not belong to the line combination, all elements in " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "L_{k}" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": " are set to 0. Otherwise, each value in " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "L_{k}" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": " is set to either 1 or " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": " to indicate where the corresponding pixel is located between the two parts divided by the " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": "th reliable line. In other words, " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "L_{k}" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": " informs how the " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": "th line splits the image into two regions. We then produce a positional feature map " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{HW\\times 2C}" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": " by applying a series of fully connected layers to " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": ". Then, we yield a compositional feature map " + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "inline_equation", + "content": "Z\\in \\mathbb{R}^{HW\\times 6C}" + }, + { + "bbox": [ + 304, + 224, + 526, + 356 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 382, + 364, + 525, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 364, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 382, + 364, + 525, + 376 + ], + "type": "interline_equation", + "content": "Z = \\left[ X _ {l}, X _ {r}, P \\right], \\tag {6}", + "image_path": "d18614e7427bda0f7384be78646852b2b205f2df9ec8aff4d3be101eb4085924.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 385, + 525, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 525, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 525, + 407 + ], + "type": "text", + "content": "which contains information about how the lines in the combination separate the image into multiple parts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 410, + 526, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 526, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 526, + 498 + ], + "type": "text", + "content": "Score regression: Lastly, using a regressor, we estimate a composition score for each line combination. Then, we declare the line combination with the highest score as the optimal group of semantic lines. The regressor takes the compositional feature map " + }, + { + "bbox": [ + 304, + 410, + 526, + 498 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 410, + 526, + 498 + ], + "type": "text", + "content": " in (6) as input and predicts a composition score " + }, + { + "bbox": [ + 304, + 410, + 526, + 498 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 410, + 526, + 498 + ], + "type": "text", + "content": " within [0, 1]. It is implemented using a bilinear interpolation layer and a series of 2D convolution layers and fully connected layers with the ReLU activation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 505, + 391, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 505, + 391, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 505, + 391, + 516 + ], + "type": "text", + "content": "3.3. Loss Functions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 522, + 474, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 522, + 474, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 474, + 533 + ], + "type": "text", + "content": "To train SLCD, we design two loss functions." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": "Semantic region separation loss: When two pixels " + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": " are located in distinct regions, they should be assigned to different region queries. In other words, their probability vectors " + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": " in the attention matrix " + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": " should be far from each other. Let " + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 536, + 525, + 610 + ], + "type": "text", + "content": " be the two regions divided by a ground-truth line. Their probability vectors are defined as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 335, + 610, + 525, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 610, + 525, + 637 + ], + "spans": [ + { + "bbox": [ + 335, + 610, + 525, + 637 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {X} = \\frac {1}{| X |} \\sum_ {i \\in X} \\mathbf {p} _ {i}, \\quad \\mathbf {p} _ {Y} = \\frac {1}{| Y |} \\sum_ {j \\in Y} \\mathbf {p} _ {j}, \\tag {7}", + "image_path": "97e99cd6e4d93422656ff1a4b0ff44617014e1929c9a0b1c13b8c617a31650d7.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 642, + 524, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 524, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 524, + 663 + ], + "type": "text", + "content": "which should be also far from each other, for the ground-truth line divides the image into two semantic regions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 664, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 664, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 305, + 664, + 525, + 686 + ], + "type": "text", + "content": "To measure the distance between probability vectors, we adopt the Kullback-Leibler divergence (KLD) [5]. Then," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "text", + "content": "28069" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 108, + 287, + 203 + ], + "blocks": [ + { + "bbox": [ + 70, + 100, + 281, + 108 + ], + "lines": [ + { + "bbox": [ + 70, + 100, + 281, + 108 + ], + "spans": [ + { + "bbox": [ + 70, + 100, + 281, + 108 + ], + "type": "text", + "content": "Horizontal Vertical Diagonal Triangle Symmetric Low Front" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 108, + 287, + 203 + ], + "lines": [ + { + "bbox": [ + 70, + 108, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 70, + 108, + 287, + 203 + ], + "type": "image", + "image_path": "9ab17d575ca7742c214e5f6b6731cb5b27e54609641d4e07740989a0c89fa3e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 208, + 289, + 249 + ], + "lines": [ + { + "bbox": [ + 68, + 208, + 289, + 249 + ], + "spans": [ + { + "bbox": [ + 68, + 208, + 289, + 249 + ], + "type": "text", + "content": "Figure 7. The proposed CDL dataset contains 9,100 images in seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. The ground-truth semantic lines are depicted in green." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 262, + 289, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 262, + 289, + 284 + ], + "spans": [ + { + "bbox": [ + 68, + 262, + 289, + 284 + ], + "type": "text", + "content": "we define the semantic region segmentation (SRS) loss, by employing all ground-truth lines, as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 291, + 289, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 291, + 289, + 321 + ], + "spans": [ + { + "bbox": [ + 91, + 291, + 289, + 321 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S R S}} = - \\sum_ {l = 1} ^ {T} \\left[ D \\left(\\mathbf {p} _ {X _ {l}} \\| \\mathbf {p} _ {Y _ {l}}\\right) + D \\left(\\mathbf {p} _ {Y _ {l}} \\| \\mathbf {p} _ {X _ {l}}\\right) \\right] \\tag {8}", + "image_path": "4097172fa9fd65c1375ef275568cc642cd011435c3e53c527166bfd12277e3c6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "spans": [ + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": " denotes the KLD, " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{X_l}" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{Y_l}" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": " are the regions divided by the " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": "th ground-truth line, and " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": " is the number of ground-truth lines. Both " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "D(\\mathbf{p}_{X_l} \\parallel \\mathbf{p}_{Y_l})" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "D(\\mathbf{p}_{Y_l} \\parallel \\mathbf{p}_{X_l})" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": " are computed because " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": " is not commutative. With " + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SRS}}" + }, + { + "bbox": [ + 68, + 327, + 289, + 405 + ], + "type": "text", + "content": ", it is possible to roughly segment an image into meaningful parts, as illustrated in Figure 5, even though no ground-truth labels for semantic segmentation are used for training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "content": "Regression loss: We also design a loss for regressing the composition score of each line combination. To this end, we employ the HIoU metric [13] that quantifies the structural layout of a line combination. Let " + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\star}" + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "content": " denote a line combination and the ground-truth combination, respectively. We compute the HIoU between " + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\star}" + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "content": " and use it as the ground-truth composition score " + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\overline{s}" + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 68, + 407, + 289, + 495 + ], + "type": "text", + "content": ". Then, the regression loss is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 501, + 288, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 501, + 288, + 515 + ], + "spans": [ + { + "bbox": [ + 147, + 501, + 288, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = (s - \\bar {s}) ^ {2} \\tag {9}", + "image_path": "24c5cbbd5c7a376b4cd7f5ceabc96815a08aa6863f639a8d6dea8416d9db1480.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "text", + "content": " is the predicted score of " + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "text", + "content": ". To reduce " + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "text", + "content": " effectively, a ranking loss " + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rank}}" + }, + { + "bbox": [ + 68, + 521, + 289, + 544 + ], + "type": "text", + "content": " is also employed as in [17]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 552, + 185, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 552, + 185, + 565 + ], + "spans": [ + { + "bbox": [ + 69, + 552, + 185, + 565 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 571, + 191, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 571, + 191, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 571, + 191, + 582 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": "We adopt ResNet50 [11] as the encoder of the proposed SLCD. We use the AdamW optimizer [19] with a learning rate " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", a weight decay of " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ". We use a batch size of two for 400,000 iterations. Training images are resized to " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": " and augmented by random horizontal flipping. We fix the number of line candidates, reliable lines, and region queries to " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "N = 1024" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "M = 8" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", respectively. We set " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "H = 60" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "W = 60" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "inline_equation", + "content": "C = 96" + }, + { + "bbox": [ + 68, + 587, + 290, + 686 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 122, + 525, + 182 + ], + "blocks": [ + { + "bbox": [ + 305, + 99, + 525, + 118 + ], + "lines": [ + { + "bbox": [ + 305, + 99, + 525, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 99, + 525, + 118 + ], + "type": "text", + "content": "Table 1. Comparison of the HIoU scores " + }, + { + "bbox": [ + 305, + 99, + 525, + 118 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 99, + 525, + 118 + ], + "type": "text", + "content": " on the SEL, SEL_Hard, NKL, and CDL datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 122, + 525, + 182 + ], + "lines": [ + { + "bbox": [ + 308, + 122, + 525, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 122, + 525, + 182 + ], + "type": "table", + "html": "
SELSEL_HardNKLCDL
SLNet [15]77.8759.7165.4957.78
DHT [9]79.6263.3969.0863.24
DRM [12]80.2368.8367.4263.96
HSLD [13]81.0365.9974.2964.98
Proposed84.0968.1576.2168.85
", + "image_path": "5f9576c44757900c226d49acc436ebe6d0bcd93e03b7ae4acc6dd6ed06099b38.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 199, + 362, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 362, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 362, + 209 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 216, + 525, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 525, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 525, + 281 + ], + "type": "text", + "content": "SEL [15]: It is the first dataset for semantic line detection. It contains 1,750 images, which are split into 1,575 training and 175 test images. Each semantic line is annotated by the coordinates of two endpoints on an image boundary. In SEL, most images are high-quality landscapes, in which semantic lines are often obvious." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 284, + 525, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 525, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 525, + 317 + ], + "type": "text", + "content": "SEL_Hard [12]: It contains 300 test images only, selected from the ADE20K segmentation dataset [37]. It is challenging because of cluttered scenes or occluded objects." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 319, + 525, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 525, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 525, + 352 + ], + "type": "text", + "content": "NKL [35]: It is a relatively large dataset of 5,200 training and 1,300 testing images. It includes both indoor and outdoor scenes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 354, + 526, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 526, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 526, + 475 + ], + "type": "text", + "content": "CDL: We construct CDL to contain 7,100 scenes with diverse contents and compositions. It is split into 6,390 training and 710 test images. As in Figure 7, the images are categorized into seven composition classes: Horizontal, Vertical, Diagonal, Triangle, Symmetric, Low, and Front. In Low and Front, humans and animals are essential parts of the image composition. In the other classes, most images are outdoor ones, as in the other datasets. To construct CDL, semantic lines were manually annotated in about 200 man-hours. More examples and the annotation process are provided in the supplemental document." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 481, + 360, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 360, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 360, + 491 + ], + "type": "text", + "content": "4.3. Metrics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 498, + 525, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 525, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 525, + 564 + ], + "type": "text", + "content": "As mentioned earlier, the harmony of detected lines is more important than individual lines in semantic line detection. Therefore, the main paper discusses only HIoU performances. The precision, recall, and f-measure results, which are metrics used in previous work, are compared in the supplemental document." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 566, + 525, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 525, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 525, + 665 + ], + "type": "text", + "content": "HIoU: An optimal group of semantic lines conveys a harmonious impression about the composition of an image. To assess the overall harmony of detection results, we use the HIoU metric [13]. It measures the consistency between the division of an image by detected lines and that by the ground truth. Let " + }, + { + "bbox": [ + 304, + 566, + 525, + 665 + ], + "type": "inline_equation", + "content": "S = \\{s_1,s_2,\\dots ,s_N\\}" + }, + { + "bbox": [ + 304, + 566, + 525, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 566, + 525, + 665 + ], + "type": "inline_equation", + "content": "T = \\{t_{1},t_{2},\\ldots ,t_{M}\\}" + }, + { + "bbox": [ + 304, + 566, + 525, + 665 + ], + "type": "text", + "content": " be the regions divided by the detected lines and the ground-truth lines, respectively. Then, HIoU is computed as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 670, + 525, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 525, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 525, + 689 + ], + "type": "interline_equation", + "content": "\\mathrm {H I o U} = \\frac {\\sum_ {i = 1} ^ {N} \\max _ {k} \\operatorname {I o U} \\left(s _ {i} , t _ {k}\\right) + \\sum_ {j = 1} ^ {M} \\max _ {k} \\operatorname {I o U} \\left(t _ {j} , s _ {k}\\right)}{N + M}. \\tag {10}", + "image_path": "6b9273a35f958a55e07a52ca42e651908ca299c45631a254ab81c99ed4382d4f.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "text", + "content": "28070" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 91, + 525, + 342 + ], + "blocks": [ + { + "bbox": [ + 72, + 91, + 525, + 342 + ], + "lines": [ + { + "bbox": [ + 72, + 91, + 525, + 342 + ], + "spans": [ + { + "bbox": [ + 72, + 91, + 525, + 342 + ], + "type": "image", + "image_path": "7f40557f8557bd0bac1cbdba040bbc06303e37f7b735fa096631ecd5a814228d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 347, + 526, + 367 + ], + "lines": [ + { + "bbox": [ + 68, + 347, + 526, + 367 + ], + "spans": [ + { + "bbox": [ + 68, + 347, + 526, + 367 + ], + "type": "text", + "content": "Figure 8. Comparison of detected semantic lines. From the left, two images are selected from each of the SEL, SEL_Hard, NKL, and CDL datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "spans": [ + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "text", + "content": "Specifically, for each " + }, + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "text", + "content": ", the most overlapping region " + }, + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "text", + "content": " is chosen and the intersection-over-union (IoU) is calculated. This is performed similarly for each " + }, + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 68, + 381, + 290, + 437 + ], + "type": "text", + "content": ". Then, the HIoU score is given by the average of these bi-directionally matching IoUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 444, + 198, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 444, + 198, + 456 + ], + "spans": [ + { + "bbox": [ + 69, + 444, + 198, + 456 + ], + "type": "text", + "content": "4.4. Comparative Assessment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 461, + 290, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 461, + 290, + 560 + ], + "spans": [ + { + "bbox": [ + 68, + 461, + 290, + 560 + ], + "type": "text", + "content": "Table 1 compares the HIoU scores of the proposed SLCD with those of the existing detectors [9, 12, 13, 15] on the SEL, SEL_Hard, NKL, and CDL datasets. Also, Figure 8 compares detection results qualitatively. The existing detectors miss correct lines or fail to remove redundant lines, yielding sub-optimal results. In contrast, SLCD detects semantic lines more precisely and represents the composition more reliably than the existing detectors do. More detection results are available in the supplemental document." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 563, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 563, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 68, + 563, + 288, + 639 + ], + "type": "text", + "content": "Comparison on SEL: In Table 1, SLCD outperforms the existing detectors on SEL. Compared with the second-best HSLD, SLCD yields a wide HIoU margin of 3.06. This indicates that SLCD finds an optimal group of semantic lines more effectively by processing all lines in each combination simultaneously, instead of performing pairwise comparisons in HSLD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 642, + 290, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 642, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 642, + 290, + 687 + ], + "type": "text", + "content": "Comparison on SEL_Hard: As done in [12], we conduct experiments on SEL_Hard using the networks trained on the SEL dataset. In Table 1, SLCD ranks 2nd on SEL_Hard. DRM provides a better result than SLCD, but it demands" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 405, + 518, + 432 + ], + "blocks": [ + { + "bbox": [ + 306, + 380, + 525, + 401 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 525, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 525, + 401 + ], + "type": "text", + "content": "Table 2. Comparison of the HIoU scores " + }, + { + "bbox": [ + 306, + 380, + 525, + 401 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 306, + 380, + 525, + 401 + ], + "type": "text", + "content": " on the CDL dataset according to the number " + }, + { + "bbox": [ + 306, + 380, + 525, + 401 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 380, + 525, + 401 + ], + "type": "text", + "content": " of region queries." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 405, + 518, + 432 + ], + "lines": [ + { + "bbox": [ + 312, + 405, + 518, + 432 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 518, + 432 + ], + "type": "table", + "html": "
M481632
H IoU66.4368.8568.1167.07
", + "image_path": "4ec32f932ef9c5ddbb6406f86fb06d037bf16468c985adf4bcf5c791363fc34d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 441, + 524, + 514 + ], + "blocks": [ + { + "bbox": [ + 307, + 441, + 524, + 514 + ], + "lines": [ + { + "bbox": [ + 307, + 441, + 524, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 524, + 514 + ], + "type": "image", + "image_path": "5cad2bf36d941e55e506c89673e99df94b5231413db42ebfcacd7e0bb5a9c9ee.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 518, + 524, + 529 + ], + "lines": [ + { + "bbox": [ + 306, + 518, + 524, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 524, + 529 + ], + "type": "text", + "content": "Figure 9. Visualization of the membership maps according to " + }, + { + "bbox": [ + 306, + 518, + 524, + 529 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 518, + 524, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 546, + 524, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 524, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 524, + 557 + ], + "type": "text", + "content": "much higher complexity, as will be discussed in Section 4.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 561, + 525, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 561, + 525, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 561, + 525, + 594 + ], + "type": "text", + "content": "Comparison on NKL: SLCD surpasses all existing detectors. For example, its HIoU score is 1.92 points higher than the second-best HSLD." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "content": "Comparison on CDL: Table 1 also lists HIoU scores on the proposed CDL dataset. For a fair comparison, we train the existing detectors on CDL using their publicly available source codes. We see that SLCD outperforms all existing detectors on CDL as well. SLNet, DHT, and DRM yield poor results because they do not consider the overall harmony of detected lines. HSLD is better than these detectors but is inferior to the proposed SLCD. We compare and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "text", + "content": "28071" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 122, + 282, + 150 + ], + "blocks": [ + { + "bbox": [ + 69, + 98, + 289, + 118 + ], + "lines": [ + { + "bbox": [ + 69, + 98, + 289, + 118 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 289, + 118 + ], + "type": "text", + "content": "Table 3. Comparison of the HIoU scores " + }, + { + "bbox": [ + 69, + 98, + 289, + 118 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 69, + 98, + 289, + 118 + ], + "type": "text", + "content": " on the CDL dataset according to the usage of the SRS loss " + }, + { + "bbox": [ + 69, + 98, + 289, + 118 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SRS}}" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 122, + 282, + 150 + ], + "lines": [ + { + "bbox": [ + 75, + 122, + 282, + 150 + ], + "spans": [ + { + "bbox": [ + 75, + 122, + 282, + 150 + ], + "type": "table", + "html": "
w/o LSRProposed
HIoU66.7168.85
", + "image_path": "cb623a46b76daa409fb2eecc2b996a1e061d719a13e7bd9b4f02b39056242901.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 75, + 183, + 282, + 238 + ], + "blocks": [ + { + "bbox": [ + 69, + 160, + 288, + 179 + ], + "lines": [ + { + "bbox": [ + 69, + 160, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 288, + 179 + ], + "type": "text", + "content": "Table 4. Ablation studies for the compositional feature extraction on the CDL dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 183, + 282, + 238 + ], + "lines": [ + { + "bbox": [ + 75, + 183, + 282, + 238 + ], + "spans": [ + { + "bbox": [ + 75, + 183, + 282, + 238 + ], + "type": "table", + "html": "
LineRegionPositionalHIoU
I✓65.12
II✓65.74
III✓✓66.37
IV✓✓✓68.85
", + "image_path": "dc6e890ec4d1aa86c06743ca92b72b5a0ed70275bd72f58b9d859f85c27635df.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 256, + 288, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 256, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 68, + 256, + 288, + 278 + ], + "type": "text", + "content": "discuss the detection results according to the composition classes in the supplemental document." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 285, + 127, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 285, + 127, + 297 + ], + "spans": [ + { + "bbox": [ + 69, + 285, + 127, + 297 + ], + "type": "text", + "content": "4.5. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 303, + 288, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 303, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 68, + 303, + 288, + 324 + ], + "type": "text", + "content": "We conduct several ablation studies to analyze each component of the proposed SLCD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "text", + "content": "The number of " + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "text", + "content": ": Table 2 lists the HIoU performances on the CDL dataset, according to the number " + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "text", + "content": " of region queries. At the default " + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "inline_equation", + "content": "M = 8" + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "text", + "content": ", SLCD achieves the best performance. When " + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 327, + 289, + 393 + ], + "type": "text", + "content": " is smaller or bigger than 8, the performance degrades due to under- or over-segmentation of semantic parts, respectively, as shown in Figure 9." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 396, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 396, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 68, + 396, + 289, + 450 + ], + "type": "text", + "content": "Efficacy of " + }, + { + "bbox": [ + 68, + 396, + 289, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SRS}}" + }, + { + "bbox": [ + 68, + 396, + 289, + 450 + ], + "type": "text", + "content": ": In Table 3, if the proposed SRS loss " + }, + { + "bbox": [ + 68, + 396, + 289, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SRS}}" + }, + { + "bbox": [ + 68, + 396, + 289, + 450 + ], + "type": "text", + "content": " in (8) is excluded from the training, the performance drops by 2.14 points. It means that the composition analysis based on the SRS loss is essential for finding optimal line combinations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "text", + "content": "Efficacy of compositional feature extraction: SLCD extracts the compositional feature map " + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "text", + "content": " for each line combination, by processing the line feature map " + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "text", + "content": ", the region feature map " + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "inline_equation", + "content": "X_{r}" + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "text", + "content": ", and the positional feature map " + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 68, + 453, + 288, + 540 + ], + "type": "text", + "content": " in (6). In Table 4, Method I uses the line feature map only, while II uses the region feature map only. Method III utilizes both feature maps. Method IV, which is the proposed SLCD, uses all three maps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 541, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 541, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 68, + 541, + 288, + 639 + ], + "type": "text", + "content": "Method I yields the worst result, for it uses only the contextual information near line pixels. Method II is slightly better than Method I, by exploiting the regional information. Using both line and region feature maps, III provides a better result. Moreover, IV further improves the performance significantly by utilizing the positional feature map. This is because the overall harmony is estimated more effectively, by combining the line structures in the positional feature map with scene contexts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 642, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 642, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 642, + 289, + 687 + ], + "type": "text", + "content": "Runtime: Table 5 compares the runtimes of SLCD and the existing detectors in seconds per frame (spf). We use a PC with AMD Ryzen 9 3900X CPU and NVIDIA RTX 2080 GPU. The proposed SLCD takes 0.114spf, adding" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 133, + 520, + 160 + ], + "blocks": [ + { + "bbox": [ + 304, + 98, + 525, + 129 + ], + "lines": [ + { + "bbox": [ + 304, + 98, + 525, + 129 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 525, + 129 + ], + "type": "text", + "content": "Table 5. Runtime comparison of the proposed SLCD and the existing detectors. The processing times are reported in seconds per frame (spf)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 133, + 520, + 160 + ], + "lines": [ + { + "bbox": [ + 309, + 133, + 520, + 160 + ], + "spans": [ + { + "bbox": [ + 309, + 133, + 520, + 160 + ], + "type": "table", + "html": "
SLNet [15]DHT [9]DRM [12]HSLD [13]Proposed
0.1360.0330.9520.0460.114
", + "image_path": "8713657656d0c36a64321c4bb0145247e2b7d6da2a8416bd1b091c353b867d0a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 310, + 185, + 520, + 232 + ], + "blocks": [ + { + "bbox": [ + 315, + 171, + 515, + 182 + ], + "lines": [ + { + "bbox": [ + 315, + 171, + 515, + 182 + ], + "spans": [ + { + "bbox": [ + 315, + 171, + 515, + 182 + ], + "type": "text", + "content": "Table 6. Comparison of the AA scores " + }, + { + "bbox": [ + 315, + 171, + 515, + 182 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 315, + 171, + 515, + 182 + ], + "type": "text", + "content": " of VP detection." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 185, + 520, + 232 + ], + "lines": [ + { + "bbox": [ + 310, + 185, + 520, + 232 + ], + "spans": [ + { + "bbox": [ + 310, + 185, + 520, + 232 + ], + "type": "table", + "html": "
AA1°AA2°AA10°
Zhou et al. [39]18.533.060.0
Jin et al. [12]8.622.968.3
Proposed16.636.978.3
", + "image_path": "df92b167f65fd904d2e83b60fa11a7750248f4645163c77d1b2915c6d82eefc3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 244, + 351, + 276 + ], + "blocks": [ + { + "bbox": [ + 307, + 244, + 351, + 276 + ], + "lines": [ + { + "bbox": [ + 307, + 244, + 351, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 351, + 276 + ], + "type": "image", + "image_path": "959a4c3b38ec86f392b4e02c6b4a0a280d95ae345be6a7250cb0e7239250a6b3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 351, + 244, + 394, + 277 + ], + "blocks": [ + { + "bbox": [ + 351, + 244, + 394, + 277 + ], + "lines": [ + { + "bbox": [ + 351, + 244, + 394, + 277 + ], + "spans": [ + { + "bbox": [ + 351, + 244, + 394, + 277 + ], + "type": "image", + "image_path": "383589e349806de2088fb487d72d2f1b3f5e24dc5e8293ef75eece014f9bfacf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 351, + 277, + 394, + 309 + ], + "blocks": [ + { + "bbox": [ + 351, + 277, + 394, + 309 + ], + "lines": [ + { + "bbox": [ + 351, + 277, + 394, + 309 + ], + "spans": [ + { + "bbox": [ + 351, + 277, + 394, + 309 + ], + "type": "image", + "image_path": "b1ce6b0a7510c7a901fe5856c739fb376930ce7d418d30d8a4d2810cc2264807.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 317, + 525, + 348 + ], + "lines": [ + { + "bbox": [ + 304, + 317, + 525, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 525, + 348 + ], + "type": "text", + "content": "Figure 10. Detected dominant parallel lines and their intersections are depicted by yellow lines and green triangles, respectively. The ground-truth VPs are depicted by red dots." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 394, + 244, + 436, + 277 + ], + "blocks": [ + { + "bbox": [ + 394, + 244, + 436, + 277 + ], + "lines": [ + { + "bbox": [ + 394, + 244, + 436, + 277 + ], + "spans": [ + { + "bbox": [ + 394, + 244, + 436, + 277 + ], + "type": "image", + "image_path": "4b50d7c4b343f89aa571b578a144ba3922851fd495db9c62dc98175268075b18.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 394, + 277, + 436, + 309 + ], + "blocks": [ + { + "bbox": [ + 394, + 277, + 436, + 309 + ], + "lines": [ + { + "bbox": [ + 394, + 277, + 436, + 309 + ], + "spans": [ + { + "bbox": [ + 394, + 277, + 436, + 309 + ], + "type": "image", + "image_path": "eade586151d69ba3145ba6a22b7a10d581dc3537733da735cd6b6994a28a0203.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 436, + 244, + 480, + 277 + ], + "blocks": [ + { + "bbox": [ + 436, + 244, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 436, + 244, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 436, + 244, + 480, + 277 + ], + "type": "image", + "image_path": "6f2b0c4b156000e45d831c8818791939272e4089ac14af15dab5ceffdeb9d111.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 436, + 277, + 480, + 309 + ], + "blocks": [ + { + "bbox": [ + 436, + 277, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 436, + 277, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 436, + 277, + 480, + 309 + ], + "type": "image", + "image_path": "130d977ad799b84a0060fcba31190943ee9501631bf296dee4a57b2a57bfa2ad.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 480, + 244, + 523, + 277 + ], + "blocks": [ + { + "bbox": [ + 480, + 244, + 523, + 277 + ], + "lines": [ + { + "bbox": [ + 480, + 244, + 523, + 277 + ], + "spans": [ + { + "bbox": [ + 480, + 244, + 523, + 277 + ], + "type": "image", + "image_path": "78a0814479b3248024a17fe086e70206428606b9f9138c8acc8dc90eca298c9a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 480, + 277, + 523, + 309 + ], + "blocks": [ + { + "bbox": [ + 480, + 277, + 523, + 309 + ], + "lines": [ + { + "bbox": [ + 480, + 277, + 523, + 309 + ], + "spans": [ + { + "bbox": [ + 480, + 277, + 523, + 309 + ], + "type": "image", + "image_path": "b4a840a90cecb107679c05a08857c0d7eade2faab03adbb0e052c821ab165c83.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 361, + 525, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 525, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 525, + 416 + ], + "type": "text", + "content": "up 0.030spf and 0.074spf for generating and assessing line combinations, respectively. DHT is the fastest detector, but it is inferior to SLCD on all datasets. On the other hand, even though DRM performs better on SEL_Hard, it is about 8.4 times slower than SLCD." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "type": "text", + "content": "5. Applications" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 449, + 525, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 525, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 525, + 515 + ], + "type": "text", + "content": "We apply the proposed SLCD to three vision tasks: dominant vanishing point detection, reflection symmetry axis detection, and composition-based image retrieval. Note that the first two tasks were considered in [12]. Due to the page limit, more details and results are described in the supplemental document." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 525, + 484, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 484, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 484, + 537 + ], + "type": "text", + "content": "5.1. Dominant Vanishing Point Detection" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 544, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 526, + 687 + ], + "type": "text", + "content": "A vanishing point (VP) facilitates understanding of the 3D geometric structure. We apply the proposed SLCD to detect dominant VPs, by identifying vanishing lines. We use the AVA landscape dataset [39], in which two dominant parallel lines are annotated for each image. To detect a dominant VP, we generate line combinations containing two lines only. We then find the best combination, whose intersecting point is declared as a VP. Table 6 compares this VP detection scheme with the existing line-based VP detectors in [12, 39]. The angle accuracies (AAs) [38] are compared. We see that SLCD is better than the existing detectors, except that its " + }, + { + "bbox": [ + 304, + 544, + 526, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{AA1}^{\\circ}" + }, + { + "bbox": [ + 304, + 544, + 526, + 687 + ], + "type": "text", + "content": " score is lower than that of Zhou et al. [39]. Figure 10 shows some VP detection results." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 318, + 745 + ], + "type": "text", + "content": "28072" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 121, + 282, + 191 + ], + "blocks": [ + { + "bbox": [ + 69, + 98, + 288, + 118 + ], + "lines": [ + { + "bbox": [ + 69, + 98, + 288, + 118 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 288, + 118 + ], + "type": "text", + "content": "Table 7. Comparison of the AUC_A scores " + }, + { + "bbox": [ + 69, + 98, + 288, + 118 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 69, + 98, + 288, + 118 + ], + "type": "text", + "content": " of symmetry axis detection." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 121, + 282, + 191 + ], + "lines": [ + { + "bbox": [ + 73, + 121, + 282, + 191 + ], + "spans": [ + { + "bbox": [ + 73, + 121, + 282, + 191 + ], + "type": "table", + "html": "
ICCVNYUSYM_Hard
Cicconet et al. [3]80.8082.8568.99
Elawady et al. [6]87.2483.8373.90
Cicconet et al. [4]87.3887.6481.04
Loy & Eklundh [20]89.7790.8581.99
Jin et al. [12]90.6092.7884.73
Proposed93.1593.5188.03
", + "image_path": "10af0201aa97ea5ea3e60f7ddb9c8d9a1523d32f4f70e1694200e791bc0a6af7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 193, + 287, + 259 + ], + "blocks": [ + { + "bbox": [ + 72, + 193, + 287, + 259 + ], + "lines": [ + { + "bbox": [ + 72, + 193, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 72, + 193, + 287, + 259 + ], + "type": "image", + "image_path": "b2c64ee54b58ceab391233324a1453dd9c9ffb0f86dd748a3ac2736c7eb081d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 262, + 289, + 302 + ], + "lines": [ + { + "bbox": [ + 68, + 262, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 68, + 262, + 289, + 302 + ], + "type": "text", + "content": "Figure 11. Symmetry axis detection results. In the top row, the ground-truth and predicted axes are depicted by dashed red and solid yellow lines, respectively. The membership maps are also visualized in the bottom row." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 311, + 244, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 311, + 244, + 323 + ], + "spans": [ + { + "bbox": [ + 69, + 311, + 244, + 323 + ], + "type": "text", + "content": "5.2. Reflection Symmetry Axis Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 324, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 324, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 68, + 324, + 290, + 478 + ], + "type": "text", + "content": "Reflection symmetry is a visual cue for analyzing object shapes and patterns. Its detection is challenging because the symmetry is only implicit in many cases. We test the proposed SLCD on three datasets: ICCV [7], NYU [4], and SYM_Hard [12]. In these datasets, each image contains a single reflection symmetry axis. Thus, in the proposed SLCD, each of the " + }, + { + "bbox": [ + 68, + 324, + 290, + 478 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 324, + 290, + 478 + ], + "type": "text", + "content": " reliable lines is regarded as a line combination. Table 7 compares the AUC_A scores [15] of the proposed SLCD and the conventional techniques [3, 4, 6, 12, 20]. SLCD outperforms all these techniques on all datasets. Figure 11 shows some detection results together with the membership maps. We see that regional parts are symmetrically divided along implied axes, enabling SLCD to identify those axes effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 486, + 244, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 486, + 244, + 498 + ], + "spans": [ + { + "bbox": [ + 69, + 486, + 244, + 498 + ], + "type": "text", + "content": "5.3. Composition-Based Image Retrieval" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 499, + 288, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 499, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 68, + 499, + 288, + 554 + ], + "type": "text", + "content": "Existing image retrieval techniques focus on the visual contents of a query image to find similar images in a database [8, 21, 24, 25]. In this work, however, we attempt to discover images with similar compositions to a query image, i.e. composition-based image retrieval." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "text", + "content": "We test the proposed SLCD on the Oxford 5k and Paris 6k retrieval dataset [23]. We first detect semantic lines in every image, while storing the positional feature map " + }, + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "text", + "content": " in (6). We filter out some images whose composition scores are lower than a threshold since a low score indicates a low-quality image with inharmonious composition in general. Then, for a randomly selected query image, we compute the " + }, + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "text", + "content": "-distances between the positional feature maps " + }, + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 68, + 554, + 288, + 663 + ], + "type": "text", + "content": " of the query and the remaining images. We determine the images with the smallest distances as retrieval results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 664, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 664, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 664, + 289, + 687 + ], + "type": "text", + "content": "Figure 12 shows the top-4 retrieval results for four query images. We see that SLCD returns structurally similar im" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 99, + 352, + 195 + ], + "blocks": [ + { + "bbox": [ + 308, + 99, + 352, + 195 + ], + "lines": [ + { + "bbox": [ + 308, + 99, + 352, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 99, + 352, + 195 + ], + "type": "image", + "image_path": "576dbbd33363f999f807e75a277b54742f5fe14c30814966efee0663584381cd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 196, + 350, + 204 + ], + "lines": [ + { + "bbox": [ + 310, + 196, + 350, + 204 + ], + "spans": [ + { + "bbox": [ + 310, + 196, + 350, + 204 + ], + "type": "text", + "content": "Query image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 354, + 99, + 524, + 195 + ], + "blocks": [ + { + "bbox": [ + 354, + 99, + 524, + 195 + ], + "lines": [ + { + "bbox": [ + 354, + 99, + 524, + 195 + ], + "spans": [ + { + "bbox": [ + 354, + 99, + 524, + 195 + ], + "type": "image", + "image_path": "dc92edc03a8701496e64266227c326dfb56783c1f8cc4e0fff47ad3a22fc319a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 196, + 471, + 204 + ], + "lines": [ + { + "bbox": [ + 406, + 196, + 471, + 204 + ], + "spans": [ + { + "bbox": [ + 406, + 196, + 471, + 204 + ], + "type": "text", + "content": "Top-4 retrieval results" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 220, + 523, + 331 + ], + "blocks": [ + { + "bbox": [ + 307, + 205, + 522, + 216 + ], + "lines": [ + { + "bbox": [ + 307, + 205, + 522, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 522, + 216 + ], + "type": "text", + "content": "Figure 12. Composition-based retrieval results for query images." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 220, + 523, + 331 + ], + "lines": [ + { + "bbox": [ + 308, + 220, + 523, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 523, + 331 + ], + "type": "image", + "image_path": "ae0a8aa009561a0ca6d71390ed852a25cbe9500eaf2042fd3f89e7d85b2315b2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 335, + 525, + 354 + ], + "lines": [ + { + "bbox": [ + 305, + 335, + 525, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 525, + 354 + ], + "type": "text", + "content": "Figure 13. t-SNE visualization [27] of the feature space for the Oxford 5k and Paris 6k dataset [23]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 363, + 525, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 525, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 525, + 473 + ], + "type": "text", + "content": "ages to the queries. This means that the positional feature maps " + }, + { + "bbox": [ + 304, + 363, + 525, + 473 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 363, + 525, + 473 + ], + "type": "text", + "content": ", containing the structural information of semantic lines, can be used to compute the compositional differences. Based on this observation, we perform data clustering by employing k-means [10] on the feature space of " + }, + { + "bbox": [ + 304, + 363, + 525, + 473 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 363, + 525, + 473 + ], + "type": "text", + "content": ". Figure 13 is t-SNE visualization [27] of the clustering results. The nine images nearest to each centroid are also shown. Note that images with similar composition are grouped into the same cluster, confirming the efficacy of SLCD in the compositional analysis of images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 479, + 376, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 376, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 376, + 490 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 493, + 525, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 525, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 525, + 604 + ], + "type": "text", + "content": "We proposed a novel semantic line detector, SLCD, which processes a combination of lines at once to estimate the overall harmony reliably. We first generated all possible line combinations from reliable lines. Then, we estimated the score of each line combination and determined the best combination. Experimental results demonstrated that the proposed SLCD can detect semantic lines reliably on existing datasets, as well as on the new dataset CDL. Furthermore, SLCD can be successfully used in vanishing point detection, symmetry axis detection, and image retrieval." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 608, + 400, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 400, + 619 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 400, + 619 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 620, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 525, + 686 + ], + "type": "text", + "content": "This work was conducted by Center for Applied Research in Artificial Intelligence (CARAI) grant funded by DAPA and ADD (UD230017TD) and was supported by the National Research Foundation of Korea (NRF) grants funded by the Korea government (MSIT) (No. NRF-2021R1A4A1031864 and No. NRF-2022R1A2B5B03002310)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "text", + "content": "28073" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 117, + 289, + 686 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 74, + 117, + 289, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 117, + 289, + 166 + ], + "spans": [ + { + "bbox": [ + 74, + 117, + 289, + 166 + ], + "type": "text", + "content": "[1] Jean-Charles Bazin, Yongduek Seo, Cedric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in Manhattan world. In Proc. IEEE CVPR, 2012. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 167, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 167, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 74, + 167, + 288, + 207 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 209, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 209, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 74, + 209, + 288, + 247 + ], + "type": "text", + "content": "[3] Marcelo Cicconet, Vighnesh Birodkar, Mads Lund, Michael Werman, and Davi Geiger. A convolutional approach to reflection symmetry. Pattern Recog. Lett., 95:44-50, 2017. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 249, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 249, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 74, + 249, + 288, + 289 + ], + "type": "text", + "content": "[4] Marcelo Cicconet, David GC Hildebrand, and Hunter Elliott. Finding mirror symmetry via registration and optimal symmetric pairwise assignment of curves: Algorithm and results. In Proc. IEEE ICCV Workshops, 2017. 1, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 290, + 288, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 290, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 74, + 290, + 288, + 309 + ], + "type": "text", + "content": "[5] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley, 2nd edition, 2006. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 310, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 310, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 74, + 310, + 288, + 350 + ], + "type": "text", + "content": "[6] Mohamed Elawady, Christophe Ducottet, Olivier Alata, Cécile Barat, and Philippe Colantoni. Wavelet-based reflection symmetry detection via textural and color histograms. In Proc. IEEE ICCV Workshops, 2017. 1, 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 351, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 351, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 74, + 351, + 288, + 390 + ], + "type": "text", + "content": "[7] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 ICCV Challenge: Detecting symmetry in the wild. In Proc. IEEE ICCV, 2017. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 392, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 392, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 74, + 392, + 288, + 422 + ], + "type": "text", + "content": "[8] Albert Gordo, Jon Almazan, Jerome Revaud, and Diane Larlus. End-to-end learning of deep visual representations for image retrieval. Int. J. Comput. Vis., 124:237-254, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 422, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 422, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 74, + 422, + 288, + 452 + ], + "type": "text", + "content": "[9] Qi Han, Kai Zhao, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. In Proc. ECCV, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 453, + 288, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 453, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 70, + 453, + 288, + 482 + ], + "type": "text", + "content": "[10] J. A. Hartigan and M. A. Wong. A K-means clustering algorithm. Journal of the Royal Statistical Society. Series C (Applied Statistics), 28(1):100-108, 1979. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 483, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 483, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 70, + 483, + 288, + 512 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE CVPR, 2016. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 514, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 514, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 70, + 514, + 288, + 544 + ], + "type": "text", + "content": "[12] Dongkwon Jin, Jun-Tae Lee, and Chang-Su Kim. Semantic line detection using mirror attention and comparative ranking and matching. In Proc. ECCV, 2020. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 544, + 288, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 544, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 70, + 544, + 288, + 583 + ], + "type": "text", + "content": "[13] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, and Chang-Su Kim. Harmonious semantic line detection via maximal weight clique selection. In Proc. IEEE CVPR, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 70, + 585, + 288, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 585, + 288, + 614 + ], + "spans": [ + { + "bbox": [ + 70, + 585, + 288, + 614 + ], + "type": "text", + "content": "[14] Florian Kluger, Hanno Ackermann, Michael Ying Yang, and Bodo Rosenhahn. Temporally consistent horizon lines. In Proc. IEEE ICRA, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 70, + 616, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 616, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 70, + 616, + 288, + 645 + ], + "type": "text", + "content": "[15] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Semantic line detection and its applications. In Proc. IEEE ICCV, 2017. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 70, + 647, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 647, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 70, + 647, + 288, + 686 + ], + "type": "text", + "content": "[16] Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Photographic composition classification and dominant geometric element detection for outdoor scenes. J. Vis. Commun. Image Represent., 55:91-105, 2018. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 524, + 685 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 101, + 524, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 101, + 524, + 130 + ], + "spans": [ + { + "bbox": [ + 307, + 101, + 524, + 130 + ], + "type": "text", + "content": "[17] Debang Li, Junge Zhang, Kaiqi Huang, and Ming-Hsuan Yang. Composing good shots by exploiting mutual relations. In Proc. IEEE CVPR, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 132, + 524, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 524, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 524, + 161 + ], + "type": "text", + "content": "[18] Ligang Liu, Renjie Chen, Lior Wolf, and Daniel Cohen-Or. Optimizing photo composition. Computer Graphics Forum, 29:469-478, 2010. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 524, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 524, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 524, + 182 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 184, + 524, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 524, + 213 + ], + "type": "text", + "content": "[20] Gareth Loy and Jan-Olof Eklundh. Detecting symmetry and symmetric constellations of features. In Proc. ECCV, 2006. 1, 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 214, + 524, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 214, + 524, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 214, + 524, + 244 + ], + "type": "text", + "content": "[21] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proc. IEEE ICCV, 2017. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 246, + 524, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 524, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 524, + 275 + ], + "type": "text", + "content": "[22] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. iDisc: Internal discretization for monocular depth estimation. In Proc. IEEE CVPR, 2023. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 277, + 524, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 524, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 524, + 316 + ], + "type": "text", + "content": "[23] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proc. IEEE CVPR, 2018. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 317, + 524, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 524, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 524, + 346 + ], + "type": "text", + "content": "[24] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell., 41:1655-1668, 2018. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 349, + 524, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 524, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 524, + 388 + ], + "type": "text", + "content": "[25] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. IEEE Trans. Pattern Anal. Mach. Intell., 22:1349-1380, 2000. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 389, + 524, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 524, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 524, + 429 + ], + "type": "text", + "content": "[26] Lucas Tabelini, Rodrigo Berriel, Thiago M. Paixao, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 430, + 524, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 524, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 524, + 460 + ], + "type": "text", + "content": "[27] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (11), 2008. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 461, + 524, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 461, + 524, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 461, + 524, + 500 + ], + "type": "text", + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 502, + 524, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 524, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 524, + 551 + ], + "type": "text", + "content": "[29] Yaoting Wang, Yongzhen Ke, Kai Wang, Jing Guo, and Shuai Yang. Spatial-invariant convolutional neural network for photographic composition prediction and automatic correction. J. Vis. Commun. Image Represent., 90:103751, 2023. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 553, + 524, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 553, + 524, + 572 + ], + "spans": [ + { + "bbox": [ + 307, + 553, + 524, + 572 + ], + "type": "text", + "content": "[30] Scott Workman, Menghua Zhai, and Nathan Jacobs. Horizon lines in the wild. In Proc. BMVC, 2016. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 574, + 524, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 574, + 524, + 594 + ], + "spans": [ + { + "bbox": [ + 307, + 574, + 524, + 594 + ], + "type": "text", + "content": "[31] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proc. IEEE ICCV, 2015. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 595, + 524, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 595, + 524, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 595, + 524, + 644 + ], + "type": "text", + "content": "[32] Qihang Yu, Huiyu Wang, Dahun Kim, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. CMT-DeepLab: Clustering mask transformers for panoptic segmentation. In Proc. IEEE CVPR, 2022. 3, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "type": "text", + "content": "[33] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. " + }, + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 307, + 646, + 524, + 685 + ], + "type": "text", + "content": "-means mask transformer. In Proc. ECCV, 2022. 3, 4" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 745 + ], + "type": "text", + "content": "28074" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 288, + 356 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 70, + 100, + 288, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 100, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 70, + 100, + 288, + 130 + ], + "type": "text", + "content": "[34] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proc. IEEE CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 132, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 132, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 70, + 132, + 288, + 170 + ], + "type": "text", + "content": "[35] Kai Zhao, Qi Han, Chang-Bin Zhang, Jun Xu, and Ming-Ming Cheng. Deep hough transform for semantic line detection. IEEE Trans. Pattern Anal. Mach. Intell., 44:4793-4806, 2021. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 172, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 172, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 70, + 172, + 288, + 212 + ], + "type": "text", + "content": "[36] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. CLRNet: Cross layer refinement network for lane detection. In Proc. IEEE CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 214, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 214, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 70, + 214, + 288, + 243 + ], + "type": "text", + "content": "[37] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In Proc. IEEE CVPR, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 245, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 245, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 70, + 245, + 288, + 274 + ], + "type": "text", + "content": "[38] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. NeurVPS: Neural vanishing point scanning via conic convolution. In Proc. NeurIPS, 2019. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 276, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 276, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 70, + 276, + 288, + 315 + ], + "type": "text", + "content": "[39] Zihan Zhou, Farshid Farhat, and James Z Wang. Detecting dominant vanishing points in natural scenes with application to composition-sensitive image retrieval. IEEE Trans. Multimedia, 19:2651-2665, 2017. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 317, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 317, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 70, + 317, + 288, + 356 + ], + "type": "text", + "content": "[40] Rui Zhu, Xingyi Yang, Yannick Hold-Geoffroy, Federico Perazzi, Jonathan Eisenmann, Kalyan Sunkavalli, and Manmohan Chandraker. Single view metrology in the wild. In Proc. ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "28075" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_content_list.json b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db4a9fff335048d69123cf829822eea5fa7605c2 --- /dev/null +++ b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_content_list.json @@ -0,0 +1,1483 @@ +[ + { + "type": "text", + "text": "Separating the \"Chirp\" from the \"Chat\": Self-supervised Visual Grounding of Sound and Language", + "text_level": 1, + "bbox": [ + 191, + 130, + 779, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mark Hamilton \nMIT, Microsoft \nmarkth@mit.edu", + "bbox": [ + 132, + 203, + 261, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Zisserman Oxford, Google", + "bbox": [ + 303, + 203, + 455, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "John R. Hershey Google", + "bbox": [ + 500, + 203, + 630, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "William T. Freeman \nMIT, Google", + "bbox": [ + 676, + 203, + 834, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbd6229aa8445331f3b0e4631e0a77170938c732de2277a7475a59f813380723.jpg", + "image_caption": [ + "Figure 1. Visual overview of the DenseAV algorithm. Two modality-specific backbones featurize audio and visual signals. We introduce a novel generalization of multi-head attention to extract attention maps that discover and separate the \"meaning\" of spoken words and the sounds an object makes. DenseAV performs this localization and decomposition solely through observing paired stimuli such as videos." + ], + "image_footnote": [], + "bbox": [ + 84, + 277, + 895, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 532, + 313, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present DenseAV, a novel dual encoder grounding architecture that learns high-resolution, semantically meaningful, and audio-visual aligned features solely through watching videos. We show that DenseAV can discover the \"meaning\" of words and the \"location\" of sounds without explicit localization supervision. Furthermore, it automatically discovers and distinguishes between these two types of associations without supervision. We show that DenseAV's localization abilities arise from a new multi-head feature aggregation operator that directly compares dense image and audio representations for contrastive learning. In contrast, many other systems that learn \"global\" audio and video representations cannot localize words and sound. Finally, we contribute two new datasets to improve the evaluation of AV representations through speech and sound prompted semantic segmentation. On these and other datasets we show DenseAV dramatically outperforms the prior art on speech and sound prompted semantic segmentation. DenseAV outperforms the current state-of-the-art, ImageBind, on cross-modal retrieval using fewer than half of the parameters. Project Page: https://aka.ms/denseav", + "bbox": [ + 75, + 565, + 473, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 532, + 630, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Associating audio and video is a fundamental aspect of human perception. As infants develop, the synchronization and correspondence of sounds enables multi-modal association – a voice with a face, and a 'moo' with a cow [50]. Later, as they acquire language, they associate spoken words with objects they represent [10, 45]. Amazingly, these association abilities, constituting speech recognition, sound event recognition, and visual object recognition, develop without much direct supervision. This work aims to create a model with this capability by learning high-resolution, semantically meaningful, audio-visual (AV) aligned representations. Features with these properties can be used to discover fine-grained correspondences between modalities without localization supervision or knowledge of the textual representation of language.", + "bbox": [ + 496, + 561, + 890, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As an example, consider the spoken caption and accompanying sounds of the image shown in Figure 1. We wish to \"ground\" both the speech and the sounds with high resolution. For instance, both the spoken word \"dog\" and the sound of a bark in the audio signal should be associated with the pixels of the dog in the visual signal if present. We seek high quality local representations where simple inner", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13117", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90aafa0b6605adfbec68b597cef114d2f07e1b8e2f7999d23fcfcce7dcfbb642.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 128, + 236, + 270 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2f1046456c605eb34400f3e15b8b967034ac0ec032c9d9eabe6a9e8c7ddb2ab.jpg", + "image_caption": [ + "Figure 2. Qualitative comparison of several modern architectures for associating audio and video modalities. Only DenseAV learns a high-resolution and semantically aligned set of local features. This allows us to perform speech and sound prompted semantic segmentation using only the inner products between deep features. Other approaches, such as ImageBind, do not show aligned local feature maps. Approaches that do show some localization capabilities, like DAVENet, do not generalize to sound and language, and do not achieve the high-resolution localization capabilities of DenseAV. Dense features are visualized using PCA as in [20]" + ], + "image_footnote": [], + "bbox": [ + 259, + 70, + 890, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "products between features exhibit this behavior, which is interestingly absent from popular approaches in the literature.", + "bbox": [ + 75, + 419, + 468, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we make three innovations. First, we introduce DenseAV, a dual-encoder architecture that computes a dense similarity volume over audio and visual features before aggregating this volume into a single similarity score. If we look at a slice of this similarity volume when word is spoken, as in Figure 1, we can visualize the AV activation strength between a word or sound and an image's pixels. The novelty we introduce is to generalize this dense similarity mechanism with multiple heads, much like those of multi-head attention. This generalization allows each head to specialize on particular types of couplings between the visual and audio modalities. Interestingly, we discover that if we give DenseAV two heads and train on a dataset that contains both language and sound, the heads naturally learn to distinguish language from more general sound using only cross-modality supervision. For example, as shown in Figure 1, head 1 focuses on what objects create a sound, like the barking of a dog, whereas head 2 focuses only on the meaning of words.", + "bbox": [ + 75, + 455, + 468, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we show the importance of the \"aggregation function\" one uses to create a similarity score between an audio clip and a video frame for contrastive learning. The traditional choices; using inner products between global representations such as class tokens [5, 13, 49] or pooled features [18, 58], do not promote AV alignment of dense local features. Because of this, several popular audio-video backbones that excel on cross-modal retrieval cannot directly associate objects and sounds using their local features. This limits their ability to be used for downstream", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks such as semantic segmentation, sound localization, or unsupervised language learning and discovery.", + "bbox": [ + 498, + 420, + 890, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Third, we introduce two semantic segmentation datasets to evaluate visual grounding with AV representations for speech and (non-speech) sounds. We build these datasets from the high-quality segmentation masks provided by the ADE20K dataset [59] and measure mean average precision (mAP) and mean intersection over union (mIoU) on a binary mask prediction task. This evaluation is simpler and more thorough than previous efforts to measure visual grounding such as the concept counting metrics of [23] and the \"pointing games\" of [2, 14, 38] that only check if a heatmap's peak occurs within a target box or segment. Furthermore, our evaluation avoids brittle word-net ontologies [34], clustering, Wu and Palmer distance [55], threshold choices, and a variety of other complicating factors.", + "bbox": [ + 496, + 454, + 892, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 519, + 669, + 870, + 683 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce DenseAV, a novel self-supervised architecture that learns high-resolution AV correspondences.", + "- We introduce a local-feature-based image similarity function that significantly improves a network's zero-shot localization ability compared to common strategies such as average pooling or CLS tokens.", + "- We introduce new datasets for evaluating speech and sound prompted semantic segmentation. We show DenseAV significantly outperforms the current state-of-the-art on these tasks as well as on cross-modal retrieval.", + "- We discover that our multi-head architecture naturally disentangles audio-visual correspondence into sound and language components using only contrastive supervision." + ], + "bbox": [ + 500, + 695, + 890, + 891 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13118", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/039a8c44a90011de72a3148624ab418a96b3fc9f8ab983fa4374d1d4b7feeec0.jpg", + "image_caption": [ + "Figure 3. Architectural overview of our multi-head attention aggregator. Dense feature maps are split into $K$ heads ( $K = 1,2$ ) in our experiments. We form an AV activation tensor by taking the inner-products of each head's features across the spatial and temporal extent of the visual and audio signals respectively as in Equation 1. We then aggregate this similarity volume into a single similarity score by max-pooling head and spatial dimensions and average-pooling audio dimensions. Our approach aims to encourage the network to identify specific shared objects between the audio and visual modalities. In particular, max-pooling of heads disentangles sound and language, and max-pooling spatial dimensions helps localize objects." + ], + "image_footnote": [], + "bbox": [ + 148, + 70, + 851, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 376, + 218, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-visual (AV), text-visual, and more general multimodal models have recently surged in popularity [60]. Broadly speaking DenseAV is an audio-video contrastive learning architecture, this class of methods learns AV representations by aligning paired signals and pushing apart negative signals [11, 27]. Of the models in this class, several stand out for their ability to localize sounds [2, 7, 42] or capture the semantics of language [23, 43]. Many models in this class compare AV signals using inner products between \"global\" representations formed by pooled deep features [18, 35, 53], or class tokens [17, 32, 42, 43, 49]. Most notably, ImageBind has gained popularity due to its state-of-the-art performance on a variety of tasks and datasets and unified class-token-based contrastive architecture. In this work we show that many of these architectures do not show strong localization properties in their local features, despite excelling at cross-modal retrieval on a \"global\" level. This limits their applicability to new out-of-domain sounds, sounds that don't have a textual representation, and low-resource languages. We diverge from these works by directly supervising local tokens. In particular, we build on previous works [2, 23] that show max-pooling improves localization capabilities and introduce a new multi-head aggregation operator that generalizes previous losses using an operator similar to self-attention [52].", + "bbox": [ + 75, + 410, + 472, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another class of methods discover structure in signals through uni- and multi-modal clustering. Early works on audio clustering [41] discovered meaningful utterances without supervision. Similar visual analyses have discovered visual objects [4, 8, 20, 28]. Recent works have applied these ideas to the AV domain [1, 21], but do not focus on extracting high-resolution AV representations.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, several works investigate generative audio-video learning. The Sound of Pixels [57] generates the sound of a specific object using a source separation loss. Newer approaches using GANs [30, 31], and diffusion models [9, 17, 33] have generated audio from video and vice versa. Here we focus on improving the local representations of contrastive learners because of their relative scalability, simplicity, and ability to learn high-quality representations.", + "bbox": [ + 496, + 378, + 893, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 512, + 599, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At a high level, DenseAV tries to determine when a given audio and visual signal belong “together” using dense audio-visual representations. To perform this task robustly, DenseAV must learn how to predict the contents of an audio signal from a visual signal and vice versa. This causes DenseAV to learn dense modality-specific features that capture the mutual information shared between the modalities [51]. Once learned, we can directly query these informative features to perform speech and sound prompted semantic segmentation as illustrated in Figure 1.", + "bbox": [ + 496, + 537, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, DenseAV is built from two modality-specific deep featurizers. These backbones produce temporally varying audio features across an audio clip and spatially varying video features for a single randomly selected frame. Our loss computes a similarity between audio and visual signals based on the intuition that two signals are similar if they have a variety of strong couplings or shared objects. More formally, we form a scalar similarity for a pair of audio and video signals by carefully aggregating a volume of pairwise inner products between dense features. We use the InfoNCE [36] contrastive loss to encourage similarity between \"positive\" pairs of signals and dissimilarity between \"negative\" pairs formed by in-batch shuffling. Figure 3 graphically depicts this loss function and subsequent", + "bbox": [ + 496, + 689, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13119", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sections detail each component of our architecture.", + "bbox": [ + 76, + 90, + 413, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Multi-Headed Aggregation of Similarities", + "text_level": 1, + "bbox": [ + 76, + 118, + 431, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DenseAV's key architectural distinction is its loss function that directly supervises the \"local\" tokens of the visual and audio featurizers. This is a significant departure from other works [5, 17, 19, 39, 46, 49] that pool modality specific information into \"global\" representations prior to the contrastive loss. Unlike prior works, our loss function aggregates the full pairwise similarities between the local tokens into an aggregate measure of similarity for a given pair of audio and visual signals. We show in Figure 2 that this architectural choice enables DenseAV's local features to align across modalities whereas other approaches such as average pooling, class tokens, and SimPool [44] do not.", + "bbox": [ + 75, + 143, + 468, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first describe our loss function informally and definite it more precisely in the next paragraph. Our loss function computes the (un-normalized) inner product between every pair of visual and audio features to form a \"volume\" of inner products. This volume represents how strongly each part of an audio signal \"couples\" to each part of a visual signal. We aim to find many large couplings between positive pairs of audio and visual signals. Ideally, these couplings should connect visual objects with their references in the audio signal. Conversely, we do not want to find couplings between negative pairs of signals. To compute a single global coupling strength for a pair of signals, we aggregate this volume of pairwise similarities into a single number. There are myriad ways to aggregate this volume ranging in \"softness\" from average-pooling to max-pooling. Average pooling yields dense gradients and can improve convergence speed and stability. However, max-pooling allows the network to focus on the best couplings regardless the object's size or a sound's duration. Our aggregation function combines the benefits of average and max pooling by max-pooling visual dimensions and average pooling audio dimensions as proposed in [23]. Intuitively speaking, this averages the strongest image couplings over an audio signal. It allows small visual objects to have large effects yet provides a strong training gradient to many regions of the signals. Finally, we draw inspiration from multi-head self-attention [52] and generalize this operation to multiple \"heads\" that we max-pool before pooling the visual and audio dimensions. This allows DenseAV to discover multiple \"ways\" to associate objects across modalities.", + "bbox": [ + 75, + 325, + 470, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More formally, let $S(a,v) \\in \\mathbb{R}$ represent the similarity between a tensor of audio features $a \\in \\mathbb{R}^{CKFT}$ of size (Channel × K-heads × Frequency × Time) and a tensor of visual features $v \\in \\mathbb{R}^{CKHW}$ of size (Channel × K-heads × Height × Width). To define this scalar similarity score, we first create a local similarity volume, $s(a,v) \\in \\mathbb{R}^{kfhw}$ . For simplicity, we consider the aggregated similarity between a single image and audio clip but note one can easily general-", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ize this to max-pool over video-frames. We define the full pairwise volume of similarities as:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns (a, v) \\in \\mathbb {R} ^ {k f t h w} = \\sum_ {c = 1} ^ {C} a [ c, k, f, t ] \\cdot v [ c, k, h, w ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 125, + 890, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $a[c,k,f,t]$ represents the value of $a$ at location $[c,k,f,t]$ and $\\cdot$ is scalar multiplication. We aggregate this similarity volume into a single score $S(a,v)\\in \\mathbb{R}$ :", + "bbox": [ + 498, + 172, + 890, + 219 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (a, v) = \\frac {1}{F T} \\sum_ {f = 1} ^ {F} \\sum_ {t = 1} ^ {T} \\max _ {k, h, w} (s (a, v) [ k, f, t, h, w ]) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 222, + 890, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note that this operation can be viewed as a multi-head generalization of the MISA loss of [23], and a multi-head multi-time generalization of the MIL loss of [2].", + "bbox": [ + 498, + 272, + 890, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Loss", + "text_level": 1, + "bbox": [ + 500, + 325, + 571, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can use the similarity between audio and visual signals defined in Equation 2 to construct a contrastive loss. We follow recent works [15, 17, 54] and use the temperature-weighted InfoNCE [36] to encourage similarity between positive pairs of signals and dissimilarity between negative pairs. In DenseAV, we form $B$ positive pairs by splitting the audio and visual components of a Batch of training data. We form $B^2 - B$ negative pairs by comparing a signal to all of the other signals in the training batch. More formally let $(a_b, v_b)_1^B$ be $B$ pairs of audio and visual signals. The visual-retrieval term of our InfoNCE loss is then:", + "bbox": [ + 496, + 348, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A \\rightarrow V} = \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {\\exp \\left(\\gamma \\mathcal {S} \\left(a _ {b} , v _ {b}\\right)\\right)}{\\sum_ {b ^ {\\prime} = 1} ^ {B} \\exp \\left(\\gamma \\mathcal {S} \\left(a _ {b} , v _ {b ^ {\\prime}}\\right)\\right)}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 518, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\gamma \\in \\mathbb{R}^{+}$ is a trainable inverse temperature parameter. We symmetrize this loss by adding the analogous audio-retrieval term, $\\mathcal{L}_{V\\to A}$ , which iterates over negative audio signals in the denominator.", + "bbox": [ + 498, + 566, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Audio and Visual Featurizers", + "text_level": 1, + "bbox": [ + 500, + 635, + 759, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core of DenseAV is two modality-specific backbone networks. We use the DINO vision transformer [5] with ImageNet pretrained weights (without labels) to provide a strong, yet fully unsupervised, vision backbone. Unlike other approaches that use CLIP [46] as a backbone, DINO does not require paired text captions and learns from unlabeled images only. Practically, we find that DINO outperforms CLIP because of its better-behaved local tokens [12], an effect we explore in the Supplement. We append an additional layer norm operation across the channel dimension [3] and a $1 \\times 1$ Convolution to DINO. The layer-norm and $1 \\times 1$ convolution ensure the architecture does not start with a saturated loss function. We use the HuBERT audio transformer [25] as DenseAV's audio backbone. HubERT operates on waveforms and is trained on the LibriSpeech [40] dataset using only self-supervision. Hubert", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13120", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "outputs a single feature per frequency, corresponding to $F = 1$ in Section 3. Though HuBERT was only trained on speech, its audio features can be fine-tuned for more general sounds, much like how vision backbones can be fine-tuned for new datasets [56]. As in the visual branch, we append a channel-wise LayerNorm block and two $3 \\times 3$ convolutions to the audio branch. These layers help the network avoid saturation and speed convergence. Furthermore, the two convolutions help the model aggregate information, which reduces the cost of the pairwise feature comparison used in our loss function. We refer to these added layers after the pretrained backbones as the \"aligners\" in later sections.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Regularizers", + "text_level": 1, + "bbox": [ + 76, + 280, + 210, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Disentanglement Regularizer, $\\mathcal{L}_{\\text{Dis}}$ : We add a small regularization term to encourage each head of Equation 1 to specialize and learn independent types of audio-visual associations. Interestingly we find that our 2-head model naturally learns to distinguish the meaning of words with one head and capture the sounds objects produce with another head. To further encourage this unsupervised discovery of concepts, we penalize the network when multiple attention heads are simultaneously active. More precisely, let $(a_{b}, v_{b})_{1}^{B}$ be a Batch of $B$ paired audio and visual signals. Our disentanglement loss for two heads is then:", + "bbox": [ + 75, + 301, + 468, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D i s} = \\operatorname {M e a n} \\left(\\left| s \\left(a _ {b}, v _ {b}\\right) [ 1 ] \\circ s \\left(a _ {b}, v _ {b}\\right) [ 2 ] \\right|\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 474, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\circ$ is elementwise multiplication and $|\\cdot|$ is the elementwise absolute value function. $[k]$ mirrors PyTorch slicing notation and refers to selecting the activations for only the $k$ th attention head. Intuitively, this loss encourages one head to be silent if the other head is active and is a \"cross-term\" generalization of the $l^2$ regularizer [24] for encouraging activation shrinkage. When $K > 2$ we average contributions from every combination of heads. We ablate this, and our decision to max-pool heads in Table 3.", + "bbox": [ + 75, + 494, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stability Regularizers, $\\mathcal{L}_{\\text {Stability }}$ : Finally, we add several other small regularization terms to encourage stable convergence. We detail and ablate these terms in the Supplement. Briefly, these terms include standard regularizers like Total Variation [48] smoothness over time and non-negative pressure to encourage the network to focus on similarity instead of dissimilarity. In addition, we add a regularizer to prevent the calibration temperature, $\\gamma$ , from drifting too quickly, and a regularizer to discourage activations during silence and noise. In the supplement we show that each regularizer alone does not have a dramatic effect on final metrics but together they can stop collapses during training.", + "bbox": [ + 75, + 631, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining these losses into a single loss function yields:", + "bbox": [ + 96, + 811, + 467, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {A \\rightarrow V} + \\mathcal {L} _ {V \\rightarrow A} + \\lambda_ {\\text {D i s}} \\mathcal {L} _ {\\text {D i s}} + \\mathcal {L} _ {\\text {S t a b i l i t y}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 832, + 468, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments we use $\\lambda_{Dis} = 0.05$ and refer interested readers to the supplement for the details of our small stability regularizer, $\\mathcal{L}_{\\text{Stability}}$ .", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training", + "text_level": 1, + "bbox": [ + 500, + 90, + 604, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments we train DenseAV and relevant baselines on the AudioSet [16] dataset for sound prompted segmentation and AudioSet retrieval. We train on the PlacesAudio [22] dataset for speech prompted segmentation, PlacesAudio retrieval, and the ablation studies of Table 4. In our disentanglement experiments of Table 3 and feature visualizations of Figures 1 and 2 we train on both AudioSet and PlacesAudio so that DenseAV can be familiar with both language, the prominent audio signal in PlacesAudio, and more general sounds from AudioSet. In these experiments we sample training data from these two corpora, so each batch has an even split between AudioSet and PlacesAudio.", + "bbox": [ + 496, + 114, + 890, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Warming up Aligners: We find that we can dramatically improve the stability by first training the added aligners (convolutions and layer norms) for 3000 steps while keeping pretrained DINO and HuBERT backbones fixed. This allows the aligners to adapt to these intelligent backbones before modifying each backbone's sensitive weights. We use random resize crops, color jitter, random flips, and random greyscale as image augmentations. We randomly sample a single video frame to feed to our visual branch. Audio clips are converted to single-channel format and are trimmed or padded with silence to create uniform 10 second clips. We re-sample audio clips according to the requirements of the backbone models used. For HuBERT, we re-sample to 16KhZ. We train on 8 V100 GPUs with an effective batch size of 80, and aggregate negative samples on all GPUs prior to computing the loss to ensure efficient parallelization. We provide additional training information and hyperparameters in the supplement.", + "bbox": [ + 496, + 296, + 892, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Training: After warming up the aligners, we train the full model for an additional 800,000 steps using the same loss, batch-size, and training logic. We train all aligner weights and fine-tune all HuBERT audio backbone weights. We use low rank adaptation (LoRA) [26] to fine-tune the \"Q\", \"K\", and \"V\" layers of the DINO visual backbone attention blocks. This allows us to efficiently adapt DINO and stabilize the training as it is quite easy to collapse the carefully trained DINO weights. We use a LoRA rank of 8.", + "bbox": [ + 496, + 569, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 722, + 633, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate AV representation quality, we perform a variety of analyses including comparative activation visualization, quantitative measurements of speech and sound prompted semantic segmentation, and cross-modal retrieval. Additionally, we quantify our observation that DenseAV can distinguish the meanings of words (language), from the sounds of objects (sound) without supervision.", + "bbox": [ + 496, + 748, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To adequately measure a representation's AV alignment quality, we found it necessary to introduce two evaluation datasets that measure speech and sound prompted seman", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13121", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/111fd3a0658521944c7dfd36b0db85af127eddf0a4f27519c9e3f261def3281f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Speech and Sound prompted semantic segmentation. We analyze the quality of local features using two prompted semantic segmentation tasks. We prompt networks with speech of the form \"a picture of a(n) [Object]\" to determine whether local feature inner products can segment objects in the ADE20K dataset by name. We create sound prompts for a given ADE20K class using a curated mapping from the ADE20K ontology to the VG-GSound ontology. DenseAV's local features perform significantly better than all baselines investigated. We bold \"first place\" results and underline \"second place\" results." + ], + "table_body": "
MethodSpeech Semseg.Sound Semseg.
mAPmIoUmAPmIoU
DAVENet [23]32.2%26.3%16.8%17.0%
CAVMAE [18]27.2%19.9%26.0%20.5%
ImageBind [17]20.2%19.7%18.3%18.1%
Ours48.7%36.8%32.4%25.5%
", + "bbox": [ + 78, + 66, + 464, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tic segmentation performance. Our two datasets introduce pairs of speech and sound prompts coupled with matching images and segmentation masks derived from ADE20K. We create these datasets because previous works [23] have not published their datasets or evaluation code. However, we use an experimental setting from the literature for our cross-modal retrieval experiments.", + "bbox": [ + 76, + 333, + 467, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare against a variety of prior art including the popular state-of-the-art multi-modal retrieval network, ImageBind [17]. We also compare against CAVMAE [18], a leading multimodal backbone trained specifically for AudioSet retrieval, and DAVENet [23], which is trained to localize the meanings of words. We include two other baselines [21, 22] which have reported cross modal retrieval metrics on Places Audio. Finally, we compare our multi-head aggregation strategy to common \"global\" retrieval methods such as inner products between classtokens, average-pooled tokens, and SimPooled[44] tokens. We note that SimPool achieves state-of-the-art localization results when compared to 14 other pooling methods. Nevertheless, our multi-head aligner yields better localization results than any of these \"global\" methods.", + "bbox": [ + 76, + 440, + 467, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Comparison of Feature Maps", + "text_level": 1, + "bbox": [ + 76, + 679, + 428, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our first experiment in Figure 2 highlights the dramatic differences in quality between DenseAV's features and other approaches in the literature. DenseAV is the only backbone whose local tokens are semantically meaningful and show cross-modal alignment for speech and sound. Though both CAVMAE and ImageBind show high-quality retrieval performance, neither shows high quality aligned local tokens. As a result, DenseAV can associate and localize both sound and language significantly better than other backbones. DAVENet shows coarse correspondences between language and visual objects but cannot associate sound with visual objects and does not match DenseAV's high resolution maps. Furthermore, the right half of Figure 1 demon", + "bbox": [ + 76, + 703, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bf4f4798533e7d28ef8b41b6594644822d3acb60ad6035e64e20f371a92ad8c8.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Cross-modal retrieval using 1000 evaluation videos from the PlacesAudio and AudioSet validation datasets. DenseAV dramatically outperforms all approaches tested in all metrics. Most notably, the state-of-the-art image retrieval foundation model, ImageBind, is incapable of recognizing speech. We note that the ImageBind authors do not publish retraining code, so we evaluate their largest pretrained model. Models with a * indicate that they have been previously reported in the literature. Other numbers are calculated by using pretrained models when available or from training with the author's official training scripts." + ], + "table_body": "
MethodPlaces Acc. @10AudioSet Acc. @10
I → AA → II → AA → I
[22]*46.3%54.8%--
[21]*54.2%56.4%--
DAVENet [23]*52.8%60.4%--
CAVMAE [18]81.7%77.7%55.7%50.7%
ImageBind [17]1.10%1.10%64.5%66.5%
Ours94.2%94.3%68.2%68.4%
", + "bbox": [ + 501, + 66, + 913, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "strates that DenseAV naturally discovers and separates word semantics from the sound of objects without labels to supervise this separation. In the supplement, we provide additional visualizations of all backbones considered across a wide range of words and sounds.", + "bbox": [ + 496, + 369, + 890, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Speech Prompted Image Segmentation", + "text_level": 1, + "bbox": [ + 498, + 454, + 831, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: We introduce a speech prompted segmentation dataset using the ADE20K dataset, which is known for its comprehensive ontology and pixel-precise annotations [59]. From this dataset, we curate an evaluation subset of image-class pairs by sampling up to 10 images for each object class in ADE20K, excluding images where the selected class was tiny ( $< 5\\%$ of pixels). We only consider classes with at least 2 images that pass the tiny object criterion. For each class and image, we formed a binary target mask by selecting the semantic segmentation mask for that class. This resulted in 3030 image-object pairs spanning 478 ADE20K classes.", + "bbox": [ + 496, + 477, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We created paired speech signals by speaking the prompt \"A picture of a(n) [object]\" where [object] is the name of the ADE20K class. We create clear, controlled, and consistent audio prompts using Microsoft's neural text to speech service [47]. This service also provides exact timing of the \"[object]\" utterance within the broader prompt and ensures each class is measured equally. Grammar was manually verified for the utterances to ensure proper singular/plural and a/an agreement with the class name. We release images, masks, and audio prompts for reproducibility.", + "bbox": [ + 496, + 643, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Measure: We evaluate methods based on how well their speech-prompted activations align with ground truth masks for the visual object's class. We quantify this with the binary Average Precision (AP) and Intersection over Union (IoU) metrics. These quantify how close activations match with the binary label mask from the ADE20K dataset. To compute an aggregate score over all of the object", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13122", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/126e8acdd47b6ae4004f1548bbb9f626573c0065461e079eb132026f839e9bb0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Quantitative ablation study of the impact of max-pooling attention heads and adding our disentanglement loss, $\\mathcal{L}_{Dis}$ . Intuitively, max-pooling attention heads allows each head to specialize on its own specific set of triggers. Our disentanglement loss further encourages the heads to operate independently and orthogonally." + ], + "table_body": "
MethodPred. Dis.Act. Dis.
No LDis, No Head Max Pool64.1%70.3%
No LDis99.9%86.5%
Ours99.9%91.2%
", + "bbox": [ + 81, + 65, + 460, + 130 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classes considered, we compute the mean average precision (mAP) and mean intersection over union (mIoU) by averaging AP scores across all object categories considered.", + "bbox": [ + 75, + 229, + 468, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The mAP is particularly well suited for evaluating feature similarities because it is unaffected by monotonic transformations of the similarity scores. This eliminates the need for arbitrary thresholding and calibration. This is particularly important because many networks' inner products are not centered at zero, and the best thresholding strategy can be nontrivial, and dependent on the network and object class. Average Precision avoids these confounding factors and ensures a fair comparison across methods. Unfortunately, unlike the mAP, the mIoU metric requires selecting a threshold. To ensure our mIoU measurement is similarly invariant to monotonic transformations we evaluate 20 uniformly spaced thresholds between the smallest and largest activations of each model. For each baseline, we report results for the best threshold to ensure a fair comparison between all networks considered.", + "bbox": [ + 75, + 275, + 468, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation: We compute image heatmaps by evaluating each modality-specific network on the image-audio pairs from our dataset. We extract dense features from the final layer of each network and form their similarity volume according to Equation 1. For DenseAV we max-pool the head dimension to properly compare with single-headed models. We average activations over the temporal extent of the \"[object]\" utterance using the word timing information from the ground truth audio clip. This creates a heatmap over the image features that can be bi-linearly resized to the original image's size. We then compare these per-pixel activation scores to ground truth object masks from our dataset.", + "bbox": [ + 75, + 517, + 468, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results: In Speech mAP and mIoU columns of Table 1 we show that DenseAV achieves a $51\\%$ $(+16.5\\mathrm{mAP})$ relative increase in speech-prompted semantic segmentation over previous methods. Approaches that use global token based contrastive strategies such as CAVMAE and Image-Bind perform particularly poorly in this task, and this observation aligns with the qualitative results of Figure 2.", + "bbox": [ + 75, + 698, + 468, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Sound Prompted Image Segmentation", + "text_level": 1, + "bbox": [ + 76, + 816, + 405, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset: To evaluate how well deep features localize sound, we build on Section 4.2 and create a dataset of sound prompts that align with ADE20K classes. We first select the same (large) image-object pairs from ADE20K. We then", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/68af70bc9db37516f62cc94dd99261f502ef13c7be73b11b9d79a5e99b965220.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Quantitative ablation of different feature aggregation strategies. Though the common practice of average pooling and using a learned CLS token to aggregate features have little effect on retrieval performance, they dramatically degrade performance on speech prompted semantic segmentation." + ], + "table_body": "
MethodSpeech mAPPlaces Acc. @10
V → AA → V
Average Pool20.1%92.0%91.2%
CLS Token20.6%86.4%89.8%
SimPool [44]35.3%92.6%92.8%
Multi-Head (Ours)48.2%93.5%93.8%
", + "bbox": [ + 501, + 65, + 893, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "create a mapping between the ADE20K and VGGSound [6] ontologies. To compute a robust mapping, we first embed ADE20K class names and VGGSound class names with the GPT Ada 2 text embedding model [37]. For each ADE20K class, we create a list of at most three candidates from the VGGSound ontology that have a cosine similarity $(> .85)$ . We then manually review these candidates to select the best VGGSound class for each ADE20K class and remove any spurious or mistaken matches. This produces a set of 95 ADE20K classes with strong matches in the VGGSound ontology. For each of our original 3030 image-object pairs we select a random VGGSound validation clip with a matching class according to our mapped ontology. This yields 106 image-object pairs across 20 ADE20K classes.", + "bbox": [ + 496, + 260, + 890, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Measure: We use the same mAP and mIoU evaluation metrics as Section 4.2, but instead average over the 20 ADE20K classes considered.", + "bbox": [ + 498, + 472, + 890, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation: We compute sound prompted image activations as in section 4.2 but with one key change: we average activations over the entire clip because we do not have ground-truth sound timing information.", + "bbox": [ + 498, + 518, + 890, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results: The \"Sound mAP and mIoU\" columns of Table 1 show that DenseAV achieves a $25\\%$ (+6.4mAP) relative improvement in sound prompted segmentation compared to the prior art. Most notably, ImageBind's features cannot localize sound despite their high cross-modal retrieval performance learned from millions of hours of sound.", + "bbox": [ + 498, + 579, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Cross-Modal Retrieval", + "text_level": 1, + "bbox": [ + 500, + 680, + 710, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show that DenseAV's representations are not only better for localization, but significantly outperform other approaches on cross-modal retrieval. We adopt the evaluation setting of [23] and measure cross modal retrieval accuracy at 1, 5, and 10 in a thousand-way retrieval task. In particular, we use the same thousand images from the validation set of [23] and also replicate this analysis on one-thousand random clips from the AudioSet validation data. Table 2 shows results for 1000-way retrieval tasks on both the Places Audio and AudioSet datasets. We show cross-modal accuracy at 10, but also show larger tables in the supplement that echo these results using accuracy at 1 and 5. DenseAV significantly outperforms all baselines across all metrics. Interest-", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13123", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ingly, DenseAV outperforms ImageBind with less than half of the trainable parameters and no reliance on text.", + "bbox": [ + 75, + 90, + 472, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Measuring Disentanglement", + "text_level": 1, + "bbox": [ + 76, + 130, + 331, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that DenseAV's heads naturally learn to differentiate audio-visual couplings that capture the meaning of words (language) and those that capture the sounds of objects (sound). Furthermore this effect generalizes to novel clips, including those with both sound and language as shown in Figure 1. We quantify this observation in two ways, the first measures if a head's average activation strength predicts whether a clip contains mainly \"language\" or \"sound\". The second method quantifies how often the \"sound\" head is incorrectly active when the \"language\" head should be active and vice versa. We leverage the fact that AudioSet dataset contains mostly clips with ambient sound and rarely contains language. In contrast, Places Audio is entirely language-based without external ambient sound. We note that these analyses are specifically for our architecture with two heads $K = 2$ and trained on both AudioSet and PlacesAudio data.", + "bbox": [ + 75, + 152, + 468, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For both measures of disentanglement, we first compute a clip's aggregated similarity for each head. In particular, we remove the max-pooling over heads in Equation 2 to create a single-head similarity, $S(a,v)_k$ . We then min-max scale the scores of each head across both datasets to lie in the [0,1] interval, which we refer to as $\\hat{S}(a,v)_k$ . Using these normalized scores, we can create metrics that capture how well a given head responds only to a specific dataset.", + "bbox": [ + 75, + 410, + 468, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our first metric measures how well a head's scores predict whether a clip is from the \"sound\" or \"language\" dataset. Let $(a_{b},v_{b})_{1}^{B}$ be tuples of paired audio and visual signals. Let $l[k^{\\prime}]_{b}$ be an indicator variable of whether the signal $(a_{b},v_{b})$ arises from the sound dataset, AudioSet, $(k^{\\prime} = 1)$ , or the language dataset Places Audio $(k^{\\prime} = 2)$ .", + "bbox": [ + 76, + 531, + 468, + 623 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {p r e d} (k, k ^ {\\prime}) = \\mathrm {A P} \\left(\\left(\\hat {\\mathcal {S}} \\left(a _ {b}, v _ {b}\\right) _ {k}\\right) _ {1} ^ {B}, \\left(l \\left[ k ^ {\\prime} \\right] _ {b}\\right) _ {1} ^ {B}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 633, + 468, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Where $AP(\\cdot, \\cdot)$ is the binary average precision with prediction and label arguments respectively. Intuitively, this measures whether the scores of head $k$ are direct predictors of whether the data is from dataset $k'$ . We can find the best assignment between heads and datasets such that each head is maximally predictive of the given dataset:", + "bbox": [ + 75, + 667, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {P r e d D i s} = \\frac {1}{2} \\max \\left(\\delta_ {p r e d} (0, 0) + \\delta_ {p r e d} (1, 1), \\right. \\\\ \\delta_ {p r e d} (1, 0) + \\delta_ {p r e d} (0, 1)) \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 766, + 468, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The prediction disentanglement score, PredDis, is a percentage that ranges from $50\\%$ for completely entangled signals to $100\\%$ if one can perfectly classify the signals using the scores of either head. The maximum over the two possible assignments makes this metric invariant to permutations", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the heads. We note that this metric is a Hungarian matching assignment [29] over two entries, a common technique to asses unsupervised classification performance [20, 28].", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our second measure quantifies \"spurious activations\" in the non-dominant head. A truly disentangled system should have a head that only fires on sound, and another head that only fires on language. We create another disentanglement measure, ActDis, by replacing $\\delta_{pred}$ in Equation 7 with:", + "bbox": [ + 496, + 136, + 890, + 212 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {a c t} \\left(k, k ^ {\\prime}\\right) = 1 - \\frac {1}{\\sum_ {b ^ {\\prime}} l \\left[ k ^ {\\prime} \\right] _ {b ^ {\\prime}}} \\sum_ {b = 1} ^ {B} \\hat {\\mathcal {S}} \\left(a _ {b}, v _ {b}\\right) _ {k} \\cdot l \\left[ k ^ {\\prime} \\right] _ {b} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 214, + 890, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, this measures the \"inactivity\" of head $k$ on dataset $k'$ . If head $k$ is totally silent on dataset $k'$ then $\\delta_{act}(k, k') = 1$ . Like PredDis, ActDis is a percentage ranging from $50\\%$ to $100\\%$ with $100\\%$ representing perfect disentanglement where the sound head is completely silent during the language clips, and vice versa.", + "bbox": [ + 496, + 261, + 890, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows that DenseAV achieves near perfect predictive (99%) and activation (91%) disentanglement. It also shows that our disentanglement regularizer and max-pooling over heads improves DenseAV's natural ability to distinguish sound from language without supervision.", + "bbox": [ + 496, + 352, + 890, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 439, + 617, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented DenseAV, a novel contrastive learning architecture that can discover the meaning of words and localize the sounds of objects using only video supervision. We are the first to observe both qualitatively and quantitatively that it's possible to disentangle the meaning of words from the sound of objects with only a contrastive learning signal. DenseAV's success stems from its novel multi-head attention aggregation mechanism that encourages its modality-specific backbones to create high-resolution, semantically meaningful, and AV aligned representations. These properties of DenseAV's representation are not seen in other state-of-the-art models in the literature. Consequently, DenseAV significantly surpasses other leading models in dense prediction tasks such as speech and sound-prompted semantic segmentation as well as in cross-modal retrieval.", + "bbox": [ + 496, + 465, + 890, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 703, + 666, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation. This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/). This work is funded by a Royal Society Research Professorship RSRP\\R\\241003, and EPSRC Programme Grant VisualAI EP/T028572/1.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13124", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020. 3", + "[2] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 2, 3, 4", + "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4", + "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 4", + "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 7", + "[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 3", + "[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 3", + "[9] Jeongsoo Choi, Joanna Hong, and Yong Man Ro. Diffv2s: Diffusion-based video-to-speech synthesis with vision-guided speaker embedding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7812-7821, 2023. 3", + "[10] Noam Chomsky. Language and problems of knowledge: The Managua lectures. MIT press, 1987. 1", + "[11] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 3", + "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023.4", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2", + "[15] Nicholas Frosst, Nicolas Papernot, and Geoffrey Hinton. Analyzing and improving representations with the soft nearest neighbor loss. In International conference on machine learning, pages 2012-2020. PMLR, 2019. 4", + "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5", + "[17] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3,4,6", + "[18] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 6", + "[19] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 4", + "[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 2, 3, 8", + "[21] David Harwath and James R Glass. Learning word-like units from joint audio-visual analysis. arXiv preprint arXiv:1701.07481, 2017. 3, 6", + "[22] David Harwath, Antonio Torralba, and James Glass. Unsupervised learning of spoken language with visual context. Advances in Neural Information Processing Systems, 29, 2016. 5, 6", + "[23] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649-665, 2018. 2, 3, 4, 6, 7", + "[24] Arthur E Hoerl and Robert W Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55-67, 1970. 5", + "[25] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 4", + "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13125", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5", + "[27] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 3", + "[28] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9865-9874, 2019. 3, 8", + "[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 8", + "[30] Neeraj Kumar, Srishti Goel, Ankur Narang, and Mujtaba Hasan. Robust one shot audio to video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 770-771, 2020. 3", + "[31] Shiguang Liu, Sijia Li, and Haonan Cheng. Towards an end-to-end visual-to-raw-audio generation with gan. IEEE Transactions on Circuits and Systems for Video Technology, 32(3): 1299–1312, 2021. 3", + "[32] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song. Active contrastive learning of audio-visual video representations. arXiv preprint arXiv:2009.09805, 2020. 3", + "[33] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3", + "[34] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2", + "[35] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871–14881, 2021. 3", + "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4", + "[37] OpenAI. Gpt-4 technical report, 2023. 7", + "[38] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Is object localization for free?-weakly-supervised learning with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 685-694, 2015. 2", + "[39] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4", + "[41] Alex S Park and James R Glass. Unsupervised pattern discovery in speech. IEEE Transactions on Audio, Speech, and Language Processing, 16(1):186-197, 2007. 3", + "[42] Puyuan Peng and David Harwath. Self-supervised representation learning for speech using visual grounding and masked language modeling. arXiv preprint arXiv:2202.03543, 2022. 3", + "[43] Puyuan Peng and David Harwath. Word discovery in visually grounded, self-supervised speech models. arXiv preprint arXiv:2203.15081, 2022. 3", + "[44] Bill Psomas, Ioannis Kakogeorgiou, Konstantinos Karantzalos, and Yannis Avrithis. Keep it simpool: Who said supervised transformers suffer from attention deficit? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5350-5360, 2023. 4, 6, 7", + "[45] Geoffrey K Pullum and Barbara C Scholz. Empirical assessment of stimulus poverty arguments. The linguistic review, 19(1-2):9-50, 2002. 1", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[47] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv preprint arXiv:2006.04558, 2020. 6", + "[48] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5", + "[49] Yi-Jen Shih, Hsuan-Fu Wang, Heng-Jui Chang, Layne Berry, Hung-yi Lee, and David Harwath. Speechclip: Integrating speech with pre-trained vision and language model. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 715-722. IEEE, 2023. 2, 3, 4", + "[50] Linda Smith and Chen Yu. Infants rapidly learn word-referent mappings via cross-situational statistics. Cognition, 106(3):1558-1568, 2008. 1", + "[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 3", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 4", + "[53] Luyu Wang and Aaron van den Oord. Multi-format contrastive learning of audio representations. arXiv preprint arXiv:2103.06508, 2021. 3", + "[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13126", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 4", + "[55] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 2", + "[56] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5", + "[57] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European conference on computer vision (ECCV), pages 570-586, 2018. 3", + "[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2", + "[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6", + "[60] Hao Zhu, Man-Di Luo, Rui Wang, Ai-Hua Zheng, and Ran He. Deep audio-visual learning: A survey. International Journal of Automation and Computing, 18:351–376, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 445 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13127", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_model.json b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1206c35a531aa8b7bc523928f0844198a98b7082 --- /dev/null +++ b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.78, + 0.177 + ], + "angle": 0, + "content": "Separating the \"Chirp\" from the \"Chat\": Self-supervised Visual Grounding of Sound and Language" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.204, + 0.262, + 0.256 + ], + "angle": 0, + "content": "Mark Hamilton \nMIT, Microsoft \nmarkth@mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.204, + 0.457, + 0.24 + ], + "angle": 0, + "content": "Andrew Zisserman Oxford, Google" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.204, + 0.632, + 0.24 + ], + "angle": 0, + "content": "John R. Hershey Google" + }, + { + "type": "text", + "bbox": [ + 0.677, + 0.204, + 0.836, + 0.24 + ], + "angle": 0, + "content": "William T. Freeman \nMIT, Google" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.279, + 0.896, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.478, + 0.894, + 0.52 + ], + "angle": 0, + "content": "Figure 1. Visual overview of the DenseAV algorithm. Two modality-specific backbones featurize audio and visual signals. We introduce a novel generalization of multi-head attention to extract attention maps that discover and separate the \"meaning\" of spoken words and the sounds an object makes. DenseAV performs this localization and decomposition solely through observing paired stimuli such as videos." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.533, + 0.314, + 0.548 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.474, + 0.884 + ], + "angle": 0, + "content": "We present DenseAV, a novel dual encoder grounding architecture that learns high-resolution, semantically meaningful, and audio-visual aligned features solely through watching videos. We show that DenseAV can discover the \"meaning\" of words and the \"location\" of sounds without explicit localization supervision. Furthermore, it automatically discovers and distinguishes between these two types of associations without supervision. We show that DenseAV's localization abilities arise from a new multi-head feature aggregation operator that directly compares dense image and audio representations for contrastive learning. In contrast, many other systems that learn \"global\" audio and video representations cannot localize words and sound. Finally, we contribute two new datasets to improve the evaluation of AV representations through speech and sound prompted semantic segmentation. On these and other datasets we show DenseAV dramatically outperforms the prior art on speech and sound prompted semantic segmentation. DenseAV outperforms the current state-of-the-art, ImageBind, on cross-modal retrieval using fewer than half of the parameters. Project Page: https://aka.ms/denseav" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.533, + 0.632, + 0.548 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Associating audio and video is a fundamental aspect of human perception. As infants develop, the synchronization and correspondence of sounds enables multi-modal association – a voice with a face, and a 'moo' with a cow [50]. Later, as they acquire language, they associate spoken words with objects they represent [10, 45]. Amazingly, these association abilities, constituting speech recognition, sound event recognition, and visual object recognition, develop without much direct supervision. This work aims to create a model with this capability by learning high-resolution, semantically meaningful, audio-visual (AV) aligned representations. Features with these properties can be used to discover fine-grained correspondences between modalities without localization supervision or knowledge of the textual representation of language." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As an example, consider the spoken caption and accompanying sounds of the image shown in Figure 1. We wish to \"ground\" both the speech and the sounds with high resolution. For instance, both the spoken word \"dog\" and the sound of a bark in the audio signal should be associated with the pixels of the dog in the visual signal if present. We seek high quality local representations where simple inner" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13117" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.13, + 0.238, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.071, + 0.892, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.333, + 0.893, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparison of several modern architectures for associating audio and video modalities. Only DenseAV learns a high-resolution and semantically aligned set of local features. This allows us to perform speech and sound prompted semantic segmentation using only the inner products between deep features. Other approaches, such as ImageBind, do not show aligned local feature maps. Approaches that do show some localization capabilities, like DAVENet, do not generalize to sound and language, and do not achieve the high-resolution localization capabilities of DenseAV. Dense features are visualized using PCA as in [20]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.469, + 0.451 + ], + "angle": 0, + "content": "products between features exhibit this behavior, which is interestingly absent from popular approaches in the literature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.744 + ], + "angle": 0, + "content": "To achieve this, we make three innovations. First, we introduce DenseAV, a dual-encoder architecture that computes a dense similarity volume over audio and visual features before aggregating this volume into a single similarity score. If we look at a slice of this similarity volume when word is spoken, as in Figure 1, we can visualize the AV activation strength between a word or sound and an image's pixels. The novelty we introduce is to generalize this dense similarity mechanism with multiple heads, much like those of multi-head attention. This generalization allows each head to specialize on particular types of couplings between the visual and audio modalities. Interestingly, we discover that if we give DenseAV two heads and train on a dataset that contains both language and sound, the heads naturally learn to distinguish language from more general sound using only cross-modality supervision. For example, as shown in Figure 1, head 1 focuses on what objects create a sound, like the barking of a dog, whereas head 2 focuses only on the meaning of words." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Second, we show the importance of the \"aggregation function\" one uses to create a similarity score between an audio clip and a video frame for contrastive learning. The traditional choices; using inner products between global representations such as class tokens [5, 13, 49] or pooled features [18, 58], do not promote AV alignment of dense local features. Because of this, several popular audio-video backbones that excel on cross-modal retrieval cannot directly associate objects and sounds using their local features. This limits their ability to be used for downstream" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.421, + 0.892, + 0.451 + ], + "angle": 0, + "content": "tasks such as semantic segmentation, sound localization, or unsupervised language learning and discovery." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.667 + ], + "angle": 0, + "content": "Third, we introduce two semantic segmentation datasets to evaluate visual grounding with AV representations for speech and (non-speech) sounds. We build these datasets from the high-quality segmentation masks provided by the ADE20K dataset [59] and measure mean average precision (mAP) and mean intersection over union (mIoU) on a binary mask prediction task. This evaluation is simpler and more thorough than previous efforts to measure visual grounding such as the concept counting metrics of [23] and the \"pointing games\" of [2, 14, 38] that only check if a heatmap's peak occurs within a target box or segment. Furthermore, our evaluation avoids brittle word-net ontologies [34], clustering, Wu and Palmer distance [55], threshold choices, and a variety of other complicating factors." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.67, + 0.871, + 0.684 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.696, + 0.89, + 0.725 + ], + "angle": 0, + "content": "- We introduce DenseAV, a novel self-supervised architecture that learns high-resolution AV correspondences." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.726, + 0.891, + 0.785 + ], + "angle": 0, + "content": "- We introduce a local-feature-based image similarity function that significantly improves a network's zero-shot localization ability compared to common strategies such as average pooling or CLS tokens." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.787, + 0.892, + 0.845 + ], + "angle": 0, + "content": "- We introduce new datasets for evaluating speech and sound prompted semantic segmentation. We show DenseAV significantly outperforms the current state-of-the-art on these tasks as well as on cross-modal retrieval." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.847, + 0.891, + 0.892 + ], + "angle": 0, + "content": "- We discover that our multi-head architecture naturally disentangles audio-visual correspondence into sound and language components using only contrastive supervision." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.696, + 0.892, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13118" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.071, + 0.852, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 3. Architectural overview of our multi-head attention aggregator. Dense feature maps are split into \\( K \\) heads (\\( K = 1,2 \\)) in our experiments. We form an AV activation tensor by taking the inner-products of each head's features across the spatial and temporal extent of the visual and audio signals respectively as in Equation 1. We then aggregate this similarity volume into a single similarity score by max-pooling head and spatial dimensions and average-pooling audio dimensions. Our approach aims to encourage the network to identify specific shared objects between the audio and visual modalities. In particular, max-pooling of heads disentangles sound and language, and max-pooling spatial dimensions helps localize objects." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.377, + 0.22, + 0.394 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.473, + 0.789 + ], + "angle": 0, + "content": "Audio-visual (AV), text-visual, and more general multimodal models have recently surged in popularity [60]. Broadly speaking DenseAV is an audio-video contrastive learning architecture, this class of methods learns AV representations by aligning paired signals and pushing apart negative signals [11, 27]. Of the models in this class, several stand out for their ability to localize sounds [2, 7, 42] or capture the semantics of language [23, 43]. Many models in this class compare AV signals using inner products between \"global\" representations formed by pooled deep features [18, 35, 53], or class tokens [17, 32, 42, 43, 49]. Most notably, ImageBind has gained popularity due to its state-of-the-art performance on a variety of tasks and datasets and unified class-token-based contrastive architecture. In this work we show that many of these architectures do not show strong localization properties in their local features, despite excelling at cross-modal retrieval on a \"global\" level. This limits their applicability to new out-of-domain sounds, sounds that don't have a textual representation, and low-resource languages. We diverge from these works by directly supervising local tokens. In particular, we build on previous works [2, 23] that show max-pooling improves localization capabilities and introduce a new multi-head aggregation operator that generalizes previous losses using an operator similar to self-attention [52]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Another class of methods discover structure in signals through uni- and multi-modal clustering. Early works on audio clustering [41] discovered meaningful utterances without supervision. Similar visual analyses have discovered visual objects [4, 8, 20, 28]. Recent works have applied these ideas to the AV domain [1, 21], but do not focus on extracting high-resolution AV representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.895, + 0.503 + ], + "angle": 0, + "content": "Finally, several works investigate generative audio-video learning. The Sound of Pixels [57] generates the sound of a specific object using a source separation loss. Newer approaches using GANs [30, 31], and diffusion models [9, 17, 33] have generated audio from video and vice versa. Here we focus on improving the local representations of contrastive learners because of their relative scalability, simplicity, and ability to learn high-quality representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.6, + 0.53 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.689 + ], + "angle": 0, + "content": "At a high level, DenseAV tries to determine when a given audio and visual signal belong “together” using dense audio-visual representations. To perform this task robustly, DenseAV must learn how to predict the contents of an audio signal from a visual signal and vice versa. This causes DenseAV to learn dense modality-specific features that capture the mutual information shared between the modalities [51]. Once learned, we can directly query these informative features to perform speech and sound prompted semantic segmentation as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.902 + ], + "angle": 0, + "content": "More specifically, DenseAV is built from two modality-specific deep featurizers. These backbones produce temporally varying audio features across an audio clip and spatially varying video features for a single randomly selected frame. Our loss computes a similarity between audio and visual signals based on the intuition that two signals are similar if they have a variety of strong couplings or shared objects. More formally, we form a scalar similarity for a pair of audio and video signals by carefully aggregating a volume of pairwise inner products between dense features. We use the InfoNCE [36] contrastive loss to encourage similarity between \"positive\" pairs of signals and dissimilarity between \"negative\" pairs formed by in-batch shuffling. Figure 3 graphically depicts this loss function and subsequent" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13119" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.414, + 0.106 + ], + "angle": 0, + "content": "sections detail each component of our architecture." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.119, + 0.433, + 0.135 + ], + "angle": 0, + "content": "3.1. Multi-Headed Aggregation of Similarities" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.144, + 0.47, + 0.325 + ], + "angle": 0, + "content": "DenseAV's key architectural distinction is its loss function that directly supervises the \"local\" tokens of the visual and audio featurizers. This is a significant departure from other works [5, 17, 19, 39, 46, 49] that pool modality specific information into \"global\" representations prior to the contrastive loss. Unlike prior works, our loss function aggregates the full pairwise similarities between the local tokens into an aggregate measure of similarity for a given pair of audio and visual signals. We show in Figure 2 that this architectural choice enables DenseAV's local features to align across modalities whereas other approaches such as average pooling, class tokens, and SimPool [44] do not." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.471, + 0.78 + ], + "angle": 0, + "content": "We first describe our loss function informally and definite it more precisely in the next paragraph. Our loss function computes the (un-normalized) inner product between every pair of visual and audio features to form a \"volume\" of inner products. This volume represents how strongly each part of an audio signal \"couples\" to each part of a visual signal. We aim to find many large couplings between positive pairs of audio and visual signals. Ideally, these couplings should connect visual objects with their references in the audio signal. Conversely, we do not want to find couplings between negative pairs of signals. To compute a single global coupling strength for a pair of signals, we aggregate this volume of pairwise similarities into a single number. There are myriad ways to aggregate this volume ranging in \"softness\" from average-pooling to max-pooling. Average pooling yields dense gradients and can improve convergence speed and stability. However, max-pooling allows the network to focus on the best couplings regardless the object's size or a sound's duration. Our aggregation function combines the benefits of average and max pooling by max-pooling visual dimensions and average pooling audio dimensions as proposed in [23]. Intuitively speaking, this averages the strongest image couplings over an audio signal. It allows small visual objects to have large effects yet provides a strong training gradient to many regions of the signals. Finally, we draw inspiration from multi-head self-attention [52] and generalize this operation to multiple \"heads\" that we max-pool before pooling the visual and audio dimensions. This allows DenseAV to discover multiple \"ways\" to associate objects across modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "More formally, let \\( S(a,v) \\in \\mathbb{R} \\) represent the similarity between a tensor of audio features \\( a \\in \\mathbb{R}^{CKFT} \\) of size (Channel × K-heads × Frequency × Time) and a tensor of visual features \\( v \\in \\mathbb{R}^{CKHW} \\) of size (Channel × K-heads × Height × Width). To define this scalar similarity score, we first create a local similarity volume, \\( s(a,v) \\in \\mathbb{R}^{kfhw} \\). For simplicity, we consider the aggregated similarity between a single image and audio clip but note one can easily general-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ize this to max-pool over video-frames. We define the full pairwise volume of similarities as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.126, + 0.892, + 0.166 + ], + "angle": 0, + "content": "\\[\ns (a, v) \\in \\mathbb {R} ^ {k f t h w} = \\sum_ {c = 1} ^ {C} a [ c, k, f, t ] \\cdot v [ c, k, h, w ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.174, + 0.892, + 0.22 + ], + "angle": 0, + "content": "Where \\(a[c,k,f,t]\\) represents the value of \\(a\\) at location \\([c,k,f,t]\\) and \\(\\cdot\\) is scalar multiplication. We aggregate this similarity volume into a single score \\(S(a,v)\\in \\mathbb{R}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.223, + 0.892, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (a, v) = \\frac {1}{F T} \\sum_ {f = 1} ^ {F} \\sum_ {t = 1} ^ {T} \\max _ {k, h, w} (s (a, v) [ k, f, t, h, w ]) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.318 + ], + "angle": 0, + "content": "We note that this operation can be viewed as a multi-head generalization of the MISA loss of [23], and a multi-head multi-time generalization of the MIL loss of [2]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.326, + 0.573, + 0.34 + ], + "angle": 0, + "content": "3.2. Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.514 + ], + "angle": 0, + "content": "We can use the similarity between audio and visual signals defined in Equation 2 to construct a contrastive loss. We follow recent works [15, 17, 54] and use the temperature-weighted InfoNCE [36] to encourage similarity between positive pairs of signals and dissimilarity between negative pairs. In DenseAV, we form \\( B \\) positive pairs by splitting the audio and visual components of a Batch of training data. We form \\( B^2 - B \\) negative pairs by comparing a signal to all of the other signals in the training batch. More formally let \\( (a_b, v_b)_1^B \\) be \\( B \\) pairs of audio and visual signals. The visual-retrieval term of our InfoNCE loss is then:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.519, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A \\rightarrow V} = \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {\\exp \\left(\\gamma \\mathcal {S} \\left(a _ {b} , v _ {b}\\right)\\right)}{\\sum_ {b ^ {\\prime} = 1} ^ {B} \\exp \\left(\\gamma \\mathcal {S} \\left(a _ {b} , v _ {b ^ {\\prime}}\\right)\\right)}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.567, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Where \\(\\gamma \\in \\mathbb{R}^{+}\\) is a trainable inverse temperature parameter. We symmetrize this loss by adding the analogous audio-retrieval term, \\(\\mathcal{L}_{V\\to A}\\), which iterates over negative audio signals in the denominator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.761, + 0.651 + ], + "angle": 0, + "content": "3.3. Audio and Visual Featurizers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The core of DenseAV is two modality-specific backbone networks. We use the DINO vision transformer [5] with ImageNet pretrained weights (without labels) to provide a strong, yet fully unsupervised, vision backbone. Unlike other approaches that use CLIP [46] as a backbone, DINO does not require paired text captions and learns from unlabeled images only. Practically, we find that DINO outperforms CLIP because of its better-behaved local tokens [12], an effect we explore in the Supplement. We append an additional layer norm operation across the channel dimension [3] and a \\( 1 \\times 1 \\) Convolution to DINO. The layer-norm and \\( 1 \\times 1 \\) convolution ensure the architecture does not start with a saturated loss function. We use the HuBERT audio transformer [25] as DenseAV's audio backbone. HubERT operates on waveforms and is trained on the LibriSpeech [40] dataset using only self-supervision. Hubert" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13120" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "outputs a single feature per frequency, corresponding to \\( F = 1 \\) in Section 3. Though HuBERT was only trained on speech, its audio features can be fine-tuned for more general sounds, much like how vision backbones can be fine-tuned for new datasets [56]. As in the visual branch, we append a channel-wise LayerNorm block and two \\( 3 \\times 3 \\) convolutions to the audio branch. These layers help the network avoid saturation and speed convergence. Furthermore, the two convolutions help the model aggregate information, which reduces the cost of the pairwise feature comparison used in our loss function. We refer to these added layers after the pretrained backbones as the \"aligners\" in later sections." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.281, + 0.212, + 0.296 + ], + "angle": 0, + "content": "3.4. Regularizers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.47, + 0.469 + ], + "angle": 0, + "content": "Disentanglement Regularizer, \\(\\mathcal{L}_{\\text{Dis}}\\): We add a small regularization term to encourage each head of Equation 1 to specialize and learn independent types of audio-visual associations. Interestingly we find that our 2-head model naturally learns to distinguish the meaning of words with one head and capture the sounds objects produce with another head. To further encourage this unsupervised discovery of concepts, we penalize the network when multiple attention heads are simultaneously active. More precisely, let \\((a_{b}, v_{b})_{1}^{B}\\) be a Batch of \\(B\\) paired audio and visual signals. Our disentanglement loss for two heads is then:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.476, + 0.469, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D i s} = \\operatorname {M e a n} \\left(\\left| s \\left(a _ {b}, v _ {b}\\right) [ 1 ] \\circ s \\left(a _ {b}, v _ {b}\\right) [ 2 ] \\right|\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Where \\( \\circ \\) is elementwise multiplication and \\( |\\cdot| \\) is the elementwise absolute value function. \\( [k] \\) mirrors PyTorch slicing notation and refers to selecting the activations for only the \\( k \\)th attention head. Intuitively, this loss encourages one head to be silent if the other head is active and is a \"cross-term\" generalization of the \\( l^2 \\) regularizer [24] for encouraging activation shrinkage. When \\( K > 2 \\) we average contributions from every combination of heads. We ablate this, and our decision to max-pool heads in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.469, + 0.813 + ], + "angle": 0, + "content": "Stability Regularizers, \\(\\mathcal{L}_{\\text {Stability }}\\): Finally, we add several other small regularization terms to encourage stable convergence. We detail and ablate these terms in the Supplement. Briefly, these terms include standard regularizers like Total Variation [48] smoothness over time and non-negative pressure to encourage the network to focus on similarity instead of dissimilarity. In addition, we add a regularizer to prevent the calibration temperature, \\(\\gamma\\), from drifting too quickly, and a regularizer to discourage activations during silence and noise. In the supplement we show that each regularizer alone does not have a dramatic effect on final metrics but together they can stop collapses during training." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.813, + 0.468, + 0.828 + ], + "angle": 0, + "content": "Combining these losses into a single loss function yields:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.833, + 0.469, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {A \\rightarrow V} + \\mathcal {L} _ {V \\rightarrow A} + \\lambda_ {\\text {D i s}} \\mathcal {L} _ {\\text {D i s}} + \\mathcal {L} _ {\\text {S t a b i l i t y}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In our experiments we use \\(\\lambda_{Dis} = 0.05\\) and refer interested readers to the supplement for the details of our small stability regularizer, \\(\\mathcal{L}_{\\text{Stability}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.605, + 0.107 + ], + "angle": 0, + "content": "3.5. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.297 + ], + "angle": 0, + "content": "In our experiments we train DenseAV and relevant baselines on the AudioSet [16] dataset for sound prompted segmentation and AudioSet retrieval. We train on the PlacesAudio [22] dataset for speech prompted segmentation, PlacesAudio retrieval, and the ablation studies of Table 4. In our disentanglement experiments of Table 3 and feature visualizations of Figures 1 and 2 we train on both AudioSet and PlacesAudio so that DenseAV can be familiar with both language, the prominent audio signal in PlacesAudio, and more general sounds from AudioSet. In these experiments we sample training data from these two corpora, so each batch has an even split between AudioSet and PlacesAudio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.893, + 0.569 + ], + "angle": 0, + "content": "Warming up Aligners: We find that we can dramatically improve the stability by first training the added aligners (convolutions and layer norms) for 3000 steps while keeping pretrained DINO and HuBERT backbones fixed. This allows the aligners to adapt to these intelligent backbones before modifying each backbone's sensitive weights. We use random resize crops, color jitter, random flips, and random greyscale as image augmentations. We randomly sample a single video frame to feed to our visual branch. Audio clips are converted to single-channel format and are trimmed or padded with silence to create uniform 10 second clips. We re-sample audio clips according to the requirements of the backbone models used. For HuBERT, we re-sample to 16KhZ. We train on 8 V100 GPUs with an effective batch size of 80, and aggregate negative samples on all GPUs prior to computing the loss to ensure efficient parallelization. We provide additional training information and hyperparameters in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Full Training: After warming up the aligners, we train the full model for an additional 800,000 steps using the same loss, batch-size, and training logic. We train all aligner weights and fine-tune all HuBERT audio backbone weights. We use low rank adaptation (LoRA) [26] to fine-tune the \"Q\", \"K\", and \"V\" layers of the DINO visual backbone attention blocks. This allows us to efficiently adapt DINO and stabilize the training as it is quite easy to collapse the carefully trained DINO weights. We use a LoRA rank of 8." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.634, + 0.74 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.854 + ], + "angle": 0, + "content": "To evaluate AV representation quality, we perform a variety of analyses including comparative activation visualization, quantitative measurements of speech and sound prompted semantic segmentation, and cross-modal retrieval. Additionally, we quantify our observation that DenseAV can distinguish the meanings of words (language), from the sounds of objects (sound) without supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To adequately measure a representation's AV alignment quality, we found it necessary to introduce two evaluation datasets that measure speech and sound prompted seman" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13121" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.067, + 0.465, + 0.173 + ], + "angle": 0, + "content": "
MethodSpeech Semseg.Sound Semseg.
mAPmIoUmAPmIoU
DAVENet [23]32.2%26.3%16.8%17.0%
CAVMAE [18]27.2%19.9%26.0%20.5%
ImageBind [17]20.2%19.7%18.3%18.1%
Ours48.7%36.8%32.4%25.5%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.175, + 0.468, + 0.313 + ], + "angle": 0, + "content": "Table 1. Speech and Sound prompted semantic segmentation. We analyze the quality of local features using two prompted semantic segmentation tasks. We prompt networks with speech of the form \"a picture of a(n) [Object]\" to determine whether local feature inner products can segment objects in the ADE20K dataset by name. We create sound prompts for a given ADE20K class using a curated mapping from the ADE20K ontology to the VG-GSound ontology. DenseAV's local features perform significantly better than all baselines investigated. We bold \"first place\" results and underline \"second place\" results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.334, + 0.468, + 0.439 + ], + "angle": 0, + "content": "tic segmentation performance. Our two datasets introduce pairs of speech and sound prompts coupled with matching images and segmentation masks derived from ADE20K. We create these datasets because previous works [23] have not published their datasets or evaluation code. However, we use an experimental setting from the literature for our cross-modal retrieval experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.441, + 0.468, + 0.668 + ], + "angle": 0, + "content": "We compare against a variety of prior art including the popular state-of-the-art multi-modal retrieval network, ImageBind [17]. We also compare against CAVMAE [18], a leading multimodal backbone trained specifically for AudioSet retrieval, and DAVENet [23], which is trained to localize the meanings of words. We include two other baselines [21, 22] which have reported cross modal retrieval metrics on Places Audio. Finally, we compare our multi-head aggregation strategy to common \"global\" retrieval methods such as inner products between classtokens, average-pooled tokens, and SimPooled[44] tokens. We note that SimPool achieves state-of-the-art localization results when compared to 14 other pooling methods. Nevertheless, our multi-head aligner yields better localization results than any of these \"global\" methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.429, + 0.697 + ], + "angle": 0, + "content": "4.1. Qualitative Comparison of Feature Maps" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.704, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our first experiment in Figure 2 highlights the dramatic differences in quality between DenseAV's features and other approaches in the literature. DenseAV is the only backbone whose local tokens are semantically meaningful and show cross-modal alignment for speech and sound. Though both CAVMAE and ImageBind show high-quality retrieval performance, neither shows high quality aligned local tokens. As a result, DenseAV can associate and localize both sound and language significantly better than other backbones. DAVENet shows coarse correspondences between language and visual objects but cannot associate sound with visual objects and does not match DenseAV's high resolution maps. Furthermore, the right half of Figure 1 demon" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.067, + 0.915, + 0.203 + ], + "angle": 0, + "content": "
MethodPlaces Acc. @10AudioSet Acc. @10
I → AA → II → AA → I
[22]*46.3%54.8%--
[21]*54.2%56.4%--
DAVENet [23]*52.8%60.4%--
CAVMAE [18]81.7%77.7%55.7%50.7%
ImageBind [17]1.10%1.10%64.5%66.5%
Ours94.2%94.3%68.2%68.4%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.346 + ], + "angle": 0, + "content": "Table 2. Cross-modal retrieval using 1000 evaluation videos from the PlacesAudio and AudioSet validation datasets. DenseAV dramatically outperforms all approaches tested in all metrics. Most notably, the state-of-the-art image retrieval foundation model, ImageBind, is incapable of recognizing speech. We note that the ImageBind authors do not publish retraining code, so we evaluate their largest pretrained model. Models with a * indicate that they have been previously reported in the literature. Other numbers are calculated by using pretrained models when available or from training with the author's official training scripts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.446 + ], + "angle": 0, + "content": "strates that DenseAV naturally discovers and separates word semantics from the sound of objects without labels to supervise this separation. In the supplement, we provide additional visualizations of all backbones considered across a wide range of words and sounds." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.455, + 0.833, + 0.472 + ], + "angle": 0, + "content": "4.2. Speech Prompted Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Dataset: We introduce a speech prompted segmentation dataset using the ADE20K dataset, which is known for its comprehensive ontology and pixel-precise annotations [59]. From this dataset, we curate an evaluation subset of image-class pairs by sampling up to 10 images for each object class in ADE20K, excluding images where the selected class was tiny (\\(< 5\\%\\) of pixels). We only consider classes with at least 2 images that pass the tiny object criterion. For each class and image, we formed a binary target mask by selecting the semantic segmentation mask for that class. This resulted in 3030 image-object pairs spanning 478 ADE20K classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We created paired speech signals by speaking the prompt \"A picture of a(n) [object]\" where [object] is the name of the ADE20K class. We create clear, controlled, and consistent audio prompts using Microsoft's neural text to speech service [47]. This service also provides exact timing of the \"[object]\" utterance within the broader prompt and ensures each class is measured equally. Grammar was manually verified for the utterances to ensure proper singular/plural and a/an agreement with the class name. We release images, masks, and audio prompts for reproducibility." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Measure: We evaluate methods based on how well their speech-prompted activations align with ground truth masks for the visual object's class. We quantify this with the binary Average Precision (AP) and Intersection over Union (IoU) metrics. These quantify how close activations match with the binary label mask from the ADE20K dataset. To compute an aggregate score over all of the object" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13122" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.066, + 0.462, + 0.131 + ], + "angle": 0, + "content": "
MethodPred. Dis.Act. Dis.
No LDis, No Head Max Pool64.1%70.3%
No LDis99.9%86.5%
Ours99.9%91.2%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.132, + 0.47, + 0.203 + ], + "angle": 0, + "content": "Table 3. Quantitative ablation study of the impact of max-pooling attention heads and adding our disentanglement loss, \\(\\mathcal{L}_{Dis}\\). Intuitively, max-pooling attention heads allows each head to specialize on its own specific set of triggers. Our disentanglement loss further encourages the heads to operate independently and orthogonally." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.469, + 0.275 + ], + "angle": 0, + "content": "classes considered, we compute the mean average precision (mAP) and mean intersection over union (mIoU) by averaging AP scores across all object categories considered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.47, + 0.517 + ], + "angle": 0, + "content": "The mAP is particularly well suited for evaluating feature similarities because it is unaffected by monotonic transformations of the similarity scores. This eliminates the need for arbitrary thresholding and calibration. This is particularly important because many networks' inner products are not centered at zero, and the best thresholding strategy can be nontrivial, and dependent on the network and object class. Average Precision avoids these confounding factors and ensures a fair comparison across methods. Unfortunately, unlike the mAP, the mIoU metric requires selecting a threshold. To ensure our mIoU measurement is similarly invariant to monotonic transformations we evaluate 20 uniformly spaced thresholds between the smallest and largest activations of each model. For each baseline, we report results for the best threshold to ensure a fair comparison between all networks considered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Implementation: We compute image heatmaps by evaluating each modality-specific network on the image-audio pairs from our dataset. We extract dense features from the final layer of each network and form their similarity volume according to Equation 1. For DenseAV we max-pool the head dimension to properly compare with single-headed models. We average activations over the temporal extent of the \"[object]\" utterance using the word timing information from the ground truth audio clip. This creates a heatmap over the image features that can be bi-linearly resized to the original image's size. We then compare these per-pixel activation scores to ground truth object masks from our dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.469, + 0.806 + ], + "angle": 0, + "content": "Results: In Speech mAP and mIoU columns of Table 1 we show that DenseAV achieves a \\(51\\%\\) \\((+16.5\\mathrm{mAP})\\) relative increase in speech-prompted semantic segmentation over previous methods. Approaches that use global token based contrastive strategies such as CAVMAE and Image-Bind perform particularly poorly in this task, and this observation aligns with the qualitative results of Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.406, + 0.833 + ], + "angle": 0, + "content": "4.3. Sound Prompted Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Dataset: To evaluate how well deep features localize sound, we build on Section 4.2 and create a dataset of sound prompts that align with ADE20K classes. We first select the same (large) image-object pairs from ADE20K. We then" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.066, + 0.895, + 0.161 + ], + "angle": 0, + "content": "
MethodSpeech mAPPlaces Acc. @10
V → AA → V
Average Pool20.1%92.0%91.2%
CLS Token20.6%86.4%89.8%
SimPool [44]35.3%92.6%92.8%
Multi-Head (Ours)48.2%93.5%93.8%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.164, + 0.894, + 0.233 + ], + "angle": 0, + "content": "Table 4. Quantitative ablation of different feature aggregation strategies. Though the common practice of average pooling and using a learned CLS token to aggregate features have little effect on retrieval performance, they dramatically degrade performance on speech prompted semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.472 + ], + "angle": 0, + "content": "create a mapping between the ADE20K and VGGSound [6] ontologies. To compute a robust mapping, we first embed ADE20K class names and VGGSound class names with the GPT Ada 2 text embedding model [37]. For each ADE20K class, we create a list of at most three candidates from the VGGSound ontology that have a cosine similarity \\((> .85)\\). We then manually review these candidates to select the best VGGSound class for each ADE20K class and remove any spurious or mistaken matches. This produces a set of 95 ADE20K classes with strong matches in the VGGSound ontology. For each of our original 3030 image-object pairs we select a random VGGSound validation clip with a matching class according to our mapped ontology. This yields 106 image-object pairs across 20 ADE20K classes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.473, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Evaluation Measure: We use the same mAP and mIoU evaluation metrics as Section 4.2, but instead average over the 20 ADE20K classes considered." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.519, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Implementation: We compute sound prompted image activations as in section 4.2 but with one key change: we average activations over the entire clip because we do not have ground-truth sound timing information." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.58, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Results: The \"Sound mAP and mIoU\" columns of Table 1 show that DenseAV achieves a \\(25\\%\\) (+6.4mAP) relative improvement in sound prompted segmentation compared to the prior art. Most notably, ImageBind's features cannot localize sound despite their high cross-modal retrieval performance learned from millions of hours of sound." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.712, + 0.696 + ], + "angle": 0, + "content": "4.4. Cross-Modal Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We show that DenseAV's representations are not only better for localization, but significantly outperform other approaches on cross-modal retrieval. We adopt the evaluation setting of [23] and measure cross modal retrieval accuracy at 1, 5, and 10 in a thousand-way retrieval task. In particular, we use the same thousand images from the validation set of [23] and also replicate this analysis on one-thousand random clips from the AudioSet validation data. Table 2 shows results for 1000-way retrieval tasks on both the Places Audio and AudioSet datasets. We show cross-modal accuracy at 10, but also show larger tables in the supplement that echo these results using accuracy at 1 and 5. DenseAV significantly outperforms all baselines across all metrics. Interest-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13123" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.123 + ], + "angle": 0, + "content": "ingly, DenseAV outperforms ImageBind with less than half of the trainable parameters and no reliance on text." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.131, + 0.332, + 0.148 + ], + "angle": 0, + "content": "4.5. Measuring Disentanglement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.47, + 0.41 + ], + "angle": 0, + "content": "We observe that DenseAV's heads naturally learn to differentiate audio-visual couplings that capture the meaning of words (language) and those that capture the sounds of objects (sound). Furthermore this effect generalizes to novel clips, including those with both sound and language as shown in Figure 1. We quantify this observation in two ways, the first measures if a head's average activation strength predicts whether a clip contains mainly \"language\" or \"sound\". The second method quantifies how often the \"sound\" head is incorrectly active when the \"language\" head should be active and vice versa. We leverage the fact that AudioSet dataset contains mostly clips with ambient sound and rarely contains language. In contrast, Places Audio is entirely language-based without external ambient sound. We note that these analyses are specifically for our architecture with two heads \\( K = 2 \\) and trained on both AudioSet and PlacesAudio data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.469, + 0.531 + ], + "angle": 0, + "content": "For both measures of disentanglement, we first compute a clip's aggregated similarity for each head. In particular, we remove the max-pooling over heads in Equation 2 to create a single-head similarity, \\( S(a,v)_k \\). We then min-max scale the scores of each head across both datasets to lie in the [0,1] interval, which we refer to as \\( \\hat{S}(a,v)_k \\). Using these normalized scores, we can create metrics that capture how well a given head responds only to a specific dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.532, + 0.47, + 0.624 + ], + "angle": 0, + "content": "Our first metric measures how well a head's scores predict whether a clip is from the \"sound\" or \"language\" dataset. Let \\((a_{b},v_{b})_{1}^{B}\\) be tuples of paired audio and visual signals. Let \\(l[k^{\\prime}]_{b}\\) be an indicator variable of whether the signal \\((a_{b},v_{b})\\) arises from the sound dataset, AudioSet, \\((k^{\\prime} = 1)\\), or the language dataset Places Audio \\((k^{\\prime} = 2)\\)." + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.635, + 0.47, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\delta_ {p r e d} (k, k ^ {\\prime}) = \\mathrm {A P} \\left(\\left(\\hat {\\mathcal {S}} \\left(a _ {b}, v _ {b}\\right) _ {k}\\right) _ {1} ^ {B}, \\left(l \\left[ k ^ {\\prime} \\right] _ {b}\\right) _ {1} ^ {B}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Where \\(AP(\\cdot, \\cdot)\\) is the binary average precision with prediction and label arguments respectively. Intuitively, this measures whether the scores of head \\(k\\) are direct predictors of whether the data is from dataset \\(k'\\). We can find the best assignment between heads and datasets such that each head is maximally predictive of the given dataset:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.767, + 0.469, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {P r e d D i s} = \\frac {1}{2} \\max \\left(\\delta_ {p r e d} (0, 0) + \\delta_ {p r e d} (1, 1), \\right. \\\\ \\delta_ {p r e d} (1, 0) + \\delta_ {p r e d} (0, 1)) \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The prediction disentanglement score, PredDis, is a percentage that ranges from \\(50\\%\\) for completely entangled signals to \\(100\\%\\) if one can perfectly classify the signals using the scores of either head. The maximum over the two possible assignments makes this metric invariant to permutations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "of the heads. We note that this metric is a Hungarian matching assignment [29] over two entries, a common technique to asses unsupervised classification performance [20, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Our second measure quantifies \"spurious activations\" in the non-dominant head. A truly disentangled system should have a head that only fires on sound, and another head that only fires on language. We create another disentanglement measure, ActDis, by replacing \\(\\delta_{pred}\\) in Equation 7 with:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.215, + 0.892, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\delta_ {a c t} \\left(k, k ^ {\\prime}\\right) = 1 - \\frac {1}{\\sum_ {b ^ {\\prime}} l \\left[ k ^ {\\prime} \\right] _ {b ^ {\\prime}}} \\sum_ {b = 1} ^ {B} \\hat {\\mathcal {S}} \\left(a _ {b}, v _ {b}\\right) _ {k} \\cdot l \\left[ k ^ {\\prime} \\right] _ {b} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.353 + ], + "angle": 0, + "content": "Intuitively, this measures the \"inactivity\" of head \\( k \\) on dataset \\( k' \\). If head \\( k \\) is totally silent on dataset \\( k' \\) then \\( \\delta_{act}(k, k') = 1 \\). Like PredDis, ActDis is a percentage ranging from \\( 50\\% \\) to \\( 100\\% \\) with \\( 100\\% \\) representing perfect disentanglement where the sound head is completely silent during the language clips, and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Table 3 shows that DenseAV achieves near perfect predictive (99%) and activation (91%) disentanglement. It also shows that our disentanglement regularizer and max-pooling over heads improves DenseAV's natural ability to distinguish sound from language without supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.619, + 0.456 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.692 + ], + "angle": 0, + "content": "We presented DenseAV, a novel contrastive learning architecture that can discover the meaning of words and localize the sounds of objects using only video supervision. We are the first to observe both qualitatively and quantitatively that it's possible to disentangle the meaning of words from the sound of objects with only a contrastive learning signal. DenseAV's success stems from its novel multi-head attention aggregation mechanism that encourages its modality-specific backbones to create high-resolution, semantically meaningful, and AV aligned representations. These properties of DenseAV's representation are not seen in other state-of-the-art models in the literature. Consequently, DenseAV significantly surpasses other leading models in dense prediction tasks such as speech and sound-prompted semantic segmentation as well as in cross-modal retrieval." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.704, + 0.668, + 0.719 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation. This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/). This work is funded by a Royal Society Research Professorship RSRP\\R\\241003, and EPSRC Programme Grant VisualAI EP/T028572/1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13124" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.471, + 0.227 + ], + "angle": 0, + "content": "[2] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.268 + ], + "angle": 0, + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Jeongsoo Choi, Joanna Hong, and Yong Man Ro. Diffv2s: Diffusion-based video-to-speech synthesis with vision-guided speaker embedding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7812-7821, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[10] Noam Chomsky. Language and problems of knowledge: The Managua lectures. MIT press, 1987. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[11] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Nicholas Frosst, Nicolas Papernot, and Geoffrey Hinton. Analyzing and improving representations with the soft nearest neighbor loss. In International conference on machine learning, pages 2012-2020. PMLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[17] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3,4,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[18] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[19] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[21] David Harwath and James R Glass. Learning word-like units from joint audio-visual analysis. arXiv preprint arXiv:1701.07481, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[22] David Harwath, Antonio Torralba, and James Glass. Unsupervised learning of spoken language with visual context. Advances in Neural Information Processing Systems, 29, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[23] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649-665, 2018. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[24] Arthur E Hoerl and Robert W Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55-67, 1970. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[25] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13125" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[27] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[28] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9865-9874, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.469, + 0.364 + ], + "angle": 0, + "content": "[30] Neeraj Kumar, Srishti Goel, Ankur Narang, and Mujtaba Hasan. Robust one shot audio to video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 770-771, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[31] Shiguang Liu, Sijia Li, and Haonan Cheng. Towards an end-to-end visual-to-raw-audio generation with gan. IEEE Transactions on Circuits and Systems for Video Technology, 32(3): 1299–1312, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[32] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song. Active contrastive learning of audio-visual video representations. arXiv preprint arXiv:2009.09805, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[33] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.528, + 0.468, + 0.555 + ], + "angle": 0, + "content": "[34] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[35] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871–14881, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.353, + 0.702 + ], + "angle": 0, + "content": "[37] OpenAI. Gpt-4 technical report, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[38] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Is object localization for free?-weakly-supervised learning with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 685-694, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[40] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Alex S Park and James R Glass. Unsupervised pattern discovery in speech. IEEE Transactions on Audio, Speech, and Language Processing, 16(1):186-197, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.257 + ], + "angle": 0, + "content": "[42] Puyuan Peng and David Harwath. Self-supervised representation learning for speech using visual grounding and masked language modeling. arXiv preprint arXiv:2202.03543, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[43] Puyuan Peng and David Harwath. Word discovery in visually grounded, self-supervised speech models. arXiv preprint arXiv:2203.15081, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[44] Bill Psomas, Ioannis Kakogeorgiou, Konstantinos Karantzalos, and Yannis Avrithis. Keep it simpool: Who said supervised transformers suffer from attention deficit? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5350-5360, 2023. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[45] Geoffrey K Pullum and Barbara C Scholz. Empirical assessment of stimulus poverty arguments. The linguistic review, 19(1-2):9-50, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[47] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv preprint arXiv:2006.04558, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[48] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[49] Yi-Jen Shih, Hsuan-Fu Wang, Heng-Jui Chang, Layne Berry, Hung-yi Lee, and David Harwath. Speechclip: Integrating speech with pre-trained vision and language model. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 715-722. IEEE, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[50] Linda Smith and Chen Yu. Infants rapidly learn word-referent mappings via cross-situational statistics. Cognition, 106(3):1558-1568, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[53] Luyu Wang and Aaron van den Oord. Multi-format contrastive learning of audio representations. arXiv preprint arXiv:2103.06508, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13126" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[55] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[56] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[57] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European conference on computer vision (ECCV), pages 570-586, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[60] Hao Zhu, Man-Di Luo, Rui Wang, Ai-Hua Zheng, and Ran He. Deep audio-visual learning: A survey. International Journal of Automation and Computing, 18:351–376, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13127" + } + ] +] \ No newline at end of file diff --git a/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_origin.pdf b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f0927bed9b331999e1d72deb59fe4ad94cde18e --- /dev/null +++ b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/7873c286-c4ef-4202-8df2-5bd3e7312a38_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d02bfb5f58b5ff2c61e299a65ff73e07c04ab24e725647cd761f8583f45cbbb3 +size 3346702 diff --git a/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/full.md b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1e6fd93ae15e8a3a948d92bbce50b272b85b3217 --- /dev/null +++ b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/full.md @@ -0,0 +1,307 @@ +# Separating the "Chirp" from the "Chat": Self-supervised Visual Grounding of Sound and Language + +Mark Hamilton +MIT, Microsoft +markth@mit.edu + +Andrew Zisserman Oxford, Google + +John R. Hershey Google + +William T. Freeman +MIT, Google + +![](images/bbd6229aa8445331f3b0e4631e0a77170938c732de2277a7475a59f813380723.jpg) +Figure 1. Visual overview of the DenseAV algorithm. Two modality-specific backbones featurize audio and visual signals. We introduce a novel generalization of multi-head attention to extract attention maps that discover and separate the "meaning" of spoken words and the sounds an object makes. DenseAV performs this localization and decomposition solely through observing paired stimuli such as videos. + +# Abstract + +We present DenseAV, a novel dual encoder grounding architecture that learns high-resolution, semantically meaningful, and audio-visual aligned features solely through watching videos. We show that DenseAV can discover the "meaning" of words and the "location" of sounds without explicit localization supervision. Furthermore, it automatically discovers and distinguishes between these two types of associations without supervision. We show that DenseAV's localization abilities arise from a new multi-head feature aggregation operator that directly compares dense image and audio representations for contrastive learning. In contrast, many other systems that learn "global" audio and video representations cannot localize words and sound. Finally, we contribute two new datasets to improve the evaluation of AV representations through speech and sound prompted semantic segmentation. On these and other datasets we show DenseAV dramatically outperforms the prior art on speech and sound prompted semantic segmentation. DenseAV outperforms the current state-of-the-art, ImageBind, on cross-modal retrieval using fewer than half of the parameters. Project Page: https://aka.ms/denseav + +# 1. Introduction + +Associating audio and video is a fundamental aspect of human perception. As infants develop, the synchronization and correspondence of sounds enables multi-modal association – a voice with a face, and a 'moo' with a cow [50]. Later, as they acquire language, they associate spoken words with objects they represent [10, 45]. Amazingly, these association abilities, constituting speech recognition, sound event recognition, and visual object recognition, develop without much direct supervision. This work aims to create a model with this capability by learning high-resolution, semantically meaningful, audio-visual (AV) aligned representations. Features with these properties can be used to discover fine-grained correspondences between modalities without localization supervision or knowledge of the textual representation of language. + +As an example, consider the spoken caption and accompanying sounds of the image shown in Figure 1. We wish to "ground" both the speech and the sounds with high resolution. For instance, both the spoken word "dog" and the sound of a bark in the audio signal should be associated with the pixels of the dog in the visual signal if present. We seek high quality local representations where simple inner + +![](images/90aafa0b6605adfbec68b597cef114d2f07e1b8e2f7999d23fcfcce7dcfbb642.jpg) + +![](images/b2f1046456c605eb34400f3e15b8b967034ac0ec032c9d9eabe6a9e8c7ddb2ab.jpg) +Figure 2. Qualitative comparison of several modern architectures for associating audio and video modalities. Only DenseAV learns a high-resolution and semantically aligned set of local features. This allows us to perform speech and sound prompted semantic segmentation using only the inner products between deep features. Other approaches, such as ImageBind, do not show aligned local feature maps. Approaches that do show some localization capabilities, like DAVENet, do not generalize to sound and language, and do not achieve the high-resolution localization capabilities of DenseAV. Dense features are visualized using PCA as in [20] + +products between features exhibit this behavior, which is interestingly absent from popular approaches in the literature. + +To achieve this, we make three innovations. First, we introduce DenseAV, a dual-encoder architecture that computes a dense similarity volume over audio and visual features before aggregating this volume into a single similarity score. If we look at a slice of this similarity volume when word is spoken, as in Figure 1, we can visualize the AV activation strength between a word or sound and an image's pixels. The novelty we introduce is to generalize this dense similarity mechanism with multiple heads, much like those of multi-head attention. This generalization allows each head to specialize on particular types of couplings between the visual and audio modalities. Interestingly, we discover that if we give DenseAV two heads and train on a dataset that contains both language and sound, the heads naturally learn to distinguish language from more general sound using only cross-modality supervision. For example, as shown in Figure 1, head 1 focuses on what objects create a sound, like the barking of a dog, whereas head 2 focuses only on the meaning of words. + +Second, we show the importance of the "aggregation function" one uses to create a similarity score between an audio clip and a video frame for contrastive learning. The traditional choices; using inner products between global representations such as class tokens [5, 13, 49] or pooled features [18, 58], do not promote AV alignment of dense local features. Because of this, several popular audio-video backbones that excel on cross-modal retrieval cannot directly associate objects and sounds using their local features. This limits their ability to be used for downstream + +tasks such as semantic segmentation, sound localization, or unsupervised language learning and discovery. + +Third, we introduce two semantic segmentation datasets to evaluate visual grounding with AV representations for speech and (non-speech) sounds. We build these datasets from the high-quality segmentation masks provided by the ADE20K dataset [59] and measure mean average precision (mAP) and mean intersection over union (mIoU) on a binary mask prediction task. This evaluation is simpler and more thorough than previous efforts to measure visual grounding such as the concept counting metrics of [23] and the "pointing games" of [2, 14, 38] that only check if a heatmap's peak occurs within a target box or segment. Furthermore, our evaluation avoids brittle word-net ontologies [34], clustering, Wu and Palmer distance [55], threshold choices, and a variety of other complicating factors. + +To summarize, our main contributions are as follows: + +- We introduce DenseAV, a novel self-supervised architecture that learns high-resolution AV correspondences. +- We introduce a local-feature-based image similarity function that significantly improves a network's zero-shot localization ability compared to common strategies such as average pooling or CLS tokens. +- We introduce new datasets for evaluating speech and sound prompted semantic segmentation. We show DenseAV significantly outperforms the current state-of-the-art on these tasks as well as on cross-modal retrieval. +- We discover that our multi-head architecture naturally disentangles audio-visual correspondence into sound and language components using only contrastive supervision. + +![](images/039a8c44a90011de72a3148624ab418a96b3fc9f8ab983fa4374d1d4b7feeec0.jpg) +Figure 3. Architectural overview of our multi-head attention aggregator. Dense feature maps are split into $K$ heads ( $K = 1,2$ ) in our experiments. We form an AV activation tensor by taking the inner-products of each head's features across the spatial and temporal extent of the visual and audio signals respectively as in Equation 1. We then aggregate this similarity volume into a single similarity score by max-pooling head and spatial dimensions and average-pooling audio dimensions. Our approach aims to encourage the network to identify specific shared objects between the audio and visual modalities. In particular, max-pooling of heads disentangles sound and language, and max-pooling spatial dimensions helps localize objects. + +# 2. Related Work + +Audio-visual (AV), text-visual, and more general multimodal models have recently surged in popularity [60]. Broadly speaking DenseAV is an audio-video contrastive learning architecture, this class of methods learns AV representations by aligning paired signals and pushing apart negative signals [11, 27]. Of the models in this class, several stand out for their ability to localize sounds [2, 7, 42] or capture the semantics of language [23, 43]. Many models in this class compare AV signals using inner products between "global" representations formed by pooled deep features [18, 35, 53], or class tokens [17, 32, 42, 43, 49]. Most notably, ImageBind has gained popularity due to its state-of-the-art performance on a variety of tasks and datasets and unified class-token-based contrastive architecture. In this work we show that many of these architectures do not show strong localization properties in their local features, despite excelling at cross-modal retrieval on a "global" level. This limits their applicability to new out-of-domain sounds, sounds that don't have a textual representation, and low-resource languages. We diverge from these works by directly supervising local tokens. In particular, we build on previous works [2, 23] that show max-pooling improves localization capabilities and introduce a new multi-head aggregation operator that generalizes previous losses using an operator similar to self-attention [52]. + +Another class of methods discover structure in signals through uni- and multi-modal clustering. Early works on audio clustering [41] discovered meaningful utterances without supervision. Similar visual analyses have discovered visual objects [4, 8, 20, 28]. Recent works have applied these ideas to the AV domain [1, 21], but do not focus on extracting high-resolution AV representations. + +Finally, several works investigate generative audio-video learning. The Sound of Pixels [57] generates the sound of a specific object using a source separation loss. Newer approaches using GANs [30, 31], and diffusion models [9, 17, 33] have generated audio from video and vice versa. Here we focus on improving the local representations of contrastive learners because of their relative scalability, simplicity, and ability to learn high-quality representations. + +# 3. Methods + +At a high level, DenseAV tries to determine when a given audio and visual signal belong “together” using dense audio-visual representations. To perform this task robustly, DenseAV must learn how to predict the contents of an audio signal from a visual signal and vice versa. This causes DenseAV to learn dense modality-specific features that capture the mutual information shared between the modalities [51]. Once learned, we can directly query these informative features to perform speech and sound prompted semantic segmentation as illustrated in Figure 1. + +More specifically, DenseAV is built from two modality-specific deep featurizers. These backbones produce temporally varying audio features across an audio clip and spatially varying video features for a single randomly selected frame. Our loss computes a similarity between audio and visual signals based on the intuition that two signals are similar if they have a variety of strong couplings or shared objects. More formally, we form a scalar similarity for a pair of audio and video signals by carefully aggregating a volume of pairwise inner products between dense features. We use the InfoNCE [36] contrastive loss to encourage similarity between "positive" pairs of signals and dissimilarity between "negative" pairs formed by in-batch shuffling. Figure 3 graphically depicts this loss function and subsequent + +sections detail each component of our architecture. + +# 3.1. Multi-Headed Aggregation of Similarities + +DenseAV's key architectural distinction is its loss function that directly supervises the "local" tokens of the visual and audio featurizers. This is a significant departure from other works [5, 17, 19, 39, 46, 49] that pool modality specific information into "global" representations prior to the contrastive loss. Unlike prior works, our loss function aggregates the full pairwise similarities between the local tokens into an aggregate measure of similarity for a given pair of audio and visual signals. We show in Figure 2 that this architectural choice enables DenseAV's local features to align across modalities whereas other approaches such as average pooling, class tokens, and SimPool [44] do not. + +We first describe our loss function informally and definite it more precisely in the next paragraph. Our loss function computes the (un-normalized) inner product between every pair of visual and audio features to form a "volume" of inner products. This volume represents how strongly each part of an audio signal "couples" to each part of a visual signal. We aim to find many large couplings between positive pairs of audio and visual signals. Ideally, these couplings should connect visual objects with their references in the audio signal. Conversely, we do not want to find couplings between negative pairs of signals. To compute a single global coupling strength for a pair of signals, we aggregate this volume of pairwise similarities into a single number. There are myriad ways to aggregate this volume ranging in "softness" from average-pooling to max-pooling. Average pooling yields dense gradients and can improve convergence speed and stability. However, max-pooling allows the network to focus on the best couplings regardless the object's size or a sound's duration. Our aggregation function combines the benefits of average and max pooling by max-pooling visual dimensions and average pooling audio dimensions as proposed in [23]. Intuitively speaking, this averages the strongest image couplings over an audio signal. It allows small visual objects to have large effects yet provides a strong training gradient to many regions of the signals. Finally, we draw inspiration from multi-head self-attention [52] and generalize this operation to multiple "heads" that we max-pool before pooling the visual and audio dimensions. This allows DenseAV to discover multiple "ways" to associate objects across modalities. + +More formally, let $S(a,v) \in \mathbb{R}$ represent the similarity between a tensor of audio features $a \in \mathbb{R}^{CKFT}$ of size (Channel × K-heads × Frequency × Time) and a tensor of visual features $v \in \mathbb{R}^{CKHW}$ of size (Channel × K-heads × Height × Width). To define this scalar similarity score, we first create a local similarity volume, $s(a,v) \in \mathbb{R}^{kfhw}$ . For simplicity, we consider the aggregated similarity between a single image and audio clip but note one can easily general- + +ize this to max-pool over video-frames. We define the full pairwise volume of similarities as: + +$$ +s (a, v) \in \mathbb {R} ^ {k f t h w} = \sum_ {c = 1} ^ {C} a [ c, k, f, t ] \cdot v [ c, k, h, w ] \tag {1} +$$ + +Where $a[c,k,f,t]$ represents the value of $a$ at location $[c,k,f,t]$ and $\cdot$ is scalar multiplication. We aggregate this similarity volume into a single score $S(a,v)\in \mathbb{R}$ : + +$$ +\mathcal {S} (a, v) = \frac {1}{F T} \sum_ {f = 1} ^ {F} \sum_ {t = 1} ^ {T} \max _ {k, h, w} (s (a, v) [ k, f, t, h, w ]) \tag {2} +$$ + +We note that this operation can be viewed as a multi-head generalization of the MISA loss of [23], and a multi-head multi-time generalization of the MIL loss of [2]. + +# 3.2. Loss + +We can use the similarity between audio and visual signals defined in Equation 2 to construct a contrastive loss. We follow recent works [15, 17, 54] and use the temperature-weighted InfoNCE [36] to encourage similarity between positive pairs of signals and dissimilarity between negative pairs. In DenseAV, we form $B$ positive pairs by splitting the audio and visual components of a Batch of training data. We form $B^2 - B$ negative pairs by comparing a signal to all of the other signals in the training batch. More formally let $(a_b, v_b)_1^B$ be $B$ pairs of audio and visual signals. The visual-retrieval term of our InfoNCE loss is then: + +$$ +\mathcal {L} _ {A \rightarrow V} = \frac {1}{2 B} \sum_ {i = 1} ^ {B} \left(\log \frac {\exp \left(\gamma \mathcal {S} \left(a _ {b} , v _ {b}\right)\right)}{\sum_ {b ^ {\prime} = 1} ^ {B} \exp \left(\gamma \mathcal {S} \left(a _ {b} , v _ {b ^ {\prime}}\right)\right)}\right) \tag {3} +$$ + +Where $\gamma \in \mathbb{R}^{+}$ is a trainable inverse temperature parameter. We symmetrize this loss by adding the analogous audio-retrieval term, $\mathcal{L}_{V\to A}$ , which iterates over negative audio signals in the denominator. + +# 3.3. Audio and Visual Featurizers + +The core of DenseAV is two modality-specific backbone networks. We use the DINO vision transformer [5] with ImageNet pretrained weights (without labels) to provide a strong, yet fully unsupervised, vision backbone. Unlike other approaches that use CLIP [46] as a backbone, DINO does not require paired text captions and learns from unlabeled images only. Practically, we find that DINO outperforms CLIP because of its better-behaved local tokens [12], an effect we explore in the Supplement. We append an additional layer norm operation across the channel dimension [3] and a $1 \times 1$ Convolution to DINO. The layer-norm and $1 \times 1$ convolution ensure the architecture does not start with a saturated loss function. We use the HuBERT audio transformer [25] as DenseAV's audio backbone. HubERT operates on waveforms and is trained on the LibriSpeech [40] dataset using only self-supervision. Hubert + +outputs a single feature per frequency, corresponding to $F = 1$ in Section 3. Though HuBERT was only trained on speech, its audio features can be fine-tuned for more general sounds, much like how vision backbones can be fine-tuned for new datasets [56]. As in the visual branch, we append a channel-wise LayerNorm block and two $3 \times 3$ convolutions to the audio branch. These layers help the network avoid saturation and speed convergence. Furthermore, the two convolutions help the model aggregate information, which reduces the cost of the pairwise feature comparison used in our loss function. We refer to these added layers after the pretrained backbones as the "aligners" in later sections. + +# 3.4. Regularizers + +Disentanglement Regularizer, $\mathcal{L}_{\text{Dis}}$ : We add a small regularization term to encourage each head of Equation 1 to specialize and learn independent types of audio-visual associations. Interestingly we find that our 2-head model naturally learns to distinguish the meaning of words with one head and capture the sounds objects produce with another head. To further encourage this unsupervised discovery of concepts, we penalize the network when multiple attention heads are simultaneously active. More precisely, let $(a_{b}, v_{b})_{1}^{B}$ be a Batch of $B$ paired audio and visual signals. Our disentanglement loss for two heads is then: + +$$ +\mathcal {L} _ {D i s} = \operatorname {M e a n} \left(\left| s \left(a _ {b}, v _ {b}\right) [ 1 ] \circ s \left(a _ {b}, v _ {b}\right) [ 2 ] \right|\right) \tag {4} +$$ + +Where $\circ$ is elementwise multiplication and $|\cdot|$ is the elementwise absolute value function. $[k]$ mirrors PyTorch slicing notation and refers to selecting the activations for only the $k$ th attention head. Intuitively, this loss encourages one head to be silent if the other head is active and is a "cross-term" generalization of the $l^2$ regularizer [24] for encouraging activation shrinkage. When $K > 2$ we average contributions from every combination of heads. We ablate this, and our decision to max-pool heads in Table 3. + +Stability Regularizers, $\mathcal{L}_{\text {Stability }}$ : Finally, we add several other small regularization terms to encourage stable convergence. We detail and ablate these terms in the Supplement. Briefly, these terms include standard regularizers like Total Variation [48] smoothness over time and non-negative pressure to encourage the network to focus on similarity instead of dissimilarity. In addition, we add a regularizer to prevent the calibration temperature, $\gamma$ , from drifting too quickly, and a regularizer to discourage activations during silence and noise. In the supplement we show that each regularizer alone does not have a dramatic effect on final metrics but together they can stop collapses during training. + +Combining these losses into a single loss function yields: + +$$ +\mathcal {L} = \mathcal {L} _ {A \rightarrow V} + \mathcal {L} _ {V \rightarrow A} + \lambda_ {\text {D i s}} \mathcal {L} _ {\text {D i s}} + \mathcal {L} _ {\text {S t a b i l i t y}} \tag {5} +$$ + +In our experiments we use $\lambda_{Dis} = 0.05$ and refer interested readers to the supplement for the details of our small stability regularizer, $\mathcal{L}_{\text{Stability}}$ . + +# 3.5. Training + +In our experiments we train DenseAV and relevant baselines on the AudioSet [16] dataset for sound prompted segmentation and AudioSet retrieval. We train on the PlacesAudio [22] dataset for speech prompted segmentation, PlacesAudio retrieval, and the ablation studies of Table 4. In our disentanglement experiments of Table 3 and feature visualizations of Figures 1 and 2 we train on both AudioSet and PlacesAudio so that DenseAV can be familiar with both language, the prominent audio signal in PlacesAudio, and more general sounds from AudioSet. In these experiments we sample training data from these two corpora, so each batch has an even split between AudioSet and PlacesAudio. + +Warming up Aligners: We find that we can dramatically improve the stability by first training the added aligners (convolutions and layer norms) for 3000 steps while keeping pretrained DINO and HuBERT backbones fixed. This allows the aligners to adapt to these intelligent backbones before modifying each backbone's sensitive weights. We use random resize crops, color jitter, random flips, and random greyscale as image augmentations. We randomly sample a single video frame to feed to our visual branch. Audio clips are converted to single-channel format and are trimmed or padded with silence to create uniform 10 second clips. We re-sample audio clips according to the requirements of the backbone models used. For HuBERT, we re-sample to 16KhZ. We train on 8 V100 GPUs with an effective batch size of 80, and aggregate negative samples on all GPUs prior to computing the loss to ensure efficient parallelization. We provide additional training information and hyperparameters in the supplement. + +Full Training: After warming up the aligners, we train the full model for an additional 800,000 steps using the same loss, batch-size, and training logic. We train all aligner weights and fine-tune all HuBERT audio backbone weights. We use low rank adaptation (LoRA) [26] to fine-tune the "Q", "K", and "V" layers of the DINO visual backbone attention blocks. This allows us to efficiently adapt DINO and stabilize the training as it is quite easy to collapse the carefully trained DINO weights. We use a LoRA rank of 8. + +# 4. Experiments + +To evaluate AV representation quality, we perform a variety of analyses including comparative activation visualization, quantitative measurements of speech and sound prompted semantic segmentation, and cross-modal retrieval. Additionally, we quantify our observation that DenseAV can distinguish the meanings of words (language), from the sounds of objects (sound) without supervision. + +To adequately measure a representation's AV alignment quality, we found it necessary to introduce two evaluation datasets that measure speech and sound prompted seman + +
MethodSpeech Semseg.Sound Semseg.
mAPmIoUmAPmIoU
DAVENet [23]32.2%26.3%16.8%17.0%
CAVMAE [18]27.2%19.9%26.0%20.5%
ImageBind [17]20.2%19.7%18.3%18.1%
Ours48.7%36.8%32.4%25.5%
+ +Table 1. Speech and Sound prompted semantic segmentation. We analyze the quality of local features using two prompted semantic segmentation tasks. We prompt networks with speech of the form "a picture of a(n) [Object]" to determine whether local feature inner products can segment objects in the ADE20K dataset by name. We create sound prompts for a given ADE20K class using a curated mapping from the ADE20K ontology to the VG-GSound ontology. DenseAV's local features perform significantly better than all baselines investigated. We bold "first place" results and underline "second place" results. + +tic segmentation performance. Our two datasets introduce pairs of speech and sound prompts coupled with matching images and segmentation masks derived from ADE20K. We create these datasets because previous works [23] have not published their datasets or evaluation code. However, we use an experimental setting from the literature for our cross-modal retrieval experiments. + +We compare against a variety of prior art including the popular state-of-the-art multi-modal retrieval network, ImageBind [17]. We also compare against CAVMAE [18], a leading multimodal backbone trained specifically for AudioSet retrieval, and DAVENet [23], which is trained to localize the meanings of words. We include two other baselines [21, 22] which have reported cross modal retrieval metrics on Places Audio. Finally, we compare our multi-head aggregation strategy to common "global" retrieval methods such as inner products between classtokens, average-pooled tokens, and SimPooled[44] tokens. We note that SimPool achieves state-of-the-art localization results when compared to 14 other pooling methods. Nevertheless, our multi-head aligner yields better localization results than any of these "global" methods. + +# 4.1. Qualitative Comparison of Feature Maps + +Our first experiment in Figure 2 highlights the dramatic differences in quality between DenseAV's features and other approaches in the literature. DenseAV is the only backbone whose local tokens are semantically meaningful and show cross-modal alignment for speech and sound. Though both CAVMAE and ImageBind show high-quality retrieval performance, neither shows high quality aligned local tokens. As a result, DenseAV can associate and localize both sound and language significantly better than other backbones. DAVENet shows coarse correspondences between language and visual objects but cannot associate sound with visual objects and does not match DenseAV's high resolution maps. Furthermore, the right half of Figure 1 demon + +
MethodPlaces Acc. @10AudioSet Acc. @10
I → AA → II → AA → I
[22]*46.3%54.8%--
[21]*54.2%56.4%--
DAVENet [23]*52.8%60.4%--
CAVMAE [18]81.7%77.7%55.7%50.7%
ImageBind [17]1.10%1.10%64.5%66.5%
Ours94.2%94.3%68.2%68.4%
+ +Table 2. Cross-modal retrieval using 1000 evaluation videos from the PlacesAudio and AudioSet validation datasets. DenseAV dramatically outperforms all approaches tested in all metrics. Most notably, the state-of-the-art image retrieval foundation model, ImageBind, is incapable of recognizing speech. We note that the ImageBind authors do not publish retraining code, so we evaluate their largest pretrained model. Models with a * indicate that they have been previously reported in the literature. Other numbers are calculated by using pretrained models when available or from training with the author's official training scripts. + +strates that DenseAV naturally discovers and separates word semantics from the sound of objects without labels to supervise this separation. In the supplement, we provide additional visualizations of all backbones considered across a wide range of words and sounds. + +# 4.2. Speech Prompted Image Segmentation + +Dataset: We introduce a speech prompted segmentation dataset using the ADE20K dataset, which is known for its comprehensive ontology and pixel-precise annotations [59]. From this dataset, we curate an evaluation subset of image-class pairs by sampling up to 10 images for each object class in ADE20K, excluding images where the selected class was tiny ( $< 5\%$ of pixels). We only consider classes with at least 2 images that pass the tiny object criterion. For each class and image, we formed a binary target mask by selecting the semantic segmentation mask for that class. This resulted in 3030 image-object pairs spanning 478 ADE20K classes. + +We created paired speech signals by speaking the prompt "A picture of a(n) [object]" where [object] is the name of the ADE20K class. We create clear, controlled, and consistent audio prompts using Microsoft's neural text to speech service [47]. This service also provides exact timing of the "[object]" utterance within the broader prompt and ensures each class is measured equally. Grammar was manually verified for the utterances to ensure proper singular/plural and a/an agreement with the class name. We release images, masks, and audio prompts for reproducibility. + +Evaluation Measure: We evaluate methods based on how well their speech-prompted activations align with ground truth masks for the visual object's class. We quantify this with the binary Average Precision (AP) and Intersection over Union (IoU) metrics. These quantify how close activations match with the binary label mask from the ADE20K dataset. To compute an aggregate score over all of the object + +
MethodPred. Dis.Act. Dis.
No LDis, No Head Max Pool64.1%70.3%
No LDis99.9%86.5%
Ours99.9%91.2%
+ +Table 3. Quantitative ablation study of the impact of max-pooling attention heads and adding our disentanglement loss, $\mathcal{L}_{Dis}$ . Intuitively, max-pooling attention heads allows each head to specialize on its own specific set of triggers. Our disentanglement loss further encourages the heads to operate independently and orthogonally. + +classes considered, we compute the mean average precision (mAP) and mean intersection over union (mIoU) by averaging AP scores across all object categories considered. + +The mAP is particularly well suited for evaluating feature similarities because it is unaffected by monotonic transformations of the similarity scores. This eliminates the need for arbitrary thresholding and calibration. This is particularly important because many networks' inner products are not centered at zero, and the best thresholding strategy can be nontrivial, and dependent on the network and object class. Average Precision avoids these confounding factors and ensures a fair comparison across methods. Unfortunately, unlike the mAP, the mIoU metric requires selecting a threshold. To ensure our mIoU measurement is similarly invariant to monotonic transformations we evaluate 20 uniformly spaced thresholds between the smallest and largest activations of each model. For each baseline, we report results for the best threshold to ensure a fair comparison between all networks considered. + +Implementation: We compute image heatmaps by evaluating each modality-specific network on the image-audio pairs from our dataset. We extract dense features from the final layer of each network and form their similarity volume according to Equation 1. For DenseAV we max-pool the head dimension to properly compare with single-headed models. We average activations over the temporal extent of the "[object]" utterance using the word timing information from the ground truth audio clip. This creates a heatmap over the image features that can be bi-linearly resized to the original image's size. We then compare these per-pixel activation scores to ground truth object masks from our dataset. + +Results: In Speech mAP and mIoU columns of Table 1 we show that DenseAV achieves a $51\%$ $(+16.5\mathrm{mAP})$ relative increase in speech-prompted semantic segmentation over previous methods. Approaches that use global token based contrastive strategies such as CAVMAE and Image-Bind perform particularly poorly in this task, and this observation aligns with the qualitative results of Figure 2. + +# 4.3. Sound Prompted Image Segmentation + +Dataset: To evaluate how well deep features localize sound, we build on Section 4.2 and create a dataset of sound prompts that align with ADE20K classes. We first select the same (large) image-object pairs from ADE20K. We then + +
MethodSpeech mAPPlaces Acc. @10
V → AA → V
Average Pool20.1%92.0%91.2%
CLS Token20.6%86.4%89.8%
SimPool [44]35.3%92.6%92.8%
Multi-Head (Ours)48.2%93.5%93.8%
+ +Table 4. Quantitative ablation of different feature aggregation strategies. Though the common practice of average pooling and using a learned CLS token to aggregate features have little effect on retrieval performance, they dramatically degrade performance on speech prompted semantic segmentation. + +create a mapping between the ADE20K and VGGSound [6] ontologies. To compute a robust mapping, we first embed ADE20K class names and VGGSound class names with the GPT Ada 2 text embedding model [37]. For each ADE20K class, we create a list of at most three candidates from the VGGSound ontology that have a cosine similarity $(> .85)$ . We then manually review these candidates to select the best VGGSound class for each ADE20K class and remove any spurious or mistaken matches. This produces a set of 95 ADE20K classes with strong matches in the VGGSound ontology. For each of our original 3030 image-object pairs we select a random VGGSound validation clip with a matching class according to our mapped ontology. This yields 106 image-object pairs across 20 ADE20K classes. + +Evaluation Measure: We use the same mAP and mIoU evaluation metrics as Section 4.2, but instead average over the 20 ADE20K classes considered. + +Implementation: We compute sound prompted image activations as in section 4.2 but with one key change: we average activations over the entire clip because we do not have ground-truth sound timing information. + +Results: The "Sound mAP and mIoU" columns of Table 1 show that DenseAV achieves a $25\%$ (+6.4mAP) relative improvement in sound prompted segmentation compared to the prior art. Most notably, ImageBind's features cannot localize sound despite their high cross-modal retrieval performance learned from millions of hours of sound. + +# 4.4. Cross-Modal Retrieval + +We show that DenseAV's representations are not only better for localization, but significantly outperform other approaches on cross-modal retrieval. We adopt the evaluation setting of [23] and measure cross modal retrieval accuracy at 1, 5, and 10 in a thousand-way retrieval task. In particular, we use the same thousand images from the validation set of [23] and also replicate this analysis on one-thousand random clips from the AudioSet validation data. Table 2 shows results for 1000-way retrieval tasks on both the Places Audio and AudioSet datasets. We show cross-modal accuracy at 10, but also show larger tables in the supplement that echo these results using accuracy at 1 and 5. DenseAV significantly outperforms all baselines across all metrics. Interest- + +ingly, DenseAV outperforms ImageBind with less than half of the trainable parameters and no reliance on text. + +# 4.5. Measuring Disentanglement + +We observe that DenseAV's heads naturally learn to differentiate audio-visual couplings that capture the meaning of words (language) and those that capture the sounds of objects (sound). Furthermore this effect generalizes to novel clips, including those with both sound and language as shown in Figure 1. We quantify this observation in two ways, the first measures if a head's average activation strength predicts whether a clip contains mainly "language" or "sound". The second method quantifies how often the "sound" head is incorrectly active when the "language" head should be active and vice versa. We leverage the fact that AudioSet dataset contains mostly clips with ambient sound and rarely contains language. In contrast, Places Audio is entirely language-based without external ambient sound. We note that these analyses are specifically for our architecture with two heads $K = 2$ and trained on both AudioSet and PlacesAudio data. + +For both measures of disentanglement, we first compute a clip's aggregated similarity for each head. In particular, we remove the max-pooling over heads in Equation 2 to create a single-head similarity, $S(a,v)_k$ . We then min-max scale the scores of each head across both datasets to lie in the [0,1] interval, which we refer to as $\hat{S}(a,v)_k$ . Using these normalized scores, we can create metrics that capture how well a given head responds only to a specific dataset. + +Our first metric measures how well a head's scores predict whether a clip is from the "sound" or "language" dataset. Let $(a_{b},v_{b})_{1}^{B}$ be tuples of paired audio and visual signals. Let $l[k^{\prime}]_{b}$ be an indicator variable of whether the signal $(a_{b},v_{b})$ arises from the sound dataset, AudioSet, $(k^{\prime} = 1)$ , or the language dataset Places Audio $(k^{\prime} = 2)$ . + +$$ +\delta_ {p r e d} (k, k ^ {\prime}) = \mathrm {A P} \left(\left(\hat {\mathcal {S}} \left(a _ {b}, v _ {b}\right) _ {k}\right) _ {1} ^ {B}, \left(l \left[ k ^ {\prime} \right] _ {b}\right) _ {1} ^ {B}\right) \tag {6} +$$ + +Where $AP(\cdot, \cdot)$ is the binary average precision with prediction and label arguments respectively. Intuitively, this measures whether the scores of head $k$ are direct predictors of whether the data is from dataset $k'$ . We can find the best assignment between heads and datasets such that each head is maximally predictive of the given dataset: + +$$ +\begin{array}{l} \mathrm {P r e d D i s} = \frac {1}{2} \max \left(\delta_ {p r e d} (0, 0) + \delta_ {p r e d} (1, 1), \right. \\ \delta_ {p r e d} (1, 0) + \delta_ {p r e d} (0, 1)) \tag {7} \\ \end{array} +$$ + +The prediction disentanglement score, PredDis, is a percentage that ranges from $50\%$ for completely entangled signals to $100\%$ if one can perfectly classify the signals using the scores of either head. The maximum over the two possible assignments makes this metric invariant to permutations + +of the heads. We note that this metric is a Hungarian matching assignment [29] over two entries, a common technique to asses unsupervised classification performance [20, 28]. + +Our second measure quantifies "spurious activations" in the non-dominant head. A truly disentangled system should have a head that only fires on sound, and another head that only fires on language. We create another disentanglement measure, ActDis, by replacing $\delta_{pred}$ in Equation 7 with: + +$$ +\delta_ {a c t} \left(k, k ^ {\prime}\right) = 1 - \frac {1}{\sum_ {b ^ {\prime}} l \left[ k ^ {\prime} \right] _ {b ^ {\prime}}} \sum_ {b = 1} ^ {B} \hat {\mathcal {S}} \left(a _ {b}, v _ {b}\right) _ {k} \cdot l \left[ k ^ {\prime} \right] _ {b} \tag {8} +$$ + +Intuitively, this measures the "inactivity" of head $k$ on dataset $k'$ . If head $k$ is totally silent on dataset $k'$ then $\delta_{act}(k, k') = 1$ . Like PredDis, ActDis is a percentage ranging from $50\%$ to $100\%$ with $100\%$ representing perfect disentanglement where the sound head is completely silent during the language clips, and vice versa. + +Table 3 shows that DenseAV achieves near perfect predictive (99%) and activation (91%) disentanglement. It also shows that our disentanglement regularizer and max-pooling over heads improves DenseAV's natural ability to distinguish sound from language without supervision. + +# 5. Conclusion + +We presented DenseAV, a novel contrastive learning architecture that can discover the meaning of words and localize the sounds of objects using only video supervision. We are the first to observe both qualitatively and quantitatively that it's possible to disentangle the meaning of words from the sound of objects with only a contrastive learning signal. DenseAV's success stems from its novel multi-head attention aggregation mechanism that encourages its modality-specific backbones to create high-resolution, semantically meaningful, and AV aligned representations. These properties of DenseAV's representation are not seen in other state-of-the-art models in the literature. Consequently, DenseAV significantly surpasses other leading models in dense prediction tasks such as speech and sound-prompted semantic segmentation as well as in cross-modal retrieval. + +# Acknowledgements + +This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation. This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/). This work is funded by a Royal Society Research Professorship RSRP\R\241003, and EPSRC Programme Grant VisualAI EP/T028572/1. + +# References + +[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020. 3 +[2] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 2, 3, 4 +[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4 +[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 4 +[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 7 +[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 3 +[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 3 +[9] Jeongsoo Choi, Joanna Hong, and Yong Man Ro. Diffv2s: Diffusion-based video-to-speech synthesis with vision-guided speaker embedding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7812-7821, 2023. 3 +[10] Noam Chomsky. Language and problems of knowledge: The Managua lectures. MIT press, 1987. 1 +[11] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 3 +[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023.4 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 + +[14] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2 +[15] Nicholas Frosst, Nicolas Papernot, and Geoffrey Hinton. Analyzing and improving representations with the soft nearest neighbor loss. In International conference on machine learning, pages 2012-2020. PMLR, 2019. 4 +[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5 +[17] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3,4,6 +[18] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 6 +[19] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 4 +[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 2, 3, 8 +[21] David Harwath and James R Glass. Learning word-like units from joint audio-visual analysis. arXiv preprint arXiv:1701.07481, 2017. 3, 6 +[22] David Harwath, Antonio Torralba, and James Glass. Unsupervised learning of spoken language with visual context. Advances in Neural Information Processing Systems, 29, 2016. 5, 6 +[23] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649-665, 2018. 2, 3, 4, 6, 7 +[24] Arthur E Hoerl and Robert W Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55-67, 1970. 5 +[25] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 4 +[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. + +Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[27] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 3 +[28] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9865-9874, 2019. 3, 8 +[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 8 +[30] Neeraj Kumar, Srishti Goel, Ankur Narang, and Mujtaba Hasan. Robust one shot audio to video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 770-771, 2020. 3 +[31] Shiguang Liu, Sijia Li, and Haonan Cheng. Towards an end-to-end visual-to-raw-audio generation with gan. IEEE Transactions on Circuits and Systems for Video Technology, 32(3): 1299–1312, 2021. 3 +[32] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song. Active contrastive learning of audio-visual video representations. arXiv preprint arXiv:2009.09805, 2020. 3 +[33] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3 +[34] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2 +[35] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871–14881, 2021. 3 +[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4 +[37] OpenAI. Gpt-4 technical report, 2023. 7 +[38] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Is object localization for free?-weakly-supervised learning with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 685-694, 2015. 2 +[39] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4 + +[40] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4 +[41] Alex S Park and James R Glass. Unsupervised pattern discovery in speech. IEEE Transactions on Audio, Speech, and Language Processing, 16(1):186-197, 2007. 3 +[42] Puyuan Peng and David Harwath. Self-supervised representation learning for speech using visual grounding and masked language modeling. arXiv preprint arXiv:2202.03543, 2022. 3 +[43] Puyuan Peng and David Harwath. Word discovery in visually grounded, self-supervised speech models. arXiv preprint arXiv:2203.15081, 2022. 3 +[44] Bill Psomas, Ioannis Kakogeorgiou, Konstantinos Karantzalos, and Yannis Avrithis. Keep it simpool: Who said supervised transformers suffer from attention deficit? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5350-5360, 2023. 4, 6, 7 +[45] Geoffrey K Pullum and Barbara C Scholz. Empirical assessment of stimulus poverty arguments. The linguistic review, 19(1-2):9-50, 2002. 1 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[47] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv preprint arXiv:2006.04558, 2020. 6 +[48] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5 +[49] Yi-Jen Shih, Hsuan-Fu Wang, Heng-Jui Chang, Layne Berry, Hung-yi Lee, and David Harwath. Speechclip: Integrating speech with pre-trained vision and language model. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 715-722. IEEE, 2023. 2, 3, 4 +[50] Linda Smith and Chen Yu. Infants rapidly learn word-referent mappings via cross-situational statistics. Cognition, 106(3):1558-1568, 2008. 1 +[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 3 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 4 +[53] Luyu Wang and Aaron van den Oord. Multi-format contrastive learning of audio representations. arXiv preprint arXiv:2103.06508, 2021. 3 +[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on + +the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 4 +[55] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 2 +[56] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5 +[57] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European conference on computer vision (ECCV), pages 570-586, 2018. 3 +[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2 +[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6 +[60] Hao Zhu, Man-Di Luo, Rui Wang, Ai-Hua Zheng, and Ran He. Deep audio-visual learning: A survey. International Journal of Automation and Computing, 18:351–376, 2021. 3 \ No newline at end of file diff --git a/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/images.zip b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03294f10542560aabc95c0556dfdc570f4cf8290 --- /dev/null +++ b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dc327ac5e67843c62d2149732549f97dfa8a229d0eb80020ff6829b00f42f26 +size 444899 diff --git a/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/layout.json b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e23e80e0d4e1f7b325d8330b0c3779ee1404489a --- /dev/null +++ b/2024/Separating the 'Chirp' from the 'Chat'_ Self-supervised Visual Grounding of Sound and Language/layout.json @@ -0,0 +1,7786 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 477, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 477, + 140 + ], + "type": "text", + "content": "Separating the \"Chirp\" from the \"Chat\": Self-supervised Visual Grounding of Sound and Language" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 161, + 160, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 161, + 160, + 202 + ], + "spans": [ + { + "bbox": [ + 81, + 161, + 160, + 202 + ], + "type": "text", + "content": "Mark Hamilton \nMIT, Microsoft \nmarkth@mit.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 161, + 279, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 161, + 279, + 190 + ], + "spans": [ + { + "bbox": [ + 186, + 161, + 279, + 190 + ], + "type": "text", + "content": "Andrew Zisserman Oxford, Google" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 161, + 386, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 161, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 386, + 190 + ], + "type": "text", + "content": "John R. Hershey Google" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 414, + 161, + 511, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 161, + 511, + 190 + ], + "spans": [ + { + "bbox": [ + 414, + 161, + 511, + 190 + ], + "type": "text", + "content": "William T. Freeman \nMIT, Google" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 52, + 220, + 548, + 373 + ], + "blocks": [ + { + "bbox": [ + 52, + 220, + 548, + 373 + ], + "lines": [ + { + "bbox": [ + 52, + 220, + 548, + 373 + ], + "spans": [ + { + "bbox": [ + 52, + 220, + 548, + 373 + ], + "type": "image", + "image_path": "bbd6229aa8445331f3b0e4631e0a77170938c732de2277a7475a59f813380723.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 378, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 547, + 411 + ], + "type": "text", + "content": "Figure 1. Visual overview of the DenseAV algorithm. Two modality-specific backbones featurize audio and visual signals. We introduce a novel generalization of multi-head attention to extract attention maps that discover and separate the \"meaning\" of spoken words and the sounds an object makes. DenseAV performs this localization and decomposition solely through observing paired stimuli such as videos." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "spans": [ + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 448, + 290, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 290, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 290, + 700 + ], + "type": "text", + "content": "We present DenseAV, a novel dual encoder grounding architecture that learns high-resolution, semantically meaningful, and audio-visual aligned features solely through watching videos. We show that DenseAV can discover the \"meaning\" of words and the \"location\" of sounds without explicit localization supervision. Furthermore, it automatically discovers and distinguishes between these two types of associations without supervision. We show that DenseAV's localization abilities arise from a new multi-head feature aggregation operator that directly compares dense image and audio representations for contrastive learning. In contrast, many other systems that learn \"global\" audio and video representations cannot localize words and sound. Finally, we contribute two new datasets to improve the evaluation of AV representations through speech and sound prompted semantic segmentation. On these and other datasets we show DenseAV dramatically outperforms the prior art on speech and sound prompted semantic segmentation. DenseAV outperforms the current state-of-the-art, ImageBind, on cross-modal retrieval using fewer than half of the parameters. Project Page: https://aka.ms/denseav" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 422, + 386, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 386, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 386, + 434 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 445, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 625 + ], + "type": "text", + "content": "Associating audio and video is a fundamental aspect of human perception. As infants develop, the synchronization and correspondence of sounds enables multi-modal association – a voice with a face, and a 'moo' with a cow [50]. Later, as they acquire language, they associate spoken words with objects they represent [10, 45]. Amazingly, these association abilities, constituting speech recognition, sound event recognition, and visual object recognition, develop without much direct supervision. This work aims to create a model with this capability by learning high-resolution, semantically meaningful, audio-visual (AV) aligned representations. Features with these properties can be used to discover fine-grained correspondences between modalities without localization supervision or knowledge of the textual representation of language." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "As an example, consider the spoken caption and accompanying sounds of the image shown in Figure 1. We wish to \"ground\" both the speech and the sounds with high resolution. For instance, both the spoken word \"dog\" and the sound of a bark in the audio signal should be associated with the pixels of the dog in the visual signal if present. We seek high quality local representations where simple inner" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13117" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 102, + 145, + 214 + ], + "blocks": [ + { + "bbox": [ + 48, + 102, + 145, + 214 + ], + "lines": [ + { + "bbox": [ + 48, + 102, + 145, + 214 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 145, + 214 + ], + "type": "image", + "image_path": "90aafa0b6605adfbec68b597cef114d2f07e1b8e2f7999d23fcfcce7dcfbb642.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 159, + 56, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 159, + 56, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 159, + 56, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 159, + 56, + 545, + 259 + ], + "type": "image", + "image_path": "b2f1046456c605eb34400f3e15b8b967034ac0ec032c9d9eabe6a9e8c7ddb2ab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 263, + 546, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 263, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 546, + 319 + ], + "type": "text", + "content": "Figure 2. Qualitative comparison of several modern architectures for associating audio and video modalities. Only DenseAV learns a high-resolution and semantically aligned set of local features. This allows us to perform speech and sound prompted semantic segmentation using only the inner products between deep features. Other approaches, such as ImageBind, do not show aligned local feature maps. Approaches that do show some localization capabilities, like DAVENet, do not generalize to sound and language, and do not achieve the high-resolution localization capabilities of DenseAV. Dense features are visualized using PCA as in [20]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 332, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 287, + 357 + ], + "type": "text", + "content": "products between features exhibit this behavior, which is interestingly absent from popular approaches in the literature." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 589 + ], + "type": "text", + "content": "To achieve this, we make three innovations. First, we introduce DenseAV, a dual-encoder architecture that computes a dense similarity volume over audio and visual features before aggregating this volume into a single similarity score. If we look at a slice of this similarity volume when word is spoken, as in Figure 1, we can visualize the AV activation strength between a word or sound and an image's pixels. The novelty we introduce is to generalize this dense similarity mechanism with multiple heads, much like those of multi-head attention. This generalization allows each head to specialize on particular types of couplings between the visual and audio modalities. Interestingly, we discover that if we give DenseAV two heads and train on a dataset that contains both language and sound, the heads naturally learn to distinguish language from more general sound using only cross-modality supervision. For example, as shown in Figure 1, head 1 focuses on what objects create a sound, like the barking of a dog, whereas head 2 focuses only on the meaning of words." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Second, we show the importance of the \"aggregation function\" one uses to create a similarity score between an audio clip and a video frame for contrastive learning. The traditional choices; using inner products between global representations such as class tokens [5, 13, 49] or pooled features [18, 58], do not promote AV alignment of dense local features. Because of this, several popular audio-video backbones that excel on cross-modal retrieval cannot directly associate objects and sounds using their local features. This limits their ability to be used for downstream" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 333, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 333, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 545, + 357 + ], + "type": "text", + "content": "tasks such as semantic segmentation, sound localization, or unsupervised language learning and discovery." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 360, + 546, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 528 + ], + "type": "text", + "content": "Third, we introduce two semantic segmentation datasets to evaluate visual grounding with AV representations for speech and (non-speech) sounds. We build these datasets from the high-quality segmentation masks provided by the ADE20K dataset [59] and measure mean average precision (mAP) and mean intersection over union (mIoU) on a binary mask prediction task. This evaluation is simpler and more thorough than previous efforts to measure visual grounding such as the concept counting metrics of [23] and the \"pointing games\" of [2, 14, 38] that only check if a heatmap's peak occurs within a target box or segment. Furthermore, our evaluation avoids brittle word-net ontologies [34], clustering, Wu and Palmer distance [55], threshold choices, and a variety of other complicating factors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 530, + 533, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 530, + 533, + 541 + ], + "spans": [ + { + "bbox": [ + 318, + 530, + 533, + 541 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 551, + 545, + 706 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 551, + 544, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 544, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 544, + 574 + ], + "type": "text", + "content": "- We introduce DenseAV, a novel self-supervised architecture that learns high-resolution AV correspondences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 574, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 545, + 621 + ], + "type": "text", + "content": "- We introduce a local-feature-based image similarity function that significantly improves a network's zero-shot localization ability compared to common strategies such as average pooling or CLS tokens." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 623, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 545, + 669 + ], + "type": "text", + "content": "- We introduce new datasets for evaluating speech and sound prompted semantic segmentation. We show DenseAV significantly outperforms the current state-of-the-art on these tasks as well as on cross-modal retrieval." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 545, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 545, + 706 + ], + "type": "text", + "content": "- We discover that our multi-head architecture naturally disentangles audio-visual correspondence into sound and language components using only contrastive supervision." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13118" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 56, + 521, + 217 + ], + "blocks": [ + { + "bbox": [ + 91, + 56, + 521, + 217 + ], + "lines": [ + { + "bbox": [ + 91, + 56, + 521, + 217 + ], + "spans": [ + { + "bbox": [ + 91, + 56, + 521, + 217 + ], + "type": "image", + "image_path": "039a8c44a90011de72a3148624ab418a96b3fc9f8ab983fa4374d1d4b7feeec0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 288 + ], + "type": "text", + "content": "Figure 3. Architectural overview of our multi-head attention aggregator. Dense feature maps are split into " + }, + { + "bbox": [ + 46, + 220, + 547, + 288 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 220, + 547, + 288 + ], + "type": "text", + "content": " heads (" + }, + { + "bbox": [ + 46, + 220, + 547, + 288 + ], + "type": "inline_equation", + "content": "K = 1,2" + }, + { + "bbox": [ + 46, + 220, + 547, + 288 + ], + "type": "text", + "content": ") in our experiments. We form an AV activation tensor by taking the inner-products of each head's features across the spatial and temporal extent of the visual and audio signals respectively as in Equation 1. We then aggregate this similarity volume into a single similarity score by max-pooling head and spatial dimensions and average-pooling audio dimensions. Our approach aims to encourage the network to identify specific shared objects between the audio and visual modalities. In particular, max-pooling of heads disentangles sound and language, and max-pooling spatial dimensions helps localize objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 298, + 134, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 134, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 134, + 312 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 289, + 624 + ], + "type": "text", + "content": "Audio-visual (AV), text-visual, and more general multimodal models have recently surged in popularity [60]. Broadly speaking DenseAV is an audio-video contrastive learning architecture, this class of methods learns AV representations by aligning paired signals and pushing apart negative signals [11, 27]. Of the models in this class, several stand out for their ability to localize sounds [2, 7, 42] or capture the semantics of language [23, 43]. Many models in this class compare AV signals using inner products between \"global\" representations formed by pooled deep features [18, 35, 53], or class tokens [17, 32, 42, 43, 49]. Most notably, ImageBind has gained popularity due to its state-of-the-art performance on a variety of tasks and datasets and unified class-token-based contrastive architecture. In this work we show that many of these architectures do not show strong localization properties in their local features, despite excelling at cross-modal retrieval on a \"global\" level. This limits their applicability to new out-of-domain sounds, sounds that don't have a textual representation, and low-resource languages. We diverge from these works by directly supervising local tokens. In particular, we build on previous works [2, 23] that show max-pooling improves localization capabilities and introduce a new multi-head aggregation operator that generalizes previous losses using an operator similar to self-attention [52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Another class of methods discover structure in signals through uni- and multi-modal clustering. Early works on audio clustering [41] discovered meaningful utterances without supervision. Similar visual analyses have discovered visual objects [4, 8, 20, 28]. Recent works have applied these ideas to the AV domain [1, 21], but do not focus on extracting high-resolution AV representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 300, + 547, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 398 + ], + "type": "text", + "content": "Finally, several works investigate generative audio-video learning. The Sound of Pixels [57] generates the sound of a specific object using a source separation loss. Newer approaches using GANs [30, 31], and diffusion models [9, 17, 33] have generated audio from video and vice versa. Here we focus on improving the local representations of contrastive learners because of their relative scalability, simplicity, and ability to learn high-quality representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 406, + 367, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 367, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 367, + 419 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "content": "At a high level, DenseAV tries to determine when a given audio and visual signal belong “together” using dense audio-visual representations. To perform this task robustly, DenseAV must learn how to predict the contents of an audio signal from a visual signal and vice versa. This causes DenseAV to learn dense modality-specific features that capture the mutual information shared between the modalities [51]. Once learned, we can directly query these informative features to perform speech and sound prompted semantic segmentation as illustrated in Figure 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "text", + "content": "More specifically, DenseAV is built from two modality-specific deep featurizers. These backbones produce temporally varying audio features across an audio clip and spatially varying video features for a single randomly selected frame. Our loss computes a similarity between audio and visual signals based on the intuition that two signals are similar if they have a variety of strong couplings or shared objects. More formally, we form a scalar similarity for a pair of audio and video signals by carefully aggregating a volume of pairwise inner products between dense features. We use the InfoNCE [36] contrastive loss to encourage similarity between \"positive\" pairs of signals and dissimilarity between \"negative\" pairs formed by in-batch shuffling. Figure 3 graphically depicts this loss function and subsequent" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13119" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 253, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 253, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 253, + 83 + ], + "type": "text", + "content": "sections detail each component of our architecture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 94, + 264, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 94, + 264, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 94, + 264, + 106 + ], + "type": "text", + "content": "3.1. Multi-Headed Aggregation of Similarities" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 114, + 287, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 114, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 114, + 287, + 257 + ], + "type": "text", + "content": "DenseAV's key architectural distinction is its loss function that directly supervises the \"local\" tokens of the visual and audio featurizers. This is a significant departure from other works [5, 17, 19, 39, 46, 49] that pool modality specific information into \"global\" representations prior to the contrastive loss. Unlike prior works, our loss function aggregates the full pairwise similarities between the local tokens into an aggregate measure of similarity for a given pair of audio and visual signals. We show in Figure 2 that this architectural choice enables DenseAV's local features to align across modalities whereas other approaches such as average pooling, class tokens, and SimPool [44] do not." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 617 + ], + "type": "text", + "content": "We first describe our loss function informally and definite it more precisely in the next paragraph. Our loss function computes the (un-normalized) inner product between every pair of visual and audio features to form a \"volume\" of inner products. This volume represents how strongly each part of an audio signal \"couples\" to each part of a visual signal. We aim to find many large couplings between positive pairs of audio and visual signals. Ideally, these couplings should connect visual objects with their references in the audio signal. Conversely, we do not want to find couplings between negative pairs of signals. To compute a single global coupling strength for a pair of signals, we aggregate this volume of pairwise similarities into a single number. There are myriad ways to aggregate this volume ranging in \"softness\" from average-pooling to max-pooling. Average pooling yields dense gradients and can improve convergence speed and stability. However, max-pooling allows the network to focus on the best couplings regardless the object's size or a sound's duration. Our aggregation function combines the benefits of average and max pooling by max-pooling visual dimensions and average pooling audio dimensions as proposed in [23]. Intuitively speaking, this averages the strongest image couplings over an audio signal. It allows small visual objects to have large effects yet provides a strong training gradient to many regions of the signals. Finally, we draw inspiration from multi-head self-attention [52] and generalize this operation to multiple \"heads\" that we max-pool before pooling the visual and audio dimensions. This allows DenseAV to discover multiple \"ways\" to associate objects across modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "More formally, let " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "S(a,v) \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": " represent the similarity between a tensor of audio features " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{R}^{CKFT}" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": " of size (Channel × K-heads × Frequency × Time) and a tensor of visual features " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^{CKHW}" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": " of size (Channel × K-heads × Height × Width). To define this scalar similarity score, we first create a local similarity volume, " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "s(a,v) \\in \\mathbb{R}^{kfhw}" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": ". For simplicity, we consider the aggregated similarity between a single image and audio clip but note one can easily general-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "ize this to max-pool over video-frames. We define the full pairwise volume of similarities as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 99, + 545, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 99, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 320, + 99, + 545, + 131 + ], + "type": "interline_equation", + "content": "s (a, v) \\in \\mathbb {R} ^ {k f t h w} = \\sum_ {c = 1} ^ {C} a [ c, k, f, t ] \\cdot v [ c, k, h, w ] \\tag {1}", + "image_path": "db7c5ff302b3df7f34f7ef008906edfcde356d0aded95b66018ed6954af8774c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "inline_equation", + "content": "a[c,k,f,t]" + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "content": " represents the value of " + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "content": " at location " + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "inline_equation", + "content": "[c,k,f,t]" + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "content": " is scalar multiplication. We aggregate this similarity volume into a single score " + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "inline_equation", + "content": "S(a,v)\\in \\mathbb{R}" + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 176, + 545, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 176, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 176, + 545, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (a, v) = \\frac {1}{F T} \\sum_ {f = 1} ^ {F} \\sum_ {t = 1} ^ {T} \\max _ {k, h, w} (s (a, v) [ k, f, t, h, w ]) \\tag {2}", + "image_path": "fcc15c8d844f6cb63760b24bbd6b6f5c8cd9cbe20a7817c9f920d19c8934b5de.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 216, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 251 + ], + "type": "text", + "content": "We note that this operation can be viewed as a multi-head generalization of the MISA loss of [23], and a multi-head multi-time generalization of the MIL loss of [2]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 258, + 350, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 350, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 350, + 269 + ], + "type": "text", + "content": "3.2. Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": "We can use the similarity between audio and visual signals defined in Equation 2 to construct a contrastive loss. We follow recent works [15, 17, 54] and use the temperature-weighted InfoNCE [36] to encourage similarity between positive pairs of signals and dissimilarity between negative pairs. In DenseAV, we form " + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": " positive pairs by splitting the audio and visual components of a Batch of training data. We form " + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "inline_equation", + "content": "B^2 - B" + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": " negative pairs by comparing a signal to all of the other signals in the training batch. More formally let " + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "inline_equation", + "content": "(a_b, v_b)_1^B" + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 276, + 545, + 407 + ], + "type": "text", + "content": " pairs of audio and visual signals. The visual-retrieval term of our InfoNCE loss is then:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 411, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A \\rightarrow V} = \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {\\exp \\left(\\gamma \\mathcal {S} \\left(a _ {b} , v _ {b}\\right)\\right)}{\\sum_ {b ^ {\\prime} = 1} ^ {B} \\exp \\left(\\gamma \\mathcal {S} \\left(a _ {b} , v _ {b ^ {\\prime}}\\right)\\right)}\\right) \\tag {3}", + "image_path": "c442caadf75a3c490a5b60573f6c8ffee451a7489aea53fa892f8eb49629af89.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "text", + "content": " is a trainable inverse temperature parameter. We symmetrize this loss by adding the analogous audio-retrieval term, " + }, + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{V\\to A}" + }, + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "text", + "content": ", which iterates over negative audio signals in the denominator." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 503, + 465, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 465, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 465, + 515 + ], + "type": "text", + "content": "3.3. Audio and Visual Featurizers" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "The core of DenseAV is two modality-specific backbone networks. We use the DINO vision transformer [5] with ImageNet pretrained weights (without labels) to provide a strong, yet fully unsupervised, vision backbone. Unlike other approaches that use CLIP [46] as a backbone, DINO does not require paired text captions and learns from unlabeled images only. Practically, we find that DINO outperforms CLIP because of its better-behaved local tokens [12], an effect we explore in the Supplement. We append an additional layer norm operation across the channel dimension [3] and a " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " Convolution to DINO. The layer-norm and " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " convolution ensure the architecture does not start with a saturated loss function. We use the HuBERT audio transformer [25] as DenseAV's audio backbone. HubERT operates on waveforms and is trained on the LibriSpeech [40] dataset using only self-supervision. Hubert" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13120" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "outputs a single feature per frequency, corresponding to " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F = 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " in Section 3. Though HuBERT was only trained on speech, its audio features can be fine-tuned for more general sounds, much like how vision backbones can be fine-tuned for new datasets [56]. As in the visual branch, we append a channel-wise LayerNorm block and two " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " convolutions to the audio branch. These layers help the network avoid saturation and speed convergence. Furthermore, the two convolutions help the model aggregate information, which reduces the cost of the pairwise feature comparison used in our loss function. We refer to these added layers after the pretrained backbones as the \"aligners\" in later sections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 222, + 129, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 222, + 129, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 129, + 234 + ], + "type": "text", + "content": "3.4. Regularizers" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "text", + "content": "Disentanglement Regularizer, " + }, + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Dis}}" + }, + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "text", + "content": ": We add a small regularization term to encourage each head of Equation 1 to specialize and learn independent types of audio-visual associations. Interestingly we find that our 2-head model naturally learns to distinguish the meaning of words with one head and capture the sounds objects produce with another head. To further encourage this unsupervised discovery of concepts, we penalize the network when multiple attention heads are simultaneously active. More precisely, let " + }, + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "inline_equation", + "content": "(a_{b}, v_{b})_{1}^{B}" + }, + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "text", + "content": " be a Batch of " + }, + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 239, + 287, + 371 + ], + "type": "text", + "content": " paired audio and visual signals. Our disentanglement loss for two heads is then:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 376, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 376, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 80, + 376, + 287, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D i s} = \\operatorname {M e a n} \\left(\\left| s \\left(a _ {b}, v _ {b}\\right) [ 1 ] \\circ s \\left(a _ {b}, v _ {b}\\right) [ 2 ] \\right|\\right) \\tag {4}", + "image_path": "10bd5cb8600740f477cdb6849e1df731321205fdc3e14c348151877a66cc52c4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": " is elementwise multiplication and " + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": " is the elementwise absolute value function. " + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "inline_equation", + "content": "[k]" + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": " mirrors PyTorch slicing notation and refers to selecting the activations for only the " + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": "th attention head. Intuitively, this loss encourages one head to be silent if the other head is active and is a \"cross-term\" generalization of the " + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "inline_equation", + "content": "l^2" + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": " regularizer [24] for encouraging activation shrinkage. When " + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "inline_equation", + "content": "K > 2" + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": " we average contributions from every combination of heads. We ablate this, and our decision to max-pool heads in Table 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "text", + "content": "Stability Regularizers, " + }, + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {Stability }}" + }, + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "text", + "content": ": Finally, we add several other small regularization terms to encourage stable convergence. We detail and ablate these terms in the Supplement. Briefly, these terms include standard regularizers like Total Variation [48] smoothness over time and non-negative pressure to encourage the network to focus on similarity instead of dissimilarity. In addition, we add a regularizer to prevent the calibration temperature, " + }, + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "text", + "content": ", from drifting too quickly, and a regularizer to discourage activations during silence and noise. In the supplement we show that each regularizer alone does not have a dramatic effect on final metrics but together they can stop collapses during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 643, + 286, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 643, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 59, + 643, + 286, + 655 + ], + "type": "text", + "content": "Combining these losses into a single loss function yields:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 659, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 659, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 63, + 659, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {A \\rightarrow V} + \\mathcal {L} _ {V \\rightarrow A} + \\lambda_ {\\text {D i s}} \\mathcal {L} _ {\\text {D i s}} + \\mathcal {L} _ {\\text {S t a b i l i t y}} \\tag {5}", + "image_path": "529a9262978f9a83810e856d24dffcdb4f941c14e4a3a0b80f00334bccc97ff0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "In our experiments we use " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{Dis} = 0.05" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " and refer interested readers to the supplement for the details of our small stability regularizer, " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Stability}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 370, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 370, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 370, + 84 + ], + "type": "text", + "content": "3.5. Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "type": "text", + "content": "In our experiments we train DenseAV and relevant baselines on the AudioSet [16] dataset for sound prompted segmentation and AudioSet retrieval. We train on the PlacesAudio [22] dataset for speech prompted segmentation, PlacesAudio retrieval, and the ablation studies of Table 4. In our disentanglement experiments of Table 3 and feature visualizations of Figures 1 and 2 we train on both AudioSet and PlacesAudio so that DenseAV can be familiar with both language, the prominent audio signal in PlacesAudio, and more general sounds from AudioSet. In these experiments we sample training data from these two corpora, so each batch has an even split between AudioSet and PlacesAudio." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 235, + 546, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 546, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 546, + 450 + ], + "type": "text", + "content": "Warming up Aligners: We find that we can dramatically improve the stability by first training the added aligners (convolutions and layer norms) for 3000 steps while keeping pretrained DINO and HuBERT backbones fixed. This allows the aligners to adapt to these intelligent backbones before modifying each backbone's sensitive weights. We use random resize crops, color jitter, random flips, and random greyscale as image augmentations. We randomly sample a single video frame to feed to our visual branch. Audio clips are converted to single-channel format and are trimmed or padded with silence to create uniform 10 second clips. We re-sample audio clips according to the requirements of the backbone models used. For HuBERT, we re-sample to 16KhZ. We train on 8 V100 GPUs with an effective batch size of 80, and aggregate negative samples on all GPUs prior to computing the loss to ensure efficient parallelization. We provide additional training information and hyperparameters in the supplement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 451, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 559 + ], + "type": "text", + "content": "Full Training: After warming up the aligners, we train the full model for an additional 800,000 steps using the same loss, batch-size, and training logic. We train all aligner weights and fine-tune all HuBERT audio backbone weights. We use low rank adaptation (LoRA) [26] to fine-tune the \"Q\", \"K\", and \"V\" layers of the DINO visual backbone attention blocks. This allows us to efficiently adapt DINO and stabilize the training as it is quite easy to collapse the carefully trained DINO weights. We use a LoRA rank of 8." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 572, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 388, + 586 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 593, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 676 + ], + "type": "text", + "content": "To evaluate AV representation quality, we perform a variety of analyses including comparative activation visualization, quantitative measurements of speech and sound prompted semantic segmentation, and cross-modal retrieval. Additionally, we quantify our observation that DenseAV can distinguish the meanings of words (language), from the sounds of objects (sound) without supervision." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To adequately measure a representation's AV alignment quality, we found it necessary to introduce two evaluation datasets that measure speech and sound prompted seman" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13121" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 53, + 284, + 137 + ], + "blocks": [ + { + "bbox": [ + 48, + 53, + 284, + 137 + ], + "lines": [ + { + "bbox": [ + 48, + 53, + 284, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 53, + 284, + 137 + ], + "type": "table", + "html": "
MethodSpeech Semseg.Sound Semseg.
mAPmIoUmAPmIoU
DAVENet [23]32.2%26.3%16.8%17.0%
CAVMAE [18]27.2%19.9%26.0%20.5%
ImageBind [17]20.2%19.7%18.3%18.1%
Ours48.7%36.8%32.4%25.5%
", + "image_path": "111fd3a0658521944c7dfd36b0db85af127eddf0a4f27519c9e3f261def3281f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 138, + 286, + 247 + ], + "lines": [ + { + "bbox": [ + 47, + 138, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 286, + 247 + ], + "type": "text", + "content": "Table 1. Speech and Sound prompted semantic segmentation. We analyze the quality of local features using two prompted semantic segmentation tasks. We prompt networks with speech of the form \"a picture of a(n) [Object]\" to determine whether local feature inner products can segment objects in the ADE20K dataset by name. We create sound prompts for a given ADE20K class using a curated mapping from the ADE20K ontology to the VG-GSound ontology. DenseAV's local features perform significantly better than all baselines investigated. We bold \"first place\" results and underline \"second place\" results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 264, + 286, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 286, + 347 + ], + "type": "text", + "content": "tic segmentation performance. Our two datasets introduce pairs of speech and sound prompts coupled with matching images and segmentation masks derived from ADE20K. We create these datasets because previous works [23] have not published their datasets or evaluation code. However, we use an experimental setting from the literature for our cross-modal retrieval experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 349, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 286, + 529 + ], + "type": "text", + "content": "We compare against a variety of prior art including the popular state-of-the-art multi-modal retrieval network, ImageBind [17]. We also compare against CAVMAE [18], a leading multimodal backbone trained specifically for AudioSet retrieval, and DAVENet [23], which is trained to localize the meanings of words. We include two other baselines [21, 22] which have reported cross modal retrieval metrics on Places Audio. Finally, we compare our multi-head aggregation strategy to common \"global\" retrieval methods such as inner products between classtokens, average-pooled tokens, and SimPooled[44] tokens. We note that SimPool achieves state-of-the-art localization results when compared to 14 other pooling methods. Nevertheless, our multi-head aligner yields better localization results than any of these \"global\" methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 262, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 262, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 262, + 552 + ], + "type": "text", + "content": "4.1. Qualitative Comparison of Feature Maps" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 557, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 286, + 713 + ], + "type": "text", + "content": "Our first experiment in Figure 2 highlights the dramatic differences in quality between DenseAV's features and other approaches in the literature. DenseAV is the only backbone whose local tokens are semantically meaningful and show cross-modal alignment for speech and sound. Though both CAVMAE and ImageBind show high-quality retrieval performance, neither shows high quality aligned local tokens. As a result, DenseAV can associate and localize both sound and language significantly better than other backbones. DAVENet shows coarse correspondences between language and visual objects but cannot associate sound with visual objects and does not match DenseAV's high resolution maps. Furthermore, the right half of Figure 1 demon" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 53, + 559, + 160 + ], + "blocks": [ + { + "bbox": [ + 307, + 53, + 559, + 160 + ], + "lines": [ + { + "bbox": [ + 307, + 53, + 559, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 53, + 559, + 160 + ], + "type": "table", + "html": "
MethodPlaces Acc. @10AudioSet Acc. @10
I → AA → II → AA → I
[22]*46.3%54.8%--
[21]*54.2%56.4%--
DAVENet [23]*52.8%60.4%--
CAVMAE [18]81.7%77.7%55.7%50.7%
ImageBind [17]1.10%1.10%64.5%66.5%
Ours94.2%94.3%68.2%68.4%
", + "image_path": "bf4f4798533e7d28ef8b41b6594644822d3acb60ad6035e64e20f371a92ad8c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 163, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 274 + ], + "type": "text", + "content": "Table 2. Cross-modal retrieval using 1000 evaluation videos from the PlacesAudio and AudioSet validation datasets. DenseAV dramatically outperforms all approaches tested in all metrics. Most notably, the state-of-the-art image retrieval foundation model, ImageBind, is incapable of recognizing speech. We note that the ImageBind authors do not publish retraining code, so we evaluate their largest pretrained model. Models with a * indicate that they have been previously reported in the literature. Other numbers are calculated by using pretrained models when available or from training with the author's official training scripts." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 293, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 353 + ], + "type": "text", + "content": "strates that DenseAV naturally discovers and separates word semantics from the sound of objects without labels to supervise this separation. In the supplement, we provide additional visualizations of all backbones considered across a wide range of words and sounds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 360, + 509, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 509, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 509, + 373 + ], + "type": "text", + "content": "4.2. Speech Prompted Image Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "type": "text", + "content": "Dataset: We introduce a speech prompted segmentation dataset using the ADE20K dataset, which is known for its comprehensive ontology and pixel-precise annotations [59]. From this dataset, we curate an evaluation subset of image-class pairs by sampling up to 10 images for each object class in ADE20K, excluding images where the selected class was tiny (" + }, + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "type": "inline_equation", + "content": "< 5\\%" + }, + { + "bbox": [ + 304, + 378, + 545, + 510 + ], + "type": "text", + "content": " of pixels). We only consider classes with at least 2 images that pass the tiny object criterion. For each class and image, we formed a binary target mask by selecting the semantic segmentation mask for that class. This resulted in 3030 image-object pairs spanning 478 ADE20K classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "content": "We created paired speech signals by speaking the prompt \"A picture of a(n) [object]\" where [object] is the name of the ADE20K class. We create clear, controlled, and consistent audio prompts using Microsoft's neural text to speech service [47]. This service also provides exact timing of the \"[object]\" utterance within the broader prompt and ensures each class is measured equally. Grammar was manually verified for the utterances to ensure proper singular/plural and a/an agreement with the class name. We release images, masks, and audio prompts for reproducibility." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Measure: We evaluate methods based on how well their speech-prompted activations align with ground truth masks for the visual object's class. We quantify this with the binary Average Precision (AP) and Intersection over Union (IoU) metrics. These quantify how close activations match with the binary label mask from the ADE20K dataset. To compute an aggregate score over all of the object" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13122" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 52, + 282, + 103 + ], + "blocks": [ + { + "bbox": [ + 50, + 52, + 282, + 103 + ], + "lines": [ + { + "bbox": [ + 50, + 52, + 282, + 103 + ], + "spans": [ + { + "bbox": [ + 50, + 52, + 282, + 103 + ], + "type": "table", + "html": "
MethodPred. Dis.Act. Dis.
No LDis, No Head Max Pool64.1%70.3%
No LDis99.9%86.5%
Ours99.9%91.2%
", + "image_path": "126e8acdd47b6ae4004f1548bbb9f626573c0065461e079eb132026f839e9bb0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 104, + 287, + 160 + ], + "lines": [ + { + "bbox": [ + 46, + 104, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 46, + 104, + 287, + 160 + ], + "type": "text", + "content": "Table 3. Quantitative ablation study of the impact of max-pooling attention heads and adding our disentanglement loss, " + }, + { + "bbox": [ + 46, + 104, + 287, + 160 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Dis}" + }, + { + "bbox": [ + 46, + 104, + 287, + 160 + ], + "type": "text", + "content": ". Intuitively, max-pooling attention heads allows each head to specialize on its own specific set of triggers. Our disentanglement loss further encourages the heads to operate independently and orthogonally." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 217 + ], + "type": "text", + "content": "classes considered, we compute the mean average precision (mAP) and mean intersection over union (mIoU) by averaging AP scores across all object categories considered." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 218, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 287, + 409 + ], + "type": "text", + "content": "The mAP is particularly well suited for evaluating feature similarities because it is unaffected by monotonic transformations of the similarity scores. This eliminates the need for arbitrary thresholding and calibration. This is particularly important because many networks' inner products are not centered at zero, and the best thresholding strategy can be nontrivial, and dependent on the network and object class. Average Precision avoids these confounding factors and ensures a fair comparison across methods. Unfortunately, unlike the mAP, the mIoU metric requires selecting a threshold. To ensure our mIoU measurement is similarly invariant to monotonic transformations we evaluate 20 uniformly spaced thresholds between the smallest and largest activations of each model. For each baseline, we report results for the best threshold to ensure a fair comparison between all networks considered." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "type": "text", + "content": "Implementation: We compute image heatmaps by evaluating each modality-specific network on the image-audio pairs from our dataset. We extract dense features from the final layer of each network and form their similarity volume according to Equation 1. For DenseAV we max-pool the head dimension to properly compare with single-headed models. We average activations over the temporal extent of the \"[object]\" utterance using the word timing information from the ground truth audio clip. This creates a heatmap over the image features that can be bi-linearly resized to the original image's size. We then compare these per-pixel activation scores to ground truth object masks from our dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 553, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 638 + ], + "type": "text", + "content": "Results: In Speech mAP and mIoU columns of Table 1 we show that DenseAV achieves a " + }, + { + "bbox": [ + 46, + 553, + 287, + 638 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 46, + 553, + 287, + 638 + ], + "type": "inline_equation", + "content": "(+16.5\\mathrm{mAP})" + }, + { + "bbox": [ + 46, + 553, + 287, + 638 + ], + "type": "text", + "content": " relative increase in speech-prompted semantic segmentation over previous methods. Approaches that use global token based contrastive strategies such as CAVMAE and Image-Bind perform particularly poorly in this task, and this observation aligns with the qualitative results of Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 248, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 248, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 248, + 659 + ], + "type": "text", + "content": "4.3. Sound Prompted Image Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Dataset: To evaluate how well deep features localize sound, we build on Section 4.2 and create a dataset of sound prompts that align with ADE20K classes. We first select the same (large) image-object pairs from ADE20K. We then" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 52, + 547, + 127 + ], + "blocks": [ + { + "bbox": [ + 307, + 52, + 547, + 127 + ], + "lines": [ + { + "bbox": [ + 307, + 52, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 52, + 547, + 127 + ], + "type": "table", + "html": "
MethodSpeech mAPPlaces Acc. @10
V → AA → V
Average Pool20.1%92.0%91.2%
CLS Token20.6%86.4%89.8%
SimPool [44]35.3%92.6%92.8%
Multi-Head (Ours)48.2%93.5%93.8%
", + "image_path": "68af70bc9db37516f62cc94dd99261f502ef13c7be73b11b9d79a5e99b965220.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 129, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 547, + 184 + ], + "type": "text", + "content": "Table 4. Quantitative ablation of different feature aggregation strategies. Though the common practice of average pooling and using a learned CLS token to aggregate features have little effect on retrieval performance, they dramatically degrade performance on speech prompted semantic segmentation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 206, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 373 + ], + "type": "text", + "content": "create a mapping between the ADE20K and VGGSound [6] ontologies. To compute a robust mapping, we first embed ADE20K class names and VGGSound class names with the GPT Ada 2 text embedding model [37]. For each ADE20K class, we create a list of at most three candidates from the VGGSound ontology that have a cosine similarity " + }, + { + "bbox": [ + 304, + 206, + 545, + 373 + ], + "type": "inline_equation", + "content": "(> .85)" + }, + { + "bbox": [ + 304, + 206, + 545, + 373 + ], + "type": "text", + "content": ". We then manually review these candidates to select the best VGGSound class for each ADE20K class and remove any spurious or mistaken matches. This produces a set of 95 ADE20K classes with strong matches in the VGGSound ontology. For each of our original 3030 image-object pairs we select a random VGGSound validation clip with a matching class according to our mapped ontology. This yields 106 image-object pairs across 20 ADE20K classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 374, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 374, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 545, + 409 + ], + "type": "text", + "content": "Evaluation Measure: We use the same mAP and mIoU evaluation metrics as Section 4.2, but instead average over the 20 ADE20K classes considered." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 411, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 458 + ], + "type": "text", + "content": "Implementation: We compute sound prompted image activations as in section 4.2 but with one key change: we average activations over the entire clip because we do not have ground-truth sound timing information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 459, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 545, + 529 + ], + "type": "text", + "content": "Results: The \"Sound mAP and mIoU\" columns of Table 1 show that DenseAV achieves a " + }, + { + "bbox": [ + 305, + 459, + 545, + 529 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 459, + 545, + 529 + ], + "type": "text", + "content": " (+6.4mAP) relative improvement in sound prompted segmentation compared to the prior art. Most notably, ImageBind's features cannot localize sound despite their high cross-modal retrieval performance learned from millions of hours of sound." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 539, + 435, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 435, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 435, + 551 + ], + "type": "text", + "content": "4.4. Cross-Modal Retrieval" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "We show that DenseAV's representations are not only better for localization, but significantly outperform other approaches on cross-modal retrieval. We adopt the evaluation setting of [23] and measure cross modal retrieval accuracy at 1, 5, and 10 in a thousand-way retrieval task. In particular, we use the same thousand images from the validation set of [23] and also replicate this analysis on one-thousand random clips from the AudioSet validation data. Table 2 shows results for 1000-way retrieval tasks on both the Places Audio and AudioSet datasets. We show cross-modal accuracy at 10, but also show larger tables in the supplement that echo these results using accuracy at 1 and 5. DenseAV significantly outperforms all baselines across all metrics. Interest-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13123" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 97 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 97 + ], + "type": "text", + "content": "ingly, DenseAV outperforms ImageBind with less than half of the trainable parameters and no reliance on text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 103, + 203, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 203, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 203, + 117 + ], + "type": "text", + "content": "4.5. Measuring Disentanglement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 324 + ], + "type": "text", + "content": "We observe that DenseAV's heads naturally learn to differentiate audio-visual couplings that capture the meaning of words (language) and those that capture the sounds of objects (sound). Furthermore this effect generalizes to novel clips, including those with both sound and language as shown in Figure 1. We quantify this observation in two ways, the first measures if a head's average activation strength predicts whether a clip contains mainly \"language\" or \"sound\". The second method quantifies how often the \"sound\" head is incorrectly active when the \"language\" head should be active and vice versa. We leverage the fact that AudioSet dataset contains mostly clips with ambient sound and rarely contains language. In contrast, Places Audio is entirely language-based without external ambient sound. We note that these analyses are specifically for our architecture with two heads " + }, + { + "bbox": [ + 46, + 121, + 287, + 324 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 121, + 287, + 324 + ], + "type": "text", + "content": " and trained on both AudioSet and PlacesAudio data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "text", + "content": "For both measures of disentanglement, we first compute a clip's aggregated similarity for each head. In particular, we remove the max-pooling over heads in Equation 2 to create a single-head similarity, " + }, + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "inline_equation", + "content": "S(a,v)_k" + }, + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "text", + "content": ". We then min-max scale the scores of each head across both datasets to lie in the [0,1] interval, which we refer to as " + }, + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\hat{S}(a,v)_k" + }, + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "text", + "content": ". Using these normalized scores, we can create metrics that capture how well a given head responds only to a specific dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "text", + "content": "Our first metric measures how well a head's scores predict whether a clip is from the \"sound\" or \"language\" dataset. Let " + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "inline_equation", + "content": "(a_{b},v_{b})_{1}^{B}" + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "text", + "content": " be tuples of paired audio and visual signals. Let " + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "inline_equation", + "content": "l[k^{\\prime}]_{b}" + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "text", + "content": " be an indicator variable of whether the signal " + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "inline_equation", + "content": "(a_{b},v_{b})" + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "text", + "content": " arises from the sound dataset, AudioSet, " + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "inline_equation", + "content": "(k^{\\prime} = 1)" + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "text", + "content": ", or the language dataset Places Audio " + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "inline_equation", + "content": "(k^{\\prime} = 2)" + }, + { + "bbox": [ + 47, + 421, + 287, + 494 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 502, + 287, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 502, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 73, + 502, + 287, + 522 + ], + "type": "interline_equation", + "content": "\\delta_ {p r e d} (k, k ^ {\\prime}) = \\mathrm {A P} \\left(\\left(\\hat {\\mathcal {S}} \\left(a _ {b}, v _ {b}\\right) _ {k}\\right) _ {1} ^ {B}, \\left(l \\left[ k ^ {\\prime} \\right] _ {b}\\right) _ {1} ^ {B}\\right) \\tag {6}", + "image_path": "9b98a0cc71e5621aa572bac56c5eb587a29fbe64b1bc44852f5fd6d1f441fa87.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "AP(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " is the binary average precision with prediction and label arguments respectively. Intuitively, this measures whether the scores of head " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " are direct predictors of whether the data is from dataset " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "k'" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": ". We can find the best assignment between heads and datasets such that each head is maximally predictive of the given dataset:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 607, + 287, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 607, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 607, + 287, + 645 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {P r e d D i s} = \\frac {1}{2} \\max \\left(\\delta_ {p r e d} (0, 0) + \\delta_ {p r e d} (1, 1), \\right. \\\\ \\delta_ {p r e d} (1, 0) + \\delta_ {p r e d} (0, 1)) \\tag {7} \\\\ \\end{array}", + "image_path": "a63c92460f9bc76f1ea04433f634c8662b373c71de204f5192038c08c4990130.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "The prediction disentanglement score, PredDis, is a percentage that ranges from " + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": " for completely entangled signals to " + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": " if one can perfectly classify the signals using the scores of either head. The maximum over the two possible assignments makes this metric invariant to permutations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "of the heads. We note that this metric is a Hungarian matching assignment [29] over two entries, a common technique to asses unsupervised classification performance [20, 28]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": "Our second measure quantifies \"spurious activations\" in the non-dominant head. A truly disentangled system should have a head that only fires on sound, and another head that only fires on language. We create another disentanglement measure, ActDis, by replacing " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\delta_{pred}" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": " in Equation 7 with:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 170, + 545, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 170, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 315, + 170, + 545, + 202 + ], + "type": "interline_equation", + "content": "\\delta_ {a c t} \\left(k, k ^ {\\prime}\\right) = 1 - \\frac {1}{\\sum_ {b ^ {\\prime}} l \\left[ k ^ {\\prime} \\right] _ {b ^ {\\prime}}} \\sum_ {b = 1} ^ {B} \\hat {\\mathcal {S}} \\left(a _ {b}, v _ {b}\\right) _ {k} \\cdot l \\left[ k ^ {\\prime} \\right] _ {b} \\tag {8}", + "image_path": "2e05a87a639a90bb4b16c19a167f214011f358a82ff9377eb34dd158d8704e15.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": "Intuitively, this measures the \"inactivity\" of head " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": " on dataset " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "k'" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": ". If head " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": " is totally silent on dataset " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "k'" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\delta_{act}(k, k') = 1" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": ". Like PredDis, ActDis is a percentage ranging from " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 207, + 545, + 279 + ], + "type": "text", + "content": " representing perfect disentanglement where the sound head is completely silent during the language clips, and vice versa." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 279, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 339 + ], + "type": "text", + "content": "Table 3 shows that DenseAV achieves near perfect predictive (99%) and activation (91%) disentanglement. It also shows that our disentanglement regularizer and max-pooling over heads improves DenseAV's natural ability to distinguish sound from language without supervision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 348, + 378, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 378, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 378, + 361 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 369, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 548 + ], + "type": "text", + "content": "We presented DenseAV, a novel contrastive learning architecture that can discover the meaning of words and localize the sounds of objects using only video supervision. We are the first to observe both qualitatively and quantitatively that it's possible to disentangle the meaning of words from the sound of objects with only a contrastive learning signal. DenseAV's success stems from its novel multi-head attention aggregation mechanism that encourages its modality-specific backbones to create high-resolution, semantically meaningful, and AV aligned representations. These properties of DenseAV's representation are not seen in other state-of-the-art models in the literature. Consequently, DenseAV significantly surpasses other leading models in dense prediction tasks such as speech and sound-prompted semantic segmentation as well as in cross-modal retrieval." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 557, + 408, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 408, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 408, + 569 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. 2021323067. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation. This work is supported by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/). This work is funded by a Royal Society Research Professorship RSRP\\R\\241003, and EPSRC Programme Grant VisualAI EP/T028572/1." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13124" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "type": "text", + "content": "[2] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 212 + ], + "type": "text", + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "text", + "content": "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 368 + ], + "type": "text", + "content": "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "text", + "content": "[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "text", + "content": "[9] Jeongsoo Choi, Joanna Hong, and Yong Man Ro. Diffv2s: Diffusion-based video-to-speech synthesis with vision-guided speaker embedding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7812-7821, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 557 + ], + "type": "text", + "content": "[10] Noam Chomsky. Language and problems of knowledge: The Managua lectures. MIT press, 1987. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 612 + ], + "type": "text", + "content": "[11] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "type": "text", + "content": "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023.4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[15] Nicholas Frosst, Nicolas Papernot, and Geoffrey Hinton. Analyzing and improving representations with the soft nearest neighbor loss. In International conference on machine learning, pages 2012-2020. PMLR, 2019. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "text", + "content": "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "text", + "content": "[17] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3,4,6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "type": "text", + "content": "[18] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "type": "text", + "content": "[19] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 545, + 453 + ], + "type": "text", + "content": "[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 2, 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 455, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 486 + ], + "type": "text", + "content": "[21] David Harwath and James R Glass. Learning word-like units from joint audio-visual analysis. arXiv preprint arXiv:1701.07481, 2017. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "text", + "content": "[22] David Harwath, Antonio Torralba, and James Glass. Unsupervised learning of spoken language with visual context. Advances in Neural Information Processing Systems, 29, 2016. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 589 + ], + "type": "text", + "content": "[23] David Harwath, Adria Recasens, Dídac Surís, Galen Chuang, Antonio Torralba, and James Glass. Jointly discovering visual objects and spoken words from raw sensory input. In Proceedings of the European conference on computer vision (ECCV), pages 649-665, 2018. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 590, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 622 + ], + "type": "text", + "content": "[24] Arthur E Hoerl and Robert W Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55-67, 1970. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 689 + ], + "type": "text", + "content": "[25] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13125" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "text", + "content": "[27] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1): 2, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 196 + ], + "type": "text", + "content": "[28] Xu Ji, João F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9865-9874, 2019. 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 232 + ], + "type": "text", + "content": "[29] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 235, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 288 + ], + "type": "text", + "content": "[30] Neeraj Kumar, Srishti Goel, Ankur Narang, and Mujtaba Hasan. Robust one shot audio to video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 770-771, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "type": "text", + "content": "[31] Shiguang Liu, Sijia Li, and Haonan Cheng. Towards an end-to-end visual-to-raw-audio generation with gan. IEEE Transactions on Circuits and Systems for Video Technology, 32(3): 1299–1312, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 369 + ], + "type": "text", + "content": "[32] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song. Active contrastive learning of audio-visual video representations. arXiv preprint arXiv:2009.09805, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "text", + "content": "[33] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 418, + 286, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 286, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 286, + 439 + ], + "type": "text", + "content": "[34] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 441, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 506 + ], + "type": "text", + "content": "[35] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871–14881, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "text", + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 216, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 216, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 216, + 555 + ], + "type": "text", + "content": "[37] OpenAI. Gpt-4 technical report, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[38] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Is object localization for free?-weakly-supervised learning with convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 685-694, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 712 + ], + "type": "text", + "content": "[39] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[40] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[41] Alex S Park and James R Glass. Unsupervised pattern discovery in speech. IEEE Transactions on Audio, Speech, and Language Processing, 16(1):186-197, 2007. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 162, + 545, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 203 + ], + "type": "text", + "content": "[42] Puyuan Peng and David Harwath. Self-supervised representation learning for speech using visual grounding and masked language modeling. arXiv preprint arXiv:2202.03543, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 237 + ], + "type": "text", + "content": "[43] Puyuan Peng and David Harwath. Word discovery in visually grounded, self-supervised speech models. arXiv preprint arXiv:2203.15081, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[44] Bill Psomas, Ioannis Kakogeorgiou, Konstantinos Karantzalos, and Yannis Avrithis. Keep it simpool: Who said supervised transformers suffer from attention deficit? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5350-5360, 2023. 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "text", + "content": "[45] Geoffrey K Pullum and Barbara C Scholz. Empirical assessment of stimulus poverty arguments. The linguistic review, 19(1-2):9-50, 2002. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 392 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "type": "text", + "content": "[47] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv preprint arXiv:2006.04558, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "text", + "content": "[48] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "type": "text", + "content": "[49] Yi-Jen Shih, Hsuan-Fu Wang, Heng-Jui Chang, Layne Berry, Hung-yi Lee, and David Harwath. Speechclip: Integrating speech with pre-trained vision and language model. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 715-722. IEEE, 2023. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 525, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 557 + ], + "type": "text", + "content": "[50] Linda Smith and Chen Yu. Infants rapidly learn word-referent mappings via cross-situational statistics. Cognition, 106(3):1558-1568, 2008. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 559, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 612 + ], + "type": "text", + "content": "[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[53] Luyu Wang and Aaron van den Oord. Multi-format contrastive learning of audio representations. arXiv preprint arXiv:2103.06508, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[54] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13126" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 353 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "text", + "content": "[55] Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. arXiv preprint cmp-lg/9406033, 1994. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 119, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 161 + ], + "type": "text", + "content": "[56] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "text", + "content": "[57] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European conference on computer vision (ECCV), pages 570-586, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[58] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "text", + "content": "[60] Hao Zhu, Man-Di Luo, Rui Wang, Ai-Hua Zheng, and Ran He. Deep audio-visual learning: A survey. International Journal of Automation and Computing, 18:351–376, 2021. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13127" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_content_list.json b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f87a62d51dbbf50b0449b6d67e8405dd385ebe6b --- /dev/null +++ b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_content_list.json @@ -0,0 +1,2273 @@ +[ + { + "type": "text", + "text": "Sequential Modeling Enables Scalable Learning for Large Vision Models", + "text_level": 1, + "bbox": [ + 117, + 130, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yutong Bai $^{1,2,*}$", + "bbox": [ + 173, + 179, + 292, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alan L. Yuille", + "bbox": [ + 179, + 198, + 302, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyang Geng $^{1,*}$", + "bbox": [ + 328, + 180, + 464, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trevor Darrell", + "bbox": [ + 334, + 198, + 460, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Karttikeya Mangalam", + "bbox": [ + 493, + 180, + 673, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jitendra Malik1", + "bbox": [ + 493, + 198, + 620, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amir Bar", + "bbox": [ + 709, + 180, + 792, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexei A. Efros", + "bbox": [ + 656, + 198, + 787, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UC Berkeley (BAIR)", + "bbox": [ + 246, + 232, + 424, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Johns Hopkins University", + "bbox": [ + 501, + 232, + 718, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, \"visual sentences\", in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions without needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion tokens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time.", + "bbox": [ + 73, + 318, + 473, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 560, + 209, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) such as GPT [10] and LLaMA [80] have taken the world by storm. What would it take to build a Large Vision Model (LVM)? From the animal world, we know that visual competences are not dependent on language. In particular, many experiments have shown that the visual world of non-human primates is remarkably similar to that of humans. So while the space of vision-language models such as LLaVA [53] is interesting and worthwhile to pursue, in this paper we seek an answer to a different question – how far can we go from pixels alone?", + "bbox": [ + 73, + 585, + 470, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key features of contemporary LLMs that we seek to emulate in LVMs are: 1) scaling in the presence of big data, and 2) flexible specification of tasks through prompting (in-context learning). How do we achieve this? As usual, there are three main components that must be specified:", + "bbox": [ + 75, + 738, + 468, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data: We want to exploit all the remarkable diversity in visual data. First of all, just raw unannotated images and videos. Next, we want to exploit the variety of annotated visual data sources that have been produced over the last", + "bbox": [ + 75, + 814, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "couple of decades – semantic segmentations, depth reconstructions, keypoints, multiple views of 3D objects, among others. We define a common format, “visual sentences”, in which to represent these different annotations without needing any meta-knowledge beyond the pixels. The total size of our training dataset is 1.64 billion images/frames.", + "bbox": [ + 496, + 287, + 893, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Architecture: We use a large transformer architecture (3 billion parameters) trained on visual data represented as sequence of tokens, using a learned tokenizer that maps each image to a string of 256 vector-quantized tokens.", + "bbox": [ + 496, + 378, + 893, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Loss function: We draw inspiration from the natural language community, where masked token modeling has given way to sequential autoregressive prediction. Once images/videos/annotated images can all be represented as sequences, we can train the model to minimize the cross-entropy loss for predicting the next token.", + "bbox": [ + 496, + 440, + 893, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With this extremely simple design, we demonstrate some noteworthy behaviors:", + "bbox": [ + 498, + 532, + 890, + 561 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Appropriate scaling behavior as one increases model size and data size.", + "- Many different vision tasks can now be \"solved\" by designing suitable prompts at test time. While the results don't show as high performance as bespoke, specifically-trained models, the fact that so many tasks are all addressed by a single vision model is quite encouraging.", + "- We see a clear benefit of the amount of unsupervised data on the performance on various standard vision tasks.", + "- We see a hint of an ability for general visual reasoning – handling out-of-distribution data, and performing novel tasks. But further investigation is needed." + ], + "bbox": [ + 500, + 563, + 893, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 757, + 640, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pretrained Vision Models. The value of using pretrained models (such as ImageNet-pretrained AlexNet [45]) has been demonstrated as far back as 2015 in R-CNN [34], and it has since become standard practice in computer vision. Self-supervised pretraining was proposed as a way to vastly increase the amount of data available for pretraining [16, 25, 37, 61, 62, 99]. Unfortunately, this was not very successful, likely because the CNN-based architectures", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. Webpage: https://yutongbai.com/lvm.html.", + "bbox": [ + 94, + 886, + 437, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22861", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of that time did not have enough capacity to absorb the data. With the introduction of Transformers [84], which have much higher capacity, researchers revisited self-supervised pretraining, and showed that transformer-based masked image reconstruction approaches, such as BEiT [7], MAE [38], SimMIM [91], MaskGIT [13] perform vastly better than their CNN-based counterparts [62]. Yet, despite their recent successes, current pretrained vision-only models have had trouble scaling up to the really large datasets, such as LAION [72].", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-task Learning and In-context Learning. From the classic one-model-per-task setups, computer vision is slowly moving toward a single model performing multiple different tasks. Various multi-task learning approaches [24, 40, 43, 73, 97] exist but they are typically limited to a fixed, predefined number of tasks. More recently, methods inspired by in-context learning in LLMs forgo any notion of tasks and instead let the model infer the task directly from the input prompt. For example, Visual Prompting [8, 87] takes in a task input/output example pair and a query image at test time, concatenates them into a single 2-by-2 image, and uses inpainting to generate the desired output. But, since the inpainting is performed using a variant of MAE [38], the same problems with scaling are inherited by these approaches.", + "bbox": [ + 75, + 246, + 472, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Auto-regressive Visual Models. The idea of using autoregressive models for synthesizing visual data goes back at least 70 years. Inspired by Shannon's use of $N$ -grams to synthesize language [74, 75], a number of works, starting with Attneave's seminal 1954 paper [5], applied this idea to sequentially synthesizing pixels [28, 31, 39, 64], image patches [27], video frames [69], and motion capture data [4, 44, 48]. As deep models became popular, newer works replaced $N$ -grams with RNNs or CNNs for pixel [81, 82] and video [66] synthesis. Most recently, transformer-based autoregressive visual generation methods have been proposed [15, 29, 94, 96], and, combined with language, have demonstrated impressive image synthesis results, e.g. Parti [95].", + "bbox": [ + 75, + 462, + 472, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Data", + "text_level": 1, + "bbox": [ + 76, + 685, + 142, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Data! Data! Data! I can't make bricks without clay!\"", + "bbox": [ + 140, + 713, + 470, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SHERLOCK HOLMES", + "bbox": [ + 336, + 737, + 467, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key requirement of any Large Pre-trained Model is that it must be trained on vast amounts of data. For language models, very large and very diverse datasets are fairly easy to obtain. For instance, the popular Common Crawl repository [1] contains 250 billion web pages spanning the entire Web, is extremely diverse, and includes \"natural demonstrations\" like language translations, question answering, etc. In computer vision, we are still very far from having a data source of comparable size and diversity. One of the", + "bbox": [ + 75, + 763, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "central contributions of our work is the first step toward curating such a dataset that we call Unified Vision Dataset v1 (UVDv1). To assemble it, we leverage many different sources of visual data: (1) unlabelled images, (2) images with visual annotations, (3) unlabelled videos, (4) videos with visual annotations, and (5) 3D synthetic objects. The unlabeled images, which represent over $80\\%$ of our data, capture a huge cross-section of our visual world, and provide the required diversity, at the cost of lower quality. Images with annotations have a much more constrained distribution, but are usually of higher quality. Video data is even more constrained (typically, to human-centric activities), but is an invaluable source of temporal data. Renderings of 3D synthetic objects are the lowest in diversity but can provide valuable hints about the behavior of 3D structures. Importantly, UVDv1 is a purely visual dataset, with no non-visual meta-data (e.g. text) included. All together, UVDv1 contains 1.64 billion images.", + "bbox": [ + 496, + 90, + 893, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another important difference from Large Language Models is that language data has a natural, unified one-dimensional structure for all the data - a stream of text. Unfortunately, this is not the case for visual data, with different sources all having different structures. In this work we propose visual sentence as a unified unit of visual data, which enables us to train scalable models from a diverse set of sources. A visual sentence is simply a sequence containing one or more images followed by an end-of-sentence (EOS) token. Figure 1 shows how the various data sources are partitioned into visual sentences. In particular:", + "bbox": [ + 496, + 363, + 895, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single images. A single image itself represents the simplest form of a visual sentence - $\\{$ image, EOS $\\}$ . We use the filtered subset of 1.49 billion images [88] from the LAION 5B [71] dataset. This is by far the largest part of our data, comprising $88.5\\%$ .", + "bbox": [ + 496, + 534, + 893, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image sequences. A sequence of images is a natural form of visual sentence. We create such sequences by sourcing video data from a wide range of existing datasets [11, 12, 21, 35, 36, 46, 50, 51, 55, 57-59, 63, 68, 76-78, 92, 93]. Visual sentences of 16 frames are formed by randomly sampling the videos at three different strides (10, 20, and 30). In addition, we utilize synthetic 3D objects from the Objaverse Dataset [22] to generate object-centric multiview sequences for a variety of objects. For each object, we sample one radius length between the object center and the camera from 1.5 to 2.2, and sample one constant elevation from -45 degrees to 45 degrees, then traverse different views of the object by changing the azimuth with a step length of 15 degrees and render 24 views. We rendered 42000 such sequences in total for training and 8000 for testing. Finally, we can also represent images belonging to the same semantic category as being (part of) a sequence. We use categories from ImageNet, concatenating together groups of images (2,4,8, or 16) from the same category into a 16-image long", + "bbox": [ + 496, + 613, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22862", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Sentences", + "text_level": 1, + "bbox": [ + 140, + 102, + 325, + 119 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d35e2bd48e9c47f6b74a859f3b83dd27df6d33f2b69a9072df35f789704ac45b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 132, + 122, + 164 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 127, + 152, + 151, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single images, e.g. LAION", + "bbox": [ + 179, + 143, + 344, + 156 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/12d009d3676495b166dbbe08977b4b1e87c87b9f7b442350b391e4ec75091050.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 167, + 388, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 390, + 189, + 413, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/05f3f0a5ab5cd60e92de770e16445b0dd90ff843edb65f124df856ab180ca62e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 204, + 387, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 390, + 203, + 413, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0f3c7a1217f7766d48c9ed6a38f238311059d4a1871e62ee48200e5d7bbaae5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 241, + 387, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 390, + 241, + 413, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 390, + 262, + 413, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image sequences, e.g. videos, 3D rotations, synthetic viewpoints", + "bbox": [ + 455, + 189, + 581, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d222d66e84a2f4a4c2b7ee1e6eabb187cdafc56075e3b39d892f11209ffec40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 275, + 158, + 306 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4e992a3888d53e2b91b820826178570d10fcd3347accfd212ca9abcd441307a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 275, + 248, + 306 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca73993924913fd58018f80182d7af16263a9549ab3f3023ec44486a49ae395a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 275, + 334, + 306 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0fd55ba213dd6da12d30d9a8d819b347f3b159d14d2fda33de8cdfbd9721d76e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 275, + 424, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "··", + "bbox": [ + 428, + 294, + 450, + 306 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/255a7cfdc4aa8e6b5e20456b7b1cfa02fa00f8723cebe2a23cb6da69e51a5cba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 311, + 158, + 344 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd8334a170b70fcb859b027fd60a94f35a8664fe7552b4741c286bdb84ea071e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 311, + 248, + 344 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/331223a6c6d61b18c8944adeab785e30c0b8f752f8dea5a60d49242c860ece55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 311, + 336, + 344 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19e215c306c745e4bf3dc6222338abf3c8daf8c5055c01aec6e19785427b50f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 311, + 424, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 428, + 311, + 450, + 321 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ebff1d7b9c084e04ddfa542085c32a0fdaf45576c3044fd9c83a087af92163f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 349, + 158, + 380 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e1c32ea914f99cb0e4497011c685eef33b60aaf9d9634450129a8f6818dd8bfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 349, + 248, + 380 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b16445fccde8375bf0353909a4fe5598943586ce1d81606ba02528237df392dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 349, + 336, + 380 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b92ce6c2efa5252166ca3c69fab5ad501908254c7fd36d8f0300aa134c5c982a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 349, + 424, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 428, + 372, + 450, + 380 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b7369a4bf92f2b4925fd4e339ac810b52f8b79b9bb9d38e815573a5af0970412.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 386, + 158, + 417 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca16e771aa9035b03660710a61f21e0b5d2e326121d2b2fe4adc643741748ebd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 386, + 248, + 417 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1907b449fc7e84c31b86d510621f04d7a10eceb1af56844e3f91317e275becb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 386, + 336, + 417 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5002f47cd7bd159b9b70ab0c1d991b928c674f1b3d1f1a21d1f4d35912b4594.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 386, + 424, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 428, + 386, + 450, + 407 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b76b3bdf4750d333bf6fa8f8c457448dfc22a49025114f1fa0e193d516651c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 422, + 197, + 454 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f4c16f46e6a1cecc9220bf9d9301c8ea722163837abae0c25d2aa9112e22591.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 424, + 325, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "··", + "bbox": [ + 330, + 443, + 352, + 450 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/74af6dd9eecf0abfa318bb078ac44187b6409168bd6e95e0fe6e730183a10b3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 459, + 158, + 489 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99a314ff597a49d586745816c389479a64fbfeb9c08b4a72e4d604cdeaf05d92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 459, + 248, + 489 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/367b5a32941f61bcd90e211a5554233c57d110ae4269682bd8506f3b635e1f19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 460, + 336, + 489 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dbd571c50da373321b7676eaa1d62eea07e7aa01458f1498d5d41cf1d0661153.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 460, + 424, + 489 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/720368fcfc2df52da3021da686678a1727eae2a00aed10853f0001553acd417f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 460, + 450, + 489 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a27f1522d2b55bd9909c3a55fe25ec8fbf11936e417a9198239d62139a1fb0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 460, + 581, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Edges with free form notation, 1. object detection + frame segmentation", + "bbox": [ + 583, + 424, + 702, + 481 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64e2176d1c23a956a0faaa839fa2b19b57db455b8435cb260d37c462d5c1ca4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 496, + 197, + 526 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/21449481cfe6c5fc8773982e32160a2b113987b694c9fa9b7328a8f5fd2ce00c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 496, + 315, + 527 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5531978c2355c6e0f79882390f17e686bc340009960ced7cdb1d28041394ffb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 496, + 424, + 527 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/95eafdfc25b5f6a6f846f153dc77094fb8ea7d2e80f91e901928e661d4cc4a6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 496, + 553, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 558, + 532, + 704, + 563 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cfba108de1d02b91068e23bb8034bd8ef6c24066b78c1ba5fa2cf42685e518d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 493, + 733, + 563 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9860739c00b2620b4053e324d17626f67e8873b110423d67fe6bfb88790310e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 294, + 888, + 463 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d3111027ee1360470b0d542ee8894e9f968d67b26f68644f5337013ca754107.jpg", + "image_caption": [ + "Figure 1. Visual sentences allow us to format diverse vision data into the unified structure of image sequences." + ], + "image_footnote": [], + "bbox": [ + 81, + 532, + 235, + 563 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f38fa1a2215d7404b1d69f6fd93355a7b16fabd50a6a1f905b7090661c3e4982.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 534, + 385, + 563 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/22307d15ebf8d2b1b32c2d6c478fb8af744d90faf7d22338022237ef0e171f3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 532, + 553, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 558, + 532, + 704, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 705, + 563, + 728, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Videos with annotation, e.g. video segmentation", + "bbox": [ + 730, + 497, + 846, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "visual sentences.", + "bbox": [ + 76, + 588, + 189, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Images with annotations. To handle different types of image annotations in a uniform way, we choose to represent all annotations as images. Some data types, e.g. semantic segmentation maps [100], edge maps [79], depth [65] and normal images [6], are already represented this way. For others we apply tailored methods for each specific annotation type: 1) Object Detection: We create annotations by overlaying a color-coded bounding box around each object, following the methodology in [14]; 2) Human Pose: Human skeletons are rendered in pixel space, adhering to the OpenPose format, utilizing MMPose [19]; 3) Depth Estimation, Surface Normal, and Edge Detection: given ImageNet and COCO images, we generate annotations in line with the protocols from [54]. 3) Style Transfer [9], De-rain [98], De-noise [85], Low Light Enhancement [89], and Stereo Datasets [33]: These are all represented as image pairs (e.g. input/output). 4) Colorization: We convert ImageNet images to greyscale, producing image pairs. 5) Inpainting: The process involves randomly adding black-colored boxes", + "bbox": [ + 75, + 613, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in images to simulate corruption, resulting in image pairs. For all the above annotation types, we can create visual sentences by concatenating 8 image pairs of the same annotation type into a 16-image visual sentence. For datasets containing $k$ different annotations for the same image we use a different approach: for each set of $1 + k$ images (input plus $k$ annotations), we randomly select $m$ elements, where $m \\leq k + 1 \\leq 16$ . These m-tuples are then concatenated to form visual sequences.", + "bbox": [ + 496, + 588, + 893, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image sequences with annotations. When converting annotated video data (VIPSeg [56], Hand14K [30], AVA [59], JHMDB [41]) to visual sentences, we apply two complementary strategies. The first is similar to how we treat image data with paired annotations: each visual sentence is constructed by concatenating frames with their annotations - {frame1,annot1,frame2,annot2,...}. The second method involves grouping multiple frames followed by their corresponding annotations - {frame1,frame2,annot1,annot2,...}.", + "bbox": [ + 496, + 731, + 893, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a detailed summary of all the data sources, annotation type and data statistics of UVDv1 in the Appendix.", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22863", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 76, + 90, + 187, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe the design of our autoregressive Large Vision Model. Unlike text data, which naturally exhibits discrete sequential structure, it is not straightforward to model image pixels in visual sentences. In this work, we take a two-stage approach: 1) train a large visual tokenizer (which operates on individual images) to convert each image into a sequence of visual tokens; 2) train an autoregressive transformer model on visual sentences, each represented as a sequence of tokens. We summarize our approach in Figure 2.", + "bbox": [ + 75, + 114, + 470, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Image Tokenization", + "text_level": 1, + "bbox": [ + 76, + 258, + 264, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the visual sentences exhibit a sequence structure between consecutive images, we don't have such natural sequence structure within an image. Therefore, in order to apply a transformer model to images, prior works typically do one of the following: either divide the image into patches in scan-line order, and treat that as a sequence [26], or use a pre-trained image tokenizer, such as VQVAE [83] or VQGAN [29], to cluster image features into a grid of discrete tokens, which, again, are turned into a sequence in scan-line order. We adopt the latter approach since the discrete categorical output from a model naturally forms a probabilistic distribution that one can easily sample from, enabling flexible conditional generation of new images within a visual sentence.", + "bbox": [ + 75, + 281, + 470, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we employ semantic tokens generated by a VQGAN model, a concept introduced by Esser et al [29]. This framework consists of an encoding and a decoding mechanism, featuring a quantization layer that assigns input images to a sequence of discrete tokens from an established codebook. Our encoders and decoders are constructed purely with convolutional layers. The encoder is equipped with several downsampling modules to contract the spatial dimension of the input, whereas the decoder is fitted with an equivalent series of upsampling modules to restore the image to its initial size. For a given image, our VQGAN tokenizer produces 256 discrete tokens.", + "bbox": [ + 75, + 493, + 470, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to note that our tokenizer operates on individual images independently, rather than on the entire visual sentence at once. This independence allows us to decouple the tokenizer training from the downstream Transformer model so that the tokenizer can be trained on a dataset of single images without having to consider the distribution of visual sentences.", + "bbox": [ + 75, + 674, + 470, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details: We adopt an off-the-shelf VQGAN architecture from Chang et al. [13]. We follow the exact configuration in Chang et al. [13], which uses a down-sampling factor of $f = 16$ and codebook size 8192. This means that for an image of size $256 \\times 256$ , our VQGAN tokenizer produces $16 \\times 16 = 256$ tokens where each can take 8192 different values. We found that using the results of an ImageNet pre-trained tokenizer did not generalize well", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c645823cc553d995a033025c581a2ab17f86f35e24183231853b9e86ec5e74e4.jpg", + "image_caption": [ + "Figure 2. Architecture of LVM. We first convert individual images from a visual sentence into discrete tokens using a VQGAN encoder. The resulting tokens from all images are then concatenated into a 1D sequence, and fed into an autoregressive Transformer model to predict the next token in the sequence. The predicted visual tokens are decoded into images using the VQGAN decoder." + ], + "image_footnote": [], + "bbox": [ + 537, + 92, + 851, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "beyond ImageNet images. Therefore, we train our own tokenizer on a 1.5B subset of the LAION 5B dataset [71].", + "bbox": [ + 496, + 414, + 893, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Sequence Modeling of Visual Sentences", + "text_level": 1, + "bbox": [ + 498, + 454, + 836, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After converting images into discrete tokens with VQGAN, we treat our visual sentence as a unified sequence by concatenating the discrete tokens from multiple images into a 1D sequence. Importantly, all visual sentences are treated equally – we do not make use of any special tokens to indicate particular tasks or formats. We train a causal Transformer model with the next token prediction objective using a cross-entropy loss, similar to the standard approach for language models [10]. Training the model the same way on all visual sentences enables the model to infer the relation between images from context instead of from task- or format-specific tokens. This gives the model an opportunity to generalize to other, unseen visual sentence structures.", + "bbox": [ + 496, + 477, + 893, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details: After tokenizing each image in a visual sentence into 256 tokens, we concatenate them to form a 1D sequence of tokens. On top of the sequences of visual tokens, our Transformer model is virtually the same as an autoregressive language model, so we adopt the Transformer architecture of LLaMA [80], a popular open-source language model with widely available implementations. We use a context length of 4096 tokens, which can fit 16 images under our VQGAN tokenizer. Similar to language models, we add a [BOS] (begin of sentence) token to the beginning of each visual sentence and an [EOS] (end of sentence) token to the end, and use sequence concatenation [18] during training time to improve efficiency. We train our model on our entire UVDv1 dataset (420 billion tokens) using one epoch (simple epoch training is standard in language models to avoid poten", + "bbox": [ + 496, + 674, + 895, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22864", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ddb3d3ccef716ea90bf68423385be11ea456d862ac1dfae940f8f1562d1cd0c1.jpg", + "image_caption": [ + "Figure 3. Training loss for the 300M, 600M, 1B, and 3B models. All models are trained on 420B tokens (1.64B images). The training scales well with model sizes. Zoom in for more details: notice how the shape of the 4 curves is very similar. This is the consequence of the deterministic training pipeline, so the loss fluctuations are determined by the intrinsic difficulty of the data stream." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 467, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tial overfitting). We train 4 models with different numbers of parameters: 300 million, 600 million, 1 billion and 3 billion, following the same training configurations. We provide the detailed training hyperparameters in Appendix 5.5.", + "bbox": [ + 75, + 412, + 472, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Inference by Visual Prompting", + "text_level": 1, + "bbox": [ + 76, + 482, + 349, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the autoregressive Transformer in our model outputs a probability distribution of the next token conditioned on previous tokens, we can easily sample from this distribution to generate new visual tokens that complete a visual sentence. To use the model for downstream tasks, one can construct a partial visual sentence that defines a task at test time, and apply the model to generate the output. This is similar to in-context learning in language models [10] or visual prompting in computer vision [8, 39].", + "bbox": [ + 75, + 505, + 472, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results and Analysis", + "text_level": 1, + "bbox": [ + 76, + 647, + 395, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the scaling abilities of our trained model, as well as its ability to understand and answer a range of diverse prompted tasks.", + "bbox": [ + 75, + 672, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Scalability", + "text_level": 1, + "bbox": [ + 76, + 726, + 194, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We investigate the scaling behavior of our model in terms of the training loss and downstream task performance as we increase the model size as well as the number of tokens seen during training.", + "bbox": [ + 75, + 750, + 468, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training loss. We first inspect the training loss of LVM with different parameter sizes, which we present in Figure 3. Since all our models are trained for only one epoch on the dataset, the model sees a given data sample just once, and therefore the training loss at any point during training is very similar to the validation loss. One can observe that as", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37117ed5e9cc178a1818577203dd9a4135c7b467e2fa5215636e7d1a628717a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 89, + 694, + 214 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f27582b2f581d86014004d20a0c3abb8f799a3556435c44f9f1389173e14116.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 861, + 214 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/848fc8927e203ecfcaf895d0584e5b9cdc766e23693371a6a9a0d859b686b350.jpg", + "image_caption": [ + "Figure 4. Larger LVMs perform better on downstream tasks. We evaluate LVMs of varying sizes on 4 different downstream tasks, following the 5 shot setting on the ImageNet validation set and report the perplexity. We find that perplexity decreases with larger models across all tasks, indicating strong scalability." + ], + "image_footnote": [], + "bbox": [ + 529, + 215, + 691, + 340 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/696a76a887e9610f75ac82333f4c1f04e0aefc80e4f2cbb8c426f4765b1c6a18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 215, + 859, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training progresses: 1) the training loss (perplexity) of the models, regardless of their size, continues to decrease; 2) as we increase the size of the model (parameter count), the loss decreases faster. These observations indicate that LVM shows strong scalability behavior with both larger models and more data.", + "bbox": [ + 496, + 426, + 893, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scalability on downstream benchmarks. While the LVM overall loss scales well during training, there is no guarantee that the better overall model would also perform better on a given specific downstream task. Therefore, we evaluate different sizes of models on 4 downstream tasks: semantic segmentation, depth estimation, surface normal estimation, and edge detection. We evaluate these tasks on the ImageNet validation set and generate all the annotations using the corresponding method described in Sec. 3. For each task, we give 5 pairs consisting of the inputs and corresponding ground-truth annotations as well as the query image as input prompt and evaluate the perplexity of the ground-truth annotation under our model's prediction of the next 256 tokens (one image). We report the results in Figure 4. We see that larger models indeed attain lower perplexity across all tasks, showcasing that our scalable overall performance does transfer to a range of downstream tasks.", + "bbox": [ + 496, + 520, + 895, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset ablation. While LVM attains better performance with larger models and more data, it is natural to ask whether each data component we collect in UVDv1 helps. To answer this question, we conduct an ablation study on our dataset by training several 3B models on subsets of our dataset, and compare their performances on downstream tasks. We use the same 4 downstream tasks and settings as before and present the results in Figure 5. We observe that each", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22865", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e20f3c458684b4b7616c4b088c3c841732a416a0d8b0950ed539514001778f5c.jpg", + "image_caption": [ + "Figure 6. Frame predictions. LVM predicts the next frames (marked in red) given previous video frames as prompt. The results reveal that the LVM can predict the video frames while considering dynamic objects and camera motion." + ], + "image_footnote": [], + "bbox": [ + 83, + 0, + 888, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data component contributes positively to the downstream tasks. LVM not only benefits from larger data, but also improves with more diversity in the dataset, which includes both annotated and unsupervised image and video data.", + "bbox": [ + 75, + 450, + 470, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Sequential Prompting", + "text_level": 1, + "bbox": [ + 76, + 513, + 282, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin with the most intuitive and straightforward approach to visually prompt the LVM: sequential reasoning. Here the prompt construction is very simple: we present the model with a sequence of 7 images and ask it to predict the next image (256 tokens).", + "bbox": [ + 75, + 537, + 470, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video frame prediction. The most direct task for sequential prompting is video prediction. Figure 6 presents several next frame prediction examples, prompted by sequences from the Kinetics-700 validation set. At the top, 7 frame prompts (blue border) are followed by the predicted frame (red border). We observe a certain degree of inferential ability regarding spatial positioning, viewpoint and object understanding. Perplexity of prediction on Kinetics val set is 49.8. The last 4 rows show predictions with longer context (15 frames) and a longer prediction (4 frames). See Figures 16 to 21 in the Appendix for many more examples.", + "bbox": [ + 75, + 613, + 470, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rotation and Category prediction. The same type of simple sequential prompting can be used in other ways as well. For example, Figure 15 shows how prompting the model with a sequence of 3D rotations of a synthetic object around an arbitrary axis allows it to predict further rotation. Or we can think of a list of items of a given category as a sequence and predict other ideas in that same category, as shown in Figure 14. Note that, while the system was trained", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on groups of images from the same ImageNet category, here the prompt consists of sketches, which have not been seen in any annotated data.", + "bbox": [ + 496, + 450, + 892, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Context length analysis. Next we ask how much temporal context is required to accurately predict the subsequent frame? We assessed the model's frame generation perplexity when prompted with a context of varying lengths (1 to 15 frames). On the Kinetics-700 val set, we see a clear improvement in perplexity from 1 to 11 frames after which it stabilizes (from $62.1 \\rightarrow 48.4$ ).", + "bbox": [ + 496, + 496, + 893, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Analogy Prompting", + "text_level": 1, + "bbox": [ + 500, + 608, + 687, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our study progresses by evaluating a more complex prompting structure, which we call 'Analogy Prompting'. This method challenges the model to comprehend analogies of arbitrary length and complexity, thereby testing its advanced interpretative abilities.", + "bbox": [ + 496, + 626, + 893, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 7 shows a sampling of qualitative results with analogy prompting on a number of tasks. The prompts consist of a sequence of 14 images giving examples of various tasks, followed by a 15th query image. Given each prompt, the next image predicted is the result. The top part of the figure shows several example prompts defining tasks that were part of the training set (but these actual images were never seen at training). The bottom part of the figure demonstrates generalization to tasks never shown at training. See Appendix for many more qualitative examples.", + "bbox": [ + 496, + 702, + 893, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unseen Tasks and Dataset. We present the results for keypoint detection on Pascal 3D+ [90], evaluated using the standard Percentage of Correct Keypoints (PCK) metric with", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22866", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8503a81df206b6a830757b15e66da6a2a3d8c51e15a6d1f1be3ce1e025f49cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 89, + 816, + 505 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afb4ba85aaea0b31ac16e2067dfdcc7c253e5658ae115e65c070e04c4410b103.jpg", + "image_caption": [ + "Figure 7. In and out of distribution prompting examples. Every row is a prompt that contains a sequence of images interleaved with annotations, followed by a query. The last image is predicted by the model (marked in red). The last 5 rows show examples where the query image is out of distribution (painting, sketch, etc) for the task it was trained for.", + "Figure 8. Task Compositionality. Examples of prompts that combine two different tasks - object rotation and keypoint tracking." + ], + "image_footnote": [], + "bbox": [ + 142, + 551, + 823, + 728 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/681745521dcc7a5e2b191e8414c8b15988e7ea771f11923cfa673f6225ed8c49.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelForeground SegmentationSingle Object DetectionColorization ↓
Split 0Split 1Split 2Split 3Split 1Split 2Split 3Split 4MSELPIPS
MAE (IN-1k)1.926.763.854.571.371.981.621.621.130.87
MAE-VQGAN (IN-1k)2.227.075.486.283.343.212.802.803.310.75
MAE (CVF)17.4225.7018.6416.535.494.985.245.840.430.55
MAE-VQGAN (CVF)27.8330.4426.1524.2524.1925.2025.3625.230.670.40
Ours48.9451.2947.6650.8248.2549.6050.0848.920.510.46
", + "bbox": [ + 80, + 750, + 467, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison with Visual Prompting [8]. For Foreground Segmentation and Single Object Detection, we report the mIOU score. For Colorization, we report the MSE and LPIPS.", + "bbox": [ + 76, + 810, + 468, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a of threshold 0.1. Remarkably, LVM achieves a PCK of 81.2 without training on this dataset, demonstrating impressive generalization capabilities. In comparison, we show", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "some existing task-specific model: StackedHourglass [60] scores 68.0 PCK, MSS-Net [42] achieves 68.9 PCK, and StarMap [101] registers 78.6 PCK.", + "bbox": [ + 498, + 750, + 890, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Visual Prompting. The closest approach to ours that also allows for defining arbitrary tasks is Visual Prompting [8]. In Table 1, we compare various visual prompting models on few-shot segmentation, object detection, and colorization tasks. Note that our sequential LVM beats previous approaches on almost all tasks.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22867", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c41df0ad173827242d02488f71f3dee33fc9650e2e50abd2e6625b5ddea3cda4.jpg", + "image_caption": [ + "Figure 9. Miscellaneous Prompting. A variety of simple vision tasks, such as object replication (top), relighting (middle), and zooming in (bottom), can be simply specified via a suitably chosen visual sentence prompt that expresses the task to the LVM." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 426, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3df5b978d7fc8aacad29b6f95fe2b47ab8a388d3e13a843d7a2a4685ede0a650.jpg", + "image_caption": [ + "Figure 10. What comes next? Tasks that are not always easily describable in language." + ], + "image_footnote": [], + "bbox": [ + 135, + 316, + 403, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e10c1d17c730dcf91646a885a62667de00a2c63b864c26f9e3da650ed17da03f.jpg", + "image_caption": [ + "Figure 11. \"Sparks of AGI?\" We prompt LVM with a masked reasoning visual sentence to infer the solution for non-verbal reasoning questions which are prevalent in IQ tests (masked image, second from the right). Left are successful cases and right is failure. We find that the model often infers and applies the abstract visual pattern correctly. So, we graciously hand over to you, our gentle reader, the task of pondering whether our modest LVM also exhibits the much-vaunted 'Sparks of AGI'." + ], + "image_footnote": [], + "bbox": [ + 81, + 458, + 467, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task Compositing. Figure 8 demonstrates compositing several tasks together within a single prompt. Here, we demonstrate the rotation task together with the novel keypoint correspondence task and ask the model to continue the pattern. The model can successfully combine these two at test time, demonstrating some degree of compositionality.", + "bbox": [ + 75, + 630, + 468, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Miscellaneous Prompts", + "text_level": 1, + "bbox": [ + 76, + 728, + 290, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we try to see how far we can push our model by offering it various prompts it has not seen before. Figure 9 shows a few such prompts that happened to work reasonably well. Figure 10 shows some prompts which are not easily describable by words – these are the type of tasks where LVMs might eventually outshine LLMs.", + "bbox": [ + 75, + 750, + 468, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 11, we show initial qualitative results on a typical visual reasoning question as found on non-verbal human IQ tests (Raven's Progressive Matrices [67]). With considerable squinting, one could imagine the LVM having a", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cee850566bfc5566a34f6a35f432eeda073e4bc33191194e46dd3300caa2f80.jpg", + "image_caption": [ + "Figure 12. Failure Cases. This figure illustrates seven examples of failure cases: (1) Task confusion: the model interprets the counting task as style transfer, resulting in a pear-shaped apple. (2) Task entanglement: the model perceives the prompt as high frequencies from the entire image, rather than just from some parts of image (objects). (3) Wrong instance: the model correctly identifies rotation but mistakenly generates a brush instead of a manta. (4) Outlier detection versus generation: rather than detecting an outlier, the model directly generates it. (5) Digits: the model fails to represent digit sequences. (6) Tokenizer: poor performance with some out-of-distribution synthetic data generation (wrong background). (7) Sequence degeneration: sometimes frame predictions go astray." + ], + "image_footnote": [], + "bbox": [ + 504, + 92, + 890, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "latent ability for grasping abstract visual patterns and applying them to extrapolate the shown visual sequence. We also perform perplexity analysis on 10 classic Raven 5-way multiple-choice Matrices in Table 2, choosing the answer with lowest perplexity under our model. It shows that our results are better than chance. To further understand this process, we generate 900 new Raven-like synthetic tests to study individual factors: color, shape, and size following [8].", + "bbox": [ + 498, + 404, + 893, + 525 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/07a0c93eea144c0693050fad70246b463787510032ad19197027fdc5b12283b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Raven IQ matricesshapecolorsize
chance20%33%33%50%
Ours30%45%42%94%
", + "bbox": [ + 544, + 527, + 846, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Performance on 10 Raven's Progressive Matrices and 900 synthetic reasoning tasks by lowest perplexity under our model.", + "bbox": [ + 498, + 566, + 890, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Failure Cases", + "text_level": 1, + "bbox": [ + 500, + 599, + 638, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 12 shows some typical failure cases of the current model. One common element, the use of visual prompt to define a task is often under-constrained (more so than in language, since images are very high-dimensional), or the requested task might be beyond the capabilities of the current system. Other, more mundane failures involve issues with the tokenizer and lack of high-quality video training data.", + "bbox": [ + 498, + 613, + 890, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: We are grateful to many friends and colleagues for discussions and comments on this work, including Yossi Gandelsman, Aleksander Holynski, Angjoo Kanazawa, Qianqian Wang, Sophia Koepke, Quoc Le, Chen Liang, Zichao Li, Ekin D. Cubuk, Assaf Shocher, Amir Zamir, Carl Vondrick, Ludwig Schmidt, Aviral Kumar, Xiaolong Wang and Miki Rubinstein. We thank the support from Google during Yutong's part-time internship. The work has been supported, in part, by ONR MURI N0001 4-22-1-2773, N00014-21-1-2801, and N00014-21-1-2812, ERC HOLI, Apple Graduate Fellowship to Yutong Bai, and compute donation via the Google TPU Research Cloud. Darrell's group is supported in part by the US DoD, BAIR&BDD industrial consortia.", + "bbox": [ + 496, + 714, + 893, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22868", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Common crawl repository. https://commoncrawl.org/.", + "[2] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1692-1700, 2018.", + "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014.", + "[4] Okan Arikan and D. A. Forsyth. Interactive motion generation from examples. ACM Trans. Graph., 21(3):483-490, 2002.", + "[5] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954.", + "[6] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13137-13146, 2021.", + "[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021.", + "[8] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022.", + "[9] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023.", + "[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020.", + "[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015.", + "[12] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019.", + "[13] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022.", + "[14] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019." + ], + "bbox": [ + 86, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020.", + "[16] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021.", + "[17] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022.", + "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023.", + "[19] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.", + "[20] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.", + "[21] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2634-2641, 2013.", + "[22] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023.", + "[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[24] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE international conference on computer vision, pages 2051-2060, 2017.", + "[25] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015.", + "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020.", + "[27] Alexei A. Efros and William T. Freeman. Image quilting for" + ], + "bbox": [ + 509, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22869", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "texture synthesis and transfer. In Proceedings of SIGGRAPH 2001, 2001.", + "[28] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In IEEE International Conference on Computer Vision, pages 1033-1038, Corfu, Greece, 1999.", + "[29] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021.", + "[30] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In CVPR 2011, pages 3281-3288. IEEE, 2011.", + "[31] David Donovan Garber. Computational Models for Texture Analysis and Texture Synthesis. PhD thesis, University of Southern California, Image Processing Institute, 1981.", + "[32] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012.", + "[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012.", + "[34] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence, 38(1):142-158, 2015.", + "[35] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017.", + "[36] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022.", + "[37] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020.", + "[38] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022.", + "[39] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001." + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In International Conference on Learning Representations, 2021.", + "[41] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In International Conf. on Computer Vision (ICCV), pages 3192-3199, 2013.", + "[42] Lipeng Ke, Ming-Ching Chang, Honggang Qi, and Siwei Lyu. Multi-scale structure-aware network for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 713-728, 2018.", + "[43] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6129-6138, 2017.", + "[44] Lucas Kovar, Michael Gleicher, and Frederic Pighin. Motion Graphs. In Proceedings of SIGGRAPH '02, San Antonio, TX, 2002.", + "[45] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012.", + "[46] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011.", + "[47] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017.", + "[48] Jehee Lee, Jinxiang Chai, Paul Reitsma, Jessica K. Hodgins, and Nancy Pollard. Interactive control of avatars animated with human motion data. ACM Transactions on Graphics (TOG), 21(3):491 - 500, 2002.", + "[49] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017.", + "[50] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018.", + "[51] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021.", + "[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22870", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024.", + "[54] Shikun Liu, Linxi Fan, Edward Johns, Zhiding Yu, Chaowei Xiao, and Anima Anandkumar. Prismer: A vision-language model with an ensemble of experts. arXiv preprint arXiv:2303.02506, 2023.", + "[55] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019.", + "[56] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022.", + "[57] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019.", + "[58] Mathew Monfort, Bowen Pan, Kandan Ramakrishnan, Alex Andonian, Barry A McNamara, Alex Lascelles, Quanfu Fan, Dan Gutfreund, Rogério Schmidt Feris, and Aude Oliva. Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9434-9445, 2021.", + "[59] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE conference on computer vision and pattern recognition, pages 2408-2415. IEEE, 2012.", + "[60] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016.", + "[61] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016.", + "[62] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016.", + "[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016.", + "[64] Kris Popat and Rosalind W. Picard. Novel cluster-based probability model for texture synthesis, classification, and compression. In Proc. SPIE Visual Comm. and Image Processing, 1993." + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.", + "[66] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014.", + "[67] John C Raven. Standardization of progressive matrices, 1938. British Journal of Medical Psychology, 1941.", + "[68] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021.", + "[69] Arno Schodl, Richard Szeliski, David H. Salesin, and Irfan Essa. Video textures. In Proceedings of SIGGRAPH '00, pages 489-498, 2000.", + "[70] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.", + "[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022.", + "[72] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022.", + "[73] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018.", + "[74] Claude E. Shannon. A mathematical theory of communication. Bell Sys. Tech. Journal, 27, 1948.", + "[75] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951.", + "[76] Gunnar A Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016.", + "[77] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018.", + "[78] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012." + ], + "bbox": [ + 509, + 92, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22871", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] X. Soria, E. Riba, and A. Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1912-1921, Los Alamitos, CA, USA, 2020. IEEE Computer Society.", + "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[81] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, koray kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016.", + "[82] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, New York, New York, USA, 2016. PMLR.", + "[83] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017.", + "[84] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017.", + "[85] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th International Conference on Machine Learning, page 1096-1103, 2008.", + "[86] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019.", + "[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023.", + "[88] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of laion-2b, 2023.", + "[89] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018.", + "[90] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2014.", + "[91] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022.", + "[92] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language." + ], + "bbox": [ + 86, + 92, + 470, + 901 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016.", + "[93] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018.", + "[94] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In International Conference on Learning Representations, 2021.", + "[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022.", + "[96] Yingchen Yu, Fangneng Zhan, Rongliang Wu, Jianxiong Pan, Kaiwen Cui, Shijian Lu, Feiying Ma, Xuansong Xie, and Chunyan Miao. Diverse image inpainting with bidirectional and autoregressive transformers. In Proceedings of the 29th ACM International Conference on Multimedia, pages 69-78, 2021.", + "[97] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018.", + "[98] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In CVPR, 2018.", + "[99] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016.", + "[100] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019.", + "[101] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018." + ], + "bbox": [ + 501, + 92, + 893, + 671 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22872", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_model.json b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..554c43b15cbba5b13bd557f836368598b8723b39 --- /dev/null +++ b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_model.json @@ -0,0 +1,3326 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.131, + 0.852, + 0.153 + ], + "angle": 0, + "content": "Sequential Modeling Enables Scalable Learning for Large Vision Models" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.18, + 0.294, + 0.199 + ], + "angle": 0, + "content": "Yutong Bai\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.199, + 0.303, + 0.216 + ], + "angle": 0, + "content": "Alan L. Yuille" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.181, + 0.465, + 0.199 + ], + "angle": 0, + "content": "Xinyang Geng\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.199, + 0.461, + 0.216 + ], + "angle": 0, + "content": "Trevor Darrell" + }, + { + "type": "text", + "bbox": [ + 0.494, + 0.181, + 0.674, + 0.199 + ], + "angle": 0, + "content": "Karttikeya Mangalam" + }, + { + "type": "text", + "bbox": [ + 0.494, + 0.199, + 0.621, + 0.216 + ], + "angle": 0, + "content": "Jitendra Malik1" + }, + { + "type": "text", + "bbox": [ + 0.71, + 0.181, + 0.794, + 0.198 + ], + "angle": 0, + "content": "Amir Bar" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.199, + 0.788, + 0.216 + ], + "angle": 0, + "content": "Alexei A. Efros" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.233, + 0.426, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)UC Berkeley (BAIR)" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.233, + 0.719, + 0.253 + ], + "angle": 0, + "content": "2Johns Hopkins University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.319, + 0.474, + 0.531 + ], + "angle": 0, + "content": "We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, \"visual sentences\", in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions without needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion tokens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.21, + 0.577 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.587, + 0.471, + 0.737 + ], + "angle": 0, + "content": "Large language models (LLMs) such as GPT [10] and LLaMA [80] have taken the world by storm. What would it take to build a Large Vision Model (LVM)? From the animal world, we know that visual competences are not dependent on language. In particular, many experiments have shown that the visual world of non-human primates is remarkably similar to that of humans. So while the space of vision-language models such as LLaVA [53] is interesting and worthwhile to pursue, in this paper we seek an answer to a different question – how far can we go from pixels alone?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.815 + ], + "angle": 0, + "content": "The key features of contemporary LLMs that we seek to emulate in LVMs are: 1) scaling in the presence of big data, and 2) flexible specification of tasks through prompting (in-context learning). How do we achieve this? As usual, there are three main components that must be specified:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.815, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Data: We want to exploit all the remarkable diversity in visual data. First of all, just raw unannotated images and videos. Next, we want to exploit the variety of annotated visual data sources that have been produced over the last" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.379 + ], + "angle": 0, + "content": "couple of decades – semantic segmentations, depth reconstructions, keypoints, multiple views of 3D objects, among others. We define a common format, “visual sentences”, in which to represent these different annotations without needing any meta-knowledge beyond the pixels. The total size of our training dataset is 1.64 billion images/frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Architecture: We use a large transformer architecture (3 billion parameters) trained on visual data represented as sequence of tokens, using a learned tokenizer that maps each image to a string of 256 vector-quantized tokens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.895, + 0.532 + ], + "angle": 0, + "content": "Loss function: We draw inspiration from the natural language community, where masked token modeling has given way to sequential autoregressive prediction. Once images/videos/annotated images can all be represented as sequences, we can train the model to minimize the cross-entropy loss for predicting the next token." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.563 + ], + "angle": 0, + "content": "With this extremely simple design, we demonstrate some noteworthy behaviors:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.564, + 0.892, + 0.593 + ], + "angle": 0, + "content": "- Appropriate scaling behavior as one increases model size and data size." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.594, + 0.894, + 0.669 + ], + "angle": 0, + "content": "- Many different vision tasks can now be \"solved\" by designing suitable prompts at test time. While the results don't show as high performance as bespoke, specifically-trained models, the fact that so many tasks are all addressed by a single vision model is quite encouraging." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.669, + 0.892, + 0.699 + ], + "angle": 0, + "content": "- We see a clear benefit of the amount of unsupervised data on the performance on various standard vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.7, + 0.895, + 0.745 + ], + "angle": 0, + "content": "- We see a hint of an ability for general visual reasoning – handling out-of-distribution data, and performing novel tasks. But further investigation is needed." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.564, + 0.895, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.758, + 0.642, + 0.773 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Pretrained Vision Models. The value of using pretrained models (such as ImageNet-pretrained AlexNet [45]) has been demonstrated as far back as 2015 in R-CNN [34], and it has since become standard practice in computer vision. Self-supervised pretraining was proposed as a way to vastly increase the amount of data available for pretraining [16, 25, 37, 61, 62, 99]. Unfortunately, this was not very successful, likely because the CNN-based architectures" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.438, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution. Webpage: https://yutongbai.com/lvm.html." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22861" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "of that time did not have enough capacity to absorb the data. With the introduction of Transformers [84], which have much higher capacity, researchers revisited self-supervised pretraining, and showed that transformer-based masked image reconstruction approaches, such as BEiT [7], MAE [38], SimMIM [91], MaskGIT [13] perform vastly better than their CNN-based counterparts [62]. Yet, despite their recent successes, current pretrained vision-only models have had trouble scaling up to the really large datasets, such as LAION [72]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.46 + ], + "angle": 0, + "content": "Multi-task Learning and In-context Learning. From the classic one-model-per-task setups, computer vision is slowly moving toward a single model performing multiple different tasks. Various multi-task learning approaches [24, 40, 43, 73, 97] exist but they are typically limited to a fixed, predefined number of tasks. More recently, methods inspired by in-context learning in LLMs forgo any notion of tasks and instead let the model infer the task directly from the input prompt. For example, Visual Prompting [8, 87] takes in a task input/output example pair and a query image at test time, concatenates them into a single 2-by-2 image, and uses inpainting to generate the desired output. But, since the inpainting is performed using a variant of MAE [38], the same problems with scaling are inherited by these approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.676 + ], + "angle": 0, + "content": "Auto-regressive Visual Models. The idea of using autoregressive models for synthesizing visual data goes back at least 70 years. Inspired by Shannon's use of \\(N\\)-grams to synthesize language [74, 75], a number of works, starting with Attneave's seminal 1954 paper [5], applied this idea to sequentially synthesizing pixels [28, 31, 39, 64], image patches [27], video frames [69], and motion capture data [4, 44, 48]. As deep models became popular, newer works replaced \\(N\\)-grams with RNNs or CNNs for pixel [81, 82] and video [66] synthesis. Most recently, transformer-based autoregressive visual generation methods have been proposed [15, 29, 94, 96], and, combined with language, have demonstrated impressive image synthesis results, e.g. Parti [95]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.686, + 0.143, + 0.701 + ], + "angle": 0, + "content": "3. Data" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.714, + 0.471, + 0.728 + ], + "angle": 0, + "content": "\"Data! Data! Data! I can't make bricks without clay!\"" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.738, + 0.468, + 0.75 + ], + "angle": 0, + "content": "SHERLOCK HOLMES" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.901 + ], + "angle": 0, + "content": "The key requirement of any Large Pre-trained Model is that it must be trained on vast amounts of data. For language models, very large and very diverse datasets are fairly easy to obtain. For instance, the popular Common Crawl repository [1] contains 250 billion web pages spanning the entire Web, is extremely diverse, and includes \"natural demonstrations\" like language translations, question answering, etc. In computer vision, we are still very far from having a data source of comparable size and diversity. One of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.363 + ], + "angle": 0, + "content": "central contributions of our work is the first step toward curating such a dataset that we call Unified Vision Dataset v1 (UVDv1). To assemble it, we leverage many different sources of visual data: (1) unlabelled images, (2) images with visual annotations, (3) unlabelled videos, (4) videos with visual annotations, and (5) 3D synthetic objects. The unlabeled images, which represent over \\(80\\%\\) of our data, capture a huge cross-section of our visual world, and provide the required diversity, at the cost of lower quality. Images with annotations have a much more constrained distribution, but are usually of higher quality. Video data is even more constrained (typically, to human-centric activities), but is an invaluable source of temporal data. Renderings of 3D synthetic objects are the lowest in diversity but can provide valuable hints about the behavior of 3D structures. Importantly, UVDv1 is a purely visual dataset, with no non-visual meta-data (e.g. text) included. All together, UVDv1 contains 1.64 billion images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.897, + 0.53 + ], + "angle": 0, + "content": "Another important difference from Large Language Models is that language data has a natural, unified one-dimensional structure for all the data - a stream of text. Unfortunately, this is not the case for visual data, with different sources all having different structures. In this work we propose visual sentence as a unified unit of visual data, which enables us to train scalable models from a diverse set of sources. A visual sentence is simply a sequence containing one or more images followed by an end-of-sentence (EOS) token. Figure 1 shows how the various data sources are partitioned into visual sentences. In particular:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Single images. A single image itself represents the simplest form of a visual sentence - \\(\\{\\) image, EOS\\(\\}\\). We use the filtered subset of 1.49 billion images [88] from the LAION 5B [71] dataset. This is by far the largest part of our data, comprising \\(88.5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Image sequences. A sequence of images is a natural form of visual sentence. We create such sequences by sourcing video data from a wide range of existing datasets [11, 12, 21, 35, 36, 46, 50, 51, 55, 57-59, 63, 68, 76-78, 92, 93]. Visual sentences of 16 frames are formed by randomly sampling the videos at three different strides (10, 20, and 30). In addition, we utilize synthetic 3D objects from the Objaverse Dataset [22] to generate object-centric multiview sequences for a variety of objects. For each object, we sample one radius length between the object center and the camera from 1.5 to 2.2, and sample one constant elevation from -45 degrees to 45 degrees, then traverse different views of the object by changing the azimuth with a step length of 15 degrees and render 24 views. We rendered 42000 such sequences in total for training and 8000 for testing. Finally, we can also represent images belonging to the same semantic category as being (part of) a sequence. We use categories from ImageNet, concatenating together groups of images (2,4,8, or 16) from the same category into a 16-image long" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22862" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.141, + 0.103, + 0.326, + 0.121 + ], + "angle": 0, + "content": "Visual Sentences" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.133, + 0.123, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.154, + 0.152, + 0.162 + ], + "angle": 0, + "content": "#" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.144, + 0.345, + 0.157 + ], + "angle": 0, + "content": "Single images, e.g. LAION" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.169, + 0.389, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.19, + 0.414, + 0.2 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.205, + 0.388, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.204, + 0.415, + 0.216 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.242, + 0.388, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.242, + 0.415, + 0.253 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.263, + 0.414, + 0.27 + ], + "angle": 0, + "content": "#" + }, + { + "type": "text", + "bbox": [ + 0.456, + 0.19, + 0.582, + 0.246 + ], + "angle": 0, + "content": "Image sequences, e.g. videos, 3D rotations, synthetic viewpoints" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.276, + 0.159, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.17, + 0.276, + 0.249, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.276, + 0.336, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.276, + 0.425, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.295, + 0.452, + 0.307 + ], + "angle": 0, + "content": "··" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.313, + 0.16, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.313, + 0.249, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.313, + 0.337, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.313, + 0.425, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.313, + 0.452, + 0.323 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.35, + 0.16, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.35, + 0.249, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.35, + 0.338, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.35, + 0.425, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.373, + 0.452, + 0.381 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.387, + 0.16, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.387, + 0.249, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.387, + 0.338, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.387, + 0.425, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.387, + 0.452, + 0.408 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.423, + 0.199, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.425, + 0.326, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.444, + 0.353, + 0.452 + ], + "angle": 0, + "content": "··" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.46, + 0.159, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.46, + 0.249, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.461, + 0.338, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.461, + 0.425, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.461, + 0.452, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.461, + 0.582, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.584, + 0.425, + 0.703, + 0.482 + ], + "angle": 0, + "content": "Edges with free form notation, 1. object detection + frame segmentation" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.497, + 0.199, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.497, + 0.316, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.497, + 0.425, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.497, + 0.554, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.534, + 0.705, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.494, + 0.734, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.295, + 0.89, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.534, + 0.236, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.535, + 0.386, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.534, + 0.554, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.534, + 0.705, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.706, + 0.564, + 0.73, + 0.569 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.731, + 0.498, + 0.848, + 0.553 + ], + "angle": 0, + "content": "Videos with annotation, e.g. video segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.569, + 0.812, + 0.583 + ], + "angle": 0, + "content": "Figure 1. Visual sentences allow us to format diverse vision data into the unified structure of image sequences." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.589, + 0.191, + 0.602 + ], + "angle": 0, + "content": "visual sentences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Images with annotations. To handle different types of image annotations in a uniform way, we choose to represent all annotations as images. Some data types, e.g. semantic segmentation maps [100], edge maps [79], depth [65] and normal images [6], are already represented this way. For others we apply tailored methods for each specific annotation type: 1) Object Detection: We create annotations by overlaying a color-coded bounding box around each object, following the methodology in [14]; 2) Human Pose: Human skeletons are rendered in pixel space, adhering to the OpenPose format, utilizing MMPose [19]; 3) Depth Estimation, Surface Normal, and Edge Detection: given ImageNet and COCO images, we generate annotations in line with the protocols from [54]. 3) Style Transfer [9], De-rain [98], De-noise [85], Low Light Enhancement [89], and Stereo Datasets [33]: These are all represented as image pairs (e.g. input/output). 4) Colorization: We convert ImageNet images to greyscale, producing image pairs. 5) Inpainting: The process involves randomly adding black-colored boxes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.895, + 0.724 + ], + "angle": 0, + "content": "in images to simulate corruption, resulting in image pairs. For all the above annotation types, we can create visual sentences by concatenating 8 image pairs of the same annotation type into a 16-image visual sentence. For datasets containing \\( k \\) different annotations for the same image we use a different approach: for each set of \\( 1 + k \\) images (input plus \\( k \\) annotations), we randomly select \\( m \\) elements, where \\( m \\leq k + 1 \\leq 16 \\). These m-tuples are then concatenated to form visual sequences." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.732, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Image sequences with annotations. When converting annotated video data (VIPSeg [56], Hand14K [30], AVA [59], JHMDB [41]) to visual sentences, we apply two complementary strategies. The first is similar to how we treat image data with paired annotations: each visual sentence is constructed by concatenating frames with their annotations - {frame1,annot1,frame2,annot2,...}. The second method involves grouping multiple frames followed by their corresponding annotations - {frame1,frame2,annot1,annot2,...}." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We present a detailed summary of all the data sources, annotation type and data statistics of UVDv1 in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22863" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.188, + 0.108 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.471, + 0.252 + ], + "angle": 0, + "content": "In this section, we describe the design of our autoregressive Large Vision Model. Unlike text data, which naturally exhibits discrete sequential structure, it is not straightforward to model image pixels in visual sentences. In this work, we take a two-stage approach: 1) train a large visual tokenizer (which operates on individual images) to convert each image into a sequence of visual tokens; 2) train an autoregressive transformer model on visual sentences, each represented as a sequence of tokens. We summarize our approach in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.259, + 0.266, + 0.275 + ], + "angle": 0, + "content": "4.1. Image Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.471, + 0.492 + ], + "angle": 0, + "content": "While the visual sentences exhibit a sequence structure between consecutive images, we don't have such natural sequence structure within an image. Therefore, in order to apply a transformer model to images, prior works typically do one of the following: either divide the image into patches in scan-line order, and treat that as a sequence [26], or use a pre-trained image tokenizer, such as VQVAE [83] or VQGAN [29], to cluster image features into a grid of discrete tokens, which, again, are turned into a sequence in scan-line order. We adopt the latter approach since the discrete categorical output from a model naturally forms a probabilistic distribution that one can easily sample from, enabling flexible conditional generation of new images within a visual sentence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.674 + ], + "angle": 0, + "content": "Specifically, we employ semantic tokens generated by a VQGAN model, a concept introduced by Esser et al [29]. This framework consists of an encoding and a decoding mechanism, featuring a quantization layer that assigns input images to a sequence of discrete tokens from an established codebook. Our encoders and decoders are constructed purely with convolutional layers. The encoder is equipped with several downsampling modules to contract the spatial dimension of the input, whereas the decoder is fitted with an equivalent series of upsampling modules to restore the image to its initial size. For a given image, our VQGAN tokenizer produces 256 discrete tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.78 + ], + "angle": 0, + "content": "It is important to note that our tokenizer operates on individual images independently, rather than on the entire visual sentence at once. This independence allows us to decouple the tokenizer training from the downstream Transformer model so that the tokenizer can be trained on a dataset of single images without having to consider the distribution of visual sentences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implementation Details: We adopt an off-the-shelf VQGAN architecture from Chang et al. [13]. We follow the exact configuration in Chang et al. [13], which uses a down-sampling factor of \\( f = 16 \\) and codebook size 8192. This means that for an image of size \\( 256 \\times 256 \\), our VQGAN tokenizer produces \\( 16 \\times 16 = 256 \\) tokens where each can take 8192 different values. We found that using the results of an ImageNet pre-trained tokenizer did not generalize well" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.093, + 0.852, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.326, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Figure 2. Architecture of LVM. We first convert individual images from a visual sentence into discrete tokens using a VQGAN encoder. The resulting tokens from all images are then concatenated into a 1D sequence, and fed into an autoregressive Transformer model to predict the next token in the sequence. The predicted visual tokens are decoded into images using the VQGAN decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.446 + ], + "angle": 0, + "content": "beyond ImageNet images. Therefore, we train our own tokenizer on a 1.5B subset of the LAION 5B dataset [71]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.455, + 0.838, + 0.471 + ], + "angle": 0, + "content": "4.2. Sequence Modeling of Visual Sentences" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.674 + ], + "angle": 0, + "content": "After converting images into discrete tokens with VQGAN, we treat our visual sentence as a unified sequence by concatenating the discrete tokens from multiple images into a 1D sequence. Importantly, all visual sentences are treated equally – we do not make use of any special tokens to indicate particular tasks or formats. We train a causal Transformer model with the next token prediction objective using a cross-entropy loss, similar to the standard approach for language models [10]. Training the model the same way on all visual sentences enables the model to infer the relation between images from context instead of from task- or format-specific tokens. This gives the model an opportunity to generalize to other, unseen visual sentence structures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Implementation Details: After tokenizing each image in a visual sentence into 256 tokens, we concatenate them to form a 1D sequence of tokens. On top of the sequences of visual tokens, our Transformer model is virtually the same as an autoregressive language model, so we adopt the Transformer architecture of LLaMA [80], a popular open-source language model with widely available implementations. We use a context length of 4096 tokens, which can fit 16 images under our VQGAN tokenizer. Similar to language models, we add a [BOS] (begin of sentence) token to the beginning of each visual sentence and an [EOS] (end of sentence) token to the end, and use sequence concatenation [18] during training time to improve efficiency. We train our model on our entire UVDv1 dataset (420 billion tokens) using one epoch (simple epoch training is standard in language models to avoid poten" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22864" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.468, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.472, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Training loss for the 300M, 600M, 1B, and 3B models. All models are trained on 420B tokens (1.64B images). The training scales well with model sizes. Zoom in for more details: notice how the shape of the 4 curves is very similar. This is the consequence of the deterministic training pipeline, so the loss fluctuations are determined by the intrinsic difficulty of the data stream." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.473, + 0.477 + ], + "angle": 0, + "content": "tial overfitting). We train 4 models with different numbers of parameters: 300 million, 600 million, 1 billion and 3 billion, following the same training configurations. We provide the detailed training hyperparameters in Appendix 5.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.351, + 0.5 + ], + "angle": 0, + "content": "4.3. Inference by Visual Prompting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.643 + ], + "angle": 0, + "content": "Since the autoregressive Transformer in our model outputs a probability distribution of the next token conditioned on previous tokens, we can easily sample from this distribution to generate new visual tokens that complete a visual sentence. To use the model for downstream tasks, one can construct a partial visual sentence that defines a task at test time, and apply the model to generate the output. This is similar to in-context learning in language models [10] or visual prompting in computer vision [8, 39]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.396, + 0.667 + ], + "angle": 0, + "content": "5. Experimental Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.719 + ], + "angle": 0, + "content": "In this section, we evaluate the scaling abilities of our trained model, as well as its ability to understand and answer a range of diverse prompted tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.196, + 0.743 + ], + "angle": 0, + "content": "5.1. Scalability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.811 + ], + "angle": 0, + "content": "We investigate the scaling behavior of our model in terms of the training loss and downstream task performance as we increase the model size as well as the number of tokens seen during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training loss. We first inspect the training loss of LVM with different parameter sizes, which we present in Figure 3. Since all our models are trained for only one epoch on the dataset, the model sees a given data sample just once, and therefore the training loss at any point during training is very similar to the validation loss. One can observe that as" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.09, + 0.695, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.862, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.217, + 0.692, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.217, + 0.86, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.343, + 0.895, + 0.414 + ], + "angle": 0, + "content": "Figure 4. Larger LVMs perform better on downstream tasks. We evaluate LVMs of varying sizes on 4 different downstream tasks, following the 5 shot setting on the ImageNet validation set and report the perplexity. We find that perplexity decreases with larger models across all tasks, indicating strong scalability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.895, + 0.517 + ], + "angle": 0, + "content": "training progresses: 1) the training loss (perplexity) of the models, regardless of their size, continues to decrease; 2) as we increase the size of the model (parameter count), the loss decreases faster. These observations indicate that LVM shows strong scalability behavior with both larger models and more data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.896, + 0.777 + ], + "angle": 0, + "content": "Scalability on downstream benchmarks. While the LVM overall loss scales well during training, there is no guarantee that the better overall model would also perform better on a given specific downstream task. Therefore, we evaluate different sizes of models on 4 downstream tasks: semantic segmentation, depth estimation, surface normal estimation, and edge detection. We evaluate these tasks on the ImageNet validation set and generate all the annotations using the corresponding method described in Sec. 3. For each task, we give 5 pairs consisting of the inputs and corresponding ground-truth annotations as well as the query image as input prompt and evaluate the perplexity of the ground-truth annotation under our model's prediction of the next 256 tokens (one image). We report the results in Figure 4. We see that larger models indeed attain lower perplexity across all tasks, showcasing that our scalable overall performance does transfer to a range of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Dataset ablation. While LVM attains better performance with larger models and more data, it is natural to ask whether each data component we collect in UVDv1 helps. To answer this question, we conduct an ablation study on our dataset by training several 3B models on subsets of our dataset, and compare their performances on downstream tasks. We use the same 4 downstream tasks and settings as before and present the results in Figure 5. We observe that each" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.0, + 0.89, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.417, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Figure 6. Frame predictions. LVM predicts the next frames (marked in red) given previous video frames as prompt. The results reveal that the LVM can predict the video frames while considering dynamic objects and camera motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.512 + ], + "angle": 0, + "content": "data component contributes positively to the downstream tasks. LVM not only benefits from larger data, but also improves with more diversity in the dataset, which includes both annotated and unsupervised image and video data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.283, + 0.531 + ], + "angle": 0, + "content": "5.2. Sequential Prompting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.613 + ], + "angle": 0, + "content": "We begin with the most intuitive and straightforward approach to visually prompt the LVM: sequential reasoning. Here the prompt construction is very simple: we present the model with a sequence of 7 images and ask it to predict the next image (256 tokens)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.472, + 0.78 + ], + "angle": 0, + "content": "Video frame prediction. The most direct task for sequential prompting is video prediction. Figure 6 presents several next frame prediction examples, prompted by sequences from the Kinetics-700 validation set. At the top, 7 frame prompts (blue border) are followed by the predicted frame (red border). We observe a certain degree of inferential ability regarding spatial positioning, viewpoint and object understanding. Perplexity of prediction on Kinetics val set is 49.8. The last 4 rows show predictions with longer context (15 frames) and a longer prediction (4 frames). See Figures 16 to 21 in the Appendix for many more examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Rotation and Category prediction. The same type of simple sequential prompting can be used in other ways as well. For example, Figure 15 shows how prompting the model with a sequence of 3D rotations of a synthetic object around an arbitrary axis allows it to predict further rotation. Or we can think of a list of items of a given category as a sequence and predict other ideas in that same category, as shown in Figure 14. Note that, while the system was trained" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.893, + 0.495 + ], + "angle": 0, + "content": "on groups of images from the same ImageNet category, here the prompt consists of sketches, which have not been seen in any annotated data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.895, + 0.603 + ], + "angle": 0, + "content": "Context length analysis. Next we ask how much temporal context is required to accurately predict the subsequent frame? We assessed the model's frame generation perplexity when prompted with a context of varying lengths (1 to 15 frames). On the Kinetics-700 val set, we see a clear improvement in perplexity from 1 to 11 frames after which it stabilizes (from \\(62.1 \\rightarrow 48.4\\))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.609, + 0.688, + 0.626 + ], + "angle": 0, + "content": "5.3. Analogy Prompting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.701 + ], + "angle": 0, + "content": "Our study progresses by evaluating a more complex prompting structure, which we call 'Analogy Prompting'. This method challenges the model to comprehend analogies of arbitrary length and complexity, thereby testing its advanced interpretative abilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Qualitative Results. Figure 7 shows a sampling of qualitative results with analogy prompting on a number of tasks. The prompts consist of a sequence of 14 images giving examples of various tasks, followed by a 15th query image. Given each prompt, the next image predicted is the result. The top part of the figure shows several example prompts defining tasks that were part of the training set (but these actual images were never seen at training). The bottom part of the figure demonstrates generalization to tasks never shown at training. See Appendix for many more qualitative examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Unseen Tasks and Dataset. We present the results for keypoint detection on Pascal 3D+ [90], evaluated using the standard Percentage of Correct Keypoints (PCK) metric with" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22866" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.09, + 0.818, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.509, + 0.895, + 0.551 + ], + "angle": 0, + "content": "Figure 7. In and out of distribution prompting examples. Every row is a prompt that contains a sequence of images interleaved with annotations, followed by a query. The last image is predicted by the model (marked in red). The last 5 rows show examples where the query image is out of distribution (painting, sketch, etc) for the task it was trained for." + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.553, + 0.825, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.732, + 0.867, + 0.746 + ], + "angle": 0, + "content": "Figure 8. Task Compositionality. Examples of prompts that combine two different tasks - object rotation and keypoint tracking." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.75, + 0.468, + 0.81 + ], + "angle": 0, + "content": "
ModelForeground SegmentationSingle Object DetectionColorization ↓
Split 0Split 1Split 2Split 3Split 1Split 2Split 3Split 4MSELPIPS
MAE (IN-1k)1.926.763.854.571.371.981.621.621.130.87
MAE-VQGAN (IN-1k)2.227.075.486.283.343.212.802.803.310.75
MAE (CVF)17.4225.7018.6416.535.494.985.245.840.430.55
MAE-VQGAN (CVF)27.8330.4426.1524.2524.1925.2025.3625.230.670.40
Ours48.9451.2947.6650.8248.2549.6050.0848.920.510.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.811, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Table 1. Comparison with Visual Prompting [8]. For Foreground Segmentation and Single Object Detection, we report the mIOU score. For Colorization, we report the MSE and LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "a of threshold 0.1. Remarkably, LVM achieves a PCK of 81.2 without training on this dataset, demonstrating impressive generalization capabilities. In comparison, we show" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.751, + 0.892, + 0.798 + ], + "angle": 0, + "content": "some existing task-specific model: StackedHourglass [60] scores 68.0 PCK, MSS-Net [42] achieves 68.9 PCK, and StarMap [101] registers 78.6 PCK." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Comparison with Visual Prompting. The closest approach to ours that also allows for defining arbitrary tasks is Visual Prompting [8]. In Table 1, we compare various visual prompting models on few-shot segmentation, object detection, and colorization tasks. Note that our sequential LVM beats previous approaches on almost all tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22867" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.088, + 0.427, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.312 + ], + "angle": 0, + "content": "Figure 9. Miscellaneous Prompting. A variety of simple vision tasks, such as object replication (top), relighting (middle), and zooming in (bottom), can be simply specified via a suitably chosen visual sentence prompt that expresses the task to the LVM." + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.318, + 0.404, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.452 + ], + "angle": 0, + "content": "Figure 10. What comes next? Tasks that are not always easily describable in language." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.459, + 0.468, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.624 + ], + "angle": 0, + "content": "Figure 11. \"Sparks of AGI?\" We prompt LVM with a masked reasoning visual sentence to infer the solution for non-verbal reasoning questions which are prevalent in IQ tests (masked image, second from the right). Left are successful cases and right is failure. We find that the model often infers and applies the abstract visual pattern correctly. So, we graciously hand over to you, our gentle reader, the task of pondering whether our modest LVM also exhibits the much-vaunted 'Sparks of AGI'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Task Compositing. Figure 8 demonstrates compositing several tasks together within a single prompt. Here, we demonstrate the rotation task together with the novel keypoint correspondence task and ask the model to continue the pattern. The model can successfully combine these two at test time, demonstrating some degree of compositionality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.729, + 0.292, + 0.745 + ], + "angle": 0, + "content": "5.4. Miscellaneous Prompts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Here we try to see how far we can push our model by offering it various prompts it has not seen before. Figure 9 shows a few such prompts that happened to work reasonably well. Figure 10 shows some prompts which are not easily describable by words – these are the type of tasks where LVMs might eventually outshine LLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Figure 11, we show initial qualitative results on a typical visual reasoning question as found on non-verbal human IQ tests (Raven's Progressive Matrices [67]). With considerable squinting, one could imagine the LVM having a" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.093, + 0.891, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.234, + 0.894, + 0.4 + ], + "angle": 0, + "content": "Figure 12. Failure Cases. This figure illustrates seven examples of failure cases: (1) Task confusion: the model interprets the counting task as style transfer, resulting in a pear-shaped apple. (2) Task entanglement: the model perceives the prompt as high frequencies from the entire image, rather than just from some parts of image (objects). (3) Wrong instance: the model correctly identifies rotation but mistakenly generates a brush instead of a manta. (4) Outlier detection versus generation: rather than detecting an outlier, the model directly generates it. (5) Digits: the model fails to represent digit sequences. (6) Tokenizer: poor performance with some out-of-distribution synthetic data generation (wrong background). (7) Sequence degeneration: sometimes frame predictions go astray." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.405, + 0.894, + 0.526 + ], + "angle": 0, + "content": "latent ability for grasping abstract visual patterns and applying them to extrapolate the shown visual sequence. We also perform perplexity analysis on 10 classic Raven 5-way multiple-choice Matrices in Table 2, choosing the answer with lowest perplexity under our model. It shows that our results are better than chance. To further understand this process, we generate 900 new Raven-like synthetic tests to study individual factors: color, shape, and size following [8]." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.528, + 0.847, + 0.564 + ], + "angle": 0, + "content": "
Raven IQ matricesshapecolorsize
chance20%33%33%50%
Ours30%45%42%94%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.568, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Table 2. Performance on 10 Raven's Progressive Matrices and 900 synthetic reasoning tasks by lowest perplexity under our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.6, + 0.64, + 0.613 + ], + "angle": 0, + "content": "5.5. Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Figure 12 shows some typical failure cases of the current model. One common element, the use of visual prompt to define a task is often under-constrained (more so than in language, since images are very high-dimensional), or the requested task might be beyond the capabilities of the current system. Other, more mundane failures involve issues with the tokenizer and lack of high-quality video training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.894, + 0.895 + ], + "angle": 0, + "content": "Acknowledgements: We are grateful to many friends and colleagues for discussions and comments on this work, including Yossi Gandelsman, Aleksander Holynski, Angjoo Kanazawa, Qianqian Wang, Sophia Koepke, Quoc Le, Chen Liang, Zichao Li, Ekin D. Cubuk, Assaf Shocher, Amir Zamir, Carl Vondrick, Ludwig Schmidt, Aviral Kumar, Xiaolong Wang and Miki Rubinstein. We thank the support from Google during Yutong's part-time internship. The work has been supported, in part, by ONR MURI N0001 4-22-1-2773, N00014-21-1-2801, and N00014-21-1-2812, ERC HOLI, Apple Graduate Fellowship to Yutong Bai, and compute donation via the Google TPU Research Cloud. Darrell's group is supported in part by the US DoD, BAIR&BDD industrial consortia." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22868" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.438, + 0.129 + ], + "angle": 0, + "content": "[1] Common crawl repository. https://commoncrawl.org/." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.131, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1692-1700, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.202, + 0.471, + 0.258 + ], + "angle": 0, + "content": "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.26, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[4] Okan Arikan and D. A. Forsyth. Interactive motion generation from examples. ACM Trans. Graph., 21(3):483-490, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.303, + 0.471, + 0.33 + ], + "angle": 0, + "content": "[5] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.333, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[6] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13137-13146, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.404, + 0.471, + 0.445 + ], + "angle": 0, + "content": "[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.447, + 0.471, + 0.502 + ], + "angle": 0, + "content": "[8] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.504, + 0.471, + 0.544 + ], + "angle": 0, + "content": "[9] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.548, + 0.471, + 0.617 + ], + "angle": 0, + "content": "[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.619, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.69, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[12] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.733, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[13] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.79, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[14] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[15] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.151, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[16] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.207, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[17] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.251, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.334, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[19] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.378, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[20] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.463, + 0.894, + 0.544 + ], + "angle": 0, + "content": "[21] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2634-2641, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.547, + 0.894, + 0.63 + ], + "angle": 0, + "content": "[22] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.632, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.689, + 0.894, + 0.743 + ], + "angle": 0, + "content": "[24] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE international conference on computer vision, pages 2051-2060, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.746, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[25] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.803, + 0.894, + 0.885 + ], + "angle": 0, + "content": "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.888, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[27] Alexei A. Efros and William T. Freeman. Image quilting for" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22869" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "texture synthesis and transfer. In Proceedings of SIGGRAPH 2001, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.123, + 0.472, + 0.165 + ], + "angle": 0, + "content": "[28] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In IEEE International Conference on Computer Vision, pages 1033-1038, Corfu, Greece, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.168, + 0.472, + 0.223 + ], + "angle": 0, + "content": "[29] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.226, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[30] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In CVPR 2011, pages 3281-3288. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.27, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[31] David Donovan Garber. Computational Models for Texture Analysis and Texture Synthesis. PhD thesis, University of Southern California, Image Processing Institute, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.315, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[32] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.373, + 0.471, + 0.428 + ], + "angle": 0, + "content": "[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.431, + 0.471, + 0.498 + ], + "angle": 0, + "content": "[34] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence, 38(1):142-158, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.502, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[35] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.602, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[36] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.471, + 0.771 + ], + "angle": 0, + "content": "[37] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.774, + 0.471, + 0.842 + ], + "angle": 0, + "content": "[38] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[39] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[40] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.177, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[41] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In International Conf. on Computer Vision (ICCV), pages 3192-3199, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.219, + 0.893, + 0.273 + ], + "angle": 0, + "content": "[42] Lipeng Ke, Ming-Ching Chang, Honggang Qi, and Siwei Lyu. Multi-scale structure-aware network for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 713-728, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.275, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[43] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6129-6138, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.345, + 0.894, + 0.384 + ], + "angle": 0, + "content": "[44] Lucas Kovar, Michael Gleicher, and Frederic Pighin. Motion Graphs. In Proceedings of SIGGRAPH '02, San Antonio, TX, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.386, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[45] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.442, + 0.894, + 0.509 + ], + "angle": 0, + "content": "[46] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.511, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[47] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.581, + 0.893, + 0.636 + ], + "angle": 0, + "content": "[48] Jehee Lee, Jinxiang Chai, Paul Reitsma, Jessica K. Hodgins, and Nancy Pollard. Interactive control of avatars animated with human motion data. ACM Transactions on Graphics (TOG), 21(3):491 - 500, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.637, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[49] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.692, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[50] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.749, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[51] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.818, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22870" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.471, + 0.189 + ], + "angle": 0, + "content": "[54] Shikun Liu, Linxi Fan, Edward Johns, Zhiding Yu, Chaowei Xiao, and Anima Anandkumar. Prismer: A vision-language model with an ensemble of experts. arXiv preprint arXiv:2303.02506, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[55] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[56] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[57] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[58] Mathew Monfort, Bowen Pan, Kandan Ramakrishnan, Alex Andonian, Barry A McNamara, Alex Lascelles, Quanfu Fan, Dan Gutfreund, Rogério Schmidt Feris, and Aude Oliva. Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9434-9445, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[59] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE conference on computer vision and pattern recognition, pages 2408-2415. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.567, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[60] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.637, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[61] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.692, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[62] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[64] Kris Popat and Rosalind W. Picard. Novel cluster-based probability model for texture synthesis, classification, and compression. In Proc. SPIE Visual Comm. and Image Processing, 1993." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[65] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[66] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[67] John C Raven. Standardization of progressive matrices, 1938. British Journal of Medical Psychology, 1941." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.233, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[68] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.316, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[69] Arno Schodl, Richard Szeliski, David H. Salesin, and Irfan Essa. Video textures. In Proceedings of SIGGRAPH '00, pages 489-498, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.359, + 0.894, + 0.427 + ], + "angle": 0, + "content": "[70] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.428, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.525, + 0.894, + 0.607 + ], + "angle": 0, + "content": "[72] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.608, + 0.894, + 0.649 + ], + "angle": 0, + "content": "[73] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.65, + 0.894, + 0.677 + ], + "angle": 0, + "content": "[74] Claude E. Shannon. A mathematical theory of communication. Bell Sys. Tech. Journal, 27, 1948." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.679, + 0.894, + 0.705 + ], + "angle": 0, + "content": "[75] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.706, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[76] Gunnar A Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.803, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[77] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.86, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[78] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22871" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.162 + ], + "angle": 0, + "content": "[79] X. Soria, E. Riba, and A. Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1912-1921, Los Alamitos, CA, USA, 2020. IEEE Computer Society." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.164, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.234, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[81] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, koray kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[82] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, New York, New York, USA, 2016. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[83] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.472, + 0.475 + ], + "angle": 0, + "content": "[84] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[85] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th International Conference on Machine Learning, page 1096-1103, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.472, + 0.602 + ], + "angle": 0, + "content": "[86] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.674, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[88] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of laion-2b, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.703, + 0.472, + 0.744 + ], + "angle": 0, + "content": "[89] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.746, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[90] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[91] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[92] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[93] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[94] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.249, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.334, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[96] Yingchen Yu, Fangneng Zhan, Rongliang Wu, Jianxiong Pan, Kaiwen Cui, Shijian Lu, Feiying Ma, Xuansong Xie, and Chunyan Miao. Diverse image inpainting with bidirectional and autoregressive transformers. In Proceedings of the 29th ACM International Conference on Multimedia, pages 69-78, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.418, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[97] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.488, + 0.894, + 0.529 + ], + "angle": 0, + "content": "[98] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.531, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[99] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[100] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[101] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22872" + } + ] +] \ No newline at end of file diff --git a/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_origin.pdf b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a1c2319daad18034508636b3d38509620347a3b --- /dev/null +++ b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/262d155e-8b9e-40b4-8485-80e9c6034a6b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:372ad29937c9e022db595f9e8d150d37340cac570b577d1771e036133854428f +size 14575705 diff --git a/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/full.md b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d09712ce46a4b162404fc5a239ef9d7bc8a9b343 --- /dev/null +++ b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/full.md @@ -0,0 +1,424 @@ +# Sequential Modeling Enables Scalable Learning for Large Vision Models + +Yutong Bai $^{1,2,*}$ + +Alan L. Yuille + +Xinyang Geng $^{1,*}$ + +Trevor Darrell + +Karttikeya Mangalam + +Jitendra Malik1 + +Amir Bar + +Alexei A. Efros + +$^{1}$ UC Berkeley (BAIR) + +2Johns Hopkins University + +# Abstract + +We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, "visual sentences", in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions without needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion tokens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time. + +# 1. Introduction + +Large language models (LLMs) such as GPT [10] and LLaMA [80] have taken the world by storm. What would it take to build a Large Vision Model (LVM)? From the animal world, we know that visual competences are not dependent on language. In particular, many experiments have shown that the visual world of non-human primates is remarkably similar to that of humans. So while the space of vision-language models such as LLaVA [53] is interesting and worthwhile to pursue, in this paper we seek an answer to a different question – how far can we go from pixels alone? + +The key features of contemporary LLMs that we seek to emulate in LVMs are: 1) scaling in the presence of big data, and 2) flexible specification of tasks through prompting (in-context learning). How do we achieve this? As usual, there are three main components that must be specified: + +Data: We want to exploit all the remarkable diversity in visual data. First of all, just raw unannotated images and videos. Next, we want to exploit the variety of annotated visual data sources that have been produced over the last + +couple of decades – semantic segmentations, depth reconstructions, keypoints, multiple views of 3D objects, among others. We define a common format, “visual sentences”, in which to represent these different annotations without needing any meta-knowledge beyond the pixels. The total size of our training dataset is 1.64 billion images/frames. + +Architecture: We use a large transformer architecture (3 billion parameters) trained on visual data represented as sequence of tokens, using a learned tokenizer that maps each image to a string of 256 vector-quantized tokens. + +Loss function: We draw inspiration from the natural language community, where masked token modeling has given way to sequential autoregressive prediction. Once images/videos/annotated images can all be represented as sequences, we can train the model to minimize the cross-entropy loss for predicting the next token. + +With this extremely simple design, we demonstrate some noteworthy behaviors: + +- Appropriate scaling behavior as one increases model size and data size. +- Many different vision tasks can now be "solved" by designing suitable prompts at test time. While the results don't show as high performance as bespoke, specifically-trained models, the fact that so many tasks are all addressed by a single vision model is quite encouraging. +- We see a clear benefit of the amount of unsupervised data on the performance on various standard vision tasks. +- We see a hint of an ability for general visual reasoning – handling out-of-distribution data, and performing novel tasks. But further investigation is needed. + +# 2. Related Work + +Pretrained Vision Models. The value of using pretrained models (such as ImageNet-pretrained AlexNet [45]) has been demonstrated as far back as 2015 in R-CNN [34], and it has since become standard practice in computer vision. Self-supervised pretraining was proposed as a way to vastly increase the amount of data available for pretraining [16, 25, 37, 61, 62, 99]. Unfortunately, this was not very successful, likely because the CNN-based architectures + +of that time did not have enough capacity to absorb the data. With the introduction of Transformers [84], which have much higher capacity, researchers revisited self-supervised pretraining, and showed that transformer-based masked image reconstruction approaches, such as BEiT [7], MAE [38], SimMIM [91], MaskGIT [13] perform vastly better than their CNN-based counterparts [62]. Yet, despite their recent successes, current pretrained vision-only models have had trouble scaling up to the really large datasets, such as LAION [72]. + +Multi-task Learning and In-context Learning. From the classic one-model-per-task setups, computer vision is slowly moving toward a single model performing multiple different tasks. Various multi-task learning approaches [24, 40, 43, 73, 97] exist but they are typically limited to a fixed, predefined number of tasks. More recently, methods inspired by in-context learning in LLMs forgo any notion of tasks and instead let the model infer the task directly from the input prompt. For example, Visual Prompting [8, 87] takes in a task input/output example pair and a query image at test time, concatenates them into a single 2-by-2 image, and uses inpainting to generate the desired output. But, since the inpainting is performed using a variant of MAE [38], the same problems with scaling are inherited by these approaches. + +Auto-regressive Visual Models. The idea of using autoregressive models for synthesizing visual data goes back at least 70 years. Inspired by Shannon's use of $N$ -grams to synthesize language [74, 75], a number of works, starting with Attneave's seminal 1954 paper [5], applied this idea to sequentially synthesizing pixels [28, 31, 39, 64], image patches [27], video frames [69], and motion capture data [4, 44, 48]. As deep models became popular, newer works replaced $N$ -grams with RNNs or CNNs for pixel [81, 82] and video [66] synthesis. Most recently, transformer-based autoregressive visual generation methods have been proposed [15, 29, 94, 96], and, combined with language, have demonstrated impressive image synthesis results, e.g. Parti [95]. + +# 3. Data + +"Data! Data! Data! I can't make bricks without clay!" + +SHERLOCK HOLMES + +The key requirement of any Large Pre-trained Model is that it must be trained on vast amounts of data. For language models, very large and very diverse datasets are fairly easy to obtain. For instance, the popular Common Crawl repository [1] contains 250 billion web pages spanning the entire Web, is extremely diverse, and includes "natural demonstrations" like language translations, question answering, etc. In computer vision, we are still very far from having a data source of comparable size and diversity. One of the + +central contributions of our work is the first step toward curating such a dataset that we call Unified Vision Dataset v1 (UVDv1). To assemble it, we leverage many different sources of visual data: (1) unlabelled images, (2) images with visual annotations, (3) unlabelled videos, (4) videos with visual annotations, and (5) 3D synthetic objects. The unlabeled images, which represent over $80\%$ of our data, capture a huge cross-section of our visual world, and provide the required diversity, at the cost of lower quality. Images with annotations have a much more constrained distribution, but are usually of higher quality. Video data is even more constrained (typically, to human-centric activities), but is an invaluable source of temporal data. Renderings of 3D synthetic objects are the lowest in diversity but can provide valuable hints about the behavior of 3D structures. Importantly, UVDv1 is a purely visual dataset, with no non-visual meta-data (e.g. text) included. All together, UVDv1 contains 1.64 billion images. + +Another important difference from Large Language Models is that language data has a natural, unified one-dimensional structure for all the data - a stream of text. Unfortunately, this is not the case for visual data, with different sources all having different structures. In this work we propose visual sentence as a unified unit of visual data, which enables us to train scalable models from a diverse set of sources. A visual sentence is simply a sequence containing one or more images followed by an end-of-sentence (EOS) token. Figure 1 shows how the various data sources are partitioned into visual sentences. In particular: + +Single images. A single image itself represents the simplest form of a visual sentence - $\{$ image, EOS $\}$ . We use the filtered subset of 1.49 billion images [88] from the LAION 5B [71] dataset. This is by far the largest part of our data, comprising $88.5\%$ . + +Image sequences. A sequence of images is a natural form of visual sentence. We create such sequences by sourcing video data from a wide range of existing datasets [11, 12, 21, 35, 36, 46, 50, 51, 55, 57-59, 63, 68, 76-78, 92, 93]. Visual sentences of 16 frames are formed by randomly sampling the videos at three different strides (10, 20, and 30). In addition, we utilize synthetic 3D objects from the Objaverse Dataset [22] to generate object-centric multiview sequences for a variety of objects. For each object, we sample one radius length between the object center and the camera from 1.5 to 2.2, and sample one constant elevation from -45 degrees to 45 degrees, then traverse different views of the object by changing the azimuth with a step length of 15 degrees and render 24 views. We rendered 42000 such sequences in total for training and 8000 for testing. Finally, we can also represent images belonging to the same semantic category as being (part of) a sequence. We use categories from ImageNet, concatenating together groups of images (2,4,8, or 16) from the same category into a 16-image long + +# Visual Sentences + +![](images/d35e2bd48e9c47f6b74a859f3b83dd27df6d33f2b69a9072df35f789704ac45b.jpg) + +# + +Single images, e.g. LAION + +![](images/12d009d3676495b166dbbe08977b4b1e87c87b9f7b442350b391e4ec75091050.jpg) + +# + +![](images/05f3f0a5ab5cd60e92de770e16445b0dd90ff843edb65f124df856ab180ca62e.jpg) + +![](images/0f3c7a1217f7766d48c9ed6a38f238311059d4a1871e62ee48200e5d7bbaae5c.jpg) + +# + +Image sequences, e.g. videos, 3D rotations, synthetic viewpoints + +![](images/1d222d66e84a2f4a4c2b7ee1e6eabb187cdafc56075e3b39d892f11209ffec40.jpg) + +![](images/4e992a3888d53e2b91b820826178570d10fcd3347accfd212ca9abcd441307a4.jpg) + +![](images/ca73993924913fd58018f80182d7af16263a9549ab3f3023ec44486a49ae395a.jpg) + +![](images/0fd55ba213dd6da12d30d9a8d819b347f3b159d14d2fda33de8cdfbd9721d76e.jpg) + +·· + +![](images/255a7cfdc4aa8e6b5e20456b7b1cfa02fa00f8723cebe2a23cb6da69e51a5cba.jpg) + +![](images/dd8334a170b70fcb859b027fd60a94f35a8664fe7552b4741c286bdb84ea071e.jpg) + +![](images/331223a6c6d61b18c8944adeab785e30c0b8f752f8dea5a60d49242c860ece55.jpg) + +![](images/19e215c306c745e4bf3dc6222338abf3c8daf8c5055c01aec6e19785427b50f2.jpg) + +![](images/ebff1d7b9c084e04ddfa542085c32a0fdaf45576c3044fd9c83a087af92163f9.jpg) + +![](images/e1c32ea914f99cb0e4497011c685eef33b60aaf9d9634450129a8f6818dd8bfe.jpg) + +![](images/b16445fccde8375bf0353909a4fe5598943586ce1d81606ba02528237df392dc.jpg) + +![](images/b92ce6c2efa5252166ca3c69fab5ad501908254c7fd36d8f0300aa134c5c982a.jpg) + +# + +![](images/b7369a4bf92f2b4925fd4e339ac810b52f8b79b9bb9d38e815573a5af0970412.jpg) + +![](images/ca16e771aa9035b03660710a61f21e0b5d2e326121d2b2fe4adc643741748ebd.jpg) + +![](images/1907b449fc7e84c31b86d510621f04d7a10eceb1af56844e3f91317e275becb9.jpg) + +![](images/f5002f47cd7bd159b9b70ab0c1d991b928c674f1b3d1f1a21d1f4d35912b4594.jpg) + +![](images/5b76b3bdf4750d333bf6fa8f8c457448dfc22a49025114f1fa0e193d516651c2.jpg) + +![](images/1f4c16f46e6a1cecc9220bf9d9301c8ea722163837abae0c25d2aa9112e22591.jpg) + +·· + +![](images/74af6dd9eecf0abfa318bb078ac44187b6409168bd6e95e0fe6e730183a10b3b.jpg) + +![](images/99a314ff597a49d586745816c389479a64fbfeb9c08b4a72e4d604cdeaf05d92.jpg) + +![](images/367b5a32941f61bcd90e211a5554233c57d110ae4269682bd8506f3b635e1f19.jpg) + +![](images/dbd571c50da373321b7676eaa1d62eea07e7aa01458f1498d5d41cf1d0661153.jpg) + +![](images/720368fcfc2df52da3021da686678a1727eae2a00aed10853f0001553acd417f.jpg) + +![](images/0a27f1522d2b55bd9909c3a55fe25ec8fbf11936e417a9198239d62139a1fb0e.jpg) + +Edges with free form notation, 1. object detection + frame segmentation + +![](images/64e2176d1c23a956a0faaa839fa2b19b57db455b8435cb260d37c462d5c1ca4e.jpg) + +![](images/21449481cfe6c5fc8773982e32160a2b113987b694c9fa9b7328a8f5fd2ce00c.jpg) + +![](images/a5531978c2355c6e0f79882390f17e686bc340009960ced7cdb1d28041394ffb.jpg) + +![](images/95eafdfc25b5f6a6f846f153dc77094fb8ea7d2e80f91e901928e661d4cc4a6a.jpg) + +![](images/cfba108de1d02b91068e23bb8034bd8ef6c24066b78c1ba5fa2cf42685e518d8.jpg) + +![](images/9860739c00b2620b4053e324d17626f67e8873b110423d67fe6bfb88790310e5.jpg) + +![](images/0d3111027ee1360470b0d542ee8894e9f968d67b26f68644f5337013ca754107.jpg) +Figure 1. Visual sentences allow us to format diverse vision data into the unified structure of image sequences. + +![](images/f38fa1a2215d7404b1d69f6fd93355a7b16fabd50a6a1f905b7090661c3e4982.jpg) + +![](images/22307d15ebf8d2b1b32c2d6c478fb8af744d90faf7d22338022237ef0e171f3b.jpg) + +Videos with annotation, e.g. video segmentation + +visual sentences. + +Images with annotations. To handle different types of image annotations in a uniform way, we choose to represent all annotations as images. Some data types, e.g. semantic segmentation maps [100], edge maps [79], depth [65] and normal images [6], are already represented this way. For others we apply tailored methods for each specific annotation type: 1) Object Detection: We create annotations by overlaying a color-coded bounding box around each object, following the methodology in [14]; 2) Human Pose: Human skeletons are rendered in pixel space, adhering to the OpenPose format, utilizing MMPose [19]; 3) Depth Estimation, Surface Normal, and Edge Detection: given ImageNet and COCO images, we generate annotations in line with the protocols from [54]. 3) Style Transfer [9], De-rain [98], De-noise [85], Low Light Enhancement [89], and Stereo Datasets [33]: These are all represented as image pairs (e.g. input/output). 4) Colorization: We convert ImageNet images to greyscale, producing image pairs. 5) Inpainting: The process involves randomly adding black-colored boxes + +in images to simulate corruption, resulting in image pairs. For all the above annotation types, we can create visual sentences by concatenating 8 image pairs of the same annotation type into a 16-image visual sentence. For datasets containing $k$ different annotations for the same image we use a different approach: for each set of $1 + k$ images (input plus $k$ annotations), we randomly select $m$ elements, where $m \leq k + 1 \leq 16$ . These m-tuples are then concatenated to form visual sequences. + +Image sequences with annotations. When converting annotated video data (VIPSeg [56], Hand14K [30], AVA [59], JHMDB [41]) to visual sentences, we apply two complementary strategies. The first is similar to how we treat image data with paired annotations: each visual sentence is constructed by concatenating frames with their annotations - {frame1,annot1,frame2,annot2,...}. The second method involves grouping multiple frames followed by their corresponding annotations - {frame1,frame2,annot1,annot2,...}. + +We present a detailed summary of all the data sources, annotation type and data statistics of UVDv1 in the Appendix. + +# 4. Approach + +In this section, we describe the design of our autoregressive Large Vision Model. Unlike text data, which naturally exhibits discrete sequential structure, it is not straightforward to model image pixels in visual sentences. In this work, we take a two-stage approach: 1) train a large visual tokenizer (which operates on individual images) to convert each image into a sequence of visual tokens; 2) train an autoregressive transformer model on visual sentences, each represented as a sequence of tokens. We summarize our approach in Figure 2. + +# 4.1. Image Tokenization + +While the visual sentences exhibit a sequence structure between consecutive images, we don't have such natural sequence structure within an image. Therefore, in order to apply a transformer model to images, prior works typically do one of the following: either divide the image into patches in scan-line order, and treat that as a sequence [26], or use a pre-trained image tokenizer, such as VQVAE [83] or VQGAN [29], to cluster image features into a grid of discrete tokens, which, again, are turned into a sequence in scan-line order. We adopt the latter approach since the discrete categorical output from a model naturally forms a probabilistic distribution that one can easily sample from, enabling flexible conditional generation of new images within a visual sentence. + +Specifically, we employ semantic tokens generated by a VQGAN model, a concept introduced by Esser et al [29]. This framework consists of an encoding and a decoding mechanism, featuring a quantization layer that assigns input images to a sequence of discrete tokens from an established codebook. Our encoders and decoders are constructed purely with convolutional layers. The encoder is equipped with several downsampling modules to contract the spatial dimension of the input, whereas the decoder is fitted with an equivalent series of upsampling modules to restore the image to its initial size. For a given image, our VQGAN tokenizer produces 256 discrete tokens. + +It is important to note that our tokenizer operates on individual images independently, rather than on the entire visual sentence at once. This independence allows us to decouple the tokenizer training from the downstream Transformer model so that the tokenizer can be trained on a dataset of single images without having to consider the distribution of visual sentences. + +Implementation Details: We adopt an off-the-shelf VQGAN architecture from Chang et al. [13]. We follow the exact configuration in Chang et al. [13], which uses a down-sampling factor of $f = 16$ and codebook size 8192. This means that for an image of size $256 \times 256$ , our VQGAN tokenizer produces $16 \times 16 = 256$ tokens where each can take 8192 different values. We found that using the results of an ImageNet pre-trained tokenizer did not generalize well + +![](images/c645823cc553d995a033025c581a2ab17f86f35e24183231853b9e86ec5e74e4.jpg) +Figure 2. Architecture of LVM. We first convert individual images from a visual sentence into discrete tokens using a VQGAN encoder. The resulting tokens from all images are then concatenated into a 1D sequence, and fed into an autoregressive Transformer model to predict the next token in the sequence. The predicted visual tokens are decoded into images using the VQGAN decoder. + +beyond ImageNet images. Therefore, we train our own tokenizer on a 1.5B subset of the LAION 5B dataset [71]. + +# 4.2. Sequence Modeling of Visual Sentences + +After converting images into discrete tokens with VQGAN, we treat our visual sentence as a unified sequence by concatenating the discrete tokens from multiple images into a 1D sequence. Importantly, all visual sentences are treated equally – we do not make use of any special tokens to indicate particular tasks or formats. We train a causal Transformer model with the next token prediction objective using a cross-entropy loss, similar to the standard approach for language models [10]. Training the model the same way on all visual sentences enables the model to infer the relation between images from context instead of from task- or format-specific tokens. This gives the model an opportunity to generalize to other, unseen visual sentence structures. + +Implementation Details: After tokenizing each image in a visual sentence into 256 tokens, we concatenate them to form a 1D sequence of tokens. On top of the sequences of visual tokens, our Transformer model is virtually the same as an autoregressive language model, so we adopt the Transformer architecture of LLaMA [80], a popular open-source language model with widely available implementations. We use a context length of 4096 tokens, which can fit 16 images under our VQGAN tokenizer. Similar to language models, we add a [BOS] (begin of sentence) token to the beginning of each visual sentence and an [EOS] (end of sentence) token to the end, and use sequence concatenation [18] during training time to improve efficiency. We train our model on our entire UVDv1 dataset (420 billion tokens) using one epoch (simple epoch training is standard in language models to avoid poten + +![](images/ddb3d3ccef716ea90bf68423385be11ea456d862ac1dfae940f8f1562d1cd0c1.jpg) +Figure 3. Training loss for the 300M, 600M, 1B, and 3B models. All models are trained on 420B tokens (1.64B images). The training scales well with model sizes. Zoom in for more details: notice how the shape of the 4 curves is very similar. This is the consequence of the deterministic training pipeline, so the loss fluctuations are determined by the intrinsic difficulty of the data stream. + +tial overfitting). We train 4 models with different numbers of parameters: 300 million, 600 million, 1 billion and 3 billion, following the same training configurations. We provide the detailed training hyperparameters in Appendix 5.5. + +# 4.3. Inference by Visual Prompting + +Since the autoregressive Transformer in our model outputs a probability distribution of the next token conditioned on previous tokens, we can easily sample from this distribution to generate new visual tokens that complete a visual sentence. To use the model for downstream tasks, one can construct a partial visual sentence that defines a task at test time, and apply the model to generate the output. This is similar to in-context learning in language models [10] or visual prompting in computer vision [8, 39]. + +# 5. Experimental Results and Analysis + +In this section, we evaluate the scaling abilities of our trained model, as well as its ability to understand and answer a range of diverse prompted tasks. + +# 5.1. Scalability + +We investigate the scaling behavior of our model in terms of the training loss and downstream task performance as we increase the model size as well as the number of tokens seen during training. + +Training loss. We first inspect the training loss of LVM with different parameter sizes, which we present in Figure 3. Since all our models are trained for only one epoch on the dataset, the model sees a given data sample just once, and therefore the training loss at any point during training is very similar to the validation loss. One can observe that as + +![](images/37117ed5e9cc178a1818577203dd9a4135c7b467e2fa5215636e7d1a628717a0.jpg) + +![](images/9f27582b2f581d86014004d20a0c3abb8f799a3556435c44f9f1389173e14116.jpg) + +![](images/848fc8927e203ecfcaf895d0584e5b9cdc766e23693371a6a9a0d859b686b350.jpg) +Figure 4. Larger LVMs perform better on downstream tasks. We evaluate LVMs of varying sizes on 4 different downstream tasks, following the 5 shot setting on the ImageNet validation set and report the perplexity. We find that perplexity decreases with larger models across all tasks, indicating strong scalability. + +![](images/696a76a887e9610f75ac82333f4c1f04e0aefc80e4f2cbb8c426f4765b1c6a18.jpg) + +training progresses: 1) the training loss (perplexity) of the models, regardless of their size, continues to decrease; 2) as we increase the size of the model (parameter count), the loss decreases faster. These observations indicate that LVM shows strong scalability behavior with both larger models and more data. + +Scalability on downstream benchmarks. While the LVM overall loss scales well during training, there is no guarantee that the better overall model would also perform better on a given specific downstream task. Therefore, we evaluate different sizes of models on 4 downstream tasks: semantic segmentation, depth estimation, surface normal estimation, and edge detection. We evaluate these tasks on the ImageNet validation set and generate all the annotations using the corresponding method described in Sec. 3. For each task, we give 5 pairs consisting of the inputs and corresponding ground-truth annotations as well as the query image as input prompt and evaluate the perplexity of the ground-truth annotation under our model's prediction of the next 256 tokens (one image). We report the results in Figure 4. We see that larger models indeed attain lower perplexity across all tasks, showcasing that our scalable overall performance does transfer to a range of downstream tasks. + +Dataset ablation. While LVM attains better performance with larger models and more data, it is natural to ask whether each data component we collect in UVDv1 helps. To answer this question, we conduct an ablation study on our dataset by training several 3B models on subsets of our dataset, and compare their performances on downstream tasks. We use the same 4 downstream tasks and settings as before and present the results in Figure 5. We observe that each + +![](images/e20f3c458684b4b7616c4b088c3c841732a416a0d8b0950ed539514001778f5c.jpg) +Figure 6. Frame predictions. LVM predicts the next frames (marked in red) given previous video frames as prompt. The results reveal that the LVM can predict the video frames while considering dynamic objects and camera motion. + +data component contributes positively to the downstream tasks. LVM not only benefits from larger data, but also improves with more diversity in the dataset, which includes both annotated and unsupervised image and video data. + +# 5.2. Sequential Prompting + +We begin with the most intuitive and straightforward approach to visually prompt the LVM: sequential reasoning. Here the prompt construction is very simple: we present the model with a sequence of 7 images and ask it to predict the next image (256 tokens). + +Video frame prediction. The most direct task for sequential prompting is video prediction. Figure 6 presents several next frame prediction examples, prompted by sequences from the Kinetics-700 validation set. At the top, 7 frame prompts (blue border) are followed by the predicted frame (red border). We observe a certain degree of inferential ability regarding spatial positioning, viewpoint and object understanding. Perplexity of prediction on Kinetics val set is 49.8. The last 4 rows show predictions with longer context (15 frames) and a longer prediction (4 frames). See Figures 16 to 21 in the Appendix for many more examples. + +Rotation and Category prediction. The same type of simple sequential prompting can be used in other ways as well. For example, Figure 15 shows how prompting the model with a sequence of 3D rotations of a synthetic object around an arbitrary axis allows it to predict further rotation. Or we can think of a list of items of a given category as a sequence and predict other ideas in that same category, as shown in Figure 14. Note that, while the system was trained + +on groups of images from the same ImageNet category, here the prompt consists of sketches, which have not been seen in any annotated data. + +Context length analysis. Next we ask how much temporal context is required to accurately predict the subsequent frame? We assessed the model's frame generation perplexity when prompted with a context of varying lengths (1 to 15 frames). On the Kinetics-700 val set, we see a clear improvement in perplexity from 1 to 11 frames after which it stabilizes (from $62.1 \rightarrow 48.4$ ). + +# 5.3. Analogy Prompting + +Our study progresses by evaluating a more complex prompting structure, which we call 'Analogy Prompting'. This method challenges the model to comprehend analogies of arbitrary length and complexity, thereby testing its advanced interpretative abilities. + +Qualitative Results. Figure 7 shows a sampling of qualitative results with analogy prompting on a number of tasks. The prompts consist of a sequence of 14 images giving examples of various tasks, followed by a 15th query image. Given each prompt, the next image predicted is the result. The top part of the figure shows several example prompts defining tasks that were part of the training set (but these actual images were never seen at training). The bottom part of the figure demonstrates generalization to tasks never shown at training. See Appendix for many more qualitative examples. + +Unseen Tasks and Dataset. We present the results for keypoint detection on Pascal 3D+ [90], evaluated using the standard Percentage of Correct Keypoints (PCK) metric with + +![](images/d8503a81df206b6a830757b15e66da6a2a3d8c51e15a6d1f1be3ce1e025f49cb.jpg) + +![](images/afb4ba85aaea0b31ac16e2067dfdcc7c253e5658ae115e65c070e04c4410b103.jpg) +Figure 7. In and out of distribution prompting examples. Every row is a prompt that contains a sequence of images interleaved with annotations, followed by a query. The last image is predicted by the model (marked in red). The last 5 rows show examples where the query image is out of distribution (painting, sketch, etc) for the task it was trained for. +Figure 8. Task Compositionality. Examples of prompts that combine two different tasks - object rotation and keypoint tracking. + +
ModelForeground SegmentationSingle Object DetectionColorization ↓
Split 0Split 1Split 2Split 3Split 1Split 2Split 3Split 4MSELPIPS
MAE (IN-1k)1.926.763.854.571.371.981.621.621.130.87
MAE-VQGAN (IN-1k)2.227.075.486.283.343.212.802.803.310.75
MAE (CVF)17.4225.7018.6416.535.494.985.245.840.430.55
MAE-VQGAN (CVF)27.8330.4426.1524.2524.1925.2025.3625.230.670.40
Ours48.9451.2947.6650.8248.2549.6050.0848.920.510.46
+ +Table 1. Comparison with Visual Prompting [8]. For Foreground Segmentation and Single Object Detection, we report the mIOU score. For Colorization, we report the MSE and LPIPS. + +a of threshold 0.1. Remarkably, LVM achieves a PCK of 81.2 without training on this dataset, demonstrating impressive generalization capabilities. In comparison, we show + +some existing task-specific model: StackedHourglass [60] scores 68.0 PCK, MSS-Net [42] achieves 68.9 PCK, and StarMap [101] registers 78.6 PCK. + +Comparison with Visual Prompting. The closest approach to ours that also allows for defining arbitrary tasks is Visual Prompting [8]. In Table 1, we compare various visual prompting models on few-shot segmentation, object detection, and colorization tasks. Note that our sequential LVM beats previous approaches on almost all tasks. + +![](images/c41df0ad173827242d02488f71f3dee33fc9650e2e50abd2e6625b5ddea3cda4.jpg) +Figure 9. Miscellaneous Prompting. A variety of simple vision tasks, such as object replication (top), relighting (middle), and zooming in (bottom), can be simply specified via a suitably chosen visual sentence prompt that expresses the task to the LVM. + +![](images/3df5b978d7fc8aacad29b6f95fe2b47ab8a388d3e13a843d7a2a4685ede0a650.jpg) +Figure 10. What comes next? Tasks that are not always easily describable in language. + +![](images/e10c1d17c730dcf91646a885a62667de00a2c63b864c26f9e3da650ed17da03f.jpg) +Figure 11. "Sparks of AGI?" We prompt LVM with a masked reasoning visual sentence to infer the solution for non-verbal reasoning questions which are prevalent in IQ tests (masked image, second from the right). Left are successful cases and right is failure. We find that the model often infers and applies the abstract visual pattern correctly. So, we graciously hand over to you, our gentle reader, the task of pondering whether our modest LVM also exhibits the much-vaunted 'Sparks of AGI'. + +Task Compositing. Figure 8 demonstrates compositing several tasks together within a single prompt. Here, we demonstrate the rotation task together with the novel keypoint correspondence task and ask the model to continue the pattern. The model can successfully combine these two at test time, demonstrating some degree of compositionality. + +# 5.4. Miscellaneous Prompts + +Here we try to see how far we can push our model by offering it various prompts it has not seen before. Figure 9 shows a few such prompts that happened to work reasonably well. Figure 10 shows some prompts which are not easily describable by words – these are the type of tasks where LVMs might eventually outshine LLMs. + +In Figure 11, we show initial qualitative results on a typical visual reasoning question as found on non-verbal human IQ tests (Raven's Progressive Matrices [67]). With considerable squinting, one could imagine the LVM having a + +![](images/2cee850566bfc5566a34f6a35f432eeda073e4bc33191194e46dd3300caa2f80.jpg) +Figure 12. Failure Cases. This figure illustrates seven examples of failure cases: (1) Task confusion: the model interprets the counting task as style transfer, resulting in a pear-shaped apple. (2) Task entanglement: the model perceives the prompt as high frequencies from the entire image, rather than just from some parts of image (objects). (3) Wrong instance: the model correctly identifies rotation but mistakenly generates a brush instead of a manta. (4) Outlier detection versus generation: rather than detecting an outlier, the model directly generates it. (5) Digits: the model fails to represent digit sequences. (6) Tokenizer: poor performance with some out-of-distribution synthetic data generation (wrong background). (7) Sequence degeneration: sometimes frame predictions go astray. + +latent ability for grasping abstract visual patterns and applying them to extrapolate the shown visual sequence. We also perform perplexity analysis on 10 classic Raven 5-way multiple-choice Matrices in Table 2, choosing the answer with lowest perplexity under our model. It shows that our results are better than chance. To further understand this process, we generate 900 new Raven-like synthetic tests to study individual factors: color, shape, and size following [8]. + +
Raven IQ matricesshapecolorsize
chance20%33%33%50%
Ours30%45%42%94%
+ +Table 2. Performance on 10 Raven's Progressive Matrices and 900 synthetic reasoning tasks by lowest perplexity under our model. + +# 5.5. Failure Cases + +Figure 12 shows some typical failure cases of the current model. One common element, the use of visual prompt to define a task is often under-constrained (more so than in language, since images are very high-dimensional), or the requested task might be beyond the capabilities of the current system. Other, more mundane failures involve issues with the tokenizer and lack of high-quality video training data. + +Acknowledgements: We are grateful to many friends and colleagues for discussions and comments on this work, including Yossi Gandelsman, Aleksander Holynski, Angjoo Kanazawa, Qianqian Wang, Sophia Koepke, Quoc Le, Chen Liang, Zichao Li, Ekin D. Cubuk, Assaf Shocher, Amir Zamir, Carl Vondrick, Ludwig Schmidt, Aviral Kumar, Xiaolong Wang and Miki Rubinstein. We thank the support from Google during Yutong's part-time internship. The work has been supported, in part, by ONR MURI N0001 4-22-1-2773, N00014-21-1-2801, and N00014-21-1-2812, ERC HOLI, Apple Graduate Fellowship to Yutong Bai, and compute donation via the Google TPU Research Cloud. Darrell's group is supported in part by the US DoD, BAIR&BDD industrial consortia. + +# References + +[1] Common crawl repository. https://commoncrawl.org/. +[2] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1692-1700, 2018. +[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. +[4] Okan Arikan and D. A. Forsyth. Interactive motion generation from examples. ACM Trans. Graph., 21(3):483-490, 2002. +[5] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954. +[6] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13137-13146, 2021. +[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. +[8] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. +[9] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. +[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. +[12] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. +[13] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. +[14] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. + +[15] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. +[16] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. +[17] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. +[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. +[19] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020. +[20] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. +[21] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2634-2641, 2013. +[22] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. +[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[24] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE international conference on computer vision, pages 2051-2060, 2017. +[25] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. +[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. +[27] Alexei A. Efros and William T. Freeman. Image quilting for + +texture synthesis and transfer. In Proceedings of SIGGRAPH 2001, 2001. +[28] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In IEEE International Conference on Computer Vision, pages 1033-1038, Corfu, Greece, 1999. +[29] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. +[30] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In CVPR 2011, pages 3281-3288. IEEE, 2011. +[31] David Donovan Garber. Computational Models for Texture Analysis and Texture Synthesis. PhD thesis, University of Southern California, Image Processing Institute, 1981. +[32] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. +[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. +[34] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence, 38(1):142-158, 2015. +[35] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. +[36] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. +[37] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. +[38] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. +[39] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. + +[40] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In International Conference on Learning Representations, 2021. +[41] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In International Conf. on Computer Vision (ICCV), pages 3192-3199, 2013. +[42] Lipeng Ke, Ming-Ching Chang, Honggang Qi, and Siwei Lyu. Multi-scale structure-aware network for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 713-728, 2018. +[43] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6129-6138, 2017. +[44] Lucas Kovar, Michael Gleicher, and Frederic Pighin. Motion Graphs. In Proceedings of SIGGRAPH '02, San Antonio, TX, 2002. +[45] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. +[46] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. +[47] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017. +[48] Jehee Lee, Jinxiang Chai, Paul Reitsma, Jessica K. Hodgins, and Nancy Pollard. Interactive control of avatars animated with human motion data. ACM Transactions on Graphics (TOG), 21(3):491 - 500, 2002. +[49] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017. +[50] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. +[51] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021. +[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. + +[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. +[54] Shikun Liu, Linxi Fan, Edward Johns, Zhiding Yu, Chaowei Xiao, and Anima Anandkumar. Prismer: A vision-language model with an ensemble of experts. arXiv preprint arXiv:2303.02506, 2023. +[55] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. +[56] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. +[57] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019. +[58] Mathew Monfort, Bowen Pan, Kandan Ramakrishnan, Alex Andonian, Barry A McNamara, Alex Lascelles, Quanfu Fan, Dan Gutfreund, Rogério Schmidt Feris, and Aude Oliva. Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9434-9445, 2021. +[59] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE conference on computer vision and pattern recognition, pages 2408-2415. IEEE, 2012. +[60] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016. +[61] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. +[62] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. +[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. +[64] Kris Popat and Rosalind W. Picard. Novel cluster-based probability model for texture synthesis, classification, and compression. In Proc. SPIE Visual Comm. and Image Processing, 1993. + +[65] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. +[66] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. +[67] John C Raven. Standardization of progressive matrices, 1938. British Journal of Medical Psychology, 1941. +[68] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. +[69] Arno Schodl, Richard Szeliski, David H. Salesin, and Irfan Essa. Video textures. In Proceedings of SIGGRAPH '00, pages 489-498, 2000. +[70] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. +[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. +[72] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. +[73] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. +[74] Claude E. Shannon. A mathematical theory of communication. Bell Sys. Tech. Journal, 27, 1948. +[75] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951. +[76] Gunnar A Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. +[77] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018. +[78] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. + +[79] X. Soria, E. Riba, and A. Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1912-1921, Los Alamitos, CA, USA, 2020. IEEE Computer Society. +[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[81] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, koray kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. +[82] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, New York, New York, USA, 2016. PMLR. +[83] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. +[84] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. +[85] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th International Conference on Machine Learning, page 1096-1103, 2008. +[86] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. +[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. +[88] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of laion-2b, 2023. +[89] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. +[90] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2014. +[91] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. +[92] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. + +In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. +[93] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. +[94] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In International Conference on Learning Representations, 2021. +[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022. +[96] Yingchen Yu, Fangneng Zhan, Rongliang Wu, Jianxiong Pan, Kaiwen Cui, Shijian Lu, Feiying Ma, Xuansong Xie, and Chunyan Miao. Diverse image inpainting with bidirectional and autoregressive transformers. In Proceedings of the 29th ACM International Conference on Multimedia, pages 69-78, 2021. +[97] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. +[98] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In CVPR, 2018. +[99] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. +[100] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. +[101] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. \ No newline at end of file diff --git a/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/images.zip b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d50c7bcba05b7f194f80d2848ea78a908b1f0a8f --- /dev/null +++ b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0340113699820bba44127cc59bc213ca4c7dc90632e0d73a7b039ccabf043bd +size 1034413 diff --git a/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/layout.json b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dcc52a8c87d5bd3bc090ba62dc4dce05fcdffee4 --- /dev/null +++ b/2024/Sequential Modeling Enables Scalable Learning for Large Vision Models/layout.json @@ -0,0 +1,10823 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 521, + 121 + ], + "type": "text", + "content": "Sequential Modeling Enables Scalable Learning for Large Vision Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 179, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 142, + 179, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 179, + 157 + ], + "type": "text", + "content": "Yutong Bai" + }, + { + "bbox": [ + 106, + 142, + 179, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 157, + 185, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 157, + 185, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 157, + 185, + 171 + ], + "type": "text", + "content": "Alan L. Yuille" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 201, + 143, + 284, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 143, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 201, + 143, + 284, + 157 + ], + "type": "text", + "content": "Xinyang Geng" + }, + { + "bbox": [ + 201, + 143, + 284, + 157 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 205, + 157, + 282, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 157, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 205, + 157, + 282, + 171 + ], + "type": "text", + "content": "Trevor Darrell" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 143, + 412, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 143, + 412, + 157 + ], + "spans": [ + { + "bbox": [ + 302, + 143, + 412, + 157 + ], + "type": "text", + "content": "Karttikeya Mangalam" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 157, + 380, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 157, + 380, + 171 + ], + "spans": [ + { + "bbox": [ + 302, + 157, + 380, + 171 + ], + "type": "text", + "content": "Jitendra Malik1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 434, + 143, + 485, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 434, + 143, + 485, + 156 + ], + "spans": [ + { + "bbox": [ + 434, + 143, + 485, + 156 + ], + "type": "text", + "content": "Amir Bar" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 402, + 157, + 482, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 157, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 402, + 157, + 482, + 171 + ], + "type": "text", + "content": "Alexei A. Efros" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 184, + 260, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 184, + 260, + 199 + ], + "spans": [ + { + "bbox": [ + 151, + 184, + 260, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 184, + 260, + 199 + ], + "type": "text", + "content": "UC Berkeley (BAIR)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 184, + 440, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 440, + 200 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 440, + 200 + ], + "type": "text", + "content": "2Johns Hopkins University" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 252, + 290, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 252, + 290, + 420 + ], + "spans": [ + { + "bbox": [ + 45, + 252, + 290, + 420 + ], + "type": "text", + "content": "We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, \"visual sentences\", in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions without needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion tokens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 444, + 128, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 128, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 128, + 456 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 45, + 464, + 288, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 464, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 45, + 464, + 288, + 583 + ], + "type": "text", + "content": "Large language models (LLMs) such as GPT [10] and LLaMA [80] have taken the world by storm. What would it take to build a Large Vision Model (LVM)? From the animal world, we know that visual competences are not dependent on language. In particular, many experiments have shown that the visual world of non-human primates is remarkably similar to that of humans. So while the space of vision-language models such as LLaVA [53] is interesting and worthwhile to pursue, in this paper we seek an answer to a different question – how far can we go from pixels alone?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 585, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 645 + ], + "type": "text", + "content": "The key features of contemporary LLMs that we seek to emulate in LVMs are: 1) scaling in the presence of big data, and 2) flexible specification of tasks through prompting (in-context learning). How do we achieve this? As usual, there are three main components that must be specified:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "content": "Data: We want to exploit all the remarkable diversity in visual data. First of all, just raw unannotated images and videos. Next, we want to exploit the variety of annotated visual data sources that have been produced over the last" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 228, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 300 + ], + "type": "text", + "content": "couple of decades – semantic segmentations, depth reconstructions, keypoints, multiple views of 3D objects, among others. We define a common format, “visual sentences”, in which to represent these different annotations without needing any meta-knowledge beyond the pixels. The total size of our training dataset is 1.64 billion images/frames." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 300, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 348 + ], + "type": "text", + "content": "Architecture: We use a large transformer architecture (3 billion parameters) trained on visual data represented as sequence of tokens, using a learned tokenizer that maps each image to a string of 256 vector-quantized tokens." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 349, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 547, + 421 + ], + "type": "text", + "content": "Loss function: We draw inspiration from the natural language community, where masked token modeling has given way to sequential autoregressive prediction. Once images/videos/annotated images can all be represented as sequences, we can train the model to minimize the cross-entropy loss for predicting the next token." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 422, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 445 + ], + "type": "text", + "content": "With this extremely simple design, we demonstrate some noteworthy behaviors:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 446, + 547, + 590 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 306, + 446, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 545, + 469 + ], + "type": "text", + "content": "- Appropriate scaling behavior as one increases model size and data size." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 470, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 547, + 529 + ], + "type": "text", + "content": "- Many different vision tasks can now be \"solved\" by designing suitable prompts at test time. While the results don't show as high performance as bespoke, specifically-trained models, the fact that so many tasks are all addressed by a single vision model is quite encouraging." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 529, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 545, + 553 + ], + "type": "text", + "content": "- We see a clear benefit of the amount of unsupervised data on the performance on various standard vision tasks." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 554, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 547, + 590 + ], + "type": "text", + "content": "- We see a hint of an ability for general visual reasoning – handling out-of-distribution data, and performing novel tasks. But further investigation is needed." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 600, + 392, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 600, + 392, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 600, + 392, + 612 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Pretrained Vision Models. The value of using pretrained models (such as ImageNet-pretrained AlexNet [45]) has been demonstrated as far back as 2015 in R-CNN [34], and it has since become standard practice in computer vision. Self-supervised pretraining was proposed as a way to vastly increase the amount of data available for pretraining [16, 25, 37, 61, 62, 99]. Unfortunately, this was not very successful, likely because the CNN-based architectures" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 268, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 268, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 268, + 713 + ], + "type": "text", + "content": "*Equal Contribution. Webpage: https://yutongbai.com/lvm.html." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22861" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "of that time did not have enough capacity to absorb the data. With the introduction of Transformers [84], which have much higher capacity, researchers revisited self-supervised pretraining, and showed that transformer-based masked image reconstruction approaches, such as BEiT [7], MAE [38], SimMIM [91], MaskGIT [13] perform vastly better than their CNN-based counterparts [62]. Yet, despite their recent successes, current pretrained vision-only models have had trouble scaling up to the really large datasets, such as LAION [72]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 364 + ], + "type": "text", + "content": "Multi-task Learning and In-context Learning. From the classic one-model-per-task setups, computer vision is slowly moving toward a single model performing multiple different tasks. Various multi-task learning approaches [24, 40, 43, 73, 97] exist but they are typically limited to a fixed, predefined number of tasks. More recently, methods inspired by in-context learning in LLMs forgo any notion of tasks and instead let the model infer the task directly from the input prompt. For example, Visual Prompting [8, 87] takes in a task input/output example pair and a query image at test time, concatenates them into a single 2-by-2 image, and uses inpainting to generate the desired output. But, since the inpainting is performed using a variant of MAE [38], the same problems with scaling are inherited by these approaches." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 366, + 289, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 535 + ], + "type": "text", + "content": "Auto-regressive Visual Models. The idea of using autoregressive models for synthesizing visual data goes back at least 70 years. Inspired by Shannon's use of " + }, + { + "bbox": [ + 46, + 366, + 289, + 535 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 366, + 289, + 535 + ], + "type": "text", + "content": "-grams to synthesize language [74, 75], a number of works, starting with Attneave's seminal 1954 paper [5], applied this idea to sequentially synthesizing pixels [28, 31, 39, 64], image patches [27], video frames [69], and motion capture data [4, 44, 48]. As deep models became popular, newer works replaced " + }, + { + "bbox": [ + 46, + 366, + 289, + 535 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 366, + 289, + 535 + ], + "type": "text", + "content": "-grams with RNNs or CNNs for pixel [81, 82] and video [66] synthesis. Most recently, transformer-based autoregressive visual generation methods have been proposed [15, 29, 94, 96], and, combined with language, have demonstrated impressive image synthesis results, e.g. Parti [95]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 543, + 87, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 87, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 87, + 555 + ], + "type": "text", + "content": "3. Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 565, + 288, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 565, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 86, + 565, + 288, + 576 + ], + "type": "text", + "content": "\"Data! Data! Data! I can't make bricks without clay!\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 584, + 286, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 584, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 206, + 584, + 286, + 594 + ], + "type": "text", + "content": "SHERLOCK HOLMES" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "type": "text", + "content": "The key requirement of any Large Pre-trained Model is that it must be trained on vast amounts of data. For language models, very large and very diverse datasets are fairly easy to obtain. For instance, the popular Common Crawl repository [1] contains 250 billion web pages spanning the entire Web, is extremely diverse, and includes \"natural demonstrations\" like language translations, question answering, etc. In computer vision, we are still very far from having a data source of comparable size and diversity. One of the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "text", + "content": "central contributions of our work is the first step toward curating such a dataset that we call Unified Vision Dataset v1 (UVDv1). To assemble it, we leverage many different sources of visual data: (1) unlabelled images, (2) images with visual annotations, (3) unlabelled videos, (4) videos with visual annotations, and (5) 3D synthetic objects. The unlabeled images, which represent over " + }, + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "text", + "content": " of our data, capture a huge cross-section of our visual world, and provide the required diversity, at the cost of lower quality. Images with annotations have a much more constrained distribution, but are usually of higher quality. Video data is even more constrained (typically, to human-centric activities), but is an invaluable source of temporal data. Renderings of 3D synthetic objects are the lowest in diversity but can provide valuable hints about the behavior of 3D structures. Importantly, UVDv1 is a purely visual dataset, with no non-visual meta-data (e.g. text) included. All together, UVDv1 contains 1.64 billion images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 288, + 548, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 548, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 548, + 419 + ], + "type": "text", + "content": "Another important difference from Large Language Models is that language data has a natural, unified one-dimensional structure for all the data - a stream of text. Unfortunately, this is not the case for visual data, with different sources all having different structures. In this work we propose visual sentence as a unified unit of visual data, which enables us to train scalable models from a diverse set of sources. A visual sentence is simply a sequence containing one or more images followed by an end-of-sentence (EOS) token. Figure 1 shows how the various data sources are partitioned into visual sentences. In particular:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "text", + "content": "Single images. A single image itself represents the simplest form of a visual sentence - " + }, + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\{" + }, + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "text", + "content": " image, EOS" + }, + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\}" + }, + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "text", + "content": ". We use the filtered subset of 1.49 billion images [88] from the LAION 5B [71] dataset. This is by far the largest part of our data, comprising " + }, + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "inline_equation", + "content": "88.5\\%" + }, + { + "bbox": [ + 304, + 423, + 547, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 548, + 713 + ], + "type": "text", + "content": "Image sequences. A sequence of images is a natural form of visual sentence. We create such sequences by sourcing video data from a wide range of existing datasets [11, 12, 21, 35, 36, 46, 50, 51, 55, 57-59, 63, 68, 76-78, 92, 93]. Visual sentences of 16 frames are formed by randomly sampling the videos at three different strides (10, 20, and 30). In addition, we utilize synthetic 3D objects from the Objaverse Dataset [22] to generate object-centric multiview sequences for a variety of objects. For each object, we sample one radius length between the object center and the camera from 1.5 to 2.2, and sample one constant elevation from -45 degrees to 45 degrees, then traverse different views of the object by changing the azimuth with a step length of 15 degrees and render 24 views. We rendered 42000 such sequences in total for training and 8000 for testing. Finally, we can also represent images belonging to the same semantic category as being (part of) a sequence. We use categories from ImageNet, concatenating together groups of images (2,4,8, or 16) from the same category into a 16-image long" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22862" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 86, + 81, + 199, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 81, + 199, + 95 + ], + "spans": [ + { + "bbox": [ + 86, + 81, + 199, + 95 + ], + "type": "text", + "content": "Visual Sentences" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 105, + 75, + 130 + ], + "blocks": [ + { + "bbox": [ + 49, + 105, + 75, + 130 + ], + "lines": [ + { + "bbox": [ + 49, + 105, + 75, + 130 + ], + "spans": [ + { + "bbox": [ + 49, + 105, + 75, + 130 + ], + "type": "image", + "image_path": "d35e2bd48e9c47f6b74a859f3b83dd27df6d33f2b69a9072df35f789704ac45b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 121, + 93, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 121, + 93, + 128 + ], + "spans": [ + { + "bbox": [ + 78, + 121, + 93, + 128 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 114, + 211, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 114, + 211, + 124 + ], + "spans": [ + { + "bbox": [ + 110, + 114, + 211, + 124 + ], + "type": "text", + "content": "Single images, e.g. LAION" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 133, + 238, + 158 + ], + "blocks": [ + { + "bbox": [ + 50, + 133, + 238, + 158 + ], + "lines": [ + { + "bbox": [ + 50, + 133, + 238, + 158 + ], + "spans": [ + { + "bbox": [ + 50, + 133, + 238, + 158 + ], + "type": "image", + "image_path": "12d009d3676495b166dbbe08977b4b1e87c87b9f7b442350b391e4ec75091050.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 150, + 253, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 150, + 253, + 158 + ], + "spans": [ + { + "bbox": [ + 239, + 150, + 253, + 158 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 162, + 237, + 187 + ], + "blocks": [ + { + "bbox": [ + 50, + 162, + 237, + 187 + ], + "lines": [ + { + "bbox": [ + 50, + 162, + 237, + 187 + ], + "spans": [ + { + "bbox": [ + 50, + 162, + 237, + 187 + ], + "type": "image", + "image_path": "05f3f0a5ab5cd60e92de770e16445b0dd90ff843edb65f124df856ab180ca62e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 239, + 161, + 253, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 161, + 253, + 171 + ], + "spans": [ + { + "bbox": [ + 239, + 161, + 253, + 171 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 191, + 237, + 215 + ], + "blocks": [ + { + "bbox": [ + 50, + 191, + 237, + 215 + ], + "lines": [ + { + "bbox": [ + 50, + 191, + 237, + 215 + ], + "spans": [ + { + "bbox": [ + 50, + 191, + 237, + 215 + ], + "type": "image", + "image_path": "0f3c7a1217f7766d48c9ed6a38f238311059d4a1871e62ee48200e5d7bbaae5c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 191, + 253, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 191, + 253, + 200 + ], + "spans": [ + { + "bbox": [ + 239, + 191, + 253, + 200 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 239, + 208, + 253, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 208, + 253, + 213 + ], + "spans": [ + { + "bbox": [ + 239, + 208, + 253, + 213 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 279, + 150, + 356, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 150, + 356, + 194 + ], + "spans": [ + { + "bbox": [ + 279, + 150, + 356, + 194 + ], + "type": "text", + "content": "Image sequences, e.g. videos, 3D rotations, synthetic viewpoints" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 218, + 97, + 243 + ], + "blocks": [ + { + "bbox": [ + 50, + 218, + 97, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 218, + 97, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 218, + 97, + 243 + ], + "type": "image", + "image_path": "1d222d66e84a2f4a4c2b7ee1e6eabb187cdafc56075e3b39d892f11209ffec40.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 104, + 218, + 152, + 243 + ], + "blocks": [ + { + "bbox": [ + 104, + 218, + 152, + 243 + ], + "lines": [ + { + "bbox": [ + 104, + 218, + 152, + 243 + ], + "spans": [ + { + "bbox": [ + 104, + 218, + 152, + 243 + ], + "type": "image", + "image_path": "4e992a3888d53e2b91b820826178570d10fcd3347accfd212ca9abcd441307a4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 158, + 218, + 205, + 243 + ], + "blocks": [ + { + "bbox": [ + 158, + 218, + 205, + 243 + ], + "lines": [ + { + "bbox": [ + 158, + 218, + 205, + 243 + ], + "spans": [ + { + "bbox": [ + 158, + 218, + 205, + 243 + ], + "type": "image", + "image_path": "ca73993924913fd58018f80182d7af16263a9549ab3f3023ec44486a49ae395a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 212, + 218, + 260, + 243 + ], + "blocks": [ + { + "bbox": [ + 212, + 218, + 260, + 243 + ], + "lines": [ + { + "bbox": [ + 212, + 218, + 260, + 243 + ], + "spans": [ + { + "bbox": [ + 212, + 218, + 260, + 243 + ], + "type": "image", + "image_path": "0fd55ba213dd6da12d30d9a8d819b347f3b159d14d2fda33de8cdfbd9721d76e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 262, + 233, + 276, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 233, + 276, + 243 + ], + "spans": [ + { + "bbox": [ + 262, + 233, + 276, + 243 + ], + "type": "text", + "content": "··" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 50, + 247, + 97, + 273 + ], + "blocks": [ + { + "bbox": [ + 50, + 247, + 97, + 273 + ], + "lines": [ + { + "bbox": [ + 50, + 247, + 97, + 273 + ], + "spans": [ + { + "bbox": [ + 50, + 247, + 97, + 273 + ], + "type": "image", + "image_path": "255a7cfdc4aa8e6b5e20456b7b1cfa02fa00f8723cebe2a23cb6da69e51a5cba.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 104, + 247, + 152, + 273 + ], + "blocks": [ + { + "bbox": [ + 104, + 247, + 152, + 273 + ], + "lines": [ + { + "bbox": [ + 104, + 247, + 152, + 273 + ], + "spans": [ + { + "bbox": [ + 104, + 247, + 152, + 273 + ], + "type": "image", + "image_path": "dd8334a170b70fcb859b027fd60a94f35a8664fe7552b4741c286bdb84ea071e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 158, + 247, + 206, + 273 + ], + "blocks": [ + { + "bbox": [ + 158, + 247, + 206, + 273 + ], + "lines": [ + { + "bbox": [ + 158, + 247, + 206, + 273 + ], + "spans": [ + { + "bbox": [ + 158, + 247, + 206, + 273 + ], + "type": "image", + "image_path": "331223a6c6d61b18c8944adeab785e30c0b8f752f8dea5a60d49242c860ece55.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 212, + 247, + 260, + 273 + ], + "blocks": [ + { + "bbox": [ + 212, + 247, + 260, + 273 + ], + "lines": [ + { + "bbox": [ + 212, + 247, + 260, + 273 + ], + "spans": [ + { + "bbox": [ + 212, + 247, + 260, + 273 + ], + "type": "image", + "image_path": "19e215c306c745e4bf3dc6222338abf3c8daf8c5055c01aec6e19785427b50f2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 262, + 247, + 276, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 247, + 276, + 255 + ], + "spans": [ + { + "bbox": [ + 262, + 247, + 276, + 255 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 50, + 277, + 97, + 301 + ], + "blocks": [ + { + "bbox": [ + 50, + 277, + 97, + 301 + ], + "lines": [ + { + "bbox": [ + 50, + 277, + 97, + 301 + ], + "spans": [ + { + "bbox": [ + 50, + 277, + 97, + 301 + ], + "type": "image", + "image_path": "ebff1d7b9c084e04ddfa542085c32a0fdaf45576c3044fd9c83a087af92163f9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 104, + 277, + 152, + 301 + ], + "blocks": [ + { + "bbox": [ + 104, + 277, + 152, + 301 + ], + "lines": [ + { + "bbox": [ + 104, + 277, + 152, + 301 + ], + "spans": [ + { + "bbox": [ + 104, + 277, + 152, + 301 + ], + "type": "image", + "image_path": "e1c32ea914f99cb0e4497011c685eef33b60aaf9d9634450129a8f6818dd8bfe.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 158, + 277, + 206, + 301 + ], + "blocks": [ + { + "bbox": [ + 158, + 277, + 206, + 301 + ], + "lines": [ + { + "bbox": [ + 158, + 277, + 206, + 301 + ], + "spans": [ + { + "bbox": [ + 158, + 277, + 206, + 301 + ], + "type": "image", + "image_path": "b16445fccde8375bf0353909a4fe5598943586ce1d81606ba02528237df392dc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 212, + 277, + 260, + 301 + ], + "blocks": [ + { + "bbox": [ + 212, + 277, + 260, + 301 + ], + "lines": [ + { + "bbox": [ + 212, + 277, + 260, + 301 + ], + "spans": [ + { + "bbox": [ + 212, + 277, + 260, + 301 + ], + "type": "image", + "image_path": "b92ce6c2efa5252166ca3c69fab5ad501908254c7fd36d8f0300aa134c5c982a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 262, + 295, + 276, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 295, + 276, + 301 + ], + "spans": [ + { + "bbox": [ + 262, + 295, + 276, + 301 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 50, + 306, + 97, + 331 + ], + "blocks": [ + { + "bbox": [ + 50, + 306, + 97, + 331 + ], + "lines": [ + { + "bbox": [ + 50, + 306, + 97, + 331 + ], + "spans": [ + { + "bbox": [ + 50, + 306, + 97, + 331 + ], + "type": "image", + "image_path": "b7369a4bf92f2b4925fd4e339ac810b52f8b79b9bb9d38e815573a5af0970412.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 104, + 306, + 152, + 331 + ], + "blocks": [ + { + "bbox": [ + 104, + 306, + 152, + 331 + ], + "lines": [ + { + "bbox": [ + 104, + 306, + 152, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 306, + 152, + 331 + ], + "type": "image", + "image_path": "ca16e771aa9035b03660710a61f21e0b5d2e326121d2b2fe4adc643741748ebd.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 158, + 306, + 206, + 331 + ], + "blocks": [ + { + "bbox": [ + 158, + 306, + 206, + 331 + ], + "lines": [ + { + "bbox": [ + 158, + 306, + 206, + 331 + ], + "spans": [ + { + "bbox": [ + 158, + 306, + 206, + 331 + ], + "type": "image", + "image_path": "1907b449fc7e84c31b86d510621f04d7a10eceb1af56844e3f91317e275becb9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 212, + 306, + 260, + 331 + ], + "blocks": [ + { + "bbox": [ + 212, + 306, + 260, + 331 + ], + "lines": [ + { + "bbox": [ + 212, + 306, + 260, + 331 + ], + "spans": [ + { + "bbox": [ + 212, + 306, + 260, + 331 + ], + "type": "image", + "image_path": "f5002f47cd7bd159b9b70ab0c1d991b928c674f1b3d1f1a21d1f4d35912b4594.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 262, + 306, + 276, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 306, + 276, + 323 + ], + "spans": [ + { + "bbox": [ + 262, + 306, + 276, + 323 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 50, + 335, + 121, + 360 + ], + "blocks": [ + { + "bbox": [ + 50, + 335, + 121, + 360 + ], + "lines": [ + { + "bbox": [ + 50, + 335, + 121, + 360 + ], + "spans": [ + { + "bbox": [ + 50, + 335, + 121, + 360 + ], + "type": "image", + "image_path": "5b76b3bdf4750d333bf6fa8f8c457448dfc22a49025114f1fa0e193d516651c2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 127, + 336, + 199, + 360 + ], + "blocks": [ + { + "bbox": [ + 127, + 336, + 199, + 360 + ], + "lines": [ + { + "bbox": [ + 127, + 336, + 199, + 360 + ], + "spans": [ + { + "bbox": [ + 127, + 336, + 199, + 360 + ], + "type": "image", + "image_path": "1f4c16f46e6a1cecc9220bf9d9301c8ea722163837abae0c25d2aa9112e22591.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 202, + 351, + 216, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 351, + 216, + 357 + ], + "spans": [ + { + "bbox": [ + 202, + 351, + 216, + 357 + ], + "type": "text", + "content": "··" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 50, + 364, + 97, + 388 + ], + "blocks": [ + { + "bbox": [ + 50, + 364, + 97, + 388 + ], + "lines": [ + { + "bbox": [ + 50, + 364, + 97, + 388 + ], + "spans": [ + { + "bbox": [ + 50, + 364, + 97, + 388 + ], + "type": "image", + "image_path": "74af6dd9eecf0abfa318bb078ac44187b6409168bd6e95e0fe6e730183a10b3b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 104, + 364, + 152, + 388 + ], + "blocks": [ + { + "bbox": [ + 104, + 364, + 152, + 388 + ], + "lines": [ + { + "bbox": [ + 104, + 364, + 152, + 388 + ], + "spans": [ + { + "bbox": [ + 104, + 364, + 152, + 388 + ], + "type": "image", + "image_path": "99a314ff597a49d586745816c389479a64fbfeb9c08b4a72e4d604cdeaf05d92.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 158, + 365, + 206, + 388 + ], + "blocks": [ + { + "bbox": [ + 158, + 365, + 206, + 388 + ], + "lines": [ + { + "bbox": [ + 158, + 365, + 206, + 388 + ], + "spans": [ + { + "bbox": [ + 158, + 365, + 206, + 388 + ], + "type": "image", + "image_path": "367b5a32941f61bcd90e211a5554233c57d110ae4269682bd8506f3b635e1f19.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 212, + 365, + 260, + 388 + ], + "blocks": [ + { + "bbox": [ + 212, + 365, + 260, + 388 + ], + "lines": [ + { + "bbox": [ + 212, + 365, + 260, + 388 + ], + "spans": [ + { + "bbox": [ + 212, + 365, + 260, + 388 + ], + "type": "image", + "image_path": "dbd571c50da373321b7676eaa1d62eea07e7aa01458f1498d5d41cf1d0661153.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 262, + 365, + 276, + 388 + ], + "blocks": [ + { + "bbox": [ + 262, + 365, + 276, + 388 + ], + "lines": [ + { + "bbox": [ + 262, + 365, + 276, + 388 + ], + "spans": [ + { + "bbox": [ + 262, + 365, + 276, + 388 + ], + "type": "image", + "image_path": "720368fcfc2df52da3021da686678a1727eae2a00aed10853f0001553acd417f.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 279, + 365, + 356, + 388 + ], + "blocks": [ + { + "bbox": [ + 279, + 365, + 356, + 388 + ], + "lines": [ + { + "bbox": [ + 279, + 365, + 356, + 388 + ], + "spans": [ + { + "bbox": [ + 279, + 365, + 356, + 388 + ], + "type": "image", + "image_path": "0a27f1522d2b55bd9909c3a55fe25ec8fbf11936e417a9198239d62139a1fb0e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 357, + 336, + 430, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 336, + 430, + 381 + ], + "spans": [ + { + "bbox": [ + 357, + 336, + 430, + 381 + ], + "type": "text", + "content": "Edges with free form notation, 1. object detection + frame segmentation" + } + ] + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 50, + 393, + 121, + 417 + ], + "blocks": [ + { + "bbox": [ + 50, + 393, + 121, + 417 + ], + "lines": [ + { + "bbox": [ + 50, + 393, + 121, + 417 + ], + "spans": [ + { + "bbox": [ + 50, + 393, + 121, + 417 + ], + "type": "image", + "image_path": "64e2176d1c23a956a0faaa839fa2b19b57db455b8435cb260d37c462d5c1ca4e.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 123, + 393, + 193, + 418 + ], + "blocks": [ + { + "bbox": [ + 123, + 393, + 193, + 418 + ], + "lines": [ + { + "bbox": [ + 123, + 393, + 193, + 418 + ], + "spans": [ + { + "bbox": [ + 123, + 393, + 193, + 418 + ], + "type": "image", + "image_path": "21449481cfe6c5fc8773982e32160a2b113987b694c9fa9b7328a8f5fd2ce00c.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 199, + 393, + 260, + 418 + ], + "blocks": [ + { + "bbox": [ + 199, + 393, + 260, + 418 + ], + "lines": [ + { + "bbox": [ + 199, + 393, + 260, + 418 + ], + "spans": [ + { + "bbox": [ + 199, + 393, + 260, + 418 + ], + "type": "image", + "image_path": "a5531978c2355c6e0f79882390f17e686bc340009960ced7cdb1d28041394ffb.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 261, + 393, + 339, + 418 + ], + "blocks": [ + { + "bbox": [ + 261, + 393, + 339, + 418 + ], + "lines": [ + { + "bbox": [ + 261, + 393, + 339, + 418 + ], + "spans": [ + { + "bbox": [ + 261, + 393, + 339, + 418 + ], + "type": "image", + "image_path": "95eafdfc25b5f6a6f846f153dc77094fb8ea7d2e80f91e901928e661d4cc4a6a.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 342, + 422, + 431, + 446 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 422, + 431, + 446 + ], + "spans": [ + { + "bbox": [ + 342, + 422, + 431, + 446 + ], + "type": "image", + "image_path": "44168d6e6bcf64997725b9b940c56a5a05b60d0fa552b7fa7917909cd30cb329.jpg" + } + ] + } + ], + "index": 46, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 432, + 391, + 449, + 446 + ], + "blocks": [ + { + "bbox": [ + 432, + 391, + 449, + 446 + ], + "lines": [ + { + "bbox": [ + 432, + 391, + 449, + 446 + ], + "spans": [ + { + "bbox": [ + 432, + 391, + 449, + 446 + ], + "type": "image", + "image_path": "cfba108de1d02b91068e23bb8034bd8ef6c24066b78c1ba5fa2cf42685e518d8.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 449, + 233, + 544, + 367 + ], + "blocks": [ + { + "bbox": [ + 449, + 233, + 544, + 367 + ], + "lines": [ + { + "bbox": [ + 449, + 233, + 544, + 367 + ], + "spans": [ + { + "bbox": [ + 449, + 233, + 544, + 367 + ], + "type": "image", + "image_path": "9860739c00b2620b4053e324d17626f67e8873b110423d67fe6bfb88790310e5.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 50, + 422, + 144, + 446 + ], + "blocks": [ + { + "bbox": [ + 50, + 422, + 144, + 446 + ], + "lines": [ + { + "bbox": [ + 50, + 422, + 144, + 446 + ], + "spans": [ + { + "bbox": [ + 50, + 422, + 144, + 446 + ], + "type": "image", + "image_path": "0d3111027ee1360470b0d542ee8894e9f968d67b26f68644f5337013ca754107.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 450, + 496, + 461 + ], + "lines": [ + { + "bbox": [ + 94, + 450, + 496, + 461 + ], + "spans": [ + { + "bbox": [ + 94, + 450, + 496, + 461 + ], + "type": "text", + "content": "Figure 1. Visual sentences allow us to format diverse vision data into the unified structure of image sequences." + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 145, + 423, + 236, + 446 + ], + "blocks": [ + { + "bbox": [ + 145, + 423, + 236, + 446 + ], + "lines": [ + { + "bbox": [ + 145, + 423, + 236, + 446 + ], + "spans": [ + { + "bbox": [ + 145, + 423, + 236, + 446 + ], + "type": "image", + "image_path": "f38fa1a2215d7404b1d69f6fd93355a7b16fabd50a6a1f905b7090661c3e4982.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 242, + 422, + 339, + 446 + ], + "blocks": [ + { + "bbox": [ + 242, + 422, + 339, + 446 + ], + "lines": [ + { + "bbox": [ + 242, + 422, + 339, + 446 + ], + "spans": [ + { + "bbox": [ + 242, + 422, + 339, + 446 + ], + "type": "image", + "image_path": "22307d15ebf8d2b1b32c2d6c478fb8af744d90faf7d22338022237ef0e171f3b.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "bbox": [ + 342, + 422, + 431, + 446 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 422, + 431, + 446 + ], + "spans": [ + { + "bbox": [ + 342, + 422, + 431, + 446 + ], + "type": "image", + "image_path": "44168d6e6bcf64997725b9b940c56a5a05b60d0fa552b7fa7917909cd30cb329.jpg" + } + ] + } + ], + "index": 52, + "type": "text" + }, + { + "bbox": [ + 432, + 446, + 446, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 446, + 446, + 450 + ], + "spans": [ + { + "bbox": [ + 432, + 446, + 446, + 450 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 447, + 394, + 518, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 447, + 394, + 518, + 437 + ], + "spans": [ + { + "bbox": [ + 447, + 394, + 518, + 437 + ], + "type": "text", + "content": "Videos with annotation, e.g. video segmentation" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 47, + 466, + 116, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 116, + 476 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 116, + 476 + ], + "type": "text", + "content": "visual sentences." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "content": "Images with annotations. To handle different types of image annotations in a uniform way, we choose to represent all annotations as images. Some data types, e.g. semantic segmentation maps [100], edge maps [79], depth [65] and normal images [6], are already represented this way. For others we apply tailored methods for each specific annotation type: 1) Object Detection: We create annotations by overlaying a color-coded bounding box around each object, following the methodology in [14]; 2) Human Pose: Human skeletons are rendered in pixel space, adhering to the OpenPose format, utilizing MMPose [19]; 3) Depth Estimation, Surface Normal, and Edge Detection: given ImageNet and COCO images, we generate annotations in line with the protocols from [54]. 3) Style Transfer [9], De-rain [98], De-noise [85], Low Light Enhancement [89], and Stereo Datasets [33]: These are all represented as image pairs (e.g. input/output). 4) Colorization: We convert ImageNet images to greyscale, producing image pairs. 5) Inpainting: The process involves randomly adding black-colored boxes" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "text", + "content": "in images to simulate corruption, resulting in image pairs. For all the above annotation types, we can create visual sentences by concatenating 8 image pairs of the same annotation type into a 16-image visual sentence. For datasets containing " + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "text", + "content": " different annotations for the same image we use a different approach: for each set of " + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "inline_equation", + "content": "1 + k" + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "text", + "content": " images (input plus " + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "text", + "content": " annotations), we randomly select " + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "text", + "content": " elements, where " + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "inline_equation", + "content": "m \\leq k + 1 \\leq 16" + }, + { + "bbox": [ + 304, + 466, + 547, + 573 + ], + "type": "text", + "content": ". These m-tuples are then concatenated to form visual sequences." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 579, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 687 + ], + "type": "text", + "content": "Image sequences with annotations. When converting annotated video data (VIPSeg [56], Hand14K [30], AVA [59], JHMDB [41]) to visual sentences, we apply two complementary strategies. The first is similar to how we treat image data with paired annotations: each visual sentence is constructed by concatenating frames with their annotations - {frame1,annot1,frame2,annot2,...}. The second method involves grouping multiple frames followed by their corresponding annotations - {frame1,frame2,annot1,annot2,...}." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "We present a detailed summary of all the data sources, annotation type and data statistics of UVDv1 in the Appendix." + } + ] + } + ], + "index": 60 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22863" + } + ] + } + ], + "index": 61 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 115, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 115, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 115, + 85 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "text", + "content": "In this section, we describe the design of our autoregressive Large Vision Model. Unlike text data, which naturally exhibits discrete sequential structure, it is not straightforward to model image pixels in visual sentences. In this work, we take a two-stage approach: 1) train a large visual tokenizer (which operates on individual images) to convert each image into a sequence of visual tokens; 2) train an autoregressive transformer model on visual sentences, each represented as a sequence of tokens. We summarize our approach in Figure 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 205, + 162, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 162, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 162, + 217 + ], + "type": "text", + "content": "4.1. Image Tokenization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 223, + 288, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 288, + 389 + ], + "type": "text", + "content": "While the visual sentences exhibit a sequence structure between consecutive images, we don't have such natural sequence structure within an image. Therefore, in order to apply a transformer model to images, prior works typically do one of the following: either divide the image into patches in scan-line order, and treat that as a sequence [26], or use a pre-trained image tokenizer, such as VQVAE [83] or VQGAN [29], to cluster image features into a grid of discrete tokens, which, again, are turned into a sequence in scan-line order. We adopt the latter approach since the discrete categorical output from a model naturally forms a probabilistic distribution that one can easily sample from, enabling flexible conditional generation of new images within a visual sentence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 391, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 533 + ], + "type": "text", + "content": "Specifically, we employ semantic tokens generated by a VQGAN model, a concept introduced by Esser et al [29]. This framework consists of an encoding and a decoding mechanism, featuring a quantization layer that assigns input images to a sequence of discrete tokens from an established codebook. Our encoders and decoders are constructed purely with convolutional layers. The encoder is equipped with several downsampling modules to contract the spatial dimension of the input, whereas the decoder is fitted with an equivalent series of upsampling modules to restore the image to its initial size. For a given image, our VQGAN tokenizer produces 256 discrete tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 617 + ], + "type": "text", + "content": "It is important to note that our tokenizer operates on individual images independently, rather than on the entire visual sentence at once. This independence allows us to decouple the tokenizer training from the downstream Transformer model so that the tokenizer can be trained on a dataset of single images without having to consider the distribution of visual sentences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Implementation Details: We adopt an off-the-shelf VQGAN architecture from Chang et al. [13]. We follow the exact configuration in Chang et al. [13], which uses a down-sampling factor of " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f = 16" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and codebook size 8192. This means that for an image of size " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", our VQGAN tokenizer produces " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "16 \\times 16 = 256" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " tokens where each can take 8192 different values. We found that using the results of an ImageNet pre-trained tokenizer did not generalize well" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 329, + 73, + 521, + 255 + ], + "blocks": [ + { + "bbox": [ + 329, + 73, + 521, + 255 + ], + "lines": [ + { + "bbox": [ + 329, + 73, + 521, + 255 + ], + "spans": [ + { + "bbox": [ + 329, + 73, + 521, + 255 + ], + "type": "image", + "image_path": "c645823cc553d995a033025c581a2ab17f86f35e24183231853b9e86ec5e74e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 258, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 324 + ], + "type": "text", + "content": "Figure 2. Architecture of LVM. We first convert individual images from a visual sentence into discrete tokens using a VQGAN encoder. The resulting tokens from all images are then concatenated into a 1D sequence, and fed into an autoregressive Transformer model to predict the next token in the sequence. The predicted visual tokens are decoded into images using the VQGAN decoder." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 328, + 547, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 353 + ], + "type": "text", + "content": "beyond ImageNet images. Therefore, we train our own tokenizer on a 1.5B subset of the LAION 5B dataset [71]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 360, + 512, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 512, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 512, + 373 + ], + "type": "text", + "content": "4.2. Sequence Modeling of Visual Sentences" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 533 + ], + "type": "text", + "content": "After converting images into discrete tokens with VQGAN, we treat our visual sentence as a unified sequence by concatenating the discrete tokens from multiple images into a 1D sequence. Importantly, all visual sentences are treated equally – we do not make use of any special tokens to indicate particular tasks or formats. We train a causal Transformer model with the next token prediction objective using a cross-entropy loss, similar to the standard approach for language models [10]. Training the model the same way on all visual sentences enables the model to infer the relation between images from context instead of from task- or format-specific tokens. This gives the model an opportunity to generalize to other, unseen visual sentence structures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "content": "Implementation Details: After tokenizing each image in a visual sentence into 256 tokens, we concatenate them to form a 1D sequence of tokens. On top of the sequences of visual tokens, our Transformer model is virtually the same as an autoregressive language model, so we adopt the Transformer architecture of LLaMA [80], a popular open-source language model with widely available implementations. We use a context length of 4096 tokens, which can fit 16 images under our VQGAN tokenizer. Similar to language models, we add a [BOS] (begin of sentence) token to the beginning of each visual sentence and an [EOS] (end of sentence) token to the end, and use sequence concatenation [18] during training time to improve efficiency. We train our model on our entire UVDv1 dataset (420 billion tokens) using one epoch (simple epoch training is standard in language models to avoid poten" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22864" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 286, + 257 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 286, + 257 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 286, + 257 + ], + "type": "image", + "image_path": "ddb3d3ccef716ea90bf68423385be11ea456d862ac1dfae940f8f1562d1cd0c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 288, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 326 + ], + "type": "text", + "content": "Figure 3. Training loss for the 300M, 600M, 1B, and 3B models. All models are trained on 420B tokens (1.64B images). The training scales well with model sizes. Zoom in for more details: notice how the shape of the 4 curves is very similar. This is the consequence of the deterministic training pipeline, so the loss fluctuations are determined by the intrinsic difficulty of the data stream." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 289, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 377 + ], + "type": "text", + "content": "tial overfitting). We train 4 models with different numbers of parameters: 300 million, 600 million, 1 billion and 3 billion, following the same training configurations. We provide the detailed training hyperparameters in Appendix 5.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 382, + 214, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 214, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 214, + 396 + ], + "type": "text", + "content": "4.3. Inference by Visual Prompting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 400, + 289, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 509 + ], + "type": "text", + "content": "Since the autoregressive Transformer in our model outputs a probability distribution of the next token conditioned on previous tokens, we can easily sample from this distribution to generate new visual tokens that complete a visual sentence. To use the model for downstream tasks, one can construct a partial visual sentence that defines a task at test time, and apply the model to generate the output. This is similar to in-context learning in language models [10] or visual prompting in computer vision [8, 39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 513, + 242, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 242, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 242, + 528 + ], + "type": "text", + "content": "5. Experimental Results and Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "type": "text", + "content": "In this section, we evaluate the scaling abilities of our trained model, as well as its ability to understand and answer a range of diverse prompted tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 119, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 119, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 119, + 588 + ], + "type": "text", + "content": "5.1. Scalability" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 642 + ], + "type": "text", + "content": "We investigate the scaling behavior of our model in terms of the training loss and downstream task performance as we increase the model size as well as the number of tokens seen during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Training loss. We first inspect the training loss of LVM with different parameter sizes, which we present in Figure 3. Since all our models are trained for only one epoch on the dataset, the model sees a given data sample just once, and therefore the training loss at any point during training is very similar to the validation loss. One can observe that as" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 71, + 425, + 170 + ], + "blocks": [ + { + "bbox": [ + 325, + 71, + 425, + 170 + ], + "lines": [ + { + "bbox": [ + 325, + 71, + 425, + 170 + ], + "spans": [ + { + "bbox": [ + 325, + 71, + 425, + 170 + ], + "type": "image", + "image_path": "37117ed5e9cc178a1818577203dd9a4135c7b467e2fa5215636e7d1a628717a0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 527, + 170 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 527, + 170 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 527, + 170 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 527, + 170 + ], + "type": "image", + "image_path": "9f27582b2f581d86014004d20a0c3abb8f799a3556435c44f9f1389173e14116.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 324, + 171, + 423, + 270 + ], + "blocks": [ + { + "bbox": [ + 324, + 171, + 423, + 270 + ], + "lines": [ + { + "bbox": [ + 324, + 171, + 423, + 270 + ], + "spans": [ + { + "bbox": [ + 324, + 171, + 423, + 270 + ], + "type": "image", + "image_path": "848fc8927e203ecfcaf895d0584e5b9cdc766e23693371a6a9a0d859b686b350.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 271, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 304, + 271, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 547, + 327 + ], + "type": "text", + "content": "Figure 4. Larger LVMs perform better on downstream tasks. We evaluate LVMs of varying sizes on 4 different downstream tasks, following the 5 shot setting on the ImageNet validation set and report the perplexity. We find that perplexity decreases with larger models across all tasks, indicating strong scalability." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 426, + 171, + 526, + 270 + ], + "blocks": [ + { + "bbox": [ + 426, + 171, + 526, + 270 + ], + "lines": [ + { + "bbox": [ + 426, + 171, + 526, + 270 + ], + "spans": [ + { + "bbox": [ + 426, + 171, + 526, + 270 + ], + "type": "image", + "image_path": "696a76a887e9610f75ac82333f4c1f04e0aefc80e4f2cbb8c426f4765b1c6a18.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 338, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 409 + ], + "type": "text", + "content": "training progresses: 1) the training loss (perplexity) of the models, regardless of their size, continues to decrease; 2) as we increase the size of the model (parameter count), the loss decreases faster. These observations indicate that LVM shows strong scalability behavior with both larger models and more data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 412, + 548, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 548, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 548, + 615 + ], + "type": "text", + "content": "Scalability on downstream benchmarks. While the LVM overall loss scales well during training, there is no guarantee that the better overall model would also perform better on a given specific downstream task. Therefore, we evaluate different sizes of models on 4 downstream tasks: semantic segmentation, depth estimation, surface normal estimation, and edge detection. We evaluate these tasks on the ImageNet validation set and generate all the annotations using the corresponding method described in Sec. 3. For each task, we give 5 pairs consisting of the inputs and corresponding ground-truth annotations as well as the query image as input prompt and evaluate the perplexity of the ground-truth annotation under our model's prediction of the next 256 tokens (one image). We report the results in Figure 4. We see that larger models indeed attain lower perplexity across all tasks, showcasing that our scalable overall performance does transfer to a range of downstream tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Dataset ablation. While LVM attains better performance with larger models and more data, it is natural to ask whether each data component we collect in UVDv1 helps. To answer this question, we conduct an ablation study on our dataset by training several 3B models on subsets of our dataset, and compare their performances on downstream tasks. We use the same 4 downstream tasks and settings as before and present the results in Figure 5. We observe that each" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22865" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 0, + 544, + 330 + ], + "blocks": [ + { + "bbox": [ + 51, + 0, + 544, + 330 + ], + "lines": [ + { + "bbox": [ + 51, + 0, + 544, + 330 + ], + "spans": [ + { + "bbox": [ + 51, + 0, + 544, + 330 + ], + "type": "image", + "image_path": "e20f3c458684b4b7616c4b088c3c841732a416a0d8b0950ed539514001778f5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 330, + 546, + 353 + ], + "lines": [ + { + "bbox": [ + 47, + 330, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 546, + 353 + ], + "type": "text", + "content": "Figure 6. Frame predictions. LVM predicts the next frames (marked in red) given previous video frames as prompt. The results reveal that the LVM can predict the video frames while considering dynamic objects and camera motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 288, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 405 + ], + "type": "text", + "content": "data component contributes positively to the downstream tasks. LVM not only benefits from larger data, but also improves with more diversity in the dataset, which includes both annotated and unsupervised image and video data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 173, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 173, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 173, + 420 + ], + "type": "text", + "content": "5.2. Sequential Prompting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 485 + ], + "type": "text", + "content": "We begin with the most intuitive and straightforward approach to visually prompt the LVM: sequential reasoning. Here the prompt construction is very simple: we present the model with a sequence of 7 images and ask it to predict the next image (256 tokens)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 617 + ], + "type": "text", + "content": "Video frame prediction. The most direct task for sequential prompting is video prediction. Figure 6 presents several next frame prediction examples, prompted by sequences from the Kinetics-700 validation set. At the top, 7 frame prompts (blue border) are followed by the predicted frame (red border). We observe a certain degree of inferential ability regarding spatial positioning, viewpoint and object understanding. Perplexity of prediction on Kinetics val set is 49.8. The last 4 rows show predictions with longer context (15 frames) and a longer prediction (4 frames). See Figures 16 to 21 in the Appendix for many more examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Rotation and Category prediction. The same type of simple sequential prompting can be used in other ways as well. For example, Figure 15 shows how prompting the model with a sequence of 3D rotations of a synthetic object around an arbitrary axis allows it to predict further rotation. Or we can think of a list of items of a given category as a sequence and predict other ideas in that same category, as shown in Figure 14. Note that, while the system was trained" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 357, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 546, + 392 + ], + "type": "text", + "content": "on groups of images from the same ImageNet category, here the prompt consists of sketches, which have not been seen in any annotated data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 393, + 547, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 477 + ], + "type": "text", + "content": "Context length analysis. Next we ask how much temporal context is required to accurately predict the subsequent frame? We assessed the model's frame generation perplexity when prompted with a context of varying lengths (1 to 15 frames). On the Kinetics-700 val set, we see a clear improvement in perplexity from 1 to 11 frames after which it stabilizes (from " + }, + { + "bbox": [ + 304, + 393, + 547, + 477 + ], + "type": "inline_equation", + "content": "62.1 \\rightarrow 48.4" + }, + { + "bbox": [ + 304, + 393, + 547, + 477 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 482, + 421, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 421, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 421, + 495 + ], + "type": "text", + "content": "5.3. Analogy Prompting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "type": "text", + "content": "Our study progresses by evaluating a more complex prompting structure, which we call 'Analogy Prompting'. This method challenges the model to comprehend analogies of arbitrary length and complexity, thereby testing its advanced interpretative abilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "type": "text", + "content": "Qualitative Results. Figure 7 shows a sampling of qualitative results with analogy prompting on a number of tasks. The prompts consist of a sequence of 14 images giving examples of various tasks, followed by a 15th query image. Given each prompt, the next image predicted is the result. The top part of the figure shows several example prompts defining tasks that were part of the training set (but these actual images were never seen at training). The bottom part of the figure demonstrates generalization to tasks never shown at training. See Appendix for many more qualitative examples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Unseen Tasks and Dataset. We present the results for keypoint detection on Pascal 3D+ [90], evaluated using the standard Percentage of Correct Keypoints (PCK) metric with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22866" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 71, + 500, + 400 + ], + "blocks": [ + { + "bbox": [ + 89, + 71, + 500, + 400 + ], + "lines": [ + { + "bbox": [ + 89, + 71, + 500, + 400 + ], + "spans": [ + { + "bbox": [ + 89, + 71, + 500, + 400 + ], + "type": "image", + "image_path": "d8503a81df206b6a830757b15e66da6a2a3d8c51e15a6d1f1be3ce1e025f49cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 87, + 437, + 504, + 577 + ], + "blocks": [ + { + "bbox": [ + 46, + 403, + 547, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 547, + 436 + ], + "type": "text", + "content": "Figure 7. In and out of distribution prompting examples. Every row is a prompt that contains a sequence of images interleaved with annotations, followed by a query. The last image is predicted by the model (marked in red). The last 5 rows show examples where the query image is out of distribution (painting, sketch, etc) for the task it was trained for." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 437, + 504, + 577 + ], + "lines": [ + { + "bbox": [ + 87, + 437, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 87, + 437, + 504, + 577 + ], + "type": "image", + "image_path": "afb4ba85aaea0b31ac16e2067dfdcc7c253e5658ae115e65c070e04c4410b103.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 579, + 530, + 590 + ], + "lines": [ + { + "bbox": [ + 62, + 579, + 530, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 530, + 590 + ], + "type": "text", + "content": "Figure 8. Task Compositionality. Examples of prompts that combine two different tasks - object rotation and keypoint tracking." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 594, + 286, + 641 + ], + "blocks": [ + { + "bbox": [ + 49, + 594, + 286, + 641 + ], + "lines": [ + { + "bbox": [ + 49, + 594, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 49, + 594, + 286, + 641 + ], + "type": "table", + "html": "
ModelForeground SegmentationSingle Object DetectionColorization ↓
Split 0Split 1Split 2Split 3Split 1Split 2Split 3Split 4MSELPIPS
MAE (IN-1k)1.926.763.854.571.371.981.621.621.130.87
MAE-VQGAN (IN-1k)2.227.075.486.283.343.212.802.803.310.75
MAE (CVF)17.4225.7018.6416.535.494.985.245.840.430.55
MAE-VQGAN (CVF)27.8330.4426.1524.2524.1925.2025.3625.230.670.40
Ours48.9451.2947.6650.8248.2549.6050.0848.920.510.46
", + "image_path": "681745521dcc7a5e2b191e8414c8b15988e7ea771f11923cfa673f6225ed8c49.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 642, + 287, + 674 + ], + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 674 + ], + "type": "text", + "content": "Table 1. Comparison with Visual Prompting [8]. For Foreground Segmentation and Single Object Detection, we report the mIOU score. For Colorization, we report the MSE and LPIPS." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "a of threshold 0.1. Remarkably, LVM achieves a PCK of 81.2 without training on this dataset, demonstrating impressive generalization capabilities. In comparison, we show" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 594, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 632 + ], + "type": "text", + "content": "some existing task-specific model: StackedHourglass [60] scores 68.0 PCK, MSS-Net [42] achieves 68.9 PCK, and StarMap [101] registers 78.6 PCK." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Comparison with Visual Prompting. The closest approach to ours that also allows for defining arbitrary tasks is Visual Prompting [8]. In Table 1, we compare various visual prompting models on few-shot segmentation, object detection, and colorization tasks. Note that our sequential LVM beats previous approaches on almost all tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22867" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 261, + 201 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 261, + 201 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 261, + 201 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 261, + 201 + ], + "type": "image", + "image_path": "c41df0ad173827242d02488f71f3dee33fc9650e2e50abd2e6625b5ddea3cda4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "type": "text", + "content": "Figure 9. Miscellaneous Prompting. A variety of simple vision tasks, such as object replication (top), relighting (middle), and zooming in (bottom), can be simply specified via a suitably chosen visual sentence prompt that expresses the task to the LVM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 251, + 247, + 335 + ], + "blocks": [ + { + "bbox": [ + 83, + 251, + 247, + 335 + ], + "lines": [ + { + "bbox": [ + 83, + 251, + 247, + 335 + ], + "spans": [ + { + "bbox": [ + 83, + 251, + 247, + 335 + ], + "type": "image", + "image_path": "3df5b978d7fc8aacad29b6f95fe2b47ab8a388d3e13a843d7a2a4685ede0a650.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 336, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 357 + ], + "type": "text", + "content": "Figure 10. What comes next? Tasks that are not always easily describable in language." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 363, + 286, + 403 + ], + "blocks": [ + { + "bbox": [ + 50, + 363, + 286, + 403 + ], + "lines": [ + { + "bbox": [ + 50, + 363, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 50, + 363, + 286, + 403 + ], + "type": "image", + "image_path": "e10c1d17c730dcf91646a885a62667de00a2c63b864c26f9e3da650ed17da03f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 406, + 287, + 494 + ], + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 494 + ], + "type": "text", + "content": "Figure 11. \"Sparks of AGI?\" We prompt LVM with a masked reasoning visual sentence to infer the solution for non-verbal reasoning questions which are prevalent in IQ tests (masked image, second from the right). Left are successful cases and right is failure. We find that the model often infers and applies the abstract visual pattern correctly. So, we graciously hand over to you, our gentle reader, the task of pondering whether our modest LVM also exhibits the much-vaunted 'Sparks of AGI'." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "content": "Task Compositing. Figure 8 demonstrates compositing several tasks together within a single prompt. Here, we demonstrate the rotation task together with the novel keypoint correspondence task and ask the model to continue the pattern. The model can successfully combine these two at test time, demonstrating some degree of compositionality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 577, + 178, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 178, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 178, + 590 + ], + "type": "text", + "content": "5.4. Miscellaneous Prompts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": "Here we try to see how far we can push our model by offering it various prompts it has not seen before. Figure 9 shows a few such prompts that happened to work reasonably well. Figure 10 shows some prompts which are not easily describable by words – these are the type of tasks where LVMs might eventually outshine LLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In Figure 11, we show initial qualitative results on a typical visual reasoning question as found on non-verbal human IQ tests (Raven's Progressive Matrices [67]). With considerable squinting, one could imagine the LVM having a" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 309, + 73, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 545, + 185 + ], + "type": "image", + "image_path": "2cee850566bfc5566a34f6a35f432eeda073e4bc33191194e46dd3300caa2f80.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 185, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 185, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 185, + 547, + 316 + ], + "type": "text", + "content": "Figure 12. Failure Cases. This figure illustrates seven examples of failure cases: (1) Task confusion: the model interprets the counting task as style transfer, resulting in a pear-shaped apple. (2) Task entanglement: the model perceives the prompt as high frequencies from the entire image, rather than just from some parts of image (objects). (3) Wrong instance: the model correctly identifies rotation but mistakenly generates a brush instead of a manta. (4) Outlier detection versus generation: rather than detecting an outlier, the model directly generates it. (5) Digits: the model fails to represent digit sequences. (6) Tokenizer: poor performance with some out-of-distribution synthetic data generation (wrong background). (7) Sequence degeneration: sometimes frame predictions go astray." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 320, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 320, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 547, + 416 + ], + "type": "text", + "content": "latent ability for grasping abstract visual patterns and applying them to extrapolate the shown visual sequence. We also perform perplexity analysis on 10 classic Raven 5-way multiple-choice Matrices in Table 2, choosing the answer with lowest perplexity under our model. It shows that our results are better than chance. To further understand this process, we generate 900 new Raven-like synthetic tests to study individual factors: color, shape, and size following [8]." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 333, + 418, + 518, + 446 + ], + "blocks": [ + { + "bbox": [ + 333, + 418, + 518, + 446 + ], + "lines": [ + { + "bbox": [ + 333, + 418, + 518, + 446 + ], + "spans": [ + { + "bbox": [ + 333, + 418, + 518, + 446 + ], + "type": "table", + "html": "
Raven IQ matricesshapecolorsize
chance20%33%33%50%
Ours30%45%42%94%
", + "image_path": "07a0c93eea144c0693050fad70246b463787510032ad19197027fdc5b12283b7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 449, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 472 + ], + "type": "text", + "content": "Table 2. Performance on 10 Raven's Progressive Matrices and 900 synthetic reasoning tasks by lowest perplexity under our model." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 475, + 391, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 391, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 391, + 485 + ], + "type": "text", + "content": "5.5. Failure Cases" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 486, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 564 + ], + "type": "text", + "content": "Figure 12 shows some typical failure cases of the current model. One common element, the use of visual prompt to define a task is often under-constrained (more so than in language, since images are very high-dimensional), or the requested task might be beyond the capabilities of the current system. Other, more mundane failures involve issues with the tokenizer and lack of high-quality video training data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 566, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 708 + ], + "type": "text", + "content": "Acknowledgements: We are grateful to many friends and colleagues for discussions and comments on this work, including Yossi Gandelsman, Aleksander Holynski, Angjoo Kanazawa, Qianqian Wang, Sophia Koepke, Quoc Le, Chen Liang, Zichao Li, Ekin D. Cubuk, Assaf Shocher, Amir Zamir, Carl Vondrick, Ludwig Schmidt, Aviral Kumar, Xiaolong Wang and Miki Rubinstein. We thank the support from Google during Yutong's part-time internship. The work has been supported, in part, by ONR MURI N0001 4-22-1-2773, N00014-21-1-2801, and N00014-21-1-2812, ERC HOLI, Apple Graduate Fellowship to Yutong Bai, and compute donation via the Google TPU Research Cloud. Darrell's group is supported in part by the US DoD, BAIR&BDD industrial consortia." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22868" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 268, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 268, + 102 + ], + "type": "text", + "content": "[1] Common crawl repository. https://commoncrawl.org/." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 103, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 287, + 158 + ], + "type": "text", + "content": "[2] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1692-1700, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 159, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 159, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 58, + 159, + 288, + 204 + ], + "type": "text", + "content": "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 58, + 205, + 288, + 237 + ], + "type": "text", + "content": "[4] Okan Arikan and D. A. Forsyth. Interactive motion generation from examples. ACM Trans. Graph., 21(3):483-490, 2002." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 239, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 239, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 288, + 261 + ], + "type": "text", + "content": "[5] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 263, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 263, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 58, + 263, + 288, + 317 + ], + "type": "text", + "content": "[6] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13137-13146, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 319, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 319, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 58, + 319, + 288, + 352 + ], + "type": "text", + "content": "[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 354, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 354, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 58, + 354, + 288, + 397 + ], + "type": "text", + "content": "[8] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 399, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 399, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 58, + 399, + 288, + 430 + ], + "type": "text", + "content": "[9] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 434, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 488 + ], + "type": "text", + "content": "[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 490, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 288, + 544 + ], + "type": "text", + "content": "[11] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 546, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 288, + 578 + ], + "type": "text", + "content": "[12] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 580, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 288, + 624 + ], + "type": "text", + "content": "[13] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 625, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 288, + 712 + ], + "type": "text", + "content": "[14] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 117 + ], + "type": "text", + "content": "[15] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 119, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 119, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 119, + 547, + 161 + ], + "type": "text", + "content": "[16] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 163, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 547, + 196 + ], + "type": "text", + "content": "[17] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 198, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 547, + 262 + ], + "type": "text", + "content": "[18] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 264, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 547, + 297 + ], + "type": "text", + "content": "[19] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 299, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 299, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 299, + 546, + 364 + ], + "type": "text", + "content": "[20] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 366, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 366, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 547, + 430 + ], + "type": "text", + "content": "[21] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2634-2641, 2013." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 433, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 433, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 312, + 433, + 547, + 498 + ], + "type": "text", + "content": "[22] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 500, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 500, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 500, + 547, + 544 + ], + "type": "text", + "content": "[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 545, + 547, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 547, + 588 + ], + "type": "text", + "content": "[24] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE international conference on computer vision, pages 2051-2060, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 590, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 590, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 312, + 590, + 547, + 633 + ], + "type": "text", + "content": "[25] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 635, + 547, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 547, + 700 + ], + "type": "text", + "content": "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 703, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 703, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 703, + 546, + 714 + ], + "type": "text", + "content": "[27] Alexei A. Efros and William T. Freeman. Image quilting for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22869" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "texture synthesis and transfer. In Proceedings of SIGGRAPH 2001, 2001." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 97, + 288, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 97, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 53, + 97, + 288, + 130 + ], + "type": "text", + "content": "[28] Alexei A. Efros and Thomas K. Leung. Texture synthesis by non-parametric sampling. In IEEE International Conference on Computer Vision, pages 1033-1038, Corfu, Greece, 1999." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 133, + 288, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 133, + 288, + 176 + ], + "spans": [ + { + "bbox": [ + 53, + 133, + 288, + 176 + ], + "type": "text", + "content": "[29] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 178, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 288, + 212 + ], + "type": "text", + "content": "[30] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In CVPR 2011, pages 3281-3288. IEEE, 2011." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 247 + ], + "type": "text", + "content": "[31] David Donovan Garber. Computational Models for Texture Analysis and Texture Synthesis. PhD thesis, University of Southern California, Image Processing Institute, 1981." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 249, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 293 + ], + "type": "text", + "content": "[32] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "text", + "content": "[33] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 394 + ], + "type": "text", + "content": "[34] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE transactions on pattern analysis and machine intelligence, 38(1):142-158, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 474 + ], + "type": "text", + "content": "[35] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 542 + ], + "type": "text", + "content": "[36] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 544, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 288, + 610 + ], + "type": "text", + "content": "[37] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 613, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 288, + 666 + ], + "type": "text", + "content": "[38] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "text", + "content": "[39] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 138 + ], + "type": "text", + "content": "[40] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 140, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 547, + 172 + ], + "type": "text", + "content": "[41] H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In International Conf. on Computer Vision (ICCV), pages 3192-3199, 2013." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 173, + 546, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 546, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 546, + 216 + ], + "type": "text", + "content": "[42] Lipeng Ke, Ming-Ching Chang, Honggang Qi, and Siwei Lyu. Multi-scale structure-aware network for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 713-728, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 217, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 547, + 271 + ], + "type": "text", + "content": "[43] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6129-6138, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 273, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 273, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 273, + 547, + 304 + ], + "type": "text", + "content": "[44] Lucas Kovar, Michael Gleicher, and Frederic Pighin. Motion Graphs. In Proceedings of SIGGRAPH '02, San Antonio, TX, 2002." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 305, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 305, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 547, + 348 + ], + "type": "text", + "content": "[45] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 350, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 547, + 403 + ], + "type": "text", + "content": "[46] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 404, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 547, + 459 + ], + "type": "text", + "content": "[47] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 460, + 546, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 546, + 503 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 546, + 503 + ], + "type": "text", + "content": "[48] Jehee Lee, Jinxiang Chai, Paul Reitsma, Jessica K. Hodgins, and Nancy Pollard. Interactive control of avatars animated with human motion data. ACM Transactions on Graphics (TOG), 21(3):491 - 500, 2002." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 504, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 547, + 547 + ], + "type": "text", + "content": "[49] Jianshu Li, Jian Zhao, Yunchao Wei, Congyan Lang, Yidong Li, Terence Sim, Shuicheng Yan, and Jiashi Feng. Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 548, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 547, + 592 + ], + "type": "text", + "content": "[50] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 593, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 546, + 647 + ], + "type": "text", + "content": "[51] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 647, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 547, + 713 + ], + "type": "text", + "content": "[52] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22870" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 149 + ], + "type": "text", + "content": "[54] Shikun Liu, Linxi Fan, Edward Johns, Zhiding Yu, Chaowei Xiao, and Anima Anandkumar. Prismer: A vision-language model with an ensemble of experts. arXiv preprint arXiv:2303.02506, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "type": "text", + "content": "[55] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "text", + "content": "[56] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 326 + ], + "type": "text", + "content": "[57] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 327, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 403 + ], + "type": "text", + "content": "[58] Mathew Monfort, Bowen Pan, Kandan Ramakrishnan, Alex Andonian, Barry A McNamara, Alex Lascelles, Quanfu Fan, Dan Gutfreund, Rogério Schmidt Feris, and Aude Oliva. Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9434-9445, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "type": "text", + "content": "[59] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE conference on computer vision and pattern recognition, pages 2408-2415. IEEE, 2012." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 449, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 288, + 503 + ], + "type": "text", + "content": "[60] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 504, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 288, + 546 + ], + "type": "text", + "content": "[61] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 548, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 601 + ], + "type": "text", + "content": "[62] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 602, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 288, + 667 + ], + "type": "text", + "content": "[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "text", + "content": "[64] Kris Popat and Rosalind W. Picard. Novel cluster-based probability model for texture synthesis, classification, and compression. In Proc. SPIE Visual Comm. and Image Processing, 1993." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 116 + ], + "type": "text", + "content": "[65] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "type": "text", + "content": "[66] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 162, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 547, + 183 + ], + "type": "text", + "content": "[67] John C Raven. Standardization of progressive matrices, 1938. British Journal of Medical Psychology, 1941." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 184, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 547, + 249 + ], + "type": "text", + "content": "[68] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 250, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 547, + 282 + ], + "type": "text", + "content": "[69] Arno Schodl, Richard Szeliski, David H. Salesin, and Irfan Essa. Video textures. In Proceedings of SIGGRAPH '00, pages 489-498, 2000." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 284, + 547, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 284, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 547, + 338 + ], + "type": "text", + "content": "[70] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 338, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 547, + 415 + ], + "type": "text", + "content": "[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 415, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 547, + 480 + ], + "type": "text", + "content": "[72] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 481, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 547, + 514 + ], + "type": "text", + "content": "[73] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 514, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 547, + 536 + ], + "type": "text", + "content": "[74] Claude E. Shannon. A mathematical theory of communication. Bell Sys. Tech. Journal, 27, 1948." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 537, + 547, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 547, + 558 + ], + "type": "text", + "content": "[75] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 559, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 547, + 634 + ], + "type": "text", + "content": "[76] Gunnar A Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 635, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 547, + 679 + ], + "type": "text", + "content": "[77] Gunnar A Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 681, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 547, + 712 + ], + "type": "text", + "content": "[78] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22871" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 288, + 128 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 288, + 128 + ], + "type": "text", + "content": "[79] X. Soria, E. Riba, and A. Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1912-1921, Los Alamitos, CA, USA, 2020. IEEE Computer Society." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 288, + 184 + ], + "type": "text", + "content": "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 288, + 239 + ], + "type": "text", + "content": "[81] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, koray kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 241, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 295 + ], + "type": "text", + "content": "[82] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, New York, New York, USA, 2016. PMLR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 297, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 330 + ], + "type": "text", + "content": "[83] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 331, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 376 + ], + "type": "text", + "content": "[84] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 376, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 288, + 431 + ], + "type": "text", + "content": "[85] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th International Conference on Machine Learning, page 1096-1103, 2008." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 433, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 288, + 476 + ], + "type": "text", + "content": "[86] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 477, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 288, + 533 + ], + "type": "text", + "content": "[87] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 288, + 555 + ], + "type": "text", + "content": "[88] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of laion-2b, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 556, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 288, + 589 + ], + "type": "text", + "content": "[89] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "type": "text", + "content": "[90] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 635, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 690 + ], + "type": "text", + "content": "[91] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 714 + ], + "type": "text", + "content": "[92] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 532 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 140 + ], + "type": "text", + "content": "[93] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 141, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 547, + 196 + ], + "type": "text", + "content": "[94] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 197, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 197, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 197, + 547, + 262 + ], + "type": "text", + "content": "[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 264, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 547, + 329 + ], + "type": "text", + "content": "[96] Yingchen Yu, Fangneng Zhan, Rongliang Wu, Jianxiong Pan, Kaiwen Cui, Shijian Lu, Feiying Ma, Xuansong Xie, and Chunyan Miao. Diverse image inpainting with bidirectional and autoregressive transformers. In Proceedings of the 29th ACM International Conference on Multimedia, pages 69-78, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 331, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 331, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 312, + 331, + 547, + 385 + ], + "type": "text", + "content": "[97] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 386, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 386, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 312, + 386, + 547, + 418 + ], + "type": "text", + "content": "[98] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In CVPR, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 420, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 420, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 312, + 420, + 547, + 442 + ], + "type": "text", + "content": "[99] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 443, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 547, + 487 + ], + "type": "text", + "content": "[100] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 487, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 547, + 532 + ], + "type": "text", + "content": "[101] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22872" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_content_list.json b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1debde0d81a8964e4f7dfc6ee25d31ceb9d7678e --- /dev/null +++ b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_content_list.json @@ -0,0 +1,1768 @@ +[ + { + "type": "text", + "text": "SfmCAD: Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations", + "text_level": 1, + "bbox": [ + 78, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\mathrm{PuLi^{1,2}}$ Jianwei Guo $^{1,2*}$ Huibin Li $^{2,1}$ Bedrich Benes $^{3}$ Dong-Ming Yan $^{1,2}$", + "bbox": [ + 161, + 200, + 807, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MAIS, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 272, + 223, + 692, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences $^{3}$ Computer Science, Purdue University", + "bbox": [ + 84, + 239, + 879, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce304b141a898683556788c5ffe5b1e21efcae42ffc29deb5e85e3bf8cbc3ca1.jpg", + "image_caption": [ + "Figure 1. SfmCAD parses voxel shapes and reconstructs them into a parameterized set of sketch-based feature modeling operations: the input voxels, neural typed sketch+path representation, and the final CAD model (each part is randomly colored)." + ], + "image_footnote": [], + "bbox": [ + 81, + 282, + 883, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 513, + 313, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces SfmCAD, a novel unsupervised network that reconstructs 3D shapes by learning the Sketch-based Feature Modeling operations commonly used in modern CAD workflows. Given a 3D shape represented as voxels, SfmCAD learns a neural-typed sketch+path parameterized representation, including 2D sketches of feature primitives and their 3D sweeping paths without supervision, for inferring feature-based CAD programs. SfmCAD employs 2D sketches for local detail representation and 3D paths to capture the overall structure, achieving a clear separation between shape details and structure. This conversion into parametric forms enables users to seamlessly adjust the shape's geometric and structural features, thus enhancing interpretability and user control. We demonstrate the effectiveness of our method by applying SfmCAD to many different types of objects, such as CAD parts, ShapeNet objects, and tree shapes. Extensive comparisons show that SfmCAD produces compact and faithful 3D reconstructions with superior quality compared to alternatives. The code is released at https://github.com/BunnySoCrazy/SfmCAD.", + "bbox": [ + 75, + 539, + 473, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 513, + 630, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing shapes into CAD representations is a fundamental problem in computer vision and graphics with numerous commercial applications in industrial design and manufacturing [39]. Recent years have seen an increased interest in leveraging deep learning techniques to learn geometric representations of 3D shapes, including implicit fields [2, 25, 28], geometric primitives [19, 22, 24, 37], boundary representation (B-Rep) [9, 17], or Constructive Solid Geometry (CSG) [3, 36, 49], each with its unique characteristics and applications, contributing to the rich tapestry of research and development in this field.", + "bbox": [ + 496, + 546, + 892, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, current CAD reconstruction methods are still lacking detail reconstruction accuracy, compactness, and interpretability of the obtained models. For example, a typical line of research is dedicated to reconstruction-oriented tasks, i.e., representing 3D shapes as parametric primitives, implicit fields, or B-rep. These methods deliver a high level of detail and accuracy by capturing intricate details and complex geometries with remarkable precision. However, constructing CAD models by assembling instances of primitives or by identifying faces, edges, and vertices is a tedious process [35]. Besides, the results often include complex and unintuitive operations that make the model difficult to inter-", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: jianwei.guo@nlpr.ia.ac.cn", + "bbox": [ + 94, + 886, + 364, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4671", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pret and edit. This lack of control poses significant challenges in scenarios where user interaction and shape manipulation are required.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A straightforward approach to CAD reconstruction is to learn CSG representation [14, 30, 36, 49] that implicitly recovers the CAD design history. A CSG 3D shape is expressed as a tree with leaves consisting of volumetric primitives (e.g., cubes, spheres, and cylinders) that are combined using Boolean operations in the CSG inner nodes. Despite the sought-after attribute of compact shape representation, these methods often fall short regarding reconstruction accuracy by using only a set of simple typed primitives.", + "bbox": [ + 75, + 138, + 467, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce SfmCAD, a novel neural model for reconstructing 3D shapes into their CAD representation. Compared to related work, SfmCAD improves the reconstruction accuracy and the level of model editability by decoupling 3D structures and local geometric details of the shape. SfmCAD is inspired by feature-based modeling, a predominant method for creating 3D models in modern CAD systems, which sequentially add features (e.g., holes, slots, bosses) that represent common manufacturing operations [6, 41, 42]. Specifically, we employ unsupervised learning to retrieve a neural typed sketch+path representation of 3D shapes. Taking 3D voxel data as input, SfmCAD learns implicit fields that embody the sketch profile and generates control points for a Bezier curve, constituting the sweeping path for the component's geometry. The basic extrusion operation has been learned in a supervised [18, 38] or unsupervised [23, 31] manner. However, our sketch+path representation supports not only extrude but also sweep, loft, and revolve operations, thus enhancing the shape representation capabilities, as demonstrated in Fig. 1.", + "bbox": [ + 75, + 276, + 467, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At the core of the SfmCAD is a two-stage learning strategy that operates in a coarse-to-fine manner to tackle the inherent complexity and time-intensiveness of simultaneous sketch profile and sweeping path learning. Initially, the network is trained to learn a coarse Box+Path representation of the shape, enabling a quick grasp of the shape's path and broad structure using box-like geometric proxies. Then, an implicit network delves deeper into the geometric details of the shape, enhancing the reconstruction accuracy. We demonstrate the capability and versatility of our method by applying SfmCAD to many different types of objects, including CAD parts [16], ShapeNet objects [1], tree branches [20], and by comparing to previous works. We claim the following contributions:", + "bbox": [ + 75, + 580, + 467, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose SfmCAD, a novel neural approach that parses 3D shapes as a set of industry-standard sketch-based CAD modeling operations. To the best of our knowledge, SfmCAD is the first unsupervised and universal neural network for learning the common CAD commands, including extrude, sweep, loft, and revolve.", + "2. We introduce novel decoupling of 3D structures and" + ], + "bbox": [ + 76, + 795, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "local geometric details, which are represented by 3D Bézier curves and implicit 2D sketches. This representation ensures accurate reconstruction and ease of editing.", + "bbox": [ + 517, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We propose a two-stage learning strategy, namely 3D Box+Path learning and 2D implicit sketch learning, which significantly accelerates the training speed in shape parser learning.", + "bbox": [ + 500, + 136, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 210, + 635, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometric primitive extraction. Geometric primitive is a commonly used representation for approximating and abstracting 3D shapes [13]. Traditional methods based on RANSAC [32], Hough transform [10], or variational optimization [4, 46] have been used to detect and fit primitives on point clouds or polygonal meshes, but they often require careful parameter tuning for each shape.", + "bbox": [ + 496, + 236, + 890, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the overwhelming complexity of traditional methods, deep neural networks addressed the primitive segmentation/detection from point clouds [19, 47], they fit parametric surfaces [22, 29, 37], or detect parametric curves and sharp corners [9, 24, 40] to achieve compact CAD reconstructions. Although achieving remarkable accuracy on reconstruction tasks, these methods only output individual primitives with limited types, which restricts their capability for reconstructing complex and more general 3D shapes. Learning CSG-based reconstruction. CSG reconstruction embodies a 3D shape of high complexity and nonconvexity with CSG boolean operations to represent the 3D shape construction process [5]. CSGNet [36] first develops a neural model that parses a shape into a sequence of CSG operations. More recent works follow the line of CSG parsing by advancing the inference without any supervision [14] or improving representation capability with a three-layer reformulation of the classic CSG-tree [30], or handling richer geometric and topological variations by introducing quadric surface primitives [49, 50]. However, CSG reconstruction tends to combine a large number of shape primitives that limit the user editing capabilities. Moreover, using only basic primitives (e.g., boxes, spheres) is insufficient for approximating complicated shapes, thus usually limiting the reconstruction accuracy of small details [50].", + "bbox": [ + 496, + 342, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning feature-based reconstruction. Modern CAD workflows use feature-based modeling that decomposes a target shape into a sequence of CAD commands [34], such as drawing a sketch followed by CAD operations such as extrusion, loft, etc. This paradigm makes modeling more efficient and in tune with how designers and engineers work. Recent work has explored the potential of deep learning to generate 2D engineering sketches [7, 27, 33], or directly learn B-Rep CAD models [6, 11, 12, 17]. Several approaches propose generative models for CAD design, predicting a sequence of CAD modeling operations to produce editable CAD models, such as DeepCAD [42], Fu", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4672", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6d6e1b6a26f7ccf15ebac890899b83926979f626c2bfbc86d5b9a25fc5388a3.jpg", + "image_caption": [ + "Extrude", + "Figure 2. Sketch-based feature modeling operators supported in SfmCAD. The resulting 3D shape of each operation is called a feature primitive." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 151, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9b7a09a24e9599bc32258379f4261cfd9cc38fc157f7429a03f39757c3441ebb.jpg", + "image_caption": [ + "Loft" + ], + "image_footnote": [], + "bbox": [ + 165, + 93, + 269, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f2ff0c5fc413f576a8ac896823b0543c66e68606a78f1172668b4e5c91dac37.jpg", + "image_caption": [ + "Sweep" + ], + "image_footnote": [], + "bbox": [ + 285, + 95, + 331, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6cd51c599307985d3e078f6b310e5835392e30f5b3ba6ed7dd2fd441302b9325.jpg", + "image_caption": [ + "Revolve" + ], + "image_footnote": [], + "bbox": [ + 354, + 95, + 452, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion360 [41], Zone graph [43], and SkexGen [44]. These works focus on parametric CAD generation and do not address the reconstruction task of CAD models with suitable geometric loss functions.", + "bbox": [ + 75, + 270, + 468, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most closely related to our work are the approaches that learn sketch+extrude operation (e.g., Point2Cyl [38], ExtrudeNet [31], SECAD-Net [23]), in which a collection of extrusions are predicted to build CAD models. While expressive within their scopes, these methods are limited to only one type of CAD operation (i.e., extrude) to simplify the search space. SfmCAD includes loft, revolve, and sweep reconstructions, broadening the shape representation capabilities. This approach enables us to test on a broader range of shapes beyond CAD.", + "bbox": [ + 75, + 330, + 468, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Statement and Overview", + "text_level": 1, + "bbox": [ + 76, + 494, + 383, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch-based feature modeling creates 2D engineering sketches and then lifts them to 3D along a path using CAD operations [8, 21]. Here, we first introduce this typed sketch+path representation for sketch-based modeling and then give an overview of its generation from a voxel model.", + "bbox": [ + 75, + 520, + 468, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.Typed Sketch+Path Modeling", + "text_level": 1, + "bbox": [ + 76, + 604, + 341, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define a profile as a closed curve, and its enclosed region is composed of one or multiple inner/outer loops. We define a sketch as the collection of one profile and its loops in a 2D plane. Let $Z$ denote the sketch and $\\mathcal{C}$ be a 3D path to extend the sketches to create a 3D solid shape. We define four principal operations that are commonly used in modern CAD workflow (see Fig. 2): Extrude extends $Z$ along a linear path, Revolve produces a model by rotating $Z$ around an axis, Sweep moves $Z$ along a specified path to generate generalized cylinder shapes, and Loft allows for variations in the sketch, interpolating between multiple distinct sketches $\\{Z_1, Z_2, \\ldots, Z_n\\}$ along the shortest distances between them to create complex shapes.", + "bbox": [ + 75, + 628, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the similarity between the above operations, i.e., lifting 2D sketches to a 3D shape along a 3D path, we use a general typed sketch $^+$ path to represent all of these operations, where the term 'typed' indicates the specific type of operation. Without loss of generality, we call the 3D", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "path a sweeping path and call the lofted or swept 3D shape a feature primitive. We further use a Bézier curve $\\Upsilon(t)$ to represent the sweeping path:", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\Upsilon (t) = \\sum_ {i = 0} ^ {n} \\binom {n} {i} (1 - t) ^ {n - i} t ^ {i} \\mathbf {P} _ {i}, \\quad t \\in [ 0, 1 ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 155, + 890, + 194 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_i$ is a control point. The Bézier cubic is expressive enough to approximate the shapes shown in the paper, but higher-order curves can be obtained by simply changing the number of control points output by the network.", + "bbox": [ + 496, + 210, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, it is noteworthy that the Extrude and Revolve operations can be regarded as general variants of the sweep operation because a Bézier curve becomes a linear path (for Extrude) if all the control points are collinear and an $n$ -piece (usually $n = 4$ ) cubic Bézier curve can approximate a circular path (for Revolve) [45]. In the following, we focus on reconstructing Sweep and Loft operations, where Sweep contains Extrude and Revolve, and the Loft is restricted to contain only two sketches with a linear path between them.", + "bbox": [ + 496, + 275, + 890, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 431, + 609, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a 3D voxel shape, we aim to reconstruct the CAD model by neurally decomposing the target shape into a sequence of typed sketch+path representation, which can be executed by the corresponding CAD commands to assemble a compact CAD model. Learning the path and sketch of a shape simultaneously is challenging and time-consuming. Thus, we propose a two-stage learning approach to improve the efficiency of network training. First, the network learns a Box+Path representation to enclose the input shape, capturing the sweeping path, basic geometry, and rotation angles. Then, we learn implicit sketches that refine the coarse box+path into a detailed shape representation.", + "bbox": [ + 496, + 458, + 890, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 3 details the two-stage learning network. We begin with a 3D convolutional encoder that transforms the voxel input into a feature vector $\\mathbf{z}$ , which is then processed through a multilayer perceptron (MLP), yielding matrices $\\mathbf{P}$ and $\\mathbf{B}$ . Here, $\\mathbf{P}$ are the control points of the Bezier curves representing the sweeping paths (Fig. 3 (b)), and $\\mathbf{B}$ encapsulates the box+path parameters (Fig. 3 (c)), including length, width, and sweep's twist angles. Once the sweeping paths are derived from Stage 1, we focus on learning the geometric details surrounding each path. The feature $\\mathbf{z}$ is then mapped to $N_{p}$ local features via another MLP. These vectors are subsequently concatenated with test points and fed into an implicit sketch network. The output from this network includes two occupancy values that represent the top and bottom sketches (Fig. 3 (d)), respectively. We then obtain the final CAD reconstruction by sweeping or lofting the detected sketches along the paths (Fig. 3 (e)).", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4673", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db4ebf5c3eeface6d083e50e13715f5614f636be0e816a07eb93f582048def8e.jpg", + "image_caption": [ + "Figure 3. Network architecture for SfmCAD: Stage 1 encodes the input voxels into feature $\\mathbf{z}$ via a 3D convolutional encoder. An MLP then processes $\\mathbf{z}$ to generate the control point coordinates $\\mathbf{P}_1, \\ldots, \\mathbf{P}_4$ for each Bézier curve, as well as the dimensions $l, w$ and twist angles $\\delta^u, \\delta^b$ for each Box proxy $\\mathcal{B}$ . In Stage 2, another MLP transforms feature $\\mathbf{z}$ into local features $\\mathbf{z}'$ , which are then concatenated with the testing point coordinate and fed into the implicit sketch network to predict the sketch SDF value. After extending the sketches along the paths, we combine all parts for the final reconstruction." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 872, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/64fe7203a5b631c604c7cf03634621b3e325522753e6b3c211bd21d291fae198.jpg", + "image_caption": [ + "Input", + "Figure 4. 2D illustration of our Box+Path representation (right) to approximate the input shape (left), which can be decomposed into a set of feature primitives (middle, the green arrows indicate the sweeping paths)." + ], + "image_footnote": [], + "bbox": [ + 104, + 393, + 200, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1fbdd88bc1a27ea19bd5f9bbf8248a677502e7ba8acb114eba69869654b27b5f.jpg", + "image_caption": [ + "Feature primitives" + ], + "image_footnote": [], + "bbox": [ + 225, + 393, + 321, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c64f2ae0f223d3741d6f3e3d667583745235a2a896926f81799696a09ba5b2c.jpg", + "image_caption": [ + "Box+Path" + ], + "image_footnote": [], + "bbox": [ + 348, + 393, + 442, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 571, + 212, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we first present the box+path representation, which can be efficiently learned to approximate the input shape by stacking a set of box-like proxies along the sweeping path. Then, we define the neural sketch+path to refine the coarse reconstruction. Finally, we introduce the two-stage learning network for inferring this coarse-to-fine reconstruction.", + "bbox": [ + 75, + 595, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Box+Path Representation", + "text_level": 1, + "bbox": [ + 76, + 696, + 310, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous unsupervised methods [23, 30, 31] for computing the Signed Distance Function (SDF) for geometric primitives (e.g., cubes, spheres, cylinders) rely on linear transformations that convert global testing point coordinates into local ones. However, the inherent non-linearity of curved paths limits the direct application of linear methods. We introduce a linear approximation method for the curved SDF computation, which we term Box+Path representation. We will first discuss the Box-Path in a 2D context to provide a more intuitive understanding before expanding it to 3D.", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Box+Path in 2D. Specifically, we approximate one feature primitive $\\mathcal{P}$ using a discrete set of $N_{s}$ rigid box proxies,", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "i.e., $\\mathcal{P} = \\{\\mathcal{B}_i\\}_{i=1}^{N_s}$ . The center position $\\mathbf{c}_i$ of each box $\\mathcal{B}_i$ is obtained by uniformly sampling the Bezier curve $\\Upsilon(t)$ :", + "bbox": [ + 498, + 395, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i} = \\Upsilon (i / N _ {s}), \\quad i = 1, \\dots , N _ {s} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 435, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The height $h_i$ and the normal vector (or heading direction) $\\mathbf{n}_i$ of each box are calculated as", + "bbox": [ + 498, + 460, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\left| \\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i} \\right|, \\mathbf {n} _ {i} = \\frac {\\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i}}{\\left| \\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i} \\right|} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 497, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By incorporating a learnable width parameter $w$ , the 2D Box+Path can be fully characterized.", + "bbox": [ + 498, + 534, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Box+Path in 3D. When extended to 3D space, we introduce the tangent vector $\\mathbf{t}_i$ and the binormal vector $\\mathbf{b}_i$ to parameterize the orientation of the $i$ -th box proxy. To allow the sketch to rotate by a prescribed angle along the sweeping path, we should learn rotation angles $\\delta^u$ and $\\delta^b$ at the start and end points of the sweeping path. Then, the intermediate rotation angles $\\delta_i$ are obtained by linear interpolation along the sweeping path. Next, we compute the tangent vector $\\mathbf{t}_i$ as the cross product of the normal $\\mathbf{n}_i$ and the positive $z$ -axis unit vector $\\mathbf{e}_z$ . We then rotate $\\mathbf{t}_i$ around $\\mathbf{n}_i$ :", + "bbox": [ + 496, + 564, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {i} = R _ {\\mathbf {n} _ {i}} \\left(\\mathbf {n} _ {i} \\times \\mathbf {e} _ {z}, \\delta_ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 724, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R_{\\mathbf{n}_i}(\\mathbf{v}, \\delta_i)$ represents a counter-clockwise rotation of vector $\\mathbf{v}$ with angle $\\delta_i$ around $\\mathbf{n}_i$ . The binormal vector $\\mathbf{b}_i$ is obtained as a cross product of $\\mathbf{n}_i \\times \\mathbf{t}_i$ . Altogether, the triplet $\\{\\mathbf{n}_i, \\mathbf{t}_i, \\mathbf{b}_i\\}$ defines the orientation for each box along the curve path. The 3D Box+Path is fully characterized by incorporating learnable dimension parameters $l, w$ .", + "bbox": [ + 496, + 748, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Neural Sketch+Path", + "text_level": 1, + "bbox": [ + 500, + 847, + 692, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present differentiable methods for lofting or sweeping 2D sketches into 3D solid shapes to accurately reconstruct", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4674", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the feature primitives. The shape of each 2D sketch is defined by an implicit representation, where we use $\\hat{S}_{\\mathrm{sk}}$ to represent the SDF of sketches. We now describe the differentiable operators that calculate the SDF of the 3D feature primitives corresponding to the sketches.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differentiable Sketch-Extrude. In the context of generating the geometry of feature primitives from sketch and extrude operations, $\\hat{S}_{\\mathrm{extrude}}^i$ denotes the $i$ -th sketch-extrude operation, $\\mathbf{x}_{i_z}$ represents the $z$ -component of the $i$ -th position vector, and $h_i$ is the height of the $i$ -th feature primitive. We introduce a set of intermediate functions $\\mathcal{M}_1, \\mathcal{M}_2$ , and $\\mathcal{M}_3$ to facilitate a clearer formulation:", + "bbox": [ + 75, + 167, + 467, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {M} _ {1} \\left(\\hat {\\mathcal {S}} _ {\\mathbf {s k}} ^ {i}\\right) = \\max \\left(\\hat {\\mathcal {S}} _ {\\mathbf {s k}} ^ {i}, \\left| \\mathbf {x} _ {i _ {z}} \\right| - h _ {i}\\right), \\\\ \\mathcal {M} _ {2} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i}\\right) = \\max \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i}, 0\\right), \\\\ \\mathcal {M} _ {3} = \\max \\left(\\left| \\mathbf {x} _ {i _ {z}} \\right| - h _ {i}, 0\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 279, + 387, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SDF of the $i$ -th extrusion primitive is then given by:", + "bbox": [ + 76, + 347, + 449, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} _ {\\text {e x t r u d e}} ^ {i} = \\min \\left(\\mathcal {M} _ {1} \\left(\\hat {S} _ {\\mathrm {s k}} ^ {i}\\right), 0\\right) + \\sqrt {\\mathcal {M} _ {2} \\left(\\hat {S} _ {\\mathrm {s k}} ^ {i}\\right) ^ {2} + \\mathcal {M} _ {3} ^ {2}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 372, + 468, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differentiable Sketch-Loft. For the loft operation, we interpolate between two distinct sketches, $\\hat{S}_{\\mathrm{sk}}^{u_i}$ and $\\hat{S}_{\\mathrm{sk}}^{l_i}$ , at the upper and lower faces of the $i$ -th feature primitive. Given the interpolation factor $\\alpha$ as $|\\mathbf{x}_{i_z}| / h_i$ , the linearly interpolated sketch, $\\hat{S}_{\\mathrm{sk}}^{i_\\alpha}$ , is computed as:", + "bbox": [ + 75, + 402, + 467, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}} = (1 - \\alpha) \\cdot \\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {l _ {i}} + \\alpha \\cdot \\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {u _ {i}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 487, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the same intermediate variables as in the extrusion case, but with $\\hat{S}_{\\mathrm{sk}}^{i}$ replaced by $\\hat{S}_{\\mathrm{sk}}^{i_{\\alpha}}$ . The SDF of the $i$ -th loft primitive can be expressed similarly to Eq. 5:", + "bbox": [ + 75, + 515, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {S}} _ {\\text {l o f t}} ^ {i} = \\min \\left(\\mathcal {M} _ {1} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}}\\right), 0\\right) + \\sqrt {\\mathcal {M} _ {2} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}}\\right) ^ {2} + \\mathcal {M} _ {3} ^ {2}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 570, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differentiable Sketch-Sweep. For sweep operation, we decompose it into a series of $N_{s}$ extrusions. We denote the predicted SDF for the $j$ -th extrude segment as $\\hat{S}_{\\mathrm{extrude}}^i$ . The SDF of the $i$ -th sweep primitive is then obtained by taking the minimum of the SDFs of all extrusion segments:", + "bbox": [ + 75, + 601, + 467, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {S}} _ {\\text {s w e e p}} ^ {i} = \\min _ {j \\in 1, \\dots , N _ {s}} \\hat {\\mathcal {S}} _ {\\text {e x t r u d e}} ^ {j}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 684, + 468, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformation to Parametric CAD. Following [23], we transform implicit sketches into parametric ones. With the control points of Bezier curves outputted by the network, we can achieve the sketch+path parametric reconstruction.", + "bbox": [ + 75, + 717, + 467, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Two-stage Network Training", + "text_level": 1, + "bbox": [ + 76, + 786, + 333, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1: Box+Path learning. In this stage, the network predicts the control point coordinates $\\{\\mathbf{P}_1,\\dots ,\\mathbf{P}_4\\}$ of the cubic Bezier curve and the parameters $\\{l_i,w_i,\\delta^u,\\delta^b\\}$ describing the $i$ -th Box $^+$ Path $\\mathcal{B}_i$ . These parameters subsequently inform the computation of the remaining parameters $\\{h_i,\\mathbf{n}_i,\\mathbf{t}_i,\\mathbf{b}_i\\}$ , as detailed in Sect. 4.1. To align the", + "bbox": [ + 75, + 809, + 467, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "path of $\\mathcal{B}$ with the target shape path, we first compute the SDF, $\\hat{S}_{\\mathrm{box}}^{ij}$ , for each box at the $i$ -th primitive and $j$ -th segment, where $i \\in [1, N_p]$ and $j \\in [1, N_s]$ , with $N_p$ and $N_s$ representing the total number of primitives and segments. Taking the union over the $\\hat{S}_{\\mathrm{box}}^{ij}$ yields the SDF for the overall box+path $\\hat{S}_{\\mathcal{B}}$ :", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {S}} _ {\\mathcal {B}} = \\bigcup_ {i = 1} ^ {N _ {p}} \\bigcup_ {j = 1} ^ {N _ {s}} \\hat {\\mathcal {S}} _ {\\text {b o x}} ^ {i j}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 191, + 890, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bigcup (x) = \\mathrm{softmin}(\\mu \\cdot x)\\cdot x$ and $\\mu$ is a hyperparameter. We then apply a differentiable converter $\\Psi$ similar to prior works [23, 31] to transform the $\\hat{S}_{\\mathcal{B}}$ into occupancy, yielding $\\hat{\\mathcal{O}}_{\\mathcal{B}} = \\Psi (\\hat{S}_{\\mathcal{B}})$ , where $\\Psi (x) = (\\tanh (-\\beta \\cdot x) + 1) / 2$ , and $\\beta$ is a hyperparameter. We then compute the Mean Squared Error (MSE) loss between $\\hat{\\mathcal{O}}_{\\mathcal{B}}$ and the ground truth occupancy $\\tilde{\\mathcal{O}}$ :", + "bbox": [ + 496, + 241, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {B}} = \\mathbb {E} _ {x \\in \\mathcal {X}} \\left[ | | \\hat {\\mathcal {O}} _ {\\mathcal {B}} - \\tilde {\\mathcal {O}} | | _ {2} ^ {2} \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 347, + 890, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X}$ is the set of all testing points. The exclusive use of $\\mathcal{L}_{\\mathcal{B}}$ may lead to potential challenges: (1) the emergence of the curve with acute angles, and (2) the dimensions of $\\mathcal{B}$ vastly outstrip the path length. Both increase the risk of shape self-intersections, which are considered invalid in CAD. To address these issues, we introduce two additional regularization terms,", + "bbox": [ + 496, + 375, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s m}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\min _ {j = 1, \\dots , N _ {s}} \\left(\\mathbf {n} _ {i, j} \\cdot \\mathbf {n} _ {i, j + 1}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 487, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l w}} = \\frac {1}{2 N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\operatorname {R e L U} \\left(h _ {i} + w _ {i} - 2 \\cdot \\Theta\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 532, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{sm}}$ penalizes the sharpest curvature in the curve by maximizing the inner product between adjacent $\\mathbf{n}$ , and $\\mathcal{L}_{\\mathrm{lw}}$ penalizes any length or width greater than a threshold $\\Theta$ . Combine with balance factors $\\lambda_{1}, \\lambda_{2}$ , the overall loss of Stage 1 is:", + "bbox": [ + 496, + 580, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b o x}} = \\mathcal {L} _ {\\mathcal {B}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {s m}} + \\lambda_ {2}. \\mathcal {L} _ {\\mathrm {l w}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 665, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Implicit sketch learning. Using the path learned in Stage 1, we progress towards fine reconstruction in Stage 2 by learning sketches. We initially utilize a Multi-Layer Perceptron (MLP) to convert the feature $\\mathbf{z}$ into $N_{p}$ local features $\\mathbf{z}^{\\prime}$ . These local features are then concatenated with the local $\\{x,y\\}$ coordinates of each point and fed into an implicit sketch network, which is essentially an MLP predicting the SDF value $\\hat{S}_{\\mathrm{sk}}^{i}$ . For the sketch-loft operation, two implicit sketch networks are used to learn $\\hat{S}_{\\mathrm{sk}}^{u_i}$ and $\\hat{S}_{\\mathrm{sk}}^{l_i}$ . Following the computation methods outlined in Sect. 4.2, we can derive the predicted overall occupancy value $\\hat{\\mathcal{O}}_{\\mathrm{F}}$ by:", + "bbox": [ + 496, + 688, + 890, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {O}} _ {\\mathrm {F}} = \\left\\{ \\begin{array}{l l} \\Psi (\\bigcup_ {i = 1} ^ {N _ {p}} \\hat {S} _ {\\text {s w e e p}} ^ {i}), & \\text {f o r s w e e p o p e r a t i o n} \\\\ \\Psi (\\bigcup_ {i = 1} ^ {N _ {p}} \\hat {S} _ {\\text {l o f t}} ^ {i}), & \\text {f o r l o f t o p e r a t i o n .} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 863, + 851, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4675", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, an MSE loss $\\mathcal{L}_{\\mathrm{rec}}$ is computed against the ground truth $\\mathcal{O}$ :", + "bbox": [ + 75, + 90, + 468, + 118 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e c}} = \\mathbb {E} _ {x \\in \\mathcal {X}} \\left[ \\left| \\left| \\hat {\\mathcal {O}} _ {\\mathrm {F}} - \\tilde {\\mathcal {O}} \\right| \\right| _ {2} ^ {2} \\right]. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 116, + 468, + 141 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The direct computation of $\\mathcal{L}_{\\mathrm{rec}}$ over the entire point set $\\mathcal{X}$ requires the repetition of $\\mathcal{X}$ corresponding to $\\hat{S}_{\\mathrm{extrude}}$ for $N_{p} \\times N_{s}$ times, which is time-consuming. To accelerate the training of the implicit sketch network, we propose a real-time sampling strategy, exploiting the sketch consistency trait inherent in the sweep operation. Specifically, we sample grid-like testing points $y_{i}$ in the base sketch plane of the $i$ -th path $\\Upsilon_{i}$ obtained in the first stage. The occupancy value of $y_{i}$ is determined by inputting it into the implicit sketch network and applying the converter $\\Psi$ . Subsequently, $y_{i}$ is moved along the curve for $N_{s} - 1$ times, with its occupancy value directly replicated, yielding $\\hat{\\mathcal{O}}_{\\mathrm{Fi}}^{*}$ . The ground truth occupancy value $\\tilde{\\mathcal{O}}_i^*$ is provided by the input voxel. After performing this procedure for all paths and conducting a union operation, the loss function corresponding to real-time sampling is given by:", + "bbox": [ + 75, + 146, + 472, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e c}} ^ {*} = \\mathbb {E} _ {y \\in \\mathcal {Y}} \\left[ | | \\hat {\\mathcal {O}} _ {\\mathrm {F}} ^ {*} - \\tilde {\\mathcal {O}} ^ {*} | | _ {2} ^ {2} \\right], \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 400, + 468, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{V}$ represents the set of testing points obtained through real-time sampling, and $*$ denotes computation using real-time sampling data.", + "bbox": [ + 75, + 436, + 468, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the SfmCAD is designed to choose between learning either loft or sweep operations, which is determined by its hyperparameters. However, enabling the network to concurrently learn both operations is also viable, with details provided in the supplementary materials.", + "bbox": [ + 75, + 481, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results and Evaluation", + "text_level": 1, + "bbox": [ + 76, + 570, + 294, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of SfmCAD on two reconstruction tasks: (1) holistic shapes of different categories and (2) semantically segmented shapes. We show the output editability of SfmCAD in the supplemental materials.", + "bbox": [ + 75, + 595, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 665, + 158, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We implemented SfmCAD in PyTorch by using an NVIDIA® TITAN RTX GPU. The Adam optimizer [15] is utilized with a learning rate 1e-4 and beta values of (0.5, 0.99). The hyperparameters in Sect. 4.3 are set to: $\\mu = 20$ , $\\beta = 50$ , $\\Theta = 0.1$ , $\\lambda_1 = 0.05$ , $\\lambda_2 = 0.05$ . In our evaluations, we employ the best-performing operation (Loft/Sweep) for each category of holistic shapes and every part of segmented shapes. SfmCAD trains for 500 epochs per stage with a batch size of 24, followed by fine-tuning each test shape for 200 iterations per stage. All baseline methods are trained for 1,000 epochs and fine-tuned for 400 iterations for a fair comparison.", + "bbox": [ + 75, + 688, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We utilize three common metrics for 3D shape reconstruction: Chamfer Distance (CD), Normal", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/efdf5ab4d3d5aa26ec6a92047f1176e01467ba81e248c0d24532a9f611a99b02.jpg", + "image_caption": [ + "Figure 5. Visual comparison on ABC dataset." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 888, + 369 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/35307c2921a3ca012040d6e698a26714fb361e804b6cfa87565cdae00f326cca.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on ABC dataset." + ], + "table_footnote": [], + "table_body": "
MethodsCD↓ECD↓NC↑
UCSG-Net [14]1.23321.7860.866
CSG-Stump [30]0.6717.7510.892
ExtrudeNet [31]0.5197.1110.885
SECAD-Net [23]0.5067.2860.884
Ours0.3955.0380.919
", + "bbox": [ + 537, + 438, + 852, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistency (NC), and Edge Chamfer Distance (ECD). All experiments generate the reconstructed shapes via Marching Cubes at a resolution of $256^3$ . To calculate CD, we sample 8,192 points from the ground truth and predicted shape surfaces. For ECD, we employ a threshold of 0.1 for normal cross-products to identify points close to the edges on the ground truth surface and 0.5 for the predicted surface, considering the lack of sharpness in shapes generated by Marching Cubes. For better visual clarity, we magnify the CD and ECD by factors of $10^3$ and $10^2$ .", + "bbox": [ + 496, + 561, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Reconstruction from Holistic Shape", + "text_level": 1, + "bbox": [ + 500, + 724, + 808, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SfmCAD with UCSG-Net [14], CSG-Stump [30], ExtrudeNet [31] and SECAD-Net [23]. We test these methods on holistic shapes from different sources, including CAD parts in ABC [16], ShapeNet objects [1], and a synthetic tree dataset. The comparison against CAPRI-Net [49] is provided in the supplementary materials.", + "bbox": [ + 496, + 748, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on ABC dataset. We employ the voxel grids and sampling point data provided by [49]. The dataset included 5,000 groups of data for training and 1,000 for testing. We select 100 shape subsets from the test set for eval", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4676", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c9f492e70ed69c1a36b65b7620e294165e499b64318137444c87d16d57e9196.jpg", + "image_caption": [ + "Figure 6. Visual comparison on ShapeNet dataset." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b10f921e597eea6b9b57a877317b589ec5cc0e2ca03aec9a5b35fe4d43c2d01.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on ShapeNet dataset." + ], + "table_footnote": [], + "table_body": "
MethodsCD↓ECD↓NC↑
UCSG-Net [14]2.11123.6160.815
CSG-Stump [30]2.16817.5920.818
ExtrudeNet [31]0.91014.1100.849
SECAD-Net [23]0.83614.8180.837
Ours0.62614.0960.867
", + "bbox": [ + 114, + 450, + 426, + 549 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58b5469b325f2844f5da6839f4ee3694aa9fa445811bcb5128d2696a0543266f.jpg", + "table_caption": [ + "Table 3. Quantitative comparison on Tree dataset." + ], + "table_footnote": [], + "table_body": "
MethodsExtrudeNet [31]SECAD-Net [23]Ours
CD↓7.2422.1970.686
NC↑0.6080.6220.722
", + "bbox": [ + 78, + 589, + 468, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "uation. The visual results are shown in Fig. 5, while the quantitative results are presented in Tab. 1.", + "bbox": [ + 75, + 671, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on ShapeNet dataset. We utilize three categories from ShapeNet: chair, table, and display. The voxel grids and sampling points are provided by [3]. We selected 120 shapes for evaluation, with 40 from each category. Visual and quantitative results are in Fig. 6 and Tab. 2.", + "bbox": [ + 75, + 702, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on tree branches. We use a procedural modeling approach of Palubicki et al. [26] to generate tree meshes voxelized into a resolution of $64^{3}$ . We adopt the method from [2] to obtain sampling points and occupancy values. This process results in 5,000 data groups, of which 4,500 are used for training and 500 for testing. We train all methods using increasingly denser voxel resolutions $(16^{3},32^{3},64^{3})$ and select 50 trees for numerical evaluation.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc7741dd422e4c70e6ec5e185ad0b0e76423ff7409b14cae5c587b893a65149d.jpg", + "image_caption": [ + "Figure 7. Visual comparison on Tree dataset." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 890, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion. Note given the cylindrical nature of tree branches, we adapt our Box+Path representation by substituting boxes with cylinders for the final reconstruction (see Fig. 7 and Tab. 3 for visual and quantitative results).", + "bbox": [ + 496, + 349, + 890, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Summary of comparative analysis. As demonstrated in Tables 1, 2, and 3, SfmCAD consistently outperforms other methods in terms of CD, ECD, and NC, suggesting that the sketch+path representation of SfmCAD offers more flexible shape expression capabilities than the geometric primitives used by UCSG-Net and CSG-Stump, as well as the simple sketch+extrusion operations employed by ExtrudeNet and SECAD. Figures 5, 6, and 7 further show SfmCAD's ability to generate detailed reconstructions (such as thin structures) across diverse shape categories.", + "bbox": [ + 496, + 411, + 892, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Reconstruction from Segmented Shape", + "text_level": 1, + "bbox": [ + 498, + 573, + 834, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic segmentation provides information for reconstructing complex shapes. Previous CAD reconstruction methods [23, 31] could not utilize this information because their single-type primitives lack the flexibility to represent each part independently. In contrast, SfmCAD successfully reconstructs multiple primitives with its sketch+path representation capability.", + "bbox": [ + 496, + 597, + 890, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the task of unsupervised CAD reconstruction utilizing semantic segmentation. The evaluation is conducted on PartNet dataset [48], which provides instance-level 3D part information. We utilize three categories: chair, table, and trashcan. We voxelize the segmented parts and sampled occupancy values at resolutions $16^{3}, 32^{3}, 64^{3}$ following the scheme by [2]. This process yields a total of 12,000 chair parts (1,229 complete), 12,000 table parts (1,491 complete), and 3,000 trashcan parts (234 complete). The train-test split for all categories was set as 5:1. We selected 90 complete shapes, including 40 tables, 40 chairs, and 10 trashcans, for evaluation.", + "bbox": [ + 496, + 703, + 892, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All methods were trained on part shapes using increas", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4677", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8f8675dfb12f8055225e7b780ce587fc1aab2e348791fb888e141a6bb401604.jpg", + "image_caption": [ + "Figure 8. Visual comparison on PartNet dataset." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 308 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/105c23a70844c235c472c4a98d9ffe2cca4c1c83afcf96f59e31bdc4999a8cac.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on PartNet dataset." + ], + "table_footnote": [], + "table_body": "
MethodsCD↓ECD↓NC↑
ExtrudeNet [31]3.2510.5490.783
SECAD-Net [23]1.8880.4720.818
Ours1.4100.4460.814
", + "bbox": [ + 114, + 369, + 426, + 436 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/208e048c6289a1c01e5084e870c110b80cfd3cc7b479a2dee1d2e3db088303a8.jpg", + "table_caption": [ + "Table 5. Ablation study on sweep twist angles and loss terms." + ], + "table_footnote": [], + "table_body": "
Settings- (δu, δl)-δl- Llw- LsmOurs
CD↓4.0923.3674.8043.2683.142
ECD↓0.1730.1390.1640.1540.133
NC↑0.6550.7140.6850.7320.776
", + "bbox": [ + 78, + 474, + 468, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ingly denser voxel resolutions $(16^{3}, 32^{3}, 64^{3})$ . For each part, we set ExtrudeNet [31] and SECAD-Net [23] to output an extruded cylinder and SfmCAD to output a sketch+path shape. After reconstructing all parts, they were reassembled to obtain the final CAD models.", + "bbox": [ + 75, + 565, + 468, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative results in Tab. 4 indicate that our method outperforms others. Compared to the closest competitor, SECAD-Net, SfmCAD improves the CD by $25.3\\%$ and the ECD by $5.5\\%$ . Fig. 8 illustrates that the sketch+extrusion reconstruction of ExtrudeNet and SECAD can only characterize the nonlinear shapes through sketches. In contrast, SfmCAD can model these complex shapes using sketches and free-form paths, enabling a more refined reconstruction.", + "bbox": [ + 75, + 642, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 786, + 230, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted an ablation study by using 40 box+path reconstructions on chair class from PartNet to evaluate the efficiency of our network design and loss terms. Initially, we disable the two sweep twist angles $\\delta^u$ , $\\delta^l$ or $\\delta^l$ . Then we separately discarded $\\mathcal{L}_{\\mathrm{sm}}$ and $\\mathcal{L}_{\\mathrm{lw}}$ . The visual results in Fig. 9 and the quantitative results in Tab. 5 suggest that", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59112db521e7246d8c7e561bac2e5cad7b7329363fce054f73854cf20fc3b3d9.jpg", + "image_caption": [ + "Figure 9. Visual comparison between reconstruction results for ablation study. From top to bottom, each column shows the predicted 3D path, vector $\\{\\mathbf{n},\\mathbf{t},\\mathbf{b}\\}$ , and the box+path representation." + ], + "image_footnote": [], + "bbox": [ + 511, + 87, + 883, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "removing any of these settings negatively affects the reconstruction performance of SfmCAD.", + "bbox": [ + 498, + 387, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 441, + 769, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented an unsupervised neural network for CAD reconstruction that infers high-level CAD modeling operations. We have also introduced a unique sketch+path representation that encapsulates the local geometric details and overall structural information of the 3D shape. A novel two-stage learning strategy, operating in a coarse-to-fine manner, was proposed to address the complexity of simultaneous sketch and path learning, thus improving the efficiency of network training. We evaluated SfmCAD using several different categories of objects and demonstrated its advantages by comparing it to state-of-the-art methods.", + "bbox": [ + 496, + 470, + 892, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In future work, we plan to explore using 2D sketches and 3D part templates to enhance the efficiency of unsupervised CAD reconstruction. Besides, given the manipulable property of SfmCAD's outputs, extending this work for generative CAD design or modeling would be meaningful.", + "bbox": [ + 496, + 640, + 893, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank the anonymous reviewers for their valuable suggestions. This work is partially funded by the Strategic Priority Research Program of the Chinese Academy of Sciences (XDB0640000), Guangdong Science and Technology Program (2023B1515120026), National Natural Science Foundation of China (62172416, 62172415, U22B2034, 62262043, 62365014), and Youth Innovation Promotion Association of the Chinese Academy of Sciences (2022131). This work was supported by NRCS grant #NR233A750004G044 to Benes. The findings and conclusions should not be construed to represent any agency determination or policy.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4678", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6", + "[2] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5939-5948, 2019. 1, 7", + "[3] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-Net: Generating compact meshes via binary space partitioning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 45–54, 2020. 1, 7", + "[4] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. In ACM SIGGRAPH 2004 Papers, pages 905–914. 2004. 2", + "[5] Tao Du, Jeevana Priya Inala, Yewen Pu, Andrew Spielberg, Adriana Schulz, Daniela Rus, Armando Solar-Lezama, and Wojciech Matusik. Inversecsg: Automatic conversion of 3d models to csg trees. ACM Trans. Graph., 37(6):1-16, 2018. 2", + "[6] Elona Dupont, Kseniya Cherenkova, Anis Kacem, Sk Aziz Ali, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Cadops-net: Jointly learning cad operation types and steps from boundary-representations. In International Conference on 3D Vision (3DV), pages 114-123. IEEE, 2022. 2", + "[7] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems (NeurIPS), pages 5885–5897, 2021. 2", + "[8] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Trans. Graph., 39(6):1-16, 2020. 3", + "[9] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: CAD reconstruction by B-rep chain complex generation. ACM Trans. Graph., 41(4): 1-18, 2022. 1, 2", + "[10] Rostislav Hulik, Michal Spanel, Pavel Smrz, and Zdenek Materna. Continuous plane detection in point-cloud data based on 3d hough transform. Journal of visual communication and image representation, 25(1):86-97, 2014. 2", + "[11] Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G Lambourne, Karl DD Willis, Thomas Davies, Hooman Shayani, and Nigel Morris. Uv-net: Learning from boundary representations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11703-11712, 2021. 2", + "[12] Pradeep Kumar Jayaraman, Joseph G Lambourne, Nishkrit Desai, Karl DD Willis, Aditya Sanghi, and Nigel JW Morris. Solidgen: An autoregressive model for direct b-rep synthesis. arXiv preprint arXiv:2203.13944, 2022. 2", + "[13] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2", + "[14] Kacper Kania, Maciej Zieba, and Tomasz Kajdanowicz. UCSG-NET-Unsupervised discovering of constructive solid" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "geometry tree. In Advances in Neural Information Processing Systems (NeurIPS), pages 8776-8786, 2020. 2, 6, 7", + "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6", + "[16] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 9601-9611, 2019. 2, 6", + "[17] Joseph G Lambourne, Karl DD Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. BRepNet: A topological message passing system for solid models. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 12773-12782, 2021. 1, 2", + "[18] Joseph George Lambourne, Karl Willis, Pradeep Kumar Jayaraman, Longfei Zhang, Aditya Sanghi, and Kamal Rahimi Malekshan. Reconstructing editable prismatic cad from rounded voxel models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2", + "[19] Eric-Tuan Lé, Minhyuk Sung, Duygu Ceylan, Radomir Mech, Tamy Boubekeur, and Niloy J Mitra. CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In IEEE International Conference on Computer Vision (ICCV), pages 7457-7466, 2021. 1, 2", + "[20] Bosheng Li, Jacek Kałuzny, Jonathan Klein, Dominik L. Michels, Wojtek Pałubicki, Bedrich Benes, and Soren Pirk. Learning to reconstruct botanical trees from single images. ACM Trans. Grap., 40(6):1-15, 2021. 2", + "[21] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Sketch2cad: Sequential cad modeling by sketching in context. ACM Trans. Graph., 39(6):1-14, 2020. 3", + "[22] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3d point clouds. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 2652-2660, 2019. 1, 2", + "[23] Pu Li, Jianwei Guo, Xiaopeng Zhang, and Dong-Ming Yan. Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 16816-16826, 2023. 2, 3, 4, 5, 6, 7, 8", + "[24] Yuanqi Li, Shun Liu, Xinran Yang, Jianwei Guo, Jie Guo, and Yanwen Guo. Surface and edge detection for primitive fitting of point clouds. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2", + "[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 4460-4470, 2019. 1", + "[26] Wojciech Palubicki, Kipp Horel, Steven Longay, Adam Runions, Brendan Lane, Radomír Měch, and Przemyslaw Prusinkiewicz. Self-organizing tree models for image synthesis. ACM Trans. Graph., 28(3):1–10, 2009. 7", + "[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. In Advances in Neural Information Processing Systems (NeurIPS), pages 5077-5088, 2021. 2" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4679", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 165-174, 2019. 1", + "[29] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 10344-10353, 2019. 2", + "[30] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, et al. CSG-stump: A learning friendly CSG-like representation for interpretable shape parsing. In IEEE International Conference on Computer Vision (ICCV), pages 12478-12487, 2021. 2, 4, 6, 7", + "[31] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude for shape parsing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 482-498, 2022. 2, 3, 4, 5, 6, 7, 8", + "[32] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient RANSAC for point-cloud shape detection. In Comput. Graph. Forum, pages 214-226, 2007. 2", + "[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In International Conference on Learning Representations (ICLR), 2022. 2", + "[34] Jami J Shah and Martti Mäntylä. Parametric and feature-based CAD/CAM: concepts, techniques, and applications. John Wiley & Sons, 1995. 2", + "[35] Vadim Shapiro. Solid modeling. Handbook of computer aided geometric design, 20:473-518, 2002. 1", + "[36] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5515-5523, 2018. 1, 2", + "[37] Gopal Sharma, Difan Liu, Subhransu Maji, Evangelos Kalogerakis, Siddhartha Chaudhuri, and Radomir Mech. ParSeNet: A parametric surface fitting network for 3d point clouds. In European Conference on Computer Vision (ECCV), pages 261-276, 2020. 1, 2", + "[38] Mikaela Angelina Uy, Yen-Yu Chang, Minhyuk Sung, Purvi Goel, Joseph G Lambourne, Tolga Birdal, and Leonidas J Guibas. Point2Cyl: Reverse engineering 3D objects from point clouds to extrusion cylinders. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11850-11860, 2022. 2, 3", + "[39] Tamas Varady and Ralph Martin. Reverse engineering. In Handbook of Computer Aided Geometric Design, pages 651-681. 2002. 1", + "[40] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. PIE-Net: Parametric inference of point cloud edges. Advances in Neural Information Processing Systems (NeurIPS), pages 20167–20178, 2020. 2", + "[41] Karl DD Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G Lambourne, Armando Solar-Lezama, and Wo" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "jciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM Trans. Graph., 40(4):1-24, 2021. 2, 3", + "[42] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In IEEE International Conference on Computer Vision (ICCV), pages 6772–6782, 2021. 2", + "[43] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl DD Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 6062-6070, 2021. 3", + "[44] Xiang Xu, Karl DD Willis, Joseph G Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, and Yasutaka Furukawa. Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In International Conference on Machine Learning, 2022. 3", + "[45] Fujio Yamaguchi. Curves and surfaces in computer aided geometric design. Springer Science & Business Media, 2012. 3", + "[46] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 2", + "[47] Siming Yan, Zhenpei Yang, Chongyang Ma, Haibin Huang, Etienne Vouga, and Qixing Huang. Hpnet: Deep primitive segmentation using hybrid representations. In IEEE International Conference on Computer Vision (ICCV), pages 2753-2762, 2021. 2", + "[48] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, page to appear, 2019. 7", + "[49] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11768–11778, 2022. 1, 2, 6", + "[50] Fenggen Yu, Qimin Chen, Maham Tanveer, Ali Mahdavi Amiri, and Hao Zhang. $\\mathrm{D}^2\\mathrm{csg}$ : Unsupervised learning of compact csg trees with dual complements and dropouts. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 670 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4680", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_model.json b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6663e891b1d5bb87d336b52c511c87bb82f3a028 --- /dev/null +++ b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.13, + 0.89, + 0.178 + ], + "angle": 0, + "content": "SfmCAD: Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.202, + 0.808, + 0.223 + ], + "angle": 0, + "content": "\\(\\mathrm{PuLi^{1,2}}\\) Jianwei Guo\\(^{1,2*}\\) Huibin Li\\(^{2,1}\\) Bedrich Benes\\(^{3}\\) Dong-Ming Yan\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.224, + 0.694, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)MAIS, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.241, + 0.88, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences \\(^{3}\\)Computer Science, Purdue University" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.283, + 0.885, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Figure 1. SfmCAD parses voxel shapes and reconstructs them into a parameterized set of sketch-based feature modeling operations: the input voxels, neural typed sketch+path representation, and the final CAD model (each part is randomly colored)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.515, + 0.314, + 0.531 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.474, + 0.859 + ], + "angle": 0, + "content": "This paper introduces SfmCAD, a novel unsupervised network that reconstructs 3D shapes by learning the Sketch-based Feature Modeling operations commonly used in modern CAD workflows. Given a 3D shape represented as voxels, SfmCAD learns a neural-typed sketch+path parameterized representation, including 2D sketches of feature primitives and their 3D sweeping paths without supervision, for inferring feature-based CAD programs. SfmCAD employs 2D sketches for local detail representation and 3D paths to capture the overall structure, achieving a clear separation between shape details and structure. This conversion into parametric forms enables users to seamlessly adjust the shape's geometric and structural features, thus enhancing interpretability and user control. We demonstrate the effectiveness of our method by applying SfmCAD to many different types of objects, such as CAD parts, ShapeNet objects, and tree shapes. Extensive comparisons show that SfmCAD produces compact and faithful 3D reconstructions with superior quality compared to alternatives. The code is released at https://github.com/BunnySoCrazy/SfmCAD." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.515, + 0.631, + 0.531 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Reconstructing shapes into CAD representations is a fundamental problem in computer vision and graphics with numerous commercial applications in industrial design and manufacturing [39]. Recent years have seen an increased interest in leveraging deep learning techniques to learn geometric representations of 3D shapes, including implicit fields [2, 25, 28], geometric primitives [19, 22, 24, 37], boundary representation (B-Rep) [9, 17], or Constructive Solid Geometry (CSG) [3, 36, 49], each with its unique characteristics and applications, contributing to the rich tapestry of research and development in this field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "However, current CAD reconstruction methods are still lacking detail reconstruction accuracy, compactness, and interpretability of the obtained models. For example, a typical line of research is dedicated to reconstruction-oriented tasks, i.e., representing 3D shapes as parametric primitives, implicit fields, or B-rep. These methods deliver a high level of detail and accuracy by capturing intricate details and complex geometries with remarkable precision. However, constructing CAD models by assembling instances of primitives or by identifying faces, edges, and vertices is a tedious process [35]. Besides, the results often include complex and unintuitive operations that make the model difficult to inter-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.365, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: jianwei.guo@nlpr.ia.ac.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4671" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "pret and edit. This lack of control poses significant challenges in scenarios where user interaction and shape manipulation are required." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.468, + 0.275 + ], + "angle": 0, + "content": "A straightforward approach to CAD reconstruction is to learn CSG representation [14, 30, 36, 49] that implicitly recovers the CAD design history. A CSG 3D shape is expressed as a tree with leaves consisting of volumetric primitives (e.g., cubes, spheres, and cylinders) that are combined using Boolean operations in the CSG inner nodes. Despite the sought-after attribute of compact shape representation, these methods often fall short regarding reconstruction accuracy by using only a set of simple typed primitives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.468, + 0.58 + ], + "angle": 0, + "content": "We introduce SfmCAD, a novel neural model for reconstructing 3D shapes into their CAD representation. Compared to related work, SfmCAD improves the reconstruction accuracy and the level of model editability by decoupling 3D structures and local geometric details of the shape. SfmCAD is inspired by feature-based modeling, a predominant method for creating 3D models in modern CAD systems, which sequentially add features (e.g., holes, slots, bosses) that represent common manufacturing operations [6, 41, 42]. Specifically, we employ unsupervised learning to retrieve a neural typed sketch+path representation of 3D shapes. Taking 3D voxel data as input, SfmCAD learns implicit fields that embody the sketch profile and generates control points for a Bezier curve, constituting the sweeping path for the component's geometry. The basic extrusion operation has been learned in a supervised [18, 38] or unsupervised [23, 31] manner. However, our sketch+path representation supports not only extrude but also sweep, loft, and revolve operations, thus enhancing the shape representation capabilities, as demonstrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.793 + ], + "angle": 0, + "content": "At the core of the SfmCAD is a two-stage learning strategy that operates in a coarse-to-fine manner to tackle the inherent complexity and time-intensiveness of simultaneous sketch profile and sweeping path learning. Initially, the network is trained to learn a coarse Box+Path representation of the shape, enabling a quick grasp of the shape's path and broad structure using box-like geometric proxies. Then, an implicit network delves deeper into the geometric details of the shape, enhancing the reconstruction accuracy. We demonstrate the capability and versatility of our method by applying SfmCAD to many different types of objects, including CAD parts [16], ShapeNet objects [1], tree branches [20], and by comparing to previous works. We claim the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.468, + 0.884 + ], + "angle": 0, + "content": "1. We propose SfmCAD, a novel neural approach that parses 3D shapes as a set of industry-standard sketch-based CAD modeling operations. To the best of our knowledge, SfmCAD is the first unsupervised and universal neural network for learning the common CAD commands, including extrude, sweep, loft, and revolve." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.9 + ], + "angle": 0, + "content": "2. We introduce novel decoupling of 3D structures and" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.796, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "local geometric details, which are represented by 3D Bézier curves and implicit 2D sketches. This representation ensures accurate reconstruction and ease of editing." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.892, + 0.198 + ], + "angle": 0, + "content": "3. We propose a two-stage learning strategy, namely 3D Box+Path learning and 2D implicit sketch learning, which significantly accelerates the training speed in shape parser learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.211, + 0.637, + 0.227 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Geometric primitive extraction. Geometric primitive is a commonly used representation for approximating and abstracting 3D shapes [13]. Traditional methods based on RANSAC [32], Hough transform [10], or variational optimization [4, 46] have been used to detect and fit primitives on point clouds or polygonal meshes, but they often require careful parameter tuning for each shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.719 + ], + "angle": 0, + "content": "To overcome the overwhelming complexity of traditional methods, deep neural networks addressed the primitive segmentation/detection from point clouds [19, 47], they fit parametric surfaces [22, 29, 37], or detect parametric curves and sharp corners [9, 24, 40] to achieve compact CAD reconstructions. Although achieving remarkable accuracy on reconstruction tasks, these methods only output individual primitives with limited types, which restricts their capability for reconstructing complex and more general 3D shapes. Learning CSG-based reconstruction. CSG reconstruction embodies a 3D shape of high complexity and nonconvexity with CSG boolean operations to represent the 3D shape construction process [5]. CSGNet [36] first develops a neural model that parses a shape into a sequence of CSG operations. More recent works follow the line of CSG parsing by advancing the inference without any supervision [14] or improving representation capability with a three-layer reformulation of the classic CSG-tree [30], or handling richer geometric and topological variations by introducing quadric surface primitives [49, 50]. However, CSG reconstruction tends to combine a large number of shape primitives that limit the user editing capabilities. Moreover, using only basic primitives (e.g., boxes, spheres) is insufficient for approximating complicated shapes, thus usually limiting the reconstruction accuracy of small details [50]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Learning feature-based reconstruction. Modern CAD workflows use feature-based modeling that decomposes a target shape into a sequence of CAD commands [34], such as drawing a sketch followed by CAD operations such as extrusion, loft, etc. This paradigm makes modeling more efficient and in tune with how designers and engineers work. Recent work has explored the potential of deep learning to generate 2D engineering sketches [7, 27, 33], or directly learn B-Rep CAD models [6, 11, 12, 17]. Several approaches propose generative models for CAD design, predicting a sequence of CAD modeling operations to produce editable CAD models, such as DeepCAD [42], Fu" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.152, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.182, + 0.143, + 0.191 + ], + "angle": 0, + "content": "Extrude" + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.094, + 0.271, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.182, + 0.23, + 0.191 + ], + "angle": 0, + "content": "Loft" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.097, + 0.332, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.182, + 0.33, + 0.192 + ], + "angle": 0, + "content": "Sweep" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.097, + 0.454, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.182, + 0.428, + 0.192 + ], + "angle": 0, + "content": "Revolve" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.203, + 0.47, + 0.245 + ], + "angle": 0, + "content": "Figure 2. Sketch-based feature modeling operators supported in SfmCAD. The resulting 3D shape of each operation is called a feature primitive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.469, + 0.331 + ], + "angle": 0, + "content": "sion360 [41], Zone graph [43], and SkexGen [44]. These works focus on parametric CAD generation and do not address the reconstruction task of CAD models with suitable geometric loss functions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.469, + 0.483 + ], + "angle": 0, + "content": "Most closely related to our work are the approaches that learn sketch+extrude operation (e.g., Point2Cyl [38], ExtrudeNet [31], SECAD-Net [23]), in which a collection of extrusions are predicted to build CAD models. While expressive within their scopes, these methods are limited to only one type of CAD operation (i.e., extrude) to simplify the search space. SfmCAD includes loft, revolve, and sweep reconstructions, broadening the shape representation capabilities. This approach enables us to test on a broader range of shapes beyond CAD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.384, + 0.511 + ], + "angle": 0, + "content": "3. Problem Statement and Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.598 + ], + "angle": 0, + "content": "Sketch-based feature modeling creates 2D engineering sketches and then lifts them to 3D along a path using CAD operations [8, 21]. Here, we first introduce this typed sketch+path representation for sketch-based modeling and then give an overview of its generation from a voxel model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.342, + 0.623 + ], + "angle": 0, + "content": "3.1.Typed Sketch+Path Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.825 + ], + "angle": 0, + "content": "We define a profile as a closed curve, and its enclosed region is composed of one or multiple inner/outer loops. We define a sketch as the collection of one profile and its loops in a 2D plane. Let \\( Z \\) denote the sketch and \\( \\mathcal{C} \\) be a 3D path to extend the sketches to create a 3D solid shape. We define four principal operations that are commonly used in modern CAD workflow (see Fig. 2): Extrude extends \\( Z \\) along a linear path, Revolve produces a model by rotating \\( Z \\) around an axis, Sweep moves \\( Z \\) along a specified path to generate generalized cylinder shapes, and Loft allows for variations in the sketch, interpolating between multiple distinct sketches \\( \\{Z_1, Z_2, \\ldots, Z_n\\} \\) along the shortest distances between them to create complex shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Considering the similarity between the above operations, i.e., lifting 2D sketches to a 3D shape along a 3D path, we use a general typed sketch \\(^+\\) path to represent all of these operations, where the term 'typed' indicates the specific type of operation. Without loss of generality, we call the 3D" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "path a sweeping path and call the lofted or swept 3D shape a feature primitive. We further use a Bézier curve \\(\\Upsilon(t)\\) to represent the sweeping path:" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.156, + 0.892, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\Upsilon (t) = \\sum_ {i = 0} ^ {n} \\binom {n} {i} (1 - t) ^ {n - i} t ^ {i} \\mathbf {P} _ {i}, \\quad t \\in [ 0, 1 ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.892, + 0.273 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_i\\) is a control point. The Bézier cubic is expressive enough to approximate the shapes shown in the paper, but higher-order curves can be obtained by simply changing the number of control points output by the network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Moreover, it is noteworthy that the Extrude and Revolve operations can be regarded as general variants of the sweep operation because a Bézier curve becomes a linear path (for Extrude) if all the control points are collinear and an \\( n \\)-piece (usually \\( n = 4 \\)) cubic Bézier curve can approximate a circular path (for Revolve) [45]. In the following, we focus on reconstructing Sweep and Loft operations, where Sweep contains Extrude and Revolve, and the Loft is restricted to contain only two sketches with a linear path between them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.432, + 0.61, + 0.447 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Given a 3D voxel shape, we aim to reconstruct the CAD model by neurally decomposing the target shape into a sequence of typed sketch+path representation, which can be executed by the corresponding CAD commands to assemble a compact CAD model. Learning the path and sketch of a shape simultaneously is challenging and time-consuming. Thus, we propose a two-stage learning approach to improve the efficiency of network training. First, the network learns a Box+Path representation to enclose the input shape, capturing the sweeping path, basic geometry, and rotation angles. Then, we learn implicit sketches that refine the coarse box+path into a detailed shape representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 details the two-stage learning network. We begin with a 3D convolutional encoder that transforms the voxel input into a feature vector \\(\\mathbf{z}\\), which is then processed through a multilayer perceptron (MLP), yielding matrices \\(\\mathbf{P}\\) and \\(\\mathbf{B}\\). Here, \\(\\mathbf{P}\\) are the control points of the Bezier curves representing the sweeping paths (Fig. 3 (b)), and \\(\\mathbf{B}\\) encapsulates the box+path parameters (Fig. 3 (c)), including length, width, and sweep's twist angles. Once the sweeping paths are derived from Stage 1, we focus on learning the geometric details surrounding each path. The feature \\(\\mathbf{z}\\) is then mapped to \\(N_{p}\\) local features via another MLP. These vectors are subsequently concatenated with test points and fed into an implicit sketch network. The output from this network includes two occupancy values that represent the top and bottom sketches (Fig. 3 (d)), respectively. We then obtain the final CAD reconstruction by sweeping or lofting the detected sketches along the paths (Fig. 3 (e))." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.873, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.897, + 0.373 + ], + "angle": 0, + "content": "Figure 3. Network architecture for SfmCAD: Stage 1 encodes the input voxels into feature \\(\\mathbf{z}\\) via a 3D convolutional encoder. An MLP then processes \\(\\mathbf{z}\\) to generate the control point coordinates \\(\\mathbf{P}_1, \\ldots, \\mathbf{P}_4\\) for each Bézier curve, as well as the dimensions \\(l, w\\) and twist angles \\(\\delta^u, \\delta^b\\) for each Box proxy \\(\\mathcal{B}\\). In Stage 2, another MLP transforms feature \\(\\mathbf{z}\\) into local features \\(\\mathbf{z}'\\), which are then concatenated with the testing point coordinate and fed into the implicit sketch network to predict the sketch SDF value. After extending the sketches along the paths, we combine all parts for the final reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.394, + 0.201, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.491, + 0.165, + 0.501 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.395, + 0.323, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.491, + 0.316, + 0.501 + ], + "angle": 0, + "content": "Feature primitives" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.395, + 0.443, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.491, + 0.419, + 0.5 + ], + "angle": 0, + "content": "Box+Path" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.563 + ], + "angle": 0, + "content": "Figure 4. 2D illustration of our Box+Path representation (right) to approximate the input shape (left), which can be decomposed into a set of feature primitives (middle, the green arrows indicate the sweeping paths)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.214, + 0.59 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Here, we first present the box+path representation, which can be efficiently learned to approximate the input shape by stacking a set of box-like proxies along the sweeping path. Then, we define the neural sketch+path to refine the coarse reconstruction. Finally, we introduce the two-stage learning network for inferring this coarse-to-fine reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.312, + 0.713 + ], + "angle": 0, + "content": "4.1. Box+Path Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.871 + ], + "angle": 0, + "content": "Previous unsupervised methods [23, 30, 31] for computing the Signed Distance Function (SDF) for geometric primitives (e.g., cubes, spheres, cylinders) rely on linear transformations that convert global testing point coordinates into local ones. However, the inherent non-linearity of curved paths limits the direct application of linear methods. We introduce a linear approximation method for the curved SDF computation, which we term Box+Path representation. We will first discuss the Box-Path in a 2D context to provide a more intuitive understanding before expanding it to 3D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Box+Path in 2D. Specifically, we approximate one feature primitive \\(\\mathcal{P}\\) using a discrete set of \\(N_{s}\\) rigid box proxies," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.396, + 0.892, + 0.428 + ], + "angle": 0, + "content": "i.e., \\(\\mathcal{P} = \\{\\mathcal{B}_i\\}_{i=1}^{N_s}\\). The center position \\(\\mathbf{c}_i\\) of each box \\(\\mathcal{B}_i\\) is obtained by uniformly sampling the Bezier curve \\(\\Upsilon(t)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.436, + 0.892, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i} = \\Upsilon (i / N _ {s}), \\quad i = 1, \\dots , N _ {s} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.491 + ], + "angle": 0, + "content": "The height \\( h_i \\) and the normal vector (or heading direction) \\( \\mathbf{n}_i \\) of each box are calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.498, + 0.892, + 0.529 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\left| \\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i} \\right|, \\mathbf {n} _ {i} = \\frac {\\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i}}{\\left| \\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i} \\right|} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.565 + ], + "angle": 0, + "content": "By incorporating a learnable width parameter \\(w\\), the 2D Box+Path can be fully characterized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Box+Path in 3D. When extended to 3D space, we introduce the tangent vector \\(\\mathbf{t}_i\\) and the binormal vector \\(\\mathbf{b}_i\\) to parameterize the orientation of the \\(i\\)-th box proxy. To allow the sketch to rotate by a prescribed angle along the sweeping path, we should learn rotation angles \\(\\delta^u\\) and \\(\\delta^b\\) at the start and end points of the sweeping path. Then, the intermediate rotation angles \\(\\delta_i\\) are obtained by linear interpolation along the sweeping path. Next, we compute the tangent vector \\(\\mathbf{t}_i\\) as the cross product of the normal \\(\\mathbf{n}_i\\) and the positive \\(z\\)-axis unit vector \\(\\mathbf{e}_z\\). We then rotate \\(\\mathbf{t}_i\\) around \\(\\mathbf{n}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.725, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {i} = R _ {\\mathbf {n} _ {i}} \\left(\\mathbf {n} _ {i} \\times \\mathbf {e} _ {z}, \\delta_ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.894, + 0.84 + ], + "angle": 0, + "content": "where \\( R_{\\mathbf{n}_i}(\\mathbf{v}, \\delta_i) \\) represents a counter-clockwise rotation of vector \\( \\mathbf{v} \\) with angle \\( \\delta_i \\) around \\( \\mathbf{n}_i \\). The binormal vector \\( \\mathbf{b}_i \\) is obtained as a cross product of \\( \\mathbf{n}_i \\times \\mathbf{t}_i \\). Altogether, the triplet \\( \\{\\mathbf{n}_i, \\mathbf{t}_i, \\mathbf{b}_i\\} \\) defines the orientation for each box along the curve path. The 3D Box+Path is fully characterized by incorporating learnable dimension parameters \\( l, w \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.693, + 0.862 + ], + "angle": 0, + "content": "4.2. Neural Sketch+Path" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present differentiable methods for lofting or sweeping 2D sketches into 3D solid shapes to accurately reconstruct" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4674" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "the feature primitives. The shape of each 2D sketch is defined by an implicit representation, where we use \\(\\hat{S}_{\\mathrm{sk}}\\) to represent the SDF of sketches. We now describe the differentiable operators that calculate the SDF of the 3D feature primitives corresponding to the sketches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.273 + ], + "angle": 0, + "content": "Differentiable Sketch-Extrude. In the context of generating the geometry of feature primitives from sketch and extrude operations, \\(\\hat{S}_{\\mathrm{extrude}}^i\\) denotes the \\(i\\)-th sketch-extrude operation, \\(\\mathbf{x}_{i_z}\\) represents the \\(z\\)-component of the \\(i\\)-th position vector, and \\(h_i\\) is the height of the \\(i\\)-th feature primitive. We introduce a set of intermediate functions \\(\\mathcal{M}_1, \\mathcal{M}_2\\), and \\(\\mathcal{M}_3\\) to facilitate a clearer formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.28, + 0.388, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {M} _ {1} \\left(\\hat {\\mathcal {S}} _ {\\mathbf {s k}} ^ {i}\\right) = \\max \\left(\\hat {\\mathcal {S}} _ {\\mathbf {s k}} ^ {i}, \\left| \\mathbf {x} _ {i _ {z}} \\right| - h _ {i}\\right), \\\\ \\mathcal {M} _ {2} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i}\\right) = \\max \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i}, 0\\right), \\\\ \\mathcal {M} _ {3} = \\max \\left(\\left| \\mathbf {x} _ {i _ {z}} \\right| - h _ {i}, 0\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.45, + 0.364 + ], + "angle": 0, + "content": "The SDF of the \\(i\\)-th extrusion primitive is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.373, + 0.469, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\hat {S} _ {\\text {e x t r u d e}} ^ {i} = \\min \\left(\\mathcal {M} _ {1} \\left(\\hat {S} _ {\\mathrm {s k}} ^ {i}\\right), 0\\right) + \\sqrt {\\mathcal {M} _ {2} \\left(\\hat {S} _ {\\mathrm {s k}} ^ {i}\\right) ^ {2} + \\mathcal {M} _ {3} ^ {2}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.468, + 0.48 + ], + "angle": 0, + "content": "Differentiable Sketch-Loft. For the loft operation, we interpolate between two distinct sketches, \\(\\hat{S}_{\\mathrm{sk}}^{u_i}\\) and \\(\\hat{S}_{\\mathrm{sk}}^{l_i}\\), at the upper and lower faces of the \\(i\\)-th feature primitive. Given the interpolation factor \\(\\alpha\\) as \\(|\\mathbf{x}_{i_z}| / h_i\\), the linearly interpolated sketch, \\(\\hat{S}_{\\mathrm{sk}}^{i_\\alpha}\\), is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.488, + 0.469, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}} = (1 - \\alpha) \\cdot \\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {l _ {i}} + \\alpha \\cdot \\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {u _ {i}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.469, + 0.562 + ], + "angle": 0, + "content": "We use the same intermediate variables as in the extrusion case, but with \\(\\hat{S}_{\\mathrm{sk}}^{i}\\) replaced by \\(\\hat{S}_{\\mathrm{sk}}^{i_{\\alpha}}\\). The SDF of the \\(i\\)-th loft primitive can be expressed similarly to Eq. 5:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.571, + 0.469, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {S}} _ {\\text {l o f t}} ^ {i} = \\min \\left(\\mathcal {M} _ {1} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}}\\right), 0\\right) + \\sqrt {\\mathcal {M} _ {2} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}}\\right) ^ {2} + \\mathcal {M} _ {3} ^ {2}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.468, + 0.678 + ], + "angle": 0, + "content": "Differentiable Sketch-Sweep. For sweep operation, we decompose it into a series of \\(N_{s}\\) extrusions. We denote the predicted SDF for the \\(j\\)-th extrude segment as \\(\\hat{S}_{\\mathrm{extrude}}^i\\). The SDF of the \\(i\\)-th sweep primitive is then obtained by taking the minimum of the SDFs of all extrusion segments:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.685, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {S}} _ {\\text {s w e e p}} ^ {i} = \\min _ {j \\in 1, \\dots , N _ {s}} \\hat {\\mathcal {S}} _ {\\text {e x t r u d e}} ^ {j}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.468, + 0.779 + ], + "angle": 0, + "content": "Transformation to Parametric CAD. Following [23], we transform implicit sketches into parametric ones. With the control points of Bezier curves outputted by the network, we can achieve the sketch+path parametric reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.334, + 0.803 + ], + "angle": 0, + "content": "4.3. Two-stage Network Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Stage 1: Box+Path learning. In this stage, the network predicts the control point coordinates \\(\\{\\mathbf{P}_1,\\dots ,\\mathbf{P}_4\\}\\) of the cubic Bezier curve and the parameters \\(\\{l_i,w_i,\\delta^u,\\delta^b\\}\\) describing the \\(i\\) -th Box \\(^+\\) Path \\(\\mathcal{B}_i\\) . These parameters subsequently inform the computation of the remaining parameters \\(\\{h_i,\\mathbf{n}_i,\\mathbf{t}_i,\\mathbf{b}_i\\}\\) , as detailed in Sect. 4.1. To align the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "path of \\(\\mathcal{B}\\) with the target shape path, we first compute the SDF, \\(\\hat{S}_{\\mathrm{box}}^{ij}\\), for each box at the \\(i\\)-th primitive and \\(j\\)-th segment, where \\(i \\in [1, N_p]\\) and \\(j \\in [1, N_s]\\), with \\(N_p\\) and \\(N_s\\) representing the total number of primitives and segments. Taking the union over the \\(\\hat{S}_{\\mathrm{box}}^{ij}\\) yields the SDF for the overall box+path \\(\\hat{S}_{\\mathcal{B}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.192, + 0.891, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {S}} _ {\\mathcal {B}} = \\bigcup_ {i = 1} ^ {N _ {p}} \\bigcup_ {j = 1} ^ {N _ {s}} \\hat {\\mathcal {S}} _ {\\text {b o x}} ^ {i j}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.892, + 0.348 + ], + "angle": 0, + "content": "where \\(\\bigcup (x) = \\mathrm{softmin}(\\mu \\cdot x)\\cdot x\\) and \\(\\mu\\) is a hyperparameter. We then apply a differentiable converter \\(\\Psi\\) similar to prior works [23, 31] to transform the \\(\\hat{S}_{\\mathcal{B}}\\) into occupancy, yielding \\(\\hat{\\mathcal{O}}_{\\mathcal{B}} = \\Psi (\\hat{S}_{\\mathcal{B}})\\), where \\(\\Psi (x) = (\\tanh (-\\beta \\cdot x) + 1) / 2\\), and \\(\\beta\\) is a hyperparameter. We then compute the Mean Squared Error (MSE) loss between \\(\\hat{\\mathcal{O}}_{\\mathcal{B}}\\) and the ground truth occupancy \\(\\tilde{\\mathcal{O}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.348, + 0.891, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {B}} = \\mathbb {E} _ {x \\in \\mathcal {X}} \\left[ | | \\hat {\\mathcal {O}} _ {\\mathcal {B}} - \\tilde {\\mathcal {O}} | | _ {2} ^ {2} \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.481 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}\\) is the set of all testing points. The exclusive use of \\(\\mathcal{L}_{\\mathcal{B}}\\) may lead to potential challenges: (1) the emergence of the curve with acute angles, and (2) the dimensions of \\(\\mathcal{B}\\) vastly outstrip the path length. Both increase the risk of shape self-intersections, which are considered invalid in CAD. To address these issues, we introduce two additional regularization terms," + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.488, + 0.891, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s m}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\min _ {j = 1, \\dots , N _ {s}} \\left(\\mathbf {n} _ {i, j} \\cdot \\mathbf {n} _ {i, j + 1}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.534, + 0.891, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l w}} = \\frac {1}{2 N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\operatorname {R e L U} \\left(h _ {i} + w _ {i} - 2 \\cdot \\Theta\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.657 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{sm}}\\) penalizes the sharpest curvature in the curve by maximizing the inner product between adjacent \\(\\mathbf{n}\\), and \\(\\mathcal{L}_{\\mathrm{lw}}\\) penalizes any length or width greater than a threshold \\(\\Theta\\). Combine with balance factors \\(\\lambda_{1}, \\lambda_{2}\\), the overall loss of Stage 1 is:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.666, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b o x}} = \\mathcal {L} _ {\\mathcal {B}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {s m}} + \\lambda_ {2}. \\mathcal {L} _ {\\mathrm {l w}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.858 + ], + "angle": 0, + "content": "Stage 2: Implicit sketch learning. Using the path learned in Stage 1, we progress towards fine reconstruction in Stage 2 by learning sketches. We initially utilize a Multi-Layer Perceptron (MLP) to convert the feature \\(\\mathbf{z}\\) into \\(N_{p}\\) local features \\(\\mathbf{z}^{\\prime}\\). These local features are then concatenated with the local \\(\\{x,y\\}\\) coordinates of each point and fed into an implicit sketch network, which is essentially an MLP predicting the SDF value \\(\\hat{S}_{\\mathrm{sk}}^{i}\\). For the sketch-loft operation, two implicit sketch networks are used to learn \\(\\hat{S}_{\\mathrm{sk}}^{u_i}\\) and \\(\\hat{S}_{\\mathrm{sk}}^{l_i}\\). Following the computation methods outlined in Sect. 4.2, we can derive the predicted overall occupancy value \\(\\hat{\\mathcal{O}}_{\\mathrm{F}}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.864, + 0.852, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {O}} _ {\\mathrm {F}} = \\left\\{ \\begin{array}{l l} \\Psi (\\bigcup_ {i = 1} ^ {N _ {p}} \\hat {S} _ {\\text {s w e e p}} ^ {i}), & \\text {f o r s w e e p o p e r a t i o n} \\\\ \\Psi (\\bigcup_ {i = 1} ^ {N _ {p}} \\hat {S} _ {\\text {l o f t}} ^ {i}), & \\text {f o r l o f t o p e r a t i o n .} \\end{array} \\right.\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "Finally, an MSE loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) is computed against the ground truth \\(\\mathcal{O}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.117, + 0.469, + 0.142 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e c}} = \\mathbb {E} _ {x \\in \\mathcal {X}} \\left[ \\left| \\left| \\hat {\\mathcal {O}} _ {\\mathrm {F}} - \\tilde {\\mathcal {O}} \\right| \\right| _ {2} ^ {2} \\right]. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.147, + 0.473, + 0.392 + ], + "angle": 0, + "content": "The direct computation of \\(\\mathcal{L}_{\\mathrm{rec}}\\) over the entire point set \\(\\mathcal{X}\\) requires the repetition of \\(\\mathcal{X}\\) corresponding to \\(\\hat{S}_{\\mathrm{extrude}}\\) for \\(N_{p} \\times N_{s}\\) times, which is time-consuming. To accelerate the training of the implicit sketch network, we propose a real-time sampling strategy, exploiting the sketch consistency trait inherent in the sweep operation. Specifically, we sample grid-like testing points \\(y_{i}\\) in the base sketch plane of the \\(i\\)-th path \\(\\Upsilon_{i}\\) obtained in the first stage. The occupancy value of \\(y_{i}\\) is determined by inputting it into the implicit sketch network and applying the converter \\(\\Psi\\). Subsequently, \\(y_{i}\\) is moved along the curve for \\(N_{s} - 1\\) times, with its occupancy value directly replicated, yielding \\(\\hat{\\mathcal{O}}_{\\mathrm{Fi}}^{*}\\). The ground truth occupancy value \\(\\tilde{\\mathcal{O}}_i^*\\) is provided by the input voxel. After performing this procedure for all paths and conducting a union operation, the loss function corresponding to real-time sampling is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.401, + 0.469, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e c}} ^ {*} = \\mathbb {E} _ {y \\in \\mathcal {Y}} \\left[ | | \\hat {\\mathcal {O}} _ {\\mathrm {F}} ^ {*} - \\tilde {\\mathcal {O}} ^ {*} | | _ {2} ^ {2} \\right], \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.482 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}\\) represents the set of testing points obtained through real-time sampling, and \\(*\\) denotes computation using real-time sampling data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Note that the SfmCAD is designed to choose between learning either loft or sweep operations, which is determined by its hyperparameters. However, enabling the network to concurrently learn both operations is also viable, with details provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.571, + 0.295, + 0.587 + ], + "angle": 0, + "content": "5. Results and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.469, + 0.658 + ], + "angle": 0, + "content": "We evaluate the performance of SfmCAD on two reconstruction tasks: (1) holistic shapes of different categories and (2) semantically segmented shapes. We show the output editability of SfmCAD in the supplemental materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.159, + 0.683 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Implementation details. We implemented SfmCAD in PyTorch by using an NVIDIA® TITAN RTX GPU. The Adam optimizer [15] is utilized with a learning rate 1e-4 and beta values of (0.5, 0.99). The hyperparameters in Sect. 4.3 are set to: \\(\\mu = 20\\), \\(\\beta = 50\\), \\(\\Theta = 0.1\\), \\(\\lambda_1 = 0.05\\), \\(\\lambda_2 = 0.05\\). In our evaluations, we employ the best-performing operation (Loft/Sweep) for each category of holistic shapes and every part of segmented shapes. SfmCAD trains for 500 epochs per stage with a batch size of 24, followed by fine-tuning each test shape for 200 iterations per stage. All baseline methods are trained for 1,000 epochs and fine-tuned for 400 iterations for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. We utilize three common metrics for 3D shape reconstruction: Chamfer Distance (CD), Normal" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.889, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.384, + 0.832, + 0.399 + ], + "angle": 0, + "content": "Figure 5. Visual comparison on ABC dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.546, + 0.414, + 0.846, + 0.429 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on ABC dataset." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.439, + 0.853, + 0.539 + ], + "angle": 0, + "content": "
MethodsCD↓ECD↓NC↑
UCSG-Net [14]1.23321.7860.866
CSG-Stump [30]0.6717.7510.892
ExtrudeNet [31]0.5197.1110.885
SECAD-Net [23]0.5067.2860.884
Ours0.3955.0380.919
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Consistency (NC), and Edge Chamfer Distance (ECD). All experiments generate the reconstructed shapes via Marching Cubes at a resolution of \\(256^3\\). To calculate CD, we sample 8,192 points from the ground truth and predicted shape surfaces. For ECD, we employ a threshold of 0.1 for normal cross-products to identify points close to the edges on the ground truth surface and 0.5 for the predicted surface, considering the lack of sharpness in shapes generated by Marching Cubes. For better visual clarity, we magnify the CD and ECD by factors of \\(10^3\\) and \\(10^2\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.81, + 0.743 + ], + "angle": 0, + "content": "5.2. Reconstruction from Holistic Shape" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.84 + ], + "angle": 0, + "content": "We compare SfmCAD with UCSG-Net [14], CSG-Stump [30], ExtrudeNet [31] and SECAD-Net [23]. We test these methods on holistic shapes from different sources, including CAD parts in ABC [16], ShapeNet objects [1], and a synthetic tree dataset. The comparison against CAPRI-Net [49] is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on ABC dataset. We employ the voxel grids and sampling point data provided by [49]. The dataset included 5,000 groups of data for training and 1,000 for testing. We select 100 shape subsets from the test set for eval" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.395, + 0.424, + 0.41 + ], + "angle": 0, + "content": "Figure 6. Visual comparison on ShapeNet dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.425, + 0.436, + 0.44 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on ShapeNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.451, + 0.428, + 0.55 + ], + "angle": 0, + "content": "
MethodsCD↓ECD↓NC↑
UCSG-Net [14]2.11123.6160.815
CSG-Stump [30]2.16817.5920.818
ExtrudeNet [31]0.91014.1100.849
SECAD-Net [23]0.83614.8180.837
Ours0.62614.0960.867
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.565, + 0.421, + 0.579 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on Tree dataset." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.59, + 0.47, + 0.645 + ], + "angle": 0, + "content": "
MethodsExtrudeNet [31]SECAD-Net [23]Ours
CD↓7.2422.1970.686
NC↑0.6080.6220.722
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.702 + ], + "angle": 0, + "content": "uation. The visual results are shown in Fig. 5, while the quantitative results are presented in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Evaluation on ShapeNet dataset. We utilize three categories from ShapeNet: chair, table, and display. The voxel grids and sampling points are provided by [3]. We selected 120 shapes for evaluation, with 40 from each category. Visual and quantitative results are in Fig. 6 and Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation on tree branches. We use a procedural modeling approach of Palubicki et al. [26] to generate tree meshes voxelized into a resolution of \\(64^{3}\\). We adopt the method from [2] to obtain sampling points and occupancy values. This process results in 5,000 data groups, of which 4,500 are used for training and 500 for testing. We train all methods using increasingly denser voxel resolutions \\((16^{3},32^{3},64^{3})\\) and select 50 trees for numerical evaluation." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.309, + 0.831, + 0.324 + ], + "angle": 0, + "content": "Figure 7. Visual comparison on Tree dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.412 + ], + "angle": 0, + "content": "tion. Note given the cylindrical nature of tree branches, we adapt our Box+Path representation by substituting boxes with cylinders for the final reconstruction (see Fig. 7 and Tab. 3 for visual and quantitative results)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.893, + 0.565 + ], + "angle": 0, + "content": "Summary of comparative analysis. As demonstrated in Tables 1, 2, and 3, SfmCAD consistently outperforms other methods in terms of CD, ECD, and NC, suggesting that the sketch+path representation of SfmCAD offers more flexible shape expression capabilities than the geometric primitives used by UCSG-Net and CSG-Stump, as well as the simple sketch+extrusion operations employed by ExtrudeNet and SECAD. Figures 5, 6, and 7 further show SfmCAD's ability to generate detailed reconstructions (such as thin structures) across diverse shape categories." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.574, + 0.836, + 0.591 + ], + "angle": 0, + "content": "5.3. Reconstruction from Segmented Shape" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Semantic segmentation provides information for reconstructing complex shapes. Previous CAD reconstruction methods [23, 31] could not utilize this information because their single-type primitives lack the flexibility to represent each part independently. In contrast, SfmCAD successfully reconstructs multiple primitives with its sketch+path representation capability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.885 + ], + "angle": 0, + "content": "We show the task of unsupervised CAD reconstruction utilizing semantic segmentation. The evaluation is conducted on PartNet dataset [48], which provides instance-level 3D part information. We utilize three categories: chair, table, and trashcan. We voxelize the segmented parts and sampled occupancy values at resolutions \\(16^{3}, 32^{3}, 64^{3}\\) following the scheme by [2]. This process yields a total of 12,000 chair parts (1,229 complete), 12,000 table parts (1,491 complete), and 3,000 trashcan parts (234 complete). The train-test split for all categories was set as 5:1. We selected 90 complete shapes, including 40 tables, 40 chairs, and 10 trashcans, for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "All methods were trained on part shapes using increas" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4677" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.316, + 0.417, + 0.331 + ], + "angle": 0, + "content": "Figure 8. Visual comparison on PartNet dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.344, + 0.431, + 0.358 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on PartNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.37, + 0.428, + 0.438 + ], + "angle": 0, + "content": "
MethodsCD↓ECD↓NC↑
ExtrudeNet [31]3.2510.5490.783
SECAD-Net [23]1.8880.4720.818
Ours1.4100.4460.814
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.45, + 0.455, + 0.465 + ], + "angle": 0, + "content": "Table 5. Ablation study on sweep twist angles and loss terms." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.476, + 0.469, + 0.541 + ], + "angle": 0, + "content": "
Settings- (δu, δl)-δl- Llw- LsmOurs
CD↓4.0923.3674.8043.2683.142
ECD↓0.1730.1390.1640.1540.133
NC↑0.6550.7140.6850.7320.776
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.469, + 0.642 + ], + "angle": 0, + "content": "ingly denser voxel resolutions \\((16^{3}, 32^{3}, 64^{3})\\). For each part, we set ExtrudeNet [31] and SECAD-Net [23] to output an extruded cylinder and SfmCAD to output a sketch+path shape. After reconstructing all parts, they were reassembled to obtain the final CAD models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.469, + 0.779 + ], + "angle": 0, + "content": "The quantitative results in Tab. 4 indicate that our method outperforms others. Compared to the closest competitor, SECAD-Net, SfmCAD improves the CD by \\(25.3\\%\\) and the ECD by \\(5.5\\%\\). Fig. 8 illustrates that the sketch+extrusion reconstruction of ExtrudeNet and SECAD can only characterize the nonlinear shapes through sketches. In contrast, SfmCAD can model these complex shapes using sketches and free-form paths, enabling a more refined reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.231, + 0.803 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We conducted an ablation study by using 40 box+path reconstructions on chair class from PartNet to evaluate the efficiency of our network design and loss terms. Initially, we disable the two sweep twist angles \\(\\delta^u\\), \\(\\delta^l\\) or \\(\\delta^l\\). Then we separately discarded \\(\\mathcal{L}_{\\mathrm{sm}}\\) and \\(\\mathcal{L}_{\\mathrm{lw}}\\). The visual results in Fig. 9 and the quantitative results in Tab. 5 suggest that" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.088, + 0.885, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.312, + 0.893, + 0.356 + ], + "angle": 0, + "content": "Figure 9. Visual comparison between reconstruction results for ablation study. From top to bottom, each column shows the predicted 3D path, vector \\(\\{\\mathbf{n},\\mathbf{t},\\mathbf{b}\\}\\), and the box+path representation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.892, + 0.419 + ], + "angle": 0, + "content": "removing any of these settings negatively affects the reconstruction performance of SfmCAD." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.442, + 0.771, + 0.459 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.637 + ], + "angle": 0, + "content": "We have presented an unsupervised neural network for CAD reconstruction that infers high-level CAD modeling operations. We have also introduced a unique sketch+path representation that encapsulates the local geometric details and overall structural information of the 3D shape. A novel two-stage learning strategy, operating in a coarse-to-fine manner, was proposed to address the complexity of simultaneous sketch and path learning, thus improving the efficiency of network training. We evaluated SfmCAD using several different categories of objects and demonstrated its advantages by comparing it to state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.894, + 0.717 + ], + "angle": 0, + "content": "In future work, we plan to explore using 2D sketches and 3D part templates to enhance the efficiency of unsupervised CAD reconstruction. Besides, given the manipulable property of SfmCAD's outputs, extending this work for generative CAD design or modeling would be meaningful." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments. We thank the anonymous reviewers for their valuable suggestions. This work is partially funded by the Strategic Priority Research Program of the Chinese Academy of Sciences (XDB0640000), Guangdong Science and Technology Program (2023B1515120026), National Natural Science Foundation of China (62172416, 62172415, U22B2034, 62262043, 62365014), and Youth Innovation Promotion Association of the Chinese Academy of Sciences (2022131). This work was supported by NRCS grant #NR233A750004G044 to Benes. The findings and conclusions should not be construed to represent any agency determination or policy." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4678" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5939-5948, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-Net: Generating compact meshes via binary space partitioning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 45–54, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[4] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. In ACM SIGGRAPH 2004 Papers, pages 905–914. 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[5] Tao Du, Jeevana Priya Inala, Yewen Pu, Andrew Spielberg, Adriana Schulz, Daniela Rus, Armando Solar-Lezama, and Wojciech Matusik. Inversecsg: Automatic conversion of 3d models to csg trees. ACM Trans. Graph., 37(6):1-16, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[6] Elona Dupont, Kseniya Cherenkova, Anis Kacem, Sk Aziz Ali, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Cadops-net: Jointly learning cad operation types and steps from boundary-representations. In International Conference on 3D Vision (3DV), pages 114-123. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[7] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems (NeurIPS), pages 5885–5897, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[8] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Trans. Graph., 39(6):1-16, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[9] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: CAD reconstruction by B-rep chain complex generation. ACM Trans. Graph., 41(4): 1-18, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[10] Rostislav Hulik, Michal Spanel, Pavel Smrz, and Zdenek Materna. Continuous plane detection in point-cloud data based on 3d hough transform. Journal of visual communication and image representation, 25(1):86-97, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[11] Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G Lambourne, Karl DD Willis, Thomas Davies, Hooman Shayani, and Nigel Morris. Uv-net: Learning from boundary representations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11703-11712, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[12] Pradeep Kumar Jayaraman, Joseph G Lambourne, Nishkrit Desai, Karl DD Willis, Aditya Sanghi, and Nigel JW Morris. Solidgen: An autoregressive model for direct b-rep synthesis. arXiv preprint arXiv:2203.13944, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[13] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Kacper Kania, Maciej Zieba, and Tomasz Kajdanowicz. UCSG-NET-Unsupervised discovering of constructive solid" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "geometry tree. In Advances in Neural Information Processing Systems (NeurIPS), pages 8776-8786, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[16] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 9601-9611, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[17] Joseph G Lambourne, Karl DD Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. BRepNet: A topological message passing system for solid models. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 12773-12782, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Joseph George Lambourne, Karl Willis, Pradeep Kumar Jayaraman, Longfei Zhang, Aditya Sanghi, and Kamal Rahimi Malekshan. Reconstructing editable prismatic cad from rounded voxel models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[19] Eric-Tuan Lé, Minhyuk Sung, Duygu Ceylan, Radomir Mech, Tamy Boubekeur, and Niloy J Mitra. CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In IEEE International Conference on Computer Vision (ICCV), pages 7457-7466, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Bosheng Li, Jacek Kałuzny, Jonathan Klein, Dominik L. Michels, Wojtek Pałubicki, Bedrich Benes, and Soren Pirk. Learning to reconstruct botanical trees from single images. ACM Trans. Grap., 40(6):1-15, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[21] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Sketch2cad: Sequential cad modeling by sketching in context. ACM Trans. Graph., 39(6):1-14, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[22] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3d point clouds. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 2652-2660, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[23] Pu Li, Jianwei Guo, Xiaopeng Zhang, and Dong-Ming Yan. Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 16816-16826, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Yuanqi Li, Shun Liu, Xinran Yang, Jianwei Guo, Jie Guo, and Yanwen Guo. Surface and edge detection for primitive fitting of point clouds. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 4460-4470, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Wojciech Palubicki, Kipp Horel, Steven Longay, Adam Runions, Brendan Lane, Radomír Měch, and Przemyslaw Prusinkiewicz. Self-organizing tree models for image synthesis. ACM Trans. Graph., 28(3):1–10, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. In Advances in Neural Information Processing Systems (NeurIPS), pages 5077-5088, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4679" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 165-174, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 10344-10353, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[30] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, et al. CSG-stump: A learning friendly CSG-like representation for interpretable shape parsing. In IEEE International Conference on Computer Vision (ICCV), pages 12478-12487, 2021. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[31] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude for shape parsing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 482-498, 2022. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.389, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[32] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient RANSAC for point-cloud shape detection. In Comput. Graph. Forum, pages 214-226, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[34] Jami J Shah and Martti Mäntylä. Parametric and feature-based CAD/CAM: concepts, techniques, and applications. John Wiley & Sons, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[35] Vadim Shapiro. Solid modeling. Handbook of computer aided geometric design, 20:473-518, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[36] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5515-5523, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[37] Gopal Sharma, Difan Liu, Subhransu Maji, Evangelos Kalogerakis, Siddhartha Chaudhuri, and Radomir Mech. ParSeNet: A parametric surface fitting network for 3d point clouds. In European Conference on Computer Vision (ECCV), pages 261-276, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[38] Mikaela Angelina Uy, Yen-Yu Chang, Minhyuk Sung, Purvi Goel, Joseph G Lambourne, Tolga Birdal, and Leonidas J Guibas. Point2Cyl: Reverse engineering 3D objects from point clouds to extrusion cylinders. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11850-11860, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[39] Tamas Varady and Ralph Martin. Reverse engineering. In Handbook of Computer Aided Geometric Design, pages 651-681. 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[40] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. PIE-Net: Parametric inference of point cloud edges. Advances in Neural Information Processing Systems (NeurIPS), pages 20167–20178, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Karl DD Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G Lambourne, Armando Solar-Lezama, and Wo" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "jciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM Trans. Graph., 40(4):1-24, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In IEEE International Conference on Computer Vision (ICCV), pages 6772–6782, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[43] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl DD Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 6062-6070, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[44] Xiang Xu, Karl DD Willis, Joseph G Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, and Yasutaka Furukawa. Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In International Conference on Machine Learning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[45] Fujio Yamaguchi. Curves and surfaces in computer aided geometric design. Springer Science & Business Media, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[46] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[47] Siming Yan, Zhenpei Yang, Chongyang Ma, Haibin Huang, Etienne Vouga, and Qixing Huang. Hpnet: Deep primitive segmentation using hybrid representations. In IEEE International Conference on Computer Vision (ICCV), pages 2753-2762, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[48] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, page to appear, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[49] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11768–11778, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[50] Fenggen Yu, Qimin Chen, Maham Tanveer, Ali Mahdavi Amiri, and Hao Zhang. \\(\\mathrm{D}^2\\mathrm{csg}\\): Unsupervised learning of compact csg trees with dual complements and dropouts. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4680" + } + ] +] \ No newline at end of file diff --git a/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_origin.pdf b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc64497e92fcd3779225027bc2fa15896495b7e1 --- /dev/null +++ b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/a4171579-c596-406f-93c6-62c45be3925d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de824133d1f30b54fd3b6b28fb9f84d54ab2198484a8b712df7b9898b6d79566 +size 10436759 diff --git a/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/full.md b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b56e3164039798fa01e38f84792a3c2bbd432b0f --- /dev/null +++ b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/full.md @@ -0,0 +1,370 @@ +# SfmCAD: Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations + +$\mathrm{PuLi^{1,2}}$ Jianwei Guo $^{1,2*}$ Huibin Li $^{2,1}$ Bedrich Benes $^{3}$ Dong-Ming Yan $^{1,2}$ + +$^{1}$ MAIS, Institute of Automation, Chinese Academy of Sciences + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences $^{3}$ Computer Science, Purdue University + +![](images/ce304b141a898683556788c5ffe5b1e21efcae42ffc29deb5e85e3bf8cbc3ca1.jpg) +Figure 1. SfmCAD parses voxel shapes and reconstructs them into a parameterized set of sketch-based feature modeling operations: the input voxels, neural typed sketch+path representation, and the final CAD model (each part is randomly colored). + +# Abstract + +This paper introduces SfmCAD, a novel unsupervised network that reconstructs 3D shapes by learning the Sketch-based Feature Modeling operations commonly used in modern CAD workflows. Given a 3D shape represented as voxels, SfmCAD learns a neural-typed sketch+path parameterized representation, including 2D sketches of feature primitives and their 3D sweeping paths without supervision, for inferring feature-based CAD programs. SfmCAD employs 2D sketches for local detail representation and 3D paths to capture the overall structure, achieving a clear separation between shape details and structure. This conversion into parametric forms enables users to seamlessly adjust the shape's geometric and structural features, thus enhancing interpretability and user control. We demonstrate the effectiveness of our method by applying SfmCAD to many different types of objects, such as CAD parts, ShapeNet objects, and tree shapes. Extensive comparisons show that SfmCAD produces compact and faithful 3D reconstructions with superior quality compared to alternatives. The code is released at https://github.com/BunnySoCrazy/SfmCAD. + +# 1. Introduction + +Reconstructing shapes into CAD representations is a fundamental problem in computer vision and graphics with numerous commercial applications in industrial design and manufacturing [39]. Recent years have seen an increased interest in leveraging deep learning techniques to learn geometric representations of 3D shapes, including implicit fields [2, 25, 28], geometric primitives [19, 22, 24, 37], boundary representation (B-Rep) [9, 17], or Constructive Solid Geometry (CSG) [3, 36, 49], each with its unique characteristics and applications, contributing to the rich tapestry of research and development in this field. + +However, current CAD reconstruction methods are still lacking detail reconstruction accuracy, compactness, and interpretability of the obtained models. For example, a typical line of research is dedicated to reconstruction-oriented tasks, i.e., representing 3D shapes as parametric primitives, implicit fields, or B-rep. These methods deliver a high level of detail and accuracy by capturing intricate details and complex geometries with remarkable precision. However, constructing CAD models by assembling instances of primitives or by identifying faces, edges, and vertices is a tedious process [35]. Besides, the results often include complex and unintuitive operations that make the model difficult to inter- + +pret and edit. This lack of control poses significant challenges in scenarios where user interaction and shape manipulation are required. + +A straightforward approach to CAD reconstruction is to learn CSG representation [14, 30, 36, 49] that implicitly recovers the CAD design history. A CSG 3D shape is expressed as a tree with leaves consisting of volumetric primitives (e.g., cubes, spheres, and cylinders) that are combined using Boolean operations in the CSG inner nodes. Despite the sought-after attribute of compact shape representation, these methods often fall short regarding reconstruction accuracy by using only a set of simple typed primitives. + +We introduce SfmCAD, a novel neural model for reconstructing 3D shapes into their CAD representation. Compared to related work, SfmCAD improves the reconstruction accuracy and the level of model editability by decoupling 3D structures and local geometric details of the shape. SfmCAD is inspired by feature-based modeling, a predominant method for creating 3D models in modern CAD systems, which sequentially add features (e.g., holes, slots, bosses) that represent common manufacturing operations [6, 41, 42]. Specifically, we employ unsupervised learning to retrieve a neural typed sketch+path representation of 3D shapes. Taking 3D voxel data as input, SfmCAD learns implicit fields that embody the sketch profile and generates control points for a Bezier curve, constituting the sweeping path for the component's geometry. The basic extrusion operation has been learned in a supervised [18, 38] or unsupervised [23, 31] manner. However, our sketch+path representation supports not only extrude but also sweep, loft, and revolve operations, thus enhancing the shape representation capabilities, as demonstrated in Fig. 1. + +At the core of the SfmCAD is a two-stage learning strategy that operates in a coarse-to-fine manner to tackle the inherent complexity and time-intensiveness of simultaneous sketch profile and sweeping path learning. Initially, the network is trained to learn a coarse Box+Path representation of the shape, enabling a quick grasp of the shape's path and broad structure using box-like geometric proxies. Then, an implicit network delves deeper into the geometric details of the shape, enhancing the reconstruction accuracy. We demonstrate the capability and versatility of our method by applying SfmCAD to many different types of objects, including CAD parts [16], ShapeNet objects [1], tree branches [20], and by comparing to previous works. We claim the following contributions: + +1. We propose SfmCAD, a novel neural approach that parses 3D shapes as a set of industry-standard sketch-based CAD modeling operations. To the best of our knowledge, SfmCAD is the first unsupervised and universal neural network for learning the common CAD commands, including extrude, sweep, loft, and revolve. +2. We introduce novel decoupling of 3D structures and + +local geometric details, which are represented by 3D Bézier curves and implicit 2D sketches. This representation ensures accurate reconstruction and ease of editing. + +3. We propose a two-stage learning strategy, namely 3D Box+Path learning and 2D implicit sketch learning, which significantly accelerates the training speed in shape parser learning. + +# 2. Related work + +Geometric primitive extraction. Geometric primitive is a commonly used representation for approximating and abstracting 3D shapes [13]. Traditional methods based on RANSAC [32], Hough transform [10], or variational optimization [4, 46] have been used to detect and fit primitives on point clouds or polygonal meshes, but they often require careful parameter tuning for each shape. + +To overcome the overwhelming complexity of traditional methods, deep neural networks addressed the primitive segmentation/detection from point clouds [19, 47], they fit parametric surfaces [22, 29, 37], or detect parametric curves and sharp corners [9, 24, 40] to achieve compact CAD reconstructions. Although achieving remarkable accuracy on reconstruction tasks, these methods only output individual primitives with limited types, which restricts their capability for reconstructing complex and more general 3D shapes. Learning CSG-based reconstruction. CSG reconstruction embodies a 3D shape of high complexity and nonconvexity with CSG boolean operations to represent the 3D shape construction process [5]. CSGNet [36] first develops a neural model that parses a shape into a sequence of CSG operations. More recent works follow the line of CSG parsing by advancing the inference without any supervision [14] or improving representation capability with a three-layer reformulation of the classic CSG-tree [30], or handling richer geometric and topological variations by introducing quadric surface primitives [49, 50]. However, CSG reconstruction tends to combine a large number of shape primitives that limit the user editing capabilities. Moreover, using only basic primitives (e.g., boxes, spheres) is insufficient for approximating complicated shapes, thus usually limiting the reconstruction accuracy of small details [50]. + +Learning feature-based reconstruction. Modern CAD workflows use feature-based modeling that decomposes a target shape into a sequence of CAD commands [34], such as drawing a sketch followed by CAD operations such as extrusion, loft, etc. This paradigm makes modeling more efficient and in tune with how designers and engineers work. Recent work has explored the potential of deep learning to generate 2D engineering sketches [7, 27, 33], or directly learn B-Rep CAD models [6, 11, 12, 17]. Several approaches propose generative models for CAD design, predicting a sequence of CAD modeling operations to produce editable CAD models, such as DeepCAD [42], Fu + +![](images/b6d6e1b6a26f7ccf15ebac890899b83926979f626c2bfbc86d5b9a25fc5388a3.jpg) +Extrude +Figure 2. Sketch-based feature modeling operators supported in SfmCAD. The resulting 3D shape of each operation is called a feature primitive. + +![](images/9b7a09a24e9599bc32258379f4261cfd9cc38fc157f7429a03f39757c3441ebb.jpg) +Loft + +![](images/8f2ff0c5fc413f576a8ac896823b0543c66e68606a78f1172668b4e5c91dac37.jpg) +Sweep + +![](images/6cd51c599307985d3e078f6b310e5835392e30f5b3ba6ed7dd2fd441302b9325.jpg) +Revolve + +sion360 [41], Zone graph [43], and SkexGen [44]. These works focus on parametric CAD generation and do not address the reconstruction task of CAD models with suitable geometric loss functions. + +Most closely related to our work are the approaches that learn sketch+extrude operation (e.g., Point2Cyl [38], ExtrudeNet [31], SECAD-Net [23]), in which a collection of extrusions are predicted to build CAD models. While expressive within their scopes, these methods are limited to only one type of CAD operation (i.e., extrude) to simplify the search space. SfmCAD includes loft, revolve, and sweep reconstructions, broadening the shape representation capabilities. This approach enables us to test on a broader range of shapes beyond CAD. + +# 3. Problem Statement and Overview + +Sketch-based feature modeling creates 2D engineering sketches and then lifts them to 3D along a path using CAD operations [8, 21]. Here, we first introduce this typed sketch+path representation for sketch-based modeling and then give an overview of its generation from a voxel model. + +# 3.1.Typed Sketch+Path Modeling + +We define a profile as a closed curve, and its enclosed region is composed of one or multiple inner/outer loops. We define a sketch as the collection of one profile and its loops in a 2D plane. Let $Z$ denote the sketch and $\mathcal{C}$ be a 3D path to extend the sketches to create a 3D solid shape. We define four principal operations that are commonly used in modern CAD workflow (see Fig. 2): Extrude extends $Z$ along a linear path, Revolve produces a model by rotating $Z$ around an axis, Sweep moves $Z$ along a specified path to generate generalized cylinder shapes, and Loft allows for variations in the sketch, interpolating between multiple distinct sketches $\{Z_1, Z_2, \ldots, Z_n\}$ along the shortest distances between them to create complex shapes. + +Considering the similarity between the above operations, i.e., lifting 2D sketches to a 3D shape along a 3D path, we use a general typed sketch $^+$ path to represent all of these operations, where the term 'typed' indicates the specific type of operation. Without loss of generality, we call the 3D + +path a sweeping path and call the lofted or swept 3D shape a feature primitive. We further use a Bézier curve $\Upsilon(t)$ to represent the sweeping path: + +$$ +\mathcal {C} = \Upsilon (t) = \sum_ {i = 0} ^ {n} \binom {n} {i} (1 - t) ^ {n - i} t ^ {i} \mathbf {P} _ {i}, \quad t \in [ 0, 1 ] \tag {1} +$$ + +where $\mathbf{P}_i$ is a control point. The Bézier cubic is expressive enough to approximate the shapes shown in the paper, but higher-order curves can be obtained by simply changing the number of control points output by the network. + +Moreover, it is noteworthy that the Extrude and Revolve operations can be regarded as general variants of the sweep operation because a Bézier curve becomes a linear path (for Extrude) if all the control points are collinear and an $n$ -piece (usually $n = 4$ ) cubic Bézier curve can approximate a circular path (for Revolve) [45]. In the following, we focus on reconstructing Sweep and Loft operations, where Sweep contains Extrude and Revolve, and the Loft is restricted to contain only two sketches with a linear path between them. + +# 3.2. Overview + +Given a 3D voxel shape, we aim to reconstruct the CAD model by neurally decomposing the target shape into a sequence of typed sketch+path representation, which can be executed by the corresponding CAD commands to assemble a compact CAD model. Learning the path and sketch of a shape simultaneously is challenging and time-consuming. Thus, we propose a two-stage learning approach to improve the efficiency of network training. First, the network learns a Box+Path representation to enclose the input shape, capturing the sweeping path, basic geometry, and rotation angles. Then, we learn implicit sketches that refine the coarse box+path into a detailed shape representation. + +Fig. 3 details the two-stage learning network. We begin with a 3D convolutional encoder that transforms the voxel input into a feature vector $\mathbf{z}$ , which is then processed through a multilayer perceptron (MLP), yielding matrices $\mathbf{P}$ and $\mathbf{B}$ . Here, $\mathbf{P}$ are the control points of the Bezier curves representing the sweeping paths (Fig. 3 (b)), and $\mathbf{B}$ encapsulates the box+path parameters (Fig. 3 (c)), including length, width, and sweep's twist angles. Once the sweeping paths are derived from Stage 1, we focus on learning the geometric details surrounding each path. The feature $\mathbf{z}$ is then mapped to $N_{p}$ local features via another MLP. These vectors are subsequently concatenated with test points and fed into an implicit sketch network. The output from this network includes two occupancy values that represent the top and bottom sketches (Fig. 3 (d)), respectively. We then obtain the final CAD reconstruction by sweeping or lofting the detected sketches along the paths (Fig. 3 (e)). + +![](images/db4ebf5c3eeface6d083e50e13715f5614f636be0e816a07eb93f582048def8e.jpg) +Figure 3. Network architecture for SfmCAD: Stage 1 encodes the input voxels into feature $\mathbf{z}$ via a 3D convolutional encoder. An MLP then processes $\mathbf{z}$ to generate the control point coordinates $\mathbf{P}_1, \ldots, \mathbf{P}_4$ for each Bézier curve, as well as the dimensions $l, w$ and twist angles $\delta^u, \delta^b$ for each Box proxy $\mathcal{B}$ . In Stage 2, another MLP transforms feature $\mathbf{z}$ into local features $\mathbf{z}'$ , which are then concatenated with the testing point coordinate and fed into the implicit sketch network to predict the sketch SDF value. After extending the sketches along the paths, we combine all parts for the final reconstruction. + +![](images/64fe7203a5b631c604c7cf03634621b3e325522753e6b3c211bd21d291fae198.jpg) +Input +Figure 4. 2D illustration of our Box+Path representation (right) to approximate the input shape (left), which can be decomposed into a set of feature primitives (middle, the green arrows indicate the sweeping paths). + +![](images/1fbdd88bc1a27ea19bd5f9bbf8248a677502e7ba8acb114eba69869654b27b5f.jpg) +Feature primitives + +![](images/0c64f2ae0f223d3741d6f3e3d667583745235a2a896926f81799696a09ba5b2c.jpg) +Box+Path + +# 4. Methodology + +Here, we first present the box+path representation, which can be efficiently learned to approximate the input shape by stacking a set of box-like proxies along the sweeping path. Then, we define the neural sketch+path to refine the coarse reconstruction. Finally, we introduce the two-stage learning network for inferring this coarse-to-fine reconstruction. + +# 4.1. Box+Path Representation + +Previous unsupervised methods [23, 30, 31] for computing the Signed Distance Function (SDF) for geometric primitives (e.g., cubes, spheres, cylinders) rely on linear transformations that convert global testing point coordinates into local ones. However, the inherent non-linearity of curved paths limits the direct application of linear methods. We introduce a linear approximation method for the curved SDF computation, which we term Box+Path representation. We will first discuss the Box-Path in a 2D context to provide a more intuitive understanding before expanding it to 3D. + +Box+Path in 2D. Specifically, we approximate one feature primitive $\mathcal{P}$ using a discrete set of $N_{s}$ rigid box proxies, + +i.e., $\mathcal{P} = \{\mathcal{B}_i\}_{i=1}^{N_s}$ . The center position $\mathbf{c}_i$ of each box $\mathcal{B}_i$ is obtained by uniformly sampling the Bezier curve $\Upsilon(t)$ : + +$$ +\mathbf {c} _ {i} = \Upsilon (i / N _ {s}), \quad i = 1, \dots , N _ {s} \tag {2} +$$ + +The height $h_i$ and the normal vector (or heading direction) $\mathbf{n}_i$ of each box are calculated as + +$$ +h _ {i} = \left| \mathbf {c} _ {i + 1} - \mathbf {c} _ {i} \right|, \mathbf {n} _ {i} = \frac {\mathbf {c} _ {i + 1} - \mathbf {c} _ {i}}{\left| \mathbf {c} _ {i + 1} - \mathbf {c} _ {i} \right|} \tag {3} +$$ + +By incorporating a learnable width parameter $w$ , the 2D Box+Path can be fully characterized. + +Box+Path in 3D. When extended to 3D space, we introduce the tangent vector $\mathbf{t}_i$ and the binormal vector $\mathbf{b}_i$ to parameterize the orientation of the $i$ -th box proxy. To allow the sketch to rotate by a prescribed angle along the sweeping path, we should learn rotation angles $\delta^u$ and $\delta^b$ at the start and end points of the sweeping path. Then, the intermediate rotation angles $\delta_i$ are obtained by linear interpolation along the sweeping path. Next, we compute the tangent vector $\mathbf{t}_i$ as the cross product of the normal $\mathbf{n}_i$ and the positive $z$ -axis unit vector $\mathbf{e}_z$ . We then rotate $\mathbf{t}_i$ around $\mathbf{n}_i$ : + +$$ +\mathbf {t} _ {i} = R _ {\mathbf {n} _ {i}} \left(\mathbf {n} _ {i} \times \mathbf {e} _ {z}, \delta_ {i}\right), \tag {4} +$$ + +where $R_{\mathbf{n}_i}(\mathbf{v}, \delta_i)$ represents a counter-clockwise rotation of vector $\mathbf{v}$ with angle $\delta_i$ around $\mathbf{n}_i$ . The binormal vector $\mathbf{b}_i$ is obtained as a cross product of $\mathbf{n}_i \times \mathbf{t}_i$ . Altogether, the triplet $\{\mathbf{n}_i, \mathbf{t}_i, \mathbf{b}_i\}$ defines the orientation for each box along the curve path. The 3D Box+Path is fully characterized by incorporating learnable dimension parameters $l, w$ . + +# 4.2. Neural Sketch+Path + +We present differentiable methods for lofting or sweeping 2D sketches into 3D solid shapes to accurately reconstruct + +the feature primitives. The shape of each 2D sketch is defined by an implicit representation, where we use $\hat{S}_{\mathrm{sk}}$ to represent the SDF of sketches. We now describe the differentiable operators that calculate the SDF of the 3D feature primitives corresponding to the sketches. + +Differentiable Sketch-Extrude. In the context of generating the geometry of feature primitives from sketch and extrude operations, $\hat{S}_{\mathrm{extrude}}^i$ denotes the $i$ -th sketch-extrude operation, $\mathbf{x}_{i_z}$ represents the $z$ -component of the $i$ -th position vector, and $h_i$ is the height of the $i$ -th feature primitive. We introduce a set of intermediate functions $\mathcal{M}_1, \mathcal{M}_2$ , and $\mathcal{M}_3$ to facilitate a clearer formulation: + +$$ +\begin{array}{l} \mathcal {M} _ {1} \left(\hat {\mathcal {S}} _ {\mathbf {s k}} ^ {i}\right) = \max \left(\hat {\mathcal {S}} _ {\mathbf {s k}} ^ {i}, \left| \mathbf {x} _ {i _ {z}} \right| - h _ {i}\right), \\ \mathcal {M} _ {2} \left(\hat {\mathcal {S}} _ {\mathrm {s k}} ^ {i}\right) = \max \left(\hat {\mathcal {S}} _ {\mathrm {s k}} ^ {i}, 0\right), \\ \mathcal {M} _ {3} = \max \left(\left| \mathbf {x} _ {i _ {z}} \right| - h _ {i}, 0\right). \\ \end{array} +$$ + +The SDF of the $i$ -th extrusion primitive is then given by: + +$$ +\hat {S} _ {\text {e x t r u d e}} ^ {i} = \min \left(\mathcal {M} _ {1} \left(\hat {S} _ {\mathrm {s k}} ^ {i}\right), 0\right) + \sqrt {\mathcal {M} _ {2} \left(\hat {S} _ {\mathrm {s k}} ^ {i}\right) ^ {2} + \mathcal {M} _ {3} ^ {2}} \tag {5} +$$ + +Differentiable Sketch-Loft. For the loft operation, we interpolate between two distinct sketches, $\hat{S}_{\mathrm{sk}}^{u_i}$ and $\hat{S}_{\mathrm{sk}}^{l_i}$ , at the upper and lower faces of the $i$ -th feature primitive. Given the interpolation factor $\alpha$ as $|\mathbf{x}_{i_z}| / h_i$ , the linearly interpolated sketch, $\hat{S}_{\mathrm{sk}}^{i_\alpha}$ , is computed as: + +$$ +\hat {\mathcal {S}} _ {\mathrm {s k}} ^ {i _ {\alpha}} = (1 - \alpha) \cdot \hat {\mathcal {S}} _ {\mathrm {s k}} ^ {l _ {i}} + \alpha \cdot \hat {\mathcal {S}} _ {\mathrm {s k}} ^ {u _ {i}}. \tag {6} +$$ + +We use the same intermediate variables as in the extrusion case, but with $\hat{S}_{\mathrm{sk}}^{i}$ replaced by $\hat{S}_{\mathrm{sk}}^{i_{\alpha}}$ . The SDF of the $i$ -th loft primitive can be expressed similarly to Eq. 5: + +$$ +\hat {\mathcal {S}} _ {\text {l o f t}} ^ {i} = \min \left(\mathcal {M} _ {1} \left(\hat {\mathcal {S}} _ {\mathrm {s k}} ^ {i _ {\alpha}}\right), 0\right) + \sqrt {\mathcal {M} _ {2} \left(\hat {\mathcal {S}} _ {\mathrm {s k}} ^ {i _ {\alpha}}\right) ^ {2} + \mathcal {M} _ {3} ^ {2}}. \tag {7} +$$ + +Differentiable Sketch-Sweep. For sweep operation, we decompose it into a series of $N_{s}$ extrusions. We denote the predicted SDF for the $j$ -th extrude segment as $\hat{S}_{\mathrm{extrude}}^i$ . The SDF of the $i$ -th sweep primitive is then obtained by taking the minimum of the SDFs of all extrusion segments: + +$$ +\hat {\mathcal {S}} _ {\text {s w e e p}} ^ {i} = \min _ {j \in 1, \dots , N _ {s}} \hat {\mathcal {S}} _ {\text {e x t r u d e}} ^ {j}. \tag {8} +$$ + +Transformation to Parametric CAD. Following [23], we transform implicit sketches into parametric ones. With the control points of Bezier curves outputted by the network, we can achieve the sketch+path parametric reconstruction. + +# 4.3. Two-stage Network Training + +Stage 1: Box+Path learning. In this stage, the network predicts the control point coordinates $\{\mathbf{P}_1,\dots ,\mathbf{P}_4\}$ of the cubic Bezier curve and the parameters $\{l_i,w_i,\delta^u,\delta^b\}$ describing the $i$ -th Box $^+$ Path $\mathcal{B}_i$ . These parameters subsequently inform the computation of the remaining parameters $\{h_i,\mathbf{n}_i,\mathbf{t}_i,\mathbf{b}_i\}$ , as detailed in Sect. 4.1. To align the + +path of $\mathcal{B}$ with the target shape path, we first compute the SDF, $\hat{S}_{\mathrm{box}}^{ij}$ , for each box at the $i$ -th primitive and $j$ -th segment, where $i \in [1, N_p]$ and $j \in [1, N_s]$ , with $N_p$ and $N_s$ representing the total number of primitives and segments. Taking the union over the $\hat{S}_{\mathrm{box}}^{ij}$ yields the SDF for the overall box+path $\hat{S}_{\mathcal{B}}$ : + +$$ +\hat {\mathcal {S}} _ {\mathcal {B}} = \bigcup_ {i = 1} ^ {N _ {p}} \bigcup_ {j = 1} ^ {N _ {s}} \hat {\mathcal {S}} _ {\text {b o x}} ^ {i j}, \tag {9} +$$ + +where $\bigcup (x) = \mathrm{softmin}(\mu \cdot x)\cdot x$ and $\mu$ is a hyperparameter. We then apply a differentiable converter $\Psi$ similar to prior works [23, 31] to transform the $\hat{S}_{\mathcal{B}}$ into occupancy, yielding $\hat{\mathcal{O}}_{\mathcal{B}} = \Psi (\hat{S}_{\mathcal{B}})$ , where $\Psi (x) = (\tanh (-\beta \cdot x) + 1) / 2$ , and $\beta$ is a hyperparameter. We then compute the Mean Squared Error (MSE) loss between $\hat{\mathcal{O}}_{\mathcal{B}}$ and the ground truth occupancy $\tilde{\mathcal{O}}$ : + +$$ +\mathcal {L} _ {\mathcal {B}} = \mathbb {E} _ {x \in \mathcal {X}} \left[ | | \hat {\mathcal {O}} _ {\mathcal {B}} - \tilde {\mathcal {O}} | | _ {2} ^ {2} \right], \tag {10} +$$ + +where $\mathcal{X}$ is the set of all testing points. The exclusive use of $\mathcal{L}_{\mathcal{B}}$ may lead to potential challenges: (1) the emergence of the curve with acute angles, and (2) the dimensions of $\mathcal{B}$ vastly outstrip the path length. Both increase the risk of shape self-intersections, which are considered invalid in CAD. To address these issues, we introduce two additional regularization terms, + +$$ +\mathcal {L} _ {\mathrm {s m}} = \frac {1}{N _ {p}} \sum_ {i = 1} ^ {N _ {p}} \left(1 - \min _ {j = 1, \dots , N _ {s}} \left(\mathbf {n} _ {i, j} \cdot \mathbf {n} _ {i, j + 1}\right)\right) \tag {11} +$$ + +$$ +\mathcal {L} _ {\mathrm {l w}} = \frac {1}{2 N _ {p}} \sum_ {i = 1} ^ {N _ {p}} \operatorname {R e L U} \left(h _ {i} + w _ {i} - 2 \cdot \Theta\right), \tag {12} +$$ + +where $\mathcal{L}_{\mathrm{sm}}$ penalizes the sharpest curvature in the curve by maximizing the inner product between adjacent $\mathbf{n}$ , and $\mathcal{L}_{\mathrm{lw}}$ penalizes any length or width greater than a threshold $\Theta$ . Combine with balance factors $\lambda_{1}, \lambda_{2}$ , the overall loss of Stage 1 is: + +$$ +\mathcal {L} _ {\text {b o x}} = \mathcal {L} _ {\mathcal {B}} + \lambda_ {1} \mathcal {L} _ {\mathrm {s m}} + \lambda_ {2}. \mathcal {L} _ {\mathrm {l w}} \tag {13} +$$ + +Stage 2: Implicit sketch learning. Using the path learned in Stage 1, we progress towards fine reconstruction in Stage 2 by learning sketches. We initially utilize a Multi-Layer Perceptron (MLP) to convert the feature $\mathbf{z}$ into $N_{p}$ local features $\mathbf{z}^{\prime}$ . These local features are then concatenated with the local $\{x,y\}$ coordinates of each point and fed into an implicit sketch network, which is essentially an MLP predicting the SDF value $\hat{S}_{\mathrm{sk}}^{i}$ . For the sketch-loft operation, two implicit sketch networks are used to learn $\hat{S}_{\mathrm{sk}}^{u_i}$ and $\hat{S}_{\mathrm{sk}}^{l_i}$ . Following the computation methods outlined in Sect. 4.2, we can derive the predicted overall occupancy value $\hat{\mathcal{O}}_{\mathrm{F}}$ by: + +$$ +\hat {\mathcal {O}} _ {\mathrm {F}} = \left\{ \begin{array}{l l} \Psi (\bigcup_ {i = 1} ^ {N _ {p}} \hat {S} _ {\text {s w e e p}} ^ {i}), & \text {f o r s w e e p o p e r a t i o n} \\ \Psi (\bigcup_ {i = 1} ^ {N _ {p}} \hat {S} _ {\text {l o f t}} ^ {i}), & \text {f o r l o f t o p e r a t i o n .} \end{array} \right. +$$ + +Finally, an MSE loss $\mathcal{L}_{\mathrm{rec}}$ is computed against the ground truth $\mathcal{O}$ : + +$$ +\mathcal {L} _ {\mathrm {r e c}} = \mathbb {E} _ {x \in \mathcal {X}} \left[ \left| \left| \hat {\mathcal {O}} _ {\mathrm {F}} - \tilde {\mathcal {O}} \right| \right| _ {2} ^ {2} \right]. \tag {14} +$$ + +The direct computation of $\mathcal{L}_{\mathrm{rec}}$ over the entire point set $\mathcal{X}$ requires the repetition of $\mathcal{X}$ corresponding to $\hat{S}_{\mathrm{extrude}}$ for $N_{p} \times N_{s}$ times, which is time-consuming. To accelerate the training of the implicit sketch network, we propose a real-time sampling strategy, exploiting the sketch consistency trait inherent in the sweep operation. Specifically, we sample grid-like testing points $y_{i}$ in the base sketch plane of the $i$ -th path $\Upsilon_{i}$ obtained in the first stage. The occupancy value of $y_{i}$ is determined by inputting it into the implicit sketch network and applying the converter $\Psi$ . Subsequently, $y_{i}$ is moved along the curve for $N_{s} - 1$ times, with its occupancy value directly replicated, yielding $\hat{\mathcal{O}}_{\mathrm{Fi}}^{*}$ . The ground truth occupancy value $\tilde{\mathcal{O}}_i^*$ is provided by the input voxel. After performing this procedure for all paths and conducting a union operation, the loss function corresponding to real-time sampling is given by: + +$$ +\mathcal {L} _ {\mathrm {r e c}} ^ {*} = \mathbb {E} _ {y \in \mathcal {Y}} \left[ | | \hat {\mathcal {O}} _ {\mathrm {F}} ^ {*} - \tilde {\mathcal {O}} ^ {*} | | _ {2} ^ {2} \right], \tag {15} +$$ + +where $\mathcal{V}$ represents the set of testing points obtained through real-time sampling, and $*$ denotes computation using real-time sampling data. + +Note that the SfmCAD is designed to choose between learning either loft or sweep operations, which is determined by its hyperparameters. However, enabling the network to concurrently learn both operations is also viable, with details provided in the supplementary materials. + +# 5. Results and Evaluation + +We evaluate the performance of SfmCAD on two reconstruction tasks: (1) holistic shapes of different categories and (2) semantically segmented shapes. We show the output editability of SfmCAD in the supplemental materials. + +# 5.1. Setup + +Implementation details. We implemented SfmCAD in PyTorch by using an NVIDIA® TITAN RTX GPU. The Adam optimizer [15] is utilized with a learning rate 1e-4 and beta values of (0.5, 0.99). The hyperparameters in Sect. 4.3 are set to: $\mu = 20$ , $\beta = 50$ , $\Theta = 0.1$ , $\lambda_1 = 0.05$ , $\lambda_2 = 0.05$ . In our evaluations, we employ the best-performing operation (Loft/Sweep) for each category of holistic shapes and every part of segmented shapes. SfmCAD trains for 500 epochs per stage with a batch size of 24, followed by fine-tuning each test shape for 200 iterations per stage. All baseline methods are trained for 1,000 epochs and fine-tuned for 400 iterations for a fair comparison. + +Evaluation metrics. We utilize three common metrics for 3D shape reconstruction: Chamfer Distance (CD), Normal + +![](images/efdf5ab4d3d5aa26ec6a92047f1176e01467ba81e248c0d24532a9f611a99b02.jpg) +Figure 5. Visual comparison on ABC dataset. + +Table 1. Quantitative comparison on ABC dataset. + +
MethodsCD↓ECD↓NC↑
UCSG-Net [14]1.23321.7860.866
CSG-Stump [30]0.6717.7510.892
ExtrudeNet [31]0.5197.1110.885
SECAD-Net [23]0.5067.2860.884
Ours0.3955.0380.919
+ +Consistency (NC), and Edge Chamfer Distance (ECD). All experiments generate the reconstructed shapes via Marching Cubes at a resolution of $256^3$ . To calculate CD, we sample 8,192 points from the ground truth and predicted shape surfaces. For ECD, we employ a threshold of 0.1 for normal cross-products to identify points close to the edges on the ground truth surface and 0.5 for the predicted surface, considering the lack of sharpness in shapes generated by Marching Cubes. For better visual clarity, we magnify the CD and ECD by factors of $10^3$ and $10^2$ . + +# 5.2. Reconstruction from Holistic Shape + +We compare SfmCAD with UCSG-Net [14], CSG-Stump [30], ExtrudeNet [31] and SECAD-Net [23]. We test these methods on holistic shapes from different sources, including CAD parts in ABC [16], ShapeNet objects [1], and a synthetic tree dataset. The comparison against CAPRI-Net [49] is provided in the supplementary materials. + +Evaluation on ABC dataset. We employ the voxel grids and sampling point data provided by [49]. The dataset included 5,000 groups of data for training and 1,000 for testing. We select 100 shape subsets from the test set for eval + +![](images/4c9f492e70ed69c1a36b65b7620e294165e499b64318137444c87d16d57e9196.jpg) +Figure 6. Visual comparison on ShapeNet dataset. + +Table 2. Quantitative comparison on ShapeNet dataset. + +
MethodsCD↓ECD↓NC↑
UCSG-Net [14]2.11123.6160.815
CSG-Stump [30]2.16817.5920.818
ExtrudeNet [31]0.91014.1100.849
SECAD-Net [23]0.83614.8180.837
Ours0.62614.0960.867
+ +Table 3. Quantitative comparison on Tree dataset. + +
MethodsExtrudeNet [31]SECAD-Net [23]Ours
CD↓7.2422.1970.686
NC↑0.6080.6220.722
+ +uation. The visual results are shown in Fig. 5, while the quantitative results are presented in Tab. 1. + +Evaluation on ShapeNet dataset. We utilize three categories from ShapeNet: chair, table, and display. The voxel grids and sampling points are provided by [3]. We selected 120 shapes for evaluation, with 40 from each category. Visual and quantitative results are in Fig. 6 and Tab. 2. + +Evaluation on tree branches. We use a procedural modeling approach of Palubicki et al. [26] to generate tree meshes voxelized into a resolution of $64^{3}$ . We adopt the method from [2] to obtain sampling points and occupancy values. This process results in 5,000 data groups, of which 4,500 are used for training and 500 for testing. We train all methods using increasingly denser voxel resolutions $(16^{3},32^{3},64^{3})$ and select 50 trees for numerical evaluation. + +![](images/cc7741dd422e4c70e6ec5e185ad0b0e76423ff7409b14cae5c587b893a65149d.jpg) +Figure 7. Visual comparison on Tree dataset. + +tion. Note given the cylindrical nature of tree branches, we adapt our Box+Path representation by substituting boxes with cylinders for the final reconstruction (see Fig. 7 and Tab. 3 for visual and quantitative results). + +Summary of comparative analysis. As demonstrated in Tables 1, 2, and 3, SfmCAD consistently outperforms other methods in terms of CD, ECD, and NC, suggesting that the sketch+path representation of SfmCAD offers more flexible shape expression capabilities than the geometric primitives used by UCSG-Net and CSG-Stump, as well as the simple sketch+extrusion operations employed by ExtrudeNet and SECAD. Figures 5, 6, and 7 further show SfmCAD's ability to generate detailed reconstructions (such as thin structures) across diverse shape categories. + +# 5.3. Reconstruction from Segmented Shape + +Semantic segmentation provides information for reconstructing complex shapes. Previous CAD reconstruction methods [23, 31] could not utilize this information because their single-type primitives lack the flexibility to represent each part independently. In contrast, SfmCAD successfully reconstructs multiple primitives with its sketch+path representation capability. + +We show the task of unsupervised CAD reconstruction utilizing semantic segmentation. The evaluation is conducted on PartNet dataset [48], which provides instance-level 3D part information. We utilize three categories: chair, table, and trashcan. We voxelize the segmented parts and sampled occupancy values at resolutions $16^{3}, 32^{3}, 64^{3}$ following the scheme by [2]. This process yields a total of 12,000 chair parts (1,229 complete), 12,000 table parts (1,491 complete), and 3,000 trashcan parts (234 complete). The train-test split for all categories was set as 5:1. We selected 90 complete shapes, including 40 tables, 40 chairs, and 10 trashcans, for evaluation. + +All methods were trained on part shapes using increas + +![](images/a8f8675dfb12f8055225e7b780ce587fc1aab2e348791fb888e141a6bb401604.jpg) +Figure 8. Visual comparison on PartNet dataset. + +Table 4. Quantitative comparison on PartNet dataset. + +
MethodsCD↓ECD↓NC↑
ExtrudeNet [31]3.2510.5490.783
SECAD-Net [23]1.8880.4720.818
Ours1.4100.4460.814
+ +Table 5. Ablation study on sweep twist angles and loss terms. + +
Settings- (δu, δl)-δl- Llw- LsmOurs
CD↓4.0923.3674.8043.2683.142
ECD↓0.1730.1390.1640.1540.133
NC↑0.6550.7140.6850.7320.776
+ +ingly denser voxel resolutions $(16^{3}, 32^{3}, 64^{3})$ . For each part, we set ExtrudeNet [31] and SECAD-Net [23] to output an extruded cylinder and SfmCAD to output a sketch+path shape. After reconstructing all parts, they were reassembled to obtain the final CAD models. + +The quantitative results in Tab. 4 indicate that our method outperforms others. Compared to the closest competitor, SECAD-Net, SfmCAD improves the CD by $25.3\%$ and the ECD by $5.5\%$ . Fig. 8 illustrates that the sketch+extrusion reconstruction of ExtrudeNet and SECAD can only characterize the nonlinear shapes through sketches. In contrast, SfmCAD can model these complex shapes using sketches and free-form paths, enabling a more refined reconstruction. + +# 5.4. Ablation Study + +We conducted an ablation study by using 40 box+path reconstructions on chair class from PartNet to evaluate the efficiency of our network design and loss terms. Initially, we disable the two sweep twist angles $\delta^u$ , $\delta^l$ or $\delta^l$ . Then we separately discarded $\mathcal{L}_{\mathrm{sm}}$ and $\mathcal{L}_{\mathrm{lw}}$ . The visual results in Fig. 9 and the quantitative results in Tab. 5 suggest that + +![](images/59112db521e7246d8c7e561bac2e5cad7b7329363fce054f73854cf20fc3b3d9.jpg) +Figure 9. Visual comparison between reconstruction results for ablation study. From top to bottom, each column shows the predicted 3D path, vector $\{\mathbf{n},\mathbf{t},\mathbf{b}\}$ , and the box+path representation. + +removing any of these settings negatively affects the reconstruction performance of SfmCAD. + +# 6. Conclusion and Future Work + +We have presented an unsupervised neural network for CAD reconstruction that infers high-level CAD modeling operations. We have also introduced a unique sketch+path representation that encapsulates the local geometric details and overall structural information of the 3D shape. A novel two-stage learning strategy, operating in a coarse-to-fine manner, was proposed to address the complexity of simultaneous sketch and path learning, thus improving the efficiency of network training. We evaluated SfmCAD using several different categories of objects and demonstrated its advantages by comparing it to state-of-the-art methods. + +In future work, we plan to explore using 2D sketches and 3D part templates to enhance the efficiency of unsupervised CAD reconstruction. Besides, given the manipulable property of SfmCAD's outputs, extending this work for generative CAD design or modeling would be meaningful. + +Acknowledgments. We thank the anonymous reviewers for their valuable suggestions. This work is partially funded by the Strategic Priority Research Program of the Chinese Academy of Sciences (XDB0640000), Guangdong Science and Technology Program (2023B1515120026), National Natural Science Foundation of China (62172416, 62172415, U22B2034, 62262043, 62365014), and Youth Innovation Promotion Association of the Chinese Academy of Sciences (2022131). This work was supported by NRCS grant #NR233A750004G044 to Benes. The findings and conclusions should not be construed to represent any agency determination or policy. + +# References + +[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6 +[2] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5939-5948, 2019. 1, 7 +[3] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-Net: Generating compact meshes via binary space partitioning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 45–54, 2020. 1, 7 +[4] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. In ACM SIGGRAPH 2004 Papers, pages 905–914. 2004. 2 +[5] Tao Du, Jeevana Priya Inala, Yewen Pu, Andrew Spielberg, Adriana Schulz, Daniela Rus, Armando Solar-Lezama, and Wojciech Matusik. Inversecsg: Automatic conversion of 3d models to csg trees. ACM Trans. Graph., 37(6):1-16, 2018. 2 +[6] Elona Dupont, Kseniya Cherenkova, Anis Kacem, Sk Aziz Ali, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Cadops-net: Jointly learning cad operation types and steps from boundary-representations. In International Conference on 3D Vision (3DV), pages 114-123. IEEE, 2022. 2 +[7] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems (NeurIPS), pages 5885–5897, 2021. 2 +[8] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Trans. Graph., 39(6):1-16, 2020. 3 +[9] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: CAD reconstruction by B-rep chain complex generation. ACM Trans. Graph., 41(4): 1-18, 2022. 1, 2 +[10] Rostislav Hulik, Michal Spanel, Pavel Smrz, and Zdenek Materna. Continuous plane detection in point-cloud data based on 3d hough transform. Journal of visual communication and image representation, 25(1):86-97, 2014. 2 +[11] Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G Lambourne, Karl DD Willis, Thomas Davies, Hooman Shayani, and Nigel Morris. Uv-net: Learning from boundary representations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11703-11712, 2021. 2 +[12] Pradeep Kumar Jayaraman, Joseph G Lambourne, Nishkrit Desai, Karl DD Willis, Aditya Sanghi, and Nigel JW Morris. Solidgen: An autoregressive model for direct b-rep synthesis. arXiv preprint arXiv:2203.13944, 2022. 2 +[13] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2 +[14] Kacper Kania, Maciej Zieba, and Tomasz Kajdanowicz. UCSG-NET-Unsupervised discovering of constructive solid + +geometry tree. In Advances in Neural Information Processing Systems (NeurIPS), pages 8776-8786, 2020. 2, 6, 7 +[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6 +[16] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 9601-9611, 2019. 2, 6 +[17] Joseph G Lambourne, Karl DD Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. BRepNet: A topological message passing system for solid models. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 12773-12782, 2021. 1, 2 +[18] Joseph George Lambourne, Karl Willis, Pradeep Kumar Jayaraman, Longfei Zhang, Aditya Sanghi, and Kamal Rahimi Malekshan. Reconstructing editable prismatic cad from rounded voxel models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2 +[19] Eric-Tuan Lé, Minhyuk Sung, Duygu Ceylan, Radomir Mech, Tamy Boubekeur, and Niloy J Mitra. CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In IEEE International Conference on Computer Vision (ICCV), pages 7457-7466, 2021. 1, 2 +[20] Bosheng Li, Jacek Kałuzny, Jonathan Klein, Dominik L. Michels, Wojtek Pałubicki, Bedrich Benes, and Soren Pirk. Learning to reconstruct botanical trees from single images. ACM Trans. Grap., 40(6):1-15, 2021. 2 +[21] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Sketch2cad: Sequential cad modeling by sketching in context. ACM Trans. Graph., 39(6):1-14, 2020. 3 +[22] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3d point clouds. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 2652-2660, 2019. 1, 2 +[23] Pu Li, Jianwei Guo, Xiaopeng Zhang, and Dong-Ming Yan. Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 16816-16826, 2023. 2, 3, 4, 5, 6, 7, 8 +[24] Yuanqi Li, Shun Liu, Xinran Yang, Jianwei Guo, Jie Guo, and Yanwen Guo. Surface and edge detection for primitive fitting of point clouds. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2 +[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 4460-4470, 2019. 1 +[26] Wojciech Palubicki, Kipp Horel, Steven Longay, Adam Runions, Brendan Lane, Radomír Měch, and Przemyslaw Prusinkiewicz. Self-organizing tree models for image synthesis. ACM Trans. Graph., 28(3):1–10, 2009. 7 +[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. In Advances in Neural Information Processing Systems (NeurIPS), pages 5077-5088, 2021. 2 + +[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 165-174, 2019. 1 +[29] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 10344-10353, 2019. 2 +[30] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, et al. CSG-stump: A learning friendly CSG-like representation for interpretable shape parsing. In IEEE International Conference on Computer Vision (ICCV), pages 12478-12487, 2021. 2, 4, 6, 7 +[31] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude for shape parsing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 482-498, 2022. 2, 3, 4, 5, 6, 7, 8 +[32] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient RANSAC for point-cloud shape detection. In Comput. Graph. Forum, pages 214-226, 2007. 2 +[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In International Conference on Learning Representations (ICLR), 2022. 2 +[34] Jami J Shah and Martti Mäntylä. Parametric and feature-based CAD/CAM: concepts, techniques, and applications. John Wiley & Sons, 1995. 2 +[35] Vadim Shapiro. Solid modeling. Handbook of computer aided geometric design, 20:473-518, 2002. 1 +[36] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5515-5523, 2018. 1, 2 +[37] Gopal Sharma, Difan Liu, Subhransu Maji, Evangelos Kalogerakis, Siddhartha Chaudhuri, and Radomir Mech. ParSeNet: A parametric surface fitting network for 3d point clouds. In European Conference on Computer Vision (ECCV), pages 261-276, 2020. 1, 2 +[38] Mikaela Angelina Uy, Yen-Yu Chang, Minhyuk Sung, Purvi Goel, Joseph G Lambourne, Tolga Birdal, and Leonidas J Guibas. Point2Cyl: Reverse engineering 3D objects from point clouds to extrusion cylinders. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11850-11860, 2022. 2, 3 +[39] Tamas Varady and Ralph Martin. Reverse engineering. In Handbook of Computer Aided Geometric Design, pages 651-681. 2002. 1 +[40] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. PIE-Net: Parametric inference of point cloud edges. Advances in Neural Information Processing Systems (NeurIPS), pages 20167–20178, 2020. 2 +[41] Karl DD Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G Lambourne, Armando Solar-Lezama, and Wo + +jciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM Trans. Graph., 40(4):1-24, 2021. 2, 3 +[42] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In IEEE International Conference on Computer Vision (ICCV), pages 6772–6782, 2021. 2 +[43] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl DD Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 6062-6070, 2021. 3 +[44] Xiang Xu, Karl DD Willis, Joseph G Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, and Yasutaka Furukawa. Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In International Conference on Machine Learning, 2022. 3 +[45] Fujio Yamaguchi. Curves and surfaces in computer aided geometric design. Springer Science & Business Media, 2012. 3 +[46] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 2 +[47] Siming Yan, Zhenpei Yang, Chongyang Ma, Haibin Huang, Etienne Vouga, and Qixing Huang. Hpnet: Deep primitive segmentation using hybrid representations. In IEEE International Conference on Computer Vision (ICCV), pages 2753-2762, 2021. 2 +[48] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, page to appear, 2019. 7 +[49] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11768–11778, 2022. 1, 2, 6 +[50] Fenggen Yu, Qimin Chen, Maham Tanveer, Ali Mahdavi Amiri, and Hao Zhang. $\mathrm{D}^2\mathrm{csg}$ : Unsupervised learning of compact csg trees with dual complements and dropouts. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2 \ No newline at end of file diff --git a/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/images.zip b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e653161da3f281eca8431e2ff2352e3d53c76149 --- /dev/null +++ b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b60b5df594e4209ec802ce9c83ae716a010ca26e537fc6cca0b4f88eb7898548 +size 593087 diff --git a/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/layout.json b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..666dbca6e9dbd0944a0254c31ac271a009547258 --- /dev/null +++ b/2024/SfmCAD_ Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations/layout.json @@ -0,0 +1,10798 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 544, + 140 + ], + "type": "text", + "content": "SfmCAD: Unsupervised CAD Reconstruction by Learning Sketch-based Feature Modeling Operations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "spans": [ + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{PuLi^{1,2}}" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "text", + "content": " Jianwei Guo" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "text", + "content": " Huibin Li" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "text", + "content": " Bedrich Benes" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "text", + "content": " Dong-Ming Yan" + }, + { + "bbox": [ + 99, + 159, + 494, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 177, + 424, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 177, + 424, + 189 + ], + "spans": [ + { + "bbox": [ + 167, + 177, + 424, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 167, + 177, + 424, + 189 + ], + "type": "text", + "content": "MAIS, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 190, + 538, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 190, + 538, + 204 + ], + "spans": [ + { + "bbox": [ + 52, + 190, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 190, + 538, + 204 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 52, + 190, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 190, + 538, + 204 + ], + "type": "text", + "content": "Computer Science, Purdue University" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 224, + 541, + 365 + ], + "blocks": [ + { + "bbox": [ + 50, + 224, + 541, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 224, + 541, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 224, + 541, + 365 + ], + "type": "image", + "image_path": "ce304b141a898683556788c5ffe5b1e21efcae42ffc29deb5e85e3bf8cbc3ca1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "type": "text", + "content": "Figure 1. SfmCAD parses voxel shapes and reconstructs them into a parameterized set of sketch-based feature modeling operations: the input voxels, neural typed sketch+path representation, and the final CAD model (each part is randomly colored)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 407, + 192, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 407, + 192, + 420 + ], + "spans": [ + { + "bbox": [ + 143, + 407, + 192, + 420 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 427, + 290, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 290, + 680 + ], + "type": "text", + "content": "This paper introduces SfmCAD, a novel unsupervised network that reconstructs 3D shapes by learning the Sketch-based Feature Modeling operations commonly used in modern CAD workflows. Given a 3D shape represented as voxels, SfmCAD learns a neural-typed sketch+path parameterized representation, including 2D sketches of feature primitives and their 3D sweeping paths without supervision, for inferring feature-based CAD programs. SfmCAD employs 2D sketches for local detail representation and 3D paths to capture the overall structure, achieving a clear separation between shape details and structure. This conversion into parametric forms enables users to seamlessly adjust the shape's geometric and structural features, thus enhancing interpretability and user control. We demonstrate the effectiveness of our method by applying SfmCAD to many different types of objects, such as CAD parts, ShapeNet objects, and tree shapes. Extensive comparisons show that SfmCAD produces compact and faithful 3D reconstructions with superior quality compared to alternatives. The code is released at https://github.com/BunnySoCrazy/SfmCAD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 407, + 386, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 386, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 386, + 420 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 433, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 546, + 564 + ], + "type": "text", + "content": "Reconstructing shapes into CAD representations is a fundamental problem in computer vision and graphics with numerous commercial applications in industrial design and manufacturing [39]. Recent years have seen an increased interest in leveraging deep learning techniques to learn geometric representations of 3D shapes, including implicit fields [2, 25, 28], geometric primitives [19, 22, 24, 37], boundary representation (B-Rep) [9, 17], or Constructive Solid Geometry (CSG) [3, 36, 49], each with its unique characteristics and applications, contributing to the rich tapestry of research and development in this field." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "However, current CAD reconstruction methods are still lacking detail reconstruction accuracy, compactness, and interpretability of the obtained models. For example, a typical line of research is dedicated to reconstruction-oriented tasks, i.e., representing 3D shapes as parametric primitives, implicit fields, or B-rep. These methods deliver a high level of detail and accuracy by capturing intricate details and complex geometries with remarkable precision. However, constructing CAD models by assembling instances of primitives or by identifying faces, edges, and vertices is a tedious process [35]. Besides, the results often include complex and unintuitive operations that make the model difficult to inter-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 223, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 223, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 223, + 713 + ], + "type": "text", + "content": "*Corresponding author: jianwei.guo@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4671" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "pret and edit. This lack of control poses significant challenges in scenarios where user interaction and shape manipulation are required." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 286, + 217 + ], + "type": "text", + "content": "A straightforward approach to CAD reconstruction is to learn CSG representation [14, 30, 36, 49] that implicitly recovers the CAD design history. A CSG 3D shape is expressed as a tree with leaves consisting of volumetric primitives (e.g., cubes, spheres, and cylinders) that are combined using Boolean operations in the CSG inner nodes. Despite the sought-after attribute of compact shape representation, these methods often fall short regarding reconstruction accuracy by using only a set of simple typed primitives." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 219, + 286, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 286, + 459 + ], + "type": "text", + "content": "We introduce SfmCAD, a novel neural model for reconstructing 3D shapes into their CAD representation. Compared to related work, SfmCAD improves the reconstruction accuracy and the level of model editability by decoupling 3D structures and local geometric details of the shape. SfmCAD is inspired by feature-based modeling, a predominant method for creating 3D models in modern CAD systems, which sequentially add features (e.g., holes, slots, bosses) that represent common manufacturing operations [6, 41, 42]. Specifically, we employ unsupervised learning to retrieve a neural typed sketch+path representation of 3D shapes. Taking 3D voxel data as input, SfmCAD learns implicit fields that embody the sketch profile and generates control points for a Bezier curve, constituting the sweeping path for the component's geometry. The basic extrusion operation has been learned in a supervised [18, 38] or unsupervised [23, 31] manner. However, our sketch+path representation supports not only extrude but also sweep, loft, and revolve operations, thus enhancing the shape representation capabilities, as demonstrated in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "type": "text", + "content": "At the core of the SfmCAD is a two-stage learning strategy that operates in a coarse-to-fine manner to tackle the inherent complexity and time-intensiveness of simultaneous sketch profile and sweeping path learning. Initially, the network is trained to learn a coarse Box+Path representation of the shape, enabling a quick grasp of the shape's path and broad structure using box-like geometric proxies. Then, an implicit network delves deeper into the geometric details of the shape, enhancing the reconstruction accuracy. We demonstrate the capability and versatility of our method by applying SfmCAD to many different types of objects, including CAD parts [16], ShapeNet objects [1], tree branches [20], and by comparing to previous works. We claim the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 630, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 630, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 286, + 700 + ], + "type": "text", + "content": "1. We propose SfmCAD, a novel neural approach that parses 3D shapes as a set of industry-standard sketch-based CAD modeling operations. To the best of our knowledge, SfmCAD is the first unsupervised and universal neural network for learning the common CAD commands, including extrude, sweep, loft, and revolve." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 712 + ], + "type": "text", + "content": "2. We introduce novel decoupling of 3D structures and" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 108 + ], + "type": "text", + "content": "local geometric details, which are represented by 3D Bézier curves and implicit 2D sketches. This representation ensures accurate reconstruction and ease of editing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 108, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 545, + 156 + ], + "type": "text", + "content": "3. We propose a two-stage learning strategy, namely 3D Box+Path learning and 2D implicit sketch learning, which significantly accelerates the training speed in shape parser learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 167, + 389, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 389, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 389, + 179 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 187, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 545, + 270 + ], + "type": "text", + "content": "Geometric primitive extraction. Geometric primitive is a commonly used representation for approximating and abstracting 3D shapes [13]. Traditional methods based on RANSAC [32], Hough transform [10], or variational optimization [4, 46] have been used to detect and fit primitives on point clouds or polygonal meshes, but they often require careful parameter tuning for each shape." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 271, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 569 + ], + "type": "text", + "content": "To overcome the overwhelming complexity of traditional methods, deep neural networks addressed the primitive segmentation/detection from point clouds [19, 47], they fit parametric surfaces [22, 29, 37], or detect parametric curves and sharp corners [9, 24, 40] to achieve compact CAD reconstructions. Although achieving remarkable accuracy on reconstruction tasks, these methods only output individual primitives with limited types, which restricts their capability for reconstructing complex and more general 3D shapes. Learning CSG-based reconstruction. CSG reconstruction embodies a 3D shape of high complexity and nonconvexity with CSG boolean operations to represent the 3D shape construction process [5]. CSGNet [36] first develops a neural model that parses a shape into a sequence of CSG operations. More recent works follow the line of CSG parsing by advancing the inference without any supervision [14] or improving representation capability with a three-layer reformulation of the classic CSG-tree [30], or handling richer geometric and topological variations by introducing quadric surface primitives [49, 50]. However, CSG reconstruction tends to combine a large number of shape primitives that limit the user editing capabilities. Moreover, using only basic primitives (e.g., boxes, spheres) is insufficient for approximating complicated shapes, thus usually limiting the reconstruction accuracy of small details [50]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Learning feature-based reconstruction. Modern CAD workflows use feature-based modeling that decomposes a target shape into a sequence of CAD commands [34], such as drawing a sketch followed by CAD operations such as extrusion, loft, etc. This paradigm makes modeling more efficient and in tune with how designers and engineers work. Recent work has explored the potential of deep learning to generate 2D engineering sketches [7, 27, 33], or directly learn B-Rep CAD models [6, 11, 12, 17]. Several approaches propose generative models for CAD design, predicting a sequence of CAD modeling operations to produce editable CAD models, such as DeepCAD [42], Fu" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4672" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 93, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 93, + 144 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 93, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 93, + 144 + ], + "type": "image", + "image_path": "b6d6e1b6a26f7ccf15ebac890899b83926979f626c2bfbc86d5b9a25fc5388a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 144, + 87, + 151 + ], + "lines": [ + { + "bbox": [ + 61, + 144, + 87, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 144, + 87, + 151 + ], + "type": "text", + "content": "Extrude" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 160, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 160, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 287, + 194 + ], + "type": "text", + "content": "Figure 2. Sketch-based feature modeling operators supported in SfmCAD. The resulting 3D shape of each operation is called a feature primitive." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 101, + 74, + 165, + 143 + ], + "blocks": [ + { + "bbox": [ + 101, + 74, + 165, + 143 + ], + "lines": [ + { + "bbox": [ + 101, + 74, + 165, + 143 + ], + "spans": [ + { + "bbox": [ + 101, + 74, + 165, + 143 + ], + "type": "image", + "image_path": "9b7a09a24e9599bc32258379f4261cfd9cc38fc157f7429a03f39757c3441ebb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 144, + 140, + 151 + ], + "lines": [ + { + "bbox": [ + 126, + 144, + 140, + 151 + ], + "spans": [ + { + "bbox": [ + 126, + 144, + 140, + 151 + ], + "type": "text", + "content": "Loft" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 76, + 203, + 143 + ], + "blocks": [ + { + "bbox": [ + 175, + 76, + 203, + 143 + ], + "lines": [ + { + "bbox": [ + 175, + 76, + 203, + 143 + ], + "spans": [ + { + "bbox": [ + 175, + 76, + 203, + 143 + ], + "type": "image", + "image_path": "8f2ff0c5fc413f576a8ac896823b0543c66e68606a78f1172668b4e5c91dac37.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 144, + 201, + 152 + ], + "lines": [ + { + "bbox": [ + 179, + 144, + 201, + 152 + ], + "spans": [ + { + "bbox": [ + 179, + 144, + 201, + 152 + ], + "type": "text", + "content": "Sweep" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 76, + 277, + 143 + ], + "blocks": [ + { + "bbox": [ + 217, + 76, + 277, + 143 + ], + "lines": [ + { + "bbox": [ + 217, + 76, + 277, + 143 + ], + "spans": [ + { + "bbox": [ + 217, + 76, + 277, + 143 + ], + "type": "image", + "image_path": "6cd51c599307985d3e078f6b310e5835392e30f5b3ba6ed7dd2fd441302b9325.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 144, + 261, + 152 + ], + "lines": [ + { + "bbox": [ + 235, + 144, + 261, + 152 + ], + "spans": [ + { + "bbox": [ + 235, + 144, + 261, + 152 + ], + "type": "text", + "content": "Revolve" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 214, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 262 + ], + "type": "text", + "content": "sion360 [41], Zone graph [43], and SkexGen [44]. These works focus on parametric CAD generation and do not address the reconstruction task of CAD models with suitable geometric loss functions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 262, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 382 + ], + "type": "text", + "content": "Most closely related to our work are the approaches that learn sketch+extrude operation (e.g., Point2Cyl [38], ExtrudeNet [31], SECAD-Net [23]), in which a collection of extrusions are predicted to build CAD models. While expressive within their scopes, these methods are limited to only one type of CAD operation (i.e., extrude) to simplify the search space. SfmCAD includes loft, revolve, and sweep reconstructions, broadening the shape representation capabilities. This approach enables us to test on a broader range of shapes beyond CAD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 392, + 235, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 235, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 235, + 404 + ], + "type": "text", + "content": "3. Problem Statement and Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 412, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 473 + ], + "type": "text", + "content": "Sketch-based feature modeling creates 2D engineering sketches and then lifts them to 3D along a path using CAD operations [8, 21]. Here, we first introduce this typed sketch+path representation for sketch-based modeling and then give an overview of its generation from a voxel model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 479, + 209, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 209, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 209, + 493 + ], + "type": "text", + "content": "3.1.Typed Sketch+Path Modeling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": "We define a profile as a closed curve, and its enclosed region is composed of one or multiple inner/outer loops. We define a sketch as the collection of one profile and its loops in a 2D plane. Let " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " denote the sketch and " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " be a 3D path to extend the sketches to create a 3D solid shape. We define four principal operations that are commonly used in modern CAD workflow (see Fig. 2): Extrude extends " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " along a linear path, Revolve produces a model by rotating " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " around an axis, Sweep moves " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " along a specified path to generate generalized cylinder shapes, and Loft allows for variations in the sketch, interpolating between multiple distinct sketches " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\{Z_1, Z_2, \\ldots, Z_n\\}" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " along the shortest distances between them to create complex shapes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Considering the similarity between the above operations, i.e., lifting 2D sketches to a 3D shape along a 3D path, we use a general typed sketch " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " path to represent all of these operations, where the term 'typed' indicates the specific type of operation. Without loss of generality, we call the 3D" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "path a sweeping path and call the lofted or swept 3D shape a feature primitive. We further use a Bézier curve " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\Upsilon(t)" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " to represent the sweeping path:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 123, + 545, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 123, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 317, + 123, + 545, + 154 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\Upsilon (t) = \\sum_ {i = 0} ^ {n} \\binom {n} {i} (1 - t) ^ {n - i} t ^ {i} \\mathbf {P} _ {i}, \\quad t \\in [ 0, 1 ] \\tag {1}", + "image_path": "7d01185a4c46d0186f2bd956222af8a8b8dd1f6def5d58e517c6ee30c61670a1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 167, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 167, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 304, + 167, + 545, + 216 + ], + "type": "text", + "content": " is a control point. The Bézier cubic is expressive enough to approximate the shapes shown in the paper, but higher-order curves can be obtained by simply changing the number of control points output by the network." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 218, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 327 + ], + "type": "text", + "content": "Moreover, it is noteworthy that the Extrude and Revolve operations can be regarded as general variants of the sweep operation because a Bézier curve becomes a linear path (for Extrude) if all the control points are collinear and an " + }, + { + "bbox": [ + 304, + 218, + 545, + 327 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 218, + 545, + 327 + ], + "type": "text", + "content": "-piece (usually " + }, + { + "bbox": [ + 304, + 218, + 545, + 327 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 304, + 218, + 545, + 327 + ], + "type": "text", + "content": ") cubic Bézier curve can approximate a circular path (for Revolve) [45]. In the following, we focus on reconstructing Sweep and Loft operations, where Sweep contains Extrude and Revolve, and the Loft is restricted to contain only two sketches with a linear path between them." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 342, + 373, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 373, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 373, + 354 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 363, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 507 + ], + "type": "text", + "content": "Given a 3D voxel shape, we aim to reconstruct the CAD model by neurally decomposing the target shape into a sequence of typed sketch+path representation, which can be executed by the corresponding CAD commands to assemble a compact CAD model. Learning the path and sketch of a shape simultaneously is challenging and time-consuming. Thus, we propose a two-stage learning approach to improve the efficiency of network training. First, the network learns a Box+Path representation to enclose the input shape, capturing the sweeping path, basic geometry, and rotation angles. Then, we learn implicit sketches that refine the coarse box+path into a detailed shape representation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Fig. 3 details the two-stage learning network. We begin with a 3D convolutional encoder that transforms the voxel input into a feature vector " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", which is then processed through a multilayer perceptron (MLP), yielding matrices " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " are the control points of the Bezier curves representing the sweeping paths (Fig. 3 (b)), and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " encapsulates the box+path parameters (Fig. 3 (c)), including length, width, and sweep's twist angles. Once the sweeping paths are derived from Stage 1, we focus on learning the geometric details surrounding each path. The feature " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is then mapped to " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " local features via another MLP. These vectors are subsequently concatenated with test points and fed into an implicit sketch network. The output from this network includes two occupancy values that represent the top and bottom sketches (Fig. 3 (d)), respectively. We then obtain the final CAD reconstruction by sweeping or lofting the detected sketches along the paths (Fig. 3 (e))." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4673" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 534, + 232 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 534, + 232 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 534, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 534, + 232 + ], + "type": "image", + "image_path": "db4ebf5c3eeface6d083e50e13715f5614f636be0e816a07eb93f582048def8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": "Figure 3. Network architecture for SfmCAD: Stage 1 encodes the input voxels into feature " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": " via a 3D convolutional encoder. An MLP then processes " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": " to generate the control point coordinates " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1, \\ldots, \\mathbf{P}_4" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": " for each Bézier curve, as well as the dimensions " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "l, w" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": " and twist angles " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\delta^u, \\delta^b" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": " for each Box proxy " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": ". In Stage 2, another MLP transforms feature " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": " into local features " + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{z}'" + }, + { + "bbox": [ + 46, + 238, + 548, + 295 + ], + "type": "text", + "content": ", which are then concatenated with the testing point coordinate and fed into the implicit sketch network to predict the sketch SDF value. After extending the sketches along the paths, we combine all parts for the final reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 64, + 312, + 123, + 388 + ], + "blocks": [ + { + "bbox": [ + 64, + 312, + 123, + 388 + ], + "lines": [ + { + "bbox": [ + 64, + 312, + 123, + 388 + ], + "spans": [ + { + "bbox": [ + 64, + 312, + 123, + 388 + ], + "type": "image", + "image_path": "64fe7203a5b631c604c7cf03634621b3e325522753e6b3c211bd21d291fae198.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 388, + 100, + 396 + ], + "lines": [ + { + "bbox": [ + 83, + 388, + 100, + 396 + ], + "spans": [ + { + "bbox": [ + 83, + 388, + 100, + 396 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 401, + 288, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 445 + ], + "type": "text", + "content": "Figure 4. 2D illustration of our Box+Path representation (right) to approximate the input shape (left), which can be decomposed into a set of feature primitives (middle, the green arrows indicate the sweeping paths)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 312, + 197, + 388 + ], + "blocks": [ + { + "bbox": [ + 138, + 312, + 197, + 388 + ], + "lines": [ + { + "bbox": [ + 138, + 312, + 197, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 197, + 388 + ], + "type": "image", + "image_path": "1fbdd88bc1a27ea19bd5f9bbf8248a677502e7ba8acb114eba69869654b27b5f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 388, + 193, + 396 + ], + "lines": [ + { + "bbox": [ + 140, + 388, + 193, + 396 + ], + "spans": [ + { + "bbox": [ + 140, + 388, + 193, + 396 + ], + "type": "text", + "content": "Feature primitives" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 213, + 312, + 271, + 388 + ], + "blocks": [ + { + "bbox": [ + 213, + 312, + 271, + 388 + ], + "lines": [ + { + "bbox": [ + 213, + 312, + 271, + 388 + ], + "spans": [ + { + "bbox": [ + 213, + 312, + 271, + 388 + ], + "type": "image", + "image_path": "0c64f2ae0f223d3741d6f3e3d667583745235a2a896926f81799696a09ba5b2c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 388, + 256, + 396 + ], + "lines": [ + { + "bbox": [ + 226, + 388, + 256, + 396 + ], + "spans": [ + { + "bbox": [ + 226, + 388, + 256, + 396 + ], + "type": "text", + "content": "Box+Path" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 453, + 130, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 130, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 130, + 467 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 472, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 545 + ], + "type": "text", + "content": "Here, we first present the box+path representation, which can be efficiently learned to approximate the input shape by stacking a set of box-like proxies along the sweeping path. Then, we define the neural sketch+path to refine the coarse reconstruction. Finally, we introduce the two-stage learning network for inferring this coarse-to-fine reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 552, + 190, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 190, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 190, + 564 + ], + "type": "text", + "content": "4.1. Box+Path Representation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "Previous unsupervised methods [23, 30, 31] for computing the Signed Distance Function (SDF) for geometric primitives (e.g., cubes, spheres, cylinders) rely on linear transformations that convert global testing point coordinates into local ones. However, the inherent non-linearity of curved paths limits the direct application of linear methods. We introduce a linear approximation method for the curved SDF computation, which we term Box+Path representation. We will first discuss the Box-Path in a 2D context to provide a more intuitive understanding before expanding it to 3D." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Box+Path in 2D. Specifically, we approximate one feature primitive " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " using a discrete set of " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " rigid box proxies," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": "i.e., " + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathcal{B}_i\\}_{i=1}^{N_s}" + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": ". The center position " + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": " of each box " + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": " is obtained by uniformly sampling the Bezier curve " + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\Upsilon(t)" + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 345, + 545, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 345, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 358, + 345, + 545, + 358 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i} = \\Upsilon (i / N _ {s}), \\quad i = 1, \\dots , N _ {s} \\tag {2}", + "image_path": "0a6e942b905688444005bf562271667f430da4e0bb23e37f31a9e8497895bf60.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "text", + "content": "The height " + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "text", + "content": " and the normal vector (or heading direction) " + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i" + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "text", + "content": " of each box are calculated as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 343, + 394, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 394, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 343, + 394, + 545, + 418 + ], + "type": "interline_equation", + "content": "h _ {i} = \\left| \\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i} \\right|, \\mathbf {n} _ {i} = \\frac {\\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i}}{\\left| \\mathbf {c} _ {i + 1} - \\mathbf {c} _ {i} \\right|} \\tag {3}", + "image_path": "d36f1da273c1b8a3fb506fbff1b24b88f8929bf3545be8f4ab0a9a1be2a2c409.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 423, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 447 + ], + "type": "text", + "content": "By incorporating a learnable width parameter " + }, + { + "bbox": [ + 305, + 423, + 545, + 447 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 423, + 545, + 447 + ], + "type": "text", + "content": ", the 2D Box+Path can be fully characterized." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": "Box+Path in 3D. When extended to 3D space, we introduce the tangent vector " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " and the binormal vector " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " to parameterize the orientation of the " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": "-th box proxy. To allow the sketch to rotate by a prescribed angle along the sweeping path, we should learn rotation angles " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\delta^u" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\delta^b" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " at the start and end points of the sweeping path. Then, the intermediate rotation angles " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " are obtained by linear interpolation along the sweeping path. Next, we compute the tangent vector " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " as the cross product of the normal " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " and the positive " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": "-axis unit vector " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_z" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": ". We then rotate " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i" + }, + { + "bbox": [ + 304, + 447, + 545, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 378, + 574, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 574, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 378, + 574, + 545, + 587 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {i} = R _ {\\mathbf {n} _ {i}} \\left(\\mathbf {n} _ {i} \\times \\mathbf {e} _ {z}, \\delta_ {i}\\right), \\tag {4}", + "image_path": "ffaf757d5c3b259ec4a6f96256f937ad43200eb6ba8355b109bfbbcff1f2de90.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{n}_i}(\\mathbf{v}, \\delta_i)" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": " represents a counter-clockwise rotation of vector " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": " with angle " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": ". The binormal vector " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": " is obtained as a cross product of " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i \\times \\mathbf{t}_i" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": ". Altogether, the triplet " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{n}_i, \\mathbf{t}_i, \\mathbf{b}_i\\}" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": " defines the orientation for each box along the curve path. The 3D Box+Path is fully characterized by incorporating learnable dimension parameters " + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "inline_equation", + "content": "l, w" + }, + { + "bbox": [ + 304, + 593, + 547, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 671, + 424, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 424, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 424, + 682 + ], + "type": "text", + "content": "4.2. Neural Sketch+Path" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We present differentiable methods for lofting or sweeping 2D sketches into 3D solid shapes to accurately reconstruct" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4674" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "the feature primitives. The shape of each 2D sketch is defined by an implicit representation, where we use " + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}" + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": " to represent the SDF of sketches. We now describe the differentiable operators that calculate the SDF of the 3D feature primitives corresponding to the sketches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": "Differentiable Sketch-Extrude. In the context of generating the geometry of feature primitives from sketch and extrude operations, " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{extrude}}^i" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": "-th sketch-extrude operation, " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i_z}" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": "-component of the " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": "-th position vector, and " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": " is the height of the " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": "-th feature primitive. We introduce a set of intermediate functions " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1, \\mathcal{M}_2" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_3" + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": " to facilitate a clearer formulation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 95, + 221, + 237, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 221, + 237, + 268 + ], + "spans": [ + { + "bbox": [ + 95, + 221, + 237, + 268 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {M} _ {1} \\left(\\hat {\\mathcal {S}} _ {\\mathbf {s k}} ^ {i}\\right) = \\max \\left(\\hat {\\mathcal {S}} _ {\\mathbf {s k}} ^ {i}, \\left| \\mathbf {x} _ {i _ {z}} \\right| - h _ {i}\\right), \\\\ \\mathcal {M} _ {2} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i}\\right) = \\max \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i}, 0\\right), \\\\ \\mathcal {M} _ {3} = \\max \\left(\\left| \\mathbf {x} _ {i _ {z}} \\right| - h _ {i}, 0\\right). \\\\ \\end{array}", + "image_path": "26a818c845471ef5f0c8012a206e0ae35dcd47f9ae44a890ff0175f8d27b2ca7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 275, + 275, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 275, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 275, + 288 + ], + "type": "text", + "content": "The SDF of the " + }, + { + "bbox": [ + 47, + 275, + 275, + 288 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 275, + 275, + 288 + ], + "type": "text", + "content": "-th extrusion primitive is then given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 287, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 287, + 314 + ], + "type": "interline_equation", + "content": "\\hat {S} _ {\\text {e x t r u d e}} ^ {i} = \\min \\left(\\mathcal {M} _ {1} \\left(\\hat {S} _ {\\mathrm {s k}} ^ {i}\\right), 0\\right) + \\sqrt {\\mathcal {M} _ {2} \\left(\\hat {S} _ {\\mathrm {s k}} ^ {i}\\right) ^ {2} + \\mathcal {M} _ {3} ^ {2}} \\tag {5}", + "image_path": "bc98106b441a3ae925be1b7d513a89ea472987efe8c288926cfe52de443bf382.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "content": "Differentiable Sketch-Loft. For the loft operation, we interpolate between two distinct sketches, " + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{u_i}" + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{l_i}" + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "content": ", at the upper and lower faces of the " + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "content": "-th feature primitive. Given the interpolation factor " + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "inline_equation", + "content": "|\\mathbf{x}_{i_z}| / h_i" + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "content": ", the linearly interpolated sketch, " + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{i_\\alpha}" + }, + { + "bbox": [ + 46, + 319, + 286, + 380 + ], + "type": "text", + "content": ", is computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 386, + 287, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 386, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 102, + 386, + 287, + 403 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}} = (1 - \\alpha) \\cdot \\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {l _ {i}} + \\alpha \\cdot \\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {u _ {i}}. \\tag {6}", + "image_path": "e88f8f61944acddf9f33af8558c3cb986d298e517e32483bf3dab5ae3adb2778.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "content": "We use the same intermediate variables as in the extrusion case, but with " + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{i}" + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "content": " replaced by " + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{i_{\\alpha}}" + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "content": ". The SDF of the " + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 408, + 287, + 445 + ], + "type": "text", + "content": "-th loft primitive can be expressed similarly to Eq. 5:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 452, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 452, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 59, + 452, + 287, + 472 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {S}} _ {\\text {l o f t}} ^ {i} = \\min \\left(\\mathcal {M} _ {1} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}}\\right), 0\\right) + \\sqrt {\\mathcal {M} _ {2} \\left(\\hat {\\mathcal {S}} _ {\\mathrm {s k}} ^ {i _ {\\alpha}}\\right) ^ {2} + \\mathcal {M} _ {3} ^ {2}}. \\tag {7}", + "image_path": "08ac69ef48922970ed14d660d2bc05be4cbbebd9cbe2ca3ba1071c69bc2cc7e6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "text", + "content": "Differentiable Sketch-Sweep. For sweep operation, we decompose it into a series of " + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "text", + "content": " extrusions. We denote the predicted SDF for the " + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "text", + "content": "-th extrude segment as " + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{extrude}}^i" + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "text", + "content": ". The SDF of the " + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 476, + 286, + 536 + ], + "type": "text", + "content": "-th sweep primitive is then obtained by taking the minimum of the SDFs of all extrusion segments:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 112, + 542, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 542, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 112, + 542, + 287, + 562 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {S}} _ {\\text {s w e e p}} ^ {i} = \\min _ {j \\in 1, \\dots , N _ {s}} \\hat {\\mathcal {S}} _ {\\text {e x t r u d e}} ^ {j}. \\tag {8}", + "image_path": "fd57f3966bb3d2da5c1c70b45020f5901a508dc9ef4e60c96fd2da56668088a1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 568, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 286, + 616 + ], + "type": "text", + "content": "Transformation to Parametric CAD. Following [23], we transform implicit sketches into parametric ones. With the control points of Bezier curves outputted by the network, we can achieve the sketch+path parametric reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 623, + 204, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 204, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 204, + 635 + ], + "type": "text", + "content": "4.3. Two-stage Network Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "content": "Stage 1: Box+Path learning. In this stage, the network predicts the control point coordinates " + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_1,\\dots ,\\mathbf{P}_4\\}" + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "content": " of the cubic Bezier curve and the parameters " + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\{l_i,w_i,\\delta^u,\\delta^b\\}" + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "content": " describing the " + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "content": " -th Box " + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "content": " Path " + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "content": " . These parameters subsequently inform the computation of the remaining parameters " + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\{h_i,\\mathbf{n}_i,\\mathbf{t}_i,\\mathbf{b}_i\\}" + }, + { + "bbox": [ + 46, + 641, + 286, + 714 + ], + "type": "text", + "content": " , as detailed in Sect. 4.1. To align the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "path of " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " with the target shape path, we first compute the SDF, " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{box}}^{ij}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", for each box at the " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "-th primitive and " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "-th segment, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "i \\in [1, N_p]" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "j \\in [1, N_s]" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " representing the total number of primitives and segments. Taking the union over the " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{box}}^{ij}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " yields the SDF for the overall box+path " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 386, + 152, + 545, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 152, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 386, + 152, + 545, + 186 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {S}} _ {\\mathcal {B}} = \\bigcup_ {i = 1} ^ {N _ {p}} \\bigcup_ {j = 1} ^ {N _ {s}} \\hat {\\mathcal {S}} _ {\\text {b o x}} ^ {i j}, \\tag {9}", + "image_path": "b3fdd9c577bc7dc60504c56e921530cbdb1ba5fb975ac141d4db5e5464503783.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\bigcup (x) = \\mathrm{softmin}(\\mu \\cdot x)\\cdot x" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": " is a hyperparameter. We then apply a differentiable converter " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": " similar to prior works [23, 31] to transform the " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": " into occupancy, yielding " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}_{\\mathcal{B}} = \\Psi (\\hat{S}_{\\mathcal{B}})" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\Psi (x) = (\\tanh (-\\beta \\cdot x) + 1) / 2" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": " is a hyperparameter. We then compute the Mean Squared Error (MSE) loss between " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": " and the ground truth occupancy " + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{O}}" + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 275, + 545, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 275, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 367, + 275, + 545, + 294 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {B}} = \\mathbb {E} _ {x \\in \\mathcal {X}} \\left[ | | \\hat {\\mathcal {O}} _ {\\mathcal {B}} - \\tilde {\\mathcal {O}} | | _ {2} ^ {2} \\right], \\tag {10}", + "image_path": "7a6526fb9a512f49e5bb0582608774ca9b4802d3b8f30f5a7176a24115e2108a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "text", + "content": " is the set of all testing points. The exclusive use of " + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "text", + "content": " may lead to potential challenges: (1) the emergence of the curve with acute angles, and (2) the dimensions of " + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 297, + 545, + 380 + ], + "type": "text", + "content": " vastly outstrip the path length. Both increase the risk of shape self-intersections, which are considered invalid in CAD. To address these issues, we introduce two additional regularization terms," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 327, + 386, + 545, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 386, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 327, + 386, + 545, + 419 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s m}} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\left(1 - \\min _ {j = 1, \\dots , N _ {s}} \\left(\\mathbf {n} _ {i, j} \\cdot \\mathbf {n} _ {i, j + 1}\\right)\\right) \\tag {11}", + "image_path": "ae6e91bed5e21f54a5889cbe58ce7a30f6c2401e77f8d83cb67d86c50bc3a495.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 328, + 422, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 328, + 422, + 545, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l w}} = \\frac {1}{2 N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\operatorname {R e L U} \\left(h _ {i} + w _ {i} - 2 \\cdot \\Theta\\right), \\tag {12}", + "image_path": "02cf78a55927eeab9acc890c2f8a1d50dff9317354fecc776df4ced6ecb2990d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sm}}" + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": " penalizes the sharpest curvature in the curve by maximizing the inner product between adjacent " + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lw}}" + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": " penalizes any length or width greater than a threshold " + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": ". Combine with balance factors " + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{1}, \\lambda_{2}" + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": ", the overall loss of Stage 1 is:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 364, + 527, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 527, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 364, + 527, + 545, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b o x}} = \\mathcal {L} _ {\\mathcal {B}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {s m}} + \\lambda_ {2}. \\mathcal {L} _ {\\mathrm {l w}} \\tag {13}", + "image_path": "85ede053162991b3b046b60cef742dfc93a4df87a5067704ede82d8e13a15915.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": "Stage 2: Implicit sketch learning. Using the path learned in Stage 1, we progress towards fine reconstruction in Stage 2 by learning sketches. We initially utilize a Multi-Layer Perceptron (MLP) to convert the feature " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": " local features " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\prime}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": ". These local features are then concatenated with the local " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\{x,y\\}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": " coordinates of each point and fed into an implicit sketch network, which is essentially an MLP predicting the SDF value " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{i}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": ". For the sketch-loft operation, two implicit sketch networks are used to learn " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{u_i}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{sk}}^{l_i}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": ". Following the computation methods outlined in Sect. 4.2, we can derive the predicted overall occupancy value " + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}_{\\mathrm{F}}" + }, + { + "bbox": [ + 304, + 545, + 545, + 679 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 328, + 684, + 521, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 684, + 521, + 716 + ], + "spans": [ + { + "bbox": [ + 328, + 684, + 521, + 716 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {O}} _ {\\mathrm {F}} = \\left\\{ \\begin{array}{l l} \\Psi (\\bigcup_ {i = 1} ^ {N _ {p}} \\hat {S} _ {\\text {s w e e p}} ^ {i}), & \\text {f o r s w e e p o p e r a t i o n} \\\\ \\Psi (\\bigcup_ {i = 1} ^ {N _ {p}} \\hat {S} _ {\\text {l o f t}} ^ {i}), & \\text {f o r l o f t o p e r a t i o n .} \\end{array} \\right.", + "image_path": "444641aba71f4a8c27c560f5b9a9132bcc09f63919da46dc4f0295fc5c0fdde4.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4675" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 94 + ], + "type": "text", + "content": "Finally, an MSE loss " + }, + { + "bbox": [ + 46, + 72, + 287, + 94 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 94 + ], + "type": "text", + "content": " is computed against the ground truth " + }, + { + "bbox": [ + 46, + 72, + 287, + 94 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 72, + 287, + 94 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 92, + 287, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 92, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 107, + 92, + 287, + 112 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e c}} = \\mathbb {E} _ {x \\in \\mathcal {X}} \\left[ \\left| \\left| \\hat {\\mathcal {O}} _ {\\mathrm {F}} - \\tilde {\\mathcal {O}} \\right| \\right| _ {2} ^ {2} \\right]. \\tag {14}", + "image_path": "50e309af2e22ce365b1042429826eba5d0d2cc5193972ea7071a56ae0e5e1b6f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": "The direct computation of " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " over the entire point set " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " requires the repetition of " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{extrude}}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{s}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " times, which is time-consuming. To accelerate the training of the implicit sketch network, we propose a real-time sampling strategy, exploiting the sketch consistency trait inherent in the sweep operation. Specifically, we sample grid-like testing points " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " in the base sketch plane of the " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": "-th path " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\Upsilon_{i}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " obtained in the first stage. The occupancy value of " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " is determined by inputting it into the implicit sketch network and applying the converter " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " is moved along the curve for " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "N_{s} - 1" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " times, with its occupancy value directly replicated, yielding " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}_{\\mathrm{Fi}}^{*}" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": ". The ground truth occupancy value " + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{O}}_i^*" + }, + { + "bbox": [ + 46, + 116, + 289, + 310 + ], + "type": "text", + "content": " is provided by the input voxel. After performing this procedure for all paths and conducting a union operation, the loss function corresponding to real-time sampling is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 317, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 317, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 287, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e c}} ^ {*} = \\mathbb {E} _ {y \\in \\mathcal {Y}} \\left[ | | \\hat {\\mathcal {O}} _ {\\mathrm {F}} ^ {*} - \\tilde {\\mathcal {O}} ^ {*} | | _ {2} ^ {2} \\right], \\tag {15}", + "image_path": "962b4c1fcd6201bcd91c9ea1e3d438452ceb1fc44142d031dfab54f36b002fcc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "text", + "content": " represents the set of testing points obtained through real-time sampling, and " + }, + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "text", + "content": " denotes computation using real-time sampling data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "type": "text", + "content": "Note that the SfmCAD is designed to choose between learning either loft or sweep operations, which is determined by its hyperparameters. However, enabling the network to concurrently learn both operations is also viable, with details provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 452, + 180, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 180, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 180, + 464 + ], + "type": "text", + "content": "5. Results and Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "type": "text", + "content": "We evaluate the performance of SfmCAD on two reconstruction tasks: (1) holistic shapes of different categories and (2) semantically segmented shapes. We show the output editability of SfmCAD in the supplemental materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 527, + 97, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 97, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 97, + 540 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": "Implementation details. We implemented SfmCAD in PyTorch by using an NVIDIA® TITAN RTX GPU. The Adam optimizer [15] is utilized with a learning rate 1e-4 and beta values of (0.5, 0.99). The hyperparameters in Sect. 4.3 are set to: " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mu = 20" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\beta = 50" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\Theta = 0.1" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.05" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 0.05" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": ". In our evaluations, we employ the best-performing operation (Loft/Sweep) for each category of holistic shapes and every part of segmented shapes. SfmCAD trains for 500 epochs per stage with a batch size of 24, followed by fine-tuning each test shape for 200 iterations per stage. All baseline methods are trained for 1,000 epochs and fine-tuned for 400 iterations for a fair comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Evaluation metrics. We utilize three common metrics for 3D shape reconstruction: Chamfer Distance (CD), Normal" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 544, + 293 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 544, + 293 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 544, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 544, + 293 + ], + "type": "image", + "image_path": "efdf5ab4d3d5aa26ec6a92047f1176e01467ba81e248c0d24532a9f611a99b02.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 304, + 509, + 316 + ], + "lines": [ + { + "bbox": [ + 342, + 304, + 509, + 316 + ], + "spans": [ + { + "bbox": [ + 342, + 304, + 509, + 316 + ], + "type": "text", + "content": "Figure 5. Visual comparison on ABC dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 329, + 347, + 522, + 426 + ], + "blocks": [ + { + "bbox": [ + 334, + 327, + 517, + 339 + ], + "lines": [ + { + "bbox": [ + 334, + 327, + 517, + 339 + ], + "spans": [ + { + "bbox": [ + 334, + 327, + 517, + 339 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on ABC dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 347, + 522, + 426 + ], + "lines": [ + { + "bbox": [ + 329, + 347, + 522, + 426 + ], + "spans": [ + { + "bbox": [ + 329, + 347, + 522, + 426 + ], + "type": "table", + "html": "
MethodsCD↓ECD↓NC↑
UCSG-Net [14]1.23321.7860.866
CSG-Stump [30]0.6717.7510.892
ExtrudeNet [31]0.5197.1110.885
SECAD-Net [23]0.5067.2860.884
Ours0.3955.0380.919
", + "image_path": "35307c2921a3ca012040d6e698a26714fb361e804b6cfa87565cdae00f326cca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "text", + "content": "Consistency (NC), and Edge Chamfer Distance (ECD). All experiments generate the reconstructed shapes via Marching Cubes at a resolution of " + }, + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "inline_equation", + "content": "256^3" + }, + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "text", + "content": ". To calculate CD, we sample 8,192 points from the ground truth and predicted shape surfaces. For ECD, we employ a threshold of 0.1 for normal cross-products to identify points close to the edges on the ground truth surface and 0.5 for the predicted surface, considering the lack of sharpness in shapes generated by Marching Cubes. For better visual clarity, we magnify the CD and ECD by factors of " + }, + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "inline_equation", + "content": "10^3" + }, + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "inline_equation", + "content": "10^2" + }, + { + "bbox": [ + 304, + 445, + 545, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 574, + 495, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 495, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 495, + 588 + ], + "type": "text", + "content": "5.2. Reconstruction from Holistic Shape" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 593, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 665 + ], + "type": "text", + "content": "We compare SfmCAD with UCSG-Net [14], CSG-Stump [30], ExtrudeNet [31] and SECAD-Net [23]. We test these methods on holistic shapes from different sources, including CAD parts in ABC [16], ShapeNet objects [1], and a synthetic tree dataset. The comparison against CAPRI-Net [49] is provided in the supplementary materials." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on ABC dataset. We employ the voxel grids and sampling point data provided by [49]. The dataset included 5,000 groups of data for training and 1,000 for testing. We select 100 shape subsets from the test set for eval" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4676" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 308 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 308 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 308 + ], + "type": "image", + "image_path": "4c9f492e70ed69c1a36b65b7620e294165e499b64318137444c87d16d57e9196.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 312, + 259, + 324 + ], + "lines": [ + { + "bbox": [ + 75, + 312, + 259, + 324 + ], + "spans": [ + { + "bbox": [ + 75, + 312, + 259, + 324 + ], + "type": "text", + "content": "Figure 6. Visual comparison on ShapeNet dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 357, + 261, + 435 + ], + "blocks": [ + { + "bbox": [ + 67, + 336, + 266, + 348 + ], + "lines": [ + { + "bbox": [ + 67, + 336, + 266, + 348 + ], + "spans": [ + { + "bbox": [ + 67, + 336, + 266, + 348 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on ShapeNet dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 357, + 261, + 435 + ], + "lines": [ + { + "bbox": [ + 70, + 357, + 261, + 435 + ], + "spans": [ + { + "bbox": [ + 70, + 357, + 261, + 435 + ], + "type": "table", + "html": "
MethodsCD↓ECD↓NC↑
UCSG-Net [14]2.11123.6160.815
CSG-Stump [30]2.16817.5920.818
ExtrudeNet [31]0.91014.1100.849
SECAD-Net [23]0.83614.8180.837
Ours0.62614.0960.867
", + "image_path": "9b10f921e597eea6b9b57a877317b589ec5cc0e2ca03aec9a5b35fe4d43c2d01.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 467, + 287, + 510 + ], + "blocks": [ + { + "bbox": [ + 76, + 447, + 257, + 458 + ], + "lines": [ + { + "bbox": [ + 76, + 447, + 257, + 458 + ], + "spans": [ + { + "bbox": [ + 76, + 447, + 257, + 458 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on Tree dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "table", + "html": "
MethodsExtrudeNet [31]SECAD-Net [23]Ours
CD↓7.2422.1970.686
NC↑0.6080.6220.722
", + "image_path": "58b5469b325f2844f5da6839f4ee3694aa9fa445811bcb5128d2696a0543266f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 532, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 555 + ], + "type": "text", + "content": "uation. The visual results are shown in Fig. 5, while the quantitative results are presented in Tab. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": "Evaluation on ShapeNet dataset. We utilize three categories from ShapeNet: chair, table, and display. The voxel grids and sampling points are provided by [3]. We selected 120 shapes for evaluation, with 40 from each category. Visual and quantitative results are in Fig. 6 and Tab. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Evaluation on tree branches. We use a procedural modeling approach of Palubicki et al. [26] to generate tree meshes voxelized into a resolution of " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ". We adopt the method from [2] to obtain sampling points and occupancy values. This process results in 5,000 data groups, of which 4,500 are used for training and 500 for testing. We train all methods using increasingly denser voxel resolutions " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(16^{3},32^{3},64^{3})" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and select 50 trees for numerical evaluation." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 545, + 241 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 241 + ], + "type": "image", + "image_path": "cc7741dd422e4c70e6ec5e185ad0b0e76423ff7409b14cae5c587b893a65149d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 244, + 508, + 256 + ], + "lines": [ + { + "bbox": [ + 343, + 244, + 508, + 256 + ], + "spans": [ + { + "bbox": [ + 343, + 244, + 508, + 256 + ], + "type": "text", + "content": "Figure 7. Visual comparison on Tree dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 277, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 326 + ], + "type": "text", + "content": "tion. Note given the cylindrical nature of tree branches, we adapt our Box+Path representation by substituting boxes with cylinders for the final reconstruction (see Fig. 7 and Tab. 3 for visual and quantitative results)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 326, + 546, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 546, + 447 + ], + "type": "text", + "content": "Summary of comparative analysis. As demonstrated in Tables 1, 2, and 3, SfmCAD consistently outperforms other methods in terms of CD, ECD, and NC, suggesting that the sketch+path representation of SfmCAD offers more flexible shape expression capabilities than the geometric primitives used by UCSG-Net and CSG-Stump, as well as the simple sketch+extrusion operations employed by ExtrudeNet and SECAD. Figures 5, 6, and 7 further show SfmCAD's ability to generate detailed reconstructions (such as thin structures) across diverse shape categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 454, + 511, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 454, + 511, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 454, + 511, + 468 + ], + "type": "text", + "content": "5.3. Reconstruction from Segmented Shape" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 557 + ], + "type": "text", + "content": "Semantic segmentation provides information for reconstructing complex shapes. Previous CAD reconstruction methods [23, 31] could not utilize this information because their single-type primitives lack the flexibility to represent each part independently. In contrast, SfmCAD successfully reconstructs multiple primitives with its sketch+path representation capability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "text", + "content": "We show the task of unsupervised CAD reconstruction utilizing semantic segmentation. The evaluation is conducted on PartNet dataset [48], which provides instance-level 3D part information. We utilize three categories: chair, table, and trashcan. We voxelize the segmented parts and sampled occupancy values at resolutions " + }, + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "inline_equation", + "content": "16^{3}, 32^{3}, 64^{3}" + }, + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "text", + "content": " following the scheme by [2]. This process yields a total of 12,000 chair parts (1,229 complete), 12,000 table parts (1,491 complete), and 3,000 trashcan parts (234 complete). The train-test split for all categories was set as 5:1. We selected 90 complete shapes, including 40 tables, 40 chairs, and 10 trashcans, for evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "All methods were trained on part shapes using increas" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4677" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 244 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 244 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 244 + ], + "type": "image", + "image_path": "a8f8675dfb12f8055225e7b780ce587fc1aab2e348791fb888e141a6bb401604.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 250, + 255, + 262 + ], + "lines": [ + { + "bbox": [ + 79, + 250, + 255, + 262 + ], + "spans": [ + { + "bbox": [ + 79, + 250, + 255, + 262 + ], + "type": "text", + "content": "Figure 8. Visual comparison on PartNet dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 293, + 261, + 346 + ], + "blocks": [ + { + "bbox": [ + 70, + 272, + 263, + 283 + ], + "lines": [ + { + "bbox": [ + 70, + 272, + 263, + 283 + ], + "spans": [ + { + "bbox": [ + 70, + 272, + 263, + 283 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on PartNet dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 293, + 261, + 346 + ], + "lines": [ + { + "bbox": [ + 70, + 293, + 261, + 346 + ], + "spans": [ + { + "bbox": [ + 70, + 293, + 261, + 346 + ], + "type": "table", + "html": "
MethodsCD↓ECD↓NC↑
ExtrudeNet [31]3.2510.5490.783
SECAD-Net [23]1.8880.4720.818
Ours1.4100.4460.814
", + "image_path": "105c23a70844c235c472c4a98d9ffe2cca4c1c83afcf96f59e31bdc4999a8cac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 376, + 287, + 428 + ], + "blocks": [ + { + "bbox": [ + 55, + 356, + 278, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 278, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 278, + 368 + ], + "type": "text", + "content": "Table 5. Ablation study on sweep twist angles and loss terms." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 376, + 287, + 428 + ], + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 428 + ], + "type": "table", + "html": "
Settings- (δu, δl)-δl- Llw- LsmOurs
CD↓4.0923.3674.8043.2683.142
ECD↓0.1730.1390.1640.1540.133
NC↑0.6550.7140.6850.7320.776
", + "image_path": "208e048c6289a1c01e5084e870c110b80cfd3cc7b479a2dee1d2e3db088303a8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 448, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 508 + ], + "type": "text", + "content": "ingly denser voxel resolutions " + }, + { + "bbox": [ + 46, + 448, + 287, + 508 + ], + "type": "inline_equation", + "content": "(16^{3}, 32^{3}, 64^{3})" + }, + { + "bbox": [ + 46, + 448, + 287, + 508 + ], + "type": "text", + "content": ". For each part, we set ExtrudeNet [31] and SECAD-Net [23] to output an extruded cylinder and SfmCAD to output a sketch+path shape. After reconstructing all parts, they were reassembled to obtain the final CAD models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 509, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 616 + ], + "type": "text", + "content": "The quantitative results in Tab. 4 indicate that our method outperforms others. Compared to the closest competitor, SECAD-Net, SfmCAD improves the CD by " + }, + { + "bbox": [ + 46, + 509, + 287, + 616 + ], + "type": "inline_equation", + "content": "25.3\\%" + }, + { + "bbox": [ + 46, + 509, + 287, + 616 + ], + "type": "text", + "content": " and the ECD by " + }, + { + "bbox": [ + 46, + 509, + 287, + 616 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 46, + 509, + 287, + 616 + ], + "type": "text", + "content": ". Fig. 8 illustrates that the sketch+extrusion reconstruction of ExtrudeNet and SECAD can only characterize the nonlinear shapes through sketches. In contrast, SfmCAD can model these complex shapes using sketches and free-form paths, enabling a more refined reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "We conducted an ablation study by using 40 box+path reconstructions on chair class from PartNet to evaluate the efficiency of our network design and loss terms. Initially, we disable the two sweep twist angles " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\delta^u" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\delta^l" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\delta^l" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": ". Then we separately discarded " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sm}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lw}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": ". The visual results in Fig. 9 and the quantitative results in Tab. 5 suggest that" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 69, + 541, + 241 + ], + "blocks": [ + { + "bbox": [ + 313, + 69, + 541, + 241 + ], + "lines": [ + { + "bbox": [ + 313, + 69, + 541, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 69, + 541, + 241 + ], + "type": "image", + "image_path": "59112db521e7246d8c7e561bac2e5cad7b7329363fce054f73854cf20fc3b3d9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 247, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 546, + 281 + ], + "type": "text", + "content": "Figure 9. Visual comparison between reconstruction results for ablation study. From top to bottom, each column shows the predicted 3D path, vector " + }, + { + "bbox": [ + 305, + 247, + 546, + 281 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{n},\\mathbf{t},\\mathbf{b}\\}" + }, + { + "bbox": [ + 305, + 247, + 546, + 281 + ], + "type": "text", + "content": ", and the box+path representation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "type": "text", + "content": "removing any of these settings negatively affects the reconstruction performance of SfmCAD." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 350, + 471, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 471, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 471, + 363 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 373, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 504 + ], + "type": "text", + "content": "We have presented an unsupervised neural network for CAD reconstruction that infers high-level CAD modeling operations. We have also introduced a unique sketch+path representation that encapsulates the local geometric details and overall structural information of the 3D shape. A novel two-stage learning strategy, operating in a coarse-to-fine manner, was proposed to address the complexity of simultaneous sketch and path learning, thus improving the efficiency of network training. We evaluated SfmCAD using several different categories of objects and demonstrated its advantages by comparing it to state-of-the-art methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 507, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 567 + ], + "type": "text", + "content": "In future work, we plan to explore using 2D sketches and 3D part templates to enhance the efficiency of unsupervised CAD reconstruction. Besides, given the manipulable property of SfmCAD's outputs, extending this work for generative CAD design or modeling would be meaningful." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgments. We thank the anonymous reviewers for their valuable suggestions. This work is partially funded by the Strategic Priority Research Program of the Chinese Academy of Sciences (XDB0640000), Guangdong Science and Technology Program (2023B1515120026), National Natural Science Foundation of China (62172416, 62172415, U22B2034, 62262043, 62365014), and Youth Innovation Promotion Association of the Chinese Academy of Sciences (2022131). This work was supported by NRCS grant #NR233A750004G044 to Benes. The findings and conclusions should not be construed to represent any agency determination or policy." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4678" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 179 + ], + "type": "text", + "content": "[2] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5939-5948, 2019. 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "text", + "content": "[3] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. BSP-Net: Generating compact meshes via binary space partitioning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 45–54, 2020. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "type": "text", + "content": "[4] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. In ACM SIGGRAPH 2004 Papers, pages 905–914. 2004. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 315 + ], + "type": "text", + "content": "[5] Tao Du, Jeevana Priya Inala, Yewen Pu, Andrew Spielberg, Adriana Schulz, Daniela Rus, Armando Solar-Lezama, and Wojciech Matusik. Inversecsg: Automatic conversion of 3d models to csg trees. ACM Trans. Graph., 37(6):1-16, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 372 + ], + "type": "text", + "content": "[6] Elona Dupont, Kseniya Cherenkova, Anis Kacem, Sk Aziz Ali, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Cadops-net: Jointly learning cad operation types and steps from boundary-representations. In International Conference on 3D Vision (3DV), pages 114-123. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "text", + "content": "[7] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems (NeurIPS), pages 5885–5897, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 452 + ], + "type": "text", + "content": "[8] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Trans. Graph., 39(6):1-16, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 496 + ], + "type": "text", + "content": "[9] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: CAD reconstruction by B-rep chain complex generation. ACM Trans. Graph., 41(4): 1-18, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "text", + "content": "[10] Rostislav Hulik, Michal Spanel, Pavel Smrz, and Zdenek Materna. Continuous plane detection in point-cloud data based on 3d hough transform. Journal of visual communication and image representation, 25(1):86-97, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[11] Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G Lambourne, Karl DD Willis, Thomas Davies, Hooman Shayani, and Nigel Morris. Uv-net: Learning from boundary representations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11703-11712, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[12] Pradeep Kumar Jayaraman, Joseph G Lambourne, Nishkrit Desai, Karl DD Willis, Aditya Sanghi, and Nigel JW Morris. Solidgen: An autoregressive model for direct b-rep synthesis. arXiv preprint arXiv:2203.13944, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[13] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Kacper Kania, Maciej Zieba, and Tomasz Kajdanowicz. UCSG-NET-Unsupervised discovering of constructive solid" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "text", + "content": "geometry tree. In Advances in Neural Information Processing Systems (NeurIPS), pages 8776-8786, 2020. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "type": "text", + "content": "[16] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 9601-9611, 2019. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "text", + "content": "[17] Joseph G Lambourne, Karl DD Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. BRepNet: A topological message passing system for solid models. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 12773-12782, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "text", + "content": "[18] Joseph George Lambourne, Karl Willis, Pradeep Kumar Jayaraman, Longfei Zhang, Aditya Sanghi, and Kamal Rahimi Malekshan. Reconstructing editable prismatic cad from rounded voxel models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "text", + "content": "[19] Eric-Tuan Lé, Minhyuk Sung, Duygu Ceylan, Radomir Mech, Tamy Boubekeur, and Niloy J Mitra. CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In IEEE International Conference on Computer Vision (ICCV), pages 7457-7466, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "text", + "content": "[20] Bosheng Li, Jacek Kałuzny, Jonathan Klein, Dominik L. Michels, Wojtek Pałubicki, Bedrich Benes, and Soren Pirk. Learning to reconstruct botanical trees from single images. ACM Trans. Grap., 40(6):1-15, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "text", + "content": "[21] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Sketch2cad: Sequential cad modeling by sketching in context. ACM Trans. Graph., 39(6):1-14, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "text", + "content": "[22] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3d point clouds. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 2652-2660, 2019. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "text", + "content": "[23] Pu Li, Jianwei Guo, Xiaopeng Zhang, and Dong-Ming Yan. Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 16816-16826, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[24] Yuanqi Li, Shun Liu, Xinran Yang, Jianwei Guo, Jie Guo, and Yanwen Guo. Surface and edge detection for primitive fitting of point clouds. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[25] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 4460-4470, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[26] Wojciech Palubicki, Kipp Horel, Steven Longay, Adam Runions, Brendan Lane, Radomír Měch, and Przemyslaw Prusinkiewicz. Self-organizing tree models for image synthesis. ACM Trans. Graph., 28(3):1–10, 2009. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. In Advances in Neural Information Processing Systems (NeurIPS), pages 5077-5088, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4679" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 165-174, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[29] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 10344-10353, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[30] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, et al. CSG-stump: A learning friendly CSG-like representation for interpretable shape parsing. In IEEE International Conference on Computer Vision (ICCV), pages 12478-12487, 2021. 2, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 306 + ], + "type": "text", + "content": "[31] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude for shape parsing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 482-498, 2022. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 341 + ], + "type": "text", + "content": "[32] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient RANSAC for point-cloud shape detection. In Comput. Graph. Forum, pages 214-226, 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "text", + "content": "[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "type": "text", + "content": "[34] Jami J Shah and Martti Mäntylä. Parametric and feature-based CAD/CAM: concepts, techniques, and applications. John Wiley & Sons, 1995. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 443 + ], + "type": "text", + "content": "[35] Vadim Shapiro. Solid modeling. Handbook of computer aided geometric design, 20:473-518, 2002. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "text", + "content": "[36] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 5515-5523, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[37] Gopal Sharma, Difan Liu, Subhransu Maji, Evangelos Kalogerakis, Siddhartha Chaudhuri, and Radomir Mech. ParSeNet: A parametric surface fitting network for 3d point clouds. In European Conference on Computer Vision (ECCV), pages 261-276, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "text", + "content": "[38] Mikaela Angelina Uy, Yen-Yu Chang, Minhyuk Sung, Purvi Goel, Joseph G Lambourne, Tolga Birdal, and Leonidas J Guibas. Point2Cyl: Reverse engineering 3D objects from point clouds to extrusion cylinders. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11850-11860, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[39] Tamas Varady and Ralph Martin. Reverse engineering. In Handbook of Computer Aided Geometric Design, pages 651-681. 2002. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[40] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. PIE-Net: Parametric inference of point cloud edges. Advances in Neural Information Processing Systems (NeurIPS), pages 20167–20178, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[41] Karl DD Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G Lambourne, Armando Solar-Lezama, and Wo" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 531 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "jciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM Trans. Graph., 40(4):1-24, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[42] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In IEEE International Conference on Computer Vision (ICCV), pages 6772–6782, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[43] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl DD Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 6062-6070, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[44] Xiang Xu, Karl DD Willis, Joseph G Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, and Yasutaka Furukawa. Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In International Conference on Machine Learning, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 284 + ], + "type": "text", + "content": "[45] Fujio Yamaguchi. Curves and surfaces in computer aided geometric design. Springer Science & Business Media, 2012. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[46] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "text", + "content": "[47] Siming Yan, Zhenpei Yang, Chongyang Ma, Haibin Huang, Etienne Vouga, and Qixing Huang. Hpnet: Deep primitive segmentation using hybrid representations. In IEEE International Conference on Computer Vision (ICCV), pages 2753-2762, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[48] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, page to appear, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "text", + "content": "[49] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Comp. Vision and Pat. Rec. (CVPR), pages 11768–11778, 2022. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "type": "text", + "content": "[50] Fenggen Yu, Qimin Chen, Maham Tanveer, Ali Mahdavi Amiri, and Hao Zhang. " + }, + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{csg}" + }, + { + "bbox": [ + 307, + 476, + 545, + 531 + ], + "type": "text", + "content": ": Unsupervised learning of compact csg trees with dual complements and dropouts. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4680" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_content_list.json b/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e0522cf6493a56f4f3088ea428457e676872dd01 --- /dev/null +++ b/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_content_list.json @@ -0,0 +1,2156 @@ +[ + { + "type": "text", + "text": "Shadow Generation for Composite Image Using Diffusion Model", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingyang Liu1, Junqi You1, Jianting Wang1, Xinhao Tao1, Bo Zhang1, Li Niu1,2* \n1 Shanghai Jiao Tong University 2 miguo.ai", + "bbox": [ + 166, + 178, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ narumimaria, yjqsjtu2022, gloryl1299, taoxinhao, bo-zhang, ustcnewly}@sjtu.edu.cn", + "bbox": [ + 142, + 218, + 818, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the realm of image composition, generating realistic shadow for the inserted foreground remains a formidable challenge. Previous works have developed image-to-image translation models which are trained on paired training data. However, they are struggling to generate shadows with accurate shapes and intensities, hindered by data scarcity and inherent task complexity. In this paper, we resort to foundation model with rich prior knowledge of natural shadow images. Specifically, we first adapt ControlNet to our task and then propose intensity modulation modules to improve the shadow intensity. Moreover, we extend the small-scale DESOBA dataset to DESOBAv2 using a novel data acquisition pipeline. Experimental results on both DESOBA and DESOBAv2 datasets as well as real composite images demonstrate the superior capability of our model for shadow generation task. The dataset, code, and model are released at https://github.com/bcmi/Object-Shadow-Generation-Dataset-DESOBAv2.", + "bbox": [ + 75, + 301, + 473, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 604, + 209, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image composition [28] aims to merge the foreground of one image with another background image to produce a composite image, which has a wide range of applications like virtual reality, artistic creation, and E-commerce. Simply pasting the foreground onto the background often results in visual inconsistencies, including the incompatible illumination between foreground and background [3], lack of foreground shadow/reflection [12, 34], and so on. In this paper, we focus on the shadow issue, i.e., the inserted foreground does not have plausible shadow on the background, which could significantly degrade the realism and quality of composite image.", + "bbox": [ + 75, + 631, + 468, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As illustrated in Figure 1, shadow generation is a challenging task because the foreground shadow is determined by many complicated factors like the lighting information and the geometry of foreground/background. The exist-", + "bbox": [ + 75, + 813, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19751b68b323e2547de1ce8434ae68dc5be14c2111fec618e4da7c3f1cabc42a.jpg", + "image_caption": [ + "Figure 1. A composite image can be obtained by pasting the foreground on the background. Shadow generation aims to generate plausible shadow for the inserted foreground in the composite image to produce a more realistic image." + ], + "image_footnote": [], + "bbox": [ + 511, + 268, + 879, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing shadow generation methods can be divided into rendering based methods [34-36] and non-rendering based methods [12, 22, 53]. Rendering based methods usually impose strict assumptions on the geometry and lighting, which could hardly be satisfied in real-world scenarios. Besides, [35, 36] require users to specify the lighting information, which hinders its direct application in our task. Non-rendering based methods usually train an image-to-image translation network, based on pairs of composite images without foreground shadows and real images with foreground shadows. However, due to the training data scarcity and task difficulty, these methods are struggling to generate shadows with reasonable shapes and intensities.", + "bbox": [ + 496, + 549, + 890, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, foundation model (e.g., stable diffusion [32]) pretrained on large-scale dataset has demonstrated unprecedented potential for image generation and editing. In previous works [44, 48] on object-guided inpainting or composition, they show that the generated foregrounds are accompanied by shadows even without considering the shadow issue, probably because of the rich prior knowledge of natural shadow images in foundation model. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly.", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8121", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We build our method upon conditional foundation model [52] and propose several key innovations. First, we modify the control encoder input and the noise loss to fit our task. Then, we observe that the generated shadow intensity (the level of darkness) is unsatisfactory. Especially when the background objects has shadows, the intensity inconsistency between foreground shadow and background shadows make the whole image unrealistic. Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, the denoising U-Net is modified to output both noise map and foreground shadow mask. The intensity encoder takes in the composite image and background shadow mask, producing the scale/bias to modulate the predicted noise within the foreground shadow region. Finally, we devise a post-processing network to rectify the color shift and background variation.", + "bbox": [ + 80, + 90, + 467, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The model training requires abundant pairs of composite images without foreground shadows and real images with foreground shadows. The existing real-world shadow generation dataset DESOBA [12] is limited by scale (i.e., 1,012 real images and 3,623 pairs) due to the high cost of manual shadow removal, which is insufficient to train our model. To ensure sufficient supervision, we design a novel data construction pipeline, which extends DESOBA to DESOBAv2 (i.e., 21,575 real images and 28,573 pairs) using object-shadow detection and inpainting techniques. Specifically, we first collect a large number of real-world images with one or more object-shadow pairs. Then, we use pretrained object-shadow detection model [41] to predict object and shadow masks for object-shadow pairs. Next, we apply pretrained inpainting model [32] to inpaint the detected shadow regions to get desshadowed images. Finally, based on real images and desshadowed images, we construct pairs of synthetic composite images and ground-truth target images.", + "bbox": [ + 80, + 333, + 467, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on both DESOBAv2 and DESOBA datasets. The results reveal remarkable improvement in shadow generation task, after leveraging the benefits of large-scale data and foundation model. Our main contributions can be summarized as follows: 1) We contribute DESOBAv2, a large-scale real-world shadow generation dataset, which could greatly facilitate the shadow generation task. 2) We propose a cutting-edge diffusion model specifically designed to produce shadows for the composite foregrounds. 3) Through comprehensive experiments, we validate the efficacy of our dataset construction pipeline and the superiority of our proposed model.", + "bbox": [ + 80, + 607, + 467, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 805, + 215, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Composition", + "text_level": 1, + "bbox": [ + 80, + 830, + 261, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image composition aims to overlay a foreground object on a background image to yield a composite result [20, 22, 42, 46, 47]. Previous research works have tackled differ", + "bbox": [ + 80, + 856, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ent issues that can compromise the quality of composite images. For instance, image blending methods [31, 42, 49, 51] target at combining the foreground and background seamlessly. Image harmonization methods [3-6, 40] aim to rectify the illumination disparity between foreground and background. Nonetheless, the above methods largely overlook the shadow cast by the foreground onto the background. Recently, generative image composition methods [38, 44, 48] can insert a foreground object into a bounding box in the background and the inserted object is likely to have shadow effect. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly.", + "bbox": [ + 503, + 90, + 888, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Shadow Generation", + "text_level": 1, + "bbox": [ + 503, + 297, + 686, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, the goal of shadow generation task is generating plausible shadow for the composite foreground. Existing methods can be broadly categorized into rendering based methods and non-rendering based methods. The rendering based methods necessitate a comprehensive understanding of factors like illumination, reflectance, material properties, and scene geometry to produce shadows for the inserted objects. However, such detailed knowledge relies on user input [15, 16, 21, 35, 36] or model prediction [1, 7, 19, 50], which is either labor-intensive or unreliable [53]. For example, [35, 36] could produce compelling results with user control. However, in the composite image, the lighting information should be inferred automatically from background instead of requested by users.", + "bbox": [ + 503, + 321, + 888, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-rendering based methods [12, 22, 25, 53] aim to translate an input composite image without foreground shadow to an output with foreground shadow, bypassing the need for explicit knowledge of the aforementioned factors. For instance, ShadowGAN [53] utilizes both global and local conditional discriminator to enhance the realism of generated shadows. ARShadowGAN [22] emphasizes the importance of background shadow and uses it to guide foreground shadow generation. SGRNet [12] encourages the information exchange between foreground and background, and employs a classic illumination model for better shadow effect. The work [25] produces multiple underexposure images and fuses them to get the final shadow region. DMASNet [39] decomposes shadow mask prediction into box prediction and shape prediction, achieving better cross-domain transferability.", + "bbox": [ + 503, + 534, + 888, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, we are the first diffusion-based method focusing on shadow generation.", + "bbox": [ + 503, + 775, + 888, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Diffusion Models", + "text_level": 1, + "bbox": [ + 503, + 816, + 665, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, diffusion models have emerged as a powerful tool in image generation and image editing. These models approach image generation as a series of stochastic transitions, moving from a basic distribution to the desired", + "bbox": [ + 503, + 840, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8122", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f81995dc9c2606ff9b9ede849f50101ef2deb2f66d4b156033fb36b684386cb5.jpg", + "image_caption": [ + "Figure 2. The pipeline of dataset construction. We use object-shadow detection model [41] to predict pairs of object and shadow masks in the real image $I_r$ . Then we obtain the union $M_s$ of all shadow masks as the inpainting mask and apply inpainting model [32] to get a desshadowed image $I_d$ . After designating a foreground object, we replace the background shadow regions $M_{bs}$ in $I_d$ with the counterparts in $I_r$ to synthesize a composite image $I_c$ , and replace all the shadow regions $M_s$ in $I_d$ with the counterparts in $I_r$ to obtain the ground-truth target image $I_g$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 897, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data distribution [11]. Diffusion models can be divided into unconditional diffusion models [11, 37] and conditional diffusion models [27, 32, 52]. Unconditional diffusion models focus on generating realistic images by capturing the distribution of natural images, without the need of any specific input conditions. Conditional diffusion models are designed to produce images under the guidance of specific conditional inputs, such as text descriptions, semantic masks, and so on. ControlNet [52] is a popular conditional diffusion model, which equips large pretrained text-to-image diffusion models with spatial-aware and task-specific conditions. We build our model upon ControlNet and propose several innovations to meet the specific requirements of shadow generation.", + "bbox": [ + 75, + 393, + 472, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Construction", + "text_level": 1, + "bbox": [ + 76, + 621, + 282, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of our dataset construction is illustrated in Figure 2, which will be detailed next.", + "bbox": [ + 75, + 647, + 468, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Shadow Image Collection", + "text_level": 1, + "bbox": [ + 76, + 689, + 308, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We harvest an extensive collection of real-world outdoor images with natural lighting across various scenes from two sources. On one hand, we crawl online images from public websites that have licenses for reuse. On the other hand, we hire photographers to capture photos in the outdoor scenes that satisfy our requirements. We only preserve the images with at least one object-shadow pair, arriving at 44,044 images.", + "bbox": [ + 75, + 713, + 468, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Shadow Removal", + "text_level": 1, + "bbox": [ + 76, + 845, + 246, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a real image $I_r$ with object-shadow pairs, we use the pretrained object-shadow detection model [41] to predict $K$", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pairs of object and shadow masks. We use $M_{o,k}$ (resp., $M_{s,k}$ ) to denote the object (resp., shadow) mask of the $k$ -th object. We refer to one detected object-shadow pair as one detected instance. We eliminate the images without any detected instance.", + "bbox": [ + 496, + 393, + 890, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, we attempt to erase all the detected shadows. We have tried some state-of-the-art shadow removal models [8, 9], but the performance in the wild is below our expectation due to poor generalization ability. Considering the recent rapid advance of image inpainting [14, 23, 29, 32, 45, 56] techniques, we resort to image inpainting to remove the shadows. Although image inpainting cannot preserve the background information precisely, we observe that the background textures in the shadow region are usually very simple, and the inpainted result has similar textures with the original background. Thus, we roughly treat the inpainted results as desshadowed results.", + "bbox": [ + 496, + 472, + 892, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We obtain the union of all detected shadow masks $M_{s} = M_{s,1} \\cup M_{s,2} \\cup \\dots \\cup M_{s,K}$ as the inpainting mask and apply the pretrained inpainting model [32] to get a deshadowed image $I_{d}$ . In practice, we observe that the inpainting model is prone to generate low-quality shadow in the inpainted region in some cases. To prevent the inpainting model from generating undesirable shadows in the inpainted region, we adopt some tricks like dilating the inpainting mask and flipping images vertically, which can effectively obstruct undesirable shadow generation during inpainting. However, there may still exist undesirable shadows or noticeable artifacts in the inpainted region.", + "bbox": [ + 496, + 656, + 893, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After inpainting, we manually filter the object-shadow pairs according to the following rules: 1) We remove the object-shadow pairs with low-quality object masks or shadow masks. 2) We remove those object-shadow pairs", + "bbox": [ + 496, + 840, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8123", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with generated shadows or noticeable artifacts in the inpainted region. After manual filtering, we refer to the remaining object-shadow pairs as valid instances. We have 21,575 images with 28,573 valid instances.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Composite Image Synthesis", + "text_level": 1, + "bbox": [ + 76, + 164, + 323, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a pair of a real image $I_r$ and a deshadowed image $I_d$ , we randomly select the $k$ -th foreground object from valid instances and synthesize the composite image. $M_{o,k}$ (resp., $M_{s,k}$ ) is referred to as the foreground object (resp., shadow) mask $M_{f_o}$ (resp., $M_{f_s}$ ). One strategy is replacing the shadow region $M_{fs}$ of this foreground object in $I_r$ with the counterpart in $I_d$ to erase the foreground shadow. However, this strategy may leave traces along the shadow boundary, in which case the model may find a shortcut to generate the shadow. Another strategy is replacing the shadow regions $M_{bs} = M_{s,1} \\cup \\dots \\cup M_{s,k-1} \\cup M_{s,k+1} \\cup \\dots \\cup M_{s,K}$ of the other objects in $I_d$ with the counterparts in $I_r$ to synthesize a composite image $I_c$ , in which only the selected foreground object does not have shadow while all the other objects have shadows. We adopt the second strategy.", + "bbox": [ + 75, + 188, + 468, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After inpainting, the background may undergo slight changes, so the background of $I_{c}$ may be slightly different from that of $I_{r}$ . To ensure consistent background, we obtain the ground-truth target image $I_{g}$ by replacing the shadow regions $M_{s}$ of all objects in $I_{d}$ with the counterparts in $I_{r}$ . Then, $I_{c}$ and $I_{g}$ form a pair of input composite image and ground-truth target image. So far, we obtain tuples in the form of $\\{I_{c}, M_{fo}, M_{fs}, M_{bs}, I_{g}\\}$ , which will be used for model training. Example images and more statistics of our dataset can be found in the supplementary.", + "bbox": [ + 75, + 416, + 470, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Background", + "text_level": 1, + "bbox": [ + 76, + 584, + 205, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stable Diffusion [32] is latent diffusion model operating in a latent space. First, $512 \\times 512$ images are converted to $64 \\times 64$ latent images using VAE [18] with encoder $E_r$ and decoder $D_r$ . The image space is projected to the latent space using $E_r$ , and back to the image space using $D_r$ . Then, the forward diffusion process and backward denoising process are performed in the latent space. The denoising U-Net [33] consists of an encoder with 12 blocks, a middle block, and a skip-connected decoder with 12 blocks.", + "bbox": [ + 75, + 609, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, random Gaussian noise $\\epsilon$ is added to the latent image $z_0$ in the denoising step $t$ , producing a noisy latent image $z_t$ . Given time step $t$ and text prompt $c_{txt}$ , the denoising U-Net with model parameters $\\epsilon_\\theta$ is trained to predict the added noise $\\epsilon$ .", + "bbox": [ + 75, + 748, + 468, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To support spatial conditional information (e.g., edge, pose, depth), ControlNet [52] integrates a control encoder $E_{c}$ with pre-trained Stable Diffusion. Specifically, the control encoder contains trainable replicas of its 12 encoding blocks and middle block across four resolutions", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(64 \\times 64, 32 \\times 32, 16 \\times 16, 8 \\times 8)$ . It takes a $512 \\times 512$ conditional image as input.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The conditional feature maps $c_{img}$ output from control encoder are used to enhance the 12 skip-connections and middle block in denoising U-Net via zero convolution layers. While the original Stable Diffusion is fixed to retain prior knowledge, control encoder could incorporate additional conditions to guide image generation. The objective could be rewritten as", + "bbox": [ + 496, + 121, + 890, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c t r l} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {t x t}, \\boldsymbol {c} _ {i m g}\\right) \\| _ {2} ^ {2} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 229, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Method", + "text_level": 1, + "bbox": [ + 500, + 261, + 589, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a composite image $I_{c}$ without foreground shadow as well as the foreground object mask $M_{fo}$ , our Shadow Generation Diffusion (SGDiffusion) model aims to produce $\\tilde{I}_{g}$ with plausible foreground shadow. We will adapt ControlNet [52] to shadow generation task in Section 5.1, and propose novel modules to improve the shadow intensity in Section 5.2. Finally, we will briefly introduce post-processing techniques to enhance the image quality in Section 5.3.", + "bbox": [ + 496, + 286, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Adapting ControlNet to Shadow Generation", + "text_level": 1, + "bbox": [ + 498, + 417, + 874, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For shadow generation task, the useful conditional information is input composite image $I_{c}$ and foreground object mask $M_{fo}$ , in which the foreground object mask indicates the target object we need to generate shadow for. We concatenate $I_{c}$ with $M_{fo}$ as the input of control encoder $E_{c}$ . The control encoder outputs the conditional feature maps $c_{sg}$ , which are injected into the denoising decoder to provide guidance. For the text prompt, we have tried several variants like \"the [object category] with shadow\", but they have no significant impact on the generated shadows. Therefore, we use null text prompt by default.", + "bbox": [ + 496, + 440, + 890, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a set of conditions including time step $t$ and conditional feature maps $c_{sg}$ , the denoising U-Net with model parameters $\\epsilon_{\\theta}$ predicts the noise $\\epsilon$ added to the noisy latent image $z_{t}$ :", + "bbox": [ + 496, + 606, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {s g}\\right)) \\| _ {2} ^ {2} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 672, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enforce the model to place more emphasis on the foreground shadow region, we introduce weighted noise loss, which assigns higher weights to the foreground shadow region. We expand the foreground shadow mask by a dilated kernel to get the expanded mask $\\hat{M}_{fs}$ . The weights in the expanded foreground shadow region are $w$ while the other weights are 1, leading to the weight map $W_{fs}$ . If we do not expand the foreground shadow region, the model will be misled to generate large shadows, overlooking the details of shadow shapes and boundaries. By applying weight map $W_{fs}$ to the noise loss, we can arrive at", + "bbox": [ + 496, + 705, + 890, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\boldsymbol {W} _ {f s} \\circ (\\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {s g})) \\| _ {2} ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 878, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8124", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/976a603a173f618365fdf99dfad9378cc01405353e58f5ff3b4c84cc834122c7.jpg", + "image_caption": [ + "Figure 3. The framework of our SGDiffusion. We adapt ControlNet (Control Encoder and Denoising U-Net) to shadow generation task. We also introduce an intensity encoder to modulate the foreground shadow region in the noise map $\\tilde{\\epsilon}$ , leading to $\\hat{\\epsilon}$ . The output noise $\\hat{\\epsilon}$ is supervised by weighted noise loss $\\mathcal{L}_{mwsg}$ based on the expanded foreground shadow mask $\\tilde{M}_{fs}$" + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 831, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\circ$ denotes element-wise multiplication.", + "bbox": [ + 76, + 470, + 377, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, to retain more information of input composite image $I_{c}$ in the initial noise, we obtain $z_{T}$ by adding noise to the latent image of $I_{c}$ , rather than directly sampling from the Gaussian distribution $\\mathcal{N}(0,1)$ .", + "bbox": [ + 75, + 486, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Shadow Intensity Modulation", + "text_level": 1, + "bbox": [ + 76, + 558, + 341, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By using the adapted ControlNet in Section 5.1, we observe that the intensity of generated foreground shadow is unsatisfactory. Especially when the background has object-shadow pairs, the generated foreground shadow is often notably darker or brighter than background shadows. Such inconsistency between foreground shadow intensity and background shadow intensity makes the whole image unrealistic.", + "bbox": [ + 75, + 582, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, we use encoder $E_{i}$ to extract intensity-relevant information. Intuitively, by observing background shadows and its surrounding unshadowed areas, we can estimate the intensity of foreground shadows. Thus, the input of intensity encoder $E_{i}$ should include the composite image $I_{c}$ and background shadow mask $M_{bs}$ . When there is no background shadow, the mask is all black. We concatenate $I_{c}$ with background shadow mask $M_{bs}$ as the input of intensity encoder.", + "bbox": [ + 75, + 688, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The intensity encoder outputs scales and biases to adjust the intensity of noise map within the foreground shadow region. The modulated noise map results in the modulated latent image, and further results in the modulated foreground", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shadow. Therefore, the intensity adjustment of noise map is finally embodied in the intensity variation of generated foreground shadow. Specifically, when the noise map has $c$ channels, $E_{i}$ outputs the $c$ -dim scale vector $\\boldsymbol{s}$ and $c$ -dim bias vector $\\boldsymbol{b}$ , containing channel-wise scales and biases. $\\boldsymbol{s}$ and $\\boldsymbol{b}$ are used to modulate the predicted noise map within the foreground shadow region.", + "bbox": [ + 496, + 470, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One problem is that the foreground shadow region is unknown in the testing stage, so we need to predict the foreground shadow mask. To avoid much extra computational cost, we take advantage of the feature maps in the denoising U-Net to predict the foreground shadow mask. Previous works usually combine different layers of feature maps in denoising U-Net for mask prediction [24, 43]. We try different layers of feature maps and find that decoder feature maps are more effective in shadow mask prediction. We also use foreground object mask, which could provide useful hints for the location of foreground shadow. We resize all decoder feature maps and foreground object mask to the same size, and concatenate them channel-wisely. The concatenation passes through several convolutional layers to predict the foreground shadow mask $\\tilde{M}_{fs}$ . $\\tilde{M}_{fs}$ is supervised with ground-truth foreground shadow mask $M_{fs}$ by Binary Cross-Entropy (BCE) loss and Dice loss [26]:", + "bbox": [ + 496, + 577, + 892, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m a s k} = \\mathcal {L} _ {b c e} (\\tilde {M} _ {f s}, M _ {f s}) + \\mathcal {L} _ {d i c e} (\\tilde {M} _ {f s}, M _ {f s}). (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 843, + 890, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When $t$ is large, $z_{t}$ is close to random noise and thus the decoder feature maps are not informative to predict shadow", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8125", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mask. Hence, we only employ the loss $\\mathcal{L}_{mask}$ when the time step $t$ is small. We set the threshold of $t$ as $\\sigma T$ , in which $T$ is the total number of steps. Accordingly, shadow intensity modulation is only applied when $t$ is smaller than the threshold $\\sigma T$ .", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Provided with the predicted foreground shadow mask $\\tilde{M}_{fs}$ , we can modulate the noise map within the foreground shadow region. Given the predicted noise map $\\tilde{\\epsilon} = \\epsilon_{\\theta}(z_t, t, c_{sg})$ , we multiply $\\tilde{\\epsilon}$ by channel-wise scales $s$ and add channel-wise biases $\\pmb{b}$ to get $\\tilde{\\epsilon}'$ . Then, based on $\\tilde{M}_{fs}$ , we combine the modulated noise map and original noise map to get the final noise map: $\\hat{\\epsilon} = \\tilde{\\epsilon}' \\circ \\tilde{M}_{fs} + \\tilde{\\epsilon} \\circ (1 - \\tilde{M}_{fs})$ .", + "bbox": [ + 76, + 166, + 467, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We replace the predicted noise map in Eqn. (3) with the final noise map $\\hat{\\epsilon}$ and get", + "bbox": [ + 76, + 273, + 467, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m w s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\boldsymbol {W} _ {f s} \\circ (\\epsilon - \\hat {\\epsilon}) \\| _ {2} ^ {2} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 313, + 467, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We summarize the mask prediction loss in Eqn. (4) and weighted noise loss in Eqn. (5) as", + "bbox": [ + 76, + 347, + 467, + 377 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda \\mathcal {L} _ {\\text {m w s g}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 388, + 467, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is a trade-off parameter.", + "bbox": [ + 76, + 415, + 294, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Post-processing", + "text_level": 1, + "bbox": [ + 76, + 438, + 233, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that the generated images could have color shift and background variation issues. Color shift means that the overall color tone deviates from the input composite image. Background variation means that some background details are changed. To solve these issues, we create a multi-task post-processing network which yields the rectified image together with the foreground shadow mask. Then, we combine input composite image and rectified image based on the predicted foreground shadow mask to produce the final image. The technical details are left to supplementary.", + "bbox": [ + 75, + 462, + 467, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 626, + 207, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 650, + 362, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on both DESOBA [12] and our contributed DESOBAv2 dataset. We split DESOBAv2 into 21,088 training images with 27,718 tuples and 487 test images with 855 tuples. Following [12], the test set contains BOS images (with background object-shadow pairs) and BOS-free images. Most of our experiments are based on DESOBAv2 dataset due to the following two concerns: 1) DESOBAv2 has larger test set which supports more comprehensive evaluation. 2) DESOBA has the artifacts caused by manual shadow removal and the existing methods (e.g., SGRNet) tend to overfit such artifacts.", + "bbox": [ + 75, + 672, + 467, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the generated results, we evaluate both image quality and mask quality. For image evaluation, following [12], we adopt RMSE and SSIM, which are calculated based on the ground-truth target image and the generated image.", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Global RMSE (GR) and Global SSIM (GS) are calculated over the whole image, while Local RMSE (LR) and Local SSIM (LS) are calculated over the ground-truth foreground shadow region. For the mask evaluation, following [12], we adopt Balanced Error Rate (BER), which is calculated based on the ground-truth binary foreground shadow mask and the predicted foreground shadow mask obtained by threshold 0.5. Global BER (GB) is calculated over the whole image, while Local BER (LB) is calculated over the ground-truth foreground shadow region. Note that diffusion model has stochastic property and shadow generation is a multi-modal task, that is, one input has multiple plausible outputs. Similar to multi-modal inpainting evaluation [54, 55], we generate 5 results for one test image with different random seeds and select the one closest to the ground-truth (the highest Local SSIM) to calculate evaluation metrics.", + "bbox": [ + 496, + 90, + 890, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 358, + 715, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We develop our method with PyTorch 1.12.1 [30]. Our model is trained using the Adam optimizer [17] with a constant learning rate of $1e^{-5}$ over 50 epochs, on four NVIDIA RTX A6000 GPUs. Our method is built upon ControlNet [52]. We employ ResNet18 [10] as the intensity encoder. The mask predictor passes the concatenation of decoder feature maps and foreground object mask through four convolutional layers, with ReLU activation following the first three layers and Sigmoid activation following the last layer. We set the hyper-parameters $w$ , $\\sigma$ , and $\\lambda$ as 10, 0.7, and 1, respectively.", + "bbox": [ + 496, + 387, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 498, + 579, + 743, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [12], we compare with ShadowGAN [53], Mask-ShadowGAN [13], ARShadowGAN [22], and SGRNet [12]. We train and test all methods on DESOBAv2 dataset. The quantitative results are summarized in Table 1. We observe that our SGDiffusion achieves the lowest GRMSE, LRMSE and the highest GSSIM, LSSIM, which demonstrates that our method could generate shadow images that are closer to the ground-truth shadow images. The best GB and LB results demonstrate that the shapes and locations of our generated shadows are more accurate.", + "bbox": [ + 496, + 607, + 890, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For qualitative comparison, we show several example results in Figure 4. Compared with the baseline methods, the shadows produced by our model have more reasonable shapes and intensities. Moreover, as shown in row 1, our method can take into account the self-occlusion of objects to generate discontinuous shadows. As shown in row 4, our method can also consider the material of the objects, producing shadows with translucency effects. We provide more examples in the supplementary.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8126", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e17d3a2538d52dc5db075b9d47d110dcbe3cfad68d4be178861b9ac64deddb78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 102, + 194, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5ad355ce11dd36cc7fb3a30f930573ad3b90c86eeeb4ae1d7fabfd4167b9bea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 102, + 290, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a03e7b338c2b6708488c996507271c8dec5af53968fac8ddd7f43251133770dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 102, + 385, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71b88ff84bb7f4d84a21af979a9f912b07f629e3943fe97c04aba36550886f20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 102, + 482, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2405d23ecb5e5a9c7c3af460adb50f176a321633030f34976a6897dfd5b737a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 102, + 580, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c38209021a2dce44d0f3ee3df1323102680dbf5bf5e7f104b391cd81c3eba8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 102, + 676, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faf3c83ffeeae0b959067e0a7226af1d1f044d0bb5ece80c7cd7cf97f15114ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 102, + 774, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27e99ab1b67aaf69575769ea323efa6646874b56a93bb6f6d26835030c554abc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 102, + 872, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b1636b5f955b36109bb158be39f9a1ffc65d1c82a4f7bce6a74d5d15f74132d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 178, + 192, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c636852e0334ef3b043660ab486ed75e753c9272f7e24ce8d410bb103816197.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 178, + 290, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb021e3f20ca4010d8e869de53546cc40161ddac343853185e6faf779aa2d6f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 178, + 385, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e5792c07c83fb3d6f2d46ebb869967d7d3ce6b18de76e68324523a50da456e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 178, + 483, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c671801a469a75b781f4293a0f45d6ca0453f65ee0d5c43c9998e5392f49a78f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 178, + 580, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59c13068a9a375b067bc7f093017cc641880acb88d08b6a07a964145553cf71d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 178, + 676, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84e69e7792d30621b7939e09ff5dd20dafb38f8e95acd984d71b26f6791e34b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 178, + 774, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4bc469259e2b7d520bceab4e9275a6925136ab0b7df4c1ad862907753123c1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 178, + 872, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d037949ea397c53342ee0e20ceecdf202b984cf9e6b3d00ee3ecb4c7968c66de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 253, + 192, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/068abca7ecd3657ba5f48bc24051bf516283bbb722d0affec1255aaa27801bc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 253, + 290, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd006c10810b049c00cfa8944b5366f1139dd728ef02f576fee6e186047e5323.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 253, + 385, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db4b852d6a9b4a7ebca692b96ef98b208edfdce63ae521a0915dcce8a703439b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 253, + 483, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee7fe548c9efd7438a343f373abb42be59134523a24368486f3b233007eb0876.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 253, + 580, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84de971930cabf625e5a41ef0540156f3242f5e745b3194cb57e6f958180900c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 253, + 676, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18b948ba09a6134f76f2054bde2f0677b154024f87bdbe72d14d4cb906bd081f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 253, + 774, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a15da3c17616dce5a306c69a508a168bd766ec5a5caacd3a6d72abbb960fb71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 253, + 872, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4be851cd4d740da59790ce93174b173f10546874b87bed2f7468e8c268ee7507.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 98, + 329, + 192, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da73d8df4e2f3de3251f85108a2828f7a2f8628a3fa9fd755e64553088a72d1a.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 194, + 329, + 290, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6623fdb43e19ee7c114059eb9863788d07b984d544b084fb073d60e84cd5b49e.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 290, + 329, + 385, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f406e079ac0961b77c2bd8c55fee98e7595757c6c16672830210232023c21121.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 387, + 329, + 483, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4dc0a8fb9212b6b99741dfc57b44987f9458198e8672a6f9551a9702e532d8f8.jpg", + "image_caption": [ + "(e)", + "Figure 4. Visual comparison of different methods on DESOBAv2 dataset. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), our SGDiffusion (g), ground-truth (h)." + ], + "image_footnote": [], + "bbox": [ + 483, + 329, + 580, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61ed1e035026ea38d9d2b601550403d1aa4151b044378f1495e7509096b06adc.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 581, + 329, + 676, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09d862ad07d39b42adfc6452f15f3f36be05c66ee2dc7226085bf954e41a71aa.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 679, + 329, + 774, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45ba09f138e0343c871944764b42420e4866ddf088c07a14d5320df44c2490d5.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 776, + 329, + 872, + 402 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/413f6a902a31e6210de757d9beee744f8253b92253c481487e6b6119370e57ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBOS Test ImagesBOS-free Test Images
GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓
ShadowGAN [53]7.51167.4640.9610.1970.4460.89017.32576.5080.9010.0600.4250.842
MaskshadowGAN [13]8.99779.4180.9510.1800.5001.00019.33894.3270.9060.0440.5001.000
ARShadowGAN [22]7.33558.0370.9610.2410.3830.76116.06763.7130.9080.1040.3490.682
SGRNet [12]7.18468.2550.9640.2060.3010.59615.59660.3500.9090.1000.2710.534
SGDiffusion6.09853.6110.9710.3700.2450.48715.11055.8740.9130.1170.2330.452
", + "bbox": [ + 78, + 511, + 911, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. The results of different methods on DESOBAv2 dataset. The best results are highlighted in boldface.", + "bbox": [ + 158, + 656, + 807, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 696, + 240, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study the impact of weighted noise loss (WL), intensity modulation (IM), and post-processing (PP) of our SGDiffusion on BOS test images from DESOBAv2. The quantitative results are summarized in Table 2.", + "bbox": [ + 75, + 727, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In row 1, we report the results of basic ControlNet without weighted noise loss. For WL, the comparison between row 3 and row 1 emphasizes the importance of paying more attention to the foreground shadow region. We also report a special case $\\dagger$ in row 2, where the foreground shadow mask is not expanded when constructing the weight map. The results in row 2 are comparable or even worse than those in", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "row 1, as the model tends to generate larger shadow size while ignoring shape and edge details. For IM, the comparison between row 1 and row 5 shows that the intensity modulation can significantly improve the shadow quality by adjusting the shadow intensity. We also report a special case in row 4, where the intensity encoder input does not contain background shadow mask. The comparison between row 4 and row 5 shows that background shadow mask is helpful, because the background shadow regions and their surrounding regions could provide useful clues to infer shadow intensity. For PP, the comparison between row 6 and row 7 demonstrates that post-processing effectively corrects color shift and background variations, substantially reducing the", + "bbox": [ + 496, + 696, + 893, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8127", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a77dbfb90c264b8c2247d452a947b5b11b357f2f9605ffa3eee469f3375b944.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 104, + 207, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a550b99cae36e0890550c8265d965f72d9faef0ca07747ea98f20e914052fe90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 104, + 316, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee63cb7b91a9ced50431d72d1d7f09168bf541328b8ddbabfd8d56ef03fca9d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 104, + 426, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ceca2d6bfaa2990472ac7e489e40718d3b1586153d1972d86e692e61537a11d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 104, + 537, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c138983ce75567ea481cc3f28c4b6db9a1fa3984e2a62a81a5257bf4fa7bab3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 104, + 647, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/720f8cf315f5c7c21e8af4b16451ff3a3fa03eb2672a9de84ac2fea14795a662.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 104, + 758, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9035dcbe44ad55d64ad412cfa7b53b83c1fe56940136a0d77c38e3244942da3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 104, + 869, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2b2fbc67b24c9ed9f97d1a8b35b288a573dae4ae29947b52671d4957560cf4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 190, + 207, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6269f88ea18a3ef521f98bc0ff94fd8d72e9e18fe9025c57737109379172661.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 190, + 316, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21a03b159c9bbb8fff0123bfd57097a82dad8c49bb2f2dc5f2034addefac9cc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 190, + 426, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94e2f4a949c725541f47dadea13d618553d0ace03fb20b9c321eb98d6b04c8de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 190, + 537, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/862abce213ed04ccc67149d11dc90fddc48331f81641df9d61e994f59ff46f8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 190, + 647, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94a3472d4d3f7b6b1b047d1a277a33d6225ff2736c2532c23cd980aefd1e5830.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 190, + 758, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f984aab0d189cec403acfa49a2dcb0987b5f5ea1957a11c0b2034f5f45833a24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 190, + 869, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed1fef56431338ad1362df518088686fff4bd6a242e9a6819a8056e85ad3ae7e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 101, + 276, + 209, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5554a5c482fdebafb1dbeaf7ecf6044a6e25687c6a5fb7e6323e0423e41d7ba5.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 210, + 276, + 316, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/060331219053558fc6bcecd058e7c6302128bd171e128583d40122fef4e607b8.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 320, + 276, + 426, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/057d50aefabba5f5c5108789ebe54ca984c8944ccc410196b2eb3ce67b58330d.jpg", + "image_caption": [ + "(d)", + "Figure 5. Visual comparison of different methods on real composite images. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), SGDiffusion (g)." + ], + "image_footnote": [], + "bbox": [ + 431, + 276, + 537, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cdd0959c16b8d6dca11b9a5849da66a80e27a483893372ba668c4da84179310.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 540, + 276, + 648, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51c7b4577166f181f30d2a083fbab261eb666f34aa7d1ed12b233ee934c1200e.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 651, + 276, + 759, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5873fc882610db1e3f148726dc21fe9d45d801e3a6a2776ab94f2a99ff934df8.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 761, + 276, + 869, + 359 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/525575a281ef4a448bc87836a6f5c2bfec609dd74cff9b58f035636f259da6d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RowWLIMPPGR↓LR↓GB↓LB↓
1--+8.28559.7530.2710.534
2†-+8.31959.4910.2820.563
3+-+7.04153.8290.2490.492
4-○+7.41056.1210.2690.536
5-++7.35754.1590.2620.526
6++-13.44755.2310.2450.487
7+++6.09853.6110.2450.487
", + "bbox": [ + 78, + 469, + 468, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation studies of our method on BOS test images from DESOBAv2 dataset. WL is short for weighted loss and $\\dagger$ means without expanding shadow mask. IM is short for intensity modulation and $\\circ$ means without using background shadow mask. PP is short for post-processing.", + "bbox": [ + 75, + 625, + 468, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "global RMSE. We also provide the visual results of ablated versions in the supplementary.", + "bbox": [ + 75, + 724, + 468, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Real Composite Images", + "text_level": 1, + "bbox": [ + 76, + 768, + 294, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare different methods on real composite images provided by [12], where background images and foreground objects are from the DESOBA [12] test set. We train all methods on DESOBAv2 and finetune them on DESOBA. The visual results of different methods are showcased in Figure 5. These results confirm that SGDiffusion adeptly synthesizes lifelike shadows with precise contours, loca", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions, and directions, which are compatible with the background object-shadow pairs and foreground object information. In contrast, previous methods often produce vague and misdirected shadows. We provide more examples in the supplementary.", + "bbox": [ + 496, + 470, + 890, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the absence of ground-truth images for real composite images, following [12], we opt for subjective evaluation, engaging 50 human raters in the user study. Each participant is presented with image pairs from the results generated by 5 methods, and asked to choose the image with more realistic foreground shadow. Using the Bradley-Terry model [2], we report the B-T scores in the supplementary, which again proves the advantage of our method.", + "bbox": [ + 496, + 547, + 892, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 617, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have contributed a large-scale shadow generation dataset DESOBAv2. We have also designed a novel diffusion-based shadow generation method. Extensive experimental results show that our method is able to generate plausible shadows for composite foregrounds, significantly surpassing previous methods.", + "bbox": [ + 496, + 704, + 890, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 805, + 679, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major/Key Project, China (Grant No. 2021SHZDZX0102).", + "bbox": [ + 496, + 829, + 890, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8128", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ibrahim Arief, Simon McCallum, and Jon Yngve Hardeberg. Realtime estimation of illumination direction for augmented reality on mobile devices. In CIC, 2012. 2", + "[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 8", + "[3] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020. 1, 2", + "[4] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In ICME, 2021.", + "[5] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In CVPR, 2022.", + "[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. TIP, 2020. 2", + "[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagné, and Jean-Francois Lalonde. Deep parametric indoor lighting estimation. In ICCV, 2019. 2", + "[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 3", + "[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, 2023. 3", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurlPS, 2020. 3", + "[12] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. AAAI, 2022. 1, 2, 6, 7, 8", + "[13] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, 2019. 6, 7, 8", + "[14] Shaozong Huang and Lan Hong. Diffusion model for mural image inpainting. In ITOEC, 2023. 3", + "[15] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM TOG, 2014. 2", + "[16] Eric Kee, James F. O'Brien, and Hany Samir Farid. Exposing photo manipulation from shading and shadows. ACM TOG, 2014. 2", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6", + "[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Bin Liao, Yao Zhu, Chao Liang, Fei Luo, and Chunxia Xiao. Illumination animating and editing in a single picture using scene structure estimation. Computers & Graphics, 82:53-64, 2019. 2", + "[20] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In CVPR, 2018. 2", + "[21] Bin Liu, Kun Xu, and Ralph R Martin. Static scene illumination estimation from videos with applications. JCST, 32 (3):430-442, 2017. 2", + "[22] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In CVPR, 2020. 1, 2, 6, 7, 8", + "[23] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 3", + "[24] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. CoRR, abs/2303.17870, 2023. 5", + "[25] Quanling Meng, Shengping Zhang, Zonglin Li, Chenyang Wang, Weigang Zhang, and Qingming Huang. Automatic shadow generation via exposure fusion. IEEE Transactions on Multimedia, 2023. 2", + "[26] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, 2016. 5", + "[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3", + "[28] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. CoRR, abs/2106.14490, 2021. 1", + "[29] Sibam Parida, Vignesh Srinivas, Bhavishya Jain, Rajesh Naik, and Neeraj Rao. Survey on diverse image inpainting using diffusion models. In PCEMS, 2023. 3", + "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NIPS, 32, 2019. 6", + "[31] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In SIGGRAPH. 2003. 2", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4", + "[34] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In CVPR, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8129", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In ECCV, 2022. 1, 2", + "[36] Yichen Sheng, Jianming Zhang, Julien Philip, Yannick Hold-Geoffroy, Xin Sun, He Zhang, Lu Ling, and Bedrich Benes. Pixht-lab: Pixel height based light effect generation for image compositing. In CVPR, 2023. 1, 2", + "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. CoRR, abs/2010.02502, 2020. 3", + "[38] Yi-Zhe Song, Zhifei Zhang, Zhe L. Lin, Scott D. Cohen, Brian L. Price, Jianming Zhang, Soo Ye Kim, and Daniel G. Aliaga. Objectstitch: Generative object compositing. In CVPR, 2023. 2", + "[39] Xinhao Tao, Junyan Cao, Yan Hong, and Li Niu. Shadow generation with decomposed mask prediction and attentive shadow filling. In AAAI, 2024. 2", + "[40] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In CVPR, 2017. 2", + "[41] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. TPAMI, 2022. 2, 3", + "[42] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In ACM MM, 2019. 2", + "[43] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 5", + "[44] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 1, 2", + "[45] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, 2023. 3", + "[46] Fangneng Zhan, Jiaxing Huang, and Shijian Lu. Adaptive composition gan towards realistic image synthesis. CoRR, abs/1905.04693, 2019. 2", + "[47] Fangneng Zhan, Shijian Lu, Changgong Zhang, Feiying Ma, and Xuansong Xie. Towards realistic 3d embedding via view alignment. CoRR, abs/2007.07066, 2020. 2", + "[48] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 1, 2", + "[49] He Zhang, Jianming Zhang, Federico Perazzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In WACV, 2021. 2", + "[50] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In CVPR, 2019. 2", + "[51] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In WACV, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 6", + "[53] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5:105-115, 2019. 1, 2, 6, 7, 8", + "[54] Lei Zhao, Qihang Mo, Sihuan Lin, Zhizhong Wang, Zhiwen Zuo, Haibo Chen, Wei Xing, and Dongming Lu. Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In CVPR, 2020. 6", + "[55] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In CVPR, 2019. 6", + "[56] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In ECCV, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8130", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_model.json b/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf651489f634f1e45a0a479d76cf652c97f69c5c --- /dev/null +++ b/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_model.json @@ -0,0 +1,2761 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Shadow Generation for Composite Image Using Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.179, + 0.807, + 0.217 + ], + "angle": 0, + "content": "Qingyang Liu1, Junqi You1, Jianting Wang1, Xinhao Tao1, Bo Zhang1, Li Niu1,2* \n1 Shanghai Jiao Tong University 2 miguo.ai" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.219, + 0.819, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)narumimaria, yjqsjtu2022, gloryl1299, taoxinhao, bo-zhang, ustcnewly}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.573 + ], + "angle": 0, + "content": "In the realm of image composition, generating realistic shadow for the inserted foreground remains a formidable challenge. Previous works have developed image-to-image translation models which are trained on paired training data. However, they are struggling to generate shadows with accurate shapes and intensities, hindered by data scarcity and inherent task complexity. In this paper, we resort to foundation model with rich prior knowledge of natural shadow images. Specifically, we first adapt ControlNet to our task and then propose intensity modulation modules to improve the shadow intensity. Moreover, we extend the small-scale DESOBA dataset to DESOBAv2 using a novel data acquisition pipeline. Experimental results on both DESOBA and DESOBAv2 datasets as well as real composite images demonstrate the superior capability of our model for shadow generation task. The dataset, code, and model are released at https://github.com/bcmi/Object-Shadow-Generation-Dataset-DESOBAv2." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.605, + 0.21, + 0.621 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Image composition [28] aims to merge the foreground of one image with another background image to produce a composite image, which has a wide range of applications like virtual reality, artistic creation, and E-commerce. Simply pasting the foreground onto the background often results in visual inconsistencies, including the incompatible illumination between foreground and background [3], lack of foreground shadow/reflection [12, 34], and so on. In this paper, we focus on the shadow issue, i.e., the inserted foreground does not have plausible shadow on the background, which could significantly degrade the realism and quality of composite image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.47, + 0.875 + ], + "angle": 0, + "content": "As illustrated in Figure 1, shadow generation is a challenging task because the foreground shadow is determined by many complicated factors like the lighting information and the geometry of foreground/background. The exist-" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.269, + 0.88, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.518 + ], + "angle": 0, + "content": "Figure 1. A composite image can be obtained by pasting the foreground on the background. Shadow generation aims to generate plausible shadow for the inserted foreground in the composite image to produce a more realistic image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.746 + ], + "angle": 0, + "content": "ing shadow generation methods can be divided into rendering based methods [34-36] and non-rendering based methods [12, 22, 53]. Rendering based methods usually impose strict assumptions on the geometry and lighting, which could hardly be satisfied in real-world scenarios. Besides, [35, 36] require users to specify the lighting information, which hinders its direct application in our task. Non-rendering based methods usually train an image-to-image translation network, based on pairs of composite images without foreground shadows and real images with foreground shadows. However, due to the training data scarcity and task difficulty, these methods are struggling to generate shadows with reasonable shapes and intensities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Recently, foundation model (e.g., stable diffusion [32]) pretrained on large-scale dataset has demonstrated unprecedented potential for image generation and editing. In previous works [44, 48] on object-guided inpainting or composition, they show that the generated foregrounds are accompanied by shadows even without considering the shadow issue, probably because of the rich prior knowledge of natural shadow images in foundation model. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8121" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.332 + ], + "angle": 0, + "content": "We build our method upon conditional foundation model [52] and propose several key innovations. First, we modify the control encoder input and the noise loss to fit our task. Then, we observe that the generated shadow intensity (the level of darkness) is unsatisfactory. Especially when the background objects has shadows, the intensity inconsistency between foreground shadow and background shadows make the whole image unrealistic. Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, the denoising U-Net is modified to output both noise map and foreground shadow mask. The intensity encoder takes in the composite image and background shadow mask, producing the scale/bias to modulate the predicted noise within the foreground shadow region. Finally, we devise a post-processing network to rectify the color shift and background variation." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.334, + 0.468, + 0.606 + ], + "angle": 0, + "content": "The model training requires abundant pairs of composite images without foreground shadows and real images with foreground shadows. The existing real-world shadow generation dataset DESOBA [12] is limited by scale (i.e., 1,012 real images and 3,623 pairs) due to the high cost of manual shadow removal, which is insufficient to train our model. To ensure sufficient supervision, we design a novel data construction pipeline, which extends DESOBA to DESOBAv2 (i.e., 21,575 real images and 28,573 pairs) using object-shadow detection and inpainting techniques. Specifically, we first collect a large number of real-world images with one or more object-shadow pairs. Then, we use pretrained object-shadow detection model [41] to predict object and shadow masks for object-shadow pairs. Next, we apply pretrained inpainting model [32] to inpaint the detected shadow regions to get desshadowed images. Finally, based on real images and desshadowed images, we construct pairs of synthetic composite images and ground-truth target images." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.608, + 0.468, + 0.787 + ], + "angle": 0, + "content": "We conduct experiments on both DESOBAv2 and DESOBA datasets. The results reveal remarkable improvement in shadow generation task, after leveraging the benefits of large-scale data and foundation model. Our main contributions can be summarized as follows: 1) We contribute DESOBAv2, a large-scale real-world shadow generation dataset, which could greatly facilitate the shadow generation task. 2) We propose a cutting-edge diffusion model specifically designed to produce shadows for the composite foregrounds. 3) Through comprehensive experiments, we validate the efficacy of our dataset construction pipeline and the superiority of our proposed model." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.806, + 0.217, + 0.82 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.832, + 0.262, + 0.847 + ], + "angle": 0, + "content": "2.1. Image Composition" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.857, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Image composition aims to overlay a foreground object on a background image to yield a composite result [20, 22, 42, 46, 47]. Previous research works have tackled differ" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.287 + ], + "angle": 0, + "content": "ent issues that can compromise the quality of composite images. For instance, image blending methods [31, 42, 49, 51] target at combining the foreground and background seamlessly. Image harmonization methods [3-6, 40] aim to rectify the illumination disparity between foreground and background. Nonetheless, the above methods largely overlook the shadow cast by the foreground onto the background. Recently, generative image composition methods [38, 44, 48] can insert a foreground object into a bounding box in the background and the inserted object is likely to have shadow effect. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.299, + 0.687, + 0.313 + ], + "angle": 0, + "content": "2.2. Shadow Generation" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.323, + 0.89, + 0.533 + ], + "angle": 0, + "content": "In this paper, the goal of shadow generation task is generating plausible shadow for the composite foreground. Existing methods can be broadly categorized into rendering based methods and non-rendering based methods. The rendering based methods necessitate a comprehensive understanding of factors like illumination, reflectance, material properties, and scene geometry to produce shadows for the inserted objects. However, such detailed knowledge relies on user input [15, 16, 21, 35, 36] or model prediction [1, 7, 19, 50], which is either labor-intensive or unreliable [53]. For example, [35, 36] could produce compelling results with user control. However, in the composite image, the lighting information should be inferred automatically from background instead of requested by users." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.535, + 0.89, + 0.774 + ], + "angle": 0, + "content": "Non-rendering based methods [12, 22, 25, 53] aim to translate an input composite image without foreground shadow to an output with foreground shadow, bypassing the need for explicit knowledge of the aforementioned factors. For instance, ShadowGAN [53] utilizes both global and local conditional discriminator to enhance the realism of generated shadows. ARShadowGAN [22] emphasizes the importance of background shadow and uses it to guide foreground shadow generation. SGRNet [12] encourages the information exchange between foreground and background, and employs a classic illumination model for better shadow effect. The work [25] produces multiple underexposure images and fuses them to get the final shadow region. DMASNet [39] decomposes shadow mask prediction into box prediction and shape prediction, achieving better cross-domain transferability." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.776, + 0.89, + 0.806 + ], + "angle": 0, + "content": "To the best of our knowledge, we are the first diffusion-based method focusing on shadow generation." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.817, + 0.666, + 0.832 + ], + "angle": 0, + "content": "2.3. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.841, + 0.89, + 0.899 + ], + "angle": 0, + "content": "In recent years, diffusion models have emerged as a powerful tool in image generation and image editing. These models approach image generation as a series of stochastic transitions, moving from a basic distribution to the desired" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8122" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.898, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Figure 2. The pipeline of dataset construction. We use object-shadow detection model [41] to predict pairs of object and shadow masks in the real image \\( I_r \\). Then we obtain the union \\( M_s \\) of all shadow masks as the inpainting mask and apply inpainting model [32] to get a desshadowed image \\( I_d \\). After designating a foreground object, we replace the background shadow regions \\( M_{bs} \\) in \\( I_d \\) with the counterparts in \\( I_r \\) to synthesize a composite image \\( I_c \\), and replace all the shadow regions \\( M_s \\) in \\( I_d \\) with the counterparts in \\( I_r \\) to obtain the ground-truth target image \\( I_g \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.606 + ], + "angle": 0, + "content": "data distribution [11]. Diffusion models can be divided into unconditional diffusion models [11, 37] and conditional diffusion models [27, 32, 52]. Unconditional diffusion models focus on generating realistic images by capturing the distribution of natural images, without the need of any specific input conditions. Conditional diffusion models are designed to produce images under the guidance of specific conditional inputs, such as text descriptions, semantic masks, and so on. ControlNet [52] is a popular conditional diffusion model, which equips large pretrained text-to-image diffusion models with spatial-aware and task-specific conditions. We build our model upon ControlNet and propose several innovations to meet the specific requirements of shadow generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.622, + 0.283, + 0.637 + ], + "angle": 0, + "content": "3. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.47, + 0.678 + ], + "angle": 0, + "content": "The pipeline of our dataset construction is illustrated in Figure 2, which will be detailed next." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.69, + 0.31, + 0.706 + ], + "angle": 0, + "content": "3.1. Shadow Image Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.835 + ], + "angle": 0, + "content": "We harvest an extensive collection of real-world outdoor images with natural lighting across various scenes from two sources. On one hand, we crawl online images from public websites that have licenses for reuse. On the other hand, we hire photographers to capture photos in the outdoor scenes that satisfy our requirements. We only preserve the images with at least one object-shadow pair, arriving at 44,044 images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.248, + 0.862 + ], + "angle": 0, + "content": "3.2. Shadow Removal" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Given a real image \\(I_r\\) with object-shadow pairs, we use the pretrained object-shadow detection model [41] to predict \\(K\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.469 + ], + "angle": 0, + "content": "pairs of object and shadow masks. We use \\( M_{o,k} \\) (resp., \\( M_{s,k} \\)) to denote the object (resp., shadow) mask of the \\( k \\)-th object. We refer to one detected object-shadow pair as one detected instance. We eliminate the images without any detected instance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.653 + ], + "angle": 0, + "content": "Subsequently, we attempt to erase all the detected shadows. We have tried some state-of-the-art shadow removal models [8, 9], but the performance in the wild is below our expectation due to poor generalization ability. Considering the recent rapid advance of image inpainting [14, 23, 29, 32, 45, 56] techniques, we resort to image inpainting to remove the shadows. Although image inpainting cannot preserve the background information precisely, we observe that the background textures in the shadow region are usually very simple, and the inpainted result has similar textures with the original background. Thus, we roughly treat the inpainted results as desshadowed results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.894, + 0.838 + ], + "angle": 0, + "content": "We obtain the union of all detected shadow masks \\( M_{s} = M_{s,1} \\cup M_{s,2} \\cup \\dots \\cup M_{s,K} \\) as the inpainting mask and apply the pretrained inpainting model [32] to get a deshadowed image \\( I_{d} \\). In practice, we observe that the inpainting model is prone to generate low-quality shadow in the inpainted region in some cases. To prevent the inpainting model from generating undesirable shadows in the inpainted region, we adopt some tricks like dilating the inpainting mask and flipping images vertically, which can effectively obstruct undesirable shadow generation during inpainting. However, there may still exist undesirable shadows or noticeable artifacts in the inpainted region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.902 + ], + "angle": 0, + "content": "After inpainting, we manually filter the object-shadow pairs according to the following rules: 1) We remove the object-shadow pairs with low-quality object masks or shadow masks. 2) We remove those object-shadow pairs" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8123" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "with generated shadows or noticeable artifacts in the inpainted region. After manual filtering, we refer to the remaining object-shadow pairs as valid instances. We have 21,575 images with 28,573 valid instances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.165, + 0.325, + 0.182 + ], + "angle": 0, + "content": "3.3. Composite Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.47, + 0.416 + ], + "angle": 0, + "content": "Given a pair of a real image \\( I_r \\) and a deshadowed image \\( I_d \\), we randomly select the \\( k \\)-th foreground object from valid instances and synthesize the composite image. \\( M_{o,k} \\) (resp., \\( M_{s,k} \\)) is referred to as the foreground object (resp., shadow) mask \\( M_{f_o} \\) (resp., \\( M_{f_s} \\)). One strategy is replacing the shadow region \\( M_{fs} \\) of this foreground object in \\( I_r \\) with the counterpart in \\( I_d \\) to erase the foreground shadow. However, this strategy may leave traces along the shadow boundary, in which case the model may find a shortcut to generate the shadow. Another strategy is replacing the shadow regions \\( M_{bs} = M_{s,1} \\cup \\dots \\cup M_{s,k-1} \\cup M_{s,k+1} \\cup \\dots \\cup M_{s,K} \\) of the other objects in \\( I_d \\) with the counterparts in \\( I_r \\) to synthesize a composite image \\( I_c \\), in which only the selected foreground object does not have shadow while all the other objects have shadows. We adopt the second strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.471, + 0.569 + ], + "angle": 0, + "content": "After inpainting, the background may undergo slight changes, so the background of \\( I_{c} \\) may be slightly different from that of \\( I_{r} \\). To ensure consistent background, we obtain the ground-truth target image \\( I_{g} \\) by replacing the shadow regions \\( M_{s} \\) of all objects in \\( I_{d} \\) with the counterparts in \\( I_{r} \\). Then, \\( I_{c} \\) and \\( I_{g} \\) form a pair of input composite image and ground-truth target image. So far, we obtain tuples in the form of \\( \\{I_{c}, M_{fo}, M_{fs}, M_{bs}, I_{g}\\} \\), which will be used for model training. Example images and more statistics of our dataset can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.585, + 0.206, + 0.603 + ], + "angle": 0, + "content": "4. Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.469, + 0.748 + ], + "angle": 0, + "content": "Stable Diffusion [32] is latent diffusion model operating in a latent space. First, \\(512 \\times 512\\) images are converted to \\(64 \\times 64\\) latent images using VAE [18] with encoder \\(E_r\\) and decoder \\(D_r\\). The image space is projected to the latent space using \\(E_r\\), and back to the image space using \\(D_r\\). Then, the forward diffusion process and backward denoising process are performed in the latent space. The denoising U-Net [33] consists of an encoder with 12 blocks, a middle block, and a skip-connected decoder with 12 blocks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.824 + ], + "angle": 0, + "content": "During training, random Gaussian noise \\(\\epsilon\\) is added to the latent image \\(z_0\\) in the denoising step \\(t\\), producing a noisy latent image \\(z_t\\). Given time step \\(t\\) and text prompt \\(c_{txt}\\), the denoising U-Net with model parameters \\(\\epsilon_\\theta\\) is trained to predict the added noise \\(\\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To support spatial conditional information (e.g., edge, pose, depth), ControlNet [52] integrates a control encoder \\( E_{c} \\) with pre-trained Stable Diffusion. Specifically, the control encoder contains trainable replicas of its 12 encoding blocks and middle block across four resolutions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "\\((64 \\times 64, 32 \\times 32, 16 \\times 16, 8 \\times 8)\\). It takes a \\(512 \\times 512\\) conditional image as input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.226 + ], + "angle": 0, + "content": "The conditional feature maps \\( c_{img} \\) output from control encoder are used to enhance the 12 skip-connections and middle block in denoising U-Net via zero convolution layers. While the original Stable Diffusion is fixed to retain prior knowledge, control encoder could incorporate additional conditions to guide image generation. The objective could be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.231, + 0.892, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c t r l} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {t x t}, \\boldsymbol {c} _ {i m g}\\right) \\| _ {2} ^ {2} \\right]. \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.262, + 0.591, + 0.277 + ], + "angle": 0, + "content": "5. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Given a composite image \\( I_{c} \\) without foreground shadow as well as the foreground object mask \\( M_{fo} \\), our Shadow Generation Diffusion (SGDiffusion) model aims to produce \\( \\tilde{I}_{g} \\) with plausible foreground shadow. We will adapt ControlNet [52] to shadow generation task in Section 5.1, and propose novel modules to improve the shadow intensity in Section 5.2. Finally, we will briefly introduce post-processing techniques to enhance the image quality in Section 5.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.418, + 0.875, + 0.433 + ], + "angle": 0, + "content": "5.1. Adapting ControlNet to Shadow Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.607 + ], + "angle": 0, + "content": "For shadow generation task, the useful conditional information is input composite image \\( I_{c} \\) and foreground object mask \\( M_{fo} \\), in which the foreground object mask indicates the target object we need to generate shadow for. We concatenate \\( I_{c} \\) with \\( M_{fo} \\) as the input of control encoder \\( E_{c} \\). The control encoder outputs the conditional feature maps \\( c_{sg} \\), which are injected into the denoising decoder to provide guidance. For the text prompt, we have tried several variants like \"the [object category] with shadow\", but they have no significant impact on the generated shadows. Therefore, we use null text prompt by default." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.667 + ], + "angle": 0, + "content": "Given a set of conditions including time step \\( t \\) and conditional feature maps \\( c_{sg} \\), the denoising U-Net with model parameters \\( \\epsilon_{\\theta} \\) predicts the noise \\( \\epsilon \\) added to the noisy latent image \\( z_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.673, + 0.891, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {s g}\\right)) \\| _ {2} ^ {2} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.872 + ], + "angle": 0, + "content": "To enforce the model to place more emphasis on the foreground shadow region, we introduce weighted noise loss, which assigns higher weights to the foreground shadow region. We expand the foreground shadow mask by a dilated kernel to get the expanded mask \\(\\hat{M}_{fs}\\). The weights in the expanded foreground shadow region are \\(w\\) while the other weights are 1, leading to the weight map \\(W_{fs}\\). If we do not expand the foreground shadow region, the model will be misled to generate large shadows, overlooking the details of shadow shapes and boundaries. By applying weight map \\(W_{fs}\\) to the noise loss, we can arrive at" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.879, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\boldsymbol {W} _ {f s} \\circ (\\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {s g})) \\| _ {2} ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8124" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.089, + 0.833, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.403, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Figure 3. The framework of our SGDiffusion. We adapt ControlNet (Control Encoder and Denoising U-Net) to shadow generation task. We also introduce an intensity encoder to modulate the foreground shadow region in the noise map \\(\\tilde{\\epsilon}\\), leading to \\(\\hat{\\epsilon}\\). The output noise \\(\\hat{\\epsilon}\\) is supervised by weighted noise loss \\(\\mathcal{L}_{mwsg}\\) based on the expanded foreground shadow mask \\(\\tilde{M}_{fs}\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.472, + 0.379, + 0.486 + ], + "angle": 0, + "content": "where \\(\\circ\\) denotes element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.549 + ], + "angle": 0, + "content": "During inference, to retain more information of input composite image \\( I_{c} \\) in the initial noise, we obtain \\( z_{T} \\) by adding noise to the latent image of \\( I_{c} \\), rather than directly sampling from the Gaussian distribution \\( \\mathcal{N}(0,1) \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.342, + 0.575 + ], + "angle": 0, + "content": "5.2. Shadow Intensity Modulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.688 + ], + "angle": 0, + "content": "By using the adapted ControlNet in Section 5.1, we observe that the intensity of generated foreground shadow is unsatisfactory. Especially when the background has object-shadow pairs, the generated foreground shadow is often notably darker or brighter than background shadows. Such inconsistency between foreground shadow intensity and background shadow intensity makes the whole image unrealistic." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, we use encoder \\( E_{i} \\) to extract intensity-relevant information. Intuitively, by observing background shadows and its surrounding unshadowed areas, we can estimate the intensity of foreground shadows. Thus, the input of intensity encoder \\( E_{i} \\) should include the composite image \\( I_{c} \\) and background shadow mask \\( M_{bs} \\). When there is no background shadow, the mask is all black. We concatenate \\( I_{c} \\) with background shadow mask \\( M_{bs} \\) as the input of intensity encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The intensity encoder outputs scales and biases to adjust the intensity of noise map within the foreground shadow region. The modulated noise map results in the modulated latent image, and further results in the modulated foreground" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.577 + ], + "angle": 0, + "content": "shadow. Therefore, the intensity adjustment of noise map is finally embodied in the intensity variation of generated foreground shadow. Specifically, when the noise map has \\(c\\) channels, \\(E_{i}\\) outputs the \\(c\\)-dim scale vector \\(\\boldsymbol{s}\\) and \\(c\\)-dim bias vector \\(\\boldsymbol{b}\\), containing channel-wise scales and biases. \\(\\boldsymbol{s}\\) and \\(\\boldsymbol{b}\\) are used to modulate the predicted noise map within the foreground shadow region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.893, + 0.835 + ], + "angle": 0, + "content": "One problem is that the foreground shadow region is unknown in the testing stage, so we need to predict the foreground shadow mask. To avoid much extra computational cost, we take advantage of the feature maps in the denoising U-Net to predict the foreground shadow mask. Previous works usually combine different layers of feature maps in denoising U-Net for mask prediction [24, 43]. We try different layers of feature maps and find that decoder feature maps are more effective in shadow mask prediction. We also use foreground object mask, which could provide useful hints for the location of foreground shadow. We resize all decoder feature maps and foreground object mask to the same size, and concatenate them channel-wisely. The concatenation passes through several convolutional layers to predict the foreground shadow mask \\(\\tilde{M}_{fs}\\). \\(\\tilde{M}_{fs}\\) is supervised with ground-truth foreground shadow mask \\(M_{fs}\\) by Binary Cross-Entropy (BCE) loss and Dice loss [26]:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.844, + 0.892, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m a s k} = \\mathcal {L} _ {b c e} (\\tilde {M} _ {f s}, M _ {f s}) + \\mathcal {L} _ {d i c e} (\\tilde {M} _ {f s}, M _ {f s}). (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "When \\( t \\) is large, \\( z_{t} \\) is close to random noise and thus the decoder feature maps are not informative to predict shadow" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "mask. Hence, we only employ the loss \\(\\mathcal{L}_{mask}\\) when the time step \\(t\\) is small. We set the threshold of \\(t\\) as \\(\\sigma T\\), in which \\(T\\) is the total number of steps. Accordingly, shadow intensity modulation is only applied when \\(t\\) is smaller than the threshold \\(\\sigma T\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.167, + 0.468, + 0.274 + ], + "angle": 0, + "content": "Provided with the predicted foreground shadow mask \\(\\tilde{M}_{fs}\\), we can modulate the noise map within the foreground shadow region. Given the predicted noise map \\(\\tilde{\\epsilon} = \\epsilon_{\\theta}(z_t, t, c_{sg})\\), we multiply \\(\\tilde{\\epsilon}\\) by channel-wise scales \\(s\\) and add channel-wise biases \\(\\pmb{b}\\) to get \\(\\tilde{\\epsilon}'\\). Then, based on \\(\\tilde{M}_{fs}\\), we combine the modulated noise map and original noise map to get the final noise map: \\(\\hat{\\epsilon} = \\tilde{\\epsilon}' \\circ \\tilde{M}_{fs} + \\tilde{\\epsilon} \\circ (1 - \\tilde{M}_{fs})\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.275, + 0.468, + 0.304 + ], + "angle": 0, + "content": "We replace the predicted noise map in Eqn. (3) with the final noise map \\(\\hat{\\epsilon}\\) and get" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.314, + 0.468, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m w s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\boldsymbol {W} _ {f s} \\circ (\\epsilon - \\hat {\\epsilon}) \\| _ {2} ^ {2} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.468, + 0.378 + ], + "angle": 0, + "content": "We summarize the mask prediction loss in Eqn. (4) and weighted noise loss in Eqn. (5) as" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.389, + 0.468, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda \\mathcal {L} _ {\\text {m w s g}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.416, + 0.295, + 0.43 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a trade-off parameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.439, + 0.234, + 0.455 + ], + "angle": 0, + "content": "5.3. Post-processing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.468, + 0.613 + ], + "angle": 0, + "content": "We observe that the generated images could have color shift and background variation issues. Color shift means that the overall color tone deviates from the input composite image. Background variation means that some background details are changed. To solve these issues, we create a multi-task post-processing network which yields the rectified image together with the foreground shadow mask. Then, we combine input composite image and rectified image based on the predicted foreground shadow mask to produce the final image. The technical details are left to supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.209, + 0.644 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.364, + 0.667 + ], + "angle": 0, + "content": "6.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.839 + ], + "angle": 0, + "content": "We conduct experiments on both DESOBA [12] and our contributed DESOBAv2 dataset. We split DESOBAv2 into 21,088 training images with 27,718 tuples and 487 test images with 855 tuples. Following [12], the test set contains BOS images (with background object-shadow pairs) and BOS-free images. Most of our experiments are based on DESOBAv2 dataset due to the following two concerns: 1) DESOBAv2 has larger test set which supports more comprehensive evaluation. 2) DESOBA has the artifacts caused by manual shadow removal and the existing methods (e.g., SGRNet) tend to overfit such artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "For the generated results, we evaluate both image quality and mask quality. For image evaluation, following [12], we adopt RMSE and SSIM, which are calculated based on the ground-truth target image and the generated image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Global RMSE (GR) and Global SSIM (GS) are calculated over the whole image, while Local RMSE (LR) and Local SSIM (LS) are calculated over the ground-truth foreground shadow region. For the mask evaluation, following [12], we adopt Balanced Error Rate (BER), which is calculated based on the ground-truth binary foreground shadow mask and the predicted foreground shadow mask obtained by threshold 0.5. Global BER (GB) is calculated over the whole image, while Local BER (LB) is calculated over the ground-truth foreground shadow region. Note that diffusion model has stochastic property and shadow generation is a multi-modal task, that is, one input has multiple plausible outputs. Similar to multi-modal inpainting evaluation [54, 55], we generate 5 results for one test image with different random seeds and select the one closest to the ground-truth (the highest Local SSIM) to calculate evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.359, + 0.716, + 0.374 + ], + "angle": 0, + "content": "6.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.553 + ], + "angle": 0, + "content": "We develop our method with PyTorch 1.12.1 [30]. Our model is trained using the Adam optimizer [17] with a constant learning rate of \\( 1e^{-5} \\) over 50 epochs, on four NVIDIA RTX A6000 GPUs. Our method is built upon ControlNet [52]. We employ ResNet18 [10] as the intensity encoder. The mask predictor passes the concatenation of decoder feature maps and foreground object mask through four convolutional layers, with ReLU activation following the first three layers and Sigmoid activation following the last layer. We set the hyper-parameters \\( w \\), \\( \\sigma \\), and \\( \\lambda \\) as 10, 0.7, and 1, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.58, + 0.744, + 0.595 + ], + "angle": 0, + "content": "6.3. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Following [12], we compare with ShadowGAN [53], Mask-ShadowGAN [13], ARShadowGAN [22], and SGRNet [12]. We train and test all methods on DESOBAv2 dataset. The quantitative results are summarized in Table 1. We observe that our SGDiffusion achieves the lowest GRMSE, LRMSE and the highest GSSIM, LSSIM, which demonstrates that our method could generate shadow images that are closer to the ground-truth shadow images. The best GB and LB results demonstrate that the shapes and locations of our generated shadows are more accurate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For qualitative comparison, we show several example results in Figure 4. Compared with the baseline methods, the shadows produced by our model have more reasonable shapes and intensities. Moreover, as shown in row 1, our method can take into account the self-occlusion of objects to generate discontinuous shadows. As shown in row 4, our method can also consider the material of the objects, producing shadows with translucency effects. We provide more examples in the supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8126" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.103, + 0.195, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.103, + 0.291, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.103, + 0.387, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.103, + 0.483, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.103, + 0.581, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.103, + 0.678, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.103, + 0.776, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.103, + 0.874, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.179, + 0.194, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.179, + 0.291, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.179, + 0.387, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.179, + 0.484, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.179, + 0.581, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.179, + 0.678, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.179, + 0.776, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.179, + 0.874, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.254, + 0.194, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.254, + 0.291, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.254, + 0.387, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.254, + 0.484, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.254, + 0.581, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.254, + 0.678, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.254, + 0.776, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.254, + 0.874, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.33, + 0.194, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.412, + 0.157, + 0.427 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.33, + 0.291, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.412, + 0.255, + 0.427 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.33, + 0.387, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.412, + 0.351, + 0.427 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.33, + 0.484, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.412, + 0.448, + 0.427 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.33, + 0.581, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.412, + 0.545, + 0.427 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.33, + 0.678, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.412, + 0.641, + 0.427 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.33, + 0.776, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.412, + 0.74, + 0.427 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.33, + 0.874, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.412, + 0.835, + 0.427 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.454, + 0.895, + 0.497 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of different methods on DESOBAv2 dataset. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), our SGDiffusion (g), ground-truth (h)." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.512, + 0.912, + 0.635 + ], + "angle": 0, + "content": "
MethodBOS Test ImagesBOS-free Test Images
GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓
ShadowGAN [53]7.51167.4640.9610.1970.4460.89017.32576.5080.9010.0600.4250.842
MaskshadowGAN [13]8.99779.4180.9510.1800.5001.00019.33894.3270.9060.0440.5001.000
ARShadowGAN [22]7.33558.0370.9610.2410.3830.76116.06763.7130.9080.1040.3490.682
SGRNet [12]7.18468.2550.9640.2060.3010.59615.59660.3500.9090.1000.2710.534
SGDiffusion6.09853.6110.9710.3700.2450.48715.11055.8740.9130.1170.2330.452
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.657, + 0.808, + 0.672 + ], + "angle": 0, + "content": "Table 1. The results of different methods on DESOBAv2 dataset. The best results are highlighted in boldface." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.698, + 0.241, + 0.712 + ], + "angle": 0, + "content": "6.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.788 + ], + "angle": 0, + "content": "We study the impact of weighted noise loss (WL), intensity modulation (IM), and post-processing (PP) of our SGDiffusion on BOS test images from DESOBAv2. The quantitative results are summarized in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In row 1, we report the results of basic ControlNet without weighted noise loss. For WL, the comparison between row 3 and row 1 emphasizes the importance of paying more attention to the foreground shadow region. We also report a special case \\(\\dagger\\) in row 2, where the foreground shadow mask is not expanded when constructing the weight map. The results in row 2 are comparable or even worse than those in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.895, + 0.895 + ], + "angle": 0, + "content": "row 1, as the model tends to generate larger shadow size while ignoring shape and edge details. For IM, the comparison between row 1 and row 5 shows that the intensity modulation can significantly improve the shadow quality by adjusting the shadow intensity. We also report a special case in row 4, where the intensity encoder input does not contain background shadow mask. The comparison between row 4 and row 5 shows that background shadow mask is helpful, because the background shadow regions and their surrounding regions could provide useful clues to infer shadow intensity. For PP, the comparison between row 6 and row 7 demonstrates that post-processing effectively corrects color shift and background variations, substantially reducing the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.106, + 0.209, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.106, + 0.318, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.106, + 0.428, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.106, + 0.538, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.106, + 0.648, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.106, + 0.759, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.106, + 0.87, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.191, + 0.209, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.191, + 0.318, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.191, + 0.428, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.191, + 0.538, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.191, + 0.648, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.191, + 0.759, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.191, + 0.87, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.277, + 0.21, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.277, + 0.318, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.277, + 0.428, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.277, + 0.538, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.277, + 0.65, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.277, + 0.76, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.277, + 0.87, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.37, + 0.167, + 0.386 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.37, + 0.278, + 0.385 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.371, + 0.386, + 0.385 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.371, + 0.498, + 0.385 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.371, + 0.608, + 0.385 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.371, + 0.716, + 0.385 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.371, + 0.828, + 0.386 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.417, + 0.894, + 0.446 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of different methods on real composite images. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), SGDiffusion (g)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.47, + 0.47, + 0.601 + ], + "angle": 0, + "content": "
RowWLIMPPGR↓LR↓GB↓LB↓
1--+8.28559.7530.2710.534
2†-+8.31959.4910.2820.563
3+-+7.04153.8290.2490.492
4-○+7.41056.1210.2690.536
5-++7.35754.1590.2620.526
6++-13.44755.2310.2450.487
7+++6.09853.6110.2450.487
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Table 2. Ablation studies of our method on BOS test images from DESOBAv2 dataset. WL is short for weighted loss and \\(\\dagger\\) means without expanding shadow mask. IM is short for intensity modulation and \\(\\circ\\) means without using background shadow mask. PP is short for post-processing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.47, + 0.756 + ], + "angle": 0, + "content": "global RMSE. We also provide the visual results of ablated versions in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.295, + 0.788 + ], + "angle": 0, + "content": "6.5. Real Composite Images" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We compare different methods on real composite images provided by [12], where background images and foreground objects are from the DESOBA [12] test set. We train all methods on DESOBAv2 and finetune them on DESOBA. The visual results of different methods are showcased in Figure 5. These results confirm that SGDiffusion adeptly synthesizes lifelike shadows with precise contours, loca" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.547 + ], + "angle": 0, + "content": "tions, and directions, which are compatible with the background object-shadow pairs and foreground object information. In contrast, previous methods often produce vague and misdirected shadows. We provide more examples in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.893, + 0.669 + ], + "angle": 0, + "content": "Given the absence of ground-truth images for real composite images, following [12], we opt for subjective evaluation, engaging 50 human raters in the user study. Each participant is presented with image pairs from the results generated by 5 methods, and asked to choose the image with more realistic foreground shadow. Using the Bradley-Terry model [2], we report the B-T scores in the supplementary, which again proves the advantage of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.619, + 0.695 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.796 + ], + "angle": 0, + "content": "In this paper, we have contributed a large-scale shadow generation dataset DESOBAv2. We have also designed a novel diffusion-based shadow generation method. Extensive experimental results show that our method is able to generate plausible shadows for composite foregrounds, significantly surpassing previous methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.806, + 0.681, + 0.824 + ], + "angle": 0, + "content": "8. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.892, + 0.889 + ], + "angle": 0, + "content": "The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major/Key Project, China (Grant No. 2021SHZDZX0102)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8128" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Ibrahim Arief, Simon McCallum, and Jon Yngve Hardeberg. Realtime estimation of illumination direction for augmented reality on mobile devices. In CIC, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[3] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[4] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In ICME, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[5] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.348, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. TIP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagné, and Jean-Francois Lalonde. Deep parametric indoor lighting estimation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.449, + 0.468, + 0.49 + ], + "angle": 0, + "content": "[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.493, + 0.468, + 0.548 + ], + "angle": 0, + "content": "[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.468, + 0.592 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.623 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurlPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.625, + 0.468, + 0.665 + ], + "angle": 0, + "content": "[12] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. AAAI, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.668, + 0.468, + 0.71 + ], + "angle": 0, + "content": "[13] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, 2019. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.712, + 0.468, + 0.74 + ], + "angle": 0, + "content": "[14] Shaozong Huang and Lan Hong. Diffusion model for mural image inpainting. In ITOEC, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.468, + 0.797 + ], + "angle": 0, + "content": "[15] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM TOG, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.8, + 0.468, + 0.84 + ], + "angle": 0, + "content": "[16] Eric Kee, James F. O'Brien, and Hany Samir Farid. Exposing photo manipulation from shading and shadows. ACM TOG, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.843, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[19] Bin Liao, Yao Zhu, Chao Liang, Fei Luo, and Chunxia Xiao. Illumination animating and editing in a single picture using scene structure estimation. Computers & Graphics, 82:53-64, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[20] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[21] Bin Liu, Kun Xu, and Ralph R Martin. Static scene illumination estimation from videos with applications. JCST, 32 (3):430-442, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[22] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In CVPR, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[23] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[24] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. CoRR, abs/2303.17870, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[25] Quanling Meng, Shengping Zhang, Zonglin Li, Chenyang Wang, Weigang Zhang, and Qingming Huang. Automatic shadow generation via exposure fusion. IEEE Transactions on Multimedia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[26] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[28] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. CoRR, abs/2106.14490, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[29] Sibam Parida, Vignesh Srinivas, Bhavishya Jain, Rajesh Naik, and Neeraj Rao. Survey on diverse image inpainting using diffusion models. In PCEMS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NIPS, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[31] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In SIGGRAPH. 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In CVPR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8129" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[35] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[36] Yichen Sheng, Jianming Zhang, Julien Philip, Yannick Hold-Geoffroy, Xin Sun, He Zhang, Lu Ling, and Bedrich Benes. Pixht-lab: Pixel height based light effect generation for image compositing. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. CoRR, abs/2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[38] Yi-Zhe Song, Zhifei Zhang, Zhe L. Lin, Scott D. Cohen, Brian L. Price, Jianming Zhang, Soo Ye Kim, and Daniel G. Aliaga. Objectstitch: Generative object compositing. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[39] Xinhao Tao, Junyan Cao, Yan Hong, and Li Niu. Shadow generation with decomposed mask prediction and attentive shadow filling. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[40] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[41] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. TPAMI, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[42] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In ACM MM, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[43] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[44] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[45] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[46] Fangneng Zhan, Jiaxing Huang, and Shijian Lu. Adaptive composition gan towards realistic image synthesis. CoRR, abs/1905.04693, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[47] Fangneng Zhan, Shijian Lu, Changgong Zhang, Feiying Ma, and Xuansong Xie. Towards realistic 3d embedding via view alignment. CoRR, abs/2007.07066, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[48] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[49] He Zhang, Jianming Zhang, Federico Perazzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In WACV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[50] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[51] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In WACV, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[53] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5:105-115, 2019. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[54] Lei Zhao, Qihang Mo, Sihuan Lin, Zhizhong Wang, Zhiwen Zuo, Haibo Chen, Wei Xing, and Dongming Lu. Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[55] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[56] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In ECCV, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8130" + } + ] +] \ No newline at end of file diff --git a/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_origin.pdf b/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5552118bc3ba6dcc331d2d83ca5b1b24a76d6e1c --- /dev/null +++ b/2024/Shadow Generation for Composite Image Using Diffusion Model/d6e83b88-6d3b-4f78-b729-db99f8dc57c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caf170da98a760dbe75a3ee60beb5b07347726591797f7a5c6492d48321afda2 +size 6313329 diff --git a/2024/Shadow Generation for Composite Image Using Diffusion Model/full.md b/2024/Shadow Generation for Composite Image Using Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..152172636544c3d121814894b830138a11e3946f --- /dev/null +++ b/2024/Shadow Generation for Composite Image Using Diffusion Model/full.md @@ -0,0 +1,404 @@ +# Shadow Generation for Composite Image Using Diffusion Model + +Qingyang Liu1, Junqi You1, Jianting Wang1, Xinhao Tao1, Bo Zhang1, Li Niu1,2* +1 Shanghai Jiao Tong University 2 miguo.ai + +$^{1}$ narumimaria, yjqsjtu2022, gloryl1299, taoxinhao, bo-zhang, ustcnewly}@sjtu.edu.cn + +# Abstract + +In the realm of image composition, generating realistic shadow for the inserted foreground remains a formidable challenge. Previous works have developed image-to-image translation models which are trained on paired training data. However, they are struggling to generate shadows with accurate shapes and intensities, hindered by data scarcity and inherent task complexity. In this paper, we resort to foundation model with rich prior knowledge of natural shadow images. Specifically, we first adapt ControlNet to our task and then propose intensity modulation modules to improve the shadow intensity. Moreover, we extend the small-scale DESOBA dataset to DESOBAv2 using a novel data acquisition pipeline. Experimental results on both DESOBA and DESOBAv2 datasets as well as real composite images demonstrate the superior capability of our model for shadow generation task. The dataset, code, and model are released at https://github.com/bcmi/Object-Shadow-Generation-Dataset-DESOBAv2. + +# 1. Introduction + +Image composition [28] aims to merge the foreground of one image with another background image to produce a composite image, which has a wide range of applications like virtual reality, artistic creation, and E-commerce. Simply pasting the foreground onto the background often results in visual inconsistencies, including the incompatible illumination between foreground and background [3], lack of foreground shadow/reflection [12, 34], and so on. In this paper, we focus on the shadow issue, i.e., the inserted foreground does not have plausible shadow on the background, which could significantly degrade the realism and quality of composite image. + +As illustrated in Figure 1, shadow generation is a challenging task because the foreground shadow is determined by many complicated factors like the lighting information and the geometry of foreground/background. The exist- + +![](images/19751b68b323e2547de1ce8434ae68dc5be14c2111fec618e4da7c3f1cabc42a.jpg) +Figure 1. A composite image can be obtained by pasting the foreground on the background. Shadow generation aims to generate plausible shadow for the inserted foreground in the composite image to produce a more realistic image. + +ing shadow generation methods can be divided into rendering based methods [34-36] and non-rendering based methods [12, 22, 53]. Rendering based methods usually impose strict assumptions on the geometry and lighting, which could hardly be satisfied in real-world scenarios. Besides, [35, 36] require users to specify the lighting information, which hinders its direct application in our task. Non-rendering based methods usually train an image-to-image translation network, based on pairs of composite images without foreground shadows and real images with foreground shadows. However, due to the training data scarcity and task difficulty, these methods are struggling to generate shadows with reasonable shapes and intensities. + +Recently, foundation model (e.g., stable diffusion [32]) pretrained on large-scale dataset has demonstrated unprecedented potential for image generation and editing. In previous works [44, 48] on object-guided inpainting or composition, they show that the generated foregrounds are accompanied by shadows even without considering the shadow issue, probably because of the rich prior knowledge of natural shadow images in foundation model. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly. + +We build our method upon conditional foundation model [52] and propose several key innovations. First, we modify the control encoder input and the noise loss to fit our task. Then, we observe that the generated shadow intensity (the level of darkness) is unsatisfactory. Especially when the background objects has shadows, the intensity inconsistency between foreground shadow and background shadows make the whole image unrealistic. Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, the denoising U-Net is modified to output both noise map and foreground shadow mask. The intensity encoder takes in the composite image and background shadow mask, producing the scale/bias to modulate the predicted noise within the foreground shadow region. Finally, we devise a post-processing network to rectify the color shift and background variation. + +The model training requires abundant pairs of composite images without foreground shadows and real images with foreground shadows. The existing real-world shadow generation dataset DESOBA [12] is limited by scale (i.e., 1,012 real images and 3,623 pairs) due to the high cost of manual shadow removal, which is insufficient to train our model. To ensure sufficient supervision, we design a novel data construction pipeline, which extends DESOBA to DESOBAv2 (i.e., 21,575 real images and 28,573 pairs) using object-shadow detection and inpainting techniques. Specifically, we first collect a large number of real-world images with one or more object-shadow pairs. Then, we use pretrained object-shadow detection model [41] to predict object and shadow masks for object-shadow pairs. Next, we apply pretrained inpainting model [32] to inpaint the detected shadow regions to get desshadowed images. Finally, based on real images and desshadowed images, we construct pairs of synthetic composite images and ground-truth target images. + +We conduct experiments on both DESOBAv2 and DESOBA datasets. The results reveal remarkable improvement in shadow generation task, after leveraging the benefits of large-scale data and foundation model. Our main contributions can be summarized as follows: 1) We contribute DESOBAv2, a large-scale real-world shadow generation dataset, which could greatly facilitate the shadow generation task. 2) We propose a cutting-edge diffusion model specifically designed to produce shadows for the composite foregrounds. 3) Through comprehensive experiments, we validate the efficacy of our dataset construction pipeline and the superiority of our proposed model. + +# 2. Related Work + +# 2.1. Image Composition + +Image composition aims to overlay a foreground object on a background image to yield a composite result [20, 22, 42, 46, 47]. Previous research works have tackled differ + +ent issues that can compromise the quality of composite images. For instance, image blending methods [31, 42, 49, 51] target at combining the foreground and background seamlessly. Image harmonization methods [3-6, 40] aim to rectify the illumination disparity between foreground and background. Nonetheless, the above methods largely overlook the shadow cast by the foreground onto the background. Recently, generative image composition methods [38, 44, 48] can insert a foreground object into a bounding box in the background and the inserted object is likely to have shadow effect. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly. + +# 2.2. Shadow Generation + +In this paper, the goal of shadow generation task is generating plausible shadow for the composite foreground. Existing methods can be broadly categorized into rendering based methods and non-rendering based methods. The rendering based methods necessitate a comprehensive understanding of factors like illumination, reflectance, material properties, and scene geometry to produce shadows for the inserted objects. However, such detailed knowledge relies on user input [15, 16, 21, 35, 36] or model prediction [1, 7, 19, 50], which is either labor-intensive or unreliable [53]. For example, [35, 36] could produce compelling results with user control. However, in the composite image, the lighting information should be inferred automatically from background instead of requested by users. + +Non-rendering based methods [12, 22, 25, 53] aim to translate an input composite image without foreground shadow to an output with foreground shadow, bypassing the need for explicit knowledge of the aforementioned factors. For instance, ShadowGAN [53] utilizes both global and local conditional discriminator to enhance the realism of generated shadows. ARShadowGAN [22] emphasizes the importance of background shadow and uses it to guide foreground shadow generation. SGRNet [12] encourages the information exchange between foreground and background, and employs a classic illumination model for better shadow effect. The work [25] produces multiple underexposure images and fuses them to get the final shadow region. DMASNet [39] decomposes shadow mask prediction into box prediction and shape prediction, achieving better cross-domain transferability. + +To the best of our knowledge, we are the first diffusion-based method focusing on shadow generation. + +# 2.3. Diffusion Models + +In recent years, diffusion models have emerged as a powerful tool in image generation and image editing. These models approach image generation as a series of stochastic transitions, moving from a basic distribution to the desired + +![](images/f81995dc9c2606ff9b9ede849f50101ef2deb2f66d4b156033fb36b684386cb5.jpg) +Figure 2. The pipeline of dataset construction. We use object-shadow detection model [41] to predict pairs of object and shadow masks in the real image $I_r$ . Then we obtain the union $M_s$ of all shadow masks as the inpainting mask and apply inpainting model [32] to get a desshadowed image $I_d$ . After designating a foreground object, we replace the background shadow regions $M_{bs}$ in $I_d$ with the counterparts in $I_r$ to synthesize a composite image $I_c$ , and replace all the shadow regions $M_s$ in $I_d$ with the counterparts in $I_r$ to obtain the ground-truth target image $I_g$ . + +data distribution [11]. Diffusion models can be divided into unconditional diffusion models [11, 37] and conditional diffusion models [27, 32, 52]. Unconditional diffusion models focus on generating realistic images by capturing the distribution of natural images, without the need of any specific input conditions. Conditional diffusion models are designed to produce images under the guidance of specific conditional inputs, such as text descriptions, semantic masks, and so on. ControlNet [52] is a popular conditional diffusion model, which equips large pretrained text-to-image diffusion models with spatial-aware and task-specific conditions. We build our model upon ControlNet and propose several innovations to meet the specific requirements of shadow generation. + +# 3. Dataset Construction + +The pipeline of our dataset construction is illustrated in Figure 2, which will be detailed next. + +# 3.1. Shadow Image Collection + +We harvest an extensive collection of real-world outdoor images with natural lighting across various scenes from two sources. On one hand, we crawl online images from public websites that have licenses for reuse. On the other hand, we hire photographers to capture photos in the outdoor scenes that satisfy our requirements. We only preserve the images with at least one object-shadow pair, arriving at 44,044 images. + +# 3.2. Shadow Removal + +Given a real image $I_r$ with object-shadow pairs, we use the pretrained object-shadow detection model [41] to predict $K$ + +pairs of object and shadow masks. We use $M_{o,k}$ (resp., $M_{s,k}$ ) to denote the object (resp., shadow) mask of the $k$ -th object. We refer to one detected object-shadow pair as one detected instance. We eliminate the images without any detected instance. + +Subsequently, we attempt to erase all the detected shadows. We have tried some state-of-the-art shadow removal models [8, 9], but the performance in the wild is below our expectation due to poor generalization ability. Considering the recent rapid advance of image inpainting [14, 23, 29, 32, 45, 56] techniques, we resort to image inpainting to remove the shadows. Although image inpainting cannot preserve the background information precisely, we observe that the background textures in the shadow region are usually very simple, and the inpainted result has similar textures with the original background. Thus, we roughly treat the inpainted results as desshadowed results. + +We obtain the union of all detected shadow masks $M_{s} = M_{s,1} \cup M_{s,2} \cup \dots \cup M_{s,K}$ as the inpainting mask and apply the pretrained inpainting model [32] to get a deshadowed image $I_{d}$ . In practice, we observe that the inpainting model is prone to generate low-quality shadow in the inpainted region in some cases. To prevent the inpainting model from generating undesirable shadows in the inpainted region, we adopt some tricks like dilating the inpainting mask and flipping images vertically, which can effectively obstruct undesirable shadow generation during inpainting. However, there may still exist undesirable shadows or noticeable artifacts in the inpainted region. + +After inpainting, we manually filter the object-shadow pairs according to the following rules: 1) We remove the object-shadow pairs with low-quality object masks or shadow masks. 2) We remove those object-shadow pairs + +with generated shadows or noticeable artifacts in the inpainted region. After manual filtering, we refer to the remaining object-shadow pairs as valid instances. We have 21,575 images with 28,573 valid instances. + +# 3.3. Composite Image Synthesis + +Given a pair of a real image $I_r$ and a deshadowed image $I_d$ , we randomly select the $k$ -th foreground object from valid instances and synthesize the composite image. $M_{o,k}$ (resp., $M_{s,k}$ ) is referred to as the foreground object (resp., shadow) mask $M_{f_o}$ (resp., $M_{f_s}$ ). One strategy is replacing the shadow region $M_{fs}$ of this foreground object in $I_r$ with the counterpart in $I_d$ to erase the foreground shadow. However, this strategy may leave traces along the shadow boundary, in which case the model may find a shortcut to generate the shadow. Another strategy is replacing the shadow regions $M_{bs} = M_{s,1} \cup \dots \cup M_{s,k-1} \cup M_{s,k+1} \cup \dots \cup M_{s,K}$ of the other objects in $I_d$ with the counterparts in $I_r$ to synthesize a composite image $I_c$ , in which only the selected foreground object does not have shadow while all the other objects have shadows. We adopt the second strategy. + +After inpainting, the background may undergo slight changes, so the background of $I_{c}$ may be slightly different from that of $I_{r}$ . To ensure consistent background, we obtain the ground-truth target image $I_{g}$ by replacing the shadow regions $M_{s}$ of all objects in $I_{d}$ with the counterparts in $I_{r}$ . Then, $I_{c}$ and $I_{g}$ form a pair of input composite image and ground-truth target image. So far, we obtain tuples in the form of $\{I_{c}, M_{fo}, M_{fs}, M_{bs}, I_{g}\}$ , which will be used for model training. Example images and more statistics of our dataset can be found in the supplementary. + +# 4. Background + +Stable Diffusion [32] is latent diffusion model operating in a latent space. First, $512 \times 512$ images are converted to $64 \times 64$ latent images using VAE [18] with encoder $E_r$ and decoder $D_r$ . The image space is projected to the latent space using $E_r$ , and back to the image space using $D_r$ . Then, the forward diffusion process and backward denoising process are performed in the latent space. The denoising U-Net [33] consists of an encoder with 12 blocks, a middle block, and a skip-connected decoder with 12 blocks. + +During training, random Gaussian noise $\epsilon$ is added to the latent image $z_0$ in the denoising step $t$ , producing a noisy latent image $z_t$ . Given time step $t$ and text prompt $c_{txt}$ , the denoising U-Net with model parameters $\epsilon_\theta$ is trained to predict the added noise $\epsilon$ . + +To support spatial conditional information (e.g., edge, pose, depth), ControlNet [52] integrates a control encoder $E_{c}$ with pre-trained Stable Diffusion. Specifically, the control encoder contains trainable replicas of its 12 encoding blocks and middle block across four resolutions + +$(64 \times 64, 32 \times 32, 16 \times 16, 8 \times 8)$ . It takes a $512 \times 512$ conditional image as input. + +The conditional feature maps $c_{img}$ output from control encoder are used to enhance the 12 skip-connections and middle block in denoising U-Net via zero convolution layers. While the original Stable Diffusion is fixed to retain prior knowledge, control encoder could incorporate additional conditions to guide image generation. The objective could be rewritten as + +$$ +\mathcal {L} _ {c t r l} = \mathbb {E} _ {t, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon - \epsilon_ {\boldsymbol {\theta}} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {c} _ {t x t}, \boldsymbol {c} _ {i m g}\right) \| _ {2} ^ {2} \right]. \tag {1} +$$ + +# 5. Method + +Given a composite image $I_{c}$ without foreground shadow as well as the foreground object mask $M_{fo}$ , our Shadow Generation Diffusion (SGDiffusion) model aims to produce $\tilde{I}_{g}$ with plausible foreground shadow. We will adapt ControlNet [52] to shadow generation task in Section 5.1, and propose novel modules to improve the shadow intensity in Section 5.2. Finally, we will briefly introduce post-processing techniques to enhance the image quality in Section 5.3. + +# 5.1. Adapting ControlNet to Shadow Generation + +For shadow generation task, the useful conditional information is input composite image $I_{c}$ and foreground object mask $M_{fo}$ , in which the foreground object mask indicates the target object we need to generate shadow for. We concatenate $I_{c}$ with $M_{fo}$ as the input of control encoder $E_{c}$ . The control encoder outputs the conditional feature maps $c_{sg}$ , which are injected into the denoising decoder to provide guidance. For the text prompt, we have tried several variants like "the [object category] with shadow", but they have no significant impact on the generated shadows. Therefore, we use null text prompt by default. + +Given a set of conditions including time step $t$ and conditional feature maps $c_{sg}$ , the denoising U-Net with model parameters $\epsilon_{\theta}$ predicts the noise $\epsilon$ added to the noisy latent image $z_{t}$ : + +$$ +\mathcal {L} _ {s g} = \mathbb {E} _ {t, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon - \epsilon_ {\boldsymbol {\theta}} \left(\boldsymbol {z} _ {t}, t, \boldsymbol {c} _ {s g}\right)) \| _ {2} ^ {2} \right]. \tag {2} +$$ + +To enforce the model to place more emphasis on the foreground shadow region, we introduce weighted noise loss, which assigns higher weights to the foreground shadow region. We expand the foreground shadow mask by a dilated kernel to get the expanded mask $\hat{M}_{fs}$ . The weights in the expanded foreground shadow region are $w$ while the other weights are 1, leading to the weight map $W_{fs}$ . If we do not expand the foreground shadow region, the model will be misled to generate large shadows, overlooking the details of shadow shapes and boundaries. By applying weight map $W_{fs}$ to the noise loss, we can arrive at + +$$ +\mathcal {L} _ {w s g} = \mathbb {E} _ {t, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \boldsymbol {W} _ {f s} \circ (\epsilon - \epsilon_ {\boldsymbol {\theta}} (\boldsymbol {z} _ {t}, t, \boldsymbol {c} _ {s g})) \| _ {2} ^ {2} \right], \tag {3} +$$ + +![](images/976a603a173f618365fdf99dfad9378cc01405353e58f5ff3b4c84cc834122c7.jpg) +Figure 3. The framework of our SGDiffusion. We adapt ControlNet (Control Encoder and Denoising U-Net) to shadow generation task. We also introduce an intensity encoder to modulate the foreground shadow region in the noise map $\tilde{\epsilon}$ , leading to $\hat{\epsilon}$ . The output noise $\hat{\epsilon}$ is supervised by weighted noise loss $\mathcal{L}_{mwsg}$ based on the expanded foreground shadow mask $\tilde{M}_{fs}$ + +where $\circ$ denotes element-wise multiplication. + +During inference, to retain more information of input composite image $I_{c}$ in the initial noise, we obtain $z_{T}$ by adding noise to the latent image of $I_{c}$ , rather than directly sampling from the Gaussian distribution $\mathcal{N}(0,1)$ . + +# 5.2. Shadow Intensity Modulation + +By using the adapted ControlNet in Section 5.1, we observe that the intensity of generated foreground shadow is unsatisfactory. Especially when the background has object-shadow pairs, the generated foreground shadow is often notably darker or brighter than background shadows. Such inconsistency between foreground shadow intensity and background shadow intensity makes the whole image unrealistic. + +Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, we use encoder $E_{i}$ to extract intensity-relevant information. Intuitively, by observing background shadows and its surrounding unshadowed areas, we can estimate the intensity of foreground shadows. Thus, the input of intensity encoder $E_{i}$ should include the composite image $I_{c}$ and background shadow mask $M_{bs}$ . When there is no background shadow, the mask is all black. We concatenate $I_{c}$ with background shadow mask $M_{bs}$ as the input of intensity encoder. + +The intensity encoder outputs scales and biases to adjust the intensity of noise map within the foreground shadow region. The modulated noise map results in the modulated latent image, and further results in the modulated foreground + +shadow. Therefore, the intensity adjustment of noise map is finally embodied in the intensity variation of generated foreground shadow. Specifically, when the noise map has $c$ channels, $E_{i}$ outputs the $c$ -dim scale vector $\boldsymbol{s}$ and $c$ -dim bias vector $\boldsymbol{b}$ , containing channel-wise scales and biases. $\boldsymbol{s}$ and $\boldsymbol{b}$ are used to modulate the predicted noise map within the foreground shadow region. + +One problem is that the foreground shadow region is unknown in the testing stage, so we need to predict the foreground shadow mask. To avoid much extra computational cost, we take advantage of the feature maps in the denoising U-Net to predict the foreground shadow mask. Previous works usually combine different layers of feature maps in denoising U-Net for mask prediction [24, 43]. We try different layers of feature maps and find that decoder feature maps are more effective in shadow mask prediction. We also use foreground object mask, which could provide useful hints for the location of foreground shadow. We resize all decoder feature maps and foreground object mask to the same size, and concatenate them channel-wisely. The concatenation passes through several convolutional layers to predict the foreground shadow mask $\tilde{M}_{fs}$ . $\tilde{M}_{fs}$ is supervised with ground-truth foreground shadow mask $M_{fs}$ by Binary Cross-Entropy (BCE) loss and Dice loss [26]: + +$$ +\mathcal {L} _ {m a s k} = \mathcal {L} _ {b c e} (\tilde {M} _ {f s}, M _ {f s}) + \mathcal {L} _ {d i c e} (\tilde {M} _ {f s}, M _ {f s}). (4) +$$ + +When $t$ is large, $z_{t}$ is close to random noise and thus the decoder feature maps are not informative to predict shadow + +mask. Hence, we only employ the loss $\mathcal{L}_{mask}$ when the time step $t$ is small. We set the threshold of $t$ as $\sigma T$ , in which $T$ is the total number of steps. Accordingly, shadow intensity modulation is only applied when $t$ is smaller than the threshold $\sigma T$ . + +Provided with the predicted foreground shadow mask $\tilde{M}_{fs}$ , we can modulate the noise map within the foreground shadow region. Given the predicted noise map $\tilde{\epsilon} = \epsilon_{\theta}(z_t, t, c_{sg})$ , we multiply $\tilde{\epsilon}$ by channel-wise scales $s$ and add channel-wise biases $\pmb{b}$ to get $\tilde{\epsilon}'$ . Then, based on $\tilde{M}_{fs}$ , we combine the modulated noise map and original noise map to get the final noise map: $\hat{\epsilon} = \tilde{\epsilon}' \circ \tilde{M}_{fs} + \tilde{\epsilon} \circ (1 - \tilde{M}_{fs})$ . + +We replace the predicted noise map in Eqn. (3) with the final noise map $\hat{\epsilon}$ and get + +$$ +\mathcal {L} _ {m w s g} = \mathbb {E} _ {t, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \boldsymbol {W} _ {f s} \circ (\epsilon - \hat {\epsilon}) \| _ {2} ^ {2} \right]. \tag {5} +$$ + +We summarize the mask prediction loss in Eqn. (4) and weighted noise loss in Eqn. (5) as + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {m a s k}} + \lambda \mathcal {L} _ {\text {m w s g}}, \tag {6} +$$ + +where $\lambda$ is a trade-off parameter. + +# 5.3. Post-processing + +We observe that the generated images could have color shift and background variation issues. Color shift means that the overall color tone deviates from the input composite image. Background variation means that some background details are changed. To solve these issues, we create a multi-task post-processing network which yields the rectified image together with the foreground shadow mask. Then, we combine input composite image and rectified image based on the predicted foreground shadow mask to produce the final image. The technical details are left to supplementary. + +# 6. Experiments + +# 6.1. Datasets and Evaluation Metrics + +We conduct experiments on both DESOBA [12] and our contributed DESOBAv2 dataset. We split DESOBAv2 into 21,088 training images with 27,718 tuples and 487 test images with 855 tuples. Following [12], the test set contains BOS images (with background object-shadow pairs) and BOS-free images. Most of our experiments are based on DESOBAv2 dataset due to the following two concerns: 1) DESOBAv2 has larger test set which supports more comprehensive evaluation. 2) DESOBA has the artifacts caused by manual shadow removal and the existing methods (e.g., SGRNet) tend to overfit such artifacts. + +For the generated results, we evaluate both image quality and mask quality. For image evaluation, following [12], we adopt RMSE and SSIM, which are calculated based on the ground-truth target image and the generated image. + +Global RMSE (GR) and Global SSIM (GS) are calculated over the whole image, while Local RMSE (LR) and Local SSIM (LS) are calculated over the ground-truth foreground shadow region. For the mask evaluation, following [12], we adopt Balanced Error Rate (BER), which is calculated based on the ground-truth binary foreground shadow mask and the predicted foreground shadow mask obtained by threshold 0.5. Global BER (GB) is calculated over the whole image, while Local BER (LB) is calculated over the ground-truth foreground shadow region. Note that diffusion model has stochastic property and shadow generation is a multi-modal task, that is, one input has multiple plausible outputs. Similar to multi-modal inpainting evaluation [54, 55], we generate 5 results for one test image with different random seeds and select the one closest to the ground-truth (the highest Local SSIM) to calculate evaluation metrics. + +# 6.2. Implementation Details + +We develop our method with PyTorch 1.12.1 [30]. Our model is trained using the Adam optimizer [17] with a constant learning rate of $1e^{-5}$ over 50 epochs, on four NVIDIA RTX A6000 GPUs. Our method is built upon ControlNet [52]. We employ ResNet18 [10] as the intensity encoder. The mask predictor passes the concatenation of decoder feature maps and foreground object mask through four convolutional layers, with ReLU activation following the first three layers and Sigmoid activation following the last layer. We set the hyper-parameters $w$ , $\sigma$ , and $\lambda$ as 10, 0.7, and 1, respectively. + +# 6.3. Comparison with Baselines + +Following [12], we compare with ShadowGAN [53], Mask-ShadowGAN [13], ARShadowGAN [22], and SGRNet [12]. We train and test all methods on DESOBAv2 dataset. The quantitative results are summarized in Table 1. We observe that our SGDiffusion achieves the lowest GRMSE, LRMSE and the highest GSSIM, LSSIM, which demonstrates that our method could generate shadow images that are closer to the ground-truth shadow images. The best GB and LB results demonstrate that the shapes and locations of our generated shadows are more accurate. + +For qualitative comparison, we show several example results in Figure 4. Compared with the baseline methods, the shadows produced by our model have more reasonable shapes and intensities. Moreover, as shown in row 1, our method can take into account the self-occlusion of objects to generate discontinuous shadows. As shown in row 4, our method can also consider the material of the objects, producing shadows with translucency effects. We provide more examples in the supplementary. + +![](images/e17d3a2538d52dc5db075b9d47d110dcbe3cfad68d4be178861b9ac64deddb78.jpg) + +![](images/b5ad355ce11dd36cc7fb3a30f930573ad3b90c86eeeb4ae1d7fabfd4167b9bea.jpg) + +![](images/a03e7b338c2b6708488c996507271c8dec5af53968fac8ddd7f43251133770dd.jpg) + +![](images/71b88ff84bb7f4d84a21af979a9f912b07f629e3943fe97c04aba36550886f20.jpg) + +![](images/c2405d23ecb5e5a9c7c3af460adb50f176a321633030f34976a6897dfd5b737a.jpg) + +![](images/1c38209021a2dce44d0f3ee3df1323102680dbf5bf5e7f104b391cd81c3eba8d.jpg) + +![](images/faf3c83ffeeae0b959067e0a7226af1d1f044d0bb5ece80c7cd7cf97f15114ba.jpg) + +![](images/27e99ab1b67aaf69575769ea323efa6646874b56a93bb6f6d26835030c554abc.jpg) + +![](images/7b1636b5f955b36109bb158be39f9a1ffc65d1c82a4f7bce6a74d5d15f74132d.jpg) + +![](images/7c636852e0334ef3b043660ab486ed75e753c9272f7e24ce8d410bb103816197.jpg) + +![](images/bb021e3f20ca4010d8e869de53546cc40161ddac343853185e6faf779aa2d6f4.jpg) + +![](images/9e5792c07c83fb3d6f2d46ebb869967d7d3ce6b18de76e68324523a50da456e2.jpg) + +![](images/c671801a469a75b781f4293a0f45d6ca0453f65ee0d5c43c9998e5392f49a78f.jpg) + +![](images/59c13068a9a375b067bc7f093017cc641880acb88d08b6a07a964145553cf71d.jpg) + +![](images/84e69e7792d30621b7939e09ff5dd20dafb38f8e95acd984d71b26f6791e34b3.jpg) + +![](images/c4bc469259e2b7d520bceab4e9275a6925136ab0b7df4c1ad862907753123c1c.jpg) + +![](images/d037949ea397c53342ee0e20ceecdf202b984cf9e6b3d00ee3ecb4c7968c66de.jpg) + +![](images/068abca7ecd3657ba5f48bc24051bf516283bbb722d0affec1255aaa27801bc4.jpg) + +![](images/dd006c10810b049c00cfa8944b5366f1139dd728ef02f576fee6e186047e5323.jpg) + +![](images/db4b852d6a9b4a7ebca692b96ef98b208edfdce63ae521a0915dcce8a703439b.jpg) + +![](images/ee7fe548c9efd7438a343f373abb42be59134523a24368486f3b233007eb0876.jpg) + +![](images/84de971930cabf625e5a41ef0540156f3242f5e745b3194cb57e6f958180900c.jpg) + +![](images/18b948ba09a6134f76f2054bde2f0677b154024f87bdbe72d14d4cb906bd081f.jpg) + +![](images/4a15da3c17616dce5a306c69a508a168bd766ec5a5caacd3a6d72abbb960fb71.jpg) + +![](images/4be851cd4d740da59790ce93174b173f10546874b87bed2f7468e8c268ee7507.jpg) +(a) + +![](images/da73d8df4e2f3de3251f85108a2828f7a2f8628a3fa9fd755e64553088a72d1a.jpg) +(b) + +![](images/6623fdb43e19ee7c114059eb9863788d07b984d544b084fb073d60e84cd5b49e.jpg) +(c) + +![](images/f406e079ac0961b77c2bd8c55fee98e7595757c6c16672830210232023c21121.jpg) +(d) + +![](images/4dc0a8fb9212b6b99741dfc57b44987f9458198e8672a6f9551a9702e532d8f8.jpg) +(e) +Figure 4. Visual comparison of different methods on DESOBAv2 dataset. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), our SGDiffusion (g), ground-truth (h). + +![](images/61ed1e035026ea38d9d2b601550403d1aa4151b044378f1495e7509096b06adc.jpg) +(f) + +![](images/09d862ad07d39b42adfc6452f15f3f36be05c66ee2dc7226085bf954e41a71aa.jpg) +(g) + +![](images/45ba09f138e0343c871944764b42420e4866ddf088c07a14d5320df44c2490d5.jpg) +(h) + +
MethodBOS Test ImagesBOS-free Test Images
GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓
ShadowGAN [53]7.51167.4640.9610.1970.4460.89017.32576.5080.9010.0600.4250.842
MaskshadowGAN [13]8.99779.4180.9510.1800.5001.00019.33894.3270.9060.0440.5001.000
ARShadowGAN [22]7.33558.0370.9610.2410.3830.76116.06763.7130.9080.1040.3490.682
SGRNet [12]7.18468.2550.9640.2060.3010.59615.59660.3500.9090.1000.2710.534
SGDiffusion6.09853.6110.9710.3700.2450.48715.11055.8740.9130.1170.2330.452
+ +Table 1. The results of different methods on DESOBAv2 dataset. The best results are highlighted in boldface. + +# 6.4. Ablation Studies + +We study the impact of weighted noise loss (WL), intensity modulation (IM), and post-processing (PP) of our SGDiffusion on BOS test images from DESOBAv2. The quantitative results are summarized in Table 2. + +In row 1, we report the results of basic ControlNet without weighted noise loss. For WL, the comparison between row 3 and row 1 emphasizes the importance of paying more attention to the foreground shadow region. We also report a special case $\dagger$ in row 2, where the foreground shadow mask is not expanded when constructing the weight map. The results in row 2 are comparable or even worse than those in + +row 1, as the model tends to generate larger shadow size while ignoring shape and edge details. For IM, the comparison between row 1 and row 5 shows that the intensity modulation can significantly improve the shadow quality by adjusting the shadow intensity. We also report a special case in row 4, where the intensity encoder input does not contain background shadow mask. The comparison between row 4 and row 5 shows that background shadow mask is helpful, because the background shadow regions and their surrounding regions could provide useful clues to infer shadow intensity. For PP, the comparison between row 6 and row 7 demonstrates that post-processing effectively corrects color shift and background variations, substantially reducing the + +![](images/5a77dbfb90c264b8c2247d452a947b5b11b357f2f9605ffa3eee469f3375b944.jpg) + +![](images/a550b99cae36e0890550c8265d965f72d9faef0ca07747ea98f20e914052fe90.jpg) + +![](images/ee63cb7b91a9ced50431d72d1d7f09168bf541328b8ddbabfd8d56ef03fca9d6.jpg) + +![](images/0ceca2d6bfaa2990472ac7e489e40718d3b1586153d1972d86e692e61537a11d.jpg) + +![](images/3c138983ce75567ea481cc3f28c4b6db9a1fa3984e2a62a81a5257bf4fa7bab3.jpg) + +![](images/720f8cf315f5c7c21e8af4b16451ff3a3fa03eb2672a9de84ac2fea14795a662.jpg) + +![](images/9035dcbe44ad55d64ad412cfa7b53b83c1fe56940136a0d77c38e3244942da3f.jpg) + +![](images/a2b2fbc67b24c9ed9f97d1a8b35b288a573dae4ae29947b52671d4957560cf4e.jpg) + +![](images/d6269f88ea18a3ef521f98bc0ff94fd8d72e9e18fe9025c57737109379172661.jpg) + +![](images/21a03b159c9bbb8fff0123bfd57097a82dad8c49bb2f2dc5f2034addefac9cc8.jpg) + +![](images/94e2f4a949c725541f47dadea13d618553d0ace03fb20b9c321eb98d6b04c8de.jpg) + +![](images/862abce213ed04ccc67149d11dc90fddc48331f81641df9d61e994f59ff46f8d.jpg) + +![](images/94a3472d4d3f7b6b1b047d1a277a33d6225ff2736c2532c23cd980aefd1e5830.jpg) + +![](images/f984aab0d189cec403acfa49a2dcb0987b5f5ea1957a11c0b2034f5f45833a24.jpg) + +![](images/ed1fef56431338ad1362df518088686fff4bd6a242e9a6819a8056e85ad3ae7e.jpg) +(a) + +![](images/5554a5c482fdebafb1dbeaf7ecf6044a6e25687c6a5fb7e6323e0423e41d7ba5.jpg) +(b) + +![](images/060331219053558fc6bcecd058e7c6302128bd171e128583d40122fef4e607b8.jpg) +(c) + +![](images/057d50aefabba5f5c5108789ebe54ca984c8944ccc410196b2eb3ce67b58330d.jpg) +(d) +Figure 5. Visual comparison of different methods on real composite images. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), SGDiffusion (g). + +![](images/5cdd0959c16b8d6dca11b9a5849da66a80e27a483893372ba668c4da84179310.jpg) +(e) + +![](images/51c7b4577166f181f30d2a083fbab261eb666f34aa7d1ed12b233ee934c1200e.jpg) +(f) + +![](images/5873fc882610db1e3f148726dc21fe9d45d801e3a6a2776ab94f2a99ff934df8.jpg) +(g) + +
RowWLIMPPGR↓LR↓GB↓LB↓
1--+8.28559.7530.2710.534
2†-+8.31959.4910.2820.563
3+-+7.04153.8290.2490.492
4-○+7.41056.1210.2690.536
5-++7.35754.1590.2620.526
6++-13.44755.2310.2450.487
7+++6.09853.6110.2450.487
+ +Table 2. Ablation studies of our method on BOS test images from DESOBAv2 dataset. WL is short for weighted loss and $\dagger$ means without expanding shadow mask. IM is short for intensity modulation and $\circ$ means without using background shadow mask. PP is short for post-processing. + +global RMSE. We also provide the visual results of ablated versions in the supplementary. + +# 6.5. Real Composite Images + +We compare different methods on real composite images provided by [12], where background images and foreground objects are from the DESOBA [12] test set. We train all methods on DESOBAv2 and finetune them on DESOBA. The visual results of different methods are showcased in Figure 5. These results confirm that SGDiffusion adeptly synthesizes lifelike shadows with precise contours, loca + +tions, and directions, which are compatible with the background object-shadow pairs and foreground object information. In contrast, previous methods often produce vague and misdirected shadows. We provide more examples in the supplementary. + +Given the absence of ground-truth images for real composite images, following [12], we opt for subjective evaluation, engaging 50 human raters in the user study. Each participant is presented with image pairs from the results generated by 5 methods, and asked to choose the image with more realistic foreground shadow. Using the Bradley-Terry model [2], we report the B-T scores in the supplementary, which again proves the advantage of our method. + +# 7. Conclusion + +In this paper, we have contributed a large-scale shadow generation dataset DESOBAv2. We have also designed a novel diffusion-based shadow generation method. Extensive experimental results show that our method is able to generate plausible shadows for composite foregrounds, significantly surpassing previous methods. + +# 8. Acknowledgement + +The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major/Key Project, China (Grant No. 2021SHZDZX0102). + +# References + +[1] Ibrahim Arief, Simon McCallum, and Jon Yngve Hardeberg. Realtime estimation of illumination direction for augmented reality on mobile devices. In CIC, 2012. 2 +[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 8 +[3] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020. 1, 2 +[4] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In ICME, 2021. +[5] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In CVPR, 2022. +[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. TIP, 2020. 2 +[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagné, and Jean-Francois Lalonde. Deep parametric indoor lighting estimation. In ICCV, 2019. 2 +[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 3 +[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, 2023. 3 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurlPS, 2020. 3 +[12] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. AAAI, 2022. 1, 2, 6, 7, 8 +[13] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, 2019. 6, 7, 8 +[14] Shaozong Huang and Lan Hong. Diffusion model for mural image inpainting. In ITOEC, 2023. 3 +[15] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM TOG, 2014. 2 +[16] Eric Kee, James F. O'Brien, and Hany Samir Farid. Exposing photo manipulation from shading and shadows. ACM TOG, 2014. 2 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6 +[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 4 + +[19] Bin Liao, Yao Zhu, Chao Liang, Fei Luo, and Chunxia Xiao. Illumination animating and editing in a single picture using scene structure estimation. Computers & Graphics, 82:53-64, 2019. 2 +[20] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In CVPR, 2018. 2 +[21] Bin Liu, Kun Xu, and Ralph R Martin. Static scene illumination estimation from videos with applications. JCST, 32 (3):430-442, 2017. 2 +[22] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In CVPR, 2020. 1, 2, 6, 7, 8 +[23] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 3 +[24] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. CoRR, abs/2303.17870, 2023. 5 +[25] Quanling Meng, Shengping Zhang, Zonglin Li, Chenyang Wang, Weigang Zhang, and Qingming Huang. Automatic shadow generation via exposure fusion. IEEE Transactions on Multimedia, 2023. 2 +[26] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, 2016. 5 +[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3 +[28] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. CoRR, abs/2106.14490, 2021. 1 +[29] Sibam Parida, Vignesh Srinivas, Bhavishya Jain, Rajesh Naik, and Neeraj Rao. Survey on diverse image inpainting using diffusion models. In PCEMS, 2023. 3 +[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NIPS, 32, 2019. 6 +[31] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In SIGGRAPH. 2003. 2 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4 +[34] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In CVPR, 2021. 1 + +[35] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In ECCV, 2022. 1, 2 +[36] Yichen Sheng, Jianming Zhang, Julien Philip, Yannick Hold-Geoffroy, Xin Sun, He Zhang, Lu Ling, and Bedrich Benes. Pixht-lab: Pixel height based light effect generation for image compositing. In CVPR, 2023. 1, 2 +[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. CoRR, abs/2010.02502, 2020. 3 +[38] Yi-Zhe Song, Zhifei Zhang, Zhe L. Lin, Scott D. Cohen, Brian L. Price, Jianming Zhang, Soo Ye Kim, and Daniel G. Aliaga. Objectstitch: Generative object compositing. In CVPR, 2023. 2 +[39] Xinhao Tao, Junyan Cao, Yan Hong, and Li Niu. Shadow generation with decomposed mask prediction and attentive shadow filling. In AAAI, 2024. 2 +[40] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In CVPR, 2017. 2 +[41] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. TPAMI, 2022. 2, 3 +[42] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In ACM MM, 2019. 2 +[43] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 5 +[44] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 1, 2 +[45] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, 2023. 3 +[46] Fangneng Zhan, Jiaxing Huang, and Shijian Lu. Adaptive composition gan towards realistic image synthesis. CoRR, abs/1905.04693, 2019. 2 +[47] Fangneng Zhan, Shijian Lu, Changgong Zhang, Feiying Ma, and Xuansong Xie. Towards realistic 3d embedding via view alignment. CoRR, abs/2007.07066, 2020. 2 +[48] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 1, 2 +[49] He Zhang, Jianming Zhang, Federico Perazzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In WACV, 2021. 2 +[50] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In CVPR, 2019. 2 +[51] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In WACV, 2020. 2 + +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 6 +[53] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5:105-115, 2019. 1, 2, 6, 7, 8 +[54] Lei Zhao, Qihang Mo, Sihuan Lin, Zhizhong Wang, Zhiwen Zuo, Haibo Chen, Wei Xing, and Dongming Lu. Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In CVPR, 2020. 6 +[55] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In CVPR, 2019. 6 +[56] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In ECCV, 2022. 3 \ No newline at end of file diff --git a/2024/Shadow Generation for Composite Image Using Diffusion Model/images.zip b/2024/Shadow Generation for Composite Image Using Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..128bba4ed4f3e8eb12118249739937d8caba4310 --- /dev/null +++ b/2024/Shadow Generation for Composite Image Using Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e33e6d55e0d5135115ef7b93394de53f81aa4320fa3edb918cf176d3608351e +size 639632 diff --git a/2024/Shadow Generation for Composite Image Using Diffusion Model/layout.json b/2024/Shadow Generation for Composite Image Using Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d52fa62a7a1aa15c62c48ced0910b7c3c01fdd41 --- /dev/null +++ b/2024/Shadow Generation for Composite Image Using Diffusion Model/layout.json @@ -0,0 +1,11541 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "text", + "content": "Shadow Generation for Composite Image Using Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 141, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 141, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 102, + 141, + 493, + 171 + ], + "type": "text", + "content": "Qingyang Liu1, Junqi You1, Jianting Wang1, Xinhao Tao1, Bo Zhang1, Li Niu1,2* \n1 Shanghai Jiao Tong University 2 miguo.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 173, + 501, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 173, + 501, + 185 + ], + "spans": [ + { + "bbox": [ + 87, + 173, + 501, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 173, + 501, + 185 + ], + "type": "text", + "content": "narumimaria, yjqsjtu2022, gloryl1299, taoxinhao, bo-zhang, ustcnewly}@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 453 + ], + "type": "text", + "content": "In the realm of image composition, generating realistic shadow for the inserted foreground remains a formidable challenge. Previous works have developed image-to-image translation models which are trained on paired training data. However, they are struggling to generate shadows with accurate shapes and intensities, hindered by data scarcity and inherent task complexity. In this paper, we resort to foundation model with rich prior knowledge of natural shadow images. Specifically, we first adapt ControlNet to our task and then propose intensity modulation modules to improve the shadow intensity. Moreover, we extend the small-scale DESOBA dataset to DESOBAv2 using a novel data acquisition pipeline. Experimental results on both DESOBA and DESOBAv2 datasets as well as real composite images demonstrate the superior capability of our model for shadow generation task. The dataset, code, and model are released at https://github.com/bcmi/Object-Shadow-Generation-Dataset-DESOBAv2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 643 + ], + "type": "text", + "content": "Image composition [28] aims to merge the foreground of one image with another background image to produce a composite image, which has a wide range of applications like virtual reality, artistic creation, and E-commerce. Simply pasting the foreground onto the background often results in visual inconsistencies, including the incompatible illumination between foreground and background [3], lack of foreground shadow/reflection [12, 34], and so on. In this paper, we focus on the shadow issue, i.e., the inserted foreground does not have plausible shadow on the background, which could significantly degrade the realism and quality of composite image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "content": "As illustrated in Figure 1, shadow generation is a challenging task because the foreground shadow is determined by many complicated factors like the lighting information and the geometry of foreground/background. The exist-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 213, + 538, + 353 + ], + "blocks": [ + { + "bbox": [ + 313, + 213, + 538, + 353 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 538, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 538, + 353 + ], + "type": "image", + "image_path": "19751b68b323e2547de1ce8434ae68dc5be14c2111fec618e4da7c3f1cabc42a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 365, + 545, + 410 + ], + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 410 + ], + "type": "text", + "content": "Figure 1. A composite image can be obtained by pasting the foreground on the background. Shadow generation aims to generate plausible shadow for the inserted foreground in the composite image to produce a more realistic image." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 590 + ], + "type": "text", + "content": "ing shadow generation methods can be divided into rendering based methods [34-36] and non-rendering based methods [12, 22, 53]. Rendering based methods usually impose strict assumptions on the geometry and lighting, which could hardly be satisfied in real-world scenarios. Besides, [35, 36] require users to specify the lighting information, which hinders its direct application in our task. Non-rendering based methods usually train an image-to-image translation network, based on pairs of composite images without foreground shadows and real images with foreground shadows. However, due to the training data scarcity and task difficulty, these methods are struggling to generate shadows with reasonable shapes and intensities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Recently, foundation model (e.g., stable diffusion [32]) pretrained on large-scale dataset has demonstrated unprecedented potential for image generation and editing. In previous works [44, 48] on object-guided inpainting or composition, they show that the generated foregrounds are accompanied by shadows even without considering the shadow issue, probably because of the rich prior knowledge of natural shadow images in foundation model. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8121" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 262 + ], + "type": "text", + "content": "We build our method upon conditional foundation model [52] and propose several key innovations. First, we modify the control encoder input and the noise loss to fit our task. Then, we observe that the generated shadow intensity (the level of darkness) is unsatisfactory. Especially when the background objects has shadows, the intensity inconsistency between foreground shadow and background shadows make the whole image unrealistic. Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, the denoising U-Net is modified to output both noise map and foreground shadow mask. The intensity encoder takes in the composite image and background shadow mask, producing the scale/bias to modulate the predicted noise within the foreground shadow region. Finally, we devise a post-processing network to rectify the color shift and background variation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 264, + 286, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 286, + 479 + ], + "type": "text", + "content": "The model training requires abundant pairs of composite images without foreground shadows and real images with foreground shadows. The existing real-world shadow generation dataset DESOBA [12] is limited by scale (i.e., 1,012 real images and 3,623 pairs) due to the high cost of manual shadow removal, which is insufficient to train our model. To ensure sufficient supervision, we design a novel data construction pipeline, which extends DESOBA to DESOBAv2 (i.e., 21,575 real images and 28,573 pairs) using object-shadow detection and inpainting techniques. Specifically, we first collect a large number of real-world images with one or more object-shadow pairs. Then, we use pretrained object-shadow detection model [41] to predict object and shadow masks for object-shadow pairs. Next, we apply pretrained inpainting model [32] to inpaint the detected shadow regions to get desshadowed images. Finally, based on real images and desshadowed images, we construct pairs of synthetic composite images and ground-truth target images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 481, + 286, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 481, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 49, + 481, + 286, + 623 + ], + "type": "text", + "content": "We conduct experiments on both DESOBAv2 and DESOBA datasets. The results reveal remarkable improvement in shadow generation task, after leveraging the benefits of large-scale data and foundation model. Our main contributions can be summarized as follows: 1) We contribute DESOBAv2, a large-scale real-world shadow generation dataset, which could greatly facilitate the shadow generation task. 2) We propose a cutting-edge diffusion model specifically designed to produce shadows for the composite foregrounds. 3) Through comprehensive experiments, we validate the efficacy of our dataset construction pipeline and the superiority of our proposed model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 638, + 132, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 638, + 132, + 649 + ], + "spans": [ + { + "bbox": [ + 49, + 638, + 132, + 649 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 658, + 160, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 160, + 670 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 160, + 670 + ], + "type": "text", + "content": "2.1. Image Composition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 678, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 678, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 678, + 286, + 712 + ], + "type": "text", + "content": "Image composition aims to overlay a foreground object on a background image to yield a composite result [20, 22, 42, 46, 47]. Previous research works have tackled differ" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 544, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 227 + ], + "type": "text", + "content": "ent issues that can compromise the quality of composite images. For instance, image blending methods [31, 42, 49, 51] target at combining the foreground and background seamlessly. Image harmonization methods [3-6, 40] aim to rectify the illumination disparity between foreground and background. Nonetheless, the above methods largely overlook the shadow cast by the foreground onto the background. Recently, generative image composition methods [38, 44, 48] can insert a foreground object into a bounding box in the background and the inserted object is likely to have shadow effect. However, they could only generate satisfactory shadows in simple cases and the object appearance could be altered unexpectedly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 236, + 420, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 236, + 420, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 236, + 420, + 247 + ], + "type": "text", + "content": "2.2. Shadow Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 255, + 544, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 544, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 544, + 422 + ], + "type": "text", + "content": "In this paper, the goal of shadow generation task is generating plausible shadow for the composite foreground. Existing methods can be broadly categorized into rendering based methods and non-rendering based methods. The rendering based methods necessitate a comprehensive understanding of factors like illumination, reflectance, material properties, and scene geometry to produce shadows for the inserted objects. However, such detailed knowledge relies on user input [15, 16, 21, 35, 36] or model prediction [1, 7, 19, 50], which is either labor-intensive or unreliable [53]. For example, [35, 36] could produce compelling results with user control. However, in the composite image, the lighting information should be inferred automatically from background instead of requested by users." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 423, + 544, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 423, + 544, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 423, + 544, + 613 + ], + "type": "text", + "content": "Non-rendering based methods [12, 22, 25, 53] aim to translate an input composite image without foreground shadow to an output with foreground shadow, bypassing the need for explicit knowledge of the aforementioned factors. For instance, ShadowGAN [53] utilizes both global and local conditional discriminator to enhance the realism of generated shadows. ARShadowGAN [22] emphasizes the importance of background shadow and uses it to guide foreground shadow generation. SGRNet [12] encourages the information exchange between foreground and background, and employs a classic illumination model for better shadow effect. The work [25] produces multiple underexposure images and fuses them to get the final shadow region. DMASNet [39] decomposes shadow mask prediction into box prediction and shape prediction, achieving better cross-domain transferability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 614, + 544, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 544, + 638 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 544, + 638 + ], + "type": "text", + "content": "To the best of our knowledge, we are the first diffusion-based method focusing on shadow generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 647, + 407, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 407, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 407, + 658 + ], + "type": "text", + "content": "2.3. Diffusion Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 666, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 666, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 666, + 544, + 712 + ], + "type": "text", + "content": "In recent years, diffusion models have emerged as a powerful tool in image generation and image editing. These models approach image generation as a series of stochastic transitions, moving from a basic distribution to the desired" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "8122" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 549, + 229 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 549, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 549, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 549, + 229 + ], + "type": "image", + "image_path": "f81995dc9c2606ff9b9ede849f50101ef2deb2f66d4b156033fb36b684386cb5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": "Figure 2. The pipeline of dataset construction. We use object-shadow detection model [41] to predict pairs of object and shadow masks in the real image " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": ". Then we obtain the union " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "M_s" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": " of all shadow masks as the inpainting mask and apply inpainting model [32] to get a desshadowed image " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": ". After designating a foreground object, we replace the background shadow regions " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "M_{bs}" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": " with the counterparts in " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": " to synthesize a composite image " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_c" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": ", and replace all the shadow regions " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "M_s" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": " with the counterparts in " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": " to obtain the ground-truth target image " + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "inline_equation", + "content": "I_g" + }, + { + "bbox": [ + 46, + 236, + 547, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 479 + ], + "type": "text", + "content": "data distribution [11]. Diffusion models can be divided into unconditional diffusion models [11, 37] and conditional diffusion models [27, 32, 52]. Unconditional diffusion models focus on generating realistic images by capturing the distribution of natural images, without the need of any specific input conditions. Conditional diffusion models are designed to produce images under the guidance of specific conditional inputs, such as text descriptions, semantic masks, and so on. ControlNet [52] is a popular conditional diffusion model, which equips large pretrained text-to-image diffusion models with spatial-aware and task-specific conditions. We build our model upon ControlNet and propose several innovations to meet the specific requirements of shadow generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 492, + 173, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 173, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 173, + 504 + ], + "type": "text", + "content": "3. Dataset Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 513, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 536 + ], + "type": "text", + "content": "The pipeline of our dataset construction is illustrated in Figure 2, which will be detailed next." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 546, + 189, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 189, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 189, + 559 + ], + "type": "text", + "content": "3.1. Shadow Image Collection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 565, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 661 + ], + "type": "text", + "content": "We harvest an extensive collection of real-world outdoor images with natural lighting across various scenes from two sources. On one hand, we crawl online images from public websites that have licenses for reuse. On the other hand, we hire photographers to capture photos in the outdoor scenes that satisfy our requirements. We only preserve the images with at least one object-shadow pair, arriving at 44,044 images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 151, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 151, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 151, + 682 + ], + "type": "text", + "content": "3.2. Shadow Removal" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Given a real image " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " with object-shadow pairs, we use the pretrained object-shadow detection model [41] to predict " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "text", + "content": "pairs of object and shadow masks. We use " + }, + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "inline_equation", + "content": "M_{o,k}" + }, + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "text", + "content": " (resp., " + }, + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "inline_equation", + "content": "M_{s,k}" + }, + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "text", + "content": ") to denote the object (resp., shadow) mask of the " + }, + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 312, + 545, + 371 + ], + "type": "text", + "content": "-th object. We refer to one detected object-shadow pair as one detected instance. We eliminate the images without any detected instance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 374, + 546, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 517 + ], + "type": "text", + "content": "Subsequently, we attempt to erase all the detected shadows. We have tried some state-of-the-art shadow removal models [8, 9], but the performance in the wild is below our expectation due to poor generalization ability. Considering the recent rapid advance of image inpainting [14, 23, 29, 32, 45, 56] techniques, we resort to image inpainting to remove the shadows. Although image inpainting cannot preserve the background information precisely, we observe that the background textures in the shadow region are usually very simple, and the inpainted result has similar textures with the original background. Thus, we roughly treat the inpainted results as desshadowed results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "content": "We obtain the union of all detected shadow masks " + }, + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "inline_equation", + "content": "M_{s} = M_{s,1} \\cup M_{s,2} \\cup \\dots \\cup M_{s,K}" + }, + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "content": " as the inpainting mask and apply the pretrained inpainting model [32] to get a deshadowed image " + }, + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "content": ". In practice, we observe that the inpainting model is prone to generate low-quality shadow in the inpainted region in some cases. To prevent the inpainting model from generating undesirable shadows in the inpainted region, we adopt some tricks like dilating the inpainting mask and flipping images vertically, which can effectively obstruct undesirable shadow generation during inpainting. However, there may still exist undesirable shadows or noticeable artifacts in the inpainted region." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": "After inpainting, we manually filter the object-shadow pairs according to the following rules: 1) We remove the object-shadow pairs with low-quality object masks or shadow masks. 2) We remove those object-shadow pairs" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8123" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "with generated shadows or noticeable artifacts in the inpainted region. After manual filtering, we refer to the remaining object-shadow pairs as valid instances. We have 21,575 images with 28,573 valid instances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 130, + 198, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 130, + 198, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 198, + 144 + ], + "type": "text", + "content": "3.3. Composite Image Synthesis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": "Given a pair of a real image " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " and a deshadowed image " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": ", we randomly select the " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": "-th foreground object from valid instances and synthesize the composite image. " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "M_{o,k}" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " (resp., " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "M_{s,k}" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": ") is referred to as the foreground object (resp., shadow) mask " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "M_{f_o}" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " (resp., " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "M_{f_s}" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": "). One strategy is replacing the shadow region " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "M_{fs}" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " of this foreground object in " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " with the counterpart in " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " to erase the foreground shadow. However, this strategy may leave traces along the shadow boundary, in which case the model may find a shortcut to generate the shadow. Another strategy is replacing the shadow regions " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "M_{bs} = M_{s,1} \\cup \\dots \\cup M_{s,k-1} \\cup M_{s,k+1} \\cup \\dots \\cup M_{s,K}" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " of the other objects in " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "I_d" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " with the counterparts in " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": " to synthesize a composite image " + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "inline_equation", + "content": "I_c" + }, + { + "bbox": [ + 46, + 149, + 287, + 329 + ], + "type": "text", + "content": ", in which only the selected foreground object does not have shadow while all the other objects have shadows. We adopt the second strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": "After inpainting, the background may undergo slight changes, so the background of " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": " may be slightly different from that of " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": ". To ensure consistent background, we obtain the ground-truth target image " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": " by replacing the shadow regions " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": " of all objects in " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": " with the counterparts in " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": " form a pair of input composite image and ground-truth target image. So far, we obtain tuples in the form of " + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\{I_{c}, M_{fo}, M_{fs}, M_{bs}, I_{g}\\}" + }, + { + "bbox": [ + 46, + 330, + 288, + 450 + ], + "type": "text", + "content": ", which will be used for model training. Example images and more statistics of our dataset can be found in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 463, + 126, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 126, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 126, + 477 + ], + "type": "text", + "content": "4. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "content": "Stable Diffusion [32] is latent diffusion model operating in a latent space. First, " + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "content": " images are converted to " + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "content": " latent images using VAE [18] with encoder " + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "inline_equation", + "content": "E_r" + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "content": ". The image space is projected to the latent space using " + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "inline_equation", + "content": "E_r" + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "content": ", and back to the image space using " + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 46, + 483, + 287, + 592 + ], + "type": "text", + "content": ". Then, the forward diffusion process and backward denoising process are performed in the latent space. The denoising U-Net [33] consists of an encoder with 12 blocks, a middle block, and a skip-connected decoder with 12 blocks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": "During training, random Gaussian noise " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": " is added to the latent image " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": " in the denoising step " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": ", producing a noisy latent image " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": ". Given time step " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": " and text prompt " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "c_{txt}" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": ", the denoising U-Net with model parameters " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": " is trained to predict the added noise " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "To support spatial conditional information (e.g., edge, pose, depth), ControlNet [52] integrates a control encoder " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " with pre-trained Stable Diffusion. Specifically, the control encoder contains trainable replicas of its 12 encoding blocks and middle block across four resolutions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "(64 \\times 64, 32 \\times 32, 16 \\times 16, 8 \\times 8)" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ". It takes a " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " conditional image as input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 96, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 178 + ], + "type": "text", + "content": "The conditional feature maps " + }, + { + "bbox": [ + 304, + 96, + 545, + 178 + ], + "type": "inline_equation", + "content": "c_{img}" + }, + { + "bbox": [ + 304, + 96, + 545, + 178 + ], + "type": "text", + "content": " output from control encoder are used to enhance the 12 skip-connections and middle block in denoising U-Net via zero convolution layers. While the original Stable Diffusion is fixed to retain prior knowledge, control encoder could incorporate additional conditions to guide image generation. The objective could be rewritten as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 182, + 545, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 182, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 545, + 203 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c t r l} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {t x t}, \\boldsymbol {c} _ {i m g}\\right) \\| _ {2} ^ {2} \\right]. \\tag {1}", + "image_path": "ce3eff3acfad5c6e4fd54e9e7b8847e789605cd532b46e548ba420f63b9ffa33.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 207, + 361, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 361, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 361, + 219 + ], + "type": "text", + "content": "5. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "text", + "content": "Given a composite image " + }, + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "text", + "content": " without foreground shadow as well as the foreground object mask " + }, + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "inline_equation", + "content": "M_{fo}" + }, + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "text", + "content": ", our Shadow Generation Diffusion (SGDiffusion) model aims to produce " + }, + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{g}" + }, + { + "bbox": [ + 304, + 227, + 545, + 324 + ], + "type": "text", + "content": " with plausible foreground shadow. We will adapt ControlNet [52] to shadow generation task in Section 5.1, and propose novel modules to improve the shadow intensity in Section 5.2. Finally, we will briefly introduce post-processing techniques to enhance the image quality in Section 5.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 331, + 535, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 535, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 535, + 342 + ], + "type": "text", + "content": "5.1. Adapting ControlNet to Shadow Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "content": "For shadow generation task, the useful conditional information is input composite image " + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "content": " and foreground object mask " + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "inline_equation", + "content": "M_{fo}" + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "content": ", in which the foreground object mask indicates the target object we need to generate shadow for. We concatenate " + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "inline_equation", + "content": "M_{fo}" + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "content": " as the input of control encoder " + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "content": ". The control encoder outputs the conditional feature maps " + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "inline_equation", + "content": "c_{sg}" + }, + { + "bbox": [ + 304, + 349, + 545, + 480 + ], + "type": "text", + "content": ", which are injected into the denoising decoder to provide guidance. For the text prompt, we have tried several variants like \"the [object category] with shadow\", but they have no significant impact on the generated shadows. Therefore, we use null text prompt by default." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "text", + "content": "Given a set of conditions including time step " + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "text", + "content": " and conditional feature maps " + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "inline_equation", + "content": "c_{sg}" + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "text", + "content": ", the denoising U-Net with model parameters " + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "text", + "content": " predicts the noise " + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "text", + "content": " added to the noisy latent image " + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 480, + 545, + 528 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 533, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 533, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 336, + 533, + 545, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {s g}\\right)) \\| _ {2} ^ {2} \\right]. \\tag {2}", + "image_path": "c285690389f66cccd9817f8af6ba71bfbaf49dbd53d0a231addcb4a6d0015425.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "text", + "content": "To enforce the model to place more emphasis on the foreground shadow region, we introduce weighted noise loss, which assigns higher weights to the foreground shadow region. We expand the foreground shadow mask by a dilated kernel to get the expanded mask " + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\hat{M}_{fs}" + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "text", + "content": ". The weights in the expanded foreground shadow region are " + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "text", + "content": " while the other weights are 1, leading to the weight map " + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "inline_equation", + "content": "W_{fs}" + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "text", + "content": ". If we do not expand the foreground shadow region, the model will be misled to generate large shadows, overlooking the details of shadow shapes and boundaries. By applying weight map " + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "inline_equation", + "content": "W_{fs}" + }, + { + "bbox": [ + 304, + 559, + 545, + 690 + ], + "type": "text", + "content": " to the noise loss, we can arrive at" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 696, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 696, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 310, + 696, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\boldsymbol {W} _ {f s} \\circ (\\epsilon - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c} _ {s g})) \\| _ {2} ^ {2} \\right], \\tag {3}", + "image_path": "d536a8484e9ae829ce829f4a0b18efaa757f7093730851ef24ca2ace3b61c162.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8124" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 509, + 312 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 509, + 312 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 509, + 312 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 509, + 312 + ], + "type": "image", + "image_path": "976a603a173f618365fdf99dfad9378cc01405353e58f5ff3b4c84cc834122c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": "Figure 3. The framework of our SGDiffusion. We adapt ControlNet (Control Encoder and Denoising U-Net) to shadow generation task. We also introduce an intensity encoder to modulate the foreground shadow region in the noise map " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": ", leading to " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": ". The output noise " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": " is supervised by weighted noise loss " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mwsg}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": " based on the expanded foreground shadow mask " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{fs}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 373, + 231, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 231, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 231, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 373, + 231, + 384 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 47, + 373, + 231, + 384 + ], + "type": "text", + "content": " denotes element-wise multiplication." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "content": "During inference, to retain more information of input composite image " + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "content": " in the initial noise, we obtain " + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "content": " by adding noise to the latent image of " + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "content": ", rather than directly sampling from the Gaussian distribution " + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 442, + 209, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 209, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 209, + 455 + ], + "type": "text", + "content": "5.2. Shadow Intensity Modulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 544 + ], + "type": "text", + "content": "By using the adapted ControlNet in Section 5.1, we observe that the intensity of generated foreground shadow is unsatisfactory. Especially when the background has object-shadow pairs, the generated foreground shadow is often notably darker or brighter than background shadows. Such inconsistency between foreground shadow intensity and background shadow intensity makes the whole image unrealistic." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": "Therefore, we introduce another intensity encoder to modulate the foreground shadow intensity. Specifically, we use encoder " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " to extract intensity-relevant information. Intuitively, by observing background shadows and its surrounding unshadowed areas, we can estimate the intensity of foreground shadows. Thus, the input of intensity encoder " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " should include the composite image " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " and background shadow mask " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "M_{bs}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": ". When there is no background shadow, the mask is all black. We concatenate " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " with background shadow mask " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "M_{bs}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " as the input of intensity encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The intensity encoder outputs scales and biases to adjust the intensity of noise map within the foreground shadow region. The modulated noise map results in the modulated latent image, and further results in the modulated foreground" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": "shadow. Therefore, the intensity adjustment of noise map is finally embodied in the intensity variation of generated foreground shadow. Specifically, when the noise map has " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " channels, " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " outputs the " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": "-dim scale vector " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": "-dim bias vector " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{b}" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": ", containing channel-wise scales and biases. " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{b}" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " are used to modulate the predicted noise map within the foreground shadow region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "text", + "content": "One problem is that the foreground shadow region is unknown in the testing stage, so we need to predict the foreground shadow mask. To avoid much extra computational cost, we take advantage of the feature maps in the denoising U-Net to predict the foreground shadow mask. Previous works usually combine different layers of feature maps in denoising U-Net for mask prediction [24, 43]. We try different layers of feature maps and find that decoder feature maps are more effective in shadow mask prediction. We also use foreground object mask, which could provide useful hints for the location of foreground shadow. We resize all decoder feature maps and foreground object mask to the same size, and concatenate them channel-wisely. The concatenation passes through several convolutional layers to predict the foreground shadow mask " + }, + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{fs}" + }, + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{fs}" + }, + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "text", + "content": " is supervised with ground-truth foreground shadow mask " + }, + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "inline_equation", + "content": "M_{fs}" + }, + { + "bbox": [ + 304, + 457, + 546, + 661 + ], + "type": "text", + "content": " by Binary Cross-Entropy (BCE) loss and Dice loss [26]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 668, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 668, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 318, + 668, + 545, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m a s k} = \\mathcal {L} _ {b c e} (\\tilde {M} _ {f s}, M _ {f s}) + \\mathcal {L} _ {d i c e} (\\tilde {M} _ {f s}, M _ {f s}). (4)", + "image_path": "7c66ecbf3b5738359bc1da1aa6b27cd871a78883a43b24452964428970c21b55.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " is large, " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " is close to random noise and thus the decoder feature maps are not informative to predict shadow" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8125" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "mask. Hence, we only employ the loss " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " when the time step " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " is small. We set the threshold of " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\sigma T" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " is the total number of steps. Accordingly, shadow intensity modulation is only applied when " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " is smaller than the threshold " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\sigma T" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": "Provided with the predicted foreground shadow mask " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{fs}" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": ", we can modulate the noise map within the foreground shadow region. Given the predicted noise map " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon} = \\epsilon_{\\theta}(z_t, t, c_{sg})" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": ", we multiply " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": " by channel-wise scales " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": " and add channel-wise biases " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "\\pmb{b}" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}'" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": ". Then, based on " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{fs}" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": ", we combine the modulated noise map and original noise map to get the final noise map: " + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = \\tilde{\\epsilon}' \\circ \\tilde{M}_{fs} + \\tilde{\\epsilon} \\circ (1 - \\tilde{M}_{fs})" + }, + { + "bbox": [ + 47, + 132, + 286, + 217 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "content": "We replace the predicted noise map in Eqn. (3) with the final noise map " + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "content": " and get" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 248, + 286, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 248, + 286, + 268 + ], + "spans": [ + { + "bbox": [ + 77, + 248, + 286, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m w s g} = \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\boldsymbol {W} _ {f s} \\circ (\\epsilon - \\hat {\\epsilon}) \\| _ {2} ^ {2} \\right]. \\tag {5}", + "image_path": "4f692bcfc1740f9664fa94f927a05fc1868c6f91a78f0cf6ecdd2fee14671304.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 275, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 286, + 299 + ], + "type": "text", + "content": "We summarize the mask prediction loss in Eqn. (4) and weighted noise loss in Eqn. (5) as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 308, + 286, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 308, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 112, + 308, + 286, + 320 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda \\mathcal {L} _ {\\text {m w s g}}, \\tag {6}", + "image_path": "60b3db09a041a1e274e4ba506c132b855e619e8858897e21901cd2d22dbeeb77.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 329, + 180, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 180, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 180, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 329, + 180, + 340 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 329, + 180, + 340 + ], + "type": "text", + "content": " is a trade-off parameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 347, + 143, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 143, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 143, + 360 + ], + "type": "text", + "content": "5.3. Post-processing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 366, + 286, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 286, + 485 + ], + "type": "text", + "content": "We observe that the generated images could have color shift and background variation issues. Color shift means that the overall color tone deviates from the input composite image. Background variation means that some background details are changed. To solve these issues, we create a multi-task post-processing network which yields the rectified image together with the foreground shadow mask. Then, we combine input composite image and rectified image based on the predicted foreground shadow mask to produce the final image. The technical details are left to supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 496, + 127, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 127, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 127, + 510 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 515, + 222, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 222, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 222, + 528 + ], + "type": "text", + "content": "6.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 533, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 664 + ], + "type": "text", + "content": "We conduct experiments on both DESOBA [12] and our contributed DESOBAv2 dataset. We split DESOBAv2 into 21,088 training images with 27,718 tuples and 487 test images with 855 tuples. Following [12], the test set contains BOS images (with background object-shadow pairs) and BOS-free images. Most of our experiments are based on DESOBAv2 dataset due to the following two concerns: 1) DESOBAv2 has larger test set which supports more comprehensive evaluation. 2) DESOBA has the artifacts caused by manual shadow removal and the existing methods (e.g., SGRNet) tend to overfit such artifacts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "For the generated results, we evaluate both image quality and mask quality. For image evaluation, following [12], we adopt RMSE and SSIM, which are calculated based on the ground-truth target image and the generated image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "type": "text", + "content": "Global RMSE (GR) and Global SSIM (GS) are calculated over the whole image, while Local RMSE (LR) and Local SSIM (LS) are calculated over the ground-truth foreground shadow region. For the mask evaluation, following [12], we adopt Balanced Error Rate (BER), which is calculated based on the ground-truth binary foreground shadow mask and the predicted foreground shadow mask obtained by threshold 0.5. Global BER (GB) is calculated over the whole image, while Local BER (LB) is calculated over the ground-truth foreground shadow region. Note that diffusion model has stochastic property and shadow generation is a multi-modal task, that is, one input has multiple plausible outputs. Similar to multi-modal inpainting evaluation [54, 55], we generate 5 results for one test image with different random seeds and select the one closest to the ground-truth (the highest Local SSIM) to calculate evaluation metrics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 284, + 438, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 284, + 438, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 438, + 296 + ], + "type": "text", + "content": "6.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "text", + "content": "We develop our method with PyTorch 1.12.1 [30]. Our model is trained using the Adam optimizer [17] with a constant learning rate of " + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "text", + "content": " over 50 epochs, on four NVIDIA RTX A6000 GPUs. Our method is built upon ControlNet [52]. We employ ResNet18 [10] as the intensity encoder. The mask predictor passes the concatenation of decoder feature maps and foreground object mask through four convolutional layers, with ReLU activation following the first three layers and Sigmoid activation following the last layer. We set the hyper-parameters " + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 307, + 545, + 437 + ], + "type": "text", + "content": " as 10, 0.7, and 1, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 459, + 455, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 455, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 455, + 471 + ], + "type": "text", + "content": "6.3. Comparison with Baselines" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 481, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 601 + ], + "type": "text", + "content": "Following [12], we compare with ShadowGAN [53], Mask-ShadowGAN [13], ARShadowGAN [22], and SGRNet [12]. We train and test all methods on DESOBAv2 dataset. The quantitative results are summarized in Table 1. We observe that our SGDiffusion achieves the lowest GRMSE, LRMSE and the highest GSSIM, LSSIM, which demonstrates that our method could generate shadow images that are closer to the ground-truth shadow images. The best GB and LB results demonstrate that the shapes and locations of our generated shadows are more accurate." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "For qualitative comparison, we show several example results in Figure 4. Compared with the baseline methods, the shadows produced by our model have more reasonable shapes and intensities. Moreover, as shown in row 1, our method can take into account the self-occlusion of objects to generate discontinuous shadows. As shown in row 4, our method can also consider the material of the objects, producing shadows with translucency effects. We provide more examples in the supplementary." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8126" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 81, + 119, + 140 + ], + "blocks": [ + { + "bbox": [ + 60, + 81, + 119, + 140 + ], + "lines": [ + { + "bbox": [ + 60, + 81, + 119, + 140 + ], + "spans": [ + { + "bbox": [ + 60, + 81, + 119, + 140 + ], + "type": "image", + "image_path": "e17d3a2538d52dc5db075b9d47d110dcbe3cfad68d4be178861b9ac64deddb78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 81, + 178, + 140 + ], + "blocks": [ + { + "bbox": [ + 119, + 81, + 178, + 140 + ], + "lines": [ + { + "bbox": [ + 119, + 81, + 178, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 81, + 178, + 140 + ], + "type": "image", + "image_path": "b5ad355ce11dd36cc7fb3a30f930573ad3b90c86eeeb4ae1d7fabfd4167b9bea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 178, + 81, + 236, + 140 + ], + "blocks": [ + { + "bbox": [ + 178, + 81, + 236, + 140 + ], + "lines": [ + { + "bbox": [ + 178, + 81, + 236, + 140 + ], + "spans": [ + { + "bbox": [ + 178, + 81, + 236, + 140 + ], + "type": "image", + "image_path": "a03e7b338c2b6708488c996507271c8dec5af53968fac8ddd7f43251133770dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 237, + 81, + 295, + 140 + ], + "blocks": [ + { + "bbox": [ + 237, + 81, + 295, + 140 + ], + "lines": [ + { + "bbox": [ + 237, + 81, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 237, + 81, + 295, + 140 + ], + "type": "image", + "image_path": "71b88ff84bb7f4d84a21af979a9f912b07f629e3943fe97c04aba36550886f20.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 296, + 81, + 355, + 140 + ], + "blocks": [ + { + "bbox": [ + 296, + 81, + 355, + 140 + ], + "lines": [ + { + "bbox": [ + 296, + 81, + 355, + 140 + ], + "spans": [ + { + "bbox": [ + 296, + 81, + 355, + 140 + ], + "type": "image", + "image_path": "c2405d23ecb5e5a9c7c3af460adb50f176a321633030f34976a6897dfd5b737a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 356, + 81, + 414, + 140 + ], + "blocks": [ + { + "bbox": [ + 356, + 81, + 414, + 140 + ], + "lines": [ + { + "bbox": [ + 356, + 81, + 414, + 140 + ], + "spans": [ + { + "bbox": [ + 356, + 81, + 414, + 140 + ], + "type": "image", + "image_path": "1c38209021a2dce44d0f3ee3df1323102680dbf5bf5e7f104b391cd81c3eba8d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 416, + 81, + 474, + 140 + ], + "blocks": [ + { + "bbox": [ + 416, + 81, + 474, + 140 + ], + "lines": [ + { + "bbox": [ + 416, + 81, + 474, + 140 + ], + "spans": [ + { + "bbox": [ + 416, + 81, + 474, + 140 + ], + "type": "image", + "image_path": "faf3c83ffeeae0b959067e0a7226af1d1f044d0bb5ece80c7cd7cf97f15114ba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 475, + 81, + 534, + 140 + ], + "blocks": [ + { + "bbox": [ + 475, + 81, + 534, + 140 + ], + "lines": [ + { + "bbox": [ + 475, + 81, + 534, + 140 + ], + "spans": [ + { + "bbox": [ + 475, + 81, + 534, + 140 + ], + "type": "image", + "image_path": "27e99ab1b67aaf69575769ea323efa6646874b56a93bb6f6d26835030c554abc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 60, + 141, + 118, + 200 + ], + "blocks": [ + { + "bbox": [ + 60, + 141, + 118, + 200 + ], + "lines": [ + { + "bbox": [ + 60, + 141, + 118, + 200 + ], + "spans": [ + { + "bbox": [ + 60, + 141, + 118, + 200 + ], + "type": "image", + "image_path": "7b1636b5f955b36109bb158be39f9a1ffc65d1c82a4f7bce6a74d5d15f74132d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 119, + 141, + 178, + 200 + ], + "blocks": [ + { + "bbox": [ + 119, + 141, + 178, + 200 + ], + "lines": [ + { + "bbox": [ + 119, + 141, + 178, + 200 + ], + "spans": [ + { + "bbox": [ + 119, + 141, + 178, + 200 + ], + "type": "image", + "image_path": "7c636852e0334ef3b043660ab486ed75e753c9272f7e24ce8d410bb103816197.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 178, + 141, + 236, + 200 + ], + "blocks": [ + { + "bbox": [ + 178, + 141, + 236, + 200 + ], + "lines": [ + { + "bbox": [ + 178, + 141, + 236, + 200 + ], + "spans": [ + { + "bbox": [ + 178, + 141, + 236, + 200 + ], + "type": "image", + "image_path": "bb021e3f20ca4010d8e869de53546cc40161ddac343853185e6faf779aa2d6f4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 237, + 141, + 296, + 200 + ], + "blocks": [ + { + "bbox": [ + 237, + 141, + 296, + 200 + ], + "lines": [ + { + "bbox": [ + 237, + 141, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 237, + 141, + 296, + 200 + ], + "type": "image", + "image_path": "9e5792c07c83fb3d6f2d46ebb869967d7d3ce6b18de76e68324523a50da456e2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 296, + 141, + 355, + 200 + ], + "blocks": [ + { + "bbox": [ + 296, + 141, + 355, + 200 + ], + "lines": [ + { + "bbox": [ + 296, + 141, + 355, + 200 + ], + "spans": [ + { + "bbox": [ + 296, + 141, + 355, + 200 + ], + "type": "image", + "image_path": "c671801a469a75b781f4293a0f45d6ca0453f65ee0d5c43c9998e5392f49a78f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 356, + 141, + 414, + 200 + ], + "blocks": [ + { + "bbox": [ + 356, + 141, + 414, + 200 + ], + "lines": [ + { + "bbox": [ + 356, + 141, + 414, + 200 + ], + "spans": [ + { + "bbox": [ + 356, + 141, + 414, + 200 + ], + "type": "image", + "image_path": "59c13068a9a375b067bc7f093017cc641880acb88d08b6a07a964145553cf71d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 416, + 141, + 474, + 200 + ], + "blocks": [ + { + "bbox": [ + 416, + 141, + 474, + 200 + ], + "lines": [ + { + "bbox": [ + 416, + 141, + 474, + 200 + ], + "spans": [ + { + "bbox": [ + 416, + 141, + 474, + 200 + ], + "type": "image", + "image_path": "84e69e7792d30621b7939e09ff5dd20dafb38f8e95acd984d71b26f6791e34b3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 475, + 141, + 534, + 200 + ], + "blocks": [ + { + "bbox": [ + 475, + 141, + 534, + 200 + ], + "lines": [ + { + "bbox": [ + 475, + 141, + 534, + 200 + ], + "spans": [ + { + "bbox": [ + 475, + 141, + 534, + 200 + ], + "type": "image", + "image_path": "c4bc469259e2b7d520bceab4e9275a6925136ab0b7df4c1ad862907753123c1c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 60, + 201, + 118, + 259 + ], + "blocks": [ + { + "bbox": [ + 60, + 201, + 118, + 259 + ], + "lines": [ + { + "bbox": [ + 60, + 201, + 118, + 259 + ], + "spans": [ + { + "bbox": [ + 60, + 201, + 118, + 259 + ], + "type": "image", + "image_path": "d037949ea397c53342ee0e20ceecdf202b984cf9e6b3d00ee3ecb4c7968c66de.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 119, + 201, + 178, + 259 + ], + "blocks": [ + { + "bbox": [ + 119, + 201, + 178, + 259 + ], + "lines": [ + { + "bbox": [ + 119, + 201, + 178, + 259 + ], + "spans": [ + { + "bbox": [ + 119, + 201, + 178, + 259 + ], + "type": "image", + "image_path": "068abca7ecd3657ba5f48bc24051bf516283bbb722d0affec1255aaa27801bc4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 178, + 201, + 236, + 259 + ], + "blocks": [ + { + "bbox": [ + 178, + 201, + 236, + 259 + ], + "lines": [ + { + "bbox": [ + 178, + 201, + 236, + 259 + ], + "spans": [ + { + "bbox": [ + 178, + 201, + 236, + 259 + ], + "type": "image", + "image_path": "dd006c10810b049c00cfa8944b5366f1139dd728ef02f576fee6e186047e5323.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 237, + 201, + 296, + 259 + ], + "blocks": [ + { + "bbox": [ + 237, + 201, + 296, + 259 + ], + "lines": [ + { + "bbox": [ + 237, + 201, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 237, + 201, + 296, + 259 + ], + "type": "image", + "image_path": "db4b852d6a9b4a7ebca692b96ef98b208edfdce63ae521a0915dcce8a703439b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 296, + 201, + 355, + 259 + ], + "blocks": [ + { + "bbox": [ + 296, + 201, + 355, + 259 + ], + "lines": [ + { + "bbox": [ + 296, + 201, + 355, + 259 + ], + "spans": [ + { + "bbox": [ + 296, + 201, + 355, + 259 + ], + "type": "image", + "image_path": "ee7fe548c9efd7438a343f373abb42be59134523a24368486f3b233007eb0876.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 356, + 201, + 414, + 259 + ], + "blocks": [ + { + "bbox": [ + 356, + 201, + 414, + 259 + ], + "lines": [ + { + "bbox": [ + 356, + 201, + 414, + 259 + ], + "spans": [ + { + "bbox": [ + 356, + 201, + 414, + 259 + ], + "type": "image", + "image_path": "84de971930cabf625e5a41ef0540156f3242f5e745b3194cb57e6f958180900c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 416, + 201, + 474, + 259 + ], + "blocks": [ + { + "bbox": [ + 416, + 201, + 474, + 259 + ], + "lines": [ + { + "bbox": [ + 416, + 201, + 474, + 259 + ], + "spans": [ + { + "bbox": [ + 416, + 201, + 474, + 259 + ], + "type": "image", + "image_path": "18b948ba09a6134f76f2054bde2f0677b154024f87bdbe72d14d4cb906bd081f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 475, + 201, + 534, + 259 + ], + "blocks": [ + { + "bbox": [ + 475, + 201, + 534, + 259 + ], + "lines": [ + { + "bbox": [ + 475, + 201, + 534, + 259 + ], + "spans": [ + { + "bbox": [ + 475, + 201, + 534, + 259 + ], + "type": "image", + "image_path": "4a15da3c17616dce5a306c69a508a168bd766ec5a5caacd3a6d72abbb960fb71.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 60, + 261, + 118, + 319 + ], + "blocks": [ + { + "bbox": [ + 60, + 261, + 118, + 319 + ], + "lines": [ + { + "bbox": [ + 60, + 261, + 118, + 319 + ], + "spans": [ + { + "bbox": [ + 60, + 261, + 118, + 319 + ], + "type": "image", + "image_path": "4be851cd4d740da59790ce93174b173f10546874b87bed2f7468e8c268ee7507.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 326, + 96, + 338 + ], + "lines": [ + { + "bbox": [ + 83, + 326, + 96, + 338 + ], + "spans": [ + { + "bbox": [ + 83, + 326, + 96, + 338 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 119, + 261, + 178, + 319 + ], + "blocks": [ + { + "bbox": [ + 119, + 261, + 178, + 319 + ], + "lines": [ + { + "bbox": [ + 119, + 261, + 178, + 319 + ], + "spans": [ + { + "bbox": [ + 119, + 261, + 178, + 319 + ], + "type": "image", + "image_path": "da73d8df4e2f3de3251f85108a2828f7a2f8628a3fa9fd755e64553088a72d1a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 326, + 156, + 338 + ], + "lines": [ + { + "bbox": [ + 141, + 326, + 156, + 338 + ], + "spans": [ + { + "bbox": [ + 141, + 326, + 156, + 338 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 178, + 261, + 236, + 319 + ], + "blocks": [ + { + "bbox": [ + 178, + 261, + 236, + 319 + ], + "lines": [ + { + "bbox": [ + 178, + 261, + 236, + 319 + ], + "spans": [ + { + "bbox": [ + 178, + 261, + 236, + 319 + ], + "type": "image", + "image_path": "6623fdb43e19ee7c114059eb9863788d07b984d544b084fb073d60e84cd5b49e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 326, + 214, + 338 + ], + "lines": [ + { + "bbox": [ + 201, + 326, + 214, + 338 + ], + "spans": [ + { + "bbox": [ + 201, + 326, + 214, + 338 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 237, + 261, + 296, + 319 + ], + "blocks": [ + { + "bbox": [ + 237, + 261, + 296, + 319 + ], + "lines": [ + { + "bbox": [ + 237, + 261, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 237, + 261, + 296, + 319 + ], + "type": "image", + "image_path": "f406e079ac0961b77c2bd8c55fee98e7595757c6c16672830210232023c21121.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 326, + 274, + 338 + ], + "lines": [ + { + "bbox": [ + 260, + 326, + 274, + 338 + ], + "spans": [ + { + "bbox": [ + 260, + 326, + 274, + 338 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 296, + 261, + 355, + 319 + ], + "blocks": [ + { + "bbox": [ + 296, + 261, + 355, + 319 + ], + "lines": [ + { + "bbox": [ + 296, + 261, + 355, + 319 + ], + "spans": [ + { + "bbox": [ + 296, + 261, + 355, + 319 + ], + "type": "image", + "image_path": "4dc0a8fb9212b6b99741dfc57b44987f9458198e8672a6f9551a9702e532d8f8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 326, + 333, + 338 + ], + "lines": [ + { + "bbox": [ + 320, + 326, + 333, + 338 + ], + "spans": [ + { + "bbox": [ + 320, + 326, + 333, + 338 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 359, + 547, + 393 + ], + "lines": [ + { + "bbox": [ + 46, + 359, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 547, + 393 + ], + "type": "text", + "content": "Figure 4. Visual comparison of different methods on DESOBAv2 dataset. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), our SGDiffusion (g), ground-truth (h)." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 356, + 261, + 414, + 319 + ], + "blocks": [ + { + "bbox": [ + 356, + 261, + 414, + 319 + ], + "lines": [ + { + "bbox": [ + 356, + 261, + 414, + 319 + ], + "spans": [ + { + "bbox": [ + 356, + 261, + 414, + 319 + ], + "type": "image", + "image_path": "61ed1e035026ea38d9d2b601550403d1aa4151b044378f1495e7509096b06adc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 326, + 392, + 338 + ], + "lines": [ + { + "bbox": [ + 380, + 326, + 392, + 338 + ], + "spans": [ + { + "bbox": [ + 380, + 326, + 392, + 338 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 416, + 261, + 474, + 319 + ], + "blocks": [ + { + "bbox": [ + 416, + 261, + 474, + 319 + ], + "lines": [ + { + "bbox": [ + 416, + 261, + 474, + 319 + ], + "spans": [ + { + "bbox": [ + 416, + 261, + 474, + 319 + ], + "type": "image", + "image_path": "09d862ad07d39b42adfc6452f15f3f36be05c66ee2dc7226085bf954e41a71aa.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 326, + 452, + 338 + ], + "lines": [ + { + "bbox": [ + 438, + 326, + 452, + 338 + ], + "spans": [ + { + "bbox": [ + 438, + 326, + 452, + 338 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 475, + 261, + 534, + 319 + ], + "blocks": [ + { + "bbox": [ + 475, + 261, + 534, + 319 + ], + "lines": [ + { + "bbox": [ + 475, + 261, + 534, + 319 + ], + "spans": [ + { + "bbox": [ + 475, + 261, + 534, + 319 + ], + "type": "image", + "image_path": "45ba09f138e0343c871944764b42420e4866ddf088c07a14d5320df44c2490d5.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 326, + 511, + 338 + ], + "lines": [ + { + "bbox": [ + 498, + 326, + 511, + 338 + ], + "spans": [ + { + "bbox": [ + 498, + 326, + 511, + 338 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 48, + 405, + 558, + 502 + ], + "blocks": [ + { + "bbox": [ + 48, + 405, + 558, + 502 + ], + "lines": [ + { + "bbox": [ + 48, + 405, + 558, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 558, + 502 + ], + "type": "table", + "html": "
MethodBOS Test ImagesBOS-free Test Images
GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓GR ↓LR ↓GS ↑LS ↑GB ↓LB ↓
ShadowGAN [53]7.51167.4640.9610.1970.4460.89017.32576.5080.9010.0600.4250.842
MaskshadowGAN [13]8.99779.4180.9510.1800.5001.00019.33894.3270.9060.0440.5001.000
ARShadowGAN [22]7.33558.0370.9610.2410.3830.76116.06763.7130.9080.1040.3490.682
SGRNet [12]7.18468.2550.9640.2060.3010.59615.59660.3500.9090.1000.2710.534
SGDiffusion6.09853.6110.9710.3700.2450.48715.11055.8740.9130.1170.2330.452
", + "image_path": "413f6a902a31e6210de757d9beee744f8253b92253c481487e6b6119370e57ff.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 97, + 520, + 494, + 532 + ], + "lines": [ + { + "bbox": [ + 97, + 520, + 494, + 532 + ], + "spans": [ + { + "bbox": [ + 97, + 520, + 494, + 532 + ], + "type": "text", + "content": "Table 1. The results of different methods on DESOBAv2 dataset. The best results are highlighted in boldface." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 552, + 147, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 147, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 147, + 563 + ], + "type": "text", + "content": "6.4. Ablation Studies" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "content": "We study the impact of weighted noise loss (WL), intensity modulation (IM), and post-processing (PP) of our SGDiffusion on BOS test images from DESOBAv2. The quantitative results are summarized in Table 2." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "In row 1, we report the results of basic ControlNet without weighted noise loss. For WL, the comparison between row 3 and row 1 emphasizes the importance of paying more attention to the foreground shadow region. We also report a special case " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " in row 2, where the foreground shadow mask is not expanded when constructing the weight map. The results in row 2 are comparable or even worse than those in" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "content": "row 1, as the model tends to generate larger shadow size while ignoring shape and edge details. For IM, the comparison between row 1 and row 5 shows that the intensity modulation can significantly improve the shadow quality by adjusting the shadow intensity. We also report a special case in row 4, where the intensity encoder input does not contain background shadow mask. The comparison between row 4 and row 5 shows that background shadow mask is helpful, because the background shadow regions and their surrounding regions could provide useful clues to infer shadow intensity. For PP, the comparison between row 6 and row 7 demonstrates that post-processing effectively corrects color shift and background variations, substantially reducing the" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8127" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 83, + 127, + 149 + ], + "blocks": [ + { + "bbox": [ + 62, + 83, + 127, + 149 + ], + "lines": [ + { + "bbox": [ + 62, + 83, + 127, + 149 + ], + "spans": [ + { + "bbox": [ + 62, + 83, + 127, + 149 + ], + "type": "image", + "image_path": "5a77dbfb90c264b8c2247d452a947b5b11b357f2f9605ffa3eee469f3375b944.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 83, + 194, + 149 + ], + "blocks": [ + { + "bbox": [ + 129, + 83, + 194, + 149 + ], + "lines": [ + { + "bbox": [ + 129, + 83, + 194, + 149 + ], + "spans": [ + { + "bbox": [ + 129, + 83, + 194, + 149 + ], + "type": "image", + "image_path": "a550b99cae36e0890550c8265d965f72d9faef0ca07747ea98f20e914052fe90.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 197, + 83, + 261, + 149 + ], + "blocks": [ + { + "bbox": [ + 197, + 83, + 261, + 149 + ], + "lines": [ + { + "bbox": [ + 197, + 83, + 261, + 149 + ], + "spans": [ + { + "bbox": [ + 197, + 83, + 261, + 149 + ], + "type": "image", + "image_path": "ee63cb7b91a9ced50431d72d1d7f09168bf541328b8ddbabfd8d56ef03fca9d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 264, + 83, + 329, + 149 + ], + "blocks": [ + { + "bbox": [ + 264, + 83, + 329, + 149 + ], + "lines": [ + { + "bbox": [ + 264, + 83, + 329, + 149 + ], + "spans": [ + { + "bbox": [ + 264, + 83, + 329, + 149 + ], + "type": "image", + "image_path": "0ceca2d6bfaa2990472ac7e489e40718d3b1586153d1972d86e692e61537a11d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 331, + 83, + 396, + 149 + ], + "blocks": [ + { + "bbox": [ + 331, + 83, + 396, + 149 + ], + "lines": [ + { + "bbox": [ + 331, + 83, + 396, + 149 + ], + "spans": [ + { + "bbox": [ + 331, + 83, + 396, + 149 + ], + "type": "image", + "image_path": "3c138983ce75567ea481cc3f28c4b6db9a1fa3984e2a62a81a5257bf4fa7bab3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 83, + 464, + 149 + ], + "blocks": [ + { + "bbox": [ + 399, + 83, + 464, + 149 + ], + "lines": [ + { + "bbox": [ + 399, + 83, + 464, + 149 + ], + "spans": [ + { + "bbox": [ + 399, + 83, + 464, + 149 + ], + "type": "image", + "image_path": "720f8cf315f5c7c21e8af4b16451ff3a3fa03eb2672a9de84ac2fea14795a662.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 466, + 83, + 532, + 149 + ], + "blocks": [ + { + "bbox": [ + 466, + 83, + 532, + 149 + ], + "lines": [ + { + "bbox": [ + 466, + 83, + 532, + 149 + ], + "spans": [ + { + "bbox": [ + 466, + 83, + 532, + 149 + ], + "type": "image", + "image_path": "9035dcbe44ad55d64ad412cfa7b53b83c1fe56940136a0d77c38e3244942da3f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 151, + 127, + 217 + ], + "blocks": [ + { + "bbox": [ + 62, + 151, + 127, + 217 + ], + "lines": [ + { + "bbox": [ + 62, + 151, + 127, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 151, + 127, + 217 + ], + "type": "image", + "image_path": "a2b2fbc67b24c9ed9f97d1a8b35b288a573dae4ae29947b52671d4957560cf4e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 129, + 151, + 194, + 217 + ], + "blocks": [ + { + "bbox": [ + 129, + 151, + 194, + 217 + ], + "lines": [ + { + "bbox": [ + 129, + 151, + 194, + 217 + ], + "spans": [ + { + "bbox": [ + 129, + 151, + 194, + 217 + ], + "type": "image", + "image_path": "d6269f88ea18a3ef521f98bc0ff94fd8d72e9e18fe9025c57737109379172661.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 196, + 151, + 261, + 217 + ], + "blocks": [ + { + "bbox": [ + 196, + 151, + 261, + 217 + ], + "lines": [ + { + "bbox": [ + 196, + 151, + 261, + 217 + ], + "spans": [ + { + "bbox": [ + 196, + 151, + 261, + 217 + ], + "type": "image", + "image_path": "21a03b159c9bbb8fff0123bfd57097a82dad8c49bb2f2dc5f2034addefac9cc8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 264, + 151, + 329, + 217 + ], + "blocks": [ + { + "bbox": [ + 264, + 151, + 329, + 217 + ], + "lines": [ + { + "bbox": [ + 264, + 151, + 329, + 217 + ], + "spans": [ + { + "bbox": [ + 264, + 151, + 329, + 217 + ], + "type": "image", + "image_path": "94e2f4a949c725541f47dadea13d618553d0ace03fb20b9c321eb98d6b04c8de.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 151, + 396, + 217 + ], + "blocks": [ + { + "bbox": [ + 331, + 151, + 396, + 217 + ], + "lines": [ + { + "bbox": [ + 331, + 151, + 396, + 217 + ], + "spans": [ + { + "bbox": [ + 331, + 151, + 396, + 217 + ], + "type": "image", + "image_path": "862abce213ed04ccc67149d11dc90fddc48331f81641df9d61e994f59ff46f8d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 399, + 151, + 464, + 217 + ], + "blocks": [ + { + "bbox": [ + 399, + 151, + 464, + 217 + ], + "lines": [ + { + "bbox": [ + 399, + 151, + 464, + 217 + ], + "spans": [ + { + "bbox": [ + 399, + 151, + 464, + 217 + ], + "type": "image", + "image_path": "94a3472d4d3f7b6b1b047d1a277a33d6225ff2736c2532c23cd980aefd1e5830.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 466, + 151, + 532, + 217 + ], + "blocks": [ + { + "bbox": [ + 466, + 151, + 532, + 217 + ], + "lines": [ + { + "bbox": [ + 466, + 151, + 532, + 217 + ], + "spans": [ + { + "bbox": [ + 466, + 151, + 532, + 217 + ], + "type": "image", + "image_path": "f984aab0d189cec403acfa49a2dcb0987b5f5ea1957a11c0b2034f5f45833a24.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 62, + 219, + 128, + 285 + ], + "blocks": [ + { + "bbox": [ + 62, + 219, + 128, + 285 + ], + "lines": [ + { + "bbox": [ + 62, + 219, + 128, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 128, + 285 + ], + "type": "image", + "image_path": "ed1fef56431338ad1362df518088686fff4bd6a242e9a6819a8056e85ad3ae7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 293, + 102, + 305 + ], + "lines": [ + { + "bbox": [ + 88, + 293, + 102, + 305 + ], + "spans": [ + { + "bbox": [ + 88, + 293, + 102, + 305 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 129, + 219, + 194, + 285 + ], + "blocks": [ + { + "bbox": [ + 129, + 219, + 194, + 285 + ], + "lines": [ + { + "bbox": [ + 129, + 219, + 194, + 285 + ], + "spans": [ + { + "bbox": [ + 129, + 219, + 194, + 285 + ], + "type": "image", + "image_path": "5554a5c482fdebafb1dbeaf7ecf6044a6e25687c6a5fb7e6323e0423e41d7ba5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 293, + 170, + 304 + ], + "lines": [ + { + "bbox": [ + 155, + 293, + 170, + 304 + ], + "spans": [ + { + "bbox": [ + 155, + 293, + 170, + 304 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 196, + 219, + 261, + 285 + ], + "blocks": [ + { + "bbox": [ + 196, + 219, + 261, + 285 + ], + "lines": [ + { + "bbox": [ + 196, + 219, + 261, + 285 + ], + "spans": [ + { + "bbox": [ + 196, + 219, + 261, + 285 + ], + "type": "image", + "image_path": "060331219053558fc6bcecd058e7c6302128bd171e128583d40122fef4e607b8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 293, + 236, + 304 + ], + "lines": [ + { + "bbox": [ + 223, + 293, + 236, + 304 + ], + "spans": [ + { + "bbox": [ + 223, + 293, + 236, + 304 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 264, + 219, + 329, + 285 + ], + "blocks": [ + { + "bbox": [ + 264, + 219, + 329, + 285 + ], + "lines": [ + { + "bbox": [ + 264, + 219, + 329, + 285 + ], + "spans": [ + { + "bbox": [ + 264, + 219, + 329, + 285 + ], + "type": "image", + "image_path": "057d50aefabba5f5c5108789ebe54ca984c8944ccc410196b2eb3ce67b58330d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 293, + 304, + 304 + ], + "lines": [ + { + "bbox": [ + 290, + 293, + 304, + 304 + ], + "spans": [ + { + "bbox": [ + 290, + 293, + 304, + 304 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 330, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 547, + 353 + ], + "type": "text", + "content": "Figure 5. Visual comparison of different methods on real composite images. From left to right are input composite image (a), foreground object mask (b), results of ShadowGAN [53] (c), MaskshadowGAN [13] (d), ARShadowGAN [22] (e), SGRNet [12] (f), SGDiffusion (g)." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 331, + 219, + 397, + 285 + ], + "blocks": [ + { + "bbox": [ + 331, + 219, + 397, + 285 + ], + "lines": [ + { + "bbox": [ + 331, + 219, + 397, + 285 + ], + "spans": [ + { + "bbox": [ + 331, + 219, + 397, + 285 + ], + "type": "image", + "image_path": "5cdd0959c16b8d6dca11b9a5849da66a80e27a483893372ba668c4da84179310.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 293, + 372, + 304 + ], + "lines": [ + { + "bbox": [ + 358, + 293, + 372, + 304 + ], + "spans": [ + { + "bbox": [ + 358, + 293, + 372, + 304 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 399, + 219, + 465, + 285 + ], + "blocks": [ + { + "bbox": [ + 399, + 219, + 465, + 285 + ], + "lines": [ + { + "bbox": [ + 399, + 219, + 465, + 285 + ], + "spans": [ + { + "bbox": [ + 399, + 219, + 465, + 285 + ], + "type": "image", + "image_path": "51c7b4577166f181f30d2a083fbab261eb666f34aa7d1ed12b233ee934c1200e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 293, + 438, + 304 + ], + "lines": [ + { + "bbox": [ + 425, + 293, + 438, + 304 + ], + "spans": [ + { + "bbox": [ + 425, + 293, + 438, + 304 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 466, + 219, + 532, + 285 + ], + "blocks": [ + { + "bbox": [ + 466, + 219, + 532, + 285 + ], + "lines": [ + { + "bbox": [ + 466, + 219, + 532, + 285 + ], + "spans": [ + { + "bbox": [ + 466, + 219, + 532, + 285 + ], + "type": "image", + "image_path": "5873fc882610db1e3f148726dc21fe9d45d801e3a6a2776ab94f2a99ff934df8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 293, + 506, + 305 + ], + "lines": [ + { + "bbox": [ + 492, + 293, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 492, + 293, + 506, + 305 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 48, + 372, + 287, + 475 + ], + "blocks": [ + { + "bbox": [ + 48, + 372, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 475 + ], + "type": "table", + "html": "
RowWLIMPPGR↓LR↓GB↓LB↓
1--+8.28559.7530.2710.534
2†-+8.31959.4910.2820.563
3+-+7.04153.8290.2490.492
4-○+7.41056.1210.2690.536
5-++7.35754.1590.2620.526
6++-13.44755.2310.2450.487
7+++6.09853.6110.2450.487
", + "image_path": "525575a281ef4a448bc87836a6f5c2bfec609dd74cff9b58f035636f259da6d7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "type": "text", + "content": "Table 2. Ablation studies of our method on BOS test images from DESOBAv2 dataset. WL is short for weighted loss and " + }, + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "type": "text", + "content": " means without expanding shadow mask. IM is short for intensity modulation and " + }, + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "type": "text", + "content": " means without using background shadow mask. PP is short for post-processing." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 574, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 598 + ], + "type": "text", + "content": "global RMSE. We also provide the visual results of ablated versions in the supplementary." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 47, + 609, + 180, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 180, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 180, + 624 + ], + "type": "text", + "content": "6.5. Real Composite Images" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "We compare different methods on real composite images provided by [12], where background images and foreground objects are from the DESOBA [12] test set. We train all methods on DESOBAv2 and finetune them on DESOBA. The visual results of different methods are showcased in Figure 5. These results confirm that SGDiffusion adeptly synthesizes lifelike shadows with precise contours, loca" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "type": "text", + "content": "tions, and directions, which are compatible with the background object-shadow pairs and foreground object information. In contrast, previous methods often produce vague and misdirected shadows. We provide more examples in the supplementary." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 434, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 546, + 529 + ], + "type": "text", + "content": "Given the absence of ground-truth images for real composite images, following [12], we opt for subjective evaluation, engaging 50 human raters in the user study. Each participant is presented with image pairs from the results generated by 5 methods, and asked to choose the image with more realistic foreground shadow. Using the Bradley-Terry model [2], we report the B-T scores in the supplementary, which again proves the advantage of our method." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 537, + 378, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 378, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 378, + 550 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": "In this paper, we have contributed a large-scale shadow generation dataset DESOBAv2. We have also designed a novel diffusion-based shadow generation method. Extensive experimental results show that our method is able to generate plausible shadows for composite foregrounds, significantly surpassing previous methods." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 638, + 416, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 416, + 652 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 416, + 652 + ], + "type": "text", + "content": "8. Acknowledgement" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 657, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 545, + 704 + ], + "type": "text", + "content": "The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major/Key Project, China (Grant No. 2021SHZDZX0102)." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8128" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Ibrahim Arief, Simon McCallum, and Jon Yngve Hardeberg. Realtime estimation of illumination direction for augmented reality on mobile devices. In CIC, 2012. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 287, + 194 + ], + "type": "text", + "content": "[3] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 195, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 228 + ], + "type": "text", + "content": "[4] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In ICME, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 287, + 273 + ], + "type": "text", + "content": "[5] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In CVPR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 275, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 275, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 275, + 287, + 308 + ], + "type": "text", + "content": "[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. TIP, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 310, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 286, + 353 + ], + "type": "text", + "content": "[7] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagné, and Jean-Francois Lalonde. Deep parametric indoor lighting estimation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 355, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 355, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 355, + 286, + 388 + ], + "type": "text", + "content": "[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 390, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 390, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 286, + 434 + ], + "type": "text", + "content": "[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 286, + 468 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 471, + 286, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 493 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurlPS, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 495, + 286, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 286, + 526 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 286, + 526 + ], + "type": "text", + "content": "[12] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. AAAI, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 529, + 286, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 286, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 286, + 562 + ], + "type": "text", + "content": "[13] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, 2019. 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 563, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 286, + 586 + ], + "type": "text", + "content": "[14] Shaozong Huang and Lan Hong. Diffusion model for mural image inpainting. In ITOEC, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 587, + 286, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 286, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 286, + 631 + ], + "type": "text", + "content": "[15] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM TOG, 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 633, + 286, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 633, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 48, + 633, + 286, + 665 + ], + "type": "text", + "content": "[16] Eric Kee, James F. O'Brien, and Hany Samir Farid. Exposing photo manipulation from shading and shadows. ACM TOG, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 667, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 667, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 667, + 286, + 689 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[19] Bin Liao, Yao Zhu, Chao Liang, Fei Luo, and Chunxia Xiao. Illumination animating and editing in a single picture using scene structure estimation. Computers & Graphics, 82:53-64, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[20] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In CVPR, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[21] Bin Liu, Kun Xu, and Ralph R Martin. Static scene illumination estimation from videos with applications. JCST, 32 (3):430-442, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[22] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In CVPR, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "text", + "content": "[23] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[24] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. CoRR, abs/2303.17870, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "text", + "content": "[25] Quanling Meng, Shengping Zhang, Zonglin Li, Chenyang Wang, Weigang Zhang, and Qingming Huang. Automatic shadow generation via exposure fusion. IEEE Transactions on Multimedia, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[26] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, 2016. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[28] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. CoRR, abs/2106.14490, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "text", + "content": "[29] Sibam Parida, Vignesh Srinivas, Bhavishya Jain, Rajesh Naik, and Neeraj Rao. Survey on diverse image inpainting using diffusion models. In PCEMS, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 522, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 577 + ], + "type": "text", + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NIPS, 32, 2019. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 578, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 601 + ], + "type": "text", + "content": "[31] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In SIGGRAPH. 2003. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[34] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In CVPR, 2021. 1" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8129" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[35] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[36] Yichen Sheng, Jianming Zhang, Julien Philip, Yannick Hold-Geoffroy, Xin Sun, He Zhang, Lu Ling, and Bedrich Benes. Pixht-lab: Pixel height based light effect generation for image compositing. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "text", + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. CoRR, abs/2010.02502, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[38] Yi-Zhe Song, Zhifei Zhang, Zhe L. Lin, Scott D. Cohen, Brian L. Price, Jianming Zhang, Soo Ye Kim, and Daniel G. Aliaga. Objectstitch: Generative object compositing. In CVPR, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[39] Xinhao Tao, Junyan Cao, Yan Hong, and Li Niu. Shadow generation with decomposed mask prediction and attentive shadow filling. In AAAI, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "text", + "content": "[40] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In CVPR, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[41] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. TPAMI, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 286, + 374 + ], + "type": "text", + "content": "[42] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In ACM MM, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "text", + "content": "[43] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[44] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "text", + "content": "[45] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[46] Fangneng Zhan, Jiaxing Huang, and Shijian Lu. Adaptive composition gan towards realistic image synthesis. CoRR, abs/1905.04693, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "text", + "content": "[47] Fangneng Zhan, Shijian Lu, Changgong Zhang, Feiying Ma, and Xuansong Xie. Towards realistic 3d embedding via view alignment. CoRR, abs/2007.07066, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 610 + ], + "type": "text", + "content": "[48] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "text", + "content": "[49] He Zhang, Jianming Zhang, Federico Perazzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In WACV, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[50] Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-François Lalonde. All-weather deep outdoor lighting estimation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[51] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In WACV, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 274 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[53] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5:105-115, 2019. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[54] Lei Zhao, Qihang Mo, Sihuan Lin, Zhizhong Wang, Zhiwen Zuo, Haibo Chen, Wei Xing, and Dongming Lu. Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In CVPR, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 197, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 217 + ], + "type": "text", + "content": "[55] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In CVPR, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[56] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In ECCV, 2022. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8130" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_content_list.json b/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e25c6039e614b44ff91a4e63cc430398e23167b5 --- /dev/null +++ b/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_content_list.json @@ -0,0 +1,2400 @@ +[ + { + "type": "text", + "text": "Shadow-Enlightened Image Outpainting", + "text_level": 1, + "bbox": [ + 277, + 130, + 691, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hang Yu $^{1}$ , Ruilin Li $^{2}$ , Shaorong Xie $^{1}$ , Jiayan Qiu $^{3*}$ $^{1}$ School of Computer Engineering and Science, Shanghai University, China \n $^{2}$ Institute of Artificial Intelligence, Shanghai University, China \n $^{3}$ School of Computing and Mathematical Sciences, University of Leicester, United Kingdom {yuhang, ruilinli, srxie}@shu.edu.cn, jiayan.qiu.1991@outlook.com", + "bbox": [ + 116, + 179, + 852, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional image outpainting methods usually treat unobserved areas as unknown and extend the scene only in terms of semantic consistency, thus overlooking the hidden information in shadows cast by unobserved areas, such as the invisible shapes and semantics. In this paper, we propose to extract and utilize the hidden information of unobserved areas from their shadows to enhance image outpainting. To this end, we propose an end-to-end deep approach that explicitly looks into the shadows within the image. Specifically, we extract shadows from the input image and identify instance-level shadow regions cast by the unobserved areas. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. In the experiments, we show that our proposed approach provides complementary cues for outpainting and achieves considerable improvement on all datasets by adopting our approach as a plug-in module.", + "bbox": [ + 75, + 335, + 473, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given an image of a complex scene, as shown in Fig. 1(a), what will its right-side unobserved area look like? Humans can, without much effort, infer the appropriate categories and rough shapes of invisible semantics (e.g. objects) and further imagine an adequately realistic overall appearance. One critical cue humans utilize to reason invisible semantics is the shadows cast in the image. In this regard, we identify the shadows of each visible object, on top of which the shadows cast by the unobserved areas are found. According to their shapes and relative spatial layout, we may reasonably infer that there is a person in the unobserved area, as shown in Fig. 1(b), which aligns with its", + "bbox": [ + 75, + 694, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b881424d830e8672e1b846d373e8f17561ae59e056d25d47c5550881cf33be1.jpg", + "image_caption": [ + "Figure 1. Given the image of a complex scene in (a) for outpainting, the extension needs to align with visible semantics. For example, the result shown in (b) is more aligned with the visible semantics from a global shadowing direction compared to the result shown in (c). For better alignment, we overcome the challenges of shadow extraction in complex scenes, such as disjoined shadows (d), overlapped shadows (e), and shadows attaching unexpected semantic information (f), respectively." + ], + "image_footnote": [], + "bbox": [ + 506, + 304, + 888, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shadow and the given scene. However, existing outpainting methods [7, 26, 57, 58, 69, 74] overlook shadows in the image, leading to inferior results, as shown in Fig. 1(c), where misalignment exists between the extension and its shadows from the shadowing direction in the scene.", + "bbox": [ + 496, + 670, + 890, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Even though prior efforts on Shape-from-Shadow [22, 32, 36, 60] also aim to extract hidden information, such as the shapes, of the semantics from their shadows in simple scenes with an ideal light source and a given semantic category, they cannot perform complex image outpainting with unknown semantics, due to there are three challenges of shadow extraction in complex scenes. 1) Given the complex spatial layout of the instances, the shadows are usually occluded by other instances, resulting in disjoined shadows of a single object, as shown in Fig. 1(d), which leads to inaccu", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7850", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rate perception or misinterpretation of invisible semantics. 2) Given the complex illumination of the scene, the shadows cast by different semantics may overlap, thus causing information loss for the unobserved area, as shown in Fig. 1(e). 3) Given the complex spatial layout of the scene, the shadows usually contain semantic information about the planes on which they are cast, which may lead to shadow-irrelevant understanding, as shown in Fig. 1(f).", + "bbox": [ + 75, + 90, + 467, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To perform image outpainting by utilizing the hidden information in the shadows, we propose a novel approach that explicitly looks into the correlation between shadows from the unobserved area and the visible semantics, the shadows with semantic removal, and the consistency between local and global alignment.", + "bbox": [ + 75, + 210, + 467, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, as depicted in Fig. 2, given an input image, we first conduct shadow extraction to obtain those shadow regions belonging to unobserved instances. Then, we utilize the shadow representations from unobserved instances to predict the scene layout of these instances, and then concatenate them with the shadow representation of instances in unobserved areas for layout-to-image processing. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas for the outpainting image. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows.", + "bbox": [ + 75, + 301, + 467, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution is therefore a novel framework designated to achieve image outpainting by utilizing the shadows, which is the first attempt to the best of our knowledge. This is accomplished by extracting the desired shadow information, followed by scene layout expansion and layout-to-image conversion to produce image outpainting. The whole pipeline is end-to-end trainable. By adopting our proposed framework as a plug-in module, the outpainting methods [69] have achieved considerable quantitative and qualitative performance improvement on all datasets.", + "bbox": [ + 75, + 484, + 467, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 648, + 217, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly review the prior works related to ours, including shadow processing and image outpainting.", + "bbox": [ + 75, + 674, + 467, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shadow Processing. Conventional shadow processing method reconstructs the surface shape of a single object, under ideal illuminance, from its shadow and visual information [40]. Then, a sequence of methods [1, 2, 13, 14, 31, 62, 66] is developed to improve its performance by involving carefully designed prior knowledge in the model. Then, due to the strong adaptability of deep learning, learning-based methods [22, 32] are proposed to deal with complicated environments, such as objects with non-Lambertian reflectance [32]. Moreover, shadows are utilized as cues for scene geometry information extraction [21, 23, 30, 36, 43-45, 47, 48, 67, 68], segmentation [12], object detection [9, 16, 17], and tracking [41]. Recently, a generative method", + "bbox": [ + 75, + 704, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[36] has been proposed to predict the shape of a single invisible object, with a specific category, from its shadow. However, none of the existing methods can infer the scene layout of the semantics from their shadows in complex scenes.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Outpainting. Conventional image outpainting methods extend the image with the background texture and partially observed objects towards specific directions [5, 34, 37, 52, 73, 74]. Then, information on the image margin is utilized to enable the outpainting in any direction [28, 56, 64]. Recently, the extrapolation of segmentation information has been adopted in the outpainting process [26] to add novel semantics in the extending areas. Meanwhile, the scene graph of the image is utilized to introduce new semantics by learning the semantic co-occurrence with Graph Neural Networks (GNN) [19, 33, 49, 50, 59, 69, 72]. Conventional outpainting commonly uses Generative Adversarial Networks (GAN) [15] for image conversion. However, due to the stability of diffusion, diffusion models [3, 18, 42, 51, 55] are being recognized as a promising family of generative models that have proven to be state-of-the-art sample quality for a variety of image generation benchmarks [8, 61, 70], including class-conditional image generation [10, 75], text-to-image generation [27, 51, 53, 71], image-to-image translation [24, 38, 54], layout-to-image generation [6, 76]. However, none of the existing methods takes advantage of the shadows to enhance the outpainting.", + "bbox": [ + 498, + 152, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 511, + 589, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we detail the working scheme of the proposed approach which comprises three stages, as shown in Fig. 2. In Stage 1, we utilize the pretrained model of shadow detection and instance shadow detection to grain rough instance-level shadow regions. Meanwhile, a pretrained shadow removal model is to remove unexpected semantic information from the shadows for the purified shadow feature. Then, we calculate the connectivity between shadow regions to merge disjoined shadows and perform overlapped segmentation on the shadows to restore overlapped shadows of instances. After iterative optimization, we obtain refined instance-level shadow regions. In Stage 2, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. In Stage 3, two discriminators are implemented to enhance alignment between the extended semantics and their shadows.", + "bbox": [ + 498, + 536, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Stage 1: Shadow Extraction", + "text_level": 1, + "bbox": [ + 500, + 801, + 751, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the input image $I$ , we first adopt a pretrained shadow detection model [77] to extract all shadow areas:", + "bbox": [ + 500, + 825, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s} = \\operatorname {D e t e c t i o n} (I), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 880, + 890, + 896 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7851", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb124024b5e96db202022757c0422ec5480f147018b0082118277ea5c3236d21.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed approach. $\\bigoplus$ denotes concatenation." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 864, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M_{s}$ denotes the mask of the shadow areas, respectively. Then, a pretrained instance shadow detection model [63] is introduced to identify the shadow areas for the visible semantics in image $I$ by:", + "bbox": [ + 75, + 455, + 468, + 516 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s} ^ {v} = \\text {I n s t a n c e - D e t e c t i o n} (I). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 526, + 468, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M_{s}^{v}$ denotes the instance-level mask of the shadow areas cast by the visible semantics. Consequently, a binary area-level mask of shadows cast by the unobserved area can be obtained with $M_{s}^{u} = M_{s} - M_{s}^{v}$ . Thus, we obtain a rough instance-level shadow region.", + "bbox": [ + 75, + 553, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, to obtain the shadow areas without the semantic information from the cast planes, we introduce a pretrained shadow removal model [39] and obtain the semantically purified shadow image $I_{s}^{p}$ by:", + "bbox": [ + 75, + 628, + 468, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {s} ^ {p} = M _ {s} \\odot (I - \\text {S h a d o w R e m o v a l} (I)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 699, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the Hadamard product.", + "bbox": [ + 76, + 727, + 346, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After obtaining the shadow image of semantic removal $I_{s}^{p}$ , we can use a convolutional net to extract the shadow feature without semantics:", + "bbox": [ + 76, + 742, + 468, + 786 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s} ^ {p} = \\operatorname {C o n v N e t} \\left(I _ {s} ^ {p}\\right), F _ {s} ^ {p} \\in \\mathbb {R} ^ {W \\times H \\times d}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 796, + 468, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the ConvNet consists of 8 Residual Blocks and preserves the resolution of the output representation to be the same as the input image. $W, H$ denotes the width and height of the input image. $d$ denotes the dimension of the shadow feature map.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering that the formation of shadows is influenced by the illumination conditions in the given scene, which can be a cue in addressing the issues of disjoined shadows and overlapped shadows, we concatenate the image $I$ and the shadow mask $M_{s}$ and fed it into a Convolutional Neural Network (CNN) to obtain the illumination feature $F^{\\mathrm{illum}} \\in \\mathbb{R}^{(W / 16) \\times (H / 16) \\times d}$ in the scene as a cue for the shadow regions optimization.", + "bbox": [ + 496, + 455, + 890, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Shadow Merging Operation", + "text_level": 1, + "bbox": [ + 498, + 594, + 748, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In complex scenes, due to the presence of instances that can occlude shadows, it can result in an instance having multiple disjointed shadows, as shown in Fig. 3.", + "bbox": [ + 496, + 618, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To classify a single instance with disjoined shadows, we first treat each connected component in $M_{s}^{u}$ as a shadow region weight mask. Let $M^{u} = \\{m_{x}^{u}\\} \\in \\mathbb{R}^{W\\times H\\times 1}, x\\in [1,N_{u}]$ be the shadow weight masks that don't belong to any visible instances, where $N_{u}$ denotes the number of connected components in $M_{s}^{u}$ . For the rough instance-level mask of visible instances in $M_{s}^{v}$ , let $M^{v} = \\{M_{y}^{v}\\} \\in \\mathbb{R}^{W\\times H\\times 1}, y\\in [1,N_{v}]$ be each shadow weight mask of visible instance, where $N_{v}$ denotes the number of visible instances. Consequently, we grain the initial instance-level shadow region $M$ :", + "bbox": [ + 496, + 662, + 890, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nN = N _ {u} + N _ {v} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 827, + 890, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM = \\left\\{m _ {z} \\right\\} = \\left\\{M ^ {u}, M ^ {v} \\right\\}, z \\in [ 1, N ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 845, + 890, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each shadow region, we extract the corresponding region-level shadow feature $F$ by the feature map by", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7852", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/78af9bcb0430208658794ee617448a99469a531a21d2fd497787f1a362cdd9b1.jpg", + "image_caption": [ + "(a) the disjoined shadow of the unobserved instance" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 271, + 194 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4d3503e3554b18a5242c2d741d0d6f7034a3b1c01ebbe23bd3bf814b036a3c5.jpg", + "image_caption": [ + "(b) the disjoined shadow of the visible instance", + "Figure 3. Samples of disjoined shadows caused by occlusion in complex scenes. (a) The shadow of an unobserved object is obstructed, creating two disjointed shadow regions. The red boxes indicate the disjoined shadows. (b) Due to the defect of the instance shadow detector, a portion of the shadow of the instance was not classified as a single instance. The blue boxes indicate the detected corresponding pair, and the yellow box indicates the undetected shadow region." + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 467, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "element-wise multiplying the shadow feature without semantics $F_{s}^{p}$ and the shadow weight mask $M$ . Then, we use the max pooling operation to resize these features to match the size of the illumination feature. Thus, we obtain the cropped feature $\\bar{F}$ by:", + "bbox": [ + 75, + 348, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF = F _ {s} ^ {p} \\odot M, F \\in \\mathbb {R} ^ {W \\times H \\times d}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 433, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {F} = \\operatorname {m a x p o o l} (F), \\bar {F} \\in \\mathbb {R} ^ {(W / 1 6) \\times (H / 1 6) \\times d}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 454, + 468, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the computation of connectivity between shadow regions, we combine the cropped features $\\bar{F}$ of two regions along with the illumination feature by concatenating them together. This concatenated input is fed into an encoder $E_{connect}$ , and then passed through a sigmoid activation function for connectivity between two shadow regions:", + "bbox": [ + 75, + 483, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {F} = \\{\\bar {f} _ {m} \\}, m \\in [ 1, N ], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 599, + 468, + 618 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f} _ {(i, j)} = \\operatorname {E} _ {\\text {c o n n e c t}} \\left(\\operatorname {c o n c a t} \\left(\\bar {f} _ {i}, \\bar {f} _ {j}, F ^ {\\text {i l l u m}}\\right)\\right), i \\neq j \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 619, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o n n e c t i v i t y} _ {(i, j)} = \\operatorname {S i g m o i d} (\\bar {f} _ {(i, j)}), i \\neq j \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 640, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the connectivity between two shadow regions is higher than a hand-setting threshold $\\alpha_{1}$ , these two shadow regions are considered to be the shadows of the same object, accordingly merging their weight masks. Conversely, they stay unchanged. Finally, we obtain the updated shadow features $F^{update}$ after merging operation:", + "bbox": [ + 75, + 667, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {u p d a t e} = \\left\\{f _ {m} \\right\\} \\in \\mathbb {R} ^ {W \\times H \\times d}, m \\in [ 1, N ^ {m} ], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 768, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N^{m}$ denotes the number of instances after merging.", + "bbox": [ + 76, + 797, + 460, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Shadow Splitting Operation", + "text_level": 1, + "bbox": [ + 76, + 830, + 326, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In complex scenes, due to the complex spatial layout in the scene, it is common to have overlapped shadows between instances, as shown in Fig. 4.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19a1e402b502588010d9ae9701cb8c53b388821d206345b1d1ae1c6f8f040369.jpg", + "image_caption": [ + "(a) overlapped with shadows cast by the unobserved instances", + "Figure 4. Samples of overlapping shadows in complex scenes. The red boxes indicate the overlapped shadows. (a) The shadows of multiple unobserved instances overlap. (b) The shadows of unobserved instances overlap with the shadows of visible instances." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 696, + 194 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/af4fb7cb5629cdb12b6b421c22c56b4d8c2cd7275beb997a4cd3c13aaaf4248f.jpg", + "image_caption": [ + "(b) overlapped with shadow cast by the visible instances" + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 888, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, each overlapped shadow is decoupled into two intersecting graph spaces. After concatenating the cropped feature $\\bar{F}$ and illumination feature $F^{illum}$ , we input them into the Overlapping BiLayers [25]. Then, we grain two weight masks $M_0'$ , and $M_1'$ , each weight mask represents the shadow region of an individual instance. Then, we can obtain the \"de-overlapped\" shadow feature $f$ of an instance by element-wise multiplying the weight mask $M'$ with the purified shadow feature $F_s^p$ :", + "bbox": [ + 496, + 286, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {0} ^ {\\prime}, M _ {1} ^ {\\prime}, \\in \\mathbb {R} ^ {W \\times H \\times 1} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 431, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {0} = M _ {0} ^ {\\prime} \\odot F _ {s} ^ {p}, f _ {1} = M _ {1} ^ {\\prime} \\odot F _ {s} ^ {p}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 453, + 890, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "when the ratio of the combined area of two weight masks to the source area is higher than a hand-setting threshold $\\alpha_{2}$ , it is considered an overlapped shadow and needs to be split. Conversely, it stays unchanged. Finally, we obtain the updated shadow features $F^{update}$ after splitting operation:", + "bbox": [ + 496, + 481, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {u p d a t e} = \\left\\{f _ {m} \\right\\} \\in \\mathbb {R} ^ {W \\times H \\times d}, m \\in [ 1, N ^ {s} ], \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 582, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N^s$ denotes the number of instances after splitting.", + "bbox": [ + 500, + 612, + 877, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Iterative Optimization", + "text_level": 1, + "bbox": [ + 500, + 645, + 709, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In complex scenes, it is common to encounter situations where disjoined shadows and overlapped shadows co-occur. Therefore, we utilize an iterative combination of merging and splitting operations to optimize the categorization of all shadow regions in the image. Thus, the issues of disjoined shadows and overlapped shadows, shown in Fig. 3 and Fig. 4, are addressed. Specifically, the iterative shadow optimization is summarized as Algorithm 1.", + "bbox": [ + 496, + 669, + 890, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stage 2: Outpaiting with Shadows", + "text_level": 1, + "bbox": [ + 500, + 799, + 797, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the refined instance-level shadow regions $M$ from Stage 1, we grain the cropped shadow representations $\\bar{f}^u$ of unobserved instances by:", + "bbox": [ + 500, + 823, + 890, + 868 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f} ^ {u} = \\operatorname {m a x p o o l} \\left(M ^ {u} \\odot F _ {s} ^ {p}\\right), M ^ {u} \\in M, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 878, + 890, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7853", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Iterative Shadow Optimization" + ], + "code_body": "Input: rough instance-level shadow weight mask \nOutput: refined instance-level shadow weight mask \n1 $M^{(0)}\\gets M$ \n2 for $t = 1$ to $T$ do \n3 $\\begin{array}{rl} & M^{(t)}\\leftarrow \\mathrm{ShadowMerging}(M^{(t - 1)})\\\\ & M^{(t)}\\leftarrow \\mathrm{ShadowSplitting}(M^{(t)})\\\\ & \\mathrm{if}\\| M^{(t)} - M^{(t - 1)}\\| < \\epsilon \\mathrm{then}\\\\ & \\mid \\mathrm{break}\\\\ & \\mathrm{end} \\end{array}$ \n5 \n6 \n7 \n8 end \n9 return $M^{(t)}$", + "bbox": [ + 81, + 111, + 467, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which are concatenated to predict the categories and layouts of new instances and outpaint the unobserved areas.", + "bbox": [ + 75, + 320, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For scene layout expansion, given a partial input image with the partial scene graph and the corresponding layout, we use GNN [11] to reveal the scene layout $L_{exp}$ of new instances. Specifically, we pass the shadow representations $\\tilde{f}^u$ from the unobserved instances through an encoder $E_h$ , and then concatenate the output with the node features $f^h$ of the masked instances from the encoder of GNN.", + "bbox": [ + 75, + 351, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} ^ {h} = \\operatorname {c o n c a t} \\left(f ^ {h}, \\mathrm {E} _ {\\mathrm {h}} \\left(\\bar {f} ^ {u}\\right)\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 468, + 468, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For layout to image, given an input image $I$ with an expanded scene layout $L_{exp}$ , we use layout-to-image model [76] to perform layout to image conversion. Specifically, we pass the shadow representations $f^u$ from the unobserved instances through an encoder $E_l$ , and then concatenate the output with scene layout features $f^l$ of the expanded instance from the encoder of the layout-to-image module.", + "bbox": [ + 75, + 500, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} ^ {l} = \\operatorname {c o n c a t} \\left(f ^ {l}, \\mathrm {E} _ {\\mathrm {l}} \\left(\\bar {f} ^ {u}\\right)\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 631, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we focus on the design of extracting shadow representation, so producing high-accuracy predictions for scene layout and high-quality image outputs is not within the main scope of this work. Thus, the graph-based and diffusion-based designs can be replaced with any scene layout expansion and layout-to-image modules if desirable.", + "bbox": [ + 75, + 662, + 468, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Stage 3: Alignment Discriminator", + "text_level": 1, + "bbox": [ + 76, + 770, + 374, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In stage 2, we outpaint the partial input image $I$ into the extended image $I_{E}$ with introduced instances. To enhance alignment between unobserved area and visible semantics, we adopt the pretrained instance shadow detector [63] to first extract the mask $M_E^u$ of the unobserved area and its corresponding shadow from $I_{E}$ , and then feed it into a local shadow-instance alignment discriminator to ensure the", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "generated alignment is visually real. This is calculated as follows:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {L D} V (L D) = \\frac {1}{2} \\mathbb {E} _ {x \\sim p _ {\\text {d a t a}} (x)} [ (L D (x) - b) ^ {2} ] \\tag {19} \\\\ + \\frac {1}{2} \\mathbb {E} _ {z \\sim p z (z)} [ (L D (M _ {E} ^ {u} \\odot G (z)) - a) ^ {2} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 128, + 890, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G} V (G) = \\frac {1}{2} \\mathbb {E} _ {z \\sim p z (z)} [ (L D (M _ {E} ^ {u} \\odot G (z)) - c) ^ {2} ], \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 213, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $LD$ denotes the local alignment discriminator. $G$ denotes the generator for outpainting. $a$ and $b$ denote the ground truth real and fake labels, respectively. $c$ denotes the value that $G$ wants $LD$ to believe for the fake data.", + "bbox": [ + 496, + 252, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the mask $M_E$ that consists of all semantics and their shadows in $I_E$ is extracted, and we feed it into a global alignment discriminator to ensure the newly generated alignment is consistent with the existing ones. This is calculated as follows:", + "bbox": [ + 496, + 311, + 890, + 386 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {G D} V (G D) = \\frac {1}{2} \\mathbb {E} _ {x \\sim p _ {\\text {d a t a}} (x)} [ (G D (x) - b) ^ {2} ] \\tag {21} \\\\ + \\frac {1}{2} \\mathbb {E} _ {z \\sim p _ {z} (z)} [ (G D (M _ {E} \\odot G (z)) - a) ^ {2} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 396, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G} V (G) = \\frac {1}{2} \\mathbb {E} _ {z \\sim p _ {z} (z)} [ (G D (M _ {E} \\odot G (z)) - c) ^ {2} ], \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 479, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $GD$ denotes the global alignment consistency discriminator.", + "bbox": [ + 496, + 518, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, shadow extraction, scene layout expansion, and layout-to-image module are end-to-end trained together, thus the visual authenticity loss may facilitate shadow representation learning. Therefore, the final objective function of our approach is:", + "bbox": [ + 496, + 549, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {1} L _ {t} + \\lambda_ {2} L _ {L - G A N} ^ {G} + \\lambda_ {3} L _ {G - G A N} ^ {G}, \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 636, + 890, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1},\\lambda_{2}$ and $\\lambda_{3}$ denote the balancing weights. $L_{t}$ denotes the loss adopted from the base model. $L_{L - GAN}^{G}$ and $L_{G - GAN}^{G}$ denote the alignment loss from the local alignment and global consistency.", + "bbox": [ + 496, + 665, + 890, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 739, + 720, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section provides details of the training settings and module implementations.", + "bbox": [ + 496, + 763, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Settings. Our approach is implemented using PyTorch [46] on four NVIDIA-RTX A6000 GPUs. During the training process, we adopt the Adam optimizer, 0.001 for the local discriminator and the global discriminator, and 0.01 for all modules in shadow representation extractors. The batch size is set to 32 and the size of the input images is $256 \\times 256$ . The training is manually stopped.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7854", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shadow Detection Models. We adopt the popular shadow segmentation method [77] as the shadow detector, which is able to segment the shadow areas in the image. Furthermore, we adopt the state-of-the-art instance shadow detection method [63] as our instance shadow detector, which is able to detect the shadow instances and their corresponding object in the image. However, it is unable to detect shadows cast by unobserved objects.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alignment Discriminator. We adopt diffusion timestep-dependent discriminators from Diffusion-GAN [65]. It not only can minimize the divergence between real and diffused data at the end of the process, but also minimizes the divergence between the diffused real data distribution and the diffused generator distribution over several timesteps. In training, the balancing weight $\\lambda_{1}$ is set to be 0.5, $\\lambda_{2}$ and $\\lambda_{3}$ are set to 0.25. Note that the architecture of the discriminator can be adjusted based on the specific type of layout-to-image module being implemented.", + "bbox": [ + 75, + 212, + 468, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Layout Expansion. We adopt GTwE [11] to predict the categories and layout of unobserved instances simultaneously. GTwE has the same aforementioned hyperparameters as [69], which are the attention hidden size $d_{atten} = 512$ , feed forward size $d_{ff} = 2048$ , multi-head number $n_{head} = 4$ , and dropout $= 0.1$ .", + "bbox": [ + 75, + 362, + 467, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Layout to Image. We adopt LayoutDiffusion [76] for better outpainting the unknown areas with the condition of layouts predicted by GTwE. Note that LayoutDiffusion can not take the visible semantics as guidance to outpaint the unknown areas. Thus, we follow RePaint [38] to modify the standard denoising process for conditioning on the given image content. In each step, we sample the known region from the input and the outpaint painted part from the DDPM [18] output.", + "bbox": [ + 75, + 454, + 467, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 602, + 207, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we provide our experimental setups and both the objective and subjective results. Since we are not aware of any existing work that performs the same task as we do, we mainly focus on showing the promise of the proposed approach. Specifically, we evaluate the effectiveness of our approach by conducting experiments with an image outpainting pipeline proposed by [69]. Our goal is, again, to show the possibility of learning effective shadow representations thus enhancing complex scene extension, rather than trying to beat the state-of-the-art shadow detection and outpainting method. Other modules with the same functionality, provided they are end-to-end trainable, can be adopted in our approach to achieve potentially better performance.", + "bbox": [ + 75, + 627, + 467, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 832, + 178, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed approach on two common datasets, Visual Genome (VG) [29] and COCO-stuff [4], which gives an adequate amount of scene layouts.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Genome [29] dataset collects 108,077 images with dense annotations of objects, attributes, and relationships. Following the setting of SG2Im [20], we divide the data into $80\\%$ , $10\\%$ , and $10\\%$ for the train, val, and test set, respectively. We select the object and relationship categories occurring at least 2000 and 500 times in the train set, respectively, leaving 178 objects and 45 relationship types, and select the images with 3 to 30 bounding boxes and ignoring all small objects. Finally, the training/validation/test set will have 62565 / 5062 / 5096 images, respectively.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "COCO-stuff [4] dataset augments a subset of the COCO dataset [35] with additional semantic categories. Thus, a total of 80 object categories (car, dog, etc.) and 91 semantic categories (sky, snow, etc.) are available, with $118\\mathrm{K} / 5\\mathrm{K}$ annotated images for training/validation. Following the setting of SG2Im [20], the ground truth object coordinates in the images are utilized for constructing the synthetic relationships, a total of six relationships are considered: left of, right of, above, below, inside, and surrounding. In addition, we use images in the train and val set with 3 to 8 objects that cover more than $2\\%$ of the image. Finally, there are 25,210 train and 3,097 val images.", + "bbox": [ + 496, + 250, + 890, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Quantitative and Qualitative Result", + "text_level": 1, + "bbox": [ + 498, + 467, + 808, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the quantitative and qualitative effectiveness of our proposed approach to the image outpainting task by adopting it as a plug-in module of the SOTA method SGT [69]. The outpainting performance can be divided into two steps, in which the first one focuses on predicting the category and layout of the unobserved semantic and then the second one aims to generate the outpainting images. Therefore, we evaluate our approach on both the two steps. As can be seen from Tab. 1, by adopting our approach as a plug-in module, SGT achieves significant improvement in unobserved semantic prediction and its relevant relationship estimation. Although our target is to enhance only the semantic prediction, the accurately predicted semantics consequently strengthen the relationship estimation, thus leading to overall improvement on all metrics. Then, as can be seen from Tab. 2, we achieve considerable improvement in unobserved layout prediction, which also explicitly impacts the alignment with visible semantics in addition to the categories of instances. Moreover, in order to evaluate the quality of outpainting images, we introduce the FID metric and the results shown in Tab. 3.", + "bbox": [ + 496, + 500, + 890, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, the visual results are shown in Fig. 5 and Fig. 6, it can be found that our proposed approach is able to enhance the outpainting in the unobserved area, which is achieved by constraining both the shape and the semantic of the outpainting to be aligned with its visible shadows.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7855", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/507c36b1fcd13bc8df4ee74a00f80144d6730383d156ee0c520f1ddf9c3fe9b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 176, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c8196d16f2b46be3ec0eeb4624f8b54332722bf9e5a137524fe7c1a8343e2b8.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 80, + 165, + 176, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f43c32b21a02dfa364d2470a0fbfab75b095fbdbee165a87f0e401404c6fe795.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 88, + 279, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fa230a843ace6fc2f17ef0ecc99a222c9218f4bcf3b49c3d4ea1732e1d19d95.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 181, + 165, + 279, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69c261e30788716cbe0e14c52e8c77001baf68dceec6ba74d0df9906f4237f38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 88, + 379, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8077ce50145808fd8a42b65b9c0dab38b19693f9ba4e41420b402b657d2a441c.jpg", + "image_caption": [ + "SGT" + ], + "image_footnote": [], + "bbox": [ + 281, + 165, + 379, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04c7263528c87473ffd15ed9a75e12714af8b43c8202a941d3e57692bbee9b5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 88, + 480, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3233d51603374a77d8811d016e9ed6b62a9b7ef4e192b55eeb17255f1d1454a7.jpg", + "image_caption": [ + "SGT+" + ], + "image_footnote": [], + "bbox": [ + 382, + 165, + 480, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec9ef1f53292cf5b6da5472d7de385e6f737a2417e199f8229ae1c54bd0aebfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 88, + 581, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b635e0cdf52dce6a52db8b709ebbd1e9f788c66c9312176bd9a818bb319b4d50.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 482, + 165, + 581, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/532f11abfa6a50293b267351d27d21a785b1a6750caffdffd81743176170b5f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 88, + 683, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da36c49a84075952fdbe4373289d492c0d25b076f4a61e0bd25da2a4fee4aeb0.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 584, + 165, + 683, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2cadabe0f39a46fd72b415d45d6501c20348df6bf45b3e5f2a52c90df58f99ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 88, + 785, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1336f4d0364d8f32870075f36565c941a1ace692e719e68c9e27941dd01ded90.jpg", + "image_caption": [ + "SGT" + ], + "image_footnote": [], + "bbox": [ + 686, + 165, + 785, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/428b2aa356eb6e5ba77e39b097dd45826fc3dff01d53813f4dc3f2d29e55e53b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 88, + 885, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5dd73909fb4fedfc21c49b21081441f4cd044f8958cb0ad1971188edddad578.jpg", + "image_caption": [ + "SGT+" + ], + "image_footnote": [], + "bbox": [ + 787, + 165, + 885, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b23e861f38b84cb2780da119efff16778bada98d4207be19d8ebe53d3d4f6e5a.jpg", + "image_caption": [ + "Figure 5. Visual results from SGT [69] on VG [29]. SGT+ denotes the result after adopting our approach as a plug-in module." + ], + "image_footnote": [], + "bbox": [ + 78, + 282, + 178, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87fe6d6696707b80d73c7896e3086d65be361b89004cbfb933a6d6b61bde8e31.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 78, + 361, + 178, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f7149815bd580316a0b206745b418ff03c326c0c59a20dad678950cd6eebb70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 282, + 279, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98c49d5d9b7ff32ae6c918ab48e23256daae9d6c0700383b5d7ad5724e13547e.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 181, + 361, + 279, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/027acb619e653579a32eafa7585b2ff2c2d11b2f364ce0cdeb8d69d3274be6af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 282, + 380, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b3dd7c59f6a32de1309e451a8203faadd1f0aabfade41ad7f767ecd8dd7214c.jpg", + "image_caption": [ + "SGT" + ], + "image_footnote": [], + "bbox": [ + 281, + 361, + 380, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/445068fe0bf90dde40aa9b1e3b67038c15dce848a0ad8f373b46528055481804.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 282, + 480, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4fc3f40336c634df678f719af7420bbbc259c9a5fca6652f9fd5465c33bd42f4.jpg", + "image_caption": [ + "SGT+" + ], + "image_footnote": [], + "bbox": [ + 382, + 361, + 480, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e108779d02e4474c5923be22214abca44dfbba988831b0ae5c695a2175dfa183.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 282, + 581, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b8da01b975ac40f746ecee93cafb33516c0cba21fb14a2c05898f6e32ab420a.jpg", + "image_caption": [ + "Input Image", + "Figure 6. Visual results from SGT [69] on COCO-stuff [4]. SGT+ denotes the result after adopting our approach as a plug-in module." + ], + "image_footnote": [], + "bbox": [ + 482, + 361, + 581, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11599049e0268ffa66c264212da71c11d14f78fdc5af01842d27a85f18fcac8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 282, + 683, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf40b14a6e16ac30b44cb143f9ff9083caafab3d745d5d1356439fd3200d3912.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 584, + 361, + 683, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a75af3b85a5c88b27f36b5946e3450eeddc31eba6d7a8d80e53417f94478f9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 282, + 785, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4025e65f0e0ffd1af6154f94e4232d2311c5be23bad8a6b4a361904a6110c881.jpg", + "image_caption": [ + "SGT" + ], + "image_footnote": [], + "bbox": [ + 686, + 361, + 785, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25354537f43adf0eb68d0a2045ffb0617b034243cf88693b7c2ed78c34ce3158.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 282, + 885, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34b2061649fd74bc43727a5fffc7a4d09d3256da8d4a73eb6f5c4b760c89acd1.jpg", + "image_caption": [ + "SGT+" + ], + "image_footnote": [], + "bbox": [ + 787, + 361, + 885, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Multiple Instances Extension Comparison", + "text_level": 1, + "bbox": [ + 76, + 479, + 434, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify whether our approach can enhance alignment with visible semantics in complex scenes, even with fewer shadows. We increase the number of instances requiring inference from unobserved areas to validate alignment between the extension and visible semantics shown in Fig 7.", + "bbox": [ + 76, + 507, + 468, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bae14e966ee2dce36cb9bdc43beefdf271106b9c9c46b7fb84d7b40286e3eb86.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 78, + 592, + 169, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2168582660cdf738a174704976c9673350d3459bb8a92954094a6fa960d47f0f.jpg", + "image_caption": [ + "the number of instances increasing" + ], + "image_footnote": [], + "bbox": [ + 181, + 592, + 272, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e548ccc6727a76f4357bf38c6e44fef3fe66245475fea6b059cb4c27d33bfd68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 592, + 370, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0a931508161131605b23e1103a3ba04c059821f901a7ede1f85a1983c1dd07f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 592, + 467, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dd784e6cca593a8780cb53d15a002c3e52885add9e0ade594f3fbbb7b2ed178.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 80, + 674, + 169, + 742 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acf32a73f6381223a58064831bbcade8710a7c8f2083b865b433f162d62cc2c5.jpg", + "image_caption": [ + "the number of instances increasing" + ], + "image_footnote": [], + "bbox": [ + 181, + 674, + 274, + 743 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c08bb8509a845af1422592eb1902b07bab48279a46365177bd6f349b03396401.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 674, + 370, + 743 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44fd4c4c432db2f9cb07fc519f42ac0a36fe004768326725e2756d020a848963.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of multiple instances extension. Each image adds one more instance from left to right." + ], + "image_footnote": [], + "bbox": [ + 377, + 674, + 467, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we can see, our method can effectively leverage shadow information in complex scenes to infer a scene layout that aligns well with the visible semantics. Specifically, instances that its shadow is not in the visible area can still align with visible semantics compared to the method without incorporating our approach, which ensures a harmonious generation of content with visible semantics.", + "bbox": [ + 76, + 795, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Shadow Extension Comparison", + "text_level": 1, + "bbox": [ + 500, + 479, + 774, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As seen in Fig. 5 and Fig. 6, our visual results demonstrate that outpainting with our module achieves excellent alignment with visible semantics. However, there is an additional issue we need to consider: the shadows of visible instances can also affect the overall shadow consistency of the image. Therefore, we have also conducted a comparison specifically addressing this concern, as shown in Fig. 8.", + "bbox": [ + 498, + 503, + 892, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/582fe5eedb59f32383079296e69635e6b5df01697dcfdf72df7667594136a855.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 616, + 596, + 690 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04e70a44dea66bf9fd6572afaa280f5252efb0d22468bc97a3dc9cc2d9b86a38.jpg", + "image_caption": [ + "Input Image", + "Figure 8. Qualitative comparison of shadow extension on the visible instance. Red boxes show the differences in the extended shadow of the visible instance." + ], + "image_footnote": [], + "bbox": [ + 501, + 691, + 591, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b254c80040bed00a89c540b014e2b1bf421b37ecc6c0115df40394551ac70c3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 616, + 692, + 690 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6170703cad1030b41ce417daca3951292be7335d3a02b1e67cd3a494b2f32cc.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 598, + 691, + 692, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7ec74152898e0c99ac9dbe70f7514fadec60ec6cfd4db5cd62e2c35e9e9c770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 616, + 790, + 690 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/481d9396ad31da0b6a3f649e651c6b3c48fb956af5d96965c5628913ec16232f.jpg", + "image_caption": [ + "SGT" + ], + "image_footnote": [], + "bbox": [ + 696, + 691, + 789, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a69c1834585ff084ff8246727b99f0238787051ee1dc2de7c7224c098960e55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 616, + 885, + 690 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b508fc767830ffeaa28ec4cfd59e63dd13666870b3014468fc6ca0d3ec81bc3a.jpg", + "image_caption": [ + "SGT+" + ], + "image_footnote": [], + "bbox": [ + 792, + 691, + 885, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we can see, our shadow extension harmonizes the generated with the visible semantics. However, the method without incorporating our module randomly elongates or abruptly truncates shadows, resulting in a lack of global alignment among the semantic-shadow pairs in the scene.", + "bbox": [ + 498, + 825, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7856", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4f4d18e46029a3eacc59f5c5a8df48fe70ca12fa75c86233dfd75dac7a81c7a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VG [29]COCO-stuff [4]
ObjectRelationshipObjectRelationship
rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑
SGT10.5122.9 / 45.75.8136.9 / 69.410.5426.4 / 52.32.9225.7 / 66.4
+Fsp9.6124.6 / 47.05.2337.7 / 72.410.0127.0 / 53.42.8925.9 / 67.1
+Fillum9.2526.4 / 49.34.8238.2 / 73.19.5427.5 / 54.42.8227.0 / 67.2
+LGAN10.123.1 / 46.55.2437.4 / 71.310.4126.3 / 53.12.9825.9 / 66.9
+LGAN9.9723.4 / 46.25.4237.2 / 71.710.7526.7 / 52.92.9326.3 / 66.5
SGT+8.9327.9 / 50.44.5939.4 / 75.28.8328.4 / 55.32.7328.3 / 68.4
", + "bbox": [ + 200, + 88, + 766, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/77df863b8edd4217c8ac283fa2aba048c0f1ff20208ac848775c2843dbddcd27.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation and ablation study of scene graph expansion on VG [29] and COCO-stuff [4] datasets. $\\uparrow$ denotes the higher is better, $\\downarrow$ the lower is better. SGT+ denotes the pipeline that adopts our approach as a plug-in module." + ], + "table_footnote": [ + "Table 2. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with mIoU metric." + ], + "table_body": "
mIoUVGCOCO-stuff
SGT22.732.6
+Fsp25.434.8
+Fillum26.436.2
+LG-LGAN23.633.4
+LG-GAN24.933.1
SGT+28.438.1
", + "bbox": [ + 81, + 306, + 274, + 452 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3248e33f61055a003b6a534dffc3abc3c27882791153319c94c4c76f3db26265.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with FID metric." + ], + "table_body": "
FIDVGCOCO-stuff
SGT20.1219.12
+Fsp19.8919.02
+Fillum19.9218.93
+LG-LGAN19.4818.60
+LG-GAN19.5218.59
SGT+19.4718.38
", + "bbox": [ + 279, + 306, + 475, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. User Studies & Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 541, + 349, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Studies. For image outpainting, any appropriate generation should be treated to be correct, thus without a specific target output. Therefore, in order to further evaluate the qualitative effectiveness of our approach, we conduct two user studies, where 91 users are involved to evaluate the outpainting correctness of our approach. In the first experiment, we send each user 20 randomized selected pairs of contents, in which the first one is the observed image with filling black in the unobserved area and the second one is the scene layout we predict to appear in the unobserved area, we then ask the users whether the predicted semantic could appear in the unobserved area. Finally, $89\\%$ of the selected pairs are scored to be true, which denotes that our proposed approach is able to predict the appropriate scene layout for the unobserved semantics.", + "bbox": [ + 75, + 566, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the second experiment, we send each user 20 randomized selected pairs of images, where one of them comes from SGT and the other one comes from adopting our approach as a plug-in module, we then ask the users which one is better in visual authenticity. Finally, images from ours achieve $87\\%$ better chosen, which shows the significant quality enhancement produced by our approach.", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. We evaluate the individual effectiveness of the semantic removal, the illumination concatenation, the local alignment discriminator, and the global consistency discriminator in this experiment. The results show in Tab. 1, Tab. 2 and Tab. 3. It can be seen that the most significant enhancement in scene layout prediction comes from illumination concatenation, which shows the importance of incorporating a comprehensive illumination understanding to tackle the challenges of shadow extraction for obtaining accurate predictions. Meanwhile, the semantic removal enhances the accuracy notably, which shows that the shadow-irrelevant semantic information in the shadow areas leads to unexpected and noisy shadow representations. And the two discriminators achieve considerable enhancement in FID, which shows the importance of ensuring the local alignment between the unobserved semantics and its visible shadows and the consistency among all semantic-shadow pairs.", + "bbox": [ + 496, + 309, + 893, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 583, + 619, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel approach that aims to perform image outpainting by utilizing shadow information, which has never been explored. This is accomplished by extracting instance-level shadows cast by the unobserved areas from the input image. Then, the instance-level shadow representations are concatenated to infer the scene layout of each instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. Subsequently, extensive objective and subjective experiments are conducted which strongly proves the proposed approach successfully captures semantic and geometry information in the shadows. In future work, we will endeavor to learn shadow representations from dynamic environments and incorporate more tasks like novel view synthesis and pedestrian trajectory prediction into the unified framework.", + "bbox": [ + 496, + 609, + 893, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement The research was supported by the Shanghai Committee of Science and Technology, China (Grant No.23ZR1423500).", + "bbox": [ + 498, + 858, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7857", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrehim Ahmed and Aly Farag. Shape from shading for hybrid surfaces. In 2007 IEEE International Conference on Image Processing, pages II-525. IEEE, 2007. 2", + "[2] Abdelrehim H Ahmed and Aly A Farag. Shape from shading under various imaging conditions. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2", + "[3] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 2", + "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6, 7, 8", + "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2", + "[6] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 2", + "[7] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 1", + "[8] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(9): 10850-10869, 2023. 2", + "[9] R. Cucchiara, C. Grana, M. Piccardi, and A. Prati. Detecting moving objects, ghosts, and shadows in video streams. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(10):1337-1342, 2003. 2", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[11] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 5, 6", + "[12] Aleksandrs Ecins, Cornelia Fermüller, and Yiannis Aloimonos. Shadow free segmentation in still images using local density measure. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-8, 2014. 2", + "[13] Jiacheng Fan, Min Chen, Jinqiu Mo, Shigang Wang, and Qinghua Liang. Variational formulation of a hybrid perspective shape from shading model. The Visual Computer, pages 1-14, 2022. 2", + "[14] Silvano Galliani, Yong Chul Ju, Michael Breuß, and Andrés Bruhn. Generalised perspective shape from shading in spherical coordinates. In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 222-233. Springer, 2013. 2", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[16] Yinsai Guo, Hang Yu, Liyan Ma, Liang Zeng, and Xiangfeng Luo. Thfe: A triple-hierarchy feature enhancement method for tiny boat detection. Engineering Applications of Artificial Intelligence, 123:106271, 2023. 2", + "[17] Yinsai Guo, Hang Yu, Shaorong Xie, Liyan Ma, Xinzhi Cao, and Xiangfeng Luo. Dsca: A dual semantic correlation alignment method for domain adaptation object detection. Pattern Recognition, page 110329, 2024. 2", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 6", + "[19] Weiqiang Jin, Biao Zhao, Hang Yu, Xi Tao, Ruiping Yin, and Guizhong Liu. Improving embedded knowledge graph multi-hop question answering by introducing relational chain reasoning. Data Mining and Knowledge Discovery, 37(1):255-288, 2023. 2", + "[20] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 6", + "[21] Imran N. Junejo and Hassan Foroosh. Estimating geotemporal location of stationary cameras using shadow trajectories. In Computer Vision – ECCV 2008, pages 318–331, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2", + "[22] Asaf Karnieli, Ohad Fried, and Yacov Hel-Or. Deepshadow: Neural shape from shadow. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part II, pages 415–430. Springer, 2022. 1, 2", + "[23] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2", + "[24] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2", + "[25] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4019–4028, 2021. 4", + "[26] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, 2021. 1, 2", + "[27] Jihyun Kim, Seong-Hun Jeong, Kyeongbo Kong, and SukJu Kang. An unified framework for language guided image completion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2568-2578, 2023. 2", + "[28] Daehyeon Kong, Kyeongbo Kong, Kyunghun Kim, Sung-Jun Min, and Suk-Ju Kang. Image-adaptive hint generation" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7858", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "via vision transformer for outpainting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3572-3581, 2022. 2", + "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7, 8", + "[30] Jean-François Lalonde, Alexei A. Efros, and Srinivasa G. Narasimhan. Estimating natural illumination from a single outdoor image. In 2009 IEEE 12th International Conference on Computer Vision, pages 183-190, 2009. 2", + "[31] Kyoung Mu Lee and C-C Jay Kuo. Shape from shading with a generalized reflectance map model. Computer vision and image understanding, 67(2):143-160, 1997. 2", + "[32] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16221-16230, 2022. 1, 2", + "[33] Pengbo Li, Hang Yu, Xiangfeng Luo, and Jia Wu. Lgmnn: A local and global aware memory-based graph neural network for fraud detection. IEEE Transactions on Big Data, 2023. 2", + "[34] Jian Liang, Chenfei Wu, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. Advances in Neural Information Processing Systems, 35:15420-15432, 2022. 2", + "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[36] Ruoshi Liu, Sachit Menon, Chengzhi Mao, Dennis Park, Simon Stent, and Carl Vondrick. What you can reconstruct from a shadow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17059-17068, 2023. 1, 2", + "[37] Chia-Ni Lu, Ya-Chu Chang, and Wei-Chen Chiu. Bridging the visual gap: Wide-range image blending. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-851, 2021. 2", + "[38] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6", + "[39] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1680-1691, 2023. 3", + "[40] Pascal Mamassian. Shape from Shadows, pages 724-725. Springer, 2014. 2" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] S. Nadimi and B. Bhanu. Physical models for moving shadow and object detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(8):1079-1087, 2004. 2", + "[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[43] Takahiro Okabe, Imari Sato, and Yoichi Sato. Attached shadow coding: Estimating surface normals from shadows under unknown reflectance and lighting conditions. In 2009 IEEE 12th International Conference on Computer Vision, pages 1693-1700. IEEE, 2009. 2", + "[44] Alexandros Panagopoulos, Dimitris Samaras, and Nikos Paragios. Robust shadow and illumination estimation using a mixture model. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 651-658, 2009.", + "[45] Alexandros Panagopoulos, Chaohui Wang, Dimitris Samaras, and Nikos Paragios. Simultaneous cast shadows, illumination and geometry inference using hypergraphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(2):437-449, 2013. 2", + "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[47] Jiayan Qiu, Xinchao Wang, Pascal Fua, and Dacheng Tao. Matching seqlets: An unsupervised approach for locality preserving sequence matching. IEEE transactions on pattern analysis and machine intelligence, 43(2):745-752, 2019. 2", + "[48] Jiayan Qiu, Xinchao Wang, Stephen J Maybank, and Dacheng Tao. World from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8493-8504, 2019. 2", + "[49] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Hallucinating visual instances in total absentia. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 264-282. Springer, 2020. 2", + "[50] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Scene essence. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8322-8333, 2021. 2", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[52] Mark Sabini and Gili Rusak. Painting outside the box: Image outpainting with gans. arXiv preprint arXiv:1808.08483, 2018.2", + "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7859", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[54] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2", + "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2", + "[56] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163, 2021. 2", + "[57] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163. IEEE, 2021. 1", + "[58] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 1", + "[59] Pinzhuo Tian and Hang Yu. Can we improve meta-learning model in few-shot learning by aligning data distributions? Knowledge-Based Systems, 277:110800, 2023. 2", + "[60] Kushagra Tiwary, Tzofi Klinghoffer, and Ramesh Raskar. Towards learning neural representations from shadows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 300-316. Springer, 2022. 1", + "[61] Anwaar Ulhaq, Naveed Akhtar, and Ganna Pogrebna. Efficient diffusion models for vision: A survey. arXiv preprint arXiv:2210.09292, 2022. 2", + "[62] Guohui Wang and Jin Cheng. Three-dimensional reconstruction of hybrid surfaces using perspective shape from shading. Optik, 127(19):7740-7751, 2016. 2", + "[63] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–14, 2022. 3, 5, 6", + "[64] Yi Wang, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Wide-context semantic image extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1399-1408, 2019. 2", + "[65] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-GAN: Training GANs with diffusion. In The Eleventh International Conference on Learning Representations, 2023. 6", + "[66] Chenyu Wu, Srinivasa G Narasimhan, and Branislav Jaramaz. A multi-image shape-from-shading framework for near-lighting perspective endoscopes. International Journal of Computer Vision, 86:211-228, 2010. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Lin Wu and Xiaochun Cao. Geo-location estimation from two shadow trajectories. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 585-590, 2010. 2", + "[68] Xiaoyu Xu, Jiayan Qiu, Xinchao Wang, and Zhou Wang. Relationship spatialization for depth estimation. In European Conference on Computer Vision, pages 615-637. Springer, 2022. 2", + "[69] Chiao-An Yang, Cheng-Yo Tan, Wan-Cyuan Fan, Cheng-Fu Yang, Meng-Lin Wu, and Yu-Chiang Frank Wang. Scene graph expansion for semantics-guided image outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15626, 2022. 1, 2, 6, 7", + "[70] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Yingxia Shao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2", + "[71] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv., 2023. Just Accepted. 2", + "[72] Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. Distilling knowledge from graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7074-7083, 2020. 2", + "[73] Zongxin Yang, Jian Dong, Ping Liu, Yi Yang, and Shuicheng Yan. Very long natural scenery image prediction by outpainting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10560-10569, 2019. 2", + "[74] Kai Yao, Penglei Gao, Xi Yang, Jie Sun, Rui Zhang, and Kaizhu Huang. Outpainting by queries. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 1, 2", + "[75] Guangcong Zheng, Shengming Li, Hui Wang, Taiping Yao, Yang Chen, Shouhong Ding, and Xi Li. Entropy-driven sampling and training scheme for conditional diffusion generation. In European Conference on Computer Vision, pages 754-769. Springer, 2022. 2", + "[76] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 5, 6", + "[77] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 824 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7860", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_model.json b/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_model.json new file mode 100644 index 0000000000000000000000000000000000000000..026c217553e98a3617690de7674753d3cd3a8be1 --- /dev/null +++ b/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_model.json @@ -0,0 +1,3412 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.279, + 0.131, + 0.692, + 0.154 + ], + "angle": 0, + "content": "Shadow-Enlightened Image Outpainting" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.18, + 0.854, + 0.273 + ], + "angle": 0, + "content": "Hang Yu\\(^{1}\\), Ruilin Li\\(^{2}\\), Shaorong Xie\\(^{1}\\), Jiayan Qiu\\(^{3*}\\) \n\\(^{1}\\)School of Computer Engineering and Science, Shanghai University, China \n\\(^{2}\\)Institute of Artificial Intelligence, Shanghai University, China \n\\(^{3}\\)School of Computing and Mathematical Sciences, University of Leicester, United Kingdom {yuhang, ruilinli, srxie}@shu.edu.cn, jiayan.qiu.1991@outlook.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.641 + ], + "angle": 0, + "content": "Conventional image outpainting methods usually treat unobserved areas as unknown and extend the scene only in terms of semantic consistency, thus overlooking the hidden information in shadows cast by unobserved areas, such as the invisible shapes and semantics. In this paper, we propose to extract and utilize the hidden information of unobserved areas from their shadows to enhance image outpainting. To this end, we propose an end-to-end deep approach that explicitly looks into the shadows within the image. Specifically, we extract shadows from the input image and identify instance-level shadow regions cast by the unobserved areas. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. In the experiments, we show that our proposed approach provides complementary cues for outpainting and achieves considerable improvement on all datasets by adopting our approach as a plug-in module." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.21, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Given an image of a complex scene, as shown in Fig. 1(a), what will its right-side unobserved area look like? Humans can, without much effort, infer the appropriate categories and rough shapes of invisible semantics (e.g. objects) and further imagine an adequately realistic overall appearance. One critical cue humans utilize to reason invisible semantics is the shadows cast in the image. In this regard, we identify the shadows of each visible object, on top of which the shadows cast by the unobserved areas are found. According to their shapes and relative spatial layout, we may reasonably infer that there is a person in the unobserved area, as shown in Fig. 1(b), which aligns with its" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.305, + 0.889, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.54, + 0.895, + 0.651 + ], + "angle": 0, + "content": "Figure 1. Given the image of a complex scene in (a) for outpainting, the extension needs to align with visible semantics. For example, the result shown in (b) is more aligned with the visible semantics from a global shadowing direction compared to the result shown in (c). For better alignment, we overcome the challenges of shadow extraction in complex scenes, such as disjoined shadows (d), overlapped shadows (e), and shadows attaching unexpected semantic information (f), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.747 + ], + "angle": 0, + "content": "shadow and the given scene. However, existing outpainting methods [7, 26, 57, 58, 69, 74] overlook shadows in the image, leading to inferior results, as shown in Fig. 1(c), where misalignment exists between the extension and its shadows from the shadowing direction in the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Even though prior efforts on Shape-from-Shadow [22, 32, 36, 60] also aim to extract hidden information, such as the shapes, of the semantics from their shadows in simple scenes with an ideal light source and a given semantic category, they cannot perform complex image outpainting with unknown semantics, due to there are three challenges of shadow extraction in complex scenes. 1) Given the complex spatial layout of the instances, the shadows are usually occluded by other instances, resulting in disjoined shadows of a single object, as shown in Fig. 1(d), which leads to inaccu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7850" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.212 + ], + "angle": 0, + "content": "rate perception or misinterpretation of invisible semantics. 2) Given the complex illumination of the scene, the shadows cast by different semantics may overlap, thus causing information loss for the unobserved area, as shown in Fig. 1(e). 3) Given the complex spatial layout of the scene, the shadows usually contain semantic information about the planes on which they are cast, which may lead to shadow-irrelevant understanding, as shown in Fig. 1(f)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.468, + 0.303 + ], + "angle": 0, + "content": "To perform image outpainting by utilizing the hidden information in the shadows, we propose a novel approach that explicitly looks into the correlation between shadows from the unobserved area and the visible semantics, the shadows with semantic removal, and the consistency between local and global alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.483 + ], + "angle": 0, + "content": "Specifically, as depicted in Fig. 2, given an input image, we first conduct shadow extraction to obtain those shadow regions belonging to unobserved instances. Then, we utilize the shadow representations from unobserved instances to predict the scene layout of these instances, and then concatenate them with the shadow representation of instances in unobserved areas for layout-to-image processing. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas for the outpainting image. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.468, + 0.635 + ], + "angle": 0, + "content": "Our contribution is therefore a novel framework designated to achieve image outpainting by utilizing the shadows, which is the first attempt to the best of our knowledge. This is accomplished by extracting the desired shadow information, followed by scene layout expansion and layout-to-image conversion to produce image outpainting. The whole pipeline is end-to-end trainable. By adopting our proposed framework as a plug-in module, the outpainting methods [69] have achieved considerable quantitative and qualitative performance improvement on all datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.218, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.704 + ], + "angle": 0, + "content": "We briefly review the prior works related to ours, including shadow processing and image outpainting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Shadow Processing. Conventional shadow processing method reconstructs the surface shape of a single object, under ideal illuminance, from its shadow and visual information [40]. Then, a sequence of methods [1, 2, 13, 14, 31, 62, 66] is developed to improve its performance by involving carefully designed prior knowledge in the model. Then, due to the strong adaptability of deep learning, learning-based methods [22, 32] are proposed to deal with complicated environments, such as objects with non-Lambertian reflectance [32]. Moreover, shadows are utilized as cues for scene geometry information extraction [21, 23, 30, 36, 43-45, 47, 48, 67, 68], segmentation [12], object detection [9, 16, 17], and tracking [41]. Recently, a generative method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "[36] has been proposed to predict the shape of a single invisible object, with a specific category, from its shadow. However, none of the existing methods can infer the scene layout of the semantics from their shadows in complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.153, + 0.892, + 0.5 + ], + "angle": 0, + "content": "Image Outpainting. Conventional image outpainting methods extend the image with the background texture and partially observed objects towards specific directions [5, 34, 37, 52, 73, 74]. Then, information on the image margin is utilized to enable the outpainting in any direction [28, 56, 64]. Recently, the extrapolation of segmentation information has been adopted in the outpainting process [26] to add novel semantics in the extending areas. Meanwhile, the scene graph of the image is utilized to introduce new semantics by learning the semantic co-occurrence with Graph Neural Networks (GNN) [19, 33, 49, 50, 59, 69, 72]. Conventional outpainting commonly uses Generative Adversarial Networks (GAN) [15] for image conversion. However, due to the stability of diffusion, diffusion models [3, 18, 42, 51, 55] are being recognized as a promising family of generative models that have proven to be state-of-the-art sample quality for a variety of image generation benchmarks [8, 61, 70], including class-conditional image generation [10, 75], text-to-image generation [27, 51, 53, 71], image-to-image translation [24, 38, 54], layout-to-image generation [6, 76]. However, none of the existing methods takes advantage of the shadows to enhance the outpainting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.59, + 0.527 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.793 + ], + "angle": 0, + "content": "In this section, we detail the working scheme of the proposed approach which comprises three stages, as shown in Fig. 2. In Stage 1, we utilize the pretrained model of shadow detection and instance shadow detection to grain rough instance-level shadow regions. Meanwhile, a pretrained shadow removal model is to remove unexpected semantic information from the shadows for the purified shadow feature. Then, we calculate the connectivity between shadow regions to merge disjoined shadows and perform overlapped segmentation on the shadows to restore overlapped shadows of instances. After iterative optimization, we obtain refined instance-level shadow regions. In Stage 2, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. In Stage 3, two discriminators are implemented to enhance alignment between the extended semantics and their shadows." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.803, + 0.753, + 0.818 + ], + "angle": 0, + "content": "3.1. Stage 1: Shadow Extraction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.826, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Given the input image \\(I\\), we first adopt a pretrained shadow detection model [77] to extract all shadow areas:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.881, + 0.892, + 0.897 + ], + "angle": 0, + "content": "\\[\nM _ {s} = \\operatorname {D e t e c t i o n} (I), \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.865, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.428, + 0.704, + 0.443 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed approach. \\(\\bigoplus\\) denotes concatenation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.517 + ], + "angle": 0, + "content": "where \\(M_{s}\\) denotes the mask of the shadow areas, respectively. Then, a pretrained instance shadow detection model [63] is introduced to identify the shadow areas for the visible semantics in image \\(I\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.527, + 0.469, + 0.545 + ], + "angle": 0, + "content": "\\[\nM _ {s} ^ {v} = \\text {I n s t a n c e - D e t e c t i o n} (I). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.469, + 0.629 + ], + "angle": 0, + "content": "where \\( M_{s}^{v} \\) denotes the instance-level mask of the shadow areas cast by the visible semantics. Consequently, a binary area-level mask of shadows cast by the unobserved area can be obtained with \\( M_{s}^{u} = M_{s} - M_{s}^{v} \\). Thus, we obtain a rough instance-level shadow region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.691 + ], + "angle": 0, + "content": "Meanwhile, to obtain the shadow areas without the semantic information from the cast planes, we introduce a pretrained shadow removal model [39] and obtain the semantically purified shadow image \\( I_{s}^{p} \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.7, + 0.469, + 0.719 + ], + "angle": 0, + "content": "\\[\nI _ {s} ^ {p} = M _ {s} \\odot (I - \\text {S h a d o w R e m o v a l} (I)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.728, + 0.347, + 0.743 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the Hadamard product." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.743, + 0.469, + 0.787 + ], + "angle": 0, + "content": "After obtaining the shadow image of semantic removal \\( I_{s}^{p} \\), we can use a convolutional net to extract the shadow feature without semantics:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.797, + 0.469, + 0.816 + ], + "angle": 0, + "content": "\\[\nF _ {s} ^ {p} = \\operatorname {C o n v N e t} \\left(I _ {s} ^ {p}\\right), F _ {s} ^ {p} \\in \\mathbb {R} ^ {W \\times H \\times d}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where the ConvNet consists of 8 Residual Blocks and preserves the resolution of the output representation to be the same as the input image. \\( W, H \\) denotes the width and height of the input image. \\( d \\) denotes the dimension of the shadow feature map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.578 + ], + "angle": 0, + "content": "Considering that the formation of shadows is influenced by the illumination conditions in the given scene, which can be a cue in addressing the issues of disjoined shadows and overlapped shadows, we concatenate the image \\(I\\) and the shadow mask \\(M_{s}\\) and fed it into a Convolutional Neural Network (CNN) to obtain the illumination feature \\(F^{\\mathrm{illum}} \\in \\mathbb{R}^{(W / 16) \\times (H / 16) \\times d}\\) in the scene as a cue for the shadow regions optimization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.595, + 0.749, + 0.611 + ], + "angle": 0, + "content": "3.1.1 Shadow Merging Operation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.892, + 0.663 + ], + "angle": 0, + "content": "In complex scenes, due to the presence of instances that can occlude shadows, it can result in an instance having multiple disjointed shadows, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.83 + ], + "angle": 0, + "content": "To classify a single instance with disjoined shadows, we first treat each connected component in \\(M_{s}^{u}\\) as a shadow region weight mask. Let \\(M^{u} = \\{m_{x}^{u}\\} \\in \\mathbb{R}^{W\\times H\\times 1}, x\\in [1,N_{u}]\\) be the shadow weight masks that don't belong to any visible instances, where \\(N_{u}\\) denotes the number of connected components in \\(M_{s}^{u}\\). For the rough instance-level mask of visible instances in \\(M_{s}^{v}\\), let \\(M^{v} = \\{M_{y}^{v}\\} \\in \\mathbb{R}^{W\\times H\\times 1}, y\\in [1,N_{v}]\\) be each shadow weight mask of visible instance, where \\(N_{v}\\) denotes the number of visible instances. Consequently, we grain the initial instance-level shadow region \\(M\\):" + }, + { + "type": "equation", + "bbox": [ + 0.646, + 0.828, + 0.891, + 0.843 + ], + "angle": 0, + "content": "\\[\nN = N _ {u} + N _ {v} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.846, + 0.891, + 0.864 + ], + "angle": 0, + "content": "\\[\nM = \\left\\{m _ {z} \\right\\} = \\left\\{M ^ {u}, M ^ {v} \\right\\}, z \\in [ 1, N ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "For each shadow region, we extract the corresponding region-level shadow feature \\( F \\) by the feature map by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7852" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.272, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.197, + 0.25, + 0.22 + ], + "angle": 0, + "content": "(a) the disjoined shadow of the unobserved instance" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.468, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.197, + 0.436, + 0.22 + ], + "angle": 0, + "content": "(b) the disjoined shadow of the visible instance" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.334 + ], + "angle": 0, + "content": "Figure 3. Samples of disjoined shadows caused by occlusion in complex scenes. (a) The shadow of an unobserved object is obstructed, creating two disjointed shadow regions. The red boxes indicate the disjoined shadows. (b) Due to the defect of the instance shadow detector, a portion of the shadow of the instance was not classified as a single instance. The blue boxes indicate the detected corresponding pair, and the yellow box indicates the undetected shadow region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.469, + 0.424 + ], + "angle": 0, + "content": "element-wise multiplying the shadow feature without semantics \\( F_{s}^{p} \\) and the shadow weight mask \\( M \\). Then, we use the max pooling operation to resize these features to match the size of the illumination feature. Thus, we obtain the cropped feature \\( \\bar{F} \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.434, + 0.469, + 0.452 + ], + "angle": 0, + "content": "\\[\nF = F _ {s} ^ {p} \\odot M, F \\in \\mathbb {R} ^ {W \\times H \\times d}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.455, + 0.469, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\bar {F} = \\operatorname {m a x p o o l} (F), \\bar {F} \\in \\mathbb {R} ^ {(W / 1 6) \\times (H / 1 6) \\times d}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.469, + 0.59 + ], + "angle": 0, + "content": "To improve the computation of connectivity between shadow regions, we combine the cropped features \\(\\bar{F}\\) of two regions along with the illumination feature by concatenating them together. This concatenated input is fed into an encoder \\(E_{connect}\\), and then passed through a sigmoid activation function for connectivity between two shadow regions:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.601, + 0.469, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\bar {F} = \\{\\bar {f} _ {m} \\}, m \\in [ 1, N ], \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.621, + 0.469, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\bar {f} _ {(i, j)} = \\operatorname {E} _ {\\text {c o n n e c t}} \\left(\\operatorname {c o n c a t} \\left(\\bar {f} _ {i}, \\bar {f} _ {j}, F ^ {\\text {i l l u m}}\\right)\\right), i \\neq j \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.641, + 0.469, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o n n e c t i v i t y} _ {(i, j)} = \\operatorname {S i g m o i d} (\\bar {f} _ {(i, j)}), i \\neq j \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.469, + 0.759 + ], + "angle": 0, + "content": "When the connectivity between two shadow regions is higher than a hand-setting threshold \\(\\alpha_{1}\\), these two shadow regions are considered to be the shadows of the same object, accordingly merging their weight masks. Conversely, they stay unchanged. Finally, we obtain the updated shadow features \\(F^{update}\\) after merging operation:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.77, + 0.469, + 0.788 + ], + "angle": 0, + "content": "\\[\nF ^ {u p d a t e} = \\left\\{f _ {m} \\right\\} \\in \\mathbb {R} ^ {W \\times H \\times d}, m \\in [ 1, N ^ {m} ], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.799, + 0.462, + 0.814 + ], + "angle": 0, + "content": "where \\( N^{m} \\) denotes the number of instances after merging." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.327, + 0.848 + ], + "angle": 0, + "content": "3.1.2 Shadow Splitting Operation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In complex scenes, due to the complex spatial layout in the scene, it is common to have overlapped shadows between instances, as shown in Fig. 4." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.697, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.197, + 0.687, + 0.221 + ], + "angle": 0, + "content": "(a) overlapped with shadows cast by the unobserved instances" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.889, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.197, + 0.868, + 0.221 + ], + "angle": 0, + "content": "(b) overlapped with shadow cast by the visible instances" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Figure 4. Samples of overlapping shadows in complex scenes. The red boxes indicate the overlapped shadows. (a) The shadows of multiple unobserved instances overlap. (b) The shadows of unobserved instances overlap with the shadows of visible instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.424 + ], + "angle": 0, + "content": "In this work, each overlapped shadow is decoupled into two intersecting graph spaces. After concatenating the cropped feature \\(\\bar{F}\\) and illumination feature \\(F^{illum}\\), we input them into the Overlapping BiLayers [25]. Then, we grain two weight masks \\(M_0'\\), and \\(M_1'\\), each weight mask represents the shadow region of an individual instance. Then, we can obtain the \"de-overlapped\" shadow feature \\(f\\) of an instance by element-wise multiplying the weight mask \\(M'\\) with the purified shadow feature \\(F_s^p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.433, + 0.891, + 0.451 + ], + "angle": 0, + "content": "\\[\nM _ {0} ^ {\\prime}, M _ {1} ^ {\\prime}, \\in \\mathbb {R} ^ {W \\times H \\times 1} \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.454, + 0.891, + 0.471 + ], + "angle": 0, + "content": "\\[\nf _ {0} = M _ {0} ^ {\\prime} \\odot F _ {s} ^ {p}, f _ {1} = M _ {1} ^ {\\prime} \\odot F _ {s} ^ {p}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.558 + ], + "angle": 0, + "content": "when the ratio of the combined area of two weight masks to the source area is higher than a hand-setting threshold \\(\\alpha_{2}\\), it is considered an overlapped shadow and needs to be split. Conversely, it stays unchanged. Finally, we obtain the updated shadow features \\(F^{update}\\) after splitting operation:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.583, + 0.891, + 0.601 + ], + "angle": 0, + "content": "\\[\nF ^ {u p d a t e} = \\left\\{f _ {m} \\right\\} \\in \\mathbb {R} ^ {W \\times H \\times d}, m \\in [ 1, N ^ {s} ], \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.613, + 0.879, + 0.629 + ], + "angle": 0, + "content": "where \\(N^s\\) denotes the number of instances after splitting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.71, + 0.661 + ], + "angle": 0, + "content": "3.1.3 Iterative Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.791 + ], + "angle": 0, + "content": "In complex scenes, it is common to encounter situations where disjoined shadows and overlapped shadows co-occur. Therefore, we utilize an iterative combination of merging and splitting operations to optimize the categorization of all shadow regions in the image. Thus, the issues of disjoined shadows and overlapped shadows, shown in Fig. 3 and Fig. 4, are addressed. Specifically, the iterative shadow optimization is summarized as Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.799, + 0.816 + ], + "angle": 0, + "content": "3.2. Stage 2: Outpaiting with Shadows" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.824, + 0.891, + 0.869 + ], + "angle": 0, + "content": "After obtaining the refined instance-level shadow regions \\(M\\) from Stage 1, we grain the cropped shadow representations \\(\\bar{f}^u\\) of unobserved instances by:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.88, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\bar {f} ^ {u} = \\operatorname {m a x p o o l} \\left(M ^ {u} \\odot F _ {s} ^ {p}\\right), M ^ {u} \\in M, \\tag {16}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7853" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.086, + 0.095, + 0.385, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Iterative Shadow Optimization" + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.112, + 0.468, + 0.287 + ], + "angle": 0, + "content": "Input: rough instance-level shadow weight mask \nOutput: refined instance-level shadow weight mask \n1 \\(M^{(0)}\\gets M\\) \n2 for \\(t = 1\\) to \\(T\\) do \n3 \\(\\begin{array}{rl} & M^{(t)}\\leftarrow \\mathrm{ShadowMerging}(M^{(t - 1)})\\\\ & M^{(t)}\\leftarrow \\mathrm{ShadowSplitting}(M^{(t)})\\\\ & \\mathrm{if}\\| M^{(t)} - M^{(t - 1)}\\| < \\epsilon \\mathrm{then}\\\\ & \\mid \\mathrm{break}\\\\ & \\mathrm{end} \\end{array}\\) \n5 \n6 \n7 \n8 end \n9 return \\(M^{(t)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.47, + 0.351 + ], + "angle": 0, + "content": "which are concatenated to predict the categories and layouts of new instances and outpaint the unobserved areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.469, + 0.457 + ], + "angle": 0, + "content": "For scene layout expansion, given a partial input image with the partial scene graph and the corresponding layout, we use GNN [11] to reveal the scene layout \\( L_{exp} \\) of new instances. Specifically, we pass the shadow representations \\( \\tilde{f}^u \\) from the unobserved instances through an encoder \\( E_h \\), and then concatenate the output with the node features \\( f^h \\) of the masked instances from the encoder of GNN." + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.469, + 0.469, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\hat {f} ^ {h} = \\operatorname {c o n c a t} \\left(f ^ {h}, \\mathrm {E} _ {\\mathrm {h}} \\left(\\bar {f} ^ {u}\\right)\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.606 + ], + "angle": 0, + "content": "For layout to image, given an input image \\(I\\) with an expanded scene layout \\(L_{exp}\\), we use layout-to-image model [76] to perform layout to image conversion. Specifically, we pass the shadow representations \\(f^u\\) from the unobserved instances through an encoder \\(E_l\\), and then concatenate the output with scene layout features \\(f^l\\) of the expanded instance from the encoder of the layout-to-image module." + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.632, + 0.469, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\hat {f} ^ {l} = \\operatorname {c o n c a t} \\left(f ^ {l}, \\mathrm {E} _ {\\mathrm {l}} \\left(\\bar {f} ^ {u}\\right)\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.469, + 0.753 + ], + "angle": 0, + "content": "Note that we focus on the design of extracting shadow representation, so producing high-accuracy predictions for scene layout and high-quality image outputs is not within the main scope of this work. Thus, the graph-based and diffusion-based designs can be replaced with any scene layout expansion and layout-to-image modules if desirable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.375, + 0.788 + ], + "angle": 0, + "content": "3.3. Stage 3: Alignment Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In stage 2, we outpaint the partial input image \\(I\\) into the extended image \\(I_{E}\\) with introduced instances. To enhance alignment between unobserved area and visible semantics, we adopt the pretrained instance shadow detector [63] to first extract the mask \\(M_E^u\\) of the unobserved area and its corresponding shadow from \\(I_{E}\\), and then feed it into a local shadow-instance alignment discriminator to ensure the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "generated alignment is visually real. This is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.129, + 0.892, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {L D} V (L D) = \\frac {1}{2} \\mathbb {E} _ {x \\sim p _ {\\text {d a t a}} (x)} [ (L D (x) - b) ^ {2} ] \\tag {19} \\\\ + \\frac {1}{2} \\mathbb {E} _ {z \\sim p z (z)} [ (L D (M _ {E} ^ {u} \\odot G (z)) - a) ^ {2} ], \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.214, + 0.892, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\min _ {G} V (G) = \\frac {1}{2} \\mathbb {E} _ {z \\sim p z (z)} [ (L D (M _ {E} ^ {u} \\odot G (z)) - c) ^ {2} ], \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.892, + 0.312 + ], + "angle": 0, + "content": "where \\( LD \\) denotes the local alignment discriminator. \\( G \\) denotes the generator for outpainting. \\( a \\) and \\( b \\) denote the ground truth real and fake labels, respectively. \\( c \\) denotes the value that \\( G \\) wants \\( LD \\) to believe for the fake data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Then, the mask \\( M_E \\) that consists of all semantics and their shadows in \\( I_E \\) is extracted, and we feed it into a global alignment discriminator to ensure the newly generated alignment is consistent with the existing ones. This is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.397, + 0.891, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {G D} V (G D) = \\frac {1}{2} \\mathbb {E} _ {x \\sim p _ {\\text {d a t a}} (x)} [ (G D (x) - b) ^ {2} ] \\tag {21} \\\\ + \\frac {1}{2} \\mathbb {E} _ {z \\sim p _ {z} (z)} [ (G D (M _ {E} \\odot G (z)) - a) ^ {2} ], \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.481, + 0.891, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\min _ {G} V (G) = \\frac {1}{2} \\mathbb {E} _ {z \\sim p _ {z} (z)} [ (G D (M _ {E} \\odot G (z)) - c) ^ {2} ], \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.891, + 0.548 + ], + "angle": 0, + "content": "where \\(GD\\) denotes the global alignment consistency discriminator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.626 + ], + "angle": 0, + "content": "In summary, shadow extraction, scene layout expansion, and layout-to-image module are end-to-end trained together, thus the visual authenticity loss may facilitate shadow representation learning. Therefore, the final objective function of our approach is:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.637, + 0.891, + 0.655 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {1} L _ {t} + \\lambda_ {2} L _ {L - G A N} ^ {G} + \\lambda_ {3} L _ {G - G A N} ^ {G}, \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.727 + ], + "angle": 0, + "content": "where \\(\\lambda_{1},\\lambda_{2}\\) and \\(\\lambda_{3}\\) denote the balancing weights. \\(L_{t}\\) denotes the loss adopted from the base model. \\(L_{L - GAN}^{G}\\) and \\(L_{G - GAN}^{G}\\) denote the alignment loss from the local alignment and global consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.722, + 0.757 + ], + "angle": 0, + "content": "4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.795 + ], + "angle": 0, + "content": "This section provides details of the training settings and module implementations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training Settings. Our approach is implemented using PyTorch [46] on four NVIDIA-RTX A6000 GPUs. During the training process, we adopt the Adam optimizer, 0.001 for the local discriminator and the global discriminator, and 0.01 for all modules in shadow representation extractors. The batch size is set to 32 and the size of the input images is \\(256 \\times 256\\). The training is manually stopped." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7854" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Shadow Detection Models. We adopt the popular shadow segmentation method [77] as the shadow detector, which is able to segment the shadow areas in the image. Furthermore, we adopt the state-of-the-art instance shadow detection method [63] as our instance shadow detector, which is able to detect the shadow instances and their corresponding object in the image. However, it is unable to detect shadows cast by unobserved objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.363 + ], + "angle": 0, + "content": "Alignment Discriminator. We adopt diffusion timestep-dependent discriminators from Diffusion-GAN [65]. It not only can minimize the divergence between real and diffused data at the end of the process, but also minimizes the divergence between the diffused real data distribution and the diffused generator distribution over several timesteps. In training, the balancing weight \\(\\lambda_{1}\\) is set to be 0.5, \\(\\lambda_{2}\\) and \\(\\lambda_{3}\\) are set to 0.25. Note that the architecture of the discriminator can be adjusted based on the specific type of layout-to-image module being implemented." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.468, + 0.453 + ], + "angle": 0, + "content": "Scene Layout Expansion. We adopt GTwE [11] to predict the categories and layout of unobserved instances simultaneously. GTwE has the same aforementioned hyperparameters as [69], which are the attention hidden size \\( d_{atten} = 512 \\), feed forward size \\( d_{ff} = 2048 \\), multi-head number \\( n_{head} = 4 \\), and dropout \\( = 0.1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.59 + ], + "angle": 0, + "content": "Layout to Image. We adopt LayoutDiffusion [76] for better outpainting the unknown areas with the condition of layouts predicted by GTwE. Note that LayoutDiffusion can not take the visible semantics as guidance to outpaint the unknown areas. Thus, we follow RePaint [38] to modify the standard denoising process for conditioning on the given image content. In each step, we sample the known region from the input and the outpaint painted part from the DDPM [18] output." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.209, + 0.619 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.468, + 0.824 + ], + "angle": 0, + "content": "In this section, we provide our experimental setups and both the objective and subjective results. Since we are not aware of any existing work that performs the same task as we do, we mainly focus on showing the promise of the proposed approach. Specifically, we evaluate the effectiveness of our approach by conducting experiments with an image outpainting pipeline proposed by [69]. Our goal is, again, to show the possibility of learning effective shadow representations thus enhancing complex scene extension, rather than trying to beat the state-of-the-art shadow detection and outpainting method. Other modules with the same functionality, provided they are end-to-end trainable, can be adopted in our approach to achieve potentially better performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.179, + 0.847 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We evaluate our proposed approach on two common datasets, Visual Genome (VG) [29] and COCO-stuff [4], which gives an adequate amount of scene layouts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Visual Genome [29] dataset collects 108,077 images with dense annotations of objects, attributes, and relationships. Following the setting of SG2Im [20], we divide the data into \\(80\\%\\), \\(10\\%\\), and \\(10\\%\\) for the train, val, and test set, respectively. We select the object and relationship categories occurring at least 2000 and 500 times in the train set, respectively, leaving 178 objects and 45 relationship types, and select the images with 3 to 30 bounding boxes and ignoring all small objects. Finally, the training/validation/test set will have 62565 / 5062 / 5096 images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.433 + ], + "angle": 0, + "content": "COCO-stuff [4] dataset augments a subset of the COCO dataset [35] with additional semantic categories. Thus, a total of 80 object categories (car, dog, etc.) and 91 semantic categories (sky, snow, etc.) are available, with \\(118\\mathrm{K} / 5\\mathrm{K}\\) annotated images for training/validation. Following the setting of SG2Im [20], the ground truth object coordinates in the images are utilized for constructing the synthetic relationships, a total of six relationships are considered: left of, right of, above, below, inside, and surrounding. In addition, we use images in the train and val set with 3 to 8 objects that cover more than \\(2\\%\\) of the image. Finally, there are 25,210 train and 3,097 val images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.468, + 0.809, + 0.484 + ], + "angle": 0, + "content": "5.2. Quantitative and Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.816 + ], + "angle": 0, + "content": "In this section, we evaluate the quantitative and qualitative effectiveness of our proposed approach to the image outpainting task by adopting it as a plug-in module of the SOTA method SGT [69]. The outpainting performance can be divided into two steps, in which the first one focuses on predicting the category and layout of the unobserved semantic and then the second one aims to generate the outpainting images. Therefore, we evaluate our approach on both the two steps. As can be seen from Tab. 1, by adopting our approach as a plug-in module, SGT achieves significant improvement in unobserved semantic prediction and its relevant relationship estimation. Although our target is to enhance only the semantic prediction, the accurately predicted semantics consequently strengthen the relationship estimation, thus leading to overall improvement on all metrics. Then, as can be seen from Tab. 2, we achieve considerable improvement in unobserved layout prediction, which also explicitly impacts the alignment with visible semantics in addition to the categories of instances. Moreover, in order to evaluate the quality of outpainting images, we introduce the FID metric and the results shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Then, the visual results are shown in Fig. 5 and Fig. 6, it can be found that our proposed approach is able to enhance the outpainting in the unobserved area, which is achieved by constraining both the shape and the semantic of the outpainting to be aligned with its visible shadows." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7855" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.178, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.166, + 0.178, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.245, + 0.162, + 0.257 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.28, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.166, + 0.28, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.245, + 0.24, + 0.255 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.089, + 0.38, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.166, + 0.38, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.245, + 0.345, + 0.255 + ], + "angle": 0, + "content": "SGT" + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.089, + 0.481, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.166, + 0.482, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.245, + 0.45, + 0.255 + ], + "angle": 0, + "content": "SGT+" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.089, + 0.582, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.166, + 0.582, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.245, + 0.566, + 0.257 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.089, + 0.684, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.166, + 0.684, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.245, + 0.645, + 0.255 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.089, + 0.786, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.166, + 0.786, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.245, + 0.749, + 0.255 + ], + "angle": 0, + "content": "SGT" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.089, + 0.887, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.166, + 0.887, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.245, + 0.854, + 0.255 + ], + "angle": 0, + "content": "SGT+" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.257, + 0.856, + 0.271 + ], + "angle": 0, + "content": "Figure 5. Visual results from SGT [69] on VG [29]. SGT+ denotes the result after adopting our approach as a plug-in module." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.284, + 0.179, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.362, + 0.179, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.441, + 0.163, + 0.452 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.284, + 0.28, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.362, + 0.28, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.441, + 0.241, + 0.451 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.284, + 0.381, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.362, + 0.381, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.441, + 0.345, + 0.451 + ], + "angle": 0, + "content": "SGT" + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.284, + 0.481, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.362, + 0.481, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.441, + 0.45, + 0.451 + ], + "angle": 0, + "content": "SGT+" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.284, + 0.582, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.362, + 0.582, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.441, + 0.566, + 0.452 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.284, + 0.684, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.362, + 0.684, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.441, + 0.645, + 0.451 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.284, + 0.786, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.362, + 0.786, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.441, + 0.749, + 0.451 + ], + "angle": 0, + "content": "SGT" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.284, + 0.887, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.362, + 0.887, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.441, + 0.854, + 0.451 + ], + "angle": 0, + "content": "SGT+" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.453, + 0.878, + 0.467 + ], + "angle": 0, + "content": "Figure 6. Visual results from SGT [69] on COCO-stuff [4]. SGT+ denotes the result after adopting our approach as a plug-in module." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.48, + 0.436, + 0.496 + ], + "angle": 0, + "content": "5.3. Multiple Instances Extension Comparison" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.508, + 0.47, + 0.584 + ], + "angle": 0, + "content": "To verify whether our approach can enhance alignment with visible semantics in complex scenes, even with fewer shadows. We increase the number of instances requiring inference from unobserved areas to validate alignment between the extension and visible semantics shown in Fig 7." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.593, + 0.17, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.664, + 0.159, + 0.675 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.593, + 0.274, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.664, + 0.373, + 0.674 + ], + "angle": 0, + "content": "the number of instances increasing" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.593, + 0.371, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.593, + 0.468, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.675, + 0.17, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.744, + 0.134, + 0.753 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.675, + 0.275, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.745, + 0.373, + 0.755 + ], + "angle": 0, + "content": "the number of instances increasing" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.675, + 0.371, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.675, + 0.468, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.757, + 0.468, + 0.784 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of multiple instances extension. Each image adds one more instance from left to right." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "As we can see, our method can effectively leverage shadow information in complex scenes to infer a scene layout that aligns well with the visible semantics. Specifically, instances that its shadow is not in the visible area can still align with visible semantics compared to the method without incorporating our approach, which ensures a harmonious generation of content with visible semantics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.48, + 0.776, + 0.496 + ], + "angle": 0, + "content": "5.4. Shadow Extension Comparison" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.504, + 0.893, + 0.61 + ], + "angle": 0, + "content": "As seen in Fig. 5 and Fig. 6, our visual results demonstrate that outpainting with our module achieves excellent alignment with visible semantics. However, there is an additional issue we need to consider: the shadows of visible instances can also affect the overall shadow consistency of the image. Therefore, we have also conducted a comparison specifically addressing this concern, as shown in Fig. 8." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.617, + 0.597, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.692, + 0.593, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.767, + 0.583, + 0.778 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.617, + 0.694, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.692, + 0.694, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.767, + 0.657, + 0.777 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.617, + 0.791, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.692, + 0.79, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.767, + 0.756, + 0.777 + ], + "angle": 0, + "content": "SGT" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.617, + 0.887, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.692, + 0.887, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.767, + 0.856, + 0.777 + ], + "angle": 0, + "content": "SGT+" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison of shadow extension on the visible instance. Red boxes show the differences in the extended shadow of the visible instance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "As we can see, our shadow extension harmonizes the generated with the visible semantics. However, the method without incorporating our module randomly elongates or abruptly truncates shadows, resulting in a lack of global alignment among the semantic-shadow pairs in the scene." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7856" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.202, + 0.089, + 0.767, + 0.252 + ], + "angle": 0, + "content": "
VG [29]COCO-stuff [4]
ObjectRelationshipObjectRelationship
rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑
SGT10.5122.9 / 45.75.8136.9 / 69.410.5426.4 / 52.32.9225.7 / 66.4
+Fsp9.6124.6 / 47.05.2337.7 / 72.410.0127.0 / 53.42.8925.9 / 67.1
+Fillum9.2526.4 / 49.34.8238.2 / 73.19.5427.5 / 54.42.8227.0 / 67.2
+LGAN10.123.1 / 46.55.2437.4 / 71.310.4126.3 / 53.12.9825.9 / 66.9
+LGAN9.9723.4 / 46.25.4237.2 / 71.710.7526.7 / 52.92.9326.3 / 66.5
SGT+8.9327.9 / 50.44.5939.4 / 75.28.8328.4 / 55.32.7328.3 / 68.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation and ablation study of scene graph expansion on VG [29] and COCO-stuff [4] datasets. \\(\\uparrow\\) denotes the higher is better, \\(\\downarrow\\) the lower is better. SGT+ denotes the pipeline that adopts our approach as a plug-in module." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.307, + 0.275, + 0.453 + ], + "angle": 0, + "content": "
mIoUVGCOCO-stuff
SGT22.732.6
+Fsp25.434.8
+Fillum26.436.2
+LG-LGAN23.633.4
+LG-GAN24.933.1
SGT+28.438.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.457, + 0.272, + 0.512 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with mIoU metric." + }, + { + "type": "table", + "bbox": [ + 0.281, + 0.307, + 0.477, + 0.453 + ], + "angle": 0, + "content": "
FIDVGCOCO-stuff
SGT20.1219.12
+Fsp19.8919.02
+Fillum19.9218.93
+LG-LGAN19.4818.60
+LG-GAN19.5218.59
SGT+19.4718.38
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.278, + 0.457, + 0.471, + 0.512 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with FID metric." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.351, + 0.559 + ], + "angle": 0, + "content": "5.5. User Studies & Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.793 + ], + "angle": 0, + "content": "User Studies. For image outpainting, any appropriate generation should be treated to be correct, thus without a specific target output. Therefore, in order to further evaluate the qualitative effectiveness of our approach, we conduct two user studies, where 91 users are involved to evaluate the outpainting correctness of our approach. In the first experiment, we send each user 20 randomized selected pairs of contents, in which the first one is the observed image with filling black in the unobserved area and the second one is the scene layout we predict to appear in the unobserved area, we then ask the users whether the predicted semantic could appear in the unobserved area. Finally, \\(89\\%\\) of the selected pairs are scored to be true, which denotes that our proposed approach is able to predict the appropriate scene layout for the unobserved semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In the second experiment, we send each user 20 randomized selected pairs of images, where one of them comes from SGT and the other one comes from adopting our approach as a plug-in module, we then ask the users which one is better in visual authenticity. Finally, images from ours achieve \\(87\\%\\) better chosen, which shows the significant quality enhancement produced by our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Ablation Study. We evaluate the individual effectiveness of the semantic removal, the illumination concatenation, the local alignment discriminator, and the global consistency discriminator in this experiment. The results show in Tab. 1, Tab. 2 and Tab. 3. It can be seen that the most significant enhancement in scene layout prediction comes from illumination concatenation, which shows the importance of incorporating a comprehensive illumination understanding to tackle the challenges of shadow extraction for obtaining accurate predictions. Meanwhile, the semantic removal enhances the accuracy notably, which shows that the shadow-irrelevant semantic information in the shadow areas leads to unexpected and noisy shadow representations. And the two discriminators achieve considerable enhancement in FID, which shows the importance of ensuring the local alignment between the unobserved semantics and its visible shadows and the consistency among all semantic-shadow pairs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.584, + 0.62, + 0.6 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.894, + 0.852 + ], + "angle": 0, + "content": "In this paper, we proposed a novel approach that aims to perform image outpainting by utilizing shadow information, which has never been explored. This is accomplished by extracting instance-level shadows cast by the unobserved areas from the input image. Then, the instance-level shadow representations are concatenated to infer the scene layout of each instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. Subsequently, extensive objective and subjective experiments are conducted which strongly proves the proposed approach successfully captures semantic and geometry information in the shadows. In future work, we will endeavor to learn shadow representations from dynamic environments and incorporate more tasks like novel view synthesis and pedestrian trajectory prediction into the unified framework." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement The research was supported by the Shanghai Committee of Science and Technology, China (Grant No.23ZR1423500)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7857" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Abdelrehim Ahmed and Aly Farag. Shape from shading for hybrid surfaces. In 2007 IEEE International Conference on Image Processing, pages II-525. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[2] Abdelrehim H Ahmed and Aly A Farag. Shape from shading under various imaging conditions. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.456 + ], + "angle": 0, + "content": "[6] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.471, + 0.528 + ], + "angle": 0, + "content": "[7] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.471, + 0.584 + ], + "angle": 0, + "content": "[8] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(9): 10850-10869, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.471, + 0.641 + ], + "angle": 0, + "content": "[9] R. Cucchiara, C. Grana, M. Piccardi, and A. Prati. Detecting moving objects, ghosts, and shadows in video streams. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(10):1337-1342, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[11] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[12] Aleksandrs Ecins, Cornelia Fermüller, and Yiannis Aloimonos. Shadow free segmentation in still images using local density measure. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-8, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.842 + ], + "angle": 0, + "content": "[13] Jiacheng Fan, Min Chen, Jinqiu Mo, Shigang Wang, and Qinghua Liang. Variational formulation of a hybrid perspective shape from shading model. The Visual Computer, pages 1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Silvano Galliani, Yong Chul Ju, Michael Breuß, and Andrés Bruhn. Generalised perspective shape from shading in spherical coordinates. In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 222-233. Springer, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Yinsai Guo, Hang Yu, Liyan Ma, Liang Zeng, and Xiangfeng Luo. Thfe: A triple-hierarchy feature enhancement method for tiny boat detection. Engineering Applications of Artificial Intelligence, 123:106271, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Yinsai Guo, Hang Yu, Shaorong Xie, Liyan Ma, Xinzhi Cao, and Xiangfeng Luo. Dsca: A dual semantic correlation alignment method for domain adaptation object detection. Pattern Recognition, page 110329, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[19] Weiqiang Jin, Biao Zhao, Hang Yu, Xi Tao, Ruiping Yin, and Guizhong Liu. Improving embedded knowledge graph multi-hop question answering by introducing relational chain reasoning. Data Mining and Knowledge Discovery, 37(1):255-288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[21] Imran N. Junejo and Hassan Foroosh. Estimating geotemporal location of stationary cameras using shadow trajectories. In Computer Vision – ECCV 2008, pages 318–331, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[22] Asaf Karnieli, Ohad Fried, and Yacov Hel-Or. Deepshadow: Neural shape from shadow. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part II, pages 415–430. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[23] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[24] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[25] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4019–4028, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[26] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Jihyun Kim, Seong-Hun Jeong, Kyeongbo Kong, and SukJu Kang. An unified framework for language guided image completion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2568-2578, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Daehyeon Kong, Kyeongbo Kong, Kyunghun Kim, Sung-Jun Min, and Suk-Ju Kang. Image-adaptive hint generation" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7858" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "via vision transformer for outpainting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3572-3581, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[30] Jean-François Lalonde, Alexei A. Efros, and Srinivasa G. Narasimhan. Estimating natural illumination from a single outdoor image. In 2009 IEEE 12th International Conference on Computer Vision, pages 183-190, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[31] Kyoung Mu Lee and C-C Jay Kuo. Shape from shading with a generalized reflectance map model. Computer vision and image understanding, 67(2):143-160, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.376 + ], + "angle": 0, + "content": "[32] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16221-16230, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[33] Pengbo Li, Hang Yu, Xiangfeng Luo, and Jia Wu. Lgmnn: A local and global aware memory-based graph neural network for fraud detection. IEEE Transactions on Big Data, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.47, + 0.504 + ], + "angle": 0, + "content": "[34] Jian Liang, Chenfei Wu, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. Advances in Neural Information Processing Systems, 35:15420-15432, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[36] Ruoshi Liu, Sachit Menon, Chengzhi Mao, Dennis Park, Simon Stent, and Carl Vondrick. What you can reconstruct from a shadow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17059-17068, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[37] Chia-Ni Lu, Ya-Chu Chang, and Wei-Chen Chiu. Bridging the visual gap: Wide-range image blending. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-851, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[38] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[39] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1680-1691, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Pascal Mamassian. Shape from Shadows, pages 724-725. Springer, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[41] S. Nadimi and B. Bhanu. Physical models for moving shadow and object detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(8):1079-1087, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.893, + 0.276 + ], + "angle": 0, + "content": "[43] Takahiro Okabe, Imari Sato, and Yoichi Sato. Attached shadow coding: Estimating surface normals from shadows under unknown reflectance and lighting conditions. In 2009 IEEE 12th International Conference on Computer Vision, pages 1693-1700. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.893, + 0.333 + ], + "angle": 0, + "content": "[44] Alexandros Panagopoulos, Dimitris Samaras, and Nikos Paragios. Robust shadow and illumination estimation using a mixture model. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 651-658, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[45] Alexandros Panagopoulos, Chaohui Wang, Dimitris Samaras, and Nikos Paragios. Simultaneous cast shadows, illumination and geometry inference using hypergraphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(2):437-449, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.893, + 0.487 + ], + "angle": 0, + "content": "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[47] Jiayan Qiu, Xinchao Wang, Pascal Fua, and Dacheng Tao. Matching seqlets: An unsupervised approach for locality preserving sequence matching. IEEE transactions on pattern analysis and machine intelligence, 43(2):745-752, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[48] Jiayan Qiu, Xinchao Wang, Stephen J Maybank, and Dacheng Tao. World from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8493-8504, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.893, + 0.673 + ], + "angle": 0, + "content": "[49] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Hallucinating visual instances in total absentia. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 264-282. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.893, + 0.73 + ], + "angle": 0, + "content": "[50] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Scene essence. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8322-8333, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.893, + 0.842 + ], + "angle": 0, + "content": "[52] Mark Sabini and Gili Rusak. Painting outside the box: Image outpainting with gans. arXiv preprint arXiv:1808.08483, 2018.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7859" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[54] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[56] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[57] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[58] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[59] Pinzhuo Tian and Hang Yu. Can we improve meta-learning model in few-shot learning by aligning data distributions? Knowledge-Based Systems, 277:110800, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[60] Kushagra Tiwary, Tzofi Klinghoffer, and Ramesh Raskar. Towards learning neural representations from shadows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 300-316. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[61] Anwaar Ulhaq, Naveed Akhtar, and Ganna Pogrebna. Efficient diffusion models for vision: A survey. arXiv preprint arXiv:2210.09292, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[62] Guohui Wang and Jin Cheng. Three-dimensional reconstruction of hybrid surfaces using perspective shape from shading. Optik, 127(19):7740-7751, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[63] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–14, 2022. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[64] Yi Wang, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Wide-context semantic image extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1399-1408, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[65] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-GAN: Training GANs with diffusion. In The Eleventh International Conference on Learning Representations, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[66] Chenyu Wu, Srinivasa G Narasimhan, and Branislav Jaramaz. A multi-image shape-from-shading framework for near-lighting perspective endoscopes. International Journal of Computer Vision, 86:211-228, 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[67] Lin Wu and Xiaochun Cao. Geo-location estimation from two shadow trajectories. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 585-590, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[68] Xiaoyu Xu, Jiayan Qiu, Xinchao Wang, and Zhou Wang. Relationship spatialization for depth estimation. In European Conference on Computer Vision, pages 615-637. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[69] Chiao-An Yang, Cheng-Yo Tan, Wan-Cyuan Fan, Cheng-Fu Yang, Meng-Lin Wu, and Yu-Chiang Frank Wang. Scene graph expansion for semantics-guided image outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15626, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[70] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Yingxia Shao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.36, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[71] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv., 2023. Just Accepted. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[72] Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. Distilling knowledge from graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7074-7083, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.501, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[73] Zongxin Yang, Jian Dong, Ping Liu, Yi Yang, and Shuicheng Yan. Very long natural scenery image prediction by outpainting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10560-10569, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[74] Kai Yao, Penglei Gao, Xi Yang, Jie Sun, Rui Zhang, and Kaizhu Huang. Outpainting by queries. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.614, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[75] Guangcong Zheng, Shengming Li, Hui Wang, Taiping Yao, Yang Chen, Shouhong Ding, and Xi Li. Entropy-driven sampling and training scheme for conditional diffusion generation. In European Conference on Computer Vision, pages 754-769. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[76] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.756, + 0.892, + 0.825 + ], + "angle": 0, + "content": "[77] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7860" + } + ] +] \ No newline at end of file diff --git a/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_origin.pdf b/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a036f00fa8aedb033df23c4fa1a4ad943e0e9e95 --- /dev/null +++ b/2024/Shadow-Enlightened Image Outpainting/50380c3f-4d94-46f3-bb7f-dbe694f86799_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0f2c24fe1a0016242e526d9b858bbeb22eff626fa30f44a23d68b410cd3c86e +size 5289502 diff --git a/2024/Shadow-Enlightened Image Outpainting/full.md b/2024/Shadow-Enlightened Image Outpainting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6541c38c5918cc0d3c8641446bb0b80008b36741 --- /dev/null +++ b/2024/Shadow-Enlightened Image Outpainting/full.md @@ -0,0 +1,521 @@ +# Shadow-Enlightened Image Outpainting + +Hang Yu $^{1}$ , Ruilin Li $^{2}$ , Shaorong Xie $^{1}$ , Jiayan Qiu $^{3*}$ $^{1}$ School of Computer Engineering and Science, Shanghai University, China + $^{2}$ Institute of Artificial Intelligence, Shanghai University, China + $^{3}$ School of Computing and Mathematical Sciences, University of Leicester, United Kingdom {yuhang, ruilinli, srxie}@shu.edu.cn, jiayan.qiu.1991@outlook.com + +# Abstract + +Conventional image outpainting methods usually treat unobserved areas as unknown and extend the scene only in terms of semantic consistency, thus overlooking the hidden information in shadows cast by unobserved areas, such as the invisible shapes and semantics. In this paper, we propose to extract and utilize the hidden information of unobserved areas from their shadows to enhance image outpainting. To this end, we propose an end-to-end deep approach that explicitly looks into the shadows within the image. Specifically, we extract shadows from the input image and identify instance-level shadow regions cast by the unobserved areas. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. In the experiments, we show that our proposed approach provides complementary cues for outpainting and achieves considerable improvement on all datasets by adopting our approach as a plug-in module. + +# 1. Introduction + +Given an image of a complex scene, as shown in Fig. 1(a), what will its right-side unobserved area look like? Humans can, without much effort, infer the appropriate categories and rough shapes of invisible semantics (e.g. objects) and further imagine an adequately realistic overall appearance. One critical cue humans utilize to reason invisible semantics is the shadows cast in the image. In this regard, we identify the shadows of each visible object, on top of which the shadows cast by the unobserved areas are found. According to their shapes and relative spatial layout, we may reasonably infer that there is a person in the unobserved area, as shown in Fig. 1(b), which aligns with its + +![](images/5b881424d830e8672e1b846d373e8f17561ae59e056d25d47c5550881cf33be1.jpg) +Figure 1. Given the image of a complex scene in (a) for outpainting, the extension needs to align with visible semantics. For example, the result shown in (b) is more aligned with the visible semantics from a global shadowing direction compared to the result shown in (c). For better alignment, we overcome the challenges of shadow extraction in complex scenes, such as disjoined shadows (d), overlapped shadows (e), and shadows attaching unexpected semantic information (f), respectively. + +shadow and the given scene. However, existing outpainting methods [7, 26, 57, 58, 69, 74] overlook shadows in the image, leading to inferior results, as shown in Fig. 1(c), where misalignment exists between the extension and its shadows from the shadowing direction in the scene. + +Even though prior efforts on Shape-from-Shadow [22, 32, 36, 60] also aim to extract hidden information, such as the shapes, of the semantics from their shadows in simple scenes with an ideal light source and a given semantic category, they cannot perform complex image outpainting with unknown semantics, due to there are three challenges of shadow extraction in complex scenes. 1) Given the complex spatial layout of the instances, the shadows are usually occluded by other instances, resulting in disjoined shadows of a single object, as shown in Fig. 1(d), which leads to inaccu + +rate perception or misinterpretation of invisible semantics. 2) Given the complex illumination of the scene, the shadows cast by different semantics may overlap, thus causing information loss for the unobserved area, as shown in Fig. 1(e). 3) Given the complex spatial layout of the scene, the shadows usually contain semantic information about the planes on which they are cast, which may lead to shadow-irrelevant understanding, as shown in Fig. 1(f). + +To perform image outpainting by utilizing the hidden information in the shadows, we propose a novel approach that explicitly looks into the correlation between shadows from the unobserved area and the visible semantics, the shadows with semantic removal, and the consistency between local and global alignment. + +Specifically, as depicted in Fig. 2, given an input image, we first conduct shadow extraction to obtain those shadow regions belonging to unobserved instances. Then, we utilize the shadow representations from unobserved instances to predict the scene layout of these instances, and then concatenate them with the shadow representation of instances in unobserved areas for layout-to-image processing. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas for the outpainting image. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. + +Our contribution is therefore a novel framework designated to achieve image outpainting by utilizing the shadows, which is the first attempt to the best of our knowledge. This is accomplished by extracting the desired shadow information, followed by scene layout expansion and layout-to-image conversion to produce image outpainting. The whole pipeline is end-to-end trainable. By adopting our proposed framework as a plug-in module, the outpainting methods [69] have achieved considerable quantitative and qualitative performance improvement on all datasets. + +# 2. Related Work + +We briefly review the prior works related to ours, including shadow processing and image outpainting. + +Shadow Processing. Conventional shadow processing method reconstructs the surface shape of a single object, under ideal illuminance, from its shadow and visual information [40]. Then, a sequence of methods [1, 2, 13, 14, 31, 62, 66] is developed to improve its performance by involving carefully designed prior knowledge in the model. Then, due to the strong adaptability of deep learning, learning-based methods [22, 32] are proposed to deal with complicated environments, such as objects with non-Lambertian reflectance [32]. Moreover, shadows are utilized as cues for scene geometry information extraction [21, 23, 30, 36, 43-45, 47, 48, 67, 68], segmentation [12], object detection [9, 16, 17], and tracking [41]. Recently, a generative method + +[36] has been proposed to predict the shape of a single invisible object, with a specific category, from its shadow. However, none of the existing methods can infer the scene layout of the semantics from their shadows in complex scenes. + +Image Outpainting. Conventional image outpainting methods extend the image with the background texture and partially observed objects towards specific directions [5, 34, 37, 52, 73, 74]. Then, information on the image margin is utilized to enable the outpainting in any direction [28, 56, 64]. Recently, the extrapolation of segmentation information has been adopted in the outpainting process [26] to add novel semantics in the extending areas. Meanwhile, the scene graph of the image is utilized to introduce new semantics by learning the semantic co-occurrence with Graph Neural Networks (GNN) [19, 33, 49, 50, 59, 69, 72]. Conventional outpainting commonly uses Generative Adversarial Networks (GAN) [15] for image conversion. However, due to the stability of diffusion, diffusion models [3, 18, 42, 51, 55] are being recognized as a promising family of generative models that have proven to be state-of-the-art sample quality for a variety of image generation benchmarks [8, 61, 70], including class-conditional image generation [10, 75], text-to-image generation [27, 51, 53, 71], image-to-image translation [24, 38, 54], layout-to-image generation [6, 76]. However, none of the existing methods takes advantage of the shadows to enhance the outpainting. + +# 3. Method + +In this section, we detail the working scheme of the proposed approach which comprises three stages, as shown in Fig. 2. In Stage 1, we utilize the pretrained model of shadow detection and instance shadow detection to grain rough instance-level shadow regions. Meanwhile, a pretrained shadow removal model is to remove unexpected semantic information from the shadows for the purified shadow feature. Then, we calculate the connectivity between shadow regions to merge disjoined shadows and perform overlapped segmentation on the shadows to restore overlapped shadows of instances. After iterative optimization, we obtain refined instance-level shadow regions. In Stage 2, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. In Stage 3, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. + +# 3.1. Stage 1: Shadow Extraction + +Given the input image $I$ , we first adopt a pretrained shadow detection model [77] to extract all shadow areas: + +$$ +M _ {s} = \operatorname {D e t e c t i o n} (I), \tag {1} +$$ + +![](images/eb124024b5e96db202022757c0422ec5480f147018b0082118277ea5c3236d21.jpg) +Figure 2. Illustration of the proposed approach. $\bigoplus$ denotes concatenation. + +where $M_{s}$ denotes the mask of the shadow areas, respectively. Then, a pretrained instance shadow detection model [63] is introduced to identify the shadow areas for the visible semantics in image $I$ by: + +$$ +M _ {s} ^ {v} = \text {I n s t a n c e - D e t e c t i o n} (I). \tag {2} +$$ + +where $M_{s}^{v}$ denotes the instance-level mask of the shadow areas cast by the visible semantics. Consequently, a binary area-level mask of shadows cast by the unobserved area can be obtained with $M_{s}^{u} = M_{s} - M_{s}^{v}$ . Thus, we obtain a rough instance-level shadow region. + +Meanwhile, to obtain the shadow areas without the semantic information from the cast planes, we introduce a pretrained shadow removal model [39] and obtain the semantically purified shadow image $I_{s}^{p}$ by: + +$$ +I _ {s} ^ {p} = M _ {s} \odot (I - \text {S h a d o w R e m o v a l} (I)), \tag {3} +$$ + +where $\odot$ denotes the Hadamard product. + +After obtaining the shadow image of semantic removal $I_{s}^{p}$ , we can use a convolutional net to extract the shadow feature without semantics: + +$$ +F _ {s} ^ {p} = \operatorname {C o n v N e t} \left(I _ {s} ^ {p}\right), F _ {s} ^ {p} \in \mathbb {R} ^ {W \times H \times d}, \tag {4} +$$ + +where the ConvNet consists of 8 Residual Blocks and preserves the resolution of the output representation to be the same as the input image. $W, H$ denotes the width and height of the input image. $d$ denotes the dimension of the shadow feature map. + +Considering that the formation of shadows is influenced by the illumination conditions in the given scene, which can be a cue in addressing the issues of disjoined shadows and overlapped shadows, we concatenate the image $I$ and the shadow mask $M_{s}$ and fed it into a Convolutional Neural Network (CNN) to obtain the illumination feature $F^{\mathrm{illum}} \in \mathbb{R}^{(W / 16) \times (H / 16) \times d}$ in the scene as a cue for the shadow regions optimization. + +# 3.1.1 Shadow Merging Operation + +In complex scenes, due to the presence of instances that can occlude shadows, it can result in an instance having multiple disjointed shadows, as shown in Fig. 3. + +To classify a single instance with disjoined shadows, we first treat each connected component in $M_{s}^{u}$ as a shadow region weight mask. Let $M^{u} = \{m_{x}^{u}\} \in \mathbb{R}^{W\times H\times 1}, x\in [1,N_{u}]$ be the shadow weight masks that don't belong to any visible instances, where $N_{u}$ denotes the number of connected components in $M_{s}^{u}$ . For the rough instance-level mask of visible instances in $M_{s}^{v}$ , let $M^{v} = \{M_{y}^{v}\} \in \mathbb{R}^{W\times H\times 1}, y\in [1,N_{v}]$ be each shadow weight mask of visible instance, where $N_{v}$ denotes the number of visible instances. Consequently, we grain the initial instance-level shadow region $M$ : + +$$ +N = N _ {u} + N _ {v} \tag {5} +$$ + +$$ +M = \left\{m _ {z} \right\} = \left\{M ^ {u}, M ^ {v} \right\}, z \in [ 1, N ], \tag {6} +$$ + +For each shadow region, we extract the corresponding region-level shadow feature $F$ by the feature map by + +![](images/78af9bcb0430208658794ee617448a99469a531a21d2fd497787f1a362cdd9b1.jpg) +(a) the disjoined shadow of the unobserved instance + +![](images/d4d3503e3554b18a5242c2d741d0d6f7034a3b1c01ebbe23bd3bf814b036a3c5.jpg) +(b) the disjoined shadow of the visible instance +Figure 3. Samples of disjoined shadows caused by occlusion in complex scenes. (a) The shadow of an unobserved object is obstructed, creating two disjointed shadow regions. The red boxes indicate the disjoined shadows. (b) Due to the defect of the instance shadow detector, a portion of the shadow of the instance was not classified as a single instance. The blue boxes indicate the detected corresponding pair, and the yellow box indicates the undetected shadow region. + +element-wise multiplying the shadow feature without semantics $F_{s}^{p}$ and the shadow weight mask $M$ . Then, we use the max pooling operation to resize these features to match the size of the illumination feature. Thus, we obtain the cropped feature $\bar{F}$ by: + +$$ +F = F _ {s} ^ {p} \odot M, F \in \mathbb {R} ^ {W \times H \times d}, \tag {7} +$$ + +$$ +\bar {F} = \operatorname {m a x p o o l} (F), \bar {F} \in \mathbb {R} ^ {(W / 1 6) \times (H / 1 6) \times d}, \tag {8} +$$ + +To improve the computation of connectivity between shadow regions, we combine the cropped features $\bar{F}$ of two regions along with the illumination feature by concatenating them together. This concatenated input is fed into an encoder $E_{connect}$ , and then passed through a sigmoid activation function for connectivity between two shadow regions: + +$$ +\bar {F} = \{\bar {f} _ {m} \}, m \in [ 1, N ], \tag {9} +$$ + +$$ +\bar {f} _ {(i, j)} = \operatorname {E} _ {\text {c o n n e c t}} \left(\operatorname {c o n c a t} \left(\bar {f} _ {i}, \bar {f} _ {j}, F ^ {\text {i l l u m}}\right)\right), i \neq j \tag {10} +$$ + +$$ +\operatorname {C o n n e c t i v i t y} _ {(i, j)} = \operatorname {S i g m o i d} (\bar {f} _ {(i, j)}), i \neq j \tag {11} +$$ + +When the connectivity between two shadow regions is higher than a hand-setting threshold $\alpha_{1}$ , these two shadow regions are considered to be the shadows of the same object, accordingly merging their weight masks. Conversely, they stay unchanged. Finally, we obtain the updated shadow features $F^{update}$ after merging operation: + +$$ +F ^ {u p d a t e} = \left\{f _ {m} \right\} \in \mathbb {R} ^ {W \times H \times d}, m \in [ 1, N ^ {m} ], \tag {12} +$$ + +where $N^{m}$ denotes the number of instances after merging. + +# 3.1.2 Shadow Splitting Operation + +In complex scenes, due to the complex spatial layout in the scene, it is common to have overlapped shadows between instances, as shown in Fig. 4. + +![](images/19a1e402b502588010d9ae9701cb8c53b388821d206345b1d1ae1c6f8f040369.jpg) +(a) overlapped with shadows cast by the unobserved instances +Figure 4. Samples of overlapping shadows in complex scenes. The red boxes indicate the overlapped shadows. (a) The shadows of multiple unobserved instances overlap. (b) The shadows of unobserved instances overlap with the shadows of visible instances. + +![](images/af4fb7cb5629cdb12b6b421c22c56b4d8c2cd7275beb997a4cd3c13aaaf4248f.jpg) +(b) overlapped with shadow cast by the visible instances + +In this work, each overlapped shadow is decoupled into two intersecting graph spaces. After concatenating the cropped feature $\bar{F}$ and illumination feature $F^{illum}$ , we input them into the Overlapping BiLayers [25]. Then, we grain two weight masks $M_0'$ , and $M_1'$ , each weight mask represents the shadow region of an individual instance. Then, we can obtain the "de-overlapped" shadow feature $f$ of an instance by element-wise multiplying the weight mask $M'$ with the purified shadow feature $F_s^p$ : + +$$ +M _ {0} ^ {\prime}, M _ {1} ^ {\prime}, \in \mathbb {R} ^ {W \times H \times 1} \tag {13} +$$ + +$$ +f _ {0} = M _ {0} ^ {\prime} \odot F _ {s} ^ {p}, f _ {1} = M _ {1} ^ {\prime} \odot F _ {s} ^ {p}, \tag {14} +$$ + +when the ratio of the combined area of two weight masks to the source area is higher than a hand-setting threshold $\alpha_{2}$ , it is considered an overlapped shadow and needs to be split. Conversely, it stays unchanged. Finally, we obtain the updated shadow features $F^{update}$ after splitting operation: + +$$ +F ^ {u p d a t e} = \left\{f _ {m} \right\} \in \mathbb {R} ^ {W \times H \times d}, m \in [ 1, N ^ {s} ], \tag {15} +$$ + +where $N^s$ denotes the number of instances after splitting. + +# 3.1.3 Iterative Optimization + +In complex scenes, it is common to encounter situations where disjoined shadows and overlapped shadows co-occur. Therefore, we utilize an iterative combination of merging and splitting operations to optimize the categorization of all shadow regions in the image. Thus, the issues of disjoined shadows and overlapped shadows, shown in Fig. 3 and Fig. 4, are addressed. Specifically, the iterative shadow optimization is summarized as Algorithm 1. + +# 3.2. Stage 2: Outpaiting with Shadows + +After obtaining the refined instance-level shadow regions $M$ from Stage 1, we grain the cropped shadow representations $\bar{f}^u$ of unobserved instances by: + +$$ +\bar {f} ^ {u} = \operatorname {m a x p o o l} \left(M ^ {u} \odot F _ {s} ^ {p}\right), M ^ {u} \in M, \tag {16} +$$ + +Algorithm 1: Iterative Shadow Optimization +Input: rough instance-level shadow weight mask +Output: refined instance-level shadow weight mask +1 $M^{(0)}\gets M$ +2 for $t = 1$ to $T$ do +3 $\begin{array}{rl} & M^{(t)}\leftarrow \mathrm{ShadowMerging}(M^{(t - 1)})\\ & M^{(t)}\leftarrow \mathrm{ShadowSplitting}(M^{(t)})\\ & \mathrm{if}\| M^{(t)} - M^{(t - 1)}\| < \epsilon \mathrm{then}\\ & \mid \mathrm{break}\\ & \mathrm{end} \end{array}$ +5 +6 +7 +8 end +9 return $M^{(t)}$ + +which are concatenated to predict the categories and layouts of new instances and outpaint the unobserved areas. + +For scene layout expansion, given a partial input image with the partial scene graph and the corresponding layout, we use GNN [11] to reveal the scene layout $L_{exp}$ of new instances. Specifically, we pass the shadow representations $\tilde{f}^u$ from the unobserved instances through an encoder $E_h$ , and then concatenate the output with the node features $f^h$ of the masked instances from the encoder of GNN. + +$$ +\hat {f} ^ {h} = \operatorname {c o n c a t} \left(f ^ {h}, \mathrm {E} _ {\mathrm {h}} \left(\bar {f} ^ {u}\right)\right) \tag {17} +$$ + +For layout to image, given an input image $I$ with an expanded scene layout $L_{exp}$ , we use layout-to-image model [76] to perform layout to image conversion. Specifically, we pass the shadow representations $f^u$ from the unobserved instances through an encoder $E_l$ , and then concatenate the output with scene layout features $f^l$ of the expanded instance from the encoder of the layout-to-image module. + +$$ +\hat {f} ^ {l} = \operatorname {c o n c a t} \left(f ^ {l}, \mathrm {E} _ {\mathrm {l}} \left(\bar {f} ^ {u}\right)\right) \tag {18} +$$ + +Note that we focus on the design of extracting shadow representation, so producing high-accuracy predictions for scene layout and high-quality image outputs is not within the main scope of this work. Thus, the graph-based and diffusion-based designs can be replaced with any scene layout expansion and layout-to-image modules if desirable. + +# 3.3. Stage 3: Alignment Discriminator + +In stage 2, we outpaint the partial input image $I$ into the extended image $I_{E}$ with introduced instances. To enhance alignment between unobserved area and visible semantics, we adopt the pretrained instance shadow detector [63] to first extract the mask $M_E^u$ of the unobserved area and its corresponding shadow from $I_{E}$ , and then feed it into a local shadow-instance alignment discriminator to ensure the + +generated alignment is visually real. This is calculated as follows: + +$$ +\begin{array}{l} \min _ {L D} V (L D) = \frac {1}{2} \mathbb {E} _ {x \sim p _ {\text {d a t a}} (x)} [ (L D (x) - b) ^ {2} ] \tag {19} \\ + \frac {1}{2} \mathbb {E} _ {z \sim p z (z)} [ (L D (M _ {E} ^ {u} \odot G (z)) - a) ^ {2} ], \\ \end{array} +$$ + +$$ +\min _ {G} V (G) = \frac {1}{2} \mathbb {E} _ {z \sim p z (z)} [ (L D (M _ {E} ^ {u} \odot G (z)) - c) ^ {2} ], \tag {20} +$$ + +where $LD$ denotes the local alignment discriminator. $G$ denotes the generator for outpainting. $a$ and $b$ denote the ground truth real and fake labels, respectively. $c$ denotes the value that $G$ wants $LD$ to believe for the fake data. + +Then, the mask $M_E$ that consists of all semantics and their shadows in $I_E$ is extracted, and we feed it into a global alignment discriminator to ensure the newly generated alignment is consistent with the existing ones. This is calculated as follows: + +$$ +\begin{array}{l} \min _ {G D} V (G D) = \frac {1}{2} \mathbb {E} _ {x \sim p _ {\text {d a t a}} (x)} [ (G D (x) - b) ^ {2} ] \tag {21} \\ + \frac {1}{2} \mathbb {E} _ {z \sim p _ {z} (z)} [ (G D (M _ {E} \odot G (z)) - a) ^ {2} ], \\ \end{array} +$$ + +$$ +\min _ {G} V (G) = \frac {1}{2} \mathbb {E} _ {z \sim p _ {z} (z)} [ (G D (M _ {E} \odot G (z)) - c) ^ {2} ], \tag {22} +$$ + +where $GD$ denotes the global alignment consistency discriminator. + +In summary, shadow extraction, scene layout expansion, and layout-to-image module are end-to-end trained together, thus the visual authenticity loss may facilitate shadow representation learning. Therefore, the final objective function of our approach is: + +$$ +L = \lambda_ {1} L _ {t} + \lambda_ {2} L _ {L - G A N} ^ {G} + \lambda_ {3} L _ {G - G A N} ^ {G}, \tag {23} +$$ + +where $\lambda_{1},\lambda_{2}$ and $\lambda_{3}$ denote the balancing weights. $L_{t}$ denotes the loss adopted from the base model. $L_{L - GAN}^{G}$ and $L_{G - GAN}^{G}$ denote the alignment loss from the local alignment and global consistency. + +# 4. Implementation Details + +This section provides details of the training settings and module implementations. + +Training Settings. Our approach is implemented using PyTorch [46] on four NVIDIA-RTX A6000 GPUs. During the training process, we adopt the Adam optimizer, 0.001 for the local discriminator and the global discriminator, and 0.01 for all modules in shadow representation extractors. The batch size is set to 32 and the size of the input images is $256 \times 256$ . The training is manually stopped. + +Shadow Detection Models. We adopt the popular shadow segmentation method [77] as the shadow detector, which is able to segment the shadow areas in the image. Furthermore, we adopt the state-of-the-art instance shadow detection method [63] as our instance shadow detector, which is able to detect the shadow instances and their corresponding object in the image. However, it is unable to detect shadows cast by unobserved objects. + +Alignment Discriminator. We adopt diffusion timestep-dependent discriminators from Diffusion-GAN [65]. It not only can minimize the divergence between real and diffused data at the end of the process, but also minimizes the divergence between the diffused real data distribution and the diffused generator distribution over several timesteps. In training, the balancing weight $\lambda_{1}$ is set to be 0.5, $\lambda_{2}$ and $\lambda_{3}$ are set to 0.25. Note that the architecture of the discriminator can be adjusted based on the specific type of layout-to-image module being implemented. + +Scene Layout Expansion. We adopt GTwE [11] to predict the categories and layout of unobserved instances simultaneously. GTwE has the same aforementioned hyperparameters as [69], which are the attention hidden size $d_{atten} = 512$ , feed forward size $d_{ff} = 2048$ , multi-head number $n_{head} = 4$ , and dropout $= 0.1$ . + +Layout to Image. We adopt LayoutDiffusion [76] for better outpainting the unknown areas with the condition of layouts predicted by GTwE. Note that LayoutDiffusion can not take the visible semantics as guidance to outpaint the unknown areas. Thus, we follow RePaint [38] to modify the standard denoising process for conditioning on the given image content. In each step, we sample the known region from the input and the outpaint painted part from the DDPM [18] output. + +# 5. Experiments + +In this section, we provide our experimental setups and both the objective and subjective results. Since we are not aware of any existing work that performs the same task as we do, we mainly focus on showing the promise of the proposed approach. Specifically, we evaluate the effectiveness of our approach by conducting experiments with an image outpainting pipeline proposed by [69]. Our goal is, again, to show the possibility of learning effective shadow representations thus enhancing complex scene extension, rather than trying to beat the state-of-the-art shadow detection and outpainting method. Other modules with the same functionality, provided they are end-to-end trainable, can be adopted in our approach to achieve potentially better performance. + +# 5.1. Datasets + +We evaluate our proposed approach on two common datasets, Visual Genome (VG) [29] and COCO-stuff [4], which gives an adequate amount of scene layouts. + +Visual Genome [29] dataset collects 108,077 images with dense annotations of objects, attributes, and relationships. Following the setting of SG2Im [20], we divide the data into $80\%$ , $10\%$ , and $10\%$ for the train, val, and test set, respectively. We select the object and relationship categories occurring at least 2000 and 500 times in the train set, respectively, leaving 178 objects and 45 relationship types, and select the images with 3 to 30 bounding boxes and ignoring all small objects. Finally, the training/validation/test set will have 62565 / 5062 / 5096 images, respectively. + +COCO-stuff [4] dataset augments a subset of the COCO dataset [35] with additional semantic categories. Thus, a total of 80 object categories (car, dog, etc.) and 91 semantic categories (sky, snow, etc.) are available, with $118\mathrm{K} / 5\mathrm{K}$ annotated images for training/validation. Following the setting of SG2Im [20], the ground truth object coordinates in the images are utilized for constructing the synthetic relationships, a total of six relationships are considered: left of, right of, above, below, inside, and surrounding. In addition, we use images in the train and val set with 3 to 8 objects that cover more than $2\%$ of the image. Finally, there are 25,210 train and 3,097 val images. + +# 5.2. Quantitative and Qualitative Result + +In this section, we evaluate the quantitative and qualitative effectiveness of our proposed approach to the image outpainting task by adopting it as a plug-in module of the SOTA method SGT [69]. The outpainting performance can be divided into two steps, in which the first one focuses on predicting the category and layout of the unobserved semantic and then the second one aims to generate the outpainting images. Therefore, we evaluate our approach on both the two steps. As can be seen from Tab. 1, by adopting our approach as a plug-in module, SGT achieves significant improvement in unobserved semantic prediction and its relevant relationship estimation. Although our target is to enhance only the semantic prediction, the accurately predicted semantics consequently strengthen the relationship estimation, thus leading to overall improvement on all metrics. Then, as can be seen from Tab. 2, we achieve considerable improvement in unobserved layout prediction, which also explicitly impacts the alignment with visible semantics in addition to the categories of instances. Moreover, in order to evaluate the quality of outpainting images, we introduce the FID metric and the results shown in Tab. 3. + +Then, the visual results are shown in Fig. 5 and Fig. 6, it can be found that our proposed approach is able to enhance the outpainting in the unobserved area, which is achieved by constraining both the shape and the semantic of the outpainting to be aligned with its visible shadows. + +![](images/507c36b1fcd13bc8df4ee74a00f80144d6730383d156ee0c520f1ddf9c3fe9b8.jpg) + +![](images/8c8196d16f2b46be3ec0eeb4624f8b54332722bf9e5a137524fe7c1a8343e2b8.jpg) +Input Image + +![](images/f43c32b21a02dfa364d2470a0fbfab75b095fbdbee165a87f0e401404c6fe795.jpg) + +![](images/2fa230a843ace6fc2f17ef0ecc99a222c9218f4bcf3b49c3d4ea1732e1d19d95.jpg) +GT + +![](images/69c261e30788716cbe0e14c52e8c77001baf68dceec6ba74d0df9906f4237f38.jpg) + +![](images/8077ce50145808fd8a42b65b9c0dab38b19693f9ba4e41420b402b657d2a441c.jpg) +SGT + +![](images/04c7263528c87473ffd15ed9a75e12714af8b43c8202a941d3e57692bbee9b5e.jpg) + +![](images/3233d51603374a77d8811d016e9ed6b62a9b7ef4e192b55eeb17255f1d1454a7.jpg) +SGT+ + +![](images/ec9ef1f53292cf5b6da5472d7de385e6f737a2417e199f8229ae1c54bd0aebfa.jpg) + +![](images/b635e0cdf52dce6a52db8b709ebbd1e9f788c66c9312176bd9a818bb319b4d50.jpg) +Input Image + +![](images/532f11abfa6a50293b267351d27d21a785b1a6750caffdffd81743176170b5f9.jpg) + +![](images/da36c49a84075952fdbe4373289d492c0d25b076f4a61e0bd25da2a4fee4aeb0.jpg) +GT + +![](images/2cadabe0f39a46fd72b415d45d6501c20348df6bf45b3e5f2a52c90df58f99ad.jpg) + +![](images/1336f4d0364d8f32870075f36565c941a1ace692e719e68c9e27941dd01ded90.jpg) +SGT + +![](images/428b2aa356eb6e5ba77e39b097dd45826fc3dff01d53813f4dc3f2d29e55e53b.jpg) + +![](images/e5dd73909fb4fedfc21c49b21081441f4cd044f8958cb0ad1971188edddad578.jpg) +SGT+ + +![](images/b23e861f38b84cb2780da119efff16778bada98d4207be19d8ebe53d3d4f6e5a.jpg) +Figure 5. Visual results from SGT [69] on VG [29]. SGT+ denotes the result after adopting our approach as a plug-in module. + +![](images/87fe6d6696707b80d73c7896e3086d65be361b89004cbfb933a6d6b61bde8e31.jpg) +Input Image + +![](images/8f7149815bd580316a0b206745b418ff03c326c0c59a20dad678950cd6eebb70.jpg) + +![](images/98c49d5d9b7ff32ae6c918ab48e23256daae9d6c0700383b5d7ad5724e13547e.jpg) +GT + +![](images/027acb619e653579a32eafa7585b2ff2c2d11b2f364ce0cdeb8d69d3274be6af.jpg) + +![](images/8b3dd7c59f6a32de1309e451a8203faadd1f0aabfade41ad7f767ecd8dd7214c.jpg) +SGT + +![](images/445068fe0bf90dde40aa9b1e3b67038c15dce848a0ad8f373b46528055481804.jpg) + +![](images/4fc3f40336c634df678f719af7420bbbc259c9a5fca6652f9fd5465c33bd42f4.jpg) +SGT+ + +![](images/e108779d02e4474c5923be22214abca44dfbba988831b0ae5c695a2175dfa183.jpg) + +![](images/8b8da01b975ac40f746ecee93cafb33516c0cba21fb14a2c05898f6e32ab420a.jpg) +Input Image +Figure 6. Visual results from SGT [69] on COCO-stuff [4]. SGT+ denotes the result after adopting our approach as a plug-in module. + +![](images/11599049e0268ffa66c264212da71c11d14f78fdc5af01842d27a85f18fcac8d.jpg) + +![](images/bf40b14a6e16ac30b44cb143f9ff9083caafab3d745d5d1356439fd3200d3912.jpg) +GT + +![](images/3a75af3b85a5c88b27f36b5946e3450eeddc31eba6d7a8d80e53417f94478f9c.jpg) + +![](images/4025e65f0e0ffd1af6154f94e4232d2311c5be23bad8a6b4a361904a6110c881.jpg) +SGT + +![](images/25354537f43adf0eb68d0a2045ffb0617b034243cf88693b7c2ed78c34ce3158.jpg) + +![](images/34b2061649fd74bc43727a5fffc7a4d09d3256da8d4a73eb6f5c4b760c89acd1.jpg) +SGT+ + +# 5.3. Multiple Instances Extension Comparison + +To verify whether our approach can enhance alignment with visible semantics in complex scenes, even with fewer shadows. We increase the number of instances requiring inference from unobserved areas to validate alignment between the extension and visible semantics shown in Fig 7. + +![](images/bae14e966ee2dce36cb9bdc43beefdf271106b9c9c46b7fb84d7b40286e3eb86.jpg) +Input Image + +![](images/2168582660cdf738a174704976c9673350d3459bb8a92954094a6fa960d47f0f.jpg) +the number of instances increasing + +![](images/e548ccc6727a76f4357bf38c6e44fef3fe66245475fea6b059cb4c27d33bfd68.jpg) + +![](images/f0a931508161131605b23e1103a3ba04c059821f901a7ede1f85a1983c1dd07f.jpg) + +![](images/8dd784e6cca593a8780cb53d15a002c3e52885add9e0ade594f3fbbb7b2ed178.jpg) +GT + +![](images/acf32a73f6381223a58064831bbcade8710a7c8f2083b865b433f162d62cc2c5.jpg) +the number of instances increasing + +![](images/c08bb8509a845af1422592eb1902b07bab48279a46365177bd6f349b03396401.jpg) + +![](images/44fd4c4c432db2f9cb07fc519f42ac0a36fe004768326725e2756d020a848963.jpg) +Figure 7. Qualitative comparison of multiple instances extension. Each image adds one more instance from left to right. + +As we can see, our method can effectively leverage shadow information in complex scenes to infer a scene layout that aligns well with the visible semantics. Specifically, instances that its shadow is not in the visible area can still align with visible semantics compared to the method without incorporating our approach, which ensures a harmonious generation of content with visible semantics. + +# 5.4. Shadow Extension Comparison + +As seen in Fig. 5 and Fig. 6, our visual results demonstrate that outpainting with our module achieves excellent alignment with visible semantics. However, there is an additional issue we need to consider: the shadows of visible instances can also affect the overall shadow consistency of the image. Therefore, we have also conducted a comparison specifically addressing this concern, as shown in Fig. 8. + +![](images/582fe5eedb59f32383079296e69635e6b5df01697dcfdf72df7667594136a855.jpg) + +![](images/04e70a44dea66bf9fd6572afaa280f5252efb0d22468bc97a3dc9cc2d9b86a38.jpg) +Input Image +Figure 8. Qualitative comparison of shadow extension on the visible instance. Red boxes show the differences in the extended shadow of the visible instance. + +![](images/b254c80040bed00a89c540b014e2b1bf421b37ecc6c0115df40394551ac70c3d.jpg) + +![](images/b6170703cad1030b41ce417daca3951292be7335d3a02b1e67cd3a494b2f32cc.jpg) +GT + +![](images/d7ec74152898e0c99ac9dbe70f7514fadec60ec6cfd4db5cd62e2c35e9e9c770.jpg) + +![](images/481d9396ad31da0b6a3f649e651c6b3c48fb956af5d96965c5628913ec16232f.jpg) +SGT + +![](images/9a69c1834585ff084ff8246727b99f0238787051ee1dc2de7c7224c098960e55.jpg) + +![](images/b508fc767830ffeaa28ec4cfd59e63dd13666870b3014468fc6ca0d3ec81bc3a.jpg) +SGT+ + +As we can see, our shadow extension harmonizes the generated with the visible semantics. However, the method without incorporating our module randomly elongates or abruptly truncates shadows, resulting in a lack of global alignment among the semantic-shadow pairs in the scene. + +
VG [29]COCO-stuff [4]
ObjectRelationshipObjectRelationship
rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑
SGT10.5122.9 / 45.75.8136.9 / 69.410.5426.4 / 52.32.9225.7 / 66.4
+Fsp9.6124.6 / 47.05.2337.7 / 72.410.0127.0 / 53.42.8925.9 / 67.1
+Fillum9.2526.4 / 49.34.8238.2 / 73.19.5427.5 / 54.42.8227.0 / 67.2
+LGAN10.123.1 / 46.55.2437.4 / 71.310.4126.3 / 53.12.9825.9 / 66.9
+LGAN9.9723.4 / 46.25.4237.2 / 71.710.7526.7 / 52.92.9326.3 / 66.5
SGT+8.9327.9 / 50.44.5939.4 / 75.28.8328.4 / 55.32.7328.3 / 68.4
+ +Table 1. Quantitative evaluation and ablation study of scene graph expansion on VG [29] and COCO-stuff [4] datasets. $\uparrow$ denotes the higher is better, $\downarrow$ the lower is better. SGT+ denotes the pipeline that adopts our approach as a plug-in module. + +
mIoUVGCOCO-stuff
SGT22.732.6
+Fsp25.434.8
+Fillum26.436.2
+LG-LGAN23.633.4
+LG-GAN24.933.1
SGT+28.438.1
+ +Table 2. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with mIoU metric. + +
FIDVGCOCO-stuff
SGT20.1219.12
+Fsp19.8919.02
+Fillum19.9218.93
+LG-LGAN19.4818.60
+LG-GAN19.5218.59
SGT+19.4718.38
+ +Table 3. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with FID metric. + +# 5.5. User Studies & Ablation Study + +User Studies. For image outpainting, any appropriate generation should be treated to be correct, thus without a specific target output. Therefore, in order to further evaluate the qualitative effectiveness of our approach, we conduct two user studies, where 91 users are involved to evaluate the outpainting correctness of our approach. In the first experiment, we send each user 20 randomized selected pairs of contents, in which the first one is the observed image with filling black in the unobserved area and the second one is the scene layout we predict to appear in the unobserved area, we then ask the users whether the predicted semantic could appear in the unobserved area. Finally, $89\%$ of the selected pairs are scored to be true, which denotes that our proposed approach is able to predict the appropriate scene layout for the unobserved semantics. + +In the second experiment, we send each user 20 randomized selected pairs of images, where one of them comes from SGT and the other one comes from adopting our approach as a plug-in module, we then ask the users which one is better in visual authenticity. Finally, images from ours achieve $87\%$ better chosen, which shows the significant quality enhancement produced by our approach. + +Ablation Study. We evaluate the individual effectiveness of the semantic removal, the illumination concatenation, the local alignment discriminator, and the global consistency discriminator in this experiment. The results show in Tab. 1, Tab. 2 and Tab. 3. It can be seen that the most significant enhancement in scene layout prediction comes from illumination concatenation, which shows the importance of incorporating a comprehensive illumination understanding to tackle the challenges of shadow extraction for obtaining accurate predictions. Meanwhile, the semantic removal enhances the accuracy notably, which shows that the shadow-irrelevant semantic information in the shadow areas leads to unexpected and noisy shadow representations. And the two discriminators achieve considerable enhancement in FID, which shows the importance of ensuring the local alignment between the unobserved semantics and its visible shadows and the consistency among all semantic-shadow pairs. + +# 6. Conclusion + +In this paper, we proposed a novel approach that aims to perform image outpainting by utilizing shadow information, which has never been explored. This is accomplished by extracting instance-level shadows cast by the unobserved areas from the input image. Then, the instance-level shadow representations are concatenated to infer the scene layout of each instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. Subsequently, extensive objective and subjective experiments are conducted which strongly proves the proposed approach successfully captures semantic and geometry information in the shadows. In future work, we will endeavor to learn shadow representations from dynamic environments and incorporate more tasks like novel view synthesis and pedestrian trajectory prediction into the unified framework. + +Acknowledgement The research was supported by the Shanghai Committee of Science and Technology, China (Grant No.23ZR1423500). + +# References + +[1] Abdelrehim Ahmed and Aly Farag. Shape from shading for hybrid surfaces. In 2007 IEEE International Conference on Image Processing, pages II-525. IEEE, 2007. 2 +[2] Abdelrehim H Ahmed and Aly A Farag. Shape from shading under various imaging conditions. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2 +[3] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 2 +[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6, 7, 8 +[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2 +[6] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 2 +[7] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 1 +[8] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(9): 10850-10869, 2023. 2 +[9] R. Cucchiara, C. Grana, M. Piccardi, and A. Prati. Detecting moving objects, ghosts, and shadows in video streams. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(10):1337-1342, 2003. 2 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[11] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 5, 6 +[12] Aleksandrs Ecins, Cornelia Fermüller, and Yiannis Aloimonos. Shadow free segmentation in still images using local density measure. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-8, 2014. 2 +[13] Jiacheng Fan, Min Chen, Jinqiu Mo, Shigang Wang, and Qinghua Liang. Variational formulation of a hybrid perspective shape from shading model. The Visual Computer, pages 1-14, 2022. 2 +[14] Silvano Galliani, Yong Chul Ju, Michael Breuß, and Andrés Bruhn. Generalised perspective shape from shading in spherical coordinates. In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013, + +Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 222-233. Springer, 2013. 2 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[16] Yinsai Guo, Hang Yu, Liyan Ma, Liang Zeng, and Xiangfeng Luo. Thfe: A triple-hierarchy feature enhancement method for tiny boat detection. Engineering Applications of Artificial Intelligence, 123:106271, 2023. 2 +[17] Yinsai Guo, Hang Yu, Shaorong Xie, Liyan Ma, Xinzhi Cao, and Xiangfeng Luo. Dsca: A dual semantic correlation alignment method for domain adaptation object detection. Pattern Recognition, page 110329, 2024. 2 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 6 +[19] Weiqiang Jin, Biao Zhao, Hang Yu, Xi Tao, Ruiping Yin, and Guizhong Liu. Improving embedded knowledge graph multi-hop question answering by introducing relational chain reasoning. Data Mining and Knowledge Discovery, 37(1):255-288, 2023. 2 +[20] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 6 +[21] Imran N. Junejo and Hassan Foroosh. Estimating geotemporal location of stationary cameras using shadow trajectories. In Computer Vision – ECCV 2008, pages 318–331, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2 +[22] Asaf Karnieli, Ohad Fried, and Yacov Hel-Or. Deepshadow: Neural shape from shadow. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part II, pages 415–430. Springer, 2022. 1, 2 +[23] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2 +[24] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2 +[25] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4019–4028, 2021. 4 +[26] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, 2021. 1, 2 +[27] Jihyun Kim, Seong-Hun Jeong, Kyeongbo Kong, and SukJu Kang. An unified framework for language guided image completion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2568-2578, 2023. 2 +[28] Daehyeon Kong, Kyeongbo Kong, Kyunghun Kim, Sung-Jun Min, and Suk-Ju Kang. Image-adaptive hint generation + +via vision transformer for outpainting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3572-3581, 2022. 2 +[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7, 8 +[30] Jean-François Lalonde, Alexei A. Efros, and Srinivasa G. Narasimhan. Estimating natural illumination from a single outdoor image. In 2009 IEEE 12th International Conference on Computer Vision, pages 183-190, 2009. 2 +[31] Kyoung Mu Lee and C-C Jay Kuo. Shape from shading with a generalized reflectance map model. Computer vision and image understanding, 67(2):143-160, 1997. 2 +[32] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16221-16230, 2022. 1, 2 +[33] Pengbo Li, Hang Yu, Xiangfeng Luo, and Jia Wu. Lgmnn: A local and global aware memory-based graph neural network for fraud detection. IEEE Transactions on Big Data, 2023. 2 +[34] Jian Liang, Chenfei Wu, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. Advances in Neural Information Processing Systems, 35:15420-15432, 2022. 2 +[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[36] Ruoshi Liu, Sachit Menon, Chengzhi Mao, Dennis Park, Simon Stent, and Carl Vondrick. What you can reconstruct from a shadow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17059-17068, 2023. 1, 2 +[37] Chia-Ni Lu, Ya-Chu Chang, and Wei-Chen Chiu. Bridging the visual gap: Wide-range image blending. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-851, 2021. 2 +[38] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6 +[39] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1680-1691, 2023. 3 +[40] Pascal Mamassian. Shape from Shadows, pages 724-725. Springer, 2014. 2 + +[41] S. Nadimi and B. Bhanu. Physical models for moving shadow and object detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(8):1079-1087, 2004. 2 +[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[43] Takahiro Okabe, Imari Sato, and Yoichi Sato. Attached shadow coding: Estimating surface normals from shadows under unknown reflectance and lighting conditions. In 2009 IEEE 12th International Conference on Computer Vision, pages 1693-1700. IEEE, 2009. 2 +[44] Alexandros Panagopoulos, Dimitris Samaras, and Nikos Paragios. Robust shadow and illumination estimation using a mixture model. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 651-658, 2009. +[45] Alexandros Panagopoulos, Chaohui Wang, Dimitris Samaras, and Nikos Paragios. Simultaneous cast shadows, illumination and geometry inference using hypergraphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(2):437-449, 2013. 2 +[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[47] Jiayan Qiu, Xinchao Wang, Pascal Fua, and Dacheng Tao. Matching seqlets: An unsupervised approach for locality preserving sequence matching. IEEE transactions on pattern analysis and machine intelligence, 43(2):745-752, 2019. 2 +[48] Jiayan Qiu, Xinchao Wang, Stephen J Maybank, and Dacheng Tao. World from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8493-8504, 2019. 2 +[49] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Hallucinating visual instances in total absentia. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 264-282. Springer, 2020. 2 +[50] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Scene essence. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8322-8333, 2021. 2 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[52] Mark Sabini and Gili Rusak. Painting outside the box: Image outpainting with gans. arXiv preprint arXiv:1808.08483, 2018.2 +[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep + +language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[54] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2 +[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2 +[56] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163, 2021. 2 +[57] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163. IEEE, 2021. 1 +[58] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 1 +[59] Pinzhuo Tian and Hang Yu. Can we improve meta-learning model in few-shot learning by aligning data distributions? Knowledge-Based Systems, 277:110800, 2023. 2 +[60] Kushagra Tiwary, Tzofi Klinghoffer, and Ramesh Raskar. Towards learning neural representations from shadows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 300-316. Springer, 2022. 1 +[61] Anwaar Ulhaq, Naveed Akhtar, and Ganna Pogrebna. Efficient diffusion models for vision: A survey. arXiv preprint arXiv:2210.09292, 2022. 2 +[62] Guohui Wang and Jin Cheng. Three-dimensional reconstruction of hybrid surfaces using perspective shape from shading. Optik, 127(19):7740-7751, 2016. 2 +[63] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–14, 2022. 3, 5, 6 +[64] Yi Wang, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Wide-context semantic image extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1399-1408, 2019. 2 +[65] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-GAN: Training GANs with diffusion. In The Eleventh International Conference on Learning Representations, 2023. 6 +[66] Chenyu Wu, Srinivasa G Narasimhan, and Branislav Jaramaz. A multi-image shape-from-shading framework for near-lighting perspective endoscopes. International Journal of Computer Vision, 86:211-228, 2010. 2 + +[67] Lin Wu and Xiaochun Cao. Geo-location estimation from two shadow trajectories. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 585-590, 2010. 2 +[68] Xiaoyu Xu, Jiayan Qiu, Xinchao Wang, and Zhou Wang. Relationship spatialization for depth estimation. In European Conference on Computer Vision, pages 615-637. Springer, 2022. 2 +[69] Chiao-An Yang, Cheng-Yo Tan, Wan-Cyuan Fan, Cheng-Fu Yang, Meng-Lin Wu, and Yu-Chiang Frank Wang. Scene graph expansion for semantics-guided image outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15626, 2022. 1, 2, 6, 7 +[70] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Yingxia Shao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2 +[71] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv., 2023. Just Accepted. 2 +[72] Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. Distilling knowledge from graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7074-7083, 2020. 2 +[73] Zongxin Yang, Jian Dong, Ping Liu, Yi Yang, and Shuicheng Yan. Very long natural scenery image prediction by outpainting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10560-10569, 2019. 2 +[74] Kai Yao, Penglei Gao, Xi Yang, Jie Sun, Rui Zhang, and Kaizhu Huang. Outpainting by queries. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 1, 2 +[75] Guangcong Zheng, Shengming Li, Hui Wang, Taiping Yao, Yang Chen, Shouhong Ding, and Xi Li. Entropy-driven sampling and training scheme for conditional diffusion generation. In European Conference on Computer Vision, pages 754-769. Springer, 2022. 2 +[76] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 5, 6 +[77] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 2, 6 \ No newline at end of file diff --git a/2024/Shadow-Enlightened Image Outpainting/images.zip b/2024/Shadow-Enlightened Image Outpainting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ce119a1c561e05c8fb4f0b7b3c2ca3bb47e752c --- /dev/null +++ b/2024/Shadow-Enlightened Image Outpainting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:856d982468d542b97ad811d7b6e0ab85fe9b41b0fb4eaad9c0979155ac89da12 +size 794410 diff --git a/2024/Shadow-Enlightened Image Outpainting/layout.json b/2024/Shadow-Enlightened Image Outpainting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..93dc19da057cdf7105d05cb28eaf5250d7a4fc95 --- /dev/null +++ b/2024/Shadow-Enlightened Image Outpainting/layout.json @@ -0,0 +1,12880 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 103, + 423, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 103, + 423, + 121 + ], + "spans": [ + { + "bbox": [ + 170, + 103, + 423, + 121 + ], + "type": "text", + "content": "Shadow-Enlightened Image Outpainting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "spans": [ + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "content": "Hang Yu" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "content": ", Ruilin Li" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "content": ", Shaorong Xie" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "content": ", Jiayan Qiu" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "content": "School of Computer Engineering and Science, Shanghai University, China \n" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "content": "Institute of Artificial Intelligence, Shanghai University, China \n" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 142, + 522, + 216 + ], + "type": "text", + "content": "School of Computing and Mathematical Sciences, University of Leicester, United Kingdom {yuhang, ruilinli, srxie}@shu.edu.cn, jiayan.qiu.1991@outlook.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 507 + ], + "type": "text", + "content": "Conventional image outpainting methods usually treat unobserved areas as unknown and extend the scene only in terms of semantic consistency, thus overlooking the hidden information in shadows cast by unobserved areas, such as the invisible shapes and semantics. In this paper, we propose to extract and utilize the hidden information of unobserved areas from their shadows to enhance image outpainting. To this end, we propose an end-to-end deep approach that explicitly looks into the shadows within the image. Specifically, we extract shadows from the input image and identify instance-level shadow regions cast by the unobserved areas. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. In the experiments, we show that our proposed approach provides complementary cues for outpainting and achieves considerable improvement on all datasets by adopting our approach as a plug-in module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 550, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 694 + ], + "type": "text", + "content": "Given an image of a complex scene, as shown in Fig. 1(a), what will its right-side unobserved area look like? Humans can, without much effort, infer the appropriate categories and rough shapes of invisible semantics (e.g. objects) and further imagine an adequately realistic overall appearance. One critical cue humans utilize to reason invisible semantics is the shadows cast in the image. In this regard, we identify the shadows of each visible object, on top of which the shadows cast by the unobserved areas are found. According to their shapes and relative spatial layout, we may reasonably infer that there is a person in the unobserved area, as shown in Fig. 1(b), which aligns with its" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 241, + 544, + 423 + ], + "blocks": [ + { + "bbox": [ + 310, + 241, + 544, + 423 + ], + "lines": [ + { + "bbox": [ + 310, + 241, + 544, + 423 + ], + "spans": [ + { + "bbox": [ + 310, + 241, + 544, + 423 + ], + "type": "image", + "image_path": "5b881424d830e8672e1b846d373e8f17561ae59e056d25d47c5550881cf33be1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 427, + 547, + 515 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 547, + 515 + ], + "type": "text", + "content": "Figure 1. Given the image of a complex scene in (a) for outpainting, the extension needs to align with visible semantics. For example, the result shown in (b) is more aligned with the visible semantics from a global shadowing direction compared to the result shown in (c). For better alignment, we overcome the challenges of shadow extraction in complex scenes, such as disjoined shadows (d), overlapped shadows (e), and shadows attaching unexpected semantic information (f), respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "content": "shadow and the given scene. However, existing outpainting methods [7, 26, 57, 58, 69, 74] overlook shadows in the image, leading to inferior results, as shown in Fig. 1(c), where misalignment exists between the extension and its shadows from the shadowing direction in the scene." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Even though prior efforts on Shape-from-Shadow [22, 32, 36, 60] also aim to extract hidden information, such as the shapes, of the semantics from their shadows in simple scenes with an ideal light source and a given semantic category, they cannot perform complex image outpainting with unknown semantics, due to there are three challenges of shadow extraction in complex scenes. 1) Given the complex spatial layout of the instances, the shadows are usually occluded by other instances, resulting in disjoined shadows of a single object, as shown in Fig. 1(d), which leads to inaccu" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7850" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "content": "rate perception or misinterpretation of invisible semantics. 2) Given the complex illumination of the scene, the shadows cast by different semantics may overlap, thus causing information loss for the unobserved area, as shown in Fig. 1(e). 3) Given the complex spatial layout of the scene, the shadows usually contain semantic information about the planes on which they are cast, which may lead to shadow-irrelevant understanding, as shown in Fig. 1(f)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 286, + 239 + ], + "type": "text", + "content": "To perform image outpainting by utilizing the hidden information in the shadows, we propose a novel approach that explicitly looks into the correlation between shadows from the unobserved area and the visible semantics, the shadows with semantic removal, and the consistency between local and global alignment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 382 + ], + "type": "text", + "content": "Specifically, as depicted in Fig. 2, given an input image, we first conduct shadow extraction to obtain those shadow regions belonging to unobserved instances. Then, we utilize the shadow representations from unobserved instances to predict the scene layout of these instances, and then concatenate them with the shadow representation of instances in unobserved areas for layout-to-image processing. Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas for the outpainting image. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 286, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 502 + ], + "type": "text", + "content": "Our contribution is therefore a novel framework designated to achieve image outpainting by utilizing the shadows, which is the first attempt to the best of our knowledge. This is accomplished by extracting the desired shadow information, followed by scene layout expansion and layout-to-image conversion to produce image outpainting. The whole pipeline is end-to-end trainable. By adopting our proposed framework as a plug-in module, the outpainting methods [69] have achieved considerable quantitative and qualitative performance improvement on all datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 133, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 133, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 133, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 557 + ], + "type": "text", + "content": "We briefly review the prior works related to ours, including shadow processing and image outpainting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": "Shadow Processing. Conventional shadow processing method reconstructs the surface shape of a single object, under ideal illuminance, from its shadow and visual information [40]. Then, a sequence of methods [1, 2, 13, 14, 31, 62, 66] is developed to improve its performance by involving carefully designed prior knowledge in the model. Then, due to the strong adaptability of deep learning, learning-based methods [22, 32] are proposed to deal with complicated environments, such as objects with non-Lambertian reflectance [32]. Moreover, shadows are utilized as cues for scene geometry information extraction [21, 23, 30, 36, 43-45, 47, 48, 67, 68], segmentation [12], object detection [9, 16, 17], and tracking [41]. Recently, a generative method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "[36] has been proposed to predict the shape of a single invisible object, with a specific category, from its shadow. However, none of the existing methods can infer the scene layout of the semantics from their shadows in complex scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 121, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 396 + ], + "type": "text", + "content": "Image Outpainting. Conventional image outpainting methods extend the image with the background texture and partially observed objects towards specific directions [5, 34, 37, 52, 73, 74]. Then, information on the image margin is utilized to enable the outpainting in any direction [28, 56, 64]. Recently, the extrapolation of segmentation information has been adopted in the outpainting process [26] to add novel semantics in the extending areas. Meanwhile, the scene graph of the image is utilized to introduce new semantics by learning the semantic co-occurrence with Graph Neural Networks (GNN) [19, 33, 49, 50, 59, 69, 72]. Conventional outpainting commonly uses Generative Adversarial Networks (GAN) [15] for image conversion. However, due to the stability of diffusion, diffusion models [3, 18, 42, 51, 55] are being recognized as a promising family of generative models that have proven to be state-of-the-art sample quality for a variety of image generation benchmarks [8, 61, 70], including class-conditional image generation [10, 75], text-to-image generation [27, 51, 53, 71], image-to-image translation [24, 38, 54], layout-to-image generation [6, 76]. However, none of the existing methods takes advantage of the shadows to enhance the outpainting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 405, + 361, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 361, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 361, + 417 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "content": "In this section, we detail the working scheme of the proposed approach which comprises three stages, as shown in Fig. 2. In Stage 1, we utilize the pretrained model of shadow detection and instance shadow detection to grain rough instance-level shadow regions. Meanwhile, a pretrained shadow removal model is to remove unexpected semantic information from the shadows for the purified shadow feature. Then, we calculate the connectivity between shadow regions to merge disjoined shadows and perform overlapped segmentation on the shadows to restore overlapped shadows of instances. After iterative optimization, we obtain refined instance-level shadow regions. In Stage 2, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas. In Stage 3, two discriminators are implemented to enhance alignment between the extended semantics and their shadows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 460, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 460, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 460, + 647 + ], + "type": "text", + "content": "3.1. Stage 1: Shadow Extraction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 654, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 545, + 677 + ], + "type": "text", + "content": "Given the input image " + }, + { + "bbox": [ + 306, + 654, + 545, + 677 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 306, + 654, + 545, + 677 + ], + "type": "text", + "content": ", we first adopt a pretrained shadow detection model [77] to extract all shadow areas:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 697, + 545, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 697, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 382, + 697, + 545, + 710 + ], + "type": "interline_equation", + "content": "M _ {s} = \\operatorname {D e t e c t i o n} (I), \\tag {1}", + "image_path": "5e832c5ada96914c5947943844e695f7da23041fb8175e34c5279f560f5e3d29.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7851" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 529, + 329 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 529, + 329 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 529, + 329 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 529, + 329 + ], + "type": "image", + "image_path": "eb124024b5e96db202022757c0422ec5480f147018b0082118277ea5c3236d21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 338, + 430, + 350 + ], + "lines": [ + { + "bbox": [ + 160, + 338, + 430, + 350 + ], + "spans": [ + { + "bbox": [ + 160, + 338, + 430, + 350 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed approach. " + }, + { + "bbox": [ + 160, + 338, + 430, + 350 + ], + "type": "inline_equation", + "content": "\\bigoplus" + }, + { + "bbox": [ + 160, + 338, + 430, + 350 + ], + "type": "text", + "content": " denotes concatenation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "text", + "content": " denotes the mask of the shadow areas, respectively. Then, a pretrained instance shadow detection model [63] is introduced to identify the shadow areas for the visible semantics in image " + }, + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 417, + 287, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 417, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 102, + 417, + 287, + 431 + ], + "type": "interline_equation", + "content": "M _ {s} ^ {v} = \\text {I n s t a n c e - D e t e c t i o n} (I). \\tag {2}", + "image_path": "f49f695b9a54ae05054a325f66df2487b1c6411b04e35a5b9c2e7140dc7dbf79.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 438, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 438, + 287, + 498 + ], + "type": "inline_equation", + "content": "M_{s}^{v}" + }, + { + "bbox": [ + 46, + 438, + 287, + 498 + ], + "type": "text", + "content": " denotes the instance-level mask of the shadow areas cast by the visible semantics. Consequently, a binary area-level mask of shadows cast by the unobserved area can be obtained with " + }, + { + "bbox": [ + 46, + 438, + 287, + 498 + ], + "type": "inline_equation", + "content": "M_{s}^{u} = M_{s} - M_{s}^{v}" + }, + { + "bbox": [ + 46, + 438, + 287, + 498 + ], + "type": "text", + "content": ". Thus, we obtain a rough instance-level shadow region." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 547 + ], + "type": "text", + "content": "Meanwhile, to obtain the shadow areas without the semantic information from the cast planes, we introduce a pretrained shadow removal model [39] and obtain the semantically purified shadow image " + }, + { + "bbox": [ + 46, + 498, + 287, + 547 + ], + "type": "inline_equation", + "content": "I_{s}^{p}" + }, + { + "bbox": [ + 46, + 498, + 287, + 547 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 554, + 287, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 554, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 82, + 554, + 287, + 569 + ], + "type": "interline_equation", + "content": "I _ {s} ^ {p} = M _ {s} \\odot (I - \\text {S h a d o w R e m o v a l} (I)), \\tag {3}", + "image_path": "544687d78060f78e1b67565a239dd11b473f5177391541d1446ce30e51795aa6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 576, + 212, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 212, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 212, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 576, + 212, + 588 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 576, + 212, + 588 + ], + "type": "text", + "content": " denotes the Hadamard product." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 588, + 287, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 287, + 623 + ], + "type": "text", + "content": "After obtaining the shadow image of semantic removal " + }, + { + "bbox": [ + 47, + 588, + 287, + 623 + ], + "type": "inline_equation", + "content": "I_{s}^{p}" + }, + { + "bbox": [ + 47, + 588, + 287, + 623 + ], + "type": "text", + "content": ", we can use a convolutional net to extract the shadow feature without semantics:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 631, + 287, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 631, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 88, + 631, + 287, + 646 + ], + "type": "interline_equation", + "content": "F _ {s} ^ {p} = \\operatorname {C o n v N e t} \\left(I _ {s} ^ {p}\\right), F _ {s} ^ {p} \\in \\mathbb {R} ^ {W \\times H \\times d}, \\tag {4}", + "image_path": "07a95c62f32c7d172ccbb877e0adad263d59769fc9f064ffb08260dd1c8c17e4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "where the ConvNet consists of 8 Residual Blocks and preserves the resolution of the output representation to be the same as the input image. " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "W, H" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " denotes the width and height of the input image. " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " denotes the dimension of the shadow feature map." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "text", + "content": "Considering that the formation of shadows is influenced by the illumination conditions in the given scene, which can be a cue in addressing the issues of disjoined shadows and overlapped shadows, we concatenate the image " + }, + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "text", + "content": " and the shadow mask " + }, + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "text", + "content": " and fed it into a Convolutional Neural Network (CNN) to obtain the illumination feature " + }, + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{illum}} \\in \\mathbb{R}^{(W / 16) \\times (H / 16) \\times d}" + }, + { + "bbox": [ + 304, + 361, + 545, + 457 + ], + "type": "text", + "content": " in the scene as a cue for the shadow regions optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 471, + 458, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 458, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 458, + 483 + ], + "type": "text", + "content": "3.1.1 Shadow Merging Operation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 490, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 545, + 525 + ], + "type": "text", + "content": "In complex scenes, due to the presence of instances that can occlude shadows, it can result in an instance having multiple disjointed shadows, as shown in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": "To classify a single instance with disjoined shadows, we first treat each connected component in " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "M_{s}^{u}" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": " as a shadow region weight mask. Let " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "M^{u} = \\{m_{x}^{u}\\} \\in \\mathbb{R}^{W\\times H\\times 1}, x\\in [1,N_{u}]" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": " be the shadow weight masks that don't belong to any visible instances, where " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "N_{u}" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": " denotes the number of connected components in " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "M_{s}^{u}" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": ". For the rough instance-level mask of visible instances in " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "M_{s}^{v}" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "M^{v} = \\{M_{y}^{v}\\} \\in \\mathbb{R}^{W\\times H\\times 1}, y\\in [1,N_{v}]" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": " be each shadow weight mask of visible instance, where " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": " denotes the number of visible instances. Consequently, we grain the initial instance-level shadow region " + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 525, + 545, + 657 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 395, + 655, + 545, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 655, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 395, + 655, + 545, + 667 + ], + "type": "interline_equation", + "content": "N = N _ {u} + N _ {v} \\tag {5}", + "image_path": "1ddac7b5869fb319a93a2a94020be80a40c2de6be0042f289faad432b39441f3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 670, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 670, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 347, + 670, + 545, + 684 + ], + "type": "interline_equation", + "content": "M = \\left\\{m _ {z} \\right\\} = \\left\\{M ^ {u}, M ^ {v} \\right\\}, z \\in [ 1, N ], \\tag {6}", + "image_path": "baa275dd6e10c9ab6a1e2156abdf0afa422d3a42ebb969db1fc0ada3c4776122.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "For each shadow region, we extract the corresponding region-level shadow feature " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " by the feature map by" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7852" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 166, + 154 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 166, + 154 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 166, + 154 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 166, + 154 + ], + "type": "image", + "image_path": "78af9bcb0430208658794ee617448a99469a531a21d2fd497787f1a362cdd9b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 156, + 153, + 174 + ], + "lines": [ + { + "bbox": [ + 65, + 156, + 153, + 174 + ], + "spans": [ + { + "bbox": [ + 65, + 156, + 153, + 174 + ], + "type": "text", + "content": "(a) the disjoined shadow of the unobserved instance" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 286, + 154 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 286, + 154 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 286, + 154 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 286, + 154 + ], + "type": "image", + "image_path": "d4d3503e3554b18a5242c2d741d0d6f7034a3b1c01ebbe23bd3bf814b036a3c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 156, + 266, + 174 + ], + "lines": [ + { + "bbox": [ + 186, + 156, + 266, + 174 + ], + "spans": [ + { + "bbox": [ + 186, + 156, + 266, + 174 + ], + "type": "text", + "content": "(b) the disjoined shadow of the visible instance" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 177, + 287, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 264 + ], + "type": "text", + "content": "Figure 3. Samples of disjoined shadows caused by occlusion in complex scenes. (a) The shadow of an unobserved object is obstructed, creating two disjointed shadow regions. The red boxes indicate the disjoined shadows. (b) Due to the defect of the instance shadow detector, a portion of the shadow of the instance was not classified as a single instance. The blue boxes indicate the detected corresponding pair, and the yellow box indicates the undetected shadow region." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "text", + "content": "element-wise multiplying the shadow feature without semantics " + }, + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "inline_equation", + "content": "F_{s}^{p}" + }, + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "text", + "content": " and the shadow weight mask " + }, + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "text", + "content": ". Then, we use the max pooling operation to resize these features to match the size of the illumination feature. Thus, we obtain the cropped feature " + }, + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 46, + 276, + 287, + 335 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 343, + 287, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 343, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 287, + 357 + ], + "type": "interline_equation", + "content": "F = F _ {s} ^ {p} \\odot M, F \\in \\mathbb {R} ^ {W \\times H \\times d}, \\tag {7}", + "image_path": "dd9355f80ec4708d48d1570505cc0542e9b376617c2f7e6f06d37819d8cd131c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 360, + 287, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 360, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 78, + 360, + 287, + 374 + ], + "type": "interline_equation", + "content": "\\bar {F} = \\operatorname {m a x p o o l} (F), \\bar {F} \\in \\mathbb {R} ^ {(W / 1 6) \\times (H / 1 6) \\times d}, \\tag {8}", + "image_path": "d09b5b000c5e7281eee0c283be316162ffa572b0f4b038b7c51c3a8c069f8159.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 383, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 467 + ], + "type": "text", + "content": "To improve the computation of connectivity between shadow regions, we combine the cropped features " + }, + { + "bbox": [ + 46, + 383, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 46, + 383, + 287, + 467 + ], + "type": "text", + "content": " of two regions along with the illumination feature by concatenating them together. This concatenated input is fed into an encoder " + }, + { + "bbox": [ + 46, + 383, + 287, + 467 + ], + "type": "inline_equation", + "content": "E_{connect}" + }, + { + "bbox": [ + 46, + 383, + 287, + 467 + ], + "type": "text", + "content": ", and then passed through a sigmoid activation function for connectivity between two shadow regions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 475, + 287, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 475, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 117, + 475, + 287, + 490 + ], + "type": "interline_equation", + "content": "\\bar {F} = \\{\\bar {f} _ {m} \\}, m \\in [ 1, N ], \\tag {9}", + "image_path": "3827a2dd3eefeba31774d4628b4f39d5a945ba8b3c05475d1226b39b7038b50a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 491, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 491, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 64, + 491, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\bar {f} _ {(i, j)} = \\operatorname {E} _ {\\text {c o n n e c t}} \\left(\\operatorname {c o n c a t} \\left(\\bar {f} _ {i}, \\bar {f} _ {j}, F ^ {\\text {i l l u m}}\\right)\\right), i \\neq j \\tag {10}", + "image_path": "9c7b72c6d9b1f9c14d5446017782cb9b22eaabd8c70618d5449dfaf60e7ad38e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 507, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 507, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 69, + 507, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\operatorname {C o n n e c t i v i t y} _ {(i, j)} = \\operatorname {S i g m o i d} (\\bar {f} _ {(i, j)}), i \\neq j \\tag {11}", + "image_path": "36c83604d04fe5614dedba1ddda0c04ca76eebd4964c385cba94b6b6ef82b492.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": "When the connectivity between two shadow regions is higher than a hand-setting threshold " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": ", these two shadow regions are considered to be the shadows of the same object, accordingly merging their weight masks. Conversely, they stay unchanged. Finally, we obtain the updated shadow features " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "F^{update}" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " after merging operation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 609, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 609, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 67, + 609, + 287, + 624 + ], + "type": "interline_equation", + "content": "F ^ {u p d a t e} = \\left\\{f _ {m} \\right\\} \\in \\mathbb {R} ^ {W \\times H \\times d}, m \\in [ 1, N ^ {m} ], \\tag {12}", + "image_path": "cbabe04a0ebe46ff638ac6d7ce48d3b5da9e8f94d82beb51d6b7214a0f2772f1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 632, + 282, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 282, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 282, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 632, + 282, + 644 + ], + "type": "inline_equation", + "content": "N^{m}" + }, + { + "bbox": [ + 47, + 632, + 282, + 644 + ], + "type": "text", + "content": " denotes the number of instances after merging." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 658, + 200, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 200, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 200, + 671 + ], + "type": "text", + "content": "3.1.2 Shadow Splitting Operation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In complex scenes, due to the complex spatial layout in the scene, it is common to have overlapped shadows between instances, as shown in Fig. 4." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 426, + 154 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 426, + 154 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 426, + 154 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 426, + 154 + ], + "type": "image", + "image_path": "19a1e402b502588010d9ae9701cb8c53b388821d206345b1d1ae1c6f8f040369.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 156, + 420, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 156, + 420, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 156, + 420, + 175 + ], + "type": "text", + "content": "(a) overlapped with shadows cast by the unobserved instances" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 177, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 221 + ], + "type": "text", + "content": "Figure 4. Samples of overlapping shadows in complex scenes. The red boxes indicate the overlapped shadows. (a) The shadows of multiple unobserved instances overlap. (b) The shadows of unobserved instances overlap with the shadows of visible instances." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 544, + 154 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 544, + 154 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 544, + 154 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 544, + 154 + ], + "type": "image", + "image_path": "af4fb7cb5629cdb12b6b421c22c56b4d8c2cd7275beb997a4cd3c13aaaf4248f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 156, + 531, + 175 + ], + "lines": [ + { + "bbox": [ + 440, + 156, + 531, + 175 + ], + "spans": [ + { + "bbox": [ + 440, + 156, + 531, + 175 + ], + "type": "text", + "content": "(b) overlapped with shadow cast by the visible instances" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": "In this work, each overlapped shadow is decoupled into two intersecting graph spaces. After concatenating the cropped feature " + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": " and illumination feature " + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "inline_equation", + "content": "F^{illum}" + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": ", we input them into the Overlapping BiLayers [25]. Then, we grain two weight masks " + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "inline_equation", + "content": "M_0'" + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "inline_equation", + "content": "M_1'" + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": ", each weight mask represents the shadow region of an individual instance. Then, we can obtain the \"de-overlapped\" shadow feature " + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": " of an instance by element-wise multiplying the weight mask " + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": " with the purified shadow feature " + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "inline_equation", + "content": "F_s^p" + }, + { + "bbox": [ + 304, + 227, + 545, + 335 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 380, + 342, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 342, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 380, + 342, + 545, + 357 + ], + "type": "interline_equation", + "content": "M _ {0} ^ {\\prime}, M _ {1} ^ {\\prime}, \\in \\mathbb {R} ^ {W \\times H \\times 1} \\tag {13}", + "image_path": "34f9eda6ba8e02335575fc683f447b72fc6d28bb30fbfd85c170138aafc16bf3.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 359, + 359, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 359, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 359, + 359, + 545, + 373 + ], + "type": "interline_equation", + "content": "f _ {0} = M _ {0} ^ {\\prime} \\odot F _ {s} ^ {p}, f _ {1} = M _ {1} ^ {\\prime} \\odot F _ {s} ^ {p}, \\tag {14}", + "image_path": "d427ca1b17240707c457dd074eac4747b16f30a43f77ac4dd0e8e75cb42a3c62.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "content": "when the ratio of the combined area of two weight masks to the source area is higher than a hand-setting threshold " + }, + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "content": ", it is considered an overlapped shadow and needs to be split. Conversely, it stays unchanged. Finally, we obtain the updated shadow features " + }, + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "inline_equation", + "content": "F^{update}" + }, + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "content": " after splitting operation:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 328, + 461, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 461, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 328, + 461, + 545, + 475 + ], + "type": "interline_equation", + "content": "F ^ {u p d a t e} = \\left\\{f _ {m} \\right\\} \\in \\mathbb {R} ^ {W \\times H \\times d}, m \\in [ 1, N ^ {s} ], \\tag {15}", + "image_path": "5d04181bfa6cffa4a7f7e6d7c7c4a416ff10f1bb767cecd114236f192def77b1.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 485, + 537, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 537, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 537, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 485, + 537, + 498 + ], + "type": "inline_equation", + "content": "N^s" + }, + { + "bbox": [ + 306, + 485, + 537, + 498 + ], + "type": "text", + "content": " denotes the number of instances after splitting." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 511, + 434, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 434, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 434, + 523 + ], + "type": "text", + "content": "3.1.3 Iterative Optimization" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 530, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 626 + ], + "type": "text", + "content": "In complex scenes, it is common to encounter situations where disjoined shadows and overlapped shadows co-occur. Therefore, we utilize an iterative combination of merging and splitting operations to optimize the categorization of all shadow regions in the image. Thus, the issues of disjoined shadows and overlapped shadows, shown in Fig. 3 and Fig. 4, are addressed. Specifically, the iterative shadow optimization is summarized as Algorithm 1." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 633, + 488, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 488, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 488, + 646 + ], + "type": "text", + "content": "3.2. Stage 2: Outpaiting with Shadows" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 652, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 652, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 652, + 545, + 688 + ], + "type": "text", + "content": "After obtaining the refined instance-level shadow regions " + }, + { + "bbox": [ + 306, + 652, + 545, + 688 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 652, + 545, + 688 + ], + "type": "text", + "content": " from Stage 1, we grain the cropped shadow representations " + }, + { + "bbox": [ + 306, + 652, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\bar{f}^u" + }, + { + "bbox": [ + 306, + 652, + 545, + 688 + ], + "type": "text", + "content": " of unobserved instances by:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 346, + 696, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 696, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 346, + 696, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\bar {f} ^ {u} = \\operatorname {m a x p o o l} \\left(M ^ {u} \\odot F _ {s} ^ {p}\\right), M ^ {u} \\in M, \\tag {16}", + "image_path": "70b550abbe639dea0bebde1a23f71713e160defaba1dbd9a97d46cfea77281d7.jpg" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7853" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 50, + 88, + 286, + 227 + ], + "blocks": [ + { + "bbox": [ + 52, + 75, + 235, + 87 + ], + "lines": [ + { + "bbox": [ + 52, + 75, + 235, + 87 + ], + "spans": [ + { + "bbox": [ + 52, + 75, + 235, + 87 + ], + "type": "text", + "content": "Algorithm 1: Iterative Shadow Optimization" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "lines": [ + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "text", + "content": "Input: rough instance-level shadow weight mask \nOutput: refined instance-level shadow weight mask \n1 " + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "inline_equation", + "content": "M^{(0)}\\gets M" + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "text", + "content": " \n2 for " + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "text", + "content": " do \n3 " + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl} & M^{(t)}\\leftarrow \\mathrm{ShadowMerging}(M^{(t - 1)})\\\\ & M^{(t)}\\leftarrow \\mathrm{ShadowSplitting}(M^{(t)})\\\\ & \\mathrm{if}\\| M^{(t)} - M^{(t - 1)}\\| < \\epsilon \\mathrm{then}\\\\ & \\mid \\mathrm{break}\\\\ & \\mathrm{end} \\end{array}" + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "text", + "content": " \n5 \n6 \n7 \n8 end \n9 return " + }, + { + "bbox": [ + 50, + 88, + 286, + 227 + ], + "type": "inline_equation", + "content": "M^{(t)}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 254, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 277 + ], + "type": "text", + "content": "which are concatenated to predict the categories and layouts of new instances and outpaint the unobserved areas." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": "For scene layout expansion, given a partial input image with the partial scene graph and the corresponding layout, we use GNN [11] to reveal the scene layout " + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "inline_equation", + "content": "L_{exp}" + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": " of new instances. Specifically, we pass the shadow representations " + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{f}^u" + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": " from the unobserved instances through an encoder " + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "inline_equation", + "content": "E_h" + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": ", and then concatenate the output with the node features " + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "inline_equation", + "content": "f^h" + }, + { + "bbox": [ + 46, + 278, + 287, + 361 + ], + "type": "text", + "content": " of the masked instances from the encoder of GNN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 371, + 287, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 371, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 111, + 371, + 287, + 385 + ], + "type": "interline_equation", + "content": "\\hat {f} ^ {h} = \\operatorname {c o n c a t} \\left(f ^ {h}, \\mathrm {E} _ {\\mathrm {h}} \\left(\\bar {f} ^ {u}\\right)\\right) \\tag {17}", + "image_path": "4532781dd161ab638ff4ca6288d31052245c4cba43fe68077e6ce9564e6a8c6a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": "For layout to image, given an input image " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " with an expanded scene layout " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "L_{exp}" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": ", we use layout-to-image model [76] to perform layout to image conversion. Specifically, we pass the shadow representations " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "f^u" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " from the unobserved instances through an encoder " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "E_l" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": ", and then concatenate the output with scene layout features " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "f^l" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " of the expanded instance from the encoder of the layout-to-image module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 500, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 500, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\hat {f} ^ {l} = \\operatorname {c o n c a t} \\left(f ^ {l}, \\mathrm {E} _ {\\mathrm {l}} \\left(\\bar {f} ^ {u}\\right)\\right) \\tag {18}", + "image_path": "9660712ed1445ebda348f889e967a2ca26ebdc129a9ef3f87e3b9e6628c8405e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 525, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 596 + ], + "type": "text", + "content": "Note that we focus on the design of extracting shadow representation, so producing high-accuracy predictions for scene layout and high-quality image outputs is not within the main scope of this work. Thus, the graph-based and diffusion-based designs can be replaced with any scene layout expansion and layout-to-image modules if desirable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 229, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 229, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 229, + 624 + ], + "type": "text", + "content": "3.3. Stage 3: Alignment Discriminator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "In stage 2, we outpaint the partial input image " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " into the extended image " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "I_{E}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " with introduced instances. To enhance alignment between unobserved area and visible semantics, we adopt the pretrained instance shadow detector [63] to first extract the mask " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "M_E^u" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " of the unobserved area and its corresponding shadow from " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "I_{E}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ", and then feed it into a local shadow-instance alignment discriminator to ensure the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "generated alignment is visually real. This is calculated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 102, + 545, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 102, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 314, + 102, + 545, + 149 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {L D} V (L D) = \\frac {1}{2} \\mathbb {E} _ {x \\sim p _ {\\text {d a t a}} (x)} [ (L D (x) - b) ^ {2} ] \\tag {19} \\\\ + \\frac {1}{2} \\mathbb {E} _ {z \\sim p z (z)} [ (L D (M _ {E} ^ {u} \\odot G (z)) - a) ^ {2} ], \\\\ \\end{array}", + "image_path": "e5b976fbe837504ae39e2eff3075c8cac55bf1c38f4287332a4ae9428f36bdeb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 169, + 545, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 169, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 311, + 169, + 545, + 192 + ], + "type": "interline_equation", + "content": "\\min _ {G} V (G) = \\frac {1}{2} \\mathbb {E} _ {z \\sim p z (z)} [ (L D (M _ {E} ^ {u} \\odot G (z)) - c) ^ {2} ], \\tag {20}", + "image_path": "32c4c99eceb62a9722611fda93170db6bccf5f225bc089feb20d83f9ffd674da.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "inline_equation", + "content": "LD" + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": " denotes the local alignment discriminator. " + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": " denotes the generator for outpainting. " + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": " denote the ground truth real and fake labels, respectively. " + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": " denotes the value that " + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": " wants " + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "inline_equation", + "content": "LD" + }, + { + "bbox": [ + 304, + 200, + 545, + 247 + ], + "type": "text", + "content": " to believe for the fake data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": "Then, the mask " + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "inline_equation", + "content": "M_E" + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": " that consists of all semantics and their shadows in " + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "inline_equation", + "content": "I_E" + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": " is extracted, and we feed it into a global alignment discriminator to ensure the newly generated alignment is consistent with the existing ones. This is calculated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 314, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 314, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 312, + 314, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {G D} V (G D) = \\frac {1}{2} \\mathbb {E} _ {x \\sim p _ {\\text {d a t a}} (x)} [ (G D (x) - b) ^ {2} ] \\tag {21} \\\\ + \\frac {1}{2} \\mathbb {E} _ {z \\sim p _ {z} (z)} [ (G D (M _ {E} \\odot G (z)) - a) ^ {2} ], \\\\ \\end{array}", + "image_path": "20aef5be5ea0c9ff7c292385c62d34abfa26e32de1c611acd50d82a33b1ca492.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 380, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 380, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 311, + 380, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\min _ {G} V (G) = \\frac {1}{2} \\mathbb {E} _ {z \\sim p _ {z} (z)} [ (G D (M _ {E} \\odot G (z)) - c) ^ {2} ], \\tag {22}", + "image_path": "90ae28104951f2a435c0438fa58653445230a97924c87aad85e2bc797029d37a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 411, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 411, + 545, + 434 + ], + "type": "inline_equation", + "content": "GD" + }, + { + "bbox": [ + 304, + 411, + 545, + 434 + ], + "type": "text", + "content": " denotes the global alignment consistency discriminator." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": "In summary, shadow extraction, scene layout expansion, and layout-to-image module are end-to-end trained together, thus the visual authenticity loss may facilitate shadow representation learning. Therefore, the final objective function of our approach is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 345, + 504, + 545, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 504, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 345, + 504, + 545, + 518 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {1} L _ {t} + \\lambda_ {2} L _ {L - G A N} ^ {G} + \\lambda_ {3} L _ {G - G A N} ^ {G}, \\tag {23}", + "image_path": "dce2638e4e7cbae7e8f7c6570107af0ce675c2e998bab6b09b965d7c4923f4f0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "text", + "content": " denote the balancing weights. " + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "inline_equation", + "content": "L_{t}" + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "text", + "content": " denotes the loss adopted from the base model. " + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "inline_equation", + "content": "L_{L - GAN}^{G}" + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "inline_equation", + "content": "L_{G - GAN}^{G}" + }, + { + "bbox": [ + 304, + 527, + 545, + 575 + ], + "type": "text", + "content": " denote the alignment loss from the local alignment and global consistency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 586, + 441, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 441, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 441, + 599 + ], + "type": "text", + "content": "4. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 629 + ], + "type": "text", + "content": "This section provides details of the training settings and module implementations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Training Settings. Our approach is implemented using PyTorch [46] on four NVIDIA-RTX A6000 GPUs. During the training process, we adopt the Adam optimizer, 0.001 for the local discriminator and the global discriminator, and 0.01 for all modules in shadow representation extractors. The batch size is set to 32 and the size of the input images is " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". The training is manually stopped." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7854" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "Shadow Detection Models. We adopt the popular shadow segmentation method [77] as the shadow detector, which is able to segment the shadow areas in the image. Furthermore, we adopt the state-of-the-art instance shadow detection method [63] as our instance shadow detector, which is able to detect the shadow instances and their corresponding object in the image. However, it is unable to detect shadows cast by unobserved objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "text", + "content": "Alignment Discriminator. We adopt diffusion timestep-dependent discriminators from Diffusion-GAN [65]. It not only can minimize the divergence between real and diffused data at the end of the process, but also minimizes the divergence between the diffused real data distribution and the diffused generator distribution over several timesteps. In training, the balancing weight " + }, + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "text", + "content": " is set to be 0.5, " + }, + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 46, + 168, + 287, + 287 + ], + "type": "text", + "content": " are set to 0.25. Note that the architecture of the discriminator can be adjusted based on the specific type of layout-to-image module being implemented." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "text", + "content": "Scene Layout Expansion. We adopt GTwE [11] to predict the categories and layout of unobserved instances simultaneously. GTwE has the same aforementioned hyperparameters as [69], which are the attention hidden size " + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "inline_equation", + "content": "d_{atten} = 512" + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "text", + "content": ", feed forward size " + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "inline_equation", + "content": "d_{ff} = 2048" + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "text", + "content": ", multi-head number " + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "inline_equation", + "content": "n_{head} = 4" + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "text", + "content": ", and dropout " + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "inline_equation", + "content": "= 0.1" + }, + { + "bbox": [ + 46, + 287, + 286, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 286, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 467 + ], + "type": "text", + "content": "Layout to Image. We adopt LayoutDiffusion [76] for better outpainting the unknown areas with the condition of layouts predicted by GTwE. Note that LayoutDiffusion can not take the visible semantics as guidance to outpaint the unknown areas. Thus, we follow RePaint [38] to modify the standard denoising process for conditioning on the given image content. In each step, we sample the known region from the input and the outpaint painted part from the DDPM [18] output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 477, + 127, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 127, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 127, + 490 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 497, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 286, + 652 + ], + "type": "text", + "content": "In this section, we provide our experimental setups and both the objective and subjective results. Since we are not aware of any existing work that performs the same task as we do, we mainly focus on showing the promise of the proposed approach. Specifically, we evaluate the effectiveness of our approach by conducting experiments with an image outpainting pipeline proposed by [69]. Our goal is, again, to show the possibility of learning effective shadow representations thus enhancing complex scene extension, rather than trying to beat the state-of-the-art shadow detection and outpainting method. Other modules with the same functionality, provided they are end-to-end trainable, can be adopted in our approach to achieve potentially better performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 109, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 109, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 109, + 670 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "We evaluate our proposed approach on two common datasets, Visual Genome (VG) [29] and COCO-stuff [4], which gives an adequate amount of scene layouts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "Visual Genome [29] dataset collects 108,077 images with dense annotations of objects, attributes, and relationships. Following the setting of SG2Im [20], we divide the data into " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " for the train, val, and test set, respectively. We select the object and relationship categories occurring at least 2000 and 500 times in the train set, respectively, leaving 178 objects and 45 relationship types, and select the images with 3 to 30 bounding boxes and ignoring all small objects. Finally, the training/validation/test set will have 62565 / 5062 / 5096 images, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": "COCO-stuff [4] dataset augments a subset of the COCO dataset [35] with additional semantic categories. Thus, a total of 80 object categories (car, dog, etc.) and 91 semantic categories (sky, snow, etc.) are available, with " + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "inline_equation", + "content": "118\\mathrm{K} / 5\\mathrm{K}" + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": " annotated images for training/validation. Following the setting of SG2Im [20], the ground truth object coordinates in the images are utilized for constructing the synthetic relationships, a total of six relationships are considered: left of, right of, above, below, inside, and surrounding. In addition, we use images in the train and val set with 3 to 8 objects that cover more than " + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": " of the image. Finally, there are 25,210 train and 3,097 val images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 370, + 495, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 495, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 495, + 383 + ], + "type": "text", + "content": "5.2. Quantitative and Qualitative Result" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 396, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 646 + ], + "type": "text", + "content": "In this section, we evaluate the quantitative and qualitative effectiveness of our proposed approach to the image outpainting task by adopting it as a plug-in module of the SOTA method SGT [69]. The outpainting performance can be divided into two steps, in which the first one focuses on predicting the category and layout of the unobserved semantic and then the second one aims to generate the outpainting images. Therefore, we evaluate our approach on both the two steps. As can be seen from Tab. 1, by adopting our approach as a plug-in module, SGT achieves significant improvement in unobserved semantic prediction and its relevant relationship estimation. Although our target is to enhance only the semantic prediction, the accurately predicted semantics consequently strengthen the relationship estimation, thus leading to overall improvement on all metrics. Then, as can be seen from Tab. 2, we achieve considerable improvement in unobserved layout prediction, which also explicitly impacts the alignment with visible semantics in addition to the categories of instances. Moreover, in order to evaluate the quality of outpainting images, we introduce the FID metric and the results shown in Tab. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Then, the visual results are shown in Fig. 5 and Fig. 6, it can be found that our proposed approach is able to enhance the outpainting in the unobserved area, which is achieved by constraining both the shape and the semantic of the outpainting to be aligned with its visible shadows." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7855" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 108, + 131 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 108, + 131 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 108, + 131 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 108, + 131 + ], + "type": "image", + "image_path": "507c36b1fcd13bc8df4ee74a00f80144d6730383d156ee0c520f1ddf9c3fe9b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 131, + 108, + 191 + ], + "blocks": [ + { + "bbox": [ + 49, + 131, + 108, + 191 + ], + "lines": [ + { + "bbox": [ + 49, + 131, + 108, + 191 + ], + "spans": [ + { + "bbox": [ + 49, + 131, + 108, + 191 + ], + "type": "image", + "image_path": "8c8196d16f2b46be3ec0eeb4624f8b54332722bf9e5a137524fe7c1a8343e2b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 194, + 99, + 203 + ], + "lines": [ + { + "bbox": [ + 59, + 194, + 99, + 203 + ], + "spans": [ + { + "bbox": [ + 59, + 194, + 99, + 203 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 111, + 70, + 171, + 131 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 171, + 131 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 171, + 131 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 171, + 131 + ], + "type": "image", + "image_path": "f43c32b21a02dfa364d2470a0fbfab75b095fbdbee165a87f0e401404c6fe795.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 111, + 131, + 171, + 191 + ], + "blocks": [ + { + "bbox": [ + 111, + 131, + 171, + 191 + ], + "lines": [ + { + "bbox": [ + 111, + 131, + 171, + 191 + ], + "spans": [ + { + "bbox": [ + 111, + 131, + 171, + 191 + ], + "type": "image", + "image_path": "2fa230a843ace6fc2f17ef0ecc99a222c9218f4bcf3b49c3d4ea1732e1d19d95.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 194, + 146, + 201 + ], + "lines": [ + { + "bbox": [ + 135, + 194, + 146, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 194, + 146, + 201 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 172, + 70, + 232, + 130 + ], + "blocks": [ + { + "bbox": [ + 172, + 70, + 232, + 130 + ], + "lines": [ + { + "bbox": [ + 172, + 70, + 232, + 130 + ], + "spans": [ + { + "bbox": [ + 172, + 70, + 232, + 130 + ], + "type": "image", + "image_path": "69c261e30788716cbe0e14c52e8c77001baf68dceec6ba74d0df9906f4237f38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 172, + 131, + 232, + 191 + ], + "blocks": [ + { + "bbox": [ + 172, + 131, + 232, + 191 + ], + "lines": [ + { + "bbox": [ + 172, + 131, + 232, + 191 + ], + "spans": [ + { + "bbox": [ + 172, + 131, + 232, + 191 + ], + "type": "image", + "image_path": "8077ce50145808fd8a42b65b9c0dab38b19693f9ba4e41420b402b657d2a441c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 194, + 211, + 201 + ], + "lines": [ + { + "bbox": [ + 195, + 194, + 211, + 201 + ], + "spans": [ + { + "bbox": [ + 195, + 194, + 211, + 201 + ], + "type": "text", + "content": "SGT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 234, + 70, + 294, + 130 + ], + "blocks": [ + { + "bbox": [ + 234, + 70, + 294, + 130 + ], + "lines": [ + { + "bbox": [ + 234, + 70, + 294, + 130 + ], + "spans": [ + { + "bbox": [ + 234, + 70, + 294, + 130 + ], + "type": "image", + "image_path": "04c7263528c87473ffd15ed9a75e12714af8b43c8202a941d3e57692bbee9b5e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 234, + 131, + 294, + 191 + ], + "blocks": [ + { + "bbox": [ + 234, + 131, + 294, + 191 + ], + "lines": [ + { + "bbox": [ + 234, + 131, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 234, + 131, + 294, + 191 + ], + "type": "image", + "image_path": "3233d51603374a77d8811d016e9ed6b62a9b7ef4e192b55eeb17255f1d1454a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 194, + 275, + 201 + ], + "lines": [ + { + "bbox": [ + 254, + 194, + 275, + 201 + ], + "spans": [ + { + "bbox": [ + 254, + 194, + 275, + 201 + ], + "type": "text", + "content": "SGT+" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 295, + 70, + 356, + 130 + ], + "blocks": [ + { + "bbox": [ + 295, + 70, + 356, + 130 + ], + "lines": [ + { + "bbox": [ + 295, + 70, + 356, + 130 + ], + "spans": [ + { + "bbox": [ + 295, + 70, + 356, + 130 + ], + "type": "image", + "image_path": "ec9ef1f53292cf5b6da5472d7de385e6f737a2417e199f8229ae1c54bd0aebfa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 295, + 131, + 356, + 191 + ], + "blocks": [ + { + "bbox": [ + 295, + 131, + 356, + 191 + ], + "lines": [ + { + "bbox": [ + 295, + 131, + 356, + 191 + ], + "spans": [ + { + "bbox": [ + 295, + 131, + 356, + 191 + ], + "type": "image", + "image_path": "b635e0cdf52dce6a52db8b709ebbd1e9f788c66c9312176bd9a818bb319b4d50.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 194, + 346, + 203 + ], + "lines": [ + { + "bbox": [ + 306, + 194, + 346, + 203 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 346, + 203 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 358, + 70, + 418, + 130 + ], + "blocks": [ + { + "bbox": [ + 358, + 70, + 418, + 130 + ], + "lines": [ + { + "bbox": [ + 358, + 70, + 418, + 130 + ], + "spans": [ + { + "bbox": [ + 358, + 70, + 418, + 130 + ], + "type": "image", + "image_path": "532f11abfa6a50293b267351d27d21a785b1a6750caffdffd81743176170b5f9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 358, + 131, + 418, + 191 + ], + "blocks": [ + { + "bbox": [ + 358, + 131, + 418, + 191 + ], + "lines": [ + { + "bbox": [ + 358, + 131, + 418, + 191 + ], + "spans": [ + { + "bbox": [ + 358, + 131, + 418, + 191 + ], + "type": "image", + "image_path": "da36c49a84075952fdbe4373289d492c0d25b076f4a61e0bd25da2a4fee4aeb0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 194, + 394, + 201 + ], + "lines": [ + { + "bbox": [ + 383, + 194, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 383, + 194, + 394, + 201 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 420, + 70, + 481, + 130 + ], + "blocks": [ + { + "bbox": [ + 420, + 70, + 481, + 130 + ], + "lines": [ + { + "bbox": [ + 420, + 70, + 481, + 130 + ], + "spans": [ + { + "bbox": [ + 420, + 70, + 481, + 130 + ], + "type": "image", + "image_path": "2cadabe0f39a46fd72b415d45d6501c20348df6bf45b3e5f2a52c90df58f99ad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 420, + 131, + 481, + 191 + ], + "blocks": [ + { + "bbox": [ + 420, + 131, + 481, + 191 + ], + "lines": [ + { + "bbox": [ + 420, + 131, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 420, + 131, + 481, + 191 + ], + "type": "image", + "image_path": "1336f4d0364d8f32870075f36565c941a1ace692e719e68c9e27941dd01ded90.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 194, + 458, + 201 + ], + "lines": [ + { + "bbox": [ + 443, + 194, + 458, + 201 + ], + "spans": [ + { + "bbox": [ + 443, + 194, + 458, + 201 + ], + "type": "text", + "content": "SGT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 482, + 70, + 542, + 130 + ], + "blocks": [ + { + "bbox": [ + 482, + 70, + 542, + 130 + ], + "lines": [ + { + "bbox": [ + 482, + 70, + 542, + 130 + ], + "spans": [ + { + "bbox": [ + 482, + 70, + 542, + 130 + ], + "type": "image", + "image_path": "428b2aa356eb6e5ba77e39b097dd45826fc3dff01d53813f4dc3f2d29e55e53b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 482, + 131, + 542, + 191 + ], + "blocks": [ + { + "bbox": [ + 482, + 131, + 542, + 191 + ], + "lines": [ + { + "bbox": [ + 482, + 131, + 542, + 191 + ], + "spans": [ + { + "bbox": [ + 482, + 131, + 542, + 191 + ], + "type": "image", + "image_path": "e5dd73909fb4fedfc21c49b21081441f4cd044f8958cb0ad1971188edddad578.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 194, + 522, + 201 + ], + "lines": [ + { + "bbox": [ + 502, + 194, + 522, + 201 + ], + "spans": [ + { + "bbox": [ + 502, + 194, + 522, + 201 + ], + "type": "text", + "content": "SGT+" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 48, + 224, + 109, + 285 + ], + "blocks": [ + { + "bbox": [ + 69, + 203, + 523, + 214 + ], + "lines": [ + { + "bbox": [ + 69, + 203, + 523, + 214 + ], + "spans": [ + { + "bbox": [ + 69, + 203, + 523, + 214 + ], + "type": "text", + "content": "Figure 5. Visual results from SGT [69] on VG [29]. SGT+ denotes the result after adopting our approach as a plug-in module." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 224, + 109, + 285 + ], + "lines": [ + { + "bbox": [ + 48, + 224, + 109, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 109, + 285 + ], + "type": "image", + "image_path": "b23e861f38b84cb2780da119efff16778bada98d4207be19d8ebe53d3d4f6e5a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 48, + 286, + 109, + 346 + ], + "blocks": [ + { + "bbox": [ + 48, + 286, + 109, + 346 + ], + "lines": [ + { + "bbox": [ + 48, + 286, + 109, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 109, + 346 + ], + "type": "image", + "image_path": "87fe6d6696707b80d73c7896e3086d65be361b89004cbfb933a6d6b61bde8e31.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 349, + 99, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 349, + 99, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 349, + 99, + 357 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 111, + 224, + 171, + 285 + ], + "blocks": [ + { + "bbox": [ + 111, + 224, + 171, + 285 + ], + "lines": [ + { + "bbox": [ + 111, + 224, + 171, + 285 + ], + "spans": [ + { + "bbox": [ + 111, + 224, + 171, + 285 + ], + "type": "image", + "image_path": "8f7149815bd580316a0b206745b418ff03c326c0c59a20dad678950cd6eebb70.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 111, + 286, + 171, + 346 + ], + "blocks": [ + { + "bbox": [ + 111, + 286, + 171, + 346 + ], + "lines": [ + { + "bbox": [ + 111, + 286, + 171, + 346 + ], + "spans": [ + { + "bbox": [ + 111, + 286, + 171, + 346 + ], + "type": "image", + "image_path": "98c49d5d9b7ff32ae6c918ab48e23256daae9d6c0700383b5d7ad5724e13547e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 349, + 147, + 357 + ], + "lines": [ + { + "bbox": [ + 135, + 349, + 147, + 357 + ], + "spans": [ + { + "bbox": [ + 135, + 349, + 147, + 357 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 172, + 224, + 233, + 285 + ], + "blocks": [ + { + "bbox": [ + 172, + 224, + 233, + 285 + ], + "lines": [ + { + "bbox": [ + 172, + 224, + 233, + 285 + ], + "spans": [ + { + "bbox": [ + 172, + 224, + 233, + 285 + ], + "type": "image", + "image_path": "027acb619e653579a32eafa7585b2ff2c2d11b2f364ce0cdeb8d69d3274be6af.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 172, + 286, + 233, + 346 + ], + "blocks": [ + { + "bbox": [ + 172, + 286, + 233, + 346 + ], + "lines": [ + { + "bbox": [ + 172, + 286, + 233, + 346 + ], + "spans": [ + { + "bbox": [ + 172, + 286, + 233, + 346 + ], + "type": "image", + "image_path": "8b3dd7c59f6a32de1309e451a8203faadd1f0aabfade41ad7f767ecd8dd7214c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 349, + 211, + 357 + ], + "lines": [ + { + "bbox": [ + 195, + 349, + 211, + 357 + ], + "spans": [ + { + "bbox": [ + 195, + 349, + 211, + 357 + ], + "type": "text", + "content": "SGT" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 234, + 224, + 294, + 285 + ], + "blocks": [ + { + "bbox": [ + 234, + 224, + 294, + 285 + ], + "lines": [ + { + "bbox": [ + 234, + 224, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 234, + 224, + 294, + 285 + ], + "type": "image", + "image_path": "445068fe0bf90dde40aa9b1e3b67038c15dce848a0ad8f373b46528055481804.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 234, + 286, + 294, + 346 + ], + "blocks": [ + { + "bbox": [ + 234, + 286, + 294, + 346 + ], + "lines": [ + { + "bbox": [ + 234, + 286, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 234, + 286, + 294, + 346 + ], + "type": "image", + "image_path": "4fc3f40336c634df678f719af7420bbbc259c9a5fca6652f9fd5465c33bd42f4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 349, + 275, + 357 + ], + "lines": [ + { + "bbox": [ + 254, + 349, + 275, + 357 + ], + "spans": [ + { + "bbox": [ + 254, + 349, + 275, + 357 + ], + "type": "text", + "content": "SGT+" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 295, + 224, + 356, + 285 + ], + "blocks": [ + { + "bbox": [ + 295, + 224, + 356, + 285 + ], + "lines": [ + { + "bbox": [ + 295, + 224, + 356, + 285 + ], + "spans": [ + { + "bbox": [ + 295, + 224, + 356, + 285 + ], + "type": "image", + "image_path": "e108779d02e4474c5923be22214abca44dfbba988831b0ae5c695a2175dfa183.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 295, + 286, + 356, + 346 + ], + "blocks": [ + { + "bbox": [ + 295, + 286, + 356, + 346 + ], + "lines": [ + { + "bbox": [ + 295, + 286, + 356, + 346 + ], + "spans": [ + { + "bbox": [ + 295, + 286, + 356, + 346 + ], + "type": "image", + "image_path": "8b8da01b975ac40f746ecee93cafb33516c0cba21fb14a2c05898f6e32ab420a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 349, + 346, + 357 + ], + "lines": [ + { + "bbox": [ + 306, + 349, + 346, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 346, + 357 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 358, + 537, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 537, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 537, + 369 + ], + "type": "text", + "content": "Figure 6. Visual results from SGT [69] on COCO-stuff [4]. SGT+ denotes the result after adopting our approach as a plug-in module." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 358, + 224, + 418, + 285 + ], + "blocks": [ + { + "bbox": [ + 358, + 224, + 418, + 285 + ], + "lines": [ + { + "bbox": [ + 358, + 224, + 418, + 285 + ], + "spans": [ + { + "bbox": [ + 358, + 224, + 418, + 285 + ], + "type": "image", + "image_path": "11599049e0268ffa66c264212da71c11d14f78fdc5af01842d27a85f18fcac8d.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 358, + 286, + 418, + 346 + ], + "blocks": [ + { + "bbox": [ + 358, + 286, + 418, + 346 + ], + "lines": [ + { + "bbox": [ + 358, + 286, + 418, + 346 + ], + "spans": [ + { + "bbox": [ + 358, + 286, + 418, + 346 + ], + "type": "image", + "image_path": "bf40b14a6e16ac30b44cb143f9ff9083caafab3d745d5d1356439fd3200d3912.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 349, + 394, + 357 + ], + "lines": [ + { + "bbox": [ + 383, + 349, + 394, + 357 + ], + "spans": [ + { + "bbox": [ + 383, + 349, + 394, + 357 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 420, + 224, + 481, + 285 + ], + "blocks": [ + { + "bbox": [ + 420, + 224, + 481, + 285 + ], + "lines": [ + { + "bbox": [ + 420, + 224, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 420, + 224, + 481, + 285 + ], + "type": "image", + "image_path": "3a75af3b85a5c88b27f36b5946e3450eeddc31eba6d7a8d80e53417f94478f9c.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 420, + 286, + 481, + 346 + ], + "blocks": [ + { + "bbox": [ + 420, + 286, + 481, + 346 + ], + "lines": [ + { + "bbox": [ + 420, + 286, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 420, + 286, + 481, + 346 + ], + "type": "image", + "image_path": "4025e65f0e0ffd1af6154f94e4232d2311c5be23bad8a6b4a361904a6110c881.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 349, + 458, + 357 + ], + "lines": [ + { + "bbox": [ + 443, + 349, + 458, + 357 + ], + "spans": [ + { + "bbox": [ + 443, + 349, + 458, + 357 + ], + "type": "text", + "content": "SGT" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 482, + 224, + 542, + 285 + ], + "blocks": [ + { + "bbox": [ + 482, + 224, + 542, + 285 + ], + "lines": [ + { + "bbox": [ + 482, + 224, + 542, + 285 + ], + "spans": [ + { + "bbox": [ + 482, + 224, + 542, + 285 + ], + "type": "image", + "image_path": "25354537f43adf0eb68d0a2045ffb0617b034243cf88693b7c2ed78c34ce3158.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 482, + 286, + 542, + 346 + ], + "blocks": [ + { + "bbox": [ + 482, + 286, + 542, + 346 + ], + "lines": [ + { + "bbox": [ + 482, + 286, + 542, + 346 + ], + "spans": [ + { + "bbox": [ + 482, + 286, + 542, + 346 + ], + "type": "image", + "image_path": "34b2061649fd74bc43727a5fffc7a4d09d3256da8d4a73eb6f5c4b760c89acd1.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 349, + 522, + 357 + ], + "lines": [ + { + "bbox": [ + 502, + 349, + 522, + 357 + ], + "spans": [ + { + "bbox": [ + 502, + 349, + 522, + 357 + ], + "type": "text", + "content": "SGT+" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 47, + 380, + 266, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 266, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 266, + 392 + ], + "type": "text", + "content": "5.3. Multiple Instances Extension Comparison" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 47, + 402, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 287, + 462 + ], + "type": "text", + "content": "To verify whether our approach can enhance alignment with visible semantics in complex scenes, even with fewer shadows. We increase the number of instances requiring inference from unobserved areas to validate alignment between the extension and visible semantics shown in Fig 7." + } + ] + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 48, + 469, + 104, + 525 + ], + "blocks": [ + { + "bbox": [ + 48, + 469, + 104, + 525 + ], + "lines": [ + { + "bbox": [ + 48, + 469, + 104, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 104, + 525 + ], + "type": "image", + "image_path": "bae14e966ee2dce36cb9bdc43beefdf271106b9c9c46b7fb84d7b40286e3eb86.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 525, + 97, + 534 + ], + "lines": [ + { + "bbox": [ + 58, + 525, + 97, + 534 + ], + "spans": [ + { + "bbox": [ + 58, + 525, + 97, + 534 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 111, + 469, + 167, + 525 + ], + "blocks": [ + { + "bbox": [ + 111, + 469, + 167, + 525 + ], + "lines": [ + { + "bbox": [ + 111, + 469, + 167, + 525 + ], + "spans": [ + { + "bbox": [ + 111, + 469, + 167, + 525 + ], + "type": "image", + "image_path": "2168582660cdf738a174704976c9673350d3459bb8a92954094a6fa960d47f0f.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 525, + 228, + 533 + ], + "lines": [ + { + "bbox": [ + 119, + 525, + 228, + 533 + ], + "spans": [ + { + "bbox": [ + 119, + 525, + 228, + 533 + ], + "type": "text", + "content": "the number of instances increasing" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 171, + 469, + 227, + 525 + ], + "blocks": [ + { + "bbox": [ + 171, + 469, + 227, + 525 + ], + "lines": [ + { + "bbox": [ + 171, + 469, + 227, + 525 + ], + "spans": [ + { + "bbox": [ + 171, + 469, + 227, + 525 + ], + "type": "image", + "image_path": "e548ccc6727a76f4357bf38c6e44fef3fe66245475fea6b059cb4c27d33bfd68.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 231, + 469, + 286, + 525 + ], + "blocks": [ + { + "bbox": [ + 231, + 469, + 286, + 525 + ], + "lines": [ + { + "bbox": [ + 231, + 469, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 231, + 469, + 286, + 525 + ], + "type": "image", + "image_path": "f0a931508161131605b23e1103a3ba04c059821f901a7ede1f85a1983c1dd07f.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 49, + 534, + 104, + 588 + ], + "blocks": [ + { + "bbox": [ + 49, + 534, + 104, + 588 + ], + "lines": [ + { + "bbox": [ + 49, + 534, + 104, + 588 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 104, + 588 + ], + "type": "image", + "image_path": "8dd784e6cca593a8780cb53d15a002c3e52885add9e0ade594f3fbbb7b2ed178.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 589, + 82, + 596 + ], + "lines": [ + { + "bbox": [ + 70, + 589, + 82, + 596 + ], + "spans": [ + { + "bbox": [ + 70, + 589, + 82, + 596 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 111, + 534, + 168, + 589 + ], + "blocks": [ + { + "bbox": [ + 111, + 534, + 168, + 589 + ], + "lines": [ + { + "bbox": [ + 111, + 534, + 168, + 589 + ], + "spans": [ + { + "bbox": [ + 111, + 534, + 168, + 589 + ], + "type": "image", + "image_path": "acf32a73f6381223a58064831bbcade8710a7c8f2083b865b433f162d62cc2c5.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 590, + 228, + 597 + ], + "lines": [ + { + "bbox": [ + 119, + 590, + 228, + 597 + ], + "spans": [ + { + "bbox": [ + 119, + 590, + 228, + 597 + ], + "type": "text", + "content": "the number of instances increasing" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 171, + 534, + 227, + 589 + ], + "blocks": [ + { + "bbox": [ + 171, + 534, + 227, + 589 + ], + "lines": [ + { + "bbox": [ + 171, + 534, + 227, + 589 + ], + "spans": [ + { + "bbox": [ + 171, + 534, + 227, + 589 + ], + "type": "image", + "image_path": "c08bb8509a845af1422592eb1902b07bab48279a46365177bd6f349b03396401.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 231, + 534, + 286, + 596 + ], + "blocks": [ + { + "bbox": [ + 231, + 534, + 286, + 596 + ], + "lines": [ + { + "bbox": [ + 231, + 534, + 286, + 596 + ], + "spans": [ + { + "bbox": [ + 231, + 534, + 286, + 596 + ], + "type": "image", + "image_path": "44fd4c4c432db2f9cb07fc519f42ac0a36fe004768326725e2756d020a848963.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 599, + 286, + 620 + ], + "lines": [ + { + "bbox": [ + 47, + 599, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 286, + 620 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of multiple instances extension. Each image adds one more instance from left to right." + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": "As we can see, our method can effectively leverage shadow information in complex scenes to infer a scene layout that aligns well with the visible semantics. Specifically, instances that its shadow is not in the visible area can still align with visible semantics compared to the method without incorporating our approach, which ensures a harmonious generation of content with visible semantics." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 306, + 380, + 474, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 474, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 474, + 392 + ], + "type": "text", + "content": "5.4. Shadow Extension Comparison" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 305, + 399, + 546, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 546, + 483 + ], + "type": "text", + "content": "As seen in Fig. 5 and Fig. 6, our visual results demonstrate that outpainting with our module achieves excellent alignment with visible semantics. However, there is an additional issue we need to consider: the shadows of visible instances can also affect the overall shadow consistency of the image. Therefore, we have also conducted a comparison specifically addressing this concern, as shown in Fig. 8." + } + ] + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 307, + 488, + 365, + 547 + ], + "blocks": [ + { + "bbox": [ + 307, + 488, + 365, + 547 + ], + "lines": [ + { + "bbox": [ + 307, + 488, + 365, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 365, + 547 + ], + "type": "image", + "image_path": "582fe5eedb59f32383079296e69635e6b5df01697dcfdf72df7667594136a855.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 307, + 548, + 362, + 604 + ], + "blocks": [ + { + "bbox": [ + 307, + 548, + 362, + 604 + ], + "lines": [ + { + "bbox": [ + 307, + 548, + 362, + 604 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 362, + 604 + ], + "type": "image", + "image_path": "04e70a44dea66bf9fd6572afaa280f5252efb0d22468bc97a3dc9cc2d9b86a38.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 607, + 356, + 616 + ], + "lines": [ + { + "bbox": [ + 316, + 607, + 356, + 616 + ], + "spans": [ + { + "bbox": [ + 316, + 607, + 356, + 616 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 618, + 545, + 650 + ], + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 650 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison of shadow extension on the visible instance. Red boxes show the differences in the extended shadow of the visible instance." + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 366, + 488, + 424, + 547 + ], + "blocks": [ + { + "bbox": [ + 366, + 488, + 424, + 547 + ], + "lines": [ + { + "bbox": [ + 366, + 488, + 424, + 547 + ], + "spans": [ + { + "bbox": [ + 366, + 488, + 424, + 547 + ], + "type": "image", + "image_path": "b254c80040bed00a89c540b014e2b1bf421b37ecc6c0115df40394551ac70c3d.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 366, + 548, + 424, + 604 + ], + "blocks": [ + { + "bbox": [ + 366, + 548, + 424, + 604 + ], + "lines": [ + { + "bbox": [ + 366, + 548, + 424, + 604 + ], + "spans": [ + { + "bbox": [ + 366, + 548, + 424, + 604 + ], + "type": "image", + "image_path": "b6170703cad1030b41ce417daca3951292be7335d3a02b1e67cd3a494b2f32cc.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 607, + 402, + 615 + ], + "lines": [ + { + "bbox": [ + 390, + 607, + 402, + 615 + ], + "spans": [ + { + "bbox": [ + 390, + 607, + 402, + 615 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 426, + 488, + 484, + 547 + ], + "blocks": [ + { + "bbox": [ + 426, + 488, + 484, + 547 + ], + "lines": [ + { + "bbox": [ + 426, + 488, + 484, + 547 + ], + "spans": [ + { + "bbox": [ + 426, + 488, + 484, + 547 + ], + "type": "image", + "image_path": "d7ec74152898e0c99ac9dbe70f7514fadec60ec6cfd4db5cd62e2c35e9e9c770.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 426, + 548, + 483, + 604 + ], + "blocks": [ + { + "bbox": [ + 426, + 548, + 483, + 604 + ], + "lines": [ + { + "bbox": [ + 426, + 548, + 483, + 604 + ], + "spans": [ + { + "bbox": [ + 426, + 548, + 483, + 604 + ], + "type": "image", + "image_path": "481d9396ad31da0b6a3f649e651c6b3c48fb956af5d96965c5628913ec16232f.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 607, + 462, + 615 + ], + "lines": [ + { + "bbox": [ + 446, + 607, + 462, + 615 + ], + "spans": [ + { + "bbox": [ + 446, + 607, + 462, + 615 + ], + "type": "text", + "content": "SGT" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 485, + 488, + 542, + 547 + ], + "blocks": [ + { + "bbox": [ + 485, + 488, + 542, + 547 + ], + "lines": [ + { + "bbox": [ + 485, + 488, + 542, + 547 + ], + "spans": [ + { + "bbox": [ + 485, + 488, + 542, + 547 + ], + "type": "image", + "image_path": "9a69c1834585ff084ff8246727b99f0238787051ee1dc2de7c7224c098960e55.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 485, + 548, + 542, + 604 + ], + "blocks": [ + { + "bbox": [ + 485, + 548, + 542, + 604 + ], + "lines": [ + { + "bbox": [ + 485, + 548, + 542, + 604 + ], + "spans": [ + { + "bbox": [ + 485, + 548, + 542, + 604 + ], + "type": "image", + "image_path": "b508fc767830ffeaa28ec4cfd59e63dd13666870b3014468fc6ca0d3ec81bc3a.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 607, + 523, + 615 + ], + "lines": [ + { + "bbox": [ + 503, + 607, + 523, + 615 + ], + "spans": [ + { + "bbox": [ + 503, + 607, + 523, + 615 + ], + "type": "text", + "content": "SGT+" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_caption" + } + ], + "index": 78 + }, + { + "bbox": [ + 305, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 546, + 713 + ], + "type": "text", + "content": "As we can see, our shadow extension harmonizes the generated with the visible semantics. However, the method without incorporating our module randomly elongates or abruptly truncates shadows, resulting in a lack of global alignment among the semantic-shadow pairs in the scene." + } + ] + } + ], + "index": 81 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7856" + } + ] + } + ], + "index": 82 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 70, + 469, + 199 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 469, + 199 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 469, + 199 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 469, + 199 + ], + "type": "table", + "html": "
VG [29]COCO-stuff [4]
ObjectRelationshipObjectRelationship
rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑rAVG↓Hit@ 1/3↑
SGT10.5122.9 / 45.75.8136.9 / 69.410.5426.4 / 52.32.9225.7 / 66.4
+Fsp9.6124.6 / 47.05.2337.7 / 72.410.0127.0 / 53.42.8925.9 / 67.1
+Fillum9.2526.4 / 49.34.8238.2 / 73.19.5427.5 / 54.42.8227.0 / 67.2
+LGAN10.123.1 / 46.55.2437.4 / 71.310.4126.3 / 53.12.9825.9 / 66.9
+LGAN9.9723.4 / 46.25.4237.2 / 71.710.7526.7 / 52.92.9326.3 / 66.5
SGT+8.9327.9 / 50.44.5939.4 / 75.28.8328.4 / 55.32.7328.3 / 68.4
", + "image_path": "4f4d18e46029a3eacc59f5c5a8df48fe70ca12fa75c86233dfd75dac7a81c7a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 243, + 168, + 358 + ], + "blocks": [ + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation and ablation study of scene graph expansion on VG [29] and COCO-stuff [4] datasets. " + }, + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "type": "text", + "content": " denotes the higher is better, " + }, + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "type": "text", + "content": " the lower is better. SGT+ denotes the pipeline that adopts our approach as a plug-in module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 243, + 168, + 358 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 168, + 358 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 168, + 358 + ], + "type": "table", + "html": "
mIoUVGCOCO-stuff
SGT22.732.6
+Fsp25.434.8
+Fillum26.436.2
+LG-LGAN23.633.4
+LG-GAN24.933.1
SGT+28.438.1
", + "image_path": "77df863b8edd4217c8ac283fa2aba048c0f1ff20208ac848775c2843dbddcd27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 361, + 166, + 405 + ], + "lines": [ + { + "bbox": [ + 47, + 361, + 166, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 166, + 405 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with mIoU metric." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 171, + 243, + 291, + 358 + ], + "blocks": [ + { + "bbox": [ + 171, + 243, + 291, + 358 + ], + "lines": [ + { + "bbox": [ + 171, + 243, + 291, + 358 + ], + "spans": [ + { + "bbox": [ + 171, + 243, + 291, + 358 + ], + "type": "table", + "html": "
FIDVGCOCO-stuff
SGT20.1219.12
+Fsp19.8919.02
+Fillum19.9218.93
+LG-LGAN19.4818.60
+LG-GAN19.5218.59
SGT+19.4718.38
", + "image_path": "3248e33f61055a003b6a534dffc3abc3c27882791153319c94c4c76f3db26265.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 170, + 361, + 288, + 405 + ], + "lines": [ + { + "bbox": [ + 170, + 361, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 170, + 361, + 288, + 405 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation and ablation study on VG [29] and COCO-stuff [4] datasets, with FID metric." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 429, + 214, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 214, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 214, + 442 + ], + "type": "text", + "content": "5.5. User Studies & Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "type": "text", + "content": "User Studies. For image outpainting, any appropriate generation should be treated to be correct, thus without a specific target output. Therefore, in order to further evaluate the qualitative effectiveness of our approach, we conduct two user studies, where 91 users are involved to evaluate the outpainting correctness of our approach. In the first experiment, we send each user 20 randomized selected pairs of contents, in which the first one is the observed image with filling black in the unobserved area and the second one is the scene layout we predict to appear in the unobserved area, we then ask the users whether the predicted semantic could appear in the unobserved area. Finally, " + }, + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 46, + 449, + 287, + 628 + ], + "type": "text", + "content": " of the selected pairs are scored to be true, which denotes that our proposed approach is able to predict the appropriate scene layout for the unobserved semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "In the second experiment, we send each user 20 randomized selected pairs of images, where one of them comes from SGT and the other one comes from adopting our approach as a plug-in module, we then ask the users which one is better in visual authenticity. Finally, images from ours achieve " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " better chosen, which shows the significant quality enhancement produced by our approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 245, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 547, + 449 + ], + "type": "text", + "content": "Ablation Study. We evaluate the individual effectiveness of the semantic removal, the illumination concatenation, the local alignment discriminator, and the global consistency discriminator in this experiment. The results show in Tab. 1, Tab. 2 and Tab. 3. It can be seen that the most significant enhancement in scene layout prediction comes from illumination concatenation, which shows the importance of incorporating a comprehensive illumination understanding to tackle the challenges of shadow extraction for obtaining accurate predictions. Meanwhile, the semantic removal enhances the accuracy notably, which shows that the shadow-irrelevant semantic information in the shadow areas leads to unexpected and noisy shadow representations. And the two discriminators achieve considerable enhancement in FID, which shows the importance of ensuring the local alignment between the unobserved semantics and its visible shadows and the consistency among all semantic-shadow pairs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 462, + 379, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 379, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 379, + 475 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 674 + ], + "type": "text", + "content": "In this paper, we proposed a novel approach that aims to perform image outpainting by utilizing shadow information, which has never been explored. This is accomplished by extracting instance-level shadows cast by the unobserved areas from the input image. Then, the instance-level shadow representations are concatenated to infer the scene layout of each instance and outpaint the unobserved areas. Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. Subsequently, extensive objective and subjective experiments are conducted which strongly proves the proposed approach successfully captures semantic and geometry information in the shadows. In future work, we will endeavor to learn shadow representations from dynamic environments and incorporate more tasks like novel view synthesis and pedestrian trajectory prediction into the unified framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 680, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 680, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgement The research was supported by the Shanghai Committee of Science and Technology, China (Grant No.23ZR1423500)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7857" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Abdelrehim Ahmed and Aly Farag. Shape from shading for hybrid surfaces. In 2007 IEEE International Conference on Image Processing, pages II-525. IEEE, 2007. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Abdelrehim H Ahmed and Aly A Farag. Shape from shading under various imaging conditions. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "text", + "content": "[3] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 259 + ], + "type": "text", + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 315 + ], + "type": "text", + "content": "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[6] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 418 + ], + "type": "text", + "content": "[7] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 462 + ], + "type": "text", + "content": "[8] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(9): 10850-10869, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "type": "text", + "content": "[9] R. Cucchiara, C. Grana, M. Piccardi, and A. Prati. Detecting moving objects, ghosts, and shadows in video streams. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(10):1337-1342, 2003. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[11] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[12] Aleksandrs Ecins, Cornelia Fermüller, and Yiannis Aloimonos. Shadow free segmentation in still images using local density measure. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-8, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[13] Jiacheng Fan, Min Chen, Jinqiu Mo, Shigang Wang, and Qinghua Liang. Variational formulation of a hybrid perspective shape from shading model. The Visual Computer, pages 1-14, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[14] Silvano Galliani, Yong Chul Ju, Michael Breuß, and Andrés Bruhn. Generalised perspective shape from shading in spherical coordinates. In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 222-233. Springer, 2013. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[16] Yinsai Guo, Hang Yu, Liyan Ma, Liang Zeng, and Xiangfeng Luo. Thfe: A triple-hierarchy feature enhancement method for tiny boat detection. Engineering Applications of Artificial Intelligence, 123:106271, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "text", + "content": "[17] Yinsai Guo, Hang Yu, Shaorong Xie, Liyan Ma, Xinzhi Cao, and Xiangfeng Luo. Dsca: A dual semantic correlation alignment method for domain adaptation object detection. Pattern Recognition, page 110329, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 262, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 315 + ], + "type": "text", + "content": "[19] Weiqiang Jin, Biao Zhao, Hang Yu, Xi Tao, Ruiping Yin, and Guizhong Liu. Improving embedded knowledge graph multi-hop question answering by introducing relational chain reasoning. Data Mining and Knowledge Discovery, 37(1):255-288, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 317, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 358 + ], + "type": "text", + "content": "[20] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "text", + "content": "[21] Imran N. Junejo and Hassan Foroosh. Estimating geotemporal location of stationary cameras using shadow trajectories. In Computer Vision – ECCV 2008, pages 318–331, Berlin, Heidelberg, 2008. Springer Berlin Heidelberg. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 458 + ], + "type": "text", + "content": "[22] Asaf Karnieli, Ohad Fried, and Yacov Hel-Or. Deepshadow: Neural shape from shadow. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part II, pages 415–430. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[23] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "text", + "content": "[24] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[25] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4019–4028, 2021. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[26] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, 2021. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[27] Jihyun Kim, Seong-Hun Jeong, Kyeongbo Kong, and SukJu Kang. An unified framework for language guided image completion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2568-2578, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Daehyeon Kong, Kyeongbo Kong, Kyunghun Kim, Sung-Jun Min, and Suk-Ju Kang. Image-adaptive hint generation" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7858" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "via vision transformer for outpainting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3572-3581, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "text", + "content": "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[30] Jean-François Lalonde, Alexei A. Efros, and Srinivasa G. Narasimhan. Estimating natural illumination from a single outdoor image. In 2009 IEEE 12th International Conference on Computer Vision, pages 183-190, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "text", + "content": "[31] Kyoung Mu Lee and C-C Jay Kuo. Shape from shading with a generalized reflectance map model. Computer vision and image understanding, 67(2):143-160, 1997. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[32] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16221-16230, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "text", + "content": "[33] Pengbo Li, Hang Yu, Xiangfeng Luo, and Jia Wu. Lgmnn: A local and global aware memory-based graph neural network for fraud detection. IEEE Transactions on Big Data, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "type": "text", + "content": "[34] Jian Liang, Chenfei Wu, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. Advances in Neural Information Processing Systems, 35:15420-15432, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 466 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "type": "text", + "content": "[36] Ruoshi Liu, Sachit Menon, Chengzhi Mao, Dennis Park, Simon Stent, and Carl Vondrick. What you can reconstruct from a shadow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17059-17068, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[37] Chia-Ni Lu, Ya-Chu Chang, and Wei-Chen Chiu. Bridging the visual gap: Wide-range image blending. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-851, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "text", + "content": "[38] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[39] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1680-1691, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[40] Pascal Mamassian. Shape from Shadows, pages 724-725. Springer, 2014. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[41] S. Nadimi and B. Bhanu. Physical models for moving shadow and object detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(8):1079-1087, 2004. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 546, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 218 + ], + "type": "text", + "content": "[43] Takahiro Okabe, Imari Sato, and Yoichi Sato. Attached shadow coding: Estimating surface normals from shadows under unknown reflectance and lighting conditions. In 2009 IEEE 12th International Conference on Computer Vision, pages 1693-1700. IEEE, 2009. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 546, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 546, + 263 + ], + "type": "text", + "content": "[44] Alexandros Panagopoulos, Dimitris Samaras, and Nikos Paragios. Robust shadow and illumination estimation using a mixture model. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 651-658, 2009." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 265, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 546, + 319 + ], + "type": "text", + "content": "[45] Alexandros Panagopoulos, Chaohui Wang, Dimitris Samaras, and Nikos Paragios. Simultaneous cast shadows, illumination and geometry inference using hypergraphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(2):437-449, 2013. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 546, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 546, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 546, + 385 + ], + "type": "text", + "content": "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 431 + ], + "type": "text", + "content": "[47] Jiayan Qiu, Xinchao Wang, Pascal Fua, and Dacheng Tao. Matching seqlets: An unsupervised approach for locality preserving sequence matching. IEEE transactions on pattern analysis and machine intelligence, 43(2):745-752, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 433, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 546, + 477 + ], + "type": "text", + "content": "[48] Jiayan Qiu, Xinchao Wang, Stephen J Maybank, and Dacheng Tao. World from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8493-8504, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 546, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 546, + 533 + ], + "type": "text", + "content": "[49] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Hallucinating visual instances in total absentia. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 264-282. Springer, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 534, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 546, + 578 + ], + "type": "text", + "content": "[50] Jiayan Qiu, Yiding Yang, Xinchao Wang, and Dacheng Tao. Scene essence. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8322-8333, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 634 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 546, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 666 + ], + "type": "text", + "content": "[52] Mark Sabini and Gili Rusak. Painting outside the box: Image outpainting with gans. arXiv preprint arXiv:1808.08483, 2018.2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 714 + ], + "type": "text", + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7859" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "text", + "content": "[54] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "text", + "content": "[55] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "text", + "content": "[56] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "text", + "content": "[57] Cheng-Yo Tan, Chiao-An Yang, Shang-Fu Chen, Meng-Lin Wu, and Yu-Chiang Frank Wang. Robust image outpainting with learnable image margins. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1159-1163. IEEE, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 380 + ], + "type": "text", + "content": "[58] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "text", + "content": "[59] Pinzhuo Tian and Hang Yu. Can we improve meta-learning model in few-shot learning by aligning data distributions? Knowledge-Based Systems, 277:110800, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[60] Kushagra Tiwary, Tzofi Klinghoffer, and Ramesh Raskar. Towards learning neural representations from shadows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 300-316. Springer, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "text", + "content": "[61] Anwaar Ulhaq, Naveed Akhtar, and Ganna Pogrebna. Efficient diffusion models for vision: A survey. arXiv preprint arXiv:2210.09292, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "text", + "content": "[62] Guohui Wang and Jin Cheng. Three-dimensional reconstruction of hybrid surfaces using perspective shape from shading. Optik, 127(19):7740-7751, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "text", + "content": "[63] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–14, 2022. 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "text", + "content": "[64] Yi Wang, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Wide-context semantic image extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1399-1408, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[65] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-GAN: Training GANs with diffusion. In The Eleventh International Conference on Learning Representations, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[66] Chenyu Wu, Srinivasa G Narasimhan, and Branislav Jaramaz. A multi-image shape-from-shading framework for near-lighting perspective endoscopes. International Journal of Computer Vision, 86:211-228, 2010. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 653 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[67] Lin Wu and Xiaochun Cao. Geo-location estimation from two shadow trajectories. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 585-590, 2010. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[68] Xiaoyu Xu, Jiayan Qiu, Xinchao Wang, and Zhou Wang. Relationship spatialization for depth estimation. In European Conference on Computer Vision, pages 615-637. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "text", + "content": "[69] Chiao-An Yang, Cheng-Yo Tan, Wan-Cyuan Fan, Cheng-Fu Yang, Meng-Lin Wu, and Yu-Chiang Frank Wang. Scene graph expansion for semantics-guided image outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15626, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "type": "text", + "content": "[70] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Yingxia Shao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 285, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 341 + ], + "type": "text", + "content": "[71] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv., 2023. Just Accepted. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 341, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 545, + 396 + ], + "type": "text", + "content": "[72] Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. Distilling knowledge from graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7074-7083, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 396, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 545, + 441 + ], + "type": "text", + "content": "[73] Zongxin Yang, Jian Dong, Ping Liu, Yi Yang, and Shuicheng Yan. Very long natural scenery image prediction by outpainting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10560-10569, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 442, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 485 + ], + "type": "text", + "content": "[74] Kai Yao, Penglei Gao, Xi Yang, Jie Sun, Rui Zhang, and Kaizhu Huang. Outpainting by queries. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 486, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 486, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 486, + 545, + 542 + ], + "type": "text", + "content": "[75] Guangcong Zheng, Shengming Li, Hui Wang, Taiping Yao, Yang Chen, Shouhong Ding, and Xi Li. Entropy-driven sampling and training scheme for conditional diffusion generation. In European Conference on Computer Vision, pages 754-769. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "type": "text", + "content": "[76] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 598, + 545, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 598, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 308, + 598, + 545, + 653 + ], + "type": "text", + "content": "[77] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 2, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7860" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_content_list.json b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f0b5ad583b873311e2e442dbeb10836e914bf2e5 --- /dev/null +++ b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_content_list.json @@ -0,0 +1,1947 @@ +[ + { + "type": "text", + "text": "Shadows Don't Lie and Lines Can't Bend! \nGenerative Models don't know Projective Geometry...for now", + "text_level": 1, + "bbox": [ + 174, + 130, + 794, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ayush Sarkar*1 Hanlin Mai*1 Amitabh Mahapatra*1 Svetlana Lazebnik", + "bbox": [ + 169, + 202, + 790, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "D.A. Forsyth1 Anand Bhattachad2", + "bbox": [ + 352, + 222, + 616, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Illinois Urbana-Champaign 2Toyota Technological Institute at Chicago", + "bbox": [ + 138, + 239, + 831, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://projective-geometry.github.io/", + "bbox": [ + 256, + 257, + 705, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a816408b7c8608b4adca114e38103aee9b4dd5338bca4b84e38dff871712a89.jpg", + "image_caption": [ + "Generated Image" + ], + "image_footnote": [], + "bbox": [ + 81, + 289, + 243, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14825c8c4e005b253de46315b0323fc37245196dfed9c0645eea5309dd1c48e1.jpg", + "image_caption": [ + "Shadow Errors" + ], + "image_footnote": [], + "bbox": [ + 243, + 289, + 406, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ecda4d8f0f8638623144ae24289797545afd64b5e07f3aa042ae87ee1725af02.jpg", + "image_caption": [ + "Detected Shadow Errors" + ], + "image_footnote": [], + "bbox": [ + 406, + 289, + 570, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4df22eacc828692978e1171d6b0d76404f455b8219035fc769521e211018f2e2.jpg", + "image_caption": [ + "Vanishing Point Errors" + ], + "image_footnote": [], + "bbox": [ + 568, + 289, + 732, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e15621b0bb55c86cdfa4356ce456c94c285d2f99d2eada6041052bf96d05e50.jpg", + "image_caption": [ + "Detected Perspective Errors", + "Figure 1. The first column presents visually compelling AI-generated images. However, a closer examination reveals fundamental inconsistencies, such as those in shadow alignment (second column) and vanishing point accuracy (fourth column). Our model's analysis, shown in the third and fifth columns, detects these shadow and perspective geometry errors. We show that these errors are systematic and can be used to identify generated images." + ], + "image_footnote": [], + "bbox": [ + 732, + 289, + 893, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 633, + 313, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative models can produce impressively realistic images. This paper demonstrates that generated images have geometric features different from those of real images. We build a set of collections of generated images, prequalified to fool simple, signal-based classifiers into believing they are real. We then show that prequalified generated images can be identified reliably by classifiers that only look at geometric properties. We use three such classifiers. All three classifiers are denied access to image pixels, and look only at derived geometric features. The first classifier looks at the perspective field of the image, the second looks at lines detected in the image, and the third looks at relations between detected objects and shadows. Our procedure detects generated images more reliably than SOTA local signal based detectors,", + "bbox": [ + 73, + 666, + 472, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for images from a number of distinct generators. Saliency maps suggest that the classifiers can identify geometric problems reliably. We conclude that current generators cannot reliably reproduce geometric properties of real images.", + "bbox": [ + 496, + 636, + 893, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 724, + 632, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Both StyleGAN [22-24] and diffusion models [36-38] are renowned for generating images that are strikingly similar to real-world photos and consistently fool people. But, as we show, generated images have distinctive geometric features, likely from a failure to fully capture projective geometry.", + "bbox": [ + 496, + 750, + 893, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bhattad et al. [5, 6], Chen et al. [10], Du et al. [13], Zhan et al. [49], and have shown generative models implicitly capture the complex scene properties, including normals, depth, albedo, and support relations. These works suggest these models \"understand\" geometry, which would", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution", + "bbox": [ + 94, + 887, + 200, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28140", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "be useful for rendering 3D scenes. Our detailed, population-level analysis of generated images suggests generative models [1, 3, 4, 12, 32] cannot fully translate this \"understanding\" into accurate geometry. Specifically, we demonstrate that generative models produce images with lines that differ from those of real images (likely due to problems aligning vanishing points); that generative models produce images with perspective fields that are unlike those of real images; and that object-shadow relations in generated images differ reliably from those in real images. We use advanced pretrained models (Line Segment Detection [31]; Perspective Fields [21]; and PointNet [34]) that inspect geometric representations to distinguish between real and generated images. We rely on derived geometry cues from SOTA methods as manual image analysis and explicit geometry-based rules like drawing lines perspective lines are not scalable or automatable.", + "bbox": [ + 75, + 90, + 472, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To guarantee the accuracy of our findings, we follow a strict process of data curation. This involves using a controlled pixel-level classifier to filter out any biases related to color, texture, and local features within our test set. This precision in data selection is crucial to isolate and accurately assess the inconsistencies in projective geometry and illumination present in generated images. By carefully screening the data, we can ensure that our results are not obscured by common artifacts that are usually found in generated images. This enhances the reliability of our conclusions. Our contributions are:", + "bbox": [ + 75, + 332, + 472, + 496 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Unearthing Geometric Discrepancies: We present a comprehensive analysis that goes beyond existing literature to both demonstrate and quantify geometric discrepancies produced by current generative models.", + "- Scalable, Data-Driven Approach: We introduce a scalable, data-driven approach by utilizing three distinct projective geometry cues to detect geometric inaccuracies.", + "- Robust and Generalizable: We demonstrate robustness and generalizability by effectively identifying geometric errors across a wide array of images from recent generative models, including those with the latest time stamps.", + "- Broadening the Scope of Model Assessment: Our approach can be used as an alternative method for evaluating models: do they get projective geometry right?" + ], + "bbox": [ + 76, + 500, + 470, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 723, + 218, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "**Generative Models:** The advancement of generative models, particularly in creating visually realistic images, marks a significant milestone in computer vision. Pioneering efforts by Karras et al. [22-24] with StyleGAN, and the emergence of diffusion models [36-38], have set new benchmarks in realism. These models, used in diverse fields from art to data augmentation, have yet to fully grasp the nuances of projective geometry, which is the focus of our analysis, primarily using open diffusion models.", + "bbox": [ + 75, + 750, + 472, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometric Understanding in Generative Models: While", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "studies like Bhattachad et al. [6], Du et al.[13], Chen et al. [10] and Zhan et al. [49] demonstrate these models' potential in understanding scene geometry, our work diverges by scrutinizing the generated images themselves, examining their adherence to the principles of projective geometry and illumination, rather than analyzing learned features.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detecting Generated Images: The realism of modern generative models has made image forensics increasingly challenging. Traditional methods focused on detecting synthetic images using signals like resampling artifacts [33] and JPEG quantization [2]. Kee et al. [25] introduced a geometric technique for detecting shadow inconsistencies, paralleling our pursuit of physical realism. However, our work extends beyond identifying photo manipulation to evaluating the overall perspective geometry and illumination consistency in images from generative models.", + "bbox": [ + 496, + 186, + 895, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zhang et al.'s work [14] focuses on detecting AI-generated images using diverse generative models and online training for future model adaptation. Our research, in contrast, assesses the projective geometry in these images, examining their ability to render scenes with accurate perspective and illumination. Boháček et al. [7], while detecting geometric inconsistencies related to shadows, align with our interest in physical realism. However, we delve deeper, thoroughly evaluating perspective geometry and illumination in generative models for a more comprehensive understanding of their geometric accuracy.", + "bbox": [ + 496, + 342, + 895, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rise of generative methods has steered image forensics towards using discriminative methods to detect synthetic content [8, 15, 19, 42, 43, 48, 50]. These advancements align with our objective of analyzing the physical and geometrical congruence of generated images. However, our work goes a step further by critically assessing whether generative models fundamentally understand and accurately replicate projective geometry, rather than simply distinguishing between real and synthetic images. This deeper level of analysis aims to unveil the intricacies and limitations of current models in faithfully rendering geometrically coherent images.", + "bbox": [ + 496, + 513, + 893, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation Metrics: Traditional metrics like the Inception Score (IS) [39] and Fréchet Inception Distance (FID) [18] focus on pixel-level fidelity. The emergence of CLIP-based scores [17, 35] and DIRE [45] offers a semantic perspective. In contrast, our approach, distinct in its focus on perspective geometry and illumination consistency, seeks to ensure comprehensive realism, bridging the gap between visual and physical authenticity.", + "bbox": [ + 496, + 684, + 893, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies like Davidsonian Scene Graph [11] and ImagenHub [26] address fine-grained evaluation inconsistencies, while the HEIM benchmark [27] assesses models across multiple aspects. Our work complements these by providing an in-depth evaluation of the physical and geometric realism of images generated by state-of-the-art models.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28141", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background on Projective Geometry", + "text_level": 1, + "bbox": [ + 76, + 89, + 408, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Projective geometry is a mathematical framework that enables the accurate representation of three-dimensional spaces in two-dimensional images. It provides the rules for perspective, which are crucial for creating realistic scenes with depth and spatial orientation [16]. In this section, we will examine the common inconsistencies that may arise during image synthesis according to projective geometry. Our evaluation framework is intended to detect and measure these discrepancies, which are essential for evaluating the realism and physical plausibility of generated images.", + "bbox": [ + 75, + 114, + 470, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inconsistent Vanishing Points. Vanishing points are fundamental to capturing the essence of perspective in images. They should align with the direction of parallel lines converging at a distance. Generated images often exhibit inconsistencies where these lines do not meet at the correct vanishing points, leading to a distorted sense of perspective.", + "bbox": [ + 75, + 267, + 470, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lighting and Shadow Inconsistencies. Accurate shadows are essential for reinforcing the position and shape of objects within a scene. Discrepancies in shadow direction, length, and softness can indicate a misalignment with the scene's light sources, disrupting the image's three-dimensionality.", + "bbox": [ + 75, + 359, + 470, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scale Discrepancies. The principle of size constancy dictates that objects of the same size should appear smaller as their distance from the observer increases. Generated images sometimes fail to maintain this scaling, resulting in a compromised depth perception.", + "bbox": [ + 75, + 435, + 470, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distortion of Geometric Figures. Geometric figures should maintain their shape when projected onto the image plane, barring intentional perspective distortion. Errors in this projection can result in circles appearing as ellipses or squares as trapezoids, indicating a flawed perspective rendering.", + "bbox": [ + 75, + 510, + 470, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth Cues. Depth perception in images is conveyed through cues such as overlapping, texture gradients, and relative size. Misrepresentation of these cues can lead to an unnatural spatial arrangement that the human eye can readily detect as artificial.", + "bbox": [ + 75, + 587, + 470, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our evaluation framework, detailed in the subsequent sections, is designed to rigorously test generated images against these projective geometry principles. While a comprehensive evaluation of projective geometry would consider all the aforementioned inconsistencies, our framework prioritizes the detection of inconsistent vanishing points and lighting and shadow inconsistencies. These elements are particularly telling indicators of an image's projective geometry realism and are often the most challenging for generative models to replicate accurately.", + "bbox": [ + 75, + 662, + 470, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Dataset Curation", + "text_level": 1, + "bbox": [ + 76, + 829, + 246, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our data curation process is carefully designed to distinguish between real images and generated images from several generative models. This process includes models that the", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "classifier has not seen during its training. Another important goal is to ensure our prequalifier effectively identifies images with recent timestamps. This helps prevent the classifier from relying on cues specific to the dataset, such as whether the image is from the training distribution or not, rather than determining whether the image is real or generated.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Real Images", + "text_level": 1, + "bbox": [ + 498, + 191, + 632, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We experiment with a diverse set of images, including:", + "bbox": [ + 496, + 214, + 861, + 229 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Indoor Scenes: A collection of 400,000 interior images featuring a variety of furniture arrangements and lighting conditions, sourced from LSUN [46]. Specifically, we used 100,000 images each from Bedroom, Dining Room, Kitchen, and Living Room categories.", + "(b) Outdoor Scenes: A dataset of 125,000 outdoor scenes with varying landscapes and urban settings, sourced from Berkeley Deep Drive 100K [47] and Mapillary Vistas [30]. We selected images that represent a wide range of weather conditions and times of day.", + "(c) Combination of Indoor and Outdoor scenes: We also analyzed a combination of above indoor and outdoor scenes to assess the performance on a more diverse dataset.", + "(d) Recent Timestamp Images: A curated test set of 500 indoor and 500 outdoor images with timestamps ranging from May 2023 to March 2024. These images were collected from various social media platforms and online sources to ensure our classifier's ability to handle recent real-world data and that the models are not obscured because of any data-source-specific biases." + ], + "bbox": [ + 496, + 229, + 893, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Image Captions", + "text_level": 1, + "bbox": [ + 498, + 540, + 658, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the ViT-bigG-14/laion2b_s39b_b160k model [20] and the BLIP model [28] in succession, to generate refined captions for real images. These models were chosen for their state-of-the-art performance in image captioning tasks. The ViT model uses a Vision Transformer architecture, while BLIP employs a multimodal pre-training approach. By using common captions, we ensure a fair evaluation of the projective geometry in generated images.", + "bbox": [ + 496, + 563, + 893, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.3. Generated Images", + "text_level": 1, + "bbox": [ + 498, + 694, + 678, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We generate images from Stable Diffusion XL v1.0 [32], Kandinsky-v3 [3], DeepFloyd IF v1.0 [12], and PixArt- $\\alpha$ v1.0 [9]. We use the same caption from the real images to generate these images with the default settings of each generative model.", + "bbox": [ + 496, + 717, + 893, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.4. Robust Prequalifier", + "text_level": 1, + "bbox": [ + 498, + 801, + 687, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to identify challenging images that a signal-based classifier may struggle to differentiate. Therefore, we need to develop a robust prequalifier. This is to ensure that our results are not affected by any false data signals. Thus, we aim to eliminate all potential factors that may influence", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28142", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d5509c7fce4085d9b8ab2ccf28b73f647ce10076f5823fffe5e1c03d71811dac.jpg", + "table_caption": [ + "Table 1. Statistical overview of the Data Curation and Filtering Process: We present the distribution of real and generated images for indoor and outdoor datasets. The ResNet50 Prequalifier helps us curate datasets, creating an 'Unconfident Set' for images with low classifier certainty and a 'Misclassified Subset' for images incorrectly labeled by the classifier. These rigorously curated sets are instrumental for our subsequent analysis, concentrating on projective geometry while mitigating the influence of signal cues. The datasets for each prequalifier—indoor, outdoor, and combined—are prepared separately to tailor the models to their specific contexts." + ], + "table_footnote": [], + "table_body": "
Indoor RealIndoor GeneratedOutdoor RealOutdoor GeneratedCombined RealCombined Generated
Total Images400,000400,000125,000125,000525,000525,000
Training and Test Sets
Training Set Size75,00075,00025,00025,000100,000100,000
Validation Set Size10,00010,0005,0005,00015,00015,000
Test Set Size315,000315,00095,00095,000410,000410,000
Post ResNet50 Prequalifier
Unconfident Set232131384014441018443922540
Misclassified Subset10399575660944323800825
", + "bbox": [ + 94, + 162, + 867, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the geometry cues we obtain because of signal weirdness in the generated images. To accomplish this, we begin by training signal-based classifiers using a vanilla ResNet-50 and primarily concentrate on identifying instances that prove challenging for these signal-based classifiers.", + "bbox": [ + 75, + 323, + 470, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We trained prequalifiers on three distinct settings - indoor scenes, outdoor scenes and a combination of the two. Each prequalifier was trained on a dataset consisting of real images and images generated by one of the four models: Stable Diffusion XL, Kandinsky-v3, DeepFloyd IF, and PixArt- $\\alpha$ . While all four types of prequalifiers performed well on their respective test sets, we found that Kandinsky exhibited superior generalization capabilities when evaluated on images generated by other models or on recent timestamp images. The fundamental experiment conducted to assess the generalizability of different generators is covered in Figure 9 in the Supplementary Material.", + "bbox": [ + 75, + 410, + 470, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For indoor scenes, our Kandinsky-based prequalifier achieves an Area Under the Curve (AUC) of 0.99 on its test set with an accuracy of 97.43 and maintained high performance on images generated by other models such as Stable Diffusion XL and PixArt- $\\alpha$ , achieving AUCs of 0.97 and 0.98 respectively. Furthermore, our Kandinsky-trained prequalifier meant for both indoor and outdoor settings combined showed robust performance on recent timestamp generated indoor and outdoor images, achieving AUC scores of 0.90, 0.95, and 0.72 on images generated by Stable Diffusion XL, PixArt- $\\alpha$ , and DeepFloyd IF, respectively.", + "bbox": [ + 75, + 602, + 470, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given their robustness and strong generalizability, we selected the Kandinsky-trained prequalifiers as the basis for training our derived geometry classifiers on Kandinsky-generated images. This choice ensures that our classifiers effectively look at geometric discrepancies and can handle various generated images. This enhances the reliability of our approach and supports our conclusion that the classifiers are not affected by spurious signal artifacts.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.5. Final Test Sets", + "text_level": 1, + "bbox": [ + 498, + 323, + 648, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We categorize our test sets into three groups based on the accuracy of the prequalifier: easy, unconfident, and misclassified. The easy test set includes images with $100\\%$ accuracy, indicating that the prequalifier can reliably distinguish between real and generated images. The unconfident test set includes images where the prequalifier performs at chance level, suggesting that it struggles to make confident predictions. Finally, the misclassified test set includes images where the prequalifier makes completely wrong predictions, either classifying real images as generated or vice versa. A summary of this split for Kandinsky-v3 is provided in Tab. 1.", + "bbox": [ + 496, + 347, + 893, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach primarily evaluates the \"hard set\" (unconfident and misclassified) for geometric and shadow inconsistencies, assessing adherence to projective geometry principles. This ensures rigorous testing of generative models' ability to reproduce geometric correctness and photometric accuracy, beyond surface-level or signal details.", + "bbox": [ + 496, + 513, + 893, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to note that projective geometry inconsistencies are prevalent in generated images but often go undetected by conventional methods. Our models, trained on geometric abstractions and projective cues, can identify subtle but critical inaccuracies that texture artifacts cannot explain. This distinction is critical, as it allows us to rigorously test our models, which, unlike the prequalifier, do not have direct access to the images.", + "bbox": [ + 496, + 604, + 893, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a suite of models to capture different facets of projective geometry, trained on datasets emphasizing geometric consistency and photometric accuracy. By combining their strengths, we comprehensively assess the quality of generated images and provide insights into the limitations and potential improvements of generative models from a projective geometry perspective.", + "bbox": [ + 496, + 724, + 893, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Analyzing Projective Geometry", + "text_level": 1, + "bbox": [ + 498, + 844, + 785, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For our analysis, we rely on three geometry cues – object shadow, line segments, and perspective fields. We train", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28143", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aaa6eefb1677ccc2dd06713ec5425e9e5f852d5983b54c9cb8963bc7331de60a.jpg", + "image_caption": [ + "Figure 2. We train three classifiers to identify discrepancies in projective geometry. These classifiers are trained using derived geometry cues such as object-shadow associations (left), perspective fields (middle), and line segments (right) without looking at image intensity. We use the ResNet architecture for Object-Shadow and Perspective Fields, and PointNet for Line Segments to process unordered data sets." + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 330, + 318 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f19d44804ed4395cdf3bef925bfbc08e184cae01bae3034fe056c47aac52e121.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 95, + 591, + 318 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/29de772733505b4f9d7466825c0dfb740ac54c9d81497a5a29d605655c8f5150.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 95, + 887, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "three separate classifiers solely based on these derived geometry cues, without utilizing any pixel information. The training process for each classifier is described below, and a schematic pipeline of these classifiers can be found in Fig. 2.", + "bbox": [ + 75, + 376, + 470, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Object-Shadow Cues", + "text_level": 1, + "bbox": [ + 76, + 446, + 274, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our first model examines object-shadow relationships to address illumination. Shadows follow projective geometry principles, and inconsistencies can reveal generated images.", + "bbox": [ + 75, + 469, + 468, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a pretrained object-shadow instance detection model [44] to identify shadows and geometric heuristics to evaluate their plausibility given the objects and their orientation. A ResNet50 classifier is trained on binary masks of object and shadow instances to score the consistency of shadows with objects. Images with implausible shadows are marked as likely generated.", + "bbox": [ + 75, + 515, + 470, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Perspective Field Cues", + "text_level": 1, + "bbox": [ + 76, + 631, + 287, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our framework's second model uses Perspective Fields [21], vector fields encoding the spatial orientation of pixels relative to vanishing points and the horizon, to assess projective geometry. We generate these fields from single images using a pretrained model.", + "bbox": [ + 75, + 654, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train a ResNet50 classifier on the Perspective Fields to differentiate real from generated images by focusing on projective geometry anomalies. The classifier evaluates the consistency of these fields with projective geometry principles, scoring images on their geometric plausibility. This method enables precise evaluation of projective geometry, enhancing the detection of subtle inconsistencies.", + "bbox": [ + 75, + 729, + 470, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Line Segment Cues", + "text_level": 1, + "bbox": [ + 76, + 845, + 261, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method also assesses projective geometry in generated images by identifying key structural lines using Deep", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LSD [31]. These lines indicate adherence to perspective rules. We then train a PointNet-like architecture [34] to classify images based on line segment patterns, differentiating real from generated images.", + "bbox": [ + 496, + 376, + 893, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PointNet's flexibility in handling unordered data makes it suitable for analyzing line segments without pre-sorting. The model assigns scores representing the likelihood of an image being real based on the spatial arrangement of its lines. Analyzing these scores reveals the model's ability to detect subtle discrepancies in line arrangements, which often indicate a generated image.", + "bbox": [ + 496, + 436, + 893, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 556, + 614, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on our prequalification analysis, we use Kandinsky as the primary source for our generated data to train our classifiers. We evaluated our three geometry-derived cues classifiers and analyzed their ability to generalize to other generated images that were not used during training. In addition, we performed a GradCam analysis [40] to identify any potential geometric discrepancies in images.", + "bbox": [ + 496, + 580, + 893, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Classifiers Results", + "text_level": 1, + "bbox": [ + 500, + 696, + 674, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 shows ROC curves for each of our methods on indoor, outdoor and combined (indoor+outdoor) scenes. In each case, classifiers are trained on images that are not prequalified and tested on prequalified scenes, meaning that performance estimates are biased low — likely training on prequalified data would lead to even more accurate classification. Each classifier is effective, with AUCs ranging from 0.72 to 0.97. Recall these classifiers see only derived geometric features and do not see the image itself.", + "bbox": [ + 496, + 719, + 893, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative examples using Grad-CAM appear in Fig. 5 and Fig. 4. Notice how images that might be acceptable to a line analysis often fail a shadow analysis. Fig. 6 shows", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28144", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/942f752e9f85aaf49934d8be94d45f48ece85e605fa1f214c084d53aeb57f06a.jpg", + "image_caption": [ + "(a) Easy Test Set (Indoor)" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 272, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4bf31823834387f2944a1e72c846da56a23f0ac817d8a152e0596f19057a64b1.jpg", + "image_caption": [ + "(b) Recent Timestamp Set (Indoor)" + ], + "image_footnote": [], + "bbox": [ + 289, + 90, + 480, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70fd52c9bc63195813b2679a2ce2d677a821558cbdefab8a3fa930033cbfaa74.jpg", + "image_caption": [ + "(c) Unconfident Test Set (Indoor)" + ], + "image_footnote": [], + "bbox": [ + 488, + 90, + 679, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f909efb9d38490dc91d45b0ae140c5ca15290232852dffa7e323b57445490ace.jpg", + "image_caption": [ + "(d) Misclassified Test Set (Indoor)" + ], + "image_footnote": [], + "bbox": [ + 696, + 90, + 887, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52292a87664af03964d6120e3917278f6594e68fbd20712ffbc9c44e466b1779.jpg", + "image_caption": [ + "(e) Easy Test Set (Outdoor)" + ], + "image_footnote": [], + "bbox": [ + 80, + 227, + 272, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3a00626eff5653177a518adf3b8cc3c10a14ef8f16067dc9d94489dd6259b6b.jpg", + "image_caption": [ + "(f) Recent Timestamp Set (Outdoor)" + ], + "image_footnote": [], + "bbox": [ + 289, + 227, + 480, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06a598880c2d633342c2dca6f2025b5bfde862e43a50d93ae53595f8ac87e21d.jpg", + "image_caption": [ + "(g) Unconfident Test Set (Outdoor)" + ], + "image_footnote": [], + "bbox": [ + 488, + 227, + 679, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08a9e15ffe4e014ee735bb2971300a9842b272034883d19d9cdf590e05ebd46c.jpg", + "image_caption": [ + "(h) Misclassified Test Set ((Outdoor))" + ], + "image_footnote": [], + "bbox": [ + 694, + 227, + 887, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/72ba2fe439ea365d2202157e111aa867d01f9eff53ff5c9a84616cbac7aa4024.jpg", + "image_caption": [ + "(i) Easy Test Set (Combined)" + ], + "image_footnote": [], + "bbox": [ + 81, + 364, + 272, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f01e36b4bb3fb13a25d434f5d1443ff28a95aaaef12e55bf41f693f79464eb9.jpg", + "image_caption": [ + "(j) Recent Timestamp Set (Combined)" + ], + "image_footnote": [], + "bbox": [ + 287, + 364, + 478, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fe1a2ef18e55c66c4a54cbf3dd9eaf37b31efeafbf68c36f028649201aedea3e.jpg", + "image_caption": [ + "Figure 3. ROC Curves Assessing Projective Geometry Cues in Generated Images trained on Kandinsky-v3. We trained separate models for indoor scenes (top row), outdoor scenes (middle row), and a combination of indoor and outdoor scenes (last row). All our derived geometry cues classifiers are trained without looking at image intensity information and can reliably detect projective geometry errors. The recent timestamp test set (second column) confirms that these models are robust. We find hard examples using a prequalifier trained on image pixels. Our derived geometry cues consistently show high AUC for finding projective geometry errors on hard test sets – the last two columns – unconfident and misclassified test sets. For the unconfident test set, where the prequalifier has an AUC of 0.51 (c), 0.51 (g), and 0.49 (h) for indoor, outdoor, and combined partition, our classifiers can still accurately identify the generated images with high AUCs – 0.82 from line segments in the indoor set, 0.84 from perspective fields in the outdoor set, and 0.80 from perspective field cues and object shadows in the combined set. Similarly, for the misclassified test set, where the prequalifier has an AUC of 0.00, as it should, our classifiers remain reliable with AUC up to 0.82. We conclude that generated images contain geometric structures not seen in real images, and these structures very reliably identify generated images by only looking at derived geometry cues." + ], + "image_footnote": [], + "bbox": [ + 485, + 364, + 679, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04e09062dca53bb2e4e90640ff2bf16401b5fa436e2d21c599e71052569f5f6b.jpg", + "image_caption": [ + "(k) Unconfident Test Set (Combined)", + "(1) Misclassified Test Set (Combined)" + ], + "image_footnote": [], + "bbox": [ + 691, + 364, + 883, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "examples to emphasize this point.", + "bbox": [ + 76, + 666, + 300, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7. Other Generators Evaluated", + "text_level": 1, + "bbox": [ + 76, + 694, + 339, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our classifiers do not see pixels, but derived geometric features. This means that one could expect a form of generalization across generators. We illustrate that this generalization occurs - ROC curves in Fig. 7 demonstrate that classifiers trained to distinguish Kandinsky images from real images can also reliably distinguish Stable Diffusion XL [3], DeepFloyd [12] and PixArt- $\\alpha$ [9] from the open-source domain. Additionally, we assess the efficacy of our models against images from proprietary generators such as OpenAI's Dalle-3[4] and Adobe's Firefly [1], representing some of the most advanced tools in image generation. Finally, we show we can detect composite made by a recent SOTA method [29]", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by looking at Object-Shadow cues in the supplement.", + "bbox": [ + 500, + 666, + 849, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "8. Discussion", + "text_level": 1, + "bbox": [ + 500, + 694, + 611, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have shown that generated images can be reliably distinguished from real images by looking only at derived geometric cues. This is likely because image generators do not fully implement the geometry one observes in real images. Producing accurate perspective geometry or accurate shadow geometry requires very tight coordination of detailed information over very long spatial scales. Our results, together with the notorious tendency of face image generators to award subjects' left and right earlobes of different shapes, suggest that doing so is beyond the capacity of current generators. Our findings have significant implications for the development of image generation models, as the inability to", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28145", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f03b11fe97913453204c040f1c223e16097ce8acd125ea5e11147387c9de8fb.jpg", + "image_caption": [ + "Generated Image" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 196, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/755700a65a7865b9578bbff74918413aadbb17b88241014f0a13be9f28b4f13e.jpg", + "image_caption": [ + "Shadow Errors" + ], + "image_footnote": [], + "bbox": [ + 197, + 90, + 310, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7e5e9612ef2d9b8a94c246a55a418a468e454602e08c2cb1a6709eea17d3694.jpg", + "image_caption": [ + "Object-Shadow (OS)" + ], + "image_footnote": [], + "bbox": [ + 312, + 90, + 426, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e161973637eee040d3749a00fbf692ad2f3ec09b10138c220522f0df99880f0.jpg", + "image_caption": [ + "OS GradCam" + ], + "image_footnote": [], + "bbox": [ + 428, + 90, + 542, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d677452df3a49e08fba7d4dd7b817719efb6ce92e03d25ae259265c8d935c590.jpg", + "image_caption": [ + "VP Errors" + ], + "image_footnote": [], + "bbox": [ + 542, + 90, + 653, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e240728e1277b5454f1d0d465048a064968d64d7695ba8d8564e8cfffcfca7d0.jpg", + "image_caption": [ + "Perspective Fields (PF)" + ], + "image_footnote": [], + "bbox": [ + 653, + 90, + 771, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a72a1de97e8ce46d5dcf5aaa3e914efbb23a3edf003ed0ad11a3f57aca727cf.jpg", + "image_caption": [ + "PF GradCam" + ], + "image_footnote": [], + "bbox": [ + 772, + 90, + 885, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b29ce764b8a4b55151a3d798bb1be69f2d35e04f66643c2f82e825a3400b105.jpg", + "image_caption": [ + "Figure 4. Grad-CAM applied to our Object-Shadow and Perspective Field classifiers reveals that the high AUCs in Fig. 3 are based on real geometric errors in indoor scenes generated by Kandinsky, Stable Diffusion XL and Dalle-3 shown in each row respectively. The second and fifth columns highlight shadow and vanishing point errors, respectively. The third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) identifies diagnostic areas for synthetic generation, such as inconsistent shadow directions (in all three rows), mismatched shadow lengths (second row). The sixth column shows Perspective Fields [21], and Grad-CAM applied to our Perspective Fields classifier (last column) reveals geometric errors in all three rows, particularly at ceilings and side walls, with noticeable errors also present in window grills in the first and second rows.", + "Generated Image", + "Figure 5. Grad-CAM results for outdoor scenes generated by Kandinsky, Stable Diffusion XL, and Adobe Firefly, shown in each row respectively. The second column highlights shadow errors, while the third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) reveals incorrect shadow shapes in the first and second rows, with shadows on the right-side pedestrians pointing in a different direction than those on the left. The fifth column shows vanishing point errors, and the sixth column presents Perspective Fields [21]. Grad-CAM applied to our Perspective Fields classifier (last column) confirms large perspective distortions on building facades and road markings, corroborating the vanishing point errors in the fifth column." + ], + "image_footnote": [], + "bbox": [ + 81, + 484, + 196, + 750 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46613e0f5ced8ad5cdbedc76e3e6de7be3eabc6e875f80cfb3b6090bbcf6579d.jpg", + "image_caption": [ + "Shadow Errors" + ], + "image_footnote": [], + "bbox": [ + 196, + 484, + 310, + 750 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0b7d98494813662f22bf3115a91e994a8cf20c7131b4da498c66861f61c54df.jpg", + "image_caption": [ + "Object-Shadow (OS)" + ], + "image_footnote": [], + "bbox": [ + 312, + 484, + 426, + 750 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8608d1484e24da2cb303234a8ed8563efa991b9221f09dd796661ca2eec3f844.jpg", + "image_caption": [ + "OS GradCam" + ], + "image_footnote": [], + "bbox": [ + 428, + 484, + 542, + 750 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eff68bc213fb46bce0f720a6a348c4fd4b5ab6690e3968091978327ca827b305.jpg", + "image_caption": [ + "VP Errors" + ], + "image_footnote": [], + "bbox": [ + 540, + 484, + 653, + 750 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd26a9c9015c1dd42bdb005ca3b7b31fdf66ca3f7d31eac38eed3745704f8b3c.jpg", + "image_caption": [ + "Perspective Fields" + ], + "image_footnote": [], + "bbox": [ + 653, + 484, + 771, + 750 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdfb7cd164f2dcdc88ae6a5169c3f92dc9c096b33ec2514a718f3335bca33c0f.jpg", + "image_caption": [ + "PF GradCam" + ], + "image_footnote": [], + "bbox": [ + 772, + 484, + 885, + 750 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28146", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fbd23d4df6975b08322542fc25e3291fc55374ebf81eb28e7e94978e7f5cbe69.jpg", + "image_caption": [ + "Generated Image" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 196, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b56195086b42baf9bf96beaa67027d1b66f01aff2c5c7d061b0a03a50f69c20f.jpg", + "image_caption": [ + "Shadow Errors" + ], + "image_footnote": [], + "bbox": [ + 197, + 90, + 426, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ea3fbeca0ff02014050f679365c6846d0740185e234e1fa3f116e12533f60a7.jpg", + "image_caption": [ + "Object-Shadow (OS)", + "OS GradCam" + ], + "image_footnote": [], + "bbox": [ + 428, + 90, + 540, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfe1453f9333e422e720876b5a23fd3a2a4ad141b9edf68270eec72282f0096c.jpg", + "image_caption": [ + "VP Errors" + ], + "image_footnote": [], + "bbox": [ + 540, + 90, + 653, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f41c45cf67d1cce83f52983cd097933a0d53cb1bfb50900d05a756b331bf4c89.jpg", + "image_caption": [ + "Perspective Fields (PF)" + ], + "image_footnote": [], + "bbox": [ + 653, + 90, + 771, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f86dd4917360c31ad438575844df4c2323858ffbceabd4f525aef4b91ed155e.jpg", + "image_caption": [ + "PF GradCam" + ], + "image_footnote": [], + "bbox": [ + 772, + 90, + 883, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f03e24a7fae66711fc8a2eab520ebd124703c5d72c8d377fc380287a85158ad.jpg", + "image_caption": [ + "Figure 6. Our projective geometry classifiers identify distinct types of problems in generated images. The top row presents an example that was classified as real by the Object-Shadow classifier but correctly identified as generated by the Perspective Fields classifier. While the shadow cast by the person appears realistic, the Perspective Fields Grad-CAM highlights the problematic geometry of the shelf on the top left. In contrast, the bottom row shows an example that was correctly identified as generated by the Object-Shadow classifier but misclassified as real by the Perspective Fields classifier. Although the perspective effects in the image appear plausible, the Grad-CAM weights correctly reveal that the two chairs are casting shadows from different light sources, indicating inconsistency in scene's illumination." + ], + "image_footnote": [], + "bbox": [ + 81, + 383, + 233, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3953b76abbd53bed9dae5fa5e41e8acdc2e2f691a617452e97f9c47210a6a623.jpg", + "image_caption": [ + "(a) Dall-E 3 (Indoor)", + "(f) Dall-E 3 (Outdoor)", + "Figure 7. Our model trained on Kandinsky-v3 generalizes in detecting projective geometry errors from images generated by various unseen models. We evaluate the model's performance on test sets generated by Dall-E 3 (a), DeepFloyd (b), Adobe Firefly (c), PixArt-α (d), and Stable Diffusion XL or SDXL (e) using the same text prompts from the 'misclassified' Kandinsky-v3 generated test set. The top row shows indoor scenes and the last row shows outdoor scenes. Our model generalizes to all evaluated generators except DeepFloyd. However, DeepFloyd-generated images can be reliably detected when the model is trained on a DeepFloyd-generated training set, but it shows poor generalization capabilities to other generators compared to Kandinsky." + ], + "image_footnote": [], + "bbox": [ + 81, + 496, + 233, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b62b039f15e4ac996405a4c3c00a030b92b852097867c2d2baa0160032439527.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 383, + 403, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6596c2f96e1841b52913430e72cb49ec4de3aa4cbe2b153a71b1fabab1b7e7ec.jpg", + "image_caption": [ + "(b) DeepFloyd (Indoor)", + "(g) DeepFloyd (Outdoor)" + ], + "image_footnote": [], + "bbox": [ + 243, + 494, + 395, + 588 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1241234b0e8d266d2a52ac7be29c687173b7e26af57407cf3501a32f750cd13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 383, + 575, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7739ab8650e5921e348f21eb161b8dce924dfd70c5ac2b7f66e17d4e7a0e8ab2.jpg", + "image_caption": [ + "(c) Adobe Firefly (Indoor)", + "(h) Adobe Firefly (Outdoor)" + ], + "image_footnote": [], + "bbox": [ + 406, + 494, + 560, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a1637bf9feef300a5872265c306f1daa0ba5b2ff248361b28234e3ccf047723.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 383, + 732, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfd3c4558800f71d84e2bdf0d815ae8181b8f7664870c6695cf6ae52030bc9d9.jpg", + "image_caption": [ + "(d) $\\mathrm{PixArt - }\\alpha$ (Indoor)", + "(i)PixArt- $\\alpha$ (Outdoor)" + ], + "image_footnote": [], + "bbox": [ + 571, + 494, + 723, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2051042f8a4d44a2fe8c89f549da0c1d2de0a1038ed8e6563bd5e60acdf4370.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 383, + 888, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4542a2733bf4625fd8ccaf3672ff13fdfc3fd03c6c77c5bda9db5a1a888aaba0.jpg", + "image_caption": [ + "(e) SDXL (Indoor)", + "(j) SDXL (Outdoor)" + ], + "image_footnote": [], + "bbox": [ + 733, + 494, + 885, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "accurately replicate projective geometry extends across various state-of-the-art models, indicating a widespread issue rather than a problem specific to a particular generator.", + "bbox": [ + 75, + 710, + 470, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We speculate that fixing this difficulty requires structural innovation in the generator, rather than simply exposing the generator to more data. The complex interplay of light, shadows, and perspective in real-world scenes may necessitate novel approaches to modeling and encoding spatial relationships within the generator's architecture. Potential avenues for improvement could include incorporating explicit geometric reasoning or developing new loss functions that prioritize the consistency of projective geometry. Fur", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "thermore, our work highlights the importance of developing robust evaluation metrics for image generation models that assess the geometric coherence of generated images. By tackling these challenges, we could create image-generation models that more faithfully capture the complex geometric relationships in real-world scenes.", + "bbox": [ + 496, + 710, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 819, + 651, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper is based on work supported in part by the National Science Foundation under Grant No. 2106825 and a gift from Boeing Corporation.", + "bbox": [ + 496, + 844, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28147", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adobe. Firefly. https://www.adobe.com/sensei/generative-ai/firefly.html, 2023. Accessed: 2023-11. 2,6", + "[2] Shruti Agarwal and Hany Farid. Photo forensics fromJPEG dimples. In 2017 IEEE workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2017. 2", + "[3] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report, 2023. 2, 3, 6", + "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 2, 6", + "[5] Anand Bhattachad and D.A. Forsyth. Stylitgan: Prompting stylegan to generate new illumination conditions. In arXiv, 2023. 1", + "[6] Anand Bhattachad, Daniel McKee, Derek Hoiem, and DA Forsyth. Stylegan knows normal, depth, albedo, and more. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2", + "[7] Matyáš Boháček and Hany Farid. A geometric and photometric exploration of gan and diffusion synthesized faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-883, 2023. 2", + "[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 2", + "[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3, 6", + "[10] Yida Chen, Fernanda Viégas, and Martin Wattenberg. Beyond surface statistics: Scene representations in a latent diffusion model. arXiv preprint arXiv:2306.05720, 2023. 1, 2", + "[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-image generation. arXiv preprint arXiv:2310.18235, 2023. 2", + "[12] Deep Floyd. Iterative filter. https://github.com/deep-floyd/IF, 2023. Accessed: 2023-11. 2, 3, 6", + "[13] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023. 1, 2", + "[14] David C. Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 382-392, 2023. 2", + "[15] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] David A Forsyth and Jean Ponce. Computer vision: a modern approach. prentice hall professional technical reference, 2002. 3", + "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2", + "[19] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In ECCV, 2018. 2", + "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3", + "[21] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 2, 5, 7, 16, 17", + "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2", + "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.", + "[24] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2", + "[25] Eric Kee, James F O'Brien, and Hany Farid. Exposing photo manipulation with inconsistent shadows. ACM Transactions on Graphics (ToG), 32(3):1-12, 2013. 2", + "[26] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 2", + "[27] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, Minguk Kang, Taesung Park, Jure Leskovec, Jun-Yan Zhu, Li Fei-Fei, Jiajun Wu, Stefano Ermon, and Percy Liang. Holistic evaluation of text-to-image models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28148", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3", + "[29] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6", + "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 5000–5009, 2017. 3", + "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5", + "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3", + "[33] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 53(2):758-767, 2005. 2", + "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 5", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022.", + "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2", + "[39] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 128(2):336-359, 2019. 5", + "[41] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18310-18319, 2023. 11", + "[42] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In ICCV, 2019. 2", + "[43] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2, 12", + "[44] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3259-3273, 2022. 5, 7, 16, 17", + "[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22445-22455, 2023. 2", + "[46] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 3", + "[47] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3", + "[48] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2", + "[49] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 1, 2", + "[50] Peng Zhou, Xintong Han, Vlad I Morariu, and Larry S Davis. Learning rich features for image manipulation detection. In CVPR, 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 727 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28149", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_model.json b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e28997677e2fc92edba36e37559c457fd71f0fa5 --- /dev/null +++ b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_model.json @@ -0,0 +1,2926 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.131, + 0.795, + 0.177 + ], + "angle": 0, + "content": "Shadows Don't Lie and Lines Can't Bend! \nGenerative Models don't know Projective Geometry...for now" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.203, + 0.792, + 0.221 + ], + "angle": 0, + "content": "Ayush Sarkar*1 Hanlin Mai*1 Amitabh Mahapatra*1 Svetlana Lazebnik" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.223, + 0.617, + 0.239 + ], + "angle": 0, + "content": "D.A. Forsyth1 Anand Bhattachad2" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.24, + 0.832, + 0.258 + ], + "angle": 0, + "content": "1University of Illinois Urbana-Champaign 2Toyota Technological Institute at Chicago" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.258, + 0.706, + 0.275 + ], + "angle": 0, + "content": "https://projective-geometry.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.29, + 0.245, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.545, + 0.209, + 0.557 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.29, + 0.408, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.545, + 0.365, + 0.556 + ], + "angle": 0, + "content": "Shadow Errors" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.29, + 0.571, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.545, + 0.551, + 0.556 + ], + "angle": 0, + "content": "Detected Shadow Errors" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.29, + 0.733, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.545, + 0.71, + 0.557 + ], + "angle": 0, + "content": "Vanishing Point Errors" + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.29, + 0.895, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.545, + 0.884, + 0.557 + ], + "angle": 0, + "content": "Detected Perspective Errors" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.566, + 0.895, + 0.623 + ], + "angle": 0, + "content": "Figure 1. The first column presents visually compelling AI-generated images. However, a closer examination reveals fundamental inconsistencies, such as those in shadow alignment (second column) and vanishing point accuracy (fourth column). Our model's analysis, shown in the third and fifth columns, detects these shadow and perspective geometry errors. We show that these errors are systematic and can be used to identify generated images." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.635, + 0.314, + 0.651 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.667, + 0.473, + 0.879 + ], + "angle": 0, + "content": "Generative models can produce impressively realistic images. This paper demonstrates that generated images have geometric features different from those of real images. We build a set of collections of generated images, prequalified to fool simple, signal-based classifiers into believing they are real. We then show that prequalified generated images can be identified reliably by classifiers that only look at geometric properties. We use three such classifiers. All three classifiers are denied access to image pixels, and look only at derived geometric features. The first classifier looks at the perspective field of the image, the second looks at lines detected in the image, and the third looks at relations between detected objects and shadows. Our procedure detects generated images more reliably than SOTA local signal based detectors," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.895, + 0.698 + ], + "angle": 0, + "content": "for images from a number of distinct generators. Saliency maps suggest that the classifiers can identify geometric problems reliably. We conclude that current generators cannot reliably reproduce geometric properties of real images." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.725, + 0.633, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.826 + ], + "angle": 0, + "content": "Both StyleGAN [22-24] and diffusion models [36-38] are renowned for generating images that are strikingly similar to real-world photos and consistently fool people. But, as we show, generated images have distinctive geometric features, likely from a failure to fully capture projective geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Bhattad et al. [5, 6], Chen et al. [10], Du et al. [13], Zhan et al. [49], and have shown generative models implicitly capture the complex scene properties, including normals, depth, albedo, and support relations. These works suggest these models \"understand\" geometry, which would" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.202, + 0.9 + ], + "angle": 0, + "content": "*equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28140" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.332 + ], + "angle": 0, + "content": "be useful for rendering 3D scenes. Our detailed, population-level analysis of generated images suggests generative models [1, 3, 4, 12, 32] cannot fully translate this \"understanding\" into accurate geometry. Specifically, we demonstrate that generative models produce images with lines that differ from those of real images (likely due to problems aligning vanishing points); that generative models produce images with perspective fields that are unlike those of real images; and that object-shadow relations in generated images differ reliably from those in real images. We use advanced pretrained models (Line Segment Detection [31]; Perspective Fields [21]; and PointNet [34]) that inspect geometric representations to distinguish between real and generated images. We rely on derived geometry cues from SOTA methods as manual image analysis and explicit geometry-based rules like drawing lines perspective lines are not scalable or automatable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.473, + 0.497 + ], + "angle": 0, + "content": "To guarantee the accuracy of our findings, we follow a strict process of data curation. This involves using a controlled pixel-level classifier to filter out any biases related to color, texture, and local features within our test set. This precision in data selection is crucial to isolate and accurately assess the inconsistencies in projective geometry and illumination present in generated images. By carefully screening the data, we can ensure that our results are not obscured by common artifacts that are usually found in generated images. This enhances the reliability of our conclusions. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.5, + 0.472, + 0.559 + ], + "angle": 0, + "content": "- Unearthing Geometric Discrepancies: We present a comprehensive analysis that goes beyond existing literature to both demonstrate and quantify geometric discrepancies produced by current generative models." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.56, + 0.472, + 0.605 + ], + "angle": 0, + "content": "- Scalable, Data-Driven Approach: We introduce a scalable, data-driven approach by utilizing three distinct projective geometry cues to detect geometric inaccuracies." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.47, + 0.664 + ], + "angle": 0, + "content": "- Robust and Generalizable: We demonstrate robustness and generalizability by effectively identifying geometric errors across a wide array of images from recent generative models, including those with the latest time stamps." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.666, + 0.47, + 0.711 + ], + "angle": 0, + "content": "- Broadening the Scope of Model Assessment: Our approach can be used as an alternative method for evaluating models: do they get projective geometry right?" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.5, + 0.472, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.22, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.885 + ], + "angle": 0, + "content": "**Generative Models:** The advancement of generative models, particularly in creating visually realistic images, marks a significant milestone in computer vision. Pioneering efforts by Karras et al. [22-24] with StyleGAN, and the emergence of diffusion models [36-38], have set new benchmarks in realism. These models, used in diverse fields from art to data augmentation, have yet to fully grasp the nuances of projective geometry, which is the focus of our analysis, primarily using open diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Geometric Understanding in Generative Models: While" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.182 + ], + "angle": 0, + "content": "studies like Bhattachad et al. [6], Du et al.[13], Chen et al. [10] and Zhan et al. [49] demonstrate these models' potential in understanding scene geometry, our work diverges by scrutinizing the generated images themselves, examining their adherence to the principles of projective geometry and illumination, rather than analyzing learned features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.187, + 0.897, + 0.338 + ], + "angle": 0, + "content": "Detecting Generated Images: The realism of modern generative models has made image forensics increasingly challenging. Traditional methods focused on detecting synthetic images using signals like resampling artifacts [33] and JPEG quantization [2]. Kee et al. [25] introduced a geometric technique for detecting shadow inconsistencies, paralleling our pursuit of physical realism. However, our work extends beyond identifying photo manipulation to evaluating the overall perspective geometry and illumination consistency in images from generative models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.897, + 0.51 + ], + "angle": 0, + "content": "Zhang et al.'s work [14] focuses on detecting AI-generated images using diverse generative models and online training for future model adaptation. Our research, in contrast, assesses the projective geometry in these images, examining their ability to render scenes with accurate perspective and illumination. Boháček et al. [7], while detecting geometric inconsistencies related to shadows, align with our interest in physical realism. However, we delve deeper, thoroughly evaluating perspective geometry and illumination in generative models for a more comprehensive understanding of their geometric accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.68 + ], + "angle": 0, + "content": "The rise of generative methods has steered image forensics towards using discriminative methods to detect synthetic content [8, 15, 19, 42, 43, 48, 50]. These advancements align with our objective of analyzing the physical and geometrical congruence of generated images. However, our work goes a step further by critically assessing whether generative models fundamentally understand and accurately replicate projective geometry, rather than simply distinguishing between real and synthetic images. This deeper level of analysis aims to unveil the intricacies and limitations of current models in faithfully rendering geometrically coherent images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.895, + 0.806 + ], + "angle": 0, + "content": "Evaluation Metrics: Traditional metrics like the Inception Score (IS) [39] and Fréchet Inception Distance (FID) [18] focus on pixel-level fidelity. The emergence of CLIP-based scores [17, 35] and DIRE [45] offers a semantic perspective. In contrast, our approach, distinct in its focus on perspective geometry and illumination consistency, seeks to ensure comprehensive realism, bridging the gap between visual and physical authenticity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Recent studies like Davidsonian Scene Graph [11] and ImagenHub [26] address fine-grained evaluation inconsistencies, while the HEIM benchmark [27] assesses models across multiple aspects. Our work complements these by providing an in-depth evaluation of the physical and geometric realism of images generated by state-of-the-art models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28141" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.409, + 0.108 + ], + "angle": 0, + "content": "3. Background on Projective Geometry" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.471, + 0.266 + ], + "angle": 0, + "content": "Projective geometry is a mathematical framework that enables the accurate representation of three-dimensional spaces in two-dimensional images. It provides the rules for perspective, which are crucial for creating realistic scenes with depth and spatial orientation [16]. In this section, we will examine the common inconsistencies that may arise during image synthesis according to projective geometry. Our evaluation framework is intended to detect and measure these discrepancies, which are essential for evaluating the realism and physical plausibility of generated images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.472, + 0.359 + ], + "angle": 0, + "content": "Inconsistent Vanishing Points. Vanishing points are fundamental to capturing the essence of perspective in images. They should align with the direction of parallel lines converging at a distance. Generated images often exhibit inconsistencies where these lines do not meet at the correct vanishing points, leading to a distorted sense of perspective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.471, + 0.435 + ], + "angle": 0, + "content": "Lighting and Shadow Inconsistencies. Accurate shadows are essential for reinforcing the position and shape of objects within a scene. Discrepancies in shadow direction, length, and softness can indicate a misalignment with the scene's light sources, disrupting the image's three-dimensionality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.472, + 0.51 + ], + "angle": 0, + "content": "Scale Discrepancies. The principle of size constancy dictates that objects of the same size should appear smaller as their distance from the observer increases. Generated images sometimes fail to maintain this scaling, resulting in a compromised depth perception." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.472, + 0.587 + ], + "angle": 0, + "content": "Distortion of Geometric Figures. Geometric figures should maintain their shape when projected onto the image plane, barring intentional perspective distortion. Errors in this projection can result in circles appearing as ellipses or squares as trapezoids, indicating a flawed perspective rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.472, + 0.661 + ], + "angle": 0, + "content": "Depth Cues. Depth perception in images is conveyed through cues such as overlapping, texture gradients, and relative size. Misrepresentation of these cues can lead to an unnatural spatial arrangement that the human eye can readily detect as artificial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.472, + 0.814 + ], + "angle": 0, + "content": "Our evaluation framework, detailed in the subsequent sections, is designed to rigorously test generated images against these projective geometry principles. While a comprehensive evaluation of projective geometry would consider all the aforementioned inconsistencies, our framework prioritizes the detection of inconsistent vanishing points and lighting and shadow inconsistencies. These elements are particularly telling indicators of an image's projective geometry realism and are often the most challenging for generative models to replicate accurately." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.248, + 0.845 + ], + "angle": 0, + "content": "4. Dataset Curation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our data curation process is carefully designed to distinguish between real images and generated images from several generative models. This process includes models that the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.183 + ], + "angle": 0, + "content": "classifier has not seen during its training. Another important goal is to ensure our prequalifier effectively identifies images with recent timestamps. This helps prevent the classifier from relying on cues specific to the dataset, such as whether the image is from the training distribution or not, rather than determining whether the image is real or generated." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.192, + 0.633, + 0.208 + ], + "angle": 0, + "content": "4.1. Real Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.862, + 0.23 + ], + "angle": 0, + "content": "We experiment with a diverse set of images, including:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.894, + 0.305 + ], + "angle": 0, + "content": "(a) Indoor Scenes: A collection of 400,000 interior images featuring a variety of furniture arrangements and lighting conditions, sourced from LSUN [46]. Specifically, we used 100,000 images each from Bedroom, Dining Room, Kitchen, and Living Room categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.894, + 0.38 + ], + "angle": 0, + "content": "(b) Outdoor Scenes: A dataset of 125,000 outdoor scenes with varying landscapes and urban settings, sourced from Berkeley Deep Drive 100K [47] and Mapillary Vistas [30]. We selected images that represent a wide range of weather conditions and times of day." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.894, + 0.425 + ], + "angle": 0, + "content": "(c) Combination of Indoor and Outdoor scenes: We also analyzed a combination of above indoor and outdoor scenes to assess the performance on a more diverse dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.894, + 0.531 + ], + "angle": 0, + "content": "(d) Recent Timestamp Images: A curated test set of 500 indoor and 500 outdoor images with timestamps ranging from May 2023 to March 2024. These images were collected from various social media platforms and online sources to ensure our classifier's ability to handle recent real-world data and that the models are not obscured because of any data-source-specific biases." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.23, + 0.894, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.541, + 0.66, + 0.558 + ], + "angle": 0, + "content": "4.2. Image Captions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.894, + 0.686 + ], + "angle": 0, + "content": "We use the ViT-bigG-14/laion2b_s39b_b160k model [20] and the BLIP model [28] in succession, to generate refined captions for real images. These models were chosen for their state-of-the-art performance in image captioning tasks. The ViT model uses a Vision Transformer architecture, while BLIP employs a multimodal pre-training approach. By using common captions, we ensure a fair evaluation of the projective geometry in generated images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.695, + 0.679, + 0.71 + ], + "angle": 0, + "content": "4.3. Generated Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.793 + ], + "angle": 0, + "content": "We generate images from Stable Diffusion XL v1.0 [32], Kandinsky-v3 [3], DeepFloyd IF v1.0 [12], and PixArt- \\(\\alpha\\) v1.0 [9]. We use the same caption from the real images to generate these images with the default settings of each generative model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.802, + 0.688, + 0.818 + ], + "angle": 0, + "content": "4.4. Robust Prequalifier" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Our goal is to identify challenging images that a signal-based classifier may struggle to differentiate. Therefore, we need to develop a robust prequalifier. This is to ensure that our results are not affected by any false data signals. Thus, we aim to eliminate all potential factors that may influence" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28142" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Table 1. Statistical overview of the Data Curation and Filtering Process: We present the distribution of real and generated images for indoor and outdoor datasets. The ResNet50 Prequalifier helps us curate datasets, creating an 'Unconfident Set' for images with low classifier certainty and a 'Misclassified Subset' for images incorrectly labeled by the classifier. These rigorously curated sets are instrumental for our subsequent analysis, concentrating on projective geometry while mitigating the influence of signal cues. The datasets for each prequalifier—indoor, outdoor, and combined—are prepared separately to tailor the models to their specific contexts." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.163, + 0.868, + 0.307 + ], + "angle": 0, + "content": "
Indoor RealIndoor GeneratedOutdoor RealOutdoor GeneratedCombined RealCombined Generated
Total Images400,000400,000125,000125,000525,000525,000
Training and Test Sets
Training Set Size75,00075,00025,00025,000100,000100,000
Validation Set Size10,00010,0005,0005,00015,00015,000
Test Set Size315,000315,00095,00095,000410,000410,000
Post ResNet50 Prequalifier
Unconfident Set232131384014441018443922540
Misclassified Subset10399575660944323800825
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.471, + 0.4 + ], + "angle": 0, + "content": "the geometry cues we obtain because of signal weirdness in the generated images. To accomplish this, we begin by training signal-based classifiers using a vanilla ResNet-50 and primarily concentrate on identifying instances that prove challenging for these signal-based classifiers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.472, + 0.593 + ], + "angle": 0, + "content": "We trained prequalifiers on three distinct settings - indoor scenes, outdoor scenes and a combination of the two. Each prequalifier was trained on a dataset consisting of real images and images generated by one of the four models: Stable Diffusion XL, Kandinsky-v3, DeepFloyd IF, and PixArt-\\(\\alpha\\). While all four types of prequalifiers performed well on their respective test sets, we found that Kandinsky exhibited superior generalization capabilities when evaluated on images generated by other models or on recent timestamp images. The fundamental experiment conducted to assess the generalizability of different generators is covered in Figure 9 in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.472, + 0.771 + ], + "angle": 0, + "content": "For indoor scenes, our Kandinsky-based prequalifier achieves an Area Under the Curve (AUC) of 0.99 on its test set with an accuracy of 97.43 and maintained high performance on images generated by other models such as Stable Diffusion XL and PixArt- \\(\\alpha\\), achieving AUCs of 0.97 and 0.98 respectively. Furthermore, our Kandinsky-trained prequalifier meant for both indoor and outdoor settings combined showed robust performance on recent timestamp generated indoor and outdoor images, achieving AUC scores of 0.90, 0.95, and 0.72 on images generated by Stable Diffusion XL, PixArt- \\(\\alpha\\), and DeepFloyd IF, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Given their robustness and strong generalizability, we selected the Kandinsky-trained prequalifiers as the basis for training our derived geometry classifiers on Kandinsky-generated images. This choice ensures that our classifiers effectively look at geometric discrepancies and can handle various generated images. This enhances the reliability of our approach and supports our conclusion that the classifiers are not affected by spurious signal artifacts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.324, + 0.65, + 0.339 + ], + "angle": 0, + "content": "4.5. Final Test Sets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.895, + 0.514 + ], + "angle": 0, + "content": "We categorize our test sets into three groups based on the accuracy of the prequalifier: easy, unconfident, and misclassified. The easy test set includes images with \\(100\\%\\) accuracy, indicating that the prequalifier can reliably distinguish between real and generated images. The unconfident test set includes images where the prequalifier performs at chance level, suggesting that it struggles to make confident predictions. Finally, the misclassified test set includes images where the prequalifier makes completely wrong predictions, either classifying real images as generated or vice versa. A summary of this split for Kandinsky-v3 is provided in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Our approach primarily evaluates the \"hard set\" (unconfident and misclassified) for geometric and shadow inconsistencies, assessing adherence to projective geometry principles. This ensures rigorous testing of generative models' ability to reproduce geometric correctness and photometric accuracy, beyond surface-level or signal details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.895, + 0.725 + ], + "angle": 0, + "content": "It is important to note that projective geometry inconsistencies are prevalent in generated images but often go undetected by conventional methods. Our models, trained on geometric abstractions and projective cues, can identify subtle but critical inaccuracies that texture artifacts cannot explain. This distinction is critical, as it allows us to rigorously test our models, which, unlike the prequalifier, do not have direct access to the images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.725, + 0.895, + 0.832 + ], + "angle": 0, + "content": "We employ a suite of models to capture different facets of projective geometry, trained on datasets emphasizing geometric consistency and photometric accuracy. By combining their strengths, we comprehensively assess the quality of generated images and provide insights into the limitations and potential improvements of generative models from a projective geometry perspective." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.845, + 0.787, + 0.864 + ], + "angle": 0, + "content": "5. Analyzing Projective Geometry" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For our analysis, we rely on three geometry cues – object shadow, line segments, and perspective fields. We train" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28143" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.096, + 0.331, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.096, + 0.593, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.097, + 0.888, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.321, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 2. We train three classifiers to identify discrepancies in projective geometry. These classifiers are trained using derived geometry cues such as object-shadow associations (left), perspective fields (middle), and line segments (right) without looking at image intensity. We use the ResNet architecture for Object-Shadow and Perspective Fields, and PointNet for Line Segments to process unordered data sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.471, + 0.439 + ], + "angle": 0, + "content": "three separate classifiers solely based on these derived geometry cues, without utilizing any pixel information. The training process for each classifier is described below, and a schematic pipeline of these classifiers can be found in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.447, + 0.276, + 0.463 + ], + "angle": 0, + "content": "5.1. Object-Shadow Cues" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.516 + ], + "angle": 0, + "content": "Our first model examines object-shadow relationships to address illumination. Shadows follow projective geometry principles, and inconsistencies can reveal generated images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.471, + 0.622 + ], + "angle": 0, + "content": "We use a pretrained object-shadow instance detection model [44] to identify shadows and geometric heuristics to evaluate their plausibility given the objects and their orientation. A ResNet50 classifier is trained on binary masks of object and shadow instances to score the consistency of shadows with objects. Images with implausible shadows are marked as likely generated." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.288, + 0.648 + ], + "angle": 0, + "content": "5.2. Perspective Field Cues" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.731 + ], + "angle": 0, + "content": "Our framework's second model uses Perspective Fields [21], vector fields encoding the spatial orientation of pixels relative to vanishing points and the horizon, to assess projective geometry. We generate these fields from single images using a pretrained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.837 + ], + "angle": 0, + "content": "We train a ResNet50 classifier on the Perspective Fields to differentiate real from generated images by focusing on projective geometry anomalies. The classifier evaluates the consistency of these fields with projective geometry principles, scoring images on their geometric plausibility. This method enables precise evaluation of projective geometry, enhancing the detection of subtle inconsistencies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.262, + 0.864 + ], + "angle": 0, + "content": "5.3. Line Segment Cues" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our method also assesses projective geometry in generated images by identifying key structural lines using Deep" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.895, + 0.437 + ], + "angle": 0, + "content": "LSD [31]. These lines indicate adherence to perspective rules. We then train a PointNet-like architecture [34] to classify images based on line segment patterns, differentiating real from generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.895, + 0.543 + ], + "angle": 0, + "content": "PointNet's flexibility in handling unordered data makes it suitable for analyzing line segments without pre-sorting. The model assigns scores representing the likelihood of an image being real based on the spatial arrangement of its lines. Analyzing these scores reveals the model's ability to detect subtle discrepancies in line arrangements, which often indicate a generated image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.616, + 0.572 + ], + "angle": 0, + "content": "6. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.894, + 0.688 + ], + "angle": 0, + "content": "Based on our prequalification analysis, we use Kandinsky as the primary source for our generated data to train our classifiers. We evaluated our three geometry-derived cues classifiers and analyzed their ability to generalize to other generated images that were not used during training. In addition, we performed a GradCam analysis [40] to identify any potential geometric discrepancies in images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.676, + 0.711 + ], + "angle": 0, + "content": "6.1. Classifiers Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.856 + ], + "angle": 0, + "content": "Fig. 3 shows ROC curves for each of our methods on indoor, outdoor and combined (indoor+outdoor) scenes. In each case, classifiers are trained on images that are not prequalified and tested on prequalified scenes, meaning that performance estimates are biased low — likely training on prequalified data would lead to even more accurate classification. Each classifier is effective, with AUCs ranging from 0.72 to 0.97. Recall these classifiers see only derived geometric features and do not see the image itself." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Qualitative examples using Grad-CAM appear in Fig. 5 and Fig. 4. Notice how images that might be acceptable to a line analysis often fail a shadow analysis. Fig. 6 shows" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.274, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.213, + 0.246, + 0.225 + ], + "angle": 0, + "content": "(a) Easy Test Set (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.091, + 0.482, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.213, + 0.479, + 0.225 + ], + "angle": 0, + "content": "(b) Recent Timestamp Set (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.091, + 0.681, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.213, + 0.672, + 0.224 + ], + "angle": 0, + "content": "(c) Unconfident Test Set (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.091, + 0.888, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.213, + 0.882, + 0.224 + ], + "angle": 0, + "content": "(d) Misclassified Test Set (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.228, + 0.274, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.35, + 0.25, + 0.362 + ], + "angle": 0, + "content": "(e) Easy Test Set (Outdoor)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.228, + 0.482, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.35, + 0.482, + 0.362 + ], + "angle": 0, + "content": "(f) Recent Timestamp Set (Outdoor)" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.228, + 0.681, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.35, + 0.676, + 0.362 + ], + "angle": 0, + "content": "(g) Unconfident Test Set (Outdoor)" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.228, + 0.888, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.35, + 0.89, + 0.362 + ], + "angle": 0, + "content": "(h) Misclassified Test Set ((Outdoor))" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.365, + 0.274, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.486, + 0.255, + 0.498 + ], + "angle": 0, + "content": "(i) Easy Test Set (Combined)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.365, + 0.48, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.486, + 0.483, + 0.498 + ], + "angle": 0, + "content": "(j) Recent Timestamp Set (Combined)" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.365, + 0.68, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.486, + 0.68, + 0.498 + ], + "angle": 0, + "content": "(k) Unconfident Test Set (Combined)" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.365, + 0.885, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.486, + 0.887, + 0.498 + ], + "angle": 0, + "content": "(1) Misclassified Test Set (Combined)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.5, + 0.893, + 0.653 + ], + "angle": 0, + "content": "Figure 3. ROC Curves Assessing Projective Geometry Cues in Generated Images trained on Kandinsky-v3. We trained separate models for indoor scenes (top row), outdoor scenes (middle row), and a combination of indoor and outdoor scenes (last row). All our derived geometry cues classifiers are trained without looking at image intensity information and can reliably detect projective geometry errors. The recent timestamp test set (second column) confirms that these models are robust. We find hard examples using a prequalifier trained on image pixels. Our derived geometry cues consistently show high AUC for finding projective geometry errors on hard test sets – the last two columns – unconfident and misclassified test sets. For the unconfident test set, where the prequalifier has an AUC of 0.51 (c), 0.51 (g), and 0.49 (h) for indoor, outdoor, and combined partition, our classifiers can still accurately identify the generated images with high AUCs – 0.82 from line segments in the indoor set, 0.84 from perspective fields in the outdoor set, and 0.80 from perspective field cues and object shadows in the combined set. Similarly, for the misclassified test set, where the prequalifier has an AUC of 0.00, as it should, our classifiers remain reliable with AUC up to 0.82. We conclude that generated images contain geometric structures not seen in real images, and these structures very reliably identify generated images by only looking at derived geometry cues." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.667, + 0.301, + 0.682 + ], + "angle": 0, + "content": "examples to emphasize this point." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.341, + 0.71 + ], + "angle": 0, + "content": "7. Other Generators Evaluated" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our classifiers do not see pixels, but derived geometric features. This means that one could expect a form of generalization across generators. We illustrate that this generalization occurs - ROC curves in Fig. 7 demonstrate that classifiers trained to distinguish Kandinsky images from real images can also reliably distinguish Stable Diffusion XL [3], DeepFloyd [12] and PixArt-\\(\\alpha\\) [9] from the open-source domain. Additionally, we assess the efficacy of our models against images from proprietary generators such as OpenAI's Dalle-3[4] and Adobe's Firefly [1], representing some of the most advanced tools in image generation. Finally, we show we can detect composite made by a recent SOTA method [29]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.667, + 0.851, + 0.682 + ], + "angle": 0, + "content": "by looking at Object-Shadow cues in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.612, + 0.71 + ], + "angle": 0, + "content": "8. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We have shown that generated images can be reliably distinguished from real images by looking only at derived geometric cues. This is likely because image generators do not fully implement the geometry one observes in real images. Producing accurate perspective geometry or accurate shadow geometry requires very tight coordination of detailed information over very long spatial scales. Our results, together with the notorious tendency of face image generators to award subjects' left and right earlobes of different shapes, suggest that doing so is beyond the capacity of current generators. Our findings have significant implications for the development of image generation models, as the inability to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28145" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.197, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.355, + 0.187, + 0.365 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.091, + 0.312, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.355, + 0.295, + 0.365 + ], + "angle": 0, + "content": "Shadow Errors" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.091, + 0.428, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.355, + 0.425, + 0.366 + ], + "angle": 0, + "content": "Object-Shadow (OS)" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.091, + 0.543, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.355, + 0.52, + 0.365 + ], + "angle": 0, + "content": "OS GradCam" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.091, + 0.655, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.355, + 0.626, + 0.365 + ], + "angle": 0, + "content": "VP Errors" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.091, + 0.772, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.355, + 0.767, + 0.366 + ], + "angle": 0, + "content": "Perspective Fields (PF)" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.091, + 0.887, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.355, + 0.863, + 0.365 + ], + "angle": 0, + "content": "PF GradCam" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.895, + 0.471 + ], + "angle": 0, + "content": "Figure 4. Grad-CAM applied to our Object-Shadow and Perspective Field classifiers reveals that the high AUCs in Fig. 3 are based on real geometric errors in indoor scenes generated by Kandinsky, Stable Diffusion XL and Dalle-3 shown in each row respectively. The second and fifth columns highlight shadow and vanishing point errors, respectively. The third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) identifies diagnostic areas for synthetic generation, such as inconsistent shadow directions (in all three rows), mismatched shadow lengths (second row). The sixth column shows Perspective Fields [21], and Grad-CAM applied to our Perspective Fields classifier (last column) reveals geometric errors in all three rows, particularly at ceilings and side walls, with noticeable errors also present in window grills in the first and second rows." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.485, + 0.197, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.753, + 0.186, + 0.765 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.485, + 0.312, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.753, + 0.295, + 0.764 + ], + "angle": 0, + "content": "Shadow Errors" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.485, + 0.428, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.753, + 0.425, + 0.764 + ], + "angle": 0, + "content": "Object-Shadow (OS)" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.485, + 0.543, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.753, + 0.52, + 0.764 + ], + "angle": 0, + "content": "OS GradCam" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.485, + 0.655, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.753, + 0.626, + 0.764 + ], + "angle": 0, + "content": "VP Errors" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.485, + 0.772, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.753, + 0.762, + 0.764 + ], + "angle": 0, + "content": "Perspective Fields" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.485, + 0.887, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.753, + 0.863, + 0.764 + ], + "angle": 0, + "content": "PF GradCam" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.77, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Figure 5. Grad-CAM results for outdoor scenes generated by Kandinsky, Stable Diffusion XL, and Adobe Firefly, shown in each row respectively. The second column highlights shadow errors, while the third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) reveals incorrect shadow shapes in the first and second rows, with shadows on the right-side pedestrians pointing in a different direction than those on the left. The fifth column shows vanishing point errors, and the sixth column presents Perspective Fields [21]. Grad-CAM applied to our Perspective Fields classifier (last column) confirms large perspective distortions on building facades and road markings, corroborating the vanishing point errors in the fifth column." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28146" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.197, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.27, + 0.187, + 0.282 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.092, + 0.428, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.295, + 0.282 + ], + "angle": 0, + "content": "Shadow Errors" + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.269, + 0.425, + 0.282 + ], + "angle": 0, + "content": "Object-Shadow (OS)" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.092, + 0.541, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.27, + 0.521, + 0.282 + ], + "angle": 0, + "content": "OS GradCam" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.092, + 0.655, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.27, + 0.626, + 0.282 + ], + "angle": 0, + "content": "VP Errors" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.092, + 0.772, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.27, + 0.767, + 0.282 + ], + "angle": 0, + "content": "Perspective Fields (PF)" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.092, + 0.885, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.27, + 0.865, + 0.281 + ], + "angle": 0, + "content": "PF GradCam" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Figure 6. Our projective geometry classifiers identify distinct types of problems in generated images. The top row presents an example that was classified as real by the Object-Shadow classifier but correctly identified as generated by the Perspective Fields classifier. While the shadow cast by the person appears realistic, the Perspective Fields Grad-CAM highlights the problematic geometry of the shelf on the top left. In contrast, the bottom row shows an example that was correctly identified as generated by the Object-Shadow classifier but misclassified as real by the Perspective Fields classifier. Although the perspective effects in the image appear plausible, the Grad-CAM weights correctly reveal that the two chairs are casting shadows from different light sources, indicating inconsistency in scene's illumination." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.384, + 0.235, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.483, + 0.214, + 0.494 + ], + "angle": 0, + "content": "(a) Dall-E 3 (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.497, + 0.234, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.594, + 0.217, + 0.606 + ], + "angle": 0, + "content": "(f) Dall-E 3 (Outdoor)" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.385, + 0.405, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.482, + 0.392, + 0.494 + ], + "angle": 0, + "content": "(b) DeepFloyd (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.496, + 0.397, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.594, + 0.388, + 0.606 + ], + "angle": 0, + "content": "(g) DeepFloyd (Outdoor)" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.385, + 0.576, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.483, + 0.569, + 0.494 + ], + "angle": 0, + "content": "(c) Adobe Firefly (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.496, + 0.562, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.594, + 0.558, + 0.606 + ], + "angle": 0, + "content": "(h) Adobe Firefly (Outdoor)" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.385, + 0.733, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.483, + 0.714, + 0.494 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{PixArt - }\\alpha\\) (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.496, + 0.724, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.594, + 0.707, + 0.606 + ], + "angle": 0, + "content": "(i)PixArt- \\(\\alpha\\) (Outdoor)" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.385, + 0.89, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.483, + 0.863, + 0.494 + ], + "angle": 0, + "content": "(e) SDXL (Indoor)" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.496, + 0.887, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.594, + 0.863, + 0.606 + ], + "angle": 0, + "content": "(j) SDXL (Outdoor)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.612, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Figure 7. Our model trained on Kandinsky-v3 generalizes in detecting projective geometry errors from images generated by various unseen models. We evaluate the model's performance on test sets generated by Dall-E 3 (a), DeepFloyd (b), Adobe Firefly (c), PixArt-α (d), and Stable Diffusion XL or SDXL (e) using the same text prompts from the 'misclassified' Kandinsky-v3 generated test set. The top row shows indoor scenes and the last row shows outdoor scenes. Our model generalizes to all evaluated generators except DeepFloyd. However, DeepFloyd-generated images can be reliably detected when the model is trained on a DeepFloyd-generated training set, but it shows poor generalization capabilities to other generators compared to Kandinsky." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.471, + 0.759 + ], + "angle": 0, + "content": "accurately replicate projective geometry extends across various state-of-the-art models, indicating a widespread issue rather than a problem specific to a particular generator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We speculate that fixing this difficulty requires structural innovation in the generator, rather than simply exposing the generator to more data. The complex interplay of light, shadows, and perspective in real-world scenes may necessitate novel approaches to modeling and encoding spatial relationships within the generator's architecture. Potential avenues for improvement could include incorporating explicit geometric reasoning or developing new loss functions that prioritize the consistency of projective geometry. Fur" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.803 + ], + "angle": 0, + "content": "thermore, our work highlights the importance of developing robust evaluation metrics for image generation models that assess the geometric coherence of generated images. By tackling these challenges, we could create image-generation models that more faithfully capture the complex geometric relationships in real-world scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.82, + 0.652, + 0.837 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.895, + 0.9 + ], + "angle": 0, + "content": "This paper is based on work supported in part by the National Science Foundation under Grant No. 2106825 and a gift from Boeing Corporation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28147" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.156 + ], + "angle": 0, + "content": "[1] Adobe. Firefly. https://www.adobe.com/sensei/generative-ai/firefly.html, 2023. Accessed: 2023-11. 2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] Shruti Agarwal and Hany Farid. Photo forensics fromJPEG dimples. In 2017 IEEE workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[3] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.356 + ], + "angle": 0, + "content": "[5] Anand Bhattachad and D.A. Forsyth. Stylitgan: Prompting stylegan to generate new illumination conditions. In arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.471, + 0.414 + ], + "angle": 0, + "content": "[6] Anand Bhattachad, Daniel McKee, Derek Hoiem, and DA Forsyth. Stylegan knows normal, depth, albedo, and more. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.472, + 0.471 + ], + "angle": 0, + "content": "[7] Matyáš Boháček and Hany Farid. A geometric and photometric exploration of gan and diffusion synthesized faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-883, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.472, + 0.585 + ], + "angle": 0, + "content": "[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[10] Yida Chen, Fernanda Viégas, and Martin Wattenberg. Beyond surface statistics: Scene representations in a latent diffusion model. arXiv preprint arXiv:2306.05720, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.471, + 0.699 + ], + "angle": 0, + "content": "[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-image generation. arXiv preprint arXiv:2310.18235, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[12] Deep Floyd. Iterative filter. https://github.com/deep-floyd/IF, 2023. Accessed: 2023-11. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[13] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[14] David C. Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 382-392, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[15] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "[16] David A Forsyth and Jean Ponce. Computer vision: a modern approach. prentice hall professional technical reference, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.895, + 0.232 + ], + "angle": 0, + "content": "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.895, + 0.303 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[19] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.895, + 0.418 + ], + "angle": 0, + "content": "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.895, + 0.502 + ], + "angle": 0, + "content": "[21] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 2, 5, 7, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.894, + 0.63 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.894, + 0.688 + ], + "angle": 0, + "content": "[24] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.893, + 0.731 + ], + "angle": 0, + "content": "[25] Eric Kee, James F O'Brien, and Hany Farid. Exposing photo manipulation with inconsistent shadows. ACM Transactions on Graphics (ToG), 32(3):1-12, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.893, + 0.789 + ], + "angle": 0, + "content": "[26] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[27] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, Minguk Kang, Taesung Park, Jure Leskovec, Jun-Yan Zhu, Li Fei-Fei, Jiajun Wu, Stefano Ermon, and Percy Liang. Holistic evaluation of text-to-image models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28148" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[29] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 5000–5009, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.348 + ], + "angle": 0, + "content": "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.472, + 0.462 + ], + "angle": 0, + "content": "[33] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 53(2):758-767, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.472, + 0.674 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.472, + 0.745 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[39] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[40] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 128(2):336-359, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[41] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18310-18319, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.276 + ], + "angle": 0, + "content": "[42] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.319 + ], + "angle": 0, + "content": "[43] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[44] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3259-3273, 2022. 5, 7, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.893, + 0.444 + ], + "angle": 0, + "content": "[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22445-22455, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.893, + 0.502 + ], + "angle": 0, + "content": "[46] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.894, + 0.585 + ], + "angle": 0, + "content": "[47] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.893, + 0.643 + ], + "angle": 0, + "content": "[48] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.893, + 0.685 + ], + "angle": 0, + "content": "[49] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.893, + 0.728 + ], + "angle": 0, + "content": "[50] Peng Zhou, Xintong Han, Vlad I Morariu, and Larry S Davis. Learning rich features for image manipulation detection. In CVPR, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28149" + } + ] +] \ No newline at end of file diff --git a/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_origin.pdf b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5636003472b2fff5c4edd12d616e903184632a32 --- /dev/null +++ b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/6c6001bd-500a-4e2e-8037-41da45d3b01b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18a503fa0508b3d2319b1693e0c571117cb95fa7716477713b1f79ac4eda9124 +size 5388655 diff --git a/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/full.md b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/full.md new file mode 100644 index 0000000000000000000000000000000000000000..162a2d446ce69ffd9d6de651594ec99d319526d9 --- /dev/null +++ b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/full.md @@ -0,0 +1,383 @@ +# Shadows Don't Lie and Lines Can't Bend! +Generative Models don't know Projective Geometry...for now + +Ayush Sarkar*1 Hanlin Mai*1 Amitabh Mahapatra*1 Svetlana Lazebnik + +D.A. Forsyth1 Anand Bhattachad2 + +1University of Illinois Urbana-Champaign 2Toyota Technological Institute at Chicago + +https://projective-geometry.github.io/ + +![](images/2a816408b7c8608b4adca114e38103aee9b4dd5338bca4b84e38dff871712a89.jpg) +Generated Image + +![](images/14825c8c4e005b253de46315b0323fc37245196dfed9c0645eea5309dd1c48e1.jpg) +Shadow Errors + +![](images/ecda4d8f0f8638623144ae24289797545afd64b5e07f3aa042ae87ee1725af02.jpg) +Detected Shadow Errors + +![](images/4df22eacc828692978e1171d6b0d76404f455b8219035fc769521e211018f2e2.jpg) +Vanishing Point Errors + +![](images/7e15621b0bb55c86cdfa4356ce456c94c285d2f99d2eada6041052bf96d05e50.jpg) +Detected Perspective Errors +Figure 1. The first column presents visually compelling AI-generated images. However, a closer examination reveals fundamental inconsistencies, such as those in shadow alignment (second column) and vanishing point accuracy (fourth column). Our model's analysis, shown in the third and fifth columns, detects these shadow and perspective geometry errors. We show that these errors are systematic and can be used to identify generated images. + +# Abstract + +Generative models can produce impressively realistic images. This paper demonstrates that generated images have geometric features different from those of real images. We build a set of collections of generated images, prequalified to fool simple, signal-based classifiers into believing they are real. We then show that prequalified generated images can be identified reliably by classifiers that only look at geometric properties. We use three such classifiers. All three classifiers are denied access to image pixels, and look only at derived geometric features. The first classifier looks at the perspective field of the image, the second looks at lines detected in the image, and the third looks at relations between detected objects and shadows. Our procedure detects generated images more reliably than SOTA local signal based detectors, + +for images from a number of distinct generators. Saliency maps suggest that the classifiers can identify geometric problems reliably. We conclude that current generators cannot reliably reproduce geometric properties of real images. + +# 1. Introduction + +Both StyleGAN [22-24] and diffusion models [36-38] are renowned for generating images that are strikingly similar to real-world photos and consistently fool people. But, as we show, generated images have distinctive geometric features, likely from a failure to fully capture projective geometry. + +Bhattad et al. [5, 6], Chen et al. [10], Du et al. [13], Zhan et al. [49], and have shown generative models implicitly capture the complex scene properties, including normals, depth, albedo, and support relations. These works suggest these models "understand" geometry, which would + +be useful for rendering 3D scenes. Our detailed, population-level analysis of generated images suggests generative models [1, 3, 4, 12, 32] cannot fully translate this "understanding" into accurate geometry. Specifically, we demonstrate that generative models produce images with lines that differ from those of real images (likely due to problems aligning vanishing points); that generative models produce images with perspective fields that are unlike those of real images; and that object-shadow relations in generated images differ reliably from those in real images. We use advanced pretrained models (Line Segment Detection [31]; Perspective Fields [21]; and PointNet [34]) that inspect geometric representations to distinguish between real and generated images. We rely on derived geometry cues from SOTA methods as manual image analysis and explicit geometry-based rules like drawing lines perspective lines are not scalable or automatable. + +To guarantee the accuracy of our findings, we follow a strict process of data curation. This involves using a controlled pixel-level classifier to filter out any biases related to color, texture, and local features within our test set. This precision in data selection is crucial to isolate and accurately assess the inconsistencies in projective geometry and illumination present in generated images. By carefully screening the data, we can ensure that our results are not obscured by common artifacts that are usually found in generated images. This enhances the reliability of our conclusions. Our contributions are: + +- Unearthing Geometric Discrepancies: We present a comprehensive analysis that goes beyond existing literature to both demonstrate and quantify geometric discrepancies produced by current generative models. +- Scalable, Data-Driven Approach: We introduce a scalable, data-driven approach by utilizing three distinct projective geometry cues to detect geometric inaccuracies. +- Robust and Generalizable: We demonstrate robustness and generalizability by effectively identifying geometric errors across a wide array of images from recent generative models, including those with the latest time stamps. +- Broadening the Scope of Model Assessment: Our approach can be used as an alternative method for evaluating models: do they get projective geometry right? + +# 2. Related Work + +**Generative Models:** The advancement of generative models, particularly in creating visually realistic images, marks a significant milestone in computer vision. Pioneering efforts by Karras et al. [22-24] with StyleGAN, and the emergence of diffusion models [36-38], have set new benchmarks in realism. These models, used in diverse fields from art to data augmentation, have yet to fully grasp the nuances of projective geometry, which is the focus of our analysis, primarily using open diffusion models. + +Geometric Understanding in Generative Models: While + +studies like Bhattachad et al. [6], Du et al.[13], Chen et al. [10] and Zhan et al. [49] demonstrate these models' potential in understanding scene geometry, our work diverges by scrutinizing the generated images themselves, examining their adherence to the principles of projective geometry and illumination, rather than analyzing learned features. + +Detecting Generated Images: The realism of modern generative models has made image forensics increasingly challenging. Traditional methods focused on detecting synthetic images using signals like resampling artifacts [33] and JPEG quantization [2]. Kee et al. [25] introduced a geometric technique for detecting shadow inconsistencies, paralleling our pursuit of physical realism. However, our work extends beyond identifying photo manipulation to evaluating the overall perspective geometry and illumination consistency in images from generative models. + +Zhang et al.'s work [14] focuses on detecting AI-generated images using diverse generative models and online training for future model adaptation. Our research, in contrast, assesses the projective geometry in these images, examining their ability to render scenes with accurate perspective and illumination. Boháček et al. [7], while detecting geometric inconsistencies related to shadows, align with our interest in physical realism. However, we delve deeper, thoroughly evaluating perspective geometry and illumination in generative models for a more comprehensive understanding of their geometric accuracy. + +The rise of generative methods has steered image forensics towards using discriminative methods to detect synthetic content [8, 15, 19, 42, 43, 48, 50]. These advancements align with our objective of analyzing the physical and geometrical congruence of generated images. However, our work goes a step further by critically assessing whether generative models fundamentally understand and accurately replicate projective geometry, rather than simply distinguishing between real and synthetic images. This deeper level of analysis aims to unveil the intricacies and limitations of current models in faithfully rendering geometrically coherent images. + +Evaluation Metrics: Traditional metrics like the Inception Score (IS) [39] and Fréchet Inception Distance (FID) [18] focus on pixel-level fidelity. The emergence of CLIP-based scores [17, 35] and DIRE [45] offers a semantic perspective. In contrast, our approach, distinct in its focus on perspective geometry and illumination consistency, seeks to ensure comprehensive realism, bridging the gap between visual and physical authenticity. + +Recent studies like Davidsonian Scene Graph [11] and ImagenHub [26] address fine-grained evaluation inconsistencies, while the HEIM benchmark [27] assesses models across multiple aspects. Our work complements these by providing an in-depth evaluation of the physical and geometric realism of images generated by state-of-the-art models. + +# 3. Background on Projective Geometry + +Projective geometry is a mathematical framework that enables the accurate representation of three-dimensional spaces in two-dimensional images. It provides the rules for perspective, which are crucial for creating realistic scenes with depth and spatial orientation [16]. In this section, we will examine the common inconsistencies that may arise during image synthesis according to projective geometry. Our evaluation framework is intended to detect and measure these discrepancies, which are essential for evaluating the realism and physical plausibility of generated images. + +Inconsistent Vanishing Points. Vanishing points are fundamental to capturing the essence of perspective in images. They should align with the direction of parallel lines converging at a distance. Generated images often exhibit inconsistencies where these lines do not meet at the correct vanishing points, leading to a distorted sense of perspective. + +Lighting and Shadow Inconsistencies. Accurate shadows are essential for reinforcing the position and shape of objects within a scene. Discrepancies in shadow direction, length, and softness can indicate a misalignment with the scene's light sources, disrupting the image's three-dimensionality. + +Scale Discrepancies. The principle of size constancy dictates that objects of the same size should appear smaller as their distance from the observer increases. Generated images sometimes fail to maintain this scaling, resulting in a compromised depth perception. + +Distortion of Geometric Figures. Geometric figures should maintain their shape when projected onto the image plane, barring intentional perspective distortion. Errors in this projection can result in circles appearing as ellipses or squares as trapezoids, indicating a flawed perspective rendering. + +Depth Cues. Depth perception in images is conveyed through cues such as overlapping, texture gradients, and relative size. Misrepresentation of these cues can lead to an unnatural spatial arrangement that the human eye can readily detect as artificial. + +Our evaluation framework, detailed in the subsequent sections, is designed to rigorously test generated images against these projective geometry principles. While a comprehensive evaluation of projective geometry would consider all the aforementioned inconsistencies, our framework prioritizes the detection of inconsistent vanishing points and lighting and shadow inconsistencies. These elements are particularly telling indicators of an image's projective geometry realism and are often the most challenging for generative models to replicate accurately. + +# 4. Dataset Curation + +Our data curation process is carefully designed to distinguish between real images and generated images from several generative models. This process includes models that the + +classifier has not seen during its training. Another important goal is to ensure our prequalifier effectively identifies images with recent timestamps. This helps prevent the classifier from relying on cues specific to the dataset, such as whether the image is from the training distribution or not, rather than determining whether the image is real or generated. + +# 4.1. Real Images + +We experiment with a diverse set of images, including: + +(a) Indoor Scenes: A collection of 400,000 interior images featuring a variety of furniture arrangements and lighting conditions, sourced from LSUN [46]. Specifically, we used 100,000 images each from Bedroom, Dining Room, Kitchen, and Living Room categories. +(b) Outdoor Scenes: A dataset of 125,000 outdoor scenes with varying landscapes and urban settings, sourced from Berkeley Deep Drive 100K [47] and Mapillary Vistas [30]. We selected images that represent a wide range of weather conditions and times of day. +(c) Combination of Indoor and Outdoor scenes: We also analyzed a combination of above indoor and outdoor scenes to assess the performance on a more diverse dataset. +(d) Recent Timestamp Images: A curated test set of 500 indoor and 500 outdoor images with timestamps ranging from May 2023 to March 2024. These images were collected from various social media platforms and online sources to ensure our classifier's ability to handle recent real-world data and that the models are not obscured because of any data-source-specific biases. + +# 4.2. Image Captions + +We use the ViT-bigG-14/laion2b_s39b_b160k model [20] and the BLIP model [28] in succession, to generate refined captions for real images. These models were chosen for their state-of-the-art performance in image captioning tasks. The ViT model uses a Vision Transformer architecture, while BLIP employs a multimodal pre-training approach. By using common captions, we ensure a fair evaluation of the projective geometry in generated images. + +# 4.3. Generated Images + +We generate images from Stable Diffusion XL v1.0 [32], Kandinsky-v3 [3], DeepFloyd IF v1.0 [12], and PixArt- $\alpha$ v1.0 [9]. We use the same caption from the real images to generate these images with the default settings of each generative model. + +# 4.4. Robust Prequalifier + +Our goal is to identify challenging images that a signal-based classifier may struggle to differentiate. Therefore, we need to develop a robust prequalifier. This is to ensure that our results are not affected by any false data signals. Thus, we aim to eliminate all potential factors that may influence + +Table 1. Statistical overview of the Data Curation and Filtering Process: We present the distribution of real and generated images for indoor and outdoor datasets. The ResNet50 Prequalifier helps us curate datasets, creating an 'Unconfident Set' for images with low classifier certainty and a 'Misclassified Subset' for images incorrectly labeled by the classifier. These rigorously curated sets are instrumental for our subsequent analysis, concentrating on projective geometry while mitigating the influence of signal cues. The datasets for each prequalifier—indoor, outdoor, and combined—are prepared separately to tailor the models to their specific contexts. + +
Indoor RealIndoor GeneratedOutdoor RealOutdoor GeneratedCombined RealCombined Generated
Total Images400,000400,000125,000125,000525,000525,000
Training and Test Sets
Training Set Size75,00075,00025,00025,000100,000100,000
Validation Set Size10,00010,0005,0005,00015,00015,000
Test Set Size315,000315,00095,00095,000410,000410,000
Post ResNet50 Prequalifier
Unconfident Set232131384014441018443922540
Misclassified Subset10399575660944323800825
+ +the geometry cues we obtain because of signal weirdness in the generated images. To accomplish this, we begin by training signal-based classifiers using a vanilla ResNet-50 and primarily concentrate on identifying instances that prove challenging for these signal-based classifiers. + +We trained prequalifiers on three distinct settings - indoor scenes, outdoor scenes and a combination of the two. Each prequalifier was trained on a dataset consisting of real images and images generated by one of the four models: Stable Diffusion XL, Kandinsky-v3, DeepFloyd IF, and PixArt- $\alpha$ . While all four types of prequalifiers performed well on their respective test sets, we found that Kandinsky exhibited superior generalization capabilities when evaluated on images generated by other models or on recent timestamp images. The fundamental experiment conducted to assess the generalizability of different generators is covered in Figure 9 in the Supplementary Material. + +For indoor scenes, our Kandinsky-based prequalifier achieves an Area Under the Curve (AUC) of 0.99 on its test set with an accuracy of 97.43 and maintained high performance on images generated by other models such as Stable Diffusion XL and PixArt- $\alpha$ , achieving AUCs of 0.97 and 0.98 respectively. Furthermore, our Kandinsky-trained prequalifier meant for both indoor and outdoor settings combined showed robust performance on recent timestamp generated indoor and outdoor images, achieving AUC scores of 0.90, 0.95, and 0.72 on images generated by Stable Diffusion XL, PixArt- $\alpha$ , and DeepFloyd IF, respectively. + +Given their robustness and strong generalizability, we selected the Kandinsky-trained prequalifiers as the basis for training our derived geometry classifiers on Kandinsky-generated images. This choice ensures that our classifiers effectively look at geometric discrepancies and can handle various generated images. This enhances the reliability of our approach and supports our conclusion that the classifiers are not affected by spurious signal artifacts. + +# 4.5. Final Test Sets + +We categorize our test sets into three groups based on the accuracy of the prequalifier: easy, unconfident, and misclassified. The easy test set includes images with $100\%$ accuracy, indicating that the prequalifier can reliably distinguish between real and generated images. The unconfident test set includes images where the prequalifier performs at chance level, suggesting that it struggles to make confident predictions. Finally, the misclassified test set includes images where the prequalifier makes completely wrong predictions, either classifying real images as generated or vice versa. A summary of this split for Kandinsky-v3 is provided in Tab. 1. + +Our approach primarily evaluates the "hard set" (unconfident and misclassified) for geometric and shadow inconsistencies, assessing adherence to projective geometry principles. This ensures rigorous testing of generative models' ability to reproduce geometric correctness and photometric accuracy, beyond surface-level or signal details. + +It is important to note that projective geometry inconsistencies are prevalent in generated images but often go undetected by conventional methods. Our models, trained on geometric abstractions and projective cues, can identify subtle but critical inaccuracies that texture artifacts cannot explain. This distinction is critical, as it allows us to rigorously test our models, which, unlike the prequalifier, do not have direct access to the images. + +We employ a suite of models to capture different facets of projective geometry, trained on datasets emphasizing geometric consistency and photometric accuracy. By combining their strengths, we comprehensively assess the quality of generated images and provide insights into the limitations and potential improvements of generative models from a projective geometry perspective. + +# 5. Analyzing Projective Geometry + +For our analysis, we rely on three geometry cues – object shadow, line segments, and perspective fields. We train + +![](images/aaa6eefb1677ccc2dd06713ec5425e9e5f852d5983b54c9cb8963bc7331de60a.jpg) +Figure 2. We train three classifiers to identify discrepancies in projective geometry. These classifiers are trained using derived geometry cues such as object-shadow associations (left), perspective fields (middle), and line segments (right) without looking at image intensity. We use the ResNet architecture for Object-Shadow and Perspective Fields, and PointNet for Line Segments to process unordered data sets. + +![](images/f19d44804ed4395cdf3bef925bfbc08e184cae01bae3034fe056c47aac52e121.jpg) + +![](images/29de772733505b4f9d7466825c0dfb740ac54c9d81497a5a29d605655c8f5150.jpg) + +three separate classifiers solely based on these derived geometry cues, without utilizing any pixel information. The training process for each classifier is described below, and a schematic pipeline of these classifiers can be found in Fig. 2. + +# 5.1. Object-Shadow Cues + +Our first model examines object-shadow relationships to address illumination. Shadows follow projective geometry principles, and inconsistencies can reveal generated images. + +We use a pretrained object-shadow instance detection model [44] to identify shadows and geometric heuristics to evaluate their plausibility given the objects and their orientation. A ResNet50 classifier is trained on binary masks of object and shadow instances to score the consistency of shadows with objects. Images with implausible shadows are marked as likely generated. + +# 5.2. Perspective Field Cues + +Our framework's second model uses Perspective Fields [21], vector fields encoding the spatial orientation of pixels relative to vanishing points and the horizon, to assess projective geometry. We generate these fields from single images using a pretrained model. + +We train a ResNet50 classifier on the Perspective Fields to differentiate real from generated images by focusing on projective geometry anomalies. The classifier evaluates the consistency of these fields with projective geometry principles, scoring images on their geometric plausibility. This method enables precise evaluation of projective geometry, enhancing the detection of subtle inconsistencies. + +# 5.3. Line Segment Cues + +Our method also assesses projective geometry in generated images by identifying key structural lines using Deep + +LSD [31]. These lines indicate adherence to perspective rules. We then train a PointNet-like architecture [34] to classify images based on line segment patterns, differentiating real from generated images. + +PointNet's flexibility in handling unordered data makes it suitable for analyzing line segments without pre-sorting. The model assigns scores representing the likelihood of an image being real based on the spatial arrangement of its lines. Analyzing these scores reveals the model's ability to detect subtle discrepancies in line arrangements, which often indicate a generated image. + +# 6. Evaluation + +Based on our prequalification analysis, we use Kandinsky as the primary source for our generated data to train our classifiers. We evaluated our three geometry-derived cues classifiers and analyzed their ability to generalize to other generated images that were not used during training. In addition, we performed a GradCam analysis [40] to identify any potential geometric discrepancies in images. + +# 6.1. Classifiers Results + +Fig. 3 shows ROC curves for each of our methods on indoor, outdoor and combined (indoor+outdoor) scenes. In each case, classifiers are trained on images that are not prequalified and tested on prequalified scenes, meaning that performance estimates are biased low — likely training on prequalified data would lead to even more accurate classification. Each classifier is effective, with AUCs ranging from 0.72 to 0.97. Recall these classifiers see only derived geometric features and do not see the image itself. + +Qualitative examples using Grad-CAM appear in Fig. 5 and Fig. 4. Notice how images that might be acceptable to a line analysis often fail a shadow analysis. Fig. 6 shows + +![](images/942f752e9f85aaf49934d8be94d45f48ece85e605fa1f214c084d53aeb57f06a.jpg) +(a) Easy Test Set (Indoor) + +![](images/4bf31823834387f2944a1e72c846da56a23f0ac817d8a152e0596f19057a64b1.jpg) +(b) Recent Timestamp Set (Indoor) + +![](images/70fd52c9bc63195813b2679a2ce2d677a821558cbdefab8a3fa930033cbfaa74.jpg) +(c) Unconfident Test Set (Indoor) + +![](images/f909efb9d38490dc91d45b0ae140c5ca15290232852dffa7e323b57445490ace.jpg) +(d) Misclassified Test Set (Indoor) + +![](images/52292a87664af03964d6120e3917278f6594e68fbd20712ffbc9c44e466b1779.jpg) +(e) Easy Test Set (Outdoor) + +![](images/c3a00626eff5653177a518adf3b8cc3c10a14ef8f16067dc9d94489dd6259b6b.jpg) +(f) Recent Timestamp Set (Outdoor) + +![](images/06a598880c2d633342c2dca6f2025b5bfde862e43a50d93ae53595f8ac87e21d.jpg) +(g) Unconfident Test Set (Outdoor) + +![](images/08a9e15ffe4e014ee735bb2971300a9842b272034883d19d9cdf590e05ebd46c.jpg) +(h) Misclassified Test Set ((Outdoor)) + +![](images/72ba2fe439ea365d2202157e111aa867d01f9eff53ff5c9a84616cbac7aa4024.jpg) +(i) Easy Test Set (Combined) + +![](images/9f01e36b4bb3fb13a25d434f5d1443ff28a95aaaef12e55bf41f693f79464eb9.jpg) +(j) Recent Timestamp Set (Combined) + +![](images/fe1a2ef18e55c66c4a54cbf3dd9eaf37b31efeafbf68c36f028649201aedea3e.jpg) +Figure 3. ROC Curves Assessing Projective Geometry Cues in Generated Images trained on Kandinsky-v3. We trained separate models for indoor scenes (top row), outdoor scenes (middle row), and a combination of indoor and outdoor scenes (last row). All our derived geometry cues classifiers are trained without looking at image intensity information and can reliably detect projective geometry errors. The recent timestamp test set (second column) confirms that these models are robust. We find hard examples using a prequalifier trained on image pixels. Our derived geometry cues consistently show high AUC for finding projective geometry errors on hard test sets – the last two columns – unconfident and misclassified test sets. For the unconfident test set, where the prequalifier has an AUC of 0.51 (c), 0.51 (g), and 0.49 (h) for indoor, outdoor, and combined partition, our classifiers can still accurately identify the generated images with high AUCs – 0.82 from line segments in the indoor set, 0.84 from perspective fields in the outdoor set, and 0.80 from perspective field cues and object shadows in the combined set. Similarly, for the misclassified test set, where the prequalifier has an AUC of 0.00, as it should, our classifiers remain reliable with AUC up to 0.82. We conclude that generated images contain geometric structures not seen in real images, and these structures very reliably identify generated images by only looking at derived geometry cues. + +![](images/04e09062dca53bb2e4e90640ff2bf16401b5fa436e2d21c599e71052569f5f6b.jpg) +(k) Unconfident Test Set (Combined) +(1) Misclassified Test Set (Combined) + +examples to emphasize this point. + +# 7. Other Generators Evaluated + +Our classifiers do not see pixels, but derived geometric features. This means that one could expect a form of generalization across generators. We illustrate that this generalization occurs - ROC curves in Fig. 7 demonstrate that classifiers trained to distinguish Kandinsky images from real images can also reliably distinguish Stable Diffusion XL [3], DeepFloyd [12] and PixArt- $\alpha$ [9] from the open-source domain. Additionally, we assess the efficacy of our models against images from proprietary generators such as OpenAI's Dalle-3[4] and Adobe's Firefly [1], representing some of the most advanced tools in image generation. Finally, we show we can detect composite made by a recent SOTA method [29] + +by looking at Object-Shadow cues in the supplement. + +# 8. Discussion + +We have shown that generated images can be reliably distinguished from real images by looking only at derived geometric cues. This is likely because image generators do not fully implement the geometry one observes in real images. Producing accurate perspective geometry or accurate shadow geometry requires very tight coordination of detailed information over very long spatial scales. Our results, together with the notorious tendency of face image generators to award subjects' left and right earlobes of different shapes, suggest that doing so is beyond the capacity of current generators. Our findings have significant implications for the development of image generation models, as the inability to + +![](images/1f03b11fe97913453204c040f1c223e16097ce8acd125ea5e11147387c9de8fb.jpg) +Generated Image + +![](images/755700a65a7865b9578bbff74918413aadbb17b88241014f0a13be9f28b4f13e.jpg) +Shadow Errors + +![](images/d7e5e9612ef2d9b8a94c246a55a418a468e454602e08c2cb1a6709eea17d3694.jpg) +Object-Shadow (OS) + +![](images/4e161973637eee040d3749a00fbf692ad2f3ec09b10138c220522f0df99880f0.jpg) +OS GradCam + +![](images/d677452df3a49e08fba7d4dd7b817719efb6ce92e03d25ae259265c8d935c590.jpg) +VP Errors + +![](images/e240728e1277b5454f1d0d465048a064968d64d7695ba8d8564e8cfffcfca7d0.jpg) +Perspective Fields (PF) + +![](images/9a72a1de97e8ce46d5dcf5aaa3e914efbb23a3edf003ed0ad11a3f57aca727cf.jpg) +PF GradCam + +![](images/7b29ce764b8a4b55151a3d798bb1be69f2d35e04f66643c2f82e825a3400b105.jpg) +Figure 4. Grad-CAM applied to our Object-Shadow and Perspective Field classifiers reveals that the high AUCs in Fig. 3 are based on real geometric errors in indoor scenes generated by Kandinsky, Stable Diffusion XL and Dalle-3 shown in each row respectively. The second and fifth columns highlight shadow and vanishing point errors, respectively. The third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) identifies diagnostic areas for synthetic generation, such as inconsistent shadow directions (in all three rows), mismatched shadow lengths (second row). The sixth column shows Perspective Fields [21], and Grad-CAM applied to our Perspective Fields classifier (last column) reveals geometric errors in all three rows, particularly at ceilings and side walls, with noticeable errors also present in window grills in the first and second rows. +Generated Image +Figure 5. Grad-CAM results for outdoor scenes generated by Kandinsky, Stable Diffusion XL, and Adobe Firefly, shown in each row respectively. The second column highlights shadow errors, while the third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) reveals incorrect shadow shapes in the first and second rows, with shadows on the right-side pedestrians pointing in a different direction than those on the left. The fifth column shows vanishing point errors, and the sixth column presents Perspective Fields [21]. Grad-CAM applied to our Perspective Fields classifier (last column) confirms large perspective distortions on building facades and road markings, corroborating the vanishing point errors in the fifth column. + +![](images/46613e0f5ced8ad5cdbedc76e3e6de7be3eabc6e875f80cfb3b6090bbcf6579d.jpg) +Shadow Errors + +![](images/b0b7d98494813662f22bf3115a91e994a8cf20c7131b4da498c66861f61c54df.jpg) +Object-Shadow (OS) + +![](images/8608d1484e24da2cb303234a8ed8563efa991b9221f09dd796661ca2eec3f844.jpg) +OS GradCam + +![](images/eff68bc213fb46bce0f720a6a348c4fd4b5ab6690e3968091978327ca827b305.jpg) +VP Errors + +![](images/bd26a9c9015c1dd42bdb005ca3b7b31fdf66ca3f7d31eac38eed3745704f8b3c.jpg) +Perspective Fields + +![](images/cdfb7cd164f2dcdc88ae6a5169c3f92dc9c096b33ec2514a718f3335bca33c0f.jpg) +PF GradCam + +![](images/fbd23d4df6975b08322542fc25e3291fc55374ebf81eb28e7e94978e7f5cbe69.jpg) +Generated Image + +![](images/b56195086b42baf9bf96beaa67027d1b66f01aff2c5c7d061b0a03a50f69c20f.jpg) +Shadow Errors + +![](images/0ea3fbeca0ff02014050f679365c6846d0740185e234e1fa3f116e12533f60a7.jpg) +Object-Shadow (OS) +OS GradCam + +![](images/bfe1453f9333e422e720876b5a23fd3a2a4ad141b9edf68270eec72282f0096c.jpg) +VP Errors + +![](images/f41c45cf67d1cce83f52983cd097933a0d53cb1bfb50900d05a756b331bf4c89.jpg) +Perspective Fields (PF) + +![](images/2f86dd4917360c31ad438575844df4c2323858ffbceabd4f525aef4b91ed155e.jpg) +PF GradCam + +![](images/9f03e24a7fae66711fc8a2eab520ebd124703c5d72c8d377fc380287a85158ad.jpg) +Figure 6. Our projective geometry classifiers identify distinct types of problems in generated images. The top row presents an example that was classified as real by the Object-Shadow classifier but correctly identified as generated by the Perspective Fields classifier. While the shadow cast by the person appears realistic, the Perspective Fields Grad-CAM highlights the problematic geometry of the shelf on the top left. In contrast, the bottom row shows an example that was correctly identified as generated by the Object-Shadow classifier but misclassified as real by the Perspective Fields classifier. Although the perspective effects in the image appear plausible, the Grad-CAM weights correctly reveal that the two chairs are casting shadows from different light sources, indicating inconsistency in scene's illumination. + +![](images/3953b76abbd53bed9dae5fa5e41e8acdc2e2f691a617452e97f9c47210a6a623.jpg) +(a) Dall-E 3 (Indoor) +(f) Dall-E 3 (Outdoor) +Figure 7. Our model trained on Kandinsky-v3 generalizes in detecting projective geometry errors from images generated by various unseen models. We evaluate the model's performance on test sets generated by Dall-E 3 (a), DeepFloyd (b), Adobe Firefly (c), PixArt-α (d), and Stable Diffusion XL or SDXL (e) using the same text prompts from the 'misclassified' Kandinsky-v3 generated test set. The top row shows indoor scenes and the last row shows outdoor scenes. Our model generalizes to all evaluated generators except DeepFloyd. However, DeepFloyd-generated images can be reliably detected when the model is trained on a DeepFloyd-generated training set, but it shows poor generalization capabilities to other generators compared to Kandinsky. + +![](images/b62b039f15e4ac996405a4c3c00a030b92b852097867c2d2baa0160032439527.jpg) + +![](images/6596c2f96e1841b52913430e72cb49ec4de3aa4cbe2b153a71b1fabab1b7e7ec.jpg) +(b) DeepFloyd (Indoor) +(g) DeepFloyd (Outdoor) + +![](images/b1241234b0e8d266d2a52ac7be29c687173b7e26af57407cf3501a32f750cd13.jpg) + +![](images/7739ab8650e5921e348f21eb161b8dce924dfd70c5ac2b7f66e17d4e7a0e8ab2.jpg) +(c) Adobe Firefly (Indoor) +(h) Adobe Firefly (Outdoor) + +![](images/0a1637bf9feef300a5872265c306f1daa0ba5b2ff248361b28234e3ccf047723.jpg) + +![](images/cfd3c4558800f71d84e2bdf0d815ae8181b8f7664870c6695cf6ae52030bc9d9.jpg) +(d) $\mathrm{PixArt - }\alpha$ (Indoor) +(i)PixArt- $\alpha$ (Outdoor) + +![](images/b2051042f8a4d44a2fe8c89f549da0c1d2de0a1038ed8e6563bd5e60acdf4370.jpg) + +![](images/4542a2733bf4625fd8ccaf3672ff13fdfc3fd03c6c77c5bda9db5a1a888aaba0.jpg) +(e) SDXL (Indoor) +(j) SDXL (Outdoor) + +accurately replicate projective geometry extends across various state-of-the-art models, indicating a widespread issue rather than a problem specific to a particular generator. + +We speculate that fixing this difficulty requires structural innovation in the generator, rather than simply exposing the generator to more data. The complex interplay of light, shadows, and perspective in real-world scenes may necessitate novel approaches to modeling and encoding spatial relationships within the generator's architecture. Potential avenues for improvement could include incorporating explicit geometric reasoning or developing new loss functions that prioritize the consistency of projective geometry. Fur + +thermore, our work highlights the importance of developing robust evaluation metrics for image generation models that assess the geometric coherence of generated images. By tackling these challenges, we could create image-generation models that more faithfully capture the complex geometric relationships in real-world scenes. + +# Acknowledgment + +This paper is based on work supported in part by the National Science Foundation under Grant No. 2106825 and a gift from Boeing Corporation. + +# References + +[1] Adobe. Firefly. https://www.adobe.com/sensei/generative-ai/firefly.html, 2023. Accessed: 2023-11. 2,6 +[2] Shruti Agarwal and Hany Farid. Photo forensics fromJPEG dimples. In 2017 IEEE workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2017. 2 +[3] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report, 2023. 2, 3, 6 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 2, 6 +[5] Anand Bhattachad and D.A. Forsyth. Stylitgan: Prompting stylegan to generate new illumination conditions. In arXiv, 2023. 1 +[6] Anand Bhattachad, Daniel McKee, Derek Hoiem, and DA Forsyth. Stylegan knows normal, depth, albedo, and more. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2 +[7] Matyáš Boháček and Hany Farid. A geometric and photometric exploration of gan and diffusion synthesized faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-883, 2023. 2 +[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 2 +[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3, 6 +[10] Yida Chen, Fernanda Viégas, and Martin Wattenberg. Beyond surface statistics: Scene representations in a latent diffusion model. arXiv preprint arXiv:2306.05720, 2023. 1, 2 +[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-image generation. arXiv preprint arXiv:2310.18235, 2023. 2 +[12] Deep Floyd. Iterative filter. https://github.com/deep-floyd/IF, 2023. Accessed: 2023-11. 2, 3, 6 +[13] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023. 1, 2 +[14] David C. Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 382-392, 2023. 2 +[15] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2 + +[16] David A Forsyth and Jean Ponce. Computer vision: a modern approach. prentice hall professional technical reference, 2002. 3 +[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2 +[19] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In ECCV, 2018. 2 +[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3 +[21] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 2, 5, 7, 16, 17 +[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2 +[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. +[24] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2 +[25] Eric Kee, James F O'Brien, and Hany Farid. Exposing photo manipulation with inconsistent shadows. ACM Transactions on Graphics (ToG), 32(3):1-12, 2013. 2 +[26] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 2 +[27] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, Minguk Kang, Taesung Park, Jure Leskovec, Jun-Yan Zhu, Li Fei-Fei, Jiajun Wu, Stefano Ermon, and Percy Liang. Holistic evaluation of text-to-image models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2 + +[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3 +[29] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6 +[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 5000–5009, 2017. 3 +[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3 +[33] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 53(2):758-767, 2005. 2 +[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 5 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. +[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2 +[39] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2 + +[40] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 128(2):336-359, 2019. 5 +[41] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18310-18319, 2023. 11 +[42] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In ICCV, 2019. 2 +[43] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2, 12 +[44] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3259-3273, 2022. 5, 7, 16, 17 +[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22445-22455, 2023. 2 +[46] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 3 +[47] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3 +[48] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2 +[49] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 1, 2 +[50] Peng Zhou, Xintong Han, Vlad I Morariu, and Larry S Davis. Learning rich features for image manipulation detection. In CVPR, 2018. 2 \ No newline at end of file diff --git a/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/images.zip b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f38a387ac7c0edcf3a40bcf0a7eabb426bba113 --- /dev/null +++ b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:662ea05e147cfa594cf68a6a458a347450bcb0f5fc2bd1ef11227807f91149a3 +size 1126703 diff --git a/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/layout.json b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc5d5ae9e0d2592415e9ac0f791b225eecb3d76c --- /dev/null +++ b/2024/Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now/layout.json @@ -0,0 +1,9487 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 486, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 486, + 140 + ], + "type": "text", + "content": "Shadows Don't Lie and Lines Can't Bend! \nGenerative Models don't know Projective Geometry...for now" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 160, + 484, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 484, + 175 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 484, + 175 + ], + "type": "text", + "content": "Ayush Sarkar*1 Hanlin Mai*1 Amitabh Mahapatra*1 Svetlana Lazebnik" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 216, + 176, + 377, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 176, + 377, + 189 + ], + "spans": [ + { + "bbox": [ + 216, + 176, + 377, + 189 + ], + "type": "text", + "content": "D.A. Forsyth1 Anand Bhattachad2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 190, + 509, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 190, + 509, + 204 + ], + "spans": [ + { + "bbox": [ + 85, + 190, + 509, + 204 + ], + "type": "text", + "content": "1University of Illinois Urbana-Champaign 2Toyota Technological Institute at Chicago" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 157, + 204, + 432, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 204, + 432, + 217 + ], + "spans": [ + { + "bbox": [ + 157, + 204, + 432, + 217 + ], + "type": "text", + "content": "https://projective-geometry.github.io/" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 229, + 149, + 430 + ], + "blocks": [ + { + "bbox": [ + 50, + 229, + 149, + 430 + ], + "lines": [ + { + "bbox": [ + 50, + 229, + 149, + 430 + ], + "spans": [ + { + "bbox": [ + 50, + 229, + 149, + 430 + ], + "type": "image", + "image_path": "2a816408b7c8608b4adca114e38103aee9b4dd5338bca4b84e38dff871712a89.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 431, + 127, + 441 + ], + "lines": [ + { + "bbox": [ + 70, + 431, + 127, + 441 + ], + "spans": [ + { + "bbox": [ + 70, + 431, + 127, + 441 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 149, + 229, + 249, + 429 + ], + "blocks": [ + { + "bbox": [ + 149, + 229, + 249, + 429 + ], + "lines": [ + { + "bbox": [ + 149, + 229, + 249, + 429 + ], + "spans": [ + { + "bbox": [ + 149, + 229, + 249, + 429 + ], + "type": "image", + "image_path": "14825c8c4e005b253de46315b0323fc37245196dfed9c0645eea5309dd1c48e1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 431, + 223, + 440 + ], + "lines": [ + { + "bbox": [ + 174, + 431, + 223, + 440 + ], + "spans": [ + { + "bbox": [ + 174, + 431, + 223, + 440 + ], + "type": "text", + "content": "Shadow Errors" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 249, + 229, + 349, + 430 + ], + "blocks": [ + { + "bbox": [ + 249, + 229, + 349, + 430 + ], + "lines": [ + { + "bbox": [ + 249, + 229, + 349, + 430 + ], + "spans": [ + { + "bbox": [ + 249, + 229, + 349, + 430 + ], + "type": "image", + "image_path": "ecda4d8f0f8638623144ae24289797545afd64b5e07f3aa042ae87ee1725af02.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 431, + 337, + 440 + ], + "lines": [ + { + "bbox": [ + 258, + 431, + 337, + 440 + ], + "spans": [ + { + "bbox": [ + 258, + 431, + 337, + 440 + ], + "type": "text", + "content": "Detected Shadow Errors" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 348, + 229, + 448, + 429 + ], + "blocks": [ + { + "bbox": [ + 348, + 229, + 448, + 429 + ], + "lines": [ + { + "bbox": [ + 348, + 229, + 448, + 429 + ], + "spans": [ + { + "bbox": [ + 348, + 229, + 448, + 429 + ], + "type": "image", + "image_path": "4df22eacc828692978e1171d6b0d76404f455b8219035fc769521e211018f2e2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 431, + 434, + 441 + ], + "lines": [ + { + "bbox": [ + 360, + 431, + 434, + 441 + ], + "spans": [ + { + "bbox": [ + 360, + 431, + 434, + 441 + ], + "type": "text", + "content": "Vanishing Point Errors" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 448, + 229, + 547, + 430 + ], + "blocks": [ + { + "bbox": [ + 448, + 229, + 547, + 430 + ], + "lines": [ + { + "bbox": [ + 448, + 229, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 448, + 229, + 547, + 430 + ], + "type": "image", + "image_path": "7e15621b0bb55c86cdfa4356ce456c94c285d2f99d2eada6041052bf96d05e50.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 431, + 541, + 441 + ], + "lines": [ + { + "bbox": [ + 451, + 431, + 541, + 441 + ], + "spans": [ + { + "bbox": [ + 451, + 431, + 541, + 441 + ], + "type": "text", + "content": "Detected Perspective Errors" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 448, + 547, + 493 + ], + "lines": [ + { + "bbox": [ + 46, + 448, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 547, + 493 + ], + "type": "text", + "content": "Figure 1. The first column presents visually compelling AI-generated images. However, a closer examination reveals fundamental inconsistencies, such as those in shadow alignment (second column) and vanishing point accuracy (fourth column). Our model's analysis, shown in the third and fifth columns, detects these shadow and perspective geometry errors. We show that these errors are systematic and can be used to identify generated images." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 143, + 502, + 192, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 502, + 192, + 515 + ], + "spans": [ + { + "bbox": [ + 143, + 502, + 192, + 515 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 45, + 528, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 528, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 45, + 528, + 289, + 696 + ], + "type": "text", + "content": "Generative models can produce impressively realistic images. This paper demonstrates that generated images have geometric features different from those of real images. We build a set of collections of generated images, prequalified to fool simple, signal-based classifiers into believing they are real. We then show that prequalified generated images can be identified reliably by classifiers that only look at geometric properties. We use three such classifiers. All three classifiers are denied access to image pixels, and look only at derived geometric features. The first classifier looks at the perspective field of the image, the second looks at lines detected in the image, and the third looks at relations between detected objects and shadows. Our procedure detects generated images more reliably than SOTA local signal based detectors," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 504, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 547, + 552 + ], + "type": "text", + "content": "for images from a number of distinct generators. Saliency maps suggest that the classifiers can identify geometric problems reliably. We conclude that current generators cannot reliably reproduce geometric properties of real images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 574, + 387, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 387, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 387, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "content": "Both StyleGAN [22-24] and diffusion models [36-38] are renowned for generating images that are strikingly similar to real-world photos and consistently fool people. But, as we show, generated images have distinctive geometric features, likely from a failure to fully capture projective geometry." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Bhattad et al. [5, 6], Chen et al. [10], Du et al. [13], Zhan et al. [49], and have shown generative models implicitly capture the complex scene properties, including normals, depth, albedo, and support relations. These works suggest these models \"understand\" geometry, which would" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 123, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 123, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 123, + 712 + ], + "type": "text", + "content": "*equal contribution" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28140" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "be useful for rendering 3D scenes. Our detailed, population-level analysis of generated images suggests generative models [1, 3, 4, 12, 32] cannot fully translate this \"understanding\" into accurate geometry. Specifically, we demonstrate that generative models produce images with lines that differ from those of real images (likely due to problems aligning vanishing points); that generative models produce images with perspective fields that are unlike those of real images; and that object-shadow relations in generated images differ reliably from those in real images. We use advanced pretrained models (Line Segment Detection [31]; Perspective Fields [21]; and PointNet [34]) that inspect geometric representations to distinguish between real and generated images. We rely on derived geometry cues from SOTA methods as manual image analysis and explicit geometry-based rules like drawing lines perspective lines are not scalable or automatable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 289, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 289, + 393 + ], + "type": "text", + "content": "To guarantee the accuracy of our findings, we follow a strict process of data curation. This involves using a controlled pixel-level classifier to filter out any biases related to color, texture, and local features within our test set. This precision in data selection is crucial to isolate and accurately assess the inconsistencies in projective geometry and illumination present in generated images. By carefully screening the data, we can ensure that our results are not obscured by common artifacts that are usually found in generated images. This enhances the reliability of our conclusions. Our contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 396, + 288, + 563 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 396, + 288, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 288, + 442 + ], + "type": "text", + "content": "- Unearthing Geometric Discrepancies: We present a comprehensive analysis that goes beyond existing literature to both demonstrate and quantify geometric discrepancies produced by current generative models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 288, + 479 + ], + "type": "text", + "content": "- Scalable, Data-Driven Approach: We introduce a scalable, data-driven approach by utilizing three distinct projective geometry cues to detect geometric inaccuracies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 525 + ], + "type": "text", + "content": "- Robust and Generalizable: We demonstrate robustness and generalizability by effectively identifying geometric errors across a wide array of images from recent generative models, including those with the latest time stamps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 527, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 287, + 563 + ], + "type": "text", + "content": "- Broadening the Scope of Model Assessment: Our approach can be used as an alternative method for evaluating models: do they get projective geometry right?" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 573, + 134, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 134, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 134, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 700 + ], + "type": "text", + "content": "**Generative Models:** The advancement of generative models, particularly in creating visually realistic images, marks a significant milestone in computer vision. Pioneering efforts by Karras et al. [22-24] with StyleGAN, and the emergence of diffusion models [36-38], have set new benchmarks in realism. These models, used in diverse fields from art to data augmentation, have yet to fully grasp the nuances of projective geometry, which is the focus of our analysis, primarily using open diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Geometric Understanding in Generative Models: While" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "studies like Bhattachad et al. [6], Du et al.[13], Chen et al. [10] and Zhan et al. [49] demonstrate these models' potential in understanding scene geometry, our work diverges by scrutinizing the generated images themselves, examining their adherence to the principles of projective geometry and illumination, rather than analyzing learned features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 148, + 548, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 548, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 548, + 267 + ], + "type": "text", + "content": "Detecting Generated Images: The realism of modern generative models has made image forensics increasingly challenging. Traditional methods focused on detecting synthetic images using signals like resampling artifacts [33] and JPEG quantization [2]. Kee et al. [25] introduced a geometric technique for detecting shadow inconsistencies, paralleling our pursuit of physical realism. However, our work extends beyond identifying photo manipulation to evaluating the overall perspective geometry and illumination consistency in images from generative models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 271, + 548, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 548, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 548, + 403 + ], + "type": "text", + "content": "Zhang et al.'s work [14] focuses on detecting AI-generated images using diverse generative models and online training for future model adaptation. Our research, in contrast, assesses the projective geometry in these images, examining their ability to render scenes with accurate perspective and illumination. Boháček et al. [7], while detecting geometric inconsistencies related to shadows, align with our interest in physical realism. However, we delve deeper, thoroughly evaluating perspective geometry and illumination in generative models for a more comprehensive understanding of their geometric accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 538 + ], + "type": "text", + "content": "The rise of generative methods has steered image forensics towards using discriminative methods to detect synthetic content [8, 15, 19, 42, 43, 48, 50]. These advancements align with our objective of analyzing the physical and geometrical congruence of generated images. However, our work goes a step further by critically assessing whether generative models fundamentally understand and accurately replicate projective geometry, rather than simply distinguishing between real and synthetic images. This deeper level of analysis aims to unveil the intricacies and limitations of current models in faithfully rendering geometrically coherent images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 542, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 638 + ], + "type": "text", + "content": "Evaluation Metrics: Traditional metrics like the Inception Score (IS) [39] and Fréchet Inception Distance (FID) [18] focus on pixel-level fidelity. The emergence of CLIP-based scores [17, 35] and DIRE [45] offers a semantic perspective. In contrast, our approach, distinct in its focus on perspective geometry and illumination consistency, seeks to ensure comprehensive realism, bridging the gap between visual and physical authenticity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Recent studies like Davidsonian Scene Graph [11] and ImagenHub [26] address fine-grained evaluation inconsistencies, while the HEIM benchmark [27] assesses models across multiple aspects. Our work complements these by providing an in-depth evaluation of the physical and geometric realism of images generated by state-of-the-art models." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28141" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 250, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 250, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 250, + 85 + ], + "type": "text", + "content": "3. Background on Projective Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 210 + ], + "type": "text", + "content": "Projective geometry is a mathematical framework that enables the accurate representation of three-dimensional spaces in two-dimensional images. It provides the rules for perspective, which are crucial for creating realistic scenes with depth and spatial orientation [16]. In this section, we will examine the common inconsistencies that may arise during image synthesis according to projective geometry. Our evaluation framework is intended to detect and measure these discrepancies, which are essential for evaluating the realism and physical plausibility of generated images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 212, + 288, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 288, + 284 + ], + "type": "text", + "content": "Inconsistent Vanishing Points. Vanishing points are fundamental to capturing the essence of perspective in images. They should align with the direction of parallel lines converging at a distance. Generated images often exhibit inconsistencies where these lines do not meet at the correct vanishing points, leading to a distorted sense of perspective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 288, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 288, + 344 + ], + "type": "text", + "content": "Lighting and Shadow Inconsistencies. Accurate shadows are essential for reinforcing the position and shape of objects within a scene. Discrepancies in shadow direction, length, and softness can indicate a misalignment with the scene's light sources, disrupting the image's three-dimensionality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 345, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 288, + 403 + ], + "type": "text", + "content": "Scale Discrepancies. The principle of size constancy dictates that objects of the same size should appear smaller as their distance from the observer increases. Generated images sometimes fail to maintain this scaling, resulting in a compromised depth perception." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 404, + 288, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 464 + ], + "type": "text", + "content": "Distortion of Geometric Figures. Geometric figures should maintain their shape when projected onto the image plane, barring intentional perspective distortion. Errors in this projection can result in circles appearing as ellipses or squares as trapezoids, indicating a flawed perspective rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 465, + 288, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 288, + 523 + ], + "type": "text", + "content": "Depth Cues. Depth perception in images is conveyed through cues such as overlapping, texture gradients, and relative size. Misrepresentation of these cues can lead to an unnatural spatial arrangement that the human eye can readily detect as artificial." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 525, + 288, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 644 + ], + "type": "text", + "content": "Our evaluation framework, detailed in the subsequent sections, is designed to rigorously test generated images against these projective geometry principles. While a comprehensive evaluation of projective geometry would consider all the aforementioned inconsistencies, our framework prioritizes the detection of inconsistent vanishing points and lighting and shadow inconsistencies. These elements are particularly telling indicators of an image's projective geometry realism and are often the most challenging for generative models to replicate accurately." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 657, + 151, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 151, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 151, + 669 + ], + "type": "text", + "content": "4. Dataset Curation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Our data curation process is carefully designed to distinguish between real images and generated images from several generative models. This process includes models that the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "classifier has not seen during its training. Another important goal is to ensure our prequalifier effectively identifies images with recent timestamps. This helps prevent the classifier from relying on cues specific to the dataset, such as whether the image is from the training distribution or not, rather than determining whether the image is real or generated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 152, + 387, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 387, + 164 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 387, + 164 + ], + "type": "text", + "content": "4.1. Real Images" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 170, + 527, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 527, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 527, + 182 + ], + "type": "text", + "content": "We experiment with a diverse set of images, including:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 182, + 547, + 420 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 182, + 547, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 241 + ], + "type": "text", + "content": "(a) Indoor Scenes: A collection of 400,000 interior images featuring a variety of furniture arrangements and lighting conditions, sourced from LSUN [46]. Specifically, we used 100,000 images each from Bedroom, Dining Room, Kitchen, and Living Room categories." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 242, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 300 + ], + "type": "text", + "content": "(b) Outdoor Scenes: A dataset of 125,000 outdoor scenes with varying landscapes and urban settings, sourced from Berkeley Deep Drive 100K [47] and Mapillary Vistas [30]. We selected images that represent a wide range of weather conditions and times of day." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 301, + 547, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 547, + 336 + ], + "type": "text", + "content": "(c) Combination of Indoor and Outdoor scenes: We also analyzed a combination of above indoor and outdoor scenes to assess the performance on a more diverse dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 337, + 547, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 420 + ], + "type": "text", + "content": "(d) Recent Timestamp Images: A curated test set of 500 indoor and 500 outdoor images with timestamps ranging from May 2023 to March 2024. These images were collected from various social media platforms and online sources to ensure our classifier's ability to handle recent real-world data and that the models are not obscured because of any data-source-specific biases." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 428, + 403, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 428, + 403, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 428, + 403, + 441 + ], + "type": "text", + "content": "4.2. Image Captions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 446, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 543 + ], + "type": "text", + "content": "We use the ViT-bigG-14/laion2b_s39b_b160k model [20] and the BLIP model [28] in succession, to generate refined captions for real images. These models were chosen for their state-of-the-art performance in image captioning tasks. The ViT model uses a Vision Transformer architecture, while BLIP employs a multimodal pre-training approach. By using common captions, we ensure a fair evaluation of the projective geometry in generated images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 550, + 415, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 550, + 415, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 415, + 562 + ], + "type": "text", + "content": "4.3. Generated Images" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": "We generate images from Stable Diffusion XL v1.0 [32], Kandinsky-v3 [3], DeepFloyd IF v1.0 [12], and PixArt- " + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": " v1.0 [9]. We use the same caption from the real images to generate these images with the default settings of each generative model." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 635, + 421, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 421, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 421, + 647 + ], + "type": "text", + "content": "4.4. Robust Prequalifier" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Our goal is to identify challenging images that a signal-based classifier may struggle to differentiate. Therefore, we need to develop a robust prequalifier. This is to ensure that our results are not affected by any false data signals. Thus, we aim to eliminate all potential factors that may influence" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28142" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 129, + 531, + 243 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "type": "text", + "content": "Table 1. Statistical overview of the Data Curation and Filtering Process: We present the distribution of real and generated images for indoor and outdoor datasets. The ResNet50 Prequalifier helps us curate datasets, creating an 'Unconfident Set' for images with low classifier certainty and a 'Misclassified Subset' for images incorrectly labeled by the classifier. These rigorously curated sets are instrumental for our subsequent analysis, concentrating on projective geometry while mitigating the influence of signal cues. The datasets for each prequalifier—indoor, outdoor, and combined—are prepared separately to tailor the models to their specific contexts." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 129, + 531, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 129, + 531, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 129, + 531, + 243 + ], + "type": "table", + "html": "
Indoor RealIndoor GeneratedOutdoor RealOutdoor GeneratedCombined RealCombined Generated
Total Images400,000400,000125,000125,000525,000525,000
Training and Test Sets
Training Set Size75,00075,00025,00025,000100,000100,000
Validation Set Size10,00010,0005,0005,00015,00015,000
Test Set Size315,000315,00095,00095,000410,000410,000
Post ResNet50 Prequalifier
Unconfident Set232131384014441018443922540
Misclassified Subset10399575660944323800825
", + "image_path": "d5509c7fce4085d9b8ab2ccf28b73f647ce10076f5823fffe5e1c03d71811dac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 256, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 288, + 316 + ], + "type": "text", + "content": "the geometry cues we obtain because of signal weirdness in the generated images. To accomplish this, we begin by training signal-based classifiers using a vanilla ResNet-50 and primarily concentrate on identifying instances that prove challenging for these signal-based classifiers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 288, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 288, + 469 + ], + "type": "text", + "content": "We trained prequalifiers on three distinct settings - indoor scenes, outdoor scenes and a combination of the two. Each prequalifier was trained on a dataset consisting of real images and images generated by one of the four models: Stable Diffusion XL, Kandinsky-v3, DeepFloyd IF, and PixArt-" + }, + { + "bbox": [ + 46, + 325, + 288, + 469 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 325, + 288, + 469 + ], + "type": "text", + "content": ". While all four types of prequalifiers performed well on their respective test sets, we found that Kandinsky exhibited superior generalization capabilities when evaluated on images generated by other models or on recent timestamp images. The fundamental experiment conducted to assess the generalizability of different generators is covered in Figure 9 in the Supplementary Material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 477, + 288, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 288, + 610 + ], + "type": "text", + "content": "For indoor scenes, our Kandinsky-based prequalifier achieves an Area Under the Curve (AUC) of 0.99 on its test set with an accuracy of 97.43 and maintained high performance on images generated by other models such as Stable Diffusion XL and PixArt- " + }, + { + "bbox": [ + 46, + 477, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 477, + 288, + 610 + ], + "type": "text", + "content": ", achieving AUCs of 0.97 and 0.98 respectively. Furthermore, our Kandinsky-trained prequalifier meant for both indoor and outdoor settings combined showed robust performance on recent timestamp generated indoor and outdoor images, achieving AUC scores of 0.90, 0.95, and 0.72 on images generated by Stable Diffusion XL, PixArt- " + }, + { + "bbox": [ + 46, + 477, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 477, + 288, + 610 + ], + "type": "text", + "content": ", and DeepFloyd IF, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Given their robustness and strong generalizability, we selected the Kandinsky-trained prequalifiers as the basis for training our derived geometry classifiers on Kandinsky-generated images. This choice ensures that our classifiers effectively look at geometric discrepancies and can handle various generated images. This enhances the reliability of our approach and supports our conclusion that the classifiers are not affected by spurious signal artifacts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 256, + 397, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 397, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 397, + 268 + ], + "type": "text", + "content": "4.5. Final Test Sets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 275, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 547, + 407 + ], + "type": "text", + "content": "We categorize our test sets into three groups based on the accuracy of the prequalifier: easy, unconfident, and misclassified. The easy test set includes images with " + }, + { + "bbox": [ + 304, + 275, + 547, + 407 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 275, + 547, + 407 + ], + "type": "text", + "content": " accuracy, indicating that the prequalifier can reliably distinguish between real and generated images. The unconfident test set includes images where the prequalifier performs at chance level, suggesting that it struggles to make confident predictions. Finally, the misclassified test set includes images where the prequalifier makes completely wrong predictions, either classifying real images as generated or vice versa. A summary of this split for Kandinsky-v3 is provided in Tab. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 407, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 478 + ], + "type": "text", + "content": "Our approach primarily evaluates the \"hard set\" (unconfident and misclassified) for geometric and shadow inconsistencies, assessing adherence to projective geometry principles. This ensures rigorous testing of generative models' ability to reproduce geometric correctness and photometric accuracy, beyond surface-level or signal details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 479, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 574 + ], + "type": "text", + "content": "It is important to note that projective geometry inconsistencies are prevalent in generated images but often go undetected by conventional methods. Our models, trained on geometric abstractions and projective cues, can identify subtle but critical inaccuracies that texture artifacts cannot explain. This distinction is critical, as it allows us to rigorously test our models, which, unlike the prequalifier, do not have direct access to the images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 574, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 547, + 658 + ], + "type": "text", + "content": "We employ a suite of models to capture different facets of projective geometry, trained on datasets emphasizing geometric consistency and photometric accuracy. By combining their strengths, we comprehensively assess the quality of generated images and provide insights into the limitations and potential improvements of generative models from a projective geometry perspective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 669, + 481, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 669, + 481, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 669, + 481, + 684 + ], + "type": "text", + "content": "5. Analyzing Projective Geometry" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "For our analysis, we rely on three geometry cues – object shadow, line segments, and perspective fields. We train" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28143" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 202, + 252 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 202, + 252 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 202, + 252 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 202, + 252 + ], + "type": "image", + "image_path": "aaa6eefb1677ccc2dd06713ec5425e9e5f852d5983b54c9cb8963bc7331de60a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 254, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 254, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 547, + 288 + ], + "type": "text", + "content": "Figure 2. We train three classifiers to identify discrepancies in projective geometry. These classifiers are trained using derived geometry cues such as object-shadow associations (left), perspective fields (middle), and line segments (right) without looking at image intensity. We use the ResNet architecture for Object-Shadow and Perspective Fields, and PointNet for Line Segments to process unordered data sets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 76, + 362, + 252 + ], + "blocks": [ + { + "bbox": [ + 205, + 76, + 362, + 252 + ], + "lines": [ + { + "bbox": [ + 205, + 76, + 362, + 252 + ], + "spans": [ + { + "bbox": [ + 205, + 76, + 362, + 252 + ], + "type": "image", + "image_path": "f19d44804ed4395cdf3bef925bfbc08e184cae01bae3034fe056c47aac52e121.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 76, + 543, + 252 + ], + "blocks": [ + { + "bbox": [ + 370, + 76, + 543, + 252 + ], + "lines": [ + { + "bbox": [ + 370, + 76, + 543, + 252 + ], + "spans": [ + { + "bbox": [ + 370, + 76, + 543, + 252 + ], + "type": "image", + "image_path": "29de772733505b4f9d7466825c0dfb740ac54c9d81497a5a29d605655c8f5150.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 298, + 288, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 288, + 347 + ], + "type": "text", + "content": "three separate classifiers solely based on these derived geometry cues, without utilizing any pixel information. The training process for each classifier is described below, and a schematic pipeline of these classifiers can be found in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 354, + 168, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 168, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 168, + 366 + ], + "type": "text", + "content": "5.1. Object-Shadow Cues" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 372, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 408 + ], + "type": "text", + "content": "Our first model examines object-shadow relationships to address illumination. Shadows follow projective geometry principles, and inconsistencies can reveal generated images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 408, + 288, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 288, + 492 + ], + "type": "text", + "content": "We use a pretrained object-shadow instance detection model [44] to identify shadows and geometric heuristics to evaluate their plausibility given the objects and their orientation. A ResNet50 classifier is trained on binary masks of object and shadow instances to score the consistency of shadows with objects. Images with implausible shadows are marked as likely generated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 500, + 176, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 176, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 176, + 513 + ], + "type": "text", + "content": "5.2. Perspective Field Cues" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 518, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 578 + ], + "type": "text", + "content": "Our framework's second model uses Perspective Fields [21], vector fields encoding the spatial orientation of pixels relative to vanishing points and the horizon, to assess projective geometry. We generate these fields from single images using a pretrained model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 578, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 662 + ], + "type": "text", + "content": "We train a ResNet50 classifier on the Perspective Fields to differentiate real from generated images by focusing on projective geometry anomalies. The classifier evaluates the consistency of these fields with projective geometry principles, scoring images on their geometric plausibility. This method enables precise evaluation of projective geometry, enhancing the detection of subtle inconsistencies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 670, + 160, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 160, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 160, + 684 + ], + "type": "text", + "content": "5.3. Line Segment Cues" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "Our method also assesses projective geometry in generated images by identifying key structural lines using Deep" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 298, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 547, + 346 + ], + "type": "text", + "content": "LSD [31]. These lines indicate adherence to perspective rules. We then train a PointNet-like architecture [34] to classify images based on line segment patterns, differentiating real from generated images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 346, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 430 + ], + "type": "text", + "content": "PointNet's flexibility in handling unordered data makes it suitable for analyzing line segments without pre-sorting. The model assigns scores representing the likelihood of an image being real based on the spatial arrangement of its lines. Analyzing these scores reveals the model's ability to detect subtle discrepancies in line arrangements, which often indicate a generated image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 441, + 376, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 376, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 376, + 453 + ], + "type": "text", + "content": "6. Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 460, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 544 + ], + "type": "text", + "content": "Based on our prequalification analysis, we use Kandinsky as the primary source for our generated data to train our classifiers. We evaluated our three geometry-derived cues classifiers and analyzed their ability to generalize to other generated images that were not used during training. In addition, we performed a GradCam analysis [40] to identify any potential geometric discrepancies in images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 552, + 413, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 413, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 413, + 563 + ], + "type": "text", + "content": "6.1. Classifiers Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": "Fig. 3 shows ROC curves for each of our methods on indoor, outdoor and combined (indoor+outdoor) scenes. In each case, classifiers are trained on images that are not prequalified and tested on prequalified scenes, meaning that performance estimates are biased low — likely training on prequalified data would lead to even more accurate classification. Each classifier is effective, with AUCs ranging from 0.72 to 0.97. Recall these classifiers see only derived geometric features and do not see the image itself." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "Qualitative examples using Grad-CAM appear in Fig. 5 and Fig. 4. Notice how images that might be acceptable to a line analysis often fail a shadow analysis. Fig. 6 shows" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28144" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 167, + 165 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 167, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 167, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 167, + 165 + ], + "type": "image", + "image_path": "942f752e9f85aaf49934d8be94d45f48ece85e605fa1f214c084d53aeb57f06a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 168, + 150, + 178 + ], + "lines": [ + { + "bbox": [ + 67, + 168, + 150, + 178 + ], + "spans": [ + { + "bbox": [ + 67, + 168, + 150, + 178 + ], + "type": "text", + "content": "(a) Easy Test Set (Indoor)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 72, + 294, + 165 + ], + "blocks": [ + { + "bbox": [ + 177, + 72, + 294, + 165 + ], + "lines": [ + { + "bbox": [ + 177, + 72, + 294, + 165 + ], + "spans": [ + { + "bbox": [ + 177, + 72, + 294, + 165 + ], + "type": "image", + "image_path": "4bf31823834387f2944a1e72c846da56a23f0ac817d8a152e0596f19057a64b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 168, + 293, + 178 + ], + "lines": [ + { + "bbox": [ + 180, + 168, + 293, + 178 + ], + "spans": [ + { + "bbox": [ + 180, + 168, + 293, + 178 + ], + "type": "text", + "content": "(b) Recent Timestamp Set (Indoor)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 416, + 165 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 416, + 165 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 416, + 165 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 416, + 165 + ], + "type": "image", + "image_path": "70fd52c9bc63195813b2679a2ce2d677a821558cbdefab8a3fa930033cbfaa74.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 168, + 411, + 177 + ], + "lines": [ + { + "bbox": [ + 304, + 168, + 411, + 177 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 411, + 177 + ], + "type": "text", + "content": "(c) Unconfident Test Set (Indoor)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 426, + 72, + 543, + 165 + ], + "blocks": [ + { + "bbox": [ + 426, + 72, + 543, + 165 + ], + "lines": [ + { + "bbox": [ + 426, + 72, + 543, + 165 + ], + "spans": [ + { + "bbox": [ + 426, + 72, + 543, + 165 + ], + "type": "image", + "image_path": "f909efb9d38490dc91d45b0ae140c5ca15290232852dffa7e323b57445490ace.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 168, + 539, + 177 + ], + "lines": [ + { + "bbox": [ + 430, + 168, + 539, + 177 + ], + "spans": [ + { + "bbox": [ + 430, + 168, + 539, + 177 + ], + "type": "text", + "content": "(d) Misclassified Test Set (Indoor)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 180, + 167, + 273 + ], + "blocks": [ + { + "bbox": [ + 49, + 180, + 167, + 273 + ], + "lines": [ + { + "bbox": [ + 49, + 180, + 167, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 180, + 167, + 273 + ], + "type": "image", + "image_path": "52292a87664af03964d6120e3917278f6594e68fbd20712ffbc9c44e466b1779.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 277, + 153, + 286 + ], + "lines": [ + { + "bbox": [ + 64, + 277, + 153, + 286 + ], + "spans": [ + { + "bbox": [ + 64, + 277, + 153, + 286 + ], + "type": "text", + "content": "(e) Easy Test Set (Outdoor)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 177, + 180, + 294, + 273 + ], + "blocks": [ + { + "bbox": [ + 177, + 180, + 294, + 273 + ], + "lines": [ + { + "bbox": [ + 177, + 180, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 177, + 180, + 294, + 273 + ], + "type": "image", + "image_path": "c3a00626eff5653177a518adf3b8cc3c10a14ef8f16067dc9d94489dd6259b6b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 277, + 294, + 286 + ], + "lines": [ + { + "bbox": [ + 178, + 277, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 178, + 277, + 294, + 286 + ], + "type": "text", + "content": "(f) Recent Timestamp Set (Outdoor)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 299, + 180, + 416, + 273 + ], + "blocks": [ + { + "bbox": [ + 299, + 180, + 416, + 273 + ], + "lines": [ + { + "bbox": [ + 299, + 180, + 416, + 273 + ], + "spans": [ + { + "bbox": [ + 299, + 180, + 416, + 273 + ], + "type": "image", + "image_path": "06a598880c2d633342c2dca6f2025b5bfde862e43a50d93ae53595f8ac87e21d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 277, + 413, + 286 + ], + "lines": [ + { + "bbox": [ + 301, + 277, + 413, + 286 + ], + "spans": [ + { + "bbox": [ + 301, + 277, + 413, + 286 + ], + "type": "text", + "content": "(g) Unconfident Test Set (Outdoor)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 180, + 543, + 273 + ], + "blocks": [ + { + "bbox": [ + 425, + 180, + 543, + 273 + ], + "lines": [ + { + "bbox": [ + 425, + 180, + 543, + 273 + ], + "spans": [ + { + "bbox": [ + 425, + 180, + 543, + 273 + ], + "type": "image", + "image_path": "08a9e15ffe4e014ee735bb2971300a9842b272034883d19d9cdf590e05ebd46c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 277, + 544, + 286 + ], + "lines": [ + { + "bbox": [ + 425, + 277, + 544, + 286 + ], + "spans": [ + { + "bbox": [ + 425, + 277, + 544, + 286 + ], + "type": "text", + "content": "(h) Misclassified Test Set ((Outdoor))" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 289, + 167, + 381 + ], + "blocks": [ + { + "bbox": [ + 50, + 289, + 167, + 381 + ], + "lines": [ + { + "bbox": [ + 50, + 289, + 167, + 381 + ], + "spans": [ + { + "bbox": [ + 50, + 289, + 167, + 381 + ], + "type": "image", + "image_path": "72ba2fe439ea365d2202157e111aa867d01f9eff53ff5c9a84616cbac7aa4024.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 384, + 156, + 394 + ], + "lines": [ + { + "bbox": [ + 61, + 384, + 156, + 394 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 156, + 394 + ], + "type": "text", + "content": "(i) Easy Test Set (Combined)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 176, + 289, + 293, + 381 + ], + "blocks": [ + { + "bbox": [ + 176, + 289, + 293, + 381 + ], + "lines": [ + { + "bbox": [ + 176, + 289, + 293, + 381 + ], + "spans": [ + { + "bbox": [ + 176, + 289, + 293, + 381 + ], + "type": "image", + "image_path": "9f01e36b4bb3fb13a25d434f5d1443ff28a95aaaef12e55bf41f693f79464eb9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 384, + 295, + 394 + ], + "lines": [ + { + "bbox": [ + 174, + 384, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 174, + 384, + 295, + 394 + ], + "type": "text", + "content": "(j) Recent Timestamp Set (Combined)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 297, + 289, + 416, + 381 + ], + "blocks": [ + { + "bbox": [ + 297, + 289, + 416, + 381 + ], + "lines": [ + { + "bbox": [ + 297, + 289, + 416, + 381 + ], + "spans": [ + { + "bbox": [ + 297, + 289, + 416, + 381 + ], + "type": "image", + "image_path": "fe1a2ef18e55c66c4a54cbf3dd9eaf37b31efeafbf68c36f028649201aedea3e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 546, + 517 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 546, + 517 + ], + "type": "text", + "content": "Figure 3. ROC Curves Assessing Projective Geometry Cues in Generated Images trained on Kandinsky-v3. We trained separate models for indoor scenes (top row), outdoor scenes (middle row), and a combination of indoor and outdoor scenes (last row). All our derived geometry cues classifiers are trained without looking at image intensity information and can reliably detect projective geometry errors. The recent timestamp test set (second column) confirms that these models are robust. We find hard examples using a prequalifier trained on image pixels. Our derived geometry cues consistently show high AUC for finding projective geometry errors on hard test sets – the last two columns – unconfident and misclassified test sets. For the unconfident test set, where the prequalifier has an AUC of 0.51 (c), 0.51 (g), and 0.49 (h) for indoor, outdoor, and combined partition, our classifiers can still accurately identify the generated images with high AUCs – 0.82 from line segments in the indoor set, 0.84 from perspective fields in the outdoor set, and 0.80 from perspective field cues and object shadows in the combined set. Similarly, for the misclassified test set, where the prequalifier has an AUC of 0.00, as it should, our classifiers remain reliable with AUC up to 0.82. We conclude that generated images contain geometric structures not seen in real images, and these structures very reliably identify generated images by only looking at derived geometry cues." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 423, + 289, + 541, + 381 + ], + "blocks": [ + { + "bbox": [ + 297, + 384, + 416, + 394 + ], + "lines": [ + { + "bbox": [ + 297, + 384, + 416, + 394 + ], + "spans": [ + { + "bbox": [ + 297, + 384, + 416, + 394 + ], + "type": "text", + "content": "(k) Unconfident Test Set (Combined)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 289, + 541, + 381 + ], + "lines": [ + { + "bbox": [ + 423, + 289, + 541, + 381 + ], + "spans": [ + { + "bbox": [ + 423, + 289, + 541, + 381 + ], + "type": "image", + "image_path": "04e09062dca53bb2e4e90640ff2bf16401b5fa436e2d21c599e71052569f5f6b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 384, + 542, + 394 + ], + "lines": [ + { + "bbox": [ + 422, + 384, + 542, + 394 + ], + "spans": [ + { + "bbox": [ + 422, + 384, + 542, + 394 + ], + "type": "text", + "content": "(1) Misclassified Test Set (Combined)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 528, + 184, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 184, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 184, + 540 + ], + "type": "text", + "content": "examples to emphasize this point." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 550, + 208, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 208, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 208, + 562 + ], + "type": "text", + "content": "7. Other Generators Evaluated" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "Our classifiers do not see pixels, but derived geometric features. This means that one could expect a form of generalization across generators. We illustrate that this generalization occurs - ROC curves in Fig. 7 demonstrate that classifiers trained to distinguish Kandinsky images from real images can also reliably distinguish Stable Diffusion XL [3], DeepFloyd [12] and PixArt-" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " [9] from the open-source domain. Additionally, we assess the efficacy of our models against images from proprietary generators such as OpenAI's Dalle-3[4] and Adobe's Firefly [1], representing some of the most advanced tools in image generation. Finally, we show we can detect composite made by a recent SOTA method [29]" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 528, + 520, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 520, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 520, + 540 + ], + "type": "text", + "content": "by looking at Object-Shadow cues in the supplement." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 550, + 374, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 374, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 374, + 562 + ], + "type": "text", + "content": "8. Discussion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "We have shown that generated images can be reliably distinguished from real images by looking only at derived geometric cues. This is likely because image generators do not fully implement the geometry one observes in real images. Producing accurate perspective geometry or accurate shadow geometry requires very tight coordination of detailed information over very long spatial scales. Our results, together with the notorious tendency of face image generators to award subjects' left and right earlobes of different shapes, suggest that doing so is beyond the capacity of current generators. Our findings have significant implications for the development of image generation models, as the inability to" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28145" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 120, + 280 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 120, + 280 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 120, + 280 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 120, + 280 + ], + "type": "image", + "image_path": "1f03b11fe97913453204c040f1c223e16097ce8acd125ea5e11147387c9de8fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 281, + 114, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 281, + 114, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 281, + 114, + 289 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 72, + 190, + 280 + ], + "blocks": [ + { + "bbox": [ + 121, + 72, + 190, + 280 + ], + "lines": [ + { + "bbox": [ + 121, + 72, + 190, + 280 + ], + "spans": [ + { + "bbox": [ + 121, + 72, + 190, + 280 + ], + "type": "image", + "image_path": "755700a65a7865b9578bbff74918413aadbb17b88241014f0a13be9f28b4f13e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 281, + 180, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 281, + 180, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 180, + 289 + ], + "type": "text", + "content": "Shadow Errors" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 191, + 72, + 261, + 281 + ], + "blocks": [ + { + "bbox": [ + 191, + 72, + 261, + 281 + ], + "lines": [ + { + "bbox": [ + 191, + 72, + 261, + 281 + ], + "spans": [ + { + "bbox": [ + 191, + 72, + 261, + 281 + ], + "type": "image", + "image_path": "d7e5e9612ef2d9b8a94c246a55a418a468e454602e08c2cb1a6709eea17d3694.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 281, + 260, + 289 + ], + "lines": [ + { + "bbox": [ + 192, + 281, + 260, + 289 + ], + "spans": [ + { + "bbox": [ + 192, + 281, + 260, + 289 + ], + "type": "text", + "content": "Object-Shadow (OS)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 262, + 72, + 332, + 281 + ], + "blocks": [ + { + "bbox": [ + 262, + 72, + 332, + 281 + ], + "lines": [ + { + "bbox": [ + 262, + 72, + 332, + 281 + ], + "spans": [ + { + "bbox": [ + 262, + 72, + 332, + 281 + ], + "type": "image", + "image_path": "4e161973637eee040d3749a00fbf692ad2f3ec09b10138c220522f0df99880f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 281, + 318, + 289 + ], + "lines": [ + { + "bbox": [ + 273, + 281, + 318, + 289 + ], + "spans": [ + { + "bbox": [ + 273, + 281, + 318, + 289 + ], + "type": "text", + "content": "OS GradCam" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 72, + 400, + 281 + ], + "blocks": [ + { + "bbox": [ + 332, + 72, + 400, + 281 + ], + "lines": [ + { + "bbox": [ + 332, + 72, + 400, + 281 + ], + "spans": [ + { + "bbox": [ + 332, + 72, + 400, + 281 + ], + "type": "image", + "image_path": "d677452df3a49e08fba7d4dd7b817719efb6ce92e03d25ae259265c8d935c590.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 281, + 383, + 289 + ], + "lines": [ + { + "bbox": [ + 350, + 281, + 383, + 289 + ], + "spans": [ + { + "bbox": [ + 350, + 281, + 383, + 289 + ], + "type": "text", + "content": "VP Errors" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 400, + 72, + 472, + 281 + ], + "blocks": [ + { + "bbox": [ + 400, + 72, + 472, + 281 + ], + "lines": [ + { + "bbox": [ + 400, + 72, + 472, + 281 + ], + "spans": [ + { + "bbox": [ + 400, + 72, + 472, + 281 + ], + "type": "image", + "image_path": "e240728e1277b5454f1d0d465048a064968d64d7695ba8d8564e8cfffcfca7d0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 281, + 469, + 289 + ], + "lines": [ + { + "bbox": [ + 403, + 281, + 469, + 289 + ], + "spans": [ + { + "bbox": [ + 403, + 281, + 469, + 289 + ], + "type": "text", + "content": "Perspective Fields (PF)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 473, + 72, + 542, + 281 + ], + "blocks": [ + { + "bbox": [ + 473, + 72, + 542, + 281 + ], + "lines": [ + { + "bbox": [ + 473, + 72, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 473, + 72, + 542, + 281 + ], + "type": "image", + "image_path": "9a72a1de97e8ce46d5dcf5aaa3e914efbb23a3edf003ed0ad11a3f57aca727cf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 281, + 528, + 289 + ], + "lines": [ + { + "bbox": [ + 485, + 281, + 528, + 289 + ], + "spans": [ + { + "bbox": [ + 485, + 281, + 528, + 289 + ], + "type": "text", + "content": "PF GradCam" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 50, + 384, + 120, + 594 + ], + "blocks": [ + { + "bbox": [ + 46, + 295, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 547, + 373 + ], + "type": "text", + "content": "Figure 4. Grad-CAM applied to our Object-Shadow and Perspective Field classifiers reveals that the high AUCs in Fig. 3 are based on real geometric errors in indoor scenes generated by Kandinsky, Stable Diffusion XL and Dalle-3 shown in each row respectively. The second and fifth columns highlight shadow and vanishing point errors, respectively. The third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) identifies diagnostic areas for synthetic generation, such as inconsistent shadow directions (in all three rows), mismatched shadow lengths (second row). The sixth column shows Perspective Fields [21], and Grad-CAM applied to our Perspective Fields classifier (last column) reveals geometric errors in all three rows, particularly at ceilings and side walls, with noticeable errors also present in window grills in the first and second rows." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 384, + 120, + 594 + ], + "lines": [ + { + "bbox": [ + 50, + 384, + 120, + 594 + ], + "spans": [ + { + "bbox": [ + 50, + 384, + 120, + 594 + ], + "type": "image", + "image_path": "7b29ce764b8a4b55151a3d798bb1be69f2d35e04f66643c2f82e825a3400b105.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 596, + 113, + 605 + ], + "lines": [ + { + "bbox": [ + 58, + 596, + 113, + 605 + ], + "spans": [ + { + "bbox": [ + 58, + 596, + 113, + 605 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 609, + 547, + 676 + ], + "lines": [ + { + "bbox": [ + 46, + 609, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 547, + 676 + ], + "type": "text", + "content": "Figure 5. Grad-CAM results for outdoor scenes generated by Kandinsky, Stable Diffusion XL, and Adobe Firefly, shown in each row respectively. The second column highlights shadow errors, while the third column overlays detected object-shadow pairs [44]. Grad-CAM applied to our Object-Shadow classifier (fourth column) reveals incorrect shadow shapes in the first and second rows, with shadows on the right-side pedestrians pointing in a different direction than those on the left. The fifth column shows vanishing point errors, and the sixth column presents Perspective Fields [21]. Grad-CAM applied to our Perspective Fields classifier (last column) confirms large perspective distortions on building facades and road markings, corroborating the vanishing point errors in the fifth column." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 120, + 384, + 190, + 594 + ], + "blocks": [ + { + "bbox": [ + 120, + 384, + 190, + 594 + ], + "lines": [ + { + "bbox": [ + 120, + 384, + 190, + 594 + ], + "spans": [ + { + "bbox": [ + 120, + 384, + 190, + 594 + ], + "type": "image", + "image_path": "46613e0f5ced8ad5cdbedc76e3e6de7be3eabc6e875f80cfb3b6090bbcf6579d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 596, + 180, + 605 + ], + "lines": [ + { + "bbox": [ + 131, + 596, + 180, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 180, + 605 + ], + "type": "text", + "content": "Shadow Errors" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 191, + 384, + 261, + 594 + ], + "blocks": [ + { + "bbox": [ + 191, + 384, + 261, + 594 + ], + "lines": [ + { + "bbox": [ + 191, + 384, + 261, + 594 + ], + "spans": [ + { + "bbox": [ + 191, + 384, + 261, + 594 + ], + "type": "image", + "image_path": "b0b7d98494813662f22bf3115a91e994a8cf20c7131b4da498c66861f61c54df.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 596, + 260, + 605 + ], + "lines": [ + { + "bbox": [ + 192, + 596, + 260, + 605 + ], + "spans": [ + { + "bbox": [ + 192, + 596, + 260, + 605 + ], + "type": "text", + "content": "Object-Shadow (OS)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 262, + 384, + 332, + 594 + ], + "blocks": [ + { + "bbox": [ + 262, + 384, + 332, + 594 + ], + "lines": [ + { + "bbox": [ + 262, + 384, + 332, + 594 + ], + "spans": [ + { + "bbox": [ + 262, + 384, + 332, + 594 + ], + "type": "image", + "image_path": "8608d1484e24da2cb303234a8ed8563efa991b9221f09dd796661ca2eec3f844.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 596, + 318, + 605 + ], + "lines": [ + { + "bbox": [ + 273, + 596, + 318, + 605 + ], + "spans": [ + { + "bbox": [ + 273, + 596, + 318, + 605 + ], + "type": "text", + "content": "OS GradCam" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 331, + 384, + 400, + 594 + ], + "blocks": [ + { + "bbox": [ + 331, + 384, + 400, + 594 + ], + "lines": [ + { + "bbox": [ + 331, + 384, + 400, + 594 + ], + "spans": [ + { + "bbox": [ + 331, + 384, + 400, + 594 + ], + "type": "image", + "image_path": "eff68bc213fb46bce0f720a6a348c4fd4b5ab6690e3968091978327ca827b305.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 596, + 383, + 605 + ], + "lines": [ + { + "bbox": [ + 350, + 596, + 383, + 605 + ], + "spans": [ + { + "bbox": [ + 350, + 596, + 383, + 605 + ], + "type": "text", + "content": "VP Errors" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 400, + 384, + 472, + 594 + ], + "blocks": [ + { + "bbox": [ + 400, + 384, + 472, + 594 + ], + "lines": [ + { + "bbox": [ + 400, + 384, + 472, + 594 + ], + "spans": [ + { + "bbox": [ + 400, + 384, + 472, + 594 + ], + "type": "image", + "image_path": "bd26a9c9015c1dd42bdb005ca3b7b31fdf66ca3f7d31eac38eed3745704f8b3c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 596, + 466, + 605 + ], + "lines": [ + { + "bbox": [ + 406, + 596, + 466, + 605 + ], + "spans": [ + { + "bbox": [ + 406, + 596, + 466, + 605 + ], + "type": "text", + "content": "Perspective Fields" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 473, + 384, + 542, + 594 + ], + "blocks": [ + { + "bbox": [ + 473, + 384, + 542, + 594 + ], + "lines": [ + { + "bbox": [ + 473, + 384, + 542, + 594 + ], + "spans": [ + { + "bbox": [ + 473, + 384, + 542, + 594 + ], + "type": "image", + "image_path": "cdfb7cd164f2dcdc88ae6a5169c3f92dc9c096b33ec2514a718f3335bca33c0f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 596, + 528, + 605 + ], + "lines": [ + { + "bbox": [ + 484, + 596, + 528, + 605 + ], + "spans": [ + { + "bbox": [ + 484, + 596, + 528, + 605 + ], + "type": "text", + "content": "PF GradCam" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28146" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 120, + 212 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 120, + 212 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 120, + 212 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 120, + 212 + ], + "type": "image", + "image_path": "fbd23d4df6975b08322542fc25e3291fc55374ebf81eb28e7e94978e7f5cbe69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 213, + 114, + 223 + ], + "lines": [ + { + "bbox": [ + 56, + 213, + 114, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 114, + 223 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 72, + 261, + 212 + ], + "blocks": [ + { + "bbox": [ + 121, + 72, + 261, + 212 + ], + "lines": [ + { + "bbox": [ + 121, + 72, + 261, + 212 + ], + "spans": [ + { + "bbox": [ + 121, + 72, + 261, + 212 + ], + "type": "image", + "image_path": "b56195086b42baf9bf96beaa67027d1b66f01aff2c5c7d061b0a03a50f69c20f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 180, + 223 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 180, + 223 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 180, + 223 + ], + "type": "text", + "content": "Shadow Errors" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 262, + 72, + 331, + 213 + ], + "blocks": [ + { + "bbox": [ + 191, + 213, + 260, + 223 + ], + "lines": [ + { + "bbox": [ + 191, + 213, + 260, + 223 + ], + "spans": [ + { + "bbox": [ + 191, + 213, + 260, + 223 + ], + "type": "text", + "content": "Object-Shadow (OS)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 262, + 72, + 331, + 213 + ], + "lines": [ + { + "bbox": [ + 262, + 72, + 331, + 213 + ], + "spans": [ + { + "bbox": [ + 262, + 72, + 331, + 213 + ], + "type": "image", + "image_path": "0ea3fbeca0ff02014050f679365c6846d0740185e234e1fa3f116e12533f60a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 213, + 318, + 223 + ], + "lines": [ + { + "bbox": [ + 273, + 213, + 318, + 223 + ], + "spans": [ + { + "bbox": [ + 273, + 213, + 318, + 223 + ], + "type": "text", + "content": "OS GradCam" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 72, + 400, + 212 + ], + "blocks": [ + { + "bbox": [ + 331, + 72, + 400, + 212 + ], + "lines": [ + { + "bbox": [ + 331, + 72, + 400, + 212 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 400, + 212 + ], + "type": "image", + "image_path": "bfe1453f9333e422e720876b5a23fd3a2a4ad141b9edf68270eec72282f0096c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 213, + 383, + 223 + ], + "lines": [ + { + "bbox": [ + 349, + 213, + 383, + 223 + ], + "spans": [ + { + "bbox": [ + 349, + 213, + 383, + 223 + ], + "type": "text", + "content": "VP Errors" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 400, + 72, + 472, + 212 + ], + "blocks": [ + { + "bbox": [ + 400, + 72, + 472, + 212 + ], + "lines": [ + { + "bbox": [ + 400, + 72, + 472, + 212 + ], + "spans": [ + { + "bbox": [ + 400, + 72, + 472, + 212 + ], + "type": "image", + "image_path": "f41c45cf67d1cce83f52983cd097933a0d53cb1bfb50900d05a756b331bf4c89.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 213, + 469, + 223 + ], + "lines": [ + { + "bbox": [ + 403, + 213, + 469, + 223 + ], + "spans": [ + { + "bbox": [ + 403, + 213, + 469, + 223 + ], + "type": "text", + "content": "Perspective Fields (PF)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 473, + 72, + 541, + 212 + ], + "blocks": [ + { + "bbox": [ + 473, + 72, + 541, + 212 + ], + "lines": [ + { + "bbox": [ + 473, + 72, + 541, + 212 + ], + "spans": [ + { + "bbox": [ + 473, + 72, + 541, + 212 + ], + "type": "image", + "image_path": "2f86dd4917360c31ad438575844df4c2323858ffbceabd4f525aef4b91ed155e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 213, + 529, + 222 + ], + "lines": [ + { + "bbox": [ + 485, + 213, + 529, + 222 + ], + "spans": [ + { + "bbox": [ + 485, + 213, + 529, + 222 + ], + "type": "text", + "content": "PF GradCam" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 304, + 143, + 378 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 545, + 294 + ], + "type": "text", + "content": "Figure 6. Our projective geometry classifiers identify distinct types of problems in generated images. The top row presents an example that was classified as real by the Object-Shadow classifier but correctly identified as generated by the Perspective Fields classifier. While the shadow cast by the person appears realistic, the Perspective Fields Grad-CAM highlights the problematic geometry of the shelf on the top left. In contrast, the bottom row shows an example that was correctly identified as generated by the Object-Shadow classifier but misclassified as real by the Perspective Fields classifier. Although the perspective effects in the image appear plausible, the Grad-CAM weights correctly reveal that the two chairs are casting shadows from different light sources, indicating inconsistency in scene's illumination." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 304, + 143, + 378 + ], + "lines": [ + { + "bbox": [ + 50, + 304, + 143, + 378 + ], + "spans": [ + { + "bbox": [ + 50, + 304, + 143, + 378 + ], + "type": "image", + "image_path": "9f03e24a7fae66711fc8a2eab520ebd124703c5d72c8d377fc380287a85158ad.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 393, + 143, + 467 + ], + "blocks": [ + { + "bbox": [ + 62, + 382, + 130, + 391 + ], + "lines": [ + { + "bbox": [ + 62, + 382, + 130, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 130, + 391 + ], + "type": "text", + "content": "(a) Dall-E 3 (Indoor)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 393, + 143, + 467 + ], + "lines": [ + { + "bbox": [ + 50, + 393, + 143, + 467 + ], + "spans": [ + { + "bbox": [ + 50, + 393, + 143, + 467 + ], + "type": "image", + "image_path": "3953b76abbd53bed9dae5fa5e41e8acdc2e2f691a617452e97f9c47210a6a623.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 470, + 132, + 479 + ], + "lines": [ + { + "bbox": [ + 59, + 470, + 132, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 470, + 132, + 479 + ], + "type": "text", + "content": "(f) Dall-E 3 (Outdoor)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 484, + 545, + 551 + ], + "lines": [ + { + "bbox": [ + 46, + 484, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 545, + 551 + ], + "type": "text", + "content": "Figure 7. Our model trained on Kandinsky-v3 generalizes in detecting projective geometry errors from images generated by various unseen models. We evaluate the model's performance on test sets generated by Dall-E 3 (a), DeepFloyd (b), Adobe Firefly (c), PixArt-α (d), and Stable Diffusion XL or SDXL (e) using the same text prompts from the 'misclassified' Kandinsky-v3 generated test set. The top row shows indoor scenes and the last row shows outdoor scenes. Our model generalizes to all evaluated generators except DeepFloyd. However, DeepFloyd-generated images can be reliably detected when the model is trained on a DeepFloyd-generated training set, but it shows poor generalization capabilities to other generators compared to Kandinsky." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 154, + 304, + 247, + 378 + ], + "blocks": [ + { + "bbox": [ + 154, + 304, + 247, + 378 + ], + "lines": [ + { + "bbox": [ + 154, + 304, + 247, + 378 + ], + "spans": [ + { + "bbox": [ + 154, + 304, + 247, + 378 + ], + "type": "image", + "image_path": "b62b039f15e4ac996405a4c3c00a030b92b852097867c2d2baa0160032439527.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 149, + 392, + 242, + 466 + ], + "blocks": [ + { + "bbox": [ + 162, + 381, + 239, + 391 + ], + "lines": [ + { + "bbox": [ + 162, + 381, + 239, + 391 + ], + "spans": [ + { + "bbox": [ + 162, + 381, + 239, + 391 + ], + "type": "text", + "content": "(b) DeepFloyd (Indoor)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 392, + 242, + 466 + ], + "lines": [ + { + "bbox": [ + 149, + 392, + 242, + 466 + ], + "spans": [ + { + "bbox": [ + 149, + 392, + 242, + 466 + ], + "type": "image", + "image_path": "6596c2f96e1841b52913430e72cb49ec4de3aa4cbe2b153a71b1fabab1b7e7ec.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 470, + 237, + 479 + ], + "lines": [ + { + "bbox": [ + 155, + 470, + 237, + 479 + ], + "spans": [ + { + "bbox": [ + 155, + 470, + 237, + 479 + ], + "type": "text", + "content": "(g) DeepFloyd (Outdoor)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 258, + 304, + 352, + 378 + ], + "blocks": [ + { + "bbox": [ + 258, + 304, + 352, + 378 + ], + "lines": [ + { + "bbox": [ + 258, + 304, + 352, + 378 + ], + "spans": [ + { + "bbox": [ + 258, + 304, + 352, + 378 + ], + "type": "image", + "image_path": "b1241234b0e8d266d2a52ac7be29c687173b7e26af57407cf3501a32f750cd13.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 249, + 392, + 343, + 467 + ], + "blocks": [ + { + "bbox": [ + 263, + 382, + 348, + 391 + ], + "lines": [ + { + "bbox": [ + 263, + 382, + 348, + 391 + ], + "spans": [ + { + "bbox": [ + 263, + 382, + 348, + 391 + ], + "type": "text", + "content": "(c) Adobe Firefly (Indoor)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 392, + 343, + 467 + ], + "lines": [ + { + "bbox": [ + 249, + 392, + 343, + 467 + ], + "spans": [ + { + "bbox": [ + 249, + 392, + 343, + 467 + ], + "type": "image", + "image_path": "7739ab8650e5921e348f21eb161b8dce924dfd70c5ac2b7f66e17d4e7a0e8ab2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 470, + 341, + 479 + ], + "lines": [ + { + "bbox": [ + 250, + 470, + 341, + 479 + ], + "spans": [ + { + "bbox": [ + 250, + 470, + 341, + 479 + ], + "type": "text", + "content": "(h) Adobe Firefly (Outdoor)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 354, + 304, + 448, + 378 + ], + "blocks": [ + { + "bbox": [ + 354, + 304, + 448, + 378 + ], + "lines": [ + { + "bbox": [ + 354, + 304, + 448, + 378 + ], + "spans": [ + { + "bbox": [ + 354, + 304, + 448, + 378 + ], + "type": "image", + "image_path": "0a1637bf9feef300a5872265c306f1daa0ba5b2ff248361b28234e3ccf047723.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 350, + 392, + 443, + 467 + ], + "blocks": [ + { + "bbox": [ + 365, + 382, + 436, + 391 + ], + "lines": [ + { + "bbox": [ + 365, + 382, + 436, + 391 + ], + "spans": [ + { + "bbox": [ + 365, + 382, + 436, + 391 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 365, + 382, + 436, + 391 + ], + "type": "inline_equation", + "content": "\\mathrm{PixArt - }\\alpha" + }, + { + "bbox": [ + 365, + 382, + 436, + 391 + ], + "type": "text", + "content": " (Indoor)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 392, + 443, + 467 + ], + "lines": [ + { + "bbox": [ + 350, + 392, + 443, + 467 + ], + "spans": [ + { + "bbox": [ + 350, + 392, + 443, + 467 + ], + "type": "image", + "image_path": "cfd3c4558800f71d84e2bdf0d815ae8181b8f7664870c6695cf6ae52030bc9d9.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 470, + 432, + 479 + ], + "lines": [ + { + "bbox": [ + 359, + 470, + 432, + 479 + ], + "spans": [ + { + "bbox": [ + 359, + 470, + 432, + 479 + ], + "type": "text", + "content": "(i)PixArt- " + }, + { + "bbox": [ + 359, + 470, + 432, + 479 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 359, + 470, + 432, + 479 + ], + "type": "text", + "content": " (Outdoor)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 451, + 304, + 544, + 378 + ], + "blocks": [ + { + "bbox": [ + 451, + 304, + 544, + 378 + ], + "lines": [ + { + "bbox": [ + 451, + 304, + 544, + 378 + ], + "spans": [ + { + "bbox": [ + 451, + 304, + 544, + 378 + ], + "type": "image", + "image_path": "b2051042f8a4d44a2fe8c89f549da0c1d2de0a1038ed8e6563bd5e60acdf4370.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 449, + 392, + 542, + 467 + ], + "blocks": [ + { + "bbox": [ + 466, + 382, + 528, + 391 + ], + "lines": [ + { + "bbox": [ + 466, + 382, + 528, + 391 + ], + "spans": [ + { + "bbox": [ + 466, + 382, + 528, + 391 + ], + "type": "text", + "content": "(e) SDXL (Indoor)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 392, + 542, + 467 + ], + "lines": [ + { + "bbox": [ + 449, + 392, + 542, + 467 + ], + "spans": [ + { + "bbox": [ + 449, + 392, + 542, + 467 + ], + "type": "image", + "image_path": "4542a2733bf4625fd8ccaf3672ff13fdfc3fd03c6c77c5bda9db5a1a888aaba0.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 470, + 528, + 479 + ], + "lines": [ + { + "bbox": [ + 462, + 470, + 528, + 479 + ], + "spans": [ + { + "bbox": [ + 462, + 470, + 528, + 479 + ], + "type": "text", + "content": "(j) SDXL (Outdoor)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 563, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 601 + ], + "type": "text", + "content": "accurately replicate projective geometry extends across various state-of-the-art models, indicating a widespread issue rather than a problem specific to a particular generator." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "We speculate that fixing this difficulty requires structural innovation in the generator, rather than simply exposing the generator to more data. The complex interplay of light, shadows, and perspective in real-world scenes may necessitate novel approaches to modeling and encoding spatial relationships within the generator's architecture. Potential avenues for improvement could include incorporating explicit geometric reasoning or developing new loss functions that prioritize the consistency of projective geometry. Fur" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": "thermore, our work highlights the importance of developing robust evaluation metrics for image generation models that assess the geometric coherence of generated images. By tackling these challenges, we could create image-generation models that more faithfully capture the complex geometric relationships in real-world scenes." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 649, + 399, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 399, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 399, + 662 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 669, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 712 + ], + "type": "text", + "content": "This paper is based on work supported in part by the National Science Foundation under Grant No. 2106825 and a gift from Boeing Corporation." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28147" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Adobe. Firefly. https://www.adobe.com/sensei/generative-ai/firefly.html, 2023. Accessed: 2023-11. 2,6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Shruti Agarwal and Hany Farid. Photo forensics fromJPEG dimples. In 2017 IEEE workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "text", + "content": "[3] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report, 2023. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 281 + ], + "type": "text", + "content": "[5] Anand Bhattachad and D.A. Forsyth. Stylitgan: Prompting stylegan to generate new illumination conditions. In arXiv, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 327 + ], + "type": "text", + "content": "[6] Anand Bhattachad, Daniel McKee, Derek Hoiem, and DA Forsyth. Stylegan knows normal, depth, albedo, and more. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 330, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 288, + 373 + ], + "type": "text", + "content": "[7] Matyáš Boháček and Hany Farid. A geometric and photometric exploration of gan and diffusion synthesized faces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-883, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 375, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 407 + ], + "type": "text", + "content": "[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 463 + ], + "type": "text", + "content": "[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "text", + "content": "[10] Yida Chen, Fernanda Viégas, and Martin Wattenberg. Beyond surface statistics: Scene representations in a latent diffusion model. arXiv preprint arXiv:2306.05720, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 553 + ], + "type": "text", + "content": "[11] Jaemin Cho, Yushi Hu, Roopal Garg, Peter Anderson, Ranjay Krishna, Jason Baldridge, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-image generation. arXiv preprint arXiv:2310.18235, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 556, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 578 + ], + "type": "text", + "content": "[12] Deep Floyd. Iterative filter. https://github.com/deep-floyd/IF, 2023. Accessed: 2023-11. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 622 + ], + "type": "text", + "content": "[13] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "text", + "content": "[14] David C. Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 382-392, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[15] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 105 + ], + "type": "text", + "content": "[16] David A Forsyth and Jean Ponce. Computer vision: a modern approach. prentice hall professional technical reference, 2002. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 183 + ], + "type": "text", + "content": "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 239 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 242, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 547, + 274 + ], + "type": "text", + "content": "[19] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In ECCV, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 331 + ], + "type": "text", + "content": "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 332, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 547, + 397 + ], + "type": "text", + "content": "[21] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 2, 5, 7, 16, 17" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 443 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 547, + 498 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 501, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 547, + 544 + ], + "type": "text", + "content": "[24] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 546, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 546, + 578 + ], + "type": "text", + "content": "[25] Eric Kee, James F O'Brien, and Hany Farid. Exposing photo manipulation with inconsistent shadows. ACM Transactions on Graphics (ToG), 32(3):1-12, 2013. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 580, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 546, + 624 + ], + "type": "text", + "content": "[26] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 713 + ], + "type": "text", + "content": "[27] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Benita Teufel, Marco Bellagente, Minguk Kang, Taesung Park, Jure Leskovec, Jun-Yan Zhu, Li Fei-Fei, Jiajun Wu, Stefano Ermon, and Percy Liang. Holistic evaluation of text-to-image models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28148" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "text", + "content": "[29] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "text", + "content": "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 5000–5009, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 275 + ], + "type": "text", + "content": "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "text", + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 365 + ], + "type": "text", + "content": "[33] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 53(2):758-767, 2005. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 366, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 288, + 421 + ], + "type": "text", + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 533 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 535, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 590 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "type": "text", + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[39] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 576 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[40] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 128(2):336-359, 2019. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[41] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18310-18319, 2023. 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 546, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 218 + ], + "type": "text", + "content": "[42] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In ICCV, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 546, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 252 + ], + "type": "text", + "content": "[43] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 253, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 547, + 297 + ], + "type": "text", + "content": "[44] Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3259-3273, 2022. 5, 7, 16, 17" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 546, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 351 + ], + "type": "text", + "content": "[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22445-22455, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 353, + 546, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 546, + 397 + ], + "type": "text", + "content": "[46] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 398, + 547, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 547, + 463 + ], + "type": "text", + "content": "[47] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 464, + 546, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 546, + 509 + ], + "type": "text", + "content": "[48] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 510, + 546, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 546, + 542 + ], + "type": "text", + "content": "[49] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 543, + 546, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 546, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 546, + 576 + ], + "type": "text", + "content": "[50] Peng Zhou, Xintong Han, Vlad I Morariu, and Larry S Davis. Learning rich features for image manipulation detection. In CVPR, 2018. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28149" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_content_list.json b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..404ba110b86f900b2d1bc1d437f1b9cb9b75839c --- /dev/null +++ b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_content_list.json @@ -0,0 +1,1811 @@ +[ + { + "type": "text", + "text": "Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification", + "text_level": 1, + "bbox": [ + 120, + 128, + 848, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Yang Jun Chen* Mang Ye*", + "bbox": [ + 351, + 203, + 625, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Engineering Research Center for Multimedia Software,", + "bbox": [ + 228, + 222, + 738, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science, Hubei Luojia Laboratory, Wuhan University, Wuhan, China", + "bbox": [ + 133, + 239, + 836, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/yangbincv/SDCL", + "bbox": [ + 330, + 258, + 625, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised visible-infrared person re-identification (US-VI-ReID) centers on learning a cross-modality retrieval model without labels, reducing the reliance on expensive cross-modality manual annotation. Previous US-VI-ReID works gravitate toward learning cross-modality information with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. The shallow feature from the shallow layers contains nuanced detail information, which is critical for effective cross-modality learning but is disregarded regrettably by the existing methods. To address the above issues, we design a Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration which are harnessed for neighbor alignment, enhancing the robustness in a dynamic manner. Furthermore, CRA associates the cross-modality labels with the ranking association between shallow and deep features, furnishing valuable supervision for cross-modality learning. Extensive experiments validate the superiority of our method, even outperforming certain supervised counterparts.", + "bbox": [ + 75, + 340, + 472, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 207, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) aims at matching the same person image captured by non-overlapping cameras. In recent times, ReID has garnered significant attention from the computer vision research community, owing to its pivotal role in the context of intelligent video surveillance applications [2, 18, 19, 23, 25, 28-30, 39, 41, 44, 46, 57,", + "bbox": [ + 75, + 786, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8709010003d7ba753fffa03e46b6e56d6c0c2fa19cc2c7548567d31034bf2688.jpg", + "image_caption": [ + "Ground Truth Cross-modality Label Pairs", + "Deep Features", + "Figure 1. Illustration of the motivation. Previous advanced works for US-VI-ReID mainly focus on learning features and associating cross-modality labels with deep features. However, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in (a). In this paper, we incorporate shallow and deep features with collaborative learning and label association to remedy these problems, as shown in (b)." + ], + "image_footnote": [], + "bbox": [ + 504, + 308, + 696, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3569712b422bf865fb2c61097d73afc29bde2ba49f540cd7f4911a0f32462460.jpg", + "image_caption": [ + "Shallow Features", + "Negative Optimization" + ], + "image_footnote": [], + "bbox": [ + 696, + 310, + 890, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "59, 62, 63]. The rank-1 accuracy has exhibited encouraging outcomes in supervised person ReID. Notably, these person images are typically acquired by visible cameras within well-illuminated environments. However, it is crucial to recognize that visible cameras often fail to capture sufficient information of individuals in low-light settings, consequently constraining the practicality of single-modality ReID in the context of night-time surveillance [1, 7, 15, 21, 40, 42, 45, 53, 61].", + "bbox": [ + 496, + 611, + 890, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, recently deployed cameras possess the capability to seamlessly transition into far/near-infrared mode during nighttime operations [50, 55]. Various techniques [12, 20, 22, 27, 34, 47, 52, 54-58, 62] have been proposed for visible-infrared person re-identification (VI-ReID), resulting in commendable accuracy levels. Notably, these achievements are facilitated by extensive human-labeled cross-modality datasets. However, annotating cross-modality datasets demands more resources compared to the annotation of single-modality ReID datasets. This", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 94, + 886, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16870", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intricate annotation task presents a formidable and pivotal challenge, ultimately giving rise to the crucial task of unsupervised VI-ReID (US-VI-ReID) for reducing the cost of expensive cross-modality annotations.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some prior research endeavors [4, 5, 24, 33, 37, 48, 50, 51] have proposed preliminary solutions for US-VI-ReID. These solutions gravitate toward learning cross-modality feature with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, i.e., intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in Fig. 1. It is important to recognize that the shallow features originating from the modality-specific shallow layers harbor a wealth of nuanced detail information pertaining to pedestrian attributes, which is critical for unsupervised cross-modality representation learning but is disregarded regrettably by the existing methods. To harness the complete potential of the shallow information, we devise a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration as the constraints for seeking reliable intramodality and inter-modality neighbor learning, thereby guaranteeing the cultivation of modality-invariant and discriminative representations in a dynamic manner. Furthermore, CRA is developed with a global perspective to exploit cross-modality ranking consistency between deep and shallow features, associating the label of two modalities in a collaborative manner and furnishing valuable cross-modality supervision for cross-modality learning. Ultimately, within the shallow-deep collaborative learning framework with CNL and CRA modules, we acquire a robust representation, effectively mitigating the cross-modality discrepancy under unsupervised conditions.", + "bbox": [ + 75, + 154, + 472, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions can be summarized as follows:", + "bbox": [ + 96, + 670, + 457, + 685 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a shallow-deep collaborative learning framework based on the transformer architecture. This framework facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow and deep features.", + "- We propose a collaborative neighbor learning module to formulate dependable intra-modality and cross-modality neighbor learning, enabling the model to capture modality-invariant and discriminative features.", + "- We propose a collaborative ranking association module to explore intra-modality and cross-modality ranking consistencies, unifying the cross-modality labels and providing invaluable cross-modality supervision.", + "- Extensive experiments validate that our SDCL frame" + ], + "bbox": [ + 76, + 689, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work surpasses existing methods on two mainstream VI-ReID benchmarks, consistently improving the unsupervised cross-modality retrieval performance.", + "bbox": [ + 511, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 159, + 640, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Supervised Visible-Infrared Person ReID", + "text_level": 1, + "bbox": [ + 500, + 186, + 849, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, advanced supervised VI-ReID focuses on bridging the gap between the two modalities and learning robust representations against modality discrepancy. Ye et al. [55] developed a dynamic tri-level relation mining (DTRM) framework simultaneously to explore the cross-modality relation cues of channel-level, part-level intra-modality, and graph-level. CAJ [56] proposed a channel-mixed learning to handle the intra-modality and cross-modality variations by randomly exchanging the color channels. SGIEL [11] proposed a shape-erased feature learning paradigm, jointly learning shape-related feature in one subspace and shape-erased features in the orthogonal complement.", + "bbox": [ + 496, + 214, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, the remarkable performances exhibited by these methods require extensive human-labeled cross-modality datasets. In this work, we pivot our focus to the realm of unsupervised visible-infrared person ReID, where the luxury of identity annotations is absent, presenting important applications for real-world VI-ReID deployments.", + "bbox": [ + 496, + 398, + 890, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unsupervised Single-Modality Person ReID", + "text_level": 1, + "bbox": [ + 500, + 508, + 870, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised single-modality ReID endeavors to train the ReID model using unlabeled data captured by visible cameras. The majority of studies have embraced cluster algorithms to derive pseudo labels for optimizing the model. Cluster Contrast [9] utilized a distinctive cluster representation to delineate each cluster, addressing the issue of cluster inconsistency. Chen et al. [3] introduced Inter-instance Contrastive Encoding (ICE), which harnesses inter-instance pairwise similarity scores to enhance preceding class-level contrastive ReID methodologies. IICS [49] tackled the unsupervised ReID challenge by decomposing the similarity computation into two stages, namely, the intra-domain and inter-domain computations, respectively. Dai et al. [8] introduced a dual-refinement approach to concurrently enhance pseudo labels during the offline clustering phase and refine features during the online training phase, augmenting label purity and feature discriminability.", + "bbox": [ + 496, + 534, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above methods are dedicated to addressing issues in single-modality ReID. When employed in the context of US-VI-ReID, these approaches encounter challenges stemming from cross-modality discrepancies and the absence of cross-modality (visible-infrared) identity labels. This impedes the learning of inter-modality feature and the generation of reliable cross-modality pseudo labels.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Unsupervised Visible-Infrared Person ReID", + "text_level": 1, + "bbox": [ + 76, + 90, + 447, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several studies [24, 37, 48, 50] represent initial endeavors in the realm of US-VI-ReID. Yang et al. [50] introduced a novel Augmented Dual-Contrastive Aggregation (ADCA) learning framework founded on the principles of homogeneous joint learning and heterogeneous aggregation, establishing a robust baseline for purely unsupervised VI-ReID. PGM [48] developed a progressive graph matching method to systematically extract cross-modality correspondences, formulating correspondence mining as a graph-matching process. H2H [24] introduced a homogeneous-to-heterogeneous approach through two-stage learning, incorporating Market1501 [64] as additional labeled data. CHCR [33] proposed a cross-modality hierarchical clustering and refinement method by promoting modality-invariant feature learning and improving the reliability of pseudo-labels.", + "bbox": [ + 75, + 113, + 468, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, the above methods primarily concentrate on feature learning with deep features. However, interfered by the intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. Our approach integrates shallow and deep features through collaborative feature learning and cross-modality label association, which is distinguished from previous works, surpassing the performance of existing US-VI-ReID methods.", + "bbox": [ + 75, + 340, + 470, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 507, + 250, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The framework of shallow-deep collaborative learning (SDCL) is shown in Fig. 2, incorporating the Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. SDCL adopts a dual-path transformer architecture with dual-contrastive learning [50], which has two shallow modality-specific patch embedding layers and a modality-shared transformer. Instance memory and cluster memory are constructed for shallow embeddings and deep features within each modality. Instance memory stores all training image features. Cluster memory is built by averaging the instance features with the same pseudo labels. With the above memories, SDCL develops the CNL module to exploit homogeneous and heterogeneous collaboration as the constraints for seeking reliable intra-modality and inter-modality neighbor learning. CRA explores the ranking consistency of cross-modality shallow embeddings and deep features to associate reliable cross-modality identities. With the collaboration of shallow embeddings and deep features within CNL and CRA modules, SDCL learns better representation for cross-modality retrieval.", + "bbox": [ + 75, + 532, + 468, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Shallow-deep Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 845, + 385, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce shallow-deep contrastive learning based on augmented dual-contrast learning [50] with a dual-path", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transformer architecture.", + "bbox": [ + 498, + 92, + 665, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cluster Memory Initialization. At the beginning of each training epoch, we construct shallow and deep cluster memories for each modality by averaging the shallow and deep feature of one cluster, which can be denoted as:", + "bbox": [ + 496, + 107, + 892, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {k} ^ {v s} = \\frac {1}{| \\mathcal {H} _ {k} ^ {v} |} \\sum_ {u _ {n} ^ {v s} \\in y _ {k} ^ {v}} u _ {n} ^ {v s}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 180, + 890, + 217 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {k} ^ {v d} = \\frac {1}{\\left| \\mathcal {H} _ {k} ^ {v} \\right|} \\sum_ {u _ {n} ^ {v d} \\in y _ {k} ^ {v}} u _ {n} ^ {v d}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 236, + 890, + 275 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {l} ^ {r s} = \\frac {1}{| \\mathcal {H} _ {l} ^ {r} |} \\sum_ {u _ {n} ^ {r s} \\in y _ {l} ^ {r}} u _ {n} ^ {r s}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 290, + 890, + 328 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {l} ^ {r d} = \\frac {1}{| \\mathcal {H} _ {l} ^ {r} |} \\sum_ {u _ {n} ^ {r d} \\in y _ {l} ^ {r}} u _ {n} ^ {r d}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 343, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $u_{n}^{r*}$ and $u_{n}^{v*}$ are infrared and visible instance features, respectively. $u_{n}^{*s}$ and $u_{n}^{*d}$ are shallow and deep instance features, respectively. $\\mathcal{H}_k$ indicates the $k$ -th cluster set and $|\\cdot|$ counts the number of samples of a set. Meanwhile, we store the instance features in shallow and deep infrared and visible instance memories $U^{rs}$ , $U^{rd}$ , $U^{vs}$ and $U^{vd}$ .", + "bbox": [ + 496, + 391, + 890, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cluster Memory Updating. During each training iteration, the deep and shallow cluster memories of two modalities by a momentum updating strategy:", + "bbox": [ + 496, + 483, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {k} ^ {(\\delta)} \\leftarrow \\alpha \\phi_ {k} ^ {(\\delta - 1)} + (1 - \\alpha) q, q \\in y _ {k}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 540, + 890, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q$ is the query features sampled from training set in each training iteration, respectively. $\\alpha$ is the momentum factor. $\\delta$ is the training iteration number.", + "bbox": [ + 496, + 575, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive Loss. Given shallow and deep visible and infrared query $q^{vs}$ , $q^{vd}$ , $q^{rs}$ , and $q^{rs}$ , we compute the contrastive loss for visible modality by the following equations:", + "bbox": [ + 496, + 621, + 890, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i d} ^ {v s} = - \\log \\frac {\\exp \\left(q ^ {v s} \\cdot \\phi_ {+} ^ {v s} / \\tau\\right)}{\\sum_ {k = 0} ^ {K} \\exp \\left(q ^ {v s} \\cdot \\phi_ {k} ^ {v s} / \\tau\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 681, + 890, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i d} ^ {v d} = - \\log \\frac {\\exp \\left(q ^ {v d} \\cdot \\phi_ {+} ^ {v d} / \\tau\\right)}{\\sum_ {k = 0} ^ {K} \\exp \\left(q ^ {v d} \\cdot \\phi_ {k} ^ {v d} / \\tau\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 731, + 890, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi_{+}$ is the positive memory corresponding to the pseudo label of $q$ and the $\\tau$ is a temperature. The infrared contrastive loss $\\mathcal{L}_{id}^{rs}$ and $\\mathcal{L}_{id}^{rd}$ are obtained by the same way.", + "bbox": [ + 496, + 777, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We optimize the modality-specific shallow embedding layers and deep modality-shared layers by combining deep and shallow contrastive loss:", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i d} = \\mathcal {L} _ {i d} ^ {v d} + \\mathcal {L} _ {i d} ^ {r d} + \\mathcal {L} _ {i d} ^ {v s} + \\mathcal {L} _ {i d} ^ {r s} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16872", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef5faaca97f99d3d306196012079f3290a78ebf352900d4c64a9899b5c521150.jpg", + "image_caption": [ + "Figure 2. Illustration of shallow-deep collaborative learning. It comprises shallow-deep contrastive learning, collaborative neighbor learning, and collaborative ranking association. SDCL finds reliable neighborhood and cross-modality labels with shallow-deep collaborations, enhancing the robustness of learned representation." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Collaborative Neighbor Learning", + "text_level": 1, + "bbox": [ + 76, + 534, + 372, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our basic rationale for the CNL module is that deep and shallow features are complementary, and their intrinsic consistency relationship can be used to constrain intra-modality and inter-modality optimization, formulating collaborative learning to enhance the robustness against intra-modality and inter-modality variations. From the above perspective, CNL explores essential homogeneous and heterogeneous shallow-deep collaborations to seek reliable intra-modality and inter-modality neighbor learning.", + "bbox": [ + 75, + 558, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Homogeneous and Heterogeneous Collaboration. Given a query $q$ , we can get the similarity $s(q_i, u_j)$ between the query and each instance in the training set by:", + "bbox": [ + 75, + 695, + 470, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(q _ {i}, u _ {j}\\right) = \\frac {q _ {i} \\cdot u _ {j}}{\\left| \\left| q _ {i} \\right| \\right| _ {2} \\left| \\left| u _ {j} \\right| \\right| _ {2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 751, + 468, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{i}$ and $u_{j}$ come from shallow or deep features of visible or infrared modality to get multiple types of homogeneous and heterogeneous shallow or deep query-instance similarity $s(q_{i}^{vs},u_{j}^{vs})$ , $s(q_{i}^{rs},u_{j}^{rs})$ , $s(q_{i}^{vs},u_{j}^{rs})$ , $s(q_{i}^{rs},u_{j}^{vs})$ , $s(q_{i}^{vd},u_{j}^{vd})$ , $s(q_{i}^{rd},u_{j}^{rd})$ , $s(q_{i}^{vd},u_{j}^{rd})$ , and $s(q_{i}^{rd},u_{j}^{vd})$ .", + "bbox": [ + 76, + 792, + 468, + 872 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the above multiple types of similarities, we seek reliable shallow and deep intra-modality and inter-modality", + "bbox": [ + 76, + 871, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "neighbors, formulating collaborative neighbor learning. The intra-modality neighbors can be defined as follows:", + "bbox": [ + 498, + 535, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right) = \\left\\{\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v s}\\right) \\cap \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v d}\\right) \\right\\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 575, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {r}\\right) = \\left\\{\\mathcal {N} ^ {r} \\left(q _ {i} ^ {r s}\\right) \\cap \\mathcal {N} ^ {r} \\left(q _ {i} ^ {r d}\\right) \\right\\}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 619, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}^v(q_i^{v*})$ and $\\mathcal{N}^r(q_i^{r*})$ are neighborhood sets searched by visible-visible and infrared-infrared similarity, respectively. $* \\in \\{s, d\\}$ represents the shallow and deep feature. The $\\mathcal{N}^v(q_i^{vs})$ and $\\mathcal{N}^r(q_i^{vd})$ can be obtained by:", + "bbox": [ + 498, + 643, + 893, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v s}\\right) = \\left\\{u _ {j} ^ {v s} \\mid s \\left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {v}} s \\left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\\right) \\right\\}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 713, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v d}\\right) = \\left\\{u _ {j} ^ {v d} \\mid s \\left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N v} s \\left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\\right) \\right\\}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 763, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $N^v$ denotes the number of visible instances and the $\\gamma$ is the positive neighbor selection range. The $\\mathcal{N}^r (q_i^{rs})$ and $\\mathcal{N}^r (q_i^{rd})$ can be calculated by similar manner.", + "bbox": [ + 496, + 796, + 890, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given query $q_{i}^{v}$ or $q_{i}^{r}$ , the inter-modality neighbors can be denoted as follows:", + "bbox": [ + 496, + 842, + 890, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v}\\right) = \\left\\{\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v s}\\right) \\cap \\mathcal {N} ^ {r} \\left(q _ {i} ^ {v d}\\right) \\right\\}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {r}\\right) = \\left\\{\\mathcal {N} ^ {v} \\left(q _ {i} ^ {r s}\\right) \\cap \\mathcal {N} ^ {v} \\left(q _ {i} ^ {r d}\\right) \\right\\}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 104, + 468, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}^r (q_i^{vs})$ and $\\mathcal{N}^r (q_i^{vd})$ are defined as follows:", + "bbox": [ + 76, + 128, + 424, + 147 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v s}\\right) = \\left\\{u _ {j} ^ {r s} \\mid s \\left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {r}} s \\left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\\right) \\right\\}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 155, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v d}\\right) = \\left\\{u _ {j} ^ {r d} \\mid s \\left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {r}} s \\left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\\right) \\right\\}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 207, + 468, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $N^r$ denotes the number of infrared instances. Similarly, we can get the $\\mathcal{N}^v(q_i^{rs})$ and $\\mathcal{N}^v(q_i^{rd})$ . With the above procedure, we get the $\\mathcal{N}^v(q_i^v), \\mathcal{N}^r(q_i^r), \\mathcal{N}^r(q_i^v)$ and $\\mathcal{N}^v(q_i^r)$ , which are the reliable neighborhood set searched by the constraints of homogeneous or heterogeneous shallow-deep collaborations.", + "bbox": [ + 76, + 239, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collaborative Neighbor Learning. With the neighborhood set $\\mathcal{N}^v(q_i^v)$ , $\\mathcal{N}^r(q_i^r)$ , $\\mathcal{N}^r(q_i^v)$ and $\\mathcal{N}^v(q_i^r)$ , we can perform neighbor learning. Given query $q_i^{vs}$ and $q_i^{vs}$ , we can obtain the expression of the visible-visible shallow and deep neighbor learning by:", + "bbox": [ + 76, + 330, + 468, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} ^ {v v} = - \\frac {1}{N ^ {b}} \\sum_ {i = 1} ^ {N ^ {b}} \\sum_ {j \\in \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right)} \\log \\frac {\\exp \\left(s \\left(q _ {i} ^ {v s} , u _ {j} ^ {v s}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N ^ {v}} \\exp \\left(s \\left(q _ {i} ^ {v s} , u _ {n} ^ {v s}\\right) / \\tau\\right)}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 417, + 468, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} ^ {v v} = - \\frac {1}{N ^ {b}} \\sum_ {i = 1} ^ {N ^ {b}} \\sum_ {j \\in \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right)} \\log \\frac {\\exp \\left(s \\left(q _ {i} ^ {v d} , u _ {j} ^ {v d}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N ^ {v}} \\exp \\left(s \\left(q _ {i} ^ {v d} , u _ {n} ^ {v d}\\right) / \\tau\\right)}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 496, + 468, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {v v} = \\mathcal {L} _ {d} ^ {v v} + \\lambda_ {s} \\mathcal {L} _ {s} ^ {v v}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 565, + 468, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N^b$ is the batch size of query $q_{i}$ . Similarly, the neighbor learning for infrared-infrared $\\mathcal{L}^{rr}$ , infrared-visible $\\mathcal{L}^{rv}$ and visible-infrared $\\mathcal{L}^{vr}$ can be obtained by similar ways. The final optimization for neighbor learning is denoted by the following combination:", + "bbox": [ + 76, + 589, + 468, + 665 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n e i g h b o r}} = \\mathcal {L} ^ {v v} + \\mathcal {L} ^ {r r} + \\mathcal {L} ^ {r v} + \\mathcal {L} ^ {v r}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 678, + 468, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss for SDCL is denoted as:", + "bbox": [ + 96, + 707, + 366, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {n e i g h b o r}} + \\mathcal {L} _ {i d} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 736, + 468, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. In contrast to neighborhood learning in [30], which exclusively relies on deep features for the selection of positive neighbors. Under the multiple discrepancies, the similarity of individual deep features is unreliable, resulting in wrong neighbor selection and negative optimization. Our method embraces collaborative neighbor learning, which dynamically refines cross-modality supervision through shallow and deep collaboration, achieving effective cross-modality feature alignment.", + "bbox": [ + 76, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Collaborative Ranking Association", + "text_level": 1, + "bbox": [ + 500, + 90, + 803, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The CRA module is underpinned by dual rationales: (1) Shallow and deep features across the two modalities should have the same ranking consistency. (2) The ranking consistencies of shallow and deep features within one modality are more accurate. Accordingly, the CRA module contains two processes, i.e., inter-modality ranking association and intra-modality ranking smoothing.", + "bbox": [ + 498, + 113, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inter-modality Ranking Association. With the similarities $s(u_{i}^{vs},u_{j}^{rs})$ and $s(u_{i}^{vd},u_{j}^{rd})$ , we can get two visible-infrared ranking lists $\\{s(u_i^{vs},u_j^{rs}),j\\in [1,N^r ]\\}$ and $\\{s(u_i^{vd},u_j^{rd}),j\\in [1,N^r ]\\}$ for $u_{i}^{vs}$ and $u_{i}^{vd}$ , which are denoted as $\\mathcal{R}_s^{vr}(u_i^{vs})$ and $\\mathcal{R}_d^{vr}(u_i^{vd})$ . The label of $k$ -th similar infrared instance in two ranking lists can be represented as:", + "bbox": [ + 498, + 220, + 890, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r s}}, u _ {j} ^ {r s} = \\mathcal {R} _ {s} ^ {v r} \\left(u _ {i} ^ {v s}\\right) [ k ], \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 329, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r d}}, u _ {j} ^ {r d} = \\mathcal {R} _ {d} ^ {v r} \\left(u _ {i} ^ {v d}\\right) [ k ], \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 369, + 890, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{y}_{u_i^{vs}}^{vr}[k]$ and $\\tilde{y}_{u_i^{vd}}^{vr}[k]$ are refined $k$ -th cross-modality labels of $u_i^{vs}$ and $u_i^{vd}$ through the ranking on visible-infrared shallow and deep similarities. Based on the rationale (1), we propose to associate the cross-modality labels with the intersection of two label sets to investigate collaborative ranking consistency:", + "bbox": [ + 498, + 397, + 890, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {u _ {i} ^ {v}} (k) = \\left\\{\\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] \\cap \\left\\{\\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ n ] \\right\\} _ {n = 1} ^ {N} \\right\\}, \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 503, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $I_{u_i^v}(k)$ records the samples with the same identity of top- $k$ identity in instances from top-1 to top- $N$ . The reliable refined cross-modality label of $u_i^v$ can be expressed by the label of maximum count in $I_{u_i^v}(k)$ :", + "bbox": [ + 498, + 537, + 890, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} _ {u _ {i} ^ {v}} ^ {c m} = \\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ], k = \\underset {k} {\\operatorname {a r g m a x}} \\left(\\left\\{\\left| I _ {u _ {i} ^ {v}} (k) \\right|, k \\in [ 1, K ] \\right\\}\\right), \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 607, + 890, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\cdot|$ denotes the counting function. The argmax operation traverses $I_{u_i^v}(k)$ and finds the labels with maximum number as cross-modality refined labels. Then, we convert the label list $\\{\\tilde{y}_{u_i^v}^{cm}, i \\in [1, N^v]\\}$ to the form of one-hot code matrix $\\tilde{Y}^{cm} \\in \\mathbb{R}^{N^v \\times C^r}$ by setting the column according to the refine labels to 1 and the rest to 0, where $C^r$ is the class number of infrared modality.", + "bbox": [ + 498, + 641, + 890, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intra-modality Collaborative Smoothing. Based on the rationale (2), two shallow and deep homogeneous shallow $P_{s}^{vv}$ and deep $P_{d}^{vv}$ similarity matrices are constructed to investigate intra-modality ranking consistency, enhancing the precision of refined cross-modality pseudo labels by:", + "bbox": [ + 498, + 752, + 890, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {s} ^ {v v} (i, j) = s \\left(u _ {i} ^ {v s}, u _ {j} ^ {v s}\\right), \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 839, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {d} ^ {v v} (i, j) = s \\left(u _ {i} ^ {v d}, u _ {j} ^ {v d}\\right), \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16874", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{*}^{vv}\\in \\mathbb{R}^{N^{v}\\times N^{v}}$ represents the intra-modality similarity structure. In order to explore the intra-modality shallow and deep collaboration, we calculate the sum of $P_{s}^{vv}(i,j)$ and $P_d^{vv}(i,j)$ by:", + "bbox": [ + 76, + 89, + 470, + 154 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {v v} (i, j) = P _ {s} ^ {v v} (i, j) + P _ {d} ^ {v v} (i, j) \\tag {29}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 160, + 468, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P^{vv}$ indicates the consistency of shallow and deep similarity matrix. We keep the 5-max values of $P^{vv}$ in each row to 1 and the rest to 0, acquiring the ranking relations. The process of intra-modality ranking smoothing is formulated as follows:", + "bbox": [ + 75, + 186, + 468, + 261 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nY ^ {c m} = P ^ {v v} \\tilde {Y} ^ {c m}, \\tag {30}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 268, + 468, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Y^{cm} \\in \\mathbb{R}^{N^v \\times C^r}$ is the final refined cross-modality label matrix of visible instance. In $Y^{cm}$ , the column number of the maximum value in each row is the refined label of samples. In our work, we refine the pseudo labels of the visible modality to the infrared modality and smooth the cross-modality labels with the shallow and deep similarity within the visible modality. Then, the infrared and refined visible labels will be used to construct a modality-shared memory for contrastive learning within two modalities [51]. Discussion. Our CRA is essentially the collaboration of shallow and deep ranking processes with a global perspective at the beginning of each training epoch, which is distinguished from previous works employing a solitary-pattern deep feature similarity [37, 48, 50, 51]. The imposition of a collaborative consistency constraint on dual-pattern similarities within CRA confers a distinct advantage, facilitating the exploration of more dependable cross-modality relations and providing valuable cross-modality supervision.", + "bbox": [ + 75, + 294, + 468, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocol", + "text_level": 1, + "bbox": [ + 76, + 604, + 369, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed SDCL framework on two widely-used visible-infrared person ReID datasets, i.e., SYSU-MM01 [45] and RegDB [32]. SYSU-MM01 dataset is collected by 6 different cameras, including 22258 visible and 11909 near-infrared images of 395 training identities. We perform ten trials of the gallery set selection [53] and calculate the average performance following existing methods with all-search and indoor-search testing mode. RegDB dataset is captured by two aligned visible and thermal camera system, which has less challenges for VI-ReID. 206 identities with 2,060 images are selected for training and this procedure is repeated 10 times following [54], and calculate the average performance with visible to thermal and thermal to visible testing mode.", + "bbox": [ + 75, + 628, + 467, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocols. Following existing works, Cumulative Matching Characteristics (CMC), mean Average Precision (mAP) and mean inverse negative penalty (mINP) [57] are calculated as the evaluation metrics.", + "bbox": [ + 75, + 839, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 90, + 718, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed framework is implemented with PyTorch. SDCL incorporates the feature extractor from TransReID [17] as the backbone network with augmented dual-contrastive learning [50]. The shallow patch embedding layers are constructed with IBN and CNN module [31]. The concatenation of shallow and deep class tokens is used to calculate the cosine similarity for retrieval. DBSCAN [10] is conducted to generate pseudo labels. The visible and infrared images are resized to $288 \\times 144$ before entering the network. We sample 8 pseudo identities and 16 instances for each pseudo identity for each modality within one batch. Channel augmentation [56], random crop, horizontal flipping, and random erasing are adopted for data augmentation. We adopt the SGD optimizer with the initial learning rate of $3.5e - 4$ to train the model. The model is trained in a total of 50 epochs. The CRA module is added in the last 20 epochs. The $\\lambda_{s}$ in Eq. 20 is set to 0.5. The $\\gamma$ in Eq. 12, Eq. 13, Eq. 16, and Eq. 17 is set to 0.9. The $N$ and $K$ in Eq. 25 and Eq. 26 are set to 20 for SYSU-MM01 and 10 for RegDB, respectively. The contrastive learning settings follow [50].", + "bbox": [ + 496, + 114, + 890, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 498, + 448, + 864, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1, our SDCL is compared with supervised and unsupervised VI-ReID methods on two benchmarks including SYSU-MM01 and RegDB.", + "bbox": [ + 496, + 473, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Unsupervised Methods. As reported in Table 1, our method exhibits superior performance compared to the current advanced unsupervised methods. More precisely, our SDCL attains a remarkable $64.49\\%$ and $86.91\\%$ rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. Compared with the best current method GUR [51], our method also exceeds the about $1\\%$ and $10\\%$ rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. In comparison to DPIS [35], a uni-semi-supervised method employing visible annotations for training, our SDCL still continues to maintain a leading position.", + "bbox": [ + 496, + 520, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Supervised Methods. Our comparison further includes 14 well-known supervised methods for reference. The comparisons with these methodologies unequivocally illustrate that our SDCL framework outperforms several supervised methods, including AGW [57], MSO [12], and DDAG [54], and achieves a competitive performance compared with MCLNet[16]. These considerable gains benefit from the insightful design of a shallow-deep collaborative learning framework. Our method facilitates the cultivation of modality-invariant features through the synergistic collaboration of shallow and deep features. Guided by our innovative solutions, SDCL surpasses prevailing unsupervised methods.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/36531d9d780d9efb107a8aef664608f0ebc5a30b9733f6c42da8511d8f27ef87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsVenueSYSU-MM01RegDB
All SearchIndoor SearchVisible to InfraredInfrared to Visible
r1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
SupervisedDDAG [54]ECCV-2054.7553.0239.6261.0267.9862.6169.3463.4649.2468.0661.8048.62
AGW [57]TPAMI-2147.5047.6535.3054.1762.9759.2370.0566.3750.1970.4965.9051.24
CA [56]ICCV-2169.8866.8953.6176.2680.3776.7985.0379.1465.3384.7577.8261.56
MPANet [47]CVPR-2170.5868.24-76.7480.95-82.880.7-83.780.9-
MSO [12]MM-2158.7056.42-63.0970.31-73.666.9-74.667.5-
AGM [26]MM-2169.6366.1152.2474.6878.3074.0088.4081.4568.5185.3481.1965.76
MCLNet [16]ICCV-2165.4061.9847.3972.5676.5872.1080.3173.0757.3975.9369.4952.63
SMCL [43]ICCV-2167.3961.78-68.8475.56-83.9379.83-83.0578.57-
FMCNet[60]CVPR-2266.3462.51-68.1574.09-89.1284.43-88.3883.86-
MAUM [27]CVPR-2271.6868.79-76.9781.94-87.8785.09-86.9584.34-
DEEN [61]CVPR-2374.771.8-80.383.3-91.185.1-89.583.4-
PMCM [34]IJCAI-2375.5471.16-81.5284.33-93.0989.57-91.4487.15-
PartMix [20]CVPR-2377.7874.62-81.5284.38-84.9382.52-85.6682.27-
SGIEL [20]CVPR-2377.1272.33-82.0782.95-95.3589.98-97.5791.41-
UnsupervisedSPCL [14]NIPS-2018.3719.3910.9926.8336.4233.0513.5914.8610.3611.7013.5610.09
MMT [13]ICLR-2021.4721.5311.5022.7931.5027.6625.6826.5119.5624.4225.5918.66
IICS [49]CVPR-2114.3915.748.4115.9124.8722.159.179.946.409.119.906.45
CAP [38]AAAI-2116.8215.717.0224.5730.7426.159.7111.568.7410.2111.347.92
Cluster Contrast [9]arXiv-2120.1622.0012.9723.3334.0130.8811.7613.889.9411.1412.998.99
ICE [3]ICCV-2120.5420.3910.2429.8138.3534.3212.9815.6411.9112.1814.8210.6
PPLR [6]CVPR-2211.9812.254.9712.7120.8117.6110.3011.948.1010.3911.237.04
OTLA [37]ECCV-2229.927.1-29.838.8-32.929.7-32.128.6-
H2H [24]TIP-2130.1529.40----23.8118.87----
ADCA [50]MM-2245.5142.7328.2950.6059.1155.1767.2064.0552.6768.4863.8149.62
DPIS [35]ICCV-2358.455.6-63.070.0-62.353.2-61.552.7-
CHCR [33]TCSVT-2359.4759.14----69.3164.74-69.9665.87-
DOTLA [5]MM-2350.3647.3632.4053.4761.7357.3585.6376.7161.5882.9174.9758.60
MBCCM [4]MM-2353.1448.1632.4155.2161.9857.1383.7977.8765.0482.8276.7461.73
PGM [48]CVPR-2357.2751.7834.9656.2362.7458.1369.4865.41-69.8565.17-
GUR [51]ICCV-2363.5161.6347.9371.1176.2372.5773.9170.2358.8875.0069.9456.21
SDCL (ours)-64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
", + "bbox": [ + 81, + 87, + 883, + 571 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/03b8730302b8916d1fed35d3632e3ad775d122e804352e4802bf7bde620a3569.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-arts on SYSU-MM01 and RegDB. Rank at $r$ accuracy(%) , mAP (%) and mINP (%) are reported." + ], + "table_footnote": [], + "table_body": "
SYSU-MM01RegDB
ComponentsAll SearchIndoor SearchVisible to InfraredInfrared to Visible
IndexBaselineLids+LrsidCNLCRAr1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
1✓49.5548.7036.9253.8560.9657.0563.1658.1242.5961.7056.7641.65
2✓✓53.2150.1440.8060.1565.2464.0269.9763.3448.9868.5463.1948.06
3✓✓✓55.2754.7941.6164.3871.1567.4076.6067.9751.3875.6468.4650.38
4✓✓✓59.4359.1848.8167.4874.3170.3080.5373.6558.1379.1370.1853.30
5✓✓✓✓✓64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
", + "bbox": [ + 78, + 601, + 888, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation studies on the SYSU-MM01 and RegDB. Rank at $r$ accuracy $\\left( \\% \\right)$ ,mAP $\\left( \\% \\right)$ and mINP $\\left( \\% \\right)$ are reported.", + "bbox": [ + 120, + 720, + 849, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 78, + 746, + 228, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we meticulously conduct ablation experiments to validate the individual efficacy of each module. The results are reported in Table 2.", + "bbox": [ + 78, + 773, + 467, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline denotes the augmented dual-contrastive learning framework [50] with a dual-path transformer architecture. The optimization for the network only uses deep features, i.e., training the model with $\\mathcal{L}_{id}^{vd} + \\mathcal{L}_{id}^{rd}$ .", + "bbox": [ + 78, + 821, + 467, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Deep and Shallow Contrastive Learn", + "bbox": [ + 78, + 885, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing. The incorporation of $\\mathcal{L}id^{vs} + \\mathcal{L}id^{rs}$ yields a $4\\% - 7\\%$ enhancement of rank-1 accuracy on SYSU-MM01 and RegDB datasets compared to the baseline. Despite shallow contrastive learning primarily optimizing the ReID model for feature learning within intra-modality, the findings underscore its capacity to augment cross-modality retrieval.", + "bbox": [ + 500, + 746, + 890, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of CNL. Compared with the baseline, the CNL improves the performance of $6\\% - 10\\%$ on SYSUMM01 and RegDB datasets. The main gain is achieved by the design of capturing credible homogeneous and hetero-", + "bbox": [ + 500, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16876", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82d50b90cfd8f8318601d1b8f5377f8b1e5bd9e44391585d0b96c9399882f90d.jpg", + "image_caption": [ + "Figure 3. Evaluation of hyper-parameters $\\lambda_{s}$ (left) and $\\gamma$ (right). The results are based on all search mode of SYSU-MM01 dataset. Rank-1 accuracy $(\\%)$ and mAP $(\\%)$ are reported." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 267, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52ce65ba4319aeb4e93719fb1ad2c644f447ab1e7deb96e68e5ad28b9cf3fa93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 89, + 446, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "genuine neighborhoods with the shallow and deep collaborative consistency, which is a dynamic optimization, reducing the impact of label noise in clustering and enhancing the robustness against multiple discrepancies.", + "bbox": [ + 75, + 270, + 468, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of CRA. In contrast to the baseline, the CRA module remarkably improves the accuracy. CRA establishes a linkage between the labels of two modalities through collaborative shallow and deep ranking relations, constituting a global association across the training set and effectively bolstering modality-invariant features.", + "bbox": [ + 75, + 334, + 468, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of SDCL. The amalgamation of CNL and CRA notably enhances Rank-1 accuracy across diverse settings. This observation substantiates the rationale behind SDCL. CRA provides cross-modality supervision at the beginning of each training epoch in an off-line manner, while CNL dynamically learns cross-modality features during training iteration in an online manner. By combining off-line label association and online feature learning, SDCL learns better representation from different perspectives.", + "bbox": [ + 75, + 428, + 468, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Further Analysis", + "text_level": 1, + "bbox": [ + 76, + 583, + 243, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper-parameter Analysis for $\\lambda_{s}$ and $\\gamma$ . We explore the influence of hyper-parameters $\\lambda_{s}$ in Eq. 20 and $\\gamma$ in Eq. 12, Eq. 13, Eq. 16, and Eq. 17, as shown in Fig. 3. The $\\lambda_{s}$ governs the balance of $\\mathcal{L}_d$ and $\\mathcal{L}_s$ in deep and shallow neighbor learning. When $\\lambda_{s} = 0$ , the $\\mathcal{L}_s$ is excluded. When $\\lambda_{base} = 0.5$ , the baseline method achieves a balance in deep and shallow neighbor learning. The $\\gamma$ is used to control the positive neighbor selection range in CNL. Setting $\\gamma = 0$ to 0 renders neighbor selection ineffective. When $\\gamma = 0.9$ , the CNL significantly reinforces the cross-modality learning.", + "bbox": [ + 75, + 609, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization. We perform feature space (t-SNE [36] map) and similarity distribution visualization for SDCL, as presented in Fig. 4. As evident in (a) and (b), SDCL brings infrared and visible positive sample points closer together, and effectively enhances the separation of cross-modality positive/negative distributions. This attests to the efficacy of our method in addressing modality discrepancies. Additionally, the visualization of shallow and deep feature map is presented in the supplementary materials.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48281d164bbfefded1b4a737da80543622f93a402eff34ad87560e760d6737e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 95, + 687, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa6d16b9d00ac1e85bac80572a898d0ba3c893bf1f69abae2e8065d565c8aef3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 95, + 870, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b9a04ad59ffbb44d8858199991ff37f5c20b0e3b6fa51c0fc973bc7a688b179.jpg", + "image_caption": [ + "(a) Baseline", + "Figure 4. The t-SNE (first row) and similarity distribution (second row) visualization of randomly selected identities. In the t-SNE visualization, identity is denoted by color, where circles represent the visible modality and triangles denote the infrared modality." + ], + "image_footnote": [], + "bbox": [ + 516, + 223, + 687, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e5168ecd89e06924d0228c49e44d4b4172a7de0934bb50e1add316e4a3f8fb7.jpg", + "image_caption": [ + "(b) SDCL" + ], + "image_footnote": [], + "bbox": [ + 697, + 223, + 870, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 443, + 617, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper investigates an extremely important and challenging problem, namely the unsupervised visible infrared person re-identification (US-VI-ReID) task, alleviating the reliance on expensive cross-modality annotation. To remedy the issues of multiple discrepancies, we propose a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer architecture, incorporating Collaborative Neighbor Learning (CNL), and Collaborative Ranking Association (CRA) module. SDCL significantly facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow patch embeddings and deep modality-shared features and capturing more discriminative representations for cross-modality retrieval. Experiments on two public benchmarks substantiate that our approach surpasses existing methodologies by a substantial margin. Moreover, it even outshines certain supervised counterparts, propelling VI-ReID toward real-world deployment.", + "bbox": [ + 496, + 468, + 890, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 753, + 658, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by the National Natural Science Foundation of China under Grants (62071338, 62176188, 62272354, 62306215), the Key Research and Development Program of Hubei Province (2022BCA009,2022BAD175), the Special Fund of Hubei Luojia Laboratory (220100015), and the Interdisciplinary Innovative Talents Foundation from Renmin Hospital of Wuhan University.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16877", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Yimin Liu, and Jianguo Jiang. Saliency and granularity: Discovering temporal coherence for video-based person re-identification. IEEE TCSVT, 2022. 1", + "[2] Cuiqun Chen, Mang Ye, Meibin Qi, and Bo Du. Sketchtrans: Disentangled prototype learning with transformer for sketchphoto recognition. IEEE TPAMI, 2023. 1", + "[3] Hao Chen, Benoit Lagadec, and Francois Bremond. Ice: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, 2021. 2, 7", + "[4] De Cheng, Lingfeng He, Nannan Wang, Shizhou Zhang, Zhen Wang, and Xinbo Gao. Efficient bilateral cross-modality cluster matching for unsupervised visible-infrared person reid. In ACM MM, 2023. 2, 7", + "[5] De Cheng, Xiaojian Huang, Nannan Wang, Lingfeng He, Zhihui Li, and Xinbo Gao. Unsupervised visible-infrared person Reid by collaborative learning with neighbor-guided label refinement. ACM MM, 2023. 2, 7", + "[6] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, 2022. 7", + "[7] Zhenyu Cui, Jiahuan Zhou, and Yuxin Peng. Dma: Dual modality-aware alignment for visible-infrared person re-identification. IEEE TIFS, 2024. 1", + "[8] Yongxing Dai, Jun Liu, Yan Bai, Zekun Tong, and Ling-Yu Duan. Dual-refinement: Joint label and feature refinement for unsupervised domain adaptive person re-identification. IEEE TIP, 2021. 2", + "[9] Zuozhuo Dai, Guangyuan Wang, Siyu Zhu, Weihao Yuan, and Ping Tan. Cluster contrast for unsupervised person re-identification. arxiv 2021. arXiv preprint arXiv:2103.11568, 2021. 2.7", + "[10] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, 1996. 6", + "[11] Jiawei Feng, Ancong Wu, and Wei-Shi Zheng. Shape-erased feature learning for visible-infrared person re-identification. In CVPR, 2023. 2", + "[12] Yajun Gao, Tengfei Liang, Yi Jin, Xiaoyan Gu, Wu Liu, Yidong Li, and Congyan Lang. Mso: Multi-feature space joint optimization network for rgb-infrared person re-identification. In ACM MM, 2021. 1, 6, 7", + "[13] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual mean-teaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. arXiv preprint arXiv:2001.01526, 2020. 7", + "[14] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. NeurIPS, 2020. 7", + "[15] Jianyang Gu, Weihua Chen, Hao Luo, Fan Wang, Hao Li, Wei Jiang, and Weijie Mao. Multi-view evolutionary training for unsupervised domain adaptive person re-identification. IEEE TIFS, 2022. 1", + "[16] Xin Hao, Sanyuan Zhao, Mang Ye, and Jianbing Shen. Cross-modality person re-identification via modality confusion and center aggregation. In ICCV, 2021. 6, 7" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, 2021. 6", + "[18] Wenke Huang, Mang Ye, Zekun Shi, and Bo Du. Generalizable heterogeneous federated cross-correlation and instance similarity learning. IEEE TPAMI, 2023. 1", + "[19] Wenke Huang, Mang Ye, Zekun Shi, Guancheng Wan, He Li, Bo Du, and Qiang Yang. A federated learning for generalization, robustness, fairness: A survey and benchmark. arXiv, 2023. 1", + "[20] Minsu Kim, Seungryong Kim, Jungin Park, Seongheon Park, and Kwanghoon Sohn. Partmix: Regularization strategy to learn part discovery for visible-infrared person re-identification. In CVPR, 2023. 1, 7", + "[21] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In AAAI, 2020. 1", + "[22] Shuang Li, Jiaxu Leng, Ji Gan, Mengjingcheng Mo, and Xinbo Gao. Shape-centered representation learning for visible-infrared person re-identification. arXiv preprint arXiv:2310.17952, 2023. 1", + "[23] Shuang Li, Fan Li, Jinxing Li, Huafeng Li, Bob Zhang, Dapeng Tao, and Xinbo Gao. Logical relation inference and multiview information interaction for domain adaptation person re-identification. IEEE TNNLS, 2023. 1", + "[24] Wenqi Liang, Guangcong Wang, Jianhuang Lai, and Xiaohua Xie. Homogeneous-to-heterogeneous: Unsupervised learning for rgb-infrared person re-identification. IEEE TIP, 2021. 2, 3, 7", + "[25] Shan Lin, Chang-Tsun Li, and Alex C. Kot. Multidomain adversarial feature generalization for person re-identification. IEEE TIP, 2021. 1", + "[26] Haojie Liu, Daoxun Xia, and Wei Jiang. Towards homogeneous modality learning and multi-granularity information exploration for visible-infrared person re-identification. IEEE Journal of Selected Topics in Signal Processing. 7", + "[27] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, 2022. 1, 7", + "[28] Yan Lu, Yue Wu, Bin Liu, Tianzhu Zhang, Baopu Li, Qi Chu, and Nenghai Yu. Cross-modality person re-identification with shared-specific feature transfer. In CVPR, 2020. 1", + "[29] Zefeng Lu, Ronghao Lin, and Haifeng Hu. Modality and camera factors bi-disentanglement for nir-vis object re-identification. IEEE TIFS, 2023.", + "[30] Chuanchen Luo, Chunfeng Song, and Zhaoxiang Zhang. Learning to adapt across dual discrepancy for cross-domain person re-identification. IEEE TPAMI, 2022. 1, 5", + "[31] Hao Luo, Pichao Wang, Yi Xu, Feng Ding, Yanxin Zhou, Fan Wang, Hao Li, and Rong Jin. Self-supervised pre-training for transformer-based person re-identification. arXiv preprint arXiv:2111.12084, 2021. 6", + "[32] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 2017. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16878", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zhiqi Pang, Chunyu Wang, Lingling Zhao, Yang Liu, and Gaurav Sharma. Cross-modality hierarchical clustering and refinement for unsupervised visible-infrared person re-identification. IEEE TCSVT, 2023. 2, 3, 7", + "[34] Zhihao Qian, Yutian Lin, and Bo Du. Visible-infrared person re-identification via patch-mixed cross-modality learning. *IJCAI*, 2023. 1, 7", + "[35] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, 2023. 6, 7", + "[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8", + "[37] Jiangming Wang, Zhizhong Zhang, Mingang Chen, Yi Zhang, Cong Wang, Bin Sheng, Yanyun Qu, and Yuan Xie. Optimal transport for label-efficient visible-infrared person re-identification. ECCV, 2022. 2, 3, 6, 7", + "[38] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, 2021. 7", + "[39] Yiming Wang, Guanqiu Qi, Shuang Li, Yi Chai, and Huafeng Li. Body part-level domain alignment for domain-adaptive person re-identification with transformer framework. IEEE TIFS, 2022. 1", + "[40] Yadi Wang, Hongyun Zhang, Duoqian Miao, and Witold Pedrycz. Multi-granularity re-ranking for visible-infrared person re-identification. CAAI Transactions on Intelligence Technology, 2023. 1", + "[41] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In CVPR, 2019. 1", + "[42] Zheng Wang, Zhixiang Wang, Yinqiang Zheng, Yang Wu, Wenjun Zeng, and Shin'ichi Satoh. Beyond intra-modality: A survey of heterogeneous person re-identification. arXiv preprint arXiv:1905.10048, 2019. 1", + "[43] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Syncretic modality collaborative learning for visible infrared person re-identification. In ICCV, 2021. 7", + "[44] Ancong Wu, Wei-Shi Zheng, and Jian-Huang Lai. Robust depth-based person re-identification. IEEE TIP, 2017. 1", + "[45] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In ICCV, 2017. 1, 6", + "[46] Dongming Wu, Mang Ye, Gaojie Lin, Xin Gao, and Jianbing Shen. Person re-identification by context-aware part attention and multi-head collaborative learning. IEEE TIFS, 2022. 1", + "[47] Qiong Wu, Pingyang Dai, Jie Chen, Chia-Wen Lin, Yongjian Wu, Feiyue Huang, Bineng Zhong, and Rongrong Ji. Discover cross-modality nuances for visible-infrared person re-identification. In CVPR, 2021. 1, 7", + "[48] Zesen Wu and Mang Ye. Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning. In CVPR, 2023. 2, 3, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, 2021. 2, 7", + "[50] Bin Yang, Mang Ye, Jun Chen, and Zesen Wu. Augmented dual-contrastive aggregation learning for unsupervised visible-infrared person re-identification. In ACM MM, 2022. 1, 2, 3, 6, 7", + "[51] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In ICCV, 2023. 2, 6, 7", + "[52] Bin Yang, Jun Chen, and Mang Ye. Top-k visual tokens transformer: Selecting tokens for visible-infrared person re-identification. In ICASSP, 2023. 1", + "[53] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE TIFS, pages 407–419, 2019. 1, 6", + "[54] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In ECCV, 2020. 1, 6, 7", + "[55] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE TIFS, 2021. 1, 2", + "[56] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In ICCV, 2021. 2, 6, 7", + "[57] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 2021. 1, 6, 7", + "[58] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE TPAMI, 2023. 1", + "[59] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 2022. 1", + "[60] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In CVPR, 2022. 7", + "[61] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, 2023. 1, 7", + "[62] Yiyuan Zhang, Yuhao Kang, Sanyuan Zhao, and Jianbing Shen. Dual-semantic consistency learning for visible-infrared person re-identification. IEEE TIFS, 2022. 1", + "[63] Yafei Zhang, Yongzeng Wang, Huafeng Li, and Shuang Li. Cross-compatible embedding and semantic consistent feature construction for sketch re-identification. In ACM MM, 2022. 1", + "[64] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, 2015. 3" + ], + "bbox": [ + 501, + 92, + 890, + 885 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16879", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_model.json b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8f1e858ee856d60adc08fc0e13358200f633456 --- /dev/null +++ b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_model.json @@ -0,0 +1,2530 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.13, + 0.849, + 0.175 + ], + "angle": 0, + "content": "Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.204, + 0.626, + 0.222 + ], + "angle": 0, + "content": "Bin Yang Jun Chen* Mang Ye*" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.223, + 0.74, + 0.239 + ], + "angle": 0, + "content": "National Engineering Research Center for Multimedia Software," + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.24, + 0.837, + 0.257 + ], + "angle": 0, + "content": "School of Computer Science, Hubei Luojia Laboratory, Wuhan University, Wuhan, China" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.26, + 0.626, + 0.274 + ], + "angle": 0, + "content": "https://github.com/yangbincv/SDCL" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.735 + ], + "angle": 0, + "content": "Unsupervised visible-infrared person re-identification (US-VI-ReID) centers on learning a cross-modality retrieval model without labels, reducing the reliance on expensive cross-modality manual annotation. Previous US-VI-ReID works gravitate toward learning cross-modality information with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. The shallow feature from the shallow layers contains nuanced detail information, which is critical for effective cross-modality learning but is disregarded regrettably by the existing methods. To address the above issues, we design a Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration which are harnessed for neighbor alignment, enhancing the robustness in a dynamic manner. Furthermore, CRA associates the cross-modality labels with the ranking association between shallow and deep features, furnishing valuable supervision for cross-modality learning. Extensive experiments validate the superiority of our method, even outperforming certain supervised counterparts." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.208, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Person re-identification (ReID) aims at matching the same person image captured by non-overlapping cameras. In recent times, ReID has garnered significant attention from the computer vision research community, owing to its pivotal role in the context of intelligent video surveillance applications [2, 18, 19, 23, 25, 28-30, 39, 41, 44, 46, 57," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.309, + 0.697, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.439, + 0.622, + 0.459 + ], + "angle": 0, + "content": "Ground Truth Cross-modality Label Pairs" + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.44, + 0.692, + 0.459 + ], + "angle": 0, + "content": "Deep Features" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.311, + 0.891, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.44, + 0.774, + 0.459 + ], + "angle": 0, + "content": "Shallow Features" + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.442, + 0.888, + 0.459 + ], + "angle": 0, + "content": "Negative Optimization" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Figure 1. Illustration of the motivation. Previous advanced works for US-VI-ReID mainly focus on learning features and associating cross-modality labels with deep features. However, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in (a). In this paper, we incorporate shallow and deep features with collaborative learning and label association to remedy these problems, as shown in (b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.748 + ], + "angle": 0, + "content": "59, 62, 63]. The rank-1 accuracy has exhibited encouraging outcomes in supervised person ReID. Notably, these person images are typically acquired by visible cameras within well-illuminated environments. However, it is crucial to recognize that visible cameras often fail to capture sufficient information of individuals in low-light settings, consequently constraining the practicality of single-modality ReID in the context of night-time surveillance [1, 7, 15, 21, 40, 42, 45, 53, 61]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In contrast, recently deployed cameras possess the capability to seamlessly transition into far/near-infrared mode during nighttime operations [50, 55]. Various techniques [12, 20, 22, 27, 34, 47, 52, 54-58, 62] have been proposed for visible-infrared person re-identification (VI-ReID), resulting in commendable accuracy levels. Notably, these achievements are facilitated by extensive human-labeled cross-modality datasets. However, annotating cross-modality datasets demands more resources compared to the annotation of single-modality ReID datasets. This" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16870" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "intricate annotation task presents a formidable and pivotal challenge, ultimately giving rise to the crucial task of unsupervised VI-ReID (US-VI-ReID) for reducing the cost of expensive cross-modality annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.473, + 0.669 + ], + "angle": 0, + "content": "Some prior research endeavors [4, 5, 24, 33, 37, 48, 50, 51] have proposed preliminary solutions for US-VI-ReID. These solutions gravitate toward learning cross-modality feature with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, i.e., intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in Fig. 1. It is important to recognize that the shallow features originating from the modality-specific shallow layers harbor a wealth of nuanced detail information pertaining to pedestrian attributes, which is critical for unsupervised cross-modality representation learning but is disregarded regrettably by the existing methods. To harness the complete potential of the shallow information, we devise a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration as the constraints for seeking reliable intramodality and inter-modality neighbor learning, thereby guaranteeing the cultivation of modality-invariant and discriminative representations in a dynamic manner. Furthermore, CRA is developed with a global perspective to exploit cross-modality ranking consistency between deep and shallow features, associating the label of two modalities in a collaborative manner and furnishing valuable cross-modality supervision for cross-modality learning. Ultimately, within the shallow-deep collaborative learning framework with CNL and CRA modules, we acquire a robust representation, effectively mitigating the cross-modality discrepancy under unsupervised conditions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.671, + 0.458, + 0.686 + ], + "angle": 0, + "content": "The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.69, + 0.468, + 0.763 + ], + "angle": 0, + "content": "- We propose a shallow-deep collaborative learning framework based on the transformer architecture. This framework facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow and deep features." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.766, + 0.468, + 0.824 + ], + "angle": 0, + "content": "- We propose a collaborative neighbor learning module to formulate dependable intra-modality and cross-modality neighbor learning, enabling the model to capture modality-invariant and discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.468, + 0.884 + ], + "angle": 0, + "content": "- We propose a collaborative ranking association module to explore intra-modality and cross-modality ranking consistencies, unifying the cross-modality labels and providing invaluable cross-modality supervision." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.887, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- Extensive experiments validate that our SDCL frame" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.69, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "work surpasses existing methods on two mainstream VI-ReID benchmarks, consistently improving the unsupervised cross-modality retrieval performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.16, + 0.642, + 0.176 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.188, + 0.851, + 0.205 + ], + "angle": 0, + "content": "2.1. Supervised Visible-Infrared Person ReID" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Currently, advanced supervised VI-ReID focuses on bridging the gap between the two modalities and learning robust representations against modality discrepancy. Ye et al. [55] developed a dynamic tri-level relation mining (DTRM) framework simultaneously to explore the cross-modality relation cues of channel-level, part-level intra-modality, and graph-level. CAJ [56] proposed a channel-mixed learning to handle the intra-modality and cross-modality variations by randomly exchanging the color channels. SGIEL [11] proposed a shape-erased feature learning paradigm, jointly learning shape-related feature in one subspace and shape-erased features in the orthogonal complement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.49 + ], + "angle": 0, + "content": "Nevertheless, the remarkable performances exhibited by these methods require extensive human-labeled cross-modality datasets. In this work, we pivot our focus to the realm of unsupervised visible-infrared person ReID, where the luxury of identity annotations is absent, presenting important applications for real-world VI-ReID deployments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.509, + 0.871, + 0.525 + ], + "angle": 0, + "content": "2.2. Unsupervised Single-Modality Person ReID" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Unsupervised single-modality ReID endeavors to train the ReID model using unlabeled data captured by visible cameras. The majority of studies have embraced cluster algorithms to derive pseudo labels for optimizing the model. Cluster Contrast [9] utilized a distinctive cluster representation to delineate each cluster, addressing the issue of cluster inconsistency. Chen et al. [3] introduced Inter-instance Contrastive Encoding (ICE), which harnesses inter-instance pairwise similarity scores to enhance preceding class-level contrastive ReID methodologies. IICS [49] tackled the unsupervised ReID challenge by decomposing the similarity computation into two stages, namely, the intra-domain and inter-domain computations, respectively. Dai et al. [8] introduced a dual-refinement approach to concurrently enhance pseudo labels during the offline clustering phase and refine features during the online training phase, augmenting label purity and feature discriminability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The above methods are dedicated to addressing issues in single-modality ReID. When employed in the context of US-VI-ReID, these approaches encounter challenges stemming from cross-modality discrepancies and the absence of cross-modality (visible-infrared) identity labels. This impedes the learning of inter-modality feature and the generation of reliable cross-modality pseudo labels." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16871" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.449, + 0.107 + ], + "angle": 0, + "content": "2.3. Unsupervised Visible-Infrared Person ReID" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.341 + ], + "angle": 0, + "content": "Several studies [24, 37, 48, 50] represent initial endeavors in the realm of US-VI-ReID. Yang et al. [50] introduced a novel Augmented Dual-Contrastive Aggregation (ADCA) learning framework founded on the principles of homogeneous joint learning and heterogeneous aggregation, establishing a robust baseline for purely unsupervised VI-ReID. PGM [48] developed a progressive graph matching method to systematically extract cross-modality correspondences, formulating correspondence mining as a graph-matching process. H2H [24] introduced a homogeneous-to-heterogeneous approach through two-stage learning, incorporating Market1501 [64] as additional labeled data. CHCR [33] proposed a cross-modality hierarchical clustering and refinement method by promoting modality-invariant feature learning and improving the reliability of pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.471, + 0.493 + ], + "angle": 0, + "content": "Nevertheless, the above methods primarily concentrate on feature learning with deep features. However, interfered by the intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. Our approach integrates shallow and deep features through collaborative feature learning and cross-modality label association, which is distinguished from previous works, surpassing the performance of existing US-VI-ReID methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.251, + 0.525 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.837 + ], + "angle": 0, + "content": "The framework of shallow-deep collaborative learning (SDCL) is shown in Fig. 2, incorporating the Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. SDCL adopts a dual-path transformer architecture with dual-contrastive learning [50], which has two shallow modality-specific patch embedding layers and a modality-shared transformer. Instance memory and cluster memory are constructed for shallow embeddings and deep features within each modality. Instance memory stores all training image features. Cluster memory is built by averaging the instance features with the same pseudo labels. With the above memories, SDCL develops the CNL module to exploit homogeneous and heterogeneous collaboration as the constraints for seeking reliable intra-modality and inter-modality neighbor learning. CRA explores the ranking consistency of cross-modality shallow embeddings and deep features to associate reliable cross-modality identities. With the collaboration of shallow embeddings and deep features within CNL and CRA modules, SDCL learns better representation for cross-modality retrieval." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.387, + 0.864 + ], + "angle": 0, + "content": "3.1. Shallow-deep Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We first introduce shallow-deep contrastive learning based on augmented dual-contrast learning [50] with a dual-path" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.093, + 0.666, + 0.105 + ], + "angle": 0, + "content": "transformer architecture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.893, + 0.168 + ], + "angle": 0, + "content": "Cluster Memory Initialization. At the beginning of each training epoch, we construct shallow and deep cluster memories for each modality by averaging the shallow and deep feature of one cluster, which can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.181, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\phi_ {k} ^ {v s} = \\frac {1}{| \\mathcal {H} _ {k} ^ {v} |} \\sum_ {u _ {n} ^ {v s} \\in y _ {k} ^ {v}} u _ {n} ^ {v s}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.237, + 0.892, + 0.276 + ], + "angle": 0, + "content": "\\[\n\\phi_ {k} ^ {v d} = \\frac {1}{\\left| \\mathcal {H} _ {k} ^ {v} \\right|} \\sum_ {u _ {n} ^ {v d} \\in y _ {k} ^ {v}} u _ {n} ^ {v d}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.291, + 0.892, + 0.329 + ], + "angle": 0, + "content": "\\[\n\\phi_ {l} ^ {r s} = \\frac {1}{| \\mathcal {H} _ {l} ^ {r} |} \\sum_ {u _ {n} ^ {r s} \\in y _ {l} ^ {r}} u _ {n} ^ {r s}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.344, + 0.892, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\phi_ {l} ^ {r d} = \\frac {1}{| \\mathcal {H} _ {l} ^ {r} |} \\sum_ {u _ {n} ^ {r d} \\in y _ {l} ^ {r}} u _ {n} ^ {r d}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.482 + ], + "angle": 0, + "content": "where \\( u_{n}^{r*} \\) and \\( u_{n}^{v*} \\) are infrared and visible instance features, respectively. \\( u_{n}^{*s} \\) and \\( u_{n}^{*d} \\) are shallow and deep instance features, respectively. \\( \\mathcal{H}_k \\) indicates the \\( k \\)-th cluster set and \\( |\\cdot| \\) counts the number of samples of a set. Meanwhile, we store the instance features in shallow and deep infrared and visible instance memories \\( U^{rs} \\), \\( U^{rd} \\), \\( U^{vs} \\) and \\( U^{vd} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Cluster Memory Updating. During each training iteration, the deep and shallow cluster memories of two modalities by a momentum updating strategy:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.541, + 0.892, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\phi_ {k} ^ {(\\delta)} \\leftarrow \\alpha \\phi_ {k} ^ {(\\delta - 1)} + (1 - \\alpha) q, q \\in y _ {k}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.621 + ], + "angle": 0, + "content": "where \\( q \\) is the query features sampled from training set in each training iteration, respectively. \\( \\alpha \\) is the momentum factor. \\( \\delta \\) is the training iteration number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Contrastive Loss. Given shallow and deep visible and infrared query \\( q^{vs} \\), \\( q^{vd} \\), \\( q^{rs} \\), and \\( q^{rs} \\), we compute the contrastive loss for visible modality by the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.682, + 0.891, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i d} ^ {v s} = - \\log \\frac {\\exp \\left(q ^ {v s} \\cdot \\phi_ {+} ^ {v s} / \\tau\\right)}{\\sum_ {k = 0} ^ {K} \\exp \\left(q ^ {v s} \\cdot \\phi_ {k} ^ {v s} / \\tau\\right)}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.732, + 0.892, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i d} ^ {v d} = - \\log \\frac {\\exp \\left(q ^ {v d} \\cdot \\phi_ {+} ^ {v d} / \\tau\\right)}{\\sum_ {k = 0} ^ {K} \\exp \\left(q ^ {v d} \\cdot \\phi_ {k} ^ {v d} / \\tau\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.826 + ], + "angle": 0, + "content": "where \\(\\phi_{+}\\) is the positive memory corresponding to the pseudo label of \\(q\\) and the \\(\\tau\\) is a temperature. The infrared contrastive loss \\(\\mathcal{L}_{id}^{rs}\\) and \\(\\mathcal{L}_{id}^{rd}\\) are obtained by the same way." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We optimize the modality-specific shallow embedding layers and deep modality-shared layers by combining deep and shallow contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i d} = \\mathcal {L} _ {i d} ^ {v d} + \\mathcal {L} _ {i d} ^ {r d} + \\mathcal {L} _ {i d} ^ {v s} + \\mathcal {L} _ {i d} ^ {r s} \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.895, + 0.511 + ], + "angle": 0, + "content": "Figure 2. Illustration of shallow-deep collaborative learning. It comprises shallow-deep contrastive learning, collaborative neighbor learning, and collaborative ranking association. SDCL finds reliable neighborhood and cross-modality labels with shallow-deep collaborations, enhancing the robustness of learned representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.373, + 0.552 + ], + "angle": 0, + "content": "3.2. Collaborative Neighbor Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.695 + ], + "angle": 0, + "content": "Our basic rationale for the CNL module is that deep and shallow features are complementary, and their intrinsic consistency relationship can be used to constrain intra-modality and inter-modality optimization, formulating collaborative learning to enhance the robustness against intra-modality and inter-modality variations. From the above perspective, CNL explores essential homogeneous and heterogeneous shallow-deep collaborations to seek reliable intra-modality and inter-modality neighbor learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.471, + 0.743 + ], + "angle": 0, + "content": "Homogeneous and Heterogeneous Collaboration. Given a query \\( q \\), we can get the similarity \\( s(q_i, u_j) \\) between the query and each instance in the training set by:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.752, + 0.47, + 0.784 + ], + "angle": 0, + "content": "\\[\ns \\left(q _ {i}, u _ {j}\\right) = \\frac {q _ {i} \\cdot u _ {j}}{\\left| \\left| q _ {i} \\right| \\right| _ {2} \\left| \\left| u _ {j} \\right| \\right| _ {2}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.793, + 0.47, + 0.873 + ], + "angle": 0, + "content": "where \\(q_{i}\\) and \\(u_{j}\\) come from shallow or deep features of visible or infrared modality to get multiple types of homogeneous and heterogeneous shallow or deep query-instance similarity \\(s(q_{i}^{vs},u_{j}^{vs})\\), \\(s(q_{i}^{rs},u_{j}^{rs})\\), \\(s(q_{i}^{vs},u_{j}^{rs})\\), \\(s(q_{i}^{rs},u_{j}^{vs})\\), \\(s(q_{i}^{vd},u_{j}^{vd})\\), \\(s(q_{i}^{rd},u_{j}^{rd})\\), \\(s(q_{i}^{vd},u_{j}^{rd})\\), and \\(s(q_{i}^{rd},u_{j}^{vd})\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "With the above multiple types of similarities, we seek reliable shallow and deep intra-modality and inter-modality" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.536, + 0.892, + 0.567 + ], + "angle": 0, + "content": "neighbors, formulating collaborative neighbor learning. The intra-modality neighbors can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.577, + 0.892, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right) = \\left\\{\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v s}\\right) \\cap \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v d}\\right) \\right\\}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.621, + 0.892, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {r}\\right) = \\left\\{\\mathcal {N} ^ {r} \\left(q _ {i} ^ {r s}\\right) \\cap \\mathcal {N} ^ {r} \\left(q _ {i} ^ {r d}\\right) \\right\\}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.645, + 0.894, + 0.707 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}^v(q_i^{v*})\\) and \\(\\mathcal{N}^r(q_i^{r*})\\) are neighborhood sets searched by visible-visible and infrared-infrared similarity, respectively. \\(* \\in \\{s, d\\}\\) represents the shallow and deep feature. The \\(\\mathcal{N}^v(q_i^{vs})\\) and \\(\\mathcal{N}^r(q_i^{vd})\\) can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.714, + 0.892, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v s}\\right) = \\left\\{u _ {j} ^ {v s} \\mid s \\left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {v}} s \\left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\\right) \\right\\}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.765, + 0.892, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v d}\\right) = \\left\\{u _ {j} ^ {v d} \\mid s \\left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N v} s \\left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\\right) \\right\\}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.892, + 0.843 + ], + "angle": 0, + "content": "where the \\(N^v\\) denotes the number of visible instances and the \\(\\gamma\\) is the positive neighbor selection range. The \\(\\mathcal{N}^r (q_i^{rs})\\) and \\(\\mathcal{N}^r (q_i^{rd})\\) can be calculated by similar manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.843, + 0.892, + 0.872 + ], + "angle": 0, + "content": "Given query \\( q_{i}^{v} \\) or \\( q_{i}^{r} \\), the inter-modality neighbors can be denoted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v}\\right) = \\left\\{\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v s}\\right) \\cap \\mathcal {N} ^ {r} \\left(q _ {i} ^ {v d}\\right) \\right\\}, \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16873" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.156, + 0.106, + 0.469, + 0.125 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {v} \\left(q _ {i} ^ {r}\\right) = \\left\\{\\mathcal {N} ^ {v} \\left(q _ {i} ^ {r s}\\right) \\cap \\mathcal {N} ^ {v} \\left(q _ {i} ^ {r d}\\right) \\right\\}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.13, + 0.425, + 0.148 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}^r (q_i^{vs})\\) and \\(\\mathcal{N}^r (q_i^{vd})\\) are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.156, + 0.469, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v s}\\right) = \\left\\{u _ {j} ^ {r s} \\mid s \\left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {r}} s \\left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\\right) \\right\\}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.208, + 0.469, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v d}\\right) = \\left\\{u _ {j} ^ {r d} \\mid s \\left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {r}} s \\left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\\right) \\right\\}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.241, + 0.469, + 0.331 + ], + "angle": 0, + "content": "where the \\(N^r\\) denotes the number of infrared instances. Similarly, we can get the \\(\\mathcal{N}^v(q_i^{rs})\\) and \\(\\mathcal{N}^v(q_i^{rd})\\). With the above procedure, we get the \\(\\mathcal{N}^v(q_i^v), \\mathcal{N}^r(q_i^r), \\mathcal{N}^r(q_i^v)\\) and \\(\\mathcal{N}^v(q_i^r)\\), which are the reliable neighborhood set searched by the constraints of homogeneous or heterogeneous shallow-deep collaborations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.469, + 0.407 + ], + "angle": 0, + "content": "Collaborative Neighbor Learning. With the neighborhood set \\(\\mathcal{N}^v(q_i^v)\\), \\(\\mathcal{N}^r(q_i^r)\\), \\(\\mathcal{N}^r(q_i^v)\\) and \\(\\mathcal{N}^v(q_i^r)\\), we can perform neighbor learning. Given query \\(q_i^{vs}\\) and \\(q_i^{vs}\\), we can obtain the expression of the visible-visible shallow and deep neighbor learning by:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.418, + 0.469, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} ^ {v v} = - \\frac {1}{N ^ {b}} \\sum_ {i = 1} ^ {N ^ {b}} \\sum_ {j \\in \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right)} \\log \\frac {\\exp \\left(s \\left(q _ {i} ^ {v s} , u _ {j} ^ {v s}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N ^ {v}} \\exp \\left(s \\left(q _ {i} ^ {v s} , u _ {n} ^ {v s}\\right) / \\tau\\right)}, \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.497, + 0.469, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} ^ {v v} = - \\frac {1}{N ^ {b}} \\sum_ {i = 1} ^ {N ^ {b}} \\sum_ {j \\in \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right)} \\log \\frac {\\exp \\left(s \\left(q _ {i} ^ {v d} , u _ {j} ^ {v d}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N ^ {v}} \\exp \\left(s \\left(q _ {i} ^ {v d} , u _ {n} ^ {v d}\\right) / \\tau\\right)}, \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.566, + 0.469, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {v v} = \\mathcal {L} _ {d} ^ {v v} + \\lambda_ {s} \\mathcal {L} _ {s} ^ {v v}, \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.59, + 0.469, + 0.666 + ], + "angle": 0, + "content": "where \\(N^b\\) is the batch size of query \\(q_{i}\\). Similarly, the neighbor learning for infrared-infrared \\(\\mathcal{L}^{rr}\\), infrared-visible \\(\\mathcal{L}^{rv}\\) and visible-infrared \\(\\mathcal{L}^{vr}\\) can be obtained by similar ways. The final optimization for neighbor learning is denoted by the following combination:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.679, + 0.469, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n e i g h b o r}} = \\mathcal {L} ^ {v v} + \\mathcal {L} ^ {r r} + \\mathcal {L} ^ {r v} + \\mathcal {L} ^ {v r}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.708, + 0.367, + 0.723 + ], + "angle": 0, + "content": "The overall loss for SDCL is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.737, + 0.469, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {n e i g h b o r}} + \\mathcal {L} _ {i d} \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Discussion. In contrast to neighborhood learning in [30], which exclusively relies on deep features for the selection of positive neighbors. Under the multiple discrepancies, the similarity of individual deep features is unreliable, resulting in wrong neighbor selection and negative optimization. Our method embraces collaborative neighbor learning, which dynamically refines cross-modality supervision through shallow and deep collaboration, achieving effective cross-modality feature alignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.804, + 0.107 + ], + "angle": 0, + "content": "3.3. Collaborative Ranking Association" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.114, + 0.892, + 0.221 + ], + "angle": 0, + "content": "The CRA module is underpinned by dual rationales: (1) Shallow and deep features across the two modalities should have the same ranking consistency. (2) The ranking consistencies of shallow and deep features within one modality are more accurate. Accordingly, the CRA module contains two processes, i.e., inter-modality ranking association and intra-modality ranking smoothing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Inter-modality Ranking Association. With the similarities \\( s(u_{i}^{vs},u_{j}^{rs}) \\) and \\( s(u_{i}^{vd},u_{j}^{rd}) \\), we can get two visible-infrared ranking lists \\( \\{s(u_i^{vs},u_j^{rs}),j\\in [1,N^r ]\\} \\) and \\( \\{s(u_i^{vd},u_j^{rd}),j\\in [1,N^r ]\\} \\) for \\( u_{i}^{vs} \\) and \\( u_{i}^{vd} \\), which are denoted as \\( \\mathcal{R}_s^{vr}(u_i^{vs}) \\) and \\( \\mathcal{R}_d^{vr}(u_i^{vd}) \\). The label of \\( k \\)-th similar infrared instance in two ranking lists can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.33, + 0.891, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r s}}, u _ {j} ^ {r s} = \\mathcal {R} _ {s} ^ {v r} \\left(u _ {i} ^ {v s}\\right) [ k ], \\tag {23}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.37, + 0.891, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r d}}, u _ {j} ^ {r d} = \\mathcal {R} _ {d} ^ {v r} \\left(u _ {i} ^ {v d}\\right) [ k ], \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.493 + ], + "angle": 0, + "content": "where \\(\\tilde{y}_{u_i^{vs}}^{vr}[k]\\) and \\(\\tilde{y}_{u_i^{vd}}^{vr}[k]\\) are refined \\(k\\)-th cross-modality labels of \\(u_i^{vs}\\) and \\(u_i^{vd}\\) through the ranking on visible-infrared shallow and deep similarities. Based on the rationale (1), we propose to associate the cross-modality labels with the intersection of two label sets to investigate collaborative ranking consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.505, + 0.891, + 0.527 + ], + "angle": 0, + "content": "\\[\nI _ {u _ {i} ^ {v}} (k) = \\left\\{\\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] \\cap \\left\\{\\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ n ] \\right\\} _ {n = 1} ^ {N} \\right\\}, \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.601 + ], + "angle": 0, + "content": "where the \\( I_{u_i^v}(k) \\) records the samples with the same identity of top- \\( k \\) identity in instances from top-1 to top- \\( N \\). The reliable refined cross-modality label of \\( u_i^v \\) can be expressed by the label of maximum count in \\( I_{u_i^v}(k) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.608, + 0.891, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} _ {u _ {i} ^ {v}} ^ {c m} = \\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ], k = \\underset {k} {\\operatorname {a r g m a x}} \\left(\\left\\{\\left| I _ {u _ {i} ^ {v}} (k) \\right|, k \\in [ 1, K ] \\right\\}\\right), \\tag {26}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.752 + ], + "angle": 0, + "content": "where \\(|\\cdot|\\) denotes the counting function. The argmax operation traverses \\(I_{u_i^v}(k)\\) and finds the labels with maximum number as cross-modality refined labels. Then, we convert the label list \\(\\{\\tilde{y}_{u_i^v}^{cm}, i \\in [1, N^v]\\}\\) to the form of one-hot code matrix \\(\\tilde{Y}^{cm} \\in \\mathbb{R}^{N^v \\times C^r}\\) by setting the column according to the refine labels to 1 and the rest to 0, where \\(C^r\\) is the class number of infrared modality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.753, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Intra-modality Collaborative Smoothing. Based on the rationale (2), two shallow and deep homogeneous shallow \\( P_{s}^{vv} \\) and deep \\( P_{d}^{vv} \\) similarity matrices are constructed to investigate intra-modality ranking consistency, enhancing the precision of refined cross-modality pseudo labels by:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.84, + 0.891, + 0.86 + ], + "angle": 0, + "content": "\\[\nP _ {s} ^ {v v} (i, j) = s \\left(u _ {i} ^ {v s}, u _ {j} ^ {v s}\\right), \\tag {27}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.884, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\nP _ {d} ^ {v v} (i, j) = s \\left(u _ {i} ^ {v d}, u _ {j} ^ {v d}\\right), \\tag {28}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16874" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.472, + 0.155 + ], + "angle": 0, + "content": "where \\(P_{*}^{vv}\\in \\mathbb{R}^{N^{v}\\times N^{v}}\\) represents the intra-modality similarity structure. In order to explore the intra-modality shallow and deep collaboration, we calculate the sum of \\(P_{s}^{vv}(i,j)\\) and \\(P_d^{vv}(i,j)\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.161, + 0.47, + 0.18 + ], + "angle": 0, + "content": "\\[\nP ^ {v v} (i, j) = P _ {s} ^ {v v} (i, j) + P _ {d} ^ {v v} (i, j) \\tag {29}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.47, + 0.262 + ], + "angle": 0, + "content": "where \\( P^{vv} \\) indicates the consistency of shallow and deep similarity matrix. We keep the 5-max values of \\( P^{vv} \\) in each row to 1 and the rest to 0, acquiring the ranking relations. The process of intra-modality ranking smoothing is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.27, + 0.469, + 0.289 + ], + "angle": 0, + "content": "\\[\nY ^ {c m} = P ^ {v v} \\tilde {Y} ^ {c m}, \\tag {30}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.569 + ], + "angle": 0, + "content": "where \\( Y^{cm} \\in \\mathbb{R}^{N^v \\times C^r} \\) is the final refined cross-modality label matrix of visible instance. In \\( Y^{cm} \\), the column number of the maximum value in each row is the refined label of samples. In our work, we refine the pseudo labels of the visible modality to the infrared modality and smooth the cross-modality labels with the shallow and deep similarity within the visible modality. Then, the infrared and refined visible labels will be used to construct a modality-shared memory for contrastive learning within two modalities [51]. Discussion. Our CRA is essentially the collaboration of shallow and deep ranking processes with a global perspective at the beginning of each training epoch, which is distinguished from previous works employing a solitary-pattern deep feature similarity [37, 48, 50, 51]. The imposition of a collaborative consistency constraint on dual-pattern similarities within CRA confers a distinct advantage, facilitating the exploration of more dependable cross-modality relations and providing valuable cross-modality supervision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.581, + 0.21, + 0.599 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.37, + 0.621 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed SDCL framework on two widely-used visible-infrared person ReID datasets, i.e., SYSU-MM01 [45] and RegDB [32]. SYSU-MM01 dataset is collected by 6 different cameras, including 22258 visible and 11909 near-infrared images of 395 training identities. We perform ten trials of the gallery set selection [53] and calculate the average performance following existing methods with all-search and indoor-search testing mode. RegDB dataset is captured by two aligned visible and thermal camera system, which has less challenges for VI-ReID. 206 identities with 2,060 images are selected for training and this procedure is repeated 10 times following [54], and calculate the average performance with visible to thermal and thermal to visible testing mode." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Evaluation Protocols. Following existing works, Cumulative Matching Characteristics (CMC), mean Average Precision (mAP) and mean inverse negative penalty (mINP) [57] are calculated as the evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.719, + 0.108 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.435 + ], + "angle": 0, + "content": "The proposed framework is implemented with PyTorch. SDCL incorporates the feature extractor from TransReID [17] as the backbone network with augmented dual-contrastive learning [50]. The shallow patch embedding layers are constructed with IBN and CNN module [31]. The concatenation of shallow and deep class tokens is used to calculate the cosine similarity for retrieval. DBSCAN [10] is conducted to generate pseudo labels. The visible and infrared images are resized to \\(288 \\times 144\\) before entering the network. We sample 8 pseudo identities and 16 instances for each pseudo identity for each modality within one batch. Channel augmentation [56], random crop, horizontal flipping, and random erasing are adopted for data augmentation. We adopt the SGD optimizer with the initial learning rate of \\(3.5e - 4\\) to train the model. The model is trained in a total of 50 epochs. The CRA module is added in the last 20 epochs. The \\(\\lambda_{s}\\) in Eq. 20 is set to 0.5. The \\(\\gamma\\) in Eq. 12, Eq. 13, Eq. 16, and Eq. 17 is set to 0.9. The \\(N\\) and \\(K\\) in Eq. 25 and Eq. 26 are set to 20 for SYSU-MM01 and 10 for RegDB, respectively. The contrastive learning settings follow [50]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.449, + 0.866, + 0.466 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.519 + ], + "angle": 0, + "content": "In Table 1, our SDCL is compared with supervised and unsupervised VI-ReID methods on two benchmarks including SYSU-MM01 and RegDB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Comparison with Unsupervised Methods. As reported in Table 1, our method exhibits superior performance compared to the current advanced unsupervised methods. More precisely, our SDCL attains a remarkable \\(64.49\\%\\) and \\(86.91\\%\\) rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. Compared with the best current method GUR [51], our method also exceeds the about \\(1\\%\\) and \\(10\\%\\) rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. In comparison to DPIS [35], a uni-semi-supervised method employing visible annotations for training, our SDCL still continues to maintain a leading position." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison with Supervised Methods. Our comparison further includes 14 well-known supervised methods for reference. The comparisons with these methodologies unequivocally illustrate that our SDCL framework outperforms several supervised methods, including AGW [57], MSO [12], and DDAG [54], and achieves a competitive performance compared with MCLNet[16]. These considerable gains benefit from the insightful design of a shallow-deep collaborative learning framework. Our method facilitates the cultivation of modality-invariant features through the synergistic collaboration of shallow and deep features. Guided by our innovative solutions, SDCL surpasses prevailing unsupervised methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16875" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.885, + 0.572 + ], + "angle": 0, + "content": "
MethodsVenueSYSU-MM01RegDB
All SearchIndoor SearchVisible to InfraredInfrared to Visible
r1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
SupervisedDDAG [54]ECCV-2054.7553.0239.6261.0267.9862.6169.3463.4649.2468.0661.8048.62
AGW [57]TPAMI-2147.5047.6535.3054.1762.9759.2370.0566.3750.1970.4965.9051.24
CA [56]ICCV-2169.8866.8953.6176.2680.3776.7985.0379.1465.3384.7577.8261.56
MPANet [47]CVPR-2170.5868.24-76.7480.95-82.880.7-83.780.9-
MSO [12]MM-2158.7056.42-63.0970.31-73.666.9-74.667.5-
AGM [26]MM-2169.6366.1152.2474.6878.3074.0088.4081.4568.5185.3481.1965.76
MCLNet [16]ICCV-2165.4061.9847.3972.5676.5872.1080.3173.0757.3975.9369.4952.63
SMCL [43]ICCV-2167.3961.78-68.8475.56-83.9379.83-83.0578.57-
FMCNet[60]CVPR-2266.3462.51-68.1574.09-89.1284.43-88.3883.86-
MAUM [27]CVPR-2271.6868.79-76.9781.94-87.8785.09-86.9584.34-
DEEN [61]CVPR-2374.771.8-80.383.3-91.185.1-89.583.4-
PMCM [34]IJCAI-2375.5471.16-81.5284.33-93.0989.57-91.4487.15-
PartMix [20]CVPR-2377.7874.62-81.5284.38-84.9382.52-85.6682.27-
SGIEL [20]CVPR-2377.1272.33-82.0782.95-95.3589.98-97.5791.41-
UnsupervisedSPCL [14]NIPS-2018.3719.3910.9926.8336.4233.0513.5914.8610.3611.7013.5610.09
MMT [13]ICLR-2021.4721.5311.5022.7931.5027.6625.6826.5119.5624.4225.5918.66
IICS [49]CVPR-2114.3915.748.4115.9124.8722.159.179.946.409.119.906.45
CAP [38]AAAI-2116.8215.717.0224.5730.7426.159.7111.568.7410.2111.347.92
Cluster Contrast [9]arXiv-2120.1622.0012.9723.3334.0130.8811.7613.889.9411.1412.998.99
ICE [3]ICCV-2120.5420.3910.2429.8138.3534.3212.9815.6411.9112.1814.8210.6
PPLR [6]CVPR-2211.9812.254.9712.7120.8117.6110.3011.948.1010.3911.237.04
OTLA [37]ECCV-2229.927.1-29.838.8-32.929.7-32.128.6-
H2H [24]TIP-2130.1529.40----23.8118.87----
ADCA [50]MM-2245.5142.7328.2950.6059.1155.1767.2064.0552.6768.4863.8149.62
DPIS [35]ICCV-2358.455.6-63.070.0-62.353.2-61.552.7-
CHCR [33]TCSVT-2359.4759.14----69.3164.74-69.9665.87-
DOTLA [5]MM-2350.3647.3632.4053.4761.7357.3585.6376.7161.5882.9174.9758.60
MBCCM [4]MM-2353.1448.1632.4155.2161.9857.1383.7977.8765.0482.8276.7461.73
PGM [48]CVPR-2357.2751.7834.9656.2362.7458.1369.4865.41-69.8565.17-
GUR [51]ICCV-2363.5161.6347.9371.1176.2372.5773.9170.2358.8875.0069.9456.21
SDCL (ours)-64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.574, + 0.885, + 0.588 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-arts on SYSU-MM01 and RegDB. Rank at \\( r \\) accuracy(%) , mAP (%) and mINP (%) are reported." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.602, + 0.889, + 0.72 + ], + "angle": 0, + "content": "
SYSU-MM01RegDB
ComponentsAll SearchIndoor SearchVisible to InfraredInfrared to Visible
IndexBaselineLids+LrsidCNLCRAr1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
1✓49.5548.7036.9253.8560.9657.0563.1658.1242.5961.7056.7641.65
2✓✓53.2150.1440.8060.1565.2464.0269.9763.3448.9868.5463.1948.06
3✓✓✓55.2754.7941.6164.3871.1567.4076.6067.9751.3875.6468.4650.38
4✓✓✓59.4359.1848.8167.4874.3170.3080.5373.6558.1379.1370.1853.30
5✓✓✓✓✓64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.722, + 0.85, + 0.735 + ], + "angle": 0, + "content": "Table 2. Ablation studies on the SYSU-MM01 and RegDB. Rank at \\( r \\) accuracy \\( \\left( \\% \\right) \\) ,mAP \\( \\left( \\% \\right) \\) and mINP \\( \\left( \\% \\right) \\) are reported." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.747, + 0.229, + 0.763 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.774, + 0.468, + 0.819 + ], + "angle": 0, + "content": "In this subsection, we meticulously conduct ablation experiments to validate the individual efficacy of each module. The results are reported in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.822, + 0.468, + 0.883 + ], + "angle": 0, + "content": "Baseline denotes the augmented dual-contrastive learning framework [50] with a dual-path transformer architecture. The optimization for the network only uses deep features, i.e., training the model with \\(\\mathcal{L}_{id}^{vd} + \\mathcal{L}_{id}^{rd}\\)." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.886, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Effectiveness of Deep and Shallow Contrastive Learn" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.747, + 0.892, + 0.838 + ], + "angle": 0, + "content": "ing. The incorporation of \\(\\mathcal{L}id^{vs} + \\mathcal{L}id^{rs}\\) yields a \\(4\\% - 7\\%\\) enhancement of rank-1 accuracy on SYSU-MM01 and RegDB datasets compared to the baseline. Despite shallow contrastive learning primarily optimizing the ReID model for feature learning within intra-modality, the findings underscore its capacity to augment cross-modality retrieval." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of CNL. Compared with the baseline, the CNL improves the performance of \\(6\\% - 10\\%\\) on SYSUMM01 and RegDB datasets. The main gain is achieved by the design of capturing credible homogeneous and hetero-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.268, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.09, + 0.447, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.218, + 0.47, + 0.26 + ], + "angle": 0, + "content": "Figure 3. Evaluation of hyper-parameters \\(\\lambda_{s}\\) (left) and \\(\\gamma\\) (right). The results are based on all search mode of SYSU-MM01 dataset. Rank-1 accuracy \\((\\%)\\) and mAP \\((\\%)\\) are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.47, + 0.332 + ], + "angle": 0, + "content": "genuine neighborhoods with the shallow and deep collaborative consistency, which is a dynamic optimization, reducing the impact of label noise in clustering and enhancing the robustness against multiple discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.469, + 0.426 + ], + "angle": 0, + "content": "Effectiveness of CRA. In contrast to the baseline, the CRA module remarkably improves the accuracy. CRA establishes a linkage between the labels of two modalities through collaborative shallow and deep ranking relations, constituting a global association across the training set and effectively bolstering modality-invariant features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Effectiveness of SDCL. The amalgamation of CNL and CRA notably enhances Rank-1 accuracy across diverse settings. This observation substantiates the rationale behind SDCL. CRA provides cross-modality supervision at the beginning of each training epoch in an off-line manner, while CNL dynamically learns cross-modality features during training iteration in an online manner. By combining off-line label association and online feature learning, SDCL learns better representation from different perspectives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.584, + 0.245, + 0.601 + ], + "angle": 0, + "content": "4.5. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Hyper-parameter Analysis for \\(\\lambda_{s}\\) and \\(\\gamma\\). We explore the influence of hyper-parameters \\(\\lambda_{s}\\) in Eq. 20 and \\(\\gamma\\) in Eq. 12, Eq. 13, Eq. 16, and Eq. 17, as shown in Fig. 3. The \\(\\lambda_{s}\\) governs the balance of \\(\\mathcal{L}_d\\) and \\(\\mathcal{L}_s\\) in deep and shallow neighbor learning. When \\(\\lambda_{s} = 0\\), the \\(\\mathcal{L}_s\\) is excluded. When \\(\\lambda_{base} = 0.5\\), the baseline method achieves a balance in deep and shallow neighbor learning. The \\(\\gamma\\) is used to control the positive neighbor selection range in CNL. Setting \\(\\gamma = 0\\) to 0 renders neighbor selection ineffective. When \\(\\gamma = 0.9\\), the CNL significantly reinforces the cross-modality learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Visualization. We perform feature space (t-SNE [36] map) and similarity distribution visualization for SDCL, as presented in Fig. 4. As evident in (a) and (b), SDCL brings infrared and visible positive sample points closer together, and effectively enhances the separation of cross-modality positive/negative distributions. This attests to the efficacy of our method in addressing modality discrepancies. Additionally, the visualization of shallow and deep feature map is presented in the supplementary materials." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.096, + 0.689, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.096, + 0.872, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.224, + 0.689, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.371, + 0.64, + 0.384 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.224, + 0.872, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.371, + 0.816, + 0.384 + ], + "angle": 0, + "content": "(b) SDCL" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Figure 4. The t-SNE (first row) and similarity distribution (second row) visualization of randomly selected identities. In the t-SNE visualization, identity is denoted by color, where circles represent the visible modality and triangles denote the infrared modality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.619, + 0.459 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.742 + ], + "angle": 0, + "content": "This paper investigates an extremely important and challenging problem, namely the unsupervised visible infrared person re-identification (US-VI-ReID) task, alleviating the reliance on expensive cross-modality annotation. To remedy the issues of multiple discrepancies, we propose a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer architecture, incorporating Collaborative Neighbor Learning (CNL), and Collaborative Ranking Association (CRA) module. SDCL significantly facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow patch embeddings and deep modality-shared features and capturing more discriminative representations for cross-modality retrieval. Experiments on two public benchmarks substantiate that our approach surpasses existing methodologies by a substantial margin. Moreover, it even outshines certain supervised counterparts, propelling VI-ReID toward real-world deployment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.659, + 0.772 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is partially supported by the National Natural Science Foundation of China under Grants (62071338, 62176188, 62272354, 62306215), the Key Research and Development Program of Hubei Province (2022BCA009,2022BAD175), the Special Fund of Hubei Luojia Laboratory (220100015), and the Interdisciplinary Innovative Talents Foundation from Renmin Hospital of Wuhan University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16877" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Yimin Liu, and Jianguo Jiang. Saliency and granularity: Discovering temporal coherence for video-based person re-identification. IEEE TCSVT, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[2] Cuiqun Chen, Mang Ye, Meibin Qi, and Bo Du. Sketchtrans: Disentangled prototype learning with transformer for sketchphoto recognition. IEEE TPAMI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[3] Hao Chen, Benoit Lagadec, and Francois Bremond. Ice: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[4] De Cheng, Lingfeng He, Nannan Wang, Shizhou Zhang, Zhen Wang, and Xinbo Gao. Efficient bilateral cross-modality cluster matching for unsupervised visible-infrared person reid. In ACM MM, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[5] De Cheng, Xiaojian Huang, Nannan Wang, Lingfeng He, Zhihui Li, and Xinbo Gao. Unsupervised visible-infrared person Reid by collaborative learning with neighbor-guided label refinement. ACM MM, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[6] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.468, + 0.45 + ], + "angle": 0, + "content": "[7] Zhenyu Cui, Jiahuan Zhou, and Yuxin Peng. Dma: Dual modality-aware alignment for visible-infrared person re-identification. IEEE TIFS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[8] Yongxing Dai, Jun Liu, Yan Bai, Zekun Tong, and Ling-Yu Duan. Dual-refinement: Joint label and feature refinement for unsupervised domain adaptive person re-identification. IEEE TIP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.468, + 0.562 + ], + "angle": 0, + "content": "[9] Zuozhuo Dai, Guangyuan Wang, Siyu Zhu, Weihao Yuan, and Ping Tan. Cluster contrast for unsupervised person re-identification. arxiv 2021. arXiv preprint arXiv:2103.11568, 2021. 2.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[10] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, 1996. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[11] Jiawei Feng, Ancong Wu, and Wei-Shi Zheng. Shape-erased feature learning for visible-infrared person re-identification. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[12] Yajun Gao, Tengfei Liang, Yi Jin, Xiaoyan Gu, Wu Liu, Yidong Li, and Congyan Lang. Mso: Multi-feature space joint optimization network for rgb-infrared person re-identification. In ACM MM, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[13] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual mean-teaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. arXiv preprint arXiv:2001.01526, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[14] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[15] Jianyang Gu, Weihua Chen, Hao Luo, Fan Wang, Hao Li, Wei Jiang, and Weijie Mao. Multi-view evolutionary training for unsupervised domain adaptive person re-identification. IEEE TIFS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[16] Xin Hao, Sanyuan Zhao, Mang Ye, and Jianbing Shen. Cross-modality person re-identification via modality confusion and center aggregation. In ICCV, 2021. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[18] Wenke Huang, Mang Ye, Zekun Shi, and Bo Du. Generalizable heterogeneous federated cross-correlation and instance similarity learning. IEEE TPAMI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[19] Wenke Huang, Mang Ye, Zekun Shi, Guancheng Wan, He Li, Bo Du, and Qiang Yang. A federated learning for generalization, robustness, fairness: A survey and benchmark. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[20] Minsu Kim, Seungryong Kim, Jungin Park, Seongheon Park, and Kwanghoon Sohn. Partmix: Regularization strategy to learn part discovery for visible-infrared person re-identification. In CVPR, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[21] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In AAAI, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[22] Shuang Li, Jiaxu Leng, Ji Gan, Mengjingcheng Mo, and Xinbo Gao. Shape-centered representation learning for visible-infrared person re-identification. arXiv preprint arXiv:2310.17952, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[23] Shuang Li, Fan Li, Jinxing Li, Huafeng Li, Bob Zhang, Dapeng Tao, and Xinbo Gao. Logical relation inference and multiview information interaction for domain adaptation person re-identification. IEEE TNNLS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[24] Wenqi Liang, Guangcong Wang, Jianhuang Lai, and Xiaohua Xie. Homogeneous-to-heterogeneous: Unsupervised learning for rgb-infrared person re-identification. IEEE TIP, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[25] Shan Lin, Chang-Tsun Li, and Alex C. Kot. Multidomain adversarial feature generalization for person re-identification. IEEE TIP, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[26] Haojie Liu, Daoxun Xia, and Wei Jiang. Towards homogeneous modality learning and multi-granularity information exploration for visible-infrared person re-identification. IEEE Journal of Selected Topics in Signal Processing. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[27] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[28] Yan Lu, Yue Wu, Bin Liu, Tianzhu Zhang, Baopu Li, Qi Chu, and Nenghai Yu. Cross-modality person re-identification with shared-specific feature transfer. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[29] Zefeng Lu, Ronghao Lin, and Haifeng Hu. Modality and camera factors bi-disentanglement for nir-vis object re-identification. IEEE TIFS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[30] Chuanchen Luo, Chunfeng Song, and Zhaoxiang Zhang. Learning to adapt across dual discrepancy for cross-domain person re-identification. IEEE TPAMI, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[31] Hao Luo, Pichao Wang, Yi Xu, Feng Ding, Yanxin Zhou, Fan Wang, Hao Li, and Rong Jin. Self-supervised pre-training for transformer-based person re-identification. arXiv preprint arXiv:2111.12084, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16878" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Zhiqi Pang, Chunyu Wang, Lingling Zhao, Yang Liu, and Gaurav Sharma. Cross-modality hierarchical clustering and refinement for unsupervised visible-infrared person re-identification. IEEE TCSVT, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[34] Zhihao Qian, Yutian Lin, and Bo Du. Visible-infrared person re-identification via patch-mixed cross-modality learning. *IJCAI*, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[35] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[37] Jiangming Wang, Zhizhong Zhang, Mingang Chen, Yi Zhang, Cong Wang, Bin Sheng, Yanyun Qu, and Yuan Xie. Optimal transport for label-efficient visible-infrared person re-identification. ECCV, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[38] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[39] Yiming Wang, Guanqiu Qi, Shuang Li, Yi Chai, and Huafeng Li. Body part-level domain alignment for domain-adaptive person re-identification with transformer framework. IEEE TIFS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[40] Yadi Wang, Hongyun Zhang, Duoqian Miao, and Witold Pedrycz. Multi-granularity re-ranking for visible-infrared person re-identification. CAAI Transactions on Intelligence Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[41] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.624 + ], + "angle": 0, + "content": "[42] Zheng Wang, Zhixiang Wang, Yinqiang Zheng, Yang Wu, Wenjun Zeng, and Shin'ichi Satoh. Beyond intra-modality: A survey of heterogeneous person re-identification. arXiv preprint arXiv:1905.10048, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.627, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[43] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Syncretic modality collaborative learning for visible infrared person re-identification. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.698 + ], + "angle": 0, + "content": "[44] Ancong Wu, Wei-Shi Zheng, and Jian-Huang Lai. Robust depth-based person re-identification. IEEE TIP, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[45] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In ICCV, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[46] Dongming Wu, Mang Ye, Gaojie Lin, Xin Gao, and Jianbing Shen. Person re-identification by context-aware part attention and multi-head collaborative learning. IEEE TIFS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[47] Qiong Wu, Pingyang Dai, Jie Chen, Chia-Wen Lin, Yongjian Wu, Feiyue Huang, Bineng Zhong, and Rongrong Ji. Discover cross-modality nuances for visible-infrared person re-identification. In CVPR, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[48] Zesen Wu and Mang Ye. Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning. In CVPR, 2023. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[50] Bin Yang, Mang Ye, Jun Chen, and Zesen Wu. Augmented dual-contrastive aggregation learning for unsupervised visible-infrared person re-identification. In ACM MM, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[51] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In ICCV, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[52] Bin Yang, Jun Chen, and Mang Ye. Top-k visual tokens transformer: Selecting tokens for visible-infrared person re-identification. In ICASSP, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[53] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE TIFS, pages 407–419, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[54] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In ECCV, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[55] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE TIFS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[56] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[57] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[58] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE TPAMI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[59] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[60] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[61] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[62] Yiyuan Zhang, Yuhao Kang, Sanyuan Zhao, and Jianbing Shen. Dual-semantic consistency learning for visible-infrared person re-identification. IEEE TIFS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[63] Yafei Zhang, Yongzeng Wang, Huafeng Li, and Shuang Li. Cross-compatible embedding and semantic consistent feature construction for sketch re-identification. In ACM MM, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.886 + ], + "angle": 0, + "content": "[64] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16879" + } + ] +] \ No newline at end of file diff --git a/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_origin.pdf b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..22ec896d95767f04be4ca43812e478603b2b0e02 --- /dev/null +++ b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/93501bbe-f504-4162-9926-acaa45b1fb43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:659390ea0c31e16d2df696d64220c82d92c46721587f825e8beb27d9bd73c646 +size 3315916 diff --git a/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/full.md b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3cab35d43e76a24bcd30afc943881cfd4e18b32b --- /dev/null +++ b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/full.md @@ -0,0 +1,406 @@ +# Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification + +Bin Yang Jun Chen* Mang Ye* + +National Engineering Research Center for Multimedia Software, + +School of Computer Science, Hubei Luojia Laboratory, Wuhan University, Wuhan, China + +https://github.com/yangbincv/SDCL + +# Abstract + +Unsupervised visible-infrared person re-identification (US-VI-ReID) centers on learning a cross-modality retrieval model without labels, reducing the reliance on expensive cross-modality manual annotation. Previous US-VI-ReID works gravitate toward learning cross-modality information with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. The shallow feature from the shallow layers contains nuanced detail information, which is critical for effective cross-modality learning but is disregarded regrettably by the existing methods. To address the above issues, we design a Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration which are harnessed for neighbor alignment, enhancing the robustness in a dynamic manner. Furthermore, CRA associates the cross-modality labels with the ranking association between shallow and deep features, furnishing valuable supervision for cross-modality learning. Extensive experiments validate the superiority of our method, even outperforming certain supervised counterparts. + +# 1. Introduction + +Person re-identification (ReID) aims at matching the same person image captured by non-overlapping cameras. In recent times, ReID has garnered significant attention from the computer vision research community, owing to its pivotal role in the context of intelligent video surveillance applications [2, 18, 19, 23, 25, 28-30, 39, 41, 44, 46, 57, + +![](images/8709010003d7ba753fffa03e46b6e56d6c0c2fa19cc2c7548567d31034bf2688.jpg) +Ground Truth Cross-modality Label Pairs +Deep Features +Figure 1. Illustration of the motivation. Previous advanced works for US-VI-ReID mainly focus on learning features and associating cross-modality labels with deep features. However, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in (a). In this paper, we incorporate shallow and deep features with collaborative learning and label association to remedy these problems, as shown in (b). + +![](images/3569712b422bf865fb2c61097d73afc29bde2ba49f540cd7f4911a0f32462460.jpg) +Shallow Features +Negative Optimization + +59, 62, 63]. The rank-1 accuracy has exhibited encouraging outcomes in supervised person ReID. Notably, these person images are typically acquired by visible cameras within well-illuminated environments. However, it is crucial to recognize that visible cameras often fail to capture sufficient information of individuals in low-light settings, consequently constraining the practicality of single-modality ReID in the context of night-time surveillance [1, 7, 15, 21, 40, 42, 45, 53, 61]. + +In contrast, recently deployed cameras possess the capability to seamlessly transition into far/near-infrared mode during nighttime operations [50, 55]. Various techniques [12, 20, 22, 27, 34, 47, 52, 54-58, 62] have been proposed for visible-infrared person re-identification (VI-ReID), resulting in commendable accuracy levels. Notably, these achievements are facilitated by extensive human-labeled cross-modality datasets. However, annotating cross-modality datasets demands more resources compared to the annotation of single-modality ReID datasets. This + +intricate annotation task presents a formidable and pivotal challenge, ultimately giving rise to the crucial task of unsupervised VI-ReID (US-VI-ReID) for reducing the cost of expensive cross-modality annotations. + +Some prior research endeavors [4, 5, 24, 33, 37, 48, 50, 51] have proposed preliminary solutions for US-VI-ReID. These solutions gravitate toward learning cross-modality feature with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, i.e., intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in Fig. 1. It is important to recognize that the shallow features originating from the modality-specific shallow layers harbor a wealth of nuanced detail information pertaining to pedestrian attributes, which is critical for unsupervised cross-modality representation learning but is disregarded regrettably by the existing methods. To harness the complete potential of the shallow information, we devise a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration as the constraints for seeking reliable intramodality and inter-modality neighbor learning, thereby guaranteeing the cultivation of modality-invariant and discriminative representations in a dynamic manner. Furthermore, CRA is developed with a global perspective to exploit cross-modality ranking consistency between deep and shallow features, associating the label of two modalities in a collaborative manner and furnishing valuable cross-modality supervision for cross-modality learning. Ultimately, within the shallow-deep collaborative learning framework with CNL and CRA modules, we acquire a robust representation, effectively mitigating the cross-modality discrepancy under unsupervised conditions. + +The main contributions can be summarized as follows: + +- We propose a shallow-deep collaborative learning framework based on the transformer architecture. This framework facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow and deep features. +- We propose a collaborative neighbor learning module to formulate dependable intra-modality and cross-modality neighbor learning, enabling the model to capture modality-invariant and discriminative features. +- We propose a collaborative ranking association module to explore intra-modality and cross-modality ranking consistencies, unifying the cross-modality labels and providing invaluable cross-modality supervision. +- Extensive experiments validate that our SDCL frame + +work surpasses existing methods on two mainstream VI-ReID benchmarks, consistently improving the unsupervised cross-modality retrieval performance. + +# 2. Related Work + +# 2.1. Supervised Visible-Infrared Person ReID + +Currently, advanced supervised VI-ReID focuses on bridging the gap between the two modalities and learning robust representations against modality discrepancy. Ye et al. [55] developed a dynamic tri-level relation mining (DTRM) framework simultaneously to explore the cross-modality relation cues of channel-level, part-level intra-modality, and graph-level. CAJ [56] proposed a channel-mixed learning to handle the intra-modality and cross-modality variations by randomly exchanging the color channels. SGIEL [11] proposed a shape-erased feature learning paradigm, jointly learning shape-related feature in one subspace and shape-erased features in the orthogonal complement. + +Nevertheless, the remarkable performances exhibited by these methods require extensive human-labeled cross-modality datasets. In this work, we pivot our focus to the realm of unsupervised visible-infrared person ReID, where the luxury of identity annotations is absent, presenting important applications for real-world VI-ReID deployments. + +# 2.2. Unsupervised Single-Modality Person ReID + +Unsupervised single-modality ReID endeavors to train the ReID model using unlabeled data captured by visible cameras. The majority of studies have embraced cluster algorithms to derive pseudo labels for optimizing the model. Cluster Contrast [9] utilized a distinctive cluster representation to delineate each cluster, addressing the issue of cluster inconsistency. Chen et al. [3] introduced Inter-instance Contrastive Encoding (ICE), which harnesses inter-instance pairwise similarity scores to enhance preceding class-level contrastive ReID methodologies. IICS [49] tackled the unsupervised ReID challenge by decomposing the similarity computation into two stages, namely, the intra-domain and inter-domain computations, respectively. Dai et al. [8] introduced a dual-refinement approach to concurrently enhance pseudo labels during the offline clustering phase and refine features during the online training phase, augmenting label purity and feature discriminability. + +The above methods are dedicated to addressing issues in single-modality ReID. When employed in the context of US-VI-ReID, these approaches encounter challenges stemming from cross-modality discrepancies and the absence of cross-modality (visible-infrared) identity labels. This impedes the learning of inter-modality feature and the generation of reliable cross-modality pseudo labels. + +# 2.3. Unsupervised Visible-Infrared Person ReID + +Several studies [24, 37, 48, 50] represent initial endeavors in the realm of US-VI-ReID. Yang et al. [50] introduced a novel Augmented Dual-Contrastive Aggregation (ADCA) learning framework founded on the principles of homogeneous joint learning and heterogeneous aggregation, establishing a robust baseline for purely unsupervised VI-ReID. PGM [48] developed a progressive graph matching method to systematically extract cross-modality correspondences, formulating correspondence mining as a graph-matching process. H2H [24] introduced a homogeneous-to-heterogeneous approach through two-stage learning, incorporating Market1501 [64] as additional labeled data. CHCR [33] proposed a cross-modality hierarchical clustering and refinement method by promoting modality-invariant feature learning and improving the reliability of pseudo-labels. + +Nevertheless, the above methods primarily concentrate on feature learning with deep features. However, interfered by the intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. Our approach integrates shallow and deep features through collaborative feature learning and cross-modality label association, which is distinguished from previous works, surpassing the performance of existing US-VI-ReID methods. + +# 3. Proposed Method + +The framework of shallow-deep collaborative learning (SDCL) is shown in Fig. 2, incorporating the Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. SDCL adopts a dual-path transformer architecture with dual-contrastive learning [50], which has two shallow modality-specific patch embedding layers and a modality-shared transformer. Instance memory and cluster memory are constructed for shallow embeddings and deep features within each modality. Instance memory stores all training image features. Cluster memory is built by averaging the instance features with the same pseudo labels. With the above memories, SDCL develops the CNL module to exploit homogeneous and heterogeneous collaboration as the constraints for seeking reliable intra-modality and inter-modality neighbor learning. CRA explores the ranking consistency of cross-modality shallow embeddings and deep features to associate reliable cross-modality identities. With the collaboration of shallow embeddings and deep features within CNL and CRA modules, SDCL learns better representation for cross-modality retrieval. + +# 3.1. Shallow-deep Contrastive Learning + +We first introduce shallow-deep contrastive learning based on augmented dual-contrast learning [50] with a dual-path + +transformer architecture. + +Cluster Memory Initialization. At the beginning of each training epoch, we construct shallow and deep cluster memories for each modality by averaging the shallow and deep feature of one cluster, which can be denoted as: + +$$ +\phi_ {k} ^ {v s} = \frac {1}{| \mathcal {H} _ {k} ^ {v} |} \sum_ {u _ {n} ^ {v s} \in y _ {k} ^ {v}} u _ {n} ^ {v s}, \tag {1} +$$ + +$$ +\phi_ {k} ^ {v d} = \frac {1}{\left| \mathcal {H} _ {k} ^ {v} \right|} \sum_ {u _ {n} ^ {v d} \in y _ {k} ^ {v}} u _ {n} ^ {v d}, \tag {2} +$$ + +$$ +\phi_ {l} ^ {r s} = \frac {1}{| \mathcal {H} _ {l} ^ {r} |} \sum_ {u _ {n} ^ {r s} \in y _ {l} ^ {r}} u _ {n} ^ {r s}, \tag {3} +$$ + +$$ +\phi_ {l} ^ {r d} = \frac {1}{| \mathcal {H} _ {l} ^ {r} |} \sum_ {u _ {n} ^ {r d} \in y _ {l} ^ {r}} u _ {n} ^ {r d}, \tag {4} +$$ + +where $u_{n}^{r*}$ and $u_{n}^{v*}$ are infrared and visible instance features, respectively. $u_{n}^{*s}$ and $u_{n}^{*d}$ are shallow and deep instance features, respectively. $\mathcal{H}_k$ indicates the $k$ -th cluster set and $|\cdot|$ counts the number of samples of a set. Meanwhile, we store the instance features in shallow and deep infrared and visible instance memories $U^{rs}$ , $U^{rd}$ , $U^{vs}$ and $U^{vd}$ . + +Cluster Memory Updating. During each training iteration, the deep and shallow cluster memories of two modalities by a momentum updating strategy: + +$$ +\phi_ {k} ^ {(\delta)} \leftarrow \alpha \phi_ {k} ^ {(\delta - 1)} + (1 - \alpha) q, q \in y _ {k}, \tag {5} +$$ + +where $q$ is the query features sampled from training set in each training iteration, respectively. $\alpha$ is the momentum factor. $\delta$ is the training iteration number. + +Contrastive Loss. Given shallow and deep visible and infrared query $q^{vs}$ , $q^{vd}$ , $q^{rs}$ , and $q^{rs}$ , we compute the contrastive loss for visible modality by the following equations: + +$$ +\mathcal {L} _ {i d} ^ {v s} = - \log \frac {\exp \left(q ^ {v s} \cdot \phi_ {+} ^ {v s} / \tau\right)}{\sum_ {k = 0} ^ {K} \exp \left(q ^ {v s} \cdot \phi_ {k} ^ {v s} / \tau\right)}, \tag {6} +$$ + +$$ +\mathcal {L} _ {i d} ^ {v d} = - \log \frac {\exp \left(q ^ {v d} \cdot \phi_ {+} ^ {v d} / \tau\right)}{\sum_ {k = 0} ^ {K} \exp \left(q ^ {v d} \cdot \phi_ {k} ^ {v d} / \tau\right)}, \tag {7} +$$ + +where $\phi_{+}$ is the positive memory corresponding to the pseudo label of $q$ and the $\tau$ is a temperature. The infrared contrastive loss $\mathcal{L}_{id}^{rs}$ and $\mathcal{L}_{id}^{rd}$ are obtained by the same way. + +We optimize the modality-specific shallow embedding layers and deep modality-shared layers by combining deep and shallow contrastive loss: + +$$ +\mathcal {L} _ {i d} = \mathcal {L} _ {i d} ^ {v d} + \mathcal {L} _ {i d} ^ {r d} + \mathcal {L} _ {i d} ^ {v s} + \mathcal {L} _ {i d} ^ {r s} \tag {8} +$$ + +![](images/ef5faaca97f99d3d306196012079f3290a78ebf352900d4c64a9899b5c521150.jpg) +Figure 2. Illustration of shallow-deep collaborative learning. It comprises shallow-deep contrastive learning, collaborative neighbor learning, and collaborative ranking association. SDCL finds reliable neighborhood and cross-modality labels with shallow-deep collaborations, enhancing the robustness of learned representation. + +# 3.2. Collaborative Neighbor Learning + +Our basic rationale for the CNL module is that deep and shallow features are complementary, and their intrinsic consistency relationship can be used to constrain intra-modality and inter-modality optimization, formulating collaborative learning to enhance the robustness against intra-modality and inter-modality variations. From the above perspective, CNL explores essential homogeneous and heterogeneous shallow-deep collaborations to seek reliable intra-modality and inter-modality neighbor learning. + +Homogeneous and Heterogeneous Collaboration. Given a query $q$ , we can get the similarity $s(q_i, u_j)$ between the query and each instance in the training set by: + +$$ +s \left(q _ {i}, u _ {j}\right) = \frac {q _ {i} \cdot u _ {j}}{\left| \left| q _ {i} \right| \right| _ {2} \left| \left| u _ {j} \right| \right| _ {2}}, \tag {9} +$$ + +where $q_{i}$ and $u_{j}$ come from shallow or deep features of visible or infrared modality to get multiple types of homogeneous and heterogeneous shallow or deep query-instance similarity $s(q_{i}^{vs},u_{j}^{vs})$ , $s(q_{i}^{rs},u_{j}^{rs})$ , $s(q_{i}^{vs},u_{j}^{rs})$ , $s(q_{i}^{rs},u_{j}^{vs})$ , $s(q_{i}^{vd},u_{j}^{vd})$ , $s(q_{i}^{rd},u_{j}^{rd})$ , $s(q_{i}^{vd},u_{j}^{rd})$ , and $s(q_{i}^{rd},u_{j}^{vd})$ . + +With the above multiple types of similarities, we seek reliable shallow and deep intra-modality and inter-modality + +neighbors, formulating collaborative neighbor learning. The intra-modality neighbors can be defined as follows: + +$$ +\mathcal {N} ^ {v} \left(q _ {i} ^ {v}\right) = \left\{\mathcal {N} ^ {v} \left(q _ {i} ^ {v s}\right) \cap \mathcal {N} ^ {v} \left(q _ {i} ^ {v d}\right) \right\}, \tag {10} +$$ + +$$ +\mathcal {N} ^ {r} \left(q _ {i} ^ {r}\right) = \left\{\mathcal {N} ^ {r} \left(q _ {i} ^ {r s}\right) \cap \mathcal {N} ^ {r} \left(q _ {i} ^ {r d}\right) \right\}, \tag {11} +$$ + +where $\mathcal{N}^v(q_i^{v*})$ and $\mathcal{N}^r(q_i^{r*})$ are neighborhood sets searched by visible-visible and infrared-infrared similarity, respectively. $* \in \{s, d\}$ represents the shallow and deep feature. The $\mathcal{N}^v(q_i^{vs})$ and $\mathcal{N}^r(q_i^{vd})$ can be obtained by: + +$$ +\mathcal {N} ^ {v} \left(q _ {i} ^ {v s}\right) = \left\{u _ {j} ^ {v s} \mid s \left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\right) > \gamma \cdot \max _ {j = 1 \dots N ^ {v}} s \left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\right) \right\}, \tag {12} +$$ + +$$ +\mathcal {N} ^ {v} \left(q _ {i} ^ {v d}\right) = \left\{u _ {j} ^ {v d} \mid s \left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\right) > \gamma \cdot \max _ {j = 1 \dots N v} s \left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\right) \right\}, \tag {13} +$$ + +where the $N^v$ denotes the number of visible instances and the $\gamma$ is the positive neighbor selection range. The $\mathcal{N}^r (q_i^{rs})$ and $\mathcal{N}^r (q_i^{rd})$ can be calculated by similar manner. + +Given query $q_{i}^{v}$ or $q_{i}^{r}$ , the inter-modality neighbors can be denoted as follows: + +$$ +\mathcal {N} ^ {r} \left(q _ {i} ^ {v}\right) = \left\{\mathcal {N} ^ {r} \left(q _ {i} ^ {v s}\right) \cap \mathcal {N} ^ {r} \left(q _ {i} ^ {v d}\right) \right\}, \tag {14} +$$ + +$$ +\mathcal {N} ^ {v} \left(q _ {i} ^ {r}\right) = \left\{\mathcal {N} ^ {v} \left(q _ {i} ^ {r s}\right) \cap \mathcal {N} ^ {v} \left(q _ {i} ^ {r d}\right) \right\}, \tag {15} +$$ + +where $\mathcal{N}^r (q_i^{vs})$ and $\mathcal{N}^r (q_i^{vd})$ are defined as follows: + +$$ +\mathcal {N} ^ {r} \left(q _ {i} ^ {v s}\right) = \left\{u _ {j} ^ {r s} \mid s \left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\right) > \gamma \cdot \max _ {j = 1 \dots N ^ {r}} s \left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\right) \right\}, \tag {16} +$$ + +$$ +\mathcal {N} ^ {r} \left(q _ {i} ^ {v d}\right) = \left\{u _ {j} ^ {r d} \mid s \left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\right) > \gamma \cdot \max _ {j = 1 \dots N ^ {r}} s \left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\right) \right\}, \tag {17} +$$ + +where the $N^r$ denotes the number of infrared instances. Similarly, we can get the $\mathcal{N}^v(q_i^{rs})$ and $\mathcal{N}^v(q_i^{rd})$ . With the above procedure, we get the $\mathcal{N}^v(q_i^v), \mathcal{N}^r(q_i^r), \mathcal{N}^r(q_i^v)$ and $\mathcal{N}^v(q_i^r)$ , which are the reliable neighborhood set searched by the constraints of homogeneous or heterogeneous shallow-deep collaborations. + +Collaborative Neighbor Learning. With the neighborhood set $\mathcal{N}^v(q_i^v)$ , $\mathcal{N}^r(q_i^r)$ , $\mathcal{N}^r(q_i^v)$ and $\mathcal{N}^v(q_i^r)$ , we can perform neighbor learning. Given query $q_i^{vs}$ and $q_i^{vs}$ , we can obtain the expression of the visible-visible shallow and deep neighbor learning by: + +$$ +\mathcal {L} _ {s} ^ {v v} = - \frac {1}{N ^ {b}} \sum_ {i = 1} ^ {N ^ {b}} \sum_ {j \in \mathcal {N} ^ {v} \left(q _ {i} ^ {v}\right)} \log \frac {\exp \left(s \left(q _ {i} ^ {v s} , u _ {j} ^ {v s}\right) / \tau\right)}{\sum_ {n = 1} ^ {N ^ {v}} \exp \left(s \left(q _ {i} ^ {v s} , u _ {n} ^ {v s}\right) / \tau\right)}, \tag {18} +$$ + +$$ +\mathcal {L} _ {d} ^ {v v} = - \frac {1}{N ^ {b}} \sum_ {i = 1} ^ {N ^ {b}} \sum_ {j \in \mathcal {N} ^ {v} \left(q _ {i} ^ {v}\right)} \log \frac {\exp \left(s \left(q _ {i} ^ {v d} , u _ {j} ^ {v d}\right) / \tau\right)}{\sum_ {n = 1} ^ {N ^ {v}} \exp \left(s \left(q _ {i} ^ {v d} , u _ {n} ^ {v d}\right) / \tau\right)}, \tag {19} +$$ + +$$ +\mathcal {L} ^ {v v} = \mathcal {L} _ {d} ^ {v v} + \lambda_ {s} \mathcal {L} _ {s} ^ {v v}, \tag {20} +$$ + +where $N^b$ is the batch size of query $q_{i}$ . Similarly, the neighbor learning for infrared-infrared $\mathcal{L}^{rr}$ , infrared-visible $\mathcal{L}^{rv}$ and visible-infrared $\mathcal{L}^{vr}$ can be obtained by similar ways. The final optimization for neighbor learning is denoted by the following combination: + +$$ +\mathcal {L} _ {\text {n e i g h b o r}} = \mathcal {L} ^ {v v} + \mathcal {L} ^ {r r} + \mathcal {L} ^ {r v} + \mathcal {L} ^ {v r}. \tag {21} +$$ + +The overall loss for SDCL is denoted as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {n e i g h b o r}} + \mathcal {L} _ {i d} \tag {22} +$$ + +Discussion. In contrast to neighborhood learning in [30], which exclusively relies on deep features for the selection of positive neighbors. Under the multiple discrepancies, the similarity of individual deep features is unreliable, resulting in wrong neighbor selection and negative optimization. Our method embraces collaborative neighbor learning, which dynamically refines cross-modality supervision through shallow and deep collaboration, achieving effective cross-modality feature alignment. + +# 3.3. Collaborative Ranking Association + +The CRA module is underpinned by dual rationales: (1) Shallow and deep features across the two modalities should have the same ranking consistency. (2) The ranking consistencies of shallow and deep features within one modality are more accurate. Accordingly, the CRA module contains two processes, i.e., inter-modality ranking association and intra-modality ranking smoothing. + +Inter-modality Ranking Association. With the similarities $s(u_{i}^{vs},u_{j}^{rs})$ and $s(u_{i}^{vd},u_{j}^{rd})$ , we can get two visible-infrared ranking lists $\{s(u_i^{vs},u_j^{rs}),j\in [1,N^r ]\}$ and $\{s(u_i^{vd},u_j^{rd}),j\in [1,N^r ]\}$ for $u_{i}^{vs}$ and $u_{i}^{vd}$ , which are denoted as $\mathcal{R}_s^{vr}(u_i^{vs})$ and $\mathcal{R}_d^{vr}(u_i^{vd})$ . The label of $k$ -th similar infrared instance in two ranking lists can be represented as: + +$$ +\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r s}}, u _ {j} ^ {r s} = \mathcal {R} _ {s} ^ {v r} \left(u _ {i} ^ {v s}\right) [ k ], \tag {23} +$$ + +$$ +\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r d}}, u _ {j} ^ {r d} = \mathcal {R} _ {d} ^ {v r} \left(u _ {i} ^ {v d}\right) [ k ], \tag {24} +$$ + +where $\tilde{y}_{u_i^{vs}}^{vr}[k]$ and $\tilde{y}_{u_i^{vd}}^{vr}[k]$ are refined $k$ -th cross-modality labels of $u_i^{vs}$ and $u_i^{vd}$ through the ranking on visible-infrared shallow and deep similarities. Based on the rationale (1), we propose to associate the cross-modality labels with the intersection of two label sets to investigate collaborative ranking consistency: + +$$ +I _ {u _ {i} ^ {v}} (k) = \left\{\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] \cap \left\{\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ n ] \right\} _ {n = 1} ^ {N} \right\}, \tag {25} +$$ + +where the $I_{u_i^v}(k)$ records the samples with the same identity of top- $k$ identity in instances from top-1 to top- $N$ . The reliable refined cross-modality label of $u_i^v$ can be expressed by the label of maximum count in $I_{u_i^v}(k)$ : + +$$ +\tilde {y} _ {u _ {i} ^ {v}} ^ {c m} = \tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ], k = \underset {k} {\operatorname {a r g m a x}} \left(\left\{\left| I _ {u _ {i} ^ {v}} (k) \right|, k \in [ 1, K ] \right\}\right), \tag {26} +$$ + +where $|\cdot|$ denotes the counting function. The argmax operation traverses $I_{u_i^v}(k)$ and finds the labels with maximum number as cross-modality refined labels. Then, we convert the label list $\{\tilde{y}_{u_i^v}^{cm}, i \in [1, N^v]\}$ to the form of one-hot code matrix $\tilde{Y}^{cm} \in \mathbb{R}^{N^v \times C^r}$ by setting the column according to the refine labels to 1 and the rest to 0, where $C^r$ is the class number of infrared modality. + +Intra-modality Collaborative Smoothing. Based on the rationale (2), two shallow and deep homogeneous shallow $P_{s}^{vv}$ and deep $P_{d}^{vv}$ similarity matrices are constructed to investigate intra-modality ranking consistency, enhancing the precision of refined cross-modality pseudo labels by: + +$$ +P _ {s} ^ {v v} (i, j) = s \left(u _ {i} ^ {v s}, u _ {j} ^ {v s}\right), \tag {27} +$$ + +$$ +P _ {d} ^ {v v} (i, j) = s \left(u _ {i} ^ {v d}, u _ {j} ^ {v d}\right), \tag {28} +$$ + +where $P_{*}^{vv}\in \mathbb{R}^{N^{v}\times N^{v}}$ represents the intra-modality similarity structure. In order to explore the intra-modality shallow and deep collaboration, we calculate the sum of $P_{s}^{vv}(i,j)$ and $P_d^{vv}(i,j)$ by: + +$$ +P ^ {v v} (i, j) = P _ {s} ^ {v v} (i, j) + P _ {d} ^ {v v} (i, j) \tag {29} +$$ + +where $P^{vv}$ indicates the consistency of shallow and deep similarity matrix. We keep the 5-max values of $P^{vv}$ in each row to 1 and the rest to 0, acquiring the ranking relations. The process of intra-modality ranking smoothing is formulated as follows: + +$$ +Y ^ {c m} = P ^ {v v} \tilde {Y} ^ {c m}, \tag {30} +$$ + +where $Y^{cm} \in \mathbb{R}^{N^v \times C^r}$ is the final refined cross-modality label matrix of visible instance. In $Y^{cm}$ , the column number of the maximum value in each row is the refined label of samples. In our work, we refine the pseudo labels of the visible modality to the infrared modality and smooth the cross-modality labels with the shallow and deep similarity within the visible modality. Then, the infrared and refined visible labels will be used to construct a modality-shared memory for contrastive learning within two modalities [51]. Discussion. Our CRA is essentially the collaboration of shallow and deep ranking processes with a global perspective at the beginning of each training epoch, which is distinguished from previous works employing a solitary-pattern deep feature similarity [37, 48, 50, 51]. The imposition of a collaborative consistency constraint on dual-pattern similarities within CRA confers a distinct advantage, facilitating the exploration of more dependable cross-modality relations and providing valuable cross-modality supervision. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Protocol + +Datasets. We evaluate the proposed SDCL framework on two widely-used visible-infrared person ReID datasets, i.e., SYSU-MM01 [45] and RegDB [32]. SYSU-MM01 dataset is collected by 6 different cameras, including 22258 visible and 11909 near-infrared images of 395 training identities. We perform ten trials of the gallery set selection [53] and calculate the average performance following existing methods with all-search and indoor-search testing mode. RegDB dataset is captured by two aligned visible and thermal camera system, which has less challenges for VI-ReID. 206 identities with 2,060 images are selected for training and this procedure is repeated 10 times following [54], and calculate the average performance with visible to thermal and thermal to visible testing mode. + +Evaluation Protocols. Following existing works, Cumulative Matching Characteristics (CMC), mean Average Precision (mAP) and mean inverse negative penalty (mINP) [57] are calculated as the evaluation metrics. + +# 4.2. Implementation Details + +The proposed framework is implemented with PyTorch. SDCL incorporates the feature extractor from TransReID [17] as the backbone network with augmented dual-contrastive learning [50]. The shallow patch embedding layers are constructed with IBN and CNN module [31]. The concatenation of shallow and deep class tokens is used to calculate the cosine similarity for retrieval. DBSCAN [10] is conducted to generate pseudo labels. The visible and infrared images are resized to $288 \times 144$ before entering the network. We sample 8 pseudo identities and 16 instances for each pseudo identity for each modality within one batch. Channel augmentation [56], random crop, horizontal flipping, and random erasing are adopted for data augmentation. We adopt the SGD optimizer with the initial learning rate of $3.5e - 4$ to train the model. The model is trained in a total of 50 epochs. The CRA module is added in the last 20 epochs. The $\lambda_{s}$ in Eq. 20 is set to 0.5. The $\gamma$ in Eq. 12, Eq. 13, Eq. 16, and Eq. 17 is set to 0.9. The $N$ and $K$ in Eq. 25 and Eq. 26 are set to 20 for SYSU-MM01 and 10 for RegDB, respectively. The contrastive learning settings follow [50]. + +# 4.3. Comparison with State-of-the-art Methods + +In Table 1, our SDCL is compared with supervised and unsupervised VI-ReID methods on two benchmarks including SYSU-MM01 and RegDB. + +Comparison with Unsupervised Methods. As reported in Table 1, our method exhibits superior performance compared to the current advanced unsupervised methods. More precisely, our SDCL attains a remarkable $64.49\%$ and $86.91\%$ rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. Compared with the best current method GUR [51], our method also exceeds the about $1\%$ and $10\%$ rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. In comparison to DPIS [35], a uni-semi-supervised method employing visible annotations for training, our SDCL still continues to maintain a leading position. + +Comparison with Supervised Methods. Our comparison further includes 14 well-known supervised methods for reference. The comparisons with these methodologies unequivocally illustrate that our SDCL framework outperforms several supervised methods, including AGW [57], MSO [12], and DDAG [54], and achieves a competitive performance compared with MCLNet[16]. These considerable gains benefit from the insightful design of a shallow-deep collaborative learning framework. Our method facilitates the cultivation of modality-invariant features through the synergistic collaboration of shallow and deep features. Guided by our innovative solutions, SDCL surpasses prevailing unsupervised methods. + +
MethodsVenueSYSU-MM01RegDB
All SearchIndoor SearchVisible to InfraredInfrared to Visible
r1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
SupervisedDDAG [54]ECCV-2054.7553.0239.6261.0267.9862.6169.3463.4649.2468.0661.8048.62
AGW [57]TPAMI-2147.5047.6535.3054.1762.9759.2370.0566.3750.1970.4965.9051.24
CA [56]ICCV-2169.8866.8953.6176.2680.3776.7985.0379.1465.3384.7577.8261.56
MPANet [47]CVPR-2170.5868.24-76.7480.95-82.880.7-83.780.9-
MSO [12]MM-2158.7056.42-63.0970.31-73.666.9-74.667.5-
AGM [26]MM-2169.6366.1152.2474.6878.3074.0088.4081.4568.5185.3481.1965.76
MCLNet [16]ICCV-2165.4061.9847.3972.5676.5872.1080.3173.0757.3975.9369.4952.63
SMCL [43]ICCV-2167.3961.78-68.8475.56-83.9379.83-83.0578.57-
FMCNet[60]CVPR-2266.3462.51-68.1574.09-89.1284.43-88.3883.86-
MAUM [27]CVPR-2271.6868.79-76.9781.94-87.8785.09-86.9584.34-
DEEN [61]CVPR-2374.771.8-80.383.3-91.185.1-89.583.4-
PMCM [34]IJCAI-2375.5471.16-81.5284.33-93.0989.57-91.4487.15-
PartMix [20]CVPR-2377.7874.62-81.5284.38-84.9382.52-85.6682.27-
SGIEL [20]CVPR-2377.1272.33-82.0782.95-95.3589.98-97.5791.41-
UnsupervisedSPCL [14]NIPS-2018.3719.3910.9926.8336.4233.0513.5914.8610.3611.7013.5610.09
MMT [13]ICLR-2021.4721.5311.5022.7931.5027.6625.6826.5119.5624.4225.5918.66
IICS [49]CVPR-2114.3915.748.4115.9124.8722.159.179.946.409.119.906.45
CAP [38]AAAI-2116.8215.717.0224.5730.7426.159.7111.568.7410.2111.347.92
Cluster Contrast [9]arXiv-2120.1622.0012.9723.3334.0130.8811.7613.889.9411.1412.998.99
ICE [3]ICCV-2120.5420.3910.2429.8138.3534.3212.9815.6411.9112.1814.8210.6
PPLR [6]CVPR-2211.9812.254.9712.7120.8117.6110.3011.948.1010.3911.237.04
OTLA [37]ECCV-2229.927.1-29.838.8-32.929.7-32.128.6-
H2H [24]TIP-2130.1529.40----23.8118.87----
ADCA [50]MM-2245.5142.7328.2950.6059.1155.1767.2064.0552.6768.4863.8149.62
DPIS [35]ICCV-2358.455.6-63.070.0-62.353.2-61.552.7-
CHCR [33]TCSVT-2359.4759.14----69.3164.74-69.9665.87-
DOTLA [5]MM-2350.3647.3632.4053.4761.7357.3585.6376.7161.5882.9174.9758.60
MBCCM [4]MM-2353.1448.1632.4155.2161.9857.1383.7977.8765.0482.8276.7461.73
PGM [48]CVPR-2357.2751.7834.9656.2362.7458.1369.4865.41-69.8565.17-
GUR [51]ICCV-2363.5161.6347.9371.1176.2372.5773.9170.2358.8875.0069.9456.21
SDCL (ours)-64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
+ +Table 1. Comparison with state-of-the-arts on SYSU-MM01 and RegDB. Rank at $r$ accuracy(%) , mAP (%) and mINP (%) are reported. + +
SYSU-MM01RegDB
ComponentsAll SearchIndoor SearchVisible to InfraredInfrared to Visible
IndexBaselineLids+LrsidCNLCRAr1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
1✓49.5548.7036.9253.8560.9657.0563.1658.1242.5961.7056.7641.65
2✓✓53.2150.1440.8060.1565.2464.0269.9763.3448.9868.5463.1948.06
3✓✓✓55.2754.7941.6164.3871.1567.4076.6067.9751.3875.6468.4650.38
4✓✓✓59.4359.1848.8167.4874.3170.3080.5373.6558.1379.1370.1853.30
5✓✓✓✓✓64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
+ +Table 2. Ablation studies on the SYSU-MM01 and RegDB. Rank at $r$ accuracy $\left( \% \right)$ ,mAP $\left( \% \right)$ and mINP $\left( \% \right)$ are reported. + +# 4.4. Ablation Study + +In this subsection, we meticulously conduct ablation experiments to validate the individual efficacy of each module. The results are reported in Table 2. + +Baseline denotes the augmented dual-contrastive learning framework [50] with a dual-path transformer architecture. The optimization for the network only uses deep features, i.e., training the model with $\mathcal{L}_{id}^{vd} + \mathcal{L}_{id}^{rd}$ . + +Effectiveness of Deep and Shallow Contrastive Learn + +ing. The incorporation of $\mathcal{L}id^{vs} + \mathcal{L}id^{rs}$ yields a $4\% - 7\%$ enhancement of rank-1 accuracy on SYSU-MM01 and RegDB datasets compared to the baseline. Despite shallow contrastive learning primarily optimizing the ReID model for feature learning within intra-modality, the findings underscore its capacity to augment cross-modality retrieval. + +Effectiveness of CNL. Compared with the baseline, the CNL improves the performance of $6\% - 10\%$ on SYSUMM01 and RegDB datasets. The main gain is achieved by the design of capturing credible homogeneous and hetero- + +![](images/82d50b90cfd8f8318601d1b8f5377f8b1e5bd9e44391585d0b96c9399882f90d.jpg) +Figure 3. Evaluation of hyper-parameters $\lambda_{s}$ (left) and $\gamma$ (right). The results are based on all search mode of SYSU-MM01 dataset. Rank-1 accuracy $(\%)$ and mAP $(\%)$ are reported. + +![](images/52ce65ba4319aeb4e93719fb1ad2c644f447ab1e7deb96e68e5ad28b9cf3fa93.jpg) + +genuine neighborhoods with the shallow and deep collaborative consistency, which is a dynamic optimization, reducing the impact of label noise in clustering and enhancing the robustness against multiple discrepancies. + +Effectiveness of CRA. In contrast to the baseline, the CRA module remarkably improves the accuracy. CRA establishes a linkage between the labels of two modalities through collaborative shallow and deep ranking relations, constituting a global association across the training set and effectively bolstering modality-invariant features. + +Effectiveness of SDCL. The amalgamation of CNL and CRA notably enhances Rank-1 accuracy across diverse settings. This observation substantiates the rationale behind SDCL. CRA provides cross-modality supervision at the beginning of each training epoch in an off-line manner, while CNL dynamically learns cross-modality features during training iteration in an online manner. By combining off-line label association and online feature learning, SDCL learns better representation from different perspectives. + +# 4.5. Further Analysis + +Hyper-parameter Analysis for $\lambda_{s}$ and $\gamma$ . We explore the influence of hyper-parameters $\lambda_{s}$ in Eq. 20 and $\gamma$ in Eq. 12, Eq. 13, Eq. 16, and Eq. 17, as shown in Fig. 3. The $\lambda_{s}$ governs the balance of $\mathcal{L}_d$ and $\mathcal{L}_s$ in deep and shallow neighbor learning. When $\lambda_{s} = 0$ , the $\mathcal{L}_s$ is excluded. When $\lambda_{base} = 0.5$ , the baseline method achieves a balance in deep and shallow neighbor learning. The $\gamma$ is used to control the positive neighbor selection range in CNL. Setting $\gamma = 0$ to 0 renders neighbor selection ineffective. When $\gamma = 0.9$ , the CNL significantly reinforces the cross-modality learning. + +Visualization. We perform feature space (t-SNE [36] map) and similarity distribution visualization for SDCL, as presented in Fig. 4. As evident in (a) and (b), SDCL brings infrared and visible positive sample points closer together, and effectively enhances the separation of cross-modality positive/negative distributions. This attests to the efficacy of our method in addressing modality discrepancies. Additionally, the visualization of shallow and deep feature map is presented in the supplementary materials. + +![](images/48281d164bbfefded1b4a737da80543622f93a402eff34ad87560e760d6737e7.jpg) + +![](images/fa6d16b9d00ac1e85bac80572a898d0ba3c893bf1f69abae2e8065d565c8aef3.jpg) + +![](images/5b9a04ad59ffbb44d8858199991ff37f5c20b0e3b6fa51c0fc973bc7a688b179.jpg) +(a) Baseline +Figure 4. The t-SNE (first row) and similarity distribution (second row) visualization of randomly selected identities. In the t-SNE visualization, identity is denoted by color, where circles represent the visible modality and triangles denote the infrared modality. + +![](images/6e5168ecd89e06924d0228c49e44d4b4172a7de0934bb50e1add316e4a3f8fb7.jpg) +(b) SDCL + +# 5. Conclusion + +This paper investigates an extremely important and challenging problem, namely the unsupervised visible infrared person re-identification (US-VI-ReID) task, alleviating the reliance on expensive cross-modality annotation. To remedy the issues of multiple discrepancies, we propose a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer architecture, incorporating Collaborative Neighbor Learning (CNL), and Collaborative Ranking Association (CRA) module. SDCL significantly facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow patch embeddings and deep modality-shared features and capturing more discriminative representations for cross-modality retrieval. Experiments on two public benchmarks substantiate that our approach surpasses existing methodologies by a substantial margin. Moreover, it even outshines certain supervised counterparts, propelling VI-ReID toward real-world deployment. + +# Acknowledgments + +This work is partially supported by the National Natural Science Foundation of China under Grants (62071338, 62176188, 62272354, 62306215), the Key Research and Development Program of Hubei Province (2022BCA009,2022BAD175), the Special Fund of Hubei Luojia Laboratory (220100015), and the Interdisciplinary Innovative Talents Foundation from Renmin Hospital of Wuhan University. + +# References + +[1] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Yimin Liu, and Jianguo Jiang. Saliency and granularity: Discovering temporal coherence for video-based person re-identification. IEEE TCSVT, 2022. 1 +[2] Cuiqun Chen, Mang Ye, Meibin Qi, and Bo Du. Sketchtrans: Disentangled prototype learning with transformer for sketchphoto recognition. IEEE TPAMI, 2023. 1 +[3] Hao Chen, Benoit Lagadec, and Francois Bremond. Ice: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, 2021. 2, 7 +[4] De Cheng, Lingfeng He, Nannan Wang, Shizhou Zhang, Zhen Wang, and Xinbo Gao. Efficient bilateral cross-modality cluster matching for unsupervised visible-infrared person reid. In ACM MM, 2023. 2, 7 +[5] De Cheng, Xiaojian Huang, Nannan Wang, Lingfeng He, Zhihui Li, and Xinbo Gao. Unsupervised visible-infrared person Reid by collaborative learning with neighbor-guided label refinement. ACM MM, 2023. 2, 7 +[6] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, 2022. 7 +[7] Zhenyu Cui, Jiahuan Zhou, and Yuxin Peng. Dma: Dual modality-aware alignment for visible-infrared person re-identification. IEEE TIFS, 2024. 1 +[8] Yongxing Dai, Jun Liu, Yan Bai, Zekun Tong, and Ling-Yu Duan. Dual-refinement: Joint label and feature refinement for unsupervised domain adaptive person re-identification. IEEE TIP, 2021. 2 +[9] Zuozhuo Dai, Guangyuan Wang, Siyu Zhu, Weihao Yuan, and Ping Tan. Cluster contrast for unsupervised person re-identification. arxiv 2021. arXiv preprint arXiv:2103.11568, 2021. 2.7 +[10] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, 1996. 6 +[11] Jiawei Feng, Ancong Wu, and Wei-Shi Zheng. Shape-erased feature learning for visible-infrared person re-identification. In CVPR, 2023. 2 +[12] Yajun Gao, Tengfei Liang, Yi Jin, Xiaoyan Gu, Wu Liu, Yidong Li, and Congyan Lang. Mso: Multi-feature space joint optimization network for rgb-infrared person re-identification. In ACM MM, 2021. 1, 6, 7 +[13] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual mean-teaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. arXiv preprint arXiv:2001.01526, 2020. 7 +[14] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. NeurIPS, 2020. 7 +[15] Jianyang Gu, Weihua Chen, Hao Luo, Fan Wang, Hao Li, Wei Jiang, and Weijie Mao. Multi-view evolutionary training for unsupervised domain adaptive person re-identification. IEEE TIFS, 2022. 1 +[16] Xin Hao, Sanyuan Zhao, Mang Ye, and Jianbing Shen. Cross-modality person re-identification via modality confusion and center aggregation. In ICCV, 2021. 6, 7 + +[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, 2021. 6 +[18] Wenke Huang, Mang Ye, Zekun Shi, and Bo Du. Generalizable heterogeneous federated cross-correlation and instance similarity learning. IEEE TPAMI, 2023. 1 +[19] Wenke Huang, Mang Ye, Zekun Shi, Guancheng Wan, He Li, Bo Du, and Qiang Yang. A federated learning for generalization, robustness, fairness: A survey and benchmark. arXiv, 2023. 1 +[20] Minsu Kim, Seungryong Kim, Jungin Park, Seongheon Park, and Kwanghoon Sohn. Partmix: Regularization strategy to learn part discovery for visible-infrared person re-identification. In CVPR, 2023. 1, 7 +[21] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In AAAI, 2020. 1 +[22] Shuang Li, Jiaxu Leng, Ji Gan, Mengjingcheng Mo, and Xinbo Gao. Shape-centered representation learning for visible-infrared person re-identification. arXiv preprint arXiv:2310.17952, 2023. 1 +[23] Shuang Li, Fan Li, Jinxing Li, Huafeng Li, Bob Zhang, Dapeng Tao, and Xinbo Gao. Logical relation inference and multiview information interaction for domain adaptation person re-identification. IEEE TNNLS, 2023. 1 +[24] Wenqi Liang, Guangcong Wang, Jianhuang Lai, and Xiaohua Xie. Homogeneous-to-heterogeneous: Unsupervised learning for rgb-infrared person re-identification. IEEE TIP, 2021. 2, 3, 7 +[25] Shan Lin, Chang-Tsun Li, and Alex C. Kot. Multidomain adversarial feature generalization for person re-identification. IEEE TIP, 2021. 1 +[26] Haojie Liu, Daoxun Xia, and Wei Jiang. Towards homogeneous modality learning and multi-granularity information exploration for visible-infrared person re-identification. IEEE Journal of Selected Topics in Signal Processing. 7 +[27] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, 2022. 1, 7 +[28] Yan Lu, Yue Wu, Bin Liu, Tianzhu Zhang, Baopu Li, Qi Chu, and Nenghai Yu. Cross-modality person re-identification with shared-specific feature transfer. In CVPR, 2020. 1 +[29] Zefeng Lu, Ronghao Lin, and Haifeng Hu. Modality and camera factors bi-disentanglement for nir-vis object re-identification. IEEE TIFS, 2023. +[30] Chuanchen Luo, Chunfeng Song, and Zhaoxiang Zhang. Learning to adapt across dual discrepancy for cross-domain person re-identification. IEEE TPAMI, 2022. 1, 5 +[31] Hao Luo, Pichao Wang, Yi Xu, Feng Ding, Yanxin Zhou, Fan Wang, Hao Li, and Rong Jin. Self-supervised pre-training for transformer-based person re-identification. arXiv preprint arXiv:2111.12084, 2021. 6 +[32] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 2017. 6 + +[33] Zhiqi Pang, Chunyu Wang, Lingling Zhao, Yang Liu, and Gaurav Sharma. Cross-modality hierarchical clustering and refinement for unsupervised visible-infrared person re-identification. IEEE TCSVT, 2023. 2, 3, 7 +[34] Zhihao Qian, Yutian Lin, and Bo Du. Visible-infrared person re-identification via patch-mixed cross-modality learning. *IJCAI*, 2023. 1, 7 +[35] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, 2023. 6, 7 +[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8 +[37] Jiangming Wang, Zhizhong Zhang, Mingang Chen, Yi Zhang, Cong Wang, Bin Sheng, Yanyun Qu, and Yuan Xie. Optimal transport for label-efficient visible-infrared person re-identification. ECCV, 2022. 2, 3, 6, 7 +[38] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, 2021. 7 +[39] Yiming Wang, Guanqiu Qi, Shuang Li, Yi Chai, and Huafeng Li. Body part-level domain alignment for domain-adaptive person re-identification with transformer framework. IEEE TIFS, 2022. 1 +[40] Yadi Wang, Hongyun Zhang, Duoqian Miao, and Witold Pedrycz. Multi-granularity re-ranking for visible-infrared person re-identification. CAAI Transactions on Intelligence Technology, 2023. 1 +[41] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In CVPR, 2019. 1 +[42] Zheng Wang, Zhixiang Wang, Yinqiang Zheng, Yang Wu, Wenjun Zeng, and Shin'ichi Satoh. Beyond intra-modality: A survey of heterogeneous person re-identification. arXiv preprint arXiv:1905.10048, 2019. 1 +[43] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Syncretic modality collaborative learning for visible infrared person re-identification. In ICCV, 2021. 7 +[44] Ancong Wu, Wei-Shi Zheng, and Jian-Huang Lai. Robust depth-based person re-identification. IEEE TIP, 2017. 1 +[45] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In ICCV, 2017. 1, 6 +[46] Dongming Wu, Mang Ye, Gaojie Lin, Xin Gao, and Jianbing Shen. Person re-identification by context-aware part attention and multi-head collaborative learning. IEEE TIFS, 2022. 1 +[47] Qiong Wu, Pingyang Dai, Jie Chen, Chia-Wen Lin, Yongjian Wu, Feiyue Huang, Bineng Zhong, and Rongrong Ji. Discover cross-modality nuances for visible-infrared person re-identification. In CVPR, 2021. 1, 7 +[48] Zesen Wu and Mang Ye. Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning. In CVPR, 2023. 2, 3, 6, 7 + +[49] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, 2021. 2, 7 +[50] Bin Yang, Mang Ye, Jun Chen, and Zesen Wu. Augmented dual-contrastive aggregation learning for unsupervised visible-infrared person re-identification. In ACM MM, 2022. 1, 2, 3, 6, 7 +[51] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In ICCV, 2023. 2, 6, 7 +[52] Bin Yang, Jun Chen, and Mang Ye. Top-k visual tokens transformer: Selecting tokens for visible-infrared person re-identification. In ICASSP, 2023. 1 +[53] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE TIFS, pages 407–419, 2019. 1, 6 +[54] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In ECCV, 2020. 1, 6, 7 +[55] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE TIFS, 2021. 1, 2 +[56] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In ICCV, 2021. 2, 6, 7 +[57] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 2021. 1, 6, 7 +[58] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE TPAMI, 2023. 1 +[59] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 2022. 1 +[60] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In CVPR, 2022. 7 +[61] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, 2023. 1, 7 +[62] Yiyuan Zhang, Yuhao Kang, Sanyuan Zhao, and Jianbing Shen. Dual-semantic consistency learning for visible-infrared person re-identification. IEEE TIFS, 2022. 1 +[63] Yafei Zhang, Yongzeng Wang, Huafeng Li, and Shuang Li. Cross-compatible embedding and semantic consistent feature construction for sketch re-identification. In ACM MM, 2022. 1 +[64] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, 2015. 3 \ No newline at end of file diff --git a/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/images.zip b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1503104368ee447f0bda38ea881c074ce566b599 --- /dev/null +++ b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78771944af2f89dc554494cf5937b16ac7959a962382af76d1cb38e444b13c62 +size 822211 diff --git a/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/layout.json b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6adca9aea5e66c1c5261766728123ff14d00a4e4 --- /dev/null +++ b/2024/Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification/layout.json @@ -0,0 +1,10369 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 519, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 519, + 138 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 519, + 138 + ], + "type": "text", + "content": "Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 161, + 383, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 161, + 383, + 175 + ], + "spans": [ + { + "bbox": [ + 215, + 161, + 383, + 175 + ], + "type": "text", + "content": "Bin Yang Jun Chen* Mang Ye*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 176, + 452, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 176, + 452, + 189 + ], + "spans": [ + { + "bbox": [ + 140, + 176, + 452, + 189 + ], + "type": "text", + "content": "National Engineering Research Center for Multimedia Software," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 190, + 512, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 190, + 512, + 203 + ], + "spans": [ + { + "bbox": [ + 82, + 190, + 512, + 203 + ], + "type": "text", + "content": "School of Computer Science, Hubei Luojia Laboratory, Wuhan University, Wuhan, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 205, + 383, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 205, + 383, + 217 + ], + "spans": [ + { + "bbox": [ + 202, + 205, + 383, + 217 + ], + "type": "text", + "content": "https://github.com/yangbincv/SDCL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 582 + ], + "type": "text", + "content": "Unsupervised visible-infrared person re-identification (US-VI-ReID) centers on learning a cross-modality retrieval model without labels, reducing the reliance on expensive cross-modality manual annotation. Previous US-VI-ReID works gravitate toward learning cross-modality information with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. The shallow feature from the shallow layers contains nuanced detail information, which is critical for effective cross-modality learning but is disregarded regrettably by the existing methods. To address the above issues, we design a Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration which are harnessed for neighbor alignment, enhancing the robustness in a dynamic manner. Furthermore, CRA associates the cross-modality labels with the ranking association between shallow and deep features, furnishing valuable supervision for cross-modality learning. Extensive experiments validate the superiority of our method, even outperforming certain supervised counterparts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 604, + 127, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 127, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 127, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 696 + ], + "type": "text", + "content": "Person re-identification (ReID) aims at matching the same person image captured by non-overlapping cameras. In recent times, ReID has garnered significant attention from the computer vision research community, owing to its pivotal role in the context of intelligent video surveillance applications [2, 18, 19, 23, 25, 28-30, 39, 41, 44, 46, 57," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 244, + 426, + 344 + ], + "blocks": [ + { + "bbox": [ + 309, + 244, + 426, + 344 + ], + "lines": [ + { + "bbox": [ + 309, + 244, + 426, + 344 + ], + "spans": [ + { + "bbox": [ + 309, + 244, + 426, + 344 + ], + "type": "image", + "image_path": "8709010003d7ba753fffa03e46b6e56d6c0c2fa19cc2c7548567d31034bf2688.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 347, + 380, + 363 + ], + "lines": [ + { + "bbox": [ + 309, + 347, + 380, + 363 + ], + "spans": [ + { + "bbox": [ + 309, + 347, + 380, + 363 + ], + "type": "text", + "content": "Ground Truth Cross-modality Label Pairs" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 348, + 423, + 363 + ], + "lines": [ + { + "bbox": [ + 390, + 348, + 423, + 363 + ], + "spans": [ + { + "bbox": [ + 390, + 348, + 423, + 363 + ], + "type": "text", + "content": "Deep Features" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 372, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 460 + ], + "type": "text", + "content": "Figure 1. Illustration of the motivation. Previous advanced works for US-VI-ReID mainly focus on learning features and associating cross-modality labels with deep features. However, interfered by the multiple discrepancies, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in (a). In this paper, we incorporate shallow and deep features with collaborative learning and label association to remedy these problems, as shown in (b)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 246, + 545, + 343 + ], + "blocks": [ + { + "bbox": [ + 426, + 246, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 426, + 246, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 426, + 246, + 545, + 343 + ], + "type": "image", + "image_path": "3569712b422bf865fb2c61097d73afc29bde2ba49f540cd7f4911a0f32462460.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 348, + 473, + 363 + ], + "lines": [ + { + "bbox": [ + 441, + 348, + 473, + 363 + ], + "spans": [ + { + "bbox": [ + 441, + 348, + 473, + 363 + ], + "type": "text", + "content": "Shallow Features" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 494, + 350, + 543, + 363 + ], + "lines": [ + { + "bbox": [ + 494, + 350, + 543, + 363 + ], + "spans": [ + { + "bbox": [ + 494, + 350, + 543, + 363 + ], + "type": "text", + "content": "Negative Optimization" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "type": "text", + "content": "59, 62, 63]. The rank-1 accuracy has exhibited encouraging outcomes in supervised person ReID. Notably, these person images are typically acquired by visible cameras within well-illuminated environments. However, it is crucial to recognize that visible cameras often fail to capture sufficient information of individuals in low-light settings, consequently constraining the practicality of single-modality ReID in the context of night-time surveillance [1, 7, 15, 21, 40, 42, 45, 53, 61]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "In contrast, recently deployed cameras possess the capability to seamlessly transition into far/near-infrared mode during nighttime operations [50, 55]. Various techniques [12, 20, 22, 27, 34, 47, 52, 54-58, 62] have been proposed for visible-infrared person re-identification (VI-ReID), resulting in commendable accuracy levels. Notably, these achievements are facilitated by extensive human-labeled cross-modality datasets. However, annotating cross-modality datasets demands more resources compared to the annotation of single-modality ReID datasets. This" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16870" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "intricate annotation task presents a formidable and pivotal challenge, ultimately giving rise to the crucial task of unsupervised VI-ReID (US-VI-ReID) for reducing the cost of expensive cross-modality annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 289, + 529 + ], + "type": "text", + "content": "Some prior research endeavors [4, 5, 24, 33, 37, 48, 50, 51] have proposed preliminary solutions for US-VI-ReID. These solutions gravitate toward learning cross-modality feature with the deep features extracted from the ultimate layer. Nevertheless, interfered by the multiple discrepancies, i.e., intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization, as shown in Fig. 1. It is important to recognize that the shallow features originating from the modality-specific shallow layers harbor a wealth of nuanced detail information pertaining to pedestrian attributes, which is critical for unsupervised cross-modality representation learning but is disregarded regrettably by the existing methods. To harness the complete potential of the shallow information, we devise a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer with shallow-deep contrastive learning, incorporating Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. Specifically, CNL unveils the intrinsic homogeneous and heterogeneous collaboration as the constraints for seeking reliable intramodality and inter-modality neighbor learning, thereby guaranteeing the cultivation of modality-invariant and discriminative representations in a dynamic manner. Furthermore, CRA is developed with a global perspective to exploit cross-modality ranking consistency between deep and shallow features, associating the label of two modalities in a collaborative manner and furnishing valuable cross-modality supervision for cross-modality learning. Ultimately, within the shallow-deep collaborative learning framework with CNL and CRA modules, we acquire a robust representation, effectively mitigating the cross-modality discrepancy under unsupervised conditions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 531, + 280, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 531, + 280, + 543 + ], + "spans": [ + { + "bbox": [ + 59, + 531, + 280, + 543 + ], + "type": "text", + "content": "The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 546, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 546, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 286, + 604 + ], + "type": "text", + "content": "- We propose a shallow-deep collaborative learning framework based on the transformer architecture. This framework facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow and deep features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 606, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 286, + 652 + ], + "type": "text", + "content": "- We propose a collaborative neighbor learning module to formulate dependable intra-modality and cross-modality neighbor learning, enabling the model to capture modality-invariant and discriminative features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 654, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 700 + ], + "type": "text", + "content": "- We propose a collaborative ranking association module to explore intra-modality and cross-modality ranking consistencies, unifying the cross-modality labels and providing invaluable cross-modality supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 702, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 286, + 712 + ], + "type": "text", + "content": "- Extensive experiments validate that our SDCL frame" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 109 + ], + "type": "text", + "content": "work surpasses existing methods on two mainstream VI-ReID benchmarks, consistently improving the unsupervised cross-modality retrieval performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 126, + 392, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 126, + 392, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 126, + 392, + 139 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 148, + 520, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 148, + 520, + 162 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 520, + 162 + ], + "type": "text", + "content": "2.1. Supervised Visible-Infrared Person ReID" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 170, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 312 + ], + "type": "text", + "content": "Currently, advanced supervised VI-ReID focuses on bridging the gap between the two modalities and learning robust representations against modality discrepancy. Ye et al. [55] developed a dynamic tri-level relation mining (DTRM) framework simultaneously to explore the cross-modality relation cues of channel-level, part-level intra-modality, and graph-level. CAJ [56] proposed a channel-mixed learning to handle the intra-modality and cross-modality variations by randomly exchanging the color channels. SGIEL [11] proposed a shape-erased feature learning paradigm, jointly learning shape-related feature in one subspace and shape-erased features in the orthogonal complement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 316, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 388 + ], + "type": "text", + "content": "Nevertheless, the remarkable performances exhibited by these methods require extensive human-labeled cross-modality datasets. In this work, we pivot our focus to the realm of unsupervised visible-infrared person ReID, where the luxury of identity annotations is absent, presenting important applications for real-world VI-ReID deployments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 403, + 533, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 533, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 533, + 415 + ], + "type": "text", + "content": "2.2. Unsupervised Single-Modality Person ReID" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 626 + ], + "type": "text", + "content": "Unsupervised single-modality ReID endeavors to train the ReID model using unlabeled data captured by visible cameras. The majority of studies have embraced cluster algorithms to derive pseudo labels for optimizing the model. Cluster Contrast [9] utilized a distinctive cluster representation to delineate each cluster, addressing the issue of cluster inconsistency. Chen et al. [3] introduced Inter-instance Contrastive Encoding (ICE), which harnesses inter-instance pairwise similarity scores to enhance preceding class-level contrastive ReID methodologies. IICS [49] tackled the unsupervised ReID challenge by decomposing the similarity computation into two stages, namely, the intra-domain and inter-domain computations, respectively. Dai et al. [8] introduced a dual-refinement approach to concurrently enhance pseudo labels during the offline clustering phase and refine features during the online training phase, augmenting label purity and feature discriminability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "The above methods are dedicated to addressing issues in single-modality ReID. When employed in the context of US-VI-ReID, these approaches encounter challenges stemming from cross-modality discrepancies and the absence of cross-modality (visible-infrared) identity labels. This impedes the learning of inter-modality feature and the generation of reliable cross-modality pseudo labels." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16871" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 274, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 274, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 274, + 84 + ], + "type": "text", + "content": "2.3. Unsupervised Visible-Infrared Person ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 270 + ], + "type": "text", + "content": "Several studies [24, 37, 48, 50] represent initial endeavors in the realm of US-VI-ReID. Yang et al. [50] introduced a novel Augmented Dual-Contrastive Aggregation (ADCA) learning framework founded on the principles of homogeneous joint learning and heterogeneous aggregation, establishing a robust baseline for purely unsupervised VI-ReID. PGM [48] developed a progressive graph matching method to systematically extract cross-modality correspondences, formulating correspondence mining as a graph-matching process. H2H [24] introduced a homogeneous-to-heterogeneous approach through two-stage learning, incorporating Market1501 [64] as additional labeled data. CHCR [33] proposed a cross-modality hierarchical clustering and refinement method by promoting modality-invariant feature learning and improving the reliability of pseudo-labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 270, + 288, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 288, + 390 + ], + "type": "text", + "content": "Nevertheless, the above methods primarily concentrate on feature learning with deep features. However, interfered by the intra-modality variation and inter-modality discrepancy, solely relying on deep features is insufficient for accurately learning modality-invariant features, resulting in negative optimization. Our approach integrates shallow and deep features through collaborative feature learning and cross-modality label association, which is distinguished from previous works, surpassing the performance of existing US-VI-ReID methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 402, + 153, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 153, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 153, + 415 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 422, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 662 + ], + "type": "text", + "content": "The framework of shallow-deep collaborative learning (SDCL) is shown in Fig. 2, incorporating the Collaborative Neighbor Learning (CNL) and Collaborative Ranking Association (CRA) module. SDCL adopts a dual-path transformer architecture with dual-contrastive learning [50], which has two shallow modality-specific patch embedding layers and a modality-shared transformer. Instance memory and cluster memory are constructed for shallow embeddings and deep features within each modality. Instance memory stores all training image features. Cluster memory is built by averaging the instance features with the same pseudo labels. With the above memories, SDCL develops the CNL module to exploit homogeneous and heterogeneous collaboration as the constraints for seeking reliable intra-modality and inter-modality neighbor learning. CRA explores the ranking consistency of cross-modality shallow embeddings and deep features to associate reliable cross-modality identities. With the collaboration of shallow embeddings and deep features within CNL and CRA modules, SDCL learns better representation for cross-modality retrieval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 236, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 236, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 236, + 684 + ], + "type": "text", + "content": "3.1. Shallow-deep Contrastive Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "We first introduce shallow-deep contrastive learning based on augmented dual-contrast learning [50] with a dual-path" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 73, + 407, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 407, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 407, + 83 + ], + "type": "text", + "content": "transformer architecture." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 85, + 546, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 546, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 546, + 133 + ], + "type": "text", + "content": "Cluster Memory Initialization. At the beginning of each training epoch, we construct shallow and deep cluster memories for each modality by averaging the shallow and deep feature of one cluster, which can be denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 143, + 545, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 143, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 376, + 143, + 545, + 172 + ], + "type": "interline_equation", + "content": "\\phi_ {k} ^ {v s} = \\frac {1}{| \\mathcal {H} _ {k} ^ {v} |} \\sum_ {u _ {n} ^ {v s} \\in y _ {k} ^ {v}} u _ {n} ^ {v s}, \\tag {1}", + "image_path": "02c54359590e36f275a8435af3b476eeaf38059ce339980c45621c258e2959c6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 376, + 187, + 545, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 187, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 376, + 187, + 545, + 218 + ], + "type": "interline_equation", + "content": "\\phi_ {k} ^ {v d} = \\frac {1}{\\left| \\mathcal {H} _ {k} ^ {v} \\right|} \\sum_ {u _ {n} ^ {v d} \\in y _ {k} ^ {v}} u _ {n} ^ {v d}, \\tag {2}", + "image_path": "fec33c81fb3e3c348e46585647195911eda316b1fac570c49b5501b0665a883c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 230, + 545, + 260 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 230, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 376, + 230, + 545, + 260 + ], + "type": "interline_equation", + "content": "\\phi_ {l} ^ {r s} = \\frac {1}{| \\mathcal {H} _ {l} ^ {r} |} \\sum_ {u _ {n} ^ {r s} \\in y _ {l} ^ {r}} u _ {n} ^ {r s}, \\tag {3}", + "image_path": "c5d176f788fce8d0ece1ab8ede75c83a4646018b431329b940ac0fa6a1691717.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 272, + 545, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 272, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 376, + 272, + 545, + 302 + ], + "type": "interline_equation", + "content": "\\phi_ {l} ^ {r d} = \\frac {1}{| \\mathcal {H} _ {l} ^ {r} |} \\sum_ {u _ {n} ^ {r d} \\in y _ {l} ^ {r}} u _ {n} ^ {r d}, \\tag {4}", + "image_path": "278b34fe04fd6cfd7d99e187b25a960ce6c0bd93a9243f9ff2fbf6391ff6b543.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "u_{n}^{r*}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "u_{n}^{v*}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": " are infrared and visible instance features, respectively. " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "u_{n}^{*s}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "u_{n}^{*d}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": " are shallow and deep instance features, respectively. " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_k" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": "-th cluster set and " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": " counts the number of samples of a set. Meanwhile, we store the instance features in shallow and deep infrared and visible instance memories " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "U^{rs}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "U^{rd}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "U^{vs}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "inline_equation", + "content": "U^{vd}" + }, + { + "bbox": [ + 304, + 310, + 545, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 383, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 418 + ], + "type": "text", + "content": "Cluster Memory Updating. During each training iteration, the deep and shallow cluster memories of two modalities by a momentum updating strategy:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 428, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 428, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 349, + 428, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\phi_ {k} ^ {(\\delta)} \\leftarrow \\alpha \\phi_ {k} ^ {(\\delta - 1)} + (1 - \\alpha) q, q \\in y _ {k}, \\tag {5}", + "image_path": "8eb88e4866493d2e953bbd01afade41306bb221ba269da6307cf4023381b4d44.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "text", + "content": " is the query features sampled from training set in each training iteration, respectively. " + }, + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "text", + "content": " is the momentum factor. " + }, + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 456, + 545, + 491 + ], + "type": "text", + "content": " is the training iteration number." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "text", + "content": "Contrastive Loss. Given shallow and deep visible and infrared query " + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "inline_equation", + "content": "q^{vs}" + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "inline_equation", + "content": "q^{vd}" + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "inline_equation", + "content": "q^{rs}" + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "inline_equation", + "content": "q^{rs}" + }, + { + "bbox": [ + 304, + 492, + 545, + 529 + ], + "type": "text", + "content": ", we compute the contrastive loss for visible modality by the following equations:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 346, + 540, + 545, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 540, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 346, + 540, + 545, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i d} ^ {v s} = - \\log \\frac {\\exp \\left(q ^ {v s} \\cdot \\phi_ {+} ^ {v s} / \\tau\\right)}{\\sum_ {k = 0} ^ {K} \\exp \\left(q ^ {v s} \\cdot \\phi_ {k} ^ {v s} / \\tau\\right)}, \\tag {6}", + "image_path": "bb987c246d99c3a8d7c3dc8829ffbdd024aabbfadc7406dfd4c9444530e3f095.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 345, + 579, + 545, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 579, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 345, + 579, + 545, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i d} ^ {v d} = - \\log \\frac {\\exp \\left(q ^ {v d} \\cdot \\phi_ {+} ^ {v d} / \\tau\\right)}{\\sum_ {k = 0} ^ {K} \\exp \\left(q ^ {v d} \\cdot \\phi_ {k} ^ {v d} / \\tau\\right)}, \\tag {7}", + "image_path": "4b0a2c0d666224b47551f3422f9d67b004eb01098948bde193cf83b74f42cc9b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\phi_{+}" + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "text", + "content": " is the positive memory corresponding to the pseudo label of " + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "text", + "content": " is a temperature. The infrared contrastive loss " + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{id}^{rs}" + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{id}^{rd}" + }, + { + "bbox": [ + 304, + 616, + 545, + 654 + ], + "type": "text", + "content": " are obtained by the same way." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "We optimize the modality-specific shallow embedding layers and deep modality-shared layers by combining deep and shallow contrastive loss:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 361, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 361, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i d} = \\mathcal {L} _ {i d} ^ {v d} + \\mathcal {L} _ {i d} ^ {r d} + \\mathcal {L} _ {i d} ^ {v s} + \\mathcal {L} _ {i d} ^ {r s} \\tag {8}", + "image_path": "0e6a7745080ffd50b2484be57dfdda5d988410e2e58c27f507064df6f6e0be4b.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "16872" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 358 + ], + "type": "image", + "image_path": "ef5faaca97f99d3d306196012079f3290a78ebf352900d4c64a9899b5c521150.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "type": "text", + "content": "Figure 2. Illustration of shallow-deep collaborative learning. It comprises shallow-deep contrastive learning, collaborative neighbor learning, and collaborative ranking association. SDCL finds reliable neighborhood and cross-modality labels with shallow-deep collaborations, enhancing the robustness of learned representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 423, + 228, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 228, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 228, + 437 + ], + "type": "text", + "content": "3.2. Collaborative Neighbor Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": "Our basic rationale for the CNL module is that deep and shallow features are complementary, and their intrinsic consistency relationship can be used to constrain intra-modality and inter-modality optimization, formulating collaborative learning to enhance the robustness against intra-modality and inter-modality variations. From the above perspective, CNL explores essential homogeneous and heterogeneous shallow-deep collaborations to seek reliable intra-modality and inter-modality neighbor learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 551, + 288, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 288, + 588 + ], + "type": "text", + "content": "Homogeneous and Heterogeneous Collaboration. Given a query " + }, + { + "bbox": [ + 46, + 551, + 288, + 588 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 551, + 288, + 588 + ], + "type": "text", + "content": ", we can get the similarity " + }, + { + "bbox": [ + 46, + 551, + 288, + 588 + ], + "type": "inline_equation", + "content": "s(q_i, u_j)" + }, + { + "bbox": [ + 46, + 551, + 288, + 588 + ], + "type": "text", + "content": " between the query and each instance in the training set by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 595, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 595, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 114, + 595, + 287, + 620 + ], + "type": "interline_equation", + "content": "s \\left(q _ {i}, u _ {j}\\right) = \\frac {q _ {i} \\cdot u _ {j}}{\\left| \\left| q _ {i} \\right| \\right| _ {2} \\left| \\left| u _ {j} \\right| \\right| _ {2}}, \\tag {9}", + "image_path": "2c2bb86d9389413d9c6c44e858a552e321d4cb54ddcd6e38c39e8acd4262f488.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "u_{j}" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": " come from shallow or deep features of visible or infrared modality to get multiple types of homogeneous and heterogeneous shallow or deep query-instance similarity " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{vs},u_{j}^{vs})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{rs},u_{j}^{rs})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{vs},u_{j}^{rs})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{rs},u_{j}^{vs})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{vd},u_{j}^{vd})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{rd},u_{j}^{rd})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{vd},u_{j}^{rd})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "inline_equation", + "content": "s(q_{i}^{rd},u_{j}^{vd})" + }, + { + "bbox": [ + 47, + 628, + 287, + 691 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 690, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 690, + 288, + 714 + ], + "type": "text", + "content": "With the above multiple types of similarities, we seek reliable shallow and deep intra-modality and inter-modality" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "content": "neighbors, formulating collaborative neighbor learning. The intra-modality neighbors can be defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 456, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 456, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 353, + 456, + 545, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right) = \\left\\{\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v s}\\right) \\cap \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v d}\\right) \\right\\}, \\tag {10}", + "image_path": "ec6347907ba44f8e0c89367d37c3caaa73fe428731dbbb777b7c12e09cb7c8f5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 491, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 491, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 354, + 491, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {r} \\left(q _ {i} ^ {r}\\right) = \\left\\{\\mathcal {N} ^ {r} \\left(q _ {i} ^ {r s}\\right) \\cap \\mathcal {N} ^ {r} \\left(q _ {i} ^ {r d}\\right) \\right\\}, \\tag {11}", + "image_path": "7ee3c7aa46582b9ce0f091ec5f51868b71aae159f2b6324ba55ee2e6c050f600.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^{v*})" + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r(q_i^{r*})" + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "text", + "content": " are neighborhood sets searched by visible-visible and infrared-infrared similarity, respectively. " + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "inline_equation", + "content": "* \\in \\{s, d\\}" + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "text", + "content": " represents the shallow and deep feature. The " + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^{vs})" + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r(q_i^{vd})" + }, + { + "bbox": [ + 305, + 510, + 547, + 559 + ], + "type": "text", + "content": " can be obtained by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 565, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 565, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 310, + 565, + 545, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v s}\\right) = \\left\\{u _ {j} ^ {v s} \\mid s \\left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {v}} s \\left(q _ {i} ^ {v s}, u _ {j} ^ {v s}\\right) \\right\\}, \\tag {12}", + "image_path": "0c04e123778b91ea888eefb0082dd74b9e8f833abd8b9df632a88881beeb507a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 605, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 605, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 309, + 605, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {v} \\left(q _ {i} ^ {v d}\\right) = \\left\\{u _ {j} ^ {v d} \\mid s \\left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N v} s \\left(q _ {i} ^ {v d}, u _ {j} ^ {v d}\\right) \\right\\}, \\tag {13}", + "image_path": "8d5011e36a8bc482a341c77387ecca85d53bdb984bd5c3ca781440b417d38ec2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "inline_equation", + "content": "N^v" + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "text", + "content": " denotes the number of visible instances and the " + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "text", + "content": " is the positive neighbor selection range. The " + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r (q_i^{rs})" + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r (q_i^{rd})" + }, + { + "bbox": [ + 304, + 631, + 545, + 667 + ], + "type": "text", + "content": " can be calculated by similar manner." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 667, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 667, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 667, + 545, + 690 + ], + "type": "text", + "content": "Given query " + }, + { + "bbox": [ + 304, + 667, + 545, + 690 + ], + "type": "inline_equation", + "content": "q_{i}^{v}" + }, + { + "bbox": [ + 304, + 667, + 545, + 690 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 667, + 545, + 690 + ], + "type": "inline_equation", + "content": "q_{i}^{r}" + }, + { + "bbox": [ + 304, + 667, + 545, + 690 + ], + "type": "text", + "content": ", the inter-modality neighbors can be denoted as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 353, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 353, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v}\\right) = \\left\\{\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v s}\\right) \\cap \\mathcal {N} ^ {r} \\left(q _ {i} ^ {v d}\\right) \\right\\}, \\tag {14}", + "image_path": "7c2be1c30aab59976cf407360fb93b650a36c8b1ca526d02c25c18f1c5622c1e.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16873" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 95, + 83, + 287, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 83, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 95, + 83, + 287, + 99 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {v} \\left(q _ {i} ^ {r}\\right) = \\left\\{\\mathcal {N} ^ {v} \\left(q _ {i} ^ {r s}\\right) \\cap \\mathcal {N} ^ {v} \\left(q _ {i} ^ {r d}\\right) \\right\\}, \\tag {15}", + "image_path": "213b4c5e2594fb32c5eafdf44c2fce84963960c03690d229340b6ff38cd047c2.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 260, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 260, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 260, + 117 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 102, + 260, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r (q_i^{vs})" + }, + { + "bbox": [ + 47, + 102, + 260, + 117 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 102, + 260, + 117 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r (q_i^{vd})" + }, + { + "bbox": [ + 47, + 102, + 260, + 117 + ], + "type": "text", + "content": " are defined as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 123, + 287, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 123, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 52, + 123, + 287, + 140 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v s}\\right) = \\left\\{u _ {j} ^ {r s} \\mid s \\left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {r}} s \\left(q _ {i} ^ {v s}, u _ {j} ^ {r s}\\right) \\right\\}, \\tag {16}", + "image_path": "53f5ec3af012bf310473a086824329d90d73d3441b3b3dff1d2f80bf8567a8db.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 164, + 287, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 164, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 51, + 164, + 287, + 182 + ], + "type": "interline_equation", + "content": "\\mathcal {N} ^ {r} \\left(q _ {i} ^ {v d}\\right) = \\left\\{u _ {j} ^ {r d} \\mid s \\left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\\right) > \\gamma \\cdot \\max _ {j = 1 \\dots N ^ {r}} s \\left(q _ {i} ^ {v d}, u _ {j} ^ {r d}\\right) \\right\\}, \\tag {17}", + "image_path": "c12df73d514319d68ef876f904e08f57fbf26c3b2c26ee3dcd7e91bd56fdde60.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "inline_equation", + "content": "N^r" + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "text", + "content": " denotes the number of infrared instances. Similarly, we can get the " + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^{rs})" + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^{rd})" + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "text", + "content": ". With the above procedure, we get the " + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^v), \\mathcal{N}^r(q_i^r), \\mathcal{N}^r(q_i^v)" + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^r)" + }, + { + "bbox": [ + 47, + 190, + 287, + 262 + ], + "type": "text", + "content": ", which are the reliable neighborhood set searched by the constraints of homogeneous or heterogeneous shallow-deep collaborations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "content": "Collaborative Neighbor Learning. With the neighborhood set " + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^v)" + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r(q_i^r)" + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^r(q_i^v)" + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^v(q_i^r)" + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "content": ", we can perform neighbor learning. Given query " + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "inline_equation", + "content": "q_i^{vs}" + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "inline_equation", + "content": "q_i^{vs}" + }, + { + "bbox": [ + 47, + 262, + 287, + 322 + ], + "type": "text", + "content": ", we can obtain the expression of the visible-visible shallow and deep neighbor learning by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 287, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 287, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} ^ {v v} = - \\frac {1}{N ^ {b}} \\sum_ {i = 1} ^ {N ^ {b}} \\sum_ {j \\in \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right)} \\log \\frac {\\exp \\left(s \\left(q _ {i} ^ {v s} , u _ {j} ^ {v s}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N ^ {v}} \\exp \\left(s \\left(q _ {i} ^ {v s} , u _ {n} ^ {v s}\\right) / \\tau\\right)}, \\tag {18}", + "image_path": "b6d510fea54b4548acb3128a154ad3bcca370f5fc0b837903f028625248f08b3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 393, + 287, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 287, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} ^ {v v} = - \\frac {1}{N ^ {b}} \\sum_ {i = 1} ^ {N ^ {b}} \\sum_ {j \\in \\mathcal {N} ^ {v} \\left(q _ {i} ^ {v}\\right)} \\log \\frac {\\exp \\left(s \\left(q _ {i} ^ {v d} , u _ {j} ^ {v d}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N ^ {v}} \\exp \\left(s \\left(q _ {i} ^ {v d} , u _ {n} ^ {v d}\\right) / \\tau\\right)}, \\tag {19}", + "image_path": "68e8d6d66b2dfad7e6c02b5c40b8df77841363ea4cf90083d03dab55a8c506f1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 448, + 287, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 448, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 123, + 448, + 287, + 461 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {v v} = \\mathcal {L} _ {d} ^ {v v} + \\lambda_ {s} \\mathcal {L} _ {s} ^ {v v}, \\tag {20}", + "image_path": "c98df8dc3760e8de739c2051705872d00d591c70b5e1092ed77cf687680b54a7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "N^b" + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "text", + "content": " is the batch size of query " + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "text", + "content": ". Similarly, the neighbor learning for infrared-infrared " + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{rr}" + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "text", + "content": ", infrared-visible " + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{rv}" + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "text", + "content": " and visible-infrared " + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{vr}" + }, + { + "bbox": [ + 47, + 467, + 287, + 527 + ], + "type": "text", + "content": " can be obtained by similar ways. The final optimization for neighbor learning is denoted by the following combination:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 537, + 287, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 537, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 88, + 537, + 287, + 551 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n e i g h b o r}} = \\mathcal {L} ^ {v v} + \\mathcal {L} ^ {r r} + \\mathcal {L} ^ {r v} + \\mathcal {L} ^ {v r}. \\tag {21}", + "image_path": "1f243bddf64539d2bd209546080fe80fb46b448f1dfbdb1ff0c0c291186993c9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 560, + 224, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 560, + 224, + 572 + ], + "spans": [ + { + "bbox": [ + 59, + 560, + 224, + 572 + ], + "type": "text", + "content": "The overall loss for SDCL is denoted as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 114, + 583, + 287, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 583, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 114, + 583, + 287, + 596 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {n e i g h b o r}} + \\mathcal {L} _ {i d} \\tag {22}", + "image_path": "29e90f9fc43cb264a6d6d6e84395aa1dc99c689ae949ce6b32c5c44672613321.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": "Discussion. In contrast to neighborhood learning in [30], which exclusively relies on deep features for the selection of positive neighbors. Under the multiple discrepancies, the similarity of individual deep features is unreliable, resulting in wrong neighbor selection and negative optimization. Our method embraces collaborative neighbor learning, which dynamically refines cross-modality supervision through shallow and deep collaboration, achieving effective cross-modality feature alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 492, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 492, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 492, + 84 + ], + "type": "text", + "content": "3.3. Collaborative Ranking Association" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 90, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 90, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 305, + 90, + 545, + 175 + ], + "type": "text", + "content": "The CRA module is underpinned by dual rationales: (1) Shallow and deep features across the two modalities should have the same ranking consistency. (2) The ranking consistencies of shallow and deep features within one modality are more accurate. Accordingly, the CRA module contains two processes, i.e., inter-modality ranking association and intra-modality ranking smoothing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": "Inter-modality Ranking Association. With the similarities " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "s(u_{i}^{vs},u_{j}^{rs})" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "s(u_{i}^{vd},u_{j}^{rd})" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": ", we can get two visible-infrared ranking lists " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "\\{s(u_i^{vs},u_j^{rs}),j\\in [1,N^r ]\\}" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "\\{s(u_i^{vd},u_j^{rd}),j\\in [1,N^r ]\\}" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "u_{i}^{vs}" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "u_{i}^{vd}" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": ", which are denoted as " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_s^{vr}(u_i^{vs})" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_d^{vr}(u_i^{vd})" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": ". The label of " + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 175, + 545, + 250 + ], + "type": "text", + "content": "-th similar infrared instance in two ranking lists can be represented as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 351, + 261, + 545, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 261, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 351, + 261, + 545, + 277 + ], + "type": "interline_equation", + "content": "\\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r s}}, u _ {j} ^ {r s} = \\mathcal {R} _ {s} ^ {v r} \\left(u _ {i} ^ {v s}\\right) [ k ], \\tag {23}", + "image_path": "9cdbb5b59ee6e8d59b40b2ccfa2485f9647d7e77a3c4fce2988eee24a427e240.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 351, + 293, + 545, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 293, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 351, + 293, + 545, + 310 + ], + "type": "interline_equation", + "content": "\\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] = y _ {u _ {j} ^ {r d}}, u _ {j} ^ {r d} = \\mathcal {R} _ {d} ^ {v r} \\left(u _ {i} ^ {v d}\\right) [ k ], \\tag {24}", + "image_path": "8b681b5582260057c6a38321a1e264c0d8f626794da53ecf3e5ff948d5607b57.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{u_i^{vs}}^{vr}[k]" + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{u_i^{vd}}^{vr}[k]" + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "text", + "content": " are refined " + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "text", + "content": "-th cross-modality labels of " + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "inline_equation", + "content": "u_i^{vs}" + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "inline_equation", + "content": "u_i^{vd}" + }, + { + "bbox": [ + 305, + 315, + 545, + 390 + ], + "type": "text", + "content": " through the ranking on visible-infrared shallow and deep similarities. Based on the rationale (1), we propose to associate the cross-modality labels with the intersection of two label sets to investigate collaborative ranking consistency:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 350, + 399, + 545, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 399, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 350, + 399, + 545, + 417 + ], + "type": "interline_equation", + "content": "I _ {u _ {i} ^ {v}} (k) = \\left\\{\\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ] \\cap \\left\\{\\tilde {y} _ {u _ {i} ^ {v s}} ^ {v r} [ n ] \\right\\} _ {n = 1} ^ {N} \\right\\}, \\tag {25}", + "image_path": "ede1acc8c91356444a5de86869211833be800c17bfabd94aa872d754adcc9a25.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "inline_equation", + "content": "I_{u_i^v}(k)" + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "text", + "content": " records the samples with the same identity of top- " + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "text", + "content": " identity in instances from top-1 to top- " + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "text", + "content": ". The reliable refined cross-modality label of " + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "inline_equation", + "content": "u_i^v" + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "text", + "content": " can be expressed by the label of maximum count in " + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "inline_equation", + "content": "I_{u_i^v}(k)" + }, + { + "bbox": [ + 305, + 426, + 545, + 475 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 481, + 545, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 481, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 315, + 481, + 545, + 499 + ], + "type": "interline_equation", + "content": "\\tilde {y} _ {u _ {i} ^ {v}} ^ {c m} = \\tilde {y} _ {u _ {i} ^ {v d}} ^ {v r} [ k ], k = \\underset {k} {\\operatorname {a r g m a x}} \\left(\\left\\{\\left| I _ {u _ {i} ^ {v}} (k) \\right|, k \\in [ 1, K ] \\right\\}\\right), \\tag {26}", + "image_path": "1d323cc6a40d5f9b5274a4cea74c30f42ac86ea11407c6581b5f299f76fc1c67.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "text", + "content": " denotes the counting function. The argmax operation traverses " + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "inline_equation", + "content": "I_{u_i^v}(k)" + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "text", + "content": " and finds the labels with maximum number as cross-modality refined labels. Then, we convert the label list " + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "inline_equation", + "content": "\\{\\tilde{y}_{u_i^v}^{cm}, i \\in [1, N^v]\\}" + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "text", + "content": " to the form of one-hot code matrix " + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "inline_equation", + "content": "\\tilde{Y}^{cm} \\in \\mathbb{R}^{N^v \\times C^r}" + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "text", + "content": " by setting the column according to the refine labels to 1 and the rest to 0, where " + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "inline_equation", + "content": "C^r" + }, + { + "bbox": [ + 305, + 508, + 545, + 595 + ], + "type": "text", + "content": " is the class number of infrared modality." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 596, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 596, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 305, + 596, + 545, + 655 + ], + "type": "text", + "content": "Intra-modality Collaborative Smoothing. Based on the rationale (2), two shallow and deep homogeneous shallow " + }, + { + "bbox": [ + 305, + 596, + 545, + 655 + ], + "type": "inline_equation", + "content": "P_{s}^{vv}" + }, + { + "bbox": [ + 305, + 596, + 545, + 655 + ], + "type": "text", + "content": " and deep " + }, + { + "bbox": [ + 305, + 596, + 545, + 655 + ], + "type": "inline_equation", + "content": "P_{d}^{vv}" + }, + { + "bbox": [ + 305, + 596, + 545, + 655 + ], + "type": "text", + "content": " similarity matrices are constructed to investigate intra-modality ranking consistency, enhancing the precision of refined cross-modality pseudo labels by:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 375, + 665, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 665, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 375, + 665, + 545, + 681 + ], + "type": "interline_equation", + "content": "P _ {s} ^ {v v} (i, j) = s \\left(u _ {i} ^ {v s}, u _ {j} ^ {v s}\\right), \\tag {27}", + "image_path": "be5419695194eeb4e70804ed57391184fa1749f62eb2450f9314a11952086627.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 375, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 375, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "P _ {d} ^ {v v} (i, j) = s \\left(u _ {i} ^ {v d}, u _ {j} ^ {v d}\\right), \\tag {28}", + "image_path": "9e8a6ddf53da6114fdd9b84490d0fce733c14a6a66dc1b716c07481b3ead9899.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16874" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "inline_equation", + "content": "P_{*}^{vv}\\in \\mathbb{R}^{N^{v}\\times N^{v}}" + }, + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "text", + "content": " represents the intra-modality similarity structure. In order to explore the intra-modality shallow and deep collaboration, we calculate the sum of " + }, + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "inline_equation", + "content": "P_{s}^{vv}(i,j)" + }, + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "inline_equation", + "content": "P_d^{vv}(i,j)" + }, + { + "bbox": [ + 47, + 71, + 288, + 122 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 127, + 287, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 127, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 97, + 127, + 287, + 142 + ], + "type": "interline_equation", + "content": "P ^ {v v} (i, j) = P _ {s} ^ {v v} (i, j) + P _ {d} ^ {v v} (i, j) \\tag {29}", + "image_path": "ce87b462e43cdcfe58b7ca394951aa3b8776a935f0f34e6a97537a782c4b904a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 287, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 148, + 287, + 207 + ], + "type": "inline_equation", + "content": "P^{vv}" + }, + { + "bbox": [ + 46, + 148, + 287, + 207 + ], + "type": "text", + "content": " indicates the consistency of shallow and deep similarity matrix. We keep the 5-max values of " + }, + { + "bbox": [ + 46, + 148, + 287, + 207 + ], + "type": "inline_equation", + "content": "P^{vv}" + }, + { + "bbox": [ + 46, + 148, + 287, + 207 + ], + "type": "text", + "content": " in each row to 1 and the rest to 0, acquiring the ranking relations. The process of intra-modality ranking smoothing is formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 213, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 213, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 129, + 213, + 287, + 228 + ], + "type": "interline_equation", + "content": "Y ^ {c m} = P ^ {v v} \\tilde {Y} ^ {c m}, \\tag {30}", + "image_path": "c5b9cdfa1cc88851dd062cc0c8d16710cd65bff2bacf06bbd5de2103b9b167f6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 233, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 233, + 287, + 450 + ], + "type": "inline_equation", + "content": "Y^{cm} \\in \\mathbb{R}^{N^v \\times C^r}" + }, + { + "bbox": [ + 46, + 233, + 287, + 450 + ], + "type": "text", + "content": " is the final refined cross-modality label matrix of visible instance. In " + }, + { + "bbox": [ + 46, + 233, + 287, + 450 + ], + "type": "inline_equation", + "content": "Y^{cm}" + }, + { + "bbox": [ + 46, + 233, + 287, + 450 + ], + "type": "text", + "content": ", the column number of the maximum value in each row is the refined label of samples. In our work, we refine the pseudo labels of the visible modality to the infrared modality and smooth the cross-modality labels with the shallow and deep similarity within the visible modality. Then, the infrared and refined visible labels will be used to construct a modality-shared memory for contrastive learning within two modalities [51]. Discussion. Our CRA is essentially the collaboration of shallow and deep ranking processes with a global perspective at the beginning of each training epoch, which is distinguished from previous works employing a solitary-pattern deep feature similarity [37, 48, 50, 51]. The imposition of a collaborative consistency constraint on dual-pattern similarities within CRA confers a distinct advantage, facilitating the exploration of more dependable cross-modality relations and providing valuable cross-modality supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 479, + 226, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 226, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 226, + 491 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocol" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 665 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed SDCL framework on two widely-used visible-infrared person ReID datasets, i.e., SYSU-MM01 [45] and RegDB [32]. SYSU-MM01 dataset is collected by 6 different cameras, including 22258 visible and 11909 near-infrared images of 395 training identities. We perform ten trials of the gallery set selection [53] and calculate the average performance following existing methods with all-search and indoor-search testing mode. RegDB dataset is captured by two aligned visible and thermal camera system, which has less challenges for VI-ReID. 206 identities with 2,060 images are selected for training and this procedure is repeated 10 times following [54], and calculate the average performance with visible to thermal and thermal to visible testing mode." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": "Evaluation Protocols. Following existing works, Cumulative Matching Characteristics (CMC), mean Average Precision (mAP) and mean inverse negative penalty (mINP) [57] are calculated as the evaluation metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 440, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 440, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 440, + 85 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "content": "The proposed framework is implemented with PyTorch. SDCL incorporates the feature extractor from TransReID [17] as the backbone network with augmented dual-contrastive learning [50]. The shallow patch embedding layers are constructed with IBN and CNN module [31]. The concatenation of shallow and deep class tokens is used to calculate the cosine similarity for retrieval. DBSCAN [10] is conducted to generate pseudo labels. The visible and infrared images are resized to " + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "inline_equation", + "content": "288 \\times 144" + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "content": " before entering the network. We sample 8 pseudo identities and 16 instances for each pseudo identity for each modality within one batch. Channel augmentation [56], random crop, horizontal flipping, and random erasing are adopted for data augmentation. We adopt the SGD optimizer with the initial learning rate of " + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "inline_equation", + "content": "3.5e - 4" + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "content": " to train the model. The model is trained in a total of 50 epochs. The CRA module is added in the last 20 epochs. The " + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "content": " in Eq. 20 is set to 0.5. The " + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "content": " in Eq. 12, Eq. 13, Eq. 16, and Eq. 17 is set to 0.9. The " + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 91, + 545, + 344 + ], + "type": "text", + "content": " in Eq. 25 and Eq. 26 are set to 20 for SYSU-MM01 and 10 for RegDB, respectively. The contrastive learning settings follow [50]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 355, + 529, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 355, + 529, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 529, + 369 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 375, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 411 + ], + "type": "text", + "content": "In Table 1, our SDCL is compared with supervised and unsupervised VI-ReID methods on two benchmarks including SYSU-MM01 and RegDB." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "text", + "content": "Comparison with Unsupervised Methods. As reported in Table 1, our method exhibits superior performance compared to the current advanced unsupervised methods. More precisely, our SDCL attains a remarkable " + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "inline_equation", + "content": "64.49\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "inline_equation", + "content": "86.91\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "text", + "content": " rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. Compared with the best current method GUR [51], our method also exceeds the about " + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 556 + ], + "type": "text", + "content": " rank-1 accuracy on SYSU-MM01 (all search) and RegDB (visible to infrared), respectively. In comparison to DPIS [35], a uni-semi-supervised method employing visible annotations for training, our SDCL still continues to maintain a leading position." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Comparison with Supervised Methods. Our comparison further includes 14 well-known supervised methods for reference. The comparisons with these methodologies unequivocally illustrate that our SDCL framework outperforms several supervised methods, including AGW [57], MSO [12], and DDAG [54], and achieves a competitive performance compared with MCLNet[16]. These considerable gains benefit from the insightful design of a shallow-deep collaborative learning framework. Our method facilitates the cultivation of modality-invariant features through the synergistic collaboration of shallow and deep features. Guided by our innovative solutions, SDCL surpasses prevailing unsupervised methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16875" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 541, + 453 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 541, + 453 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 541, + 453 + ], + "type": "table", + "html": "
MethodsVenueSYSU-MM01RegDB
All SearchIndoor SearchVisible to InfraredInfrared to Visible
r1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
SupervisedDDAG [54]ECCV-2054.7553.0239.6261.0267.9862.6169.3463.4649.2468.0661.8048.62
AGW [57]TPAMI-2147.5047.6535.3054.1762.9759.2370.0566.3750.1970.4965.9051.24
CA [56]ICCV-2169.8866.8953.6176.2680.3776.7985.0379.1465.3384.7577.8261.56
MPANet [47]CVPR-2170.5868.24-76.7480.95-82.880.7-83.780.9-
MSO [12]MM-2158.7056.42-63.0970.31-73.666.9-74.667.5-
AGM [26]MM-2169.6366.1152.2474.6878.3074.0088.4081.4568.5185.3481.1965.76
MCLNet [16]ICCV-2165.4061.9847.3972.5676.5872.1080.3173.0757.3975.9369.4952.63
SMCL [43]ICCV-2167.3961.78-68.8475.56-83.9379.83-83.0578.57-
FMCNet[60]CVPR-2266.3462.51-68.1574.09-89.1284.43-88.3883.86-
MAUM [27]CVPR-2271.6868.79-76.9781.94-87.8785.09-86.9584.34-
DEEN [61]CVPR-2374.771.8-80.383.3-91.185.1-89.583.4-
PMCM [34]IJCAI-2375.5471.16-81.5284.33-93.0989.57-91.4487.15-
PartMix [20]CVPR-2377.7874.62-81.5284.38-84.9382.52-85.6682.27-
SGIEL [20]CVPR-2377.1272.33-82.0782.95-95.3589.98-97.5791.41-
UnsupervisedSPCL [14]NIPS-2018.3719.3910.9926.8336.4233.0513.5914.8610.3611.7013.5610.09
MMT [13]ICLR-2021.4721.5311.5022.7931.5027.6625.6826.5119.5624.4225.5918.66
IICS [49]CVPR-2114.3915.748.4115.9124.8722.159.179.946.409.119.906.45
CAP [38]AAAI-2116.8215.717.0224.5730.7426.159.7111.568.7410.2111.347.92
Cluster Contrast [9]arXiv-2120.1622.0012.9723.3334.0130.8811.7613.889.9411.1412.998.99
ICE [3]ICCV-2120.5420.3910.2429.8138.3534.3212.9815.6411.9112.1814.8210.6
PPLR [6]CVPR-2211.9812.254.9712.7120.8117.6110.3011.948.1010.3911.237.04
OTLA [37]ECCV-2229.927.1-29.838.8-32.929.7-32.128.6-
H2H [24]TIP-2130.1529.40----23.8118.87----
ADCA [50]MM-2245.5142.7328.2950.6059.1155.1767.2064.0552.6768.4863.8149.62
DPIS [35]ICCV-2358.455.6-63.070.0-62.353.2-61.552.7-
CHCR [33]TCSVT-2359.4759.14----69.3164.74-69.9665.87-
DOTLA [5]MM-2350.3647.3632.4053.4761.7357.3585.6376.7161.5882.9174.9758.60
MBCCM [4]MM-2353.1448.1632.4155.2161.9857.1383.7977.8765.0482.8276.7461.73
PGM [48]CVPR-2357.2751.7834.9656.2362.7458.1369.4865.41-69.8565.17-
GUR [51]ICCV-2363.5161.6347.9371.1176.2372.5773.9170.2358.8875.0069.9456.21
SDCL (ours)-64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
", + "image_path": "36531d9d780d9efb107a8aef664608f0ebc5a30b9733f6c42da8511d8f27ef87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 476, + 544, + 570 + ], + "blocks": [ + { + "bbox": [ + 52, + 454, + 541, + 465 + ], + "lines": [ + { + "bbox": [ + 52, + 454, + 541, + 465 + ], + "spans": [ + { + "bbox": [ + 52, + 454, + 541, + 465 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-arts on SYSU-MM01 and RegDB. Rank at " + }, + { + "bbox": [ + 52, + 454, + 541, + 465 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 52, + 454, + 541, + 465 + ], + "type": "text", + "content": " accuracy(%) , mAP (%) and mINP (%) are reported." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 476, + 544, + 570 + ], + "lines": [ + { + "bbox": [ + 48, + 476, + 544, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 544, + 570 + ], + "type": "table", + "html": "
SYSU-MM01RegDB
ComponentsAll SearchIndoor SearchVisible to InfraredInfrared to Visible
IndexBaselineLids+LrsidCNLCRAr1mAPmINPr1mAPmINPr1mAPmINPr1mAPmINP
1✓49.5548.7036.9253.8560.9657.0563.1658.1242.5961.7056.7641.65
2✓✓53.2150.1440.8060.1565.2464.0269.9763.3448.9868.5463.1948.06
3✓✓✓55.2754.7941.6164.3871.1567.4076.6067.9751.3875.6468.4650.38
4✓✓✓59.4359.1848.8167.4874.3170.3080.5373.6558.1379.1370.1853.30
5✓✓✓✓✓64.4963.2451.0671.3776.9073.5086.9178.9262.8385.7677.2559.57
", + "image_path": "03b8730302b8916d1fed35d3632e3ad775d122e804352e4802bf7bde620a3569.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "lines": [ + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "spans": [ + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "text", + "content": "Table 2. Ablation studies on the SYSU-MM01 and RegDB. Rank at " + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "text", + "content": " accuracy " + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "inline_equation", + "content": "\\left( \\% \\right)" + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "text", + "content": " ,mAP " + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "inline_equation", + "content": "\\left( \\% \\right)" + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "text", + "content": " and mINP " + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "inline_equation", + "content": "\\left( \\% \\right)" + }, + { + "bbox": [ + 74, + 571, + 520, + 582 + ], + "type": "text", + "content": " are reported." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 48, + 591, + 140, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 140, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 140, + 604 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 613, + 286, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 648 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 648 + ], + "type": "text", + "content": "In this subsection, we meticulously conduct ablation experiments to validate the individual efficacy of each module. The results are reported in Table 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 651, + 286, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 651, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 651, + 286, + 699 + ], + "type": "text", + "content": "Baseline denotes the augmented dual-contrastive learning framework [50] with a dual-path transformer architecture. The optimization for the network only uses deep features, i.e., training the model with " + }, + { + "bbox": [ + 48, + 651, + 286, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{id}^{vd} + \\mathcal{L}_{id}^{rd}" + }, + { + "bbox": [ + 48, + 651, + 286, + 699 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 701, + 286, + 712 + ], + "type": "text", + "content": "Effectiveness of Deep and Shallow Contrastive Learn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 591, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 591, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 591, + 545, + 663 + ], + "type": "text", + "content": "ing. The incorporation of " + }, + { + "bbox": [ + 306, + 591, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}id^{vs} + \\mathcal{L}id^{rs}" + }, + { + "bbox": [ + 306, + 591, + 545, + 663 + ], + "type": "text", + "content": " yields a " + }, + { + "bbox": [ + 306, + 591, + 545, + 663 + ], + "type": "inline_equation", + "content": "4\\% - 7\\%" + }, + { + "bbox": [ + 306, + 591, + 545, + 663 + ], + "type": "text", + "content": " enhancement of rank-1 accuracy on SYSU-MM01 and RegDB datasets compared to the baseline. Despite shallow contrastive learning primarily optimizing the ReID model for feature learning within intra-modality, the findings underscore its capacity to augment cross-modality retrieval." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of CNL. Compared with the baseline, the CNL improves the performance of " + }, + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "6\\% - 10\\%" + }, + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "type": "text", + "content": " on SYSUMM01 and RegDB datasets. The main gain is achieved by the design of capturing credible homogeneous and hetero-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16876" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 164, + 173 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 164, + 173 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 164, + 173 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 164, + 173 + ], + "type": "image", + "image_path": "82d50b90cfd8f8318601d1b8f5377f8b1e5bd9e44391585d0b96c9399882f90d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "text", + "content": "Figure 3. Evaluation of hyper-parameters " + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "text", + "content": " (left) and " + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "text", + "content": " (right). The results are based on all search mode of SYSU-MM01 dataset. Rank-1 accuracy " + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "text", + "content": " and mAP " + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 172, + 287, + 205 + ], + "type": "text", + "content": " are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 71, + 273, + 173 + ], + "blocks": [ + { + "bbox": [ + 170, + 71, + 273, + 173 + ], + "lines": [ + { + "bbox": [ + 170, + 71, + 273, + 173 + ], + "spans": [ + { + "bbox": [ + 170, + 71, + 273, + 173 + ], + "type": "image", + "image_path": "52ce65ba4319aeb4e93719fb1ad2c644f447ab1e7deb96e68e5ad28b9cf3fa93.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 214, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 262 + ], + "type": "text", + "content": "genuine neighborhoods with the shallow and deep collaborative consistency, which is a dynamic optimization, reducing the impact of label noise in clustering and enhancing the robustness against multiple discrepancies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "type": "text", + "content": "Effectiveness of CRA. In contrast to the baseline, the CRA module remarkably improves the accuracy. CRA establishes a linkage between the labels of two modalities through collaborative shallow and deep ranking relations, constituting a global association across the training set and effectively bolstering modality-invariant features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 339, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 448 + ], + "type": "text", + "content": "Effectiveness of SDCL. The amalgamation of CNL and CRA notably enhances Rank-1 accuracy across diverse settings. This observation substantiates the rationale behind SDCL. CRA provides cross-modality supervision at the beginning of each training epoch in an off-line manner, while CNL dynamically learns cross-modality features during training iteration in an online manner. By combining off-line label association and online feature learning, SDCL learns better representation from different perspectives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 462, + 149, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 149, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 149, + 475 + ], + "type": "text", + "content": "4.5. Further Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": "Hyper-parameter Analysis for " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": ". We explore the influence of hyper-parameters " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " in Eq. 20 and " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " in Eq. 12, Eq. 13, Eq. 16, and Eq. 17, as shown in Fig. 3. The " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " governs the balance of " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " in deep and shallow neighbor learning. When " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{s} = 0" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " is excluded. When " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{base} = 0.5" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": ", the baseline method achieves a balance in deep and shallow neighbor learning. The " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " is used to control the positive neighbor selection range in CNL. Setting " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\gamma = 0" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": " to 0 renders neighbor selection ineffective. When " + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\gamma = 0.9" + }, + { + "bbox": [ + 46, + 483, + 287, + 604 + ], + "type": "text", + "content": ", the CNL significantly reinforces the cross-modality learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Visualization. We perform feature space (t-SNE [36] map) and similarity distribution visualization for SDCL, as presented in Fig. 4. As evident in (a) and (b), SDCL brings infrared and visible positive sample points closer together, and effectively enhances the separation of cross-modality positive/negative distributions. This attests to the efficacy of our method in addressing modality discrepancies. Additionally, the visualization of shallow and deep feature map is presented in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 76, + 421, + 171 + ], + "blocks": [ + { + "bbox": [ + 317, + 76, + 421, + 171 + ], + "lines": [ + { + "bbox": [ + 317, + 76, + 421, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 76, + 421, + 171 + ], + "type": "image", + "image_path": "48281d164bbfefded1b4a737da80543622f93a402eff34ad87560e760d6737e7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 76, + 533, + 171 + ], + "blocks": [ + { + "bbox": [ + 427, + 76, + 533, + 171 + ], + "lines": [ + { + "bbox": [ + 427, + 76, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 427, + 76, + 533, + 171 + ], + "type": "image", + "image_path": "fa6d16b9d00ac1e85bac80572a898d0ba3c893bf1f69abae2e8065d565c8aef3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 177, + 421, + 287 + ], + "blocks": [ + { + "bbox": [ + 316, + 177, + 421, + 287 + ], + "lines": [ + { + "bbox": [ + 316, + 177, + 421, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 421, + 287 + ], + "type": "image", + "image_path": "5b9a04ad59ffbb44d8858199991ff37f5c20b0e3b6fa51c0fc973bc7a688b179.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 293, + 391, + 304 + ], + "lines": [ + { + "bbox": [ + 349, + 293, + 391, + 304 + ], + "spans": [ + { + "bbox": [ + 349, + 293, + 391, + 304 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 304, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 348 + ], + "type": "text", + "content": "Figure 4. The t-SNE (first row) and similarity distribution (second row) visualization of randomly selected identities. In the t-SNE visualization, identity is denoted by color, where circles represent the visible modality and triangles denote the infrared modality." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 427, + 177, + 533, + 287 + ], + "blocks": [ + { + "bbox": [ + 427, + 177, + 533, + 287 + ], + "lines": [ + { + "bbox": [ + 427, + 177, + 533, + 287 + ], + "spans": [ + { + "bbox": [ + 427, + 177, + 533, + 287 + ], + "type": "image", + "image_path": "6e5168ecd89e06924d0228c49e44d4b4172a7de0934bb50e1add316e4a3f8fb7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 293, + 499, + 304 + ], + "lines": [ + { + "bbox": [ + 463, + 293, + 499, + 304 + ], + "spans": [ + { + "bbox": [ + 463, + 293, + 499, + 304 + ], + "type": "text", + "content": "(b) SDCL" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 351, + 378, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 378, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 378, + 363 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 371, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 587 + ], + "type": "text", + "content": "This paper investigates an extremely important and challenging problem, namely the unsupervised visible infrared person re-identification (US-VI-ReID) task, alleviating the reliance on expensive cross-modality annotation. To remedy the issues of multiple discrepancies, we propose a comprehensive Shallow-Deep Collaborative Learning (SDCL) framework based on the transformer architecture, incorporating Collaborative Neighbor Learning (CNL), and Collaborative Ranking Association (CRA) module. SDCL significantly facilitates the learning of robust representation, effectively countering the cross-modality discrepancy through the collaboration of shallow patch embeddings and deep modality-shared features and capturing more discriminative representations for cross-modality retrieval. Experiments on two public benchmarks substantiate that our approach surpasses existing methodologies by a substantial margin. Moreover, it even outshines certain supervised counterparts, propelling VI-ReID toward real-world deployment." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 597, + 403, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 403, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 403, + 611 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "This work is partially supported by the National Natural Science Foundation of China under Grants (62071338, 62176188, 62272354, 62306215), the Key Research and Development Program of Hubei Province (2022BCA009,2022BAD175), the Special Fund of Hubei Luojia Laboratory (220100015), and the Interdisciplinary Innovative Talents Foundation from Renmin Hospital of Wuhan University." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16877" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Yimin Liu, and Jianguo Jiang. Saliency and granularity: Discovering temporal coherence for video-based person re-identification. IEEE TCSVT, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 168 + ], + "type": "text", + "content": "[2] Cuiqun Chen, Mang Ye, Meibin Qi, and Bo Du. Sketchtrans: Disentangled prototype learning with transformer for sketchphoto recognition. IEEE TPAMI, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "text", + "content": "[3] Hao Chen, Benoit Lagadec, and Francois Bremond. Ice: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "type": "text", + "content": "[4] De Cheng, Lingfeng He, Nannan Wang, Shizhou Zhang, Zhen Wang, and Xinbo Gao. Efficient bilateral cross-modality cluster matching for unsupervised visible-infrared person reid. In ACM MM, 2023. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "text", + "content": "[5] De Cheng, Xiaojian Huang, Nannan Wang, Lingfeng He, Zhihui Li, and Xinbo Gao. Unsupervised visible-infrared person Reid by collaborative learning with neighbor-guided label refinement. ACM MM, 2023. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 292, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 323 + ], + "type": "text", + "content": "[6] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 325, + 286, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 286, + 356 + ], + "type": "text", + "content": "[7] Zhenyu Cui, Jiahuan Zhou, and Yuxin Peng. Dma: Dual modality-aware alignment for visible-infrared person re-identification. IEEE TIFS, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "type": "text", + "content": "[8] Yongxing Dai, Jun Liu, Yan Bai, Zekun Tong, and Ling-Yu Duan. Dual-refinement: Joint label and feature refinement for unsupervised domain adaptive person re-identification. IEEE TIP, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 286, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 445 + ], + "type": "text", + "content": "[9] Zuozhuo Dai, Guangyuan Wang, Siyu Zhu, Weihao Yuan, and Ping Tan. Cluster contrast for unsupervised person re-identification. arxiv 2021. arXiv preprint arXiv:2103.11568, 2021. 2.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 286, + 479 + ], + "type": "text", + "content": "[10] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, 1996. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 286, + 512 + ], + "type": "text", + "content": "[11] Jiawei Feng, Ancong Wu, and Wei-Shi Zheng. Shape-erased feature learning for visible-infrared person re-identification. In CVPR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "text", + "content": "[12] Yajun Gao, Tengfei Liang, Yi Jin, Xiaoyan Gu, Wu Liu, Yidong Li, and Congyan Lang. Mso: Multi-feature space joint optimization network for rgb-infrared person re-identification. In ACM MM, 2021. 1, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "text", + "content": "[13] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual mean-teaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. arXiv preprint arXiv:2001.01526, 2020. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 635 + ], + "type": "text", + "content": "[14] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. NeurIPS, 2020. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "type": "text", + "content": "[15] Jianyang Gu, Weihua Chen, Hao Luo, Fan Wang, Hao Li, Wei Jiang, and Weijie Mao. Multi-view evolutionary training for unsupervised domain adaptive person re-identification. IEEE TIFS, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[16] Xin Hao, Sanyuan Zhao, Mang Ye, and Jianbing Shen. Cross-modality person re-identification via modality confusion and center aggregation. In ICCV, 2021. 6, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[18] Wenke Huang, Mang Ye, Zekun Shi, and Bo Du. Generalizable heterogeneous federated cross-correlation and instance similarity learning. IEEE TPAMI, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[19] Wenke Huang, Mang Ye, Zekun Shi, Guancheng Wan, He Li, Bo Du, and Qiang Yang. A federated learning for generalization, robustness, fairness: A survey and benchmark. arXiv, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[20] Minsu Kim, Seungryong Kim, Jungin Park, Seongheon Park, and Kwanghoon Sohn. Partmix: Regularization strategy to learn part discovery for visible-infrared person re-identification. In CVPR, 2023. 1, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "text", + "content": "[21] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In AAAI, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "text", + "content": "[22] Shuang Li, Jiaxu Leng, Ji Gan, Mengjingcheng Mo, and Xinbo Gao. Shape-centered representation learning for visible-infrared person re-identification. arXiv preprint arXiv:2310.17952, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[23] Shuang Li, Fan Li, Jinxing Li, Huafeng Li, Bob Zhang, Dapeng Tao, and Xinbo Gao. Logical relation inference and multiview information interaction for domain adaptation person re-identification. IEEE TNNLS, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[24] Wenqi Liang, Guangcong Wang, Jianhuang Lai, and Xiaohua Xie. Homogeneous-to-heterogeneous: Unsupervised learning for rgb-infrared person re-identification. IEEE TIP, 2021. 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[25] Shan Lin, Chang-Tsun Li, and Alex C. Kot. Multidomain adversarial feature generalization for person re-identification. IEEE TIP, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[26] Haojie Liu, Daoxun Xia, and Wei Jiang. Towards homogeneous modality learning and multi-granularity information exploration for visible-infrared person re-identification. IEEE Journal of Selected Topics in Signal Processing. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[27] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, 2022. 1, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[28] Yan Lu, Yue Wu, Bin Liu, Tianzhu Zhang, Baopu Li, Qi Chu, and Nenghai Yu. Cross-modality person re-identification with shared-specific feature transfer. In CVPR, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "type": "text", + "content": "[29] Zefeng Lu, Ronghao Lin, and Haifeng Hu. Modality and camera factors bi-disentanglement for nir-vis object re-identification. IEEE TIFS, 2023." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 624 + ], + "type": "text", + "content": "[30] Chuanchen Luo, Chunfeng Song, and Zhaoxiang Zhang. Learning to adapt across dual discrepancy for cross-domain person re-identification. IEEE TPAMI, 2022. 1, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[31] Hao Luo, Pichao Wang, Yi Xu, Feng Ding, Yanxin Zhou, Fan Wang, Hao Li, and Rong Jin. Self-supervised pre-training for transformer-based person re-identification. arXiv preprint arXiv:2111.12084, 2021. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[32] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 2017. 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16878" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Zhiqi Pang, Chunyu Wang, Lingling Zhao, Yang Liu, and Gaurav Sharma. Cross-modality hierarchical clustering and refinement for unsupervised visible-infrared person re-identification. IEEE TCSVT, 2023. 2, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "text", + "content": "[34] Zhihao Qian, Yutian Lin, and Bo Du. Visible-infrared person re-identification via patch-mixed cross-modality learning. *IJCAI*, 2023. 1, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 207 + ], + "type": "text", + "content": "[35] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, 2023. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 232 + ], + "type": "text", + "content": "[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 287, + 277 + ], + "type": "text", + "content": "[37] Jiangming Wang, Zhizhong Zhang, Mingang Chen, Yi Zhang, Cong Wang, Bin Sheng, Yanyun Qu, and Yuan Xie. Optimal transport for label-efficient visible-infrared person re-identification. ECCV, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "text", + "content": "[38] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 357 + ], + "type": "text", + "content": "[39] Yiming Wang, Guanqiu Qi, Shuang Li, Yi Chai, and Huafeng Li. Body part-level domain alignment for domain-adaptive person re-identification with transformer framework. IEEE TIFS, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 287, + 403 + ], + "type": "text", + "content": "[40] Yadi Wang, Hongyun Zhang, Duoqian Miao, and Witold Pedrycz. Multi-granularity re-ranking for visible-infrared person re-identification. CAAI Transactions on Intelligence Technology, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "text", + "content": "[41] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In CVPR, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 450, + 287, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 494 + ], + "type": "text", + "content": "[42] Zheng Wang, Zhixiang Wang, Yinqiang Zheng, Yang Wu, Wenjun Zeng, and Shin'ichi Satoh. Beyond intra-modality: A survey of heterogeneous person re-identification. arXiv preprint arXiv:1905.10048, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 496, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 496, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 496, + 287, + 529 + ], + "type": "text", + "content": "[43] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Syncretic modality collaborative learning for visible infrared person re-identification. In ICCV, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 531, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 552 + ], + "type": "text", + "content": "[44] Ancong Wu, Wei-Shi Zheng, and Jian-Huang Lai. Robust depth-based person re-identification. IEEE TIP, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "text", + "content": "[45] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In ICCV, 2017. 1, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[46] Dongming Wu, Mang Ye, Gaojie Lin, Xin Gao, and Jianbing Shen. Person re-identification by context-aware part attention and multi-head collaborative learning. IEEE TIFS, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[47] Qiong Wu, Pingyang Dai, Jie Chen, Chia-Wen Lin, Yongjian Wu, Feiyue Huang, Bineng Zhong, and Rongrong Ji. Discover cross-modality nuances for visible-infrared person re-identification. In CVPR, 2021. 1, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[48] Zesen Wu and Mang Ye. Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning. In CVPR, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 701 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, 2021. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[50] Bin Yang, Mang Ye, Jun Chen, and Zesen Wu. Augmented dual-contrastive aggregation learning for unsupervised visible-infrared person re-identification. In ACM MM, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[51] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In ICCV, 2023. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[52] Bin Yang, Jun Chen, and Mang Ye. Top-k visual tokens transformer: Selecting tokens for visible-infrared person re-identification. In ICASSP, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[53] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE TIFS, pages 407–419, 2019. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "text", + "content": "[54] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In ECCV, 2020. 1, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[55] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE TIFS, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[56] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "text", + "content": "[57] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 2021. 1, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "text", + "content": "[58] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE TPAMI, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[59] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[60] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In CVPR, 2022. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "text", + "content": "[61] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, 2023. 1, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[62] Yiyuan Zhang, Yuhao Kang, Sanyuan Zhao, and Jianbing Shen. Dual-semantic consistency learning for visible-infrared person re-identification. IEEE TIFS, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "text", + "content": "[63] Yafei Zhang, Yongzeng Wang, Huafeng Li, and Shuang Li. Cross-compatible embedding and semantic consistent feature construction for sketch re-identification. In ACM MM, 2022. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 669, + 545, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 701 + ], + "type": "text", + "content": "[64] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In ICCV, 2015. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16879" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_content_list.json b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..31852afe86e901a6412646ad764018959e653f7b --- /dev/null +++ b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_content_list.json @@ -0,0 +1,1651 @@ +[ + { + "type": "text", + "text": "ShapeMatcher: Self-Supervised Joint Shape Canonicalization, Segmentation, Retrieval and Deformation", + "text_level": 1, + "bbox": [ + 197, + 161, + 774, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yan Di $^{1*}$ , Chenyangguang Zhang $^{2*}$ , Chaowei Wang $^{3*}$ , Ruida Zhang $^{2}$ , Guangyao Zhai $^{1}$ , Yanyan Li $^{1}$ , Bowen Fu $^{2}$ , Xiangyang Ji $^{2}$ and Shan Gao $^{†3}$", + "bbox": [ + 127, + 227, + 844, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Technical University of Munich, 2Tsinghua University, 3Northwestern Polytechnical University", + "bbox": [ + 140, + 260, + 831, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shangbuhuan13@gmail.com, gaoshan@nwpu.edu.cn", + "bbox": [ + 307, + 279, + 663, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 340, + 328, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present ShapeMaker, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonicalize the object by extracting point-wise affine-invariant features, disentangling inherent structure of the object with its pose and size. These learned features are then leveraged to predict semantically consistent part segmentation and corresponding part centers. Next, our lightweight retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. The key insight of ShapeMaker is the simultaneous training of the four highly-associated processes: canonicalization, segmentation, retrieval, and deformation, leveraging cross-task consistency losses for mutual supervision. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that ShapeMaker surpasses competitors by a large margin. Code is released at https://github.com/Det1999/ShapeMaker.", + "bbox": [ + 111, + 369, + 473, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 726, + 233, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been a notable surge in research interest focused on generating high-quality 3D models from scans of complex scenes [6, 15, 16, 21, 35, 49, 73]. This technology encourages extensive applications in both artistic creation [56, 57], robotics [66, 67] and 3D scene perception [41, 69]. Existing methods [41, 54] typically directly utilize deep neural networks to reconstruct 3D", + "bbox": [ + 111, + 750, + 470, + 847 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75860b2b8ebd284198f99c6a65466feff5cafc810e2e938a9b587a757931ea73.jpg", + "image_caption": [ + "Figure 1. Illustration of ShapeMatcher. Objects obtained from real-world scans are typically noisy, partial and exhibit various poses, making it challenging to conduct an effective $R \\& D$ process (Red 'X' on the left). To address this issue, we propose Shape-Matcher that first canonicalizes the objects and then segments them into semantic parts, facilitating R&D processes (Green '√' on the right)." + ], + "image_footnote": [], + "bbox": [ + 503, + 339, + 859, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models from imperfect scans. However, the presence of noise and occlusions poses a significant challenge in accurately capturing fine-grained geometric structures. To overcome this, Retrieval and Deformation (R&D) techniques [10,28,40,46,56-58,64] have been developed. These methods generally involve two key steps: first, identifying the most geometrically similar source shape from a precurated 3D database; and second, deforming the retrieved shape to achieve precise alignment with the target input. The R&D approach is particularly effective in producing 3D models enriched with fine details from source shapes.", + "bbox": [ + 496, + 627, + 859, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing R&D methods usually encounter two primary challenges that make them susceptible to noise, occlusion and pose variations, and difficult to be practically utilized. 1) Most R&D techniques [10, 17, 28, 40, 46, 56-58, 64] operate under the assumption that target shapes are aligned in a pre-processed canonical space. Typ", + "bbox": [ + 496, + 782, + 859, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 138, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 259, + 42, + 781, + 56 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 338, + 56, + 702, + 70 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 310, + 71, + 732, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors with equal contributions. † Corresponding author.", + "bbox": [ + 133, + 853, + 423, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21017", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ically, these methods are trained and tested on datasets where shapes have been manually adjusted to this canonical state. However, when these methods are deployed in real-world settings, they necessitate either manual alignment of scanned objects or the use of additional pose estimation networks [14, 18, 48, 71, 72]. Such procedures are not only time-intensive and laborious but also prone to yielding inconsistent results. This limitation significantly impedes the direct application of these methods in real-world scenarios. 2) Previous methods [56, 57] do not design specially for dealing with partially-observed shapes, making it difficult to handle occluded objects. Although U-RED [17] considers the partial target shapes as input in the R&D process, it directly encodes the shape as a global embedding, which is not robust when dealing with significant occlusion.", + "bbox": [ + 109, + 126, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned two challenges, in this paper, we present ShapeMatcher, a novel framework that extends traditional R&D pipeline to joint self-supervised learning of object canonicalization, segmentation, retrieval and deformation. Our core contribution lies in that the four highly-associated processes can be trained simultaneously and supervise each other via constructing several cross-task consistency losses (Fig. 1). Specifically, given a partially-observed object scan in an arbitrary pose, ShapeMatcher processes the objects in four steps. First, we follow [30], which is based on Vector Neurons [13], to extract SE(3)-invariant point-wise features by progressively separating translation and rotation. We further follow [8] to normalize the features to disassociate the object scale. Until here, we successfully obtain affine-invariant point-wise features by disentangling object's inherent structure with its pose and size. This facilitates the Canonicalization of the observed object based on these intrinsic characteristics (Fig. 2 (B)). Then we predict semantically consistent part segmentation and corresponding part centers by feeding the learned features into our Segmentation module (Fig. 2 (C)). Based on the part segmentation, in the Retrieval module (Fig. 2 (D)), we aggregate features within each part and collect them together as a comprehensive retrieval token of the object. For partial objects, we introduce a region-weighted strategy, which assigns a weight to each part according to the point inside it. Parts with more points are assigned higher weights during retrieval, which is proved to be robust to occlusions. We compare the tokens of the target object with each shape in the pre-constructed database to identify the most geometrically similar (most similar tokens) source shape. In the final Deformation module (Fig. 2 (E)), the retrieved source shape is deformed to tightly match the target object via part center guided neural cage deformation [64].", + "bbox": [ + 109, + 333, + 472, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are:", + "bbox": [ + 130, + 803, + 388, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce ShapeMatcher, a novel self-supervised framework for joint shape canonicalization, segmentation, retrieval and deformation, handling partial target", + "bbox": [ + 130, + 824, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inputs under arbitrary poses. Extensive experiments on the synthetic and real-world datasets demonstrate that ShapeMatcher surpasses existing state-of-the-art approaches by a large margin.", + "bbox": [ + 527, + 126, + 857, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate that the four highly-associated tasks: canonicalization, segmentation, retrieval and deformation, can be effectively trained simultaneously and supervise each other via constructing consistencies.", + "- We develop the region-weighted retrieval method to mitigate the impact of occlusions in the R&D process." + ], + "bbox": [ + 516, + 188, + 857, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 291, + 635, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Shape Representation. The compact representation of 3D shapes in latent space, based on deep learning, has been a focal point for many researchers. Some attempts, such as [9, 27, 36, 37, 42, 45, 61, 70], employ neural networks to construct an implicit function, while others [1, 38, 52, 62, 68] directly model the shape of objects explicitly using generative models. Another common architecture in 3D shape representation learning, as seen in [12, 29, 44, 60, 63], is to use an encoder-decoder approach to generate latent representation vectors for various shapes. Although these methods have demonstrated impressive representation performance, they often struggle to generate fine-grained shapes when dealing with occlusion and noise.", + "bbox": [ + 496, + 314, + 857, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CAD Model Retrieval and Deformation. Retrieval and Deformation (R&D) methods lead another way to recover fine-grained geometric structures. Previous works directly retrieve the most similar CAD models by comparing the similarity of expression vectors in either descriptor space [5,46] or the latent space of neural networks [4,10,22,34]. Considering the subsequent deformation error, recent efforts introduce deformation-aware embeddings [56] or proposed new optimization objectives [24] to better capture the fine structure of deformed target objects. Nevertheless, these methods yield deteriorated performance when facing partial and pose-agnostic target shapes in real world. [17] achieves an one-to-many retrieval module for addressing the issue caused by partial observations, however, it receives canonicalized target shapes as input, which limit its applicability facing pose-agnostic target shapes in real world. As the retrieved models often exhibit some deviation from the target shape, the deformation module is used to minimize this discrepancy. Traditional approaches [20, 23, 47] aim to fit the target shape by directly optimizing the deformed shape. Neural network based techniques attempt to learn a set of deformation priors from a database of models. They represent deformations as volume warping [25,32], cage deformations [64], vertex offsets [58], or flows [28,65]. These methods typically constrain two shapes are aligned in the same coordinate system, making them challenging to apply in real-world scenarios.", + "bbox": [ + 496, + 493, + 857, + 864 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21018", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SO(3)-Equivariant Methods. An increasing body of work [2,19,31,55,59] has initiated research on SO(3) equivariance. These efforts are mostly based on steerable convolutional kernels [33]. On the other hand, another set of works achieves equivariance through pose estimation. [43] estimates the object's pose to factor out SO(3) transformations, achieving approximate equivariance. While [51] learns pose estimation in a fully unsupervised manner, the equivariant backbone they employ [50] achieves equivariance primarily through data augmentation, leading to limited generalization. In this paper, we employ Vector Neural Multi-Layer Perceptron [13] as the backbone to get neural invariant features for object canonicalization. It achieves SO(3) equivariance by lifting traditional scalar neurons to vector neurons.", + "bbox": [ + 111, + 125, + 473, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 112, + 344, + 194, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. ShapeMatcher consists of 4 modules, corresponding to the 4 highly-associated tasks. Each of the first 3 modules: Canonicalization, Segmentation and Retrieval modules have two parallel branches, one for complete point cloud (in orange background) and the other for partial object input (in blue background). As shown in Fig. 2, given a partial target shape $S_{tgt} \\in \\mathbb{R}^{N \\times 3}$ in an arbitrary pose, in the Canonicalization module, we progressively decouples its inherent shape with rotation $R_{tgt} \\in SO(3)$ , translation $T_{tgt} \\in \\mathbb{R}^3$ and the 3D metric size $s_{tgt} \\in \\mathbb{R}^3$ via VN-MLP [8, 13, 30], yielding the affine-invariant point-wise features $F_{tgt}$ . The object can then be canonicalized via inverse transformation based on intrinsics $\\{R_{tgt}, T_{tgt}, s_{tgt}\\}$ . In the Segmentation module, $F_{tgt}$ is fed into a 4-layer MLP network to predict $M$ parts and corresponding part centers $\\{K_{tgt}^1, K_{tgt}^2, ..., K_{tgt}^M\\}$ . The segmentation is semantically consistent across each category and thus can be matched and compared for R&D. In the Retrieval module, inside each region $M^i$ , we aggregate the features of all points inside it as its retrieval token $Q^i$ . The retrieval token for the object is then represented as $\\mathbf{Q}_{\\mathrm{tg}} = \\{Q_{tgt}^1, Q_{tgt}^2, ..., Q_{tgt}^M\\}$ . Similarly, during training, we obtain the intrinsics $\\{R_{src}, T_{src}, s_{src}\\}$ , part centers $\\{K_{src}^1, K_{src}^2, ..., K_{src}^M\\}$ and retrieval tokens $\\mathbf{Q}_{\\mathrm{src}} = \\{Q_{src}^1, Q_{src}^2, ..., Q_{src}^M\\}$ via the branch for complete point cloud. By comparing $\\mathbf{Q}_{\\mathrm{tg}}$ and $\\mathbf{Q}_{\\mathrm{src}}$ of each source shape inside the database, we identify the most geometrically similar source shape $S_r$ . In the final Deformation module, $K_{tgt}$ and $K_{src}$ are leveraged to guide the neural cage deformation [64] to deform the retrieved $S_r$ towards $S_{tgt}$ , yiling $S_{src}^{df_m}$ .", + "bbox": [ + 111, + 367, + 473, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Canonicalization", + "text_level": 1, + "bbox": [ + 112, + 803, + 266, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (B), the Canonicalization module takes the target point cloud $S_{tgt}$ as input and disentangle the inherent structure of $S_{tgt}$ with the intrinsics", + "bbox": [ + 111, + 824, + 470, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\{R_{tgt}, T_{tgt}, s_{tgt}\\}$ , yielding a point-wise affine-invariant feature $F_{tgt}$ . Specifically, we follow VN-MLP [8, 13, 30] to first decouple translation via VNT [30] and then extract rotation via VNN [13, 30]. We further follow [8] to normalize the SE(3)-invariant features obtained above the remove the influence of scaling, yielding $F_{tgt}$ as follows,", + "bbox": [ + 498, + 125, + 857, + 209 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {t g t}, T _ {t g t}, F _ {t g t} ^ {*} = \\mathbf {V N - M L P} \\left(S _ {t g t}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 218, + 857, + 234 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t g t}, F _ {t g t} = \\text {n o r m a l i z e} \\left(F _ {t g t} ^ {*}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 255, + 856, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{tgt}^{*}$ denotes the SE(3)-invariant features and $F_{tgt}$ denotes the affine-invariant features. Thereby, the object can be canonicalized with intrinsics as,", + "bbox": [ + 498, + 276, + 857, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t g t} ^ {c} = s _ {t g t} R _ {t g t} S _ {t g t} + T _ {t g t} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 327, + 857, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S_{tgt}^{c}$ denotes the normalized and canonicalized shape of $S^{tgt}$ . During training, in order to ensure that $F_{tgt}$ fully encapsulate the geometric information of $S_{tgt}$ , we integrate a supplementary reconstruction branch which takes $F_{tgt}^{*}$ as input and reconstruct $S_{tgt}^{c}$ in the affine-invariant space [30]. Please refer to the Supplementary Material for details. For source shape $S_{src}$ from the database, we follow the same procedures to extract $F_{src}$ .", + "bbox": [ + 496, + 351, + 857, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Segmentation", + "text_level": 1, + "bbox": [ + 500, + 470, + 629, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the affine-invariant features $F_{tgt}$ , we segment the input point cloud $S_{tgt}$ into $M$ semantically consistent parts. We use a 4-layer MLP $\\Theta_l$ (Fig. 2 (C)) to predict a one-hot segmentation label for each point and use another 4-layer MLP $\\Theta_c$ to predict $M$ part centers $\\{K_{src}^1, K_{src}^2, \\dots, K_{src}^M\\}$ . Noteworthy, we don't need any ground truth annotations in this segmentation process. Our experiments show that the network can automatically learn semantically consistent segmentation solely through consistency supervision from the other three tasks. For source shape $S_{src}$ , we follow a similar process to obtain $\\{K_{src}^1, K_{src}^2, \\dots, K_{src}^M\\}$ .", + "bbox": [ + 496, + 491, + 857, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Retrieval", + "text_level": 1, + "bbox": [ + 500, + 651, + 598, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The retrieval network aims to identify the model $S_{src}$ from an existing database that bears the closest resemblance to the target object $S_{tgt}$ after deformation. Traditional methods [17, 57] directly extract the global features of objects for retrieval, which typically struggles with heavy occlusion since the global features are susceptible to noise and occlusion, and prone to producing erroneous retrieval results. In contrast, we employ a novel region-weighted retrieval method to explicitly encode independent and semantically consistent regions of the shape. This allows us to accurately handle partial shapes by identifying the visible regions to retrieve models most similar to the target.", + "bbox": [ + 496, + 672, + 857, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the part segmentation network $\\Theta_l$ takes $F_{tgt}$ as input to predict $M$ regions $F_{seg} = [C_1, C_2, C_3, \\dots, C_N]^\\top$ .", + "bbox": [ + 498, + 838, + 857, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21019", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f6fef8bdde9557f419481033fdd3cc5b156b3ee396d4175f767dadd913338958.jpg", + "image_caption": [ + "Figure 2. The pipeline of ShapeMatcher. Given a target point cloud obtained from a single-view scan and a pre-established database (A), ShapeMatcher generates the fine-grained reconstruction result using the joint 4 modules including Canonicalization (B), Segmentation (C), Retrieval (D) and Deformation (E), where the first three contains the partial branch for target processing and the full branch for source processing. Specifically, the target and source inputs are first canonicalized into the same affine-invariant space (B). Then, the semantic-consistent region segmentation is yielded from the affine-invariant features (C). The segmented regions are fed to the region-weight retrieval module (C) and the part center guided neural cage deformation module (E) for occlusion-robust R&D process. During training, the partial-full consistency losses (F) are enforced for the two branches." + ], + "image_footnote": [], + "bbox": [ + 135, + 125, + 436, + 412 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/849bcfa63bab6cd20b47261ff51c697d1e8394d7a6a4f4d131345f8c6d5e5365.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 125, + 678, + 410 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48f3f357bba1587a2666c132a58a1a3314a714b1f74abdc338be666efec4650a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 125, + 836, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of $S_{tgt}^{c}$ , where $F_{seg} \\in \\mathbb{R}^{N \\times M}$ , $C_i \\in \\mathbb{R}^M$ represents the probability of point $i$ belonging to each part center. Then we use another 4-layer feature aggregator $\\Theta_f$ (Fig. 2 (D)) to extract the retrieval tokens $Q$ of all parts as follows,", + "bbox": [ + 111, + 520, + 473, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c l s} = F _ {s e g} ^ {\\top} * \\Theta_ {f} (F _ {t g t}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 584, + 472, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {t g t} = F _ {c l s} / \\left(\\sum_ {n = 1} ^ {N} F _ {s e g} ^ {(n)}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 619, + 470, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{tgt} \\in \\mathbb{R}^{M \\times C}$ contains the $C$ -dimensional retrieval tokens for all the $M$ regions. Here we employ a soft assignment strategy where each point inside $S_{tgt}$ is estimated $M$ values describing the probabilities belonging to each of the $M$ parts. Therefore, we first aggregate features $F_{cls}$ on all points belonging to each part and then normalize $F_{cls}$ using the sum of probabilities of points in each part, as in Eq. 4 and Eq. 5. Following a similar strategy, we can obtain $Q_{src}$ for each source model in the pre-curated database. We just need to compare $Q_{tgt}$ with the retrieval tokens $Q_{src}$ of all source shapes using the weighted $\\mathcal{L}_1$ distance,", + "bbox": [ + 112, + 661, + 473, + 815 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD i s = \\omega \\sum \\mathcal {L} _ {1} \\left(Q _ {t g t} - Q _ {s r c}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 821, + 472, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where vector $\\omega \\in \\mathbb{R}^{1\\times M}$ stores the ratio of point number", + "bbox": [ + 112, + 851, + 473, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of each part with respect to the total point number $N$ . Intuitively, parts with smaller point numbers contribute less in calculating the distance score, which reduces the influence of noise and occlusion. The source shape $S_{r}$ with the smallest distance score is identified as the best retrieval.", + "bbox": [ + 496, + 521, + 859, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Deformation", + "text_level": 1, + "bbox": [ + 500, + 598, + 624, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Deformation module aims to deform the retrieved shape $S_{r}$ to tightly match the target shape $S_{tgt}$ . We utilize the neural cage scaffolding strategy as in [26, 64]. First, the neural cages $C_{src}$ for $S_{r}$ is pre-calculated. We utilize the part centers $(K_{tgt}$ and $K_{src})$ to control the vertex offsets $C_{src2tgt}$ of the neural cage $C_{src}$ to match $S_{tgt}$ . In particular, we employ a neural network $\\Theta_I$ to predict an influence vector $I\\in \\mathbb{R}^{N_c\\times M}$ for each point concerning all cage vertices by $I = \\Theta_I(\\text{concat}(F_{tgt},F_{src}))$ , where $N_{c}$ denotes the number of vertices used in $C_{src}$ . $C_{src2tgt}$ is computed through the influence vectors $I$ and the differences between region centers (Fig. 2 (E)):", + "bbox": [ + 496, + 619, + 859, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {s r c 2 t g t} = C _ {s r c} + \\sum_ {i = 1} ^ {M} I _ {i} \\left(K _ {s r c} ^ {(i)} - K _ {t g t} ^ {(i)}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 792, + 859, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we employ a sparse cage scaffolding strategy [26, 64] to achieve the deformation field of $S_{src}$ . The deformed", + "bbox": [ + 498, + 838, + 859, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21020", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "shape $S_{src2tgt}$ of $S_{src}$ can be expressed as follows:", + "bbox": [ + 112, + 126, + 421, + 140 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {s r c 2 t g t} = S _ {s r c} + \\Psi \\left(C _ {s r c}, C _ {s r c 2 t g t}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 148, + 470, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Psi$ computes the displacement of each point in $S_{src}$ by evaluating the differences between $C_{src}$ and $C_{src2tgt}$ , thereby achieving deformation.", + "bbox": [ + 111, + 171, + 470, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. ShapeMatcher: Joint Training", + "text_level": 1, + "bbox": [ + 112, + 220, + 362, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our core insight in ShapeMatcher is that the four highly-associated tasks: Canonicalization, Segmentation, Retrieval and Deformation can be trained simultaneously and supervise each other via introducing cross-task consistency terms. We mainly introduce two types of losses here, i.e. partial-full consistency losses and task-oriented loss. For more details, please refer to the Supplementary Material.", + "bbox": [ + 111, + 241, + 470, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task-Oriented Loss. In the Canonicalization, we mainly use Chamfer Distance to constrain the canonicalized $S_{tgt}^{c}$ and $\\hat{S}_{tgt}^{c}$ predicted in the affine-invariant space by the supplementary reconstruction branch, to enforce the affine-invariety of $F_{tgt}^{*}$ in Sec. 3.1, by", + "bbox": [ + 111, + 338, + 470, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c a n}} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(S _ {t g t} ^ {c}, \\hat {S} _ {t g t} ^ {c}\\right) + \\operatorname {o r t h} \\left(R _ {t g t}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 420, + 470, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\text{orth}(R_{tgt})$ serving the purpose of enforcing the orthogonality of matrices.", + "bbox": [ + 111, + 444, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the Segmentation, to keep consistency between the part segmentation and the predicted part center, we jointly train $\\Theta_{l}$ and $\\Theta_{c}$ with the following loss, which enforce that each predicted part center $K_{tgt}^{i}$ approximately lies in the center of all points belonging to the part $M_{i}$ ,", + "bbox": [ + 111, + 472, + 470, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e g} = \\sum_ {m = 1} ^ {M} \\| K _ {t g t} ^ {(m)} - \\left(F _ {s e g} ^ {\\top} * S _ {t g t} ^ {c}\\right) ^ {(m)} \\| _ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 547, + 470, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the Retrieval and Deformation simultaneously, for an input target $S_{tgt}$ , we randomly select a source model $S_{src}$ from the database for training. Specifically, to eliminate the influence of occlusion, we do not directly use the global Chamfer Distance of $S_{tgt}$ and $S_{src}$ as ground truth. Instead, we employ a regional supervision strategy, ensuring that occluded areas do not contribute to the training of retrieval network. Taking the $i$ -th region as an example, $S_{tgt}^{i}$ represents all points in $S_{tgt}$ that belong to the $i$ -th region. We calculate the average of the nearest distances $D_{i}$ from each point in $S_{tgt}^{i}$ to the deformed shape $S_{src}^{dfm}$ to enforce the learning of the regional retrieval tokens by", + "bbox": [ + 111, + 592, + 470, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e t r i e v a l}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} M S E \\left(Q _ {t g t} ^ {(i)} - Q _ {s r c} ^ {(i)}, D _ {i}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 766, + 470, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The deformation loss is achieved by directly constraining the Chamfer Distance between $S_{tgt}$ and $S_{src}^{dfm}$ , expressed as:", + "bbox": [ + 111, + 810, + 470, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e f o r m}} = \\left. d i s _ {\\text {c h a m}} \\left(S _ {t g t}, S _ {s r c} ^ {d f m}\\right) + \\| I \\| _ {2} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 849, + 470, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we regularize $I$ using the L2 norm.", + "bbox": [ + 500, + 126, + 751, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Partial-Full Consistency Losses. In the first two modules: Canonicalization, Segmentation, the full branch serves as a guidance to enhance the learning of the partial branch. Therefore, in each module, we can enforce corresponding consistency terms between the outputs of the two parallel branches (Fig. 2 (F)).", + "bbox": [ + 496, + 138, + 857, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the consistency training process, for randomly selected full input $S_{full}$ , we generate a mask $U_{f2p} \\in \\mathbb{R}^N$ to crop it to simulate the situation of a partial input $S_{partial} = S_{full}U_{f2p}$ .", + "bbox": [ + 496, + 223, + 857, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the Canonicalization module, to enforce the consistency in the affine-invariant space between the two branches, we apply the same transformation $U_{f2p}$ to $S_{full}^{c}$ in the affine-invariant space as before and then use the chamfer distance to constrain its distance to $S_{partial}^{c}$ :", + "bbox": [ + 496, + 277, + 857, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c c a n} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(S _ {\\text {p a r t i a l}} ^ {c}, S _ {\\text {f u l l}} ^ {c} U _ {f 2 p}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 356, + 857, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, in the Segmentation module, for the consistency constraint of the part center prediction network $\\Theta_{c}$ , we directly use the Chamfer Distance to constrain the region centers detected by the two branches:", + "bbox": [ + 496, + 380, + 857, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c c e n} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(K _ {\\text {p a r t i a l}}, K _ {\\text {f u l l}}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 444, + 857, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the segmentation network $\\Theta_{l}$ , we mask the segmentation results of the full branch $F_{seg}^{(full)}$ and compare them with the results of the partial branch $F_{seg}^{(partial)}$ :", + "bbox": [ + 496, + 469, + 857, + 515 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c s e g} = d i s _ {c h a m} \\left(F _ {s e g} ^ {(f u l l)} U _ {f 2 p}, F _ {s e g} ^ {(p a r t i a l)}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 521, + 857, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint Training. Generally, the joint training of Shape-Matcher is divided into three stages. First, we train the full branch by $\\mathcal{L}_{can}$ and $\\mathcal{L}_{seg}$ for construction of Canonicalization and Segmentation ability. Second, the partial branch is introduced and trained by both the task-oriented losses for Canonicalization and Segmentation $\\mathcal{L}_{can}$ and $\\mathcal{L}_{seg}$ and the partial-full consistency loss terms $\\mathcal{L}_{ccan}$ , $\\mathcal{L}_{ccen}$ and $\\mathcal{L}_{cseg}$ . Finally, after training Canonicalization and Segmentation of the both branches, $\\mathcal{L}_{\\text{retrieval}}$ and $\\mathcal{L}_{\\text{deform}}$ are adopted for joint Retrieval and Deformation training simultaneously utilizing the both branches to handle partial target inputs and full source inputs respectively.", + "bbox": [ + 496, + 547, + 857, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 724, + 619, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 748, + 671, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we mainly focus on R&D experiments, which better reflects the overall performance of the system. The ablations and analysis also demonstrate the effectiveness of considering joint Canonicalization and Part Segmentation.", + "bbox": [ + 496, + 768, + 857, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the effectiveness of our joint framework using three datasets: two synthetic datasets,", + "bbox": [ + 496, + 838, + 857, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21021", + "bbox": [ + 480, + 890, + 514, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PartNet [39] and ComplementMe [53], and one real-world dataset, Scan2CAD [3]. For datasets PartNet and ComplementMe, we follow the same database splits as in [57], separating their target inputs into training and testing sets. In our training process, we exclusively employ mesh models and do not utilize part segmentations as in [57], since the process of ShapeMatcher is fully self-supervised and does not need any additional annotations. The shapes used in PartNet and ComplementMe datasets are sourced from ShapeNet [7]. PartNet comprises 1,419 source models in the database, with 11,433 target models in the training set and 2,861 in the testing set. In ComplementMe, the numbers are 400, 11,311 and 2,825 respectively. In the synthetic cases, three categories of tables, chairs, and cabinets are evaluated on both datasets. Scan2CAD [3] is a real-world dataset developed based on ScanNet [11] with capacity of 14,225 objects. The input point cloud data on Scan2CAD is generated by reverse-projecting the depth images. In the real-world cases, we conduct training on the categories of tables, chairs, and cabinets from PartNet and directly testing on Scan2CAD.", + "bbox": [ + 109, + 125, + 472, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Both baseline methods, Uy et al. [57] and U-RED [17], are trained using the same data partitioning strategy stated above. To ensure fairness in comparison with ShapeMatcher, we augment the training data with pose variations, keeping other hyperparameters consistent with the original paper. During testing, we evaluate scenarios where target observations with arbitrary poses are directly used as input. Additionally, we test scenarios where the inputs are transformed using an offline pose estimation method [18], simulating the two-stage route of traditional methods with pre-canonicalizing (Uy et al. [57] + PE and U-RED [17] + PE). For experiments on Scan2CAD, we directly use the baseline models trained on PartNet with the 25% occlusion setting to conduct zero-shot testing, since real-world ground-truth models are inaccessible for training.", + "bbox": [ + 109, + 417, + 472, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We utilize Chamfer Distance (CD) on the magnitude of $10^{-2}$ to assess both full shape scenarios and partial shape scenarios. We calculate the metrics following [57] to use the best result among the top 10 candidate objects. The final average metrics are obtained by averaging the results across all instances.", + "bbox": [ + 109, + 628, + 470, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. During training, we uniformly sample objects to obtain point clouds with $M = 2500$ points to represent shapes. We directly generate partial point clouds from the corresponding full point clouds by random cropping for the partial branch inputs. We apply random pose augmentation to the input shape, specifically with random translations $T_{rand} \\in [-0.1, 0.1]$ and random rotations $R_{rand} \\in [-1, 1]$ on three Eulerian angles respectively. We set the initial learning rate to $1e - 3$ and train ShapeMatcher for 200 epochs in every training stage of Sec. 3.5. Regarding the weight of the loss, in the first", + "bbox": [ + 109, + 713, + 472, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stage considering only the full branch, $\\mathcal{L}_{can}$ and $\\mathcal{L}_{seg}$ are equally weighted. In the second stage introducing the partial branch, we primarily emphasize the partial-full consistency losses, assigning significant weights to $\\mathcal{L}_{ccan}$ , $\\mathcal{L}_{ccen}$ and $\\mathcal{L}_{cseg}$ with weights set as 5, 2, and 2 respectively, while keeping the remaining weights default at 1. In the final stage for joint R&D, both $\\mathcal{L}_{retrieval}$ and $\\mathcal{L}_{deform}$ are equally weighted.", + "bbox": [ + 496, + 126, + 857, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Synthetic Cases", + "text_level": 1, + "bbox": [ + 498, + 247, + 643, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the ability of ShapeMatcher tackling the challenge of arbitrary poses and occlusions, we first use synthetic datasets to simulate this scenario. We evaluate all methods [17, 57] where object observations with arbitrary poses are directly used as input. Additionally, we also report results where inputs are transformed and canonicalized using an offline pose estimation method [18] for baseline methods (Uy et al. + PE and U-RED + PE). Moreover, we analyze inference time of ShapeMatcher against the R&D baselines in the Supplementary Material.", + "bbox": [ + 496, + 268, + 857, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct two types of inputs for evaluation: full inputs using the PartNet and ComplementMe datasets, and partial input tests using $10\\%$ , $25\\%$ , and $50\\%$ occlusion rates on the PartNet dataset. The results of the full input tests are detailed in Table 1. In PartNet, our ShapeMatcher significantly outperforms the current leading competitors. For the Chamfer Distance on three categories, ShapeMatcher measures at 0.197, 0.150, and 0.519, maintaining the leading position. Even when the processed PE results are used as input, the baselines' results still fall short of ShapeMatcher. This demonstrates the effectiveness of adopting the affine-invariant features in the joint Canonicalization step. Results from ComplementMe supports the same conclusion, where ShapeMatcher reports significantly better results compared to the baseline methods. ShapeMatcher surpasses the top-performing Uy et al. + PE by $85.2\\%$ . Such superior results yielded by ShapeMatcher demonstrates that the joint consideration of all four steps improves the matching accuracy a lot.", + "bbox": [ + 496, + 409, + 857, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For evaluation on partial inputs, we control the occlusion rates of partial point clouds by controlling the position of the cropping planes onto the full point clouds. The evaluation on partial inputs are presented in Table 2. Concretely, ShapeMatcher outperforms the current top method handling partial inputs U-RED by 5.018, 5.666 and 7.241 at the occlusion rate of $10\\%$ , $25\\%$ and $50\\%$ respectively. As the occlusion rate increases, the superiority of the ShapeMatcher method grows. Considering PE adopting, the same trend is exhibited. ShapeMatcher surpasses the U-RED + PE by 0.525, 0.949, 2.434 under three occlusion rates. It demonstrates that the proposed region-weighted retrieval brings strong robustness of our method against occlusion.", + "bbox": [ + 496, + 671, + 857, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 3 and 4, our results show more geo", + "bbox": [ + 517, + 852, + 857, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21022", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8ea4bf55a8d4fee32a5a5e3b98b7e6e4aa4784108df2c440eee28b72ccf1bee5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PartNet [39]
MethodChairTableCabinetAverage
Uy et al. [57]4.2696.3024.1185.271
Uy et al. [57] + PE1.5073.0061.0702.219
U-RED [17]5.3314.9809.1415.463
U-RED [17] + PE1.0250.3591.4230.725
Ours0.1970.1500.5190.200
ComplementMe [53]
MethodChairTableCabinetAverage
Uy et al. [57]4.0185.480-4.825
Uy et al. [57] + PE1.4392.454-1.999
U-RED [17]8.5755.800-7.044
U-RED [17] + PE4.9540.847-2.688
Ours0.2530.328-0.294
", + "bbox": [ + 117, + 123, + 467, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1d0415bf3cb7cea2cb8a5e6c19063b8ae65efd82195d0fb3e8ccfb0300ae069.jpg", + "table_caption": [ + "Table 1. The Chamfer Distance metrics for joint R&D results on full shapes under arbitrary poses." + ], + "table_footnote": [], + "table_body": "
OcclusionMethodChairTableCabinetAverage
10%Uy et al. [57]4.3726.3954.1795.365
Uy et al. + PE [57]1.5232.9821.1332.219
U-RED [17]6.0255.3755.2695.640
U-RED [17] + PE1.2071.0121.6691.147
Ours0.6760.4811.2120.622
25%Uy et al. [57]4.6546.9274.7505.795
Uy et al. + PE [57]1.8033.1951.6072.481
U-RED [17]5.1967.2158.1646.442
U-RED [17] + PE1.6841.3872.7951.625
Ours0.8780.6431.0710.776
50%Uy et al. [57]6.0709.3227.9297.841
Uy et al. + PE [57]3.3145.0304.5844.272
U-RED [17]8.6968.3877.6138.455
U-RED [17] + PE4.7222.0157.9033.628
Ours1.1971.0791.8721.194
", + "bbox": [ + 117, + 359, + 467, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. The Chamfer Distance metrics for joint R&D results on partial shapes under arbitrary poses of PartNet dataset.", + "bbox": [ + 112, + 547, + 470, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "metric resemblance to the targets compared to other methods. This is attributed to the suitable joint consideration of the four highly-associated processes, which accurately decouples the input poses, mapping them to a consistent space for accurate $R\\& D$ . The region-weighted retrieval we employ explicitly eliminates the influence of occluded areas, allowing for a more precise matching with the source model.", + "bbox": [ + 111, + 599, + 470, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Real-world Cases", + "text_level": 1, + "bbox": [ + 112, + 719, + 267, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We test the effectiveness of ShapeMatcher on real-world datasets. In such case, ShapeMatcher is trained on the synthetic PartNet with $25\\%$ occlusion and directly tested on the partial scans of the real-world dataset Scan2CAD without manual pose adjustments. Table 3 displays our results, where our method significantly outperforms existing competitors. Particularly, compared to the U-RED, in three categories, the reported Chamfer Distance are reduced by $92\\%$ , $96\\%$ , and $94\\%$ respectively. In comparison to Uy et al., the", + "bbox": [ + 111, + 741, + 470, + 866 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9d19a29390a066187b4ac1a2dca0d7d5cd538e53dc1ab1f2296fc2976ac0159.jpg", + "image_caption": [ + "Figure 3. Qualitative R&D results with full target inputs on Part-Net." + ], + "image_footnote": [], + "bbox": [ + 501, + 123, + 857, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8280cc161639c2877722806059ab8e09ffdc639c2697c9bd5b740ee01ba81e4.jpg", + "image_caption": [ + "Figure 4. Qualitative R&D results with partial target inputs on the occlusion rate of $25\\%$ on PartNet." + ], + "image_footnote": [], + "bbox": [ + 501, + 342, + 854, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reported Chamfer Distance are reduced by $91\\%$ , $98\\%$ , and $92\\%$ . Even compared with the use of PE on the two baseline methods, there remains a significant leap forward of Shape-Matcher. This zero-shot experiments on real-world further prove that the procedure of joint Canonicalization, Segmentation, Retrieval and Deformation possesses a strong domain adaptation ability, showcasing the great potential of ShapeMatcher in real-world applications. The visualizations on Scan2CAD are provided in the supplementary material.", + "bbox": [ + 496, + 599, + 857, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 747, + 648, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on PartNet, mainly on two aspects. First, in the Canonicalization, we investigate the importance of disentangling different pose intrinsics in Table 4, and demonstrate the effectiveness of joint considering Canonicalization. Second, in Table 5, we ablate the region-weighted Retrieval and the part center guided Deformation. Moreover, the robustness against occlusion is", + "bbox": [ + 496, + 768, + 857, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21023", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e7cf98b08fcd1cb144ea04c18f6751fbfc81583521edfdb7526e9014cd6f6d6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodChairTableCabinetAverage
Uy et al. [57]4.8867.6058.3356.181
Uy et al. [57] + PE3.3626.6577.2614.905
U-RED [17]5.4905.13110.0915.945
U-RED [17] + PE2.8933.1645.9573.354
Ours0.4230.1860.6540.375
", + "bbox": [ + 117, + 123, + 467, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "analyzed in Table 2.", + "text_level": 1, + "bbox": [ + 112, + 268, + 236, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Canonicalization Capability. To study the impact of different intrinsics of object poses in the Canonicalization process, we conduct ablations on decoupling of translations, rotations and scales. The results are presented in Table 4. Specifically, in row (1), we make no adjustments to the input poses. Thanks to the regional-level $R\\& D$ process, it shows decent performance. However, there is still a noticeable gap compared to row (4), indicating the significance of our proposed joint Canonicalization. In row (2), we solely decouple the input translations, resulting in a decrease of $14\\%$ in reported metrics. Moving to row (3), upon this foundation, we add decoupling for rotation, leading to a substantial decrease in reported Chamfer Distance, averaging at 0.244. In row (4), we introduce scale decoupling, resulting in another decrease in the reported metrics. It is evident that accurate decoupling of rotation is a crucial aspect of the success of the Canonicalization process. Moreover, it demonstrates that to integrate Canonicalization and $R\\& D$ process is indispensable for the ShapeMatcher process.", + "bbox": [ + 111, + 282, + 473, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Deformation and Retrieval Ability. To validate the effectiveness of our proposed region-weighted Retrieval and the part center guided neural cage Deformation, we conduct an ablation study on the PartNet dataset with the $25\\%$ occlusion rate. The results are presented in Table 5. In row (1), we conduct experiments using global retrieval and global deformation. This means we directly use an MLP network to extract overall point cloud features as the retrieval vector [57]. In the deformation network, similarly, we directly use an MLP network to generate neural cage offsets for deformation [26, 64]. Due to the lack of extraction of local information, the reported Chamfer Distance is more than twice of row (4). In row (2), we employ the global retrieval and the part center guided neural cage deformation. This improvement allows much more tightly-matched deformation by the retrieved source model, resulting in a $14\\%$ decrease in reported metrics. In row (3), we conduct experiments using the regional retrieval and the global deformation. The proposed regional-weighted retrieval handles occluded objects, reducing the impact of occluded parts and resulting in a substantial decrease in Chamfer Distance, down to 0.973.", + "bbox": [ + 111, + 547, + 473, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Occlusion Robustness. We test shapes at different occlusion levels by altering the occlusion ratio in the input.", + "bbox": [ + 112, + 838, + 472, + 867 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/659e313f0673f9c5e78354fdbe00538640be07c1d639d0f5f49f3f49a57bc908.jpg", + "table_caption": [ + "Table 3. The Chamfer Distance metrics for joint R&D results in real-world Scan2CAD [3]." + ], + "table_footnote": [], + "table_body": "
Trans.Rot.Scal.ChairTableCabinetAverage
(1)0.5710.5021.2330.590
(2)✓0.4680.4421.0960.506
(3)✓✓0.2130.2000.6740.244
(4)✓✓✓0.1970.1500.5190.200
", + "bbox": [ + 504, + 123, + 856, + 185 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a5bade35b59c7661e0a75017982ed6175c9951318615cab3e14d279d3358d531.jpg", + "table_caption": [ + "Table 4. Ablations on the Canonicalization process, we demonstrate the effectiveness of the joint Canonicalization by ablating different pose intrinsics. Here, Trans. denotes the decoupling of translation, Rot. represents the rotation, and Scal. signifies the scale." + ], + "table_footnote": [], + "table_body": "
Gl. R.Re. R.Gl. D.Re. D.ChairTableCabinetAverage
(1)✓✓1.6721.4461.8001.570
(2)✓✓1.5391.3051.6411.431
(3)✓✓1.0420.8741.2230.973
(4)✓✓0.8780.6431.0710.776
", + "bbox": [ + 503, + 268, + 856, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablations of the $R\\& D$ process. GL. R. denotes the global feature based retrieval [57], Re. R. represents the proposed region-weighted retrieval, GL. D. signifies direct neural cage deformation using global features [26, 64], and Re. D. denotes the adopted regional part center guided neural cage deformation.", + "bbox": [ + 498, + 332, + 857, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For each specific occlusion ratio, we deliberately crop a portion of the complete point cloud to simulate occlusion. We test scenarios with occlusion ratios of $10\\%$ , $25\\%$ , and $50\\%$ , and the results are presented in Table 2. Observably, as the occluded regions increased, the Chamfer Distance significantly rises for the baseline methods. Taking the U-RED + PE as an example, its reported metrics increase from 1.147 to 3.628, doubling in value, as the occluded area expands. In contrast, our method increases by less than 1-fold, which exhibits strong robustness against occlusion.", + "bbox": [ + 498, + 402, + 859, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 553, + 607, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present ShapeMatcher, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonize the object by extracting point-wise affine-invariant features. Then, the affine-invariant features are leveraged to predict semantically consistent part segmentation and corresponding part centers. Afterwards, the lightweight region-weighted retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that Shape-Matcher surpasses competitors by a large margin. In the future, we plan to further apply I-RED to various downstream tasks like robotic grasping. Limitations are discussed in the Supplementary Material.", + "bbox": [ + 496, + 575, + 859, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21024", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 125, + 200, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2", + "[2] Brandon Anderson, Truong Son Hy, and Risi Kondor. Cormorant: Covariant molecular neural networks. Advances in neural information processing systems, 32, 2019. 3", + "[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6, 8", + "[4] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 2", + "[5] Frederic Bosche and Carl T Haas. Automated retrieval of 3d cad model objects in construction range images. Automation in Construction, 17(4):499-512, 2008. 2", + "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1", + "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6", + "[8] Yunlu Chen, Basura Fernando, Hakan Bilen, Matthias Nießner, and Efstratios Gavves. 3d equivariant graph implicit functions. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 2, 3", + "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2", + "[10] Manuel Dahnert, Angela Dai, Leonidas J Guibas, and Matthias Nießner. Joint embedding of 3d scan and cad objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8749-8758, 2019. 1, 2", + "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6", + "[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868-5877, 2017. 2", + "[13] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 3" + ], + "bbox": [ + 114, + 147, + 472, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-pose: Exploiting self-occlusion for direct 6d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12396-12405, 2021. 2", + "[15] Yan Di, Henrique Morimitsu, Shan Gao, and Xiangyang Ji. Monocular piecewise depth estimation in dynamic scenes by exploiting superpixel relations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4363-4372, 2019. 1", + "[16] Yan Di, Henrique Morimitsu, Zhiqiang Lou, and Xiangyang Ji. A unified framework for piecewise semantic reconstruction in dynamic scenes via exploiting superpixel relations. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10737-10743. IEEE, 2020. 1", + "[17] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. arXiv preprint arXiv:2308.06383, 2023. 1, 2, 3, 6, 7, 8", + "[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 2, 6", + "[19] Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning so (3) equivariant representations with spherical cnns. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52–68, 2018. 3", + "[20] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 2018 International Conference on 3D Vision (3DV), pages 672–681. IEEE, 2018. 2", + "[21] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9785-9795, 2019. 1", + "[22] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2", + "[23] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, volume 27, pages 1449-1457. Wiley Online Library, 2008. 2", + "[24] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 599-628. Springer, 2020. 2", + "[25] Dominic Jack, Jhony K Pontes, Sridha Sridharan, Clinton Fookes, Sareh Shirazi, Frederic Maire, and Anders Eriksson. Learning free-form deformations for 3d object reconstruction. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6," + ], + "bbox": [ + 501, + 126, + 857, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21025", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2018, Revised Selected Papers, Part II 14, pages 317-333. Springer, 2019. 2", + "[26] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3d keypoint discovery for shape control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12783-12792, 2021. 4, 8", + "[27] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2", + "[28] Chiyu Jiang, Jingwei Huang, Andrea Tagliasacchi, and Leonidas J Guibas. Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems, 33:9745-9757, 2020. 1, 2", + "[29] Jincen Jiang, Xuequan Lu, Lizhi Zhao, Richard Dazaley, and Meili Wang. Masked autoencoders in 3d point cloud representation learning. IEEE Transactions on Multimedia, 2023. 2", + "[30] Oren Katzir, Dani Lischinski, and Daniel Cohen-Or. Shape-pose disentanglement using se (3)-equivariant vector neurons. In European Conference on Computer Vision, pages 468-484. Springer, 2022. 2, 3", + "[31] Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch-gordan nets: a fully fourier space spherical convolutional neural network. Advances in Neural Information Processing Systems, 31, 2018. 3", + "[32] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 2", + "[33] Leon Lang and Maurice Weiler. A wigner-echart theorem for group equivariant convolution kernels. arXiv preprint arXiv:2010.10952, 2020. 3", + "[34] Yangyan Li, Hao Su, Charles Ruizhongtai Qi, Noa Fish, Daniel Cohen-Or, and Leonidas J Guibas. Joint embeddings of shapes and images via cnn image purification. ACM transactions on graphics (TOG), 34(6):1-12, 2015. 2", + "[35] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1", + "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2", + "[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + ], + "bbox": [ + 114, + 126, + 470, + 864 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2", + "[39] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6, 7", + "[40] Liangliang Nan, Ke Xie, and Andrei Sharf. A search-classify approach for cluttered indoor scene understanding. ACM Transactions on Graphics (TOG), 31(6):1-10, 2012. 1", + "[41] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1", + "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2", + "[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3", + "[44] Yuchen Rao, Yinyu Nie, and Angela Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems, 35:34436-34450, 2022. 2", + "[45] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2", + "[46] Adriana Schulz, Ariel Shamir, Ilya Baran, David IW Levin, Pitchaya Sitthi-Amorn, and Wojciech Matusik. Retrieval on parametric shape collections. ACM Transactions on Graphics (TOG), 36(1):1-14, 2017. 1, 2", + "[47] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116. CiteSeer, 2007. 2", + "[48] Yongzhi Su, Yan Di, Guangyao Zhai, Fabian Manhardt, Jason Rambach, Benjamin Busam, Didier Stricker, and Federico Tombari. Opa-3d: Occlusion-aware pixel-wise aggregation for monocular 3d object detection. IEEE Robotics and Automation Letters, 8(3):1327–1334, 2023. 2", + "[49] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1", + "[50] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization" + ], + "bbox": [ + 501, + 127, + 857, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21026", + "bbox": [ + 480, + 891, + 517, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11286-11295, 2020. 3", + "[51] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 1, 2021. 3", + "[52] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua Siegel, and Sanjay Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 61–70, 2020. 2", + "[53] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 6, 7", + "[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3405-3414, 2019. 1", + "[55] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3", + "[56] Mikaela Angelina Uy, Jingwei Huang, Minhyuk Sung, Tolga Birdal, and Leonidas Guibas. Deformation-aware 3d model embedding and retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 397-413. Springer, 2020. 1, 2", + "[57] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 1, 2, 3, 6, 7, 8", + "[58] Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. 3dn: 3d deformation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1038-1046, 2019. 1, 2", + "[59] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco S Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. Advances in Neural Information Processing Systems, 31, 2018. 3", + "[60] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2", + "[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2" + ], + "bbox": [ + 114, + 126, + 470, + 866 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 2", + "[63] Shuo Yang, Min Xu, Haozhe Xie, Stuart Perry, and Jiahao Xia. Single-view 3d object reconstruction from shape priors in memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3152-3161, 2021. 2", + "[64] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 1, 2, 3, 4, 8", + "[65] Michela Zaccaria, Fabian Manhardt, Yan Di, Federico Tombari, Jacopo Aleotti, and Mikhail Giorgini. Self-supervised category-level 6d object pose estimation with optical flow consistency. IEEE Robotics and Automation Letters, 8(5):2510-2517, 2023. 2", + "[66] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. arXiv preprint arXiv:2309.12188, 2023. 1", + "[67] Guangyao Zhai, Dianye Huang, Shun-Cheng Wu, HyunJun Jung, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Monograspnet: 6-dof grasping with a single rgb image. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1708-1714. IEEE, 2023. 1", + "[68] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. arXiv preprint arXiv:2305.16283, 2023. 2", + "[69] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1", + "[70] Chenyangguang Zhang, Yan Di, Ruida Zhang, Guangyao Zhai, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ddf-ho: Hand-held object reconstruction via conditional directed distance field. arXiv preprint arXiv:2308.08231, 2023. 2", + "[71] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European Conference on Computer Vision, pages 655–672. Springer, 2022. 2", + "[72] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-posed: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452-7459. IEEE, 2022. 2" + ], + "bbox": [ + 501, + 127, + 857, + 864 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21027", + "bbox": [ + 480, + 890, + 517, + 902 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu-jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022.", + "bbox": [ + 114, + 126, + 470, + 198 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21028", + "bbox": [ + 480, + 890, + 517, + 901 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_model.json b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_model.json new file mode 100644 index 0000000000000000000000000000000000000000..918357d09ac54971d5157bf503ac75b6a82d1044 --- /dev/null +++ b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_model.json @@ -0,0 +1,2424 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.26, + 0.044, + 0.782, + 0.058 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.339, + 0.058, + 0.704, + 0.071 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.312, + 0.072, + 0.733, + 0.085 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.162, + 0.775, + 0.202 + ], + "angle": 0, + "content": "ShapeMatcher: Self-Supervised Joint Shape Canonicalization, Segmentation, Retrieval and Deformation" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.228, + 0.846, + 0.261 + ], + "angle": 0, + "content": "Yan Di\\(^{1*}\\), Chenyangguang Zhang\\(^{2*}\\), Chaowei Wang\\(^{3*}\\), Ruida Zhang\\(^{2}\\), Guangyao Zhai\\(^{1}\\), Yanyan Li\\(^{1}\\), Bowen Fu\\(^{2}\\), Xiangyang Ji\\(^{2}\\) and Shan Gao\\(^{†3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.261, + 0.833, + 0.279 + ], + "angle": 0, + "content": "1Technical University of Munich, 2Tsinghua University, 3Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.28, + 0.665, + 0.294 + ], + "angle": 0, + "content": "shangbuhuan13@gmail.com, gaoshan@nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.342, + 0.33, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.371, + 0.474, + 0.704 + ], + "angle": 0, + "content": "In this paper, we present ShapeMaker, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonicalize the object by extracting point-wise affine-invariant features, disentangling inherent structure of the object with its pose and size. These learned features are then leveraged to predict semantically consistent part segmentation and corresponding part centers. Next, our lightweight retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. The key insight of ShapeMaker is the simultaneous training of the four highly-associated processes: canonicalization, segmentation, retrieval, and deformation, leveraging cross-task consistency losses for mutual supervision. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that ShapeMaker surpasses competitors by a large margin. Code is released at https://github.com/Det1999/ShapeMaker." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.727, + 0.234, + 0.741 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.75, + 0.472, + 0.848 + ], + "angle": 0, + "content": "In recent years, there has been a notable surge in research interest focused on generating high-quality 3D models from scans of complex scenes [6, 15, 16, 21, 35, 49, 73]. This technology encourages extensive applications in both artistic creation [56, 57], robotics [66, 67] and 3D scene perception [41, 69]. Existing methods [41, 54] typically directly utilize deep neural networks to reconstruct 3D" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.34, + 0.86, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.524, + 0.861, + 0.614 + ], + "angle": 0, + "content": "Figure 1. Illustration of ShapeMatcher. Objects obtained from real-world scans are typically noisy, partial and exhibit various poses, making it challenging to conduct an effective \\( R \\& D \\) process (Red 'X' on the left). To address this issue, we propose Shape-Matcher that first canonicalizes the objects and then segments them into semantic parts, facilitating R&D processes (Green '√' on the right)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.86, + 0.781 + ], + "angle": 0, + "content": "models from imperfect scans. However, the presence of noise and occlusions poses a significant challenge in accurately capturing fine-grained geometric structures. To overcome this, Retrieval and Deformation (R&D) techniques [10,28,40,46,56-58,64] have been developed. These methods generally involve two key steps: first, identifying the most geometrically similar source shape from a precurated 3D database; and second, deforming the retrieved shape to achieve precise alignment with the target input. The R&D approach is particularly effective in producing 3D models enriched with fine details from source shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.861, + 0.868 + ], + "angle": 0, + "content": "However, existing R&D methods usually encounter two primary challenges that make them susceptible to noise, occlusion and pose variations, and difficult to be practically utilized. 1) Most R&D techniques [10, 17, 28, 40, 46, 56-58, 64] operate under the assumption that target shapes are aligned in a pre-processed canonical space. Typ" + }, + { + "type": "page_footnote", + "bbox": [ + 0.134, + 0.854, + 0.424, + 0.867 + ], + "angle": 0, + "content": "*Authors with equal contributions. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "21017" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.127, + 0.473, + 0.333 + ], + "angle": 0, + "content": "ically, these methods are trained and tested on datasets where shapes have been manually adjusted to this canonical state. However, when these methods are deployed in real-world settings, they necessitate either manual alignment of scanned objects or the use of additional pose estimation networks [14, 18, 48, 71, 72]. Such procedures are not only time-intensive and laborious but also prone to yielding inconsistent results. This limitation significantly impedes the direct application of these methods in real-world scenarios. 2) Previous methods [56, 57] do not design specially for dealing with partially-observed shapes, making it difficult to handle occluded objects. Although U-RED [17] considers the partial target shapes as input in the R&D process, it directly encodes the shape as a global embedding, which is not robust when dealing with significant occlusion." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.334, + 0.473, + 0.803 + ], + "angle": 0, + "content": "To address the aforementioned two challenges, in this paper, we present ShapeMatcher, a novel framework that extends traditional R&D pipeline to joint self-supervised learning of object canonicalization, segmentation, retrieval and deformation. Our core contribution lies in that the four highly-associated processes can be trained simultaneously and supervise each other via constructing several cross-task consistency losses (Fig. 1). Specifically, given a partially-observed object scan in an arbitrary pose, ShapeMatcher processes the objects in four steps. First, we follow [30], which is based on Vector Neurons [13], to extract SE(3)-invariant point-wise features by progressively separating translation and rotation. We further follow [8] to normalize the features to disassociate the object scale. Until here, we successfully obtain affine-invariant point-wise features by disentangling object's inherent structure with its pose and size. This facilitates the Canonicalization of the observed object based on these intrinsic characteristics (Fig. 2 (B)). Then we predict semantically consistent part segmentation and corresponding part centers by feeding the learned features into our Segmentation module (Fig. 2 (C)). Based on the part segmentation, in the Retrieval module (Fig. 2 (D)), we aggregate features within each part and collect them together as a comprehensive retrieval token of the object. For partial objects, we introduce a region-weighted strategy, which assigns a weight to each part according to the point inside it. Parts with more points are assigned higher weights during retrieval, which is proved to be robust to occlusions. We compare the tokens of the target object with each shape in the pre-constructed database to identify the most geometrically similar (most similar tokens) source shape. In the final Deformation module (Fig. 2 (E)), the retrieved source shape is deformed to tightly match the target object via part center guided neural cage deformation [64]." + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.804, + 0.39, + 0.816 + ], + "angle": 0, + "content": "To summarize, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.825, + 0.473, + 0.868 + ], + "angle": 0, + "content": "- We introduce ShapeMatcher, a novel self-supervised framework for joint shape canonicalization, segmentation, retrieval and deformation, handling partial target" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.127, + 0.858, + 0.182 + ], + "angle": 0, + "content": "inputs under arbitrary poses. Extensive experiments on the synthetic and real-world datasets demonstrate that ShapeMatcher surpasses existing state-of-the-art approaches by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.189, + 0.858, + 0.245 + ], + "angle": 0, + "content": "- We demonstrate that the four highly-associated tasks: canonicalization, segmentation, retrieval and deformation, can be effectively trained simultaneously and supervise each other via constructing consistencies." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.253, + 0.858, + 0.281 + ], + "angle": 0, + "content": "- We develop the region-weighted retrieval method to mitigate the impact of occlusions in the R&D process." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.189, + 0.858, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.292, + 0.636, + 0.306 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.859, + 0.493 + ], + "angle": 0, + "content": "Neural Shape Representation. The compact representation of 3D shapes in latent space, based on deep learning, has been a focal point for many researchers. Some attempts, such as [9, 27, 36, 37, 42, 45, 61, 70], employ neural networks to construct an implicit function, while others [1, 38, 52, 62, 68] directly model the shape of objects explicitly using generative models. Another common architecture in 3D shape representation learning, as seen in [12, 29, 44, 60, 63], is to use an encoder-decoder approach to generate latent representation vectors for various shapes. Although these methods have demonstrated impressive representation performance, they often struggle to generate fine-grained shapes when dealing with occlusion and noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.859, + 0.866 + ], + "angle": 0, + "content": "CAD Model Retrieval and Deformation. Retrieval and Deformation (R&D) methods lead another way to recover fine-grained geometric structures. Previous works directly retrieve the most similar CAD models by comparing the similarity of expression vectors in either descriptor space [5,46] or the latent space of neural networks [4,10,22,34]. Considering the subsequent deformation error, recent efforts introduce deformation-aware embeddings [56] or proposed new optimization objectives [24] to better capture the fine structure of deformed target objects. Nevertheless, these methods yield deteriorated performance when facing partial and pose-agnostic target shapes in real world. [17] achieves an one-to-many retrieval module for addressing the issue caused by partial observations, however, it receives canonicalized target shapes as input, which limit its applicability facing pose-agnostic target shapes in real world. As the retrieved models often exhibit some deviation from the target shape, the deformation module is used to minimize this discrepancy. Traditional approaches [20, 23, 47] aim to fit the target shape by directly optimizing the deformed shape. Neural network based techniques attempt to learn a set of deformation priors from a database of models. They represent deformations as volume warping [25,32], cage deformations [64], vertex offsets [58], or flows [28,65]. These methods typically constrain two shapes are aligned in the same coordinate system, making them challenging to apply in real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "21018" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.126, + 0.474, + 0.333 + ], + "angle": 0, + "content": "SO(3)-Equivariant Methods. An increasing body of work [2,19,31,55,59] has initiated research on SO(3) equivariance. These efforts are mostly based on steerable convolutional kernels [33]. On the other hand, another set of works achieves equivariance through pose estimation. [43] estimates the object's pose to factor out SO(3) transformations, achieving approximate equivariance. While [51] learns pose estimation in a fully unsupervised manner, the equivariant backbone they employ [50] achieves equivariance primarily through data augmentation, leading to limited generalization. In this paper, we employ Vector Neural Multi-Layer Perceptron [13] as the backbone to get neural invariant features for object canonicalization. It achieves SO(3) equivariance by lifting traditional scalar neurons to vector neurons." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.345, + 0.196, + 0.359 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.368, + 0.474, + 0.797 + ], + "angle": 0, + "content": "Overview. ShapeMatcher consists of 4 modules, corresponding to the 4 highly-associated tasks. Each of the first 3 modules: Canonicalization, Segmentation and Retrieval modules have two parallel branches, one for complete point cloud (in orange background) and the other for partial object input (in blue background). As shown in Fig. 2, given a partial target shape \\( S_{tgt} \\in \\mathbb{R}^{N \\times 3} \\) in an arbitrary pose, in the Canonicalization module, we progressively decouples its inherent shape with rotation \\( R_{tgt} \\in SO(3) \\), translation \\( T_{tgt} \\in \\mathbb{R}^3 \\) and the 3D metric size \\( s_{tgt} \\in \\mathbb{R}^3 \\) via VN-MLP [8, 13, 30], yielding the affine-invariant point-wise features \\( F_{tgt} \\). The object can then be canonicalized via inverse transformation based on intrinsics \\( \\{R_{tgt}, T_{tgt}, s_{tgt}\\} \\). In the Segmentation module, \\( F_{tgt} \\) is fed into a 4-layer MLP network to predict \\( M \\) parts and corresponding part centers \\( \\{K_{tgt}^1, K_{tgt}^2, ..., K_{tgt}^M\\} \\). The segmentation is semantically consistent across each category and thus can be matched and compared for R&D. In the Retrieval module, inside each region \\( M^i \\), we aggregate the features of all points inside it as its retrieval token \\( Q^i \\). The retrieval token for the object is then represented as \\( \\mathbf{Q}_{\\mathrm{tg}} = \\{Q_{tgt}^1, Q_{tgt}^2, ..., Q_{tgt}^M\\} \\). Similarly, during training, we obtain the intrinsics \\( \\{R_{src}, T_{src}, s_{src}\\} \\), part centers \\( \\{K_{src}^1, K_{src}^2, ..., K_{src}^M\\} \\) and retrieval tokens \\( \\mathbf{Q}_{\\mathrm{src}} = \\{Q_{src}^1, Q_{src}^2, ..., Q_{src}^M\\} \\) via the branch for complete point cloud. By comparing \\( \\mathbf{Q}_{\\mathrm{tg}} \\) and \\( \\mathbf{Q}_{\\mathrm{src}} \\) of each source shape inside the database, we identify the most geometrically similar source shape \\( S_r \\). In the final Deformation module, \\( K_{tgt} \\) and \\( K_{src} \\) are leveraged to guide the neural cage deformation [64] to deform the retrieved \\( S_r \\) towards \\( S_{tgt} \\), yiling \\( S_{src}^{df_m} \\)." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.804, + 0.267, + 0.818 + ], + "angle": 0, + "content": "3.1. Canonicalization" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.825, + 0.472, + 0.868 + ], + "angle": 0, + "content": "As shown in Fig. 2 (B), the Canonicalization module takes the target point cloud \\( S_{tgt} \\) as input and disentangle the inherent structure of \\( S_{tgt} \\) with the intrinsics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.126, + 0.858, + 0.21 + ], + "angle": 0, + "content": "\\(\\{R_{tgt}, T_{tgt}, s_{tgt}\\}\\), yielding a point-wise affine-invariant feature \\(F_{tgt}\\). Specifically, we follow VN-MLP [8, 13, 30] to first decouple translation via VNT [30] and then extract rotation via VNN [13, 30]. We further follow [8] to normalize the SE(3)-invariant features obtained above the remove the influence of scaling, yielding \\(F_{tgt}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.219, + 0.858, + 0.235 + ], + "angle": 0, + "content": "\\[\nR _ {t g t}, T _ {t g t}, F _ {t g t} ^ {*} = \\mathbf {V N - M L P} \\left(S _ {t g t}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.256, + 0.857, + 0.273 + ], + "angle": 0, + "content": "\\[\ns _ {t g t}, F _ {t g t} = \\text {n o r m a l i z e} \\left(F _ {t g t} ^ {*}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.859, + 0.318 + ], + "angle": 0, + "content": "where \\(F_{tgt}^{*}\\) denotes the SE(3)-invariant features and \\(F_{tgt}\\) denotes the affine-invariant features. Thereby, the object can be canonicalized with intrinsics as," + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.328, + 0.858, + 0.345 + ], + "angle": 0, + "content": "\\[\nS _ {t g t} ^ {c} = s _ {t g t} R _ {t g t} S _ {t g t} + T _ {t g t} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.859, + 0.463 + ], + "angle": 0, + "content": "where \\( S_{tgt}^{c} \\) denotes the normalized and canonicalized shape of \\( S^{tgt} \\). During training, in order to ensure that \\( F_{tgt} \\) fully encapsulate the geometric information of \\( S_{tgt} \\), we integrate a supplementary reconstruction branch which takes \\( F_{tgt}^{*} \\) as input and reconstruct \\( S_{tgt}^{c} \\) in the affine-invariant space [30]. Please refer to the Supplementary Material for details. For source shape \\( S_{src} \\) from the database, we follow the same procedures to extract \\( F_{src} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.63, + 0.485 + ], + "angle": 0, + "content": "3.2. Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.859, + 0.646 + ], + "angle": 0, + "content": "Given the affine-invariant features \\(F_{tgt}\\), we segment the input point cloud \\(S_{tgt}\\) into \\(M\\) semantically consistent parts. We use a 4-layer MLP \\(\\Theta_l\\) (Fig. 2 (C)) to predict a one-hot segmentation label for each point and use another 4-layer MLP \\(\\Theta_c\\) to predict \\(M\\) part centers \\(\\{K_{src}^1, K_{src}^2, \\dots, K_{src}^M\\}\\). Noteworthy, we don't need any ground truth annotations in this segmentation process. Our experiments show that the network can automatically learn semantically consistent segmentation solely through consistency supervision from the other three tasks. For source shape \\(S_{src}\\), we follow a similar process to obtain \\(\\{K_{src}^1, K_{src}^2, \\dots, K_{src}^M\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.599, + 0.666 + ], + "angle": 0, + "content": "3.3. Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.859, + 0.838 + ], + "angle": 0, + "content": "The retrieval network aims to identify the model \\( S_{src} \\) from an existing database that bears the closest resemblance to the target object \\( S_{tgt} \\) after deformation. Traditional methods [17, 57] directly extract the global features of objects for retrieval, which typically struggles with heavy occlusion since the global features are susceptible to noise and occlusion, and prone to producing erroneous retrieval results. In contrast, we employ a novel region-weighted retrieval method to explicitly encode independent and semantically consistent regions of the shape. This allows us to accurately handle partial shapes by identifying the visible regions to retrieve models most similar to the target." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.858, + 0.868 + ], + "angle": 0, + "content": "Specifically, the part segmentation network \\(\\Theta_l\\) takes \\(F_{tgt}\\) as input to predict \\(M\\) regions \\(F_{seg} = [C_1, C_2, C_3, \\dots, C_N]^\\top\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "21019" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.125, + 0.437, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.125, + 0.679, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.125, + 0.838, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.425, + 0.862, + 0.514 + ], + "angle": 0, + "content": "Figure 2. The pipeline of ShapeMatcher. Given a target point cloud obtained from a single-view scan and a pre-established database (A), ShapeMatcher generates the fine-grained reconstruction result using the joint 4 modules including Canonicalization (B), Segmentation (C), Retrieval (D) and Deformation (E), where the first three contains the partial branch for target processing and the full branch for source processing. Specifically, the target and source inputs are first canonicalized into the same affine-invariant space (B). Then, the semantic-consistent region segmentation is yielded from the affine-invariant features (C). The segmented regions are fed to the region-weight retrieval module (C) and the part center guided neural cage deformation module (E) for occlusion-robust R&D process. During training, the partial-full consistency losses (F) are enforced for the two branches." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.521, + 0.475, + 0.579 + ], + "angle": 0, + "content": "of \\(S_{tgt}^{c}\\), where \\(F_{seg} \\in \\mathbb{R}^{N \\times M}\\), \\(C_i \\in \\mathbb{R}^M\\) represents the probability of point \\(i\\) belonging to each part center. Then we use another 4-layer feature aggregator \\(\\Theta_f\\) (Fig. 2 (D)) to extract the retrieval tokens \\(Q\\) of all parts as follows," + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.585, + 0.473, + 0.603 + ], + "angle": 0, + "content": "\\[\nF _ {c l s} = F _ {s e g} ^ {\\top} * \\Theta_ {f} (F _ {t g t}), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.62, + 0.472, + 0.657 + ], + "angle": 0, + "content": "\\[\nQ _ {t g t} = F _ {c l s} / \\left(\\sum_ {n = 1} ^ {N} F _ {s e g} ^ {(n)}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.474, + 0.816 + ], + "angle": 0, + "content": "where \\( Q_{tgt} \\in \\mathbb{R}^{M \\times C} \\) contains the \\( C \\)-dimensional retrieval tokens for all the \\( M \\) regions. Here we employ a soft assignment strategy where each point inside \\( S_{tgt} \\) is estimated \\( M \\) values describing the probabilities belonging to each of the \\( M \\) parts. Therefore, we first aggregate features \\( F_{cls} \\) on all points belonging to each part and then normalize \\( F_{cls} \\) using the sum of probabilities of points in each part, as in Eq. 4 and Eq. 5. Following a similar strategy, we can obtain \\( Q_{src} \\) for each source model in the pre-curated database. We just need to compare \\( Q_{tgt} \\) with the retrieval tokens \\( Q_{src} \\) of all source shapes using the weighted \\( \\mathcal{L}_1 \\) distance," + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.823, + 0.473, + 0.844 + ], + "angle": 0, + "content": "\\[\nD i s = \\omega \\sum \\mathcal {L} _ {1} \\left(Q _ {t g t} - Q _ {s r c}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.852, + 0.474, + 0.868 + ], + "angle": 0, + "content": "where vector \\(\\omega \\in \\mathbb{R}^{1\\times M}\\) stores the ratio of point number" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.86, + 0.591 + ], + "angle": 0, + "content": "of each part with respect to the total point number \\(N\\). Intuitively, parts with smaller point numbers contribute less in calculating the distance score, which reduces the influence of noise and occlusion. The source shape \\(S_{r}\\) with the smallest distance score is identified as the best retrieval." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.625, + 0.613 + ], + "angle": 0, + "content": "3.4. Deformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.86, + 0.787 + ], + "angle": 0, + "content": "The Deformation module aims to deform the retrieved shape \\(S_{r}\\) to tightly match the target shape \\(S_{tgt}\\). We utilize the neural cage scaffolding strategy as in [26, 64]. First, the neural cages \\(C_{src}\\) for \\(S_{r}\\) is pre-calculated. We utilize the part centers \\((K_{tgt}\\) and \\(K_{src})\\) to control the vertex offsets \\(C_{src2tgt}\\) of the neural cage \\(C_{src}\\) to match \\(S_{tgt}\\). In particular, we employ a neural network \\(\\Theta_I\\) to predict an influence vector \\(I\\in \\mathbb{R}^{N_c\\times M}\\) for each point concerning all cage vertices by \\(I = \\Theta_I(\\text{concat}(F_{tgt},F_{src}))\\), where \\(N_{c}\\) denotes the number of vertices used in \\(C_{src}\\). \\(C_{src2tgt}\\) is computed through the influence vectors \\(I\\) and the differences between region centers (Fig. 2 (E)):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.794, + 0.86, + 0.832 + ], + "angle": 0, + "content": "\\[\nC _ {s r c 2 t g t} = C _ {s r c} + \\sum_ {i = 1} ^ {M} I _ {i} \\left(K _ {s r c} ^ {(i)} - K _ {t g t} ^ {(i)}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.86, + 0.868 + ], + "angle": 0, + "content": "Finally, we employ a sparse cage scaffolding strategy [26, 64] to achieve the deformation field of \\( S_{src} \\). The deformed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "21020" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.127, + 0.423, + 0.141 + ], + "angle": 0, + "content": "shape \\(S_{src2tgt}\\) of \\(S_{src}\\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.149, + 0.472, + 0.165 + ], + "angle": 0, + "content": "\\[\nS _ {s r c 2 t g t} = S _ {s r c} + \\Psi \\left(C _ {s r c}, C _ {s r c 2 t g t}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.172, + 0.472, + 0.214 + ], + "angle": 0, + "content": "where \\(\\Psi\\) computes the displacement of each point in \\(S_{src}\\) by evaluating the differences between \\(C_{src}\\) and \\(C_{src2tgt}\\), thereby achieving deformation." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.221, + 0.363, + 0.236 + ], + "angle": 0, + "content": "3.5. ShapeMatcher: Joint Training" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.242, + 0.472, + 0.338 + ], + "angle": 0, + "content": "Our core insight in ShapeMatcher is that the four highly-associated tasks: Canonicalization, Segmentation, Retrieval and Deformation can be trained simultaneously and supervise each other via introducing cross-task consistency terms. We mainly introduce two types of losses here, i.e. partial-full consistency losses and task-oriented loss. For more details, please refer to the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.339, + 0.472, + 0.412 + ], + "angle": 0, + "content": "Task-Oriented Loss. In the Canonicalization, we mainly use Chamfer Distance to constrain the canonicalized \\( S_{tgt}^{c} \\) and \\( \\hat{S}_{tgt}^{c} \\) predicted in the affine-invariant space by the supplementary reconstruction branch, to enforce the affine-invariety of \\( F_{tgt}^{*} \\) in Sec. 3.1, by" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.421, + 0.472, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c a n}} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(S _ {t g t} ^ {c}, \\hat {S} _ {t g t} ^ {c}\\right) + \\operatorname {o r t h} \\left(R _ {t g t}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.445, + 0.472, + 0.472 + ], + "angle": 0, + "content": "with \\( \\text{orth}(R_{tgt}) \\) serving the purpose of enforcing the orthogonality of matrices." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.473, + 0.472, + 0.542 + ], + "angle": 0, + "content": "In the Segmentation, to keep consistency between the part segmentation and the predicted part center, we jointly train \\(\\Theta_{l}\\) and \\(\\Theta_{c}\\) with the following loss, which enforce that each predicted part center \\(K_{tgt}^{i}\\) approximately lies in the center of all points belonging to the part \\(M_{i}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.549, + 0.472, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e g} = \\sum_ {m = 1} ^ {M} \\| K _ {t g t} ^ {(m)} - \\left(F _ {s e g} ^ {\\top} * S _ {t g t} ^ {c}\\right) ^ {(m)} \\| _ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.593, + 0.472, + 0.759 + ], + "angle": 0, + "content": "To train the Retrieval and Deformation simultaneously, for an input target \\( S_{tgt} \\), we randomly select a source model \\( S_{src} \\) from the database for training. Specifically, to eliminate the influence of occlusion, we do not directly use the global Chamfer Distance of \\( S_{tgt} \\) and \\( S_{src} \\) as ground truth. Instead, we employ a regional supervision strategy, ensuring that occluded areas do not contribute to the training of retrieval network. Taking the \\( i \\)-th region as an example, \\( S_{tgt}^{i} \\) represents all points in \\( S_{tgt} \\) that belong to the \\( i \\)-th region. We calculate the average of the nearest distances \\( D_{i} \\) from each point in \\( S_{tgt}^{i} \\) to the deformed shape \\( S_{src}^{dfm} \\) to enforce the learning of the regional retrieval tokens by" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.767, + 0.472, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e t r i e v a l}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} M S E \\left(Q _ {t g t} ^ {(i)} - Q _ {s r c} ^ {(i)}, D _ {i}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.811, + 0.472, + 0.851 + ], + "angle": 0, + "content": "The deformation loss is achieved by directly constraining the Chamfer Distance between \\( S_{tgt} \\) and \\( S_{src}^{dfm} \\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.851, + 0.472, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e f o r m}} = \\left. d i s _ {\\text {c h a m}} \\left(S _ {t g t}, S _ {s r c} ^ {d f m}\\right) + \\| I \\| _ {2} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.127, + 0.753, + 0.139 + ], + "angle": 0, + "content": "where we regularize \\(I\\) using the L2 norm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.858, + 0.223 + ], + "angle": 0, + "content": "Partial-Full Consistency Losses. In the first two modules: Canonicalization, Segmentation, the full branch serves as a guidance to enhance the learning of the partial branch. Therefore, in each module, we can enforce corresponding consistency terms between the outputs of the two parallel branches (Fig. 2 (F))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.224, + 0.858, + 0.279 + ], + "angle": 0, + "content": "During the consistency training process, for randomly selected full input \\( S_{full} \\), we generate a mask \\( U_{f2p} \\in \\mathbb{R}^N \\) to crop it to simulate the situation of a partial input \\( S_{partial} = S_{full}U_{f2p} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.858, + 0.35 + ], + "angle": 0, + "content": "In the Canonicalization module, to enforce the consistency in the affine-invariant space between the two branches, we apply the same transformation \\( U_{f2p} \\) to \\( S_{full}^{c} \\) in the affine-invariant space as before and then use the chamfer distance to constrain its distance to \\( S_{partial}^{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.357, + 0.858, + 0.374 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c c a n} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(S _ {\\text {p a r t i a l}} ^ {c}, S _ {\\text {f u l l}} ^ {c} U _ {f 2 p}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.858, + 0.436 + ], + "angle": 0, + "content": "Similarly, in the Segmentation module, for the consistency constraint of the part center prediction network \\(\\Theta_{c}\\), we directly use the Chamfer Distance to constrain the region centers detected by the two branches:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.445, + 0.858, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c c e n} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(K _ {\\text {p a r t i a l}}, K _ {\\text {f u l l}}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.858, + 0.516 + ], + "angle": 0, + "content": "In the segmentation network \\(\\Theta_{l}\\), we mask the segmentation results of the full branch \\(F_{seg}^{(full)}\\) and compare them with the results of the partial branch \\(F_{seg}^{(partial)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.522, + 0.858, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c s e g} = d i s _ {c h a m} \\left(F _ {s e g} ^ {(f u l l)} U _ {f 2 p}, F _ {s e g} ^ {(p a r t i a l)}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.859, + 0.716 + ], + "angle": 0, + "content": "Joint Training. Generally, the joint training of Shape-Matcher is divided into three stages. First, we train the full branch by \\(\\mathcal{L}_{can}\\) and \\(\\mathcal{L}_{seg}\\) for construction of Canonicalization and Segmentation ability. Second, the partial branch is introduced and trained by both the task-oriented losses for Canonicalization and Segmentation \\(\\mathcal{L}_{can}\\) and \\(\\mathcal{L}_{seg}\\) and the partial-full consistency loss terms \\(\\mathcal{L}_{ccan}\\), \\(\\mathcal{L}_{ccen}\\) and \\(\\mathcal{L}_{cseg}\\). Finally, after training Canonicalization and Segmentation of the both branches, \\(\\mathcal{L}_{\\text{retrieval}}\\) and \\(\\mathcal{L}_{\\text{deform}}\\) are adopted for joint Retrieval and Deformation training simultaneously utilizing the both branches to handle partial target inputs and full source inputs respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.62, + 0.741 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.749, + 0.673, + 0.764 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.858, + 0.837 + ], + "angle": 0, + "content": "In this section, we mainly focus on R&D experiments, which better reflects the overall performance of the system. The ablations and analysis also demonstrate the effectiveness of considering joint Canonicalization and Part Segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.858, + 0.867 + ], + "angle": 0, + "content": "Datasets. We evaluate the effectiveness of our joint framework using three datasets: two synthetic datasets," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.516, + 0.903 + ], + "angle": 0, + "content": "21021" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.126, + 0.473, + 0.416 + ], + "angle": 0, + "content": "PartNet [39] and ComplementMe [53], and one real-world dataset, Scan2CAD [3]. For datasets PartNet and ComplementMe, we follow the same database splits as in [57], separating their target inputs into training and testing sets. In our training process, we exclusively employ mesh models and do not utilize part segmentations as in [57], since the process of ShapeMatcher is fully self-supervised and does not need any additional annotations. The shapes used in PartNet and ComplementMe datasets are sourced from ShapeNet [7]. PartNet comprises 1,419 source models in the database, with 11,433 target models in the training set and 2,861 in the testing set. In ComplementMe, the numbers are 400, 11,311 and 2,825 respectively. In the synthetic cases, three categories of tables, chairs, and cabinets are evaluated on both datasets. Scan2CAD [3] is a real-world dataset developed based on ScanNet [11] with capacity of 14,225 objects. The input point cloud data on Scan2CAD is generated by reverse-projecting the depth images. In the real-world cases, we conduct training on the categories of tables, chairs, and cabinets from PartNet and directly testing on Scan2CAD." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.419, + 0.473, + 0.626 + ], + "angle": 0, + "content": "Baselines. Both baseline methods, Uy et al. [57] and U-RED [17], are trained using the same data partitioning strategy stated above. To ensure fairness in comparison with ShapeMatcher, we augment the training data with pose variations, keeping other hyperparameters consistent with the original paper. During testing, we evaluate scenarios where target observations with arbitrary poses are directly used as input. Additionally, we test scenarios where the inputs are transformed using an offline pose estimation method [18], simulating the two-stage route of traditional methods with pre-canonicalizing (Uy et al. [57] + PE and U-RED [17] + PE). For experiments on Scan2CAD, we directly use the baseline models trained on PartNet with the 25% occlusion setting to conduct zero-shot testing, since real-world ground-truth models are inaccessible for training." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.629, + 0.472, + 0.712 + ], + "angle": 0, + "content": "Evaluation Metrics. We utilize Chamfer Distance (CD) on the magnitude of \\(10^{-2}\\) to assess both full shape scenarios and partial shape scenarios. We calculate the metrics following [57] to use the best result among the top 10 candidate objects. The final average metrics are obtained by averaging the results across all instances." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.714, + 0.473, + 0.867 + ], + "angle": 0, + "content": "Implementation Details. During training, we uniformly sample objects to obtain point clouds with \\( M = 2500 \\) points to represent shapes. We directly generate partial point clouds from the corresponding full point clouds by random cropping for the partial branch inputs. We apply random pose augmentation to the input shape, specifically with random translations \\( T_{rand} \\in [-0.1, 0.1] \\) and random rotations \\( R_{rand} \\in [-1, 1] \\) on three Eulerian angles respectively. We set the initial learning rate to \\( 1e - 3 \\) and train ShapeMatcher for 200 epochs in every training stage of Sec. 3.5. Regarding the weight of the loss, in the first" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.858, + 0.238 + ], + "angle": 0, + "content": "stage considering only the full branch, \\(\\mathcal{L}_{can}\\) and \\(\\mathcal{L}_{seg}\\) are equally weighted. In the second stage introducing the partial branch, we primarily emphasize the partial-full consistency losses, assigning significant weights to \\(\\mathcal{L}_{ccan}\\), \\(\\mathcal{L}_{ccen}\\) and \\(\\mathcal{L}_{cseg}\\) with weights set as 5, 2, and 2 respectively, while keeping the remaining weights default at 1. In the final stage for joint R&D, both \\(\\mathcal{L}_{retrieval}\\) and \\(\\mathcal{L}_{deform}\\) are equally weighted." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.248, + 0.644, + 0.263 + ], + "angle": 0, + "content": "4.2. Synthetic Cases" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.859, + 0.409 + ], + "angle": 0, + "content": "To validate the ability of ShapeMatcher tackling the challenge of arbitrary poses and occlusions, we first use synthetic datasets to simulate this scenario. We evaluate all methods [17, 57] where object observations with arbitrary poses are directly used as input. Additionally, we also report results where inputs are transformed and canonicalized using an offline pose estimation method [18] for baseline methods (Uy et al. + PE and U-RED + PE). Moreover, we analyze inference time of ShapeMatcher against the R&D baselines in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.859, + 0.672 + ], + "angle": 0, + "content": "We conduct two types of inputs for evaluation: full inputs using the PartNet and ComplementMe datasets, and partial input tests using \\(10\\%\\), \\(25\\%\\), and \\(50\\%\\) occlusion rates on the PartNet dataset. The results of the full input tests are detailed in Table 1. In PartNet, our ShapeMatcher significantly outperforms the current leading competitors. For the Chamfer Distance on three categories, ShapeMatcher measures at 0.197, 0.150, and 0.519, maintaining the leading position. Even when the processed PE results are used as input, the baselines' results still fall short of ShapeMatcher. This demonstrates the effectiveness of adopting the affine-invariant features in the joint Canonicalization step. Results from ComplementMe supports the same conclusion, where ShapeMatcher reports significantly better results compared to the baseline methods. ShapeMatcher surpasses the top-performing Uy et al. + PE by \\(85.2\\%\\). Such superior results yielded by ShapeMatcher demonstrates that the joint consideration of all four steps improves the matching accuracy a lot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.859, + 0.852 + ], + "angle": 0, + "content": "For evaluation on partial inputs, we control the occlusion rates of partial point clouds by controlling the position of the cropping planes onto the full point clouds. The evaluation on partial inputs are presented in Table 2. Concretely, ShapeMatcher outperforms the current top method handling partial inputs U-RED by 5.018, 5.666 and 7.241 at the occlusion rate of \\(10\\%\\), \\(25\\%\\) and \\(50\\%\\) respectively. As the occlusion rate increases, the superiority of the ShapeMatcher method grows. Considering PE adopting, the same trend is exhibited. ShapeMatcher surpasses the U-RED + PE by 0.525, 0.949, 2.434 under three occlusion rates. It demonstrates that the proposed region-weighted retrieval brings strong robustness of our method against occlusion." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.853, + 0.858, + 0.867 + ], + "angle": 0, + "content": "As shown in Fig. 3 and 4, our results show more geo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "21022" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.124, + 0.468, + 0.312 + ], + "angle": 0, + "content": "
PartNet [39]
MethodChairTableCabinetAverage
Uy et al. [57]4.2696.3024.1185.271
Uy et al. [57] + PE1.5073.0061.0702.219
U-RED [17]5.3314.9809.1415.463
U-RED [17] + PE1.0250.3591.4230.725
Ours0.1970.1500.5190.200
ComplementMe [53]
MethodChairTableCabinetAverage
Uy et al. [57]4.0185.480-4.825
Uy et al. [57] + PE1.4392.454-1.999
U-RED [17]8.5755.800-7.044
U-RED [17] + PE4.9540.847-2.688
Ours0.2530.328-0.294
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.322, + 0.472, + 0.348 + ], + "angle": 0, + "content": "Table 1. The Chamfer Distance metrics for joint R&D results on full shapes under arbitrary poses." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.361, + 0.468, + 0.538 + ], + "angle": 0, + "content": "
OcclusionMethodChairTableCabinetAverage
10%Uy et al. [57]4.3726.3954.1795.365
Uy et al. + PE [57]1.5232.9821.1332.219
U-RED [17]6.0255.3755.2695.640
U-RED [17] + PE1.2071.0121.6691.147
Ours0.6760.4811.2120.622
25%Uy et al. [57]4.6546.9274.7505.795
Uy et al. + PE [57]1.8033.1951.6072.481
U-RED [17]5.1967.2158.1646.442
U-RED [17] + PE1.6841.3872.7951.625
Ours0.8780.6431.0710.776
50%Uy et al. [57]6.0709.3227.9297.841
Uy et al. + PE [57]3.3145.0304.5844.272
U-RED [17]8.6968.3877.6138.455
U-RED [17] + PE4.7222.0157.9033.628
Ours1.1971.0791.8721.194
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.548, + 0.472, + 0.575 + ], + "angle": 0, + "content": "Table 2. The Chamfer Distance metrics for joint R&D results on partial shapes under arbitrary poses of PartNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.6, + 0.472, + 0.709 + ], + "angle": 0, + "content": "metric resemblance to the targets compared to other methods. This is attributed to the suitable joint consideration of the four highly-associated processes, which accurately decouples the input poses, mapping them to a consistent space for accurate \\( R\\& D \\). The region-weighted retrieval we employ explicitly eliminates the influence of occluded areas, allowing for a more precise matching with the source model." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.72, + 0.269, + 0.734 + ], + "angle": 0, + "content": "4.3. Real-world Cases" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.742, + 0.472, + 0.867 + ], + "angle": 0, + "content": "We test the effectiveness of ShapeMatcher on real-world datasets. In such case, ShapeMatcher is trained on the synthetic PartNet with \\(25\\%\\) occlusion and directly tested on the partial scans of the real-world dataset Scan2CAD without manual pose adjustments. Table 3 displays our results, where our method significantly outperforms existing competitors. Particularly, compared to the U-RED, in three categories, the reported Chamfer Distance are reduced by \\(92\\%\\), \\(96\\%\\), and \\(94\\%\\) respectively. In comparison to Uy et al., the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.124, + 0.858, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.305, + 0.858, + 0.329 + ], + "angle": 0, + "content": "Figure 3. Qualitative R&D results with full target inputs on Part-Net." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.343, + 0.855, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.549, + 0.858, + 0.573 + ], + "angle": 0, + "content": "Figure 4. Qualitative R&D results with partial target inputs on the occlusion rate of \\(25\\%\\) on PartNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.859, + 0.737 + ], + "angle": 0, + "content": "reported Chamfer Distance are reduced by \\(91\\%\\), \\(98\\%\\), and \\(92\\%\\). Even compared with the use of PE on the two baseline methods, there remains a significant leap forward of Shape-Matcher. This zero-shot experiments on real-world further prove that the procedure of joint Canonicalization, Segmentation, Retrieval and Deformation possesses a strong domain adaptation ability, showcasing the great potential of ShapeMatcher in real-world applications. The visualizations on Scan2CAD are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.748, + 0.65, + 0.761 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.859, + 0.867 + ], + "angle": 0, + "content": "We conduct ablation experiments on PartNet, mainly on two aspects. First, in the Canonicalization, we investigate the importance of disentangling different pose intrinsics in Table 4, and demonstrate the effectiveness of joint considering Canonicalization. Second, in Table 5, we ablate the region-weighted Retrieval and the part center guided Deformation. Moreover, the robustness against occlusion is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "21023" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.124, + 0.468, + 0.206 + ], + "angle": 0, + "content": "
MethodChairTableCabinetAverage
Uy et al. [57]4.8867.6058.3356.181
Uy et al. [57] + PE3.3626.6577.2614.905
U-RED [17]5.4905.13110.0915.945
U-RED [17] + PE2.8933.1645.9573.354
Ours0.4230.1860.6540.375
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.215, + 0.472, + 0.242 + ], + "angle": 0, + "content": "Table 3. The Chamfer Distance metrics for joint R&D results in real-world Scan2CAD [3]." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.269, + 0.238, + 0.282 + ], + "angle": 0, + "content": "analyzed in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.284, + 0.474, + 0.546 + ], + "angle": 0, + "content": "Canonicalization Capability. To study the impact of different intrinsics of object poses in the Canonicalization process, we conduct ablations on decoupling of translations, rotations and scales. The results are presented in Table 4. Specifically, in row (1), we make no adjustments to the input poses. Thanks to the regional-level \\( R\\& D \\) process, it shows decent performance. However, there is still a noticeable gap compared to row (4), indicating the significance of our proposed joint Canonicalization. In row (2), we solely decouple the input translations, resulting in a decrease of \\( 14\\% \\) in reported metrics. Moving to row (3), upon this foundation, we add decoupling for rotation, leading to a substantial decrease in reported Chamfer Distance, averaging at 0.244. In row (4), we introduce scale decoupling, resulting in another decrease in the reported metrics. It is evident that accurate decoupling of rotation is a crucial aspect of the success of the Canonicalization process. Moreover, it demonstrates that to integrate Canonicalization and \\( R\\& D \\) process is indispensable for the ShapeMatcher process." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.548, + 0.474, + 0.837 + ], + "angle": 0, + "content": "Deformation and Retrieval Ability. To validate the effectiveness of our proposed region-weighted Retrieval and the part center guided neural cage Deformation, we conduct an ablation study on the PartNet dataset with the \\(25\\%\\) occlusion rate. The results are presented in Table 5. In row (1), we conduct experiments using global retrieval and global deformation. This means we directly use an MLP network to extract overall point cloud features as the retrieval vector [57]. In the deformation network, similarly, we directly use an MLP network to generate neural cage offsets for deformation [26, 64]. Due to the lack of extraction of local information, the reported Chamfer Distance is more than twice of row (4). In row (2), we employ the global retrieval and the part center guided neural cage deformation. This improvement allows much more tightly-matched deformation by the retrieved source model, resulting in a \\(14\\%\\) decrease in reported metrics. In row (3), we conduct experiments using the regional retrieval and the global deformation. The proposed regional-weighted retrieval handles occluded objects, reducing the impact of occluded parts and resulting in a substantial decrease in Chamfer Distance, down to 0.973." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.839, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Occlusion Robustness. We test shapes at different occlusion levels by altering the occlusion ratio in the input." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.124, + 0.857, + 0.186 + ], + "angle": 0, + "content": "
Trans.Rot.Scal.ChairTableCabinetAverage
(1)0.5710.5021.2330.590
(2)✓0.4680.4421.0960.506
(3)✓✓0.2130.2000.6740.244
(4)✓✓✓0.1970.1500.5190.200
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.195, + 0.859, + 0.258 + ], + "angle": 0, + "content": "Table 4. Ablations on the Canonicalization process, we demonstrate the effectiveness of the joint Canonicalization by ablating different pose intrinsics. Here, Trans. denotes the decoupling of translation, Rot. represents the rotation, and Scal. signifies the scale." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.27, + 0.857, + 0.323 + ], + "angle": 0, + "content": "
Gl. R.Re. R.Gl. D.Re. D.ChairTableCabinetAverage
(1)✓✓1.6721.4461.8001.570
(2)✓✓1.5391.3051.6411.431
(3)✓✓1.0420.8741.2230.973
(4)✓✓0.8780.6431.0710.776
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.333, + 0.859, + 0.397 + ], + "angle": 0, + "content": "Table 5. Ablations of the \\( R\\& D \\) process. GL. R. denotes the global feature based retrieval [57], Re. R. represents the proposed region-weighted retrieval, GL. D. signifies direct neural cage deformation using global features [26, 64], and Re. D. denotes the adopted regional part center guided neural cage deformation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.86, + 0.542 + ], + "angle": 0, + "content": "For each specific occlusion ratio, we deliberately crop a portion of the complete point cloud to simulate occlusion. We test scenarios with occlusion ratios of \\(10\\%\\), \\(25\\%\\), and \\(50\\%\\), and the results are presented in Table 2. Observably, as the occluded regions increased, the Chamfer Distance significantly rises for the baseline methods. Taking the U-RED + PE as an example, its reported metrics increase from 1.147 to 3.628, doubling in value, as the occluded area expands. In contrast, our method increases by less than 1-fold, which exhibits strong robustness against occlusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.554, + 0.608, + 0.568 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.86, + 0.868 + ], + "angle": 0, + "content": "In this paper, we present ShapeMatcher, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonize the object by extracting point-wise affine-invariant features. Then, the affine-invariant features are leveraged to predict semantically consistent part segmentation and corresponding part centers. Afterwards, the lightweight region-weighted retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that Shape-Matcher surpasses competitors by a large margin. In the future, we plan to further apply I-RED to various downstream tasks like robotic grasping. Limitations are discussed in the Supplementary Material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "21024" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.125, + 0.201, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.148, + 0.473, + 0.2 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.201, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Brandon Anderson, Truong Son Hy, and Risi Kondor. Cormorant: Covariant molecular neural networks. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.241, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.306, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[4] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.359, + 0.472, + 0.396 + ], + "angle": 0, + "content": "[5] Frederic Bosche and Carl T Haas. Automated retrieval of 3d cad model objects in construction range images. Automation in Construction, 17(4):499-512, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.398, + 0.473, + 0.449 + ], + "angle": 0, + "content": "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.451, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.515, + 0.472, + 0.568 + ], + "angle": 0, + "content": "[8] Yunlu Chen, Basura Fernando, Hakan Bilen, Matthias Nießner, and Efstratios Gavves. 3d equivariant graph implicit functions. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.568, + 0.472, + 0.619 + ], + "angle": 0, + "content": "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.621, + 0.472, + 0.672 + ], + "angle": 0, + "content": "[10] Manuel Dahnert, Angela Dai, Leonidas J Guibas, and Matthias Nießner. Joint embedding of 3d scan and cad objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8749-8758, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.673, + 0.472, + 0.737 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.738, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868-5877, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.803, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[13] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.148, + 0.473, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.858, + 0.19 + ], + "angle": 0, + "content": "[14] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-pose: Exploiting self-occlusion for direct 6d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12396-12405, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.858, + 0.255 + ], + "angle": 0, + "content": "[15] Yan Di, Henrique Morimitsu, Shan Gao, and Xiangyang Ji. Monocular piecewise depth estimation in dynamic scenes by exploiting superpixel relations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4363-4372, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.255, + 0.858, + 0.319 + ], + "angle": 0, + "content": "[16] Yan Di, Henrique Morimitsu, Zhiqiang Lou, and Xiangyang Ji. A unified framework for piecewise semantic reconstruction in dynamic scenes via exploiting superpixel relations. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10737-10743. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.858, + 0.383 + ], + "angle": 0, + "content": "[17] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. arXiv preprint arXiv:2308.06383, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.858, + 0.458 + ], + "angle": 0, + "content": "[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.858, + 0.522 + ], + "angle": 0, + "content": "[19] Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning so (3) equivariant representations with spherical cnns. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52–68, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.523, + 0.858, + 0.585 + ], + "angle": 0, + "content": "[20] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 2018 International Conference on 3D Vision (3DV), pages 672–681. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.858, + 0.625 + ], + "angle": 0, + "content": "[21] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9785-9795, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.626, + 0.858, + 0.675 + ], + "angle": 0, + "content": "[22] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.858, + 0.727 + ], + "angle": 0, + "content": "[23] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, volume 27, pages 1449-1457. Wiley Online Library, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.858, + 0.803 + ], + "angle": 0, + "content": "[24] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 599-628. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.858, + 0.867 + ], + "angle": 0, + "content": "[25] Dominic Jack, Jhony K Pontes, Sridha Sridharan, Clinton Fookes, Sareh Shirazi, Frederic Maire, and Anders Eriksson. Learning free-form deformations for 3d object reconstruction. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.858, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "21025" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.145, + 0.127, + 0.471, + 0.154 + ], + "angle": 0, + "content": "2018, Revised Selected Papers, Part II 14, pages 317-333. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.155, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[26] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3d keypoint discovery for shape control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12783-12792, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.234, + 0.472, + 0.285 + ], + "angle": 0, + "content": "[27] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.287, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[28] Chiyu Jiang, Jingwei Huang, Andrea Tagliasacchi, and Leonidas J Guibas. Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems, 33:9745-9757, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.34, + 0.472, + 0.39 + ], + "angle": 0, + "content": "[29] Jincen Jiang, Xuequan Lu, Lizhi Zhao, Richard Dazaley, and Meili Wang. Masked autoencoders in 3d point cloud representation learning. IEEE Transactions on Multimedia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.393, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[30] Oren Katzir, Dani Lischinski, and Daniel Cohen-Or. Shape-pose disentanglement using se (3)-equivariant vector neurons. In European Conference on Computer Vision, pages 468-484. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.446, + 0.472, + 0.497 + ], + "angle": 0, + "content": "[31] Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch-gordan nets: a fully fourier space spherical convolutional neural network. Advances in Neural Information Processing Systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.499, + 0.472, + 0.576 + ], + "angle": 0, + "content": "[32] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.578, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[33] Leon Lang and Maurice Weiler. A wigner-echart theorem for group equivariant convolution kernels. arXiv preprint arXiv:2010.10952, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.618, + 0.472, + 0.67 + ], + "angle": 0, + "content": "[34] Yangyan Li, Hao Su, Charles Ruizhongtai Qi, Noa Fish, Daniel Cohen-Or, and Leonidas J Guibas. Joint embeddings of shapes and images via cnn image purification. ACM transactions on graphics (TOG), 34(6):1-12, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.671, + 0.472, + 0.735 + ], + "angle": 0, + "content": "[35] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.737, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.803, + 0.472, + 0.866 + ], + "angle": 0, + "content": "[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.127, + 0.472, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.858, + 0.178 + ], + "angle": 0, + "content": "[38] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.858, + 0.255 + ], + "angle": 0, + "content": "[39] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.257, + 0.858, + 0.295 + ], + "angle": 0, + "content": "[40] Liangliang Nan, Ke Xie, and Andrei Sharf. A search-classify approach for cluttered indoor scene understanding. ACM Transactions on Graphics (TOG), 31(6):1-10, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.858, + 0.372 + ], + "angle": 0, + "content": "[41] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.858, + 0.438 + ], + "angle": 0, + "content": "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.858, + 0.501 + ], + "angle": 0, + "content": "[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.858, + 0.554 + ], + "angle": 0, + "content": "[44] Yuchen Rao, Yinyu Nie, and Angela Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems, 35:34436-34450, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.858, + 0.618 + ], + "angle": 0, + "content": "[45] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.858, + 0.671 + ], + "angle": 0, + "content": "[46] Adriana Schulz, Ariel Shamir, Ilya Baran, David IW Levin, Pitchaya Sitthi-Amorn, and Wojciech Matusik. Retrieval on parametric shape collections. ACM Transactions on Graphics (TOG), 36(1):1-14, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.858, + 0.711 + ], + "angle": 0, + "content": "[47] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116. CiteSeer, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.858, + 0.775 + ], + "angle": 0, + "content": "[48] Yongzhi Su, Yan Di, Guangyao Zhai, Fabian Manhardt, Jason Rambach, Benjamin Busam, Didier Stricker, and Federico Tombari. Opa-3d: Occlusion-aware pixel-wise aggregation for monocular 3d object detection. IEEE Robotics and Automation Letters, 8(3):1327–1334, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.777, + 0.858, + 0.84 + ], + "angle": 0, + "content": "[49] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.858, + 0.867 + ], + "angle": 0, + "content": "[50] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.858, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.892, + 0.518, + 0.902 + ], + "angle": 0, + "content": "21026" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.144, + 0.127, + 0.472, + 0.166 + ], + "angle": 0, + "content": "for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11286-11295, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.168, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[51] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 1, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.219, + 0.472, + 0.282 + ], + "angle": 0, + "content": "[52] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua Siegel, and Sanjay Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 61–70, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.284, + 0.472, + 0.334 + ], + "angle": 0, + "content": "[53] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.336, + 0.472, + 0.4 + ], + "angle": 0, + "content": "[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3405-3414, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.402, + 0.472, + 0.464 + ], + "angle": 0, + "content": "[55] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.466, + 0.472, + 0.541 + ], + "angle": 0, + "content": "[56] Mikaela Angelina Uy, Jingwei Huang, Minhyuk Sung, Tolga Birdal, and Leonidas Guibas. Deformation-aware 3d model embedding and retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 397-413. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.543, + 0.472, + 0.619 + ], + "angle": 0, + "content": "[57] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.621, + 0.472, + 0.672 + ], + "angle": 0, + "content": "[58] Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. 3dn: 3d deformation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1038-1046, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.673, + 0.472, + 0.735 + ], + "angle": 0, + "content": "[59] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco S Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. Advances in Neural Information Processing Systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.738, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[60] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.804, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.127, + 0.472, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.858, + 0.19 + ], + "angle": 0, + "content": "[62] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.858, + 0.255 + ], + "angle": 0, + "content": "[63] Shuo Yang, Min Xu, Haozhe Xie, Stuart Perry, and Jiahao Xia. Single-view 3d object reconstruction from shape priors in memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3152-3161, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.257, + 0.858, + 0.321 + ], + "angle": 0, + "content": "[64] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.858, + 0.385 + ], + "angle": 0, + "content": "[65] Michela Zaccaria, Fabian Manhardt, Yan Di, Federico Tombari, Jacopo Aleotti, and Mikhail Giorgini. Self-supervised category-level 6d object pose estimation with optical flow consistency. IEEE Robotics and Automation Letters, 8(5):2510-2517, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.858, + 0.451 + ], + "angle": 0, + "content": "[66] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. arXiv preprint arXiv:2309.12188, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.858, + 0.528 + ], + "angle": 0, + "content": "[67] Guangyao Zhai, Dianye Huang, Shun-Cheng Wu, HyunJun Jung, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Monograspnet: 6-dof grasping with a single rgb image. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1708-1714. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.858, + 0.592 + ], + "angle": 0, + "content": "[68] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. arXiv preprint arXiv:2305.16283, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.858, + 0.659 + ], + "angle": 0, + "content": "[69] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.858, + 0.723 + ], + "angle": 0, + "content": "[70] Chenyangguang Zhang, Yan Di, Ruida Zhang, Guangyao Zhai, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ddf-ho: Hand-held object reconstruction via conditional directed distance field. arXiv preprint arXiv:2308.08231, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.725, + 0.858, + 0.789 + ], + "angle": 0, + "content": "[71] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European Conference on Computer Vision, pages 655–672. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.858, + 0.865 + ], + "angle": 0, + "content": "[72] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-posed: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452-7459. IEEE, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.858, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.903 + ], + "angle": 0, + "content": "21027" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.127, + 0.472, + 0.199 + ], + "angle": 0, + "content": "[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu-jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "21028" + } + ] +] \ No newline at end of file diff --git a/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_origin.pdf b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b1b4cda183b3b62edaf1317a3eee9e953159efe4 --- /dev/null +++ b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/8238426b-d0c8-4a68-b942-49f91f1e7a27_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc60f16bd2a4e91f8e3e7ff2128c0cde13be59e899f023e552a969dff77100de +size 1363797 diff --git a/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/full.md b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..01853e043cc1d909ce733965a4f5d0f057a68723 --- /dev/null +++ b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/full.md @@ -0,0 +1,344 @@ +# ShapeMatcher: Self-Supervised Joint Shape Canonicalization, Segmentation, Retrieval and Deformation + +Yan Di $^{1*}$ , Chenyangguang Zhang $^{2*}$ , Chaowei Wang $^{3*}$ , Ruida Zhang $^{2}$ , Guangyao Zhai $^{1}$ , Yanyan Li $^{1}$ , Bowen Fu $^{2}$ , Xiangyang Ji $^{2}$ and Shan Gao $^{†3}$ + +1Technical University of Munich, 2Tsinghua University, 3Northwestern Polytechnical University + +shangbuhuan13@gmail.com, gaoshan@nwpu.edu.cn + +# Abstract + +In this paper, we present ShapeMaker, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonicalize the object by extracting point-wise affine-invariant features, disentangling inherent structure of the object with its pose and size. These learned features are then leveraged to predict semantically consistent part segmentation and corresponding part centers. Next, our lightweight retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. The key insight of ShapeMaker is the simultaneous training of the four highly-associated processes: canonicalization, segmentation, retrieval, and deformation, leveraging cross-task consistency losses for mutual supervision. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that ShapeMaker surpasses competitors by a large margin. Code is released at https://github.com/Det1999/ShapeMaker. + +# 1. Introduction + +In recent years, there has been a notable surge in research interest focused on generating high-quality 3D models from scans of complex scenes [6, 15, 16, 21, 35, 49, 73]. This technology encourages extensive applications in both artistic creation [56, 57], robotics [66, 67] and 3D scene perception [41, 69]. Existing methods [41, 54] typically directly utilize deep neural networks to reconstruct 3D + +![](images/75860b2b8ebd284198f99c6a65466feff5cafc810e2e938a9b587a757931ea73.jpg) +Figure 1. Illustration of ShapeMatcher. Objects obtained from real-world scans are typically noisy, partial and exhibit various poses, making it challenging to conduct an effective $R \& D$ process (Red 'X' on the left). To address this issue, we propose Shape-Matcher that first canonicalizes the objects and then segments them into semantic parts, facilitating R&D processes (Green '√' on the right). + +models from imperfect scans. However, the presence of noise and occlusions poses a significant challenge in accurately capturing fine-grained geometric structures. To overcome this, Retrieval and Deformation (R&D) techniques [10,28,40,46,56-58,64] have been developed. These methods generally involve two key steps: first, identifying the most geometrically similar source shape from a precurated 3D database; and second, deforming the retrieved shape to achieve precise alignment with the target input. The R&D approach is particularly effective in producing 3D models enriched with fine details from source shapes. + +However, existing R&D methods usually encounter two primary challenges that make them susceptible to noise, occlusion and pose variations, and difficult to be practically utilized. 1) Most R&D techniques [10, 17, 28, 40, 46, 56-58, 64] operate under the assumption that target shapes are aligned in a pre-processed canonical space. Typ + +ically, these methods are trained and tested on datasets where shapes have been manually adjusted to this canonical state. However, when these methods are deployed in real-world settings, they necessitate either manual alignment of scanned objects or the use of additional pose estimation networks [14, 18, 48, 71, 72]. Such procedures are not only time-intensive and laborious but also prone to yielding inconsistent results. This limitation significantly impedes the direct application of these methods in real-world scenarios. 2) Previous methods [56, 57] do not design specially for dealing with partially-observed shapes, making it difficult to handle occluded objects. Although U-RED [17] considers the partial target shapes as input in the R&D process, it directly encodes the shape as a global embedding, which is not robust when dealing with significant occlusion. + +To address the aforementioned two challenges, in this paper, we present ShapeMatcher, a novel framework that extends traditional R&D pipeline to joint self-supervised learning of object canonicalization, segmentation, retrieval and deformation. Our core contribution lies in that the four highly-associated processes can be trained simultaneously and supervise each other via constructing several cross-task consistency losses (Fig. 1). Specifically, given a partially-observed object scan in an arbitrary pose, ShapeMatcher processes the objects in four steps. First, we follow [30], which is based on Vector Neurons [13], to extract SE(3)-invariant point-wise features by progressively separating translation and rotation. We further follow [8] to normalize the features to disassociate the object scale. Until here, we successfully obtain affine-invariant point-wise features by disentangling object's inherent structure with its pose and size. This facilitates the Canonicalization of the observed object based on these intrinsic characteristics (Fig. 2 (B)). Then we predict semantically consistent part segmentation and corresponding part centers by feeding the learned features into our Segmentation module (Fig. 2 (C)). Based on the part segmentation, in the Retrieval module (Fig. 2 (D)), we aggregate features within each part and collect them together as a comprehensive retrieval token of the object. For partial objects, we introduce a region-weighted strategy, which assigns a weight to each part according to the point inside it. Parts with more points are assigned higher weights during retrieval, which is proved to be robust to occlusions. We compare the tokens of the target object with each shape in the pre-constructed database to identify the most geometrically similar (most similar tokens) source shape. In the final Deformation module (Fig. 2 (E)), the retrieved source shape is deformed to tightly match the target object via part center guided neural cage deformation [64]. + +To summarize, our main contributions are: + +- We introduce ShapeMatcher, a novel self-supervised framework for joint shape canonicalization, segmentation, retrieval and deformation, handling partial target + +inputs under arbitrary poses. Extensive experiments on the synthetic and real-world datasets demonstrate that ShapeMatcher surpasses existing state-of-the-art approaches by a large margin. + +- We demonstrate that the four highly-associated tasks: canonicalization, segmentation, retrieval and deformation, can be effectively trained simultaneously and supervise each other via constructing consistencies. +- We develop the region-weighted retrieval method to mitigate the impact of occlusions in the R&D process. + +# 2. Related Works + +Neural Shape Representation. The compact representation of 3D shapes in latent space, based on deep learning, has been a focal point for many researchers. Some attempts, such as [9, 27, 36, 37, 42, 45, 61, 70], employ neural networks to construct an implicit function, while others [1, 38, 52, 62, 68] directly model the shape of objects explicitly using generative models. Another common architecture in 3D shape representation learning, as seen in [12, 29, 44, 60, 63], is to use an encoder-decoder approach to generate latent representation vectors for various shapes. Although these methods have demonstrated impressive representation performance, they often struggle to generate fine-grained shapes when dealing with occlusion and noise. + +CAD Model Retrieval and Deformation. Retrieval and Deformation (R&D) methods lead another way to recover fine-grained geometric structures. Previous works directly retrieve the most similar CAD models by comparing the similarity of expression vectors in either descriptor space [5,46] or the latent space of neural networks [4,10,22,34]. Considering the subsequent deformation error, recent efforts introduce deformation-aware embeddings [56] or proposed new optimization objectives [24] to better capture the fine structure of deformed target objects. Nevertheless, these methods yield deteriorated performance when facing partial and pose-agnostic target shapes in real world. [17] achieves an one-to-many retrieval module for addressing the issue caused by partial observations, however, it receives canonicalized target shapes as input, which limit its applicability facing pose-agnostic target shapes in real world. As the retrieved models often exhibit some deviation from the target shape, the deformation module is used to minimize this discrepancy. Traditional approaches [20, 23, 47] aim to fit the target shape by directly optimizing the deformed shape. Neural network based techniques attempt to learn a set of deformation priors from a database of models. They represent deformations as volume warping [25,32], cage deformations [64], vertex offsets [58], or flows [28,65]. These methods typically constrain two shapes are aligned in the same coordinate system, making them challenging to apply in real-world scenarios. + +SO(3)-Equivariant Methods. An increasing body of work [2,19,31,55,59] has initiated research on SO(3) equivariance. These efforts are mostly based on steerable convolutional kernels [33]. On the other hand, another set of works achieves equivariance through pose estimation. [43] estimates the object's pose to factor out SO(3) transformations, achieving approximate equivariance. While [51] learns pose estimation in a fully unsupervised manner, the equivariant backbone they employ [50] achieves equivariance primarily through data augmentation, leading to limited generalization. In this paper, we employ Vector Neural Multi-Layer Perceptron [13] as the backbone to get neural invariant features for object canonicalization. It achieves SO(3) equivariance by lifting traditional scalar neurons to vector neurons. + +# 3. Method + +Overview. ShapeMatcher consists of 4 modules, corresponding to the 4 highly-associated tasks. Each of the first 3 modules: Canonicalization, Segmentation and Retrieval modules have two parallel branches, one for complete point cloud (in orange background) and the other for partial object input (in blue background). As shown in Fig. 2, given a partial target shape $S_{tgt} \in \mathbb{R}^{N \times 3}$ in an arbitrary pose, in the Canonicalization module, we progressively decouples its inherent shape with rotation $R_{tgt} \in SO(3)$ , translation $T_{tgt} \in \mathbb{R}^3$ and the 3D metric size $s_{tgt} \in \mathbb{R}^3$ via VN-MLP [8, 13, 30], yielding the affine-invariant point-wise features $F_{tgt}$ . The object can then be canonicalized via inverse transformation based on intrinsics $\{R_{tgt}, T_{tgt}, s_{tgt}\}$ . In the Segmentation module, $F_{tgt}$ is fed into a 4-layer MLP network to predict $M$ parts and corresponding part centers $\{K_{tgt}^1, K_{tgt}^2, ..., K_{tgt}^M\}$ . The segmentation is semantically consistent across each category and thus can be matched and compared for R&D. In the Retrieval module, inside each region $M^i$ , we aggregate the features of all points inside it as its retrieval token $Q^i$ . The retrieval token for the object is then represented as $\mathbf{Q}_{\mathrm{tg}} = \{Q_{tgt}^1, Q_{tgt}^2, ..., Q_{tgt}^M\}$ . Similarly, during training, we obtain the intrinsics $\{R_{src}, T_{src}, s_{src}\}$ , part centers $\{K_{src}^1, K_{src}^2, ..., K_{src}^M\}$ and retrieval tokens $\mathbf{Q}_{\mathrm{src}} = \{Q_{src}^1, Q_{src}^2, ..., Q_{src}^M\}$ via the branch for complete point cloud. By comparing $\mathbf{Q}_{\mathrm{tg}}$ and $\mathbf{Q}_{\mathrm{src}}$ of each source shape inside the database, we identify the most geometrically similar source shape $S_r$ . In the final Deformation module, $K_{tgt}$ and $K_{src}$ are leveraged to guide the neural cage deformation [64] to deform the retrieved $S_r$ towards $S_{tgt}$ , yiling $S_{src}^{df_m}$ . + +# 3.1. Canonicalization + +As shown in Fig. 2 (B), the Canonicalization module takes the target point cloud $S_{tgt}$ as input and disentangle the inherent structure of $S_{tgt}$ with the intrinsics + +$\{R_{tgt}, T_{tgt}, s_{tgt}\}$ , yielding a point-wise affine-invariant feature $F_{tgt}$ . Specifically, we follow VN-MLP [8, 13, 30] to first decouple translation via VNT [30] and then extract rotation via VNN [13, 30]. We further follow [8] to normalize the SE(3)-invariant features obtained above the remove the influence of scaling, yielding $F_{tgt}$ as follows, + +$$ +R _ {t g t}, T _ {t g t}, F _ {t g t} ^ {*} = \mathbf {V N - M L P} \left(S _ {t g t}\right) \tag {1} +$$ + +$$ +s _ {t g t}, F _ {t g t} = \text {n o r m a l i z e} \left(F _ {t g t} ^ {*}\right) \tag {2} +$$ + +where $F_{tgt}^{*}$ denotes the SE(3)-invariant features and $F_{tgt}$ denotes the affine-invariant features. Thereby, the object can be canonicalized with intrinsics as, + +$$ +S _ {t g t} ^ {c} = s _ {t g t} R _ {t g t} S _ {t g t} + T _ {t g t} \tag {3} +$$ + +where $S_{tgt}^{c}$ denotes the normalized and canonicalized shape of $S^{tgt}$ . During training, in order to ensure that $F_{tgt}$ fully encapsulate the geometric information of $S_{tgt}$ , we integrate a supplementary reconstruction branch which takes $F_{tgt}^{*}$ as input and reconstruct $S_{tgt}^{c}$ in the affine-invariant space [30]. Please refer to the Supplementary Material for details. For source shape $S_{src}$ from the database, we follow the same procedures to extract $F_{src}$ . + +# 3.2. Segmentation + +Given the affine-invariant features $F_{tgt}$ , we segment the input point cloud $S_{tgt}$ into $M$ semantically consistent parts. We use a 4-layer MLP $\Theta_l$ (Fig. 2 (C)) to predict a one-hot segmentation label for each point and use another 4-layer MLP $\Theta_c$ to predict $M$ part centers $\{K_{src}^1, K_{src}^2, \dots, K_{src}^M\}$ . Noteworthy, we don't need any ground truth annotations in this segmentation process. Our experiments show that the network can automatically learn semantically consistent segmentation solely through consistency supervision from the other three tasks. For source shape $S_{src}$ , we follow a similar process to obtain $\{K_{src}^1, K_{src}^2, \dots, K_{src}^M\}$ . + +# 3.3. Retrieval + +The retrieval network aims to identify the model $S_{src}$ from an existing database that bears the closest resemblance to the target object $S_{tgt}$ after deformation. Traditional methods [17, 57] directly extract the global features of objects for retrieval, which typically struggles with heavy occlusion since the global features are susceptible to noise and occlusion, and prone to producing erroneous retrieval results. In contrast, we employ a novel region-weighted retrieval method to explicitly encode independent and semantically consistent regions of the shape. This allows us to accurately handle partial shapes by identifying the visible regions to retrieve models most similar to the target. + +Specifically, the part segmentation network $\Theta_l$ takes $F_{tgt}$ as input to predict $M$ regions $F_{seg} = [C_1, C_2, C_3, \dots, C_N]^\top$ . + +![](images/f6fef8bdde9557f419481033fdd3cc5b156b3ee396d4175f767dadd913338958.jpg) +Figure 2. The pipeline of ShapeMatcher. Given a target point cloud obtained from a single-view scan and a pre-established database (A), ShapeMatcher generates the fine-grained reconstruction result using the joint 4 modules including Canonicalization (B), Segmentation (C), Retrieval (D) and Deformation (E), where the first three contains the partial branch for target processing and the full branch for source processing. Specifically, the target and source inputs are first canonicalized into the same affine-invariant space (B). Then, the semantic-consistent region segmentation is yielded from the affine-invariant features (C). The segmented regions are fed to the region-weight retrieval module (C) and the part center guided neural cage deformation module (E) for occlusion-robust R&D process. During training, the partial-full consistency losses (F) are enforced for the two branches. + +![](images/849bcfa63bab6cd20b47261ff51c697d1e8394d7a6a4f4d131345f8c6d5e5365.jpg) + +![](images/48f3f357bba1587a2666c132a58a1a3314a714b1f74abdc338be666efec4650a.jpg) + +of $S_{tgt}^{c}$ , where $F_{seg} \in \mathbb{R}^{N \times M}$ , $C_i \in \mathbb{R}^M$ represents the probability of point $i$ belonging to each part center. Then we use another 4-layer feature aggregator $\Theta_f$ (Fig. 2 (D)) to extract the retrieval tokens $Q$ of all parts as follows, + +$$ +F _ {c l s} = F _ {s e g} ^ {\top} * \Theta_ {f} (F _ {t g t}), \tag {4} +$$ + +$$ +Q _ {t g t} = F _ {c l s} / \left(\sum_ {n = 1} ^ {N} F _ {s e g} ^ {(n)}\right). \tag {5} +$$ + +where $Q_{tgt} \in \mathbb{R}^{M \times C}$ contains the $C$ -dimensional retrieval tokens for all the $M$ regions. Here we employ a soft assignment strategy where each point inside $S_{tgt}$ is estimated $M$ values describing the probabilities belonging to each of the $M$ parts. Therefore, we first aggregate features $F_{cls}$ on all points belonging to each part and then normalize $F_{cls}$ using the sum of probabilities of points in each part, as in Eq. 4 and Eq. 5. Following a similar strategy, we can obtain $Q_{src}$ for each source model in the pre-curated database. We just need to compare $Q_{tgt}$ with the retrieval tokens $Q_{src}$ of all source shapes using the weighted $\mathcal{L}_1$ distance, + +$$ +D i s = \omega \sum \mathcal {L} _ {1} \left(Q _ {t g t} - Q _ {s r c}\right) \tag {6} +$$ + +where vector $\omega \in \mathbb{R}^{1\times M}$ stores the ratio of point number + +of each part with respect to the total point number $N$ . Intuitively, parts with smaller point numbers contribute less in calculating the distance score, which reduces the influence of noise and occlusion. The source shape $S_{r}$ with the smallest distance score is identified as the best retrieval. + +# 3.4. Deformation + +The Deformation module aims to deform the retrieved shape $S_{r}$ to tightly match the target shape $S_{tgt}$ . We utilize the neural cage scaffolding strategy as in [26, 64]. First, the neural cages $C_{src}$ for $S_{r}$ is pre-calculated. We utilize the part centers $(K_{tgt}$ and $K_{src})$ to control the vertex offsets $C_{src2tgt}$ of the neural cage $C_{src}$ to match $S_{tgt}$ . In particular, we employ a neural network $\Theta_I$ to predict an influence vector $I\in \mathbb{R}^{N_c\times M}$ for each point concerning all cage vertices by $I = \Theta_I(\text{concat}(F_{tgt},F_{src}))$ , where $N_{c}$ denotes the number of vertices used in $C_{src}$ . $C_{src2tgt}$ is computed through the influence vectors $I$ and the differences between region centers (Fig. 2 (E)): + +$$ +C _ {s r c 2 t g t} = C _ {s r c} + \sum_ {i = 1} ^ {M} I _ {i} \left(K _ {s r c} ^ {(i)} - K _ {t g t} ^ {(i)}\right), \tag {7} +$$ + +Finally, we employ a sparse cage scaffolding strategy [26, 64] to achieve the deformation field of $S_{src}$ . The deformed + +shape $S_{src2tgt}$ of $S_{src}$ can be expressed as follows: + +$$ +S _ {s r c 2 t g t} = S _ {s r c} + \Psi \left(C _ {s r c}, C _ {s r c 2 t g t}\right), \tag {8} +$$ + +where $\Psi$ computes the displacement of each point in $S_{src}$ by evaluating the differences between $C_{src}$ and $C_{src2tgt}$ , thereby achieving deformation. + +# 3.5. ShapeMatcher: Joint Training + +Our core insight in ShapeMatcher is that the four highly-associated tasks: Canonicalization, Segmentation, Retrieval and Deformation can be trained simultaneously and supervise each other via introducing cross-task consistency terms. We mainly introduce two types of losses here, i.e. partial-full consistency losses and task-oriented loss. For more details, please refer to the Supplementary Material. + +Task-Oriented Loss. In the Canonicalization, we mainly use Chamfer Distance to constrain the canonicalized $S_{tgt}^{c}$ and $\hat{S}_{tgt}^{c}$ predicted in the affine-invariant space by the supplementary reconstruction branch, to enforce the affine-invariety of $F_{tgt}^{*}$ in Sec. 3.1, by + +$$ +\mathcal {L} _ {\text {c a n}} = \operatorname {d i s} _ {\text {c h a m}} \left(S _ {t g t} ^ {c}, \hat {S} _ {t g t} ^ {c}\right) + \operatorname {o r t h} \left(R _ {t g t}\right), \tag {9} +$$ + +with $\text{orth}(R_{tgt})$ serving the purpose of enforcing the orthogonality of matrices. + +In the Segmentation, to keep consistency between the part segmentation and the predicted part center, we jointly train $\Theta_{l}$ and $\Theta_{c}$ with the following loss, which enforce that each predicted part center $K_{tgt}^{i}$ approximately lies in the center of all points belonging to the part $M_{i}$ , + +$$ +\mathcal {L} _ {s e g} = \sum_ {m = 1} ^ {M} \| K _ {t g t} ^ {(m)} - \left(F _ {s e g} ^ {\top} * S _ {t g t} ^ {c}\right) ^ {(m)} \| _ {2} \tag {10} +$$ + +To train the Retrieval and Deformation simultaneously, for an input target $S_{tgt}$ , we randomly select a source model $S_{src}$ from the database for training. Specifically, to eliminate the influence of occlusion, we do not directly use the global Chamfer Distance of $S_{tgt}$ and $S_{src}$ as ground truth. Instead, we employ a regional supervision strategy, ensuring that occluded areas do not contribute to the training of retrieval network. Taking the $i$ -th region as an example, $S_{tgt}^{i}$ represents all points in $S_{tgt}$ that belong to the $i$ -th region. We calculate the average of the nearest distances $D_{i}$ from each point in $S_{tgt}^{i}$ to the deformed shape $S_{src}^{dfm}$ to enforce the learning of the regional retrieval tokens by + +$$ +\mathcal {L} _ {\text {r e t r i e v a l}} = \frac {1}{M} \sum_ {i = 1} ^ {M} M S E \left(Q _ {t g t} ^ {(i)} - Q _ {s r c} ^ {(i)}, D _ {i}\right). \tag {11} +$$ + +The deformation loss is achieved by directly constraining the Chamfer Distance between $S_{tgt}$ and $S_{src}^{dfm}$ , expressed as: + +$$ +\mathcal {L} _ {\text {d e f o r m}} = \left. d i s _ {\text {c h a m}} \left(S _ {t g t}, S _ {s r c} ^ {d f m}\right) + \| I \| _ {2} \right. \tag {12} +$$ + +where we regularize $I$ using the L2 norm. + +Partial-Full Consistency Losses. In the first two modules: Canonicalization, Segmentation, the full branch serves as a guidance to enhance the learning of the partial branch. Therefore, in each module, we can enforce corresponding consistency terms between the outputs of the two parallel branches (Fig. 2 (F)). + +During the consistency training process, for randomly selected full input $S_{full}$ , we generate a mask $U_{f2p} \in \mathbb{R}^N$ to crop it to simulate the situation of a partial input $S_{partial} = S_{full}U_{f2p}$ . + +In the Canonicalization module, to enforce the consistency in the affine-invariant space between the two branches, we apply the same transformation $U_{f2p}$ to $S_{full}^{c}$ in the affine-invariant space as before and then use the chamfer distance to constrain its distance to $S_{partial}^{c}$ : + +$$ +\mathcal {L} _ {c c a n} = \operatorname {d i s} _ {\text {c h a m}} \left(S _ {\text {p a r t i a l}} ^ {c}, S _ {\text {f u l l}} ^ {c} U _ {f 2 p}\right) \tag {13} +$$ + +Similarly, in the Segmentation module, for the consistency constraint of the part center prediction network $\Theta_{c}$ , we directly use the Chamfer Distance to constrain the region centers detected by the two branches: + +$$ +\mathcal {L} _ {c c e n} = \operatorname {d i s} _ {\text {c h a m}} \left(K _ {\text {p a r t i a l}}, K _ {\text {f u l l}}\right). \tag {14} +$$ + +In the segmentation network $\Theta_{l}$ , we mask the segmentation results of the full branch $F_{seg}^{(full)}$ and compare them with the results of the partial branch $F_{seg}^{(partial)}$ : + +$$ +\mathcal {L} _ {c s e g} = d i s _ {c h a m} \left(F _ {s e g} ^ {(f u l l)} U _ {f 2 p}, F _ {s e g} ^ {(p a r t i a l)}\right). \tag {15} +$$ + +Joint Training. Generally, the joint training of Shape-Matcher is divided into three stages. First, we train the full branch by $\mathcal{L}_{can}$ and $\mathcal{L}_{seg}$ for construction of Canonicalization and Segmentation ability. Second, the partial branch is introduced and trained by both the task-oriented losses for Canonicalization and Segmentation $\mathcal{L}_{can}$ and $\mathcal{L}_{seg}$ and the partial-full consistency loss terms $\mathcal{L}_{ccan}$ , $\mathcal{L}_{ccen}$ and $\mathcal{L}_{cseg}$ . Finally, after training Canonicalization and Segmentation of the both branches, $\mathcal{L}_{\text{retrieval}}$ and $\mathcal{L}_{\text{deform}}$ are adopted for joint Retrieval and Deformation training simultaneously utilizing the both branches to handle partial target inputs and full source inputs respectively. + +# 4. Experiments + +# 4.1. Experimental Setup + +In this section, we mainly focus on R&D experiments, which better reflects the overall performance of the system. The ablations and analysis also demonstrate the effectiveness of considering joint Canonicalization and Part Segmentation. + +Datasets. We evaluate the effectiveness of our joint framework using three datasets: two synthetic datasets, + +PartNet [39] and ComplementMe [53], and one real-world dataset, Scan2CAD [3]. For datasets PartNet and ComplementMe, we follow the same database splits as in [57], separating their target inputs into training and testing sets. In our training process, we exclusively employ mesh models and do not utilize part segmentations as in [57], since the process of ShapeMatcher is fully self-supervised and does not need any additional annotations. The shapes used in PartNet and ComplementMe datasets are sourced from ShapeNet [7]. PartNet comprises 1,419 source models in the database, with 11,433 target models in the training set and 2,861 in the testing set. In ComplementMe, the numbers are 400, 11,311 and 2,825 respectively. In the synthetic cases, three categories of tables, chairs, and cabinets are evaluated on both datasets. Scan2CAD [3] is a real-world dataset developed based on ScanNet [11] with capacity of 14,225 objects. The input point cloud data on Scan2CAD is generated by reverse-projecting the depth images. In the real-world cases, we conduct training on the categories of tables, chairs, and cabinets from PartNet and directly testing on Scan2CAD. + +Baselines. Both baseline methods, Uy et al. [57] and U-RED [17], are trained using the same data partitioning strategy stated above. To ensure fairness in comparison with ShapeMatcher, we augment the training data with pose variations, keeping other hyperparameters consistent with the original paper. During testing, we evaluate scenarios where target observations with arbitrary poses are directly used as input. Additionally, we test scenarios where the inputs are transformed using an offline pose estimation method [18], simulating the two-stage route of traditional methods with pre-canonicalizing (Uy et al. [57] + PE and U-RED [17] + PE). For experiments on Scan2CAD, we directly use the baseline models trained on PartNet with the 25% occlusion setting to conduct zero-shot testing, since real-world ground-truth models are inaccessible for training. + +Evaluation Metrics. We utilize Chamfer Distance (CD) on the magnitude of $10^{-2}$ to assess both full shape scenarios and partial shape scenarios. We calculate the metrics following [57] to use the best result among the top 10 candidate objects. The final average metrics are obtained by averaging the results across all instances. + +Implementation Details. During training, we uniformly sample objects to obtain point clouds with $M = 2500$ points to represent shapes. We directly generate partial point clouds from the corresponding full point clouds by random cropping for the partial branch inputs. We apply random pose augmentation to the input shape, specifically with random translations $T_{rand} \in [-0.1, 0.1]$ and random rotations $R_{rand} \in [-1, 1]$ on three Eulerian angles respectively. We set the initial learning rate to $1e - 3$ and train ShapeMatcher for 200 epochs in every training stage of Sec. 3.5. Regarding the weight of the loss, in the first + +stage considering only the full branch, $\mathcal{L}_{can}$ and $\mathcal{L}_{seg}$ are equally weighted. In the second stage introducing the partial branch, we primarily emphasize the partial-full consistency losses, assigning significant weights to $\mathcal{L}_{ccan}$ , $\mathcal{L}_{ccen}$ and $\mathcal{L}_{cseg}$ with weights set as 5, 2, and 2 respectively, while keeping the remaining weights default at 1. In the final stage for joint R&D, both $\mathcal{L}_{retrieval}$ and $\mathcal{L}_{deform}$ are equally weighted. + +# 4.2. Synthetic Cases + +To validate the ability of ShapeMatcher tackling the challenge of arbitrary poses and occlusions, we first use synthetic datasets to simulate this scenario. We evaluate all methods [17, 57] where object observations with arbitrary poses are directly used as input. Additionally, we also report results where inputs are transformed and canonicalized using an offline pose estimation method [18] for baseline methods (Uy et al. + PE and U-RED + PE). Moreover, we analyze inference time of ShapeMatcher against the R&D baselines in the Supplementary Material. + +We conduct two types of inputs for evaluation: full inputs using the PartNet and ComplementMe datasets, and partial input tests using $10\%$ , $25\%$ , and $50\%$ occlusion rates on the PartNet dataset. The results of the full input tests are detailed in Table 1. In PartNet, our ShapeMatcher significantly outperforms the current leading competitors. For the Chamfer Distance on three categories, ShapeMatcher measures at 0.197, 0.150, and 0.519, maintaining the leading position. Even when the processed PE results are used as input, the baselines' results still fall short of ShapeMatcher. This demonstrates the effectiveness of adopting the affine-invariant features in the joint Canonicalization step. Results from ComplementMe supports the same conclusion, where ShapeMatcher reports significantly better results compared to the baseline methods. ShapeMatcher surpasses the top-performing Uy et al. + PE by $85.2\%$ . Such superior results yielded by ShapeMatcher demonstrates that the joint consideration of all four steps improves the matching accuracy a lot. + +For evaluation on partial inputs, we control the occlusion rates of partial point clouds by controlling the position of the cropping planes onto the full point clouds. The evaluation on partial inputs are presented in Table 2. Concretely, ShapeMatcher outperforms the current top method handling partial inputs U-RED by 5.018, 5.666 and 7.241 at the occlusion rate of $10\%$ , $25\%$ and $50\%$ respectively. As the occlusion rate increases, the superiority of the ShapeMatcher method grows. Considering PE adopting, the same trend is exhibited. ShapeMatcher surpasses the U-RED + PE by 0.525, 0.949, 2.434 under three occlusion rates. It demonstrates that the proposed region-weighted retrieval brings strong robustness of our method against occlusion. + +As shown in Fig. 3 and 4, our results show more geo + +
PartNet [39]
MethodChairTableCabinetAverage
Uy et al. [57]4.2696.3024.1185.271
Uy et al. [57] + PE1.5073.0061.0702.219
U-RED [17]5.3314.9809.1415.463
U-RED [17] + PE1.0250.3591.4230.725
Ours0.1970.1500.5190.200
ComplementMe [53]
MethodChairTableCabinetAverage
Uy et al. [57]4.0185.480-4.825
Uy et al. [57] + PE1.4392.454-1.999
U-RED [17]8.5755.800-7.044
U-RED [17] + PE4.9540.847-2.688
Ours0.2530.328-0.294
+ +Table 1. The Chamfer Distance metrics for joint R&D results on full shapes under arbitrary poses. + +
OcclusionMethodChairTableCabinetAverage
10%Uy et al. [57]4.3726.3954.1795.365
Uy et al. + PE [57]1.5232.9821.1332.219
U-RED [17]6.0255.3755.2695.640
U-RED [17] + PE1.2071.0121.6691.147
Ours0.6760.4811.2120.622
25%Uy et al. [57]4.6546.9274.7505.795
Uy et al. + PE [57]1.8033.1951.6072.481
U-RED [17]5.1967.2158.1646.442
U-RED [17] + PE1.6841.3872.7951.625
Ours0.8780.6431.0710.776
50%Uy et al. [57]6.0709.3227.9297.841
Uy et al. + PE [57]3.3145.0304.5844.272
U-RED [17]8.6968.3877.6138.455
U-RED [17] + PE4.7222.0157.9033.628
Ours1.1971.0791.8721.194
+ +Table 2. The Chamfer Distance metrics for joint R&D results on partial shapes under arbitrary poses of PartNet dataset. + +metric resemblance to the targets compared to other methods. This is attributed to the suitable joint consideration of the four highly-associated processes, which accurately decouples the input poses, mapping them to a consistent space for accurate $R\& D$ . The region-weighted retrieval we employ explicitly eliminates the influence of occluded areas, allowing for a more precise matching with the source model. + +# 4.3. Real-world Cases + +We test the effectiveness of ShapeMatcher on real-world datasets. In such case, ShapeMatcher is trained on the synthetic PartNet with $25\%$ occlusion and directly tested on the partial scans of the real-world dataset Scan2CAD without manual pose adjustments. Table 3 displays our results, where our method significantly outperforms existing competitors. Particularly, compared to the U-RED, in three categories, the reported Chamfer Distance are reduced by $92\%$ , $96\%$ , and $94\%$ respectively. In comparison to Uy et al., the + +![](images/b9d19a29390a066187b4ac1a2dca0d7d5cd538e53dc1ab1f2296fc2976ac0159.jpg) +Figure 3. Qualitative R&D results with full target inputs on Part-Net. + +![](images/e8280cc161639c2877722806059ab8e09ffdc639c2697c9bd5b740ee01ba81e4.jpg) +Figure 4. Qualitative R&D results with partial target inputs on the occlusion rate of $25\%$ on PartNet. + +reported Chamfer Distance are reduced by $91\%$ , $98\%$ , and $92\%$ . Even compared with the use of PE on the two baseline methods, there remains a significant leap forward of Shape-Matcher. This zero-shot experiments on real-world further prove that the procedure of joint Canonicalization, Segmentation, Retrieval and Deformation possesses a strong domain adaptation ability, showcasing the great potential of ShapeMatcher in real-world applications. The visualizations on Scan2CAD are provided in the supplementary material. + +# 4.4. Ablation Studies + +We conduct ablation experiments on PartNet, mainly on two aspects. First, in the Canonicalization, we investigate the importance of disentangling different pose intrinsics in Table 4, and demonstrate the effectiveness of joint considering Canonicalization. Second, in Table 5, we ablate the region-weighted Retrieval and the part center guided Deformation. Moreover, the robustness against occlusion is + +
MethodChairTableCabinetAverage
Uy et al. [57]4.8867.6058.3356.181
Uy et al. [57] + PE3.3626.6577.2614.905
U-RED [17]5.4905.13110.0915.945
U-RED [17] + PE2.8933.1645.9573.354
Ours0.4230.1860.6540.375
+ +# analyzed in Table 2. + +Canonicalization Capability. To study the impact of different intrinsics of object poses in the Canonicalization process, we conduct ablations on decoupling of translations, rotations and scales. The results are presented in Table 4. Specifically, in row (1), we make no adjustments to the input poses. Thanks to the regional-level $R\& D$ process, it shows decent performance. However, there is still a noticeable gap compared to row (4), indicating the significance of our proposed joint Canonicalization. In row (2), we solely decouple the input translations, resulting in a decrease of $14\%$ in reported metrics. Moving to row (3), upon this foundation, we add decoupling for rotation, leading to a substantial decrease in reported Chamfer Distance, averaging at 0.244. In row (4), we introduce scale decoupling, resulting in another decrease in the reported metrics. It is evident that accurate decoupling of rotation is a crucial aspect of the success of the Canonicalization process. Moreover, it demonstrates that to integrate Canonicalization and $R\& D$ process is indispensable for the ShapeMatcher process. + +Deformation and Retrieval Ability. To validate the effectiveness of our proposed region-weighted Retrieval and the part center guided neural cage Deformation, we conduct an ablation study on the PartNet dataset with the $25\%$ occlusion rate. The results are presented in Table 5. In row (1), we conduct experiments using global retrieval and global deformation. This means we directly use an MLP network to extract overall point cloud features as the retrieval vector [57]. In the deformation network, similarly, we directly use an MLP network to generate neural cage offsets for deformation [26, 64]. Due to the lack of extraction of local information, the reported Chamfer Distance is more than twice of row (4). In row (2), we employ the global retrieval and the part center guided neural cage deformation. This improvement allows much more tightly-matched deformation by the retrieved source model, resulting in a $14\%$ decrease in reported metrics. In row (3), we conduct experiments using the regional retrieval and the global deformation. The proposed regional-weighted retrieval handles occluded objects, reducing the impact of occluded parts and resulting in a substantial decrease in Chamfer Distance, down to 0.973. + +Occlusion Robustness. We test shapes at different occlusion levels by altering the occlusion ratio in the input. + +Table 3. The Chamfer Distance metrics for joint R&D results in real-world Scan2CAD [3]. + +
Trans.Rot.Scal.ChairTableCabinetAverage
(1)0.5710.5021.2330.590
(2)✓0.4680.4421.0960.506
(3)✓✓0.2130.2000.6740.244
(4)✓✓✓0.1970.1500.5190.200
+ +Table 4. Ablations on the Canonicalization process, we demonstrate the effectiveness of the joint Canonicalization by ablating different pose intrinsics. Here, Trans. denotes the decoupling of translation, Rot. represents the rotation, and Scal. signifies the scale. + +
Gl. R.Re. R.Gl. D.Re. D.ChairTableCabinetAverage
(1)✓✓1.6721.4461.8001.570
(2)✓✓1.5391.3051.6411.431
(3)✓✓1.0420.8741.2230.973
(4)✓✓0.8780.6431.0710.776
+ +Table 5. Ablations of the $R\& D$ process. GL. R. denotes the global feature based retrieval [57], Re. R. represents the proposed region-weighted retrieval, GL. D. signifies direct neural cage deformation using global features [26, 64], and Re. D. denotes the adopted regional part center guided neural cage deformation. + +For each specific occlusion ratio, we deliberately crop a portion of the complete point cloud to simulate occlusion. We test scenarios with occlusion ratios of $10\%$ , $25\%$ , and $50\%$ , and the results are presented in Table 2. Observably, as the occluded regions increased, the Chamfer Distance significantly rises for the baseline methods. Taking the U-RED + PE as an example, its reported metrics increase from 1.147 to 3.628, doubling in value, as the occluded area expands. In contrast, our method increases by less than 1-fold, which exhibits strong robustness against occlusion. + +# 5. Conclusion + +In this paper, we present ShapeMatcher, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonize the object by extracting point-wise affine-invariant features. Then, the affine-invariant features are leveraged to predict semantically consistent part segmentation and corresponding part centers. Afterwards, the lightweight region-weighted retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that Shape-Matcher surpasses competitors by a large margin. In the future, we plan to further apply I-RED to various downstream tasks like robotic grasping. Limitations are discussed in the Supplementary Material. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2 +[2] Brandon Anderson, Truong Son Hy, and Risi Kondor. Cormorant: Covariant molecular neural networks. Advances in neural information processing systems, 32, 2019. 3 +[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6, 8 +[4] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 2 +[5] Frederic Bosche and Carl T Haas. Automated retrieval of 3d cad model objects in construction range images. Automation in Construction, 17(4):499-512, 2008. 2 +[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1 +[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6 +[8] Yunlu Chen, Basura Fernando, Hakan Bilen, Matthias Nießner, and Efstratios Gavves. 3d equivariant graph implicit functions. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 2, 3 +[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2 +[10] Manuel Dahnert, Angela Dai, Leonidas J Guibas, and Matthias Nießner. Joint embedding of 3d scan and cad objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8749-8758, 2019. 1, 2 +[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6 +[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868-5877, 2017. 2 +[13] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 3 + +[14] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-pose: Exploiting self-occlusion for direct 6d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12396-12405, 2021. 2 +[15] Yan Di, Henrique Morimitsu, Shan Gao, and Xiangyang Ji. Monocular piecewise depth estimation in dynamic scenes by exploiting superpixel relations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4363-4372, 2019. 1 +[16] Yan Di, Henrique Morimitsu, Zhiqiang Lou, and Xiangyang Ji. A unified framework for piecewise semantic reconstruction in dynamic scenes via exploiting superpixel relations. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10737-10743. IEEE, 2020. 1 +[17] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. arXiv preprint arXiv:2308.06383, 2023. 1, 2, 3, 6, 7, 8 +[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 2, 6 +[19] Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning so (3) equivariant representations with spherical cnns. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52–68, 2018. 3 +[20] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 2018 International Conference on 3D Vision (3DV), pages 672–681. IEEE, 2018. 2 +[21] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9785-9795, 2019. 1 +[22] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2 +[23] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, volume 27, pages 1449-1457. Wiley Online Library, 2008. 2 +[24] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 599-628. Springer, 2020. 2 +[25] Dominic Jack, Jhony K Pontes, Sridha Sridharan, Clinton Fookes, Sareh Shirazi, Frederic Maire, and Anders Eriksson. Learning free-form deformations for 3d object reconstruction. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, + +2018, Revised Selected Papers, Part II 14, pages 317-333. Springer, 2019. 2 +[26] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3d keypoint discovery for shape control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12783-12792, 2021. 4, 8 +[27] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2 +[28] Chiyu Jiang, Jingwei Huang, Andrea Tagliasacchi, and Leonidas J Guibas. Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems, 33:9745-9757, 2020. 1, 2 +[29] Jincen Jiang, Xuequan Lu, Lizhi Zhao, Richard Dazaley, and Meili Wang. Masked autoencoders in 3d point cloud representation learning. IEEE Transactions on Multimedia, 2023. 2 +[30] Oren Katzir, Dani Lischinski, and Daniel Cohen-Or. Shape-pose disentanglement using se (3)-equivariant vector neurons. In European Conference on Computer Vision, pages 468-484. Springer, 2022. 2, 3 +[31] Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch-gordan nets: a fully fourier space spherical convolutional neural network. Advances in Neural Information Processing Systems, 31, 2018. 3 +[32] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 2 +[33] Leon Lang and Maurice Weiler. A wigner-echart theorem for group equivariant convolution kernels. arXiv preprint arXiv:2010.10952, 2020. 3 +[34] Yangyan Li, Hao Su, Charles Ruizhongtai Qi, Noa Fish, Daniel Cohen-Or, and Leonidas J Guibas. Joint embeddings of shapes and images via cnn image purification. ACM transactions on graphics (TOG), 34(6):1-12, 2015. 2 +[35] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1 +[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2 +[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 + +[38] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2 +[39] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6, 7 +[40] Liangliang Nan, Ke Xie, and Andrei Sharf. A search-classify approach for cluttered indoor scene understanding. ACM Transactions on Graphics (TOG), 31(6):1-10, 2012. 1 +[41] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1 +[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2 +[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3 +[44] Yuchen Rao, Yinyu Nie, and Angela Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems, 35:34436-34450, 2022. 2 +[45] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2 +[46] Adriana Schulz, Ariel Shamir, Ilya Baran, David IW Levin, Pitchaya Sitthi-Amorn, and Wojciech Matusik. Retrieval on parametric shape collections. ACM Transactions on Graphics (TOG), 36(1):1-14, 2017. 1, 2 +[47] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116. CiteSeer, 2007. 2 +[48] Yongzhi Su, Yan Di, Guangyao Zhai, Fabian Manhardt, Jason Rambach, Benjamin Busam, Didier Stricker, and Federico Tombari. Opa-3d: Occlusion-aware pixel-wise aggregation for monocular 3d object detection. IEEE Robotics and Automation Letters, 8(3):1327–1334, 2023. 2 +[49] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1 +[50] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization + +for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11286-11295, 2020. 3 +[51] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 1, 2021. 3 +[52] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua Siegel, and Sanjay Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 61–70, 2020. 2 +[53] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 6, 7 +[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3405-3414, 2019. 1 +[55] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3 +[56] Mikaela Angelina Uy, Jingwei Huang, Minhyuk Sung, Tolga Birdal, and Leonidas Guibas. Deformation-aware 3d model embedding and retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 397-413. Springer, 2020. 1, 2 +[57] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 1, 2, 3, 6, 7, 8 +[58] Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. 3dn: 3d deformation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1038-1046, 2019. 1, 2 +[59] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco S Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. Advances in Neural Information Processing Systems, 31, 2018. 3 +[60] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2 +[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2 + +[62] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 2 +[63] Shuo Yang, Min Xu, Haozhe Xie, Stuart Perry, and Jiahao Xia. Single-view 3d object reconstruction from shape priors in memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3152-3161, 2021. 2 +[64] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 1, 2, 3, 4, 8 +[65] Michela Zaccaria, Fabian Manhardt, Yan Di, Federico Tombari, Jacopo Aleotti, and Mikhail Giorgini. Self-supervised category-level 6d object pose estimation with optical flow consistency. IEEE Robotics and Automation Letters, 8(5):2510-2517, 2023. 2 +[66] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. arXiv preprint arXiv:2309.12188, 2023. 1 +[67] Guangyao Zhai, Dianye Huang, Shun-Cheng Wu, HyunJun Jung, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Monograspnet: 6-dof grasping with a single rgb image. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1708-1714. IEEE, 2023. 1 +[68] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. arXiv preprint arXiv:2305.16283, 2023. 2 +[69] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1 +[70] Chenyangguang Zhang, Yan Di, Ruida Zhang, Guangyao Zhai, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ddf-ho: Hand-held object reconstruction via conditional directed distance field. arXiv preprint arXiv:2308.08231, 2023. 2 +[71] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European Conference on Computer Vision, pages 655–672. Springer, 2022. 2 +[72] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-posed: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452-7459. IEEE, 2022. 2 + +[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu-jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. \ No newline at end of file diff --git a/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/images.zip b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..43e531d45f0d8f3f709df7dfc6a3d0657ab1af72 --- /dev/null +++ b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:078da551f082aba5b0a85ad67e40e3f4700c005affc1df41aeb65384ac39cade +size 444090 diff --git a/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/layout.json b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..defe4fdf9dcefe393ff432794d04ab875b9969f3 --- /dev/null +++ b/2024/ShapeMatcher_ Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation/layout.json @@ -0,0 +1,11031 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 128, + 474, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 128, + 474, + 159 + ], + "spans": [ + { + "bbox": [ + 121, + 128, + 474, + 159 + ], + "type": "text", + "content": "ShapeMatcher: Self-Supervised Joint Shape Canonicalization, Segmentation, Retrieval and Deformation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "spans": [ + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": "Yan Di" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": ", Chenyangguang Zhang" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": ", Chaowei Wang" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": ", Ruida Zhang" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": ", Guangyao Zhai" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": ", Yanyan Li" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": ", Bowen Fu" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": ", Xiangyang Ji" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "text", + "content": " and Shan Gao" + }, + { + "bbox": [ + 78, + 180, + 517, + 206 + ], + "type": "inline_equation", + "content": "^{†3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 206, + 509, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 206, + 509, + 220 + ], + "spans": [ + { + "bbox": [ + 86, + 206, + 509, + 220 + ], + "type": "text", + "content": "1Technical University of Munich, 2Tsinghua University, 3Northwestern Polytechnical University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 221, + 406, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 221, + 406, + 232 + ], + "spans": [ + { + "bbox": [ + 188, + 221, + 406, + 232 + ], + "type": "text", + "content": "shangbuhuan13@gmail.com, gaoshan@nwpu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 157, + 270, + 201, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 270, + 201, + 281 + ], + "spans": [ + { + "bbox": [ + 157, + 270, + 201, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 293, + 290, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 293, + 290, + 557 + ], + "spans": [ + { + "bbox": [ + 68, + 293, + 290, + 557 + ], + "type": "text", + "content": "In this paper, we present ShapeMaker, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonicalize the object by extracting point-wise affine-invariant features, disentangling inherent structure of the object with its pose and size. These learned features are then leveraged to predict semantically consistent part segmentation and corresponding part centers. Next, our lightweight retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. The key insight of ShapeMaker is the simultaneous training of the four highly-associated processes: canonicalization, segmentation, retrieval, and deformation, leveraging cross-task consistency losses for mutual supervision. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that ShapeMaker surpasses competitors by a large margin. Code is released at https://github.com/Det1999/ShapeMaker." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 575, + 143, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 575, + 143, + 586 + ], + "spans": [ + { + "bbox": [ + 69, + 575, + 143, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 594, + 288, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 594, + 288, + 671 + ], + "spans": [ + { + "bbox": [ + 68, + 594, + 288, + 671 + ], + "type": "text", + "content": "In recent years, there has been a notable surge in research interest focused on generating high-quality 3D models from scans of complex scenes [6, 15, 16, 21, 35, 49, 73]. This technology encourages extensive applications in both artistic creation [56, 57], robotics [66, 67] and 3D scene perception [41, 69]. Existing methods [41, 54] typically directly utilize deep neural networks to reconstruct 3D" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 269, + 526, + 401 + ], + "blocks": [ + { + "bbox": [ + 308, + 269, + 526, + 401 + ], + "lines": [ + { + "bbox": [ + 308, + 269, + 526, + 401 + ], + "spans": [ + { + "bbox": [ + 308, + 269, + 526, + 401 + ], + "type": "image", + "image_path": "75860b2b8ebd284198f99c6a65466feff5cafc810e2e938a9b587a757931ea73.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 415, + 526, + 486 + ], + "lines": [ + { + "bbox": [ + 304, + 415, + 526, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 526, + 486 + ], + "type": "text", + "content": "Figure 1. Illustration of ShapeMatcher. Objects obtained from real-world scans are typically noisy, partial and exhibit various poses, making it challenging to conduct an effective " + }, + { + "bbox": [ + 304, + 415, + 526, + 486 + ], + "type": "inline_equation", + "content": "R \\& D" + }, + { + "bbox": [ + 304, + 415, + 526, + 486 + ], + "type": "text", + "content": " process (Red 'X' on the left). To address this issue, we propose Shape-Matcher that first canonicalizes the objects and then segments them into semantic parts, facilitating R&D processes (Green '√' on the right)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 526, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 526, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 526, + 618 + ], + "type": "text", + "content": "models from imperfect scans. However, the presence of noise and occlusions poses a significant challenge in accurately capturing fine-grained geometric structures. To overcome this, Retrieval and Deformation (R&D) techniques [10,28,40,46,56-58,64] have been developed. These methods generally involve two key steps: first, identifying the most geometrically similar source shape from a precurated 3D database; and second, deforming the retrieved shape to achieve precise alignment with the target input. The R&D approach is particularly effective in producing 3D models enriched with fine details from source shapes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 620, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 526, + 687 + ], + "type": "text", + "content": "However, existing R&D methods usually encounter two primary challenges that make them susceptible to noise, occlusion and pose variations, and difficult to be practically utilized. 1) Most R&D techniques [10, 17, 28, 40, 46, 56-58, 64] operate under the assumption that target shapes are aligned in a pre-processed canonical space. Typ" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "spans": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 207, + 45, + 430, + 56 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 45, + 430, + 56 + ], + "spans": [ + { + "bbox": [ + 207, + 45, + 430, + 56 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 190, + 57, + 448, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 57, + 448, + 67 + ], + "spans": [ + { + "bbox": [ + 190, + 57, + 448, + 67 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 676, + 259, + 686 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 676, + 259, + 686 + ], + "spans": [ + { + "bbox": [ + 82, + 676, + 259, + 686 + ], + "type": "text", + "content": "*Authors with equal contributions. † Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "21017" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 100, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 100, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 67, + 100, + 289, + 263 + ], + "type": "text", + "content": "ically, these methods are trained and tested on datasets where shapes have been manually adjusted to this canonical state. However, when these methods are deployed in real-world settings, they necessitate either manual alignment of scanned objects or the use of additional pose estimation networks [14, 18, 48, 71, 72]. Such procedures are not only time-intensive and laborious but also prone to yielding inconsistent results. This limitation significantly impedes the direct application of these methods in real-world scenarios. 2) Previous methods [56, 57] do not design specially for dealing with partially-observed shapes, making it difficult to handle occluded objects. Although U-RED [17] considers the partial target shapes as input in the R&D process, it directly encodes the shape as a global embedding, which is not robust when dealing with significant occlusion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 264, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 264, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 264, + 289, + 635 + ], + "type": "text", + "content": "To address the aforementioned two challenges, in this paper, we present ShapeMatcher, a novel framework that extends traditional R&D pipeline to joint self-supervised learning of object canonicalization, segmentation, retrieval and deformation. Our core contribution lies in that the four highly-associated processes can be trained simultaneously and supervise each other via constructing several cross-task consistency losses (Fig. 1). Specifically, given a partially-observed object scan in an arbitrary pose, ShapeMatcher processes the objects in four steps. First, we follow [30], which is based on Vector Neurons [13], to extract SE(3)-invariant point-wise features by progressively separating translation and rotation. We further follow [8] to normalize the features to disassociate the object scale. Until here, we successfully obtain affine-invariant point-wise features by disentangling object's inherent structure with its pose and size. This facilitates the Canonicalization of the observed object based on these intrinsic characteristics (Fig. 2 (B)). Then we predict semantically consistent part segmentation and corresponding part centers by feeding the learned features into our Segmentation module (Fig. 2 (C)). Based on the part segmentation, in the Retrieval module (Fig. 2 (D)), we aggregate features within each part and collect them together as a comprehensive retrieval token of the object. For partial objects, we introduce a region-weighted strategy, which assigns a weight to each part according to the point inside it. Parts with more points are assigned higher weights during retrieval, which is proved to be robust to occlusions. We compare the tokens of the target object with each shape in the pre-constructed database to identify the most geometrically similar (most similar tokens) source shape. In the final Deformation module (Fig. 2 (E)), the retrieved source shape is deformed to tightly match the target object via part center guided neural cage deformation [64]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 636, + 238, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 636, + 238, + 646 + ], + "spans": [ + { + "bbox": [ + 80, + 636, + 238, + 646 + ], + "type": "text", + "content": "To summarize, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 653, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 653, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 80, + 653, + 289, + 687 + ], + "type": "text", + "content": "- We introduce ShapeMatcher, a novel self-supervised framework for joint shape canonicalization, segmentation, retrieval and deformation, handling partial target" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 323, + 100, + 525, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 100, + 525, + 144 + ], + "spans": [ + { + "bbox": [ + 323, + 100, + 525, + 144 + ], + "type": "text", + "content": "inputs under arbitrary poses. Extensive experiments on the synthetic and real-world datasets demonstrate that ShapeMatcher surpasses existing state-of-the-art approaches by a large margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 149, + 525, + 222 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 316, + 149, + 525, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 149, + 525, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 149, + 525, + 194 + ], + "type": "text", + "content": "- We demonstrate that the four highly-associated tasks: canonicalization, segmentation, retrieval and deformation, can be effectively trained simultaneously and supervise each other via constructing consistencies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 200, + 525, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 525, + 222 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 525, + 222 + ], + "type": "text", + "content": "- We develop the region-weighted retrieval method to mitigate the impact of occlusions in the R&D process." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 231, + 389, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 389, + 242 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 389, + 242 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 249, + 525, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 525, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 525, + 390 + ], + "type": "text", + "content": "Neural Shape Representation. The compact representation of 3D shapes in latent space, based on deep learning, has been a focal point for many researchers. Some attempts, such as [9, 27, 36, 37, 42, 45, 61, 70], employ neural networks to construct an implicit function, while others [1, 38, 52, 62, 68] directly model the shape of objects explicitly using generative models. Another common architecture in 3D shape representation learning, as seen in [12, 29, 44, 60, 63], is to use an encoder-decoder approach to generate latent representation vectors for various shapes. Although these methods have demonstrated impressive representation performance, they often struggle to generate fine-grained shapes when dealing with occlusion and noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 391, + 525, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 525, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 525, + 685 + ], + "type": "text", + "content": "CAD Model Retrieval and Deformation. Retrieval and Deformation (R&D) methods lead another way to recover fine-grained geometric structures. Previous works directly retrieve the most similar CAD models by comparing the similarity of expression vectors in either descriptor space [5,46] or the latent space of neural networks [4,10,22,34]. Considering the subsequent deformation error, recent efforts introduce deformation-aware embeddings [56] or proposed new optimization objectives [24] to better capture the fine structure of deformed target objects. Nevertheless, these methods yield deteriorated performance when facing partial and pose-agnostic target shapes in real world. [17] achieves an one-to-many retrieval module for addressing the issue caused by partial observations, however, it receives canonicalized target shapes as input, which limit its applicability facing pose-agnostic target shapes in real world. As the retrieved models often exhibit some deviation from the target shape, the deformation module is used to minimize this discrepancy. Traditional approaches [20, 23, 47] aim to fit the target shape by directly optimizing the deformed shape. Neural network based techniques attempt to learn a set of deformation priors from a database of models. They represent deformations as volume warping [25,32], cage deformations [64], vertex offsets [58], or flows [28,65]. These methods typically constrain two shapes are aligned in the same coordinate system, making them challenging to apply in real-world scenarios." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "21018" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 99, + 290, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 99, + 290, + 263 + ], + "spans": [ + { + "bbox": [ + 68, + 99, + 290, + 263 + ], + "type": "text", + "content": "SO(3)-Equivariant Methods. An increasing body of work [2,19,31,55,59] has initiated research on SO(3) equivariance. These efforts are mostly based on steerable convolutional kernels [33]. On the other hand, another set of works achieves equivariance through pose estimation. [43] estimates the object's pose to factor out SO(3) transformations, achieving approximate equivariance. While [51] learns pose estimation in a fully unsupervised manner, the equivariant backbone they employ [50] achieves equivariance primarily through data augmentation, leading to limited generalization. In this paper, we employ Vector Neural Multi-Layer Perceptron [13] as the backbone to get neural invariant features for object canonicalization. It achieves SO(3) equivariance by lifting traditional scalar neurons to vector neurons." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 273, + 119, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 273, + 119, + 284 + ], + "spans": [ + { + "bbox": [ + 69, + 273, + 119, + 284 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "spans": [ + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": "Overview. ShapeMatcher consists of 4 modules, corresponding to the 4 highly-associated tasks. Each of the first 3 modules: Canonicalization, Segmentation and Retrieval modules have two parallel branches, one for complete point cloud (in orange background) and the other for partial object input (in blue background). As shown in Fig. 2, given a partial target shape " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "S_{tgt} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " in an arbitrary pose, in the Canonicalization module, we progressively decouples its inherent shape with rotation " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "R_{tgt} \\in SO(3)" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ", translation " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "T_{tgt} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " and the 3D metric size " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "s_{tgt} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " via VN-MLP [8, 13, 30], yielding the affine-invariant point-wise features " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ". The object can then be canonicalized via inverse transformation based on intrinsics " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\{R_{tgt}, T_{tgt}, s_{tgt}\\}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ". In the Segmentation module, " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " is fed into a 4-layer MLP network to predict " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " parts and corresponding part centers " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\{K_{tgt}^1, K_{tgt}^2, ..., K_{tgt}^M\\}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ". The segmentation is semantically consistent across each category and thus can be matched and compared for R&D. In the Retrieval module, inside each region " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "M^i" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ", we aggregate the features of all points inside it as its retrieval token " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "Q^i" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ". The retrieval token for the object is then represented as " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tg}} = \\{Q_{tgt}^1, Q_{tgt}^2, ..., Q_{tgt}^M\\}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ". Similarly, during training, we obtain the intrinsics " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\{R_{src}, T_{src}, s_{src}\\}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ", part centers " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\{K_{src}^1, K_{src}^2, ..., K_{src}^M\\}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " and retrieval tokens " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{src}} = \\{Q_{src}^1, Q_{src}^2, ..., Q_{src}^M\\}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " via the branch for complete point cloud. By comparing " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tg}}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{src}}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " of each source shape inside the database, we identify the most geometrically similar source shape " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ". In the final Deformation module, " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "K_{tgt}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "K_{src}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " are leveraged to guide the neural cage deformation [64] to deform the retrieved " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": " towards " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": ", yiling " + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "inline_equation", + "content": "S_{src}^{df_m}" + }, + { + "bbox": [ + 68, + 291, + 290, + 631 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 636, + 163, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 636, + 163, + 647 + ], + "spans": [ + { + "bbox": [ + 69, + 636, + 163, + 647 + ], + "type": "text", + "content": "3.1. Canonicalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 653, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 653, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 653, + 288, + 687 + ], + "type": "text", + "content": "As shown in Fig. 2 (B), the Canonicalization module takes the target point cloud " + }, + { + "bbox": [ + 68, + 653, + 288, + 687 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 68, + 653, + 288, + 687 + ], + "type": "text", + "content": " as input and disentangle the inherent structure of " + }, + { + "bbox": [ + 68, + 653, + 288, + 687 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 68, + 653, + 288, + 687 + ], + "type": "text", + "content": " with the intrinsics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "type": "inline_equation", + "content": "\\{R_{tgt}, T_{tgt}, s_{tgt}\\}" + }, + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "type": "text", + "content": ", yielding a point-wise affine-invariant feature " + }, + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "type": "text", + "content": ". Specifically, we follow VN-MLP [8, 13, 30] to first decouple translation via VNT [30] and then extract rotation via VNN [13, 30]. We further follow [8] to normalize the SE(3)-invariant features obtained above the remove the influence of scaling, yielding " + }, + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 305, + 99, + 525, + 166 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 347, + 173, + 525, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 173, + 525, + 186 + ], + "spans": [ + { + "bbox": [ + 347, + 173, + 525, + 186 + ], + "type": "interline_equation", + "content": "R _ {t g t}, T _ {t g t}, F _ {t g t} ^ {*} = \\mathbf {V N - M L P} \\left(S _ {t g t}\\right) \\tag {1}", + "image_path": "6cd1e36928e5501a69e065e881fc5faedb1c486a3f52801f5e66d788309e5d46.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 358, + 202, + 524, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 202, + 524, + 216 + ], + "spans": [ + { + "bbox": [ + 358, + 202, + 524, + 216 + ], + "type": "interline_equation", + "content": "s _ {t g t}, F _ {t g t} = \\text {n o r m a l i z e} \\left(F _ {t g t} ^ {*}\\right) \\tag {2}", + "image_path": "49bafbda4e01c7bb34dcf45372a6b527d563359e15f01f1bda10062da43e3982.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 219, + 525, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 525, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 525, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 219, + 525, + 251 + ], + "type": "inline_equation", + "content": "F_{tgt}^{*}" + }, + { + "bbox": [ + 305, + 219, + 525, + 251 + ], + "type": "text", + "content": " denotes the SE(3)-invariant features and " + }, + { + "bbox": [ + 305, + 219, + 525, + 251 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 305, + 219, + 525, + 251 + ], + "type": "text", + "content": " denotes the affine-invariant features. Thereby, the object can be canonicalized with intrinsics as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 364, + 259, + 525, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 259, + 525, + 273 + ], + "spans": [ + { + "bbox": [ + 364, + 259, + 525, + 273 + ], + "type": "interline_equation", + "content": "S _ {t g t} ^ {c} = s _ {t g t} R _ {t g t} S _ {t g t} + T _ {t g t} \\tag {3}", + "image_path": "8632f0326d28662d946a00d0475bb8262368dfd4beb3792b4f80a9d1ce7f3745.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "S_{tgt}^{c}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": " denotes the normalized and canonicalized shape of " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "S^{tgt}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": ". During training, in order to ensure that " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": " fully encapsulate the geometric information of " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": ", we integrate a supplementary reconstruction branch which takes " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "F_{tgt}^{*}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": " as input and reconstruct " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "S_{tgt}^{c}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": " in the affine-invariant space [30]. Please refer to the Supplementary Material for details. For source shape " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": " from the database, we follow the same procedures to extract " + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "inline_equation", + "content": "F_{src}" + }, + { + "bbox": [ + 304, + 278, + 525, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 373, + 385, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 385, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 385, + 384 + ], + "type": "text", + "content": "3.2. Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": "Given the affine-invariant features " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": ", we segment the input point cloud " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": " semantically consistent parts. We use a 4-layer MLP " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "\\Theta_l" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": " (Fig. 2 (C)) to predict a one-hot segmentation label for each point and use another 4-layer MLP " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "\\Theta_c" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": " part centers " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "\\{K_{src}^1, K_{src}^2, \\dots, K_{src}^M\\}" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": ". Noteworthy, we don't need any ground truth annotations in this segmentation process. Our experiments show that the network can automatically learn semantically consistent segmentation solely through consistency supervision from the other three tasks. For source shape " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": ", we follow a similar process to obtain " + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "inline_equation", + "content": "\\{K_{src}^1, K_{src}^2, \\dots, K_{src}^M\\}" + }, + { + "bbox": [ + 304, + 389, + 525, + 511 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 516, + 366, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 366, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 366, + 527 + ], + "type": "text", + "content": "3.3. Retrieval" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 525, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 525, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 525, + 663 + ], + "type": "text", + "content": "The retrieval network aims to identify the model " + }, + { + "bbox": [ + 304, + 533, + 525, + 663 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 304, + 533, + 525, + 663 + ], + "type": "text", + "content": " from an existing database that bears the closest resemblance to the target object " + }, + { + "bbox": [ + 304, + 533, + 525, + 663 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 304, + 533, + 525, + 663 + ], + "type": "text", + "content": " after deformation. Traditional methods [17, 57] directly extract the global features of objects for retrieval, which typically struggles with heavy occlusion since the global features are susceptible to noise and occlusion, and prone to producing erroneous retrieval results. In contrast, we employ a novel region-weighted retrieval method to explicitly encode independent and semantically consistent regions of the shape. This allows us to accurately handle partial shapes by identifying the visible regions to retrieve models most similar to the target." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "content": "Specifically, the part segmentation network " + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\Theta_l" + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "content": " as input to predict " + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "content": " regions " + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "inline_equation", + "content": "F_{seg} = [C_1, C_2, C_3, \\dots, C_N]^\\top" + }, + { + "bbox": [ + 305, + 664, + 525, + 687 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "21019" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 99, + 267, + 327 + ], + "blocks": [ + { + "bbox": [ + 83, + 99, + 267, + 327 + ], + "lines": [ + { + "bbox": [ + 83, + 99, + 267, + 327 + ], + "spans": [ + { + "bbox": [ + 83, + 99, + 267, + 327 + ], + "type": "image", + "image_path": "f6fef8bdde9557f419481033fdd3cc5b156b3ee396d4175f767dadd913338958.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 336, + 527, + 407 + ], + "lines": [ + { + "bbox": [ + 68, + 336, + 527, + 407 + ], + "spans": [ + { + "bbox": [ + 68, + 336, + 527, + 407 + ], + "type": "text", + "content": "Figure 2. The pipeline of ShapeMatcher. Given a target point cloud obtained from a single-view scan and a pre-established database (A), ShapeMatcher generates the fine-grained reconstruction result using the joint 4 modules including Canonicalization (B), Segmentation (C), Retrieval (D) and Deformation (E), where the first three contains the partial branch for target processing and the full branch for source processing. Specifically, the target and source inputs are first canonicalized into the same affine-invariant space (B). Then, the semantic-consistent region segmentation is yielded from the affine-invariant features (C). The segmented regions are fed to the region-weight retrieval module (C) and the part center guided neural cage deformation module (E) for occlusion-robust R&D process. During training, the partial-full consistency losses (F) are enforced for the two branches." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 268, + 99, + 415, + 325 + ], + "blocks": [ + { + "bbox": [ + 268, + 99, + 415, + 325 + ], + "lines": [ + { + "bbox": [ + 268, + 99, + 415, + 325 + ], + "spans": [ + { + "bbox": [ + 268, + 99, + 415, + 325 + ], + "type": "image", + "image_path": "849bcfa63bab6cd20b47261ff51c697d1e8394d7a6a4f4d131345f8c6d5e5365.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 417, + 99, + 512, + 325 + ], + "blocks": [ + { + "bbox": [ + 417, + 99, + 512, + 325 + ], + "lines": [ + { + "bbox": [ + 417, + 99, + 512, + 325 + ], + "spans": [ + { + "bbox": [ + 417, + 99, + 512, + 325 + ], + "type": "image", + "image_path": "48f3f357bba1587a2666c132a58a1a3314a714b1f74abdc338be666efec4650a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "spans": [ + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "inline_equation", + "content": "S_{tgt}^{c}" + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "inline_equation", + "content": "F_{seg} \\in \\mathbb{R}^{N \\times M}" + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "inline_equation", + "content": "C_i \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "content": " represents the probability of point " + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "content": " belonging to each part center. Then we use another 4-layer feature aggregator " + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "inline_equation", + "content": "\\Theta_f" + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "content": " (Fig. 2 (D)) to extract the retrieval tokens " + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 68, + 412, + 290, + 458 + ], + "type": "text", + "content": " of all parts as follows," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 463, + 289, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 289, + 477 + ], + "type": "interline_equation", + "content": "F _ {c l s} = F _ {s e g} ^ {\\top} * \\Theta_ {f} (F _ {t g t}), \\tag {4}", + "image_path": "eb825a92595d06d4dd31ea89e29752acaf740a8c39f522c8bae20a92a7b77fa0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 491, + 288, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 288, + 520 + ], + "type": "interline_equation", + "content": "Q _ {t g t} = F _ {c l s} / \\left(\\sum_ {n = 1} ^ {N} F _ {s e g} ^ {(n)}\\right). \\tag {5}", + "image_path": "2dec1a4f26efb244fd74cd01ffd3e4d093d05d2e0f51511ab7fa3773cc6866ee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "spans": [ + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "Q_{tgt} \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " contains the " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": "-dimensional retrieval tokens for all the " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " regions. Here we employ a soft assignment strategy where each point inside " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " is estimated " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " values describing the probabilities belonging to each of the " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " parts. Therefore, we first aggregate features " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "F_{cls}" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " on all points belonging to each part and then normalize " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "F_{cls}" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " using the sum of probabilities of points in each part, as in Eq. 4 and Eq. 5. Following a similar strategy, we can obtain " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "Q_{src}" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " for each source model in the pre-curated database. We just need to compare " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "Q_{tgt}" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " with the retrieval tokens " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "Q_{src}" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " of all source shapes using the weighted " + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 69, + 524, + 290, + 646 + ], + "type": "text", + "content": " distance," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 651, + 289, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 651, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 121, + 651, + 289, + 668 + ], + "type": "interline_equation", + "content": "D i s = \\omega \\sum \\mathcal {L} _ {1} \\left(Q _ {t g t} - Q _ {s r c}\\right) \\tag {6}", + "image_path": "6376c4ce44fa512fee572c46d4d5df7ae70d4c422e7888526c77cbac1ee052ea.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 674, + 290, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 290, + 687 + ], + "type": "text", + "content": "where vector " + }, + { + "bbox": [ + 69, + 674, + 290, + 687 + ], + "type": "inline_equation", + "content": "\\omega \\in \\mathbb{R}^{1\\times M}" + }, + { + "bbox": [ + 69, + 674, + 290, + 687 + ], + "type": "text", + "content": " stores the ratio of point number" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 413, + 526, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 526, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 526, + 468 + ], + "type": "text", + "content": "of each part with respect to the total point number " + }, + { + "bbox": [ + 304, + 413, + 526, + 468 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 413, + 526, + 468 + ], + "type": "text", + "content": ". Intuitively, parts with smaller point numbers contribute less in calculating the distance score, which reduces the influence of noise and occlusion. The source shape " + }, + { + "bbox": [ + 304, + 413, + 526, + 468 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 304, + 413, + 526, + 468 + ], + "type": "text", + "content": " with the smallest distance score is identified as the best retrieval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 474, + 382, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 382, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 382, + 485 + ], + "type": "text", + "content": "3.4. Deformation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": "The Deformation module aims to deform the retrieved shape " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " to tightly match the target shape " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": ". We utilize the neural cage scaffolding strategy as in [26, 64]. First, the neural cages " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "C_{src}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " is pre-calculated. We utilize the part centers " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "(K_{tgt}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "K_{src})" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " to control the vertex offsets " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "C_{src2tgt}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " of the neural cage " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "C_{src}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " to match " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": ". In particular, we employ a neural network " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "\\Theta_I" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " to predict an influence vector " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "I\\in \\mathbb{R}^{N_c\\times M}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " for each point concerning all cage vertices by " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "I = \\Theta_I(\\text{concat}(F_{tgt},F_{src}))" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " denotes the number of vertices used in " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "C_{src}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "C_{src2tgt}" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " is computed through the influence vectors " + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 491, + 526, + 623 + ], + "type": "text", + "content": " and the differences between region centers (Fig. 2 (E)):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 628, + 526, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 628, + 526, + 658 + ], + "spans": [ + { + "bbox": [ + 339, + 628, + 526, + 658 + ], + "type": "interline_equation", + "content": "C _ {s r c 2 t g t} = C _ {s r c} + \\sum_ {i = 1} ^ {M} I _ {i} \\left(K _ {s r c} ^ {(i)} - K _ {t g t} ^ {(i)}\\right), \\tag {7}", + "image_path": "a6e18e633344223c490d66ae93ccfa39e5e3625738c23e0513e0e86f7048aadd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 664, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 664, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 664, + 526, + 687 + ], + "type": "text", + "content": "Finally, we employ a sparse cage scaffolding strategy [26, 64] to achieve the deformation field of " + }, + { + "bbox": [ + 305, + 664, + 526, + 687 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 305, + 664, + 526, + 687 + ], + "type": "text", + "content": ". The deformed" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "21020" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 100, + 258, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 100, + 258, + 111 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 258, + 111 + ], + "type": "text", + "content": "shape " + }, + { + "bbox": [ + 69, + 100, + 258, + 111 + ], + "type": "inline_equation", + "content": "S_{src2tgt}" + }, + { + "bbox": [ + 69, + 100, + 258, + 111 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 69, + 100, + 258, + 111 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 69, + 100, + 258, + 111 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 108, + 118, + 288, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 118, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 108, + 118, + 288, + 130 + ], + "type": "interline_equation", + "content": "S _ {s r c 2 t g t} = S _ {s r c} + \\Psi \\left(C _ {s r c}, C _ {s r c 2 t g t}\\right), \\tag {8}", + "image_path": "e8bff375f4c2715e59d790d8176da7b263319705c823e4e6097a8f6e0c3cb68d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "text", + "content": " computes the displacement of each point in " + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "text", + "content": " by evaluating the differences between " + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "inline_equation", + "content": "C_{src}" + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "inline_equation", + "content": "C_{src2tgt}" + }, + { + "bbox": [ + 68, + 136, + 288, + 169 + ], + "type": "text", + "content": ", thereby achieving deformation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 175, + 222, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 175, + 222, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 175, + 222, + 186 + ], + "type": "text", + "content": "3.5. ShapeMatcher: Joint Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 191, + 288, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 191, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 68, + 191, + 288, + 267 + ], + "type": "text", + "content": "Our core insight in ShapeMatcher is that the four highly-associated tasks: Canonicalization, Segmentation, Retrieval and Deformation can be trained simultaneously and supervise each other via introducing cross-task consistency terms. We mainly introduce two types of losses here, i.e. partial-full consistency losses and task-oriented loss. For more details, please refer to the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "text", + "content": "Task-Oriented Loss. In the Canonicalization, we mainly use Chamfer Distance to constrain the canonicalized " + }, + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "inline_equation", + "content": "S_{tgt}^{c}" + }, + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\hat{S}_{tgt}^{c}" + }, + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "text", + "content": " predicted in the affine-invariant space by the supplementary reconstruction branch, to enforce the affine-invariety of " + }, + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "inline_equation", + "content": "F_{tgt}^{*}" + }, + { + "bbox": [ + 68, + 268, + 288, + 326 + ], + "type": "text", + "content": " in Sec. 3.1, by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 333, + 288, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 333, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 96, + 333, + 288, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c a n}} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(S _ {t g t} ^ {c}, \\hat {S} _ {t g t} ^ {c}\\right) + \\operatorname {o r t h} \\left(R _ {t g t}\\right), \\tag {9}", + "image_path": "3f44466d08996e9ad0ab065f12f20e8d020b1aaa3e2e2d7af202a38b1fa72756.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 352, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 352, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 68, + 352, + 288, + 373 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 68, + 352, + 288, + 373 + ], + "type": "inline_equation", + "content": "\\text{orth}(R_{tgt})" + }, + { + "bbox": [ + 68, + 352, + 288, + 373 + ], + "type": "text", + "content": " serving the purpose of enforcing the orthogonality of matrices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "text", + "content": "In the Segmentation, to keep consistency between the part segmentation and the predicted part center, we jointly train " + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "inline_equation", + "content": "\\Theta_{l}" + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "text", + "content": " with the following loss, which enforce that each predicted part center " + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "inline_equation", + "content": "K_{tgt}^{i}" + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "text", + "content": " approximately lies in the center of all points belonging to the part " + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 68, + 374, + 288, + 429 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 434, + 288, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 434, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 100, + 434, + 288, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e g} = \\sum_ {m = 1} ^ {M} \\| K _ {t g t} ^ {(m)} - \\left(F _ {s e g} ^ {\\top} * S _ {t g t} ^ {c}\\right) ^ {(m)} \\| _ {2} \\tag {10}", + "image_path": "3c6b004c60aea2005eb5437be30abdac599f4fceec3d1bb936d0001abf29c3ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": "To train the Retrieval and Deformation simultaneously, for an input target " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": ", we randomly select a source model " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " from the database for training. Specifically, to eliminate the influence of occlusion, we do not directly use the global Chamfer Distance of " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{src}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " as ground truth. Instead, we employ a regional supervision strategy, ensuring that occluded areas do not contribute to the training of retrieval network. Taking the " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": "-th region as an example, " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{tgt}^{i}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " represents all points in " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " that belong to the " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": "-th region. We calculate the average of the nearest distances " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " from each point in " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{tgt}^{i}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " to the deformed shape " + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "inline_equation", + "content": "S_{src}^{dfm}" + }, + { + "bbox": [ + 68, + 469, + 288, + 601 + ], + "type": "text", + "content": " to enforce the learning of the regional retrieval tokens by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 84, + 607, + 288, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 607, + 288, + 637 + ], + "spans": [ + { + "bbox": [ + 84, + 607, + 288, + 637 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e t r i e v a l}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} M S E \\left(Q _ {t g t} ^ {(i)} - Q _ {s r c} ^ {(i)}, D _ {i}\\right). \\tag {11}", + "image_path": "36d61e41f9a2390a723c53aa75e04e2b9690592734a892100dc8531d53bcef2f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 642, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 642, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 68, + 642, + 288, + 673 + ], + "type": "text", + "content": "The deformation loss is achieved by directly constraining the Chamfer Distance between " + }, + { + "bbox": [ + 68, + 642, + 288, + 673 + ], + "type": "inline_equation", + "content": "S_{tgt}" + }, + { + "bbox": [ + 68, + 642, + 288, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 642, + 288, + 673 + ], + "type": "inline_equation", + "content": "S_{src}^{dfm}" + }, + { + "bbox": [ + 68, + 642, + 288, + 673 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 102, + 673, + 288, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 673, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 102, + 673, + 288, + 688 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e f o r m}} = \\left. d i s _ {\\text {c h a m}} \\left(S _ {t g t}, S _ {s r c} ^ {d f m}\\right) + \\| I \\| _ {2} \\right. \\tag {12}", + "image_path": "78ea0be63e1920e63a3730d5d7f802caa0a95706c612de384a1dfc1b6bdb0006.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 100, + 460, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 100, + 460, + 110 + ], + "spans": [ + { + "bbox": [ + 306, + 100, + 460, + 110 + ], + "type": "text", + "content": "where we regularize " + }, + { + "bbox": [ + 306, + 100, + 460, + 110 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 306, + 100, + 460, + 110 + ], + "type": "text", + "content": " using the L2 norm." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 110, + 525, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 525, + 176 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 525, + 176 + ], + "type": "text", + "content": "Partial-Full Consistency Losses. In the first two modules: Canonicalization, Segmentation, the full branch serves as a guidance to enhance the learning of the partial branch. Therefore, in each module, we can enforce corresponding consistency terms between the outputs of the two parallel branches (Fig. 2 (F))." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "text", + "content": "During the consistency training process, for randomly selected full input " + }, + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "inline_equation", + "content": "S_{full}" + }, + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "text", + "content": ", we generate a mask " + }, + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "inline_equation", + "content": "U_{f2p} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "text", + "content": " to crop it to simulate the situation of a partial input " + }, + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "inline_equation", + "content": "S_{partial} = S_{full}U_{f2p}" + }, + { + "bbox": [ + 304, + 177, + 525, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "text", + "content": "In the Canonicalization module, to enforce the consistency in the affine-invariant space between the two branches, we apply the same transformation " + }, + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "inline_equation", + "content": "U_{f2p}" + }, + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "inline_equation", + "content": "S_{full}^{c}" + }, + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "text", + "content": " in the affine-invariant space as before and then use the chamfer distance to constrain its distance to " + }, + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "inline_equation", + "content": "S_{partial}^{c}" + }, + { + "bbox": [ + 304, + 220, + 525, + 277 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 282, + 525, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 282, + 525, + 296 + ], + "spans": [ + { + "bbox": [ + 342, + 282, + 525, + 296 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c c a n} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(S _ {\\text {p a r t i a l}} ^ {c}, S _ {\\text {f u l l}} ^ {c} U _ {f 2 p}\\right) \\tag {13}", + "image_path": "e33de267bc8251c611d38746450a4b26a8d714000cb6281e5a9257d9d2da5c9d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 301, + 525, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 525, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 525, + 345 + ], + "type": "text", + "content": "Similarly, in the Segmentation module, for the consistency constraint of the part center prediction network " + }, + { + "bbox": [ + 304, + 301, + 525, + 345 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 304, + 301, + 525, + 345 + ], + "type": "text", + "content": ", we directly use the Chamfer Distance to constrain the region centers detected by the two branches:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 348, + 352, + 525, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 352, + 525, + 365 + ], + "spans": [ + { + "bbox": [ + 348, + 352, + 525, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c c e n} = \\operatorname {d i s} _ {\\text {c h a m}} \\left(K _ {\\text {p a r t i a l}}, K _ {\\text {f u l l}}\\right). \\tag {14}", + "image_path": "af85bb6cfc7494018bb91d3793e1a9c3b769c8aae2a9677f9ee64bacaf68cd2f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "text", + "content": "In the segmentation network " + }, + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "inline_equation", + "content": "\\Theta_{l}" + }, + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "text", + "content": ", we mask the segmentation results of the full branch " + }, + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "inline_equation", + "content": "F_{seg}^{(full)}" + }, + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "text", + "content": " and compare them with the results of the partial branch " + }, + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "inline_equation", + "content": "F_{seg}^{(partial)}" + }, + { + "bbox": [ + 304, + 372, + 525, + 408 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 327, + 413, + 525, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 413, + 525, + 429 + ], + "spans": [ + { + "bbox": [ + 327, + 413, + 525, + 429 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c s e g} = d i s _ {c h a m} \\left(F _ {s e g} ^ {(f u l l)} U _ {f 2 p}, F _ {s e g} ^ {(p a r t i a l)}\\right). \\tag {15}", + "image_path": "f2c4153859087d49d280df641c5407c9305e837f62a5a86600e5771e13b8e47a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": "Joint Training. Generally, the joint training of Shape-Matcher is divided into three stages. First, we train the full branch by " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{can}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": " for construction of Canonicalization and Segmentation ability. Second, the partial branch is introduced and trained by both the task-oriented losses for Canonicalization and Segmentation " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{can}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": " and the partial-full consistency loss terms " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccan}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccen}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cseg}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": ". Finally, after training Canonicalization and Segmentation of the both branches, " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{retrieval}}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{deform}}" + }, + { + "bbox": [ + 304, + 434, + 525, + 567 + ], + "type": "text", + "content": " are adopted for joint Retrieval and Deformation training simultaneously utilizing the both branches to handle partial target inputs and full source inputs respectively." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 574, + 379, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 379, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 379, + 586 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 593, + 411, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 593, + 411, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 593, + 411, + 605 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 609, + 525, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 525, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 525, + 662 + ], + "type": "text", + "content": "In this section, we mainly focus on R&D experiments, which better reflects the overall performance of the system. The ablations and analysis also demonstrate the effectiveness of considering joint Canonicalization and Part Segmentation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 664, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 525, + 686 + ], + "type": "text", + "content": "Datasets. We evaluate the effectiveness of our joint framework using three datasets: two synthetic datasets," + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 315, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 315, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 315, + 715 + ], + "type": "text", + "content": "21021" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 99, + 289, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 289, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 289, + 329 + ], + "type": "text", + "content": "PartNet [39] and ComplementMe [53], and one real-world dataset, Scan2CAD [3]. For datasets PartNet and ComplementMe, we follow the same database splits as in [57], separating their target inputs into training and testing sets. In our training process, we exclusively employ mesh models and do not utilize part segmentations as in [57], since the process of ShapeMatcher is fully self-supervised and does not need any additional annotations. The shapes used in PartNet and ComplementMe datasets are sourced from ShapeNet [7]. PartNet comprises 1,419 source models in the database, with 11,433 target models in the training set and 2,861 in the testing set. In ComplementMe, the numbers are 400, 11,311 and 2,825 respectively. In the synthetic cases, three categories of tables, chairs, and cabinets are evaluated on both datasets. Scan2CAD [3] is a real-world dataset developed based on ScanNet [11] with capacity of 14,225 objects. The input point cloud data on Scan2CAD is generated by reverse-projecting the depth images. In the real-world cases, we conduct training on the categories of tables, chairs, and cabinets from PartNet and directly testing on Scan2CAD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 331, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 331, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 67, + 331, + 289, + 495 + ], + "type": "text", + "content": "Baselines. Both baseline methods, Uy et al. [57] and U-RED [17], are trained using the same data partitioning strategy stated above. To ensure fairness in comparison with ShapeMatcher, we augment the training data with pose variations, keeping other hyperparameters consistent with the original paper. During testing, we evaluate scenarios where target observations with arbitrary poses are directly used as input. Additionally, we test scenarios where the inputs are transformed using an offline pose estimation method [18], simulating the two-stage route of traditional methods with pre-canonicalizing (Uy et al. [57] + PE and U-RED [17] + PE). For experiments on Scan2CAD, we directly use the baseline models trained on PartNet with the 25% occlusion setting to conduct zero-shot testing, since real-world ground-truth models are inaccessible for training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 498, + 288, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 498, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 67, + 498, + 288, + 563 + ], + "type": "text", + "content": "Evaluation Metrics. We utilize Chamfer Distance (CD) on the magnitude of " + }, + { + "bbox": [ + 67, + 498, + 288, + 563 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 67, + 498, + 288, + 563 + ], + "type": "text", + "content": " to assess both full shape scenarios and partial shape scenarios. We calculate the metrics following [57] to use the best result among the top 10 candidate objects. The final average metrics are obtained by averaging the results across all instances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "text", + "content": "Implementation Details. During training, we uniformly sample objects to obtain point clouds with " + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "inline_equation", + "content": "M = 2500" + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "text", + "content": " points to represent shapes. We directly generate partial point clouds from the corresponding full point clouds by random cropping for the partial branch inputs. We apply random pose augmentation to the input shape, specifically with random translations " + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "inline_equation", + "content": "T_{rand} \\in [-0.1, 0.1]" + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "text", + "content": " and random rotations " + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "inline_equation", + "content": "R_{rand} \\in [-1, 1]" + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "text", + "content": " on three Eulerian angles respectively. We set the initial learning rate to " + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 67, + 565, + 289, + 686 + ], + "type": "text", + "content": " and train ShapeMatcher for 200 epochs in every training stage of Sec. 3.5. Regarding the weight of the loss, in the first" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": "stage considering only the full branch, " + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{can}" + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": " are equally weighted. In the second stage introducing the partial branch, we primarily emphasize the partial-full consistency losses, assigning significant weights to " + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccan}" + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccen}" + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cseg}" + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": " with weights set as 5, 2, and 2 respectively, while keeping the remaining weights default at 1. In the final stage for joint R&D, both " + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{retrieval}" + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{deform}" + }, + { + "bbox": [ + 304, + 100, + 525, + 188 + ], + "type": "text", + "content": " are equally weighted." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 196, + 394, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 196, + 394, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 394, + 208 + ], + "type": "text", + "content": "4.2. Synthetic Cases" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 213, + 525, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 525, + 323 + ], + "type": "text", + "content": "To validate the ability of ShapeMatcher tackling the challenge of arbitrary poses and occlusions, we first use synthetic datasets to simulate this scenario. We evaluate all methods [17, 57] where object observations with arbitrary poses are directly used as input. Additionally, we also report results where inputs are transformed and canonicalized using an offline pose estimation method [18] for baseline methods (Uy et al. + PE and U-RED + PE). Moreover, we analyze inference time of ShapeMatcher against the R&D baselines in the Supplementary Material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "text", + "content": "We conduct two types of inputs for evaluation: full inputs using the PartNet and ComplementMe datasets, and partial input tests using " + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "text", + "content": " occlusion rates on the PartNet dataset. The results of the full input tests are detailed in Table 1. In PartNet, our ShapeMatcher significantly outperforms the current leading competitors. For the Chamfer Distance on three categories, ShapeMatcher measures at 0.197, 0.150, and 0.519, maintaining the leading position. Even when the processed PE results are used as input, the baselines' results still fall short of ShapeMatcher. This demonstrates the effectiveness of adopting the affine-invariant features in the joint Canonicalization step. Results from ComplementMe supports the same conclusion, where ShapeMatcher reports significantly better results compared to the baseline methods. ShapeMatcher surpasses the top-performing Uy et al. + PE by " + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 304, + 324, + 525, + 532 + ], + "type": "text", + "content": ". Such superior results yielded by ShapeMatcher demonstrates that the joint consideration of all four steps improves the matching accuracy a lot." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "text", + "content": "For evaluation on partial inputs, we control the occlusion rates of partial point clouds by controlling the position of the cropping planes onto the full point clouds. The evaluation on partial inputs are presented in Table 2. Concretely, ShapeMatcher outperforms the current top method handling partial inputs U-RED by 5.018, 5.666 and 7.241 at the occlusion rate of " + }, + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 532, + 525, + 674 + ], + "type": "text", + "content": " respectively. As the occlusion rate increases, the superiority of the ShapeMatcher method grows. Considering PE adopting, the same trend is exhibited. ShapeMatcher surpasses the U-RED + PE by 0.525, 0.949, 2.434 under three occlusion rates. It demonstrates that the proposed region-weighted retrieval brings strong robustness of our method against occlusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 675, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 675, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 317, + 675, + 525, + 686 + ], + "type": "text", + "content": "As shown in Fig. 3 and 4, our results show more geo" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "21022" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 98, + 286, + 247 + ], + "blocks": [ + { + "bbox": [ + 72, + 98, + 286, + 247 + ], + "lines": [ + { + "bbox": [ + 72, + 98, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 72, + 98, + 286, + 247 + ], + "type": "table", + "html": "
PartNet [39]
MethodChairTableCabinetAverage
Uy et al. [57]4.2696.3024.1185.271
Uy et al. [57] + PE1.5073.0061.0702.219
U-RED [17]5.3314.9809.1415.463
U-RED [17] + PE1.0250.3591.4230.725
Ours0.1970.1500.5190.200
ComplementMe [53]
MethodChairTableCabinetAverage
Uy et al. [57]4.0185.480-4.825
Uy et al. [57] + PE1.4392.454-1.999
U-RED [17]8.5755.800-7.044
U-RED [17] + PE4.9540.847-2.688
Ours0.2530.328-0.294
", + "image_path": "8ea4bf55a8d4fee32a5a5e3b98b7e6e4aa4784108df2c440eee28b72ccf1bee5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 285, + 286, + 426 + ], + "blocks": [ + { + "bbox": [ + 69, + 255, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 69, + 255, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 69, + 255, + 288, + 275 + ], + "type": "text", + "content": "Table 1. The Chamfer Distance metrics for joint R&D results on full shapes under arbitrary poses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 285, + 286, + 426 + ], + "lines": [ + { + "bbox": [ + 72, + 285, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 72, + 285, + 286, + 426 + ], + "type": "table", + "html": "
OcclusionMethodChairTableCabinetAverage
10%Uy et al. [57]4.3726.3954.1795.365
Uy et al. + PE [57]1.5232.9821.1332.219
U-RED [17]6.0255.3755.2695.640
U-RED [17] + PE1.2071.0121.6691.147
Ours0.6760.4811.2120.622
25%Uy et al. [57]4.6546.9274.7505.795
Uy et al. + PE [57]1.8033.1951.6072.481
U-RED [17]5.1967.2158.1646.442
U-RED [17] + PE1.6841.3872.7951.625
Ours0.8780.6431.0710.776
50%Uy et al. [57]6.0709.3227.9297.841
Uy et al. + PE [57]3.3145.0304.5844.272
U-RED [17]8.6968.3877.6138.455
U-RED [17] + PE4.7222.0157.9033.628
Ours1.1971.0791.8721.194
", + "image_path": "a1d0415bf3cb7cea2cb8a5e6c19063b8ae65efd82195d0fb3e8ccfb0300ae069.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 434, + 288, + 455 + ], + "lines": [ + { + "bbox": [ + 69, + 434, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 434, + 288, + 455 + ], + "type": "text", + "content": "Table 2. The Chamfer Distance metrics for joint R&D results on partial shapes under arbitrary poses of PartNet dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 475, + 288, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 475, + 288, + 561 + ], + "spans": [ + { + "bbox": [ + 68, + 475, + 288, + 561 + ], + "type": "text", + "content": "metric resemblance to the targets compared to other methods. This is attributed to the suitable joint consideration of the four highly-associated processes, which accurately decouples the input poses, mapping them to a consistent space for accurate " + }, + { + "bbox": [ + 68, + 475, + 288, + 561 + ], + "type": "inline_equation", + "content": "R\\& D" + }, + { + "bbox": [ + 68, + 475, + 288, + 561 + ], + "type": "text", + "content": ". The region-weighted retrieval we employ explicitly eliminates the influence of occluded areas, allowing for a more precise matching with the source model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 570, + 164, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 570, + 164, + 581 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 164, + 581 + ], + "type": "text", + "content": "4.3. Real-world Cases" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "text", + "content": "We test the effectiveness of ShapeMatcher on real-world datasets. In such case, ShapeMatcher is trained on the synthetic PartNet with " + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "text", + "content": " occlusion and directly tested on the partial scans of the real-world dataset Scan2CAD without manual pose adjustments. Table 3 displays our results, where our method significantly outperforms existing competitors. Particularly, compared to the U-RED, in three categories, the reported Chamfer Distance are reduced by " + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 68, + 587, + 288, + 686 + ], + "type": "text", + "content": " respectively. In comparison to Uy et al., the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 98, + 525, + 232 + ], + "blocks": [ + { + "bbox": [ + 307, + 98, + 525, + 232 + ], + "lines": [ + { + "bbox": [ + 307, + 98, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 98, + 525, + 232 + ], + "type": "image", + "image_path": "b9d19a29390a066187b4ac1a2dca0d7d5cd538e53dc1ab1f2296fc2976ac0159.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 241, + 525, + 260 + ], + "lines": [ + { + "bbox": [ + 306, + 241, + 525, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 525, + 260 + ], + "type": "text", + "content": "Figure 3. Qualitative R&D results with full target inputs on Part-Net." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 271, + 523, + 426 + ], + "blocks": [ + { + "bbox": [ + 307, + 271, + 523, + 426 + ], + "lines": [ + { + "bbox": [ + 307, + 271, + 523, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 523, + 426 + ], + "type": "image", + "image_path": "e8280cc161639c2877722806059ab8e09ffdc639c2697c9bd5b740ee01ba81e4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 434, + 525, + 453 + ], + "lines": [ + { + "bbox": [ + 306, + 434, + 525, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 525, + 453 + ], + "type": "text", + "content": "Figure 4. Qualitative R&D results with partial target inputs on the occlusion rate of " + }, + { + "bbox": [ + 306, + 434, + 525, + 453 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 306, + 434, + 525, + 453 + ], + "type": "text", + "content": " on PartNet." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "text", + "content": "reported Chamfer Distance are reduced by " + }, + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 304, + 475, + 525, + 583 + ], + "type": "text", + "content": ". Even compared with the use of PE on the two baseline methods, there remains a significant leap forward of Shape-Matcher. This zero-shot experiments on real-world further prove that the procedure of joint Canonicalization, Segmentation, Retrieval and Deformation possesses a strong domain adaptation ability, showcasing the great potential of ShapeMatcher in real-world applications. The visualizations on Scan2CAD are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 592, + 397, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 397, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 397, + 602 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 609, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 525, + 686 + ], + "type": "text", + "content": "We conduct ablation experiments on PartNet, mainly on two aspects. First, in the Canonicalization, we investigate the importance of disentangling different pose intrinsics in Table 4, and demonstrate the effectiveness of joint considering Canonicalization. Second, in Table 5, we ablate the region-weighted Retrieval and the part center guided Deformation. Moreover, the robustness against occlusion is" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "21023" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 98, + 286, + 163 + ], + "blocks": [ + { + "bbox": [ + 72, + 98, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 72, + 98, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 72, + 98, + 286, + 163 + ], + "type": "table", + "html": "
MethodChairTableCabinetAverage
Uy et al. [57]4.8867.6058.3356.181
Uy et al. [57] + PE3.3626.6577.2614.905
U-RED [17]5.4905.13110.0915.945
U-RED [17] + PE2.8933.1645.9573.354
Ours0.4230.1860.6540.375
", + "image_path": "6e7cf98b08fcd1cb144ea04c18f6751fbfc81583521edfdb7526e9014cd6f6d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 213, + 145, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 213, + 145, + 223 + ], + "spans": [ + { + "bbox": [ + 69, + 213, + 145, + 223 + ], + "type": "text", + "content": "analyzed in Table 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "text", + "content": "Canonicalization Capability. To study the impact of different intrinsics of object poses in the Canonicalization process, we conduct ablations on decoupling of translations, rotations and scales. The results are presented in Table 4. Specifically, in row (1), we make no adjustments to the input poses. Thanks to the regional-level " + }, + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "inline_equation", + "content": "R\\& D" + }, + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "text", + "content": " process, it shows decent performance. However, there is still a noticeable gap compared to row (4), indicating the significance of our proposed joint Canonicalization. In row (2), we solely decouple the input translations, resulting in a decrease of " + }, + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "text", + "content": " in reported metrics. Moving to row (3), upon this foundation, we add decoupling for rotation, leading to a substantial decrease in reported Chamfer Distance, averaging at 0.244. In row (4), we introduce scale decoupling, resulting in another decrease in the reported metrics. It is evident that accurate decoupling of rotation is a crucial aspect of the success of the Canonicalization process. Moreover, it demonstrates that to integrate Canonicalization and " + }, + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "inline_equation", + "content": "R\\& D" + }, + { + "bbox": [ + 68, + 224, + 290, + 432 + ], + "type": "text", + "content": " process is indispensable for the ShapeMatcher process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 434, + 290, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 434, + 290, + 662 + ], + "spans": [ + { + "bbox": [ + 68, + 434, + 290, + 662 + ], + "type": "text", + "content": "Deformation and Retrieval Ability. To validate the effectiveness of our proposed region-weighted Retrieval and the part center guided neural cage Deformation, we conduct an ablation study on the PartNet dataset with the " + }, + { + "bbox": [ + 68, + 434, + 290, + 662 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 68, + 434, + 290, + 662 + ], + "type": "text", + "content": " occlusion rate. The results are presented in Table 5. In row (1), we conduct experiments using global retrieval and global deformation. This means we directly use an MLP network to extract overall point cloud features as the retrieval vector [57]. In the deformation network, similarly, we directly use an MLP network to generate neural cage offsets for deformation [26, 64]. Due to the lack of extraction of local information, the reported Chamfer Distance is more than twice of row (4). In row (2), we employ the global retrieval and the part center guided neural cage deformation. This improvement allows much more tightly-matched deformation by the retrieved source model, resulting in a " + }, + { + "bbox": [ + 68, + 434, + 290, + 662 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 68, + 434, + 290, + 662 + ], + "type": "text", + "content": " decrease in reported metrics. In row (3), we conduct experiments using the regional retrieval and the global deformation. The proposed regional-weighted retrieval handles occluded objects, reducing the impact of occluded parts and resulting in a substantial decrease in Chamfer Distance, down to 0.973." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 664, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 664, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 664, + 289, + 687 + ], + "type": "text", + "content": "Occlusion Robustness. We test shapes at different occlusion levels by altering the occlusion ratio in the input." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 98, + 524, + 147 + ], + "blocks": [ + { + "bbox": [ + 69, + 170, + 288, + 191 + ], + "lines": [ + { + "bbox": [ + 69, + 170, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 69, + 170, + 288, + 191 + ], + "type": "text", + "content": "Table 3. The Chamfer Distance metrics for joint R&D results in real-world Scan2CAD [3]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 98, + 524, + 147 + ], + "lines": [ + { + "bbox": [ + 309, + 98, + 524, + 147 + ], + "spans": [ + { + "bbox": [ + 309, + 98, + 524, + 147 + ], + "type": "table", + "html": "
Trans.Rot.Scal.ChairTableCabinetAverage
(1)0.5710.5021.2330.590
(2)✓0.4680.4421.0960.506
(3)✓✓0.2130.2000.6740.244
(4)✓✓✓0.1970.1500.5190.200
", + "image_path": "659e313f0673f9c5e78354fdbe00538640be07c1d639d0f5f49f3f49a57bc908.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 213, + 524, + 255 + ], + "blocks": [ + { + "bbox": [ + 305, + 154, + 525, + 204 + ], + "lines": [ + { + "bbox": [ + 305, + 154, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 154, + 525, + 204 + ], + "type": "text", + "content": "Table 4. Ablations on the Canonicalization process, we demonstrate the effectiveness of the joint Canonicalization by ablating different pose intrinsics. Here, Trans. denotes the decoupling of translation, Rot. represents the rotation, and Scal. signifies the scale." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 213, + 524, + 255 + ], + "lines": [ + { + "bbox": [ + 308, + 213, + 524, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 213, + 524, + 255 + ], + "type": "table", + "html": "
Gl. R.Re. R.Gl. D.Re. D.ChairTableCabinetAverage
(1)✓✓1.6721.4461.8001.570
(2)✓✓1.5391.3051.6411.431
(3)✓✓1.0420.8741.2230.973
(4)✓✓0.8780.6431.0710.776
", + "image_path": "a5bade35b59c7661e0a75017982ed6175c9951318615cab3e14d279d3358d531.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 263, + 525, + 314 + ], + "lines": [ + { + "bbox": [ + 305, + 263, + 525, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 525, + 314 + ], + "type": "text", + "content": "Table 5. Ablations of the " + }, + { + "bbox": [ + 305, + 263, + 525, + 314 + ], + "type": "inline_equation", + "content": "R\\& D" + }, + { + "bbox": [ + 305, + 263, + 525, + 314 + ], + "type": "text", + "content": " process. GL. R. denotes the global feature based retrieval [57], Re. R. represents the proposed region-weighted retrieval, GL. D. signifies direct neural cage deformation using global features [26, 64], and Re. D. denotes the adopted regional part center guided neural cage deformation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "text", + "content": "For each specific occlusion ratio, we deliberately crop a portion of the complete point cloud to simulate occlusion. We test scenarios with occlusion ratios of " + }, + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 319, + 526, + 429 + ], + "type": "text", + "content": ", and the results are presented in Table 2. Observably, as the occluded regions increased, the Chamfer Distance significantly rises for the baseline methods. Taking the U-RED + PE as an example, its reported metrics increase from 1.147 to 3.628, doubling in value, as the occluded area expands. In contrast, our method increases by less than 1-fold, which exhibits strong robustness against occlusion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 438, + 372, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 372, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 372, + 449 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 456, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 526, + 687 + ], + "type": "text", + "content": "In this paper, we present ShapeMatcher, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonize the object by extracting point-wise affine-invariant features. Then, the affine-invariant features are leveraged to predict semantically consistent part segmentation and corresponding part centers. Afterwards, the lightweight region-weighted retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that Shape-Matcher surpasses competitors by a large margin. In the future, we plan to further apply I-RED to various downstream tasks like robotic grasping. Limitations are discussed in the Supplementary Material." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "21024" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 117, + 289, + 686 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 74, + 117, + 289, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 117, + 289, + 158 + ], + "spans": [ + { + "bbox": [ + 74, + 117, + 289, + 158 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 159, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 159, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 74, + 159, + 288, + 190 + ], + "type": "text", + "content": "[2] Brandon Anderson, Truong Son Hy, and Risi Kondor. Cormorant: Covariant molecular neural networks. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 190, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 190, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 74, + 190, + 288, + 241 + ], + "type": "text", + "content": "[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 242, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 242, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 74, + 242, + 288, + 282 + ], + "type": "text", + "content": "[4] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 284, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 284, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 74, + 284, + 288, + 313 + ], + "type": "text", + "content": "[5] Frederic Bosche and Carl T Haas. Automated retrieval of 3d cad model objects in construction range images. Automation in Construction, 17(4):499-512, 2008. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 315, + 289, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 315, + 289, + 355 + ], + "spans": [ + { + "bbox": [ + 74, + 315, + 289, + 355 + ], + "type": "text", + "content": "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 357, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 357, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 74, + 357, + 288, + 407 + ], + "type": "text", + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 407, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 407, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 74, + 407, + 288, + 449 + ], + "type": "text", + "content": "[8] Yunlu Chen, Basura Fernando, Hakan Bilen, Matthias Nießner, and Efstratios Gavves. 3d equivariant graph implicit functions. In European Conference on Computer Vision, pages 485-502. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 449, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 449, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 74, + 449, + 288, + 490 + ], + "type": "text", + "content": "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 491, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 491, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 70, + 491, + 288, + 532 + ], + "type": "text", + "content": "[10] Manuel Dahnert, Angela Dai, Leonidas J Guibas, and Matthias Nießner. Joint embedding of 3d scan and cad objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8749-8758, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 533, + 288, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 533, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 70, + 533, + 288, + 583 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 584, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 584, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 70, + 584, + 288, + 634 + ], + "type": "text", + "content": "[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868-5877, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 635, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 635, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 70, + 635, + 288, + 686 + ], + "type": "text", + "content": "[13] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 100, + 525, + 686 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 100, + 525, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 100, + 525, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 100, + 525, + 150 + ], + "type": "text", + "content": "[14] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-pose: Exploiting self-occlusion for direct 6d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12396-12405, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 525, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 525, + 201 + ], + "type": "text", + "content": "[15] Yan Di, Henrique Morimitsu, Shan Gao, and Xiangyang Ji. Monocular piecewise depth estimation in dynamic scenes by exploiting superpixel relations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4363-4372, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 201, + 525, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 525, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 525, + 252 + ], + "type": "text", + "content": "[16] Yan Di, Henrique Morimitsu, Zhiqiang Lou, and Xiangyang Ji. A unified framework for piecewise semantic reconstruction in dynamic scenes via exploiting superpixel relations. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10737-10743. IEEE, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 525, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 525, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 525, + 303 + ], + "type": "text", + "content": "[17] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. arXiv preprint arXiv:2308.06383, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 303, + 525, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 525, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 525, + 362 + ], + "type": "text", + "content": "[18] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-pose: Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 363, + 525, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 525, + 413 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 525, + 413 + ], + "type": "text", + "content": "[19] Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning so (3) equivariant representations with spherical cnns. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52–68, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 414, + 525, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 414, + 525, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 414, + 525, + 463 + ], + "type": "text", + "content": "[20] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 2018 International Conference on 3D Vision (3DV), pages 672–681. IEEE, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 525, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 525, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 525, + 495 + ], + "type": "text", + "content": "[21] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9785-9795, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 495, + 525, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 495, + 525, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 495, + 525, + 534 + ], + "type": "text", + "content": "[22] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 535, + 525, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 525, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 525, + 575 + ], + "type": "text", + "content": "[23] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, volume 27, pages 1449-1457. Wiley Online Library, 2008. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 576, + 525, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 525, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 525, + 635 + ], + "type": "text", + "content": "[24] Vladislav Ishimtsev, Alexey Bokhovkin, Alexey Artemov, Savva Ignatyev, Matthias Niessner, Denis Zorin, and Evgeny Burnaev. Cad-deform: Deformable fitting of cad models to 3d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 599-628. Springer, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 636, + 525, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 525, + 686 + ], + "type": "text", + "content": "[25] Dominic Jack, Jhony K Pontes, Sridha Sridharan, Clinton Fookes, Sareh Shirazi, Frederic Maire, and Anders Eriksson. Learning free-form deformations for 3d object reconstruction. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "21025" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 288, + 685 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 88, + 100, + 288, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 100, + 288, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 100, + 288, + 121 + ], + "type": "text", + "content": "2018, Revised Selected Papers, Part II 14, pages 317-333. Springer, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 122, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 122, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 71, + 122, + 288, + 183 + ], + "type": "text", + "content": "[26] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3d keypoint discovery for shape control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12783-12792, 2021. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 185, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 185, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 70, + 185, + 288, + 225 + ], + "type": "text", + "content": "[27] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 227, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 227, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 70, + 227, + 288, + 267 + ], + "type": "text", + "content": "[28] Chiyu Jiang, Jingwei Huang, Andrea Tagliasacchi, and Leonidas J Guibas. Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems, 33:9745-9757, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 269, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 269, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 70, + 269, + 288, + 308 + ], + "type": "text", + "content": "[29] Jincen Jiang, Xuequan Lu, Lizhi Zhao, Richard Dazaley, and Meili Wang. Masked autoencoders in 3d point cloud representation learning. IEEE Transactions on Multimedia, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 311, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 311, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 70, + 311, + 288, + 352 + ], + "type": "text", + "content": "[30] Oren Katzir, Dani Lischinski, and Daniel Cohen-Or. Shape-pose disentanglement using se (3)-equivariant vector neurons. In European Conference on Computer Vision, pages 468-484. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 353, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 353, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 70, + 353, + 288, + 393 + ], + "type": "text", + "content": "[31] Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch-gordan nets: a fully fourier space spherical convolutional neural network. Advances in Neural Information Processing Systems, 31, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 395, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 395, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 70, + 395, + 288, + 456 + ], + "type": "text", + "content": "[32] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 457, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 457, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 70, + 457, + 288, + 488 + ], + "type": "text", + "content": "[33] Leon Lang and Maurice Weiler. A wigner-echart theorem for group equivariant convolution kernels. arXiv preprint arXiv:2010.10952, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 489, + 288, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 489, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 70, + 489, + 288, + 530 + ], + "type": "text", + "content": "[34] Yangyan Li, Hao Su, Charles Ruizhongtai Qi, Noa Fish, Daniel Cohen-Or, and Leonidas J Guibas. Joint embeddings of shapes and images via cnn image purification. ACM transactions on graphics (TOG), 34(6):1-12, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 531, + 288, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 531, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 70, + 531, + 288, + 582 + ], + "type": "text", + "content": "[35] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 583, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 583, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 70, + 583, + 288, + 635 + ], + "type": "text", + "content": "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 635, + 288, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 635, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 70, + 635, + 288, + 685 + ], + "type": "text", + "content": "[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 525, + 686 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 101, + 525, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 101, + 525, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 101, + 525, + 140 + ], + "type": "text", + "content": "[38] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 142, + 525, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 525, + 201 + ], + "type": "text", + "content": "[39] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 203, + 525, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 203, + 525, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 203, + 525, + 233 + ], + "type": "text", + "content": "[40] Liangliang Nan, Ke Xie, and Andrei Sharf. A search-classify approach for cluttered indoor scene understanding. ACM Transactions on Graphics (TOG), 31(6):1-10, 2012. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 235, + 525, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 525, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 525, + 294 + ], + "type": "text", + "content": "[41] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 296, + 525, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 525, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 525, + 346 + ], + "type": "text", + "content": "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 347, + 525, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 525, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 525, + 396 + ], + "type": "text", + "content": "[43] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 398, + 525, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 525, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 525, + 438 + ], + "type": "text", + "content": "[44] Yuchen Rao, Yinyu Nie, and Angela Dai. Patchcomplete: Learning multi-resolution patch priors for 3d shape completion on unseen categories. Advances in Neural Information Processing Systems, 35:34436-34450, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 439, + 525, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 525, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 525, + 489 + ], + "type": "text", + "content": "[45] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. Advances in Neural Information Processing Systems, 33:22468-22478, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 491, + 525, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 525, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 525, + 531 + ], + "type": "text", + "content": "[46] Adriana Schulz, Ariel Shamir, Ilya Baran, David IW Levin, Pitchaya Sitthi-Amorn, and Wojciech Matusik. Retrieval on parametric shape collections. ACM Transactions on Graphics (TOG), 36(1):1-14, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 533, + 525, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 525, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 525, + 563 + ], + "type": "text", + "content": "[47] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116. CiteSeer, 2007. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 563, + 525, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 525, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 525, + 613 + ], + "type": "text", + "content": "[48] Yongzhi Su, Yan Di, Guangyao Zhai, Fabian Manhardt, Jason Rambach, Benjamin Busam, Didier Stricker, and Federico Tombari. Opa-3d: Occlusion-aware pixel-wise aggregation for monocular 3d object detection. IEEE Robotics and Automation Letters, 8(3):1327–1334, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 615, + 525, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 615, + 525, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 615, + 525, + 665 + ], + "type": "text", + "content": "[49] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 666, + 525, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 525, + 686 + ], + "type": "text", + "content": "[50] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 706, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 706, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 706, + 317, + 714 + ], + "type": "text", + "content": "21026" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 288, + 686 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 88, + 100, + 288, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 100, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 88, + 100, + 288, + 131 + ], + "type": "text", + "content": "for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11286-11295, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 133, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 133, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 70, + 133, + 288, + 172 + ], + "type": "text", + "content": "[51] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 1, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 173, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 173, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 70, + 173, + 288, + 223 + ], + "type": "text", + "content": "[52] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua Siegel, and Sanjay Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 61–70, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 224, + 288, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 224, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 70, + 224, + 288, + 264 + ], + "type": "text", + "content": "[53] Minhyuk Sung, Hao Su, Vladimir G Kim, Siddhartha Chaudhuri, and Leonidas Guibas. Complementme: Weakly-supervised component suggestions for 3d modeling. ACM Transactions on Graphics (TOG), 36(6):1-12, 2017. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 266, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 266, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 70, + 266, + 288, + 316 + ], + "type": "text", + "content": "[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3405-3414, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 318, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 318, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 70, + 318, + 288, + 367 + ], + "type": "text", + "content": "[55] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 369, + 288, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 369, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 70, + 369, + 288, + 428 + ], + "type": "text", + "content": "[56] Mikaela Angelina Uy, Jingwei Huang, Minhyuk Sung, Tolga Birdal, and Leonidas Guibas. Deformation-aware 3d model embedding and retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 397-413. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 430, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 430, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 70, + 430, + 288, + 490 + ], + "type": "text", + "content": "[57] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 491, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 491, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 70, + 491, + 288, + 532 + ], + "type": "text", + "content": "[58] Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. 3dn: 3d deformation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1038-1046, 2019. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 533, + 288, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 533, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 70, + 533, + 288, + 582 + ], + "type": "text", + "content": "[59] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco S Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. Advances in Neural Information Processing Systems, 31, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 584, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 584, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 70, + 584, + 288, + 635 + ], + "type": "text", + "content": "[60] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 636, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 636, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 70, + 636, + 288, + 686 + ], + "type": "text", + "content": "[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 525, + 685 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 101, + 525, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 101, + 525, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 101, + 525, + 150 + ], + "type": "text", + "content": "[62] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 152, + 525, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 525, + 201 + ], + "type": "text", + "content": "[63] Shuo Yang, Min Xu, Haozhe Xie, Stuart Perry, and Jiahao Xia. Single-view 3d object reconstruction from shape priors in memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3152-3161, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 203, + 525, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 203, + 525, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 203, + 525, + 254 + ], + "type": "text", + "content": "[64] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 255, + 525, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 525, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 525, + 304 + ], + "type": "text", + "content": "[65] Michela Zaccaria, Fabian Manhardt, Yan Di, Federico Tombari, Jacopo Aleotti, and Mikhail Giorgini. Self-supervised category-level 6d object pose estimation with optical flow consistency. IEEE Robotics and Automation Letters, 8(5):2510-2517, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 306, + 525, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 306, + 525, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 525, + 357 + ], + "type": "text", + "content": "[66] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. arXiv preprint arXiv:2309.12188, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 358, + 525, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 525, + 418 + ], + "type": "text", + "content": "[67] Guangyao Zhai, Dianye Huang, Shun-Cheng Wu, HyunJun Jung, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Monograspnet: 6-dof grasping with a single rgb image. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1708-1714. IEEE, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 419, + 525, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 525, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 525, + 468 + ], + "type": "text", + "content": "[68] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graphs. arXiv preprint arXiv:2305.16283, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 471, + 525, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 525, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 525, + 521 + ], + "type": "text", + "content": "[69] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 522, + 525, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 525, + 572 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 525, + 572 + ], + "type": "text", + "content": "[70] Chenyangguang Zhang, Yan Di, Ruida Zhang, Guangyao Zhai, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ddf-ho: Hand-held object reconstruction via conditional directed distance field. arXiv preprint arXiv:2308.08231, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 574, + 525, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 574, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 574, + 525, + 624 + ], + "type": "text", + "content": "[71] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European Conference on Computer Vision, pages 655–672. Springer, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 625, + 525, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 525, + 685 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 525, + 685 + ], + "type": "text", + "content": "[72] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-posed: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452-7459. IEEE, 2022. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 715 + ], + "type": "text", + "content": "21027" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 288, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 100, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 70, + 100, + 288, + 157 + ], + "type": "text", + "content": "[73] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu-jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022." + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "spans": [ + { + "bbox": [ + 294, + 705, + 317, + 714 + ], + "type": "text", + "content": "21028" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_content_list.json b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f2da75d5c836666bab6a454a8c4b240bf04ce17a --- /dev/null +++ b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_content_list.json @@ -0,0 +1,1437 @@ +[ + { + "type": "text", + "text": "ShapeWalk: Compositional Shape Editing through Language-Guided Chains", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Habib Slim KAUST", + "bbox": [ + 272, + 181, + 366, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "habib.smil@kaust.edu.sa", + "bbox": [ + 217, + 219, + 423, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohamed Elhoseiny KAUST", + "bbox": [ + 532, + 181, + 699, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mohamed.elhoseiny@kaust.edu.sa", + "bbox": [ + 478, + 219, + 746, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d771d94127f775e4a915ab24a00efa8254a1261303d26dd0eed586958a7c2c0.jpg", + "image_caption": [ + "Figure 1. Visualizing shape chains from ShapeWalk. Our dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. We illustrate the interpolation process by coloring shapes based on their proximity to the starting shape (blue), and the ending shape (orange). Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. For each shape transition, we also provide a precise edit vector $\\overline{\\theta_{ij}}$ describing the parameter changes necessary to transition from one shape to the next." + ], + "image_footnote": [], + "bbox": [ + 93, + 273, + 859, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 694, + 313, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Editing 3D shapes through natural language instructions is a challenging task that requires the comprehension of both language semantics and fine-grained geometric details. To bridge this gap, we introduce ShapeWalk, a carefully designed synthetic dataset designed to advance the field of language-guided shape editing. The dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. We synthesize edit chains by reconstructing and interpolating shapes sampled from a realistic CAD-designed 3D dataset in the parameter", + "bbox": [ + 73, + 719, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "space of the GeoCode shape program. We leverage rule-based methods and language models to generate accurate and realistic natural language prompts corresponding to each edit. To illustrate the practicality of our contribution, we train neural editor modules in the latent space of shape autoencoders, and demonstrate the ability of our dataset to enable a variety of language-guided shape edits. Finally, we introduce multi-step editing metrics to benchmark the capacity of our models to perform recursive shape edits. We hope that our work will enable further study of compositional language-guided shape editing, and finds application in 3D CAD design and interactive modeling.", + "bbox": [ + 496, + 696, + 893, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22574", + "bbox": [ + 478, + 945, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97c4c99fd451ec83127abeaac8f74e0458fe56c6de1e3219674814d51eea4fbc.jpg", + "image_caption": [ + "Figure 2. Comparing ShapeTalk with our work. We compare the ShapeTalk [3] dataset (top) with our work (bottom). For an equivalent edit instruction (in green), ShapeTalk provides pairs of shapes with many factors of variation, while we generate synthetic pairs of shapes with a single clear varying factor." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 467, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 356, + 209, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Whether in the realms of computer-aided design, virtual reality environments, or digital content creation, the process of refining and enhancing 3D visual data often involves intricate adjustments to geometric shapes, textures, and lighting. Furthermore, the necessity for precise modifications adds an additional layer of complexity, as inaccuracies can have profound implications on the final output. This labor-intensive nature of 3D data editing not only hinders workflow efficiency but also poses a barrier for individuals without specialized skills, limiting the democratization of 3D content manipulation. As such, there exists a compelling need for innovative solutions that streamline and democratize the 3D editing process, making it more accessible to a broader range of users. But how to train models with the ability to perform these complex edits?", + "bbox": [ + 75, + 383, + 468, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To that end, we introduce ShapeWalk, a dataset aiming at addressing the challenges inherent in the intricate and skill-intensive nature of 3D data editing. Our dataset consists of synthesized chains of 3D shapes, connected through edit vectors associated with precise language instructions describing the applied edits. It contains 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Our generation method is scalable and can yield an indefinite number of realistic shape chains, and extended to any 3D domain and shape program. To our knowledge, our dataset is the first to provide a large-scale collection of realistic 3D shape edit chains with precise language instructions, and matching ground-truth edited shapes. Our method yields specific edits undiluted by other factors of variation, as illustrated in Figure 2.", + "bbox": [ + 75, + 611, + 468, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the usefulness of our dataset, we train neural editor modules in the latent space of shape autoencoders [1], and demonstrate the ability of our dataset to train models capable of performing a variety of shape edits. We learn", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "latent mappings in the space of a frozen shape autoencoder, and show that our method can be applied on shapes unseen during the training of the neural editor. Taking advantage of the ground-truth shape chains provided by our dataset, we introduce multi-step editing metrics inspired from the field of trajectory prediction [4, 26] to evaluate the quality of our models.", + "bbox": [ + 496, + 90, + 890, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 500, + 198, + 823, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Chained Shape-Editing Dataset: We introduce a synthetic dataset of chained edits, with associated language instructions and ground-truth edited shapes. This is the first dataset of its kind, and is designed to facilitate the study of compositional shape editing.", + "- Neural Shape Editor: We introduce a neural editor module trained on our dataset, and demonstrate its ability to perform a variety of shape edits. We show that our method can be applied on shapes unseen during the training of the neural editor.", + "- Chained-Editing Metrics: With the introduction of ground-truth shape chains, we propose multi-step editing metrics to evaluate the ability of neural editors to perform recursively applied edits." + ], + "bbox": [ + 500, + 224, + 890, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 478, + 635, + 493 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Shape Programs. A significant body of work [9, 11, 12, 14, 19] is dedicated to exploring the use of procedural programs for 3D shape representation. Representing 3D objects as visual programs has many advantages. Representing shapes as visual primitives enhances interpretability [11, 14], as programs are by definition human-readable and thus understandable by human experts. Shape programs are more compact [9, 27] than their usual 3D shape modalities, and can be used to represent shapes more efficiently. Programs can also be composed to create shapes [11, 14], or decomposed into smaller programs. Ideally, shape programs can be edited [13] to modify the underlying geometry they represent.", + "bbox": [ + 496, + 503, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our dataset consists of a collection of synthetic shapes created utilizing a backbone mesh-generating shape program. Specifically, we build from GeoCode [17], a 3D shape synthesis technique addressing the challenge of mapping high-fidelity geometry to an editable parameter space. GeoCode introduces a procedural program enabling the generation of high-quality mesh outputs with a balanced blend of interpretability and fine control.", + "bbox": [ + 496, + 702, + 890, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-Guided 3D Shape Editing. Various works have explored the use of natural language instructions to guide 3D shape editing. Recent efforts [16, 20, 21] propose to leverage pre-trained CLIP [24] models to align", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22575", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D shapes with a given text prompt. Early efforts in that direction [16] included optimizing meshes to progressively align them with CLIP embeddings, leading to important computational overhead. CLIP-Sculptor [21], a more recent work, leverages a voxelized representation and a discrete latent space conditioned on CLIP's image-text embeddings to perform fast and fidel shape edits without shape optimization.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shape Editing Datasets. Other works leverage large text-aligned 3D shape datasets and require significant manual annotation effort. ShapeCrafter [8] generates 3D shapes incrementally from text using a neural network, evolving with additional phrases. ShapeCrafter is designed for recursive shape edition and utilizes a VQ-VAE [23] model to represent shapes as discrete codes. This method exhibits consistent shape-text alignment with gradual evolution. ChangeIt3D [3] introduces ShapeTalk, a large dataset for describing 3D shape differences. The framework facilitates language-based editing of 3D models without requiring 2D to 3D conversion methods, and learns a shape editing model by learning contrasts between sampled shape pairs.", + "bbox": [ + 75, + 231, + 472, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ShapeTalk [3] is the most similar work to our proposed dataset, and is the only publicly available dataset specifically designed for language-guided 3D shape-to-shape editing. ShapeTalk is a remarkable contribution in the field of language-guided shape editing, and is one of the first works to leverage large-scale 3D shape datasets to facilitate the study of language-guided shape editing. However, this work has some limitations that we attempt to address in our work. Collecting a dataset of shape differences is a challenging task requiring considerable manual annotation effort, and gathering a large number of 3D shapes may not be feasible for all domains. Furthermore, ShapeTalk is composed of edit contexts (i.e. shape pairs and edit instructions) which are hard to separate into fine-grained, composable edits. Most of the time, singular shape edits do not have an exact ground-truth in the form of a source and target shape pair. We illustrate this important distinction in Figure 2. The availability of singular shape edits with an associated ground-truth would ease benchmarking the quality of shape editing methods, and facilitate the training process of editing models. ShapeTalk also does not provide a mechanism to generate edit chains, which are necessary to study compositional shape editing. Finally, ShapeTalk does not easily enable the study of shape editing in a compositional manner. In contrast, our dataset is 1) synthesized by augmenting a small set of diverse shapes and can easily be scaled up, 2) composed of fine-grained edits and coarse-grained edits with exact ground-truths, and 3) separated into edit chains, which are designed to facilitate the study of compositional shape editing.", + "bbox": [ + 75, + 446, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. ShapeWalk", + "text_level": 1, + "bbox": [ + 500, + 95, + 620, + 113 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our dataset contains 158K shapes split into a random and realistic set. With each edge connecting two consecutive shapes, we also produce a text instruction generated using the parameter changes necessary to transition from one shape to the next. We detail the generation method of our dataset here, and summarize the process in Figure 3.", + "bbox": [ + 496, + 122, + 893, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 224, + 596, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition. Our dataset can be defined as a collection of directed graph paths (dipaths), denoted as $\\mathcal{P}^{(k)} = (S^{(k)},E^{(k)},f^{(k)})$ . For each shape chain $\\mathcal{P}^{(k)}$ of length $l$ , we denote:", + "bbox": [ + 496, + 248, + 893, + 306 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $S^{(k)} = \\{s_{\\theta_1}^{(k)}, \\ldots, s_{\\theta_N}^{(k)}\\}$ , the set of distinct shape nodes composing the chain. Each shape is defined by a set of parameters $\\theta_i \\subseteq \\Theta$ , where $\\Theta$ is the linear parameter space of our shape program.", + "- $E^{(k)} \\subseteq \\{(s_{\\theta_{\\mathbf{i}}}^{(k)}, s_{\\theta_{\\mathbf{j}}}^{(k)}) \\mid s_{\\theta_{\\mathbf{i}}}^{(k)}, s_{\\theta_{\\mathbf{j}}}^{(k)} \\in V^{(k)}, j = i + 1\\}$ , the set of edges linking each consecutive pair of shapes.", + "- $f^{(k)}: E^{(k)} \\mapsto \\{(\\overline{\\theta_{ij}}, \\mathbf{p_{ij}}) \\mid \\overline{\\theta_{ij}} \\subseteq \\Theta, \\mathbf{p_{ij}} \\in \\Sigma \\}$ , a function mapping each edge to a vector $\\overline{\\theta_{ij}} \\subseteq \\Theta$ and a set of text instructions $\\mathbf{p_{ij}}$ . $\\overline{\\theta_{ij}}$ defines the parameter changes, or edits necessary, to go from shape $i$ to $j$ . $\\mathbf{p_{ij}}$ is a set of text instructions describing this edit in natural language." + ], + "bbox": [ + 500, + 311, + 893, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generation. To generate our dataset, we start by reconstructing a set $S \\subseteq S$ of realistic 3D CAD shapes into the space of shapes covered by our shape program $S_{\\Theta} \\subset S$ . To that end, we utilize shapes from the 3DCOMPAT++ [15, 22] dataset, a realistic, industry-based 3D CAD dataset. We employ this dataset to avoid overlap with ShapeNet [5], and to ensure diversity and visual quality of the shapes.", + "bbox": [ + 496, + 537, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define a visual similarity function $d: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\mathbb{R}$ between two shapes $s_i, s_j \\in S$ as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder $\\phi_R: \\mathbb{R}^{h \\times w \\times 3} \\mapsto \\mathbb{R}^d$ .", + "bbox": [ + 496, + 645, + 890, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(s _ {i}, s _ {j}\\right) = \\left\\| \\phi_ {R} \\left(f _ {R} \\left(s _ {i}\\right)\\right) - \\phi_ {R} \\left(f _ {R} \\left(s _ {j}\\right)\\right) \\right\\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 733, + 836, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $f_{R}: S \\mapsto \\mathbb{R}^{3 \\times h \\times w}$ is a rendering function. The set of reconstructed shapes can then be defined as:", + "bbox": [ + 496, + 762, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname *{arg min}_{\\substack{S_{R}\\subseteq S\\\\ |S_{R}| = (1 - \\alpha)|S|}}\\sum_{s\\in S_{R}}d(s,\\widehat{s})\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 805, + 782, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $\\widehat{s}$ is the reconstructed shape, defined as:", + "bbox": [ + 500, + 856, + 808, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {s} = f _ {\\Theta} \\circ \\phi_ {\\Theta} (s)\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 883, + 750, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22576", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0de3b1f5fee953afb7187fafed52c153142556b1a000eebdb57aaa90f4ec887f.jpg", + "image_caption": [ + "Figure 3. Detailling the shape generation pipeline of ShapeWalk. Realistic 3D CAD shapes are reconstructed from the 3DCOMPAT++ [22] dataset into the GeoCode [17] shape program parameter space, using a mapping function $\\phi_{\\Theta}$ . Reconstructed shapes with an error over a fixed threshold are discarded using a visual similarity function $d$ , which is based on rendered feature similarity. Filtered pairs of shapes are then interpolated in the parameter space of the shape program, to generate shape chains $\\mathcal{P}^{(k)}$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 888, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With $f_{\\Theta} : \\Theta \\mapsto S$ the shape program mapping shape parameters to a 3D mesh, and $\\phi_{\\Theta} : S \\mapsto \\Theta$ an encoder based on DGCNN [17, 25] which samples points from a shape surface and regresses corresponding shape parameters in $\\Theta$ .", + "bbox": [ + 75, + 325, + 468, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We reconstruct shapes by first sampling pointclouds from the surface of the original meshes, and then feeding them into the DGCNN [25] encoder fine-tuned [17] to regress the parameters of the shape program. We discard a ratio of $\\alpha = 0.08$ of the shapes with the highest reconstruction error, which we measure as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder. This threshold is selected empirically by visually inspecting the shapes in the upper part of the reconstruction error distribution. In our dataset instance building on 3DComPAT++ shapes, we filter out a total of 89 shapes during this process, leading to a set of $|S_R| = 1113$ reconstructed shapes. This process can be scaled up to any dataset size and any shape domain, as long as a corresponding shape program is available.", + "bbox": [ + 75, + 386, + 470, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shape interpolation. After reconstructing a set $S_R$ of realistic shapes, we generate a set of shape chains $\\mathcal{P}^{(k)}$ by interpolating between the parameters of shape pairs.", + "bbox": [ + 75, + 643, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To that end, we first sample a set of shape pairs $(s_{\\theta_{\\mathrm{i}}}, s_{\\theta_{\\mathrm{j}}}) \\in S_{P} \\subseteq S_{R} \\times S_{R}$ , and partition them into $L = 10$ levels of proximity. These proximity levels are based on the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder.", + "bbox": [ + 75, + 689, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the set of shape pairs $S_P^{(l)}$ at level $l \\in [[1, L]]$ as:", + "bbox": [ + 76, + 792, + 460, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {P} ^ {(l)} = \\{(s _ {\\theta_ {\\mathrm {i}}}, s _ {\\theta_ {\\mathrm {j}}}) \\in \\mathcal {S} _ {P} | \\overline {{d (s _ {\\theta_ {\\mathrm {i}}} , s _ {\\theta_ {\\mathrm {j}}})}} \\in \\left[ \\frac {l - 1}{L}, \\frac {l}{L} \\right] \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 821, + 444, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\overline{d}$ is a visual similarity function, min-max normalized across all pairs of shapes.", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each pair $(s_{\\theta_i}, s_{\\theta_j}) \\in S_P^{(l)}$ , we then interpolate between the shape parameters $\\theta_i$ and $\\theta_j$ to generate a set of intermediate shape parameters $(\\theta_i, \\dots, \\theta_{i + (N - 2)}, \\theta_j) \\in \\Theta^N$ . The number of intermediate shapes $N$ is defined as the number of differing parameters between $\\theta_i$ and $\\theta_j$ , that is:", + "bbox": [ + 496, + 324, + 893, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN = \\left| \\left\\{k \\mid \\theta_ {\\mathbf {i}} [ k ] \\neq \\theta_ {\\mathbf {j}} [ k ] \\right\\} \\right|\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 414, + 782, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the proximity level is low, edits necessary to transition from one shape to another are larger in intensity as the shapes are more different, and smaller when $l$ is smaller. The ordering of intermediate edits is randomly sampled using a dependency-aware algorithm (for example, parameters relating to armrest height or width are sampled after the addition of armrests).", + "bbox": [ + 496, + 443, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 1, this generation process leads to plausible interpolated shapes, and to fine-grained and realistic sequences of edits with detailed metadata.", + "bbox": [ + 496, + 549, + 890, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We summarize the process of generating realistic shape chains in Figure 3.", + "bbox": [ + 496, + 594, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Random set. While the realistic set is based on reconstructed shapes from the 3DCOMPAT++ [15, 22] dataset, the random set is based on a collection of random shapes generated using the shape program. This alternative set aims at covering a large space of the parameter space of the GeoCode shape program. This subset is generated by systematically sampling various combinations of parameters within defined ranges determined by a minimum edit intensity, but does not necessarily result in realistic shapes.", + "bbox": [ + 496, + 640, + 892, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Text Instructions", + "text_level": 1, + "bbox": [ + 500, + 786, + 668, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rule-based generation. Given an edit vector $\\overline{\\theta_{ij}}$ describing the parameter changes necessary to transition from shape $i$ to $j$ , we generate a set of text instructions $\\mathbf{p}_{ij}$ describing this edit in natural language. We first map each parameter $\\theta_{i}[k]$ composing the edit vector to a natural language name, and map the magnitude of the", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22577", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f747f3fb70df19623f37cb0c8fd5e2a1d25ecf62b3e6fedd4a46afcb2766a31.jpg", + "image_caption": [ + "Figure 4. Dataset chains samples. We feature in this figure shape chains from our dataset with associated generated text instructions. Each shape chain is composed of a sequence of shapes, with each consecutive pair of shapes associated to a set of text instructions describing the edits necessary to transition from one shape to the next. Overall, generated text instructions are accurate and provide a detailed description of the edits applied to the shapes. We include in the supplementary material additional samples of generated shape chains, alongside shape chains generated for two additional table and vase categories." + ], + "image_footnote": [], + "bbox": [ + 112, + 97, + 851, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "parameter change to a natural language intensity depending on the parameter type. Boolean parameters are also described appropriately, for example by using the word add or remove when describing the addition or removal of a part.", + "bbox": [ + 75, + 603, + 470, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Augmentation. This rule-based instruction generation leads to perfectly accurate instructions, but may lack diversity. To alleviate this issue, we 1) randomly sample adjectives and adverbs to characterize the intensity of edits, 2) randomly invert parameter names and magnitude directions (e.g. \"increase armrests straightness\" will be augmented to \"decrease armrests bend\" for the same edit). As a final augmentation, we utilize a T5 [18] transformer model finetuned for paraphrasing to generate additional instructions for each edit vector. We use the Parrot library [6] to filter generated paraphrases based on a fluency and adequacy score.", + "bbox": [ + 75, + 684, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarize the process of generating text instructions in the supplementary material.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Neural Shape Editing", + "text_level": 1, + "bbox": [ + 500, + 601, + 705, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Problem", + "text_level": 1, + "bbox": [ + 500, + 628, + 602, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We are concerned with the task of editing a shape $s_i$ given a sequence of natural language edit instructions. We want to learn an editing function $f_E: S \\times \\Sigma \\mapsto S$ , where $S$ is the space of input shapes, and $\\Sigma$ is the space of natural language edit instructions, able to compose edits from a starting shape $s_1$ to an ending shape $s_N$ , given a sequence of edit instructions $\\{\\mathbf{p_t}\\}_{t=1}^N$ .", + "bbox": [ + 496, + 654, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {s _ {N}} = f _ {E} (\\phi_ {T} (\\mathbf {p _ {1 2}}), f _ {E} (\\phi_ {T} (\\mathbf {p _ {2 3}}), \\ldots f _ {E} (\\phi_ {T} (\\mathbf {p _ {k N}}), s _ {1})))\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 776, + 893, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $N = |\\mathcal{P}|$ is the length of the edit chain, and $\\phi_T: \\Sigma \\mapsto \\mathbb{R}^{D_T}$ is a text encoder. At the end of the edit sequence, we want to recover a shape $\\widehat{s_N}$ as close as possible to the ground-truth shape $s_N$ . Our proposed dataset provides a ground-truth for every intermediate shape $s_k$ , which we leverage to train and evaluate our method.", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22578", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1669385cc877170fbfb73361aecfca58418d7e9a4a3fdbb0984ef34ad3312bb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 66, + 869, + 330 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/48027af70a79e59ef61ea06eed686f69fc495704f45fda8d2ab7eff75e458833.jpg", + "image_caption": [ + "Figure 5. Visualizing synthesized pairs of edits using our latent editors. We illustrate the diverse range of edits generated by our proposed latent mapper, showcasing its ability to capture nuanced variations in the input data, for 3DS2VS latents (top two rows) and PC-AE latents (bottom two rows). Each pair demonstrates the original input on the left and the corresponding synthesized output on the right, with the associated caption. We abbreviate instructions which are too long to fit in the figure, and indicate when text is omitted with an ellipsis. For more subtle changes, we circle the areas of interest in both source and target shapes. Overall, edits are consistent with the input instruction, preserve shape identity while generating plausible output shapes." + ], + "image_footnote": [], + "bbox": [ + 89, + 339, + 864, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Objective", + "text_level": 1, + "bbox": [ + 76, + 724, + 187, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One of the main advantages of training a shape editing model with ShapeWalk compared to other works [2, 8] is the availability of exact ground-truth edited shapes for each edit instruction (see Section 2). We thus formulate the objective function of our learning problem as a simple $\\mathcal{L}_2$ loss between the latents of the synthesized shape and the ground-truth shape. For all pairs of shapes $(s_i,s_j)$ and their corresponding edit instructions $\\mathbf{p_{ij}}$ , we minimize the following loss:", + "bbox": [ + 75, + 748, + 468, + 882 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| f _ {E} \\left(\\phi_ {E} \\left(s _ {i}\\right), \\mathbf {p} _ {\\mathbf {i j}}\\right) - \\phi_ {E} \\left(s _ {j}\\right) \\right\\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 883, + 392, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\phi_{AE} = \\phi_E\\circ \\phi_D$ is a shape autoencoder, and $\\phi_E:$ $S\\mapsto \\mathbb{R}^{D_E}$ denotes its encoder component.", + "bbox": [ + 498, + 724, + 890, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Models", + "text_level": 1, + "bbox": [ + 500, + 768, + 591, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For $\\phi_{AE}$ , we experiment with PC-AE [1] and 3D2VS [28] models both pre-trained on ShapeNet [5]. Note that our method is agnostic to the choice of shape autoencoder and can be adapted to a variety of shape representations. For $\\phi_T$ , we use a pretrained BERT [7] model.", + "bbox": [ + 496, + 792, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both $\\phi_{AE}$ and $\\phi_T$ are frozen during training, and we only train the parameters of our latent mapper $f_{E}$ .", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22579", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For PC-AE, we formulate our latent mapper as a neural module which for each tuple $(\\phi_E(s_i), \\mathbf{p_{ij}}, \\phi_E(s_j))$ predicts an edit vector $\\widehat{\\theta_{ij}} \\in \\mathbb{R}^{D_E}$ in the feature space of $\\phi_E$ , and adds it to the latent of the input shape $s_i$ to generate the latent of the output shape $s_j$ :", + "bbox": [ + 75, + 90, + 468, + 170 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf _ {E} \\left(\\phi_ {E} \\left(s _ {i}\\right), \\mathbf {p} _ {\\mathbf {i j}}\\right) = \\widehat {\\theta_ {i j}} + \\phi_ {E} \\left(s _ {i}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 183, + 385, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We utilize two variants of our latent mapper $f_{E}$ : one in which the edit vector $\\widehat{\\theta_{ij}}$ is predicted directly, and one in which the edit vector $\\widehat{\\theta_{ij}}$ is predicted as a product of a normalized edit direction $\\widehat{v_{ij}} \\in \\mathbb{R}^{D_E}$ and a magnitude $\\widehat{m_{ij}} \\in \\mathbb{R}$ separately.", + "bbox": [ + 75, + 217, + 468, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For 3D2VS which encodes shapes as latents sets of higher resolutions, we use a transformer-based latent diffusion model instead to generate edited latents. We concatenate the input shape latent $\\phi_E(s_i)$ with the noised edited shape latent $\\phi_E(s_j + \\epsilon_t)$ and feed the BERT text embeddings to the cross-attention layers of the transformer blocks to predict the added noise. We illustrate our architecture for the latent mapper in the supplementary material.", + "bbox": [ + 75, + 299, + 468, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 436, + 209, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Chained Shape Editing", + "text_level": 1, + "bbox": [ + 76, + 462, + 292, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison Models. For PC-AE, we experiment with three variants of our latent mapper module:", + "bbox": [ + 75, + 487, + 468, + 518 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DIRECTGEN directly predicts the edited shape latent without regressing an edit vector.", + "- LATEFUSION follows the network proposed in [3], and first passes the shape embeddings through an encoder. Encoded shape features are then concatenated with text embeddings and fed into a neural module which finally predicts the edit vector.", + "- Ours is a simple multi-layer perceptron (MLP) which directly takes the context (shape and text embeddings) as input to predict an edit vector." + ], + "bbox": [ + 76, + 531, + 468, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For all of these variants, we experiment with various bottleneck dimensions and number of layers (in subscript).", + "bbox": [ + 75, + 715, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further decompose these variants into coupled and decoupled versions, where we predict a normalized edit direction $\\widehat{v_{ij}}\\in \\mathbb{R}^{D_E}$ and a magnitude $\\widehat{m_{ij}}\\in \\mathbb{R}$ separately. We provide in appendix the full architecture details and training hyperparameters employed to train our models.", + "bbox": [ + 75, + 748, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that we do not compare with ChangeIt3D [2] as their neural-listener distillation method is not applicable in our context: we train our models with a direct feedback signal from the ground-truth edited shapes.", + "bbox": [ + 75, + 824, + 468, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single-step metrics. We propose to measure the quality of our editing steps both before and after shape reconstruction. Since we experiment on point clouds, we use the scaled Chamfer Distance between the reconstructed original shape $s_i$ and the reconstructed edited shape $\\widehat{s_j}$ :", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {C D}} = \\frac {1}{| \\widehat {s _ {j}} |} \\sum_ {x \\in \\widehat {s _ {j}}} \\min _ {y \\in s _ {i}} \\| x - y \\| _ {2} + \\frac {1}{| s _ {i} |} \\sum_ {y \\in s _ {i}} \\min _ {x \\in \\widehat {s _ {j}}} \\| x - y \\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 176, + 880, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also use the $\\mathcal{L}_2$ distance between the latents of the original shape $s_i$ and the edited shape $s_j$ :", + "bbox": [ + 498, + 224, + 888, + 255 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathcal {L} _ {2}} = \\left\\| \\phi_ {E} (s _ {i}) - \\phi_ {E} (s _ {j}) \\right\\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 263, + 790, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-step metrics. To extend the single-step metrics to chained shape generation, we take inspiration from the trajectory prediction literature [4, 26] and propose appropriate metrics for our task.", + "bbox": [ + 496, + 291, + 890, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a shape distance function $d: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\mathbb{R}$ , a chain of ground-truth reconstructed shapes $\\{s_1, \\ldots, s_N\\}$ , and a set of recursively generated shapes $\\{\\widehat{s}_1, \\ldots, \\widehat{s_N}\\}$ , we define the average edit error as the average distance between the generated shapes and the corresponding ground-truth shapes:", + "bbox": [ + 496, + 351, + 890, + 441 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {d} = \\frac {1}{N} \\sum_ {t = 1} ^ {N} d \\left(s _ {t}, \\widehat {s} _ {t}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 441, + 769, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "And the final edit error as the distance between the last shape in the chain and the corresponding ground-truth shape:", + "bbox": [ + 496, + 486, + 890, + 531 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F} _ {d} = d \\left(s _ {N}, \\widehat {s _ {N}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 531, + 751, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report results with both metrics for $d = d_{\\mathrm{CD}}$ and $d = d_{\\mathcal{L}_2}$ .", + "bbox": [ + 496, + 553, + 890, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We provide qualitative results of edited pairs generated by our latent mapper in Figure 5. Overall, we observe that our method is able to generate plausible edits that are consistent with the input instruction, and preserve the identity of the input shape. However, our method is limited by the quality of the shape autoencoder. We notice that the PC-AE editor is unable to perform very fine-grained edits which are not properly captured by the autoencoder, while visible in the ground-truth shapes. We explore the possible causes of these limitations in Section 5.2.", + "bbox": [ + 496, + 598, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide quantitative results for our chained editing metrics averaged across $|\\mathcal{P}| \\in \\{10, 15, 20\\}$ in Table 1, and detail per-chain length results in the appendix. Directly predicting the edited shape latent (DIRECTGEN) leads to poor edit predictions across all metrics, although increasing the number of layers may improve the results. Overall, decoupling the edit direction and magnitude leads to better results. Our coupled $\\mathrm{OURS}_{512 \\times 4}$ variant performs better than the coupled $\\mathrm{LATEFUSION}_{512}$ variant on the Chamfer Distance, but worse on $\\mathcal{L}_2$ distance.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "footer", + "text": "22580", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c51a4417244b7178f536fed1e2674294ca4ddf3d0e36ea1933755b6ab20c6573.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modeldecoupled?Averaged ∀|P|
FCD×1e4ACD×1e4FL2AL2
LATEFUSION1024×2.8562.6211.4441.251
LATEFUSION512×2.7192.6091.4621.290
LATEFUSION256×2.8742.6511.5091.283
OURS512×8×3.2082.8221.7031.437
OURS512×4×2.9902.7031.5891.351
LATEFUSION1024×2.7112.5681.4051.245
LATEFUSION512×3.0022.7081.4511.254
LATEFUSION256×3.3092.8481.5521.324
OURS512×8×2.7822.5841.4971.290
OURS512×4×2.6702.5241.4471.266
", + "bbox": [ + 76, + 75, + 470, + 313 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b9ef74325056a1084144054ab1a36c8b0cd66489816d2231f14432add2ad183.jpg", + "table_caption": [ + "Table 1. Chained shape editing ablation. We report baseline results for our proposed chained shape editing task, averaged across all chain lengths $|\\mathcal{P}| \\in \\{10, 15, 20\\}$ . Both the average final and average edit error are reported for the Chamfer Distance (CD) and $\\mathcal{L}_2$ distance $(\\mathcal{L}_2)$ metrics." + ], + "table_footnote": [], + "table_body": "
ParameterAccuracy
seat height1.000
backrest curvature1.000
object width/height/depth0.978
seat roundness0.978
top bar thickness/height0.961
legs thickness0.956
legs bending/curvature0.944
adding/removing handle cushions0.750
number of legs/backrest rails0.663
legs roundness/indentation0.587
AVG (RANDOM)0.884
AVG (REALISTIC)0.969
", + "bbox": [ + 96, + 406, + 450, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Parameter-wise accuracy for shape edit recognition. We report the accuracy of our edit detector for each parameter in the random set, alongside the average accuracy for the full random and realistic sets. Overall, we observe that our classifier is able to recognize edits corresponding to high change edits, but fails to differentiate between subtle changes.", + "bbox": [ + 75, + 628, + 468, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Recognizing Shape Edits", + "text_level": 1, + "bbox": [ + 76, + 726, + 305, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Problem. To provide additional insight into the quality of our latent mapper and the difficulty of predicting specific edits, we train a binary classifier to discriminate from a pair of shapes $(s_i, s_j)$ on whether the corresponding edit instruction $\\mathbf{p}_{\\mathrm{ij}}$ was applied to $s_i$ or $s_j$ . We train a binary classifier $f_C: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\{0,1\\}$ , where $f_C(s_i, s_j) = 1$ if the edit instruction $\\mathbf{p}_{\\mathrm{ij}}$ was applied to $s_i$ , and 0 otherwise.", + "bbox": [ + 75, + 750, + 468, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Method. Similarly to our latent mapper architecture for PC-AE, we use an MLP-based shape latent encoder to encode", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the input shape latents into the bottleneck dimension, and project text features using a single linear layer to the same dimension. We then concatenate the two feature vectors and pass them through an MLP predictor outputting probabilities for the two classes $s_i$ and $s_j$ . We train our classifier using a binary cross-entropy loss.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training and evaluation. We train our classifier on the realistic set and evaluate it on the full random shape set, which has a restricted vocabulary of edit instructions. We evaluate the accuracy of the classifier on the full set, as well as on subsets of the data corresponding to specific types of edits.", + "bbox": [ + 496, + 181, + 893, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. In Table 2, we provide accuracy results for shape edit recognition, depending on the edit type. We also provide the global accuracy on the random and realistic sets. We observe that our classifier is able to recognize edits corresponding to shape parameters associated with a high degree of shape variation (like seat height and width/height of the shape), but struggles with parameters which are more subtle (like legs roundness, handle cushions). Our classifier is able to detect edits related to global width/height/depth adjustments with an accuracy of $97.8\\%$ . This accuracy drops to $94.4\\%$ for legs bending/curvature, which is a more subtle change in the shape, and to $58.7\\%$ for legs roundness/indentation. We also note that the model does not perform well on edits related to the number of parts, i.e. adding or removing a number $k$ of legs or backrest rails.", + "bbox": [ + 496, + 272, + 892, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We impute these discrepancies to several factors. The PC-AE shape autoencoder we employ may be unable to properly reconstruct fine-grained details in the ground-truth and edited shapes, which renders the task of detecting subtle changes more difficult. Pointcloud resolution is also a limitation, as fine-grained parameters like leg indentation may not be properly sampled. Finally, learning a latent mapper and a classifier that can accurately count the number of parts to add or remove may also be a difficult task of its own, and require more domain-specific inductive biases.", + "bbox": [ + 496, + 500, + 893, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 619, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, this paper introduces the ShapeWalk dataset, designed for advancing compositional shape editing guided by natural language instructions. The dataset comprises 158K unique shapes connected through 26K edit chains, synthesized from a realistic CAD-designed 3D dataset. Language instructions for applied edits are provided, alongside exact ground-truth edited shapes. Our method requires zero human annotation effort, and can be scaled up indefinitely. The usefulness of the dataset is demonstrated by training neural editor modules in the latent space of shape autoencoders [1]. Evaluation metrics inspired by trajectory prediction literature [4, 26] are introduced, offering a quantitative assessment of the quality of recursively edited shapes.", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2, 6, 8", + "[2] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 6, 7", + "[3] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. ShapeTalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 2, 3, 7", + "[4] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961–971, 2016. 2, 7, 8", + "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. In arXiv, 2015. 3, 6", + "[6] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 5", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 6", + "[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2023. 3, 6", + "[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 4", + "[11] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics, 39(6):1-20, 2020. 2", + "[12] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: macro operation discovery for 3d shape programs. ACM Transactions on Graphics, 40(4), 2021. 2", + "[13] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2", + "[14] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics, 42(4), 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yuchen Li, Ujjwal Upadhyay, Habib Slim, Ahmed Abdelreheem, Arpit Prajapati, Suhail Pothigara, Peter Wonka, and Mohamed Elhoseiny. 3D CoMPaT: Composition of Materials on Parts of 3D Things. In ECCV, 2022. 3, 4", + "[16] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3", + "[17] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. In arXiv preprint arxiv:2212.11715, 2022. 2, 4", + "[18] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In arXiv, 2023. 5", + "[19] Daniel Ritchie, Paul Guerrero, R. Kenny Jones, Niloy J. Mitra, Adriana Schulz, Karl D. D. Willis, and Jiajun Wu. Neurosymbolic models for computer graphics. Computer Graphics Forum, 42(2), 2023. 2", + "[20] Aditya Sanghi, Hang Chu, Joseph G. Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2", + "[21] Aditya Sanghi, Rao Fu, Vivian Liu, Karl D.D. Willis, Hooman Shayani, Amir H. Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 2, 3", + "[22] Habib Slim, Xiang Li, Mahmoud Ahmed Yuchen Li, Mohamed Ayman, Ujjwal Upadhyay Ahmed Abdelreheem, Suhail Pothigara Arpit Prajapati, Peter Wonka, and Mohamed Elhoseiny. 3DCoMPaT++: An improved large-scale 3d vision dataset for compositional recognition. In arXiv preprint arxiv:2212.11715, 2023. 3, 4", + "[23] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. CoRR, abs/1711.00937, 2017. 3", + "[24] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2", + "[25] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. In SIGGRAPH, 2019. 4", + "[26] Kaiping Xu, Zheng Qin, Guolong Wang, Kai Huang, Shuxiong Ye, and Huidi Zhang. Collision-Free LSTM for Human Trajectory Prediction, pages 106-116. 2018. 2, 7, 8", + "[27] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22582", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[28] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics, 42(4):1-16, 2023. 6", + "bbox": [ + 78, + 90, + 470, + 147 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22583", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_model.json b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_model.json new file mode 100644 index 0000000000000000000000000000000000000000..637c8cdc473b2966ea9ea46cc5fedf24262a24f6 --- /dev/null +++ b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_model.json @@ -0,0 +1,1815 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "ShapeWalk: Compositional Shape Editing through Language-Guided Chains" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.182, + 0.367, + 0.216 + ], + "angle": 0, + "content": "Habib Slim KAUST" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.22, + 0.424, + 0.233 + ], + "angle": 0, + "content": "habib.smil@kaust.edu.sa" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.182, + 0.7, + 0.216 + ], + "angle": 0, + "content": "Mohamed Elhoseiny KAUST" + }, + { + "type": "text", + "bbox": [ + 0.48, + 0.22, + 0.748, + 0.234 + ], + "angle": 0, + "content": "mohamed.elhoseiny@kaust.edu.sa" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.275, + 0.86, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.605, + 0.896, + 0.685 + ], + "angle": 0, + "content": "Figure 1. Visualizing shape chains from ShapeWalk. Our dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. We illustrate the interpolation process by coloring shapes based on their proximity to the starting shape (blue), and the ending shape (orange). Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. For each shape transition, we also provide a precise edit vector \\(\\overline{\\theta_{ij}}\\) describing the parameter changes necessary to transition from one shape to the next." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.695, + 0.314, + 0.71 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Editing 3D shapes through natural language instructions is a challenging task that requires the comprehension of both language semantics and fine-grained geometric details. To bridge this gap, we introduce ShapeWalk, a carefully designed synthetic dataset designed to advance the field of language-guided shape editing. The dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. We synthesize edit chains by reconstructing and interpolating shapes sampled from a realistic CAD-designed 3D dataset in the parameter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.895, + 0.879 + ], + "angle": 0, + "content": "space of the GeoCode shape program. We leverage rule-based methods and language models to generate accurate and realistic natural language prompts corresponding to each edit. To illustrate the practicality of our contribution, we train neural editor modules in the latent space of shape autoencoders, and demonstrate the ability of our dataset to enable a variety of language-guided shape edits. Finally, we introduce multi-step editing metrics to benchmark the capacity of our models to perform recursive shape edits. We hope that our work will enable further study of compositional language-guided shape editing, and finds application in 3D CAD design and interactive modeling." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22574" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.468, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.472, + 0.334 + ], + "angle": 0, + "content": "Figure 2. Comparing ShapeTalk with our work. We compare the ShapeTalk [3] dataset (top) with our work (bottom). For an equivalent edit instruction (in green), ShapeTalk provides pairs of shapes with many factors of variation, while we generate synthetic pairs of shapes with a single clear varying factor." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.357, + 0.21, + 0.372 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Whether in the realms of computer-aided design, virtual reality environments, or digital content creation, the process of refining and enhancing 3D visual data often involves intricate adjustments to geometric shapes, textures, and lighting. Furthermore, the necessity for precise modifications adds an additional layer of complexity, as inaccuracies can have profound implications on the final output. This labor-intensive nature of 3D data editing not only hinders workflow efficiency but also poses a barrier for individuals without specialized skills, limiting the democratization of 3D content manipulation. As such, there exists a compelling need for innovative solutions that streamline and democratize the 3D editing process, making it more accessible to a broader range of users. But how to train models with the ability to perform these complex edits?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.839 + ], + "angle": 0, + "content": "To that end, we introduce ShapeWalk, a dataset aiming at addressing the challenges inherent in the intricate and skill-intensive nature of 3D data editing. Our dataset consists of synthesized chains of 3D shapes, connected through edit vectors associated with precise language instructions describing the applied edits. It contains 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Our generation method is scalable and can yield an indefinite number of realistic shape chains, and extended to any 3D domain and shape program. To our knowledge, our dataset is the first to provide a large-scale collection of realistic 3D shape edit chains with precise language instructions, and matching ground-truth edited shapes. Our method yields specific edits undiluted by other factors of variation, as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To validate the usefulness of our dataset, we train neural editor modules in the latent space of shape autoencoders [1], and demonstrate the ability of our dataset to train models capable of performing a variety of shape edits. We learn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.196 + ], + "angle": 0, + "content": "latent mappings in the space of a frozen shape autoencoder, and show that our method can be applied on shapes unseen during the training of the neural editor. Taking advantage of the ground-truth shape chains provided by our dataset, we introduce multi-step editing metrics inspired from the field of trajectory prediction [4, 26] to evaluate the quality of our models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.199, + 0.825, + 0.213 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.225, + 0.892, + 0.301 + ], + "angle": 0, + "content": "- Chained Shape-Editing Dataset: We introduce a synthetic dataset of chained edits, with associated language instructions and ground-truth edited shapes. This is the first dataset of its kind, and is designed to facilitate the study of compositional shape editing." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.313, + 0.892, + 0.387 + ], + "angle": 0, + "content": "- Neural Shape Editor: We introduce a neural editor module trained on our dataset, and demonstrate its ability to perform a variety of shape edits. We show that our method can be applied on shapes unseen during the training of the neural editor." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.4, + 0.892, + 0.461 + ], + "angle": 0, + "content": "- Chained-Editing Metrics: With the introduction of ground-truth shape chains, we propose multi-step editing metrics to evaluate the ability of neural editors to perform recursively applied edits." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.225, + 0.892, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.479, + 0.637, + 0.494 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.702 + ], + "angle": 0, + "content": "3D Shape Programs. A significant body of work [9, 11, 12, 14, 19] is dedicated to exploring the use of procedural programs for 3D shape representation. Representing 3D objects as visual programs has many advantages. Representing shapes as visual primitives enhances interpretability [11, 14], as programs are by definition human-readable and thus understandable by human experts. Shape programs are more compact [9, 27] than their usual 3D shape modalities, and can be used to represent shapes more efficiently. Programs can also be composed to create shapes [11, 14], or decomposed into smaller programs. Ideally, shape programs can be edited [13] to modify the underlying geometry they represent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Our dataset consists of a collection of synthetic shapes created utilizing a backbone mesh-generating shape program. Specifically, we build from GeoCode [17], a 3D shape synthesis technique addressing the challenge of mapping high-fidelity geometry to an editable parameter space. GeoCode introduces a procedural program enabling the generation of high-quality mesh outputs with a balanced blend of interpretability and fine control." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Language-Guided 3D Shape Editing. Various works have explored the use of natural language instructions to guide 3D shape editing. Recent efforts [16, 20, 21] propose to leverage pre-trained CLIP [24] models to align" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "3D shapes with a given text prompt. Early efforts in that direction [16] included optimizing meshes to progressively align them with CLIP embeddings, leading to important computational overhead. CLIP-Sculptor [21], a more recent work, leverages a voxelized representation and a discrete latent space conditioned on CLIP's image-text embeddings to perform fast and fidel shape edits without shape optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.473, + 0.445 + ], + "angle": 0, + "content": "Shape Editing Datasets. Other works leverage large text-aligned 3D shape datasets and require significant manual annotation effort. ShapeCrafter [8] generates 3D shapes incrementally from text using a neural network, evolving with additional phrases. ShapeCrafter is designed for recursive shape edition and utilizes a VQ-VAE [23] model to represent shapes as discrete codes. This method exhibits consistent shape-text alignment with gradual evolution. ChangeIt3D [3] introduces ShapeTalk, a large dataset for describing 3D shape differences. The framework facilitates language-based editing of 3D models without requiring 2D to 3D conversion methods, and learns a shape editing model by learning contrasts between sampled shape pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.903 + ], + "angle": 0, + "content": "ShapeTalk [3] is the most similar work to our proposed dataset, and is the only publicly available dataset specifically designed for language-guided 3D shape-to-shape editing. ShapeTalk is a remarkable contribution in the field of language-guided shape editing, and is one of the first works to leverage large-scale 3D shape datasets to facilitate the study of language-guided shape editing. However, this work has some limitations that we attempt to address in our work. Collecting a dataset of shape differences is a challenging task requiring considerable manual annotation effort, and gathering a large number of 3D shapes may not be feasible for all domains. Furthermore, ShapeTalk is composed of edit contexts (i.e. shape pairs and edit instructions) which are hard to separate into fine-grained, composable edits. Most of the time, singular shape edits do not have an exact ground-truth in the form of a source and target shape pair. We illustrate this important distinction in Figure 2. The availability of singular shape edits with an associated ground-truth would ease benchmarking the quality of shape editing methods, and facilitate the training process of editing models. ShapeTalk also does not provide a mechanism to generate edit chains, which are necessary to study compositional shape editing. Finally, ShapeTalk does not easily enable the study of shape editing in a compositional manner. In contrast, our dataset is 1) synthesized by augmenting a small set of diverse shapes and can easily be scaled up, 2) composed of fine-grained edits and coarse-grained edits with exact ground-truths, and 3) separated into edit chains, which are designed to facilitate the study of compositional shape editing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.097, + 0.622, + 0.114 + ], + "angle": 0, + "content": "3. ShapeWalk" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.894, + 0.215 + ], + "angle": 0, + "content": "Our dataset contains 158K shapes split into a random and realistic set. With each edge connecting two consecutive shapes, we also produce a text instruction generated using the parameter changes necessary to transition from one shape to the next. We detail the generation method of our dataset here, and summarize the process in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.225, + 0.597, + 0.24 + ], + "angle": 0, + "content": "3.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.894, + 0.308 + ], + "angle": 0, + "content": "Definition. Our dataset can be defined as a collection of directed graph paths (dipaths), denoted as \\(\\mathcal{P}^{(k)} = (S^{(k)},E^{(k)},f^{(k)})\\). For each shape chain \\(\\mathcal{P}^{(k)}\\) of length \\(l\\), we denote:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.312, + 0.894, + 0.377 + ], + "angle": 0, + "content": "- \\( S^{(k)} = \\{s_{\\theta_1}^{(k)}, \\ldots, s_{\\theta_N}^{(k)}\\} \\), the set of distinct shape nodes composing the chain. Each shape is defined by a set of parameters \\( \\theta_i \\subseteq \\Theta \\), where \\( \\Theta \\) is the linear parameter space of our shape program." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.387, + 0.892, + 0.422 + ], + "angle": 0, + "content": "- \\( E^{(k)} \\subseteq \\{(s_{\\theta_{\\mathbf{i}}}^{(k)}, s_{\\theta_{\\mathbf{j}}}^{(k)}) \\mid s_{\\theta_{\\mathbf{i}}}^{(k)}, s_{\\theta_{\\mathbf{j}}}^{(k)} \\in V^{(k)}, j = i + 1\\} \\), the set of edges linking each consecutive pair of shapes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.431, + 0.894, + 0.524 + ], + "angle": 0, + "content": "- \\( f^{(k)}: E^{(k)} \\mapsto \\{(\\overline{\\theta_{ij}}, \\mathbf{p_{ij}}) \\mid \\overline{\\theta_{ij}} \\subseteq \\Theta, \\mathbf{p_{ij}} \\in \\Sigma \\} \\), a function mapping each edge to a vector \\( \\overline{\\theta_{ij}} \\subseteq \\Theta \\) and a set of text instructions \\( \\mathbf{p_{ij}} \\). \\( \\overline{\\theta_{ij}} \\) defines the parameter changes, or edits necessary, to go from shape \\( i \\) to \\( j \\). \\( \\mathbf{p_{ij}} \\) is a set of text instructions describing this edit in natural language." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.312, + 0.894, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.645 + ], + "angle": 0, + "content": "Generation. To generate our dataset, we start by reconstructing a set \\(S \\subseteq S\\) of realistic 3D CAD shapes into the space of shapes covered by our shape program \\(S_{\\Theta} \\subset S\\). To that end, we utilize shapes from the 3DCOMPAT++ [15, 22] dataset, a realistic, industry-based 3D CAD dataset. We employ this dataset to avoid overlap with ShapeNet [5], and to ensure diversity and visual quality of the shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.721 + ], + "angle": 0, + "content": "We define a visual similarity function \\(d: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\mathbb{R}\\) between two shapes \\(s_i, s_j \\in S\\) as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder \\(\\phi_R: \\mathbb{R}^{h \\times w \\times 3} \\mapsto \\mathbb{R}^d\\)." + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.734, + 0.838, + 0.752 + ], + "angle": 0, + "content": "\\[\nd \\left(s _ {i}, s _ {j}\\right) = \\left\\| \\phi_ {R} \\left(f _ {R} \\left(s _ {i}\\right)\\right) - \\phi_ {R} \\left(f _ {R} \\left(s _ {j}\\right)\\right) \\right\\| _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Where \\( f_{R}: S \\mapsto \\mathbb{R}^{3 \\times h \\times w} \\) is a rendering function. The set of reconstructed shapes can then be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.806, + 0.783, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\operatorname *{arg min}_{\\substack{S_{R}\\subseteq S\\\\ |S_{R}| = (1 - \\alpha)|S|}}\\sum_{s\\in S_{R}}d(s,\\widehat{s})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.857, + 0.81, + 0.872 + ], + "angle": 0, + "content": "Where \\(\\widehat{s}\\) is the reconstructed shape, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.885, + 0.751, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\widehat {s} = f _ {\\Theta} \\circ \\phi_ {\\Theta} (s)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22576" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.096, + 0.889, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.896, + 0.3 + ], + "angle": 0, + "content": "Figure 3. Detailling the shape generation pipeline of ShapeWalk. Realistic 3D CAD shapes are reconstructed from the 3DCOMPAT++ [22] dataset into the GeoCode [17] shape program parameter space, using a mapping function \\(\\phi_{\\Theta}\\). Reconstructed shapes with an error over a fixed threshold are discarded using a visual similarity function \\(d\\), which is based on rendered feature similarity. Filtered pairs of shapes are then interpolated in the parameter space of the shape program, to generate shape chains \\(\\mathcal{P}^{(k)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.386 + ], + "angle": 0, + "content": "With \\( f_{\\Theta} : \\Theta \\mapsto S \\) the shape program mapping shape parameters to a 3D mesh, and \\( \\phi_{\\Theta} : S \\mapsto \\Theta \\) an encoder based on DGCNN [17, 25] which samples points from a shape surface and regresses corresponding shape parameters in \\( \\Theta \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.471, + 0.629 + ], + "angle": 0, + "content": "We reconstruct shapes by first sampling pointclouds from the surface of the original meshes, and then feeding them into the DGCNN [25] encoder fine-tuned [17] to regress the parameters of the shape program. We discard a ratio of \\(\\alpha = 0.08\\) of the shapes with the highest reconstruction error, which we measure as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder. This threshold is selected empirically by visually inspecting the shapes in the upper part of the reconstruction error distribution. In our dataset instance building on 3DComPAT++ shapes, we filter out a total of 89 shapes during this process, leading to a set of \\(|S_R| = 1113\\) reconstructed shapes. This process can be scaled up to any dataset size and any shape domain, as long as a corresponding shape program is available." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Shape interpolation. After reconstructing a set \\( S_R \\) of realistic shapes, we generate a set of shape chains \\( \\mathcal{P}^{(k)} \\) by interpolating between the parameters of shape pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.781 + ], + "angle": 0, + "content": "To that end, we first sample a set of shape pairs \\((s_{\\theta_{\\mathrm{i}}}, s_{\\theta_{\\mathrm{j}}}) \\in S_{P} \\subseteq S_{R} \\times S_{R}\\), and partition them into \\(L = 10\\) levels of proximity. These proximity levels are based on the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.793, + 0.462, + 0.813 + ], + "angle": 0, + "content": "We define the set of shape pairs \\(S_P^{(l)}\\) at level \\(l \\in [[1, L]]\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.823, + 0.446, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {P} ^ {(l)} = \\{(s _ {\\theta_ {\\mathrm {i}}}, s _ {\\theta_ {\\mathrm {j}}}) \\in \\mathcal {S} _ {P} | \\overline {{d (s _ {\\theta_ {\\mathrm {i}}} , s _ {\\theta_ {\\mathrm {j}}})}} \\in \\left[ \\frac {l - 1}{L}, \\frac {l}{L} \\right] \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Where \\(\\overline{d}\\) is a visual similarity function, min-max normalized across all pairs of shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.404 + ], + "angle": 0, + "content": "For each pair \\((s_{\\theta_i}, s_{\\theta_j}) \\in S_P^{(l)}\\), we then interpolate between the shape parameters \\(\\theta_i\\) and \\(\\theta_j\\) to generate a set of intermediate shape parameters \\((\\theta_i, \\dots, \\theta_{i + (N - 2)}, \\theta_j) \\in \\Theta^N\\). The number of intermediate shapes \\(N\\) is defined as the number of differing parameters between \\(\\theta_i\\) and \\(\\theta_j\\), that is:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.415, + 0.783, + 0.433 + ], + "angle": 0, + "content": "\\[\nN = \\left| \\left\\{k \\mid \\theta_ {\\mathbf {i}} [ k ] \\neq \\theta_ {\\mathbf {j}} [ k ] \\right\\} \\right|\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.548 + ], + "angle": 0, + "content": "When the proximity level is low, edits necessary to transition from one shape to another are larger in intensity as the shapes are more different, and smaller when \\( l \\) is smaller. The ordering of intermediate edits is randomly sampled using a dependency-aware algorithm (for example, parameters relating to armrest height or width are sampled after the addition of armrests)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.594 + ], + "angle": 0, + "content": "As illustrated in Figure 1, this generation process leads to plausible interpolated shapes, and to fine-grained and realistic sequences of edits with detailed metadata." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.626 + ], + "angle": 0, + "content": "We summarize the process of generating realistic shape chains in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.893, + 0.777 + ], + "angle": 0, + "content": "Random set. While the realistic set is based on reconstructed shapes from the 3DCOMPAT++ [15, 22] dataset, the random set is based on a collection of random shapes generated using the shape program. This alternative set aims at covering a large space of the parameter space of the GeoCode shape program. This subset is generated by systematically sampling various combinations of parameters within defined ranges determined by a minimum edit intensity, but does not necessarily result in realistic shapes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.669, + 0.802 + ], + "angle": 0, + "content": "3.2. Text Instructions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Rule-based generation. Given an edit vector \\(\\overline{\\theta_{ij}}\\) describing the parameter changes necessary to transition from shape \\(i\\) to \\(j\\), we generate a set of text instructions \\(\\mathbf{p}_{ij}\\) describing this edit in natural language. We first map each parameter \\(\\theta_{i}[k]\\) composing the edit vector to a natural language name, and map the magnitude of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22577" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.098, + 0.852, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Figure 4. Dataset chains samples. We feature in this figure shape chains from our dataset with associated generated text instructions. Each shape chain is composed of a sequence of shapes, with each consecutive pair of shapes associated to a set of text instructions describing the edits necessary to transition from one shape to the next. Overall, generated text instructions are accurate and provide a detailed description of the edits applied to the shapes. We include in the supplementary material additional samples of generated shape chains, alongside shape chains generated for two additional table and vase categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.472, + 0.667 + ], + "angle": 0, + "content": "parameter change to a natural language intensity depending on the parameter type. Boolean parameters are also described appropriately, for example by using the word add or remove when describing the addition or removal of a part." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Augmentation. This rule-based instruction generation leads to perfectly accurate instructions, but may lack diversity. To alleviate this issue, we 1) randomly sample adjectives and adverbs to characterize the intensity of edits, 2) randomly invert parameter names and magnitude directions (e.g. \"increase armrests straightness\" will be augmented to \"decrease armrests bend\" for the same edit). As a final augmentation, we utilize a T5 [18] transformer model finetuned for paraphrasing to generate additional instructions for each edit vector. We use the Parrot library [6] to filter generated paraphrases based on a fluency and adequacy score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We summarize the process of generating text instructions in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.707, + 0.621 + ], + "angle": 0, + "content": "4. Neural Shape Editing" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.604, + 0.645 + ], + "angle": 0, + "content": "4.1. Problem" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.763 + ], + "angle": 0, + "content": "We are concerned with the task of editing a shape \\( s_i \\) given a sequence of natural language edit instructions. We want to learn an editing function \\( f_E: S \\times \\Sigma \\mapsto S \\), where \\( S \\) is the space of input shapes, and \\( \\Sigma \\) is the space of natural language edit instructions, able to compose edits from a starting shape \\( s_1 \\) to an ending shape \\( s_N \\), given a sequence of edit instructions \\( \\{\\mathbf{p_t}\\}_{t=1}^N \\)." + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.777, + 0.895, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\widehat {s _ {N}} = f _ {E} (\\phi_ {T} (\\mathbf {p _ {1 2}}), f _ {E} (\\phi_ {T} (\\mathbf {p _ {2 3}}), \\ldots f _ {E} (\\phi_ {T} (\\mathbf {p _ {k N}}), s _ {1})))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Where \\(N = |\\mathcal{P}|\\) is the length of the edit chain, and \\(\\phi_T: \\Sigma \\mapsto \\mathbb{R}^{D_T}\\) is a text encoder. At the end of the edit sequence, we want to recover a shape \\(\\widehat{s_N}\\) as close as possible to the ground-truth shape \\(s_N\\). Our proposed dataset provides a ground-truth for every intermediate shape \\(s_k\\), which we leverage to train and evaluate our method." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.068, + 0.87, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.34, + 0.866, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.615, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Figure 5. Visualizing synthesized pairs of edits using our latent editors. We illustrate the diverse range of edits generated by our proposed latent mapper, showcasing its ability to capture nuanced variations in the input data, for 3DS2VS latents (top two rows) and PC-AE latents (bottom two rows). Each pair demonstrates the original input on the left and the corresponding synthesized output on the right, with the associated caption. We abbreviate instructions which are too long to fit in the figure, and indicate when text is omitted with an ellipsis. For more subtle changes, we circle the areas of interest in both source and target shapes. Overall, edits are consistent with the input instruction, preserve shape identity while generating plausible output shapes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.189, + 0.741 + ], + "angle": 0, + "content": "4.2. Objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.883 + ], + "angle": 0, + "content": "One of the main advantages of training a shape editing model with ShapeWalk compared to other works [2, 8] is the availability of exact ground-truth edited shapes for each edit instruction (see Section 2). We thus formulate the objective function of our learning problem as a simple \\(\\mathcal{L}_2\\) loss between the latents of the synthesized shape and the ground-truth shape. For all pairs of shapes \\((s_i,s_j)\\) and their corresponding edit instructions \\(\\mathbf{p_{ij}}\\), we minimize the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.885, + 0.393, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| f _ {E} \\left(\\phi_ {E} \\left(s _ {i}\\right), \\mathbf {p} _ {\\mathbf {i j}}\\right) - \\phi_ {E} \\left(s _ {j}\\right) \\right\\| _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.726, + 0.892, + 0.757 + ], + "angle": 0, + "content": "Where \\(\\phi_{AE} = \\phi_E\\circ \\phi_D\\) is a shape autoencoder, and \\(\\phi_E:\\) \\(S\\mapsto \\mathbb{R}^{D_E}\\) denotes its encoder component." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.593, + 0.784 + ], + "angle": 0, + "content": "4.3. Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.892, + 0.87 + ], + "angle": 0, + "content": "For \\(\\phi_{AE}\\), we experiment with PC-AE [1] and 3D2VS [28] models both pre-trained on ShapeNet [5]. Note that our method is agnostic to the choice of shape autoencoder and can be adapted to a variety of shape representations. For \\(\\phi_T\\), we use a pretrained BERT [7] model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Both \\(\\phi_{AE}\\) and \\(\\phi_T\\) are frozen during training, and we only train the parameters of our latent mapper \\(f_{E}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22579" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.171 + ], + "angle": 0, + "content": "For PC-AE, we formulate our latent mapper as a neural module which for each tuple \\((\\phi_E(s_i), \\mathbf{p_{ij}}, \\phi_E(s_j))\\) predicts an edit vector \\(\\widehat{\\theta_{ij}} \\in \\mathbb{R}^{D_E}\\) in the feature space of \\(\\phi_E\\), and adds it to the latent of the input shape \\(s_i\\) to generate the latent of the output shape \\(s_j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.184, + 0.387, + 0.205 + ], + "angle": 0, + "content": "\\[\nf _ {E} \\left(\\phi_ {E} \\left(s _ {i}\\right), \\mathbf {p} _ {\\mathbf {i j}}\\right) = \\widehat {\\theta_ {i j}} + \\phi_ {E} \\left(s _ {i}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.469, + 0.299 + ], + "angle": 0, + "content": "We utilize two variants of our latent mapper \\( f_{E} \\): one in which the edit vector \\( \\widehat{\\theta_{ij}} \\) is predicted directly, and one in which the edit vector \\( \\widehat{\\theta_{ij}} \\) is predicted as a product of a normalized edit direction \\( \\widehat{v_{ij}} \\in \\mathbb{R}^{D_E} \\) and a magnitude \\( \\widehat{m_{ij}} \\in \\mathbb{R} \\) separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.47, + 0.421 + ], + "angle": 0, + "content": "For 3D2VS which encodes shapes as latents sets of higher resolutions, we use a transformer-based latent diffusion model instead to generate edited latents. We concatenate the input shape latent \\(\\phi_E(s_i)\\) with the noised edited shape latent \\(\\phi_E(s_j + \\epsilon_t)\\) and feed the BERT text embeddings to the cross-attention layers of the transformer blocks to predict the added noise. We illustrate our architecture for the latent mapper in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.438, + 0.21, + 0.455 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.463, + 0.294, + 0.48 + ], + "angle": 0, + "content": "5.1. Chained Shape Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.469, + 0.52 + ], + "angle": 0, + "content": "Comparison Models. For PC-AE, we experiment with three variants of our latent mapper module:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.532, + 0.47, + 0.563 + ], + "angle": 0, + "content": "- DIRECTGEN directly predicts the edited shape latent without regressing an edit vector." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.576, + 0.469, + 0.651 + ], + "angle": 0, + "content": "- LATEFUSION follows the network proposed in [3], and first passes the shape embeddings through an encoder. Encoded shape features are then concatenated with text embeddings and fed into a neural module which finally predicts the edit vector." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.664, + 0.469, + 0.71 + ], + "angle": 0, + "content": "- Ours is a simple multi-layer perceptron (MLP) which directly takes the context (shape and text embeddings) as input to predict an edit vector." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.532, + 0.47, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.748 + ], + "angle": 0, + "content": "For all of these variants, we experiment with various bottleneck dimensions and number of layers (in subscript)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.824 + ], + "angle": 0, + "content": "We further decompose these variants into coupled and decoupled versions, where we predict a normalized edit direction \\(\\widehat{v_{ij}}\\in \\mathbb{R}^{D_E}\\) and a magnitude \\(\\widehat{m_{ij}}\\in \\mathbb{R}\\) separately. We provide in appendix the full architecture details and training hyperparameters employed to train our models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.886 + ], + "angle": 0, + "content": "Note that we do not compare with ChangeIt3D [2] as their neural-listener distillation method is not applicable in our context: we train our models with a direct feedback signal from the ground-truth edited shapes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.169 + ], + "angle": 0, + "content": "Single-step metrics. We propose to measure the quality of our editing steps both before and after shape reconstruction. Since we experiment on point clouds, we use the scaled Chamfer Distance between the reconstructed original shape \\( s_i \\) and the reconstructed edited shape \\( \\widehat{s_j} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.177, + 0.882, + 0.215 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {C D}} = \\frac {1}{| \\widehat {s _ {j}} |} \\sum_ {x \\in \\widehat {s _ {j}}} \\min _ {y \\in s _ {i}} \\| x - y \\| _ {2} + \\frac {1}{| s _ {i} |} \\sum_ {y \\in s _ {i}} \\min _ {x \\in \\widehat {s _ {j}}} \\| x - y \\| _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.225, + 0.89, + 0.256 + ], + "angle": 0, + "content": "We also use the \\(\\mathcal{L}_2\\) distance between the latents of the original shape \\(s_i\\) and the edited shape \\(s_j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.265, + 0.792, + 0.283 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathcal {L} _ {2}} = \\left\\| \\phi_ {E} (s _ {i}) - \\phi_ {E} (s _ {j}) \\right\\| _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Multi-step metrics. To extend the single-step metrics to chained shape generation, we take inspiration from the trajectory prediction literature [4, 26] and propose appropriate metrics for our task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Given a shape distance function \\(d: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\mathbb{R}\\), a chain of ground-truth reconstructed shapes \\(\\{s_1, \\ldots, s_N\\}\\), and a set of recursively generated shapes \\(\\{\\widehat{s}_1, \\ldots, \\widehat{s_N}\\}\\), we define the average edit error as the average distance between the generated shapes and the corresponding ground-truth shapes:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.442, + 0.771, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {d} = \\frac {1}{N} \\sum_ {t = 1} ^ {N} d \\left(s _ {t}, \\widehat {s} _ {t}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.532 + ], + "angle": 0, + "content": "And the final edit error as the distance between the last shape in the chain and the corresponding ground-truth shape:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.532, + 0.752, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F} _ {d} = d \\left(s _ {N}, \\widehat {s _ {N}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.586 + ], + "angle": 0, + "content": "We report results with both metrics for \\( d = d_{\\mathrm{CD}} \\) and \\( d = d_{\\mathcal{L}_2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Results. We provide qualitative results of edited pairs generated by our latent mapper in Figure 5. Overall, we observe that our method is able to generate plausible edits that are consistent with the input instruction, and preserve the identity of the input shape. However, our method is limited by the quality of the shape autoencoder. We notice that the PC-AE editor is unable to perform very fine-grained edits which are not properly captured by the autoencoder, while visible in the ground-truth shapes. We explore the possible causes of these limitations in Section 5.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We provide quantitative results for our chained editing metrics averaged across \\( |\\mathcal{P}| \\in \\{10, 15, 20\\} \\) in Table 1, and detail per-chain length results in the appendix. Directly predicting the edited shape latent (DIRECTGEN) leads to poor edit predictions across all metrics, although increasing the number of layers may improve the results. Overall, decoupling the edit direction and magnitude leads to better results. Our coupled \\( \\mathrm{OURS}_{512 \\times 4} \\) variant performs better than the coupled \\( \\mathrm{LATEFUSION}_{512} \\) variant on the Chamfer Distance, but worse on \\( \\mathcal{L}_2 \\) distance." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22580" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.077, + 0.472, + 0.314 + ], + "angle": 0, + "content": "
Modeldecoupled?Averaged ∀|P|
FCD×1e4ACD×1e4FL2AL2
LATEFUSION1024×2.8562.6211.4441.251
LATEFUSION512×2.7192.6091.4621.290
LATEFUSION256×2.8742.6511.5091.283
OURS512×8×3.2082.8221.7031.437
OURS512×4×2.9902.7031.5891.351
LATEFUSION1024×2.7112.5681.4051.245
LATEFUSION512×3.0022.7081.4511.254
LATEFUSION256×3.3092.8481.5521.324
OURS512×8×2.7822.5841.4971.290
OURS512×4×2.6702.5241.4471.266
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.323, + 0.472, + 0.395 + ], + "angle": 0, + "content": "Table 1. Chained shape editing ablation. We report baseline results for our proposed chained shape editing task, averaged across all chain lengths \\( |\\mathcal{P}| \\in \\{10, 15, 20\\} \\). Both the average final and average edit error are reported for the Chamfer Distance (CD) and \\( \\mathcal{L}_2 \\) distance \\( (\\mathcal{L}_2) \\) metrics." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.407, + 0.451, + 0.619 + ], + "angle": 0, + "content": "
ParameterAccuracy
seat height1.000
backrest curvature1.000
object width/height/depth0.978
seat roundness0.978
top bar thickness/height0.961
legs thickness0.956
legs bending/curvature0.944
adding/removing handle cushions0.750
number of legs/backrest rails0.663
legs roundness/indentation0.587
AVG (RANDOM)0.884
AVG (REALISTIC)0.969
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.713 + ], + "angle": 0, + "content": "Table 2. Parameter-wise accuracy for shape edit recognition. We report the accuracy of our edit detector for each parameter in the random set, alongside the average accuracy for the full random and realistic sets. Overall, we observe that our classifier is able to recognize edits corresponding to high change edits, but fails to differentiate between subtle changes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.307, + 0.743 + ], + "angle": 0, + "content": "5.2. Recognizing Shape Edits" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.858 + ], + "angle": 0, + "content": "Problem. To provide additional insight into the quality of our latent mapper and the difficulty of predicting specific edits, we train a binary classifier to discriminate from a pair of shapes \\((s_i, s_j)\\) on whether the corresponding edit instruction \\(\\mathbf{p}_{\\mathrm{ij}}\\) was applied to \\(s_i\\) or \\(s_j\\). We train a binary classifier \\(f_C: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\{0,1\\}\\), where \\(f_C(s_i, s_j) = 1\\) if the edit instruction \\(\\mathbf{p}_{\\mathrm{ij}}\\) was applied to \\(s_i\\), and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Method. Similarly to our latent mapper architecture for PC-AE, we use an MLP-based shape latent encoder to encode" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "the input shape latents into the bottleneck dimension, and project text features using a single linear layer to the same dimension. We then concatenate the two feature vectors and pass them through an MLP predictor outputting probabilities for the two classes \\( s_i \\) and \\( s_j \\). We train our classifier using a binary cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.895, + 0.271 + ], + "angle": 0, + "content": "Training and evaluation. We train our classifier on the realistic set and evaluate it on the full random shape set, which has a restricted vocabulary of edit instructions. We evaluate the accuracy of the classifier on the full set, as well as on subsets of the data corresponding to specific types of edits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.893, + 0.499 + ], + "angle": 0, + "content": "Results. In Table 2, we provide accuracy results for shape edit recognition, depending on the edit type. We also provide the global accuracy on the random and realistic sets. We observe that our classifier is able to recognize edits corresponding to shape parameters associated with a high degree of shape variation (like seat height and width/height of the shape), but struggles with parameters which are more subtle (like legs roundness, handle cushions). Our classifier is able to detect edits related to global width/height/depth adjustments with an accuracy of \\(97.8\\%\\). This accuracy drops to \\(94.4\\%\\) for legs bending/curvature, which is a more subtle change in the shape, and to \\(58.7\\%\\) for legs roundness/indentation. We also note that the model does not perform well on edits related to the number of parts, i.e. adding or removing a number \\(k\\) of legs or backrest rails." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.894, + 0.651 + ], + "angle": 0, + "content": "We impute these discrepancies to several factors. The PC-AE shape autoencoder we employ may be unable to properly reconstruct fine-grained details in the ground-truth and edited shapes, which renders the task of detecting subtle changes more difficult. Pointcloud resolution is also a limitation, as fine-grained parameters like leg indentation may not be properly sampled. Finally, learning a latent mapper and a classifier that can accurately count the number of parts to add or remove may also be a difficult task of its own, and require more domain-specific inductive biases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.62, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In conclusion, this paper introduces the ShapeWalk dataset, designed for advancing compositional shape editing guided by natural language instructions. The dataset comprises 158K unique shapes connected through 26K edit chains, synthesized from a realistic CAD-designed 3D dataset. Language instructions for applied edits are provided, alongside exact ground-truth edited shapes. Our method requires zero human annotation effort, and can be scaled up indefinitely. The usefulness of the dataset is demonstrated by training neural editor modules in the latent space of shape autoencoders [1]. Evaluation metrics inspired by trajectory prediction literature [4, 26] are introduced, offering a quantitative assessment of the quality of recursively edited shapes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22581" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. ShapeTalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961–971, 2016. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. In arXiv, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.47, + 0.456 + ], + "angle": 0, + "content": "[6] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.603, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.646, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[11] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics, 39(6):1-20, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.717, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[12] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: macro operation discovery for 3d shape programs. ACM Transactions on Graphics, 40(4), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics, 42(4), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Yuchen Li, Ujjwal Upadhyay, Habib Slim, Ahmed Abdelreheem, Arpit Prajapati, Suhail Pothigara, Peter Wonka, and Mohamed Elhoseiny. 3D CoMPaT: Composition of Materials on Parts of 3D Things. In ECCV, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.894, + 0.207 + ], + "angle": 0, + "content": "[16] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[17] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. In arXiv preprint arxiv:2212.11715, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.254, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[18] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In arXiv, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.311, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[19] Daniel Ritchie, Paul Guerrero, R. Kenny Jones, Niloy J. Mitra, Adriana Schulz, Karl D. D. Willis, and Jiajun Wu. Neurosymbolic models for computer graphics. Computer Graphics Forum, 42(2), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[20] Aditya Sanghi, Hang Chu, Joseph G. Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[21] Aditya Sanghi, Rao Fu, Vivian Liu, Karl D.D. Willis, Hooman Shayani, Amir H. Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.527, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[22] Habib Slim, Xiang Li, Mahmoud Ahmed Yuchen Li, Mohamed Ayman, Ujjwal Upadhyay Ahmed Abdelreheem, Suhail Pothigara Arpit Prajapati, Peter Wonka, and Mohamed Elhoseiny. 3DCoMPaT++: An improved large-scale 3d vision dataset for compositional recognition. In arXiv preprint arxiv:2212.11715, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.613, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[23] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. CoRR, abs/1711.00937, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[24] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[25] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. In SIGGRAPH, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.787, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Kaiping Xu, Zheng Qin, Guolong Wang, Kai Huang, Shuxiong Ye, and Huidi Zhang. Collision-Free LSTM for Human Trajectory Prediction, pages 106-116. 2018. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22582" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.148 + ], + "angle": 0, + "content": "[28] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics, 42(4):1-16, 2023. 6" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "22583" + } + ] +] \ No newline at end of file diff --git a/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_origin.pdf b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0cc3721cf535b7e7ac5535516a6d01382b1f6be9 --- /dev/null +++ b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/571b12bc-0855-45a3-8d18-3317ec9acf46_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c7393093351067a582a1912a04dc61b08773002af9ba0add9f88efa0512bbc1 +size 9750229 diff --git a/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/full.md b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/full.md new file mode 100644 index 0000000000000000000000000000000000000000..42fc203d2a19d324a7108b5aaece473307e2ab8d --- /dev/null +++ b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/full.md @@ -0,0 +1,279 @@ +# ShapeWalk: Compositional Shape Editing through Language-Guided Chains + +Habib Slim KAUST + +habib.smil@kaust.edu.sa + +Mohamed Elhoseiny KAUST + +mohamed.elhoseiny@kaust.edu.sa + +![](images/8d771d94127f775e4a915ab24a00efa8254a1261303d26dd0eed586958a7c2c0.jpg) +Figure 1. Visualizing shape chains from ShapeWalk. Our dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. We illustrate the interpolation process by coloring shapes based on their proximity to the starting shape (blue), and the ending shape (orange). Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. For each shape transition, we also provide a precise edit vector $\overline{\theta_{ij}}$ describing the parameter changes necessary to transition from one shape to the next. + +# Abstract + +Editing 3D shapes through natural language instructions is a challenging task that requires the comprehension of both language semantics and fine-grained geometric details. To bridge this gap, we introduce ShapeWalk, a carefully designed synthetic dataset designed to advance the field of language-guided shape editing. The dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. We synthesize edit chains by reconstructing and interpolating shapes sampled from a realistic CAD-designed 3D dataset in the parameter + +space of the GeoCode shape program. We leverage rule-based methods and language models to generate accurate and realistic natural language prompts corresponding to each edit. To illustrate the practicality of our contribution, we train neural editor modules in the latent space of shape autoencoders, and demonstrate the ability of our dataset to enable a variety of language-guided shape edits. Finally, we introduce multi-step editing metrics to benchmark the capacity of our models to perform recursive shape edits. We hope that our work will enable further study of compositional language-guided shape editing, and finds application in 3D CAD design and interactive modeling. + +![](images/97c4c99fd451ec83127abeaac8f74e0458fe56c6de1e3219674814d51eea4fbc.jpg) +Figure 2. Comparing ShapeTalk with our work. We compare the ShapeTalk [3] dataset (top) with our work (bottom). For an equivalent edit instruction (in green), ShapeTalk provides pairs of shapes with many factors of variation, while we generate synthetic pairs of shapes with a single clear varying factor. + +# 1. Introduction + +Whether in the realms of computer-aided design, virtual reality environments, or digital content creation, the process of refining and enhancing 3D visual data often involves intricate adjustments to geometric shapes, textures, and lighting. Furthermore, the necessity for precise modifications adds an additional layer of complexity, as inaccuracies can have profound implications on the final output. This labor-intensive nature of 3D data editing not only hinders workflow efficiency but also poses a barrier for individuals without specialized skills, limiting the democratization of 3D content manipulation. As such, there exists a compelling need for innovative solutions that streamline and democratize the 3D editing process, making it more accessible to a broader range of users. But how to train models with the ability to perform these complex edits? + +To that end, we introduce ShapeWalk, a dataset aiming at addressing the challenges inherent in the intricate and skill-intensive nature of 3D data editing. Our dataset consists of synthesized chains of 3D shapes, connected through edit vectors associated with precise language instructions describing the applied edits. It contains 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Our generation method is scalable and can yield an indefinite number of realistic shape chains, and extended to any 3D domain and shape program. To our knowledge, our dataset is the first to provide a large-scale collection of realistic 3D shape edit chains with precise language instructions, and matching ground-truth edited shapes. Our method yields specific edits undiluted by other factors of variation, as illustrated in Figure 2. + +To validate the usefulness of our dataset, we train neural editor modules in the latent space of shape autoencoders [1], and demonstrate the ability of our dataset to train models capable of performing a variety of shape edits. We learn + +latent mappings in the space of a frozen shape autoencoder, and show that our method can be applied on shapes unseen during the training of the neural editor. Taking advantage of the ground-truth shape chains provided by our dataset, we introduce multi-step editing metrics inspired from the field of trajectory prediction [4, 26] to evaluate the quality of our models. + +Our contributions can be summarized as follows: + +- Chained Shape-Editing Dataset: We introduce a synthetic dataset of chained edits, with associated language instructions and ground-truth edited shapes. This is the first dataset of its kind, and is designed to facilitate the study of compositional shape editing. +- Neural Shape Editor: We introduce a neural editor module trained on our dataset, and demonstrate its ability to perform a variety of shape edits. We show that our method can be applied on shapes unseen during the training of the neural editor. +- Chained-Editing Metrics: With the introduction of ground-truth shape chains, we propose multi-step editing metrics to evaluate the ability of neural editors to perform recursively applied edits. + +# 2. Related work + +3D Shape Programs. A significant body of work [9, 11, 12, 14, 19] is dedicated to exploring the use of procedural programs for 3D shape representation. Representing 3D objects as visual programs has many advantages. Representing shapes as visual primitives enhances interpretability [11, 14], as programs are by definition human-readable and thus understandable by human experts. Shape programs are more compact [9, 27] than their usual 3D shape modalities, and can be used to represent shapes more efficiently. Programs can also be composed to create shapes [11, 14], or decomposed into smaller programs. Ideally, shape programs can be edited [13] to modify the underlying geometry they represent. + +Our dataset consists of a collection of synthetic shapes created utilizing a backbone mesh-generating shape program. Specifically, we build from GeoCode [17], a 3D shape synthesis technique addressing the challenge of mapping high-fidelity geometry to an editable parameter space. GeoCode introduces a procedural program enabling the generation of high-quality mesh outputs with a balanced blend of interpretability and fine control. + +Language-Guided 3D Shape Editing. Various works have explored the use of natural language instructions to guide 3D shape editing. Recent efforts [16, 20, 21] propose to leverage pre-trained CLIP [24] models to align + +3D shapes with a given text prompt. Early efforts in that direction [16] included optimizing meshes to progressively align them with CLIP embeddings, leading to important computational overhead. CLIP-Sculptor [21], a more recent work, leverages a voxelized representation and a discrete latent space conditioned on CLIP's image-text embeddings to perform fast and fidel shape edits without shape optimization. + +Shape Editing Datasets. Other works leverage large text-aligned 3D shape datasets and require significant manual annotation effort. ShapeCrafter [8] generates 3D shapes incrementally from text using a neural network, evolving with additional phrases. ShapeCrafter is designed for recursive shape edition and utilizes a VQ-VAE [23] model to represent shapes as discrete codes. This method exhibits consistent shape-text alignment with gradual evolution. ChangeIt3D [3] introduces ShapeTalk, a large dataset for describing 3D shape differences. The framework facilitates language-based editing of 3D models without requiring 2D to 3D conversion methods, and learns a shape editing model by learning contrasts between sampled shape pairs. + +ShapeTalk [3] is the most similar work to our proposed dataset, and is the only publicly available dataset specifically designed for language-guided 3D shape-to-shape editing. ShapeTalk is a remarkable contribution in the field of language-guided shape editing, and is one of the first works to leverage large-scale 3D shape datasets to facilitate the study of language-guided shape editing. However, this work has some limitations that we attempt to address in our work. Collecting a dataset of shape differences is a challenging task requiring considerable manual annotation effort, and gathering a large number of 3D shapes may not be feasible for all domains. Furthermore, ShapeTalk is composed of edit contexts (i.e. shape pairs and edit instructions) which are hard to separate into fine-grained, composable edits. Most of the time, singular shape edits do not have an exact ground-truth in the form of a source and target shape pair. We illustrate this important distinction in Figure 2. The availability of singular shape edits with an associated ground-truth would ease benchmarking the quality of shape editing methods, and facilitate the training process of editing models. ShapeTalk also does not provide a mechanism to generate edit chains, which are necessary to study compositional shape editing. Finally, ShapeTalk does not easily enable the study of shape editing in a compositional manner. In contrast, our dataset is 1) synthesized by augmenting a small set of diverse shapes and can easily be scaled up, 2) composed of fine-grained edits and coarse-grained edits with exact ground-truths, and 3) separated into edit chains, which are designed to facilitate the study of compositional shape editing. + +# 3. ShapeWalk + +Our dataset contains 158K shapes split into a random and realistic set. With each edge connecting two consecutive shapes, we also produce a text instruction generated using the parameter changes necessary to transition from one shape to the next. We detail the generation method of our dataset here, and summarize the process in Figure 3. + +# 3.1. Dataset + +Definition. Our dataset can be defined as a collection of directed graph paths (dipaths), denoted as $\mathcal{P}^{(k)} = (S^{(k)},E^{(k)},f^{(k)})$ . For each shape chain $\mathcal{P}^{(k)}$ of length $l$ , we denote: + +- $S^{(k)} = \{s_{\theta_1}^{(k)}, \ldots, s_{\theta_N}^{(k)}\}$ , the set of distinct shape nodes composing the chain. Each shape is defined by a set of parameters $\theta_i \subseteq \Theta$ , where $\Theta$ is the linear parameter space of our shape program. +- $E^{(k)} \subseteq \{(s_{\theta_{\mathbf{i}}}^{(k)}, s_{\theta_{\mathbf{j}}}^{(k)}) \mid s_{\theta_{\mathbf{i}}}^{(k)}, s_{\theta_{\mathbf{j}}}^{(k)} \in V^{(k)}, j = i + 1\}$ , the set of edges linking each consecutive pair of shapes. +- $f^{(k)}: E^{(k)} \mapsto \{(\overline{\theta_{ij}}, \mathbf{p_{ij}}) \mid \overline{\theta_{ij}} \subseteq \Theta, \mathbf{p_{ij}} \in \Sigma \}$ , a function mapping each edge to a vector $\overline{\theta_{ij}} \subseteq \Theta$ and a set of text instructions $\mathbf{p_{ij}}$ . $\overline{\theta_{ij}}$ defines the parameter changes, or edits necessary, to go from shape $i$ to $j$ . $\mathbf{p_{ij}}$ is a set of text instructions describing this edit in natural language. + +Generation. To generate our dataset, we start by reconstructing a set $S \subseteq S$ of realistic 3D CAD shapes into the space of shapes covered by our shape program $S_{\Theta} \subset S$ . To that end, we utilize shapes from the 3DCOMPAT++ [15, 22] dataset, a realistic, industry-based 3D CAD dataset. We employ this dataset to avoid overlap with ShapeNet [5], and to ensure diversity and visual quality of the shapes. + +We define a visual similarity function $d: \mathcal{S} \times \mathcal{S} \mapsto \mathbb{R}$ between two shapes $s_i, s_j \in S$ as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder $\phi_R: \mathbb{R}^{h \times w \times 3} \mapsto \mathbb{R}^d$ . + +$$ +d \left(s _ {i}, s _ {j}\right) = \left\| \phi_ {R} \left(f _ {R} \left(s _ {i}\right)\right) - \phi_ {R} \left(f _ {R} \left(s _ {j}\right)\right) \right\| _ {2} +$$ + +Where $f_{R}: S \mapsto \mathbb{R}^{3 \times h \times w}$ is a rendering function. The set of reconstructed shapes can then be defined as: + +$$ +\operatorname *{arg min}_{\substack{S_{R}\subseteq S\\ |S_{R}| = (1 - \alpha)|S|}}\sum_{s\in S_{R}}d(s,\widehat{s}) +$$ + +Where $\widehat{s}$ is the reconstructed shape, defined as: + +$$ +\widehat {s} = f _ {\Theta} \circ \phi_ {\Theta} (s) +$$ + +![](images/0de3b1f5fee953afb7187fafed52c153142556b1a000eebdb57aaa90f4ec887f.jpg) +Figure 3. Detailling the shape generation pipeline of ShapeWalk. Realistic 3D CAD shapes are reconstructed from the 3DCOMPAT++ [22] dataset into the GeoCode [17] shape program parameter space, using a mapping function $\phi_{\Theta}$ . Reconstructed shapes with an error over a fixed threshold are discarded using a visual similarity function $d$ , which is based on rendered feature similarity. Filtered pairs of shapes are then interpolated in the parameter space of the shape program, to generate shape chains $\mathcal{P}^{(k)}$ . + +With $f_{\Theta} : \Theta \mapsto S$ the shape program mapping shape parameters to a 3D mesh, and $\phi_{\Theta} : S \mapsto \Theta$ an encoder based on DGCNN [17, 25] which samples points from a shape surface and regresses corresponding shape parameters in $\Theta$ . + +We reconstruct shapes by first sampling pointclouds from the surface of the original meshes, and then feeding them into the DGCNN [25] encoder fine-tuned [17] to regress the parameters of the shape program. We discard a ratio of $\alpha = 0.08$ of the shapes with the highest reconstruction error, which we measure as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder. This threshold is selected empirically by visually inspecting the shapes in the upper part of the reconstruction error distribution. In our dataset instance building on 3DComPAT++ shapes, we filter out a total of 89 shapes during this process, leading to a set of $|S_R| = 1113$ reconstructed shapes. This process can be scaled up to any dataset size and any shape domain, as long as a corresponding shape program is available. + +Shape interpolation. After reconstructing a set $S_R$ of realistic shapes, we generate a set of shape chains $\mathcal{P}^{(k)}$ by interpolating between the parameters of shape pairs. + +To that end, we first sample a set of shape pairs $(s_{\theta_{\mathrm{i}}}, s_{\theta_{\mathrm{j}}}) \in S_{P} \subseteq S_{R} \times S_{R}$ , and partition them into $L = 10$ levels of proximity. These proximity levels are based on the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder. + +We define the set of shape pairs $S_P^{(l)}$ at level $l \in [[1, L]]$ as: + +$$ +\mathcal {S} _ {P} ^ {(l)} = \{(s _ {\theta_ {\mathrm {i}}}, s _ {\theta_ {\mathrm {j}}}) \in \mathcal {S} _ {P} | \overline {{d (s _ {\theta_ {\mathrm {i}}} , s _ {\theta_ {\mathrm {j}}})}} \in \left[ \frac {l - 1}{L}, \frac {l}{L} \right] \} +$$ + +Where $\overline{d}$ is a visual similarity function, min-max normalized across all pairs of shapes. + +For each pair $(s_{\theta_i}, s_{\theta_j}) \in S_P^{(l)}$ , we then interpolate between the shape parameters $\theta_i$ and $\theta_j$ to generate a set of intermediate shape parameters $(\theta_i, \dots, \theta_{i + (N - 2)}, \theta_j) \in \Theta^N$ . The number of intermediate shapes $N$ is defined as the number of differing parameters between $\theta_i$ and $\theta_j$ , that is: + +$$ +N = \left| \left\{k \mid \theta_ {\mathbf {i}} [ k ] \neq \theta_ {\mathbf {j}} [ k ] \right\} \right| +$$ + +When the proximity level is low, edits necessary to transition from one shape to another are larger in intensity as the shapes are more different, and smaller when $l$ is smaller. The ordering of intermediate edits is randomly sampled using a dependency-aware algorithm (for example, parameters relating to armrest height or width are sampled after the addition of armrests). + +As illustrated in Figure 1, this generation process leads to plausible interpolated shapes, and to fine-grained and realistic sequences of edits with detailed metadata. + +We summarize the process of generating realistic shape chains in Figure 3. + +Random set. While the realistic set is based on reconstructed shapes from the 3DCOMPAT++ [15, 22] dataset, the random set is based on a collection of random shapes generated using the shape program. This alternative set aims at covering a large space of the parameter space of the GeoCode shape program. This subset is generated by systematically sampling various combinations of parameters within defined ranges determined by a minimum edit intensity, but does not necessarily result in realistic shapes. + +# 3.2. Text Instructions + +Rule-based generation. Given an edit vector $\overline{\theta_{ij}}$ describing the parameter changes necessary to transition from shape $i$ to $j$ , we generate a set of text instructions $\mathbf{p}_{ij}$ describing this edit in natural language. We first map each parameter $\theta_{i}[k]$ composing the edit vector to a natural language name, and map the magnitude of the + +![](images/8f747f3fb70df19623f37cb0c8fd5e2a1d25ecf62b3e6fedd4a46afcb2766a31.jpg) +Figure 4. Dataset chains samples. We feature in this figure shape chains from our dataset with associated generated text instructions. Each shape chain is composed of a sequence of shapes, with each consecutive pair of shapes associated to a set of text instructions describing the edits necessary to transition from one shape to the next. Overall, generated text instructions are accurate and provide a detailed description of the edits applied to the shapes. We include in the supplementary material additional samples of generated shape chains, alongside shape chains generated for two additional table and vase categories. + +parameter change to a natural language intensity depending on the parameter type. Boolean parameters are also described appropriately, for example by using the word add or remove when describing the addition or removal of a part. + +Augmentation. This rule-based instruction generation leads to perfectly accurate instructions, but may lack diversity. To alleviate this issue, we 1) randomly sample adjectives and adverbs to characterize the intensity of edits, 2) randomly invert parameter names and magnitude directions (e.g. "increase armrests straightness" will be augmented to "decrease armrests bend" for the same edit). As a final augmentation, we utilize a T5 [18] transformer model finetuned for paraphrasing to generate additional instructions for each edit vector. We use the Parrot library [6] to filter generated paraphrases based on a fluency and adequacy score. + +We summarize the process of generating text instructions in the supplementary material. + +# 4. Neural Shape Editing + +# 4.1. Problem + +We are concerned with the task of editing a shape $s_i$ given a sequence of natural language edit instructions. We want to learn an editing function $f_E: S \times \Sigma \mapsto S$ , where $S$ is the space of input shapes, and $\Sigma$ is the space of natural language edit instructions, able to compose edits from a starting shape $s_1$ to an ending shape $s_N$ , given a sequence of edit instructions $\{\mathbf{p_t}\}_{t=1}^N$ . + +$$ +\widehat {s _ {N}} = f _ {E} (\phi_ {T} (\mathbf {p _ {1 2}}), f _ {E} (\phi_ {T} (\mathbf {p _ {2 3}}), \ldots f _ {E} (\phi_ {T} (\mathbf {p _ {k N}}), s _ {1}))) +$$ + +Where $N = |\mathcal{P}|$ is the length of the edit chain, and $\phi_T: \Sigma \mapsto \mathbb{R}^{D_T}$ is a text encoder. At the end of the edit sequence, we want to recover a shape $\widehat{s_N}$ as close as possible to the ground-truth shape $s_N$ . Our proposed dataset provides a ground-truth for every intermediate shape $s_k$ , which we leverage to train and evaluate our method. + +![](images/1669385cc877170fbfb73361aecfca58418d7e9a4a3fdbb0984ef34ad3312bb3.jpg) + +![](images/48027af70a79e59ef61ea06eed686f69fc495704f45fda8d2ab7eff75e458833.jpg) +Figure 5. Visualizing synthesized pairs of edits using our latent editors. We illustrate the diverse range of edits generated by our proposed latent mapper, showcasing its ability to capture nuanced variations in the input data, for 3DS2VS latents (top two rows) and PC-AE latents (bottom two rows). Each pair demonstrates the original input on the left and the corresponding synthesized output on the right, with the associated caption. We abbreviate instructions which are too long to fit in the figure, and indicate when text is omitted with an ellipsis. For more subtle changes, we circle the areas of interest in both source and target shapes. Overall, edits are consistent with the input instruction, preserve shape identity while generating plausible output shapes. + +# 4.2. Objective + +One of the main advantages of training a shape editing model with ShapeWalk compared to other works [2, 8] is the availability of exact ground-truth edited shapes for each edit instruction (see Section 2). We thus formulate the objective function of our learning problem as a simple $\mathcal{L}_2$ loss between the latents of the synthesized shape and the ground-truth shape. For all pairs of shapes $(s_i,s_j)$ and their corresponding edit instructions $\mathbf{p_{ij}}$ , we minimize the following loss: + +$$ +\mathcal {L} = \left\| f _ {E} \left(\phi_ {E} \left(s _ {i}\right), \mathbf {p} _ {\mathbf {i j}}\right) - \phi_ {E} \left(s _ {j}\right) \right\| _ {2} +$$ + +Where $\phi_{AE} = \phi_E\circ \phi_D$ is a shape autoencoder, and $\phi_E:$ $S\mapsto \mathbb{R}^{D_E}$ denotes its encoder component. + +# 4.3. Models + +For $\phi_{AE}$ , we experiment with PC-AE [1] and 3D2VS [28] models both pre-trained on ShapeNet [5]. Note that our method is agnostic to the choice of shape autoencoder and can be adapted to a variety of shape representations. For $\phi_T$ , we use a pretrained BERT [7] model. + +Both $\phi_{AE}$ and $\phi_T$ are frozen during training, and we only train the parameters of our latent mapper $f_{E}$ . + +For PC-AE, we formulate our latent mapper as a neural module which for each tuple $(\phi_E(s_i), \mathbf{p_{ij}}, \phi_E(s_j))$ predicts an edit vector $\widehat{\theta_{ij}} \in \mathbb{R}^{D_E}$ in the feature space of $\phi_E$ , and adds it to the latent of the input shape $s_i$ to generate the latent of the output shape $s_j$ : + +$$ +f _ {E} \left(\phi_ {E} \left(s _ {i}\right), \mathbf {p} _ {\mathbf {i j}}\right) = \widehat {\theta_ {i j}} + \phi_ {E} \left(s _ {i}\right) +$$ + +We utilize two variants of our latent mapper $f_{E}$ : one in which the edit vector $\widehat{\theta_{ij}}$ is predicted directly, and one in which the edit vector $\widehat{\theta_{ij}}$ is predicted as a product of a normalized edit direction $\widehat{v_{ij}} \in \mathbb{R}^{D_E}$ and a magnitude $\widehat{m_{ij}} \in \mathbb{R}$ separately. + +For 3D2VS which encodes shapes as latents sets of higher resolutions, we use a transformer-based latent diffusion model instead to generate edited latents. We concatenate the input shape latent $\phi_E(s_i)$ with the noised edited shape latent $\phi_E(s_j + \epsilon_t)$ and feed the BERT text embeddings to the cross-attention layers of the transformer blocks to predict the added noise. We illustrate our architecture for the latent mapper in the supplementary material. + +# 5. Experiments + +# 5.1. Chained Shape Editing + +Comparison Models. For PC-AE, we experiment with three variants of our latent mapper module: + +- DIRECTGEN directly predicts the edited shape latent without regressing an edit vector. +- LATEFUSION follows the network proposed in [3], and first passes the shape embeddings through an encoder. Encoded shape features are then concatenated with text embeddings and fed into a neural module which finally predicts the edit vector. +- Ours is a simple multi-layer perceptron (MLP) which directly takes the context (shape and text embeddings) as input to predict an edit vector. + +For all of these variants, we experiment with various bottleneck dimensions and number of layers (in subscript). + +We further decompose these variants into coupled and decoupled versions, where we predict a normalized edit direction $\widehat{v_{ij}}\in \mathbb{R}^{D_E}$ and a magnitude $\widehat{m_{ij}}\in \mathbb{R}$ separately. We provide in appendix the full architecture details and training hyperparameters employed to train our models. + +Note that we do not compare with ChangeIt3D [2] as their neural-listener distillation method is not applicable in our context: we train our models with a direct feedback signal from the ground-truth edited shapes. + +Single-step metrics. We propose to measure the quality of our editing steps both before and after shape reconstruction. Since we experiment on point clouds, we use the scaled Chamfer Distance between the reconstructed original shape $s_i$ and the reconstructed edited shape $\widehat{s_j}$ : + +$$ +d _ {\mathrm {C D}} = \frac {1}{| \widehat {s _ {j}} |} \sum_ {x \in \widehat {s _ {j}}} \min _ {y \in s _ {i}} \| x - y \| _ {2} + \frac {1}{| s _ {i} |} \sum_ {y \in s _ {i}} \min _ {x \in \widehat {s _ {j}}} \| x - y \| _ {2} +$$ + +We also use the $\mathcal{L}_2$ distance between the latents of the original shape $s_i$ and the edited shape $s_j$ : + +$$ +d _ {\mathcal {L} _ {2}} = \left\| \phi_ {E} (s _ {i}) - \phi_ {E} (s _ {j}) \right\| _ {2} +$$ + +Multi-step metrics. To extend the single-step metrics to chained shape generation, we take inspiration from the trajectory prediction literature [4, 26] and propose appropriate metrics for our task. + +Given a shape distance function $d: \mathcal{S} \times \mathcal{S} \mapsto \mathbb{R}$ , a chain of ground-truth reconstructed shapes $\{s_1, \ldots, s_N\}$ , and a set of recursively generated shapes $\{\widehat{s}_1, \ldots, \widehat{s_N}\}$ , we define the average edit error as the average distance between the generated shapes and the corresponding ground-truth shapes: + +$$ +\mathbf {A} _ {d} = \frac {1}{N} \sum_ {t = 1} ^ {N} d \left(s _ {t}, \widehat {s} _ {t}\right) +$$ + +And the final edit error as the distance between the last shape in the chain and the corresponding ground-truth shape: + +$$ +\mathrm {F} _ {d} = d \left(s _ {N}, \widehat {s _ {N}}\right) +$$ + +We report results with both metrics for $d = d_{\mathrm{CD}}$ and $d = d_{\mathcal{L}_2}$ . + +Results. We provide qualitative results of edited pairs generated by our latent mapper in Figure 5. Overall, we observe that our method is able to generate plausible edits that are consistent with the input instruction, and preserve the identity of the input shape. However, our method is limited by the quality of the shape autoencoder. We notice that the PC-AE editor is unable to perform very fine-grained edits which are not properly captured by the autoencoder, while visible in the ground-truth shapes. We explore the possible causes of these limitations in Section 5.2. + +We provide quantitative results for our chained editing metrics averaged across $|\mathcal{P}| \in \{10, 15, 20\}$ in Table 1, and detail per-chain length results in the appendix. Directly predicting the edited shape latent (DIRECTGEN) leads to poor edit predictions across all metrics, although increasing the number of layers may improve the results. Overall, decoupling the edit direction and magnitude leads to better results. Our coupled $\mathrm{OURS}_{512 \times 4}$ variant performs better than the coupled $\mathrm{LATEFUSION}_{512}$ variant on the Chamfer Distance, but worse on $\mathcal{L}_2$ distance. + +
Modeldecoupled?Averaged ∀|P|
FCD×1e4ACD×1e4FL2AL2
LATEFUSION1024×2.8562.6211.4441.251
LATEFUSION512×2.7192.6091.4621.290
LATEFUSION256×2.8742.6511.5091.283
OURS512×8×3.2082.8221.7031.437
OURS512×4×2.9902.7031.5891.351
LATEFUSION1024×2.7112.5681.4051.245
LATEFUSION512×3.0022.7081.4511.254
LATEFUSION256×3.3092.8481.5521.324
OURS512×8×2.7822.5841.4971.290
OURS512×4×2.6702.5241.4471.266
+ +Table 1. Chained shape editing ablation. We report baseline results for our proposed chained shape editing task, averaged across all chain lengths $|\mathcal{P}| \in \{10, 15, 20\}$ . Both the average final and average edit error are reported for the Chamfer Distance (CD) and $\mathcal{L}_2$ distance $(\mathcal{L}_2)$ metrics. + +
ParameterAccuracy
seat height1.000
backrest curvature1.000
object width/height/depth0.978
seat roundness0.978
top bar thickness/height0.961
legs thickness0.956
legs bending/curvature0.944
adding/removing handle cushions0.750
number of legs/backrest rails0.663
legs roundness/indentation0.587
AVG (RANDOM)0.884
AVG (REALISTIC)0.969
+ +Table 2. Parameter-wise accuracy for shape edit recognition. We report the accuracy of our edit detector for each parameter in the random set, alongside the average accuracy for the full random and realistic sets. Overall, we observe that our classifier is able to recognize edits corresponding to high change edits, but fails to differentiate between subtle changes. + +# 5.2. Recognizing Shape Edits + +Problem. To provide additional insight into the quality of our latent mapper and the difficulty of predicting specific edits, we train a binary classifier to discriminate from a pair of shapes $(s_i, s_j)$ on whether the corresponding edit instruction $\mathbf{p}_{\mathrm{ij}}$ was applied to $s_i$ or $s_j$ . We train a binary classifier $f_C: \mathcal{S} \times \mathcal{S} \mapsto \{0,1\}$ , where $f_C(s_i, s_j) = 1$ if the edit instruction $\mathbf{p}_{\mathrm{ij}}$ was applied to $s_i$ , and 0 otherwise. + +Method. Similarly to our latent mapper architecture for PC-AE, we use an MLP-based shape latent encoder to encode + +the input shape latents into the bottleneck dimension, and project text features using a single linear layer to the same dimension. We then concatenate the two feature vectors and pass them through an MLP predictor outputting probabilities for the two classes $s_i$ and $s_j$ . We train our classifier using a binary cross-entropy loss. + +Training and evaluation. We train our classifier on the realistic set and evaluate it on the full random shape set, which has a restricted vocabulary of edit instructions. We evaluate the accuracy of the classifier on the full set, as well as on subsets of the data corresponding to specific types of edits. + +Results. In Table 2, we provide accuracy results for shape edit recognition, depending on the edit type. We also provide the global accuracy on the random and realistic sets. We observe that our classifier is able to recognize edits corresponding to shape parameters associated with a high degree of shape variation (like seat height and width/height of the shape), but struggles with parameters which are more subtle (like legs roundness, handle cushions). Our classifier is able to detect edits related to global width/height/depth adjustments with an accuracy of $97.8\%$ . This accuracy drops to $94.4\%$ for legs bending/curvature, which is a more subtle change in the shape, and to $58.7\%$ for legs roundness/indentation. We also note that the model does not perform well on edits related to the number of parts, i.e. adding or removing a number $k$ of legs or backrest rails. + +We impute these discrepancies to several factors. The PC-AE shape autoencoder we employ may be unable to properly reconstruct fine-grained details in the ground-truth and edited shapes, which renders the task of detecting subtle changes more difficult. Pointcloud resolution is also a limitation, as fine-grained parameters like leg indentation may not be properly sampled. Finally, learning a latent mapper and a classifier that can accurately count the number of parts to add or remove may also be a difficult task of its own, and require more domain-specific inductive biases. + +# 6. Conclusion + +In conclusion, this paper introduces the ShapeWalk dataset, designed for advancing compositional shape editing guided by natural language instructions. The dataset comprises 158K unique shapes connected through 26K edit chains, synthesized from a realistic CAD-designed 3D dataset. Language instructions for applied edits are provided, alongside exact ground-truth edited shapes. Our method requires zero human annotation effort, and can be scaled up indefinitely. The usefulness of the dataset is demonstrated by training neural editor modules in the latent space of shape autoencoders [1]. Evaluation metrics inspired by trajectory prediction literature [4, 26] are introduced, offering a quantitative assessment of the quality of recursively edited shapes. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2, 6, 8 +[2] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 6, 7 +[3] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. ShapeTalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 2, 3, 7 +[4] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961–971, 2016. 2, 7, 8 +[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. In arXiv, 2015. 3, 6 +[6] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 5 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 6 +[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2023. 3, 6 +[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 4 +[11] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics, 39(6):1-20, 2020. 2 +[12] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: macro operation discovery for 3d shape programs. ACM Transactions on Graphics, 40(4), 2021. 2 +[13] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2 +[14] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics, 42(4), 2023. 2 + +[15] Yuchen Li, Ujjwal Upadhyay, Habib Slim, Ahmed Abdelreheem, Arpit Prajapati, Suhail Pothigara, Peter Wonka, and Mohamed Elhoseiny. 3D CoMPaT: Composition of Materials on Parts of 3D Things. In ECCV, 2022. 3, 4 +[16] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3 +[17] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. In arXiv preprint arxiv:2212.11715, 2022. 2, 4 +[18] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In arXiv, 2023. 5 +[19] Daniel Ritchie, Paul Guerrero, R. Kenny Jones, Niloy J. Mitra, Adriana Schulz, Karl D. D. Willis, and Jiajun Wu. Neurosymbolic models for computer graphics. Computer Graphics Forum, 42(2), 2023. 2 +[20] Aditya Sanghi, Hang Chu, Joseph G. Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2 +[21] Aditya Sanghi, Rao Fu, Vivian Liu, Karl D.D. Willis, Hooman Shayani, Amir H. Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 2, 3 +[22] Habib Slim, Xiang Li, Mahmoud Ahmed Yuchen Li, Mohamed Ayman, Ujjwal Upadhyay Ahmed Abdelreheem, Suhail Pothigara Arpit Prajapati, Peter Wonka, and Mohamed Elhoseiny. 3DCoMPaT++: An improved large-scale 3d vision dataset for compositional recognition. In arXiv preprint arxiv:2212.11715, 2023. 3, 4 +[23] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. CoRR, abs/1711.00937, 2017. 3 +[24] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2 +[25] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. In SIGGRAPH, 2019. 4 +[26] Kaiping Xu, Zheng Qin, Guolong Wang, Kai Huang, Shuxiong Ye, and Huidi Zhang. Collision-Free LSTM for Human Trajectory Prediction, pages 106-116. 2018. 2, 7, 8 +[27] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2 + +[28] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics, 42(4):1-16, 2023. 6 \ No newline at end of file diff --git a/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/images.zip b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cb237f82aa4f763813e42ddc488d6ea42b7bccd1 --- /dev/null +++ b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b18826915b083e0cd0b1bc6959b949f3f4476e5337b70124c1c24e1cfdab1996 +size 460993 diff --git a/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/layout.json b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe281151a3ecb57172cbc7a5373f9ee28219752b --- /dev/null +++ b/2024/ShapeWalk_ Compositional Shape Editing Through Language-Guided Chains/layout.json @@ -0,0 +1,7860 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "ShapeWalk: Compositional Shape Editing through Language-Guided Chains" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 144, + 224, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 144, + 224, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 144, + 224, + 171 + ], + "type": "text", + "content": "Habib Slim KAUST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 259, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 259, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 259, + 184 + ], + "type": "text", + "content": "habib.smil@kaust.edu.sa" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 144, + 428, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 144, + 428, + 171 + ], + "spans": [ + { + "bbox": [ + 326, + 144, + 428, + 171 + ], + "type": "text", + "content": "Mohamed Elhoseiny KAUST" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 293, + 174, + 457, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 174, + 457, + 185 + ], + "spans": [ + { + "bbox": [ + 293, + 174, + 457, + 185 + ], + "type": "text", + "content": "mohamed.elhoseiny@kaust.edu.sa" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 217, + 526, + 473 + ], + "blocks": [ + { + "bbox": [ + 57, + 217, + 526, + 473 + ], + "lines": [ + { + "bbox": [ + 57, + 217, + 526, + 473 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 526, + 473 + ], + "type": "image", + "image_path": "8d771d94127f775e4a915ab24a00efa8254a1261303d26dd0eed586958a7c2c0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 479, + 548, + 542 + ], + "lines": [ + { + "bbox": [ + 45, + 479, + 548, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 479, + 548, + 542 + ], + "type": "text", + "content": "Figure 1. Visualizing shape chains from ShapeWalk. Our dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. We illustrate the interpolation process by coloring shapes based on their proximity to the starting shape (blue), and the ending shape (orange). Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. For each shape transition, we also provide a precise edit vector " + }, + { + "bbox": [ + 45, + 479, + 548, + 542 + ], + "type": "inline_equation", + "content": "\\overline{\\theta_{ij}}" + }, + { + "bbox": [ + 45, + 479, + 548, + 542 + ], + "type": "text", + "content": " describing the parameter changes necessary to transition from one shape to the next." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 550, + 192, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 550, + 192, + 562 + ], + "spans": [ + { + "bbox": [ + 143, + 550, + 192, + 562 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 570, + 289, + 715 + ], + "type": "text", + "content": "Editing 3D shapes through natural language instructions is a challenging task that requires the comprehension of both language semantics and fine-grained geometric details. To bridge this gap, we introduce ShapeWalk, a carefully designed synthetic dataset designed to advance the field of language-guided shape editing. The dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Each consecutive pair of shapes is associated with precise language instructions describing the applied edits. We synthesize edit chains by reconstructing and interpolating shapes sampled from a realistic CAD-designed 3D dataset in the parameter" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 552, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 696 + ], + "type": "text", + "content": "space of the GeoCode shape program. We leverage rule-based methods and language models to generate accurate and realistic natural language prompts corresponding to each edit. To illustrate the practicality of our contribution, we train neural editor modules in the latent space of shape autoencoders, and demonstrate the ability of our dataset to enable a variety of language-guided shape edits. Finally, we introduce multi-step editing metrics to benchmark the capacity of our models to perform recursive shape edits. We hope that our work will enable further study of compositional language-guided shape editing, and finds application in 3D CAD design and interactive modeling." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "text", + "content": "22574" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 286, + 194 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 286, + 194 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 286, + 194 + ], + "type": "image", + "image_path": "97c4c99fd451ec83127abeaac8f74e0458fe56c6de1e3219674814d51eea4fbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 208, + 288, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 288, + 264 + ], + "type": "text", + "content": "Figure 2. Comparing ShapeTalk with our work. We compare the ShapeTalk [3] dataset (top) with our work (bottom). For an equivalent edit instruction (in green), ShapeTalk provides pairs of shapes with many factors of variation, while we generate synthetic pairs of shapes with a single clear varying factor." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 282, + 128, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 128, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 128, + 294 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 483 + ], + "type": "text", + "content": "Whether in the realms of computer-aided design, virtual reality environments, or digital content creation, the process of refining and enhancing 3D visual data often involves intricate adjustments to geometric shapes, textures, and lighting. Furthermore, the necessity for precise modifications adds an additional layer of complexity, as inaccuracies can have profound implications on the final output. This labor-intensive nature of 3D data editing not only hinders workflow efficiency but also poses a barrier for individuals without specialized skills, limiting the democratization of 3D content manipulation. As such, there exists a compelling need for innovative solutions that streamline and democratize the 3D editing process, making it more accessible to a broader range of users. But how to train models with the ability to perform these complex edits?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 484, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 664 + ], + "type": "text", + "content": "To that end, we introduce ShapeWalk, a dataset aiming at addressing the challenges inherent in the intricate and skill-intensive nature of 3D data editing. Our dataset consists of synthesized chains of 3D shapes, connected through edit vectors associated with precise language instructions describing the applied edits. It contains 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes. Our generation method is scalable and can yield an indefinite number of realistic shape chains, and extended to any 3D domain and shape program. To our knowledge, our dataset is the first to provide a large-scale collection of realistic 3D shape edit chains with precise language instructions, and matching ground-truth edited shapes. Our method yields specific edits undiluted by other factors of variation, as illustrated in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "To validate the usefulness of our dataset, we train neural editor modules in the latent space of shape autoencoders [1], and demonstrate the ability of our dataset to train models capable of performing a variety of shape edits. We learn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "content": "latent mappings in the space of a frozen shape autoencoder, and show that our method can be applied on shapes unseen during the training of the neural editor. Taking advantage of the ground-truth shape chains provided by our dataset, we introduce multi-step editing metrics inspired from the field of trajectory prediction [4, 26] to evaluate the quality of our models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 157, + 504, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 157, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 504, + 168 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 178, + 545, + 365 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 178, + 545, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 545, + 238 + ], + "type": "text", + "content": "- Chained Shape-Editing Dataset: We introduce a synthetic dataset of chained edits, with associated language instructions and ground-truth edited shapes. This is the first dataset of its kind, and is designed to facilitate the study of compositional shape editing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 247, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 545, + 306 + ], + "type": "text", + "content": "- Neural Shape Editor: We introduce a neural editor module trained on our dataset, and demonstrate its ability to perform a variety of shape edits. We show that our method can be applied on shapes unseen during the training of the neural editor." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 316, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 545, + 365 + ], + "type": "text", + "content": "- Chained-Editing Metrics: With the introduction of ground-truth shape chains, we propose multi-step editing metrics to evaluate the ability of neural editors to perform recursively applied edits." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 379, + 389, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 389, + 391 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 389, + 391 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 399, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 555 + ], + "type": "text", + "content": "3D Shape Programs. A significant body of work [9, 11, 12, 14, 19] is dedicated to exploring the use of procedural programs for 3D shape representation. Representing 3D objects as visual programs has many advantages. Representing shapes as visual primitives enhances interpretability [11, 14], as programs are by definition human-readable and thus understandable by human experts. Shape programs are more compact [9, 27] than their usual 3D shape modalities, and can be used to represent shapes more efficiently. Programs can also be composed to create shapes [11, 14], or decomposed into smaller programs. Ideally, shape programs can be edited [13] to modify the underlying geometry they represent." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "type": "text", + "content": "Our dataset consists of a collection of synthetic shapes created utilizing a backbone mesh-generating shape program. Specifically, we build from GeoCode [17], a 3D shape synthesis technique addressing the challenge of mapping high-fidelity geometry to an editable parameter space. GeoCode introduces a procedural program enabling the generation of high-quality mesh outputs with a balanced blend of interpretability and fine control." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Language-Guided 3D Shape Editing. Various works have explored the use of natural language instructions to guide 3D shape editing. Recent efforts [16, 20, 21] propose to leverage pre-trained CLIP [24] models to align" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22575" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "3D shapes with a given text prompt. Early efforts in that direction [16] included optimizing meshes to progressively align them with CLIP embeddings, leading to important computational overhead. CLIP-Sculptor [21], a more recent work, leverages a voxelized representation and a discrete latent space conditioned on CLIP's image-text embeddings to perform fast and fidel shape edits without shape optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 289, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 289, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 289, + 352 + ], + "type": "text", + "content": "Shape Editing Datasets. Other works leverage large text-aligned 3D shape datasets and require significant manual annotation effort. ShapeCrafter [8] generates 3D shapes incrementally from text using a neural network, evolving with additional phrases. ShapeCrafter is designed for recursive shape edition and utilizes a VQ-VAE [23] model to represent shapes as discrete codes. This method exhibits consistent shape-text alignment with gradual evolution. ChangeIt3D [3] introduces ShapeTalk, a large dataset for describing 3D shape differences. The framework facilitates language-based editing of 3D models without requiring 2D to 3D conversion methods, and learns a shape editing model by learning contrasts between sampled shape pairs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 354, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 715 + ], + "type": "text", + "content": "ShapeTalk [3] is the most similar work to our proposed dataset, and is the only publicly available dataset specifically designed for language-guided 3D shape-to-shape editing. ShapeTalk is a remarkable contribution in the field of language-guided shape editing, and is one of the first works to leverage large-scale 3D shape datasets to facilitate the study of language-guided shape editing. However, this work has some limitations that we attempt to address in our work. Collecting a dataset of shape differences is a challenging task requiring considerable manual annotation effort, and gathering a large number of 3D shapes may not be feasible for all domains. Furthermore, ShapeTalk is composed of edit contexts (i.e. shape pairs and edit instructions) which are hard to separate into fine-grained, composable edits. Most of the time, singular shape edits do not have an exact ground-truth in the form of a source and target shape pair. We illustrate this important distinction in Figure 2. The availability of singular shape edits with an associated ground-truth would ease benchmarking the quality of shape editing methods, and facilitate the training process of editing models. ShapeTalk also does not provide a mechanism to generate edit chains, which are necessary to study compositional shape editing. Finally, ShapeTalk does not easily enable the study of shape editing in a compositional manner. In contrast, our dataset is 1) synthesized by augmenting a small set of diverse shapes and can easily be scaled up, 2) composed of fine-grained edits and coarse-grained edits with exact ground-truths, and 3) separated into edit chains, which are designed to facilitate the study of compositional shape editing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 76, + 380, + 90 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 76, + 380, + 90 + ], + "spans": [ + { + "bbox": [ + 306, + 76, + 380, + 90 + ], + "type": "text", + "content": "3. ShapeWalk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 97, + 547, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 547, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 547, + 170 + ], + "type": "text", + "content": "Our dataset contains 158K shapes split into a random and realistic set. With each edge connecting two consecutive shapes, we also produce a text instruction generated using the parameter changes necessary to transition from one shape to the next. We detail the generation method of our dataset here, and summarize the process in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 178, + 365, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 365, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 365, + 190 + ], + "type": "text", + "content": "3.1. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "text", + "content": "Definition. Our dataset can be defined as a collection of directed graph paths (dipaths), denoted as " + }, + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(k)} = (S^{(k)},E^{(k)},f^{(k)})" + }, + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "text", + "content": ". For each shape chain " + }, + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(k)}" + }, + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 197, + 547, + 243 + ], + "type": "text", + "content": ", we denote:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 247, + 547, + 415 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "inline_equation", + "content": "S^{(k)} = \\{s_{\\theta_1}^{(k)}, \\ldots, s_{\\theta_N}^{(k)}\\}" + }, + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "text", + "content": ", the set of distinct shape nodes composing the chain. Each shape is defined by a set of parameters " + }, + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "inline_equation", + "content": "\\theta_i \\subseteq \\Theta" + }, + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 306, + 247, + 547, + 298 + ], + "type": "text", + "content": " is the linear parameter space of our shape program." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 306, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 545, + 334 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 306, + 306, + 545, + 334 + ], + "type": "inline_equation", + "content": "E^{(k)} \\subseteq \\{(s_{\\theta_{\\mathbf{i}}}^{(k)}, s_{\\theta_{\\mathbf{j}}}^{(k)}) \\mid s_{\\theta_{\\mathbf{i}}}^{(k)}, s_{\\theta_{\\mathbf{j}}}^{(k)} \\in V^{(k)}, j = i + 1\\}" + }, + { + "bbox": [ + 306, + 306, + 545, + 334 + ], + "type": "text", + "content": ", the set of edges linking each consecutive pair of shapes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "inline_equation", + "content": "f^{(k)}: E^{(k)} \\mapsto \\{(\\overline{\\theta_{ij}}, \\mathbf{p_{ij}}) \\mid \\overline{\\theta_{ij}} \\subseteq \\Theta, \\mathbf{p_{ij}} \\in \\Sigma \\}" + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": ", a function mapping each edge to a vector " + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\overline{\\theta_{ij}} \\subseteq \\Theta" + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": " and a set of text instructions " + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{p_{ij}}" + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\overline{\\theta_{ij}}" + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": " defines the parameter changes, or edits necessary, to go from shape " + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{p_{ij}}" + }, + { + "bbox": [ + 306, + 341, + 547, + 415 + ], + "type": "text", + "content": " is a set of text instructions describing this edit in natural language." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": "Generation. To generate our dataset, we start by reconstructing a set " + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "S \\subseteq S" + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": " of realistic 3D CAD shapes into the space of shapes covered by our shape program " + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "S_{\\Theta} \\subset S" + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": ". To that end, we utilize shapes from the 3DCOMPAT++ [15, 22] dataset, a realistic, industry-based 3D CAD dataset. We employ this dataset to avoid overlap with ShapeNet [5], and to ensure diversity and visual quality of the shapes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "text", + "content": "We define a visual similarity function " + }, + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "inline_equation", + "content": "d: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\mathbb{R}" + }, + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "text", + "content": " between two shapes " + }, + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "inline_equation", + "content": "s_i, s_j \\in S" + }, + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "text", + "content": " as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder " + }, + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\phi_R: \\mathbb{R}^{h \\times w \\times 3} \\mapsto \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 511, + 545, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 581, + 512, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 581, + 512, + 595 + ], + "spans": [ + { + "bbox": [ + 339, + 581, + 512, + 595 + ], + "type": "interline_equation", + "content": "d \\left(s _ {i}, s _ {j}\\right) = \\left\\| \\phi_ {R} \\left(f _ {R} \\left(s _ {i}\\right)\\right) - \\phi_ {R} \\left(f _ {R} \\left(s _ {j}\\right)\\right) \\right\\| _ {2}", + "image_path": "a56926076212070a21a0be3ea00d7b8469de065c24a2c49800d966e7bc48e4b3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "inline_equation", + "content": "f_{R}: S \\mapsto \\mathbb{R}^{3 \\times h \\times w}" + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "content": " is a rendering function. The set of reconstructed shapes can then be defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 638, + 479, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 638, + 479, + 669 + ], + "spans": [ + { + "bbox": [ + 373, + 638, + 479, + 669 + ], + "type": "interline_equation", + "content": "\\operatorname *{arg min}_{\\substack{S_{R}\\subseteq S\\\\ |S_{R}| = (1 - \\alpha)|S|}}\\sum_{s\\in S_{R}}d(s,\\widehat{s})", + "image_path": "27db30f3c3a78d7ab5a90b8513a6570ccf6256f1c0356b833b9df6af78c25e75.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 678, + 495, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 678, + 495, + 690 + ], + "spans": [ + { + "bbox": [ + 306, + 678, + 495, + 690 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 306, + 678, + 495, + 690 + ], + "type": "inline_equation", + "content": "\\widehat{s}" + }, + { + "bbox": [ + 306, + 678, + 495, + 690 + ], + "type": "text", + "content": " is the reconstructed shape, defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 392, + 700, + 459, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 700, + 459, + 714 + ], + "spans": [ + { + "bbox": [ + 392, + 700, + 459, + 714 + ], + "type": "interline_equation", + "content": "\\widehat {s} = f _ {\\Theta} \\circ \\phi_ {\\Theta} (s)", + "image_path": "5f0eb3d312e1429cb5a01286b5c106f4b30f1a27f91c597eafcd6bf613871706.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22576" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 544, + 182 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 544, + 182 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 544, + 182 + ], + "type": "image", + "image_path": "0de3b1f5fee953afb7187fafed52c153142556b1a000eebdb57aaa90f4ec887f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "type": "text", + "content": "Figure 3. Detailling the shape generation pipeline of ShapeWalk. Realistic 3D CAD shapes are reconstructed from the 3DCOMPAT++ [22] dataset into the GeoCode [17] shape program parameter space, using a mapping function " + }, + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "type": "inline_equation", + "content": "\\phi_{\\Theta}" + }, + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "type": "text", + "content": ". Reconstructed shapes with an error over a fixed threshold are discarded using a visual similarity function " + }, + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "type": "text", + "content": ", which is based on rendered feature similarity. Filtered pairs of shapes are then interpolated in the parameter space of the shape program, to generate shape chains " + }, + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(k)}" + }, + { + "bbox": [ + 46, + 192, + 548, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "inline_equation", + "content": "f_{\\Theta} : \\Theta \\mapsto S" + }, + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "text", + "content": " the shape program mapping shape parameters to a 3D mesh, and " + }, + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\phi_{\\Theta} : S \\mapsto \\Theta" + }, + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "text", + "content": " an encoder based on DGCNN [17, 25] which samples points from a shape surface and regresses corresponding shape parameters in " + }, + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 258, + 287, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 288, + 498 + ], + "type": "text", + "content": "We reconstruct shapes by first sampling pointclouds from the surface of the original meshes, and then feeding them into the DGCNN [25] encoder fine-tuned [17] to regress the parameters of the shape program. We discard a ratio of " + }, + { + "bbox": [ + 46, + 306, + 288, + 498 + ], + "type": "inline_equation", + "content": "\\alpha = 0.08" + }, + { + "bbox": [ + 46, + 306, + 288, + 498 + ], + "type": "text", + "content": " of the shapes with the highest reconstruction error, which we measure as the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder. This threshold is selected empirically by visually inspecting the shapes in the upper part of the reconstruction error distribution. In our dataset instance building on 3DComPAT++ shapes, we filter out a total of 89 shapes during this process, leading to a set of " + }, + { + "bbox": [ + 46, + 306, + 288, + 498 + ], + "type": "inline_equation", + "content": "|S_R| = 1113" + }, + { + "bbox": [ + 46, + 306, + 288, + 498 + ], + "type": "text", + "content": " reconstructed shapes. This process can be scaled up to any dataset size and any shape domain, as long as a corresponding shape program is available." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "content": "Shape interpolation. After reconstructing a set " + }, + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "inline_equation", + "content": "S_R" + }, + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "content": " of realistic shapes, we generate a set of shape chains " + }, + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(k)}" + }, + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "content": " by interpolating between the parameters of shape pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 618 + ], + "type": "text", + "content": "To that end, we first sample a set of shape pairs " + }, + { + "bbox": [ + 46, + 546, + 287, + 618 + ], + "type": "inline_equation", + "content": "(s_{\\theta_{\\mathrm{i}}}, s_{\\theta_{\\mathrm{j}}}) \\in S_{P} \\subseteq S_{R} \\times S_{R}" + }, + { + "bbox": [ + 46, + 546, + 287, + 618 + ], + "type": "text", + "content": ", and partition them into " + }, + { + "bbox": [ + 46, + 546, + 287, + 618 + ], + "type": "inline_equation", + "content": "L = 10" + }, + { + "bbox": [ + 46, + 546, + 287, + 618 + ], + "type": "text", + "content": " levels of proximity. These proximity levels are based on the feature-wise mean squared error between the original and reconstructed meshes' renderings, in the space of a pre-trained ResNet-50 [10] encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 628, + 282, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 282, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 282, + 643 + ], + "type": "text", + "content": "We define the set of shape pairs " + }, + { + "bbox": [ + 47, + 628, + 282, + 643 + ], + "type": "inline_equation", + "content": "S_P^{(l)}" + }, + { + "bbox": [ + 47, + 628, + 282, + 643 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 47, + 628, + 282, + 643 + ], + "type": "inline_equation", + "content": "l \\in [[1, L]]" + }, + { + "bbox": [ + 47, + 628, + 282, + 643 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 651, + 272, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 651, + 272, + 680 + ], + "spans": [ + { + "bbox": [ + 61, + 651, + 272, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {P} ^ {(l)} = \\{(s _ {\\theta_ {\\mathrm {i}}}, s _ {\\theta_ {\\mathrm {j}}}) \\in \\mathcal {S} _ {P} | \\overline {{d (s _ {\\theta_ {\\mathrm {i}}} , s _ {\\theta_ {\\mathrm {j}}})}} \\in \\left[ \\frac {l - 1}{L}, \\frac {l}{L} \\right] \\}", + "image_path": "61bdec1d2fb4372ce4be6987e5a7ec75137b3e8a4fb40100c328ef48fef6845d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\overline{d}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is a visual similarity function, min-max normalized across all pairs of shapes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": "For each pair " + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "inline_equation", + "content": "(s_{\\theta_i}, s_{\\theta_j}) \\in S_P^{(l)}" + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": ", we then interpolate between the shape parameters " + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\theta_j" + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": " to generate a set of intermediate shape parameters " + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "inline_equation", + "content": "(\\theta_i, \\dots, \\theta_{i + (N - 2)}, \\theta_j) \\in \\Theta^N" + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": ". The number of intermediate shapes " + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": " is defined as the number of differing parameters between " + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\theta_j" + }, + { + "bbox": [ + 304, + 257, + 547, + 319 + ], + "type": "text", + "content": ", that is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 328, + 479, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 328, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 372, + 328, + 479, + 342 + ], + "type": "interline_equation", + "content": "N = \\left| \\left\\{k \\mid \\theta_ {\\mathbf {i}} [ k ] \\neq \\theta_ {\\mathbf {j}} [ k ] \\right\\} \\right|", + "image_path": "3bc202309bb9cb19e55540602b045eb4b761331c596499a1dce87b82a2dbd15f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": "When the proximity level is low, edits necessary to transition from one shape to another are larger in intensity as the shapes are more different, and smaller when " + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": " is smaller. The ordering of intermediate edits is randomly sampled using a dependency-aware algorithm (for example, parameters relating to armrest height or width are sampled after the addition of armrests)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 435, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 470 + ], + "type": "text", + "content": "As illustrated in Figure 1, this generation process leads to plausible interpolated shapes, and to fine-grained and realistic sequences of edits with detailed metadata." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 471, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 495 + ], + "type": "text", + "content": "We summarize the process of generating realistic shape chains in Figure 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 507, + 546, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 546, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 546, + 615 + ], + "type": "text", + "content": "Random set. While the realistic set is based on reconstructed shapes from the 3DCOMPAT++ [15, 22] dataset, the random set is based on a collection of random shapes generated using the shape program. This alternative set aims at covering a large space of the parameter space of the GeoCode shape program. This subset is generated by systematically sampling various combinations of parameters within defined ranges determined by a minimum edit intensity, but does not necessarily result in realistic shapes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 623, + 409, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 409, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 409, + 635 + ], + "type": "text", + "content": "3.2. Text Instructions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Rule-based generation. Given an edit vector " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\overline{\\theta_{ij}}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " describing the parameter changes necessary to transition from shape " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ", we generate a set of text instructions " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " describing this edit in natural language. We first map each parameter " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{i}[k]" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " composing the edit vector to a natural language name, and map the magnitude of the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22577" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 77, + 521, + 390 + ], + "blocks": [ + { + "bbox": [ + 69, + 77, + 521, + 390 + ], + "lines": [ + { + "bbox": [ + 69, + 77, + 521, + 390 + ], + "spans": [ + { + "bbox": [ + 69, + 77, + 521, + 390 + ], + "type": "image", + "image_path": "8f747f3fb70df19623f37cb0c8fd5e2a1d25ecf62b3e6fedd4a46afcb2766a31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 402, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 547, + 460 + ], + "type": "text", + "content": "Figure 4. Dataset chains samples. We feature in this figure shape chains from our dataset with associated generated text instructions. Each shape chain is composed of a sequence of shapes, with each consecutive pair of shapes associated to a set of text instructions describing the edits necessary to transition from one shape to the next. Overall, generated text instructions are accurate and provide a detailed description of the edits applied to the shapes. We include in the supplementary material additional samples of generated shape chains, alongside shape chains generated for two additional table and vase categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 478, + 288, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 288, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 288, + 528 + ], + "type": "text", + "content": "parameter change to a natural language intensity depending on the parameter type. Boolean parameters are also described appropriately, for example by using the word add or remove when describing the addition or removal of a part." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 542, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 685 + ], + "type": "text", + "content": "Augmentation. This rule-based instruction generation leads to perfectly accurate instructions, but may lack diversity. To alleviate this issue, we 1) randomly sample adjectives and adverbs to characterize the intensity of edits, 2) randomly invert parameter names and magnitude directions (e.g. \"increase armrests straightness\" will be augmented to \"decrease armrests bend\" for the same edit). As a final augmentation, we utilize a T5 [18] transformer model finetuned for paraphrasing to generate additional instructions for each edit vector. We use the Parrot library [6] to filter generated paraphrases based on a fluency and adequacy score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "We summarize the process of generating text instructions in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 476, + 432, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 432, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 432, + 491 + ], + "type": "text", + "content": "4. Neural Shape Editing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 498, + 369, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 369, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 369, + 510 + ], + "type": "text", + "content": "4.1. Problem" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": "We are concerned with the task of editing a shape " + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": " given a sequence of natural language edit instructions. We want to learn an editing function " + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "inline_equation", + "content": "f_E: S \\times \\Sigma \\mapsto S" + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": " is the space of input shapes, and " + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": " is the space of natural language edit instructions, able to compose edits from a starting shape " + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": " to an ending shape " + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "inline_equation", + "content": "s_N" + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": ", given a sequence of edit instructions " + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p_t}\\}_{t=1}^N" + }, + { + "bbox": [ + 304, + 518, + 545, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 615, + 547, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 615, + 547, + 630 + ], + "spans": [ + { + "bbox": [ + 306, + 615, + 547, + 630 + ], + "type": "interline_equation", + "content": "\\widehat {s _ {N}} = f _ {E} (\\phi_ {T} (\\mathbf {p _ {1 2}}), f _ {E} (\\phi_ {T} (\\mathbf {p _ {2 3}}), \\ldots f _ {E} (\\phi_ {T} (\\mathbf {p _ {k N}}), s _ {1})))", + "image_path": "b1f07edeeef06c1d75bdc06daeac109c2bc1b1ba2d9b5e9bde07f10e23e1b3d3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "N = |\\mathcal{P}|" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " is the length of the edit chain, and " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi_T: \\Sigma \\mapsto \\mathbb{R}^{D_T}" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " is a text encoder. At the end of the edit sequence, we want to recover a shape " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{s_N}" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " as close as possible to the ground-truth shape " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "s_N" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": ". Our proposed dataset provides a ground-truth for every intermediate shape " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": ", which we leverage to train and evaluate our method." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22578" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 53, + 532, + 262 + ], + "blocks": [ + { + "bbox": [ + 60, + 53, + 532, + 262 + ], + "lines": [ + { + "bbox": [ + 60, + 53, + 532, + 262 + ], + "spans": [ + { + "bbox": [ + 60, + 53, + 532, + 262 + ], + "type": "image", + "image_path": "1669385cc877170fbfb73361aecfca58418d7e9a4a3fdbb0984ef34ad3312bb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 269, + 529, + 476 + ], + "blocks": [ + { + "bbox": [ + 55, + 269, + 529, + 476 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 529, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 529, + 476 + ], + "type": "image", + "image_path": "48027af70a79e59ef61ea06eed686f69fc495704f45fda8d2ab7eff75e458833.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 487, + 545, + 555 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 545, + 555 + ], + "type": "text", + "content": "Figure 5. Visualizing synthesized pairs of edits using our latent editors. We illustrate the diverse range of edits generated by our proposed latent mapper, showcasing its ability to capture nuanced variations in the input data, for 3DS2VS latents (top two rows) and PC-AE latents (bottom two rows). Each pair demonstrates the original input on the left and the corresponding synthesized output on the right, with the associated caption. We abbreviate instructions which are too long to fit in the figure, and indicate when text is omitted with an ellipsis. For more subtle changes, we circle the areas of interest in both source and target shapes. Overall, edits are consistent with the input instruction, preserve shape identity while generating plausible output shapes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 574, + 115, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 115, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 115, + 586 + ], + "type": "text", + "content": "4.2. Objective" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": "One of the main advantages of training a shape editing model with ShapeWalk compared to other works [2, 8] is the availability of exact ground-truth edited shapes for each edit instruction (see Section 2). We thus formulate the objective function of our learning problem as a simple " + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": " loss between the latents of the synthesized shape and the ground-truth shape. For all pairs of shapes " + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "inline_equation", + "content": "(s_i,s_j)" + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": " and their corresponding edit instructions " + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{p_{ij}}" + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": ", we minimize the following loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 700, + 240, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 700, + 240, + 715 + ], + "spans": [ + { + "bbox": [ + 93, + 700, + 240, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| f _ {E} \\left(\\phi_ {E} \\left(s _ {i}\\right), \\mathbf {p} _ {\\mathbf {i j}}\\right) - \\phi_ {E} \\left(s _ {j}\\right) \\right\\| _ {2}", + "image_path": "245c457480f32559e53d0f51dd4a74be2d227a5f8f9fb6fc7784a6e11b195d1c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "type": "inline_equation", + "content": "\\phi_{AE} = \\phi_E\\circ \\phi_D" + }, + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "type": "text", + "content": " is a shape autoencoder, and " + }, + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "type": "inline_equation", + "content": "\\phi_E:" + }, + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "type": "inline_equation", + "content": "S\\mapsto \\mathbb{R}^{D_E}" + }, + { + "bbox": [ + 305, + 574, + 545, + 599 + ], + "type": "text", + "content": " denotes its encoder component." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 609, + 362, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 362, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 362, + 620 + ], + "type": "text", + "content": "4.3. Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 628, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 689 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 304, + 628, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\phi_{AE}" + }, + { + "bbox": [ + 304, + 628, + 545, + 689 + ], + "type": "text", + "content": ", we experiment with PC-AE [1] and 3D2VS [28] models both pre-trained on ShapeNet [5]. Note that our method is agnostic to the choice of shape autoencoder and can be adapted to a variety of shape representations. For " + }, + { + "bbox": [ + 304, + 628, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\phi_T" + }, + { + "bbox": [ + 304, + 628, + 545, + 689 + ], + "type": "text", + "content": ", we use a pretrained BERT [7] model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "Both " + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi_{AE}" + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi_T" + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": " are frozen during training, and we only train the parameters of our latent mapper " + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22579" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "text", + "content": "For PC-AE, we formulate our latent mapper as a neural module which for each tuple " + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "(\\phi_E(s_i), \\mathbf{p_{ij}}, \\phi_E(s_j))" + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "text", + "content": " predicts an edit vector " + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "\\widehat{\\theta_{ij}} \\in \\mathbb{R}^{D_E}" + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "text", + "content": " in the feature space of " + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "\\phi_E" + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "text", + "content": ", and adds it to the latent of the input shape " + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "text", + "content": " to generate the latent of the output shape " + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 72, + 287, + 135 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 145, + 236, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 145, + 236, + 162 + ], + "spans": [ + { + "bbox": [ + 99, + 145, + 236, + 162 + ], + "type": "interline_equation", + "content": "f _ {E} \\left(\\phi_ {E} \\left(s _ {i}\\right), \\mathbf {p} _ {\\mathbf {i j}}\\right) = \\widehat {\\theta_ {i j}} + \\phi_ {E} \\left(s _ {i}\\right)", + "image_path": "b80d9ae63ed46aa445bd56afac370daad82cbdedc7b36f6184ec17edf176afc9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "text", + "content": "We utilize two variants of our latent mapper " + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "text", + "content": ": one in which the edit vector " + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\widehat{\\theta_{ij}}" + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "text", + "content": " is predicted directly, and one in which the edit vector " + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\widehat{\\theta_{ij}}" + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "text", + "content": " is predicted as a product of a normalized edit direction " + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\widehat{v_{ij}} \\in \\mathbb{R}^{D_E}" + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "text", + "content": " and a magnitude " + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\widehat{m_{ij}} \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 172, + 287, + 236 + ], + "type": "text", + "content": " separately." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 287, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 333 + ], + "type": "text", + "content": "For 3D2VS which encodes shapes as latents sets of higher resolutions, we use a transformer-based latent diffusion model instead to generate edited latents. We concatenate the input shape latent " + }, + { + "bbox": [ + 46, + 237, + 287, + 333 + ], + "type": "inline_equation", + "content": "\\phi_E(s_i)" + }, + { + "bbox": [ + 46, + 237, + 287, + 333 + ], + "type": "text", + "content": " with the noised edited shape latent " + }, + { + "bbox": [ + 46, + 237, + 287, + 333 + ], + "type": "inline_equation", + "content": "\\phi_E(s_j + \\epsilon_t)" + }, + { + "bbox": [ + 46, + 237, + 287, + 333 + ], + "type": "text", + "content": " and feed the BERT text embeddings to the cross-attention layers of the transformer blocks to predict the added noise. We illustrate our architecture for the latent mapper in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 346, + 128, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 128, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 128, + 360 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 366, + 179, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 179, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 179, + 380 + ], + "type": "text", + "content": "5.1. Chained Shape Editing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 386, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 411 + ], + "type": "text", + "content": "Comparison Models. For PC-AE, we experiment with three variants of our latent mapper module:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 421, + 287, + 562 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 47, + 421, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 445 + ], + "type": "text", + "content": "- DIRECTGEN directly predicts the edited shape latent without regressing an edit vector." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 456, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 515 + ], + "type": "text", + "content": "- LATEFUSION follows the network proposed in [3], and first passes the shape embeddings through an encoder. Encoded shape features are then concatenated with text embeddings and fed into a neural module which finally predicts the edit vector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 525, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 287, + 562 + ], + "type": "text", + "content": "- Ours is a simple multi-layer perceptron (MLP) which directly takes the context (shape and text embeddings) as input to predict an edit vector." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 567, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 592 + ], + "type": "text", + "content": "For all of these variants, we experiment with various bottleneck dimensions and number of layers (in subscript)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": "We further decompose these variants into coupled and decoupled versions, where we predict a normalized edit direction " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\widehat{v_{ij}}\\in \\mathbb{R}^{D_E}" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": " and a magnitude " + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\widehat{m_{ij}}\\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": " separately. We provide in appendix the full architecture details and training hyperparameters employed to train our models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "type": "text", + "content": "Note that we do not compare with ChangeIt3D [2] as their neural-listener distillation method is not applicable in our context: we train our models with a direct feedback signal from the ground-truth edited shapes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "Single-step metrics. We propose to measure the quality of our editing steps both before and after shape reconstruction. Since we experiment on point clouds, we use the scaled Chamfer Distance between the reconstructed original shape " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and the reconstructed edited shape " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\widehat{s_j}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 140, + 539, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 539, + 170 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 539, + 170 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {C D}} = \\frac {1}{| \\widehat {s _ {j}} |} \\sum_ {x \\in \\widehat {s _ {j}}} \\min _ {y \\in s _ {i}} \\| x - y \\| _ {2} + \\frac {1}{| s _ {i} |} \\sum_ {y \\in s _ {i}} \\min _ {x \\in \\widehat {s _ {j}}} \\| x - y \\| _ {2}", + "image_path": "53667eff713eadc3a2068c22920f1d63c731bd536a63e11b34972385c38fd48d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "text", + "content": "We also use the " + }, + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "text", + "content": " distance between the latents of the original shape " + }, + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "text", + "content": " and the edited shape " + }, + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 305, + 178, + 544, + 202 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 209, + 484, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 209, + 484, + 224 + ], + "spans": [ + { + "bbox": [ + 367, + 209, + 484, + 224 + ], + "type": "interline_equation", + "content": "d _ {\\mathcal {L} _ {2}} = \\left\\| \\phi_ {E} (s _ {i}) - \\phi_ {E} (s _ {j}) \\right\\| _ {2}", + "image_path": "1bf940b9413040d3c05e9d2b5ee413909b7f80585490557804310daac1c53d76.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 231, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 277 + ], + "type": "text", + "content": "Multi-step metrics. To extend the single-step metrics to chained shape generation, we take inspiration from the trajectory prediction literature [4, 26] and propose appropriate metrics for our task." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": "Given a shape distance function " + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "inline_equation", + "content": "d: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\mathbb{R}" + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": ", a chain of ground-truth reconstructed shapes " + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\{s_1, \\ldots, s_N\\}" + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": ", and a set of recursively generated shapes " + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\{\\widehat{s}_1, \\ldots, \\widehat{s_N}\\}" + }, + { + "bbox": [ + 304, + 278, + 545, + 350 + ], + "type": "text", + "content": ", we define the average edit error as the average distance between the generated shapes and the corresponding ground-truth shapes:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 380, + 350, + 471, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 350, + 471, + 381 + ], + "spans": [ + { + "bbox": [ + 380, + 350, + 471, + 381 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {d} = \\frac {1}{N} \\sum_ {t = 1} ^ {N} d \\left(s _ {t}, \\widehat {s} _ {t}\\right)", + "image_path": "fc7ecec03c30ced55a5560d816ae80fa192da525fdf478a584898422d927b929.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 385, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 421 + ], + "type": "text", + "content": "And the final edit error as the distance between the last shape in the chain and the corresponding ground-truth shape:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 391, + 421, + 460, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 421, + 460, + 434 + ], + "spans": [ + { + "bbox": [ + 391, + 421, + 460, + 434 + ], + "type": "interline_equation", + "content": "\\mathrm {F} _ {d} = d \\left(s _ {N}, \\widehat {s _ {N}}\\right)", + "image_path": "294c8e87eff8783e9aabca44f6490c2277d33db5674e97b74ba9b9b7620652f6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 438, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 464 + ], + "type": "text", + "content": "We report results with both metrics for " + }, + { + "bbox": [ + 304, + 438, + 545, + 464 + ], + "type": "inline_equation", + "content": "d = d_{\\mathrm{CD}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 545, + 464 + ], + "type": "inline_equation", + "content": "d = d_{\\mathcal{L}_2}" + }, + { + "bbox": [ + 304, + 438, + 545, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "type": "text", + "content": "Results. We provide qualitative results of edited pairs generated by our latent mapper in Figure 5. Overall, we observe that our method is able to generate plausible edits that are consistent with the input instruction, and preserve the identity of the input shape. However, our method is limited by the quality of the shape autoencoder. We notice that the PC-AE editor is unable to perform very fine-grained edits which are not properly captured by the autoencoder, while visible in the ground-truth shapes. We explore the possible causes of these limitations in Section 5.2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We provide quantitative results for our chained editing metrics averaged across " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}| \\in \\{10, 15, 20\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " in Table 1, and detail per-chain length results in the appendix. Directly predicting the edited shape latent (DIRECTGEN) leads to poor edit predictions across all metrics, although increasing the number of layers may improve the results. Overall, decoupling the edit direction and magnitude leads to better results. Our coupled " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{OURS}_{512 \\times 4}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " variant performs better than the coupled " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LATEFUSION}_{512}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " variant on the Chamfer Distance, but worse on " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " distance." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22580" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 60, + 288, + 248 + ], + "blocks": [ + { + "bbox": [ + 47, + 60, + 288, + 248 + ], + "lines": [ + { + "bbox": [ + 47, + 60, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 60, + 288, + 248 + ], + "type": "table", + "html": "
Modeldecoupled?Averaged ∀|P|
FCD×1e4ACD×1e4FL2AL2
LATEFUSION1024×2.8562.6211.4441.251
LATEFUSION512×2.7192.6091.4621.290
LATEFUSION256×2.8742.6511.5091.283
OURS512×8×3.2082.8221.7031.437
OURS512×4×2.9902.7031.5891.351
LATEFUSION1024×2.7112.5681.4051.245
LATEFUSION512×3.0022.7081.4511.254
LATEFUSION256×3.3092.8481.5521.324
OURS512×8×2.7822.5841.4971.290
OURS512×4×2.6702.5241.4471.266
", + "image_path": "c51a4417244b7178f536fed1e2674294ca4ddf3d0e36ea1933755b6ab20c6573.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 322, + 276, + 490 + ], + "blocks": [ + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "type": "text", + "content": "Table 1. Chained shape editing ablation. We report baseline results for our proposed chained shape editing task, averaged across all chain lengths " + }, + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}| \\in \\{10, 15, 20\\}" + }, + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "type": "text", + "content": ". Both the average final and average edit error are reported for the Chamfer Distance (CD) and " + }, + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "type": "text", + "content": " distance " + }, + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_2)" + }, + { + "bbox": [ + 46, + 255, + 288, + 312 + ], + "type": "text", + "content": " metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 322, + 276, + 490 + ], + "lines": [ + { + "bbox": [ + 59, + 322, + 276, + 490 + ], + "spans": [ + { + "bbox": [ + 59, + 322, + 276, + 490 + ], + "type": "table", + "html": "
ParameterAccuracy
seat height1.000
backrest curvature1.000
object width/height/depth0.978
seat roundness0.978
top bar thickness/height0.961
legs thickness0.956
legs bending/curvature0.944
adding/removing handle cushions0.750
number of legs/backrest rails0.663
legs roundness/indentation0.587
AVG (RANDOM)0.884
AVG (REALISTIC)0.969
", + "image_path": "9b9ef74325056a1084144054ab1a36c8b0cd66489816d2231f14432add2ad183.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 287, + 564 + ], + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 564 + ], + "type": "text", + "content": "Table 2. Parameter-wise accuracy for shape edit recognition. We report the accuracy of our edit detector for each parameter in the random set, alongside the average accuracy for the full random and realistic sets. Overall, we observe that our classifier is able to recognize edits corresponding to high change edits, but fails to differentiate between subtle changes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 575, + 187, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 187, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 187, + 588 + ], + "type": "text", + "content": "5.2. Recognizing Shape Edits" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": "Problem. To provide additional insight into the quality of our latent mapper and the difficulty of predicting specific edits, we train a binary classifier to discriminate from a pair of shapes " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "(s_i, s_j)" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": " on whether the corresponding edit instruction " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{ij}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": " was applied to " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": ". We train a binary classifier " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "f_C: \\mathcal{S} \\times \\mathcal{S} \\mapsto \\{0,1\\}" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "f_C(s_i, s_j) = 1" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": " if the edit instruction " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{ij}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": " was applied to " + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 594, + 287, + 679 + ], + "type": "text", + "content": ", and 0 otherwise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Method. Similarly to our latent mapper architecture for PC-AE, we use an MLP-based shape latent encoder to encode" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "the input shape latents into the bottleneck dimension, and project text features using a single linear layer to the same dimension. We then concatenate the two feature vectors and pass them through an MLP predictor outputting probabilities for the two classes " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". We train our classifier using a binary cross-entropy loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 547, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 214 + ], + "type": "text", + "content": "Training and evaluation. We train our classifier on the realistic set and evaluate it on the full random shape set, which has a restricted vocabulary of edit instructions. We evaluate the accuracy of the classifier on the full set, as well as on subsets of the data corresponding to specific types of edits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "text", + "content": "Results. In Table 2, we provide accuracy results for shape edit recognition, depending on the edit type. We also provide the global accuracy on the random and realistic sets. We observe that our classifier is able to recognize edits corresponding to shape parameters associated with a high degree of shape variation (like seat height and width/height of the shape), but struggles with parameters which are more subtle (like legs roundness, handle cushions). Our classifier is able to detect edits related to global width/height/depth adjustments with an accuracy of " + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "inline_equation", + "content": "97.8\\%" + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "text", + "content": ". This accuracy drops to " + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "inline_equation", + "content": "94.4\\%" + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "text", + "content": " for legs bending/curvature, which is a more subtle change in the shape, and to " + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "inline_equation", + "content": "58.7\\%" + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "text", + "content": " for legs roundness/indentation. We also note that the model does not perform well on edits related to the number of parts, i.e. adding or removing a number " + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 216, + 546, + 395 + ], + "type": "text", + "content": " of legs or backrest rails." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 515 + ], + "type": "text", + "content": "We impute these discrepancies to several factors. The PC-AE shape autoencoder we employ may be unable to properly reconstruct fine-grained details in the ground-truth and edited shapes, which renders the task of detecting subtle changes more difficult. Pointcloud resolution is also a limitation, as fine-grained parameters like leg indentation may not be properly sampled. Finally, learning a latent mapper and a classifier that can accurately count the number of parts to add or remove may also be a difficult task of its own, and require more domain-specific inductive biases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 526, + 379, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 379, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 379, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "In conclusion, this paper introduces the ShapeWalk dataset, designed for advancing compositional shape editing guided by natural language instructions. The dataset comprises 158K unique shapes connected through 26K edit chains, synthesized from a realistic CAD-designed 3D dataset. Language instructions for applied edits are provided, alongside exact ground-truth edited shapes. Our method requires zero human annotation effort, and can be scaled up indefinitely. The usefulness of the dataset is demonstrated by training neural editor modules in the latent space of shape autoencoders [1]. Evaluation metrics inspired by trajectory prediction literature [4, 26] are introduced, offering a quantitative assessment of the quality of recursively edited shapes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22581" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 225 + ], + "type": "text", + "content": "[3] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. ShapeTalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 282 + ], + "type": "text", + "content": "[4] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961–971, 2016. 2, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "type": "text", + "content": "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. In arXiv, 2015. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 287, + 361 + ], + "type": "text", + "content": "[6] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 396 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "text", + "content": "[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2023. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "text", + "content": "[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "text", + "content": "[11] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics, 39(6):1-20, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[12] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: macro operation discovery for 3d shape programs. ACM Transactions on Graphics, 40(4), 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[13] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[14] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics, 42(4), 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Yuchen Li, Ujjwal Upadhyay, Habib Slim, Ahmed Abdelreheem, Arpit Prajapati, Suhail Pothigara, Peter Wonka, and Mohamed Elhoseiny. 3D CoMPaT: Composition of Materials on Parts of 3D Things. In ECCV, 2022. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 547, + 163 + ], + "type": "text", + "content": "[16] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 165, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 198 + ], + "type": "text", + "content": "[17] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. In arXiv preprint arxiv:2212.11715, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "text", + "content": "[18] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. In arXiv, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 246, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 246, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 545, + 290 + ], + "type": "text", + "content": "[19] Daniel Ritchie, Paul Guerrero, R. Kenny Jones, Niloy J. Mitra, Adriana Schulz, Karl D. D. Willis, and Jiajun Wu. Neurosymbolic models for computer graphics. Computer Graphics Forum, 42(2), 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 292, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 347 + ], + "type": "text", + "content": "[20] Aditya Sanghi, Hang Chu, Joseph G. Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 415 + ], + "type": "text", + "content": "[21] Aditya Sanghi, Rao Fu, Vivian Liu, Karl D.D. Willis, Hooman Shayani, Amir H. Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 417, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 417, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 417, + 545, + 483 + ], + "type": "text", + "content": "[22] Habib Slim, Xiang Li, Mahmoud Ahmed Yuchen Li, Mohamed Ayman, Ujjwal Upadhyay Ahmed Abdelreheem, Suhail Pothigara Arpit Prajapati, Peter Wonka, and Mohamed Elhoseiny. 3DCoMPaT++: An improved large-scale 3d vision dataset for compositional recognition. In arXiv preprint arxiv:2212.11715, 2023. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 485, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 485, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 485, + 545, + 518 + ], + "type": "text", + "content": "[23] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. CoRR, abs/1711.00937, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 520, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 545, + 575 + ], + "type": "text", + "content": "[24] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 577, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 620 + ], + "type": "text", + "content": "[25] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. In SIGGRAPH, 2019. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 623, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 623, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 623, + 545, + 657 + ], + "type": "text", + "content": "[26] Kaiping Xu, Zheng Qin, Guolong Wang, Kai Huang, Shuxiong Ye, and Huidi Zhang. Collision-Free LSTM for Human Trajectory Prediction, pages 106-116. 2018. 2, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[27] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22582" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 117 + ], + "type": "text", + "content": "[28] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics, 42(4):1-16, 2023. 6" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22583" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_content_list.json b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..91975b402c9fdc958678cbc9b79b1119e99a653c --- /dev/null +++ b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_content_list.json @@ -0,0 +1,1511 @@ +[ + { + "type": "text", + "text": "Sharingan: A Transformer Architecture for Multi-Person Gaze Following", + "text_level": 1, + "bbox": [ + 112, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samy Tafasca Anshul Gupta Jean-Marc Odobez Idiap Research Institute, Switzerland École Polytechnique Fédérale de Lausanne, Switzerland {stafasca, agupta, odobez}@idiap.ch", + "bbox": [ + 233, + 180, + 733, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f31d57e4dd09686e237ee1ef97f6c1f37f60b39a2b1a85d8bb06b848286d2c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 287, + 316, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/989b488159efff31f59072335aa643d96bdc887478fff84ccb4a7c45e417a723.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 287, + 653, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd20632f1bc399cfd863e8a463ed244cf2a229ad3a2b88e1715ba2a71a44faf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 287, + 890, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2c396e6a6b18277e75d65260aca90dc22199cc609ac48bde274451a140d7c20.jpg", + "image_caption": [ + "Figure 1. Predictions of Sharingan on naturalistic images from the internet with different people, activities, interactions, postures, and environments (indoors and outdoors). We provide more qualitative samples in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 78, + 411, + 359, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c1241e016785f2cc88308d28fd101a2defd70ca443e4625dcf319653ab1df70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 411, + 607, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbc3aed12cb3eda5443dc888f733066640a2dac7b542c62dcb64e0182e2bf9f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 411, + 890, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 585, + 312, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaze is a powerful form of non-verbal communication that humans develop from an early age. As such, modeling this behavior is an important task that can benefit a broad set of application domains ranging from robotics to sociology. In particular, the gaze following task in computer vision is defined as the prediction of the 2D pixel coordinates where a person in the image is looking. Previous attempts in this area have primarily centered on CNN-based architectures, but they have been constrained by the need to process one person at a time, which proves to be highly inefficient. In this paper, we introduce a novel and effective multi-person transformer-based architecture for gaze prediction. While there exist prior works using transformers for multi-person gaze prediction [38, 39], they use a fixed set of learnable embeddings to decode both the person and its gaze target, which requires a matching step afterward to link the predictions with the annotations. Thus, it is difficult to quantitatively evaluate these methods reliably with", + "bbox": [ + 73, + 628, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the available benchmarks, or integrate them into a larger human behavior understanding system. Instead, we are the first to propose a multi-person transformer-based architecture that maintains the original task formulation and ensures control over the people fed as input. Our main contribution lies in encoding the person-specific information into a single controlled token to be processed alongside image tokens and using its output for prediction based on a novel multiscale decoding mechanism. Our new architecture achieves state-of-the-art results on the GazeFollow, VideoAttentionTarget, and ChildPlay datasets and outperforms comparable multi-person architectures with a notable margin. Our code, checkpoints, and data extractions will be made publicly available soon.", + "bbox": [ + 496, + 588, + 893, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 829, + 632, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaze is an important form of communication and was extensively studied across different domains and applications such as consumer behavior understanding [4, 19, 36], soci", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2008", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6cfe831da9f5f7d6da0f6136e99a53f846c71a1cc33530cda5032e46d98bd8c1.jpg", + "image_caption": [ + "Figure 2. Illustration of aspects relevant to gaze following." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 467, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ology by analyzing different gaze behaviors (e.g. joint attention, eye contact) [10, 24-27], robotics through humanrobot interactions [1, 17, 32] and clinical research for the study of neurodevelopmental disorders [7, 20, 34], etc.", + "bbox": [ + 75, + 262, + 467, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike traditional works on gaze analytics proposed by the computer vision community which focused mainly on predicting 3D gaze directions from the eyes [33, 40] or the face [18] of a person, gaze following [29] tackles the task in a more general form where the goal is to infer the 2D location in the image where a person is looking without the need for any assumptions or wearable devices. This formulation is particularly interesting in the context of analyzing social scenes and human interactions given the important role that gaze behavior plays in social dynamics.", + "bbox": [ + 75, + 323, + 467, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is important to emphasize that this task is very difficult to solve. It hinges upon understanding two aspects simultaneously: (i) the target person (e.g. head pose) to infer a general gaze direction, and (ii) the context of the scene (e.g. social interaction) to identify regions of saliency (cf. Figure 2). Once this is achieved, the rest is essentially a selection process of the gaze target after merging information from the previous steps. This also explains why most architectures have a 2-tower design with one branch to process the scene and a second one to focus on the target person.", + "bbox": [ + 75, + 476, + 467, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One major issue with this dual-stream approach is that the input instance is a person, not an image. Therefore, it requires multiple forward passes to predict gaze for all people in the same scene, which makes the inference process extremely inefficient. This is made even worse when other modalities are involved (e.g. depth [11, 35] or pose [12]). Moreover, most previous works focus on how to build the person-specific gaze representation, the scene saliency maps, and the fusion between them but pay very little attention to the decoding mechanism that predicts the final gaze heatmap. For example, [7, 11, 17] all use a decoding module based on a few convolutional layers followed by a set of transposed convolutions. The input to this module is often very low dimensional (e.g. $7 \\times 7$ ) which forces the prediction to be based on coarse information.", + "bbox": [ + 75, + 628, + 467, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to tackle these challenges by addressing the multi-person gaze-following task while maintaining the original problem formulation. To this end,", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we propose Sharingan, a novel, effective, and efficient transformer-based architecture to predict the gaze target of multiple people simultaneously. A key component of this architecture is to represent the person's gaze information by a single gaze token produced by a gaze backbone and processed alongside the image tokens. This is in stark contrast to previous methods that represent intermediate person-specific gaze information as a visual attention map [7] or gaze cone [11, 12, 21, 35]. We show in our ablations that this is not only unnecessary but can also hinder performance in the context of a transformer. Furthermore, we introduce Conditional DPT: a more sophisticated lightweight multi-scale gaze decoding mechanism that helps improve performance by providing a finer-grained understanding of the scene for gaze target selection. This also has the benefit of producing heatmaps that better capture uncertainty when it is difficult to decide where a person is looking (cf. qualitative results in the supplementary material).", + "bbox": [ + 496, + 90, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive ablations and evaluations, we find that Sharingan achieves good performance on all public benchmarks, and even transfers well to other gaze-related tasks such as shared attention and mutual gaze.", + "bbox": [ + 496, + 363, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 436, + 638, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we present several relevant research topics.", + "bbox": [ + 500, + 462, + 883, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaze Following. The task of gaze following was first introduced in the seminal work of Recasens et al. [29]. The idea is to predict the pixel-wise 2D location in the image corresponding to where a target person is looking. The main advantage of this formulation is the lack of constraints which allows methods trained this way to generalize to arbitrary settings (i.e. scene properties, camera parameters, image conditions, etc.). It was later extended by Chong et al. [7] to also include the prediction of whether the given person is looking inside the image frame or somewhere outside.", + "bbox": [ + 496, + 478, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional methods for gaze following [7, 11, 12, 16, 17, 21, 29] typically rely on convolutional networks and follow a 2-tower architecture. The first branch processes the scene image to highlight salient regions, while the second branch processes the head crop of the target person to infer a general gaze direction. A fusion mechanism then combines information from both parts to produce the final prediction.", + "bbox": [ + 496, + 628, + 890, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The gaze following task is often framed as the prediction of a gaze heatmap where pixels with high intensity represent spatial areas with higher prediction confidence. We devote a section later to discuss the alternative formulation of regressing the 2D location directly (cf. Section 5).", + "bbox": [ + 496, + 734, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Person Gaze Following. A major downside of the traditional formulation of gaze following is the need for multiple forward passes when predicting the gaze of different people in the same image. This problem motivated the need for architectures that can natively handle the prediction of gaze for multiple people with a single forward", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2009", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pass. Jin et al. [16] first proposed a simple convolution-based architecture to handle the multi-person setting where a scene backbone computes a fixed person-agnostic feature representation. This is then fused repetitively with head features computed from the different people using another head backbone before decoding each into its corresponding gaze heatmap. Aside from the architectural differences and limited performance, one of the main drawbacks of this method is that the computation for each person is done independently from the others, which ignores potential interactions between people. Recently, Tu et al. [39] and Tonini et al. [38] proposed transformer-based architectures to perform multi-person gaze target prediction. Their methods only take the image as input and simultaneously predict both the head box and gaze target (among others) for every person in the scene. Their work is inspired by the DETR architecture [5], where the task is formulated as a set prediction problem. Instead of reinventing the wheel, our method focuses solely on the gaze prediction part (i.e. given that heads are easily and accurately obtainable using off-the-shelf detectors), and naturally adapts the transformer architecture to the original task formulation by introducing gaze tokens to capture person-specific gaze and head location information and can be directly decoded later into gaze predictions.", + "bbox": [ + 75, + 90, + 472, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sharingan Architecture", + "text_level": 1, + "bbox": [ + 76, + 469, + 302, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Sharingan architecture is illustrated in Figure 3. The main idea is to use a transformer that lets scene tokens and person-specific gaze tokens interact within an attention framework to jointly predict the 2D gaze heatmap of each individual. Thus, the inputs are the image and the head crops that we assume are available. We introduce below the different components of this architecture.", + "bbox": [ + 75, + 494, + 468, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image tokens", + "text_level": 1, + "bbox": [ + 76, + 612, + 217, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow a standard ViT architecture to produce image tokens. The input scene image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}$ goes through a patch projection $\\mathcal{P}_{\\mathrm{img}}$ to produce image tokens that we equip with positional information $\\mathbf{x}^{\\mathrm{img}} \\in \\mathbb{R}^{N \\times D}$ , where $N$ is the number of patches, and $D$ is the token dimension.", + "bbox": [ + 75, + 636, + 468, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Gaze tokens", + "text_level": 1, + "bbox": [ + 76, + 723, + 207, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main purpose of a gaze token is to map the gaze information of a person into a token embedded in the same space as the image tokens, which can interact with scene tokens to select the relevant content for prediction. For simplicity, we first introduce this process for a single person.", + "bbox": [ + 75, + 747, + 468, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single Person Case. Let $\\mathbf{h}_{\\mathrm{crop}} \\in \\mathbb{R}^{h \\times w \\times C}$ denote the head crop of a person and $\\mathbf{h}_{\\mathrm{bbox}} = (x_{\\mathrm{min}}, y_{\\mathrm{min}}, x_{\\mathrm{max}}, y_{\\mathrm{max}}) \\in [0,1]^4$ her head bounding box. The mapping works as follows. The head crop $\\mathbf{h}_{\\mathrm{crop}}$ is fed to a gaze backbone $\\mathcal{G}$ to produce a gaze embedding $\\mathbf{g}^{\\mathrm{emb}} \\in \\mathbb{R}^{d_{\\mathrm{emb}}}$ . This embedding", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is used in two ways. First, it goes through a gaze regressor (i.e. MLP) $\\mathcal{O}_{\\mathrm{gv}}$ to predict a 2D gaze vector $\\mathbf{g}_{\\mathrm{v}} = \\mathcal{O}_{\\mathrm{gv}}(\\mathbf{g}^{\\mathrm{emb}})$ . This output is supervised using an angular gaze loss.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, the gaze embedding is projected to the token dimension using a learnable linear projection $\\mathcal{P}_{\\mathrm{gaze}}$ , resulting in the gaze token $\\mathbf{x}^{\\mathrm{emb}} = \\mathcal{P}_{\\mathrm{gaze}}(\\mathbf{g}^{\\mathrm{emb}}) \\in \\mathbb{R}^{D}$ . As we want to incorporate information about the person's location (and size), we also project the head bounding box $\\mathbf{h}_{\\mathrm{bbox}}$ into a bounding box embedding $\\mathbf{x}^{\\mathrm{bbox}}$ using a learnable linear projection $\\mathcal{P}_{\\mathrm{bbox}}$ : $\\mathbf{x}^{\\mathrm{bbox}} = \\mathcal{P}_{\\mathrm{bbox}}(\\mathbf{h}_{\\mathrm{bbox}}) \\in \\mathbb{R}^{D}$ . Finally, we add this embedding to the gaze token to obtain the final location-aware gaze token:", + "bbox": [ + 496, + 137, + 893, + 272 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\mathrm {g}} = \\mathbf {x} ^ {\\text {e m b}} + \\mathbf {x} ^ {\\text {b b o x}} \\in \\mathbb {R} ^ {D} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 285, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-person case. When $N_{p}$ persons are detected, the architecture will produce a set of $N_{p}$ gaze tokens, following the same process described above for each person. Thus, if $\\mathbf{h}_{\\mathrm{bbox}}^i$ and $\\mathbf{h}_{\\mathrm{crop}}^i$ denote the bounding-box and head crop of person $i$ , the above process will generate a gaze token $\\mathbf{x}_i^{\\mathrm{g}}$ for this person. To simplify notation, we will also denote by $\\mathbf{x}^{\\mathrm{g}}$ the set of gaze tokens of all people in the scene, with $\\mathbf{x}^{\\mathrm{g}} = \\mathbf{x}_1^{\\mathrm{g}} \\oplus \\ldots \\oplus \\mathbf{x}_{N_p}^{\\mathrm{g}}$ , where $\\oplus$ is the concatenation operator.", + "bbox": [ + 496, + 311, + 890, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modality Encoding. Given the different nature of gaze tokens compared to image tokens, we need to encode modality-specific information to distinguish between them. Rather than using an explicit scheme, in practice we expect this modality information to be captured by the bias terms of the different projection operators $\\mathcal{P}_{\\mathrm{gaze}}$ and $\\mathcal{P}_{\\mathrm{img}}$ .", + "bbox": [ + 496, + 436, + 890, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Transformer Encoder", + "text_level": 1, + "bbox": [ + 500, + 535, + 702, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transformer encoder is a standard ViT [8]. It takes as input the concatenation of the scene tokens $\\mathbf{x}^{\\mathrm{img}}$ , the gaze token(s) $\\mathbf{x}^{\\mathrm{g}}$ , according to $\\mathbf{x} = \\mathbf{x}^{\\mathrm{img}} \\oplus \\mathbf{x}^{\\mathrm{g}} \\in \\mathbb{R}^{N_t \\times D}$ , where $N_t = N + N_p$ . The set of input tokens goes through a series of $L$ transformer blocks to obtain an output sequence of similar shape, denoted by $\\mathbf{x}^{\\mathrm{out}} = \\mathbf{x}^{(L)} \\in \\mathbb{R}^{N_t \\times D}$ .", + "bbox": [ + 496, + 558, + 890, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Gaze Decoder", + "text_level": 1, + "bbox": [ + 500, + 659, + 643, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of the gaze decoder $\\mathcal{D}_{\\mathrm{gaze}}$ is to predict a set of gaze heatmaps. Our Conditional DPT (cf. Figure 4) takes four intermediate representations of the image tokens $\\mathbf{x}_{(i)}^{\\mathrm{img}}$ and gaze tokens $\\mathbf{x}_{(i)}^{\\mathrm{g}}$ and combines them progressively at different simulated resolutions, where lower resolutions have more channels and correspond to deeper layers of the encoder. This can be viewed as the isotropic equivalent of a Feature Pyramid Network [22].", + "bbox": [ + 496, + 681, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our design is inspired by DPT [28], which can only handle decoding image tokens alone. In our case, we need this decoding to be conditioned on each person. To this end, after each block of layers, the image tokens $\\mathbf{x}_{(l_k)}^{\\mathrm{img}}$ , $k \\in \\{4, 8, 16, 32\\}$ are reassembled into an image-like representation at resolution $\\left(\\frac{H}{k}, \\frac{W}{k}\\right)$ and dimension $d_k$ . The gaze to", + "bbox": [ + 496, + 806, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2010", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6217bc5303033b18b549d5ac042ec60e5c2335598f84db871b70d0e02251c30.jpg", + "image_caption": [ + "Figure 3. Overview of our Sharingan architecture. A. The input image is projected into image tokens (red squares). B. The head crops and head box coordinates are processed to generate location and size-aware person-specific tokens (blue squares) as follows. First, the head crop is fed to a gaze backbone to produce a gaze embedding used to (i) predict a normalized 2D gaze vector that is supervised using an angular loss; and (ii) produce a gaze token by projecting it to the token dimension. Second, head bounding box coordinates are projected to obtain an embedding which, added to the gaze token, produces a person gaze token. C. Image tokens and gaze tokens are fed to the transformer encoder, and the output tokens corresponding to input people are all decoded using a Conditional DPT decoder to predict each person's gaze heatmaps. In addition, input and output gaze tokens are concatenated together to predict the in- vs-out label." + ], + "image_footnote": [], + "bbox": [ + 122, + 93, + 851, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "kens $\\mathbf{x}_{(l_k)}^{\\mathrm{g}}$ at that layer are also projected to the same dimension. Next, we duplicate the image feature maps $N_{p}$ times, and apply an element-wise dot-product between each gaze token and a copy of the image feature map. Finally, these person-specific image features are stacked, and we merge the batch and person dimensions to produce a final output of dimension $(B\\times N_p,d_k,\\frac{H}{k},\\frac{W}{k})$ . This tensor is passed to a fusion module where it is processed by a small residual convnet and added to the output from the previous fusion block. The result goes through another residual convnet, an upsampling stage to double the resolution, and a projection, leading to a tensor of dimension $(B\\times N_p,d_k,\\frac{2H}{k},\\frac{2W}{k})$ . At the end of this process, we get a tensor of dimension $(B\\times N_p,d_{\\mathrm{out}},\\frac{H}{2},\\frac{W}{2})$ , which goes through a convolutional head that predicts the heatmaps by bringing the channel dimension down to 1 and resizing the spatial dimension to that of the gaze heatmap. Finally, we separate the batch and person dimensions such that the final shape of the output is $(B,N_{p},1,H_{\\mathrm{hm}},W_{\\mathrm{hm}})$ . The rationale behind this design is to gather information from different layers and resolutions, which is important for dense prediction tasks. In this case, it is particularly useful for gaze tokens where information from the early layers might retain more scene-independent gaze cues due to their proximity to the gaze encoder.", + "bbox": [ + 76, + 537, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. In-Out prediction", + "text_level": 1, + "bbox": [ + 500, + 537, + 671, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The In-Out classifier head $\\mathcal{O}_{\\mathrm{MLP}}$ consists of an MLP with 7 layers. It is fed the concatenation of input and output gaze tokens to predict a binary in-vs-out label for each person.", + "bbox": [ + 498, + 561, + 890, + 608 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {o} = \\mathcal {O} _ {\\mathrm {M L P}} \\left(\\left[ \\mathbf {x} _ {(L)} ^ {\\mathrm {g}}, \\mathbf {x} ^ {\\mathrm {g}} \\right]\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 621, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.6. Loss and implementation details", + "text_level": 1, + "bbox": [ + 500, + 655, + 784, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train our model using a combination of three losses:", + "bbox": [ + 500, + 679, + 870, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Heatmap Loss $(\\mathcal{L}_{reg})$ . The heatmap loss is the pixel-wise MSE between the GT heatmap and the predicted heatmap: $\\mathcal{L}_{hm} = \\sum_{x,y}^{W_{hm},H_{hm}}||\\mathcal{A}_{x,y}^{\\mathrm{gt}} - \\mathcal{A}_{x,y}^{\\mathrm{pred}}||_2^2$", + "bbox": [ + 500, + 696, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Angular Loss $(\\mathcal{L}_{ang})$ . The angular loss drives the prediction of a normalized gaze direction vector. It maximizes the cosine of the angle between the predicted and ground truth gaze vectors according to: $\\mathcal{L}_{ang} = 1 - < \\mathbf{g}_{\\mathrm{v}}^{gt}, \\mathbf{g}_{\\mathrm{v}}^{pred}>$ where $< a, b >$ denotes the inner product between $a$ and $b$ .", + "bbox": [ + 498, + 744, + 890, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In-Out Loss $(\\mathcal{L}_{io})$ . The in-out loss is the standard binary cross-entropy for in-vs-out prediction.", + "bbox": [ + 498, + 821, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global loss. The final loss is a given by:", + "bbox": [ + 500, + 854, + 769, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {r e g} \\mathcal {L} _ {r e g} + \\lambda_ {a n g} \\mathcal {L} _ {a n g} + \\lambda_ {i o} \\mathcal {L} _ {i o} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 885, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2011", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/13c062be1ece1f26981f9b509f2b2a935b8091ef8b8e33c2eb83cf3ec8147348.jpg", + "image_caption": [ + "Figure 4. Overview of our proposed Conditional DPT decoder." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 854, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 308, + 209, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 334, + 181, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GazeFollow. GazeFollow [30] is an image-based dataset annotated mostly with head bounding boxes, 2D gaze points. Overall, it has around $130\\mathrm{K}$ annotated instances in $122\\mathrm{K}$ images. The test set comprises 4782 gaze instances, each of which is labeled by multiple annotators $(\\sim 10)$ .", + "bbox": [ + 75, + 358, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VideoAttentionTarget. VideoAttentionTarget [7] is a video-based dataset consisting of 1331 clips from 50 TV shows. It is also annotated with the head bounding boxes, 2D gaze points, and in vs out labels. Overall, it contains 164K instances in 71K frames.", + "bbox": [ + 76, + 435, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ChildPlay. ChildPlay [35] is a video dataset consisting of 401 clips from 95 YouTube videos of children engaged in play activities. On top of the standard annotations, the authors extend the in vs out label to include other gaze classes (e.g. gaze shift), which we don't use in this paper. Overall, it contains 257K instances in 120K frames.", + "bbox": [ + 75, + 512, + 470, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 76, + 613, + 173, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use four metrics to evaluate our gaze following models. The first three are AUC, Distance, and AP which are all standard in the literature [7]. Particularly, the AUC adaptation introduced by [7] for datasets with a single point annotation is not very informative, so we decided to replace it with PLAH, which we define below.", + "bbox": [ + 75, + 637, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PLAH. Recently introduced in [35], this metric computes the Precision of looking at people's heads to incorporate semantic information in the evaluation of gaze models. Unlike [35], we consider a prediction positive if the predicted and annotated gaze points fall within the same head box.", + "bbox": [ + 75, + 729, + 468, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Experimental Protocol", + "text_level": 1, + "bbox": [ + 76, + 815, + 287, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Context People. The training of Sharingan relies on processing multiple people at the same time (cf. section 5), but available benchmarks often annotate 1 person per image. To circumvent this problem, we apply an off-the-shelf head", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "detector $^{1}$ trained on the CrowdHuman dataset [31]. We discard detections with a confidence score lower than 0.5 and detections with an IOU score higher than 0.5 with the annotated target person. During training, due to batch constraints, we set $N_{p}^{tr}$ and keep it fixed. For each image, we use the person with the GT annotation and randomly sample $N_{p}^{tr} - 1$ (detected) heads when available, otherwise we use padding for the box and head crop. Incidentally, the loss is computed and propagated solely from the annotated person. At evaluation, for each image $i$ we set $N_{p}^{i}$ to the number of all people in it and process them in a single forward pass using a batch size of 1. Note that $N_{p}^{tr}$ is a property of the training process, not the architecture, and thus doesn't restrict the $N_{p}^{i}$ that can be used during inference. Unless stated otherwise, we use $N_{p}^{tr} = 2$ in all experiments.", + "bbox": [ + 496, + 309, + 893, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Sharingan processes the input scene image and head crop at a resolution of $224 \\times 224$ , while the output heatmap is $64 \\times 64$ . The gaze backbone $\\mathcal{G}$ is a ResNet-18 [13] pretrained on Gaze360 [18], and the transformer encoder is a ViT-base model [9] initialized with weights from a multimodal MAE [2].", + "bbox": [ + 496, + 541, + 893, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. The models are trained for 20 epochs on GazeFollow. For VideoAttentionTarget and ChildPlay, following standard practices, we take the trained GazeFollow model, freeze everything except the gaze decoder and In-Out classifier, and fine-tune them separately for 2 epochs each. We use the AdamW optimizer [23] with a learning rate of $3e - 5$ and a cosine annealing schedule. The fine-tuning uses a learning rate of $1e - 6$ for the gaze decoder and $3e - 4$ for the In-Out classifier. We also make use of Stochastic Weight Averaging [15] to stabilize training on GazeFollow. The loss coefficients are $\\lambda_{reg} = 1000$ and $\\lambda_{ang} = 3$ .", + "bbox": [ + 496, + 636, + 893, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Validation. Since GazeFollow [30] and VideoAttentionTarget [7] do not propose any validation splits, we use the train-val splits proposed by [35]. The best model on the validation set is selected based on the distance metric.", + "bbox": [ + 496, + 806, + 893, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/deepakcrk/yolov5-crowdhuman", + "bbox": [ + 517, + 886, + 792, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2012", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Comparison with the State-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 90, + 398, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We summarize our quantitative results on the GazeFollow and VideoAttentionTarget datasets in Table 1, and on ChildPlay in Table 2, compared to previous works. Our model sets a new state-of-the-art on all 3 datasets on most metrics and outperforms the only comparable multi-person method by 0.013 on the Avg. Dist. metric. Moreover, the only method that comes close to our results is [12], which we slightly outperform on GazeFollow in both Avg. Dist. and AUC. However, unlike Sharingan, this method uses 2 other modalities (i.e. depth, pose), has a very complex and costly training protocol (i.e. modality-specific backbones are pretrained separately for the task), and is resource-intensive during inference (i.e. single-person). Moreover, it generalizes poorly as evidenced by a cross-dataset evaluation (i.e. Dist. 0.113 vs 0.134 on VideoAttentionTarget, and 0.109 vs 0.142 on ChildPlay). Incidentally, the image version of this method is also significantly worse (i.e. 0.134 vs 0.113). In terms of multi-person comparison, we beat [16] by 0.013 and 0.02 in Avg. Dist. and Min. Dist. respectively.", + "bbox": [ + 75, + 112, + 472, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also see similar results on video datasets (VideoAttentionTarget and ChildPlay), and it is interesting to note that the model trained on GazeFollow is achieving remarkable cross-dataset performance without fine-tuning. Surprisingly, it is already improving on its multi-person competitor by a large margin (i.e. Dist. 0.113 vs 0.134), which is a testament to the generalization ability of our model. We also note that the model from [35] is on par with ours on the video datasets. We believe the reason is these datasets have high-quality images compared to GazeFollow, so the depth-based geometric prior in [35] proves useful, but that method remains multimodal and single-person. We provide more details, discussions, and experiments in the supplementary material to further assess robustness and generalization.", + "bbox": [ + 75, + 400, + 472, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablation Experiments", + "text_level": 1, + "bbox": [ + 76, + 619, + 284, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Person Encoding. An important aspect of Sharingan is the way people are encoded into the architecture. Previous methods often represent the head location as a binary mask, and gaze information as a visual attention map [7] or gaze cone [11, 12, 35]. We modify Sharingan to experiment with 3 variants, and opt for single-person training to alleviate the computational cost from some of these formulations: 1. Instead of projecting the box coordinates, we use a head location mask that we tokenize and add the resulting head position embeddings to the image tokens (i.e. Head Mask Embed). 2. Using the same head mask embedding, we tokenize the head crop and append the head tokens to the image ones (i.e. Head Crop Tokens). 3. Instead of processing the head directly, we use the gaze backbone to regress a gaze", + "bbox": [ + 75, + 643, + 472, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "vector which is used to build a gaze cone that we tokenize (i.e. Gaze Cone Tokens). Please note that formulations 2 and 3 double the number of tokens, and we decode the output image tokens into a gaze heatmap using a normal DPT [28]. As we can see from Table 3 (top), the location-aware gaze token is both efficient and performs the best.", + "bbox": [ + 500, + 90, + 893, + 183 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/143acfd47c91d46933f0025d135ef4f91c430de6ffac47a8f1005f777c60217b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAUC ↑Avg. D. ↓Min. D. ↓
Head Mask Embed0.9400.1170.060
Head Crop Tokens0.9330.1380.076
Gaze Cone Tokens0.9340.1330.073
Gaze Token0.9440.1130.057
Token to Heatmap0.6470.3020.234
Dot-Product0.9230.1200.062
Up & Dot-Product0.9340.1160.059
Conditional DPT0.9440.1130.057
", + "bbox": [ + 503, + 193, + 888, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablation results for person encoding (top) and gaze decoding (bottom).", + "bbox": [ + 498, + 392, + 890, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gaze Decoder. We also perform ablations to assess the suitability of our Conditional DPT gaze decoder by comparing it to other baselines as shown in Table 3 (bottom). The first baseline (i.e. Token to Heatmap) regresses a gaze heatmap directly from the output person token using an MLP. This is also the decoding approach undertaken by [39] and [38]. The second baseline (cf. Dot-Product) projects the tokens, then performs a dot-product between each person token and each image token before resizing the output to $64 \\times 64$ to get the final heatmap. The third baseline (cf. Up & Dot-Product) upscales the image representation first, then performs the dot-product. We note that our Conditional DPT outperforms all the other methods, justifying the need for a more sophisticated decoding mechanism.", + "bbox": [ + 496, + 438, + 890, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Angular Loss. Our experiments show that the angular loss doesn't affect the final performance, but we decided to keep it anyway to always have a reliable gaze direction, even when the person is looking outside the frame.", + "bbox": [ + 498, + 651, + 893, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 723, + 612, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Efficiency. Unlike most previous methods, Sharingan's ability to predict the gaze of multiple people at the same time makes it very efficient for real-world applications. Figure 5 (left) shows a fairly large improvement in inference time compared to a lightweight baseline when $N_{p}$ increases. It's worth noting that while Sharingan does most of its processing once within the transformer, the small gaze encoder and decoder themselves are executed for each person independently by combining the batch and person dimensions. This explains why the curve is not constant.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2We omit [38, 39] from the table because their evaluation protocol is different, which makes them incomparable to the rest of the methods. Results are taken from [35].", + "bbox": [ + 76, + 862, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2013", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/00c555a24a697d0170d2fa1b183274af0044ed932210f48349b3bc9f745d8471.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMultiModalityGazeFollowVideoAttentionTarget
AUC ↑Avg. Dist. ↓Min. Dist. ↓PLAH ↑Dist. ↓PLAH ↑AP ↑
Recasens [29]XI0.8780.1900.113————
Chong [6]XI+T0.8960.1870.112—0.171—0.712
Lian [21]XI0.9060.1450.081————
Chong [7]XI+T0.9210.1370.077—0.134—0.853
Fang [11]XI+D+E0.9220.1240.067—0.108—0.896
Fang [11]XI+D————0.124—0.872
Jin [17]XI+D0.9200.1180.063—0.109—0.897
Jin [17]XI0.9090.1370.077————
Tonini [37]XI+D0.9270.141——0.129——
Gupta [12]XI+D+P0.9430.1140.056—0.110—0.879
Gupta [12]XI0.9330.1340.071—0.122—0.864
Bao [3]XI+D+P0.9280.122——0.120—0.869
Hu [14]XI+D+O0.9230.1280.069—0.118—0.881
Tafasca [35]XI+D0.9360.1250.0640.6220.1090.7520.834
Jin [16]✓I0.9190.1260.076—0.134—0.880
\\(Sharingan^†\\)✓I0.9440.1130.0570.6670.1130.748—
Sharingan✓I0.9440.1130.0570.6670.1070.7380.891
", + "bbox": [ + 78, + 88, + 895, + 435 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a3f2e3f7bc80522e4d97203b7081ee6f2e8b3f051b8108c795b90edbceca096e.jpg", + "table_caption": [ + "Table 1. Results of our Sharingan architecture on the GazeFollow and VideoAttentionTarget datasets. The best scores for multi-person models are given in bold, while the best scores in general are underlined. The $\\dagger$ symbol means that the model was trained on GazeFollow and evaluated without fine-tuning. The modality column uses the codes I (image), T (time), D (depth), E (eyes), P (pose), and O (objects)." + ], + "table_footnote": [], + "table_body": "
MethodMultiDist. ↓PLAH ↑AP ↑
Gupta [12]X0.113—0.983
Tafasca [35]X0.1070.5900.986
Sharingan†✓0.1090.600—
Sharingan✓0.1060.6000.990
", + "bbox": [ + 96, + 508, + 452, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on the ChildPlay dataset.", + "bbox": [ + 147, + 626, + 398, + 640 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03f804cb84deb5704863d28c658bb3b0b49042ceef98977b0d68f315e3b896f8.jpg", + "image_caption": [ + "Figure 5. (Left): Comparison of CPU inference time based on $N_{p}$ . (Right): Comparison of single- and multi-person models by $N_{p}^{i}$ during evaluation on GazeFollow." + ], + "image_footnote": [], + "bbox": [ + 78, + 654, + 264, + 806 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd306b9510cf2f4edf3d285b7d31737290e5080fe2e0522590070c51194ba350.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 654, + 455, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of People. One consideration of particular importance in the context of Sharingan is the influence of the number of people on training and evaluation. Table 4 shows that increasing $N_{p}^{tr}$ during training does not influence performance. We believe that since the gaze tokens are interchangeable and play a symmetric role (i.e. no order encoding), as long as $N_{p}^{tr} > 1$ , the model will be forced to learn how to accommodate an arbitrary number of people in the scene during inference by sharing the learned image representations among them. However, the same cannot be said for the single-person setting ( $N_{p}^{tr} = 1$ ), which is fundamentally different from multi-person ( $N_{p}^{tr} > 1$ ). In this case, the model inevitably learns to tailor its image representations to a single person. To illustrate this behavior, we evaluate a single-person ( $N_{p}^{tr} = 1$ ) and a multi-person ( $N_{p}^{tr} = 2$ ) models twice: first using one person ( $N_{p}^{i} = 1$ ) and a second time by processing all people in the scene ( $N_{p}^{i} = all$ ). Figure 5 (right) shows that the multi-person model is perfectly able to do both single-person and multi-person prediction. However, the single-person model experiences a significant degradation when attempting multi-person prediction.", + "bbox": [ + 496, + 530, + 890, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Heatmap vs 2D Point. Gaze following has always been framed as a heatmap prediction task. To the best of our", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2014", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0d3e3c5c55dd2b1d23e16129be368d7eb81f1c17892c269a0f6b55bee6ae50cb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AUCAvg. D. ↓Min. D. ↓
Ntrp=20.9440.1130.057
Ntrp=30.9430.1140.058
Ntrp=40.9420.1130.057
", + "bbox": [ + 120, + 88, + 426, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "knowledge, Lian et al. [21] were the only authors to experiment with a 2D point regression objective in their ablation study. In this section, we explore this formulation further to gain a deeper understanding of the task. To this end, we trained a Sharingan model by replacing the Conditional DPT decoder with a simple MLP to directly regress $(x,y)$ coordinates from the output person token. Since the 2D Point model can only predict a single value, it can't represent the entire distribution over the 2D space of the image. Instead, we believe that it converges to an expectation of this posterior probability. When this distribution is multimodal (i.e. more than one probable gaze target), the expectation becomes unlikely under that posterior distribution. We illustrate this behavior in Figure 6 by comparing the person-specific attention map from the last layer of the encoder of the 2D Point model to the predicted gaze heatmap from the Heatmap variant. It is clear that both models capture the different modes quite well, but the former outputs $(x,y)$ coordinates resembling a weighted average of these modes, which ends up distant from all of them.", + "bbox": [ + 75, + 243, + 470, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This phenomenon leads to the results given in Table 5, showing that the $2D$ point model achieves better average distance and PLAH, but lags behind the Heatmap model in terms of minimum distance and RLAH. This is not surprising since the average distance of the GazeFollow test set is a distance to an average of modes (i.e. multiple annotated points), which loosely resembles the objective that the $2D$ Point model is optimizing for. Consequently, we believe that the average distance should never be the only metric for evaluating gaze following methods, a claim also supported by [3]. On the other hand, positive LAH predictions for the $2D$ Point model mean that it is very confident, which explains the better PLAH value. However, this also means that the number of false negatives will be high, hence the big gap in RLAH because the predicted point will often be slightly off-target when the model is not confident (e.g. marginally outside the head area when looking at a person).", + "bbox": [ + 75, + 547, + 468, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. Sharingan processes both image and gaze tokens simultaneously in the transformer. This means that the same weights operate on both types of tokens, which makes it difficult to understand how the model is combining information. One idea worth exploring is to disentangle scene and person processing, and selectively fuse their informa", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ad760e11f5772874226da3ab0c96ff792e051dad9995a9b70db34c5b4a2421d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 694, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e8c3fb0a92bbcefdad3466c262b701e4b8429c26a2c514cc86177e4f8f064ccb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 890, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16ff9ec24986c8120aaa8d2fb77cac8da92fd5d71aae494bfee4a8b42df59506.jpg", + "image_caption": [ + "Figure 6. Comparison of Heatmap (i.e. left) and 2D Point regression model (i.e. right) models. The heatmap on the right is obtained by computing the attention weights (i.e. last encoder layer) of the person's gaze token with the image tokens." + ], + "image_footnote": [], + "bbox": [ + 501, + 174, + 694, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c4edd67233700f4bf730cd3e1cdf2bc535da07ec8902fe40c09a37783689696.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 174, + 890, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2efa76fd030d22461c21fd272920fd766e7ac2d3d3cee039a90eb38221cc898c.jpg", + "table_caption": [ + "Table 4. Influence of the number of people during training on GazeFollow. The evaluation uses all available people." + ], + "table_footnote": [], + "table_body": "
Avg. D.Min. D.PLAHRLAH
2D Point0.1060.0660.6830.368
Heatmap0.1130.0570.6670.571
", + "bbox": [ + 519, + 335, + 874, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison between the heatmap and the 2D point training objectives on GazeFollow. RLAH is the recall of LAH.", + "bbox": [ + 500, + 412, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion along the architecture. This may help improve the stability of predictions on videos, and allow people to interact together in a more meaningful way.", + "bbox": [ + 498, + 470, + 890, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 536, + 617, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced Sharingan, a novel and efficient transformer-based architecture for gaze target prediction that is designed to support an arbitrary number of people out of the box. Sharingan stands out for its efficiency both in training and inference, delivering SOTA results on public benchmarks. Furthermore, it demonstrates strong generalization when tested on other datasets and naturalistic scenes. We also validated architectural decisions through extensive ablation experiments and discussed key aspects related to the model and the task.", + "bbox": [ + 496, + 563, + 890, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Beyond gaze following, Sharingan's intuitive design makes it suitable for other research areas in human behavior understanding. Specifically, the architecture can be extended to perform a multi-faceted analysis of social scenes by integrating different modalities (e.g. depth, semantics), and producing more outputs (e.g. gestures, interactions). We intend to explore this direction further in future work.", + "bbox": [ + 496, + 715, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research has been supported by the AI4Autism project (Digital Phenotyping of Autism Spectrum Disorders in children, grant agreement no. CR-SII5 202235 / 1) of the Sinergia interdisciplinary program of the SNSF.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2015", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henny Admoni and Brian Scassellati. Social eye gaze in human-robot interaction: a review. Journal of Human-Robot Interaction, 6(1):25-63, 2017. 2", + "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 5", + "[3] Jun Bao, Buyu Liu, and Jun Yu. Escnet: Gaze target detection with the understanding of 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14126-14135, 2022. 7, 8", + "[4] Bridget K Behe, Patricia T Huddleston, Kevin L Childs, Jiaoping Chen, and Iago S Muraro. Seeing through the forest: The gaze path to purchase. Plos one, 15(10):e0240179, 2020. 1", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3", + "[6] Eunji Chong, Nataniel Ruiz, Yongxin Wang, Yun Zhang, Agata Rozga, and James M Rehg. Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In Proceedings of the European conference on computer vision (ECCV), pages 383-398, 2018. 7", + "[7] Eunji Chong, Yongxin Wang, Nataniel Ruiz, and James M Rehg. Detecting attended visual targets in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5406, 2020. 2, 5, 6, 7", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5", + "[10] Lifeng Fan, Yixin Chen, Ping Wei, Wenguan Wang, and Song-Chun Zhu. Inferring shared attention in social scene videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6460-6468, 2018. 2, 5", + "[11] Yi Fang, Jiapeng Tang, Wang Shen, Wei Shen, Xiao Gu, Li Song, and Guangtao Zhai. Dual attention guided gaze target detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11390-11399, 2021. 2, 6, 7", + "[12] Anshul Gupta, Samy Tafasca, and Jean-Marc Odobez. A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5041-5050, 2022. 2, 6, 7", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[14] Zhengxi Hu, Kunxu Zhao, Bohan Zhou, Hang Guo, Shichao Wu, Yuxue Yang, and Jingtai Liu. Gaze target estimation inspired by interactive attention. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8524-8536, 2022. 7", + "[15] P Izmailov, AG Wilson, D Podoprikhin, D Vetrov, and T Garipov. Averaging weights leads to wider optima and better generalization. In 34th Conference on Uncertainty in Artificial Intelligence 2018, UAI 2018, pages 876-885, 2018. 5", + "[16] Tianlei Jin, Zheyuan Lin, Shiqiang Zhu, Wen Wang, and Shunda Hu. Multi-person gaze-following with numerical coordinate regression. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 01-08. IEEE, 2021. 2, 3, 6, 7", + "[17] Tianlei Jin, Qizhi Yu, Shiqiang Zhu, Zheyuan Lin, Jie Ren, Yuanhai Zhou, and Wei Song. Depth-aware gaze-following via auxiliary networks for robotics. Engineering Applications of Artificial Intelligence, 113:104924, 2022. 2, 7", + "[18] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2, 5", + "[19] Nayeon Kim and Hyunsoo Lee. Assessing consumer attention and arousal using eye-tracking technology in virtual retail environment. Frontiers in Psychology, 12:665658, 2021. 1", + "[20] Jing Li, Zejin Chen, Yihao Zhong, Hak-Keung Lam, Junxia Han, Gaoxiang Ouyang, Xiaoli Li, and Honghai Liu. Appearance-based gaze estimation for asd diagnosis. IEEE Transactions on Cybernetics, 52(7):6504-6517, 2022. 2", + "[21] Dongze Lian, Zehao Yu, and Shenghua Gao. Believe it or not, we know what you are looking at! In Asian Conference on Computer Vision, pages 35-50. Springer, 2018. 2, 7, 8, 5", + "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3", + "[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5", + "[24] Manuel Jesús Marin-Jimenez, Andrew Zisserman, Marcin Eichner, and Vittorio Ferrari. Detecting people looking at each other in videos. International Journal of Computer Vision, 106(3):282-296, 2014. 2", + "[25] Manuel J Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net: revisiting people looking at each other in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3477-3485, 2019. 5" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2016", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Skanda Muralidhar, Rémy Siegfried, Jean-Marc Odobez, and Daniel Gatica-Perez. Facing employers and customers: What do gaze and expressions tell about soft skills? In Proceedings of the 17th International Conference on Mobile and Ubiquitous Multimedia, MUM 2018, Cairo, Egypt, November 25-28, 2018, pages 121-126, 2018.", + "[27] Catharine Oertel, Patrik Jonell, Dimosthenis Kontogiorgos, Kenneth Funes Mora, Jean-Marc Odobez, and Joakim Gustafson. Towards an engagement-aware attentive artificial listener for multi-party interactions. Frontiers in Robotics and AI, 8:189, 2021. 2", + "[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 3, 6", + "[29] Adria Recasens*, Aditya Khosla*, Carl Vondrick, and Antonio Torralba. Where are they looking? In Advances in Neural Information Processing Systems (NIPS), 2015. * indicates equal contribution. 2, 7", + "[30] Adria Recasens, Carl Vondrick, Aditya Khosla, and Antonio Torralba. Following gaze in video. In Proceedings of the IEEE International Conference on Computer Vision, pages 1435-1443, 2017. 5", + "[31] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 5", + "[32] Samira Sheikhi and Jean-Marc Odobez. Combining dynamic head pose-gaze mapping with the robot conversational state for attention recognition in human-robot interactions. Pattern Recognition Letters, 66:81–90, 2015. 2", + "[33] R. Siegfried and J.-M. Odobez. Robust unsupervised gaze calibration using conversation and manipulation attention priors. ACM Transactions on Multimedia Computing, Communications, and Applications, 18(1):1-27, 2022. 2", + "[34] S. Tafasca, A. Gupta, N. Kojovic, M. Gelsomini, T. Maillart, M. Papandrea, M. Schaer, and J.-M. Odobez. The ai4autism project: A multimodal and interdisciplinary approach to autism diagnosis and stratification. In Proceeding of the Int. Conference on Multimodal Interaction workshop, Paris, 2023. 2", + "[35] Samy Tafasca, Anshul Gupta, and Jean-Marc Odobez. Childplay: A new benchmark for understanding children's gaze behaviour. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20935-20946, 2023. 2, 5, 6, 7, 1", + "[36] Henri Tomas, Marcus Reyes, Raimarc Dionido, Mark Ty, Jonric Mirando, Joel Casimiro, Rowel Atienza, and Richard Guinto. Goo: A dataset for gaze object prediction in retail environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3125-3133, 2021. 1", + "[37] Francesco Tonini, Cigdem Beyer, and Elisa Ricci. Multimodal across domains gaze target detection. In Proceedings of the 2022 International Conference on Multimodal Interaction, pages 420-431, 2022. 7", + "[38] Francesco Tonini, Nicola Dall'Asen, Cigdem Beyer, and Elisa Ricci. Object-aware gaze target detection. In Proceed-" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF International Conference on Computer Vision, pages 21860-21869, 2023. 1, 3, 6, 5", + "[39] Danyang Tu, Xiongkuo Min, Huiyu Duan, Guodong Guo, Guangtao Zhai, and Wei Shen. End-to-end human-gazetarget detection with transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2192-2200. IEEE, 2022. 1, 3, 6, 5", + "[40] Yu Yu, Gang Liu, and Jean-Marc Odobez. Deep multitask gaze estimation with a constrained landmark-gaze model. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 247 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2017", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_model.json b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..060c19e513bed254910f6955214e2f627ea3f632 --- /dev/null +++ b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "Sharingan: A Transformer Architecture for Multi-Person Gaze Following" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.181, + 0.734, + 0.253 + ], + "angle": 0, + "content": "Samy Tafasca Anshul Gupta Jean-Marc Odobez Idiap Research Institute, Switzerland École Polytechnique Fédérale de Lausanne, Switzerland {stafasca, agupta, odobez}@idiap.ch" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.288, + 0.317, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.289, + 0.655, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.289, + 0.892, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.412, + 0.361, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.412, + 0.609, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.412, + 0.892, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.546, + 0.893, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Predictions of Sharingan on naturalistic images from the internet with different people, activities, interactions, postures, and environments (indoors and outdoors). We provide more qualitative samples in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.587, + 0.313, + 0.602 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Gaze is a powerful form of non-verbal communication that humans develop from an early age. As such, modeling this behavior is an important task that can benefit a broad set of application domains ranging from robotics to sociology. In particular, the gaze following task in computer vision is defined as the prediction of the 2D pixel coordinates where a person in the image is looking. Previous attempts in this area have primarily centered on CNN-based architectures, but they have been constrained by the need to process one person at a time, which proves to be highly inefficient. In this paper, we introduce a novel and effective multi-person transformer-based architecture for gaze prediction. While there exist prior works using transformers for multi-person gaze prediction [38, 39], they use a fixed set of learnable embeddings to decode both the person and its gaze target, which requires a matching step afterward to link the predictions with the annotations. Thus, it is difficult to quantitatively evaluate these methods reliably with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.895, + 0.801 + ], + "angle": 0, + "content": "the available benchmarks, or integrate them into a larger human behavior understanding system. Instead, we are the first to propose a multi-person transformer-based architecture that maintains the original task formulation and ensures control over the people fed as input. Our main contribution lies in encoding the person-specific information into a single controlled token to be processed alongside image tokens and using its output for prediction based on a novel multiscale decoding mechanism. Our new architecture achieves state-of-the-art results on the GazeFollow, VideoAttentionTarget, and ChildPlay datasets and outperforms comparable multi-person architectures with a notable margin. Our code, checkpoints, and data extractions will be made publicly available soon." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.83, + 0.633, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Gaze is an important form of communication and was extensively studied across different domains and applications such as consumer behavior understanding [4, 19, 36], soci" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2008" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.468, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.221, + 0.446, + 0.235 + ], + "angle": 0, + "content": "Figure 2. Illustration of aspects relevant to gaze following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.468, + 0.323 + ], + "angle": 0, + "content": "ology by analyzing different gaze behaviors (e.g. joint attention, eye contact) [10, 24-27], robotics through humanrobot interactions [1, 17, 32] and clinical research for the study of neurodevelopmental disorders [7, 20, 34], etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.468, + 0.476 + ], + "angle": 0, + "content": "Unlike traditional works on gaze analytics proposed by the computer vision community which focused mainly on predicting 3D gaze directions from the eyes [33, 40] or the face [18] of a person, gaze following [29] tackles the task in a more general form where the goal is to infer the 2D location in the image where a person is looking without the need for any assumptions or wearable devices. This formulation is particularly interesting in the context of analyzing social scenes and human interactions given the important role that gaze behavior plays in social dynamics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.468, + 0.627 + ], + "angle": 0, + "content": "It is important to emphasize that this task is very difficult to solve. It hinges upon understanding two aspects simultaneously: (i) the target person (e.g. head pose) to infer a general gaze direction, and (ii) the context of the scene (e.g. social interaction) to identify regions of saliency (cf. Figure 2). Once this is achieved, the rest is essentially a selection process of the gaze target after merging information from the previous steps. This also explains why most architectures have a 2-tower design with one branch to process the scene and a second one to focus on the target person." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.853 + ], + "angle": 0, + "content": "One major issue with this dual-stream approach is that the input instance is a person, not an image. Therefore, it requires multiple forward passes to predict gaze for all people in the same scene, which makes the inference process extremely inefficient. This is made even worse when other modalities are involved (e.g. depth [11, 35] or pose [12]). Moreover, most previous works focus on how to build the person-specific gaze representation, the scene saliency maps, and the fusion between them but pay very little attention to the decoding mechanism that predicts the final gaze heatmap. For example, [7, 11, 17] all use a decoding module based on a few convolutional layers followed by a set of transposed convolutions. The input to this module is often very low dimensional (e.g. \\(7 \\times 7\\)) which forces the prediction to be based on coarse information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this work, we aim to tackle these challenges by addressing the multi-person gaze-following task while maintaining the original problem formulation. To this end," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.363 + ], + "angle": 0, + "content": "we propose Sharingan, a novel, effective, and efficient transformer-based architecture to predict the gaze target of multiple people simultaneously. A key component of this architecture is to represent the person's gaze information by a single gaze token produced by a gaze backbone and processed alongside the image tokens. This is in stark contrast to previous methods that represent intermediate person-specific gaze information as a visual attention map [7] or gaze cone [11, 12, 21, 35]. We show in our ablations that this is not only unnecessary but can also hinder performance in the context of a transformer. Furthermore, we introduce Conditional DPT: a more sophisticated lightweight multi-scale gaze decoding mechanism that helps improve performance by providing a finer-grained understanding of the scene for gaze target selection. This also has the benefit of producing heatmaps that better capture uncertainty when it is difficult to decide where a person is looking (cf. qualitative results in the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Through extensive ablations and evaluations, we find that Sharingan achieves good performance on all public benchmarks, and even transfers well to other gaze-related tasks such as shared attention and mutual gaze." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.64, + 0.453 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.463, + 0.885, + 0.478 + ], + "angle": 0, + "content": "In this section, we present several relevant research topics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Gaze Following. The task of gaze following was first introduced in the seminal work of Recasens et al. [29]. The idea is to predict the pixel-wise 2D location in the image corresponding to where a target person is looking. The main advantage of this formulation is the lack of constraints which allows methods trained this way to generalize to arbitrary settings (i.e. scene properties, camera parameters, image conditions, etc.). It was later extended by Chong et al. [7] to also include the prediction of whether the given person is looking inside the image frame or somewhere outside." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Traditional methods for gaze following [7, 11, 12, 16, 17, 21, 29] typically rely on convolutional networks and follow a 2-tower architecture. The first branch processes the scene image to highlight salient regions, while the second branch processes the head crop of the target person to infer a general gaze direction. A fusion mechanism then combines information from both parts to produce the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.81 + ], + "angle": 0, + "content": "The gaze following task is often framed as the prediction of a gaze heatmap where pixels with high intensity represent spatial areas with higher prediction confidence. We devote a section later to discuss the alternative formulation of regressing the 2D location directly (cf. Section 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Multi-Person Gaze Following. A major downside of the traditional formulation of gaze following is the need for multiple forward passes when predicting the gaze of different people in the same image. This problem motivated the need for architectures that can natively handle the prediction of gaze for multiple people with a single forward" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2009" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "pass. Jin et al. [16] first proposed a simple convolution-based architecture to handle the multi-person setting where a scene backbone computes a fixed person-agnostic feature representation. This is then fused repetitively with head features computed from the different people using another head backbone before decoding each into its corresponding gaze heatmap. Aside from the architectural differences and limited performance, one of the main drawbacks of this method is that the computation for each person is done independently from the others, which ignores potential interactions between people. Recently, Tu et al. [39] and Tonini et al. [38] proposed transformer-based architectures to perform multi-person gaze target prediction. Their methods only take the image as input and simultaneously predict both the head box and gaze target (among others) for every person in the scene. Their work is inspired by the DETR architecture [5], where the task is formulated as a set prediction problem. Instead of reinventing the wheel, our method focuses solely on the gaze prediction part (i.e. given that heads are easily and accurately obtainable using off-the-shelf detectors), and naturally adapts the transformer architecture to the original task formulation by introducing gaze tokens to capture person-specific gaze and head location information and can be directly decoded later into gaze predictions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.303, + 0.487 + ], + "angle": 0, + "content": "3. Sharingan Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Our Sharingan architecture is illustrated in Figure 3. The main idea is to use a transformer that lets scene tokens and person-specific gaze tokens interact within an attention framework to jointly predict the 2D gaze heatmap of each individual. Thus, the inputs are the image and the head crops that we assume are available. We introduce below the different components of this architecture." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.613, + 0.218, + 0.629 + ], + "angle": 0, + "content": "3.1. Image tokens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.713 + ], + "angle": 0, + "content": "We follow a standard ViT architecture to produce image tokens. The input scene image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}\\) goes through a patch projection \\(\\mathcal{P}_{\\mathrm{img}}\\) to produce image tokens that we equip with positional information \\(\\mathbf{x}^{\\mathrm{img}} \\in \\mathbb{R}^{N \\times D}\\), where \\(N\\) is the number of patches, and \\(D\\) is the token dimension." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.208, + 0.739 + ], + "angle": 0, + "content": "3.2. Gaze tokens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.824 + ], + "angle": 0, + "content": "The main purpose of a gaze token is to map the gaze information of a person into a token embedded in the same space as the image tokens, which can interact with scene tokens to select the relevant content for prediction. For simplicity, we first introduce this process for a single person." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Single Person Case. Let \\(\\mathbf{h}_{\\mathrm{crop}} \\in \\mathbb{R}^{h \\times w \\times C}\\) denote the head crop of a person and \\(\\mathbf{h}_{\\mathrm{bbox}} = (x_{\\mathrm{min}}, y_{\\mathrm{min}}, x_{\\mathrm{max}}, y_{\\mathrm{max}}) \\in [0,1]^4\\) her head bounding box. The mapping works as follows. The head crop \\(\\mathbf{h}_{\\mathrm{crop}}\\) is fed to a gaze backbone \\(\\mathcal{G}\\) to produce a gaze embedding \\(\\mathbf{g}^{\\mathrm{emb}} \\in \\mathbb{R}^{d_{\\mathrm{emb}}}\\). This embedding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "is used in two ways. First, it goes through a gaze regressor (i.e. MLP) \\(\\mathcal{O}_{\\mathrm{gv}}\\) to predict a 2D gaze vector \\(\\mathbf{g}_{\\mathrm{v}} = \\mathcal{O}_{\\mathrm{gv}}(\\mathbf{g}^{\\mathrm{emb}})\\). This output is supervised using an angular gaze loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Secondly, the gaze embedding is projected to the token dimension using a learnable linear projection \\(\\mathcal{P}_{\\mathrm{gaze}}\\), resulting in the gaze token \\(\\mathbf{x}^{\\mathrm{emb}} = \\mathcal{P}_{\\mathrm{gaze}}(\\mathbf{g}^{\\mathrm{emb}}) \\in \\mathbb{R}^{D}\\). As we want to incorporate information about the person's location (and size), we also project the head bounding box \\(\\mathbf{h}_{\\mathrm{bbox}}\\) into a bounding box embedding \\(\\mathbf{x}^{\\mathrm{bbox}}\\) using a learnable linear projection \\(\\mathcal{P}_{\\mathrm{bbox}}\\): \\(\\mathbf{x}^{\\mathrm{bbox}} = \\mathcal{P}_{\\mathrm{bbox}}(\\mathbf{h}_{\\mathrm{bbox}}) \\in \\mathbb{R}^{D}\\). Finally, we add this embedding to the gaze token to obtain the final location-aware gaze token:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.286, + 0.892, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\mathrm {g}} = \\mathbf {x} ^ {\\text {e m b}} + \\mathbf {x} ^ {\\text {b b o x}} \\in \\mathbb {R} ^ {D} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.436 + ], + "angle": 0, + "content": "Multi-person case. When \\(N_{p}\\) persons are detected, the architecture will produce a set of \\(N_{p}\\) gaze tokens, following the same process described above for each person. Thus, if \\(\\mathbf{h}_{\\mathrm{bbox}}^i\\) and \\(\\mathbf{h}_{\\mathrm{crop}}^i\\) denote the bounding-box and head crop of person \\(i\\), the above process will generate a gaze token \\(\\mathbf{x}_i^{\\mathrm{g}}\\) for this person. To simplify notation, we will also denote by \\(\\mathbf{x}^{\\mathrm{g}}\\) the set of gaze tokens of all people in the scene, with \\(\\mathbf{x}^{\\mathrm{g}} = \\mathbf{x}_1^{\\mathrm{g}} \\oplus \\ldots \\oplus \\mathbf{x}_{N_p}^{\\mathrm{g}}\\), where \\(\\oplus\\) is the concatenation operator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Modality Encoding. Given the different nature of gaze tokens compared to image tokens, we need to encode modality-specific information to distinguish between them. Rather than using an explicit scheme, in practice we expect this modality information to be captured by the bias terms of the different projection operators \\(\\mathcal{P}_{\\mathrm{gaze}}\\) and \\(\\mathcal{P}_{\\mathrm{img}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.704, + 0.551 + ], + "angle": 0, + "content": "3.3. Transformer Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.892, + 0.651 + ], + "angle": 0, + "content": "The transformer encoder is a standard ViT [8]. It takes as input the concatenation of the scene tokens \\(\\mathbf{x}^{\\mathrm{img}}\\), the gaze token(s) \\(\\mathbf{x}^{\\mathrm{g}}\\), according to \\(\\mathbf{x} = \\mathbf{x}^{\\mathrm{img}} \\oplus \\mathbf{x}^{\\mathrm{g}} \\in \\mathbb{R}^{N_t \\times D}\\), where \\(N_t = N + N_p\\). The set of input tokens goes through a series of \\(L\\) transformer blocks to obtain an output sequence of similar shape, denoted by \\(\\mathbf{x}^{\\mathrm{out}} = \\mathbf{x}^{(L)} \\in \\mathbb{R}^{N_t \\times D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.644, + 0.674 + ], + "angle": 0, + "content": "3.4. Gaze Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.808 + ], + "angle": 0, + "content": "The goal of the gaze decoder \\( \\mathcal{D}_{\\mathrm{gaze}} \\) is to predict a set of gaze heatmaps. Our Conditional DPT (cf. Figure 4) takes four intermediate representations of the image tokens \\( \\mathbf{x}_{(i)}^{\\mathrm{img}} \\) and gaze tokens \\( \\mathbf{x}_{(i)}^{\\mathrm{g}} \\) and combines them progressively at different simulated resolutions, where lower resolutions have more channels and correspond to deeper layers of the encoder. This can be viewed as the isotropic equivalent of a Feature Pyramid Network [22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Our design is inspired by DPT [28], which can only handle decoding image tokens alone. In our case, we need this decoding to be conditioned on each person. To this end, after each block of layers, the image tokens \\(\\mathbf{x}_{(l_k)}^{\\mathrm{img}}\\), \\(k \\in \\{4, 8, 16, 32\\}\\) are reassembled into an image-like representation at resolution \\(\\left(\\frac{H}{k}, \\frac{W}{k}\\right)\\) and dimension \\(d_k\\). The gaze to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.094, + 0.852, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.415, + 0.892, + 0.513 + ], + "angle": 0, + "content": "Figure 3. Overview of our Sharingan architecture. A. The input image is projected into image tokens (red squares). B. The head crops and head box coordinates are processed to generate location and size-aware person-specific tokens (blue squares) as follows. First, the head crop is fed to a gaze backbone to produce a gaze embedding used to (i) predict a normalized 2D gaze vector that is supervised using an angular loss; and (ii) produce a gaze token by projecting it to the token dimension. Second, head bounding box coordinates are projected to obtain an embedding which, added to the gaze token, produces a person gaze token. C. Image tokens and gaze tokens are fed to the transformer encoder, and the output tokens corresponding to input people are all decoded using a Conditional DPT decoder to predict each person's gaze heatmaps. In addition, input and output gaze tokens are concatenated together to predict the in- vs-out label." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.473, + 0.903 + ], + "angle": 0, + "content": "kens \\(\\mathbf{x}_{(l_k)}^{\\mathrm{g}}\\) at that layer are also projected to the same dimension. Next, we duplicate the image feature maps \\(N_{p}\\) times, and apply an element-wise dot-product between each gaze token and a copy of the image feature map. Finally, these person-specific image features are stacked, and we merge the batch and person dimensions to produce a final output of dimension \\((B\\times N_p,d_k,\\frac{H}{k},\\frac{W}{k})\\). This tensor is passed to a fusion module where it is processed by a small residual convnet and added to the output from the previous fusion block. The result goes through another residual convnet, an upsampling stage to double the resolution, and a projection, leading to a tensor of dimension \\((B\\times N_p,d_k,\\frac{2H}{k},\\frac{2W}{k})\\). At the end of this process, we get a tensor of dimension \\((B\\times N_p,d_{\\mathrm{out}},\\frac{H}{2},\\frac{W}{2})\\), which goes through a convolutional head that predicts the heatmaps by bringing the channel dimension down to 1 and resizing the spatial dimension to that of the gaze heatmap. Finally, we separate the batch and person dimensions such that the final shape of the output is \\((B,N_{p},1,H_{\\mathrm{hm}},W_{\\mathrm{hm}})\\). The rationale behind this design is to gather information from different layers and resolutions, which is important for dense prediction tasks. In this case, it is particularly useful for gaze tokens where information from the early layers might retain more scene-independent gaze cues due to their proximity to the gaze encoder." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.673, + 0.554 + ], + "angle": 0, + "content": "3.5. In-Out prediction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.563, + 0.892, + 0.609 + ], + "angle": 0, + "content": "The In-Out classifier head \\(\\mathcal{O}_{\\mathrm{MLP}}\\) consists of an MLP with 7 layers. It is fed the concatenation of input and output gaze tokens to predict a binary in-vs-out label for each person." + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.622, + 0.891, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathbf {o} = \\mathcal {O} _ {\\mathrm {M L P}} \\left(\\left[ \\mathbf {x} _ {(L)} ^ {\\mathrm {g}}, \\mathbf {x} ^ {\\mathrm {g}} \\right]\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.656, + 0.785, + 0.672 + ], + "angle": 0, + "content": "3.6. Loss and implementation details" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.68, + 0.871, + 0.696 + ], + "angle": 0, + "content": "We train our model using a combination of three losses:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.698, + 0.891, + 0.746 + ], + "angle": 0, + "content": "Heatmap Loss \\((\\mathcal{L}_{reg})\\). The heatmap loss is the pixel-wise MSE between the GT heatmap and the predicted heatmap: \\(\\mathcal{L}_{hm} = \\sum_{x,y}^{W_{hm},H_{hm}}||\\mathcal{A}_{x,y}^{\\mathrm{gt}} - \\mathcal{A}_{x,y}^{\\mathrm{pred}}||_2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.746, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Angular Loss \\((\\mathcal{L}_{ang})\\). The angular loss drives the prediction of a normalized gaze direction vector. It maximizes the cosine of the angle between the predicted and ground truth gaze vectors according to: \\(\\mathcal{L}_{ang} = 1 - < \\mathbf{g}_{\\mathrm{v}}^{gt}, \\mathbf{g}_{\\mathrm{v}}^{pred}>\\) where \\(< a, b >\\) denotes the inner product between \\(a\\) and \\(b\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.823, + 0.891, + 0.854 + ], + "angle": 0, + "content": "In-Out Loss \\((\\mathcal{L}_{io})\\). The in-out loss is the standard binary cross-entropy for in-vs-out prediction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.771, + 0.871 + ], + "angle": 0, + "content": "Global loss. The final loss is a given by:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.886, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {r e g} \\mathcal {L} _ {r e g} + \\lambda_ {a n g} \\mathcal {L} _ {a n g} + \\lambda_ {i o} \\mathcal {L} _ {i o} \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2011" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.855, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.27, + 0.673, + 0.284 + ], + "angle": 0, + "content": "Figure 4. Overview of our proposed Conditional DPT decoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.309, + 0.21, + 0.327 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.335, + 0.182, + 0.35 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.435 + ], + "angle": 0, + "content": "GazeFollow. GazeFollow [30] is an image-based dataset annotated mostly with head bounding boxes, 2D gaze points. Overall, it has around \\(130\\mathrm{K}\\) annotated instances in \\(122\\mathrm{K}\\) images. The test set comprises 4782 gaze instances, each of which is labeled by multiple annotators \\((\\sim 10)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.436, + 0.47, + 0.51 + ], + "angle": 0, + "content": "VideoAttentionTarget. VideoAttentionTarget [7] is a video-based dataset consisting of 1331 clips from 50 TV shows. It is also annotated with the head bounding boxes, 2D gaze points, and in vs out labels. Overall, it contains 164K instances in 71K frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.471, + 0.603 + ], + "angle": 0, + "content": "ChildPlay. ChildPlay [35] is a video dataset consisting of 401 clips from 95 YouTube videos of children engaged in play activities. On top of the standard annotations, the authors extend the in vs out label to include other gaze classes (e.g. gaze shift), which we don't use in this paper. Overall, it contains 257K instances in 120K frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.614, + 0.174, + 0.629 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.728 + ], + "angle": 0, + "content": "We use four metrics to evaluate our gaze following models. The first three are AUC, Distance, and AP which are all standard in the literature [7]. Particularly, the AUC adaptation introduced by [7] for datasets with a single point annotation is not very informative, so we decided to replace it with PLAH, which we define below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.806 + ], + "angle": 0, + "content": "PLAH. Recently introduced in [35], this metric computes the Precision of looking at people's heads to incorporate semantic information in the evaluation of gaze models. Unlike [35], we consider a prediction positive if the predicted and annotated gaze points fall within the same head box." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.288, + 0.832 + ], + "angle": 0, + "content": "4.3. Experimental Protocol" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Context People. The training of Sharingan relies on processing multiple people at the same time (cf. section 5), but available benchmarks often annotate 1 person per image. To circumvent this problem, we apply an off-the-shelf head" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.895, + 0.539 + ], + "angle": 0, + "content": "detector\\(^{1}\\) trained on the CrowdHuman dataset [31]. We discard detections with a confidence score lower than 0.5 and detections with an IOU score higher than 0.5 with the annotated target person. During training, due to batch constraints, we set \\(N_{p}^{tr}\\) and keep it fixed. For each image, we use the person with the GT annotation and randomly sample \\(N_{p}^{tr} - 1\\) (detected) heads when available, otherwise we use padding for the box and head crop. Incidentally, the loss is computed and propagated solely from the annotated person. At evaluation, for each image \\(i\\) we set \\(N_{p}^{i}\\) to the number of all people in it and process them in a single forward pass using a batch size of 1. Note that \\(N_{p}^{tr}\\) is a property of the training process, not the architecture, and thus doesn't restrict the \\(N_{p}^{i}\\) that can be used during inference. Unless stated otherwise, we use \\(N_{p}^{tr} = 2\\) in all experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.895, + 0.633 + ], + "angle": 0, + "content": "Implementation Details. Sharingan processes the input scene image and head crop at a resolution of \\(224 \\times 224\\), while the output heatmap is \\(64 \\times 64\\). The gaze backbone \\(\\mathcal{G}\\) is a ResNet-18 [13] pretrained on Gaze360 [18], and the transformer encoder is a ViT-base model [9] initialized with weights from a multimodal MAE [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.895, + 0.804 + ], + "angle": 0, + "content": "Training. The models are trained for 20 epochs on GazeFollow. For VideoAttentionTarget and ChildPlay, following standard practices, we take the trained GazeFollow model, freeze everything except the gaze decoder and In-Out classifier, and fine-tune them separately for 2 epochs each. We use the AdamW optimizer [23] with a learning rate of \\(3e - 5\\) and a cosine annealing schedule. The fine-tuning uses a learning rate of \\(1e - 6\\) for the gaze decoder and \\(3e - 4\\) for the In-Out classifier. We also make use of Stochastic Weight Averaging [15] to stabilize training on GazeFollow. The loss coefficients are \\(\\lambda_{reg} = 1000\\) and \\(\\lambda_{ang} = 3\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Validation. Since GazeFollow [30] and VideoAttentionTarget [7] do not propose any validation splits, we use the train-val splits proposed by [35]. The best model on the validation set is selected based on the distance metric." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.794, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/deepakcrk/yolov5-crowdhuman" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2012" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.4, + 0.108 + ], + "angle": 0, + "content": "4.4. Comparison with the State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.473, + 0.401 + ], + "angle": 0, + "content": "We summarize our quantitative results on the GazeFollow and VideoAttentionTarget datasets in Table 1, and on ChildPlay in Table 2, compared to previous works. Our model sets a new state-of-the-art on all 3 datasets on most metrics and outperforms the only comparable multi-person method by 0.013 on the Avg. Dist. metric. Moreover, the only method that comes close to our results is [12], which we slightly outperform on GazeFollow in both Avg. Dist. and AUC. However, unlike Sharingan, this method uses 2 other modalities (i.e. depth, pose), has a very complex and costly training protocol (i.e. modality-specific backbones are pretrained separately for the task), and is resource-intensive during inference (i.e. single-person). Moreover, it generalizes poorly as evidenced by a cross-dataset evaluation (i.e. Dist. 0.113 vs 0.134 on VideoAttentionTarget, and 0.109 vs 0.142 on ChildPlay). Incidentally, the image version of this method is also significantly worse (i.e. 0.134 vs 0.113). In terms of multi-person comparison, we beat [16] by 0.013 and 0.02 in Avg. Dist. and Min. Dist. respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.473, + 0.614 + ], + "angle": 0, + "content": "We also see similar results on video datasets (VideoAttentionTarget and ChildPlay), and it is interesting to note that the model trained on GazeFollow is achieving remarkable cross-dataset performance without fine-tuning. Surprisingly, it is already improving on its multi-person competitor by a large margin (i.e. Dist. 0.113 vs 0.134), which is a testament to the generalization ability of our model. We also note that the model from [35] is on par with ours on the video datasets. We believe the reason is these datasets have high-quality images compared to GazeFollow, so the depth-based geometric prior in [35] proves useful, but that method remains multimodal and single-person. We provide more details, discussions, and experiments in the supplementary material to further assess robustness and generalization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.285, + 0.637 + ], + "angle": 0, + "content": "4.5. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.858 + ], + "angle": 0, + "content": "Person Encoding. An important aspect of Sharingan is the way people are encoded into the architecture. Previous methods often represent the head location as a binary mask, and gaze information as a visual attention map [7] or gaze cone [11, 12, 35]. We modify Sharingan to experiment with 3 variants, and opt for single-person training to alleviate the computational cost from some of these formulations: 1. Instead of projecting the box coordinates, we use a head location mask that we tokenize and add the resulting head position embeddings to the image tokens (i.e. Head Mask Embed). 2. Using the same head mask embedding, we tokenize the head crop and append the head tokens to the image ones (i.e. Head Crop Tokens). 3. Instead of processing the head directly, we use the gaze backbone to regress a gaze" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "vector which is used to build a gaze cone that we tokenize (i.e. Gaze Cone Tokens). Please note that formulations 2 and 3 double the number of tokens, and we decode the output image tokens into a gaze heatmap using a normal DPT [28]. As we can see from Table 3 (top), the location-aware gaze token is both efficient and performs the best." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.194, + 0.89, + 0.383 + ], + "angle": 0, + "content": "
MethodAUC ↑Avg. D. ↓Min. D. ↓
Head Mask Embed0.9400.1170.060
Head Crop Tokens0.9330.1380.076
Gaze Cone Tokens0.9340.1330.073
Gaze Token0.9440.1130.057
Token to Heatmap0.6470.3020.234
Dot-Product0.9230.1200.062
Up & Dot-Product0.9340.1160.059
Conditional DPT0.9440.1130.057
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.393, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Table 3. Ablation results for person encoding (top) and gaze decoding (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Gaze Decoder. We also perform ablations to assess the suitability of our Conditional DPT gaze decoder by comparing it to other baselines as shown in Table 3 (bottom). The first baseline (i.e. Token to Heatmap) regresses a gaze heatmap directly from the output person token using an MLP. This is also the decoding approach undertaken by [39] and [38]. The second baseline (cf. Dot-Product) projects the tokens, then performs a dot-product between each person token and each image token before resizing the output to \\(64 \\times 64\\) to get the final heatmap. The third baseline (cf. Up & Dot-Product) upscales the image representation first, then performs the dot-product. We note that our Conditional DPT outperforms all the other methods, justifying the need for a more sophisticated decoding mechanism." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.652, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Angular Loss. Our experiments show that the angular loss doesn't affect the final performance, but we decided to keep it anyway to always have a reliable gaze direction, even when the person is looking outside the frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.614, + 0.74 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Model Efficiency. Unlike most previous methods, Sharingan's ability to predict the gaze of multiple people at the same time makes it very efficient for real-world applications. Figure 5 (left) shows a fairly large improvement in inference time compared to a lightweight baseline when \\( N_{p} \\) increases. It's worth noting that while Sharingan does most of its processing once within the transformer, the small gaze encoder and decoder themselves are executed for each person independently by combining the batch and person dimensions. This explains why the curve is not constant." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.468, + 0.901 + ], + "angle": 0, + "content": "2We omit [38, 39] from the table because their evaluation protocol is different, which makes them incomparable to the rest of the methods. Results are taken from [35]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2013" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.897, + 0.436 + ], + "angle": 0, + "content": "
MethodMultiModalityGazeFollowVideoAttentionTarget
AUC ↑Avg. Dist. ↓Min. Dist. ↓PLAH ↑Dist. ↓PLAH ↑AP ↑
Recasens [29]XI0.8780.1900.113————
Chong [6]XI+T0.8960.1870.112—0.171—0.712
Lian [21]XI0.9060.1450.081————
Chong [7]XI+T0.9210.1370.077—0.134—0.853
Fang [11]XI+D+E0.9220.1240.067—0.108—0.896
Fang [11]XI+D————0.124—0.872
Jin [17]XI+D0.9200.1180.063—0.109—0.897
Jin [17]XI0.9090.1370.077————
Tonini [37]XI+D0.9270.141——0.129——
Gupta [12]XI+D+P0.9430.1140.056—0.110—0.879
Gupta [12]XI0.9330.1340.071—0.122—0.864
Bao [3]XI+D+P0.9280.122——0.120—0.869
Hu [14]XI+D+O0.9230.1280.069—0.118—0.881
Tafasca [35]XI+D0.9360.1250.0640.6220.1090.7520.834
Jin [16]✓I0.9190.1260.076—0.134—0.880
\\(Sharingan^†\\)✓I0.9440.1130.0570.6670.1130.748—
Sharingan✓I0.9440.1130.0570.6670.1070.7380.891
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.444, + 0.895, + 0.489 + ], + "angle": 0, + "content": "Table 1. Results of our Sharingan architecture on the GazeFollow and VideoAttentionTarget datasets. The best scores for multi-person models are given in bold, while the best scores in general are underlined. The \\(\\dagger\\) symbol means that the model was trained on GazeFollow and evaluated without fine-tuning. The modality column uses the codes I (image), T (time), D (depth), E (eyes), P (pose), and O (objects)." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.51, + 0.454, + 0.617 + ], + "angle": 0, + "content": "
MethodMultiDist. ↓PLAH ↑AP ↑
Gupta [12]X0.113—0.983
Tafasca [35]X0.1070.5900.986
Sharingan†✓0.1090.600—
Sharingan✓0.1060.6000.990
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.627, + 0.399, + 0.641 + ], + "angle": 0, + "content": "Table 2. Results on the ChildPlay dataset." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.655, + 0.266, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.655, + 0.457, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.816, + 0.47, + 0.859 + ], + "angle": 0, + "content": "Figure 5. (Left): Comparison of CPU inference time based on \\( N_{p} \\). (Right): Comparison of single- and multi-person models by \\( N_{p}^{i} \\) during evaluation on GazeFollow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Number of People. One consideration of particular importance in the context of Sharingan is the influence of the number of people on training and evaluation. Table 4 shows that increasing \\( N_{p}^{tr} \\) during training does not influence performance. We believe that since the gaze tokens are interchangeable and play a symmetric role (i.e. no order encoding), as long as \\( N_{p}^{tr} > 1 \\), the model will be forced to learn how to accommodate an arbitrary number of people in the scene during inference by sharing the learned image representations among them. However, the same cannot be said for the single-person setting (\\( N_{p}^{tr} = 1 \\)), which is fundamentally different from multi-person (\\( N_{p}^{tr} > 1 \\)). In this case, the model inevitably learns to tailor its image representations to a single person. To illustrate this behavior, we evaluate a single-person (\\( N_{p}^{tr} = 1 \\)) and a multi-person (\\( N_{p}^{tr} = 2 \\)) models twice: first using one person (\\( N_{p}^{i} = 1 \\)) and a second time by processing all people in the scene (\\( N_{p}^{i} = all \\)). Figure 5 (right) shows that the multi-person model is perfectly able to do both single-person and multi-person prediction. However, the single-person model experiences a significant degradation when attempting multi-person prediction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Heatmap vs 2D Point. Gaze following has always been framed as a heatmap prediction task. To the best of our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2014" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.428, + 0.172 + ], + "angle": 0, + "content": "
AUCAvg. D. ↓Min. D. ↓
Ntrp=20.9440.1130.057
Ntrp=30.9430.1140.058
Ntrp=40.9420.1130.057
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.183, + 0.47, + 0.211 + ], + "angle": 0, + "content": "Table 4. Influence of the number of people during training on GazeFollow. The evaluation uses all available people." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.471, + 0.546 + ], + "angle": 0, + "content": "knowledge, Lian et al. [21] were the only authors to experiment with a 2D point regression objective in their ablation study. In this section, we explore this formulation further to gain a deeper understanding of the task. To this end, we trained a Sharingan model by replacing the Conditional DPT decoder with a simple MLP to directly regress \\((x,y)\\) coordinates from the output person token. Since the 2D Point model can only predict a single value, it can't represent the entire distribution over the 2D space of the image. Instead, we believe that it converges to an expectation of this posterior probability. When this distribution is multimodal (i.e. more than one probable gaze target), the expectation becomes unlikely under that posterior distribution. We illustrate this behavior in Figure 6 by comparing the person-specific attention map from the last layer of the encoder of the 2D Point model to the predicted gaze heatmap from the Heatmap variant. It is clear that both models capture the different modes quite well, but the former outputs \\((x,y)\\) coordinates resembling a weighted average of these modes, which ends up distant from all of them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.807 + ], + "angle": 0, + "content": "This phenomenon leads to the results given in Table 5, showing that the \\(2D\\) point model achieves better average distance and PLAH, but lags behind the Heatmap model in terms of minimum distance and RLAH. This is not surprising since the average distance of the GazeFollow test set is a distance to an average of modes (i.e. multiple annotated points), which loosely resembles the objective that the \\(2D\\) Point model is optimizing for. Consequently, we believe that the average distance should never be the only metric for evaluating gaze following methods, a claim also supported by [3]. On the other hand, positive LAH predictions for the \\(2D\\) Point model mean that it is very confident, which explains the better PLAH value. However, this also means that the number of false negatives will be high, hence the big gap in RLAH because the predicted point will often be slightly off-target when the model is not confident (e.g. marginally outside the head area when looking at a person)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Limitation. Sharingan processes both image and gaze tokens simultaneously in the transformer. This means that the same weights operate on both types of tokens, which makes it difficult to understand how the model is combining information. One idea worth exploring is to disentangle scene and person processing, and selectively fuse their informa" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.695, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.892, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.175, + 0.695, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.175, + 0.892, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Figure 6. Comparison of Heatmap (i.e. left) and 2D Point regression model (i.e. right) models. The heatmap on the right is obtained by computing the attention weights (i.e. last encoder layer) of the person's gaze token with the image tokens." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.336, + 0.875, + 0.404 + ], + "angle": 0, + "content": "
Avg. D.Min. D.PLAHRLAH
2D Point0.1060.0660.6830.368
Heatmap0.1130.0570.6670.571
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.413, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Table 5. Comparison between the heatmap and the 2D point training objectives on GazeFollow. RLAH is the recall of LAH." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.518 + ], + "angle": 0, + "content": "tion along the architecture. This may help improve the stability of predictions on videos, and allow people to interact together in a more meaningful way." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.619, + 0.552 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.715 + ], + "angle": 0, + "content": "In this paper, we introduced Sharingan, a novel and efficient transformer-based architecture for gaze target prediction that is designed to support an arbitrary number of people out of the box. Sharingan stands out for its efficiency both in training and inference, delivering SOTA results on public benchmarks. Furthermore, it demonstrates strong generalization when tested on other datasets and naturalistic scenes. We also validated architectural decisions through extensive ablation experiments and discussed key aspects related to the model and the task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Beyond gaze following, Sharingan's intuitive design makes it suitable for other research areas in human behavior understanding. Specifically, the architecture can be extended to perform a multi-faceted analysis of social scenes by integrating different modalities (e.g. depth, semantics), and producing more outputs (e.g. gestures, interactions). We intend to explore this direction further in future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This research has been supported by the AI4Autism project (Digital Phenotyping of Autism Spectrum Disorders in children, grant agreement no. CR-SII5 202235 / 1) of the Sinergia interdisciplinary program of the SNSF." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2015" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Henny Admoni and Brian Scassellati. Social eye gaze in human-robot interaction: a review. Journal of Human-Robot Interaction, 6(1):25-63, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[3] Jun Bao, Buyu Liu, and Jun Yu. Escnet: Gaze target detection with the understanding of 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14126-14135, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[4] Bridget K Behe, Patricia T Huddleston, Kevin L Childs, Jiaoping Chen, and Iago S Muraro. Seeing through the forest: The gaze path to purchase. Plos one, 15(10):e0240179, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.472, + 0.424 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.472, + 0.507 + ], + "angle": 0, + "content": "[6] Eunji Chong, Nataniel Ruiz, Yongxin Wang, Yun Zhang, Agata Rozga, and James M Rehg. Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In Proceedings of the European conference on computer vision (ECCV), pages 383-398, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[7] Eunji Chong, Yongxin Wang, Nataniel Ruiz, and James M Rehg. Detecting attended visual targets in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5406, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[10] Lifeng Fan, Yixin Chen, Ping Wei, Wenguan Wang, and Song-Chun Zhu. Inferring shared attention in social scene videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6460-6468, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[11] Yi Fang, Jiapeng Tang, Wang Shen, Wei Shen, Xiao Gu, Li Song, and Guangtao Zhai. Dual attention guided gaze target detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11390-11399, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Anshul Gupta, Samy Tafasca, and Jean-Marc Odobez. A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5041-5050, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[14] Zhengxi Hu, Kunxu Zhao, Bohan Zhou, Hang Guo, Shichao Wu, Yuxue Yang, and Jingtai Liu. Gaze target estimation inspired by interactive attention. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8524-8536, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[15] P Izmailov, AG Wilson, D Podoprikhin, D Vetrov, and T Garipov. Averaging weights leads to wider optima and better generalization. In 34th Conference on Uncertainty in Artificial Intelligence 2018, UAI 2018, pages 876-885, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[16] Tianlei Jin, Zheyuan Lin, Shiqiang Zhu, Wen Wang, and Shunda Hu. Multi-person gaze-following with numerical coordinate regression. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 01-08. IEEE, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[17] Tianlei Jin, Qizhi Yu, Shiqiang Zhu, Zheyuan Lin, Jie Ren, Yuanhai Zhou, and Wei Song. Depth-aware gaze-following via auxiliary networks for robotics. Engineering Applications of Artificial Intelligence, 113:104924, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[18] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[19] Nayeon Kim and Hyunsoo Lee. Assessing consumer attention and arousal using eye-tracking technology in virtual retail environment. Frontiers in Psychology, 12:665658, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[20] Jing Li, Zejin Chen, Yihao Zhong, Hak-Keung Lam, Junxia Han, Gaoxiang Ouyang, Xiaoli Li, and Honghai Liu. Appearance-based gaze estimation for asd diagnosis. IEEE Transactions on Cybernetics, 52(7):6504-6517, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[21] Dongze Lian, Zehao Yu, and Shenghua Gao. Believe it or not, we know what you are looking at! In Asian Conference on Computer Vision, pages 35-50. Springer, 2018. 2, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Manuel Jesús Marin-Jimenez, Andrew Zisserman, Marcin Eichner, and Vittorio Ferrari. Detecting people looking at each other in videos. International Journal of Computer Vision, 106(3):282-296, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[25] Manuel J Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net: revisiting people looking at each other in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3477-3485, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2016" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[26] Skanda Muralidhar, Rémy Siegfried, Jean-Marc Odobez, and Daniel Gatica-Perez. Facing employers and customers: What do gaze and expressions tell about soft skills? In Proceedings of the 17th International Conference on Mobile and Ubiquitous Multimedia, MUM 2018, Cairo, Egypt, November 25-28, 2018, pages 121-126, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[27] Catharine Oertel, Patrik Jonell, Dimosthenis Kontogiorgos, Kenneth Funes Mora, Jean-Marc Odobez, and Joakim Gustafson. Towards an engagement-aware attentive artificial listener for multi-party interactions. Frontiers in Robotics and AI, 8:189, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[29] Adria Recasens*, Aditya Khosla*, Carl Vondrick, and Antonio Torralba. Where are they looking? In Advances in Neural Information Processing Systems (NIPS), 2015. * indicates equal contribution. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[30] Adria Recasens, Carl Vondrick, Aditya Khosla, and Antonio Torralba. Following gaze in video. In Proceedings of the IEEE International Conference on Computer Vision, pages 1435-1443, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[31] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[32] Samira Sheikhi and Jean-Marc Odobez. Combining dynamic head pose-gaze mapping with the robot conversational state for attention recognition in human-robot interactions. Pattern Recognition Letters, 66:81–90, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[33] R. Siegfried and J.-M. Odobez. Robust unsupervised gaze calibration using conversation and manipulation attention priors. ACM Transactions on Multimedia Computing, Communications, and Applications, 18(1):1-27, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[34] S. Tafasca, A. Gupta, N. Kojovic, M. Gelsomini, T. Maillart, M. Papandrea, M. Schaer, and J.-M. Odobez. The ai4autism project: A multimodal and interdisciplinary approach to autism diagnosis and stratification. In Proceeding of the Int. Conference on Multimodal Interaction workshop, Paris, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[35] Samy Tafasca, Anshul Gupta, and Jean-Marc Odobez. Childplay: A new benchmark for understanding children's gaze behaviour. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20935-20946, 2023. 2, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[36] Henri Tomas, Marcus Reyes, Raimarc Dionido, Mark Ty, Jonric Mirando, Joel Casimiro, Rowel Atienza, and Richard Guinto. Goo: A dataset for gaze object prediction in retail environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3125-3133, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[37] Francesco Tonini, Cigdem Beyer, and Elisa Ricci. Multimodal across domains gaze target detection. In Proceedings of the 2022 International Conference on Multimodal Interaction, pages 420-431, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[38] Francesco Tonini, Nicola Dall'Asen, Cigdem Beyer, and Elisa Ricci. Object-aware gaze target detection. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ings of the IEEE/CVF International Conference on Computer Vision, pages 21860-21869, 2023. 1, 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[39] Danyang Tu, Xiongkuo Min, Huiyu Duan, Guodong Guo, Guangtao Zhai, and Wei Shen. End-to-end human-gazetarget detection with transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2192-2200. IEEE, 2022. 1, 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[40] Yu Yu, Gang Liu, and Jean-Marc Odobez. Deep multitask gaze estimation with a constrained landmark-gaze model. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2017" + } + ] +] \ No newline at end of file diff --git a/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_origin.pdf b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76859d7056d6acebf699226c165d8aacc250bfa1 --- /dev/null +++ b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/df6417d4-023a-4e78-a931-75b2f738abf7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cd6f2ba340d14c9ef4a78e2850d603959053a4246dd31e70b08d45565877e17 +size 9054906 diff --git a/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/full.md b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f20e706a2260c6de4005be2d9608b7330dc43a6 --- /dev/null +++ b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/full.md @@ -0,0 +1,275 @@ +# Sharingan: A Transformer Architecture for Multi-Person Gaze Following + +Samy Tafasca Anshul Gupta Jean-Marc Odobez Idiap Research Institute, Switzerland École Polytechnique Fédérale de Lausanne, Switzerland {stafasca, agupta, odobez}@idiap.ch + +![](images/1f31d57e4dd09686e237ee1ef97f6c1f37f60b39a2b1a85d8bb06b848286d2c6.jpg) + +![](images/989b488159efff31f59072335aa643d96bdc887478fff84ccb4a7c45e417a723.jpg) + +![](images/dd20632f1bc399cfd863e8a463ed244cf2a229ad3a2b88e1715ba2a71a44faf8.jpg) + +![](images/e2c396e6a6b18277e75d65260aca90dc22199cc609ac48bde274451a140d7c20.jpg) +Figure 1. Predictions of Sharingan on naturalistic images from the internet with different people, activities, interactions, postures, and environments (indoors and outdoors). We provide more qualitative samples in the supplementary material. + +![](images/2c1241e016785f2cc88308d28fd101a2defd70ca443e4625dcf319653ab1df70.jpg) + +![](images/bbc3aed12cb3eda5443dc888f733066640a2dac7b542c62dcb64e0182e2bf9f8.jpg) + +# Abstract + +Gaze is a powerful form of non-verbal communication that humans develop from an early age. As such, modeling this behavior is an important task that can benefit a broad set of application domains ranging from robotics to sociology. In particular, the gaze following task in computer vision is defined as the prediction of the 2D pixel coordinates where a person in the image is looking. Previous attempts in this area have primarily centered on CNN-based architectures, but they have been constrained by the need to process one person at a time, which proves to be highly inefficient. In this paper, we introduce a novel and effective multi-person transformer-based architecture for gaze prediction. While there exist prior works using transformers for multi-person gaze prediction [38, 39], they use a fixed set of learnable embeddings to decode both the person and its gaze target, which requires a matching step afterward to link the predictions with the annotations. Thus, it is difficult to quantitatively evaluate these methods reliably with + +the available benchmarks, or integrate them into a larger human behavior understanding system. Instead, we are the first to propose a multi-person transformer-based architecture that maintains the original task formulation and ensures control over the people fed as input. Our main contribution lies in encoding the person-specific information into a single controlled token to be processed alongside image tokens and using its output for prediction based on a novel multiscale decoding mechanism. Our new architecture achieves state-of-the-art results on the GazeFollow, VideoAttentionTarget, and ChildPlay datasets and outperforms comparable multi-person architectures with a notable margin. Our code, checkpoints, and data extractions will be made publicly available soon. + +# 1. Introduction + +Gaze is an important form of communication and was extensively studied across different domains and applications such as consumer behavior understanding [4, 19, 36], soci + +![](images/6cfe831da9f5f7d6da0f6136e99a53f846c71a1cc33530cda5032e46d98bd8c1.jpg) +Figure 2. Illustration of aspects relevant to gaze following. + +ology by analyzing different gaze behaviors (e.g. joint attention, eye contact) [10, 24-27], robotics through humanrobot interactions [1, 17, 32] and clinical research for the study of neurodevelopmental disorders [7, 20, 34], etc. + +Unlike traditional works on gaze analytics proposed by the computer vision community which focused mainly on predicting 3D gaze directions from the eyes [33, 40] or the face [18] of a person, gaze following [29] tackles the task in a more general form where the goal is to infer the 2D location in the image where a person is looking without the need for any assumptions or wearable devices. This formulation is particularly interesting in the context of analyzing social scenes and human interactions given the important role that gaze behavior plays in social dynamics. + +It is important to emphasize that this task is very difficult to solve. It hinges upon understanding two aspects simultaneously: (i) the target person (e.g. head pose) to infer a general gaze direction, and (ii) the context of the scene (e.g. social interaction) to identify regions of saliency (cf. Figure 2). Once this is achieved, the rest is essentially a selection process of the gaze target after merging information from the previous steps. This also explains why most architectures have a 2-tower design with one branch to process the scene and a second one to focus on the target person. + +One major issue with this dual-stream approach is that the input instance is a person, not an image. Therefore, it requires multiple forward passes to predict gaze for all people in the same scene, which makes the inference process extremely inefficient. This is made even worse when other modalities are involved (e.g. depth [11, 35] or pose [12]). Moreover, most previous works focus on how to build the person-specific gaze representation, the scene saliency maps, and the fusion between them but pay very little attention to the decoding mechanism that predicts the final gaze heatmap. For example, [7, 11, 17] all use a decoding module based on a few convolutional layers followed by a set of transposed convolutions. The input to this module is often very low dimensional (e.g. $7 \times 7$ ) which forces the prediction to be based on coarse information. + +In this work, we aim to tackle these challenges by addressing the multi-person gaze-following task while maintaining the original problem formulation. To this end, + +we propose Sharingan, a novel, effective, and efficient transformer-based architecture to predict the gaze target of multiple people simultaneously. A key component of this architecture is to represent the person's gaze information by a single gaze token produced by a gaze backbone and processed alongside the image tokens. This is in stark contrast to previous methods that represent intermediate person-specific gaze information as a visual attention map [7] or gaze cone [11, 12, 21, 35]. We show in our ablations that this is not only unnecessary but can also hinder performance in the context of a transformer. Furthermore, we introduce Conditional DPT: a more sophisticated lightweight multi-scale gaze decoding mechanism that helps improve performance by providing a finer-grained understanding of the scene for gaze target selection. This also has the benefit of producing heatmaps that better capture uncertainty when it is difficult to decide where a person is looking (cf. qualitative results in the supplementary material). + +Through extensive ablations and evaluations, we find that Sharingan achieves good performance on all public benchmarks, and even transfers well to other gaze-related tasks such as shared attention and mutual gaze. + +# 2. Related Work + +In this section, we present several relevant research topics. + +Gaze Following. The task of gaze following was first introduced in the seminal work of Recasens et al. [29]. The idea is to predict the pixel-wise 2D location in the image corresponding to where a target person is looking. The main advantage of this formulation is the lack of constraints which allows methods trained this way to generalize to arbitrary settings (i.e. scene properties, camera parameters, image conditions, etc.). It was later extended by Chong et al. [7] to also include the prediction of whether the given person is looking inside the image frame or somewhere outside. + +Traditional methods for gaze following [7, 11, 12, 16, 17, 21, 29] typically rely on convolutional networks and follow a 2-tower architecture. The first branch processes the scene image to highlight salient regions, while the second branch processes the head crop of the target person to infer a general gaze direction. A fusion mechanism then combines information from both parts to produce the final prediction. + +The gaze following task is often framed as the prediction of a gaze heatmap where pixels with high intensity represent spatial areas with higher prediction confidence. We devote a section later to discuss the alternative formulation of regressing the 2D location directly (cf. Section 5). + +Multi-Person Gaze Following. A major downside of the traditional formulation of gaze following is the need for multiple forward passes when predicting the gaze of different people in the same image. This problem motivated the need for architectures that can natively handle the prediction of gaze for multiple people with a single forward + +pass. Jin et al. [16] first proposed a simple convolution-based architecture to handle the multi-person setting where a scene backbone computes a fixed person-agnostic feature representation. This is then fused repetitively with head features computed from the different people using another head backbone before decoding each into its corresponding gaze heatmap. Aside from the architectural differences and limited performance, one of the main drawbacks of this method is that the computation for each person is done independently from the others, which ignores potential interactions between people. Recently, Tu et al. [39] and Tonini et al. [38] proposed transformer-based architectures to perform multi-person gaze target prediction. Their methods only take the image as input and simultaneously predict both the head box and gaze target (among others) for every person in the scene. Their work is inspired by the DETR architecture [5], where the task is formulated as a set prediction problem. Instead of reinventing the wheel, our method focuses solely on the gaze prediction part (i.e. given that heads are easily and accurately obtainable using off-the-shelf detectors), and naturally adapts the transformer architecture to the original task formulation by introducing gaze tokens to capture person-specific gaze and head location information and can be directly decoded later into gaze predictions. + +# 3. Sharingan Architecture + +Our Sharingan architecture is illustrated in Figure 3. The main idea is to use a transformer that lets scene tokens and person-specific gaze tokens interact within an attention framework to jointly predict the 2D gaze heatmap of each individual. Thus, the inputs are the image and the head crops that we assume are available. We introduce below the different components of this architecture. + +# 3.1. Image tokens + +We follow a standard ViT architecture to produce image tokens. The input scene image $\mathbf{I} \in \mathbb{R}^{H \times W \times C}$ goes through a patch projection $\mathcal{P}_{\mathrm{img}}$ to produce image tokens that we equip with positional information $\mathbf{x}^{\mathrm{img}} \in \mathbb{R}^{N \times D}$ , where $N$ is the number of patches, and $D$ is the token dimension. + +# 3.2. Gaze tokens + +The main purpose of a gaze token is to map the gaze information of a person into a token embedded in the same space as the image tokens, which can interact with scene tokens to select the relevant content for prediction. For simplicity, we first introduce this process for a single person. + +Single Person Case. Let $\mathbf{h}_{\mathrm{crop}} \in \mathbb{R}^{h \times w \times C}$ denote the head crop of a person and $\mathbf{h}_{\mathrm{bbox}} = (x_{\mathrm{min}}, y_{\mathrm{min}}, x_{\mathrm{max}}, y_{\mathrm{max}}) \in [0,1]^4$ her head bounding box. The mapping works as follows. The head crop $\mathbf{h}_{\mathrm{crop}}$ is fed to a gaze backbone $\mathcal{G}$ to produce a gaze embedding $\mathbf{g}^{\mathrm{emb}} \in \mathbb{R}^{d_{\mathrm{emb}}}$ . This embedding + +is used in two ways. First, it goes through a gaze regressor (i.e. MLP) $\mathcal{O}_{\mathrm{gv}}$ to predict a 2D gaze vector $\mathbf{g}_{\mathrm{v}} = \mathcal{O}_{\mathrm{gv}}(\mathbf{g}^{\mathrm{emb}})$ . This output is supervised using an angular gaze loss. + +Secondly, the gaze embedding is projected to the token dimension using a learnable linear projection $\mathcal{P}_{\mathrm{gaze}}$ , resulting in the gaze token $\mathbf{x}^{\mathrm{emb}} = \mathcal{P}_{\mathrm{gaze}}(\mathbf{g}^{\mathrm{emb}}) \in \mathbb{R}^{D}$ . As we want to incorporate information about the person's location (and size), we also project the head bounding box $\mathbf{h}_{\mathrm{bbox}}$ into a bounding box embedding $\mathbf{x}^{\mathrm{bbox}}$ using a learnable linear projection $\mathcal{P}_{\mathrm{bbox}}$ : $\mathbf{x}^{\mathrm{bbox}} = \mathcal{P}_{\mathrm{bbox}}(\mathbf{h}_{\mathrm{bbox}}) \in \mathbb{R}^{D}$ . Finally, we add this embedding to the gaze token to obtain the final location-aware gaze token: + +$$ +\mathbf {x} ^ {\mathrm {g}} = \mathbf {x} ^ {\text {e m b}} + \mathbf {x} ^ {\text {b b o x}} \in \mathbb {R} ^ {D} \tag {1} +$$ + +Multi-person case. When $N_{p}$ persons are detected, the architecture will produce a set of $N_{p}$ gaze tokens, following the same process described above for each person. Thus, if $\mathbf{h}_{\mathrm{bbox}}^i$ and $\mathbf{h}_{\mathrm{crop}}^i$ denote the bounding-box and head crop of person $i$ , the above process will generate a gaze token $\mathbf{x}_i^{\mathrm{g}}$ for this person. To simplify notation, we will also denote by $\mathbf{x}^{\mathrm{g}}$ the set of gaze tokens of all people in the scene, with $\mathbf{x}^{\mathrm{g}} = \mathbf{x}_1^{\mathrm{g}} \oplus \ldots \oplus \mathbf{x}_{N_p}^{\mathrm{g}}$ , where $\oplus$ is the concatenation operator. + +Modality Encoding. Given the different nature of gaze tokens compared to image tokens, we need to encode modality-specific information to distinguish between them. Rather than using an explicit scheme, in practice we expect this modality information to be captured by the bias terms of the different projection operators $\mathcal{P}_{\mathrm{gaze}}$ and $\mathcal{P}_{\mathrm{img}}$ . + +# 3.3. Transformer Encoder + +The transformer encoder is a standard ViT [8]. It takes as input the concatenation of the scene tokens $\mathbf{x}^{\mathrm{img}}$ , the gaze token(s) $\mathbf{x}^{\mathrm{g}}$ , according to $\mathbf{x} = \mathbf{x}^{\mathrm{img}} \oplus \mathbf{x}^{\mathrm{g}} \in \mathbb{R}^{N_t \times D}$ , where $N_t = N + N_p$ . The set of input tokens goes through a series of $L$ transformer blocks to obtain an output sequence of similar shape, denoted by $\mathbf{x}^{\mathrm{out}} = \mathbf{x}^{(L)} \in \mathbb{R}^{N_t \times D}$ . + +# 3.4. Gaze Decoder + +The goal of the gaze decoder $\mathcal{D}_{\mathrm{gaze}}$ is to predict a set of gaze heatmaps. Our Conditional DPT (cf. Figure 4) takes four intermediate representations of the image tokens $\mathbf{x}_{(i)}^{\mathrm{img}}$ and gaze tokens $\mathbf{x}_{(i)}^{\mathrm{g}}$ and combines them progressively at different simulated resolutions, where lower resolutions have more channels and correspond to deeper layers of the encoder. This can be viewed as the isotropic equivalent of a Feature Pyramid Network [22]. + +Our design is inspired by DPT [28], which can only handle decoding image tokens alone. In our case, we need this decoding to be conditioned on each person. To this end, after each block of layers, the image tokens $\mathbf{x}_{(l_k)}^{\mathrm{img}}$ , $k \in \{4, 8, 16, 32\}$ are reassembled into an image-like representation at resolution $\left(\frac{H}{k}, \frac{W}{k}\right)$ and dimension $d_k$ . The gaze to + +![](images/b6217bc5303033b18b549d5ac042ec60e5c2335598f84db871b70d0e02251c30.jpg) +Figure 3. Overview of our Sharingan architecture. A. The input image is projected into image tokens (red squares). B. The head crops and head box coordinates are processed to generate location and size-aware person-specific tokens (blue squares) as follows. First, the head crop is fed to a gaze backbone to produce a gaze embedding used to (i) predict a normalized 2D gaze vector that is supervised using an angular loss; and (ii) produce a gaze token by projecting it to the token dimension. Second, head bounding box coordinates are projected to obtain an embedding which, added to the gaze token, produces a person gaze token. C. Image tokens and gaze tokens are fed to the transformer encoder, and the output tokens corresponding to input people are all decoded using a Conditional DPT decoder to predict each person's gaze heatmaps. In addition, input and output gaze tokens are concatenated together to predict the in- vs-out label. + +kens $\mathbf{x}_{(l_k)}^{\mathrm{g}}$ at that layer are also projected to the same dimension. Next, we duplicate the image feature maps $N_{p}$ times, and apply an element-wise dot-product between each gaze token and a copy of the image feature map. Finally, these person-specific image features are stacked, and we merge the batch and person dimensions to produce a final output of dimension $(B\times N_p,d_k,\frac{H}{k},\frac{W}{k})$ . This tensor is passed to a fusion module where it is processed by a small residual convnet and added to the output from the previous fusion block. The result goes through another residual convnet, an upsampling stage to double the resolution, and a projection, leading to a tensor of dimension $(B\times N_p,d_k,\frac{2H}{k},\frac{2W}{k})$ . At the end of this process, we get a tensor of dimension $(B\times N_p,d_{\mathrm{out}},\frac{H}{2},\frac{W}{2})$ , which goes through a convolutional head that predicts the heatmaps by bringing the channel dimension down to 1 and resizing the spatial dimension to that of the gaze heatmap. Finally, we separate the batch and person dimensions such that the final shape of the output is $(B,N_{p},1,H_{\mathrm{hm}},W_{\mathrm{hm}})$ . The rationale behind this design is to gather information from different layers and resolutions, which is important for dense prediction tasks. In this case, it is particularly useful for gaze tokens where information from the early layers might retain more scene-independent gaze cues due to their proximity to the gaze encoder. + +# 3.5. In-Out prediction + +The In-Out classifier head $\mathcal{O}_{\mathrm{MLP}}$ consists of an MLP with 7 layers. It is fed the concatenation of input and output gaze tokens to predict a binary in-vs-out label for each person. + +$$ +\mathbf {o} = \mathcal {O} _ {\mathrm {M L P}} \left(\left[ \mathbf {x} _ {(L)} ^ {\mathrm {g}}, \mathbf {x} ^ {\mathrm {g}} \right]\right) \tag {2} +$$ + +# 3.6. Loss and implementation details + +We train our model using a combination of three losses: + +Heatmap Loss $(\mathcal{L}_{reg})$ . The heatmap loss is the pixel-wise MSE between the GT heatmap and the predicted heatmap: $\mathcal{L}_{hm} = \sum_{x,y}^{W_{hm},H_{hm}}||\mathcal{A}_{x,y}^{\mathrm{gt}} - \mathcal{A}_{x,y}^{\mathrm{pred}}||_2^2$ + +Angular Loss $(\mathcal{L}_{ang})$ . The angular loss drives the prediction of a normalized gaze direction vector. It maximizes the cosine of the angle between the predicted and ground truth gaze vectors according to: $\mathcal{L}_{ang} = 1 - < \mathbf{g}_{\mathrm{v}}^{gt}, \mathbf{g}_{\mathrm{v}}^{pred}>$ where $< a, b >$ denotes the inner product between $a$ and $b$ . + +In-Out Loss $(\mathcal{L}_{io})$ . The in-out loss is the standard binary cross-entropy for in-vs-out prediction. + +Global loss. The final loss is a given by: + +$$ +\mathcal {L} = \lambda_ {r e g} \mathcal {L} _ {r e g} + \lambda_ {a n g} \mathcal {L} _ {a n g} + \lambda_ {i o} \mathcal {L} _ {i o} \tag {3} +$$ + +![](images/13c062be1ece1f26981f9b509f2b2a935b8091ef8b8e33c2eb83cf3ec8147348.jpg) +Figure 4. Overview of our proposed Conditional DPT decoder. + +# 4. Experiments + +# 4.1. Datasets + +GazeFollow. GazeFollow [30] is an image-based dataset annotated mostly with head bounding boxes, 2D gaze points. Overall, it has around $130\mathrm{K}$ annotated instances in $122\mathrm{K}$ images. The test set comprises 4782 gaze instances, each of which is labeled by multiple annotators $(\sim 10)$ . + +VideoAttentionTarget. VideoAttentionTarget [7] is a video-based dataset consisting of 1331 clips from 50 TV shows. It is also annotated with the head bounding boxes, 2D gaze points, and in vs out labels. Overall, it contains 164K instances in 71K frames. + +ChildPlay. ChildPlay [35] is a video dataset consisting of 401 clips from 95 YouTube videos of children engaged in play activities. On top of the standard annotations, the authors extend the in vs out label to include other gaze classes (e.g. gaze shift), which we don't use in this paper. Overall, it contains 257K instances in 120K frames. + +# 4.2. Metrics + +We use four metrics to evaluate our gaze following models. The first three are AUC, Distance, and AP which are all standard in the literature [7]. Particularly, the AUC adaptation introduced by [7] for datasets with a single point annotation is not very informative, so we decided to replace it with PLAH, which we define below. + +PLAH. Recently introduced in [35], this metric computes the Precision of looking at people's heads to incorporate semantic information in the evaluation of gaze models. Unlike [35], we consider a prediction positive if the predicted and annotated gaze points fall within the same head box. + +# 4.3. Experimental Protocol + +Context People. The training of Sharingan relies on processing multiple people at the same time (cf. section 5), but available benchmarks often annotate 1 person per image. To circumvent this problem, we apply an off-the-shelf head + +detector $^{1}$ trained on the CrowdHuman dataset [31]. We discard detections with a confidence score lower than 0.5 and detections with an IOU score higher than 0.5 with the annotated target person. During training, due to batch constraints, we set $N_{p}^{tr}$ and keep it fixed. For each image, we use the person with the GT annotation and randomly sample $N_{p}^{tr} - 1$ (detected) heads when available, otherwise we use padding for the box and head crop. Incidentally, the loss is computed and propagated solely from the annotated person. At evaluation, for each image $i$ we set $N_{p}^{i}$ to the number of all people in it and process them in a single forward pass using a batch size of 1. Note that $N_{p}^{tr}$ is a property of the training process, not the architecture, and thus doesn't restrict the $N_{p}^{i}$ that can be used during inference. Unless stated otherwise, we use $N_{p}^{tr} = 2$ in all experiments. + +Implementation Details. Sharingan processes the input scene image and head crop at a resolution of $224 \times 224$ , while the output heatmap is $64 \times 64$ . The gaze backbone $\mathcal{G}$ is a ResNet-18 [13] pretrained on Gaze360 [18], and the transformer encoder is a ViT-base model [9] initialized with weights from a multimodal MAE [2]. + +Training. The models are trained for 20 epochs on GazeFollow. For VideoAttentionTarget and ChildPlay, following standard practices, we take the trained GazeFollow model, freeze everything except the gaze decoder and In-Out classifier, and fine-tune them separately for 2 epochs each. We use the AdamW optimizer [23] with a learning rate of $3e - 5$ and a cosine annealing schedule. The fine-tuning uses a learning rate of $1e - 6$ for the gaze decoder and $3e - 4$ for the In-Out classifier. We also make use of Stochastic Weight Averaging [15] to stabilize training on GazeFollow. The loss coefficients are $\lambda_{reg} = 1000$ and $\lambda_{ang} = 3$ . + +Validation. Since GazeFollow [30] and VideoAttentionTarget [7] do not propose any validation splits, we use the train-val splits proposed by [35]. The best model on the validation set is selected based on the distance metric. + +# 4.4. Comparison with the State-of-the-art + +We summarize our quantitative results on the GazeFollow and VideoAttentionTarget datasets in Table 1, and on ChildPlay in Table 2, compared to previous works. Our model sets a new state-of-the-art on all 3 datasets on most metrics and outperforms the only comparable multi-person method by 0.013 on the Avg. Dist. metric. Moreover, the only method that comes close to our results is [12], which we slightly outperform on GazeFollow in both Avg. Dist. and AUC. However, unlike Sharingan, this method uses 2 other modalities (i.e. depth, pose), has a very complex and costly training protocol (i.e. modality-specific backbones are pretrained separately for the task), and is resource-intensive during inference (i.e. single-person). Moreover, it generalizes poorly as evidenced by a cross-dataset evaluation (i.e. Dist. 0.113 vs 0.134 on VideoAttentionTarget, and 0.109 vs 0.142 on ChildPlay). Incidentally, the image version of this method is also significantly worse (i.e. 0.134 vs 0.113). In terms of multi-person comparison, we beat [16] by 0.013 and 0.02 in Avg. Dist. and Min. Dist. respectively. + +We also see similar results on video datasets (VideoAttentionTarget and ChildPlay), and it is interesting to note that the model trained on GazeFollow is achieving remarkable cross-dataset performance without fine-tuning. Surprisingly, it is already improving on its multi-person competitor by a large margin (i.e. Dist. 0.113 vs 0.134), which is a testament to the generalization ability of our model. We also note that the model from [35] is on par with ours on the video datasets. We believe the reason is these datasets have high-quality images compared to GazeFollow, so the depth-based geometric prior in [35] proves useful, but that method remains multimodal and single-person. We provide more details, discussions, and experiments in the supplementary material to further assess robustness and generalization. + +# 4.5. Ablation Experiments + +Person Encoding. An important aspect of Sharingan is the way people are encoded into the architecture. Previous methods often represent the head location as a binary mask, and gaze information as a visual attention map [7] or gaze cone [11, 12, 35]. We modify Sharingan to experiment with 3 variants, and opt for single-person training to alleviate the computational cost from some of these formulations: 1. Instead of projecting the box coordinates, we use a head location mask that we tokenize and add the resulting head position embeddings to the image tokens (i.e. Head Mask Embed). 2. Using the same head mask embedding, we tokenize the head crop and append the head tokens to the image ones (i.e. Head Crop Tokens). 3. Instead of processing the head directly, we use the gaze backbone to regress a gaze + +vector which is used to build a gaze cone that we tokenize (i.e. Gaze Cone Tokens). Please note that formulations 2 and 3 double the number of tokens, and we decode the output image tokens into a gaze heatmap using a normal DPT [28]. As we can see from Table 3 (top), the location-aware gaze token is both efficient and performs the best. + +
MethodAUC ↑Avg. D. ↓Min. D. ↓
Head Mask Embed0.9400.1170.060
Head Crop Tokens0.9330.1380.076
Gaze Cone Tokens0.9340.1330.073
Gaze Token0.9440.1130.057
Token to Heatmap0.6470.3020.234
Dot-Product0.9230.1200.062
Up & Dot-Product0.9340.1160.059
Conditional DPT0.9440.1130.057
+ +Table 3. Ablation results for person encoding (top) and gaze decoding (bottom). + +Gaze Decoder. We also perform ablations to assess the suitability of our Conditional DPT gaze decoder by comparing it to other baselines as shown in Table 3 (bottom). The first baseline (i.e. Token to Heatmap) regresses a gaze heatmap directly from the output person token using an MLP. This is also the decoding approach undertaken by [39] and [38]. The second baseline (cf. Dot-Product) projects the tokens, then performs a dot-product between each person token and each image token before resizing the output to $64 \times 64$ to get the final heatmap. The third baseline (cf. Up & Dot-Product) upscales the image representation first, then performs the dot-product. We note that our Conditional DPT outperforms all the other methods, justifying the need for a more sophisticated decoding mechanism. + +Angular Loss. Our experiments show that the angular loss doesn't affect the final performance, but we decided to keep it anyway to always have a reliable gaze direction, even when the person is looking outside the frame. + +# 5. Discussion + +Model Efficiency. Unlike most previous methods, Sharingan's ability to predict the gaze of multiple people at the same time makes it very efficient for real-world applications. Figure 5 (left) shows a fairly large improvement in inference time compared to a lightweight baseline when $N_{p}$ increases. It's worth noting that while Sharingan does most of its processing once within the transformer, the small gaze encoder and decoder themselves are executed for each person independently by combining the batch and person dimensions. This explains why the curve is not constant. + +
MethodMultiModalityGazeFollowVideoAttentionTarget
AUC ↑Avg. Dist. ↓Min. Dist. ↓PLAH ↑Dist. ↓PLAH ↑AP ↑
Recasens [29]XI0.8780.1900.113————
Chong [6]XI+T0.8960.1870.112—0.171—0.712
Lian [21]XI0.9060.1450.081————
Chong [7]XI+T0.9210.1370.077—0.134—0.853
Fang [11]XI+D+E0.9220.1240.067—0.108—0.896
Fang [11]XI+D————0.124—0.872
Jin [17]XI+D0.9200.1180.063—0.109—0.897
Jin [17]XI0.9090.1370.077————
Tonini [37]XI+D0.9270.141——0.129——
Gupta [12]XI+D+P0.9430.1140.056—0.110—0.879
Gupta [12]XI0.9330.1340.071—0.122—0.864
Bao [3]XI+D+P0.9280.122——0.120—0.869
Hu [14]XI+D+O0.9230.1280.069—0.118—0.881
Tafasca [35]XI+D0.9360.1250.0640.6220.1090.7520.834
Jin [16]✓I0.9190.1260.076—0.134—0.880
\(Sharingan^†\)✓I0.9440.1130.0570.6670.1130.748—
Sharingan✓I0.9440.1130.0570.6670.1070.7380.891
+ +Table 1. Results of our Sharingan architecture on the GazeFollow and VideoAttentionTarget datasets. The best scores for multi-person models are given in bold, while the best scores in general are underlined. The $\dagger$ symbol means that the model was trained on GazeFollow and evaluated without fine-tuning. The modality column uses the codes I (image), T (time), D (depth), E (eyes), P (pose), and O (objects). + +
MethodMultiDist. ↓PLAH ↑AP ↑
Gupta [12]X0.113—0.983
Tafasca [35]X0.1070.5900.986
Sharingan†✓0.1090.600—
Sharingan✓0.1060.6000.990
+ +Table 2. Results on the ChildPlay dataset. + +![](images/03f804cb84deb5704863d28c658bb3b0b49042ceef98977b0d68f315e3b896f8.jpg) +Figure 5. (Left): Comparison of CPU inference time based on $N_{p}$ . (Right): Comparison of single- and multi-person models by $N_{p}^{i}$ during evaluation on GazeFollow. + +![](images/dd306b9510cf2f4edf3d285b7d31737290e5080fe2e0522590070c51194ba350.jpg) + +Number of People. One consideration of particular importance in the context of Sharingan is the influence of the number of people on training and evaluation. Table 4 shows that increasing $N_{p}^{tr}$ during training does not influence performance. We believe that since the gaze tokens are interchangeable and play a symmetric role (i.e. no order encoding), as long as $N_{p}^{tr} > 1$ , the model will be forced to learn how to accommodate an arbitrary number of people in the scene during inference by sharing the learned image representations among them. However, the same cannot be said for the single-person setting ( $N_{p}^{tr} = 1$ ), which is fundamentally different from multi-person ( $N_{p}^{tr} > 1$ ). In this case, the model inevitably learns to tailor its image representations to a single person. To illustrate this behavior, we evaluate a single-person ( $N_{p}^{tr} = 1$ ) and a multi-person ( $N_{p}^{tr} = 2$ ) models twice: first using one person ( $N_{p}^{i} = 1$ ) and a second time by processing all people in the scene ( $N_{p}^{i} = all$ ). Figure 5 (right) shows that the multi-person model is perfectly able to do both single-person and multi-person prediction. However, the single-person model experiences a significant degradation when attempting multi-person prediction. + +Heatmap vs 2D Point. Gaze following has always been framed as a heatmap prediction task. To the best of our + +
AUCAvg. D. ↓Min. D. ↓
Ntrp=20.9440.1130.057
Ntrp=30.9430.1140.058
Ntrp=40.9420.1130.057
+ +knowledge, Lian et al. [21] were the only authors to experiment with a 2D point regression objective in their ablation study. In this section, we explore this formulation further to gain a deeper understanding of the task. To this end, we trained a Sharingan model by replacing the Conditional DPT decoder with a simple MLP to directly regress $(x,y)$ coordinates from the output person token. Since the 2D Point model can only predict a single value, it can't represent the entire distribution over the 2D space of the image. Instead, we believe that it converges to an expectation of this posterior probability. When this distribution is multimodal (i.e. more than one probable gaze target), the expectation becomes unlikely under that posterior distribution. We illustrate this behavior in Figure 6 by comparing the person-specific attention map from the last layer of the encoder of the 2D Point model to the predicted gaze heatmap from the Heatmap variant. It is clear that both models capture the different modes quite well, but the former outputs $(x,y)$ coordinates resembling a weighted average of these modes, which ends up distant from all of them. + +This phenomenon leads to the results given in Table 5, showing that the $2D$ point model achieves better average distance and PLAH, but lags behind the Heatmap model in terms of minimum distance and RLAH. This is not surprising since the average distance of the GazeFollow test set is a distance to an average of modes (i.e. multiple annotated points), which loosely resembles the objective that the $2D$ Point model is optimizing for. Consequently, we believe that the average distance should never be the only metric for evaluating gaze following methods, a claim also supported by [3]. On the other hand, positive LAH predictions for the $2D$ Point model mean that it is very confident, which explains the better PLAH value. However, this also means that the number of false negatives will be high, hence the big gap in RLAH because the predicted point will often be slightly off-target when the model is not confident (e.g. marginally outside the head area when looking at a person). + +Limitation. Sharingan processes both image and gaze tokens simultaneously in the transformer. This means that the same weights operate on both types of tokens, which makes it difficult to understand how the model is combining information. One idea worth exploring is to disentangle scene and person processing, and selectively fuse their informa + +![](images/0ad760e11f5772874226da3ab0c96ff792e051dad9995a9b70db34c5b4a2421d.jpg) + +![](images/e8c3fb0a92bbcefdad3466c262b701e4b8429c26a2c514cc86177e4f8f064ccb.jpg) + +![](images/16ff9ec24986c8120aaa8d2fb77cac8da92fd5d71aae494bfee4a8b42df59506.jpg) +Figure 6. Comparison of Heatmap (i.e. left) and 2D Point regression model (i.e. right) models. The heatmap on the right is obtained by computing the attention weights (i.e. last encoder layer) of the person's gaze token with the image tokens. + +![](images/3c4edd67233700f4bf730cd3e1cdf2bc535da07ec8902fe40c09a37783689696.jpg) + +Table 4. Influence of the number of people during training on GazeFollow. The evaluation uses all available people. + +
Avg. D.Min. D.PLAHRLAH
2D Point0.1060.0660.6830.368
Heatmap0.1130.0570.6670.571
+ +Table 5. Comparison between the heatmap and the 2D point training objectives on GazeFollow. RLAH is the recall of LAH. + +tion along the architecture. This may help improve the stability of predictions on videos, and allow people to interact together in a more meaningful way. + +# 6. Conclusion + +In this paper, we introduced Sharingan, a novel and efficient transformer-based architecture for gaze target prediction that is designed to support an arbitrary number of people out of the box. Sharingan stands out for its efficiency both in training and inference, delivering SOTA results on public benchmarks. Furthermore, it demonstrates strong generalization when tested on other datasets and naturalistic scenes. We also validated architectural decisions through extensive ablation experiments and discussed key aspects related to the model and the task. + +Beyond gaze following, Sharingan's intuitive design makes it suitable for other research areas in human behavior understanding. Specifically, the architecture can be extended to perform a multi-faceted analysis of social scenes by integrating different modalities (e.g. depth, semantics), and producing more outputs (e.g. gestures, interactions). We intend to explore this direction further in future work. + +Acknowledgement. This research has been supported by the AI4Autism project (Digital Phenotyping of Autism Spectrum Disorders in children, grant agreement no. CR-SII5 202235 / 1) of the Sinergia interdisciplinary program of the SNSF. + +# References + +[1] Henny Admoni and Brian Scassellati. Social eye gaze in human-robot interaction: a review. Journal of Human-Robot Interaction, 6(1):25-63, 2017. 2 +[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 5 +[3] Jun Bao, Buyu Liu, and Jun Yu. Escnet: Gaze target detection with the understanding of 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14126-14135, 2022. 7, 8 +[4] Bridget K Behe, Patricia T Huddleston, Kevin L Childs, Jiaoping Chen, and Iago S Muraro. Seeing through the forest: The gaze path to purchase. Plos one, 15(10):e0240179, 2020. 1 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3 +[6] Eunji Chong, Nataniel Ruiz, Yongxin Wang, Yun Zhang, Agata Rozga, and James M Rehg. Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In Proceedings of the European conference on computer vision (ECCV), pages 383-398, 2018. 7 +[7] Eunji Chong, Yongxin Wang, Nataniel Ruiz, and James M Rehg. Detecting attended visual targets in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5406, 2020. 2, 5, 6, 7 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5 +[10] Lifeng Fan, Yixin Chen, Ping Wei, Wenguan Wang, and Song-Chun Zhu. Inferring shared attention in social scene videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6460-6468, 2018. 2, 5 +[11] Yi Fang, Jiapeng Tang, Wang Shen, Wei Shen, Xiao Gu, Li Song, and Guangtao Zhai. Dual attention guided gaze target detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11390-11399, 2021. 2, 6, 7 +[12] Anshul Gupta, Samy Tafasca, and Jean-Marc Odobez. A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5041-5050, 2022. 2, 6, 7 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[14] Zhengxi Hu, Kunxu Zhao, Bohan Zhou, Hang Guo, Shichao Wu, Yuxue Yang, and Jingtai Liu. Gaze target estimation inspired by interactive attention. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8524-8536, 2022. 7 +[15] P Izmailov, AG Wilson, D Podoprikhin, D Vetrov, and T Garipov. Averaging weights leads to wider optima and better generalization. In 34th Conference on Uncertainty in Artificial Intelligence 2018, UAI 2018, pages 876-885, 2018. 5 +[16] Tianlei Jin, Zheyuan Lin, Shiqiang Zhu, Wen Wang, and Shunda Hu. Multi-person gaze-following with numerical coordinate regression. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 01-08. IEEE, 2021. 2, 3, 6, 7 +[17] Tianlei Jin, Qizhi Yu, Shiqiang Zhu, Zheyuan Lin, Jie Ren, Yuanhai Zhou, and Wei Song. Depth-aware gaze-following via auxiliary networks for robotics. Engineering Applications of Artificial Intelligence, 113:104924, 2022. 2, 7 +[18] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2, 5 +[19] Nayeon Kim and Hyunsoo Lee. Assessing consumer attention and arousal using eye-tracking technology in virtual retail environment. Frontiers in Psychology, 12:665658, 2021. 1 +[20] Jing Li, Zejin Chen, Yihao Zhong, Hak-Keung Lam, Junxia Han, Gaoxiang Ouyang, Xiaoli Li, and Honghai Liu. Appearance-based gaze estimation for asd diagnosis. IEEE Transactions on Cybernetics, 52(7):6504-6517, 2022. 2 +[21] Dongze Lian, Zehao Yu, and Shenghua Gao. Believe it or not, we know what you are looking at! In Asian Conference on Computer Vision, pages 35-50. Springer, 2018. 2, 7, 8, 5 +[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3 +[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5 +[24] Manuel Jesús Marin-Jimenez, Andrew Zisserman, Marcin Eichner, and Vittorio Ferrari. Detecting people looking at each other in videos. International Journal of Computer Vision, 106(3):282-296, 2014. 2 +[25] Manuel J Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net: revisiting people looking at each other in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3477-3485, 2019. 5 + +[26] Skanda Muralidhar, Rémy Siegfried, Jean-Marc Odobez, and Daniel Gatica-Perez. Facing employers and customers: What do gaze and expressions tell about soft skills? In Proceedings of the 17th International Conference on Mobile and Ubiquitous Multimedia, MUM 2018, Cairo, Egypt, November 25-28, 2018, pages 121-126, 2018. +[27] Catharine Oertel, Patrik Jonell, Dimosthenis Kontogiorgos, Kenneth Funes Mora, Jean-Marc Odobez, and Joakim Gustafson. Towards an engagement-aware attentive artificial listener for multi-party interactions. Frontiers in Robotics and AI, 8:189, 2021. 2 +[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 3, 6 +[29] Adria Recasens*, Aditya Khosla*, Carl Vondrick, and Antonio Torralba. Where are they looking? In Advances in Neural Information Processing Systems (NIPS), 2015. * indicates equal contribution. 2, 7 +[30] Adria Recasens, Carl Vondrick, Aditya Khosla, and Antonio Torralba. Following gaze in video. In Proceedings of the IEEE International Conference on Computer Vision, pages 1435-1443, 2017. 5 +[31] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 5 +[32] Samira Sheikhi and Jean-Marc Odobez. Combining dynamic head pose-gaze mapping with the robot conversational state for attention recognition in human-robot interactions. Pattern Recognition Letters, 66:81–90, 2015. 2 +[33] R. Siegfried and J.-M. Odobez. Robust unsupervised gaze calibration using conversation and manipulation attention priors. ACM Transactions on Multimedia Computing, Communications, and Applications, 18(1):1-27, 2022. 2 +[34] S. Tafasca, A. Gupta, N. Kojovic, M. Gelsomini, T. Maillart, M. Papandrea, M. Schaer, and J.-M. Odobez. The ai4autism project: A multimodal and interdisciplinary approach to autism diagnosis and stratification. In Proceeding of the Int. Conference on Multimodal Interaction workshop, Paris, 2023. 2 +[35] Samy Tafasca, Anshul Gupta, and Jean-Marc Odobez. Childplay: A new benchmark for understanding children's gaze behaviour. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20935-20946, 2023. 2, 5, 6, 7, 1 +[36] Henri Tomas, Marcus Reyes, Raimarc Dionido, Mark Ty, Jonric Mirando, Joel Casimiro, Rowel Atienza, and Richard Guinto. Goo: A dataset for gaze object prediction in retail environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3125-3133, 2021. 1 +[37] Francesco Tonini, Cigdem Beyer, and Elisa Ricci. Multimodal across domains gaze target detection. In Proceedings of the 2022 International Conference on Multimodal Interaction, pages 420-431, 2022. 7 +[38] Francesco Tonini, Nicola Dall'Asen, Cigdem Beyer, and Elisa Ricci. Object-aware gaze target detection. In Proceed- + +ings of the IEEE/CVF International Conference on Computer Vision, pages 21860-21869, 2023. 1, 3, 6, 5 +[39] Danyang Tu, Xiongkuo Min, Huiyu Duan, Guodong Guo, Guangtao Zhai, and Wei Shen. End-to-end human-gazetarget detection with transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2192-2200. IEEE, 2022. 1, 3, 6, 5 +[40] Yu Yu, Gang Liu, and Jean-Marc Odobez. Deep multitask gaze estimation with a constrained landmark-gaze model. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2 \ No newline at end of file diff --git a/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/images.zip b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e2bad0f34b039708f8c53bf36cfeeea474acf61 --- /dev/null +++ b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44cc8ce95a137b419bf11ee2cbc7286aa54345e1ec0a5241c4c7ab3f712a9656 +size 594518 diff --git a/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/layout.json b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f84944e82eb9a3f5d5b09279a0dc610b48cea455 --- /dev/null +++ b/2024/Sharingan_ A Transformer Architecture for Multi-Person Gaze Following/layout.json @@ -0,0 +1,7963 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "text", + "content": "Sharingan: A Transformer Architecture for Multi-Person Gaze Following" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 143, + 449, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 143, + 449, + 200 + ], + "spans": [ + { + "bbox": [ + 143, + 143, + 449, + 200 + ], + "type": "text", + "content": "Samy Tafasca Anshul Gupta Jean-Marc Odobez Idiap Research Institute, Switzerland École Polytechnique Fédérale de Lausanne, Switzerland {stafasca, agupta, odobez}@idiap.ch" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 228, + 194, + 324 + ], + "blocks": [ + { + "bbox": [ + 48, + 228, + 194, + 324 + ], + "lines": [ + { + "bbox": [ + 48, + 228, + 194, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 194, + 324 + ], + "type": "image", + "image_path": "1f31d57e4dd09686e237ee1ef97f6c1f37f60b39a2b1a85d8bb06b848286d2c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 228, + 400, + 324 + ], + "blocks": [ + { + "bbox": [ + 194, + 228, + 400, + 324 + ], + "lines": [ + { + "bbox": [ + 194, + 228, + 400, + 324 + ], + "spans": [ + { + "bbox": [ + 194, + 228, + 400, + 324 + ], + "type": "image", + "image_path": "989b488159efff31f59072335aa643d96bdc887478fff84ccb4a7c45e417a723.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 401, + 228, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 401, + 228, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 401, + 228, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 401, + 228, + 545, + 324 + ], + "type": "image", + "image_path": "dd20632f1bc399cfd863e8a463ed244cf2a229ad3a2b88e1715ba2a71a44faf8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 326, + 220, + 423 + ], + "blocks": [ + { + "bbox": [ + 48, + 326, + 220, + 423 + ], + "lines": [ + { + "bbox": [ + 48, + 326, + 220, + 423 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 220, + 423 + ], + "type": "image", + "image_path": "e2c396e6a6b18277e75d65260aca90dc22199cc609ac48bde274451a140d7c20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 432, + 546, + 455 + ], + "lines": [ + { + "bbox": [ + 45, + 432, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 45, + 432, + 546, + 455 + ], + "type": "text", + "content": "Figure 1. Predictions of Sharingan on naturalistic images from the internet with different people, activities, interactions, postures, and environments (indoors and outdoors). We provide more qualitative samples in the supplementary material." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 221, + 326, + 372, + 423 + ], + "blocks": [ + { + "bbox": [ + 221, + 326, + 372, + 423 + ], + "lines": [ + { + "bbox": [ + 221, + 326, + 372, + 423 + ], + "spans": [ + { + "bbox": [ + 221, + 326, + 372, + 423 + ], + "type": "image", + "image_path": "2c1241e016785f2cc88308d28fd101a2defd70ca443e4625dcf319653ab1df70.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 373, + 326, + 545, + 423 + ], + "blocks": [ + { + "bbox": [ + 373, + 326, + 545, + 423 + ], + "lines": [ + { + "bbox": [ + 373, + 326, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 373, + 326, + 545, + 423 + ], + "type": "image", + "image_path": "bbc3aed12cb3eda5443dc888f733066640a2dac7b542c62dcb64e0182e2bf9f8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 464, + 191, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 464, + 191, + 476 + ], + "spans": [ + { + "bbox": [ + 143, + 464, + 191, + 476 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "type": "text", + "content": "Gaze is a powerful form of non-verbal communication that humans develop from an early age. As such, modeling this behavior is an important task that can benefit a broad set of application domains ranging from robotics to sociology. In particular, the gaze following task in computer vision is defined as the prediction of the 2D pixel coordinates where a person in the image is looking. Previous attempts in this area have primarily centered on CNN-based architectures, but they have been constrained by the need to process one person at a time, which proves to be highly inefficient. In this paper, we introduce a novel and effective multi-person transformer-based architecture for gaze prediction. While there exist prior works using transformers for multi-person gaze prediction [38, 39], they use a fixed set of learnable embeddings to decode both the person and its gaze target, which requires a matching step afterward to link the predictions with the annotations. Thus, it is difficult to quantitatively evaluate these methods reliably with" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 466, + 547, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 547, + 634 + ], + "type": "text", + "content": "the available benchmarks, or integrate them into a larger human behavior understanding system. Instead, we are the first to propose a multi-person transformer-based architecture that maintains the original task formulation and ensures control over the people fed as input. Our main contribution lies in encoding the person-specific information into a single controlled token to be processed alongside image tokens and using its output for prediction based on a novel multiscale decoding mechanism. Our new architecture achieves state-of-the-art results on the GazeFollow, VideoAttentionTarget, and ChildPlay datasets and outperforms comparable multi-person architectures with a notable margin. Our code, checkpoints, and data extractions will be made publicly available soon." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 657, + 387, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 387, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 387, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Gaze is an important form of communication and was extensively studied across different domains and applications such as consumer behavior understanding [4, 19, 36], soci" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2008" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 286, + 166 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 286, + 166 + ], + "type": "image", + "image_path": "6cfe831da9f5f7d6da0f6136e99a53f846c71a1cc33530cda5032e46d98bd8c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 175, + 272, + 186 + ], + "lines": [ + { + "bbox": [ + 60, + 175, + 272, + 186 + ], + "spans": [ + { + "bbox": [ + 60, + 175, + 272, + 186 + ], + "type": "text", + "content": "Figure 2. Illustration of aspects relevant to gaze following." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 208, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 286, + 255 + ], + "type": "text", + "content": "ology by analyzing different gaze behaviors (e.g. joint attention, eye contact) [10, 24-27], robotics through humanrobot interactions [1, 17, 32] and clinical research for the study of neurodevelopmental disorders [7, 20, 34], etc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 286, + 376 + ], + "type": "text", + "content": "Unlike traditional works on gaze analytics proposed by the computer vision community which focused mainly on predicting 3D gaze directions from the eyes [33, 40] or the face [18] of a person, gaze following [29] tackles the task in a more general form where the goal is to infer the 2D location in the image where a person is looking without the need for any assumptions or wearable devices. This formulation is particularly interesting in the context of analyzing social scenes and human interactions given the important role that gaze behavior plays in social dynamics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 286, + 496 + ], + "type": "text", + "content": "It is important to emphasize that this task is very difficult to solve. It hinges upon understanding two aspects simultaneously: (i) the target person (e.g. head pose) to infer a general gaze direction, and (ii) the context of the scene (e.g. social interaction) to identify regions of saliency (cf. Figure 2). Once this is achieved, the rest is essentially a selection process of the gaze target after merging information from the previous steps. This also explains why most architectures have a 2-tower design with one branch to process the scene and a second one to focus on the target person." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 675 + ], + "type": "text", + "content": "One major issue with this dual-stream approach is that the input instance is a person, not an image. Therefore, it requires multiple forward passes to predict gaze for all people in the same scene, which makes the inference process extremely inefficient. This is made even worse when other modalities are involved (e.g. depth [11, 35] or pose [12]). Moreover, most previous works focus on how to build the person-specific gaze representation, the scene saliency maps, and the fusion between them but pay very little attention to the decoding mechanism that predicts the final gaze heatmap. For example, [7, 11, 17] all use a decoding module based on a few convolutional layers followed by a set of transposed convolutions. The input to this module is often very low dimensional (e.g. " + }, + { + "bbox": [ + 46, + 498, + 286, + 675 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 46, + 498, + 286, + 675 + ], + "type": "text", + "content": ") which forces the prediction to be based on coarse information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "In this work, we aim to tackle these challenges by addressing the multi-person gaze-following task while maintaining the original problem formulation. To this end," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "content": "we propose Sharingan, a novel, effective, and efficient transformer-based architecture to predict the gaze target of multiple people simultaneously. A key component of this architecture is to represent the person's gaze information by a single gaze token produced by a gaze backbone and processed alongside the image tokens. This is in stark contrast to previous methods that represent intermediate person-specific gaze information as a visual attention map [7] or gaze cone [11, 12, 21, 35]. We show in our ablations that this is not only unnecessary but can also hinder performance in the context of a transformer. Furthermore, we introduce Conditional DPT: a more sophisticated lightweight multi-scale gaze decoding mechanism that helps improve performance by providing a finer-grained understanding of the scene for gaze target selection. This also has the benefit of producing heatmaps that better capture uncertainty when it is difficult to decide where a person is looking (cf. qualitative results in the supplementary material)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": "Through extensive ablations and evaluations, we find that Sharingan achieves good performance on all public benchmarks, and even transfers well to other gaze-related tasks such as shared attention and mutual gaze." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 346, + 391, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 391, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 391, + 358 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 366, + 541, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 541, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 541, + 378 + ], + "type": "text", + "content": "In this section, we present several relevant research topics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "content": "Gaze Following. The task of gaze following was first introduced in the seminal work of Recasens et al. [29]. The idea is to predict the pixel-wise 2D location in the image corresponding to where a target person is looking. The main advantage of this formulation is the lack of constraints which allows methods trained this way to generalize to arbitrary settings (i.e. scene properties, camera parameters, image conditions, etc.). It was later extended by Chong et al. [7] to also include the prediction of whether the given person is looking inside the image frame or somewhere outside." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "type": "text", + "content": "Traditional methods for gaze following [7, 11, 12, 16, 17, 21, 29] typically rely on convolutional networks and follow a 2-tower architecture. The first branch processes the scene image to highlight salient regions, while the second branch processes the head crop of the target person to infer a general gaze direction. A fusion mechanism then combines information from both parts to produce the final prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": "The gaze following task is often framed as the prediction of a gaze heatmap where pixels with high intensity represent spatial areas with higher prediction confidence. We devote a section later to discuss the alternative formulation of regressing the 2D location directly (cf. Section 5)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Multi-Person Gaze Following. A major downside of the traditional formulation of gaze following is the need for multiple forward passes when predicting the gaze of different people in the same image. This problem motivated the need for architectures that can natively handle the prediction of gaze for multiple people with a single forward" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2009" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 360 + ], + "type": "text", + "content": "pass. Jin et al. [16] first proposed a simple convolution-based architecture to handle the multi-person setting where a scene backbone computes a fixed person-agnostic feature representation. This is then fused repetitively with head features computed from the different people using another head backbone before decoding each into its corresponding gaze heatmap. Aside from the architectural differences and limited performance, one of the main drawbacks of this method is that the computation for each person is done independently from the others, which ignores potential interactions between people. Recently, Tu et al. [39] and Tonini et al. [38] proposed transformer-based architectures to perform multi-person gaze target prediction. Their methods only take the image as input and simultaneously predict both the head box and gaze target (among others) for every person in the scene. Their work is inspired by the DETR architecture [5], where the task is formulated as a set prediction problem. Instead of reinventing the wheel, our method focuses solely on the gaze prediction part (i.e. given that heads are easily and accurately obtainable using off-the-shelf detectors), and naturally adapts the transformer architecture to the original task formulation by introducing gaze tokens to capture person-specific gaze and head location information and can be directly decoded later into gaze predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 372, + 185, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 185, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 185, + 385 + ], + "type": "text", + "content": "3. Sharingan Architecture" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 392, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 477 + ], + "type": "text", + "content": "Our Sharingan architecture is illustrated in Figure 3. The main idea is to use a transformer that lets scene tokens and person-specific gaze tokens interact within an attention framework to jointly predict the 2D gaze heatmap of each individual. Thus, the inputs are the image and the head crops that we assume are available. We introduce below the different components of this architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 485, + 133, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 133, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 133, + 498 + ], + "type": "text", + "content": "3.1. Image tokens" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "content": "We follow a standard ViT architecture to produce image tokens. The input scene image " + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "content": " goes through a patch projection " + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "content": " to produce image tokens that we equip with positional information " + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{img}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "content": " is the number of patches, and " + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "content": " is the token dimension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "type": "text", + "content": "3.2. Gaze tokens" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "type": "text", + "content": "The main purpose of a gaze token is to map the gaze information of a person into a token embedded in the same space as the image tokens, which can interact with scene tokens to select the relevant content for prediction. For simplicity, we first introduce this process for a single person." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Single Person Case. Let " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{crop}} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " denote the head crop of a person and " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{bbox}} = (x_{\\mathrm{min}}, y_{\\mathrm{min}}, x_{\\mathrm{max}}, y_{\\mathrm{max}}) \\in [0,1]^4" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " her head bounding box. The mapping works as follows. The head crop " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{crop}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is fed to a gaze backbone " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " to produce a gaze embedding " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathrm{emb}} \\in \\mathbb{R}^{d_{\\mathrm{emb}}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": ". This embedding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "is used in two ways. First, it goes through a gaze regressor (i.e. MLP) " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{gv}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " to predict a 2D gaze vector " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{\\mathrm{v}} = \\mathcal{O}_{\\mathrm{gv}}(\\mathbf{g}^{\\mathrm{emb}})" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": ". This output is supervised using an angular gaze loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "content": "Secondly, the gaze embedding is projected to the token dimension using a learnable linear projection " + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{gaze}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "content": ", resulting in the gaze token " + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{emb}} = \\mathcal{P}_{\\mathrm{gaze}}(\\mathbf{g}^{\\mathrm{emb}}) \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "content": ". As we want to incorporate information about the person's location (and size), we also project the head bounding box " + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "content": " into a bounding box embedding " + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{bbox}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "content": " using a learnable linear projection " + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{bbox}} = \\mathcal{P}_{\\mathrm{bbox}}(\\mathbf{h}_{\\mathrm{bbox}}) \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 109, + 547, + 216 + ], + "type": "text", + "content": ". Finally, we add this embedding to the gaze token to obtain the final location-aware gaze token:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 226, + 545, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 226, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 373, + 226, + 545, + 240 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\mathrm {g}} = \\mathbf {x} ^ {\\text {e m b}} + \\mathbf {x} ^ {\\text {b b o x}} \\in \\mathbb {R} ^ {D} \\tag {1}", + "image_path": "ec35f6098e2d0cc66b4855836d4dc8bbdaba2cd795ed90136be7e044ff37a19c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": "Multi-person case. When " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": " persons are detected, the architecture will produce a set of " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": " gaze tokens, following the same process described above for each person. Thus, if " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{bbox}}^i" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{crop}}^i" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": " denote the bounding-box and head crop of person " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": ", the above process will generate a gaze token " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{\\mathrm{g}}" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": " for this person. To simplify notation, we will also denote by " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{g}}" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": " the set of gaze tokens of all people in the scene, with " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{g}} = \\mathbf{x}_1^{\\mathrm{g}} \\oplus \\ldots \\oplus \\mathbf{x}_{N_p}^{\\mathrm{g}}" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 247, + 545, + 345 + ], + "type": "text", + "content": " is the concatenation operator." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "content": "Modality Encoding. Given the different nature of gaze tokens compared to image tokens, we need to encode modality-specific information to distinguish between them. Rather than using an explicit scheme, in practice we expect this modality information to be captured by the bias terms of the different projection operators " + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{gaze}}" + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{img}}" + }, + { + "bbox": [ + 304, + 346, + 545, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 424, + 430, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 430, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 430, + 436 + ], + "type": "text", + "content": "3.3. Transformer Encoder" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "content": "The transformer encoder is a standard ViT [8]. It takes as input the concatenation of the scene tokens " + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{img}}" + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "content": ", the gaze token(s) " + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{g}}" + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "content": ", according to " + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\mathbf{x}^{\\mathrm{img}} \\oplus \\mathbf{x}^{\\mathrm{g}} \\in \\mathbb{R}^{N_t \\times D}" + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "inline_equation", + "content": "N_t = N + N_p" + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "content": ". The set of input tokens goes through a series of " + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "content": " transformer blocks to obtain an output sequence of similar shape, denoted by " + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{out}} = \\mathbf{x}^{(L)} \\in \\mathbb{R}^{N_t \\times D}" + }, + { + "bbox": [ + 304, + 442, + 545, + 515 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 522, + 394, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 394, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 394, + 533 + ], + "type": "text", + "content": "3.4. Gaze Decoder" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "text", + "content": "The goal of the gaze decoder " + }, + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{gaze}}" + }, + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "text", + "content": " is to predict a set of gaze heatmaps. Our Conditional DPT (cf. Figure 4) takes four intermediate representations of the image tokens " + }, + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{(i)}^{\\mathrm{img}}" + }, + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "text", + "content": " and gaze tokens " + }, + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{(i)}^{\\mathrm{g}}" + }, + { + "bbox": [ + 304, + 540, + 545, + 639 + ], + "type": "text", + "content": " and combines them progressively at different simulated resolutions, where lower resolutions have more channels and correspond to deeper layers of the encoder. This can be viewed as the isotropic equivalent of a Feature Pyramid Network [22]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": "Our design is inspired by DPT [28], which can only handle decoding image tokens alone. In our case, we need this decoding to be conditioned on each person. To this end, after each block of layers, the image tokens " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{(l_k)}^{\\mathrm{img}}" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "k \\in \\{4, 8, 16, 32\\}" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": " are reassembled into an image-like representation at resolution " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\left(\\frac{H}{k}, \\frac{W}{k}\\right)" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": " and dimension " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": ". The gaze to" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2010" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 74, + 521, + 318 + ], + "blocks": [ + { + "bbox": [ + 75, + 74, + 521, + 318 + ], + "lines": [ + { + "bbox": [ + 75, + 74, + 521, + 318 + ], + "spans": [ + { + "bbox": [ + 75, + 74, + 521, + 318 + ], + "type": "image", + "image_path": "b6217bc5303033b18b549d5ac042ec60e5c2335598f84db871b70d0e02251c30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 328, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 45, + 328, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 45, + 328, + 545, + 406 + ], + "type": "text", + "content": "Figure 3. Overview of our Sharingan architecture. A. The input image is projected into image tokens (red squares). B. The head crops and head box coordinates are processed to generate location and size-aware person-specific tokens (blue squares) as follows. First, the head crop is fed to a gaze backbone to produce a gaze embedding used to (i) predict a normalized 2D gaze vector that is supervised using an angular loss; and (ii) produce a gaze token by projecting it to the token dimension. Second, head bounding box coordinates are projected to obtain an embedding which, added to the gaze token, produces a person gaze token. C. Image tokens and gaze tokens are fed to the transformer encoder, and the output tokens corresponding to input people are all decoded using a Conditional DPT decoder to predict each person's gaze heatmaps. In addition, input and output gaze tokens are concatenated together to predict the in- vs-out label." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": "kens " + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{(l_k)}^{\\mathrm{g}}" + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": " at that layer are also projected to the same dimension. Next, we duplicate the image feature maps " + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": " times, and apply an element-wise dot-product between each gaze token and a copy of the image feature map. Finally, these person-specific image features are stacked, and we merge the batch and person dimensions to produce a final output of dimension " + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(B\\times N_p,d_k,\\frac{H}{k},\\frac{W}{k})" + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": ". This tensor is passed to a fusion module where it is processed by a small residual convnet and added to the output from the previous fusion block. The result goes through another residual convnet, an upsampling stage to double the resolution, and a projection, leading to a tensor of dimension " + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(B\\times N_p,d_k,\\frac{2H}{k},\\frac{2W}{k})" + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": ". At the end of this process, we get a tensor of dimension " + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(B\\times N_p,d_{\\mathrm{out}},\\frac{H}{2},\\frac{W}{2})" + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": ", which goes through a convolutional head that predicts the heatmaps by bringing the channel dimension down to 1 and resizing the spatial dimension to that of the gaze heatmap. Finally, we separate the batch and person dimensions such that the final shape of the output is " + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "(B,N_{p},1,H_{\\mathrm{hm}},W_{\\mathrm{hm}})" + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": ". The rationale behind this design is to gather information from different layers and resolutions, which is important for dense prediction tasks. In this case, it is particularly useful for gaze tokens where information from the early layers might retain more scene-independent gaze cues due to their proximity to the gaze encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 426, + 411, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 411, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 411, + 438 + ], + "type": "text", + "content": "3.5. In-Out prediction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 445, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 482 + ], + "type": "text", + "content": "The In-Out classifier head " + }, + { + "bbox": [ + 305, + 445, + 545, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{MLP}}" + }, + { + "bbox": [ + 305, + 445, + 545, + 482 + ], + "type": "text", + "content": " consists of an MLP with 7 layers. It is fed the concatenation of input and output gaze tokens to predict a binary in-vs-out label for each person." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 380, + 492, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 492, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 380, + 492, + 545, + 509 + ], + "type": "interline_equation", + "content": "\\mathbf {o} = \\mathcal {O} _ {\\mathrm {M L P}} \\left(\\left[ \\mathbf {x} _ {(L)} ^ {\\mathrm {g}}, \\mathbf {x} ^ {\\mathrm {g}} \\right]\\right) \\tag {2}", + "image_path": "8aadaef148c60a261ebee35b8fcca9e5380bdb13ff45c85bf4614be3e819c0a8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 519, + 480, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 519, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 519, + 480, + 532 + ], + "type": "text", + "content": "3.6. Loss and implementation details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 538, + 533, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 533, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 533, + 551 + ], + "type": "text", + "content": "We train our model using a combination of three losses:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 552, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 545, + 590 + ], + "type": "text", + "content": "Heatmap Loss " + }, + { + "bbox": [ + 306, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{reg})" + }, + { + "bbox": [ + 306, + 552, + 545, + 590 + ], + "type": "text", + "content": ". The heatmap loss is the pixel-wise MSE between the GT heatmap and the predicted heatmap: " + }, + { + "bbox": [ + 306, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hm} = \\sum_{x,y}^{W_{hm},H_{hm}}||\\mathcal{A}_{x,y}^{\\mathrm{gt}} - \\mathcal{A}_{x,y}^{\\mathrm{pred}}||_2^2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "text", + "content": "Angular Loss " + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ang})" + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "text", + "content": ". The angular loss drives the prediction of a normalized gaze direction vector. It maximizes the cosine of the angle between the predicted and ground truth gaze vectors according to: " + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ang} = 1 - < \\mathbf{g}_{\\mathrm{v}}^{gt}, \\mathbf{g}_{\\mathrm{v}}^{pred}>" + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "inline_equation", + "content": "< a, b >" + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "text", + "content": " denotes the inner product between " + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 305, + 590, + 545, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 651, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 545, + 676 + ], + "type": "text", + "content": "In-Out Loss " + }, + { + "bbox": [ + 305, + 651, + 545, + 676 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{io})" + }, + { + "bbox": [ + 305, + 651, + 545, + 676 + ], + "type": "text", + "content": ". The in-out loss is the standard binary cross-entropy for in-vs-out prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 677, + 471, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 471, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 471, + 689 + ], + "type": "text", + "content": "Global loss. The final loss is a given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 701, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 349, + 701, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {r e g} \\mathcal {L} _ {r e g} + \\lambda_ {a n g} \\mathcal {L} _ {a n g} + \\lambda_ {i o} \\mathcal {L} _ {i o} \\tag {3}", + "image_path": "0e2afc1fbd776cd13db5b79d853580612ec08a529064b6978b6696839f0e62ad.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2011" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 523, + 205 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 523, + 205 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 523, + 205 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 523, + 205 + ], + "type": "image", + "image_path": "13c062be1ece1f26981f9b509f2b2a935b8091ef8b8e33c2eb83cf3ec8147348.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 213, + 411, + 224 + ], + "lines": [ + { + "bbox": [ + 181, + 213, + 411, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 213, + 411, + 224 + ], + "type": "text", + "content": "Figure 4. Overview of our proposed Conditional DPT decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 244, + 128, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 244, + 128, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 244, + 128, + 258 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 265, + 111, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 111, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 111, + 277 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "text", + "content": "GazeFollow. GazeFollow [30] is an image-based dataset annotated mostly with head bounding boxes, 2D gaze points. Overall, it has around " + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "inline_equation", + "content": "130\\mathrm{K}" + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "text", + "content": " annotated instances in " + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "inline_equation", + "content": "122\\mathrm{K}" + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "text", + "content": " images. The test set comprises 4782 gaze instances, each of which is labeled by multiple annotators " + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "inline_equation", + "content": "(\\sim 10)" + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 345, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 345, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 287, + 403 + ], + "type": "text", + "content": "VideoAttentionTarget. VideoAttentionTarget [7] is a video-based dataset consisting of 1331 clips from 50 TV shows. It is also annotated with the head bounding boxes, 2D gaze points, and in vs out labels. Overall, it contains 164K instances in 71K frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 406, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 288, + 477 + ], + "type": "text", + "content": "ChildPlay. ChildPlay [35] is a video dataset consisting of 401 clips from 95 YouTube videos of children engaged in play activities. On top of the standard annotations, the authors extend the in vs out label to include other gaze classes (e.g. gaze shift), which we don't use in this paper. Overall, it contains 257K instances in 120K frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 486, + 106, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 106, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 106, + 498 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 505, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 576 + ], + "type": "text", + "content": "We use four metrics to evaluate our gaze following models. The first three are AUC, Distance, and AP which are all standard in the literature [7]. Particularly, the AUC adaptation introduced by [7] for datasets with a single point annotation is not very informative, so we decided to replace it with PLAH, which we define below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": "PLAH. Recently introduced in [35], this metric computes the Precision of looking at people's heads to incorporate semantic information in the evaluation of gaze models. Unlike [35], we consider a prediction positive if the predicted and annotated gaze points fall within the same head box." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 646, + 176, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 176, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 176, + 658 + ], + "type": "text", + "content": "4.3. Experimental Protocol" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Context People. The training of Sharingan relies on processing multiple people at the same time (cf. section 5), but available benchmarks often annotate 1 person per image. To circumvent this problem, we apply an off-the-shelf head" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": "detector" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " trained on the CrowdHuman dataset [31]. We discard detections with a confidence score lower than 0.5 and detections with an IOU score higher than 0.5 with the annotated target person. During training, due to batch constraints, we set " + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "N_{p}^{tr}" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " and keep it fixed. For each image, we use the person with the GT annotation and randomly sample " + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "N_{p}^{tr} - 1" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " (detected) heads when available, otherwise we use padding for the box and head crop. Incidentally, the loss is computed and propagated solely from the annotated person. At evaluation, for each image " + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " we set " + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "N_{p}^{i}" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " to the number of all people in it and process them in a single forward pass using a batch size of 1. Note that " + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "N_{p}^{tr}" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " is a property of the training process, not the architecture, and thus doesn't restrict the " + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "N_{p}^{i}" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " that can be used during inference. Unless stated otherwise, we use " + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "inline_equation", + "content": "N_{p}^{tr} = 2" + }, + { + "bbox": [ + 304, + 245, + 547, + 426 + ], + "type": "text", + "content": " in all experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "text", + "content": "Implementation Details. Sharingan processes the input scene image and head crop at a resolution of " + }, + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "text", + "content": ", while the output heatmap is " + }, + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "text", + "content": ". The gaze backbone " + }, + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 429, + 547, + 501 + ], + "type": "text", + "content": " is a ResNet-18 [13] pretrained on Gaze360 [18], and the transformer encoder is a ViT-base model [9] initialized with weights from a multimodal MAE [2]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "text", + "content": "Training. The models are trained for 20 epochs on GazeFollow. For VideoAttentionTarget and ChildPlay, following standard practices, we take the trained GazeFollow model, freeze everything except the gaze decoder and In-Out classifier, and fine-tune them separately for 2 epochs each. We use the AdamW optimizer [23] with a learning rate of " + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "inline_equation", + "content": "3e - 5" + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "text", + "content": " and a cosine annealing schedule. The fine-tuning uses a learning rate of " + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "inline_equation", + "content": "1e - 6" + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "text", + "content": " for the gaze decoder and " + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "inline_equation", + "content": "3e - 4" + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "text", + "content": " for the In-Out classifier. We also make use of Stochastic Weight Averaging [15] to stabilize training on GazeFollow. The loss coefficients are " + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "inline_equation", + "content": "\\lambda_{reg} = 1000" + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "inline_equation", + "content": "\\lambda_{ang} = 3" + }, + { + "bbox": [ + 304, + 504, + 547, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 639, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 687 + ], + "type": "text", + "content": "Validation. Since GazeFollow [30] and VideoAttentionTarget [7] do not propose any validation splits, we use the train-val splits proposed by [35]. The best model on the validation set is selected based on the distance metric." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 485, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 485, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 485, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 317, + 702, + 485, + 713 + ], + "type": "text", + "content": "https://github.com/deepakcrk/yolov5-crowdhuman" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2012" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 244, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 244, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 244, + 85 + ], + "type": "text", + "content": "4.4. Comparison with the State-of-the-art" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 289, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 289, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 289, + 317 + ], + "type": "text", + "content": "We summarize our quantitative results on the GazeFollow and VideoAttentionTarget datasets in Table 1, and on ChildPlay in Table 2, compared to previous works. Our model sets a new state-of-the-art on all 3 datasets on most metrics and outperforms the only comparable multi-person method by 0.013 on the Avg. Dist. metric. Moreover, the only method that comes close to our results is [12], which we slightly outperform on GazeFollow in both Avg. Dist. and AUC. However, unlike Sharingan, this method uses 2 other modalities (i.e. depth, pose), has a very complex and costly training protocol (i.e. modality-specific backbones are pretrained separately for the task), and is resource-intensive during inference (i.e. single-person). Moreover, it generalizes poorly as evidenced by a cross-dataset evaluation (i.e. Dist. 0.113 vs 0.134 on VideoAttentionTarget, and 0.109 vs 0.142 on ChildPlay). Incidentally, the image version of this method is also significantly worse (i.e. 0.134 vs 0.113). In terms of multi-person comparison, we beat [16] by 0.013 and 0.02 in Avg. Dist. and Min. Dist. respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 317, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 289, + 486 + ], + "type": "text", + "content": "We also see similar results on video datasets (VideoAttentionTarget and ChildPlay), and it is interesting to note that the model trained on GazeFollow is achieving remarkable cross-dataset performance without fine-tuning. Surprisingly, it is already improving on its multi-person competitor by a large margin (i.e. Dist. 0.113 vs 0.134), which is a testament to the generalization ability of our model. We also note that the model from [35] is on par with ours on the video datasets. We believe the reason is these datasets have high-quality images compared to GazeFollow, so the depth-based geometric prior in [35] proves useful, but that method remains multimodal and single-person. We provide more details, discussions, and experiments in the supplementary material to further assess robustness and generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 491, + 174, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 174, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 174, + 504 + ], + "type": "text", + "content": "4.5. Ablation Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 289, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 679 + ], + "type": "text", + "content": "Person Encoding. An important aspect of Sharingan is the way people are encoded into the architecture. Previous methods often represent the head location as a binary mask, and gaze information as a visual attention map [7] or gaze cone [11, 12, 35]. We modify Sharingan to experiment with 3 variants, and opt for single-person training to alleviate the computational cost from some of these formulations: 1. Instead of projecting the box coordinates, we use a head location mask that we tokenize and add the resulting head position embeddings to the image tokens (i.e. Head Mask Embed). 2. Using the same head mask embedding, we tokenize the head crop and append the head tokens to the image ones (i.e. Head Crop Tokens). 3. Instead of processing the head directly, we use the gaze backbone to regress a gaze" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 145 + ], + "type": "text", + "content": "vector which is used to build a gaze cone that we tokenize (i.e. Gaze Cone Tokens). Please note that formulations 2 and 3 double the number of tokens, and we decode the output image tokens into a gaze heatmap using a normal DPT [28]. As we can see from Table 3 (top), the location-aware gaze token is both efficient and performs the best." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 153, + 544, + 303 + ], + "blocks": [ + { + "bbox": [ + 308, + 153, + 544, + 303 + ], + "lines": [ + { + "bbox": [ + 308, + 153, + 544, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 544, + 303 + ], + "type": "table", + "html": "
MethodAUC ↑Avg. D. ↓Min. D. ↓
Head Mask Embed0.9400.1170.060
Head Crop Tokens0.9330.1380.076
Gaze Cone Tokens0.9340.1330.073
Gaze Token0.9440.1130.057
Token to Heatmap0.6470.3020.234
Dot-Product0.9230.1200.062
Up & Dot-Product0.9340.1160.059
Conditional DPT0.9440.1130.057
", + "image_path": "143acfd47c91d46933f0025d135ef4f91c430de6ffac47a8f1005f777c60217b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 311, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 305, + 311, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 545, + 334 + ], + "type": "text", + "content": "Table 3. Ablation results for person encoding (top) and gaze decoding (bottom)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": "Gaze Decoder. We also perform ablations to assess the suitability of our Conditional DPT gaze decoder by comparing it to other baselines as shown in Table 3 (bottom). The first baseline (i.e. Token to Heatmap) regresses a gaze heatmap directly from the output person token using an MLP. This is also the decoding approach undertaken by [39] and [38]. The second baseline (cf. Dot-Product) projects the tokens, then performs a dot-product between each person token and each image token before resizing the output to " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " to get the final heatmap. The third baseline (cf. Up & Dot-Product) upscales the image representation first, then performs the dot-product. We note that our Conditional DPT outperforms all the other methods, justifying the need for a more sophisticated decoding mechanism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 516, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 516, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 516, + 547, + 564 + ], + "type": "text", + "content": "Angular Loss. Our experiments show that the angular loss doesn't affect the final performance, but we decided to keep it anyway to always have a reliable gaze direction, even when the person is looking outside the frame." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 573, + 375, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 375, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 375, + 586 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Model Efficiency. Unlike most previous methods, Sharingan's ability to predict the gaze of multiple people at the same time makes it very efficient for real-world applications. Figure 5 (left) shows a fairly large improvement in inference time compared to a lightweight baseline when " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " increases. It's worth noting that while Sharingan does most of its processing once within the transformer, the small gaze encoder and decoder themselves are executed for each person independently by combining the batch and person dimensions. This explains why the curve is not constant." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 286, + 713 + ], + "type": "text", + "content": "2We omit [38, 39] from the table because their evaluation protocol is different, which makes them incomparable to the rest of the methods. Results are taken from [35]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2013" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 548, + 345 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 548, + 345 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 548, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 548, + 345 + ], + "type": "table", + "html": "
MethodMultiModalityGazeFollowVideoAttentionTarget
AUC ↑Avg. Dist. ↓Min. Dist. ↓PLAH ↑Dist. ↓PLAH ↑AP ↑
Recasens [29]XI0.8780.1900.113————
Chong [6]XI+T0.8960.1870.112—0.171—0.712
Lian [21]XI0.9060.1450.081————
Chong [7]XI+T0.9210.1370.077—0.134—0.853
Fang [11]XI+D+E0.9220.1240.067—0.108—0.896
Fang [11]XI+D————0.124—0.872
Jin [17]XI+D0.9200.1180.063—0.109—0.897
Jin [17]XI0.9090.1370.077————
Tonini [37]XI+D0.9270.141——0.129——
Gupta [12]XI+D+P0.9430.1140.056—0.110—0.879
Gupta [12]XI0.9330.1340.071—0.122—0.864
Bao [3]XI+D+P0.9280.122——0.120—0.869
Hu [14]XI+D+O0.9230.1280.069—0.118—0.881
Tafasca [35]XI+D0.9360.1250.0640.6220.1090.7520.834
Jin [16]✓I0.9190.1260.076—0.134—0.880
\\(Sharingan^†\\)✓I0.9440.1130.0570.6670.1130.748—
Sharingan✓I0.9440.1130.0570.6670.1070.7380.891
", + "image_path": "00c555a24a697d0170d2fa1b183274af0044ed932210f48349b3bc9f745d8471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 403, + 277, + 488 + ], + "blocks": [ + { + "bbox": [ + 46, + 351, + 547, + 387 + ], + "lines": [ + { + "bbox": [ + 46, + 351, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 547, + 387 + ], + "type": "text", + "content": "Table 1. Results of our Sharingan architecture on the GazeFollow and VideoAttentionTarget datasets. The best scores for multi-person models are given in bold, while the best scores in general are underlined. The " + }, + { + "bbox": [ + 46, + 351, + 547, + 387 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 351, + 547, + 387 + ], + "type": "text", + "content": " symbol means that the model was trained on GazeFollow and evaluated without fine-tuning. The modality column uses the codes I (image), T (time), D (depth), E (eyes), P (pose), and O (objects)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 403, + 277, + 488 + ], + "lines": [ + { + "bbox": [ + 59, + 403, + 277, + 488 + ], + "spans": [ + { + "bbox": [ + 59, + 403, + 277, + 488 + ], + "type": "table", + "html": "
MethodMultiDist. ↓PLAH ↑AP ↑
Gupta [12]X0.113—0.983
Tafasca [35]X0.1070.5900.986
Sharingan†✓0.1090.600—
Sharingan✓0.1060.6000.990
", + "image_path": "a3f2e3f7bc80522e4d97203b7081ee6f2e8b3f051b8108c795b90edbceca096e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 496, + 244, + 507 + ], + "lines": [ + { + "bbox": [ + 90, + 496, + 244, + 507 + ], + "spans": [ + { + "bbox": [ + 90, + 496, + 244, + 507 + ], + "type": "text", + "content": "Table 2. Results on the ChildPlay dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 518, + 162, + 639 + ], + "blocks": [ + { + "bbox": [ + 48, + 518, + 162, + 639 + ], + "lines": [ + { + "bbox": [ + 48, + 518, + 162, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 518, + 162, + 639 + ], + "type": "image", + "image_path": "03f804cb84deb5704863d28c658bb3b0b49042ceef98977b0d68f315e3b896f8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 646, + 287, + 680 + ], + "lines": [ + { + "bbox": [ + 46, + 646, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 287, + 680 + ], + "type": "text", + "content": "Figure 5. (Left): Comparison of CPU inference time based on " + }, + { + "bbox": [ + 46, + 646, + 287, + 680 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 46, + 646, + 287, + 680 + ], + "type": "text", + "content": ". (Right): Comparison of single- and multi-person models by " + }, + { + "bbox": [ + 46, + 646, + 287, + 680 + ], + "type": "inline_equation", + "content": "N_{p}^{i}" + }, + { + "bbox": [ + 46, + 646, + 287, + 680 + ], + "type": "text", + "content": " during evaluation on GazeFollow." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 164, + 518, + 279, + 639 + ], + "blocks": [ + { + "bbox": [ + 164, + 518, + 279, + 639 + ], + "lines": [ + { + "bbox": [ + 164, + 518, + 279, + 639 + ], + "spans": [ + { + "bbox": [ + 164, + 518, + 279, + 639 + ], + "type": "image", + "image_path": "dd306b9510cf2f4edf3d285b7d31737290e5080fe2e0522590070c51194ba350.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": "Number of People. One consideration of particular importance in the context of Sharingan is the influence of the number of people on training and evaluation. Table 4 shows that increasing " + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{tr}" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": " during training does not influence performance. We believe that since the gaze tokens are interchangeable and play a symmetric role (i.e. no order encoding), as long as " + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{tr} > 1" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": ", the model will be forced to learn how to accommodate an arbitrary number of people in the scene during inference by sharing the learned image representations among them. However, the same cannot be said for the single-person setting (" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{tr} = 1" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": "), which is fundamentally different from multi-person (" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{tr} > 1" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": "). In this case, the model inevitably learns to tailor its image representations to a single person. To illustrate this behavior, we evaluate a single-person (" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{tr} = 1" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": ") and a multi-person (" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{tr} = 2" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": ") models twice: first using one person (" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{i} = 1" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": ") and a second time by processing all people in the scene (" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "inline_equation", + "content": "N_{p}^{i} = all" + }, + { + "bbox": [ + 304, + 420, + 545, + 685 + ], + "type": "text", + "content": "). Figure 5 (right) shows that the multi-person model is perfectly able to do both single-person and multi-person prediction. However, the single-person model experiences a significant degradation when attempting multi-person prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Heatmap vs 2D Point. Gaze following has always been framed as a heatmap prediction task. To the best of our" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2014" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 261, + 136 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 261, + 136 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 261, + 136 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 261, + 136 + ], + "type": "table", + "html": "
AUCAvg. D. ↓Min. D. ↓
Ntrp=20.9440.1130.057
Ntrp=30.9430.1140.058
Ntrp=40.9420.1130.057
", + "image_path": "0d3e3c5c55dd2b1d23e16129be368d7eb81f1c17892c269a0f6b55bee6ae50cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 288, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 288, + 432 + ], + "type": "text", + "content": "knowledge, Lian et al. [21] were the only authors to experiment with a 2D point regression objective in their ablation study. In this section, we explore this formulation further to gain a deeper understanding of the task. To this end, we trained a Sharingan model by replacing the Conditional DPT decoder with a simple MLP to directly regress " + }, + { + "bbox": [ + 46, + 193, + 288, + 432 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 193, + 288, + 432 + ], + "type": "text", + "content": " coordinates from the output person token. Since the 2D Point model can only predict a single value, it can't represent the entire distribution over the 2D space of the image. Instead, we believe that it converges to an expectation of this posterior probability. When this distribution is multimodal (i.e. more than one probable gaze target), the expectation becomes unlikely under that posterior distribution. We illustrate this behavior in Figure 6 by comparing the person-specific attention map from the last layer of the encoder of the 2D Point model to the predicted gaze heatmap from the Heatmap variant. It is clear that both models capture the different modes quite well, but the former outputs " + }, + { + "bbox": [ + 46, + 193, + 288, + 432 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 193, + 288, + 432 + ], + "type": "text", + "content": " coordinates resembling a weighted average of these modes, which ends up distant from all of them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "text", + "content": "This phenomenon leads to the results given in Table 5, showing that the " + }, + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "text", + "content": " point model achieves better average distance and PLAH, but lags behind the Heatmap model in terms of minimum distance and RLAH. This is not surprising since the average distance of the GazeFollow test set is a distance to an average of modes (i.e. multiple annotated points), which loosely resembles the objective that the " + }, + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "text", + "content": " Point model is optimizing for. Consequently, we believe that the average distance should never be the only metric for evaluating gaze following methods, a claim also supported by [3]. On the other hand, positive LAH predictions for the " + }, + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 46, + 434, + 287, + 639 + ], + "type": "text", + "content": " Point model mean that it is very confident, which explains the better PLAH value. However, this also means that the number of false negatives will be high, hence the big gap in RLAH because the predicted point will often be slightly off-target when the model is not confident (e.g. marginally outside the head area when looking at a person)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Limitation. Sharingan processes both image and gaze tokens simultaneously in the transformer. This means that the same weights operate on both types of tokens, which makes it difficult to understand how the model is combining information. One idea worth exploring is to disentangle scene and person processing, and selectively fuse their informa" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 425, + 137 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 425, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 425, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 425, + 137 + ], + "type": "image", + "image_path": "0ad760e11f5772874226da3ab0c96ff792e051dad9995a9b70db34c5b4a2421d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 545, + 137 + ], + "type": "image", + "image_path": "e8c3fb0a92bbcefdad3466c262b701e4b8429c26a2c514cc86177e4f8f064ccb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 138, + 425, + 205 + ], + "blocks": [ + { + "bbox": [ + 307, + 138, + 425, + 205 + ], + "lines": [ + { + "bbox": [ + 307, + 138, + 425, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 138, + 425, + 205 + ], + "type": "image", + "image_path": "16ff9ec24986c8120aaa8d2fb77cac8da92fd5d71aae494bfee4a8b42df59506.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 213, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 258 + ], + "type": "text", + "content": "Figure 6. Comparison of Heatmap (i.e. left) and 2D Point regression model (i.e. right) models. The heatmap on the right is obtained by computing the attention weights (i.e. last encoder layer) of the person's gaze token with the image tokens." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 138, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 427, + 138, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 427, + 138, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 427, + 138, + 545, + 205 + ], + "type": "image", + "image_path": "3c4edd67233700f4bf730cd3e1cdf2bc535da07ec8902fe40c09a37783689696.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 266, + 535, + 319 + ], + "blocks": [ + { + "bbox": [ + 47, + 144, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 167 + ], + "type": "text", + "content": "Table 4. Influence of the number of people during training on GazeFollow. The evaluation uses all available people." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 266, + 535, + 319 + ], + "lines": [ + { + "bbox": [ + 318, + 266, + 535, + 319 + ], + "spans": [ + { + "bbox": [ + 318, + 266, + 535, + 319 + ], + "type": "table", + "html": "
Avg. D.Min. D.PLAHRLAH
2D Point0.1060.0660.6830.368
Heatmap0.1130.0570.6670.571
", + "image_path": "2efa76fd030d22461c21fd272920fd766e7ac2d3d3cee039a90eb38221cc898c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 327, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 306, + 327, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 545, + 350 + ], + "type": "text", + "content": "Table 5. Comparison between the heatmap and the 2D point training objectives on GazeFollow. RLAH is the recall of LAH." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 373, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 410 + ], + "type": "text", + "content": "tion along the architecture. This may help improve the stability of predictions on videos, and allow people to interact together in a more meaningful way." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 378, + 437 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 446, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 566 + ], + "type": "text", + "content": "In this paper, we introduced Sharingan, a novel and efficient transformer-based architecture for gaze target prediction that is designed to support an arbitrary number of people out of the box. Sharingan stands out for its efficiency both in training and inference, delivering SOTA results on public benchmarks. Furthermore, it demonstrates strong generalization when tested on other datasets and naturalistic scenes. We also validated architectural decisions through extensive ablation experiments and discussed key aspects related to the model and the task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": "Beyond gaze following, Sharingan's intuitive design makes it suitable for other research areas in human behavior understanding. Specifically, the architecture can be extended to perform a multi-faceted analysis of social scenes by integrating different modalities (e.g. depth, semantics), and producing more outputs (e.g. gestures, interactions). We intend to explore this direction further in future work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This research has been supported by the AI4Autism project (Digital Phenotyping of Autism Spectrum Disorders in children, grant agreement no. CR-SII5 202235 / 1) of the Sinergia interdisciplinary program of the SNSF." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2015" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Henny Admoni and Brian Scassellati. Social eye gaze in human-robot interaction: a review. Journal of Human-Robot Interaction, 6(1):25-63, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "text", + "content": "[2] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 348-367. Springer, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "text", + "content": "[3] Jun Bao, Buyu Liu, and Jun Yu. Escnet: Gaze target detection with the understanding of 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14126-14135, 2022. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 267 + ], + "type": "text", + "content": "[4] Bridget K Behe, Patricia T Huddleston, Kevin L Childs, Jiaoping Chen, and Iago S Muraro. Seeing through the forest: The gaze path to purchase. Plos one, 15(10):e0240179, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 335 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 401 + ], + "type": "text", + "content": "[6] Eunji Chong, Nataniel Ruiz, Yongxin Wang, Yun Zhang, Agata Rozga, and James M Rehg. Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In Proceedings of the European conference on computer vision (ECCV), pages 383-398, 2018. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "text", + "content": "[7] Eunji Chong, Yongxin Wang, Nataniel Ruiz, and James M Rehg. Detecting attended visual targets in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5406, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 513 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 579 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 624 + ], + "type": "text", + "content": "[10] Lifeng Fan, Yixin Chen, Ping Wei, Wenguan Wang, and Song-Chun Zhu. Inferring shared attention in social scene videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6460-6468, 2018. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[11] Yi Fang, Jiapeng Tang, Wang Shen, Wei Shen, Xiao Gu, Li Song, and Guangtao Zhai. Dual attention guided gaze target detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11390-11399, 2021. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[12] Anshul Gupta, Samy Tafasca, and Jean-Marc Odobez. A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5041-5050, 2022. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[14] Zhengxi Hu, Kunxu Zhao, Bohan Zhou, Hang Guo, Shichao Wu, Yuxue Yang, and Jingtai Liu. Gaze target estimation inspired by interactive attention. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8524-8536, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[15] P Izmailov, AG Wilson, D Podoprikhin, D Vetrov, and T Garipov. Averaging weights leads to wider optima and better generalization. In 34th Conference on Uncertainty in Artificial Intelligence 2018, UAI 2018, pages 876-885, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "type": "text", + "content": "[16] Tianlei Jin, Zheyuan Lin, Shiqiang Zhu, Wen Wang, and Shunda Hu. Multi-person gaze-following with numerical coordinate regression. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 01-08. IEEE, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[17] Tianlei Jin, Qizhi Yu, Shiqiang Zhu, Zheyuan Lin, Jie Ren, Yuanhai Zhou, and Wei Song. Depth-aware gaze-following via auxiliary networks for robotics. Engineering Applications of Artificial Intelligence, 113:104924, 2022. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "text", + "content": "[18] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Ma-tusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[19] Nayeon Kim and Hyunsoo Lee. Assessing consumer attention and arousal using eye-tracking technology in virtual retail environment. Frontiers in Psychology, 12:665658, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 444, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 488 + ], + "type": "text", + "content": "[20] Jing Li, Zejin Chen, Yihao Zhong, Hak-Keung Lam, Junxia Han, Gaoxiang Ouyang, Xiaoli Li, and Honghai Liu. Appearance-based gaze estimation for asd diagnosis. IEEE Transactions on Cybernetics, 52(7):6504-6517, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 489, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 522 + ], + "type": "text", + "content": "[21] Dongze Lian, Zehao Yu, and Shenghua Gao. Believe it or not, we know what you are looking at! In Asian Conference on Computer Vision, pages 35-50. Springer, 2018. 2, 7, 8, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "text", + "content": "[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "text", + "content": "[24] Manuel Jesús Marin-Jimenez, Andrew Zisserman, Marcin Eichner, and Vittorio Ferrari. Detecting people looking at each other in videos. International Journal of Computer Vision, 106(3):282-296, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 714 + ], + "type": "text", + "content": "[25] Manuel J Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net: revisiting people looking at each other in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3477-3485, 2019. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2016" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[26] Skanda Muralidhar, Rémy Siegfried, Jean-Marc Odobez, and Daniel Gatica-Perez. Facing employers and customers: What do gaze and expressions tell about soft skills? In Proceedings of the 17th International Conference on Mobile and Ubiquitous Multimedia, MUM 2018, Cairo, Egypt, November 25-28, 2018, pages 121-126, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "text", + "content": "[27] Catharine Oertel, Patrik Jonell, Dimosthenis Kontogiorgos, Kenneth Funes Mora, Jean-Marc Odobez, and Joakim Gustafson. Towards an engagement-aware attentive artificial listener for multi-party interactions. Frontiers in Robotics and AI, 8:189, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "text", + "content": "[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "text", + "content": "[29] Adria Recasens*, Aditya Khosla*, Carl Vondrick, and Antonio Torralba. Where are they looking? In Advances in Neural Information Processing Systems (NIPS), 2015. * indicates equal contribution. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 284, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 288, + 326 + ], + "type": "text", + "content": "[30] Adria Recasens, Carl Vondrick, Aditya Khosla, and Antonio Torralba. Following gaze in video. In Proceedings of the IEEE International Conference on Computer Vision, pages 1435-1443, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "text", + "content": "[31] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "text", + "content": "[32] Samira Sheikhi and Jean-Marc Odobez. Combining dynamic head pose-gaze mapping with the robot conversational state for attention recognition in human-robot interactions. Pattern Recognition Letters, 66:81–90, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "type": "text", + "content": "[33] R. Siegfried and J.-M. Odobez. Robust unsupervised gaze calibration using conversation and manipulation attention priors. ACM Transactions on Multimedia Computing, Communications, and Applications, 18(1):1-27, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "type": "text", + "content": "[34] S. Tafasca, A. Gupta, N. Kojovic, M. Gelsomini, T. Maillart, M. Papandrea, M. Schaer, and J.-M. Odobez. The ai4autism project: A multimodal and interdisciplinary approach to autism diagnosis and stratification. In Proceeding of the Int. Conference on Multimodal Interaction workshop, Paris, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "text", + "content": "[35] Samy Tafasca, Anshul Gupta, and Jean-Marc Odobez. Childplay: A new benchmark for understanding children's gaze behaviour. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20935-20946, 2023. 2, 5, 6, 7, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 647 + ], + "type": "text", + "content": "[36] Henri Tomas, Marcus Reyes, Raimarc Dionido, Mark Ty, Jonric Mirando, Joel Casimiro, Rowel Atienza, and Richard Guinto. Goo: A dataset for gaze object prediction in retail environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3125-3133, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[37] Francesco Tonini, Cigdem Beyer, and Elisa Ricci. Multimodal across domains gaze target detection. In Proceedings of the 2022 International Conference on Multimodal Interaction, pages 420-431, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[38] Francesco Tonini, Nicola Dall'Asen, Cigdem Beyer, and Elisa Ricci. Object-aware gaze target detection. In Proceed-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 196 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF International Conference on Computer Vision, pages 21860-21869, 2023. 1, 3, 6, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[39] Danyang Tu, Xiongkuo Min, Huiyu Duan, Guodong Guo, Guangtao Zhai, and Wei Shen. End-to-end human-gazetarget detection with transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2192-2200. IEEE, 2022. 1, 3, 6, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 196 + ], + "type": "text", + "content": "[40] Yu Yu, Gang Liu, and Jean-Marc Odobez. Deep multitask gaze estimation with a constrained landmark-gaze model. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2017" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_content_list.json b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b2d6cbda7bef0475ab5f6d3ae2de17566903dc95 --- /dev/null +++ b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_content_list.json @@ -0,0 +1,1398 @@ +[ + { + "type": "text", + "text": "Sheared Backpropagation for Fine-tuning Foundation Models", + "text_level": 1, + "bbox": [ + 171, + 130, + 797, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyuan Yu $^{1}$ , Li Shen $^{*,2}$ , Liang Ding $^{2}$ , Xinmei Tian $^{1}$ , Yixin Chen $^{3}$ , and Dacheng Tao $^{4}$ $^{1}$ University of Science and Technology of China, Hefei, China; $^{2}$ JD Explore Academy, Beijing, China; $^{3}$ Washington University in St. Louis, MO, USA; $^{4}$ Nanyang Technological University, Singapore \nyuzhiyuan@mail.ustc.edu.cn, xinmei@ustc.edu.cn, chen@cse.wustl.edu \n{mathshenli, liangding.liam, dacheng.tao}@gmail.com", + "bbox": [ + 81, + 191, + 888, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 329, + 313, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fine-tuning is the process of extending the training of pre-trained models on specific target tasks, thereby significantly enhancing their performance across various applications. However, fine-tuning often demands large memory consumption, posing a challenge for low-memory devices that some previous memory-efficient fine-tuning methods attempted to mitigate by pruning activations for gradient computation, albeit at the cost of significant computational overhead from the pruning processes during training. To address these challenges, we introduce PreBackRazor, a novel activation pruning scheme offering both computational and memory efficiency through a sparsified backpropagation strategy, which judiciously avoids unnecessary activation pruning and storage and gradient computation. Before activation pruning, our approach samples a probability of selecting a portion of parameters to freeze, utilizing a bandit method for updates to prioritize impactful gradients on convergence. During the feed-forward pass, each model layer adjusts adaptively based on parameter activation status, obviating the need for sparsification and storage of redundant activations for subsequent backpropagation. Benchmarking on fine-tuning foundation models, our approach maintains baseline accuracy across diverse tasks, yielding over $20\\%$ speedup and around $10\\%$ memory reduction. Moreover, integrating with an advanced CUDA kernel achieves up to $60\\%$ speedup without extra memory costs or accuracy loss, significantly enhancing the efficiency of fine-tuning foundation models on memory-constrained devices.", + "bbox": [ + 76, + 349, + 472, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 779, + 209, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nowadays, Transformer-based models [50] have achieved remarkable results in various domains of deep learning [19, 20]. The trend towards scaling up these models for enhanced performance necessitates substantial data and computational resources, often beyond the reach of many", + "bbox": [ + 75, + 804, + 468, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "researchers [44]. Consequently, fine-tuning pre-trained Transformer models, which are a prominent example of foundation models [5], has become a prevalent strategy. This approach allows for adapting and continuing the training of these models at reduced costs while retaining high performance. Despite the advantages, fine-tuning foundation models on low-end devices, such as personal computers with consumer-level GPUs and other edge devices, presents considerable challenges. These challenges stem primarily from the limited memory and computational capacity of such devices, restricting the feasible model and batch sizes during fine-tuning. This limitation often results in prolonged training periods or compromised performance. Our paper aims to address this issue by exploring efficient methodologies for fine-tuning Transformer models on consumer-level devices, striving to balance performance with training efficiency.", + "bbox": [ + 500, + 330, + 893, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One pervasive challenge in fine-tuning models is effectively managing the training pipeline within the confines of limited device memory [41]. Beyond loading trainable parameters, memory usage in model training includes storing intermediate tensors like gradients and activations. These elements are crucial in backpropagation and become particularly memory-intensive with large input batch sizes or model scales. Although efficient training techniques, such as low-precision and pruning, have been proposed to save memory on model parameters and are widely adopted [23, 31], recent research has shifted focus towards reducing memory consumption of activations, identified as a primary memory bottleneck [7]. Notably, since activations are stored solely for backward computation, recent studies on activation reduction [8, 24, 37] have proposed an asymmetric backpropagation pipeline. This approach maintains a dense forward pass and compresses activations only for gradient computation during backpropagation. Among these, BACKRAZOR by Jiang et al. [24] employs top- $k$ pruning in the forward pass to sparsify activa", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5883", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions, thereby reducing memory requirements for their storage. This method then re-densifies them during the backward pass for gradient computation, achieving significant memory savings. However, since activation pruning methods do not retain the original activations, they typically require additional computations to determine optimal pruning locations, in order to maintain accuracy. This introduces a trade-off between memory reduction and computational overhead, necessitating a careful balance. Furthermore, it presents a challenge in enhancing the efficiency of current computation-intensive activation pruning methods.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our work, addressing this unresolved challenge, we present PREBACKRAZOR, a memory and computation-efficient activation pruning scheme. It aligns with the asymmetric sparsified backpropagation approach, where sparsified activations are preserved solely for the backpropagation process. PREBACKRAZOR operates by dynamically freezing non-essential activation coordinates, thus minimizing the necessity for their sparsification and storage. This method effectively reduces computational load, thereby boosting the overall computation efficiency. Before initiating each iteration's forward pass, we strategically assign a dropout probability to every activation coordinate. This probability is informed by the gradient values associated with these coordinates, derived from prior iterations. This process enables us to selectively freeze certain coordinates during the ensuing backward pass. The chosen probabilities aim to preserve those gradients that are crucial for model convergence. To continuously refine this probability distribution, we utilize a bandit method, an approach inspired by JOINTSPAR [30]. Consequently, during the forward pass, we bypass the pruning and storage of activations that are deemed non-essential for backpropagation. This method effectively diminishes the computational burden typically associated with activation pruning, while simultaneously upholding memory efficiency and training accuracy.", + "bbox": [ + 75, + 261, + 472, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments, conducted on a single RTX 4090 GPU, have shown significant improvements in training efficiency and memory management during the fine-tuning of Vision Transformer (ViT) and BERT models. For ViT, an extensive comparative analysis using the CIFAR datasets [26] revealed that our method, PREBACKRAZOR, outperforms traditional activation pruning techniques like BACKRAZOR in terms of computational efficiency. Notably, PREBACKRAZOR achieves an up to $1.7\\times$ increase in training speed for linear layers, without incurring additional memory overhead. Furthermore, our carefully tailored adaptation of the FLASHATTENTION kernel has significantly accelerated the fine-tuning process, ensuring efficient memory usage. This customized integration results in an impressive $1.7\\times$ boost in training speed, maintaining a balanced computational budget without sacrificing accuracy. In the context of BERT fine-tuning, our approach also results in a modest improve", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ment in time efficiency and final accuracy.", + "bbox": [ + 498, + 90, + 779, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are three-fold:", + "bbox": [ + 517, + 106, + 852, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel activation pruning pipeline to further reduce the computational overhead of previous methods. In addition, we establish the convergence analysis for the proposed algorithm.", + "- We design improved activation storage schemes that jointly reduce the compute and memory. We employ bandit methods to update gradients across activation coordinates of the model layers adaptively. This selective update mechanism efficiently manages activation pruning and gradient computation, thereby reducing memory and computational costs based on BACKRAZOR's asymmetric activation-self-sparsified backpropagation.", + "- We conduct experiments on fine-tuning Transformer-based ViT and BERT, adapt the current code library, and combine the state-of-the-art FLASHATTENTION 2 CUDA [16] kernels to boost the training, making the pipeline more welcome by nowadays consumer-level devices. Our experiments show that our method reduces over $30\\%$ of training time without loss of accuracy." + ], + "bbox": [ + 500, + 123, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 416, + 640, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly review existing memory and computation-efficient fine-tuning methods.", + "bbox": [ + 498, + 441, + 890, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Memory-efficient fine-tuning methods. In memory-limited environments, strategies like partitioning and compression leverage trade-offs in computation, communication, and storage to reduce memory usage. Gradient accumulation lowers memory usage at the cost of increased communication, while gradient checkpointing [12] achieves memory savings through recomputation. CPU offloading [42], GPU partitioning [41], and FLASHATTENTION [16, 17] balance memory optimization against communication overhead. Several memory-efficient optimizers [32, 43, 49] have been proposed to approximate backpropagation to minimize stored information. Recent approaches include computing only forward passes with estimated gradients [34] and integrating gradient computation with parameter updates [33]. Additionally, low or mixed-precision training [27, 35] and other sparse backpropagation methods [13] have been integrated, trading some numerical accuracy for enhanced memory or communication efficiency. There are also related works focusing on communication cost reduction in distributed systems that use these memory-efficient methods for gradient compression [1, 4, 45], which is different from our settings. Additionally, while quantization [9, 18] and pruning [47] methods are commonly employed to compress gradients and model weights, our research concentrates on the pruning of activations. In the realm of activation footprint reduction, Cai et al. [7] suggest freezing layer weights while updating biases and adding lightweight residual modules. Other", + "bbox": [ + 496, + 477, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5884", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods involve compression of activations [2, 11, 54]. To maintain accuracy, recent research proposes asymmetric sparsified backpropagation which stores compressed activations, using original dense inputs for forward computation and retrieving activations for backward gradient computations. They employ techniques like low-precision compression [8, 37], or utilizing top- $k$ masks for pruning activations [24].", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Computation-efficient fine-tuning methods. Existing computation-efficient fine-tuning methods include PEFT, knowledge distillation [22], few-shot learning [52], as well as quantization and pruning implementations. Our research primarily focuses on techniques that dynamically adjust training and seamlessly integrate with our sparse backpropagation pipeline. In the realm of computation-efficient fine-tuning methods, we find adaptive learning rates [25, 49] relevant for their role in facilitating faster convergence [10, 46, 56]. Also, dynamic sparse algorithms [29, 36] are notable for introducing sparsity during training. For the aspect of sparse backpropagation, the implementation of top-k masking over gradients by Sun et al. [48] to induce sparsity in backpropagation is in line with our approach to asymmetric pruning. Structured sparsity patterns [21, 55] and layer freezing techniques [6, 40], along with selective weight freezing strategies [28], offer valuable insights into reducing computational load. We also draw specific elements from the JOINTSPAR algorithm [30], which selectively drops gradient vectors, a principle we adapt to target the computational challenges in activation pruning.", + "bbox": [ + 75, + 220, + 472, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Rethink the Backpropagation Redundancy", + "text_level": 1, + "bbox": [ + 76, + 553, + 460, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present explanations of the memory cost of activations, and the asymmetric self-sparsified activations as introduced in BACKRAZOR algorithm [24]. Then we show the memory and computation redundancy in the current backpropagation scheme.", + "bbox": [ + 75, + 579, + 468, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 664, + 207, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider the nonconvex optimization problem of fin-tuning foundation models: $\\min_{\\theta \\in \\Theta} \\mathcal{L}(\\theta)$ , where $\\mathcal{L}$ is the training loss calculated with the model's weight $\\theta$ , and $\\Theta$ is the parameter space. Without loss of generality, the forward pass of a given neural network with $n$ layers can be expressed as follows:", + "bbox": [ + 75, + 688, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\mathbf {x} _ {0}; \\boldsymbol {\\theta}\\right) = f _ {n} \\left(f _ {n - 1} \\left(\\dots f _ {0} \\left(\\mathbf {x} _ {0}; \\boldsymbol {\\theta} _ {0}\\right) \\dots ; \\boldsymbol {\\theta} _ {n - 1}; \\boldsymbol {\\theta} _ {n}\\right), \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 792, + 467, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL = \\mathcal {L} (f (\\mathbf {x} _ {0}; \\boldsymbol {\\theta}), y) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 811, + 467, + 827 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_0$ represents the input features at the 0-th layer, and $y$ corresponds to the input label. The functions $f_{i}$ and parameters $\\theta_{i}$ are associated with the model's $i$ -th layer. The expression $f(\\mathbf{x}_0; \\boldsymbol{\\theta})$ indicates the model's final prediction.", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, the symbol $L$ stands for the loss between the prediction and the label, computed using the specified loss function $\\mathcal{L}$ . We define the activation $\\mathbf{x}_i$ as the output of the $(i - 1)$ -th layer, i.e., $\\mathbf{x}_i = f_{i - 1}(\\mathbf{x}_{i - 1};\\pmb{\\theta}_{i - 1})$ .", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoting $h_i^{(\\mathbf{x})}(\\mathbf{x}_i,\\pmb {\\theta}_i)$ and $h_i^{(\\pmb \\theta)}(\\mathbf{x}_i,\\pmb {\\theta}_i)$ as the derivatives of $\\mathbf{x}_{i + 1}$ with respect to $\\mathbf{x}_i$ and $\\pmb{\\theta}_i$ respectively. Normally, the backward process can be defined as:", + "bbox": [ + 496, + 152, + 890, + 199 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\mathbf {x})} \\left(\\mathbf {x} _ {i}, \\boldsymbol {\\theta} _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 210, + 890, + 241 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\boldsymbol {\\theta})} (\\mathbf {x} _ {i}, \\boldsymbol {\\theta} _ {i}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 243, + 890, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For instance, considering a linear layer with parameters $\\theta_{i}$ , which consist of a weight matrix $\\mathbf{W}_i$ and a bias vector $\\mathbf{b}_i$ , its forward pass function for computing the output $\\mathbf{x}_{i + 1}$ is expressed as $\\mathbf{x}_{i + 1} = \\mathbf{x}_i\\mathbf{W}_i^\\top +\\mathbf{b}_i$ . All of the variables $\\mathbf{x}_i$ , $\\mathbf{W}_i$ , and $\\mathbf{b}_i$ are retained for gradient computation before the layer produces the output $\\mathbf{x}_{i + 1}$ . Subsequently, during the backward pass of this layer in the loss backpropagation process, the gradients are calculated as:", + "bbox": [ + 496, + 286, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\mathbf {W} _ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 417, + 890, + 449 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial \\mathbf {W} _ {i}} = \\left(\\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}}\\right) ^ {\\top} \\mathbf {x} _ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 452, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial \\mathbf {b} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 484, + 890, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, the activation $\\mathbf{x_i}$ is involved in computing the gradient for updating the weight $\\mathbf{W}_i$ . Consequently, the activation value must be stored in memory during the forward pass. This storage occupies a significant portion of memory, often exceeding the memory required for saving the weights. Additionally, the size of the activations increases linearly with the batch size.", + "bbox": [ + 496, + 526, + 890, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Jiang et al. [24] propose BACKRAZOR, a method that only prunes the activation saved for backpropagation and keeps it dense in the forward pass. The activation $\\mathbf{x}_i$ is pruned and stored in the context as a sparse version $\\tilde{\\mathbf{x}}_i$ . The backward process with pruned activations turns into:", + "bbox": [ + 496, + 632, + 890, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\mathbf {x})} \\left(\\widetilde {\\mathbf {x}} _ {i}, \\boldsymbol {\\theta} _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 719, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\boldsymbol {\\theta})} (\\widetilde {\\mathbf {x}} _ {i}, \\boldsymbol {\\theta} _ {i}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 753, + 890, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As for their activation pruning, a straightforward top- $k$ pruning method is applied, which identifies and removes the $k$ smallest magnitude values within the activation. Denoting $v$ as the total number of values of the activation, to introduce a targeted sparsity level, a fixed parameter $\\lambda$ is set (e.g., $\\lambda = 0.9$ ), and the pruning budget is determined by setting $k = \\lambda v$ . Consequently, all activation values below the $k$ -th", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5885", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f72e5ae042119908f4351ea7777ec6778dec1c770995bac378f60d0d7fe36005.jpg", + "image_caption": [ + "Figure 1. Sparsity vs computation time cost between BackRazor (element-wise, self-sparsified) and proposed PreBackRazor (column-wise, joint-sparsified)." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 424, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "smallest value are zeroed out. After this pruning operation, the pruned activation is stored as a binary bitmap, known as the mask. This mask indicates the positions of the nonzero elements. Additionally, a dense matrix is maintained to retain the values of the non-zero elements.", + "bbox": [ + 75, + 305, + 467, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Rethink the sparsity in BACKRAZOR", + "text_level": 1, + "bbox": [ + 76, + 391, + 382, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Redundancy in full-gradient computation. Common backpropagation activation sparsification schemes involve restoring compressed activations, saved within the model context, back to their dense format during backpropagation. In the case of BACKRAZOR, this process entails densifying the sparse matrix to its original shape for gradient computation. Consequently, the full gradient is computed using the full-size activation, and the sparsity of the pruned activation is not utilized to expedite computation.", + "bbox": [ + 75, + 422, + 467, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computation overhead in pruning. A prevalent pruning approach involves generating an element-wise self-sparsified top- $k$ mask to eliminate smaller values. This method, when employed for activation pruning, incurs additional computational costs. These costs include sorting to find the $k$ -th smallest activation value and performing element-wise comparisons across all activation elements to construct a sparse tensor. Collectively, these operations contribute to significant computational overhead, potentially extending training time beyond that of traditional fine-tuning. Such a trade-off for memory reduction may not be justifiable. Exploring alternative sparsity patterns and pruning schemes, such as column-wise and joint-sparse could lead to the discovery of faster algorithms, potentially replacing the current self-sparsified top- $k$ activation pruning method and thus reducing computation costs.", + "bbox": [ + 75, + 568, + 467, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the observed inefficiencies in existing methods, we sought to investigate whether alternative sparsity patterns could offer reduced computation. Specifically, we compared the performance of column-wise sparsity with element-wise sparsity within the same linear layer, across various sparsity ratios, as depicted in Figure 1. Our find", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/909f9ee58bea531031d905889bef4a4bdb1837f9d38e2df499118edf5a6ff301.jpg", + "image_caption": [ + "Figure 2. Pipeline of the proposed PreBackRazor." + ], + "image_footnote": [], + "bbox": [ + 535, + 85, + 893, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ings indicate notable up to $1.7 \\times$ speedups, aligning with our initial hypothesis. Additionally, the column-wise activation sparsity pattern facilitates the use of existing joint sparse algorithms, which leverage gradient information to generate sparse masks previous to the BACKRAZOR pruning pipeline. This approach incurs lower computational costs. Given that the original sparsity is unstructured and of a low ratio, the integration of a structured sparsity pattern also holds promise for further acceleration, leveraging modern sparse computation software and hardware advancements [14, 21, 39].", + "bbox": [ + 496, + 340, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 523, + 633, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first propose the PREBACKRAZOR method. After that, we introduce the bandit method from JOINTSPAR algorithm [30] that we use to adjust the probability of dropping computations. Lastly, we examine the convergence of our proposed method.", + "bbox": [ + 496, + 549, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overall pipeline of PREBACKRAZOR", + "text_level": 1, + "bbox": [ + 498, + 635, + 802, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our jointly sparsified activation pruning approach, PREBACKRAZOR, employs asymmetric backpropagation, maintaining dense computations in the forward pass while pruning activations for the backward pass, as depicted in Fig. 2. Inspired by the JOINTSPAR algorithm [30], our method goes a step further by incorporating gradient sparsity into the activation pruning, aiming to significantly boost computational efficiency. In PREBACKRAZOR, we manage a probability distribution over activation coordinates to selectively prune redundant indices. This process effectively reduces the computation of related gradients and halts the updating of corresponding weights, thereby enhancing training efficiency and minimizing computational requirements. The decision-making process hinges on the analysis of gradient norms, leveraging the joint sparsity of gradients and activations. This approach", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5886", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prioritizes the retention of the most impactful gradient coordinates for weight updates. The implementation details of this method are outlined in Algorithm 1.", + "bbox": [ + 75, + 90, + 470, + 137 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 PREBACKRAZOR" + ], + "code_body": "1: Initialize all $p_0 = s / D$ \n2: for $t = 1$ to $T$ do \n3: Sample $\\mathbf{x}_0, \\mathbf{y}$ from $\\mathcal{D}$ \n4: for $i = 1$ to $n$ do \n5: Instantiate $Z_{t,d}$ , with $P(Z_{t,d} = 1) = p_{t,d}$ \n6: Determine $S_t = \\{d : Z_{t,d} = 1\\}$ \n7: $\\mathbf{x}_i = f_i(\\mathbf{x}_{i-1}, \\boldsymbol{\\theta}_i)$ \n8: mask = a binary bitmap of size $D$ \n9: for parameter $\\boldsymbol{\\theta}_{[d]}$ of $\\boldsymbol{\\theta}_i$ do \n10: if $d \\in S_t$ then \n11: $\\text{mask}_d = \\text{True}$ \n12: end if \n13: end for \n14: $\\tilde{\\mathbf{x}}_{i-1} = \\text{Sparsify}(\\mathbf{x}_{i-1}, \\text{mask})$ \n15: SaveForBackward( $\\tilde{\\mathbf{x}}_{i-1}, \\text{mask}, \\boldsymbol{\\theta}_i$ ) \n16: end for \n17: Compute loss $L = \\mathcal{L}(\\mathbf{x}_{n+1}, \\mathbf{y})$ \n18: for $i = n-1$ to 1 do \n19: $\\text{Densify}(\\tilde{\\mathbf{x}}_{i-1}, \\text{mask})$ \n20: $\\frac{\\partial L}{\\partial \\mathbf{x}_i} = \\frac{\\partial L}{\\partial \\mathbf{x}_{i+1}} h_i^{(x)}(\\tilde{\\mathbf{x}}_i, \\boldsymbol{\\theta}_i)$ \n21: for parameter $\\boldsymbol{\\theta}_{[d]}$ with $d \\in S_t$ do \n22: $\\frac{\\partial L}{\\partial \\boldsymbol{\\theta}_{[d]}} = \\frac{\\partial L}{\\partial \\mathbf{x}_{i+1}} h_i^{(\\theta)}(\\tilde{\\mathbf{x}}_i, \\boldsymbol{\\theta}_{[d]})$ \n23: end for \n24: $p_{t+1} = \\text{Update}(p_t, S_t, \\{\\mathbf{g}_{t,[d]}\\}_{d \\in S_t})$ \n25: end for \n26: end for", + "bbox": [ + 80, + 172, + 468, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a given input activation $\\mathbf{x}_i$ in the $i$ -th layer of the model, we decompose the activations column-wise, represented as $\\mathbf{x}_i = [\\mathbf{x}_{i,1},\\dots ,\\mathbf{x}_{i,d},\\dots ,\\mathbf{x}_{i,D}]$ , where $d$ and $D$ respectively denote the $d$ -th coordinate out of all $D$ column indices. Considering a linear layer, its weight parameter $\\pmb{\\theta}$ is similarly decomposed into $[\\pmb{\\theta}_{[1]},\\dots ,\\pmb{\\theta}_{[d]},\\dots ,\\pmb{\\theta}_{[D]}$ . To differentiate from the layer-wise parameter set $\\pmb{\\theta}_i$ (as discussed in Section 3.1), we use square brackets to denote each coordinate, represented as $\\pmb{\\theta}_{[d]}$ . The gradient of the loss with respect to $\\pmb{\\theta}_{[d]}$ is denoted by $\\pmb{g}_{t,d}$ , and we define $\\pmb{g}_{t,[d]}$ as a vector with all elements zero except the $d$ -th element: $\\pmb{g}_{t,[d]} = [0,\\dots ,\\pmb{g}_{t,d},\\dots ,0]$ , with $\\pmb{g}_t$ being the sum of all such vectors, $\\pmb{g}_t = \\sum_{d = 1}^D\\pmb{g}_{t,[d]}$ .", + "bbox": [ + 75, + 561, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each iteration $t$ , a probability distribution $p_t = [p_{t,1}, \\dots, p_{t,d}, \\dots, p_{t,D}]$ is maintained over the joint coordinates of the parameters $[\\pmb{\\theta}_{[1]}, \\dots, \\pmb{\\theta}_{[d]}, \\dots, \\pmb{\\theta}_{[D]}\\]$ and activations $[\\mathbf{x}_1, \\dots, \\mathbf{x}_d, \\dots, \\mathbf{x}_D]$ . For each coordinate, a Bernoulli random variable $Z_{t,d}$ is sampled with $P(Z_{t,d} = 1) = p_{t,d}$ and $P(Z_{t,d} = 0) = 1 - p_{t,d}$ , deciding whether to include $(Z_{t,d} = 1)$ or exclude $(Z_{t,d} = 0)$ the gradient of $\\pmb{\\theta}_d$ during backpropagation.", + "bbox": [ + 75, + 763, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initially, each $p_t$ is set uniformly to $\\frac{s}{D}$ , aligning with the", + "bbox": [ + 96, + 885, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "desired sparsity budget $s$ , where $s$ signifies the number of columns involved in the backward process. At the onset of each iteration $t$ , the Bernoulli random variable $Z_{t,d}$ is sampled to form the active set $S_{t}$ of coordinates that retain gradients in activations. Post gradient computation in backpropagation, this gradient information informs the update of the probability distribution, employing a bandit method which we will introduce in the subsequent section.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integrating the FLASHATTENTION kernel. The fine-tuning of Transformer-based models on consumer-grade devices presents a unique opportunity to leverage modern, efficient kernels for improved training efficiency. In our approach, we have adapted and integrated the FLASHATTENTION kernel, substituting the original, more segmented sparsified attention implementation in BACKRAZOR. To address the increased memory usage resultant from this integration, we have incorporated PREBACKRAZOR within the FLASHATTENTION function. This inclusion not only facilitates pruning but also employs recomputation strategies to mitigate precision loss. We designate this enhanced methodology as PREBACKRAZOR++. This innovative approach ensures memory usage remains on par with BACKRAZOR, while markedly accelerating the training process. We further combine JOINTSPAR over model parameters with our method as the PREBACKRAZOR++.JOINTSPAR for best memory and computation efficiency. Details on implementation are omitted here for brevity.", + "bbox": [ + 496, + 219, + 892, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Bandit method for gradient sparsification", + "text_level": 1, + "bbox": [ + 500, + 517, + 854, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To update the probability distribution $p_t$ over $D$ coordinates at each iteration $t$ , we utilize a bandit method akin to that in JOINTSPAR [30], originally introduced by Auer et al. [3], as detailed in Algorithm 2.", + "bbox": [ + 496, + 540, + 890, + 601 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Bandit [30] for Updating Distribution $p_t$" + ], + "code_body": "1: Algorithm: UPDATE $(p_t, S_t, \\{g_{t,[d]}\\}_{d \\in S_t})$ \n2: Input: $p_t$ : probability distribution, $S_t$ : set of active parameters, $\\{g_{t,[d]}\\}_{d \\in S_t}$ : gradients \n3: for $d = 1$ to $D$ do \n4: if $d \\in S_t$ then \n5: $\\tilde{l}_{t,d} = -\\frac{\\|g_{t,[d]}\\|^2}{(p_{t,d})^2} + \\frac{G^2}{p_{\\min}^2}$ \n6: else \n7: $\\tilde{l}_{t,d} = 0$ \n8: end if \n9: $w_{t,d} = p_{t,d} \\exp(-\\alpha_p \\tilde{l}_{t,d} / p_{t,d})$ \n10: end for \n11: $p_{t+1} = \\arg \\min_{q \\in \\mathcal{P}} D_{\\mathrm{KL}}(q \\| w_t)$ \n12: return $p_{t+1}$", + "bbox": [ + 506, + 633, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a set $\\mathcal{P}$ as $\\mathcal{P} = \\{p\\in \\mathbb{R}^D:\\sum_{d = 1}^D p_d = s,p_d\\geq p_{\\mathrm{min}},\\forall 1\\leq d\\leq D\\}$ , where $p_{\\mathrm{min}}$ is a constant lower bound. This set helps in guiding the update of the probability distribution. The sum of the probabilities across all", + "bbox": [ + 496, + 838, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5887", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "parameters equals the desired sparsity budget $s$ , with each probability bounded below by a constant $p_{\\mathrm{min}}$ . Note that $s$ and $p_{\\mathrm{min}}$ are parameters defined by the user, adhering to the conditions $0 < s \\leq D$ and $0 < p_{\\mathrm{min}} \\leq 1$ . Recall that $g_{t,[d]}$ is defined as $g_{t,[d]} = [0,\\dots ,g_{t,d},\\dots ,0]$ , and $g_{t} = \\sum_{d = 1}^{D}g_{t,[d]}$ . The term $\\alpha_{p}$ denotes the step size, while $D_{\\mathrm{KL}}(q||w)$ represents the KL divergence between distributions $q$ and $w$ .", + "bbox": [ + 76, + 90, + 472, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Theoretical results", + "text_level": 1, + "bbox": [ + 76, + 222, + 256, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we analyze the convergence of our proposed method PREBACKRAZOR++JOINTSPAR in the context of linear networks. We use $g_{t}$ to represent the gradient of the loss function $\\mathcal{L}$ with respect to the parameter $\\theta_{t}$ at iteration $t$ , calculated using all samples in the training dataset. In contrast, $\\tilde{g}_{t}$ denotes the stochastic gradient computed using a single training sample. To support our analysis, we adopt the classical assumptions as detailed by Jiang et al. [24]:", + "bbox": [ + 75, + 244, + 468, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assumption 1. (1) The objective function $\\mathcal{L}$ is $S_{d}$ -smooth with respect to $\\pmb{\\theta}_{[d]}$ ; (2) The gradient $\\pmb{g}_{t,[d]}$ is upper bounded by $G$ , i.e., $G \\geq \\| \\pmb{g}_{t,[d]} \\|$ . $\\forall t, d$ ; (3) The objective function $\\mathcal{L}$ is bounded below by $L^{*}$ ; (4) For each iteration $t$ , the stochastic gradient $\\tilde{\\pmb{g}}_t$ adheres to the following: $\\mathbb{E}[\\tilde{\\pmb{g}}_t] = \\pmb{g}_t$ , $\\mathbb{E}[(\\tilde{\\pmb{g}}_t - \\pmb{g}_t)_d^2] \\leq \\sigma_d^2$ , $\\forall d \\in \\{1, 2, \\dots, D\\}$ , with $D$ denoting the number of coordinates and the constant $\\sigma_d^2$ representing the variance corresponding to coordinate $d$ .", + "bbox": [ + 75, + 366, + 470, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 1 [24]. In the context of linear neural networks, the gradient of parameters at iteration $t$ can be decomposed as $\\tilde{g}_t = \\tilde{\\mathbf{x}}_t\\tilde{\\mathbf{A}}_t^\\top$ , where $\\tilde{\\mathbf{x}}_t$ denotes the precise activation, and $\\tilde{\\mathbf{A}}_t$ is referred to as the transformation matrix. After pruning the activations, only the pruned activations $\\tilde{\\mathbf{x}}_t'$ are reserved and used for backpropagation, and the gradient approximation becomes $\\tilde{g}_t' = \\tilde{\\mathbf{x}}_t'\\tilde{\\mathbf{A}}_t^\\top$ .", + "bbox": [ + 75, + 493, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 1 (Convergence). Denote $\\alpha$ as the step size. Under Assumptions 1-4, the following bound holds:", + "bbox": [ + 75, + 604, + 468, + 635 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} \\left[ \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\| \\boldsymbol {g} _ {t, i} \\| ^ {2} \\right] \\leq \\frac {\\mathcal {L} \\left(\\theta_ {0}\\right) - L ^ {*}}{\\alpha T} \\\\ + \\frac {S _ {\\infty} \\alpha}{2 T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\mathbb {E} \\left[ \\sum_ {d = 1} ^ {D} \\frac {\\| \\boldsymbol {g} _ {t , [ d ]} \\| ^ {2}}{p _ {t , d}} \\right], \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 642, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T$ is the total number of iterations, $S_{\\infty} = \\max_d S_d$ , $\\theta_0$ is the initial model weights.", + "bbox": [ + 75, + 734, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the training progresses, the first term on the right-hand side of the equation progressively diminishes to zero, whereas the second term remains confined within the bounds set by the bandit method. This theorem establishes that our proposed method converges towards a bounded neighborhood around a stationary point. While our theoretical analysis here is confined to linear networks, the extension to non-linear networks remains an area for future investigation. In the meantime, we have empirically validated", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that our proposed method effectively accelerates the finetuning of Transformer-based models. For those interested in a more thorough exploration of the underlying mathematics, the detailed derivation of these results is provided in the appendix. Additionally, the proofs and analyses in [24] and [30] offer further insight into these concepts.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 194, + 633, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform experiments on PyTorch [38] with RTX 4090 GPU with 24GB of VRAM. We apply our method on top of BACKRAZOR's code library to report the performance gains, following its training settings on ViT and BERT. We report the actual memory allocation on GPU measured with PyTorch CUDA tools. For better readability, we employ the abbreviation JS to denote JOINTSPAR, and PREBR as a shorthand for PREBACKRAZOR.", + "bbox": [ + 496, + 219, + 890, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. PREBACKRAZOR with Vision Transformer", + "text_level": 1, + "bbox": [ + 498, + 349, + 849, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. For ViT training, we initiate the process by loading the checkpoint of the ViT-B/16 model that has been pre-trained on ImageNet-22K as our starting point. Subsequently, we fine-tune the model using the CIFAR-10 and CIFAR-100 datasets [26]. The fine-tuning process encompasses 20,000 steps and employs cosine learning rate decay, with an initial 500 steps dedicated to learning rate warmup. We utilize the standard SGD optimizer and conduct validation every 1,000 steps. The maximum learning rate is set at 0.01, and the batch size during training is 128.", + "bbox": [ + 496, + 372, + 890, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d8d823856cf3a1fe7e2487de151bbc0e993c3a0eb23fbf6c88e0693c4d65ddd.jpg", + "image_caption": [ + "(a) Training." + ], + "image_footnote": [], + "bbox": [ + 506, + 537, + 691, + 642 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/72fecaff58b37f1945b8863c0267e4238335b69ba4f6f0c35b304587e8f76042.jpg", + "image_caption": [ + "(b) Validation.", + "Figure 3. Fine-tuning PREBACKRAZOR++@90% vs. BACKRAZOR@90% on CIFAR-100. (a) Training loss with wall-clock time. (b) Validation accuracy with steps." + ], + "image_footnote": [], + "bbox": [ + 700, + 537, + 883, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Convergence comparison. The results of fine-tuning PREBACKRAZOR++ with parameter-wise JOINTSPAR in comparison to BACKRAZOR at a pruning rate of $90\\%$ on CIFAR-100 are illustrated in Fig. 3. In Fig. 3a, the convergence speed of PREBACKRAZOR@90%++JOINTSPAR is comparable to that of BACKRAZOR@90%, with a slight advantage in the early stages of training. Notably, we observe a marginal improvement in accuracy, specifically $[0.33\\%, 0.05\\%]$ , when adjusting the initial probabilities of JOINTSPAR to $[70\\%, 50\\%]$ , respectively. This improvement is in contrast to the performance of BACKRAZOR@90%, as illustrated in Fig. 3b.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5888", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3754d4a9c6c30ce2a73e8da5c46a3e0d27b8163bffab6f87dcf72f96714bdd3e.jpg", + "table_caption": [ + "Table 1. Comparison of fine-tuning methods for ViT." + ], + "table_footnote": [], + "table_body": "
MethodMemoryEpoch TimeSpeedupCIFAR-100CIFAR-10CUB-200
AllocatedReduction
FULL-FT19980MB1.0×2min59s-93.099.085.5
BACKRAZOR7540MB2.6×6min19s1.0×92.198.686.6
(ours) PREBR++JS@70%7440MB2.7×4min37s1.4×92.498.885.0
(ours) PREBR++JS@50%6880MB2.9×4min02s1.6×92.198.685.6
", + "bbox": [ + 119, + 104, + 854, + 205 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b66f23c08c66f0a2a843be9f30a99f61fbb985b7eac13dc84620dcfd3b1b91a3.jpg", + "table_caption": [ + "Table 2. Performance of different attention implementations fine-tuning ViT on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
DeviceMethodAttention TypeMemoryEpoch TimeSpeedupAccuracy(%)
RTX 4090FULL-FTvanilla19980MB2min59s-93.0
BACKRAZORBACKRAZOR7540MB6min19s1.0×92.1
(ours) PREBR++JSBACKRAZOR6570MB4min37s1.4×91.4
(ours) PREBR++JSFLASHATTENTION7960MB3min40s1.7×92.2
(ours) PREBR++JSsparse FLASHATTN6880MB4min02s1.6×92.1
RTX 2080 TiFULL-FTvanillaout of memory
BACKRAZORBACKRAZOR7780MB19min47s1.0×-
(ours) PREBR+JSBACKRAZOR6870MB15min13s1.3×-
", + "bbox": [ + 119, + 228, + 854, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speedup and memory reduction. In Tab. 1, we present a comprehensive comparison among PREBACKRAZOR@90%++JOINTSPAR@[70%, 50%], BACKRAZOR@90%, and FT-Full (fine-tuning the full network) from the perspective of memory usage, speedup relative to BACKRAZOR@90%, and final accuracy on CIFAR-100, CIFAR-10 and CUB-200 for each method. Here we present the average memory allocated by CUDA during training. Compared with FT-full, BACKRAZOR@90% achieves a noteworthy $2.6\\times$ reduction in memory usage. However, it demands twice as much training time due to the additional computations required for processing activations. In contrast, PREBR++JS achieves superior computational efficiency by avoiding unnecessary computations through joint sparsification of gradients and activations. This leads to significantly $1.6\\times$ faster training without incurring additional memory costs or compromising accuracy. Additionally, Tab. 3 showcases the performance gains of forward and backward computation achieved by transitioning from a self-sparsified top- $k$ pruning pattern to a column-wise joint-sparse pattern as in Fig. 1. These measurements are conducted on a linear layer characterized by input tensors with the shape [128, 197, 768] and of type float32. Tab. 4 illustrates the performance improvements achieved by substituting the element-wise mask in BACKRAZOR with a column-wise mask. This approach, referred to as PREBACKRAZOR-, employs a self-sparsified scheme that eschews the use of the bandit method for predicting pruning masks based on gradient information. Despite this, PREBACKRAZOR- still delivers commendable results, including a modest reduction in memory usage and over a $10\\%$ increase in speed, while preserving nearly identical accuracy, particularly at a moderate pruning ratio of $80\\%$ .", + "bbox": [ + 76, + 397, + 472, + 896 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/09340828259a715b237c6158a6219fcd560caf4bbd137a09efa75ebf0ca6a374.jpg", + "table_caption": [ + "Table 3. Acceleration on a linear layer." + ], + "table_footnote": [], + "table_body": "
SparsityFWD time (ms)SpeedupFWD + BWD (ms)Speedup
BACKRAZORPREBRBACKRAZORPREBR
0.14.233.141.34×9.027.721.17×
0.24.354.241.03×9.209.031.02×
0.34.454.351.02×9.349.181.02×
0.47.974.451.79×16.529.321.77×
0.58.947.971.12×18.4516.511.12×
", + "bbox": [ + 501, + 411, + 890, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/823ae95b2771f501d9d94a4c73a702587ac0751a50e751fa33961a589247339f.jpg", + "table_caption": [ + "Table 4. Comparison of self-sparsified sparsity patterns." + ], + "table_footnote": [], + "table_body": "
Pruning RatioMethodMemory Peak UsageEpoch TimeAccuracy(%)
-FULL-FT19980MB2min59s93.0
90%BACKRAZOR (ours) PREBR-7540MB6min19s92.1
7220MB5min25s91.7
80%BACKRAZOR (ours) PREBR-8210MB6min28s92.9
8100MB5min49s92.9
", + "bbox": [ + 501, + 516, + 890, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Integration of the FLASHATTENTION kernel. We present Tab. 2 that details the performance when switching between the original FLASHATTENTION kernel (without memory savings), sparsified FLASHATTENTION (our modified implementation), and no FLASHATTENTION (using BackRazor's attention implementation). We use (PRE)BACKRAZOR@90% with parameter-wise sparsified JOINTSPAR@50% for these settings, note that the incorporation of JOINTSPAR is specifically aimed at offsetting the dense FLASHATTENTION kernel's memory usage. Additionally, we evaluate the performance of these methods on RTX 2080 Ti GPU with 11GB of VRAM, with (PRE)BACKRAZOR@90% and JOINTSPAR@30%. Note that both FULL-FT and FLASHATTENTION-2 are not available for testing on RTX 2080 Ti as they exceed the VRAM capacity and require Ampere architecture [15], respectively. Our initiative to integrate the FLASHATTENTION kernel into the training process aims to adapt a state-of-the-art", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5889", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "solution for enhancing training on consumer-level devices. This integration yields a significant speedup of $1.7 \\times$ during training, effectively mitigating the impact of additional computation introduced when compared to vanilla implementation as in FULL-FT. However, the replacement of BACKRAZOR's sparsified attention implementation does come with more memory costs up to 1GB. In response to this challenge, we have customized FLASHATTENTION by incorporating the PREBACKRAZOR method to prune the gradients and activations. This customization serves to further reduce activations and counterbalance the increased memory usage, ultimately falling below the memory footprint of BACKRAZOR.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. PREBACKRAZOR with BERT", + "text_level": 1, + "bbox": [ + 76, + 297, + 323, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Settings. For BERT training, we conduct fine-tuning on the bert-base-uncased model based on BACKRAZOR adaptations that are implemented within Huggingface's Transformers Repository [53]. We note that the model currently does not support FLASHATTENTION-2 to the best of our knowledge, from the involvement of an input attention mask in its attention algorithm which is not compatible with FLASHATTENTION kernel. We closely monitor and record the model's performance throughout 5 training epochs with batch size [8, 32] on a single GPU, which is a common case for fine-tuning on low-end devices. The selected fine-tuning task is Recognizing Textual Entailment (RTE), part of the GLUE benchmark [51]. The sparsity ratio for PREBACK-RAZOR and BACKRAZOR is set to $[80\\%, 90\\%]$ .", + "bbox": [ + 75, + 320, + 472, + 532 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0114f561c220836586f809a4bc4652d398153284239bb1bd64cc0cbe5296fb07.jpg", + "table_caption": [ + "Table 5. Performance of different fine-tuning schemes for RTE from GLUE Benchmark with BERT base." + ], + "table_footnote": [], + "table_body": "
Batch SizePruning RatioMethodMemory Peak UsageFinish TimeAccuracy(%)
32-FULL-FT16625MB2min50s70.40
90%BACKRAZOR (ours) PREBR11865MB6min26s66.43
11871MB6min24s65.70
80%BACKRAZOR (ours) PREBR12888MB6min41s65.70
13006MB6min19s66.07
8-FULL-FT5575MB4min28s70.04
90%BACKRAZOR (ours) PREBR4734MB16min59s66.43
4747MB16min05s69.68
80%BACKRAZOR (ours) PREBR4812MB17min39s69.31
4840MB16min36s72.92
", + "bbox": [ + 78, + 584, + 470, + 742 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ba4eafeeecedb43e55350069966295fd5527bca1798640b5d4e76eab4841b0b.jpg", + "image_caption": [ + "Figure 4. Training loss during fine-tuning BERT-base for RTE." + ], + "image_footnote": [], + "bbox": [ + 99, + 768, + 444, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance. We provide a comparison Tab. 5 in which we compare fine-tuning methods, including fine-tuning the full model, element-wise BACKRAZOR, and column-wise PREBACKRAZOR. Both pruning methods successfully reduce memory usage compared to the Full-FT approach. For instance, at the pruning ratio of $90\\%$ and batch size of 32, BACKRAZOR and PREBR show similar memory usage significantly lower than Full-FT, yielding a saved memory of $28.6\\%$ (4.8GB). The wall clock time for both pruning methods is higher compared to Full-FT. However, our proposed PREBACKRAZOR tends to have a slightly shorter time than BACKRAZOR in all cases, up to $9.7\\% \\times$ speedup. For $bs = 8$ which is a common case for fine-tuning large language models (LLM) on consumer-level devices, the introduction of element-wise self-sparsified activation pruning methods leads to an increased time cost of up to $5\\times$ which is not justified by the benefits obtained ( $12\\%$ memory reduction). PREBACKRAZOR demonstrates a slight improvement in convergence speed and reduction in loss compared to PREBACKRAZOR as shown in Fig. 4, highlighting its enhanced efficiency in fine-tuning, and outperforms not only BACKRAZOR but also Full-FT at an $80\\%$ pruning ratio, achieving the highest accuracy of $72.92\\%$ . Overall, while both pruning methods effectively reduce memory usage over Full-FT, PREBACKRAZOR tends to maintain or achieve slightly higher accuracy and speedup than BACKRAZOR, demonstrates a better balance between memory efficiency, completion time, and accuracy, particularly at lower batch sizes and higher pruning ratios.", + "bbox": [ + 496, + 90, + 893, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 542, + 617, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we reveal the inefficiency of the existing element-wise self-sparsified activation pruning method. Based on this finding, we develop a computation and memory-efficient activation pruning framework called PREBACKRAZOR for jointly pruning activations and gradients. We have demonstrated that integrating PREBACKRAZOR with efficient attention kernels can significantly enhance the accuracy of existing methods. In addition, we also theoretically prove that our method has the same convergence rate as BACKRAZOR. However, while PREBACKRAZOR has exhibited performance improvements in ViT and BERT fine-tuning, its applicability to larger foundation models warrants further investigation. Additionally, theoretical proofs and analyses for its effectiveness in such models require continued research efforts.", + "bbox": [ + 496, + 568, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 808, + 668, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK349000005.", + "bbox": [ + 496, + 832, + 893, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5890", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dan Alistarh, Demjan Grubic, Jerry Li, Ryota Tomioka, and Milan Vojnovic. Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in neural information processing systems, 30, 2017. 2", + "[2] Arash Ardakani, Carlo Condo, and Warren J Gross. Activation pruning of deep convolutional neural networks. In 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1325-1329. IEEE, 2017. 3", + "[3] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem. Machine learning, 47:235-256, 2002. 5", + "[4] Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Animashree Anandkumar. *signsgd: Compressed optimisation for non-convex problems*. In International Conference on Machine Learning, pages 560–569. PMLR, 2018. 2", + "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1", + "[6] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Freezeout: Accelerate training by progressively freezing layers. arXiv preprint arXiv:1706.04983, 2017. 3", + "[7] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce activations, not trainable parameters for efficient ondevice learning. arXiv preprint arXiv:2007.11622, 2020. 1, 2", + "[8] Ayan Chakrabarti and Benjamin Moseley. Backprop with approximate activations for memory-efficient network training. Advances in Neural Information Processing Systems, 32, 2019. 1, 3", + "[9] Congliang Chen, Li Shen, Haozhi Huang, and Wei Liu. Quantized adam with error feedback. ACM Transactions on Intelligent Systems and Technology (TIST), 12(5):1-26, 2021. 2", + "[10] Congliang Chen, Li Shen, Fangyu Zou, and Wei Liu. Towards practical adam: Non-convexity, convergence theory, and mini-batch acceleration. Journal of Machine Learning Research, 23(229):1-47, 2022. 3", + "[11] Jianfei Chen, Lianmin Zheng, Zhewei Yao, Dequan Wang, Ion Stoica, Michael Mahoney, and Joseph Gonzalez. Actnn: Reducing training memory footprint via 2-bit activation compressed training. In International Conference on Machine Learning, pages 1803-1813. PMLR, 2021. 3", + "[12] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2", + "[13] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022. 2", + "[14] Jack Choquette, Wishwesh Gandhi, Olivier Giroux, Nick Stam, and Ronny Krashinsky. Nvidia a100 tensor coregpu:" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Performance and innovation. IEEE Micro, 41(2):29-35, 2021. 4", + "[15] Jack Choquette, Edward Lee, Ronny Krashinsky, Vishnu Balan, and Brucek Khailany. 3.2 the a100 datacenter gpu and ampere architecture. In 2021 IEEE International Solid-State Circuits Conference (ISSCC), pages 48-50. IEEE, 2021. 7", + "[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 2", + "[17] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2", + "[18] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314, 2023. 2", + "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[21] Chao Fang, Wei Sun, Aojun Zhou, and Zhongfeng Wang. Efficient n: M sparse dnn training using algorithm, architecture, and dataflow co-design. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 3, 4", + "[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[23] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research, 18(1): 6869-6898, 2017. 1", + "[24] Ziyu Jiang, Xuxi Chen, Xueqin Huang, Xianzhi Du, Denny Zhou, and Zhangyang Wang. Back razor: Memory-efficient transfer learning by self-sparsified backpropagation. Advances in Neural Information Processing Systems, 35: 29248-29261, 2022. 1, 3, 6", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6", + "[27] Bingrui Li, Jianfei Chen, and Jun Zhu. Memory efficient optimizers with 4-bit states. arXiv preprint arXiv:2309.01507, 2023. 2", + "[28] Tianjian Li, Haoran Xu, Philipp Koehn, and Kenton Murray. Efficiently harnessing parameter importance for better training. 2023. 3", + "[29] Junjie Liu, Zhe Xu, Runbin Shi, Ray CC Cheung, and Hayden KH So. Dynamic sparse training: Find efficient sparse" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5891", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "network from scratch with trainable masked layers. arXiv preprint arXiv:2005.06870, 2020.3", + "[30] Rui Liu and Barzan Mozafari. Communication-efficient distributed learning for large batch optimization. In International Conference on Machine Learning, pages 13925-13946. PMLR, 2022. 2, 3, 4, 5, 6", + "[31] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1", + "[32] Kai Lv, Hang Yan, Qipeng Guo, Hajjun Lv, and Xipeng Qiu. Adalomo: Low-memory optimization with adaptive learning rate. arXiv preprint arXiv:2310.10195, 2023. 2", + "[33] Kai Lv, Yuqing Yang, Tengxiao Liu, Qinghui Gao, Qipeng Guo, and Xipeng Qiu. Full parameter fine-tuning for large language models with limited resources. arXiv preprint arXiv:2306.09782, 2023. 2", + "[34] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. arXiv preprint arXiv:2305.17333, 2023. 2", + "[35] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 2", + "[36] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In International Conference on Machine Learning, pages 4646-4655. PMLR, 2019. 3", + "[37] Zizheng Pan, Peng Chen, Haoyu He, Jing Liu, Jianfei Cai, and Bohan Zhuang. Mesa: A memory-saving training framework for transformers. arXiv preprint arXiv:2111.11124, 2021. 1, 3", + "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[39] Jeff Pool. Accelerating sparsity in the nvidia ampere architecture. GTC 2020, 2020. 4", + "[40] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3", + "[41] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1, 2", + "[42] Jie Ren, Samyam Rajbhandari, Reza Yazdani Aminabadi, Olatunj Ruwase, Shuangyan Yang, Minjia Zhang, Dong Li, and Yuxiong He. {ZeRO-Offload}: Democratizing {BillionScale} model training. In 2021 USENIX Annual Technical Conference (USENIX ATC 21), pages 551–564, 2021. 2", + "[43] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 2", + "[44] Li Shen, Yan Sun, Zhiyuan Yu, Liang Ding, Xinmei Tian, and Dacheng Tao. On efficient training of large-scale deep learning models: A literature review. arXiv preprint arXiv:2304.03589, 2023. 1", + "[45] Sebastian U Stich, Jean-Baptiste Cordonnier, and Martin Jaggi. Sparsified sgd with memory. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. Neural Networks, 169:506-519, 2024. 3", + "[47] Mingjie Sun, Zhuang Liu, Anna Bair, and J. Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 2", + "[48] Xu Sun, Xuancheng Ren, Shuming Ma, and Hofeng Wang. meprop: Sparsified back propagation for accelerated deep learning with reduced overfitting. In International Conference on Machine Learning, pages 3299-3308. PMLR, 2017. 3", + "[49] Tijmen Tieleman, Geoffrey Hinton, et al. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26-31, 2012. 2, 3", + "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1", + "[51] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 8", + "[52] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 3", + "[53] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 8", + "[54] Qing Yang, Wei Wen, Zuoguan Wang, and Hai Li. Joint regularization on activations and weights for efficient neural network pruning. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 790-797. IEEE, 2019. 3", + "[55] Yuxin Zhang, Yiting Luo, Mingbao Lin, Yunshan Zhong, Jingjing Xie, Fei Chao, and Rongrong Ji. Bi-directional masks for efficient n: M sparse training. arXiv preprint arXiv:2302.06058, 2023. 3", + "[56] Fangyu Zou, Li Shen, Zequn Jie, Weizhong Zhang, and Wei Liu. A sufficient condition for convergences of adam and rm-sprop. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11127-11135, 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 895 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5892", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_model.json b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d3dafe55deaf7eee0e78c176d64b1c2b92c9a523 --- /dev/null +++ b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.798, + 0.153 + ], + "angle": 0, + "content": "Sheared Backpropagation for Fine-tuning Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.193, + 0.89, + 0.285 + ], + "angle": 0, + "content": "Zhiyuan Yu\\(^{1}\\), Li Shen\\(^{*,2}\\), Liang Ding\\(^{2}\\), Xinmei Tian\\(^{1}\\), Yixin Chen\\(^{3}\\), and Dacheng Tao\\(^{4}\\) \n\\(^{1}\\)University of Science and Technology of China, Hefei, China; \\(^{2}\\)JD Explore Academy, Beijing, China; \\(^{3}\\)Washington University in St. Louis, MO, USA; \\(^{4}\\)Nanyang Technological University, Singapore \nyuzhiyuan@mail.ustc.edu.cn, xinmei@ustc.edu.cn, chen@cse.wustl.edu \n{mathshenli, liangding.liam, dacheng.tao}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.33, + 0.314, + 0.346 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.351, + 0.473, + 0.775 + ], + "angle": 0, + "content": "Fine-tuning is the process of extending the training of pre-trained models on specific target tasks, thereby significantly enhancing their performance across various applications. However, fine-tuning often demands large memory consumption, posing a challenge for low-memory devices that some previous memory-efficient fine-tuning methods attempted to mitigate by pruning activations for gradient computation, albeit at the cost of significant computational overhead from the pruning processes during training. To address these challenges, we introduce PreBackRazor, a novel activation pruning scheme offering both computational and memory efficiency through a sparsified backpropagation strategy, which judiciously avoids unnecessary activation pruning and storage and gradient computation. Before activation pruning, our approach samples a probability of selecting a portion of parameters to freeze, utilizing a bandit method for updates to prioritize impactful gradients on convergence. During the feed-forward pass, each model layer adjusts adaptively based on parameter activation status, obviating the need for sparsification and storage of redundant activations for subsequent backpropagation. Benchmarking on fine-tuning foundation models, our approach maintains baseline accuracy across diverse tasks, yielding over \\(20\\%\\) speedup and around \\(10\\%\\) memory reduction. Moreover, integrating with an advanced CUDA kernel achieves up to \\(60\\%\\) speedup without extra memory costs or accuracy loss, significantly enhancing the efficiency of fine-tuning foundation models on memory-constrained devices." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.78, + 0.21, + 0.795 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.805, + 0.47, + 0.882 + ], + "angle": 0, + "content": "Nowadays, Transformer-based models [50] have achieved remarkable results in various domains of deep learning [19, 20]. The trend towards scaling up these models for enhanced performance necessitates substantial data and computational resources, often beyond the reach of many" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.331, + 0.895, + 0.589 + ], + "angle": 0, + "content": "researchers [44]. Consequently, fine-tuning pre-trained Transformer models, which are a prominent example of foundation models [5], has become a prevalent strategy. This approach allows for adapting and continuing the training of these models at reduced costs while retaining high performance. Despite the advantages, fine-tuning foundation models on low-end devices, such as personal computers with consumer-level GPUs and other edge devices, presents considerable challenges. These challenges stem primarily from the limited memory and computational capacity of such devices, restricting the feasible model and batch sizes during fine-tuning. This limitation often results in prolonged training periods or compromised performance. Our paper aims to address this issue by exploring efficient methodologies for fine-tuning Transformer models on consumer-level devices, striving to balance performance with training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "One pervasive challenge in fine-tuning models is effectively managing the training pipeline within the confines of limited device memory [41]. Beyond loading trainable parameters, memory usage in model training includes storing intermediate tensors like gradients and activations. These elements are crucial in backpropagation and become particularly memory-intensive with large input batch sizes or model scales. Although efficient training techniques, such as low-precision and pruning, have been proposed to save memory on model parameters and are widely adopted [23, 31], recent research has shifted focus towards reducing memory consumption of activations, identified as a primary memory bottleneck [7]. Notably, since activations are stored solely for backward computation, recent studies on activation reduction [8, 24, 37] have proposed an asymmetric backpropagation pipeline. This approach maintains a dense forward pass and compresses activations only for gradient computation during backpropagation. Among these, BACKRAZOR by Jiang et al. [24] employs top-\\(k\\) pruning in the forward pass to sparsify activa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "tions, thereby reducing memory requirements for their storage. This method then re-densifies them during the backward pass for gradient computation, achieving significant memory savings. However, since activation pruning methods do not retain the original activations, they typically require additional computations to determine optimal pruning locations, in order to maintain accuracy. This introduces a trade-off between memory reduction and computational overhead, necessitating a careful balance. Furthermore, it presents a challenge in enhancing the efficiency of current computation-intensive activation pruning methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.641 + ], + "angle": 0, + "content": "In our work, addressing this unresolved challenge, we present PREBACKRAZOR, a memory and computation-efficient activation pruning scheme. It aligns with the asymmetric sparsified backpropagation approach, where sparsified activations are preserved solely for the backpropagation process. PREBACKRAZOR operates by dynamically freezing non-essential activation coordinates, thus minimizing the necessity for their sparsification and storage. This method effectively reduces computational load, thereby boosting the overall computation efficiency. Before initiating each iteration's forward pass, we strategically assign a dropout probability to every activation coordinate. This probability is informed by the gradient values associated with these coordinates, derived from prior iterations. This process enables us to selectively freeze certain coordinates during the ensuing backward pass. The chosen probabilities aim to preserve those gradients that are crucial for model convergence. To continuously refine this probability distribution, we utilize a bandit method, an approach inspired by JOINTSPAR [30]. Consequently, during the forward pass, we bypass the pruning and storage of activations that are deemed non-essential for backpropagation. This method effectively diminishes the computational burden typically associated with activation pruning, while simultaneously upholding memory efficiency and training accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our experiments, conducted on a single RTX 4090 GPU, have shown significant improvements in training efficiency and memory management during the fine-tuning of Vision Transformer (ViT) and BERT models. For ViT, an extensive comparative analysis using the CIFAR datasets [26] revealed that our method, PREBACKRAZOR, outperforms traditional activation pruning techniques like BACKRAZOR in terms of computational efficiency. Notably, PREBACKRAZOR achieves an up to \\(1.7\\times\\) increase in training speed for linear layers, without incurring additional memory overhead. Furthermore, our carefully tailored adaptation of the FLASHATTENTION kernel has significantly accelerated the fine-tuning process, ensuring efficient memory usage. This customized integration results in an impressive \\(1.7\\times\\) boost in training speed, maintaining a balanced computational budget without sacrificing accuracy. In the context of BERT fine-tuning, our approach also results in a modest improve" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.78, + 0.106 + ], + "angle": 0, + "content": "ment in time efficiency and final accuracy." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.107, + 0.854, + 0.121 + ], + "angle": 0, + "content": "In summary, our main contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.124, + 0.892, + 0.183 + ], + "angle": 0, + "content": "- We propose a novel activation pruning pipeline to further reduce the computational overhead of previous methods. In addition, we establish the convergence analysis for the proposed algorithm." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.184, + 0.892, + 0.304 + ], + "angle": 0, + "content": "- We design improved activation storage schemes that jointly reduce the compute and memory. We employ bandit methods to update gradients across activation coordinates of the model layers adaptively. This selective update mechanism efficiently manages activation pruning and gradient computation, thereby reducing memory and computational costs based on BACKRAZOR's asymmetric activation-self-sparsified backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.305, + 0.892, + 0.41 + ], + "angle": 0, + "content": "- We conduct experiments on fine-tuning Transformer-based ViT and BERT, adapt the current code library, and combine the state-of-the-art FLASHATTENTION 2 CUDA [16] kernels to boost the training, making the pipeline more welcome by nowadays consumer-level devices. Our experiments show that our method reduces over \\(30\\%\\) of training time without loss of accuracy." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.124, + 0.892, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.642, + 0.433 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.442, + 0.892, + 0.473 + ], + "angle": 0, + "content": "In this section, we briefly review existing memory and computation-efficient fine-tuning methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Memory-efficient fine-tuning methods. In memory-limited environments, strategies like partitioning and compression leverage trade-offs in computation, communication, and storage to reduce memory usage. Gradient accumulation lowers memory usage at the cost of increased communication, while gradient checkpointing [12] achieves memory savings through recomputation. CPU offloading [42], GPU partitioning [41], and FLASHATTENTION [16, 17] balance memory optimization against communication overhead. Several memory-efficient optimizers [32, 43, 49] have been proposed to approximate backpropagation to minimize stored information. Recent approaches include computing only forward passes with estimated gradients [34] and integrating gradient computation with parameter updates [33]. Additionally, low or mixed-precision training [27, 35] and other sparse backpropagation methods [13] have been integrated, trading some numerical accuracy for enhanced memory or communication efficiency. There are also related works focusing on communication cost reduction in distributed systems that use these memory-efficient methods for gradient compression [1, 4, 45], which is different from our settings. Additionally, while quantization [9, 18] and pruning [47] methods are commonly employed to compress gradients and model weights, our research concentrates on the pruning of activations. In the realm of activation footprint reduction, Cai et al. [7] suggest freezing layer weights while updating biases and adding lightweight residual modules. Other" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5884" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "methods involve compression of activations [2, 11, 54]. To maintain accuracy, recent research proposes asymmetric sparsified backpropagation which stores compressed activations, using original dense inputs for forward computation and retrieving activations for backward gradient computations. They employ techniques like low-precision compression [8, 37], or utilizing top-\\(k\\) masks for pruning activations [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.473, + 0.541 + ], + "angle": 0, + "content": "Computation-efficient fine-tuning methods. Existing computation-efficient fine-tuning methods include PEFT, knowledge distillation [22], few-shot learning [52], as well as quantization and pruning implementations. Our research primarily focuses on techniques that dynamically adjust training and seamlessly integrate with our sparse backpropagation pipeline. In the realm of computation-efficient fine-tuning methods, we find adaptive learning rates [25, 49] relevant for their role in facilitating faster convergence [10, 46, 56]. Also, dynamic sparse algorithms [29, 36] are notable for introducing sparsity during training. For the aspect of sparse backpropagation, the implementation of top-k masking over gradients by Sun et al. [48] to induce sparsity in backpropagation is in line with our approach to asymmetric pruning. Structured sparsity patterns [21, 55] and layer freezing techniques [6, 40], along with selective weight freezing strategies [28], offer valuable insights into reducing computational load. We also draw specific elements from the JOINTSPAR algorithm [30], which selectively drops gradient vectors, a principle we adapt to target the computational challenges in activation pruning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.462, + 0.573 + ], + "angle": 0, + "content": "3. Rethink the Backpropagation Redundancy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.655 + ], + "angle": 0, + "content": "In this section, we first present explanations of the memory cost of activations, and the asymmetric self-sparsified activations as introduced in BACKRAZOR algorithm [24]. Then we show the memory and computation redundancy in the current backpropagation scheme." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.208, + 0.682 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Consider the nonconvex optimization problem of fin-tuning foundation models: \\(\\min_{\\theta \\in \\Theta} \\mathcal{L}(\\theta)\\), where \\(\\mathcal{L}\\) is the training loss calculated with the model's weight \\(\\theta\\), and \\(\\Theta\\) is the parameter space. Without loss of generality, the forward pass of a given neural network with \\(n\\) layers can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.793, + 0.468, + 0.809 + ], + "angle": 0, + "content": "\\[\nf \\left(\\mathbf {x} _ {0}; \\boldsymbol {\\theta}\\right) = f _ {n} \\left(f _ {n - 1} \\left(\\dots f _ {0} \\left(\\mathbf {x} _ {0}; \\boldsymbol {\\theta} _ {0}\\right) \\dots ; \\boldsymbol {\\theta} _ {n - 1}; \\boldsymbol {\\theta} _ {n}\\right), \\right. \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.813, + 0.468, + 0.828 + ], + "angle": 0, + "content": "\\[\nL = \\mathcal {L} (f (\\mathbf {x} _ {0}; \\boldsymbol {\\theta}), y) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_0\\) represents the input features at the 0-th layer, and \\(y\\) corresponds to the input label. The functions \\(f_{i}\\) and parameters \\(\\theta_{i}\\) are associated with the model's \\(i\\)-th layer. The expression \\(f(\\mathbf{x}_0; \\boldsymbol{\\theta})\\) indicates the model's final prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "In this section, the symbol \\(L\\) stands for the loss between the prediction and the label, computed using the specified loss function \\(\\mathcal{L}\\). We define the activation \\(\\mathbf{x}_i\\) as the output of the \\((i - 1)\\)-th layer, i.e., \\(\\mathbf{x}_i = f_{i - 1}(\\mathbf{x}_{i - 1};\\pmb{\\theta}_{i - 1})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.2 + ], + "angle": 0, + "content": "Denoting \\(h_i^{(\\mathbf{x})}(\\mathbf{x}_i,\\pmb {\\theta}_i)\\) and \\(h_i^{(\\pmb \\theta)}(\\mathbf{x}_i,\\pmb {\\theta}_i)\\) as the derivatives of \\(\\mathbf{x}_{i + 1}\\) with respect to \\(\\mathbf{x}_i\\) and \\(\\pmb{\\theta}_i\\) respectively. Normally, the backward process can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.211, + 0.891, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\mathbf {x})} \\left(\\mathbf {x} _ {i}, \\boldsymbol {\\theta} _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.244, + 0.891, + 0.276 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\boldsymbol {\\theta})} (\\mathbf {x} _ {i}, \\boldsymbol {\\theta} _ {i}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.407 + ], + "angle": 0, + "content": "For instance, considering a linear layer with parameters \\(\\theta_{i}\\), which consist of a weight matrix \\(\\mathbf{W}_i\\) and a bias vector \\(\\mathbf{b}_i\\), its forward pass function for computing the output \\(\\mathbf{x}_{i + 1}\\) is expressed as \\(\\mathbf{x}_{i + 1} = \\mathbf{x}_i\\mathbf{W}_i^\\top +\\mathbf{b}_i\\). All of the variables \\(\\mathbf{x}_i\\), \\(\\mathbf{W}_i\\), and \\(\\mathbf{b}_i\\) are retained for gradient computation before the layer produces the output \\(\\mathbf{x}_{i + 1}\\). Subsequently, during the backward pass of this layer in the loss backpropagation process, the gradients are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.418, + 0.891, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\mathbf {W} _ {i}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.453, + 0.891, + 0.483 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial \\mathbf {W} _ {i}} = \\left(\\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}}\\right) ^ {\\top} \\mathbf {x} _ {i}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.486, + 0.891, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial \\mathbf {b} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Here, the activation \\(\\mathbf{x_i}\\) is involved in computing the gradient for updating the weight \\(\\mathbf{W}_i\\). Consequently, the activation value must be stored in memory during the forward pass. This storage occupies a significant portion of memory, often exceeding the memory required for saving the weights. Additionally, the size of the activations increases linearly with the batch size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Jiang et al. [24] propose BACKRAZOR, a method that only prunes the activation saved for backpropagation and keeps it dense in the forward pass. The activation \\(\\mathbf{x}_i\\) is pruned and stored in the context as a sparse version \\(\\tilde{\\mathbf{x}}_i\\). The backward process with pruned activations turns into:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.72, + 0.891, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\mathbf {x})} \\left(\\widetilde {\\mathbf {x}} _ {i}, \\boldsymbol {\\theta} _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.754, + 0.891, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\boldsymbol {\\theta})} (\\widetilde {\\mathbf {x}} _ {i}, \\boldsymbol {\\theta} _ {i}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As for their activation pruning, a straightforward top- \\(k\\) pruning method is applied, which identifies and removes the \\(k\\) smallest magnitude values within the activation. Denoting \\(v\\) as the total number of values of the activation, to introduce a targeted sparsity level, a fixed parameter \\(\\lambda\\) is set (e.g., \\(\\lambda = 0.9\\)), and the pruning budget is determined by setting \\(k = \\lambda v\\). Consequently, all activation values below the \\(k\\)-th" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5885" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.091, + 0.426, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.3 + ], + "angle": 0, + "content": "Figure 1. Sparsity vs computation time cost between BackRazor (element-wise, self-sparsified) and proposed PreBackRazor (column-wise, joint-sparsified)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.468, + 0.38 + ], + "angle": 0, + "content": "smallest value are zeroed out. After this pruning operation, the pruned activation is stored as a binary bitmap, known as the mask. This mask indicates the positions of the nonzero elements. Additionally, a dense matrix is maintained to retain the values of the non-zero elements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.392, + 0.383, + 0.407 + ], + "angle": 0, + "content": "3.2. Rethink the sparsity in BACKRAZOR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.468, + 0.56 + ], + "angle": 0, + "content": "Redundancy in full-gradient computation. Common backpropagation activation sparsification schemes involve restoring compressed activations, saved within the model context, back to their dense format during backpropagation. In the case of BACKRAZOR, this process entails densifying the sparse matrix to its original shape for gradient computation. Consequently, the full gradient is computed using the full-size activation, and the sparsity of the pruned activation is not utilized to expedite computation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Computation overhead in pruning. A prevalent pruning approach involves generating an element-wise self-sparsified top-\\(k\\) mask to eliminate smaller values. This method, when employed for activation pruning, incurs additional computational costs. These costs include sorting to find the \\(k\\)-th smallest activation value and performing element-wise comparisons across all activation elements to construct a sparse tensor. Collectively, these operations contribute to significant computational overhead, potentially extending training time beyond that of traditional fine-tuning. Such a trade-off for memory reduction may not be justifiable. Exploring alternative sparsity patterns and pruning schemes, such as column-wise and joint-sparse could lead to the discovery of faster algorithms, potentially replacing the current self-sparsified top-\\(k\\) activation pruning method and thus reducing computation costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Inspired by the observed inefficiencies in existing methods, we sought to investigate whether alternative sparsity patterns could offer reduced computation. Specifically, we compared the performance of column-wise sparsity with element-wise sparsity within the same linear layer, across various sparsity ratios, as depicted in Figure 1. Our find" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.087, + 0.895, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.316, + 0.843, + 0.33 + ], + "angle": 0, + "content": "Figure 2. Pipeline of the proposed PreBackRazor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.51 + ], + "angle": 0, + "content": "ings indicate notable up to \\(1.7 \\times\\) speedups, aligning with our initial hypothesis. Additionally, the column-wise activation sparsity pattern facilitates the use of existing joint sparse algorithms, which leverage gradient information to generate sparse masks previous to the BACKRAZOR pruning pipeline. This approach incurs lower computational costs. Given that the original sparsity is unstructured and of a low ratio, the integration of a structured sparsity pattern also holds promise for further acceleration, leveraging modern sparse computation software and hardware advancements [14, 21, 39]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.634, + 0.541 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.891, + 0.626 + ], + "angle": 0, + "content": "In this section, we first propose the PREBACKRAZOR method. After that, we introduce the bandit method from JOINTSPAR algorithm [30] that we use to adjust the probability of dropping computations. Lastly, we examine the convergence of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.803, + 0.652 + ], + "angle": 0, + "content": "4.1. Overall pipeline of PREBACKRAZOR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our jointly sparsified activation pruning approach, PREBACKRAZOR, employs asymmetric backpropagation, maintaining dense computations in the forward pass while pruning activations for the backward pass, as depicted in Fig. 2. Inspired by the JOINTSPAR algorithm [30], our method goes a step further by incorporating gradient sparsity into the activation pruning, aiming to significantly boost computational efficiency. In PREBACKRAZOR, we manage a probability distribution over activation coordinates to selectively prune redundant indices. This process effectively reduces the computation of related gradients and halts the updating of corresponding weights, thereby enhancing training efficiency and minimizing computational requirements. The decision-making process hinges on the analysis of gradient norms, leveraging the joint sparsity of gradients and activations. This approach" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5886" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.138 + ], + "angle": 0, + "content": "prioritizes the retention of the most impactful gradient coordinates for weight updates. The implementation details of this method are outlined in Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.081, + 0.155, + 0.29, + 0.169 + ], + "angle": 0, + "content": "Algorithm 1 PREBACKRAZOR" + }, + { + "type": "algorithm", + "bbox": [ + 0.081, + 0.173, + 0.47, + 0.542 + ], + "angle": 0, + "content": "1: Initialize all \\(p_0 = s / D\\) \n2: for \\(t = 1\\) to \\(T\\) do \n3: Sample \\(\\mathbf{x}_0, \\mathbf{y}\\) from \\(\\mathcal{D}\\) \n4: for \\(i = 1\\) to \\(n\\) do \n5: Instantiate \\(Z_{t,d}\\), with \\(P(Z_{t,d} = 1) = p_{t,d}\\) \n6: Determine \\(S_t = \\{d : Z_{t,d} = 1\\}\\) \n7: \\(\\mathbf{x}_i = f_i(\\mathbf{x}_{i-1}, \\boldsymbol{\\theta}_i)\\) \n8: mask = a binary bitmap of size \\(D\\) \n9: for parameter \\(\\boldsymbol{\\theta}_{[d]}\\) of \\(\\boldsymbol{\\theta}_i\\) do \n10: if \\(d \\in S_t\\) then \n11: \\(\\text{mask}_d = \\text{True}\\) \n12: end if \n13: end for \n14: \\(\\tilde{\\mathbf{x}}_{i-1} = \\text{Sparsify}(\\mathbf{x}_{i-1}, \\text{mask})\\) \n15: SaveForBackward(\\(\\tilde{\\mathbf{x}}_{i-1}, \\text{mask}, \\boldsymbol{\\theta}_i\\)) \n16: end for \n17: Compute loss \\(L = \\mathcal{L}(\\mathbf{x}_{n+1}, \\mathbf{y})\\) \n18: for \\(i = n-1\\) to 1 do \n19: \\(\\text{Densify}(\\tilde{\\mathbf{x}}_{i-1}, \\text{mask})\\) \n20: \\(\\frac{\\partial L}{\\partial \\mathbf{x}_i} = \\frac{\\partial L}{\\partial \\mathbf{x}_{i+1}} h_i^{(x)}(\\tilde{\\mathbf{x}}_i, \\boldsymbol{\\theta}_i)\\) \n21: for parameter \\(\\boldsymbol{\\theta}_{[d]}\\) with \\(d \\in S_t\\) do \n22: \\(\\frac{\\partial L}{\\partial \\boldsymbol{\\theta}_{[d]}} = \\frac{\\partial L}{\\partial \\mathbf{x}_{i+1}} h_i^{(\\theta)}(\\tilde{\\mathbf{x}}_i, \\boldsymbol{\\theta}_{[d]})\\) \n23: end for \n24: \\(p_{t+1} = \\text{Update}(p_t, S_t, \\{\\mathbf{g}_{t,[d]}\\}_{d \\in S_t})\\) \n25: end for \n26: end for" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.764 + ], + "angle": 0, + "content": "For a given input activation \\(\\mathbf{x}_i\\) in the \\(i\\)-th layer of the model, we decompose the activations column-wise, represented as \\(\\mathbf{x}_i = [\\mathbf{x}_{i,1},\\dots ,\\mathbf{x}_{i,d},\\dots ,\\mathbf{x}_{i,D}]\\), where \\(d\\) and \\(D\\) respectively denote the \\(d\\)-th coordinate out of all \\(D\\) column indices. Considering a linear layer, its weight parameter \\(\\pmb{\\theta}\\) is similarly decomposed into \\([\\pmb{\\theta}_{[1]},\\dots ,\\pmb{\\theta}_{[d]},\\dots ,\\pmb{\\theta}_{[D]}\\). To differentiate from the layer-wise parameter set \\(\\pmb{\\theta}_i\\) (as discussed in Section 3.1), we use square brackets to denote each coordinate, represented as \\(\\pmb{\\theta}_{[d]}\\). The gradient of the loss with respect to \\(\\pmb{\\theta}_{[d]}\\) is denoted by \\(\\pmb{g}_{t,d}\\), and we define \\(\\pmb{g}_{t,[d]}\\) as a vector with all elements zero except the \\(d\\)-th element: \\(\\pmb{g}_{t,[d]} = [0,\\dots ,\\pmb{g}_{t,d},\\dots ,0]\\), with \\(\\pmb{g}_t\\) being the sum of all such vectors, \\(\\pmb{g}_t = \\sum_{d = 1}^D\\pmb{g}_{t,[d]}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.884 + ], + "angle": 0, + "content": "At each iteration \\(t\\), a probability distribution \\(p_t = [p_{t,1}, \\dots, p_{t,d}, \\dots, p_{t,D}]\\) is maintained over the joint coordinates of the parameters \\([\\pmb{\\theta}_{[1]}, \\dots, \\pmb{\\theta}_{[d]}, \\dots, \\pmb{\\theta}_{[D]}\\]\\) and activations \\([\\mathbf{x}_1, \\dots, \\mathbf{x}_d, \\dots, \\mathbf{x}_D]\\). For each coordinate, a Bernoulli random variable \\(Z_{t,d}\\) is sampled with \\(P(Z_{t,d} = 1) = p_{t,d}\\) and \\(P(Z_{t,d} = 0) = 1 - p_{t,d}\\), deciding whether to include \\((Z_{t,d} = 1)\\) or exclude \\((Z_{t,d} = 0)\\) the gradient of \\(\\pmb{\\theta}_d\\) during backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.904 + ], + "angle": 0, + "content": "Initially, each \\( p_t \\) is set uniformly to \\( \\frac{s}{D} \\), aligning with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "desired sparsity budget \\( s \\), where \\( s \\) signifies the number of columns involved in the backward process. At the onset of each iteration \\( t \\), the Bernoulli random variable \\( Z_{t,d} \\) is sampled to form the active set \\( S_{t} \\) of coordinates that retain gradients in activations. Post gradient computation in backpropagation, this gradient information informs the update of the probability distribution, employing a bandit method which we will introduce in the subsequent section." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.893, + 0.508 + ], + "angle": 0, + "content": "Integrating the FLASHATTENTION kernel. The fine-tuning of Transformer-based models on consumer-grade devices presents a unique opportunity to leverage modern, efficient kernels for improved training efficiency. In our approach, we have adapted and integrated the FLASHATTENTION kernel, substituting the original, more segmented sparsified attention implementation in BACKRAZOR. To address the increased memory usage resultant from this integration, we have incorporated PREBACKRAZOR within the FLASHATTENTION function. This inclusion not only facilitates pruning but also employs recomputation strategies to mitigate precision loss. We designate this enhanced methodology as PREBACKRAZOR++. This innovative approach ensures memory usage remains on par with BACKRAZOR, while markedly accelerating the training process. We further combine JOINTSPAR over model parameters with our method as the PREBACKRAZOR++.JOINTSPAR for best memory and computation efficiency. Details on implementation are omitted here for brevity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.855, + 0.535 + ], + "angle": 0, + "content": "4.2. Bandit method for gradient sparsification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.602 + ], + "angle": 0, + "content": "To update the probability distribution \\( p_t \\) over \\( D \\) coordinates at each iteration \\( t \\), we utilize a bandit method akin to that in JOINTSPAR [30], originally introduced by Auer et al. [3], as detailed in Algorithm 2." + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.616, + 0.856, + 0.632 + ], + "angle": 0, + "content": "Algorithm 2 Bandit [30] for Updating Distribution \\( p_t \\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.508, + 0.634, + 0.892, + 0.822 + ], + "angle": 0, + "content": "1: Algorithm: UPDATE \\((p_t, S_t, \\{g_{t,[d]}\\}_{d \\in S_t})\\) \n2: Input: \\(p_t\\): probability distribution, \\(S_t\\): set of active parameters, \\(\\{g_{t,[d]}\\}_{d \\in S_t}\\): gradients \n3: for \\(d = 1\\) to \\(D\\) do \n4: if \\(d \\in S_t\\) then \n5: \\(\\tilde{l}_{t,d} = -\\frac{\\|g_{t,[d]}\\|^2}{(p_{t,d})^2} + \\frac{G^2}{p_{\\min}^2}\\) \n6: else \n7: \\(\\tilde{l}_{t,d} = 0\\) \n8: end if \n9: \\(w_{t,d} = p_{t,d} \\exp(-\\alpha_p \\tilde{l}_{t,d} / p_{t,d})\\) \n10: end for \n11: \\(p_{t+1} = \\arg \\min_{q \\in \\mathcal{P}} D_{\\mathrm{KL}}(q \\| w_t)\\) \n12: return \\(p_{t+1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce a set \\(\\mathcal{P}\\) as \\(\\mathcal{P} = \\{p\\in \\mathbb{R}^D:\\sum_{d = 1}^D p_d = s,p_d\\geq p_{\\mathrm{min}},\\forall 1\\leq d\\leq D\\}\\), where \\(p_{\\mathrm{min}}\\) is a constant lower bound. This set helps in guiding the update of the probability distribution. The sum of the probabilities across all" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5887" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.218 + ], + "angle": 0, + "content": "parameters equals the desired sparsity budget \\( s \\), with each probability bounded below by a constant \\( p_{\\mathrm{min}} \\). Note that \\( s \\) and \\( p_{\\mathrm{min}} \\) are parameters defined by the user, adhering to the conditions \\( 0 < s \\leq D \\) and \\( 0 < p_{\\mathrm{min}} \\leq 1 \\). Recall that \\( g_{t,[d]} \\) is defined as \\( g_{t,[d]} = [0,\\dots ,g_{t,d},\\dots ,0] \\), and \\( g_{t} = \\sum_{d = 1}^{D}g_{t,[d]} \\). The term \\( \\alpha_{p} \\) denotes the step size, while \\( D_{\\mathrm{KL}}(q||w) \\) represents the KL divergence between distributions \\( q \\) and \\( w \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.223, + 0.258, + 0.238 + ], + "angle": 0, + "content": "4.3. Theoretical results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.47, + 0.367 + ], + "angle": 0, + "content": "In this section, we analyze the convergence of our proposed method PREBACKRAZOR++JOINTSPAR in the context of linear networks. We use \\( g_{t} \\) to represent the gradient of the loss function \\( \\mathcal{L} \\) with respect to the parameter \\( \\theta_{t} \\) at iteration \\( t \\), calculated using all samples in the training dataset. In contrast, \\( \\tilde{g}_{t} \\) denotes the stochastic gradient computed using a single training sample. To support our analysis, we adopt the classical assumptions as detailed by Jiang et al. [24]:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.472, + 0.489 + ], + "angle": 0, + "content": "Assumption 1. (1) The objective function \\(\\mathcal{L}\\) is \\(S_{d}\\)-smooth with respect to \\(\\pmb{\\theta}_{[d]}\\); (2) The gradient \\(\\pmb{g}_{t,[d]}\\) is upper bounded by \\(G\\), i.e., \\(G \\geq \\| \\pmb{g}_{t,[d]} \\|\\). \\(\\forall t, d\\); (3) The objective function \\(\\mathcal{L}\\) is bounded below by \\(L^{*}\\); (4) For each iteration \\(t\\), the stochastic gradient \\(\\tilde{\\pmb{g}}_t\\) adheres to the following: \\(\\mathbb{E}[\\tilde{\\pmb{g}}_t] = \\pmb{g}_t\\), \\(\\mathbb{E}[(\\tilde{\\pmb{g}}_t - \\pmb{g}_t)_d^2] \\leq \\sigma_d^2\\), \\(\\forall d \\in \\{1, 2, \\dots, D\\}\\), with \\(D\\) denoting the number of coordinates and the constant \\(\\sigma_d^2\\) representing the variance corresponding to coordinate \\(d\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Lemma 1 [24]. In the context of linear neural networks, the gradient of parameters at iteration \\(t\\) can be decomposed as \\(\\tilde{g}_t = \\tilde{\\mathbf{x}}_t\\tilde{\\mathbf{A}}_t^\\top\\), where \\(\\tilde{\\mathbf{x}}_t\\) denotes the precise activation, and \\(\\tilde{\\mathbf{A}}_t\\) is referred to as the transformation matrix. After pruning the activations, only the pruned activations \\(\\tilde{\\mathbf{x}}_t'\\) are reserved and used for backpropagation, and the gradient approximation becomes \\(\\tilde{g}_t' = \\tilde{\\mathbf{x}}_t'\\tilde{\\mathbf{A}}_t^\\top\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.636 + ], + "angle": 0, + "content": "Theorem 1 (Convergence). Denote \\(\\alpha\\) as the step size. Under Assumptions 1-4, the following bound holds:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.643, + 0.469, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} \\left[ \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\| \\boldsymbol {g} _ {t, i} \\| ^ {2} \\right] \\leq \\frac {\\mathcal {L} \\left(\\theta_ {0}\\right) - L ^ {*}}{\\alpha T} \\\\ + \\frac {S _ {\\infty} \\alpha}{2 T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\mathbb {E} \\left[ \\sum_ {d = 1} ^ {D} \\frac {\\| \\boldsymbol {g} _ {t , [ d ]} \\| ^ {2}}{p _ {t , d}} \\right], \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.765 + ], + "angle": 0, + "content": "where \\(T\\) is the total number of iterations, \\(S_{\\infty} = \\max_d S_d\\), \\(\\theta_0\\) is the initial model weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As the training progresses, the first term on the right-hand side of the equation progressively diminishes to zero, whereas the second term remains confined within the bounds set by the bandit method. This theorem establishes that our proposed method converges towards a bounded neighborhood around a stationary point. While our theoretical analysis here is confined to linear networks, the extension to non-linear networks remains an area for future investigation. In the meantime, we have empirically validated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "that our proposed method effectively accelerates the finetuning of Transformer-based models. For those interested in a more thorough exploration of the underlying mathematics, the detailed derivation of these results is provided in the appendix. Additionally, the proofs and analyses in [24] and [30] offer further insight into these concepts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.195, + 0.634, + 0.212 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.34 + ], + "angle": 0, + "content": "We perform experiments on PyTorch [38] with RTX 4090 GPU with 24GB of VRAM. We apply our method on top of BACKRAZOR's code library to report the performance gains, following its training settings on ViT and BERT. We report the actual memory allocation on GPU measured with PyTorch CUDA tools. For better readability, we employ the abbreviation JS to denote JOINTSPAR, and PREBR as a shorthand for PREBACKRAZOR." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.35, + 0.85, + 0.365 + ], + "angle": 0, + "content": "5.1. PREBACKRAZOR with Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.525 + ], + "angle": 0, + "content": "Settings. For ViT training, we initiate the process by loading the checkpoint of the ViT-B/16 model that has been pre-trained on ImageNet-22K as our starting point. Subsequently, we fine-tune the model using the CIFAR-10 and CIFAR-100 datasets [26]. The fine-tuning process encompasses 20,000 steps and employs cosine learning rate decay, with an initial 500 steps dedicated to learning rate warmup. We utilize the standard SGD optimizer and conduct validation every 1,000 steps. The maximum learning rate is set at 0.01, and the batch size during training is 128." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.538, + 0.692, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.649, + 0.634, + 0.661 + ], + "angle": 0, + "content": "(a) Training." + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.539, + 0.885, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.649, + 0.832, + 0.661 + ], + "angle": 0, + "content": "(b) Validation." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.673, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Figure 3. Fine-tuning PREBACKRAZOR++@90% vs. BACKRAZOR@90% on CIFAR-100. (a) Training loss with wall-clock time. (b) Validation accuracy with steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Convergence comparison. The results of fine-tuning PREBACKRAZOR++ with parameter-wise JOINTSPAR in comparison to BACKRAZOR at a pruning rate of \\(90\\%\\) on CIFAR-100 are illustrated in Fig. 3. In Fig. 3a, the convergence speed of PREBACKRAZOR@90%++JOINTSPAR is comparable to that of BACKRAZOR@90%, with a slight advantage in the early stages of training. Notably, we observe a marginal improvement in accuracy, specifically \\([0.33\\%, 0.05\\%]\\), when adjusting the initial probabilities of JOINTSPAR to \\([70\\%, 50\\%]\\), respectively. This improvement is in contrast to the performance of BACKRAZOR@90%, as illustrated in Fig. 3b." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5888" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.328, + 0.09, + 0.644, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of fine-tuning methods for ViT." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.105, + 0.855, + 0.206 + ], + "angle": 0, + "content": "
MethodMemoryEpoch TimeSpeedupCIFAR-100CIFAR-10CUB-200
AllocatedReduction
FULL-FT19980MB1.0×2min59s-93.099.085.5
BACKRAZOR7540MB2.6×6min19s1.0×92.198.686.6
(ours) PREBR++JS@70%7440MB2.7×4min37s1.4×92.498.885.0
(ours) PREBR++JS@50%6880MB2.9×4min02s1.6×92.198.685.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.211, + 0.213, + 0.759, + 0.227 + ], + "angle": 0, + "content": "Table 2. Performance of different attention implementations fine-tuning ViT on CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.229, + 0.855, + 0.387 + ], + "angle": 0, + "content": "
DeviceMethodAttention TypeMemoryEpoch TimeSpeedupAccuracy(%)
RTX 4090FULL-FTvanilla19980MB2min59s-93.0
BACKRAZORBACKRAZOR7540MB6min19s1.0×92.1
(ours) PREBR++JSBACKRAZOR6570MB4min37s1.4×91.4
(ours) PREBR++JSFLASHATTENTION7960MB3min40s1.7×92.2
(ours) PREBR++JSsparse FLASHATTN6880MB4min02s1.6×92.1
RTX 2080 TiFULL-FTvanillaout of memory
BACKRAZORBACKRAZOR7780MB19min47s1.0×-
(ours) PREBR+JSBACKRAZOR6870MB15min13s1.3×-
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.398, + 0.473, + 0.897 + ], + "angle": 0, + "content": "Speedup and memory reduction. In Tab. 1, we present a comprehensive comparison among PREBACKRAZOR@90%++JOINTSPAR@[70%, 50%], BACKRAZOR@90%, and FT-Full (fine-tuning the full network) from the perspective of memory usage, speedup relative to BACKRAZOR@90%, and final accuracy on CIFAR-100, CIFAR-10 and CUB-200 for each method. Here we present the average memory allocated by CUDA during training. Compared with FT-full, BACKRAZOR@90% achieves a noteworthy \\(2.6\\times\\) reduction in memory usage. However, it demands twice as much training time due to the additional computations required for processing activations. In contrast, PREBR++JS achieves superior computational efficiency by avoiding unnecessary computations through joint sparsification of gradients and activations. This leads to significantly \\(1.6\\times\\) faster training without incurring additional memory costs or compromising accuracy. Additionally, Tab. 3 showcases the performance gains of forward and backward computation achieved by transitioning from a self-sparsified top-\\(k\\) pruning pattern to a column-wise joint-sparse pattern as in Fig. 1. These measurements are conducted on a linear layer characterized by input tensors with the shape [128, 197, 768] and of type float32. Tab. 4 illustrates the performance improvements achieved by substituting the element-wise mask in BACKRAZOR with a column-wise mask. This approach, referred to as PREBACKRAZOR-, employs a self-sparsified scheme that eschews the use of the bandit method for predicting pruning masks based on gradient information. Despite this, PREBACKRAZOR- still delivers commendable results, including a modest reduction in memory usage and over a \\(10\\%\\) increase in speed, while preserving nearly identical accuracy, particularly at a moderate pruning ratio of \\(80\\%\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.397, + 0.813, + 0.41 + ], + "angle": 0, + "content": "Table 3. Acceleration on a linear layer." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.412, + 0.892, + 0.496 + ], + "angle": 0, + "content": "
SparsityFWD time (ms)SpeedupFWD + BWD (ms)Speedup
BACKRAZORPREBRBACKRAZORPREBR
0.14.233.141.34×9.027.721.17×
0.24.354.241.03×9.209.031.02×
0.34.454.351.02×9.349.181.02×
0.47.974.451.79×16.529.321.77×
0.58.947.971.12×18.4516.511.12×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.501, + 0.862, + 0.515 + ], + "angle": 0, + "content": "Table 4. Comparison of self-sparsified sparsity patterns." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.517, + 0.892, + 0.619 + ], + "angle": 0, + "content": "
Pruning RatioMethodMemory Peak UsageEpoch TimeAccuracy(%)
-FULL-FT19980MB2min59s93.0
90%BACKRAZOR (ours) PREBR-7540MB6min19s92.1
7220MB5min25s91.7
80%BACKRAZOR (ours) PREBR-8210MB6min28s92.9
8100MB5min49s92.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Integration of the FLASHATTENTION kernel. We present Tab. 2 that details the performance when switching between the original FLASHATTENTION kernel (without memory savings), sparsified FLASHATTENTION (our modified implementation), and no FLASHATTENTION (using BackRazor's attention implementation). We use (PRE)BACKRAZOR@90% with parameter-wise sparsified JOINTSPAR@50% for these settings, note that the incorporation of JOINTSPAR is specifically aimed at offsetting the dense FLASHATTENTION kernel's memory usage. Additionally, we evaluate the performance of these methods on RTX 2080 Ti GPU with 11GB of VRAM, with (PRE)BACKRAZOR@90% and JOINTSPAR@30%. Note that both FULL-FT and FLASHATTENTION-2 are not available for testing on RTX 2080 Ti as they exceed the VRAM capacity and require Ampere architecture [15], respectively. Our initiative to integrate the FLASHATTENTION kernel into the training process aims to adapt a state-of-the-art" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5889" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "solution for enhancing training on consumer-level devices. This integration yields a significant speedup of \\(1.7 \\times\\) during training, effectively mitigating the impact of additional computation introduced when compared to vanilla implementation as in FULL-FT. However, the replacement of BACKRAZOR's sparsified attention implementation does come with more memory costs up to 1GB. In response to this challenge, we have customized FLASHATTENTION by incorporating the PREBACKRAZOR method to prune the gradients and activations. This customization serves to further reduce activations and counterbalance the increased memory usage, ultimately falling below the memory footprint of BACKRAZOR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.298, + 0.325, + 0.312 + ], + "angle": 0, + "content": "5.2. PREBACKRAZOR with BERT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.473, + 0.534 + ], + "angle": 0, + "content": "Settings. For BERT training, we conduct fine-tuning on the bert-base-uncased model based on BACKRAZOR adaptations that are implemented within Huggingface's Transformers Repository [53]. We note that the model currently does not support FLASHATTENTION-2 to the best of our knowledge, from the involvement of an input attention mask in its attention algorithm which is not compatible with FLASHATTENTION kernel. We closely monitor and record the model's performance throughout 5 training epochs with batch size [8, 32] on a single GPU, which is a common case for fine-tuning on low-end devices. The selected fine-tuning task is Recognizing Textual Entailment (RTE), part of the GLUE benchmark [51]. The sparsity ratio for PREBACK-RAZOR and BACKRAZOR is set to \\([80\\%, 90\\%]\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.545, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Table 5. Performance of different fine-tuning schemes for RTE from GLUE Benchmark with BERT base." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.585, + 0.472, + 0.743 + ], + "angle": 0, + "content": "
Batch SizePruning RatioMethodMemory Peak UsageFinish TimeAccuracy(%)
32-FULL-FT16625MB2min50s70.40
90%BACKRAZOR (ours) PREBR11865MB6min26s66.43
11871MB6min24s65.70
80%BACKRAZOR (ours) PREBR12888MB6min41s65.70
13006MB6min19s66.07
8-FULL-FT5575MB4min28s70.04
90%BACKRAZOR (ours) PREBR4734MB16min59s66.43
4747MB16min05s69.68
80%BACKRAZOR (ours) PREBR4812MB17min39s69.31
4840MB16min36s72.92
" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.77, + 0.445, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.9, + 0.46, + 0.915 + ], + "angle": 0, + "content": "Figure 4. Training loss during fine-tuning BERT-base for RTE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.532 + ], + "angle": 0, + "content": "Performance. We provide a comparison Tab. 5 in which we compare fine-tuning methods, including fine-tuning the full model, element-wise BACKRAZOR, and column-wise PREBACKRAZOR. Both pruning methods successfully reduce memory usage compared to the Full-FT approach. For instance, at the pruning ratio of \\(90\\%\\) and batch size of 32, BACKRAZOR and PREBR show similar memory usage significantly lower than Full-FT, yielding a saved memory of \\(28.6\\%\\) (4.8GB). The wall clock time for both pruning methods is higher compared to Full-FT. However, our proposed PREBACKRAZOR tends to have a slightly shorter time than BACKRAZOR in all cases, up to \\(9.7\\% \\times\\) speedup. For \\(bs = 8\\) which is a common case for fine-tuning large language models (LLM) on consumer-level devices, the introduction of element-wise self-sparsified activation pruning methods leads to an increased time cost of up to \\(5\\times\\) which is not justified by the benefits obtained (\\(12\\%\\) memory reduction). PREBACKRAZOR demonstrates a slight improvement in convergence speed and reduction in loss compared to PREBACKRAZOR as shown in Fig. 4, highlighting its enhanced efficiency in fine-tuning, and outperforms not only BACKRAZOR but also Full-FT at an \\(80\\%\\) pruning ratio, achieving the highest accuracy of \\(72.92\\%\\). Overall, while both pruning methods effectively reduce memory usage over Full-FT, PREBACKRAZOR tends to maintain or achieve slightly higher accuracy and speedup than BACKRAZOR, demonstrates a better balance between memory efficiency, completion time, and accuracy, particularly at lower batch sizes and higher pruning ratios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.619, + 0.558 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this work, we reveal the inefficiency of the existing element-wise self-sparsified activation pruning method. Based on this finding, we develop a computation and memory-efficient activation pruning framework called PREBACKRAZOR for jointly pruning activations and gradients. We have demonstrated that integrating PREBACKRAZOR with efficient attention kernels can significantly enhance the accuracy of existing methods. In addition, we also theoretically prove that our method has the same convergence rate as BACKRAZOR. However, while PREBACKRAZOR has exhibited performance improvements in ViT and BERT fine-tuning, its applicability to larger foundation models warrants further investigation. Additionally, theoretical proofs and analyses for its effectiveness in such models require continued research efforts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.809, + 0.669, + 0.826 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.833, + 0.894, + 0.894 + ], + "angle": 0, + "content": "This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK349000005." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5890" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Dan Alistarh, Demjan Grubic, Jerry Li, Ryota Tomioka, and Milan Vojnovic. Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Arash Ardakani, Carlo Condo, and Warren J Gross. Activation pruning of deep convolutional neural networks. In 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1325-1329. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem. Machine learning, 47:235-256, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Animashree Anandkumar. *signsgd: Compressed optimisation for non-convex problems*. In International Conference on Machine Learning, pages 560–569. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[6] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Freezeout: Accelerate training by progressively freezing layers. arXiv preprint arXiv:1706.04983, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[7] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce activations, not trainable parameters for efficient ondevice learning. arXiv preprint arXiv:2007.11622, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[8] Ayan Chakrabarti and Benjamin Moseley. Backprop with approximate activations for memory-efficient network training. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[9] Congliang Chen, Li Shen, Haozhi Huang, and Wei Liu. Quantized adam with error feedback. ACM Transactions on Intelligent Systems and Technology (TIST), 12(5):1-26, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[10] Congliang Chen, Li Shen, Fangyu Zou, and Wei Liu. Towards practical adam: Non-convexity, convergence theory, and mini-batch acceleration. Journal of Machine Learning Research, 23(229):1-47, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[11] Jianfei Chen, Lianmin Zheng, Zhewei Yao, Dequan Wang, Ion Stoica, Michael Mahoney, and Joseph Gonzalez. Actnn: Reducing training memory footprint via 2-bit activation compressed training. In International Conference on Machine Learning, pages 1803-1813. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[12] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[14] Jack Choquette, Wishwesh Gandhi, Olivier Giroux, Nick Stam, and Ronny Krashinsky. Nvidia a100 tensor coregpu:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Performance and innovation. IEEE Micro, 41(2):29-35, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] Jack Choquette, Edward Lee, Ronny Krashinsky, Vishnu Balan, and Brucek Khailany. 3.2 the a100 datacenter gpu and ampere architecture. In 2021 IEEE International Solid-State Circuits Conference (ISSCC), pages 48-50. IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[17] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[18] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[21] Chao Fang, Wei Sun, Aojun Zhou, and Zhongfeng Wang. Efficient n: M sparse dnn training using algorithm, architecture, and dataflow co-design. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[23] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research, 18(1): 6869-6898, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[24] Ziyu Jiang, Xuxi Chen, Xueqin Huang, Xianzhi Du, Denny Zhou, and Zhangyang Wang. Back razor: Memory-efficient transfer learning by self-sparsified backpropagation. Advances in Neural Information Processing Systems, 35: 29248-29261, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[27] Bingrui Li, Jianfei Chen, and Jun Zhu. Memory efficient optimizers with 4-bit states. arXiv preprint arXiv:2309.01507, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[28] Tianjian Li, Haoran Xu, Philipp Koehn, and Kenton Murray. Efficiently harnessing parameter importance for better training. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[29] Junjie Liu, Zhe Xu, Runbin Shi, Ray CC Cheung, and Hayden KH So. Dynamic sparse training: Find efficient sparse" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5891" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "network from scratch with trainable masked layers. arXiv preprint arXiv:2005.06870, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[30] Rui Liu and Barzan Mozafari. Communication-efficient distributed learning for large batch optimization. In International Conference on Machine Learning, pages 13925-13946. PMLR, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[31] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[32] Kai Lv, Hang Yan, Qipeng Guo, Hajjun Lv, and Xipeng Qiu. Adalomo: Low-memory optimization with adaptive learning rate. arXiv preprint arXiv:2310.10195, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[33] Kai Lv, Yuqing Yang, Tengxiao Liu, Qinghui Gao, Qipeng Guo, and Xipeng Qiu. Full parameter fine-tuning for large language models with limited resources. arXiv preprint arXiv:2306.09782, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[34] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. arXiv preprint arXiv:2305.17333, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.439 + ], + "angle": 0, + "content": "[35] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[36] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In International Conference on Machine Learning, pages 4646-4655. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[37] Zizheng Pan, Peng Chen, Haoyu He, Jing Liu, Jianfei Cai, and Bohan Zhuang. Mesa: A memory-saving training framework for transformers. arXiv preprint arXiv:2111.11124, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[39] Jeff Pool. Accelerating sparsity in the nvidia ampere architecture. GTC 2020, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[40] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[41] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[42] Jie Ren, Samyam Rajbhandari, Reza Yazdani Aminabadi, Olatunj Ruwase, Shuangyan Yang, Minjia Zhang, Dong Li, and Yuxiong He. {ZeRO-Offload}: Democratizing {BillionScale} model training. In 2021 USENIX Annual Technical Conference (USENIX ATC 21), pages 551–564, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[44] Li Shen, Yan Sun, Zhiyuan Yu, Liang Ding, Xinmei Tian, and Dacheng Tao. On efficient training of large-scale deep learning models: A literature review. arXiv preprint arXiv:2304.03589, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[45] Sebastian U Stich, Jean-Baptiste Cordonnier, and Martin Jaggi. Sparsified sgd with memory. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. Neural Networks, 169:506-519, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[47] Mingjie Sun, Zhuang Liu, Anna Bair, and J. Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[48] Xu Sun, Xuancheng Ren, Shuming Ma, and Hofeng Wang. meprop: Sparsified back propagation for accelerated deep learning with reduced overfitting. In International Conference on Machine Learning, pages 3299-3308. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[49] Tijmen Tieleman, Geoffrey Hinton, et al. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26-31, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[51] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[52] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[53] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[54] Qing Yang, Wei Wen, Zuoguan Wang, and Hai Li. Joint regularization on activations and weights for efficient neural network pruning. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 790-797. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[55] Yuxin Zhang, Yiting Luo, Mingbao Lin, Yunshan Zhong, Jingjing Xie, Fei Chao, and Rongrong Ji. Bi-directional masks for efficient n: M sparse training. arXiv preprint arXiv:2302.06058, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.828, + 0.892, + 0.896 + ], + "angle": 0, + "content": "[56] Fangyu Zou, Li Shen, Zequn Jie, Weizhong Zhang, and Wei Liu. A sufficient condition for convergences of adam and rm-sprop. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11127-11135, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5892" + } + ] +] \ No newline at end of file diff --git a/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_origin.pdf b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76bc3402df75a4bbe6c6de35d056edcaf59fd3d1 --- /dev/null +++ b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/362c0568-56d1-4aa5-ab6f-aa648bbaa531_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e646700bd1ad4939810380f6dcac305638b490f33916a544fa15dbeb6a9732b +size 527190 diff --git a/2024/Sheared Backpropagation for Fine-tuning Foundation Models/full.md b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d8d5562d29ae43fc58a8a58fc3af98bc0aa22c2 --- /dev/null +++ b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/full.md @@ -0,0 +1,332 @@ +# Sheared Backpropagation for Fine-tuning Foundation Models + +Zhiyuan Yu $^{1}$ , Li Shen $^{*,2}$ , Liang Ding $^{2}$ , Xinmei Tian $^{1}$ , Yixin Chen $^{3}$ , and Dacheng Tao $^{4}$ $^{1}$ University of Science and Technology of China, Hefei, China; $^{2}$ JD Explore Academy, Beijing, China; $^{3}$ Washington University in St. Louis, MO, USA; $^{4}$ Nanyang Technological University, Singapore +yuzhiyuan@mail.ustc.edu.cn, xinmei@ustc.edu.cn, chen@cse.wustl.edu +{mathshenli, liangding.liam, dacheng.tao}@gmail.com + +# Abstract + +Fine-tuning is the process of extending the training of pre-trained models on specific target tasks, thereby significantly enhancing their performance across various applications. However, fine-tuning often demands large memory consumption, posing a challenge for low-memory devices that some previous memory-efficient fine-tuning methods attempted to mitigate by pruning activations for gradient computation, albeit at the cost of significant computational overhead from the pruning processes during training. To address these challenges, we introduce PreBackRazor, a novel activation pruning scheme offering both computational and memory efficiency through a sparsified backpropagation strategy, which judiciously avoids unnecessary activation pruning and storage and gradient computation. Before activation pruning, our approach samples a probability of selecting a portion of parameters to freeze, utilizing a bandit method for updates to prioritize impactful gradients on convergence. During the feed-forward pass, each model layer adjusts adaptively based on parameter activation status, obviating the need for sparsification and storage of redundant activations for subsequent backpropagation. Benchmarking on fine-tuning foundation models, our approach maintains baseline accuracy across diverse tasks, yielding over $20\%$ speedup and around $10\%$ memory reduction. Moreover, integrating with an advanced CUDA kernel achieves up to $60\%$ speedup without extra memory costs or accuracy loss, significantly enhancing the efficiency of fine-tuning foundation models on memory-constrained devices. + +# 1. Introduction + +Nowadays, Transformer-based models [50] have achieved remarkable results in various domains of deep learning [19, 20]. The trend towards scaling up these models for enhanced performance necessitates substantial data and computational resources, often beyond the reach of many + +researchers [44]. Consequently, fine-tuning pre-trained Transformer models, which are a prominent example of foundation models [5], has become a prevalent strategy. This approach allows for adapting and continuing the training of these models at reduced costs while retaining high performance. Despite the advantages, fine-tuning foundation models on low-end devices, such as personal computers with consumer-level GPUs and other edge devices, presents considerable challenges. These challenges stem primarily from the limited memory and computational capacity of such devices, restricting the feasible model and batch sizes during fine-tuning. This limitation often results in prolonged training periods or compromised performance. Our paper aims to address this issue by exploring efficient methodologies for fine-tuning Transformer models on consumer-level devices, striving to balance performance with training efficiency. + +One pervasive challenge in fine-tuning models is effectively managing the training pipeline within the confines of limited device memory [41]. Beyond loading trainable parameters, memory usage in model training includes storing intermediate tensors like gradients and activations. These elements are crucial in backpropagation and become particularly memory-intensive with large input batch sizes or model scales. Although efficient training techniques, such as low-precision and pruning, have been proposed to save memory on model parameters and are widely adopted [23, 31], recent research has shifted focus towards reducing memory consumption of activations, identified as a primary memory bottleneck [7]. Notably, since activations are stored solely for backward computation, recent studies on activation reduction [8, 24, 37] have proposed an asymmetric backpropagation pipeline. This approach maintains a dense forward pass and compresses activations only for gradient computation during backpropagation. Among these, BACKRAZOR by Jiang et al. [24] employs top- $k$ pruning in the forward pass to sparsify activa + +tions, thereby reducing memory requirements for their storage. This method then re-densifies them during the backward pass for gradient computation, achieving significant memory savings. However, since activation pruning methods do not retain the original activations, they typically require additional computations to determine optimal pruning locations, in order to maintain accuracy. This introduces a trade-off between memory reduction and computational overhead, necessitating a careful balance. Furthermore, it presents a challenge in enhancing the efficiency of current computation-intensive activation pruning methods. + +In our work, addressing this unresolved challenge, we present PREBACKRAZOR, a memory and computation-efficient activation pruning scheme. It aligns with the asymmetric sparsified backpropagation approach, where sparsified activations are preserved solely for the backpropagation process. PREBACKRAZOR operates by dynamically freezing non-essential activation coordinates, thus minimizing the necessity for their sparsification and storage. This method effectively reduces computational load, thereby boosting the overall computation efficiency. Before initiating each iteration's forward pass, we strategically assign a dropout probability to every activation coordinate. This probability is informed by the gradient values associated with these coordinates, derived from prior iterations. This process enables us to selectively freeze certain coordinates during the ensuing backward pass. The chosen probabilities aim to preserve those gradients that are crucial for model convergence. To continuously refine this probability distribution, we utilize a bandit method, an approach inspired by JOINTSPAR [30]. Consequently, during the forward pass, we bypass the pruning and storage of activations that are deemed non-essential for backpropagation. This method effectively diminishes the computational burden typically associated with activation pruning, while simultaneously upholding memory efficiency and training accuracy. + +Our experiments, conducted on a single RTX 4090 GPU, have shown significant improvements in training efficiency and memory management during the fine-tuning of Vision Transformer (ViT) and BERT models. For ViT, an extensive comparative analysis using the CIFAR datasets [26] revealed that our method, PREBACKRAZOR, outperforms traditional activation pruning techniques like BACKRAZOR in terms of computational efficiency. Notably, PREBACKRAZOR achieves an up to $1.7\times$ increase in training speed for linear layers, without incurring additional memory overhead. Furthermore, our carefully tailored adaptation of the FLASHATTENTION kernel has significantly accelerated the fine-tuning process, ensuring efficient memory usage. This customized integration results in an impressive $1.7\times$ boost in training speed, maintaining a balanced computational budget without sacrificing accuracy. In the context of BERT fine-tuning, our approach also results in a modest improve + +ment in time efficiency and final accuracy. + +In summary, our main contributions are three-fold: + +- We propose a novel activation pruning pipeline to further reduce the computational overhead of previous methods. In addition, we establish the convergence analysis for the proposed algorithm. +- We design improved activation storage schemes that jointly reduce the compute and memory. We employ bandit methods to update gradients across activation coordinates of the model layers adaptively. This selective update mechanism efficiently manages activation pruning and gradient computation, thereby reducing memory and computational costs based on BACKRAZOR's asymmetric activation-self-sparsified backpropagation. +- We conduct experiments on fine-tuning Transformer-based ViT and BERT, adapt the current code library, and combine the state-of-the-art FLASHATTENTION 2 CUDA [16] kernels to boost the training, making the pipeline more welcome by nowadays consumer-level devices. Our experiments show that our method reduces over $30\%$ of training time without loss of accuracy. + +# 2. Related Work + +In this section, we briefly review existing memory and computation-efficient fine-tuning methods. + +Memory-efficient fine-tuning methods. In memory-limited environments, strategies like partitioning and compression leverage trade-offs in computation, communication, and storage to reduce memory usage. Gradient accumulation lowers memory usage at the cost of increased communication, while gradient checkpointing [12] achieves memory savings through recomputation. CPU offloading [42], GPU partitioning [41], and FLASHATTENTION [16, 17] balance memory optimization against communication overhead. Several memory-efficient optimizers [32, 43, 49] have been proposed to approximate backpropagation to minimize stored information. Recent approaches include computing only forward passes with estimated gradients [34] and integrating gradient computation with parameter updates [33]. Additionally, low or mixed-precision training [27, 35] and other sparse backpropagation methods [13] have been integrated, trading some numerical accuracy for enhanced memory or communication efficiency. There are also related works focusing on communication cost reduction in distributed systems that use these memory-efficient methods for gradient compression [1, 4, 45], which is different from our settings. Additionally, while quantization [9, 18] and pruning [47] methods are commonly employed to compress gradients and model weights, our research concentrates on the pruning of activations. In the realm of activation footprint reduction, Cai et al. [7] suggest freezing layer weights while updating biases and adding lightweight residual modules. Other + +methods involve compression of activations [2, 11, 54]. To maintain accuracy, recent research proposes asymmetric sparsified backpropagation which stores compressed activations, using original dense inputs for forward computation and retrieving activations for backward gradient computations. They employ techniques like low-precision compression [8, 37], or utilizing top- $k$ masks for pruning activations [24]. + +Computation-efficient fine-tuning methods. Existing computation-efficient fine-tuning methods include PEFT, knowledge distillation [22], few-shot learning [52], as well as quantization and pruning implementations. Our research primarily focuses on techniques that dynamically adjust training and seamlessly integrate with our sparse backpropagation pipeline. In the realm of computation-efficient fine-tuning methods, we find adaptive learning rates [25, 49] relevant for their role in facilitating faster convergence [10, 46, 56]. Also, dynamic sparse algorithms [29, 36] are notable for introducing sparsity during training. For the aspect of sparse backpropagation, the implementation of top-k masking over gradients by Sun et al. [48] to induce sparsity in backpropagation is in line with our approach to asymmetric pruning. Structured sparsity patterns [21, 55] and layer freezing techniques [6, 40], along with selective weight freezing strategies [28], offer valuable insights into reducing computational load. We also draw specific elements from the JOINTSPAR algorithm [30], which selectively drops gradient vectors, a principle we adapt to target the computational challenges in activation pruning. + +# 3. Rethink the Backpropagation Redundancy + +In this section, we first present explanations of the memory cost of activations, and the asymmetric self-sparsified activations as introduced in BACKRAZOR algorithm [24]. Then we show the memory and computation redundancy in the current backpropagation scheme. + +# 3.1. Preliminary + +Consider the nonconvex optimization problem of fin-tuning foundation models: $\min_{\theta \in \Theta} \mathcal{L}(\theta)$ , where $\mathcal{L}$ is the training loss calculated with the model's weight $\theta$ , and $\Theta$ is the parameter space. Without loss of generality, the forward pass of a given neural network with $n$ layers can be expressed as follows: + +$$ +f \left(\mathbf {x} _ {0}; \boldsymbol {\theta}\right) = f _ {n} \left(f _ {n - 1} \left(\dots f _ {0} \left(\mathbf {x} _ {0}; \boldsymbol {\theta} _ {0}\right) \dots ; \boldsymbol {\theta} _ {n - 1}; \boldsymbol {\theta} _ {n}\right), \right. \tag {1} +$$ + +$$ +L = \mathcal {L} (f (\mathbf {x} _ {0}; \boldsymbol {\theta}), y) \tag {2} +$$ + +where $\mathbf{x}_0$ represents the input features at the 0-th layer, and $y$ corresponds to the input label. The functions $f_{i}$ and parameters $\theta_{i}$ are associated with the model's $i$ -th layer. The expression $f(\mathbf{x}_0; \boldsymbol{\theta})$ indicates the model's final prediction. + +In this section, the symbol $L$ stands for the loss between the prediction and the label, computed using the specified loss function $\mathcal{L}$ . We define the activation $\mathbf{x}_i$ as the output of the $(i - 1)$ -th layer, i.e., $\mathbf{x}_i = f_{i - 1}(\mathbf{x}_{i - 1};\pmb{\theta}_{i - 1})$ . + +Denoting $h_i^{(\mathbf{x})}(\mathbf{x}_i,\pmb {\theta}_i)$ and $h_i^{(\pmb \theta)}(\mathbf{x}_i,\pmb {\theta}_i)$ as the derivatives of $\mathbf{x}_{i + 1}$ with respect to $\mathbf{x}_i$ and $\pmb{\theta}_i$ respectively. Normally, the backward process can be defined as: + +$$ +\frac {\partial L}{\partial \mathbf {x} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} \frac {\partial \mathbf {x} _ {i + 1}}{\partial \mathbf {x} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} h _ {i} ^ {(\mathbf {x})} \left(\mathbf {x} _ {i}, \boldsymbol {\theta} _ {i}\right), \tag {3} +$$ + +$$ +\frac {\partial L}{\partial \boldsymbol {\theta} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} \frac {\partial \mathbf {x} _ {i + 1}}{\partial \boldsymbol {\theta} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} h _ {i} ^ {(\boldsymbol {\theta})} (\mathbf {x} _ {i}, \boldsymbol {\theta} _ {i}). \tag {4} +$$ + +For instance, considering a linear layer with parameters $\theta_{i}$ , which consist of a weight matrix $\mathbf{W}_i$ and a bias vector $\mathbf{b}_i$ , its forward pass function for computing the output $\mathbf{x}_{i + 1}$ is expressed as $\mathbf{x}_{i + 1} = \mathbf{x}_i\mathbf{W}_i^\top +\mathbf{b}_i$ . All of the variables $\mathbf{x}_i$ , $\mathbf{W}_i$ , and $\mathbf{b}_i$ are retained for gradient computation before the layer produces the output $\mathbf{x}_{i + 1}$ . Subsequently, during the backward pass of this layer in the loss backpropagation process, the gradients are calculated as: + +$$ +\frac {\partial L}{\partial \mathbf {x} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} \frac {\partial \mathbf {x} _ {i + 1}}{\partial \mathbf {x} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} \mathbf {W} _ {i}, \tag {5} +$$ + +$$ +\frac {\partial L}{\partial \mathbf {W} _ {i}} = \left(\frac {\partial L}{\partial \mathbf {x} _ {i + 1}}\right) ^ {\top} \mathbf {x} _ {i}, \tag {6} +$$ + +$$ +\frac {\partial L}{\partial \mathbf {b} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}}. \tag {7} +$$ + +Here, the activation $\mathbf{x_i}$ is involved in computing the gradient for updating the weight $\mathbf{W}_i$ . Consequently, the activation value must be stored in memory during the forward pass. This storage occupies a significant portion of memory, often exceeding the memory required for saving the weights. Additionally, the size of the activations increases linearly with the batch size. + +Jiang et al. [24] propose BACKRAZOR, a method that only prunes the activation saved for backpropagation and keeps it dense in the forward pass. The activation $\mathbf{x}_i$ is pruned and stored in the context as a sparse version $\tilde{\mathbf{x}}_i$ . The backward process with pruned activations turns into: + +$$ +\frac {\partial L}{\partial \mathbf {x} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} h _ {i} ^ {(\mathbf {x})} \left(\widetilde {\mathbf {x}} _ {i}, \boldsymbol {\theta} _ {i}\right), \tag {8} +$$ + +$$ +\frac {\partial L}{\partial \boldsymbol {\theta} _ {i}} = \frac {\partial L}{\partial \mathbf {x} _ {i + 1}} h _ {i} ^ {(\boldsymbol {\theta})} (\widetilde {\mathbf {x}} _ {i}, \boldsymbol {\theta} _ {i}). \tag {9} +$$ + +As for their activation pruning, a straightforward top- $k$ pruning method is applied, which identifies and removes the $k$ smallest magnitude values within the activation. Denoting $v$ as the total number of values of the activation, to introduce a targeted sparsity level, a fixed parameter $\lambda$ is set (e.g., $\lambda = 0.9$ ), and the pruning budget is determined by setting $k = \lambda v$ . Consequently, all activation values below the $k$ -th + +![](images/f72e5ae042119908f4351ea7777ec6778dec1c770995bac378f60d0d7fe36005.jpg) +Figure 1. Sparsity vs computation time cost between BackRazor (element-wise, self-sparsified) and proposed PreBackRazor (column-wise, joint-sparsified). + +smallest value are zeroed out. After this pruning operation, the pruned activation is stored as a binary bitmap, known as the mask. This mask indicates the positions of the nonzero elements. Additionally, a dense matrix is maintained to retain the values of the non-zero elements. + +# 3.2. Rethink the sparsity in BACKRAZOR + +Redundancy in full-gradient computation. Common backpropagation activation sparsification schemes involve restoring compressed activations, saved within the model context, back to their dense format during backpropagation. In the case of BACKRAZOR, this process entails densifying the sparse matrix to its original shape for gradient computation. Consequently, the full gradient is computed using the full-size activation, and the sparsity of the pruned activation is not utilized to expedite computation. + +Computation overhead in pruning. A prevalent pruning approach involves generating an element-wise self-sparsified top- $k$ mask to eliminate smaller values. This method, when employed for activation pruning, incurs additional computational costs. These costs include sorting to find the $k$ -th smallest activation value and performing element-wise comparisons across all activation elements to construct a sparse tensor. Collectively, these operations contribute to significant computational overhead, potentially extending training time beyond that of traditional fine-tuning. Such a trade-off for memory reduction may not be justifiable. Exploring alternative sparsity patterns and pruning schemes, such as column-wise and joint-sparse could lead to the discovery of faster algorithms, potentially replacing the current self-sparsified top- $k$ activation pruning method and thus reducing computation costs. + +Inspired by the observed inefficiencies in existing methods, we sought to investigate whether alternative sparsity patterns could offer reduced computation. Specifically, we compared the performance of column-wise sparsity with element-wise sparsity within the same linear layer, across various sparsity ratios, as depicted in Figure 1. Our find + +![](images/909f9ee58bea531031d905889bef4a4bdb1837f9d38e2df499118edf5a6ff301.jpg) +Figure 2. Pipeline of the proposed PreBackRazor. + +ings indicate notable up to $1.7 \times$ speedups, aligning with our initial hypothesis. Additionally, the column-wise activation sparsity pattern facilitates the use of existing joint sparse algorithms, which leverage gradient information to generate sparse masks previous to the BACKRAZOR pruning pipeline. This approach incurs lower computational costs. Given that the original sparsity is unstructured and of a low ratio, the integration of a structured sparsity pattern also holds promise for further acceleration, leveraging modern sparse computation software and hardware advancements [14, 21, 39]. + +# 4. Methodology + +In this section, we first propose the PREBACKRAZOR method. After that, we introduce the bandit method from JOINTSPAR algorithm [30] that we use to adjust the probability of dropping computations. Lastly, we examine the convergence of our proposed method. + +# 4.1. Overall pipeline of PREBACKRAZOR + +Our jointly sparsified activation pruning approach, PREBACKRAZOR, employs asymmetric backpropagation, maintaining dense computations in the forward pass while pruning activations for the backward pass, as depicted in Fig. 2. Inspired by the JOINTSPAR algorithm [30], our method goes a step further by incorporating gradient sparsity into the activation pruning, aiming to significantly boost computational efficiency. In PREBACKRAZOR, we manage a probability distribution over activation coordinates to selectively prune redundant indices. This process effectively reduces the computation of related gradients and halts the updating of corresponding weights, thereby enhancing training efficiency and minimizing computational requirements. The decision-making process hinges on the analysis of gradient norms, leveraging the joint sparsity of gradients and activations. This approach + +prioritizes the retention of the most impactful gradient coordinates for weight updates. The implementation details of this method are outlined in Algorithm 1. + +Algorithm 1 PREBACKRAZOR +1: Initialize all $p_0 = s / D$ +2: for $t = 1$ to $T$ do +3: Sample $\mathbf{x}_0, \mathbf{y}$ from $\mathcal{D}$ +4: for $i = 1$ to $n$ do +5: Instantiate $Z_{t,d}$ , with $P(Z_{t,d} = 1) = p_{t,d}$ +6: Determine $S_t = \{d : Z_{t,d} = 1\}$ +7: $\mathbf{x}_i = f_i(\mathbf{x}_{i-1}, \boldsymbol{\theta}_i)$ +8: mask = a binary bitmap of size $D$ +9: for parameter $\boldsymbol{\theta}_{[d]}$ of $\boldsymbol{\theta}_i$ do +10: if $d \in S_t$ then +11: $\text{mask}_d = \text{True}$ +12: end if +13: end for +14: $\tilde{\mathbf{x}}_{i-1} = \text{Sparsify}(\mathbf{x}_{i-1}, \text{mask})$ +15: SaveForBackward( $\tilde{\mathbf{x}}_{i-1}, \text{mask}, \boldsymbol{\theta}_i$ ) +16: end for +17: Compute loss $L = \mathcal{L}(\mathbf{x}_{n+1}, \mathbf{y})$ +18: for $i = n-1$ to 1 do +19: $\text{Densify}(\tilde{\mathbf{x}}_{i-1}, \text{mask})$ +20: $\frac{\partial L}{\partial \mathbf{x}_i} = \frac{\partial L}{\partial \mathbf{x}_{i+1}} h_i^{(x)}(\tilde{\mathbf{x}}_i, \boldsymbol{\theta}_i)$ +21: for parameter $\boldsymbol{\theta}_{[d]}$ with $d \in S_t$ do +22: $\frac{\partial L}{\partial \boldsymbol{\theta}_{[d]}} = \frac{\partial L}{\partial \mathbf{x}_{i+1}} h_i^{(\theta)}(\tilde{\mathbf{x}}_i, \boldsymbol{\theta}_{[d]})$ +23: end for +24: $p_{t+1} = \text{Update}(p_t, S_t, \{\mathbf{g}_{t,[d]}\}_{d \in S_t})$ +25: end for +26: end for + +For a given input activation $\mathbf{x}_i$ in the $i$ -th layer of the model, we decompose the activations column-wise, represented as $\mathbf{x}_i = [\mathbf{x}_{i,1},\dots ,\mathbf{x}_{i,d},\dots ,\mathbf{x}_{i,D}]$ , where $d$ and $D$ respectively denote the $d$ -th coordinate out of all $D$ column indices. Considering a linear layer, its weight parameter $\pmb{\theta}$ is similarly decomposed into $[\pmb{\theta}_{[1]},\dots ,\pmb{\theta}_{[d]},\dots ,\pmb{\theta}_{[D]}$ . To differentiate from the layer-wise parameter set $\pmb{\theta}_i$ (as discussed in Section 3.1), we use square brackets to denote each coordinate, represented as $\pmb{\theta}_{[d]}$ . The gradient of the loss with respect to $\pmb{\theta}_{[d]}$ is denoted by $\pmb{g}_{t,d}$ , and we define $\pmb{g}_{t,[d]}$ as a vector with all elements zero except the $d$ -th element: $\pmb{g}_{t,[d]} = [0,\dots ,\pmb{g}_{t,d},\dots ,0]$ , with $\pmb{g}_t$ being the sum of all such vectors, $\pmb{g}_t = \sum_{d = 1}^D\pmb{g}_{t,[d]}$ . + +At each iteration $t$ , a probability distribution $p_t = [p_{t,1}, \dots, p_{t,d}, \dots, p_{t,D}]$ is maintained over the joint coordinates of the parameters $[\pmb{\theta}_{[1]}, \dots, \pmb{\theta}_{[d]}, \dots, \pmb{\theta}_{[D]}\]$ and activations $[\mathbf{x}_1, \dots, \mathbf{x}_d, \dots, \mathbf{x}_D]$ . For each coordinate, a Bernoulli random variable $Z_{t,d}$ is sampled with $P(Z_{t,d} = 1) = p_{t,d}$ and $P(Z_{t,d} = 0) = 1 - p_{t,d}$ , deciding whether to include $(Z_{t,d} = 1)$ or exclude $(Z_{t,d} = 0)$ the gradient of $\pmb{\theta}_d$ during backpropagation. + +Initially, each $p_t$ is set uniformly to $\frac{s}{D}$ , aligning with the + +desired sparsity budget $s$ , where $s$ signifies the number of columns involved in the backward process. At the onset of each iteration $t$ , the Bernoulli random variable $Z_{t,d}$ is sampled to form the active set $S_{t}$ of coordinates that retain gradients in activations. Post gradient computation in backpropagation, this gradient information informs the update of the probability distribution, employing a bandit method which we will introduce in the subsequent section. + +Integrating the FLASHATTENTION kernel. The fine-tuning of Transformer-based models on consumer-grade devices presents a unique opportunity to leverage modern, efficient kernels for improved training efficiency. In our approach, we have adapted and integrated the FLASHATTENTION kernel, substituting the original, more segmented sparsified attention implementation in BACKRAZOR. To address the increased memory usage resultant from this integration, we have incorporated PREBACKRAZOR within the FLASHATTENTION function. This inclusion not only facilitates pruning but also employs recomputation strategies to mitigate precision loss. We designate this enhanced methodology as PREBACKRAZOR++. This innovative approach ensures memory usage remains on par with BACKRAZOR, while markedly accelerating the training process. We further combine JOINTSPAR over model parameters with our method as the PREBACKRAZOR++.JOINTSPAR for best memory and computation efficiency. Details on implementation are omitted here for brevity. + +# 4.2. Bandit method for gradient sparsification + +To update the probability distribution $p_t$ over $D$ coordinates at each iteration $t$ , we utilize a bandit method akin to that in JOINTSPAR [30], originally introduced by Auer et al. [3], as detailed in Algorithm 2. + +Algorithm 2 Bandit [30] for Updating Distribution $p_t$ +1: Algorithm: UPDATE $(p_t, S_t, \{g_{t,[d]}\}_{d \in S_t})$ +2: Input: $p_t$ : probability distribution, $S_t$ : set of active parameters, $\{g_{t,[d]}\}_{d \in S_t}$ : gradients +3: for $d = 1$ to $D$ do +4: if $d \in S_t$ then +5: $\tilde{l}_{t,d} = -\frac{\|g_{t,[d]}\|^2}{(p_{t,d})^2} + \frac{G^2}{p_{\min}^2}$ +6: else +7: $\tilde{l}_{t,d} = 0$ +8: end if +9: $w_{t,d} = p_{t,d} \exp(-\alpha_p \tilde{l}_{t,d} / p_{t,d})$ +10: end for +11: $p_{t+1} = \arg \min_{q \in \mathcal{P}} D_{\mathrm{KL}}(q \| w_t)$ +12: return $p_{t+1}$ + +We introduce a set $\mathcal{P}$ as $\mathcal{P} = \{p\in \mathbb{R}^D:\sum_{d = 1}^D p_d = s,p_d\geq p_{\mathrm{min}},\forall 1\leq d\leq D\}$ , where $p_{\mathrm{min}}$ is a constant lower bound. This set helps in guiding the update of the probability distribution. The sum of the probabilities across all + +parameters equals the desired sparsity budget $s$ , with each probability bounded below by a constant $p_{\mathrm{min}}$ . Note that $s$ and $p_{\mathrm{min}}$ are parameters defined by the user, adhering to the conditions $0 < s \leq D$ and $0 < p_{\mathrm{min}} \leq 1$ . Recall that $g_{t,[d]}$ is defined as $g_{t,[d]} = [0,\dots ,g_{t,d},\dots ,0]$ , and $g_{t} = \sum_{d = 1}^{D}g_{t,[d]}$ . The term $\alpha_{p}$ denotes the step size, while $D_{\mathrm{KL}}(q||w)$ represents the KL divergence between distributions $q$ and $w$ . + +# 4.3. Theoretical results + +In this section, we analyze the convergence of our proposed method PREBACKRAZOR++JOINTSPAR in the context of linear networks. We use $g_{t}$ to represent the gradient of the loss function $\mathcal{L}$ with respect to the parameter $\theta_{t}$ at iteration $t$ , calculated using all samples in the training dataset. In contrast, $\tilde{g}_{t}$ denotes the stochastic gradient computed using a single training sample. To support our analysis, we adopt the classical assumptions as detailed by Jiang et al. [24]: + +Assumption 1. (1) The objective function $\mathcal{L}$ is $S_{d}$ -smooth with respect to $\pmb{\theta}_{[d]}$ ; (2) The gradient $\pmb{g}_{t,[d]}$ is upper bounded by $G$ , i.e., $G \geq \| \pmb{g}_{t,[d]} \|$ . $\forall t, d$ ; (3) The objective function $\mathcal{L}$ is bounded below by $L^{*}$ ; (4) For each iteration $t$ , the stochastic gradient $\tilde{\pmb{g}}_t$ adheres to the following: $\mathbb{E}[\tilde{\pmb{g}}_t] = \pmb{g}_t$ , $\mathbb{E}[(\tilde{\pmb{g}}_t - \pmb{g}_t)_d^2] \leq \sigma_d^2$ , $\forall d \in \{1, 2, \dots, D\}$ , with $D$ denoting the number of coordinates and the constant $\sigma_d^2$ representing the variance corresponding to coordinate $d$ . + +Lemma 1 [24]. In the context of linear neural networks, the gradient of parameters at iteration $t$ can be decomposed as $\tilde{g}_t = \tilde{\mathbf{x}}_t\tilde{\mathbf{A}}_t^\top$ , where $\tilde{\mathbf{x}}_t$ denotes the precise activation, and $\tilde{\mathbf{A}}_t$ is referred to as the transformation matrix. After pruning the activations, only the pruned activations $\tilde{\mathbf{x}}_t'$ are reserved and used for backpropagation, and the gradient approximation becomes $\tilde{g}_t' = \tilde{\mathbf{x}}_t'\tilde{\mathbf{A}}_t^\top$ . + +Theorem 1 (Convergence). Denote $\alpha$ as the step size. Under Assumptions 1-4, the following bound holds: + +$$ +\begin{array}{l} \mathbb {E} \left[ \frac {1}{T} \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {n} \| \boldsymbol {g} _ {t, i} \| ^ {2} \right] \leq \frac {\mathcal {L} \left(\theta_ {0}\right) - L ^ {*}}{\alpha T} \\ + \frac {S _ {\infty} \alpha}{2 T} \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {n} \mathbb {E} \left[ \sum_ {d = 1} ^ {D} \frac {\| \boldsymbol {g} _ {t , [ d ]} \| ^ {2}}{p _ {t , d}} \right], \tag {10} \\ \end{array} +$$ + +where $T$ is the total number of iterations, $S_{\infty} = \max_d S_d$ , $\theta_0$ is the initial model weights. + +As the training progresses, the first term on the right-hand side of the equation progressively diminishes to zero, whereas the second term remains confined within the bounds set by the bandit method. This theorem establishes that our proposed method converges towards a bounded neighborhood around a stationary point. While our theoretical analysis here is confined to linear networks, the extension to non-linear networks remains an area for future investigation. In the meantime, we have empirically validated + +that our proposed method effectively accelerates the finetuning of Transformer-based models. For those interested in a more thorough exploration of the underlying mathematics, the detailed derivation of these results is provided in the appendix. Additionally, the proofs and analyses in [24] and [30] offer further insight into these concepts. + +# 5. Experiments + +We perform experiments on PyTorch [38] with RTX 4090 GPU with 24GB of VRAM. We apply our method on top of BACKRAZOR's code library to report the performance gains, following its training settings on ViT and BERT. We report the actual memory allocation on GPU measured with PyTorch CUDA tools. For better readability, we employ the abbreviation JS to denote JOINTSPAR, and PREBR as a shorthand for PREBACKRAZOR. + +# 5.1. PREBACKRAZOR with Vision Transformer + +Settings. For ViT training, we initiate the process by loading the checkpoint of the ViT-B/16 model that has been pre-trained on ImageNet-22K as our starting point. Subsequently, we fine-tune the model using the CIFAR-10 and CIFAR-100 datasets [26]. The fine-tuning process encompasses 20,000 steps and employs cosine learning rate decay, with an initial 500 steps dedicated to learning rate warmup. We utilize the standard SGD optimizer and conduct validation every 1,000 steps. The maximum learning rate is set at 0.01, and the batch size during training is 128. + +![](images/0d8d823856cf3a1fe7e2487de151bbc0e993c3a0eb23fbf6c88e0693c4d65ddd.jpg) +(a) Training. + +![](images/72fecaff58b37f1945b8863c0267e4238335b69ba4f6f0c35b304587e8f76042.jpg) +(b) Validation. +Figure 3. Fine-tuning PREBACKRAZOR++@90% vs. BACKRAZOR@90% on CIFAR-100. (a) Training loss with wall-clock time. (b) Validation accuracy with steps. + +Convergence comparison. The results of fine-tuning PREBACKRAZOR++ with parameter-wise JOINTSPAR in comparison to BACKRAZOR at a pruning rate of $90\%$ on CIFAR-100 are illustrated in Fig. 3. In Fig. 3a, the convergence speed of PREBACKRAZOR@90%++JOINTSPAR is comparable to that of BACKRAZOR@90%, with a slight advantage in the early stages of training. Notably, we observe a marginal improvement in accuracy, specifically $[0.33\%, 0.05\%]$ , when adjusting the initial probabilities of JOINTSPAR to $[70\%, 50\%]$ , respectively. This improvement is in contrast to the performance of BACKRAZOR@90%, as illustrated in Fig. 3b. + +Table 1. Comparison of fine-tuning methods for ViT. + +
MethodMemoryEpoch TimeSpeedupCIFAR-100CIFAR-10CUB-200
AllocatedReduction
FULL-FT19980MB1.0×2min59s-93.099.085.5
BACKRAZOR7540MB2.6×6min19s1.0×92.198.686.6
(ours) PREBR++JS@70%7440MB2.7×4min37s1.4×92.498.885.0
(ours) PREBR++JS@50%6880MB2.9×4min02s1.6×92.198.685.6
+ +Table 2. Performance of different attention implementations fine-tuning ViT on CIFAR-100. + +
DeviceMethodAttention TypeMemoryEpoch TimeSpeedupAccuracy(%)
RTX 4090FULL-FTvanilla19980MB2min59s-93.0
BACKRAZORBACKRAZOR7540MB6min19s1.0×92.1
(ours) PREBR++JSBACKRAZOR6570MB4min37s1.4×91.4
(ours) PREBR++JSFLASHATTENTION7960MB3min40s1.7×92.2
(ours) PREBR++JSsparse FLASHATTN6880MB4min02s1.6×92.1
RTX 2080 TiFULL-FTvanillaout of memory
BACKRAZORBACKRAZOR7780MB19min47s1.0×-
(ours) PREBR+JSBACKRAZOR6870MB15min13s1.3×-
+ +Speedup and memory reduction. In Tab. 1, we present a comprehensive comparison among PREBACKRAZOR@90%++JOINTSPAR@[70%, 50%], BACKRAZOR@90%, and FT-Full (fine-tuning the full network) from the perspective of memory usage, speedup relative to BACKRAZOR@90%, and final accuracy on CIFAR-100, CIFAR-10 and CUB-200 for each method. Here we present the average memory allocated by CUDA during training. Compared with FT-full, BACKRAZOR@90% achieves a noteworthy $2.6\times$ reduction in memory usage. However, it demands twice as much training time due to the additional computations required for processing activations. In contrast, PREBR++JS achieves superior computational efficiency by avoiding unnecessary computations through joint sparsification of gradients and activations. This leads to significantly $1.6\times$ faster training without incurring additional memory costs or compromising accuracy. Additionally, Tab. 3 showcases the performance gains of forward and backward computation achieved by transitioning from a self-sparsified top- $k$ pruning pattern to a column-wise joint-sparse pattern as in Fig. 1. These measurements are conducted on a linear layer characterized by input tensors with the shape [128, 197, 768] and of type float32. Tab. 4 illustrates the performance improvements achieved by substituting the element-wise mask in BACKRAZOR with a column-wise mask. This approach, referred to as PREBACKRAZOR-, employs a self-sparsified scheme that eschews the use of the bandit method for predicting pruning masks based on gradient information. Despite this, PREBACKRAZOR- still delivers commendable results, including a modest reduction in memory usage and over a $10\%$ increase in speed, while preserving nearly identical accuracy, particularly at a moderate pruning ratio of $80\%$ . + +Table 3. Acceleration on a linear layer. + +
SparsityFWD time (ms)SpeedupFWD + BWD (ms)Speedup
BACKRAZORPREBRBACKRAZORPREBR
0.14.233.141.34×9.027.721.17×
0.24.354.241.03×9.209.031.02×
0.34.454.351.02×9.349.181.02×
0.47.974.451.79×16.529.321.77×
0.58.947.971.12×18.4516.511.12×
+ +Table 4. Comparison of self-sparsified sparsity patterns. + +
Pruning RatioMethodMemory Peak UsageEpoch TimeAccuracy(%)
-FULL-FT19980MB2min59s93.0
90%BACKRAZOR (ours) PREBR-7540MB6min19s92.1
7220MB5min25s91.7
80%BACKRAZOR (ours) PREBR-8210MB6min28s92.9
8100MB5min49s92.9
+ +Integration of the FLASHATTENTION kernel. We present Tab. 2 that details the performance when switching between the original FLASHATTENTION kernel (without memory savings), sparsified FLASHATTENTION (our modified implementation), and no FLASHATTENTION (using BackRazor's attention implementation). We use (PRE)BACKRAZOR@90% with parameter-wise sparsified JOINTSPAR@50% for these settings, note that the incorporation of JOINTSPAR is specifically aimed at offsetting the dense FLASHATTENTION kernel's memory usage. Additionally, we evaluate the performance of these methods on RTX 2080 Ti GPU with 11GB of VRAM, with (PRE)BACKRAZOR@90% and JOINTSPAR@30%. Note that both FULL-FT and FLASHATTENTION-2 are not available for testing on RTX 2080 Ti as they exceed the VRAM capacity and require Ampere architecture [15], respectively. Our initiative to integrate the FLASHATTENTION kernel into the training process aims to adapt a state-of-the-art + +solution for enhancing training on consumer-level devices. This integration yields a significant speedup of $1.7 \times$ during training, effectively mitigating the impact of additional computation introduced when compared to vanilla implementation as in FULL-FT. However, the replacement of BACKRAZOR's sparsified attention implementation does come with more memory costs up to 1GB. In response to this challenge, we have customized FLASHATTENTION by incorporating the PREBACKRAZOR method to prune the gradients and activations. This customization serves to further reduce activations and counterbalance the increased memory usage, ultimately falling below the memory footprint of BACKRAZOR. + +# 5.2. PREBACKRAZOR with BERT + +Settings. For BERT training, we conduct fine-tuning on the bert-base-uncased model based on BACKRAZOR adaptations that are implemented within Huggingface's Transformers Repository [53]. We note that the model currently does not support FLASHATTENTION-2 to the best of our knowledge, from the involvement of an input attention mask in its attention algorithm which is not compatible with FLASHATTENTION kernel. We closely monitor and record the model's performance throughout 5 training epochs with batch size [8, 32] on a single GPU, which is a common case for fine-tuning on low-end devices. The selected fine-tuning task is Recognizing Textual Entailment (RTE), part of the GLUE benchmark [51]. The sparsity ratio for PREBACK-RAZOR and BACKRAZOR is set to $[80\%, 90\%]$ . + +Table 5. Performance of different fine-tuning schemes for RTE from GLUE Benchmark with BERT base. + +
Batch SizePruning RatioMethodMemory Peak UsageFinish TimeAccuracy(%)
32-FULL-FT16625MB2min50s70.40
90%BACKRAZOR (ours) PREBR11865MB6min26s66.43
11871MB6min24s65.70
80%BACKRAZOR (ours) PREBR12888MB6min41s65.70
13006MB6min19s66.07
8-FULL-FT5575MB4min28s70.04
90%BACKRAZOR (ours) PREBR4734MB16min59s66.43
4747MB16min05s69.68
80%BACKRAZOR (ours) PREBR4812MB17min39s69.31
4840MB16min36s72.92
+ +![](images/3ba4eafeeecedb43e55350069966295fd5527bca1798640b5d4e76eab4841b0b.jpg) +Figure 4. Training loss during fine-tuning BERT-base for RTE. + +Performance. We provide a comparison Tab. 5 in which we compare fine-tuning methods, including fine-tuning the full model, element-wise BACKRAZOR, and column-wise PREBACKRAZOR. Both pruning methods successfully reduce memory usage compared to the Full-FT approach. For instance, at the pruning ratio of $90\%$ and batch size of 32, BACKRAZOR and PREBR show similar memory usage significantly lower than Full-FT, yielding a saved memory of $28.6\%$ (4.8GB). The wall clock time for both pruning methods is higher compared to Full-FT. However, our proposed PREBACKRAZOR tends to have a slightly shorter time than BACKRAZOR in all cases, up to $9.7\% \times$ speedup. For $bs = 8$ which is a common case for fine-tuning large language models (LLM) on consumer-level devices, the introduction of element-wise self-sparsified activation pruning methods leads to an increased time cost of up to $5\times$ which is not justified by the benefits obtained ( $12\%$ memory reduction). PREBACKRAZOR demonstrates a slight improvement in convergence speed and reduction in loss compared to PREBACKRAZOR as shown in Fig. 4, highlighting its enhanced efficiency in fine-tuning, and outperforms not only BACKRAZOR but also Full-FT at an $80\%$ pruning ratio, achieving the highest accuracy of $72.92\%$ . Overall, while both pruning methods effectively reduce memory usage over Full-FT, PREBACKRAZOR tends to maintain or achieve slightly higher accuracy and speedup than BACKRAZOR, demonstrates a better balance between memory efficiency, completion time, and accuracy, particularly at lower batch sizes and higher pruning ratios. + +# 6. Conclusion + +In this work, we reveal the inefficiency of the existing element-wise self-sparsified activation pruning method. Based on this finding, we develop a computation and memory-efficient activation pruning framework called PREBACKRAZOR for jointly pruning activations and gradients. We have demonstrated that integrating PREBACKRAZOR with efficient attention kernels can significantly enhance the accuracy of existing methods. In addition, we also theoretically prove that our method has the same convergence rate as BACKRAZOR. However, while PREBACKRAZOR has exhibited performance improvements in ViT and BERT fine-tuning, its applicability to larger foundation models warrants further investigation. Additionally, theoretical proofs and analyses for its effectiveness in such models require continued research efforts. + +# Acknowledgements + +This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK349000005. + +# References + +[1] Dan Alistarh, Demjan Grubic, Jerry Li, Ryota Tomioka, and Milan Vojnovic. Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in neural information processing systems, 30, 2017. 2 +[2] Arash Ardakani, Carlo Condo, and Warren J Gross. Activation pruning of deep convolutional neural networks. In 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1325-1329. IEEE, 2017. 3 +[3] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem. Machine learning, 47:235-256, 2002. 5 +[4] Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Animashree Anandkumar. *signsgd: Compressed optimisation for non-convex problems*. In International Conference on Machine Learning, pages 560–569. PMLR, 2018. 2 +[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1 +[6] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Freezeout: Accelerate training by progressively freezing layers. arXiv preprint arXiv:1706.04983, 2017. 3 +[7] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce activations, not trainable parameters for efficient ondevice learning. arXiv preprint arXiv:2007.11622, 2020. 1, 2 +[8] Ayan Chakrabarti and Benjamin Moseley. Backprop with approximate activations for memory-efficient network training. Advances in Neural Information Processing Systems, 32, 2019. 1, 3 +[9] Congliang Chen, Li Shen, Haozhi Huang, and Wei Liu. Quantized adam with error feedback. ACM Transactions on Intelligent Systems and Technology (TIST), 12(5):1-26, 2021. 2 +[10] Congliang Chen, Li Shen, Fangyu Zou, and Wei Liu. Towards practical adam: Non-convexity, convergence theory, and mini-batch acceleration. Journal of Machine Learning Research, 23(229):1-47, 2022. 3 +[11] Jianfei Chen, Lianmin Zheng, Zhewei Yao, Dequan Wang, Ion Stoica, Michael Mahoney, and Joseph Gonzalez. Actnn: Reducing training memory footprint via 2-bit activation compressed training. In International Conference on Machine Learning, pages 1803-1813. PMLR, 2021. 3 +[12] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2 +[13] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022. 2 +[14] Jack Choquette, Wishwesh Gandhi, Olivier Giroux, Nick Stam, and Ronny Krashinsky. Nvidia a100 tensor coregpu: + +Performance and innovation. IEEE Micro, 41(2):29-35, 2021. 4 +[15] Jack Choquette, Edward Lee, Ronny Krashinsky, Vishnu Balan, and Brucek Khailany. 3.2 the a100 datacenter gpu and ampere architecture. In 2021 IEEE International Solid-State Circuits Conference (ISSCC), pages 48-50. IEEE, 2021. 7 +[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 2 +[17] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2 +[18] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314, 2023. 2 +[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[21] Chao Fang, Wei Sun, Aojun Zhou, and Zhongfeng Wang. Efficient n: M sparse dnn training using algorithm, architecture, and dataflow co-design. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 3, 4 +[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[23] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research, 18(1): 6869-6898, 2017. 1 +[24] Ziyu Jiang, Xuxi Chen, Xueqin Huang, Xianzhi Du, Denny Zhou, and Zhangyang Wang. Back razor: Memory-efficient transfer learning by self-sparsified backpropagation. Advances in Neural Information Processing Systems, 35: 29248-29261, 2022. 1, 3, 6 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6 +[27] Bingrui Li, Jianfei Chen, and Jun Zhu. Memory efficient optimizers with 4-bit states. arXiv preprint arXiv:2309.01507, 2023. 2 +[28] Tianjian Li, Haoran Xu, Philipp Koehn, and Kenton Murray. Efficiently harnessing parameter importance for better training. 2023. 3 +[29] Junjie Liu, Zhe Xu, Runbin Shi, Ray CC Cheung, and Hayden KH So. Dynamic sparse training: Find efficient sparse + +network from scratch with trainable masked layers. arXiv preprint arXiv:2005.06870, 2020.3 +[30] Rui Liu and Barzan Mozafari. Communication-efficient distributed learning for large batch optimization. In International Conference on Machine Learning, pages 13925-13946. PMLR, 2022. 2, 3, 4, 5, 6 +[31] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1 +[32] Kai Lv, Hang Yan, Qipeng Guo, Hajjun Lv, and Xipeng Qiu. Adalomo: Low-memory optimization with adaptive learning rate. arXiv preprint arXiv:2310.10195, 2023. 2 +[33] Kai Lv, Yuqing Yang, Tengxiao Liu, Qinghui Gao, Qipeng Guo, and Xipeng Qiu. Full parameter fine-tuning for large language models with limited resources. arXiv preprint arXiv:2306.09782, 2023. 2 +[34] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. arXiv preprint arXiv:2305.17333, 2023. 2 +[35] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 2 +[36] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In International Conference on Machine Learning, pages 4646-4655. PMLR, 2019. 3 +[37] Zizheng Pan, Peng Chen, Haoyu He, Jing Liu, Jianfei Cai, and Bohan Zhuang. Mesa: A memory-saving training framework for transformers. arXiv preprint arXiv:2111.11124, 2021. 1, 3 +[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[39] Jeff Pool. Accelerating sparsity in the nvidia ampere architecture. GTC 2020, 2020. 4 +[40] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3 +[41] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1, 2 +[42] Jie Ren, Samyam Rajbhandari, Reza Yazdani Aminabadi, Olatunj Ruwase, Shuangyan Yang, Minjia Zhang, Dong Li, and Yuxiong He. {ZeRO-Offload}: Democratizing {BillionScale} model training. In 2021 USENIX Annual Technical Conference (USENIX ATC 21), pages 551–564, 2021. 2 +[43] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International + +Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 2 +[44] Li Shen, Yan Sun, Zhiyuan Yu, Liang Ding, Xinmei Tian, and Dacheng Tao. On efficient training of large-scale deep learning models: A literature review. arXiv preprint arXiv:2304.03589, 2023. 1 +[45] Sebastian U Stich, Jean-Baptiste Cordonnier, and Martin Jaggi. Sparsified sgd with memory. Advances in Neural Information Processing Systems, 31, 2018. 2 +[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. Neural Networks, 169:506-519, 2024. 3 +[47] Mingjie Sun, Zhuang Liu, Anna Bair, and J. Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 2 +[48] Xu Sun, Xuancheng Ren, Shuming Ma, and Hofeng Wang. meprop: Sparsified back propagation for accelerated deep learning with reduced overfitting. In International Conference on Machine Learning, pages 3299-3308. PMLR, 2017. 3 +[49] Tijmen Tieleman, Geoffrey Hinton, et al. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26-31, 2012. 2, 3 +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1 +[51] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 8 +[52] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 3 +[53] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 8 +[54] Qing Yang, Wei Wen, Zuoguan Wang, and Hai Li. Joint regularization on activations and weights for efficient neural network pruning. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 790-797. IEEE, 2019. 3 +[55] Yuxin Zhang, Yiting Luo, Mingbao Lin, Yunshan Zhong, Jingjing Xie, Fei Chao, and Rongrong Ji. Bi-directional masks for efficient n: M sparse training. arXiv preprint arXiv:2302.06058, 2023. 3 +[56] Fangyu Zou, Li Shen, Zequn Jie, Weizhong Zhang, and Wei Liu. A sufficient condition for convergences of adam and rm-sprop. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11127-11135, 2019. 3 \ No newline at end of file diff --git a/2024/Sheared Backpropagation for Fine-tuning Foundation Models/images.zip b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0dbeafe2b27b07e423747af7011b77228cc35d8f --- /dev/null +++ b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75069fe9f4b347122e3412c5457a283f8a3683d8dc05b410b6aa3a790d31ca05 +size 360708 diff --git a/2024/Sheared Backpropagation for Fine-tuning Foundation Models/layout.json b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..df78a85de378de532c2639c4d8e113ee4017fd7a --- /dev/null +++ b/2024/Sheared Backpropagation for Fine-tuning Foundation Models/layout.json @@ -0,0 +1,10476 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 488, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 488, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 488, + 121 + ], + "type": "text", + "content": "Sheared Backpropagation for Fine-tuning Foundation Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "spans": [ + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": "Zhiyuan Yu" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": ", Li Shen" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{*,2}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": ", Liang Ding" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": ", Xinmei Tian" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": ", Yixin Chen" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": ", and Dacheng Tao" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": "University of Science and Technology of China, Hefei, China; " + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": "JD Explore Academy, Beijing, China; " + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": "Washington University in St. Louis, MO, USA; " + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 50, + 152, + 544, + 225 + ], + "type": "text", + "content": "Nanyang Technological University, Singapore \nyuzhiyuan@mail.ustc.edu.cn, xinmei@ustc.edu.cn, chen@cse.wustl.edu \n{mathshenli, liangding.liam, dacheng.tao}@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 261, + 192, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 261, + 192, + 274 + ], + "spans": [ + { + "bbox": [ + 143, + 261, + 192, + 274 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "text", + "content": "Fine-tuning is the process of extending the training of pre-trained models on specific target tasks, thereby significantly enhancing their performance across various applications. However, fine-tuning often demands large memory consumption, posing a challenge for low-memory devices that some previous memory-efficient fine-tuning methods attempted to mitigate by pruning activations for gradient computation, albeit at the cost of significant computational overhead from the pruning processes during training. To address these challenges, we introduce PreBackRazor, a novel activation pruning scheme offering both computational and memory efficiency through a sparsified backpropagation strategy, which judiciously avoids unnecessary activation pruning and storage and gradient computation. Before activation pruning, our approach samples a probability of selecting a portion of parameters to freeze, utilizing a bandit method for updates to prioritize impactful gradients on convergence. During the feed-forward pass, each model layer adjusts adaptively based on parameter activation status, obviating the need for sparsification and storage of redundant activations for subsequent backpropagation. Benchmarking on fine-tuning foundation models, our approach maintains baseline accuracy across diverse tasks, yielding over " + }, + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "text", + "content": " speedup and around " + }, + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "text", + "content": " memory reduction. Moreover, integrating with an advanced CUDA kernel achieves up to " + }, + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 47, + 277, + 289, + 613 + ], + "type": "text", + "content": " speedup without extra memory costs or accuracy loss, significantly enhancing the efficiency of fine-tuning foundation models on memory-constrained devices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 617, + 128, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 128, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 128, + 629 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 637, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 287, + 698 + ], + "type": "text", + "content": "Nowadays, Transformer-based models [50] have achieved remarkable results in various domains of deep learning [19, 20]. The trend towards scaling up these models for enhanced performance necessitates substantial data and computational resources, often beyond the reach of many" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 262, + 547, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 547, + 466 + ], + "type": "text", + "content": "researchers [44]. Consequently, fine-tuning pre-trained Transformer models, which are a prominent example of foundation models [5], has become a prevalent strategy. This approach allows for adapting and continuing the training of these models at reduced costs while retaining high performance. Despite the advantages, fine-tuning foundation models on low-end devices, such as personal computers with consumer-level GPUs and other edge devices, presents considerable challenges. These challenges stem primarily from the limited memory and computational capacity of such devices, restricting the feasible model and batch sizes during fine-tuning. This limitation often results in prolonged training periods or compromised performance. Our paper aims to address this issue by exploring efficient methodologies for fine-tuning Transformer models on consumer-level devices, striving to balance performance with training efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "One pervasive challenge in fine-tuning models is effectively managing the training pipeline within the confines of limited device memory [41]. Beyond loading trainable parameters, memory usage in model training includes storing intermediate tensors like gradients and activations. These elements are crucial in backpropagation and become particularly memory-intensive with large input batch sizes or model scales. Although efficient training techniques, such as low-precision and pruning, have been proposed to save memory on model parameters and are widely adopted [23, 31], recent research has shifted focus towards reducing memory consumption of activations, identified as a primary memory bottleneck [7]. Notably, since activations are stored solely for backward computation, recent studies on activation reduction [8, 24, 37] have proposed an asymmetric backpropagation pipeline. This approach maintains a dense forward pass and compresses activations only for gradient computation during backpropagation. Among these, BACKRAZOR by Jiang et al. [24] employs top-" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": " pruning in the forward pass to sparsify activa" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5883" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "tions, thereby reducing memory requirements for their storage. This method then re-densifies them during the backward pass for gradient computation, achieving significant memory savings. However, since activation pruning methods do not retain the original activations, they typically require additional computations to determine optimal pruning locations, in order to maintain accuracy. This introduces a trade-off between memory reduction and computational overhead, necessitating a careful balance. Furthermore, it presents a challenge in enhancing the efficiency of current computation-intensive activation pruning methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 507 + ], + "type": "text", + "content": "In our work, addressing this unresolved challenge, we present PREBACKRAZOR, a memory and computation-efficient activation pruning scheme. It aligns with the asymmetric sparsified backpropagation approach, where sparsified activations are preserved solely for the backpropagation process. PREBACKRAZOR operates by dynamically freezing non-essential activation coordinates, thus minimizing the necessity for their sparsification and storage. This method effectively reduces computational load, thereby boosting the overall computation efficiency. Before initiating each iteration's forward pass, we strategically assign a dropout probability to every activation coordinate. This probability is informed by the gradient values associated with these coordinates, derived from prior iterations. This process enables us to selectively freeze certain coordinates during the ensuing backward pass. The chosen probabilities aim to preserve those gradients that are crucial for model convergence. To continuously refine this probability distribution, we utilize a bandit method, an approach inspired by JOINTSPAR [30]. Consequently, during the forward pass, we bypass the pruning and storage of activations that are deemed non-essential for backpropagation. This method effectively diminishes the computational burden typically associated with activation pruning, while simultaneously upholding memory efficiency and training accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Our experiments, conducted on a single RTX 4090 GPU, have shown significant improvements in training efficiency and memory management during the fine-tuning of Vision Transformer (ViT) and BERT models. For ViT, an extensive comparative analysis using the CIFAR datasets [26] revealed that our method, PREBACKRAZOR, outperforms traditional activation pruning techniques like BACKRAZOR in terms of computational efficiency. Notably, PREBACKRAZOR achieves an up to " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "1.7\\times" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " increase in training speed for linear layers, without incurring additional memory overhead. Furthermore, our carefully tailored adaptation of the FLASHATTENTION kernel has significantly accelerated the fine-tuning process, ensuring efficient memory usage. This customized integration results in an impressive " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "1.7\\times" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " boost in training speed, maintaining a balanced computational budget without sacrificing accuracy. In the context of BERT fine-tuning, our approach also results in a modest improve" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 477, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 477, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 477, + 83 + ], + "type": "text", + "content": "ment in time efficiency and final accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 84, + 522, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 84, + 522, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 84, + 522, + 95 + ], + "type": "text", + "content": "In summary, our main contributions are three-fold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 98, + 545, + 324 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 98, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 98, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 98, + 545, + 144 + ], + "type": "text", + "content": "- We propose a novel activation pruning pipeline to further reduce the computational overhead of previous methods. In addition, we establish the convergence analysis for the proposed algorithm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 145, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 145, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 145, + 545, + 240 + ], + "type": "text", + "content": "- We design improved activation storage schemes that jointly reduce the compute and memory. We employ bandit methods to update gradients across activation coordinates of the model layers adaptively. This selective update mechanism efficiently manages activation pruning and gradient computation, thereby reducing memory and computational costs based on BACKRAZOR's asymmetric activation-self-sparsified backpropagation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 241, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 545, + 324 + ], + "type": "text", + "content": "- We conduct experiments on fine-tuning Transformer-based ViT and BERT, adapt the current code library, and combine the state-of-the-art FLASHATTENTION 2 CUDA [16] kernels to boost the training, making the pipeline more welcome by nowadays consumer-level devices. Our experiments show that our method reduces over " + }, + { + "bbox": [ + 306, + 241, + 545, + 324 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 306, + 241, + 545, + 324 + ], + "type": "text", + "content": " of training time without loss of accuracy." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 330, + 392, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 392, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 392, + 342 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 350, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 374 + ], + "type": "text", + "content": "In this section, we briefly review existing memory and computation-efficient fine-tuning methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 713 + ], + "type": "text", + "content": "Memory-efficient fine-tuning methods. In memory-limited environments, strategies like partitioning and compression leverage trade-offs in computation, communication, and storage to reduce memory usage. Gradient accumulation lowers memory usage at the cost of increased communication, while gradient checkpointing [12] achieves memory savings through recomputation. CPU offloading [42], GPU partitioning [41], and FLASHATTENTION [16, 17] balance memory optimization against communication overhead. Several memory-efficient optimizers [32, 43, 49] have been proposed to approximate backpropagation to minimize stored information. Recent approaches include computing only forward passes with estimated gradients [34] and integrating gradient computation with parameter updates [33]. Additionally, low or mixed-precision training [27, 35] and other sparse backpropagation methods [13] have been integrated, trading some numerical accuracy for enhanced memory or communication efficiency. There are also related works focusing on communication cost reduction in distributed systems that use these memory-efficient methods for gradient compression [1, 4, 45], which is different from our settings. Additionally, while quantization [9, 18] and pruning [47] methods are commonly employed to compress gradients and model weights, our research concentrates on the pruning of activations. In the realm of activation footprint reduction, Cai et al. [7] suggest freezing layer weights while updating biases and adding lightweight residual modules. Other" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5884" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "methods involve compression of activations [2, 11, 54]. To maintain accuracy, recent research proposes asymmetric sparsified backpropagation which stores compressed activations, using original dense inputs for forward computation and retrieving activations for backward gradient computations. They employ techniques like low-precision compression [8, 37], or utilizing top-" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " masks for pruning activations [24]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 175, + 289, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 428 + ], + "type": "text", + "content": "Computation-efficient fine-tuning methods. Existing computation-efficient fine-tuning methods include PEFT, knowledge distillation [22], few-shot learning [52], as well as quantization and pruning implementations. Our research primarily focuses on techniques that dynamically adjust training and seamlessly integrate with our sparse backpropagation pipeline. In the realm of computation-efficient fine-tuning methods, we find adaptive learning rates [25, 49] relevant for their role in facilitating faster convergence [10, 46, 56]. Also, dynamic sparse algorithms [29, 36] are notable for introducing sparsity during training. For the aspect of sparse backpropagation, the implementation of top-k masking over gradients by Sun et al. [48] to induce sparsity in backpropagation is in line with our approach to asymmetric pruning. Structured sparsity patterns [21, 55] and layer freezing techniques [6, 40], along with selective weight freezing strategies [28], offer valuable insights into reducing computational load. We also draw specific elements from the JOINTSPAR algorithm [30], which selectively drops gradient vectors, a principle we adapt to target the computational challenges in activation pruning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 438, + 282, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 282, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 282, + 453 + ], + "type": "text", + "content": "3. Rethink the Backpropagation Redundancy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "type": "text", + "content": "In this section, we first present explanations of the memory cost of activations, and the asymmetric self-sparsified activations as introduced in BACKRAZOR algorithm [24]. Then we show the memory and computation redundancy in the current backpropagation scheme." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 526, + 127, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 127, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 127, + 540 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": "Consider the nonconvex optimization problem of fin-tuning foundation models: " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\min_{\\theta \\in \\Theta} \\mathcal{L}(\\theta)" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " is the training loss calculated with the model's weight " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " is the parameter space. Without loss of generality, the forward pass of a given neural network with " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " layers can be expressed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 628, + 286, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 628, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 59, + 628, + 286, + 640 + ], + "type": "interline_equation", + "content": "f \\left(\\mathbf {x} _ {0}; \\boldsymbol {\\theta}\\right) = f _ {n} \\left(f _ {n - 1} \\left(\\dots f _ {0} \\left(\\mathbf {x} _ {0}; \\boldsymbol {\\theta} _ {0}\\right) \\dots ; \\boldsymbol {\\theta} _ {n - 1}; \\boldsymbol {\\theta} _ {n}\\right), \\right. \\tag {1}", + "image_path": "67bcbdb3fc9dec20fcbce377533b8a81e8b222573cfc1b0627c49197b21f9852.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 643, + 286, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 643, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 127, + 643, + 286, + 655 + ], + "type": "interline_equation", + "content": "L = \\mathcal {L} (f (\\mathbf {x} _ {0}; \\boldsymbol {\\theta}), y) \\tag {2}", + "image_path": "92db52f479b26caf02131256fb8a3194261628645db7398aa4496a5880346188.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " represents the input features at the 0-th layer, and " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " corresponds to the input label. The functions " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " and parameters " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " are associated with the model's " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "-th layer. The expression " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_0; \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " indicates the model's final prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "In this section, the symbol " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " stands for the loss between the prediction and the label, computed using the specified loss function " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ". We define the activation " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " as the output of the " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "-th layer, i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f_{i - 1}(\\mathbf{x}_{i - 1};\\pmb{\\theta}_{i - 1})" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "text", + "content": "Denoting " + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "inline_equation", + "content": "h_i^{(\\mathbf{x})}(\\mathbf{x}_i,\\pmb {\\theta}_i)" + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "inline_equation", + "content": "h_i^{(\\pmb \\theta)}(\\mathbf{x}_i,\\pmb {\\theta}_i)" + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "text", + "content": " as the derivatives of " + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_i" + }, + { + "bbox": [ + 304, + 121, + 545, + 158 + ], + "type": "text", + "content": " respectively. Normally, the backward process can be defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 338, + 167, + 545, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 167, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 338, + 167, + 545, + 191 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\mathbf {x})} \\left(\\mathbf {x} _ {i}, \\boldsymbol {\\theta} _ {i}\\right), \\tag {3}", + "image_path": "d71d3ba17c3dc2e02bba623d61ff1876052eda311d74d723756372ee1ed9db59.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 193, + 545, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 193, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 339, + 193, + 545, + 218 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\boldsymbol {\\theta})} (\\mathbf {x} _ {i}, \\boldsymbol {\\theta} _ {i}). \\tag {4}", + "image_path": "6e9e0bfdb21dbfa07110229e191cf9672786f77f1e2497916280cd214409780c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": "For instance, considering a linear layer with parameters " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": ", which consist of a weight matrix " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": " and a bias vector " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": ", its forward pass function for computing the output " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": " is expressed as " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1} = \\mathbf{x}_i\\mathbf{W}_i^\\top +\\mathbf{b}_i" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": ". All of the variables " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": " are retained for gradient computation before the layer produces the output " + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 304, + 227, + 545, + 322 + ], + "type": "text", + "content": ". Subsequently, during the backward pass of this layer in the loss backpropagation process, the gradients are calculated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 331, + 545, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 331, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 354, + 331, + 545, + 356 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\frac {\\partial \\mathbf {x} _ {i + 1}}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} \\mathbf {W} _ {i}, \\tag {5}", + "image_path": "b501d7dc8cd69d55b54c0805b9ff49b63029910b63e9f6a97909150236a70703.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 354, + 358, + 545, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 358, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 354, + 358, + 545, + 382 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial \\mathbf {W} _ {i}} = \\left(\\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}}\\right) ^ {\\top} \\mathbf {x} _ {i}, \\tag {6}", + "image_path": "fa02c9b7c42006a49fa57341bf0b8f6dfd74cf422dda163939c6c52e05d88d56.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 384, + 545, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 384, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 355, + 384, + 545, + 410 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial \\mathbf {b} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}}. \\tag {7}", + "image_path": "7309d0d6c47dad27b1a6b1d89f0f2287cda290328c163623f0143d9e961dacbd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "content": "Here, the activation " + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x_i}" + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "content": " is involved in computing the gradient for updating the weight " + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "content": ". Consequently, the activation value must be stored in memory during the forward pass. This storage occupies a significant portion of memory, often exceeding the memory required for saving the weights. Additionally, the size of the activations increases linearly with the batch size." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "content": "Jiang et al. [24] propose BACKRAZOR, a method that only prunes the activation saved for backpropagation and keeps it dense in the forward pass. The activation " + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "content": " is pruned and stored in the context as a sparse version " + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "content": ". The backward process with pruned activations turns into:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 372, + 570, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 570, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 372, + 570, + 545, + 594 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial \\mathbf {x} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\mathbf {x})} \\left(\\widetilde {\\mathbf {x}} _ {i}, \\boldsymbol {\\theta} _ {i}\\right), \\tag {8}", + "image_path": "ef5032991cedc58b95d4daee4a81daee222c95148c119226e7cae467563f7066.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 372, + 597, + 545, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 597, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 372, + 597, + 545, + 621 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial \\boldsymbol {\\theta} _ {i}} = \\frac {\\partial L}{\\partial \\mathbf {x} _ {i + 1}} h _ {i} ^ {(\\boldsymbol {\\theta})} (\\widetilde {\\mathbf {x}} _ {i}, \\boldsymbol {\\theta} _ {i}). \\tag {9}", + "image_path": "48be520dcdee28b327b28835df95c3e0c4e1e106ad07a9a69dec1126baba7635.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "As for their activation pruning, a straightforward top- " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " pruning method is applied, which identifies and removes the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " smallest magnitude values within the activation. Denoting " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " as the total number of values of the activation, to introduce a targeted sparsity level, a fixed parameter " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is set (e.g., " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0.9" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "), and the pruning budget is determined by setting " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "k = \\lambda v" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Consequently, all activation values below the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "-th" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5885" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 260, + 198 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 260, + 198 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 260, + 198 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 260, + 198 + ], + "type": "image", + "image_path": "f72e5ae042119908f4351ea7777ec6778dec1c770995bac378f60d0d7fe36005.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 237 + ], + "type": "text", + "content": "Figure 1. Sparsity vs computation time cost between BackRazor (element-wise, self-sparsified) and proposed PreBackRazor (column-wise, joint-sparsified)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 286, + 300 + ], + "type": "text", + "content": "smallest value are zeroed out. After this pruning operation, the pruned activation is stored as a binary bitmap, known as the mask. This mask indicates the positions of the nonzero elements. Additionally, a dense matrix is maintained to retain the values of the non-zero elements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 310, + 234, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 234, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 234, + 322 + ], + "type": "text", + "content": "3.2. Rethink the sparsity in BACKRAZOR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "type": "text", + "content": "Redundancy in full-gradient computation. Common backpropagation activation sparsification schemes involve restoring compressed activations, saved within the model context, back to their dense format during backpropagation. In the case of BACKRAZOR, this process entails densifying the sparse matrix to its original shape for gradient computation. Consequently, the full gradient is computed using the full-size activation, and the sparsity of the pruned activation is not utilized to expedite computation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "text", + "content": "Computation overhead in pruning. A prevalent pruning approach involves generating an element-wise self-sparsified top-" + }, + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "text", + "content": " mask to eliminate smaller values. This method, when employed for activation pruning, incurs additional computational costs. These costs include sorting to find the " + }, + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "text", + "content": "-th smallest activation value and performing element-wise comparisons across all activation elements to construct a sparse tensor. Collectively, these operations contribute to significant computational overhead, potentially extending training time beyond that of traditional fine-tuning. Such a trade-off for memory reduction may not be justifiable. Exploring alternative sparsity patterns and pruning schemes, such as column-wise and joint-sparse could lead to the discovery of faster algorithms, potentially replacing the current self-sparsified top-" + }, + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 450, + 286, + 641 + ], + "type": "text", + "content": " activation pruning method and thus reducing computation costs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "Inspired by the observed inefficiencies in existing methods, we sought to investigate whether alternative sparsity patterns could offer reduced computation. Specifically, we compared the performance of column-wise sparsity with element-wise sparsity within the same linear layer, across various sparsity ratios, as depicted in Figure 1. Our find" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 68, + 547, + 242 + ], + "blocks": [ + { + "bbox": [ + 328, + 68, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 328, + 68, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 328, + 68, + 547, + 242 + ], + "type": "image", + "image_path": "909f9ee58bea531031d905889bef4a4bdb1837f9d38e2df499118edf5a6ff301.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 250, + 515, + 261 + ], + "lines": [ + { + "bbox": [ + 334, + 250, + 515, + 261 + ], + "spans": [ + { + "bbox": [ + 334, + 250, + 515, + 261 + ], + "type": "text", + "content": "Figure 2. Pipeline of the proposed PreBackRazor." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 270, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 403 + ], + "type": "text", + "content": "ings indicate notable up to " + }, + { + "bbox": [ + 304, + 270, + 545, + 403 + ], + "type": "inline_equation", + "content": "1.7 \\times" + }, + { + "bbox": [ + 304, + 270, + 545, + 403 + ], + "type": "text", + "content": " speedups, aligning with our initial hypothesis. Additionally, the column-wise activation sparsity pattern facilitates the use of existing joint sparse algorithms, which leverage gradient information to generate sparse masks previous to the BACKRAZOR pruning pipeline. This approach incurs lower computational costs. Given that the original sparsity is unstructured and of a low ratio, the integration of a structured sparsity pattern also holds promise for further acceleration, leveraging modern sparse computation software and hardware advancements [14, 21, 39]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 415, + 388, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 388, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 388, + 428 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": "In this section, we first propose the PREBACKRAZOR method. After that, we introduce the bandit method from JOINTSPAR algorithm [30] that we use to adjust the probability of dropping computations. Lastly, we examine the convergence of our proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 503, + 491, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 491, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 491, + 516 + ], + "type": "text", + "content": "4.1. Overall pipeline of PREBACKRAZOR" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Our jointly sparsified activation pruning approach, PREBACKRAZOR, employs asymmetric backpropagation, maintaining dense computations in the forward pass while pruning activations for the backward pass, as depicted in Fig. 2. Inspired by the JOINTSPAR algorithm [30], our method goes a step further by incorporating gradient sparsity into the activation pruning, aiming to significantly boost computational efficiency. In PREBACKRAZOR, we manage a probability distribution over activation coordinates to selectively prune redundant indices. This process effectively reduces the computation of related gradients and halts the updating of corresponding weights, thereby enhancing training efficiency and minimizing computational requirements. The decision-making process hinges on the analysis of gradient norms, leveraging the joint sparsity of gradients and activations. This approach" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5886" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "type": "text", + "content": "prioritizes the retention of the most impactful gradient coordinates for weight updates. The implementation details of this method are outlined in Algorithm 1." + } + ] + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 49, + 137, + 287, + 429 + ], + "blocks": [ + { + "bbox": [ + 49, + 122, + 177, + 133 + ], + "lines": [ + { + "bbox": [ + 49, + 122, + 177, + 133 + ], + "spans": [ + { + "bbox": [ + 49, + 122, + 177, + 133 + ], + "type": "text", + "content": "Algorithm 1 PREBACKRAZOR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "lines": [ + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": "1: Initialize all " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "p_0 = s / D" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " do \n3: Sample " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0, \\mathbf{y}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n4: for " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " do \n5: Instantiate " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "Z_{t,d}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "P(Z_{t,d} = 1) = p_{t,d}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n6: Determine " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "S_t = \\{d : Z_{t,d} = 1\\}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f_i(\\mathbf{x}_{i-1}, \\boldsymbol{\\theta}_i)" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n8: mask = a binary bitmap of size " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n9: for parameter " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}_{[d]}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}_i" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " do \n10: if " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "d \\in S_t" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " then \n11: " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\text{mask}_d = \\text{True}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n12: end if \n13: end for \n14: " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{i-1} = \\text{Sparsify}(\\mathbf{x}_{i-1}, \\text{mask})" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n15: SaveForBackward(" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{i-1}, \\text{mask}, \\boldsymbol{\\theta}_i" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": ") \n16: end for \n17: Compute loss " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "L = \\mathcal{L}(\\mathbf{x}_{n+1}, \\mathbf{y})" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n18: for " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "i = n-1" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " to 1 do \n19: " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\text{Densify}(\\tilde{\\mathbf{x}}_{i-1}, \\text{mask})" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n20: " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial \\mathbf{x}_i} = \\frac{\\partial L}{\\partial \\mathbf{x}_{i+1}} h_i^{(x)}(\\tilde{\\mathbf{x}}_i, \\boldsymbol{\\theta}_i)" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n21: for parameter " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}_{[d]}" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "d \\in S_t" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " do \n22: " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial \\boldsymbol{\\theta}_{[d]}} = \\frac{\\partial L}{\\partial \\mathbf{x}_{i+1}} h_i^{(\\theta)}(\\tilde{\\mathbf{x}}_i, \\boldsymbol{\\theta}_{[d]})" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n23: end for \n24: " + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "inline_equation", + "content": "p_{t+1} = \\text{Update}(p_t, S_t, \\{\\mathbf{g}_{t,[d]}\\}_{d \\in S_t})" + }, + { + "bbox": [ + 49, + 137, + 287, + 429 + ], + "type": "text", + "content": " \n25: end for \n26: end for" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": "For a given input activation " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": "-th layer of the model, we decompose the activations column-wise, represented as " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = [\\mathbf{x}_{i,1},\\dots ,\\mathbf{x}_{i,d},\\dots ,\\mathbf{x}_{i,D}]" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " respectively denote the " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": "-th coordinate out of all " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " column indices. Considering a linear layer, its weight parameter " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " is similarly decomposed into " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "[\\pmb{\\theta}_{[1]},\\dots ,\\pmb{\\theta}_{[d]},\\dots ,\\pmb{\\theta}_{[D]}" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": ". To differentiate from the layer-wise parameter set " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_i" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " (as discussed in Section 3.1), we use square brackets to denote each coordinate, represented as " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{[d]}" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": ". The gradient of the loss with respect to " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{[d]}" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_{t,d}" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": ", and we define " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_{t,[d]}" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " as a vector with all elements zero except the " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": "-th element: " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_{t,[d]} = [0,\\dots ,\\pmb{g}_{t,d},\\dots ,0]" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": " being the sum of all such vectors, " + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_t = \\sum_{d = 1}^D\\pmb{g}_{t,[d]}" + }, + { + "bbox": [ + 46, + 445, + 287, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": "At each iteration " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": ", a probability distribution " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "p_t = [p_{t,1}, \\dots, p_{t,d}, \\dots, p_{t,D}]" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " is maintained over the joint coordinates of the parameters " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "[\\pmb{\\theta}_{[1]}, \\dots, \\pmb{\\theta}_{[d]}, \\dots, \\pmb{\\theta}_{[D]}\\]" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " and activations " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "[\\mathbf{x}_1, \\dots, \\mathbf{x}_d, \\dots, \\mathbf{x}_D]" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": ". For each coordinate, a Bernoulli random variable " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "Z_{t,d}" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " is sampled with " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "P(Z_{t,d} = 1) = p_{t,d}" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "P(Z_{t,d} = 0) = 1 - p_{t,d}" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": ", deciding whether to include " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "(Z_{t,d} = 1)" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " or exclude " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "(Z_{t,d} = 0)" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " the gradient of " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_d" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " during backpropagation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 715 + ], + "type": "text", + "content": "Initially, each " + }, + { + "bbox": [ + 59, + 701, + 287, + 715 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 59, + 701, + 287, + 715 + ], + "type": "text", + "content": " is set uniformly to " + }, + { + "bbox": [ + 59, + 701, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\frac{s}{D}" + }, + { + "bbox": [ + 59, + 701, + 287, + 715 + ], + "type": "text", + "content": ", aligning with the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "desired sparsity budget " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " signifies the number of columns involved in the backward process. At the onset of each iteration " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", the Bernoulli random variable " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "Z_{t,d}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " is sampled to form the active set " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " of coordinates that retain gradients in activations. Post gradient computation in backpropagation, this gradient information informs the update of the probability distribution, employing a bandit method which we will introduce in the subsequent section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 174, + 546, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 546, + 402 + ], + "type": "text", + "content": "Integrating the FLASHATTENTION kernel. The fine-tuning of Transformer-based models on consumer-grade devices presents a unique opportunity to leverage modern, efficient kernels for improved training efficiency. In our approach, we have adapted and integrated the FLASHATTENTION kernel, substituting the original, more segmented sparsified attention implementation in BACKRAZOR. To address the increased memory usage resultant from this integration, we have incorporated PREBACKRAZOR within the FLASHATTENTION function. This inclusion not only facilitates pruning but also employs recomputation strategies to mitigate precision loss. We designate this enhanced methodology as PREBACKRAZOR++. This innovative approach ensures memory usage remains on par with BACKRAZOR, while markedly accelerating the training process. We further combine JOINTSPAR over model parameters with our method as the PREBACKRAZOR++.JOINTSPAR for best memory and computation efficiency. Details on implementation are omitted here for brevity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 410, + 523, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 523, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 523, + 423 + ], + "type": "text", + "content": "4.2. Bandit method for gradient sparsification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "text", + "content": "To update the probability distribution " + }, + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "text", + "content": " coordinates at each iteration " + }, + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 428, + 545, + 476 + ], + "type": "text", + "content": ", we utilize a bandit method akin to that in JOINTSPAR [30], originally introduced by Auer et al. [3], as detailed in Algorithm 2." + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 310, + 502, + 545, + 651 + ], + "blocks": [ + { + "bbox": [ + 307, + 487, + 523, + 500 + ], + "lines": [ + { + "bbox": [ + 307, + 487, + 523, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 523, + 500 + ], + "type": "text", + "content": "Algorithm 2 Bandit [30] for Updating Distribution " + }, + { + "bbox": [ + 307, + 487, + 523, + 500 + ], + "type": "inline_equation", + "content": "p_t" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "lines": [ + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": "1: Algorithm: UPDATE " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "(p_t, S_t, \\{g_{t,[d]}\\}_{d \\in S_t})" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " \n2: Input: " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": ": probability distribution, " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": ": set of active parameters, " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\{g_{t,[d]}\\}_{d \\in S_t}" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": ": gradients \n3: for " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "d = 1" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " do \n4: if " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "d \\in S_t" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " then \n5: " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\tilde{l}_{t,d} = -\\frac{\\|g_{t,[d]}\\|^2}{(p_{t,d})^2} + \\frac{G^2}{p_{\\min}^2}" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " \n6: else \n7: " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\tilde{l}_{t,d} = 0" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " \n8: end if \n9: " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "w_{t,d} = p_{t,d} \\exp(-\\alpha_p \\tilde{l}_{t,d} / p_{t,d})" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " \n10: end for \n11: " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "p_{t+1} = \\arg \\min_{q \\in \\mathcal{P}} D_{\\mathrm{KL}}(q \\| w_t)" + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "text", + "content": " \n12: return " + }, + { + "bbox": [ + 310, + 502, + 545, + 651 + ], + "type": "inline_equation", + "content": "p_{t+1}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "text", + "content": "We introduce a set " + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p\\in \\mathbb{R}^D:\\sum_{d = 1}^D p_d = s,p_d\\geq p_{\\mathrm{min}},\\forall 1\\leq d\\leq D\\}" + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{min}}" + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "text", + "content": " is a constant lower bound. This set helps in guiding the update of the probability distribution. The sum of the probabilities across all" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5887" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": "parameters equals the desired sparsity budget " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": ", with each probability bounded below by a constant " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{min}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{min}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": " are parameters defined by the user, adhering to the conditions " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "0 < s \\leq D" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "0 < p_{\\mathrm{min}} \\leq 1" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": ". Recall that " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "g_{t,[d]}" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "g_{t,[d]} = [0,\\dots ,g_{t,d},\\dots ,0]" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "g_{t} = \\sum_{d = 1}^{D}g_{t,[d]}" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\alpha_{p}" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": " denotes the step size, while " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}(q||w)" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": " represents the KL divergence between distributions " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 72, + 289, + 172 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 176, + 157, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 157, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 157, + 188 + ], + "type": "text", + "content": "4.3. Theoretical results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "text", + "content": "In this section, we analyze the convergence of our proposed method PREBACKRAZOR++JOINTSPAR in the context of linear networks. We use " + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "text", + "content": " to represent the gradient of the loss function " + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "text", + "content": " with respect to the parameter " + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "text", + "content": ", calculated using all samples in the training dataset. In contrast, " + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{t}" + }, + { + "bbox": [ + 46, + 194, + 287, + 290 + ], + "type": "text", + "content": " denotes the stochastic gradient computed using a single training sample. To support our analysis, we adopt the classical assumptions as detailed by Jiang et al. [24]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": "Assumption 1. (1) The objective function " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "S_{d}" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": "-smooth with respect to " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{[d]}" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": "; (2) The gradient " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\pmb{g}_{t,[d]}" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": " is upper bounded by " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "G \\geq \\| \\pmb{g}_{t,[d]} \\|" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\forall t, d" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": "; (3) The objective function " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": " is bounded below by " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "L^{*}" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": "; (4) For each iteration " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": ", the stochastic gradient " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{g}}_t" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": " adheres to the following: " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\tilde{\\pmb{g}}_t] = \\pmb{g}_t" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[(\\tilde{\\pmb{g}}_t - \\pmb{g}_t)_d^2] \\leq \\sigma_d^2" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\forall d \\in \\{1, 2, \\dots, D\\}" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": " denoting the number of coordinates and the constant " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\sigma_d^2" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": " representing the variance corresponding to coordinate " + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 290, + 288, + 387 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "content": "Lemma 1 [24]. In the context of linear neural networks, the gradient of parameters at iteration " + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "content": " can be decomposed as " + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{g}_t = \\tilde{\\mathbf{x}}_t\\tilde{\\mathbf{A}}_t^\\top" + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_t" + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "content": " denotes the precise activation, and " + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}}_t" + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "content": " is referred to as the transformation matrix. After pruning the activations, only the pruned activations " + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_t'" + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "content": " are reserved and used for backpropagation, and the gradient approximation becomes " + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{g}_t' = \\tilde{\\mathbf{x}}_t'\\tilde{\\mathbf{A}}_t^\\top" + }, + { + "bbox": [ + 46, + 391, + 287, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "content": "Theorem 1 (Convergence). Denote " + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "content": " as the step size. Under Assumptions 1-4, the following bound holds:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 509, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 509, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 85, + 509, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} \\left[ \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\| \\boldsymbol {g} _ {t, i} \\| ^ {2} \\right] \\leq \\frac {\\mathcal {L} \\left(\\theta_ {0}\\right) - L ^ {*}}{\\alpha T} \\\\ + \\frac {S _ {\\infty} \\alpha}{2 T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {n} \\mathbb {E} \\left[ \\sum_ {d = 1} ^ {D} \\frac {\\| \\boldsymbol {g} _ {t , [ d ]} \\| ^ {2}}{p _ {t , d}} \\right], \\tag {10} \\\\ \\end{array}", + "image_path": "6db21ba9e7c66ceb0a7edcd143b61de3e33395a309f97880fb6bffa8a0989e42.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": " is the total number of iterations, " + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "inline_equation", + "content": "S_{\\infty} = \\max_d S_d" + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": " is the initial model weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "As the training progresses, the first term on the right-hand side of the equation progressively diminishes to zero, whereas the second term remains confined within the bounds set by the bandit method. This theorem establishes that our proposed method converges towards a bounded neighborhood around a stationary point. While our theoretical analysis here is confined to linear networks, the extension to non-linear networks remains an area for future investigation. In the meantime, we have empirically validated" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "that our proposed method effectively accelerates the finetuning of Transformer-based models. For those interested in a more thorough exploration of the underlying mathematics, the detailed derivation of these results is provided in the appendix. Additionally, the proofs and analyses in [24] and [30] offer further insight into these concepts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 154, + 388, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 154, + 388, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 154, + 388, + 167 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 174, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 269 + ], + "type": "text", + "content": "We perform experiments on PyTorch [38] with RTX 4090 GPU with 24GB of VRAM. We apply our method on top of BACKRAZOR's code library to report the performance gains, following its training settings on ViT and BERT. We report the actual memory allocation on GPU measured with PyTorch CUDA tools. For better readability, we employ the abbreviation JS to denote JOINTSPAR, and PREBR as a shorthand for PREBACKRAZOR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 277, + 520, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 520, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 520, + 289 + ], + "type": "text", + "content": "5.1. PREBACKRAZOR with Vision Transformer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 295, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 415 + ], + "type": "text", + "content": "Settings. For ViT training, we initiate the process by loading the checkpoint of the ViT-B/16 model that has been pre-trained on ImageNet-22K as our starting point. Subsequently, we fine-tune the model using the CIFAR-10 and CIFAR-100 datasets [26]. The fine-tuning process encompasses 20,000 steps and employs cosine learning rate decay, with an initial 500 steps dedicated to learning rate warmup. We utilize the standard SGD optimizer and conduct validation every 1,000 steps. The maximum learning rate is set at 0.01, and the batch size during training is 128." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 426, + 423, + 509 + ], + "blocks": [ + { + "bbox": [ + 310, + 426, + 423, + 509 + ], + "lines": [ + { + "bbox": [ + 310, + 426, + 423, + 509 + ], + "spans": [ + { + "bbox": [ + 310, + 426, + 423, + 509 + ], + "type": "image", + "image_path": "0d8d823856cf3a1fe7e2487de151bbc0e993c3a0eb23fbf6c88e0693c4d65ddd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 514, + 388, + 523 + ], + "lines": [ + { + "bbox": [ + 345, + 514, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 345, + 514, + 388, + 523 + ], + "type": "text", + "content": "(a) Training." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 429, + 426, + 541, + 510 + ], + "blocks": [ + { + "bbox": [ + 429, + 426, + 541, + 510 + ], + "lines": [ + { + "bbox": [ + 429, + 426, + 541, + 510 + ], + "spans": [ + { + "bbox": [ + 429, + 426, + 541, + 510 + ], + "type": "image", + "image_path": "72fecaff58b37f1945b8863c0267e4238335b69ba4f6f0c35b304587e8f76042.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 514, + 509, + 523 + ], + "lines": [ + { + "bbox": [ + 461, + 514, + 509, + 523 + ], + "spans": [ + { + "bbox": [ + 461, + 514, + 509, + 523 + ], + "type": "text", + "content": "(b) Validation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 533, + 545, + 567 + ], + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 567 + ], + "type": "text", + "content": "Figure 3. Fine-tuning PREBACKRAZOR++@90% vs. BACKRAZOR@90% on CIFAR-100. (a) Training loss with wall-clock time. (b) Validation accuracy with steps." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Convergence comparison. The results of fine-tuning PREBACKRAZOR++ with parameter-wise JOINTSPAR in comparison to BACKRAZOR at a pruning rate of " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " on CIFAR-100 are illustrated in Fig. 3. In Fig. 3a, the convergence speed of PREBACKRAZOR@90%++JOINTSPAR is comparable to that of BACKRAZOR@90%, with a slight advantage in the early stages of training. Notably, we observe a marginal improvement in accuracy, specifically " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "[0.33\\%, 0.05\\%]" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", when adjusting the initial probabilities of JOINTSPAR to " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "[70\\%, 50\\%]" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", respectively. This improvement is in contrast to the performance of BACKRAZOR@90%, as illustrated in Fig. 3b." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5888" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 83, + 523, + 163 + ], + "blocks": [ + { + "bbox": [ + 200, + 71, + 394, + 82 + ], + "lines": [ + { + "bbox": [ + 200, + 71, + 394, + 82 + ], + "spans": [ + { + "bbox": [ + 200, + 71, + 394, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of fine-tuning methods for ViT." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 83, + 523, + 163 + ], + "lines": [ + { + "bbox": [ + 73, + 83, + 523, + 163 + ], + "spans": [ + { + "bbox": [ + 73, + 83, + 523, + 163 + ], + "type": "table", + "html": "
MethodMemoryEpoch TimeSpeedupCIFAR-100CIFAR-10CUB-200
AllocatedReduction
FULL-FT19980MB1.0×2min59s-93.099.085.5
BACKRAZOR7540MB2.6×6min19s1.0×92.198.686.6
(ours) PREBR++JS@70%7440MB2.7×4min37s1.4×92.498.885.0
(ours) PREBR++JS@50%6880MB2.9×4min02s1.6×92.198.685.6
", + "image_path": "3754d4a9c6c30ce2a73e8da5c46a3e0d27b8163bffab6f87dcf72f96714bdd3e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 73, + 181, + 523, + 306 + ], + "blocks": [ + { + "bbox": [ + 129, + 168, + 464, + 179 + ], + "lines": [ + { + "bbox": [ + 129, + 168, + 464, + 179 + ], + "spans": [ + { + "bbox": [ + 129, + 168, + 464, + 179 + ], + "type": "text", + "content": "Table 2. Performance of different attention implementations fine-tuning ViT on CIFAR-100." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 181, + 523, + 306 + ], + "lines": [ + { + "bbox": [ + 73, + 181, + 523, + 306 + ], + "spans": [ + { + "bbox": [ + 73, + 181, + 523, + 306 + ], + "type": "table", + "html": "
DeviceMethodAttention TypeMemoryEpoch TimeSpeedupAccuracy(%)
RTX 4090FULL-FTvanilla19980MB2min59s-93.0
BACKRAZORBACKRAZOR7540MB6min19s1.0×92.1
(ours) PREBR++JSBACKRAZOR6570MB4min37s1.4×91.4
(ours) PREBR++JSFLASHATTENTION7960MB3min40s1.7×92.2
(ours) PREBR++JSsparse FLASHATTN6880MB4min02s1.6×92.1
RTX 2080 TiFULL-FTvanillaout of memory
BACKRAZORBACKRAZOR7780MB19min47s1.0×-
(ours) PREBR+JSBACKRAZOR6870MB15min13s1.3×-
", + "image_path": "b66f23c08c66f0a2a843be9f30a99f61fbb985b7eac13dc84620dcfd3b1b91a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "text", + "content": "Speedup and memory reduction. In Tab. 1, we present a comprehensive comparison among PREBACKRAZOR@90%++JOINTSPAR@[70%, 50%], BACKRAZOR@90%, and FT-Full (fine-tuning the full network) from the perspective of memory usage, speedup relative to BACKRAZOR@90%, and final accuracy on CIFAR-100, CIFAR-10 and CUB-200 for each method. Here we present the average memory allocated by CUDA during training. Compared with FT-full, BACKRAZOR@90% achieves a noteworthy " + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "inline_equation", + "content": "2.6\\times" + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "text", + "content": " reduction in memory usage. However, it demands twice as much training time due to the additional computations required for processing activations. In contrast, PREBR++JS achieves superior computational efficiency by avoiding unnecessary computations through joint sparsification of gradients and activations. This leads to significantly " + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "inline_equation", + "content": "1.6\\times" + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "text", + "content": " faster training without incurring additional memory costs or compromising accuracy. Additionally, Tab. 3 showcases the performance gains of forward and backward computation achieved by transitioning from a self-sparsified top-" + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "text", + "content": " pruning pattern to a column-wise joint-sparse pattern as in Fig. 1. These measurements are conducted on a linear layer characterized by input tensors with the shape [128, 197, 768] and of type float32. Tab. 4 illustrates the performance improvements achieved by substituting the element-wise mask in BACKRAZOR with a column-wise mask. This approach, referred to as PREBACKRAZOR-, employs a self-sparsified scheme that eschews the use of the bandit method for predicting pruning masks based on gradient information. Despite this, PREBACKRAZOR- still delivers commendable results, including a modest reduction in memory usage and over a " + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "text", + "content": " increase in speed, while preserving nearly identical accuracy, particularly at a moderate pruning ratio of " + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 47, + 315, + 289, + 710 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 326, + 545, + 392 + ], + "blocks": [ + { + "bbox": [ + 354, + 314, + 497, + 324 + ], + "lines": [ + { + "bbox": [ + 354, + 314, + 497, + 324 + ], + "spans": [ + { + "bbox": [ + 354, + 314, + 497, + 324 + ], + "type": "text", + "content": "Table 3. Acceleration on a linear layer." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 326, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 307, + 326, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 545, + 392 + ], + "type": "table", + "html": "
SparsityFWD time (ms)SpeedupFWD + BWD (ms)Speedup
BACKRAZORPREBRBACKRAZORPREBR
0.14.233.141.34×9.027.721.17×
0.24.354.241.03×9.209.031.02×
0.34.454.351.02×9.349.181.02×
0.47.974.451.79×16.529.321.77×
0.58.947.971.12×18.4516.511.12×
", + "image_path": "09340828259a715b237c6158a6219fcd560caf4bbd137a09efa75ebf0ca6a374.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 409, + 545, + 490 + ], + "blocks": [ + { + "bbox": [ + 324, + 396, + 527, + 407 + ], + "lines": [ + { + "bbox": [ + 324, + 396, + 527, + 407 + ], + "spans": [ + { + "bbox": [ + 324, + 396, + 527, + 407 + ], + "type": "text", + "content": "Table 4. Comparison of self-sparsified sparsity patterns." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 409, + 545, + 490 + ], + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 490 + ], + "type": "table", + "html": "
Pruning RatioMethodMemory Peak UsageEpoch TimeAccuracy(%)
-FULL-FT19980MB2min59s93.0
90%BACKRAZOR (ours) PREBR-7540MB6min19s92.1
7220MB5min25s91.7
80%BACKRAZOR (ours) PREBR-8210MB6min28s92.9
8100MB5min49s92.9
", + "image_path": "823ae95b2771f501d9d94a4c73a702587ac0751a50e751fa33961a589247339f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "Integration of the FLASHATTENTION kernel. We present Tab. 2 that details the performance when switching between the original FLASHATTENTION kernel (without memory savings), sparsified FLASHATTENTION (our modified implementation), and no FLASHATTENTION (using BackRazor's attention implementation). We use (PRE)BACKRAZOR@90% with parameter-wise sparsified JOINTSPAR@50% for these settings, note that the incorporation of JOINTSPAR is specifically aimed at offsetting the dense FLASHATTENTION kernel's memory usage. Additionally, we evaluate the performance of these methods on RTX 2080 Ti GPU with 11GB of VRAM, with (PRE)BACKRAZOR@90% and JOINTSPAR@30%. Note that both FULL-FT and FLASHATTENTION-2 are not available for testing on RTX 2080 Ti as they exceed the VRAM capacity and require Ampere architecture [15], respectively. Our initiative to integrate the FLASHATTENTION kernel into the training process aims to adapt a state-of-the-art" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5889" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "solution for enhancing training on consumer-level devices. This integration yields a significant speedup of " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "1.7 \\times" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " during training, effectively mitigating the impact of additional computation introduced when compared to vanilla implementation as in FULL-FT. However, the replacement of BACKRAZOR's sparsified attention implementation does come with more memory costs up to 1GB. In response to this challenge, we have customized FLASHATTENTION by incorporating the PREBACKRAZOR method to prune the gradients and activations. This customization serves to further reduce activations and counterbalance the increased memory usage, ultimately falling below the memory footprint of BACKRAZOR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 236, + 198, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 198, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 198, + 247 + ], + "type": "text", + "content": "5.2. PREBACKRAZOR with BERT" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 254, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 289, + 422 + ], + "type": "text", + "content": "Settings. For BERT training, we conduct fine-tuning on the bert-base-uncased model based on BACKRAZOR adaptations that are implemented within Huggingface's Transformers Repository [53]. We note that the model currently does not support FLASHATTENTION-2 to the best of our knowledge, from the involvement of an input attention mask in its attention algorithm which is not compatible with FLASHATTENTION kernel. We closely monitor and record the model's performance throughout 5 training epochs with batch size [8, 32] on a single GPU, which is a common case for fine-tuning on low-end devices. The selected fine-tuning task is Recognizing Textual Entailment (RTE), part of the GLUE benchmark [51]. The sparsity ratio for PREBACK-RAZOR and BACKRAZOR is set to " + }, + { + "bbox": [ + 46, + 254, + 289, + 422 + ], + "type": "inline_equation", + "content": "[80\\%, 90\\%]" + }, + { + "bbox": [ + 46, + 254, + 289, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 463, + 288, + 588 + ], + "blocks": [ + { + "bbox": [ + 47, + 431, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 47, + 431, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 287, + 453 + ], + "type": "text", + "content": "Table 5. Performance of different fine-tuning schemes for RTE from GLUE Benchmark with BERT base." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 463, + 288, + 588 + ], + "lines": [ + { + "bbox": [ + 48, + 463, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 288, + 588 + ], + "type": "table", + "html": "
Batch SizePruning RatioMethodMemory Peak UsageFinish TimeAccuracy(%)
32-FULL-FT16625MB2min50s70.40
90%BACKRAZOR (ours) PREBR11865MB6min26s66.43
11871MB6min24s65.70
80%BACKRAZOR (ours) PREBR12888MB6min41s65.70
13006MB6min19s66.07
8-FULL-FT5575MB4min28s70.04
90%BACKRAZOR (ours) PREBR4734MB16min59s66.43
4747MB16min05s69.68
80%BACKRAZOR (ours) PREBR4812MB17min39s69.31
4840MB16min36s72.92
", + "image_path": "0114f561c220836586f809a4bc4652d398153284239bb1bd64cc0cbe5296fb07.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 609, + 272, + 711 + ], + "blocks": [ + { + "bbox": [ + 61, + 609, + 272, + 711 + ], + "lines": [ + { + "bbox": [ + 61, + 609, + 272, + 711 + ], + "spans": [ + { + "bbox": [ + 61, + 609, + 272, + 711 + ], + "type": "image", + "image_path": "3ba4eafeeecedb43e55350069966295fd5527bca1798640b5d4e76eab4841b0b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 712, + 281, + 724 + ], + "lines": [ + { + "bbox": [ + 52, + 712, + 281, + 724 + ], + "spans": [ + { + "bbox": [ + 52, + 712, + 281, + 724 + ], + "type": "text", + "content": "Figure 4. Training loss during fine-tuning BERT-base for RTE." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": "Performance. We provide a comparison Tab. 5 in which we compare fine-tuning methods, including fine-tuning the full model, element-wise BACKRAZOR, and column-wise PREBACKRAZOR. Both pruning methods successfully reduce memory usage compared to the Full-FT approach. For instance, at the pruning ratio of " + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": " and batch size of 32, BACKRAZOR and PREBR show similar memory usage significantly lower than Full-FT, yielding a saved memory of " + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "28.6\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": " (4.8GB). The wall clock time for both pruning methods is higher compared to Full-FT. However, our proposed PREBACKRAZOR tends to have a slightly shorter time than BACKRAZOR in all cases, up to " + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "9.7\\% \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": " speedup. For " + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "bs = 8" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": " which is a common case for fine-tuning large language models (LLM) on consumer-level devices, the introduction of element-wise self-sparsified activation pruning methods leads to an increased time cost of up to " + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": " which is not justified by the benefits obtained (" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": " memory reduction). PREBACKRAZOR demonstrates a slight improvement in convergence speed and reduction in loss compared to PREBACKRAZOR as shown in Fig. 4, highlighting its enhanced efficiency in fine-tuning, and outperforms not only BACKRAZOR but also Full-FT at an " + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": " pruning ratio, achieving the highest accuracy of " + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "inline_equation", + "content": "72.92\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": ". Overall, while both pruning methods effectively reduce memory usage over Full-FT, PREBACKRAZOR tends to maintain or achieve slightly higher accuracy and speedup than BACKRAZOR, demonstrates a better balance between memory efficiency, completion time, and accuracy, particularly at lower batch sizes and higher pruning ratios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": "In this work, we reveal the inefficiency of the existing element-wise self-sparsified activation pruning method. Based on this finding, we develop a computation and memory-efficient activation pruning framework called PREBACKRAZOR for jointly pruning activations and gradients. We have demonstrated that integrating PREBACKRAZOR with efficient attention kernels can significantly enhance the accuracy of existing methods. In addition, we also theoretically prove that our method has the same convergence rate as BACKRAZOR. However, while PREBACKRAZOR has exhibited performance improvements in ViT and BERT fine-tuning, its applicability to larger foundation models warrants further investigation. Additionally, theoretical proofs and analyses for its effectiveness in such models require continued research efforts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 640, + 409, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 640, + 409, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 640, + 409, + 654 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 659, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 659, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 659, + 547, + 708 + ], + "type": "text", + "content": "This work is supported by STI 2030—Major Projects (No. 2021ZD0201405), NSFC No. 62222117, the Fundamental Research Funds for the Central Universities under contract WK349000005." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5890" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Dan Alistarh, Demjan Grubic, Jerry Li, Ryota Tomioka, and Milan Vojnovic. Qsgd: Communication-efficient sgd via gradient quantization and encoding. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Arash Ardakani, Carlo Condo, and Warren J Gross. Activation pruning of deep convolutional neural networks. In 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1325-1329. IEEE, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "text", + "content": "[3] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem. Machine learning, 47:235-256, 2002. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 270 + ], + "type": "text", + "content": "[4] Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Animashree Anandkumar. *signsgd: Compressed optimisation for non-convex problems*. In International Conference on Machine Learning, pages 560–569. PMLR, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 327 + ], + "type": "text", + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 361 + ], + "type": "text", + "content": "[6] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Freezeout: Accelerate training by progressively freezing layers. arXiv preprint arXiv:1706.04983, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 406 + ], + "type": "text", + "content": "[7] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce activations, not trainable parameters for efficient ondevice learning. arXiv preprint arXiv:2007.11622, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 452 + ], + "type": "text", + "content": "[8] Ayan Chakrabarti and Benjamin Moseley. Backprop with approximate activations for memory-efficient network training. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 496 + ], + "type": "text", + "content": "[9] Congliang Chen, Li Shen, Haozhi Huang, and Wei Liu. Quantized adam with error feedback. ACM Transactions on Intelligent Systems and Technology (TIST), 12(5):1-26, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "text", + "content": "[10] Congliang Chen, Li Shen, Fangyu Zou, and Wei Liu. Towards practical adam: Non-convexity, convergence theory, and mini-batch acceleration. Journal of Machine Learning Research, 23(229):1-47, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[11] Jianfei Chen, Lianmin Zheng, Zhewei Yao, Dequan Wang, Ion Stoica, Michael Mahoney, and Joseph Gonzalez. Actnn: Reducing training memory footprint via 2-bit activation compressed training. In International Conference on Machine Learning, pages 1803-1813. PMLR, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[12] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[13] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[14] Jack Choquette, Wishwesh Gandhi, Olivier Giroux, Nick Stam, and Ronny Krashinsky. Nvidia a100 tensor coregpu:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Performance and innovation. IEEE Micro, 41(2):29-35, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[15] Jack Choquette, Edward Lee, Ronny Krashinsky, Vishnu Balan, and Brucek Khailany. 3.2 the a100 datacenter gpu and ampere architecture. In 2021 IEEE International Solid-State Circuits Conference (ISSCC), pages 48-50. IEEE, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "text", + "content": "[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "text", + "content": "[17] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 253 + ], + "type": "text", + "content": "[18] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "text", + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "text", + "content": "[21] Chao Fang, Wei Sun, Aojun Zhou, and Zhongfeng Wang. Efficient n: M sparse dnn training using algorithm, architecture, and dataflow co-design. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "type": "text", + "content": "[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 455, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 509 + ], + "type": "text", + "content": "[23] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran ElYaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research, 18(1): 6869-6898, 2017. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "text", + "content": "[24] Ziyu Jiang, Xuxi Chen, Xueqin Huang, Xianzhi Du, Denny Zhou, and Zhangyang Wang. Back razor: Memory-efficient transfer learning by self-sparsified backpropagation. Advances in Neural Information Processing Systems, 35: 29248-29261, 2022. 1, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 601, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 623 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "text", + "content": "[27] Bingrui Li, Jianfei Chen, and Jun Zhu. Memory efficient optimizers with 4-bit states. arXiv preprint arXiv:2309.01507, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[28] Tianjian Li, Haoran Xu, Philipp Koehn, and Kenton Murray. Efficiently harnessing parameter importance for better training. 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[29] Junjie Liu, Zhe Xu, Runbin Shi, Ray CC Cheung, and Hayden KH So. Dynamic sparse training: Find efficient sparse" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5891" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "network from scratch with trainable masked layers. arXiv preprint arXiv:2005.06870, 2020.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[30] Rui Liu and Barzan Mozafari. Communication-efficient distributed learning for large batch optimization. In International Conference on Machine Learning, pages 13925-13946. PMLR, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[31] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "text", + "content": "[32] Kai Lv, Hang Yan, Qipeng Guo, Hajjun Lv, and Xipeng Qiu. Adalomo: Low-memory optimization with adaptive learning rate. arXiv preprint arXiv:2310.10195, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "type": "text", + "content": "[33] Kai Lv, Yuqing Yang, Tengxiao Liu, Qinghui Gao, Qipeng Guo, and Xipeng Qiu. Full parameter fine-tuning for large language models with limited resources. arXiv preprint arXiv:2306.09782, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "text", + "content": "[34] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. arXiv preprint arXiv:2305.17333, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "type": "text", + "content": "[35] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "text", + "content": "[36] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In International Conference on Machine Learning, pages 4646-4655. PMLR, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[37] Zizheng Pan, Peng Chen, Haoyu He, Jing Liu, Jianfei Cai, and Bohan Zhuang. Mesa: A memory-saving training framework for transformers. arXiv preprint arXiv:2111.11124, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "text", + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 525 + ], + "type": "text", + "content": "[39] Jeff Pool. Accelerating sparsity in the nvidia ampere architecture. GTC 2020, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "text", + "content": "[40] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[41] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "text", + "content": "[42] Jie Ren, Samyam Rajbhandari, Reza Yazdani Aminabadi, Olatunj Ruwase, Shuangyan Yang, Minjia Zhang, Dong Li, and Yuxiong He. {ZeRO-Offload}: Democratizing {BillionScale} model training. In 2021 USENIX Annual Technical Conference (USENIX ATC 21), pages 551–564, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[43] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 709 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[44] Li Shen, Yan Sun, Zhiyuan Yu, Liang Ding, Xinmei Tian, and Dacheng Tao. On efficient training of large-scale deep learning models: A literature review. arXiv preprint arXiv:2304.03589, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 140, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 174 + ], + "type": "text", + "content": "[45] Sebastian U Stich, Jean-Baptiste Cordonnier, and Martin Jaggi. Sparsified sgd with memory. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[46] Hao Sun, Li Shen, Qihuang Zhong, Liang Ding, Shixiang Chen, Jingwei Sun, Jing Li, Guangzhong Sun, and Dacheng Tao. Adasam: Boosting sharpness-aware minimization with adaptive learning rate and momentum for training deep neural networks. Neural Networks, 169:506-519, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 230, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 263 + ], + "type": "text", + "content": "[47] Mingjie Sun, Zhuang Liu, Anna Bair, and J. Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 318 + ], + "type": "text", + "content": "[48] Xu Sun, Xuancheng Ren, Shuming Ma, and Hofeng Wang. meprop: Sparsified back propagation for accelerated deep learning with reduced overfitting. In International Conference on Machine Learning, pages 3299-3308. PMLR, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "text", + "content": "[49] Tijmen Tieleman, Geoffrey Hinton, et al. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26-31, 2012. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "text", + "content": "[51] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "text", + "content": "[52] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 498, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 554 + ], + "type": "text", + "content": "[53] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[54] Qing Yang, Wei Wen, Zuoguan Wang, and Hai Li. Joint regularization on activations and weights for efficient neural network pruning. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 790-797. IEEE, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "text", + "content": "[55] Yuxin Zhang, Yiting Luo, Mingbao Lin, Yunshan Zhong, Jingjing Xie, Fei Chao, and Rongrong Ji. Bi-directional masks for efficient n: M sparse training. arXiv preprint arXiv:2302.06058, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 655, + 545, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 655, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 307, + 655, + 545, + 709 + ], + "type": "text", + "content": "[56] Fangyu Zou, Li Shen, Zequn Jie, Weizhong Zhang, and Wei Liu. A sufficient condition for convergences of adam and rm-sprop. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11127-11135, 2019. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5892" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_content_list.json b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9dd16e2930f2872e4eda24c5f60fb2fa19edefbb --- /dev/null +++ b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_content_list.json @@ -0,0 +1,1665 @@ +[ + { + "type": "text", + "text": "Sherpa3D: Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior", + "text_level": 1, + "bbox": [ + 91, + 130, + 877, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fangfu Liu $^{1}$ , Diankun Wu $^{1}$ , Yi Wei $^{1}$ , Yongming Rao $^{2}$ , Yueqi Duan $^{1\\dagger}$ , Tsinghua University, BAAI", + "bbox": [ + 212, + 179, + 754, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8c3f96632c6ebab61211bbb34c256e0ab87fc38cead4172796f0e738cebb26c.jpg", + "image_caption": [ + "Figure 1. Gallery of Sherpa3D: Blender rendering for various textured meshes from Sherpa3D, which is able to generate high-fidelity, diverse, and multi-view consistent 3D contents with input text prompts. Our method is also compatible with popular graphics engines." + ], + "image_footnote": [], + "bbox": [ + 80, + 226, + 888, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 736, + 313, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, 3D content creation from text prompts has demonstrated remarkable progress by utilizing 2D and 3D diffusion models. While 3D diffusion models ensure great multi-view consistency, their ability to generate high-quality and diverse 3D assets is hindered by the limited 3D data. In contrast, 2D diffusion models find a distillation approach that achieves excellent generalization", + "bbox": [ + 75, + 768, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and rich details without any 3D data. However, 2D lifting methods suffer from inherent view-agnostic ambiguity thereby leading to serious multi-face Janus issues, where text prompts fail to provide sufficient guidance to learn coherent 3D results. Instead of retraining a costly viewpoint-aware model, we study how to fully exploit easily accessible coarse 3D knowledge to enhance the prompts and guide 2D lifting optimization for refinement. In this paper, we propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric", + "bbox": [ + 496, + 738, + 893, + 891 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 101, + 888, + 228, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20763", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "consistency simultaneously. Specifically, we design a pair of guiding strategies derived from the coarse 3D prior generated by the 3D diffusion model: a structural guidance for geometric fidelity and a semantic guidance for 3D coherence. Employing the two types of guidance, the 2D diffusion model enriches the 3D content with diversified and high-quality results. Extensive experiments show the superiority of our Sherpa3D over the state-of-the-art text-to-3D methods in terms of quality and 3D consistency. Project page: https://liuff19.github.io/Sherpa3D/.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 271, + 209, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D content generation [34, 37, 56, 76] finds a broad range of applications, including games, movies, virtual/augmented reality and robots. However, the conventional process of creating premium 3D assets is still expensive and challenging as it requires multiple labor-intensive and time-consuming stages [31]. Fortunately, this challenge has prompted the development of recent text-to-3D methods [10, 25, 27, 35, 37, 47, 50, 56, 74]. Only using text prompts to automate 3D generation, these techniques pave a promising way towards streamlining 3D creation.", + "bbox": [ + 75, + 296, + 468, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Powered by the great breakthroughs in diffusion models [52, 60, 61, 81], two research lines of rationalization have recently emerged in text-to-3D: inference-only 3D diffusion methods and optimization-based 2D lifting methods. Specifically, the inference-only methods [20, 28, 52] seek to directly generate 3D-consistent assets by extensively training a new diffusion model on 3D data. However, due to the scarcity of 3D datasets compared to accessible 2D images or text data, these 3D diffusion models suffer from low quality and limited generalizability. Without requiring any 3D data for training, 2D lifting methods [4, 10, 37, 47, 56, 72, 74] can produce high-quality and diversified 3D results by distilling 3D knowledge from pre-trained 2D diffusion models [60, 61, 81], also known as Score Distillation Sampling (SDS). Yet lifting 2D observations into 3D is inherently ambiguous without sufficient 3D guidance from text prompts, leading to notorious multi-view inconsistency (e.g., Janus problems) in 2D lifting methods.", + "bbox": [ + 75, + 448, + 468, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These findings motivate us to think: is it possible to bridge the two aforementioned streams to achieve generalizability, high-fidelity, and geometric consistency simultaneously? An intuitive idea is to leverage more 3D data [11, 12] to fine-tune a view-point aware diffusion model, but it requires substantial computational resources and is prone to overfitting due to data bias [40, 67]. In contrast, our key insight is to utilize the easily accessible 3D diffusion model as guidance and study how to fully exploit coarse 3D knowledge to guide 2D lifting optimization for refinement. In particular, when maintaining the quality and generalizability of the original 2D diffusion model, we hope", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the 2D lifting awareness can be guided by the strong 3D geometric information from the 3D diffusion model. However, it is non-trivial in pursuit of this balance. Relying too heavily on the coarse 3D priors from the 3D diffusion model may degrade the generation quality, whereas little 3D guidance could result in a lack of geometric awareness, leading to multi-view inconsistency.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards this end, we propose Sherpa3D in this paper, which greatly boosts high-fidelity and highly diversified text-to-3D generation with geometric consistency. Our method begins by employing a 3D diffusion model to craft a basic 3D guide with limited details. Building upon the coarse 3D prior, we introduce two guiding strategies to inform 2D diffusion model throughout lifting optimization: a structural guide for geometric fidelity and a semantic guide for 3D coherence. Specifically, the structural guide leverages the first-order gradient information of the normals from the 3D prior to supervise the optimization of the structure. These normals are then integrated into the input of a pre-trained 2D diffusion model, refining the geometric details. Concurrently, our semantic guide extracts high-level features from multi-views of the 3D prior. These features guide the 2D lifting optimization to perceive the geometric consistency under the preservation of original generalizability and quality. Furthermore, we design an annealing function, which modulates the influence of the 3D guidance to better preserve the capabilities of 2D and 3D diffusion models. As a result, our Sherpa3D is aware of the geometric consistency with rich details and generalizes well across diverse text prompts. Extensive experiments verify the efficacy of our framework and show that our Sherpa3D outperforms existing methods for high-fidelity and geometric consistency (see qualitative results gallery in Figure 1 and quantitative results in Table 2).", + "bbox": [ + 496, + 198, + 892, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 619, + 640, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-image Generation", + "text_level": 1, + "bbox": [ + 500, + 645, + 733, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, text-to-image models such as unCLIP [59], Imagegen [61], and Stable Diffusion [60] have shown remarkable capability of generating high-quality and creative images given text prompts. Such significant progress is powered by advances in diffusion models [13, 24, 53, 69], which can be pre-trained on billions of image-text pairs [62, 64] and understands general objects with complex semantic concepts (nouns, artistic styles, etc.) [60]. Despite the great success of photorealistic and diversified image generation, using language to generate different viewpoints of the same object with 3D coherence remains a challenging problem [75].", + "bbox": [ + 496, + 669, + 890, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 500, + 845, + 709, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on promising text-to-image diffusion models, there has been a surge of studies in text-to-3D generation.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20764", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is non-trivial due to the scarcity of diverse 3D data [8, 12, 77] compared to 2D. Existing 3D native diffusion models [20, 28, 43, 52, 80, 83] usually work on a limited object category and struggle with generating in-the-wild 3D assets. To achieve generalizable 3D generation, pioneering works DreamFusion [56] and SJC [72] propose to distill the score of image distribution from pre-trained 2D diffusion models [60, 61] and show impressive results. Following works [10, 26, 36, 37, 47, 70, 71, 74, 79, 84] continue to enhance various aspects such as generation fidelity and optimization stability or explore more application scenarios [58, 68, 85]. As it is inherently ambiguous to lift 2D observations into 3D, they may suffer from multi-face issues. Although some methods use prompt engineering [4] or train a costly viewpoint-aware model [40, 67] to alleviate such problems, they fail to generate high-quality results [10] or easily overfit to domain-specific data [12, 67]. In this work, we bridge the gap between 3D and 2D diffusion models through meticulously designed 3D guidance, which leads the 2D lifting process to achieve high-fidelity, diversified, and coherent 3D generation.", + "bbox": [ + 76, + 90, + 472, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 3D Generative Models", + "text_level": 1, + "bbox": [ + 76, + 416, + 284, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive research has been conducted in the field of 3D generative modeling, exploring diverse 3D representations like 3D voxel grids [15, 22, 44], point clouds [3, 45, 49], and meshes [16, 82]. The majority of these approaches rely on training data presented in the form of 3D assets, which proves challenging to obtain at a large scale. Drawing inspiration from the success of neural volume rendering, recent studies have shifted towards investing in 3D-aware image synthesis [6, 7, 18, 21, 54, 63]. This approach offers the advantage of directly learning 3D generative models from images. However, volume rendering networks typically exhibit slow querying speeds, resulting in a trade-off between extended training times and a lack of multi-view consistency. Recently, benefitted from 2D diffusion models, some works generate multi-view images with single-view input [39-42, 66, 78]. As one of the pioneering works, Zero-1-to-3 [40] uses a synthetic dataset to fine-tune the pretrained diffusion models, aiming to learn controls of the relative camera viewpoint. Beyond Zero-1-to-3, SyncDreamer [41] employs a synchronized multiview diffusion model to capture the joint probability distribution of multiview images. This model facilitates the generation of multiview-consistent images through a unified reverse process. Different from these methods, we focus on text-to-3D synthesis, with the goal of generating multi-view consistent 3D contents with text prompts.", + "bbox": [ + 76, + 439, + 472, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 844, + 166, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a text prompt, our goal is to generate 3D assets with high quality, generalizability, and multi-view consistency.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework can be divided into three stages: (1) build coarse 3D prior from the 3D diffusion model (Sec. 3.2); (2) formulate two guiding strategies (e.g., structural and semantic guidance) for 2D lifting process (Sec. 3.3); (3) incorporate both 3D guidance and SDS loss with an annealing technique in optimization and generate the final 3D object (Sec. 3.4). In this way, we can leverage the full power of state-of-the-art 3D and 2D diffusion models to obtain 3D coherence as 3D models, retaining intricate details and creative freedom as 2D models. Our pipeline is depicted in Figure 2. Before introducing our Sherpa3D in detail, we first review the theory of Score Distillation Sampling (SDS).", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 281, + 640, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS). As one of the most representative 2D lifting methods, Dreamfusion [56] first presents the concept of Score Distillation Sampling (SDS), which is an algorithm to optimize a 3D representation such that the image rendered from any view maintains a high likelihood as evaluated by the 2D diffusion model given text prompts. SDS consists of two key components: (1) a 3D representation with parameter $\\theta$ , which can produce an image $x$ at desired camera $\\mathbf{c}$ through a parametric function $\\mathbf{x} = g(\\theta; \\mathbf{c})$ ; (2) a pre-trained text-to-image 2D diffusion model $\\phi$ with a score function $\\epsilon_{\\phi}(\\mathbf{x}_t; y, t)$ that predicts the sample noise $\\epsilon$ given noisy image $\\mathbf{x}_t$ , noise level $t$ and text embedding $y$ . The score function guides the direction of the gradient for updating $\\theta$ to reside rendered images in high-density areas conditioned on text $y$ . The gradient is calculated by SDS as:", + "bbox": [ + 496, + 303, + 893, + 546 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\mathbf {x} _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 554, + 890, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w(t)$ is a weighting function. In practice, the denoising score function $\\epsilon_{\\phi}$ is often replaced with another function $\\tilde{\\epsilon}_{\\phi}$ that uses classifier-free guidance [23] that controls the strength of the text condition (see Supplementary).", + "bbox": [ + 496, + 595, + 893, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sculpting a Coarse 3D Prior", + "text_level": 1, + "bbox": [ + 500, + 665, + 754, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate text-to-3D generation, most existing methods [26, 47, 56] rely on implicit 3D representations such as Neural Radiance Fields (NeRF) [48] and its variants [5, 51]. However, it is difficult for NeRF-based modeling to extract the high-quality surface with material and texture [73]. To address this, we adopt the hybrid scene representation of DMTet [65], including a deformable tetrahedral grid that encodes a signed distance function (SDF) and a differentiable marching tetrahedra (MT) layer that extracts explicit surface mesh. Equipped with the hybrid representation, we sculpt a coarse 3D prior from 3D diffusion model $G_{3D}$ (e.g., Shap-E [28]) by the following procedure. Given text prompts $y$ , we first use the 3D diffusion model $G_{3D}$ to generate 3D results $M_0$ and employ multi-layer perceptions", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20765", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5d0d0087271a6944901772de5dcc1d22d6223fdaabab22d3023c32bf817b304.jpg", + "image_caption": [ + "Figure 2. Pipeline of our Sherpa3D. Given a text as input, we first prompt 3D diffusion to build a coarse 3D prior $M$ encoded in the geometry model (e.g., DMTet). Next, we render the normal map of the extracted mesh in DMTet and derive two guiding strategies from $M$ . (a) Structural Guidance: we utilize the structural descriptor to compute salient geometric features for preserving geometry fidelity (e.g., without a pockmarked face problem). (b) Semantic Guidance: we leverage a semantic encoder (e.g., CLIP) to extract high-level information for keeping 3D consistency (e.g., without multi-face issues). Employing the two guidance in 2D lifting process, we use the normal map as shape encoding of the 2D diffusion model and unleash its power to generate high-quality and diversified results with 3D coherence. Then we achieve the final 3D results via photorealistic rendering through appearance modeling. (\"Everest's summit eludes many without Sherpa.\")" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(MLPs) to query SDF values for each vertex along a regular grid. Next we sample a point set $\\mathcal{P} = \\{\\pmb {p}_i\\in \\mathbb{R}^3\\}$ from $M_0$ with their SDF values $\\{SDF(\\pmb {p}_i)\\}$ . For each $\\pmb{p}_i$ , the DMTet network $\\mathcal{F}$ can predict SDF value $s(\\pmb {p}_i)$ , and a position offset $\\Delta p_{i}$ by:", + "bbox": [ + 76, + 449, + 470, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(s \\left(\\boldsymbol {p} _ {i}\\right), \\Delta \\boldsymbol {p} _ {i}\\right) = \\mathcal {F} \\left(\\boldsymbol {p} _ {i}; \\theta\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 537, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ is the parameters of network $\\mathcal{F}$ . Then, we incorporate 3D priors into the DMTet network $\\mathcal{F}_{\\theta}$ with the point set derived from 3D diffusion model by minimizing:", + "bbox": [ + 76, + 568, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D F} = \\sum_ {\\boldsymbol {p} _ {i} \\in \\mathcal {P}} | s (\\boldsymbol {p} _ {i}) - S D F (\\boldsymbol {p} _ {i}) | ^ {2} + \\lambda_ {d e f} \\sum_ {\\boldsymbol {p} _ {i} \\in \\mathcal {P}} \\| \\Delta \\boldsymbol {p} _ {i} \\| _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 625, + 468, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{def}$ is the hyperparameter controlling $L_{2}$ regularization strengths on offsets to avoid artifacts. Finally, we apply the MT layer to extract mesh representation $M$ . Now, we have leveraged the knowledge from the 3D diffusion model to construct a coarse 3D prior, which is encoded implicitly in DMTet $\\mathcal{F}_{\\theta}$ and represented explicitly by mesh $M$ . Next, we will discuss how to utilize the coarse 3D prior $M$ as a guide during the subsequent 2D diffusion lifting optimization to refine a high-quality result with 3D coherence.", + "bbox": [ + 75, + 667, + 468, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 3D Guidance for 2D Lifting Optimization", + "text_level": 1, + "bbox": [ + 76, + 816, + 431, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What knowledge can serve as guidance? The purpose of introducing a 3D prior as guidance is to address the prevalent issue of viewpoint inconsistency both in geometry and appearance. Through empirical studies, we have identified", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "geometric inconsistency as the main cause of 3D incoherence, leading to multi-face Janus problem [35, 67]. In contrast, appearance inconsistency emerges in much more extreme scenarios with lesser significance. Therefore, we disentangle the geometry from the 3D model and fully leverage coarse prior $M$ to guide 2D lifting geometry optimization with view-point awareness. Our analysis of the coarse 3D prior indicates that it contains the essential geometric structures and captures the basic categorical attributes, keeping semantic rationality across different views. Building upon these observations, a natural insight is to preserve such inherent 3D knowledge as guidance and continuously benefit the 2D lifting process. For example, given text prompts \"a head of the Terracotta Army,\" we hope the knowledge in the guidance can prevent issues such as a pockmarked face or the unrealistic scenario of having a face on the back (e.g., Janus problem). To this end, we have designed two guiding strategies derived from $M$ : structural guidance for geometric fidelity and semantic guidance for 3D coherence.", + "bbox": [ + 496, + 449, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structural guidance. Given the current DMTet net $\\mathcal{F}$ with parameters $\\theta$ that encodes the coarse 3D prior $M$ , we apply a differentiable render $f_{n}$ (e.g., nvidiffrast [32]) to generate a set of normal maps $\\mathcal{N} = \\{\\pmb{n}_i|\\pmb{n}_i = f_n(\\mathcal{F}_{\\theta},\\mathbf{c}_i),i = 1,\\dots n\\}$ , where $\\mathbf{c}_i$ is the camera position randomly sampled in spherical coordinates. To extract the salient geometric structure features, we first use a Gaussian filter with a kernel standard deviation $\\sigma$", + "bbox": [ + 496, + 739, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (x, y) = \\frac {1}{2 \\pi \\sigma^ {2}} e ^ {- \\frac {x ^ {2} + y ^ {2}}{2 \\sigma^ {2}}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 875, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20766", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to reduce the noise impact and obtain $\\{\\sigma (\\pmb {n}_i)\\}$ . As gradients are simple but effective tools for revealing the geometric contours and salient structures [14, 29], we then compute the structural descriptor sets $\\{G_{\\sigma}(\\pmb {n}_i)\\}$ by", + "bbox": [ + 75, + 90, + 468, + 154 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\sigma} (\\boldsymbol {n} _ {i}) = \\sqrt {(\\frac {\\partial \\sigma (\\boldsymbol {n} _ {i})}{\\partial x}) ^ {2} + (\\frac {\\partial \\sigma (\\boldsymbol {n} _ {i})}{\\partial y}) ^ {2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 164, + 468, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ and $y$ are the coordinate directions of the normal map $\\pmb{n}_i$ . Throughout the 2D lifting process of updating $\\mathcal{F}_{\\theta}$ with newly rendered normal maps $\\tilde{\\mathcal{N}} = \\{\\tilde{n}_i\\}$ , it should follow the structural guidance as:", + "bbox": [ + 76, + 215, + 470, + 279 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} _ {\\text {s t r u c}} := \\sum_ {i = 1} ^ {n} \\left\\| G _ {\\sigma} \\left(\\boldsymbol {n} _ {i}\\right) - G _ {\\sigma} \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right) \\right\\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 289, + 468, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which enables the 2D lifting process to preserve geometric fidelity and a well-aligned structure with the coarse 3D prior when generating rich details.", + "bbox": [ + 76, + 340, + 468, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic guidance. While structural guidance maintains low-level geometric perception from coarse 3D prior, semantic guidance extracts high-level features for 3D coherence. We first apply the pre-trained CLIP [57] model as semantic encoder $\\psi$ to the normal set $\\mathcal{N}$ and obtain semantic feature maps $\\mathcal{N}_c = \\{\\psi(\\boldsymbol{n}_i)\\}$ , proven to effectively capture semantic attributes like facial expressions or view categories [17]. Following the notation as above, we then define the semantic guidance with cosine similarity:", + "bbox": [ + 75, + 386, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} _ {\\text {s e m}} := \\sum_ {i = 1} ^ {n} \\frac {\\psi \\left(\\boldsymbol {n} _ {i}\\right) \\cdot \\psi \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right)}{\\| \\psi \\left(\\boldsymbol {n} _ {i}\\right) \\| \\| \\psi \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right) \\|.} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 535, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Employing this guidance, we ensure that different views retain inherent high-level information throughout the 2D lifting optimization process. Experiments show that it can effectively mitigate multi-face problems, keeping 3D content semantically plausible from all viewing angles.", + "bbox": [ + 76, + 585, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 76, + 672, + 217, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we incorporate both structural and semantic guidance derived from coarse 3D prior to 2D lifting optimization so that it can produce vivid and diversified objects with multi-view consistency. For the disentangled geometry modeling, we use the randomly sampled normal map $\\pmb{n}$ as the input, bridging the gap between 3D and 2D diffusion. To update the geometry model DMTet network $\\mathcal{F}_{\\theta}$ , we choose to use the publicly available Stable Diffusion [60] as pre-trained 2D diffusion model $\\phi$ and compute the gradient of the SDS loss similar in Eq. 1:", + "bbox": [ + 75, + 696, + 468, + 848 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\theta , \\boldsymbol {n}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\boldsymbol {z} _ {t} ^ {\\boldsymbol {n}}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\boldsymbol {z} _ {t} ^ {\\boldsymbol {n}}}{\\partial \\boldsymbol {n}} \\frac {\\partial \\boldsymbol {n}}{\\partial \\theta} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 857, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\partial z_t^n /\\partial n$ calculates the gradient of the encoder in the latent diffusion model (LDM) [60]. Additionally, we introduce a step annealing technique to balance the influence of the 3D guidance during 2D lifting optimization:", + "bbox": [ + 496, + 90, + 893, + 152 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma (\\lambda) = \\lambda e ^ {- \\beta \\max (0, n _ {\\mathrm {c u r}} - m)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 157, + 890, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_{\\mathrm{cur}}$ is the current epoch and $\\{\\beta, m, \\lambda\\}$ are the hyperparameters that control how $\\gamma$ decreased. Therefore, the total loss $\\mathcal{L}_{\\mathrm{geo}}$ to lift 2D geometry optimization with 3D guidance is a weighted sum of three loss terms:", + "bbox": [ + 496, + 183, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g e o}} (\\theta , \\boldsymbol {n}) = \\mathcal {L} _ {\\mathrm {S D S}} + \\gamma \\left(\\lambda_ {\\mathrm {s t r u c}}\\right) \\mathcal {L} _ {\\mathrm {s t r u c}} + \\gamma \\left(\\lambda_ {\\mathrm {s e m}}\\right) \\mathcal {L} _ {\\mathrm {s e m}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 253, + 890, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which not only enables the 3D content generation without multi-view inconsistency issues but also preserves the generalization and quality in 2D diffusion model $\\phi$ . As our pipeline can be integrated into any appearance model [9, 10, 33], we adopt a similar approach as Fantasia3D [10] to better align our text and 3D object. Denote $\\mathcal{T}$ with parameters $\\eta$ as our appearance model, we have the rendered image $\\mathbf{x} = \\mathcal{T}_{\\eta}(\\mathcal{F}_{\\theta}, c_{i})$ . To update $\\eta$ , we again apply the SDS loss for the final complete generated 3D object with detailed texture and coherent geometry:", + "bbox": [ + 496, + 277, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\eta} \\mathcal {L} _ {\\mathrm {a p p}} (\\eta , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\boldsymbol {z} _ {t} ^ {\\mathbf {x}}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\boldsymbol {z} _ {t} ^ {\\mathbf {x}}}{\\partial \\mathbf {x}} \\frac {\\partial \\mathbf {x}}{\\partial \\eta} \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 434, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which shares similar notations defined in Eq. 8. Finally, through the tailored 3D structural and semantic guidance that bridges the 2D and 3D diffusion models, our Sherpa3D can mitigate the multi-face problem and achieve high-fidelity and diversified results.", + "bbox": [ + 496, + 469, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 551, + 718, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply the multilayer perceptron (MLP) comprising of three hidden layers to approximate $\\mathcal{F}_{\\theta}$ and $\\mathcal{T}_{\\eta}$ . Adam optimizer [30] is used to update $\\mathcal{F}_{\\theta}$ and $\\mathcal{T}_{\\eta}$ with an initial learning rates of $1e - 3$ decaying into $5e - 4$ . For 3D representations, we use textured mesh with a DMTet resolution of 128 to achieve a balance between quality and generation speed. We sample random camera poses at a fixed radius of 2.5, y-axis FOV of $45^{\\circ}$ , with the azimuth in $[-180^{\\circ}, 180^{\\circ}]$ and elevation in $[-30^{\\circ}, 30^{\\circ}]$ . We load Shap-E from [46] for 3D diffusion model and choose stabilityai/stable-diffsuion-2-1-base [60] for 2D diffusion model. For weighting factors, we follow the same strategy as [26] to tune $w(t)$ . $\\lambda_{\\mathrm{struc}}$ is set to 10 and $\\lambda_{\\mathrm{sem}}$ is 30 to balance the magnitude of SDS loss. Notably, our method only needs a single NVIDIA RTX3090 (24GB) GPU within 25 minutes. More details of optimization, architecture design, and hyperparameter settings can be found in the supplementary.", + "bbox": [ + 496, + 575, + 890, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 844, + 633, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive experiments to evaluate our text-to-3D framework Sherpa3D and show", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20767", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d6ff6248e87c28dde9bb2fbbea9a9faedaab9b7317e158a271c995b649fc3b9.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons with baseline methods across different views $(0^{\\circ}$ and $180^{\\circ})$ . We can observe that baseline methods suffer from severe multi-face issues while our Sherpa3D can achieve better quality and 3D coherence." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 851, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comparison results against other text-to-3D baseline methods. We first present qualitative results compared with five SOTA baselines from different viewpoints. Then we report the quantitative results with a user study. Finally, we carry out ablation studies to further verify the efficacy of our framework design. Please refer to the supplementary for more comparisons, visualizations, and detailed analysis.", + "bbox": [ + 75, + 590, + 472, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 709, + 253, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We extensively compare our method Sherpa3D against five baselines: Shap-E [28], DreamFusion [56], Magic3D [37], ProlificDreamer [74], and Fantasis3D [10]. Due to various reasons, we can't obtain the original implementation of some baselines. For DreamFusion, Magic3D, and ProlificDreamer, we utilize their implementations in the Threestudio library [19] for comparison. For Shap-E and Fantasia3D, we follow their official implementation. We consider these implementations to be the most reliable and comprehensive open-source option available in the field. To ensure a fair comparison, we use the Stable Diffusion [60]", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model as 2D diffusion prior by default.", + "bbox": [ + 498, + 590, + 759, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We will show our results with notable comparisons to other baselines through visualization. As there is no Ground-Truth 3D content corresponding to the text prompt, reference-based metrics like Chamfer Distance are difficult to apply to zero-shot text-to-3D generation. Following [27, 56], we evaluate the CLIP R-Precision [55], which can measure how well the rendered images of generated 3D content align with the input text. We use 100 prompts from the Common Objects in Context (COCO) dataset [38] as DreamFusion [56]. we also conduct a user study to further demonstrate the multi-view consistency and overall generation quality of our method,", + "bbox": [ + 496, + 607, + 890, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 498, + 800, + 728, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first demonstrate vivid and diversified text-to-3D results generated from our Sherpa3D in the gallery as shown in Figure 1. Then we compare our method with five baseline method: Shap-E [28], DreamFusion [56], Magic3D [37], Fantasia3D [10] and ProlificDreamer [74]. Figure 3 and 4", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20768", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32c1b7f6ece146318a53c2deb9f2af304970f0c375b5b1d6dc48b7a188d45f94.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons with baseline methods across different views $(-30^{\\circ}$ and $150^{\\circ})$" + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 844, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "give the comparative results with the same text prompt for each object generation. We observe that the Shap-E [28] only generates coarse shapes while other 2D lifting methods suffer from multi-face problems. In contrast, our Sherpa3D produces high-fidelity 3D assets with compelling texture quality and multi-view consistency. Notably, our framework is more efficient than other baselines with less time to optimize. Specifically, it only takes within 25 minutes from a text prompt to a high-quality 3D model ready to be used in graphic engines.", + "bbox": [ + 75, + 563, + 472, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 724, + 316, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we report the CLIP R-Precision for Sherpa3D and several baselines. It shows that our method outperforms other baselines consistently across different CLIP models, and approaches the performance of ground truth (GT) images. For the user study, we render 360-degree rotating videos of 3D models generated from a collection of 120 images. Each volunteer is shown 10 samples of rendered video from a random method and rates in two aspects: multi-view consistency and overall generation quality. We collect results from 50 volunteers shown in Table 2. We observe", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9502c457b05f78d972e5b140b2e1a61f85941e2fd328d7e8fdbbce9bcaef2dcd.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on generation renderings with text prompts using different CLIP retrieval models. We compared to ground-truth images, Shap-E [28], Dreamfusion [56], Magic3D [37], evaluated on object-centric COCO as in [56]." + ], + "table_footnote": [], + "table_body": "
MethodR-Precision (%) ↑
CLIP B/32CLIP B/16CLIP L/14
GT Images77.379.2-
Shape-E [28]41.142.546.4
DreamFusion [56]70.373.275.0
Magic3D [37]71.573.876.1
Sherpa3D (Ours)72.375.679.3
", + "bbox": [ + 504, + 627, + 890, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that most users consider our results with much higher viewpoints consistency and overall generation fidelity.", + "bbox": [ + 498, + 760, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 500, + 800, + 756, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We carry out ablation studies on the design of our Sherpa3D framework in Figure 5 using an example text prompt \"a head of the Terracotta Army\". Specifically, we perform ablation on three aspects of our method: structural guidance, semantic guidance, and the step annealing strategy. The re", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20769", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0621e3cec3b6f38e082b6003b1dbebfc7010780acae19a872404a42fe7c600af.jpg", + "image_caption": [ + "$0^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 109, + 94, + 169, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d532502f20d283fa6c957c729d2c36f6976412953337cb654564a775c702abe.jpg", + "image_caption": [ + "$180^{\\circ}$", + "Figure 5. Ablation study of our method. The generation is based on the text prompt \"a head of the Terracotta Army\". We ablate the design choices of structural guidance, semantic guidance (Sec. 3.3), and the step annealing technique (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 106, + 185, + 176, + 266 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a6437b306a74baebd0daaae842bd2fa2abf600edf6987a693fff13a3a4888f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 93, + 263, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ce9efb194bca24646ce7a8547a9b70ce0084f61b01cb523fec302eb84c074da.jpg", + "image_caption": [ + "Full" + ], + "image_footnote": [], + "bbox": [ + 205, + 185, + 277, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bb268e3aadc1079fc2d41590fc43c7fca663e82448599b1fa7ffc08dadb9c02.jpg", + "image_caption": [ + "appearance" + ], + "image_footnote": [], + "bbox": [ + 310, + 104, + 377, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acd3e382ceab519f97c67094d081af11890c03013593f8301aa21ed422821d4a.jpg", + "image_caption": [ + "w/o Structural Guidance" + ], + "image_footnote": [], + "bbox": [ + 310, + 186, + 380, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a55cbdabe9a2eaa10b3612361317d091ff3b53cb162617d7fdea7b64dedcdd37.jpg", + "image_caption": [ + "geometry" + ], + "image_footnote": [], + "bbox": [ + 416, + 104, + 480, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bcce458f5808737039f5f7d0e383faefc4e04ac16487e2a694dd807de118ff2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 190, + 483, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/515e8be1d68da195e576a82d90e9cb58f8cc7993bf98f15a7a12ae07292f0239.jpg", + "image_caption": [ + "appearance" + ], + "image_footnote": [], + "bbox": [ + 516, + 104, + 584, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6efdc9de646c27c0e24b3715d1b810ecfc6a472c9cff292182dac52245b187e3.jpg", + "image_caption": [ + "w/o Semantic Guidance" + ], + "image_footnote": [], + "bbox": [ + 513, + 184, + 589, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c12160555f8f2eaf666930b4047a12b562db56ee815e0f7c5abb4987d0360254.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 94, + 669, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9391c2def14aaca71bd70c965cbd8fa888490440fd2006fade33619100de7ff9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 184, + 691, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b17721830ddb10af0cbe66dcfe5c8b1ed558b00a902d014c09a9a8a4049f6607.jpg", + "image_caption": [ + "appearance" + ], + "image_footnote": [], + "bbox": [ + 725, + 104, + 784, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f2b55d271bc0fe6744cc06b0b829d23a76c39101a7a377156d8953e6de1b03c.jpg", + "image_caption": [ + "w/o step annealing" + ], + "image_footnote": [], + "bbox": [ + 722, + 185, + 785, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d030b334723685502f8e5c2c480fc47c1aed397a586e49831506a498632d2e05.jpg", + "image_caption": [ + "geometry" + ], + "image_footnote": [], + "bbox": [ + 826, + 104, + 885, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/588137a4dabf0da9dbd143790d45e7cfbef9055cf99cbe9b0eabe14661580629.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 184, + 890, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cf47d90964a50a1ed764411d3af336145bad1ddf546b477fa5112aed246d7530.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on the multi-view consistency and overall generation quality score in a user study, rated on a scale of 1-10, with higher scores indicating better performance." + ], + "table_footnote": [], + "table_body": "
MethodMulti-view Consistency ↑Overall Quality ↑
Shap-E [28]6.093.33
DreamFusion [56]4.586.05
Magic3D [37]5.256.73
Fantasia3D [10]3.835.90
ProlificDreamer [74]5.787.02
Sherpa3D(Ours)8.958.74
", + "bbox": [ + 81, + 398, + 468, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sults reveal that the omission of any of these elements leads to a degradation in terms of quality and consistency. Notably, the absence of structural guidance leads to a loss of geometric fidelity in the \"army\", leading to a pockmarked face; without semantic guidance, there's a loss of semantic rationality across different views, resulting in the multiview Janus problem. The lack of a balanced step annealing results in an excessive influence of guidance with a rough final output. This illustrates the effectiveness of our overall framework (Figure 2), which drives geometric fidelity, multi-view consistency, and optimization balance steered by the 3D guidance and annealing strategy.", + "bbox": [ + 75, + 513, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further demonstrate our generalizability, we compare our method in Figure 6 with the Zero123 [40] which uses more 3D data [40] to finetune a 2D diffusion model to be viewpoint-aware. However, such a finetuning-based method easily overfits to 3D training data and suffers from severe performance degradation with unseen input of the training set. In contrast, our method is more generalizable to open-vocabulary text prompts.", + "bbox": [ + 75, + 696, + 468, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 829, + 194, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Sherpa3D, a new framework that simultaneously achieves high-quality, diversified, and 3D consistent text-to-3D generation. By fully exploiting eas", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c0d347e447eabfbaaf36cf29e02ba7e20bba97abe3b98204bc6c0a5a9e62498.jpg", + "image_caption": [ + "Sherpa3D (Ours)", + "\"A deer draped in exquisite ancient Chinese fabric, oil painting style\"", + "Figure 6. Comparison with Zero123 [40]. We use the front view of our generated 3D model as the input of Zero123 with open-vocabulary text prompts." + ], + "image_footnote": [], + "bbox": [ + 501, + 358, + 581, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc971454719d6accdf8ac61ae0a3297419f96e83ef13edeb02b77d1da4e55052.jpg", + "image_caption": [ + "Zero123" + ], + "image_footnote": [], + "bbox": [ + 607, + 361, + 691, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8b4c2d0b228ecdbd7e2b490f92777f8f240dd187890f4ab26edc4d6815366a6.jpg", + "image_caption": [ + "Sherpa3D (Ours)", + "\"A head of the Terracotta Army\"" + ], + "image_footnote": [], + "bbox": [ + 723, + 362, + 774, + 424 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bb17022eb364810a80b375f803a624dc71d37f5f790542a481ec2cf325097dd.jpg", + "image_caption": [ + "Zero123" + ], + "image_footnote": [], + "bbox": [ + 813, + 373, + 890, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ily obtained coarse 3D knowledge from the 3D diffusion model, we derive structural guidance and semantic guidance to enhance the prompts and provide continuous guidance with geometric fidelity and 3D coherence throughout the 2D lifting optimization. To further improve the overall performance, we incorporate a step annealing strategy that modulates the impact of 3D guidance and 2D refinement. Therefore, our framework bridges the gap between 2D and 3D diffusion models, preserving multi-view coherent generation while maintaining the generalizability and fidelity of 2D models. Extensive experiments verify the remarkable improvement of our Sherpa3D on text-to-3D generation.", + "bbox": [ + 496, + 522, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future works. Although our Sherpa3D achieves remarkable text-to-3D results, the quality still seems to be limited to the backbone itself as we choose Shap-E [28] and Stable Diffusion v2.1 base model in this work. We expect them to be solved with a larger diffusion model, such as SDXL [1] and DeepFloyd [2]. In future work, we are interested in extending our insight to more creative text-to-4D generation. We believe that Sherpa3D provides a promising research path for user-friendly and more accessible 3D content creation.", + "bbox": [ + 496, + 704, + 890, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant 62206147, and in part by Deng Feng Fund.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20770", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] stable-diffusion-xl-base-1.0. https://huggingface.co/stabilityai/stable-diffusion-x1-base-1.0. Accessed:2023-08-29.8", + "[2] Deepfloyd. https://huggingface.co/DeepFloyd. Accessed: 2023-08-25. 8", + "[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, pages 40-49. PMLR, 2018. 3", + "[4] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 3", + "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3", + "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 3", + "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3", + "[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3", + "[9] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023.5", + "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 2, 3, 5, 6, 8", + "[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 2", + "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5", + "[15] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 3DV, pages 402-411. IEEE, 2017. 3", + "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurlPS, 35:31841-31854, 2022. 3", + "[17] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 6(3):e30, 2021. 5", + "[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3", + "[19] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6", + "[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 2, 3", + "[21] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Gancraft: Unsupervised 3d neural rendering of northwest world. In ICCV, pages 14072-14082, 2021. 3", + "[22] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In ICCV, pages 9984-9993, 2019. 3", + "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[25] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989, 2023. 2", + "[26] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 5", + "[27] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2, 6", + "[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6, 7, 8", + "[29] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20771", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[31] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6:124, 2011. 2", + "[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4", + "[33] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 5", + "[34] Chenghao Li, Chaoning Zhang, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2", + "[35] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2,4", + "[36] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3", + "[37] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 300–309, 2023. 2, 3, 6, 7, 8", + "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[39] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 3", + "[40] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2, 3, 8", + "[41] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3", + "[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Sin-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gle image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 3", + "[43] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 3", + "[44] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3", + "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 3", + "[46] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5", + "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 3", + "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 3", + "[49] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 3", + "[50] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2", + "[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3", + "[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2, 3", + "[53] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[54] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, pages 13503–13513, 2022. 3", + "[55] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6", + "[56] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20772", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Eleventh International Conference on Learning Representations (ICLR), 2023. 2, 3, 6, 7, 8", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[58] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3", + "[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3, 5, 6", + "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2, 3", + "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[63] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. NeurlPS, 35:33999-34011, 2022. 3", + "[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2", + "[65] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3", + "[66] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3", + "[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4", + "[68] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea" + ], + "bbox": [ + 78, + 92, + 468, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3", + "[69] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3", + "[71] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3", + "[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 2, 3", + "[73] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3", + "[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2, 3, 6, 8", + "[75] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2", + "[76] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2", + "[77] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 3", + "[78] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 3", + "[79] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6841–6850, 2023. 3", + "[80] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3", + "[81] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20773", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. arXiv preprint arXiv:2010.09125, 2020. 3", + "[83] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461, 2023. 3", + "[84] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 3", + "[85] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 303 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20774", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_model.json b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7faf358aeb5eeac59bd7ff5ed846ba477cba31f --- /dev/null +++ b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_model.json @@ -0,0 +1,2699 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.878, + 0.153 + ], + "angle": 0, + "content": "Sherpa3D: Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.756, + 0.218 + ], + "angle": 0, + "content": "Fangfu Liu\\(^{1}\\), Diankun Wu\\(^{1}\\), Yi Wei\\(^{1}\\), Yongming Rao\\(^{2}\\), Yueqi Duan\\(^{1\\dagger}\\), Tsinghua University, BAAI" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.227, + 0.89, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.689, + 0.893, + 0.719 + ], + "angle": 0, + "content": "Figure 1. Gallery of Sherpa3D: Blender rendering for various textured meshes from Sherpa3D, which is able to generate high-fidelity, diverse, and multi-view consistent 3D contents with input text prompts. Our method is also compatible with popular graphics engines." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.737, + 0.314, + 0.753 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.471, + 0.877 + ], + "angle": 0, + "content": "Recently, 3D content creation from text prompts has demonstrated remarkable progress by utilizing 2D and 3D diffusion models. While 3D diffusion models ensure great multi-view consistency, their ability to generate high-quality and diverse 3D assets is hindered by the limited 3D data. In contrast, 2D diffusion models find a distillation approach that achieves excellent generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.895, + 0.892 + ], + "angle": 0, + "content": "and rich details without any 3D data. However, 2D lifting methods suffer from inherent view-agnostic ambiguity thereby leading to serious multi-face Janus issues, where text prompts fail to provide sufficient guidance to learn coherent 3D results. Instead of retraining a costly viewpoint-aware model, we study how to fully exploit easily accessible coarse 3D knowledge to enhance the prompts and guide 2D lifting optimization for refinement. In this paper, we propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.889, + 0.229, + 0.902 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20763" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "consistency simultaneously. Specifically, we design a pair of guiding strategies derived from the coarse 3D prior generated by the 3D diffusion model: a structural guidance for geometric fidelity and a semantic guidance for 3D coherence. Employing the two types of guidance, the 2D diffusion model enriches the 3D content with diversified and high-quality results. Extensive experiments show the superiority of our Sherpa3D over the state-of-the-art text-to-3D methods in terms of quality and 3D consistency. Project page: https://liuff19.github.io/Sherpa3D/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.272, + 0.21, + 0.287 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.47, + 0.448 + ], + "angle": 0, + "content": "3D content generation [34, 37, 56, 76] finds a broad range of applications, including games, movies, virtual/augmented reality and robots. However, the conventional process of creating premium 3D assets is still expensive and challenging as it requires multiple labor-intensive and time-consuming stages [31]. Fortunately, this challenge has prompted the development of recent text-to-3D methods [10, 25, 27, 35, 37, 47, 50, 56, 74]. Only using text prompts to automate 3D generation, these techniques pave a promising way towards streamlining 3D creation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.719 + ], + "angle": 0, + "content": "Powered by the great breakthroughs in diffusion models [52, 60, 61, 81], two research lines of rationalization have recently emerged in text-to-3D: inference-only 3D diffusion methods and optimization-based 2D lifting methods. Specifically, the inference-only methods [20, 28, 52] seek to directly generate 3D-consistent assets by extensively training a new diffusion model on 3D data. However, due to the scarcity of 3D datasets compared to accessible 2D images or text data, these 3D diffusion models suffer from low quality and limited generalizability. Without requiring any 3D data for training, 2D lifting methods [4, 10, 37, 47, 56, 72, 74] can produce high-quality and diversified 3D results by distilling 3D knowledge from pre-trained 2D diffusion models [60, 61, 81], also known as Score Distillation Sampling (SDS). Yet lifting 2D observations into 3D is inherently ambiguous without sufficient 3D guidance from text prompts, leading to notorious multi-view inconsistency (e.g., Janus problems) in 2D lifting methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "These findings motivate us to think: is it possible to bridge the two aforementioned streams to achieve generalizability, high-fidelity, and geometric consistency simultaneously? An intuitive idea is to leverage more 3D data [11, 12] to fine-tune a view-point aware diffusion model, but it requires substantial computational resources and is prone to overfitting due to data bias [40, 67]. In contrast, our key insight is to utilize the easily accessible 3D diffusion model as guidance and study how to fully exploit coarse 3D knowledge to guide 2D lifting optimization for refinement. In particular, when maintaining the quality and generalizability of the original 2D diffusion model, we hope" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "the 2D lifting awareness can be guided by the strong 3D geometric information from the 3D diffusion model. However, it is non-trivial in pursuit of this balance. Relying too heavily on the coarse 3D priors from the 3D diffusion model may degrade the generation quality, whereas little 3D guidance could result in a lack of geometric awareness, leading to multi-view inconsistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.606 + ], + "angle": 0, + "content": "Towards this end, we propose Sherpa3D in this paper, which greatly boosts high-fidelity and highly diversified text-to-3D generation with geometric consistency. Our method begins by employing a 3D diffusion model to craft a basic 3D guide with limited details. Building upon the coarse 3D prior, we introduce two guiding strategies to inform 2D diffusion model throughout lifting optimization: a structural guide for geometric fidelity and a semantic guide for 3D coherence. Specifically, the structural guide leverages the first-order gradient information of the normals from the 3D prior to supervise the optimization of the structure. These normals are then integrated into the input of a pre-trained 2D diffusion model, refining the geometric details. Concurrently, our semantic guide extracts high-level features from multi-views of the 3D prior. These features guide the 2D lifting optimization to perceive the geometric consistency under the preservation of original generalizability and quality. Furthermore, we design an annealing function, which modulates the influence of the 3D guidance to better preserve the capabilities of 2D and 3D diffusion models. As a result, our Sherpa3D is aware of the geometric consistency with rich details and generalizes well across diverse text prompts. Extensive experiments verify the efficacy of our framework and show that our Sherpa3D outperforms existing methods for high-fidelity and geometric consistency (see qualitative results gallery in Figure 1 and quantitative results in Table 2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.642, + 0.636 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.734, + 0.662 + ], + "angle": 0, + "content": "2.1. Text-to-image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Recently, text-to-image models such as unCLIP [59], Imagegen [61], and Stable Diffusion [60] have shown remarkable capability of generating high-quality and creative images given text prompts. Such significant progress is powered by advances in diffusion models [13, 24, 53, 69], which can be pre-trained on billions of image-text pairs [62, 64] and understands general objects with complex semantic concepts (nouns, artistic styles, etc.) [60]. Despite the great success of photorealistic and diversified image generation, using language to generate different viewpoints of the same object with 3D coherence remains a challenging problem [75]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.71, + 0.862 + ], + "angle": 0, + "content": "2.2. Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Building on promising text-to-image diffusion models, there has been a surge of studies in text-to-3D generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20764" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.41 + ], + "angle": 0, + "content": "However, it is non-trivial due to the scarcity of diverse 3D data [8, 12, 77] compared to 2D. Existing 3D native diffusion models [20, 28, 43, 52, 80, 83] usually work on a limited object category and struggle with generating in-the-wild 3D assets. To achieve generalizable 3D generation, pioneering works DreamFusion [56] and SJC [72] propose to distill the score of image distribution from pre-trained 2D diffusion models [60, 61] and show impressive results. Following works [10, 26, 36, 37, 47, 70, 71, 74, 79, 84] continue to enhance various aspects such as generation fidelity and optimization stability or explore more application scenarios [58, 68, 85]. As it is inherently ambiguous to lift 2D observations into 3D, they may suffer from multi-face issues. Although some methods use prompt engineering [4] or train a costly viewpoint-aware model [40, 67] to alleviate such problems, they fail to generate high-quality results [10] or easily overfit to domain-specific data [12, 67]. In this work, we bridge the gap between 3D and 2D diffusion models through meticulously designed 3D guidance, which leads the 2D lifting process to achieve high-fidelity, diversified, and coherent 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.417, + 0.285, + 0.431 + ], + "angle": 0, + "content": "2.3. 3D Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.473, + 0.835 + ], + "angle": 0, + "content": "Extensive research has been conducted in the field of 3D generative modeling, exploring diverse 3D representations like 3D voxel grids [15, 22, 44], point clouds [3, 45, 49], and meshes [16, 82]. The majority of these approaches rely on training data presented in the form of 3D assets, which proves challenging to obtain at a large scale. Drawing inspiration from the success of neural volume rendering, recent studies have shifted towards investing in 3D-aware image synthesis [6, 7, 18, 21, 54, 63]. This approach offers the advantage of directly learning 3D generative models from images. However, volume rendering networks typically exhibit slow querying speeds, resulting in a trade-off between extended training times and a lack of multi-view consistency. Recently, benefitted from 2D diffusion models, some works generate multi-view images with single-view input [39-42, 66, 78]. As one of the pioneering works, Zero-1-to-3 [40] uses a synthetic dataset to fine-tune the pretrained diffusion models, aiming to learn controls of the relative camera viewpoint. Beyond Zero-1-to-3, SyncDreamer [41] employs a synchronized multiview diffusion model to capture the joint probability distribution of multiview images. This model facilitates the generation of multiview-consistent images through a unified reverse process. Different from these methods, we focus on text-to-3D synthesis, with the goal of generating multi-view consistent 3D contents with text prompts." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.168, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given a text prompt, our goal is to generate 3D assets with high quality, generalizability, and multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.273 + ], + "angle": 0, + "content": "Our framework can be divided into three stages: (1) build coarse 3D prior from the 3D diffusion model (Sec. 3.2); (2) formulate two guiding strategies (e.g., structural and semantic guidance) for 2D lifting process (Sec. 3.3); (3) incorporate both 3D guidance and SDS loss with an annealing technique in optimization and generate the final 3D object (Sec. 3.4). In this way, we can leverage the full power of state-of-the-art 3D and 2D diffusion models to obtain 3D coherence as 3D models, retaining intricate details and creative freedom as 2D models. Our pipeline is depicted in Figure 2. Before introducing our Sherpa3D in detail, we first review the theory of Score Distillation Sampling (SDS)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.282, + 0.642, + 0.296 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.547 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS). As one of the most representative 2D lifting methods, Dreamfusion [56] first presents the concept of Score Distillation Sampling (SDS), which is an algorithm to optimize a 3D representation such that the image rendered from any view maintains a high likelihood as evaluated by the 2D diffusion model given text prompts. SDS consists of two key components: (1) a 3D representation with parameter \\(\\theta\\), which can produce an image \\(x\\) at desired camera \\(\\mathbf{c}\\) through a parametric function \\(\\mathbf{x} = g(\\theta; \\mathbf{c})\\); (2) a pre-trained text-to-image 2D diffusion model \\(\\phi\\) with a score function \\(\\epsilon_{\\phi}(\\mathbf{x}_t; y, t)\\) that predicts the sample noise \\(\\epsilon\\) given noisy image \\(\\mathbf{x}_t\\), noise level \\(t\\) and text embedding \\(y\\). The score function guides the direction of the gradient for updating \\(\\theta\\) to reside rendered images in high-density areas conditioned on text \\(y\\). The gradient is calculated by SDS as:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.555, + 0.892, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\mathbf {x} _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.894, + 0.659 + ], + "angle": 0, + "content": "where \\( w(t) \\) is a weighting function. In practice, the denoising score function \\( \\epsilon_{\\phi} \\) is often replaced with another function \\( \\tilde{\\epsilon}_{\\phi} \\) that uses classifier-free guidance [23] that controls the strength of the text condition (see Supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.755, + 0.683 + ], + "angle": 0, + "content": "3.2. Sculpting a Coarse 3D Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To facilitate text-to-3D generation, most existing methods [26, 47, 56] rely on implicit 3D representations such as Neural Radiance Fields (NeRF) [48] and its variants [5, 51]. However, it is difficult for NeRF-based modeling to extract the high-quality surface with material and texture [73]. To address this, we adopt the hybrid scene representation of DMTet [65], including a deformable tetrahedral grid that encodes a signed distance function (SDF) and a differentiable marching tetrahedra (MT) layer that extracts explicit surface mesh. Equipped with the hybrid representation, we sculpt a coarse 3D prior from 3D diffusion model \\( G_{3D} \\) (e.g., Shap-E [28]) by the following procedure. Given text prompts \\( y \\), we first use the 3D diffusion model \\( G_{3D} \\) to generate 3D results \\( M_0 \\) and employ multi-layer perceptions" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20765" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.896, + 0.436 + ], + "angle": 0, + "content": "Figure 2. Pipeline of our Sherpa3D. Given a text as input, we first prompt 3D diffusion to build a coarse 3D prior \\(M\\) encoded in the geometry model (e.g., DMTet). Next, we render the normal map of the extracted mesh in DMTet and derive two guiding strategies from \\(M\\). (a) Structural Guidance: we utilize the structural descriptor to compute salient geometric features for preserving geometry fidelity (e.g., without a pockmarked face problem). (b) Semantic Guidance: we leverage a semantic encoder (e.g., CLIP) to extract high-level information for keeping 3D consistency (e.g., without multi-face issues). Employing the two guidance in 2D lifting process, we use the normal map as shape encoding of the 2D diffusion model and unleash its power to generate high-quality and diversified results with 3D coherence. Then we achieve the final 3D results via photorealistic rendering through appearance modeling. (\"Everest's summit eludes many without Sherpa.\")" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.471, + 0.526 + ], + "angle": 0, + "content": "(MLPs) to query SDF values for each vertex along a regular grid. Next we sample a point set \\(\\mathcal{P} = \\{\\pmb {p}_i\\in \\mathbb{R}^3\\}\\) from \\(M_0\\) with their SDF values \\(\\{SDF(\\pmb {p}_i)\\}\\). For each \\(\\pmb{p}_i\\), the DMTet network \\(\\mathcal{F}\\) can predict SDF value \\(s(\\pmb {p}_i)\\), and a position offset \\(\\Delta p_{i}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.539, + 0.47, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\left(s \\left(\\boldsymbol {p} _ {i}\\right), \\Delta \\boldsymbol {p} _ {i}\\right) = \\mathcal {F} \\left(\\boldsymbol {p} _ {i}; \\theta\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.47, + 0.615 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the parameters of network \\(\\mathcal{F}\\). Then, we incorporate 3D priors into the DMTet network \\(\\mathcal{F}_{\\theta}\\) with the point set derived from 3D diffusion model by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.626, + 0.47, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D F} = \\sum_ {\\boldsymbol {p} _ {i} \\in \\mathcal {P}} | s (\\boldsymbol {p} _ {i}) - S D F (\\boldsymbol {p} _ {i}) | ^ {2} + \\lambda_ {d e f} \\sum_ {\\boldsymbol {p} _ {i} \\in \\mathcal {P}} \\| \\Delta \\boldsymbol {p} _ {i} \\| _ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.804 + ], + "angle": 0, + "content": "where \\(\\lambda_{def}\\) is the hyperparameter controlling \\(L_{2}\\) regularization strengths on offsets to avoid artifacts. Finally, we apply the MT layer to extract mesh representation \\(M\\). Now, we have leveraged the knowledge from the 3D diffusion model to construct a coarse 3D prior, which is encoded implicitly in DMTet \\(\\mathcal{F}_{\\theta}\\) and represented explicitly by mesh \\(M\\). Next, we will discuss how to utilize the coarse 3D prior \\(M\\) as a guide during the subsequent 2D diffusion lifting optimization to refine a high-quality result with 3D coherence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.433, + 0.833 + ], + "angle": 0, + "content": "3.3. 3D Guidance for 2D Lifting Optimization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "What knowledge can serve as guidance? The purpose of introducing a 3D prior as guidance is to address the prevalent issue of viewpoint inconsistency both in geometry and appearance. Through empirical studies, we have identified" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.736 + ], + "angle": 0, + "content": "geometric inconsistency as the main cause of 3D incoherence, leading to multi-face Janus problem [35, 67]. In contrast, appearance inconsistency emerges in much more extreme scenarios with lesser significance. Therefore, we disentangle the geometry from the 3D model and fully leverage coarse prior \\( M \\) to guide 2D lifting geometry optimization with view-point awareness. Our analysis of the coarse 3D prior indicates that it contains the essential geometric structures and captures the basic categorical attributes, keeping semantic rationality across different views. Building upon these observations, a natural insight is to preserve such inherent 3D knowledge as guidance and continuously benefit the 2D lifting process. For example, given text prompts \"a head of the Terracotta Army,\" we hope the knowledge in the guidance can prevent issues such as a pockmarked face or the unrealistic scenario of having a face on the back (e.g., Janus problem). To this end, we have designed two guiding strategies derived from \\( M \\): structural guidance for geometric fidelity and semantic guidance for 3D coherence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.86 + ], + "angle": 0, + "content": "Structural guidance. Given the current DMTet net \\(\\mathcal{F}\\) with parameters \\(\\theta\\) that encodes the coarse 3D prior \\(M\\), we apply a differentiable render \\(f_{n}\\) (e.g., nvidiffrast [32]) to generate a set of normal maps \\(\\mathcal{N} = \\{\\pmb{n}_i|\\pmb{n}_i = f_n(\\mathcal{F}_{\\theta},\\mathbf{c}_i),i = 1,\\dots n\\}\\), where \\(\\mathbf{c}_i\\) is the camera position randomly sampled in spherical coordinates. To extract the salient geometric structure features, we first use a Gaussian filter with a kernel standard deviation \\(\\sigma\\)" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.875, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nG (x, y) = \\frac {1}{2 \\pi \\sigma^ {2}} e ^ {- \\frac {x ^ {2} + y ^ {2}}{2 \\sigma^ {2}}} \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20766" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.155 + ], + "angle": 0, + "content": "to reduce the noise impact and obtain \\(\\{\\sigma (\\pmb {n}_i)\\}\\). As gradients are simple but effective tools for revealing the geometric contours and salient structures [14, 29], we then compute the structural descriptor sets \\(\\{G_{\\sigma}(\\pmb {n}_i)\\}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.165, + 0.47, + 0.207 + ], + "angle": 0, + "content": "\\[\nG _ {\\sigma} (\\boldsymbol {n} _ {i}) = \\sqrt {(\\frac {\\partial \\sigma (\\boldsymbol {n} _ {i})}{\\partial x}) ^ {2} + (\\frac {\\partial \\sigma (\\boldsymbol {n} _ {i})}{\\partial y}) ^ {2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.217, + 0.471, + 0.28 + ], + "angle": 0, + "content": "where \\(x\\) and \\(y\\) are the coordinate directions of the normal map \\(\\pmb{n}_i\\). Throughout the 2D lifting process of updating \\(\\mathcal{F}_{\\theta}\\) with newly rendered normal maps \\(\\tilde{\\mathcal{N}} = \\{\\tilde{n}_i\\}\\), it should follow the structural guidance as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.29, + 0.47, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} _ {\\text {s t r u c}} := \\sum_ {i = 1} ^ {n} \\left\\| G _ {\\sigma} \\left(\\boldsymbol {n} _ {i}\\right) - G _ {\\sigma} \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right) \\right\\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.47, + 0.386 + ], + "angle": 0, + "content": "which enables the 2D lifting process to preserve geometric fidelity and a well-aligned structure with the coarse 3D prior when generating rich details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.524 + ], + "angle": 0, + "content": "Semantic guidance. While structural guidance maintains low-level geometric perception from coarse 3D prior, semantic guidance extracts high-level features for 3D coherence. We first apply the pre-trained CLIP [57] model as semantic encoder \\(\\psi\\) to the normal set \\(\\mathcal{N}\\) and obtain semantic feature maps \\(\\mathcal{N}_c = \\{\\psi(\\boldsymbol{n}_i)\\}\\), proven to effectively capture semantic attributes like facial expressions or view categories [17]. Following the notation as above, we then define the semantic guidance with cosine similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.536, + 0.47, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} _ {\\text {s e m}} := \\sum_ {i = 1} ^ {n} \\frac {\\psi \\left(\\boldsymbol {n} _ {i}\\right) \\cdot \\psi \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right)}{\\| \\psi \\left(\\boldsymbol {n} _ {i}\\right) \\| \\| \\psi \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right) \\|.} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.587, + 0.47, + 0.664 + ], + "angle": 0, + "content": "Employing this guidance, we ensure that different views retain inherent high-level information throughout the 2D lifting optimization process. Experiments show that it can effectively mitigate multi-face problems, keeping 3D content semantically plausible from all viewing angles." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.673, + 0.218, + 0.69 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.849 + ], + "angle": 0, + "content": "In this subsection, we incorporate both structural and semantic guidance derived from coarse 3D prior to 2D lifting optimization so that it can produce vivid and diversified objects with multi-view consistency. For the disentangled geometry modeling, we use the randomly sampled normal map \\( \\pmb{n} \\) as the input, bridging the gap between 3D and 2D diffusion. To update the geometry model DMTet network \\( \\mathcal{F}_{\\theta} \\), we choose to use the publicly available Stable Diffusion [60] as pre-trained 2D diffusion model \\( \\phi \\) and compute the gradient of the SDS loss similar in Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.858, + 0.47, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\theta , \\boldsymbol {n}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\boldsymbol {z} _ {t} ^ {\\boldsymbol {n}}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\boldsymbol {z} _ {t} ^ {\\boldsymbol {n}}}{\\partial \\boldsymbol {n}} \\frac {\\partial \\boldsymbol {n}}{\\partial \\theta} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.153 + ], + "angle": 0, + "content": "where \\(\\partial z_t^n /\\partial n\\) calculates the gradient of the encoder in the latent diffusion model (LDM) [60]. Additionally, we introduce a step annealing technique to balance the influence of the 3D guidance during 2D lifting optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.158, + 0.892, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\gamma (\\lambda) = \\lambda e ^ {- \\beta \\max (0, n _ {\\mathrm {c u r}} - m)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.245 + ], + "angle": 0, + "content": "where \\(n_{\\mathrm{cur}}\\) is the current epoch and \\(\\{\\beta, m, \\lambda\\}\\) are the hyperparameters that control how \\(\\gamma\\) decreased. Therefore, the total loss \\(\\mathcal{L}_{\\mathrm{geo}}\\) to lift 2D geometry optimization with 3D guidance is a weighted sum of three loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.254, + 0.892, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g e o}} (\\theta , \\boldsymbol {n}) = \\mathcal {L} _ {\\mathrm {S D S}} + \\gamma \\left(\\lambda_ {\\mathrm {s t r u c}}\\right) \\mathcal {L} _ {\\mathrm {s t r u c}} + \\gamma \\left(\\lambda_ {\\mathrm {s e m}}\\right) \\mathcal {L} _ {\\mathrm {s e m}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.43 + ], + "angle": 0, + "content": "which not only enables the 3D content generation without multi-view inconsistency issues but also preserves the generalization and quality in 2D diffusion model \\(\\phi\\). As our pipeline can be integrated into any appearance model [9, 10, 33], we adopt a similar approach as Fantasia3D [10] to better align our text and 3D object. Denote \\(\\mathcal{T}\\) with parameters \\(\\eta\\) as our appearance model, we have the rendered image \\(\\mathbf{x} = \\mathcal{T}_{\\eta}(\\mathcal{F}_{\\theta}, c_{i})\\). To update \\(\\eta\\), we again apply the SDS loss for the final complete generated 3D object with detailed texture and coherent geometry:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.435, + 0.892, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\eta} \\mathcal {L} _ {\\mathrm {a p p}} (\\eta , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\boldsymbol {z} _ {t} ^ {\\mathbf {x}}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\boldsymbol {z} _ {t} ^ {\\mathbf {x}}}{\\partial \\mathbf {x}} \\frac {\\partial \\mathbf {x}}{\\partial \\eta} \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.545 + ], + "angle": 0, + "content": "which shares similar notations defined in Eq. 8. Finally, through the tailored 3D structural and semantic guidance that bridges the 2D and 3D diffusion models, our Sherpa3D can mitigate the multi-face problem and achieve high-fidelity and diversified results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.719, + 0.57 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.835 + ], + "angle": 0, + "content": "We apply the multilayer perceptron (MLP) comprising of three hidden layers to approximate \\(\\mathcal{F}_{\\theta}\\) and \\(\\mathcal{T}_{\\eta}\\). Adam optimizer [30] is used to update \\(\\mathcal{F}_{\\theta}\\) and \\(\\mathcal{T}_{\\eta}\\) with an initial learning rates of \\(1e - 3\\) decaying into \\(5e - 4\\). For 3D representations, we use textured mesh with a DMTet resolution of 128 to achieve a balance between quality and generation speed. We sample random camera poses at a fixed radius of 2.5, y-axis FOV of \\(45^{\\circ}\\), with the azimuth in \\([-180^{\\circ}, 180^{\\circ}]\\) and elevation in \\([-30^{\\circ}, 30^{\\circ}]\\). We load Shap-E from [46] for 3D diffusion model and choose stabilityai/stable-diffsuion-2-1-base [60] for 2D diffusion model. For weighting factors, we follow the same strategy as [26] to tune \\(w(t)\\). \\(\\lambda_{\\mathrm{struc}}\\) is set to 10 and \\(\\lambda_{\\mathrm{sem}}\\) is 30 to balance the magnitude of SDS loss. Notably, our method only needs a single NVIDIA RTX3090 (24GB) GPU within 25 minutes. More details of optimization, architecture design, and hyperparameter settings can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.634, + 0.863 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive experiments to evaluate our text-to-3D framework Sherpa3D and show" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20767" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.086, + 0.852, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.537, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons with baseline methods across different views \\((0^{\\circ}\\) and \\(180^{\\circ})\\). We can observe that baseline methods suffer from severe multi-face issues while our Sherpa3D can achieve better quality and 3D coherence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.473, + 0.698 + ], + "angle": 0, + "content": "comparison results against other text-to-3D baseline methods. We first present qualitative results compared with five SOTA baselines from different viewpoints. Then we report the quantitative results with a user study. Finally, we carry out ablation studies to further verify the efficacy of our framework design. Please refer to the supplementary for more comparisons, visualizations, and detailed analysis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.254, + 0.727 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Baselines. We extensively compare our method Sherpa3D against five baselines: Shap-E [28], DreamFusion [56], Magic3D [37], ProlificDreamer [74], and Fantasis3D [10]. Due to various reasons, we can't obtain the original implementation of some baselines. For DreamFusion, Magic3D, and ProlificDreamer, we utilize their implementations in the Threestudio library [19] for comparison. For Shap-E and Fantasia3D, we follow their official implementation. We consider these implementations to be the most reliable and comprehensive open-source option available in the field. To ensure a fair comparison, we use the Stable Diffusion [60]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.76, + 0.607 + ], + "angle": 0, + "content": "model as 2D diffusion prior by default." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Metrics. We will show our results with notable comparisons to other baselines through visualization. As there is no Ground-Truth 3D content corresponding to the text prompt, reference-based metrics like Chamfer Distance are difficult to apply to zero-shot text-to-3D generation. Following [27, 56], we evaluate the CLIP R-Precision [55], which can measure how well the rendered images of generated 3D content align with the input text. We use 100 prompts from the Common Objects in Context (COCO) dataset [38] as DreamFusion [56]. we also conduct a user study to further demonstrate the multi-view consistency and overall generation quality of our method," + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.801, + 0.729, + 0.818 + ], + "angle": 0, + "content": "4.2. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We first demonstrate vivid and diversified text-to-3D results generated from our Sherpa3D in the gallery as shown in Figure 1. Then we compare our method with five baseline method: Shap-E [28], DreamFusion [56], Magic3D [37], Fantasia3D [10] and ProlificDreamer [74]. Figure 3 and 4" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20768" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.09, + 0.845, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.523, + 0.781, + 0.538 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons with baseline methods across different views \\((-30^{\\circ}\\) and \\(150^{\\circ})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.473, + 0.716 + ], + "angle": 0, + "content": "give the comparative results with the same text prompt for each object generation. We observe that the Shap-E [28] only generates coarse shapes while other 2D lifting methods suffer from multi-face problems. In contrast, our Sherpa3D produces high-fidelity 3D assets with compelling texture quality and multi-view consistency. Notably, our framework is more efficient than other baselines with less time to optimize. Specifically, it only takes within 25 minutes from a text prompt to a high-quality 3D model ready to be used in graphic engines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.317, + 0.743 + ], + "angle": 0, + "content": "4.3. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.472, + 0.901 + ], + "angle": 0, + "content": "In Table 1, we report the CLIP R-Precision for Sherpa3D and several baselines. It shows that our method outperforms other baselines consistently across different CLIP models, and approaches the performance of ground truth (GT) images. For the user study, we render 360-degree rotating videos of 3D models generated from a collection of 120 images. Each volunteer is shown 10 samples of rendered video from a random method and rates in two aspects: multi-view consistency and overall generation quality. We collect results from 50 volunteers shown in Table 2. We observe" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.562, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on generation renderings with text prompts using different CLIP retrieval models. We compared to ground-truth images, Shap-E [28], Dreamfusion [56], Magic3D [37], evaluated on object-centric COCO as in [56]." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.628, + 0.892, + 0.733 + ], + "angle": 0, + "content": "
MethodR-Precision (%) ↑
CLIP B/32CLIP B/16CLIP L/14
GT Images77.379.2-
Shape-E [28]41.142.546.4
DreamFusion [56]70.373.275.0
Magic3D [37]71.573.876.1
Sherpa3D (Ours)72.375.679.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.761, + 0.892, + 0.791 + ], + "angle": 0, + "content": "that most users consider our results with much higher viewpoints consistency and overall generation fidelity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.757, + 0.818 + ], + "angle": 0, + "content": "4.4. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We carry out ablation studies on the design of our Sherpa3D framework in Figure 5 using an example text prompt \"a head of the Terracotta Army\". Specifically, we perform ablation on three aspects of our method: structural guidance, semantic guidance, and the step annealing strategy. The re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20769" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.135, + 0.1, + 0.145 + ], + "angle": 0, + "content": "\\(0^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.217, + 0.107, + 0.229 + ], + "angle": 0, + "content": "\\(180^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.095, + 0.171, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.186, + 0.177, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.094, + 0.264, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.186, + 0.278, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.268, + 0.209, + 0.278 + ], + "angle": 0, + "content": "Full" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.094, + 0.367, + 0.103 + ], + "angle": 0, + "content": "appearance" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.105, + 0.378, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.187, + 0.382, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.268, + 0.462, + 0.278 + ], + "angle": 0, + "content": "w/o Structural Guidance" + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.095, + 0.462, + 0.103 + ], + "angle": 0, + "content": "geometry" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.106, + 0.482, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.191, + 0.484, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.095, + 0.576, + 0.104 + ], + "angle": 0, + "content": "appearance" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.105, + 0.585, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.185, + 0.59, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.268, + 0.666, + 0.278 + ], + "angle": 0, + "content": "w/o Semantic Guidance" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.095, + 0.671, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.185, + 0.692, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.095, + 0.774, + 0.103 + ], + "angle": 0, + "content": "appearance" + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.105, + 0.785, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.186, + 0.787, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.268, + 0.861, + 0.279 + ], + "angle": 0, + "content": "w/o step annealing" + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.095, + 0.871, + 0.104 + ], + "angle": 0, + "content": "geometry" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.105, + 0.887, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.185, + 0.891, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.294, + 0.893, + 0.323 + ], + "angle": 0, + "content": "Figure 5. Ablation study of our method. The generation is based on the text prompt \"a head of the Terracotta Army\". We ablate the design choices of structural guidance, semantic guidance (Sec. 3.3), and the step annealing technique (Sec. 3.4)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.347, + 0.47, + 0.389 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on the multi-view consistency and overall generation quality score in a user study, rated on a scale of 1-10, with higher scores indicating better performance." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.399, + 0.47, + 0.49 + ], + "angle": 0, + "content": "
MethodMulti-view Consistency ↑Overall Quality ↑
Shap-E [28]6.093.33
DreamFusion [56]4.586.05
Magic3D [37]5.256.73
Fantasia3D [10]3.835.90
ProlificDreamer [74]5.787.02
Sherpa3D(Ours)8.958.74
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.697 + ], + "angle": 0, + "content": "sults reveal that the omission of any of these elements leads to a degradation in terms of quality and consistency. Notably, the absence of structural guidance leads to a loss of geometric fidelity in the \"army\", leading to a pockmarked face; without semantic guidance, there's a loss of semantic rationality across different views, resulting in the multiview Janus problem. The lack of a balanced step annealing results in an excessive influence of guidance with a rough final output. This illustrates the effectiveness of our overall framework (Figure 2), which drives geometric fidelity, multi-view consistency, and optimization balance steered by the 3D guidance and annealing strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.817 + ], + "angle": 0, + "content": "To further demonstrate our generalizability, we compare our method in Figure 6 with the Zero123 [40] which uses more 3D data [40] to finetune a 2D diffusion model to be viewpoint-aware. However, such a finetuning-based method easily overfits to 3D training data and suffers from severe performance degradation with unseen input of the training set. In contrast, our method is more generalizable to open-vocabulary text prompts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.196, + 0.846 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present Sherpa3D, a new framework that simultaneously achieves high-quality, diversified, and 3D consistent text-to-3D generation. By fully exploiting eas" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.346, + 0.585, + 0.357 + ], + "angle": 0, + "content": "Sherpa3D (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.359, + 0.582, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.435, + 0.712, + 0.452 + ], + "angle": 0, + "content": "\"A deer draped in exquisite ancient Chinese fabric, oil painting style\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.347, + 0.663, + 0.357 + ], + "angle": 0, + "content": "Zero123" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.362, + 0.692, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.347, + 0.8, + 0.358 + ], + "angle": 0, + "content": "Sherpa3D (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.363, + 0.776, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.435, + 0.85, + 0.453 + ], + "angle": 0, + "content": "\"A head of the Terracotta Army\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.347, + 0.879, + 0.357 + ], + "angle": 0, + "content": "Zero123" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.374, + 0.891, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.464, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Figure 6. Comparison with Zero123 [40]. We use the front view of our generated 3D model as the input of Zero123 with open-vocabulary text prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.704 + ], + "angle": 0, + "content": "ily obtained coarse 3D knowledge from the 3D diffusion model, we derive structural guidance and semantic guidance to enhance the prompts and provide continuous guidance with geometric fidelity and 3D coherence throughout the 2D lifting optimization. To further improve the overall performance, we incorporate a step annealing strategy that modulates the impact of 3D guidance and 2D refinement. Therefore, our framework bridges the gap between 2D and 3D diffusion models, preserving multi-view coherent generation while maintaining the generalizability and fidelity of 2D models. Extensive experiments verify the remarkable improvement of our Sherpa3D on text-to-3D generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Limitations and future works. Although our Sherpa3D achieves remarkable text-to-3D results, the quality still seems to be limited to the backbone itself as we choose Shap-E [28] and Stable Diffusion v2.1 base model in this work. We expect them to be solved with a larger diffusion model, such as SDXL [1] and DeepFloyd [2]. In future work, we are interested in extending our insight to more creative text-to-4D generation. We believe that Sherpa3D provides a promising research path for user-friendly and more accessible 3D content creation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant 62206147, and in part by Deng Feng Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20770" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] stable-diffusion-xl-base-1.0. https://huggingface.co/stabilityai/stable-diffusion-x1-base-1.0. Accessed:2023-08-29.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.184 + ], + "angle": 0, + "content": "[2] Deepfloyd. https://huggingface.co/DeepFloyd. Accessed: 2023-08-25. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, pages 40-49. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[4] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.452 + ], + "angle": 0, + "content": "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.468, + 0.591 + ], + "angle": 0, + "content": "[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[9] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. arXiv preprint arXiv:2307.05663, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[14] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[15] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 3DV, pages 402-411. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurlPS, 35:31841-31854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 6(3):e30, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[21] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Gancraft: Unsupervised 3d neural rendering of northwest world. In ICCV, pages 14072-14082, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[22] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In ICCV, pages 9984-9993, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[25] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[26] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[27] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[29] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20771" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[31] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6:124, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[33] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[34] Chenghao Li, Chaoning Zhang, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[35] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[36] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.589 + ], + "angle": 0, + "content": "[37] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 300–309, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[39] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[40] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[41] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Sin-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "gle image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[43] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[44] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.894, + 0.334 + ], + "angle": 0, + "content": "[46] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.894, + 0.405 + ], + "angle": 0, + "content": "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.894, + 0.532 + ], + "angle": 0, + "content": "[49] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.894, + 0.589 + ], + "angle": 0, + "content": "[50] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.894, + 0.688 + ], + "angle": 0, + "content": "[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[53] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[54] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, pages 13503–13513, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[55] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[56] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20772" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Eleventh International Conference on Learning Representations (ICLR), 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[58] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[63] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. NeurlPS, 35:33999-34011, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[65] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[66] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[68] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[69] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[71] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[73] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[75] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[76] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[77] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[78] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[79] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6841–6850, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[80] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[81] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20773" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[82] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. arXiv preprint arXiv:2010.09125, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[83] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[84] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.304 + ], + "angle": 0, + "content": "[85] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20774" + } + ] +] \ No newline at end of file diff --git a/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_origin.pdf b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..19533bdc6982ec542a8192d5a8bcd72cdf2c1deb --- /dev/null +++ b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/4ac2d1ab-faf4-4d9c-949b-ff0f163e6a1b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8e97d1754126fbf9adb79c1f2882ea7f63d2ba3fc341280874a2eea4819ac60 +size 5427640 diff --git a/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/full.md b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c9702cc1738e857eda9b5a4632ee51faff14cc4 --- /dev/null +++ b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/full.md @@ -0,0 +1,359 @@ +# Sherpa3D: Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior + +Fangfu Liu $^{1}$ , Diankun Wu $^{1}$ , Yi Wei $^{1}$ , Yongming Rao $^{2}$ , Yueqi Duan $^{1\dagger}$ , Tsinghua University, BAAI + +![](images/b8c3f96632c6ebab61211bbb34c256e0ab87fc38cead4172796f0e738cebb26c.jpg) +Figure 1. Gallery of Sherpa3D: Blender rendering for various textured meshes from Sherpa3D, which is able to generate high-fidelity, diverse, and multi-view consistent 3D contents with input text prompts. Our method is also compatible with popular graphics engines. + +# Abstract + +Recently, 3D content creation from text prompts has demonstrated remarkable progress by utilizing 2D and 3D diffusion models. While 3D diffusion models ensure great multi-view consistency, their ability to generate high-quality and diverse 3D assets is hindered by the limited 3D data. In contrast, 2D diffusion models find a distillation approach that achieves excellent generalization + +and rich details without any 3D data. However, 2D lifting methods suffer from inherent view-agnostic ambiguity thereby leading to serious multi-face Janus issues, where text prompts fail to provide sufficient guidance to learn coherent 3D results. Instead of retraining a costly viewpoint-aware model, we study how to fully exploit easily accessible coarse 3D knowledge to enhance the prompts and guide 2D lifting optimization for refinement. In this paper, we propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric + +consistency simultaneously. Specifically, we design a pair of guiding strategies derived from the coarse 3D prior generated by the 3D diffusion model: a structural guidance for geometric fidelity and a semantic guidance for 3D coherence. Employing the two types of guidance, the 2D diffusion model enriches the 3D content with diversified and high-quality results. Extensive experiments show the superiority of our Sherpa3D over the state-of-the-art text-to-3D methods in terms of quality and 3D consistency. Project page: https://liuff19.github.io/Sherpa3D/. + +# 1. Introduction + +3D content generation [34, 37, 56, 76] finds a broad range of applications, including games, movies, virtual/augmented reality and robots. However, the conventional process of creating premium 3D assets is still expensive and challenging as it requires multiple labor-intensive and time-consuming stages [31]. Fortunately, this challenge has prompted the development of recent text-to-3D methods [10, 25, 27, 35, 37, 47, 50, 56, 74]. Only using text prompts to automate 3D generation, these techniques pave a promising way towards streamlining 3D creation. + +Powered by the great breakthroughs in diffusion models [52, 60, 61, 81], two research lines of rationalization have recently emerged in text-to-3D: inference-only 3D diffusion methods and optimization-based 2D lifting methods. Specifically, the inference-only methods [20, 28, 52] seek to directly generate 3D-consistent assets by extensively training a new diffusion model on 3D data. However, due to the scarcity of 3D datasets compared to accessible 2D images or text data, these 3D diffusion models suffer from low quality and limited generalizability. Without requiring any 3D data for training, 2D lifting methods [4, 10, 37, 47, 56, 72, 74] can produce high-quality and diversified 3D results by distilling 3D knowledge from pre-trained 2D diffusion models [60, 61, 81], also known as Score Distillation Sampling (SDS). Yet lifting 2D observations into 3D is inherently ambiguous without sufficient 3D guidance from text prompts, leading to notorious multi-view inconsistency (e.g., Janus problems) in 2D lifting methods. + +These findings motivate us to think: is it possible to bridge the two aforementioned streams to achieve generalizability, high-fidelity, and geometric consistency simultaneously? An intuitive idea is to leverage more 3D data [11, 12] to fine-tune a view-point aware diffusion model, but it requires substantial computational resources and is prone to overfitting due to data bias [40, 67]. In contrast, our key insight is to utilize the easily accessible 3D diffusion model as guidance and study how to fully exploit coarse 3D knowledge to guide 2D lifting optimization for refinement. In particular, when maintaining the quality and generalizability of the original 2D diffusion model, we hope + +the 2D lifting awareness can be guided by the strong 3D geometric information from the 3D diffusion model. However, it is non-trivial in pursuit of this balance. Relying too heavily on the coarse 3D priors from the 3D diffusion model may degrade the generation quality, whereas little 3D guidance could result in a lack of geometric awareness, leading to multi-view inconsistency. + +Towards this end, we propose Sherpa3D in this paper, which greatly boosts high-fidelity and highly diversified text-to-3D generation with geometric consistency. Our method begins by employing a 3D diffusion model to craft a basic 3D guide with limited details. Building upon the coarse 3D prior, we introduce two guiding strategies to inform 2D diffusion model throughout lifting optimization: a structural guide for geometric fidelity and a semantic guide for 3D coherence. Specifically, the structural guide leverages the first-order gradient information of the normals from the 3D prior to supervise the optimization of the structure. These normals are then integrated into the input of a pre-trained 2D diffusion model, refining the geometric details. Concurrently, our semantic guide extracts high-level features from multi-views of the 3D prior. These features guide the 2D lifting optimization to perceive the geometric consistency under the preservation of original generalizability and quality. Furthermore, we design an annealing function, which modulates the influence of the 3D guidance to better preserve the capabilities of 2D and 3D diffusion models. As a result, our Sherpa3D is aware of the geometric consistency with rich details and generalizes well across diverse text prompts. Extensive experiments verify the efficacy of our framework and show that our Sherpa3D outperforms existing methods for high-fidelity and geometric consistency (see qualitative results gallery in Figure 1 and quantitative results in Table 2). + +# 2. Related Work + +# 2.1. Text-to-image Generation + +Recently, text-to-image models such as unCLIP [59], Imagegen [61], and Stable Diffusion [60] have shown remarkable capability of generating high-quality and creative images given text prompts. Such significant progress is powered by advances in diffusion models [13, 24, 53, 69], which can be pre-trained on billions of image-text pairs [62, 64] and understands general objects with complex semantic concepts (nouns, artistic styles, etc.) [60]. Despite the great success of photorealistic and diversified image generation, using language to generate different viewpoints of the same object with 3D coherence remains a challenging problem [75]. + +# 2.2. Text-to-3D Generation + +Building on promising text-to-image diffusion models, there has been a surge of studies in text-to-3D generation. + +However, it is non-trivial due to the scarcity of diverse 3D data [8, 12, 77] compared to 2D. Existing 3D native diffusion models [20, 28, 43, 52, 80, 83] usually work on a limited object category and struggle with generating in-the-wild 3D assets. To achieve generalizable 3D generation, pioneering works DreamFusion [56] and SJC [72] propose to distill the score of image distribution from pre-trained 2D diffusion models [60, 61] and show impressive results. Following works [10, 26, 36, 37, 47, 70, 71, 74, 79, 84] continue to enhance various aspects such as generation fidelity and optimization stability or explore more application scenarios [58, 68, 85]. As it is inherently ambiguous to lift 2D observations into 3D, they may suffer from multi-face issues. Although some methods use prompt engineering [4] or train a costly viewpoint-aware model [40, 67] to alleviate such problems, they fail to generate high-quality results [10] or easily overfit to domain-specific data [12, 67]. In this work, we bridge the gap between 3D and 2D diffusion models through meticulously designed 3D guidance, which leads the 2D lifting process to achieve high-fidelity, diversified, and coherent 3D generation. + +# 2.3. 3D Generative Models + +Extensive research has been conducted in the field of 3D generative modeling, exploring diverse 3D representations like 3D voxel grids [15, 22, 44], point clouds [3, 45, 49], and meshes [16, 82]. The majority of these approaches rely on training data presented in the form of 3D assets, which proves challenging to obtain at a large scale. Drawing inspiration from the success of neural volume rendering, recent studies have shifted towards investing in 3D-aware image synthesis [6, 7, 18, 21, 54, 63]. This approach offers the advantage of directly learning 3D generative models from images. However, volume rendering networks typically exhibit slow querying speeds, resulting in a trade-off between extended training times and a lack of multi-view consistency. Recently, benefitted from 2D diffusion models, some works generate multi-view images with single-view input [39-42, 66, 78]. As one of the pioneering works, Zero-1-to-3 [40] uses a synthetic dataset to fine-tune the pretrained diffusion models, aiming to learn controls of the relative camera viewpoint. Beyond Zero-1-to-3, SyncDreamer [41] employs a synchronized multiview diffusion model to capture the joint probability distribution of multiview images. This model facilitates the generation of multiview-consistent images through a unified reverse process. Different from these methods, we focus on text-to-3D synthesis, with the goal of generating multi-view consistent 3D contents with text prompts. + +# 3. Method + +Given a text prompt, our goal is to generate 3D assets with high quality, generalizability, and multi-view consistency. + +Our framework can be divided into three stages: (1) build coarse 3D prior from the 3D diffusion model (Sec. 3.2); (2) formulate two guiding strategies (e.g., structural and semantic guidance) for 2D lifting process (Sec. 3.3); (3) incorporate both 3D guidance and SDS loss with an annealing technique in optimization and generate the final 3D object (Sec. 3.4). In this way, we can leverage the full power of state-of-the-art 3D and 2D diffusion models to obtain 3D coherence as 3D models, retaining intricate details and creative freedom as 2D models. Our pipeline is depicted in Figure 2. Before introducing our Sherpa3D in detail, we first review the theory of Score Distillation Sampling (SDS). + +# 3.1. Preliminaries + +Score Distillation Sampling (SDS). As one of the most representative 2D lifting methods, Dreamfusion [56] first presents the concept of Score Distillation Sampling (SDS), which is an algorithm to optimize a 3D representation such that the image rendered from any view maintains a high likelihood as evaluated by the 2D diffusion model given text prompts. SDS consists of two key components: (1) a 3D representation with parameter $\theta$ , which can produce an image $x$ at desired camera $\mathbf{c}$ through a parametric function $\mathbf{x} = g(\theta; \mathbf{c})$ ; (2) a pre-trained text-to-image 2D diffusion model $\phi$ with a score function $\epsilon_{\phi}(\mathbf{x}_t; y, t)$ that predicts the sample noise $\epsilon$ given noisy image $\mathbf{x}_t$ , noise level $t$ and text embedding $y$ . The score function guides the direction of the gradient for updating $\theta$ to reside rendered images in high-density areas conditioned on text $y$ . The gradient is calculated by SDS as: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\phi , \mathbf {x}) = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\phi} \left(\mathbf {x} _ {t}; y, t\right) - \epsilon\right) \frac {\partial \mathbf {x}}{\partial \theta} \right], \tag {1} +$$ + +where $w(t)$ is a weighting function. In practice, the denoising score function $\epsilon_{\phi}$ is often replaced with another function $\tilde{\epsilon}_{\phi}$ that uses classifier-free guidance [23] that controls the strength of the text condition (see Supplementary). + +# 3.2. Sculpting a Coarse 3D Prior + +To facilitate text-to-3D generation, most existing methods [26, 47, 56] rely on implicit 3D representations such as Neural Radiance Fields (NeRF) [48] and its variants [5, 51]. However, it is difficult for NeRF-based modeling to extract the high-quality surface with material and texture [73]. To address this, we adopt the hybrid scene representation of DMTet [65], including a deformable tetrahedral grid that encodes a signed distance function (SDF) and a differentiable marching tetrahedra (MT) layer that extracts explicit surface mesh. Equipped with the hybrid representation, we sculpt a coarse 3D prior from 3D diffusion model $G_{3D}$ (e.g., Shap-E [28]) by the following procedure. Given text prompts $y$ , we first use the 3D diffusion model $G_{3D}$ to generate 3D results $M_0$ and employ multi-layer perceptions + +![](images/a5d0d0087271a6944901772de5dcc1d22d6223fdaabab22d3023c32bf817b304.jpg) +Figure 2. Pipeline of our Sherpa3D. Given a text as input, we first prompt 3D diffusion to build a coarse 3D prior $M$ encoded in the geometry model (e.g., DMTet). Next, we render the normal map of the extracted mesh in DMTet and derive two guiding strategies from $M$ . (a) Structural Guidance: we utilize the structural descriptor to compute salient geometric features for preserving geometry fidelity (e.g., without a pockmarked face problem). (b) Semantic Guidance: we leverage a semantic encoder (e.g., CLIP) to extract high-level information for keeping 3D consistency (e.g., without multi-face issues). Employing the two guidance in 2D lifting process, we use the normal map as shape encoding of the 2D diffusion model and unleash its power to generate high-quality and diversified results with 3D coherence. Then we achieve the final 3D results via photorealistic rendering through appearance modeling. ("Everest's summit eludes many without Sherpa.") + +(MLPs) to query SDF values for each vertex along a regular grid. Next we sample a point set $\mathcal{P} = \{\pmb {p}_i\in \mathbb{R}^3\}$ from $M_0$ with their SDF values $\{SDF(\pmb {p}_i)\}$ . For each $\pmb{p}_i$ , the DMTet network $\mathcal{F}$ can predict SDF value $s(\pmb {p}_i)$ , and a position offset $\Delta p_{i}$ by: + +$$ +\left(s \left(\boldsymbol {p} _ {i}\right), \Delta \boldsymbol {p} _ {i}\right) = \mathcal {F} \left(\boldsymbol {p} _ {i}; \theta\right), \tag {2} +$$ + +where $\theta$ is the parameters of network $\mathcal{F}$ . Then, we incorporate 3D priors into the DMTet network $\mathcal{F}_{\theta}$ with the point set derived from 3D diffusion model by minimizing: + +$$ +\mathcal {L} _ {S D F} = \sum_ {\boldsymbol {p} _ {i} \in \mathcal {P}} | s (\boldsymbol {p} _ {i}) - S D F (\boldsymbol {p} _ {i}) | ^ {2} + \lambda_ {d e f} \sum_ {\boldsymbol {p} _ {i} \in \mathcal {P}} \| \Delta \boldsymbol {p} _ {i} \| _ {2}, \tag {3} +$$ + +where $\lambda_{def}$ is the hyperparameter controlling $L_{2}$ regularization strengths on offsets to avoid artifacts. Finally, we apply the MT layer to extract mesh representation $M$ . Now, we have leveraged the knowledge from the 3D diffusion model to construct a coarse 3D prior, which is encoded implicitly in DMTet $\mathcal{F}_{\theta}$ and represented explicitly by mesh $M$ . Next, we will discuss how to utilize the coarse 3D prior $M$ as a guide during the subsequent 2D diffusion lifting optimization to refine a high-quality result with 3D coherence. + +# 3.3. 3D Guidance for 2D Lifting Optimization + +What knowledge can serve as guidance? The purpose of introducing a 3D prior as guidance is to address the prevalent issue of viewpoint inconsistency both in geometry and appearance. Through empirical studies, we have identified + +geometric inconsistency as the main cause of 3D incoherence, leading to multi-face Janus problem [35, 67]. In contrast, appearance inconsistency emerges in much more extreme scenarios with lesser significance. Therefore, we disentangle the geometry from the 3D model and fully leverage coarse prior $M$ to guide 2D lifting geometry optimization with view-point awareness. Our analysis of the coarse 3D prior indicates that it contains the essential geometric structures and captures the basic categorical attributes, keeping semantic rationality across different views. Building upon these observations, a natural insight is to preserve such inherent 3D knowledge as guidance and continuously benefit the 2D lifting process. For example, given text prompts "a head of the Terracotta Army," we hope the knowledge in the guidance can prevent issues such as a pockmarked face or the unrealistic scenario of having a face on the back (e.g., Janus problem). To this end, we have designed two guiding strategies derived from $M$ : structural guidance for geometric fidelity and semantic guidance for 3D coherence. + +Structural guidance. Given the current DMTet net $\mathcal{F}$ with parameters $\theta$ that encodes the coarse 3D prior $M$ , we apply a differentiable render $f_{n}$ (e.g., nvidiffrast [32]) to generate a set of normal maps $\mathcal{N} = \{\pmb{n}_i|\pmb{n}_i = f_n(\mathcal{F}_{\theta},\mathbf{c}_i),i = 1,\dots n\}$ , where $\mathbf{c}_i$ is the camera position randomly sampled in spherical coordinates. To extract the salient geometric structure features, we first use a Gaussian filter with a kernel standard deviation $\sigma$ + +$$ +G (x, y) = \frac {1}{2 \pi \sigma^ {2}} e ^ {- \frac {x ^ {2} + y ^ {2}}{2 \sigma^ {2}}} \tag {4} +$$ + +to reduce the noise impact and obtain $\{\sigma (\pmb {n}_i)\}$ . As gradients are simple but effective tools for revealing the geometric contours and salient structures [14, 29], we then compute the structural descriptor sets $\{G_{\sigma}(\pmb {n}_i)\}$ by + +$$ +G _ {\sigma} (\boldsymbol {n} _ {i}) = \sqrt {(\frac {\partial \sigma (\boldsymbol {n} _ {i})}{\partial x}) ^ {2} + (\frac {\partial \sigma (\boldsymbol {n} _ {i})}{\partial y}) ^ {2}}, \tag {5} +$$ + +where $x$ and $y$ are the coordinate directions of the normal map $\pmb{n}_i$ . Throughout the 2D lifting process of updating $\mathcal{F}_{\theta}$ with newly rendered normal maps $\tilde{\mathcal{N}} = \{\tilde{n}_i\}$ , it should follow the structural guidance as: + +$$ +\min _ {\theta} \mathcal {L} _ {\text {s t r u c}} := \sum_ {i = 1} ^ {n} \left\| G _ {\sigma} \left(\boldsymbol {n} _ {i}\right) - G _ {\sigma} \left(\tilde {\boldsymbol {n}} _ {i}\right) \right\| _ {2} ^ {2}, \tag {6} +$$ + +which enables the 2D lifting process to preserve geometric fidelity and a well-aligned structure with the coarse 3D prior when generating rich details. + +Semantic guidance. While structural guidance maintains low-level geometric perception from coarse 3D prior, semantic guidance extracts high-level features for 3D coherence. We first apply the pre-trained CLIP [57] model as semantic encoder $\psi$ to the normal set $\mathcal{N}$ and obtain semantic feature maps $\mathcal{N}_c = \{\psi(\boldsymbol{n}_i)\}$ , proven to effectively capture semantic attributes like facial expressions or view categories [17]. Following the notation as above, we then define the semantic guidance with cosine similarity: + +$$ +\min _ {\theta} \mathcal {L} _ {\text {s e m}} := \sum_ {i = 1} ^ {n} \frac {\psi \left(\boldsymbol {n} _ {i}\right) \cdot \psi \left(\tilde {\boldsymbol {n}} _ {i}\right)}{\| \psi \left(\boldsymbol {n} _ {i}\right) \| \| \psi \left(\tilde {\boldsymbol {n}} _ {i}\right) \|.} \tag {7} +$$ + +Employing this guidance, we ensure that different views retain inherent high-level information throughout the 2D lifting optimization process. Experiments show that it can effectively mitigate multi-face problems, keeping 3D content semantically plausible from all viewing angles. + +# 3.4. Optimization + +In this subsection, we incorporate both structural and semantic guidance derived from coarse 3D prior to 2D lifting optimization so that it can produce vivid and diversified objects with multi-view consistency. For the disentangled geometry modeling, we use the randomly sampled normal map $\pmb{n}$ as the input, bridging the gap between 3D and 2D diffusion. To update the geometry model DMTet network $\mathcal{F}_{\theta}$ , we choose to use the publicly available Stable Diffusion [60] as pre-trained 2D diffusion model $\phi$ and compute the gradient of the SDS loss similar in Eq. 1: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\theta , \boldsymbol {n}) = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\phi} \left(\boldsymbol {z} _ {t} ^ {\boldsymbol {n}}; y, t\right) - \epsilon\right) \frac {\partial \boldsymbol {z} _ {t} ^ {\boldsymbol {n}}}{\partial \boldsymbol {n}} \frac {\partial \boldsymbol {n}}{\partial \theta} \right], \tag {8} +$$ + +where $\partial z_t^n /\partial n$ calculates the gradient of the encoder in the latent diffusion model (LDM) [60]. Additionally, we introduce a step annealing technique to balance the influence of the 3D guidance during 2D lifting optimization: + +$$ +\gamma (\lambda) = \lambda e ^ {- \beta \max (0, n _ {\mathrm {c u r}} - m)}, \tag {9} +$$ + +where $n_{\mathrm{cur}}$ is the current epoch and $\{\beta, m, \lambda\}$ are the hyperparameters that control how $\gamma$ decreased. Therefore, the total loss $\mathcal{L}_{\mathrm{geo}}$ to lift 2D geometry optimization with 3D guidance is a weighted sum of three loss terms: + +$$ +\mathcal {L} _ {\mathrm {g e o}} (\theta , \boldsymbol {n}) = \mathcal {L} _ {\mathrm {S D S}} + \gamma \left(\lambda_ {\mathrm {s t r u c}}\right) \mathcal {L} _ {\mathrm {s t r u c}} + \gamma \left(\lambda_ {\mathrm {s e m}}\right) \mathcal {L} _ {\mathrm {s e m}}, \tag {10} +$$ + +which not only enables the 3D content generation without multi-view inconsistency issues but also preserves the generalization and quality in 2D diffusion model $\phi$ . As our pipeline can be integrated into any appearance model [9, 10, 33], we adopt a similar approach as Fantasia3D [10] to better align our text and 3D object. Denote $\mathcal{T}$ with parameters $\eta$ as our appearance model, we have the rendered image $\mathbf{x} = \mathcal{T}_{\eta}(\mathcal{F}_{\theta}, c_{i})$ . To update $\eta$ , we again apply the SDS loss for the final complete generated 3D object with detailed texture and coherent geometry: + +$$ +\nabla_ {\eta} \mathcal {L} _ {\mathrm {a p p}} (\eta , \mathbf {x}) = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\phi} \left(\boldsymbol {z} _ {t} ^ {\mathbf {x}}; y, t\right) - \epsilon\right) \frac {\partial \boldsymbol {z} _ {t} ^ {\mathbf {x}}}{\partial \mathbf {x}} \frac {\partial \mathbf {x}}{\partial \eta} \right], \tag {11} +$$ + +which shares similar notations defined in Eq. 8. Finally, through the tailored 3D structural and semantic guidance that bridges the 2D and 3D diffusion models, our Sherpa3D can mitigate the multi-face problem and achieve high-fidelity and diversified results. + +# 3.5. Implementation Details + +We apply the multilayer perceptron (MLP) comprising of three hidden layers to approximate $\mathcal{F}_{\theta}$ and $\mathcal{T}_{\eta}$ . Adam optimizer [30] is used to update $\mathcal{F}_{\theta}$ and $\mathcal{T}_{\eta}$ with an initial learning rates of $1e - 3$ decaying into $5e - 4$ . For 3D representations, we use textured mesh with a DMTet resolution of 128 to achieve a balance between quality and generation speed. We sample random camera poses at a fixed radius of 2.5, y-axis FOV of $45^{\circ}$ , with the azimuth in $[-180^{\circ}, 180^{\circ}]$ and elevation in $[-30^{\circ}, 30^{\circ}]$ . We load Shap-E from [46] for 3D diffusion model and choose stabilityai/stable-diffsuion-2-1-base [60] for 2D diffusion model. For weighting factors, we follow the same strategy as [26] to tune $w(t)$ . $\lambda_{\mathrm{struc}}$ is set to 10 and $\lambda_{\mathrm{sem}}$ is 30 to balance the magnitude of SDS loss. Notably, our method only needs a single NVIDIA RTX3090 (24GB) GPU within 25 minutes. More details of optimization, architecture design, and hyperparameter settings can be found in the supplementary. + +# 4. Experiments + +In this section, we conduct comprehensive experiments to evaluate our text-to-3D framework Sherpa3D and show + +![](images/7d6ff6248e87c28dde9bb2fbbea9a9faedaab9b7317e158a271c995b649fc3b9.jpg) +Figure 3. Qualitative comparisons with baseline methods across different views $(0^{\circ}$ and $180^{\circ})$ . We can observe that baseline methods suffer from severe multi-face issues while our Sherpa3D can achieve better quality and 3D coherence. + +comparison results against other text-to-3D baseline methods. We first present qualitative results compared with five SOTA baselines from different viewpoints. Then we report the quantitative results with a user study. Finally, we carry out ablation studies to further verify the efficacy of our framework design. Please refer to the supplementary for more comparisons, visualizations, and detailed analysis. + +# 4.1. Experiment Setup + +Baselines. We extensively compare our method Sherpa3D against five baselines: Shap-E [28], DreamFusion [56], Magic3D [37], ProlificDreamer [74], and Fantasis3D [10]. Due to various reasons, we can't obtain the original implementation of some baselines. For DreamFusion, Magic3D, and ProlificDreamer, we utilize their implementations in the Threestudio library [19] for comparison. For Shap-E and Fantasia3D, we follow their official implementation. We consider these implementations to be the most reliable and comprehensive open-source option available in the field. To ensure a fair comparison, we use the Stable Diffusion [60] + +model as 2D diffusion prior by default. + +Metrics. We will show our results with notable comparisons to other baselines through visualization. As there is no Ground-Truth 3D content corresponding to the text prompt, reference-based metrics like Chamfer Distance are difficult to apply to zero-shot text-to-3D generation. Following [27, 56], we evaluate the CLIP R-Precision [55], which can measure how well the rendered images of generated 3D content align with the input text. We use 100 prompts from the Common Objects in Context (COCO) dataset [38] as DreamFusion [56]. we also conduct a user study to further demonstrate the multi-view consistency and overall generation quality of our method, + +# 4.2. Qualitative Comparisons + +We first demonstrate vivid and diversified text-to-3D results generated from our Sherpa3D in the gallery as shown in Figure 1. Then we compare our method with five baseline method: Shap-E [28], DreamFusion [56], Magic3D [37], Fantasia3D [10] and ProlificDreamer [74]. Figure 3 and 4 + +![](images/32c1b7f6ece146318a53c2deb9f2af304970f0c375b5b1d6dc48b7a188d45f94.jpg) +Figure 4. Qualitative comparisons with baseline methods across different views $(-30^{\circ}$ and $150^{\circ})$ + +give the comparative results with the same text prompt for each object generation. We observe that the Shap-E [28] only generates coarse shapes while other 2D lifting methods suffer from multi-face problems. In contrast, our Sherpa3D produces high-fidelity 3D assets with compelling texture quality and multi-view consistency. Notably, our framework is more efficient than other baselines with less time to optimize. Specifically, it only takes within 25 minutes from a text prompt to a high-quality 3D model ready to be used in graphic engines. + +# 4.3. Quantitative Comparisons + +In Table 1, we report the CLIP R-Precision for Sherpa3D and several baselines. It shows that our method outperforms other baselines consistently across different CLIP models, and approaches the performance of ground truth (GT) images. For the user study, we render 360-degree rotating videos of 3D models generated from a collection of 120 images. Each volunteer is shown 10 samples of rendered video from a random method and rates in two aspects: multi-view consistency and overall generation quality. We collect results from 50 volunteers shown in Table 2. We observe + +Table 1. Quantitative comparisons on generation renderings with text prompts using different CLIP retrieval models. We compared to ground-truth images, Shap-E [28], Dreamfusion [56], Magic3D [37], evaluated on object-centric COCO as in [56]. + +
MethodR-Precision (%) ↑
CLIP B/32CLIP B/16CLIP L/14
GT Images77.379.2-
Shape-E [28]41.142.546.4
DreamFusion [56]70.373.275.0
Magic3D [37]71.573.876.1
Sherpa3D (Ours)72.375.679.3
+ +that most users consider our results with much higher viewpoints consistency and overall generation fidelity. + +# 4.4. Ablation Study and Analysis + +We carry out ablation studies on the design of our Sherpa3D framework in Figure 5 using an example text prompt "a head of the Terracotta Army". Specifically, we perform ablation on three aspects of our method: structural guidance, semantic guidance, and the step annealing strategy. The re + +![](images/0621e3cec3b6f38e082b6003b1dbebfc7010780acae19a872404a42fe7c600af.jpg) +$0^{\circ}$ + +![](images/5d532502f20d283fa6c957c729d2c36f6976412953337cb654564a775c702abe.jpg) +$180^{\circ}$ +Figure 5. Ablation study of our method. The generation is based on the text prompt "a head of the Terracotta Army". We ablate the design choices of structural guidance, semantic guidance (Sec. 3.3), and the step annealing technique (Sec. 3.4). + +![](images/3a6437b306a74baebd0daaae842bd2fa2abf600edf6987a693fff13a3a4888f8.jpg) + +![](images/5ce9efb194bca24646ce7a8547a9b70ce0084f61b01cb523fec302eb84c074da.jpg) +Full + +![](images/0bb268e3aadc1079fc2d41590fc43c7fca663e82448599b1fa7ffc08dadb9c02.jpg) +appearance + +![](images/acd3e382ceab519f97c67094d081af11890c03013593f8301aa21ed422821d4a.jpg) +w/o Structural Guidance + +![](images/a55cbdabe9a2eaa10b3612361317d091ff3b53cb162617d7fdea7b64dedcdd37.jpg) +geometry + +![](images/6bcce458f5808737039f5f7d0e383faefc4e04ac16487e2a694dd807de118ff2.jpg) + +![](images/515e8be1d68da195e576a82d90e9cb58f8cc7993bf98f15a7a12ae07292f0239.jpg) +appearance + +![](images/6efdc9de646c27c0e24b3715d1b810ecfc6a472c9cff292182dac52245b187e3.jpg) +w/o Semantic Guidance + +![](images/c12160555f8f2eaf666930b4047a12b562db56ee815e0f7c5abb4987d0360254.jpg) + +![](images/9391c2def14aaca71bd70c965cbd8fa888490440fd2006fade33619100de7ff9.jpg) + +![](images/b17721830ddb10af0cbe66dcfe5c8b1ed558b00a902d014c09a9a8a4049f6607.jpg) +appearance + +![](images/9f2b55d271bc0fe6744cc06b0b829d23a76c39101a7a377156d8953e6de1b03c.jpg) +w/o step annealing + +![](images/d030b334723685502f8e5c2c480fc47c1aed397a586e49831506a498632d2e05.jpg) +geometry + +![](images/588137a4dabf0da9dbd143790d45e7cfbef9055cf99cbe9b0eabe14661580629.jpg) + +Table 2. Quantitative comparisons on the multi-view consistency and overall generation quality score in a user study, rated on a scale of 1-10, with higher scores indicating better performance. + +
MethodMulti-view Consistency ↑Overall Quality ↑
Shap-E [28]6.093.33
DreamFusion [56]4.586.05
Magic3D [37]5.256.73
Fantasia3D [10]3.835.90
ProlificDreamer [74]5.787.02
Sherpa3D(Ours)8.958.74
+ +sults reveal that the omission of any of these elements leads to a degradation in terms of quality and consistency. Notably, the absence of structural guidance leads to a loss of geometric fidelity in the "army", leading to a pockmarked face; without semantic guidance, there's a loss of semantic rationality across different views, resulting in the multiview Janus problem. The lack of a balanced step annealing results in an excessive influence of guidance with a rough final output. This illustrates the effectiveness of our overall framework (Figure 2), which drives geometric fidelity, multi-view consistency, and optimization balance steered by the 3D guidance and annealing strategy. + +To further demonstrate our generalizability, we compare our method in Figure 6 with the Zero123 [40] which uses more 3D data [40] to finetune a 2D diffusion model to be viewpoint-aware. However, such a finetuning-based method easily overfits to 3D training data and suffers from severe performance degradation with unseen input of the training set. In contrast, our method is more generalizable to open-vocabulary text prompts. + +# 5. Conclusion + +In this paper, we present Sherpa3D, a new framework that simultaneously achieves high-quality, diversified, and 3D consistent text-to-3D generation. By fully exploiting eas + +![](images/7c0d347e447eabfbaaf36cf29e02ba7e20bba97abe3b98204bc6c0a5a9e62498.jpg) +Sherpa3D (Ours) +"A deer draped in exquisite ancient Chinese fabric, oil painting style" +Figure 6. Comparison with Zero123 [40]. We use the front view of our generated 3D model as the input of Zero123 with open-vocabulary text prompts. + +![](images/bc971454719d6accdf8ac61ae0a3297419f96e83ef13edeb02b77d1da4e55052.jpg) +Zero123 + +![](images/a8b4c2d0b228ecdbd7e2b490f92777f8f240dd187890f4ab26edc4d6815366a6.jpg) +Sherpa3D (Ours) +"A head of the Terracotta Army" + +![](images/7bb17022eb364810a80b375f803a624dc71d37f5f790542a481ec2cf325097dd.jpg) +Zero123 + +ily obtained coarse 3D knowledge from the 3D diffusion model, we derive structural guidance and semantic guidance to enhance the prompts and provide continuous guidance with geometric fidelity and 3D coherence throughout the 2D lifting optimization. To further improve the overall performance, we incorporate a step annealing strategy that modulates the impact of 3D guidance and 2D refinement. Therefore, our framework bridges the gap between 2D and 3D diffusion models, preserving multi-view coherent generation while maintaining the generalizability and fidelity of 2D models. Extensive experiments verify the remarkable improvement of our Sherpa3D on text-to-3D generation. + +Limitations and future works. Although our Sherpa3D achieves remarkable text-to-3D results, the quality still seems to be limited to the backbone itself as we choose Shap-E [28] and Stable Diffusion v2.1 base model in this work. We expect them to be solved with a larger diffusion model, such as SDXL [1] and DeepFloyd [2]. In future work, we are interested in extending our insight to more creative text-to-4D generation. We believe that Sherpa3D provides a promising research path for user-friendly and more accessible 3D content creation. + +Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant 62206147, and in part by Deng Feng Fund. + +# References + +[1] stable-diffusion-xl-base-1.0. https://huggingface.co/stabilityai/stable-diffusion-x1-base-1.0. Accessed:2023-08-29.8 +[2] Deepfloyd. https://huggingface.co/DeepFloyd. Accessed: 2023-08-25. 8 +[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, pages 40-49. PMLR, 2018. 3 +[4] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 3 +[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3 +[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 3 +[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3 +[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3 +[9] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023.5 +[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 2, 3, 5, 6, 8 +[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. 2 +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 + +[14] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5 +[15] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 3DV, pages 402-411. IEEE, 2017. 3 +[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurlPS, 35:31841-31854, 2022. 3 +[17] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 6(3):e30, 2021. 5 +[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3 +[19] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6 +[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 2, 3 +[21] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Gancraft: Unsupervised 3d neural rendering of northwest world. In ICCV, pages 14072-14082, 2021. 3 +[22] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In ICCV, pages 9984-9993, 2019. 3 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[25] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989, 2023. 2 +[26] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 5 +[27] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2, 6 +[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6, 7, 8 +[29] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5 + +[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[31] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6:124, 2011. 2 +[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4 +[33] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 5 +[34] Chenghao Li, Chaoning Zhang, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2 +[35] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2,4 +[36] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3 +[37] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 300–309, 2023. 2, 3, 6, 7, 8 +[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[39] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 3 +[40] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2, 3, 8 +[41] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3 +[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Sin- + +gle image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 3 +[43] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 3 +[44] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3 +[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 3 +[46] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5 +[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 3 +[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 3 +[49] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 3 +[50] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2 +[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3 +[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2, 3 +[53] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[54] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, pages 13503–13513, 2022. 3 +[55] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6 +[56] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The + +Eleventh International Conference on Learning Representations (ICLR), 2023. 2, 3, 6, 7, 8 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[58] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3 +[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3, 5, 6 +[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2, 3 +[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[63] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. NeurlPS, 35:33999-34011, 2022. 3 +[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2 +[65] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3 +[66] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3 +[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4 +[68] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea + +Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3 +[69] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3 +[71] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3 +[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 2, 3 +[73] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3 +[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2, 3, 6, 8 +[75] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2 +[76] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2 +[77] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 3 +[78] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 3 +[79] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6841–6850, 2023. 3 +[80] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3 +[81] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2 + +[82] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. arXiv preprint arXiv:2010.09125, 2020. 3 +[83] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461, 2023. 3 +[84] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 3 +[85] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3 \ No newline at end of file diff --git a/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/images.zip b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2bd1e22653a9193285571afb04f4a29eac400010 --- /dev/null +++ b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd7863af7d9c9fa89fafd2b39a47d9e215982c156d6f5d0a0830dc3d4865bbc9 +size 736707 diff --git a/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/layout.json b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..365492b69234c1cd51c190dc920378fbead579f0 --- /dev/null +++ b/2024/Sherpa3D_ Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior/layout.json @@ -0,0 +1,10149 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 537, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 537, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 537, + 121 + ], + "type": "text", + "content": "Sherpa3D: Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": "Fangfu Liu" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": ", Diankun Wu" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": ", Yi Wei" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": ", Yongming Rao" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": ", Yueqi Duan" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 142, + 462, + 172 + ], + "type": "text", + "content": ", Tsinghua University, BAAI" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 179, + 544, + 534 + ], + "blocks": [ + { + "bbox": [ + 49, + 179, + 544, + 534 + ], + "lines": [ + { + "bbox": [ + 49, + 179, + 544, + 534 + ], + "spans": [ + { + "bbox": [ + 49, + 179, + 544, + 534 + ], + "type": "image", + "image_path": "b8c3f96632c6ebab61211bbb34c256e0ab87fc38cead4172796f0e738cebb26c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 545, + 546, + 569 + ], + "lines": [ + { + "bbox": [ + 46, + 545, + 546, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 546, + 569 + ], + "type": "text", + "content": "Figure 1. Gallery of Sherpa3D: Blender rendering for various textured meshes from Sherpa3D, which is able to generate high-fidelity, diverse, and multi-view consistent 3D contents with input text prompts. Our method is also compatible with popular graphics engines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 583, + 192, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 583, + 192, + 596 + ], + "spans": [ + { + "bbox": [ + 143, + 583, + 192, + 596 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 609, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 288, + 694 + ], + "type": "text", + "content": "Recently, 3D content creation from text prompts has demonstrated remarkable progress by utilizing 2D and 3D diffusion models. While 3D diffusion models ensure great multi-view consistency, their ability to generate high-quality and diverse 3D assets is hindered by the limited 3D data. In contrast, 2D diffusion models find a distillation approach that achieves excellent generalization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 585, + 547, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 547, + 706 + ], + "type": "text", + "content": "and rich details without any 3D data. However, 2D lifting methods suffer from inherent view-agnostic ambiguity thereby leading to serious multi-face Janus issues, where text prompts fail to provide sufficient guidance to learn coherent 3D results. Instead of retraining a costly viewpoint-aware model, we study how to fully exploit easily accessible coarse 3D knowledge to enhance the prompts and guide 2D lifting optimization for refinement. In this paper, we propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 704, + 140, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 704, + 140, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 704, + 140, + 714 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20763" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "consistency simultaneously. Specifically, we design a pair of guiding strategies derived from the coarse 3D prior generated by the 3D diffusion model: a structural guidance for geometric fidelity and a semantic guidance for 3D coherence. Employing the two types of guidance, the 2D diffusion model enriches the 3D content with diversified and high-quality results. Extensive experiments show the superiority of our Sherpa3D over the state-of-the-art text-to-3D methods in terms of quality and 3D consistency. Project page: https://liuff19.github.io/Sherpa3D/." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 215, + 128, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 128, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 128, + 227 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 235, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 354 + ], + "type": "text", + "content": "3D content generation [34, 37, 56, 76] finds a broad range of applications, including games, movies, virtual/augmented reality and robots. However, the conventional process of creating premium 3D assets is still expensive and challenging as it requires multiple labor-intensive and time-consuming stages [31]. Fortunately, this challenge has prompted the development of recent text-to-3D methods [10, 25, 27, 35, 37, 47, 50, 56, 74]. Only using text prompts to automate 3D generation, these techniques pave a promising way towards streamlining 3D creation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 569 + ], + "type": "text", + "content": "Powered by the great breakthroughs in diffusion models [52, 60, 61, 81], two research lines of rationalization have recently emerged in text-to-3D: inference-only 3D diffusion methods and optimization-based 2D lifting methods. Specifically, the inference-only methods [20, 28, 52] seek to directly generate 3D-consistent assets by extensively training a new diffusion model on 3D data. However, due to the scarcity of 3D datasets compared to accessible 2D images or text data, these 3D diffusion models suffer from low quality and limited generalizability. Without requiring any 3D data for training, 2D lifting methods [4, 10, 37, 47, 56, 72, 74] can produce high-quality and diversified 3D results by distilling 3D knowledge from pre-trained 2D diffusion models [60, 61, 81], also known as Score Distillation Sampling (SDS). Yet lifting 2D observations into 3D is inherently ambiguous without sufficient 3D guidance from text prompts, leading to notorious multi-view inconsistency (e.g., Janus problems) in 2D lifting methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "These findings motivate us to think: is it possible to bridge the two aforementioned streams to achieve generalizability, high-fidelity, and geometric consistency simultaneously? An intuitive idea is to leverage more 3D data [11, 12] to fine-tune a view-point aware diffusion model, but it requires substantial computational resources and is prone to overfitting due to data bias [40, 67]. In contrast, our key insight is to utilize the easily accessible 3D diffusion model as guidance and study how to fully exploit coarse 3D knowledge to guide 2D lifting optimization for refinement. In particular, when maintaining the quality and generalizability of the original 2D diffusion model, we hope" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "the 2D lifting awareness can be guided by the strong 3D geometric information from the 3D diffusion model. However, it is non-trivial in pursuit of this balance. Relying too heavily on the coarse 3D priors from the 3D diffusion model may degrade the generation quality, whereas little 3D guidance could result in a lack of geometric awareness, leading to multi-view inconsistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 157, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 479 + ], + "type": "text", + "content": "Towards this end, we propose Sherpa3D in this paper, which greatly boosts high-fidelity and highly diversified text-to-3D generation with geometric consistency. Our method begins by employing a 3D diffusion model to craft a basic 3D guide with limited details. Building upon the coarse 3D prior, we introduce two guiding strategies to inform 2D diffusion model throughout lifting optimization: a structural guide for geometric fidelity and a semantic guide for 3D coherence. Specifically, the structural guide leverages the first-order gradient information of the normals from the 3D prior to supervise the optimization of the structure. These normals are then integrated into the input of a pre-trained 2D diffusion model, refining the geometric details. Concurrently, our semantic guide extracts high-level features from multi-views of the 3D prior. These features guide the 2D lifting optimization to perceive the geometric consistency under the preservation of original generalizability and quality. Furthermore, we design an annealing function, which modulates the influence of the 3D guidance to better preserve the capabilities of 2D and 3D diffusion models. As a result, our Sherpa3D is aware of the geometric consistency with rich details and generalizes well across diverse text prompts. Extensive experiments verify the efficacy of our framework and show that our Sherpa3D outperforms existing methods for high-fidelity and geometric consistency (see qualitative results gallery in Figure 1 and quantitative results in Table 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 491, + 392, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 392, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 392, + 503 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 511, + 449, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 449, + 524 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 449, + 524 + ], + "type": "text", + "content": "2.1. Text-to-image Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 530, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 662 + ], + "type": "text", + "content": "Recently, text-to-image models such as unCLIP [59], Imagegen [61], and Stable Diffusion [60] have shown remarkable capability of generating high-quality and creative images given text prompts. Such significant progress is powered by advances in diffusion models [13, 24, 53, 69], which can be pre-trained on billions of image-text pairs [62, 64] and understands general objects with complex semantic concepts (nouns, artistic styles, etc.) [60]. Despite the great success of photorealistic and diversified image generation, using language to generate different viewpoints of the same object with 3D coherence remains a challenging problem [75]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 670, + 434, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 434, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 434, + 682 + ], + "type": "text", + "content": "2.2. Text-to-3D Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Building on promising text-to-image diffusion models, there has been a surge of studies in text-to-3D generation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20764" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "content": "However, it is non-trivial due to the scarcity of diverse 3D data [8, 12, 77] compared to 2D. Existing 3D native diffusion models [20, 28, 43, 52, 80, 83] usually work on a limited object category and struggle with generating in-the-wild 3D assets. To achieve generalizable 3D generation, pioneering works DreamFusion [56] and SJC [72] propose to distill the score of image distribution from pre-trained 2D diffusion models [60, 61] and show impressive results. Following works [10, 26, 36, 37, 47, 70, 71, 74, 79, 84] continue to enhance various aspects such as generation fidelity and optimization stability or explore more application scenarios [58, 68, 85]. As it is inherently ambiguous to lift 2D observations into 3D, they may suffer from multi-face issues. Although some methods use prompt engineering [4] or train a costly viewpoint-aware model [40, 67] to alleviate such problems, they fail to generate high-quality results [10] or easily overfit to domain-specific data [12, 67]. In this work, we bridge the gap between 3D and 2D diffusion models through meticulously designed 3D guidance, which leads the 2D lifting process to achieve high-fidelity, diversified, and coherent 3D generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 330, + 174, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 174, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 174, + 341 + ], + "type": "text", + "content": "2.3. 3D Generative Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 348, + 289, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 289, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 289, + 661 + ], + "type": "text", + "content": "Extensive research has been conducted in the field of 3D generative modeling, exploring diverse 3D representations like 3D voxel grids [15, 22, 44], point clouds [3, 45, 49], and meshes [16, 82]. The majority of these approaches rely on training data presented in the form of 3D assets, which proves challenging to obtain at a large scale. Drawing inspiration from the success of neural volume rendering, recent studies have shifted towards investing in 3D-aware image synthesis [6, 7, 18, 21, 54, 63]. This approach offers the advantage of directly learning 3D generative models from images. However, volume rendering networks typically exhibit slow querying speeds, resulting in a trade-off between extended training times and a lack of multi-view consistency. Recently, benefitted from 2D diffusion models, some works generate multi-view images with single-view input [39-42, 66, 78]. As one of the pioneering works, Zero-1-to-3 [40] uses a synthetic dataset to fine-tune the pretrained diffusion models, aiming to learn controls of the relative camera viewpoint. Beyond Zero-1-to-3, SyncDreamer [41] employs a synchronized multiview diffusion model to capture the joint probability distribution of multiview images. This model facilitates the generation of multiview-consistent images through a unified reverse process. Different from these methods, we focus on text-to-3D synthesis, with the goal of generating multi-view consistent 3D contents with text prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 669, + 102, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 102, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 102, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Given a text prompt, our goal is to generate 3D assets with high quality, generalizability, and multi-view consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "Our framework can be divided into three stages: (1) build coarse 3D prior from the 3D diffusion model (Sec. 3.2); (2) formulate two guiding strategies (e.g., structural and semantic guidance) for 2D lifting process (Sec. 3.3); (3) incorporate both 3D guidance and SDS loss with an annealing technique in optimization and generate the final 3D object (Sec. 3.4). In this way, we can leverage the full power of state-of-the-art 3D and 2D diffusion models to obtain 3D coherence as 3D models, retaining intricate details and creative freedom as 2D models. Our pipeline is depicted in Figure 2. Before introducing our Sherpa3D in detail, we first review the theory of Score Distillation Sampling (SDS)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 223, + 392, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 223, + 392, + 234 + ], + "spans": [ + { + "bbox": [ + 306, + 223, + 392, + 234 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS). As one of the most representative 2D lifting methods, Dreamfusion [56] first presents the concept of Score Distillation Sampling (SDS), which is an algorithm to optimize a 3D representation such that the image rendered from any view maintains a high likelihood as evaluated by the 2D diffusion model given text prompts. SDS consists of two key components: (1) a 3D representation with parameter " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": ", which can produce an image " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": " at desired camera " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": " through a parametric function " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\theta; \\mathbf{c})" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": "; (2) a pre-trained text-to-image 2D diffusion model " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": " with a score function " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\mathbf{x}_t; y, t)" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": " that predicts the sample noise " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": " given noisy image " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": ", noise level " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": ". The score function guides the direction of the gradient for updating " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": " to reside rendered images in high-density areas conditioned on text " + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 240, + 547, + 433 + ], + "type": "text", + "content": ". The gradient is calculated by SDS as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 439, + 545, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 545, + 466 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\mathbf {x} _ {t}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {1}", + "image_path": "9cb5a234864404a13128483fc19cddce8e6026d7c6a0a43683044de6c8cccf66.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "content": " is a weighting function. In practice, the denoising score function " + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "content": " is often replaced with another function " + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_{\\phi}" + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "content": " that uses classifier-free guidance [23] that controls the strength of the text condition (see Supplementary)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 527, + 462, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 462, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 462, + 540 + ], + "type": "text", + "content": "3.2. Sculpting a Coarse 3D Prior" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "To facilitate text-to-3D generation, most existing methods [26, 47, 56] rely on implicit 3D representations such as Neural Radiance Fields (NeRF) [48] and its variants [5, 51]. However, it is difficult for NeRF-based modeling to extract the high-quality surface with material and texture [73]. To address this, we adopt the hybrid scene representation of DMTet [65], including a deformable tetrahedral grid that encodes a signed distance function (SDF) and a differentiable marching tetrahedra (MT) layer that extracts explicit surface mesh. Equipped with the hybrid representation, we sculpt a coarse 3D prior from 3D diffusion model " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "G_{3D}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " (e.g., Shap-E [28]) by the following procedure. Given text prompts " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ", we first use the 3D diffusion model " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "G_{3D}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " to generate 3D results " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and employ multi-layer perceptions" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20765" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 246 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 246 + ], + "type": "image", + "image_path": "a5d0d0087271a6944901772de5dcc1d22d6223fdaabab22d3023c32bf817b304.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 256, + 548, + 345 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 548, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 548, + 345 + ], + "type": "text", + "content": "Figure 2. Pipeline of our Sherpa3D. Given a text as input, we first prompt 3D diffusion to build a coarse 3D prior " + }, + { + "bbox": [ + 46, + 256, + 548, + 345 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 256, + 548, + 345 + ], + "type": "text", + "content": " encoded in the geometry model (e.g., DMTet). Next, we render the normal map of the extracted mesh in DMTet and derive two guiding strategies from " + }, + { + "bbox": [ + 46, + 256, + 548, + 345 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 256, + 548, + 345 + ], + "type": "text", + "content": ". (a) Structural Guidance: we utilize the structural descriptor to compute salient geometric features for preserving geometry fidelity (e.g., without a pockmarked face problem). (b) Semantic Guidance: we leverage a semantic encoder (e.g., CLIP) to extract high-level information for keeping 3D consistency (e.g., without multi-face issues). Employing the two guidance in 2D lifting process, we use the normal map as shape encoding of the 2D diffusion model and unleash its power to generate high-quality and diversified results with 3D coherence. Then we achieve the final 3D results via photorealistic rendering through appearance modeling. (\"Everest's summit eludes many without Sherpa.\")" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": "(MLPs) to query SDF values for each vertex along a regular grid. Next we sample a point set " + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\pmb {p}_i\\in \\mathbb{R}^3\\}" + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": " with their SDF values " + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\{SDF(\\pmb {p}_i)\\}" + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": ", the DMTet network " + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": " can predict SDF value " + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "inline_equation", + "content": "s(\\pmb {p}_i)" + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": ", and a position offset " + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "inline_equation", + "content": "\\Delta p_{i}" + }, + { + "bbox": [ + 47, + 356, + 288, + 416 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 426, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 426, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 426, + 287, + 441 + ], + "type": "interline_equation", + "content": "\\left(s \\left(\\boldsymbol {p} _ {i}\\right), \\Delta \\boldsymbol {p} _ {i}\\right) = \\mathcal {F} \\left(\\boldsymbol {p} _ {i}; \\theta\\right), \\tag {2}", + "image_path": "628395e998e8fe33993060af55b375179f434038b5c3892c64c713c29b6898c8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "text", + "content": " is the parameters of network " + }, + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "text", + "content": ". Then, we incorporate 3D priors into the DMTet network " + }, + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 47, + 450, + 287, + 487 + ], + "type": "text", + "content": " with the point set derived from 3D diffusion model by minimizing:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 495, + 287, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 495, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 52, + 495, + 287, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D F} = \\sum_ {\\boldsymbol {p} _ {i} \\in \\mathcal {P}} | s (\\boldsymbol {p} _ {i}) - S D F (\\boldsymbol {p} _ {i}) | ^ {2} + \\lambda_ {d e f} \\sum_ {\\boldsymbol {p} _ {i} \\in \\mathcal {P}} \\| \\Delta \\boldsymbol {p} _ {i} \\| _ {2}, \\tag {3}", + "image_path": "4009315b0f2d8fd50ca30211df6cff2e47253a35b41c0714214d55776230598a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\lambda_{def}" + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": " is the hyperparameter controlling " + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": " regularization strengths on offsets to avoid artifacts. Finally, we apply the MT layer to extract mesh representation " + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": ". Now, we have leveraged the knowledge from the 3D diffusion model to construct a coarse 3D prior, which is encoded implicitly in DMTet " + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": " and represented explicitly by mesh " + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": ". Next, we will discuss how to utilize the coarse 3D prior " + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": " as a guide during the subsequent 2D diffusion lifting optimization to refine a high-quality result with 3D coherence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 264, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 264, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 264, + 659 + ], + "type": "text", + "content": "3.3. 3D Guidance for 2D Lifting Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "What knowledge can serve as guidance? The purpose of introducing a 3D prior as guidance is to address the prevalent issue of viewpoint inconsistency both in geometry and appearance. Through empirical studies, we have identified" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 356, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 582 + ], + "type": "text", + "content": "geometric inconsistency as the main cause of 3D incoherence, leading to multi-face Janus problem [35, 67]. In contrast, appearance inconsistency emerges in much more extreme scenarios with lesser significance. Therefore, we disentangle the geometry from the 3D model and fully leverage coarse prior " + }, + { + "bbox": [ + 304, + 356, + 545, + 582 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 356, + 545, + 582 + ], + "type": "text", + "content": " to guide 2D lifting geometry optimization with view-point awareness. Our analysis of the coarse 3D prior indicates that it contains the essential geometric structures and captures the basic categorical attributes, keeping semantic rationality across different views. Building upon these observations, a natural insight is to preserve such inherent 3D knowledge as guidance and continuously benefit the 2D lifting process. For example, given text prompts \"a head of the Terracotta Army,\" we hope the knowledge in the guidance can prevent issues such as a pockmarked face or the unrealistic scenario of having a face on the back (e.g., Janus problem). To this end, we have designed two guiding strategies derived from " + }, + { + "bbox": [ + 304, + 356, + 545, + 582 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 356, + 545, + 582 + ], + "type": "text", + "content": ": structural guidance for geometric fidelity and semantic guidance for 3D coherence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": "Structural guidance. Given the current DMTet net " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": " that encodes the coarse 3D prior " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": ", we apply a differentiable render " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": " (e.g., nvidiffrast [32]) to generate a set of normal maps " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{\\pmb{n}_i|\\pmb{n}_i = f_n(\\mathcal{F}_{\\theta},\\mathbf{c}_i),i = 1,\\dots n\\}" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": " is the camera position randomly sampled in spherical coordinates. To extract the salient geometric structure features, we first use a Gaussian filter with a kernel standard deviation " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\sigma" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 693, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 693, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 372, + 693, + 545, + 717 + ], + "type": "interline_equation", + "content": "G (x, y) = \\frac {1}{2 \\pi \\sigma^ {2}} e ^ {- \\frac {x ^ {2} + y ^ {2}}{2 \\sigma^ {2}}} \\tag {4}", + "image_path": "10004a404b0cb619a9fe52a84cd5cbc6a16199b9441549dbbcbdfb1099df0343.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20766" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "content": "to reduce the noise impact and obtain " + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "\\{\\sigma (\\pmb {n}_i)\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "content": ". As gradients are simple but effective tools for revealing the geometric contours and salient structures [14, 29], we then compute the structural descriptor sets " + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "\\{G_{\\sigma}(\\pmb {n}_i)\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 122 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 130, + 287, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 85, + 130, + 287, + 163 + ], + "type": "interline_equation", + "content": "G _ {\\sigma} (\\boldsymbol {n} _ {i}) = \\sqrt {(\\frac {\\partial \\sigma (\\boldsymbol {n} _ {i})}{\\partial x}) ^ {2} + (\\frac {\\partial \\sigma (\\boldsymbol {n} _ {i})}{\\partial y}) ^ {2}}, \\tag {5}", + "image_path": "8694201a40e48c570d545017b324961a8c1e9ea5379302767d881af3677dbe83.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "text", + "content": " are the coordinate directions of the normal map " + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i" + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "text", + "content": ". Throughout the 2D lifting process of updating " + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "text", + "content": " with newly rendered normal maps " + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{N}} = \\{\\tilde{n}_i\\}" + }, + { + "bbox": [ + 47, + 171, + 288, + 221 + ], + "type": "text", + "content": ", it should follow the structural guidance as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 229, + 287, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 229, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 82, + 229, + 287, + 261 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} _ {\\text {s t r u c}} := \\sum_ {i = 1} ^ {n} \\left\\| G _ {\\sigma} \\left(\\boldsymbol {n} _ {i}\\right) - G _ {\\sigma} \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right) \\right\\| _ {2} ^ {2}, \\tag {6}", + "image_path": "721caaf2f2d3eede924f98caee539a4cc83256cfafb1474cb3500f32fd47556e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 270, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 305 + ], + "type": "text", + "content": "which enables the 2D lifting process to preserve geometric fidelity and a well-aligned structure with the coarse 3D prior when generating rich details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "text", + "content": "Semantic guidance. While structural guidance maintains low-level geometric perception from coarse 3D prior, semantic guidance extracts high-level features for 3D coherence. We first apply the pre-trained CLIP [57] model as semantic encoder " + }, + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "text", + "content": " to the normal set " + }, + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "text", + "content": " and obtain semantic feature maps " + }, + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c = \\{\\psi(\\boldsymbol{n}_i)\\}" + }, + { + "bbox": [ + 46, + 306, + 287, + 415 + ], + "type": "text", + "content": ", proven to effectively capture semantic attributes like facial expressions or view categories [17]. Following the notation as above, we then define the semantic guidance with cosine similarity:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 424, + 287, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 424, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 93, + 424, + 287, + 455 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} _ {\\text {s e m}} := \\sum_ {i = 1} ^ {n} \\frac {\\psi \\left(\\boldsymbol {n} _ {i}\\right) \\cdot \\psi \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right)}{\\| \\psi \\left(\\boldsymbol {n} _ {i}\\right) \\| \\| \\psi \\left(\\tilde {\\boldsymbol {n}} _ {i}\\right) \\|.} \\tag {7}", + "image_path": "18215ae54a6dfe4d3c9aa1bf3d3be59af84aadc0850286ea1ac911a107cdf941.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 464, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 287, + 525 + ], + "type": "text", + "content": "Employing this guidance, we ensure that different views retain inherent high-level information throughout the 2D lifting optimization process. Experiments show that it can effectively mitigate multi-face problems, keeping 3D content semantically plausible from all viewing angles." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 533, + 133, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 133, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 133, + 546 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "content": "In this subsection, we incorporate both structural and semantic guidance derived from coarse 3D prior to 2D lifting optimization so that it can produce vivid and diversified objects with multi-view consistency. For the disentangled geometry modeling, we use the randomly sampled normal map " + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "content": " as the input, bridging the gap between 3D and 2D diffusion. To update the geometry model DMTet network " + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "content": ", we choose to use the publicly available Stable Diffusion [60] as pre-trained 2D diffusion model " + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "content": " and compute the gradient of the SDS loss similar in Eq. 1:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 679, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 679, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 679, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\theta , \\boldsymbol {n}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\boldsymbol {z} _ {t} ^ {\\boldsymbol {n}}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\boldsymbol {z} _ {t} ^ {\\boldsymbol {n}}}{\\partial \\boldsymbol {n}} \\frac {\\partial \\boldsymbol {n}}{\\partial \\theta} \\right], \\tag {8}", + "image_path": "fc3ded7297696c11e8e8c33df2f8a911235b42dbdbbd9148d28b13a01d461f6d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "\\partial z_t^n /\\partial n" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": " calculates the gradient of the encoder in the latent diffusion model (LDM) [60]. Additionally, we introduce a step annealing technique to balance the influence of the 3D guidance during 2D lifting optimization:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 368, + 125, + 545, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 125, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 368, + 125, + 545, + 140 + ], + "type": "interline_equation", + "content": "\\gamma (\\lambda) = \\lambda e ^ {- \\beta \\max (0, n _ {\\mathrm {c u r}} - m)}, \\tag {9}", + "image_path": "235f7766bbe4c28d140434ddb7cae0925b40a204af190eb2d6f37d7d67fc937c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{cur}}" + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "text", + "content": " is the current epoch and " + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\{\\beta, m, \\lambda\\}" + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "text", + "content": " are the hyperparameters that control how " + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "text", + "content": " decreased. Therefore, the total loss " + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 304, + 145, + 545, + 194 + ], + "type": "text", + "content": " to lift 2D geometry optimization with 3D guidance is a weighted sum of three loss terms:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 201, + 545, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 201, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 201, + 545, + 215 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g e o}} (\\theta , \\boldsymbol {n}) = \\mathcal {L} _ {\\mathrm {S D S}} + \\gamma \\left(\\lambda_ {\\mathrm {s t r u c}}\\right) \\mathcal {L} _ {\\mathrm {s t r u c}} + \\gamma \\left(\\lambda_ {\\mathrm {s e m}}\\right) \\mathcal {L} _ {\\mathrm {s e m}}, \\tag {10}", + "image_path": "1848874e0657c10b75b6767300afd66aed21b5ec6010a5f68ee3b81c23768675.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": "which not only enables the 3D content generation without multi-view inconsistency issues but also preserves the generalization and quality in 2D diffusion model " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": ". As our pipeline can be integrated into any appearance model [9, 10, 33], we adopt a similar approach as Fantasia3D [10] to better align our text and 3D object. Denote " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": " as our appearance model, we have the rendered image " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\mathcal{T}_{\\eta}(\\mathcal{F}_{\\theta}, c_{i})" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": ". To update " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": ", we again apply the SDS loss for the final complete generated 3D object with detailed texture and coherent geometry:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 344, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 344, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 310, + 344, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\eta} \\mathcal {L} _ {\\mathrm {a p p}} (\\eta , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\phi} \\left(\\boldsymbol {z} _ {t} ^ {\\mathbf {x}}; y, t\\right) - \\epsilon\\right) \\frac {\\partial \\boldsymbol {z} _ {t} ^ {\\mathbf {x}}}{\\partial \\mathbf {x}} \\frac {\\partial \\mathbf {x}}{\\partial \\eta} \\right], \\tag {11}", + "image_path": "847636414f13d9acd2f2d44f8e06849c99fcff11512f5cf537b3139a721ab728.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 372, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 431 + ], + "type": "text", + "content": "which shares similar notations defined in Eq. 8. Finally, through the tailored 3D structural and semantic guidance that bridges the 2D and 3D diffusion models, our Sherpa3D can mitigate the multi-face problem and achieve high-fidelity and diversified results." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 437, + 440, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 440, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 440, + 451 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": "We apply the multilayer perceptron (MLP) comprising of three hidden layers to approximate " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\eta}" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": ". Adam optimizer [30] is used to update " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\eta}" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": " with an initial learning rates of " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": " decaying into " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": ". For 3D representations, we use textured mesh with a DMTet resolution of 128 to achieve a balance between quality and generation speed. We sample random camera poses at a fixed radius of 2.5, y-axis FOV of " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": ", with the azimuth in " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "[-180^{\\circ}, 180^{\\circ}]" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": " and elevation in " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "[-30^{\\circ}, 30^{\\circ}]" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": ". We load Shap-E from [46] for 3D diffusion model and choose stabilityai/stable-diffsuion-2-1-base [60] for 2D diffusion model. For weighting factors, we follow the same strategy as [26] to tune " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{struc}}" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": " is set to 10 and " + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sem}}" + }, + { + "bbox": [ + 304, + 456, + 545, + 661 + ], + "type": "text", + "content": " is 30 to balance the magnitude of SDS loss. Notably, our method only needs a single NVIDIA RTX3090 (24GB) GPU within 25 minutes. More details of optimization, architecture design, and hyperparameter settings can be found in the supplementary." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 669, + 388, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 388, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 388, + 683 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "In this section, we conduct comprehensive experiments to evaluate our text-to-3D framework Sherpa3D and show" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20767" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 521, + 417 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 521, + 417 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 521, + 417 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 521, + 417 + ], + "type": "image", + "image_path": "7d6ff6248e87c28dde9bb2fbbea9a9faedaab9b7317e158a271c995b649fc3b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons with baseline methods across different views " + }, + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "type": "inline_equation", + "content": "(0^{\\circ}" + }, + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "type": "inline_equation", + "content": "180^{\\circ})" + }, + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "type": "text", + "content": ". We can observe that baseline methods suffer from severe multi-face issues while our Sherpa3D can achieve better quality and 3D coherence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 468, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 289, + 552 + ], + "type": "text", + "content": "comparison results against other text-to-3D baseline methods. We first present qualitative results compared with five SOTA baselines from different viewpoints. Then we report the quantitative results with a user study. Finally, we carry out ablation studies to further verify the efficacy of our framework design. Please refer to the supplementary for more comparisons, visualizations, and detailed analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 562, + 155, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 155, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 155, + 575 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Baselines. We extensively compare our method Sherpa3D against five baselines: Shap-E [28], DreamFusion [56], Magic3D [37], ProlificDreamer [74], and Fantasis3D [10]. Due to various reasons, we can't obtain the original implementation of some baselines. For DreamFusion, Magic3D, and ProlificDreamer, we utilize their implementations in the Threestudio library [19] for comparison. For Shap-E and Fantasia3D, we follow their official implementation. We consider these implementations to be the most reliable and comprehensive open-source option available in the field. To ensure a fair comparison, we use the Stable Diffusion [60]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 468, + 465, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 465, + 480 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 465, + 480 + ], + "type": "text", + "content": "model as 2D diffusion prior by default." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 481, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 625 + ], + "type": "text", + "content": "Metrics. We will show our results with notable comparisons to other baselines through visualization. As there is no Ground-Truth 3D content corresponding to the text prompt, reference-based metrics like Chamfer Distance are difficult to apply to zero-shot text-to-3D generation. Following [27, 56], we evaluate the CLIP R-Precision [55], which can measure how well the rendered images of generated 3D content align with the input text. We use 100 prompts from the Common Objects in Context (COCO) dataset [38] as DreamFusion [56]. we also conduct a user study to further demonstrate the multi-view consistency and overall generation quality of our method," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 634, + 446, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 634, + 446, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 634, + 446, + 647 + ], + "type": "text", + "content": "4.2. Qualitative Comparisons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "We first demonstrate vivid and diversified text-to-3D results generated from our Sherpa3D in the gallery as shown in Figure 1. Then we compare our method with five baseline method: Shap-E [28], DreamFusion [56], Magic3D [37], Fantasia3D [10] and ProlificDreamer [74]. Figure 3 and 4" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20768" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 517, + 404 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 517, + 404 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 517, + 404 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 517, + 404 + ], + "type": "image", + "image_path": "32c1b7f6ece146318a53c2deb9f2af304970f0c375b5b1d6dc48b7a188d45f94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 414, + 477, + 426 + ], + "lines": [ + { + "bbox": [ + 115, + 414, + 477, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 477, + 426 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons with baseline methods across different views " + }, + { + "bbox": [ + 115, + 414, + 477, + 426 + ], + "type": "inline_equation", + "content": "(-30^{\\circ}" + }, + { + "bbox": [ + 115, + 414, + 477, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 115, + 414, + 477, + 426 + ], + "type": "inline_equation", + "content": "150^{\\circ})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 446, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 289, + 567 + ], + "type": "text", + "content": "give the comparative results with the same text prompt for each object generation. We observe that the Shap-E [28] only generates coarse shapes while other 2D lifting methods suffer from multi-face problems. In contrast, our Sherpa3D produces high-fidelity 3D assets with compelling texture quality and multi-view consistency. Notably, our framework is more efficient than other baselines with less time to optimize. Specifically, it only takes within 25 minutes from a text prompt to a high-quality 3D model ready to be used in graphic engines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 574, + 194, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 194, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 194, + 588 + ], + "type": "text", + "content": "4.3. Quantitative Comparisons" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "In Table 1, we report the CLIP R-Precision for Sherpa3D and several baselines. It shows that our method outperforms other baselines consistently across different CLIP models, and approaches the performance of ground truth (GT) images. For the user study, we render 360-degree rotating videos of 3D models generated from a collection of 120 images. Each volunteer is shown 10 samples of rendered video from a random method and rates in two aspects: multi-view consistency and overall generation quality. We collect results from 50 volunteers shown in Table 2. We observe" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 497, + 545, + 580 + ], + "blocks": [ + { + "bbox": [ + 305, + 445, + 545, + 490 + ], + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 490 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on generation renderings with text prompts using different CLIP retrieval models. We compared to ground-truth images, Shap-E [28], Dreamfusion [56], Magic3D [37], evaluated on object-centric COCO as in [56]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 497, + 545, + 580 + ], + "lines": [ + { + "bbox": [ + 309, + 497, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 309, + 497, + 545, + 580 + ], + "type": "table", + "html": "
MethodR-Precision (%) ↑
CLIP B/32CLIP B/16CLIP L/14
GT Images77.379.2-
Shape-E [28]41.142.546.4
DreamFusion [56]70.373.275.0
Magic3D [37]71.573.876.1
Sherpa3D (Ours)72.375.679.3
", + "image_path": "9502c457b05f78d972e5b140b2e1a61f85941e2fd328d7e8fdbbce9bcaef2dcd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 602, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 602, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 305, + 602, + 545, + 626 + ], + "type": "text", + "content": "that most users consider our results with much higher viewpoints consistency and overall generation fidelity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 634, + 463, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 463, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 463, + 647 + ], + "type": "text", + "content": "4.4. Ablation Study and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "We carry out ablation studies on the design of our Sherpa3D framework in Figure 5 using an example text prompt \"a head of the Terracotta Army\". Specifically, we perform ablation on three aspects of our method: structural guidance, semantic guidance, and the step annealing strategy. The re" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20769" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 75, + 104, + 144 + ], + "blocks": [ + { + "bbox": [ + 49, + 106, + 61, + 114 + ], + "lines": [ + { + "bbox": [ + 49, + 106, + 61, + 114 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 61, + 114 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 75, + 104, + 144 + ], + "lines": [ + { + "bbox": [ + 67, + 75, + 104, + 144 + ], + "spans": [ + { + "bbox": [ + 67, + 75, + 104, + 144 + ], + "type": "image", + "image_path": "0621e3cec3b6f38e082b6003b1dbebfc7010780acae19a872404a42fe7c600af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 65, + 147, + 108, + 211 + ], + "blocks": [ + { + "bbox": [ + 48, + 171, + 65, + 181 + ], + "lines": [ + { + "bbox": [ + 48, + 171, + 65, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 65, + 181 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 147, + 108, + 211 + ], + "lines": [ + { + "bbox": [ + 65, + 147, + 108, + 211 + ], + "spans": [ + { + "bbox": [ + 65, + 147, + 108, + 211 + ], + "type": "image", + "image_path": "5d532502f20d283fa6c957c729d2c36f6976412953337cb654564a775c702abe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 546, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 546, + 255 + ], + "type": "text", + "content": "Figure 5. Ablation study of our method. The generation is based on the text prompt \"a head of the Terracotta Army\". We ablate the design choices of structural guidance, semantic guidance (Sec. 3.3), and the step annealing technique (Sec. 3.4)." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 74, + 161, + 144 + ], + "blocks": [ + { + "bbox": [ + 135, + 74, + 161, + 144 + ], + "lines": [ + { + "bbox": [ + 135, + 74, + 161, + 144 + ], + "spans": [ + { + "bbox": [ + 135, + 74, + 161, + 144 + ], + "type": "image", + "image_path": "3a6437b306a74baebd0daaae842bd2fa2abf600edf6987a693fff13a3a4888f8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 126, + 147, + 170, + 210 + ], + "blocks": [ + { + "bbox": [ + 126, + 147, + 170, + 210 + ], + "lines": [ + { + "bbox": [ + 126, + 147, + 170, + 210 + ], + "spans": [ + { + "bbox": [ + 126, + 147, + 170, + 210 + ], + "type": "image", + "image_path": "5ce9efb194bca24646ce7a8547a9b70ce0084f61b01cb523fec302eb84c074da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 212, + 127, + 220 + ], + "lines": [ + { + "bbox": [ + 114, + 212, + 127, + 220 + ], + "spans": [ + { + "bbox": [ + 114, + 212, + 127, + 220 + ], + "type": "text", + "content": "Full" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 190, + 83, + 231, + 144 + ], + "blocks": [ + { + "bbox": [ + 195, + 74, + 224, + 81 + ], + "lines": [ + { + "bbox": [ + 195, + 74, + 224, + 81 + ], + "spans": [ + { + "bbox": [ + 195, + 74, + 224, + 81 + ], + "type": "text", + "content": "appearance" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 83, + 231, + 144 + ], + "lines": [ + { + "bbox": [ + 190, + 83, + 231, + 144 + ], + "spans": [ + { + "bbox": [ + 190, + 83, + 231, + 144 + ], + "type": "image", + "image_path": "0bb268e3aadc1079fc2d41590fc43c7fca663e82448599b1fa7ffc08dadb9c02.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 148, + 233, + 210 + ], + "blocks": [ + { + "bbox": [ + 190, + 148, + 233, + 210 + ], + "lines": [ + { + "bbox": [ + 190, + 148, + 233, + 210 + ], + "spans": [ + { + "bbox": [ + 190, + 148, + 233, + 210 + ], + "type": "image", + "image_path": "acd3e382ceab519f97c67094d081af11890c03013593f8301aa21ed422821d4a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 212, + 282, + 220 + ], + "lines": [ + { + "bbox": [ + 207, + 212, + 282, + 220 + ], + "spans": [ + { + "bbox": [ + 207, + 212, + 282, + 220 + ], + "type": "text", + "content": "w/o Structural Guidance" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 255, + 83, + 294, + 145 + ], + "blocks": [ + { + "bbox": [ + 258, + 75, + 282, + 81 + ], + "lines": [ + { + "bbox": [ + 258, + 75, + 282, + 81 + ], + "spans": [ + { + "bbox": [ + 258, + 75, + 282, + 81 + ], + "type": "text", + "content": "geometry" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 83, + 294, + 145 + ], + "lines": [ + { + "bbox": [ + 255, + 83, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 255, + 83, + 294, + 145 + ], + "type": "image", + "image_path": "a55cbdabe9a2eaa10b3612361317d091ff3b53cb162617d7fdea7b64dedcdd37.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 255, + 151, + 296, + 210 + ], + "blocks": [ + { + "bbox": [ + 255, + 151, + 296, + 210 + ], + "lines": [ + { + "bbox": [ + 255, + 151, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 255, + 151, + 296, + 210 + ], + "type": "image", + "image_path": "6bcce458f5808737039f5f7d0e383faefc4e04ac16487e2a694dd807de118ff2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 83, + 358, + 144 + ], + "blocks": [ + { + "bbox": [ + 323, + 75, + 352, + 82 + ], + "lines": [ + { + "bbox": [ + 323, + 75, + 352, + 82 + ], + "spans": [ + { + "bbox": [ + 323, + 75, + 352, + 82 + ], + "type": "text", + "content": "appearance" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 83, + 358, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 83, + 358, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 83, + 358, + 144 + ], + "type": "image", + "image_path": "515e8be1d68da195e576a82d90e9cb58f8cc7993bf98f15a7a12ae07292f0239.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 314, + 146, + 361, + 210 + ], + "blocks": [ + { + "bbox": [ + 314, + 146, + 361, + 210 + ], + "lines": [ + { + "bbox": [ + 314, + 146, + 361, + 210 + ], + "spans": [ + { + "bbox": [ + 314, + 146, + 361, + 210 + ], + "type": "image", + "image_path": "6efdc9de646c27c0e24b3715d1b810ecfc6a472c9cff292182dac52245b187e3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 212, + 407, + 220 + ], + "lines": [ + { + "bbox": [ + 333, + 212, + 407, + 220 + ], + "spans": [ + { + "bbox": [ + 333, + 212, + 407, + 220 + ], + "type": "text", + "content": "w/o Semantic Guidance" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 386, + 75, + 410, + 144 + ], + "blocks": [ + { + "bbox": [ + 386, + 75, + 410, + 144 + ], + "lines": [ + { + "bbox": [ + 386, + 75, + 410, + 144 + ], + "spans": [ + { + "bbox": [ + 386, + 75, + 410, + 144 + ], + "type": "image", + "image_path": "c12160555f8f2eaf666930b4047a12b562db56ee815e0f7c5abb4987d0360254.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 376, + 146, + 423, + 210 + ], + "blocks": [ + { + "bbox": [ + 376, + 146, + 423, + 210 + ], + "lines": [ + { + "bbox": [ + 376, + 146, + 423, + 210 + ], + "spans": [ + { + "bbox": [ + 376, + 146, + 423, + 210 + ], + "type": "image", + "image_path": "9391c2def14aaca71bd70c965cbd8fa888490440fd2006fade33619100de7ff9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 444, + 83, + 480, + 144 + ], + "blocks": [ + { + "bbox": [ + 445, + 75, + 473, + 81 + ], + "lines": [ + { + "bbox": [ + 445, + 75, + 473, + 81 + ], + "spans": [ + { + "bbox": [ + 445, + 75, + 473, + 81 + ], + "type": "text", + "content": "appearance" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 83, + 480, + 144 + ], + "lines": [ + { + "bbox": [ + 444, + 83, + 480, + 144 + ], + "spans": [ + { + "bbox": [ + 444, + 83, + 480, + 144 + ], + "type": "image", + "image_path": "b17721830ddb10af0cbe66dcfe5c8b1ed558b00a902d014c09a9a8a4049f6607.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 442, + 147, + 481, + 210 + ], + "blocks": [ + { + "bbox": [ + 442, + 147, + 481, + 210 + ], + "lines": [ + { + "bbox": [ + 442, + 147, + 481, + 210 + ], + "spans": [ + { + "bbox": [ + 442, + 147, + 481, + 210 + ], + "type": "image", + "image_path": "9f2b55d271bc0fe6744cc06b0b829d23a76c39101a7a377156d8953e6de1b03c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 212, + 526, + 220 + ], + "lines": [ + { + "bbox": [ + 468, + 212, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 468, + 212, + 526, + 220 + ], + "type": "text", + "content": "w/o step annealing" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 506, + 83, + 542, + 144 + ], + "blocks": [ + { + "bbox": [ + 509, + 75, + 533, + 82 + ], + "lines": [ + { + "bbox": [ + 509, + 75, + 533, + 82 + ], + "spans": [ + { + "bbox": [ + 509, + 75, + 533, + 82 + ], + "type": "text", + "content": "geometry" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 506, + 83, + 542, + 144 + ], + "lines": [ + { + "bbox": [ + 506, + 83, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 506, + 83, + 542, + 144 + ], + "type": "image", + "image_path": "d030b334723685502f8e5c2c480fc47c1aed397a586e49831506a498632d2e05.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 504, + 146, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 504, + 146, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 504, + 146, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 504, + 146, + 545, + 210 + ], + "type": "image", + "image_path": "588137a4dabf0da9dbd143790d45e7cfbef9055cf99cbe9b0eabe14661580629.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 50, + 316, + 287, + 388 + ], + "blocks": [ + { + "bbox": [ + 46, + 274, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 308 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on the multi-view consistency and overall generation quality score in a user study, rated on a scale of 1-10, with higher scores indicating better performance." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 316, + 287, + 388 + ], + "lines": [ + { + "bbox": [ + 50, + 316, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 50, + 316, + 287, + 388 + ], + "type": "table", + "html": "
MethodMulti-view Consistency ↑Overall Quality ↑
Shap-E [28]6.093.33
DreamFusion [56]4.586.05
Magic3D [37]5.256.73
Fantasia3D [10]3.835.90
ProlificDreamer [74]5.787.02
Sherpa3D(Ours)8.958.74
", + "image_path": "cf47d90964a50a1ed764411d3af336145bad1ddf546b477fa5112aed246d7530.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 407, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 552 + ], + "type": "text", + "content": "sults reveal that the omission of any of these elements leads to a degradation in terms of quality and consistency. Notably, the absence of structural guidance leads to a loss of geometric fidelity in the \"army\", leading to a pockmarked face; without semantic guidance, there's a loss of semantic rationality across different views, resulting in the multiview Janus problem. The lack of a balanced step annealing results in an excessive influence of guidance with a rough final output. This illustrates the effectiveness of our overall framework (Figure 2), which drives geometric fidelity, multi-view consistency, and optimization balance steered by the 3D guidance and annealing strategy." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "text", + "content": "To further demonstrate our generalizability, we compare our method in Figure 6 with the Zero123 [40] which uses more 3D data [40] to finetune a 2D diffusion model to be viewpoint-aware. However, such a finetuning-based method easily overfits to 3D training data and suffers from severe performance degradation with unseen input of the training set. In contrast, our method is more generalizable to open-vocabulary text prompts." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 47, + 657, + 119, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 119, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 119, + 670 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "In this paper, we present Sherpa3D, a new framework that simultaneously achieves high-quality, diversified, and 3D consistent text-to-3D generation. By fully exploiting eas" + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 307, + 284, + 356, + 342 + ], + "blocks": [ + { + "bbox": [ + 306, + 274, + 358, + 282 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 358, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 358, + 282 + ], + "type": "text", + "content": "Sherpa3D (Ours)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 284, + 356, + 342 + ], + "lines": [ + { + "bbox": [ + 307, + 284, + 356, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 356, + 342 + ], + "type": "image", + "image_path": "7c0d347e447eabfbaaf36cf29e02ba7e20bba97abe3b98204bc6c0a5a9e62498.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 344, + 435, + 357 + ], + "lines": [ + { + "bbox": [ + 317, + 344, + 435, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 344, + 435, + 357 + ], + "type": "text", + "content": "\"A deer draped in exquisite ancient Chinese fabric, oil painting style\"" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "type": "text", + "content": "Figure 6. Comparison with Zero123 [40]. We use the front view of our generated 3D model as the input of Zero123 with open-vocabulary text prompts." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 372, + 286, + 423, + 339 + ], + "blocks": [ + { + "bbox": [ + 380, + 274, + 405, + 282 + ], + "lines": [ + { + "bbox": [ + 380, + 274, + 405, + 282 + ], + "spans": [ + { + "bbox": [ + 380, + 274, + 405, + 282 + ], + "type": "text", + "content": "Zero123" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 286, + 423, + 339 + ], + "lines": [ + { + "bbox": [ + 372, + 286, + 423, + 339 + ], + "spans": [ + { + "bbox": [ + 372, + 286, + 423, + 339 + ], + "type": "image", + "image_path": "bc971454719d6accdf8ac61ae0a3297419f96e83ef13edeb02b77d1da4e55052.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 443, + 287, + 474, + 336 + ], + "blocks": [ + { + "bbox": [ + 439, + 274, + 489, + 283 + ], + "lines": [ + { + "bbox": [ + 439, + 274, + 489, + 283 + ], + "spans": [ + { + "bbox": [ + 439, + 274, + 489, + 283 + ], + "type": "text", + "content": "Sherpa3D (Ours)" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 287, + 474, + 336 + ], + "lines": [ + { + "bbox": [ + 443, + 287, + 474, + 336 + ], + "spans": [ + { + "bbox": [ + 443, + 287, + 474, + 336 + ], + "type": "image", + "image_path": "a8b4c2d0b228ecdbd7e2b490f92777f8f240dd187890f4ab26edc4d6815366a6.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 344, + 520, + 358 + ], + "lines": [ + { + "bbox": [ + 457, + 344, + 520, + 358 + ], + "spans": [ + { + "bbox": [ + 457, + 344, + 520, + 358 + ], + "type": "text", + "content": "\"A head of the Terracotta Army\"" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 498, + 296, + 545, + 336 + ], + "blocks": [ + { + "bbox": [ + 511, + 274, + 537, + 282 + ], + "lines": [ + { + "bbox": [ + 511, + 274, + 537, + 282 + ], + "spans": [ + { + "bbox": [ + 511, + 274, + 537, + 282 + ], + "type": "text", + "content": "Zero123" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 296, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 498, + 296, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 498, + 296, + 545, + 336 + ], + "type": "image", + "image_path": "7bb17022eb364810a80b375f803a624dc71d37f5f790542a481ec2cf325097dd.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 414, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 557 + ], + "type": "text", + "content": "ily obtained coarse 3D knowledge from the 3D diffusion model, we derive structural guidance and semantic guidance to enhance the prompts and provide continuous guidance with geometric fidelity and 3D coherence throughout the 2D lifting optimization. To further improve the overall performance, we incorporate a step annealing strategy that modulates the impact of 3D guidance and 2D refinement. Therefore, our framework bridges the gap between 2D and 3D diffusion models, preserving multi-view coherent generation while maintaining the generalizability and fidelity of 2D models. Extensive experiments verify the remarkable improvement of our Sherpa3D on text-to-3D generation." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "content": "Limitations and future works. Although our Sherpa3D achieves remarkable text-to-3D results, the quality still seems to be limited to the backbone itself as we choose Shap-E [28] and Stable Diffusion v2.1 base model in this work. We expect them to be solved with a larger diffusion model, such as SDXL [1] and DeepFloyd [2]. In future work, we are interested in extending our insight to more creative text-to-4D generation. We believe that Sherpa3D provides a promising research path for user-friendly and more accessible 3D content creation." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant 62206147, and in part by Deng Feng Fund." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20770" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] stable-diffusion-xl-base-1.0. https://huggingface.co/stabilityai/stable-diffusion-x1-base-1.0. Accessed:2023-08-29.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 145 + ], + "type": "text", + "content": "[2] Deepfloyd. https://huggingface.co/DeepFloyd. Accessed: 2023-08-25. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 190 + ], + "type": "text", + "content": "[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, pages 40-49. PMLR, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 286, + 246 + ], + "type": "text", + "content": "[4] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 312 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 357 + ], + "type": "text", + "content": "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 413 + ], + "type": "text", + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 286, + 468 + ], + "type": "text", + "content": "[8] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 286, + 512 + ], + "type": "text", + "content": "[9] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "text", + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "text", + "content": "[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 679 + ], + "type": "text", + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[14] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 127 + ], + "type": "text", + "content": "[15] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In 3DV, pages 402-411. IEEE, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "text", + "content": "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurlPS, 35:31841-31854, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[17] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 6(3):e30, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 337 + ], + "type": "text", + "content": "[19] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "type": "text", + "content": "[21] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Gancraft: Unsupervised 3d neural rendering of northwest world. In ICCV, pages 14072-14082, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "text", + "content": "[22] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In ICCV, pages 9984-9993, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 459 + ], + "type": "text", + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[25] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[26] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[27] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "text", + "content": "[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[29] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20771" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": "[31] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6:124, 2011. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "text", + "content": "[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "text", + "content": "[33] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "text", + "content": "[34] Chenghao Li, Chaoning Zhang, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "type": "text", + "content": "[35] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 2,4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "text", + "content": "[36] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 466 + ], + "type": "text", + "content": "[37] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 300–309, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 533 + ], + "type": "text", + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[39] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 635 + ], + "type": "text", + "content": "[40] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2, 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[41] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Sin-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "gle image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[43] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[44] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "type": "text", + "content": "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 547, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 264 + ], + "type": "text", + "content": "[46] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 547, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 547, + 320 + ], + "type": "text", + "content": "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 547, + 376 + ], + "type": "text", + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 547, + 421 + ], + "type": "text", + "content": "[49] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 547, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 466 + ], + "type": "text", + "content": "[50] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "type": "text", + "content": "[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 512, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 547, + 544 + ], + "type": "text", + "content": "[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 547, + 589 + ], + "type": "text", + "content": "[53] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 635 + ], + "type": "text", + "content": "[54] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, pages 13503–13513, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "type": "text", + "content": "[55] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[56] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20772" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Eleventh International Conference on Learning Representations (ICLR), 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 218 + ], + "type": "text", + "content": "[58] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 374 + ], + "type": "text", + "content": "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 442 + ], + "type": "text", + "content": "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 477 + ], + "type": "text", + "content": "[63] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. NeurlPS, 35:33999-34011, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 543 + ], + "type": "text", + "content": "[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[65] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[66] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 657, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 690 + ], + "type": "text", + "content": "[67] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[68] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[69] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[70] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[71] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "type": "text", + "content": "[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "type": "text", + "content": "[73] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "text", + "content": "[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 366, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 410 + ], + "type": "text", + "content": "[75] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[76] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "type": "text", + "content": "[77] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[78] Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hongdong Li. Consistnet: Enforcing 3d consistency for multiview images diffusion. arXiv preprint arXiv:2310.10343, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "text", + "content": "[79] Chaohui Yu, Qiang Zhou, Jingliang Li, Zhe Zhang, Zhibin Wang, and Fan Wang. Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6841–6850, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[80] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[81] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20773" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 240 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[82] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yinan Zhang, Antonio Torralba, and Sanja Fidler. Image gans meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. arXiv preprint arXiv:2010.09125, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[83] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 205 + ], + "type": "text", + "content": "[84] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "type": "text", + "content": "[85] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20774" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_content_list.json b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bb1505722998733d4d61580a812743481036bb74 --- /dev/null +++ b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_content_list.json @@ -0,0 +1,1807 @@ +[ + { + "type": "text", + "text": "SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion", + "text_level": 1, + "bbox": [ + 222, + 130, + 746, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hsuan-I Ho", + "bbox": [ + 251, + 203, + 349, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jie Song", + "bbox": [ + 426, + 203, + 500, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Otmar Hilliges", + "bbox": [ + 575, + 203, + 696, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, ETH Zürich", + "bbox": [ + 299, + 229, + 671, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ait.ethz.ch/sith", + "bbox": [ + 374, + 251, + 589, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c22fb8934c9c87c6cb9274b9744967f2112240041104535705c47084e1e5ae9.jpg", + "image_caption": [ + "Real Photo" + ], + "image_footnote": [], + "bbox": [ + 96, + 301, + 197, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/11ae73ac30fb082476415d5050abcc1f16b7882d4caa311255e0347ffc497ca1.jpg", + "image_caption": [ + "Back-view Hallucination" + ], + "image_footnote": [], + "bbox": [ + 205, + 306, + 352, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fae3d85f80dfe4f4d8c6753ce1922658ee8e849510dd406bd2944197ec4ffe81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 290, + 467, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de9dc06b1659efad7fed662abd7483f403cb744c7bf87492e3b62c48f7593acd.jpg", + "image_caption": [ + "Mesh Reconstruction", + "AI-generated Image", + "Figure 1. Single-view textured human reconstruction. SiTH is a novel pipeline for creating high-quality and fully textured 3D human meshes from single images. We first hallucinate back-view appearances through an image-conditioned diffusion model, followed by the reconstruction of full-body textured meshes using both the front and back-view images. Our pipeline enables the creation of lifelike and diverse 3D humans from unseen photos (left) and AI-generated images (right)." + ], + "image_footnote": [], + "bbox": [ + 362, + 380, + 467, + 464 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7b2fea9a0ad8bfd1fe7e46abd9ffdc7432b798ab74d9a7e8afa3d064d91a1a8.jpg", + "image_caption": [ + "\"A woman in a white T-shirt and a blue tennis skirt\"" + ], + "image_footnote": [], + "bbox": [ + 498, + 303, + 738, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74a63743f32178e548ce7612b440d25fd45baf500a3cbafc46a9b9db40945de7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 290, + 852, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/801585909fa333b7f9bd7ad9dbd445032270cfb27b65bef790b7bb0d11dcb800.jpg", + "image_caption": [ + "e Back-view Hallucination", + "Mesh Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 751, + 378, + 851, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 556, + 313, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A long-standing goal of 3D human reconstruction is to create lifelike and fully detailed 3D humans from single-view images. The main challenge lies in inferring unknown body shapes, appearances, and clothing details in areas not visible in the images. To address this, we propose SiTH, a novel pipeline that uniquely integrates an image-conditioned diffusion model into a 3D mesh reconstruction workflow. At the core of our method lies the decomposition of the challenging single-view reconstruction problem into generative hallucination and reconstruction subproblems. For the former, we employ a powerful generative diffusion model to hallucinate unseen back-view appearance based on the input images. For the latter, we leverage skinned body meshes as guidance to recover full-body texture meshes from the input and back-view images. SiTH requires as few as 500 3D human scans for training while maintaining its generality and robustness to diverse images. Extensive evaluations on two 3D human benchmarks, including our newly created one, highlighted our method's superior accuracy and perceptual quality in 3D textured human reconstruction.", + "bbox": [ + 73, + 593, + 472, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 556, + 630, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the growing popularity of 3D and virtual reality applications, there has been increasing interest in creating realistic 3D human models. In general, crafting 3D humans is labor-intensive, time-consuming, and requires collaboration from highly skilled professionals. To bring lifelike 3D humans to reality and to support both expert and amateur creators in this task, it is essential to enable users to create textured 3D humans from simple 2D images or photos.", + "bbox": [ + 496, + 582, + 890, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing a fully textured human mesh from a single-view image presents an ill-posed problem with two major challenges. Firstly, the appearance information required for generating texture in unobserved regions is missing. Secondly, 3D information for mesh reconstruction, such as depth, surface, and body pose, becomes ambiguous in a 2D image. Previous efforts [4, 60, 79] attempted to tackle these challenges in a data-driven manner, focusing on training neural networks with image-mesh pairs. However, these approaches struggle with images featuring unseen appearances or poses, due to limited 3D human training data. More recent studies [61, 73, 79] introduced additional 3D reasoning modules to enhance robustness against unseen", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "538", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "poses. Yet, generating realistic and full-body textures from unseen appearances still remains an unsolved problem.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, we propose SiTH, a novel pipeline that integrates an image-conditioned diffusion model to reconstruct lifelike 3D textured humans from monocular images. At the core of our approach is the decomposition of the challenging single-view problem into two subproblems: generative back-view hallucination and mesh reconstruction. This decomposition enables us to exploit the generative capability of pretrained diffusion models to guide full-body mesh and texture reconstruction. The workflow is depicted in Fig. 1. Given a front-view image, the first stage involves hallucinating a perceptually consistent back-view image using image-conditioned diffusion. The second stage reconstructs full-body mesh and texture, utilizing both the front and back-view images as guidance.", + "bbox": [ + 75, + 125, + 468, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, we employ the generative capabilities of pretrained diffusion models (e.g. Stable Diffusion [57]) to infer unobserved back-view appearances for full-body 3D reconstruction. The primary challenge in ensuring the realism of 3D meshes lies in generating images that depict spatially aligned body shapes and perceptually consistent appearances with the input images. While diffusion models demonstrate impressive generative abilities with text conditioning, they are limited in producing desired back-view images using the frontal images as image conditions. To overcome this, we adapt the network architecture to enable conditioning on frontal images and introduce additional trainable components following ControlNet [77] to provide pose and mask control. To fully tailor this model to our task while retaining its original generative power, we carefully fine-tune the diffusion model using multi-view images rendered from 3D human scans. Complementing this generative model, we develop a mesh reconstruction module to recover full-body textured mesh from front and back-view images. We follow prior work in handling 3D ambiguity through normal [61] and skinned body [73, 79] guidance. It is worth noting that the models for both subproblems are trained using the same public THuman2.0 [75] dataset, which consists of as few as 500 scans.", + "bbox": [ + 75, + 339, + 470, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To advance research in single-view human reconstruction, we created a new benchmark based on the high-quality CustomHumans [22] dataset and conducted comprehensive evaluations against state-of-the-art methods. Compared to existing end-to-end methods [4, 60, 79], our two-stage pipeline can recover full-body textured meshes, including back-view details, and demonstrates robustness to unseen images. In contrast to time-intensive diffusion-based optimization methods [24, 35, 41], our pipeline efficiently produces high-quality textured meshes in under two minutes. Moreover, we explored applications combining text-guided diffusion models, showing SiTH's versatility in 3D human creation. Our contributions are summarized as follows:", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SiTH, a single-view human reconstruction pipeline capable of producing high-quality, fully textured 3D human meshes within two minutes.", + "- Through decomposing the single-view reconstruction task, SiTH can be efficiently trained with public 3D human scans and is more robust to unseen images.", + "- We establish a new benchmark featuring more diverse subjects for evaluating textured human reconstruction." + ], + "bbox": [ + 500, + 90, + 890, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 229, + 640, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-view human mesh reconstruction. Reconstructing 3D humans from monocular inputs [14, 17, 26, 28, 29, 62, 67, 71] has gained more popularity in research. In this context, we focus on methods that recover 3D human shapes, garments, and textures from a single image. As a seminal work, Saito et al. [60] first proposed a data-driven method with pixel-aligned features and neural fields [72]. Its follow-up work PIFuHD [61] further improved this framework with high-res normal guidance. Later approaches extended this framework with additional human body priors. For instance, PaMIR [79] and ICON [73] utilized skinned body models [42, 51] to guide 3D reconstruction. ARCH [25], $\\mathrm{ARCH} + +$ [21], and CAR [39] transformed global coordinates into the canonical coordinates to allow for reposing. PHOHRUM [4] and S3F [12] further disentangled shading and albedo to enable relighting. Another line of work replaced the neural representations with conventional Poisson surface reconstruction [33, 34]. ECON [74] and 2K2K [18] trained normal and depth predictors to generate front and back 2.5D point clouds. The human mesh is obtained by fusing these point clouds with body priors and 3D heuristics. However, none of these methods produce realistic full-body texture and geometry in the unobserved regions. Our pipeline addresses this problem by incorporating a generative diffusion model into the 3D human reconstruction workflow.", + "bbox": [ + 496, + 250, + 890, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D generation with 2D diffusion models. Diffusion models [52, 56, 57, 59] trained with large collections of images have demonstrated unprecedented capability in creating 3D objects from text prompts. Most prior work [11, 19, 40, 47, 53, 68] followed an optimization workflow to update 3D representations (e.g. NeRF [48], SDF tetrahedron [63]) via neural rendering [66] and a score distillation sampling (SDS) [53] loss. While some methods [3, 10, 24, 27, 64, 76] applied this workflow to human bodies, they cannot produce accurate human bodies and appearances due to the ambiguity of text-conditioning. More recent work [41, 54] also tried to extend this workflow with more accurate image-conditioning. However, we show that they struggle to recover human clothing details and require a long optimization time. Most related to our work is Chupa [35], which also decomposes its pipeline into two stages. Note that Chupa is an optimization-based approach that relies on texts", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "539", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ecbf5bd8e7652b382bdf6e6652f9c9a93828a0744e17fc309b49a8cbbb915d4.jpg", + "image_caption": [ + "Figure 2. Method overview. SiTH is a two-stage pipeline composed of back-view hallucination and mesh reconstruction. The back-view hallucination module samples perceptually consistent back-view images through an iterative denoising process conditioned on the input image, UV map, and silhouette mask (Sec. 3.1). Based on the input and generated back-view images, the mesh reconstruction module recovers a full-body mesh and textures leveraging a skinned body prior as guidance (Sec. 3.2). Note that both modules in the pipeline can be trained with the same public 3D human dataset and generalize unseen images." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 883, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and cannot model colors. We address these issues by introducing an image-conditioning strategy and model. Most importantly, our method swiftly reconstructs full-texture human meshes without any optimization process.", + "bbox": [ + 75, + 388, + 468, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models adaptation. Foundation models [8, 13, 20, 36] trained on large-scale datasets have been shown to be adaptable to various downstream tasks. Following this trend, pretrained diffusion models [52, 56, 57, 59] have become common backbones for generative modeling. For instance, they can be customized by finetuning with a small collection of images [23, 37, 58]. ControlNet [77] introduced additional trainable plugins to enable image conditioning such as body skeletons. While these strategies have been widely adopted, none of them directly fit our objective. More relevant to our task is DreamPose [31], which utilizes DensePose [16] images as conditions to repose input images. However, it cannot handle out-of-distribution images due to overfitting. Similarly, Zero-1-to-3 [41] finetunes a diffusion model with multi-view images to allow for viewpoint control. However, we show that viewpoint conditioning is not sufficient for generating consistent human bodies. Our model addresses this issue by providing accurate body pose and mask conditions for back-view hallucination.", + "bbox": [ + 75, + 450, + 470, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 744, + 210, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Method overview. Given an input image of a human body and estimated SMPL-X [51] parameters, SiTH produces a full-body textured mesh. This mesh not only captures the observed appearances but also recovers geometric and textural details in unseen regions, such as clothing wrinkles on the back. The pipeline is composed of two modules and is summarized in Fig. 2. In the first stage, we hallucinate unobserved appearances leveraging the generative power of an image-conditioned diffusion model (Sec. 3.1). In the second", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "stage, we reconstruct a full-body textured mesh given the input front-view image and the generated back-view image as guidance (Sec. 3.2). Notably, both modules are efficiently trained with 500 textured human scans in THuman2.0 [75].", + "bbox": [ + 496, + 388, + 892, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Back-view Hallucination", + "text_level": 1, + "bbox": [ + 498, + 455, + 725, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. Given an input front-view image $I^{F} \\in \\mathbb{R}^{H \\times W \\times 3}$ , our goal is to infer a back-view image $I^{B} \\in \\mathbb{R}^{H \\times W \\times 3}$ which depicts unobserved body appearances. This task is under-constrained since there are multiple possible solutions to the same input images. Taking this perspective into account, we leverage a latent diffusion model (LDM) [57] to learn a conditional distribution of back-view images given a front-view image. First, a VAE autoencoder, consisting of an encoder $\\mathcal{E}$ and a decoder $\\mathcal{D}$ , is pretrained on a corpus of 2D natural images through image reconstruction, i.e., $\\tilde{I} = \\mathcal{D}(\\mathcal{E}(I))$ . Afterwards, an LDM learns to produce a latent code $z$ within the VAE latent distribution $z = \\mathcal{E}(I)$ from randomly sampled noise. To sample an image, a latent code $\\tilde{z}$ is obtained by iteratively denoising Gaussian noise. The final image is reconstructed through the decoder, i.e., $\\tilde{I} = \\mathcal{D}(\\tilde{z})$ .", + "bbox": [ + 496, + 472, + 892, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-conditioned diffusion model. Simply applying the LDM architecture to our task is not sufficient since our goal is to learn a conditional distribution of back-view images given an input conditional image. To this end, we make several adaptations to allow for image-conditioning as shown in Fig. 3. First, we utilize the pretrained CLIP [55] image encoder and VAE encoder $\\mathcal{E}$ to extract image features from the front-view image (i.e., $I^{F}$ ). These image features are used for conditioning the LDM, ensuring the output image shares a consistent appearance with the input image. Second, we follow the idea of ControlNet [77] and propose to use a UV map $(I_{UV}^{B} \\in \\mathbb{R}^{H \\times W \\times 3})$ and a silhou", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "540", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2067ee22552ec0b33f3bce63075eb345743003e883df4eadbb64da2f4bd71c9f.jpg", + "image_caption": [ + "Figure 3. Training of back-view hallucination module. We employ a pretrained LDM and ControlNet architecture to enable image conditioning. To train our model, we render training pairs of conditional images $I^{F}$ and ground-truth images $I^{B}$ from 3D human scans. Given a noisy image latent $z_{t}$ , the model predicts added noise $\\epsilon$ given the conditional image $I^{F}$ , UV map $I_{UV}^{B}$ , and mask $I_{M}^{B}$ as conditions. We train the ControlNet model and cross-attention layers while keeping other parameters frozen." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 467, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ette mask $(I_M^B\\in \\mathbb{R}^{H\\times W})$ from the back view as additional conditions. These conditional signals provide additional information that ensures the output image has a similar body shape and pose to the conditional input image.", + "bbox": [ + 75, + 415, + 468, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning hallucination from pretraining. Another challenge in training an image-conditioned LDM is data. Training the model from scratch is infeasible due to the requirement of a large number of paired images rendered from 3D textured human scans. Inspired by the concept of learning from large-scale pretraining [13, 20], we build our image-conditioned LDM on top of a pretrained diffusion U-Net [57]. We utilize the finetuning strategy [37, 77] to optimize cross-attention layers and ControlNet parameters while keeping most of the other parameters frozen (see Fig. 3). The design and training strategy of our image-conditioned diffusion model enables hallucinating plausible back-view images that are cosistent with the frontal inputs.", + "bbox": [ + 75, + 479, + 472, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training and inference. To generate pairwise training images from 3D human scans, we sample camera view angles and use orthographic projection to render RGBA images from 3D scans and UV maps from their SMPL-X fits. Given a pair of images rendered by a frontal and its corresponding back camera, the first image serves as the conditional input $I^{F}$ while the other one is the ground-truth image $I^{B}$ . During training, the ground-truth latent code $z_{0} = \\mathcal{E}(I^{B})$ is perturbed by the diffusion process in t time steps, resulting in a noisy latent $z_{t}$ . The image-conditioned LDM model $\\epsilon_{\\theta}$ aims to predict the added noise $\\epsilon$ given the noisy latent $z_{t}$ , the time step $t \\sim [0,1000]$ , the conditional image $I^{F}$ , the silhouette mask $I_{M}^{B}$ , and the UV map $I_{UV}^{B}$ (See Fig. 3). The objective function for fine-tuning can be", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8b77da9a94f91b4745165cdb3f3c431a3d3a5c8fb7de9ee1ee9816e5081cf82e.jpg", + "image_caption": [ + "Figure 4. Mesh reconstruction module. Given front and back-view images $(I^{F}, I^{B})$ we predict their normal images $(N^{F}, N^{B})$ through a learned normal predictor. A 3D point $\\mathbf{x}$ is projected onto these images for querying pixel-aligned features $(\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x})$ . To leverage human body mesh as guidance, we embed the point $\\mathbf{x}$ into the local UV coordinates $\\mathbf{u}_c$ , vector $\\mathbf{n_c}$ , distance $d_c$ , and visibility $v_c$ . Finally, two decoders $(H_d, H_r)$ predict SDF and RGB values at $\\mathbf{x}$ given the positional embedding and pixel-aligned features." + ], + "image_footnote": [], + "bbox": [ + 542, + 93, + 846, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "represented as:", + "bbox": [ + 500, + 426, + 602, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {z \\sim \\mathcal {E} (I), t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\\right) \\| _ {2} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 454, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At test time, we obtain $I_{UV}^{B}, I_{M}^{B}$ from an off-the-shelf pose predictor [9] and segmentation model [36]. To infer a back-view image, we sample a latent $\\tilde{z}_0$ by performing the iterative denoising process starting from a Gaussian noise $z_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The back-view image can be obtained by:", + "bbox": [ + 496, + 491, + 893, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {I} ^ {B} = \\mathcal {D} (\\tilde {z} _ {0}) = \\mathcal {D} \\left(f _ {\\theta} \\left(z _ {T}, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 580, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta}$ is a function representing the iterative denoising process of our image-conditioned LDM (See Fig. 2 left).", + "bbox": [ + 496, + 611, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Human Mesh Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 654, + 763, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the back-view image, our goal is to construct a full-body human mesh and its textures using the input and back-view image as guidance. We follow the literature [60, 61] to model this task with a data-driven method. Given pairwise training data (i.e., front/back-view images and 3D scans), we learn a data-driven model that maps these images to a 3D representation (e.g., a signed distance field (SDF)). We define this mapping as below:", + "bbox": [ + 496, + 672, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Phi : \\mathbb {R} ^ {H \\times W \\times 3} \\times \\mathbb {R} ^ {H \\times W \\times 3} \\times \\mathbb {R} ^ {3} \\rightarrow \\mathbb {R} \\times \\mathbb {R} ^ {3} \\tag {3} \\\\ (I ^ {F}, I ^ {B}, \\mathbf {x}) \\mapsto d _ {x}, \\mathbf {r} _ {x}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 804, + 890, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is the 3D coordinate of a query point, and $d_x, \\mathbf{r}_x$ denote the signed distance and RGB color value at point $\\mathbf{x}$ . The network components we used for learning the mapping function are depicted in Fig. 4.", + "bbox": [ + 496, + 856, + 893, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "541", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b1ebc1e3bb2f1b8e8dbe76773adfb901a8c15154e5c7b9c351d313e6e735471.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 83, + 99, + 163, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51b68c462af8841dc78e08284c8c6bf4b6a0c2cacab6f92fbd9dae42cac2c4f3.jpg", + "image_caption": [ + "PaMIR" + ], + "image_footnote": [], + "bbox": [ + 176, + 102, + 356, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f2f38dfba8c3297aab94ddfe51f92f9f62ae4bc3026f994d4c0ba0343781720.jpg", + "image_caption": [ + "PHORHUM" + ], + "image_footnote": [], + "bbox": [ + 359, + 102, + 535, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b2fc4efbbc189511464cfe449720b5b54869f3fcbc7a7e4c15961bfdedbe1f76.jpg", + "image_caption": [ + "SiTH (Ours)" + ], + "image_footnote": [], + "bbox": [ + 539, + 102, + 707, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/213325818320f1edd3af26deb3df9d5751bec21f3b7d9c7146f1451dce178187.jpg", + "image_caption": [ + "GT References" + ], + "image_footnote": [], + "bbox": [ + 710, + 102, + 883, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ca033b379b2315af5852f306aa9940595d3877c11d403404dc131f21aa8f3dd.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 81, + 247, + 176, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1def554032d40cb086d01cb1b9cc9f496729d20005c53fc5c4401aa9dddcaa6.jpg", + "image_caption": [ + "ECON" + ], + "image_footnote": [], + "bbox": [ + 176, + 247, + 367, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b02915f2e73fcd1206dafbf49afcc646577574ab699f7e4b6b2b9f81c281d78f.jpg", + "image_caption": [ + "PIFuHD" + ], + "image_footnote": [], + "bbox": [ + 367, + 247, + 545, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/281ae0046b6f7412000e4eadbe5585bbd456a1f54bba6fccfa3d63159ed6cdeb.jpg", + "image_caption": [ + "SiTH (Ours)", + "Figure 5. Qualitative comparison on CustomHumans. Top: Results of methods generating mesh and texture. Bottom: Results of methods generating mesh only. Note that single-view reconstruction is not possible to replicate exact back-view texture and geometry. Our method generates realistic texture and clothing wrinkles perceptually close to the real scans while other baselines only produce smooth colors and surfaces in the back regions. Best viewed in color and zoom in." + ], + "image_footnote": [], + "bbox": [ + 545, + 247, + 720, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/464fc5cbf953c8d367ba6074c1d569e7bdfc086a9653a29b800bdcda1a5a1f76.jpg", + "image_caption": [ + "GT References" + ], + "image_footnote": [], + "bbox": [ + 715, + 247, + 890, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local feature querying. To learn a generic mapping function that is robust to unseen images, it is important that the model is conditioned solely on local image information with respect to the position of $\\mathbf{x}$ . Therefore, we employ the idea of pixel-aligned feature querying [60, 61] and separate our model into two branches, i.e., color and geometry. Our model contains a normal predictor that converts the RGB image pair $(I^{F}, I^{B})$ into normal maps $(N^{F}, N^{B})$ . Two image feature encoders $G_{d}, G_{r}$ then extract color and geometry feature maps $(\\mathbf{f}_{d}, \\mathbf{f}_{r}) \\in \\mathbb{R}^{H' \\times W' \\times D}$ from the images and normal maps respectively (for simplicity we describe the process for a single image and leave out the superscripts, but both front and back images are treated the same). Finally, we project the query point $\\mathbf{x}$ onto the image coordinate (Fig. 4 red points) to retrieve the local features $(\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x}) \\in \\mathbb{R}^{D}$ :", + "bbox": [ + 75, + 446, + 472, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} _ {d, x} = \\mathcal {B} (\\mathbf {f} _ {d}, \\pi (\\mathbf {x})) = \\mathcal {B} (G _ {d} (N), \\pi (\\mathbf {x})), \\\\ \\mathbf {f} _ {r, x} = \\mathcal {B} (\\mathbf {f} _ {r}, \\pi (\\mathbf {x})) = \\mathcal {B} (G _ {r} (I), \\pi (\\mathbf {x})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 699, + 403, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ is a local feature querying operation using bilinear interpolation and $\\pi (\\cdot)$ denotes orthographic projection.", + "bbox": [ + 75, + 743, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local positional embedding with skinned body prior.", + "text_level": 1, + "bbox": [ + 75, + 780, + 459, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1, a major difficulty in mesh reconstruction is 3D ambiguity where a model has to infer unknown depth information between the front and back images. To address this issue, we follow prior work [22, 73, 79] leveraging a skinned body mesh [51] for guiding the reconstruction task. This body mesh is regarded as an anchor that provides an approximate 3D shape of the human body.", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To exploit this body prior, we devise a local positional embedding function that transforms the query point $\\mathbf{x}$ into the local body mesh coordinate system. We look for the closest point $\\mathbf{x}_c^*$ on the body mesh (Fig. 4 blue point), i.e.,", + "bbox": [ + 496, + 446, + 892, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {c} ^ {*} = \\arg \\min _ {\\mathbf {x} _ {c}} \\| \\mathbf {x} - \\mathbf {x} _ {c} \\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 521, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_c$ are points on the skinned body mesh $\\mathcal{M}$ . Our positional embedding $\\mathbf{p}$ constitutes four elements: a signed distance value $d_c$ between $\\mathbf{x}_c^*$ and $\\mathbf{x}$ , a vector $\\mathbf{n}_c = (\\mathbf{x} - \\mathbf{x}_c^*)$ , the UV coordinates $\\mathbf{u}_c \\in [0,1]^2$ of the point $\\mathbf{x}_c^*$ , and a visibility label $v_c \\in \\{1, -1, 0\\}$ that indicates whether $\\mathbf{x}_c^*$ is visible in the front/back image or neither. Finally, two separate MLPs $H_d, H_r$ take the positional embedding $\\mathbf{p} = [d_c, \\mathbf{n}_c, \\mathbf{u}_c, v_c]$ and the local texture/geometry features $(\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x})$ as inputs to predict the final SDF and RGB values at point $\\mathbf{x}$ :", + "bbox": [ + 496, + 555, + 892, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d _ {x} = H _ {d} \\left(\\mathbf {f} _ {d, x} ^ {F}, \\mathbf {f} _ {d, x} ^ {B}, \\mathbf {p}\\right), \\\\ \\mathbf {r} _ {x} = H _ {r} \\left(\\mathbf {f} _ {r, x} ^ {F}, \\mathbf {f} _ {r, x} ^ {B}, \\mathbf {p}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 718, + 772, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and inference. We used the same 3D dataset described in Sec. 3.1 to render training image pairs $(I^{F}, I^{B})$ from the 3D textured scans. For each training scan, query points $\\mathbf{x}$ are sampled within a 3-dimensional cube $[-1, 1]^{3}$ . For each point, we compute the ground-truth signed distance values $d$ to the scan surface, closest texture RGB values $r$ , and surface normal $\\mathbf{n}$ . Finally, we jointly optimized the normal predictors, the image encoders, and the MLPs in", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "542", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d9048fb13619c9b6317dbce4311b17106ab9e02e94f15d318739b4868935d226.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CAPE [45]CustomHuman [22]
MethodCD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F (×10-2) ↓CD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F / B (×10-2) ↓
PIFu [60]2.368 / 3.7630.77833.8422.7202.209 / 2.5820.80534.8816.073 / 8.496
PIFuHD [61]2.401 / 3.5220.77235.706-2.107 / 2.2280.80439.076-
PaMIR [79]2.190 / 2.8060.80436.7252.0852.181 / 2.5070.81335.8474.646 / 7.152
2K2K [18]2.478 / 3.6830.78228.700-2.488 / 3.2920.79630.186-
FOF [15]2.196 / 4.0400.77734.227-2.079 / 2.6440.80836.013-
ICON [73]2.516 / 3.0790.78629.630-2.256 / 2.7950.79130.437-
ECON [74]2.475 / 2.9700.78830.488-2.483 / 2.6800.79730.894-
SiTH (Ours)1.899 / 2.2610.81637.7631.9771.871 / 2.0450.82637.0293.929 / 6.803
", + "bbox": [ + 119, + 88, + 851, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Single-view human reconstruction benchmarks. We report Chamfer distance (CD), normal consistency (NC), and f-score between ground truth and predicted meshes. To evaluate texture reconstruction quality, we compute LPIPS between the image rendering of GT and generated textures. The best and the second best methods are highlighted in bold and underlined respectively. Note that gray color denotes models trained on more commercial 3D human scans while the others are trained on the public THuman2.0 dataset.", + "bbox": [ + 75, + 280, + 890, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "both branches with the following reconstruction losses:", + "bbox": [ + 76, + 349, + 442, + 364 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\left\\| d - d _ {x} \\right\\| _ {1} + \\lambda_ {n} \\left(1 - \\mathbf {n} \\cdot \\nabla_ {\\mathbf {x}} d _ {x}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 372, + 468, + 388 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\left\\| \\mathbf {r} - \\mathbf {r} _ {x} \\right\\| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 397, + 468, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that $\\nabla_{\\mathbf{x}}$ indicates numerical finite differences for computing local normals at point $\\mathbf{x}$ and $\\lambda_{n}$ is a hyperparameter.", + "bbox": [ + 75, + 417, + 468, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, we use the input image $I^{F}$ and the back-view image $\\tilde{I}^{B}$ obtained from Sec. 3.1 to reconstruct 3D mesh and textures. First, we align both images with the estimated body mesh $\\mathcal{M}$ to ensure that image features can be properly queried around the 3D anchor. We adopt a similar strategy of SMPLify [7] to optimize the scale and the offset of the body mesh with silhouette and 2D joint errors. Finally, we perform the marching cube algorithm [43] by querying SDF and RGB values within a dense voxel grid via Eq. (3) (see Fig. 2 right).", + "bbox": [ + 75, + 448, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 604, + 209, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 627, + 266, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Previous work relied on training data from commercial datasets such as RenderPeople [1]. While these datasets offer high-quality textured meshes, they also limit reproducibility due to limited accessibility. For fair comparisons, we follow ICON [73] by training our method on the public 3D dataset THuman2.0 [75] and using the CAPE [45] dataset for evaluation. However, we observed potential biases in the evaluation due to the low-res groundtruth meshes and image rendering defects in the CAPE dataset (for a detailed discussion, please refer to SuppSec. 6). Consequently, we further create a new benchmark that evaluates the baselines on a higher-quality 3D human dataset CustomHumans [22]. In the following, we provide a summary of the datasets used in our experiments:", + "bbox": [ + 75, + 643, + 468, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- THuman2.0 [75] contains approximately 500 scans of humans wearing 150 different garments in various poses. We use these 3D scans as the training data.", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CAPE [45] contains 15 subjects in 8 types of tight outfits. The test set, provided by ICON, consists of 100 meshes. We use CAPE for the quantitative evaluation (Sec. 4.2).", + "- CustomHumans [22] contains 600 higher-quality scans of 80 subjects in 120 different garments and varied poses. We selected 60 subjects for all quantitative experiments, user studies, and ablation studies. (Sec. 4.2 - Sec. 4.4)" + ], + "bbox": [ + 500, + 349, + 890, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. We follow the evaluation protocol in OccNet [46] and ICON [73] to compute 3D metrics Chamfer distance (CD), normal consistency (NC), and f-Score [65] on the generated meshes. To evaluate reconstructed mesh texture, we report LPIPS [78] of front and back texture rendering. In user studies, 30 participants rank the meshes obtained by four different methods. We report the average ranking ranging from 1 (best) to 4 (worst).", + "bbox": [ + 496, + 454, + 890, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Single-view Human Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 580, + 807, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark evaluation. We compared SiTH with state-of-the-art single-view human reconstruction methods, including PIFu [60], PIFuHD [61], PaMIR [79], FOF [15], ICON [73], PHORHUM [4], 2K2K [18], and ECON [74] on CAPE and CustomHumans. Note that PHORHUM is only used for qualitative comparison since a different camera system is used, leading to the misalignment with ground-truth meshes. We visualize the generated mesh texture and normals in Fig. 5. Existing methods produce oversmoothed texture and normals, particularly in the back. Our method not only generates photorealistic and perceptually consistent appearances in unobserved regions but also recovers underlying geometric details like clothing wrinkles.", + "bbox": [ + 496, + 598, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative results are summarized in Tab. 1. It's worth noting that most methods are trained with commercial datasets (gray color in Tab. 1), while the others are trained on the public THuman2.0 dataset. To evaluate the methods leveraging a skinned body prior (i.e., PaMIR, ICON, ECON, FOF, and SiTH), we use the same pose alignment procedure in their original implementations for a fair", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "543", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e4f28041b59e1a8da72898ae9c70f8ee695b29116f6c8dadd633686dea730304.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 166, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee2961317ec3147f70fb3982c5f1270386614287fd779d481d5eeb72b4073aca.jpg", + "image_caption": [ + "Pix2PixHD" + ], + "image_footnote": [], + "bbox": [ + 171, + 88, + 259, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cb24378c11b722c53a0558be43d6a0283e39592288fb718f98c06d959e0e775.jpg", + "image_caption": [ + "DreamPose" + ], + "image_footnote": [], + "bbox": [ + 264, + 88, + 367, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54f3dbe7a06ed222c79335b776006682ad7a9bf66a1802f85b3bcf7964e9c566.jpg", + "image_caption": [ + "ControlNet + Interrogate" + ], + "image_footnote": [], + "bbox": [ + 369, + 90, + 392, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06031aa23c7972a25329344d79444a8a64042082486556994610c74404c97bcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 90, + 478, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5d0c263d25c30f6af5b9c636f40a4a0b19cb63a5289e17fef21a1603c211985.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 90, + 524, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c9572de8d4770389f1a6d658f3cf55826d8b1c88c5f780058571c6382233a7b.jpg", + "image_caption": [ + "Zero-1-to-3" + ], + "image_footnote": [], + "bbox": [ + 529, + 90, + 553, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20124d72a1f43ff6b18ee3bab5c9ca1e2f3aa22a4709edc97c23f49d6c1c9282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 90, + 660, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05d6c49f43214c6ad6b794f9972be85e9e779abe61042e9acc7787f233b6d86b.jpg", + "image_caption": [ + "SiTH (Ours)" + ], + "image_footnote": [], + "bbox": [ + 663, + 90, + 746, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3afae08893a53dc535c0287a13358930cb20b76981048e8d882565cd52ff36e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 90, + 794, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dc0dcbdddfadc7f7af2c21016afc2ff1dede7f97a91c1c404b092e89b117b5b.jpg", + "image_caption": [ + "GT References" + ], + "image_footnote": [], + "bbox": [ + 797, + 90, + 883, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84327b293fce65fc3f6028d4454a36766aed3ccd017ee8b542144837ac58bd3f.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of back-view hallucination. We visualize back-view images generated by the baseline methods. Note that the three different images are sampled from different random seeds. Our results are perceptually close to the ground-truth image in terms of appearances and poses. Moreover, our method also preserves generative stochasticity for handling tiny wrinkle changes.", + "Figure 7. Comparison with optimization-based methods. Compared to methods that utilize diffusion models for optimization, our result is closer to the input image and contains local geometric details. Note that Chupa is not conditioned on the image, and fails to generate correct clothing from the text prompts." + ], + "image_footnote": [], + "bbox": [ + 94, + 277, + 455, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "comparison. Results in Tab. 1 show that the method using a body prior (PaMIR) outperformed the end-to-end method (PIFuHD) on tight clothing and challenging poses in CAPE. However, it falls short in handling diverse outfits in CustomHumans. Moreover, the methods trained on commercial datasets achieve better performance than those trained with public data (ICON, ECON). Notably, our method is robust across both benchmarks, achieving performance comparable to the methods trained on high-quality commercial data.", + "bbox": [ + 75, + 559, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared with optimization-based methods. We compared SiTH with methods that use pretrained diffusion models and a score distillation sampling loss [53] to optimize 3D meshes. In the case of Zero-1-to-3 [41], we used the input image to optimize an instant NGP [49] radiance field, and for Magic-123 [54], we provided additional text prompts to optimize an SDF tetrahedron [63]. From Fig. 7, we see that while both methods can handle full-body textures, they struggle with reasoning the underlying geometry and clothing details. It is worth noting that Zero-1-to-3 and Magic-123 require 10 minutes and 6 hours in optimization, respectively, while our method takes under 2 minutes to generate a textured mesh with a marching cube of $512^{3}$ resolution.", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a7f501f2e6457aca7a2553824bca44924c84897ddeabba2cc518405cdd2b42e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ICONECONPIFuHDOurs
Front Geometry3.1272.7202.6071.547
Back Geometry3.1932.5133.0931.200
Similarity3.0932.6602.7801.456
PIFuPaMIRPHOHRUMOurs
Front Texture3.0672.1533.4501.327
Back Texture3.4502.3553.1401.054
Similarity3.3072.1923.4161.093
ChupaOurs
User Preference36.0%64.0%
", + "bbox": [ + 504, + 275, + 883, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. User study results. Top: 30 users are asked to rank the quality of surface normal images from best (1) to worst (4). We report the average ranking of each method. Middle: Similar to the first task, users are asked to rank the quality of RGB textures. Bottom: We ask users to choose the mesh with a better quality.", + "bbox": [ + 498, + 468, + 890, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, more similar to our method is Chupa [35], which generates front/back-view normals for mesh reconstruction. Note that Chupa is not conditioned on images and does not generate texture. Instead, we provided body poses and text prompts generated by an image-to-text interrogator [38] as their conditional inputs. From Fig. 7, it's clear that text-conditioning is less accurate than image-conditioning, and the method struggles to generate unseen clothing styles such as coats. By contrast, our method can reconstruct correct clothing geometry and texture from unseen images. We present more discussions and comparisons with optimization-based methods in Supp-Sec. 9.1.", + "bbox": [ + 498, + 545, + 890, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User study. The above metrics may not fully capture the quality of 3D meshes in terms of realism and local details. To address this, we conducted a user study to compare the texture and geometry quality among various baselines. We invited 30 users to rank the front/back-view texture and normal renderings of 3D meshes generated by four different methods. Additionally, we asked the users to assess the similarity between the input images and the generated meshes. The results (Tab. 2) support our claim that existing methods struggle to efficiently generate desirable back-view texture/geometry from single-view images. Our method,", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "544", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b3ca412680ab6266f9b762cc192bfb13f4b747b730eaeb799b5af966f987d1b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCD (cm)↓NC↑f-Score↑
W/o Body Mesh2.4710.80133.244
W/o Hallucination1.9600.84036.677
Full Pipeline1.9580.82637.029
W/ GT Body Mesh1.1720.89158.858
W/ GT Body and I^B1.0590.91463.356
", + "bbox": [ + 99, + 88, + 450, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on CustomHumans. We ablate the hallucination module and the skinned body mesh in our pipeline. Please refer to our discussion in Sec. 4.4.", + "bbox": [ + 75, + 205, + 468, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which leverages the generative capability of diffusion models, consistently outperforms each baseline. It also produces more preferred front-view textures and geometries, as evidenced by higher user rankings. We also conducted a user study with Chupa (in Tab. 2 bottom) which also indicates more users prefer the 3D meshes generated by our method.", + "bbox": [ + 75, + 251, + 470, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Generative Capability", + "text_level": 1, + "bbox": [ + 76, + 347, + 284, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image quality comparison. Our hallucination module is a unique and essential component that generates spatially aligned human images to guide 3D mesh reconstruction. Given that our focus is on back-view hallucination, we compare the quality of generated images with the relevant generative methods in Fig. 6. We trained a baseline Pix2PixHD [69] model, which produced smooth and blurry results on unseen images due to overfitting to 500 subjects. Another method closely related to ours is DreamPose [31], which conditions the model with DensePose images and finetunes the diffusion model with paired data. However, their model failed to handle unseen images, in contrast to our approach. While Zero-1-to-3 [41] can generalize to unseen images, their method faces challenges in generating consistent body poses given the same back-view camera. Moreover, we designed another baseline that provides ControlNet [77] for corresponding text prompts using an image-to-text interrogator [38]. However, without proper image conditioning and fine-tuning, such a method cannot generate images that faithfully match the input appearances. Our method not only addresses these issues but also handles stochastic appearances (e.g., tiny differences in wrinkles) from different random seeds. We report 2D generative evaluation metrics and more results in Supp-Sec. 8.1.", + "bbox": [ + 75, + 364, + 472, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 732, + 230, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted controlled experiments to validate the effectiveness of our proposed modules. As shown in Fig. 8, the skinned body mesh is a crucial component for 3D human reconstruction. Without this body mesh as a 3D anchor, the output mesh contains an incorrect body shape due to the depth ambiguity issue. Conversely, removing the hallucination module has minimal impact on 3D reconstruction metrics, though it slightly degrades normal consistency. However, the overall quality in both texture and geometry is incomparable with our full model (see Fig. 8 right). This is", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c192e9142d0984e52a76ce26c54d1a8619edce1a63c6b0c8574f8f72e1274b69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 87, + 893, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/181415364ee9a723664cc01615b1cc12cd1c0cb6851e80eea47f528b4ed42532.jpg", + "image_caption": [ + "Figure 8. Ablation study. We visualize back and side-view rendering of the reconstructed meshes. Our full model produced a correct body shape and more realistic clothing geometry.", + "Figure 9. Applications. Our pipeline can be incorporated with text-to-image diffusion models for 3D human creation." + ], + "image_footnote": [], + "bbox": [ + 504, + 239, + 890, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consistent with our findings in user studies, indicating that 3D metrics may not accurately reflect the perceptual quality of 3D meshes. Finally, we tested two additional variants, leveraging ground-truth body meshes and real back-view images in our full pipeline, representing the upper bound of our method. As shown in Tab. 3 bottom, this additional information notably improves the 3D metrics. These results highlight the persistent challenges in the single-view reconstruction problem, including pose ambiguity and the stochastic nature of clothing geometry. For more experiments on our design choices, please refer to Supp-Sec. 8.5.", + "bbox": [ + 496, + 383, + 892, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Applications", + "text_level": 1, + "bbox": [ + 500, + 556, + 633, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inheriting the generative capability of LDM, SiTH is robust to diverse inputs, such as out-of-distribution or AI-generated images. We demonstrate a unique solution to link photo-realistic AI photos and high-fidelity 3D humans. In Fig. 9, we introduce a 3D creation workflow integrating powerful text-to-image generative models. Given a body pose, we generate a front-view image using Stable Diffusion and ControlNet using text prompts. SiTH then creates a full-body textured human from the AI-generated image.", + "bbox": [ + 496, + 574, + 890, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 714, + 619, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose an innovative pipeline designed to create fully textured 3D humans from single-view images. Our approach seamlessly integrates an image-conditioned diffusion model into the existing data-driven 3D reconstruction workflow. Leveraging the generative capabilities of the diffusion model, our method efficiently produces lifelike 3D humans from a diverse range of unseen images in under two minutes. We expect our work will advance the application of generative AI in 3D human creation.", + "bbox": [ + 496, + 733, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was partially supported by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\".", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "545", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Renderpeople, https://renderpeople.com/.6", + "[2] Stable diffusion image variations. huggingface . co / lambdaabs / stable - diffusion - image -conditioned.1", + "[3] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia, 2023. 2, 7", + "[4] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 4", + "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4", + "[6] Mikolaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD gans. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2", + "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Proceedings of the European Conference on Computer Vision (ECCV). Springer International Publishing, 2016. 6", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 3", + "[9] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4", + "[10] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 2", + "[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2", + "[12] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing relightable and animatable avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 4" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6, 5", + "[16] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human pose estimation in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[17] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[18] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4", + "[19] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2", + "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4", + "[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2", + "[22] Hsuan-I Ho, Lixin Xue, Jie Song, and Otmar Hilliges. Learning locally editable virtual humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 1", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 3", + "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2, 7", + "[25] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[26] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Selfrecon: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "546", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2", + "[28] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[30] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 2", + "[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 8, 2", + "[32] Manuel Kaufmann, Velko Vechev, and Dario Mylonopoulos. aitviewer, 2022. 4", + "[33] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (TOG), 32(3):1-13, 2013. 2", + "[34] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2", + "[35] Byungjun Kim, Patrick Kwon, Kwangho Lee, Myunggi Lee, Sookwan Han, Daesik Kim, and Hanbyul Joo. Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 7", + "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4", + "[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4", + "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 7, 8", + "[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 7, 8, 1", + "[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 2", + "[43] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm, 1998. 6", + "[44] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2", + "[45] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 1, 5", + "[46] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[48] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 7", + "[50] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2", + "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "547", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3", + "[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 7", + "[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2, 7", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3", + "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3", + "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4", + "[58] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35, 2022. 2, 3", + "[60] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2, 4, 5, 6", + "[61] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 4, 5, 6", + "[62] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[63] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid represent" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 7", + "[64] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2", + "[65] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[66] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 703-735. Wiley Online Library, 2022. 2", + "[67] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2", + "[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[69] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8, 2", + "[70] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. 2", + "[71] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 2", + "[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 641-676. Wiley Online Library, 2022. 2", + "[73] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6, 4", + "[74] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "548", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4, 5", + "[75] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 6, 1", + "[76] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 2", + "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 8, 1", + "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 2", + "[79] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 5, 6" + ], + "bbox": [ + 78, + 92, + 468, + 455 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "549", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_model.json b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56744f7f4e9d01932f98f5bb7c382007c44d04ba --- /dev/null +++ b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_model.json @@ -0,0 +1,2875 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.748, + 0.177 + ], + "angle": 0, + "content": "SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.204, + 0.351, + 0.221 + ], + "angle": 0, + "content": "Hsuan-I Ho" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.204, + 0.5, + 0.222 + ], + "angle": 0, + "content": "Jie Song" + }, + { + "type": "text", + "bbox": [ + 0.576, + 0.204, + 0.697, + 0.222 + ], + "angle": 0, + "content": "Otmar Hilliges" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.231, + 0.672, + 0.249 + ], + "angle": 0, + "content": "Department of Computer Science, ETH Zürich" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.252, + 0.591, + 0.265 + ], + "angle": 0, + "content": "https://ait.ethz.ch/sith" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.303, + 0.199, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.472, + 0.18, + 0.483 + ], + "angle": 0, + "content": "Real Photo" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.308, + 0.353, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.471, + 0.347, + 0.483 + ], + "angle": 0, + "content": "Back-view Hallucination" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.291, + 0.468, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.381, + 0.468, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.471, + 0.474, + 0.483 + ], + "angle": 0, + "content": "Mesh Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.285, + 0.604, + 0.303 + ], + "angle": 0, + "content": "\"A woman in a white T-shirt and a blue tennis skirt\"" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.304, + 0.74, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.471, + 0.594, + 0.483 + ], + "angle": 0, + "content": "AI-generated Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.471, + 0.737, + 0.483 + ], + "angle": 0, + "content": "e Back-view Hallucination" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.291, + 0.853, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.379, + 0.852, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.471, + 0.86, + 0.483 + ], + "angle": 0, + "content": "Mesh Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.489, + 0.894, + 0.545 + ], + "angle": 0, + "content": "Figure 1. Single-view textured human reconstruction. SiTH is a novel pipeline for creating high-quality and fully textured 3D human meshes from single images. We first hallucinate back-view appearances through an image-conditioned diffusion model, followed by the reconstruction of full-body textured meshes using both the front and back-view images. Our pipeline enables the creation of lifelike and diverse 3D humans from unseen photos (left) and AI-generated images (right)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.558, + 0.314, + 0.573 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.594, + 0.473, + 0.897 + ], + "angle": 0, + "content": "A long-standing goal of 3D human reconstruction is to create lifelike and fully detailed 3D humans from single-view images. The main challenge lies in inferring unknown body shapes, appearances, and clothing details in areas not visible in the images. To address this, we propose SiTH, a novel pipeline that uniquely integrates an image-conditioned diffusion model into a 3D mesh reconstruction workflow. At the core of our method lies the decomposition of the challenging single-view reconstruction problem into generative hallucination and reconstruction subproblems. For the former, we employ a powerful generative diffusion model to hallucinate unseen back-view appearance based on the input images. For the latter, we leverage skinned body meshes as guidance to recover full-body texture meshes from the input and back-view images. SiTH requires as few as 500 3D human scans for training while maintaining its generality and robustness to diverse images. Extensive evaluations on two 3D human benchmarks, including our newly created one, highlighted our method's superior accuracy and perceptual quality in 3D textured human reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.558, + 0.631, + 0.573 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.704 + ], + "angle": 0, + "content": "With the growing popularity of 3D and virtual reality applications, there has been increasing interest in creating realistic 3D human models. In general, crafting 3D humans is labor-intensive, time-consuming, and requires collaboration from highly skilled professionals. To bring lifelike 3D humans to reality and to support both expert and amateur creators in this task, it is essential to enable users to create textured 3D humans from simple 2D images or photos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Reconstructing a fully textured human mesh from a single-view image presents an ill-posed problem with two major challenges. Firstly, the appearance information required for generating texture in unobserved regions is missing. Secondly, 3D information for mesh reconstruction, such as depth, surface, and body pose, becomes ambiguous in a 2D image. Previous efforts [4, 60, 79] attempted to tackle these challenges in a data-driven manner, focusing on training neural networks with image-mesh pairs. However, these approaches struggle with images featuring unseen appearances or poses, due to limited 3D human training data. More recent studies [61, 73, 79] introduced additional 3D reasoning modules to enhance robustness against unseen" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "538" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "poses. Yet, generating realistic and full-body textures from unseen appearances still remains an unsolved problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.125, + 0.47, + 0.337 + ], + "angle": 0, + "content": "To address the above challenges, we propose SiTH, a novel pipeline that integrates an image-conditioned diffusion model to reconstruct lifelike 3D textured humans from monocular images. At the core of our approach is the decomposition of the challenging single-view problem into two subproblems: generative back-view hallucination and mesh reconstruction. This decomposition enables us to exploit the generative capability of pretrained diffusion models to guide full-body mesh and texture reconstruction. The workflow is depicted in Fig. 1. Given a front-view image, the first stage involves hallucinating a perceptually consistent back-view image using image-conditioned diffusion. The second stage reconstructs full-body mesh and texture, utilizing both the front and back-view images as guidance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.471, + 0.701 + ], + "angle": 0, + "content": "More specifically, we employ the generative capabilities of pretrained diffusion models (e.g. Stable Diffusion [57]) to infer unobserved back-view appearances for full-body 3D reconstruction. The primary challenge in ensuring the realism of 3D meshes lies in generating images that depict spatially aligned body shapes and perceptually consistent appearances with the input images. While diffusion models demonstrate impressive generative abilities with text conditioning, they are limited in producing desired back-view images using the frontal images as image conditions. To overcome this, we adapt the network architecture to enable conditioning on frontal images and introduce additional trainable components following ControlNet [77] to provide pose and mask control. To fully tailor this model to our task while retaining its original generative power, we carefully fine-tune the diffusion model using multi-view images rendered from 3D human scans. Complementing this generative model, we develop a mesh reconstruction module to recover full-body textured mesh from front and back-view images. We follow prior work in handling 3D ambiguity through normal [61] and skinned body [73, 79] guidance. It is worth noting that the models for both subproblems are trained using the same public THuman2.0 [75] dataset, which consists of as few as 500 scans." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To advance research in single-view human reconstruction, we created a new benchmark based on the high-quality CustomHumans [22] dataset and conducted comprehensive evaluations against state-of-the-art methods. Compared to existing end-to-end methods [4, 60, 79], our two-stage pipeline can recover full-body textured meshes, including back-view details, and demonstrates robustness to unseen images. In contrast to time-intensive diffusion-based optimization methods [24, 35, 41], our pipeline efficiently produces high-quality textured meshes in under two minutes. Moreover, we explored applications combining text-guided diffusion models, showing SiTH's versatility in 3D human creation. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "- We introduce SiTH, a single-view human reconstruction pipeline capable of producing high-quality, fully textured 3D human meshes within two minutes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.892, + 0.182 + ], + "angle": 0, + "content": "- Through decomposing the single-view reconstruction task, SiTH can be efficiently trained with public 3D human scans and is more robust to unseen images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.183, + 0.892, + 0.213 + ], + "angle": 0, + "content": "- We establish a new benchmark featuring more diverse subjects for evaluating textured human reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.23, + 0.642, + 0.246 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Single-view human mesh reconstruction. Reconstructing 3D humans from monocular inputs [14, 17, 26, 28, 29, 62, 67, 71] has gained more popularity in research. In this context, we focus on methods that recover 3D human shapes, garments, and textures from a single image. As a seminal work, Saito et al. [60] first proposed a data-driven method with pixel-aligned features and neural fields [72]. Its follow-up work PIFuHD [61] further improved this framework with high-res normal guidance. Later approaches extended this framework with additional human body priors. For instance, PaMIR [79] and ICON [73] utilized skinned body models [42, 51] to guide 3D reconstruction. ARCH [25], \\(\\mathrm{ARCH} + +\\) [21], and CAR [39] transformed global coordinates into the canonical coordinates to allow for reposing. PHOHRUM [4] and S3F [12] further disentangled shading and albedo to enable relighting. Another line of work replaced the neural representations with conventional Poisson surface reconstruction [33, 34]. ECON [74] and 2K2K [18] trained normal and depth predictors to generate front and back 2.5D point clouds. The human mesh is obtained by fusing these point clouds with body priors and 3D heuristics. However, none of these methods produce realistic full-body texture and geometry in the unobserved regions. Our pipeline addresses this problem by incorporating a generative diffusion model into the 3D human reconstruction workflow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D generation with 2D diffusion models. Diffusion models [52, 56, 57, 59] trained with large collections of images have demonstrated unprecedented capability in creating 3D objects from text prompts. Most prior work [11, 19, 40, 47, 53, 68] followed an optimization workflow to update 3D representations (e.g. NeRF [48], SDF tetrahedron [63]) via neural rendering [66] and a score distillation sampling (SDS) [53] loss. While some methods [3, 10, 24, 27, 64, 76] applied this workflow to human bodies, they cannot produce accurate human bodies and appearances due to the ambiguity of text-conditioning. More recent work [41, 54] also tried to extend this workflow with more accurate image-conditioning. However, we show that they struggle to recover human clothing details and require a long optimization time. Most related to our work is Chupa [35], which also decomposes its pipeline into two stages. Note that Chupa is an optimization-based approach that relies on texts" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "539" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.885, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.893, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Method overview. SiTH is a two-stage pipeline composed of back-view hallucination and mesh reconstruction. The back-view hallucination module samples perceptually consistent back-view images through an iterative denoising process conditioned on the input image, UV map, and silhouette mask (Sec. 3.1). Based on the input and generated back-view images, the mesh reconstruction module recovers a full-body mesh and textures leveraging a skinned body prior as guidance (Sec. 3.2). Note that both modules in the pipeline can be trained with the same public 3D human dataset and generalize unseen images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.451 + ], + "angle": 0, + "content": "and cannot model colors. We address these issues by introducing an image-conditioning strategy and model. Most importantly, our method swiftly reconstructs full-texture human meshes without any optimization process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.737 + ], + "angle": 0, + "content": "Diffusion models adaptation. Foundation models [8, 13, 20, 36] trained on large-scale datasets have been shown to be adaptable to various downstream tasks. Following this trend, pretrained diffusion models [52, 56, 57, 59] have become common backbones for generative modeling. For instance, they can be customized by finetuning with a small collection of images [23, 37, 58]. ControlNet [77] introduced additional trainable plugins to enable image conditioning such as body skeletons. While these strategies have been widely adopted, none of them directly fit our objective. More relevant to our task is DreamPose [31], which utilizes DensePose [16] images as conditions to repose input images. However, it cannot handle out-of-distribution images due to overfitting. Similarly, Zero-1-to-3 [41] finetunes a diffusion model with multi-view images to allow for viewpoint control. However, we show that viewpoint conditioning is not sufficient for generating consistent human bodies. Our model addresses this issue by providing accurate body pose and mask conditions for back-view hallucination." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.746, + 0.212, + 0.763 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Method overview. Given an input image of a human body and estimated SMPL-X [51] parameters, SiTH produces a full-body textured mesh. This mesh not only captures the observed appearances but also recovers geometric and textural details in unseen regions, such as clothing wrinkles on the back. The pipeline is composed of two modules and is summarized in Fig. 2. In the first stage, we hallucinate unobserved appearances leveraging the generative power of an image-conditioned diffusion model (Sec. 3.1). In the second" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.893, + 0.452 + ], + "angle": 0, + "content": "stage, we reconstruct a full-body textured mesh given the input front-view image and the generated back-view image as guidance (Sec. 3.2). Notably, both modules are efficiently trained with 500 textured human scans in THuman2.0 [75]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.456, + 0.726, + 0.47 + ], + "angle": 0, + "content": "3.1. Back-view Hallucination" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.716 + ], + "angle": 0, + "content": "Preliminaries. Given an input front-view image \\( I^{F} \\in \\mathbb{R}^{H \\times W \\times 3} \\), our goal is to infer a back-view image \\( I^{B} \\in \\mathbb{R}^{H \\times W \\times 3} \\) which depicts unobserved body appearances. This task is under-constrained since there are multiple possible solutions to the same input images. Taking this perspective into account, we leverage a latent diffusion model (LDM) [57] to learn a conditional distribution of back-view images given a front-view image. First, a VAE autoencoder, consisting of an encoder \\( \\mathcal{E} \\) and a decoder \\( \\mathcal{D} \\), is pretrained on a corpus of 2D natural images through image reconstruction, i.e., \\( \\tilde{I} = \\mathcal{D}(\\mathcal{E}(I)) \\). Afterwards, an LDM learns to produce a latent code \\( z \\) within the VAE latent distribution \\( z = \\mathcal{E}(I) \\) from randomly sampled noise. To sample an image, a latent code \\( \\tilde{z} \\) is obtained by iteratively denoising Gaussian noise. The final image is reconstructed through the decoder, i.e., \\( \\tilde{I} = \\mathcal{D}(\\tilde{z}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image-conditioned diffusion model. Simply applying the LDM architecture to our task is not sufficient since our goal is to learn a conditional distribution of back-view images given an input conditional image. To this end, we make several adaptations to allow for image-conditioning as shown in Fig. 3. First, we utilize the pretrained CLIP [55] image encoder and VAE encoder \\(\\mathcal{E}\\) to extract image features from the front-view image (i.e., \\(I^{F}\\)). These image features are used for conditioning the LDM, ensuring the output image shares a consistent appearance with the input image. Second, we follow the idea of ControlNet [77] and propose to use a UV map \\((I_{UV}^{B} \\in \\mathbb{R}^{H \\times W \\times 3})\\) and a silhou" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "540" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.468, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Figure 3. Training of back-view hallucination module. We employ a pretrained LDM and ControlNet architecture to enable image conditioning. To train our model, we render training pairs of conditional images \\( I^{F} \\) and ground-truth images \\( I^{B} \\) from 3D human scans. Given a noisy image latent \\( z_{t} \\), the model predicts added noise \\( \\epsilon \\) given the conditional image \\( I^{F} \\), UV map \\( I_{UV}^{B} \\), and mask \\( I_{M}^{B} \\) as conditions. We train the ControlNet model and cross-attention layers while keeping other parameters frozen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.478 + ], + "angle": 0, + "content": "ette mask \\((I_M^B\\in \\mathbb{R}^{H\\times W})\\) from the back view as additional conditions. These conditional signals provide additional information that ensures the output image has a similar body shape and pose to the conditional input image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.473, + 0.68 + ], + "angle": 0, + "content": "Learning hallucination from pretraining. Another challenge in training an image-conditioned LDM is data. Training the model from scratch is infeasible due to the requirement of a large number of paired images rendered from 3D textured human scans. Inspired by the concept of learning from large-scale pretraining [13, 20], we build our image-conditioned LDM on top of a pretrained diffusion U-Net [57]. We utilize the finetuning strategy [37, 77] to optimize cross-attention layers and ControlNet parameters while keeping most of the other parameters frozen (see Fig. 3). The design and training strategy of our image-conditioned diffusion model enables hallucinating plausible back-view images that are cosistent with the frontal inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Training and inference. To generate pairwise training images from 3D human scans, we sample camera view angles and use orthographic projection to render RGBA images from 3D scans and UV maps from their SMPL-X fits. Given a pair of images rendered by a frontal and its corresponding back camera, the first image serves as the conditional input \\( I^{F} \\) while the other one is the ground-truth image \\( I^{B} \\). During training, the ground-truth latent code \\( z_{0} = \\mathcal{E}(I^{B}) \\) is perturbed by the diffusion process in t time steps, resulting in a noisy latent \\( z_{t} \\). The image-conditioned LDM model \\( \\epsilon_{\\theta} \\) aims to predict the added noise \\( \\epsilon \\) given the noisy latent \\( z_{t} \\), the time step \\( t \\sim [0,1000] \\), the conditional image \\( I^{F} \\), the silhouette mask \\( I_{M}^{B} \\), and the UV map \\( I_{UV}^{B} \\) (See Fig. 3). The objective function for fine-tuning can be" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.094, + 0.848, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Figure 4. Mesh reconstruction module. Given front and back-view images \\((I^{F}, I^{B})\\) we predict their normal images \\((N^{F}, N^{B})\\) through a learned normal predictor. A 3D point \\(\\mathbf{x}\\) is projected onto these images for querying pixel-aligned features \\((\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x})\\). To leverage human body mesh as guidance, we embed the point \\(\\mathbf{x}\\) into the local UV coordinates \\(\\mathbf{u}_c\\), vector \\(\\mathbf{n_c}\\), distance \\(d_c\\), and visibility \\(v_c\\). Finally, two decoders \\((H_d, H_r)\\) predict SDF and RGB values at \\(\\mathbf{x}\\) given the positional embedding and pixel-aligned features." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.428, + 0.604, + 0.442 + ], + "angle": 0, + "content": "represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.455, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {z \\sim \\mathcal {E} (I), t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\\right) \\| _ {2} ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.895, + 0.571 + ], + "angle": 0, + "content": "At test time, we obtain \\( I_{UV}^{B}, I_{M}^{B} \\) from an off-the-shelf pose predictor [9] and segmentation model [36]. To infer a back-view image, we sample a latent \\( \\tilde{z}_0 \\) by performing the iterative denoising process starting from a Gaussian noise \\( z_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\). The back-view image can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.581, + 0.892, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\tilde {I} ^ {B} = \\mathcal {D} (\\tilde {z} _ {0}) = \\mathcal {D} \\left(f _ {\\theta} \\left(z _ {T}, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.644 + ], + "angle": 0, + "content": "where \\( f_{\\theta} \\) is a function representing the iterative denoising process of our image-conditioned LDM (See Fig. 2 left)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.764, + 0.669 + ], + "angle": 0, + "content": "3.2. Human Mesh Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.794 + ], + "angle": 0, + "content": "After obtaining the back-view image, our goal is to construct a full-body human mesh and its textures using the input and back-view image as guidance. We follow the literature [60, 61] to model this task with a data-driven method. Given pairwise training data (i.e., front/back-view images and 3D scans), we learn a data-driven model that maps these images to a 3D representation (e.g., a signed distance field (SDF)). We define this mapping as below:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.805, + 0.892, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Phi : \\mathbb {R} ^ {H \\times W \\times 3} \\times \\mathbb {R} ^ {H \\times W \\times 3} \\times \\mathbb {R} ^ {3} \\rightarrow \\mathbb {R} \\times \\mathbb {R} ^ {3} \\tag {3} \\\\ (I ^ {F}, I ^ {B}, \\mathbf {x}) \\mapsto d _ {x}, \\mathbf {r} _ {x}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.857, + 0.895, + 0.918 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is the 3D coordinate of a query point, and \\(d_x, \\mathbf{r}_x\\) denote the signed distance and RGB color value at point \\(\\mathbf{x}\\). The network components we used for learning the mapping function are depicted in Fig. 4." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.101, + 0.165, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.228, + 0.165, + 0.243 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.103, + 0.357, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.228, + 0.296, + 0.242 + ], + "angle": 0, + "content": "PaMIR" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.103, + 0.536, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.228, + 0.484, + 0.242 + ], + "angle": 0, + "content": "PHORHUM" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.103, + 0.708, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.228, + 0.663, + 0.242 + ], + "angle": 0, + "content": "SiTH (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.103, + 0.885, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.228, + 0.843, + 0.242 + ], + "angle": 0, + "content": "GT References" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.248, + 0.177, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.363, + 0.165, + 0.377 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.248, + 0.368, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.363, + 0.296, + 0.375 + ], + "angle": 0, + "content": "ECON" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.248, + 0.546, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.363, + 0.472, + 0.375 + ], + "angle": 0, + "content": "PIFuHD" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.248, + 0.721, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.363, + 0.663, + 0.376 + ], + "angle": 0, + "content": "SiTH (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.248, + 0.892, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.363, + 0.843, + 0.375 + ], + "angle": 0, + "content": "GT References" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.894, + 0.442 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on CustomHumans. Top: Results of methods generating mesh and texture. Bottom: Results of methods generating mesh only. Note that single-view reconstruction is not possible to replicate exact back-view texture and geometry. Our method generates realistic texture and clothing wrinkles perceptually close to the real scans while other baselines only produce smooth colors and surfaces in the back regions. Best viewed in color and zoom in." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.692 + ], + "angle": 0, + "content": "Local feature querying. To learn a generic mapping function that is robust to unseen images, it is important that the model is conditioned solely on local image information with respect to the position of \\(\\mathbf{x}\\). Therefore, we employ the idea of pixel-aligned feature querying [60, 61] and separate our model into two branches, i.e., color and geometry. Our model contains a normal predictor that converts the RGB image pair \\((I^{F}, I^{B})\\) into normal maps \\((N^{F}, N^{B})\\). Two image feature encoders \\(G_{d}, G_{r}\\) then extract color and geometry feature maps \\((\\mathbf{f}_{d}, \\mathbf{f}_{r}) \\in \\mathbb{R}^{H' \\times W' \\times D}\\) from the images and normal maps respectively (for simplicity we describe the process for a single image and leave out the superscripts, but both front and back images are treated the same). Finally, we project the query point \\(\\mathbf{x}\\) onto the image coordinate (Fig. 4 red points) to retrieve the local features \\((\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x}) \\in \\mathbb{R}^{D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.7, + 0.405, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} _ {d, x} = \\mathcal {B} (\\mathbf {f} _ {d}, \\pi (\\mathbf {x})) = \\mathcal {B} (G _ {d} (N), \\pi (\\mathbf {x})), \\\\ \\mathbf {f} _ {r, x} = \\mathcal {B} (\\mathbf {f} _ {r}, \\pi (\\mathbf {x})) = \\mathcal {B} (G _ {r} (I), \\pi (\\mathbf {x})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.775 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) is a local feature querying operation using bilinear interpolation and \\(\\pi (\\cdot)\\) denotes orthographic projection." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.781, + 0.46, + 0.796 + ], + "angle": 0, + "content": "Local positional embedding with skinned body prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As mentioned in Sec. 1, a major difficulty in mesh reconstruction is 3D ambiguity where a model has to infer unknown depth information between the front and back images. To address this issue, we follow prior work [22, 73, 79] leveraging a skinned body mesh [51] for guiding the reconstruction task. This body mesh is regarded as an anchor that provides an approximate 3D shape of the human body." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.51 + ], + "angle": 0, + "content": "To exploit this body prior, we devise a local positional embedding function that transforms the query point \\(\\mathbf{x}\\) into the local body mesh coordinate system. We look for the closest point \\(\\mathbf{x}_c^*\\) on the body mesh (Fig. 4 blue point), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.522, + 0.892, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {c} ^ {*} = \\arg \\min _ {\\mathbf {x} _ {c}} \\| \\mathbf {x} - \\mathbf {x} _ {c} \\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.893, + 0.707 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_c\\) are points on the skinned body mesh \\(\\mathcal{M}\\). Our positional embedding \\(\\mathbf{p}\\) constitutes four elements: a signed distance value \\(d_c\\) between \\(\\mathbf{x}_c^*\\) and \\(\\mathbf{x}\\), a vector \\(\\mathbf{n}_c = (\\mathbf{x} - \\mathbf{x}_c^*)\\), the UV coordinates \\(\\mathbf{u}_c \\in [0,1]^2\\) of the point \\(\\mathbf{x}_c^*\\), and a visibility label \\(v_c \\in \\{1, -1, 0\\}\\) that indicates whether \\(\\mathbf{x}_c^*\\) is visible in the front/back image or neither. Finally, two separate MLPs \\(H_d, H_r\\) take the positional embedding \\(\\mathbf{p} = [d_c, \\mathbf{n}_c, \\mathbf{u}_c, v_c]\\) and the local texture/geometry features \\((\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x})\\) as inputs to predict the final SDF and RGB values at point \\(\\mathbf{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.719, + 0.773, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d _ {x} = H _ {d} \\left(\\mathbf {f} _ {d, x} ^ {F}, \\mathbf {f} _ {d, x} ^ {B}, \\mathbf {p}\\right), \\\\ \\mathbf {r} _ {x} = H _ {r} \\left(\\mathbf {f} _ {r, x} ^ {F}, \\mathbf {f} _ {r, x} ^ {B}, \\mathbf {p}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Training and inference. We used the same 3D dataset described in Sec. 3.1 to render training image pairs \\((I^{F}, I^{B})\\) from the 3D textured scans. For each training scan, query points \\(\\mathbf{x}\\) are sampled within a 3-dimensional cube \\([-1, 1]^{3}\\). For each point, we compute the ground-truth signed distance values \\(d\\) to the scan surface, closest texture RGB values \\(r\\), and surface normal \\(\\mathbf{n}\\). Finally, we jointly optimized the normal predictors, the image encoders, and the MLPs in" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "542" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.852, + 0.272 + ], + "angle": 0, + "content": "
CAPE [45]CustomHuman [22]
MethodCD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F (×10-2) ↓CD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F / B (×10-2) ↓
PIFu [60]2.368 / 3.7630.77833.8422.7202.209 / 2.5820.80534.8816.073 / 8.496
PIFuHD [61]2.401 / 3.5220.77235.706-2.107 / 2.2280.80439.076-
PaMIR [79]2.190 / 2.8060.80436.7252.0852.181 / 2.5070.81335.8474.646 / 7.152
2K2K [18]2.478 / 3.6830.78228.700-2.488 / 3.2920.79630.186-
FOF [15]2.196 / 4.0400.77734.227-2.079 / 2.6440.80836.013-
ICON [73]2.516 / 3.0790.78629.630-2.256 / 2.7950.79130.437-
ECON [74]2.475 / 2.9700.78830.488-2.483 / 2.6800.79730.894-
SiTH (Ours)1.899 / 2.2610.81637.7631.9771.871 / 2.0450.82637.0293.929 / 6.803
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.281, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Table 1. Single-view human reconstruction benchmarks. We report Chamfer distance (CD), normal consistency (NC), and f-score between ground truth and predicted meshes. To evaluate texture reconstruction quality, we compute LPIPS between the image rendering of GT and generated textures. The best and the second best methods are highlighted in bold and underlined respectively. Note that gray color denotes models trained on more commercial 3D human scans while the others are trained on the public THuman2.0 dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.35, + 0.443, + 0.365 + ], + "angle": 0, + "content": "both branches with the following reconstruction losses:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.373, + 0.469, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\left\\| d - d _ {x} \\right\\| _ {1} + \\lambda_ {n} \\left(1 - \\mathbf {n} \\cdot \\nabla_ {\\mathbf {x}} d _ {x}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.398, + 0.469, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\left\\| \\mathbf {r} - \\mathbf {r} _ {x} \\right\\| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.469, + 0.448 + ], + "angle": 0, + "content": "Note that \\(\\nabla_{\\mathbf{x}}\\) indicates numerical finite differences for computing local normals at point \\(\\mathbf{x}\\) and \\(\\lambda_{n}\\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.469, + 0.601 + ], + "angle": 0, + "content": "During inference, we use the input image \\( I^{F} \\) and the back-view image \\( \\tilde{I}^{B} \\) obtained from Sec. 3.1 to reconstruct 3D mesh and textures. First, we align both images with the estimated body mesh \\( \\mathcal{M} \\) to ensure that image features can be properly queried around the 3D anchor. We adopt a similar strategy of SMPLify [7] to optimize the scale and the offset of the body mesh with silhouette and 2D joint errors. Finally, we perform the marching cube algorithm [43] by querying SDF and RGB values within a dense voxel grid via Eq. (3) (see Fig. 2 right)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.21, + 0.621 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.628, + 0.267, + 0.643 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Dataset. Previous work relied on training data from commercial datasets such as RenderPeople [1]. While these datasets offer high-quality textured meshes, they also limit reproducibility due to limited accessibility. For fair comparisons, we follow ICON [73] by training our method on the public 3D dataset THuman2.0 [75] and using the CAPE [45] dataset for evaluation. However, we observed potential biases in the evaluation due to the low-res groundtruth meshes and image rendering defects in the CAPE dataset (for a detailed discussion, please refer to SuppSec. 6). Consequently, we further create a new benchmark that evaluates the baselines on a higher-quality 3D human dataset CustomHumans [22]. In the following, we provide a summary of the datasets used in our experiments:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- THuman2.0 [75] contains approximately 500 scans of humans wearing 150 different garments in various poses. We use these 3D scans as the training data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.35, + 0.89, + 0.395 + ], + "angle": 0, + "content": "- CAPE [45] contains 15 subjects in 8 types of tight outfits. The test set, provided by ICON, consists of 100 meshes. We use CAPE for the quantitative evaluation (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.396, + 0.891, + 0.455 + ], + "angle": 0, + "content": "- CustomHumans [22] contains 600 higher-quality scans of 80 subjects in 120 different garments and varied poses. We selected 60 subjects for all quantitative experiments, user studies, and ablation studies. (Sec. 4.2 - Sec. 4.4)" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.35, + 0.891, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Evaluation protocol. We follow the evaluation protocol in OccNet [46] and ICON [73] to compute 3D metrics Chamfer distance (CD), normal consistency (NC), and f-Score [65] on the generated meshes. To evaluate reconstructed mesh texture, we report LPIPS [78] of front and back texture rendering. In user studies, 30 participants rank the meshes obtained by four different methods. We report the average ranking ranging from 1 (best) to 4 (worst)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.582, + 0.808, + 0.597 + ], + "angle": 0, + "content": "4.2. Single-view Human Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Benchmark evaluation. We compared SiTH with state-of-the-art single-view human reconstruction methods, including PIFu [60], PIFuHD [61], PaMIR [79], FOF [15], ICON [73], PHORHUM [4], 2K2K [18], and ECON [74] on CAPE and CustomHumans. Note that PHORHUM is only used for qualitative comparison since a different camera system is used, leading to the misalignment with ground-truth meshes. We visualize the generated mesh texture and normals in Fig. 5. Existing methods produce oversmoothed texture and normals, particularly in the back. Our method not only generates photorealistic and perceptually consistent appearances in unobserved regions but also recovers underlying geometric details like clothing wrinkles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The quantitative results are summarized in Tab. 1. It's worth noting that most methods are trained with commercial datasets (gray color in Tab. 1), while the others are trained on the public THuman2.0 dataset. To evaluate the methods leveraging a skinned body prior (i.e., PaMIR, ICON, ECON, FOF, and SiTH), we use the same pose alignment procedure in their original implementations for a fair" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "543" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.168, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.208, + 0.162, + 0.221 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.089, + 0.261, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.208, + 0.248, + 0.22 + ], + "angle": 0, + "content": "Pix2PixHD" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.089, + 0.369, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.208, + 0.358, + 0.22 + ], + "angle": 0, + "content": "DreamPose" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.091, + 0.393, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.091, + 0.48, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.091, + 0.526, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.091, + 0.555, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.091, + 0.661, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.208, + 0.532, + 0.221 + ], + "angle": 0, + "content": "ControlNet + Interrogate" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.208, + 0.63, + 0.22 + ], + "angle": 0, + "content": "Zero-1-to-3" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.091, + 0.748, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.208, + 0.766, + 0.22 + ], + "angle": 0, + "content": "SiTH (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.091, + 0.795, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.091, + 0.885, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.208, + 0.882, + 0.22 + ], + "angle": 0, + "content": "GT References" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.229, + 0.894, + 0.272 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of back-view hallucination. We visualize back-view images generated by the baseline methods. Note that the three different images are sampled from different random seeds. Our results are perceptually close to the ground-truth image in terms of appearances and poses. Moreover, our method also preserves generative stochasticity for handling tiny wrinkle changes." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.279, + 0.457, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Figure 7. Comparison with optimization-based methods. Compared to methods that utilize diffusion models for optimization, our result is closer to the input image and contains local geometric details. Note that Chupa is not conditioned on the image, and fails to generate correct clothing from the text prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.697 + ], + "angle": 0, + "content": "comparison. Results in Tab. 1 show that the method using a body prior (PaMIR) outperformed the end-to-end method (PIFuHD) on tight clothing and challenging poses in CAPE. However, it falls short in handling diverse outfits in CustomHumans. Moreover, the methods trained on commercial datasets achieve better performance than those trained with public data (ICON, ECON). Notably, our method is robust across both benchmarks, achieving performance comparable to the methods trained on high-quality commercial data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Compared with optimization-based methods. We compared SiTH with methods that use pretrained diffusion models and a score distillation sampling loss [53] to optimize 3D meshes. In the case of Zero-1-to-3 [41], we used the input image to optimize an instant NGP [49] radiance field, and for Magic-123 [54], we provided additional text prompts to optimize an SDF tetrahedron [63]. From Fig. 7, we see that while both methods can handle full-body textures, they struggle with reasoning the underlying geometry and clothing details. It is worth noting that Zero-1-to-3 and Magic-123 require 10 minutes and 6 hours in optimization, respectively, while our method takes under 2 minutes to generate a textured mesh with a marching cube of \\(512^{3}\\) resolution." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.276, + 0.885, + 0.464 + ], + "angle": 0, + "content": "
ICONECONPIFuHDOurs
Front Geometry3.1272.7202.6071.547
Back Geometry3.1932.5133.0931.200
Similarity3.0932.6602.7801.456
PIFuPaMIRPHOHRUMOurs
Front Texture3.0672.1533.4501.327
Back Texture3.4502.3553.1401.054
Similarity3.3072.1923.4161.093
ChupaOurs
User Preference36.0%64.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Table 2. User study results. Top: 30 users are asked to rank the quality of surface normal images from best (1) to worst (4). We report the average ranking of each method. Middle: Similar to the first task, users are asked to rank the quality of RGB textures. Bottom: We ask users to choose the mesh with a better quality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.546, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Additionally, more similar to our method is Chupa [35], which generates front/back-view normals for mesh reconstruction. Note that Chupa is not conditioned on images and does not generate texture. Instead, we provided body poses and text prompts generated by an image-to-text interrogator [38] as their conditional inputs. From Fig. 7, it's clear that text-conditioning is less accurate than image-conditioning, and the method struggles to generate unseen clothing styles such as coats. By contrast, our method can reconstruct correct clothing geometry and texture from unseen images. We present more discussions and comparisons with optimization-based methods in Supp-Sec. 9.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "User study. The above metrics may not fully capture the quality of 3D meshes in terms of realism and local details. To address this, we conducted a user study to compare the texture and geometry quality among various baselines. We invited 30 users to rank the front/back-view texture and normal renderings of 3D meshes generated by four different methods. Additionally, we asked the users to assess the similarity between the input images and the generated meshes. The results (Tab. 2) support our claim that existing methods struggle to efficiently generate desirable back-view texture/geometry from single-view images. Our method," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "544" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.451, + 0.196 + ], + "angle": 0, + "content": "
MethodCD (cm)↓NC↑f-Score↑
W/o Body Mesh2.4710.80133.244
W/o Hallucination1.9600.84036.677
Full Pipeline1.9580.82637.029
W/ GT Body Mesh1.1720.89158.858
W/ GT Body and I^B1.0590.91463.356
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.206, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Table 3. Ablation study on CustomHumans. We ablate the hallucination module and the skinned body mesh in our pipeline. Please refer to our discussion in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.471, + 0.344 + ], + "angle": 0, + "content": "which leverages the generative capability of diffusion models, consistently outperforms each baseline. It also produces more preferred front-view textures and geometries, as evidenced by higher user rankings. We also conducted a user study with Chupa (in Tab. 2 bottom) which also indicates more users prefer the 3D meshes generated by our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.285, + 0.364 + ], + "angle": 0, + "content": "4.3. Generative Capability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.729 + ], + "angle": 0, + "content": "Image quality comparison. Our hallucination module is a unique and essential component that generates spatially aligned human images to guide 3D mesh reconstruction. Given that our focus is on back-view hallucination, we compare the quality of generated images with the relevant generative methods in Fig. 6. We trained a baseline Pix2PixHD [69] model, which produced smooth and blurry results on unseen images due to overfitting to 500 subjects. Another method closely related to ours is DreamPose [31], which conditions the model with DensePose images and finetunes the diffusion model with paired data. However, their model failed to handle unseen images, in contrast to our approach. While Zero-1-to-3 [41] can generalize to unseen images, their method faces challenges in generating consistent body poses given the same back-view camera. Moreover, we designed another baseline that provides ControlNet [77] for corresponding text prompts using an image-to-text interrogator [38]. However, without proper image conditioning and fine-tuning, such a method cannot generate images that faithfully match the input appearances. Our method not only addresses these issues but also handles stochastic appearances (e.g., tiny differences in wrinkles) from different random seeds. We report 2D generative evaluation metrics and more results in Supp-Sec. 8.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.733, + 0.232, + 0.749 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We conducted controlled experiments to validate the effectiveness of our proposed modules. As shown in Fig. 8, the skinned body mesh is a crucial component for 3D human reconstruction. Without this body mesh as a 3D anchor, the output mesh contains an incorrect body shape due to the depth ambiguity issue. Conversely, removing the hallucination module has minimal impact on 3D reconstruction metrics, though it slightly degrades normal consistency. However, the overall quality in both texture and geometry is incomparable with our full model (see Fig. 8 right). This is" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.895, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.196, + 0.895, + 0.239 + ], + "angle": 0, + "content": "Figure 8. Ablation study. We visualize back and side-view rendering of the reconstructed meshes. Our full model produced a correct body shape and more realistic clothing geometry." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.241, + 0.891, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Figure 9. Applications. Our pipeline can be incorporated with text-to-image diffusion models for 3D human creation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.893, + 0.55 + ], + "angle": 0, + "content": "consistent with our findings in user studies, indicating that 3D metrics may not accurately reflect the perceptual quality of 3D meshes. Finally, we tested two additional variants, leveraging ground-truth body meshes and real back-view images in our full pipeline, representing the upper bound of our method. As shown in Tab. 3 bottom, this additional information notably improves the 3D metrics. These results highlight the persistent challenges in the single-view reconstruction problem, including pose ambiguity and the stochastic nature of clothing geometry. For more experiments on our design choices, please refer to Supp-Sec. 8.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.635, + 0.573 + ], + "angle": 0, + "content": "4.5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Inheriting the generative capability of LDM, SiTH is robust to diverse inputs, such as out-of-distribution or AI-generated images. We demonstrate a unique solution to link photo-realistic AI photos and high-fidelity 3D humans. In Fig. 9, we introduce a 3D creation workflow integrating powerful text-to-image generative models. Given a body pose, we generate a front-view image using Stable Diffusion and ControlNet using text prompts. SiTH then creates a full-body textured human from the AI-generated image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.715, + 0.62, + 0.73 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We propose an innovative pipeline designed to create fully textured 3D humans from single-view images. Our approach seamlessly integrates an image-conditioned diffusion model into the existing data-driven 3D reconstruction workflow. Leveraging the generative capabilities of the diffusion model, our method efficiently produces lifelike 3D humans from a diverse range of unseen images in under two minutes. We expect our work will advance the application of generative AI in 3D human creation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was partially supported by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\"." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "545" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.441, + 0.129 + ], + "angle": 0, + "content": "[1] Renderpeople, https://renderpeople.com/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.468, + 0.17 + ], + "angle": 0, + "content": "[2] Stable diffusion image variations. huggingface . co / lambdaabs / stable - diffusion - image -conditioned.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.469, + 0.226 + ], + "angle": 0, + "content": "[3] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[4] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.469, + 0.34 + ], + "angle": 0, + "content": "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.342, + 0.468, + 0.396 + ], + "angle": 0, + "content": "[6] Mikolaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD gans. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Proceedings of the European Conference on Computer Vision (ECCV). Springer International Publishing, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.566, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[9] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[10] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[12] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing relightable and animatable avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[16] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human pose estimation in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[18] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[19] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Hsuan-I Ho, Lixin Xue, Jie Song, and Otmar Hilliges. Learning locally editable virtual humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[25] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Selfrecon: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "546" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[28] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[30] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[32] Manuel Kaufmann, Velko Vechev, and Dario Mylonopoulos. aitviewer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[33] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (TOG), 32(3):1-13, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.519 + ], + "angle": 0, + "content": "[34] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[35] Byungjun Kim, Patrick Kwon, Kwangho Lee, Myunggi Lee, Sookwan Han, Daesik Kim, and Hanbyul Joo. Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[43] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[44] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[45] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[46] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[48] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[50] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "547" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[58] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.664 + ], + "angle": 0, + "content": "[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[60] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[61] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[62] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[63] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid represent" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "tation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.134, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[64] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.204, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[65] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[66] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 703-735. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[67] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[69] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[70] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[71] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 641-676. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[73] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[74] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "548" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[75] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[76] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[79] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "549" + } + ] +] \ No newline at end of file diff --git a/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_origin.pdf b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4960052fccd02d3579eb23125ee69d1e45fed12c --- /dev/null +++ b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/978debbe-87ee-4742-aad1-3f1d3627166b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47ca63fcd3f141822df6e324f9ed0bce752aba7c8a03f01f1a276ab5400cdf3c +size 5509054 diff --git a/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/full.md b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..51e91f9d21f01f239ad18c3f95d430af583a2b30 --- /dev/null +++ b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/full.md @@ -0,0 +1,377 @@ +# SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion + +Hsuan-I Ho + +Jie Song + +Otmar Hilliges + +Department of Computer Science, ETH Zürich + +https://ait.ethz.ch/sith + +![](images/2c22fb8934c9c87c6cb9274b9744967f2112240041104535705c47084e1e5ae9.jpg) +Real Photo + +![](images/11ae73ac30fb082476415d5050abcc1f16b7882d4caa311255e0347ffc497ca1.jpg) +Back-view Hallucination + +![](images/fae3d85f80dfe4f4d8c6753ce1922658ee8e849510dd406bd2944197ec4ffe81.jpg) + +![](images/de9dc06b1659efad7fed662abd7483f403cb744c7bf87492e3b62c48f7593acd.jpg) +Mesh Reconstruction +AI-generated Image +Figure 1. Single-view textured human reconstruction. SiTH is a novel pipeline for creating high-quality and fully textured 3D human meshes from single images. We first hallucinate back-view appearances through an image-conditioned diffusion model, followed by the reconstruction of full-body textured meshes using both the front and back-view images. Our pipeline enables the creation of lifelike and diverse 3D humans from unseen photos (left) and AI-generated images (right). + +![](images/a7b2fea9a0ad8bfd1fe7e46abd9ffdc7432b798ab74d9a7e8afa3d064d91a1a8.jpg) +"A woman in a white T-shirt and a blue tennis skirt" + +![](images/74a63743f32178e548ce7612b440d25fd45baf500a3cbafc46a9b9db40945de7.jpg) + +![](images/801585909fa333b7f9bd7ad9dbd445032270cfb27b65bef790b7bb0d11dcb800.jpg) +e Back-view Hallucination +Mesh Reconstruction + +# Abstract + +A long-standing goal of 3D human reconstruction is to create lifelike and fully detailed 3D humans from single-view images. The main challenge lies in inferring unknown body shapes, appearances, and clothing details in areas not visible in the images. To address this, we propose SiTH, a novel pipeline that uniquely integrates an image-conditioned diffusion model into a 3D mesh reconstruction workflow. At the core of our method lies the decomposition of the challenging single-view reconstruction problem into generative hallucination and reconstruction subproblems. For the former, we employ a powerful generative diffusion model to hallucinate unseen back-view appearance based on the input images. For the latter, we leverage skinned body meshes as guidance to recover full-body texture meshes from the input and back-view images. SiTH requires as few as 500 3D human scans for training while maintaining its generality and robustness to diverse images. Extensive evaluations on two 3D human benchmarks, including our newly created one, highlighted our method's superior accuracy and perceptual quality in 3D textured human reconstruction. + +# 1. Introduction + +With the growing popularity of 3D and virtual reality applications, there has been increasing interest in creating realistic 3D human models. In general, crafting 3D humans is labor-intensive, time-consuming, and requires collaboration from highly skilled professionals. To bring lifelike 3D humans to reality and to support both expert and amateur creators in this task, it is essential to enable users to create textured 3D humans from simple 2D images or photos. + +Reconstructing a fully textured human mesh from a single-view image presents an ill-posed problem with two major challenges. Firstly, the appearance information required for generating texture in unobserved regions is missing. Secondly, 3D information for mesh reconstruction, such as depth, surface, and body pose, becomes ambiguous in a 2D image. Previous efforts [4, 60, 79] attempted to tackle these challenges in a data-driven manner, focusing on training neural networks with image-mesh pairs. However, these approaches struggle with images featuring unseen appearances or poses, due to limited 3D human training data. More recent studies [61, 73, 79] introduced additional 3D reasoning modules to enhance robustness against unseen + +poses. Yet, generating realistic and full-body textures from unseen appearances still remains an unsolved problem. + +To address the above challenges, we propose SiTH, a novel pipeline that integrates an image-conditioned diffusion model to reconstruct lifelike 3D textured humans from monocular images. At the core of our approach is the decomposition of the challenging single-view problem into two subproblems: generative back-view hallucination and mesh reconstruction. This decomposition enables us to exploit the generative capability of pretrained diffusion models to guide full-body mesh and texture reconstruction. The workflow is depicted in Fig. 1. Given a front-view image, the first stage involves hallucinating a perceptually consistent back-view image using image-conditioned diffusion. The second stage reconstructs full-body mesh and texture, utilizing both the front and back-view images as guidance. + +More specifically, we employ the generative capabilities of pretrained diffusion models (e.g. Stable Diffusion [57]) to infer unobserved back-view appearances for full-body 3D reconstruction. The primary challenge in ensuring the realism of 3D meshes lies in generating images that depict spatially aligned body shapes and perceptually consistent appearances with the input images. While diffusion models demonstrate impressive generative abilities with text conditioning, they are limited in producing desired back-view images using the frontal images as image conditions. To overcome this, we adapt the network architecture to enable conditioning on frontal images and introduce additional trainable components following ControlNet [77] to provide pose and mask control. To fully tailor this model to our task while retaining its original generative power, we carefully fine-tune the diffusion model using multi-view images rendered from 3D human scans. Complementing this generative model, we develop a mesh reconstruction module to recover full-body textured mesh from front and back-view images. We follow prior work in handling 3D ambiguity through normal [61] and skinned body [73, 79] guidance. It is worth noting that the models for both subproblems are trained using the same public THuman2.0 [75] dataset, which consists of as few as 500 scans. + +To advance research in single-view human reconstruction, we created a new benchmark based on the high-quality CustomHumans [22] dataset and conducted comprehensive evaluations against state-of-the-art methods. Compared to existing end-to-end methods [4, 60, 79], our two-stage pipeline can recover full-body textured meshes, including back-view details, and demonstrates robustness to unseen images. In contrast to time-intensive diffusion-based optimization methods [24, 35, 41], our pipeline efficiently produces high-quality textured meshes in under two minutes. Moreover, we explored applications combining text-guided diffusion models, showing SiTH's versatility in 3D human creation. Our contributions are summarized as follows: + +- We introduce SiTH, a single-view human reconstruction pipeline capable of producing high-quality, fully textured 3D human meshes within two minutes. +- Through decomposing the single-view reconstruction task, SiTH can be efficiently trained with public 3D human scans and is more robust to unseen images. +- We establish a new benchmark featuring more diverse subjects for evaluating textured human reconstruction. + +# 2. Related Work + +Single-view human mesh reconstruction. Reconstructing 3D humans from monocular inputs [14, 17, 26, 28, 29, 62, 67, 71] has gained more popularity in research. In this context, we focus on methods that recover 3D human shapes, garments, and textures from a single image. As a seminal work, Saito et al. [60] first proposed a data-driven method with pixel-aligned features and neural fields [72]. Its follow-up work PIFuHD [61] further improved this framework with high-res normal guidance. Later approaches extended this framework with additional human body priors. For instance, PaMIR [79] and ICON [73] utilized skinned body models [42, 51] to guide 3D reconstruction. ARCH [25], $\mathrm{ARCH} + +$ [21], and CAR [39] transformed global coordinates into the canonical coordinates to allow for reposing. PHOHRUM [4] and S3F [12] further disentangled shading and albedo to enable relighting. Another line of work replaced the neural representations with conventional Poisson surface reconstruction [33, 34]. ECON [74] and 2K2K [18] trained normal and depth predictors to generate front and back 2.5D point clouds. The human mesh is obtained by fusing these point clouds with body priors and 3D heuristics. However, none of these methods produce realistic full-body texture and geometry in the unobserved regions. Our pipeline addresses this problem by incorporating a generative diffusion model into the 3D human reconstruction workflow. + +3D generation with 2D diffusion models. Diffusion models [52, 56, 57, 59] trained with large collections of images have demonstrated unprecedented capability in creating 3D objects from text prompts. Most prior work [11, 19, 40, 47, 53, 68] followed an optimization workflow to update 3D representations (e.g. NeRF [48], SDF tetrahedron [63]) via neural rendering [66] and a score distillation sampling (SDS) [53] loss. While some methods [3, 10, 24, 27, 64, 76] applied this workflow to human bodies, they cannot produce accurate human bodies and appearances due to the ambiguity of text-conditioning. More recent work [41, 54] also tried to extend this workflow with more accurate image-conditioning. However, we show that they struggle to recover human clothing details and require a long optimization time. Most related to our work is Chupa [35], which also decomposes its pipeline into two stages. Note that Chupa is an optimization-based approach that relies on texts + +![](images/3ecbf5bd8e7652b382bdf6e6652f9c9a93828a0744e17fc309b49a8cbbb915d4.jpg) +Figure 2. Method overview. SiTH is a two-stage pipeline composed of back-view hallucination and mesh reconstruction. The back-view hallucination module samples perceptually consistent back-view images through an iterative denoising process conditioned on the input image, UV map, and silhouette mask (Sec. 3.1). Based on the input and generated back-view images, the mesh reconstruction module recovers a full-body mesh and textures leveraging a skinned body prior as guidance (Sec. 3.2). Note that both modules in the pipeline can be trained with the same public 3D human dataset and generalize unseen images. + +and cannot model colors. We address these issues by introducing an image-conditioning strategy and model. Most importantly, our method swiftly reconstructs full-texture human meshes without any optimization process. + +Diffusion models adaptation. Foundation models [8, 13, 20, 36] trained on large-scale datasets have been shown to be adaptable to various downstream tasks. Following this trend, pretrained diffusion models [52, 56, 57, 59] have become common backbones for generative modeling. For instance, they can be customized by finetuning with a small collection of images [23, 37, 58]. ControlNet [77] introduced additional trainable plugins to enable image conditioning such as body skeletons. While these strategies have been widely adopted, none of them directly fit our objective. More relevant to our task is DreamPose [31], which utilizes DensePose [16] images as conditions to repose input images. However, it cannot handle out-of-distribution images due to overfitting. Similarly, Zero-1-to-3 [41] finetunes a diffusion model with multi-view images to allow for viewpoint control. However, we show that viewpoint conditioning is not sufficient for generating consistent human bodies. Our model addresses this issue by providing accurate body pose and mask conditions for back-view hallucination. + +# 3. Methodology + +Method overview. Given an input image of a human body and estimated SMPL-X [51] parameters, SiTH produces a full-body textured mesh. This mesh not only captures the observed appearances but also recovers geometric and textural details in unseen regions, such as clothing wrinkles on the back. The pipeline is composed of two modules and is summarized in Fig. 2. In the first stage, we hallucinate unobserved appearances leveraging the generative power of an image-conditioned diffusion model (Sec. 3.1). In the second + +stage, we reconstruct a full-body textured mesh given the input front-view image and the generated back-view image as guidance (Sec. 3.2). Notably, both modules are efficiently trained with 500 textured human scans in THuman2.0 [75]. + +# 3.1. Back-view Hallucination + +Preliminaries. Given an input front-view image $I^{F} \in \mathbb{R}^{H \times W \times 3}$ , our goal is to infer a back-view image $I^{B} \in \mathbb{R}^{H \times W \times 3}$ which depicts unobserved body appearances. This task is under-constrained since there are multiple possible solutions to the same input images. Taking this perspective into account, we leverage a latent diffusion model (LDM) [57] to learn a conditional distribution of back-view images given a front-view image. First, a VAE autoencoder, consisting of an encoder $\mathcal{E}$ and a decoder $\mathcal{D}$ , is pretrained on a corpus of 2D natural images through image reconstruction, i.e., $\tilde{I} = \mathcal{D}(\mathcal{E}(I))$ . Afterwards, an LDM learns to produce a latent code $z$ within the VAE latent distribution $z = \mathcal{E}(I)$ from randomly sampled noise. To sample an image, a latent code $\tilde{z}$ is obtained by iteratively denoising Gaussian noise. The final image is reconstructed through the decoder, i.e., $\tilde{I} = \mathcal{D}(\tilde{z})$ . + +Image-conditioned diffusion model. Simply applying the LDM architecture to our task is not sufficient since our goal is to learn a conditional distribution of back-view images given an input conditional image. To this end, we make several adaptations to allow for image-conditioning as shown in Fig. 3. First, we utilize the pretrained CLIP [55] image encoder and VAE encoder $\mathcal{E}$ to extract image features from the front-view image (i.e., $I^{F}$ ). These image features are used for conditioning the LDM, ensuring the output image shares a consistent appearance with the input image. Second, we follow the idea of ControlNet [77] and propose to use a UV map $(I_{UV}^{B} \in \mathbb{R}^{H \times W \times 3})$ and a silhou + +![](images/2067ee22552ec0b33f3bce63075eb345743003e883df4eadbb64da2f4bd71c9f.jpg) +Figure 3. Training of back-view hallucination module. We employ a pretrained LDM and ControlNet architecture to enable image conditioning. To train our model, we render training pairs of conditional images $I^{F}$ and ground-truth images $I^{B}$ from 3D human scans. Given a noisy image latent $z_{t}$ , the model predicts added noise $\epsilon$ given the conditional image $I^{F}$ , UV map $I_{UV}^{B}$ , and mask $I_{M}^{B}$ as conditions. We train the ControlNet model and cross-attention layers while keeping other parameters frozen. + +ette mask $(I_M^B\in \mathbb{R}^{H\times W})$ from the back view as additional conditions. These conditional signals provide additional information that ensures the output image has a similar body shape and pose to the conditional input image. + +Learning hallucination from pretraining. Another challenge in training an image-conditioned LDM is data. Training the model from scratch is infeasible due to the requirement of a large number of paired images rendered from 3D textured human scans. Inspired by the concept of learning from large-scale pretraining [13, 20], we build our image-conditioned LDM on top of a pretrained diffusion U-Net [57]. We utilize the finetuning strategy [37, 77] to optimize cross-attention layers and ControlNet parameters while keeping most of the other parameters frozen (see Fig. 3). The design and training strategy of our image-conditioned diffusion model enables hallucinating plausible back-view images that are cosistent with the frontal inputs. + +Training and inference. To generate pairwise training images from 3D human scans, we sample camera view angles and use orthographic projection to render RGBA images from 3D scans and UV maps from their SMPL-X fits. Given a pair of images rendered by a frontal and its corresponding back camera, the first image serves as the conditional input $I^{F}$ while the other one is the ground-truth image $I^{B}$ . During training, the ground-truth latent code $z_{0} = \mathcal{E}(I^{B})$ is perturbed by the diffusion process in t time steps, resulting in a noisy latent $z_{t}$ . The image-conditioned LDM model $\epsilon_{\theta}$ aims to predict the added noise $\epsilon$ given the noisy latent $z_{t}$ , the time step $t \sim [0,1000]$ , the conditional image $I^{F}$ , the silhouette mask $I_{M}^{B}$ , and the UV map $I_{UV}^{B}$ (See Fig. 3). The objective function for fine-tuning can be + +![](images/8b77da9a94f91b4745165cdb3f3c431a3d3a5c8fb7de9ee1ee9816e5081cf82e.jpg) +Figure 4. Mesh reconstruction module. Given front and back-view images $(I^{F}, I^{B})$ we predict their normal images $(N^{F}, N^{B})$ through a learned normal predictor. A 3D point $\mathbf{x}$ is projected onto these images for querying pixel-aligned features $(\mathbf{f}_{d,x}, \mathbf{f}_{r,x})$ . To leverage human body mesh as guidance, we embed the point $\mathbf{x}$ into the local UV coordinates $\mathbf{u}_c$ , vector $\mathbf{n_c}$ , distance $d_c$ , and visibility $v_c$ . Finally, two decoders $(H_d, H_r)$ predict SDF and RGB values at $\mathbf{x}$ given the positional embedding and pixel-aligned features. + +represented as: + +$$ +\min _ {\theta} \mathbb {E} _ {z \sim \mathcal {E} (I), t, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\right) \| _ {2} ^ {2}. \tag {1} +$$ + +At test time, we obtain $I_{UV}^{B}, I_{M}^{B}$ from an off-the-shelf pose predictor [9] and segmentation model [36]. To infer a back-view image, we sample a latent $\tilde{z}_0$ by performing the iterative denoising process starting from a Gaussian noise $z_T \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . The back-view image can be obtained by: + +$$ +\tilde {I} ^ {B} = \mathcal {D} (\tilde {z} _ {0}) = \mathcal {D} \left(f _ {\theta} \left(z _ {T}, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\right)\right), \tag {2} +$$ + +where $f_{\theta}$ is a function representing the iterative denoising process of our image-conditioned LDM (See Fig. 2 left). + +# 3.2. Human Mesh Reconstruction + +After obtaining the back-view image, our goal is to construct a full-body human mesh and its textures using the input and back-view image as guidance. We follow the literature [60, 61] to model this task with a data-driven method. Given pairwise training data (i.e., front/back-view images and 3D scans), we learn a data-driven model that maps these images to a 3D representation (e.g., a signed distance field (SDF)). We define this mapping as below: + +$$ +\begin{array}{l} \Phi : \mathbb {R} ^ {H \times W \times 3} \times \mathbb {R} ^ {H \times W \times 3} \times \mathbb {R} ^ {3} \rightarrow \mathbb {R} \times \mathbb {R} ^ {3} \tag {3} \\ (I ^ {F}, I ^ {B}, \mathbf {x}) \mapsto d _ {x}, \mathbf {r} _ {x}, \\ \end{array} +$$ + +where $\mathbf{x}$ is the 3D coordinate of a query point, and $d_x, \mathbf{r}_x$ denote the signed distance and RGB color value at point $\mathbf{x}$ . The network components we used for learning the mapping function are depicted in Fig. 4. + +![](images/1b1ebc1e3bb2f1b8e8dbe76773adfb901a8c15154e5c7b9c351d313e6e735471.jpg) +Input Image + +![](images/51b68c462af8841dc78e08284c8c6bf4b6a0c2cacab6f92fbd9dae42cac2c4f3.jpg) +PaMIR + +![](images/9f2f38dfba8c3297aab94ddfe51f92f9f62ae4bc3026f994d4c0ba0343781720.jpg) +PHORHUM + +![](images/b2fc4efbbc189511464cfe449720b5b54869f3fcbc7a7e4c15961bfdedbe1f76.jpg) +SiTH (Ours) + +![](images/213325818320f1edd3af26deb3df9d5751bec21f3b7d9c7146f1451dce178187.jpg) +GT References + +![](images/1ca033b379b2315af5852f306aa9940595d3877c11d403404dc131f21aa8f3dd.jpg) +Input Image + +![](images/a1def554032d40cb086d01cb1b9cc9f496729d20005c53fc5c4401aa9dddcaa6.jpg) +ECON + +![](images/b02915f2e73fcd1206dafbf49afcc646577574ab699f7e4b6b2b9f81c281d78f.jpg) +PIFuHD + +![](images/281ae0046b6f7412000e4eadbe5585bbd456a1f54bba6fccfa3d63159ed6cdeb.jpg) +SiTH (Ours) +Figure 5. Qualitative comparison on CustomHumans. Top: Results of methods generating mesh and texture. Bottom: Results of methods generating mesh only. Note that single-view reconstruction is not possible to replicate exact back-view texture and geometry. Our method generates realistic texture and clothing wrinkles perceptually close to the real scans while other baselines only produce smooth colors and surfaces in the back regions. Best viewed in color and zoom in. + +![](images/464fc5cbf953c8d367ba6074c1d569e7bdfc086a9653a29b800bdcda1a5a1f76.jpg) +GT References + +Local feature querying. To learn a generic mapping function that is robust to unseen images, it is important that the model is conditioned solely on local image information with respect to the position of $\mathbf{x}$ . Therefore, we employ the idea of pixel-aligned feature querying [60, 61] and separate our model into two branches, i.e., color and geometry. Our model contains a normal predictor that converts the RGB image pair $(I^{F}, I^{B})$ into normal maps $(N^{F}, N^{B})$ . Two image feature encoders $G_{d}, G_{r}$ then extract color and geometry feature maps $(\mathbf{f}_{d}, \mathbf{f}_{r}) \in \mathbb{R}^{H' \times W' \times D}$ from the images and normal maps respectively (for simplicity we describe the process for a single image and leave out the superscripts, but both front and back images are treated the same). Finally, we project the query point $\mathbf{x}$ onto the image coordinate (Fig. 4 red points) to retrieve the local features $(\mathbf{f}_{d,x}, \mathbf{f}_{r,x}) \in \mathbb{R}^{D}$ : + +$$ +\begin{array}{l} \mathbf {f} _ {d, x} = \mathcal {B} (\mathbf {f} _ {d}, \pi (\mathbf {x})) = \mathcal {B} (G _ {d} (N), \pi (\mathbf {x})), \\ \mathbf {f} _ {r, x} = \mathcal {B} (\mathbf {f} _ {r}, \pi (\mathbf {x})) = \mathcal {B} (G _ {r} (I), \pi (\mathbf {x})), \\ \end{array} +$$ + +where $\mathcal{B}$ is a local feature querying operation using bilinear interpolation and $\pi (\cdot)$ denotes orthographic projection. + +# Local positional embedding with skinned body prior. + +As mentioned in Sec. 1, a major difficulty in mesh reconstruction is 3D ambiguity where a model has to infer unknown depth information between the front and back images. To address this issue, we follow prior work [22, 73, 79] leveraging a skinned body mesh [51] for guiding the reconstruction task. This body mesh is regarded as an anchor that provides an approximate 3D shape of the human body. + +To exploit this body prior, we devise a local positional embedding function that transforms the query point $\mathbf{x}$ into the local body mesh coordinate system. We look for the closest point $\mathbf{x}_c^*$ on the body mesh (Fig. 4 blue point), i.e., + +$$ +\mathbf {x} _ {c} ^ {*} = \arg \min _ {\mathbf {x} _ {c}} \| \mathbf {x} - \mathbf {x} _ {c} \| _ {2}, \tag {5} +$$ + +where $\mathbf{x}_c$ are points on the skinned body mesh $\mathcal{M}$ . Our positional embedding $\mathbf{p}$ constitutes four elements: a signed distance value $d_c$ between $\mathbf{x}_c^*$ and $\mathbf{x}$ , a vector $\mathbf{n}_c = (\mathbf{x} - \mathbf{x}_c^*)$ , the UV coordinates $\mathbf{u}_c \in [0,1]^2$ of the point $\mathbf{x}_c^*$ , and a visibility label $v_c \in \{1, -1, 0\}$ that indicates whether $\mathbf{x}_c^*$ is visible in the front/back image or neither. Finally, two separate MLPs $H_d, H_r$ take the positional embedding $\mathbf{p} = [d_c, \mathbf{n}_c, \mathbf{u}_c, v_c]$ and the local texture/geometry features $(\mathbf{f}_{d,x}, \mathbf{f}_{r,x})$ as inputs to predict the final SDF and RGB values at point $\mathbf{x}$ : + +$$ +\begin{array}{l} d _ {x} = H _ {d} \left(\mathbf {f} _ {d, x} ^ {F}, \mathbf {f} _ {d, x} ^ {B}, \mathbf {p}\right), \\ \mathbf {r} _ {x} = H _ {r} \left(\mathbf {f} _ {r, x} ^ {F}, \mathbf {f} _ {r, x} ^ {B}, \mathbf {p}\right). \\ \end{array} +$$ + +Training and inference. We used the same 3D dataset described in Sec. 3.1 to render training image pairs $(I^{F}, I^{B})$ from the 3D textured scans. For each training scan, query points $\mathbf{x}$ are sampled within a 3-dimensional cube $[-1, 1]^{3}$ . For each point, we compute the ground-truth signed distance values $d$ to the scan surface, closest texture RGB values $r$ , and surface normal $\mathbf{n}$ . Finally, we jointly optimized the normal predictors, the image encoders, and the MLPs in + +
CAPE [45]CustomHuman [22]
MethodCD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F (×10-2) ↓CD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F / B (×10-2) ↓
PIFu [60]2.368 / 3.7630.77833.8422.7202.209 / 2.5820.80534.8816.073 / 8.496
PIFuHD [61]2.401 / 3.5220.77235.706-2.107 / 2.2280.80439.076-
PaMIR [79]2.190 / 2.8060.80436.7252.0852.181 / 2.5070.81335.8474.646 / 7.152
2K2K [18]2.478 / 3.6830.78228.700-2.488 / 3.2920.79630.186-
FOF [15]2.196 / 4.0400.77734.227-2.079 / 2.6440.80836.013-
ICON [73]2.516 / 3.0790.78629.630-2.256 / 2.7950.79130.437-
ECON [74]2.475 / 2.9700.78830.488-2.483 / 2.6800.79730.894-
SiTH (Ours)1.899 / 2.2610.81637.7631.9771.871 / 2.0450.82637.0293.929 / 6.803
+ +Table 1. Single-view human reconstruction benchmarks. We report Chamfer distance (CD), normal consistency (NC), and f-score between ground truth and predicted meshes. To evaluate texture reconstruction quality, we compute LPIPS between the image rendering of GT and generated textures. The best and the second best methods are highlighted in bold and underlined respectively. Note that gray color denotes models trained on more commercial 3D human scans while the others are trained on the public THuman2.0 dataset. + +both branches with the following reconstruction losses: + +$$ +\mathcal {L} _ {d} = \left\| d - d _ {x} \right\| _ {1} + \lambda_ {n} \left(1 - \mathbf {n} \cdot \nabla_ {\mathbf {x}} d _ {x}\right), \tag {7} +$$ + +$$ +\mathcal {L} _ {r} = \left\| \mathbf {r} - \mathbf {r} _ {x} \right\| _ {1}. \tag {8} +$$ + +Note that $\nabla_{\mathbf{x}}$ indicates numerical finite differences for computing local normals at point $\mathbf{x}$ and $\lambda_{n}$ is a hyperparameter. + +During inference, we use the input image $I^{F}$ and the back-view image $\tilde{I}^{B}$ obtained from Sec. 3.1 to reconstruct 3D mesh and textures. First, we align both images with the estimated body mesh $\mathcal{M}$ to ensure that image features can be properly queried around the 3D anchor. We adopt a similar strategy of SMPLify [7] to optimize the scale and the offset of the body mesh with silhouette and 2D joint errors. Finally, we perform the marching cube algorithm [43] by querying SDF and RGB values within a dense voxel grid via Eq. (3) (see Fig. 2 right). + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. Previous work relied on training data from commercial datasets such as RenderPeople [1]. While these datasets offer high-quality textured meshes, they also limit reproducibility due to limited accessibility. For fair comparisons, we follow ICON [73] by training our method on the public 3D dataset THuman2.0 [75] and using the CAPE [45] dataset for evaluation. However, we observed potential biases in the evaluation due to the low-res groundtruth meshes and image rendering defects in the CAPE dataset (for a detailed discussion, please refer to SuppSec. 6). Consequently, we further create a new benchmark that evaluates the baselines on a higher-quality 3D human dataset CustomHumans [22]. In the following, we provide a summary of the datasets used in our experiments: + +- THuman2.0 [75] contains approximately 500 scans of humans wearing 150 different garments in various poses. We use these 3D scans as the training data. + +- CAPE [45] contains 15 subjects in 8 types of tight outfits. The test set, provided by ICON, consists of 100 meshes. We use CAPE for the quantitative evaluation (Sec. 4.2). +- CustomHumans [22] contains 600 higher-quality scans of 80 subjects in 120 different garments and varied poses. We selected 60 subjects for all quantitative experiments, user studies, and ablation studies. (Sec. 4.2 - Sec. 4.4) + +Evaluation protocol. We follow the evaluation protocol in OccNet [46] and ICON [73] to compute 3D metrics Chamfer distance (CD), normal consistency (NC), and f-Score [65] on the generated meshes. To evaluate reconstructed mesh texture, we report LPIPS [78] of front and back texture rendering. In user studies, 30 participants rank the meshes obtained by four different methods. We report the average ranking ranging from 1 (best) to 4 (worst). + +# 4.2. Single-view Human Reconstruction + +Benchmark evaluation. We compared SiTH with state-of-the-art single-view human reconstruction methods, including PIFu [60], PIFuHD [61], PaMIR [79], FOF [15], ICON [73], PHORHUM [4], 2K2K [18], and ECON [74] on CAPE and CustomHumans. Note that PHORHUM is only used for qualitative comparison since a different camera system is used, leading to the misalignment with ground-truth meshes. We visualize the generated mesh texture and normals in Fig. 5. Existing methods produce oversmoothed texture and normals, particularly in the back. Our method not only generates photorealistic and perceptually consistent appearances in unobserved regions but also recovers underlying geometric details like clothing wrinkles. + +The quantitative results are summarized in Tab. 1. It's worth noting that most methods are trained with commercial datasets (gray color in Tab. 1), while the others are trained on the public THuman2.0 dataset. To evaluate the methods leveraging a skinned body prior (i.e., PaMIR, ICON, ECON, FOF, and SiTH), we use the same pose alignment procedure in their original implementations for a fair + +![](images/e4f28041b59e1a8da72898ae9c70f8ee695b29116f6c8dadd633686dea730304.jpg) +Input Image + +![](images/ee2961317ec3147f70fb3982c5f1270386614287fd779d481d5eeb72b4073aca.jpg) +Pix2PixHD + +![](images/7cb24378c11b722c53a0558be43d6a0283e39592288fb718f98c06d959e0e775.jpg) +DreamPose + +![](images/54f3dbe7a06ed222c79335b776006682ad7a9bf66a1802f85b3bcf7964e9c566.jpg) +ControlNet + Interrogate + +![](images/06031aa23c7972a25329344d79444a8a64042082486556994610c74404c97bcb.jpg) + +![](images/f5d0c263d25c30f6af5b9c636f40a4a0b19cb63a5289e17fef21a1603c211985.jpg) + +![](images/1c9572de8d4770389f1a6d658f3cf55826d8b1c88c5f780058571c6382233a7b.jpg) +Zero-1-to-3 + +![](images/20124d72a1f43ff6b18ee3bab5c9ca1e2f3aa22a4709edc97c23f49d6c1c9282.jpg) + +![](images/05d6c49f43214c6ad6b794f9972be85e9e779abe61042e9acc7787f233b6d86b.jpg) +SiTH (Ours) + +![](images/3afae08893a53dc535c0287a13358930cb20b76981048e8d882565cd52ff36e6.jpg) + +![](images/6dc0dcbdddfadc7f7af2c21016afc2ff1dede7f97a91c1c404b092e89b117b5b.jpg) +GT References + +![](images/84327b293fce65fc3f6028d4454a36766aed3ccd017ee8b542144837ac58bd3f.jpg) +Figure 6. Qualitative comparison of back-view hallucination. We visualize back-view images generated by the baseline methods. Note that the three different images are sampled from different random seeds. Our results are perceptually close to the ground-truth image in terms of appearances and poses. Moreover, our method also preserves generative stochasticity for handling tiny wrinkle changes. +Figure 7. Comparison with optimization-based methods. Compared to methods that utilize diffusion models for optimization, our result is closer to the input image and contains local geometric details. Note that Chupa is not conditioned on the image, and fails to generate correct clothing from the text prompts. + +comparison. Results in Tab. 1 show that the method using a body prior (PaMIR) outperformed the end-to-end method (PIFuHD) on tight clothing and challenging poses in CAPE. However, it falls short in handling diverse outfits in CustomHumans. Moreover, the methods trained on commercial datasets achieve better performance than those trained with public data (ICON, ECON). Notably, our method is robust across both benchmarks, achieving performance comparable to the methods trained on high-quality commercial data. + +Compared with optimization-based methods. We compared SiTH with methods that use pretrained diffusion models and a score distillation sampling loss [53] to optimize 3D meshes. In the case of Zero-1-to-3 [41], we used the input image to optimize an instant NGP [49] radiance field, and for Magic-123 [54], we provided additional text prompts to optimize an SDF tetrahedron [63]. From Fig. 7, we see that while both methods can handle full-body textures, they struggle with reasoning the underlying geometry and clothing details. It is worth noting that Zero-1-to-3 and Magic-123 require 10 minutes and 6 hours in optimization, respectively, while our method takes under 2 minutes to generate a textured mesh with a marching cube of $512^{3}$ resolution. + +
ICONECONPIFuHDOurs
Front Geometry3.1272.7202.6071.547
Back Geometry3.1932.5133.0931.200
Similarity3.0932.6602.7801.456
PIFuPaMIRPHOHRUMOurs
Front Texture3.0672.1533.4501.327
Back Texture3.4502.3553.1401.054
Similarity3.3072.1923.4161.093
ChupaOurs
User Preference36.0%64.0%
+ +Table 2. User study results. Top: 30 users are asked to rank the quality of surface normal images from best (1) to worst (4). We report the average ranking of each method. Middle: Similar to the first task, users are asked to rank the quality of RGB textures. Bottom: We ask users to choose the mesh with a better quality. + +Additionally, more similar to our method is Chupa [35], which generates front/back-view normals for mesh reconstruction. Note that Chupa is not conditioned on images and does not generate texture. Instead, we provided body poses and text prompts generated by an image-to-text interrogator [38] as their conditional inputs. From Fig. 7, it's clear that text-conditioning is less accurate than image-conditioning, and the method struggles to generate unseen clothing styles such as coats. By contrast, our method can reconstruct correct clothing geometry and texture from unseen images. We present more discussions and comparisons with optimization-based methods in Supp-Sec. 9.1. + +User study. The above metrics may not fully capture the quality of 3D meshes in terms of realism and local details. To address this, we conducted a user study to compare the texture and geometry quality among various baselines. We invited 30 users to rank the front/back-view texture and normal renderings of 3D meshes generated by four different methods. Additionally, we asked the users to assess the similarity between the input images and the generated meshes. The results (Tab. 2) support our claim that existing methods struggle to efficiently generate desirable back-view texture/geometry from single-view images. Our method, + +
MethodCD (cm)↓NC↑f-Score↑
W/o Body Mesh2.4710.80133.244
W/o Hallucination1.9600.84036.677
Full Pipeline1.9580.82637.029
W/ GT Body Mesh1.1720.89158.858
W/ GT Body and I^B1.0590.91463.356
+ +Table 3. Ablation study on CustomHumans. We ablate the hallucination module and the skinned body mesh in our pipeline. Please refer to our discussion in Sec. 4.4. + +which leverages the generative capability of diffusion models, consistently outperforms each baseline. It also produces more preferred front-view textures and geometries, as evidenced by higher user rankings. We also conducted a user study with Chupa (in Tab. 2 bottom) which also indicates more users prefer the 3D meshes generated by our method. + +# 4.3. Generative Capability + +Image quality comparison. Our hallucination module is a unique and essential component that generates spatially aligned human images to guide 3D mesh reconstruction. Given that our focus is on back-view hallucination, we compare the quality of generated images with the relevant generative methods in Fig. 6. We trained a baseline Pix2PixHD [69] model, which produced smooth and blurry results on unseen images due to overfitting to 500 subjects. Another method closely related to ours is DreamPose [31], which conditions the model with DensePose images and finetunes the diffusion model with paired data. However, their model failed to handle unseen images, in contrast to our approach. While Zero-1-to-3 [41] can generalize to unseen images, their method faces challenges in generating consistent body poses given the same back-view camera. Moreover, we designed another baseline that provides ControlNet [77] for corresponding text prompts using an image-to-text interrogator [38]. However, without proper image conditioning and fine-tuning, such a method cannot generate images that faithfully match the input appearances. Our method not only addresses these issues but also handles stochastic appearances (e.g., tiny differences in wrinkles) from different random seeds. We report 2D generative evaluation metrics and more results in Supp-Sec. 8.1. + +# 4.4. Ablation Study + +We conducted controlled experiments to validate the effectiveness of our proposed modules. As shown in Fig. 8, the skinned body mesh is a crucial component for 3D human reconstruction. Without this body mesh as a 3D anchor, the output mesh contains an incorrect body shape due to the depth ambiguity issue. Conversely, removing the hallucination module has minimal impact on 3D reconstruction metrics, though it slightly degrades normal consistency. However, the overall quality in both texture and geometry is incomparable with our full model (see Fig. 8 right). This is + +![](images/c192e9142d0984e52a76ce26c54d1a8619edce1a63c6b0c8574f8f72e1274b69.jpg) + +![](images/181415364ee9a723664cc01615b1cc12cd1c0cb6851e80eea47f528b4ed42532.jpg) +Figure 8. Ablation study. We visualize back and side-view rendering of the reconstructed meshes. Our full model produced a correct body shape and more realistic clothing geometry. +Figure 9. Applications. Our pipeline can be incorporated with text-to-image diffusion models for 3D human creation. + +consistent with our findings in user studies, indicating that 3D metrics may not accurately reflect the perceptual quality of 3D meshes. Finally, we tested two additional variants, leveraging ground-truth body meshes and real back-view images in our full pipeline, representing the upper bound of our method. As shown in Tab. 3 bottom, this additional information notably improves the 3D metrics. These results highlight the persistent challenges in the single-view reconstruction problem, including pose ambiguity and the stochastic nature of clothing geometry. For more experiments on our design choices, please refer to Supp-Sec. 8.5. + +# 4.5. Applications + +Inheriting the generative capability of LDM, SiTH is robust to diverse inputs, such as out-of-distribution or AI-generated images. We demonstrate a unique solution to link photo-realistic AI photos and high-fidelity 3D humans. In Fig. 9, we introduce a 3D creation workflow integrating powerful text-to-image generative models. Given a body pose, we generate a front-view image using Stable Diffusion and ControlNet using text prompts. SiTH then creates a full-body textured human from the AI-generated image. + +# 5. Conclusion + +We propose an innovative pipeline designed to create fully textured 3D humans from single-view images. Our approach seamlessly integrates an image-conditioned diffusion model into the existing data-driven 3D reconstruction workflow. Leveraging the generative capabilities of the diffusion model, our method efficiently produces lifelike 3D humans from a diverse range of unseen images in under two minutes. We expect our work will advance the application of generative AI in 3D human creation. + +Acknowledgements. This work was partially supported by the Swiss SERI Consolidation Grant "AI-PERCEIVE". + +# References + +[1] Renderpeople, https://renderpeople.com/.6 +[2] Stable diffusion image variations. huggingface . co / lambdaabs / stable - diffusion - image -conditioned.1 +[3] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia, 2023. 2, 7 +[4] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 4 +[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4 +[6] Mikolaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD gans. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2 +[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Proceedings of the European Conference on Computer Vision (ECCV). Springer International Publishing, 2016. 6 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 3 +[9] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4 +[10] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 2 +[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2 +[12] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing relightable and animatable avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 4 + +[14] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6, 5 +[16] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human pose estimation in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[17] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[18] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4 +[19] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2 +[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4 +[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2 +[22] Hsuan-I Ho, Lixin Xue, Jie Song, and Otmar Hilliges. Learning locally editable virtual humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 1 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 3 +[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2, 7 +[25] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[26] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Selfrecon: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 + +[27] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2 +[28] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[30] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 2 +[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 8, 2 +[32] Manuel Kaufmann, Velko Vechev, and Dario Mylonopoulos. aitviewer, 2022. 4 +[33] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (TOG), 32(3):1-13, 2013. 2 +[34] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2 +[35] Byungjun Kim, Patrick Kwon, Kwangho Lee, Myunggi Lee, Sookwan Han, Daesik Kim, and Hanbyul Joo. Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 7 +[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4 +[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4 +[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 7, 8 +[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 + +[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 7, 8, 1 +[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 2 +[43] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm, 1998. 6 +[44] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2 +[45] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 1, 5 +[46] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[48] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 7 +[50] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2 +[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 5 + +[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3 +[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 7 +[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2, 7 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3 +[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3 +[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4 +[58] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35, 2022. 2, 3 +[60] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2, 4, 5, 6 +[61] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 4, 5, 6 +[62] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[63] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid represent + +tation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 7 +[64] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2 +[65] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[66] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 703-735. Wiley Online Library, 2022. 2 +[67] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2 +[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[69] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8, 2 +[70] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. 2 +[71] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 2 +[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 641-676. Wiley Online Library, 2022. 2 +[73] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6, 4 +[74] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE + +Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4, 5 +[75] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 6, 1 +[76] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 2 +[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 8, 1 +[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 2 +[79] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 5, 6 \ No newline at end of file diff --git a/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/images.zip b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fdb32625f1fd14367f3d979efe030f4970f6d906 --- /dev/null +++ b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51d5e2ef4232c2a0bbe3dbe46db071bece2a8163b2f91bf7b5732730cd9b4fdb +size 691399 diff --git a/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/layout.json b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae5499e15883f9a9dda75d4fd1115e254126cf46 --- /dev/null +++ b/2024/SiTH_ Single-view Textured Human Reconstruction with Image-Conditioned Diffusion/layout.json @@ -0,0 +1,10749 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 457, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 457, + 140 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 457, + 140 + ], + "type": "text", + "content": "SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 161, + 214, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 161, + 214, + 175 + ], + "spans": [ + { + "bbox": [ + 154, + 161, + 214, + 175 + ], + "type": "text", + "content": "Hsuan-I Ho" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 161, + 306, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 161, + 306, + 175 + ], + "spans": [ + { + "bbox": [ + 261, + 161, + 306, + 175 + ], + "type": "text", + "content": "Jie Song" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 352, + 161, + 426, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 161, + 426, + 175 + ], + "spans": [ + { + "bbox": [ + 352, + 161, + 426, + 175 + ], + "type": "text", + "content": "Otmar Hilliges" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 183, + 182, + 411, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 182, + 411, + 197 + ], + "spans": [ + { + "bbox": [ + 183, + 182, + 411, + 197 + ], + "type": "text", + "content": "Department of Computer Science, ETH Zürich" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 199, + 361, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 199, + 361, + 209 + ], + "spans": [ + { + "bbox": [ + 229, + 199, + 361, + 209 + ], + "type": "text", + "content": "https://ait.ethz.ch/sith" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 239, + 121, + 366 + ], + "blocks": [ + { + "bbox": [ + 59, + 239, + 121, + 366 + ], + "lines": [ + { + "bbox": [ + 59, + 239, + 121, + 366 + ], + "spans": [ + { + "bbox": [ + 59, + 239, + 121, + 366 + ], + "type": "image", + "image_path": "2c22fb8934c9c87c6cb9274b9744967f2112240041104535705c47084e1e5ae9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 373, + 110, + 382 + ], + "lines": [ + { + "bbox": [ + 71, + 373, + 110, + 382 + ], + "spans": [ + { + "bbox": [ + 71, + 373, + 110, + 382 + ], + "type": "text", + "content": "Real Photo" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 126, + 243, + 216, + 361 + ], + "blocks": [ + { + "bbox": [ + 126, + 243, + 216, + 361 + ], + "lines": [ + { + "bbox": [ + 126, + 243, + 216, + 361 + ], + "spans": [ + { + "bbox": [ + 126, + 243, + 216, + 361 + ], + "type": "image", + "image_path": "11ae73ac30fb082476415d5050abcc1f16b7882d4caa311255e0347ffc497ca1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 373, + 212, + 382 + ], + "lines": [ + { + "bbox": [ + 129, + 373, + 212, + 382 + ], + "spans": [ + { + "bbox": [ + 129, + 373, + 212, + 382 + ], + "type": "text", + "content": "Back-view Hallucination" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 221, + 230, + 286, + 299 + ], + "blocks": [ + { + "bbox": [ + 221, + 230, + 286, + 299 + ], + "lines": [ + { + "bbox": [ + 221, + 230, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 221, + 230, + 286, + 299 + ], + "type": "image", + "image_path": "fae3d85f80dfe4f4d8c6753ce1922658ee8e849510dd406bd2944197ec4ffe81.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 222, + 301, + 286, + 368 + ], + "blocks": [ + { + "bbox": [ + 222, + 301, + 286, + 368 + ], + "lines": [ + { + "bbox": [ + 222, + 301, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 222, + 301, + 286, + 368 + ], + "type": "image", + "image_path": "de9dc06b1659efad7fed662abd7483f403cb744c7bf87492e3b62c48f7593acd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 373, + 290, + 382 + ], + "lines": [ + { + "bbox": [ + 219, + 373, + 290, + 382 + ], + "spans": [ + { + "bbox": [ + 219, + 373, + 290, + 382 + ], + "type": "text", + "content": "Mesh Reconstruction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 373, + 363, + 382 + ], + "lines": [ + { + "bbox": [ + 297, + 373, + 363, + 382 + ], + "spans": [ + { + "bbox": [ + 297, + 373, + 363, + 382 + ], + "type": "text", + "content": "AI-generated Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 387, + 547, + 431 + ], + "lines": [ + { + "bbox": [ + 45, + 387, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 45, + 387, + 547, + 431 + ], + "type": "text", + "content": "Figure 1. Single-view textured human reconstruction. SiTH is a novel pipeline for creating high-quality and fully textured 3D human meshes from single images. We first hallucinate back-view appearances through an image-conditioned diffusion model, followed by the reconstruction of full-body textured meshes using both the front and back-view images. Our pipeline enables the creation of lifelike and diverse 3D humans from unseen photos (left) and AI-generated images (right)." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 305, + 240, + 452, + 360 + ], + "blocks": [ + { + "bbox": [ + 294, + 225, + 369, + 239 + ], + "lines": [ + { + "bbox": [ + 294, + 225, + 369, + 239 + ], + "spans": [ + { + "bbox": [ + 294, + 225, + 369, + 239 + ], + "type": "text", + "content": "\"A woman in a white T-shirt and a blue tennis skirt\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 240, + 452, + 360 + ], + "lines": [ + { + "bbox": [ + 305, + 240, + 452, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 452, + 360 + ], + "type": "image", + "image_path": "a7b2fea9a0ad8bfd1fe7e46abd9ffdc7432b798ab74d9a7e8afa3d064d91a1a8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 460, + 230, + 522, + 299 + ], + "blocks": [ + { + "bbox": [ + 460, + 230, + 522, + 299 + ], + "lines": [ + { + "bbox": [ + 460, + 230, + 522, + 299 + ], + "spans": [ + { + "bbox": [ + 460, + 230, + 522, + 299 + ], + "type": "image", + "image_path": "74a63743f32178e548ce7612b440d25fd45baf500a3cbafc46a9b9db40945de7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 460, + 300, + 521, + 368 + ], + "blocks": [ + { + "bbox": [ + 361, + 373, + 451, + 382 + ], + "lines": [ + { + "bbox": [ + 361, + 373, + 451, + 382 + ], + "spans": [ + { + "bbox": [ + 361, + 373, + 451, + 382 + ], + "type": "text", + "content": "e Back-view Hallucination" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 300, + 521, + 368 + ], + "lines": [ + { + "bbox": [ + 460, + 300, + 521, + 368 + ], + "spans": [ + { + "bbox": [ + 460, + 300, + 521, + 368 + ], + "type": "image", + "image_path": "801585909fa333b7f9bd7ad9dbd445032270cfb27b65bef790b7bb0d11dcb800.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 373, + 526, + 382 + ], + "lines": [ + { + "bbox": [ + 456, + 373, + 526, + 382 + ], + "spans": [ + { + "bbox": [ + 456, + 373, + 526, + 382 + ], + "type": "text", + "content": "Mesh Reconstruction" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 143, + 441, + 192, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 441, + 192, + 453 + ], + "spans": [ + { + "bbox": [ + 143, + 441, + 192, + 453 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 45, + 470, + 289, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 470, + 289, + 710 + ], + "spans": [ + { + "bbox": [ + 45, + 470, + 289, + 710 + ], + "type": "text", + "content": "A long-standing goal of 3D human reconstruction is to create lifelike and fully detailed 3D humans from single-view images. The main challenge lies in inferring unknown body shapes, appearances, and clothing details in areas not visible in the images. To address this, we propose SiTH, a novel pipeline that uniquely integrates an image-conditioned diffusion model into a 3D mesh reconstruction workflow. At the core of our method lies the decomposition of the challenging single-view reconstruction problem into generative hallucination and reconstruction subproblems. For the former, we employ a powerful generative diffusion model to hallucinate unseen back-view appearance based on the input images. For the latter, we leverage skinned body meshes as guidance to recover full-body texture meshes from the input and back-view images. SiTH requires as few as 500 3D human scans for training while maintaining its generality and robustness to diverse images. Extensive evaluations on two 3D human benchmarks, including our newly created one, highlighted our method's superior accuracy and perceptual quality in 3D textured human reconstruction." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 441, + 386, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 386, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 386, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 461, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 557 + ], + "type": "text", + "content": "With the growing popularity of 3D and virtual reality applications, there has been increasing interest in creating realistic 3D human models. In general, crafting 3D humans is labor-intensive, time-consuming, and requires collaboration from highly skilled professionals. To bring lifelike 3D humans to reality and to support both expert and amateur creators in this task, it is essential to enable users to create textured 3D humans from simple 2D images or photos." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Reconstructing a fully textured human mesh from a single-view image presents an ill-posed problem with two major challenges. Firstly, the appearance information required for generating texture in unobserved regions is missing. Secondly, 3D information for mesh reconstruction, such as depth, surface, and body pose, becomes ambiguous in a 2D image. Previous efforts [4, 60, 79] attempted to tackle these challenges in a data-driven manner, focusing on training neural networks with image-mesh pairs. However, these approaches struggle with images featuring unseen appearances or poses, due to limited 3D human training data. More recent studies [61, 73, 79] introduced additional 3D reasoning modules to enhance robustness against unseen" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "538" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "poses. Yet, generating realistic and full-body textures from unseen appearances still remains an unsolved problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 287, + 266 + ], + "type": "text", + "content": "To address the above challenges, we propose SiTH, a novel pipeline that integrates an image-conditioned diffusion model to reconstruct lifelike 3D textured humans from monocular images. At the core of our approach is the decomposition of the challenging single-view problem into two subproblems: generative back-view hallucination and mesh reconstruction. This decomposition enables us to exploit the generative capability of pretrained diffusion models to guide full-body mesh and texture reconstruction. The workflow is depicted in Fig. 1. Given a front-view image, the first stage involves hallucinating a perceptually consistent back-view image using image-conditioned diffusion. The second stage reconstructs full-body mesh and texture, utilizing both the front and back-view images as guidance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 269, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 288, + 555 + ], + "type": "text", + "content": "More specifically, we employ the generative capabilities of pretrained diffusion models (e.g. Stable Diffusion [57]) to infer unobserved back-view appearances for full-body 3D reconstruction. The primary challenge in ensuring the realism of 3D meshes lies in generating images that depict spatially aligned body shapes and perceptually consistent appearances with the input images. While diffusion models demonstrate impressive generative abilities with text conditioning, they are limited in producing desired back-view images using the frontal images as image conditions. To overcome this, we adapt the network architecture to enable conditioning on frontal images and introduce additional trainable components following ControlNet [77] to provide pose and mask control. To fully tailor this model to our task while retaining its original generative power, we carefully fine-tune the diffusion model using multi-view images rendered from 3D human scans. Complementing this generative model, we develop a mesh reconstruction module to recover full-body textured mesh from front and back-view images. We follow prior work in handling 3D ambiguity through normal [61] and skinned body [73, 79] guidance. It is worth noting that the models for both subproblems are trained using the same public THuman2.0 [75] dataset, which consists of as few as 500 scans." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "To advance research in single-view human reconstruction, we created a new benchmark based on the high-quality CustomHumans [22] dataset and conducted comprehensive evaluations against state-of-the-art methods. Compared to existing end-to-end methods [4, 60, 79], our two-stage pipeline can recover full-body textured meshes, including back-view details, and demonstrates robustness to unseen images. In contrast to time-intensive diffusion-based optimization methods [24, 35, 41], our pipeline efficiently produces high-quality textured meshes in under two minutes. Moreover, we explored applications combining text-guided diffusion models, showing SiTH's versatility in 3D human creation. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 545, + 168 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "type": "text", + "content": "- We introduce SiTH, a single-view human reconstruction pipeline capable of producing high-quality, fully textured 3D human meshes within two minutes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "text", + "content": "- Through decomposing the single-view reconstruction task, SiTH can be efficiently trained with public 3D human scans and is more robust to unseen images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 144, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 168 + ], + "type": "text", + "content": "- We establish a new benchmark featuring more diverse subjects for evaluating textured human reconstruction." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 182, + 392, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 182, + 392, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 392, + 194 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 507 + ], + "type": "text", + "content": "Single-view human mesh reconstruction. Reconstructing 3D humans from monocular inputs [14, 17, 26, 28, 29, 62, 67, 71] has gained more popularity in research. In this context, we focus on methods that recover 3D human shapes, garments, and textures from a single image. As a seminal work, Saito et al. [60] first proposed a data-driven method with pixel-aligned features and neural fields [72]. Its follow-up work PIFuHD [61] further improved this framework with high-res normal guidance. Later approaches extended this framework with additional human body priors. For instance, PaMIR [79] and ICON [73] utilized skinned body models [42, 51] to guide 3D reconstruction. ARCH [25], " + }, + { + "bbox": [ + 304, + 198, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{ARCH} + +" + }, + { + "bbox": [ + 304, + 198, + 545, + 507 + ], + "type": "text", + "content": " [21], and CAR [39] transformed global coordinates into the canonical coordinates to allow for reposing. PHOHRUM [4] and S3F [12] further disentangled shading and albedo to enable relighting. Another line of work replaced the neural representations with conventional Poisson surface reconstruction [33, 34]. ECON [74] and 2K2K [18] trained normal and depth predictors to generate front and back 2.5D point clouds. The human mesh is obtained by fusing these point clouds with body priors and 3D heuristics. However, none of these methods produce realistic full-body texture and geometry in the unobserved regions. Our pipeline addresses this problem by incorporating a generative diffusion model into the 3D human reconstruction workflow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "3D generation with 2D diffusion models. Diffusion models [52, 56, 57, 59] trained with large collections of images have demonstrated unprecedented capability in creating 3D objects from text prompts. Most prior work [11, 19, 40, 47, 53, 68] followed an optimization workflow to update 3D representations (e.g. NeRF [48], SDF tetrahedron [63]) via neural rendering [66] and a score distillation sampling (SDS) [53] loss. While some methods [3, 10, 24, 27, 64, 76] applied this workflow to human bodies, they cannot produce accurate human bodies and appearances due to the ambiguity of text-conditioning. More recent work [41, 54] also tried to extend this workflow with more accurate image-conditioning. However, we show that they struggle to recover human clothing details and require a long optimization time. Most related to our work is Chupa [35], which also decomposes its pipeline into two stages. Note that Chupa is an optimization-based approach that relies on texts" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "539" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 541, + 245 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 541, + 245 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 541, + 245 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 541, + 245 + ], + "type": "image", + "image_path": "3ecbf5bd8e7652b382bdf6e6652f9c9a93828a0744e17fc309b49a8cbbb915d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 546, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 546, + 304 + ], + "type": "text", + "content": "Figure 2. Method overview. SiTH is a two-stage pipeline composed of back-view hallucination and mesh reconstruction. The back-view hallucination module samples perceptually consistent back-view images through an iterative denoising process conditioned on the input image, UV map, and silhouette mask (Sec. 3.1). Based on the input and generated back-view images, the mesh reconstruction module recovers a full-body mesh and textures leveraging a skinned body prior as guidance (Sec. 3.2). Note that both modules in the pipeline can be trained with the same public 3D human dataset and generalize unseen images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "text", + "content": "and cannot model colors. We address these issues by introducing an image-conditioning strategy and model. Most importantly, our method swiftly reconstructs full-texture human meshes without any optimization process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 288, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 583 + ], + "type": "text", + "content": "Diffusion models adaptation. Foundation models [8, 13, 20, 36] trained on large-scale datasets have been shown to be adaptable to various downstream tasks. Following this trend, pretrained diffusion models [52, 56, 57, 59] have become common backbones for generative modeling. For instance, they can be customized by finetuning with a small collection of images [23, 37, 58]. ControlNet [77] introduced additional trainable plugins to enable image conditioning such as body skeletons. While these strategies have been widely adopted, none of them directly fit our objective. More relevant to our task is DreamPose [31], which utilizes DensePose [16] images as conditions to repose input images. However, it cannot handle out-of-distribution images due to overfitting. Similarly, Zero-1-to-3 [41] finetunes a diffusion model with multi-view images to allow for viewpoint control. However, we show that viewpoint conditioning is not sufficient for generating consistent human bodies. Our model addresses this issue by providing accurate body pose and mask conditions for back-view hallucination." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 590, + 129, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 129, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 129, + 604 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Method overview. Given an input image of a human body and estimated SMPL-X [51] parameters, SiTH produces a full-body textured mesh. This mesh not only captures the observed appearances but also recovers geometric and textural details in unseen regions, such as clothing wrinkles on the back. The pipeline is composed of two modules and is summarized in Fig. 2. In the first stage, we hallucinate unobserved appearances leveraging the generative power of an image-conditioned diffusion model (Sec. 3.1). In the second" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 308, + 546, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 546, + 357 + ], + "type": "text", + "content": "stage, we reconstruct a full-body textured mesh given the input front-view image and the generated back-view image as guidance (Sec. 3.2). Notably, both modules are efficiently trained with 500 textured human scans in THuman2.0 [75]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 361, + 444, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 444, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 444, + 372 + ], + "type": "text", + "content": "3.1. Back-view Hallucination" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": "Preliminaries. Given an input front-view image " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "I^{F} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": ", our goal is to infer a back-view image " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "I^{B} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": " which depicts unobserved body appearances. This task is under-constrained since there are multiple possible solutions to the same input images. Taking this perspective into account, we leverage a latent diffusion model (LDM) [57] to learn a conditional distribution of back-view images given a front-view image. First, a VAE autoencoder, consisting of an encoder " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": ", is pretrained on a corpus of 2D natural images through image reconstruction, i.e., " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{I} = \\mathcal{D}(\\mathcal{E}(I))" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": ". Afterwards, an LDM learns to produce a latent code " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": " within the VAE latent distribution " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(I)" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": " from randomly sampled noise. To sample an image, a latent code " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": " is obtained by iteratively denoising Gaussian noise. The final image is reconstructed through the decoder, i.e., " + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{I} = \\mathcal{D}(\\tilde{z})" + }, + { + "bbox": [ + 304, + 374, + 546, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Image-conditioned diffusion model. Simply applying the LDM architecture to our task is not sufficient since our goal is to learn a conditional distribution of back-view images given an input conditional image. To this end, we make several adaptations to allow for image-conditioning as shown in Fig. 3. First, we utilize the pretrained CLIP [55] image encoder and VAE encoder " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " to extract image features from the front-view image (i.e., " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "I^{F}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "). These image features are used for conditioning the LDM, ensuring the output image shares a consistent appearance with the input image. Second, we follow the idea of ControlNet [77] and propose to use a UV map " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "(I_{UV}^{B} \\in \\mathbb{R}^{H \\times W \\times 3})" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and a silhou" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "540" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 286, + 232 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 286, + 232 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 286, + 232 + ], + "type": "image", + "image_path": "2067ee22552ec0b33f3bce63075eb345743003e883df4eadbb64da2f4bd71c9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": "Figure 3. Training of back-view hallucination module. We employ a pretrained LDM and ControlNet architecture to enable image conditioning. To train our model, we render training pairs of conditional images " + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "inline_equation", + "content": "I^{F}" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": " and ground-truth images " + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "inline_equation", + "content": "I^{B}" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": " from 3D human scans. Given a noisy image latent " + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": ", the model predicts added noise " + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": " given the conditional image " + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "inline_equation", + "content": "I^{F}" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": ", UV map " + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "inline_equation", + "content": "I_{UV}^{B}" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": ", and mask " + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "inline_equation", + "content": "I_{M}^{B}" + }, + { + "bbox": [ + 46, + 235, + 289, + 324 + ], + "type": "text", + "content": " as conditions. We train the ControlNet model and cross-attention layers while keeping other parameters frozen." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 329, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 378 + ], + "type": "text", + "content": "ette mask " + }, + { + "bbox": [ + 46, + 329, + 287, + 378 + ], + "type": "inline_equation", + "content": "(I_M^B\\in \\mathbb{R}^{H\\times W})" + }, + { + "bbox": [ + 46, + 329, + 287, + 378 + ], + "type": "text", + "content": " from the back view as additional conditions. These conditional signals provide additional information that ensures the output image has a similar body shape and pose to the conditional input image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 380, + 289, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 289, + 538 + ], + "type": "text", + "content": "Learning hallucination from pretraining. Another challenge in training an image-conditioned LDM is data. Training the model from scratch is infeasible due to the requirement of a large number of paired images rendered from 3D textured human scans. Inspired by the concept of learning from large-scale pretraining [13, 20], we build our image-conditioned LDM on top of a pretrained diffusion U-Net [57]. We utilize the finetuning strategy [37, 77] to optimize cross-attention layers and ControlNet parameters while keeping most of the other parameters frozen (see Fig. 3). The design and training strategy of our image-conditioned diffusion model enables hallucinating plausible back-view images that are cosistent with the frontal inputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Training and inference. To generate pairwise training images from 3D human scans, we sample camera view angles and use orthographic projection to render RGBA images from 3D scans and UV maps from their SMPL-X fits. Given a pair of images rendered by a frontal and its corresponding back camera, the first image serves as the conditional input " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "I^{F}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " while the other one is the ground-truth image " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "I^{B}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". During training, the ground-truth latent code " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{0} = \\mathcal{E}(I^{B})" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " is perturbed by the diffusion process in t time steps, resulting in a noisy latent " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". The image-conditioned LDM model " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " aims to predict the added noise " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " given the noisy latent " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", the time step " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "t \\sim [0,1000]" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", the conditional image " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "I^{F}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", the silhouette mask " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "I_{M}^{B}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ", and the UV map " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "I_{UV}^{B}" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " (See Fig. 3). The objective function for fine-tuning can be" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 332, + 74, + 518, + 241 + ], + "blocks": [ + { + "bbox": [ + 332, + 74, + 518, + 241 + ], + "lines": [ + { + "bbox": [ + 332, + 74, + 518, + 241 + ], + "spans": [ + { + "bbox": [ + 332, + 74, + 518, + 241 + ], + "type": "image", + "image_path": "8b77da9a94f91b4745165cdb3f3c431a3d3a5c8fb7de9ee1ee9816e5081cf82e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": "Figure 4. Mesh reconstruction module. Given front and back-view images " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "(I^{F}, I^{B})" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": " we predict their normal images " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "(N^{F}, N^{B})" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": " through a learned normal predictor. A 3D point " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": " is projected onto these images for querying pixel-aligned features " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x})" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": ". To leverage human body mesh as guidance, we embed the point " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": " into the local UV coordinates " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_c" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": ", vector " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{n_c}" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": ", distance " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "d_c" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": ", and visibility " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "v_c" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": ". Finally, two decoders " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "(H_d, H_r)" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": " predict SDF and RGB values at " + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 242, + 547, + 331 + ], + "type": "text", + "content": " given the positional embedding and pixel-aligned features." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 338, + 369, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 369, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 369, + 350 + ], + "type": "text", + "content": "represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 360, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 545, + 389 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {z \\sim \\mathcal {E} (I), t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\\right) \\| _ {2} ^ {2}. \\tag {1}", + "image_path": "f019279285e69778a7f75b9e5ee8a43bf13466715237f70b7f0693b3b9ac376e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "text", + "content": "At test time, we obtain " + }, + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "inline_equation", + "content": "I_{UV}^{B}, I_{M}^{B}" + }, + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "text", + "content": " from an off-the-shelf pose predictor [9] and segmentation model [36]. To infer a back-view image, we sample a latent " + }, + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "inline_equation", + "content": "\\tilde{z}_0" + }, + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "text", + "content": " by performing the iterative denoising process starting from a Gaussian noise " + }, + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "inline_equation", + "content": "z_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 389, + 547, + 452 + ], + "type": "text", + "content": ". The back-view image can be obtained by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 460, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 460, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 341, + 460, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\tilde {I} ^ {B} = \\mathcal {D} (\\tilde {z} _ {0}) = \\mathcal {D} \\left(f _ {\\theta} \\left(z _ {T}, I ^ {F}, I _ {U V} ^ {B}, I _ {M} ^ {B}\\right)\\right), \\tag {2}", + "image_path": "f60d50176108eeb9228689d0a30d439362b3320e033bf93d58ead1fde24bce9c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 484, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 484, + 545, + 510 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 484, + 545, + 510 + ], + "type": "text", + "content": " is a function representing the iterative denoising process of our image-conditioned LDM (See Fig. 2 left)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 518, + 467, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 467, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 467, + 529 + ], + "type": "text", + "content": "3.2. Human Mesh Reconstruction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "content": "After obtaining the back-view image, our goal is to construct a full-body human mesh and its textures using the input and back-view image as guidance. We follow the literature [60, 61] to model this task with a data-driven method. Given pairwise training data (i.e., front/back-view images and 3D scans), we learn a data-driven model that maps these images to a 3D representation (e.g., a signed distance field (SDF)). We define this mapping as below:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 637, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 637, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 334, + 637, + 545, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Phi : \\mathbb {R} ^ {H \\times W \\times 3} \\times \\mathbb {R} ^ {H \\times W \\times 3} \\times \\mathbb {R} ^ {3} \\rightarrow \\mathbb {R} \\times \\mathbb {R} ^ {3} \\tag {3} \\\\ (I ^ {F}, I ^ {B}, \\mathbf {x}) \\mapsto d _ {x}, \\mathbf {r} _ {x}, \\\\ \\end{array}", + "image_path": "ae84054dc3d0c5e03305adb05cff06a329e6939fece41f745fda60b9cbd2cdaf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "spans": [ + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "text", + "content": " is the 3D coordinate of a query point, and " + }, + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "inline_equation", + "content": "d_x, \\mathbf{r}_x" + }, + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "text", + "content": " denote the signed distance and RGB color value at point " + }, + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 678, + 547, + 727 + ], + "type": "text", + "content": ". The network components we used for learning the mapping function are depicted in Fig. 4." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "541" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 79, + 100, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 79, + 100, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 79, + 100, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 79, + 100, + 178 + ], + "type": "image", + "image_path": "1b1ebc1e3bb2f1b8e8dbe76773adfb901a8c15154e5c7b9c351d313e6e735471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 180, + 100, + 192 + ], + "lines": [ + { + "bbox": [ + 54, + 180, + 100, + 192 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 100, + 192 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 81, + 218, + 178 + ], + "blocks": [ + { + "bbox": [ + 108, + 81, + 218, + 178 + ], + "lines": [ + { + "bbox": [ + 108, + 81, + 218, + 178 + ], + "spans": [ + { + "bbox": [ + 108, + 81, + 218, + 178 + ], + "type": "image", + "image_path": "51b68c462af8841dc78e08284c8c6bf4b6a0c2cacab6f92fbd9dae42cac2c4f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 180, + 181, + 191 + ], + "lines": [ + { + "bbox": [ + 153, + 180, + 181, + 191 + ], + "spans": [ + { + "bbox": [ + 153, + 180, + 181, + 191 + ], + "type": "text", + "content": "PaMIR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 81, + 328, + 178 + ], + "blocks": [ + { + "bbox": [ + 220, + 81, + 328, + 178 + ], + "lines": [ + { + "bbox": [ + 220, + 81, + 328, + 178 + ], + "spans": [ + { + "bbox": [ + 220, + 81, + 328, + 178 + ], + "type": "image", + "image_path": "9f2f38dfba8c3297aab94ddfe51f92f9f62ae4bc3026f994d4c0ba0343781720.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 180, + 296, + 191 + ], + "lines": [ + { + "bbox": [ + 249, + 180, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 249, + 180, + 296, + 191 + ], + "type": "text", + "content": "PHORHUM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 330, + 81, + 433, + 178 + ], + "blocks": [ + { + "bbox": [ + 330, + 81, + 433, + 178 + ], + "lines": [ + { + "bbox": [ + 330, + 81, + 433, + 178 + ], + "spans": [ + { + "bbox": [ + 330, + 81, + 433, + 178 + ], + "type": "image", + "image_path": "b2fc4efbbc189511464cfe449720b5b54869f3fcbc7a7e4c15961bfdedbe1f76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 180, + 405, + 191 + ], + "lines": [ + { + "bbox": [ + 358, + 180, + 405, + 191 + ], + "spans": [ + { + "bbox": [ + 358, + 180, + 405, + 191 + ], + "type": "text", + "content": "SiTH (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 435, + 81, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 435, + 81, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 435, + 81, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 435, + 81, + 541, + 178 + ], + "type": "image", + "image_path": "213325818320f1edd3af26deb3df9d5751bec21f3b7d9c7146f1451dce178187.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 180, + 515, + 191 + ], + "lines": [ + { + "bbox": [ + 460, + 180, + 515, + 191 + ], + "spans": [ + { + "bbox": [ + 460, + 180, + 515, + 191 + ], + "type": "text", + "content": "GT References" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 196, + 108, + 286 + ], + "blocks": [ + { + "bbox": [ + 50, + 196, + 108, + 286 + ], + "lines": [ + { + "bbox": [ + 50, + 196, + 108, + 286 + ], + "spans": [ + { + "bbox": [ + 50, + 196, + 108, + 286 + ], + "type": "image", + "image_path": "1ca033b379b2315af5852f306aa9940595d3877c11d403404dc131f21aa8f3dd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 287, + 100, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 100, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 100, + 298 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 108, + 196, + 225, + 286 + ], + "blocks": [ + { + "bbox": [ + 108, + 196, + 225, + 286 + ], + "lines": [ + { + "bbox": [ + 108, + 196, + 225, + 286 + ], + "spans": [ + { + "bbox": [ + 108, + 196, + 225, + 286 + ], + "type": "image", + "image_path": "a1def554032d40cb086d01cb1b9cc9f496729d20005c53fc5c4401aa9dddcaa6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 287, + 181, + 297 + ], + "lines": [ + { + "bbox": [ + 154, + 287, + 181, + 297 + ], + "spans": [ + { + "bbox": [ + 154, + 287, + 181, + 297 + ], + "type": "text", + "content": "ECON" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 225, + 196, + 334, + 286 + ], + "blocks": [ + { + "bbox": [ + 225, + 196, + 334, + 286 + ], + "lines": [ + { + "bbox": [ + 225, + 196, + 334, + 286 + ], + "spans": [ + { + "bbox": [ + 225, + 196, + 334, + 286 + ], + "type": "image", + "image_path": "b02915f2e73fcd1206dafbf49afcc646577574ab699f7e4b6b2b9f81c281d78f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 287, + 288, + 297 + ], + "lines": [ + { + "bbox": [ + 256, + 287, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 256, + 287, + 288, + 297 + ], + "type": "text", + "content": "PIFuHD" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 334, + 196, + 441, + 286 + ], + "blocks": [ + { + "bbox": [ + 334, + 196, + 441, + 286 + ], + "lines": [ + { + "bbox": [ + 334, + 196, + 441, + 286 + ], + "spans": [ + { + "bbox": [ + 334, + 196, + 441, + 286 + ], + "type": "image", + "image_path": "281ae0046b6f7412000e4eadbe5585bbd456a1f54bba6fccfa3d63159ed6cdeb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 287, + 405, + 297 + ], + "lines": [ + { + "bbox": [ + 359, + 287, + 405, + 297 + ], + "spans": [ + { + "bbox": [ + 359, + 287, + 405, + 297 + ], + "type": "text", + "content": "SiTH (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on CustomHumans. Top: Results of methods generating mesh and texture. Bottom: Results of methods generating mesh only. Note that single-view reconstruction is not possible to replicate exact back-view texture and geometry. Our method generates realistic texture and clothing wrinkles perceptually close to the real scans while other baselines only produce smooth colors and surfaces in the back regions. Best viewed in color and zoom in." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 438, + 196, + 545, + 286 + ], + "blocks": [ + { + "bbox": [ + 438, + 196, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 438, + 196, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 438, + 196, + 545, + 286 + ], + "type": "image", + "image_path": "464fc5cbf953c8d367ba6074c1d569e7bdfc086a9653a29b800bdcda1a5a1f76.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 287, + 515, + 297 + ], + "lines": [ + { + "bbox": [ + 460, + 287, + 515, + 297 + ], + "spans": [ + { + "bbox": [ + 460, + 287, + 515, + 297 + ], + "type": "text", + "content": "GT References" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": "Local feature querying. To learn a generic mapping function that is robust to unseen images, it is important that the model is conditioned solely on local image information with respect to the position of " + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": ". Therefore, we employ the idea of pixel-aligned feature querying [60, 61] and separate our model into two branches, i.e., color and geometry. Our model contains a normal predictor that converts the RGB image pair " + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "inline_equation", + "content": "(I^{F}, I^{B})" + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": " into normal maps " + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "inline_equation", + "content": "(N^{F}, N^{B})" + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": ". Two image feature encoders " + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "inline_equation", + "content": "G_{d}, G_{r}" + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": " then extract color and geometry feature maps " + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{d}, \\mathbf{f}_{r}) \\in \\mathbb{R}^{H' \\times W' \\times D}" + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": " from the images and normal maps respectively (for simplicity we describe the process for a single image and leave out the superscripts, but both front and back images are treated the same). Finally, we project the query point " + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": " onto the image coordinate (Fig. 4 red points) to retrieve the local features " + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x}) \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 46, + 354, + 289, + 548 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 85, + 554, + 247, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 554, + 247, + 582 + ], + "spans": [ + { + "bbox": [ + 85, + 554, + 247, + 582 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} _ {d, x} = \\mathcal {B} (\\mathbf {f} _ {d}, \\pi (\\mathbf {x})) = \\mathcal {B} (G _ {d} (N), \\pi (\\mathbf {x})), \\\\ \\mathbf {f} _ {r, x} = \\mathcal {B} (\\mathbf {f} _ {r}, \\pi (\\mathbf {x})) = \\mathcal {B} (G _ {r} (I), \\pi (\\mathbf {x})), \\\\ \\end{array}", + "image_path": "394b73bb874f59ba63dfbdf926f8deb21397fdb02c409b05e6c805f09cfd17bd.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "content": " is a local feature querying operation using bilinear interpolation and " + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "content": " denotes orthographic projection." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 618, + 281, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 281, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 281, + 630 + ], + "type": "text", + "content": "Local positional embedding with skinned body prior." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "As mentioned in Sec. 1, a major difficulty in mesh reconstruction is 3D ambiguity where a model has to infer unknown depth information between the front and back images. To address this issue, we follow prior work [22, 73, 79] leveraging a skinned body mesh [51] for guiding the reconstruction task. This body mesh is regarded as an anchor that provides an approximate 3D shape of the human body." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 354, + 546, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 403 + ], + "type": "text", + "content": "To exploit this body prior, we devise a local positional embedding function that transforms the query point " + }, + { + "bbox": [ + 304, + 354, + 546, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 354, + 546, + 403 + ], + "type": "text", + "content": " into the local body mesh coordinate system. We look for the closest point " + }, + { + "bbox": [ + 304, + 354, + 546, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c^*" + }, + { + "bbox": [ + 304, + 354, + 546, + 403 + ], + "type": "text", + "content": " on the body mesh (Fig. 4 blue point), i.e.," + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 372, + 413, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 413, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 372, + 413, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {c} ^ {*} = \\arg \\min _ {\\mathbf {x} _ {c}} \\| \\mathbf {x} - \\mathbf {x} _ {c} \\| _ {2}, \\tag {5}", + "image_path": "e7416853593667735be6f0613a3ee5afc9987f39e4469c832345778b71573430.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " are points on the skinned body mesh " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": ". Our positional embedding " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " constitutes four elements: a signed distance value " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "d_c" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c^*" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": ", a vector " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_c = (\\mathbf{x} - \\mathbf{x}_c^*)" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": ", the UV coordinates " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_c \\in [0,1]^2" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " of the point " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c^*" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": ", and a visibility label " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "v_c \\in \\{1, -1, 0\\}" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " that indicates whether " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c^*" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " is visible in the front/back image or neither. Finally, two separate MLPs " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "H_d, H_r" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " take the positional embedding " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [d_c, \\mathbf{n}_c, \\mathbf{u}_c, v_c]" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " and the local texture/geometry features " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{d,x}, \\mathbf{f}_{r,x})" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": " as inputs to predict the final SDF and RGB values at point " + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 440, + 546, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 377, + 569, + 473, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 569, + 473, + 601 + ], + "spans": [ + { + "bbox": [ + 377, + 569, + 473, + 601 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d _ {x} = H _ {d} \\left(\\mathbf {f} _ {d, x} ^ {F}, \\mathbf {f} _ {d, x} ^ {B}, \\mathbf {p}\\right), \\\\ \\mathbf {r} _ {x} = H _ {r} \\left(\\mathbf {f} _ {r, x} ^ {F}, \\mathbf {f} _ {r, x} ^ {B}, \\mathbf {p}\\right). \\\\ \\end{array}", + "image_path": "038e60cbed06eb8877f5e14e1787b1b99dda394c2997ac8bf2009b1b009e8bfa.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Training and inference. We used the same 3D dataset described in Sec. 3.1 to render training image pairs " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "(I^{F}, I^{B})" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " from the 3D textured scans. For each training scan, query points " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " are sampled within a 3-dimensional cube " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "[-1, 1]^{3}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ". For each point, we compute the ground-truth signed distance values " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " to the scan surface, closest texture RGB values " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ", and surface normal " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ". Finally, we jointly optimized the normal predictors, the image encoders, and the MLPs in" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "542" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 521, + 215 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 521, + 215 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 521, + 215 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 521, + 215 + ], + "type": "table", + "html": "
CAPE [45]CustomHuman [22]
MethodCD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F (×10-2) ↓CD: P-to-S / S-to-P (cm)↓NC↑f-Score↑LPIPS: F / B (×10-2) ↓
PIFu [60]2.368 / 3.7630.77833.8422.7202.209 / 2.5820.80534.8816.073 / 8.496
PIFuHD [61]2.401 / 3.5220.77235.706-2.107 / 2.2280.80439.076-
PaMIR [79]2.190 / 2.8060.80436.7252.0852.181 / 2.5070.81335.8474.646 / 7.152
2K2K [18]2.478 / 3.6830.78228.700-2.488 / 3.2920.79630.186-
FOF [15]2.196 / 4.0400.77734.227-2.079 / 2.6440.80836.013-
ICON [73]2.516 / 3.0790.78629.630-2.256 / 2.7950.79130.437-
ECON [74]2.475 / 2.9700.78830.488-2.483 / 2.6800.79730.894-
SiTH (Ours)1.899 / 2.2610.81637.7631.9771.871 / 2.0450.82637.0293.929 / 6.803
", + "image_path": "d9048fb13619c9b6317dbce4311b17106ab9e02e94f15d318739b4868935d226.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 222, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 545, + 269 + ], + "type": "text", + "content": "Table 1. Single-view human reconstruction benchmarks. We report Chamfer distance (CD), normal consistency (NC), and f-score between ground truth and predicted meshes. To evaluate texture reconstruction quality, we compute LPIPS between the image rendering of GT and generated textures. The best and the second best methods are highlighted in bold and underlined respectively. Note that gray color denotes models trained on more commercial 3D human scans while the others are trained on the public THuman2.0 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 277, + 271, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 271, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 271, + 289 + ], + "type": "text", + "content": "both branches with the following reconstruction losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 295, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 295, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 87, + 295, + 287, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\left\\| d - d _ {x} \\right\\| _ {1} + \\lambda_ {n} \\left(1 - \\mathbf {n} \\cdot \\nabla_ {\\mathbf {x}} d _ {x}\\right), \\tag {7}", + "image_path": "e4c7fa2a60cbee48dc270be527d80a8dbbed588885728622bf45359aae070994.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 315, + 287, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 315, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 287, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\left\\| \\mathbf {r} - \\mathbf {r} _ {x} \\right\\| _ {1}. \\tag {8}", + "image_path": "31aaac586af52d8a74b623e9c44830f7ba70d46ce7e65af1e7f9b218c97856b2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "text", + "content": " indicates numerical finite differences for computing local normals at point " + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\lambda_{n}" + }, + { + "bbox": [ + 46, + 331, + 287, + 354 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "text", + "content": "During inference, we use the input image " + }, + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "inline_equation", + "content": "I^{F}" + }, + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "text", + "content": " and the back-view image " + }, + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{I}^{B}" + }, + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "text", + "content": " obtained from Sec. 3.1 to reconstruct 3D mesh and textures. First, we align both images with the estimated body mesh " + }, + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 355, + 287, + 475 + ], + "type": "text", + "content": " to ensure that image features can be properly queried around the 3D anchor. We adopt a similar strategy of SMPLify [7] to optimize the scale and the offset of the body mesh with silhouette and 2D joint errors. Finally, we perform the marching cube algorithm [43] by querying SDF and RGB values within a dense voxel grid via Eq. (3) (see Fig. 2 right)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 497, + 163, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 163, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 163, + 509 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 676 + ], + "type": "text", + "content": "Dataset. Previous work relied on training data from commercial datasets such as RenderPeople [1]. While these datasets offer high-quality textured meshes, they also limit reproducibility due to limited accessibility. For fair comparisons, we follow ICON [73] by training our method on the public 3D dataset THuman2.0 [75] and using the CAPE [45] dataset for evaluation. However, we observed potential biases in the evaluation due to the low-res groundtruth meshes and image rendering defects in the CAPE dataset (for a detailed discussion, please refer to SuppSec. 6). Consequently, we further create a new benchmark that evaluates the baselines on a higher-quality 3D human dataset CustomHumans [22]. In the following, we provide a summary of the datasets used in our experiments:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "- THuman2.0 [75] contains approximately 500 scans of humans wearing 150 different garments in various poses. We use these 3D scans as the training data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 277, + 545, + 360 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 277, + 544, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 544, + 312 + ], + "type": "text", + "content": "- CAPE [45] contains 15 subjects in 8 types of tight outfits. The test set, provided by ICON, consists of 100 meshes. We use CAPE for the quantitative evaluation (Sec. 4.2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 313, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 313, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 313, + 545, + 360 + ], + "type": "text", + "content": "- CustomHumans [22] contains 600 higher-quality scans of 80 subjects in 120 different garments and varied poses. We selected 60 subjects for all quantitative experiments, user studies, and ablation studies. (Sec. 4.2 - Sec. 4.4)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 360, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 456 + ], + "type": "text", + "content": "Evaluation protocol. We follow the evaluation protocol in OccNet [46] and ICON [73] to compute 3D metrics Chamfer distance (CD), normal consistency (NC), and f-Score [65] on the generated meshes. To evaluate reconstructed mesh texture, we report LPIPS [78] of front and back texture rendering. In user studies, 30 participants rank the meshes obtained by four different methods. We report the average ranking ranging from 1 (best) to 4 (worst)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 460, + 494, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 494, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 494, + 472 + ], + "type": "text", + "content": "4.2. Single-view Human Reconstruction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "type": "text", + "content": "Benchmark evaluation. We compared SiTH with state-of-the-art single-view human reconstruction methods, including PIFu [60], PIFuHD [61], PaMIR [79], FOF [15], ICON [73], PHORHUM [4], 2K2K [18], and ECON [74] on CAPE and CustomHumans. Note that PHORHUM is only used for qualitative comparison since a different camera system is used, leading to the misalignment with ground-truth meshes. We visualize the generated mesh texture and normals in Fig. 5. Existing methods produce oversmoothed texture and normals, particularly in the back. Our method not only generates photorealistic and perceptually consistent appearances in unobserved regions but also recovers underlying geometric details like clothing wrinkles." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "The quantitative results are summarized in Tab. 1. It's worth noting that most methods are trained with commercial datasets (gray color in Tab. 1), while the others are trained on the public THuman2.0 dataset. To evaluate the methods leveraging a skinned body prior (i.e., PaMIR, ICON, ECON, FOF, and SiTH), we use the same pose alignment procedure in their original implementations for a fair" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "543" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 102, + 159 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 102, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 102, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 102, + 159 + ], + "type": "image", + "image_path": "e4f28041b59e1a8da72898ae9c70f8ee695b29116f6c8dadd633686dea730304.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 164, + 99, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 164, + 99, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 99, + 175 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 105, + 70, + 159, + 160 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 159, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 159, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 159, + 160 + ], + "type": "image", + "image_path": "ee2961317ec3147f70fb3982c5f1270386614287fd779d481d5eeb72b4073aca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 164, + 151, + 174 + ], + "lines": [ + { + "bbox": [ + 111, + 164, + 151, + 174 + ], + "spans": [ + { + "bbox": [ + 111, + 164, + 151, + 174 + ], + "type": "text", + "content": "Pix2PixHD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 162, + 70, + 225, + 160 + ], + "blocks": [ + { + "bbox": [ + 162, + 70, + 225, + 160 + ], + "lines": [ + { + "bbox": [ + 162, + 70, + 225, + 160 + ], + "spans": [ + { + "bbox": [ + 162, + 70, + 225, + 160 + ], + "type": "image", + "image_path": "7cb24378c11b722c53a0558be43d6a0283e39592288fb718f98c06d959e0e775.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 164, + 219, + 174 + ], + "lines": [ + { + "bbox": [ + 179, + 164, + 219, + 174 + ], + "spans": [ + { + "bbox": [ + 179, + 164, + 219, + 174 + ], + "type": "text", + "content": "DreamPose" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 226, + 72, + 240, + 159 + ], + "blocks": [ + { + "bbox": [ + 226, + 72, + 240, + 159 + ], + "lines": [ + { + "bbox": [ + 226, + 72, + 240, + 159 + ], + "spans": [ + { + "bbox": [ + 226, + 72, + 240, + 159 + ], + "type": "image", + "image_path": "54f3dbe7a06ed222c79335b776006682ad7a9bf66a1802f85b3bcf7964e9c566.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 164, + 325, + 175 + ], + "lines": [ + { + "bbox": [ + 239, + 164, + 325, + 175 + ], + "spans": [ + { + "bbox": [ + 239, + 164, + 325, + 175 + ], + "type": "text", + "content": "ControlNet + Interrogate" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 242, + 72, + 293, + 159 + ], + "blocks": [ + { + "bbox": [ + 242, + 72, + 293, + 159 + ], + "lines": [ + { + "bbox": [ + 242, + 72, + 293, + 159 + ], + "spans": [ + { + "bbox": [ + 242, + 72, + 293, + 159 + ], + "type": "image", + "image_path": "06031aa23c7972a25329344d79444a8a64042082486556994610c74404c97bcb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 294, + 72, + 321, + 159 + ], + "blocks": [ + { + "bbox": [ + 294, + 72, + 321, + 159 + ], + "lines": [ + { + "bbox": [ + 294, + 72, + 321, + 159 + ], + "spans": [ + { + "bbox": [ + 294, + 72, + 321, + 159 + ], + "type": "image", + "image_path": "f5d0c263d25c30f6af5b9c636f40a4a0b19cb63a5289e17fef21a1603c211985.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 72, + 339, + 159 + ], + "blocks": [ + { + "bbox": [ + 324, + 72, + 339, + 159 + ], + "lines": [ + { + "bbox": [ + 324, + 72, + 339, + 159 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 339, + 159 + ], + "type": "image", + "image_path": "1c9572de8d4770389f1a6d658f3cf55826d8b1c88c5f780058571c6382233a7b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 164, + 385, + 174 + ], + "lines": [ + { + "bbox": [ + 343, + 164, + 385, + 174 + ], + "spans": [ + { + "bbox": [ + 343, + 164, + 385, + 174 + ], + "type": "text", + "content": "Zero-1-to-3" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 341, + 72, + 404, + 159 + ], + "blocks": [ + { + "bbox": [ + 341, + 72, + 404, + 159 + ], + "lines": [ + { + "bbox": [ + 341, + 72, + 404, + 159 + ], + "spans": [ + { + "bbox": [ + 341, + 72, + 404, + 159 + ], + "type": "image", + "image_path": "20124d72a1f43ff6b18ee3bab5c9ca1e2f3aa22a4709edc97c23f49d6c1c9282.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 406, + 72, + 457, + 159 + ], + "blocks": [ + { + "bbox": [ + 406, + 72, + 457, + 159 + ], + "lines": [ + { + "bbox": [ + 406, + 72, + 457, + 159 + ], + "spans": [ + { + "bbox": [ + 406, + 72, + 457, + 159 + ], + "type": "image", + "image_path": "05d6c49f43214c6ad6b794f9972be85e9e779abe61042e9acc7787f233b6d86b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 164, + 468, + 174 + ], + "lines": [ + { + "bbox": [ + 424, + 164, + 468, + 174 + ], + "spans": [ + { + "bbox": [ + 424, + 164, + 468, + 174 + ], + "type": "text", + "content": "SiTH (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 459, + 72, + 486, + 159 + ], + "blocks": [ + { + "bbox": [ + 459, + 72, + 486, + 159 + ], + "lines": [ + { + "bbox": [ + 459, + 72, + 486, + 159 + ], + "spans": [ + { + "bbox": [ + 459, + 72, + 486, + 159 + ], + "type": "image", + "image_path": "3afae08893a53dc535c0287a13358930cb20b76981048e8d882565cd52ff36e6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 488, + 72, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 488, + 72, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 488, + 72, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 488, + 72, + 541, + 159 + ], + "type": "image", + "image_path": "6dc0dcbdddfadc7f7af2c21016afc2ff1dede7f97a91c1c404b092e89b117b5b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 164, + 539, + 174 + ], + "lines": [ + { + "bbox": [ + 488, + 164, + 539, + 174 + ], + "spans": [ + { + "bbox": [ + 488, + 164, + 539, + 174 + ], + "type": "text", + "content": "GT References" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 58, + 220, + 279, + 378 + ], + "blocks": [ + { + "bbox": [ + 46, + 181, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 181, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 547, + 215 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of back-view hallucination. We visualize back-view images generated by the baseline methods. Note that the three different images are sampled from different random seeds. Our results are perceptually close to the ground-truth image in terms of appearances and poses. Moreover, our method also preserves generative stochasticity for handling tiny wrinkle changes." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 220, + 279, + 378 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 279, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 279, + 378 + ], + "type": "image", + "image_path": "84327b293fce65fc3f6028d4454a36766aed3ccd017ee8b542144837ac58bd3f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 381, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 437 + ], + "type": "text", + "content": "Figure 7. Comparison with optimization-based methods. Compared to methods that utilize diffusion models for optimization, our result is closer to the input image and contains local geometric details. Note that Chupa is not conditioned on the image, and fails to generate correct clothing from the text prompts." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 552 + ], + "type": "text", + "content": "comparison. Results in Tab. 1 show that the method using a body prior (PaMIR) outperformed the end-to-end method (PIFuHD) on tight clothing and challenging poses in CAPE. However, it falls short in handling diverse outfits in CustomHumans. Moreover, the methods trained on commercial datasets achieve better performance than those trained with public data (ICON, ECON). Notably, our method is robust across both benchmarks, achieving performance comparable to the methods trained on high-quality commercial data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Compared with optimization-based methods. We compared SiTH with methods that use pretrained diffusion models and a score distillation sampling loss [53] to optimize 3D meshes. In the case of Zero-1-to-3 [41], we used the input image to optimize an instant NGP [49] radiance field, and for Magic-123 [54], we provided additional text prompts to optimize an SDF tetrahedron [63]. From Fig. 7, we see that while both methods can handle full-body textures, they struggle with reasoning the underlying geometry and clothing details. It is worth noting that Zero-1-to-3 and Magic-123 require 10 minutes and 6 hours in optimization, respectively, while our method takes under 2 minutes to generate a textured mesh with a marching cube of " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "512^{3}" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 309, + 218, + 541, + 367 + ], + "blocks": [ + { + "bbox": [ + 309, + 218, + 541, + 367 + ], + "lines": [ + { + "bbox": [ + 309, + 218, + 541, + 367 + ], + "spans": [ + { + "bbox": [ + 309, + 218, + 541, + 367 + ], + "type": "table", + "html": "
ICONECONPIFuHDOurs
Front Geometry3.1272.7202.6071.547
Back Geometry3.1932.5133.0931.200
Similarity3.0932.6602.7801.456
PIFuPaMIRPHOHRUMOurs
Front Texture3.0672.1533.4501.327
Back Texture3.4502.3553.1401.054
Similarity3.3072.1923.4161.093
ChupaOurs
User Preference36.0%64.0%
", + "image_path": "0a7f501f2e6457aca7a2553824bca44924c84897ddeabba2cc518405cdd2b42e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 371, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 427 + ], + "type": "text", + "content": "Table 2. User study results. Top: 30 users are asked to rank the quality of surface normal images from best (1) to worst (4). We report the average ranking of each method. Middle: Similar to the first task, users are asked to rank the quality of RGB textures. Bottom: We ask users to choose the mesh with a better quality." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 432, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 432, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 545, + 576 + ], + "type": "text", + "content": "Additionally, more similar to our method is Chupa [35], which generates front/back-view normals for mesh reconstruction. Note that Chupa is not conditioned on images and does not generate texture. Instead, we provided body poses and text prompts generated by an image-to-text interrogator [38] as their conditional inputs. From Fig. 7, it's clear that text-conditioning is less accurate than image-conditioning, and the method struggles to generate unseen clothing styles such as coats. By contrast, our method can reconstruct correct clothing geometry and texture from unseen images. We present more discussions and comparisons with optimization-based methods in Supp-Sec. 9.1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "User study. The above metrics may not fully capture the quality of 3D meshes in terms of realism and local details. To address this, we conducted a user study to compare the texture and geometry quality among various baselines. We invited 30 users to rank the front/back-view texture and normal renderings of 3D meshes generated by four different methods. Additionally, we asked the users to assess the similarity between the input images and the generated meshes. The results (Tab. 2) support our claim that existing methods struggle to efficiently generate desirable back-view texture/geometry from single-view images. Our method," + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "544" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 276, + 155 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 276, + 155 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 276, + 155 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 276, + 155 + ], + "type": "table", + "html": "
MethodCD (cm)↓NC↑f-Score↑
W/o Body Mesh2.4710.80133.244
W/o Hallucination1.9600.84036.677
Full Pipeline1.9580.82637.029
W/ GT Body Mesh1.1720.89158.858
W/ GT Body and I^B1.0590.91463.356
", + "image_path": "7b3ca412680ab6266f9b762cc192bfb13f4b747b730eaeb799b5af966f987d1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 163, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 287, + 196 + ], + "type": "text", + "content": "Table 3. Ablation study on CustomHumans. We ablate the hallucination module and the skinned body mesh in our pipeline. Please refer to our discussion in Sec. 4.4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 199, + 288, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 288, + 272 + ], + "type": "text", + "content": "which leverages the generative capability of diffusion models, consistently outperforms each baseline. It also produces more preferred front-view textures and geometries, as evidenced by higher user rankings. We also conducted a user study with Chupa (in Tab. 2 bottom) which also indicates more users prefer the 3D meshes generated by our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 275, + 174, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 174, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 174, + 288 + ], + "type": "text", + "content": "4.3. Generative Capability" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 289, + 289, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 577 + ], + "type": "text", + "content": "Image quality comparison. Our hallucination module is a unique and essential component that generates spatially aligned human images to guide 3D mesh reconstruction. Given that our focus is on back-view hallucination, we compare the quality of generated images with the relevant generative methods in Fig. 6. We trained a baseline Pix2PixHD [69] model, which produced smooth and blurry results on unseen images due to overfitting to 500 subjects. Another method closely related to ours is DreamPose [31], which conditions the model with DensePose images and finetunes the diffusion model with paired data. However, their model failed to handle unseen images, in contrast to our approach. While Zero-1-to-3 [41] can generalize to unseen images, their method faces challenges in generating consistent body poses given the same back-view camera. Moreover, we designed another baseline that provides ControlNet [77] for corresponding text prompts using an image-to-text interrogator [38]. However, without proper image conditioning and fine-tuning, such a method cannot generate images that faithfully match the input appearances. Our method not only addresses these issues but also handles stochastic appearances (e.g., tiny differences in wrinkles) from different random seeds. We report 2D generative evaluation metrics and more results in Supp-Sec. 8.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 580, + 141, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 141, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 141, + 593 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "We conducted controlled experiments to validate the effectiveness of our proposed modules. As shown in Fig. 8, the skinned body mesh is a crucial component for 3D human reconstruction. Without this body mesh as a 3D anchor, the output mesh contains an incorrect body shape due to the depth ambiguity issue. Conversely, removing the hallucination module has minimal impact on 3D reconstruction metrics, though it slightly degrades normal consistency. However, the overall quality in both texture and geometry is incomparable with our full model (see Fig. 8 right). This is" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 547, + 150 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 547, + 150 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 547, + 150 + ], + "type": "image", + "image_path": "c192e9142d0984e52a76ce26c54d1a8619edce1a63c6b0c8574f8f72e1274b69.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 190, + 545, + 275 + ], + "blocks": [ + { + "bbox": [ + 304, + 155, + 547, + 189 + ], + "lines": [ + { + "bbox": [ + 304, + 155, + 547, + 189 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 547, + 189 + ], + "type": "text", + "content": "Figure 8. Ablation study. We visualize back and side-view rendering of the reconstructed meshes. Our full model produced a correct body shape and more realistic clothing geometry." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 190, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 309, + 190, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 309, + 190, + 545, + 275 + ], + "type": "image", + "image_path": "181415364ee9a723664cc01615b1cc12cd1c0cb6851e80eea47f528b4ed42532.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 278, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 300 + ], + "type": "text", + "content": "Figure 9. Applications. Our pipeline can be incorporated with text-to-image diffusion models for 3D human creation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 435 + ], + "type": "text", + "content": "consistent with our findings in user studies, indicating that 3D metrics may not accurately reflect the perceptual quality of 3D meshes. Finally, we tested two additional variants, leveraging ground-truth body meshes and real back-view images in our full pipeline, representing the upper bound of our method. As shown in Tab. 3 bottom, this additional information notably improves the 3D metrics. These results highlight the persistent challenges in the single-view reconstruction problem, including pose ambiguity and the stochastic nature of clothing geometry. For more experiments on our design choices, please refer to Supp-Sec. 8.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 388, + 453 + ], + "type": "text", + "content": "4.5. Applications" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "content": "Inheriting the generative capability of LDM, SiTH is robust to diverse inputs, such as out-of-distribution or AI-generated images. We demonstrate a unique solution to link photo-realistic AI photos and high-fidelity 3D humans. In Fig. 9, we introduce a 3D creation workflow integrating powerful text-to-image generative models. Given a body pose, we generate a front-view image using Stable Diffusion and ControlNet using text prompts. SiTH then creates a full-body textured human from the AI-generated image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 566, + 379, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 566, + 379, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 566, + 379, + 578 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": "We propose an innovative pipeline designed to create fully textured 3D humans from single-view images. Our approach seamlessly integrates an image-conditioned diffusion model into the existing data-driven 3D reconstruction workflow. Leveraging the generative capabilities of the diffusion model, our method efficiently produces lifelike 3D humans from a diverse range of unseen images in under two minutes. We expect our work will advance the application of generative AI in 3D human creation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was partially supported by the Swiss SERI Consolidation Grant \"AI-PERCEIVE\"." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "545" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 269, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 269, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 269, + 102 + ], + "type": "text", + "content": "[1] Renderpeople, https://renderpeople.com/.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 286, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 286, + 134 + ], + "type": "text", + "content": "[2] Stable diffusion image variations. huggingface . co / lambdaabs / stable - diffusion - image -conditioned.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 178 + ], + "type": "text", + "content": "[3] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia, 2023. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 181, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 287, + 235 + ], + "type": "text", + "content": "[4] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 269 + ], + "type": "text", + "content": "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 286, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 313 + ], + "type": "text", + "content": "[6] Mikolaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD gans. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 315, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 287, + 379 + ], + "type": "text", + "content": "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Proceedings of the European Conference on Computer Vision (ECCV). Springer International Publishing, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 381, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 287, + 445 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 448, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 287, + 512 + ], + "type": "text", + "content": "[9] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "text", + "content": "[10] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "text", + "content": "[11] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[12] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing relightable and animatable avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Zijian Dong, Chen Guo, Jie Song, Xu Chen, Andreas Geiger, and Otmar Hilliges. Pina: Learning a personalized implicit neural avatar from a single rgb-d video sequence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "text", + "content": "[16] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human pose estimation in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "text", + "content": "[17] Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "text", + "content": "[18] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[19] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "type": "text", + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "text", + "content": "[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[22] Hsuan-I Ho, Lixin Xue, Jie Song, and Otmar Hilliges. Learning locally editable virtual humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 526, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 580 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[25] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[26] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Selfrecon: Self reconstruction your digital avatar from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "546" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Ruixiang Jiang, Can Wang, Jingbo Zhang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "text", + "content": "[28] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[30] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 8, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 342 + ], + "type": "text", + "content": "[32] Manuel Kaufmann, Velko Vechev, and Dario Mylonopoulos. aitviewer, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "text", + "content": "[33] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (TOG), 32(3):1-13, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 411 + ], + "type": "text", + "content": "[34] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "type": "text", + "content": "[35] Byungjun Kim, Patrick Kwon, Kwangho Lee, Myunggi Lee, Sookwan Han, Daesik Kim, and Hanbyul Joo. Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "text", + "content": "[36] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "text", + "content": "[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[39] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "type": "text", + "content": "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 7, 8, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "text", + "content": "[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 239, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 261 + ], + "type": "text", + "content": "[43] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm, 1998. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 338 + ], + "type": "text", + "content": "[44] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[45] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to Dress 3D People in Generative Clothing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 393, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 448 + ], + "type": "text", + "content": "[46] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "text", + "content": "[48] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "text", + "content": "[50] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "547" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[53] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 239 + ], + "type": "text", + "content": "[54] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "text", + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[58] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 459, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 525 + ], + "type": "text", + "content": "[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "text", + "content": "[60] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[61] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "text", + "content": "[62] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[63] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid represent" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "tation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "type": "text", + "content": "[64] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 161, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 161, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 161, + 545, + 217 + ], + "type": "text", + "content": "[65] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 217, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 282 + ], + "type": "text", + "content": "[66] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 703-735. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 282, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 545, + 327 + ], + "type": "text", + "content": "[67] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "text", + "content": "[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[69] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 8, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "text", + "content": "[70] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "text", + "content": "[71] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "type": "text", + "content": "[72] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Annual Conference of the European Association for Computer Graphics (EUROGRAPHICS), pages 641-676. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 624, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 679 + ], + "type": "text", + "content": "[73] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 5, 6, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[74] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "548" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 361 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[75] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 6, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[76] Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 8, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "text", + "content": "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "text", + "content": "[79] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "549" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_content_list.json b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..734b98c46c772b97cfee9150e7e5660758d1c22f --- /dev/null +++ b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_content_list.json @@ -0,0 +1,1705 @@ +[ + { + "type": "text", + "text": "Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding", + "text_level": 1, + "bbox": [ + 187, + 130, + 782, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chaolei Tan $^{1}$ Jianhuang Lai $^{1,2,3}$ Wei-Shi Zheng $^{1,2,3}$ Jian-Fang Hu $^{1,2,3*}$", + "bbox": [ + 169, + 202, + 841, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China", + "bbox": [ + 169, + 222, + 810, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Guangdong Province Key Laboratory of Information Security Technology, China", + "bbox": [ + 169, + 238, + 828, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 114, + 256, + 890, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tanchlei@mail2.sysu.edu.cn, stsljh@mail.sysu.edu.cn, wszheng@ieee.org, hujf5@mail.sysu.edu.cn", + "bbox": [ + 86, + 277, + 910, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 300, + 313, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Paragraph Grounding (VPG) is an emerging task in video-language understanding, which aims at localizing multiple sentences with semantic relations and temporal order from an untrimmed video. However, existing VPG approaches are heavily reliant on a considerable number of temporal labels that are laborious and time-consuming to acquire. In this work, we introduce and explore Weakly-Supervised Video Paragraph Grounding (WSVPG) to eliminate the need of temporal annotations. Different from previous weakly-supervised grounding frameworks based on multiple instance learning or reconstruction learning for two-stage candidate ranking, we propose a novel siamese learning framework that jointly learns the cross-modal feature alignment and temporal coordinate regression without timestamp labels to achieve concise one-stage localization for WSVPG. Specifically, we devise a Siamese Grounding Transformer (SiamGTR) consisting of two weight-sharing branches for learning complementary supervision. An Augmentation Branch is utilized for directly regressing the temporal boundaries of a complete paragraph within a pseudo video, and an Inference Branch is designed to capture the order-guided feature correspondence for localizing multiple sentences in a normal video. We demonstrate by extensive experiments that our paradigm has superior practicability and flexibility to achieve efficient weakly-supervised or semi-supervised learning, outperforming state-of-the-art methods trained with the same or stronger supervision.", + "bbox": [ + 75, + 330, + 473, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 748, + 209, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Natural Language Video Grounding (NLVG) is an essential area in vision-language understanding, which has received increasing attention due to its wide range of real-world applications such as video retrieval [3, 17, 21, 23, 80, 94], video summarization [50, 55, 59, 60, 83, 99], action segmentation [19, 27, 35, 37, 45, 72], video question answering [30, 36, 38, 41, 82, 89], etc. Most previous works focus", + "bbox": [ + 75, + 772, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a07f72f59db4d4e9d5779d04b2bdc7fd7225047d51fddad602d14601e61d026.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 506, + 300, + 890, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dcd0c75ad0246d487780dbf293a74255f905341c913c5972890fa3ecc999c344.jpg", + "image_caption": [ + "Figure 1. (a) Chronological cross-modal alignment in a video and its paired sentences. (b) Pseudo boundary supervision for regressing paragraph timestamps in a composed pseudo video." + ], + "image_footnote": [], + "bbox": [ + 504, + 400, + 888, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on tackling Video Sentence Grounding (VSG) proposed in [1, 22], which targets at localizing the temporal boundaries of an individual sentence from an untrimmed video. However, localizing single sentences can be ambiguous because the contextual information conveyed by multiple sentences is necessary to uniquely determine the temporal locations of input queries. To alleviate such issue, Bao et al. [4] proposed to contextualize video grounding into localizing multiple events indicated by sentences of a paragraph from the video, which is called Video Paragraph Grounding (VPG).", + "bbox": [ + 496, + 566, + 890, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although remarkable progress has been attained in tackling VSG and VPG problems under a fully-supervised setting, the extremely prohibitive overheads of manually annotating temporal boundaries for language queries limit these methods to utilize large-scale video-text data. In addition, the subjectivity of annotators inevitably brings label noises that may adversely affect model training. Recently, weakly-supervised methods free of temporal annotations have become increasingly popular in the area of video grounding, aiming to address the above limitations. These approaches can be mainly categorized into multiple instance learning methods and reconstruction learning methods and are typ-", + "bbox": [ + 496, + 719, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13569", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ically based on a propose-and-rank pipeline. Nevertheless, both of these paradigms assume the contributions of proposals to the contrastive or reconstruction loss accurately represent the proposal quality, which is not necessarily the case during model learning. Moreover, the quadratic complexity of proposal schemes prevents them from scaling up in parameters or training data, and the adopted supervision of video-level contrast or query-based reconstruction in prior works is not temporal-sensitive thus suffering from a huge gap with fully-supervised temporal guidance. Besides, all of the weakly-supervised methods in video grounding are tailored for tackling VSG while the weakly-supervised setting of VPG (i.e., WSVPG) has been understudied so far.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To circumvent the aforementioned drawbacks and explore an efficient weakly-supervised framework for VPG, we seek to mine the unique characteristics and underlying supervision from the intrinsic structure of video-paragraph pairs for model training. On the one hand, as observed in Figure 1 (a), the temporal location of an event is highly correlated with the position of the sentence describing that event in the paragraph. For example, the sentence appearing in the middle of the paragraph tends to have stronger relevance with visual content located around the temporal midpoint of the video. On the other hand, dense visual events mentioned in the paragraph approximately represent the global video content that unambiguously distinguishes itself from another video, which can be observed in Figure 1 (b). Therefore, inserting the query-related video into another irrelevant video automatically generates pseudo boundary labels close to the ground-truth when regarding the complete paragraph as language query for video grounding.", + "bbox": [ + 75, + 289, + 472, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above observation, we propose a novel Siamese Grounding TRansformer (SiamGTR) for WSVPG. It jointly learns the cross-modal alignment and boundary regression via two siamese branches without generating proposals. Specifically, we propose to construct an Augmentation Branch (AB) which takes as input a pseudo video and adopts a complete paragraph as the language query to learn high-quality boundary supervision for localization. Also, an Inference Branch (IB) is designed to receive a normal video as input and is enforced to capture the order-guided cross-modal correspondence for attending over the specific video content relevant to each sentence. The two weight-sharing branches are effective to transfer complementary supervision for joint boundary prediction and feature association, which yields a weakly-supervised model with superior generalization through a concise one-stage pipeline. Extensive experiments verify the effectiveness of our model and show our method with the same or weaker supervision surpasses prior state-of-the-arts. In summary, our contributions are:", + "bbox": [ + 75, + 565, + 472, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce the task of Weakly-Supervised Video Paragraph Grounding (WSVPG), which aims to train a model for localizing multiple events indicated by queries with", + "bbox": [ + 76, + 854, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "out the supervision of timestamp labels.", + "bbox": [ + 511, + 90, + 777, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Siamese Grounding TRansformer (SiamGTR) for concise and efficient one-stage weakly-supervised learning of video paragraph grounding. It is composed of two weight-sharing branches including an Augmentation Branch (AB) for learning boundary regression of pseudo boundaries and an Inference Branch (IB) for learning order-guided cross-modal feature alignment.", + "- Extensive experiments verify the efficacy of our method, and demonstrate that our framework under the same or weaker supervision outperforms state-of-the-arts." + ], + "bbox": [ + 500, + 107, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 270, + 640, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Sentence Grounding", + "text_level": 1, + "bbox": [ + 500, + 291, + 741, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully-Supervised Video Sentence Grounding. Plenty of approaches [1, 7, 12, 22, 29, 40, 43, 44, 47, 53, 64, 73, 74, 78, 79, 84-86, 90, 91, 93, 95-97, 100-103] have been proposed to address Fully-Supervised Video Sentence Grounding (FSVSG). In general, these works can be roughly categorized into proposal-based and proposal-free methods. Specifically, proposal-based methods [1, 22, 40, 43, 44, 73, 74, 78, 85, 86, 90, 93, 95, 101, 102] involve a proposal generation stage using sliding windows [22], anchor proposals [12, 43, 44, 47, 74, 90, 93, 95] or 2D temporal maps [64, 73, 78, 85, 101, 102], after which the generated proposals are ranked according to the query matching scores with potential post-processing like Non-Maximum Suppression (NMS). In contrast, proposal-free methods [7, 53, 79, 91, 96, 97, 100, 103] remove the dense proposal generation and score ranking process by directly regressing timestamps [7, 53, 84, 91, 103], predicting boundary distributions [96, 97, 100] or using reinforcement learning [79], which improves the computation efficiency and scenario adaptability. Following the line of proposal-free works, we propose a novel weakly-supervised regression-based framework that shows superior performance and practicability.", + "bbox": [ + 496, + 311, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-Supervised Video Sentence Grounding. Weakly-Supervised Video Sentence Grounding (WSVSG) [8, 11, 26, 28, 42, 49, 52, 65, 68, 75-77, 88, 104-107] has become a popular research area because of the severe dependence of FSVSG approaches on laborious and expensive manual temporal annotations. Most of existing WSVSG methods are based on a two-stage pipeline using multiple instance learning [26, 49, 75, 88], reconstruction learning [42], or the combination of both [106, 107]. In particular, Chen et al. [11] have proposed a video composition strategy to generate pseudo temporal labels for WSVSG, which is the most related work to ours. However, several inherent drawbacks are involved in this approach. Firstly, individual sentences only describe local video content, thus viewing the starting/ending locations of foreground video as temporal boundaries of an individual sentence produces a weak and noisy temporal alignment, which is unsuitable for accurate", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13570", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "boundary supervision. Moreover, it simply adapts an existing fully-supervised proposal-based framework for weakly-supervised training, which leads to inferior generalization caused by the large train-test discrepancy. Distinct from all of the above works, we design a novel siamese framework to capture the essential characteristics of video paragraph grounding for weakly-supervised learning.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Video Paragraph Grounding", + "text_level": 1, + "bbox": [ + 76, + 205, + 334, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Paragraph Grounding (VPG) is introduced by Bao et al. [4], which aims to jointly localize multiple sentences of a paragraph from an untrimmed video. Shi et al. [62] presented an end-to-end network by re-purposing transformers into language-conditioned regressors. Jiang et al. [31] proposed to employ contrastive encoders for contrastive learning between video-paragraph pairs. Tan et al. [67] proposed a hierarchical semantic correspondence network for modeling hierarchical video-language alignment and grounding multiple levels of language queries in the video. Particularly, Jiang et al. [31] first explored Semi-Supervised Video Paragraph Grounding (SSVPG) to relieve the annotation burden of temporal labels. However, SSVPG is still not quite practical considering the expensive cost of temporal annotations in untrimmed videos. Besides, these semi-supervised methods still require a considerable proportion of temporal labels up to at least $10\\%$ [31] for training. To thoroughly get rid of the temporal annotations, we pioneer to explore the weakly-supervised setting in VPG.", + "bbox": [ + 75, + 227, + 473, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Siamese Networks", + "text_level": 1, + "bbox": [ + 76, + 523, + 253, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Siamese networks [6] are weight-sharing neural networks. There have been a wide range of scenarios where siamese networks are applied for achieving different purposes, such as face verification [66], image recognition [33], object tracking [5], etc. In particular, siamese networks are commonly used in contrastive self-supervised learning methods [2, 10, 13-15, 24, 25, 69, 81, 108], in which augmented views of the same or different instance are forwarded into multiple weight-sharing network copies for learning a generalizable visual representation via instance-level discrimination. In this work, we explore a new way to combine the transferability of siamese architectures and the flexibility of transformer architectures for concise and efficient weakly-supervised learning of video paragraph grounding.", + "bbox": [ + 75, + 545, + 470, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 773, + 210, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 800, + 187, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Formulation. Given an untrimmed video $\\mathcal{V}$ and a paragraph query $\\mathcal{P}$ consisting of $N$ temporally ordered sentences $\\{\\mathcal{S}_i\\}_{i=1}^N$ , the goal of Video Paragraph Grounding (VPG) is to simultaneously localize the temporal intervals $\\mathcal{T} = \\{(\\tau_i^{\\mathrm{st}}, \\tau_i^{\\mathrm{ed}})\\}_{i=1}^N$ of all the events described by sentences", + "bbox": [ + 76, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2e2781a658a58acd039e4de131bab5d48e8d0bb526f09b19cb8e008e7e33bfc8.jpg", + "image_caption": [ + "Figure 2. Our siamese framework for joint alignment and regression." + ], + "image_footnote": [], + "bbox": [ + 550, + 88, + 709, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/71bc0739af9c28e8c4279d701bbd3063cc32d1ac2149ffe969a1d6ab057af9f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 88, + 839, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the paragraph, where $\\tau_{i}^{\\mathrm{st}}$ and $\\tau_{i}^{\\mathrm{ed}}$ are the starting and ending timestamps for the $i$ -th sentence $S_{i}$ , respectively.", + "bbox": [ + 496, + 229, + 890, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Siamese Learning. An overview of our siamese framework is shown in Figure 2. Overall, we jointly train an augmentation branch and an inference branch with shared parameters and structures for learning the complementary abilities of cross-modal feature alignment and temporal boundary regression. These two branches follow the same workflow to first encode the text queries and input video into unimodal features, after which the query features are iteratively used in the transformer decoder to extract relevant information from the video features for timestamp decoding. For testing, we only keep the pipeline of inference branch for boundary prediction. More architectural details are illustrated in Figure 3 and are elaborated in the following sections.", + "bbox": [ + 496, + 261, + 892, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Extraction", + "text_level": 1, + "bbox": [ + 500, + 467, + 681, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Feature Extraction. For each input video, we divide it into consecutive clips consisting of a fixed number of frames for feature extraction. Specifically, a frozen pretrained 3D Convolutional Neural Network (3D-CNN) [70] and a linear projection layer are successively employed to obtain a 1D feature vector for each short clip, resulting in a video feature sequence $\\mathcal{F}_{\\mathrm{v}} \\in \\mathbb{R}^{L \\times D}$ , where $L$ and $D$ are the sequence length and hidden dimension, respectively.", + "bbox": [ + 496, + 489, + 890, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text Feature Extraction. For each input paragraph consisting of $N$ sentences, we first utilize a frozen pre-trained word embedding model tookenize and embed the text into a sequence of word vectors. Then, a bidirectional Gated Recurrent Unit (GRU) [16] is employed on each sentence, and the last hidden states in both directions are concatenated and then projected by a linear layer to construct the sentence features $\\mathcal{F}_{\\mathrm{s}}\\in \\mathbb{R}^{N\\times D}$ , where $D$ is the hidden dimension.", + "bbox": [ + 496, + 611, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Augmentation Branch", + "text_level": 1, + "bbox": [ + 500, + 741, + 705, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The augmentation branch aims to learn accurate boundary regression from pseudo videos with paragraph queries, which naturally transfers to the inference branch via the shared feature space established by the siamese structure.", + "bbox": [ + 496, + 763, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pseudo Data Generation. To drive end-to-end weakly-supervised regression learning, the input stream of the augmentation branch should provide reliable and direct boundary supervision. Drawing inspiration from the boundary-sensitive video pretext tasks [11, 54, 87], we propose to uti", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a245b9fc51cd6f2fbd98c7cda54eff7d73eaf65fa8de3049543f3ba4f0ae5b6d.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed Siamese Grounding TRansformer (SiamGTR) architecture. The augmentation branch (abbreviated as A.B.) takes the pseudo video features derived from randomly inserting the query-related video features into irrelevant video features. It learns to temporally regress the interval of interest from the pseudo video with the paragraph as query. The Inference Branch (abbreviated as I.B.) receives normal video features for learning the cross-modal feature alignment among multiple sentences in the video." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lize the synthesized videos paired with paragraph queries to serve as a well-suited source of surrogate boundary supervision. Specifically, for each input video $\\nu$ and its paragraph description $\\mathcal{P}$ , we randomly sample a background video $\\nu_{\\mathrm{bg}}$ to obtain irrelevant video content to $\\mathcal{P}$ . Because $\\mathcal{P}$ could be treated as a unique referring expression specific to $\\nu$ , the search space of background videos can be the entire training set for increasing the data diversity. Denote video features of $\\nu$ and $\\nu_{\\mathrm{bg}}$ as $\\mathcal{F}_{\\mathrm{v}}$ and $\\mathcal{F}_{\\mathrm{bg}}$ respectively, we construct the pseudo video features $\\mathcal{F}_{\\mathrm{aug}}$ as:", + "bbox": [ + 75, + 363, + 472, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathrm {a u g}} = \\operatorname {N R S} \\left(\\operatorname {C o n c a t} \\left(\\mathcal {F} _ {\\mathrm {b g}} ^ {1: I}, \\operatorname {R R S} \\left(\\mathcal {F} _ {\\mathrm {v}}\\right), \\mathcal {F} _ {\\mathrm {b g}} ^ {I + 1:}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 521, + 468, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I$ is a randomly generated inserting index within the feature sequence $\\mathcal{F}_{\\mathrm{bg}}$ and Concat ( $\\cdot$ ) denotes the temporal concatenation. RRS ( $\\cdot$ ) and NRS ( $\\cdot$ ) respectively represent a Random Re-Sampling operation to stochastically re-scale the length of $\\mathcal{F}_{\\mathrm{v}}$ and a Normalized Re-Sampling operation that converts the length of $\\mathcal{F}_{\\mathrm{aug}}$ into a fixed number of $T$ . Thereafter, we compute the synthesized temporal boundaries ( $\\tau_{\\mathrm{aug}}^{\\mathrm{st}}, \\tau_{\\mathrm{aug}}^{\\mathrm{ed}}$ ) in the pseudo video $\\mathcal{V}_{\\mathrm{aug}}$ as follows:", + "bbox": [ + 75, + 551, + 472, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {\\mathrm {a u g}} ^ {\\mathrm {s t}} = \\frac {I + \\Delta I ^ {\\mathrm {s t}}}{r L + L _ {\\mathrm {b g}}}, \\tau_ {\\mathrm {a u g}} ^ {\\mathrm {e d}} = \\frac {I + r L - \\Delta I ^ {\\mathrm {e d}}}{r L + L _ {\\mathrm {b g}}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 680, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L$ and $L_{\\mathrm{bg}}$ respectively indicate the length of $\\mathcal{F}_{\\mathrm{v}}$ and $\\mathcal{F}_{\\mathrm{bg}}$ with $r$ being the random re-scaling factor of RRS $(\\cdot)$ . To alleviate potential synthesis artifacts and boundary uncertainty, we further introduce a Random Boundary Shifting (RBS) strategy that incorporates small random offsets $\\Delta I^{\\mathrm{st}}$ and $\\Delta I^{\\mathrm{ed}}$ into computing the pseudo labels, which is simple yet effective to boost the quality of boundary supervision.", + "bbox": [ + 75, + 719, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Encoder. Based on the obtained pseudo video features $\\mathcal{F}_{\\mathrm{aug}}$ , we then encode the temporal contextual information across multiple clips by feeding $\\mathcal{F}_{\\mathrm{aug}}$ to a video encoder, which follows a similar architecture of DETR encoders [9, 46, 71], i.e., each encoder layer consists of a", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "multi-head self-attention module equipped with additive sinusoidal positional encodings and a feed-forward network. For simplicity, we omit the layer index and denote the input of each video encoder layer as $\\mathcal{X}_{\\mathrm{aug}}$ , then a set of Modulated Positional Encodings (MPE) are given as:", + "bbox": [ + 496, + 363, + 890, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {\\mathrm {p e}} = \\operatorname {M L P} \\left(\\mathcal {X} _ {\\mathrm {a u g}}\\right) \\odot \\operatorname {P E} \\left(\\mathcal {X} _ {\\mathrm {a u g}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 444, + 890, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{\\mathrm{pe}} \\in \\mathbb{R}^{T \\times D}$ and $\\odot$ denotes element-wise multiplication. PE $(\\cdot)$ is the sinusoidal function [71] with respect to temporal locations of $\\mathcal{X}_{\\mathrm{aug}}$ , MLP $(\\cdot)$ denotes a two-layer feed-forward network computed on $\\mathcal{X}_{\\mathrm{aug}}$ . In each encoder layer, $\\mathcal{X}_{\\mathrm{pe}}$ is only added to the input of projection layers of queries and keys in the self-attention mechanism. The encoded video features of the last video encoder layer, also called the encoder memory features $\\mathcal{M}_{\\mathrm{aug}}$ , are further fed to the query-guided decoder for localization prediction.", + "bbox": [ + 496, + 464, + 892, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query Encoder. Since there are strong contextual correlations among multiple sentences in the paragraph query, here we utilize a vanilla transformer encoder [71] to help reason the semantic and chronological relationships of events. To extract a global representation of the complete paragraph, we initialize an extra learnable query token $\\mathcal{F}_{\\mathrm{p}}$ and integrate it with the sentence features $\\mathcal{F}_{\\mathrm{s}}$ by concatenation. Thus, the input of the query encoder is constructed as follows:", + "bbox": [ + 496, + 601, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {\\mathrm {q}} = \\left[ \\mathcal {F} _ {\\mathrm {p}}, \\mathcal {F} _ {\\mathrm {s}} ^ {1}, \\mathcal {F} _ {\\mathrm {s}} ^ {2}, \\dots , \\mathcal {F} _ {\\mathrm {s}} ^ {i}, \\dots \\mathcal {F} _ {\\mathrm {s}} ^ {N} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 724, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{\\mathrm{q}} \\in \\mathbb{R}^{(N + 1) \\times D}$ and $\\mathcal{F}_{\\mathrm{s}}^{i}$ is the $i$ -th sentence feature in $\\mathcal{F}_{\\mathrm{s}}$ . We first normalize $\\mathcal{X}_{\\mathrm{q}}$ , add fixed sinusoidal positional encodings to it, and then iteratively employ self-attention modules and feed-forward networks to contextualize the global and local tokens that respectively represent the query semantics of the entire paragraph and the sentences. The output features of the last query encoder layer are denoted as $\\mathcal{Z}_{\\mathrm{q}}$ and are further forwarded to the query decoder.", + "bbox": [ + 496, + 748, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conceptual Semantic Connector. Although the siamese network structure can implicitly transfer boundary knowl-", + "bbox": [ + 496, + 871, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13572", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "edge from the augmentation branch to the inference branch, there is still a certain semantic gap between the query representations of short sentences and long paragraphs. To narrow this gap, we develop a Conceptual Semantic Connector (CSC) module for explicit semantic guidance. Specifically, we first collect high-frequency linguistic concepts, including verbs and nouns from the training corpus, and then construct a set of dictionary features by selecting and projecting the Glove vectors [56]. The loss $\\mathcal{L}_{\\mathrm{csc}}$ is computed as:", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c s c}} = \\mathrm {B C E} \\left(y _ {\\text {c p t}} ^ {\\mathrm {p}}, \\hat {y} _ {\\text {c e p t}} ^ {\\mathrm {p}}\\right) + \\mathrm {B C E} \\left(y _ {\\text {s c t}} ^ {\\mathrm {s}}, \\hat {y} _ {\\text {c e p t}} ^ {\\mathrm {s}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 227, + 468, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{BCE}(\\cdot)$ is the binary cross-entropy loss. $y_{\\mathrm{cept}}^{\\mathrm{p}}$ and $y_{\\mathrm{cept}}^{\\mathrm{s}}$ are multi-hot labels indicating semantic concepts contained by the paragraph and sentence queries, respectively. $\\hat{y}_{\\mathrm{cept}}^{\\mathrm{s}}$ and $\\hat{y}_{\\mathrm{cept}}^{\\mathrm{p}}$ are the concept predictions obtained by dot-product between the conceptual dictionary features and the textual query features $\\mathcal{Z}_{\\mathrm{q}}$ with a sigmoid activation.", + "bbox": [ + 75, + 246, + 468, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Query Decoder. Inspired by the spiritual ideas of dynamic-anchor DETR decoders [39, 46, 51, 98], we design a novel query decoder that enables dynamic position-aware decoding of language queries. Specifically, each query decoder layer consists of a self-attention module, a cross-attention module, and a feed-forward network, with dynamically adjustable anchor boxes to indicate the query-specific location information. Initially, the input anchor boxes are set to all zeros, and an MLP is used to estimate a set of seed anchor boxes based on the cross-modal interactions between $\\mathcal{M}_{\\mathrm{aug}}$ and $\\mathcal{Z}_{\\mathrm{q}}$ , i.e., the output features of the first decoder layer $\\mathcal{Q}_{\\mathrm{aug}}^{(1)} \\in \\mathbb{R}^{(N + 1) \\times D}$ are mapped into the seed anchor boxes $\\mathcal{A}_{\\mathrm{aug}}^{(1)} \\in \\mathbb{R}^{(N + 1) \\times 2}$ by the MLP. For the $(i + 1)$ -th decoder layer, we first convert the box coordinates of the input anchors $\\mathcal{A}_{\\mathrm{aug}}^{(i)}$ into high-dimensional sinusoidal embeddings $\\mathcal{F}_{\\mathrm{a}}^{(i)} \\in \\mathbb{R}^{(N + 1) \\times D}$ by a sinusoidal function [46, 71] and further obtain $\\mathcal{H}_{\\mathrm{a}}^{(i)}$ by projecting $\\mathcal{F}_{\\mathrm{a}}^{(i)}$ with an MLP. Afterwards, we conduct the self-attention operation and update the query features in the decoder as follows:", + "bbox": [ + 75, + 335, + 472, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)} \\leftarrow \\text {S e l f - A t t n} \\left\\{ \\begin{array}{l} Q = \\varphi_ {q} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) + \\varphi_ {q} ^ {p} \\left(\\mathcal {H} _ {\\mathrm {a}} ^ {(i)}\\right) \\\\ K = \\varphi_ {k} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) + \\varphi_ {k} ^ {p} \\left(\\mathcal {H} _ {\\mathrm {a}} ^ {(i)}\\right) \\\\ V = \\varphi_ {v} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 635, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}$ is the updated query features after the self-attention operation in the $(i + 1)$ -th decoder layer. The above series of $\\varphi$ functions are used to indicate different linear projection layers for the content part or position part of the queries, keys, or values in the self-attention mechanism. Then, we conduct a cross-attention operation to extract useful cross-modal interactive information as follows:", + "bbox": [ + 75, + 709, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)} \\leftarrow \\text {C r o s s - A t t n} \\left\\{ \\begin{array}{l} Q = \\left[ \\varphi_ {q} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)}\\right); \\varphi_ {q} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {a}} ^ {(i)}\\right) \\right] \\\\ K = \\left[ \\varphi_ {k} ^ {c} \\left(\\mathcal {M} _ {\\mathrm {a u g}}\\right); \\varphi_ {k} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {p e}} ^ {\\mathcal {M}}\\right) \\right] \\\\ V = \\varphi_ {v} ^ {c} \\left(\\mathcal {M} _ {\\mathrm {a u g}}\\right) + \\varphi_ {v} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {p e}} ^ {\\mathcal {M}}\\right) \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 829, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{pe}}^{\\mathcal{M}} = \\mathrm{PE}\\left(\\mathcal{M}_{\\mathrm{aug}}\\right)$ and the query features $\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}$ is then further updated by a feed-forward network as the feature output of the $(i + 1)$ -th decoder layer. Then the anchor boxes are dynamically updated as $\\mathcal{A}_{\\mathrm{aug}}^{(i + 1)} \\gets \\mathcal{A}_{\\mathrm{aug}}^{(i)} + \\Delta \\mathcal{A}_{\\mathrm{aug}}^{(i)}$ where we utilize an MLP layer to predict the relative offsets, i.e., $\\Delta \\mathcal{A}_{\\mathrm{aug}}^{(i)} \\in \\mathbb{R}^{(N + 1) \\times 2}$ , based on the updated query features $\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}$ . $\\mathcal{A}_{\\mathrm{aug}}^{(i + 1)}$ continues to be forwarded to the next decoder layer for computing $\\mathcal{F}_{\\mathrm{a}}^{(i + 1)}$ and $\\mathcal{H}_{\\mathrm{a}}^{(i + 1)}$ .", + "bbox": [ + 496, + 89, + 890, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Boundary Prediction. Based on the output features and attention weights of the last query decoder layer, we simply use an MLP predictor to predict the paragraph timestamps, i.e., $\\hat{\\mathcal{T}}_{\\mathrm{aug}}^{\\mathrm{p}} = (\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{ed}})$ . Similarly, the sentence timestamps $\\widehat{\\mathcal{T}}_{\\mathrm{inf}}^{\\mathrm{s}} = \\left\\{(\\hat{\\tau}_{\\mathrm{j}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{j}}^{\\mathrm{ed}})\\right\\}_{j = 1}^{N}$ can also be obtained by feeding last-layer output from the inference branch to the same MLP.", + "bbox": [ + 496, + 223, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-Consistent Boundary Regression. We improve the attention-agnostic regression loss $\\mathcal{L}_{\\mathrm{reg}}$ to make it aware of the model's self-consistent scores, where the main idea is to selectively optimize the regression loss of self-consistent samples for better weakly-supervised regression learning. Self-consistent samples have high attention weights over the pseudo ground-truth intervals, which are more suitable for learning less noisy supervision for accurate boundary prediction. Specifically, the self-consistent boundary regression loss $\\mathcal{L}_{\\mathrm{sreg}}$ is defined as:", + "bbox": [ + 496, + 327, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s c r e g}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\mathrm {L} 1} + \\mathcal {L} _ {\\mathrm {G I o U}}, & \\text {i f} \\quad s _ {\\mathrm {a t t}} > \\beta \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 487, + 890, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{L1}}$ and $\\mathcal{L}_{\\mathrm{GIoU}}$ respectively represent L1 and Generalized Intersection over Union (GIoU) [58] loss computed between $(\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{ed}})$ and $(\\tau_{\\mathrm{aug}}^{\\mathrm{st}},\\tau_{\\mathrm{aug}}^{\\mathrm{ed}})$ . $\\beta$ is set to 0.5 and $s_{\\mathrm{att}}$ is the attention sum over the pseudo ground-truth interval.", + "bbox": [ + 496, + 535, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference Branch", + "text_level": 1, + "bbox": [ + 500, + 604, + 669, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our siamese framework, the inference branch shares the same parameter weights and network structure with the augmentation branch, i.e., a video encoder, a query encoder, and a query decoder. The only difference lies in the input streams and objectives, i.e., the inference branch receives a normal video during training to learn in-domain cross-modal correspondence that cannot be acquired through the pseudo video stream, which significantly improves the generalization ability of the model. Specifically, it receives the encoded normal video features $\\mathcal{M}_{\\mathrm{inf}}$ and the encoded text query features $\\mathcal{Z}_{\\mathrm{q}}$ as the decoder input and generates the hidden query features $\\mathcal{Q}_{\\mathrm{inf}}$ for boundary prediction.", + "bbox": [ + 496, + 628, + 890, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Order-guided Attention Loss. The chronological prior given by the sentence order provides explicit guidance for learning cross-modal alignment between the video and language features during decoding. To learn the order-guided cross-modal correspondence, we constrain cross-modal attention weights in the decoder as follows:", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o g a}} = \\max \\left(0, \\Delta m T + \\sum_ {t = 1} ^ {T} t \\alpha_ {\\mathrm {s}} ^ {j} (t) - \\sum_ {t = 1} ^ {T} t \\alpha_ {\\mathrm {s}} ^ {j + 1} (t)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 88, + 468, + 125 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{\\mathrm{s}}^{j}(t)$ and $\\alpha_{\\mathrm{s}}^{j + 1}(t)$ are attention weights over video features for the $j$ -th and $(j + 1)$ -th sentence from the last decoder layer, respectively. $\\Delta m$ is the minimal distance between attention centroids. This loss is mainly contributed by the inference branch with $\\Delta m = \\frac{1}{2N}$ and partly contributed by the augmentation branch with $\\Delta m = \\frac{1}{4N}$ .", + "bbox": [ + 76, + 125, + 467, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Auxiliary Losses. To exploit more guidance for weakly-supervised representation learning, we employ three auxiliary losses including a cross-branch loss $\\mathcal{L}_{\\mathrm{cb}}$ , an anchor ranking loss $\\mathcal{L}_{\\mathrm{ar}}$ and a pseudo attention loss $\\mathcal{L}_{\\mathrm{pa}}$ . Specifically, $\\mathcal{L}_{\\mathrm{cb}}$ utilizes the semantic consistency constraint of output features across siamese branches, which is calculated analogous to MoCo [25]. $\\mathcal{L}_{\\mathrm{ar}}$ is used for inducing the decoder to learn a set of order-preserving anchor boxes, which is computed on the anchor boxes like in the equation (9). $\\mathcal{L}_{\\mathrm{pa}}$ makes use of attention supervision between the pseudo video and language features from the augmentation branch, which follows the calculation proposed in [62, 91].", + "bbox": [ + 76, + 218, + 467, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 407, + 290, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weakly-Supervised Learning. The weakly-supervised loss of our proposed framework can be formulated as a weighted sum of the two losses from the siamese branches as $\\mathcal{L}_{\\mathrm{WS}} = \\lambda_{\\mathrm{sreg}}\\mathcal{L}_{\\mathrm{sreg}} + \\lambda_{\\mathrm{oga}}\\mathcal{L}_{\\mathrm{oga}}$ , where $\\lambda_{\\mathrm{sreg}}$ and $\\lambda_{\\mathrm{oga}}$ are scalar weights to balance the contributions of the two different losses. The overall loss for the weakly-supervised model is defined as $\\mathcal{L} = \\mathcal{L}_{\\mathrm{WS}} + \\lambda_{\\mathrm{csc}}\\mathcal{L}_{\\mathrm{csc}} + \\mathcal{L}_{\\mathrm{aux}}$ .", + "bbox": [ + 76, + 430, + 467, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semi-Supervised Learning. Although our framework is initially designed for weakly-supervised learning, it can be easily adapted for end-to-end semi-supervised learning. Specifically, in addition to $\\mathcal{L}_{\\mathrm{WS}}$ which is calculated on all training samples, we only need to employ an extra fully-supervised loss $\\mathcal{L}_{\\mathrm{FS}}$ on those fully-annotated samples without changing any part of the network structure. The semi-supervised loss $\\mathcal{L}_{\\mathrm{SS}}$ is defined as $\\mathcal{L}_{\\mathrm{SS}} = \\mathcal{L}_{\\mathrm{WS}} + \\mathcal{L}_{\\mathrm{FS}}$ , where $\\mathcal{L}_{\\mathrm{FS}}$ consists of a regression loss and an attention loss and is calculated on the labeled samples in the inference branch.", + "bbox": [ + 76, + 537, + 467, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Inference. As mentioned, the augmentation branch and conceptual semantic connector are discarded, while other inference branch modules are preserved for testing.", + "bbox": [ + 76, + 686, + 467, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 746, + 200, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 771, + 274, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ActivityNet-Captions. ActivityNet-Captions [34] dataset is a large-scale dataset with diverse open-domain content sourced from ActivityNet dataset [34]. There are 14,926 videos and 19,811 localized video-paragraph pairs in total. Each video lasts for 117.60 seconds and each paragraph consists of 3.63 sentences on average. The entire dataset is divided into train/val_1/val_2 sets containing", + "bbox": [ + 76, + 794, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "10,009/4,917/4,885 video-paragraph pairs, respectively. We follow prior works [4, 31] to use val_2 set for testing.", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Charades-CD-OOD. Charades-STA dataset [22] is built from the Charades dataset [63] with indoor activities. Following the previous work [31], we adopt a reorganized version of Charades-STA named Charades-CD-OOD proposed in [92]. It is divided into train/val/test_ood sets consisting of 4,564/333/1,440 video-paragraph pairs, respectively. Specifically, the average video duration is 30.78 seconds and the average paragraph length is 2.41 sentences.", + "bbox": [ + 498, + 121, + 890, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TACoS. TACoS dataset [57] is constructed from the MPII corpus [61] tailored for cooking activities and kitchen scenarios. There are 127 videos in total with each video paired with multiple paragraphs at different granularities. Concretely, there are 1,107,418, and 380 video-paragraph pairs for training, validation, and testing, respectively. The average video length and number of sentences in the paragraph are 4.79 minutes and 8.75 in this dataset, respectively.", + "bbox": [ + 498, + 242, + 890, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous works [4, 31], we adopt mean Intersection over Union (i.e., mIoU) and recall under IoU threshold of $m$ (i.e., $\\mathbf{R}@\\mathbf{m}$ ) as our evaluation metrics. The metrics are averaged over all sentences and $m$ is set to be $\\{0.3, 0.5, 0.7\\}$ for ActivityNet-Captions and Charades-CD-OOD, and $\\{0.1, 0.3, 0.5\\}$ for TACoS.", + "bbox": [ + 498, + 364, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 465, + 715, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fair comparison with existing works [4, 62], we adopt the same C3D network [70] and Glove model [56] as feature extractors. The number of sampled video clips $T$ is set to be 256, 128, and 512 for ActivityNet-Captions, Charades-CD-OOD, and TACoS datasets, respectively. We train the model using Adam [32] optimizer with a fixed learning rate of 0.0001 and a batch size of 32, 32, and 16 for ActivityNet-Captions, Charades-CD-OOD and TACoS, respectively. We select top-100 high-frequency concepts for each dataset, and the loss weights $\\{\\lambda_{\\mathrm{screg}}, \\lambda_{\\mathrm{oga}}, \\lambda_{\\mathrm{csc}}\\}$ are set to $\\{2, 1, 10\\}$ . The number of encoder and decoder layers is set to be 3, and the hidden size $D$ is 256 in all settings.", + "bbox": [ + 498, + 488, + 890, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 500, + 680, + 799, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed SiamGTR with existing state-of-the-art methods for VPG to demonstrate the superiority of our framework. Specifically, 3D-TPN [4, 101], DepNet [4], PRVG [62], SVPTR [31] and HSCNet [67] are fully-supervised approaches requiring temporal annotations for the entire dataset. Besides, the semi-supervised setting has been studied in [31] with several methods developed. For fair comparison with our method, we regard the reconstruction learning method WSSL [18] as one baseline and further develop a more competitive model called Weakly-Supervised Temporal Paragraph Network (WSTPN) by incorporating Beam Search [20] into WSTAN [75]. Specifically, WSTPN utilizes a complete paragraph for multiple", + "bbox": [ + 498, + 703, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13574", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e669ab1585cf9bbff748335194304b6399d417614989cd6822d85eb315a52ec3.jpg", + "table_caption": [ + "Table 1. Comparison on ActivityNet-Captions dataset." + ], + "table_footnote": [], + "table_body": "
MethodSettingR@0.3R@0.5R@0.7mIoU
3D-TPN [101]FS67.5651.4930.92-
DepNet [4]FS72.8155.9133.46-
PRVG [62]FS78.2761.1537.8355.62
SVPTR [31]FS78.0761.7038.3655.91
HSCNet [67]FS81.8966.5744.0359.71
DepNet [4]SS61.4645.1426.7844.11
VPTR [31]SS72.8053.1429.0750.08
SVPTR [31]SS73.3956.7232.7851.98
SiamGTR (Ours)SS78.7559.1134.1254.57
WSSL [18]WS41.9823.34-28.33
WSTPN [75]WS57.7433.0213.6238.54
SiamGTR (Ours)WS75.4357.2330.5652.32
", + "bbox": [ + 78, + 104, + 467, + 289 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8d7e1eadff7d7c8d55267deddba3a7987a96b6d5c3e67067c26f4f622c70bda.jpg", + "table_caption": [ + "Table 2. Comparison on Charades-CD-OOD dataset." + ], + "table_footnote": [], + "table_body": "
MethodSettingR@0.3R@0.5R@0.7mIoU
DepNet [4]FS45.6127.5910.6929.30
STLG [48]FS48.3030.399.79-
SVPTR [31]FS55.1432.4415.5336.01
DepNet [4]SS43.0325.0710.1428.09
STLG [48]SS46.1529.439.38-
VPTR [31]SS45.1324.9810.2228.92
SVPTR [31]SS50.3128.5012.2732.13
SiamGTR (Ours)SS59.0735.4714.9538.87
WSSL [18]WS35.8623.678.27-
WSTPN [75]WS48.6129.2710.7933.49
SiamGTR (Ours)WS57.3333.8712.3137.21
", + "bbox": [ + 78, + 308, + 467, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "instance learning and searches the best sequence of timestamps with the highest overall confidence while maintaining a consistent temporal order with the input sentences.", + "bbox": [ + 75, + 486, + 468, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comprehensive results over three different datasets are shown in Table 1, Table 2, and Table 3, where FS/SS/WS are used to indicate fully/semi/weakly-supervised settings of video paragraph grounding, respectively. First of all, our SiamGTR remarkably surpasses all the other methods under the same supervision in all metrics over the three datasets. Concretely, our framework outperforms WSTPN by $13.78\\%$ , $3.72\\%$ , and $11.16\\%$ in mIoU on ActivityNetCaptions, Charades-CD-OOD and TACoS datasets, respectively. Compared to semi-supervised methods using a considerable number of temporal labels, our weakly-supervised method is also able to achieve comparable or even better results, which demonstrates the effectiveness of our framework in efficient weakly-supervised learning. Furthermore, our framework is flexible and can be easily adapted to semi-supervised learning for further gains. As shown, our semi-supervised model outstrips all semi-supervised state-of-the-art's by a large margin, and it performs better or on par with the fully-supervised SVPTR on all three datasets.", + "bbox": [ + 75, + 531, + 468, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 830, + 230, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies to investigate the contributions of different components on ActivityNet-Captions dataset.", + "bbox": [ + 75, + 854, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of data augmentation. To evaluate the in", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4f063427878e236bc543ab524c52f06c811c8b02b8a85eba086ae5ac519a6b25.jpg", + "table_caption": [ + "Table 3. Comparison on TACoS dataset." + ], + "table_footnote": [], + "table_body": "
MethodSettingR@0.1R@0.3R@0.5mIoU
3D-TPN [101]FS55.0540.3126.54-
DepNet [4]FS56.1041.3427.16-
PRVG [62]FS61.6445.4026.3729.18
SVPTR [31]FS67.9147.8928.2231.42
HSCNet [67]FS76.2859.7442.0040.61
DepNet [4]SS40.2726.9516.5418.68
VPTR [31]SS61.3140.5921.3926.59
SVPTR [31]SS63.0640.1920.0526.10
SiamGTR (Ours)SS67.3049.3531.6932.81
WSTPN [75]WS28.5910.044.769.32
SiamGTR (Ours)WS61.5126.2210.5320.48
", + "bbox": [ + 501, + 104, + 890, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d021248bb2f651fb0e0cc2ea69b8e928c47e14d3ce7d09d682c67235a13bc58b.jpg", + "table_caption": [ + "Table 4. Ablation studies on component designs of our framework. Experimental results are marked from ID (a) $\\sim$ (l). RBS and RRS respectively denote the random boundary shifting and random resampling operations for pseudo data generation. MPE, CSC and DAB stand for the modulated positional encodings, conceptual semantic connector, and dynamic anchor boxes, respectively." + ], + "table_footnote": [], + "table_body": "
IDRBSRRSMPECSCDABR@0.5mIoU
(a)✓✓✓47.1046.04
(b)✓✓✓✓48.1946.96
(c)✓✓✓✓54.2450.94
(d)✓✓✓✓✓57.2352.32
(e)✓✓45.2544.46
(f)✓✓✓46.4745.47
(g)✓✓✓46.9645.23
(h)✓✓✓51.3448.52
(i)✓✓✓✓52.1248.75
(j)✓✓✓✓53.5750.06
(k)✓✓✓✓54.0950.84
(l)✓✓✓✓✓57.2352.32
", + "bbox": [ + 504, + 369, + 890, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fluences of the random boundary shifting and random resampling operations for data augmentation, we remove one or both of them from the training pipeline and the results are shown in Table 4 (a) $\\sim$ (d). The model performance clearly degrades after the removal, which demonstrates the necessity of increasing sample diversity and alleviating overfitting for weakly-supervised cross-modal regression learning.", + "bbox": [ + 496, + 564, + 890, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on module designs. As shown in Table 4 (e) $\\sim$ (l), we conduct detailed ablation studies on module designs to validate the rationality of the proposed model. As observed, the designs of modulated positional encodings and dynamic anchor boxes in the encoder and decoder are consistently beneficial to improving the model capacity for better performance. The conceptual semantic connector that bridges two types of queries is also effective, and it further boosts the performance by $2.26\\%$ in mIoU (50.06% vs. 52.32%) even though the network has been equipped with strong encoders and decoders. We notice the dynamic anchors for decoding are the most crucial component given the sharpest performance drop of $3.57\\%$ with its removal, which indicates the importance to explicitly represent the intermediate location information for video grounding.", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80c4b9256a8f9828148f203399908a2d1dacd8fb3394acadc6c007b0baeaa670.jpg", + "table_caption": [ + "Table 5. Ablation studies on different weakly-supervised losses." + ], + "table_footnote": [], + "table_body": "
ID\\( {\\mathcal{L}}_{\\text{screg }} \\)\\( {\\mathcal{L}}_{\\text{oga }} \\)R@0.3R@0.5R@0.7mIoU
(a)45.8529.8210.9330.97
(b)✓46.9029.3812.0631.82
(c)✓73.5854.5828.8250.62
(d)✓✓75.4357.2330.5652.32
", + "bbox": [ + 104, + 104, + 433, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e0ce66939a2fb303545ac8d741df3ac043c1665050a8aedd330e0433ac7c597f.jpg", + "table_caption": [ + "Table 6. Impact of different auxiliary losses." + ], + "table_footnote": [], + "table_body": "
MethodR@0.3R@0.5R@0.7mIoU
w/o Lcb71.4949.7824.6448.07
w/o Lar72.9650.7226.4249.14
w/o Lpa74.3456.0130.2951.76
Full Model75.4357.2330.5652.32
", + "bbox": [ + 120, + 196, + 418, + 272 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/787dbf9380367c350f50d6c4a07825ef543c490eb5aadb781ca49d6cb34fca18.jpg", + "table_caption": [ + "Table 7. Evaluation on different types of paragraph representation." + ], + "table_footnote": [], + "table_body": "
MethodR@0.3R@0.5R@0.7mIoU
Max-Pooling72.6152.6127.0349.55
Mean-Pooling74.3153.1328.6450.57
Word Concat73.7952.2727.3049.96
Learnable75.4357.2330.5652.32
", + "bbox": [ + 114, + 291, + 424, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on weakly-supervised losses. Ablation studies on contributions of two weakly-supervised losses $\\mathcal{L}_{\\mathrm{sreg}}$ and $\\mathcal{L}_{\\mathrm{oga}}$ are shown in Table 5. In experiment (a) and (c), we remove $\\mathcal{L}_{\\mathrm{sreg}}$ but use a plain regression loss $\\mathcal{L}_{\\mathrm{reg}}$ for ablation analysis. Firstly, simply employing regressive supervision attains inferior performance because the feature-level cross-modal correspondence can hardly be learned with coordinate-level supervision. Furthermore, we find $\\mathcal{L}_{\\mathrm{oga}}$ is critical for learning precise temporal localization since it explicitly guides the model to align video features and language features that are highly likely to be correlated. Besides, we observe that using $\\mathcal{L}_{\\mathrm{sreg}}$ for selecting high-quality regression samples always brings gains to the performance. Impact of auxiliary losses. We investigate the influences of auxiliary losses on the model performance, which involve $\\mathcal{L}_{\\mathrm{cb}}$ for exploiting the cross-branch knowledge, $\\mathcal{L}_{\\mathrm{ar}}$ as guidance to learn a set of order-preserving anchor boxes and $\\mathcal{L}_{\\mathrm{pa}}$ to make use of the feature alignment supervision from the augmentation branch. As presented in Table 6, all three auxiliary losses bring positive impacts, with $\\mathcal{L}_{\\mathrm{cb}}$ being the most effective to improve the mIoU metric by $4.25\\%$ . The reason might lie in the rich complementary knowledge and consistency supervision across the siamese branches.", + "bbox": [ + 76, + 369, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the paragraph representation. The comparison of different types of paragraph representation is shown in Table 7. Four different schemes are included, i.e., mean-pooling or max-pooling the sentence features, embedding the sequence of all word tokens in the paragraph, and using a learnable query token to extract global information. It is clear that adaptively learning a paragraph representation with our method achieves the best performance and the max-pooling scheme performs the worst because it loses too much detailed information. The mean-pooling scheme performs slightly better than the word-concat scheme, which may attribute to the advantage of late fusion at feature level.", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d70722cea3e2ab9e991d32d0a59993de8774329bdbe587033a5dae94b7b101c5.jpg", + "image_caption": [ + "① A woman is sitting down on a floral glass table drilling a design into the pumpkin. ② As she is drilling, two boys are standing next to her watching her and then they suddenly leave. ③ The person behind the camera then picks up the top of the pumpkin to show its empty contents before the lady closes it and continues drilling." + ], + "image_footnote": [], + "bbox": [ + 506, + 123, + 887, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4d57aea25de7fce1ea6fb500717e1d4e5a760b6a6c79aa12cb69df43b98ca9e.jpg", + "image_caption": [ + "A woman is seen standing in a circle and looking over her shoulder. She throws an object off into the distance and is shown again. 3 She throws her arms up and walks away afterwards.", + "Figure 4. Visualization of prediction results from different models." + ], + "image_footnote": [], + "bbox": [ + 506, + 260, + 887, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 500, + 404, + 635, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 4, we intuitively visualize the predicted timestamps from WSTPN and our proposed model trained under the weakly-supervised or semi-supervised setting. Overall, the predicted timestamps given by WSTPN coarsely capture the sentence order relations but with inaccurate boundaries. In contrast, our weakly-supervised model achieves much better results. It is notable that our semi-supervised model generates more fine-grained boundaries, which shows the advantage of our siamese framework in jointly leveraging fewer and weaker labels for efficient learning.", + "bbox": [ + 496, + 428, + 890, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 588, + 617, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we explore the weakly-supervised setting in video paragraph grounding (i.e., WSVPG) to eliminate the dependence of the temporal annotations. To achieve this goal, we propose a novel siamese learning framework to jointly learn the cross-modal feature alignment and temporal coordinate regression without ground-truth supervision. Specifically, we design a novel Siamese Grounding Transformer (SiamGTR) consisting of an augmentation branch and an inference branch. The augmentation branch utilizes the boundary supervision provided by temporally regressing a complete paragraph in a pseudo video, and the inference branch learns the order-guided cross-modal correspondence of multiple sentences in a normal video. Extensive experiments verify the effectiveness of our framework.", + "bbox": [ + 496, + 613, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009), and Guangzhou Science and Technology Plan Project (202201011134).", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13576", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 1, 2", + "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1", + "[4] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, 2021. 1, 3, 6, 7", + "[5] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV Workshops, 2016. 3", + "[6] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. In NeurIPS, 1993. 3", + "[7] Meng Cao, Long Chen, Mike Zheng Shou, Can Zhang, and Yuexian Zou. On pursuit of designing multi-modal transformer for video grounding. In EMNLP, 2021. 2", + "[8] Meng Cao, Fangyun Wei, Can Xu, Xiubo Geng, Long Chen, Can Zhang, Yuexian Zou, Tao Shen, and Daxin Jiang. Iterative proposal refinement for weakly-supervised video grounding. In CVPR, 2023. 2", + "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 4", + "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3", + "[11] Jiaming Chen, Weixin Luo, Wei Zhang, and Lin Ma. Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In AAAI, 2022. 2, 3", + "[12] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2", + "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3", + "[14] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020.", + "[15] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3", + "[16] Kyunghyun Cho, B van Merrienboer, Caglar Gulcehre, F Bougares, H Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP, 2014. 3" + ], + "bbox": [ + 86, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In ACMMM, 2022. 1", + "[18] Xuguang Duan, Wenbing Huang, Chuang Gan, Jingdong Wang, Wenwu Zhu, and Junzhou Huang. Weakly supervised dense event captioning in videos. In NeurIPS, 2018. 6, 7", + "[19] Yazan Abu Farha and Jurgen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 1", + "[20] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 6", + "[21] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020. 1", + "[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 1, 2, 6", + "[23] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022. 1", + "[24] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. arXiv preprint arXiv:2305.14344, 2023.3", + "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3, 6", + "[26] Jiabo Huang, Yang Liu, Shaogang Gong, and Hailin Jin. Cross-sentence temporal and semantic relations in video activity localisation. In ICCV, 2021. 2", + "[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In CVPR, 2020. 1", + "[28] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In CVPR, 2023. 2", + "[29] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, 2023. 2", + "[30] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, 2017. 1", + "[31] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, 2022. 3, 6, 7", + "[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[33] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Workshops, 2015. 3", + "[34] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 6" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13577", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In CVPR, 2022. 1", + "[36] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, 2020. 1", + "[37] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, 2017. 1", + "[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 1", + "[39] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 5", + "[40] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In ICCV, 2023. 2", + "[41] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 1", + "[42] Zhijie Lin, Zhou Zhao, Zhu Zhang, Qi Wang, and Huasheng Liu. Weakly-supervised video moment retrieval via semantic completion network. In AAAI, 2020. 2", + "[43] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In ACMMM, 2022. 2", + "[44] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In AAAI, 2022. 2", + "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 1", + "[46] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In ICLR, 2021. 4, 5", + "[47] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In EMNLP, 2019. 2", + "[48] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. TMM, 2022. 7", + "[49] Minuk Ma, Sunjae Yoon, Junyeong Kim, Youngjoon Lee, Sunghun Kang, and Chang D Yoo. Vlanet: Video-language alignment network for weakly-supervised video moment retrieval. In ECCV, 2020. 2", + "[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1", + "[51] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 5" + ], + "bbox": [ + 86, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Niluthpol Chowdhury Mithun, Sujoy Paul, and Amit K Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2", + "[53] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 2", + "[54] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. Bassl: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 3", + "[55] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 1", + "[56] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5, 6", + "[57] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 2013. 6", + "[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5", + "[59] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 1", + "[60] Mrigank Rochan, Linwei Ye, and Yang Wang. Video summarization using fully convolutional sequence networks. In ECCV, 2018. 1", + "[61] Marcus Rohrbach, Michaela Regneri, Mykhaylo Andriluka, Sikandar Amin, Manfred Pinkal, and Bernt Schiele. Script data for attribute-based recognition of composite activities. In ECCV, 2012. 6", + "[62] Fengyuan Shi, Limin Wang, and Weilin Huang. End-to-end dense video grounding via parallel regression. arXiv preprint arXiv:2109.11265, 2021. 3, 6, 7", + "[63] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 6", + "[64] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. Vlg-net: Video-language graph matching network for video grounding. In CVPR, 2021. 2", + "[65] Yijun Song, Jingwen Wang, Lin Ma, Zhou Yu, and Jun Yu. Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048, 2020. 2", + "[66] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In CVPR, 2014. 3", + "[67] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Wei-Shi Zheng, and Jianhuang Lai. Hierarchical semantic correspondence networks for video paragraph grounding. In CVPR, 2023. 3, 6, 7", + "[68] Reuben Tan, Huijuan Xu, Kate Saenko, and Bryan A Plummer. Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In WACV, 2021. 2" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13578", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In CVPR, 2023. 3", + "[70] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 3, 6", + "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4, 5", + "[72] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. In AAAI, 2021. 1", + "[73] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In CVPR, 2021. 2", + "[74] Jingwen Wang, Lin Ma, and Wenhao Jiang. Temporally grounding language queries in videos by contextual boundary-aware prediction. In AAAI, 2020. 2", + "[75] Yuechen Wang, Jiajun Deng, Wengang Zhou, and Houqiang Li. Weakly supervised temporal adjacent network for language grounding. TMM, 2022. 2, 6, 7", + "[76] Yunxiao Wang, Meng Liu, Yinwei Wei, Zhiyong Cheng, Yinglong Wang, and Liqiang Nie. Siamese alignment network for weakly supervised video moment retrieval. TMM, 2023.", + "[77] Zheng Wang, Jingjing Chen, and Yu-Gang Jiang. Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In ACMMM, 2021. 2", + "[78] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2", + "[79] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In AAAI, 2020. 2", + "[80] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In CVPR, 2023. 1", + "[81] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3", + "[82] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 1", + "[83] Shuwen Xiao, Zhou Zhao, Zijian Zhang, Xiaohui Yan, and Min Yang. Convolutional hierarchical attention network for query-focused video summarization. In AAAI, 2020. 1", + "[84] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. In EMNLP, 2021. 2", + "[85] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In AAAI, 2021. 2", + "[86] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In AAAI, 2019. 2" + ], + "bbox": [ + 86, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021. 3", + "[88] Wenfei Yang, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Local correspondence network for weakly supervised temporal sentence grounding. TIP, 2021. 2", + "[89] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 1", + "[90] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In NeurIPS, 2019. 2", + "[91] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In AAAI, 2019. 2, 6", + "[92] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In ACMMM Workshops, 2021. 6", + "[93] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, 2020. 2", + "[94] Chenchi Zhang, Wenbo Ma, Jun Xiao, Hanwang Zhang, Jian Shao, Yueting Zhuang, and Long Chen. Vl-nms: Breaking proposal bottlenecks in two-stage visual-language matching. TOMM, 2023. 1", + "[95] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, 2019. 2", + "[96] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 2", + "[97] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. TPAMI, 2021. 2", + "[98] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 5", + "[99] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 1", + "[100] Mingxing Zhang, Yang Yang, Xinghan Chen, Yanli Ji, Xing Xu, Jingjing Li, and Heng Tao Shen. Multi-stage aggregated transformer network for temporal language localization in videos. In CVPR, 2021. 2", + "[101] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 2, 6, 7", + "[102] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. TPAMI, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13579", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[103] Yimeng Zhang, Xin Chen, Jinghan Jia, Sijia Liu, and Ke Ding. Text-visual prompting for efficient 2d temporal video grounding. In CVPR, 2023. 2", + "[104] Zhu Zhang, Zhijie Lin, Zhou Zhao, Jieming Zhu, and Xiuqiang He. Regularized two-branch proposal networks for weakly-supervised moment retrieval in videos. In ACMMM, 2020. 2", + "[105] Zhu Zhang, Zhou Zhao, Zhijie Lin, Xiuqiang He, et al. Counterfactual contrastive learning for weakly-supervised vision-language grounding. In NeurIPS, 2020.", + "[106] Minghang Zheng, Yanjie Huang, Qingchao Chen, and Yang Liu. Weakly supervised video moment localization with contrastive negative sample mining. In AAAI, 2022. 2", + "[107] Minghang Zheng, Yanjie Huang, Qingchao Chen, Yuxin Peng, and Yang Liu. Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In CVPR, 2022. 2", + "[108] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2021. 3" + ], + "bbox": [ + 78, + 90, + 467, + 375 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13580", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_model.json b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8ff357632fbef2361edd3e3624f5cd8ce501b326 --- /dev/null +++ b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_model.json @@ -0,0 +1,2875 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.131, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.203, + 0.842, + 0.223 + ], + "angle": 0, + "content": "Chaolei Tan\\(^{1}\\) Jianhuang Lai\\(^{1,2,3}\\) Wei-Shi Zheng\\(^{1,2,3}\\) Jian-Fang Hu\\(^{1,2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.223, + 0.812, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.239, + 0.829, + 0.257 + ], + "angle": 0, + "content": "2Guangdong Province Key Laboratory of Information Security Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.257, + 0.892, + 0.275 + ], + "angle": 0, + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.278, + 0.911, + 0.293 + ], + "angle": 0, + "content": "tanchlei@mail2.sysu.edu.cn, stsljh@mail.sysu.edu.cn, wszheng@ieee.org, hujf5@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.301, + 0.314, + 0.316 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.474, + 0.741 + ], + "angle": 0, + "content": "Video Paragraph Grounding (VPG) is an emerging task in video-language understanding, which aims at localizing multiple sentences with semantic relations and temporal order from an untrimmed video. However, existing VPG approaches are heavily reliant on a considerable number of temporal labels that are laborious and time-consuming to acquire. In this work, we introduce and explore Weakly-Supervised Video Paragraph Grounding (WSVPG) to eliminate the need of temporal annotations. Different from previous weakly-supervised grounding frameworks based on multiple instance learning or reconstruction learning for two-stage candidate ranking, we propose a novel siamese learning framework that jointly learns the cross-modal feature alignment and temporal coordinate regression without timestamp labels to achieve concise one-stage localization for WSVPG. Specifically, we devise a Siamese Grounding Transformer (SiamGTR) consisting of two weight-sharing branches for learning complementary supervision. An Augmentation Branch is utilized for directly regressing the temporal boundaries of a complete paragraph within a pseudo video, and an Inference Branch is designed to capture the order-guided feature correspondence for localizing multiple sentences in a normal video. We demonstrate by extensive experiments that our paradigm has superior practicability and flexibility to achieve efficient weakly-supervised or semi-supervised learning, outperforming state-of-the-art methods trained with the same or stronger supervision." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.749, + 0.21, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Natural Language Video Grounding (NLVG) is an essential area in vision-language understanding, which has received increasing attention due to its wide range of real-world applications such as video retrieval [3, 17, 21, 23, 80, 94], video summarization [50, 55, 59, 60, 83, 99], action segmentation [19, 27, 35, 37, 45, 72], video question answering [30, 36, 38, 41, 82, 89], etc. Most previous works focus" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.301, + 0.891, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.39, + 0.706, + 0.4 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.401, + 0.89, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.518, + 0.893, + 0.561 + ], + "angle": 0, + "content": "Figure 1. (a) Chronological cross-modal alignment in a video and its paired sentences. (b) Pseudo boundary supervision for regressing paragraph timestamps in a composed pseudo video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.719 + ], + "angle": 0, + "content": "on tackling Video Sentence Grounding (VSG) proposed in [1, 22], which targets at localizing the temporal boundaries of an individual sentence from an untrimmed video. However, localizing single sentences can be ambiguous because the contextual information conveyed by multiple sentences is necessary to uniquely determine the temporal locations of input queries. To alleviate such issue, Bao et al. [4] proposed to contextualize video grounding into localizing multiple events indicated by sentences of a paragraph from the video, which is called Video Paragraph Grounding (VPG)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Although remarkable progress has been attained in tackling VSG and VPG problems under a fully-supervised setting, the extremely prohibitive overheads of manually annotating temporal boundaries for language queries limit these methods to utilize large-scale video-text data. In addition, the subjectivity of annotators inevitably brings label noises that may adversely affect model training. Recently, weakly-supervised methods free of temporal annotations have become increasingly popular in the area of video grounding, aiming to address the above limitations. These approaches can be mainly categorized into multiple instance learning methods and reconstruction learning methods and are typ-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13569" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "ically based on a propose-and-rank pipeline. Nevertheless, both of these paradigms assume the contributions of proposals to the contrastive or reconstruction loss accurately represent the proposal quality, which is not necessarily the case during model learning. Moreover, the quadratic complexity of proposal schemes prevents them from scaling up in parameters or training data, and the adopted supervision of video-level contrast or query-based reconstruction in prior works is not temporal-sensitive thus suffering from a huge gap with fully-supervised temporal guidance. Besides, all of the weakly-supervised methods in video grounding are tailored for tackling VSG while the weakly-supervised setting of VPG (i.e., WSVPG) has been understudied so far." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.563 + ], + "angle": 0, + "content": "To circumvent the aforementioned drawbacks and explore an efficient weakly-supervised framework for VPG, we seek to mine the unique characteristics and underlying supervision from the intrinsic structure of video-paragraph pairs for model training. On the one hand, as observed in Figure 1 (a), the temporal location of an event is highly correlated with the position of the sentence describing that event in the paragraph. For example, the sentence appearing in the middle of the paragraph tends to have stronger relevance with visual content located around the temporal midpoint of the video. On the other hand, dense visual events mentioned in the paragraph approximately represent the global video content that unambiguously distinguishes itself from another video, which can be observed in Figure 1 (b). Therefore, inserting the query-related video into another irrelevant video automatically generates pseudo boundary labels close to the ground-truth when regarding the complete paragraph as language query for video grounding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.473, + 0.852 + ], + "angle": 0, + "content": "Motivated by the above observation, we propose a novel Siamese Grounding TRansformer (SiamGTR) for WSVPG. It jointly learns the cross-modal alignment and boundary regression via two siamese branches without generating proposals. Specifically, we propose to construct an Augmentation Branch (AB) which takes as input a pseudo video and adopts a complete paragraph as the language query to learn high-quality boundary supervision for localization. Also, an Inference Branch (IB) is designed to receive a normal video as input and is enforced to capture the order-guided cross-modal correspondence for attending over the specific video content relevant to each sentence. The two weight-sharing branches are effective to transfer complementary supervision for joint boundary prediction and feature association, which yields a weakly-supervised model with superior generalization through a concise one-stage pipeline. Extensive experiments verify the effectiveness of our model and show our method with the same or weaker supervision surpasses prior state-of-the-arts. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- We introduce the task of Weakly-Supervised Video Paragraph Grounding (WSVPG), which aims to train a model for localizing multiple events indicated by queries with" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.779, + 0.107 + ], + "angle": 0, + "content": "out the supervision of timestamp labels." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.108, + 0.894, + 0.212 + ], + "angle": 0, + "content": "- We propose a novel Siamese Grounding TRansformer (SiamGTR) for concise and efficient one-stage weakly-supervised learning of video paragraph grounding. It is composed of two weight-sharing branches including an Augmentation Branch (AB) for learning boundary regression of pseudo boundaries and an Inference Branch (IB) for learning order-guided cross-modal feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.213, + 0.894, + 0.258 + ], + "angle": 0, + "content": "- Extensive experiments verify the efficacy of our method, and demonstrate that our framework under the same or weaker supervision outperforms state-of-the-arts." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.108, + 0.894, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.271, + 0.642, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.292, + 0.743, + 0.309 + ], + "angle": 0, + "content": "2.1. Video Sentence Grounding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Fully-Supervised Video Sentence Grounding. Plenty of approaches [1, 7, 12, 22, 29, 40, 43, 44, 47, 53, 64, 73, 74, 78, 79, 84-86, 90, 91, 93, 95-97, 100-103] have been proposed to address Fully-Supervised Video Sentence Grounding (FSVSG). In general, these works can be roughly categorized into proposal-based and proposal-free methods. Specifically, proposal-based methods [1, 22, 40, 43, 44, 73, 74, 78, 85, 86, 90, 93, 95, 101, 102] involve a proposal generation stage using sliding windows [22], anchor proposals [12, 43, 44, 47, 74, 90, 93, 95] or 2D temporal maps [64, 73, 78, 85, 101, 102], after which the generated proposals are ranked according to the query matching scores with potential post-processing like Non-Maximum Suppression (NMS). In contrast, proposal-free methods [7, 53, 79, 91, 96, 97, 100, 103] remove the dense proposal generation and score ranking process by directly regressing timestamps [7, 53, 84, 91, 103], predicting boundary distributions [96, 97, 100] or using reinforcement learning [79], which improves the computation efficiency and scenario adaptability. Following the line of proposal-free works, we propose a novel weakly-supervised regression-based framework that shows superior performance and practicability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Weakly-Supervised Video Sentence Grounding. Weakly-Supervised Video Sentence Grounding (WSVSG) [8, 11, 26, 28, 42, 49, 52, 65, 68, 75-77, 88, 104-107] has become a popular research area because of the severe dependence of FSVSG approaches on laborious and expensive manual temporal annotations. Most of existing WSVSG methods are based on a two-stage pipeline using multiple instance learning [26, 49, 75, 88], reconstruction learning [42], or the combination of both [106, 107]. In particular, Chen et al. [11] have proposed a video composition strategy to generate pseudo temporal labels for WSVSG, which is the most related work to ours. However, several inherent drawbacks are involved in this approach. Firstly, individual sentences only describe local video content, thus viewing the starting/ending locations of foreground video as temporal boundaries of an individual sentence produces a weak and noisy temporal alignment, which is unsuitable for accurate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13570" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "boundary supervision. Moreover, it simply adapts an existing fully-supervised proposal-based framework for weakly-supervised training, which leads to inferior generalization caused by the large train-test discrepancy. Distinct from all of the above works, we design a novel siamese framework to capture the essential characteristics of video paragraph grounding for weakly-supervised learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.207, + 0.336, + 0.224 + ], + "angle": 0, + "content": "2.2. Video Paragraph Grounding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.474, + 0.516 + ], + "angle": 0, + "content": "Video Paragraph Grounding (VPG) is introduced by Bao et al. [4], which aims to jointly localize multiple sentences of a paragraph from an untrimmed video. Shi et al. [62] presented an end-to-end network by re-purposing transformers into language-conditioned regressors. Jiang et al. [31] proposed to employ contrastive encoders for contrastive learning between video-paragraph pairs. Tan et al. [67] proposed a hierarchical semantic correspondence network for modeling hierarchical video-language alignment and grounding multiple levels of language queries in the video. Particularly, Jiang et al. [31] first explored Semi-Supervised Video Paragraph Grounding (SSVPG) to relieve the annotation burden of temporal labels. However, SSVPG is still not quite practical considering the expensive cost of temporal annotations in untrimmed videos. Besides, these semi-supervised methods still require a considerable proportion of temporal labels up to at least \\(10\\%\\) [31] for training. To thoroughly get rid of the temporal annotations, we pioneer to explore the weakly-supervised setting in VPG." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.525, + 0.254, + 0.539 + ], + "angle": 0, + "content": "2.3. Siamese Networks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.472, + 0.759 + ], + "angle": 0, + "content": "Siamese networks [6] are weight-sharing neural networks. There have been a wide range of scenarios where siamese networks are applied for achieving different purposes, such as face verification [66], image recognition [33], object tracking [5], etc. In particular, siamese networks are commonly used in contrastive self-supervised learning methods [2, 10, 13-15, 24, 25, 69, 81, 108], in which augmented views of the same or different instance are forwarded into multiple weight-sharing network copies for learning a generalizable visual representation via instance-level discrimination. In this work, we explore a new way to combine the transferability of siamese architectures and the flexibility of transformer architectures for concise and efficient weakly-supervised learning of video paragraph grounding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.775, + 0.212, + 0.792 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.188, + 0.815 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Task Formulation. Given an untrimmed video \\(\\mathcal{V}\\) and a paragraph query \\(\\mathcal{P}\\) consisting of \\(N\\) temporally ordered sentences \\(\\{\\mathcal{S}_i\\}_{i=1}^N\\), the goal of Video Paragraph Grounding (VPG) is to simultaneously localize the temporal intervals \\(\\mathcal{T} = \\{(\\tau_i^{\\mathrm{st}}, \\tau_i^{\\mathrm{ed}})\\}_{i=1}^N\\) of all the events described by sentences" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.089, + 0.71, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.089, + 0.84, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.213, + 0.88, + 0.227 + ], + "angle": 0, + "content": "Figure 2. Our siamese framework for joint alignment and regression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.262 + ], + "angle": 0, + "content": "in the paragraph, where \\(\\tau_{i}^{\\mathrm{st}}\\) and \\(\\tau_{i}^{\\mathrm{ed}}\\) are the starting and ending timestamps for the \\(i\\)-th sentence \\(S_{i}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.893, + 0.458 + ], + "angle": 0, + "content": "Siamese Learning. An overview of our siamese framework is shown in Figure 2. Overall, we jointly train an augmentation branch and an inference branch with shared parameters and structures for learning the complementary abilities of cross-modal feature alignment and temporal boundary regression. These two branches follow the same workflow to first encode the text queries and input video into unimodal features, after which the query features are iteratively used in the transformer decoder to extract relevant information from the video features for timestamp decoding. For testing, we only keep the pipeline of inference branch for boundary prediction. More architectural details are illustrated in Figure 3 and are elaborated in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.682, + 0.483 + ], + "angle": 0, + "content": "3.2. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Video Feature Extraction. For each input video, we divide it into consecutive clips consisting of a fixed number of frames for feature extraction. Specifically, a frozen pretrained 3D Convolutional Neural Network (3D-CNN) [70] and a linear projection layer are successively employed to obtain a 1D feature vector for each short clip, resulting in a video feature sequence \\(\\mathcal{F}_{\\mathrm{v}} \\in \\mathbb{R}^{L \\times D}\\), where \\(L\\) and \\(D\\) are the sequence length and hidden dimension, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Text Feature Extraction. For each input paragraph consisting of \\(N\\) sentences, we first utilize a frozen pre-trained word embedding model tookenize and embed the text into a sequence of word vectors. Then, a bidirectional Gated Recurrent Unit (GRU) [16] is employed on each sentence, and the last hidden states in both directions are concatenated and then projected by a linear layer to construct the sentence features \\(\\mathcal{F}_{\\mathrm{s}}\\in \\mathbb{R}^{N\\times D}\\), where \\(D\\) is the hidden dimension." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.707, + 0.758 + ], + "angle": 0, + "content": "3.3. Augmentation Branch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.825 + ], + "angle": 0, + "content": "The augmentation branch aims to learn accurate boundary regression from pseudo videos with paragraph queries, which naturally transfers to the inference branch via the shared feature space established by the siamese structure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pseudo Data Generation. To drive end-to-end weakly-supervised regression learning, the input stream of the augmentation branch should provide reliable and direct boundary supervision. Drawing inspiration from the boundary-sensitive video pretext tasks [11, 54, 87], we propose to uti" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed Siamese Grounding TRansformer (SiamGTR) architecture. The augmentation branch (abbreviated as A.B.) takes the pseudo video features derived from randomly inserting the query-related video features into irrelevant video features. It learns to temporally regress the interval of interest from the pseudo video with the paragraph as query. The Inference Branch (abbreviated as I.B.) receives normal video features for learning the cross-modal feature alignment among multiple sentences in the video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.517 + ], + "angle": 0, + "content": "lize the synthesized videos paired with paragraph queries to serve as a well-suited source of surrogate boundary supervision. Specifically, for each input video \\(\\nu\\) and its paragraph description \\(\\mathcal{P}\\), we randomly sample a background video \\(\\nu_{\\mathrm{bg}}\\) to obtain irrelevant video content to \\(\\mathcal{P}\\). Because \\(\\mathcal{P}\\) could be treated as a unique referring expression specific to \\(\\nu\\), the search space of background videos can be the entire training set for increasing the data diversity. Denote video features of \\(\\nu\\) and \\(\\nu_{\\mathrm{bg}}\\) as \\(\\mathcal{F}_{\\mathrm{v}}\\) and \\(\\mathcal{F}_{\\mathrm{bg}}\\) respectively, we construct the pseudo video features \\(\\mathcal{F}_{\\mathrm{aug}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.522, + 0.47, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathrm {a u g}} = \\operatorname {N R S} \\left(\\operatorname {C o n c a t} \\left(\\mathcal {F} _ {\\mathrm {b g}} ^ {1: I}, \\operatorname {R R S} \\left(\\mathcal {F} _ {\\mathrm {v}}\\right), \\mathcal {F} _ {\\mathrm {b g}} ^ {I + 1:}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.473, + 0.676 + ], + "angle": 0, + "content": "where \\(I\\) is a randomly generated inserting index within the feature sequence \\(\\mathcal{F}_{\\mathrm{bg}}\\) and Concat (\\(\\cdot\\)) denotes the temporal concatenation. RRS (\\(\\cdot\\)) and NRS (\\(\\cdot\\)) respectively represent a Random Re-Sampling operation to stochastically re-scale the length of \\(\\mathcal{F}_{\\mathrm{v}}\\) and a Normalized Re-Sampling operation that converts the length of \\(\\mathcal{F}_{\\mathrm{aug}}\\) into a fixed number of \\(T\\). Thereafter, we compute the synthesized temporal boundaries (\\(\\tau_{\\mathrm{aug}}^{\\mathrm{st}}, \\tau_{\\mathrm{aug}}^{\\mathrm{ed}}\\)) in the pseudo video \\(\\mathcal{V}_{\\mathrm{aug}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.681, + 0.47, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\tau_ {\\mathrm {a u g}} ^ {\\mathrm {s t}} = \\frac {I + \\Delta I ^ {\\mathrm {s t}}}{r L + L _ {\\mathrm {b g}}}, \\tau_ {\\mathrm {a u g}} ^ {\\mathrm {e d}} = \\frac {I + r L - \\Delta I ^ {\\mathrm {e d}}}{r L + L _ {\\mathrm {b g}}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.825 + ], + "angle": 0, + "content": "where \\(L\\) and \\(L_{\\mathrm{bg}}\\) respectively indicate the length of \\(\\mathcal{F}_{\\mathrm{v}}\\) and \\(\\mathcal{F}_{\\mathrm{bg}}\\) with \\(r\\) being the random re-scaling factor of RRS \\((\\cdot)\\). To alleviate potential synthesis artifacts and boundary uncertainty, we further introduce a Random Boundary Shifting (RBS) strategy that incorporates small random offsets \\(\\Delta I^{\\mathrm{st}}\\) and \\(\\Delta I^{\\mathrm{ed}}\\) into computing the pseudo labels, which is simple yet effective to boost the quality of boundary supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Video Encoder. Based on the obtained pseudo video features \\(\\mathcal{F}_{\\mathrm{aug}}\\), we then encode the temporal contextual information across multiple clips by feeding \\(\\mathcal{F}_{\\mathrm{aug}}\\) to a video encoder, which follows a similar architecture of DETR encoders [9, 46, 71], i.e., each encoder layer consists of a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.44 + ], + "angle": 0, + "content": "multi-head self-attention module equipped with additive sinusoidal positional encodings and a feed-forward network. For simplicity, we omit the layer index and denote the input of each video encoder layer as \\(\\mathcal{X}_{\\mathrm{aug}}\\), then a set of Modulated Positional Encodings (MPE) are given as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.445, + 0.892, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {\\mathrm {p e}} = \\operatorname {M L P} \\left(\\mathcal {X} _ {\\mathrm {a u g}}\\right) \\odot \\operatorname {P E} \\left(\\mathcal {X} _ {\\mathrm {a u g}}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.893, + 0.602 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{\\mathrm{pe}} \\in \\mathbb{R}^{T \\times D}\\) and \\(\\odot\\) denotes element-wise multiplication. PE \\((\\cdot)\\) is the sinusoidal function [71] with respect to temporal locations of \\(\\mathcal{X}_{\\mathrm{aug}}\\), MLP \\((\\cdot)\\) denotes a two-layer feed-forward network computed on \\(\\mathcal{X}_{\\mathrm{aug}}\\). In each encoder layer, \\(\\mathcal{X}_{\\mathrm{pe}}\\) is only added to the input of projection layers of queries and keys in the self-attention mechanism. The encoded video features of the last video encoder layer, also called the encoder memory features \\(\\mathcal{M}_{\\mathrm{aug}}\\), are further fed to the query-guided decoder for localization prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.722 + ], + "angle": 0, + "content": "Query Encoder. Since there are strong contextual correlations among multiple sentences in the paragraph query, here we utilize a vanilla transformer encoder [71] to help reason the semantic and chronological relationships of events. To extract a global representation of the complete paragraph, we initialize an extra learnable query token \\(\\mathcal{F}_{\\mathrm{p}}\\) and integrate it with the sentence features \\(\\mathcal{F}_{\\mathrm{s}}\\) by concatenation. Thus, the input of the query encoder is constructed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.725, + 0.891, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {\\mathrm {q}} = \\left[ \\mathcal {F} _ {\\mathrm {p}}, \\mathcal {F} _ {\\mathrm {s}} ^ {1}, \\mathcal {F} _ {\\mathrm {s}} ^ {2}, \\dots , \\mathcal {F} _ {\\mathrm {s}} ^ {i}, \\dots \\mathcal {F} _ {\\mathrm {s}} ^ {N} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.871 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{\\mathrm{q}} \\in \\mathbb{R}^{(N + 1) \\times D}\\) and \\(\\mathcal{F}_{\\mathrm{s}}^{i}\\) is the \\(i\\)-th sentence feature in \\(\\mathcal{F}_{\\mathrm{s}}\\). We first normalize \\(\\mathcal{X}_{\\mathrm{q}}\\), add fixed sinusoidal positional encodings to it, and then iteratively employ self-attention modules and feed-forward networks to contextualize the global and local tokens that respectively represent the query semantics of the entire paragraph and the sentences. The output features of the last query encoder layer are denoted as \\(\\mathcal{Z}_{\\mathrm{q}}\\) and are further forwarded to the query decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Conceptual Semantic Connector. Although the siamese network structure can implicitly transfer boundary knowl-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13572" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "edge from the augmentation branch to the inference branch, there is still a certain semantic gap between the query representations of short sentences and long paragraphs. To narrow this gap, we develop a Conceptual Semantic Connector (CSC) module for explicit semantic guidance. Specifically, we first collect high-frequency linguistic concepts, including verbs and nouns from the training corpus, and then construct a set of dictionary features by selecting and projecting the Glove vectors [56]. The loss \\(\\mathcal{L}_{\\mathrm{csc}}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.228, + 0.47, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c s c}} = \\mathrm {B C E} \\left(y _ {\\text {c p t}} ^ {\\mathrm {p}}, \\hat {y} _ {\\text {c e p t}} ^ {\\mathrm {p}}\\right) + \\mathrm {B C E} \\left(y _ {\\text {s c t}} ^ {\\mathrm {s}}, \\hat {y} _ {\\text {c e p t}} ^ {\\mathrm {s}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.47, + 0.337 + ], + "angle": 0, + "content": "where \\(\\mathrm{BCE}(\\cdot)\\) is the binary cross-entropy loss. \\(y_{\\mathrm{cept}}^{\\mathrm{p}}\\) and \\(y_{\\mathrm{cept}}^{\\mathrm{s}}\\) are multi-hot labels indicating semantic concepts contained by the paragraph and sentence queries, respectively. \\(\\hat{y}_{\\mathrm{cept}}^{\\mathrm{s}}\\) and \\(\\hat{y}_{\\mathrm{cept}}^{\\mathrm{p}}\\) are the concept predictions obtained by dot-product between the conceptual dictionary features and the textual query features \\(\\mathcal{Z}_{\\mathrm{q}}\\) with a sigmoid activation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.636 + ], + "angle": 0, + "content": "Query Decoder. Inspired by the spiritual ideas of dynamic-anchor DETR decoders [39, 46, 51, 98], we design a novel query decoder that enables dynamic position-aware decoding of language queries. Specifically, each query decoder layer consists of a self-attention module, a cross-attention module, and a feed-forward network, with dynamically adjustable anchor boxes to indicate the query-specific location information. Initially, the input anchor boxes are set to all zeros, and an MLP is used to estimate a set of seed anchor boxes based on the cross-modal interactions between \\(\\mathcal{M}_{\\mathrm{aug}}\\) and \\(\\mathcal{Z}_{\\mathrm{q}}\\), i.e., the output features of the first decoder layer \\(\\mathcal{Q}_{\\mathrm{aug}}^{(1)} \\in \\mathbb{R}^{(N + 1) \\times D}\\) are mapped into the seed anchor boxes \\(\\mathcal{A}_{\\mathrm{aug}}^{(1)} \\in \\mathbb{R}^{(N + 1) \\times 2}\\) by the MLP. For the \\((i + 1)\\)-th decoder layer, we first convert the box coordinates of the input anchors \\(\\mathcal{A}_{\\mathrm{aug}}^{(i)}\\) into high-dimensional sinusoidal embeddings \\(\\mathcal{F}_{\\mathrm{a}}^{(i)} \\in \\mathbb{R}^{(N + 1) \\times D}\\) by a sinusoidal function [46, 71] and further obtain \\(\\mathcal{H}_{\\mathrm{a}}^{(i)}\\) by projecting \\(\\mathcal{F}_{\\mathrm{a}}^{(i)}\\) with an MLP. Afterwards, we conduct the self-attention operation and update the query features in the decoder as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.636, + 0.47, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)} \\leftarrow \\text {S e l f - A t t n} \\left\\{ \\begin{array}{l} Q = \\varphi_ {q} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) + \\varphi_ {q} ^ {p} \\left(\\mathcal {H} _ {\\mathrm {a}} ^ {(i)}\\right) \\\\ K = \\varphi_ {k} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) + \\varphi_ {k} ^ {p} \\left(\\mathcal {H} _ {\\mathrm {a}} ^ {(i)}\\right) \\\\ V = \\varphi_ {v} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.818 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}\\) is the updated query features after the self-attention operation in the \\((i + 1)\\)-th decoder layer. The above series of \\(\\varphi\\) functions are used to indicate different linear projection layers for the content part or position part of the queries, keys, or values in the self-attention mechanism. Then, we conduct a cross-attention operation to extract useful cross-modal interactive information as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.83, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)} \\leftarrow \\text {C r o s s - A t t n} \\left\\{ \\begin{array}{l} Q = \\left[ \\varphi_ {q} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)}\\right); \\varphi_ {q} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {a}} ^ {(i)}\\right) \\right] \\\\ K = \\left[ \\varphi_ {k} ^ {c} \\left(\\mathcal {M} _ {\\mathrm {a u g}}\\right); \\varphi_ {k} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {p e}} ^ {\\mathcal {M}}\\right) \\right] \\\\ V = \\varphi_ {v} ^ {c} \\left(\\mathcal {M} _ {\\mathrm {a u g}}\\right) + \\varphi_ {v} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {p e}} ^ {\\mathcal {M}}\\right) \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.892, + 0.224 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{pe}}^{\\mathcal{M}} = \\mathrm{PE}\\left(\\mathcal{M}_{\\mathrm{aug}}\\right)\\) and the query features \\(\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}\\) is then further updated by a feed-forward network as the feature output of the \\((i + 1)\\)-th decoder layer. Then the anchor boxes are dynamically updated as \\(\\mathcal{A}_{\\mathrm{aug}}^{(i + 1)} \\gets \\mathcal{A}_{\\mathrm{aug}}^{(i)} + \\Delta \\mathcal{A}_{\\mathrm{aug}}^{(i)}\\) where we utilize an MLP layer to predict the relative offsets, i.e., \\(\\Delta \\mathcal{A}_{\\mathrm{aug}}^{(i)} \\in \\mathbb{R}^{(N + 1) \\times 2}\\), based on the updated query features \\(\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}\\). \\(\\mathcal{A}_{\\mathrm{aug}}^{(i + 1)}\\) continues to be forwarded to the next decoder layer for computing \\(\\mathcal{F}_{\\mathrm{a}}^{(i + 1)}\\) and \\(\\mathcal{H}_{\\mathrm{a}}^{(i + 1)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Boundary Prediction. Based on the output features and attention weights of the last query decoder layer, we simply use an MLP predictor to predict the paragraph timestamps, i.e., \\(\\hat{\\mathcal{T}}_{\\mathrm{aug}}^{\\mathrm{p}} = (\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{ed}})\\). Similarly, the sentence timestamps \\(\\widehat{\\mathcal{T}}_{\\mathrm{inf}}^{\\mathrm{s}} = \\left\\{(\\hat{\\tau}_{\\mathrm{j}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{j}}^{\\mathrm{ed}})\\right\\}_{j = 1}^{N}\\) can also be obtained by feeding last-layer output from the inference branch to the same MLP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Self-Consistent Boundary Regression. We improve the attention-agnostic regression loss \\(\\mathcal{L}_{\\mathrm{reg}}\\) to make it aware of the model's self-consistent scores, where the main idea is to selectively optimize the regression loss of self-consistent samples for better weakly-supervised regression learning. Self-consistent samples have high attention weights over the pseudo ground-truth intervals, which are more suitable for learning less noisy supervision for accurate boundary prediction. Specifically, the self-consistent boundary regression loss \\(\\mathcal{L}_{\\mathrm{sreg}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.488, + 0.891, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s c r e g}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\mathrm {L} 1} + \\mathcal {L} _ {\\mathrm {G I o U}}, & \\text {i f} \\quad s _ {\\mathrm {a t t}} > \\beta \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.597 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{L1}}\\) and \\(\\mathcal{L}_{\\mathrm{GIoU}}\\) respectively represent L1 and Generalized Intersection over Union (GIoU) [58] loss computed between \\((\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{ed}})\\) and \\((\\tau_{\\mathrm{aug}}^{\\mathrm{st}},\\tau_{\\mathrm{aug}}^{\\mathrm{ed}})\\). \\(\\beta\\) is set to 0.5 and \\(s_{\\mathrm{att}}\\) is the attention sum over the pseudo ground-truth interval." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.671, + 0.621 + ], + "angle": 0, + "content": "3.4. Inference Branch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.81 + ], + "angle": 0, + "content": "In our siamese framework, the inference branch shares the same parameter weights and network structure with the augmentation branch, i.e., a video encoder, a query encoder, and a query decoder. The only difference lies in the input streams and objectives, i.e., the inference branch receives a normal video during training to learn in-domain cross-modal correspondence that cannot be acquired through the pseudo video stream, which significantly improves the generalization ability of the model. Specifically, it receives the encoded normal video features \\(\\mathcal{M}_{\\mathrm{inf}}\\) and the encoded text query features \\(\\mathcal{Z}_{\\mathrm{q}}\\) as the decoder input and generates the hidden query features \\(\\mathcal{Q}_{\\mathrm{inf}}\\) for boundary prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Order-guided Attention Loss. The chronological prior given by the sentence order provides explicit guidance for learning cross-modal alignment between the video and language features during decoding. To learn the order-guided cross-modal correspondence, we constrain cross-modal attention weights in the decoder as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13573" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.096, + 0.089, + 0.47, + 0.126 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o g a}} = \\max \\left(0, \\Delta m T + \\sum_ {t = 1} ^ {T} t \\alpha_ {\\mathrm {s}} ^ {j} (t) - \\sum_ {t = 1} ^ {T} t \\alpha_ {\\mathrm {s}} ^ {j + 1} (t)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.468, + 0.219 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\mathrm{s}}^{j}(t)\\) and \\(\\alpha_{\\mathrm{s}}^{j + 1}(t)\\) are attention weights over video features for the \\(j\\)-th and \\((j + 1)\\)-th sentence from the last decoder layer, respectively. \\(\\Delta m\\) is the minimal distance between attention centroids. This loss is mainly contributed by the inference branch with \\(\\Delta m = \\frac{1}{2N}\\) and partly contributed by the augmentation branch with \\(\\Delta m = \\frac{1}{4N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.219, + 0.468, + 0.399 + ], + "angle": 0, + "content": "Auxiliary Losses. To exploit more guidance for weakly-supervised representation learning, we employ three auxiliary losses including a cross-branch loss \\(\\mathcal{L}_{\\mathrm{cb}}\\), an anchor ranking loss \\(\\mathcal{L}_{\\mathrm{ar}}\\) and a pseudo attention loss \\(\\mathcal{L}_{\\mathrm{pa}}\\). Specifically, \\(\\mathcal{L}_{\\mathrm{cb}}\\) utilizes the semantic consistency constraint of output features across siamese branches, which is calculated analogous to MoCo [25]. \\(\\mathcal{L}_{\\mathrm{ar}}\\) is used for inducing the decoder to learn a set of order-preserving anchor boxes, which is computed on the anchor boxes like in the equation (9). \\(\\mathcal{L}_{\\mathrm{pa}}\\) makes use of attention supervision between the pseudo video and language features from the augmentation branch, which follows the calculation proposed in [62, 91]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.408, + 0.292, + 0.424 + ], + "angle": 0, + "content": "3.5. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.468, + 0.537 + ], + "angle": 0, + "content": "Weakly-Supervised Learning. The weakly-supervised loss of our proposed framework can be formulated as a weighted sum of the two losses from the siamese branches as \\(\\mathcal{L}_{\\mathrm{WS}} = \\lambda_{\\mathrm{sreg}}\\mathcal{L}_{\\mathrm{sreg}} + \\lambda_{\\mathrm{oga}}\\mathcal{L}_{\\mathrm{oga}}\\), where \\(\\lambda_{\\mathrm{sreg}}\\) and \\(\\lambda_{\\mathrm{oga}}\\) are scalar weights to balance the contributions of the two different losses. The overall loss for the weakly-supervised model is defined as \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{WS}} + \\lambda_{\\mathrm{csc}}\\mathcal{L}_{\\mathrm{csc}} + \\mathcal{L}_{\\mathrm{aux}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.468, + 0.687 + ], + "angle": 0, + "content": "Semi-Supervised Learning. Although our framework is initially designed for weakly-supervised learning, it can be easily adapted for end-to-end semi-supervised learning. Specifically, in addition to \\(\\mathcal{L}_{\\mathrm{WS}}\\) which is calculated on all training samples, we only need to employ an extra fully-supervised loss \\(\\mathcal{L}_{\\mathrm{FS}}\\) on those fully-annotated samples without changing any part of the network structure. The semi-supervised loss \\(\\mathcal{L}_{\\mathrm{SS}}\\) is defined as \\(\\mathcal{L}_{\\mathrm{SS}} = \\mathcal{L}_{\\mathrm{WS}} + \\mathcal{L}_{\\mathrm{FS}}\\), where \\(\\mathcal{L}_{\\mathrm{FS}}\\) consists of a regression loss and an attention loss and is calculated on the labeled samples in the inference branch." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.688, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Model Inference. As mentioned, the augmentation branch and conceptual semantic connector are discarded, while other inference branch modules are preserved for testing." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.201, + 0.763 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.275, + 0.786 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "ActivityNet-Captions. ActivityNet-Captions [34] dataset is a large-scale dataset with diverse open-domain content sourced from ActivityNet dataset [34]. There are 14,926 videos and 19,811 localized video-paragraph pairs in total. Each video lasts for 117.60 seconds and each paragraph consists of 3.63 sentences on average. The entire dataset is divided into train/val_1/val_2 sets containing" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "10,009/4,917/4,885 video-paragraph pairs, respectively. We follow prior works [4, 31] to use val_2 set for testing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Charades-CD-OOD. Charades-STA dataset [22] is built from the Charades dataset [63] with indoor activities. Following the previous work [31], we adopt a reorganized version of Charades-STA named Charades-CD-OOD proposed in [92]. It is divided into train/val/test_ood sets consisting of 4,564/333/1,440 video-paragraph pairs, respectively. Specifically, the average video duration is 30.78 seconds and the average paragraph length is 2.41 sentences." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.364 + ], + "angle": 0, + "content": "TACoS. TACoS dataset [57] is constructed from the MPII corpus [61] tailored for cooking activities and kitchen scenarios. There are 127 videos in total with each video paired with multiple paragraphs at different granularities. Concretely, there are 1,107,418, and 380 video-paragraph pairs for training, validation, and testing, respectively. The average video length and number of sentences in the paragraph are 4.79 minutes and 8.75 in this dataset, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous works [4, 31], we adopt mean Intersection over Union (i.e., mIoU) and recall under IoU threshold of \\(m\\) (i.e., \\(\\mathbf{R}@\\mathbf{m}\\)) as our evaluation metrics. The metrics are averaged over all sentences and \\(m\\) is set to be \\(\\{0.3, 0.5, 0.7\\}\\) for ActivityNet-Captions and Charades-CD-OOD, and \\(\\{0.1, 0.3, 0.5\\}\\) for TACoS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.716, + 0.481 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.671 + ], + "angle": 0, + "content": "For fair comparison with existing works [4, 62], we adopt the same C3D network [70] and Glove model [56] as feature extractors. The number of sampled video clips \\( T \\) is set to be 256, 128, and 512 for ActivityNet-Captions, Charades-CD-OOD, and TACoS datasets, respectively. We train the model using Adam [32] optimizer with a fixed learning rate of 0.0001 and a batch size of 32, 32, and 16 for ActivityNet-Captions, Charades-CD-OOD and TACoS, respectively. We select top-100 high-frequency concepts for each dataset, and the loss weights \\( \\{\\lambda_{\\mathrm{screg}}, \\lambda_{\\mathrm{oga}}, \\lambda_{\\mathrm{csc}}\\} \\) are set to \\( \\{2, 1, 10\\} \\). The number of encoder and decoder layers is set to be 3, and the hidden size \\( D \\) is 256 in all settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.8, + 0.697 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare the proposed SiamGTR with existing state-of-the-art methods for VPG to demonstrate the superiority of our framework. Specifically, 3D-TPN [4, 101], DepNet [4], PRVG [62], SVPTR [31] and HSCNet [67] are fully-supervised approaches requiring temporal annotations for the entire dataset. Besides, the semi-supervised setting has been studied in [31] with several methods developed. For fair comparison with our method, we regard the reconstruction learning method WSSL [18] as one baseline and further develop a more competitive model called Weakly-Supervised Temporal Paragraph Network (WSTPN) by incorporating Beam Search [20] into WSTAN [75]. Specifically, WSTPN utilizes a complete paragraph for multiple" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13574" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.09, + 0.435, + 0.103 + ], + "angle": 0, + "content": "Table 1. Comparison on ActivityNet-Captions dataset." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.105, + 0.468, + 0.29 + ], + "angle": 0, + "content": "
MethodSettingR@0.3R@0.5R@0.7mIoU
3D-TPN [101]FS67.5651.4930.92-
DepNet [4]FS72.8155.9133.46-
PRVG [62]FS78.2761.1537.8355.62
SVPTR [31]FS78.0761.7038.3655.91
HSCNet [67]FS81.8966.5744.0359.71
DepNet [4]SS61.4645.1426.7844.11
VPTR [31]SS72.8053.1429.0750.08
SVPTR [31]SS73.3956.7232.7851.98
SiamGTR (Ours)SS78.7559.1134.1254.57
WSSL [18]WS41.9823.34-28.33
WSTPN [75]WS57.7433.0213.6238.54
SiamGTR (Ours)WS75.4357.2330.5652.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.293, + 0.43, + 0.307 + ], + "angle": 0, + "content": "Table 2. Comparison on Charades-CD-OOD dataset." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.309, + 0.468, + 0.482 + ], + "angle": 0, + "content": "
MethodSettingR@0.3R@0.5R@0.7mIoU
DepNet [4]FS45.6127.5910.6929.30
STLG [48]FS48.3030.399.79-
SVPTR [31]FS55.1432.4415.5336.01
DepNet [4]SS43.0325.0710.1428.09
STLG [48]SS46.1529.439.38-
VPTR [31]SS45.1324.9810.2228.92
SVPTR [31]SS50.3128.5012.2732.13
SiamGTR (Ours)SS59.0735.4714.9538.87
WSSL [18]WS35.8623.678.27-
WSTPN [75]WS48.6129.2710.7933.49
SiamGTR (Ours)WS57.3333.8712.3137.21
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.469, + 0.531 + ], + "angle": 0, + "content": "instance learning and searches the best sequence of timestamps with the highest overall confidence while maintaining a consistent temporal order with the input sentences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.469, + 0.819 + ], + "angle": 0, + "content": "Comprehensive results over three different datasets are shown in Table 1, Table 2, and Table 3, where FS/SS/WS are used to indicate fully/semi/weakly-supervised settings of video paragraph grounding, respectively. First of all, our SiamGTR remarkably surpasses all the other methods under the same supervision in all metrics over the three datasets. Concretely, our framework outperforms WSTPN by \\(13.78\\%\\), \\(3.72\\%\\), and \\(11.16\\%\\) in mIoU on ActivityNetCaptions, Charades-CD-OOD and TACoS datasets, respectively. Compared to semi-supervised methods using a considerable number of temporal labels, our weakly-supervised method is also able to achieve comparable or even better results, which demonstrates the effectiveness of our framework in efficient weakly-supervised learning. Furthermore, our framework is flexible and can be easily adapted to semi-supervised learning for further gains. As shown, our semi-supervised model outstrips all semi-supervised state-of-the-art's by a large margin, and it performs better or on par with the fully-supervised SVPTR on all three datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.231, + 0.847 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.469, + 0.884 + ], + "angle": 0, + "content": "We conduct ablation studies to investigate the contributions of different components on ActivityNet-Captions dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of data augmentation. To evaluate the in" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.09, + 0.816, + 0.103 + ], + "angle": 0, + "content": "Table 3. Comparison on TACoS dataset." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.105, + 0.892, + 0.279 + ], + "angle": 0, + "content": "
MethodSettingR@0.1R@0.3R@0.5mIoU
3D-TPN [101]FS55.0540.3126.54-
DepNet [4]FS56.1041.3427.16-
PRVG [62]FS61.6445.4026.3729.18
SVPTR [31]FS67.9147.8928.2231.42
HSCNet [67]FS76.2859.7442.0040.61
DepNet [4]SS40.2726.9516.5418.68
VPTR [31]SS61.3140.5921.3926.59
SVPTR [31]SS63.0640.1920.0526.10
SiamGTR (Ours)SS67.3049.3531.6932.81
WSTPN [75]WS28.5910.044.769.32
SiamGTR (Ours)WS61.5126.2210.5320.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.284, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Table 4. Ablation studies on component designs of our framework. Experimental results are marked from ID (a) \\(\\sim\\) (l). RBS and RRS respectively denote the random boundary shifting and random resampling operations for pseudo data generation. MPE, CSC and DAB stand for the modulated positional encodings, conceptual semantic connector, and dynamic anchor boxes, respectively." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.37, + 0.891, + 0.558 + ], + "angle": 0, + "content": "
IDRBSRRSMPECSCDABR@0.5mIoU
(a)✓✓✓47.1046.04
(b)✓✓✓✓48.1946.96
(c)✓✓✓✓54.2450.94
(d)✓✓✓✓✓57.2352.32
(e)✓✓45.2544.46
(f)✓✓✓46.4745.47
(g)✓✓✓46.9645.23
(h)✓✓✓51.3448.52
(i)✓✓✓✓52.1248.75
(j)✓✓✓✓53.5750.06
(k)✓✓✓✓54.0950.84
(l)✓✓✓✓✓57.2352.32
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.672 + ], + "angle": 0, + "content": "fluences of the random boundary shifting and random resampling operations for data augmentation, we remove one or both of them from the training pipeline and the results are shown in Table 4 (a) \\(\\sim\\) (d). The model performance clearly degrades after the removal, which demonstrates the necessity of increasing sample diversity and alleviating overfitting for weakly-supervised cross-modal regression learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Ablation on module designs. As shown in Table 4 (e) \\(\\sim\\) (l), we conduct detailed ablation studies on module designs to validate the rationality of the proposed model. As observed, the designs of modulated positional encodings and dynamic anchor boxes in the encoder and decoder are consistently beneficial to improving the model capacity for better performance. The conceptual semantic connector that bridges two types of queries is also effective, and it further boosts the performance by \\(2.26\\%\\) in mIoU (50.06% vs. 52.32%) even though the network has been equipped with strong encoders and decoders. We notice the dynamic anchors for decoding are the most crucial component given the sharpest performance drop of \\(3.57\\%\\) with its removal, which indicates the importance to explicitly represent the intermediate location information for video grounding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13575" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.09, + 0.462, + 0.103 + ], + "angle": 0, + "content": "Table 5. Ablation studies on different weakly-supervised losses." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.106, + 0.434, + 0.178 + ], + "angle": 0, + "content": "
ID\\( {\\mathcal{L}}_{\\text{screg }} \\)\\( {\\mathcal{L}}_{\\text{oga }} \\)R@0.3R@0.5R@0.7mIoU
(a)45.8529.8210.9330.97
(b)✓46.9029.3812.0631.82
(c)✓73.5854.5828.8250.62
(d)✓✓75.4357.2330.5652.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.142, + 0.182, + 0.404, + 0.195 + ], + "angle": 0, + "content": "Table 6. Impact of different auxiliary losses." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.197, + 0.419, + 0.273 + ], + "angle": 0, + "content": "
MethodR@0.3R@0.5R@0.7mIoU
w/o Lcb71.4949.7824.6448.07
w/o Lar72.9650.7226.4249.14
w/o Lpa74.3456.0130.2951.76
Full Model75.4357.2330.5652.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.277, + 0.468, + 0.291 + ], + "angle": 0, + "content": "Table 7. Evaluation on different types of paragraph representation." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.292, + 0.426, + 0.368 + ], + "angle": 0, + "content": "
MethodR@0.3R@0.5R@0.7mIoU
Max-Pooling72.6152.6127.0349.55
Mean-Pooling74.3153.1328.6450.57
Word Concat73.7952.2727.3049.96
Learnable75.4357.2330.5652.32
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.469, + 0.718 + ], + "angle": 0, + "content": "Analysis on weakly-supervised losses. Ablation studies on contributions of two weakly-supervised losses \\(\\mathcal{L}_{\\mathrm{sreg}}\\) and \\(\\mathcal{L}_{\\mathrm{oga}}\\) are shown in Table 5. In experiment (a) and (c), we remove \\(\\mathcal{L}_{\\mathrm{sreg}}\\) but use a plain regression loss \\(\\mathcal{L}_{\\mathrm{reg}}\\) for ablation analysis. Firstly, simply employing regressive supervision attains inferior performance because the feature-level cross-modal correspondence can hardly be learned with coordinate-level supervision. Furthermore, we find \\(\\mathcal{L}_{\\mathrm{oga}}\\) is critical for learning precise temporal localization since it explicitly guides the model to align video features and language features that are highly likely to be correlated. Besides, we observe that using \\(\\mathcal{L}_{\\mathrm{sreg}}\\) for selecting high-quality regression samples always brings gains to the performance. Impact of auxiliary losses. We investigate the influences of auxiliary losses on the model performance, which involve \\(\\mathcal{L}_{\\mathrm{cb}}\\) for exploiting the cross-branch knowledge, \\(\\mathcal{L}_{\\mathrm{ar}}\\) as guidance to learn a set of order-preserving anchor boxes and \\(\\mathcal{L}_{\\mathrm{pa}}\\) to make use of the feature alignment supervision from the augmentation branch. As presented in Table 6, all three auxiliary losses bring positive impacts, with \\(\\mathcal{L}_{\\mathrm{cb}}\\) being the most effective to improve the mIoU metric by \\(4.25\\%\\). The reason might lie in the rich complementary knowledge and consistency supervision across the siamese branches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Impact of the paragraph representation. The comparison of different types of paragraph representation is shown in Table 7. Four different schemes are included, i.e., mean-pooling or max-pooling the sentence features, embedding the sequence of all word tokens in the paragraph, and using a learnable query token to extract global information. It is clear that adaptively learning a paragraph representation with our method achieves the best performance and the max-pooling scheme performs the worst because it loses too much detailed information. The mean-pooling scheme performs slightly better than the word-concat scheme, which may attribute to the advantage of late fusion at feature level." + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.089, + 0.889, + 0.123 + ], + "angle": 0, + "content": "① A woman is sitting down on a floral glass table drilling a design into the pumpkin. ② As she is drilling, two boys are standing next to her watching her and then they suddenly leave. ③ The person behind the camera then picks up the top of the pumpkin to show its empty contents before the lady closes it and continues drilling." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.124, + 0.888, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.244, + 0.88, + 0.262 + ], + "angle": 0, + "content": "A woman is seen standing in a circle and looking over her shoulder. She throws an object off into the distance and is shown again. 3 She throws her arms up and walks away afterwards." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.261, + 0.888, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.383, + 0.89, + 0.397 + ], + "angle": 0, + "content": "Figure 4. Visualization of prediction results from different models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.637, + 0.419 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.58 + ], + "angle": 0, + "content": "In Figure 4, we intuitively visualize the predicted timestamps from WSTPN and our proposed model trained under the weakly-supervised or semi-supervised setting. Overall, the predicted timestamps given by WSTPN coarsely capture the sentence order relations but with inaccurate boundaries. In contrast, our weakly-supervised model achieves much better results. It is notable that our semi-supervised model generates more fine-grained boundaries, which shows the advantage of our siamese framework in jointly leveraging fewer and weaker labels for efficient learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.619, + 0.604 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In this work, we explore the weakly-supervised setting in video paragraph grounding (i.e., WSVPG) to eliminate the dependence of the temporal annotations. To achieve this goal, we propose a novel siamese learning framework to jointly learn the cross-modal feature alignment and temporal coordinate regression without ground-truth supervision. Specifically, we design a novel Siamese Grounding Transformer (SiamGTR) consisting of an augmentation branch and an inference branch. The augmentation branch utilizes the boundary supervision provided by temporally regressing a complete paragraph in a pseudo video, and the inference branch learns the order-guided cross-modal correspondence of multiple sentences in a normal video. Extensive experiments verify the effectiveness of our framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009), and Guangzhou Science and Technology Plan Project (202201011134)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13576" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.173, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.23, + 0.469, + 0.27 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.273, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[4] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.303, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[5] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV Workshops, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.346, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[6] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. In NeurIPS, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.403, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[7] Meng Cao, Long Chen, Mike Zheng Shou, Can Zhang, and Yuexian Zou. On pursuit of designing multi-modal transformer for video grounding. In EMNLP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.446, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[8] Meng Cao, Fangyun Wei, Can Xu, Xiubo Geng, Long Chen, Can Zhang, Yuexian Zou, Tao Shen, and Daxin Jiang. Iterative proposal refinement for weakly-supervised video grounding. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.503, + 0.469, + 0.555 + ], + "angle": 0, + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.56, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.617, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[11] Jiaming Chen, Weixin Luo, Wei Zhang, and Lin Ma. Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In AAAI, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.674, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[12] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.731, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.774, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[14] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[15] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[16] Kyunghyun Cho, B van Merrienboer, Caglar Gulcehre, F Bougares, H Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP, 2014. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[17] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In ACMMM, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[18] Xuguang Duan, Wenbing Huang, Chuang Gan, Jingdong Wang, Wenwu Zhu, and Junzhou Huang. Weakly supervised dense event captioning in videos. In NeurIPS, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.194, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[19] Yazan Abu Farha and Jurgen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.238, + 0.892, + 0.264 + ], + "angle": 0, + "content": "[20] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.267, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[21] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.311, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.354, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[23] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.412, + 0.892, + 0.45 + ], + "angle": 0, + "content": "[24] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. arXiv preprint arXiv:2305.14344, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.455, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.499, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[26] Jiabo Huang, Yang Liu, Shaogang Gong, and Hailin Jin. Cross-sentence temporal and semantic relations in video activity localisation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.542, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.585, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[28] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.629, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[29] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.673, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[30] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.716, + 0.892, + 0.768 + ], + "angle": 0, + "content": "[31] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.773, + 0.892, + 0.812 + ], + "angle": 0, + "content": "[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[33] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Workshops, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[34] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13577" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[35] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.151, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[36] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.194, + 0.47, + 0.235 + ], + "angle": 0, + "content": "[37] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.238, + 0.47, + 0.278 + ], + "angle": 0, + "content": "[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.281, + 0.47, + 0.322 + ], + "angle": 0, + "content": "[39] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.325, + 0.47, + 0.379 + ], + "angle": 0, + "content": "[40] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[41] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[42] Zhijie Lin, Zhou Zhao, Zhu Zhang, Qi Wang, and Huasheng Liu. Weakly-supervised video moment retrieval via semantic completion network. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[43] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In ACMMM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.513, + 0.47, + 0.554 + ], + "angle": 0, + "content": "[44] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.556, + 0.47, + 0.597 + ], + "angle": 0, + "content": "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.6, + 0.47, + 0.653 + ], + "angle": 0, + "content": "[46] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In ICLR, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.657, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[47] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In EMNLP, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.701, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[48] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. TMM, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.744, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[49] Minuk Ma, Sunjae Yoon, Junyeong Kim, Youngjoon Lee, Sunghun Kang, and Chang D Yoo. Vlanet: Video-language alignment network for weakly-supervised video moment retrieval. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[51] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[52] Niluthpol Chowdhury Mithun, Sujoy Paul, and Amit K Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[53] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.18, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[54] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. Bassl: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.238, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[55] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[56] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.325, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[57] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.368, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.426, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[59] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.455, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[60] Mrigank Rochan, Linwei Ye, and Yang Wang. Video summarization using fully convolutional sequence networks. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.499, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[61] Marcus Rohrbach, Michaela Regneri, Mykhaylo Andriluka, Sikandar Amin, Manfred Pinkal, and Bernt Schiele. Script data for attribute-based recognition of composite activities. In ECCV, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.556, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[62] Fengyuan Shi, Limin Wang, and Weilin Huang. End-to-end dense video grounding via parallel regression. arXiv preprint arXiv:2109.11265, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.6, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[63] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.657, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[64] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. Vlg-net: Video-language graph matching network for video grounding. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.701, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[65] Yijun Song, Jingwen Wang, Lin Ma, Zhou Yu, and Jun Yu. Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.759, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[66] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In CVPR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[67] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Wei-Shi Zheng, and Jianhuang Lai. Hierarchical semantic correspondence networks for video paragraph grounding. In CVPR, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[68] Reuben Tan, Huijuan Xu, Kate Saenko, and Bryan A Plummer. Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In WACV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13578" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[69] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[70] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[72] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. In AAAI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[73] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[74] Jingwen Wang, Lin Ma, and Wenhao Jiang. Temporally grounding language queries in videos by contextual boundary-aware prediction. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[75] Yuechen Wang, Jiajun Deng, Wengang Zhou, and Houqiang Li. Weakly supervised temporal adjacent network for language grounding. TMM, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.414, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[76] Yunxiao Wang, Meng Liu, Yinwei Wei, Zhiyong Cheng, Yinglong Wang, and Liqiang Nie. Siamese alignment network for weakly supervised video moment retrieval. TMM, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[77] Zheng Wang, Jingjing Chen, and Yu-Gang Jiang. Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In ACMMM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[78] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.553, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[79] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[80] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.637, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[81] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.679, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[82] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.47, + 0.762 + ], + "angle": 0, + "content": "[83] Shuwen Xiao, Zhou Zhao, Zijian Zhang, Xiaohui Yan, and Min Yang. Convolutional hierarchical attention network for query-focused video summarization. In AAAI, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.763, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[84] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. In EMNLP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.805, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[85] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[86] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In AAAI, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[87] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[88] Wenfei Yang, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Local correspondence network for weakly supervised temporal sentence grounding. TIP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.192, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[89] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[90] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[91] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In AAAI, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[92] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In ACMMM Workshops, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[93] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[94] Chenchi Zhang, Wenbo Ma, Jun Xiao, Hanwang Zhang, Jian Shao, Yueting Zhuang, and Long Chen. Vl-nms: Breaking proposal bottlenecks in two-stage visual-language matching. TOMM, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[95] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[96] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.59, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[97] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[98] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.704, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[99] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[100] Mingxing Zhang, Yang Yang, Xinghan Chen, Yanli Ji, Xing Xu, Jingjing Li, and Heng Tao Shen. Multi-stage aggregated transformer network for temporal language localization in videos. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[101] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[102] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. TPAMI, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13579" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[103] Yimeng Zhang, Xin Chen, Jinghan Jia, Sijia Liu, and Ke Ding. Text-visual prompting for efficient 2d temporal video grounding. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[104] Zhu Zhang, Zhijie Lin, Zhou Zhao, Jieming Zhu, and Xiuqiang He. Regularized two-branch proposal networks for weakly-supervised moment retrieval in videos. In ACMMM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[105] Zhu Zhang, Zhou Zhao, Zhijie Lin, Xiuqiang He, et al. Counterfactual contrastive learning for weakly-supervised vision-language grounding. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.276 + ], + "angle": 0, + "content": "[106] Minghang Zheng, Yanjie Huang, Qingchao Chen, and Yang Liu. Weakly supervised video moment localization with contrastive negative sample mining. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[107] Minghang Zheng, Yanjie Huang, Qingchao Chen, Yuxin Peng, and Yang Liu. Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[108] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13580" + } + ] +] \ No newline at end of file diff --git a/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_origin.pdf b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3b62529918233f55c477972398764c5355b9cd25 --- /dev/null +++ b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/b2a21229-501b-4026-b12f-315fb210d141_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05034845892423214e8e455e0d252864f506aaad6462dd1776fc4f907c649330 +size 1892746 diff --git a/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/full.md b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..77a8ee4152da31880bd721a023a5a127c002965d --- /dev/null +++ b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/full.md @@ -0,0 +1,380 @@ +# Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding + +Chaolei Tan $^{1}$ Jianhuang Lai $^{1,2,3}$ Wei-Shi Zheng $^{1,2,3}$ Jian-Fang Hu $^{1,2,3*}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China + +2Guangdong Province Key Laboratory of Information Security Technology, China + +3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +tanchlei@mail2.sysu.edu.cn, stsljh@mail.sysu.edu.cn, wszheng@ieee.org, hujf5@mail.sysu.edu.cn + +# Abstract + +Video Paragraph Grounding (VPG) is an emerging task in video-language understanding, which aims at localizing multiple sentences with semantic relations and temporal order from an untrimmed video. However, existing VPG approaches are heavily reliant on a considerable number of temporal labels that are laborious and time-consuming to acquire. In this work, we introduce and explore Weakly-Supervised Video Paragraph Grounding (WSVPG) to eliminate the need of temporal annotations. Different from previous weakly-supervised grounding frameworks based on multiple instance learning or reconstruction learning for two-stage candidate ranking, we propose a novel siamese learning framework that jointly learns the cross-modal feature alignment and temporal coordinate regression without timestamp labels to achieve concise one-stage localization for WSVPG. Specifically, we devise a Siamese Grounding Transformer (SiamGTR) consisting of two weight-sharing branches for learning complementary supervision. An Augmentation Branch is utilized for directly regressing the temporal boundaries of a complete paragraph within a pseudo video, and an Inference Branch is designed to capture the order-guided feature correspondence for localizing multiple sentences in a normal video. We demonstrate by extensive experiments that our paradigm has superior practicability and flexibility to achieve efficient weakly-supervised or semi-supervised learning, outperforming state-of-the-art methods trained with the same or stronger supervision. + +# 1. Introduction + +Natural Language Video Grounding (NLVG) is an essential area in vision-language understanding, which has received increasing attention due to its wide range of real-world applications such as video retrieval [3, 17, 21, 23, 80, 94], video summarization [50, 55, 59, 60, 83, 99], action segmentation [19, 27, 35, 37, 45, 72], video question answering [30, 36, 38, 41, 82, 89], etc. Most previous works focus + +![](images/3a07f72f59db4d4e9d5779d04b2bdc7fd7225047d51fddad602d14601e61d026.jpg) +(a) + +![](images/dcd0c75ad0246d487780dbf293a74255f905341c913c5972890fa3ecc999c344.jpg) +Figure 1. (a) Chronological cross-modal alignment in a video and its paired sentences. (b) Pseudo boundary supervision for regressing paragraph timestamps in a composed pseudo video. + +on tackling Video Sentence Grounding (VSG) proposed in [1, 22], which targets at localizing the temporal boundaries of an individual sentence from an untrimmed video. However, localizing single sentences can be ambiguous because the contextual information conveyed by multiple sentences is necessary to uniquely determine the temporal locations of input queries. To alleviate such issue, Bao et al. [4] proposed to contextualize video grounding into localizing multiple events indicated by sentences of a paragraph from the video, which is called Video Paragraph Grounding (VPG). + +Although remarkable progress has been attained in tackling VSG and VPG problems under a fully-supervised setting, the extremely prohibitive overheads of manually annotating temporal boundaries for language queries limit these methods to utilize large-scale video-text data. In addition, the subjectivity of annotators inevitably brings label noises that may adversely affect model training. Recently, weakly-supervised methods free of temporal annotations have become increasingly popular in the area of video grounding, aiming to address the above limitations. These approaches can be mainly categorized into multiple instance learning methods and reconstruction learning methods and are typ- + +ically based on a propose-and-rank pipeline. Nevertheless, both of these paradigms assume the contributions of proposals to the contrastive or reconstruction loss accurately represent the proposal quality, which is not necessarily the case during model learning. Moreover, the quadratic complexity of proposal schemes prevents them from scaling up in parameters or training data, and the adopted supervision of video-level contrast or query-based reconstruction in prior works is not temporal-sensitive thus suffering from a huge gap with fully-supervised temporal guidance. Besides, all of the weakly-supervised methods in video grounding are tailored for tackling VSG while the weakly-supervised setting of VPG (i.e., WSVPG) has been understudied so far. + +To circumvent the aforementioned drawbacks and explore an efficient weakly-supervised framework for VPG, we seek to mine the unique characteristics and underlying supervision from the intrinsic structure of video-paragraph pairs for model training. On the one hand, as observed in Figure 1 (a), the temporal location of an event is highly correlated with the position of the sentence describing that event in the paragraph. For example, the sentence appearing in the middle of the paragraph tends to have stronger relevance with visual content located around the temporal midpoint of the video. On the other hand, dense visual events mentioned in the paragraph approximately represent the global video content that unambiguously distinguishes itself from another video, which can be observed in Figure 1 (b). Therefore, inserting the query-related video into another irrelevant video automatically generates pseudo boundary labels close to the ground-truth when regarding the complete paragraph as language query for video grounding. + +Motivated by the above observation, we propose a novel Siamese Grounding TRansformer (SiamGTR) for WSVPG. It jointly learns the cross-modal alignment and boundary regression via two siamese branches without generating proposals. Specifically, we propose to construct an Augmentation Branch (AB) which takes as input a pseudo video and adopts a complete paragraph as the language query to learn high-quality boundary supervision for localization. Also, an Inference Branch (IB) is designed to receive a normal video as input and is enforced to capture the order-guided cross-modal correspondence for attending over the specific video content relevant to each sentence. The two weight-sharing branches are effective to transfer complementary supervision for joint boundary prediction and feature association, which yields a weakly-supervised model with superior generalization through a concise one-stage pipeline. Extensive experiments verify the effectiveness of our model and show our method with the same or weaker supervision surpasses prior state-of-the-arts. In summary, our contributions are: + +- We introduce the task of Weakly-Supervised Video Paragraph Grounding (WSVPG), which aims to train a model for localizing multiple events indicated by queries with + +out the supervision of timestamp labels. + +- We propose a novel Siamese Grounding TRansformer (SiamGTR) for concise and efficient one-stage weakly-supervised learning of video paragraph grounding. It is composed of two weight-sharing branches including an Augmentation Branch (AB) for learning boundary regression of pseudo boundaries and an Inference Branch (IB) for learning order-guided cross-modal feature alignment. +- Extensive experiments verify the efficacy of our method, and demonstrate that our framework under the same or weaker supervision outperforms state-of-the-arts. + +# 2. Related Work + +# 2.1. Video Sentence Grounding + +Fully-Supervised Video Sentence Grounding. Plenty of approaches [1, 7, 12, 22, 29, 40, 43, 44, 47, 53, 64, 73, 74, 78, 79, 84-86, 90, 91, 93, 95-97, 100-103] have been proposed to address Fully-Supervised Video Sentence Grounding (FSVSG). In general, these works can be roughly categorized into proposal-based and proposal-free methods. Specifically, proposal-based methods [1, 22, 40, 43, 44, 73, 74, 78, 85, 86, 90, 93, 95, 101, 102] involve a proposal generation stage using sliding windows [22], anchor proposals [12, 43, 44, 47, 74, 90, 93, 95] or 2D temporal maps [64, 73, 78, 85, 101, 102], after which the generated proposals are ranked according to the query matching scores with potential post-processing like Non-Maximum Suppression (NMS). In contrast, proposal-free methods [7, 53, 79, 91, 96, 97, 100, 103] remove the dense proposal generation and score ranking process by directly regressing timestamps [7, 53, 84, 91, 103], predicting boundary distributions [96, 97, 100] or using reinforcement learning [79], which improves the computation efficiency and scenario adaptability. Following the line of proposal-free works, we propose a novel weakly-supervised regression-based framework that shows superior performance and practicability. + +Weakly-Supervised Video Sentence Grounding. Weakly-Supervised Video Sentence Grounding (WSVSG) [8, 11, 26, 28, 42, 49, 52, 65, 68, 75-77, 88, 104-107] has become a popular research area because of the severe dependence of FSVSG approaches on laborious and expensive manual temporal annotations. Most of existing WSVSG methods are based on a two-stage pipeline using multiple instance learning [26, 49, 75, 88], reconstruction learning [42], or the combination of both [106, 107]. In particular, Chen et al. [11] have proposed a video composition strategy to generate pseudo temporal labels for WSVSG, which is the most related work to ours. However, several inherent drawbacks are involved in this approach. Firstly, individual sentences only describe local video content, thus viewing the starting/ending locations of foreground video as temporal boundaries of an individual sentence produces a weak and noisy temporal alignment, which is unsuitable for accurate + +boundary supervision. Moreover, it simply adapts an existing fully-supervised proposal-based framework for weakly-supervised training, which leads to inferior generalization caused by the large train-test discrepancy. Distinct from all of the above works, we design a novel siamese framework to capture the essential characteristics of video paragraph grounding for weakly-supervised learning. + +# 2.2. Video Paragraph Grounding + +Video Paragraph Grounding (VPG) is introduced by Bao et al. [4], which aims to jointly localize multiple sentences of a paragraph from an untrimmed video. Shi et al. [62] presented an end-to-end network by re-purposing transformers into language-conditioned regressors. Jiang et al. [31] proposed to employ contrastive encoders for contrastive learning between video-paragraph pairs. Tan et al. [67] proposed a hierarchical semantic correspondence network for modeling hierarchical video-language alignment and grounding multiple levels of language queries in the video. Particularly, Jiang et al. [31] first explored Semi-Supervised Video Paragraph Grounding (SSVPG) to relieve the annotation burden of temporal labels. However, SSVPG is still not quite practical considering the expensive cost of temporal annotations in untrimmed videos. Besides, these semi-supervised methods still require a considerable proportion of temporal labels up to at least $10\%$ [31] for training. To thoroughly get rid of the temporal annotations, we pioneer to explore the weakly-supervised setting in VPG. + +# 2.3. Siamese Networks + +Siamese networks [6] are weight-sharing neural networks. There have been a wide range of scenarios where siamese networks are applied for achieving different purposes, such as face verification [66], image recognition [33], object tracking [5], etc. In particular, siamese networks are commonly used in contrastive self-supervised learning methods [2, 10, 13-15, 24, 25, 69, 81, 108], in which augmented views of the same or different instance are forwarded into multiple weight-sharing network copies for learning a generalizable visual representation via instance-level discrimination. In this work, we explore a new way to combine the transferability of siamese architectures and the flexibility of transformer architectures for concise and efficient weakly-supervised learning of video paragraph grounding. + +# 3. Methodology + +# 3.1. Overview + +Task Formulation. Given an untrimmed video $\mathcal{V}$ and a paragraph query $\mathcal{P}$ consisting of $N$ temporally ordered sentences $\{\mathcal{S}_i\}_{i=1}^N$ , the goal of Video Paragraph Grounding (VPG) is to simultaneously localize the temporal intervals $\mathcal{T} = \{(\tau_i^{\mathrm{st}}, \tau_i^{\mathrm{ed}})\}_{i=1}^N$ of all the events described by sentences + +![](images/2e2781a658a58acd039e4de131bab5d48e8d0bb526f09b19cb8e008e7e33bfc8.jpg) +Figure 2. Our siamese framework for joint alignment and regression. + +![](images/71bc0739af9c28e8c4279d701bbd3063cc32d1ac2149ffe969a1d6ab057af9f0.jpg) + +in the paragraph, where $\tau_{i}^{\mathrm{st}}$ and $\tau_{i}^{\mathrm{ed}}$ are the starting and ending timestamps for the $i$ -th sentence $S_{i}$ , respectively. + +Siamese Learning. An overview of our siamese framework is shown in Figure 2. Overall, we jointly train an augmentation branch and an inference branch with shared parameters and structures for learning the complementary abilities of cross-modal feature alignment and temporal boundary regression. These two branches follow the same workflow to first encode the text queries and input video into unimodal features, after which the query features are iteratively used in the transformer decoder to extract relevant information from the video features for timestamp decoding. For testing, we only keep the pipeline of inference branch for boundary prediction. More architectural details are illustrated in Figure 3 and are elaborated in the following sections. + +# 3.2. Feature Extraction + +Video Feature Extraction. For each input video, we divide it into consecutive clips consisting of a fixed number of frames for feature extraction. Specifically, a frozen pretrained 3D Convolutional Neural Network (3D-CNN) [70] and a linear projection layer are successively employed to obtain a 1D feature vector for each short clip, resulting in a video feature sequence $\mathcal{F}_{\mathrm{v}} \in \mathbb{R}^{L \times D}$ , where $L$ and $D$ are the sequence length and hidden dimension, respectively. + +Text Feature Extraction. For each input paragraph consisting of $N$ sentences, we first utilize a frozen pre-trained word embedding model tookenize and embed the text into a sequence of word vectors. Then, a bidirectional Gated Recurrent Unit (GRU) [16] is employed on each sentence, and the last hidden states in both directions are concatenated and then projected by a linear layer to construct the sentence features $\mathcal{F}_{\mathrm{s}}\in \mathbb{R}^{N\times D}$ , where $D$ is the hidden dimension. + +# 3.3. Augmentation Branch + +The augmentation branch aims to learn accurate boundary regression from pseudo videos with paragraph queries, which naturally transfers to the inference branch via the shared feature space established by the siamese structure. + +Pseudo Data Generation. To drive end-to-end weakly-supervised regression learning, the input stream of the augmentation branch should provide reliable and direct boundary supervision. Drawing inspiration from the boundary-sensitive video pretext tasks [11, 54, 87], we propose to uti + +![](images/a245b9fc51cd6f2fbd98c7cda54eff7d73eaf65fa8de3049543f3ba4f0ae5b6d.jpg) +Figure 3. Illustration of the proposed Siamese Grounding TRansformer (SiamGTR) architecture. The augmentation branch (abbreviated as A.B.) takes the pseudo video features derived from randomly inserting the query-related video features into irrelevant video features. It learns to temporally regress the interval of interest from the pseudo video with the paragraph as query. The Inference Branch (abbreviated as I.B.) receives normal video features for learning the cross-modal feature alignment among multiple sentences in the video. + +lize the synthesized videos paired with paragraph queries to serve as a well-suited source of surrogate boundary supervision. Specifically, for each input video $\nu$ and its paragraph description $\mathcal{P}$ , we randomly sample a background video $\nu_{\mathrm{bg}}$ to obtain irrelevant video content to $\mathcal{P}$ . Because $\mathcal{P}$ could be treated as a unique referring expression specific to $\nu$ , the search space of background videos can be the entire training set for increasing the data diversity. Denote video features of $\nu$ and $\nu_{\mathrm{bg}}$ as $\mathcal{F}_{\mathrm{v}}$ and $\mathcal{F}_{\mathrm{bg}}$ respectively, we construct the pseudo video features $\mathcal{F}_{\mathrm{aug}}$ as: + +$$ +\mathcal {F} _ {\mathrm {a u g}} = \operatorname {N R S} \left(\operatorname {C o n c a t} \left(\mathcal {F} _ {\mathrm {b g}} ^ {1: I}, \operatorname {R R S} \left(\mathcal {F} _ {\mathrm {v}}\right), \mathcal {F} _ {\mathrm {b g}} ^ {I + 1:}\right)\right) \tag {1} +$$ + +where $I$ is a randomly generated inserting index within the feature sequence $\mathcal{F}_{\mathrm{bg}}$ and Concat ( $\cdot$ ) denotes the temporal concatenation. RRS ( $\cdot$ ) and NRS ( $\cdot$ ) respectively represent a Random Re-Sampling operation to stochastically re-scale the length of $\mathcal{F}_{\mathrm{v}}$ and a Normalized Re-Sampling operation that converts the length of $\mathcal{F}_{\mathrm{aug}}$ into a fixed number of $T$ . Thereafter, we compute the synthesized temporal boundaries ( $\tau_{\mathrm{aug}}^{\mathrm{st}}, \tau_{\mathrm{aug}}^{\mathrm{ed}}$ ) in the pseudo video $\mathcal{V}_{\mathrm{aug}}$ as follows: + +$$ +\tau_ {\mathrm {a u g}} ^ {\mathrm {s t}} = \frac {I + \Delta I ^ {\mathrm {s t}}}{r L + L _ {\mathrm {b g}}}, \tau_ {\mathrm {a u g}} ^ {\mathrm {e d}} = \frac {I + r L - \Delta I ^ {\mathrm {e d}}}{r L + L _ {\mathrm {b g}}} \tag {2} +$$ + +where $L$ and $L_{\mathrm{bg}}$ respectively indicate the length of $\mathcal{F}_{\mathrm{v}}$ and $\mathcal{F}_{\mathrm{bg}}$ with $r$ being the random re-scaling factor of RRS $(\cdot)$ . To alleviate potential synthesis artifacts and boundary uncertainty, we further introduce a Random Boundary Shifting (RBS) strategy that incorporates small random offsets $\Delta I^{\mathrm{st}}$ and $\Delta I^{\mathrm{ed}}$ into computing the pseudo labels, which is simple yet effective to boost the quality of boundary supervision. + +Video Encoder. Based on the obtained pseudo video features $\mathcal{F}_{\mathrm{aug}}$ , we then encode the temporal contextual information across multiple clips by feeding $\mathcal{F}_{\mathrm{aug}}$ to a video encoder, which follows a similar architecture of DETR encoders [9, 46, 71], i.e., each encoder layer consists of a + +multi-head self-attention module equipped with additive sinusoidal positional encodings and a feed-forward network. For simplicity, we omit the layer index and denote the input of each video encoder layer as $\mathcal{X}_{\mathrm{aug}}$ , then a set of Modulated Positional Encodings (MPE) are given as: + +$$ +\mathcal {X} _ {\mathrm {p e}} = \operatorname {M L P} \left(\mathcal {X} _ {\mathrm {a u g}}\right) \odot \operatorname {P E} \left(\mathcal {X} _ {\mathrm {a u g}}\right) \tag {3} +$$ + +where $\mathcal{X}_{\mathrm{pe}} \in \mathbb{R}^{T \times D}$ and $\odot$ denotes element-wise multiplication. PE $(\cdot)$ is the sinusoidal function [71] with respect to temporal locations of $\mathcal{X}_{\mathrm{aug}}$ , MLP $(\cdot)$ denotes a two-layer feed-forward network computed on $\mathcal{X}_{\mathrm{aug}}$ . In each encoder layer, $\mathcal{X}_{\mathrm{pe}}$ is only added to the input of projection layers of queries and keys in the self-attention mechanism. The encoded video features of the last video encoder layer, also called the encoder memory features $\mathcal{M}_{\mathrm{aug}}$ , are further fed to the query-guided decoder for localization prediction. + +Query Encoder. Since there are strong contextual correlations among multiple sentences in the paragraph query, here we utilize a vanilla transformer encoder [71] to help reason the semantic and chronological relationships of events. To extract a global representation of the complete paragraph, we initialize an extra learnable query token $\mathcal{F}_{\mathrm{p}}$ and integrate it with the sentence features $\mathcal{F}_{\mathrm{s}}$ by concatenation. Thus, the input of the query encoder is constructed as follows: + +$$ +\mathcal {X} _ {\mathrm {q}} = \left[ \mathcal {F} _ {\mathrm {p}}, \mathcal {F} _ {\mathrm {s}} ^ {1}, \mathcal {F} _ {\mathrm {s}} ^ {2}, \dots , \mathcal {F} _ {\mathrm {s}} ^ {i}, \dots \mathcal {F} _ {\mathrm {s}} ^ {N} \right] \tag {4} +$$ + +where $\mathcal{X}_{\mathrm{q}} \in \mathbb{R}^{(N + 1) \times D}$ and $\mathcal{F}_{\mathrm{s}}^{i}$ is the $i$ -th sentence feature in $\mathcal{F}_{\mathrm{s}}$ . We first normalize $\mathcal{X}_{\mathrm{q}}$ , add fixed sinusoidal positional encodings to it, and then iteratively employ self-attention modules and feed-forward networks to contextualize the global and local tokens that respectively represent the query semantics of the entire paragraph and the sentences. The output features of the last query encoder layer are denoted as $\mathcal{Z}_{\mathrm{q}}$ and are further forwarded to the query decoder. + +Conceptual Semantic Connector. Although the siamese network structure can implicitly transfer boundary knowl- + +edge from the augmentation branch to the inference branch, there is still a certain semantic gap between the query representations of short sentences and long paragraphs. To narrow this gap, we develop a Conceptual Semantic Connector (CSC) module for explicit semantic guidance. Specifically, we first collect high-frequency linguistic concepts, including verbs and nouns from the training corpus, and then construct a set of dictionary features by selecting and projecting the Glove vectors [56]. The loss $\mathcal{L}_{\mathrm{csc}}$ is computed as: + +$$ +\mathcal {L} _ {\mathrm {c s c}} = \mathrm {B C E} \left(y _ {\text {c p t}} ^ {\mathrm {p}}, \hat {y} _ {\text {c e p t}} ^ {\mathrm {p}}\right) + \mathrm {B C E} \left(y _ {\text {s c t}} ^ {\mathrm {s}}, \hat {y} _ {\text {c e p t}} ^ {\mathrm {s}}\right) \tag {5} +$$ + +where $\mathrm{BCE}(\cdot)$ is the binary cross-entropy loss. $y_{\mathrm{cept}}^{\mathrm{p}}$ and $y_{\mathrm{cept}}^{\mathrm{s}}$ are multi-hot labels indicating semantic concepts contained by the paragraph and sentence queries, respectively. $\hat{y}_{\mathrm{cept}}^{\mathrm{s}}$ and $\hat{y}_{\mathrm{cept}}^{\mathrm{p}}$ are the concept predictions obtained by dot-product between the conceptual dictionary features and the textual query features $\mathcal{Z}_{\mathrm{q}}$ with a sigmoid activation. + +Query Decoder. Inspired by the spiritual ideas of dynamic-anchor DETR decoders [39, 46, 51, 98], we design a novel query decoder that enables dynamic position-aware decoding of language queries. Specifically, each query decoder layer consists of a self-attention module, a cross-attention module, and a feed-forward network, with dynamically adjustable anchor boxes to indicate the query-specific location information. Initially, the input anchor boxes are set to all zeros, and an MLP is used to estimate a set of seed anchor boxes based on the cross-modal interactions between $\mathcal{M}_{\mathrm{aug}}$ and $\mathcal{Z}_{\mathrm{q}}$ , i.e., the output features of the first decoder layer $\mathcal{Q}_{\mathrm{aug}}^{(1)} \in \mathbb{R}^{(N + 1) \times D}$ are mapped into the seed anchor boxes $\mathcal{A}_{\mathrm{aug}}^{(1)} \in \mathbb{R}^{(N + 1) \times 2}$ by the MLP. For the $(i + 1)$ -th decoder layer, we first convert the box coordinates of the input anchors $\mathcal{A}_{\mathrm{aug}}^{(i)}$ into high-dimensional sinusoidal embeddings $\mathcal{F}_{\mathrm{a}}^{(i)} \in \mathbb{R}^{(N + 1) \times D}$ by a sinusoidal function [46, 71] and further obtain $\mathcal{H}_{\mathrm{a}}^{(i)}$ by projecting $\mathcal{F}_{\mathrm{a}}^{(i)}$ with an MLP. Afterwards, we conduct the self-attention operation and update the query features in the decoder as follows: + +$$ +\mathcal {Q} _ {\mathrm {a u g}} ^ {(i + 1)} \leftarrow \text {S e l f - A t t n} \left\{ \begin{array}{l} Q = \varphi_ {q} ^ {c} \left(\mathcal {Q} _ {\mathrm {a u g}} ^ {(i)}\right) + \varphi_ {q} ^ {p} \left(\mathcal {H} _ {\mathrm {a}} ^ {(i)}\right) \\ K = \varphi_ {k} ^ {c} \left(\mathcal {Q} _ {\mathrm {a u g}} ^ {(i)}\right) + \varphi_ {k} ^ {p} \left(\mathcal {H} _ {\mathrm {a}} ^ {(i)}\right) \\ V = \varphi_ {v} ^ {c} \left(\mathcal {Q} _ {\mathrm {a u g}} ^ {(i)}\right) \end{array} \right. \tag {6} +$$ + +where $\mathcal{Q}_{\mathrm{aug}}^{(i + 1)}$ is the updated query features after the self-attention operation in the $(i + 1)$ -th decoder layer. The above series of $\varphi$ functions are used to indicate different linear projection layers for the content part or position part of the queries, keys, or values in the self-attention mechanism. Then, we conduct a cross-attention operation to extract useful cross-modal interactive information as follows: + +$$ +\mathcal {Q} _ {\mathrm {a u g}} ^ {(i + 1)} \leftarrow \text {C r o s s - A t t n} \left\{ \begin{array}{l} Q = \left[ \varphi_ {q} ^ {c} \left(\mathcal {Q} _ {\mathrm {a u g}} ^ {(i + 1)}\right); \varphi_ {q} ^ {p} \left(\mathcal {F} _ {\mathrm {a}} ^ {(i)}\right) \right] \\ K = \left[ \varphi_ {k} ^ {c} \left(\mathcal {M} _ {\mathrm {a u g}}\right); \varphi_ {k} ^ {p} \left(\mathcal {F} _ {\mathrm {p e}} ^ {\mathcal {M}}\right) \right] \\ V = \varphi_ {v} ^ {c} \left(\mathcal {M} _ {\mathrm {a u g}}\right) + \varphi_ {v} ^ {p} \left(\mathcal {F} _ {\mathrm {p e}} ^ {\mathcal {M}}\right) \end{array} \right. \tag {7} +$$ + +where $\mathcal{F}_{\mathrm{pe}}^{\mathcal{M}} = \mathrm{PE}\left(\mathcal{M}_{\mathrm{aug}}\right)$ and the query features $\mathcal{Q}_{\mathrm{aug}}^{(i + 1)}$ is then further updated by a feed-forward network as the feature output of the $(i + 1)$ -th decoder layer. Then the anchor boxes are dynamically updated as $\mathcal{A}_{\mathrm{aug}}^{(i + 1)} \gets \mathcal{A}_{\mathrm{aug}}^{(i)} + \Delta \mathcal{A}_{\mathrm{aug}}^{(i)}$ where we utilize an MLP layer to predict the relative offsets, i.e., $\Delta \mathcal{A}_{\mathrm{aug}}^{(i)} \in \mathbb{R}^{(N + 1) \times 2}$ , based on the updated query features $\mathcal{Q}_{\mathrm{aug}}^{(i + 1)}$ . $\mathcal{A}_{\mathrm{aug}}^{(i + 1)}$ continues to be forwarded to the next decoder layer for computing $\mathcal{F}_{\mathrm{a}}^{(i + 1)}$ and $\mathcal{H}_{\mathrm{a}}^{(i + 1)}$ . + +Boundary Prediction. Based on the output features and attention weights of the last query decoder layer, we simply use an MLP predictor to predict the paragraph timestamps, i.e., $\hat{\mathcal{T}}_{\mathrm{aug}}^{\mathrm{p}} = (\hat{\tau}_{\mathrm{aug}}^{\mathrm{st}},\hat{\tau}_{\mathrm{aug}}^{\mathrm{ed}})$ . Similarly, the sentence timestamps $\widehat{\mathcal{T}}_{\mathrm{inf}}^{\mathrm{s}} = \left\{(\hat{\tau}_{\mathrm{j}}^{\mathrm{st}},\hat{\tau}_{\mathrm{j}}^{\mathrm{ed}})\right\}_{j = 1}^{N}$ can also be obtained by feeding last-layer output from the inference branch to the same MLP. + +Self-Consistent Boundary Regression. We improve the attention-agnostic regression loss $\mathcal{L}_{\mathrm{reg}}$ to make it aware of the model's self-consistent scores, where the main idea is to selectively optimize the regression loss of self-consistent samples for better weakly-supervised regression learning. Self-consistent samples have high attention weights over the pseudo ground-truth intervals, which are more suitable for learning less noisy supervision for accurate boundary prediction. Specifically, the self-consistent boundary regression loss $\mathcal{L}_{\mathrm{sreg}}$ is defined as: + +$$ +\mathcal {L} _ {\mathrm {s c r e g}} = \left\{ \begin{array}{l l} \mathcal {L} _ {\mathrm {L} 1} + \mathcal {L} _ {\mathrm {G I o U}}, & \text {i f} \quad s _ {\mathrm {a t t}} > \beta \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{L1}}$ and $\mathcal{L}_{\mathrm{GIoU}}$ respectively represent L1 and Generalized Intersection over Union (GIoU) [58] loss computed between $(\hat{\tau}_{\mathrm{aug}}^{\mathrm{st}},\hat{\tau}_{\mathrm{aug}}^{\mathrm{ed}})$ and $(\tau_{\mathrm{aug}}^{\mathrm{st}},\tau_{\mathrm{aug}}^{\mathrm{ed}})$ . $\beta$ is set to 0.5 and $s_{\mathrm{att}}$ is the attention sum over the pseudo ground-truth interval. + +# 3.4. Inference Branch + +In our siamese framework, the inference branch shares the same parameter weights and network structure with the augmentation branch, i.e., a video encoder, a query encoder, and a query decoder. The only difference lies in the input streams and objectives, i.e., the inference branch receives a normal video during training to learn in-domain cross-modal correspondence that cannot be acquired through the pseudo video stream, which significantly improves the generalization ability of the model. Specifically, it receives the encoded normal video features $\mathcal{M}_{\mathrm{inf}}$ and the encoded text query features $\mathcal{Z}_{\mathrm{q}}$ as the decoder input and generates the hidden query features $\mathcal{Q}_{\mathrm{inf}}$ for boundary prediction. + +Order-guided Attention Loss. The chronological prior given by the sentence order provides explicit guidance for learning cross-modal alignment between the video and language features during decoding. To learn the order-guided cross-modal correspondence, we constrain cross-modal attention weights in the decoder as follows: + +$$ +\mathcal {L} _ {\mathrm {o g a}} = \max \left(0, \Delta m T + \sum_ {t = 1} ^ {T} t \alpha_ {\mathrm {s}} ^ {j} (t) - \sum_ {t = 1} ^ {T} t \alpha_ {\mathrm {s}} ^ {j + 1} (t)\right) \tag {9} +$$ + +where $\alpha_{\mathrm{s}}^{j}(t)$ and $\alpha_{\mathrm{s}}^{j + 1}(t)$ are attention weights over video features for the $j$ -th and $(j + 1)$ -th sentence from the last decoder layer, respectively. $\Delta m$ is the minimal distance between attention centroids. This loss is mainly contributed by the inference branch with $\Delta m = \frac{1}{2N}$ and partly contributed by the augmentation branch with $\Delta m = \frac{1}{4N}$ . + +Auxiliary Losses. To exploit more guidance for weakly-supervised representation learning, we employ three auxiliary losses including a cross-branch loss $\mathcal{L}_{\mathrm{cb}}$ , an anchor ranking loss $\mathcal{L}_{\mathrm{ar}}$ and a pseudo attention loss $\mathcal{L}_{\mathrm{pa}}$ . Specifically, $\mathcal{L}_{\mathrm{cb}}$ utilizes the semantic consistency constraint of output features across siamese branches, which is calculated analogous to MoCo [25]. $\mathcal{L}_{\mathrm{ar}}$ is used for inducing the decoder to learn a set of order-preserving anchor boxes, which is computed on the anchor boxes like in the equation (9). $\mathcal{L}_{\mathrm{pa}}$ makes use of attention supervision between the pseudo video and language features from the augmentation branch, which follows the calculation proposed in [62, 91]. + +# 3.5. Training and Inference + +Weakly-Supervised Learning. The weakly-supervised loss of our proposed framework can be formulated as a weighted sum of the two losses from the siamese branches as $\mathcal{L}_{\mathrm{WS}} = \lambda_{\mathrm{sreg}}\mathcal{L}_{\mathrm{sreg}} + \lambda_{\mathrm{oga}}\mathcal{L}_{\mathrm{oga}}$ , where $\lambda_{\mathrm{sreg}}$ and $\lambda_{\mathrm{oga}}$ are scalar weights to balance the contributions of the two different losses. The overall loss for the weakly-supervised model is defined as $\mathcal{L} = \mathcal{L}_{\mathrm{WS}} + \lambda_{\mathrm{csc}}\mathcal{L}_{\mathrm{csc}} + \mathcal{L}_{\mathrm{aux}}$ . + +Semi-Supervised Learning. Although our framework is initially designed for weakly-supervised learning, it can be easily adapted for end-to-end semi-supervised learning. Specifically, in addition to $\mathcal{L}_{\mathrm{WS}}$ which is calculated on all training samples, we only need to employ an extra fully-supervised loss $\mathcal{L}_{\mathrm{FS}}$ on those fully-annotated samples without changing any part of the network structure. The semi-supervised loss $\mathcal{L}_{\mathrm{SS}}$ is defined as $\mathcal{L}_{\mathrm{SS}} = \mathcal{L}_{\mathrm{WS}} + \mathcal{L}_{\mathrm{FS}}$ , where $\mathcal{L}_{\mathrm{FS}}$ consists of a regression loss and an attention loss and is calculated on the labeled samples in the inference branch. + +Model Inference. As mentioned, the augmentation branch and conceptual semantic connector are discarded, while other inference branch modules are preserved for testing. + +# 4. Experiment + +# 4.1. Datasets and Metrics + +ActivityNet-Captions. ActivityNet-Captions [34] dataset is a large-scale dataset with diverse open-domain content sourced from ActivityNet dataset [34]. There are 14,926 videos and 19,811 localized video-paragraph pairs in total. Each video lasts for 117.60 seconds and each paragraph consists of 3.63 sentences on average. The entire dataset is divided into train/val_1/val_2 sets containing + +10,009/4,917/4,885 video-paragraph pairs, respectively. We follow prior works [4, 31] to use val_2 set for testing. + +Charades-CD-OOD. Charades-STA dataset [22] is built from the Charades dataset [63] with indoor activities. Following the previous work [31], we adopt a reorganized version of Charades-STA named Charades-CD-OOD proposed in [92]. It is divided into train/val/test_ood sets consisting of 4,564/333/1,440 video-paragraph pairs, respectively. Specifically, the average video duration is 30.78 seconds and the average paragraph length is 2.41 sentences. + +TACoS. TACoS dataset [57] is constructed from the MPII corpus [61] tailored for cooking activities and kitchen scenarios. There are 127 videos in total with each video paired with multiple paragraphs at different granularities. Concretely, there are 1,107,418, and 380 video-paragraph pairs for training, validation, and testing, respectively. The average video length and number of sentences in the paragraph are 4.79 minutes and 8.75 in this dataset, respectively. + +Evaluation Metrics. Following previous works [4, 31], we adopt mean Intersection over Union (i.e., mIoU) and recall under IoU threshold of $m$ (i.e., $\mathbf{R}@\mathbf{m}$ ) as our evaluation metrics. The metrics are averaged over all sentences and $m$ is set to be $\{0.3, 0.5, 0.7\}$ for ActivityNet-Captions and Charades-CD-OOD, and $\{0.1, 0.3, 0.5\}$ for TACoS. + +# 4.2. Implementation Details + +For fair comparison with existing works [4, 62], we adopt the same C3D network [70] and Glove model [56] as feature extractors. The number of sampled video clips $T$ is set to be 256, 128, and 512 for ActivityNet-Captions, Charades-CD-OOD, and TACoS datasets, respectively. We train the model using Adam [32] optimizer with a fixed learning rate of 0.0001 and a batch size of 32, 32, and 16 for ActivityNet-Captions, Charades-CD-OOD and TACoS, respectively. We select top-100 high-frequency concepts for each dataset, and the loss weights $\{\lambda_{\mathrm{screg}}, \lambda_{\mathrm{oga}}, \lambda_{\mathrm{csc}}\}$ are set to $\{2, 1, 10\}$ . The number of encoder and decoder layers is set to be 3, and the hidden size $D$ is 256 in all settings. + +# 4.3. Comparison with State-of-the-arts + +We compare the proposed SiamGTR with existing state-of-the-art methods for VPG to demonstrate the superiority of our framework. Specifically, 3D-TPN [4, 101], DepNet [4], PRVG [62], SVPTR [31] and HSCNet [67] are fully-supervised approaches requiring temporal annotations for the entire dataset. Besides, the semi-supervised setting has been studied in [31] with several methods developed. For fair comparison with our method, we regard the reconstruction learning method WSSL [18] as one baseline and further develop a more competitive model called Weakly-Supervised Temporal Paragraph Network (WSTPN) by incorporating Beam Search [20] into WSTAN [75]. Specifically, WSTPN utilizes a complete paragraph for multiple + +Table 1. Comparison on ActivityNet-Captions dataset. + +
MethodSettingR@0.3R@0.5R@0.7mIoU
3D-TPN [101]FS67.5651.4930.92-
DepNet [4]FS72.8155.9133.46-
PRVG [62]FS78.2761.1537.8355.62
SVPTR [31]FS78.0761.7038.3655.91
HSCNet [67]FS81.8966.5744.0359.71
DepNet [4]SS61.4645.1426.7844.11
VPTR [31]SS72.8053.1429.0750.08
SVPTR [31]SS73.3956.7232.7851.98
SiamGTR (Ours)SS78.7559.1134.1254.57
WSSL [18]WS41.9823.34-28.33
WSTPN [75]WS57.7433.0213.6238.54
SiamGTR (Ours)WS75.4357.2330.5652.32
+ +Table 2. Comparison on Charades-CD-OOD dataset. + +
MethodSettingR@0.3R@0.5R@0.7mIoU
DepNet [4]FS45.6127.5910.6929.30
STLG [48]FS48.3030.399.79-
SVPTR [31]FS55.1432.4415.5336.01
DepNet [4]SS43.0325.0710.1428.09
STLG [48]SS46.1529.439.38-
VPTR [31]SS45.1324.9810.2228.92
SVPTR [31]SS50.3128.5012.2732.13
SiamGTR (Ours)SS59.0735.4714.9538.87
WSSL [18]WS35.8623.678.27-
WSTPN [75]WS48.6129.2710.7933.49
SiamGTR (Ours)WS57.3333.8712.3137.21
+ +instance learning and searches the best sequence of timestamps with the highest overall confidence while maintaining a consistent temporal order with the input sentences. + +Comprehensive results over three different datasets are shown in Table 1, Table 2, and Table 3, where FS/SS/WS are used to indicate fully/semi/weakly-supervised settings of video paragraph grounding, respectively. First of all, our SiamGTR remarkably surpasses all the other methods under the same supervision in all metrics over the three datasets. Concretely, our framework outperforms WSTPN by $13.78\%$ , $3.72\%$ , and $11.16\%$ in mIoU on ActivityNetCaptions, Charades-CD-OOD and TACoS datasets, respectively. Compared to semi-supervised methods using a considerable number of temporal labels, our weakly-supervised method is also able to achieve comparable or even better results, which demonstrates the effectiveness of our framework in efficient weakly-supervised learning. Furthermore, our framework is flexible and can be easily adapted to semi-supervised learning for further gains. As shown, our semi-supervised model outstrips all semi-supervised state-of-the-art's by a large margin, and it performs better or on par with the fully-supervised SVPTR on all three datasets. + +# 4.4. Ablation Study + +We conduct ablation studies to investigate the contributions of different components on ActivityNet-Captions dataset. + +Effectiveness of data augmentation. To evaluate the in + +Table 3. Comparison on TACoS dataset. + +
MethodSettingR@0.1R@0.3R@0.5mIoU
3D-TPN [101]FS55.0540.3126.54-
DepNet [4]FS56.1041.3427.16-
PRVG [62]FS61.6445.4026.3729.18
SVPTR [31]FS67.9147.8928.2231.42
HSCNet [67]FS76.2859.7442.0040.61
DepNet [4]SS40.2726.9516.5418.68
VPTR [31]SS61.3140.5921.3926.59
SVPTR [31]SS63.0640.1920.0526.10
SiamGTR (Ours)SS67.3049.3531.6932.81
WSTPN [75]WS28.5910.044.769.32
SiamGTR (Ours)WS61.5126.2210.5320.48
+ +Table 4. Ablation studies on component designs of our framework. Experimental results are marked from ID (a) $\sim$ (l). RBS and RRS respectively denote the random boundary shifting and random resampling operations for pseudo data generation. MPE, CSC and DAB stand for the modulated positional encodings, conceptual semantic connector, and dynamic anchor boxes, respectively. + +
IDRBSRRSMPECSCDABR@0.5mIoU
(a)✓✓✓47.1046.04
(b)✓✓✓✓48.1946.96
(c)✓✓✓✓54.2450.94
(d)✓✓✓✓✓57.2352.32
(e)✓✓45.2544.46
(f)✓✓✓46.4745.47
(g)✓✓✓46.9645.23
(h)✓✓✓51.3448.52
(i)✓✓✓✓52.1248.75
(j)✓✓✓✓53.5750.06
(k)✓✓✓✓54.0950.84
(l)✓✓✓✓✓57.2352.32
+ +fluences of the random boundary shifting and random resampling operations for data augmentation, we remove one or both of them from the training pipeline and the results are shown in Table 4 (a) $\sim$ (d). The model performance clearly degrades after the removal, which demonstrates the necessity of increasing sample diversity and alleviating overfitting for weakly-supervised cross-modal regression learning. + +Ablation on module designs. As shown in Table 4 (e) $\sim$ (l), we conduct detailed ablation studies on module designs to validate the rationality of the proposed model. As observed, the designs of modulated positional encodings and dynamic anchor boxes in the encoder and decoder are consistently beneficial to improving the model capacity for better performance. The conceptual semantic connector that bridges two types of queries is also effective, and it further boosts the performance by $2.26\%$ in mIoU (50.06% vs. 52.32%) even though the network has been equipped with strong encoders and decoders. We notice the dynamic anchors for decoding are the most crucial component given the sharpest performance drop of $3.57\%$ with its removal, which indicates the importance to explicitly represent the intermediate location information for video grounding. + +Table 5. Ablation studies on different weakly-supervised losses. + +
ID\( {\mathcal{L}}_{\text{screg }} \)\( {\mathcal{L}}_{\text{oga }} \)R@0.3R@0.5R@0.7mIoU
(a)45.8529.8210.9330.97
(b)✓46.9029.3812.0631.82
(c)✓73.5854.5828.8250.62
(d)✓✓75.4357.2330.5652.32
+ +Table 6. Impact of different auxiliary losses. + +
MethodR@0.3R@0.5R@0.7mIoU
w/o Lcb71.4949.7824.6448.07
w/o Lar72.9650.7226.4249.14
w/o Lpa74.3456.0130.2951.76
Full Model75.4357.2330.5652.32
+ +Table 7. Evaluation on different types of paragraph representation. + +
MethodR@0.3R@0.5R@0.7mIoU
Max-Pooling72.6152.6127.0349.55
Mean-Pooling74.3153.1328.6450.57
Word Concat73.7952.2727.3049.96
Learnable75.4357.2330.5652.32
+ +Analysis on weakly-supervised losses. Ablation studies on contributions of two weakly-supervised losses $\mathcal{L}_{\mathrm{sreg}}$ and $\mathcal{L}_{\mathrm{oga}}$ are shown in Table 5. In experiment (a) and (c), we remove $\mathcal{L}_{\mathrm{sreg}}$ but use a plain regression loss $\mathcal{L}_{\mathrm{reg}}$ for ablation analysis. Firstly, simply employing regressive supervision attains inferior performance because the feature-level cross-modal correspondence can hardly be learned with coordinate-level supervision. Furthermore, we find $\mathcal{L}_{\mathrm{oga}}$ is critical for learning precise temporal localization since it explicitly guides the model to align video features and language features that are highly likely to be correlated. Besides, we observe that using $\mathcal{L}_{\mathrm{sreg}}$ for selecting high-quality regression samples always brings gains to the performance. Impact of auxiliary losses. We investigate the influences of auxiliary losses on the model performance, which involve $\mathcal{L}_{\mathrm{cb}}$ for exploiting the cross-branch knowledge, $\mathcal{L}_{\mathrm{ar}}$ as guidance to learn a set of order-preserving anchor boxes and $\mathcal{L}_{\mathrm{pa}}$ to make use of the feature alignment supervision from the augmentation branch. As presented in Table 6, all three auxiliary losses bring positive impacts, with $\mathcal{L}_{\mathrm{cb}}$ being the most effective to improve the mIoU metric by $4.25\%$ . The reason might lie in the rich complementary knowledge and consistency supervision across the siamese branches. + +Impact of the paragraph representation. The comparison of different types of paragraph representation is shown in Table 7. Four different schemes are included, i.e., mean-pooling or max-pooling the sentence features, embedding the sequence of all word tokens in the paragraph, and using a learnable query token to extract global information. It is clear that adaptively learning a paragraph representation with our method achieves the best performance and the max-pooling scheme performs the worst because it loses too much detailed information. The mean-pooling scheme performs slightly better than the word-concat scheme, which may attribute to the advantage of late fusion at feature level. + +![](images/d70722cea3e2ab9e991d32d0a59993de8774329bdbe587033a5dae94b7b101c5.jpg) +① A woman is sitting down on a floral glass table drilling a design into the pumpkin. ② As she is drilling, two boys are standing next to her watching her and then they suddenly leave. ③ The person behind the camera then picks up the top of the pumpkin to show its empty contents before the lady closes it and continues drilling. + +![](images/c4d57aea25de7fce1ea6fb500717e1d4e5a760b6a6c79aa12cb69df43b98ca9e.jpg) +A woman is seen standing in a circle and looking over her shoulder. She throws an object off into the distance and is shown again. 3 She throws her arms up and walks away afterwards. +Figure 4. Visualization of prediction results from different models. + +# 4.5. Visualization + +In Figure 4, we intuitively visualize the predicted timestamps from WSTPN and our proposed model trained under the weakly-supervised or semi-supervised setting. Overall, the predicted timestamps given by WSTPN coarsely capture the sentence order relations but with inaccurate boundaries. In contrast, our weakly-supervised model achieves much better results. It is notable that our semi-supervised model generates more fine-grained boundaries, which shows the advantage of our siamese framework in jointly leveraging fewer and weaker labels for efficient learning. + +# 5. Conclusion + +In this work, we explore the weakly-supervised setting in video paragraph grounding (i.e., WSVPG) to eliminate the dependence of the temporal annotations. To achieve this goal, we propose a novel siamese learning framework to jointly learn the cross-modal feature alignment and temporal coordinate regression without ground-truth supervision. Specifically, we design a novel Siamese Grounding Transformer (SiamGTR) consisting of an augmentation branch and an inference branch. The augmentation branch utilizes the boundary supervision provided by temporally regressing a complete paragraph in a pseudo video, and the inference branch learns the order-guided cross-modal correspondence of multiple sentences in a normal video. Extensive experiments verify the effectiveness of our framework. + +Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009), and Guangzhou Science and Technology Plan Project (202201011134). + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 1, 2 +[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1 +[4] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, 2021. 1, 3, 6, 7 +[5] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV Workshops, 2016. 3 +[6] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a" siamese" time delay neural network. In NeurIPS, 1993. 3 +[7] Meng Cao, Long Chen, Mike Zheng Shou, Can Zhang, and Yuexian Zou. On pursuit of designing multi-modal transformer for video grounding. In EMNLP, 2021. 2 +[8] Meng Cao, Fangyun Wei, Can Xu, Xiubo Geng, Long Chen, Can Zhang, Yuexian Zou, Tao Shen, and Daxin Jiang. Iterative proposal refinement for weakly-supervised video grounding. In CVPR, 2023. 2 +[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 4 +[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3 +[11] Jiaming Chen, Weixin Luo, Wei Zhang, and Lin Ma. Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In AAAI, 2022. 2, 3 +[12] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2 +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3 +[14] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020. +[15] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3 +[16] Kyunghyun Cho, B van Merrienboer, Caglar Gulcehre, F Bougares, H Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP, 2014. 3 + +[17] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In ACMMM, 2022. 1 +[18] Xuguang Duan, Wenbing Huang, Chuang Gan, Jingdong Wang, Wenwu Zhu, and Junzhou Huang. Weakly supervised dense event captioning in videos. In NeurIPS, 2018. 6, 7 +[19] Yazan Abu Farha and Jurgen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 1 +[20] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 6 +[21] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020. 1 +[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 1, 2, 6 +[23] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022. 1 +[24] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. arXiv preprint arXiv:2305.14344, 2023.3 +[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3, 6 +[26] Jiabo Huang, Yang Liu, Shaogang Gong, and Hailin Jin. Cross-sentence temporal and semantic relations in video activity localisation. In ICCV, 2021. 2 +[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In CVPR, 2020. 1 +[28] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In CVPR, 2023. 2 +[29] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, 2023. 2 +[30] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, 2017. 1 +[31] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, 2022. 3, 6, 7 +[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[33] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Workshops, 2015. 3 +[34] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 6 + +[35] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In CVPR, 2022. 1 +[36] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, 2020. 1 +[37] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, 2017. 1 +[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 1 +[39] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 5 +[40] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In ICCV, 2023. 2 +[41] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 1 +[42] Zhijie Lin, Zhou Zhao, Zhu Zhang, Qi Wang, and Huasheng Liu. Weakly-supervised video moment retrieval via semantic completion network. In AAAI, 2020. 2 +[43] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In ACMMM, 2022. 2 +[44] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In AAAI, 2022. 2 +[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 1 +[46] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In ICLR, 2021. 4, 5 +[47] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In EMNLP, 2019. 2 +[48] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. TMM, 2022. 7 +[49] Minuk Ma, Sunjae Yoon, Junyeong Kim, Youngjoon Lee, Sunghun Kang, and Chang D Yoo. Vlanet: Video-language alignment network for weakly-supervised video moment retrieval. In ECCV, 2020. 2 +[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1 +[51] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 5 + +[52] Niluthpol Chowdhury Mithun, Sujoy Paul, and Amit K Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2 +[53] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 2 +[54] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. Bassl: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 3 +[55] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 1 +[56] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5, 6 +[57] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 2013. 6 +[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5 +[59] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 1 +[60] Mrigank Rochan, Linwei Ye, and Yang Wang. Video summarization using fully convolutional sequence networks. In ECCV, 2018. 1 +[61] Marcus Rohrbach, Michaela Regneri, Mykhaylo Andriluka, Sikandar Amin, Manfred Pinkal, and Bernt Schiele. Script data for attribute-based recognition of composite activities. In ECCV, 2012. 6 +[62] Fengyuan Shi, Limin Wang, and Weilin Huang. End-to-end dense video grounding via parallel regression. arXiv preprint arXiv:2109.11265, 2021. 3, 6, 7 +[63] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 6 +[64] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. Vlg-net: Video-language graph matching network for video grounding. In CVPR, 2021. 2 +[65] Yijun Song, Jingwen Wang, Lin Ma, Zhou Yu, and Jun Yu. Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048, 2020. 2 +[66] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In CVPR, 2014. 3 +[67] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Wei-Shi Zheng, and Jianhuang Lai. Hierarchical semantic correspondence networks for video paragraph grounding. In CVPR, 2023. 3, 6, 7 +[68] Reuben Tan, Huijuan Xu, Kate Saenko, and Bryan A Plummer. Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In WACV, 2021. 2 + +[69] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In CVPR, 2023. 3 +[70] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 3, 6 +[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4, 5 +[72] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. In AAAI, 2021. 1 +[73] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In CVPR, 2021. 2 +[74] Jingwen Wang, Lin Ma, and Wenhao Jiang. Temporally grounding language queries in videos by contextual boundary-aware prediction. In AAAI, 2020. 2 +[75] Yuechen Wang, Jiajun Deng, Wengang Zhou, and Houqiang Li. Weakly supervised temporal adjacent network for language grounding. TMM, 2022. 2, 6, 7 +[76] Yunxiao Wang, Meng Liu, Yinwei Wei, Zhiyong Cheng, Yinglong Wang, and Liqiang Nie. Siamese alignment network for weakly supervised video moment retrieval. TMM, 2023. +[77] Zheng Wang, Jingjing Chen, and Yu-Gang Jiang. Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In ACMMM, 2021. 2 +[78] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2 +[79] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In AAAI, 2020. 2 +[80] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In CVPR, 2023. 1 +[81] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3 +[82] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 1 +[83] Shuwen Xiao, Zhou Zhao, Zijian Zhang, Xiaohui Yan, and Min Yang. Convolutional hierarchical attention network for query-focused video summarization. In AAAI, 2020. 1 +[84] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. In EMNLP, 2021. 2 +[85] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In AAAI, 2021. 2 +[86] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In AAAI, 2019. 2 + +[87] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021. 3 +[88] Wenfei Yang, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Local correspondence network for weakly supervised temporal sentence grounding. TIP, 2021. 2 +[89] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 1 +[90] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In NeurIPS, 2019. 2 +[91] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In AAAI, 2019. 2, 6 +[92] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In ACMMM Workshops, 2021. 6 +[93] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, 2020. 2 +[94] Chenchi Zhang, Wenbo Ma, Jun Xiao, Hanwang Zhang, Jian Shao, Yueting Zhuang, and Long Chen. Vl-nms: Breaking proposal bottlenecks in two-stage visual-language matching. TOMM, 2023. 1 +[95] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, 2019. 2 +[96] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 2 +[97] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. TPAMI, 2021. 2 +[98] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 5 +[99] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 1 +[100] Mingxing Zhang, Yang Yang, Xinghan Chen, Yanli Ji, Xing Xu, Jingjing Li, and Heng Tao Shen. Multi-stage aggregated transformer network for temporal language localization in videos. In CVPR, 2021. 2 +[101] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 2, 6, 7 +[102] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. TPAMI, 2021. 2 + +[103] Yimeng Zhang, Xin Chen, Jinghan Jia, Sijia Liu, and Ke Ding. Text-visual prompting for efficient 2d temporal video grounding. In CVPR, 2023. 2 +[104] Zhu Zhang, Zhijie Lin, Zhou Zhao, Jieming Zhu, and Xiuqiang He. Regularized two-branch proposal networks for weakly-supervised moment retrieval in videos. In ACMMM, 2020. 2 +[105] Zhu Zhang, Zhou Zhao, Zhijie Lin, Xiuqiang He, et al. Counterfactual contrastive learning for weakly-supervised vision-language grounding. In NeurIPS, 2020. +[106] Minghang Zheng, Yanjie Huang, Qingchao Chen, and Yang Liu. Weakly supervised video moment localization with contrastive negative sample mining. In AAAI, 2022. 2 +[107] Minghang Zheng, Yanjie Huang, Qingchao Chen, Yuxin Peng, and Yang Liu. Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In CVPR, 2022. 2 +[108] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2021. 3 \ No newline at end of file diff --git a/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/images.zip b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..97df614ecb494f47e1ceb6a3c51fbcc2ec40d020 --- /dev/null +++ b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01d362af7d976edc6e914a36df7d3675ac5e845ccf3c55a0561b84215b2a15ff +size 560458 diff --git a/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/layout.json b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ed1e1e24c584c4bf43e2f67939aba40db9b6edea --- /dev/null +++ b/2024/Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding/layout.json @@ -0,0 +1,11897 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 479, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 479, + 140 + ], + "type": "text", + "content": "Siamese Learning with Joint Alignment and Regression for Weakly-Supervised Video Paragraph Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "text", + "content": "Chaolei Tan" + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "text", + "content": " Jianhuang Lai" + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "text", + "content": " Wei-Shi Zheng" + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "text", + "content": " Jian-Fang Hu" + }, + { + "bbox": [ + 104, + 160, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 176, + 496, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 176, + 496, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 496, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 176, + 496, + 189 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 189, + 507, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 189, + 507, + 203 + ], + "spans": [ + { + "bbox": [ + 104, + 189, + 507, + 203 + ], + "type": "text", + "content": "2Guangdong Province Key Laboratory of Information Security Technology, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 203, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 203, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 70, + 203, + 545, + 217 + ], + "type": "text", + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 220, + 557, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 557, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 557, + 232 + ], + "type": "text", + "content": "tanchlei@mail2.sysu.edu.cn, stsljh@mail.sysu.edu.cn, wszheng@ieee.org, hujf5@mail.sysu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 238, + 192, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 238, + 192, + 250 + ], + "spans": [ + { + "bbox": [ + 143, + 238, + 192, + 250 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 262, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 290, + 586 + ], + "type": "text", + "content": "Video Paragraph Grounding (VPG) is an emerging task in video-language understanding, which aims at localizing multiple sentences with semantic relations and temporal order from an untrimmed video. However, existing VPG approaches are heavily reliant on a considerable number of temporal labels that are laborious and time-consuming to acquire. In this work, we introduce and explore Weakly-Supervised Video Paragraph Grounding (WSVPG) to eliminate the need of temporal annotations. Different from previous weakly-supervised grounding frameworks based on multiple instance learning or reconstruction learning for two-stage candidate ranking, we propose a novel siamese learning framework that jointly learns the cross-modal feature alignment and temporal coordinate regression without timestamp labels to achieve concise one-stage localization for WSVPG. Specifically, we devise a Siamese Grounding Transformer (SiamGTR) consisting of two weight-sharing branches for learning complementary supervision. An Augmentation Branch is utilized for directly regressing the temporal boundaries of a complete paragraph within a pseudo video, and an Inference Branch is designed to capture the order-guided feature correspondence for localizing multiple sentences in a normal video. We demonstrate by extensive experiments that our paradigm has superior practicability and flexibility to achieve efficient weakly-supervised or semi-supervised learning, outperforming state-of-the-art methods trained with the same or stronger supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 593, + 128, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 593, + 128, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 593, + 128, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 612, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 287, + 696 + ], + "type": "text", + "content": "Natural Language Video Grounding (NLVG) is an essential area in vision-language understanding, which has received increasing attention due to its wide range of real-world applications such as video retrieval [3, 17, 21, 23, 80, 94], video summarization [50, 55, 59, 60, 83, 99], action segmentation [19, 27, 35, 37, 45, 72], video question answering [30, 36, 38, 41, 82, 89], etc. Most previous works focus" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 238, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 310, + 238, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 310, + 238, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 310, + 238, + 545, + 311 + ], + "type": "image", + "image_path": "3a07f72f59db4d4e9d5779d04b2bdc7fd7225047d51fddad602d14601e61d026.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 308, + 432, + 316 + ], + "lines": [ + { + "bbox": [ + 420, + 308, + 432, + 316 + ], + "spans": [ + { + "bbox": [ + 420, + 308, + 432, + 316 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 317, + 544, + 409 + ], + "blocks": [ + { + "bbox": [ + 309, + 317, + 544, + 409 + ], + "lines": [ + { + "bbox": [ + 309, + 317, + 544, + 409 + ], + "spans": [ + { + "bbox": [ + 309, + 317, + 544, + 409 + ], + "type": "image", + "image_path": "dcd0c75ad0246d487780dbf293a74255f905341c913c5972890fa3ecc999c344.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 410, + 546, + 444 + ], + "lines": [ + { + "bbox": [ + 305, + 410, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 546, + 444 + ], + "type": "text", + "content": "Figure 1. (a) Chronological cross-modal alignment in a video and its paired sentences. (b) Pseudo boundary supervision for regressing paragraph timestamps in a composed pseudo video." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "text", + "content": "on tackling Video Sentence Grounding (VSG) proposed in [1, 22], which targets at localizing the temporal boundaries of an individual sentence from an untrimmed video. However, localizing single sentences can be ambiguous because the contextual information conveyed by multiple sentences is necessary to uniquely determine the temporal locations of input queries. To alleviate such issue, Bao et al. [4] proposed to contextualize video grounding into localizing multiple events indicated by sentences of a paragraph from the video, which is called Video Paragraph Grounding (VPG)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 714 + ], + "type": "text", + "content": "Although remarkable progress has been attained in tackling VSG and VPG problems under a fully-supervised setting, the extremely prohibitive overheads of manually annotating temporal boundaries for language queries limit these methods to utilize large-scale video-text data. In addition, the subjectivity of annotators inevitably brings label noises that may adversely affect model training. Recently, weakly-supervised methods free of temporal annotations have become increasingly popular in the area of video grounding, aiming to address the above limitations. These approaches can be mainly categorized into multiple instance learning methods and reconstruction learning methods and are typ-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13569" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "ically based on a propose-and-rank pipeline. Nevertheless, both of these paradigms assume the contributions of proposals to the contrastive or reconstruction loss accurately represent the proposal quality, which is not necessarily the case during model learning. Moreover, the quadratic complexity of proposal schemes prevents them from scaling up in parameters or training data, and the adopted supervision of video-level contrast or query-based reconstruction in prior works is not temporal-sensitive thus suffering from a huge gap with fully-supervised temporal guidance. Besides, all of the weakly-supervised methods in video grounding are tailored for tackling VSG while the weakly-supervised setting of VPG (i.e., WSVPG) has been understudied so far." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 445 + ], + "type": "text", + "content": "To circumvent the aforementioned drawbacks and explore an efficient weakly-supervised framework for VPG, we seek to mine the unique characteristics and underlying supervision from the intrinsic structure of video-paragraph pairs for model training. On the one hand, as observed in Figure 1 (a), the temporal location of an event is highly correlated with the position of the sentence describing that event in the paragraph. For example, the sentence appearing in the middle of the paragraph tends to have stronger relevance with visual content located around the temporal midpoint of the video. On the other hand, dense visual events mentioned in the paragraph approximately represent the global video content that unambiguously distinguishes itself from another video, which can be observed in Figure 1 (b). Therefore, inserting the query-related video into another irrelevant video automatically generates pseudo boundary labels close to the ground-truth when regarding the complete paragraph as language query for video grounding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 448, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 289, + 674 + ], + "type": "text", + "content": "Motivated by the above observation, we propose a novel Siamese Grounding TRansformer (SiamGTR) for WSVPG. It jointly learns the cross-modal alignment and boundary regression via two siamese branches without generating proposals. Specifically, we propose to construct an Augmentation Branch (AB) which takes as input a pseudo video and adopts a complete paragraph as the language query to learn high-quality boundary supervision for localization. Also, an Inference Branch (IB) is designed to receive a normal video as input and is enforced to capture the order-guided cross-modal correspondence for attending over the specific video content relevant to each sentence. The two weight-sharing branches are effective to transfer complementary supervision for joint boundary prediction and feature association, which yields a weakly-supervised model with superior generalization through a concise one-stage pipeline. Extensive experiments verify the effectiveness of our model and show our method with the same or weaker supervision surpasses prior state-of-the-arts. In summary, our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": "- We introduce the task of Weakly-Supervised Video Paragraph Grounding (WSVPG), which aims to train a model for localizing multiple events indicated by queries with" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "type": "text", + "content": "out the supervision of timestamp labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 85, + 547, + 204 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 85, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 85, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 547, + 167 + ], + "type": "text", + "content": "- We propose a novel Siamese Grounding TRansformer (SiamGTR) for concise and efficient one-stage weakly-supervised learning of video paragraph grounding. It is composed of two weight-sharing branches including an Augmentation Branch (AB) for learning boundary regression of pseudo boundaries and an Inference Branch (IB) for learning order-guided cross-modal feature alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 168, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 547, + 204 + ], + "type": "text", + "content": "- Extensive experiments verify the efficacy of our method, and demonstrate that our framework under the same or weaker supervision outperforms state-of-the-arts." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 214, + 392, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 392, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 392, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 231, + 454, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 454, + 244 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 454, + 244 + ], + "type": "text", + "content": "2.1. Video Sentence Grounding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 510 + ], + "type": "text", + "content": "Fully-Supervised Video Sentence Grounding. Plenty of approaches [1, 7, 12, 22, 29, 40, 43, 44, 47, 53, 64, 73, 74, 78, 79, 84-86, 90, 91, 93, 95-97, 100-103] have been proposed to address Fully-Supervised Video Sentence Grounding (FSVSG). In general, these works can be roughly categorized into proposal-based and proposal-free methods. Specifically, proposal-based methods [1, 22, 40, 43, 44, 73, 74, 78, 85, 86, 90, 93, 95, 101, 102] involve a proposal generation stage using sliding windows [22], anchor proposals [12, 43, 44, 47, 74, 90, 93, 95] or 2D temporal maps [64, 73, 78, 85, 101, 102], after which the generated proposals are ranked according to the query matching scores with potential post-processing like Non-Maximum Suppression (NMS). In contrast, proposal-free methods [7, 53, 79, 91, 96, 97, 100, 103] remove the dense proposal generation and score ranking process by directly regressing timestamps [7, 53, 84, 91, 103], predicting boundary distributions [96, 97, 100] or using reinforcement learning [79], which improves the computation efficiency and scenario adaptability. Following the line of proposal-free works, we propose a novel weakly-supervised regression-based framework that shows superior performance and practicability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Weakly-Supervised Video Sentence Grounding. Weakly-Supervised Video Sentence Grounding (WSVSG) [8, 11, 26, 28, 42, 49, 52, 65, 68, 75-77, 88, 104-107] has become a popular research area because of the severe dependence of FSVSG approaches on laborious and expensive manual temporal annotations. Most of existing WSVSG methods are based on a two-stage pipeline using multiple instance learning [26, 49, 75, 88], reconstruction learning [42], or the combination of both [106, 107]. In particular, Chen et al. [11] have proposed a video composition strategy to generate pseudo temporal labels for WSVSG, which is the most related work to ours. However, several inherent drawbacks are involved in this approach. Firstly, individual sentences only describe local video content, thus viewing the starting/ending locations of foreground video as temporal boundaries of an individual sentence produces a weak and noisy temporal alignment, which is unsuitable for accurate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13570" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "boundary supervision. Moreover, it simply adapts an existing fully-supervised proposal-based framework for weakly-supervised training, which leads to inferior generalization caused by the large train-test discrepancy. Distinct from all of the above works, we design a novel siamese framework to capture the essential characteristics of video paragraph grounding for weakly-supervised learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 205, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 205, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 205, + 177 + ], + "type": "text", + "content": "2.2. Video Paragraph Grounding" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 290, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 290, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 290, + 408 + ], + "type": "text", + "content": "Video Paragraph Grounding (VPG) is introduced by Bao et al. [4], which aims to jointly localize multiple sentences of a paragraph from an untrimmed video. Shi et al. [62] presented an end-to-end network by re-purposing transformers into language-conditioned regressors. Jiang et al. [31] proposed to employ contrastive encoders for contrastive learning between video-paragraph pairs. Tan et al. [67] proposed a hierarchical semantic correspondence network for modeling hierarchical video-language alignment and grounding multiple levels of language queries in the video. Particularly, Jiang et al. [31] first explored Semi-Supervised Video Paragraph Grounding (SSVPG) to relieve the annotation burden of temporal labels. However, SSVPG is still not quite practical considering the expensive cost of temporal annotations in untrimmed videos. Besides, these semi-supervised methods still require a considerable proportion of temporal labels up to at least " + }, + { + "bbox": [ + 46, + 180, + 290, + 408 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 180, + 290, + 408 + ], + "type": "text", + "content": " [31] for training. To thoroughly get rid of the temporal annotations, we pioneer to explore the weakly-supervised setting in VPG." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 155, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 155, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 155, + 426 + ], + "type": "text", + "content": "2.3. Siamese Networks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 432, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 288, + 601 + ], + "type": "text", + "content": "Siamese networks [6] are weight-sharing neural networks. There have been a wide range of scenarios where siamese networks are applied for achieving different purposes, such as face verification [66], image recognition [33], object tracking [5], etc. In particular, siamese networks are commonly used in contrastive self-supervised learning methods [2, 10, 13-15, 24, 25, 69, 81, 108], in which augmented views of the same or different instance are forwarded into multiple weight-sharing network copies for learning a generalizable visual representation via instance-level discrimination. In this work, we explore a new way to combine the transferability of siamese architectures and the flexibility of transformer architectures for concise and efficient weakly-supervised learning of video paragraph grounding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 613, + 129, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 129, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 129, + 627 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 115, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 115, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 115, + 645 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": "Task Formulation. Given an untrimmed video " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " and a paragraph query " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " temporally ordered sentences " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{S}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": ", the goal of Video Paragraph Grounding (VPG) is to simultaneously localize the temporal intervals " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{(\\tau_i^{\\mathrm{st}}, \\tau_i^{\\mathrm{ed}})\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " of all the events described by sentences" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 337, + 70, + 434, + 167 + ], + "blocks": [ + { + "bbox": [ + 337, + 70, + 434, + 167 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 434, + 167 + ], + "type": "image", + "image_path": "2e2781a658a58acd039e4de131bab5d48e8d0bb526f09b19cb8e008e7e33bfc8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 168, + 538, + 179 + ], + "lines": [ + { + "bbox": [ + 312, + 168, + 538, + 179 + ], + "spans": [ + { + "bbox": [ + 312, + 168, + 538, + 179 + ], + "type": "text", + "content": "Figure 2. Our siamese framework for joint alignment and regression." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 449, + 70, + 514, + 167 + ], + "blocks": [ + { + "bbox": [ + 449, + 70, + 514, + 167 + ], + "lines": [ + { + "bbox": [ + 449, + 70, + 514, + 167 + ], + "spans": [ + { + "bbox": [ + 449, + 70, + 514, + 167 + ], + "type": "image", + "image_path": "71bc0739af9c28e8c4279d701bbd3063cc32d1ac2149ffe969a1d6ab057af9f0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "text", + "content": "in the paragraph, where " + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\tau_{i}^{\\mathrm{st}}" + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\tau_{i}^{\\mathrm{ed}}" + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "text", + "content": " are the starting and ending timestamps for the " + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "text", + "content": "-th sentence " + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 304, + 182, + 545, + 207 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 207, + 546, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 546, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 546, + 362 + ], + "type": "text", + "content": "Siamese Learning. An overview of our siamese framework is shown in Figure 2. Overall, we jointly train an augmentation branch and an inference branch with shared parameters and structures for learning the complementary abilities of cross-modal feature alignment and temporal boundary regression. These two branches follow the same workflow to first encode the text queries and input video into unimodal features, after which the query features are iteratively used in the transformer decoder to extract relevant information from the video features for timestamp decoding. For testing, we only keep the pipeline of inference branch for boundary prediction. More architectural details are illustrated in Figure 3 and are elaborated in the following sections." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 370, + 417, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 417, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 417, + 382 + ], + "type": "text", + "content": "3.2. Feature Extraction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": "Video Feature Extraction. For each input video, we divide it into consecutive clips consisting of a fixed number of frames for feature extraction. Specifically, a frozen pretrained 3D Convolutional Neural Network (3D-CNN) [70] and a linear projection layer are successively employed to obtain a 1D feature vector for each short clip, resulting in a video feature sequence " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{v}} \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " are the sequence length and hidden dimension, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": "Text Feature Extraction. For each input paragraph consisting of " + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": " sentences, we first utilize a frozen pre-trained word embedding model tookenize and embed the text into a sequence of word vectors. Then, a bidirectional Gated Recurrent Unit (GRU) [16] is employed on each sentence, and the last hidden states in both directions are concatenated and then projected by a linear layer to construct the sentence features " + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{s}}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": " is the hidden dimension." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 587, + 432, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 432, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 432, + 600 + ], + "type": "text", + "content": "3.3. Augmentation Branch" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": "The augmentation branch aims to learn accurate boundary regression from pseudo videos with paragraph queries, which naturally transfers to the inference branch via the shared feature space established by the siamese structure." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Pseudo Data Generation. To drive end-to-end weakly-supervised regression learning, the input stream of the augmentation branch should provide reliable and direct boundary supervision. Drawing inspiration from the boundary-sensitive video pretext tasks [11, 54, 87], we propose to uti" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13571" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 234 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 234 + ], + "type": "image", + "image_path": "a245b9fc51cd6f2fbd98c7cda54eff7d73eaf65fa8de3049543f3ba4f0ae5b6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 282 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed Siamese Grounding TRansformer (SiamGTR) architecture. The augmentation branch (abbreviated as A.B.) takes the pseudo video features derived from randomly inserting the query-related video features into irrelevant video features. It learns to temporally regress the interval of interest from the pseudo video with the paragraph as query. The Inference Branch (abbreviated as I.B.) receives normal video features for learning the cross-modal feature alignment among multiple sentences in the video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": "lize the synthesized videos paired with paragraph queries to serve as a well-suited source of surrogate boundary supervision. Specifically, for each input video " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " and its paragraph description " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": ", we randomly sample a background video " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " to obtain irrelevant video content to " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": ". Because " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " could be treated as a unique referring expression specific to " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": ", the search space of background videos can be the entire training set for increasing the data diversity. Denote video features of " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{v}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " respectively, we construct the pseudo video features " + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 409 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 413, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 413, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 61, + 413, + 287, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathrm {a u g}} = \\operatorname {N R S} \\left(\\operatorname {C o n c a t} \\left(\\mathcal {F} _ {\\mathrm {b g}} ^ {1: I}, \\operatorname {R R S} \\left(\\mathcal {F} _ {\\mathrm {v}}\\right), \\mathcal {F} _ {\\mathrm {b g}} ^ {I + 1:}\\right)\\right) \\tag {1}", + "image_path": "c955235c594fc9fe90083f7688e9ebcbfd2c2c514b41232d5819f519dc92dbe1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": " is a randomly generated inserting index within the feature sequence " + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": " and Concat (" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": ") denotes the temporal concatenation. RRS (" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": ") and NRS (" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": ") respectively represent a Random Re-Sampling operation to stochastically re-scale the length of " + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{v}}" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": " and a Normalized Re-Sampling operation that converts the length of " + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": " into a fixed number of " + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": ". Thereafter, we compute the synthesized temporal boundaries (" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{aug}}^{\\mathrm{st}}, \\tau_{\\mathrm{aug}}^{\\mathrm{ed}}" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": ") in the pseudo video " + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 437, + 289, + 535 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 539, + 287, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 539, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 82, + 539, + 287, + 567 + ], + "type": "interline_equation", + "content": "\\tau_ {\\mathrm {a u g}} ^ {\\mathrm {s t}} = \\frac {I + \\Delta I ^ {\\mathrm {s t}}}{r L + L _ {\\mathrm {b g}}}, \\tau_ {\\mathrm {a u g}} ^ {\\mathrm {e d}} = \\frac {I + r L - \\Delta I ^ {\\mathrm {e d}}}{r L + L _ {\\mathrm {b g}}} \\tag {2}", + "image_path": "0cd5281f2d5fc1a5d8856aaf01c710d3b563d2ab32bf89604c1d3b8190ff766d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " respectively indicate the length of " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{v}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " being the random re-scaling factor of RRS " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": ". To alleviate potential synthesis artifacts and boundary uncertainty, we further introduce a Random Boundary Shifting (RBS) strategy that incorporates small random offsets " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\Delta I^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\Delta I^{\\mathrm{ed}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " into computing the pseudo labels, which is simple yet effective to boost the quality of boundary supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "Video Encoder. Based on the obtained pseudo video features " + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": ", we then encode the temporal contextual information across multiple clips by feeding " + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": " to a video encoder, which follows a similar architecture of DETR encoders [9, 46, 71], i.e., each encoder layer consists of a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 288, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 348 + ], + "type": "text", + "content": "multi-head self-attention module equipped with additive sinusoidal positional encodings and a feed-forward network. For simplicity, we omit the layer index and denote the input of each video encoder layer as " + }, + { + "bbox": [ + 304, + 288, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 304, + 288, + 545, + 348 + ], + "type": "text", + "content": ", then a set of Modulated Positional Encodings (MPE) are given as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 352, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 352, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 361, + 352, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {\\mathrm {p e}} = \\operatorname {M L P} \\left(\\mathcal {X} _ {\\mathrm {a u g}}\\right) \\odot \\operatorname {P E} \\left(\\mathcal {X} _ {\\mathrm {a u g}}\\right) \\tag {3}", + "image_path": "2bd9d3e1dafaddc89f5288dfacc6b73c56e1217c144e673474569c0ed832fb69.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{pe}} \\in \\mathbb{R}^{T \\times D}" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": " denotes element-wise multiplication. PE " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": " is the sinusoidal function [71] with respect to temporal locations of " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": ", MLP " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": " denotes a two-layer feed-forward network computed on " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": ". In each encoder layer, " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{pe}}" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": " is only added to the input of projection layers of queries and keys in the self-attention mechanism. The encoded video features of the last video encoder layer, also called the encoder memory features " + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 304, + 368, + 546, + 476 + ], + "type": "text", + "content": ", are further fed to the query-guided decoder for localization prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 476, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 571 + ], + "type": "text", + "content": "Query Encoder. Since there are strong contextual correlations among multiple sentences in the paragraph query, here we utilize a vanilla transformer encoder [71] to help reason the semantic and chronological relationships of events. To extract a global representation of the complete paragraph, we initialize an extra learnable query token " + }, + { + "bbox": [ + 304, + 476, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{p}}" + }, + { + "bbox": [ + 304, + 476, + 545, + 571 + ], + "type": "text", + "content": " and integrate it with the sentence features " + }, + { + "bbox": [ + 304, + 476, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 476, + 545, + 571 + ], + "type": "text", + "content": " by concatenation. Thus, the input of the query encoder is constructed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 574, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 574, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 351, + 574, + 545, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {\\mathrm {q}} = \\left[ \\mathcal {F} _ {\\mathrm {p}}, \\mathcal {F} _ {\\mathrm {s}} ^ {1}, \\mathcal {F} _ {\\mathrm {s}} ^ {2}, \\dots , \\mathcal {F} _ {\\mathrm {s}} ^ {i}, \\dots \\mathcal {F} _ {\\mathrm {s}} ^ {N} \\right] \\tag {4}", + "image_path": "1ac50f6d8be376b7e1b6cf9a907b8a9dbcf55fbe671cfc21ecec2ec9aca6b648.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{q}} \\in \\mathbb{R}^{(N + 1) \\times D}" + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{s}}^{i}" + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": "-th sentence feature in " + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": ". We first normalize " + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": ", add fixed sinusoidal positional encodings to it, and then iteratively employ self-attention modules and feed-forward networks to contextualize the global and local tokens that respectively represent the query semantics of the entire paragraph and the sentences. The output features of the last query encoder layer are denoted as " + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": " and are further forwarded to the query decoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 690, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 690, + 545, + 713 + ], + "type": "text", + "content": "Conceptual Semantic Connector. Although the siamese network structure can implicitly transfer boundary knowl-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "13572" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "edge from the augmentation branch to the inference branch, there is still a certain semantic gap between the query representations of short sentences and long paragraphs. To narrow this gap, we develop a Conceptual Semantic Connector (CSC) module for explicit semantic guidance. Specifically, we first collect high-frequency linguistic concepts, including verbs and nouns from the training corpus, and then construct a set of dictionary features by selecting and projecting the Glove vectors [56]. The loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{csc}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 180, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 180, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 76, + 180, + 287, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c s c}} = \\mathrm {B C E} \\left(y _ {\\text {c p t}} ^ {\\mathrm {p}}, \\hat {y} _ {\\text {c e p t}} ^ {\\mathrm {p}}\\right) + \\mathrm {B C E} \\left(y _ {\\text {s c t}} ^ {\\mathrm {s}}, \\hat {y} _ {\\text {c e p t}} ^ {\\mathrm {s}}\\right) \\tag {5}", + "image_path": "81d694368e6f1b7c6356a5b7e50f561bf52f6d04b358d7bbdde84ea3dbc2ef37.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{BCE}(\\cdot)" + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": " is the binary cross-entropy loss. " + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{cept}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{cept}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": " are multi-hot labels indicating semantic concepts contained by the paragraph and sentence queries, respectively. " + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\mathrm{cept}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\mathrm{cept}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": " are the concept predictions obtained by dot-product between the conceptual dictionary features and the textual query features " + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathrm{q}}" + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": " with a sigmoid activation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": "Query Decoder. Inspired by the spiritual ideas of dynamic-anchor DETR decoders [39, 46, 51, 98], we design a novel query decoder that enables dynamic position-aware decoding of language queries. Specifically, each query decoder layer consists of a self-attention module, a cross-attention module, and a feed-forward network, with dynamically adjustable anchor boxes to indicate the query-specific location information. Initially, the input anchor boxes are set to all zeros, and an MLP is used to estimate a set of seed anchor boxes based on the cross-modal interactions between " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathrm{q}}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": ", i.e., the output features of the first decoder layer " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{aug}}^{(1)} \\in \\mathbb{R}^{(N + 1) \\times D}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": " are mapped into the seed anchor boxes " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{aug}}^{(1)} \\in \\mathbb{R}^{(N + 1) \\times 2}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": " by the MLP. For the " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": "-th decoder layer, we first convert the box coordinates of the input anchors " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{aug}}^{(i)}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": " into high-dimensional sinusoidal embeddings " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{a}}^{(i)} \\in \\mathbb{R}^{(N + 1) \\times D}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": " by a sinusoidal function [46, 71] and further obtain " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{a}}^{(i)}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": " by projecting " + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{a}}^{(i)}" + }, + { + "bbox": [ + 46, + 266, + 289, + 503 + ], + "type": "text", + "content": " with an MLP. Afterwards, we conduct the self-attention operation and update the query features in the decoder as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 503, + 287, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 287, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)} \\leftarrow \\text {S e l f - A t t n} \\left\\{ \\begin{array}{l} Q = \\varphi_ {q} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) + \\varphi_ {q} ^ {p} \\left(\\mathcal {H} _ {\\mathrm {a}} ^ {(i)}\\right) \\\\ K = \\varphi_ {k} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) + \\varphi_ {k} ^ {p} \\left(\\mathcal {H} _ {\\mathrm {a}} ^ {(i)}\\right) \\\\ V = \\varphi_ {v} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i)}\\right) \\end{array} \\right. \\tag {6}", + "image_path": "fc1449c22cf90bd049165a5d1f8b0f10049d0b7eda622da69a0b8d5af5f4ef90.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}" + }, + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "text", + "content": " is the updated query features after the self-attention operation in the " + }, + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "text", + "content": "-th decoder layer. The above series of " + }, + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 562, + 287, + 647 + ], + "type": "text", + "content": " functions are used to indicate different linear projection layers for the content part or position part of the queries, keys, or values in the self-attention mechanism. Then, we conduct a cross-attention operation to extract useful cross-modal interactive information as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 657, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 657, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 53, + 657, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)} \\leftarrow \\text {C r o s s - A t t n} \\left\\{ \\begin{array}{l} Q = \\left[ \\varphi_ {q} ^ {c} \\left(\\mathcal {Q} _ {\\mathrm {a u g}} ^ {(i + 1)}\\right); \\varphi_ {q} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {a}} ^ {(i)}\\right) \\right] \\\\ K = \\left[ \\varphi_ {k} ^ {c} \\left(\\mathcal {M} _ {\\mathrm {a u g}}\\right); \\varphi_ {k} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {p e}} ^ {\\mathcal {M}}\\right) \\right] \\\\ V = \\varphi_ {v} ^ {c} \\left(\\mathcal {M} _ {\\mathrm {a u g}}\\right) + \\varphi_ {v} ^ {p} \\left(\\mathcal {F} _ {\\mathrm {p e}} ^ {\\mathcal {M}}\\right) \\end{array} \\right. \\tag {7}", + "image_path": "4316398d9bb7fbff59ec55fc719d91f4e47889de1e80ee84c1aba00b4a3221e8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{pe}}^{\\mathcal{M}} = \\mathrm{PE}\\left(\\mathcal{M}_{\\mathrm{aug}}\\right)" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": " and the query features " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": " is then further updated by a feed-forward network as the feature output of the " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": "-th decoder layer. Then the anchor boxes are dynamically updated as " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{aug}}^{(i + 1)} \\gets \\mathcal{A}_{\\mathrm{aug}}^{(i)} + \\Delta \\mathcal{A}_{\\mathrm{aug}}^{(i)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": " where we utilize an MLP layer to predict the relative offsets, i.e., " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{A}_{\\mathrm{aug}}^{(i)} \\in \\mathbb{R}^{(N + 1) \\times 2}" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": ", based on the updated query features " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{aug}}^{(i + 1)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{aug}}^{(i + 1)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": " continues to be forwarded to the next decoder layer for computing " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{a}}^{(i + 1)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{a}}^{(i + 1)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 177 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 177, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 258 + ], + "type": "text", + "content": "Boundary Prediction. Based on the output features and attention weights of the last query decoder layer, we simply use an MLP predictor to predict the paragraph timestamps, i.e., " + }, + { + "bbox": [ + 304, + 177, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{T}}_{\\mathrm{aug}}^{\\mathrm{p}} = (\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{ed}})" + }, + { + "bbox": [ + 304, + 177, + 545, + 258 + ], + "type": "text", + "content": ". Similarly, the sentence timestamps " + }, + { + "bbox": [ + 304, + 177, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{T}}_{\\mathrm{inf}}^{\\mathrm{s}} = \\left\\{(\\hat{\\tau}_{\\mathrm{j}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{j}}^{\\mathrm{ed}})\\right\\}_{j = 1}^{N}" + }, + { + "bbox": [ + 304, + 177, + 545, + 258 + ], + "type": "text", + "content": " can also be obtained by feeding last-layer output from the inference branch to the same MLP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 259, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 380 + ], + "type": "text", + "content": "Self-Consistent Boundary Regression. We improve the attention-agnostic regression loss " + }, + { + "bbox": [ + 304, + 259, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 304, + 259, + 545, + 380 + ], + "type": "text", + "content": " to make it aware of the model's self-consistent scores, where the main idea is to selectively optimize the regression loss of self-consistent samples for better weakly-supervised regression learning. Self-consistent samples have high attention weights over the pseudo ground-truth intervals, which are more suitable for learning less noisy supervision for accurate boundary prediction. Specifically, the self-consistent boundary regression loss " + }, + { + "bbox": [ + 304, + 259, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sreg}}" + }, + { + "bbox": [ + 304, + 259, + 545, + 380 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 386, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 386, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 340, + 386, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s c r e g}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\mathrm {L} 1} + \\mathcal {L} _ {\\mathrm {G I o U}}, & \\text {i f} \\quad s _ {\\mathrm {a t t}} > \\beta \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}", + "image_path": "77f7333c646dfb99f0f57ce07609fc56cb942fc5aa2929f3706863351d54374e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{L1}}" + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GIoU}}" + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "content": " respectively represent L1 and Generalized Intersection over Union (GIoU) [58] loss computed between " + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "inline_equation", + "content": "(\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{st}},\\hat{\\tau}_{\\mathrm{aug}}^{\\mathrm{ed}})" + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "inline_equation", + "content": "(\\tau_{\\mathrm{aug}}^{\\mathrm{st}},\\tau_{\\mathrm{aug}}^{\\mathrm{ed}})" + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "content": " is set to 0.5 and " + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{att}}" + }, + { + "bbox": [ + 304, + 424, + 545, + 472 + ], + "type": "text", + "content": " is the attention sum over the pseudo ground-truth interval." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 479, + 410, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 410, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 410, + 491 + ], + "type": "text", + "content": "3.4. Inference Branch" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": "In our siamese framework, the inference branch shares the same parameter weights and network structure with the augmentation branch, i.e., a video encoder, a query encoder, and a query decoder. The only difference lies in the input streams and objectives, i.e., the inference branch receives a normal video during training to learn in-domain cross-modal correspondence that cannot be acquired through the pseudo video stream, which significantly improves the generalization ability of the model. Specifically, it receives the encoded normal video features " + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{inf}}" + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": " and the encoded text query features " + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": " as the decoder input and generates the hidden query features " + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{inf}}" + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": " for boundary prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Order-guided Attention Loss. The chronological prior given by the sentence order provides explicit guidance for learning cross-modal alignment between the video and language features during decoding. To learn the order-guided cross-modal correspondence, we constrain cross-modal attention weights in the decoder as follows:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13573" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 70, + 287, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 70, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 287, + 99 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o g a}} = \\max \\left(0, \\Delta m T + \\sum_ {t = 1} ^ {T} t \\alpha_ {\\mathrm {s}} ^ {j} (t) - \\sum_ {t = 1} ^ {T} t \\alpha_ {\\mathrm {s}} ^ {j + 1} (t)\\right) \\tag {9}", + "image_path": "2e6602f5df886268f9c07f3e408eef58552d88cea2be912cd56d38216c5f26f2.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{s}}^{j}(t)" + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{s}}^{j + 1}(t)" + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": " are attention weights over video features for the " + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "inline_equation", + "content": "(j + 1)" + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": "-th sentence from the last decoder layer, respectively. " + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "inline_equation", + "content": "\\Delta m" + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": " is the minimal distance between attention centroids. This loss is mainly contributed by the inference branch with " + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "inline_equation", + "content": "\\Delta m = \\frac{1}{2N}" + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": " and partly contributed by the augmentation branch with " + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "inline_equation", + "content": "\\Delta m = \\frac{1}{4N}" + }, + { + "bbox": [ + 47, + 99, + 286, + 173 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "content": "Auxiliary Losses. To exploit more guidance for weakly-supervised representation learning, we employ three auxiliary losses including a cross-branch loss " + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cb}}" + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "content": ", an anchor ranking loss " + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ar}}" + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "content": " and a pseudo attention loss " + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pa}}" + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cb}}" + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "content": " utilizes the semantic consistency constraint of output features across siamese branches, which is calculated analogous to MoCo [25]. " + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ar}}" + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "content": " is used for inducing the decoder to learn a set of order-preserving anchor boxes, which is computed on the anchor boxes like in the equation (9). " + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pa}}" + }, + { + "bbox": [ + 47, + 173, + 286, + 316 + ], + "type": "text", + "content": " makes use of attention supervision between the pseudo video and language features from the augmentation branch, which follows the calculation proposed in [62, 91]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 178, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 178, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 178, + 335 + ], + "type": "text", + "content": "3.5. Training and Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "text", + "content": "Weakly-Supervised Learning. The weakly-supervised loss of our proposed framework can be formulated as a weighted sum of the two losses from the siamese branches as " + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{WS}} = \\lambda_{\\mathrm{sreg}}\\mathcal{L}_{\\mathrm{sreg}} + \\lambda_{\\mathrm{oga}}\\mathcal{L}_{\\mathrm{oga}}" + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sreg}}" + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{oga}}" + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "text", + "content": " are scalar weights to balance the contributions of the two different losses. The overall loss for the weakly-supervised model is defined as " + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{WS}} + \\lambda_{\\mathrm{csc}}\\mathcal{L}_{\\mathrm{csc}} + \\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 47, + 341, + 286, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "text", + "content": "Semi-Supervised Learning. Although our framework is initially designed for weakly-supervised learning, it can be easily adapted for end-to-end semi-supervised learning. Specifically, in addition to " + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{WS}}" + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "text", + "content": " which is calculated on all training samples, we only need to employ an extra fully-supervised loss " + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FS}}" + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "text", + "content": " on those fully-annotated samples without changing any part of the network structure. The semi-supervised loss " + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SS}}" + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SS}} = \\mathcal{L}_{\\mathrm{WS}} + \\mathcal{L}_{\\mathrm{FS}}" + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FS}}" + }, + { + "bbox": [ + 47, + 426, + 286, + 544 + ], + "type": "text", + "content": " consists of a regression loss and an attention loss and is calculated on the labeled samples in the inference branch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 544, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 286, + 581 + ], + "type": "text", + "content": "Model Inference. As mentioned, the augmentation branch and conceptual semantic connector are discarded, while other inference branch modules are preserved for testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 591, + 123, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 123, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 123, + 604 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 611, + 168, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 168, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 168, + 622 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 286, + 713 + ], + "type": "text", + "content": "ActivityNet-Captions. ActivityNet-Captions [34] dataset is a large-scale dataset with diverse open-domain content sourced from ActivityNet dataset [34]. There are 14,926 videos and 19,811 localized video-paragraph pairs in total. Each video lasts for 117.60 seconds and each paragraph consists of 3.63 sentences on average. The entire dataset is divided into train/val_1/val_2 sets containing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "10,009/4,917/4,885 video-paragraph pairs, respectively. We follow prior works [4, 31] to use val_2 set for testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 96, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 191 + ], + "type": "text", + "content": "Charades-CD-OOD. Charades-STA dataset [22] is built from the Charades dataset [63] with indoor activities. Following the previous work [31], we adopt a reorganized version of Charades-STA named Charades-CD-OOD proposed in [92]. It is divided into train/val/test_ood sets consisting of 4,564/333/1,440 video-paragraph pairs, respectively. Specifically, the average video duration is 30.78 seconds and the average paragraph length is 2.41 sentences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 192, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 288 + ], + "type": "text", + "content": "TACoS. TACoS dataset [57] is constructed from the MPII corpus [61] tailored for cooking activities and kitchen scenarios. There are 127 videos in total with each video paired with multiple paragraphs at different granularities. Concretely, there are 1,107,418, and 380 video-paragraph pairs for training, validation, and testing, respectively. The average video length and number of sentences in the paragraph are 4.79 minutes and 8.75 in this dataset, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous works [4, 31], we adopt mean Intersection over Union (i.e., mIoU) and recall under IoU threshold of " + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{R}@\\mathbf{m}" + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "text", + "content": ") as our evaluation metrics. The metrics are averaged over all sentences and " + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "text", + "content": " is set to be " + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\{0.3, 0.5, 0.7\\}" + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "text", + "content": " for ActivityNet-Captions and Charades-CD-OOD, and " + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.3, 0.5\\}" + }, + { + "bbox": [ + 305, + 289, + 545, + 361 + ], + "type": "text", + "content": " for TACoS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 369, + 438, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 438, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 438, + 380 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "text", + "content": "For fair comparison with existing works [4, 62], we adopt the same C3D network [70] and Glove model [56] as feature extractors. The number of sampled video clips " + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "text", + "content": " is set to be 256, 128, and 512 for ActivityNet-Captions, Charades-CD-OOD, and TACoS datasets, respectively. We train the model using Adam [32] optimizer with a fixed learning rate of 0.0001 and a batch size of 32, 32, and 16 for ActivityNet-Captions, Charades-CD-OOD and TACoS, respectively. We select top-100 high-frequency concepts for each dataset, and the loss weights " + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\{\\lambda_{\\mathrm{screg}}, \\lambda_{\\mathrm{oga}}, \\lambda_{\\mathrm{csc}}\\}" + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\{2, 1, 10\\}" + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "text", + "content": ". The number of encoder and decoder layers is set to be 3, and the hidden size " + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 387, + 545, + 531 + ], + "type": "text", + "content": " is 256 in all settings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 539, + 489, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 489, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 489, + 552 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "content": "We compare the proposed SiamGTR with existing state-of-the-art methods for VPG to demonstrate the superiority of our framework. Specifically, 3D-TPN [4, 101], DepNet [4], PRVG [62], SVPTR [31] and HSCNet [67] are fully-supervised approaches requiring temporal annotations for the entire dataset. Besides, the semi-supervised setting has been studied in [31] with several methods developed. For fair comparison with our method, we regard the reconstruction learning method WSSL [18] as one baseline and further develop a more competitive model called Weakly-Supervised Temporal Paragraph Network (WSTPN) by incorporating Beam Search [20] into WSTAN [75]. Specifically, WSTPN utilizes a complete paragraph for multiple" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13574" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 83, + 286, + 229 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 266, + 81 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 266, + 81 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 266, + 81 + ], + "type": "text", + "content": "Table 1. Comparison on ActivityNet-Captions dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 83, + 286, + 229 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 286, + 229 + ], + "type": "table", + "html": "
MethodSettingR@0.3R@0.5R@0.7mIoU
3D-TPN [101]FS67.5651.4930.92-
DepNet [4]FS72.8155.9133.46-
PRVG [62]FS78.2761.1537.8355.62
SVPTR [31]FS78.0761.7038.3655.91
HSCNet [67]FS81.8966.5744.0359.71
DepNet [4]SS61.4645.1426.7844.11
VPTR [31]SS72.8053.1429.0750.08
SVPTR [31]SS73.3956.7232.7851.98
SiamGTR (Ours)SS78.7559.1134.1254.57
WSSL [18]WS41.9823.34-28.33
WSTPN [75]WS57.7433.0213.6238.54
SiamGTR (Ours)WS75.4357.2330.5652.32
", + "image_path": "e669ab1585cf9bbff748335194304b6399d417614989cd6822d85eb315a52ec3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 244, + 286, + 381 + ], + "blocks": [ + { + "bbox": [ + 71, + 232, + 263, + 243 + ], + "lines": [ + { + "bbox": [ + 71, + 232, + 263, + 243 + ], + "spans": [ + { + "bbox": [ + 71, + 232, + 263, + 243 + ], + "type": "text", + "content": "Table 2. Comparison on Charades-CD-OOD dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 244, + 286, + 381 + ], + "lines": [ + { + "bbox": [ + 48, + 244, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 286, + 381 + ], + "type": "table", + "html": "
MethodSettingR@0.3R@0.5R@0.7mIoU
DepNet [4]FS45.6127.5910.6929.30
STLG [48]FS48.3030.399.79-
SVPTR [31]FS55.1432.4415.5336.01
DepNet [4]SS43.0325.0710.1428.09
STLG [48]SS46.1529.439.38-
VPTR [31]SS45.1324.9810.2228.92
SVPTR [31]SS50.3128.5012.2732.13
SiamGTR (Ours)SS59.0735.4714.9538.87
WSSL [18]WS35.8623.678.27-
WSTPN [75]WS48.6129.2710.7933.49
SiamGTR (Ours)WS57.3333.8712.3137.21
", + "image_path": "a8d7e1eadff7d7c8d55267deddba3a7987a96b6d5c3e67067c26f4f622c70bda.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 420 + ], + "type": "text", + "content": "instance learning and searches the best sequence of timestamps with the highest overall confidence while maintaining a consistent temporal order with the input sentences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "text", + "content": "Comprehensive results over three different datasets are shown in Table 1, Table 2, and Table 3, where FS/SS/WS are used to indicate fully/semi/weakly-supervised settings of video paragraph grounding, respectively. First of all, our SiamGTR remarkably surpasses all the other methods under the same supervision in all metrics over the three datasets. Concretely, our framework outperforms WSTPN by " + }, + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "inline_equation", + "content": "13.78\\%" + }, + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "inline_equation", + "content": "3.72\\%" + }, + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "inline_equation", + "content": "11.16\\%" + }, + { + "bbox": [ + 46, + 421, + 287, + 648 + ], + "type": "text", + "content": " in mIoU on ActivityNetCaptions, Charades-CD-OOD and TACoS datasets, respectively. Compared to semi-supervised methods using a considerable number of temporal labels, our weakly-supervised method is also able to achieve comparable or even better results, which demonstrates the effectiveness of our framework in efficient weakly-supervised learning. Furthermore, our framework is flexible and can be easily adapted to semi-supervised learning for further gains. As shown, our semi-supervised model outstrips all semi-supervised state-of-the-art's by a large margin, and it performs better or on par with the fully-supervised SVPTR on all three datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 141, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 141, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 141, + 670 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 700 + ], + "type": "text", + "content": "We conduct ablation studies to investigate the contributions of different components on ActivityNet-Captions dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Effectiveness of data augmentation. To evaluate the in" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 83, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 351, + 71, + 499, + 81 + ], + "lines": [ + { + "bbox": [ + 351, + 71, + 499, + 81 + ], + "spans": [ + { + "bbox": [ + 351, + 71, + 499, + 81 + ], + "type": "text", + "content": "Table 3. Comparison on TACoS dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 83, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 545, + 220 + ], + "type": "table", + "html": "
MethodSettingR@0.1R@0.3R@0.5mIoU
3D-TPN [101]FS55.0540.3126.54-
DepNet [4]FS56.1041.3427.16-
PRVG [62]FS61.6445.4026.3729.18
SVPTR [31]FS67.9147.8928.2231.42
HSCNet [67]FS76.2859.7442.0040.61
DepNet [4]SS40.2726.9516.5418.68
VPTR [31]SS61.3140.5921.3926.59
SVPTR [31]SS63.0640.1920.0526.10
SiamGTR (Ours)SS67.3049.3531.6932.81
WSTPN [75]WS28.5910.044.769.32
SiamGTR (Ours)WS61.5126.2210.5320.48
", + "image_path": "4f063427878e236bc543ab524c52f06c811c8b02b8a85eba086ae5ac519a6b25.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 293, + 545, + 441 + ], + "blocks": [ + { + "bbox": [ + 305, + 224, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 291 + ], + "type": "text", + "content": "Table 4. Ablation studies on component designs of our framework. Experimental results are marked from ID (a) " + }, + { + "bbox": [ + 305, + 224, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 305, + 224, + 545, + 291 + ], + "type": "text", + "content": " (l). RBS and RRS respectively denote the random boundary shifting and random resampling operations for pseudo data generation. MPE, CSC and DAB stand for the modulated positional encodings, conceptual semantic connector, and dynamic anchor boxes, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 293, + 545, + 441 + ], + "lines": [ + { + "bbox": [ + 309, + 293, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 309, + 293, + 545, + 441 + ], + "type": "table", + "html": "
IDRBSRRSMPECSCDABR@0.5mIoU
(a)✓✓✓47.1046.04
(b)✓✓✓✓48.1946.96
(c)✓✓✓✓54.2450.94
(d)✓✓✓✓✓57.2352.32
(e)✓✓45.2544.46
(f)✓✓✓46.4745.47
(g)✓✓✓46.9645.23
(h)✓✓✓51.3448.52
(i)✓✓✓✓52.1248.75
(j)✓✓✓✓53.5750.06
(k)✓✓✓✓54.0950.84
(l)✓✓✓✓✓57.2352.32
", + "image_path": "d021248bb2f651fb0e0cc2ea69b8e928c47e14d3ce7d09d682c67235a13bc58b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 532 + ], + "type": "text", + "content": "fluences of the random boundary shifting and random resampling operations for data augmentation, we remove one or both of them from the training pipeline and the results are shown in Table 4 (a) " + }, + { + "bbox": [ + 304, + 447, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 304, + 447, + 545, + 532 + ], + "type": "text", + "content": " (d). The model performance clearly degrades after the removal, which demonstrates the necessity of increasing sample diversity and alleviating overfitting for weakly-supervised cross-modal regression learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Ablation on module designs. As shown in Table 4 (e) " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " (l), we conduct detailed ablation studies on module designs to validate the rationality of the proposed model. As observed, the designs of modulated positional encodings and dynamic anchor boxes in the encoder and decoder are consistently beneficial to improving the model capacity for better performance. The conceptual semantic connector that bridges two types of queries is also effective, and it further boosts the performance by " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.26\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " in mIoU (50.06% vs. 52.32%) even though the network has been equipped with strong encoders and decoders. We notice the dynamic anchors for decoding are the most crucial component given the sharpest performance drop of " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "3.57\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " with its removal, which indicates the importance to explicitly represent the intermediate location information for video grounding." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13575" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 83, + 265, + 140 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 282, + 81 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 282, + 81 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 282, + 81 + ], + "type": "text", + "content": "Table 5. Ablation studies on different weakly-supervised losses." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 83, + 265, + 140 + ], + "lines": [ + { + "bbox": [ + 64, + 83, + 265, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 83, + 265, + 140 + ], + "type": "table", + "html": "
ID\\( {\\mathcal{L}}_{\\text{screg }} \\)\\( {\\mathcal{L}}_{\\text{oga }} \\)R@0.3R@0.5R@0.7mIoU
(a)45.8529.8210.9330.97
(b)✓46.9029.3812.0631.82
(c)✓73.5854.5828.8250.62
(d)✓✓75.4357.2330.5652.32
", + "image_path": "80c4b9256a8f9828148f203399908a2d1dacd8fb3394acadc6c007b0baeaa670.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 74, + 156, + 256, + 216 + ], + "blocks": [ + { + "bbox": [ + 86, + 144, + 247, + 154 + ], + "lines": [ + { + "bbox": [ + 86, + 144, + 247, + 154 + ], + "spans": [ + { + "bbox": [ + 86, + 144, + 247, + 154 + ], + "type": "text", + "content": "Table 6. Impact of different auxiliary losses." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 156, + 256, + 216 + ], + "lines": [ + { + "bbox": [ + 74, + 156, + 256, + 216 + ], + "spans": [ + { + "bbox": [ + 74, + 156, + 256, + 216 + ], + "type": "table", + "html": "
MethodR@0.3R@0.5R@0.7mIoU
w/o Lcb71.4949.7824.6448.07
w/o Lar72.9650.7226.4249.14
w/o Lpa74.3456.0130.2951.76
Full Model75.4357.2330.5652.32
", + "image_path": "e0ce66939a2fb303545ac8d741df3ac043c1665050a8aedd330e0433ac7c597f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 70, + 231, + 260, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 219, + 286, + 230 + ], + "lines": [ + { + "bbox": [ + 47, + 219, + 286, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 286, + 230 + ], + "type": "text", + "content": "Table 7. Evaluation on different types of paragraph representation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 231, + 260, + 291 + ], + "lines": [ + { + "bbox": [ + 70, + 231, + 260, + 291 + ], + "spans": [ + { + "bbox": [ + 70, + 231, + 260, + 291 + ], + "type": "table", + "html": "
MethodR@0.3R@0.5R@0.7mIoU
Max-Pooling72.6152.6127.0349.55
Mean-Pooling74.3153.1328.6450.57
Word Concat73.7952.2727.3049.96
Learnable75.4357.2330.5652.32
", + "image_path": "787dbf9380367c350f50d6c4a07825ef543c490eb5aadb781ca49d6cb34fca18.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": "Analysis on weakly-supervised losses. Ablation studies on contributions of two weakly-supervised losses " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sreg}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{oga}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " are shown in Table 5. In experiment (a) and (c), we remove " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sreg}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " but use a plain regression loss " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " for ablation analysis. Firstly, simply employing regressive supervision attains inferior performance because the feature-level cross-modal correspondence can hardly be learned with coordinate-level supervision. Furthermore, we find " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{oga}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " is critical for learning precise temporal localization since it explicitly guides the model to align video features and language features that are highly likely to be correlated. Besides, we observe that using " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sreg}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " for selecting high-quality regression samples always brings gains to the performance. Impact of auxiliary losses. We investigate the influences of auxiliary losses on the model performance, which involve " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cb}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " for exploiting the cross-branch knowledge, " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ar}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " as guidance to learn a set of order-preserving anchor boxes and " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pa}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " to make use of the feature alignment supervision from the augmentation branch. As presented in Table 6, all three auxiliary losses bring positive impacts, with " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cb}}" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": " being the most effective to improve the mIoU metric by " + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "inline_equation", + "content": "4.25\\%" + }, + { + "bbox": [ + 47, + 293, + 287, + 568 + ], + "type": "text", + "content": ". The reason might lie in the rich complementary knowledge and consistency supervision across the siamese branches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Impact of the paragraph representation. The comparison of different types of paragraph representation is shown in Table 7. Four different schemes are included, i.e., mean-pooling or max-pooling the sentence features, embedding the sequence of all word tokens in the paragraph, and using a learnable query token to extract global information. It is clear that adaptively learning a paragraph representation with our method achieves the best performance and the max-pooling scheme performs the worst because it loses too much detailed information. The mean-pooling scheme performs slightly better than the word-concat scheme, which may attribute to the advantage of late fusion at feature level." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 98, + 543, + 191 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 544, + 97 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 544, + 97 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 544, + 97 + ], + "type": "text", + "content": "① A woman is sitting down on a floral glass table drilling a design into the pumpkin. ② As she is drilling, two boys are standing next to her watching her and then they suddenly leave. ③ The person behind the camera then picks up the top of the pumpkin to show its empty contents before the lady closes it and continues drilling." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 98, + 543, + 191 + ], + "lines": [ + { + "bbox": [ + 310, + 98, + 543, + 191 + ], + "spans": [ + { + "bbox": [ + 310, + 98, + 543, + 191 + ], + "type": "image", + "image_path": "d70722cea3e2ab9e991d32d0a59993de8774329bdbe587033a5dae94b7b101c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 206, + 543, + 299 + ], + "blocks": [ + { + "bbox": [ + 321, + 193, + 538, + 207 + ], + "lines": [ + { + "bbox": [ + 321, + 193, + 538, + 207 + ], + "spans": [ + { + "bbox": [ + 321, + 193, + 538, + 207 + ], + "type": "text", + "content": "A woman is seen standing in a circle and looking over her shoulder. She throws an object off into the distance and is shown again. 3 She throws her arms up and walks away afterwards." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 206, + 543, + 299 + ], + "lines": [ + { + "bbox": [ + 310, + 206, + 543, + 299 + ], + "spans": [ + { + "bbox": [ + 310, + 206, + 543, + 299 + ], + "type": "image", + "image_path": "c4d57aea25de7fce1ea6fb500717e1d4e5a760b6a6c79aa12cb69df43b98ca9e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 303, + 544, + 314 + ], + "lines": [ + { + "bbox": [ + 306, + 303, + 544, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 544, + 314 + ], + "type": "text", + "content": "Figure 4. Visualization of prediction results from different models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 320, + 389, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 389, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 389, + 331 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 339, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 459 + ], + "type": "text", + "content": "In Figure 4, we intuitively visualize the predicted timestamps from WSTPN and our proposed model trained under the weakly-supervised or semi-supervised setting. Overall, the predicted timestamps given by WSTPN coarsely capture the sentence order relations but with inaccurate boundaries. In contrast, our weakly-supervised model achieves much better results. It is notable that our semi-supervised model generates more fine-grained boundaries, which shows the advantage of our siamese framework in jointly leveraging fewer and weaker labels for efficient learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 466, + 378, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 378, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 378, + 478 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": "In this work, we explore the weakly-supervised setting in video paragraph grounding (i.e., WSVPG) to eliminate the dependence of the temporal annotations. To achieve this goal, we propose a novel siamese learning framework to jointly learn the cross-modal feature alignment and temporal coordinate regression without ground-truth supervision. Specifically, we design a novel Siamese Grounding Transformer (SiamGTR) consisting of an augmentation branch and an inference branch. The augmentation branch utilizes the boundary supervision provided by temporally regressing a complete paragraph in a pseudo video, and the inference branch learns the order-guided cross-modal correspondence of multiple sentences in a normal video. Extensive experiments verify the effectiveness of our framework." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009), and Guangzhou Science and Technology Plan Project (202201011134)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13576" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 137, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 137, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 287, + 179 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 182, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 182, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 287, + 213 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 216, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 216, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 58, + 216, + 287, + 237 + ], + "type": "text", + "content": "[4] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 287, + 271 + ], + "type": "text", + "content": "[5] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV Workshops, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 274, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 274, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 58, + 274, + 287, + 315 + ], + "type": "text", + "content": "[6] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. In NeurIPS, 1993. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 319, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 319, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 319, + 287, + 350 + ], + "type": "text", + "content": "[7] Meng Cao, Long Chen, Mike Zheng Shou, Can Zhang, and Yuexian Zou. On pursuit of designing multi-modal transformer for video grounding. In EMNLP, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 58, + 353, + 287, + 396 + ], + "type": "text", + "content": "[8] Meng Cao, Fangyun Wei, Can Xu, Xiubo Geng, Long Chen, Can Zhang, Yuexian Zou, Tao Shen, and Daxin Jiang. Iterative proposal refinement for weakly-supervised video grounding. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 398, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 398, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 58, + 398, + 287, + 439 + ], + "type": "text", + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 443, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 485 + ], + "type": "text", + "content": "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 488, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 287, + 530 + ], + "type": "text", + "content": "[11] Jiaming Chen, Weixin Luo, Wei Zhang, and Lin Ma. Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In AAAI, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 533, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 287, + 575 + ], + "type": "text", + "content": "[12] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 578, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 287, + 609 + ], + "type": "text", + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 287, + 644 + ], + "type": "text", + "content": "[14] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 647, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 667 + ], + "type": "text", + "content": "[15] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[16] Kyunghyun Cho, B van Merrienboer, Caglar Gulcehre, F Bougares, H Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP, 2014. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "text", + "content": "[17] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In ACMMM, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 150 + ], + "type": "text", + "content": "[18] Xuguang Duan, Wenbing Huang, Chuang Gan, Jingdong Wang, Wenwu Zhu, and Junzhou Huang. Weakly supervised dense event captioning in videos. In NeurIPS, 2018. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 153, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 153, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 312, + 153, + 545, + 185 + ], + "type": "text", + "content": "[19] Yazan Abu Farha and Jurgen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 188, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 188, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 312, + 188, + 545, + 209 + ], + "type": "text", + "content": "[20] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 211, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 211, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 312, + 211, + 545, + 242 + ], + "type": "text", + "content": "[21] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 246, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 246, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 246, + 545, + 277 + ], + "type": "text", + "content": "[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, 2017. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 312, + 280, + 545, + 323 + ], + "type": "text", + "content": "[23] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 326, + 545, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 326, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 312, + 326, + 545, + 356 + ], + "type": "text", + "content": "[24] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. arXiv preprint arXiv:2305.14344, 2023.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 360, + 545, + 392 + ], + "type": "text", + "content": "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 395, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 395, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 395, + 545, + 426 + ], + "type": "text", + "content": "[26] Jiabo Huang, Yang Liu, Shaogang Gong, and Hailin Jin. Cross-sentence temporal and semantic relations in video activity localisation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 429, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 429, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 312, + 429, + 545, + 460 + ], + "type": "text", + "content": "[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In CVPR, 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 463, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 463, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 312, + 463, + 545, + 495 + ], + "type": "text", + "content": "[28] Yifei Huang, Lijin Yang, and Yoichi Sato. Weakly supervised temporal sentence grounding with uncertainty-guided self-training. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 498, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 545, + 529 + ], + "type": "text", + "content": "[29] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 533, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 545, + 564 + ], + "type": "text", + "content": "[30] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, 2017. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 567, + 545, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 567, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 312, + 567, + 545, + 608 + ], + "type": "text", + "content": "[31] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, 2022. 3, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 612, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 612, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 312, + 612, + 545, + 643 + ], + "type": "text", + "content": "[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 545, + 678 + ], + "type": "text", + "content": "[33] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Workshops, 2015. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "text", + "content": "[34] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13577" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "text", + "content": "[35] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In CVPR, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 119, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 119, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 53, + 119, + 287, + 152 + ], + "type": "text", + "content": "[36] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 153, + 287, + 186 + ], + "type": "text", + "content": "[37] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 188, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 188, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 287, + 220 + ], + "type": "text", + "content": "[38] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In EMNLP, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 222, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 222, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 222, + 287, + 255 + ], + "type": "text", + "content": "[39] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 257, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 257, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 257, + 287, + 300 + ], + "type": "text", + "content": "[40] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In ICCV, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "text", + "content": "[41] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "type": "text", + "content": "[42] Zhijie Lin, Zhou Zhao, Zhu Zhang, Qi Wang, and Huasheng Liu. Weakly-supervised video moment retrieval via semantic completion network. In AAAI, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 287, + 403 + ], + "type": "text", + "content": "[43] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In ACMMM, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 406, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 406, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 287, + 438 + ], + "type": "text", + "content": "[44] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In AAAI, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 440, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 440, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 440, + 287, + 472 + ], + "type": "text", + "content": "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In ICCV, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 475, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 517 + ], + "type": "text", + "content": "[46] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In ICLR, 2021. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 520, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 520, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 520, + 287, + 553 + ], + "type": "text", + "content": "[47] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In EMNLP, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 287, + 587 + ], + "type": "text", + "content": "[48] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. TMM, 2022. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 53, + 589, + 287, + 632 + ], + "type": "text", + "content": "[49] Minuk Ma, Sunjae Yoon, Junyeong Kim, Youngjoon Lee, Sunghun Kang, and Chang D Yoo. Vlanet: Video-language alignment network for weakly-supervised video moment retrieval. In ECCV, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "type": "text", + "content": "[50] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In CVPR, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[51] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[52] Niluthpol Chowdhury Mithun, Sujoy Paul, and Amit K Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 139 + ], + "type": "text", + "content": "[53] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 142, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 142, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 312, + 142, + 545, + 186 + ], + "type": "text", + "content": "[54] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. Bassl: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 188, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 188, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 312, + 188, + 545, + 220 + ], + "type": "text", + "content": "[55] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 222, + 545, + 255 + ], + "type": "text", + "content": "[56] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 257, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 257, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 312, + 257, + 545, + 289 + ], + "type": "text", + "content": "[57] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 2013. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 291, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 291, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 312, + 291, + 545, + 334 + ], + "type": "text", + "content": "[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 337, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 337, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 337, + 545, + 358 + ], + "type": "text", + "content": "[59] Mrigank Rochan and Yang Wang. Video summarization by learning from unpaired data. In CVPR, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 360, + 545, + 392 + ], + "type": "text", + "content": "[60] Mrigank Rochan, Linwei Ye, and Yang Wang. Video summarization using fully convolutional sequence networks. In ECCV, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 395, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 395, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 395, + 545, + 437 + ], + "type": "text", + "content": "[61] Marcus Rohrbach, Michaela Regneri, Mykhaylo Andriluka, Sikandar Amin, Manfred Pinkal, and Bernt Schiele. Script data for attribute-based recognition of composite activities. In ECCV, 2012. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 440, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 440, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 312, + 440, + 545, + 473 + ], + "type": "text", + "content": "[62] Fengyuan Shi, Limin Wang, and Weilin Huang. End-to-end dense video grounding via parallel regression. arXiv preprint arXiv:2109.11265, 2021. 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 312, + 475, + 545, + 518 + ], + "type": "text", + "content": "[63] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 520, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 520, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 520, + 545, + 553 + ], + "type": "text", + "content": "[64] Mattia Soldan, Mengmeng Xu, Sisi Qu, Jesper Tegner, and Bernard Ghanem. Vlg-net: Video-language graph matching network for video grounding. In CVPR, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 545, + 598 + ], + "type": "text", + "content": "[65] Yijun Song, Jingwen Wang, Lin Ma, Zhou Yu, and Jun Yu. Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 601, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 601, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 312, + 601, + 545, + 632 + ], + "type": "text", + "content": "[66] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In CVPR, 2014. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 677 + ], + "type": "text", + "content": "[67] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Wei-Shi Zheng, and Jianhuang Lai. Hierarchical semantic correspondence networks for video paragraph grounding. In CVPR, 2023. 3, 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 680, + 545, + 713 + ], + "type": "text", + "content": "[68] Reuben Tan, Huijuan Xu, Kate Saenko, and Bryan A Plummer. Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In WACV, 2021. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13578" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[69] Chenxin Tao, Xizhou Zhu, Weijie Su, Gao Huang, Bin Li, Jie Zhou, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. In CVPR, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 149 + ], + "type": "text", + "content": "[70] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 287, + 194 + ], + "type": "text", + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 227 + ], + "type": "text", + "content": "[72] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. In AAAI, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 260 + ], + "type": "text", + "content": "[73] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "text", + "content": "[74] Jingwen Wang, Lin Ma, and Wenhao Jiang. Temporally grounding language queries in videos by contextual boundary-aware prediction. In AAAI, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 294, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 287, + 327 + ], + "type": "text", + "content": "[75] Yuechen Wang, Jiajun Deng, Wengang Zhou, and Houqiang Li. Weakly supervised temporal adjacent network for language grounding. TMM, 2022. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 287, + 369 + ], + "type": "text", + "content": "[76] Yunxiao Wang, Meng Liu, Yinwei Wei, Zhiyong Cheng, Yinglong Wang, and Liqiang Nie. Siamese alignment network for weakly supervised video moment retrieval. TMM, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 287, + 403 + ], + "type": "text", + "content": "[77] Zheng Wang, Jingjing Chen, and Yu-Gang Jiang. Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In ACMMM, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "text", + "content": "[78] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. In AAAI, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 287, + 470 + ], + "type": "text", + "content": "[79] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In AAAI, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 287, + 503 + ], + "type": "text", + "content": "[80] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In CVPR, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 504, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 287, + 536 + ], + "type": "text", + "content": "[81] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 537, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 287, + 568 + ], + "type": "text", + "content": "[82] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 570, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 287, + 603 + ], + "type": "text", + "content": "[83] Shuwen Xiao, Zhou Zhao, Zijian Zhang, Xiaohui Yan, and Min Yang. Convolutional hierarchical attention network for query-focused video summarization. In AAAI, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 604, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 604, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 604, + 287, + 635 + ], + "type": "text", + "content": "[84] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. In EMNLP, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 637, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 637, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 637, + 287, + 678 + ], + "type": "text", + "content": "[85] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In AAAI, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "type": "text", + "content": "[86] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In AAAI, 2019. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[87] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 150 + ], + "type": "text", + "content": "[88] Wenfei Yang, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Local correspondence network for weakly supervised temporal sentence grounding. TIP, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 194 + ], + "type": "text", + "content": "[89] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 197, + 545, + 229 + ], + "type": "text", + "content": "[90] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 545, + 262 + ], + "type": "text", + "content": "[91] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In AAAI, 2019. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "text", + "content": "[92] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In ACMMM Workshops, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 312, + 309, + 545, + 342 + ], + "type": "text", + "content": "[93] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 545, + 387 + ], + "type": "text", + "content": "[94] Chenchi Zhang, Wenbo Ma, Jun Xiao, Hanwang Zhang, Jian Shao, Yueting Zhuang, and Long Chen. Vl-nms: Breaking proposal bottlenecks in two-stage visual-language matching. TOMM, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 545, + 430 + ], + "type": "text", + "content": "[95] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 312, + 433, + 545, + 464 + ], + "type": "text", + "content": "[96] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 545, + 510 + ], + "type": "text", + "content": "[97] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. TPAMI, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 512, + 545, + 555 + ], + "type": "text", + "content": "[98] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 557, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 545, + 589 + ], + "type": "text", + "content": "[99] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In ECCV, 2016. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "type": "text", + "content": "[100] Mingxing Zhang, Yang Yang, Xinghan Chen, Yanli Ji, Xing Xu, Jingjing Li, and Heng Tao Shen. Multi-stage aggregated transformer network for temporal language localization in videos. In CVPR, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "text", + "content": "[101] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 2, 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[102] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. TPAMI, 2021. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13579" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 297 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[103] Yimeng Zhang, Xin Chen, Jinghan Jia, Sijia Liu, and Ke Ding. Text-visual prompting for efficient 2d temporal video grounding. In CVPR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[104] Zhu Zhang, Zhijie Lin, Zhou Zhao, Jieming Zhu, and Xiuqiang He. Regularized two-branch proposal networks for weakly-supervised moment retrieval in videos. In ACMMM, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "text", + "content": "[105] Zhu Zhang, Zhou Zhao, Zhijie Lin, Xiuqiang He, et al. Counterfactual contrastive learning for weakly-supervised vision-language grounding. In NeurIPS, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 218 + ], + "type": "text", + "content": "[106] Minghang Zheng, Yanjie Huang, Qingchao Chen, and Yang Liu. Weakly supervised video moment localization with contrastive negative sample mining. In AAAI, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 262 + ], + "type": "text", + "content": "[107] Minghang Zheng, Yanjie Huang, Qingchao Chen, Yuxin Peng, and Yang Liu. Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 286, + 297 + ], + "type": "text", + "content": "[108] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2021. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13580" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_content_list.json b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9a8d0322f3faee953f28120ce0f9263359de17b0 --- /dev/null +++ b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_content_list.json @@ -0,0 +1,1463 @@ +[ + { + "type": "text", + "text": "Sieve: Multimodal Dataset Pruning Using Image Captioning Models", + "text_level": 1, + "bbox": [ + 140, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anas Mahmoud1,2* Mostafa Elhoushi1 Amro Abbas† Yu Yang3", + "bbox": [ + 215, + 179, + 753, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Newsha Ardalani1 Hugh Leather1 Ari S. Morcos4", + "bbox": [ + 272, + 199, + 694, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FAIR at Meta $^{2}$ University of Toronto $^{3}$ UC Los Angeles $^{4}$ DatologyAI", + "bbox": [ + 181, + 215, + 789, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning. We argue that this approach suffers from multiple limitations including: false positives and negatives due to CLIP's pretraining on noisy labels. We propose a pruning signal, Sieve, that employs synthetic captions generated by image captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on unlabeled text corpus. Using DataComp, a multimodal dataset filtering benchmark, when evaluating on 38 downstream tasks, our pruning approach, surpasses CLIPScore by $2.6\\%$ and $1.7\\%$ on medium and large scale respectively. In addition, on retrieval tasks, Sieve leads to a significant improvement of $2.7\\%$ and $4.5\\%$ on medium and large scale respectively.", + "bbox": [ + 75, + 300, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 690, + 209, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) [32] models have shown great success in solving zero-shot image classification and multimodal retrieval tasks. In addition, many foundational Vision-Language Models (VLMs) use pretrained CLIP encoders to condition image generation on CLIP text embeddings [33] in retrieval augmented vision-language models [13, 42], and to align modalities including audio, depth, and thermal with language through CLIP image embeddings [10]. Therefore, the quality of CLIP representations can influence the performance of many VLMs.", + "bbox": [ + 75, + 715, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To pretrain CLIP, billions of image-text pairs are collected using common crawl. The raw data is highly diverse but contains many noisy image-text pairs, including low quality images, low quality alternative text (alt-text), and misaligned image-text pairs. Pretraining CLIP models on noisy data can have adverse effects on the learned representations, thus leading to poor performance on downstream tasks [1].", + "bbox": [ + 500, + 268, + 890, + 391 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge, researchers have developed data pruning methods to remove low quality image-text pairs. Heuristics that filter out image-text pairs based on image dimensions, aspect ratio, alt-text length, and complexity are commonly used [9, 36] to reduce noise, but can also limit the diversity of the dataset [26]. Methods that use images or class names from datasets, like ImageNet, to sample semantically similar image-text pairs can lead to higher accuracy on downstream tasks [40], but limit the diversity of the selected samples as they sample image-text pairs close to a specific dataset.", + "bbox": [ + 496, + 393, + 892, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the most effective pruning methods, CLIP-Score [11, 36], computes the cosine similarity between image and text embeddings using a pretrained CLIP model. This score is then used to rank the alignment of image-text pairs. However, as shown in Figure 1, using CLIPScore can lead to false positives – samples that are poorly aligned but have high CLIPScore (i.e., bad samples) due to spurious correlations learned by the pretrained CLIP model [41]. In addition, using CLIPScore can lead to false negatives – samples that are aligned but have low CLIPScore (i.e., hard samples) due to the poor discrimination between hard and bad samples. Excluding hard samples and including bad samples can negatively affect the generalization of CLIP image and text encoders.", + "bbox": [ + 496, + 564, + 893, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of this work is to reduce both false negatives and positives induced by CLIPScore ranking by relying on an image-captioning model pretrained on small, diverse, and well-aligned image-text pairs. As shown in Figure 1, Sieve can reduce false positives or false negatives in different scenarios, such as samples where CLIPScore focuses on geographic or chronological context of an image rather than the content of an image. As depicted in Figure 2, we evalu", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at Meta", + "bbox": [ + 94, + 875, + 300, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Work done during an AI residency at Meta", + "bbox": [ + 96, + 887, + 328, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22423", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a48f7a47ad075a945510158eab7694247ca5a2e6914b0d4a11dc4c331e56fef3.jpg", + "image_caption": [], + "image_footnote": [ + "ALT-Text:", + "The Best Places to Stay in Kaikoura, New", + "Zealand", + "CLIPScore:0.30", + "False Positive", + "Generated Caption:", + "a living room with a large window and a couch", + "Sieve: 0.12", + "True Negative" + ], + "bbox": [ + 84, + 95, + 302, + 193 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8a7d643f141acfb48b68768d4a7d88168c82c715f12adccc613339d9c86140a2.jpg", + "image_caption": [ + "Figure 1. Examples of image-text pairs in which the scores of pretrained CLIP models, a commonly used image-text data filtering approach, fail to measure their alignment. Our proposed approach, Sieve, provides an accurate alignment score using a caption generator and sentence transformer. Top left and bottom left: Examples of false positives where alt-text describes concepts that are not found or unrelated to the image. CLIP is trained on similar noisy image-text pairs, thus, it assigns a relatively high score. Sieve can detect that such image-text pairs are misaligned. Top right and bottom right: Examples of false negatives where images are aligned with the alt-text but are assigned low CLIP scores, either due to the low likelihood of these pairs in the pretraining data, or because CLIP may have seen similar images aligned with other noisy labels. Sieve can detect that such image-text pairs are well-aligned and selects them for pretraining." + ], + "image_footnote": [ + "ALT-Text:", + "The area were evacuated and Keiserville Rd.", + "in Washington Twp. Pa., was closed as", + "crews worked on the Marcellus street", + "CLIPScore: 0.31", + "False Positive", + "Generated Caption:", + "a view of a construction site with water", + "coming out of it", + "Sieve: 0.21", + "True Negative" + ], + "bbox": [ + 88, + 198, + 297, + 306 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/712cfe9d6b8953759f4d7f9329818b324e3fda5c89e6c4445affd74b42657645.jpg", + "image_caption": [], + "image_footnote": [ + "ALT-Text:", + "The Great Sphinx of Giza", + "CLIPScore:0.25", + "False Negative", + "Generated Caption:", + "the great sphinx and the pyramids of giza", + "Sieve: 0.91", + "True Positive" + ], + "bbox": [ + 669, + 94, + 883, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0e55db0e340610e54f5dfcd653d294a5079a7e17f780f6bdf510d955a688a2e0.jpg", + "image_caption": [], + "image_footnote": [ + "ALT-Text:", + "A view of the interior pool", + "CLIPScore: 0.23", + "False Negative", + "Generated Caption:", + "a large indoor swimming pool with a skylight", + "Sieve: 0.65", + "True Positive" + ], + "bbox": [ + 666, + 210, + 883, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ate the alignment of web-crawled image-text pairs by first, generating multiple captions for each image using nucleus sampling [12], followed by removing phrases that describe the medium (e.g., \"an image of\", \"a photo of\") rather than visual concepts. Finally, to evaluate semantic similarity between the limited diversity of generated captions and the high diversity of alt-text, we utilize the embedding space of a lightweight sentence transformer pretrained on unlabeled text corpus. The alignment between the generated captions and the alt-text is then used as a proxy for image-text alignment. Our goal is to fuse this alignment signal with CLIP-Score to minimize false positives and negatives, leading to a more aligned pretraining dataset. To evaluate the effectiveness of our proposed pruning method, we utilize the DataComp [9] benchmark, which fixes the pretraining hyperparameters of CLIP and provides multiple candidate pools of noisy image-text data for pretraining CLIP models. The goal is to select a subset of noisy image-text data that leads to the best performance on 38 downstream tasks. Using image-captioning model alignment scores fused with CLIP-Score, we surpass CLIPScore filtering by $2.6\\%$ and $1.7\\%$ on the average of 38 downstream tasks on medium and large scale pool respectively. In addition, on multimodal retrieval tasks, our approach leads to an improvement of $2.7\\%$ and $4.5\\%$ on medium and large scale respectively.", + "bbox": [ + 75, + 448, + 472, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 843, + 218, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Heuristics are basic filtering methods including: filtering non-English alt-text using fastText [15], filtering alt-text", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with a few words [9, 36], filtering alt-text with low text complexity [30], and filtering images by size or aspect ratio [9]. A combination of these unimodal filtering approaches has been explored by DataComp [9]. An example of a multimodal filtering approach is text spotting: detecting and recognizing text in images and filtering image-text pairs with high overlap between spotted text (text detected in image) and alt-text (associated label of image) [30].", + "bbox": [ + 496, + 446, + 890, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets as Priors was proposed in DataComp [9], relying on sampling image-text pairs that are semantically similar to diverse and curated datasets like ImageNet [5]. Text-based sampling selects image-text pairs with alt-text overlapping one of the ImageNet classes. CiT [40] uses cosine similarity to filter alt-text that are similar to ImageNet classes. Image-based sampling approaches encode images from the unfiltered candidate pool using the OpenAI CLIP's ViT-L/14 vision encoder, and clusters the images into 100,000 groups using FAISS [14]. Then, embeddings of ImageNet training samples are used to keep the closest cluster to each sample. The main limitation of such approaches is that they bias the CLIP model and may not generalize well to new downstream tasks. We argue that selecting samples that match the distribution of downstream tasks encourages overfitting to the evaluation set and, thus limits generalization to other downstream tasks. Our approach, Sieve, does not use any dataset as a prior.", + "bbox": [ + 496, + 575, + 892, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pretrained VLMs One of the most successful methods for evaluating image-text alignment is CLIPScore [11]. LAION filtering [36] uses an OpenAI CLIP model [32] pre", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22424", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained on 400 million image-text pairs to evaluate image-text alignment of large webscale datasets, and filter out samples with the lowest CLIPScore. Filtering using CLIP-Score can suffer from false negatives, which leads to filtering out hard informative samples, and false positives, which leads to including misaligned samples. Another approach proposes a non-filtering approach that utilizes pretrained VLMs [26], using large image-captioning models like BLIP2 [21] to replace alt-text labels with descriptive synthetic captions. The synthetic captions are then used to train CLIP models. The authors [26] demonstrate that at scale, the improvement of synthetic captions is capped by the limited diversity of generated captions compared to the high diversity of noisy text labels. Compared to [26], we do not alter the original alt-text and thus our focus is on the dataset pruning challenge.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent Work Inspired by text spotting [30], T-MARS [24] is concurrent work that detects and masks text regions in images before computing CLIPScore, resulting in improved visual representations. We reason that T-MARS is orthogonal to our approach as we can apply the same masking before calculating Sieve. In addition, Devil is In the Details (DID) [44], utilizes a combination of multiple filters and sampling approaches, which is orthogonal to our approach as Sieve can be combined to other filters. Moreover, most of DID's accuracy improvements is from aligning the selected data distribution with the DataComp [9] evaluation set (downstream tasks), which overfits downstream tasks as discussed in previous sections. More importantly, T-MARS [24] and DID [9] only perform pruning experiments on the medium scale of DataComp [9], while Sieve shows results on both medium and large scales. Finally, DFN [8], uses CLIPScore from a CLIP model pretrained on a private 357 million human verified image-text labels, which is $25 \\times$ bigger than the dataset our selected captioning models were pretrained on, and is bigger than DataComp's medium scale pool.", + "bbox": [ + 75, + 333, + 472, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 666, + 212, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{D} = \\{(I_i,T_i)\\}_{i = 1}^N$ denote an uncurated dataset consisting of $N$ image-text pairs crawled from the web. Our goal is to curate a dataset, $\\mathcal{D}' = \\{(I_{i'},T_{i'})\\}_{i' = 1}^{N'}$ , that is a subset of the pool, $\\mathcal{D}'\\subseteq \\mathcal{D}$ $N^{\\prime}\\leq N$ , to train a new CLIP model from uninitialized weights, $\\Theta_0$ , to new weights, $\\Theta^{\\prime}$ :", + "bbox": [ + 75, + 691, + 468, + 767 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {\\prime} = \\operatorname {t r a i n} \\left(\\Theta_ {0}, \\mathcal {D} ^ {\\prime}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 780, + 468, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a given scoring function, $f$ , that maps an image-text pair to a scalar value, $s = f(I_i, T_i)$ , we express a pruning function that selects a fraction, $k$ , of dataset, $\\mathcal{D}$ , using function, $f$ :", + "bbox": [ + 75, + 811, + 468, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Pr n e _ {f} (\\mathcal {D}, k) \\quad \\text {s . t .} 0 \\leq k \\leq 1, f: (I _ {i}, T _ {i}) \\rightarrow \\mathbb {R} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 883, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{prune}_f(\\mathcal{D},k)$ applies function, $f$ , on each image-text sample in $\\mathcal{D}$ to obtain a score for each sample, ranks the scores in descending order, and returns a set of the top $k$ portion of the samples.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One common approach for pruning is CLIPScore [9, 36]. Let $E$ be a CLIP model consisting of an image encoder, $E_{\\mathrm{image}}$ , that maps an image, $I$ , to an embedding vector, $E_{\\mathrm{image}}(I) \\in \\mathbb{R}^d$ , and a text encoder, $E_{\\mathrm{text}}$ , that maps a text sample, $T$ , to an embedding vector, $E_{\\mathrm{text}}(T) \\in \\mathbb{R}^d$ . CLIP-Score is a measure of alignment between $I$ and $T$ , and is defined as:", + "bbox": [ + 498, + 152, + 893, + 256 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {C L I P}} (I, T) = \\left\\langle E _ {\\text {i m a g e}} (I), E _ {\\text {t e x t}} (T) \\right\\rangle \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 268, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle \\mathbf{x},\\mathbf{y}\\rangle$ is the cosine similarity between two vectors, $\\mathbf{x}$ and $\\mathbf{y}$ , which is defined as the dot product of the $l_{2}$ normalized vectors. The most common CLIP model used for pruning is pretrained on 400 million noisy image-text pairs [36]. Our proposed pruning method, Sieve attempts to minimize the false positives and negatives induced by CLIPScore filtering. Sieve consists of two main components: ImageCaptioning and Sentence Transformer.", + "bbox": [ + 496, + 295, + 890, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-Captioning Let $G$ be a captioning model that generates text, $T_{i}^{G}$ , describing the content of image, $I_{i}$ :", + "bbox": [ + 498, + 416, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} ^ {G} = G \\left(I _ {i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 455, + 890, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a captioning model pre-trained on a small, representative and well-aligned dataset of image-text pairs, we are interested in estimating the alignment between image-text pairs sampled from a very large, diverse but noisy dataset. The alignment score can then be used as a ranking metric for dataset filtering. We hypothesize that:", + "bbox": [ + 496, + 484, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The probability of generating a caption that is semantically similar to the alt-text from an aligned pair is much higher than that from a misaligned pair.", + "- The probability of generating a caption that is semantically similar to a hard alt-text is higher than generating a caption that is semantically similar to a misaligned alt-text. Here, a hard alt-text is a text label with low likelihood with respect to the captioning model, but is aligned with the image content." + ], + "bbox": [ + 500, + 575, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As images can contain multiple objects with complex attributes and relationships, there exist multiple ways to describe their content. Given the inherent many-to-many relationship between images and text labels, our goal is to increase the probability of generating a caption that matches an aligned alt-text. To achieve this, we utilize nucleus sampling [12], a decoding strategy used to sample multiple captions, $r$ , per-image:", + "bbox": [ + 496, + 710, + 890, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (I, r) = \\left\\{T _ {0} ^ {G}, T _ {1} ^ {G}, \\dots , T _ {r - 1} ^ {G} \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 840, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sentence Transformer Given an image, its alt-text, and a set of generated captions, our goal is to estimate the align-", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22425", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bcb5e268942e6704ce3b5bd2b457eb99bd8b594b3c6be2cbcb09b9befc031b68.jpg", + "image_caption": [ + "Figure 2. Our proposed framework enables dataset pruning using image captioning models. To evaluate the alignment of a noisy image-text pair, we generate multiple captions per image using nucleus sampling. Then medium phrases, like \"an image of\" or \"a photo of\", are masked from alt-text and generated captions. Finally, a lightweight sentence encoder is used to semantically compare the generated captions with alt-text." + ], + "image_footnote": [], + "bbox": [ + 127, + 97, + 844, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ment between the generated captions and the alt-text. However, there is a very large diversity gap between the generated captions and the highly diverse alt-text as measured by the number of unique nouns and trigrams [26]. On the other hand, constructing a large, diverse and curated image-text dataset is expensive, which limits the diversity of the generated captions. We propose to bridge this gap by utilizing a light-weight sentence similarity model to encode the alt-text and the generated captions. We expect the semantically similar alt-text and generated caption pairs to be closely clustered in the embedding space compared to semantically distinct pairs. We reason that the rich semantic textual embedding space of the sentence similarity model enables pretraining the captioning model only on a small but curated image-text dataset. Thus, we rely on the semantic understanding of the sentence similarity model to bridge the gap between the limited diversity in the generated captions and the highly diverse alt-text labels.", + "bbox": [ + 75, + 388, + 472, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate the alignment score, we compute the cosine similarity between embeddings of each generated caption and text label. Let $S$ be a language model that encodes a text sample, $T$ , to a vector, $S(T) \\in \\mathbb{R}^d$ . We define the alignment between two text samples, $T_a$ and $T_b$ , as the cosine similarity between their language model encodings:", + "bbox": [ + 75, + 661, + 470, + 752 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\langle S \\left(T _ {a}\\right), S \\left(T _ {b}\\right) \\rangle \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 762, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This estimate can then be used as a proxy for the image-text alignment of an image, $I$ , and text, $T$ :", + "bbox": [ + 76, + 787, + 470, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\langle S (G (I)), S (T) \\rangle\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 828, + 331, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If the image captioning model generates $r$ different caption candidates for an image, $I$ , we can use the maximum alignment between each of the generated captions,", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$G(I,r) = \\{T_0^G,T_1^G,\\ldots ,T_{r - 1}^G\\}$ , and a text sample, $T$ :", + "bbox": [ + 500, + 387, + 861, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {T _ {j} ^ {G} \\in G (I, r)} \\langle S \\left(T _ {j} ^ {G}\\right), S (T) \\rangle \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 412, + 890, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In literature, there are different models and approaches to obtain text embeddings. [16] uses the average of N-gram features of each word in a text sample to obtain an embedding. A more common option is to use the logits of the last token generated by a decoder-only language model, which is the approach taken with CLIP's text encoder [32], in GPT-1 [31], as well as in [1]. Encoder-only models, such as BERT [6] or RoBERTa [23], can also be used, where the embedding vector may be either the logits of the classification token, or the average pool of the logits of all tokens. Although such language models may have strong generation or classification capabilities, they were not optimized for sentence similarity tasks, but either for next word prediction (i.e., causal language modeling) or masked word prediction (i.e., masked language modeling) tasks. Therefore, their embeddings may not be ideal to measure alignment between sentences. More importantly, such models are large in size and hence slow to infer on large datasets. A language model finetuned on a sentence similarity task, such as SNLI [3], aligns with the goal of estimating semantic textual similarity between alt-text and captions. We find that sentence similarity models [34] pretrained using a self-supervised instance discrimination task on unlabeled corpus of sentences perform well in estimating the alignment between text pairs, and are lightweight in size and latency (e.g., tens of millions of parameters in contrast to billions of parameters of performant decoder-only large language models).", + "bbox": [ + 496, + 446, + 892, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking Medium Words Phrases such as \"image of\", \"picture of\", or \"photo of\" can appear in either alt-text or", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22426", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generated captions. We refer to such phrases as \"medium phrases\", as they describe the medium rather than the contents of an image. We notice that the existence of such medium phrases adds noise to the sentence similarity score, as shown in Figure 3. A pair of sentences that each have a medium phrase are assigned a misleadingly high sentence similarity score by a sentence transformer, as they have been trained on a wide and diverse corpus of text, rather than on image captions. Hence, the existence of medium phrases may increase their attention to the topic of images or media, rather than the topic of the content of such images. Therefore, we neutralize the effect of medium phrases by removing them from both alt-text and generated captions. We express the operation of masking medium words in a text sample, $M(T)$ , on text, $T$ , as masking all possible contiguous subsequences of the text, where masking on a phrase, $t$ , removes it if it is in the pre-determined list of medium phrases, $\\mathcal{M} = \\{\\text{\"image of\" }, \\text{\"picture of\" }, \\text{\"photo of\" }, \\ldots \\}$ , or keeps it otherwise.", + "bbox": [ + 75, + 90, + 472, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Putting it all together, we define the Sieve score function as:", + "bbox": [ + 76, + 378, + 470, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {S i e v e}} (I, T) = \\max _ {T _ {j} ^ {G} \\in G (I, r)} \\left\\langle S \\left(M \\left(T _ {j} ^ {G}\\right)\\right), S (M (T)) \\right\\rangle \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 417, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dataset pruned using Sieve with the top $k$ portion of its samples can be expressed as:", + "bbox": [ + 76, + 454, + 468, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\text {S i e v e}, k} = \\operatorname {p r u n e} _ {f _ {\\text {S i e v e}}} (\\mathcal {D}, k) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 496, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarize our approach in Figure 2 and as psuedocode in Algorithm 1 in the Appendix.", + "bbox": [ + 76, + 522, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 565, + 209, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training and Evaluation", + "text_level": 1, + "bbox": [ + 76, + 590, + 302, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the DataComp benchmark to evaluate the utility of image-captioning models for multimodal dataset pruning. Two candidate pools are considered, the medium and the large scale, consisting of 128 million and 1.28 billion image-text pairs, respectively. To train CLIP models, we use DataComp's hyperparameters and architectures to standardize training [9]: $5 \\times 10^{-4}$ learning rate, 500 iterations warmup, AdamW optimizer, for medium scale: ViT-B/32 image encoder [7], batch size 4096, 128M training samples as a compute budget, and for large scale: ViT-B/16 image encoder, batch size 8192, 1.28B training samples as a compute budget. We evaluate the zero-shot performance on 38 downstream tasks, including classification and retrieval tasks [18, 32, 45].", + "bbox": [ + 75, + 613, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For our captioning model, we utilize BLIP with ViT-B/16 image encoder pretrained on 14 million image-text pairs [20], including highly curated web datasets: Conceptual Captions [37], Conceptual 12M [4] and SBU captions [29]; as well as small human-annotated datasets:", + "bbox": [ + 75, + 825, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "COCO, and Visual Genome [17]. To compute the alignment between generated captions and alt-text, we use a lightweight distilled sentence transformer, all-MiniLM-L6-v2 [39], further finetuned using self-supervised contrastive learning on unlabeled text corpus.", + "bbox": [ + 496, + 90, + 892, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 498, + 175, + 638, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 reports multiple baselines from DataComp [9], including applying no filtering, basic filtering, and CLIP-Score filtering. On the medium scale, Sieve with an image-captioning model pretrained on 30 times less but curated data surpasses CLIPScore by $1.8\\%$ on average. In addition, we demonstrate that Sieve's alignment score provides a complimentary signal to CLIPScore by fusing Sieve with CLIPScore. We apply min-max normalization to Sieve alignment scores and CLIPScore independently, then take the per-sample weighted average of both scores:", + "bbox": [ + 496, + 198, + 892, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {\\text {S i e v e} + \\text {C L I P}} (I, T) = \\\\ (1 - \\alpha) \\times \\bar {f} _ {\\text {S i e v e}} (I, T) + \\alpha \\times \\bar {f} _ {\\text {C L I P}} (I, T) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 378, + 874, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} \\text {s . t .} \\quad \\overline {{f}} (I, T) = \\end{array} \\\\ \\frac {f (I , T) - \\min _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i})}{\\max _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i}) - \\min _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i})} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 441, + 874, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the weight $\\alpha$ used in the reported results is 0.5. Finally, we select the top $20\\%$ of samples. We observe that the fused approach improves CLIPScore average performance by $2.6\\%$ and $1.7\\%$ on medium and large scale respectively, demonstrating the complementarity of Sieve's alignment score. Moreover, Sieve without fusion achieves the best performance on retrieval tasks on both medium and large scale experiments. Relative to CLIPScore, a significant portion of Sieve's retrieval gain is effectively transferred when fusing Sieve with CLIPScore (i.e., $+2.7\\%$ and $+4.5\\%$ on medium and large scale). This is also promising as retrieval performance is critical for retrieval augmented VLMs [13, 42].", + "bbox": [ + 496, + 506, + 892, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Per-Task Performance", + "text_level": 1, + "bbox": [ + 498, + 710, + 709, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 shows the change in accuracy introduced by Sieve as well as Sieve+CLIPScore on each task compared to CLIPScore on medium scale, and Figure 6 of the Appendix shows for the large scale. We observe that in addition to outperforming on image retrieval tasks, Flickr [43], and MS COCO [22], Sieve's greatest performance boost comes from WingoGAViL [2], a retrieval task which requires diverse reasoning skills, including general knowledge, common sense, and abstraction. This high performance can be attributed to Sieve's preference towards keeping samples where the alt-text correctly describes visual concepts and", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22427", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eaf9ab012b2e130c30e3b372567c1addb373a1d1e6e5326d2612736c654e1174.jpg", + "image_caption": [ + "Figure 3. Masking medium phrases improves the ranking of sentence similarity scores. On the left, sentence pairs with misleadingly high (or low) sentence similarity due to the existence (or absence) of medium phrases are highlighted in dark red (or light red). On the right, similarity scores that are more aligned with semantics are highlighted in dark green. The sentence similarity scores are computed using the all-MiniLM-L6-v2 sentence transformer [39]." + ], + "image_footnote": [], + "bbox": [ + 217, + 101, + 754, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3e0dae1e0f73786287d9f6e3a13a2ff694df1091dfdd4c5487accdfb3c320178.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScaleFilteringDataset SizeImageNetImageNet dist. shiftsVTABRetrievalAverage over 38 datasets
Medium (128 Million)No Filtering128M17.615.225.921.925.8
Basic Filtering30M22.619.328.425.128.5
LAION Filtering13M23.019.830.723.329.2
CLIPScore38M27.323.033.825.132.8
Sieve24M29.425.035.228.934.6
Sieve+CLIPScore24M30.325.436.227.835.4
Large (1.28 Billion)No Filtering1.28B45.937.842.641.943.7
Basic Filtering298M51.642.344.648.045.8
LAION Filtering130M55.345.351.049.550.1
CLIPScore384M57.847.453.846.652.9
Sieve235M57.347.852.052.052.3
Sieve+CLIPScore235M59.749.154.851.154.6
", + "bbox": [ + 78, + 338, + 890, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Zero-shot performance of CLIP models pretrained using various filtering strategies on medium and large scale pools of the DataComp benchmark. Sieve fused with CLIPScore beats CLIPScore by $2.6\\%$ and $1.7\\%$ on medium and large scale respectively. In addition, on retrieval tasks, Sieve achieves best performance on both scales and after fusing with CLIPScore, leads an improvement on retrieval over CLIPScore of $2.7\\%$ and $4.5\\%$ on medium and large scale respectively.", + "bbox": [ + 75, + 595, + 893, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "their attributes and relations. Sieve, especially when combined with CLIPScore, significantly outperforms on medical diagnosis tasks, Camelyon17 and PatchCamelyon.", + "bbox": [ + 75, + 678, + 470, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sieve mainly underperforms in tasks requiring parsing text from images, such as MNIST [19], SVHN [25], and Rendered SST-2 [28], concluding that Sieve is less likely to select image-text pairs that are useful for OCR tasks. In addition, Sieve underperforms CLIPScore on context-based tasks like Country211 [27], a task assessing the geolocation capability of visual representations, demonstrating Sieve's preference towards selecting samples based on the alignment of alt-text with visual concepts rather than context. Interestingly, when fusing with CLIPScore, we improve the performance of all these tasks while retaining the advantage", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of Sieve especially at large scale (see Figure 6).", + "bbox": [ + 500, + 678, + 816, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 707, + 665, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct studies on the medium scale pool and report the average of three runs per experiment.", + "bbox": [ + 498, + 732, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining data-distribution We study the effect of the pretraining data distribution used to train the captioning model on the quality of the alignment score. This is measured based on the downstream performance of the CLIP model trained using the selected image-text pairs. Two pretraining data distributions proposed by BLIP [20] are investigated. The first uses 14 million curated image-text pairs, while the second uses an additional 115 million web images with noisy alt-text [36]. Although the original BLIP", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22428", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e807ca87d0e3eb1912fd0a8bb0a0737556557f42142380d5cf415a637b80b51.jpg", + "image_caption": [ + "Figure 4. Evaluating CLIP models pretrained on different fractions of the top ranked samples based on our proposed approach (Sieve), CLIPScore, and fusing Sieve with CLIPscore (Sieve+CLIPSCore), on medium scale." + ], + "image_footnote": [], + "bbox": [ + 217, + 107, + 470, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50f0e847567f9ef0efad18ef1657a8f14727f96870d55466c32991801dd49059.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 108, + 738, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b6dfc36c7f8768e6032601f46909d1e6ed328909a3a97093248e1d2d5d8b6ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Percentage KeptCaption Generator Pretraining DataImageNetImageNet dist. shiftVTABRetrievalAverage over 38 datasets
10BLIP-129M23.0020.6030.2021.4030.00
BLIP-14M26.5022.5032.1023.7530.90
15BLIP-129M25.9522.8032.8024.4032.40
BLIP-14M29.0524.6033.3526.9532.70
20BLIP-129M27.8523.6533.4526.3533.05
BLIP-14M29.6024.9335.0728.5734.03
", + "bbox": [ + 160, + 313, + 810, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "work reports higher captioning performance when pretraining on 115 million samples, our results in Table 2 indicate that for the purpose of dataset pruning, using curated image-text pairs results in a better alignment score than using a much larger noisy dataset. This highlights the importance of using a captioning model pretrained on higher quality data for large-scale dataset pruning.", + "bbox": [ + 75, + 512, + 468, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Captioning models We conduct experiments to investigate the generalizability of Sieve to other captioning models. In Table 3, we compare GIT [38], which utilizes an image encoder, and a text decoder captioning model, pretrained on 10 million image-text pairs to BLIP [20], which utilizes an image encoder, text encoder and a text decoder and is pretrained on 14 million image-text pairs. We observe that fusing Sieve with CLIPScore based on GIT and BLIP improves retrieval performance by $3.13\\%$ and $3.83\\%$ and improves average performance on 38 downstream tasks by $1.07\\%$ and $2.52\\%$ respectively.", + "bbox": [ + 75, + 623, + 468, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text embedding space In Table 4, we ablate over embeddings from different text models and show that embeddings from our selected sentence transformer perform better than embeddings from CLIP and BLIP text encoders. The CLIP text encoder was pretrained along with the CLIP vision encoder to map text and images to the same embedding space, and is used in diffusion models to condition image genera", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/808cf490816859cda521407ef5197fe4326cbe647d6a52e7d17e8ee14e54daec.jpg", + "table_caption": [ + "Table 2. Effect of a caption generator's pretraining data-distribution on Sieve. The 14M pretraining dataset consists of curated image-text pairs, while the 129M dataset includes an additional 115M noisy image-text pairs from LAION [36]." + ], + "table_footnote": [], + "table_body": "
FilteringImageNetVTABRetrievalAvg
CLIPScore27.1333.9024.1732.63
Sieve-GIT27.4733.2027.2732.40
+CLIPScore28.9034.8327.3033.70
Sieve-BLIP29.6035.0728.5734.03
+CLIPScore30.3535.9028.0035.15
", + "bbox": [ + 504, + 508, + 888, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Effect of captioning models, GIT [38] pretrained on 10M image-text pairs, and BLIP [20] pretrained on 14M pairs", + "bbox": [ + 498, + 637, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion [33]. However, we observe that the CLIP text encoder suffers from poor semantic textual understanding, leading to a large drop in accuracy when used as a caption similarity measure. BLIP's text encoder performs better than that of CLIP, but the lightweight sentence transformer specifically pretrained on aligning semantically similar texts performs significantly better with $\\geq 2\\%$ improvements across various task types. In Figure 7 in the Appendix we show how cosine similarities of sentence similarity models result in better semantic textual clustering compared to CLIP and BLIP text encoders.", + "bbox": [ + 496, + 684, + 890, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pruning percentage We study the effect of the fraction of samples selected for pretraining. For each experiment, we compute the Sieve alignment score and CLIPScore for each", + "bbox": [ + 498, + 854, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22429", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e95e62bd6f1e0f6eda52720c2fce408e4211dd9bbc7ca2052d9f79518ff85b9d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Text EncoderImageNetImageNet dist shiftVTABRetrievalAverage over 38 datasets
CLIP18.0015.6526.9020.9525.90
BLIP27.2022.1032.7025.4531.85
Sentence Transformer29.6024.9335.0728.5734.03
", + "bbox": [ + 200, + 88, + 771, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7adffcfebeef1609f6edb9e435b1a71c90fdbee32c6ff793493257c422c9b779.jpg", + "image_caption": [ + "(a) Sieve gain over CLIPScore on medium scale pool." + ], + "image_footnote": [], + "bbox": [ + 83, + 253, + 464, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7399286e9c74e23defc786d289f598219c4806b89eb130de48644823e04d2078.jpg", + "image_caption": [ + "(b) Sieve+CLIPScore gain over CLIPScore on medium scale pool.", + "Figure 5. The relative performance gain of Sieve and Sieve+CLIPScore relative to CLIPScore on 38 downstream tasks on the medium scale pool." + ], + "image_footnote": [], + "bbox": [ + 81, + 450, + 464, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sample. The top- $k\\%$ and pretraining CLIP models are then selected. Here, $k\\%$ is set to $10\\%$ , $15\\%$ , $20\\%$ , $25\\%$ and $30\\%$ . Finally, we report the zero-shot performance on ImageNet and the average on 38 tasks in Figure 4. We observe that Sieve achieves the best performance using $20\\%$ of the data, while CLIPScore peaks at $30\\%$ (similar to results reported in 9). Hence, pruning using Sieve achieves better performance with less data, compared to CLIPScore.", + "bbox": [ + 75, + 714, + 468, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of generated captions and fusion with CLIP-Score We study the effect of using multiple captions per image to maximize the alignment of the generated captions with the alt-text. For nucleus sampling [12], we set the", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/221f6635e42131526baefbed106bd2b8e4b357dde5af81c9a1aa69971c795938.jpg", + "table_caption": [ + "Table 4. Effect of the sentence encoder on the performance of Sieve. CLIP uses the text encoder pretrained on 400M samples [36], BLIP uses the text encoder pretrained on the curated 14M samples defined in [20], and Sentence Transformer uses a language model pretrained on unlabeled text corpus [39]. Each encoder encodes generated captions and the alt-text where the textual semantic alignment is computed." + ], + "table_footnote": [], + "table_body": "
Generated CaptionsCLIP weightImageNetAverage over 38 datasets
10.028.6032.50
20.029.0033.40
40.029.5333.70
80.029.6034.03
0.330.1034.40
80.530.3535.15
0.730.2534.35
", + "bbox": [ + 524, + 251, + 867, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Effect of number of generated captions and weight of CLIPScore on zero-shot performance of pretrained CLIP models", + "bbox": [ + 498, + 409, + 892, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cumulative probability of the smallest set of words to 0.9, and the minimum and maximum sequence lengths to 5 and 20, respectively. We study the effect of sampling 1, 2, 4, and 8 captions. For each input image-text pair, we assign the maximum alignment score between the alt-text and the generated captions. We observe in Table 5 that increasing the number of generated captions improves the performance on downstream tasks. We reason that due to the many-to-many relationship between images and captions, generating more captions increases the probability of matching a hard aligned alt-text. We also investigate the effect of fusing the Sieve alignment score with CLIPScore in Table 5. Each score is independently normalized, and a weighted average is applied between the two scores. Finally, the top $20\\%$ of samples ranked by Sieve+CLIPScore are selected. We observe that a weight of 0.5 achieves the best performance.", + "bbox": [ + 496, + 448, + 892, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 707, + 617, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a novel method, Sieve, that enables pruning large-scale noisy web-crawled image-text datasets. We propose utilizing synthetic captions from image captioning models pretrained on small, diverse, and curated datasets to evaluate the alignment of noisy image-text pairs. Using the embedding space of a lightweight sentence transformer, we compute an alignment score between generated captions and alt-text. We demonstrate that Sieve provides a complementary pruning signal to CLIPScore, effectively minimizing false positives and negatives, leading to improved zero-shot classification and retrieval performance.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22430", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S. Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication, 2023. 1, 4", + "[2] Yonatan Bitton, Nitzan Bitton Guetta, Ron Yosef, Yuval Elovici, Mohit Bansal, Gabriel Stanovsky, and Roy Schwartz. WinoGAVil: Gamified association benchmark to challenge vision-and-language models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5", + "[3] Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326, 2015. 4", + "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 5", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 2", + "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5", + "[8] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 3", + "[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortzman, Dhruba Ghosh, Jieyu Zhang, et al. Dat acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1, 2, 3, 5, 8, 12", + "[10] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1", + "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1, 2", + "[12] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In ICLR. OpenReview.net, 2020. 2, 3, 8", + "[13] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23369-23379, 2023. 1, 5", + "[14] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 2", + "[15] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 2", + "[16] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics, 2017. 4", + "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5", + "[18] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022. 5", + "[19] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 11", + "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5, 6, 7, 8", + "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5", + "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 4", + "[24] Pratyush Maini, Sachin Goyal, Zachary C Lipton, J Zico Kolter, and Aditi Raghunathan. T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132, 2023. 3, 11", + "[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6, 11" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22431", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. arXiv preprint arXiv:2307.10350, 2023. 1, 3, 4", + "[27] OpenAI. The country211 dataset, .6", + "[28] OpenAI. The rendered sst2 dataset, .6", + "[29] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5", + "[30] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6967-6977, 2023. 2, 3, 11", + "[31] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 4", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 7", + "[34] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 4", + "[35] William A Gaviria Rojas, Sudnya Diamos, Keertan Ranjan Kini, David Kanter, Vijay Janapa Reddi, and Cody Coleman. The dollar street dataset: Images representing the geographic and socioeconomic diversity of the world. In Thirtysixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 11", + "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 3, 6, 7, 8", + "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5", + "[38] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 7", + "[39] Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, 2020. 5, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. 2023. 1, 2", + "[41] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In Proceedings of the 40th International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 1", + "[42] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. 2023. 1, 5", + "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5", + "[44] Haichao Yu, Yu Tian, Sateesh Kumar, Linjie Yang, and Heng Wang. The devil is in the details: A deep dive into the rabbit hole of data filtering. arXiv preprint arXiv:2309.15954, 2023. 3", + "[45] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5" + ], + "bbox": [ + 501, + 92, + 890, + 484 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22432", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_model.json b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f9c6d2d0c2c852eb471ef68cb423a522e17d55fc --- /dev/null +++ b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.83, + 0.153 + ], + "angle": 0, + "content": "Sieve: Multimodal Dataset Pruning Using Image Captioning Models" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.18, + 0.754, + 0.2 + ], + "angle": 0, + "content": "Anas Mahmoud1,2* Mostafa Elhoushi1 Amro Abbas† Yu Yang3" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.2, + 0.695, + 0.216 + ], + "angle": 0, + "content": "Newsha Ardalani1 Hugh Leather1 Ari S. Morcos4" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.216, + 0.79, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)FAIR at Meta \\(^{2}\\)University of Toronto \\(^{3}\\)UC Los Angeles \\(^{4}\\)DatologyAI" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning. We argue that this approach suffers from multiple limitations including: false positives and negatives due to CLIP's pretraining on noisy labels. We propose a pruning signal, Sieve, that employs synthetic captions generated by image captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on unlabeled text corpus. Using DataComp, a multimodal dataset filtering benchmark, when evaluating on 38 downstream tasks, our pruning approach, surpasses CLIPScore by \\(2.6\\%\\) and \\(1.7\\%\\) on medium and large scale respectively. In addition, on retrieval tasks, Sieve leads to a significant improvement of \\(2.7\\%\\) and \\(4.5\\%\\) on medium and large scale respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.691, + 0.21, + 0.707 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) [32] models have shown great success in solving zero-shot image classification and multimodal retrieval tasks. In addition, many foundational Vision-Language Models (VLMs) use pretrained CLIP encoders to condition image generation on CLIP text embeddings [33] in retrieval augmented vision-language models [13, 42], and to align modalities including audio, depth, and thermal with language through CLIP image embeddings [10]. Therefore, the quality of CLIP representations can influence the performance of many VLMs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.892, + 0.392 + ], + "angle": 0, + "content": "To pretrain CLIP, billions of image-text pairs are collected using common crawl. The raw data is highly diverse but contains many noisy image-text pairs, including low quality images, low quality alternative text (alt-text), and misaligned image-text pairs. Pretraining CLIP models on noisy data can have adverse effects on the learned representations, thus leading to poor performance on downstream tasks [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.893, + 0.562 + ], + "angle": 0, + "content": "To address this challenge, researchers have developed data pruning methods to remove low quality image-text pairs. Heuristics that filter out image-text pairs based on image dimensions, aspect ratio, alt-text length, and complexity are commonly used [9, 36] to reduce noise, but can also limit the diversity of the dataset [26]. Methods that use images or class names from datasets, like ImageNet, to sample semantically similar image-text pairs can lead to higher accuracy on downstream tasks [40], but limit the diversity of the selected samples as they sample image-text pairs close to a specific dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.777 + ], + "angle": 0, + "content": "One of the most effective pruning methods, CLIP-Score [11, 36], computes the cosine similarity between image and text embeddings using a pretrained CLIP model. This score is then used to rank the alignment of image-text pairs. However, as shown in Figure 1, using CLIPScore can lead to false positives – samples that are poorly aligned but have high CLIPScore (i.e., bad samples) due to spurious correlations learned by the pretrained CLIP model [41]. In addition, using CLIPScore can lead to false negatives – samples that are aligned but have low CLIPScore (i.e., hard samples) due to the poor discrimination between hard and bad samples. Excluding hard samples and including bad samples can negatively affect the generalization of CLIP image and text encoders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The goal of this work is to reduce both false negatives and positives induced by CLIPScore ranking by relying on an image-captioning model pretrained on small, diverse, and well-aligned image-text pairs. As shown in Figure 1, Sieve can reduce false positives or false negatives in different scenarios, such as samples where CLIPScore focuses on geographic or chronological context of an image rather than the content of an image. As depicted in Figure 2, we evalu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.302, + 0.888 + ], + "angle": 0, + "content": "*Work done during internship at Meta" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.33, + 0.9 + ], + "angle": 0, + "content": "†Work done during an AI residency at Meta" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.33, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22423" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.097, + 0.303, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.199, + 0.299, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.322, + 0.097, + 0.359, + 0.105 + ], + "angle": 0, + "content": "ALT-Text:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.106, + 0.461, + 0.112 + ], + "angle": 0, + "content": "The Best Places to Stay in Kaikoura, New" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.113, + 0.352, + 0.119 + ], + "angle": 0, + "content": "Zealand" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.124, + 0.382, + 0.131 + ], + "angle": 0, + "content": "CLIPScore:0.30" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.132, + 0.374, + 0.139 + ], + "angle": 0, + "content": "False Positive" + }, + { + "type": "list", + "bbox": [ + 0.323, + 0.124, + 0.382, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.151, + 0.392, + 0.158 + ], + "angle": 0, + "content": "Generated Caption:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.159, + 0.475, + 0.165 + ], + "angle": 0, + "content": "a living room with a large window and a couch" + }, + { + "type": "list", + "bbox": [ + 0.323, + 0.151, + 0.475, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.178, + 0.363, + 0.185 + ], + "angle": 0, + "content": "Sieve: 0.12" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.186, + 0.373, + 0.193 + ], + "angle": 0, + "content": "True Negative" + }, + { + "type": "list", + "bbox": [ + 0.323, + 0.178, + 0.373, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.207, + 0.358, + 0.213 + ], + "angle": 0, + "content": "ALT-Text:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.215, + 0.47, + 0.222 + ], + "angle": 0, + "content": "The area were evacuated and Keiserville Rd." + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.223, + 0.452, + 0.229 + ], + "angle": 0, + "content": "in Washington Twp. Pa., was closed as" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.23, + 0.446, + 0.235 + ], + "angle": 0, + "content": "crews worked on the Marcellus street" + }, + { + "type": "list", + "bbox": [ + 0.323, + 0.207, + 0.47, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.241, + 0.382, + 0.248 + ], + "angle": 0, + "content": "CLIPScore: 0.31" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.249, + 0.374, + 0.255 + ], + "angle": 0, + "content": "False Positive" + }, + { + "type": "list", + "bbox": [ + 0.323, + 0.241, + 0.382, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.265, + 0.393, + 0.272 + ], + "angle": 0, + "content": "Generated Caption:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.273, + 0.45, + 0.279 + ], + "angle": 0, + "content": "a view of a construction site with water" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.279, + 0.43, + 0.285 + ], + "angle": 0, + "content": "coming out of it" + }, + { + "type": "list", + "bbox": [ + 0.323, + 0.265, + 0.45, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.292, + 0.362, + 0.299 + ], + "angle": 0, + "content": "Sieve: 0.21" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.299, + 0.373, + 0.306 + ], + "angle": 0, + "content": "True Negative" + }, + { + "type": "list", + "bbox": [ + 0.323, + 0.292, + 0.373, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.616, + 0.099, + 0.651, + 0.105 + ], + "angle": 0, + "content": "ALT-Text:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.567, + 0.106, + 0.651, + 0.113 + ], + "angle": 0, + "content": "The Great Sphinx of Giza" + }, + { + "type": "image_footnote", + "bbox": [ + 0.594, + 0.127, + 0.651, + 0.134 + ], + "angle": 0, + "content": "CLIPScore:0.25" + }, + { + "type": "image_footnote", + "bbox": [ + 0.599, + 0.134, + 0.65, + 0.141 + ], + "angle": 0, + "content": "False Negative" + }, + { + "type": "image_footnote", + "bbox": [ + 0.582, + 0.155, + 0.651, + 0.161 + ], + "angle": 0, + "content": "Generated Caption:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.515, + 0.162, + 0.651, + 0.169 + ], + "angle": 0, + "content": "the great sphinx and the pyramids of giza" + }, + { + "type": "image_footnote", + "bbox": [ + 0.611, + 0.175, + 0.651, + 0.181 + ], + "angle": 0, + "content": "Sieve: 0.91" + }, + { + "type": "image_footnote", + "bbox": [ + 0.603, + 0.182, + 0.651, + 0.188 + ], + "angle": 0, + "content": "True Positive" + }, + { + "type": "image_footnote", + "bbox": [ + 0.616, + 0.211, + 0.651, + 0.217 + ], + "angle": 0, + "content": "ALT-Text:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.567, + 0.218, + 0.651, + 0.224 + ], + "angle": 0, + "content": "A view of the interior pool" + }, + { + "type": "image_footnote", + "bbox": [ + 0.594, + 0.245, + 0.651, + 0.251 + ], + "angle": 0, + "content": "CLIPScore: 0.23" + }, + { + "type": "image_footnote", + "bbox": [ + 0.598, + 0.252, + 0.65, + 0.258 + ], + "angle": 0, + "content": "False Negative" + }, + { + "type": "image_footnote", + "bbox": [ + 0.582, + 0.269, + 0.651, + 0.275 + ], + "angle": 0, + "content": "Generated Caption:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.507, + 0.276, + 0.651, + 0.283 + ], + "angle": 0, + "content": "a large indoor swimming pool with a skylight" + }, + { + "type": "image_footnote", + "bbox": [ + 0.61, + 0.289, + 0.651, + 0.295 + ], + "angle": 0, + "content": "Sieve: 0.65" + }, + { + "type": "image_footnote", + "bbox": [ + 0.604, + 0.296, + 0.65, + 0.302 + ], + "angle": 0, + "content": "True Positive" + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.095, + 0.885, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.212, + 0.885, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Figure 1. Examples of image-text pairs in which the scores of pretrained CLIP models, a commonly used image-text data filtering approach, fail to measure their alignment. Our proposed approach, Sieve, provides an accurate alignment score using a caption generator and sentence transformer. Top left and bottom left: Examples of false positives where alt-text describes concepts that are not found or unrelated to the image. CLIP is trained on similar noisy image-text pairs, thus, it assigns a relatively high score. Sieve can detect that such image-text pairs are misaligned. Top right and bottom right: Examples of false negatives where images are aligned with the alt-text but are assigned low CLIP scores, either due to the low likelihood of these pairs in the pretraining data, or because CLIP may have seen similar images aligned with other noisy labels. Sieve can detect that such image-text pairs are well-aligned and selects them for pretraining." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.473, + 0.827 + ], + "angle": 0, + "content": "ate the alignment of web-crawled image-text pairs by first, generating multiple captions for each image using nucleus sampling [12], followed by removing phrases that describe the medium (e.g., \"an image of\", \"a photo of\") rather than visual concepts. Finally, to evaluate semantic similarity between the limited diversity of generated captions and the high diversity of alt-text, we utilize the embedding space of a lightweight sentence transformer pretrained on unlabeled text corpus. The alignment between the generated captions and the alt-text is then used as a proxy for image-text alignment. Our goal is to fuse this alignment signal with CLIP-Score to minimize false positives and negatives, leading to a more aligned pretraining dataset. To evaluate the effectiveness of our proposed pruning method, we utilize the DataComp [9] benchmark, which fixes the pretraining hyperparameters of CLIP and provides multiple candidate pools of noisy image-text data for pretraining CLIP models. The goal is to select a subset of noisy image-text data that leads to the best performance on 38 downstream tasks. Using image-captioning model alignment scores fused with CLIP-Score, we surpass CLIPScore filtering by \\(2.6\\%\\) and \\(1.7\\%\\) on the average of 38 downstream tasks on medium and large scale pool respectively. In addition, on multimodal retrieval tasks, our approach leads to an improvement of \\(2.7\\%\\) and \\(4.5\\%\\) on medium and large scale respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.844, + 0.219, + 0.859 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Heuristics are basic filtering methods including: filtering non-English alt-text using fastText [15], filtering alt-text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.57 + ], + "angle": 0, + "content": "with a few words [9, 36], filtering alt-text with low text complexity [30], and filtering images by size or aspect ratio [9]. A combination of these unimodal filtering approaches has been explored by DataComp [9]. An example of a multimodal filtering approach is text spotting: detecting and recognizing text in images and filtering image-text pairs with high overlap between spotted text (text detected in image) and alt-text (associated label of image) [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.893, + 0.85 + ], + "angle": 0, + "content": "Datasets as Priors was proposed in DataComp [9], relying on sampling image-text pairs that are semantically similar to diverse and curated datasets like ImageNet [5]. Text-based sampling selects image-text pairs with alt-text overlapping one of the ImageNet classes. CiT [40] uses cosine similarity to filter alt-text that are similar to ImageNet classes. Image-based sampling approaches encode images from the unfiltered candidate pool using the OpenAI CLIP's ViT-L/14 vision encoder, and clusters the images into 100,000 groups using FAISS [14]. Then, embeddings of ImageNet training samples are used to keep the closest cluster to each sample. The main limitation of such approaches is that they bias the CLIP model and may not generalize well to new downstream tasks. We argue that selecting samples that match the distribution of downstream tasks encourages overfitting to the evaluation set and, thus limits generalization to other downstream tasks. Our approach, Sieve, does not use any dataset as a prior." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Pretrained VLMs One of the most successful methods for evaluating image-text alignment is CLIPScore [11]. LAION filtering [36] uses an OpenAI CLIP model [32] pre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22424" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "trained on 400 million image-text pairs to evaluate image-text alignment of large webscale datasets, and filter out samples with the lowest CLIPScore. Filtering using CLIP-Score can suffer from false negatives, which leads to filtering out hard informative samples, and false positives, which leads to including misaligned samples. Another approach proposes a non-filtering approach that utilizes pretrained VLMs [26], using large image-captioning models like BLIP2 [21] to replace alt-text labels with descriptive synthetic captions. The synthetic captions are then used to train CLIP models. The authors [26] demonstrate that at scale, the improvement of synthetic captions is capped by the limited diversity of generated captions compared to the high diversity of noisy text labels. Compared to [26], we do not alter the original alt-text and thus our focus is on the dataset pruning challenge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.651 + ], + "angle": 0, + "content": "Concurrent Work Inspired by text spotting [30], T-MARS [24] is concurrent work that detects and masks text regions in images before computing CLIPScore, resulting in improved visual representations. We reason that T-MARS is orthogonal to our approach as we can apply the same masking before calculating Sieve. In addition, Devil is In the Details (DID) [44], utilizes a combination of multiple filters and sampling approaches, which is orthogonal to our approach as Sieve can be combined to other filters. Moreover, most of DID's accuracy improvements is from aligning the selected data distribution with the DataComp [9] evaluation set (downstream tasks), which overfits downstream tasks as discussed in previous sections. More importantly, T-MARS [24] and DID [9] only perform pruning experiments on the medium scale of DataComp [9], while Sieve shows results on both medium and large scales. Finally, DFN [8], uses CLIPScore from a CLIP model pretrained on a private 357 million human verified image-text labels, which is \\(25 \\times\\) bigger than the dataset our selected captioning models were pretrained on, and is bigger than DataComp's medium scale pool." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.667, + 0.214, + 0.684 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Let \\(\\mathcal{D} = \\{(I_i,T_i)\\}_{i = 1}^N\\) denote an uncurated dataset consisting of \\(N\\) image-text pairs crawled from the web. Our goal is to curate a dataset, \\(\\mathcal{D}' = \\{(I_{i'},T_{i'})\\}_{i' = 1}^{N'}\\), that is a subset of the pool, \\(\\mathcal{D}'\\subseteq \\mathcal{D}\\) \\(N^{\\prime}\\leq N\\), to train a new CLIP model from uninitialized weights, \\(\\Theta_0\\), to new weights, \\(\\Theta^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.781, + 0.47, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {\\prime} = \\operatorname {t r a i n} \\left(\\Theta_ {0}, \\mathcal {D} ^ {\\prime}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.47, + 0.873 + ], + "angle": 0, + "content": "For a given scoring function, \\( f \\), that maps an image-text pair to a scalar value, \\( s = f(I_i, T_i) \\), we express a pruning function that selects a fraction, \\( k \\), of dataset, \\( \\mathcal{D} \\), using function, \\( f \\):" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.885, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\Pr n e _ {f} (\\mathcal {D}, k) \\quad \\text {s . t .} 0 \\leq k \\leq 1, f: (I _ {i}, T _ {i}) \\rightarrow \\mathbb {R} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "where \\(\\mathrm{prune}_f(\\mathcal{D},k)\\) applies function, \\(f\\), on each image-text sample in \\(\\mathcal{D}\\) to obtain a score for each sample, ranks the scores in descending order, and returns a set of the top \\(k\\) portion of the samples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.153, + 0.894, + 0.257 + ], + "angle": 0, + "content": "One common approach for pruning is CLIPScore [9, 36]. Let \\( E \\) be a CLIP model consisting of an image encoder, \\( E_{\\mathrm{image}} \\), that maps an image, \\( I \\), to an embedding vector, \\( E_{\\mathrm{image}}(I) \\in \\mathbb{R}^d \\), and a text encoder, \\( E_{\\mathrm{text}} \\), that maps a text sample, \\( T \\), to an embedding vector, \\( E_{\\mathrm{text}}(T) \\in \\mathbb{R}^d \\). CLIP-Score is a measure of alignment between \\( I \\) and \\( T \\), and is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.269, + 0.892, + 0.288 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {C L I P}} (I, T) = \\left\\langle E _ {\\text {i m a g e}} (I), E _ {\\text {t e x t}} (T) \\right\\rangle \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.417 + ], + "angle": 0, + "content": "where \\(\\langle \\mathbf{x},\\mathbf{y}\\rangle\\) is the cosine similarity between two vectors, \\(\\mathbf{x}\\) and \\(\\mathbf{y}\\), which is defined as the dot product of the \\(l_{2}\\) normalized vectors. The most common CLIP model used for pruning is pretrained on 400 million noisy image-text pairs [36]. Our proposed pruning method, Sieve attempts to minimize the false positives and negatives induced by CLIPScore filtering. Sieve consists of two main components: ImageCaptioning and Sentence Transformer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Image-Captioning Let \\( G \\) be a captioning model that generates text, \\( T_{i}^{G} \\), describing the content of image, \\( I_{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.456, + 0.892, + 0.475 + ], + "angle": 0, + "content": "\\[\nT _ {i} ^ {G} = G \\left(I _ {i}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Given a captioning model pre-trained on a small, representative and well-aligned dataset of image-text pairs, we are interested in estimating the alignment between image-text pairs sampled from a very large, diverse but noisy dataset. The alignment score can then be used as a ranking metric for dataset filtering. We hypothesize that:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.576, + 0.892, + 0.621 + ], + "angle": 0, + "content": "- The probability of generating a caption that is semantically similar to the alt-text from an aligned pair is much higher than that from a misaligned pair." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.622, + 0.892, + 0.711 + ], + "angle": 0, + "content": "- The probability of generating a caption that is semantically similar to a hard alt-text is higher than generating a caption that is semantically similar to a misaligned alt-text. Here, a hard alt-text is a text label with low likelihood with respect to the captioning model, but is aligned with the image content." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.576, + 0.892, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.833 + ], + "angle": 0, + "content": "As images can contain multiple objects with complex attributes and relationships, there exist multiple ways to describe their content. Given the inherent many-to-many relationship between images and text labels, our goal is to increase the probability of generating a caption that matches an aligned alt-text. To achieve this, we utilize nucleus sampling [12], a decoding strategy used to sample multiple captions, \\( r \\), per-image:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.842, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\nG (I, r) = \\left\\{T _ {0} ^ {G}, T _ {1} ^ {G}, \\dots , T _ {r - 1} ^ {G} \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Sentence Transformer Given an image, its alt-text, and a set of generated captions, our goal is to estimate the align-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22425" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.098, + 0.845, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.895, + 0.365 + ], + "angle": 0, + "content": "Figure 2. Our proposed framework enables dataset pruning using image captioning models. To evaluate the alignment of a noisy image-text pair, we generate multiple captions per image using nucleus sampling. Then medium phrases, like \"an image of\" or \"a photo of\", are masked from alt-text and generated captions. Finally, a lightweight sentence encoder is used to semantically compare the generated captions with alt-text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.661 + ], + "angle": 0, + "content": "ment between the generated captions and the alt-text. However, there is a very large diversity gap between the generated captions and the highly diverse alt-text as measured by the number of unique nouns and trigrams [26]. On the other hand, constructing a large, diverse and curated image-text dataset is expensive, which limits the diversity of the generated captions. We propose to bridge this gap by utilizing a light-weight sentence similarity model to encode the alt-text and the generated captions. We expect the semantically similar alt-text and generated caption pairs to be closely clustered in the embedding space compared to semantically distinct pairs. We reason that the rich semantic textual embedding space of the sentence similarity model enables pretraining the captioning model only on a small but curated image-text dataset. Thus, we rely on the semantic understanding of the sentence similarity model to bridge the gap between the limited diversity in the generated captions and the highly diverse alt-text labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.472, + 0.753 + ], + "angle": 0, + "content": "To estimate the alignment score, we compute the cosine similarity between embeddings of each generated caption and text label. Let \\( S \\) be a language model that encodes a text sample, \\( T \\), to a vector, \\( S(T) \\in \\mathbb{R}^d \\). We define the alignment between two text samples, \\( T_a \\) and \\( T_b \\), as the cosine similarity between their language model encodings:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.763, + 0.47, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\langle S \\left(T _ {a}\\right), S \\left(T _ {b}\\right) \\rangle \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.789, + 0.471, + 0.82 + ], + "angle": 0, + "content": "This estimate can then be used as a proxy for the image-text alignment of an image, \\(I\\), and text, \\(T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.829, + 0.332, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\langle S (G (I)), S (T) \\rangle\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "If the image captioning model generates \\( r \\) different caption candidates for an image, \\( I \\), we can use the maximum alignment between each of the generated captions," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.388, + 0.862, + 0.407 + ], + "angle": 0, + "content": "\\(G(I,r) = \\{T_0^G,T_1^G,\\ldots ,T_{r - 1}^G\\}\\), and a text sample, \\(T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.413, + 0.892, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\max _ {T _ {j} ^ {G} \\in G (I, r)} \\langle S \\left(T _ {j} ^ {G}\\right), S (T) \\rangle \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.869 + ], + "angle": 0, + "content": "In literature, there are different models and approaches to obtain text embeddings. [16] uses the average of N-gram features of each word in a text sample to obtain an embedding. A more common option is to use the logits of the last token generated by a decoder-only language model, which is the approach taken with CLIP's text encoder [32], in GPT-1 [31], as well as in [1]. Encoder-only models, such as BERT [6] or RoBERTa [23], can also be used, where the embedding vector may be either the logits of the classification token, or the average pool of the logits of all tokens. Although such language models may have strong generation or classification capabilities, they were not optimized for sentence similarity tasks, but either for next word prediction (i.e., causal language modeling) or masked word prediction (i.e., masked language modeling) tasks. Therefore, their embeddings may not be ideal to measure alignment between sentences. More importantly, such models are large in size and hence slow to infer on large datasets. A language model finetuned on a sentence similarity task, such as SNLI [3], aligns with the goal of estimating semantic textual similarity between alt-text and captions. We find that sentence similarity models [34] pretrained using a self-supervised instance discrimination task on unlabeled corpus of sentences perform well in estimating the alignment between text pairs, and are lightweight in size and latency (e.g., tens of millions of parameters in contrast to billions of parameters of performant decoder-only large language models)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Masking Medium Words Phrases such as \"image of\", \"picture of\", or \"photo of\" can appear in either alt-text or" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22426" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.378 + ], + "angle": 0, + "content": "generated captions. We refer to such phrases as \"medium phrases\", as they describe the medium rather than the contents of an image. We notice that the existence of such medium phrases adds noise to the sentence similarity score, as shown in Figure 3. A pair of sentences that each have a medium phrase are assigned a misleadingly high sentence similarity score by a sentence transformer, as they have been trained on a wide and diverse corpus of text, rather than on image captions. Hence, the existence of medium phrases may increase their attention to the topic of images or media, rather than the topic of the content of such images. Therefore, we neutralize the effect of medium phrases by removing them from both alt-text and generated captions. We express the operation of masking medium words in a text sample, \\( M(T) \\), on text, \\( T \\), as masking all possible contiguous subsequences of the text, where masking on a phrase, \\( t \\), removes it if it is in the pre-determined list of medium phrases, \\( \\mathcal{M} = \\{\\text{\"image of\" }, \\text{\"picture of\" }, \\text{\"photo of\" }, \\ldots \\} \\), or keeps it otherwise." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.471, + 0.409 + ], + "angle": 0, + "content": "Putting it all together, we define the Sieve score function as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.418, + 0.47, + 0.448 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {S i e v e}} (I, T) = \\max _ {T _ {j} ^ {G} \\in G (I, r)} \\left\\langle S \\left(M \\left(T _ {j} ^ {G}\\right)\\right), S (M (T)) \\right\\rangle \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.455, + 0.47, + 0.487 + ], + "angle": 0, + "content": "The dataset pruned using Sieve with the top \\( k \\) portion of its samples can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.497, + 0.47, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\text {S i e v e}, k} = \\operatorname {p r u n e} _ {f _ {\\text {S i e v e}}} (\\mathcal {D}, k) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.523, + 0.47, + 0.556 + ], + "angle": 0, + "content": "We summarize our approach in Figure 2 and as psuedocode in Algorithm 1 in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.566, + 0.21, + 0.584 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.303, + 0.607 + ], + "angle": 0, + "content": "4.1. Training and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.825 + ], + "angle": 0, + "content": "We utilize the DataComp benchmark to evaluate the utility of image-captioning models for multimodal dataset pruning. Two candidate pools are considered, the medium and the large scale, consisting of 128 million and 1.28 billion image-text pairs, respectively. To train CLIP models, we use DataComp's hyperparameters and architectures to standardize training [9]: \\(5 \\times 10^{-4}\\) learning rate, 500 iterations warmup, AdamW optimizer, for medium scale: ViT-B/32 image encoder [7], batch size 4096, 128M training samples as a compute budget, and for large scale: ViT-B/16 image encoder, batch size 8192, 1.28B training samples as a compute budget. We evaluate the zero-shot performance on 38 downstream tasks, including classification and retrieval tasks [18, 32, 45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.901 + ], + "angle": 0, + "content": "For our captioning model, we utilize BLIP with ViT-B/16 image encoder pretrained on 14 million image-text pairs [20], including highly curated web datasets: Conceptual Captions [37], Conceptual 12M [4] and SBU captions [29]; as well as small human-annotated datasets:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.168 + ], + "angle": 0, + "content": "COCO, and Visual Genome [17]. To compute the alignment between generated captions and alt-text, we use a lightweight distilled sentence transformer, all-MiniLM-L6-v2 [39], further finetuned using self-supervised contrastive learning on unlabeled text corpus." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.176, + 0.64, + 0.191 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.351 + ], + "angle": 0, + "content": "Table 1 reports multiple baselines from DataComp [9], including applying no filtering, basic filtering, and CLIP-Score filtering. On the medium scale, Sieve with an image-captioning model pretrained on 30 times less but curated data surpasses CLIPScore by \\(1.8\\%\\) on average. In addition, we demonstrate that Sieve's alignment score provides a complimentary signal to CLIPScore by fusing Sieve with CLIPScore. We apply min-max normalization to Sieve alignment scores and CLIPScore independently, then take the per-sample weighted average of both scores:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.38, + 0.875, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {\\text {S i e v e} + \\text {C L I P}} (I, T) = \\\\ (1 - \\alpha) \\times \\bar {f} _ {\\text {S i e v e}} (I, T) + \\alpha \\times \\bar {f} _ {\\text {C L I P}} (I, T) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.443, + 0.875, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} \\text {s . t .} \\quad \\overline {{f}} (I, T) = \\end{array} \\\\ \\frac {f (I , T) - \\min _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i})}{\\max _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i}) - \\min _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i})} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.704 + ], + "angle": 0, + "content": "where the weight \\(\\alpha\\) used in the reported results is 0.5. Finally, we select the top \\(20\\%\\) of samples. We observe that the fused approach improves CLIPScore average performance by \\(2.6\\%\\) and \\(1.7\\%\\) on medium and large scale respectively, demonstrating the complementarity of Sieve's alignment score. Moreover, Sieve without fusion achieves the best performance on retrieval tasks on both medium and large scale experiments. Relative to CLIPScore, a significant portion of Sieve's retrieval gain is effectively transferred when fusing Sieve with CLIPScore (i.e., \\(+2.7\\%\\) and \\(+4.5\\%\\) on medium and large scale). This is also promising as retrieval performance is critical for retrieval augmented VLMs [13, 42]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.712, + 0.71, + 0.727 + ], + "angle": 0, + "content": "4.3. Per-Task Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Figure 5 shows the change in accuracy introduced by Sieve as well as Sieve+CLIPScore on each task compared to CLIPScore on medium scale, and Figure 6 of the Appendix shows for the large scale. We observe that in addition to outperforming on image retrieval tasks, Flickr [43], and MS COCO [22], Sieve's greatest performance boost comes from WingoGAViL [2], a retrieval task which requires diverse reasoning skills, including general knowledge, common sense, and abstraction. This high performance can be attributed to Sieve's preference towards keeping samples where the alt-text correctly describes visual concepts and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22427" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.218, + 0.102, + 0.755, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 3. Masking medium phrases improves the ranking of sentence similarity scores. On the left, sentence pairs with misleadingly high (or low) sentence similarity due to the existence (or absence) of medium phrases are highlighted in dark red (or light red). On the right, similarity scores that are more aligned with semantics are highlighted in dark green. The sentence similarity scores are computed using the all-MiniLM-L6-v2 sentence transformer [39]." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.339, + 0.892, + 0.587 + ], + "angle": 0, + "content": "
ScaleFilteringDataset SizeImageNetImageNet dist. shiftsVTABRetrievalAverage over 38 datasets
Medium (128 Million)No Filtering128M17.615.225.921.925.8
Basic Filtering30M22.619.328.425.128.5
LAION Filtering13M23.019.830.723.329.2
CLIPScore38M27.323.033.825.132.8
Sieve24M29.425.035.228.934.6
Sieve+CLIPScore24M30.325.436.227.835.4
Large (1.28 Billion)No Filtering1.28B45.937.842.641.943.7
Basic Filtering298M51.642.344.648.045.8
LAION Filtering130M55.345.351.049.550.1
CLIPScore384M57.847.453.846.652.9
Sieve235M57.347.852.052.052.3
Sieve+CLIPScore235M59.749.154.851.154.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.596, + 0.895, + 0.655 + ], + "angle": 0, + "content": "Table 1. Zero-shot performance of CLIP models pretrained using various filtering strategies on medium and large scale pools of the DataComp benchmark. Sieve fused with CLIPScore beats CLIPScore by \\(2.6\\%\\) and \\(1.7\\%\\) on medium and large scale respectively. In addition, on retrieval tasks, Sieve achieves best performance on both scales and after fusing with CLIPScore, leads an improvement on retrieval over CLIPScore of \\(2.7\\%\\) and \\(4.5\\%\\) on medium and large scale respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.471, + 0.725 + ], + "angle": 0, + "content": "their attributes and relations. Sieve, especially when combined with CLIPScore, significantly outperforms on medical diagnosis tasks, Camelyon17 and PatchCamelyon." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Sieve mainly underperforms in tasks requiring parsing text from images, such as MNIST [19], SVHN [25], and Rendered SST-2 [28], concluding that Sieve is less likely to select image-text pairs that are useful for OCR tasks. In addition, Sieve underperforms CLIPScore on context-based tasks like Country211 [27], a task assessing the geolocation capability of visual representations, demonstrating Sieve's preference towards selecting samples based on the alignment of alt-text with visual concepts rather than context. Interestingly, when fusing with CLIPScore, we improve the performance of all these tasks while retaining the advantage" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.679, + 0.818, + 0.695 + ], + "angle": 0, + "content": "of Sieve especially at large scale (see Figure 6)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.666, + 0.723 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.733, + 0.892, + 0.763 + ], + "angle": 0, + "content": "We conduct studies on the medium scale pool and report the average of three runs per experiment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Pretraining data-distribution We study the effect of the pretraining data distribution used to train the captioning model on the quality of the alignment score. This is measured based on the downstream performance of the CLIP model trained using the selected image-text pairs. Two pretraining data distributions proposed by BLIP [20] are investigated. The first uses 14 million curated image-text pairs, while the second uses an additional 115 million web images with noisy alt-text [36]. Although the original BLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22428" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.218, + 0.108, + 0.472, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.109, + 0.74, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.302 + ], + "angle": 0, + "content": "Figure 4. Evaluating CLIP models pretrained on different fractions of the top ranked samples based on our proposed approach (Sieve), CLIPScore, and fusing Sieve with CLIPscore (Sieve+CLIPSCore), on medium scale." + }, + { + "type": "table", + "bbox": [ + 0.161, + 0.314, + 0.812, + 0.449 + ], + "angle": 0, + "content": "
Percentage KeptCaption Generator Pretraining DataImageNetImageNet dist. shiftVTABRetrievalAverage over 38 datasets
10BLIP-129M23.0020.6030.2021.4030.00
BLIP-14M26.5022.5032.1023.7530.90
15BLIP-129M25.9522.8032.8024.4032.40
BLIP-14M29.0524.6033.3526.9532.70
20BLIP-129M27.8523.6533.4526.3533.05
BLIP-14M29.6024.9335.0728.5734.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.458, + 0.893, + 0.488 + ], + "angle": 0, + "content": "Table 2. Effect of a caption generator's pretraining data-distribution on Sieve. The 14M pretraining dataset consists of curated image-text pairs, while the 129M dataset includes an additional 115M noisy image-text pairs from LAION [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.62 + ], + "angle": 0, + "content": "work reports higher captioning performance when pretraining on 115 million samples, our results in Table 2 indicate that for the purpose of dataset pruning, using curated image-text pairs results in a better alignment score than using a much larger noisy dataset. This highlights the importance of using a captioning model pretrained on higher quality data for large-scale dataset pruning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Captioning models We conduct experiments to investigate the generalizability of Sieve to other captioning models. In Table 3, we compare GIT [38], which utilizes an image encoder, and a text decoder captioning model, pretrained on 10 million image-text pairs to BLIP [20], which utilizes an image encoder, text encoder and a text decoder and is pretrained on 14 million image-text pairs. We observe that fusing Sieve with CLIPScore based on GIT and BLIP improves retrieval performance by \\(3.13\\%\\) and \\(3.83\\%\\) and improves average performance on 38 downstream tasks by \\(1.07\\%\\) and \\(2.52\\%\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Text embedding space In Table 4, we ablate over embeddings from different text models and show that embeddings from our selected sentence transformer perform better than embeddings from CLIP and BLIP text encoders. The CLIP text encoder was pretrained along with the CLIP vision encoder to map text and images to the same embedding space, and is used in diffusion models to condition image genera" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.51, + 0.889, + 0.628 + ], + "angle": 0, + "content": "
FilteringImageNetVTABRetrievalAvg
CLIPScore27.1333.9024.1732.63
Sieve-GIT27.4733.2027.2732.40
+CLIPScore28.9034.8327.3033.70
Sieve-BLIP29.6035.0728.5734.03
+CLIPScore30.3535.9028.0035.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.638, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Table 3. Effect of captioning models, GIT [38] pretrained on 10M image-text pairs, and BLIP [20] pretrained on 14M pairs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.85 + ], + "angle": 0, + "content": "tion [33]. However, we observe that the CLIP text encoder suffers from poor semantic textual understanding, leading to a large drop in accuracy when used as a caption similarity measure. BLIP's text encoder performs better than that of CLIP, but the lightweight sentence transformer specifically pretrained on aligning semantically similar texts performs significantly better with \\(\\geq 2\\%\\) improvements across various task types. In Figure 7 in the Appendix we show how cosine similarities of sentence similarity models result in better semantic textual clustering compared to CLIP and BLIP text encoders." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Pruning percentage We study the effect of the fraction of samples selected for pretraining. For each experiment, we compute the Sieve alignment score and CLIPScore for each" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22429" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.202, + 0.089, + 0.772, + 0.176 + ], + "angle": 0, + "content": "
Text EncoderImageNetImageNet dist shiftVTABRetrievalAverage over 38 datasets
CLIP18.0015.6526.9020.9525.90
BLIP27.2022.1032.7025.4531.85
Sentence Transformer29.6024.9335.0728.5734.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.186, + 0.895, + 0.23 + ], + "angle": 0, + "content": "Table 4. Effect of the sentence encoder on the performance of Sieve. CLIP uses the text encoder pretrained on 400M samples [36], BLIP uses the text encoder pretrained on the curated 14M samples defined in [20], and Sentence Transformer uses a language model pretrained on unlabeled text corpus [39]. Each encoder encodes generated captions and the alt-text where the textual semantic alignment is computed." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.255, + 0.465, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.435, + 0.414, + 0.448 + ], + "angle": 0, + "content": "(a) Sieve gain over CLIPScore on medium scale pool." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.452, + 0.465, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.631, + 0.447, + 0.644 + ], + "angle": 0, + "content": "(b) Sieve+CLIPScore gain over CLIPScore on medium scale pool." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Figure 5. The relative performance gain of Sieve and Sieve+CLIPScore relative to CLIPScore on 38 downstream tasks on the medium scale pool." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.469, + 0.837 + ], + "angle": 0, + "content": "sample. The top-\\(k\\%\\) and pretraining CLIP models are then selected. Here, \\(k\\%\\) is set to \\(10\\%\\), \\(15\\%\\), \\(20\\%\\), \\(25\\%\\) and \\(30\\%\\). Finally, we report the zero-shot performance on ImageNet and the average on 38 tasks in Figure 4. We observe that Sieve achieves the best performance using \\(20\\%\\) of the data, while CLIPScore peaks at \\(30\\%\\) (similar to results reported in 9). Hence, pruning using Sieve achieves better performance with less data, compared to CLIPScore." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Number of generated captions and fusion with CLIP-Score We study the effect of using multiple captions per image to maximize the alignment of the generated captions with the alt-text. For nucleus sampling [12], we set the" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.252, + 0.869, + 0.399 + ], + "angle": 0, + "content": "
Generated CaptionsCLIP weightImageNetAverage over 38 datasets
10.028.6032.50
20.029.0033.40
40.029.5333.70
80.029.6034.03
0.330.1034.40
80.530.3535.15
0.730.2534.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.41, + 0.893, + 0.438 + ], + "angle": 0, + "content": "Table 5. Effect of number of generated captions and weight of CLIPScore on zero-shot performance of pretrained CLIP models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.893, + 0.692 + ], + "angle": 0, + "content": "cumulative probability of the smallest set of words to 0.9, and the minimum and maximum sequence lengths to 5 and 20, respectively. We study the effect of sampling 1, 2, 4, and 8 captions. For each input image-text pair, we assign the maximum alignment score between the alt-text and the generated captions. We observe in Table 5 that increasing the number of generated captions improves the performance on downstream tasks. We reason that due to the many-to-many relationship between images and captions, generating more captions increases the probability of matching a hard aligned alt-text. We also investigate the effect of fusing the Sieve alignment score with CLIPScore in Table 5. Each score is independently normalized, and a weighted average is applied between the two scores. Finally, the top \\(20\\%\\) of samples ranked by Sieve+CLIPScore are selected. We observe that a weight of 0.5 achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.619, + 0.724 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce a novel method, Sieve, that enables pruning large-scale noisy web-crawled image-text datasets. We propose utilizing synthetic captions from image captioning models pretrained on small, diverse, and curated datasets to evaluate the alignment of noisy image-text pairs. Using the embedding space of a lightweight sentence transformer, we compute an alignment score between generated captions and alt-text. We demonstrate that Sieve provides a complementary pruning signal to CLIPScore, effectively minimizing false positives and negatives, leading to improved zero-shot classification and retrieval performance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22430" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S. Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Yonatan Bitton, Nitzan Bitton Guetta, Ron Yosef, Yuval Elovici, Mohit Bansal, Gabriel Stanovsky, and Roy Schwartz. WinoGAVil: Gamified association benchmark to challenge vision-and-language models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.241, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[8] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortzman, Dhruba Ghosh, Jieyu Zhang, et al. Dat acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1, 2, 3, 5, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.69, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[10] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.774, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.83, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[12] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In ICLR. OpenReview.net, 2020. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[13] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23369-23379, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[16] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[18] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[19] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Pratyush Maini, Sachin Goyal, Zachary C Lipton, J Zico Kolter, and Aditi Raghunathan. T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132, 2023. 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6, 11" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22431" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. arXiv preprint arXiv:2307.10350, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.333, + 0.162 + ], + "angle": 0, + "content": "[27] OpenAI. The country211 dataset, .6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.344, + 0.176 + ], + "angle": 0, + "content": "[28] OpenAI. The rendered sst2 dataset, .6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[29] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.234, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[30] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6967-6977, 2023. 2, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[31] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.468, + 0.496 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[34] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.54, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[35] William A Gaviria Rojas, Sudnya Diamos, Keertan Ranjan Kini, David Kanter, Vijay Janapa Reddi, and Cody Coleman. The dollar street dataset: Images representing the geographic and socioeconomic diversity of the world. In Thirtysixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.693, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.79, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[38] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, 2020. 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[40] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[41] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In Proceedings of the 40th International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[42] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[44] Haichao Yu, Yu Tian, Sateesh Kumar, Linjie Yang, and Heng Wang. The devil is in the details: A deep dive into the rabbit hole of data filtering. arXiv preprint arXiv:2309.15954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[45] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22432" + } + ] +] \ No newline at end of file diff --git a/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_origin.pdf b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7666c239b75080c833cd5f291397ad8a83dcf3c6 --- /dev/null +++ b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/6ff5f5b2-831e-43d1-8c97-18e5f39eb692_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f338e548cd7cfdb2748d9b1ada56986ef4bfa7f68a4352ab9ab628b94b41c51 +size 9546033 diff --git a/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/full.md b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a66a1931f5deca58ebf1e914c02de3efdad3ea6 --- /dev/null +++ b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/full.md @@ -0,0 +1,319 @@ +# Sieve: Multimodal Dataset Pruning Using Image Captioning Models + +Anas Mahmoud1,2* Mostafa Elhoushi1 Amro Abbas† Yu Yang3 + +Newsha Ardalani1 Hugh Leather1 Ari S. Morcos4 + +$^{1}$ FAIR at Meta $^{2}$ University of Toronto $^{3}$ UC Los Angeles $^{4}$ DatologyAI + +# Abstract + +Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning. We argue that this approach suffers from multiple limitations including: false positives and negatives due to CLIP's pretraining on noisy labels. We propose a pruning signal, Sieve, that employs synthetic captions generated by image captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on unlabeled text corpus. Using DataComp, a multimodal dataset filtering benchmark, when evaluating on 38 downstream tasks, our pruning approach, surpasses CLIPScore by $2.6\%$ and $1.7\%$ on medium and large scale respectively. In addition, on retrieval tasks, Sieve leads to a significant improvement of $2.7\%$ and $4.5\%$ on medium and large scale respectively. + +# 1. Introduction + +Contrastive Language-Image Pre-training (CLIP) [32] models have shown great success in solving zero-shot image classification and multimodal retrieval tasks. In addition, many foundational Vision-Language Models (VLMs) use pretrained CLIP encoders to condition image generation on CLIP text embeddings [33] in retrieval augmented vision-language models [13, 42], and to align modalities including audio, depth, and thermal with language through CLIP image embeddings [10]. Therefore, the quality of CLIP representations can influence the performance of many VLMs. + +To pretrain CLIP, billions of image-text pairs are collected using common crawl. The raw data is highly diverse but contains many noisy image-text pairs, including low quality images, low quality alternative text (alt-text), and misaligned image-text pairs. Pretraining CLIP models on noisy data can have adverse effects on the learned representations, thus leading to poor performance on downstream tasks [1]. + +To address this challenge, researchers have developed data pruning methods to remove low quality image-text pairs. Heuristics that filter out image-text pairs based on image dimensions, aspect ratio, alt-text length, and complexity are commonly used [9, 36] to reduce noise, but can also limit the diversity of the dataset [26]. Methods that use images or class names from datasets, like ImageNet, to sample semantically similar image-text pairs can lead to higher accuracy on downstream tasks [40], but limit the diversity of the selected samples as they sample image-text pairs close to a specific dataset. + +One of the most effective pruning methods, CLIP-Score [11, 36], computes the cosine similarity between image and text embeddings using a pretrained CLIP model. This score is then used to rank the alignment of image-text pairs. However, as shown in Figure 1, using CLIPScore can lead to false positives – samples that are poorly aligned but have high CLIPScore (i.e., bad samples) due to spurious correlations learned by the pretrained CLIP model [41]. In addition, using CLIPScore can lead to false negatives – samples that are aligned but have low CLIPScore (i.e., hard samples) due to the poor discrimination between hard and bad samples. Excluding hard samples and including bad samples can negatively affect the generalization of CLIP image and text encoders. + +The goal of this work is to reduce both false negatives and positives induced by CLIPScore ranking by relying on an image-captioning model pretrained on small, diverse, and well-aligned image-text pairs. As shown in Figure 1, Sieve can reduce false positives or false negatives in different scenarios, such as samples where CLIPScore focuses on geographic or chronological context of an image rather than the content of an image. As depicted in Figure 2, we evalu + +![](images/a48f7a47ad075a945510158eab7694247ca5a2e6914b0d4a11dc4c331e56fef3.jpg) +ALT-Text: +The Best Places to Stay in Kaikoura, New +Zealand +CLIPScore:0.30 +False Positive +Generated Caption: +a living room with a large window and a couch +Sieve: 0.12 +True Negative + +Figure 1. Examples of image-text pairs in which the scores of pretrained CLIP models, a commonly used image-text data filtering approach, fail to measure their alignment. Our proposed approach, Sieve, provides an accurate alignment score using a caption generator and sentence transformer. Top left and bottom left: Examples of false positives where alt-text describes concepts that are not found or unrelated to the image. CLIP is trained on similar noisy image-text pairs, thus, it assigns a relatively high score. Sieve can detect that such image-text pairs are misaligned. Top right and bottom right: Examples of false negatives where images are aligned with the alt-text but are assigned low CLIP scores, either due to the low likelihood of these pairs in the pretraining data, or because CLIP may have seen similar images aligned with other noisy labels. Sieve can detect that such image-text pairs are well-aligned and selects them for pretraining. +![](images/8a7d643f141acfb48b68768d4a7d88168c82c715f12adccc613339d9c86140a2.jpg) +ALT-Text: +The area were evacuated and Keiserville Rd. +in Washington Twp. Pa., was closed as +crews worked on the Marcellus street +CLIPScore: 0.31 +False Positive +Generated Caption: +a view of a construction site with water +coming out of it +Sieve: 0.21 +True Negative + +![](images/712cfe9d6b8953759f4d7f9329818b324e3fda5c89e6c4445affd74b42657645.jpg) +ALT-Text: +The Great Sphinx of Giza +CLIPScore:0.25 +False Negative +Generated Caption: +the great sphinx and the pyramids of giza +Sieve: 0.91 +True Positive + +![](images/0e55db0e340610e54f5dfcd653d294a5079a7e17f780f6bdf510d955a688a2e0.jpg) +ALT-Text: +A view of the interior pool +CLIPScore: 0.23 +False Negative +Generated Caption: +a large indoor swimming pool with a skylight +Sieve: 0.65 +True Positive + +ate the alignment of web-crawled image-text pairs by first, generating multiple captions for each image using nucleus sampling [12], followed by removing phrases that describe the medium (e.g., "an image of", "a photo of") rather than visual concepts. Finally, to evaluate semantic similarity between the limited diversity of generated captions and the high diversity of alt-text, we utilize the embedding space of a lightweight sentence transformer pretrained on unlabeled text corpus. The alignment between the generated captions and the alt-text is then used as a proxy for image-text alignment. Our goal is to fuse this alignment signal with CLIP-Score to minimize false positives and negatives, leading to a more aligned pretraining dataset. To evaluate the effectiveness of our proposed pruning method, we utilize the DataComp [9] benchmark, which fixes the pretraining hyperparameters of CLIP and provides multiple candidate pools of noisy image-text data for pretraining CLIP models. The goal is to select a subset of noisy image-text data that leads to the best performance on 38 downstream tasks. Using image-captioning model alignment scores fused with CLIP-Score, we surpass CLIPScore filtering by $2.6\%$ and $1.7\%$ on the average of 38 downstream tasks on medium and large scale pool respectively. In addition, on multimodal retrieval tasks, our approach leads to an improvement of $2.7\%$ and $4.5\%$ on medium and large scale respectively. + +# 2. Related Work + +Heuristics are basic filtering methods including: filtering non-English alt-text using fastText [15], filtering alt-text + +with a few words [9, 36], filtering alt-text with low text complexity [30], and filtering images by size or aspect ratio [9]. A combination of these unimodal filtering approaches has been explored by DataComp [9]. An example of a multimodal filtering approach is text spotting: detecting and recognizing text in images and filtering image-text pairs with high overlap between spotted text (text detected in image) and alt-text (associated label of image) [30]. + +Datasets as Priors was proposed in DataComp [9], relying on sampling image-text pairs that are semantically similar to diverse and curated datasets like ImageNet [5]. Text-based sampling selects image-text pairs with alt-text overlapping one of the ImageNet classes. CiT [40] uses cosine similarity to filter alt-text that are similar to ImageNet classes. Image-based sampling approaches encode images from the unfiltered candidate pool using the OpenAI CLIP's ViT-L/14 vision encoder, and clusters the images into 100,000 groups using FAISS [14]. Then, embeddings of ImageNet training samples are used to keep the closest cluster to each sample. The main limitation of such approaches is that they bias the CLIP model and may not generalize well to new downstream tasks. We argue that selecting samples that match the distribution of downstream tasks encourages overfitting to the evaluation set and, thus limits generalization to other downstream tasks. Our approach, Sieve, does not use any dataset as a prior. + +Pretrained VLMs One of the most successful methods for evaluating image-text alignment is CLIPScore [11]. LAION filtering [36] uses an OpenAI CLIP model [32] pre + +trained on 400 million image-text pairs to evaluate image-text alignment of large webscale datasets, and filter out samples with the lowest CLIPScore. Filtering using CLIP-Score can suffer from false negatives, which leads to filtering out hard informative samples, and false positives, which leads to including misaligned samples. Another approach proposes a non-filtering approach that utilizes pretrained VLMs [26], using large image-captioning models like BLIP2 [21] to replace alt-text labels with descriptive synthetic captions. The synthetic captions are then used to train CLIP models. The authors [26] demonstrate that at scale, the improvement of synthetic captions is capped by the limited diversity of generated captions compared to the high diversity of noisy text labels. Compared to [26], we do not alter the original alt-text and thus our focus is on the dataset pruning challenge. + +Concurrent Work Inspired by text spotting [30], T-MARS [24] is concurrent work that detects and masks text regions in images before computing CLIPScore, resulting in improved visual representations. We reason that T-MARS is orthogonal to our approach as we can apply the same masking before calculating Sieve. In addition, Devil is In the Details (DID) [44], utilizes a combination of multiple filters and sampling approaches, which is orthogonal to our approach as Sieve can be combined to other filters. Moreover, most of DID's accuracy improvements is from aligning the selected data distribution with the DataComp [9] evaluation set (downstream tasks), which overfits downstream tasks as discussed in previous sections. More importantly, T-MARS [24] and DID [9] only perform pruning experiments on the medium scale of DataComp [9], while Sieve shows results on both medium and large scales. Finally, DFN [8], uses CLIPScore from a CLIP model pretrained on a private 357 million human verified image-text labels, which is $25 \times$ bigger than the dataset our selected captioning models were pretrained on, and is bigger than DataComp's medium scale pool. + +# 3. Methodology + +Let $\mathcal{D} = \{(I_i,T_i)\}_{i = 1}^N$ denote an uncurated dataset consisting of $N$ image-text pairs crawled from the web. Our goal is to curate a dataset, $\mathcal{D}' = \{(I_{i'},T_{i'})\}_{i' = 1}^{N'}$ , that is a subset of the pool, $\mathcal{D}'\subseteq \mathcal{D}$ $N^{\prime}\leq N$ , to train a new CLIP model from uninitialized weights, $\Theta_0$ , to new weights, $\Theta^{\prime}$ : + +$$ +\Theta^ {\prime} = \operatorname {t r a i n} \left(\Theta_ {0}, \mathcal {D} ^ {\prime}\right) \tag {1} +$$ + +For a given scoring function, $f$ , that maps an image-text pair to a scalar value, $s = f(I_i, T_i)$ , we express a pruning function that selects a fraction, $k$ , of dataset, $\mathcal{D}$ , using function, $f$ : + +$$ +\Pr n e _ {f} (\mathcal {D}, k) \quad \text {s . t .} 0 \leq k \leq 1, f: (I _ {i}, T _ {i}) \rightarrow \mathbb {R} \tag {2} +$$ + +where $\mathrm{prune}_f(\mathcal{D},k)$ applies function, $f$ , on each image-text sample in $\mathcal{D}$ to obtain a score for each sample, ranks the scores in descending order, and returns a set of the top $k$ portion of the samples. + +One common approach for pruning is CLIPScore [9, 36]. Let $E$ be a CLIP model consisting of an image encoder, $E_{\mathrm{image}}$ , that maps an image, $I$ , to an embedding vector, $E_{\mathrm{image}}(I) \in \mathbb{R}^d$ , and a text encoder, $E_{\mathrm{text}}$ , that maps a text sample, $T$ , to an embedding vector, $E_{\mathrm{text}}(T) \in \mathbb{R}^d$ . CLIP-Score is a measure of alignment between $I$ and $T$ , and is defined as: + +$$ +f _ {\mathrm {C L I P}} (I, T) = \left\langle E _ {\text {i m a g e}} (I), E _ {\text {t e x t}} (T) \right\rangle \tag {3} +$$ + +where $\langle \mathbf{x},\mathbf{y}\rangle$ is the cosine similarity between two vectors, $\mathbf{x}$ and $\mathbf{y}$ , which is defined as the dot product of the $l_{2}$ normalized vectors. The most common CLIP model used for pruning is pretrained on 400 million noisy image-text pairs [36]. Our proposed pruning method, Sieve attempts to minimize the false positives and negatives induced by CLIPScore filtering. Sieve consists of two main components: ImageCaptioning and Sentence Transformer. + +Image-Captioning Let $G$ be a captioning model that generates text, $T_{i}^{G}$ , describing the content of image, $I_{i}$ : + +$$ +T _ {i} ^ {G} = G \left(I _ {i}\right) \tag {4} +$$ + +Given a captioning model pre-trained on a small, representative and well-aligned dataset of image-text pairs, we are interested in estimating the alignment between image-text pairs sampled from a very large, diverse but noisy dataset. The alignment score can then be used as a ranking metric for dataset filtering. We hypothesize that: + +- The probability of generating a caption that is semantically similar to the alt-text from an aligned pair is much higher than that from a misaligned pair. +- The probability of generating a caption that is semantically similar to a hard alt-text is higher than generating a caption that is semantically similar to a misaligned alt-text. Here, a hard alt-text is a text label with low likelihood with respect to the captioning model, but is aligned with the image content. + +As images can contain multiple objects with complex attributes and relationships, there exist multiple ways to describe their content. Given the inherent many-to-many relationship between images and text labels, our goal is to increase the probability of generating a caption that matches an aligned alt-text. To achieve this, we utilize nucleus sampling [12], a decoding strategy used to sample multiple captions, $r$ , per-image: + +$$ +G (I, r) = \left\{T _ {0} ^ {G}, T _ {1} ^ {G}, \dots , T _ {r - 1} ^ {G} \right\} \tag {5} +$$ + +Sentence Transformer Given an image, its alt-text, and a set of generated captions, our goal is to estimate the align- + +![](images/bcb5e268942e6704ce3b5bd2b457eb99bd8b594b3c6be2cbcb09b9befc031b68.jpg) +Figure 2. Our proposed framework enables dataset pruning using image captioning models. To evaluate the alignment of a noisy image-text pair, we generate multiple captions per image using nucleus sampling. Then medium phrases, like "an image of" or "a photo of", are masked from alt-text and generated captions. Finally, a lightweight sentence encoder is used to semantically compare the generated captions with alt-text. + +ment between the generated captions and the alt-text. However, there is a very large diversity gap between the generated captions and the highly diverse alt-text as measured by the number of unique nouns and trigrams [26]. On the other hand, constructing a large, diverse and curated image-text dataset is expensive, which limits the diversity of the generated captions. We propose to bridge this gap by utilizing a light-weight sentence similarity model to encode the alt-text and the generated captions. We expect the semantically similar alt-text and generated caption pairs to be closely clustered in the embedding space compared to semantically distinct pairs. We reason that the rich semantic textual embedding space of the sentence similarity model enables pretraining the captioning model only on a small but curated image-text dataset. Thus, we rely on the semantic understanding of the sentence similarity model to bridge the gap between the limited diversity in the generated captions and the highly diverse alt-text labels. + +To estimate the alignment score, we compute the cosine similarity between embeddings of each generated caption and text label. Let $S$ be a language model that encodes a text sample, $T$ , to a vector, $S(T) \in \mathbb{R}^d$ . We define the alignment between two text samples, $T_a$ and $T_b$ , as the cosine similarity between their language model encodings: + +$$ +\langle S \left(T _ {a}\right), S \left(T _ {b}\right) \rangle \tag {6} +$$ + +This estimate can then be used as a proxy for the image-text alignment of an image, $I$ , and text, $T$ : + +$$ +\langle S (G (I)), S (T) \rangle +$$ + +If the image captioning model generates $r$ different caption candidates for an image, $I$ , we can use the maximum alignment between each of the generated captions, + +$G(I,r) = \{T_0^G,T_1^G,\ldots ,T_{r - 1}^G\}$ , and a text sample, $T$ : + +$$ +\max _ {T _ {j} ^ {G} \in G (I, r)} \langle S \left(T _ {j} ^ {G}\right), S (T) \rangle \tag {7} +$$ + +In literature, there are different models and approaches to obtain text embeddings. [16] uses the average of N-gram features of each word in a text sample to obtain an embedding. A more common option is to use the logits of the last token generated by a decoder-only language model, which is the approach taken with CLIP's text encoder [32], in GPT-1 [31], as well as in [1]. Encoder-only models, such as BERT [6] or RoBERTa [23], can also be used, where the embedding vector may be either the logits of the classification token, or the average pool of the logits of all tokens. Although such language models may have strong generation or classification capabilities, they were not optimized for sentence similarity tasks, but either for next word prediction (i.e., causal language modeling) or masked word prediction (i.e., masked language modeling) tasks. Therefore, their embeddings may not be ideal to measure alignment between sentences. More importantly, such models are large in size and hence slow to infer on large datasets. A language model finetuned on a sentence similarity task, such as SNLI [3], aligns with the goal of estimating semantic textual similarity between alt-text and captions. We find that sentence similarity models [34] pretrained using a self-supervised instance discrimination task on unlabeled corpus of sentences perform well in estimating the alignment between text pairs, and are lightweight in size and latency (e.g., tens of millions of parameters in contrast to billions of parameters of performant decoder-only large language models). + +Masking Medium Words Phrases such as "image of", "picture of", or "photo of" can appear in either alt-text or + +generated captions. We refer to such phrases as "medium phrases", as they describe the medium rather than the contents of an image. We notice that the existence of such medium phrases adds noise to the sentence similarity score, as shown in Figure 3. A pair of sentences that each have a medium phrase are assigned a misleadingly high sentence similarity score by a sentence transformer, as they have been trained on a wide and diverse corpus of text, rather than on image captions. Hence, the existence of medium phrases may increase their attention to the topic of images or media, rather than the topic of the content of such images. Therefore, we neutralize the effect of medium phrases by removing them from both alt-text and generated captions. We express the operation of masking medium words in a text sample, $M(T)$ , on text, $T$ , as masking all possible contiguous subsequences of the text, where masking on a phrase, $t$ , removes it if it is in the pre-determined list of medium phrases, $\mathcal{M} = \{\text{"image of" }, \text{"picture of" }, \text{"photo of" }, \ldots \}$ , or keeps it otherwise. + +Putting it all together, we define the Sieve score function as: + +$$ +f _ {\text {S i e v e}} (I, T) = \max _ {T _ {j} ^ {G} \in G (I, r)} \left\langle S \left(M \left(T _ {j} ^ {G}\right)\right), S (M (T)) \right\rangle \tag {8} +$$ + +The dataset pruned using Sieve with the top $k$ portion of its samples can be expressed as: + +$$ +\mathcal {D} _ {\text {S i e v e}, k} = \operatorname {p r u n e} _ {f _ {\text {S i e v e}}} (\mathcal {D}, k) \tag {9} +$$ + +We summarize our approach in Figure 2 and as psuedocode in Algorithm 1 in the Appendix. + +# 4. Experiments + +# 4.1. Training and Evaluation + +We utilize the DataComp benchmark to evaluate the utility of image-captioning models for multimodal dataset pruning. Two candidate pools are considered, the medium and the large scale, consisting of 128 million and 1.28 billion image-text pairs, respectively. To train CLIP models, we use DataComp's hyperparameters and architectures to standardize training [9]: $5 \times 10^{-4}$ learning rate, 500 iterations warmup, AdamW optimizer, for medium scale: ViT-B/32 image encoder [7], batch size 4096, 128M training samples as a compute budget, and for large scale: ViT-B/16 image encoder, batch size 8192, 1.28B training samples as a compute budget. We evaluate the zero-shot performance on 38 downstream tasks, including classification and retrieval tasks [18, 32, 45]. + +For our captioning model, we utilize BLIP with ViT-B/16 image encoder pretrained on 14 million image-text pairs [20], including highly curated web datasets: Conceptual Captions [37], Conceptual 12M [4] and SBU captions [29]; as well as small human-annotated datasets: + +COCO, and Visual Genome [17]. To compute the alignment between generated captions and alt-text, we use a lightweight distilled sentence transformer, all-MiniLM-L6-v2 [39], further finetuned using self-supervised contrastive learning on unlabeled text corpus. + +# 4.2. Main Results + +Table 1 reports multiple baselines from DataComp [9], including applying no filtering, basic filtering, and CLIP-Score filtering. On the medium scale, Sieve with an image-captioning model pretrained on 30 times less but curated data surpasses CLIPScore by $1.8\%$ on average. In addition, we demonstrate that Sieve's alignment score provides a complimentary signal to CLIPScore by fusing Sieve with CLIPScore. We apply min-max normalization to Sieve alignment scores and CLIPScore independently, then take the per-sample weighted average of both scores: + +$$ +\begin{array}{l} f _ {\text {S i e v e} + \text {C L I P}} (I, T) = \\ (1 - \alpha) \times \bar {f} _ {\text {S i e v e}} (I, T) + \alpha \times \bar {f} _ {\text {C L I P}} (I, T) \\ \end{array} +$$ + +$$ +\begin{array}{l} \begin{array}{l} \text {s . t .} \quad \overline {{f}} (I, T) = \end{array} \\ \frac {f (I , T) - \min _ {(I _ {i} , T _ {i}) \in \mathcal {D}} f (I _ {i} , T _ {i})}{\max _ {(I _ {i} , T _ {i}) \in \mathcal {D}} f (I _ {i} , T _ {i}) - \min _ {(I _ {i} , T _ {i}) \in \mathcal {D}} f (I _ {i} , T _ {i})} \\ \end{array} +$$ + +where the weight $\alpha$ used in the reported results is 0.5. Finally, we select the top $20\%$ of samples. We observe that the fused approach improves CLIPScore average performance by $2.6\%$ and $1.7\%$ on medium and large scale respectively, demonstrating the complementarity of Sieve's alignment score. Moreover, Sieve without fusion achieves the best performance on retrieval tasks on both medium and large scale experiments. Relative to CLIPScore, a significant portion of Sieve's retrieval gain is effectively transferred when fusing Sieve with CLIPScore (i.e., $+2.7\%$ and $+4.5\%$ on medium and large scale). This is also promising as retrieval performance is critical for retrieval augmented VLMs [13, 42]. + +# 4.3. Per-Task Performance + +Figure 5 shows the change in accuracy introduced by Sieve as well as Sieve+CLIPScore on each task compared to CLIPScore on medium scale, and Figure 6 of the Appendix shows for the large scale. We observe that in addition to outperforming on image retrieval tasks, Flickr [43], and MS COCO [22], Sieve's greatest performance boost comes from WingoGAViL [2], a retrieval task which requires diverse reasoning skills, including general knowledge, common sense, and abstraction. This high performance can be attributed to Sieve's preference towards keeping samples where the alt-text correctly describes visual concepts and + +![](images/eaf9ab012b2e130c30e3b372567c1addb373a1d1e6e5326d2612736c654e1174.jpg) +Figure 3. Masking medium phrases improves the ranking of sentence similarity scores. On the left, sentence pairs with misleadingly high (or low) sentence similarity due to the existence (or absence) of medium phrases are highlighted in dark red (or light red). On the right, similarity scores that are more aligned with semantics are highlighted in dark green. The sentence similarity scores are computed using the all-MiniLM-L6-v2 sentence transformer [39]. + +
ScaleFilteringDataset SizeImageNetImageNet dist. shiftsVTABRetrievalAverage over 38 datasets
Medium (128 Million)No Filtering128M17.615.225.921.925.8
Basic Filtering30M22.619.328.425.128.5
LAION Filtering13M23.019.830.723.329.2
CLIPScore38M27.323.033.825.132.8
Sieve24M29.425.035.228.934.6
Sieve+CLIPScore24M30.325.436.227.835.4
Large (1.28 Billion)No Filtering1.28B45.937.842.641.943.7
Basic Filtering298M51.642.344.648.045.8
LAION Filtering130M55.345.351.049.550.1
CLIPScore384M57.847.453.846.652.9
Sieve235M57.347.852.052.052.3
Sieve+CLIPScore235M59.749.154.851.154.6
+ +Table 1. Zero-shot performance of CLIP models pretrained using various filtering strategies on medium and large scale pools of the DataComp benchmark. Sieve fused with CLIPScore beats CLIPScore by $2.6\%$ and $1.7\%$ on medium and large scale respectively. In addition, on retrieval tasks, Sieve achieves best performance on both scales and after fusing with CLIPScore, leads an improvement on retrieval over CLIPScore of $2.7\%$ and $4.5\%$ on medium and large scale respectively. + +their attributes and relations. Sieve, especially when combined with CLIPScore, significantly outperforms on medical diagnosis tasks, Camelyon17 and PatchCamelyon. + +Sieve mainly underperforms in tasks requiring parsing text from images, such as MNIST [19], SVHN [25], and Rendered SST-2 [28], concluding that Sieve is less likely to select image-text pairs that are useful for OCR tasks. In addition, Sieve underperforms CLIPScore on context-based tasks like Country211 [27], a task assessing the geolocation capability of visual representations, demonstrating Sieve's preference towards selecting samples based on the alignment of alt-text with visual concepts rather than context. Interestingly, when fusing with CLIPScore, we improve the performance of all these tasks while retaining the advantage + +of Sieve especially at large scale (see Figure 6). + +# 4.4. Ablation Studies + +We conduct studies on the medium scale pool and report the average of three runs per experiment. + +Pretraining data-distribution We study the effect of the pretraining data distribution used to train the captioning model on the quality of the alignment score. This is measured based on the downstream performance of the CLIP model trained using the selected image-text pairs. Two pretraining data distributions proposed by BLIP [20] are investigated. The first uses 14 million curated image-text pairs, while the second uses an additional 115 million web images with noisy alt-text [36]. Although the original BLIP + +![](images/4e807ca87d0e3eb1912fd0a8bb0a0737556557f42142380d5cf415a637b80b51.jpg) +Figure 4. Evaluating CLIP models pretrained on different fractions of the top ranked samples based on our proposed approach (Sieve), CLIPScore, and fusing Sieve with CLIPscore (Sieve+CLIPSCore), on medium scale. + +![](images/50f0e847567f9ef0efad18ef1657a8f14727f96870d55466c32991801dd49059.jpg) + +
Percentage KeptCaption Generator Pretraining DataImageNetImageNet dist. shiftVTABRetrievalAverage over 38 datasets
10BLIP-129M23.0020.6030.2021.4030.00
BLIP-14M26.5022.5032.1023.7530.90
15BLIP-129M25.9522.8032.8024.4032.40
BLIP-14M29.0524.6033.3526.9532.70
20BLIP-129M27.8523.6533.4526.3533.05
BLIP-14M29.6024.9335.0728.5734.03
+ +work reports higher captioning performance when pretraining on 115 million samples, our results in Table 2 indicate that for the purpose of dataset pruning, using curated image-text pairs results in a better alignment score than using a much larger noisy dataset. This highlights the importance of using a captioning model pretrained on higher quality data for large-scale dataset pruning. + +Captioning models We conduct experiments to investigate the generalizability of Sieve to other captioning models. In Table 3, we compare GIT [38], which utilizes an image encoder, and a text decoder captioning model, pretrained on 10 million image-text pairs to BLIP [20], which utilizes an image encoder, text encoder and a text decoder and is pretrained on 14 million image-text pairs. We observe that fusing Sieve with CLIPScore based on GIT and BLIP improves retrieval performance by $3.13\%$ and $3.83\%$ and improves average performance on 38 downstream tasks by $1.07\%$ and $2.52\%$ respectively. + +Text embedding space In Table 4, we ablate over embeddings from different text models and show that embeddings from our selected sentence transformer perform better than embeddings from CLIP and BLIP text encoders. The CLIP text encoder was pretrained along with the CLIP vision encoder to map text and images to the same embedding space, and is used in diffusion models to condition image genera + +Table 2. Effect of a caption generator's pretraining data-distribution on Sieve. The 14M pretraining dataset consists of curated image-text pairs, while the 129M dataset includes an additional 115M noisy image-text pairs from LAION [36]. + +
FilteringImageNetVTABRetrievalAvg
CLIPScore27.1333.9024.1732.63
Sieve-GIT27.4733.2027.2732.40
+CLIPScore28.9034.8327.3033.70
Sieve-BLIP29.6035.0728.5734.03
+CLIPScore30.3535.9028.0035.15
+ +Table 3. Effect of captioning models, GIT [38] pretrained on 10M image-text pairs, and BLIP [20] pretrained on 14M pairs + +tion [33]. However, we observe that the CLIP text encoder suffers from poor semantic textual understanding, leading to a large drop in accuracy when used as a caption similarity measure. BLIP's text encoder performs better than that of CLIP, but the lightweight sentence transformer specifically pretrained on aligning semantically similar texts performs significantly better with $\geq 2\%$ improvements across various task types. In Figure 7 in the Appendix we show how cosine similarities of sentence similarity models result in better semantic textual clustering compared to CLIP and BLIP text encoders. + +Pruning percentage We study the effect of the fraction of samples selected for pretraining. For each experiment, we compute the Sieve alignment score and CLIPScore for each + +
Text EncoderImageNetImageNet dist shiftVTABRetrievalAverage over 38 datasets
CLIP18.0015.6526.9020.9525.90
BLIP27.2022.1032.7025.4531.85
Sentence Transformer29.6024.9335.0728.5734.03
+ +![](images/7adffcfebeef1609f6edb9e435b1a71c90fdbee32c6ff793493257c422c9b779.jpg) +(a) Sieve gain over CLIPScore on medium scale pool. + +![](images/7399286e9c74e23defc786d289f598219c4806b89eb130de48644823e04d2078.jpg) +(b) Sieve+CLIPScore gain over CLIPScore on medium scale pool. +Figure 5. The relative performance gain of Sieve and Sieve+CLIPScore relative to CLIPScore on 38 downstream tasks on the medium scale pool. + +sample. The top- $k\%$ and pretraining CLIP models are then selected. Here, $k\%$ is set to $10\%$ , $15\%$ , $20\%$ , $25\%$ and $30\%$ . Finally, we report the zero-shot performance on ImageNet and the average on 38 tasks in Figure 4. We observe that Sieve achieves the best performance using $20\%$ of the data, while CLIPScore peaks at $30\%$ (similar to results reported in 9). Hence, pruning using Sieve achieves better performance with less data, compared to CLIPScore. + +Number of generated captions and fusion with CLIP-Score We study the effect of using multiple captions per image to maximize the alignment of the generated captions with the alt-text. For nucleus sampling [12], we set the + +Table 4. Effect of the sentence encoder on the performance of Sieve. CLIP uses the text encoder pretrained on 400M samples [36], BLIP uses the text encoder pretrained on the curated 14M samples defined in [20], and Sentence Transformer uses a language model pretrained on unlabeled text corpus [39]. Each encoder encodes generated captions and the alt-text where the textual semantic alignment is computed. + +
Generated CaptionsCLIP weightImageNetAverage over 38 datasets
10.028.6032.50
20.029.0033.40
40.029.5333.70
80.029.6034.03
0.330.1034.40
80.530.3535.15
0.730.2534.35
+ +Table 5. Effect of number of generated captions and weight of CLIPScore on zero-shot performance of pretrained CLIP models + +cumulative probability of the smallest set of words to 0.9, and the minimum and maximum sequence lengths to 5 and 20, respectively. We study the effect of sampling 1, 2, 4, and 8 captions. For each input image-text pair, we assign the maximum alignment score between the alt-text and the generated captions. We observe in Table 5 that increasing the number of generated captions improves the performance on downstream tasks. We reason that due to the many-to-many relationship between images and captions, generating more captions increases the probability of matching a hard aligned alt-text. We also investigate the effect of fusing the Sieve alignment score with CLIPScore in Table 5. Each score is independently normalized, and a weighted average is applied between the two scores. Finally, the top $20\%$ of samples ranked by Sieve+CLIPScore are selected. We observe that a weight of 0.5 achieves the best performance. + +# 5. Conclusion + +We introduce a novel method, Sieve, that enables pruning large-scale noisy web-crawled image-text datasets. We propose utilizing synthetic captions from image captioning models pretrained on small, diverse, and curated datasets to evaluate the alignment of noisy image-text pairs. Using the embedding space of a lightweight sentence transformer, we compute an alignment score between generated captions and alt-text. We demonstrate that Sieve provides a complementary pruning signal to CLIPScore, effectively minimizing false positives and negatives, leading to improved zero-shot classification and retrieval performance. + +# References + +[1] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S. Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication, 2023. 1, 4 +[2] Yonatan Bitton, Nitzan Bitton Guetta, Ron Yosef, Yuval Elovici, Mohit Bansal, Gabriel Stanovsky, and Roy Schwartz. WinoGAVil: Gamified association benchmark to challenge vision-and-language models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5 +[3] Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326, 2015. 4 +[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 5 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 2 +[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5 +[8] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 3 +[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortzman, Dhruba Ghosh, Jieyu Zhang, et al. Dat acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1, 2, 3, 5, 8, 12 +[10] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1 +[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1, 2 +[12] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In ICLR. OpenReview.net, 2020. 2, 3, 8 +[13] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and + +Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23369-23379, 2023. 1, 5 +[14] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 2 +[15] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 2 +[16] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics, 2017. 4 +[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5 +[18] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022. 5 +[19] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 11 +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5, 6, 7, 8 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5 +[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 4 +[24] Pratyush Maini, Sachin Goyal, Zachary C Lipton, J Zico Kolter, and Aditi Raghunathan. T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132, 2023. 3, 11 +[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6, 11 + +[26] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. arXiv preprint arXiv:2307.10350, 2023. 1, 3, 4 +[27] OpenAI. The country211 dataset, .6 +[28] OpenAI. The rendered sst2 dataset, .6 +[29] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5 +[30] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6967-6977, 2023. 2, 3, 11 +[31] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 4 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 7 +[34] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 4 +[35] William A Gaviria Rojas, Sudnya Diamos, Keertan Ranjan Kini, David Kanter, Vijay Janapa Reddi, and Cody Coleman. The dollar street dataset: Images representing the geographic and socioeconomic diversity of the world. In Thirtysixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 11 +[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 3, 6, 7, 8 +[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5 +[38] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 7 +[39] Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, 2020. 5, 6, 8 + +[40] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. 2023. 1, 2 +[41] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In Proceedings of the 40th International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 1 +[42] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. 2023. 1, 5 +[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5 +[44] Haichao Yu, Yu Tian, Sateesh Kumar, Linjie Yang, and Heng Wang. The devil is in the details: A deep dive into the rabbit hole of data filtering. arXiv preprint arXiv:2309.15954, 2023. 3 +[45] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5 \ No newline at end of file diff --git a/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/images.zip b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f0eb2aeb557e4d92053564a6d463c508202d9d47 --- /dev/null +++ b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3eebca739cf98e2073dc7e9f4b433a2129ee5767cc7cd29c02350ac6c15e3b93 +size 511468 diff --git a/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/layout.json b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ca443ffad578797990626fac12f89695cb8cd044 --- /dev/null +++ b/2024/Sieve_ Multimodal Dataset Pruning using Image Captioning Models/layout.json @@ -0,0 +1,8898 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "text", + "content": "Sieve: Multimodal Dataset Pruning Using Image Captioning Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 142, + 461, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 142, + 461, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 461, + 158 + ], + "type": "text", + "content": "Anas Mahmoud1,2* Mostafa Elhoushi1 Amro Abbas† Yu Yang3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 158, + 425, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 158, + 425, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 158, + 425, + 171 + ], + "type": "text", + "content": "Newsha Ardalani1 Hugh Leather1 Ari S. Morcos4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "spans": [ + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "text", + "content": "FAIR at Meta " + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "text", + "content": "University of Toronto " + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "text", + "content": "UC Los Angeles " + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 171, + 483, + 186 + ], + "type": "text", + "content": "DatologyAI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning. We argue that this approach suffers from multiple limitations including: false positives and negatives due to CLIP's pretraining on noisy labels. We propose a pruning signal, Sieve, that employs synthetic captions generated by image captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on unlabeled text corpus. Using DataComp, a multimodal dataset filtering benchmark, when evaluating on 38 downstream tasks, our pruning approach, surpasses CLIPScore by " + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": " on medium and large scale respectively. In addition, on retrieval tasks, Sieve leads to a significant improvement of " + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": " on medium and large scale respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 547, + 128, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 128, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 128, + 559 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) [32] models have shown great success in solving zero-shot image classification and multimodal retrieval tasks. In addition, many foundational Vision-Language Models (VLMs) use pretrained CLIP encoders to condition image generation on CLIP text embeddings [33] in retrieval augmented vision-language models [13, 42], and to align modalities including audio, depth, and thermal with language through CLIP image embeddings [10]. Therefore, the quality of CLIP representations can influence the performance of many VLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 213, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 545, + 310 + ], + "type": "text", + "content": "To pretrain CLIP, billions of image-text pairs are collected using common crawl. The raw data is highly diverse but contains many noisy image-text pairs, including low quality images, low quality alternative text (alt-text), and misaligned image-text pairs. Pretraining CLIP models on noisy data can have adverse effects on the learned representations, thus leading to poor performance on downstream tasks [1]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 312, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 546, + 445 + ], + "type": "text", + "content": "To address this challenge, researchers have developed data pruning methods to remove low quality image-text pairs. Heuristics that filter out image-text pairs based on image dimensions, aspect ratio, alt-text length, and complexity are commonly used [9, 36] to reduce noise, but can also limit the diversity of the dataset [26]. Methods that use images or class names from datasets, like ImageNet, to sample semantically similar image-text pairs can lead to higher accuracy on downstream tasks [40], but limit the diversity of the selected samples as they sample image-text pairs close to a specific dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 447, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 615 + ], + "type": "text", + "content": "One of the most effective pruning methods, CLIP-Score [11, 36], computes the cosine similarity between image and text embeddings using a pretrained CLIP model. This score is then used to rank the alignment of image-text pairs. However, as shown in Figure 1, using CLIPScore can lead to false positives – samples that are poorly aligned but have high CLIPScore (i.e., bad samples) due to spurious correlations learned by the pretrained CLIP model [41]. In addition, using CLIPScore can lead to false negatives – samples that are aligned but have low CLIPScore (i.e., hard samples) due to the poor discrimination between hard and bad samples. Excluding hard samples and including bad samples can negatively affect the generalization of CLIP image and text encoders." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "The goal of this work is to reduce both false negatives and positives induced by CLIPScore ranking by relying on an image-captioning model pretrained on small, diverse, and well-aligned image-text pairs. As shown in Figure 1, Sieve can reduce false positives or false negatives in different scenarios, such as samples where CLIPScore focuses on geographic or chronological context of an image rather than the content of an image. As depicted in Figure 2, we evalu" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 184, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 184, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 184, + 703 + ], + "type": "text", + "content": "*Work done during internship at Meta" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 201, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 201, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 201, + 712 + ], + "type": "text", + "content": "†Work done during an AI residency at Meta" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22423" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 76, + 185, + 153 + ], + "blocks": [ + { + "bbox": [ + 52, + 76, + 185, + 153 + ], + "lines": [ + { + "bbox": [ + 52, + 76, + 185, + 153 + ], + "spans": [ + { + "bbox": [ + 52, + 76, + 185, + 153 + ], + "type": "image", + "image_path": "a48f7a47ad075a945510158eab7694247ca5a2e6914b0d4a11dc4c331e56fef3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 76, + 219, + 83 + ], + "lines": [ + { + "bbox": [ + 197, + 76, + 219, + 83 + ], + "spans": [ + { + "bbox": [ + 197, + 76, + 219, + 83 + ], + "type": "text", + "content": "ALT-Text:" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 83, + 282, + 88 + ], + "lines": [ + { + "bbox": [ + 197, + 83, + 282, + 88 + ], + "spans": [ + { + "bbox": [ + 197, + 83, + 282, + 88 + ], + "type": "text", + "content": "The Best Places to Stay in Kaikoura, New" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 89, + 215, + 94 + ], + "lines": [ + { + "bbox": [ + 197, + 89, + 215, + 94 + ], + "spans": [ + { + "bbox": [ + 197, + 89, + 215, + 94 + ], + "type": "text", + "content": "Zealand" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 98, + 233, + 103 + ], + "lines": [ + { + "bbox": [ + 197, + 98, + 233, + 103 + ], + "spans": [ + { + "bbox": [ + 197, + 98, + 233, + 103 + ], + "type": "text", + "content": "CLIPScore:0.30" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 104, + 228, + 110 + ], + "lines": [ + { + "bbox": [ + 197, + 104, + 228, + 110 + ], + "spans": [ + { + "bbox": [ + 197, + 104, + 228, + 110 + ], + "type": "text", + "content": "False Positive" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 119, + 239, + 125 + ], + "lines": [ + { + "bbox": [ + 197, + 119, + 239, + 125 + ], + "spans": [ + { + "bbox": [ + 197, + 119, + 239, + 125 + ], + "type": "text", + "content": "Generated Caption:" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 125, + 290, + 130 + ], + "lines": [ + { + "bbox": [ + 197, + 125, + 290, + 130 + ], + "spans": [ + { + "bbox": [ + 197, + 125, + 290, + 130 + ], + "type": "text", + "content": "a living room with a large window and a couch" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 140, + 222, + 146 + ], + "lines": [ + { + "bbox": [ + 197, + 140, + 222, + 146 + ], + "spans": [ + { + "bbox": [ + 197, + 140, + 222, + 146 + ], + "type": "text", + "content": "Sieve: 0.12" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 147, + 228, + 152 + ], + "lines": [ + { + "bbox": [ + 197, + 147, + 228, + 152 + ], + "spans": [ + { + "bbox": [ + 197, + 147, + 228, + 152 + ], + "type": "text", + "content": "True Negative" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 157, + 182, + 243 + ], + "blocks": [ + { + "bbox": [ + 54, + 157, + 182, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 157, + 182, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 157, + 182, + 243 + ], + "type": "image", + "image_path": "8a7d643f141acfb48b68768d4a7d88168c82c715f12adccc613339d9c86140a2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 163, + 219, + 168 + ], + "lines": [ + { + "bbox": [ + 197, + 163, + 219, + 168 + ], + "spans": [ + { + "bbox": [ + 197, + 163, + 219, + 168 + ], + "type": "text", + "content": "ALT-Text:" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 170, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 197, + 170, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 197, + 170, + 287, + 175 + ], + "type": "text", + "content": "The area were evacuated and Keiserville Rd." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 176, + 276, + 181 + ], + "lines": [ + { + "bbox": [ + 197, + 176, + 276, + 181 + ], + "spans": [ + { + "bbox": [ + 197, + 176, + 276, + 181 + ], + "type": "text", + "content": "in Washington Twp. Pa., was closed as" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 182, + 272, + 186 + ], + "lines": [ + { + "bbox": [ + 197, + 182, + 272, + 186 + ], + "spans": [ + { + "bbox": [ + 197, + 182, + 272, + 186 + ], + "type": "text", + "content": "crews worked on the Marcellus street" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 190, + 233, + 196 + ], + "lines": [ + { + "bbox": [ + 197, + 190, + 233, + 196 + ], + "spans": [ + { + "bbox": [ + 197, + 190, + 233, + 196 + ], + "type": "text", + "content": "CLIPScore: 0.31" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 197, + 228, + 201 + ], + "lines": [ + { + "bbox": [ + 197, + 197, + 228, + 201 + ], + "spans": [ + { + "bbox": [ + 197, + 197, + 228, + 201 + ], + "type": "text", + "content": "False Positive" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 209, + 240, + 215 + ], + "lines": [ + { + "bbox": [ + 197, + 209, + 240, + 215 + ], + "spans": [ + { + "bbox": [ + 197, + 209, + 240, + 215 + ], + "type": "text", + "content": "Generated Caption:" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 216, + 275, + 220 + ], + "lines": [ + { + "bbox": [ + 197, + 216, + 275, + 220 + ], + "spans": [ + { + "bbox": [ + 197, + 216, + 275, + 220 + ], + "type": "text", + "content": "a view of a construction site with water" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 220, + 263, + 225 + ], + "lines": [ + { + "bbox": [ + 197, + 220, + 263, + 225 + ], + "spans": [ + { + "bbox": [ + 197, + 220, + 263, + 225 + ], + "type": "text", + "content": "coming out of it" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 231, + 221, + 236 + ], + "lines": [ + { + "bbox": [ + 197, + 231, + 221, + 236 + ], + "spans": [ + { + "bbox": [ + 197, + 231, + 221, + 236 + ], + "type": "text", + "content": "Sieve: 0.21" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 236, + 228, + 242 + ], + "lines": [ + { + "bbox": [ + 197, + 236, + 228, + 242 + ], + "spans": [ + { + "bbox": [ + 197, + 236, + 228, + 242 + ], + "type": "text", + "content": "True Negative" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 256, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 545, + 334 + ], + "type": "text", + "content": "Figure 1. Examples of image-text pairs in which the scores of pretrained CLIP models, a commonly used image-text data filtering approach, fail to measure their alignment. Our proposed approach, Sieve, provides an accurate alignment score using a caption generator and sentence transformer. Top left and bottom left: Examples of false positives where alt-text describes concepts that are not found or unrelated to the image. CLIP is trained on similar noisy image-text pairs, thus, it assigns a relatively high score. Sieve can detect that such image-text pairs are misaligned. Top right and bottom right: Examples of false negatives where images are aligned with the alt-text but are assigned low CLIP scores, either due to the low likelihood of these pairs in the pretraining data, or because CLIP may have seen similar images aligned with other noisy labels. Sieve can detect that such image-text pairs are well-aligned and selects them for pretraining." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 410, + 75, + 541, + 163 + ], + "blocks": [ + { + "bbox": [ + 376, + 78, + 398, + 83 + ], + "lines": [ + { + "bbox": [ + 376, + 78, + 398, + 83 + ], + "spans": [ + { + "bbox": [ + 376, + 78, + 398, + 83 + ], + "type": "text", + "content": "ALT-Text:" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 347, + 83, + 398, + 89 + ], + "lines": [ + { + "bbox": [ + 347, + 83, + 398, + 89 + ], + "spans": [ + { + "bbox": [ + 347, + 83, + 398, + 89 + ], + "type": "text", + "content": "The Great Sphinx of Giza" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 363, + 100, + 398, + 106 + ], + "lines": [ + { + "bbox": [ + 363, + 100, + 398, + 106 + ], + "spans": [ + { + "bbox": [ + 363, + 100, + 398, + 106 + ], + "type": "text", + "content": "CLIPScore:0.25" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 366, + 106, + 397, + 111 + ], + "lines": [ + { + "bbox": [ + 366, + 106, + 397, + 111 + ], + "spans": [ + { + "bbox": [ + 366, + 106, + 397, + 111 + ], + "type": "text", + "content": "False Negative" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 356, + 122, + 398, + 127 + ], + "lines": [ + { + "bbox": [ + 356, + 122, + 398, + 127 + ], + "spans": [ + { + "bbox": [ + 356, + 122, + 398, + 127 + ], + "type": "text", + "content": "Generated Caption:" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 315, + 128, + 398, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 128, + 398, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 128, + 398, + 133 + ], + "type": "text", + "content": "the great sphinx and the pyramids of giza" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 373, + 138, + 398, + 143 + ], + "lines": [ + { + "bbox": [ + 373, + 138, + 398, + 143 + ], + "spans": [ + { + "bbox": [ + 373, + 138, + 398, + 143 + ], + "type": "text", + "content": "Sieve: 0.91" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 369, + 144, + 398, + 148 + ], + "lines": [ + { + "bbox": [ + 369, + 144, + 398, + 148 + ], + "spans": [ + { + "bbox": [ + 369, + 144, + 398, + 148 + ], + "type": "text", + "content": "True Positive" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 410, + 75, + 541, + 163 + ], + "lines": [ + { + "bbox": [ + 410, + 75, + 541, + 163 + ], + "spans": [ + { + "bbox": [ + 410, + 75, + 541, + 163 + ], + "type": "image", + "image_path": "712cfe9d6b8953759f4d7f9329818b324e3fda5c89e6c4445affd74b42657645.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 408, + 167, + 541, + 243 + ], + "blocks": [ + { + "bbox": [ + 376, + 167, + 398, + 171 + ], + "lines": [ + { + "bbox": [ + 376, + 167, + 398, + 171 + ], + "spans": [ + { + "bbox": [ + 376, + 167, + 398, + 171 + ], + "type": "text", + "content": "ALT-Text:" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 347, + 172, + 398, + 177 + ], + "lines": [ + { + "bbox": [ + 347, + 172, + 398, + 177 + ], + "spans": [ + { + "bbox": [ + 347, + 172, + 398, + 177 + ], + "type": "text", + "content": "A view of the interior pool" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 363, + 194, + 398, + 198 + ], + "lines": [ + { + "bbox": [ + 363, + 194, + 398, + 198 + ], + "spans": [ + { + "bbox": [ + 363, + 194, + 398, + 198 + ], + "type": "text", + "content": "CLIPScore: 0.23" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 365, + 199, + 397, + 204 + ], + "lines": [ + { + "bbox": [ + 365, + 199, + 397, + 204 + ], + "spans": [ + { + "bbox": [ + 365, + 199, + 397, + 204 + ], + "type": "text", + "content": "False Negative" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 356, + 213, + 398, + 217 + ], + "lines": [ + { + "bbox": [ + 356, + 213, + 398, + 217 + ], + "spans": [ + { + "bbox": [ + 356, + 213, + 398, + 217 + ], + "type": "text", + "content": "Generated Caption:" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 310, + 218, + 398, + 224 + ], + "lines": [ + { + "bbox": [ + 310, + 218, + 398, + 224 + ], + "spans": [ + { + "bbox": [ + 310, + 218, + 398, + 224 + ], + "type": "text", + "content": "a large indoor swimming pool with a skylight" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 373, + 228, + 398, + 233 + ], + "lines": [ + { + "bbox": [ + 373, + 228, + 398, + 233 + ], + "spans": [ + { + "bbox": [ + 373, + 228, + 398, + 233 + ], + "type": "text", + "content": "Sieve: 0.65" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 369, + 234, + 397, + 239 + ], + "lines": [ + { + "bbox": [ + 369, + 234, + 397, + 239 + ], + "spans": [ + { + "bbox": [ + 369, + 234, + 397, + 239 + ], + "type": "text", + "content": "True Positive" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 408, + 167, + 541, + 243 + ], + "lines": [ + { + "bbox": [ + 408, + 167, + 541, + 243 + ], + "spans": [ + { + "bbox": [ + 408, + 167, + 541, + 243 + ], + "type": "image", + "image_path": "0e55db0e340610e54f5dfcd653d294a5079a7e17f780f6bdf510d955a688a2e0.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "text", + "content": "ate the alignment of web-crawled image-text pairs by first, generating multiple captions for each image using nucleus sampling [12], followed by removing phrases that describe the medium (e.g., \"an image of\", \"a photo of\") rather than visual concepts. Finally, to evaluate semantic similarity between the limited diversity of generated captions and the high diversity of alt-text, we utilize the embedding space of a lightweight sentence transformer pretrained on unlabeled text corpus. The alignment between the generated captions and the alt-text is then used as a proxy for image-text alignment. Our goal is to fuse this alignment signal with CLIP-Score to minimize false positives and negatives, leading to a more aligned pretraining dataset. To evaluate the effectiveness of our proposed pruning method, we utilize the DataComp [9] benchmark, which fixes the pretraining hyperparameters of CLIP and provides multiple candidate pools of noisy image-text data for pretraining CLIP models. The goal is to select a subset of noisy image-text data that leads to the best performance on 38 downstream tasks. Using image-captioning model alignment scores fused with CLIP-Score, we surpass CLIPScore filtering by " + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "text", + "content": " on the average of 38 downstream tasks on medium and large scale pool respectively. In addition, on multimodal retrieval tasks, our approach leads to an improvement of " + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 355, + 289, + 654 + ], + "type": "text", + "content": " on medium and large scale respectively." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 47, + 668, + 134, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 134, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 134, + 680 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Heuristics are basic filtering methods including: filtering non-English alt-text using fastText [15], filtering alt-text" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 354, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 451 + ], + "type": "text", + "content": "with a few words [9, 36], filtering alt-text with low text complexity [30], and filtering images by size or aspect ratio [9]. A combination of these unimodal filtering approaches has been explored by DataComp [9]. An example of a multimodal filtering approach is text spotting: detecting and recognizing text in images and filtering image-text pairs with high overlap between spotted text (text detected in image) and alt-text (associated label of image) [30]." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 456, + 546, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 546, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 546, + 673 + ], + "type": "text", + "content": "Datasets as Priors was proposed in DataComp [9], relying on sampling image-text pairs that are semantically similar to diverse and curated datasets like ImageNet [5]. Text-based sampling selects image-text pairs with alt-text overlapping one of the ImageNet classes. CiT [40] uses cosine similarity to filter alt-text that are similar to ImageNet classes. Image-based sampling approaches encode images from the unfiltered candidate pool using the OpenAI CLIP's ViT-L/14 vision encoder, and clusters the images into 100,000 groups using FAISS [14]. Then, embeddings of ImageNet training samples are used to keep the closest cluster to each sample. The main limitation of such approaches is that they bias the CLIP model and may not generalize well to new downstream tasks. We argue that selecting samples that match the distribution of downstream tasks encourages overfitting to the evaluation set and, thus limits generalization to other downstream tasks. Our approach, Sieve, does not use any dataset as a prior." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "Pretrained VLMs One of the most successful methods for evaluating image-text alignment is CLIPScore [11]. LAION filtering [36] uses an OpenAI CLIP model [32] pre" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22424" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "trained on 400 million image-text pairs to evaluate image-text alignment of large webscale datasets, and filter out samples with the lowest CLIPScore. Filtering using CLIP-Score can suffer from false negatives, which leads to filtering out hard informative samples, and false positives, which leads to including misaligned samples. Another approach proposes a non-filtering approach that utilizes pretrained VLMs [26], using large image-captioning models like BLIP2 [21] to replace alt-text labels with descriptive synthetic captions. The synthetic captions are then used to train CLIP models. The authors [26] demonstrate that at scale, the improvement of synthetic captions is capped by the limited diversity of generated captions compared to the high diversity of noisy text labels. Compared to [26], we do not alter the original alt-text and thus our focus is on the dataset pruning challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "content": "Concurrent Work Inspired by text spotting [30], T-MARS [24] is concurrent work that detects and masks text regions in images before computing CLIPScore, resulting in improved visual representations. We reason that T-MARS is orthogonal to our approach as we can apply the same masking before calculating Sieve. In addition, Devil is In the Details (DID) [44], utilizes a combination of multiple filters and sampling approaches, which is orthogonal to our approach as Sieve can be combined to other filters. Moreover, most of DID's accuracy improvements is from aligning the selected data distribution with the DataComp [9] evaluation set (downstream tasks), which overfits downstream tasks as discussed in previous sections. More importantly, T-MARS [24] and DID [9] only perform pruning experiments on the medium scale of DataComp [9], while Sieve shows results on both medium and large scales. Finally, DFN [8], uses CLIPScore from a CLIP model pretrained on a private 357 million human verified image-text labels, which is " + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "inline_equation", + "content": "25 \\times" + }, + { + "bbox": [ + 46, + 264, + 289, + 515 + ], + "type": "text", + "content": " bigger than the dataset our selected captioning models were pretrained on, and is bigger than DataComp's medium scale pool." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 528, + 130, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 130, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 130, + 541 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(I_i,T_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "content": " denote an uncurated dataset consisting of " + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "content": " image-text pairs crawled from the web. Our goal is to curate a dataset, " + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{D}' = \\{(I_{i'},T_{i'})\\}_{i' = 1}^{N'}" + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "content": ", that is a subset of the pool, " + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'\\subseteq \\mathcal{D}" + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "inline_equation", + "content": "N^{\\prime}\\leq N" + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "content": ", to train a new CLIP model from uninitialized weights, " + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "content": ", to new weights, " + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\Theta^{\\prime}" + }, + { + "bbox": [ + 46, + 548, + 287, + 608 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 618, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 618, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 126, + 618, + 287, + 632 + ], + "type": "interline_equation", + "content": "\\Theta^ {\\prime} = \\operatorname {t r a i n} \\left(\\Theta_ {0}, \\mathcal {D} ^ {\\prime}\\right) \\tag {1}", + "image_path": "dfa2aea65735c4292e8222d4625c7acfcf7ebbc0b1aed7b4ba6683d323b4b822.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "text", + "content": "For a given scoring function, " + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "text", + "content": ", that maps an image-text pair to a scalar value, " + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "inline_equation", + "content": "s = f(I_i, T_i)" + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "text", + "content": ", we express a pruning function that selects a fraction, " + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "text", + "content": ", of dataset, " + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "text", + "content": ", using function, " + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 643, + 287, + 691 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\Pr n e _ {f} (\\mathcal {D}, k) \\quad \\text {s . t .} 0 \\leq k \\leq 1, f: (I _ {i}, T _ {i}) \\rightarrow \\mathbb {R} \\tag {2}", + "image_path": "6e78f89932429b37ff8a8535a46f5323090b0bacd4920b2a8dd8a939b5a58fc1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathrm{prune}_f(\\mathcal{D},k)" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " applies function, " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", on each image-text sample in " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " to obtain a score for each sample, ranks the scores in descending order, and returns a set of the top " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " portion of the samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": "One common approach for pruning is CLIPScore [9, 36]. Let " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": " be a CLIP model consisting of an image encoder, " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{image}}" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": ", that maps an image, " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": ", to an embedding vector, " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{image}}(I) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": ", and a text encoder, " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{text}}" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": ", that maps a text sample, " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": ", to an embedding vector, " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{text}}(T) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": ". CLIP-Score is a measure of alignment between " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 121, + 547, + 203 + ], + "type": "text", + "content": ", and is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 213, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 213, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 353, + 213, + 545, + 228 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {C L I P}} (I, T) = \\left\\langle E _ {\\text {i m a g e}} (I), E _ {\\text {t e x t}} (T) \\right\\rangle \\tag {3}", + "image_path": "922cd347aa9ef5c685bfe5078231b345d197b7bd3586d98bff1b423e51e1aa6a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{x},\\mathbf{y}\\rangle" + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "text", + "content": " is the cosine similarity between two vectors, " + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "text", + "content": ", which is defined as the dot product of the " + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 234, + 545, + 330 + ], + "type": "text", + "content": " normalized vectors. The most common CLIP model used for pruning is pretrained on 400 million noisy image-text pairs [36]. Our proposed pruning method, Sieve attempts to minimize the false positives and negatives induced by CLIPScore filtering. Sieve consists of two main components: ImageCaptioning and Sentence Transformer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "text", + "content": "Image-Captioning Let " + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "text", + "content": " be a captioning model that generates text, " + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "inline_equation", + "content": "T_{i}^{G}" + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "text", + "content": ", describing the content of image, " + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 399, + 361, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 361, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 399, + 361, + 545, + 376 + ], + "type": "interline_equation", + "content": "T _ {i} ^ {G} = G \\left(I _ {i}\\right) \\tag {4}", + "image_path": "e8c6d1d81547f2e78e1eeb092904c14ebf73e787418e07c6f5a16bc67b8dd83a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "content": "Given a captioning model pre-trained on a small, representative and well-aligned dataset of image-text pairs, we are interested in estimating the alignment between image-text pairs sampled from a very large, diverse but noisy dataset. The alignment score can then be used as a ranking metric for dataset filtering. We hypothesize that:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 456, + 545, + 563 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 456, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 545, + 491 + ], + "type": "text", + "content": "- The probability of generating a caption that is semantically similar to the alt-text from an aligned pair is much higher than that from a misaligned pair." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 492, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 492, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 545, + 563 + ], + "type": "text", + "content": "- The probability of generating a caption that is semantically similar to a hard alt-text is higher than generating a caption that is semantically similar to a misaligned alt-text. Here, a hard alt-text is a text label with low likelihood with respect to the captioning model, but is aligned with the image content." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "type": "text", + "content": "As images can contain multiple objects with complex attributes and relationships, there exist multiple ways to describe their content. Given the inherent many-to-many relationship between images and text labels, our goal is to increase the probability of generating a caption that matches an aligned alt-text. To achieve this, we utilize nucleus sampling [12], a decoding strategy used to sample multiple captions, " + }, + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "type": "text", + "content": ", per-image:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 359, + 666, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 666, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 359, + 666, + 545, + 681 + ], + "type": "interline_equation", + "content": "G (I, r) = \\left\\{T _ {0} ^ {G}, T _ {1} ^ {G}, \\dots , T _ {r - 1} ^ {G} \\right\\} \\tag {5}", + "image_path": "53c4c71c383c4eb550c241b6d00d80c54da35fcf87826835afb317639c05c906.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Sentence Transformer Given an image, its alt-text, and a set of generated captions, our goal is to estimate the align-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22425" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 77, + 517, + 231 + ], + "blocks": [ + { + "bbox": [ + 78, + 77, + 517, + 231 + ], + "lines": [ + { + "bbox": [ + 78, + 77, + 517, + 231 + ], + "spans": [ + { + "bbox": [ + 78, + 77, + 517, + 231 + ], + "type": "image", + "image_path": "bcb5e268942e6704ce3b5bd2b457eb99bd8b594b3c6be2cbcb09b9befc031b68.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "type": "text", + "content": "Figure 2. Our proposed framework enables dataset pruning using image captioning models. To evaluate the alignment of a noisy image-text pair, we generate multiple captions per image using nucleus sampling. Then medium phrases, like \"an image of\" or \"a photo of\", are masked from alt-text and generated captions. Finally, a lightweight sentence encoder is used to semantically compare the generated captions with alt-text." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 289, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 523 + ], + "type": "text", + "content": "ment between the generated captions and the alt-text. However, there is a very large diversity gap between the generated captions and the highly diverse alt-text as measured by the number of unique nouns and trigrams [26]. On the other hand, constructing a large, diverse and curated image-text dataset is expensive, which limits the diversity of the generated captions. We propose to bridge this gap by utilizing a light-weight sentence similarity model to encode the alt-text and the generated captions. We expect the semantically similar alt-text and generated caption pairs to be closely clustered in the embedding space compared to semantically distinct pairs. We reason that the rich semantic textual embedding space of the sentence similarity model enables pretraining the captioning model only on a small but curated image-text dataset. Thus, we rely on the semantic understanding of the sentence similarity model to bridge the gap between the limited diversity in the generated captions and the highly diverse alt-text labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "text", + "content": "To estimate the alignment score, we compute the cosine similarity between embeddings of each generated caption and text label. Let " + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "text", + "content": " be a language model that encodes a text sample, " + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "text", + "content": ", to a vector, " + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "inline_equation", + "content": "S(T) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "text", + "content": ". We define the alignment between two text samples, " + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "inline_equation", + "content": "T_a" + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "inline_equation", + "content": "T_b" + }, + { + "bbox": [ + 46, + 524, + 288, + 596 + ], + "type": "text", + "content": ", as the cosine similarity between their language model encodings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 604, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 604, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 135, + 604, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\langle S \\left(T _ {a}\\right), S \\left(T _ {b}\\right) \\rangle \\tag {6}", + "image_path": "46991b9c1e3f93372a935d991135b93628d155e032dd108f33293232ac687caf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 624, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 288, + 649 + ], + "type": "text", + "content": "This estimate can then be used as a proxy for the image-text alignment of an image, " + }, + { + "bbox": [ + 47, + 624, + 288, + 649 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 624, + 288, + 649 + ], + "type": "text", + "content": ", and text, " + }, + { + "bbox": [ + 47, + 624, + 288, + 649 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 624, + 288, + 649 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 656, + 203, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 656, + 203, + 671 + ], + "spans": [ + { + "bbox": [ + 131, + 656, + 203, + 671 + ], + "type": "interline_equation", + "content": "\\langle S (G (I)), S (T) \\rangle", + "image_path": "49ee4e3392e1f11cb3de78cc262a8ad24481fa21e025e222f1cbae868209bcbe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "If the image captioning model generates " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " different caption candidates for an image, " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": ", we can use the maximum alignment between each of the generated captions," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 307, + 527, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 527, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 527, + 322 + ], + "type": "inline_equation", + "content": "G(I,r) = \\{T_0^G,T_1^G,\\ldots ,T_{r - 1}^G\\}" + }, + { + "bbox": [ + 306, + 307, + 527, + 322 + ], + "type": "text", + "content": ", and a text sample, " + }, + { + "bbox": [ + 306, + 307, + 527, + 322 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 307, + 527, + 322 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 373, + 327, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 327, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 373, + 327, + 545, + 350 + ], + "type": "interline_equation", + "content": "\\max _ {T _ {j} ^ {G} \\in G (I, r)} \\langle S \\left(T _ {j} ^ {G}\\right), S (T) \\rangle \\tag {7}", + "image_path": "4bd56365cb1e31a21d6dd7ac8832335258f2a9e4c8258e1366ee64b46c032c21.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 354, + 546, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 688 + ], + "type": "text", + "content": "In literature, there are different models and approaches to obtain text embeddings. [16] uses the average of N-gram features of each word in a text sample to obtain an embedding. A more common option is to use the logits of the last token generated by a decoder-only language model, which is the approach taken with CLIP's text encoder [32], in GPT-1 [31], as well as in [1]. Encoder-only models, such as BERT [6] or RoBERTa [23], can also be used, where the embedding vector may be either the logits of the classification token, or the average pool of the logits of all tokens. Although such language models may have strong generation or classification capabilities, they were not optimized for sentence similarity tasks, but either for next word prediction (i.e., causal language modeling) or masked word prediction (i.e., masked language modeling) tasks. Therefore, their embeddings may not be ideal to measure alignment between sentences. More importantly, such models are large in size and hence slow to infer on large datasets. A language model finetuned on a sentence similarity task, such as SNLI [3], aligns with the goal of estimating semantic textual similarity between alt-text and captions. We find that sentence similarity models [34] pretrained using a self-supervised instance discrimination task on unlabeled corpus of sentences perform well in estimating the alignment between text pairs, and are lightweight in size and latency (e.g., tens of millions of parameters in contrast to billions of parameters of performant decoder-only large language models)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Masking Medium Words Phrases such as \"image of\", \"picture of\", or \"photo of\" can appear in either alt-text or" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22426" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": "generated captions. We refer to such phrases as \"medium phrases\", as they describe the medium rather than the contents of an image. We notice that the existence of such medium phrases adds noise to the sentence similarity score, as shown in Figure 3. A pair of sentences that each have a medium phrase are assigned a misleadingly high sentence similarity score by a sentence transformer, as they have been trained on a wide and diverse corpus of text, rather than on image captions. Hence, the existence of medium phrases may increase their attention to the topic of images or media, rather than the topic of the content of such images. Therefore, we neutralize the effect of medium phrases by removing them from both alt-text and generated captions. We express the operation of masking medium words in a text sample, " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "M(T)" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", on text, " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", as masking all possible contiguous subsequences of the text, where masking on a phrase, " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", removes it if it is in the pre-determined list of medium phrases, " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\text{\"image of\" }, \\text{\"picture of\" }, \\text{\"photo of\" }, \\ldots \\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", or keeps it otherwise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 300, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 288, + 323 + ], + "type": "text", + "content": "Putting it all together, we define the Sieve score function as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 331, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 287, + 354 + ], + "type": "interline_equation", + "content": "f _ {\\text {S i e v e}} (I, T) = \\max _ {T _ {j} ^ {G} \\in G (I, r)} \\left\\langle S \\left(M \\left(T _ {j} ^ {G}\\right)\\right), S (M (T)) \\right\\rangle \\tag {8}", + "image_path": "35ea3db6287af5640c023721cb2e7a01b5883b435fe78244de904975888de861.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 287, + 385 + ], + "type": "text", + "content": "The dataset pruned using Sieve with the top " + }, + { + "bbox": [ + 47, + 360, + 287, + 385 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 360, + 287, + 385 + ], + "type": "text", + "content": " portion of its samples can be expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 393, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 393, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 111, + 393, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\text {S i e v e}, k} = \\operatorname {p r u n e} _ {f _ {\\text {S i e v e}}} (\\mathcal {D}, k) \\tag {9}", + "image_path": "58452ab7cf474e0ba83c629a99d2c6b11ade9bfbb04195d6bbc34bb748741bff.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 414, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 287, + 440 + ], + "type": "text", + "content": "We summarize our approach in Figure 2 and as psuedocode in Algorithm 1 in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 448, + 128, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 128, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 128, + 462 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 468, + 185, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 185, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 185, + 480 + ], + "type": "text", + "content": "4.1. Training and Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "text", + "content": "We utilize the DataComp benchmark to evaluate the utility of image-captioning models for multimodal dataset pruning. Two candidate pools are considered, the medium and the large scale, consisting of 128 million and 1.28 billion image-text pairs, respectively. To train CLIP models, we use DataComp's hyperparameters and architectures to standardize training [9]: " + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "text", + "content": " learning rate, 500 iterations warmup, AdamW optimizer, for medium scale: ViT-B/32 image encoder [7], batch size 4096, 128M training samples as a compute budget, and for large scale: ViT-B/16 image encoder, batch size 8192, 1.28B training samples as a compute budget. We evaluate the zero-shot performance on 38 downstream tasks, including classification and retrieval tasks [18, 32, 45]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "content": "For our captioning model, we utilize BLIP with ViT-B/16 image encoder pretrained on 14 million image-text pairs [20], including highly curated web datasets: Conceptual Captions [37], Conceptual 12M [4] and SBU captions [29]; as well as small human-annotated datasets:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 546, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 133 + ], + "type": "text", + "content": "COCO, and Visual Genome [17]. To compute the alignment between generated captions and alt-text, we use a lightweight distilled sentence transformer, all-MiniLM-L6-v2 [39], further finetuned using self-supervised contrastive learning on unlabeled text corpus." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 139, + 391, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 139, + 391, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 139, + 391, + 151 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 157, + 546, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 277 + ], + "type": "text", + "content": "Table 1 reports multiple baselines from DataComp [9], including applying no filtering, basic filtering, and CLIP-Score filtering. On the medium scale, Sieve with an image-captioning model pretrained on 30 times less but curated data surpasses CLIPScore by " + }, + { + "bbox": [ + 304, + 157, + 546, + 277 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 304, + 157, + 546, + 277 + ], + "type": "text", + "content": " on average. In addition, we demonstrate that Sieve's alignment score provides a complimentary signal to CLIPScore by fusing Sieve with CLIPScore. We apply min-max normalization to Sieve alignment scores and CLIPScore independently, then take the per-sample weighted average of both scores:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 300, + 535, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 300, + 535, + 331 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 535, + 331 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {\\text {S i e v e} + \\text {C L I P}} (I, T) = \\\\ (1 - \\alpha) \\times \\bar {f} _ {\\text {S i e v e}} (I, T) + \\alpha \\times \\bar {f} _ {\\text {C L I P}} (I, T) \\\\ \\end{array}", + "image_path": "b498d2e66a705bc817859c140b73759e0930b393c0fd5b9d2ea68e114fdc84ff.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 350, + 535, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 350, + 535, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 535, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} \\text {s . t .} \\quad \\overline {{f}} (I, T) = \\end{array} \\\\ \\frac {f (I , T) - \\min _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i})}{\\max _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i}) - \\min _ {(I _ {i} , T _ {i}) \\in \\mathcal {D}} f (I _ {i} , T _ {i})} \\\\ \\end{array}", + "image_path": "6fbf1200e46eb0938682f120da67a805839f5226f23b825b376d98a5c47612e3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "content": "where the weight " + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "content": " used in the reported results is 0.5. Finally, we select the top " + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "content": " of samples. We observe that the fused approach improves CLIPScore average performance by " + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "content": " on medium and large scale respectively, demonstrating the complementarity of Sieve's alignment score. Moreover, Sieve without fusion achieves the best performance on retrieval tasks on both medium and large scale experiments. Relative to CLIPScore, a significant portion of Sieve's retrieval gain is effectively transferred when fusing Sieve with CLIPScore (i.e., " + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "inline_equation", + "content": "+2.7\\%" + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "inline_equation", + "content": "+4.5\\%" + }, + { + "bbox": [ + 304, + 401, + 546, + 557 + ], + "type": "text", + "content": " on medium and large scale). This is also promising as retrieval performance is critical for retrieval augmented VLMs [13, 42]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 563, + 434, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 434, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 434, + 575 + ], + "type": "text", + "content": "4.3. Per-Task Performance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Figure 5 shows the change in accuracy introduced by Sieve as well as Sieve+CLIPScore on each task compared to CLIPScore on medium scale, and Figure 6 of the Appendix shows for the large scale. We observe that in addition to outperforming on image retrieval tasks, Flickr [43], and MS COCO [22], Sieve's greatest performance boost comes from WingoGAViL [2], a retrieval task which requires diverse reasoning skills, including general knowledge, common sense, and abstraction. This high performance can be attributed to Sieve's preference towards keeping samples where the alt-text correctly describes visual concepts and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22427" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 80, + 462, + 196 + ], + "blocks": [ + { + "bbox": [ + 133, + 80, + 462, + 196 + ], + "lines": [ + { + "bbox": [ + 133, + 80, + 462, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 80, + 462, + 196 + ], + "type": "image", + "image_path": "eaf9ab012b2e130c30e3b372567c1addb373a1d1e6e5326d2612736c654e1174.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": "Figure 3. Masking medium phrases improves the ranking of sentence similarity scores. On the left, sentence pairs with misleadingly high (or low) sentence similarity due to the existence (or absence) of medium phrases are highlighted in dark red (or light red). On the right, similarity scores that are more aligned with semantics are highlighted in dark green. The sentence similarity scores are computed using the all-MiniLM-L6-v2 sentence transformer [39]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 268, + 545, + 464 + ], + "blocks": [ + { + "bbox": [ + 48, + 268, + 545, + 464 + ], + "lines": [ + { + "bbox": [ + 48, + 268, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 545, + 464 + ], + "type": "table", + "html": "
ScaleFilteringDataset SizeImageNetImageNet dist. shiftsVTABRetrievalAverage over 38 datasets
Medium (128 Million)No Filtering128M17.615.225.921.925.8
Basic Filtering30M22.619.328.425.128.5
LAION Filtering13M23.019.830.723.329.2
CLIPScore38M27.323.033.825.132.8
Sieve24M29.425.035.228.934.6
Sieve+CLIPScore24M30.325.436.227.835.4
Large (1.28 Billion)No Filtering1.28B45.937.842.641.943.7
Basic Filtering298M51.642.344.648.045.8
LAION Filtering130M55.345.351.049.550.1
CLIPScore384M57.847.453.846.652.9
Sieve235M57.347.852.052.052.3
Sieve+CLIPScore235M59.749.154.851.154.6
", + "image_path": "3e0dae1e0f73786287d9f6e3a13a2ff694df1091dfdd4c5487accdfb3c320178.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "text", + "content": "Table 1. Zero-shot performance of CLIP models pretrained using various filtering strategies on medium and large scale pools of the DataComp benchmark. Sieve fused with CLIPScore beats CLIPScore by " + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "text", + "content": " on medium and large scale respectively. In addition, on retrieval tasks, Sieve achieves best performance on both scales and after fusing with CLIPScore, leads an improvement on retrieval over CLIPScore of " + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 472, + 547, + 518 + ], + "type": "text", + "content": " on medium and large scale respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 537, + 288, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 288, + 574 + ], + "type": "text", + "content": "their attributes and relations. Sieve, especially when combined with CLIPScore, significantly outperforms on medical diagnosis tasks, Camelyon17 and PatchCamelyon." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Sieve mainly underperforms in tasks requiring parsing text from images, such as MNIST [19], SVHN [25], and Rendered SST-2 [28], concluding that Sieve is less likely to select image-text pairs that are useful for OCR tasks. In addition, Sieve underperforms CLIPScore on context-based tasks like Country211 [27], a task assessing the geolocation capability of visual representations, demonstrating Sieve's preference towards selecting samples based on the alignment of alt-text with visual concepts rather than context. Interestingly, when fusing with CLIPScore, we improve the performance of all these tasks while retaining the advantage" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 537, + 500, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 500, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 500, + 550 + ], + "type": "text", + "content": "of Sieve especially at large scale (see Figure 6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 560, + 407, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 407, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 407, + 572 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "content": "We conduct studies on the medium scale pool and report the average of three runs per experiment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Pretraining data-distribution We study the effect of the pretraining data distribution used to train the captioning model on the quality of the alignment score. This is measured based on the downstream performance of the CLIP model trained using the selected image-text pairs. Two pretraining data distributions proposed by BLIP [20] are investigated. The first uses 14 million curated image-text pairs, while the second uses an additional 115 million web images with noisy alt-text [36]. Although the original BLIP" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22428" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 85, + 288, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 85, + 288, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 85, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 85, + 288, + 205 + ], + "type": "image", + "image_path": "4e807ca87d0e3eb1912fd0a8bb0a0737556557f42142380d5cf415a637b80b51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": "Figure 4. Evaluating CLIP models pretrained on different fractions of the top ranked samples based on our proposed approach (Sieve), CLIPScore, and fusing Sieve with CLIPscore (Sieve+CLIPSCore), on medium scale." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 86, + 452, + 205 + ], + "blocks": [ + { + "bbox": [ + 294, + 86, + 452, + 205 + ], + "lines": [ + { + "bbox": [ + 294, + 86, + 452, + 205 + ], + "spans": [ + { + "bbox": [ + 294, + 86, + 452, + 205 + ], + "type": "image", + "image_path": "50f0e847567f9ef0efad18ef1657a8f14727f96870d55466c32991801dd49059.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 98, + 248, + 496, + 355 + ], + "blocks": [ + { + "bbox": [ + 98, + 248, + 496, + 355 + ], + "lines": [ + { + "bbox": [ + 98, + 248, + 496, + 355 + ], + "spans": [ + { + "bbox": [ + 98, + 248, + 496, + 355 + ], + "type": "table", + "html": "
Percentage KeptCaption Generator Pretraining DataImageNetImageNet dist. shiftVTABRetrievalAverage over 38 datasets
10BLIP-129M23.0020.6030.2021.4030.00
BLIP-14M26.5022.5032.1023.7530.90
15BLIP-129M25.9522.8032.8024.4032.40
BLIP-14M29.0524.6033.3526.9532.70
20BLIP-129M27.8523.6533.4526.3533.05
BLIP-14M29.6024.9335.0728.5734.03
", + "image_path": "6b6dfc36c7f8768e6032601f46909d1e6ed328909a3a97093248e1d2d5d8b6ba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 491 + ], + "type": "text", + "content": "work reports higher captioning performance when pretraining on 115 million samples, our results in Table 2 indicate that for the purpose of dataset pruning, using curated image-text pairs results in a better alignment score than using a much larger noisy dataset. This highlights the importance of using a captioning model pretrained on higher quality data for large-scale dataset pruning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "content": "Captioning models We conduct experiments to investigate the generalizability of Sieve to other captioning models. In Table 3, we compare GIT [38], which utilizes an image encoder, and a text decoder captioning model, pretrained on 10 million image-text pairs to BLIP [20], which utilizes an image encoder, text encoder and a text decoder and is pretrained on 14 million image-text pairs. We observe that fusing Sieve with CLIPScore based on GIT and BLIP improves retrieval performance by " + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "inline_equation", + "content": "3.13\\%" + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "inline_equation", + "content": "3.83\\%" + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "content": " and improves average performance on 38 downstream tasks by " + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "inline_equation", + "content": "2.52\\%" + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Text embedding space In Table 4, we ablate over embeddings from different text models and show that embeddings from our selected sentence transformer perform better than embeddings from CLIP and BLIP text encoders. The CLIP text encoder was pretrained along with the CLIP vision encoder to map text and images to the same embedding space, and is used in diffusion models to condition image genera" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 403, + 544, + 497 + ], + "blocks": [ + { + "bbox": [ + 46, + 362, + 546, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 362, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 546, + 386 + ], + "type": "text", + "content": "Table 2. Effect of a caption generator's pretraining data-distribution on Sieve. The 14M pretraining dataset consists of curated image-text pairs, while the 129M dataset includes an additional 115M noisy image-text pairs from LAION [36]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 403, + 544, + 497 + ], + "lines": [ + { + "bbox": [ + 309, + 403, + 544, + 497 + ], + "spans": [ + { + "bbox": [ + 309, + 403, + 544, + 497 + ], + "type": "table", + "html": "
FilteringImageNetVTABRetrievalAvg
CLIPScore27.1333.9024.1732.63
Sieve-GIT27.4733.2027.2732.40
+CLIPScore28.9034.8327.3033.70
Sieve-BLIP29.6035.0728.5734.03
+CLIPScore30.3535.9028.0035.15
", + "image_path": "808cf490816859cda521407ef5197fe4326cbe647d6a52e7d17e8ee14e54daec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 505, + 545, + 529 + ], + "lines": [ + { + "bbox": [ + 305, + 505, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 505, + 545, + 529 + ], + "type": "text", + "content": "Table 3. Effect of captioning models, GIT [38] pretrained on 10M image-text pairs, and BLIP [20] pretrained on 14M pairs" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 542, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 673 + ], + "type": "text", + "content": "tion [33]. However, we observe that the CLIP text encoder suffers from poor semantic textual understanding, leading to a large drop in accuracy when used as a caption similarity measure. BLIP's text encoder performs better than that of CLIP, but the lightweight sentence transformer specifically pretrained on aligning semantically similar texts performs significantly better with " + }, + { + "bbox": [ + 304, + 542, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\geq 2\\%" + }, + { + "bbox": [ + 304, + 542, + 545, + 673 + ], + "type": "text", + "content": " improvements across various task types. In Figure 7 in the Appendix we show how cosine similarities of sentence similarity models result in better semantic textual clustering compared to CLIP and BLIP text encoders." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": "Pruning percentage We study the effect of the fraction of samples selected for pretraining. For each experiment, we compute the Sieve alignment score and CLIPScore for each" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22429" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 70, + 472, + 139 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 472, + 139 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 472, + 139 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 472, + 139 + ], + "type": "table", + "html": "
Text EncoderImageNetImageNet dist shiftVTABRetrievalAverage over 38 datasets
CLIP18.0015.6526.9020.9525.90
BLIP27.2022.1032.7025.4531.85
Sentence Transformer29.6024.9335.0728.5734.03
", + "image_path": "e95e62bd6f1e0f6eda52720c2fce408e4211dd9bbc7ca2052d9f79518ff85b9d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 201, + 284, + 339 + ], + "blocks": [ + { + "bbox": [ + 51, + 201, + 284, + 339 + ], + "lines": [ + { + "bbox": [ + 51, + 201, + 284, + 339 + ], + "spans": [ + { + "bbox": [ + 51, + 201, + 284, + 339 + ], + "type": "image", + "image_path": "7adffcfebeef1609f6edb9e435b1a71c90fdbee32c6ff793493257c422c9b779.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 344, + 253, + 354 + ], + "lines": [ + { + "bbox": [ + 81, + 344, + 253, + 354 + ], + "spans": [ + { + "bbox": [ + 81, + 344, + 253, + 354 + ], + "type": "text", + "content": "(a) Sieve gain over CLIPScore on medium scale pool." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 357, + 284, + 495 + ], + "blocks": [ + { + "bbox": [ + 50, + 357, + 284, + 495 + ], + "lines": [ + { + "bbox": [ + 50, + 357, + 284, + 495 + ], + "spans": [ + { + "bbox": [ + 50, + 357, + 284, + 495 + ], + "type": "image", + "image_path": "7399286e9c74e23defc786d289f598219c4806b89eb130de48644823e04d2078.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 499, + 273, + 510 + ], + "lines": [ + { + "bbox": [ + 61, + 499, + 273, + 510 + ], + "spans": [ + { + "bbox": [ + 61, + 499, + 273, + 510 + ], + "type": "text", + "content": "(b) Sieve+CLIPScore gain over CLIPScore on medium scale pool." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "type": "text", + "content": "Figure 5. The relative performance gain of Sieve and Sieve+CLIPScore relative to CLIPScore on 38 downstream tasks on the medium scale pool." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": "sample. The top-" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": " and pretraining CLIP models are then selected. Here, " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": ". Finally, we report the zero-shot performance on ImageNet and the average on 38 tasks in Figure 4. We observe that Sieve achieves the best performance using " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": " of the data, while CLIPScore peaks at " + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 566, + 287, + 662 + ], + "type": "text", + "content": " (similar to results reported in 9). Hence, pruning using Sieve achieves better performance with less data, compared to CLIPScore." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Number of generated captions and fusion with CLIP-Score We study the effect of using multiple captions per image to maximize the alignment of the generated captions with the alt-text. For nucleus sampling [12], we set the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 199, + 531, + 316 + ], + "blocks": [ + { + "bbox": [ + 46, + 147, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 547, + 182 + ], + "type": "text", + "content": "Table 4. Effect of the sentence encoder on the performance of Sieve. CLIP uses the text encoder pretrained on 400M samples [36], BLIP uses the text encoder pretrained on the curated 14M samples defined in [20], and Sentence Transformer uses a language model pretrained on unlabeled text corpus [39]. Each encoder encodes generated captions and the alt-text where the textual semantic alignment is computed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 199, + 531, + 316 + ], + "lines": [ + { + "bbox": [ + 321, + 199, + 531, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 199, + 531, + 316 + ], + "type": "table", + "html": "
Generated CaptionsCLIP weightImageNetAverage over 38 datasets
10.028.6032.50
20.029.0033.40
40.029.5333.70
80.029.6034.03
0.330.1034.40
80.530.3535.15
0.730.2534.35
", + "image_path": "221f6635e42131526baefbed106bd2b8e4b357dde5af81c9a1aa69971c795938.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 324, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 305, + 324, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 546, + 346 + ], + "type": "text", + "content": "Table 5. Effect of number of generated captions and weight of CLIPScore on zero-shot performance of pretrained CLIP models" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 355, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 546, + 548 + ], + "type": "text", + "content": "cumulative probability of the smallest set of words to 0.9, and the minimum and maximum sequence lengths to 5 and 20, respectively. We study the effect of sampling 1, 2, 4, and 8 captions. For each input image-text pair, we assign the maximum alignment score between the alt-text and the generated captions. We observe in Table 5 that increasing the number of generated captions improves the performance on downstream tasks. We reason that due to the many-to-many relationship between images and captions, generating more captions increases the probability of matching a hard aligned alt-text. We also investigate the effect of fusing the Sieve alignment score with CLIPScore in Table 5. Each score is independently normalized, and a weighted average is applied between the two scores. Finally, the top " + }, + { + "bbox": [ + 304, + 355, + 546, + 548 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 355, + 546, + 548 + ], + "type": "text", + "content": " of samples ranked by Sieve+CLIPScore are selected. We observe that a weight of 0.5 achieves the best performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 560, + 378, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 378, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 378, + 573 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We introduce a novel method, Sieve, that enables pruning large-scale noisy web-crawled image-text datasets. We propose utilizing synthetic captions from image captioning models pretrained on small, diverse, and curated datasets to evaluate the alignment of noisy image-text pairs. Using the embedding space of a lightweight sentence transformer, we compute an alignment score between generated captions and alt-text. We demonstrate that Sieve provides a complementary pruning signal to CLIPScore, effectively minimizing false positives and negatives, leading to improved zero-shot classification and retrieval performance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22430" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S. Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication, 2023. 1, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "text", + "content": "[2] Yonatan Bitton, Nitzan Bitton Guetta, Ron Yosef, Yuval Elovici, Mohit Bansal, Gabriel Stanovsky, and Roy Schwartz. WinoGAVil: Gamified association benchmark to challenge vision-and-language models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 190, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 288, + 235 + ], + "type": "text", + "content": "[3] Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 291 + ], + "type": "text", + "content": "[4] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "text", + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 480 + ], + "type": "text", + "content": "[8] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 545 + ], + "type": "text", + "content": "[9] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortzman, Dhruba Ghosh, Jieyu Zhang, et al. Dat acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 1, 2, 3, 5, 8, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 288, + 612 + ], + "type": "text", + "content": "[10] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "text", + "content": "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 657, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 691 + ], + "type": "text", + "content": "[12] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In ICLR. OpenReview.net, 2020. 2, 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[13] Ziniu Hu, Ahmet Iscen, Chen Sun, Zirui Wang, Kai-Wei Chang, Yizhou Sun, Cordelia Schmid, David A Ross, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 127 + ], + "type": "text", + "content": "Alireza Fathi. Reveal: Retrieval-augmented visual-language pre-training with multi-source multimodal knowledge memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23369-23379, 2023. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "text", + "content": "[14] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[15] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 263 + ], + "type": "text", + "content": "[16] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics, 2017. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "type": "text", + "content": "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "text", + "content": "[18] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "text", + "content": "[19] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6, 11" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 511 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 567 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "text", + "content": "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "text", + "content": "[24] Pratyush Maini, Sachin Goyal, Zachary C Lipton, J Zico Kolter, and Aditi Raghunathan. T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132, 2023. 3, 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6, 11" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22431" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. arXiv preprint arXiv:2307.10350, 2023. 1, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 203, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 203, + 128 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 203, + 128 + ], + "type": "text", + "content": "[27] OpenAI. The country211 dataset, .6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 129, + 210, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 210, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 210, + 139 + ], + "type": "text", + "content": "[28] OpenAI. The rendered sst2 dataset, .6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "text", + "content": "[29] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 185, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 185, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 185, + 287, + 258 + ], + "type": "text", + "content": "[30] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6967-6977, 2023. 2, 3, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 261, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 283 + ], + "type": "text", + "content": "[31] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 284, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 348 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 350, + 286, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 286, + 392 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 394, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 426 + ], + "type": "text", + "content": "[34] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 427, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 427, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 427, + 287, + 491 + ], + "type": "text", + "content": "[35] William A Gaviria Rojas, Sudnya Diamos, Keertan Ranjan Kini, David Kanter, Vijay Janapa Reddi, and Cody Coleman. The dollar street dataset: Images representing the geographic and socioeconomic diversity of the world. In Thirtysixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 493, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 493, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 49, + 493, + 287, + 548 + ], + "type": "text", + "content": "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 548, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 624 + ], + "type": "text", + "content": "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 49, + 625, + 287, + 669 + ], + "type": "text", + "content": "[38] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[39] Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, 2020. 5, 6, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 384 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[40] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "type": "text", + "content": "[41] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In Proceedings of the 40th International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 217 + ], + "type": "text", + "content": "[42] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Richard James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. 2023. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "text", + "content": "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 317 + ], + "type": "text", + "content": "[44] Haichao Yu, Yu Tian, Sateesh Kumar, Linjie Yang, and Heng Wang. The devil is in the details: A deep dive into the rabbit hole of data filtering. arXiv preprint arXiv:2309.15954, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 319, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 384 + ], + "type": "text", + "content": "[45] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22432" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_content_list.json b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e27f86a94a9f751679ee5baa50fb779d20808466 --- /dev/null +++ b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_content_list.json @@ -0,0 +1,1593 @@ +[ + { + "type": "text", + "text": "SignGraph: A Sign Sequence is Worth Graphs of Nodes", + "text_level": 1, + "bbox": [ + 202, + 127, + 767, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shiwei Gan† Yafeng Yin†* Zhiwei Jiang† Hongkai Wen‡ Lei Xie† Sanglu Lu†", + "bbox": [ + 153, + 164, + 815, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† State Key Laboratory for Novel Software Technology, Nanjing University, China", + "bbox": [ + 238, + 186, + 728, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\ddagger$ Department of Computer Science, The University of Warwick, UK", + "bbox": [ + 281, + 203, + 687, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sw@smail.nju.edu.cn {yafeng, jzw, lxie, sanglu}@nju.edu.cn hongkai.wen@warwick.ac.uk", + "bbox": [ + 114, + 222, + 848, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 272, + 313, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the recent success of sign language research, the widely adopted CNN-based backbones are mainly migrated from other computer vision tasks, in which the contours and texture of objects are crucial for identifying objects. They usually treat sign frames as grids and may fail to capture effective cross-region features. In fact, sign language tasks need to focus on the correlation of different regions in one frame and the interaction of different regions among adjacent frames for identifying a sign sequence. In this paper, we propose to represent a sign sequence as graphs and introduce a simple yet effective graph-based sign language processing architecture named SignGraph, to extract cross-region features at the graph level. SignGraph consists of two basic modules: Local Sign Graph (LSG) module for learning the correlation of intra-frame cross-region features in one frame and Temporal Sign Graph (TSG) module for tracking the interaction of inter-frame cross-region features among adjacent frames. With LSG and TSG, we build our model in a multiscale manner to ensure that the representation of nodes can capture cross-region features at different granularities. Extensive experiments on current public sign language datasets demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues. Code and models are available at: https://github.com/gswycf/SignGraph.", + "bbox": [ + 75, + 303, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 717, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer vision technology and natural language processing technology have greatly advanced sign language (SL) research, including Sign Language Recognition (SLR) and Sign Language Translation (SLT). The former task aims at recognizing an isolated sign/continuous signs as a corresponding gloss or gloss sequence, while the latter task aims at translating continuous signs into spoken language [3]. In the modern sign language processing models (SLR models and SLT models), one of the key points of these deep learn", + "bbox": [ + 75, + 742, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/028e330c5cb10bd0a08b0e6b2bb532a6b1bc9f1021d3bb53a297117ab6c920d0.jpg", + "image_caption": [ + "Figure 1. Comparison of feature extraction between convolutional network and our graph convolutional network." + ], + "image_footnote": [], + "bbox": [ + 521, + 272, + 635, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5471e091b92d6d53f96f20954b39fa1e7c4f81793988af4e7a0c10d470d6af87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 282, + 754, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96a2f22030b91665634dfd0a25b05b452c2a74ff0817dc153247813e291942db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 282, + 864, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing based methods is to learn sign-related features through training with extensive sign video data. As a matter of course, 2D CNNs and 3D CNNs are used to be the defacto standard visual backbones [7, 10] for extracting local sign frame features, while 1D CNN, LSTM and Transformer [3, 11, 37] are used as the temporal module to capture dynamic changes of sign frames.", + "bbox": [ + 496, + 416, + 892, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, as shown in Figure 1, the widely-used CNN-based backbones that are mainly migrated from image recognition and object detection tasks, treat the sign frame/sign clips as grids and adopt 'kernels' to extract local features by sliding across sign frames. These CNNs are good at capturing contour-based representation and texture-based representation [12], but they may fail to capture the explicit collaboration of signs in different regions. In fact, sign language tasks need to focus on manual and non-manual features, especially the collaboration of these cues in different regions [35]. Besides, temporal modules like LSTM and Transformer layer usually regard a sign frame as a whole region and learn the overall variance of whole frames, while ignoring the interaction of subregions in adjacent frames. Other temporal methods like 1D CNN can only focus on the same regions in adjacent frames, which may fail to track dynamic motions across different regions. In fact, the dynamic movements in the body, hands, and face which are reflected in the same/different regions between adjacent frames, are primary elements of identifying a sign for sign language tasks.", + "bbox": [ + 496, + 522, + 892, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to make up for the shortcomings of typical CNN backbones explicitly or implicitly and allow the model to learn robust sign-related features, exiting work committed to designing various CNN-based backbones by injecting do", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Yafeng Yin is the corresponding author.", + "bbox": [ + 94, + 887, + 312, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13470", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "main knowledge i.e., skeleton [11], depth images [30] or local areas [37]. Besides, some work trained their backbones with more training samples by back translation [36], cross modality augmentation [28] and pre-training models [6], while others added extra constraints on backbones by using knowledge distillation [16], contrastive learning [10].", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from previous work, we build a simple yet effective sign graph neural network (SignGraph for short) for continuous sign language recognition (CSLR) task, in which we take sign frames as graphs of nodes and build dynamic graphs to learn cross-region features. The main intuitions are: (1) Sign-related features involve the correlation of different regions in one frame, such as face region and hand regions. Thus, we propose a local sign graph $LSG$ module by dynamically building graphs based on nodes in each frame, to learn intra-frame cross-region features. (2) The dynamic movements in the body, hands, and face are reflected in different regions among adjacent frames. Thus we propose a temporal sign graph $TSG$ module by dynamically connecting sign-related regions based on nodes in adjacent frames, to learn inter-frame cross-region features. (3) Considering that patches with a single size may not capture local features effectively, we build our SignGraph in a multiscale manner by merging small patches into larger-size patches, and use different-size patches to capture cross-region features at different granularities. We make the following contributions:", + "bbox": [ + 75, + 185, + 473, + 503 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Different from CNN-based SL models, we propose a graph-based CSLR model (SignGraph), which processes sign frames as a set of nodes and dynamically builds graphs from one frame and adjacent frames to capture cross-region features related to signs, thus can aggregate sign-related features from different regions for better feature representation.", + "- We propose a new local sign graph module $LSG$ that dynamically builds edges based on node features to learn the correlation of cross-region features in one frame, and a new temporal sign graph module $TSG$ that dynamically connects sign-related regions to learn the interaction of cross-region features among adjacent frames.", + "- We build our model in a multiscale manner with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities.", + "- The extensive experiments on CSLR task demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues." + ], + "bbox": [ + 76, + 512, + 468, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 840, + 218, + 857 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We mainly review CSLR technologies and graph convolutional networks in sign language.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continuous Sign Language Recognition. The goal of CSLR is to interpret the continuous sign sequence as corresponding gloss sequence [24], thus the core of current CSLR tasks is to identify signs. The widely adopted CNN-based backbones in CSLR which are migrated from the image recognition task, are good at capturing the texture and contours of objects [12]. However, identifying a sign needs to focus on the correlation of different regions (e.g., hand regions and face region) in one frame and the interaction of the same or different regions (e.g., hand motions, facial expression changes) among adjacent frames. As a matter of course, to extract robust sign-related features with CNN-based backbones, some sign language models introduced prior knowledge (or expert knowledge) to guide the model to focus on sign-related features. For example, extra information like local areas [2, 4], skeletons [4, 7, 38] and depth images [30] was manually injected into model. However, in these models, on the one hand, to obtain such extra knowledge, additional algorithms need to be introduced, which will cause additional computational overhead. On the other hand, integrating such knowledge will lead to a complex network structure. Besides, some work turned its attention to obtaining more sign language samples by back translation [36], cross modality augmentation [28], and pretraining with other sign language related datasets [6, 7]. While others were dedicated to designing complex backbones with attention mechanism [18] or adding extra constraints on CNN backbones by adopting knowledge distillation [16, 26], contrastive learning [10].", + "bbox": [ + 496, + 90, + 893, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph Convolutional Network. Compared with sequences and grids, graphs are more flexible data structures that model the relationships of a set of objects with nodes and edges. In order to learn representation from the graph, graph convolutional network (GCN) [22] was proposed, and it has been widely applied in action recognition tasks [34] and has also been extended to sign language tasks [11, 17, 21]. To apply GCN to SL tasks, the existing models often relied on extracted skeleton, whose connections between joints make skeleton data naturally fit GCN. Specifically, some work utilized GCN to guide CNN-based backbones to extract skeleton-related features [11, 27]. While others ignored RGB features and applied GCN to explore sign features directly with skeleton [20, 21]. Recently, in addition to applying GCN with skeletons, the usage of applying GCN on images emerged. Specifically, VIG [15] represented an image as a graph structure to capture irregular and complex objects, and demonstrated the effectiveness of graph representation of images. Inspired by VIG, we find that cross-region features are important for identifying a sign and can be captured by graphs. Thus unlike the previous work which adopted CNN-based backbones or relied on skeletons to apply GCN, we directly treat sign frames as graphs of nodes and design a simple GCN-based framework", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d1360906ec4fa4240be220f9e174fedc1c399cb3a2fd003ad726e89ffc757af1.jpg", + "image_caption": [ + "Figure 2. The proposed SignGraph architecture." + ], + "image_footnote": [], + "bbox": [ + 120, + 92, + 851, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for learning sign-related features.", + "bbox": [ + 76, + 332, + 299, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 359, + 168, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first give an overview of the proposed graph-based sign language processing architecture. Then we describe the proposed Local Sign Graph $LSG$ module for capturing intra-frame cross-region features and Temporal Sign Graph $TSG$ module for capturing inter-frame cross-region features. Finally, with $LSG$ and $TSG$ , we describe our multiscale SignGraph for CSLR.", + "bbox": [ + 75, + 383, + 468, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 76, + 497, + 266, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a sign language video $f = \\{f_i\\}_{i=1}^{\\theta}$ with $\\theta$ frames, the target of CSLR model is to get a recognized gloss sequence $g = \\{g_i\\}_{i=1}^{\\vartheta}$ with $\\vartheta$ glosses based on input $f$ .", + "bbox": [ + 75, + 520, + 468, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, our SignGraph model consists of two key modules, i.e. Local Sign Graph $LSG$ module and Temporal Sign Graph $TSG$ module. First, we convert each frame $f_{i}$ into $N$ patches with a patchify stem. Then each frame $f_{i}$ is represented by a set of patches (i.e., nodes) $v_{i} = \\{v_{ij}\\}_{j=1}^{N}$ . Second, by learning the correlation between each node and connecting $K$ nearest neighbors for each node, we build a dynamic local graph $G_{i}^{l}$ for each frame $f_{i}$ . Then a local graph convolutional layer is leveraged to capture cross-region features within each frame. Third, by connecting sign-related regions between two adjacent frames, we build a dynamic temporal graph $G_{i}^{t}$ for adjacent frames $f_{i}$ and $f_{i+1}$ . Then, a temporal graph convolutional layer is adopted to track the interaction of cross-region features among adjacent frames.", + "bbox": [ + 75, + 566, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Further, we merge patches into larger size patches by downsampling frames with a factor of 2, then each frame $f_{i}$ can be represented by another set of patches (nodes) $\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}$ . Following that, we repeatedly adopt LSG module and TSG module to learn cross-region features with different-size patches. After that, following previous sign language models [10, 26], we adopt a global feature module", + "bbox": [ + 75, + 792, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to learn global changes of whole frames. Finally, a classifier and a widely-used CTC loss [13] are adopted to predict the probability $p(g|f)$ of target gloss sequence.", + "bbox": [ + 496, + 332, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sign Graph Learning", + "text_level": 1, + "bbox": [ + 498, + 386, + 702, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patchify Stem. In order to build graphs based on sign frames instead of skeleton data, we need to convert RGB data into nodes. Following previous work [9, 32], for each frame $f_{i} \\in \\mathbb{R}^{[H \\times W \\times 3]}$ with height $H$ and width $W$ , we first convert it into a number of patches (or nodes) $v_{i} = \\{v_{ij}\\}_{j=1}^{N}$ with a patchify stem $\\mathcal{PS}$ .", + "bbox": [ + 496, + 409, + 890, + 501 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} = \\left\\{\\mu_ {i j} \\right\\} _ {j = 1} ^ {N} = \\mathcal {P S} (f _ {i}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 513, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $v_{ij}$ represents the $j$ th node in the $i$ -th frame $f_i$ , $\\mu_{ij} \\in \\mathbb{R}^D$ denotes node features of $v_{ij}$ , $D$ is the feature dimension, $N = HW / P^2$ is the number of patches, and $P$ is the patch size.", + "bbox": [ + 496, + 537, + 893, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local Sign Graph Learning. For a single frame in the sign video, sign language models should focus on the correlation of different regions, such as face region and hand regions. In order to learn intra-frame cross-region features, we need to build a graph for frame $f_{i}$ , in which regions with correlation (e.g., hand regions and face region) should be connected.", + "bbox": [ + 496, + 602, + 890, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 3, for a set of nodes $v_{i}$ , we first adopt a linear projection layer with learnable weights $W_{1}^{l}$ to map node features $\\mu_{i}$ to the space where the distance function is applied: $\\mu_{i}' = \\mu_{i}W_{1}^{l}$ . Then we adopt $K$ -Nearest Neighbors (KNN) algorithm to find top $\\mathsf{K}_l$ nearest neighbors $\\mathcal{N}_{\\mathsf{K}_l}(v_{ij})$ for each node based on distance function $\\mathcal{D}\\mathcal{I}\\mathcal{S}$ [15].", + "bbox": [ + 496, + 693, + 890, + 784 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D I S} \\left(v _ {i j}, v _ {i k}\\right) = \\sqrt {\\sum \\left(\\mu_ {i j} ^ {\\prime} - \\mu_ {i k} ^ {\\prime}\\right) ^ {2}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 792, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} _ {\\mathrm {K} _ {l}} \\left(v _ {i j}\\right) = \\left\\{v _ {i k} \\mid v _ {i k} \\in \\operatorname {T o p} K \\left(\\left\\{\\mathcal {D I S} \\left(v _ {i j}, v _ {i k}\\right) \\right\\} _ {k = 1} ^ {N}\\right) \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 838, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $TopK$ function outputs the top $K$ nearest neighbors for $v_{ij}$ based on node distances. For each node $v_{ij}$ ,", + "bbox": [ + 496, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13472", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c4e5cec7923a3e95cbb2a3c5b72987a0ad3e17c8516918f6da2ba13472f69af5.jpg", + "image_caption": [ + "Figure 3. Graph construction in local sign graph module." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 457, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we add an edge $e(v_{ij}, v_{ik})$ between $v_{ik}$ and $v_{ij}$ , where $v_{ik} \\in \\mathcal{N}_{\\mathsf{K}_l}(v_{ij})$ . Thus, for frame $f_i$ , we can get the edge set $e_i^l = \\{e(v_{ij}, v_{ik}), |j \\in [1, N], v_{ik} \\in \\mathcal{N}_{\\mathsf{K}_l}(v_{ij})\\}$ , and build a local graph $G_i^l = \\{v_i, e_i^l\\}$ . After that, we adopt a graph convolutional layer $\\mathcal{GCN}_l$ and a linear projection layer with learnable weights $W_2^l$ to aggregate features from cross regions for each frame.", + "bbox": [ + 75, + 215, + 472, + 321 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} = \\mu_ {i} + \\mathcal {G C N} _ {l} \\left(\\mu_ {i} ^ {\\prime}, e _ {i} ^ {l}\\right) W _ {2} ^ {l} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 329, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The LSG module contains one graph convolutional layer and two linear projection layers. We do not add extra modules (e.g., feed-forward network) since we aim to keep our model simple, and we show that our model can achieve a convincing performance even with the simple design.", + "bbox": [ + 75, + 356, + 468, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Sign Graph Learning. The dynamic movements in body, hands, and face are essential elements to identify a sign, which are often reflected in different regions among adjacent frames. Hence, we design a temporal graph module, which can dynamically build edges between regions in two adjacent frames and learn inter-frame cross-region features.", + "bbox": [ + 75, + 436, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For two adjacent frames $f_{i}$ and $f_{i+1}$ , we have $2N$ nodes $\\{v_{ij}\\}_{j=1}^{N}$ and $\\{v_{(i+1)k}\\}_{k=1}^{N}$ . Similarly, we first adopt a linear projection layer with weights $W_{1}^{t}$ to map node features $\\mu_{i}$ to the space where the distance function is applied: $\\mu_{i}^{\\prime \\prime} = \\mu_{i}W_{1}^{t}$ . Then we compute distances between $v_{ij}$ and $v_{(i+1)k}$ , thus we have a $N \\times N$ distance matrix $M$ , where $\\{M[j,k] = \\mathcal{DLS}(v_{ij},v_{(i+1)k})|j\\in [1,N],k\\in [1,N]\\}$ . When building the temporal graph, unlike the LSG module where we build edges for each node $v_{ij}$ (i.e., building a dense graph), we only choose top $K_{t}$ node pairs between two adjacent frames and connect them (i.e., building a sparse graph). To avoid establishing edges between meaningless node pairs, we set a threshold of 0.05 on normalized distance matrix $M$ to filter out patch pairs in the background, since these pairs often have a high and constant similarity (i.e., $M[j,k] < 0.05$ ). Then we get the temporal edge set $e_{i}^{t}$ between frame $f_{i}$ and $f_{i+1}$ and temporal graph $G_{i}^{t} = \\{(v_{i},v_{i+1}),e_{i}^{t}\\}$ .", + "bbox": [ + 76, + 542, + 472, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} ^ {t} = \\left\\{e \\left(v _ {i j}, v _ {(i + 1) k}\\right) \\mid \\mathcal {D I S} \\left(v _ {i j}, v _ {(i + 1) k}\\right) \\in T o p K (M) \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 823, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, we adopt a temporal graph convolutional layer $\\mathcal{GCN}_t$ and another linear projection $W_2^t$ to aggregate cross-region features from adjacent frames, as follows.", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d48394fb8a432a0326745c25dbc5604ceee138eb1f26c76632d773909248d2b1.jpg", + "image_caption": [ + "Figure 4. Graph construction in temporal sign graph module." + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 870, + 200 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mu_ {i} \\right\\} _ {i = 1} ^ {\\theta} = \\left\\{\\mu_ {i} \\right\\} _ {i = 1} ^ {\\theta} + \\mathcal {G C N} _ {t} \\left(\\left\\{\\mu_ {i} ^ {\\prime \\prime} \\right\\} _ {i = 1} ^ {\\theta}, \\left\\{e _ {i} ^ {t} \\right\\} _ {i = 1} ^ {\\theta - 1}\\right) W _ {2} ^ {t} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 255, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intuition of building the sparse graph for $TSG$ module rather than building the dense graph like $LSG$ module is to avoid the over-smoothing phenomenon [19], which may decrease the distinctiveness of node features.", + "bbox": [ + 496, + 287, + 892, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multiscale Sign Graph Convolutional Network", + "text_level": 1, + "bbox": [ + 498, + 356, + 892, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patches with a fixed window may not capture sign features in one region effectively (e.g., hand regions may be split by patches). Thus, similar to Swin Transformer [25], we propose a multiscale SignGraph, which builds our model in multiscale manners to ensure that the representation of nodes can capture cross-region features at different granularities.", + "bbox": [ + 496, + 380, + 892, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patch Merging. To produce a multiscale representation, we further design a patch merging module to reduce the number of patches (nodes) and expand the receptive field of patches. Specifically, for $N$ patches (or nodes) with the size of $16 \\times 16$ in frame $f_{i}$ , the patch merging module applies convolutional layers to downsample frames by a factor of 2. Then we get a set of nodes $\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}$ for frame $i$ , and its corresponding node features $\\{\\bar{\\mu}_{ij}\\}_{j=1}^{N/4}$ where $\\bar{\\mu}_{ij} \\in \\mathbb{R}^{2D}$ .", + "bbox": [ + 496, + 492, + 893, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multiscale Sign Graph Learning. After getting the set of nodes $\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}$ with a larger size for frame $i$ by patch merging, we adopt the another LSG module with $\\mathsf{K}_l$ edges for each node in one frame and TSG module with $\\mathsf{K}_t$ edges for two adjacent frames. For convenience, we name the ith LSG, TSG module as LSG $_i$ , TSG $_i$ , and the corresponding hyperparameters as $\\mathsf{K}_l^i$ , $\\mathsf{K}_t^i$ . According to experimental results in Section 4.3, we adopt two LSG, TSG modules in our baseline settings.", + "bbox": [ + 496, + 625, + 893, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 777, + 633, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 801, + 602, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the previous work [16, 36], we evaluate our model on three publicly available SL datasets. (1) PHOENIX14 [23] is a widely used German SL dataset for CSLR with a vocabulary of 1295 glosses from 9 signers. It contains 5672, 540, and 629 weather forecast sam", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "plies for training, validation, and testing respectively. (2) PHOENIX14T [3] is another German SL dataset with both gloss annotations and translation annotations. It contains 7096, 519, and 642 samples from 9 signers for training, validation, and testing respectively, and it has a vocabulary of 1066 different signs for CSLR. (3) CSL-Daily is a Chinese SL dataset, which contains 18401, 1077, and 1176 labeled videos from 10 signers for training, validation, and testing respectively, and it has gloss annotations with a vocabulary of 2000 glosses for CSLR.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Setting", + "text_level": 1, + "bbox": [ + 76, + 252, + 277, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we describe the baseline settings for our architecture.", + "bbox": [ + 76, + 275, + 467, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Preprocessing. For data preprocessing, we follow the existing work [18, 26] and apply data augmentation during training, including: resizing frames to $256 \\times 256$ pixels, random cropping frames to $224 \\times 224$ pixels, random horizontal flipping with a probability of 0.5, random temporal scaling ( $\\pm 20\\%$ ). During testing, we resize frames to $256 \\times 256$ pixels and use center cropping to get frames with $224 \\times 224$ pixels.", + "bbox": [ + 75, + 297, + 467, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture Setting. Our architecture is implemented by PyTorch 1.11. (1) Patchify stem: Considering that sign frames have a strong 2D local structure, i.e., spatially neighboring pixels are usually highly correlated, we do not adopt non-overlapping patches, which are obtained by linear projection [9] and lack modeling power for 2D local spatial context [32, 33]. Instead, we obtain initial patches (nodes) embedding by using convolutional-based patchify stem. Specifically, we adopt the first four stages of ResNet18 as our patchify stem. The dimension of initial patches $D$ is set to 512. (2) Patch merging: Similar to the patchify stem, we adopt two convolutional blocks with kernel size $3 \\times 3$ to downsample frames and get different-size patches after LSG and TSG modules. (3) Sign Graph learning: There are two LSG and TSG modules in our baseline settings, and the initial $\\mathsf{K}_l^1$ , $\\mathsf{K}_l^2$ in LSG modules are set to 4 and initial $\\mathsf{K}_t^1$ , $\\mathsf{K}_t^2$ in TSG modules are set to 49 (49 is the number of nodes in $TSG_2$ ). We adopt the implementation of GCN-Conv [22] as initial graph convolutional layer (GCN layer). (4) Distance function: We adopt Euclidean distance to measure the distance between two nodes in the baseline setting. (5) Global feature module: The Global feature module is the combination of two 1D convolution blocks, 2-layer BiLSTM with hidden size 1024 for global feature modeling, and a fully connected layer for final prediction.", + "bbox": [ + 75, + 425, + 472, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Setting. For a fair comparison, we adopt the same training settings as in previous work [26]. We adopt the Adam optimizer with a weight decay of 0.0001 to train our model for 40 epochs on 1 GeForce RTX 3090 GPU. The initial learning rate is 0.0001 with a decay factor of 0.5 at epoch 20 and 30, and the batch size is set to 4. CTC loss", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a965897ba2644c582d46a0d174e12d8ba15de43bf02676caa6542a8807554cbb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
\\( LSG_1 \\)\\( TSG_1 \\)\\( LSG_2 \\)\\( TSG_2 \\)DevTest
WERDel/InsWERDel/ins
✘✘✘✘22.38.4/2.522.28.1/2.7
✓19.25.6/2.221.04.8/2.3
✓19.65.6/2.121.55.1/2.5
✓19.35.3/2.120.85.8/2.0
✓19.56.6/1.721.25.4/2.4
✓✓18.45.4/1.620.15.1/2.1
✓✓19.15.6/1.820.84.8/2.2
✓✓18.75.1 /2.320.65.2/1.7
✓✓18.64.3 /1.820.25.5/1.7
✓✓✓✓18.15.4/1.420.16.1/1.8
", + "bbox": [ + 513, + 88, + 883, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Ablation study on Phoenix14T dataset.", + "bbox": [ + 552, + 284, + 836, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is applied after both 1D convolution blocks and the fully connected layer as loss functions.", + "bbox": [ + 498, + 311, + 890, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation setting. To evaluate our model, we adopt the Word Error Rate (WER) as CSLR performance metric,", + "bbox": [ + 498, + 354, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW E R = \\frac {\\# S + \\# I + \\# D}{\\# N} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 396, + 890, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\# S, \\# I, \\# D$ denote the minimum number of substitution, insertion and deletion operations needed to convert a predicted gloss sequence to the ground truth. $\\# N$ is the number of glosses in ground truth.", + "bbox": [ + 498, + 429, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 501, + 650, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the previous work [3], we perform ablation studies on Phoenix14T dataset to verify the effectiveness of the proposed SignGraph.", + "bbox": [ + 498, + 526, + 890, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effects of Proposed Sign Graph Module. As shown in Table 1, results demonstrate that both the proposed $LSG$ and $TSG$ modules contribute to higher performance for CSLR. Specifically, our baseline with ResNet18 as backbone achieves $22.3\\%$ WER on Phoenix14T dev set. When we replace backbone with the designed local sign graph module $LSG_{1}$ , $LSG_{2}$ , WER on dev set decreases by $3.1\\%$ and $3.0\\%$ respectively. When we replace backbone with designed temporal sign graph module $TSG_{1}$ , $TSG_{2}$ , WER decreases by $2.7\\%$ and $2.8\\%$ respectively, which verifies that both intra-frame cross-region features and inter-frame cross-region features are crucial for learning sign-related features.", + "bbox": [ + 496, + 582, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we combine $LSG$ and $TSG$ modules, WER decreases by $3.9\\%$ when applying $LSG_{1}$ and $TSG_{1}$ , $3.2\\%$ when applying $LSG_{2}$ and $TSG_{2}$ respectively, which demonstrates that combination of cross-region features in one frame and those among adjacent frames are beneficial for identifying signs. When adopting two stages of $LSG$ and $TSG$ , performance can achieve further improvement (i.e., WER decreases by $4.2\\% / 2.1\\%$ on dev/test set).", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13474", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/952d95c66642a9268df30f240dcaf6836965ff800d18afbf06512cd14e033477.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
\\( LSG_1 \\)\\( TSG_1 \\)DevTest
WERDel/InsWERDel/ins
DenseSparse18.45.4/1.620.15.1/2.1
DenseDense20.24.9/2.921.85.5/2.6
SparseSparse19.15.0/1.921.75.3/2.7
SparseDense20.95.4/2.421.96.1/2.5
", + "bbox": [ + 109, + 88, + 437, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation study on graph types.", + "bbox": [ + 153, + 195, + 392, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It shows that leveraging multiscale representation of nodes can capture cross-region features at different granularities.", + "bbox": [ + 75, + 222, + 468, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Graph Type. In SignGraph, we adopt dense graphs (i.e., connecting top $K$ neighbors for each node) in the LSG module, while adopting sparse graphs (only connecting top $K$ node pairs) in the TSG module. To verify the effectiveness of our design, we test the effect of different combinations of graph types in LSG and TSG modules. For convenience, we only adopt one LSG module and one TSG module. As shown in Table 2, when we adopt the dense graph in TSG module or adopt the sparse graph in LSG module, there is a noticeable performance drop (i.e., WER on dev set rises by $1.8\\%$ or $0.7\\%$ ). Our module with dense graph in LSG and sparse graph in TSG achieves the best recognition performance, which demonstrates the effectiveness of our model.", + "bbox": [ + 75, + 258, + 470, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of Hyperparameters K. In the proposed multiscale SignGraph, there are four hyperparameters, i.e., $\\mathsf{K}_l^1$ , $\\mathsf{K}_l^2$ in LSG modules and $\\mathsf{K}_t^1$ , $\\mathsf{K}_t^2$ in TSG modules. Here, to set proper hyperparameters K for our model, we test recognition performance under different $\\mathsf{K}_l$ ranging ranging from 2 to 9 and under different $\\mathsf{K}_t$ ranging from 7 to 91. Since there are many combinations of these hyperparameters, it is impossible to enumerate all possible values to find the globally optimal setting, thus we fix the other 3 parameters and adjust one to obtain suitable parameter settings. According to Figure 5, we set $\\mathsf{K}_l^1$ , $\\mathsf{K}_l^2$ , $\\mathsf{K}_t^1$ , $\\mathsf{K}_t^2$ to 3, 4, 49, 49 respectively based on WER performance on DEV set.", + "bbox": [ + 75, + 478, + 468, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of Different Backbones. To verify the effectiveness of the proposed SignGraph model, we replace our backbone with other patch-based networks (i.e., CvT [33], Swin transformer [25] and VIG [15]). According to Table 3a, simply adopting the current patch-based SOTA backbones will not gain appealing performance on CSLR task, while our proposed SignGraph backbone achieves satisfactory performance by effectively capturing intra-frame cross-region features and inter-frame cross-region features.", + "bbox": [ + 75, + 666, + 468, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of Distance Function. In the SignGraph, we adopt KNN algorithm to find the nearest neighbors based on the distance of two nodes. Here, we show performances with commonly used distance functions including Cosine distance, Chebyshev distance and Euclidean distance in Table 3b. There are subtle performance gaps between different", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71a4ca41c005b0bb9a7d28c7d142b36bc76f1dcc2fe0423398ed3abd28b41721.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 89, + 689, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1367bac6a034ff9617870c5b76b955d1040d55c540aedfde196433b7ee515148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 89, + 880, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bff04dab6e61bfe6e11f9ac902a6b13a4b809161366a10cdaddd062232944084.jpg", + "image_caption": [ + "Figure 5. Effects of $\\mathsf{K}_l$ and $\\mathsf{K}_t$ ." + ], + "image_footnote": [], + "bbox": [ + 503, + 193, + 687, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/600fa3fd1c0c550f42d3822f2f838cedc8fbb8892123a4855ea19c9086ebeca3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 193, + 879, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "distance functions and we adopt Euclidean distance in our model for better performances.", + "bbox": [ + 498, + 328, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of GCN Layer. We also test the representative variants of graph convolution, including GATv2Conv [1], SAGEConv [14], GCNConv [22] and EdgeConv [31]. From Table 3c, we can see that our model can achieve excellent performances with different GCN layer, indicating the flexibility of SignGraph model. Among them, EdgeConv achieves the best recognition performance (i.e., $17.8\\%$ WER). In the rest of the experiments, we use EdgeConv layer in our module by default unless specifically stated.", + "bbox": [ + 498, + 369, + 892, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Patch Sizes. Patches with different sizes can capture different local features and further affect model performance. To test the effect of patch sizes, we adopt only one LSG module and one TSG module. As shown in Table 3d, model with a smaller patch size (i.e., $8 \\times 8$ ) achieves the worst performance, while model with too large patch size may also degenerate local features.", + "bbox": [ + 498, + 515, + 890, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Multiscale SignGraph. To leverage the scale-invariant property of images [15], PyVIG adopts pyramid architecture that gradually increases patch size from 4 to 32, by gradually shrinking the spatial size of feature maps. Similarly, we also gradually add $LSG$ and $TSG$ modules in the early stage of the patchify stem and add patch merging module at the end of each $TSG$ module for downsampling. In regard to the added LSG and TSG modules in each stage, we also choose appropriate hyperparameters $K_{l}, K_{t}$ for them through extensive experiments. As shown in Table 3e, adding more $LSG$ and $TSG$ modules does not always bring more performance gain, thus we adopt two $LSG$ and $TSG$ modules that increase patch size from 16 to 32 in our model.", + "bbox": [ + 498, + 632, + 892, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of DropEdge. DropEdge [29] tackles oversmoothing and over-fitting problems for dense graph by randomly removing a certain number of edges from the input graph at each training epoch. Considering that we", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13475", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8cdd8844da732c97c63e895de0fdffe5cf66053b80f1d77cfa79e0d4a6a5cc9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackBoneWERDelIns
CvT [32]39.215.30.9
SwinT [25]45.416.31.3
PyVIG [15]35.412.11.3
Ours17.95.11.5
", + "bbox": [ + 109, + 77, + 307, + 145 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3b8800fe9ce9cbd19698d543f0280a672271e0ebb21be22d8340514ef2e9f0f.jpg", + "table_caption": [ + "(a) Backbone. Comparison of different backbones." + ], + "table_footnote": [], + "table_body": "
PatchSizeWERDelIns
819.14.82.6
1618.45.41.6
3218.75.71.7
", + "bbox": [ + 109, + 175, + 305, + 228 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e64c38a4150ce9f32b1380a06b07c6409fecebfcd0038e8f1a307f0b913627fa.jpg", + "table_caption": [ + "(d) Patch size. Comparison of different patch sizes with one LSG and TSG modules." + ], + "table_footnote": [], + "table_body": "
DistanceWERDelIns
Cosine18.05.01.5
Chebyshev18.24.81.6
Euclidean17.95.11.5
", + "bbox": [ + 370, + 77, + 602, + 132 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3b015c0ea1b3565f24c02a7f1de42af46353d30bf7479d02bcec0eda6ecfa12b.jpg", + "table_caption": [ + "(b) Distance function. Comparison of distance functions." + ], + "table_footnote": [], + "table_body": "
StagesWERDelDel
16→3217.85.01.6
8→16→3218.45.31.7
4→8→16→3218.14.71.9
", + "bbox": [ + 374, + 174, + 599, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b523f2e340e8995134a0bc7b222f3e360103b09655b7935663e1c23cf9530364.jpg", + "table_caption": [ + "(e) Multiscale SignGraph. Effect of the number of stages in multiscale SignGraph." + ], + "table_footnote": [], + "table_body": "
GraphConvWERDelIns
GATv2Conv [1]18.65.21.8
SAGEConv [14]18.04.81.7
GCNConv [22]17.95.11.5
EdgeConv [31]17.85.01.6
", + "bbox": [ + 645, + 77, + 897, + 145 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/feb42feb581fd4b577d00ef5457dadf5b33ad7fe4402861dcb973af37eeb7258.jpg", + "table_caption": [ + "(c) Graph Convolution Effect of different GCN layers." + ], + "table_footnote": [], + "table_body": "
DropRateWERDelIns
017.85.01.6
15%19.26.71.3
30%21.35.91.7
", + "bbox": [ + 645, + 174, + 898, + 228 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(f) Drop edge. Adding DropEdge [29] in the local graph does not improve performance.", + "bbox": [ + 638, + 232, + 887, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7196f2e33b9ae4497c2a4e47eab83a841d05d857f5d4fcfa67f7afcd834b1067.jpg", + "image_caption": [ + "Table 3. Ablation experiments of SignGraph on Phoenix14T dev set.", + "Figure 6. Visualization of graph construction of $LSG_{1}$ and $TSG_{1}$ in the first row, $LSG_{2}$ and $TSG_{2}$ in the second row. The graph in $LSG$ module is shown in yellow, and the graph in $TSG$ is shown in blue. We also show some 'unimportant' edges between nodes in the background with red color." + ], + "image_footnote": [], + "bbox": [ + 133, + 292, + 836, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "build dense graphs for $LSG$ module, we also add DropEdge with different drop rates to $LSG$ module and show the recognition performance in Table 3f. However, adding DropEdge does not improve our model performance, which means that our model can achieve excellent performance without relying on external modules, and can handle over-smoothing and over-fitting problems well.", + "bbox": [ + 75, + 614, + 468, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of SignGraph. To verify whether our model can effectively capture sign-related features, we select a sign video from Phoenix14T test set and visualize the constructed graph structure of $LSG$ and $TSG$ modules in both two stages. In Figure 6, we show the graphs of two stages and we only show part of the edges for a better display. In the $LSG$ module, our model can link the nodes with similar content and semantic representation (e.g., hand regions and face regions) for extracting better intra-frame cross-region features. In the $TSG$ module, edges can be built for tracking dynamic changes in gestures and facial expres", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sions, thus can better identify a sign with inter-frame cross-region features. We also show some 'unimportant' edges between nodes in the background with red color. As can be seen, background nodes in $LSG$ are naturally connected to their neighboring nodes, and there are still a few background nodes connected in $TSG$ . Fortunately, nodes in the background will not 'disturb' nodes in sign-related regions, which demonstrates the effectiveness of our model.", + "bbox": [ + 496, + 614, + 890, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 753, + 638, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Phoenix14T dataset. As shown in Table 4, we compare our model with existing models on CSLR performance, and we provide both the performances on the validation set (i.e., 'DEV') and test set (i.e., 'TEST'). Most of the current models adopt existing CNN-based backbones, and achieve convincing performances by injecting extra cues [7, 35, 38], adding extra constraints [10, 16, 26] or introducing attention mechanism [18]. The GCN-based", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13476", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0c1e9d2c9f7411027d7375990a3e30d1bd57e75ddca2fcd63658b30da3860d41.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBackboneExtra cuesPhoenixPhoenixT
F/MHSPDEVTESTDEVTEST
WERdel/insWERdel/insWERWER
STMC [35] (AAAI'20)VGG11✓✓✓21.17.7/3.420.77.4/2.619.621.0
C2SLR [38](CVPR'22)ResNet18✓20.5-/-20.4-/-20.220.4
TwoStream [7](NeurIPS'22)S3D✓✓✓✓*18.4-/-18.8-/-17.719.3
SLT [5] (CVPR'20)GooleNet✓----24.624.5
TwoStream [7] (NeurIPS'22)S3D✓*22.4-/-23.3-/-21.122.4
VAC [26] (ICCV'21)ResNet18✓21.27.9/2.522.38.4/2.6--
SMKD [16] (ICCV'21)ResNet18✓20.86.8/2.521.06.3/2.320.822.4
CorrNet [18] (CVPR'23)ResNet18✓18.85.6/2.819.45.7/2.318.920.5
FCN [8] (ECCV'20)customed23.7-/-23.9-/-
Contrastive [10] (IJCAI'23)ResNet1819.65.1/2.719.85.8/3.020.020.1
HST-GNN [21] (WACV'22)Customized(CNN+GCN)✓✓✓✓19.5-/-19.8-/-19.519.8
CoSign [20] (ICCV'23)ST-GCN(GCN)✓✓✓19.7-/-20.1-/-19.520.1
SignGraphCustomized(GCN)✓18.45.6/1.820.15.4/2.118.320.0
MultiSignGraphCustomized(GCN)✓18.24.9/2.019.15.3/1.917.819.1
", + "bbox": [ + 99, + 85, + 870, + 314 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa252f2a559f8155d5660f11ea92651eca9fadac773cd2f009b4fc6ed799557f.jpg", + "table_caption": [ + "Table 4. Comparison of CSLR performance on Phoenix14 and Phoenix14T datasets. (F: face, M: mouth, H: hands, S: skeleton, P: pretraining backbone with ImageNet, $\\sqrt{}^{*}$ : pretraining on other SL-related datasets.)" + ], + "table_footnote": [], + "table_body": "
ModelBackboneExtra cuesDEVTEST
SPWERdel/insWERdel/ins
SLT [5](CVPR'20)GoogleNet✓33.110.3/4.432.09.6/4.1
TwoStream [7](NeurIPS'22)S3D✓✓*25.4-/-25.3-/-
TwoStream [7](NeurIPS'22)S3D✓*28.9-/-28.5-/-
BN-TIN [36](CVPR21)GoogLeNet✓33.613.9/3.433.113.5/3.0
CorrNet [18](CVPR'23)ResNet18✓30.6-/-30.1-/-
CoSign [20] (ICCV'23)ST-GCN(GCN)✓28.1-/-27.2-/-
SignGraphcustomized(GCN)✓28.48.8/2.427.48.2/2.1
MultiSignGraphcustomized(GCN)✓27.37.9/2.326.47.8/2.1
", + "bbox": [ + 80, + 359, + 465, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison of CSLR performance on CSL-daily dataset.", + "bbox": [ + 76, + 479, + 468, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model CoSign [20] mainly relies on pre-processed fine-grained skeleton data, and achieves $19.5\\%$ , $20.1\\%$ WER on dev, test set. While HST-GCN [21] adopts both CNN-based backbone and GCN-based backbone to extract RGB features and skeleton features respectively, achieving $19.5\\%$ , $19.8\\%$ WER on dev, test set. It is worth mentioning that the state-of-the-art (SOTA) model TwoStream utilizes both RGB features and fine-grained skeleton features (i.e., keypoints in hands, body and face), and pretrains its backbone on other SL-related datasets, achieving $17.7\\%$ , $19.3\\%$ WER on dev, test set respectively. Without using any extract cues, our best model still achieves comparable performance with TwoStream on dev set. While on test set, our model even outperforms TwoStream model by $0.2\\%$ WER.", + "bbox": [ + 75, + 503, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on Phoenix14 dataset. We also compare our model with existing models of CSLR performance on Phoenix14 dataset. As shown in Table 4, with the simple yet powerful multiscale SignGraph, our model can achieve the best performance (i.e., $18.2\\%$ WER) on dev set and a comparable performance with SOTA model on test set, which demonstrates the effectiveness of our model.", + "bbox": [ + 75, + 724, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on CSL-daily dataset. We also provide CSLR performance comparisons with other methods on CSL-daily dataset. As shown in Table 5, the WER of our model on dev and test set is $27.3\\%$ and $26.4\\%$ , respectively. Our model", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outperforms most existing models, and achieves a comparable performance with TwoStream model.", + "bbox": [ + 498, + 362, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 420, + 617, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet effective SignGraph architecture, which represents a sign sequence as a graph structure to extract sign-related features at the graph level. Specifically, to learn the correlation of cross-region features in one frame, we propose a Local Sign Graph $LSG$ module, which dynamically builds edges between nodes in one frame and aggregates intra-frame cross-region features. To track the interaction of cross-region features among adjacent frames, we propose a Temporal Sign Graph $TSG$ module, which dynamically builds edges between nodes among adjacent frames and captures inter-frame cross-region features. Besides, we also build our model in multiscale manners with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities. The extensive experiments on public datasets demonstrate the superiority of the proposed SignGraph, which achieves comparable performances with the SOTA model without using any extra cues.", + "bbox": [ + 496, + 450, + 890, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 750, + 679, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Key Research and Development Program of China under Grant No. 2022YFB3303900; Jiangsu Provincial Key Research and Development Program under Grant BE2020001-4; National Natural Science Foundation of China under Grant Nos. 62172208, 62272216. This work is partially supported by Collaborative Innovation Center of Novel Software Technology and Industrialization.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13477", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021. 6, 7", + "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In ICCV, pages 3075-3084. IEEE, 2017. 2", + "[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, pages 7784-7793, 2018. 1, 5", + "[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In ECCV, pages 301-319. Springer, 2020. 2", + "[5] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, pages 10023-10033, 2020. 8", + "[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 2", + "[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 35:17043-17056, 2022. 1, 2, 7, 8", + "[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV. Springer, 2020. 8", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 5", + "[10] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Kang Xia, Lei Xie, and Sanglu Lu. Contrastive learning for sign language recognition and translation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 763-772. 1, 2, 3, 7, 8", + "[11] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Lei Xie, and Sanglu Lu. Skeleton-aware neural sign language translation. In MM, pages 4353-4361, 2021. 1, 2", + "[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2", + "[13] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In ICML, pages 369-376, 2006. 3", + "[14] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2, 3, 6, 7", + "[16] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In ICCV, pages 11303-11312, 2021. 2, 4, 7, 8", + "[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[18] Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Continuous sign language recognition with correlation network. In CVPR, pages 2529-2539, 2023. 2, 5, 7, 8", + "[19] Wenbing Huang, Yu Rong, Tingyang Xu, Fuchun Sun, and Junzhou Huang. Tackling over-smoothing for general graph convolutional networks. arXiv preprint arXiv:2008.09864, 2020. 4", + "[20] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 8", + "[21] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 8", + "[22] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 5, 6, 7", + "[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. CVIU, 141:108-125, 2015. 4", + "[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In CVPR, pages 4297-4305, 2017. 2", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4, 6, 7", + "[26] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In ICCV, pages 11542-11551, 2021. 2, 3, 5, 7, 8", + "[27] Maria Parelli, Katerina Papadimitriou, Gerasimos Potamianos, Georgios Pavlakos, and Petros Maragos. Spatiotemporal graph convolutional networks for continuous sign language recognition. In ICASSP. IEEE, 2022. 2", + "[28] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In MM, pages 1497-1505, 2020. 2", + "[29] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In International Conference on Learning Representations, 2020. 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13478", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. TMM, 2021. 2", + "[31] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7", + "[32] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 3, 5, 7", + "[33] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 5, 6", + "[34] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2", + "[35] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, 2020. 1, 7, 8", + "[36] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, pages 1316–1325, 2021. 2, 4, 8", + "[37] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, 2021. 1, 2", + "[38] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2, 7, 8" + ], + "bbox": [ + 78, + 90, + 467, + 542 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13479", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_model.json b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..67bb73b45ef30df0882b5179c11eff663099cdf7 --- /dev/null +++ b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.128, + 0.768, + 0.15 + ], + "angle": 0, + "content": "SignGraph: A Sign Sequence is Worth Graphs of Nodes" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.165, + 0.816, + 0.185 + ], + "angle": 0, + "content": "Shiwei Gan† Yafeng Yin†* Zhiwei Jiang† Hongkai Wen‡ Lei Xie† Sanglu Lu†" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.187, + 0.73, + 0.202 + ], + "angle": 0, + "content": "† State Key Laboratory for Novel Software Technology, Nanjing University, China" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.204, + 0.689, + 0.22 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Department of Computer Science, The University of Warwick, UK" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.223, + 0.849, + 0.238 + ], + "angle": 0, + "content": "sw@smail.nju.edu.cn {yafeng, jzw, lxie, sanglu}@nju.edu.cn hongkai.wen@warwick.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.273, + 0.314, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Despite the recent success of sign language research, the widely adopted CNN-based backbones are mainly migrated from other computer vision tasks, in which the contours and texture of objects are crucial for identifying objects. They usually treat sign frames as grids and may fail to capture effective cross-region features. In fact, sign language tasks need to focus on the correlation of different regions in one frame and the interaction of different regions among adjacent frames for identifying a sign sequence. In this paper, we propose to represent a sign sequence as graphs and introduce a simple yet effective graph-based sign language processing architecture named SignGraph, to extract cross-region features at the graph level. SignGraph consists of two basic modules: Local Sign Graph (LSG) module for learning the correlation of intra-frame cross-region features in one frame and Temporal Sign Graph (TSG) module for tracking the interaction of inter-frame cross-region features among adjacent frames. With LSG and TSG, we build our model in a multiscale manner to ensure that the representation of nodes can capture cross-region features at different granularities. Extensive experiments on current public sign language datasets demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues. Code and models are available at: https://github.com/gswycf/SignGraph." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.718, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Computer vision technology and natural language processing technology have greatly advanced sign language (SL) research, including Sign Language Recognition (SLR) and Sign Language Translation (SLT). The former task aims at recognizing an isolated sign/continuous signs as a corresponding gloss or gloss sequence, while the latter task aims at translating continuous signs into spoken language [3]. In the modern sign language processing models (SLR models and SLT models), one of the key points of these deep learn" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.273, + 0.637, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.283, + 0.756, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.283, + 0.865, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.379, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Figure 1. Comparison of feature extraction between convolutional network and our graph convolutional network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.893, + 0.522 + ], + "angle": 0, + "content": "ing based methods is to learn sign-related features through training with extensive sign video data. As a matter of course, 2D CNNs and 3D CNNs are used to be the defacto standard visual backbones [7, 10] for extracting local sign frame features, while 1D CNN, LSTM and Transformer [3, 11, 37] are used as the temporal module to capture dynamic changes of sign frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.893, + 0.84 + ], + "angle": 0, + "content": "However, as shown in Figure 1, the widely-used CNN-based backbones that are mainly migrated from image recognition and object detection tasks, treat the sign frame/sign clips as grids and adopt 'kernels' to extract local features by sliding across sign frames. These CNNs are good at capturing contour-based representation and texture-based representation [12], but they may fail to capture the explicit collaboration of signs in different regions. In fact, sign language tasks need to focus on manual and non-manual features, especially the collaboration of these cues in different regions [35]. Besides, temporal modules like LSTM and Transformer layer usually regard a sign frame as a whole region and learn the overall variance of whole frames, while ignoring the interaction of subregions in adjacent frames. Other temporal methods like 1D CNN can only focus on the same regions in adjacent frames, which may fail to track dynamic motions across different regions. In fact, the dynamic movements in the body, hands, and face which are reflected in the same/different regions between adjacent frames, are primary elements of identifying a sign for sign language tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In order to make up for the shortcomings of typical CNN backbones explicitly or implicitly and allow the model to learn robust sign-related features, exiting work committed to designing various CNN-based backbones by injecting do" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.313, + 0.901 + ], + "angle": 0, + "content": "*Yafeng Yin is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.183 + ], + "angle": 0, + "content": "main knowledge i.e., skeleton [11], depth images [30] or local areas [37]. Besides, some work trained their backbones with more training samples by back translation [36], cross modality augmentation [28] and pre-training models [6], while others added extra constraints on backbones by using knowledge distillation [16], contrastive learning [10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.474, + 0.504 + ], + "angle": 0, + "content": "Different from previous work, we build a simple yet effective sign graph neural network (SignGraph for short) for continuous sign language recognition (CSLR) task, in which we take sign frames as graphs of nodes and build dynamic graphs to learn cross-region features. The main intuitions are: (1) Sign-related features involve the correlation of different regions in one frame, such as face region and hand regions. Thus, we propose a local sign graph \\( LSG \\) module by dynamically building graphs based on nodes in each frame, to learn intra-frame cross-region features. (2) The dynamic movements in the body, hands, and face are reflected in different regions among adjacent frames. Thus we propose a temporal sign graph \\( TSG \\) module by dynamically connecting sign-related regions based on nodes in adjacent frames, to learn inter-frame cross-region features. (3) Considering that patches with a single size may not capture local features effectively, we build our SignGraph in a multiscale manner by merging small patches into larger-size patches, and use different-size patches to capture cross-region features at different granularities. We make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.513, + 0.47, + 0.62 + ], + "angle": 0, + "content": "- Different from CNN-based SL models, we propose a graph-based CSLR model (SignGraph), which processes sign frames as a set of nodes and dynamically builds graphs from one frame and adjacent frames to capture cross-region features related to signs, thus can aggregate sign-related features from different regions for better feature representation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.62, + 0.47, + 0.71 + ], + "angle": 0, + "content": "- We propose a new local sign graph module \\( LSG \\) that dynamically builds edges based on node features to learn the correlation of cross-region features in one frame, and a new temporal sign graph module \\( TSG \\) that dynamically connects sign-related regions to learn the interaction of cross-region features among adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.711, + 0.47, + 0.756 + ], + "angle": 0, + "content": "- We build our model in a multiscale manner with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.756, + 0.47, + 0.817 + ], + "angle": 0, + "content": "- The extensive experiments on CSLR task demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.513, + 0.47, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.841, + 0.22, + 0.858 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We mainly review CSLR technologies and graph convolutional networks in sign language." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.532 + ], + "angle": 0, + "content": "Continuous Sign Language Recognition. The goal of CSLR is to interpret the continuous sign sequence as corresponding gloss sequence [24], thus the core of current CSLR tasks is to identify signs. The widely adopted CNN-based backbones in CSLR which are migrated from the image recognition task, are good at capturing the texture and contours of objects [12]. However, identifying a sign needs to focus on the correlation of different regions (e.g., hand regions and face region) in one frame and the interaction of the same or different regions (e.g., hand motions, facial expression changes) among adjacent frames. As a matter of course, to extract robust sign-related features with CNN-based backbones, some sign language models introduced prior knowledge (or expert knowledge) to guide the model to focus on sign-related features. For example, extra information like local areas [2, 4], skeletons [4, 7, 38] and depth images [30] was manually injected into model. However, in these models, on the one hand, to obtain such extra knowledge, additional algorithms need to be introduced, which will cause additional computational overhead. On the other hand, integrating such knowledge will lead to a complex network structure. Besides, some work turned its attention to obtaining more sign language samples by back translation [36], cross modality augmentation [28], and pretraining with other sign language related datasets [6, 7]. While others were dedicated to designing complex backbones with attention mechanism [18] or adding extra constraints on CNN backbones by adopting knowledge distillation [16, 26], contrastive learning [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Graph Convolutional Network. Compared with sequences and grids, graphs are more flexible data structures that model the relationships of a set of objects with nodes and edges. In order to learn representation from the graph, graph convolutional network (GCN) [22] was proposed, and it has been widely applied in action recognition tasks [34] and has also been extended to sign language tasks [11, 17, 21]. To apply GCN to SL tasks, the existing models often relied on extracted skeleton, whose connections between joints make skeleton data naturally fit GCN. Specifically, some work utilized GCN to guide CNN-based backbones to extract skeleton-related features [11, 27]. While others ignored RGB features and applied GCN to explore sign features directly with skeleton [20, 21]. Recently, in addition to applying GCN with skeletons, the usage of applying GCN on images emerged. Specifically, VIG [15] represented an image as a graph structure to capture irregular and complex objects, and demonstrated the effectiveness of graph representation of images. Inspired by VIG, we find that cross-region features are important for identifying a sign and can be captured by graphs. Thus unlike the previous work which adopted CNN-based backbones or relied on skeletons to apply GCN, we directly treat sign frames as graphs of nodes and design a simple GCN-based framework" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.093, + 0.852, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.303, + 0.629, + 0.318 + ], + "angle": 0, + "content": "Figure 2. The proposed SignGraph architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.3, + 0.349 + ], + "angle": 0, + "content": "for learning sign-related features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.36, + 0.169, + 0.375 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.492 + ], + "angle": 0, + "content": "In this section, we first give an overview of the proposed graph-based sign language processing architecture. Then we describe the proposed Local Sign Graph \\( LSG \\) module for capturing intra-frame cross-region features and Temporal Sign Graph \\( TSG \\) module for capturing inter-frame cross-region features. Finally, with \\( LSG \\) and \\( TSG \\), we describe our multiscale SignGraph for CSLR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.267, + 0.513 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.568 + ], + "angle": 0, + "content": "For a sign language video \\( f = \\{f_i\\}_{i=1}^{\\theta} \\) with \\( \\theta \\) frames, the target of CSLR model is to get a recognized gloss sequence \\( g = \\{g_i\\}_{i=1}^{\\vartheta} \\) with \\( \\vartheta \\) glosses based on input \\( f \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.793 + ], + "angle": 0, + "content": "As shown in Figure 2, our SignGraph model consists of two key modules, i.e. Local Sign Graph \\( LSG \\) module and Temporal Sign Graph \\( TSG \\) module. First, we convert each frame \\( f_{i} \\) into \\( N \\) patches with a patchify stem. Then each frame \\( f_{i} \\) is represented by a set of patches (i.e., nodes) \\( v_{i} = \\{v_{ij}\\}_{j=1}^{N} \\). Second, by learning the correlation between each node and connecting \\( K \\) nearest neighbors for each node, we build a dynamic local graph \\( G_{i}^{l} \\) for each frame \\( f_{i} \\). Then a local graph convolutional layer is leveraged to capture cross-region features within each frame. Third, by connecting sign-related regions between two adjacent frames, we build a dynamic temporal graph \\( G_{i}^{t} \\) for adjacent frames \\( f_{i} \\) and \\( f_{i+1} \\). Then, a temporal graph convolutional layer is adopted to track the interaction of cross-region features among adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Further, we merge patches into larger size patches by downsampling frames with a factor of 2, then each frame \\( f_{i} \\) can be represented by another set of patches (nodes) \\( \\{\\bar{v}_{ij}\\}_{j=1}^{N/4} \\). Following that, we repeatedly adopt LSG module and TSG module to learn cross-region features with different-size patches. After that, following previous sign language models [10, 26], we adopt a global feature module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.38 + ], + "angle": 0, + "content": "to learn global changes of whole frames. Finally, a classifier and a widely-used CTC loss [13] are adopted to predict the probability \\( p(g|f) \\) of target gloss sequence." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.387, + 0.704, + 0.404 + ], + "angle": 0, + "content": "3.2. Sign Graph Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Patchify Stem. In order to build graphs based on sign frames instead of skeleton data, we need to convert RGB data into nodes. Following previous work [9, 32], for each frame \\( f_{i} \\in \\mathbb{R}^{[H \\times W \\times 3]} \\) with height \\( H \\) and width \\( W \\), we first convert it into a number of patches (or nodes) \\( v_{i} = \\{v_{ij}\\}_{j=1}^{N} \\) with a patchify stem \\( \\mathcal{PS} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.514, + 0.891, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} = \\left\\{\\mu_ {i j} \\right\\} _ {j = 1} ^ {N} = \\mathcal {P S} (f _ {i}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.894, + 0.598 + ], + "angle": 0, + "content": "Here, \\( v_{ij} \\) represents the \\( j \\)th node in the \\( i \\)-th frame \\( f_i \\), \\( \\mu_{ij} \\in \\mathbb{R}^D \\) denotes node features of \\( v_{ij} \\), \\( D \\) is the feature dimension, \\( N = HW / P^2 \\) is the number of patches, and \\( P \\) is the patch size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Local Sign Graph Learning. For a single frame in the sign video, sign language models should focus on the correlation of different regions, such as face region and hand regions. In order to learn intra-frame cross-region features, we need to build a graph for frame \\( f_{i} \\), in which regions with correlation (e.g., hand regions and face region) should be connected." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.785 + ], + "angle": 0, + "content": "As shown in Figure 3, for a set of nodes \\( v_{i} \\), we first adopt a linear projection layer with learnable weights \\( W_{1}^{l} \\) to map node features \\( \\mu_{i} \\) to the space where the distance function is applied: \\( \\mu_{i}' = \\mu_{i}W_{1}^{l} \\). Then we adopt \\( K \\)-Nearest Neighbors (KNN) algorithm to find top \\( \\mathsf{K}_l \\) nearest neighbors \\( \\mathcal{N}_{\\mathsf{K}_l}(v_{ij}) \\) for each node based on distance function \\( \\mathcal{D}\\mathcal{I}\\mathcal{S} \\) [15]." + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.793, + 0.892, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D I S} \\left(v _ {i j}, v _ {i k}\\right) = \\sqrt {\\sum \\left(\\mu_ {i j} ^ {\\prime} - \\mu_ {i k} ^ {\\prime}\\right) ^ {2}} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.839, + 0.891, + 0.871 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} _ {\\mathrm {K} _ {l}} \\left(v _ {i j}\\right) = \\left\\{v _ {i k} \\mid v _ {i k} \\in \\operatorname {T o p} K \\left(\\left\\{\\mathcal {D I S} \\left(v _ {i j}, v _ {i k}\\right) \\right\\} _ {k = 1} ^ {N}\\right) \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Here, \\(TopK\\) function outputs the top \\(K\\) nearest neighbors for \\(v_{ij}\\) based on node distances. For each node \\(v_{ij}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.458, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.192, + 0.444, + 0.207 + ], + "angle": 0, + "content": "Figure 3. Graph construction in local sign graph module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.473, + 0.323 + ], + "angle": 0, + "content": "we add an edge \\( e(v_{ij}, v_{ik}) \\) between \\( v_{ik} \\) and \\( v_{ij} \\), where \\( v_{ik} \\in \\mathcal{N}_{\\mathsf{K}_l}(v_{ij}) \\). Thus, for frame \\( f_i \\), we can get the edge set \\( e_i^l = \\{e(v_{ij}, v_{ik}), |j \\in [1, N], v_{ik} \\in \\mathcal{N}_{\\mathsf{K}_l}(v_{ij})\\} \\), and build a local graph \\( G_i^l = \\{v_i, e_i^l\\} \\). After that, we adopt a graph convolutional layer \\( \\mathcal{GCN}_l \\) and a linear projection layer with learnable weights \\( W_2^l \\) to aggregate features from cross regions for each frame." + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.33, + 0.47, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} = \\mu_ {i} + \\mathcal {G C N} _ {l} \\left(\\mu_ {i} ^ {\\prime}, e _ {i} ^ {l}\\right) W _ {2} ^ {l} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.435 + ], + "angle": 0, + "content": "The LSG module contains one graph convolutional layer and two linear projection layers. We do not add extra modules (e.g., feed-forward network) since we aim to keep our model simple, and we show that our model can achieve a convincing performance even with the simple design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.544 + ], + "angle": 0, + "content": "Temporal Sign Graph Learning. The dynamic movements in body, hands, and face are essential elements to identify a sign, which are often reflected in different regions among adjacent frames. Hence, we design a temporal graph module, which can dynamically build edges between regions in two adjacent frames and learn inter-frame cross-region features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.544, + 0.473, + 0.818 + ], + "angle": 0, + "content": "For two adjacent frames \\( f_{i} \\) and \\( f_{i+1} \\), we have \\( 2N \\) nodes \\( \\{v_{ij}\\}_{j=1}^{N} \\) and \\( \\{v_{(i+1)k}\\}_{k=1}^{N} \\). Similarly, we first adopt a linear projection layer with weights \\( W_{1}^{t} \\) to map node features \\( \\mu_{i} \\) to the space where the distance function is applied: \\( \\mu_{i}^{\\prime \\prime} = \\mu_{i}W_{1}^{t} \\). Then we compute distances between \\( v_{ij} \\) and \\( v_{(i+1)k} \\), thus we have a \\( N \\times N \\) distance matrix \\( M \\), where \\( \\{M[j,k] = \\mathcal{DLS}(v_{ij},v_{(i+1)k})|j\\in [1,N],k\\in [1,N]\\} \\). When building the temporal graph, unlike the LSG module where we build edges for each node \\( v_{ij} \\) (i.e., building a dense graph), we only choose top \\( K_{t} \\) node pairs between two adjacent frames and connect them (i.e., building a sparse graph). To avoid establishing edges between meaningless node pairs, we set a threshold of 0.05 on normalized distance matrix \\( M \\) to filter out patch pairs in the background, since these pairs often have a high and constant similarity (i.e., \\( M[j,k] < 0.05 \\)). Then we get the temporal edge set \\( e_{i}^{t} \\) between frame \\( f_{i} \\) and \\( f_{i+1} \\) and temporal graph \\( G_{i}^{t} = \\{(v_{i},v_{i+1}),e_{i}^{t}\\} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.824, + 0.47, + 0.856 + ], + "angle": 0, + "content": "\\[\ne _ {i} ^ {t} = \\left\\{e \\left(v _ {i j}, v _ {(i + 1) k}\\right) \\mid \\mathcal {D I S} \\left(v _ {i j}, v _ {(i + 1) k}\\right) \\in T o p K (M) \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "After that, we adopt a temporal graph convolutional layer \\(\\mathcal{GCN}_t\\) and another linear projection \\(W_2^t\\) to aggregate cross-region features from adjacent frames, as follows." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.872, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.208, + 0.878, + 0.223 + ], + "angle": 0, + "content": "Figure 4. Graph construction in temporal sign graph module." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.256, + 0.892, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mu_ {i} \\right\\} _ {i = 1} ^ {\\theta} = \\left\\{\\mu_ {i} \\right\\} _ {i = 1} ^ {\\theta} + \\mathcal {G C N} _ {t} \\left(\\left\\{\\mu_ {i} ^ {\\prime \\prime} \\right\\} _ {i = 1} ^ {\\theta}, \\left\\{e _ {i} ^ {t} \\right\\} _ {i = 1} ^ {\\theta - 1}\\right) W _ {2} ^ {t} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.893, + 0.349 + ], + "angle": 0, + "content": "The intuition of building the sparse graph for \\( TSG \\) module rather than building the dense graph like \\( LSG \\) module is to avoid the over-smoothing phenomenon [19], which may decrease the distinctiveness of node features." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.357, + 0.893, + 0.374 + ], + "angle": 0, + "content": "3.3. Multiscale Sign Graph Convolutional Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.893, + 0.485 + ], + "angle": 0, + "content": "Patches with a fixed window may not capture sign features in one region effectively (e.g., hand regions may be split by patches). Thus, similar to Swin Transformer [25], we propose a multiscale SignGraph, which builds our model in multiscale manners to ensure that the representation of nodes can capture cross-region features at different granularities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.624 + ], + "angle": 0, + "content": "Patch Merging. To produce a multiscale representation, we further design a patch merging module to reduce the number of patches (nodes) and expand the receptive field of patches. Specifically, for \\(N\\) patches (or nodes) with the size of \\(16 \\times 16\\) in frame \\(f_{i}\\), the patch merging module applies convolutional layers to downsample frames by a factor of 2. Then we get a set of nodes \\(\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}\\) for frame \\(i\\), and its corresponding node features \\(\\{\\bar{\\mu}_{ij}\\}_{j=1}^{N/4}\\) where \\(\\bar{\\mu}_{ij} \\in \\mathbb{R}^{2D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.766 + ], + "angle": 0, + "content": "Multiscale Sign Graph Learning. After getting the set of nodes \\(\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}\\) with a larger size for frame \\(i\\) by patch merging, we adopt the another LSG module with \\(\\mathsf{K}_l\\) edges for each node in one frame and TSG module with \\(\\mathsf{K}_t\\) edges for two adjacent frames. For convenience, we name the ith LSG, TSG module as LSG\\(_i\\), TSG\\(_i\\), and the corresponding hyperparameters as \\(\\mathsf{K}_l^i\\), \\(\\mathsf{K}_t^i\\). According to experimental results in Section 4.3, we adopt two LSG, TSG modules in our baseline settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.634, + 0.795 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.603, + 0.817 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Following the previous work [16, 36], we evaluate our model on three publicly available SL datasets. (1) PHOENIX14 [23] is a widely used German SL dataset for CSLR with a vocabulary of 1295 glosses from 9 signers. It contains 5672, 540, and 629 weather forecast sam" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13473" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "plies for training, validation, and testing respectively. (2) PHOENIX14T [3] is another German SL dataset with both gloss annotations and translation annotations. It contains 7096, 519, and 642 samples from 9 signers for training, validation, and testing respectively, and it has a vocabulary of 1066 different signs for CSLR. (3) CSL-Daily is a Chinese SL dataset, which contains 18401, 1077, and 1176 labeled videos from 10 signers for training, validation, and testing respectively, and it has gloss annotations with a vocabulary of 2000 glosses for CSLR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.253, + 0.279, + 0.269 + ], + "angle": 0, + "content": "4.2. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.276, + 0.468, + 0.292 + ], + "angle": 0, + "content": "Here, we describe the baseline settings for our architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.468, + 0.419 + ], + "angle": 0, + "content": "Data Preprocessing. For data preprocessing, we follow the existing work [18, 26] and apply data augmentation during training, including: resizing frames to \\(256 \\times 256\\) pixels, random cropping frames to \\(224 \\times 224\\) pixels, random horizontal flipping with a probability of 0.5, random temporal scaling (\\(\\pm 20\\%\\)). During testing, we resize frames to \\(256 \\times 256\\) pixels and use center cropping to get frames with \\(224 \\times 224\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.804 + ], + "angle": 0, + "content": "Architecture Setting. Our architecture is implemented by PyTorch 1.11. (1) Patchify stem: Considering that sign frames have a strong 2D local structure, i.e., spatially neighboring pixels are usually highly correlated, we do not adopt non-overlapping patches, which are obtained by linear projection [9] and lack modeling power for 2D local spatial context [32, 33]. Instead, we obtain initial patches (nodes) embedding by using convolutional-based patchify stem. Specifically, we adopt the first four stages of ResNet18 as our patchify stem. The dimension of initial patches \\( D \\) is set to 512. (2) Patch merging: Similar to the patchify stem, we adopt two convolutional blocks with kernel size \\( 3 \\times 3 \\) to downsample frames and get different-size patches after LSG and TSG modules. (3) Sign Graph learning: There are two LSG and TSG modules in our baseline settings, and the initial \\( \\mathsf{K}_l^1 \\), \\( \\mathsf{K}_l^2 \\) in LSG modules are set to 4 and initial \\( \\mathsf{K}_t^1 \\), \\( \\mathsf{K}_t^2 \\) in TSG modules are set to 49 (49 is the number of nodes in \\( TSG_2 \\)). We adopt the implementation of GCN-Conv [22] as initial graph convolutional layer (GCN layer). (4) Distance function: We adopt Euclidean distance to measure the distance between two nodes in the baseline setting. (5) Global feature module: The Global feature module is the combination of two 1D convolution blocks, 2-layer BiLSTM with hidden size 1024 for global feature modeling, and a fully connected layer for final prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Training Setting. For a fair comparison, we adopt the same training settings as in previous work [26]. We adopt the Adam optimizer with a weight decay of 0.0001 to train our model for 40 epochs on 1 GeForce RTX 3090 GPU. The initial learning rate is 0.0001 with a decay factor of 0.5 at epoch 20 and 30, and the batch size is set to 4. CTC loss" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.089, + 0.885, + 0.282 + ], + "angle": 0, + "content": "
\\( LSG_1 \\)\\( TSG_1 \\)\\( LSG_2 \\)\\( TSG_2 \\)DevTest
WERDel/InsWERDel/ins
✘✘✘✘22.38.4/2.522.28.1/2.7
✓19.25.6/2.221.04.8/2.3
✓19.65.6/2.121.55.1/2.5
✓19.35.3/2.120.85.8/2.0
✓19.56.6/1.721.25.4/2.4
✓✓18.45.4/1.620.15.1/2.1
✓✓19.15.6/1.820.84.8/2.2
✓✓18.75.1 /2.320.65.2/1.7
✓✓18.64.3 /1.820.25.5/1.7
✓✓✓✓18.15.4/1.420.16.1/1.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.285, + 0.838, + 0.299 + ], + "angle": 0, + "content": "Table 1. Ablation study on Phoenix14T dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.344 + ], + "angle": 0, + "content": "is applied after both 1D convolution blocks and the fully connected layer as loss functions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Evaluation setting. To evaluate our model, we adopt the Word Error Rate (WER) as CSLR performance metric," + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.397, + 0.892, + 0.426 + ], + "angle": 0, + "content": "\\[\nW E R = \\frac {\\# S + \\# I + \\# D}{\\# N} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.43, + 0.892, + 0.492 + ], + "angle": 0, + "content": "where \\(\\# S, \\# I, \\# D\\) denote the minimum number of substitution, insertion and deletion operations needed to convert a predicted gloss sequence to the ground truth. \\(\\# N\\) is the number of glosses in ground truth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.502, + 0.651, + 0.518 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Following the previous work [3], we perform ablation studies on Phoenix14T dataset to verify the effectiveness of the proposed SignGraph." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Effects of Proposed Sign Graph Module. As shown in Table 1, results demonstrate that both the proposed \\( LSG \\) and \\( TSG \\) modules contribute to higher performance for CSLR. Specifically, our baseline with ResNet18 as backbone achieves \\( 22.3\\% \\) WER on Phoenix14T dev set. When we replace backbone with the designed local sign graph module \\( LSG_{1} \\), \\( LSG_{2} \\), WER on dev set decreases by \\( 3.1\\% \\) and \\( 3.0\\% \\) respectively. When we replace backbone with designed temporal sign graph module \\( TSG_{1} \\), \\( TSG_{2} \\), WER decreases by \\( 2.7\\% \\) and \\( 2.8\\% \\) respectively, which verifies that both intra-frame cross-region features and inter-frame cross-region features are crucial for learning sign-related features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Then we combine \\( LSG \\) and \\( TSG \\) modules, WER decreases by \\( 3.9\\% \\) when applying \\( LSG_{1} \\) and \\( TSG_{1} \\), \\( 3.2\\% \\) when applying \\( LSG_{2} \\) and \\( TSG_{2} \\) respectively, which demonstrates that combination of cross-region features in one frame and those among adjacent frames are beneficial for identifying signs. When adopting two stages of \\( LSG \\) and \\( TSG \\), performance can achieve further improvement (i.e., WER decreases by \\( 4.2\\% / 2.1\\% \\) on dev/test set)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13474" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.438, + 0.193 + ], + "angle": 0, + "content": "
\\( LSG_1 \\)\\( TSG_1 \\)DevTest
WERDel/InsWERDel/ins
DenseSparse18.45.4/1.620.15.1/2.1
DenseDense20.24.9/2.921.85.5/2.6
SparseSparse19.15.0/1.921.75.3/2.7
SparseDense20.95.4/2.421.96.1/2.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.196, + 0.393, + 0.21 + ], + "angle": 0, + "content": "Table 2. Ablation study on graph types." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.47, + 0.253 + ], + "angle": 0, + "content": "It shows that leveraging multiscale representation of nodes can capture cross-region features at different granularities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.471, + 0.471 + ], + "angle": 0, + "content": "Effect of Graph Type. In SignGraph, we adopt dense graphs (i.e., connecting top \\( K \\) neighbors for each node) in the LSG module, while adopting sparse graphs (only connecting top \\( K \\) node pairs) in the TSG module. To verify the effectiveness of our design, we test the effect of different combinations of graph types in LSG and TSG modules. For convenience, we only adopt one LSG module and one TSG module. As shown in Table 2, when we adopt the dense graph in TSG module or adopt the sparse graph in LSG module, there is a noticeable performance drop (i.e., WER on dev set rises by \\( 1.8\\% \\) or \\( 0.7\\% \\)). Our module with dense graph in LSG and sparse graph in TSG achieves the best recognition performance, which demonstrates the effectiveness of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.661 + ], + "angle": 0, + "content": "Effects of Hyperparameters K. In the proposed multiscale SignGraph, there are four hyperparameters, i.e., \\(\\mathsf{K}_l^1\\), \\(\\mathsf{K}_l^2\\) in LSG modules and \\(\\mathsf{K}_t^1\\), \\(\\mathsf{K}_t^2\\) in TSG modules. Here, to set proper hyperparameters K for our model, we test recognition performance under different \\(\\mathsf{K}_l\\) ranging ranging from 2 to 9 and under different \\(\\mathsf{K}_t\\) ranging from 7 to 91. Since there are many combinations of these hyperparameters, it is impossible to enumerate all possible values to find the globally optimal setting, thus we fix the other 3 parameters and adjust one to obtain suitable parameter settings. According to Figure 5, we set \\(\\mathsf{K}_l^1\\), \\(\\mathsf{K}_l^2\\), \\(\\mathsf{K}_t^1\\), \\(\\mathsf{K}_t^2\\) to 3, 4, 49, 49 respectively based on WER performance on DEV set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.804 + ], + "angle": 0, + "content": "Effects of Different Backbones. To verify the effectiveness of the proposed SignGraph model, we replace our backbone with other patch-based networks (i.e., CvT [33], Swin transformer [25] and VIG [15]). According to Table 3a, simply adopting the current patch-based SOTA backbones will not gain appealing performance on CSLR task, while our proposed SignGraph backbone achieves satisfactory performance by effectively capturing intra-frame cross-region features and inter-frame cross-region features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effects of Distance Function. In the SignGraph, we adopt KNN algorithm to find the nearest neighbors based on the distance of two nodes. Here, we show performances with commonly used distance functions including Cosine distance, Chebyshev distance and Euclidean distance in Table 3b. There are subtle performance gaps between different" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.691, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.09, + 0.882, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.194, + 0.689, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.194, + 0.88, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.301, + 0.788, + 0.315 + ], + "angle": 0, + "content": "Figure 5. Effects of \\(\\mathsf{K}_l\\) and \\(\\mathsf{K}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.329, + 0.892, + 0.36 + ], + "angle": 0, + "content": "distance functions and we adopt Euclidean distance in our model for better performances." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Effects of GCN Layer. We also test the representative variants of graph convolution, including GATv2Conv [1], SAGEConv [14], GCNConv [22] and EdgeConv [31]. From Table 3c, we can see that our model can achieve excellent performances with different GCN layer, indicating the flexibility of SignGraph model. Among them, EdgeConv achieves the best recognition performance (i.e., \\(17.8\\%\\) WER). In the rest of the experiments, we use EdgeConv layer in our module by default unless specifically stated." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Effect of Patch Sizes. Patches with different sizes can capture different local features and further affect model performance. To test the effect of patch sizes, we adopt only one LSG module and one TSG module. As shown in Table 3d, model with a smaller patch size (i.e., \\(8 \\times 8\\)) achieves the worst performance, while model with too large patch size may also degenerate local features." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.633, + 0.893, + 0.83 + ], + "angle": 0, + "content": "Effect of Multiscale SignGraph. To leverage the scale-invariant property of images [15], PyVIG adopts pyramid architecture that gradually increases patch size from 4 to 32, by gradually shrinking the spatial size of feature maps. Similarly, we also gradually add \\( LSG \\) and \\( TSG \\) modules in the early stage of the patchify stem and add patch merging module at the end of each \\( TSG \\) module for downsampling. In regard to the added LSG and TSG modules in each stage, we also choose appropriate hyperparameters \\( K_{l}, K_{t} \\) for them through extensive experiments. As shown in Table 3e, adding more \\( LSG \\) and \\( TSG \\) modules does not always bring more performance gain, thus we adopt two \\( LSG \\) and \\( TSG \\) modules that increase patch size from 16 to 32 in our model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of DropEdge. DropEdge [29] tackles oversmoothing and over-fitting problems for dense graph by randomly removing a certain number of edges from the input graph at each training epoch. Considering that we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13475" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.078, + 0.308, + 0.146 + ], + "angle": 0, + "content": "
BackBoneWERDelIns
CvT [32]39.215.30.9
SwinT [25]45.416.31.3
PyVIG [15]35.412.11.3
Ours17.95.11.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.149, + 0.329, + 0.172 + ], + "angle": 0, + "content": "(a) Backbone. Comparison of different backbones." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.176, + 0.307, + 0.229 + ], + "angle": 0, + "content": "
PatchSizeWERDelIns
819.14.82.6
1618.45.41.6
3218.75.71.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.233, + 0.33, + 0.257 + ], + "angle": 0, + "content": "(d) Patch size. Comparison of different patch sizes with one LSG and TSG modules." + }, + { + "type": "table", + "bbox": [ + 0.371, + 0.078, + 0.603, + 0.133 + ], + "angle": 0, + "content": "
DistanceWERDelIns
Cosine18.05.01.5
Chebyshev18.24.81.6
Euclidean17.95.11.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.148, + 0.608, + 0.172 + ], + "angle": 0, + "content": "(b) Distance function. Comparison of distance functions." + }, + { + "type": "table", + "bbox": [ + 0.375, + 0.175, + 0.601, + 0.23 + ], + "angle": 0, + "content": "
StagesWERDelDel
16→3217.85.01.6
8→16→3218.45.31.7
4→8→16→3218.14.71.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.233, + 0.608, + 0.258 + ], + "angle": 0, + "content": "(e) Multiscale SignGraph. Effect of the number of stages in multiscale SignGraph." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.078, + 0.898, + 0.146 + ], + "angle": 0, + "content": "
GraphConvWERDelIns
GATv2Conv [1]18.65.21.8
SAGEConv [14]18.04.81.7
GCNConv [22]17.95.11.5
EdgeConv [31]17.85.01.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.641, + 0.149, + 0.888, + 0.173 + ], + "angle": 0, + "content": "(c) Graph Convolution Effect of different GCN layers." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.175, + 0.899, + 0.229 + ], + "angle": 0, + "content": "
DropRateWERDelIns
017.85.01.6
15%19.26.71.3
30%21.35.91.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.233, + 0.888, + 0.258 + ], + "angle": 0, + "content": "(f) Drop edge. Adding DropEdge [29] in the local graph does not improve performance." + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.266, + 0.698, + 0.28 + ], + "angle": 0, + "content": "Table 3. Ablation experiments of SignGraph on Phoenix14T dev set." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.293, + 0.837, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.562, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Figure 6. Visualization of graph construction of \\( LSG_{1} \\) and \\( TSG_{1} \\) in the first row, \\( LSG_{2} \\) and \\( TSG_{2} \\) in the second row. The graph in \\( LSG \\) module is shown in yellow, and the graph in \\( TSG \\) is shown in blue. We also show some 'unimportant' edges between nodes in the background with red color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.722 + ], + "angle": 0, + "content": "build dense graphs for \\( LSG \\) module, we also add DropEdge with different drop rates to \\( LSG \\) module and show the recognition performance in Table 3f. However, adding DropEdge does not improve our model performance, which means that our model can achieve excellent performance without relying on external modules, and can handle over-smoothing and over-fitting problems well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Visualization of SignGraph. To verify whether our model can effectively capture sign-related features, we select a sign video from Phoenix14T test set and visualize the constructed graph structure of \\( LSG \\) and \\( TSG \\) modules in both two stages. In Figure 6, we show the graphs of two stages and we only show part of the edges for a better display. In the \\( LSG \\) module, our model can link the nodes with similar content and semantic representation (e.g., hand regions and face regions) for extracting better intra-frame cross-region features. In the \\( TSG \\) module, edges can be built for tracking dynamic changes in gestures and facial expres" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.737 + ], + "angle": 0, + "content": "sions, thus can better identify a sign with inter-frame cross-region features. We also show some 'unimportant' edges between nodes in the background with red color. As can be seen, background nodes in \\( LSG \\) are naturally connected to their neighboring nodes, and there are still a few background nodes connected in \\( TSG \\). Fortunately, nodes in the background will not 'disturb' nodes in sign-related regions, which demonstrates the effectiveness of our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.64, + 0.771 + ], + "angle": 0, + "content": "4.4. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on Phoenix14T dataset. As shown in Table 4, we compare our model with existing models on CSLR performance, and we provide both the performances on the validation set (i.e., 'DEV') and test set (i.e., 'TEST'). Most of the current models adopt existing CNN-based backbones, and achieve convincing performances by injecting extra cues [7, 35, 38], adding extra constraints [10, 16, 26] or introducing attention mechanism [18]. The GCN-based" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13476" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.087, + 0.872, + 0.315 + ], + "angle": 0, + "content": "
ModelBackboneExtra cuesPhoenixPhoenixT
F/MHSPDEVTESTDEVTEST
WERdel/insWERdel/insWERWER
STMC [35] (AAAI'20)VGG11✓✓✓21.17.7/3.420.77.4/2.619.621.0
C2SLR [38](CVPR'22)ResNet18✓20.5-/-20.4-/-20.220.4
TwoStream [7](NeurIPS'22)S3D✓✓✓✓*18.4-/-18.8-/-17.719.3
SLT [5] (CVPR'20)GooleNet✓----24.624.5
TwoStream [7] (NeurIPS'22)S3D✓*22.4-/-23.3-/-21.122.4
VAC [26] (ICCV'21)ResNet18✓21.27.9/2.522.38.4/2.6--
SMKD [16] (ICCV'21)ResNet18✓20.86.8/2.521.06.3/2.320.822.4
CorrNet [18] (CVPR'23)ResNet18✓18.85.6/2.819.45.7/2.318.920.5
FCN [8] (ECCV'20)customed23.7-/-23.9-/-
Contrastive [10] (IJCAI'23)ResNet1819.65.1/2.719.85.8/3.020.020.1
HST-GNN [21] (WACV'22)Customized(CNN+GCN)✓✓✓✓19.5-/-19.8-/-19.519.8
CoSign [20] (ICCV'23)ST-GCN(GCN)✓✓✓19.7-/-20.1-/-19.520.1
SignGraphCustomized(GCN)✓18.45.6/1.820.15.4/2.118.320.0
MultiSignGraphCustomized(GCN)✓18.24.9/2.019.15.3/1.917.819.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.319, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Table 4. Comparison of CSLR performance on Phoenix14 and Phoenix14T datasets. (F: face, M: mouth, H: hands, S: skeleton, P: pretraining backbone with ImageNet, \\(\\sqrt{}^{*}\\): pretraining on other SL-related datasets.)" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.36, + 0.466, + 0.475 + ], + "angle": 0, + "content": "
ModelBackboneExtra cuesDEVTEST
SPWERdel/insWERdel/ins
SLT [5](CVPR'20)GoogleNet✓33.110.3/4.432.09.6/4.1
TwoStream [7](NeurIPS'22)S3D✓✓*25.4-/-25.3-/-
TwoStream [7](NeurIPS'22)S3D✓*28.9-/-28.5-/-
BN-TIN [36](CVPR21)GoogLeNet✓33.613.9/3.433.113.5/3.0
CorrNet [18](CVPR'23)ResNet18✓30.6-/-30.1-/-
CoSign [20] (ICCV'23)ST-GCN(GCN)✓28.1-/-27.2-/-
SignGraphcustomized(GCN)✓28.48.8/2.427.48.2/2.1
MultiSignGraphcustomized(GCN)✓27.37.9/2.326.47.8/2.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.48, + 0.469, + 0.495 + ], + "angle": 0, + "content": "Table 5. Comparison of CSLR performance on CSL-daily dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.716 + ], + "angle": 0, + "content": "model CoSign [20] mainly relies on pre-processed fine-grained skeleton data, and achieves \\(19.5\\%\\), \\(20.1\\%\\) WER on dev, test set. While HST-GCN [21] adopts both CNN-based backbone and GCN-based backbone to extract RGB features and skeleton features respectively, achieving \\(19.5\\%\\), \\(19.8\\%\\) WER on dev, test set. It is worth mentioning that the state-of-the-art (SOTA) model TwoStream utilizes both RGB features and fine-grained skeleton features (i.e., keypoints in hands, body and face), and pretrains its backbone on other SL-related datasets, achieving \\(17.7\\%\\), \\(19.3\\%\\) WER on dev, test set respectively. Without using any extract cues, our best model still achieves comparable performance with TwoStream on dev set. While on test set, our model even outperforms TwoStream model by \\(0.2\\%\\) WER." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Evaluation on Phoenix14 dataset. We also compare our model with existing models of CSLR performance on Phoenix14 dataset. As shown in Table 4, with the simple yet powerful multiscale SignGraph, our model can achieve the best performance (i.e., \\(18.2\\%\\) WER) on dev set and a comparable performance with SOTA model on test set, which demonstrates the effectiveness of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Evaluation on CSL-daily dataset. We also provide CSLR performance comparisons with other methods on CSL-daily dataset. As shown in Table 5, the WER of our model on dev and test set is \\(27.3\\%\\) and \\(26.4\\%\\), respectively. Our model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.363, + 0.892, + 0.394 + ], + "angle": 0, + "content": "outperforms most existing models, and achieves a comparable performance with TwoStream model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.619, + 0.436 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.892, + 0.724 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet effective SignGraph architecture, which represents a sign sequence as a graph structure to extract sign-related features at the graph level. Specifically, to learn the correlation of cross-region features in one frame, we propose a Local Sign Graph \\( LSG \\) module, which dynamically builds edges between nodes in one frame and aggregates intra-frame cross-region features. To track the interaction of cross-region features among adjacent frames, we propose a Temporal Sign Graph \\( TSG \\) module, which dynamically builds edges between nodes among adjacent frames and captures inter-frame cross-region features. Besides, we also build our model in multiscale manners with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities. The extensive experiments on public datasets demonstrate the superiority of the proposed SignGraph, which achieves comparable performances with the SOTA model without using any extra cues." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.751, + 0.681, + 0.768 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by National Key Research and Development Program of China under Grant No. 2022YFB3303900; Jiangsu Provincial Key Research and Development Program under Grant BE2020001-4; National Natural Science Foundation of China under Grant Nos. 62172208, 62272216. This work is partially supported by Collaborative Innovation Center of Novel Software Technology and Industrialization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13477" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In ICCV, pages 3075-3084. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, pages 7784-7793, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In ECCV, pages 301-319. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[5] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, pages 10023-10033, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.487 + ], + "angle": 0, + "content": "[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 35:17043-17056, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.531 + ], + "angle": 0, + "content": "[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Kang Xia, Lei Xie, and Sanglu Lu. Contrastive learning for sign language recognition and translation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 763-772. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[11] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Lei Xie, and Sanglu Lu. Skeleton-aware neural sign language translation. In MM, pages 4353-4361, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[13] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In ICML, pages 369-376, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In ICCV, pages 11303-11312, 2021. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[18] Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Continuous sign language recognition with correlation network. In CVPR, pages 2529-2539, 2023. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[19] Wenbing Huang, Yu Rong, Tingyang Xu, Fuchun Sun, and Junzhou Huang. Tackling over-smoothing for general graph convolutional networks. arXiv preprint arXiv:2008.09864, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[20] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[22] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. CVIU, 141:108-125, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In CVPR, pages 4297-4305, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[26] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In ICCV, pages 11542-11551, 2021. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[27] Maria Parelli, Katerina Papadimitriou, Gerasimos Potamianos, Georgios Pavlakos, and Petros Maragos. Spatiotemporal graph convolutional networks for continuous sign language recognition. In ICASSP. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[28] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In MM, pages 1497-1505, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In International Conference on Learning Representations, 2020. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13478" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[30] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. TMM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[31] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[32] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[33] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[34] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[35] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, 2020. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[36] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, pages 1316–1325, 2021. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[37] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[38] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13479" + } + ] +] \ No newline at end of file diff --git a/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_origin.pdf b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c8047ec1c0ff934da259692302243916feacf8ff --- /dev/null +++ b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/5f8bdf16-8e1b-41c3-9a76-ee437197dbea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24c2a940657044b047ab00b22c0a4282ccfa86134f9c88b07d5053386fd11c87 +size 1521319 diff --git a/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/full.md b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4fdfe1c54d17ad7c65d967eb18ff04ecf966ab8b --- /dev/null +++ b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/full.md @@ -0,0 +1,306 @@ +# SignGraph: A Sign Sequence is Worth Graphs of Nodes + +Shiwei Gan† Yafeng Yin†* Zhiwei Jiang† Hongkai Wen‡ Lei Xie† Sanglu Lu† + +† State Key Laboratory for Novel Software Technology, Nanjing University, China + +$\ddagger$ Department of Computer Science, The University of Warwick, UK + +sw@smail.nju.edu.cn {yafeng, jzw, lxie, sanglu}@nju.edu.cn hongkai.wen@warwick.ac.uk + +# Abstract + +Despite the recent success of sign language research, the widely adopted CNN-based backbones are mainly migrated from other computer vision tasks, in which the contours and texture of objects are crucial for identifying objects. They usually treat sign frames as grids and may fail to capture effective cross-region features. In fact, sign language tasks need to focus on the correlation of different regions in one frame and the interaction of different regions among adjacent frames for identifying a sign sequence. In this paper, we propose to represent a sign sequence as graphs and introduce a simple yet effective graph-based sign language processing architecture named SignGraph, to extract cross-region features at the graph level. SignGraph consists of two basic modules: Local Sign Graph (LSG) module for learning the correlation of intra-frame cross-region features in one frame and Temporal Sign Graph (TSG) module for tracking the interaction of inter-frame cross-region features among adjacent frames. With LSG and TSG, we build our model in a multiscale manner to ensure that the representation of nodes can capture cross-region features at different granularities. Extensive experiments on current public sign language datasets demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues. Code and models are available at: https://github.com/gswycf/SignGraph. + +# 1. Introduction + +Computer vision technology and natural language processing technology have greatly advanced sign language (SL) research, including Sign Language Recognition (SLR) and Sign Language Translation (SLT). The former task aims at recognizing an isolated sign/continuous signs as a corresponding gloss or gloss sequence, while the latter task aims at translating continuous signs into spoken language [3]. In the modern sign language processing models (SLR models and SLT models), one of the key points of these deep learn + +![](images/028e330c5cb10bd0a08b0e6b2bb532a6b1bc9f1021d3bb53a297117ab6c920d0.jpg) +Figure 1. Comparison of feature extraction between convolutional network and our graph convolutional network. + +![](images/5471e091b92d6d53f96f20954b39fa1e7c4f81793988af4e7a0c10d470d6af87.jpg) + +![](images/96a2f22030b91665634dfd0a25b05b452c2a74ff0817dc153247813e291942db.jpg) + +ing based methods is to learn sign-related features through training with extensive sign video data. As a matter of course, 2D CNNs and 3D CNNs are used to be the defacto standard visual backbones [7, 10] for extracting local sign frame features, while 1D CNN, LSTM and Transformer [3, 11, 37] are used as the temporal module to capture dynamic changes of sign frames. + +However, as shown in Figure 1, the widely-used CNN-based backbones that are mainly migrated from image recognition and object detection tasks, treat the sign frame/sign clips as grids and adopt 'kernels' to extract local features by sliding across sign frames. These CNNs are good at capturing contour-based representation and texture-based representation [12], but they may fail to capture the explicit collaboration of signs in different regions. In fact, sign language tasks need to focus on manual and non-manual features, especially the collaboration of these cues in different regions [35]. Besides, temporal modules like LSTM and Transformer layer usually regard a sign frame as a whole region and learn the overall variance of whole frames, while ignoring the interaction of subregions in adjacent frames. Other temporal methods like 1D CNN can only focus on the same regions in adjacent frames, which may fail to track dynamic motions across different regions. In fact, the dynamic movements in the body, hands, and face which are reflected in the same/different regions between adjacent frames, are primary elements of identifying a sign for sign language tasks. + +In order to make up for the shortcomings of typical CNN backbones explicitly or implicitly and allow the model to learn robust sign-related features, exiting work committed to designing various CNN-based backbones by injecting do + +main knowledge i.e., skeleton [11], depth images [30] or local areas [37]. Besides, some work trained their backbones with more training samples by back translation [36], cross modality augmentation [28] and pre-training models [6], while others added extra constraints on backbones by using knowledge distillation [16], contrastive learning [10]. + +Different from previous work, we build a simple yet effective sign graph neural network (SignGraph for short) for continuous sign language recognition (CSLR) task, in which we take sign frames as graphs of nodes and build dynamic graphs to learn cross-region features. The main intuitions are: (1) Sign-related features involve the correlation of different regions in one frame, such as face region and hand regions. Thus, we propose a local sign graph $LSG$ module by dynamically building graphs based on nodes in each frame, to learn intra-frame cross-region features. (2) The dynamic movements in the body, hands, and face are reflected in different regions among adjacent frames. Thus we propose a temporal sign graph $TSG$ module by dynamically connecting sign-related regions based on nodes in adjacent frames, to learn inter-frame cross-region features. (3) Considering that patches with a single size may not capture local features effectively, we build our SignGraph in a multiscale manner by merging small patches into larger-size patches, and use different-size patches to capture cross-region features at different granularities. We make the following contributions: + +- Different from CNN-based SL models, we propose a graph-based CSLR model (SignGraph), which processes sign frames as a set of nodes and dynamically builds graphs from one frame and adjacent frames to capture cross-region features related to signs, thus can aggregate sign-related features from different regions for better feature representation. +- We propose a new local sign graph module $LSG$ that dynamically builds edges based on node features to learn the correlation of cross-region features in one frame, and a new temporal sign graph module $TSG$ that dynamically connects sign-related regions to learn the interaction of cross-region features among adjacent frames. +- We build our model in a multiscale manner with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities. +- The extensive experiments on CSLR task demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues. + +# 2. Related Work + +We mainly review CSLR technologies and graph convolutional networks in sign language. + +Continuous Sign Language Recognition. The goal of CSLR is to interpret the continuous sign sequence as corresponding gloss sequence [24], thus the core of current CSLR tasks is to identify signs. The widely adopted CNN-based backbones in CSLR which are migrated from the image recognition task, are good at capturing the texture and contours of objects [12]. However, identifying a sign needs to focus on the correlation of different regions (e.g., hand regions and face region) in one frame and the interaction of the same or different regions (e.g., hand motions, facial expression changes) among adjacent frames. As a matter of course, to extract robust sign-related features with CNN-based backbones, some sign language models introduced prior knowledge (or expert knowledge) to guide the model to focus on sign-related features. For example, extra information like local areas [2, 4], skeletons [4, 7, 38] and depth images [30] was manually injected into model. However, in these models, on the one hand, to obtain such extra knowledge, additional algorithms need to be introduced, which will cause additional computational overhead. On the other hand, integrating such knowledge will lead to a complex network structure. Besides, some work turned its attention to obtaining more sign language samples by back translation [36], cross modality augmentation [28], and pretraining with other sign language related datasets [6, 7]. While others were dedicated to designing complex backbones with attention mechanism [18] or adding extra constraints on CNN backbones by adopting knowledge distillation [16, 26], contrastive learning [10]. + +Graph Convolutional Network. Compared with sequences and grids, graphs are more flexible data structures that model the relationships of a set of objects with nodes and edges. In order to learn representation from the graph, graph convolutional network (GCN) [22] was proposed, and it has been widely applied in action recognition tasks [34] and has also been extended to sign language tasks [11, 17, 21]. To apply GCN to SL tasks, the existing models often relied on extracted skeleton, whose connections between joints make skeleton data naturally fit GCN. Specifically, some work utilized GCN to guide CNN-based backbones to extract skeleton-related features [11, 27]. While others ignored RGB features and applied GCN to explore sign features directly with skeleton [20, 21]. Recently, in addition to applying GCN with skeletons, the usage of applying GCN on images emerged. Specifically, VIG [15] represented an image as a graph structure to capture irregular and complex objects, and demonstrated the effectiveness of graph representation of images. Inspired by VIG, we find that cross-region features are important for identifying a sign and can be captured by graphs. Thus unlike the previous work which adopted CNN-based backbones or relied on skeletons to apply GCN, we directly treat sign frames as graphs of nodes and design a simple GCN-based framework + +![](images/d1360906ec4fa4240be220f9e174fedc1c399cb3a2fd003ad726e89ffc757af1.jpg) +Figure 2. The proposed SignGraph architecture. + +for learning sign-related features. + +# 3. Method + +In this section, we first give an overview of the proposed graph-based sign language processing architecture. Then we describe the proposed Local Sign Graph $LSG$ module for capturing intra-frame cross-region features and Temporal Sign Graph $TSG$ module for capturing inter-frame cross-region features. Finally, with $LSG$ and $TSG$ , we describe our multiscale SignGraph for CSLR. + +# 3.1. Overall Framework + +For a sign language video $f = \{f_i\}_{i=1}^{\theta}$ with $\theta$ frames, the target of CSLR model is to get a recognized gloss sequence $g = \{g_i\}_{i=1}^{\vartheta}$ with $\vartheta$ glosses based on input $f$ . + +As shown in Figure 2, our SignGraph model consists of two key modules, i.e. Local Sign Graph $LSG$ module and Temporal Sign Graph $TSG$ module. First, we convert each frame $f_{i}$ into $N$ patches with a patchify stem. Then each frame $f_{i}$ is represented by a set of patches (i.e., nodes) $v_{i} = \{v_{ij}\}_{j=1}^{N}$ . Second, by learning the correlation between each node and connecting $K$ nearest neighbors for each node, we build a dynamic local graph $G_{i}^{l}$ for each frame $f_{i}$ . Then a local graph convolutional layer is leveraged to capture cross-region features within each frame. Third, by connecting sign-related regions between two adjacent frames, we build a dynamic temporal graph $G_{i}^{t}$ for adjacent frames $f_{i}$ and $f_{i+1}$ . Then, a temporal graph convolutional layer is adopted to track the interaction of cross-region features among adjacent frames. + +Further, we merge patches into larger size patches by downsampling frames with a factor of 2, then each frame $f_{i}$ can be represented by another set of patches (nodes) $\{\bar{v}_{ij}\}_{j=1}^{N/4}$ . Following that, we repeatedly adopt LSG module and TSG module to learn cross-region features with different-size patches. After that, following previous sign language models [10, 26], we adopt a global feature module + +to learn global changes of whole frames. Finally, a classifier and a widely-used CTC loss [13] are adopted to predict the probability $p(g|f)$ of target gloss sequence. + +# 3.2. Sign Graph Learning + +Patchify Stem. In order to build graphs based on sign frames instead of skeleton data, we need to convert RGB data into nodes. Following previous work [9, 32], for each frame $f_{i} \in \mathbb{R}^{[H \times W \times 3]}$ with height $H$ and width $W$ , we first convert it into a number of patches (or nodes) $v_{i} = \{v_{ij}\}_{j=1}^{N}$ with a patchify stem $\mathcal{PS}$ . + +$$ +\mu_ {i} = \left\{\mu_ {i j} \right\} _ {j = 1} ^ {N} = \mathcal {P S} (f _ {i}) \tag {1} +$$ + +Here, $v_{ij}$ represents the $j$ th node in the $i$ -th frame $f_i$ , $\mu_{ij} \in \mathbb{R}^D$ denotes node features of $v_{ij}$ , $D$ is the feature dimension, $N = HW / P^2$ is the number of patches, and $P$ is the patch size. + +Local Sign Graph Learning. For a single frame in the sign video, sign language models should focus on the correlation of different regions, such as face region and hand regions. In order to learn intra-frame cross-region features, we need to build a graph for frame $f_{i}$ , in which regions with correlation (e.g., hand regions and face region) should be connected. + +As shown in Figure 3, for a set of nodes $v_{i}$ , we first adopt a linear projection layer with learnable weights $W_{1}^{l}$ to map node features $\mu_{i}$ to the space where the distance function is applied: $\mu_{i}' = \mu_{i}W_{1}^{l}$ . Then we adopt $K$ -Nearest Neighbors (KNN) algorithm to find top $\mathsf{K}_l$ nearest neighbors $\mathcal{N}_{\mathsf{K}_l}(v_{ij})$ for each node based on distance function $\mathcal{D}\mathcal{I}\mathcal{S}$ [15]. + +$$ +\mathcal {D I S} \left(v _ {i j}, v _ {i k}\right) = \sqrt {\sum \left(\mu_ {i j} ^ {\prime} - \mu_ {i k} ^ {\prime}\right) ^ {2}} \tag {2} +$$ + +$$ +\mathcal {N} _ {\mathrm {K} _ {l}} \left(v _ {i j}\right) = \left\{v _ {i k} \mid v _ {i k} \in \operatorname {T o p} K \left(\left\{\mathcal {D I S} \left(v _ {i j}, v _ {i k}\right) \right\} _ {k = 1} ^ {N}\right) \right\} \tag {3} +$$ + +Here, $TopK$ function outputs the top $K$ nearest neighbors for $v_{ij}$ based on node distances. For each node $v_{ij}$ , + +![](images/c4e5cec7923a3e95cbb2a3c5b72987a0ad3e17c8516918f6da2ba13472f69af5.jpg) +Figure 3. Graph construction in local sign graph module. + +we add an edge $e(v_{ij}, v_{ik})$ between $v_{ik}$ and $v_{ij}$ , where $v_{ik} \in \mathcal{N}_{\mathsf{K}_l}(v_{ij})$ . Thus, for frame $f_i$ , we can get the edge set $e_i^l = \{e(v_{ij}, v_{ik}), |j \in [1, N], v_{ik} \in \mathcal{N}_{\mathsf{K}_l}(v_{ij})\}$ , and build a local graph $G_i^l = \{v_i, e_i^l\}$ . After that, we adopt a graph convolutional layer $\mathcal{GCN}_l$ and a linear projection layer with learnable weights $W_2^l$ to aggregate features from cross regions for each frame. + +$$ +\mu_ {i} = \mu_ {i} + \mathcal {G C N} _ {l} \left(\mu_ {i} ^ {\prime}, e _ {i} ^ {l}\right) W _ {2} ^ {l} \tag {4} +$$ + +The LSG module contains one graph convolutional layer and two linear projection layers. We do not add extra modules (e.g., feed-forward network) since we aim to keep our model simple, and we show that our model can achieve a convincing performance even with the simple design. + +Temporal Sign Graph Learning. The dynamic movements in body, hands, and face are essential elements to identify a sign, which are often reflected in different regions among adjacent frames. Hence, we design a temporal graph module, which can dynamically build edges between regions in two adjacent frames and learn inter-frame cross-region features. + +For two adjacent frames $f_{i}$ and $f_{i+1}$ , we have $2N$ nodes $\{v_{ij}\}_{j=1}^{N}$ and $\{v_{(i+1)k}\}_{k=1}^{N}$ . Similarly, we first adopt a linear projection layer with weights $W_{1}^{t}$ to map node features $\mu_{i}$ to the space where the distance function is applied: $\mu_{i}^{\prime \prime} = \mu_{i}W_{1}^{t}$ . Then we compute distances between $v_{ij}$ and $v_{(i+1)k}$ , thus we have a $N \times N$ distance matrix $M$ , where $\{M[j,k] = \mathcal{DLS}(v_{ij},v_{(i+1)k})|j\in [1,N],k\in [1,N]\}$ . When building the temporal graph, unlike the LSG module where we build edges for each node $v_{ij}$ (i.e., building a dense graph), we only choose top $K_{t}$ node pairs between two adjacent frames and connect them (i.e., building a sparse graph). To avoid establishing edges between meaningless node pairs, we set a threshold of 0.05 on normalized distance matrix $M$ to filter out patch pairs in the background, since these pairs often have a high and constant similarity (i.e., $M[j,k] < 0.05$ ). Then we get the temporal edge set $e_{i}^{t}$ between frame $f_{i}$ and $f_{i+1}$ and temporal graph $G_{i}^{t} = \{(v_{i},v_{i+1}),e_{i}^{t}\}$ . + +$$ +e _ {i} ^ {t} = \left\{e \left(v _ {i j}, v _ {(i + 1) k}\right) \mid \mathcal {D I S} \left(v _ {i j}, v _ {(i + 1) k}\right) \in T o p K (M) \right\} \tag {5} +$$ + +After that, we adopt a temporal graph convolutional layer $\mathcal{GCN}_t$ and another linear projection $W_2^t$ to aggregate cross-region features from adjacent frames, as follows. + +![](images/d48394fb8a432a0326745c25dbc5604ceee138eb1f26c76632d773909248d2b1.jpg) +Figure 4. Graph construction in temporal sign graph module. + +$$ +\left\{\mu_ {i} \right\} _ {i = 1} ^ {\theta} = \left\{\mu_ {i} \right\} _ {i = 1} ^ {\theta} + \mathcal {G C N} _ {t} \left(\left\{\mu_ {i} ^ {\prime \prime} \right\} _ {i = 1} ^ {\theta}, \left\{e _ {i} ^ {t} \right\} _ {i = 1} ^ {\theta - 1}\right) W _ {2} ^ {t} \tag {6} +$$ + +The intuition of building the sparse graph for $TSG$ module rather than building the dense graph like $LSG$ module is to avoid the over-smoothing phenomenon [19], which may decrease the distinctiveness of node features. + +# 3.3. Multiscale Sign Graph Convolutional Network + +Patches with a fixed window may not capture sign features in one region effectively (e.g., hand regions may be split by patches). Thus, similar to Swin Transformer [25], we propose a multiscale SignGraph, which builds our model in multiscale manners to ensure that the representation of nodes can capture cross-region features at different granularities. + +Patch Merging. To produce a multiscale representation, we further design a patch merging module to reduce the number of patches (nodes) and expand the receptive field of patches. Specifically, for $N$ patches (or nodes) with the size of $16 \times 16$ in frame $f_{i}$ , the patch merging module applies convolutional layers to downsample frames by a factor of 2. Then we get a set of nodes $\{\bar{v}_{ij}\}_{j=1}^{N/4}$ for frame $i$ , and its corresponding node features $\{\bar{\mu}_{ij}\}_{j=1}^{N/4}$ where $\bar{\mu}_{ij} \in \mathbb{R}^{2D}$ . + +Multiscale Sign Graph Learning. After getting the set of nodes $\{\bar{v}_{ij}\}_{j=1}^{N/4}$ with a larger size for frame $i$ by patch merging, we adopt the another LSG module with $\mathsf{K}_l$ edges for each node in one frame and TSG module with $\mathsf{K}_t$ edges for two adjacent frames. For convenience, we name the ith LSG, TSG module as LSG $_i$ , TSG $_i$ , and the corresponding hyperparameters as $\mathsf{K}_l^i$ , $\mathsf{K}_t^i$ . According to experimental results in Section 4.3, we adopt two LSG, TSG modules in our baseline settings. + +# 4. Experiments + +# 4.1. Datasets + +Following the previous work [16, 36], we evaluate our model on three publicly available SL datasets. (1) PHOENIX14 [23] is a widely used German SL dataset for CSLR with a vocabulary of 1295 glosses from 9 signers. It contains 5672, 540, and 629 weather forecast sam + +plies for training, validation, and testing respectively. (2) PHOENIX14T [3] is another German SL dataset with both gloss annotations and translation annotations. It contains 7096, 519, and 642 samples from 9 signers for training, validation, and testing respectively, and it has a vocabulary of 1066 different signs for CSLR. (3) CSL-Daily is a Chinese SL dataset, which contains 18401, 1077, and 1176 labeled videos from 10 signers for training, validation, and testing respectively, and it has gloss annotations with a vocabulary of 2000 glosses for CSLR. + +# 4.2. Experimental Setting + +Here, we describe the baseline settings for our architecture. + +Data Preprocessing. For data preprocessing, we follow the existing work [18, 26] and apply data augmentation during training, including: resizing frames to $256 \times 256$ pixels, random cropping frames to $224 \times 224$ pixels, random horizontal flipping with a probability of 0.5, random temporal scaling ( $\pm 20\%$ ). During testing, we resize frames to $256 \times 256$ pixels and use center cropping to get frames with $224 \times 224$ pixels. + +Architecture Setting. Our architecture is implemented by PyTorch 1.11. (1) Patchify stem: Considering that sign frames have a strong 2D local structure, i.e., spatially neighboring pixels are usually highly correlated, we do not adopt non-overlapping patches, which are obtained by linear projection [9] and lack modeling power for 2D local spatial context [32, 33]. Instead, we obtain initial patches (nodes) embedding by using convolutional-based patchify stem. Specifically, we adopt the first four stages of ResNet18 as our patchify stem. The dimension of initial patches $D$ is set to 512. (2) Patch merging: Similar to the patchify stem, we adopt two convolutional blocks with kernel size $3 \times 3$ to downsample frames and get different-size patches after LSG and TSG modules. (3) Sign Graph learning: There are two LSG and TSG modules in our baseline settings, and the initial $\mathsf{K}_l^1$ , $\mathsf{K}_l^2$ in LSG modules are set to 4 and initial $\mathsf{K}_t^1$ , $\mathsf{K}_t^2$ in TSG modules are set to 49 (49 is the number of nodes in $TSG_2$ ). We adopt the implementation of GCN-Conv [22] as initial graph convolutional layer (GCN layer). (4) Distance function: We adopt Euclidean distance to measure the distance between two nodes in the baseline setting. (5) Global feature module: The Global feature module is the combination of two 1D convolution blocks, 2-layer BiLSTM with hidden size 1024 for global feature modeling, and a fully connected layer for final prediction. + +Training Setting. For a fair comparison, we adopt the same training settings as in previous work [26]. We adopt the Adam optimizer with a weight decay of 0.0001 to train our model for 40 epochs on 1 GeForce RTX 3090 GPU. The initial learning rate is 0.0001 with a decay factor of 0.5 at epoch 20 and 30, and the batch size is set to 4. CTC loss + +
\( LSG_1 \)\( TSG_1 \)\( LSG_2 \)\( TSG_2 \)DevTest
WERDel/InsWERDel/ins
✘✘✘✘22.38.4/2.522.28.1/2.7
✓19.25.6/2.221.04.8/2.3
✓19.65.6/2.121.55.1/2.5
✓19.35.3/2.120.85.8/2.0
✓19.56.6/1.721.25.4/2.4
✓✓18.45.4/1.620.15.1/2.1
✓✓19.15.6/1.820.84.8/2.2
✓✓18.75.1 /2.320.65.2/1.7
✓✓18.64.3 /1.820.25.5/1.7
✓✓✓✓18.15.4/1.420.16.1/1.8
+ +Table 1. Ablation study on Phoenix14T dataset. + +is applied after both 1D convolution blocks and the fully connected layer as loss functions. + +Evaluation setting. To evaluate our model, we adopt the Word Error Rate (WER) as CSLR performance metric, + +$$ +W E R = \frac {\# S + \# I + \# D}{\# N} \tag {7} +$$ + +where $\# S, \# I, \# D$ denote the minimum number of substitution, insertion and deletion operations needed to convert a predicted gloss sequence to the ground truth. $\# N$ is the number of glosses in ground truth. + +# 4.3. Ablation study + +Following the previous work [3], we perform ablation studies on Phoenix14T dataset to verify the effectiveness of the proposed SignGraph. + +Effects of Proposed Sign Graph Module. As shown in Table 1, results demonstrate that both the proposed $LSG$ and $TSG$ modules contribute to higher performance for CSLR. Specifically, our baseline with ResNet18 as backbone achieves $22.3\%$ WER on Phoenix14T dev set. When we replace backbone with the designed local sign graph module $LSG_{1}$ , $LSG_{2}$ , WER on dev set decreases by $3.1\%$ and $3.0\%$ respectively. When we replace backbone with designed temporal sign graph module $TSG_{1}$ , $TSG_{2}$ , WER decreases by $2.7\%$ and $2.8\%$ respectively, which verifies that both intra-frame cross-region features and inter-frame cross-region features are crucial for learning sign-related features. + +Then we combine $LSG$ and $TSG$ modules, WER decreases by $3.9\%$ when applying $LSG_{1}$ and $TSG_{1}$ , $3.2\%$ when applying $LSG_{2}$ and $TSG_{2}$ respectively, which demonstrates that combination of cross-region features in one frame and those among adjacent frames are beneficial for identifying signs. When adopting two stages of $LSG$ and $TSG$ , performance can achieve further improvement (i.e., WER decreases by $4.2\% / 2.1\%$ on dev/test set). + +
\( LSG_1 \)\( TSG_1 \)DevTest
WERDel/InsWERDel/ins
DenseSparse18.45.4/1.620.15.1/2.1
DenseDense20.24.9/2.921.85.5/2.6
SparseSparse19.15.0/1.921.75.3/2.7
SparseDense20.95.4/2.421.96.1/2.5
+ +Table 2. Ablation study on graph types. + +It shows that leveraging multiscale representation of nodes can capture cross-region features at different granularities. + +Effect of Graph Type. In SignGraph, we adopt dense graphs (i.e., connecting top $K$ neighbors for each node) in the LSG module, while adopting sparse graphs (only connecting top $K$ node pairs) in the TSG module. To verify the effectiveness of our design, we test the effect of different combinations of graph types in LSG and TSG modules. For convenience, we only adopt one LSG module and one TSG module. As shown in Table 2, when we adopt the dense graph in TSG module or adopt the sparse graph in LSG module, there is a noticeable performance drop (i.e., WER on dev set rises by $1.8\%$ or $0.7\%$ ). Our module with dense graph in LSG and sparse graph in TSG achieves the best recognition performance, which demonstrates the effectiveness of our model. + +Effects of Hyperparameters K. In the proposed multiscale SignGraph, there are four hyperparameters, i.e., $\mathsf{K}_l^1$ , $\mathsf{K}_l^2$ in LSG modules and $\mathsf{K}_t^1$ , $\mathsf{K}_t^2$ in TSG modules. Here, to set proper hyperparameters K for our model, we test recognition performance under different $\mathsf{K}_l$ ranging ranging from 2 to 9 and under different $\mathsf{K}_t$ ranging from 7 to 91. Since there are many combinations of these hyperparameters, it is impossible to enumerate all possible values to find the globally optimal setting, thus we fix the other 3 parameters and adjust one to obtain suitable parameter settings. According to Figure 5, we set $\mathsf{K}_l^1$ , $\mathsf{K}_l^2$ , $\mathsf{K}_t^1$ , $\mathsf{K}_t^2$ to 3, 4, 49, 49 respectively based on WER performance on DEV set. + +Effects of Different Backbones. To verify the effectiveness of the proposed SignGraph model, we replace our backbone with other patch-based networks (i.e., CvT [33], Swin transformer [25] and VIG [15]). According to Table 3a, simply adopting the current patch-based SOTA backbones will not gain appealing performance on CSLR task, while our proposed SignGraph backbone achieves satisfactory performance by effectively capturing intra-frame cross-region features and inter-frame cross-region features. + +Effects of Distance Function. In the SignGraph, we adopt KNN algorithm to find the nearest neighbors based on the distance of two nodes. Here, we show performances with commonly used distance functions including Cosine distance, Chebyshev distance and Euclidean distance in Table 3b. There are subtle performance gaps between different + +![](images/71a4ca41c005b0bb9a7d28c7d142b36bc76f1dcc2fe0423398ed3abd28b41721.jpg) + +![](images/1367bac6a034ff9617870c5b76b955d1040d55c540aedfde196433b7ee515148.jpg) + +![](images/bff04dab6e61bfe6e11f9ac902a6b13a4b809161366a10cdaddd062232944084.jpg) +Figure 5. Effects of $\mathsf{K}_l$ and $\mathsf{K}_t$ . + +![](images/600fa3fd1c0c550f42d3822f2f838cedc8fbb8892123a4855ea19c9086ebeca3.jpg) + +distance functions and we adopt Euclidean distance in our model for better performances. + +Effects of GCN Layer. We also test the representative variants of graph convolution, including GATv2Conv [1], SAGEConv [14], GCNConv [22] and EdgeConv [31]. From Table 3c, we can see that our model can achieve excellent performances with different GCN layer, indicating the flexibility of SignGraph model. Among them, EdgeConv achieves the best recognition performance (i.e., $17.8\%$ WER). In the rest of the experiments, we use EdgeConv layer in our module by default unless specifically stated. + +Effect of Patch Sizes. Patches with different sizes can capture different local features and further affect model performance. To test the effect of patch sizes, we adopt only one LSG module and one TSG module. As shown in Table 3d, model with a smaller patch size (i.e., $8 \times 8$ ) achieves the worst performance, while model with too large patch size may also degenerate local features. + +Effect of Multiscale SignGraph. To leverage the scale-invariant property of images [15], PyVIG adopts pyramid architecture that gradually increases patch size from 4 to 32, by gradually shrinking the spatial size of feature maps. Similarly, we also gradually add $LSG$ and $TSG$ modules in the early stage of the patchify stem and add patch merging module at the end of each $TSG$ module for downsampling. In regard to the added LSG and TSG modules in each stage, we also choose appropriate hyperparameters $K_{l}, K_{t}$ for them through extensive experiments. As shown in Table 3e, adding more $LSG$ and $TSG$ modules does not always bring more performance gain, thus we adopt two $LSG$ and $TSG$ modules that increase patch size from 16 to 32 in our model. + +Effect of DropEdge. DropEdge [29] tackles oversmoothing and over-fitting problems for dense graph by randomly removing a certain number of edges from the input graph at each training epoch. Considering that we + +
BackBoneWERDelIns
CvT [32]39.215.30.9
SwinT [25]45.416.31.3
PyVIG [15]35.412.11.3
Ours17.95.11.5
+ +(a) Backbone. Comparison of different backbones. + +
PatchSizeWERDelIns
819.14.82.6
1618.45.41.6
3218.75.71.7
+ +(d) Patch size. Comparison of different patch sizes with one LSG and TSG modules. + +
DistanceWERDelIns
Cosine18.05.01.5
Chebyshev18.24.81.6
Euclidean17.95.11.5
+ +(b) Distance function. Comparison of distance functions. + +
StagesWERDelDel
16→3217.85.01.6
8→16→3218.45.31.7
4→8→16→3218.14.71.9
+ +(e) Multiscale SignGraph. Effect of the number of stages in multiscale SignGraph. + +
GraphConvWERDelIns
GATv2Conv [1]18.65.21.8
SAGEConv [14]18.04.81.7
GCNConv [22]17.95.11.5
EdgeConv [31]17.85.01.6
+ +(c) Graph Convolution Effect of different GCN layers. + +
DropRateWERDelIns
017.85.01.6
15%19.26.71.3
30%21.35.91.7
+ +(f) Drop edge. Adding DropEdge [29] in the local graph does not improve performance. + +![](images/7196f2e33b9ae4497c2a4e47eab83a841d05d857f5d4fcfa67f7afcd834b1067.jpg) +Table 3. Ablation experiments of SignGraph on Phoenix14T dev set. +Figure 6. Visualization of graph construction of $LSG_{1}$ and $TSG_{1}$ in the first row, $LSG_{2}$ and $TSG_{2}$ in the second row. The graph in $LSG$ module is shown in yellow, and the graph in $TSG$ is shown in blue. We also show some 'unimportant' edges between nodes in the background with red color. + +build dense graphs for $LSG$ module, we also add DropEdge with different drop rates to $LSG$ module and show the recognition performance in Table 3f. However, adding DropEdge does not improve our model performance, which means that our model can achieve excellent performance without relying on external modules, and can handle over-smoothing and over-fitting problems well. + +Visualization of SignGraph. To verify whether our model can effectively capture sign-related features, we select a sign video from Phoenix14T test set and visualize the constructed graph structure of $LSG$ and $TSG$ modules in both two stages. In Figure 6, we show the graphs of two stages and we only show part of the edges for a better display. In the $LSG$ module, our model can link the nodes with similar content and semantic representation (e.g., hand regions and face regions) for extracting better intra-frame cross-region features. In the $TSG$ module, edges can be built for tracking dynamic changes in gestures and facial expres + +sions, thus can better identify a sign with inter-frame cross-region features. We also show some 'unimportant' edges between nodes in the background with red color. As can be seen, background nodes in $LSG$ are naturally connected to their neighboring nodes, and there are still a few background nodes connected in $TSG$ . Fortunately, nodes in the background will not 'disturb' nodes in sign-related regions, which demonstrates the effectiveness of our model. + +# 4.4. Comparisons + +Evaluation on Phoenix14T dataset. As shown in Table 4, we compare our model with existing models on CSLR performance, and we provide both the performances on the validation set (i.e., 'DEV') and test set (i.e., 'TEST'). Most of the current models adopt existing CNN-based backbones, and achieve convincing performances by injecting extra cues [7, 35, 38], adding extra constraints [10, 16, 26] or introducing attention mechanism [18]. The GCN-based + +
ModelBackboneExtra cuesPhoenixPhoenixT
F/MHSPDEVTESTDEVTEST
WERdel/insWERdel/insWERWER
STMC [35] (AAAI'20)VGG11✓✓✓21.17.7/3.420.77.4/2.619.621.0
C2SLR [38](CVPR'22)ResNet18✓20.5-/-20.4-/-20.220.4
TwoStream [7](NeurIPS'22)S3D✓✓✓✓*18.4-/-18.8-/-17.719.3
SLT [5] (CVPR'20)GooleNet✓----24.624.5
TwoStream [7] (NeurIPS'22)S3D✓*22.4-/-23.3-/-21.122.4
VAC [26] (ICCV'21)ResNet18✓21.27.9/2.522.38.4/2.6--
SMKD [16] (ICCV'21)ResNet18✓20.86.8/2.521.06.3/2.320.822.4
CorrNet [18] (CVPR'23)ResNet18✓18.85.6/2.819.45.7/2.318.920.5
FCN [8] (ECCV'20)customed23.7-/-23.9-/-
Contrastive [10] (IJCAI'23)ResNet1819.65.1/2.719.85.8/3.020.020.1
HST-GNN [21] (WACV'22)Customized(CNN+GCN)✓✓✓✓19.5-/-19.8-/-19.519.8
CoSign [20] (ICCV'23)ST-GCN(GCN)✓✓✓19.7-/-20.1-/-19.520.1
SignGraphCustomized(GCN)✓18.45.6/1.820.15.4/2.118.320.0
MultiSignGraphCustomized(GCN)✓18.24.9/2.019.15.3/1.917.819.1
+ +Table 4. Comparison of CSLR performance on Phoenix14 and Phoenix14T datasets. (F: face, M: mouth, H: hands, S: skeleton, P: pretraining backbone with ImageNet, $\sqrt{}^{*}$ : pretraining on other SL-related datasets.) + +
ModelBackboneExtra cuesDEVTEST
SPWERdel/insWERdel/ins
SLT [5](CVPR'20)GoogleNet✓33.110.3/4.432.09.6/4.1
TwoStream [7](NeurIPS'22)S3D✓✓*25.4-/-25.3-/-
TwoStream [7](NeurIPS'22)S3D✓*28.9-/-28.5-/-
BN-TIN [36](CVPR21)GoogLeNet✓33.613.9/3.433.113.5/3.0
CorrNet [18](CVPR'23)ResNet18✓30.6-/-30.1-/-
CoSign [20] (ICCV'23)ST-GCN(GCN)✓28.1-/-27.2-/-
SignGraphcustomized(GCN)✓28.48.8/2.427.48.2/2.1
MultiSignGraphcustomized(GCN)✓27.37.9/2.326.47.8/2.1
+ +Table 5. Comparison of CSLR performance on CSL-daily dataset. + +model CoSign [20] mainly relies on pre-processed fine-grained skeleton data, and achieves $19.5\%$ , $20.1\%$ WER on dev, test set. While HST-GCN [21] adopts both CNN-based backbone and GCN-based backbone to extract RGB features and skeleton features respectively, achieving $19.5\%$ , $19.8\%$ WER on dev, test set. It is worth mentioning that the state-of-the-art (SOTA) model TwoStream utilizes both RGB features and fine-grained skeleton features (i.e., keypoints in hands, body and face), and pretrains its backbone on other SL-related datasets, achieving $17.7\%$ , $19.3\%$ WER on dev, test set respectively. Without using any extract cues, our best model still achieves comparable performance with TwoStream on dev set. While on test set, our model even outperforms TwoStream model by $0.2\%$ WER. + +Evaluation on Phoenix14 dataset. We also compare our model with existing models of CSLR performance on Phoenix14 dataset. As shown in Table 4, with the simple yet powerful multiscale SignGraph, our model can achieve the best performance (i.e., $18.2\%$ WER) on dev set and a comparable performance with SOTA model on test set, which demonstrates the effectiveness of our model. + +Evaluation on CSL-daily dataset. We also provide CSLR performance comparisons with other methods on CSL-daily dataset. As shown in Table 5, the WER of our model on dev and test set is $27.3\%$ and $26.4\%$ , respectively. Our model + +outperforms most existing models, and achieves a comparable performance with TwoStream model. + +# 5. Conclusion + +In this paper, we propose a simple yet effective SignGraph architecture, which represents a sign sequence as a graph structure to extract sign-related features at the graph level. Specifically, to learn the correlation of cross-region features in one frame, we propose a Local Sign Graph $LSG$ module, which dynamically builds edges between nodes in one frame and aggregates intra-frame cross-region features. To track the interaction of cross-region features among adjacent frames, we propose a Temporal Sign Graph $TSG$ module, which dynamically builds edges between nodes among adjacent frames and captures inter-frame cross-region features. Besides, we also build our model in multiscale manners with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities. The extensive experiments on public datasets demonstrate the superiority of the proposed SignGraph, which achieves comparable performances with the SOTA model without using any extra cues. + +# 6. Acknowledgments + +This work is supported by National Key Research and Development Program of China under Grant No. 2022YFB3303900; Jiangsu Provincial Key Research and Development Program under Grant BE2020001-4; National Natural Science Foundation of China under Grant Nos. 62172208, 62272216. This work is partially supported by Collaborative Innovation Center of Novel Software Technology and Industrialization. + +# References + +[1] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021. 6, 7 +[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In ICCV, pages 3075-3084. IEEE, 2017. 2 +[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, pages 7784-7793, 2018. 1, 5 +[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In ECCV, pages 301-319. Springer, 2020. 2 +[5] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, pages 10023-10033, 2020. 8 +[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 2 +[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 35:17043-17056, 2022. 1, 2, 7, 8 +[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV. Springer, 2020. 8 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 5 +[10] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Kang Xia, Lei Xie, and Sanglu Lu. Contrastive learning for sign language recognition and translation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 763-772. 1, 2, 3, 7, 8 +[11] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Lei Xie, and Sanglu Lu. Skeleton-aware neural sign language translation. In MM, pages 4353-4361, 2021. 1, 2 +[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2 +[13] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In ICML, pages 369-376, 2006. 3 +[14] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7 + +[15] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2, 3, 6, 7 +[16] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In ICCV, pages 11303-11312, 2021. 2, 4, 7, 8 +[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[18] Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Continuous sign language recognition with correlation network. In CVPR, pages 2529-2539, 2023. 2, 5, 7, 8 +[19] Wenbing Huang, Yu Rong, Tingyang Xu, Fuchun Sun, and Junzhou Huang. Tackling over-smoothing for general graph convolutional networks. arXiv preprint arXiv:2008.09864, 2020. 4 +[20] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 8 +[21] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 8 +[22] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 5, 6, 7 +[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. CVIU, 141:108-125, 2015. 4 +[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In CVPR, pages 4297-4305, 2017. 2 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4, 6, 7 +[26] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In ICCV, pages 11542-11551, 2021. 2, 3, 5, 7, 8 +[27] Maria Parelli, Katerina Papadimitriou, Gerasimos Potamianos, Georgios Pavlakos, and Petros Maragos. Spatiotemporal graph convolutional networks for continuous sign language recognition. In ICASSP. IEEE, 2022. 2 +[28] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In MM, pages 1497-1505, 2020. 2 +[29] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In International Conference on Learning Representations, 2020. 6, 7 + +[30] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. TMM, 2021. 2 +[31] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7 +[32] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 3, 5, 7 +[33] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 5, 6 +[34] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2 +[35] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, 2020. 1, 7, 8 +[36] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, pages 1316–1325, 2021. 2, 4, 8 +[37] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, 2021. 1, 2 +[38] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2, 7, 8 \ No newline at end of file diff --git a/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/images.zip b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ef709806024e6898b6ba88522220beda9611f486 --- /dev/null +++ b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b761a3a8e5209d681e15d7d42e570c10d746eff73999cb92e15f83e8234ea896 +size 659949 diff --git a/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/layout.json b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bd08521e5b1651f90487b695074b87f5d471b299 --- /dev/null +++ b/2024/SignGraph_ A Sign Sequence is Worth Graphs of Nodes/layout.json @@ -0,0 +1,10317 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 101, + 470, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 101, + 470, + 118 + ], + "spans": [ + { + "bbox": [ + 124, + 101, + 470, + 118 + ], + "type": "text", + "content": "SignGraph: A Sign Sequence is Worth Graphs of Nodes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 130, + 499, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 130, + 499, + 146 + ], + "spans": [ + { + "bbox": [ + 94, + 130, + 499, + 146 + ], + "type": "text", + "content": "Shiwei Gan† Yafeng Yin†* Zhiwei Jiang† Hongkai Wen‡ Lei Xie† Sanglu Lu†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 148, + 446, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 148, + 446, + 159 + ], + "spans": [ + { + "bbox": [ + 146, + 148, + 446, + 159 + ], + "type": "text", + "content": "† State Key Laboratory for Novel Software Technology, Nanjing University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 161, + 421, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 161, + 421, + 174 + ], + "spans": [ + { + "bbox": [ + 172, + 161, + 421, + 174 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 172, + 161, + 421, + 174 + ], + "type": "text", + "content": " Department of Computer Science, The University of Warwick, UK" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 176, + 519, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 176, + 519, + 188 + ], + "spans": [ + { + "bbox": [ + 70, + 176, + 519, + 188 + ], + "type": "text", + "content": "sw@smail.nju.edu.cn {yafeng, jzw, lxie, sanglu}@nju.edu.cn hongkai.wen@warwick.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 240, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 290, + 553 + ], + "type": "text", + "content": "Despite the recent success of sign language research, the widely adopted CNN-based backbones are mainly migrated from other computer vision tasks, in which the contours and texture of objects are crucial for identifying objects. They usually treat sign frames as grids and may fail to capture effective cross-region features. In fact, sign language tasks need to focus on the correlation of different regions in one frame and the interaction of different regions among adjacent frames for identifying a sign sequence. In this paper, we propose to represent a sign sequence as graphs and introduce a simple yet effective graph-based sign language processing architecture named SignGraph, to extract cross-region features at the graph level. SignGraph consists of two basic modules: Local Sign Graph (LSG) module for learning the correlation of intra-frame cross-region features in one frame and Temporal Sign Graph (TSG) module for tracking the interaction of inter-frame cross-region features among adjacent frames. With LSG and TSG, we build our model in a multiscale manner to ensure that the representation of nodes can capture cross-region features at different granularities. Extensive experiments on current public sign language datasets demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues. Code and models are available at: https://github.com/gswycf/SignGraph." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 568, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 588, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 696 + ], + "type": "text", + "content": "Computer vision technology and natural language processing technology have greatly advanced sign language (SL) research, including Sign Language Recognition (SLR) and Sign Language Translation (SLT). The former task aims at recognizing an isolated sign/continuous signs as a corresponding gloss or gloss sequence, while the latter task aims at translating continuous signs into spoken language [3]. In the modern sign language processing models (SLR models and SLT models), one of the key points of these deep learn" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 216, + 389, + 297 + ], + "blocks": [ + { + "bbox": [ + 319, + 216, + 389, + 297 + ], + "lines": [ + { + "bbox": [ + 319, + 216, + 389, + 297 + ], + "spans": [ + { + "bbox": [ + 319, + 216, + 389, + 297 + ], + "type": "image", + "image_path": "028e330c5cb10bd0a08b0e6b2bb532a6b1bc9f1021d3bb53a297117ab6c920d0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 300, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 322 + ], + "type": "text", + "content": "Figure 1. Comparison of feature extraction between convolutional network and our graph convolutional network." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 398, + 224, + 462, + 289 + ], + "blocks": [ + { + "bbox": [ + 398, + 224, + 462, + 289 + ], + "lines": [ + { + "bbox": [ + 398, + 224, + 462, + 289 + ], + "spans": [ + { + "bbox": [ + 398, + 224, + 462, + 289 + ], + "type": "image", + "image_path": "5471e091b92d6d53f96f20954b39fa1e7c4f81793988af4e7a0c10d470d6af87.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 465, + 224, + 529, + 289 + ], + "blocks": [ + { + "bbox": [ + 465, + 224, + 529, + 289 + ], + "lines": [ + { + "bbox": [ + 465, + 224, + 529, + 289 + ], + "spans": [ + { + "bbox": [ + 465, + 224, + 529, + 289 + ], + "type": "image", + "image_path": "96a2f22030b91665634dfd0a25b05b452c2a74ff0817dc153247813e291942db.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 330, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 546, + 413 + ], + "type": "text", + "content": "ing based methods is to learn sign-related features through training with extensive sign video data. As a matter of course, 2D CNNs and 3D CNNs are used to be the defacto standard visual backbones [7, 10] for extracting local sign frame features, while 1D CNN, LSTM and Transformer [3, 11, 37] are used as the temporal module to capture dynamic changes of sign frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 414, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 546, + 665 + ], + "type": "text", + "content": "However, as shown in Figure 1, the widely-used CNN-based backbones that are mainly migrated from image recognition and object detection tasks, treat the sign frame/sign clips as grids and adopt 'kernels' to extract local features by sliding across sign frames. These CNNs are good at capturing contour-based representation and texture-based representation [12], but they may fail to capture the explicit collaboration of signs in different regions. In fact, sign language tasks need to focus on manual and non-manual features, especially the collaboration of these cues in different regions [35]. Besides, temporal modules like LSTM and Transformer layer usually regard a sign frame as a whole region and learn the overall variance of whole frames, while ignoring the interaction of subregions in adjacent frames. Other temporal methods like 1D CNN can only focus on the same regions in adjacent frames, which may fail to track dynamic motions across different regions. In fact, the dynamic movements in the body, hands, and face which are reflected in the same/different regions between adjacent frames, are primary elements of identifying a sign for sign language tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "In order to make up for the shortcomings of typical CNN backbones explicitly or implicitly and allow the model to learn robust sign-related features, exiting work committed to designing various CNN-based backbones by injecting do" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 191, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 191, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 191, + 713 + ], + "type": "text", + "content": "*Yafeng Yin is the corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13470" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "main knowledge i.e., skeleton [11], depth images [30] or local areas [37]. Besides, some work trained their backbones with more training samples by back translation [36], cross modality augmentation [28] and pre-training models [6], while others added extra constraints on backbones by using knowledge distillation [16], contrastive learning [10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 290, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 290, + 399 + ], + "type": "text", + "content": "Different from previous work, we build a simple yet effective sign graph neural network (SignGraph for short) for continuous sign language recognition (CSLR) task, in which we take sign frames as graphs of nodes and build dynamic graphs to learn cross-region features. The main intuitions are: (1) Sign-related features involve the correlation of different regions in one frame, such as face region and hand regions. Thus, we propose a local sign graph " + }, + { + "bbox": [ + 46, + 147, + 290, + 399 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 147, + 290, + 399 + ], + "type": "text", + "content": " module by dynamically building graphs based on nodes in each frame, to learn intra-frame cross-region features. (2) The dynamic movements in the body, hands, and face are reflected in different regions among adjacent frames. Thus we propose a temporal sign graph " + }, + { + "bbox": [ + 46, + 147, + 290, + 399 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 46, + 147, + 290, + 399 + ], + "type": "text", + "content": " module by dynamically connecting sign-related regions based on nodes in adjacent frames, to learn inter-frame cross-region features. (3) Considering that patches with a single size may not capture local features effectively, we build our SignGraph in a multiscale manner by merging small patches into larger-size patches, and use different-size patches to capture cross-region features at different granularities. We make the following contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 406, + 287, + 647 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 406, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 287, + 491 + ], + "type": "text", + "content": "- Different from CNN-based SL models, we propose a graph-based CSLR model (SignGraph), which processes sign frames as a set of nodes and dynamically builds graphs from one frame and adjacent frames to capture cross-region features related to signs, thus can aggregate sign-related features from different regions for better feature representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 491, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 287, + 562 + ], + "type": "text", + "content": "- We propose a new local sign graph module " + }, + { + "bbox": [ + 47, + 491, + 287, + 562 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 47, + 491, + 287, + 562 + ], + "type": "text", + "content": " that dynamically builds edges based on node features to learn the correlation of cross-region features in one frame, and a new temporal sign graph module " + }, + { + "bbox": [ + 47, + 491, + 287, + 562 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 47, + 491, + 287, + 562 + ], + "type": "text", + "content": " that dynamically connects sign-related regions to learn the interaction of cross-region features among adjacent frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 563, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 598 + ], + "type": "text", + "content": "- We build our model in a multiscale manner with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 598, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 647 + ], + "type": "text", + "content": "- The extensive experiments on CSLR task demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 666, + 134, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 134, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 134, + 679 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "We mainly review CSLR technologies and graph convolutional networks in sign language." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 421 + ], + "type": "text", + "content": "Continuous Sign Language Recognition. The goal of CSLR is to interpret the continuous sign sequence as corresponding gloss sequence [24], thus the core of current CSLR tasks is to identify signs. The widely adopted CNN-based backbones in CSLR which are migrated from the image recognition task, are good at capturing the texture and contours of objects [12]. However, identifying a sign needs to focus on the correlation of different regions (e.g., hand regions and face region) in one frame and the interaction of the same or different regions (e.g., hand motions, facial expression changes) among adjacent frames. As a matter of course, to extract robust sign-related features with CNN-based backbones, some sign language models introduced prior knowledge (or expert knowledge) to guide the model to focus on sign-related features. For example, extra information like local areas [2, 4], skeletons [4, 7, 38] and depth images [30] was manually injected into model. However, in these models, on the one hand, to obtain such extra knowledge, additional algorithms need to be introduced, which will cause additional computational overhead. On the other hand, integrating such knowledge will lead to a complex network structure. Besides, some work turned its attention to obtaining more sign language samples by back translation [36], cross modality augmentation [28], and pretraining with other sign language related datasets [6, 7]. While others were dedicated to designing complex backbones with attention mechanism [18] or adding extra constraints on CNN backbones by adopting knowledge distillation [16, 26], contrastive learning [10]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Graph Convolutional Network. Compared with sequences and grids, graphs are more flexible data structures that model the relationships of a set of objects with nodes and edges. In order to learn representation from the graph, graph convolutional network (GCN) [22] was proposed, and it has been widely applied in action recognition tasks [34] and has also been extended to sign language tasks [11, 17, 21]. To apply GCN to SL tasks, the existing models often relied on extracted skeleton, whose connections between joints make skeleton data naturally fit GCN. Specifically, some work utilized GCN to guide CNN-based backbones to extract skeleton-related features [11, 27]. While others ignored RGB features and applied GCN to explore sign features directly with skeleton [20, 21]. Recently, in addition to applying GCN with skeletons, the usage of applying GCN on images emerged. Specifically, VIG [15] represented an image as a graph structure to capture irregular and complex objects, and demonstrated the effectiveness of graph representation of images. Inspired by VIG, we find that cross-region features are important for identifying a sign and can be captured by graphs. Thus unlike the previous work which adopted CNN-based backbones or relied on skeletons to apply GCN, we directly treat sign frames as graphs of nodes and design a simple GCN-based framework" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13471" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 73, + 521, + 235 + ], + "blocks": [ + { + "bbox": [ + 74, + 73, + 521, + 235 + ], + "lines": [ + { + "bbox": [ + 74, + 73, + 521, + 235 + ], + "spans": [ + { + "bbox": [ + 74, + 73, + 521, + 235 + ], + "type": "image", + "image_path": "d1360906ec4fa4240be220f9e174fedc1c399cb3a2fd003ad726e89ffc757af1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 239, + 384, + 251 + ], + "lines": [ + { + "bbox": [ + 208, + 239, + 384, + 251 + ], + "spans": [ + { + "bbox": [ + 208, + 239, + 384, + 251 + ], + "type": "text", + "content": "Figure 2. The proposed SignGraph architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 263, + 183, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 183, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 183, + 276 + ], + "type": "text", + "content": "for learning sign-related features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 285, + 103, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 103, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 103, + 297 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "content": "In this section, we first give an overview of the proposed graph-based sign language processing architecture. Then we describe the proposed Local Sign Graph " + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "content": " module for capturing intra-frame cross-region features and Temporal Sign Graph " + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "content": " module for capturing inter-frame cross-region features. Finally, with " + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "content": ", we describe our multiscale SignGraph for CSLR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 394, + 163, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 163, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 163, + 406 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "text", + "content": "For a sign language video " + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "inline_equation", + "content": "f = \\{f_i\\}_{i=1}^{\\theta}" + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "text", + "content": " frames, the target of CSLR model is to get a recognized gloss sequence " + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "inline_equation", + "content": "g = \\{g_i\\}_{i=1}^{\\vartheta}" + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "text", + "content": " glosses based on input " + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 412, + 287, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": "As shown in Figure 2, our SignGraph model consists of two key modules, i.e. Local Sign Graph " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " module and Temporal Sign Graph " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " module. First, we convert each frame " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " patches with a patchify stem. Then each frame " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " is represented by a set of patches (i.e., nodes) " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "v_{i} = \\{v_{ij}\\}_{j=1}^{N}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": ". Second, by learning the correlation between each node and connecting " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " nearest neighbors for each node, we build a dynamic local graph " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "G_{i}^{l}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": ". Then a local graph convolutional layer is leveraged to capture cross-region features within each frame. Third, by connecting sign-related regions between two adjacent frames, we build a dynamic temporal graph " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "G_{i}^{t}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " for adjacent frames " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "inline_equation", + "content": "f_{i+1}" + }, + { + "bbox": [ + 46, + 449, + 288, + 628 + ], + "type": "text", + "content": ". Then, a temporal graph convolutional layer is adopted to track the interaction of cross-region features among adjacent frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": "Further, we merge patches into larger size patches by downsampling frames with a factor of 2, then each frame " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " can be represented by another set of patches (nodes) " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": ". Following that, we repeatedly adopt LSG module and TSG module to learn cross-region features with different-size patches. After that, following previous sign language models [10, 26], we adopt a global feature module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "type": "text", + "content": "to learn global changes of whole frames. Finally, a classifier and a widely-used CTC loss [13] are adopted to predict the probability " + }, + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "type": "inline_equation", + "content": "p(g|f)" + }, + { + "bbox": [ + 304, + 263, + 545, + 300 + ], + "type": "text", + "content": " of target gloss sequence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 306, + 430, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 430, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 430, + 319 + ], + "type": "text", + "content": "3.2. Sign Graph Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "content": "Patchify Stem. In order to build graphs based on sign frames instead of skeleton data, we need to convert RGB data into nodes. Following previous work [9, 32], for each frame " + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "inline_equation", + "content": "f_{i} \\in \\mathbb{R}^{[H \\times W \\times 3]}" + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "content": " with height " + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "content": ", we first convert it into a number of patches (or nodes) " + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "inline_equation", + "content": "v_{i} = \\{v_{ij}\\}_{j=1}^{N}" + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "content": " with a patchify stem " + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{PS}" + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 407, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 407, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 372, + 407, + 545, + 422 + ], + "type": "interline_equation", + "content": "\\mu_ {i} = \\left\\{\\mu_ {i j} \\right\\} _ {j = 1} ^ {N} = \\mathcal {P S} (f _ {i}) \\tag {1}", + "image_path": "02f45d5d3b75ff2b44c071a16f608373b1b8aea65be7bca5cb67126ed7bce7bd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": "th node in the " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": "-th frame " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mu_{ij} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": " denotes node features of " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": " is the feature dimension, " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "N = HW / P^2" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": " is the number of patches, and " + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 426, + 547, + 473 + ], + "type": "text", + "content": " is the patch size." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "type": "text", + "content": "Local Sign Graph Learning. For a single frame in the sign video, sign language models should focus on the correlation of different regions, such as face region and hand regions. In order to learn intra-frame cross-region features, we need to build a graph for frame " + }, + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 304, + 477, + 545, + 548 + ], + "type": "text", + "content": ", in which regions with correlation (e.g., hand regions and face region) should be connected." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": "As shown in Figure 3, for a set of nodes " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": ", we first adopt a linear projection layer with learnable weights " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "W_{1}^{l}" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": " to map node features " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": " to the space where the distance function is applied: " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\mu_{i}' = \\mu_{i}W_{1}^{l}" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": ". Then we adopt " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": "-Nearest Neighbors (KNN) algorithm to find top " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": " nearest neighbors " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\mathsf{K}_l}(v_{ij})" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": " for each node based on distance function " + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\mathcal{I}\\mathcal{S}" + }, + { + "bbox": [ + 304, + 549, + 545, + 621 + ], + "type": "text", + "content": " [15]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 350, + 628, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 628, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 350, + 628, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {D I S} \\left(v _ {i j}, v _ {i k}\\right) = \\sqrt {\\sum \\left(\\mu_ {i j} ^ {\\prime} - \\mu_ {i k} ^ {\\prime}\\right) ^ {2}} \\tag {2}", + "image_path": "8b6b00b383c5c1aef4914ceda94cf9c9a20f50fb73f2145b67588c4b2b7fd1fb.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 664, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 664, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 664, + 545, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {N} _ {\\mathrm {K} _ {l}} \\left(v _ {i j}\\right) = \\left\\{v _ {i k} \\mid v _ {i k} \\in \\operatorname {T o p} K \\left(\\left\\{\\mathcal {D I S} \\left(v _ {i j}, v _ {i k}\\right) \\right\\} _ {k = 1} ^ {N}\\right) \\right\\} \\tag {3}", + "image_path": "3a2dcf3b0fe7d180d7b0077df091de07ef9cb2050949eb669517406e115f7088.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "TopK" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " function outputs the top " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " nearest neighbors for " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " based on node distances. For each node " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "13472" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 280, + 145 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 280, + 145 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 280, + 145 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 280, + 145 + ], + "type": "image", + "image_path": "c4e5cec7923a3e95cbb2a3c5b72987a0ad3e17c8516918f6da2ba13472f69af5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 152, + 271, + 163 + ], + "lines": [ + { + "bbox": [ + 63, + 152, + 271, + 163 + ], + "spans": [ + { + "bbox": [ + 63, + 152, + 271, + 163 + ], + "type": "text", + "content": "Figure 3. Graph construction in local sign graph module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": "we add an edge " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "e(v_{ij}, v_{ik})" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "v_{ik}" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "v_{ik} \\in \\mathcal{N}_{\\mathsf{K}_l}(v_{ij})" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": ". Thus, for frame " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": ", we can get the edge set " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "e_i^l = \\{e(v_{ij}, v_{ik}), |j \\in [1, N], v_{ik} \\in \\mathcal{N}_{\\mathsf{K}_l}(v_{ij})\\}" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": ", and build a local graph " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "G_i^l = \\{v_i, e_i^l\\}" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": ". After that, we adopt a graph convolutional layer " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{GCN}_l" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": " and a linear projection layer with learnable weights " + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "inline_equation", + "content": "W_2^l" + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": " to aggregate features from cross regions for each frame." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 261, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 261, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 108, + 261, + 287, + 277 + ], + "type": "interline_equation", + "content": "\\mu_ {i} = \\mu_ {i} + \\mathcal {G C N} _ {l} \\left(\\mu_ {i} ^ {\\prime}, e _ {i} ^ {l}\\right) W _ {2} ^ {l} \\tag {4}", + "image_path": "015d13ca43fe4032a2dba3f3c45c3b7035723ff236aae49c67fb709461b137d2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 282, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 344 + ], + "type": "text", + "content": "The LSG module contains one graph convolutional layer and two linear projection layers. We do not add extra modules (e.g., feed-forward network) since we aim to keep our model simple, and we show that our model can achieve a convincing performance even with the simple design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 346, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 430 + ], + "type": "text", + "content": "Temporal Sign Graph Learning. The dynamic movements in body, hands, and face are essential elements to identify a sign, which are often reflected in different regions among adjacent frames. Hence, we design a temporal graph module, which can dynamically build edges between regions in two adjacent frames and learn inter-frame cross-region features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": "For two adjacent frames " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "f_{i+1}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " nodes " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\{v_{ij}\\}_{j=1}^{N}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\{v_{(i+1)k}\\}_{k=1}^{N}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": ". Similarly, we first adopt a linear projection layer with weights " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "W_{1}^{t}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " to map node features " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " to the space where the distance function is applied: " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\mu_{i}^{\\prime \\prime} = \\mu_{i}W_{1}^{t}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": ". Then we compute distances between " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "v_{(i+1)k}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": ", thus we have a " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " distance matrix " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\{M[j,k] = \\mathcal{DLS}(v_{ij},v_{(i+1)k})|j\\in [1,N],k\\in [1,N]\\}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": ". When building the temporal graph, unlike the LSG module where we build edges for each node " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " (i.e., building a dense graph), we only choose top " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " node pairs between two adjacent frames and connect them (i.e., building a sparse graph). To avoid establishing edges between meaningless node pairs, we set a threshold of 0.05 on normalized distance matrix " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " to filter out patch pairs in the background, since these pairs often have a high and constant similarity (i.e., " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "M[j,k] < 0.05" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": "). Then we get the temporal edge set " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "e_{i}^{t}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " between frame " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "f_{i+1}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": " and temporal graph " + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "inline_equation", + "content": "G_{i}^{t} = \\{(v_{i},v_{i+1}),e_{i}^{t}\\}" + }, + { + "bbox": [ + 47, + 430, + 289, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 652, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 652, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 652, + 287, + 677 + ], + "type": "interline_equation", + "content": "e _ {i} ^ {t} = \\left\\{e \\left(v _ {i j}, v _ {(i + 1) k}\\right) \\mid \\mathcal {D I S} \\left(v _ {i j}, v _ {(i + 1) k}\\right) \\in T o p K (M) \\right\\} \\tag {5}", + "image_path": "2f47b3de041e649f5cdfe5346d692c968d868927eadf7e58916ef8d1d2148245.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "After that, we adopt a temporal graph convolutional layer " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{GCN}_t" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": " and another linear projection " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "W_2^t" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": " to aggregate cross-region features from adjacent frames, as follows." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 533, + 159 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 533, + 159 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 533, + 159 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 533, + 159 + ], + "type": "image", + "image_path": "d48394fb8a432a0326745c25dbc5604ceee138eb1f26c76632d773909248d2b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 164, + 537, + 176 + ], + "lines": [ + { + "bbox": [ + 314, + 164, + 537, + 176 + ], + "spans": [ + { + "bbox": [ + 314, + 164, + 537, + 176 + ], + "type": "text", + "content": "Figure 4. Graph construction in temporal sign graph module." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 202, + 545, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 202, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 202, + 545, + 227 + ], + "type": "interline_equation", + "content": "\\left\\{\\mu_ {i} \\right\\} _ {i = 1} ^ {\\theta} = \\left\\{\\mu_ {i} \\right\\} _ {i = 1} ^ {\\theta} + \\mathcal {G C N} _ {t} \\left(\\left\\{\\mu_ {i} ^ {\\prime \\prime} \\right\\} _ {i = 1} ^ {\\theta}, \\left\\{e _ {i} ^ {t} \\right\\} _ {i = 1} ^ {\\theta - 1}\\right) W _ {2} ^ {t} \\tag {6}", + "image_path": "3284f6495f6145824db4a99db5ec6af9912d90f1a1e87ab87b2fe2bb2b2af489.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 546, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 276 + ], + "type": "text", + "content": "The intuition of building the sparse graph for " + }, + { + "bbox": [ + 304, + 228, + 546, + 276 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 304, + 228, + 546, + 276 + ], + "type": "text", + "content": " module rather than building the dense graph like " + }, + { + "bbox": [ + 304, + 228, + 546, + 276 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 304, + 228, + 546, + 276 + ], + "type": "text", + "content": " module is to avoid the over-smoothing phenomenon [19], which may decrease the distinctiveness of node features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 282, + 546, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 546, + 296 + ], + "type": "text", + "content": "3.3. Multiscale Sign Graph Convolutional Network" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 301, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 546, + 384 + ], + "type": "text", + "content": "Patches with a fixed window may not capture sign features in one region effectively (e.g., hand regions may be split by patches). Thus, similar to Swin Transformer [25], we propose a multiscale SignGraph, which builds our model in multiscale manners to ensure that the representation of nodes can capture cross-region features at different granularities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": "Patch Merging. To produce a multiscale representation, we further design a patch merging module to reduce the number of patches (nodes) and expand the receptive field of patches. Specifically, for " + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": " patches (or nodes) with the size of " + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": ", the patch merging module applies convolutional layers to downsample frames by a factor of 2. Then we get a set of nodes " + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}" + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": " for frame " + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": ", and its corresponding node features " + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mu}_{ij}\\}_{j=1}^{N/4}" + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mu}_{ij} \\in \\mathbb{R}^{2D}" + }, + { + "bbox": [ + 304, + 390, + 547, + 494 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": "Multiscale Sign Graph Learning. After getting the set of nodes " + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\{\\bar{v}_{ij}\\}_{j=1}^{N/4}" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": " with a larger size for frame " + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": " by patch merging, we adopt the another LSG module with " + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": " edges for each node in one frame and TSG module with " + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": " edges for two adjacent frames. For convenience, we name the ith LSG, TSG module as LSG" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "_i" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": ", TSG" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "_i" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": ", and the corresponding hyperparameters as " + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l^i" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t^i" + }, + { + "bbox": [ + 304, + 495, + 547, + 606 + ], + "type": "text", + "content": ". According to experimental results in Section 4.3, we adopt two LSG, TSG modules in our baseline settings." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 616, + 388, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 388, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 388, + 629 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 635, + 369, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 369, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 369, + 647 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Following the previous work [16, 36], we evaluate our model on three publicly available SL datasets. (1) PHOENIX14 [23] is a widely used German SL dataset for CSLR with a vocabulary of 1295 glosses from 9 signers. It contains 5672, 540, and 629 weather forecast sam" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13473" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "plies for training, validation, and testing respectively. (2) PHOENIX14T [3] is another German SL dataset with both gloss annotations and translation annotations. It contains 7096, 519, and 642 samples from 9 signers for training, validation, and testing respectively, and it has a vocabulary of 1066 different signs for CSLR. (3) CSL-Daily is a Chinese SL dataset, which contains 18401, 1077, and 1176 labeled videos from 10 signers for training, validation, and testing respectively, and it has gloss annotations with a vocabulary of 2000 glosses for CSLR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 170, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 170, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 170, + 213 + ], + "type": "text", + "content": "4.2. Experimental Setting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 218, + 286, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 286, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 286, + 231 + ], + "type": "text", + "content": "Here, we describe the baseline settings for our architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "text", + "content": "Data Preprocessing. For data preprocessing, we follow the existing work [18, 26] and apply data augmentation during training, including: resizing frames to " + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "text", + "content": " pixels, random cropping frames to " + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "text", + "content": " pixels, random horizontal flipping with a probability of 0.5, random temporal scaling (" + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "inline_equation", + "content": "\\pm 20\\%" + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "text", + "content": "). During testing, we resize frames to " + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "text", + "content": " pixels and use center cropping to get frames with " + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 236, + 286, + 331 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": "Architecture Setting. Our architecture is implemented by PyTorch 1.11. (1) Patchify stem: Considering that sign frames have a strong 2D local structure, i.e., spatially neighboring pixels are usually highly correlated, we do not adopt non-overlapping patches, which are obtained by linear projection [9] and lack modeling power for 2D local spatial context [32, 33]. Instead, we obtain initial patches (nodes) embedding by using convolutional-based patchify stem. Specifically, we adopt the first four stages of ResNet18 as our patchify stem. The dimension of initial patches " + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": " is set to 512. (2) Patch merging: Similar to the patchify stem, we adopt two convolutional blocks with kernel size " + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": " to downsample frames and get different-size patches after LSG and TSG modules. (3) Sign Graph learning: There are two LSG and TSG modules in our baseline settings, and the initial " + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l^1" + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l^2" + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": " in LSG modules are set to 4 and initial " + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t^1" + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t^2" + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": " in TSG modules are set to 49 (49 is the number of nodes in " + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "inline_equation", + "content": "TSG_2" + }, + { + "bbox": [ + 46, + 337, + 289, + 636 + ], + "type": "text", + "content": "). We adopt the implementation of GCN-Conv [22] as initial graph convolutional layer (GCN layer). (4) Distance function: We adopt Euclidean distance to measure the distance between two nodes in the baseline setting. (5) Global feature module: The Global feature module is the combination of two 1D convolution blocks, 2-layer BiLSTM with hidden size 1024 for global feature modeling, and a fully connected layer for final prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Training Setting. For a fair comparison, we adopt the same training settings as in previous work [26]. We adopt the Adam optimizer with a weight decay of 0.0001 to train our model for 40 epochs on 1 GeForce RTX 3090 GPU. The initial learning rate is 0.0001 with a decay factor of 0.5 at epoch 20 and 30, and the batch size is set to 4. CTC loss" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 314, + 70, + 541, + 223 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 541, + 223 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 541, + 223 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 541, + 223 + ], + "type": "table", + "html": "
\\( LSG_1 \\)\\( TSG_1 \\)\\( LSG_2 \\)\\( TSG_2 \\)DevTest
WERDel/InsWERDel/ins
✘✘✘✘22.38.4/2.522.28.1/2.7
✓19.25.6/2.221.04.8/2.3
✓19.65.6/2.121.55.1/2.5
✓19.35.3/2.120.85.8/2.0
✓19.56.6/1.721.25.4/2.4
✓✓18.45.4/1.620.15.1/2.1
✓✓19.15.6/1.820.84.8/2.2
✓✓18.75.1 /2.320.65.2/1.7
✓✓18.64.3 /1.820.25.5/1.7
✓✓✓✓18.15.4/1.420.16.1/1.8
", + "image_path": "a965897ba2644c582d46a0d174e12d8ba15de43bf02676caa6542a8807554cbb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 338, + 225, + 512, + 236 + ], + "lines": [ + { + "bbox": [ + 338, + 225, + 512, + 236 + ], + "spans": [ + { + "bbox": [ + 338, + 225, + 512, + 236 + ], + "type": "text", + "content": "Table 1. Ablation study on Phoenix14T dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "text", + "content": "is applied after both 1D convolution blocks and the fully connected layer as loss functions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 281, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 304 + ], + "type": "text", + "content": "Evaluation setting. To evaluate our model, we adopt the Word Error Rate (WER) as CSLR performance metric," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 314, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 314, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 372, + 314, + 545, + 337 + ], + "type": "interline_equation", + "content": "W E R = \\frac {\\# S + \\# I + \\# D}{\\# N} \\tag {7}", + "image_path": "61e774e72d9d5a03659169134f7e0dc7c2fdc0c415a0a6047b55514e576504be.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 340, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 340, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 545, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 340, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\# S, \\# I, \\# D" + }, + { + "bbox": [ + 305, + 340, + 545, + 389 + ], + "type": "text", + "content": " denote the minimum number of substitution, insertion and deletion operations needed to convert a predicted gloss sequence to the ground truth. " + }, + { + "bbox": [ + 305, + 340, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\# N" + }, + { + "bbox": [ + 305, + 340, + 545, + 389 + ], + "type": "text", + "content": " is the number of glosses in ground truth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 397, + 398, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 398, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 398, + 410 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 545, + 453 + ], + "type": "text", + "content": "Following the previous work [3], we perform ablation studies on Phoenix14T dataset to verify the effectiveness of the proposed SignGraph." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": "Effects of Proposed Sign Graph Module. As shown in Table 1, results demonstrate that both the proposed " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": " modules contribute to higher performance for CSLR. Specifically, our baseline with ResNet18 as backbone achieves " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": " WER on Phoenix14T dev set. When we replace backbone with the designed local sign graph module " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "LSG_{1}" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "LSG_{2}" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": ", WER on dev set decreases by " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": " respectively. When we replace backbone with designed temporal sign graph module " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "TSG_{1}" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "TSG_{2}" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": ", WER decreases by " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": " respectively, which verifies that both intra-frame cross-region features and inter-frame cross-region features are crucial for learning sign-related features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Then we combine " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " modules, WER decreases by " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " when applying " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "LSG_{1}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "TSG_{1}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " when applying " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "LSG_{2}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "TSG_{2}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " respectively, which demonstrates that combination of cross-region features in one frame and those among adjacent frames are beneficial for identifying signs. When adopting two stages of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", performance can achieve further improvement (i.e., WER decreases by " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.2\\% / 2.1\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " on dev/test set)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13474" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 268, + 152 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 268, + 152 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 268, + 152 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 268, + 152 + ], + "type": "table", + "html": "
\\( LSG_1 \\)\\( TSG_1 \\)DevTest
WERDel/InsWERDel/ins
DenseSparse18.45.4/1.620.15.1/2.1
DenseDense20.24.9/2.921.85.5/2.6
SparseSparse19.15.0/1.921.75.3/2.7
SparseDense20.95.4/2.421.96.1/2.5
", + "image_path": "952d95c66642a9268df30f240dcaf6836965ff800d18afbf06512cd14e033477.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 155, + 240, + 166 + ], + "lines": [ + { + "bbox": [ + 94, + 155, + 240, + 166 + ], + "spans": [ + { + "bbox": [ + 94, + 155, + 240, + 166 + ], + "type": "text", + "content": "Table 2. Ablation study on graph types." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 176, + 287, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 200 + ], + "type": "text", + "content": "It shows that leveraging multiscale representation of nodes can capture cross-region features at different granularities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "content": "Effect of Graph Type. In SignGraph, we adopt dense graphs (i.e., connecting top " + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "content": " neighbors for each node) in the LSG module, while adopting sparse graphs (only connecting top " + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "content": " node pairs) in the TSG module. To verify the effectiveness of our design, we test the effect of different combinations of graph types in LSG and TSG modules. For convenience, we only adopt one LSG module and one TSG module. As shown in Table 2, when we adopt the dense graph in TSG module or adopt the sparse graph in LSG module, there is a noticeable performance drop (i.e., WER on dev set rises by " + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 205, + 288, + 373 + ], + "type": "text", + "content": "). Our module with dense graph in LSG and sparse graph in TSG achieves the best recognition performance, which demonstrates the effectiveness of our model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": "Effects of Hyperparameters K. In the proposed multiscale SignGraph, there are four hyperparameters, i.e., " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l^1" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l^2" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": " in LSG modules and " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t^1" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t^2" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": " in TSG modules. Here, to set proper hyperparameters K for our model, we test recognition performance under different " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": " ranging ranging from 2 to 9 and under different " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": " ranging from 7 to 91. Since there are many combinations of these hyperparameters, it is impossible to enumerate all possible values to find the globally optimal setting, thus we fix the other 3 parameters and adjust one to obtain suitable parameter settings. According to Figure 5, we set " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l^1" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l^2" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t^1" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t^2" + }, + { + "bbox": [ + 46, + 379, + 287, + 523 + ], + "type": "text", + "content": " to 3, 4, 49, 49 respectively based on WER performance on DEV set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 636 + ], + "type": "text", + "content": "Effects of Different Backbones. To verify the effectiveness of the proposed SignGraph model, we replace our backbone with other patch-based networks (i.e., CvT [33], Swin transformer [25] and VIG [15]). According to Table 3a, simply adopting the current patch-based SOTA backbones will not gain appealing performance on CSLR task, while our proposed SignGraph backbone achieves satisfactory performance by effectively capturing intra-frame cross-region features and inter-frame cross-region features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Effects of Distance Function. In the SignGraph, we adopt KNN algorithm to find the nearest neighbors based on the distance of two nodes. Here, we show performances with commonly used distance functions including Cosine distance, Chebyshev distance and Euclidean distance in Table 3b. There are subtle performance gaps between different" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 422, + 152 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 422, + 152 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 422, + 152 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 422, + 152 + ], + "type": "image", + "image_path": "71a4ca41c005b0bb9a7d28c7d142b36bc76f1dcc2fe0423398ed3abd28b41721.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 71, + 539, + 152 + ], + "blocks": [ + { + "bbox": [ + 426, + 71, + 539, + 152 + ], + "lines": [ + { + "bbox": [ + 426, + 71, + 539, + 152 + ], + "spans": [ + { + "bbox": [ + 426, + 71, + 539, + 152 + ], + "type": "image", + "image_path": "1367bac6a034ff9617870c5b76b955d1040d55c540aedfde196433b7ee515148.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 153, + 421, + 235 + ], + "blocks": [ + { + "bbox": [ + 308, + 153, + 421, + 235 + ], + "lines": [ + { + "bbox": [ + 308, + 153, + 421, + 235 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 421, + 235 + ], + "type": "image", + "image_path": "bff04dab6e61bfe6e11f9ac902a6b13a4b809161366a10cdaddd062232944084.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 238, + 482, + 249 + ], + "lines": [ + { + "bbox": [ + 369, + 238, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 369, + 238, + 482, + 249 + ], + "type": "text", + "content": "Figure 5. Effects of " + }, + { + "bbox": [ + 369, + 238, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_l" + }, + { + "bbox": [ + 369, + 238, + 482, + 249 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 369, + 238, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\mathsf{K}_t" + }, + { + "bbox": [ + 369, + 238, + 482, + 249 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 153, + 538, + 234 + ], + "blocks": [ + { + "bbox": [ + 425, + 153, + 538, + 234 + ], + "lines": [ + { + "bbox": [ + 425, + 153, + 538, + 234 + ], + "spans": [ + { + "bbox": [ + 425, + 153, + 538, + 234 + ], + "type": "image", + "image_path": "600fa3fd1c0c550f42d3822f2f838cedc8fbb8892123a4855ea19c9086ebeca3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "content": "distance functions and we adopt Euclidean distance in our model for better performances." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 293, + 546, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 546, + 401 + ], + "type": "text", + "content": "Effects of GCN Layer. We also test the representative variants of graph convolution, including GATv2Conv [1], SAGEConv [14], GCNConv [22] and EdgeConv [31]. From Table 3c, we can see that our model can achieve excellent performances with different GCN layer, indicating the flexibility of SignGraph model. Among them, EdgeConv achieves the best recognition performance (i.e., " + }, + { + "bbox": [ + 305, + 293, + 546, + 401 + ], + "type": "inline_equation", + "content": "17.8\\%" + }, + { + "bbox": [ + 305, + 293, + 546, + 401 + ], + "type": "text", + "content": " WER). In the rest of the experiments, we use EdgeConv layer in our module by default unless specifically stated." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 408, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 493 + ], + "type": "text", + "content": "Effect of Patch Sizes. Patches with different sizes can capture different local features and further affect model performance. To test the effect of patch sizes, we adopt only one LSG module and one TSG module. As shown in Table 3d, model with a smaller patch size (i.e., " + }, + { + "bbox": [ + 305, + 408, + 545, + 493 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 305, + 408, + 545, + 493 + ], + "type": "text", + "content": ") achieves the worst performance, while model with too large patch size may also degenerate local features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": "Effect of Multiscale SignGraph. To leverage the scale-invariant property of images [15], PyVIG adopts pyramid architecture that gradually increases patch size from 4 to 32, by gradually shrinking the spatial size of feature maps. Similarly, we also gradually add " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " modules in the early stage of the patchify stem and add patch merging module at the end of each " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " module for downsampling. In regard to the added LSG and TSG modules in each stage, we also choose appropriate hyperparameters " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "K_{l}, K_{t}" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " for them through extensive experiments. As shown in Table 3e, adding more " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " modules does not always bring more performance gain, thus we adopt two " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 305, + 501, + 546, + 657 + ], + "type": "text", + "content": " modules that increase patch size from 16 to 32 in our model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Effect of DropEdge. DropEdge [29] tackles oversmoothing and over-fitting problems for dense graph by randomly removing a certain number of edges from the input graph at each training epoch. Considering that we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13475" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 61, + 188, + 115 + ], + "blocks": [ + { + "bbox": [ + 67, + 61, + 188, + 115 + ], + "lines": [ + { + "bbox": [ + 67, + 61, + 188, + 115 + ], + "spans": [ + { + "bbox": [ + 67, + 61, + 188, + 115 + ], + "type": "table", + "html": "
BackBoneWERDelIns
CvT [32]39.215.30.9
SwinT [25]45.416.31.3
PyVIG [15]35.412.11.3
Ours17.95.11.5
", + "image_path": "8cdd8844da732c97c63e895de0fdffe5cf66053b80f1d77cfa79e0d4a6a5cc9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 139, + 187, + 181 + ], + "blocks": [ + { + "bbox": [ + 49, + 118, + 201, + 136 + ], + "lines": [ + { + "bbox": [ + 49, + 118, + 201, + 136 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 201, + 136 + ], + "type": "text", + "content": "(a) Backbone. Comparison of different backbones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 139, + 187, + 181 + ], + "lines": [ + { + "bbox": [ + 67, + 139, + 187, + 181 + ], + "spans": [ + { + "bbox": [ + 67, + 139, + 187, + 181 + ], + "type": "table", + "html": "
PatchSizeWERDelIns
819.14.82.6
1618.45.41.6
3218.75.71.7
", + "image_path": "c3b8800fe9ce9cbd19698d543f0280a672271e0ebb21be22d8340514ef2e9f0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 227, + 61, + 369, + 105 + ], + "blocks": [ + { + "bbox": [ + 50, + 184, + 201, + 203 + ], + "lines": [ + { + "bbox": [ + 50, + 184, + 201, + 203 + ], + "spans": [ + { + "bbox": [ + 50, + 184, + 201, + 203 + ], + "type": "text", + "content": "(d) Patch size. Comparison of different patch sizes with one LSG and TSG modules." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 61, + 369, + 105 + ], + "lines": [ + { + "bbox": [ + 227, + 61, + 369, + 105 + ], + "spans": [ + { + "bbox": [ + 227, + 61, + 369, + 105 + ], + "type": "table", + "html": "
DistanceWERDelIns
Cosine18.05.01.5
Chebyshev18.24.81.6
Euclidean17.95.11.5
", + "image_path": "e64c38a4150ce9f32b1380a06b07c6409fecebfcd0038e8f1a307f0b913627fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 229, + 138, + 367, + 182 + ], + "blocks": [ + { + "bbox": [ + 220, + 117, + 372, + 136 + ], + "lines": [ + { + "bbox": [ + 220, + 117, + 372, + 136 + ], + "spans": [ + { + "bbox": [ + 220, + 117, + 372, + 136 + ], + "type": "text", + "content": "(b) Distance function. Comparison of distance functions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 229, + 138, + 367, + 182 + ], + "lines": [ + { + "bbox": [ + 229, + 138, + 367, + 182 + ], + "spans": [ + { + "bbox": [ + 229, + 138, + 367, + 182 + ], + "type": "table", + "html": "
StagesWERDelDel
16→3217.85.01.6
8→16→3218.45.31.7
4→8→16→3218.14.71.9
", + "image_path": "3b015c0ea1b3565f24c02a7f1de42af46353d30bf7479d02bcec0eda6ecfa12b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 395, + 61, + 549, + 115 + ], + "blocks": [ + { + "bbox": [ + 220, + 184, + 372, + 204 + ], + "lines": [ + { + "bbox": [ + 220, + 184, + 372, + 204 + ], + "spans": [ + { + "bbox": [ + 220, + 184, + 372, + 204 + ], + "type": "text", + "content": "(e) Multiscale SignGraph. Effect of the number of stages in multiscale SignGraph." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 395, + 61, + 549, + 115 + ], + "lines": [ + { + "bbox": [ + 395, + 61, + 549, + 115 + ], + "spans": [ + { + "bbox": [ + 395, + 61, + 549, + 115 + ], + "type": "table", + "html": "
GraphConvWERDelIns
GATv2Conv [1]18.65.21.8
SAGEConv [14]18.04.81.7
GCNConv [22]17.95.11.5
EdgeConv [31]17.85.01.6
", + "image_path": "b523f2e340e8995134a0bc7b222f3e360103b09655b7935663e1c23cf9530364.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 395, + 138, + 550, + 181 + ], + "blocks": [ + { + "bbox": [ + 392, + 118, + 543, + 137 + ], + "lines": [ + { + "bbox": [ + 392, + 118, + 543, + 137 + ], + "spans": [ + { + "bbox": [ + 392, + 118, + 543, + 137 + ], + "type": "text", + "content": "(c) Graph Convolution Effect of different GCN layers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 395, + 138, + 550, + 181 + ], + "lines": [ + { + "bbox": [ + 395, + 138, + 550, + 181 + ], + "spans": [ + { + "bbox": [ + 395, + 138, + 550, + 181 + ], + "type": "table", + "html": "
DropRateWERDelIns
017.85.01.6
15%19.26.71.3
30%21.35.91.7
", + "image_path": "feb42feb581fd4b577d00ef5457dadf5b33ad7fe4402861dcb973af37eeb7258.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 391, + 184, + 543, + 204 + ], + "lines": [ + { + "bbox": [ + 391, + 184, + 543, + 204 + ], + "spans": [ + { + "bbox": [ + 391, + 184, + 543, + 204 + ], + "type": "text", + "content": "(f) Drop edge. Adding DropEdge [29] in the local graph does not improve performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 82, + 232, + 512, + 437 + ], + "blocks": [ + { + "bbox": [ + 165, + 210, + 427, + 221 + ], + "lines": [ + { + "bbox": [ + 165, + 210, + 427, + 221 + ], + "spans": [ + { + "bbox": [ + 165, + 210, + 427, + 221 + ], + "type": "text", + "content": "Table 3. Ablation experiments of SignGraph on Phoenix14T dev set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 232, + 512, + 437 + ], + "lines": [ + { + "bbox": [ + 82, + 232, + 512, + 437 + ], + "spans": [ + { + "bbox": [ + 82, + 232, + 512, + 437 + ], + "type": "image", + "image_path": "7196f2e33b9ae4497c2a4e47eab83a841d05d857f5d4fcfa67f7afcd834b1067.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "text", + "content": "Figure 6. Visualization of graph construction of " + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "inline_equation", + "content": "LSG_{1}" + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "inline_equation", + "content": "TSG_{1}" + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "text", + "content": " in the first row, " + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "inline_equation", + "content": "LSG_{2}" + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "inline_equation", + "content": "TSG_{2}" + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "text", + "content": " in the second row. The graph in " + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "text", + "content": " module is shown in yellow, and the graph in " + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 46, + 445, + 545, + 478 + ], + "type": "text", + "content": " is shown in blue. We also show some 'unimportant' edges between nodes in the background with red color." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 487, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 571 + ], + "type": "text", + "content": "build dense graphs for " + }, + { + "bbox": [ + 46, + 487, + 287, + 571 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 487, + 287, + 571 + ], + "type": "text", + "content": " module, we also add DropEdge with different drop rates to " + }, + { + "bbox": [ + 46, + 487, + 287, + 571 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 487, + 287, + 571 + ], + "type": "text", + "content": " module and show the recognition performance in Table 3f. However, adding DropEdge does not improve our model performance, which means that our model can achieve excellent performance without relying on external modules, and can handle over-smoothing and over-fitting problems well." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Visualization of SignGraph. To verify whether our model can effectively capture sign-related features, we select a sign video from Phoenix14T test set and visualize the constructed graph structure of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " modules in both two stages. In Figure 6, we show the graphs of two stages and we only show part of the edges for a better display. In the " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " module, our model can link the nodes with similar content and semantic representation (e.g., hand regions and face regions) for extracting better intra-frame cross-region features. In the " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " module, edges can be built for tracking dynamic changes in gestures and facial expres" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 487, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 583 + ], + "type": "text", + "content": "sions, thus can better identify a sign with inter-frame cross-region features. We also show some 'unimportant' edges between nodes in the background with red color. As can be seen, background nodes in " + }, + { + "bbox": [ + 304, + 487, + 545, + 583 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 304, + 487, + 545, + 583 + ], + "type": "text", + "content": " are naturally connected to their neighboring nodes, and there are still a few background nodes connected in " + }, + { + "bbox": [ + 304, + 487, + 545, + 583 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 304, + 487, + 545, + 583 + ], + "type": "text", + "content": ". Fortunately, nodes in the background will not 'disturb' nodes in sign-related regions, which demonstrates the effectiveness of our model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 597, + 391, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 391, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 391, + 610 + ], + "type": "text", + "content": "4.4. Comparisons" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on Phoenix14T dataset. As shown in Table 4, we compare our model with existing models on CSLR performance, and we provide both the performances on the validation set (i.e., 'DEV') and test set (i.e., 'TEST'). Most of the current models adopt existing CNN-based backbones, and achieve convincing performances by injecting extra cues [7, 35, 38], adding extra constraints [10, 16, 26] or introducing attention mechanism [18]. The GCN-based" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13476" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 68, + 533, + 249 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 533, + 249 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 533, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 533, + 249 + ], + "type": "table", + "html": "
ModelBackboneExtra cuesPhoenixPhoenixT
F/MHSPDEVTESTDEVTEST
WERdel/insWERdel/insWERWER
STMC [35] (AAAI'20)VGG11✓✓✓21.17.7/3.420.77.4/2.619.621.0
C2SLR [38](CVPR'22)ResNet18✓20.5-/-20.4-/-20.220.4
TwoStream [7](NeurIPS'22)S3D✓✓✓✓*18.4-/-18.8-/-17.719.3
SLT [5] (CVPR'20)GooleNet✓----24.624.5
TwoStream [7] (NeurIPS'22)S3D✓*22.4-/-23.3-/-21.122.4
VAC [26] (ICCV'21)ResNet18✓21.27.9/2.522.38.4/2.6--
SMKD [16] (ICCV'21)ResNet18✓20.86.8/2.521.06.3/2.320.822.4
CorrNet [18] (CVPR'23)ResNet18✓18.85.6/2.819.45.7/2.318.920.5
FCN [8] (ECCV'20)customed23.7-/-23.9-/-
Contrastive [10] (IJCAI'23)ResNet1819.65.1/2.719.85.8/3.020.020.1
HST-GNN [21] (WACV'22)Customized(CNN+GCN)✓✓✓✓19.5-/-19.8-/-19.519.8
CoSign [20] (ICCV'23)ST-GCN(GCN)✓✓✓19.7-/-20.1-/-19.520.1
SignGraphCustomized(GCN)✓18.45.6/1.820.15.4/2.118.320.0
MultiSignGraphCustomized(GCN)✓18.24.9/2.019.15.3/1.917.819.1
", + "image_path": "0c1e9d2c9f7411027d7375990a3e30d1bd57e75ddca2fcd63658b30da3860d41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 285, + 285, + 376 + ], + "blocks": [ + { + "bbox": [ + 46, + 252, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 545, + 275 + ], + "type": "text", + "content": "Table 4. Comparison of CSLR performance on Phoenix14 and Phoenix14T datasets. (F: face, M: mouth, H: hands, S: skeleton, P: pretraining backbone with ImageNet, " + }, + { + "bbox": [ + 46, + 252, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\sqrt{}^{*}" + }, + { + "bbox": [ + 46, + 252, + 545, + 275 + ], + "type": "text", + "content": ": pretraining on other SL-related datasets.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 285, + 285, + 376 + ], + "lines": [ + { + "bbox": [ + 49, + 285, + 285, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 285, + 376 + ], + "type": "table", + "html": "
ModelBackboneExtra cuesDEVTEST
SPWERdel/insWERdel/ins
SLT [5](CVPR'20)GoogleNet✓33.110.3/4.432.09.6/4.1
TwoStream [7](NeurIPS'22)S3D✓✓*25.4-/-25.3-/-
TwoStream [7](NeurIPS'22)S3D✓*28.9-/-28.5-/-
BN-TIN [36](CVPR21)GoogLeNet✓33.613.9/3.433.113.5/3.0
CorrNet [18](CVPR'23)ResNet18✓30.6-/-30.1-/-
CoSign [20] (ICCV'23)ST-GCN(GCN)✓28.1-/-27.2-/-
SignGraphcustomized(GCN)✓28.48.8/2.427.48.2/2.1
MultiSignGraphcustomized(GCN)✓27.37.9/2.326.47.8/2.1
", + "image_path": "aa252f2a559f8155d5660f11ea92651eca9fadac773cd2f009b4fc6ed799557f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 380, + 287, + 392 + ], + "lines": [ + { + "bbox": [ + 47, + 380, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 287, + 392 + ], + "type": "text", + "content": "Table 5. Comparison of CSLR performance on CSL-daily dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": "model CoSign [20] mainly relies on pre-processed fine-grained skeleton data, and achieves " + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "inline_equation", + "content": "19.5\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "inline_equation", + "content": "20.1\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": " WER on dev, test set. While HST-GCN [21] adopts both CNN-based backbone and GCN-based backbone to extract RGB features and skeleton features respectively, achieving " + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "inline_equation", + "content": "19.5\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": " WER on dev, test set. It is worth mentioning that the state-of-the-art (SOTA) model TwoStream utilizes both RGB features and fine-grained skeleton features (i.e., keypoints in hands, body and face), and pretrains its backbone on other SL-related datasets, achieving " + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "inline_equation", + "content": "17.7\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "inline_equation", + "content": "19.3\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": " WER on dev, test set respectively. Without using any extract cues, our best model still achieves comparable performance with TwoStream on dev set. While on test set, our model even outperforms TwoStream model by " + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": " WER." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "text", + "content": "Evaluation on Phoenix14 dataset. We also compare our model with existing models of CSLR performance on Phoenix14 dataset. As shown in Table 4, with the simple yet powerful multiscale SignGraph, our model can achieve the best performance (i.e., " + }, + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "inline_equation", + "content": "18.2\\%" + }, + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "text", + "content": " WER) on dev set and a comparable performance with SOTA model on test set, which demonstrates the effectiveness of our model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Evaluation on CSL-daily dataset. We also provide CSLR performance comparisons with other methods on CSL-daily dataset. As shown in Table 5, the WER of our model on dev and test set is " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "27.3\\%" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "26.4\\%" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", respectively. Our model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 287, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 545, + 312 + ], + "type": "text", + "content": "outperforms most existing models, and achieves a comparable performance with TwoStream model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 333, + 378, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 378, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 378, + 345 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "text", + "content": "In this paper, we propose a simple yet effective SignGraph architecture, which represents a sign sequence as a graph structure to extract sign-related features at the graph level. Specifically, to learn the correlation of cross-region features in one frame, we propose a Local Sign Graph " + }, + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "inline_equation", + "content": "LSG" + }, + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "text", + "content": " module, which dynamically builds edges between nodes in one frame and aggregates intra-frame cross-region features. To track the interaction of cross-region features among adjacent frames, we propose a Temporal Sign Graph " + }, + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "inline_equation", + "content": "TSG" + }, + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "text", + "content": " module, which dynamically builds edges between nodes among adjacent frames and captures inter-frame cross-region features. Besides, we also build our model in multiscale manners with different patch sizes to ensure that the representation of nodes can capture cross-region features at different granularities. The extensive experiments on public datasets demonstrate the superiority of the proposed SignGraph, which achieves comparable performances with the SOTA model without using any extra cues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 594, + 416, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 416, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 416, + 608 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by National Key Research and Development Program of China under Grant No. 2022YFB3303900; Jiangsu Provincial Key Research and Development Program under Grant BE2020001-4; National Natural Science Foundation of China under Grant Nos. 62172208, 62272216. This work is partially supported by Collaborative Innovation Center of Novel Software Technology and Industrialization." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13477" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In ICCV, pages 3075-3084. IEEE, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "text", + "content": "[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, pages 7784-7793, 2018. 1, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "text", + "content": "[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In ECCV, pages 301-319. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 295 + ], + "type": "text", + "content": "[5] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, pages 10023-10033, 2020. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "type": "text", + "content": "[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 385 + ], + "type": "text", + "content": "[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 35:17043-17056, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 420 + ], + "type": "text", + "content": "[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV. Springer, 2020. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 421, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 487 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 488, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 543 + ], + "type": "text", + "content": "[10] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Kang Xia, Lei Xie, and Sanglu Lu. Contrastive learning for sign language recognition and translation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 763-772. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[11] Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Lei Xie, and Sanglu Lu. Skeleton-aware neural sign language translation. In MM, pages 4353-4361, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 634 + ], + "type": "text", + "content": "[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[13] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In ICML, pages 369-376, 2006. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in Neural Information Processing Systems, 35:8291-8303, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[16] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In ICCV, pages 11303-11312, 2021. 2, 4, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "text", + "content": "[18] Lianyu Hu, Liqing Gao, Zekang Liu, and Wei Feng. Continuous sign language recognition with correlation network. In CVPR, pages 2529-2539, 2023. 2, 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[19] Wenbing Huang, Yu Rong, Tingyang Xu, Fuchun Sun, and Junzhou Huang. Tackling over-smoothing for general graph convolutional networks. arXiv preprint arXiv:2008.09864, 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[20] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 316, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 381 + ], + "type": "text", + "content": "[21] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "text", + "content": "[22] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "text", + "content": "[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. CVIU, 141:108-125, 2015. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "text", + "content": "[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In CVPR, pages 4297-4305, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 548 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "text", + "content": "[26] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In ICCV, pages 11542-11551, 2021. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "text", + "content": "[27] Maria Parelli, Katerina Papadimitriou, Gerasimos Potamianos, Georgios Pavlakos, and Petros Maragos. Spatiotemporal graph convolutional networks for continuous sign language recognition. In ICASSP. IEEE, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "text", + "content": "[28] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In MM, pages 1497-1505, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[29] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In International Conference on Learning Representations, 2020. 6, 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13478" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 430 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[30] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. TMM, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[31] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 207 + ], + "type": "text", + "content": "[32] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 3, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 286, + 251 + ], + "type": "text", + "content": "[33] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 286, + 285 + ], + "type": "text", + "content": "[34] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "type": "text", + "content": "[35] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, 2020. 1, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 286, + 363 + ], + "type": "text", + "content": "[36] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, pages 1316–1325, 2021. 2, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 286, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 397 + ], + "type": "text", + "content": "[37] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "type": "text", + "content": "[38] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2, 7, 8" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13479" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_content_list.json b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f73a32d83597c878e3e96ebf8c735c6a47de7820 --- /dev/null +++ b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_content_list.json @@ -0,0 +1,1911 @@ +[ + { + "type": "text", + "text": "SimAC: A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models", + "text_level": 1, + "bbox": [ + 109, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feifei Wang $^{1,2,*}$ ; Zhentao Tan $^{2,1}$ , Tianyi Wei $^{1}$ , Yue Wu $^{2}$ , Qidong Huang $^{1,\\dagger}$", + "bbox": [ + 171, + 202, + 803, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China 2Alibaba Cloud", + "bbox": [ + 220, + 220, + 750, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangfeifei@, tzt@, bestwty@, hqd0037@}mail.ustc.edu.cn matthew.wy@alibaba-inc.com", + "bbox": [ + 114, + 241, + 849, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the success of diffusion-based customization methods on visual content creation, increasing concerns have been raised about such techniques from both privacy and political perspectives. To tackle this issue, several anti-customization methods have been proposed in very recent months, predominantly grounded in adversarial attacks. Unfortunately, most of these methods adopt straightforward designs, such as end-to-end optimization with a focus on adversarially maximizing the original training loss, thereby neglecting nuanced internal properties intrinsic to the diffusion model, and even leading to ineffective optimization in some diffusion time steps. In this paper, we strive to bridge this gap by undertaking a comprehensive exploration of these inherent properties, to boost the performance of current anti-customization approaches. Two aspects of properties are investigated: 1) We examine the relationship between time step selection and the model's perception in the frequency domain of images and find that lower time steps can give much more contributions to adversarial noises. This inspires us to propose an adaptive greedy search for optimal time steps that seamlessly integrates with existing anti-customization methods. 2) We scrutinize the roles of features at different layers during denoising and devise a sophisticated feature-based optimization framework for anti-customization. Experiments on facial benchmarks demonstrate that our approach significantly increases identity disruption, thereby protecting user privacy and copyright. Our code is available at: https://github.com/somuchtoe/SimAC.", + "bbox": [ + 75, + 323, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 791, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Latent Diffusion model (LDM) [12, 31, 32] has been recently proven as a strong paradigm for photorealistic visual content generation. The emergence of open-source Stable", + "bbox": [ + 75, + 816, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18743d4cd9e1f11606bc64b573c72a40db7ef86b3c365e6716f6a3e13a57acfa.jpg", + "image_caption": [ + "Figure 1. Comparison between Anti-DreamBooth before and after adding SimAC. Our method further boosts its ability to de-identity." + ], + "image_footnote": [], + "bbox": [ + 506, + 290, + 890, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion encourages users to explore creative possibilities with LDMs. Users only need to provide several images representing the same subject along with a rare identifier to customize their diffusion models [7, 13, 22, 28]. After finetuning the model or optimizing the text embedding of rare identifiers, it can flexibly generate high-quality images containing the specified object.", + "bbox": [ + 496, + 594, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The convenience of customizing large-scale text-to-image models also allows the malicious users to generate forgery images that violate the truth. Some of them may use such technique to steal others' painting styles and generate new painting content without permission. While some of them may collect one's portrait photos that are published on the social platform, and generate fake images of this person through customization. In other words, the lawbreakers can easily produce fake news that contains celebrity photos [2], greatly misleading the public. The infringement poses threats to user privacy and intellectual property. Hence, it is essential to devise effective countermeasures to safeguard users against such malicious usage.", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship in Alibaba Cloud.", + "bbox": [ + 94, + 875, + 367, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12047", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1fa719306813ca22bcd59973e0e26be945ca4a932f2abc81a5869de3fe3a215.jpg", + "image_caption": [ + "(a) the number of absolute gradients $< 1e^{-10}$" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 331, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/288d32bee9a12b2e79eca9fc28e50aeb7a5acea7ee5c2e1273d04f5f15d64010.jpg", + "image_caption": [ + "(b) the maximum absolute gradients", + "Figure 2. Distribution of Anti-DreamBooth attack gradients on different diffusion timesteps, where (a) counts the number of absolute gradients below the threshold $1e^{-10}$ at each timestep, (b) presents the maximum absolute gradients at different timesteps, and (c) demonstrates how the mean and median of the absolute gradients change over timesteps. Apparently, the absolute gradients show great discrepancy on the varying timesteps and nearly zero values appear at large timesteps, which leads to ineffective noise optimization." + ], + "image_footnote": [], + "bbox": [ + 334, + 87, + 591, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/20f4c6fcb3346b7083d62a5bad82c4d7864bbe5e3390bb5b37cd36f9f494df94.jpg", + "image_caption": [ + "(c) the mean and median of the absolute gradients" + ], + "image_footnote": [], + "bbox": [ + 594, + 88, + 883, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current anti-customization methods are generally based on adversarial attack [9, 33]. AdvDM [23] is the pioneering work that uses adversarial noise to protect user images from being customized by diffusion models. It ingeniously combines the diffusion model with adversarial samples, firstly achieving user privacy protection. Anti-DreamBooth [36] further enhances the protective effect by employing alternate training. Both of them randomly select the timesteps from (0, max denoising steps) regarding the noise added to LDM latent, and directly employ the maximization of the LDM training loss as the optimization objective.", + "bbox": [ + 75, + 368, + 468, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as showcased in Figure 2, we notice that the gradient of the perturbed images is quite small and even zero when the randomly sampled timesteps are relatively large, i.e., the noisy latent is closer to Gaussian noise. This implies that within the limited steps of an adversarial attack, a large portion of steps are ineffective since the perturbed image cannot optimize the objective with such zero gradients, leading to a decrease in both protection effectiveness and efficiency. Hence, the images protected by these current methods, when customized, still allow the model to capture many details from user-input images and leak the user's privacy, as shown in Figure 1.", + "bbox": [ + 75, + 536, + 468, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The fundamental reason is that these methods fail to combine the adversarial attacks with properties inherent in diffusion models. We aim to conduct an in-depth analysis of why the gradients of perturbed images have such discrepancy at different time steps and how diffusion models perceive input images at different intermediate layers. Then, based on the in-depth analysis, we propose improvements to existing customization methods from both the temporal and feature dimensions.", + "bbox": [ + 75, + 718, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To observe the relationship between timesteps and the gradient of the perturbed images, we first reconstruct images predicted from noisy latents at different time steps and", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "compare them with the input images. Considering regular adversarial noises mainly affect the high frequency of images, our comparison is conducted in the frequency domain, aiming at investigating whether the model's perception lies on lower-level or higher-level. We observe that, the model focuses on the higher frequency components of images when selecting smaller time steps, and vice versa. Therefore, introducing adversarial noises at larger timesteps is ineffective, since the subtle changes perturbed on images can hardly affect the low-frequency of the generated images. To improve the effectiveness, we propose an adaptive timestep selection method to find optimal time intervals, where we iteratively update the range of selection in a greedy way. To explore the effect of different layers' features in the U-Net decoder during denoising, we employ PCA (Principal Component Analysis) to visualize them and show the discrepancy. We clearly find that within the decoder, the feature extraction gradually shifts from low-frequency to high-frequency as the layer goes deeper. It indicates the higher layers concentrates more on the texture of images, while the lower layers focus on the structure. Consequently, we select the features representing high-frequency information during optimization, since regular adversarial noises concentrate on the high frequency of images. And we construct the feature interference loss and integrate it with the diffusion denoising loss as the objective function, to improve the ability of identity interference.", + "bbox": [ + 496, + 368, + 890, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 517, + 777, + 774, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal the inadequate optimization steps that exist in current anti-customization methods, and gives detailed analysis regarding the perceptual discrepancy of diffusion models at different timesteps and intermediate layers, we have better aligned the optimization of adversarial gradients with diffusion models.", + "- Based on analysis, we propose a simple but effective anti" + ], + "bbox": [ + 500, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12048", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "customization method, including adaptive greedy time selection and a feature interference loss to improve the protection ability. Our method can be easily generalized to different anti-customization frameworks and improve their performance.", + "- Extensive evaluations on two face datasets demonstrate that our method achieves more obvious disruption of the user's identity and provides better privacy protection." + ], + "bbox": [ + 76, + 90, + 467, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 227, + 212, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Generative Models and Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 252, + 426, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational autoencoders (VAEs) [21] and Generative adversarial networks (GANs) [8] are popular frameworks among generative models which have strong generative ability. These models encode the data $x$ as latent variables $z$ and model the joint distribution $p_{\\theta}(x,z)$ . However, the quality of VAE samples is not competitive with GANs which are suffering from training instability [10]. Since diffusion probabilistic models (DM) [31] progressively add noise to the data from the joint distribution $q_{\\theta}(x_{0:T}|x_0)$ and denoise step by step, the efficiency of training and quality of samples have achieved state-of-work.", + "bbox": [ + 76, + 275, + 467, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unconditional generative models cannot produce desired samples and then models take different input as guidance have sprung up. Based on GAN, cGAN [25] generates images conditioned on the given labels $y$ and Cycle-GAN [38] implements unpaired image translation considering the given image. Equipped with some techniques like classifier-free guidance [11], the diffusion model gains the ability to follow diverse prompts as conditions during generation.", + "bbox": [ + 76, + 441, + 467, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The open-sourced Latent Diffusion Models (LDMs) [27] operate images in the latent space of low dimensions rather than pixel space which greatly reduces training computation. To support different condition inputs, they add cross-attention layers into the underlying U-Net backbone as conditioning mechanisms $\\tau_{\\theta}$ . The above delicate designs make diffusion more friendly for users to create what they need. It can also assist them in designing their diffusions which synthesize visual contents that contain specific concepts (e.g., objects or styles) given by users during inference.", + "bbox": [ + 76, + 564, + 467, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.Customization", + "text_level": 1, + "bbox": [ + 76, + 724, + 223, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DreamBooth [28] stands out as a popular diffusion-based method for customizing text-to-image generation. This approach involves presenting $3 \\sim 5$ images depicting a specific concept (e.g., a particular dog) alongside a corresponding identifier (e.g., \"a sks dog\"). DreamBooth utilizes these images to fine-tune the pre-trained Stable Diffusion model. This fine-tuning process encourages the model to \"memorize\" the concept and its identifier, enabling it to reproduce this concept in new contexts during inference.", + "bbox": [ + 76, + 750, + 467, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, Textual-inversion [7] employs a dif", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ferent approach. This method freezes the U-Net and exclusively optimizes the text embedding of unique identifiers (e.g. \"sks\") to represent the input concepts.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by DreamBooth, numerous works have been mushroomed such as custom-diffusion [22], Sine [37], among others. In the quest for more efficient fine-tuning, DreamBooth has a successful integration with LoRA [13] and has become a very influential customization project in the community. LoRA specifically decomposes the attention layer of the vision transformer model [6] into low-rank matrices reducing the cost associated with fine-tuning.", + "bbox": [ + 498, + 137, + 890, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Privacy Protection for Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 268, + 841, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the issue that private images are misused by Stable Diffusion based customization, so-called \"anticustomization\", some researchers [23, 30, 36] have recently delved into privacy protection for diffusion models. AdvDM [23] misleads the feature extraction of diffusion models. They analyze the training objective of fine-tuning and propose a direct way that uses the gradient of denoising loss as guidance to optimize the latent variables sampled from the denoising process. The generated adversarial examples degrade the generation ability of DreamBooth or other DM-based customized approaches. Inspired by [14], Antidreambooth [36] uses the alternating surrogate and perturbation learning (ASPL) to approximate the real trained models. They train the model on clean data and use these models as the surrogate model to compute the noise added to the user-provided images. The perturbed images then as the training data for surrogate model fine-tuning which mimics the real scenes. Photoguard [29] specially focuses on unauthorized image inpainting which misleads the public and does harm to personal reputation. They use both VAE encoder attack and UNet attack targeted on a gray image to hinder infringers from creating fake news.", + "bbox": [ + 498, + 292, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some concerns limit the application of current anticom customization methods in more practical scenarios, such as ineffective optimization, poor identity disruption, and simply using the reconstruction loss as guidance to generate adversarial examples [15-19]. Our paper aims to fill these blanks by figuring out the internal mechanisms that affects the performance of protection.", + "bbox": [ + 498, + 625, + 890, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 744, + 589, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 771, + 638, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model (DM) DM is a probabilistic generative model that samples from Gaussian distribution and then progressive denoise to learn the desired data distribution. Given $x_0 \\sim q(x)$ , the forward process adds increasing noise to the input images at each time-step $t \\in (0, T)$ which produces a sequence of noisy latent, $\\{x_0, x_1, \\dots, x_T\\}$ . The backward process trains model $\\epsilon_{\\theta}(x_t, t, c)$ to predict the", + "bbox": [ + 498, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12049", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe297b61c3dc13b8033d9e30ff1cae5949f1d05dc5be630413f952a08339a1d1.jpg", + "image_caption": [ + "Figure 3. Pipeline of SimAC. We first greedily select the time step with our adaptive strategy during the feed forward phase. Then we integrate the feature interference loss with the vanilla training loss as the final objective. The noise is iteratively optimized." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 441, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noise added in $x_{t}$ to infer $x_{t-1}$ . During denoising, the training loss is $l_{2}$ distance, as shown in Eq.(1), Eq.(2). Although there are many implementations of text-to-image diffusion models, Stable Diffusion is one of the few open-sourced diffusion models and is widely used in the community. Thus, our method is mainly based on the checkpoints of stable diffusion provided in HuggingFace.", + "bbox": [ + 75, + 409, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u n c o n d} (\\theta , x _ {0}) = \\mathbb {E} _ {x _ {0}, t, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t + 1}, t) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 527, + 467, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n d}} (\\theta , x _ {0}) = \\mathbb {E} _ {x _ {0}, t, c, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t + 1}, t, c) \\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 547, + 468, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ is the input image, $t$ is the corresponding timestep, $c$ is condition input, (e.g., text or image), $\\epsilon$ is the noise term.", + "bbox": [ + 75, + 578, + 468, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial Attack The adversarial examples for classification are crafted to mislead the model to classify the given input to the wrong labels. However, the traditional attack strategies are not effective when dealing with generative models. For diffusion models, adversarial examples are some images that are added on imperceptible noise, causing diffusion models to consider them out of the generated distribution. In detail, this noise heightens the challenge of image reconstruction and hinders the customization capabilities of applications based on diffusion models (DMs). The optimized noise is often typically constrained to be smaller than a constant value $\\eta$ . The $\\delta$ is determined through the following formation:", + "bbox": [ + 75, + 609, + 468, + 804 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {a d v} = \\arg \\max _ {\\| \\delta \\| _ {p} < \\eta} L \\left(f _ {\\theta} (x + \\delta), y _ {r e a l}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 821, + 468, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ is the input image, $y_{real}$ is the real images, and $L$ is the loss function used to evaluate the performance of adversarial examples.", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Projected Gradient Descent (PGD) [24] is a widely used method to iteratively optimize adversarial examples. The process is formulated as", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {t + 1} = \\prod_ {(x, \\eta)} \\left(x ^ {t} + \\alpha \\operatorname {s g n} \\left(\\nabla_ {x} L \\left(f _ {\\theta} (x + \\delta), y _ {\\text {r e a l}}\\right)\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 147, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x^0 = x$ and $x$ is the input image, $\\operatorname{sgn}(\\cdot)$ is sign function, $(\\nabla_x L(f_\\theta(x + \\delta), y_{real})$ is the gradient of the loss function with respect to $x + \\delta$ . $\\alpha$ represents the step size during each iteration and $t$ is the iteration number. With the operation $\\Pi_{(x,\\eta)}$ , the noise is limited to a $\\eta$ -ball ensuring the adversarial examples are acceptable.", + "bbox": [ + 498, + 191, + 890, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 292, + 609, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we delve into the properties of LDMs and analyze the potential vulnerabilities that can benefit the attack. In Sec. 3.3, we give a comprehensive analysis of the properties. In Sec. 3.4 and 3.5, we propose our method based on these analyses, mainly including two components, i.e., adaptive greedy time interval selection and the feature interference loss. The overall pipeline is illustrated in Figure 3.", + "bbox": [ + 498, + 315, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Analysis on Properties of LDMs", + "text_level": 1, + "bbox": [ + 500, + 430, + 781, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differences at different time step. Our exploration focuses on analyzing gradients across various time intervals from a statistical standpoint. We've established a threshold of $1e^{-10}$ to assess the number of gradients with absolute values below this threshold in perturbed images across different timesteps. Figure 1 illustrates a notable trend: during the latter part of the total time interval $(0, \\text{MaxTrainStep})$ , there's a sharp increase in the count of absolute gradients below the specified threshold. Additionally, Figure 1 presents the maximum value, mean, and median of absolute gradient decreases in the perturbed images when timesteps get larger.", + "bbox": [ + 498, + 454, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown Eq.(5), the forward process admits sampling $x_{t}$ at an arbitrary timestep $t$ . The magnitude of the noise corresponding to timestep is determined by a pre-defined noise schedule. To ensure that the final latent $x_{T}$ conforms to a standard normal distribution, the amplitude of noise injection gradually increases with the timestep.", + "bbox": [ + 498, + 635, + 890, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 736, + 890, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During denoising, we reconstruct the input image $x_0$ based on the noisy latent $z_t$ and visualize the difference between the reconstructed image and the input image in the frequency domain at different timesteps. As shown in Figure 4, when $t \\in [0,100]$ , the main difference between the original and reconstructed images lies in the high-frequency components. Our initial expectation is to disrupt the reconstruction of the original image by introducing high-frequency adversarial noise, aiming to achieve anti-customization. How-", + "bbox": [ + 498, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12050", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b180f286dbb87f5ade78d118a771268a1860d5b4599e12bb08361b550acdbde6.jpg", + "image_caption": [ + "Figure 4. Frequency domain residual analysis. The X-axis is the distance from the low-frequency center (from low frequency to high frequency), and the Y-axis represents total magnitude of frequency residual. It can be seen that when $t$ is small, the high-frequency difference exceeds the low-frequency difference. As $t$ increasing, the low-frequency difference gradually dominated." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 202, + 202 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf95b0908637bcbb0d6b2c087e6fd4f286d1dfaa99c1866df11db03e544a4f07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 334, + 202 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/530f3e19b45f68a81b84385c5e4f2b2e6afc233aa9738b0e20bbd6b0968633e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 88, + 468, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ever, when $t$ increases, such as $t \\in [700,800]$ , the difference in low-frequency components between the reconstructed images and the input images dominates most of the frequency domain differences between the two. Therefore, this can explain why, when the timestep of noise scheduler is large, the gradient of the perturbed image is close to zero and result in ineffective noise optimization.", + "bbox": [ + 75, + 321, + 467, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differences at Different UNet Layers. Inspired by [35], we utilize PCA to visualize the output features of each layer in the U-Net decoder blocks during denoising. The decoder block consists of self-attention, cross-attention, and residual blocks, we select features of residual blocks. There are 11 layers in total, and the layers are visualized at timestep=500. As Figure 5 shows, with the output feature of UNet decoder blocks increasing, the visualized features gradually change from depicting structures and other low-frequency information to capturing texture and similar high-frequency information. Since our noise is intended to disrupt high-frequency information, the deeper features are good perturbation objects for our adversarial noises to reinforce interference which attempts to perturb high-frequency components in generated images.", + "bbox": [ + 75, + 429, + 468, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Adaptive Greedy Time Interval Selection", + "text_level": 1, + "bbox": [ + 76, + 665, + 426, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve more effective privacy protection within a limited number of noise injection steps, we propose a fast adaptive time interval selection strategy. Firstly, based on the input image, we randomly select five timesteps from the interval (0, max training step). For each time step, we compute the gradient concerning the input image. The absolute values of the gradients are summed. If the sum of absolute gradient values is the minimum at timestep $t$ , the corresponding interval $(t - 20, t + 20)$ is then removed. For the timestep with maximal sum, noise is computed and added to the image and this noised image becomes the input for the next round of gradient computation. This process is repeated until the final time interval length is no greater than 100. We obtain the ultimate interval used for noise injection.", + "bbox": [ + 75, + 688, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method relies on PGD[24] (Projected Gradient Descent), an iterative approach for adversarial attacks. At each iteration, computing the gradient becomes pivotal as it signifies the rate of change of the objective function concerning the perturbed images. The gradient approaching zero implies the possibility of iterative optimization but with minimal alterations to the objective function. This ultimately leads to reduced attack efficiency and diminished effectiveness. Thus, randomly selecting timesteps makes it more challenging to optimize the objective function under the same noise budget and results in decreased efficiency and effectiveness in countering customization. This highlights the crucial need for employing our adaptive, greedy time interval selection for perturbation.", + "bbox": [ + 496, + 90, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5784d6a4494f50a286559a152acb8e71b348da39e5cf7385719f39a0840455b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 319, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis on efficacy. Here we simply analyze our adaptive greedy selection for timestep $t$ (the above algorithm). Its key steps in each iteration are randomly sampling 5 timesteps $T_{s} = \\{t_{j}\\}_{j = 1}^{5}$ and rescaling the timestep range by deleting the interval of $t \\in T_{s}$ , which leads to the smallest gradient absolute value. Thus avoiding very small or even zero gradients during optimization.", + "bbox": [ + 496, + 718, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. Suppose the timestep range is rescaled from $T = A \\cup B$ to $T' = (A \\setminus \\Delta A) \\cup (B \\setminus \\Delta B)$ in some iteration, where timestep $t \\in B$ leads to zero gradients, satisfying $A \\cap B = \\emptyset$ , and $\\Delta A \\cup \\Delta B$ denotes the deleted interval, then we have $E_{t \\sim T} |\\nabla_x \\mathcal{L}_{DM}| < E_{T_s \\sim T} E_{t \\sim T'} |\\nabla_x \\mathcal{L}_{DM}|$ .", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5d9fe4e4f9da23b2e0b620234bd8f65b8e09fd3ef4d9a870a1c10b089610a02b.jpg", + "image_caption": [ + "Figure 5. PCA visualization of features. The output feature of residual blocks are visualized at timestep=500. As feature goes deeper, the high-frequency information captured by the feature becomes more significant." + ], + "image_footnote": [], + "bbox": [ + 73, + 88, + 138, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/116e7b02e75e4334c2c7946a0286cff766c215cb7ae32d3e2410c5cc77029158.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 142, + 89, + 202, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64177697146243cc0dbea35ed76c7bb8aef09cd84df0cfbf8a0d5751ec0746ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 89, + 263, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba0c9dc5d1db9af27eb35b89aa83c6737af78410f1b0a6192e115d83dcc865c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 89, + 325, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff8485302aef3db6bb0e14dd75463e7a6a17e9a5fe7b33375f8167b7933f82e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 89, + 388, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81bfe356bbb90efecb14e5f8d0d82c13c5f694a294c14c9e52f41be9fec551bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 89, + 449, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0147ae25280bb7572f5d0f4bd8955839919131077fd2cb927db82f64e5b75593.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 89, + 514, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44c8de6128b2fe3ca14c8003032019eee927ca19a442137ec4087ffb0b4c08d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 89, + 573, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c451000ed1b6dafe575399d49c13e852f8479262c99e82229359c54fe18a4524.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 89, + 638, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16a4572380c47ad0d89b00869fc0ce44339af38e5d748ad942696b998b7c2e5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 89, + 700, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/156a615e14979e42c83e29cb699f7a3a154d5e710e3169e218733d1910c0c385.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 89, + 759, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e1e4a09a348669b0044e1ac20048400d708a77443437b1f4a8c384bbf26bd4b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 89, + 823, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/819a421bc3b6409f91a1d95fc267038b34da33f1643dccc4dac8564bab384305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 89, + 883, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "portrait image", + "bbox": [ + 71, + 141, + 147, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 0", + "bbox": [ + 151, + 142, + 189, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 1", + "bbox": [ + 215, + 142, + 251, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 2", + "bbox": [ + 277, + 142, + 313, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 3", + "bbox": [ + 339, + 142, + 377, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 4", + "bbox": [ + 401, + 142, + 439, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 5", + "bbox": [ + 465, + 142, + 501, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 6", + "bbox": [ + 526, + 142, + 563, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 7", + "bbox": [ + 589, + 142, + 625, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 8", + "bbox": [ + 651, + 142, + 687, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 9", + "bbox": [ + 714, + 142, + 750, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 10", + "bbox": [ + 772, + 142, + 815, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layer 11", + "bbox": [ + 834, + 142, + 875, + 151 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f5d71dedf0c5c2c86784552847d978356aff5774f8c33246ed53803c47c790a.jpg", + "image_caption": [ + "Figure 6. Diffusion feature over different timesteps. As the timestep of the noise scheduler increasing, part of the high-frequency information is lost and it is not good for adding high-frequency adversarial noise. This means that choosing an appropriate time sequence is also beneficial to feature interference loss." + ], + "image_footnote": [ + "portrait image t ∈ [100,200] t ∈ [300,400] t ∈ [500,600] t ∈ [700,800]" + ], + "bbox": [ + 78, + 209, + 151, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63ef16559c62b14d2d9da95e2c0eac01d95f7b5765ddf45a1b328f079e374610.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 156, + 209, + 228, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d49296633c15a35e5ab55cbae7120afe17a1cb24d3179a52794e25f16014ede.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 209, + 305, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d515fe086cc9e37c4c5937e54750a786496f96bde09de713c2182bb422bc7169.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 209, + 385, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e20609d520f2196d2fc86389f1cb4e3564fd9d1e37b1b5735af331f0f67e7e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 209, + 462, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. We can easily obtain $E_{T_s \\sim T} \\frac{|\\Delta A|}{|A|} < E_{T_s \\sim T} \\frac{|\\Delta B|}{|B|}$ via proof by contradiction. For simplicity, we denote $\\nabla_x \\mathcal{L}_{DM}$ as $g(t)$ , $LHS$ and $RHS$ denote left and right, then", + "bbox": [ + 75, + 376, + 468, + 426 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L H S = p (t \\in A) E _ {t \\sim A} | g (t) | + p (t \\in B) E _ {t \\sim B} | g (t) | \\\\ = \\frac {| A |}{| A | + | B |} E _ {t \\sim A} | g (t) | \\approx E _ {T _ {s} \\sim T} \\frac {| A | \\cdot E _ {t \\sim A \\backslash \\Delta A} | g (t) |}{| A | + | B |} \\\\ < E _ {T _ {s} \\sim T} \\frac {| A | - | \\Delta A |}{| A | + | B | - | \\Delta A | - | \\Delta B |} E _ {t \\sim A \\backslash \\Delta A} | g (t) | \\\\ = E _ {T _ {s} \\sim T} E _ {t \\sim A \\backslash \\Delta A} p (t \\in A \\backslash \\Delta A) | g (t) | = R H S \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 433, + 485, + 539 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b71cf575c980efa973f49ea1af1636b642ade1055e787c1a3a3fec55774f825.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 547, + 467, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Feature Interference Loss", + "text_level": 1, + "bbox": [ + 76, + 570, + 310, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the analysis in Sec 3.3, we propose the feature interference loss and integrate it with the vanilla training loss as the overall objective for optimization. This loss calculates the layer-wise euclidean distance between the intermediate features in some specially selected layers, i.e.,", + "bbox": [ + 75, + 593, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\mathbb {E} \\| f _ {l} ^ {t ^ {*}} - f _ {l} ^ {t} \\| _ {2} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 672, + 468, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n d}} + \\lambda \\mathcal {L} _ {f} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 695, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $f_{l}^{t^{*}}$ is the output features of the selected layers sets $l$ at timestep $t$ for the input image, $f_{l}^{t}$ is the output features for current perturbed images, and $\\lambda$ represents the weighting coefficient for the feature interference loss.", + "bbox": [ + 75, + 718, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We visualized the output features of layer 7 at different time steps. As Figure 6 shows, when the timesteps increase, more noise is added to the latent and the response to the high-frequency components of the input image gradually decreases. Therefore, utilizing feature interference loss as the optimization objective for adversarial noise in small time intervals is a better choice to keep the effectiveness of adversarial noise disturbing high-frequency information.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 210, + 630, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 236, + 581, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset We utilized two facial datasets for experiments including Celeb-HQ [20] and VGGFace2 dataset[3]. The dataset comprises about 50 individuals and aligns with the Anti-DreamBooth, with each individual including at least 15 clear face images for customization. Since Anti-dreambooth needs alternating training, the dataset is divided into three sets, including set A, set B, and set C and each set contains 5 portrait images.", + "bbox": [ + 496, + 261, + 890, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Since the open-sourced stable diffusion is the most popular implementation of latent diffusion among the community, our experiments mainly use the SD-v2.1 [1]. To test the performance of our method in a black-box scenario, we assume the versions of Stable Diffusion between anticustomization and customization are the same or different.", + "bbox": [ + 496, + 383, + 890, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline We compare several open-source methods that employ adversarial attacks on diffusion models to protect user images from being misused by text-to-image diffusion models, including Photoguard [29], AdvDM [23] and AntiDreamBooth [36]. Due to the high GPU memory requirements of the complete PhotoGuard, we only utilize the VAE encoder attack strategy in its paper for comparison.", + "bbox": [ + 496, + 474, + 890, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric We utilize a face detector named Retinaface [5] to detect if there is a face in the image and use the Face Detection Failure Rate (FDFR) to assess the level of disruption to the generated faces. Upon detecting a face, we encode it using ArcFace [4] and calculate the cosine similarity between the protected image and clean input, measuring the identity resemblance between the detected face in the generated image and that of users. This matrix is defined as Identity Score Matching (ISM). In addition, the image quality is quantified by BRISQE [26], and the quality of detected facial images is measured through SER-FIQ [34].", + "bbox": [ + 496, + 582, + 890, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details We set the same noise budget for all methods $\\eta = 16 / 255$ . Additionally, the optimization steps and step size align with the settings specified in each baseline. The number of training epochs is 50 and each epoch includes 3 steps for surrogate model training and 9 steps for adversarial noise optimization. The step size for adding noise is set as 0.005 and the learning rate for the training model is set as 5e-7 which is appropriate for human faces. The maximum adaptive greedy search steps are set as 50 by default. After training, we used the noised im", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12052", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2f0a66adeb55d3abb46e41b0465943a5f7bb61688439127cff21993f33c1b2a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CelebA-HQ
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2541.0919.380.55
AdvDM [23]0.3270.4838.170.20
Anti-DB [36]0.2877.2837.430.20
Anti-DB + SimAC0.3187.0738.860.21
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2028.5029.330.59
AdvDM [23]0.2565.3737.860.41
Anti-DB [36]0.1986.8038.900.27
Anti-DB + SimAC0.1296.8742.100.15
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.1830.0726.960.40
AdvDM[23]0.2935.1036.460.36
Anti-DB[36]0.2242.8640.340.28
Anti-DB + SimAC0.1291.9043.970.06
Method“a photo of sks person in front of eiffel tower”
ISM ↓FDFR ↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.0850.9532.820.40
AdvDM[23]0.0938.1036.020.30
Anti-DB[36]0.0656.2641.350.22
Anti-DB + SimAC0.0566.1942.770.12
", + "bbox": [ + 86, + 88, + 455, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "age for customization. The default base model is stable diffusion v2.1 combined with DreamBooth. According to the above analysis of the middle layer of the model, we select the 9, 10, and 11 layer features that can best represent the high-frequency information of the image for additional disturbance. The weight $\\lambda$ of feature interference loss is set to 1. After finetuning 1000 steps, we save the model checkpoints and conduct inference. For each prompt, 30 images in .png format are generated for metric calculation.", + "bbox": [ + 75, + 474, + 468, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with Baseline Methods", + "text_level": 1, + "bbox": [ + 76, + 619, + 387, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results To test the effectiveness of our approach in enhancing the protection of users' portrait images, we conduct a quantitative comparison under three prompts in Table 1 and Table 2. To be more practical, we list four text prompts for inference, the first one \"a photo of sks person\" is the same as the prompt used in training, while the other three were prompts unseen before. For each prompt, we randomly select 30 generated images to compute each metric, reporting their average values.", + "bbox": [ + 75, + 643, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can find that our method greatly improves the performance of Anti-DreamBooth and outperforms other baselines across all prompts. Due to our analysis of the unique properties of the diffusion model during denoising and effective addition of high-frequency noise, the face detection failure rates greatly increase, and the identity matching scores between the detected face and the input image are the lowest among all methods. This implies that our method is", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3d7f416f3f3fa48d5c1c5f56953b193404017df76b952c2f5211d9f6e54f893.jpg", + "table_caption": [ + "Table 1. Comparison with other open-sourced anti-customization methods on CelebA-HQ. We evaluate the performance under four different prompts during customization." + ], + "table_footnote": [], + "table_body": "
VGG-Face2
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2929.2720.670.47
AdvDM [23]0.3263.0738.510.21
Anti-DB[36]0.3064.6737.890.22
Anti-DB + SimAC0.3180.2740.710.22
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2517.3328.520.55
AdvDM [23]0.3067.8037.580.35
Anti-DB[36]0.2377.4738.790.29
Anti-DB + SimAC0.1196.3341.780.12
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1740.8730.100.32
AdvDM[23]0.2737.7335.430.29
Anti-DB[36]0.2545.0039.250.27
Anti-DB + SimAC0.1389.6044.970.07
Method“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1351.0730.690.41
AdvDM[23]0.1438.6735.990.31
Anti-DB[36]0.1054.9341.130.23
Anti-DB + SimAC0.0961.2042.170.10
", + "bbox": [ + 516, + 88, + 874, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with other open-sourced anti-customization methods on VGG-Face2. We evaluate the performance under four different prompts during customization.", + "bbox": [ + 498, + 417, + 890, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "more effective in resisting abuse from customization.", + "bbox": [ + 498, + 474, + 851, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results We show some results in Figure 7. It's evident that SimAC combined with Anti-DreamBooth achieves a strong image disruption effect, providing the best privacy protection for the input portrait. Since PhotoGuard operates an attack on the latent space, the resulting images tend to generate patterns similar to the target latent. This attack doesn't effectively reduce the probability of facial appearances or improve the quality of generated images.", + "bbox": [ + 496, + 491, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Both AdvDM and Anti-DreamBooth use DM's training loss as an objective for optimizing noise, which, as been seen, yields similar results. The two methods aim to disrupt high-frequency components in images via adversarial noise to create artifacts and raise the bar for potential misuse of users' photos. Although the quality of generated images decreases, the results of both AdvDM and Anti-DreamBooth retain many details from the user input images. This indicates a leak of privacy for some users and isn't user-friendly. According to our analysis, both of these methods follow the training process of DM, randomly selecting timestamps for optimization. This leads to a waste of some noise addition steps, thereby reducing the efficiency of the attack and the effectiveness of protection.", + "bbox": [ + 496, + 611, + 892, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Black-Box Performance", + "text_level": 1, + "bbox": [ + 500, + 832, + 718, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt Mismatch When attackers use stable diffusion to customize concepts, the prompt they use might differ from our assumptions when adding noise. Thus, we use the", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4057c7074a43fff26ebeaa1d6ae7e6bcc73229e3bb5269b885eba01db66293d.jpg", + "image_caption": [ + "Portrait Image", + "Figure 7. Visualization results (four prompts) on CelebA-HQ. From the first row to the last row is \"a photo of sks person\", \"a dslr portrait of sks person\", \"a photo of sks person looking at the mirror\", \"a photo of sks person in front of eiffel tower\"." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 151, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14a263c12d2f12c6adc67fd88ba240969415fdb73d44fd01acd80ae499582741.jpg", + "image_caption": [ + "PhotoGuard" + ], + "image_footnote": [], + "bbox": [ + 156, + 88, + 230, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d06941ba839fd1308cac69c8f2c70dd5c47c38a2cbf8983787b5af56c84ceef3.jpg", + "image_caption": [ + "AdvDM" + ], + "image_footnote": [], + "bbox": [ + 235, + 88, + 310, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e80d6edf48e8bf88993ce14eb806fe367f93fa99f29eaa509ea2f9a3c2cbff57.jpg", + "image_caption": [ + "Anti-DB" + ], + "image_footnote": [], + "bbox": [ + 315, + 88, + 388, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e4653885dff5433de2cbc5800881cb6eaf6642f2a264a7f7a6717dfd402da65.jpg", + "image_caption": [ + "Anti-DB+SimAC" + ], + "image_footnote": [], + "bbox": [ + 393, + 88, + 468, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"a photo of sks person\" during perturbation learning and change the rare identifier \"sks\" to \"t@t\" during DreamBooth model finetuing. The results in Table 3 indicate a decrease in performance in this scenario, but we notice that the Identity Score Matching(ISM), representing identity similarity remains consistent.", + "bbox": [ + 75, + 636, + 468, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Mismatch The models used to add adversarial noise may also mismatch with the model fine-tuned by DreamBooth. We examine the effectiveness of adversarial noise learned on stable diffusion v2.1 against customization based on stable diffusion v2.1 and v1.4, or vice versa in Table 4. The conclusion of model mismatch resembles those of prompt mismatch. Although there is an overall decline in performance, the critical matrix, ISM, remains considerable. This also implies that in a model mismatch scenario, SimAC + Anti-DreamBooth can still protect user portrait privacy.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6e7a32a55323b8a230886b4cf6f579483eea2f8e75e000418218fb5118a83174.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train [v]Test [v]“a photo of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.3187.0738.860.21
skst@t0.3081.3639.670.31
Train [v]Test [v]“a dslr portrait of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1296.8742.100.15
skst@t0.2354.4237.770.52
Train [v]Test [v]“a photo of [v] person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1291.9043.970.06
skst@t0.1630.5440.630.25
Train [v]Test [v]“a photo of [v] person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.0566.1942.770.12
sksst@t0.0623.2037.740.28
", + "bbox": [ + 509, + 88, + 880, + 318 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1ef84162be17a85cf6588ed1c620435cd72c5dcf2658bca3d5f0fcd8c7e061af.jpg", + "table_caption": [ + "Table 3. Prompt mismatch between training and testing on CelebA-HQ. The training prompt is \"a photo of sks person\" and the inference prompt uses rare identifier \"sks\" or \"t@t\"." + ], + "table_footnote": [], + "table_body": "
TrainTest“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.3187.0738.860.21
v1.4v2.10.3870.0739.220.35
v2.1v1.40.0199.8662.090.02
TrainTest“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1296.8742.100.15
v1.4v2.10.1495.1740.950.23
v2.1v1.40.2620.5416.440.57
TrainTest“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1291.9043.970.06
v1.4v2.10.1679.3942.520.13
v2.1v1.40.2082.7240.120.16
TrainTest“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.0566.1942.770.12
v1.4v2.10.0661.3641.510.10
v2.1v1.40.0853.6720.560.22
", + "bbox": [ + 524, + 371, + 864, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Model version mismatch during training and testing on CelebA-HQ. The training prompt is \"a photo of sks person\"", + "bbox": [ + 498, + 643, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 691, + 617, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Current anti-customization methods, primarily relying on adversarial attacks, often overlook crucial internal properties of the diffusion model, leading to ineffective optimization. This paper addresses this issue by exploring inherent properties to enhance anti-customization. Two key aspects are examined: Analyzing the relationship between timestep selection and the model's perception in the frequency domain inspires an adaptive greedy search for optimal timesteps. Scrutinizing feature roles during denoising, resulting in a sophisticated feature-based optimization framework. Experiments show increased identity disruption, enhancing user privacy and security.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12054", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stable-diffusion-2-1. https://huggingface.co/stabilityai/stable-diffusion-2-1.6", + "[2] Making pictures of trump getting arrested while waiting for trump's arrest. https://twitter.com/EliotHiggins/status/1637927681734987777. 1", + "[3] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6", + "[4] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 6", + "[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 6", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3", + "[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015. 2", + "[10] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 3", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 3", + "[14] Qidong Huang, Jie Zhang, Wenbo Zhou, Weiming Zhang, and Nenghai Yu. Initiative defense against facial manipula" + ], + "bbox": [ + 78, + 114, + 467, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1619-1627, 2021. 3", + "[15] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, and Nenghai Yu. Shape-invariant 3d adversarial point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15335-15344, 2022. 3", + "[16] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, Kui Zhang, Gang Hua, and Nenghai Yu. Pointcat: Contrastive adversarial training for robust point cloud recognition. arXiv preprint arXiv:2209.07788, 2022.", + "[17] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Yinpeng Chen, Lu Yuan, Gang Hua, Weiming Zhang, and Nenghai Yu. Improving adversarial robustness of masked autoencoders via test-time frequency-domain prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1600-1610, 2023.", + "[18] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Weiming Zhang, Feifei Wang, Gang Hua, and Nenghai Yu. Diversity-aware meta visual prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10878-10887, 2023.", + "[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. arXiv preprint arXiv:2311.17911, 2023. 3", + "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 6", + "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3", + "[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, pages 20763-20786. 2, 3, 6, 7", + "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In 6th International Conference on Learning Representations, ICLR 2018. 4, 5", + "[25] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 3", + "[26] A. Mittal, A. K. Moorthy, and A. C. Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, page 4695-4708, 2012. 6", + "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12055", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 3", + "[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, pages 29894-29918. 3, 6, 7", + "[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 3", + "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3", + "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR, 2021. 1", + "[33] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2013. 2", + "[34] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. IEEE, 2020. 6", + "[35] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5", + "[36] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3, 6, 7", + "[37] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 3", + "[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + ], + "bbox": [ + 78, + 90, + 468, + 878 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12056", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_model.json b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5014feb75398aa8a94d262ec3d74d450d0229b21 --- /dev/null +++ b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "SimAC: A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.203, + 0.804, + 0.222 + ], + "angle": 0, + "content": "Feifei Wang\\(^{1,2,*}\\); Zhentao Tan\\(^{2,1}\\), Tianyi Wei\\(^{1}\\), Yue Wu\\(^{2}\\), Qidong Huang\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.222, + 0.75, + 0.24 + ], + "angle": 0, + "content": "1University of Science and Technology of China 2Alibaba Cloud" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.242, + 0.85, + 0.258 + ], + "angle": 0, + "content": "{wangfeifei@, tzt@, bestwty@, hqd0037@}mail.ustc.edu.cn matthew.wy@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.765 + ], + "angle": 0, + "content": "Despite the success of diffusion-based customization methods on visual content creation, increasing concerns have been raised about such techniques from both privacy and political perspectives. To tackle this issue, several anti-customization methods have been proposed in very recent months, predominantly grounded in adversarial attacks. Unfortunately, most of these methods adopt straightforward designs, such as end-to-end optimization with a focus on adversarially maximizing the original training loss, thereby neglecting nuanced internal properties intrinsic to the diffusion model, and even leading to ineffective optimization in some diffusion time steps. In this paper, we strive to bridge this gap by undertaking a comprehensive exploration of these inherent properties, to boost the performance of current anti-customization approaches. Two aspects of properties are investigated: 1) We examine the relationship between time step selection and the model's perception in the frequency domain of images and find that lower time steps can give much more contributions to adversarial noises. This inspires us to propose an adaptive greedy search for optimal time steps that seamlessly integrates with existing anti-customization methods. 2) We scrutinize the roles of features at different layers during denoising and devise a sophisticated feature-based optimization framework for anti-customization. Experiments on facial benchmarks demonstrate that our approach significantly increases identity disruption, thereby protecting user privacy and copyright. Our code is available at: https://github.com/somuchtoe/SimAC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.792, + 0.21, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.47, + 0.865 + ], + "angle": 0, + "content": "Latent Diffusion model (LDM) [12, 31, 32] has been recently proven as a strong paradigm for photorealistic visual content generation. The emergence of open-source Stable" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.291, + 0.892, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.533, + 0.894, + 0.562 + ], + "angle": 0, + "content": "Figure 1. Comparison between Anti-DreamBooth before and after adding SimAC. Our method further boosts its ability to de-identity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Diffusion encourages users to explore creative possibilities with LDMs. Users only need to provide several images representing the same subject along with a rare identifier to customize their diffusion models [7, 13, 22, 28]. After finetuning the model or optimizing the text embedding of rare identifiers, it can flexibly generate high-quality images containing the specified object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The convenience of customizing large-scale text-to-image models also allows the malicious users to generate forgery images that violate the truth. Some of them may use such technique to steal others' painting styles and generate new painting content without permission. While some of them may collect one's portrait photos that are published on the social platform, and generate fake images of this person through customization. In other words, the lawbreakers can easily produce fake news that contains celebrity photos [2], greatly misleading the public. The infringement poses threats to user privacy and intellectual property. Hence, it is essential to devise effective countermeasures to safeguard users against such malicious usage." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.368, + 0.888 + ], + "angle": 0, + "content": "*Work done during an internship in Alibaba Cloud." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.368, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12047" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.333, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.248, + 0.324, + 0.261 + ], + "angle": 0, + "content": "(a) the number of absolute gradients \\(< 1e^{-10}\\)" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.088, + 0.593, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.251, + 0.555, + 0.263 + ], + "angle": 0, + "content": "(b) the maximum absolute gradients" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.089, + 0.885, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.251, + 0.863, + 0.262 + ], + "angle": 0, + "content": "(c) the mean and median of the absolute gradients" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Figure 2. Distribution of Anti-DreamBooth attack gradients on different diffusion timesteps, where (a) counts the number of absolute gradients below the threshold \\(1e^{-10}\\) at each timestep, (b) presents the maximum absolute gradients at different timesteps, and (c) demonstrates how the mean and median of the absolute gradients change over timesteps. Apparently, the absolute gradients show great discrepancy on the varying timesteps and nearly zero values appear at large timesteps, which leads to ineffective noise optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.535 + ], + "angle": 0, + "content": "Current anti-customization methods are generally based on adversarial attack [9, 33]. AdvDM [23] is the pioneering work that uses adversarial noise to protect user images from being customized by diffusion models. It ingeniously combines the diffusion model with adversarial samples, firstly achieving user privacy protection. Anti-DreamBooth [36] further enhances the protective effect by employing alternate training. Both of them randomly select the timesteps from (0, max denoising steps) regarding the noise added to LDM latent, and directly employ the maximization of the LDM training loss as the optimization objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.717 + ], + "angle": 0, + "content": "However, as showcased in Figure 2, we notice that the gradient of the perturbed images is quite small and even zero when the randomly sampled timesteps are relatively large, i.e., the noisy latent is closer to Gaussian noise. This implies that within the limited steps of an adversarial attack, a large portion of steps are ineffective since the perturbed image cannot optimize the objective with such zero gradients, leading to a decrease in both protection effectiveness and efficiency. Hence, the images protected by these current methods, when customized, still allow the model to capture many details from user-input images and leak the user's privacy, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.853 + ], + "angle": 0, + "content": "The fundamental reason is that these methods fail to combine the adversarial attacks with properties inherent in diffusion models. We aim to conduct an in-depth analysis of why the gradients of perturbed images have such discrepancy at different time steps and how diffusion models perceive input images at different intermediate layers. Then, based on the in-depth analysis, we propose improvements to existing customization methods from both the temporal and feature dimensions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To observe the relationship between timesteps and the gradient of the perturbed images, we first reconstruct images predicted from noisy latents at different time steps and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.776 + ], + "angle": 0, + "content": "compare them with the input images. Considering regular adversarial noises mainly affect the high frequency of images, our comparison is conducted in the frequency domain, aiming at investigating whether the model's perception lies on lower-level or higher-level. We observe that, the model focuses on the higher frequency components of images when selecting smaller time steps, and vice versa. Therefore, introducing adversarial noises at larger timesteps is ineffective, since the subtle changes perturbed on images can hardly affect the low-frequency of the generated images. To improve the effectiveness, we propose an adaptive timestep selection method to find optimal time intervals, where we iteratively update the range of selection in a greedy way. To explore the effect of different layers' features in the U-Net decoder during denoising, we employ PCA (Principal Component Analysis) to visualize them and show the discrepancy. We clearly find that within the decoder, the feature extraction gradually shifts from low-frequency to high-frequency as the layer goes deeper. It indicates the higher layers concentrates more on the texture of images, while the lower layers focus on the structure. Consequently, we select the features representing high-frequency information during optimization, since regular adversarial noises concentrate on the high frequency of images. And we construct the feature interference loss and integrate it with the diffusion denoising loss as the objective function, to improve the ability of identity interference." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.779, + 0.775, + 0.793 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.796, + 0.892, + 0.884 + ], + "angle": 0, + "content": "- We reveal the inadequate optimization steps that exist in current anti-customization methods, and gives detailed analysis regarding the perceptual discrepancy of diffusion models at different timesteps and intermediate layers, we have better aligned the optimization of adversarial gradients with diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.887, + 0.891, + 0.901 + ], + "angle": 0, + "content": "- Based on analysis, we propose a simple but effective anti" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12048" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "customization method, including adaptive greedy time selection and a feature interference loss to improve the protection ability. Our method can be easily generalized to different anti-customization frameworks and improve their performance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.168, + 0.468, + 0.213 + ], + "angle": 0, + "content": "- Extensive evaluations on two face datasets demonstrate that our method achieves more obvious disruption of the user's identity and provides better privacy protection." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.092, + 0.468, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.228, + 0.214, + 0.243 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.253, + 0.428, + 0.267 + ], + "angle": 0, + "content": "2.1. Generative Models and Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.276, + 0.468, + 0.441 + ], + "angle": 0, + "content": "Variational autoencoders (VAEs) [21] and Generative adversarial networks (GANs) [8] are popular frameworks among generative models which have strong generative ability. These models encode the data \\( x \\) as latent variables \\( z \\) and model the joint distribution \\( p_{\\theta}(x,z) \\). However, the quality of VAE samples is not competitive with GANs which are suffering from training instability [10]. Since diffusion probabilistic models (DM) [31] progressively add noise to the data from the joint distribution \\( q_{\\theta}(x_{0:T}|x_0) \\) and denoise step by step, the efficiency of training and quality of samples have achieved state-of-work." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.443, + 0.468, + 0.563 + ], + "angle": 0, + "content": "Unconditional generative models cannot produce desired samples and then models take different input as guidance have sprung up. Based on GAN, cGAN [25] generates images conditioned on the given labels \\( y \\) and Cycle-GAN [38] implements unpaired image translation considering the given image. Equipped with some techniques like classifier-free guidance [11], the diffusion model gains the ability to follow diverse prompts as conditions during generation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.565, + 0.468, + 0.715 + ], + "angle": 0, + "content": "The open-sourced Latent Diffusion Models (LDMs) [27] operate images in the latent space of low dimensions rather than pixel space which greatly reduces training computation. To support different condition inputs, they add cross-attention layers into the underlying U-Net backbone as conditioning mechanisms \\(\\tau_{\\theta}\\). The above delicate designs make diffusion more friendly for users to create what they need. It can also assist them in designing their diffusions which synthesize visual contents that contain specific concepts (e.g., objects or styles) given by users during inference." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.726, + 0.225, + 0.74 + ], + "angle": 0, + "content": "2.2.Customization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.468, + 0.884 + ], + "angle": 0, + "content": "DreamBooth [28] stands out as a popular diffusion-based method for customizing text-to-image generation. This approach involves presenting \\(3 \\sim 5\\) images depicting a specific concept (e.g., a particular dog) alongside a corresponding identifier (e.g., \"a sks dog\"). DreamBooth utilizes these images to fine-tune the pre-trained Stable Diffusion model. This fine-tuning process encourages the model to \"memorize\" the concept and its identifier, enabling it to reproduce this concept in new contexts during inference." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "On the other hand, Textual-inversion [7] employs a dif" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "ferent approach. This method freezes the U-Net and exclusively optimizes the text embedding of unique identifiers (e.g. \"sks\") to represent the input concepts." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.138, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Inspired by DreamBooth, numerous works have been mushroomed such as custom-diffusion [22], Sine [37], among others. In the quest for more efficient fine-tuning, DreamBooth has a successful integration with LoRA [13] and has become a very influential customization project in the community. LoRA specifically decomposes the attention layer of the vision transformer model [6] into low-rank matrices reducing the cost associated with fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.269, + 0.842, + 0.285 + ], + "angle": 0, + "content": "2.3. Privacy Protection for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.293, + 0.892, + 0.625 + ], + "angle": 0, + "content": "To alleviate the issue that private images are misused by Stable Diffusion based customization, so-called \"anticustomization\", some researchers [23, 30, 36] have recently delved into privacy protection for diffusion models. AdvDM [23] misleads the feature extraction of diffusion models. They analyze the training objective of fine-tuning and propose a direct way that uses the gradient of denoising loss as guidance to optimize the latent variables sampled from the denoising process. The generated adversarial examples degrade the generation ability of DreamBooth or other DM-based customized approaches. Inspired by [14], Antidreambooth [36] uses the alternating surrogate and perturbation learning (ASPL) to approximate the real trained models. They train the model on clean data and use these models as the surrogate model to compute the noise added to the user-provided images. The perturbed images then as the training data for surrogate model fine-tuning which mimics the real scenes. Photoguard [29] specially focuses on unauthorized image inpainting which misleads the public and does harm to personal reputation. They use both VAE encoder attack and UNet attack targeted on a gray image to hinder infringers from creating fake news." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.626, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Some concerns limit the application of current anticom customization methods in more practical scenarios, such as ineffective optimization, poor identity disruption, and simply using the reconstruction loss as guidance to generate adversarial examples [15-19]. Our paper aims to fill these blanks by figuring out the internal mechanisms that affects the performance of protection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.746, + 0.591, + 0.761 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.64, + 0.786 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Diffusion Model (DM) DM is a probabilistic generative model that samples from Gaussian distribution and then progressive denoise to learn the desired data distribution. Given \\( x_0 \\sim q(x) \\), the forward process adds increasing noise to the input images at each time-step \\( t \\in (0, T) \\) which produces a sequence of noisy latent, \\( \\{x_0, x_1, \\dots, x_T\\} \\). The backward process trains model \\( \\epsilon_{\\theta}(x_t, t, c) \\) to predict the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12049" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.442, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.326, + 0.471, + 0.381 + ], + "angle": 0, + "content": "Figure 3. Pipeline of SimAC. We first greedily select the time step with our adaptive strategy during the feed forward phase. Then we integrate the feature interference loss with the vanilla training loss as the final objective. The noise is iteratively optimized." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.515 + ], + "angle": 0, + "content": "noise added in \\( x_{t} \\) to infer \\( x_{t-1} \\). During denoising, the training loss is \\( l_{2} \\) distance, as shown in Eq.(1), Eq.(2). Although there are many implementations of text-to-image diffusion models, Stable Diffusion is one of the few open-sourced diffusion models and is widely used in the community. Thus, our method is mainly based on the checkpoints of stable diffusion provided in HuggingFace." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.528, + 0.468, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u n c o n d} (\\theta , x _ {0}) = \\mathbb {E} _ {x _ {0}, t, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t + 1}, t) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.549, + 0.469, + 0.566 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n d}} (\\theta , x _ {0}) = \\mathbb {E} _ {x _ {0}, t, c, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t + 1}, t, c) \\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.608 + ], + "angle": 0, + "content": "where \\( x \\) is the input image, \\( t \\) is the corresponding timestep, \\( c \\) is condition input, (e.g., text or image), \\( \\epsilon \\) is the noise term." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.805 + ], + "angle": 0, + "content": "Adversarial Attack The adversarial examples for classification are crafted to mislead the model to classify the given input to the wrong labels. However, the traditional attack strategies are not effective when dealing with generative models. For diffusion models, adversarial examples are some images that are added on imperceptible noise, causing diffusion models to consider them out of the generated distribution. In detail, this noise heightens the challenge of image reconstruction and hinders the customization capabilities of applications based on diffusion models (DMs). The optimized noise is often typically constrained to be smaller than a constant value \\(\\eta\\). The \\(\\delta\\) is determined through the following formation:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.822, + 0.469, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\delta_ {a d v} = \\arg \\max _ {\\| \\delta \\| _ {p} < \\eta} L \\left(f _ {\\theta} (x + \\delta), y _ {r e a l}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\( x \\) is the input image, \\( y_{real} \\) is the real images, and \\( L \\) is the loss function used to evaluate the performance of adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "Projected Gradient Descent (PGD) [24] is a widely used method to iteratively optimize adversarial examples. The process is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.148, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\[\nx ^ {t + 1} = \\prod_ {(x, \\eta)} \\left(x ^ {t} + \\alpha \\operatorname {s g n} \\left(\\nabla_ {x} L \\left(f _ {\\theta} (x + \\delta), y _ {\\text {r e a l}}\\right)\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.193, + 0.892, + 0.284 + ], + "angle": 0, + "content": "where \\( x^0 = x \\) and \\( x \\) is the input image, \\( \\operatorname{sgn}(\\cdot) \\) is sign function, \\( (\\nabla_x L(f_\\theta(x + \\delta), y_{real}) \\) is the gradient of the loss function with respect to \\( x + \\delta \\). \\( \\alpha \\) represents the step size during each iteration and \\( t \\) is the iteration number. With the operation \\( \\Pi_{(x,\\eta)} \\), the noise is limited to a \\( \\eta \\)-ball ensuring the adversarial examples are acceptable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.293, + 0.61, + 0.307 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Here we delve into the properties of LDMs and analyze the potential vulnerabilities that can benefit the attack. In Sec. 3.3, we give a comprehensive analysis of the properties. In Sec. 3.4 and 3.5, we propose our method based on these analyses, mainly including two components, i.e., adaptive greedy time interval selection and the feature interference loss. The overall pipeline is illustrated in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.782, + 0.447 + ], + "angle": 0, + "content": "3.3. Analysis on Properties of LDMs" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.455, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Differences at different time step. Our exploration focuses on analyzing gradients across various time intervals from a statistical standpoint. We've established a threshold of \\(1e^{-10}\\) to assess the number of gradients with absolute values below this threshold in perturbed images across different timesteps. Figure 1 illustrates a notable trend: during the latter part of the total time interval \\((0, \\text{MaxTrainStep})\\), there's a sharp increase in the count of absolute gradients below the specified threshold. Additionally, Figure 1 presents the maximum value, mean, and median of absolute gradient decreases in the perturbed images when timesteps get larger." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.727 + ], + "angle": 0, + "content": "As shown Eq.(5), the forward process admits sampling \\( x_{t} \\) at an arbitrary timestep \\( t \\). The magnitude of the noise corresponding to timestep is determined by a pre-defined noise schedule. To ensure that the final latent \\( x_{T} \\) conforms to a standard normal distribution, the amplitude of noise injection gradually increases with the timestep." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.737, + 0.892, + 0.755 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "During denoising, we reconstruct the input image \\( x_0 \\) based on the noisy latent \\( z_t \\) and visualize the difference between the reconstructed image and the input image in the frequency domain at different timesteps. As shown in Figure 4, when \\( t \\in [0,100] \\), the main difference between the original and reconstructed images lies in the high-frequency components. Our initial expectation is to disrupt the reconstruction of the original image by introducing high-frequency adversarial noise, aiming to achieve anti-customization. How-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12050" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.203, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.336, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.089, + 0.47, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Frequency domain residual analysis. The X-axis is the distance from the low-frequency center (from low frequency to high frequency), and the Y-axis represents total magnitude of frequency residual. It can be seen that when \\( t \\) is small, the high-frequency difference exceeds the low-frequency difference. As \\( t \\) increasing, the low-frequency difference gradually dominated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.429 + ], + "angle": 0, + "content": "ever, when \\( t \\) increases, such as \\( t \\in [700,800] \\), the difference in low-frequency components between the reconstructed images and the input images dominates most of the frequency domain differences between the two. Therefore, this can explain why, when the timestep of noise scheduler is large, the gradient of the perturbed image is close to zero and result in ineffective noise optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Differences at Different UNet Layers. Inspired by [35], we utilize PCA to visualize the output features of each layer in the U-Net decoder blocks during denoising. The decoder block consists of self-attention, cross-attention, and residual blocks, we select features of residual blocks. There are 11 layers in total, and the layers are visualized at timestep=500. As Figure 5 shows, with the output feature of UNet decoder blocks increasing, the visualized features gradually change from depicting structures and other low-frequency information to capturing texture and similar high-frequency information. Since our noise is intended to disrupt high-frequency information, the deeper features are good perturbation objects for our adversarial noises to reinforce interference which attempts to perturb high-frequency components in generated images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.427, + 0.683 + ], + "angle": 0, + "content": "3.4. Adaptive Greedy Time Interval Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To achieve more effective privacy protection within a limited number of noise injection steps, we propose a fast adaptive time interval selection strategy. Firstly, based on the input image, we randomly select five timesteps from the interval (0, max training step). For each time step, we compute the gradient concerning the input image. The absolute values of the gradients are summed. If the sum of absolute gradient values is the minimum at timestep \\(t\\), the corresponding interval \\((t - 20, t + 20)\\) is then removed. For the timestep with maximal sum, noise is computed and added to the image and this noised image becomes the input for the next round of gradient computation. This process is repeated until the final time interval length is no greater than 100. We obtain the ultimate interval used for noise injection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Our method relies on PGD[24] (Projected Gradient Descent), an iterative approach for adversarial attacks. At each iteration, computing the gradient becomes pivotal as it signifies the rate of change of the objective function concerning the perturbed images. The gradient approaching zero implies the possibility of iterative optimization but with minimal alterations to the objective function. This ultimately leads to reduced attack efficiency and diminished effectiveness. Thus, randomly selecting timesteps makes it more challenging to optimize the objective function under the same noise budget and results in decreased efficiency and effectiveness in countering customization. This highlights the crucial need for employing our adaptive, greedy time interval selection for perturbation." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.32, + 0.892, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Analysis on efficacy. Here we simply analyze our adaptive greedy selection for timestep \\( t \\) (the above algorithm). Its key steps in each iteration are randomly sampling 5 timesteps \\( T_{s} = \\{t_{j}\\}_{j = 1}^{5} \\) and rescaling the timestep range by deleting the interval of \\( t \\in T_{s} \\), which leads to the smallest gradient absolute value. Thus avoiding very small or even zero gradients during optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Theorem 1. Suppose the timestep range is rescaled from \\( T = A \\cup B \\) to \\( T' = (A \\setminus \\Delta A) \\cup (B \\setminus \\Delta B) \\) in some iteration, where timestep \\( t \\in B \\) leads to zero gradients, satisfying \\( A \\cap B = \\emptyset \\), and \\( \\Delta A \\cup \\Delta B \\) denotes the deleted interval, then we have \\( E_{t \\sim T} |\\nabla_x \\mathcal{L}_{DM}| < E_{T_s \\sim T} E_{t \\sim T'} |\\nabla_x \\mathcal{L}_{DM}| \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.075, + 0.089, + 0.139, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.09, + 0.203, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.09, + 0.264, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.09, + 0.326, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.09, + 0.389, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.09, + 0.45, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.09, + 0.515, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.574, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.09, + 0.639, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.09, + 0.701, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.09, + 0.761, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.09, + 0.825, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.09, + 0.885, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.073, + 0.142, + 0.148, + 0.153 + ], + "angle": 0, + "content": "portrait image" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.143, + 0.19, + 0.152 + ], + "angle": 0, + "content": "layer 0" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.143, + 0.252, + 0.152 + ], + "angle": 0, + "content": "layer 1" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.143, + 0.315, + 0.152 + ], + "angle": 0, + "content": "layer 2" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.143, + 0.378, + 0.152 + ], + "angle": 0, + "content": "layer 3" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.143, + 0.44, + 0.152 + ], + "angle": 0, + "content": "layer 4" + }, + { + "type": "text", + "bbox": [ + 0.466, + 0.143, + 0.502, + 0.152 + ], + "angle": 0, + "content": "layer 5" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.143, + 0.564, + 0.152 + ], + "angle": 0, + "content": "layer 6" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.143, + 0.626, + 0.152 + ], + "angle": 0, + "content": "layer 7" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.143, + 0.689, + 0.152 + ], + "angle": 0, + "content": "layer 8" + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.143, + 0.75, + 0.152 + ], + "angle": 0, + "content": "layer 9" + }, + { + "type": "text", + "bbox": [ + 0.774, + 0.143, + 0.816, + 0.152 + ], + "angle": 0, + "content": "layer 10" + }, + { + "type": "text", + "bbox": [ + 0.836, + 0.143, + 0.877, + 0.152 + ], + "angle": 0, + "content": "layer 11" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.157, + 0.892, + 0.186 + ], + "angle": 0, + "content": "Figure 5. PCA visualization of features. The output feature of residual blocks are visualized at timestep=500. As feature goes deeper, the high-frequency information captured by the feature becomes more significant." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.21, + 0.152, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.21, + 0.23, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.21, + 0.307, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.21, + 0.386, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.21, + 0.464, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.078, + 0.272, + 0.464, + 0.284 + ], + "angle": 0, + "content": "portrait image t ∈ [100,200] t ∈ [300,400] t ∈ [500,600] t ∈ [700,800]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.469, + 0.357 + ], + "angle": 0, + "content": "Figure 6. Diffusion feature over different timesteps. As the timestep of the noise scheduler increasing, part of the high-frequency information is lost and it is not good for adding high-frequency adversarial noise. This means that choosing an appropriate time sequence is also beneficial to feature interference loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.469, + 0.427 + ], + "angle": 0, + "content": "Proof. We can easily obtain \\( E_{T_s \\sim T} \\frac{|\\Delta A|}{|A|} < E_{T_s \\sim T} \\frac{|\\Delta B|}{|B|} \\) via proof by contradiction. For simplicity, we denote \\( \\nabla_x \\mathcal{L}_{DM} \\) as \\( g(t) \\), \\( LHS \\) and \\( RHS \\) denote left and right, then" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.434, + 0.486, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L H S = p (t \\in A) E _ {t \\sim A} | g (t) | + p (t \\in B) E _ {t \\sim B} | g (t) | \\\\ = \\frac {| A |}{| A | + | B |} E _ {t \\sim A} | g (t) | \\approx E _ {T _ {s} \\sim T} \\frac {| A | \\cdot E _ {t \\sim A \\backslash \\Delta A} | g (t) |}{| A | + | B |} \\\\ < E _ {T _ {s} \\sim T} \\frac {| A | - | \\Delta A |}{| A | + | B | - | \\Delta A | - | \\Delta B |} E _ {t \\sim A \\backslash \\Delta A} | g (t) | \\\\ = E _ {T _ {s} \\sim T} E _ {t \\sim A \\backslash \\Delta A} p (t \\in A \\backslash \\Delta A) | g (t) | = R H S \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.549, + 0.468, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.571, + 0.311, + 0.585 + ], + "angle": 0, + "content": "3.5. Feature Interference Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.469, + 0.669 + ], + "angle": 0, + "content": "Based on the analysis in Sec 3.3, we propose the feature interference loss and integrate it with the vanilla training loss as the overall objective for optimization. This loss calculates the layer-wise euclidean distance between the intermediate features in some specially selected layers, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.674, + 0.469, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\mathbb {E} \\| f _ {l} ^ {t ^ {*}} - f _ {l} ^ {t} \\| _ {2} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.696, + 0.469, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n d}} + \\lambda \\mathcal {L} _ {f} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Where \\( f_{l}^{t^{*}} \\) is the output features of the selected layers sets \\( l \\) at timestep \\( t \\) for the input image, \\( f_{l}^{t} \\) is the output features for current perturbed images, and \\( \\lambda \\) represents the weighting coefficient for the feature interference loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We visualized the output features of layer 7 at different time steps. As Figure 6 shows, when the timesteps increase, more noise is added to the latent and the response to the high-frequency components of the input image gradually decreases. Therefore, utilizing feature interference loss as the optimization objective for adversarial noise in small time intervals is a better choice to keep the effectiveness of adversarial noise disturbing high-frequency information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.211, + 0.632, + 0.228 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.237, + 0.582, + 0.253 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Dataset We utilized two facial datasets for experiments including Celeb-HQ [20] and VGGFace2 dataset[3]. The dataset comprises about 50 individuals and aligns with the Anti-DreamBooth, with each individual including at least 15 clear face images for customization. Since Anti-dreambooth needs alternating training, the dataset is divided into three sets, including set A, set B, and set C and each set contains 5 portrait images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Model Since the open-sourced stable diffusion is the most popular implementation of latent diffusion among the community, our experiments mainly use the SD-v2.1 [1]. To test the performance of our method in a black-box scenario, we assume the versions of Stable Diffusion between anticustomization and customization are the same or different." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Baseline We compare several open-source methods that employ adversarial attacks on diffusion models to protect user images from being misused by text-to-image diffusion models, including Photoguard [29], AdvDM [23] and AntiDreamBooth [36]. Due to the high GPU memory requirements of the complete PhotoGuard, we only utilize the VAE encoder attack strategy in its paper for comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Metric We utilize a face detector named Retinaface [5] to detect if there is a face in the image and use the Face Detection Failure Rate (FDFR) to assess the level of disruption to the generated faces. Upon detecting a face, we encode it using ArcFace [4] and calculate the cosine similarity between the protected image and clean input, measuring the identity resemblance between the detected face in the generated image and that of users. This matrix is defined as Identity Score Matching (ISM). In addition, the image quality is quantified by BRISQE [26], and the quality of detected facial images is measured through SER-FIQ [34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details We set the same noise budget for all methods \\(\\eta = 16 / 255\\). Additionally, the optimization steps and step size align with the settings specified in each baseline. The number of training epochs is 50 and each epoch includes 3 steps for surrogate model training and 9 steps for adversarial noise optimization. The step size for adding noise is set as 0.005 and the learning rate for the training model is set as 5e-7 which is appropriate for human faces. The maximum adaptive greedy search steps are set as 50 by default. After training, we used the noised im" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12052" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.457, + 0.415 + ], + "angle": 0, + "content": "
CelebA-HQ
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2541.0919.380.55
AdvDM [23]0.3270.4838.170.20
Anti-DB [36]0.2877.2837.430.20
Anti-DB + SimAC0.3187.0738.860.21
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2028.5029.330.59
AdvDM [23]0.2565.3737.860.41
Anti-DB [36]0.1986.8038.900.27
Anti-DB + SimAC0.1296.8742.100.15
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.1830.0726.960.40
AdvDM[23]0.2935.1036.460.36
Anti-DB[36]0.2242.8640.340.28
Anti-DB + SimAC0.1291.9043.970.06
Method“a photo of sks person in front of eiffel tower”
ISM ↓FDFR ↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.0850.9532.820.40
AdvDM[23]0.0938.1036.020.30
Anti-DB[36]0.0656.2641.350.22
Anti-DB + SimAC0.0566.1942.770.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.418, + 0.472, + 0.461 + ], + "angle": 0, + "content": "Table 1. Comparison with other open-sourced anti-customization methods on CelebA-HQ. We evaluate the performance under four different prompts during customization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.613 + ], + "angle": 0, + "content": "age for customization. The default base model is stable diffusion v2.1 combined with DreamBooth. According to the above analysis of the middle layer of the model, we select the 9, 10, and 11 layer features that can best represent the high-frequency information of the image for additional disturbance. The weight \\(\\lambda\\) of feature interference loss is set to 1. After finetuning 1000 steps, we save the model checkpoints and conduct inference. For each prompt, 30 images in .png format are generated for metric calculation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.388, + 0.637 + ], + "angle": 0, + "content": "4.2. Comparison with Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Quantitative Results To test the effectiveness of our approach in enhancing the protection of users' portrait images, we conduct a quantitative comparison under three prompts in Table 1 and Table 2. To be more practical, we list four text prompts for inference, the first one \"a photo of sks person\" is the same as the prompt used in training, while the other three were prompts unseen before. For each prompt, we randomly select 30 generated images to compute each metric, reporting their average values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We can find that our method greatly improves the performance of Anti-DreamBooth and outperforms other baselines across all prompts. Due to our analysis of the unique properties of the diffusion model during denoising and effective addition of high-frequency noise, the face detection failure rates greatly increase, and the identity matching scores between the detected face and the input image are the lowest among all methods. This implies that our method is" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.875, + 0.415 + ], + "angle": 0, + "content": "
VGG-Face2
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2929.2720.670.47
AdvDM [23]0.3263.0738.510.21
Anti-DB[36]0.3064.6737.890.22
Anti-DB + SimAC0.3180.2740.710.22
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2517.3328.520.55
AdvDM [23]0.3067.8037.580.35
Anti-DB[36]0.2377.4738.790.29
Anti-DB + SimAC0.1196.3341.780.12
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1740.8730.100.32
AdvDM[23]0.2737.7335.430.29
Anti-DB[36]0.2545.0039.250.27
Anti-DB + SimAC0.1389.6044.970.07
Method“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1351.0730.690.41
AdvDM[23]0.1438.6735.990.31
Anti-DB[36]0.1054.9341.130.23
Anti-DB + SimAC0.0961.2042.170.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Table 2. Comparison with other open-sourced anti-customization methods on VGG-Face2. We evaluate the performance under four different prompts during customization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.476, + 0.852, + 0.491 + ], + "angle": 0, + "content": "more effective in resisting abuse from customization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Qualitative Results We show some results in Figure 7. It's evident that SimAC combined with Anti-DreamBooth achieves a strong image disruption effect, providing the best privacy protection for the input portrait. Since PhotoGuard operates an attack on the latent space, the resulting images tend to generate patterns similar to the target latent. This attack doesn't effectively reduce the probability of facial appearances or improve the quality of generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Both AdvDM and Anti-DreamBooth use DM's training loss as an objective for optimizing noise, which, as been seen, yields similar results. The two methods aim to disrupt high-frequency components in images via adversarial noise to create artifacts and raise the bar for potential misuse of users' photos. Although the quality of generated images decreases, the results of both AdvDM and Anti-DreamBooth retain many details from the user input images. This indicates a leak of privacy for some users and isn't user-friendly. According to our analysis, both of these methods follow the training process of DM, randomly selecting timestamps for optimization. This leads to a waste of some noise addition steps, thereby reducing the efficiency of the attack and the effectiveness of protection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.72, + 0.848 + ], + "angle": 0, + "content": "4.3. Black-Box Performance" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Prompt Mismatch When attackers use stable diffusion to customize concepts, the prompt they use might differ from our assumptions when adding noise. Thus, we use the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12053" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.152, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.535, + 0.154, + 0.547 + ], + "angle": 0, + "content": "Portrait Image" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.089, + 0.231, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.536, + 0.226, + 0.547 + ], + "angle": 0, + "content": "PhotoGuard" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.089, + 0.311, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.536, + 0.294, + 0.546 + ], + "angle": 0, + "content": "AdvDM" + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.089, + 0.39, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.536, + 0.372, + 0.546 + ], + "angle": 0, + "content": "Anti-DB" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.089, + 0.47, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.536, + 0.472, + 0.547 + ], + "angle": 0, + "content": "Anti-DB+SimAC" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.607 + ], + "angle": 0, + "content": "Figure 7. Visualization results (four prompts) on CelebA-HQ. From the first row to the last row is \"a photo of sks person\", \"a dslr portrait of sks person\", \"a photo of sks person looking at the mirror\", \"a photo of sks person in front of eiffel tower\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.728 + ], + "angle": 0, + "content": "\"a photo of sks person\" during perturbation learning and change the rare identifier \"sks\" to \"t@t\" during DreamBooth model finetuing. The results in Table 3 indicate a decrease in performance in this scenario, but we notice that the Identity Score Matching(ISM), representing identity similarity remains consistent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Model Mismatch The models used to add adversarial noise may also mismatch with the model fine-tuned by DreamBooth. We examine the effectiveness of adversarial noise learned on stable diffusion v2.1 against customization based on stable diffusion v2.1 and v1.4, or vice versa in Table 4. The conclusion of model mismatch resembles those of prompt mismatch. Although there is an overall decline in performance, the critical matrix, ISM, remains considerable. This also implies that in a model mismatch scenario, SimAC + Anti-DreamBooth can still protect user portrait privacy." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.089, + 0.881, + 0.319 + ], + "angle": 0, + "content": "
Train [v]Test [v]“a photo of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.3187.0738.860.21
skst@t0.3081.3639.670.31
Train [v]Test [v]“a dslr portrait of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1296.8742.100.15
skst@t0.2354.4237.770.52
Train [v]Test [v]“a photo of [v] person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1291.9043.970.06
skst@t0.1630.5440.630.25
Train [v]Test [v]“a photo of [v] person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.0566.1942.770.12
sksst@t0.0623.2037.740.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Table 3. Prompt mismatch between training and testing on CelebA-HQ. The training prompt is \"a photo of sks person\" and the inference prompt uses rare identifier \"sks\" or \"t@t\"." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.372, + 0.866, + 0.642 + ], + "angle": 0, + "content": "
TrainTest“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.3187.0738.860.21
v1.4v2.10.3870.0739.220.35
v2.1v1.40.0199.8662.090.02
TrainTest“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1296.8742.100.15
v1.4v2.10.1495.1740.950.23
v2.1v1.40.2620.5416.440.57
TrainTest“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1291.9043.970.06
v1.4v2.10.1679.3942.520.13
v2.1v1.40.2082.7240.120.16
TrainTest“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.0566.1942.770.12
v1.4v2.10.0661.3641.510.10
v2.1v1.40.0853.6720.560.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.644, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Table 4. Model version mismatch during training and testing on CelebA-HQ. The training prompt is \"a photo of sks person\"" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.693, + 0.619, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Current anti-customization methods, primarily relying on adversarial attacks, often overlook crucial internal properties of the diffusion model, leading to ineffective optimization. This paper addresses this issue by exploring inherent properties to enhance anti-customization. Two key aspects are examined: Analyzing the relationship between timestep selection and the model's perception in the frequency domain inspires an adaptive greedy search for optimal timesteps. Scrutinizing feature roles during denoising, resulting in a sophisticated feature-based optimization framework. Experiments show increased identity disruption, enhancing user privacy and security." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12054" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Stable-diffusion-2-1. https://huggingface.co/stabilityai/stable-diffusion-2-1.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.197 + ], + "angle": 0, + "content": "[2] Making pictures of trump getting arrested while waiting for trump's arrest. https://twitter.com/EliotHiggins/status/1637927681734987777. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.468, + 0.269 + ], + "angle": 0, + "content": "[3] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[4] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.468, + 0.492 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.468, + 0.563 + ], + "angle": 0, + "content": "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[10] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.468, + 0.774 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.872 + ], + "angle": 0, + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[14] Qidong Huang, Jie Zhang, Wenbo Zhou, Weiming Zhang, and Nenghai Yu. Initiative defense against facial manipula" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1619-1627, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[15] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, and Nenghai Yu. Shape-invariant 3d adversarial point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15335-15344, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, Kui Zhang, Gang Hua, and Nenghai Yu. Pointcat: Contrastive adversarial training for robust point cloud recognition. arXiv preprint arXiv:2209.07788, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Yinpeng Chen, Lu Yuan, Gang Hua, Weiming Zhang, and Nenghai Yu. Improving adversarial robustness of masked autoencoders via test-time frequency-domain prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1600-1610, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Weiming Zhang, Feifei Wang, Gang Hua, and Nenghai Yu. Diversity-aware meta visual prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10878-10887, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. arXiv preprint arXiv:2311.17911, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, pages 20763-20786. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In 6th International Conference on Learning Representations, ICLR 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[25] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] A. Mittal, A. K. Moorthy, and A. C. Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, page 4695-4708, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12055" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, pages 29894-29918. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.4 + ], + "angle": 0, + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[33] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[34] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. IEEE, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[35] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[36] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.469, + 0.81 + ], + "angle": 0, + "content": "[37] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.812, + 0.469, + 0.88 + ], + "angle": 0, + "content": "[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12056" + } + ] +] \ No newline at end of file diff --git a/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_origin.pdf b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e077b7a04fe2b8912e1fc49712af50d5ab625791 --- /dev/null +++ b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/95390555-fbca-4ec8-a42e-5840969340e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b12005bb0baafc66ee5674deb684efa0824f778943f713536c751792072ce6f +size 1655363 diff --git a/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/full.md b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..58bf4d3cef1518dfca9bc1a1f445270e9054e822 --- /dev/null +++ b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/full.md @@ -0,0 +1,358 @@ +# SimAC: A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models + +Feifei Wang $^{1,2,*}$ ; Zhentao Tan $^{2,1}$ , Tianyi Wei $^{1}$ , Yue Wu $^{2}$ , Qidong Huang $^{1,\dagger}$ + +1University of Science and Technology of China 2Alibaba Cloud + +{wangfeifei@, tzt@, bestwty@, hqd0037@}mail.ustc.edu.cn matthew.wy@alibaba-inc.com + +# Abstract + +Despite the success of diffusion-based customization methods on visual content creation, increasing concerns have been raised about such techniques from both privacy and political perspectives. To tackle this issue, several anti-customization methods have been proposed in very recent months, predominantly grounded in adversarial attacks. Unfortunately, most of these methods adopt straightforward designs, such as end-to-end optimization with a focus on adversarially maximizing the original training loss, thereby neglecting nuanced internal properties intrinsic to the diffusion model, and even leading to ineffective optimization in some diffusion time steps. In this paper, we strive to bridge this gap by undertaking a comprehensive exploration of these inherent properties, to boost the performance of current anti-customization approaches. Two aspects of properties are investigated: 1) We examine the relationship between time step selection and the model's perception in the frequency domain of images and find that lower time steps can give much more contributions to adversarial noises. This inspires us to propose an adaptive greedy search for optimal time steps that seamlessly integrates with existing anti-customization methods. 2) We scrutinize the roles of features at different layers during denoising and devise a sophisticated feature-based optimization framework for anti-customization. Experiments on facial benchmarks demonstrate that our approach significantly increases identity disruption, thereby protecting user privacy and copyright. Our code is available at: https://github.com/somuchtoe/SimAC. + +# 1. Introduction + +Latent Diffusion model (LDM) [12, 31, 32] has been recently proven as a strong paradigm for photorealistic visual content generation. The emergence of open-source Stable + +![](images/18743d4cd9e1f11606bc64b573c72a40db7ef86b3c365e6716f6a3e13a57acfa.jpg) +Figure 1. Comparison between Anti-DreamBooth before and after adding SimAC. Our method further boosts its ability to de-identity. + +Diffusion encourages users to explore creative possibilities with LDMs. Users only need to provide several images representing the same subject along with a rare identifier to customize their diffusion models [7, 13, 22, 28]. After finetuning the model or optimizing the text embedding of rare identifiers, it can flexibly generate high-quality images containing the specified object. + +The convenience of customizing large-scale text-to-image models also allows the malicious users to generate forgery images that violate the truth. Some of them may use such technique to steal others' painting styles and generate new painting content without permission. While some of them may collect one's portrait photos that are published on the social platform, and generate fake images of this person through customization. In other words, the lawbreakers can easily produce fake news that contains celebrity photos [2], greatly misleading the public. The infringement poses threats to user privacy and intellectual property. Hence, it is essential to devise effective countermeasures to safeguard users against such malicious usage. + +![](images/b1fa719306813ca22bcd59973e0e26be945ca4a932f2abc81a5869de3fe3a215.jpg) +(a) the number of absolute gradients $< 1e^{-10}$ + +![](images/288d32bee9a12b2e79eca9fc28e50aeb7a5acea7ee5c2e1273d04f5f15d64010.jpg) +(b) the maximum absolute gradients +Figure 2. Distribution of Anti-DreamBooth attack gradients on different diffusion timesteps, where (a) counts the number of absolute gradients below the threshold $1e^{-10}$ at each timestep, (b) presents the maximum absolute gradients at different timesteps, and (c) demonstrates how the mean and median of the absolute gradients change over timesteps. Apparently, the absolute gradients show great discrepancy on the varying timesteps and nearly zero values appear at large timesteps, which leads to ineffective noise optimization. + +![](images/20f4c6fcb3346b7083d62a5bad82c4d7864bbe5e3390bb5b37cd36f9f494df94.jpg) +(c) the mean and median of the absolute gradients + +Current anti-customization methods are generally based on adversarial attack [9, 33]. AdvDM [23] is the pioneering work that uses adversarial noise to protect user images from being customized by diffusion models. It ingeniously combines the diffusion model with adversarial samples, firstly achieving user privacy protection. Anti-DreamBooth [36] further enhances the protective effect by employing alternate training. Both of them randomly select the timesteps from (0, max denoising steps) regarding the noise added to LDM latent, and directly employ the maximization of the LDM training loss as the optimization objective. + +However, as showcased in Figure 2, we notice that the gradient of the perturbed images is quite small and even zero when the randomly sampled timesteps are relatively large, i.e., the noisy latent is closer to Gaussian noise. This implies that within the limited steps of an adversarial attack, a large portion of steps are ineffective since the perturbed image cannot optimize the objective with such zero gradients, leading to a decrease in both protection effectiveness and efficiency. Hence, the images protected by these current methods, when customized, still allow the model to capture many details from user-input images and leak the user's privacy, as shown in Figure 1. + +The fundamental reason is that these methods fail to combine the adversarial attacks with properties inherent in diffusion models. We aim to conduct an in-depth analysis of why the gradients of perturbed images have such discrepancy at different time steps and how diffusion models perceive input images at different intermediate layers. Then, based on the in-depth analysis, we propose improvements to existing customization methods from both the temporal and feature dimensions. + +To observe the relationship between timesteps and the gradient of the perturbed images, we first reconstruct images predicted from noisy latents at different time steps and + +compare them with the input images. Considering regular adversarial noises mainly affect the high frequency of images, our comparison is conducted in the frequency domain, aiming at investigating whether the model's perception lies on lower-level or higher-level. We observe that, the model focuses on the higher frequency components of images when selecting smaller time steps, and vice versa. Therefore, introducing adversarial noises at larger timesteps is ineffective, since the subtle changes perturbed on images can hardly affect the low-frequency of the generated images. To improve the effectiveness, we propose an adaptive timestep selection method to find optimal time intervals, where we iteratively update the range of selection in a greedy way. To explore the effect of different layers' features in the U-Net decoder during denoising, we employ PCA (Principal Component Analysis) to visualize them and show the discrepancy. We clearly find that within the decoder, the feature extraction gradually shifts from low-frequency to high-frequency as the layer goes deeper. It indicates the higher layers concentrates more on the texture of images, while the lower layers focus on the structure. Consequently, we select the features representing high-frequency information during optimization, since regular adversarial noises concentrate on the high frequency of images. And we construct the feature interference loss and integrate it with the diffusion denoising loss as the objective function, to improve the ability of identity interference. + +Our main contributions are as follows: + +- We reveal the inadequate optimization steps that exist in current anti-customization methods, and gives detailed analysis regarding the perceptual discrepancy of diffusion models at different timesteps and intermediate layers, we have better aligned the optimization of adversarial gradients with diffusion models. +- Based on analysis, we propose a simple but effective anti + +customization method, including adaptive greedy time selection and a feature interference loss to improve the protection ability. Our method can be easily generalized to different anti-customization frameworks and improve their performance. +- Extensive evaluations on two face datasets demonstrate that our method achieves more obvious disruption of the user's identity and provides better privacy protection. + +# 2. Related work + +# 2.1. Generative Models and Diffusion Models + +Variational autoencoders (VAEs) [21] and Generative adversarial networks (GANs) [8] are popular frameworks among generative models which have strong generative ability. These models encode the data $x$ as latent variables $z$ and model the joint distribution $p_{\theta}(x,z)$ . However, the quality of VAE samples is not competitive with GANs which are suffering from training instability [10]. Since diffusion probabilistic models (DM) [31] progressively add noise to the data from the joint distribution $q_{\theta}(x_{0:T}|x_0)$ and denoise step by step, the efficiency of training and quality of samples have achieved state-of-work. + +Unconditional generative models cannot produce desired samples and then models take different input as guidance have sprung up. Based on GAN, cGAN [25] generates images conditioned on the given labels $y$ and Cycle-GAN [38] implements unpaired image translation considering the given image. Equipped with some techniques like classifier-free guidance [11], the diffusion model gains the ability to follow diverse prompts as conditions during generation. + +The open-sourced Latent Diffusion Models (LDMs) [27] operate images in the latent space of low dimensions rather than pixel space which greatly reduces training computation. To support different condition inputs, they add cross-attention layers into the underlying U-Net backbone as conditioning mechanisms $\tau_{\theta}$ . The above delicate designs make diffusion more friendly for users to create what they need. It can also assist them in designing their diffusions which synthesize visual contents that contain specific concepts (e.g., objects or styles) given by users during inference. + +# 2.2.Customization + +DreamBooth [28] stands out as a popular diffusion-based method for customizing text-to-image generation. This approach involves presenting $3 \sim 5$ images depicting a specific concept (e.g., a particular dog) alongside a corresponding identifier (e.g., "a sks dog"). DreamBooth utilizes these images to fine-tune the pre-trained Stable Diffusion model. This fine-tuning process encourages the model to "memorize" the concept and its identifier, enabling it to reproduce this concept in new contexts during inference. + +On the other hand, Textual-inversion [7] employs a dif + +ferent approach. This method freezes the U-Net and exclusively optimizes the text embedding of unique identifiers (e.g. "sks") to represent the input concepts. + +Inspired by DreamBooth, numerous works have been mushroomed such as custom-diffusion [22], Sine [37], among others. In the quest for more efficient fine-tuning, DreamBooth has a successful integration with LoRA [13] and has become a very influential customization project in the community. LoRA specifically decomposes the attention layer of the vision transformer model [6] into low-rank matrices reducing the cost associated with fine-tuning. + +# 2.3. Privacy Protection for Diffusion Models + +To alleviate the issue that private images are misused by Stable Diffusion based customization, so-called "anticustomization", some researchers [23, 30, 36] have recently delved into privacy protection for diffusion models. AdvDM [23] misleads the feature extraction of diffusion models. They analyze the training objective of fine-tuning and propose a direct way that uses the gradient of denoising loss as guidance to optimize the latent variables sampled from the denoising process. The generated adversarial examples degrade the generation ability of DreamBooth or other DM-based customized approaches. Inspired by [14], Antidreambooth [36] uses the alternating surrogate and perturbation learning (ASPL) to approximate the real trained models. They train the model on clean data and use these models as the surrogate model to compute the noise added to the user-provided images. The perturbed images then as the training data for surrogate model fine-tuning which mimics the real scenes. Photoguard [29] specially focuses on unauthorized image inpainting which misleads the public and does harm to personal reputation. They use both VAE encoder attack and UNet attack targeted on a gray image to hinder infringers from creating fake news. + +Some concerns limit the application of current anticom customization methods in more practical scenarios, such as ineffective optimization, poor identity disruption, and simply using the reconstruction loss as guidance to generate adversarial examples [15-19]. Our paper aims to fill these blanks by figuring out the internal mechanisms that affects the performance of protection. + +# 3. Method + +# 3.1. Preliminaries + +Diffusion Model (DM) DM is a probabilistic generative model that samples from Gaussian distribution and then progressive denoise to learn the desired data distribution. Given $x_0 \sim q(x)$ , the forward process adds increasing noise to the input images at each time-step $t \in (0, T)$ which produces a sequence of noisy latent, $\{x_0, x_1, \dots, x_T\}$ . The backward process trains model $\epsilon_{\theta}(x_t, t, c)$ to predict the + +![](images/fe297b61c3dc13b8033d9e30ff1cae5949f1d05dc5be630413f952a08339a1d1.jpg) +Figure 3. Pipeline of SimAC. We first greedily select the time step with our adaptive strategy during the feed forward phase. Then we integrate the feature interference loss with the vanilla training loss as the final objective. The noise is iteratively optimized. + +noise added in $x_{t}$ to infer $x_{t-1}$ . During denoising, the training loss is $l_{2}$ distance, as shown in Eq.(1), Eq.(2). Although there are many implementations of text-to-image diffusion models, Stable Diffusion is one of the few open-sourced diffusion models and is widely used in the community. Thus, our method is mainly based on the checkpoints of stable diffusion provided in HuggingFace. + +$$ +\mathcal {L} _ {u n c o n d} (\theta , x _ {0}) = \mathbb {E} _ {x _ {0}, t, \epsilon \in \mathcal {N} (0, 1)} \| \epsilon - \epsilon_ {\theta} (x _ {t + 1}, t) \| _ {2} ^ {2}, \tag {1} +$$ + +$$ +\mathcal {L} _ {\text {c o n d}} (\theta , x _ {0}) = \mathbb {E} _ {x _ {0}, t, c, \epsilon \in \mathcal {N} (0, 1)} \| \epsilon - \epsilon_ {\theta} (x _ {t + 1}, t, c) \| _ {2} ^ {2}, \tag {2} +$$ + +where $x$ is the input image, $t$ is the corresponding timestep, $c$ is condition input, (e.g., text or image), $\epsilon$ is the noise term. + +Adversarial Attack The adversarial examples for classification are crafted to mislead the model to classify the given input to the wrong labels. However, the traditional attack strategies are not effective when dealing with generative models. For diffusion models, adversarial examples are some images that are added on imperceptible noise, causing diffusion models to consider them out of the generated distribution. In detail, this noise heightens the challenge of image reconstruction and hinders the customization capabilities of applications based on diffusion models (DMs). The optimized noise is often typically constrained to be smaller than a constant value $\eta$ . The $\delta$ is determined through the following formation: + +$$ +\delta_ {a d v} = \arg \max _ {\| \delta \| _ {p} < \eta} L \left(f _ {\theta} (x + \delta), y _ {r e a l}\right), \tag {3} +$$ + +where $x$ is the input image, $y_{real}$ is the real images, and $L$ is the loss function used to evaluate the performance of adversarial examples. + +Projected Gradient Descent (PGD) [24] is a widely used method to iteratively optimize adversarial examples. The process is formulated as + +$$ +x ^ {t + 1} = \prod_ {(x, \eta)} \left(x ^ {t} + \alpha \operatorname {s g n} \left(\nabla_ {x} L \left(f _ {\theta} (x + \delta), y _ {\text {r e a l}}\right)\right)\right) \tag {4} +$$ + +where $x^0 = x$ and $x$ is the input image, $\operatorname{sgn}(\cdot)$ is sign function, $(\nabla_x L(f_\theta(x + \delta), y_{real})$ is the gradient of the loss function with respect to $x + \delta$ . $\alpha$ represents the step size during each iteration and $t$ is the iteration number. With the operation $\Pi_{(x,\eta)}$ , the noise is limited to a $\eta$ -ball ensuring the adversarial examples are acceptable. + +# 3.2. Overview + +Here we delve into the properties of LDMs and analyze the potential vulnerabilities that can benefit the attack. In Sec. 3.3, we give a comprehensive analysis of the properties. In Sec. 3.4 and 3.5, we propose our method based on these analyses, mainly including two components, i.e., adaptive greedy time interval selection and the feature interference loss. The overall pipeline is illustrated in Figure 3. + +# 3.3. Analysis on Properties of LDMs + +Differences at different time step. Our exploration focuses on analyzing gradients across various time intervals from a statistical standpoint. We've established a threshold of $1e^{-10}$ to assess the number of gradients with absolute values below this threshold in perturbed images across different timesteps. Figure 1 illustrates a notable trend: during the latter part of the total time interval $(0, \text{MaxTrainStep})$ , there's a sharp increase in the count of absolute gradients below the specified threshold. Additionally, Figure 1 presents the maximum value, mean, and median of absolute gradient decreases in the perturbed images when timesteps get larger. + +As shown Eq.(5), the forward process admits sampling $x_{t}$ at an arbitrary timestep $t$ . The magnitude of the noise corresponding to timestep is determined by a pre-defined noise schedule. To ensure that the final latent $x_{T}$ conforms to a standard normal distribution, the amplitude of noise injection gradually increases with the timestep. + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right) \tag {5} +$$ + +During denoising, we reconstruct the input image $x_0$ based on the noisy latent $z_t$ and visualize the difference between the reconstructed image and the input image in the frequency domain at different timesteps. As shown in Figure 4, when $t \in [0,100]$ , the main difference between the original and reconstructed images lies in the high-frequency components. Our initial expectation is to disrupt the reconstruction of the original image by introducing high-frequency adversarial noise, aiming to achieve anti-customization. How- + +![](images/b180f286dbb87f5ade78d118a771268a1860d5b4599e12bb08361b550acdbde6.jpg) +Figure 4. Frequency domain residual analysis. The X-axis is the distance from the low-frequency center (from low frequency to high frequency), and the Y-axis represents total magnitude of frequency residual. It can be seen that when $t$ is small, the high-frequency difference exceeds the low-frequency difference. As $t$ increasing, the low-frequency difference gradually dominated. + +![](images/cf95b0908637bcbb0d6b2c087e6fd4f286d1dfaa99c1866df11db03e544a4f07.jpg) + +![](images/530f3e19b45f68a81b84385c5e4f2b2e6afc233aa9738b0e20bbd6b0968633e3.jpg) + +ever, when $t$ increases, such as $t \in [700,800]$ , the difference in low-frequency components between the reconstructed images and the input images dominates most of the frequency domain differences between the two. Therefore, this can explain why, when the timestep of noise scheduler is large, the gradient of the perturbed image is close to zero and result in ineffective noise optimization. + +Differences at Different UNet Layers. Inspired by [35], we utilize PCA to visualize the output features of each layer in the U-Net decoder blocks during denoising. The decoder block consists of self-attention, cross-attention, and residual blocks, we select features of residual blocks. There are 11 layers in total, and the layers are visualized at timestep=500. As Figure 5 shows, with the output feature of UNet decoder blocks increasing, the visualized features gradually change from depicting structures and other low-frequency information to capturing texture and similar high-frequency information. Since our noise is intended to disrupt high-frequency information, the deeper features are good perturbation objects for our adversarial noises to reinforce interference which attempts to perturb high-frequency components in generated images. + +# 3.4. Adaptive Greedy Time Interval Selection + +To achieve more effective privacy protection within a limited number of noise injection steps, we propose a fast adaptive time interval selection strategy. Firstly, based on the input image, we randomly select five timesteps from the interval (0, max training step). For each time step, we compute the gradient concerning the input image. The absolute values of the gradients are summed. If the sum of absolute gradient values is the minimum at timestep $t$ , the corresponding interval $(t - 20, t + 20)$ is then removed. For the timestep with maximal sum, noise is computed and added to the image and this noised image becomes the input for the next round of gradient computation. This process is repeated until the final time interval length is no greater than 100. We obtain the ultimate interval used for noise injection. + +Our method relies on PGD[24] (Projected Gradient Descent), an iterative approach for adversarial attacks. At each iteration, computing the gradient becomes pivotal as it signifies the rate of change of the objective function concerning the perturbed images. The gradient approaching zero implies the possibility of iterative optimization but with minimal alterations to the objective function. This ultimately leads to reduced attack efficiency and diminished effectiveness. Thus, randomly selecting timesteps makes it more challenging to optimize the objective function under the same noise budget and results in decreased efficiency and effectiveness in countering customization. This highlights the crucial need for employing our adaptive, greedy time interval selection for perturbation. + +![](images/5784d6a4494f50a286559a152acb8e71b348da39e5cf7385719f39a0840455b1.jpg) + +Analysis on efficacy. Here we simply analyze our adaptive greedy selection for timestep $t$ (the above algorithm). Its key steps in each iteration are randomly sampling 5 timesteps $T_{s} = \{t_{j}\}_{j = 1}^{5}$ and rescaling the timestep range by deleting the interval of $t \in T_{s}$ , which leads to the smallest gradient absolute value. Thus avoiding very small or even zero gradients during optimization. + +Theorem 1. Suppose the timestep range is rescaled from $T = A \cup B$ to $T' = (A \setminus \Delta A) \cup (B \setminus \Delta B)$ in some iteration, where timestep $t \in B$ leads to zero gradients, satisfying $A \cap B = \emptyset$ , and $\Delta A \cup \Delta B$ denotes the deleted interval, then we have $E_{t \sim T} |\nabla_x \mathcal{L}_{DM}| < E_{T_s \sim T} E_{t \sim T'} |\nabla_x \mathcal{L}_{DM}|$ . + +![](images/5d9fe4e4f9da23b2e0b620234bd8f65b8e09fd3ef4d9a870a1c10b089610a02b.jpg) +Figure 5. PCA visualization of features. The output feature of residual blocks are visualized at timestep=500. As feature goes deeper, the high-frequency information captured by the feature becomes more significant. + +![](images/116e7b02e75e4334c2c7946a0286cff766c215cb7ae32d3e2410c5cc77029158.jpg) + +![](images/64177697146243cc0dbea35ed76c7bb8aef09cd84df0cfbf8a0d5751ec0746ea.jpg) + +![](images/ba0c9dc5d1db9af27eb35b89aa83c6737af78410f1b0a6192e115d83dcc865c8.jpg) + +![](images/ff8485302aef3db6bb0e14dd75463e7a6a17e9a5fe7b33375f8167b7933f82e7.jpg) + +![](images/81bfe356bbb90efecb14e5f8d0d82c13c5f694a294c14c9e52f41be9fec551bc.jpg) + +![](images/0147ae25280bb7572f5d0f4bd8955839919131077fd2cb927db82f64e5b75593.jpg) + +![](images/44c8de6128b2fe3ca14c8003032019eee927ca19a442137ec4087ffb0b4c08d2.jpg) + +![](images/c451000ed1b6dafe575399d49c13e852f8479262c99e82229359c54fe18a4524.jpg) + +![](images/16a4572380c47ad0d89b00869fc0ce44339af38e5d748ad942696b998b7c2e5d.jpg) + +![](images/156a615e14979e42c83e29cb699f7a3a154d5e710e3169e218733d1910c0c385.jpg) + +![](images/e1e4a09a348669b0044e1ac20048400d708a77443437b1f4a8c384bbf26bd4b1.jpg) + +![](images/819a421bc3b6409f91a1d95fc267038b34da33f1643dccc4dac8564bab384305.jpg) + +portrait image + +layer 0 + +layer 1 + +layer 2 + +layer 3 + +layer 4 + +layer 5 + +layer 6 + +layer 7 + +layer 8 + +layer 9 + +layer 10 + +layer 11 + +Figure 6. Diffusion feature over different timesteps. As the timestep of the noise scheduler increasing, part of the high-frequency information is lost and it is not good for adding high-frequency adversarial noise. This means that choosing an appropriate time sequence is also beneficial to feature interference loss. +![](images/2f5d71dedf0c5c2c86784552847d978356aff5774f8c33246ed53803c47c790a.jpg) +portrait image t ∈ [100,200] t ∈ [300,400] t ∈ [500,600] t ∈ [700,800] + +![](images/63ef16559c62b14d2d9da95e2c0eac01d95f7b5765ddf45a1b328f079e374610.jpg) + +![](images/6d49296633c15a35e5ab55cbae7120afe17a1cb24d3179a52794e25f16014ede.jpg) + +![](images/d515fe086cc9e37c4c5937e54750a786496f96bde09de713c2182bb422bc7169.jpg) + +![](images/5e20609d520f2196d2fc86389f1cb4e3564fd9d1e37b1b5735af331f0f67e7e2.jpg) + +Proof. We can easily obtain $E_{T_s \sim T} \frac{|\Delta A|}{|A|} < E_{T_s \sim T} \frac{|\Delta B|}{|B|}$ via proof by contradiction. For simplicity, we denote $\nabla_x \mathcal{L}_{DM}$ as $g(t)$ , $LHS$ and $RHS$ denote left and right, then + +$$ +\begin{array}{l} L H S = p (t \in A) E _ {t \sim A} | g (t) | + p (t \in B) E _ {t \sim B} | g (t) | \\ = \frac {| A |}{| A | + | B |} E _ {t \sim A} | g (t) | \approx E _ {T _ {s} \sim T} \frac {| A | \cdot E _ {t \sim A \backslash \Delta A} | g (t) |}{| A | + | B |} \\ < E _ {T _ {s} \sim T} \frac {| A | - | \Delta A |}{| A | + | B | - | \Delta A | - | \Delta B |} E _ {t \sim A \backslash \Delta A} | g (t) | \\ = E _ {T _ {s} \sim T} E _ {t \sim A \backslash \Delta A} p (t \in A \backslash \Delta A) | g (t) | = R H S \\ \end{array} +$$ + +![](images/4b71cf575c980efa973f49ea1af1636b642ade1055e787c1a3a3fec55774f825.jpg) + +# 3.5. Feature Interference Loss + +Based on the analysis in Sec 3.3, we propose the feature interference loss and integrate it with the vanilla training loss as the overall objective for optimization. This loss calculates the layer-wise euclidean distance between the intermediate features in some specially selected layers, i.e., + +$$ +\mathcal {L} _ {f} = \mathbb {E} \| f _ {l} ^ {t ^ {*}} - f _ {l} ^ {t} \| _ {2} ^ {2} \tag {6} +$$ + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o n d}} + \lambda \mathcal {L} _ {f} \tag {7} +$$ + +Where $f_{l}^{t^{*}}$ is the output features of the selected layers sets $l$ at timestep $t$ for the input image, $f_{l}^{t}$ is the output features for current perturbed images, and $\lambda$ represents the weighting coefficient for the feature interference loss. + +We visualized the output features of layer 7 at different time steps. As Figure 6 shows, when the timesteps increase, more noise is added to the latent and the response to the high-frequency components of the input image gradually decreases. Therefore, utilizing feature interference loss as the optimization objective for adversarial noise in small time intervals is a better choice to keep the effectiveness of adversarial noise disturbing high-frequency information. + +# 4. Experiments + +# 4.1. Setup + +Dataset We utilized two facial datasets for experiments including Celeb-HQ [20] and VGGFace2 dataset[3]. The dataset comprises about 50 individuals and aligns with the Anti-DreamBooth, with each individual including at least 15 clear face images for customization. Since Anti-dreambooth needs alternating training, the dataset is divided into three sets, including set A, set B, and set C and each set contains 5 portrait images. + +Model Since the open-sourced stable diffusion is the most popular implementation of latent diffusion among the community, our experiments mainly use the SD-v2.1 [1]. To test the performance of our method in a black-box scenario, we assume the versions of Stable Diffusion between anticustomization and customization are the same or different. + +Baseline We compare several open-source methods that employ adversarial attacks on diffusion models to protect user images from being misused by text-to-image diffusion models, including Photoguard [29], AdvDM [23] and AntiDreamBooth [36]. Due to the high GPU memory requirements of the complete PhotoGuard, we only utilize the VAE encoder attack strategy in its paper for comparison. + +Metric We utilize a face detector named Retinaface [5] to detect if there is a face in the image and use the Face Detection Failure Rate (FDFR) to assess the level of disruption to the generated faces. Upon detecting a face, we encode it using ArcFace [4] and calculate the cosine similarity between the protected image and clean input, measuring the identity resemblance between the detected face in the generated image and that of users. This matrix is defined as Identity Score Matching (ISM). In addition, the image quality is quantified by BRISQE [26], and the quality of detected facial images is measured through SER-FIQ [34]. + +Implementation Details We set the same noise budget for all methods $\eta = 16 / 255$ . Additionally, the optimization steps and step size align with the settings specified in each baseline. The number of training epochs is 50 and each epoch includes 3 steps for surrogate model training and 9 steps for adversarial noise optimization. The step size for adding noise is set as 0.005 and the learning rate for the training model is set as 5e-7 which is appropriate for human faces. The maximum adaptive greedy search steps are set as 50 by default. After training, we used the noised im + +
CelebA-HQ
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2541.0919.380.55
AdvDM [23]0.3270.4838.170.20
Anti-DB [36]0.2877.2837.430.20
Anti-DB + SimAC0.3187.0738.860.21
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2028.5029.330.59
AdvDM [23]0.2565.3737.860.41
Anti-DB [36]0.1986.8038.900.27
Anti-DB + SimAC0.1296.8742.100.15
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.1830.0726.960.40
AdvDM[23]0.2935.1036.460.36
Anti-DB[36]0.2242.8640.340.28
Anti-DB + SimAC0.1291.9043.970.06
Method“a photo of sks person in front of eiffel tower”
ISM ↓FDFR ↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.0850.9532.820.40
AdvDM[23]0.0938.1036.020.30
Anti-DB[36]0.0656.2641.350.22
Anti-DB + SimAC0.0566.1942.770.12
+ +age for customization. The default base model is stable diffusion v2.1 combined with DreamBooth. According to the above analysis of the middle layer of the model, we select the 9, 10, and 11 layer features that can best represent the high-frequency information of the image for additional disturbance. The weight $\lambda$ of feature interference loss is set to 1. After finetuning 1000 steps, we save the model checkpoints and conduct inference. For each prompt, 30 images in .png format are generated for metric calculation. + +# 4.2. Comparison with Baseline Methods + +Quantitative Results To test the effectiveness of our approach in enhancing the protection of users' portrait images, we conduct a quantitative comparison under three prompts in Table 1 and Table 2. To be more practical, we list four text prompts for inference, the first one "a photo of sks person" is the same as the prompt used in training, while the other three were prompts unseen before. For each prompt, we randomly select 30 generated images to compute each metric, reporting their average values. + +We can find that our method greatly improves the performance of Anti-DreamBooth and outperforms other baselines across all prompts. Due to our analysis of the unique properties of the diffusion model during denoising and effective addition of high-frequency noise, the face detection failure rates greatly increase, and the identity matching scores between the detected face and the input image are the lowest among all methods. This implies that our method is + +Table 1. Comparison with other open-sourced anti-customization methods on CelebA-HQ. We evaluate the performance under four different prompts during customization. + +
VGG-Face2
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2929.2720.670.47
AdvDM [23]0.3263.0738.510.21
Anti-DB[36]0.3064.6737.890.22
Anti-DB + SimAC0.3180.2740.710.22
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2517.3328.520.55
AdvDM [23]0.3067.8037.580.35
Anti-DB[36]0.2377.4738.790.29
Anti-DB + SimAC0.1196.3341.780.12
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1740.8730.100.32
AdvDM[23]0.2737.7335.430.29
Anti-DB[36]0.2545.0039.250.27
Anti-DB + SimAC0.1389.6044.970.07
Method“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1351.0730.690.41
AdvDM[23]0.1438.6735.990.31
Anti-DB[36]0.1054.9341.130.23
Anti-DB + SimAC0.0961.2042.170.10
+ +Table 2. Comparison with other open-sourced anti-customization methods on VGG-Face2. We evaluate the performance under four different prompts during customization. + +more effective in resisting abuse from customization. + +Qualitative Results We show some results in Figure 7. It's evident that SimAC combined with Anti-DreamBooth achieves a strong image disruption effect, providing the best privacy protection for the input portrait. Since PhotoGuard operates an attack on the latent space, the resulting images tend to generate patterns similar to the target latent. This attack doesn't effectively reduce the probability of facial appearances or improve the quality of generated images. + +Both AdvDM and Anti-DreamBooth use DM's training loss as an objective for optimizing noise, which, as been seen, yields similar results. The two methods aim to disrupt high-frequency components in images via adversarial noise to create artifacts and raise the bar for potential misuse of users' photos. Although the quality of generated images decreases, the results of both AdvDM and Anti-DreamBooth retain many details from the user input images. This indicates a leak of privacy for some users and isn't user-friendly. According to our analysis, both of these methods follow the training process of DM, randomly selecting timestamps for optimization. This leads to a waste of some noise addition steps, thereby reducing the efficiency of the attack and the effectiveness of protection. + +# 4.3. Black-Box Performance + +Prompt Mismatch When attackers use stable diffusion to customize concepts, the prompt they use might differ from our assumptions when adding noise. Thus, we use the + +![](images/c4057c7074a43fff26ebeaa1d6ae7e6bcc73229e3bb5269b885eba01db66293d.jpg) +Portrait Image +Figure 7. Visualization results (four prompts) on CelebA-HQ. From the first row to the last row is "a photo of sks person", "a dslr portrait of sks person", "a photo of sks person looking at the mirror", "a photo of sks person in front of eiffel tower". + +![](images/14a263c12d2f12c6adc67fd88ba240969415fdb73d44fd01acd80ae499582741.jpg) +PhotoGuard + +![](images/d06941ba839fd1308cac69c8f2c70dd5c47c38a2cbf8983787b5af56c84ceef3.jpg) +AdvDM + +![](images/e80d6edf48e8bf88993ce14eb806fe367f93fa99f29eaa509ea2f9a3c2cbff57.jpg) +Anti-DB + +![](images/4e4653885dff5433de2cbc5800881cb6eaf6642f2a264a7f7a6717dfd402da65.jpg) +Anti-DB+SimAC + +"a photo of sks person" during perturbation learning and change the rare identifier "sks" to "t@t" during DreamBooth model finetuing. The results in Table 3 indicate a decrease in performance in this scenario, but we notice that the Identity Score Matching(ISM), representing identity similarity remains consistent. + +Model Mismatch The models used to add adversarial noise may also mismatch with the model fine-tuned by DreamBooth. We examine the effectiveness of adversarial noise learned on stable diffusion v2.1 against customization based on stable diffusion v2.1 and v1.4, or vice versa in Table 4. The conclusion of model mismatch resembles those of prompt mismatch. Although there is an overall decline in performance, the critical matrix, ISM, remains considerable. This also implies that in a model mismatch scenario, SimAC + Anti-DreamBooth can still protect user portrait privacy. + +
Train [v]Test [v]“a photo of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.3187.0738.860.21
skst@t0.3081.3639.670.31
Train [v]Test [v]“a dslr portrait of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1296.8742.100.15
skst@t0.2354.4237.770.52
Train [v]Test [v]“a photo of [v] person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1291.9043.970.06
skst@t0.1630.5440.630.25
Train [v]Test [v]“a photo of [v] person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.0566.1942.770.12
sksst@t0.0623.2037.740.28
+ +Table 3. Prompt mismatch between training and testing on CelebA-HQ. The training prompt is "a photo of sks person" and the inference prompt uses rare identifier "sks" or "t@t". + +
TrainTest“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.3187.0738.860.21
v1.4v2.10.3870.0739.220.35
v2.1v1.40.0199.8662.090.02
TrainTest“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1296.8742.100.15
v1.4v2.10.1495.1740.950.23
v2.1v1.40.2620.5416.440.57
TrainTest“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1291.9043.970.06
v1.4v2.10.1679.3942.520.13
v2.1v1.40.2082.7240.120.16
TrainTest“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.0566.1942.770.12
v1.4v2.10.0661.3641.510.10
v2.1v1.40.0853.6720.560.22
+ +Table 4. Model version mismatch during training and testing on CelebA-HQ. The training prompt is "a photo of sks person" + +# 5. Conclusion + +Current anti-customization methods, primarily relying on adversarial attacks, often overlook crucial internal properties of the diffusion model, leading to ineffective optimization. This paper addresses this issue by exploring inherent properties to enhance anti-customization. Two key aspects are examined: Analyzing the relationship between timestep selection and the model's perception in the frequency domain inspires an adaptive greedy search for optimal timesteps. Scrutinizing feature roles during denoising, resulting in a sophisticated feature-based optimization framework. Experiments show increased identity disruption, enhancing user privacy and security. + +# References + +[1] Stable-diffusion-2-1. https://huggingface.co/stabilityai/stable-diffusion-2-1.6 +[2] Making pictures of trump getting arrested while waiting for trump's arrest. https://twitter.com/EliotHiggins/status/1637927681734987777. 1 +[3] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6 +[4] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 6 +[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 6 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3 +[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015. 2 +[10] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 3 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 3 +[14] Qidong Huang, Jie Zhang, Wenbo Zhou, Weiming Zhang, and Nenghai Yu. Initiative defense against facial manipula + +tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1619-1627, 2021. 3 +[15] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, and Nenghai Yu. Shape-invariant 3d adversarial point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15335-15344, 2022. 3 +[16] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, Kui Zhang, Gang Hua, and Nenghai Yu. Pointcat: Contrastive adversarial training for robust point cloud recognition. arXiv preprint arXiv:2209.07788, 2022. +[17] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Yinpeng Chen, Lu Yuan, Gang Hua, Weiming Zhang, and Nenghai Yu. Improving adversarial robustness of masked autoencoders via test-time frequency-domain prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1600-1610, 2023. +[18] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Weiming Zhang, Feifei Wang, Gang Hua, and Nenghai Yu. Diversity-aware meta visual prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10878-10887, 2023. +[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. arXiv preprint arXiv:2311.17911, 2023. 3 +[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 6 +[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3 +[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, pages 20763-20786. 2, 3, 6, 7 +[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In 6th International Conference on Learning Representations, ICLR 2018. 4, 5 +[25] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 3 +[26] A. Mittal, A. K. Moorthy, and A. C. Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, page 4695-4708, 2012. 6 +[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image + +synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 3 +[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, pages 29894-29918. 3, 6, 7 +[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 3 +[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3 +[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR, 2021. 1 +[33] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2013. 2 +[34] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. IEEE, 2020. 6 +[35] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5 +[36] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3, 6, 7 +[37] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 3 +[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3 \ No newline at end of file diff --git a/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/images.zip b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..32b1a0b6f4be7635fdf5d6ebfbb38e58727e326d --- /dev/null +++ b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:878bf35ba62a36e58efb040b013d4dba23a69219309ce2a8b2a54011f92e2d03 +size 788652 diff --git a/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/layout.json b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..435c7637f27080a3f072f2d72a53e5378288fbac --- /dev/null +++ b/2024/SimAC_ A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models/layout.json @@ -0,0 +1,9131 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "text", + "content": "SimAC: A Simple Anti-Customization Method for Protecting Face Privacy against Text-to-Image Synthesis of Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "text", + "content": "Feifei Wang" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "text", + "content": "; Zhentao Tan" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "text", + "content": ", Tianyi Wei" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "text", + "content": ", Yue Wu" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "text", + "content": ", Qidong Huang" + }, + { + "bbox": [ + 105, + 160, + 492, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 175, + 459, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 175, + 459, + 190 + ], + "spans": [ + { + "bbox": [ + 135, + 175, + 459, + 190 + ], + "type": "text", + "content": "1University of Science and Technology of China 2Alibaba Cloud" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 191, + 520, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 191, + 520, + 204 + ], + "spans": [ + { + "bbox": [ + 70, + 191, + 520, + 204 + ], + "type": "text", + "content": "{wangfeifei@, tzt@, bestwty@, hqd0037@}mail.ustc.edu.cn matthew.wy@alibaba-inc.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "type": "text", + "content": "Despite the success of diffusion-based customization methods on visual content creation, increasing concerns have been raised about such techniques from both privacy and political perspectives. To tackle this issue, several anti-customization methods have been proposed in very recent months, predominantly grounded in adversarial attacks. Unfortunately, most of these methods adopt straightforward designs, such as end-to-end optimization with a focus on adversarially maximizing the original training loss, thereby neglecting nuanced internal properties intrinsic to the diffusion model, and even leading to ineffective optimization in some diffusion time steps. In this paper, we strive to bridge this gap by undertaking a comprehensive exploration of these inherent properties, to boost the performance of current anti-customization approaches. Two aspects of properties are investigated: 1) We examine the relationship between time step selection and the model's perception in the frequency domain of images and find that lower time steps can give much more contributions to adversarial noises. This inspires us to propose an adaptive greedy search for optimal time steps that seamlessly integrates with existing anti-customization methods. 2) We scrutinize the roles of features at different layers during denoising and devise a sophisticated feature-based optimization framework for anti-customization. Experiments on facial benchmarks demonstrate that our approach significantly increases identity disruption, thereby protecting user privacy and copyright. Our code is available at: https://github.com/somuchtoe/SimAC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 647, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 685 + ], + "type": "text", + "content": "Latent Diffusion model (LDM) [12, 31, 32] has been recently proven as a strong paradigm for photorealistic visual content generation. The emergence of open-source Stable" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 230, + 545, + 421 + ], + "blocks": [ + { + "bbox": [ + 310, + 230, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 310, + 230, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 310, + 230, + 545, + 421 + ], + "type": "image", + "image_path": "18743d4cd9e1f11606bc64b573c72a40db7ef86b3c365e6716f6a3e13a57acfa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 422, + 547, + 445 + ], + "lines": [ + { + "bbox": [ + 305, + 422, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 547, + 445 + ], + "type": "text", + "content": "Figure 1. Comparison between Anti-DreamBooth before and after adding SimAC. Our method further boosts its ability to de-identity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "type": "text", + "content": "Diffusion encourages users to explore creative possibilities with LDMs. Users only need to provide several images representing the same subject along with a rare identifier to customize their diffusion models [7, 13, 22, 28]. After finetuning the model or optimizing the text embedding of rare identifiers, it can flexibly generate high-quality images containing the specified object." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "The convenience of customizing large-scale text-to-image models also allows the malicious users to generate forgery images that violate the truth. Some of them may use such technique to steal others' painting styles and generate new painting content without permission. While some of them may collect one's portrait photos that are published on the social platform, and generate fake images of this person through customization. In other words, the lawbreakers can easily produce fake news that contains celebrity photos [2], greatly misleading the public. The infringement poses threats to user privacy and intellectual property. Hence, it is essential to devise effective countermeasures to safeguard users against such malicious usage." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 225, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 225, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 225, + 703 + ], + "type": "text", + "content": "*Work done during an internship in Alibaba Cloud." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12047" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 203, + 193 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 203, + 193 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 203, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 203, + 193 + ], + "type": "image", + "image_path": "b1fa719306813ca22bcd59973e0e26be945ca4a932f2abc81a5869de3fe3a215.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 196, + 198, + 206 + ], + "lines": [ + { + "bbox": [ + 53, + 196, + 198, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 198, + 206 + ], + "type": "text", + "content": "(a) the number of absolute gradients " + }, + { + "bbox": [ + 53, + 196, + 198, + 206 + ], + "type": "inline_equation", + "content": "< 1e^{-10}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 69, + 362, + 194 + ], + "blocks": [ + { + "bbox": [ + 205, + 69, + 362, + 194 + ], + "lines": [ + { + "bbox": [ + 205, + 69, + 362, + 194 + ], + "spans": [ + { + "bbox": [ + 205, + 69, + 362, + 194 + ], + "type": "image", + "image_path": "288d32bee9a12b2e79eca9fc28e50aeb7a5acea7ee5c2e1273d04f5f15d64010.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 198, + 339, + 208 + ], + "lines": [ + { + "bbox": [ + 229, + 198, + 339, + 208 + ], + "spans": [ + { + "bbox": [ + 229, + 198, + 339, + 208 + ], + "type": "text", + "content": "(b) the maximum absolute gradients" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 212, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 257 + ], + "type": "text", + "content": "Figure 2. Distribution of Anti-DreamBooth attack gradients on different diffusion timesteps, where (a) counts the number of absolute gradients below the threshold " + }, + { + "bbox": [ + 46, + 212, + 547, + 257 + ], + "type": "inline_equation", + "content": "1e^{-10}" + }, + { + "bbox": [ + 46, + 212, + 547, + 257 + ], + "type": "text", + "content": " at each timestep, (b) presents the maximum absolute gradients at different timesteps, and (c) demonstrates how the mean and median of the absolute gradients change over timesteps. Apparently, the absolute gradients show great discrepancy on the varying timesteps and nearly zero values appear at large timesteps, which leads to ineffective noise optimization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 364, + 70, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 364, + 70, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 364, + 70, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 364, + 70, + 541, + 194 + ], + "type": "image", + "image_path": "20f4c6fcb3346b7083d62a5bad82c4d7864bbe5e3390bb5b37cd36f9f494df94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 198, + 528, + 207 + ], + "lines": [ + { + "bbox": [ + 377, + 198, + 528, + 207 + ], + "spans": [ + { + "bbox": [ + 377, + 198, + 528, + 207 + ], + "type": "text", + "content": "(c) the mean and median of the absolute gradients" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 292, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 423 + ], + "type": "text", + "content": "Current anti-customization methods are generally based on adversarial attack [9, 33]. AdvDM [23] is the pioneering work that uses adversarial noise to protect user images from being customized by diffusion models. It ingeniously combines the diffusion model with adversarial samples, firstly achieving user privacy protection. Anti-DreamBooth [36] further enhances the protective effect by employing alternate training. Both of them randomly select the timesteps from (0, max denoising steps) regarding the noise added to LDM latent, and directly employ the maximization of the LDM training loss as the optimization objective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 425, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 567 + ], + "type": "text", + "content": "However, as showcased in Figure 2, we notice that the gradient of the perturbed images is quite small and even zero when the randomly sampled timesteps are relatively large, i.e., the noisy latent is closer to Gaussian noise. This implies that within the limited steps of an adversarial attack, a large portion of steps are ineffective since the perturbed image cannot optimize the objective with such zero gradients, leading to a decrease in both protection effectiveness and efficiency. Hence, the images protected by these current methods, when customized, still allow the model to capture many details from user-input images and leak the user's privacy, as shown in Figure 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 569, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 675 + ], + "type": "text", + "content": "The fundamental reason is that these methods fail to combine the adversarial attacks with properties inherent in diffusion models. We aim to conduct an in-depth analysis of why the gradients of perturbed images have such discrepancy at different time steps and how diffusion models perceive input images at different intermediate layers. Then, based on the in-depth analysis, we propose improvements to existing customization methods from both the temporal and feature dimensions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "To observe the relationship between timesteps and the gradient of the perturbed images, we first reconstruct images predicted from noisy latents at different time steps and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 292, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 614 + ], + "type": "text", + "content": "compare them with the input images. Considering regular adversarial noises mainly affect the high frequency of images, our comparison is conducted in the frequency domain, aiming at investigating whether the model's perception lies on lower-level or higher-level. We observe that, the model focuses on the higher frequency components of images when selecting smaller time steps, and vice versa. Therefore, introducing adversarial noises at larger timesteps is ineffective, since the subtle changes perturbed on images can hardly affect the low-frequency of the generated images. To improve the effectiveness, we propose an adaptive timestep selection method to find optimal time intervals, where we iteratively update the range of selection in a greedy way. To explore the effect of different layers' features in the U-Net decoder during denoising, we employ PCA (Principal Component Analysis) to visualize them and show the discrepancy. We clearly find that within the decoder, the feature extraction gradually shifts from low-frequency to high-frequency as the layer goes deeper. It indicates the higher layers concentrates more on the texture of images, while the lower layers focus on the structure. Consequently, we select the features representing high-frequency information during optimization, since regular adversarial noises concentrate on the high frequency of images. And we construct the feature interference loss and integrate it with the diffusion denoising loss as the objective function, to improve the ability of identity interference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 616, + 474, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 616, + 474, + 628 + ], + "spans": [ + { + "bbox": [ + 317, + 616, + 474, + 628 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 630, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 630, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 630, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 306, + 630, + 545, + 700 + ], + "type": "text", + "content": "- We reveal the inadequate optimization steps that exist in current anti-customization methods, and gives detailed analysis regarding the perceptual discrepancy of diffusion models at different timesteps and intermediate layers, we have better aligned the optimization of adversarial gradients with diffusion models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 702, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 545, + 713 + ], + "type": "text", + "content": "- Based on analysis, we propose a simple but effective anti" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12048" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 168 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 131 + ], + "type": "text", + "content": "customization method, including adaptive greedy time selection and a feature interference loss to improve the protection ability. Our method can be easily generalized to different anti-customization frameworks and improve their performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 286, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 286, + 168 + ], + "type": "text", + "content": "- Extensive evaluations on two face datasets demonstrate that our method achieves more obvious disruption of the user's identity and provides better privacy protection." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 180, + 130, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 130, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 130, + 192 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 200, + 261, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 261, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 261, + 211 + ], + "type": "text", + "content": "2.1. Generative Models and Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "text", + "content": "Variational autoencoders (VAEs) [21] and Generative adversarial networks (GANs) [8] are popular frameworks among generative models which have strong generative ability. These models encode the data " + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "text", + "content": " as latent variables " + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "text", + "content": " and model the joint distribution " + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x,z)" + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "text", + "content": ". However, the quality of VAE samples is not competitive with GANs which are suffering from training instability [10]. Since diffusion probabilistic models (DM) [31] progressively add noise to the data from the joint distribution " + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "inline_equation", + "content": "q_{\\theta}(x_{0:T}|x_0)" + }, + { + "bbox": [ + 47, + 218, + 286, + 349 + ], + "type": "text", + "content": " and denoise step by step, the efficiency of training and quality of samples have achieved state-of-work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 350, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 286, + 445 + ], + "type": "text", + "content": "Unconditional generative models cannot produce desired samples and then models take different input as guidance have sprung up. Based on GAN, cGAN [25] generates images conditioned on the given labels " + }, + { + "bbox": [ + 47, + 350, + 286, + 445 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 350, + 286, + 445 + ], + "type": "text", + "content": " and Cycle-GAN [38] implements unpaired image translation considering the given image. Equipped with some techniques like classifier-free guidance [11], the diffusion model gains the ability to follow diverse prompts as conditions during generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 447, + 286, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 286, + 566 + ], + "type": "text", + "content": "The open-sourced Latent Diffusion Models (LDMs) [27] operate images in the latent space of low dimensions rather than pixel space which greatly reduces training computation. To support different condition inputs, they add cross-attention layers into the underlying U-Net backbone as conditioning mechanisms " + }, + { + "bbox": [ + 47, + 447, + 286, + 566 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 47, + 447, + 286, + 566 + ], + "type": "text", + "content": ". The above delicate designs make diffusion more friendly for users to create what they need. It can also assist them in designing their diffusions which synthesize visual contents that contain specific concepts (e.g., objects or styles) given by users during inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 574, + 137, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 137, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 137, + 586 + ], + "type": "text", + "content": "2.2.Customization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 594, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 700 + ], + "type": "text", + "content": "DreamBooth [28] stands out as a popular diffusion-based method for customizing text-to-image generation. This approach involves presenting " + }, + { + "bbox": [ + 47, + 594, + 286, + 700 + ], + "type": "inline_equation", + "content": "3 \\sim 5" + }, + { + "bbox": [ + 47, + 594, + 286, + 700 + ], + "type": "text", + "content": " images depicting a specific concept (e.g., a particular dog) alongside a corresponding identifier (e.g., \"a sks dog\"). DreamBooth utilizes these images to fine-tune the pre-trained Stable Diffusion model. This fine-tuning process encourages the model to \"memorize\" the concept and its identifier, enabling it to reproduce this concept in new contexts during inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "On the other hand, Textual-inversion [7] employs a dif" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "ferent approach. This method freezes the U-Net and exclusively optimizes the text embedding of unique identifiers (e.g. \"sks\") to represent the input concepts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 109, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 545, + 204 + ], + "type": "text", + "content": "Inspired by DreamBooth, numerous works have been mushroomed such as custom-diffusion [22], Sine [37], among others. In the quest for more efficient fine-tuning, DreamBooth has a successful integration with LoRA [13] and has become a very influential customization project in the community. LoRA specifically decomposes the attention layer of the vision transformer model [6] into low-rank matrices reducing the cost associated with fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 213, + 515, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 515, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 515, + 225 + ], + "type": "text", + "content": "2.3. Privacy Protection for Diffusion Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 232, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 495 + ], + "type": "text", + "content": "To alleviate the issue that private images are misused by Stable Diffusion based customization, so-called \"anticustomization\", some researchers [23, 30, 36] have recently delved into privacy protection for diffusion models. AdvDM [23] misleads the feature extraction of diffusion models. They analyze the training objective of fine-tuning and propose a direct way that uses the gradient of denoising loss as guidance to optimize the latent variables sampled from the denoising process. The generated adversarial examples degrade the generation ability of DreamBooth or other DM-based customized approaches. Inspired by [14], Antidreambooth [36] uses the alternating surrogate and perturbation learning (ASPL) to approximate the real trained models. They train the model on clean data and use these models as the surrogate model to compute the noise added to the user-provided images. The perturbed images then as the training data for surrogate model fine-tuning which mimics the real scenes. Photoguard [29] specially focuses on unauthorized image inpainting which misleads the public and does harm to personal reputation. They use both VAE encoder attack and UNet attack targeted on a gray image to hinder infringers from creating fake news." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 495, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 545, + 578 + ], + "type": "text", + "content": "Some concerns limit the application of current anticom customization methods in more practical scenarios, such as ineffective optimization, poor identity disruption, and simply using the reconstruction loss as guidance to generate adversarial examples [15-19]. Our paper aims to fill these blanks by figuring out the internal mechanisms that affects the performance of protection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 590, + 361, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 361, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 361, + 602 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 611, + 391, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 391, + 622 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 391, + 622 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": "Diffusion Model (DM) DM is a probabilistic generative model that samples from Gaussian distribution and then progressive denoise to learn the desired data distribution. Given " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_0 \\sim q(x)" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": ", the forward process adds increasing noise to the input images at each time-step " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "t \\in (0, T)" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " which produces a sequence of noisy latent, " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{x_0, x_1, \\dots, x_T\\}" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": ". The backward process trains model " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, t, c)" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " to predict the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12049" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 270, + 247 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 270, + 247 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 270, + 247 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 270, + 247 + ], + "type": "image", + "image_path": "fe297b61c3dc13b8033d9e30ff1cae5949f1d05dc5be630413f952a08339a1d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 288, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 301 + ], + "type": "text", + "content": "Figure 3. Pipeline of SimAC. We first greedily select the time step with our adaptive strategy during the feed forward phase. Then we integrate the feature interference loss with the vanilla training loss as the final objective. The noise is iteratively optimized." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "text", + "content": "noise added in " + }, + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "text", + "content": " to infer " + }, + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "text", + "content": ". During denoising, the training loss is " + }, + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 324, + 287, + 407 + ], + "type": "text", + "content": " distance, as shown in Eq.(1), Eq.(2). Although there are many implementations of text-to-image diffusion models, Stable Diffusion is one of the few open-sourced diffusion models and is widely used in the community. Thus, our method is mainly based on the checkpoints of stable diffusion provided in HuggingFace." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 418, + 286, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 418, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 59, + 418, + 286, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u n c o n d} (\\theta , x _ {0}) = \\mathbb {E} _ {x _ {0}, t, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t + 1}, t) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "3ce0bc7c3829d44479d9d96f567f8458a23de54d2e726154bdea06f2e2ada3f8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 434, + 287, + 448 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 287, + 448 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n d}} (\\theta , x _ {0}) = \\mathbb {E} _ {x _ {0}, t, c, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t + 1}, t, c) \\| _ {2} ^ {2}, \\tag {2}", + "image_path": "9bd3bd4d25f662598be702a7079fcb5a1ca0ef81dcd7354837674ab49ee6e194.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "text", + "content": " is the input image, " + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "text", + "content": " is the corresponding timestep, " + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "text", + "content": " is condition input, (e.g., text or image), " + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 458, + 287, + 481 + ], + "type": "text", + "content": " is the noise term." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 637 + ], + "type": "text", + "content": "Adversarial Attack The adversarial examples for classification are crafted to mislead the model to classify the given input to the wrong labels. However, the traditional attack strategies are not effective when dealing with generative models. For diffusion models, adversarial examples are some images that are added on imperceptible noise, causing diffusion models to consider them out of the generated distribution. In detail, this noise heightens the challenge of image reconstruction and hinders the customization capabilities of applications based on diffusion models (DMs). The optimized noise is often typically constrained to be smaller than a constant value " + }, + { + "bbox": [ + 46, + 483, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 483, + 287, + 637 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 483, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 483, + 287, + 637 + ], + "type": "text", + "content": " is determined through the following formation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 651, + 287, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 651, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 86, + 651, + 287, + 671 + ], + "type": "interline_equation", + "content": "\\delta_ {a d v} = \\arg \\max _ {\\| \\delta \\| _ {p} < \\eta} L \\left(f _ {\\theta} (x + \\delta), y _ {r e a l}\\right), \\tag {3}", + "image_path": "1637c36bdf80eb2dcdff3e87d5555a28d90b960f0935f947fc96f8677d0464dc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " is the input image, " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "y_{real}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " is the real images, and " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " is the loss function used to evaluate the performance of adversarial examples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "Projected Gradient Descent (PGD) [24] is a widely used method to iteratively optimize adversarial examples. The process is formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 117, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 117, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 117, + 545, + 144 + ], + "type": "interline_equation", + "content": "x ^ {t + 1} = \\prod_ {(x, \\eta)} \\left(x ^ {t} + \\alpha \\operatorname {s g n} \\left(\\nabla_ {x} L \\left(f _ {\\theta} (x + \\delta), y _ {\\text {r e a l}}\\right)\\right)\\right) \\tag {4}", + "image_path": "b9e790c85c35ac91817f236f49ee2dfdcb86144df50ce9f8d6257886101348aa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "x^0 = x" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": " is the input image, " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\operatorname{sgn}(\\cdot)" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": " is sign function, " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "(\\nabla_x L(f_\\theta(x + \\delta), y_{real})" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": " is the gradient of the loss function with respect to " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "x + \\delta" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": " represents the step size during each iteration and " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": " is the iteration number. With the operation " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\Pi_{(x,\\eta)}" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": ", the noise is limited to a " + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 152, + 545, + 224 + ], + "type": "text", + "content": "-ball ensuring the adversarial examples are acceptable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 232, + 373, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 373, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 373, + 243 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 250, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 334 + ], + "type": "text", + "content": "Here we delve into the properties of LDMs and analyze the potential vulnerabilities that can benefit the attack. In Sec. 3.3, we give a comprehensive analysis of the properties. In Sec. 3.4 and 3.5, we propose our method based on these analyses, mainly including two components, i.e., adaptive greedy time interval selection and the feature interference loss. The overall pipeline is illustrated in Figure 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 341, + 478, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 478, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 478, + 354 + ], + "type": "text", + "content": "3.3. Analysis on Properties of LDMs" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 360, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 545, + 502 + ], + "type": "text", + "content": "Differences at different time step. Our exploration focuses on analyzing gradients across various time intervals from a statistical standpoint. We've established a threshold of " + }, + { + "bbox": [ + 305, + 360, + 545, + 502 + ], + "type": "inline_equation", + "content": "1e^{-10}" + }, + { + "bbox": [ + 305, + 360, + 545, + 502 + ], + "type": "text", + "content": " to assess the number of gradients with absolute values below this threshold in perturbed images across different timesteps. Figure 1 illustrates a notable trend: during the latter part of the total time interval " + }, + { + "bbox": [ + 305, + 360, + 545, + 502 + ], + "type": "inline_equation", + "content": "(0, \\text{MaxTrainStep})" + }, + { + "bbox": [ + 305, + 360, + 545, + 502 + ], + "type": "text", + "content": ", there's a sharp increase in the count of absolute gradients below the specified threshold. Additionally, Figure 1 presents the maximum value, mean, and median of absolute gradient decreases in the perturbed images when timesteps get larger." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "text", + "content": "As shown Eq.(5), the forward process admits sampling " + }, + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "text", + "content": " at an arbitrary timestep " + }, + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "text", + "content": ". The magnitude of the noise corresponding to timestep is determined by a pre-defined noise schedule. To ensure that the final latent " + }, + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 305, + 503, + 545, + 575 + ], + "type": "text", + "content": " conforms to a standard normal distribution, the amplitude of noise injection gradually increases with the timestep." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 341, + 583, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 583, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 341, + 583, + 545, + 597 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right) \\tag {5}", + "image_path": "2a757ae2137b58e893bc4fb30907937bde06883074e763580a8932d77f1f6362.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": "During denoising, we reconstruct the input image " + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": " based on the noisy latent " + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": " and visualize the difference between the reconstructed image and the input image in the frequency domain at different timesteps. As shown in Figure 4, when " + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "t \\in [0,100]" + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": ", the main difference between the original and reconstructed images lies in the high-frequency components. Our initial expectation is to disrupt the reconstruction of the original image by introducing high-frequency adversarial noise, aiming to achieve anti-customization. How-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12050" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 124, + 160 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 124, + 160 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 124, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 124, + 160 + ], + "type": "image", + "image_path": "b180f286dbb87f5ade78d118a771268a1860d5b4599e12bb08361b550acdbde6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "text", + "content": "Figure 4. Frequency domain residual analysis. The X-axis is the distance from the low-frequency center (from low frequency to high frequency), and the Y-axis represents total magnitude of frequency residual. It can be seen that when " + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "text", + "content": " is small, the high-frequency difference exceeds the low-frequency difference. As " + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "text", + "content": " increasing, the low-frequency difference gradually dominated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 205, + 160 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 205, + 160 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 205, + 160 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 205, + 160 + ], + "type": "image", + "image_path": "cf95b0908637bcbb0d6b2c087e6fd4f286d1dfaa99c1866df11db03e544a4f07.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 210, + 70, + 287, + 160 + ], + "blocks": [ + { + "bbox": [ + 210, + 70, + 287, + 160 + ], + "lines": [ + { + "bbox": [ + 210, + 70, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 210, + 70, + 287, + 160 + ], + "type": "image", + "image_path": "530f3e19b45f68a81b84385c5e4f2b2e6afc233aa9738b0e20bbd6b0968633e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 286, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 339 + ], + "type": "text", + "content": "ever, when " + }, + { + "bbox": [ + 46, + 255, + 286, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 255, + 286, + 339 + ], + "type": "text", + "content": " increases, such as " + }, + { + "bbox": [ + 46, + 255, + 286, + 339 + ], + "type": "inline_equation", + "content": "t \\in [700,800]" + }, + { + "bbox": [ + 46, + 255, + 286, + 339 + ], + "type": "text", + "content": ", the difference in low-frequency components between the reconstructed images and the input images dominates most of the frequency domain differences between the two. Therefore, this can explain why, when the timestep of noise scheduler is large, the gradient of the perturbed image is close to zero and result in ineffective noise optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "text", + "content": "Differences at Different UNet Layers. Inspired by [35], we utilize PCA to visualize the output features of each layer in the U-Net decoder blocks during denoising. The decoder block consists of self-attention, cross-attention, and residual blocks, we select features of residual blocks. There are 11 layers in total, and the layers are visualized at timestep=500. As Figure 5 shows, with the output feature of UNet decoder blocks increasing, the visualized features gradually change from depicting structures and other low-frequency information to capturing texture and similar high-frequency information. Since our noise is intended to disrupt high-frequency information, the deeper features are good perturbation objects for our adversarial noises to reinforce interference which attempts to perturb high-frequency components in generated images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 261, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 261, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 261, + 540 + ], + "type": "text", + "content": "3.4. Adaptive Greedy Time Interval Selection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": "To achieve more effective privacy protection within a limited number of noise injection steps, we propose a fast adaptive time interval selection strategy. Firstly, based on the input image, we randomly select five timesteps from the interval (0, max training step). For each time step, we compute the gradient concerning the input image. The absolute values of the gradients are summed. If the sum of absolute gradient values is the minimum at timestep " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": ", the corresponding interval " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "(t - 20, t + 20)" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": " is then removed. For the timestep with maximal sum, noise is computed and added to the image and this noised image becomes the input for the next round of gradient computation. This process is repeated until the final time interval length is no greater than 100. We obtain the ultimate interval used for noise injection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "Our method relies on PGD[24] (Projected Gradient Descent), an iterative approach for adversarial attacks. At each iteration, computing the gradient becomes pivotal as it signifies the rate of change of the objective function concerning the perturbed images. The gradient approaching zero implies the possibility of iterative optimization but with minimal alterations to the objective function. This ultimately leads to reduced attack efficiency and diminished effectiveness. Thus, randomly selecting timesteps makes it more challenging to optimize the objective function under the same noise budget and results in decreased efficiency and effectiveness in countering customization. This highlights the crucial need for employing our adaptive, greedy time interval selection for perturbation." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 253, + 545, + 552 + ], + "blocks": [ + { + "bbox": [ + 306, + 253, + 545, + 552 + ], + "lines": [ + { + "bbox": [ + 306, + 253, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 545, + 552 + ], + "type": "image", + "image_path": "5784d6a4494f50a286559a152acb8e71b348da39e5cf7385719f39a0840455b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": "Analysis on efficacy. Here we simply analyze our adaptive greedy selection for timestep " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": " (the above algorithm). Its key steps in each iteration are randomly sampling 5 timesteps " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "T_{s} = \\{t_{j}\\}_{j = 1}^{5}" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": " and rescaling the timestep range by deleting the interval of " + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "inline_equation", + "content": "t \\in T_{s}" + }, + { + "bbox": [ + 304, + 569, + 545, + 653 + ], + "type": "text", + "content": ", which leads to the smallest gradient absolute value. Thus avoiding very small or even zero gradients during optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Theorem 1. Suppose the timestep range is rescaled from " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "T = A \\cup B" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "T' = (A \\setminus \\Delta A) \\cup (B \\setminus \\Delta B)" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " in some iteration, where timestep " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "t \\in B" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " leads to zero gradients, satisfying " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "A \\cap B = \\emptyset" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\Delta A \\cup \\Delta B" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " denotes the deleted interval, then we have " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "E_{t \\sim T} |\\nabla_x \\mathcal{L}_{DM}| < E_{T_s \\sim T} E_{t \\sim T'} |\\nabla_x \\mathcal{L}_{DM}|" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12051" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 45, + 70, + 85, + 107 + ], + "blocks": [ + { + "bbox": [ + 45, + 70, + 85, + 107 + ], + "lines": [ + { + "bbox": [ + 45, + 70, + 85, + 107 + ], + "spans": [ + { + "bbox": [ + 45, + 70, + 85, + 107 + ], + "type": "image", + "image_path": "5d9fe4e4f9da23b2e0b620234bd8f65b8e09fd3ef4d9a870a1c10b089610a02b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 124, + 545, + 147 + ], + "lines": [ + { + "bbox": [ + 45, + 124, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 45, + 124, + 545, + 147 + ], + "type": "text", + "content": "Figure 5. PCA visualization of features. The output feature of residual blocks are visualized at timestep=500. As feature goes deeper, the high-frequency information captured by the feature becomes more significant." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 87, + 71, + 124, + 107 + ], + "blocks": [ + { + "bbox": [ + 87, + 71, + 124, + 107 + ], + "lines": [ + { + "bbox": [ + 87, + 71, + 124, + 107 + ], + "spans": [ + { + "bbox": [ + 87, + 71, + 124, + 107 + ], + "type": "image", + "image_path": "116e7b02e75e4334c2c7946a0286cff766c215cb7ae32d3e2410c5cc77029158.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 126, + 71, + 161, + 107 + ], + "blocks": [ + { + "bbox": [ + 126, + 71, + 161, + 107 + ], + "lines": [ + { + "bbox": [ + 126, + 71, + 161, + 107 + ], + "spans": [ + { + "bbox": [ + 126, + 71, + 161, + 107 + ], + "type": "image", + "image_path": "64177697146243cc0dbea35ed76c7bb8aef09cd84df0cfbf8a0d5751ec0746ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 162, + 71, + 199, + 107 + ], + "blocks": [ + { + "bbox": [ + 162, + 71, + 199, + 107 + ], + "lines": [ + { + "bbox": [ + 162, + 71, + 199, + 107 + ], + "spans": [ + { + "bbox": [ + 162, + 71, + 199, + 107 + ], + "type": "image", + "image_path": "ba0c9dc5d1db9af27eb35b89aa83c6737af78410f1b0a6192e115d83dcc865c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 202, + 71, + 238, + 107 + ], + "blocks": [ + { + "bbox": [ + 202, + 71, + 238, + 107 + ], + "lines": [ + { + "bbox": [ + 202, + 71, + 238, + 107 + ], + "spans": [ + { + "bbox": [ + 202, + 71, + 238, + 107 + ], + "type": "image", + "image_path": "ff8485302aef3db6bb0e14dd75463e7a6a17e9a5fe7b33375f8167b7933f82e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 71, + 275, + 107 + ], + "blocks": [ + { + "bbox": [ + 239, + 71, + 275, + 107 + ], + "lines": [ + { + "bbox": [ + 239, + 71, + 275, + 107 + ], + "spans": [ + { + "bbox": [ + 239, + 71, + 275, + 107 + ], + "type": "image", + "image_path": "81bfe356bbb90efecb14e5f8d0d82c13c5f694a294c14c9e52f41be9fec551bc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 277, + 71, + 315, + 107 + ], + "blocks": [ + { + "bbox": [ + 277, + 71, + 315, + 107 + ], + "lines": [ + { + "bbox": [ + 277, + 71, + 315, + 107 + ], + "spans": [ + { + "bbox": [ + 277, + 71, + 315, + 107 + ], + "type": "image", + "image_path": "0147ae25280bb7572f5d0f4bd8955839919131077fd2cb927db82f64e5b75593.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 351, + 107 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 351, + 107 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 351, + 107 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 351, + 107 + ], + "type": "image", + "image_path": "44c8de6128b2fe3ca14c8003032019eee927ca19a442137ec4087ffb0b4c08d2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 353, + 71, + 391, + 107 + ], + "blocks": [ + { + "bbox": [ + 353, + 71, + 391, + 107 + ], + "lines": [ + { + "bbox": [ + 353, + 71, + 391, + 107 + ], + "spans": [ + { + "bbox": [ + 353, + 71, + 391, + 107 + ], + "type": "image", + "image_path": "c451000ed1b6dafe575399d49c13e852f8479262c99e82229359c54fe18a4524.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 429, + 107 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 429, + 107 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 429, + 107 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 429, + 107 + ], + "type": "image", + "image_path": "16a4572380c47ad0d89b00869fc0ce44339af38e5d748ad942696b998b7c2e5d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 430, + 71, + 465, + 107 + ], + "blocks": [ + { + "bbox": [ + 430, + 71, + 465, + 107 + ], + "lines": [ + { + "bbox": [ + 430, + 71, + 465, + 107 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 465, + 107 + ], + "type": "image", + "image_path": "156a615e14979e42c83e29cb699f7a3a154d5e710e3169e218733d1910c0c385.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 467, + 71, + 504, + 107 + ], + "blocks": [ + { + "bbox": [ + 467, + 71, + 504, + 107 + ], + "lines": [ + { + "bbox": [ + 467, + 71, + 504, + 107 + ], + "spans": [ + { + "bbox": [ + 467, + 71, + 504, + 107 + ], + "type": "image", + "image_path": "e1e4a09a348669b0044e1ac20048400d708a77443437b1f4a8c384bbf26bd4b1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 506, + 71, + 541, + 107 + ], + "blocks": [ + { + "bbox": [ + 506, + 71, + 541, + 107 + ], + "lines": [ + { + "bbox": [ + 506, + 71, + 541, + 107 + ], + "spans": [ + { + "bbox": [ + 506, + 71, + 541, + 107 + ], + "type": "image", + "image_path": "819a421bc3b6409f91a1d95fc267038b34da33f1643dccc4dac8564bab384305.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 44, + 112, + 90, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 112, + 90, + 121 + ], + "spans": [ + { + "bbox": [ + 44, + 112, + 90, + 121 + ], + "type": "text", + "content": "portrait image" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 93, + 113, + 116, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 113, + 116, + 120 + ], + "spans": [ + { + "bbox": [ + 93, + 113, + 116, + 120 + ], + "type": "text", + "content": "layer 0" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 113, + 154, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 113, + 154, + 120 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 154, + 120 + ], + "type": "text", + "content": "layer 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 170, + 113, + 192, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 113, + 192, + 120 + ], + "spans": [ + { + "bbox": [ + 170, + 113, + 192, + 120 + ], + "type": "text", + "content": "layer 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 208, + 113, + 231, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 113, + 231, + 120 + ], + "spans": [ + { + "bbox": [ + 208, + 113, + 231, + 120 + ], + "type": "text", + "content": "layer 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 246, + 113, + 269, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 113, + 269, + 120 + ], + "spans": [ + { + "bbox": [ + 246, + 113, + 269, + 120 + ], + "type": "text", + "content": "layer 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 285, + 113, + 307, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 113, + 307, + 120 + ], + "spans": [ + { + "bbox": [ + 285, + 113, + 307, + 120 + ], + "type": "text", + "content": "layer 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 322, + 113, + 345, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 113, + 345, + 120 + ], + "spans": [ + { + "bbox": [ + 322, + 113, + 345, + 120 + ], + "type": "text", + "content": "layer 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 361, + 113, + 383, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 113, + 383, + 120 + ], + "spans": [ + { + "bbox": [ + 361, + 113, + 383, + 120 + ], + "type": "text", + "content": "layer 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 399, + 113, + 421, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 113, + 421, + 120 + ], + "spans": [ + { + "bbox": [ + 399, + 113, + 421, + 120 + ], + "type": "text", + "content": "layer 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 437, + 113, + 459, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 113, + 459, + 120 + ], + "spans": [ + { + "bbox": [ + 437, + 113, + 459, + 120 + ], + "type": "text", + "content": "layer 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 473, + 113, + 499, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 473, + 113, + 499, + 120 + ], + "spans": [ + { + "bbox": [ + 473, + 113, + 499, + 120 + ], + "type": "text", + "content": "layer 10" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 511, + 113, + 536, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 511, + 113, + 536, + 120 + ], + "spans": [ + { + "bbox": [ + 511, + 113, + 536, + 120 + ], + "type": "text", + "content": "layer 11" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 48, + 166, + 93, + 210 + ], + "blocks": [ + { + "bbox": [ + 48, + 166, + 93, + 210 + ], + "lines": [ + { + "bbox": [ + 48, + 166, + 93, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 166, + 93, + 210 + ], + "type": "image", + "image_path": "2f5d71dedf0c5c2c86784552847d978356aff5774f8c33246ed53803c47c790a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 215, + 283, + 224 + ], + "lines": [ + { + "bbox": [ + 47, + 215, + 283, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 283, + 224 + ], + "type": "text", + "content": "portrait image t ∈ [100,200] t ∈ [300,400] t ∈ [500,600] t ∈ [700,800]" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 228, + 287, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 282 + ], + "type": "text", + "content": "Figure 6. Diffusion feature over different timesteps. As the timestep of the noise scheduler increasing, part of the high-frequency information is lost and it is not good for adding high-frequency adversarial noise. This means that choosing an appropriate time sequence is also beneficial to feature interference loss." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 96, + 166, + 140, + 210 + ], + "blocks": [ + { + "bbox": [ + 96, + 166, + 140, + 210 + ], + "lines": [ + { + "bbox": [ + 96, + 166, + 140, + 210 + ], + "spans": [ + { + "bbox": [ + 96, + 166, + 140, + 210 + ], + "type": "image", + "image_path": "63ef16559c62b14d2d9da95e2c0eac01d95f7b5765ddf45a1b328f079e374610.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 143, + 166, + 187, + 210 + ], + "blocks": [ + { + "bbox": [ + 143, + 166, + 187, + 210 + ], + "lines": [ + { + "bbox": [ + 143, + 166, + 187, + 210 + ], + "spans": [ + { + "bbox": [ + 143, + 166, + 187, + 210 + ], + "type": "image", + "image_path": "6d49296633c15a35e5ab55cbae7120afe17a1cb24d3179a52794e25f16014ede.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 192, + 166, + 236, + 210 + ], + "blocks": [ + { + "bbox": [ + 192, + 166, + 236, + 210 + ], + "lines": [ + { + "bbox": [ + 192, + 166, + 236, + 210 + ], + "spans": [ + { + "bbox": [ + 192, + 166, + 236, + 210 + ], + "type": "image", + "image_path": "d515fe086cc9e37c4c5937e54750a786496f96bde09de713c2182bb422bc7169.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 239, + 166, + 283, + 210 + ], + "blocks": [ + { + "bbox": [ + 239, + 166, + 283, + 210 + ], + "lines": [ + { + "bbox": [ + 239, + 166, + 283, + 210 + ], + "spans": [ + { + "bbox": [ + 239, + 166, + 283, + 210 + ], + "type": "image", + "image_path": "5e20609d520f2196d2fc86389f1cb4e3564fd9d1e37b1b5735af331f0f67e7e2.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "text", + "content": "Proof. We can easily obtain " + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "inline_equation", + "content": "E_{T_s \\sim T} \\frac{|\\Delta A|}{|A|} < E_{T_s \\sim T} \\frac{|\\Delta B|}{|B|}" + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "text", + "content": " via proof by contradiction. For simplicity, we denote " + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\nabla_x \\mathcal{L}_{DM}" + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "inline_equation", + "content": "LHS" + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "inline_equation", + "content": "RHS" + }, + { + "bbox": [ + 46, + 298, + 287, + 338 + ], + "type": "text", + "content": " denote left and right, then" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 47, + 343, + 297, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 297, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 297, + 427 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L H S = p (t \\in A) E _ {t \\sim A} | g (t) | + p (t \\in B) E _ {t \\sim B} | g (t) | \\\\ = \\frac {| A |}{| A | + | B |} E _ {t \\sim A} | g (t) | \\approx E _ {T _ {s} \\sim T} \\frac {| A | \\cdot E _ {t \\sim A \\backslash \\Delta A} | g (t) |}{| A | + | B |} \\\\ < E _ {T _ {s} \\sim T} \\frac {| A | - | \\Delta A |}{| A | + | B | - | \\Delta A | - | \\Delta B |} E _ {t \\sim A \\backslash \\Delta A} | g (t) | \\\\ = E _ {T _ {s} \\sim T} E _ {t \\sim A \\backslash \\Delta A} p (t \\in A \\backslash \\Delta A) | g (t) | = R H S \\\\ \\end{array}", + "image_path": "f1e03d780da00d9951afa727c077cce31f7d0436876fa4c5d402439999e05307.jpg" + } + ] + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 277, + 434, + 286, + 442 + ], + "blocks": [ + { + "bbox": [ + 277, + 434, + 286, + 442 + ], + "lines": [ + { + "bbox": [ + 277, + 434, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 277, + 434, + 286, + 442 + ], + "type": "image", + "image_path": "4b71cf575c980efa973f49ea1af1636b642ade1055e787c1a3a3fec55774f825.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 47, + 452, + 190, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 190, + 463 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 190, + 463 + ], + "type": "text", + "content": "3.5. Feature Interference Loss" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 470, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 529 + ], + "type": "text", + "content": "Based on the analysis in Sec 3.3, we propose the feature interference loss and integrate it with the vanilla training loss as the overall objective for optimization. This loss calculates the layer-wise euclidean distance between the intermediate features in some specially selected layers, i.e.," + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 124, + 533, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 533, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 124, + 533, + 287, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\mathbb {E} \\| f _ {l} ^ {t ^ {*}} - f _ {l} ^ {t} \\| _ {2} ^ {2} \\tag {6}", + "image_path": "2d37320acb0f794ea5f59510effaf09884775ea1bfb0a7e19bf2b80fb908fd84.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 130, + 551, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o n d}} + \\lambda \\mathcal {L} _ {f} \\tag {7}", + "image_path": "56a8d8df4676c33e86f71b6b2d9e558c5721491935d250921e03ab53928c0b9c.jpg" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "inline_equation", + "content": "f_{l}^{t^{*}}" + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": " is the output features of the selected layers sets " + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": " for the input image, " + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "inline_equation", + "content": "f_{l}^{t}" + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": " is the output features for current perturbed images, and " + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": " represents the weighting coefficient for the feature interference loss." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "We visualized the output features of layer 7 at different time steps. As Figure 6 shows, when the timesteps increase, more noise is added to the latent and the response to the high-frequency components of the input image gradually decreases. Therefore, utilizing feature interference loss as the optimization objective for adversarial noise in small time intervals is a better choice to keep the effectiveness of adversarial noise disturbing high-frequency information." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 167, + 386, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 386, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 386, + 180 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 306, + 187, + 356, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 187, + 356, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 187, + 356, + 200 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 207, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 302 + ], + "type": "text", + "content": "Dataset We utilized two facial datasets for experiments including Celeb-HQ [20] and VGGFace2 dataset[3]. The dataset comprises about 50 individuals and aligns with the Anti-DreamBooth, with each individual including at least 15 clear face images for customization. Since Anti-dreambooth needs alternating training, the dataset is divided into three sets, including set A, set B, and set C and each set contains 5 portrait images." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 304, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 375 + ], + "type": "text", + "content": "Model Since the open-sourced stable diffusion is the most popular implementation of latent diffusion among the community, our experiments mainly use the SD-v2.1 [1]. To test the performance of our method in a black-box scenario, we assume the versions of Stable Diffusion between anticustomization and customization are the same or different." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "content": "Baseline We compare several open-source methods that employ adversarial attacks on diffusion models to protect user images from being misused by text-to-image diffusion models, including Photoguard [29], AdvDM [23] and AntiDreamBooth [36]. Due to the high GPU memory requirements of the complete PhotoGuard, we only utilize the VAE encoder attack strategy in its paper for comparison." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "type": "text", + "content": "Metric We utilize a face detector named Retinaface [5] to detect if there is a face in the image and use the Face Detection Failure Rate (FDFR) to assess the level of disruption to the generated faces. Upon detecting a face, we encode it using ArcFace [4] and calculate the cosine similarity between the protected image and clean input, measuring the identity resemblance between the detected face in the generated image and that of users. This matrix is defined as Identity Score Matching (ISM). In addition, the image quality is quantified by BRISQE [26], and the quality of detected facial images is measured through SER-FIQ [34]." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details We set the same noise budget for all methods " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\eta = 16 / 255" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". Additionally, the optimization steps and step size align with the settings specified in each baseline. The number of training epochs is 50 and each epoch includes 3 steps for surrogate model training and 9 steps for adversarial noise optimization. The step size for adding noise is set as 0.005 and the learning rate for the training model is set as 5e-7 which is appropriate for human faces. The maximum adaptive greedy search steps are set as 50 by default. After training, we used the noised im" + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12052" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 279, + 328 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 279, + 328 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 279, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 279, + 328 + ], + "type": "table", + "html": "
CelebA-HQ
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2541.0919.380.55
AdvDM [23]0.3270.4838.170.20
Anti-DB [36]0.2877.2837.430.20
Anti-DB + SimAC0.3187.0738.860.21
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad [29]0.2028.5029.330.59
AdvDM [23]0.2565.3737.860.41
Anti-DB [36]0.1986.8038.900.27
Anti-DB + SimAC0.1296.8742.100.15
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.1830.0726.960.40
AdvDM[23]0.2935.1036.460.36
Anti-DB[36]0.2242.8640.340.28
Anti-DB + SimAC0.1291.9043.970.06
Method“a photo of sks person in front of eiffel tower”
ISM ↓FDFR ↑BRISQUE ↑SER-FQA↓
PhotoGurad[29]0.0850.9532.820.40
AdvDM[23]0.0938.1036.020.30
Anti-DB[36]0.0656.2641.350.22
Anti-DB + SimAC0.0566.1942.770.12
", + "image_path": "2f0a66adeb55d3abb46e41b0465943a5f7bb61688439127cff21993f33c1b2a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 376, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 485 + ], + "type": "text", + "content": "age for customization. The default base model is stable diffusion v2.1 combined with DreamBooth. According to the above analysis of the middle layer of the model, we select the 9, 10, and 11 layer features that can best represent the high-frequency information of the image for additional disturbance. The weight " + }, + { + "bbox": [ + 46, + 376, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 376, + 287, + 485 + ], + "type": "text", + "content": " of feature interference loss is set to 1. After finetuning 1000 steps, we save the model checkpoints and conduct inference. For each prompt, 30 images in .png format are generated for metric calculation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 491, + 237, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 237, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 237, + 504 + ], + "type": "text", + "content": "4.2. Comparison with Baseline Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": "Quantitative Results To test the effectiveness of our approach in enhancing the protection of users' portrait images, we conduct a quantitative comparison under three prompts in Table 1 and Table 2. To be more practical, we list four text prompts for inference, the first one \"a photo of sks person\" is the same as the prompt used in training, while the other three were prompts unseen before. For each prompt, we randomly select 30 generated images to compute each metric, reporting their average values." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "We can find that our method greatly improves the performance of Anti-DreamBooth and outperforms other baselines across all prompts. Due to our analysis of the unique properties of the diffusion model during denoising and effective addition of high-frequency noise, the face detection failure rates greatly increase, and the identity matching scores between the detected face and the input image are the lowest among all methods. This implies that our method is" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 535, + 328 + ], + "blocks": [ + { + "bbox": [ + 47, + 331, + 288, + 365 + ], + "lines": [ + { + "bbox": [ + 47, + 331, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 288, + 365 + ], + "type": "text", + "content": "Table 1. Comparison with other open-sourced anti-customization methods on CelebA-HQ. We evaluate the performance under four different prompts during customization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 535, + 328 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 535, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 535, + 328 + ], + "type": "table", + "html": "
VGG-Face2
Method“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2929.2720.670.47
AdvDM [23]0.3263.0738.510.21
Anti-DB[36]0.3064.6737.890.22
Anti-DB + SimAC0.3180.2740.710.22
Method“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad [29]0.2517.3328.520.55
AdvDM [23]0.3067.8037.580.35
Anti-DB[36]0.2377.4738.790.29
Anti-DB + SimAC0.1196.3341.780.12
Method“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1740.8730.100.32
AdvDM[23]0.2737.7335.430.29
Anti-DB[36]0.2545.0039.250.27
Anti-DB + SimAC0.1389.6044.970.07
Method“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
PhotoGurad[29]0.1351.0730.690.41
AdvDM[23]0.1438.6735.990.31
Anti-DB[36]0.1054.9341.130.23
Anti-DB + SimAC0.0961.2042.170.10
", + "image_path": "c3d7f416f3f3fa48d5c1c5f56953b193404017df76b952c2f5211d9f6e54f893.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 331, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 365 + ], + "type": "text", + "content": "Table 2. Comparison with other open-sourced anti-customization methods on VGG-Face2. We evaluate the performance under four different prompts during customization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 376, + 521, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 521, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 521, + 388 + ], + "type": "text", + "content": "more effective in resisting abuse from customization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 389, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 484 + ], + "type": "text", + "content": "Qualitative Results We show some results in Figure 7. It's evident that SimAC combined with Anti-DreamBooth achieves a strong image disruption effect, providing the best privacy protection for the input portrait. Since PhotoGuard operates an attack on the latent space, the resulting images tend to generate patterns similar to the target latent. This attack doesn't effectively reduce the probability of facial appearances or improve the quality of generated images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 546, + 652 + ], + "type": "text", + "content": "Both AdvDM and Anti-DreamBooth use DM's training loss as an objective for optimizing noise, which, as been seen, yields similar results. The two methods aim to disrupt high-frequency components in images via adversarial noise to create artifacts and raise the bar for potential misuse of users' photos. Although the quality of generated images decreases, the results of both AdvDM and Anti-DreamBooth retain many details from the user input images. This indicates a leak of privacy for some users and isn't user-friendly. According to our analysis, both of these methods follow the training process of DM, randomly selecting timestamps for optimization. This leads to a waste of some noise addition steps, thereby reducing the efficiency of the attack and the effectiveness of protection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 659, + 440, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 440, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 440, + 671 + ], + "type": "text", + "content": "4.3. Black-Box Performance" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Prompt Mismatch When attackers use stable diffusion to customize concepts, the prompt they use might differ from our assumptions when adding noise. Thus, we use the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12053" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 93, + 421 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 93, + 421 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 93, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 93, + 421 + ], + "type": "image", + "image_path": "c4057c7074a43fff26ebeaa1d6ae7e6bcc73229e3bb5269b885eba01db66293d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 423, + 94, + 433 + ], + "lines": [ + { + "bbox": [ + 47, + 423, + 94, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 94, + 433 + ], + "type": "text", + "content": "Portrait Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 437, + 287, + 480 + ], + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 480 + ], + "type": "text", + "content": "Figure 7. Visualization results (four prompts) on CelebA-HQ. From the first row to the last row is \"a photo of sks person\", \"a dslr portrait of sks person\", \"a photo of sks person looking at the mirror\", \"a photo of sks person in front of eiffel tower\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 96, + 70, + 141, + 421 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 141, + 421 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 141, + 421 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 141, + 421 + ], + "type": "image", + "image_path": "14a263c12d2f12c6adc67fd88ba240969415fdb73d44fd01acd80ae499582741.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 424, + 138, + 433 + ], + "lines": [ + { + "bbox": [ + 100, + 424, + 138, + 433 + ], + "spans": [ + { + "bbox": [ + 100, + 424, + 138, + 433 + ], + "type": "text", + "content": "PhotoGuard" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 70, + 190, + 421 + ], + "blocks": [ + { + "bbox": [ + 144, + 70, + 190, + 421 + ], + "lines": [ + { + "bbox": [ + 144, + 70, + 190, + 421 + ], + "spans": [ + { + "bbox": [ + 144, + 70, + 190, + 421 + ], + "type": "image", + "image_path": "d06941ba839fd1308cac69c8f2c70dd5c47c38a2cbf8983787b5af56c84ceef3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 424, + 179, + 432 + ], + "lines": [ + { + "bbox": [ + 153, + 424, + 179, + 432 + ], + "spans": [ + { + "bbox": [ + 153, + 424, + 179, + 432 + ], + "type": "text", + "content": "AdvDM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 193, + 70, + 238, + 421 + ], + "blocks": [ + { + "bbox": [ + 193, + 70, + 238, + 421 + ], + "lines": [ + { + "bbox": [ + 193, + 70, + 238, + 421 + ], + "spans": [ + { + "bbox": [ + 193, + 70, + 238, + 421 + ], + "type": "image", + "image_path": "e80d6edf48e8bf88993ce14eb806fe367f93fa99f29eaa509ea2f9a3c2cbff57.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 424, + 227, + 432 + ], + "lines": [ + { + "bbox": [ + 201, + 424, + 227, + 432 + ], + "spans": [ + { + "bbox": [ + 201, + 424, + 227, + 432 + ], + "type": "text", + "content": "Anti-DB" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 241, + 70, + 287, + 421 + ], + "blocks": [ + { + "bbox": [ + 241, + 70, + 287, + 421 + ], + "lines": [ + { + "bbox": [ + 241, + 70, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 241, + 70, + 287, + 421 + ], + "type": "image", + "image_path": "4e4653885dff5433de2cbc5800881cb6eaf6642f2a264a7f7a6717dfd402da65.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 424, + 288, + 433 + ], + "lines": [ + { + "bbox": [ + 236, + 424, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 236, + 424, + 288, + 433 + ], + "type": "text", + "content": "Anti-DB+SimAC" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 504, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 576 + ], + "type": "text", + "content": "\"a photo of sks person\" during perturbation learning and change the rare identifier \"sks\" to \"t@t\" during DreamBooth model finetuing. The results in Table 3 indicate a decrease in performance in this scenario, but we notice that the Identity Score Matching(ISM), representing identity similarity remains consistent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Model Mismatch The models used to add adversarial noise may also mismatch with the model fine-tuned by DreamBooth. We examine the effectiveness of adversarial noise learned on stable diffusion v2.1 against customization based on stable diffusion v2.1 and v1.4, or vice versa in Table 4. The conclusion of model mismatch resembles those of prompt mismatch. Although there is an overall decline in performance, the critical matrix, ISM, remains considerable. This also implies that in a model mismatch scenario, SimAC + Anti-DreamBooth can still protect user portrait privacy." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 539, + 252 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 539, + 252 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 539, + 252 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 539, + 252 + ], + "type": "table", + "html": "
Train [v]Test [v]“a photo of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.3187.0738.860.21
skst@t0.3081.3639.670.31
Train [v]Test [v]“a dslr portrait of [v] person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1296.8742.100.15
skst@t0.2354.4237.770.52
Train [v]Test [v]“a photo of [v] person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.1291.9043.970.06
skst@t0.1630.5440.630.25
Train [v]Test [v]“a photo of [v] person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
skssks0.0566.1942.770.12
sksst@t0.0623.2037.740.28
", + "image_path": "6e7a32a55323b8a230886b4cf6f579483eea2f8e75e000418218fb5118a83174.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 321, + 294, + 529, + 508 + ], + "blocks": [ + { + "bbox": [ + 305, + 254, + 545, + 288 + ], + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 288 + ], + "type": "text", + "content": "Table 3. Prompt mismatch between training and testing on CelebA-HQ. The training prompt is \"a photo of sks person\" and the inference prompt uses rare identifier \"sks\" or \"t@t\"." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 294, + 529, + 508 + ], + "lines": [ + { + "bbox": [ + 321, + 294, + 529, + 508 + ], + "spans": [ + { + "bbox": [ + 321, + 294, + 529, + 508 + ], + "type": "table", + "html": "
TrainTest“a photo of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.3187.0738.860.21
v1.4v2.10.3870.0739.220.35
v2.1v1.40.0199.8662.090.02
TrainTest“a dslr portrait of sks person”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1296.8742.100.15
v1.4v2.10.1495.1740.950.23
v2.1v1.40.2620.5416.440.57
TrainTest“a photo of sks person looking at the mirror”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.1291.9043.970.06
v1.4v2.10.1679.3942.520.13
v2.1v1.40.2082.7240.120.16
TrainTest“a photo of sks person in front of eiffel tower”
ISM↓FDFR↑BRISQUE ↑SER-FIQ↓
v2.1v2.10.0566.1942.770.12
v1.4v2.10.0661.3641.510.10
v2.1v1.40.0853.6720.560.22
", + "image_path": "1ef84162be17a85cf6588ed1c620435cd72c5dcf2658bca3d5f0fcd8c7e061af.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 510, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 305, + 510, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 545, + 533 + ], + "type": "text", + "content": "Table 4. Model version mismatch during training and testing on CelebA-HQ. The training prompt is \"a photo of sks person\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 548, + 378, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 378, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 378, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Current anti-customization methods, primarily relying on adversarial attacks, often overlook crucial internal properties of the diffusion model, leading to ineffective optimization. This paper addresses this issue by exploring inherent properties to enhance anti-customization. Two key aspects are examined: Analyzing the relationship between timestep selection and the model's perception in the frequency domain inspires an adaptive greedy search for optimal timesteps. Scrutinizing feature roles during denoising, resulting in a sophisticated feature-based optimization framework. Experiments show increased identity disruption, enhancing user privacy and security." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12054" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Stable-diffusion-2-1. https://huggingface.co/stabilityai/stable-diffusion-2-1.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 156 + ], + "type": "text", + "content": "[2] Making pictures of trump getting arrested while waiting for trump's arrest. https://twitter.com/EliotHiggins/status/1637927681734987777. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 213 + ], + "type": "text", + "content": "[3] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 286, + 267 + ], + "type": "text", + "content": "[4] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 286, + 323 + ], + "type": "text", + "content": "[5] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5203-5212, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 286, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 286, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 286, + 389 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 286, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 286, + 445 + ], + "type": "text", + "content": "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 286, + 490 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 286, + 533 + ], + "type": "text", + "content": "[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 286, + 578 + ], + "type": "text", + "content": "[10] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 286, + 613 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 690 + ], + "type": "text", + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "text", + "content": "[14] Qidong Huang, Jie Zhang, Wenbo Zhou, Weiming Zhang, and Nenghai Yu. Initiative defense against facial manipula" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1619-1627, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[15] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, and Nenghai Yu. Shape-invariant 3d adversarial point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15335-15344, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "text", + "content": "[16] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Hang Zhou, Weiming Zhang, Kui Zhang, Gang Hua, and Nenghai Yu. Pointcat: Contrastive adversarial training for robust point cloud recognition. arXiv preprint arXiv:2209.07788, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 259 + ], + "type": "text", + "content": "[17] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Yinpeng Chen, Lu Yuan, Gang Hua, Weiming Zhang, and Nenghai Yu. Improving adversarial robustness of masked autoencoders via test-time frequency-domain prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1600-1610, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "text", + "content": "[18] Qidong Huang, Xiaoyi Dong, Dongdong Chen, Weiming Zhang, Feifei Wang, Gang Hua, and Nenghai Yu. Diversity-aware meta visual prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10878-10887, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "text", + "content": "[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. arXiv preprint arXiv:2311.17911, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 382, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 447 + ], + "type": "text", + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 590 + ], + "type": "text", + "content": "[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, pages 20763-20786. 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In 6th International Conference on Learning Representations, ICLR 2018. 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 657 + ], + "type": "text", + "content": "[25] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[26] A. Mittal, A. K. Moorthy, and A. C. Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, page 4695-4708, 2012. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12055" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 696 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 171 + ], + "type": "text", + "content": "[28] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, ICML 2023, pages 29894-29918. 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "text", + "content": "[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. arXiv preprint arXiv:2302.04222, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 316 + ], + "type": "text", + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "text", + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "text", + "content": "[33] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 473 + ], + "type": "text", + "content": "[34] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. IEEE, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "type": "text", + "content": "[35] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "text", + "content": "[36] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 587, + 287, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 287, + 641 + ], + "type": "text", + "content": "[37] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 643, + 287, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 287, + 696 + ], + "type": "text", + "content": "[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12056" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_content_list.json b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d45d6b8259128e383cff81803e22bb38df3ebcb --- /dev/null +++ b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_content_list.json @@ -0,0 +1,2034 @@ +[ + { + "type": "text", + "text": "SimDA: Simple Diffusion Adapter for Efficient Video Generation", + "text_level": 1, + "bbox": [ + 155, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhen Xing1,2", + "bbox": [ + 184, + 179, + 292, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Dai3", + "bbox": [ + 316, + 180, + 379, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Han Hu3", + "bbox": [ + 405, + 180, + 475, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zuxuan $\\mathbf{W}\\mathbf{u}^{1,2\\dagger}$", + "bbox": [ + 500, + 180, + 617, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu-Gang Jiang1,2", + "bbox": [ + 643, + 180, + 781, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University", + "bbox": [ + 174, + 200, + 800, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shanghai Collaborative Innovation Center of Intelligent Visual Computing", + "bbox": [ + 184, + 218, + 784, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Microsoft Research Asia", + "bbox": [ + 380, + 234, + 589, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://chenhsing.github.io/SimDA", + "bbox": [ + 285, + 253, + 678, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9efabbe8c1a1418a10d2477a63515b1f16e8973f54cf5bdd450b7cf92f3b4974.jpg", + "image_caption": [ + "A horse galloping through van Gogh's 'Starry Night'." + ], + "image_footnote": [], + "bbox": [ + 125, + 303, + 210, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54c5a14d39d31c8c500088b7d2158eb63e88d6297bed8daff155c8efc050de3f.jpg", + "image_caption": [ + "oping through van Gogh's 'Starry Night'." + ], + "image_footnote": [], + "bbox": [ + 212, + 304, + 297, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51af8c497a376e1777b5bd3aed89a8c27b33625e91aa7dd3c924f8f875d44881.jpg", + "image_caption": [ + "van Gogh's 'Starry Night'." + ], + "image_footnote": [], + "bbox": [ + 300, + 304, + 383, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c84413c441e44bab9c4cecc026c5b864f8da99d6e2ad6cd749926d0354d9419.jpg", + "image_caption": [ + "Fireworks being displayed for a crowd of people." + ], + "image_footnote": [], + "bbox": [ + 127, + 388, + 210, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/497d9f3f758f7784f7197e9a9b2b4e023c0c6122604da8da70f4e9cd76e4d851.jpg", + "image_caption": [ + "being displayed for a crowd of people." + ], + "image_footnote": [], + "bbox": [ + 212, + 388, + 297, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a71a8d13641763cbfb4f9ea0c5d59e96d814421a6d1b59b26443a384f84cb45.jpg", + "image_caption": [ + "A dog wearing virtual reality goggles on the grass." + ], + "image_footnote": [], + "bbox": [ + 127, + 470, + 210, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ec2deaf54f4f75a87fd1dcca5989858860949ebad426d8f921202a9f9df494d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 470, + 299, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81e6591cb1bcf9611d73738d3441118fddd7180bd0fee80e070b0ffaadd37fac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 470, + 385, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed2f153bbf69a6e5061b9bdadd40a643807d6b92eeec96add1c443bb8291b524.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 470, + 470, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2eee7dfcfd07a85522d6080da5f5b32d23af35741f91587126d25457c5bd8990.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 129, + 554, + 212, + 618 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0eed4c60c3aa769d57c76e794b23caf49aae0f1367931cd7ac55d9d38769b7a0.jpg", + "image_caption": [ + "Time lapse at a fantasy landscape, 4k, high resolution.", + "(a) Text-to-Video Generation", + "Figure 1. Examples of our SimDA results: (a) The results of open-wild Text-to-Video Generation. (b) Text-guided video Editing results using one text-video pair tuning." + ], + "image_footnote": [], + "bbox": [ + 215, + 554, + 299, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a fantasy landscape, 4k, high resolution.", + "bbox": [ + 214, + 619, + 468, + 632 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d7f70d1a3f76912b31a1d576a6092f1ca3ff2690b3f87936a4ad8737e16d5d6.jpg", + "image_caption": [ + "A jeep car is moving on the road." + ], + "image_footnote": [], + "bbox": [ + 493, + 303, + 578, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/291b22a2c11e7846b00f67c115996cc826d74b9ea9d7ca57455c62d949679b60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 304, + 665, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e3acac173ff9afee5172d47a6479db6709d6de6496690e5b258cb74ad195f78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 303, + 751, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87b4e1eb9dc8a50c6d93a08dbfe8c3ff156ca310a94347f0626b10f4e7815e45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 303, + 839, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b79fc1e4952cd6408806056f28c03154bc627a890ad8d90b7ee6cfa4b1343ae.jpg", + "image_caption": [ + "An AE86 car is moving on the road." + ], + "image_footnote": [], + "bbox": [ + 493, + 388, + 578, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4aabf6227c40db2689d68ceee1620c306d33587ecb03cf0b622c3494052043a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 388, + 665, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/914252461b512bbb7c69bf1dbb979ee32eebebf9b49eb2b96f1737ab39b970e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 388, + 751, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e08fe1255e39a05305ad73d066cbb8f8a2e58a28ac39997077803791be343a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 388, + 839, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1c4c33c835723bd0f72bb7a0ef3ebeaa4d8187b4607467eb3ddfae07c577294.jpg", + "image_caption": [ + "A jeep car is moving on the road, cartoon style." + ], + "image_footnote": [], + "bbox": [ + 493, + 470, + 578, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d842fdbf571bfd8b4a42392bfe3eb8d54702bf51774490192f1bf6172cfb7b68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 470, + 665, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0dc5e77d4502d5f352f1dd8638b66163cf74f4c3546e1239395f01bfb9d569b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 470, + 751, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/317364731f1dec199aac8890ff178e90ad9506449ebbba50c1bffa99aed63e05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 470, + 839, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de8c7e2cb38fb0d6d1b91210d4337244cb11def8557c33a4580a93e8b3b62463.jpg", + "image_caption": [ + "Keep car is moving in a forest, in Autumn.", + "(b) Text-guided Video Editing" + ], + "image_footnote": [], + "bbox": [ + 491, + 554, + 575, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 700, + 313, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent wave of AI-generated content has witnessed the great development and success of Text-to-Image (T2I) technologies. By contrast, Text-to-Video (T2V) still falls short of expectations though attracting increasing interest. Existing works either train from scratch or adapt large T2I model to videos, both of which are computation and resource expensive. In this work, we propose a Simple Diffusion Adapter (SimDA) that fine-tunes only 24M out of 1.1B parameters of a strong T2I model, adapting it to video generation in a parameter-efficient way. In particular, we turn the T2I model for T2V by designing light-weight spa", + "bbox": [ + 73, + 734, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tial and temporal adapters for transfer learning. Besides, we change the original spatial attention to the proposed Latent-Shift Attention (LSA) for temporal consistency. With a similar model architecture, we further train a video superresolution model to generate high-definition $(1024\\times 1024)$ videos. In addition to T2V generation in the wild, SimDA could also be utilized in one-shot video editing with only 2 minutes tuning. Doing so, our method could minimize the training effort with extremely few tunable parameters for model adaptation.", + "bbox": [ + 496, + 702, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 524, + 886, + 655, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7827", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image generation stands on top of the recent AIGC wave. It not only has a significant impact on the academic community but also achieves tremendous success in various applications, such as computer graphics, art and culture, medical imaging, etc. The approaches in this area mainly include methods based on generative adversarial networks (GANs) [29, 49-51, 84], auto-regressive transformers [20, 77, 125], and the latest diffusion models [17, 23, 38, 40, 58, 67, 68, 78, 79, 82, 91, 120]. Among them, diffusion models are the most popular owing to their strong controllability, simple stability, and amazing realism. However, video generation research lags behind due to challenges like the scarcity of publicly available datasets, difficulty in modeling temporal information, and high training costs, hindering progress in this area.", + "bbox": [ + 76, + 114, + 467, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There have been several research endeavors dedicated to exploring video synthesis [3, 9, 12, 15, 24, 27, 30, 31, 42, 48, 52, 55, 60, 62, 64, 70, 83, 89, 94, 100, 102, 107, 108, 121, 126]. In addition, some studies have employed popular diffusion models for video generation [32, 41, 43, 61, 101, 106, 116, 122, 135, 137]. However, most of them involve training models from scratch, which can be time-consuming due to the complex video data. Early attempts were also constrained by GPU memory or hardware limitations.", + "bbox": [ + 76, + 340, + 467, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, a small number of T2V (Text-to-Video) approaches have emerged, aiming to fine-tune well-established T2I (Text-to-Image) models [1, 8, 28, 39, 61, 104, 136]. They have incorporated temporal modeling modules (e.g.Imagen video [39], Video LDM [8]) into T2I models, which effectively accelerate the model convergence. However, it should be noted that training such models is still a challenging task due to the massive number of parameters (4B or even 16B) involved in the network architecture.", + "bbox": [ + 76, + 477, + 467, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the NLP field, state-of-the-art results of various tasks are generally achieved by adaptation from large pretrained models (i.e., BERT [16], LLMs [14, 74, 76, 132]). However, with the advent of increasingly larger and more powerful foundation models (e.g., GPT-4 with 100T parameters), conducting full fine-tuning of the entire models has become prohibitively expensive and infeasible in terms of training cost and GPU storage. To address the issue, numerous methods based on efficient fine-tuning have emerged rapidly in NLP [44, 45, 53, 54] and computer vision [13, 69, 124, 128].", + "bbox": [ + 76, + 628, + 467, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a parameter-efficient video diffusion model, namely Simple Diffusion Adapter (SimDA), that fine-tunes the large T2I (i.e. Stable Diffusion [79]) model for improved video generation. We only add $2\\%$ parameters compared to the T2I model. During training, we freeze the original T2I model, and only tune the newly added modules. We further propose a Latent-Shift Attention", + "bbox": [ + 76, + 795, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(LSA) to replace the original spatial attention, which significantly improves the temporal modeling capability and retains consistency without adding new parameters. To this end, our model demands less than 8GB GPU memory for training with a resolution of $16 \\times 256 \\times 256$ , while the inference time speeds up by $39 \\times$ compared to the autoregressive method CogVideo [42]. Besides, we turn an image super resolution framework into the video counterpart with similar architecture, which allows generating high-definition videos of $1024 \\times 1024$ . Our model can also be extended to the recently popular diffusion-based video editing [110], achieving significant $3 \\times$ faster training while retaining comparable results, as evidenced by the editing examples presented in Fig 1 (b). In conclusion, the contributions of this work can be summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore the simple adaptation from image diffusion to video diffusion, exhibiting that tuning extremely few parameters can achieve surprisingly good results.", + "- With the helpful light-weight adapters and the proposed latent-shift attention, our method can effectively model the temporal relations with negligible cost.", + "- Our diffusion adapter could be extended to text-guided video super-resolution and video editing, significantly facilitating the model training.", + "- SimDA significantly alleviates the training cost and speeds up the inference time, while remaining competitive results compared to other methods." + ], + "bbox": [ + 500, + 316, + 890, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 512, + 638, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Video Generation Similar to the advancements in Text-to-Image (T2I) generation [18, 71, 79, 130], early approaches for Text-to-Video (T2V) generation [55, 64, 70] were based on Generative Adversarial Networks (GANs) and primarily applied to domain-specific videos such as simple human actions [92] or clouds moving [117].", + "bbox": [ + 496, + 537, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, T2V methods are most based on fine tuning the T2I [78, 79] diffusion models. For instance, Make-A-Video [88] proposes fine-tuning a pretrained DALLE2 [78] model solely on video data to learn motion patterns, enabling T2V generation without explicitly training on text-video pairs. Video Diffusion Models [41] and Imagen Video [39] perform joint text-image and text-video training, treating images as independent frames and disabling temporal layers in the U-Net [80] architecture. Besides, Video LDM [8], Latent-Shift [1], VideoFactory [104], MagicVideo [136] and our methods utilize the popular open-sourced T2I Stable Diffusion [79] model. While the progress in video generation is impressive, the parameters of video generation can be highly large. As shown in Table 1, Make-A-Video [88] requires six models and 9.7B parameters and Imagen Video [39] utilizes eight models with 16.3B parameters, which limits the training efficiency of T2V models.", + "bbox": [ + 496, + 628, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text guided Video Editing In the realm of content gen", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7828", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eration, an alternative avenue is the manipulation of existing images [10, 37, 63, 98] and videos [6, 21, 59, 73, 87, 110, 115, 123, 134] using textual input as a means of control, rather than relying solely on unbridled text-based generation. SDEdit [63] introduces noise to images and then reconstructs them for the purpose of editing. Prompt-to-prompt [37] and Plug-and-Play [98] modify the cross-attention map by altering the textual description, thus influencing the editing process. When it comes to video editing, Tune-A-Video [110] fine-tunes the T2I model on a single video, enabling the generation of new videos with similar motion patterns. Video-P2P [59] and FateZero [73] extend the concept of Prompt-to-prompt editing to videos. Text2Live [6] divides videos into layers and enables separate editing of each layer based on text. MotionEditor [97] edits the motion of human while keeping the background.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter-Efficient Transfer Learning In NLP, parameter efficient fine-tuning [33, 44, 45, 53, 54, 93, 127] were initially proposed to address the heavy computation of full fine-tuning LLMs for various downstream tasks. These techniques aim to reduce the number of trainable parameters, thereby lowering computation costs, while still achieving or surpassing the performance of full fine-tuning. Recently, parameter-efficient transfer learning has also been explored in the field of computer vision [4, 13, 22, 26, 46, 47, 69, 95, 112, 113, 124, 128]. These methods mainly focus on adapting models within simple classification or detection tasks. In contrast, our approach focuses on adapting a T2I model for T2V generation task.", + "bbox": [ + 75, + 332, + 472, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Shift Module TSM [56] pioneered the introduction of the temporal shift module for action recognition, employing a partial channel shift along the temporal dimension. This approach seamlessly integrates temporal cues from both preceding and succeeding frames into the current frame without incurring additional computational overhead. Subsequently, TokShift [129] implemented channel shifting along the temporal dimension for transformer architectures [19, 133]. TPS [111] further shifted patches instead of channels to model the temporal correlations. However, such direct patch shifting would lead to inconsistency in generation tasks. Additionally, Latent-shift [1] and TSB [66] adapted shift module as TSM [56] within convolution blocks for video generation tasks. In this work, our latent-shift attention (LSA) employs the patch-level shifting manner. In contrast to TPS, we further propose to involve all tokens in the current frame as the keys and values, which guarantees temporal consistency during generation and significantly improves the video quality.", + "bbox": [ + 75, + 529, + 472, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 829, + 169, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the preliminaries of Latent Diffusion Model [79] in Sec. 3.1. The pipeline of SimDA is described in Sec. 3.2. Then we detail the proposed spa", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tial and temporal adapters as well as latent-shift attention in Sec. 3.3. Finally, we introduce the super resolution and text-guided video editing model in Sec. 3.4.", + "bbox": [ + 496, + 90, + 893, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries of Stable Diffusion", + "text_level": 1, + "bbox": [ + 498, + 146, + 787, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this subsection, we introduce the preliminaries of Stable Diffusion [79] model. It is a latent diffusion model that operates in the latent space of an autoencoder $\\mathcal{D}(\\mathcal{E}(\\cdot))$ , where $\\mathcal{E}$ is the encoder and $\\mathcal{D}$ is the decoder. In this model, for an image $I$ with its corresponding latent feature $\\pmb{x}_0 = \\mathcal{E}(I)$ , the diffusion forward process involves iteratively adding noise to the latent space,", + "bbox": [ + 496, + 169, + 893, + 275 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 287, + 890, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\in \\{1, \\dots, T\\}$ is the time step, $q(\\pmb{x}_t | \\pmb{x}_{t-1})$ is the conditional density of $\\pmb{x}_t$ given $\\pmb{x}_{t-1}$ , $\\mathbf{I}$ is identity matrix, and $\\alpha_t$ is hyperparameter. Alternatively, we can directly sample $\\pmb{x}_t$ at any time step from $\\pmb{x}_0$ with,", + "bbox": [ + 496, + 316, + 893, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 388, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i$ .", + "bbox": [ + 500, + 416, + 645, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the diffusion backward process, a U-Net denoted as $\\epsilon_{\\theta}$ is trained to predict the noise in the latent space, aiming to iteratively recover $x_0$ from $x_{T}$ . In this process, as the diffusion progresses and approaches a large value of $T$ , $x_0$ is completely disrupted and the latent representation $x_{T}$ approximates a standard Gaussian distribution. Consequently, the U-Net $\\epsilon_{\\theta}$ is trained to infer meaningful and valid $x_0$ from random Gaussian noises. The training object can be simplified as,", + "bbox": [ + 496, + 434, + 890, + 569 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), t} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t) \\| _ {2} ^ {2} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 580, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c$ is the embedding of condition text.", + "bbox": [ + 500, + 609, + 787, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the inference stage, it samples a valid latent representation $\\mathbf{x}_0$ from the standard Gaussian noise $\\mathbf{x}_T = \\mathbf{z}_T, \\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ using DDIM [90] sampling. Then, the model can decode $\\mathbf{x}_0$ using the decoder $\\mathcal{D}$ to generate the final image $I = \\mathcal{D}(\\mathbf{x}_0)$ . This process could generate diverse and high-quality images based on the sampled latent representations. In contrast, our method focuses on more challenging high-quality video generation.", + "bbox": [ + 496, + 626, + 893, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pipeline", + "text_level": 1, + "bbox": [ + 500, + 756, + 599, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SimDA, as shown in Fig. 2, is built upon the previously introduced Stable Diffusion [79]. For a video clip with $t$ frames, denoted as $\\{I_i\\}_{i=1}^t$ , we first pass it through a pretrained encoder $\\mathcal{E}$ to obtain the corresponding latent feature $\\{\\pmb{x}_i\\}_{i=1}^t$ . We then input the latent features to the forward diffusion process, where noise is incrementally added to the latents. In the backward diffusion process, we utilize an inflated U-Net architecture to predict the noise for the", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7829", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a79880bfbce85f15b13557ea3e5988d6d7172f34fb2a5e913e8cce5d29dfdcf.jpg", + "image_caption": [ + "Figure 2. Pipeline of our Parameter-Efficient Text-to-Video Framework. We utilize the pre-trained auto-encoder as in Stable Diffusion [79] to obtain latent representation. During training, we only update the parameters of the newly added adapter module, highlighted in yellow. The parameters of other modules are frozen, highlighted in blue." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/82fd11d50902bd68bd061b549257a918b6f17ec43245a98ca48fa7940a05a818.jpg", + "table_caption": [ + "Table 1. Model size and inference speed comparisons. The speed is measured in seconds on one A100 (80GB) GPU. The majority of results are sourced from [1]." + ], + "table_footnote": [], + "table_body": "
MethodParameters (Billion)Speed (s)
T2V CoreAuto EncoderText EncoderPrior ModelSuper ResolutionFrame InterpolationOverallTuned
CogVideo [42]7.70.10---7.715.515.5434.53
Make-A-Video [88]3.1-0.121.31.4 + 0.73.19.729.72-
Imagen Video [39]5.6-4.6-1.2 + 1.4 + 0.341.7 + 0.78 + 0.6316.2516.25-
Video LDM [8]1.510.080.12-0.981.514.202.65-
Latent-VDM [1]0.920.080.58---1.580.9228.62
Latent-Shift [1]0.880.080.58---1.530.8823.40
LVDM [35]0.960.080.12---1.161.0421.23
SimDA (Ours)0.880.080.12---1.080.02511.20
", + "bbox": [ + 78, + 474, + 893, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noisy video latents. Specifically, for the Convolution block, we inflate the 2D ResNet [34] block to a 3D block to accommodate video inputs. Additionally, we incorporate a lightweight Temporal Adapter module for temporal modeling. In the Attention block, we employ a latent-shift attention mechanism for spatial self-attention and introduce two spatial adapter modules to facilitate the transfer of video information. Further details are presented in Sec. 3.3. During inference, we employ DDIM [90] sampling to progressively denoise the latent representation sampled from a standard Gaussian distribution. Finally, we utilize a pre-trained decoder $\\mathcal{D}$ to reconstruct the video from the denoised latent.", + "bbox": [ + 75, + 612, + 470, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Modeling", + "text_level": 1, + "bbox": [ + 76, + 801, + 187, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe the proposed Spatial Adapter, Temporal Adapter, and Latent-Shift Attention in detail, which are the key components of our model.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Adapter The large-scale text-image pre-trained T2I model exhibits significant transferability, as evidenced by", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "its remarkable accomplishments in tasks such as personalized T2I generation [65, 81] and image editing [37, 131]. Consequently, we believe that employing a lightweight fine-tuning approach can effectively harness spatial information in the realm of video generation. Inspired by efficient fine-tuning techniques in NLP [45, 54] and vision tasks [13, 124], we adopt adapters due to their simplicity.", + "bbox": [ + 496, + 612, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our T2V framework, we design two types of spatial adapters (i.e., Attention Adapter and FFN Adapter) for transferring video spatial information. As shown in the bottom right of Fig. 2, both adapters employ a bottleneck architecture consisting of two fully connected (FC) layers with an intermediate activation layer. The first FC layer maps the input to a lower-dimensional space, while the second FC layer maps it back to the original dimensional. Formally, for an input feature matrix $\\mathbf{X} \\in \\mathbb{R}^{N \\times d}$ , the spatial adapter could be written as:", + "bbox": [ + 496, + 718, + 892, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S - A d p a t e r} (\\mathbf {X}) = \\mathbf {X} + \\mathbf {W} _ {\\text {u p}} \\left(\\operatorname {G E L U} \\left(\\mathbf {W} _ {\\text {d o w n}} (\\mathbf {X})\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 883, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7830", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/050f80f7babddd762d6b1e01d812c3c5ceb7cd1856cdb4f51bfc20fccf658aea.jpg", + "image_caption": [ + "Figure 3. The overview of Temporal Latent-shift Attention module. It is noted that the Latent-shift attention is performed on latent space, but the visualization overview is shown on the image level for clear presentation." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 472, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathrm{up}}$ and $\\mathbf{W}_{\\mathrm{down}}$ are the learnable matrix with dimension $d\\times l$ and $l\\times d, l < d$ . To preserve the structure of the original network and the pretrained weights, we initialize the second FC layer $\\mathbf{W}_{\\mathrm{down}}$ with zeros. To adapt to the spatial features of videos, we incorporate the adapter after the latent-shift attention layer. Additionally, we observe that adding an adapter to the feed-forward network (FFN) also helps the network transfer spatial information to videos. We will provide examples in Sec. 4.4. During training, all layers of the attention block are fixed, and only the adapters are updated.", + "bbox": [ + 75, + 299, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Adapter While the spatial adapter effectively transfers spatial information to video, modeling temporal information is crucial for T2V generation tasks. Previous approaches incorporate temporal convolution [11, 85, 86] or temporal attention [8, 88, 104] modules to capture temporal relationships. Although these modules are effective in modeling temporal dynamics, they often come with a huge number of parameters and high-dimensional input feature, resulting in significant computational and training costs.", + "bbox": [ + 75, + 465, + 468, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we utilize the temporal adapter module for temporal modeling as [69, 128]. In contrast to conventional spatial adapter modules, the temporal adapter module employs depth-wise 3D convolution instead of an intermediate activation layer [36]. The temporal adapter could be formally written as:", + "bbox": [ + 75, + 601, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T - A d p a t e r} (\\mathbf {X}) = \\mathbf {X} + \\mathbf {W} _ {\\mathrm {u p}} (3 \\mathrm {D} - \\operatorname {C o n v} (\\mathbf {W} _ {\\mathrm {d o w n}} (\\mathbf {X}))). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 703, + 467, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By utilizing 3D convolutions in lower-dimensional input, our approach significantly alleviates the complexity of temporal modeling. As a result, our method achieves efficient memory usage during training and exhibits the fastest inference speed among competitive approaches, as demonstrated in Table 1.", + "bbox": [ + 75, + 734, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Latent-Shift Attention In the original T2I framework, the attention block of the U-Net only performs self-attention for individual frames, neglecting the information from other frames. While joint-space-time attention, as demonstrated in [2, 7, 96, 114], can effectively model tem", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e848e44034601bf1ae52de4940bee4d1035b1ee513ab20c5550eec8195d33631.jpg", + "image_caption": [ + "Figure 4. During inference, we sample a novel video from the latent noise inverted from the input video, guided by an edited prompt (e.g., \"a Porsche sports car is moving on the road\")." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 885, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "poral dependencies, it introduces a quadratic complexity in terms of attention calculation. For a video with $L$ frames and $N$ tokens, the complexity of global spatial-temporal attention becomes $O(L^2 N^2)$ . To address this issue, we propose a latent-shift attention module as shown in Fig. 3. In addition to considering tokens within the current frame, we further conduct a patch-level shifting operation along the temporal dimension to shift tokens from the preceding $T$ frames onto the current frame, thereby composing a new latent feature frame. We concatenate the latent feature of the current frame with the temporally shifted latent feature, forming the keys and values for attention calculation. The latent-shift attention can be formally written as:", + "bbox": [ + 496, + 244, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {W} _ {\\mathrm {q}} \\left(\\boldsymbol {x} _ {z _ {i}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 449, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} = \\mathbf {W} _ {\\mathrm {k}} \\left[ \\boldsymbol {x} _ {z _ {i}}, \\boldsymbol {x} _ {z _ {s h i f t}} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 469, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\mathbf {W} _ {\\mathrm {v}} \\left[ \\boldsymbol {x} _ {z _ {i}}, \\boldsymbol {x} _ {z _ {\\text {s h i f t}}} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 487, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\boldsymbol{x}_{z_i}$ denotes the query frame and $[\\cdot]$ means concatenate. This approach reduces the complexity of attention to $O(2LN^2)$ , significantly lowering the computational burden compared to global attention. Moreover, it allows the model to learn the relationships between adjacent frames, ensuring better temporal consistency in video generation.", + "bbox": [ + 496, + 507, + 890, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Super Resolution and Editing Models", + "text_level": 1, + "bbox": [ + 500, + 604, + 823, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Super Resolution (SR) Due to constraints of limited GPU memory and the lack of high-resolution video-text datasets, most existing methods [1, 8, 35], including ours, are only able to generate images at a resolution of $256 \\times 256$ . To overcome this limitation and generate higher-resolution outputs, we adopt a two-stage training approach similar to cascaded Diffusion Models [40]. In the first stage, we generate videos with a $256 \\times 256$ resolution using our SimDA methods. In the second stage, we employ an LDM upsampler [79] to enhance the resolution of the videos to $1024 \\times 1024$ . We incorporate noise augmentation and noise level conditioning, and train a super-resolution model using the following equation:", + "bbox": [ + 496, + 627, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I}), t} [ \\| \\epsilon - \\epsilon_ {\\theta} ([ \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {l o w} ], \\boldsymbol {c}, t) \\| _ {2} ^ {2} ], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 830, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}_{low}$ is the low-resolution video, we concatenate it with $\\pmb{x}_t$ frame by frame following Video LDM [8]. The architecture of SR is similar to T2V model in the first stage,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7831", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/187c701cb8345bbe82218a001a6f8f596ba54c3c5baaae46b165c38bb0fdbb1d.jpg", + "image_caption": [ + "Figure 5. Text-to-Video generation comparison on the user study evaluation set." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 395, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29abc1c5a9912f55f43909fcd0118f536c739ea79705b3248e5efee5732948a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 89, + 620, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6876e3673f27beef49514899611a8b0df912bb716ca87fab64f3fc3d09dddb3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 89, + 846, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c1341273cc81340b790ccfdeb437a5bc9ab7154ed89d02fda54a8331742435b9.jpg", + "table_caption": [ + "Table 2. Text-to-Video generation comparison on MSR-VTT [118] dataset. We report the Fréchet Video Distance (FVD) scores and CLIPSIM scores." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataResolutionZero-shotParams(B)FVD(↓)CLIPSIM(↑)
GODIVA [108]MSR-VTT128x128No--0.2402
NüWA [109]MSR-VTT128x128No0.87-0.2439
Make-A-Video [88]WebVid-10M + HD-VILA-10M256x256Yes9.72-0.3049
VideoFactory [104]WebVid-10M + HD-VG-130M256x256Yes2.04-0.3005
LVDM [35]WebVid-2M256x256Yes1.167420.2381
MMVG [25]WebVid-2.5M256x256Yes--0.2644
CogVideo [42]WebVid-5.4M256x256Yes15.512940.2631
ED-T2V [57]WebVid-10M256x256Yes1.30-0.2763
MagicVideo [136]WebVid-10M256x256Yes-998-
Video-LDM [8]WebVid-10M256x256Yes4.20-0.2929
VideoComposer [105]WebVid-10M256x256Yes1.855800.2932
Latent-Shift [1]WebVid-10M256x256Yes1.53-0.2773
VideoFusion [61]WebVid-10M256x256Yes1.835810.2795
SimDA (Ours)WebVid-10M256x256Yes1.084560.2945
", + "bbox": [ + 104, + 385, + 859, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we change the original U-Net block by adding Spatial and Temporal Adapters as described in Sec. 3.3 and only finetune the newly added modules.", + "bbox": [ + 75, + 638, + 468, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-guided Video Editing In addition to performing T2V generation, our method could turn into one-shot tuning for text-guided video editing following Tune-A-Video [110]. The training pipeline of editing model is the same as our T2V method. However, for the inference stage, we adopt the DDIM inversion latents instead of random noisy latents together with edited prompt for novel video generation as shown in Fig. 4. By doing so, the pixel-level information control could remain in the inversion latent as demonstrated in [110]. Owing to the light-weight module and efficient pipeline of our method, SimDA needs fewer training steps (200 steps compared to 500 steps) and thus the training time and inference time is much faster than Tune-A-Video [110].", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 638, + 633, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 664, + 718, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our T2V method is composed of two-stage models. The first model predicts video frames with a resolution $256 \\times 256$ (with a latent size of $32 \\times 32$ ), while the second model is a $4 \\times$ upsampler, producing a resolution of $1024 \\times 1024$ . We train the general T2V model on WebVid-10M [5] dataset following [1, 8]. We follow previous methods [1, 104, 136] to report the CLIP score [75] and FVD (Frechet Video Distance) score [99] on MSR-VTT [118]. Besides, we compare the FVD score and CLIP score on the evaluation set of WebVid [5] as in VideoFactory [104]. We also compare the parameter scale and inference speed of our method with some open-sourced methods [1, 35, 61]. Finally, we also provide a user study between our work and VDM [41], Latent-shift [1], Video-Fusion [61] and LVDM [35].", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7832", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on Text-to-Video Generation", + "text_level": 1, + "bbox": [ + 76, + 90, + 419, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To fully evaluate the generation performance of our SimDA, we conduct automatic evaluations on two distinct datasets: WebVid [5] (Val), which shares the same domain as the training data and MSR-VTT [118] in a zero-shot setting.", + "bbox": [ + 75, + 113, + 467, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on MSR-VTT As shown in Table 2, we evaluate CLIPSIM [75] and FVD [99] on the widely used video generation benchmarks, MSR-VTT [118]. We randomly select one text prompt per example from MSR-VTT [118] and generate a total of 2,990 videos. Despite being a zero-shot setting, our method achieves an average CLIPSIM of 0.2945 that surpasses most of the competitors, indicating a strong semantic alignment between the generated videos and the input text. Though Make-A-Video [88] and VideoFactory [104] offer higher CLIP scores, they utilize additional large-scale HD-VILA [119] datasets for training.", + "bbox": [ + 75, + 176, + 467, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on WebVid As shown in Table 3, we create a validation set consisting of 4,476 randomly extracted text-video pairs from WebVid-10M. These pairs are not included in the training data following [104]. We conduct evaluations on this validation set and obtain impressive results. Our method achieves an FVD score of 363.98 and a CLIP-SIM score of 0.3054. These scores are significantly higher than those achieved by existing methods such as ModelScope [103] and LVDM [35]. Besides, our method shows competitive results compared to VideoFactory [104] which is trained with much larger datasets. These results clearly demonstrate the superiority of our approach.", + "bbox": [ + 75, + 347, + 467, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Evaluation In order to address the limitations of existing metrics and assess the performance of our SimDA from a human perspective, we conduct an extensive user study. The study involves comparing our method with four state-of-the-art methods. Specifically, we select two publicly available models, namely ModelScope [103] and LVDM [35]. Additionally, we consider two methods with similar scale parameters, VDM [41] and Latent-shift [1], which only showcase some samples on their websites.", + "bbox": [ + 75, + 531, + 467, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each case, participants were provided with two video samples, one is generated by our method and the other is from a competitor. They were then asked to compare the two samples in terms of video quality and text-video similarity. To ensure fairness in the comparisons, we also report the ratio of network parameters compared to ours. The results, along with the parameter ratios, are presented in Table 4. The user study approach allows us to gain in-depth insights into the subjective evaluation of our method.", + "bbox": [ + 75, + 667, + 467, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results The visualization of T2V generation results are shown in Fig. 1(a). Besides, we show the comparison results in Fig. 5. More examples can be found in our supplementary material.", + "bbox": [ + 75, + 806, + 467, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameter Size and Inference Speed We conduct a comparison of the number of parameters and inference speed,", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d1a0e71ddc0c4c763a0204953cf74cf8a0aa78275bf141796247594ed2d50d0.jpg", + "table_caption": [ + "Table 3. Text-to-video generation on the validation set of WebVid [5]. We report the FVD and CLIPSIM scores." + ], + "table_footnote": [], + "table_body": "
MethodParams(B)FVD(↓)CLIPSIM(↑)
LVDM [35]1.16455.530.2751
ModelScope [103]1.83414.110.3000
VideoFactory [104]2.04292.350.3070
SimDA (Ours)1.08363.980.3054
", + "bbox": [ + 506, + 119, + 883, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e7f65041e7ea81639f3c773e75d60e502a7346e09238311be35aca67e826fb6.jpg", + "table_caption": [ + "Table 4. User preference is depicted as a percentage indicating the proportion of individuals favoring our method over the compared approach. Param Ratio means the ratio of the network parameter v.s. Ours." + ], + "table_footnote": [], + "table_body": "
SampleMethodParam RatioQualityFaithfulness
Open WebsiteVDM [41]0.83×85.2%81.4%
Latent-Shift [1]1.41×81.5%79.3%
Pretrained ModelModelScope [103]1.69×78.3%79.5%
LVDM [35]1.07×83.4%84.7%
", + "bbox": [ + 503, + 279, + 903, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the results are presented in Table 1. For the speed comparison, we select CogVideo [42], Latent-Shift [1] and LVDM [35]. SimDA, on the other hand, stands out as it is significantly smaller than previous works and exhibits faster inference speed compared to other methods. Despite having fewer parameters, SimDA achieves superior performance in various benchmarks when compared to other methods. This validation further highlights our advantages in terms of model efficiency and performance.", + "bbox": [ + 496, + 380, + 890, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on Text-guided Video Editing", + "text_level": 1, + "bbox": [ + 498, + 527, + 849, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the methodology of previous studies [110], we employ CLIP score [75] and a user study to evaluate the performance of different methods in terms of frame consistency and textual alignment.", + "bbox": [ + 496, + 551, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we calculate the CLIP image embedding for all frames in the edited videos to measure frame consistency. The average cosine similarity between pairs of video frames is reported. Additionally, to assess textual faithfulness, we compute the average CLIP score between frames of the output videos and the corresponding edited prompts. A total of 15 videos from the dataset [72] were selected and edited based on object, background and style, resulting in 75 edited videos for each model. The average results, presented in Table 5, highlight our method's exceptional ability to achieve semantic alignment.", + "bbox": [ + 496, + 613, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Secondly, we conduct a user study involving videos and text prompts. Participants were asked to vote for the edited videos that exhibited the best temporal consistency and those most accurately matched the textual description. Table 5 demonstrates that our method, SimDA, receives the highest number of votes in both aspects, indicating superior editing quality and a strong preference from users in practical scenarios.", + "bbox": [ + 496, + 780, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7833", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1be6338bb456d98ffb5e7c4cf890fba0eccf9c81f2b71b55bc1941951d0a1b23.jpg", + "table_caption": [ + "Table 5. Quantitative comparison with evaluated baseline [110]. The \"Training\" refers to the process of optimization, and \"Memory\" refers to the GPU memory." + ], + "table_footnote": [], + "table_body": "
MethodFrame consistencyTextual alignmentRuntime [min]Memory [Gib]Params [Mb]
CLIP Score↑User Vote↑CLIP Score↑User Vote↑Training↓Inference↓Training↓Inference↓Tuned↓
Tune-A-Video [110]94.131.2%31.839.5%9.30.831.311.474.4
SimDA(Ours)94.968.8%31.960.5%2.50.428.68.824.9
", + "bbox": [ + 84, + 121, + 883, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd079eb50d409e473d462a68bbb8b1ed5ac4b844647e7953f339b5058e968f17.jpg", + "image_caption": [ + "Figure 6. Ablation of T2V generation results. TA, SA, AA, FA and LSA refer to Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-Shift Attention." + ], + "image_footnote": [], + "bbox": [ + 99, + 205, + 446, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 626, + 232, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we discuss the effect of each module. We perform experiments on $1K$ samples from the WebVid validation set.", + "bbox": [ + 75, + 650, + 468, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Temporal Adapter Temporal modeling is a crucial component of video generation. In our video editing task, when compared to methods that rely on temporal attention modeling like Tune-A-Video [110], we observe that our temporal adapter is more lightweight and achieves superior editing results as in Table 5. Additionally, we conduct ablation experiments, as shown in Table 6 and Fig. 6, where the lack of Temporal Adapter (TA) results in significantly higher FVD and chaotic temporal sequences in the generated videos.", + "bbox": [ + 75, + 685, + 468, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Spatial Adapter We also validate the effectiveness of the Spatial Adapter (SA) in transferring spatial knowledge of videos. As shown in Table 6, without the Attention Adapter (AA) and FFN Adapter (FA), the model's", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3f66cc8eec8f34997bd22341b227b2dffb7476438295b0ae0257165bc546e1ff.jpg", + "table_caption": [ + "Table 6. Ablation study on different modules. We report the FVD [99] and CLIPSIM [75] on $1K$ samples from the validation set of WebVid-10M [5]. TA, SA, AA, FA and LSA represent Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-shift Attention, respectively." + ], + "table_footnote": [], + "table_body": "
TAAAFALSAFVD(↓)CLIPSIM(↑)
w/o TA✓✓✓1470.10.2629
w/o SA✓✓811.30.2822
w/o AA✓✓✓764.80.2851
w/o FA✓✓✓623.70.2962
w/o LSA✓✓✓618.20.3011
Ours✓✓✓✓530.20.3034
", + "bbox": [ + 503, + 284, + 895, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FVD and CLIPSIM scores for generated videos will become worse. Additionally, it can be observed from the Fig. 6 that the model exhibits misconceptions in understanding the text prompt without the spatial adapter.", + "bbox": [ + 498, + 416, + 893, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Latent Shift Attention To investigate the impact of Latent-shift Attention (LSA), we replace it with regular single-frame spatial attention. Besides observing a decline in FVD and text alignment CLIPSIM scores in Table 6, we also test the CLIPSIM of each frame within the same video, which decreased from 96.4 to 94.5. This demonstrates that our LSA module can effectively model the relationship of adjacent frames, leading to more consistent videos.", + "bbox": [ + 498, + 482, + 893, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 614, + 619, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed SimDA, a parameter-efficient video diffusion model for text-guided video generation and editing. With the proposed light-weight spatial and temporal adapters, our method not only transferred from powerful spatial information but also modeled temporal relationships with the least new parameters. The experimental results demonstrated that our approach has the fastest training and inference speed while maintaining competitive generation and editing results. Our work is the first parameter-efficient video diffusion method serving as an efficient T2V fine-tuning baseline and paved the way for future research.", + "bbox": [ + 496, + 642, + 890, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledge This work was supported by National Science and Technology Major Project (No. 2021ZD0112805) and in part by National Natural Science Foundation of China (No. 62032006 and No. 62102092.). The computations in this research were performed using the CFFFplatform of Fudan University.", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7834", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, JiaBin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2, 3, 4, 5, 6, 7", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 5", + "[3] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. arXiv preprint arXiv:1710.11252, 2017. 2", + "[4] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6, 7, 8", + "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3", + "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5", + "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 4, 5, 6", + "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 2022. 2", + "[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3", + "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 5", + "[12] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In ICCV, 2019. 2", + "[13] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 2, 3, 4", + "[14] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 2", + "[15] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 2", + "[16] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.2", + "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2", + "[18] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 2022. 2", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2", + "[21] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3", + "[22] Qijun Feng, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Fdgaussian: Fast gaussian splatting from single image via geometric-aware diffusion model. arXiv preprint arXiv:2403.10242, 2024. 3", + "[23] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 2", + "[24] Jean-Yves Franceschi, Edouard Delasalles, Mickael Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In ICML, 2020. 2", + "[25] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 6", + "[26] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3", + "[27] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2", + "[28] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474, 2023. 2", + "[29] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2", + "[30] Sonam Gupta, Arti Keshari, and Sukhendu Das. Rvgan: Recurrent gan for unconditional video generation. In CVPR, 2022. 2" + ], + "bbox": [ + 509, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7835", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2", + "[32] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2", + "[33] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. arXiv preprint arXiv:2110.04366, 2021. 3", + "[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4", + "[35] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 4, 5, 6, 7", + "[36] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[37] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. *ICLR*, 2023. 3, 4", + "[38] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2", + "[39] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 4", + "[40] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2, 5", + "[41] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 6, 7", + "[42] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 4, 6, 7", + "[43] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022.2", + "[44] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3", + "[45] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4", + "[46] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Shibo Jie and Zhi-Hong Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 3", + "[48] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 2", + "[49] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2", + "[50] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020.", + "[51] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 2", + "[52] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 2", + "[53] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3", + "[54] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4", + "[55] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2", + "[56] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 3", + "[57] Jiawei Liu, Weining Wang, Wei Liu, Qian He, and Jing Liu. Ed-t2v: An efficient training framework for diffusion-based text-to-video generation. In IJCNN, 2023. 6", + "[58] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 2", + "[59] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3", + "[60] Pauline Luc, Aidan Clark, Sander Dieleman, Diego de Las Casas, Yotam Doron, Albin Cassirer, and Karen Simonyan. Transformation-based adversarial video prediction on large-scale data. arXiv preprint arXiv:2003.04035, 2020. 2", + "[61] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 6", + "[62] Tanya Marwah, Gaurav Mittal, and Vineeth N Balasubramanian. Attentive semantic video generation using captions. In ICCV, 2017. 2", + "[63] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. ICLR, 2022. 3", + "[64] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync-draw: Automatic video generation using" + ], + "bbox": [ + 509, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7836", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep recurrent attentive architectures. In ACM Multimedia, 2017. 2", + "[65] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4", + "[66] Andres Munoz, Mohammadreza Zolfaghari, Max Argus, and Thomas Brox. Temporal shift gan for large scale video generation. In WACV, 2021. 3", + "[67] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[68] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2", + "[69] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. NeurIPS, 2022. 2, 3, 5", + "[70] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2", + "[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. In CVPR, 2024. 2", + "[72] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7", + "[73] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 3", + "[74] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7, 8", + "[76] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 2", + "[77] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2", + "[78] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5", + "[80] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2", + "[81] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 4", + "[82] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2", + "[83] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. IJCV, 2020. 2", + "[84] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH, 2022. 2", + "[85] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 5", + "[86] Baifeng Shi, Qi Dai, Judy Hoffman, Kate Saenko, Trevor Darrell, and Huijuan Xu. Temporal action detection with multi-level supervision. In ICCV, 2021. 5", + "[87] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. 2023. 3", + "[88] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. *ICLR*, 2023. 2, 4, 5, 6, 7", + "[89] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2", + "[90] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4", + "[91] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[92] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2", + "[93] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. NeurIPS, 2021. 3", + "[94] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 2" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7837", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Shuyuan Tu, Tianzhen Guan, and Li Kuang. Multiple biological granularities network for person re-identification. In ICMR, 2022. 3", + "[96] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. In ICCV, 2023. 5", + "[97] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 3", + "[98] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3", + "[99] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7, 8", + "[100] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 2", + "[101] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Masked conditional video diffusion for prediction, generation, and interpolation. arXiv preprint arXiv:2205.09853, 2022. 2", + "[102] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 2016. 2", + "[103] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 7", + "[104] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2, 5, 6, 7", + "[105] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 6", + "[106] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, Kai Qiu, Yuhui Yuan, Chuanxin Tang, Xiaoyan Sun, Chong Luo, and Baining Guo. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2", + "[107] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2", + "[108] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 6", + "[109] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[110] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 6, 7, 8", + "[111] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In ECCV,2022.3", + "[112] Zhen Xing, Yijiang Chen, Zhixin Ling, Xiangdong Zhou, and Yu Xiang. Few-shot single-view 3d reconstruction with memory prior contrastive network. In ECCV, 2022. 3", + "[113] Zhen Xing, Hengduo Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised single-view 3d reconstruction via prototype shape priors. In ECCV, 2022. 3", + "[114] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. In CVPR, 2023. 5", + "[115] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3", + "[116] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 2", + "[117] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2", + "[118] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 6, 7", + "[119] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 7", + "[120] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295, 2023. 2", + "[121] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2", + "[122] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022. 2", + "[123] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3", + "[124] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. ICLR, 2023. 2, 3, 4", + "[125] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku," + ], + "bbox": [ + 501, + 90, + 890, + 901 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7838", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2", + "[126] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2", + "[127] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021. 3", + "[128] Bowen Zhang, Xiaojie Jin, Weibo Gong, Kai Xu, Zhao Zhang, Peng Wang, Xiaohui Shen, and Jiashi Feng. Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 2, 3, 5", + "[129] Hao Zhang, Yanbin Hao, and Chong-Wah Ngo. Token shift transformer for video classification. In ACM Multimedia, 2021. 3", + "[130] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 2", + "[131] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4", + "[132] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2", + "[133] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 3", + "[134] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. arXiv preprint arXiv:2305.17431, 2023. 3", + "[135] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In CVPR, 2024. 2", + "[136] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 6", + "[137] Rui Zhu, Yingwei Pan, Yehao Li, Ting Yao, Zhenglong Sun, Tao Mei, and Chang Wen Chen. Sd-dit: Unleashing the power of self-supervised discrimination in diffusion transformer. In CVPR, 2024. 2" + ], + "bbox": [ + 78, + 90, + 467, + 784 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "7839", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_model.json b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d743b0f4d95d47b80253e4bacf8f8b829930e39 --- /dev/null +++ b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_model.json @@ -0,0 +1,3603 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "SimDA: Simple Diffusion Adapter for Efficient Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.18, + 0.293, + 0.199 + ], + "angle": 0, + "content": "Zhen Xing1,2" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.181, + 0.38, + 0.198 + ], + "angle": 0, + "content": "Qi Dai3" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.181, + 0.477, + 0.198 + ], + "angle": 0, + "content": "Han Hu3" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.181, + 0.619, + 0.198 + ], + "angle": 0, + "content": "Zuxuan \\(\\mathbf{W}\\mathbf{u}^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.181, + 0.782, + 0.199 + ], + "angle": 0, + "content": "Yu-Gang Jiang1,2" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.201, + 0.802, + 0.219 + ], + "angle": 0, + "content": "1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.219, + 0.785, + 0.237 + ], + "angle": 0, + "content": "2 Shanghai Collaborative Innovation Center of Intelligent Visual Computing" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.236, + 0.59, + 0.252 + ], + "angle": 0, + "content": "3 Microsoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.255, + 0.679, + 0.271 + ], + "angle": 0, + "content": "https://chenhsing.github.io/SimDA" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.304, + 0.211, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.37, + 0.46, + 0.383 + ], + "angle": 0, + "content": "A horse galloping through van Gogh's 'Starry Night'." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.305, + 0.298, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.369, + 0.46, + 0.383 + ], + "angle": 0, + "content": "oping through van Gogh's 'Starry Night'." + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.305, + 0.385, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.369, + 0.469, + 0.383 + ], + "angle": 0, + "content": "van Gogh's 'Starry Night'." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.39, + 0.211, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.455, + 0.449, + 0.466 + ], + "angle": 0, + "content": "Fireworks being displayed for a crowd of people." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.39, + 0.298, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.455, + 0.472, + 0.466 + ], + "angle": 0, + "content": "being displayed for a crowd of people." + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.471, + 0.212, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.537, + 0.457, + 0.551 + ], + "angle": 0, + "content": "A dog wearing virtual reality goggles on the grass." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.471, + 0.3, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.471, + 0.386, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.471, + 0.472, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.13, + 0.555, + 0.213, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.62, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Time lapse at a fantasy landscape, 4k, high resolution." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.555, + 0.3, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.62, + 0.47, + 0.633 + ], + "angle": 0, + "content": "a fantasy landscape, 4k, high resolution." + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.644, + 0.371, + 0.657 + ], + "angle": 0, + "content": "(a) Text-to-Video Generation" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.304, + 0.579, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.371, + 0.773, + 0.383 + ], + "angle": 0, + "content": "A jeep car is moving on the road." + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.305, + 0.666, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.304, + 0.753, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.304, + 0.84, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.389, + 0.58, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.389, + 0.666, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.389, + 0.753, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.389, + 0.84, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.455, + 0.786, + 0.467 + ], + "angle": 0, + "content": "An AE86 car is moving on the road." + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.471, + 0.579, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.471, + 0.666, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.471, + 0.753, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.471, + 0.84, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.537, + 0.83, + 0.55 + ], + "angle": 0, + "content": "A jeep car is moving on the road, cartoon style." + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.555, + 0.576, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.62, + 0.812, + 0.633 + ], + "angle": 0, + "content": "Keep car is moving in a forest, in Autumn." + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.644, + 0.753, + 0.658 + ], + "angle": 0, + "content": "(b) Text-guided Video Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.664, + 0.894, + 0.692 + ], + "angle": 0, + "content": "Figure 1. Examples of our SimDA results: (a) The results of open-wild Text-to-Video Generation. (b) Text-guided video Editing results using one text-video pair tuning." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.701, + 0.314, + 0.717 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.735, + 0.474, + 0.903 + ], + "angle": 0, + "content": "The recent wave of AI-generated content has witnessed the great development and success of Text-to-Image (T2I) technologies. By contrast, Text-to-Video (T2V) still falls short of expectations though attracting increasing interest. Existing works either train from scratch or adapt large T2I model to videos, both of which are computation and resource expensive. In this work, we propose a Simple Diffusion Adapter (SimDA) that fine-tunes only 24M out of 1.1B parameters of a strong T2I model, adapting it to video generation in a parameter-efficient way. In particular, we turn the T2I model for T2V by designing light-weight spa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.854 + ], + "angle": 0, + "content": "tial and temporal adapters for transfer learning. Besides, we change the original spatial attention to the proposed Latent-Shift Attention (LSA) for temporal consistency. With a similar model architecture, we further train a video superresolution model to generate high-definition \\((1024\\times 1024)\\) videos. In addition to T2V generation in the wild, SimDA could also be utilized in one-shot video editing with only 2 minutes tuning. Doing so, our method could minimize the training effort with extremely few tunable parameters for model adaptation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.526, + 0.887, + 0.656, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7827" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.468, + 0.341 + ], + "angle": 0, + "content": "Image generation stands on top of the recent AIGC wave. It not only has a significant impact on the academic community but also achieves tremendous success in various applications, such as computer graphics, art and culture, medical imaging, etc. The approaches in this area mainly include methods based on generative adversarial networks (GANs) [29, 49-51, 84], auto-regressive transformers [20, 77, 125], and the latest diffusion models [17, 23, 38, 40, 58, 67, 68, 78, 79, 82, 91, 120]. Among them, diffusion models are the most popular owing to their strong controllability, simple stability, and amazing realism. However, video generation research lags behind due to challenges like the scarcity of publicly available datasets, difficulty in modeling temporal information, and high training costs, hindering progress in this area." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.468, + 0.477 + ], + "angle": 0, + "content": "There have been several research endeavors dedicated to exploring video synthesis [3, 9, 12, 15, 24, 27, 30, 31, 42, 48, 52, 55, 60, 62, 64, 70, 83, 89, 94, 100, 102, 107, 108, 121, 126]. In addition, some studies have employed popular diffusion models for video generation [32, 41, 43, 61, 101, 106, 116, 122, 135, 137]. However, most of them involve training models from scratch, which can be time-consuming due to the complex video data. Early attempts were also constrained by GPU memory or hardware limitations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.478, + 0.468, + 0.628 + ], + "angle": 0, + "content": "More recently, a small number of T2V (Text-to-Video) approaches have emerged, aiming to fine-tune well-established T2I (Text-to-Image) models [1, 8, 28, 39, 61, 104, 136]. They have incorporated temporal modeling modules (e.g.Imagen video [39], Video LDM [8]) into T2I models, which effectively accelerate the model convergence. However, it should be noted that training such models is still a challenging task due to the massive number of parameters (4B or even 16B) involved in the network architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.63, + 0.468, + 0.794 + ], + "angle": 0, + "content": "In the NLP field, state-of-the-art results of various tasks are generally achieved by adaptation from large pretrained models (i.e., BERT [16], LLMs [14, 74, 76, 132]). However, with the advent of increasingly larger and more powerful foundation models (e.g., GPT-4 with 100T parameters), conducting full fine-tuning of the entire models has become prohibitively expensive and infeasible in terms of training cost and GPU storage. To address the issue, numerous methods based on efficient fine-tuning have emerged rapidly in NLP [44, 45, 53, 54] and computer vision [13, 69, 124, 128]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a parameter-efficient video diffusion model, namely Simple Diffusion Adapter (SimDA), that fine-tunes the large T2I (i.e. Stable Diffusion [79]) model for improved video generation. We only add \\(2\\%\\) parameters compared to the T2I model. During training, we freeze the original T2I model, and only tune the newly added modules. We further propose a Latent-Shift Attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.317 + ], + "angle": 0, + "content": "(LSA) to replace the original spatial attention, which significantly improves the temporal modeling capability and retains consistency without adding new parameters. To this end, our model demands less than 8GB GPU memory for training with a resolution of \\(16 \\times 256 \\times 256\\), while the inference time speeds up by \\(39 \\times\\) compared to the autoregressive method CogVideo [42]. Besides, we turn an image super resolution framework into the video counterpart with similar architecture, which allows generating high-definition videos of \\(1024 \\times 1024\\). Our model can also be extended to the recently popular diffusion-based video editing [110], achieving significant \\(3 \\times\\) faster training while retaining comparable results, as evidenced by the editing examples presented in Fig 1 (b). In conclusion, the contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.318, + 0.891, + 0.363 + ], + "angle": 0, + "content": "- We explore the simple adaptation from image diffusion to video diffusion, exhibiting that tuning extremely few parameters can achieve surprisingly good results." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.891, + 0.408 + ], + "angle": 0, + "content": "- With the helpful light-weight adapters and the proposed latent-shift attention, our method can effectively model the temporal relations with negligible cost." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.409, + 0.891, + 0.453 + ], + "angle": 0, + "content": "- Our diffusion adapter could be extended to text-guided video super-resolution and video editing, significantly facilitating the model training." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.455, + 0.891, + 0.499 + ], + "angle": 0, + "content": "- SimDA significantly alleviates the training cost and speeds up the inference time, while remaining competitive results compared to other methods." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.318, + 0.891, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.64, + 0.528 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Text-to-Video Generation Similar to the advancements in Text-to-Image (T2I) generation [18, 71, 79, 130], early approaches for Text-to-Video (T2V) generation [55, 64, 70] were based on Generative Adversarial Networks (GANs) and primarily applied to domain-specific videos such as simple human actions [92] or clouds moving [117]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Recently, T2V methods are most based on fine tuning the T2I [78, 79] diffusion models. For instance, Make-A-Video [88] proposes fine-tuning a pretrained DALLE2 [78] model solely on video data to learn motion patterns, enabling T2V generation without explicitly training on text-video pairs. Video Diffusion Models [41] and Imagen Video [39] perform joint text-image and text-video training, treating images as independent frames and disabling temporal layers in the U-Net [80] architecture. Besides, Video LDM [8], Latent-Shift [1], VideoFactory [104], MagicVideo [136] and our methods utilize the popular open-sourced T2I Stable Diffusion [79] model. While the progress in video generation is impressive, the parameters of video generation can be highly large. As shown in Table 1, Make-A-Video [88] requires six models and 9.7B parameters and Imagen Video [39] utilizes eight models with 16.3B parameters, which limits the training efficiency of T2V models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Text guided Video Editing In the realm of content gen" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "eration, an alternative avenue is the manipulation of existing images [10, 37, 63, 98] and videos [6, 21, 59, 73, 87, 110, 115, 123, 134] using textual input as a means of control, rather than relying solely on unbridled text-based generation. SDEdit [63] introduces noise to images and then reconstructs them for the purpose of editing. Prompt-to-prompt [37] and Plug-and-Play [98] modify the cross-attention map by altering the textual description, thus influencing the editing process. When it comes to video editing, Tune-A-Video [110] fine-tunes the T2I model on a single video, enabling the generation of new videos with similar motion patterns. Video-P2P [59] and FateZero [73] extend the concept of Prompt-to-prompt editing to videos. Text2Live [6] divides videos into layers and enables separate editing of each layer based on text. MotionEditor [97] edits the motion of human while keeping the background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.473, + 0.529 + ], + "angle": 0, + "content": "Parameter-Efficient Transfer Learning In NLP, parameter efficient fine-tuning [33, 44, 45, 53, 54, 93, 127] were initially proposed to address the heavy computation of full fine-tuning LLMs for various downstream tasks. These techniques aim to reduce the number of trainable parameters, thereby lowering computation costs, while still achieving or surpassing the performance of full fine-tuning. Recently, parameter-efficient transfer learning has also been explored in the field of computer vision [4, 13, 22, 26, 46, 47, 69, 95, 112, 113, 124, 128]. These methods mainly focus on adapting models within simple classification or detection tasks. In contrast, our approach focuses on adapting a T2I model for T2V generation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.818 + ], + "angle": 0, + "content": "Temporal Shift Module TSM [56] pioneered the introduction of the temporal shift module for action recognition, employing a partial channel shift along the temporal dimension. This approach seamlessly integrates temporal cues from both preceding and succeeding frames into the current frame without incurring additional computational overhead. Subsequently, TokShift [129] implemented channel shifting along the temporal dimension for transformer architectures [19, 133]. TPS [111] further shifted patches instead of channels to model the temporal correlations. However, such direct patch shifting would lead to inconsistency in generation tasks. Additionally, Latent-shift [1] and TSB [66] adapted shift module as TSM [56] within convolution blocks for video generation tasks. In this work, our latent-shift attention (LSA) employs the patch-level shifting manner. In contrast to TPS, we further propose to involve all tokens in the current frame as the keys and values, which guarantees temporal consistency during generation and significantly improves the video quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.17, + 0.847 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this section, we first introduce the preliminaries of Latent Diffusion Model [79] in Sec. 3.1. The pipeline of SimDA is described in Sec. 3.2. Then we detail the proposed spa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.138 + ], + "angle": 0, + "content": "tial and temporal adapters as well as latent-shift attention in Sec. 3.3. Finally, we introduce the super resolution and text-guided video editing model in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.147, + 0.788, + 0.162 + ], + "angle": 0, + "content": "3.1. Preliminaries of Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.895, + 0.276 + ], + "angle": 0, + "content": "In this subsection, we introduce the preliminaries of Stable Diffusion [79] model. It is a latent diffusion model that operates in the latent space of an autoencoder \\(\\mathcal{D}(\\mathcal{E}(\\cdot))\\), where \\(\\mathcal{E}\\) is the encoder and \\(\\mathcal{D}\\) is the decoder. In this model, for an image \\(I\\) with its corresponding latent feature \\(\\pmb{x}_0 = \\mathcal{E}(I)\\), the diffusion forward process involves iteratively adding noise to the latent space," + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.289, + 0.892, + 0.306 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.894, + 0.378 + ], + "angle": 0, + "content": "where \\(t \\in \\{1, \\dots, T\\}\\) is the time step, \\(q(\\pmb{x}_t | \\pmb{x}_{t-1})\\) is the conditional density of \\(\\pmb{x}_t\\) given \\(\\pmb{x}_{t-1}\\), \\(\\mathbf{I}\\) is identity matrix, and \\(\\alpha_t\\) is hyperparameter. Alternatively, we can directly sample \\(\\pmb{x}_t\\) at any time step from \\(\\pmb{x}_0\\) with," + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.389, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.417, + 0.646, + 0.435 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.57 + ], + "angle": 0, + "content": "In the diffusion backward process, a U-Net denoted as \\(\\epsilon_{\\theta}\\) is trained to predict the noise in the latent space, aiming to iteratively recover \\(x_0\\) from \\(x_{T}\\). In this process, as the diffusion progresses and approaches a large value of \\(T\\), \\(x_0\\) is completely disrupted and the latent representation \\(x_{T}\\) approximates a standard Gaussian distribution. Consequently, the U-Net \\(\\epsilon_{\\theta}\\) is trained to infer meaningful and valid \\(x_0\\) from random Gaussian noises. The training object can be simplified as," + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.581, + 0.892, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), t} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t) \\| _ {2} ^ {2} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.611, + 0.789, + 0.626 + ], + "angle": 0, + "content": "where \\( c \\) is the embedding of condition text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.748 + ], + "angle": 0, + "content": "During the inference stage, it samples a valid latent representation \\( \\mathbf{x}_0 \\) from the standard Gaussian noise \\( \\mathbf{x}_T = \\mathbf{z}_T, \\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) using DDIM [90] sampling. Then, the model can decode \\( \\mathbf{x}_0 \\) using the decoder \\( \\mathcal{D} \\) to generate the final image \\( I = \\mathcal{D}(\\mathbf{x}_0) \\). This process could generate diverse and high-quality images based on the sampled latent representations. In contrast, our method focuses on more challenging high-quality video generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.6, + 0.773 + ], + "angle": 0, + "content": "3.2. Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "SimDA, as shown in Fig. 2, is built upon the previously introduced Stable Diffusion [79]. For a video clip with \\( t \\) frames, denoted as \\( \\{I_i\\}_{i=1}^t \\), we first pass it through a pretrained encoder \\( \\mathcal{E} \\) to obtain the corresponding latent feature \\( \\{\\pmb{x}_i\\}_{i=1}^t \\). We then input the latent features to the forward diffusion process, where noise is incrementally added to the latents. In the backward diffusion process, we utilize an inflated U-Net architecture to predict the noise for the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.087, + 0.891, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.893, + 0.432 + ], + "angle": 0, + "content": "Figure 2. Pipeline of our Parameter-Efficient Text-to-Video Framework. We utilize the pre-trained auto-encoder as in Stable Diffusion [79] to obtain latent representation. During training, we only update the parameters of the newly added adapter module, highlighted in yellow. The parameters of other modules are frozen, highlighted in blue." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.445, + 0.893, + 0.472 + ], + "angle": 0, + "content": "Table 1. Model size and inference speed comparisons. The speed is measured in seconds on one A100 (80GB) GPU. The majority of results are sourced from [1]." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.475, + 0.894, + 0.603 + ], + "angle": 0, + "content": "
MethodParameters (Billion)Speed (s)
T2V CoreAuto EncoderText EncoderPrior ModelSuper ResolutionFrame InterpolationOverallTuned
CogVideo [42]7.70.10---7.715.515.5434.53
Make-A-Video [88]3.1-0.121.31.4 + 0.73.19.729.72-
Imagen Video [39]5.6-4.6-1.2 + 1.4 + 0.341.7 + 0.78 + 0.6316.2516.25-
Video LDM [8]1.510.080.12-0.981.514.202.65-
Latent-VDM [1]0.920.080.58---1.580.9228.62
Latent-Shift [1]0.880.080.58---1.530.8823.40
LVDM [35]0.960.080.12---1.161.0421.23
SimDA (Ours)0.880.080.12---1.080.02511.20
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.793 + ], + "angle": 0, + "content": "noisy video latents. Specifically, for the Convolution block, we inflate the 2D ResNet [34] block to a 3D block to accommodate video inputs. Additionally, we incorporate a lightweight Temporal Adapter module for temporal modeling. In the Attention block, we employ a latent-shift attention mechanism for spatial self-attention and introduce two spatial adapter modules to facilitate the transfer of video information. Further details are presented in Sec. 3.3. During inference, we employ DDIM [90] sampling to progressively denoise the latent representation sampled from a standard Gaussian distribution. Finally, we utilize a pre-trained decoder \\(\\mathcal{D}\\) to reconstruct the video from the denoised latent." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.188, + 0.818 + ], + "angle": 0, + "content": "3.3. Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.871 + ], + "angle": 0, + "content": "In this section, we describe the proposed Spatial Adapter, Temporal Adapter, and Latent-Shift Attention in detail, which are the key components of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Spatial Adapter The large-scale text-image pre-trained T2I model exhibits significant transferability, as evidenced by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.718 + ], + "angle": 0, + "content": "its remarkable accomplishments in tasks such as personalized T2I generation [65, 81] and image editing [37, 131]. Consequently, we believe that employing a lightweight fine-tuning approach can effectively harness spatial information in the realm of video generation. Inspired by efficient fine-tuning techniques in NLP [45, 54] and vision tasks [13, 124], we adopt adapters due to their simplicity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.893, + 0.87 + ], + "angle": 0, + "content": "In our T2V framework, we design two types of spatial adapters (i.e., Attention Adapter and FFN Adapter) for transferring video spatial information. As shown in the bottom right of Fig. 2, both adapters employ a bottleneck architecture consisting of two fully connected (FC) layers with an intermediate activation layer. The first FC layer maps the input to a lower-dimensional space, while the second FC layer maps it back to the original dimensional. Formally, for an input feature matrix \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times d}\\), the spatial adapter could be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S - A d p a t e r} (\\mathbf {X}) = \\mathbf {X} + \\mathbf {W} _ {\\text {u p}} \\left(\\operatorname {G E L U} \\left(\\mathbf {W} _ {\\text {d o w n}} (\\mathbf {X})\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7830" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.236, + 0.471, + 0.292 + ], + "angle": 0, + "content": "Figure 3. The overview of Temporal Latent-shift Attention module. It is noted that the Latent-shift attention is performed on latent space, but the visualization overview is shown on the image level for clear presentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.47, + 0.466 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathrm{up}}\\) and \\(\\mathbf{W}_{\\mathrm{down}}\\) are the learnable matrix with dimension \\(d\\times l\\) and \\(l\\times d, l < d\\). To preserve the structure of the original network and the pretrained weights, we initialize the second FC layer \\(\\mathbf{W}_{\\mathrm{down}}\\) with zeros. To adapt to the spatial features of videos, we incorporate the adapter after the latent-shift attention layer. Additionally, we observe that adding an adapter to the feed-forward network (FFN) also helps the network transfer spatial information to videos. We will provide examples in Sec. 4.4. During training, all layers of the attention block are fixed, and only the adapters are updated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.602 + ], + "angle": 0, + "content": "Temporal Adapter While the spatial adapter effectively transfers spatial information to video, modeling temporal information is crucial for T2V generation tasks. Previous approaches incorporate temporal convolution [11, 85, 86] or temporal attention [8, 88, 104] modules to capture temporal relationships. Although these modules are effective in modeling temporal dynamics, they often come with a huge number of parameters and high-dimensional input feature, resulting in significant computational and training costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.469, + 0.692 + ], + "angle": 0, + "content": "To address this issue, we utilize the temporal adapter module for temporal modeling as [69, 128]. In contrast to conventional spatial adapter modules, the temporal adapter module employs depth-wise 3D convolution instead of an intermediate activation layer [36]. The temporal adapter could be formally written as:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.704, + 0.468, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T - A d p a t e r} (\\mathbf {X}) = \\mathbf {X} + \\mathbf {W} _ {\\mathrm {u p}} (3 \\mathrm {D} - \\operatorname {C o n v} (\\mathbf {W} _ {\\mathrm {d o w n}} (\\mathbf {X}))). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.824 + ], + "angle": 0, + "content": "By utilizing 3D convolutions in lower-dimensional input, our approach significantly alleviates the complexity of temporal modeling. As a result, our method achieves efficient memory usage during training and exhibits the fastest inference speed among competitive approaches, as demonstrated in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Temporal Latent-Shift Attention In the original T2I framework, the attention block of the U-Net only performs self-attention for individual frames, neglecting the information from other frames. While joint-space-time attention, as demonstrated in [2, 7, 96, 114], can effectively model tem" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.887, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Figure 4. During inference, we sample a novel video from the latent noise inverted from the input video, guided by an edited prompt (e.g., \"a Porsche sports car is moving on the road\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.442 + ], + "angle": 0, + "content": "poral dependencies, it introduces a quadratic complexity in terms of attention calculation. For a video with \\( L \\) frames and \\( N \\) tokens, the complexity of global spatial-temporal attention becomes \\( O(L^2 N^2) \\). To address this issue, we propose a latent-shift attention module as shown in Fig. 3. In addition to considering tokens within the current frame, we further conduct a patch-level shifting operation along the temporal dimension to shift tokens from the preceding \\( T \\) frames onto the current frame, thereby composing a new latent feature frame. We concatenate the latent feature of the current frame with the temporally shifted latent feature, forming the keys and values for attention calculation. The latent-shift attention can be formally written as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.45, + 0.891, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {W} _ {\\mathrm {q}} \\left(\\boldsymbol {x} _ {z _ {i}}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.47, + 0.891, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} = \\mathbf {W} _ {\\mathrm {k}} \\left[ \\boldsymbol {x} _ {z _ {i}}, \\boldsymbol {x} _ {z _ {s h i f t}} \\right], \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.488, + 0.891, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\mathbf {W} _ {\\mathrm {v}} \\left[ \\boldsymbol {x} _ {z _ {i}}, \\boldsymbol {x} _ {z _ {\\text {s h i f t}}} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.598 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{x}_{z_i} \\) denotes the query frame and \\( [\\cdot] \\) means concatenate. This approach reduces the complexity of attention to \\( O(2LN^2) \\), significantly lowering the computational burden compared to global attention. Moreover, it allows the model to learn the relationships between adjacent frames, ensuring better temporal consistency in video generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.824, + 0.621 + ], + "angle": 0, + "content": "3.4. Super Resolution and Editing Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Super Resolution (SR) Due to constraints of limited GPU memory and the lack of high-resolution video-text datasets, most existing methods [1, 8, 35], including ours, are only able to generate images at a resolution of \\(256 \\times 256\\). To overcome this limitation and generate higher-resolution outputs, we adopt a two-stage training approach similar to cascaded Diffusion Models [40]. In the first stage, we generate videos with a \\(256 \\times 256\\) resolution using our SimDA methods. In the second stage, we employ an LDM upsampler [79] to enhance the resolution of the videos to \\(1024 \\times 1024\\). We incorporate noise augmentation and noise level conditioning, and train a super-resolution model using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.832, + 0.891, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I}), t} [ \\| \\epsilon - \\epsilon_ {\\theta} ([ \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {l o w} ], \\boldsymbol {c}, t) \\| _ {2} ^ {2} ], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( \\pmb{x}_{low} \\) is the low-resolution video, we concatenate it with \\( \\pmb{x}_t \\) frame by frame following Video LDM [8]. The architecture of SR is similar to T2V model in the first stage," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7831" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.396, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.09, + 0.621, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.09, + 0.848, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.34, + 0.722, + 0.353 + ], + "angle": 0, + "content": "Figure 5. Text-to-Video generation comparison on the user study evaluation set." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.357, + 0.894, + 0.384 + ], + "angle": 0, + "content": "Table 2. Text-to-Video generation comparison on MSR-VTT [118] dataset. We report the Fréchet Video Distance (FVD) scores and CLIPSIM scores." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.386, + 0.86, + 0.617 + ], + "angle": 0, + "content": "
MethodTraining DataResolutionZero-shotParams(B)FVD(↓)CLIPSIM(↑)
GODIVA [108]MSR-VTT128x128No--0.2402
NüWA [109]MSR-VTT128x128No0.87-0.2439
Make-A-Video [88]WebVid-10M + HD-VILA-10M256x256Yes9.72-0.3049
VideoFactory [104]WebVid-10M + HD-VG-130M256x256Yes2.04-0.3005
LVDM [35]WebVid-2M256x256Yes1.167420.2381
MMVG [25]WebVid-2.5M256x256Yes--0.2644
CogVideo [42]WebVid-5.4M256x256Yes15.512940.2631
ED-T2V [57]WebVid-10M256x256Yes1.30-0.2763
MagicVideo [136]WebVid-10M256x256Yes-998-
Video-LDM [8]WebVid-10M256x256Yes4.20-0.2929
VideoComposer [105]WebVid-10M256x256Yes1.855800.2932
Latent-Shift [1]WebVid-10M256x256Yes1.53-0.2773
VideoFusion [61]WebVid-10M256x256Yes1.835810.2795
SimDA (Ours)WebVid-10M256x256Yes1.084560.2945
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.686 + ], + "angle": 0, + "content": "we change the original U-Net block by adding Spatial and Temporal Adapters as described in Sec. 3.3 and only finetune the newly added modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Text-guided Video Editing In addition to performing T2V generation, our method could turn into one-shot tuning for text-guided video editing following Tune-A-Video [110]. The training pipeline of editing model is the same as our T2V method. However, for the inference stage, we adopt the DDIM inversion latents instead of random noisy latents together with edited prompt for novel video generation as shown in Fig. 4. By doing so, the pixel-level information control could remain in the inversion latent as demonstrated in [110]. Owing to the light-weight module and efficient pipeline of our method, SimDA needs fewer training steps (200 steps compared to 500 steps) and thus the training time and inference time is much faster than Tune-A-Video [110]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.639, + 0.634, + 0.656 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.719, + 0.681 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our T2V method is composed of two-stage models. The first model predicts video frames with a resolution \\(256 \\times 256\\) (with a latent size of \\(32 \\times 32\\)), while the second model is a \\(4 \\times\\) upsampler, producing a resolution of \\(1024 \\times 1024\\). We train the general T2V model on WebVid-10M [5] dataset following [1, 8]. We follow previous methods [1, 104, 136] to report the CLIP score [75] and FVD (Frechet Video Distance) score [99] on MSR-VTT [118]. Besides, we compare the FVD score and CLIP score on the evaluation set of WebVid [5] as in VideoFactory [104]. We also compare the parameter scale and inference speed of our method with some open-sourced methods [1, 35, 61]. Finally, we also provide a user study between our work and VDM [41], Latent-shift [1], Video-Fusion [61] and LVDM [35]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7832" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.421, + 0.108 + ], + "angle": 0, + "content": "4.2. Evaluation on Text-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.468, + 0.176 + ], + "angle": 0, + "content": "To fully evaluate the generation performance of our SimDA, we conduct automatic evaluations on two distinct datasets: WebVid [5] (Val), which shares the same domain as the training data and MSR-VTT [118] in a zero-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.178, + 0.468, + 0.344 + ], + "angle": 0, + "content": "Evaluation on MSR-VTT As shown in Table 2, we evaluate CLIPSIM [75] and FVD [99] on the widely used video generation benchmarks, MSR-VTT [118]. We randomly select one text prompt per example from MSR-VTT [118] and generate a total of 2,990 videos. Despite being a zero-shot setting, our method achieves an average CLIPSIM of 0.2945 that surpasses most of the competitors, indicating a strong semantic alignment between the generated videos and the input text. Though Make-A-Video [88] and VideoFactory [104] offer higher CLIP scores, they utilize additional large-scale HD-VILA [119] datasets for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.468, + 0.529 + ], + "angle": 0, + "content": "Evaluation on WebVid As shown in Table 3, we create a validation set consisting of 4,476 randomly extracted text-video pairs from WebVid-10M. These pairs are not included in the training data following [104]. We conduct evaluations on this validation set and obtain impressive results. Our method achieves an FVD score of 363.98 and a CLIP-SIM score of 0.3054. These scores are significantly higher than those achieved by existing methods such as ModelScope [103] and LVDM [35]. Besides, our method shows competitive results compared to VideoFactory [104] which is trained with much larger datasets. These results clearly demonstrate the superiority of our approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.468, + 0.667 + ], + "angle": 0, + "content": "Human Evaluation In order to address the limitations of existing metrics and assess the performance of our SimDA from a human perspective, we conduct an extensive user study. The study involves comparing our method with four state-of-the-art methods. Specifically, we select two publicly available models, namely ModelScope [103] and LVDM [35]. Additionally, we consider two methods with similar scale parameters, VDM [41] and Latent-shift [1], which only showcase some samples on their websites." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.468, + 0.804 + ], + "angle": 0, + "content": "For each case, participants were provided with two video samples, one is generated by our method and the other is from a competitor. They were then asked to compare the two samples in terms of video quality and text-video similarity. To ensure fairness in the comparisons, we also report the ratio of network parameters compared to ours. The results, along with the parameter ratios, are presented in Table 4. The user study approach allows us to gain in-depth insights into the subjective evaluation of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.468, + 0.867 + ], + "angle": 0, + "content": "Qualitative Results The visualization of T2V generation results are shown in Fig. 1(a). Besides, we show the comparison results in Fig. 5. More examples can be found in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Parameter Size and Inference Speed We conduct a comparison of the number of parameters and inference speed," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 3. Text-to-video generation on the validation set of WebVid [5]. We report the FVD and CLIPSIM scores." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.121, + 0.885, + 0.215 + ], + "angle": 0, + "content": "
MethodParams(B)FVD(↓)CLIPSIM(↑)
LVDM [35]1.16455.530.2751
ModelScope [103]1.83414.110.3000
VideoFactory [104]2.04292.350.3070
SimDA (Ours)1.08363.980.3054
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.224, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Table 4. User preference is depicted as a percentage indicating the proportion of individuals favoring our method over the compared approach. Param Ratio means the ratio of the network parameter v.s. Ours." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.28, + 0.905, + 0.371 + ], + "angle": 0, + "content": "
SampleMethodParam RatioQualityFaithfulness
Open WebsiteVDM [41]0.83×85.2%81.4%
Latent-Shift [1]1.41×81.5%79.3%
Pretrained ModelModelScope [103]1.69×78.3%79.5%
LVDM [35]1.07×83.4%84.7%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.518 + ], + "angle": 0, + "content": "and the results are presented in Table 1. For the speed comparison, we select CogVideo [42], Latent-Shift [1] and LVDM [35]. SimDA, on the other hand, stands out as it is significantly smaller than previous works and exhibits faster inference speed compared to other methods. Despite having fewer parameters, SimDA achieves superior performance in various benchmarks when compared to other methods. This validation further highlights our advantages in terms of model efficiency and performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.851, + 0.545 + ], + "angle": 0, + "content": "4.3. Evaluation on Text-guided Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Following the methodology of previous studies [110], we employ CLIP score [75] and a user study to evaluate the performance of different methods in terms of frame consistency and textual alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.78 + ], + "angle": 0, + "content": "First, we calculate the CLIP image embedding for all frames in the edited videos to measure frame consistency. The average cosine similarity between pairs of video frames is reported. Additionally, to assess textual faithfulness, we compute the average CLIP score between frames of the output videos and the corresponding edited prompts. A total of 15 videos from the dataset [72] were selected and edited based on object, background and style, resulting in 75 edited videos for each model. The average results, presented in Table 5, highlight our method's exceptional ability to achieve semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Secondly, we conduct a user study involving videos and text prompts. Participants were asked to vote for the edited videos that exhibited the best temporal consistency and those most accurately matched the textual description. Table 5 demonstrates that our method, SimDA, receives the highest number of votes in both aspects, indicating superior editing quality and a strong preference from users in practical scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7833" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison with evaluated baseline [110]. The \"Training\" refers to the process of optimization, and \"Memory\" refers to the GPU memory." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.122, + 0.885, + 0.193 + ], + "angle": 0, + "content": "
MethodFrame consistencyTextual alignmentRuntime [min]Memory [Gib]Params [Mb]
CLIP Score↑User Vote↑CLIP Score↑User Vote↑Training↓Inference↓Training↓Inference↓Tuned↓
Tune-A-Video [110]94.131.2%31.839.5%9.30.831.311.474.4
SimDA(Ours)94.968.8%31.960.5%2.50.428.68.824.9
" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.207, + 0.447, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.575, + 0.472, + 0.618 + ], + "angle": 0, + "content": "Figure 6. Ablation of T2V generation results. TA, SA, AA, FA and LSA refer to Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-Shift Attention." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.233, + 0.644 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.681 + ], + "angle": 0, + "content": "Here we discuss the effect of each module. We perform experiments on \\(1K\\) samples from the WebVid validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Effect of Temporal Adapter Temporal modeling is a crucial component of video generation. In our video editing task, when compared to methods that rely on temporal attention modeling like Tune-A-Video [110], we observe that our temporal adapter is more lightweight and achieves superior editing results as in Table 5. Additionally, we conduct ablation experiments, as shown in Table 6 and Fig. 6, where the lack of Temporal Adapter (TA) results in significantly higher FVD and chaotic temporal sequences in the generated videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of Spatial Adapter We also validate the effectiveness of the Spatial Adapter (SA) in transferring spatial knowledge of videos. As shown in Table 6, without the Attention Adapter (AA) and FFN Adapter (FA), the model's" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.208, + 0.895, + 0.278 + ], + "angle": 0, + "content": "Table 6. Ablation study on different modules. We report the FVD [99] and CLIPSIM [75] on \\(1K\\) samples from the validation set of WebVid-10M [5]. TA, SA, AA, FA and LSA represent Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-shift Attention, respectively." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.285, + 0.897, + 0.406 + ], + "angle": 0, + "content": "
TAAAFALSAFVD(↓)CLIPSIM(↑)
w/o TA✓✓✓1470.10.2629
w/o SA✓✓811.30.2822
w/o AA✓✓✓764.80.2851
w/o FA✓✓✓623.70.2962
w/o LSA✓✓✓618.20.3011
Ours✓✓✓✓530.20.3034
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.895, + 0.479 + ], + "angle": 0, + "content": "FVD and CLIPSIM scores for generated videos will become worse. Additionally, it can be observed from the Fig. 6 that the model exhibits misconceptions in understanding the text prompt without the spatial adapter." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.483, + 0.895, + 0.605 + ], + "angle": 0, + "content": "Effect of Latent Shift Attention To investigate the impact of Latent-shift Attention (LSA), we replace it with regular single-frame spatial attention. Besides observing a decline in FVD and text alignment CLIPSIM scores in Table 6, we also test the CLIPSIM of each frame within the same video, which decreased from 96.4 to 94.5. This demonstrates that our LSA module can effectively model the relationship of adjacent frames, leading to more consistent videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.615, + 0.62, + 0.631 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.809 + ], + "angle": 0, + "content": "In this paper, we proposed SimDA, a parameter-efficient video diffusion model for text-guided video generation and editing. With the proposed light-weight spatial and temporal adapters, our method not only transferred from powerful spatial information but also modeled temporal relationships with the least new parameters. The experimental results demonstrated that our approach has the fastest training and inference speed while maintaining competitive generation and editing results. Our work is the first parameter-efficient video diffusion method serving as an efficient T2V fine-tuning baseline and paved the way for future research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Acknowledge This work was supported by National Science and Technology Major Project (No. 2021ZD0112805) and in part by National Natural Science Foundation of China (No. 62032006 and No. 62102092.). The computations in this research were performed using the CFFFplatform of Fudan University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7834" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, JiaBin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.186, + 0.469, + 0.225 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.228, + 0.469, + 0.28 + ], + "angle": 0, + "content": "[3] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. arXiv preprint arXiv:1710.11252, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.284, + 0.469, + 0.337 + ], + "angle": 0, + "content": "[4] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.341, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.396, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.439, + 0.469, + 0.478 + ], + "angle": 0, + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.481, + 0.469, + 0.548 + ], + "angle": 0, + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.551, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.607, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.649, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.691, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[12] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[13] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[14] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[15] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.124, + 0.892, + 0.15 + ], + "angle": 0, + "content": "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.153, + 0.892, + 0.193 + ], + "angle": 0, + "content": "[18] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.197, + 0.892, + 0.266 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.269, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.312, + 0.892, + 0.368 + ], + "angle": 0, + "content": "[21] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.37, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[22] Qijun Feng, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Fdgaussian: Fast gaussian splatting from single image via geometric-aware diffusion model. arXiv preprint arXiv:2403.10242, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[23] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.499, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[24] Jean-Yves Franceschi, Edouard Delasalles, Mickael Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.543, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[25] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.615, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[26] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.673, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[27] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.731, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[28] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[29] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Sonam Gupta, Arti Keshari, and Sukhendu Das. Rvgan: Recurrent gan for unconditional video generation. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7835" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[31] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.471, + 0.188 + ], + "angle": 0, + "content": "[32] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[33] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. arXiv preprint arXiv:2110.04366, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[35] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[36] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.471, + 0.425 + ], + "angle": 0, + "content": "[37] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. *ICLR*, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.427, + 0.471, + 0.454 + ], + "angle": 0, + "content": "[38] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[39] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[40] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.581, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[41] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.636, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[42] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.692, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[43] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.748, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[44] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.804, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[45] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[46] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[47] Shibo Jie and Zhi-Hong Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[48] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.178, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[49] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[50] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.263, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[51] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.306, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[52] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.362, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[53] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[54] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.448, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[55] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.49, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[56] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[57] Jiawei Liu, Weining Wang, Wei Liu, Qian He, and Jing Liu. Ed-t2v: An efficient training framework for diffusion-based text-to-video generation. In IJCNN, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[58] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[59] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.647, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[60] Pauline Luc, Aidan Clark, Sander Dieleman, Diego de Las Casas, Yotam Doron, Albin Cassirer, and Karen Simonyan. Transformation-based adversarial video prediction on large-scale data. arXiv preprint arXiv:2003.04035, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.717, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[61] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[62] Tanya Marwah, Gaurav Mittal, and Vineeth N Balasubramanian. Attentive semantic video generation using captions. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.816, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[63] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[64] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync-draw: Automatic video generation using" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7836" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "deep recurrent attentive architectures. In ACM Multimedia, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[65] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.192, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[66] Andres Munoz, Mohammadreza Zolfaghari, Max Argus, and Thomas Brox. Temporal shift gan for large scale video generation. In WACV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.236, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[67] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.305, + 0.471, + 0.334 + ], + "angle": 0, + "content": "[68] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[69] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. NeurIPS, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.391, + 0.471, + 0.433 + ], + "angle": 0, + "content": "[70] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.435, + 0.471, + 0.476 + ], + "angle": 0, + "content": "[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.478, + 0.471, + 0.533 + ], + "angle": 0, + "content": "[72] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.534, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[73] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.591, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[74] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.647, + 0.471, + 0.716 + ], + "angle": 0, + "content": "[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.718, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[76] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.789, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[77] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[78] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[80] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[81] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[82] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[83] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. IJCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.376, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[84] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[85] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[86] Baifeng Shi, Qi Dai, Judy Hoffman, Kate Saenko, Trevor Darrell, and Huijuan Xu. Temporal action detection with multi-level supervision. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.504, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[87] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[88] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. *ICLR*, 2023. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.604, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[89] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.66, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[90] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[91] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[92] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.817, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[93] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[94] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7837" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[95] Shuyuan Tu, Tianzhen Guan, and Li Kuang. Multiple biological granularities network for person re-identification. In ICMR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[96] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.177, + 0.468, + 0.23 + ], + "angle": 0, + "content": "[97] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.468, + 0.273 + ], + "angle": 0, + "content": "[98] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[99] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[100] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[101] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Masked conditional video diffusion for prediction, generation, and interpolation. arXiv preprint arXiv:2205.09853, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[102] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[103] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[104] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[105] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[106] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, Kai Qiu, Yuhui Yuan, Chuanxin Tang, Xiaoyan Sun, Chong Luo, and Baining Guo. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[107] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[108] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[109] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[110] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[111] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In ECCV,2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[112] Zhen Xing, Yijiang Chen, Zhixin Ling, Xiangdong Zhou, and Yu Xiang. Few-shot single-view 3d reconstruction with memory prior contrastive network. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[113] Zhen Xing, Hengduo Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised single-view 3d reconstruction via prototype shape priors. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[114] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.352, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[115] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[116] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.468, + 0.892, + 0.521 + ], + "angle": 0, + "content": "[117] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[118] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[119] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[120] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.684, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[121] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[122] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.826 + ], + "angle": 0, + "content": "[123] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[124] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. ICLR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[125] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7838" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[126] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[127] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.304 + ], + "angle": 0, + "content": "[128] Bowen Zhang, Xiaojie Jin, Weibo Gong, Kai Xu, Zhao Zhang, Peng Wang, Xiaohui Shen, and Jiashi Feng. Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[129] Hao Zhang, Yanbin Hao, and Chong-Wah Ngo. Token shift transformer for video classification. In ACM Multimedia, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[130] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[131] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.468, + 0.501 + ], + "angle": 0, + "content": "[132] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.468, + 0.558 + ], + "angle": 0, + "content": "[133] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.561, + 0.468, + 0.615 + ], + "angle": 0, + "content": "[134] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. arXiv preprint arXiv:2305.17431, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.617, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[135] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.674, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[136] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[137] Rui Zhu, Yingwei Pan, Yehao Li, Ting Yao, Zhenglong Sun, Tao Mei, and Chang Wen Chen. Sd-dit: Unleashing the power of self-supervised discrimination in diffusion transformer. In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7839" + } + ] +] \ No newline at end of file diff --git a/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_origin.pdf b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b48c43c7a8656f8c1a8d9306efaed7bf9a6af017 --- /dev/null +++ b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/2fcde3ae-cff4-454a-8644-ea50caebb3c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b231dc56fcb918de532e2eee3eea55b02f12bb726e5bb527b6c66b25363ec316 +size 10123708 diff --git a/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/full.md b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5565cbf49c165f4207b1987b694880f902d7e81d --- /dev/null +++ b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/full.md @@ -0,0 +1,462 @@ +# SimDA: Simple Diffusion Adapter for Efficient Video Generation + +Zhen Xing1,2 + +Qi Dai3 + +Han Hu3 + +Zuxuan $\mathbf{W}\mathbf{u}^{1,2\dagger}$ + +Yu-Gang Jiang1,2 + +1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University + +2 Shanghai Collaborative Innovation Center of Intelligent Visual Computing + +3 Microsoft Research Asia + +https://chenhsing.github.io/SimDA + +![](images/9efabbe8c1a1418a10d2477a63515b1f16e8973f54cf5bdd450b7cf92f3b4974.jpg) +A horse galloping through van Gogh's 'Starry Night'. + +![](images/54c5a14d39d31c8c500088b7d2158eb63e88d6297bed8daff155c8efc050de3f.jpg) +oping through van Gogh's 'Starry Night'. + +![](images/51af8c497a376e1777b5bd3aed89a8c27b33625e91aa7dd3c924f8f875d44881.jpg) +van Gogh's 'Starry Night'. + +![](images/1c84413c441e44bab9c4cecc026c5b864f8da99d6e2ad6cd749926d0354d9419.jpg) +Fireworks being displayed for a crowd of people. + +![](images/497d9f3f758f7784f7197e9a9b2b4e023c0c6122604da8da70f4e9cd76e4d851.jpg) +being displayed for a crowd of people. + +![](images/3a71a8d13641763cbfb4f9ea0c5d59e96d814421a6d1b59b26443a384f84cb45.jpg) +A dog wearing virtual reality goggles on the grass. + +![](images/1ec2deaf54f4f75a87fd1dcca5989858860949ebad426d8f921202a9f9df494d.jpg) + +![](images/81e6591cb1bcf9611d73738d3441118fddd7180bd0fee80e070b0ffaadd37fac.jpg) + +![](images/ed2f153bbf69a6e5061b9bdadd40a643807d6b92eeec96add1c443bb8291b524.jpg) + +![](images/2eee7dfcfd07a85522d6080da5f5b32d23af35741f91587126d25457c5bd8990.jpg) + +![](images/0eed4c60c3aa769d57c76e794b23caf49aae0f1367931cd7ac55d9d38769b7a0.jpg) +Time lapse at a fantasy landscape, 4k, high resolution. +(a) Text-to-Video Generation +Figure 1. Examples of our SimDA results: (a) The results of open-wild Text-to-Video Generation. (b) Text-guided video Editing results using one text-video pair tuning. + +a fantasy landscape, 4k, high resolution. + +![](images/0d7f70d1a3f76912b31a1d576a6092f1ca3ff2690b3f87936a4ad8737e16d5d6.jpg) +A jeep car is moving on the road. + +![](images/291b22a2c11e7846b00f67c115996cc826d74b9ea9d7ca57455c62d949679b60.jpg) + +![](images/5e3acac173ff9afee5172d47a6479db6709d6de6496690e5b258cb74ad195f78.jpg) + +![](images/87b4e1eb9dc8a50c6d93a08dbfe8c3ff156ca310a94347f0626b10f4e7815e45.jpg) + +![](images/1b79fc1e4952cd6408806056f28c03154bc627a890ad8d90b7ee6cfa4b1343ae.jpg) +An AE86 car is moving on the road. + +![](images/4aabf6227c40db2689d68ceee1620c306d33587ecb03cf0b622c3494052043a2.jpg) + +![](images/914252461b512bbb7c69bf1dbb979ee32eebebf9b49eb2b96f1737ab39b970e1.jpg) + +![](images/0e08fe1255e39a05305ad73d066cbb8f8a2e58a28ac39997077803791be343a2.jpg) + +![](images/a1c4c33c835723bd0f72bb7a0ef3ebeaa4d8187b4607467eb3ddfae07c577294.jpg) +A jeep car is moving on the road, cartoon style. + +![](images/d842fdbf571bfd8b4a42392bfe3eb8d54702bf51774490192f1bf6172cfb7b68.jpg) + +![](images/0dc5e77d4502d5f352f1dd8638b66163cf74f4c3546e1239395f01bfb9d569b6.jpg) + +![](images/317364731f1dec199aac8890ff178e90ad9506449ebbba50c1bffa99aed63e05.jpg) + +![](images/de8c7e2cb38fb0d6d1b91210d4337244cb11def8557c33a4580a93e8b3b62463.jpg) +Keep car is moving in a forest, in Autumn. +(b) Text-guided Video Editing + +# Abstract + +The recent wave of AI-generated content has witnessed the great development and success of Text-to-Image (T2I) technologies. By contrast, Text-to-Video (T2V) still falls short of expectations though attracting increasing interest. Existing works either train from scratch or adapt large T2I model to videos, both of which are computation and resource expensive. In this work, we propose a Simple Diffusion Adapter (SimDA) that fine-tunes only 24M out of 1.1B parameters of a strong T2I model, adapting it to video generation in a parameter-efficient way. In particular, we turn the T2I model for T2V by designing light-weight spa + +tial and temporal adapters for transfer learning. Besides, we change the original spatial attention to the proposed Latent-Shift Attention (LSA) for temporal consistency. With a similar model architecture, we further train a video superresolution model to generate high-definition $(1024\times 1024)$ videos. In addition to T2V generation in the wild, SimDA could also be utilized in one-shot video editing with only 2 minutes tuning. Doing so, our method could minimize the training effort with extremely few tunable parameters for model adaptation. + +# 1. Introduction + +Image generation stands on top of the recent AIGC wave. It not only has a significant impact on the academic community but also achieves tremendous success in various applications, such as computer graphics, art and culture, medical imaging, etc. The approaches in this area mainly include methods based on generative adversarial networks (GANs) [29, 49-51, 84], auto-regressive transformers [20, 77, 125], and the latest diffusion models [17, 23, 38, 40, 58, 67, 68, 78, 79, 82, 91, 120]. Among them, diffusion models are the most popular owing to their strong controllability, simple stability, and amazing realism. However, video generation research lags behind due to challenges like the scarcity of publicly available datasets, difficulty in modeling temporal information, and high training costs, hindering progress in this area. + +There have been several research endeavors dedicated to exploring video synthesis [3, 9, 12, 15, 24, 27, 30, 31, 42, 48, 52, 55, 60, 62, 64, 70, 83, 89, 94, 100, 102, 107, 108, 121, 126]. In addition, some studies have employed popular diffusion models for video generation [32, 41, 43, 61, 101, 106, 116, 122, 135, 137]. However, most of them involve training models from scratch, which can be time-consuming due to the complex video data. Early attempts were also constrained by GPU memory or hardware limitations. + +More recently, a small number of T2V (Text-to-Video) approaches have emerged, aiming to fine-tune well-established T2I (Text-to-Image) models [1, 8, 28, 39, 61, 104, 136]. They have incorporated temporal modeling modules (e.g.Imagen video [39], Video LDM [8]) into T2I models, which effectively accelerate the model convergence. However, it should be noted that training such models is still a challenging task due to the massive number of parameters (4B or even 16B) involved in the network architecture. + +In the NLP field, state-of-the-art results of various tasks are generally achieved by adaptation from large pretrained models (i.e., BERT [16], LLMs [14, 74, 76, 132]). However, with the advent of increasingly larger and more powerful foundation models (e.g., GPT-4 with 100T parameters), conducting full fine-tuning of the entire models has become prohibitively expensive and infeasible in terms of training cost and GPU storage. To address the issue, numerous methods based on efficient fine-tuning have emerged rapidly in NLP [44, 45, 53, 54] and computer vision [13, 69, 124, 128]. + +In this work, we propose a parameter-efficient video diffusion model, namely Simple Diffusion Adapter (SimDA), that fine-tunes the large T2I (i.e. Stable Diffusion [79]) model for improved video generation. We only add $2\%$ parameters compared to the T2I model. During training, we freeze the original T2I model, and only tune the newly added modules. We further propose a Latent-Shift Attention + +(LSA) to replace the original spatial attention, which significantly improves the temporal modeling capability and retains consistency without adding new parameters. To this end, our model demands less than 8GB GPU memory for training with a resolution of $16 \times 256 \times 256$ , while the inference time speeds up by $39 \times$ compared to the autoregressive method CogVideo [42]. Besides, we turn an image super resolution framework into the video counterpart with similar architecture, which allows generating high-definition videos of $1024 \times 1024$ . Our model can also be extended to the recently popular diffusion-based video editing [110], achieving significant $3 \times$ faster training while retaining comparable results, as evidenced by the editing examples presented in Fig 1 (b). In conclusion, the contributions of this work can be summarized as follows: + +- We explore the simple adaptation from image diffusion to video diffusion, exhibiting that tuning extremely few parameters can achieve surprisingly good results. +- With the helpful light-weight adapters and the proposed latent-shift attention, our method can effectively model the temporal relations with negligible cost. +- Our diffusion adapter could be extended to text-guided video super-resolution and video editing, significantly facilitating the model training. +- SimDA significantly alleviates the training cost and speeds up the inference time, while remaining competitive results compared to other methods. + +# 2. Related Work + +Text-to-Video Generation Similar to the advancements in Text-to-Image (T2I) generation [18, 71, 79, 130], early approaches for Text-to-Video (T2V) generation [55, 64, 70] were based on Generative Adversarial Networks (GANs) and primarily applied to domain-specific videos such as simple human actions [92] or clouds moving [117]. + +Recently, T2V methods are most based on fine tuning the T2I [78, 79] diffusion models. For instance, Make-A-Video [88] proposes fine-tuning a pretrained DALLE2 [78] model solely on video data to learn motion patterns, enabling T2V generation without explicitly training on text-video pairs. Video Diffusion Models [41] and Imagen Video [39] perform joint text-image and text-video training, treating images as independent frames and disabling temporal layers in the U-Net [80] architecture. Besides, Video LDM [8], Latent-Shift [1], VideoFactory [104], MagicVideo [136] and our methods utilize the popular open-sourced T2I Stable Diffusion [79] model. While the progress in video generation is impressive, the parameters of video generation can be highly large. As shown in Table 1, Make-A-Video [88] requires six models and 9.7B parameters and Imagen Video [39] utilizes eight models with 16.3B parameters, which limits the training efficiency of T2V models. + +Text guided Video Editing In the realm of content gen + +eration, an alternative avenue is the manipulation of existing images [10, 37, 63, 98] and videos [6, 21, 59, 73, 87, 110, 115, 123, 134] using textual input as a means of control, rather than relying solely on unbridled text-based generation. SDEdit [63] introduces noise to images and then reconstructs them for the purpose of editing. Prompt-to-prompt [37] and Plug-and-Play [98] modify the cross-attention map by altering the textual description, thus influencing the editing process. When it comes to video editing, Tune-A-Video [110] fine-tunes the T2I model on a single video, enabling the generation of new videos with similar motion patterns. Video-P2P [59] and FateZero [73] extend the concept of Prompt-to-prompt editing to videos. Text2Live [6] divides videos into layers and enables separate editing of each layer based on text. MotionEditor [97] edits the motion of human while keeping the background. + +Parameter-Efficient Transfer Learning In NLP, parameter efficient fine-tuning [33, 44, 45, 53, 54, 93, 127] were initially proposed to address the heavy computation of full fine-tuning LLMs for various downstream tasks. These techniques aim to reduce the number of trainable parameters, thereby lowering computation costs, while still achieving or surpassing the performance of full fine-tuning. Recently, parameter-efficient transfer learning has also been explored in the field of computer vision [4, 13, 22, 26, 46, 47, 69, 95, 112, 113, 124, 128]. These methods mainly focus on adapting models within simple classification or detection tasks. In contrast, our approach focuses on adapting a T2I model for T2V generation task. + +Temporal Shift Module TSM [56] pioneered the introduction of the temporal shift module for action recognition, employing a partial channel shift along the temporal dimension. This approach seamlessly integrates temporal cues from both preceding and succeeding frames into the current frame without incurring additional computational overhead. Subsequently, TokShift [129] implemented channel shifting along the temporal dimension for transformer architectures [19, 133]. TPS [111] further shifted patches instead of channels to model the temporal correlations. However, such direct patch shifting would lead to inconsistency in generation tasks. Additionally, Latent-shift [1] and TSB [66] adapted shift module as TSM [56] within convolution blocks for video generation tasks. In this work, our latent-shift attention (LSA) employs the patch-level shifting manner. In contrast to TPS, we further propose to involve all tokens in the current frame as the keys and values, which guarantees temporal consistency during generation and significantly improves the video quality. + +# 3. Method + +In this section, we first introduce the preliminaries of Latent Diffusion Model [79] in Sec. 3.1. The pipeline of SimDA is described in Sec. 3.2. Then we detail the proposed spa + +tial and temporal adapters as well as latent-shift attention in Sec. 3.3. Finally, we introduce the super resolution and text-guided video editing model in Sec. 3.4. + +# 3.1. Preliminaries of Stable Diffusion + +In this subsection, we introduce the preliminaries of Stable Diffusion [79] model. It is a latent diffusion model that operates in the latent space of an autoencoder $\mathcal{D}(\mathcal{E}(\cdot))$ , where $\mathcal{E}$ is the encoder and $\mathcal{D}$ is the decoder. In this model, for an image $I$ with its corresponding latent feature $\pmb{x}_0 = \mathcal{E}(I)$ , the diffusion forward process involves iteratively adding noise to the latent space, + +$$ +q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t}; \sqrt {\alpha_ {t}} \boldsymbol {x} _ {t - 1}, (1 - \alpha_ {t}) \mathbf {I}\right), \tag {1} +$$ + +where $t \in \{1, \dots, T\}$ is the time step, $q(\pmb{x}_t | \pmb{x}_{t-1})$ is the conditional density of $\pmb{x}_t$ given $\pmb{x}_{t-1}$ , $\mathbf{I}$ is identity matrix, and $\alpha_t$ is hyperparameter. Alternatively, we can directly sample $\pmb{x}_t$ at any time step from $\pmb{x}_0$ with, + +$$ +q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \tag {2} +$$ + +where $\bar{\alpha}_t = \prod_{i=1}^t \alpha_i$ . + +In the diffusion backward process, a U-Net denoted as $\epsilon_{\theta}$ is trained to predict the noise in the latent space, aiming to iteratively recover $x_0$ from $x_{T}$ . In this process, as the diffusion progresses and approaches a large value of $T$ , $x_0$ is completely disrupted and the latent representation $x_{T}$ approximates a standard Gaussian distribution. Consequently, the U-Net $\epsilon_{\theta}$ is trained to infer meaningful and valid $x_0$ from random Gaussian noises. The training object can be simplified as, + +$$ +\mathbb {E} _ {\boldsymbol {x}, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), t} [ \| \epsilon - \epsilon_ {\theta} (\boldsymbol {x} _ {t}, \boldsymbol {c}, t) \| _ {2} ^ {2} ], \tag {3} +$$ + +where $c$ is the embedding of condition text. + +During the inference stage, it samples a valid latent representation $\mathbf{x}_0$ from the standard Gaussian noise $\mathbf{x}_T = \mathbf{z}_T, \mathbf{z}_T \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ using DDIM [90] sampling. Then, the model can decode $\mathbf{x}_0$ using the decoder $\mathcal{D}$ to generate the final image $I = \mathcal{D}(\mathbf{x}_0)$ . This process could generate diverse and high-quality images based on the sampled latent representations. In contrast, our method focuses on more challenging high-quality video generation. + +# 3.2. Pipeline + +SimDA, as shown in Fig. 2, is built upon the previously introduced Stable Diffusion [79]. For a video clip with $t$ frames, denoted as $\{I_i\}_{i=1}^t$ , we first pass it through a pretrained encoder $\mathcal{E}$ to obtain the corresponding latent feature $\{\pmb{x}_i\}_{i=1}^t$ . We then input the latent features to the forward diffusion process, where noise is incrementally added to the latents. In the backward diffusion process, we utilize an inflated U-Net architecture to predict the noise for the + +![](images/0a79880bfbce85f15b13557ea3e5988d6d7172f34fb2a5e913e8cce5d29dfdcf.jpg) +Figure 2. Pipeline of our Parameter-Efficient Text-to-Video Framework. We utilize the pre-trained auto-encoder as in Stable Diffusion [79] to obtain latent representation. During training, we only update the parameters of the newly added adapter module, highlighted in yellow. The parameters of other modules are frozen, highlighted in blue. + +Table 1. Model size and inference speed comparisons. The speed is measured in seconds on one A100 (80GB) GPU. The majority of results are sourced from [1]. + +
MethodParameters (Billion)Speed (s)
T2V CoreAuto EncoderText EncoderPrior ModelSuper ResolutionFrame InterpolationOverallTuned
CogVideo [42]7.70.10---7.715.515.5434.53
Make-A-Video [88]3.1-0.121.31.4 + 0.73.19.729.72-
Imagen Video [39]5.6-4.6-1.2 + 1.4 + 0.341.7 + 0.78 + 0.6316.2516.25-
Video LDM [8]1.510.080.12-0.981.514.202.65-
Latent-VDM [1]0.920.080.58---1.580.9228.62
Latent-Shift [1]0.880.080.58---1.530.8823.40
LVDM [35]0.960.080.12---1.161.0421.23
SimDA (Ours)0.880.080.12---1.080.02511.20
+ +noisy video latents. Specifically, for the Convolution block, we inflate the 2D ResNet [34] block to a 3D block to accommodate video inputs. Additionally, we incorporate a lightweight Temporal Adapter module for temporal modeling. In the Attention block, we employ a latent-shift attention mechanism for spatial self-attention and introduce two spatial adapter modules to facilitate the transfer of video information. Further details are presented in Sec. 3.3. During inference, we employ DDIM [90] sampling to progressively denoise the latent representation sampled from a standard Gaussian distribution. Finally, we utilize a pre-trained decoder $\mathcal{D}$ to reconstruct the video from the denoised latent. + +# 3.3. Modeling + +In this section, we describe the proposed Spatial Adapter, Temporal Adapter, and Latent-Shift Attention in detail, which are the key components of our model. + +Spatial Adapter The large-scale text-image pre-trained T2I model exhibits significant transferability, as evidenced by + +its remarkable accomplishments in tasks such as personalized T2I generation [65, 81] and image editing [37, 131]. Consequently, we believe that employing a lightweight fine-tuning approach can effectively harness spatial information in the realm of video generation. Inspired by efficient fine-tuning techniques in NLP [45, 54] and vision tasks [13, 124], we adopt adapters due to their simplicity. + +In our T2V framework, we design two types of spatial adapters (i.e., Attention Adapter and FFN Adapter) for transferring video spatial information. As shown in the bottom right of Fig. 2, both adapters employ a bottleneck architecture consisting of two fully connected (FC) layers with an intermediate activation layer. The first FC layer maps the input to a lower-dimensional space, while the second FC layer maps it back to the original dimensional. Formally, for an input feature matrix $\mathbf{X} \in \mathbb{R}^{N \times d}$ , the spatial adapter could be written as: + +$$ +\operatorname {S - A d p a t e r} (\mathbf {X}) = \mathbf {X} + \mathbf {W} _ {\text {u p}} \left(\operatorname {G E L U} \left(\mathbf {W} _ {\text {d o w n}} (\mathbf {X})\right)\right), \tag {4} +$$ + +![](images/050f80f7babddd762d6b1e01d812c3c5ceb7cd1856cdb4f51bfc20fccf658aea.jpg) +Figure 3. The overview of Temporal Latent-shift Attention module. It is noted that the Latent-shift attention is performed on latent space, but the visualization overview is shown on the image level for clear presentation. + +where $\mathbf{W}_{\mathrm{up}}$ and $\mathbf{W}_{\mathrm{down}}$ are the learnable matrix with dimension $d\times l$ and $l\times d, l < d$ . To preserve the structure of the original network and the pretrained weights, we initialize the second FC layer $\mathbf{W}_{\mathrm{down}}$ with zeros. To adapt to the spatial features of videos, we incorporate the adapter after the latent-shift attention layer. Additionally, we observe that adding an adapter to the feed-forward network (FFN) also helps the network transfer spatial information to videos. We will provide examples in Sec. 4.4. During training, all layers of the attention block are fixed, and only the adapters are updated. + +Temporal Adapter While the spatial adapter effectively transfers spatial information to video, modeling temporal information is crucial for T2V generation tasks. Previous approaches incorporate temporal convolution [11, 85, 86] or temporal attention [8, 88, 104] modules to capture temporal relationships. Although these modules are effective in modeling temporal dynamics, they often come with a huge number of parameters and high-dimensional input feature, resulting in significant computational and training costs. + +To address this issue, we utilize the temporal adapter module for temporal modeling as [69, 128]. In contrast to conventional spatial adapter modules, the temporal adapter module employs depth-wise 3D convolution instead of an intermediate activation layer [36]. The temporal adapter could be formally written as: + +$$ +\operatorname {T - A d p a t e r} (\mathbf {X}) = \mathbf {X} + \mathbf {W} _ {\mathrm {u p}} (3 \mathrm {D} - \operatorname {C o n v} (\mathbf {W} _ {\mathrm {d o w n}} (\mathbf {X}))). \tag {5} +$$ + +By utilizing 3D convolutions in lower-dimensional input, our approach significantly alleviates the complexity of temporal modeling. As a result, our method achieves efficient memory usage during training and exhibits the fastest inference speed among competitive approaches, as demonstrated in Table 1. + +Temporal Latent-Shift Attention In the original T2I framework, the attention block of the U-Net only performs self-attention for individual frames, neglecting the information from other frames. While joint-space-time attention, as demonstrated in [2, 7, 96, 114], can effectively model tem + +![](images/e848e44034601bf1ae52de4940bee4d1035b1ee513ab20c5550eec8195d33631.jpg) +Figure 4. During inference, we sample a novel video from the latent noise inverted from the input video, guided by an edited prompt (e.g., "a Porsche sports car is moving on the road"). + +poral dependencies, it introduces a quadratic complexity in terms of attention calculation. For a video with $L$ frames and $N$ tokens, the complexity of global spatial-temporal attention becomes $O(L^2 N^2)$ . To address this issue, we propose a latent-shift attention module as shown in Fig. 3. In addition to considering tokens within the current frame, we further conduct a patch-level shifting operation along the temporal dimension to shift tokens from the preceding $T$ frames onto the current frame, thereby composing a new latent feature frame. We concatenate the latent feature of the current frame with the temporally shifted latent feature, forming the keys and values for attention calculation. The latent-shift attention can be formally written as: + +$$ +\mathbf {Q} = \mathbf {W} _ {\mathrm {q}} \left(\boldsymbol {x} _ {z _ {i}}\right), \tag {6} +$$ + +$$ +\mathbf {K} = \mathbf {W} _ {\mathrm {k}} \left[ \boldsymbol {x} _ {z _ {i}}, \boldsymbol {x} _ {z _ {s h i f t}} \right], \tag {7} +$$ + +$$ +\mathbf {V} = \mathbf {W} _ {\mathrm {v}} \left[ \boldsymbol {x} _ {z _ {i}}, \boldsymbol {x} _ {z _ {\text {s h i f t}}} \right], \tag {8} +$$ + +where $\boldsymbol{x}_{z_i}$ denotes the query frame and $[\cdot]$ means concatenate. This approach reduces the complexity of attention to $O(2LN^2)$ , significantly lowering the computational burden compared to global attention. Moreover, it allows the model to learn the relationships between adjacent frames, ensuring better temporal consistency in video generation. + +# 3.4. Super Resolution and Editing Models + +Super Resolution (SR) Due to constraints of limited GPU memory and the lack of high-resolution video-text datasets, most existing methods [1, 8, 35], including ours, are only able to generate images at a resolution of $256 \times 256$ . To overcome this limitation and generate higher-resolution outputs, we adopt a two-stage training approach similar to cascaded Diffusion Models [40]. In the first stage, we generate videos with a $256 \times 256$ resolution using our SimDA methods. In the second stage, we employ an LDM upsampler [79] to enhance the resolution of the videos to $1024 \times 1024$ . We incorporate noise augmentation and noise level conditioning, and train a super-resolution model using the following equation: + +$$ +\mathbb {E} _ {\boldsymbol {x}, \epsilon \sim \mathcal {N} (\boldsymbol {0}, \mathbf {I}), t} [ \| \epsilon - \epsilon_ {\theta} ([ \boldsymbol {x} _ {t}, \boldsymbol {x} _ {l o w} ], \boldsymbol {c}, t) \| _ {2} ^ {2} ], \tag {9} +$$ + +where $\pmb{x}_{low}$ is the low-resolution video, we concatenate it with $\pmb{x}_t$ frame by frame following Video LDM [8]. The architecture of SR is similar to T2V model in the first stage, + +![](images/187c701cb8345bbe82218a001a6f8f596ba54c3c5baaae46b165c38bb0fdbb1d.jpg) +Figure 5. Text-to-Video generation comparison on the user study evaluation set. + +![](images/29abc1c5a9912f55f43909fcd0118f536c739ea79705b3248e5efee5732948a2.jpg) + +![](images/6876e3673f27beef49514899611a8b0df912bb716ca87fab64f3fc3d09dddb3d.jpg) + +Table 2. Text-to-Video generation comparison on MSR-VTT [118] dataset. We report the Fréchet Video Distance (FVD) scores and CLIPSIM scores. + +
MethodTraining DataResolutionZero-shotParams(B)FVD(↓)CLIPSIM(↑)
GODIVA [108]MSR-VTT128x128No--0.2402
NüWA [109]MSR-VTT128x128No0.87-0.2439
Make-A-Video [88]WebVid-10M + HD-VILA-10M256x256Yes9.72-0.3049
VideoFactory [104]WebVid-10M + HD-VG-130M256x256Yes2.04-0.3005
LVDM [35]WebVid-2M256x256Yes1.167420.2381
MMVG [25]WebVid-2.5M256x256Yes--0.2644
CogVideo [42]WebVid-5.4M256x256Yes15.512940.2631
ED-T2V [57]WebVid-10M256x256Yes1.30-0.2763
MagicVideo [136]WebVid-10M256x256Yes-998-
Video-LDM [8]WebVid-10M256x256Yes4.20-0.2929
VideoComposer [105]WebVid-10M256x256Yes1.855800.2932
Latent-Shift [1]WebVid-10M256x256Yes1.53-0.2773
VideoFusion [61]WebVid-10M256x256Yes1.835810.2795
SimDA (Ours)WebVid-10M256x256Yes1.084560.2945
+ +we change the original U-Net block by adding Spatial and Temporal Adapters as described in Sec. 3.3 and only finetune the newly added modules. + +Text-guided Video Editing In addition to performing T2V generation, our method could turn into one-shot tuning for text-guided video editing following Tune-A-Video [110]. The training pipeline of editing model is the same as our T2V method. However, for the inference stage, we adopt the DDIM inversion latents instead of random noisy latents together with edited prompt for novel video generation as shown in Fig. 4. By doing so, the pixel-level information control could remain in the inversion latent as demonstrated in [110]. Owing to the light-weight module and efficient pipeline of our method, SimDA needs fewer training steps (200 steps compared to 500 steps) and thus the training time and inference time is much faster than Tune-A-Video [110]. + +# 4. Experiments + +# 4.1. Implementation Details + +Our T2V method is composed of two-stage models. The first model predicts video frames with a resolution $256 \times 256$ (with a latent size of $32 \times 32$ ), while the second model is a $4 \times$ upsampler, producing a resolution of $1024 \times 1024$ . We train the general T2V model on WebVid-10M [5] dataset following [1, 8]. We follow previous methods [1, 104, 136] to report the CLIP score [75] and FVD (Frechet Video Distance) score [99] on MSR-VTT [118]. Besides, we compare the FVD score and CLIP score on the evaluation set of WebVid [5] as in VideoFactory [104]. We also compare the parameter scale and inference speed of our method with some open-sourced methods [1, 35, 61]. Finally, we also provide a user study between our work and VDM [41], Latent-shift [1], Video-Fusion [61] and LVDM [35]. + +# 4.2. Evaluation on Text-to-Video Generation + +To fully evaluate the generation performance of our SimDA, we conduct automatic evaluations on two distinct datasets: WebVid [5] (Val), which shares the same domain as the training data and MSR-VTT [118] in a zero-shot setting. + +Evaluation on MSR-VTT As shown in Table 2, we evaluate CLIPSIM [75] and FVD [99] on the widely used video generation benchmarks, MSR-VTT [118]. We randomly select one text prompt per example from MSR-VTT [118] and generate a total of 2,990 videos. Despite being a zero-shot setting, our method achieves an average CLIPSIM of 0.2945 that surpasses most of the competitors, indicating a strong semantic alignment between the generated videos and the input text. Though Make-A-Video [88] and VideoFactory [104] offer higher CLIP scores, they utilize additional large-scale HD-VILA [119] datasets for training. + +Evaluation on WebVid As shown in Table 3, we create a validation set consisting of 4,476 randomly extracted text-video pairs from WebVid-10M. These pairs are not included in the training data following [104]. We conduct evaluations on this validation set and obtain impressive results. Our method achieves an FVD score of 363.98 and a CLIP-SIM score of 0.3054. These scores are significantly higher than those achieved by existing methods such as ModelScope [103] and LVDM [35]. Besides, our method shows competitive results compared to VideoFactory [104] which is trained with much larger datasets. These results clearly demonstrate the superiority of our approach. + +Human Evaluation In order to address the limitations of existing metrics and assess the performance of our SimDA from a human perspective, we conduct an extensive user study. The study involves comparing our method with four state-of-the-art methods. Specifically, we select two publicly available models, namely ModelScope [103] and LVDM [35]. Additionally, we consider two methods with similar scale parameters, VDM [41] and Latent-shift [1], which only showcase some samples on their websites. + +For each case, participants were provided with two video samples, one is generated by our method and the other is from a competitor. They were then asked to compare the two samples in terms of video quality and text-video similarity. To ensure fairness in the comparisons, we also report the ratio of network parameters compared to ours. The results, along with the parameter ratios, are presented in Table 4. The user study approach allows us to gain in-depth insights into the subjective evaluation of our method. + +Qualitative Results The visualization of T2V generation results are shown in Fig. 1(a). Besides, we show the comparison results in Fig. 5. More examples can be found in our supplementary material. + +Parameter Size and Inference Speed We conduct a comparison of the number of parameters and inference speed, + +Table 3. Text-to-video generation on the validation set of WebVid [5]. We report the FVD and CLIPSIM scores. + +
MethodParams(B)FVD(↓)CLIPSIM(↑)
LVDM [35]1.16455.530.2751
ModelScope [103]1.83414.110.3000
VideoFactory [104]2.04292.350.3070
SimDA (Ours)1.08363.980.3054
+ +Table 4. User preference is depicted as a percentage indicating the proportion of individuals favoring our method over the compared approach. Param Ratio means the ratio of the network parameter v.s. Ours. + +
SampleMethodParam RatioQualityFaithfulness
Open WebsiteVDM [41]0.83×85.2%81.4%
Latent-Shift [1]1.41×81.5%79.3%
Pretrained ModelModelScope [103]1.69×78.3%79.5%
LVDM [35]1.07×83.4%84.7%
+ +and the results are presented in Table 1. For the speed comparison, we select CogVideo [42], Latent-Shift [1] and LVDM [35]. SimDA, on the other hand, stands out as it is significantly smaller than previous works and exhibits faster inference speed compared to other methods. Despite having fewer parameters, SimDA achieves superior performance in various benchmarks when compared to other methods. This validation further highlights our advantages in terms of model efficiency and performance. + +# 4.3. Evaluation on Text-guided Video Editing + +Following the methodology of previous studies [110], we employ CLIP score [75] and a user study to evaluate the performance of different methods in terms of frame consistency and textual alignment. + +First, we calculate the CLIP image embedding for all frames in the edited videos to measure frame consistency. The average cosine similarity between pairs of video frames is reported. Additionally, to assess textual faithfulness, we compute the average CLIP score between frames of the output videos and the corresponding edited prompts. A total of 15 videos from the dataset [72] were selected and edited based on object, background and style, resulting in 75 edited videos for each model. The average results, presented in Table 5, highlight our method's exceptional ability to achieve semantic alignment. + +Secondly, we conduct a user study involving videos and text prompts. Participants were asked to vote for the edited videos that exhibited the best temporal consistency and those most accurately matched the textual description. Table 5 demonstrates that our method, SimDA, receives the highest number of votes in both aspects, indicating superior editing quality and a strong preference from users in practical scenarios. + +Table 5. Quantitative comparison with evaluated baseline [110]. The "Training" refers to the process of optimization, and "Memory" refers to the GPU memory. + +
MethodFrame consistencyTextual alignmentRuntime [min]Memory [Gib]Params [Mb]
CLIP Score↑User Vote↑CLIP Score↑User Vote↑Training↓Inference↓Training↓Inference↓Tuned↓
Tune-A-Video [110]94.131.2%31.839.5%9.30.831.311.474.4
SimDA(Ours)94.968.8%31.960.5%2.50.428.68.824.9
+ +![](images/dd079eb50d409e473d462a68bbb8b1ed5ac4b844647e7953f339b5058e968f17.jpg) +Figure 6. Ablation of T2V generation results. TA, SA, AA, FA and LSA refer to Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-Shift Attention. + +# 4.4. Ablation Study + +Here we discuss the effect of each module. We perform experiments on $1K$ samples from the WebVid validation set. + +Effect of Temporal Adapter Temporal modeling is a crucial component of video generation. In our video editing task, when compared to methods that rely on temporal attention modeling like Tune-A-Video [110], we observe that our temporal adapter is more lightweight and achieves superior editing results as in Table 5. Additionally, we conduct ablation experiments, as shown in Table 6 and Fig. 6, where the lack of Temporal Adapter (TA) results in significantly higher FVD and chaotic temporal sequences in the generated videos. + +Effect of Spatial Adapter We also validate the effectiveness of the Spatial Adapter (SA) in transferring spatial knowledge of videos. As shown in Table 6, without the Attention Adapter (AA) and FFN Adapter (FA), the model's + +Table 6. Ablation study on different modules. We report the FVD [99] and CLIPSIM [75] on $1K$ samples from the validation set of WebVid-10M [5]. TA, SA, AA, FA and LSA represent Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-shift Attention, respectively. + +
TAAAFALSAFVD(↓)CLIPSIM(↑)
w/o TA✓✓✓1470.10.2629
w/o SA✓✓811.30.2822
w/o AA✓✓✓764.80.2851
w/o FA✓✓✓623.70.2962
w/o LSA✓✓✓618.20.3011
Ours✓✓✓✓530.20.3034
+ +FVD and CLIPSIM scores for generated videos will become worse. Additionally, it can be observed from the Fig. 6 that the model exhibits misconceptions in understanding the text prompt without the spatial adapter. + +Effect of Latent Shift Attention To investigate the impact of Latent-shift Attention (LSA), we replace it with regular single-frame spatial attention. Besides observing a decline in FVD and text alignment CLIPSIM scores in Table 6, we also test the CLIPSIM of each frame within the same video, which decreased from 96.4 to 94.5. This demonstrates that our LSA module can effectively model the relationship of adjacent frames, leading to more consistent videos. + +# 5. Conclusion + +In this paper, we proposed SimDA, a parameter-efficient video diffusion model for text-guided video generation and editing. With the proposed light-weight spatial and temporal adapters, our method not only transferred from powerful spatial information but also modeled temporal relationships with the least new parameters. The experimental results demonstrated that our approach has the fastest training and inference speed while maintaining competitive generation and editing results. Our work is the first parameter-efficient video diffusion method serving as an efficient T2V fine-tuning baseline and paved the way for future research. + +Acknowledge This work was supported by National Science and Technology Major Project (No. 2021ZD0112805) and in part by National Natural Science Foundation of China (No. 62032006 and No. 62102092.). The computations in this research were performed using the CFFFplatform of Fudan University. + +# References + +[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, JiaBin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2, 3, 4, 5, 6, 7 +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 5 +[3] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. arXiv preprint arXiv:1710.11252, 2017. 2 +[4] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6, 7, 8 +[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3 +[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5 +[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 4, 5, 6 +[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 2022. 2 +[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3 +[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 5 +[12] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In ICCV, 2019. 2 +[13] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 2, 3, 4 +[14] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 2 +[15] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 2 +[16] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional + +transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.2 +[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2 +[18] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 2022. 2 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2 +[21] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3 +[22] Qijun Feng, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Fdgaussian: Fast gaussian splatting from single image via geometric-aware diffusion model. arXiv preprint arXiv:2403.10242, 2024. 3 +[23] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 2 +[24] Jean-Yves Franceschi, Edouard Delasalles, Mickael Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In ICML, 2020. 2 +[25] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 6 +[26] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3 +[27] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2 +[28] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474, 2023. 2 +[29] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2 +[30] Sonam Gupta, Arti Keshari, and Sukhendu Das. Rvgan: Recurrent gan for unconditional video generation. In CVPR, 2022. 2 + +[31] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2 +[32] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2 +[33] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. arXiv preprint arXiv:2110.04366, 2021. 3 +[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4 +[35] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 4, 5, 6, 7 +[36] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[37] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. *ICLR*, 2023. 3, 4 +[38] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2 +[39] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 4 +[40] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2, 5 +[41] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 6, 7 +[42] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 4, 6, 7 +[43] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022.2 +[44] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3 +[45] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4 +[46] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3 + +[47] Shibo Jie and Zhi-Hong Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 3 +[48] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 2 +[49] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2 +[50] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. +[51] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 2 +[52] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 2 +[53] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3 +[54] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4 +[55] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2 +[56] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 3 +[57] Jiawei Liu, Weining Wang, Wei Liu, Qian He, and Jing Liu. Ed-t2v: An efficient training framework for diffusion-based text-to-video generation. In IJCNN, 2023. 6 +[58] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 2 +[59] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3 +[60] Pauline Luc, Aidan Clark, Sander Dieleman, Diego de Las Casas, Yotam Doron, Albin Cassirer, and Karen Simonyan. Transformation-based adversarial video prediction on large-scale data. arXiv preprint arXiv:2003.04035, 2020. 2 +[61] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 6 +[62] Tanya Marwah, Gaurav Mittal, and Vineeth N Balasubramanian. Attentive semantic video generation using captions. In ICCV, 2017. 2 +[63] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. ICLR, 2022. 3 +[64] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync-draw: Automatic video generation using + +deep recurrent attentive architectures. In ACM Multimedia, 2017. 2 +[65] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4 +[66] Andres Munoz, Mohammadreza Zolfaghari, Max Argus, and Thomas Brox. Temporal shift gan for large scale video generation. In WACV, 2021. 3 +[67] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[68] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2 +[69] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. NeurIPS, 2022. 2, 3, 5 +[70] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2 +[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. In CVPR, 2024. 2 +[72] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7 +[73] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 3 +[74] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7, 8 +[76] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 2 +[77] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2 +[78] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 + +[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5 +[80] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2 +[81] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 4 +[82] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2 +[83] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. IJCV, 2020. 2 +[84] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH, 2022. 2 +[85] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 5 +[86] Baifeng Shi, Qi Dai, Judy Hoffman, Kate Saenko, Trevor Darrell, and Huijuan Xu. Temporal action detection with multi-level supervision. In ICCV, 2021. 5 +[87] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. 2023. 3 +[88] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. *ICLR*, 2023. 2, 4, 5, 6, 7 +[89] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2 +[90] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4 +[91] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[92] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2 +[93] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. NeurIPS, 2021. 3 +[94] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 2 + +[95] Shuyuan Tu, Tianzhen Guan, and Li Kuang. Multiple biological granularities network for person re-identification. In ICMR, 2022. 3 +[96] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. In ICCV, 2023. 5 +[97] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 3 +[98] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3 +[99] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7, 8 +[100] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 2 +[101] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Masked conditional video diffusion for prediction, generation, and interpolation. arXiv preprint arXiv:2205.09853, 2022. 2 +[102] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 2016. 2 +[103] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 7 +[104] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2, 5, 6, 7 +[105] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 6 +[106] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, Kai Qiu, Yuhui Yuan, Chuanxin Tang, Xiaoyan Sun, Chong Luo, and Baining Guo. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2 +[107] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2 +[108] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 6 +[109] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6 + +[110] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 6, 7, 8 +[111] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In ECCV,2022.3 +[112] Zhen Xing, Yijiang Chen, Zhixin Ling, Xiangdong Zhou, and Yu Xiang. Few-shot single-view 3d reconstruction with memory prior contrastive network. In ECCV, 2022. 3 +[113] Zhen Xing, Hengduo Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised single-view 3d reconstruction via prototype shape priors. In ECCV, 2022. 3 +[114] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. In CVPR, 2023. 5 +[115] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3 +[116] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 2 +[117] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2 +[118] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 6, 7 +[119] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 7 +[120] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295, 2023. 2 +[121] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2 +[122] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022. 2 +[123] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3 +[124] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. ICLR, 2023. 2, 3, 4 +[125] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, + +Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2 +[126] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2 +[127] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021. 3 +[128] Bowen Zhang, Xiaojie Jin, Weibo Gong, Kai Xu, Zhao Zhang, Peng Wang, Xiaohui Shen, and Jiashi Feng. Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 2, 3, 5 +[129] Hao Zhang, Yanbin Hao, and Chong-Wah Ngo. Token shift transformer for video classification. In ACM Multimedia, 2021. 3 +[130] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 2 +[131] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4 +[132] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2 +[133] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 3 +[134] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. arXiv preprint arXiv:2305.17431, 2023. 3 +[135] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In CVPR, 2024. 2 +[136] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 6 +[137] Rui Zhu, Yingwei Pan, Yehao Li, Ting Yao, Zhenglong Sun, Tao Mei, and Chang Wen Chen. Sd-dit: Unleashing the power of self-supervised discrimination in diffusion transformer. In CVPR, 2024. 2 \ No newline at end of file diff --git a/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/images.zip b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..140ae5f62044043f9ebf8cb68139bdeef8412d20 --- /dev/null +++ b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:742eebc49a43736da2dca826dd04a2520824f9f66954d26aeb73a4f355109359 +size 835694 diff --git a/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/layout.json b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..884c298a7138d98a03f7f295784b721dfe86bb75 --- /dev/null +++ b/2024/SimDA_ Simple Diffusion Adapter for Efficient Video Generation/layout.json @@ -0,0 +1,12413 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "type": "text", + "content": "SimDA: Simple Diffusion Adapter for Efficient Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 142, + 179, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 142, + 179, + 157 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 179, + 157 + ], + "type": "text", + "content": "Zhen Xing1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 143, + 232, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 143, + 232, + 156 + ], + "spans": [ + { + "bbox": [ + 194, + 143, + 232, + 156 + ], + "type": "text", + "content": "Qi Dai3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 248, + 143, + 291, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 143, + 291, + 156 + ], + "spans": [ + { + "bbox": [ + 248, + 143, + 291, + 156 + ], + "type": "text", + "content": "Han Hu3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 143, + 378, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 378, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 378, + 156 + ], + "type": "text", + "content": "Zuxuan " + }, + { + "bbox": [ + 306, + 143, + 378, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{1,2\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 394, + 143, + 478, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 143, + 478, + 157 + ], + "spans": [ + { + "bbox": [ + 394, + 143, + 478, + 157 + ], + "type": "text", + "content": "Yu-Gang Jiang1,2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 159, + 490, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 159, + 490, + 173 + ], + "spans": [ + { + "bbox": [ + 107, + 159, + 490, + 173 + ], + "type": "text", + "content": "1 Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 173, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 173, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 113, + 173, + 480, + 187 + ], + "type": "text", + "content": "2 Shanghai Collaborative Innovation Center of Intelligent Visual Computing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 186, + 361, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 186, + 361, + 199 + ], + "spans": [ + { + "bbox": [ + 233, + 186, + 361, + 199 + ], + "type": "text", + "content": "3 Microsoft Research Asia" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 175, + 201, + 415, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 201, + 415, + 214 + ], + "spans": [ + { + "bbox": [ + 175, + 201, + 415, + 214 + ], + "type": "text", + "content": "https://chenhsing.github.io/SimDA" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 77, + 240, + 129, + 292 + ], + "blocks": [ + { + "bbox": [ + 77, + 240, + 129, + 292 + ], + "lines": [ + { + "bbox": [ + 77, + 240, + 129, + 292 + ], + "spans": [ + { + "bbox": [ + 77, + 240, + 129, + 292 + ], + "type": "image", + "image_path": "9efabbe8c1a1418a10d2477a63515b1f16e8973f54cf5bdd450b7cf92f3b4974.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 293, + 281, + 303 + ], + "lines": [ + { + "bbox": [ + 84, + 293, + 281, + 303 + ], + "spans": [ + { + "bbox": [ + 84, + 293, + 281, + 303 + ], + "type": "text", + "content": "A horse galloping through van Gogh's 'Starry Night'." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 130, + 241, + 182, + 292 + ], + "blocks": [ + { + "bbox": [ + 130, + 241, + 182, + 292 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 182, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 182, + 292 + ], + "type": "image", + "image_path": "54c5a14d39d31c8c500088b7d2158eb63e88d6297bed8daff155c8efc050de3f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 292, + 281, + 303 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 281, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 281, + 303 + ], + "type": "text", + "content": "oping through van Gogh's 'Starry Night'." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 184, + 241, + 235, + 292 + ], + "blocks": [ + { + "bbox": [ + 184, + 241, + 235, + 292 + ], + "lines": [ + { + "bbox": [ + 184, + 241, + 235, + 292 + ], + "spans": [ + { + "bbox": [ + 184, + 241, + 235, + 292 + ], + "type": "image", + "image_path": "51af8c497a376e1777b5bd3aed89a8c27b33625e91aa7dd3c924f8f875d44881.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 292, + 287, + 303 + ], + "lines": [ + { + "bbox": [ + 184, + 292, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 184, + 292, + 287, + 303 + ], + "type": "text", + "content": "van Gogh's 'Starry Night'." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 78, + 308, + 129, + 359 + ], + "blocks": [ + { + "bbox": [ + 78, + 308, + 129, + 359 + ], + "lines": [ + { + "bbox": [ + 78, + 308, + 129, + 359 + ], + "spans": [ + { + "bbox": [ + 78, + 308, + 129, + 359 + ], + "type": "image", + "image_path": "1c84413c441e44bab9c4cecc026c5b864f8da99d6e2ad6cd749926d0354d9419.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 360, + 274, + 369 + ], + "lines": [ + { + "bbox": [ + 90, + 360, + 274, + 369 + ], + "spans": [ + { + "bbox": [ + 90, + 360, + 274, + 369 + ], + "type": "text", + "content": "Fireworks being displayed for a crowd of people." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 130, + 308, + 182, + 359 + ], + "blocks": [ + { + "bbox": [ + 130, + 308, + 182, + 359 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 182, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 182, + 359 + ], + "type": "image", + "image_path": "497d9f3f758f7784f7197e9a9b2b4e023c0c6122604da8da70f4e9cd76e4d851.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 360, + 288, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 360, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 288, + 369 + ], + "type": "text", + "content": "being displayed for a crowd of people." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 78, + 373, + 129, + 424 + ], + "blocks": [ + { + "bbox": [ + 78, + 373, + 129, + 424 + ], + "lines": [ + { + "bbox": [ + 78, + 373, + 129, + 424 + ], + "spans": [ + { + "bbox": [ + 78, + 373, + 129, + 424 + ], + "type": "image", + "image_path": "3a71a8d13641763cbfb4f9ea0c5d59e96d814421a6d1b59b26443a384f84cb45.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 425, + 279, + 436 + ], + "lines": [ + { + "bbox": [ + 87, + 425, + 279, + 436 + ], + "spans": [ + { + "bbox": [ + 87, + 425, + 279, + 436 + ], + "type": "text", + "content": "A dog wearing virtual reality goggles on the grass." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 130, + 373, + 183, + 424 + ], + "blocks": [ + { + "bbox": [ + 130, + 373, + 183, + 424 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 183, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 183, + 424 + ], + "type": "image", + "image_path": "1ec2deaf54f4f75a87fd1dcca5989858860949ebad426d8f921202a9f9df494d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 184, + 373, + 236, + 424 + ], + "blocks": [ + { + "bbox": [ + 184, + 373, + 236, + 424 + ], + "lines": [ + { + "bbox": [ + 184, + 373, + 236, + 424 + ], + "spans": [ + { + "bbox": [ + 184, + 373, + 236, + 424 + ], + "type": "image", + "image_path": "81e6591cb1bcf9611d73738d3441118fddd7180bd0fee80e070b0ffaadd37fac.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 237, + 373, + 288, + 424 + ], + "blocks": [ + { + "bbox": [ + 237, + 373, + 288, + 424 + ], + "lines": [ + { + "bbox": [ + 237, + 373, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 237, + 373, + 288, + 424 + ], + "type": "image", + "image_path": "ed2f153bbf69a6e5061b9bdadd40a643807d6b92eeec96add1c443bb8291b524.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 79, + 439, + 130, + 490 + ], + "blocks": [ + { + "bbox": [ + 79, + 439, + 130, + 490 + ], + "lines": [ + { + "bbox": [ + 79, + 439, + 130, + 490 + ], + "spans": [ + { + "bbox": [ + 79, + 439, + 130, + 490 + ], + "type": "image", + "image_path": "2eee7dfcfd07a85522d6080da5f5b32d23af35741f91587126d25457c5bd8990.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 132, + 439, + 183, + 491 + ], + "blocks": [ + { + "bbox": [ + 82, + 491, + 287, + 501 + ], + "lines": [ + { + "bbox": [ + 82, + 491, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 82, + 491, + 287, + 501 + ], + "type": "text", + "content": "Time lapse at a fantasy landscape, 4k, high resolution." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 439, + 183, + 491 + ], + "lines": [ + { + "bbox": [ + 132, + 439, + 183, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 183, + 491 + ], + "type": "image", + "image_path": "0eed4c60c3aa769d57c76e794b23caf49aae0f1367931cd7ac55d9d38769b7a0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 510, + 227, + 520 + ], + "lines": [ + { + "bbox": [ + 127, + 510, + 227, + 520 + ], + "spans": [ + { + "bbox": [ + 127, + 510, + 227, + 520 + ], + "type": "text", + "content": "(a) Text-to-Video Generation" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 525, + 547, + 548 + ], + "lines": [ + { + "bbox": [ + 46, + 525, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 547, + 548 + ], + "type": "text", + "content": "Figure 1. Examples of our SimDA results: (a) The results of open-wild Text-to-Video Generation. (b) Text-guided video Editing results using one text-video pair tuning." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 131, + 491, + 287, + 501 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 287, + 501 + ], + "type": "text", + "content": "a fantasy landscape, 4k, high resolution." + } + ] + } + ], + "index": 30, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 302, + 240, + 354, + 292 + ], + "blocks": [ + { + "bbox": [ + 302, + 240, + 354, + 292 + ], + "lines": [ + { + "bbox": [ + 302, + 240, + 354, + 292 + ], + "spans": [ + { + "bbox": [ + 302, + 240, + 354, + 292 + ], + "type": "image", + "image_path": "0d7f70d1a3f76912b31a1d576a6092f1ca3ff2690b3f87936a4ad8737e16d5d6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 293, + 473, + 303 + ], + "lines": [ + { + "bbox": [ + 343, + 293, + 473, + 303 + ], + "spans": [ + { + "bbox": [ + 343, + 293, + 473, + 303 + ], + "type": "text", + "content": "A jeep car is moving on the road." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 356, + 241, + 407, + 292 + ], + "blocks": [ + { + "bbox": [ + 356, + 241, + 407, + 292 + ], + "lines": [ + { + "bbox": [ + 356, + 241, + 407, + 292 + ], + "spans": [ + { + "bbox": [ + 356, + 241, + 407, + 292 + ], + "type": "image", + "image_path": "291b22a2c11e7846b00f67c115996cc826d74b9ea9d7ca57455c62d949679b60.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 408, + 240, + 460, + 292 + ], + "blocks": [ + { + "bbox": [ + 408, + 240, + 460, + 292 + ], + "lines": [ + { + "bbox": [ + 408, + 240, + 460, + 292 + ], + "spans": [ + { + "bbox": [ + 408, + 240, + 460, + 292 + ], + "type": "image", + "image_path": "5e3acac173ff9afee5172d47a6479db6709d6de6496690e5b258cb74ad195f78.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 462, + 240, + 514, + 292 + ], + "blocks": [ + { + "bbox": [ + 462, + 240, + 514, + 292 + ], + "lines": [ + { + "bbox": [ + 462, + 240, + 514, + 292 + ], + "spans": [ + { + "bbox": [ + 462, + 240, + 514, + 292 + ], + "type": "image", + "image_path": "87b4e1eb9dc8a50c6d93a08dbfe8c3ff156ca310a94347f0626b10f4e7815e45.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 302, + 308, + 354, + 359 + ], + "blocks": [ + { + "bbox": [ + 302, + 308, + 354, + 359 + ], + "lines": [ + { + "bbox": [ + 302, + 308, + 354, + 359 + ], + "spans": [ + { + "bbox": [ + 302, + 308, + 354, + 359 + ], + "type": "image", + "image_path": "1b79fc1e4952cd6408806056f28c03154bc627a890ad8d90b7ee6cfa4b1343ae.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 360, + 481, + 369 + ], + "lines": [ + { + "bbox": [ + 344, + 360, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 344, + 360, + 481, + 369 + ], + "type": "text", + "content": "An AE86 car is moving on the road." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 356, + 308, + 407, + 359 + ], + "blocks": [ + { + "bbox": [ + 356, + 308, + 407, + 359 + ], + "lines": [ + { + "bbox": [ + 356, + 308, + 407, + 359 + ], + "spans": [ + { + "bbox": [ + 356, + 308, + 407, + 359 + ], + "type": "image", + "image_path": "4aabf6227c40db2689d68ceee1620c306d33587ecb03cf0b622c3494052043a2.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 408, + 308, + 460, + 359 + ], + "blocks": [ + { + "bbox": [ + 408, + 308, + 460, + 359 + ], + "lines": [ + { + "bbox": [ + 408, + 308, + 460, + 359 + ], + "spans": [ + { + "bbox": [ + 408, + 308, + 460, + 359 + ], + "type": "image", + "image_path": "914252461b512bbb7c69bf1dbb979ee32eebebf9b49eb2b96f1737ab39b970e1.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 462, + 308, + 514, + 359 + ], + "blocks": [ + { + "bbox": [ + 462, + 308, + 514, + 359 + ], + "lines": [ + { + "bbox": [ + 462, + 308, + 514, + 359 + ], + "spans": [ + { + "bbox": [ + 462, + 308, + 514, + 359 + ], + "type": "image", + "image_path": "0e08fe1255e39a05305ad73d066cbb8f8a2e58a28ac39997077803791be343a2.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 302, + 373, + 354, + 424 + ], + "blocks": [ + { + "bbox": [ + 302, + 373, + 354, + 424 + ], + "lines": [ + { + "bbox": [ + 302, + 373, + 354, + 424 + ], + "spans": [ + { + "bbox": [ + 302, + 373, + 354, + 424 + ], + "type": "image", + "image_path": "a1c4c33c835723bd0f72bb7a0ef3ebeaa4d8187b4607467eb3ddfae07c577294.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 425, + 507, + 435 + ], + "lines": [ + { + "bbox": [ + 323, + 425, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 323, + 425, + 507, + 435 + ], + "type": "text", + "content": "A jeep car is moving on the road, cartoon style." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 356, + 373, + 407, + 424 + ], + "blocks": [ + { + "bbox": [ + 356, + 373, + 407, + 424 + ], + "lines": [ + { + "bbox": [ + 356, + 373, + 407, + 424 + ], + "spans": [ + { + "bbox": [ + 356, + 373, + 407, + 424 + ], + "type": "image", + "image_path": "d842fdbf571bfd8b4a42392bfe3eb8d54702bf51774490192f1bf6172cfb7b68.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 408, + 373, + 460, + 424 + ], + "blocks": [ + { + "bbox": [ + 408, + 373, + 460, + 424 + ], + "lines": [ + { + "bbox": [ + 408, + 373, + 460, + 424 + ], + "spans": [ + { + "bbox": [ + 408, + 373, + 460, + 424 + ], + "type": "image", + "image_path": "0dc5e77d4502d5f352f1dd8638b66163cf74f4c3546e1239395f01bfb9d569b6.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 462, + 373, + 514, + 424 + ], + "blocks": [ + { + "bbox": [ + 462, + 373, + 514, + 424 + ], + "lines": [ + { + "bbox": [ + 462, + 373, + 514, + 424 + ], + "spans": [ + { + "bbox": [ + 462, + 373, + 514, + 424 + ], + "type": "image", + "image_path": "317364731f1dec199aac8890ff178e90ad9506449ebbba50c1bffa99aed63e05.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 301, + 439, + 352, + 490 + ], + "blocks": [ + { + "bbox": [ + 301, + 439, + 352, + 490 + ], + "lines": [ + { + "bbox": [ + 301, + 439, + 352, + 490 + ], + "spans": [ + { + "bbox": [ + 301, + 439, + 352, + 490 + ], + "type": "image", + "image_path": "de8c7e2cb38fb0d6d1b91210d4337244cb11def8557c33a4580a93e8b3b62463.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 491, + 496, + 501 + ], + "lines": [ + { + "bbox": [ + 343, + 491, + 496, + 501 + ], + "spans": [ + { + "bbox": [ + 343, + 491, + 496, + 501 + ], + "type": "text", + "content": "Keep car is moving in a forest, in Autumn." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 510, + 460, + 521 + ], + "lines": [ + { + "bbox": [ + 359, + 510, + 460, + 521 + ], + "spans": [ + { + "bbox": [ + 359, + 510, + 460, + 521 + ], + "type": "text", + "content": "(b) Text-guided Video Editing" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 143, + 555, + 192, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 555, + 192, + 567 + ], + "spans": [ + { + "bbox": [ + 143, + 555, + 192, + 567 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 45, + 582, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 582, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 582, + 290, + 715 + ], + "type": "text", + "content": "The recent wave of AI-generated content has witnessed the great development and success of Text-to-Image (T2I) technologies. By contrast, Text-to-Video (T2V) still falls short of expectations though attracting increasing interest. Existing works either train from scratch or adapt large T2I model to videos, both of which are computation and resource expensive. In this work, we propose a Simple Diffusion Adapter (SimDA) that fine-tunes only 24M out of 1.1B parameters of a strong T2I model, adapting it to video generation in a parameter-efficient way. In particular, we turn the T2I model for T2V by designing light-weight spa" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "type": "text", + "content": "tial and temporal adapters for transfer learning. Besides, we change the original spatial attention to the proposed Latent-Shift Attention (LSA) for temporal consistency. With a similar model architecture, we further train a video superresolution model to generate high-definition " + }, + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "type": "inline_equation", + "content": "(1024\\times 1024)" + }, + { + "bbox": [ + 304, + 556, + 547, + 676 + ], + "type": "text", + "content": " videos. In addition to T2V generation in the wild, SimDA could also be utilized in one-shot video editing with only 2 minutes tuning. Doing so, our method could minimize the training effort with extremely few tunable parameters for model adaptation." + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 321, + 702, + 401, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 702, + 401, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 702, + 401, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7827" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 270 + ], + "type": "text", + "content": "Image generation stands on top of the recent AIGC wave. It not only has a significant impact on the academic community but also achieves tremendous success in various applications, such as computer graphics, art and culture, medical imaging, etc. The approaches in this area mainly include methods based on generative adversarial networks (GANs) [29, 49-51, 84], auto-regressive transformers [20, 77, 125], and the latest diffusion models [17, 23, 38, 40, 58, 67, 68, 78, 79, 82, 91, 120]. Among them, diffusion models are the most popular owing to their strong controllability, simple stability, and amazing realism. However, video generation research lags behind due to challenges like the scarcity of publicly available datasets, difficulty in modeling temporal information, and high training costs, hindering progress in this area." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 270, + 286, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 286, + 377 + ], + "type": "text", + "content": "There have been several research endeavors dedicated to exploring video synthesis [3, 9, 12, 15, 24, 27, 30, 31, 42, 48, 52, 55, 60, 62, 64, 70, 83, 89, 94, 100, 102, 107, 108, 121, 126]. In addition, some studies have employed popular diffusion models for video generation [32, 41, 43, 61, 101, 106, 116, 122, 135, 137]. However, most of them involve training models from scratch, which can be time-consuming due to the complex video data. Early attempts were also constrained by GPU memory or hardware limitations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 378, + 286, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 286, + 497 + ], + "type": "text", + "content": "More recently, a small number of T2V (Text-to-Video) approaches have emerged, aiming to fine-tune well-established T2I (Text-to-Image) models [1, 8, 28, 39, 61, 104, 136]. They have incorporated temporal modeling modules (e.g.Imagen video [39], Video LDM [8]) into T2I models, which effectively accelerate the model convergence. However, it should be noted that training such models is still a challenging task due to the massive number of parameters (4B or even 16B) involved in the network architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 286, + 628 + ], + "type": "text", + "content": "In the NLP field, state-of-the-art results of various tasks are generally achieved by adaptation from large pretrained models (i.e., BERT [16], LLMs [14, 74, 76, 132]). However, with the advent of increasingly larger and more powerful foundation models (e.g., GPT-4 with 100T parameters), conducting full fine-tuning of the entire models has become prohibitively expensive and infeasible in terms of training cost and GPU storage. To address the issue, numerous methods based on efficient fine-tuning have emerged rapidly in NLP [44, 45, 53, 54] and computer vision [13, 69, 124, 128]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "content": "In this work, we propose a parameter-efficient video diffusion model, namely Simple Diffusion Adapter (SimDA), that fine-tunes the large T2I (i.e. Stable Diffusion [79]) model for improved video generation. We only add " + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 47, + 630, + 286, + 713 + ], + "type": "text", + "content": " parameters compared to the T2I model. During training, we freeze the original T2I model, and only tune the newly added modules. We further propose a Latent-Shift Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "(LSA) to replace the original spatial attention, which significantly improves the temporal modeling capability and retains consistency without adding new parameters. To this end, our model demands less than 8GB GPU memory for training with a resolution of " + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "16 \\times 256 \\times 256" + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": ", while the inference time speeds up by " + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "39 \\times" + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": " compared to the autoregressive method CogVideo [42]. Besides, we turn an image super resolution framework into the video counterpart with similar architecture, which allows generating high-definition videos of " + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": ". Our model can also be extended to the recently popular diffusion-based video editing [110], achieving significant " + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": " faster training while retaining comparable results, as evidenced by the editing examples presented in Fig 1 (b). In conclusion, the contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 251, + 545, + 395 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 251, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 545, + 287 + ], + "type": "text", + "content": "- We explore the simple adaptation from image diffusion to video diffusion, exhibiting that tuning extremely few parameters can achieve surprisingly good results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 288, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 323 + ], + "type": "text", + "content": "- With the helpful light-weight adapters and the proposed latent-shift attention, our method can effectively model the temporal relations with negligible cost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 323, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 545, + 358 + ], + "type": "text", + "content": "- Our diffusion adapter could be extended to text-guided video super-resolution and video editing, significantly facilitating the model training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 360, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 545, + 395 + ], + "type": "text", + "content": "- SimDA significantly alleviates the training cost and speeds up the inference time, while remaining competitive results compared to other methods." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 406, + 391, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 391, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 391, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 498 + ], + "type": "text", + "content": "Text-to-Video Generation Similar to the advancements in Text-to-Image (T2I) generation [18, 71, 79, 130], early approaches for Text-to-Video (T2V) generation [55, 64, 70] were based on Generative Adversarial Networks (GANs) and primarily applied to domain-specific videos such as simple human actions [92] or clouds moving [117]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 700 + ], + "type": "text", + "content": "Recently, T2V methods are most based on fine tuning the T2I [78, 79] diffusion models. For instance, Make-A-Video [88] proposes fine-tuning a pretrained DALLE2 [78] model solely on video data to learn motion patterns, enabling T2V generation without explicitly training on text-video pairs. Video Diffusion Models [41] and Imagen Video [39] perform joint text-image and text-video training, treating images as independent frames and disabling temporal layers in the U-Net [80] architecture. Besides, Video LDM [8], Latent-Shift [1], VideoFactory [104], MagicVideo [136] and our methods utilize the popular open-sourced T2I Stable Diffusion [79] model. While the progress in video generation is impressive, the parameters of video generation can be highly large. As shown in Table 1, Make-A-Video [88] requires six models and 9.7B parameters and Imagen Video [39] utilizes eight models with 16.3B parameters, which limits the training efficiency of T2V models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Text guided Video Editing In the realm of content gen" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7828" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "eration, an alternative avenue is the manipulation of existing images [10, 37, 63, 98] and videos [6, 21, 59, 73, 87, 110, 115, 123, 134] using textual input as a means of control, rather than relying solely on unbridled text-based generation. SDEdit [63] introduces noise to images and then reconstructs them for the purpose of editing. Prompt-to-prompt [37] and Plug-and-Play [98] modify the cross-attention map by altering the textual description, thus influencing the editing process. When it comes to video editing, Tune-A-Video [110] fine-tunes the T2I model on a single video, enabling the generation of new videos with similar motion patterns. Video-P2P [59] and FateZero [73] extend the concept of Prompt-to-prompt editing to videos. Text2Live [6] divides videos into layers and enables separate editing of each layer based on text. MotionEditor [97] edits the motion of human while keeping the background." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 289, + 418 + ], + "type": "text", + "content": "Parameter-Efficient Transfer Learning In NLP, parameter efficient fine-tuning [33, 44, 45, 53, 54, 93, 127] were initially proposed to address the heavy computation of full fine-tuning LLMs for various downstream tasks. These techniques aim to reduce the number of trainable parameters, thereby lowering computation costs, while still achieving or surpassing the performance of full fine-tuning. Recently, parameter-efficient transfer learning has also been explored in the field of computer vision [4, 13, 22, 26, 46, 47, 69, 95, 112, 113, 124, 128]. These methods mainly focus on adapting models within simple classification or detection tasks. In contrast, our approach focuses on adapting a T2I model for T2V generation task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 419, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 647 + ], + "type": "text", + "content": "Temporal Shift Module TSM [56] pioneered the introduction of the temporal shift module for action recognition, employing a partial channel shift along the temporal dimension. This approach seamlessly integrates temporal cues from both preceding and succeeding frames into the current frame without incurring additional computational overhead. Subsequently, TokShift [129] implemented channel shifting along the temporal dimension for transformer architectures [19, 133]. TPS [111] further shifted patches instead of channels to model the temporal correlations. However, such direct patch shifting would lead to inconsistency in generation tasks. Additionally, Latent-shift [1] and TSB [66] adapted shift module as TSM [56] within convolution blocks for video generation tasks. In this work, our latent-shift attention (LSA) employs the patch-level shifting manner. In contrast to TPS, we further propose to involve all tokens in the current frame as the keys and values, which guarantees temporal consistency during generation and significantly improves the video quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 657, + 104, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 104, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 104, + 670 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "In this section, we first introduce the preliminaries of Latent Diffusion Model [79] in Sec. 3.1. The pipeline of SimDA is described in Sec. 3.2. Then we detail the proposed spa" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 109 + ], + "type": "text", + "content": "tial and temporal adapters as well as latent-shift attention in Sec. 3.3. Finally, we introduce the super resolution and text-guided video editing model in Sec. 3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 116, + 482, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 116, + 482, + 128 + ], + "spans": [ + { + "bbox": [ + 305, + 116, + 482, + 128 + ], + "type": "text", + "content": "3.1. Preliminaries of Stable Diffusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": "In this subsection, we introduce the preliminaries of Stable Diffusion [79] model. It is a latent diffusion model that operates in the latent space of an autoencoder " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{E}(\\cdot))" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": " is the encoder and " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": " is the decoder. In this model, for an image " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": " with its corresponding latent feature " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 = \\mathcal{E}(I)" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": ", the diffusion forward process involves iteratively adding noise to the latent space," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 335, + 228, + 545, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 228, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 335, + 228, + 545, + 242 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right), \\tag {1}", + "image_path": "bb4e095b822615906fbb0487830980f9c8a1bf653595a12d9cdf678ab3832d1d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": " is the time step, " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t | \\pmb{x}_{t-1})" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": " is the conditional density of " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1}" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": " is identity matrix, and " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": " is hyperparameter. Alternatively, we can directly sample " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": " at any time step from " + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 304, + 251, + 547, + 299 + ], + "type": "text", + "content": " with," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 308, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 308, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 345, + 308, + 545, + 322 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {2}", + "image_path": "d2a6250187040300b09f00ac17194c5b46ed160ea4a81240dea072f09ad2df31.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 330, + 395, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 395, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 395, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 330, + 395, + 344 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i" + }, + { + "bbox": [ + 306, + 330, + 395, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": "In the diffusion backward process, a U-Net denoted as " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": " is trained to predict the noise in the latent space, aiming to iteratively recover " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": ". In this process, as the diffusion progresses and approaches a large value of " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": " is completely disrupted and the latent representation " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": " approximates a standard Gaussian distribution. Consequently, the U-Net " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": " is trained to infer meaningful and valid " + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 344, + 545, + 451 + ], + "type": "text", + "content": " from random Gaussian noises. The training object can be simplified as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 460, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 460, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 354, + 460, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), t} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t) \\| _ {2} ^ {2} ], \\tag {3}", + "image_path": "d60ed3aa188ce279c81d377e2567b10d9211f17d28cd403c3abf82ce3bc2708d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 483, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 482, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 483, + 482, + 495 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 306, + 483, + 482, + 495 + ], + "type": "text", + "content": " is the embedding of condition text." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "text", + "content": "During the inference stage, it samples a valid latent representation " + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "text", + "content": " from the standard Gaussian noise " + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T = \\mathbf{z}_T, \\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "text", + "content": " using DDIM [90] sampling. Then, the model can decode " + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "text", + "content": " using the decoder " + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "text", + "content": " to generate the final image " + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "inline_equation", + "content": "I = \\mathcal{D}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 304, + 496, + 547, + 592 + ], + "type": "text", + "content": ". This process could generate diverse and high-quality images based on the sampled latent representations. In contrast, our method focuses on more challenging high-quality video generation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 599, + 367, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 367, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 367, + 612 + ], + "type": "text", + "content": "3.2. Pipeline" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "SimDA, as shown in Fig. 2, is built upon the previously introduced Stable Diffusion [79]. For a video clip with " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " frames, denoted as " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^t" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": ", we first pass it through a pretrained encoder " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " to obtain the corresponding latent feature " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_i\\}_{i=1}^t" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": ". We then input the latent features to the forward diffusion process, where noise is incrementally added to the latents. In the backward diffusion process, we utilize an inflated U-Net architecture to predict the noise for the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7829" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 545, + 297 + ], + "type": "image", + "image_path": "0a79880bfbce85f15b13557ea3e5988d6d7172f34fb2a5e913e8cce5d29dfdcf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 546, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 546, + 342 + ], + "type": "text", + "content": "Figure 2. Pipeline of our Parameter-Efficient Text-to-Video Framework. We utilize the pre-trained auto-encoder as in Stable Diffusion [79] to obtain latent representation. During training, we only update the parameters of the newly added adapter module, highlighted in yellow. The parameters of other modules are frozen, highlighted in blue." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 376, + 547, + 477 + ], + "blocks": [ + { + "bbox": [ + 46, + 352, + 546, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 546, + 373 + ], + "type": "text", + "content": "Table 1. Model size and inference speed comparisons. The speed is measured in seconds on one A100 (80GB) GPU. The majority of results are sourced from [1]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 376, + 547, + 477 + ], + "lines": [ + { + "bbox": [ + 48, + 376, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 547, + 477 + ], + "type": "table", + "html": "
MethodParameters (Billion)Speed (s)
T2V CoreAuto EncoderText EncoderPrior ModelSuper ResolutionFrame InterpolationOverallTuned
CogVideo [42]7.70.10---7.715.515.5434.53
Make-A-Video [88]3.1-0.121.31.4 + 0.73.19.729.72-
Imagen Video [39]5.6-4.6-1.2 + 1.4 + 0.341.7 + 0.78 + 0.6316.2516.25-
Video LDM [8]1.510.080.12-0.981.514.202.65-
Latent-VDM [1]0.920.080.58---1.580.9228.62
Latent-Shift [1]0.880.080.58---1.530.8823.40
LVDM [35]0.960.080.12---1.161.0421.23
SimDA (Ours)0.880.080.12---1.080.02511.20
", + "image_path": "82fd11d50902bd68bd061b549257a918b6f17ec43245a98ca48fa7940a05a818.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "content": "noisy video latents. Specifically, for the Convolution block, we inflate the 2D ResNet [34] block to a 3D block to accommodate video inputs. Additionally, we incorporate a lightweight Temporal Adapter module for temporal modeling. In the Attention block, we employ a latent-shift attention mechanism for spatial self-attention and introduce two spatial adapter modules to facilitate the transfer of video information. Further details are presented in Sec. 3.3. During inference, we employ DDIM [90] sampling to progressively denoise the latent representation sampled from a standard Gaussian distribution. Finally, we utilize a pre-trained decoder " + }, + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "content": " to reconstruct the video from the denoised latent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 115, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 115, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 115, + 647 + ], + "type": "text", + "content": "3.3. Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "In this section, we describe the proposed Spatial Adapter, Temporal Adapter, and Latent-Shift Attention in detail, which are the key components of our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Spatial Adapter The large-scale text-image pre-trained T2I model exhibits significant transferability, as evidenced by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "type": "text", + "content": "its remarkable accomplishments in tasks such as personalized T2I generation [65, 81] and image editing [37, 131]. Consequently, we believe that employing a lightweight fine-tuning approach can effectively harness spatial information in the realm of video generation. Inspired by efficient fine-tuning techniques in NLP [45, 54] and vision tasks [13, 124], we adopt adapters due to their simplicity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 569, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 546, + 689 + ], + "type": "text", + "content": "In our T2V framework, we design two types of spatial adapters (i.e., Attention Adapter and FFN Adapter) for transferring video spatial information. As shown in the bottom right of Fig. 2, both adapters employ a bottleneck architecture consisting of two fully connected (FC) layers with an intermediate activation layer. The first FC layer maps the input to a lower-dimensional space, while the second FC layer maps it back to the original dimensional. Formally, for an input feature matrix " + }, + { + "bbox": [ + 304, + 569, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 304, + 569, + 546, + 689 + ], + "type": "text", + "content": ", the spatial adapter could be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\operatorname {S - A d p a t e r} (\\mathbf {X}) = \\mathbf {X} + \\mathbf {W} _ {\\text {u p}} \\left(\\operatorname {G E L U} \\left(\\mathbf {W} _ {\\text {d o w n}} (\\mathbf {X})\\right)\\right), \\tag {4}", + "image_path": "e78833d62deb0e44825e83f9cc405e05a7f8e459dfe06082be3619b32ddee4f3.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7830" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 289, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 182 + ], + "type": "image", + "image_path": "050f80f7babddd762d6b1e01d812c3c5ceb7cd1856cdb4f51bfc20fccf658aea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 288, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 231 + ], + "type": "text", + "content": "Figure 3. The overview of Temporal Latent-shift Attention module. It is noted that the Latent-shift attention is performed on latent space, but the visualization overview is shown on the image level for clear presentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{up}}" + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down}}" + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "text", + "content": " are the learnable matrix with dimension " + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "inline_equation", + "content": "d\\times l" + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "inline_equation", + "content": "l\\times d, l < d" + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "text", + "content": ". To preserve the structure of the original network and the pretrained weights, we initialize the second FC layer " + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down}}" + }, + { + "bbox": [ + 46, + 237, + 287, + 369 + ], + "type": "text", + "content": " with zeros. To adapt to the spatial features of videos, we incorporate the adapter after the latent-shift attention layer. Additionally, we observe that adding an adapter to the feed-forward network (FFN) also helps the network transfer spatial information to videos. We will provide examples in Sec. 4.4. During training, all layers of the attention block are fixed, and only the adapters are updated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 287, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 476 + ], + "type": "text", + "content": "Temporal Adapter While the spatial adapter effectively transfers spatial information to video, modeling temporal information is crucial for T2V generation tasks. Previous approaches incorporate temporal convolution [11, 85, 86] or temporal attention [8, 88, 104] modules to capture temporal relationships. Although these modules are effective in modeling temporal dynamics, they often come with a huge number of parameters and high-dimensional input feature, resulting in significant computational and training costs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 476, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 548 + ], + "type": "text", + "content": "To address this issue, we utilize the temporal adapter module for temporal modeling as [69, 128]. In contrast to conventional spatial adapter modules, the temporal adapter module employs depth-wise 3D convolution instead of an intermediate activation layer [36]. The temporal adapter could be formally written as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 557, + 286, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 557, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 50, + 557, + 286, + 581 + ], + "type": "interline_equation", + "content": "\\operatorname {T - A d p a t e r} (\\mathbf {X}) = \\mathbf {X} + \\mathbf {W} _ {\\mathrm {u p}} (3 \\mathrm {D} - \\operatorname {C o n v} (\\mathbf {W} _ {\\mathrm {d o w n}} (\\mathbf {X}))). \\tag {5}", + "image_path": "8850f752b363b69688f1b39b7958c1ddbfc936a84bf17884a68b79979f7445d9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "content": "By utilizing 3D convolutions in lower-dimensional input, our approach significantly alleviates the complexity of temporal modeling. As a result, our method achieves efficient memory usage during training and exhibits the fastest inference speed among competitive approaches, as demonstrated in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Temporal Latent-Shift Attention In the original T2I framework, the attention block of the U-Net only performs self-attention for individual frames, neglecting the information from other frames. While joint-space-time attention, as demonstrated in [2, 7, 96, 114], can effectively model tem" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 542, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 542, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 542, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 542, + 152 + ], + "type": "image", + "image_path": "e848e44034601bf1ae52de4940bee4d1035b1ee513ab20c5550eec8195d33631.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 158, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 191 + ], + "type": "text", + "content": "Figure 4. During inference, we sample a novel video from the latent noise inverted from the input video, guided by an edited prompt (e.g., \"a Porsche sports car is moving on the road\")." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "text", + "content": "poral dependencies, it introduces a quadratic complexity in terms of attention calculation. For a video with " + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "text", + "content": " tokens, the complexity of global spatial-temporal attention becomes " + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "inline_equation", + "content": "O(L^2 N^2)" + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "text", + "content": ". To address this issue, we propose a latent-shift attention module as shown in Fig. 3. In addition to considering tokens within the current frame, we further conduct a patch-level shifting operation along the temporal dimension to shift tokens from the preceding " + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 194, + 545, + 350 + ], + "type": "text", + "content": " frames onto the current frame, thereby composing a new latent feature frame. We concatenate the latent feature of the current frame with the temporally shifted latent feature, forming the keys and values for attention calculation. The latent-shift attention can be formally written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 356, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 356, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 392, + 356, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {W} _ {\\mathrm {q}} \\left(\\boldsymbol {x} _ {z _ {i}}\\right), \\tag {6}", + "image_path": "675ffd504694f2f710f49c3ff59ac52c9380b3d42a9c1d5dd0610db07cce5a42.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 372, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 372, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 377, + 372, + 545, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {K} = \\mathbf {W} _ {\\mathrm {k}} \\left[ \\boldsymbol {x} _ {z _ {i}}, \\boldsymbol {x} _ {z _ {s h i f t}} \\right], \\tag {7}", + "image_path": "59786a2adb4ab7bca58e68d2f95b3cf5a7a668ed72339739f0c025eb58476593.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 386, + 545, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 386, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 378, + 386, + 545, + 399 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\mathbf {W} _ {\\mathrm {v}} \\left[ \\boldsymbol {x} _ {z _ {i}}, \\boldsymbol {x} _ {z _ {\\text {s h i f t}}} \\right], \\tag {8}", + "image_path": "9d04c1bf57e108bc58dafdccd7ac1fc1f7aa70f8ba57bdf9ae25581a64a78840.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{z_i}" + }, + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "text", + "content": " denotes the query frame and " + }, + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "text", + "content": " means concatenate. This approach reduces the complexity of attention to " + }, + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "inline_equation", + "content": "O(2LN^2)" + }, + { + "bbox": [ + 304, + 402, + 545, + 473 + ], + "type": "text", + "content": ", significantly lowering the computational burden compared to global attention. Moreover, it allows the model to learn the relationships between adjacent frames, ensuring better temporal consistency in video generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 479, + 504, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 504, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 504, + 491 + ], + "type": "text", + "content": "3.4. Super Resolution and Editing Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "text", + "content": "Super Resolution (SR) Due to constraints of limited GPU memory and the lack of high-resolution video-text datasets, most existing methods [1, 8, 35], including ours, are only able to generate images at a resolution of " + }, + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "text", + "content": ". To overcome this limitation and generate higher-resolution outputs, we adopt a two-stage training approach similar to cascaded Diffusion Models [40]. In the first stage, we generate videos with a " + }, + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "text", + "content": " resolution using our SimDA methods. In the second stage, we employ an LDM upsampler [79] to enhance the resolution of the videos to " + }, + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 497, + 545, + 653 + ], + "type": "text", + "content": ". We incorporate noise augmentation and noise level conditioning, and train a super-resolution model using the following equation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 658, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 658, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 339, + 658, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I}), t} [ \\| \\epsilon - \\epsilon_ {\\theta} ([ \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {l o w} ], \\boldsymbol {c}, t) \\| _ {2} ^ {2} ], \\tag {9}", + "image_path": "d42ab6ab9b540fb1c71dc7f49d40c2fbfa594fe8d261c4a211450b99e90c390f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{low}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the low-resolution video, we concatenate it with " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " frame by frame following Video LDM [8]. The architecture of SR is similar to T2V model in the first stage," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7831" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 242, + 266 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 242, + 266 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 242, + 266 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 242, + 266 + ], + "type": "image", + "image_path": "187c701cb8345bbe82218a001a6f8f596ba54c3c5baaae46b165c38bb0fdbb1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 269, + 441, + 279 + ], + "lines": [ + { + "bbox": [ + 151, + 269, + 441, + 279 + ], + "spans": [ + { + "bbox": [ + 151, + 269, + 441, + 279 + ], + "type": "text", + "content": "Figure 5. Text-to-Video generation comparison on the user study evaluation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 243, + 71, + 380, + 266 + ], + "blocks": [ + { + "bbox": [ + 243, + 71, + 380, + 266 + ], + "lines": [ + { + "bbox": [ + 243, + 71, + 380, + 266 + ], + "spans": [ + { + "bbox": [ + 243, + 71, + 380, + 266 + ], + "type": "image", + "image_path": "29abc1c5a9912f55f43909fcd0118f536c739ea79705b3248e5efee5732948a2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 382, + 71, + 518, + 266 + ], + "blocks": [ + { + "bbox": [ + 382, + 71, + 518, + 266 + ], + "lines": [ + { + "bbox": [ + 382, + 71, + 518, + 266 + ], + "spans": [ + { + "bbox": [ + 382, + 71, + 518, + 266 + ], + "type": "image", + "image_path": "6876e3673f27beef49514899611a8b0df912bb716ca87fab64f3fc3d09dddb3d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 64, + 305, + 526, + 488 + ], + "blocks": [ + { + "bbox": [ + 46, + 282, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 547, + 304 + ], + "type": "text", + "content": "Table 2. Text-to-Video generation comparison on MSR-VTT [118] dataset. We report the Fréchet Video Distance (FVD) scores and CLIPSIM scores." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 305, + 526, + 488 + ], + "lines": [ + { + "bbox": [ + 64, + 305, + 526, + 488 + ], + "spans": [ + { + "bbox": [ + 64, + 305, + 526, + 488 + ], + "type": "table", + "html": "
MethodTraining DataResolutionZero-shotParams(B)FVD(↓)CLIPSIM(↑)
GODIVA [108]MSR-VTT128x128No--0.2402
NüWA [109]MSR-VTT128x128No0.87-0.2439
Make-A-Video [88]WebVid-10M + HD-VILA-10M256x256Yes9.72-0.3049
VideoFactory [104]WebVid-10M + HD-VG-130M256x256Yes2.04-0.3005
LVDM [35]WebVid-2M256x256Yes1.167420.2381
MMVG [25]WebVid-2.5M256x256Yes--0.2644
CogVideo [42]WebVid-5.4M256x256Yes15.512940.2631
ED-T2V [57]WebVid-10M256x256Yes1.30-0.2763
MagicVideo [136]WebVid-10M256x256Yes-998-
Video-LDM [8]WebVid-10M256x256Yes4.20-0.2929
VideoComposer [105]WebVid-10M256x256Yes1.855800.2932
Latent-Shift [1]WebVid-10M256x256Yes1.53-0.2773
VideoFusion [61]WebVid-10M256x256Yes1.835810.2795
SimDA (Ours)WebVid-10M256x256Yes1.084560.2945
", + "image_path": "c1341273cc81340b790ccfdeb437a5bc9ab7154ed89d02fda54a8331742435b9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 543 + ], + "type": "text", + "content": "we change the original U-Net block by adding Spatial and Temporal Adapters as described in Sec. 3.3 and only finetune the newly added modules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "Text-guided Video Editing In addition to performing T2V generation, our method could turn into one-shot tuning for text-guided video editing following Tune-A-Video [110]. The training pipeline of editing model is the same as our T2V method. However, for the inference stage, we adopt the DDIM inversion latents instead of random noisy latents together with edited prompt for novel video generation as shown in Fig. 4. By doing so, the pixel-level information control could remain in the inversion latent as demonstrated in [110]. Owing to the light-weight module and efficient pipeline of our method, SimDA needs fewer training steps (200 steps compared to 500 steps) and thus the training time and inference time is much faster than Tune-A-Video [110]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 506, + 388, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 388, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 388, + 519 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 526, + 440, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 440, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 440, + 539 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Our T2V method is composed of two-stage models. The first model predicts video frames with a resolution " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " (with a latent size of " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "), while the second model is a " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " upsampler, producing a resolution of " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ". We train the general T2V model on WebVid-10M [5] dataset following [1, 8]. We follow previous methods [1, 104, 136] to report the CLIP score [75] and FVD (Frechet Video Distance) score [99] on MSR-VTT [118]. Besides, we compare the FVD score and CLIP score on the evaluation set of WebVid [5] as in VideoFactory [104]. We also compare the parameter scale and inference speed of our method with some open-sourced methods [1, 35, 61]. Finally, we also provide a user study between our work and VDM [41], Latent-shift [1], Video-Fusion [61] and LVDM [35]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7832" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 257, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 257, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 257, + 85 + ], + "type": "text", + "content": "4.2. Evaluation on Text-to-Video Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 286, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 286, + 139 + ], + "type": "text", + "content": "To fully evaluate the generation performance of our SimDA, we conduct automatic evaluations on two distinct datasets: WebVid [5] (Val), which shares the same domain as the training data and MSR-VTT [118] in a zero-shot setting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 140, + 286, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 140, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 286, + 272 + ], + "type": "text", + "content": "Evaluation on MSR-VTT As shown in Table 2, we evaluate CLIPSIM [75] and FVD [99] on the widely used video generation benchmarks, MSR-VTT [118]. We randomly select one text prompt per example from MSR-VTT [118] and generate a total of 2,990 videos. Despite being a zero-shot setting, our method achieves an average CLIPSIM of 0.2945 that surpasses most of the competitors, indicating a strong semantic alignment between the generated videos and the input text. Though Make-A-Video [88] and VideoFactory [104] offer higher CLIP scores, they utilize additional large-scale HD-VILA [119] datasets for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 275, + 286, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 286, + 418 + ], + "type": "text", + "content": "Evaluation on WebVid As shown in Table 3, we create a validation set consisting of 4,476 randomly extracted text-video pairs from WebVid-10M. These pairs are not included in the training data following [104]. We conduct evaluations on this validation set and obtain impressive results. Our method achieves an FVD score of 363.98 and a CLIP-SIM score of 0.3054. These scores are significantly higher than those achieved by existing methods such as ModelScope [103] and LVDM [35]. Besides, our method shows competitive results compared to VideoFactory [104] which is trained with much larger datasets. These results clearly demonstrate the superiority of our approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 421, + 286, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 286, + 528 + ], + "type": "text", + "content": "Human Evaluation In order to address the limitations of existing metrics and assess the performance of our SimDA from a human perspective, we conduct an extensive user study. The study involves comparing our method with four state-of-the-art methods. Specifically, we select two publicly available models, namely ModelScope [103] and LVDM [35]. Additionally, we consider two methods with similar scale parameters, VDM [41] and Latent-shift [1], which only showcase some samples on their websites." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 529, + 286, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 286, + 636 + ], + "type": "text", + "content": "For each case, participants were provided with two video samples, one is generated by our method and the other is from a competitor. They were then asked to compare the two samples in terms of video quality and text-video similarity. To ensure fairness in the comparisons, we also report the ratio of network parameters compared to ours. The results, along with the parameter ratios, are presented in Table 4. The user study approach allows us to gain in-depth insights into the subjective evaluation of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 639, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 286, + 686 + ], + "type": "text", + "content": "Qualitative Results The visualization of T2V generation results are shown in Fig. 1(a). Besides, we show the comparison results in Fig. 5. More examples can be found in our supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "Parameter Size and Inference Speed We conduct a comparison of the number of parameters and inference speed," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 95, + 541, + 170 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 3. Text-to-video generation on the validation set of WebVid [5]. We report the FVD and CLIPSIM scores." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 95, + 541, + 170 + ], + "lines": [ + { + "bbox": [ + 310, + 95, + 541, + 170 + ], + "spans": [ + { + "bbox": [ + 310, + 95, + 541, + 170 + ], + "type": "table", + "html": "
MethodParams(B)FVD(↓)CLIPSIM(↑)
LVDM [35]1.16455.530.2751
ModelScope [103]1.83414.110.3000
VideoFactory [104]2.04292.350.3070
SimDA (Ours)1.08363.980.3054
", + "image_path": "6d1a0e71ddc0c4c763a0204953cf74cf8a0aa78275bf141796247594ed2d50d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 221, + 553, + 293 + ], + "blocks": [ + { + "bbox": [ + 305, + 177, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 177, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 177, + 545, + 220 + ], + "type": "text", + "content": "Table 4. User preference is depicted as a percentage indicating the proportion of individuals favoring our method over the compared approach. Param Ratio means the ratio of the network parameter v.s. Ours." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 221, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 308, + 221, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 221, + 553, + 293 + ], + "type": "table", + "html": "
SampleMethodParam RatioQualityFaithfulness
Open WebsiteVDM [41]0.83×85.2%81.4%
Latent-Shift [1]1.41×81.5%79.3%
Pretrained ModelModelScope [103]1.69×78.3%79.5%
LVDM [35]1.07×83.4%84.7%
", + "image_path": "0e7f65041e7ea81639f3c773e75d60e502a7346e09238311be35aca67e826fb6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 301, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 410 + ], + "type": "text", + "content": "and the results are presented in Table 1. For the speed comparison, we select CogVideo [42], Latent-Shift [1] and LVDM [35]. SimDA, on the other hand, stands out as it is significantly smaller than previous works and exhibits faster inference speed compared to other methods. Despite having fewer parameters, SimDA achieves superior performance in various benchmarks when compared to other methods. This validation further highlights our advantages in terms of model efficiency and performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 418, + 520, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 520, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 520, + 431 + ], + "type": "text", + "content": "4.3. Evaluation on Text-guided Video Editing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 485 + ], + "type": "text", + "content": "Following the methodology of previous studies [110], we employ CLIP score [75] and a user study to evaluate the performance of different methods in terms of frame consistency and textual alignment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 617 + ], + "type": "text", + "content": "First, we calculate the CLIP image embedding for all frames in the edited videos to measure frame consistency. The average cosine similarity between pairs of video frames is reported. Additionally, to assess textual faithfulness, we compute the average CLIP score between frames of the output videos and the corresponding edited prompts. A total of 15 videos from the dataset [72] were selected and edited based on object, background and style, resulting in 75 edited videos for each model. The average results, presented in Table 5, highlight our method's exceptional ability to achieve semantic alignment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 618, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 712 + ], + "type": "text", + "content": "Secondly, we conduct a user study involving videos and text prompts. Participants were asked to vote for the edited videos that exhibited the best temporal consistency and those most accurately matched the textual description. Table 5 demonstrates that our method, SimDA, receives the highest number of votes in both aspects, indicating superior editing quality and a strong preference from users in practical scenarios." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7833" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 96, + 541, + 152 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 5. Quantitative comparison with evaluated baseline [110]. The \"Training\" refers to the process of optimization, and \"Memory\" refers to the GPU memory." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 96, + 541, + 152 + ], + "lines": [ + { + "bbox": [ + 52, + 96, + 541, + 152 + ], + "spans": [ + { + "bbox": [ + 52, + 96, + 541, + 152 + ], + "type": "table", + "html": "
MethodFrame consistencyTextual alignmentRuntime [min]Memory [Gib]Params [Mb]
CLIP Score↑User Vote↑CLIP Score↑User Vote↑Training↓Inference↓Training↓Inference↓Tuned↓
Tune-A-Video [110]94.131.2%31.839.5%9.30.831.311.474.4
SimDA(Ours)94.968.8%31.960.5%2.50.428.68.824.9
", + "image_path": "1be6338bb456d98ffb5e7c4cf890fba0eccf9c81f2b71b55bc1941951d0a1b23.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 163, + 273, + 452 + ], + "blocks": [ + { + "bbox": [ + 61, + 163, + 273, + 452 + ], + "lines": [ + { + "bbox": [ + 61, + 163, + 273, + 452 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 273, + 452 + ], + "type": "image", + "image_path": "dd079eb50d409e473d462a68bbb8b1ed5ac4b844647e7953f339b5058e968f17.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 455, + 288, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 455, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 288, + 489 + ], + "type": "text", + "content": "Figure 6. Ablation of T2V generation results. TA, SA, AA, FA and LSA refer to Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-Shift Attention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 496, + 142, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 142, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 142, + 510 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "text", + "content": "Here we discuss the effect of each module. We perform experiments on " + }, + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "inline_equation", + "content": "1K" + }, + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "text", + "content": " samples from the WebVid validation set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "type": "text", + "content": "Effect of Temporal Adapter Temporal modeling is a crucial component of video generation. In our video editing task, when compared to methods that rely on temporal attention modeling like Tune-A-Video [110], we observe that our temporal adapter is more lightweight and achieves superior editing results as in Table 5. Additionally, we conduct ablation experiments, as shown in Table 6 and Fig. 6, where the lack of Temporal Adapter (TA) results in significantly higher FVD and chaotic temporal sequences in the generated videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Effect of Spatial Adapter We also validate the effectiveness of the Spatial Adapter (SA) in transferring spatial knowledge of videos. As shown in Table 6, without the Attention Adapter (AA) and FFN Adapter (FA), the model's" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 225, + 548, + 321 + ], + "blocks": [ + { + "bbox": [ + 305, + 164, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 164, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 547, + 220 + ], + "type": "text", + "content": "Table 6. Ablation study on different modules. We report the FVD [99] and CLIPSIM [75] on " + }, + { + "bbox": [ + 305, + 164, + 547, + 220 + ], + "type": "inline_equation", + "content": "1K" + }, + { + "bbox": [ + 305, + 164, + 547, + 220 + ], + "type": "text", + "content": " samples from the validation set of WebVid-10M [5]. TA, SA, AA, FA and LSA represent Temporal Adapter, Spatial Adapter, Attention Adapter, FFN Adapter and Latent-shift Attention, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 225, + 548, + 321 + ], + "lines": [ + { + "bbox": [ + 308, + 225, + 548, + 321 + ], + "spans": [ + { + "bbox": [ + 308, + 225, + 548, + 321 + ], + "type": "table", + "html": "
TAAAFALSAFVD(↓)CLIPSIM(↑)
w/o TA✓✓✓1470.10.2629
w/o SA✓✓811.30.2822
w/o AA✓✓✓764.80.2851
w/o FA✓✓✓623.70.2962
w/o LSA✓✓✓618.20.3011
Ours✓✓✓✓530.20.3034
", + "image_path": "3f66cc8eec8f34997bd22341b227b2dffb7476438295b0ae0257165bc546e1ff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 330, + 547, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 547, + 379 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 547, + 379 + ], + "type": "text", + "content": "FVD and CLIPSIM scores for generated videos will become worse. Additionally, it can be observed from the Fig. 6 that the model exhibits misconceptions in understanding the text prompt without the spatial adapter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 382, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 382, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 547, + 479 + ], + "type": "text", + "content": "Effect of Latent Shift Attention To investigate the impact of Latent-shift Attention (LSA), we replace it with regular single-frame spatial attention. Besides observing a decline in FVD and text alignment CLIPSIM scores in Table 6, we also test the CLIPSIM of each frame within the same video, which decreased from 96.4 to 94.5. This demonstrates that our LSA module can effectively model the relationship of adjacent frames, leading to more consistent videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 487, + 379, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 379, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 379, + 499 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 640 + ], + "type": "text", + "content": "In this paper, we proposed SimDA, a parameter-efficient video diffusion model for text-guided video generation and editing. With the proposed light-weight spatial and temporal adapters, our method not only transferred from powerful spatial information but also modeled temporal relationships with the least new parameters. The experimental results demonstrated that our approach has the fastest training and inference speed while maintaining competitive generation and editing results. Our work is the first parameter-efficient video diffusion method serving as an efficient T2V fine-tuning baseline and paved the way for future research." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Acknowledge This work was supported by National Science and Technology Major Project (No. 2021ZD0112805) and in part by National Natural Science Foundation of China (No. 62032006 and No. 62102092.). The computations in this research were performed using the CFFFplatform of Fudan University." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7834" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, JiaBin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 147, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 147, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 147, + 287, + 178 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 180, + 287, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 180, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 58, + 180, + 287, + 221 + ], + "type": "text", + "content": "[3] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. arXiv preprint arXiv:1710.11252, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 224, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 224, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 287, + 266 + ], + "type": "text", + "content": "[4] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 270, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 270, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 270, + 287, + 311 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 313, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 313, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 313, + 287, + 346 + ], + "type": "text", + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 347, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 347, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 347, + 287, + 378 + ], + "type": "text", + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 380, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 380, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 58, + 380, + 287, + 434 + ], + "type": "text", + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 58, + 436, + 287, + 479 + ], + "type": "text", + "content": "[9] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 512 + ], + "type": "text", + "content": "[10] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 514, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 545 + ], + "type": "text", + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 547, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 547, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 547, + 287, + 578 + ], + "type": "text", + "content": "[12] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In ICCV, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 287, + 623 + ], + "type": "text", + "content": "[13] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "type": "text", + "content": "[14] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 670, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 691 + ], + "type": "text", + "content": "[15] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In ICML, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[16] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 331, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 74, + 545, + 95 + ], + "type": "text", + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 98, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 98, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 312, + 98, + 545, + 118 + ], + "type": "text", + "content": "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 121, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 121, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 312, + 121, + 545, + 152 + ], + "type": "text", + "content": "[18] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 156, + 545, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 156, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 312, + 156, + 545, + 210 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 213, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 213, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 312, + 213, + 545, + 244 + ], + "type": "text", + "content": "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 247, + 545, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 247, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 312, + 247, + 545, + 291 + ], + "type": "text", + "content": "[21] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 293, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 293, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 545, + 335 + ], + "type": "text", + "content": "[22] Qijun Feng, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Fdgaussian: Fast gaussian splatting from single image via geometric-aware diffusion model. arXiv preprint arXiv:2403.10242, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 392 + ], + "type": "text", + "content": "[23] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 395, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 395, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 312, + 395, + 545, + 427 + ], + "type": "text", + "content": "[24] Jean-Yves Franceschi, Edouard Delasalles, Mickael Chen, Sylvain Lamprier, and Patrick Gallinari. Stochastic latent residual video prediction. In ICML, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 430, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 430, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 312, + 430, + 545, + 483 + ], + "type": "text", + "content": "[25] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In CVPR, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 487, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 487, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 312, + 487, + 545, + 530 + ], + "type": "text", + "content": "[26] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 533, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 545, + 575 + ], + "type": "text", + "content": "[27] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 578, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 578, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 312, + 578, + 545, + 632 + ], + "type": "text", + "content": "[28] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 677 + ], + "type": "text", + "content": "[29] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 680, + 545, + 712 + ], + "type": "text", + "content": "[30] Sonam Gupta, Arti Keshari, and Sukhendu Das. Rvgan: Recurrent gan for unconditional video generation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7835" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[31] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 148 + ], + "type": "text", + "content": "[32] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "text", + "content": "[33] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. arXiv preprint arXiv:2110.04366, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "text", + "content": "[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[35] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 294 + ], + "type": "text", + "content": "[36] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 294, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 336 + ], + "type": "text", + "content": "[37] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. *ICLR*, 2023. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 338, + 288, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 359 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 359 + ], + "type": "text", + "content": "[38] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 360, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 288, + 415 + ], + "type": "text", + "content": "[39] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 288, + 458 + ], + "type": "text", + "content": "[40] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 460, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 288, + 502 + ], + "type": "text", + "content": "[41] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 503, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 288, + 547 + ], + "type": "text", + "content": "[42] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 548, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 590 + ], + "type": "text", + "content": "[43] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022.2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 592, + 288, + 635 + ], + "type": "text", + "content": "[44] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 636, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 288, + 680 + ], + "type": "text", + "content": "[45] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "text", + "content": "[46] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[47] Shibo Jie and Zhi-Hong Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 139 + ], + "type": "text", + "content": "[48] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 140, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 173 + ], + "type": "text", + "content": "[49] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 207 + ], + "type": "text", + "content": "[50] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 208, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 208, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 312, + 208, + 545, + 241 + ], + "type": "text", + "content": "[51] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 242, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 242, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 312, + 242, + 545, + 284 + ], + "type": "text", + "content": "[52] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 545, + 319 + ], + "type": "text", + "content": "[53] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 320, + 545, + 353 + ], + "type": "text", + "content": "[54] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 386 + ], + "type": "text", + "content": "[55] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 388, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 545, + 410 + ], + "type": "text", + "content": "[56] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "type": "text", + "content": "[57] Jiawei Liu, Weining Wang, Wei Liu, Qian He, and Jing Liu. Ed-t2v: An efficient training framework for diffusion-based text-to-video generation. In IJCNN, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "type": "text", + "content": "[58] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 478, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 511 + ], + "type": "text", + "content": "[59] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. arXiv preprint arXiv:2303.04761, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 512, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 512, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 312, + 512, + 545, + 566 + ], + "type": "text", + "content": "[60] Pauline Luc, Aidan Clark, Sander Dieleman, Diego de Las Casas, Yotam Doron, Albin Cassirer, and Karen Simonyan. Transformation-based adversarial video prediction on large-scale data. arXiv preprint arXiv:2003.04035, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 567, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 567, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 312, + 567, + 545, + 612 + ], + "type": "text", + "content": "[61] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 545, + 644 + ], + "type": "text", + "content": "[62] Tanya Marwah, Gaurav Mittal, and Vineeth N Balasubramanian. Attentive semantic video generation using captions. In ICCV, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 545, + 689 + ], + "type": "text", + "content": "[63] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. ICLR, 2022. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 690, + 545, + 714 + ], + "type": "text", + "content": "[64] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync-draw: Automatic video generation using" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7836" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 286, + 94 + ], + "type": "text", + "content": "deep recurrent attentive architectures. In ACM Multimedia, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 150 + ], + "type": "text", + "content": "[65] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "type": "text", + "content": "[66] Andres Munoz, Mohammadreza Zolfaghari, Max Argus, and Thomas Brox. Temporal shift gan for large scale video generation. In WACV, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 240 + ], + "type": "text", + "content": "[67] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 241, + 288, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 264 + ], + "type": "text", + "content": "[68] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "text", + "content": "[69] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. NeurIPS, 2022. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 309, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 309, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 309, + 288, + 342 + ], + "type": "text", + "content": "[70] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 344, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 288, + 376 + ], + "type": "text", + "content": "[71] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize, diagnose, and optimize: Towards fine-grained vision-language understanding. In CVPR, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 378, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 378, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 378, + 288, + 422 + ], + "type": "text", + "content": "[72] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 288, + 465 + ], + "type": "text", + "content": "[73] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 510 + ], + "type": "text", + "content": "[74] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 512, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 288, + 567 + ], + "type": "text", + "content": "[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 568, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 568, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 568, + 288, + 622 + ], + "type": "text", + "content": "[76] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 624, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 288, + 667 + ], + "type": "text", + "content": "[77] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "text", + "content": "[78] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 150 + ], + "type": "text", + "content": "[80] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "type": "text", + "content": "[81] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 197, + 545, + 251 + ], + "type": "text", + "content": "[82] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 312, + 253, + 545, + 296 + ], + "type": "text", + "content": "[83] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan. IJCV, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 297, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 297, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 545, + 330 + ], + "type": "text", + "content": "[84] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 331, + 545, + 364 + ], + "type": "text", + "content": "[85] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 312, + 365, + 545, + 398 + ], + "type": "text", + "content": "[86] Baifeng Shi, Qi Dai, Judy Hoffman, Kate Saenko, Trevor Darrell, and Huijuan Xu. Temporal action detection with multi-level supervision. In ICCV, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 545, + 432 + ], + "type": "text", + "content": "[87] Chaehun Shin, Heeseung Kim, Che Hyun Lee, Sang-gil Lee, and Sungroh Yoon. Edit-a-video: Single video editing with object-aware consistency. 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 312, + 433, + 545, + 476 + ], + "type": "text", + "content": "[88] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. *ICLR*, 2023. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 478, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 520 + ], + "type": "text", + "content": "[89] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 545, + 555 + ], + "type": "text", + "content": "[90] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "type": "text", + "content": "[91] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "type": "text", + "content": "[92] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 647, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 669 + ], + "type": "text", + "content": "[93] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "text", + "content": "[94] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7837" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 105 + ], + "type": "text", + "content": "[95] Shuyuan Tu, Tianzhen Guan, and Li Kuang. Multiple biological granularities network for person re-identification. In ICMR, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 287, + 139 + ], + "type": "text", + "content": "[96] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. In ICCV, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 286, + 182 + ], + "type": "text", + "content": "[97] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 183, + 286, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 286, + 216 + ], + "type": "text", + "content": "[98] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 286, + 270 + ], + "type": "text", + "content": "[99] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "type": "text", + "content": "[100] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "text", + "content": "[101] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. Masked conditional video diffusion for prediction, generation, and interpolation. arXiv preprint arXiv:2205.09853, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 381 + ], + "type": "text", + "content": "[102] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. NeurIPS, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "text", + "content": "[103] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "type": "text", + "content": "[104] Wenjing Wang, Huan Yang, Zixi Tuo, Huiguo He, Junchen Zhu, Jianlong Fu, and Jiaying Liu. Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv preprint arXiv:2305.10874, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[105] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "type": "text", + "content": "[106] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, Kai Qiu, Yuhui Yuan, Chuanxin Tang, Xiaoyan Sun, Chong Luo, and Baining Guo. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[107] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[108] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[109] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 126 + ], + "type": "text", + "content": "[110] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[111] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In ECCV,2022.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[112] Zhen Xing, Yijiang Chen, Zhixin Ling, Xiangdong Zhou, and Yu Xiang. Few-shot single-view 3d reconstruction with memory prior contrastive network. In ECCV, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "text", + "content": "[113] Zhen Xing, Hengduo Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised single-view 3d reconstruction via prototype shape priors. In ECCV, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 244, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 277 + ], + "type": "text", + "content": "[114] Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. In CVPR, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 278, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 278, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 278, + 545, + 323 + ], + "type": "text", + "content": "[115] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 324, + 545, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 367 + ], + "type": "text", + "content": "[116] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 370, + 545, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 370, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 545, + 412 + ], + "type": "text", + "content": "[117] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 415, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 448 + ], + "type": "text", + "content": "[118] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 450, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 450, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 545, + 494 + ], + "type": "text", + "content": "[119] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 545, + 540 + ], + "type": "text", + "content": "[120] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 541, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 541, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 541, + 545, + 574 + ], + "type": "text", + "content": "[121] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "type": "text", + "content": "[122] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "text", + "content": "[123] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 656, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 689 + ], + "type": "text", + "content": "[124] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. ICLR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[125] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku," + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7838" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 621 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 106 + ], + "type": "text", + "content": "Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 150 + ], + "type": "text", + "content": "[126] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "text", + "content": "[127] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 240 + ], + "type": "text", + "content": "[128] Bowen Zhang, Xiaojie Jin, Weibo Gong, Kai Xu, Zhao Zhang, Peng Wang, Xiaohui Shen, and Jiashi Feng. Multimodal video adapter for parameter efficient video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "text", + "content": "[129] Hao Zhang, Yanbin Hao, and Chong-Wah Ngo. Token shift transformer for video classification. In ACM Multimedia, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "type": "text", + "content": "[130] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 341 + ], + "type": "text", + "content": "[131] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 342, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 286, + 396 + ], + "type": "text", + "content": "[132] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 286, + 441 + ], + "type": "text", + "content": "[133] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In ICLR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 444, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 444, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 444, + 286, + 487 + ], + "type": "text", + "content": "[134] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. arXiv preprint arXiv:2305.17431, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 488, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 488, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 49, + 488, + 286, + 531 + ], + "type": "text", + "content": "[135] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 533, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 286, + 576 + ], + "type": "text", + "content": "[136] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 578, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 286, + 621 + ], + "type": "text", + "content": "[137] Rui Zhu, Yingwei Pan, Yehao Li, Ting Yao, Zhenglong Sun, Tao Mei, and Chang Wen Chen. Sd-dit: Unleashing the power of self-supervised discrimination in diffusion transformer. In CVPR, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7839" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_content_list.json b/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c3a68f6fc0961307de7e5f2adf6be11fbdf99bc8 --- /dev/null +++ b/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_content_list.json @@ -0,0 +1,1607 @@ +[ + { + "type": "text", + "text": "Simple Semantic-Aided Few-Shot Learning", + "text_level": 1, + "bbox": [ + 264, + 130, + 702, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hai Zhang $^{1*}$ Junzhe Xu $^{1,2*‡}$ Shanlin Jiang $^{3}$ Zhenan He $^{1\\dagger}$ \\\n $^{1}$ College of Computer Science, Sichuan University\\\n $^{2}$ Banma Network Technology, Alibaba Group\\\n $^{3}$ Naveen Jindal School of Management, University of Texas at Dallas\\\nzhanghi@stu.scu.edu.cn, xujunzhe.xjz@alibaba-inc.com\\\nshanlin.jiang@utdallas.edu, zhenan@scu.edu.cn", + "bbox": [ + 212, + 178, + 756, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning from a limited amount of data, namely Few-Shot Learning, stands out as a challenging computer vision task. Several works exploit semantics and design complicated semantic fusion mechanisms to compensate for rare representative features within restricted data. However, relying on naive semantics such as class names introduces biases due to their brevity, while acquiring extensive semantics from external knowledge takes a huge time and effort. This limitation severely constrains the potential of semantics in Few-Shot Learning. In this paper, we design an automatic way called Semantic Evolution to generate high-quality semantics. The incorporation of high-quality semantics alleviates the need for complex network structures and learning algorithms used in previous works. Hence, we employ a simple two-layer network termed Semantic Alignment Network to transform semantics and visual features into robust class prototypes with rich discriminative features for few-shot classification. The experimental results show our framework outperforms all previous methods on six benchmarks, demonstrating a simple network with high-quality semantics can beat intricate multi-modal modules on few-shot classification tasks. Code is available at https://github.com/zhangdoudou123/SemFew.", + "bbox": [ + 75, + 354, + 472, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models have witnessed substantial advancements by leveraging extensive annotated data [8]. Nevertheless, in numerous real-world scenarios, the shortage of labeled data restricts the applicability of conventional deep learning approaches. In contrast, humans possess a remarkable cognitive ability to learn new concepts and recognize", + "bbox": [ + 75, + 772, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "categories from just a few examples [3]. Motivated by this, Few-Shot Learning (FSL) [46] is proposed to mimic the ability of humans to learn from a few labeled samples.", + "bbox": [ + 500, + 323, + 890, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In a typical setting of FSL [46], the support set, which consists of $N$ novel classes and each class contains $K$ samples, is provided for the model. The model learned on the support set is required to accurately classify test samples, which are labeled as the query set, into one of the $N$ categories. A conventional classification approach involves projecting both support and query sets onto a pre-established metric space. Then, query images are classified by finding the nearest support image. However, few support samples may not contain sufficient discriminative features for recognition. These samples are typically located on the periphery of the sample cluster in the metric space, resulting in unstable distance evaluations and inaccurate classification.", + "bbox": [ + 496, + 369, + 892, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, researchers start from visual-based methods [7, 11, 13, 16, 19, 29, 32, 37, 46, 49, 53, 60], aiming at extracting class-related features from periphery image, reducing the intra-class variation among samples, and thus constructing robust features for classification. Although visual-based methods have undergone extensive investigation and achieved great success, this type of approach still struggles to handle situations when there is only one hard periphery sample with minimal semantic features per class, i.e., the one-shot learning task [25, 53]. Hence, some researchers turned to applying semantics as auxiliary information to help the model better understand periphery images, leveraging the synergies between language and vision modalities.", + "bbox": [ + 496, + 566, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From this perspective, semantic-based methods [5, 28, 38, 44, 50, 56-59, 61] dedicated to exploring different types of semantics to enhance periphery samples. One kind of method [57-59] applies the class name as the semantic source. However, class name is not the best way to understand a novel class. For instance, If a person never saw a zebra, it would be easier to identify a zebra by the definition, i.e., a horse with stripes, than by the name zebra.", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. † Corresponding Author. \n‡ Work done during master's degree at Sichuan University.", + "bbox": [ + 93, + 875, + 403, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28588", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The American robin is renowned for its striking appearance, with a rust-red to orange breast and abdomen, a dark grayish-blue upper body, and a white eye ring. This bird plays a pivotal ecological role as a seed disperser, primarily for fruits and berries,...", + "bbox": [ + 83, + 95, + 464, + 135 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/71f35d14f53a0d0f7579004a8d0e8a8c1906588581f2cf88234f41ca965cb527.jpg", + "image_caption": [ + "Figure 1. The introduction of how high-quality semantics reconstruct the class prototype through complementarity between modalities. Periphery image is an image with fewer discriminative features and prototype image represents images with concrete and enriched representative features." + ], + "image_footnote": [], + "bbox": [ + 78, + 154, + 468, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, the class name contains ambiguity in some scenarios. For example, the class `ear` is easily misidentified as the human organ, but images in MiniImageNet [53] belong to this class are actually photos of corn. Another kind of semantic exploits discrete attribute labels as the substitution of mere class names and achieves good performance [61]. Nevertheless, the collection process is time-consuming and expert assistance requiring. Moreover, discrete attribute labels lack robust representation and struggle to benefit from context. Therefore, effectively collecting and leveraging high-quality semantics becomes an urgent research need.", + "bbox": [ + 75, + 371, + 468, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple framework called Semantic-Aided Few-Shot Learning (SemFew). To begin with, we identify the vital role of high-quality semantics in FSL and innovatively propose an automatic step-by-step Semantic Evolution (SemEvo) process to acquire detailed and accurate semantics. It first converts the class name into a short description that matches the image content of each class. Next, to further enrich recognizable characteristics, we expand and paraphrase short descriptions. The paraphrased descriptions contain more class-related knowledge than class names and short descriptions, making them better encompass a vast array of visual features. To illustrate, Fig. 1 showcases an instance that the paraphrased semantic contains a large number of detailed descriptions matching visual features e.g. organ characteristics and eating habitat. With this information, the model can better reconstruct an image with periphery feature (left image), into an image with prototype feature (right image), which benefits few-shot classification. Then, to fully exploit the advantage of paraphrased semantics, we design a network called Semantic Alignment Network (SemAlign). Different from previous methods, SemAlign does not adopt complex semantic understanding modules but utilizes a basic two-layer network. After accepting semantics and visual features", + "bbox": [ + 75, + 537, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as input, it reconstructs prototype features for subsequent classification. Experimental results on six widely adopted benchmarks demonstrate the effectiveness of our method, providing a simple insight into future FSL research. Our contributions can be summarized as three points:", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to consider the automatic way of collecting high-quality semantics and applying them in FSL.", + "- We design a simple and efficient way to translate high-quality semantics and visual features into prototypes, without any intricate semantic understanding modules.", + "- Our approach achieves state-of-the-art performance across six benchmarks in FSL research, underscoring that a basic network can obtain excellent performance when supported by high-quality semantics." + ], + "bbox": [ + 496, + 167, + 890, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 330, + 640, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The significant issue in FSL is how to effectively and accurately extract class-related features from limited data. Targeting this issue, existing research explored two different types of approaches, i.e., visual-based methods and semantic-based methods.", + "bbox": [ + 496, + 356, + 890, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual-Based Methods focus on extracting class-related features from images for classification. There are two main types of this approach. The approaches with the optimization principle behind [1, 13, 40] concentrated on acquiring a set of initial model parameters capable of rapid adaptation to novel classes. However, updating the entire model with a limited amount of labeled data may result in metaoverfitting [12, 22, 35, 42, 65]. The approaches considered metric design aiming at training a metric space, where inter-class distances are maximized, while intra-class distances are minimized. Within this type of approach, one attracting method focuses on integrating self-supervised loss to pretrain the backbone network [10, 15, 19, 32, 49]. Other methods are dedicated to improving the metric, and previous research spanning from initially manually designed metrics [2, 25, 46, 62] to subsequently model-based metrics [48, 53], all of which have yielded encouraging outcomes.", + "bbox": [ + 496, + 431, + 890, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic-Based Methods sought to enhance visual recognition performance by fusing complementary information from both visual and textual modalities [5, 28, 57, 58, 61], and related research introducing intricate network frameworks to effectively utilize textual information. Chen et al. [57] introduced an adaptive fusion mechanism to integrate visual prototypes with semantic prototypes obtained through word embeddings of class labels. Zhang et al. [61] designed a framework for prototype completion, leveraging rare attribute-level information from labels to construct representative class prototypes. Xu and Le [58] proposed the use of the Conditional Variational Autoencoder (CVAE) [47] model to generate visual features based on semantic embeddings for further classification. Li et al. [28] pre", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28589", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sented a task-relevant adaptive margin loss based on the semantic similarity between class names. Chen et al. [5] devised complementary mechanisms for inserting semantic prompts into the feature extractor, allowing the feature extractor to better focus on class-specific features.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous research emphasized the exploitation of class names and designed a complicated network to integrate semantics with visual information. In contrast, our primary concern lies in the inherent quality of semantics, and how it can relieve the burden of designing intricate network structures and learning algorithms. Remarkably, we exhibit that high-quality semantics information has the ability to yield superior outcomes even within a simple network.", + "bbox": [ + 75, + 167, + 470, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 297, + 210, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section first revisits the background of FSL, then introduces two components of SemFew: explaining how semantics are gradually improved through SemEvo, and demonstrating the details of our proposed SemAlign.", + "bbox": [ + 75, + 323, + 468, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 76, + 391, + 209, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The dataset of FSL includes two components: a base set $\\mathcal{D}_{base} = \\{(x,y)|x\\in \\mathcal{X}_{base},y\\in \\mathcal{C}_{base}\\}$ for training a metric space and a novel set $\\mathcal{D}_{novel} = \\{(x,y)|x\\in \\mathcal{X}_{novel},y\\in \\mathcal{C}_{novel}\\}$ for testing, where $x$ denotes the image and $y$ represents the label. It is important to note that the label space for both sets is disjoint, i.e., $\\mathcal{C}_{base}\\cap \\mathcal{C}_{novel} = \\emptyset$ . During the testing process, the support set $S = \\{(x_i,y_i)\\}_{i = 0}^{N\\times K}$ is randomly sampled from $\\mathcal{D}_{novel}$ , which contains $N$ classes with each class consisting of $K$ samples. Then, the model is required to correctly classify images in the query set $\\mathcal{Q} = \\{(x_i,y_i)\\}_{i = 0}^{N\\times M}$ into one of $N$ classes in the support set $S$ , where $M$ is the number of query samples in each class. Generally, this type of classification task is named as $N$ -way $K$ -shot task.", + "bbox": [ + 75, + 414, + 468, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A simple but effective FSL method is to calculate the prototype of each support class, and then classify query samples by finding the nearest prototype in the metric space. To be specific, the prototype of class $t$ , labeled as $p_t$ , is the mean vector of all samples in the support set belonging to class $t$ :", + "bbox": [ + 75, + 626, + 468, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} = \\frac {1}{\\left| \\mathcal {S} _ {t} \\right|} \\sum_ {x _ {i} \\in \\mathcal {S} _ {t}} f \\left(x _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 712, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S_{t}$ is a set of all samples belonging to class $t$ in the support set $S$ and $f$ denotes the backbone network. Then, the probability of a query sample $q$ belonging to class $t$ is related to the distance between $q$ and prototype of class $t$ [46], which is calculated as follows:", + "bbox": [ + 75, + 750, + 468, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP (y = t) = \\frac {\\exp d \\left(q , p _ {t}\\right)}{\\sum_ {i = 1} ^ {N} \\exp d \\left(q , p _ {i}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 830, + 468, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ means distance metric, e.g., cosine distance, and $N$ denotes the number of classes.", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b24ad8af9db3e7a6085fff5b1863d6db76c335e90d50d2aa121387db69c6079.jpg", + "image_caption": [ + "Figure 2. The illustration of how Semantic Evolution converts the simple name and the definition into the high-quality description." + ], + "image_footnote": [], + "bbox": [ + 503, + 88, + 893, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantic Evolution", + "text_level": 1, + "bbox": [ + 500, + 361, + 686, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section explains how we automatically generate high-quality semantics for few-shot classification by the designed process termed Semantic Evolution. This process first converts class names to short descriptions and then paraphrases them into high-quality semantics.", + "bbox": [ + 496, + 383, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we illustrated before, adopting class names as semantics [5, 39, 58] unavoidably introduces ambiguity. Therefore, we first retrieve the definition of each class name from WordNet [33] that matches image content, where definitions are more detailed and concrete than mere class names. However, even though definitions are able to precisely describe a class to some extent, brief descriptions may still overlook key visual information, due to the fact that visual features are highly variable.", + "bbox": [ + 496, + 460, + 892, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To comprehensively conclude variable visual features, we leverage the extensive knowledge stored in the pretrained Large Language Models (LLMs) [4] to augment definitions. The prompt we used is: $\\{\\text{WordNet Definition}\\}$ is the definition of the $\\{\\text{Class Name}\\}$ . Please rewrite and expand this definition to make it more detailed and consistent with scientific fact. Briefness is required, using only one paragraph. The illustration of the Semantic Evolution process is shown in Fig. 2.", + "bbox": [ + 496, + 595, + 890, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The paraphrased class description generated by LLMs precisely describes the corresponding class itself, which does not apply to other categories. This characteristic ensures that semantic descriptions do not leak information to new categories, which follows the paradigm of FSL.", + "bbox": [ + 496, + 732, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Semantic Alignment Network", + "text_level": 1, + "bbox": [ + 500, + 816, + 764, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After obtaining paraphrased semantics in the previous section, we devise the Semantic Alignment Network (SemAlign), aiming at translating hard periphery samples into robust class prototypes.", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28590", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac749bd011964ce9970d2992531a46ccc7543a6aec1a5d9cdfc48ed5d932d7a0.jpg", + "image_caption": [ + "Figure 3. The framework of our proposed SemFew. During the training stage, images and paraphrased semantics are encoded and fed into SemAlign $h$ , with the objective of reducing the distance between the output of $h$ and the class prototype in the visual space. During the testing stage, images in the support set are transformed into class prototypes by $h$ , and query images are classified by identifying the nearest prototype. The symbol $\\oplus$ denotes a concatenation operation." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 872, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a cluster of samples, the cluster center best encapsulates the characteristic of a category and has been regarded as a prototype in previous work [46]. Suppose $S$ represents semantic features, and $C$ stands for the center of each class, the simplest implementation of SemAlign is to learn the alignment between semantics $S$ and centers $C$ :", + "bbox": [ + 75, + 479, + 468, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {W _ {1}, W _ {2}} L \\left(\\sigma \\left(S ^ {\\top} W _ {1}\\right) W _ {2}, C\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 578, + 468, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{1}$ and $W_{2}$ are parameters of network, $\\sigma$ is an activation function.", + "bbox": [ + 76, + 613, + 468, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that each modality views things from a different and complementary perspective, this phenomenon can be reasonably utilized to obtain better prototypes. Therefore, we fuse visual features with high-quality semantics, and then align multi-modal output with the cluster center, the details of the whole pipeline can be found in Fig. 3. Specifically, suppose the $i$ -th image in the dataset is represented as $x_{i}$ , its label and semantic are denoted as $y_{i}$ and $s_i$ , respectively. First, $x_{i}$ is encoded by a vision encoder $f$ , which is pre-trained on the base set. Also, semantic description $s_i$ is encoded by an off-the-shelf text encoder $g$ . Then, the encoded visual and text features are concatenated together and fed into SemAlign $h$ to reconstruct the prototype. The training loss supervises the distance between the reconstructed prototype towards the cluster center:", + "bbox": [ + 75, + 643, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {W _ {1}, W _ {2}} \\mathbb {E} [ \\| h (f (x _ {i}), g (s _ {i})) - c _ {y} \\| _ {1} ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 878, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h(f(x_i), g(s_i)) = \\sigma([f(x_i) \\cdot g(s_i)]^\\top W_1)W_2$ is SemAlign with learnable weights $W_1$ and $W_2$ , $[\\cdot]$ means concatenation operation.", + "bbox": [ + 498, + 478, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the testing process, a $N$ -way $K$ -shot support set is randomly sampled in the novel set. For class $t$ , we obtain its paraphrased description $s_t$ by Semantic Evolution proposed in Sec. 3.2. Then, support image $x$ and description $s_t$ are encoded by $f$ and $g$ , respectively, and fed into $h$ to get the reconstructed prototype:", + "bbox": [ + 498, + 527, + 892, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} h \\left(f \\left(x _ {i}\\right), g \\left(s _ {t}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 636, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r_t$ is the reconstructed prototype. To ensure the reconstructed prototype is further aligned with the groundtruth class prototype, it will be fused with support images in a convex combination manner [57]:", + "bbox": [ + 498, + 695, + 890, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} = k r _ {t} + (1 - k) u _ {t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 775, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_t$ means the classification prototype of class $t$ , $u_t = \\frac{1}{K} \\sum_{i=1}^{K} f(x_i)$ is the mean vector of support samples, $r_t$ is the reconstructed prototype produced by SemAlign, and $k \\in [0,1]$ is a manually controlled fusion factor. Then, Eq. (2) is utilized to assign label for each query sample, where $d(\\cdot, \\cdot)$ is set as cosine distance.", + "bbox": [ + 498, + 810, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4a5521e2ab8d1d53d2e45d1e4aede357cec9ecf7d23d7431a3faad569e52fa72.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueBackboneMiniImageNetTieredImageNet
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
Visual-BasedMatchNet [53]NeurIPS'16ResNet-1265.64 ± 0.2078.72 ± 0.1568.50 ± 0.9280.60 ± 0.71
ProtoNet [46]NeurIPS'17ResNet-1262.39 ± 0.2180.53 ± 0.1468.23 ± 0.2384.03 ± 0.16
MAML [13]ICML'17ResNet-1249.24 ± 0.2158.05 ± 0.1067.92 ± 0.1772.41 ± 0.20
TADAM [37]NeurIPS'18ResNet-1258.50 ± 0.3076.70 ± 0.3062.13 ± 0.3181.92 ± 0.30
CAN [20]NeurIPS'19ResNet-1263.85 ± 0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
CTM [29]CVPR'19ResNet-1864.12 ± 0.8280.51 ± 0.1368.41 ± 0.3984.28 ± 1.73
RFS [49]ECCV'20ResNet-1262.02 ± 0.6379.64 ± 0.4469.74 ± 0.7284.41 ± 0.55
FEAT [60]CVPR'20ResNet-1266.78 ± 0.2082.05 ± 0.1470.80 ± 0.2384.79 ± 0.16
Meta-Baseline [6]ICCV'21ResNet-1263.17 ± 0.2379.26 ± 0.1768.62 ± 0.2783.29 ± 0.18
SUN [11]ECCV'22ViT-S67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [19]NeurIPS'22Swin-T72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
FGFL [7]ICCV'23ResNet-1269.14 ± 0.8086.01 ± 0.6273.21 ± 0.8887.21 ± 0.61
Meta-AdaM [45]NeurIPS'23ResNet-1259.89 ± 0.4977.92 ± 0.4365.31 ± 0.4885.24 ± 0.35
Semantic-BasedKTN [38]ICCV'19Conv-12864.42 ± 0.7274.16 ± 0.5674.16 ± 0.56-
AM3 [57]NeurIPS'19ResNet-1265.30 ± 0.4978.10 ± 0.3669.08 ± 0.4782.58 ± 0.31
TRAML[28]CVPR'20ResNet-1267.10 ± 0.5279.54 ± 0.60--
AM3-BERT [59]ICMR'21ResNet-1268.42 ± 0.5181.29 ± 0.5977.03 ± 0.8587.20 ± 0.70
SVAE-Proto [58]CVPR'22ResNet-1274.84 ± 0.2383.28 ± 0.4076.98 ± 0.6585.77 ± 0.50
SP-CLIP [5]CVPR'23Visformer-T72.31 ± 0.4083.42 ± 0.3078.03 ± 0.4688.55 ± 0.32
SemFewOursResNet-1277.63 ± 0.6383.04 ± 0.4878.96 ± 0.8085.88 ± 0.62
SemFew-TransOursSwin-T78.94 ± 0.6686.49 ± 0.5082.37 ± 0.7789.89 ± 0.52
", + "bbox": [ + 83, + 89, + 888, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results (%) on MiniImageNet and TieredImageNet. The ± shows 95% confidence intervals. The best results are shown in bold.", + "bbox": [ + 80, + 474, + 885, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 503, + 207, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We exhibit experimental results in this section, including benchmarks introduction, implementation details, and few-shot classification performance.", + "bbox": [ + 76, + 530, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 589, + 179, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The study evaluates the proposed method on six established FSL datasets: MiniImageNet [53], TieredImageNet [41], CIFAR-FS [27], FC100 [37], Places [63], and CUB [54]. MiniImageNet and TieredImageNet are both subsets of ImageNet [8]. MiniImageNet consists of 64 training classes, 16 validating classes, and 20 testing classes. TieredImageNet encompasses 351 training classes, 97 validating classes, and 160 testing classes. CIFAR-FS and FC100 are derived from the CIFAR-100 [26]. CIFAR-FS employs a random partitioning strategy with 64 training classes, 16 validating classes, and 20 testing classes. FC100 introduces a unique superclass partitioning approach, where the training set comprises 12 subclasses, namely 60 classes, while both the validation and test sets include four subclasses, totaling 20 classes. Places and Caltech-UCSD Birds-200-2011 (CUB) are datasets for cross-domain scenario testing. In a cross-domain evaluation [51], the model is initially trained on MiniImageNet's base split and tested on the corresponding test split.", + "bbox": [ + 75, + 613, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 503, + 718, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. We employ a pre-trained ResNet-12 [18] and Swin-T [30] as the vision encoders. Additionally, we apply two types of text encoder: the text encoder from ViT-B/16 CLIP [39] where output dimension is 512, and the BERT-Base text encoder [9] which has 768 output dimension. The network $h$ that combines visual and textual features consists of two fully connected layers and a LeakyReLU [31] activation function. The hidden layer has a dimension of 4,096. In the context of Semantic Evolution, the utilized Large Language Model is GPT-3.5-turbo [36].", + "bbox": [ + 496, + 531, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. A two-step training procedure is performed in experiments. In step one, two kinds of vision encoder are pre-trained, the ResNet-12 [18] follows the same pretraining settings as the FEAT [60], while the Swin-T [30] adopts the identical pretraining strategy as FewTURE [19]. Next, the network $h$ is trained 50 epochs. This training is conducted with a batch size of 128 and employs an Adam optimizer [24] with a learning rate of 1e-4.", + "bbox": [ + 496, + 686, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocol. The proposed method is evaluated under 5-way 1-shot and 5-way 5-shot settings. For each setting, 600 classification tasks are uniformly sampled from the novel set. In each task, there are 15 query samples per class. The mean and $95\\%$ confidence interval of the accuracy are reported.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28592", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/074e3ca630f83d9ce4f8a842a218a98aa9c012aad9bddb81a73dab69239aad02.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueBackboneCIFAR-FSFC100
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
ProtoNet [46]NeurIPS'17ResNet-1272.20 ± 0.7083.50 ± 0.5041.54 ± 0.7657.08 ± 0.76
TADAM [37]NeurIPS'18ResNet-12--40.10 ± 0.4056.10 ± 0.40
MetaOptNet [27]CVPR'19ResNet-1272.80 ± 0.7084.30 ± 0.5047.20 ± 0.6055.50 ± 0.60
MABAS [23]ECCV'20ResNet-1273.51 ± 0.9285.65 ± 0.6542.31 ± 0.7558.16 ± 0.78
RFS [49]ECCV'20ResNet-1271.50 ± 0.8086.00 ± 0.5042.60 ± 0.7059.10 ± 0.60
SUN [11]ECCV'22ViT-S78.37 ± 0.4688.84 ± 0.32--
FewTURE [19]NeurIPS'22Swin-T77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
Meta-AdaM [45]NeurIPS'23ResNet-12--41.12 ± 0.4956.14 ± 0.49
SP-CLIP [5]CVPR'23Visformer-T82.18 ± 0.4088.24 ± 0.3248.53 ± 0.3861.55 ± 0.41
SemFewOursResNet-1283.65 ± 0.7087.66 ± 0.6054.36 ± 0.7162.79 ± 0.74
SemFew-TransOursSwin-T84.34 ± 0.6789.11 ± 0.5454.27 ± 0.7765.02 ± 0.72
", + "bbox": [ + 102, + 88, + 870, + 315 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2f469ac7825f599475b5c87c5788187dc29c329e98d90b951cbcb5955106f080.jpg", + "table_caption": [ + "Table 2. Results (%) on CIFAR-FS and FC100. The ± shows 95% confidence intervals. The best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
MethodVenueCUBPlaces
1-shot5-shot1-shot5-shot
GNN [43]ICLR'1845.6962.2553.1070.84
S2M2 [32]WACV'2048.2470.44--
FT [51]ICLR'2047.4766.9855.7773.94
ATA [55]IJCAI'2145.0066.2253.5775.48
AFA [21]ECCV'2246.8668.2554.0476.21
StyleAdv [14]CVPR'2348.4968.7258.5877.73
LDP-net [64]CVPR'2349.8270.3953.8272.90
SemFew-NameOurs57.5872.2663.2274.54
SemFewOurs59.0772.4764.0174.70
", + "bbox": [ + 81, + 352, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Average results (\\%) on cross-domain scenarios. SemFew-Name denotes that semantics are class names.", + "bbox": [ + 76, + 523, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 561, + 279, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Few-Shot Classification", + "text_level": 1, + "bbox": [ + 76, + 585, + 294, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The comparative results of our proposed approach against previous methods across four benchmarks are shown in Tabs. 1 and 2. Specifically, in the 1-shot task, SemFew exhibits notable superiority, surpassing SP-CLIP [5] $0.93\\% - 5.83\\%$ on four benchmarks. Recent studies [5, 11, 19] have explored different Transformer architectures as backbones. Consequently, we conduct experiments with a Transformer backbone, resulting in SemFew-Trans. This variant further elevates performance, outperforming SP-CLIP by $2.16\\% - 6.63\\%$ in the 1-shot setting. In summary, SemFew-Trans achieves state-of-the-art results compared to both previous visual-based and semantic-based approaches across all 1-shot and 5-shot tasks.", + "bbox": [ + 75, + 609, + 470, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Cross-Domain Classification", + "text_level": 1, + "bbox": [ + 76, + 829, + 328, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate whether semantics serve as an effective bridge across novel tasks, we test SemFew on completely novel datasets. Such a task is also known as Cross-Domain Few-", + "bbox": [ + 76, + 854, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shot Learning (CDFSL) [51], which involves extending the capacity to recognize novel categories to domains distinct from the training domain. All the results are reported in Tab. 3, where the experimental setup is the same as FT [51]. Tab. 3 indicates that SemFew significantly surpassing the state-of-the-art methods, where it outperforms best compared methods by $9.25\\%$ (1-shot) and $1.87\\%$ (5-shot) on CUB. It also surpasses the lastest method [64] $10.19\\%$ (1-shot) and $1.80\\%$ (5-shot) on Places. Conversely, SemFew-Name, which adopts a template (i.e., The Photo of a bird called {Class Name}) as the semantic, performs worse than SemFew. This observation highlights the inadequacy of using solely names as semantics.", + "bbox": [ + 496, + 354, + 893, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 561, + 663, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablation experiments on three datasets, i.e., MiniImageNet, CIFAR-FS, and FC100, and all experiments are performed using the ResNet-12 network architecture.", + "bbox": [ + 496, + 584, + 890, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.1 Alignment Source", + "text_level": 1, + "bbox": [ + 500, + 650, + 679, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given that our framework is primarily responsible for translating visual samples $V$ and semantics descriptions $S$ into ground-truth class prototype (i.e., $V + S \\Rightarrow C$ ), we conduct ablation studies by removing $S$ or $V$ . The results can be found in Tab. 4. First, we remove semantics and force pure visual samples to reconstruct prototypes (i.e., $V \\Rightarrow C$ ). The result shows that it gets the worst performance, indicating that reconstructing a prototype from pure visual features is challenging. Then, we abandon visual features to implement our proposed principle in Sec. 3.3 (i.e., $S \\Rightarrow C$ ). As we can see, it achieves substantial performance gains, reflecting the rich class representative features within semantics. However, our method combining visual and semantic acquires the best performance, which indicates the complementarity between visual and semantic modalities does en", + "bbox": [ + 496, + 672, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/44b74f08e7a152ced310721429ff9ea401b4373bafe661549ede1b26148b88a6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method5-way 1-shot5-way 5-shot
MiniImageNetCIFAR-FSFC100MiniImageNetCIFAR-FSFC100
V ⇒ C66.27 ± 0.8174.43 ± 0.9044.72 ± 0.7581.70 ± 0.5586.74 ± 0.6160.58 ± 0.76
S ⇒ C76.77 ± 0.6682.27 ± 0.7152.12 ± 0.7382.92 ± 0.4887.42 ± 0.6062.22 ± 0.75
V + S ⇒ C77.63 ± 0.6383.65 ± 0.7054.36 ± 0.7183.04 ± 0.4887.66 ± 0.6062.79 ± 0.74
", + "bbox": [ + 106, + 89, + 864, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/51d58ea90200c55f05fb7f080edb4863c2a5fe783de914d08bee44412f813f10.jpg", + "table_caption": [ + "Table 4. Results (%) on different alignment sources. $V$ means visual features, $S$ represents semantic features, $C$ stands for class prototypes." + ], + "table_footnote": [], + "table_body": "
CenterMiniImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Cluster77.1682.9478.3685.1983.3987.6153.3062.22
Mean77.6383.0478.9685.8883.6587.6654.3662.79
", + "bbox": [ + 80, + 234, + 468, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/85a6fe121fe6717aae5acfe607292d73266575ba2b2f443325838182059e36be.jpg", + "table_caption": [ + "Table 5. Average results (\\%) on different prototypes. Mean represents the mean vector of sample clusters. Cluster means the prototype is determined by Clustering Algorithm." + ], + "table_footnote": [], + "table_body": "
classifierMinImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
LR77.5582.7078.6684.7283.0886.7051.5357.49
EU77.5683.1578.6585.2682.8386.8754.1462.54
CO77.6383.0478.9685.8883.6587.6654.3662.79
", + "bbox": [ + 80, + 354, + 468, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Average results $(\\%)$ on different classifiers. LR: linear logistic regression classifier. EU: euclidean-distance nearest prototype classifier. CO: cosine-distance nearest prototype classifier.", + "bbox": [ + 76, + 426, + 468, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hance the robustness of the reconstructed prototype.", + "bbox": [ + 76, + 493, + 419, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Prototype and Classifier Selection Strategy", + "text_level": 1, + "bbox": [ + 76, + 529, + 428, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 displays the performance associated with various learning targets. Results reveal that employing the mean vector as a prototype yields slightly better results compared to prototypes identified by the clustering algorithm [17]. This suggests that the mean vector consolidates representative visual information, making it a highly effective class prototype. As for the classification head design, we present the results in Tab. 6, which indicates the use of cosine distance as the metric to achieve nearest neighbor classification gets the best results, euclidean distance gets similar but lower outcomes, and linear logistic classifier gets the worst performance.", + "bbox": [ + 75, + 553, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3 Semantic Evolution", + "text_level": 1, + "bbox": [ + 76, + 755, + 266, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use different kinds of semantics and text encoders to evaluate the influence of semantics, where results can be found in Fig. 4. In most scenarios, our model attains optimal results when aided by paraphrased semantics, and the model trained on simple definitions secures the second-best performance. This trend is reasonable since paraphrased descriptions contain more knowledge than simple definitions and class names. Between the comparison of different text", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aac258d6026627afec3d05617817eb1fa140f66da681af56a215c01ea3daebc9.jpg", + "image_caption": [ + "(a) MiniImageNet, 1-shot" + ], + "image_footnote": [], + "bbox": [ + 501, + 234, + 694, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9f8ba4222802d95e8b212a91c57036ac921ec4d66476c9fd68395abdac56354.jpg", + "image_caption": [ + "(c) CIFAR-FS, 1-shot" + ], + "image_footnote": [], + "bbox": [ + 501, + 362, + 694, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0732e6ff299f45084b3b59a268bed241b2d5caa132b551144b24fe98a11cc78a.jpg", + "image_caption": [ + "(e) FC100, 1-shot", + "Figure 4. Average results $(\\%)$ on different semantics." + ], + "image_footnote": [], + "bbox": [ + 501, + 486, + 694, + 598 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bce1ec77520057a58b9132c2c2eb36e0804e5c0c28afcb7cae2d25a56ac4a42.jpg", + "image_caption": [ + "(b) MiniImageNet, 5-shot" + ], + "image_footnote": [], + "bbox": [ + 696, + 234, + 888, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46af499085228bacf53f45433249ebe175794bac562d87a095b371f055036ae4.jpg", + "image_caption": [ + "(d) CIFAR-FS, 5-shot" + ], + "image_footnote": [], + "bbox": [ + 697, + 361, + 888, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5178b19cf2cdb301a06e4cb9ea709a7b5101eabd689ba726a4053b139a8c77c.jpg", + "image_caption": [ + "(f) FC100, 5-shot" + ], + "image_footnote": [], + "bbox": [ + 697, + 486, + 888, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "encoders, the model trained on CLIP encoded semantics greatly outperforms the model trained on BERT encoded semantics. This is because CLIP is trained in a multi-modal manner, rendering it better capable of understanding images. Although a multi-modal text encoder has been widely adopted in FSL research, our model obtains competitive results on BERT, demonstrating that uni-modal semantics can also achieve excellent results.", + "bbox": [ + 496, + 652, + 890, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualization Analysis", + "text_level": 1, + "bbox": [ + 500, + 786, + 705, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We visualize visual space features through the t-SNE algorithm [52], where the results are presented in Fig. 5. We select three novel classes on the MiniImageNet dataset, and then randomly sampled three support samples and one hundred query samples, where the distribution of support and query samples in visual space can be found in Figs. 5a", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28594", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/946115ba043e80bfa1f2164936e653aa653280766b8b4e8f594f2a12b1079258.jpg", + "image_caption": [ + "(a) Support" + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 272, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/260f067000755f8f22ca3e4a24b50ad123eacd792bfe095f3066fab05d9b50c9.jpg", + "image_caption": [ + "(b) Query" + ], + "image_footnote": [], + "bbox": [ + 282, + 87, + 478, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ba299f6f11c5aab3394802512653b41d24d57df33ef1370aad699f2003e974f.jpg", + "image_caption": [ + "(c) Real Prototype" + ], + "image_footnote": [], + "bbox": [ + 493, + 87, + 687, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28ac5315e67b252f1f1ba15ece9c403d9ce876c550f2031b52a4450b41b94755.jpg", + "image_caption": [ + "(d) Reconstructed Prototype" + ], + "image_footnote": [], + "bbox": [ + 700, + 88, + 893, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d373e0caa9759a29b73689bae818291e4e338c01c941d5b6ab4c682f97c15d3a.jpg", + "image_caption": [ + "Figure 5. Visualization results on the MiniImageNet dataset. Different colors or shapes represent different classes. The $\\star$ represents the class prototypes, and the $\\diamond$ denotes the prototypes reconstructed by our method." + ], + "image_footnote": [], + "bbox": [ + 78, + 303, + 272, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f37cf55549b011c9f554ab973ce627e53bbffc76c1407b9cec4a90bd3b1584a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 303, + 468, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab76e6586d5fdbe335edca0a6d35503d49a6b29cdc6a971f6dbea040ed0e5628.jpg", + "image_caption": [ + "(a) MiniImageNet,CLIP" + ], + "image_footnote": [], + "bbox": [ + 78, + 429, + 272, + 540 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0cfb4a20ced74d64b5cec0bc59a9a404c98f608f3daf84004c038c2b4702128.jpg", + "image_caption": [ + "(b) MiniImageNet, BERT" + ], + "image_footnote": [], + "bbox": [ + 274, + 429, + 468, + 540 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fcbbd8bafd95971bf9721b4734ad71b72feaa93bad3fec9fda3a3a24a13bbd05.jpg", + "image_caption": [ + "(c) CIFAR-FS,CLIP", + "(e) FC100, CLIP", + "Figure 6. Average results $(\\%)$ on different fusion factor $k$" + ], + "image_footnote": [], + "bbox": [ + 78, + 555, + 272, + 664 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/374e297665525f1366bed75e7c82c4e4bbf2abb9e9ae9eb526f286bf22771652.jpg", + "image_caption": [ + "(d) CIFAR-FS, BERT", + "(f) FC100, BERT" + ], + "image_footnote": [], + "bbox": [ + 274, + 555, + 468, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and 5b. In Fig. 5c, we put support samples, i.e., red dots, mean vector of each class, i.e., blue stars, and query samples in the same picture. As we can see, support samples are close to the cluster edge rather than the mean center, which leads to difficulties in correctly classifying most samples in the query set. However, after being transformed by SemFew, the reconstructed prototypes are closer to the mean center of each cluster compared to the support set samples. This phenomenon demonstrates that our method can reasonably transform visual samples into class prototypes corresponding to their respective categories.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Fusion Factor Analysis", + "text_level": 1, + "bbox": [ + 500, + 306, + 714, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The fusion factor, denoted as $k$ in Eq. (6), significantly influences performance. To demonstrate, we systematically vary the value of $k$ between 0 and 1, using a step size of 0.01, and present the accuracy trends across three benchmarks in Fig. 6. Notably, our model, i.e., blue line, gets the highest accuracy when $k$ is large. Given that $k$ represents the percentage of the reconstructed prototype $r_t$ during fusion in Eq. (6), the experiment highlights that our model produces high-quality prototypes. This observation is further supported by the accuracy comparison when $k = 1$ . In such cases, our model consistently outperforms others, providing strong evidence for the effectiveness of our reconstructed prototypes.", + "bbox": [ + 496, + 329, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we add $S \\Rightarrow V$ experiment for comparison, which generates visual samples rather than prototypes through a Conditional Generative Adversarial Network [34]. This experiment allows us to explore whether we can build robust prototypes indirectly by expanding the sample of the support set. As we can see, $S \\Rightarrow V$ , i.e., yellow line, gets relatively low performance in all settings. The experiments indicate that producing the prototype by expanding then averaging the support set is not as effective as directly generating class prototypes.", + "bbox": [ + 496, + 526, + 892, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 693, + 619, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To unleash the potential of semantics in FSL, we propose the Semantic Evolution to automatically generate high-quality semantics from simple class names. Then, we design a simple Semantic Alignment Network to transform the concatenated visual features and high-quality semantics features into robust class prototypes for classification. The experimental results show that with high-quality semantics, the basic network can easily achieve greater performance compared to other state-of-the-art methods.", + "bbox": [ + 496, + 719, + 890, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by 2023YFF1204901, NSFC-62076172, 2021DQ02-0903, 2023YFG0116.", + "bbox": [ + 496, + 854, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28595", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. In NeurIPS, 2016. 2", + "[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 2", + "[3] Irving Biederman. Recognition-by-components: a theory of human image understanding. *Psychol Rev*, 94(2):115, 1987. 1", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3", + "[5] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. In CVPR, pages 23581-23591, 2023. 1, 2, 3, 5, 6", + "[6] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In ICCV, pages 9062-9071, 2021. 5", + "[7] Hao Cheng, Siyuan Yang, Joel Tianyi Zhou, Lanqing Guo, and Bihan Wen. Frequency guidance matters in few-shot learning. In ICCV, pages 11814-11824, 2023. 1, 5", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 1, 5", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[10] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. In NeurIPS, pages 21981-21993, 2020. 2", + "[11] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In ECCV, pages 329-347. Springer, 2022. 1, 5, 6", + "[12] Thomas Elsken, Benedikt Staffler, Jan Hendrik Metzen, and Frank Hutter. Meta-learning of neural architectures for few-shot learning. In CVPR, pages 12365–12375, 2020. 2", + "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Int. Conf. Mach. Learn., pages 1126-1135. PMLR, 2017. 1, 2, 5", + "[14] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In CVPR, pages 24575-24584, 2023. 6", + "[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In ICCV, pages 8059-8068, 2019. 2", + "[16] Fusheng Hao, Fengxiang He, Liu Liu, Fuxiang Wu, Dacheng Tao, and Jun Cheng. Class-aware patch embedding adaptation for few-shot image classification. In ICCV, pages 18905-18915, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. J R Stat Soc Ser C Appl Stat, 28(1):100-108, 1979. 7", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5", + "[19] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. NeurIPS, 35:3582-3595, 2022. 1, 2, 5, 6", + "[20] Ruibing Hou, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot classification. In NeurIPS, 2019. 5", + "[21] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 6", + "[22] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In CVPR, pages 11719-11727, 2019. 2", + "[23] Jaekyeom Kim, Hyoungseok Kim, and Gunhee Kim. Model-agnostic boundary-adversarial sampling for test-time generalization in few-shot learning. In ECCV, pages 599–617. Springer, 2020. 6", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[25] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Deep Learn. workshop. Lille, 2015. 1, 2", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In CVPR, pages 10657-10665, 2019. 5, 6", + "[28] Aoxue Li, Weiran Huang, Xu Lan, Jiashi Feng, Zhenguo Li, and Liwei Wang. Boosting few-shot learning with adaptive margin loss. In CVPR, pages 12576-12584, 2020. 1, 2, 5", + "[29] Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features for few-shot learning by category traversal. In CVPR, pages 1-10, 2019. 1, 5", + "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5", + "[31] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, Georgia, USA, 2013. 5", + "[32] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In IEEE Win. Conf. Appl. Comput. Vis., pages 2218-2227, 2020. 1, 2, 6", + "[33] George A Miller. Wordnet: a lexical database for english. Commun. ACM, 38(11):39-41, 1995. 3", + "[34] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28596", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. arXiv preprint arXiv:1707.03141, 2017. 2", + "[36] OpenAI. Chatgpt. https://chat.openai.com, 2023.5", + "[37] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 1, 5, 6", + "[38] Zhimao Peng, Zechao Li, Junge Zhang, Yan Li, Guo-Jun Qi, and Jinhui Tang. Few-shot image recognition with knowledge transfer. In ICCV, pages 441-449, 2019. 1, 5", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763. PMLR, 2021. 3, 5", + "[40] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2", + "[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5", + "[42] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 2", + "[43] Victor Garcia Satorras and Joan Bruna Estrach. Few-shot learning with graph neural networks. In ICLR, 2018. 6", + "[44] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, pages 8247-8255, 2019. 1", + "[45] Hongyang Gao Siyuan Sun. Meta-adam: A meta-learned adaptive optimizer with momentum for few-shot learning. In NeurIPS, 2023. 5, 6", + "[46] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6", + "[47] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2", + "[48] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199–1208, 2018. 2", + "[49] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, pages 266-282. Springer, 2020. 1, 2, 5, 6", + "[50] Pavel Tokmakov, Yu-Xiong Wang, and Martial Hebert. Learning compositional representations for few-shot recognition. In ICCV, pages 6372-6381, 2019. 1", + "[51] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. J Mach Learn Res, 9(11), 2008. 7", + "[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In NeurIPS, 2016. 1, 2, 5", + "[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[55] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In IJCAI, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. 6", + "[56] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, pages 1831-1840, 2019. 1", + "[57] Chen Xing, Nagar Rostamzadeh, Boris Oreshkin, and Pedro O O Pinheiro. Adaptive cross-modal few-shot learning. In NeurIPS, 2019. 1, 2, 4, 5", + "[58] Jingyi Xu and Hieu Le. Generating representative samples for few-shot classification. In CVPR, pages 9003-9013, 2022. 2, 3, 5", + "[59] Kun Yan, Zied Bouraoui, Ping Wang, Shoaib Jameel, and Steven Schockaert. Aligning visual prototypes with bert embeddings for few-shot learning. In ICMR, pages 367-375, 2021. 1, 5", + "[60] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, pages 8808-8817, 2020. 1, 5", + "[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In CVPR, pages 3754-3762, 2021. 1, 2", + "[62] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Differentiable earth mover's distance for few-shot learning. IEEE TPAMI, 2022. 2", + "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2017. 5", + "[64] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In CVPR, pages 20061-20070, 2023. 6", + "[65] Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast context adaptation via meta-learning. In Int. Conf. Mach. Learn., pages 7693-7702. PMLR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 757 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_model.json b/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f5589a4bfe355097bdcdfc94478b31f04845ffab --- /dev/null +++ b/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.266, + 0.131, + 0.704, + 0.154 + ], + "angle": 0, + "content": "Simple Semantic-Aided Few-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.179, + 0.757, + 0.285 + ], + "angle": 0, + "content": "Hai Zhang\\(^{1*}\\) Junzhe Xu\\(^{1,2*‡}\\) Shanlin Jiang\\(^{3}\\) Zhenan He\\(^{1\\dagger}\\)\\\n\\(^{1}\\)College of Computer Science, Sichuan University\\\n\\(^{2}\\)Banma Network Technology, Alibaba Group\\\n\\(^{3}\\)Naveen Jindal School of Management, University of Texas at Dallas\\\nzhanghi@stu.scu.edu.cn, xujunzhe.xjz@alibaba-inc.com\\\nshanlin.jiang@utdallas.edu, zhenan@scu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Learning from a limited amount of data, namely Few-Shot Learning, stands out as a challenging computer vision task. Several works exploit semantics and design complicated semantic fusion mechanisms to compensate for rare representative features within restricted data. However, relying on naive semantics such as class names introduces biases due to their brevity, while acquiring extensive semantics from external knowledge takes a huge time and effort. This limitation severely constrains the potential of semantics in Few-Shot Learning. In this paper, we design an automatic way called Semantic Evolution to generate high-quality semantics. The incorporation of high-quality semantics alleviates the need for complex network structures and learning algorithms used in previous works. Hence, we employ a simple two-layer network termed Semantic Alignment Network to transform semantics and visual features into robust class prototypes with rich discriminative features for few-shot classification. The experimental results show our framework outperforms all previous methods on six benchmarks, demonstrating a simple network with high-quality semantics can beat intricate multi-modal modules on few-shot classification tasks. Code is available at https://github.com/zhangdoudou123/SemFew." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Deep learning models have witnessed substantial advancements by leveraging extensive annotated data [8]. Nevertheless, in numerous real-world scenarios, the shortage of labeled data restricts the applicability of conventional deep learning approaches. In contrast, humans possess a remarkable cognitive ability to learn new concepts and recognize" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.891, + 0.369 + ], + "angle": 0, + "content": "categories from just a few examples [3]. Motivated by this, Few-Shot Learning (FSL) [46] is proposed to mimic the ability of humans to learn from a few labeled samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.893, + 0.566 + ], + "angle": 0, + "content": "In a typical setting of FSL [46], the support set, which consists of \\(N\\) novel classes and each class contains \\(K\\) samples, is provided for the model. The model learned on the support set is required to accurately classify test samples, which are labeled as the query set, into one of the \\(N\\) categories. A conventional classification approach involves projecting both support and query sets onto a pre-established metric space. Then, query images are classified by finding the nearest support image. However, few support samples may not contain sufficient discriminative features for recognition. These samples are typically located on the periphery of the sample cluster in the metric space, resulting in unstable distance evaluations and inaccurate classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.779 + ], + "angle": 0, + "content": "To address this issue, researchers start from visual-based methods [7, 11, 13, 16, 19, 29, 32, 37, 46, 49, 53, 60], aiming at extracting class-related features from periphery image, reducing the intra-class variation among samples, and thus constructing robust features for classification. Although visual-based methods have undergone extensive investigation and achieved great success, this type of approach still struggles to handle situations when there is only one hard periphery sample with minimal semantic features per class, i.e., the one-shot learning task [25, 53]. Hence, some researchers turned to applying semantics as auxiliary information to help the model better understand periphery images, leveraging the synergies between language and vision modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "From this perspective, semantic-based methods [5, 28, 38, 44, 50, 56-59, 61] dedicated to exploring different types of semantics to enhance periphery samples. One kind of method [57-59] applies the class name as the semantic source. However, class name is not the best way to understand a novel class. For instance, If a person never saw a zebra, it would be easier to identify a zebra by the definition, i.e., a horse with stripes, than by the name zebra." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.404, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution. † Corresponding Author. \n‡ Work done during master's degree at Sichuan University." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28588" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.084, + 0.096, + 0.465, + 0.136 + ], + "angle": 0, + "content": "The American robin is renowned for its striking appearance, with a rust-red to orange breast and abdomen, a dark grayish-blue upper body, and a white eye ring. This bird plays a pivotal ecological role as a seed disperser, primarily for fruits and berries,..." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.155, + 0.47, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.471, + 0.346 + ], + "angle": 0, + "content": "Figure 1. The introduction of how high-quality semantics reconstruct the class prototype through complementarity between modalities. Periphery image is an image with fewer discriminative features and prototype image represents images with concrete and enriched representative features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Meanwhile, the class name contains ambiguity in some scenarios. For example, the class `ear` is easily misidentified as the human organ, but images in MiniImageNet [53] belong to this class are actually photos of corn. Another kind of semantic exploits discrete attribute labels as the substitution of mere class names and achieves good performance [61]. Nevertheless, the collection process is time-consuming and expert assistance requiring. Moreover, discrete attribute labels lack robust representation and struggle to benefit from context. Therefore, effectively collecting and leveraging high-quality semantics becomes an urgent research need." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a simple framework called Semantic-Aided Few-Shot Learning (SemFew). To begin with, we identify the vital role of high-quality semantics in FSL and innovatively propose an automatic step-by-step Semantic Evolution (SemEvo) process to acquire detailed and accurate semantics. It first converts the class name into a short description that matches the image content of each class. Next, to further enrich recognizable characteristics, we expand and paraphrase short descriptions. The paraphrased descriptions contain more class-related knowledge than class names and short descriptions, making them better encompass a vast array of visual features. To illustrate, Fig. 1 showcases an instance that the paraphrased semantic contains a large number of detailed descriptions matching visual features e.g. organ characteristics and eating habitat. With this information, the model can better reconstruct an image with periphery feature (left image), into an image with prototype feature (right image), which benefits few-shot classification. Then, to fully exploit the advantage of paraphrased semantics, we design a network called Semantic Alignment Network (SemAlign). Different from previous methods, SemAlign does not adopt complex semantic understanding modules but utilizes a basic two-layer network. After accepting semantics and visual features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "as input, it reconstructs prototype features for subsequent classification. Experimental results on six widely adopted benchmarks demonstrate the effectiveness of our method, providing a simple insight into future FSL research. Our contributions can be summarized as three points:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.212 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to consider the automatic way of collecting high-quality semantics and applying them in FSL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.258 + ], + "angle": 0, + "content": "- We design a simple and efficient way to translate high-quality semantics and visual features into prototypes, without any intricate semantic understanding modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.319 + ], + "angle": 0, + "content": "- Our approach achieves state-of-the-art performance across six benchmarks in FSL research, underscoring that a basic network can obtain excellent performance when supported by high-quality semantics." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.642, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.432 + ], + "angle": 0, + "content": "The significant issue in FSL is how to effectively and accurately extract class-related features from limited data. Targeting this issue, existing research explored two different types of approaches, i.e., visual-based methods and semantic-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Visual-Based Methods focus on extracting class-related features from images for classification. There are two main types of this approach. The approaches with the optimization principle behind [1, 13, 40] concentrated on acquiring a set of initial model parameters capable of rapid adaptation to novel classes. However, updating the entire model with a limited amount of labeled data may result in metaoverfitting [12, 22, 35, 42, 65]. The approaches considered metric design aiming at training a metric space, where inter-class distances are maximized, while intra-class distances are minimized. Within this type of approach, one attracting method focuses on integrating self-supervised loss to pretrain the backbone network [10, 15, 19, 32, 49]. Other methods are dedicated to improving the metric, and previous research spanning from initially manually designed metrics [2, 25, 46, 62] to subsequently model-based metrics [48, 53], all of which have yielded encouraging outcomes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Semantic-Based Methods sought to enhance visual recognition performance by fusing complementary information from both visual and textual modalities [5, 28, 57, 58, 61], and related research introducing intricate network frameworks to effectively utilize textual information. Chen et al. [57] introduced an adaptive fusion mechanism to integrate visual prototypes with semantic prototypes obtained through word embeddings of class labels. Zhang et al. [61] designed a framework for prototype completion, leveraging rare attribute-level information from labels to construct representative class prototypes. Xu and Le [58] proposed the use of the Conditional Variational Autoencoder (CVAE) [47] model to generate visual features based on semantic embeddings for further classification. Li et al. [28] pre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28589" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "sented a task-relevant adaptive margin loss based on the semantic similarity between class names. Chen et al. [5] devised complementary mechanisms for inserting semantic prompts into the feature extractor, allowing the feature extractor to better focus on class-specific features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.289 + ], + "angle": 0, + "content": "Previous research emphasized the exploitation of class names and designed a complicated network to integrate semantics with visual information. In contrast, our primary concern lies in the inherent quality of semantics, and how it can relieve the burden of designing intricate network structures and learning algorithms. Remarkably, we exhibit that high-quality semantics information has the ability to yield superior outcomes even within a simple network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.299, + 0.212, + 0.316 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.386 + ], + "angle": 0, + "content": "This section first revisits the background of FSL, then introduces two components of SemFew: explaining how semantics are gradually improved through SemEvo, and demonstrating the details of our proposed SemAlign." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.392, + 0.21, + 0.407 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.626 + ], + "angle": 0, + "content": "The dataset of FSL includes two components: a base set \\(\\mathcal{D}_{base} = \\{(x,y)|x\\in \\mathcal{X}_{base},y\\in \\mathcal{C}_{base}\\}\\) for training a metric space and a novel set \\(\\mathcal{D}_{novel} = \\{(x,y)|x\\in \\mathcal{X}_{novel},y\\in \\mathcal{C}_{novel}\\}\\) for testing, where \\(x\\) denotes the image and \\(y\\) represents the label. It is important to note that the label space for both sets is disjoint, i.e., \\(\\mathcal{C}_{base}\\cap \\mathcal{C}_{novel} = \\emptyset\\). During the testing process, the support set \\(S = \\{(x_i,y_i)\\}_{i = 0}^{N\\times K}\\) is randomly sampled from \\(\\mathcal{D}_{novel}\\), which contains \\(N\\) classes with each class consisting of \\(K\\) samples. Then, the model is required to correctly classify images in the query set \\(\\mathcal{Q} = \\{(x_i,y_i)\\}_{i = 0}^{N\\times M}\\) into one of \\(N\\) classes in the support set \\(S\\), where \\(M\\) is the number of query samples in each class. Generally, this type of classification task is named as \\(N\\)-way \\(K\\)-shot task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.715 + ], + "angle": 0, + "content": "A simple but effective FSL method is to calculate the prototype of each support class, and then classify query samples by finding the nearest prototype in the metric space. To be specific, the prototype of class \\( t \\), labeled as \\( p_t \\), is the mean vector of all samples in the support set belonging to class \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.713, + 0.47, + 0.75 + ], + "angle": 0, + "content": "\\[\np _ {t} = \\frac {1}{\\left| \\mathcal {S} _ {t} \\right|} \\sum_ {x _ {i} \\in \\mathcal {S} _ {t}} f \\left(x _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.751, + 0.47, + 0.827 + ], + "angle": 0, + "content": "where \\(S_{t}\\) is a set of all samples belonging to class \\(t\\) in the support set \\(S\\) and \\(f\\) denotes the backbone network. Then, the probability of a query sample \\(q\\) belonging to class \\(t\\) is related to the distance between \\(q\\) and prototype of class \\(t\\) [46], which is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.831, + 0.469, + 0.868 + ], + "angle": 0, + "content": "\\[\nP (y = t) = \\frac {\\exp d \\left(q , p _ {t}\\right)}{\\sum_ {i = 1} ^ {N} \\exp d \\left(q , p _ {i}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) means distance metric, e.g., cosine distance, and \\(N\\) denotes the number of classes." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.895, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Figure 2. The illustration of how Semantic Evolution converts the simple name and the definition into the high-quality description." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.687, + 0.377 + ], + "angle": 0, + "content": "3.2. Semantic Evolution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.461 + ], + "angle": 0, + "content": "This section explains how we automatically generate high-quality semantics for few-shot classification by the designed process termed Semantic Evolution. This process first converts class names to short descriptions and then paraphrases them into high-quality semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.597 + ], + "angle": 0, + "content": "As we illustrated before, adopting class names as semantics [5, 39, 58] unavoidably introduces ambiguity. Therefore, we first retrieve the definition of each class name from WordNet [33] that matches image content, where definitions are more detailed and concrete than mere class names. However, even though definitions are able to precisely describe a class to some extent, brief descriptions may still overlook key visual information, due to the fact that visual features are highly variable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.733 + ], + "angle": 0, + "content": "To comprehensively conclude variable visual features, we leverage the extensive knowledge stored in the pretrained Large Language Models (LLMs) [4] to augment definitions. The prompt we used is: \\(\\{\\text{WordNet Definition}\\}\\) is the definition of the \\(\\{\\text{Class Name}\\}\\). Please rewrite and expand this definition to make it more detailed and consistent with scientific fact. Briefness is required, using only one paragraph. The illustration of the Semantic Evolution process is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.809 + ], + "angle": 0, + "content": "The paraphrased class description generated by LLMs precisely describes the corresponding class itself, which does not apply to other categories. This characteristic ensures that semantic descriptions do not leak information to new categories, which follows the paradigm of FSL." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.765, + 0.834 + ], + "angle": 0, + "content": "3.3. Semantic Alignment Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "After obtaining paraphrased semantics in the previous section, we devise the Semantic Alignment Network (SemAlign), aiming at translating hard periphery samples into robust class prototypes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28590" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.873, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.397, + 0.893, + 0.455 + ], + "angle": 0, + "content": "Figure 3. The framework of our proposed SemFew. During the training stage, images and paraphrased semantics are encoded and fed into SemAlign \\( h \\), with the objective of reducing the distance between the output of \\( h \\) and the class prototype in the visual space. During the testing stage, images in the support set are transformed into class prototypes by \\( h \\), and query images are classified by identifying the nearest prototype. The symbol \\( \\oplus \\) denotes a concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.47, + 0.571 + ], + "angle": 0, + "content": "For a cluster of samples, the cluster center best encapsulates the characteristic of a category and has been regarded as a prototype in previous work [46]. Suppose \\( S \\) represents semantic features, and \\( C \\) stands for the center of each class, the simplest implementation of SemAlign is to learn the alignment between semantics \\( S \\) and centers \\( C \\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.579, + 0.469, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\min _ {W _ {1}, W _ {2}} L \\left(\\sigma \\left(S ^ {\\top} W _ {1}\\right) W _ {2}, C\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.469, + 0.642 + ], + "angle": 0, + "content": "where \\( W_{1} \\) and \\( W_{2} \\) are parameters of network, \\( \\sigma \\) is an activation function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Considering that each modality views things from a different and complementary perspective, this phenomenon can be reasonably utilized to obtain better prototypes. Therefore, we fuse visual features with high-quality semantics, and then align multi-modal output with the cluster center, the details of the whole pipeline can be found in Fig. 3. Specifically, suppose the \\(i\\)-th image in the dataset is represented as \\(x_{i}\\), its label and semantic are denoted as \\(y_{i}\\) and \\(s_i\\), respectively. First, \\(x_{i}\\) is encoded by a vision encoder \\(f\\), which is pre-trained on the base set. Also, semantic description \\(s_i\\) is encoded by an off-the-shelf text encoder \\(g\\). Then, the encoded visual and text features are concatenated together and fed into SemAlign \\(h\\) to reconstruct the prototype. The training loss supervises the distance between the reconstructed prototype towards the cluster center:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.88, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {W _ {1}, W _ {2}} \\mathbb {E} [ \\| h (f (x _ {i}), g (s _ {i})) - c _ {y} \\| _ {1} ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.892, + 0.526 + ], + "angle": 0, + "content": "where \\( h(f(x_i), g(s_i)) = \\sigma([f(x_i) \\cdot g(s_i)]^\\top W_1)W_2 \\) is SemAlign with learnable weights \\( W_1 \\) and \\( W_2 \\), \\([\\cdot]\\) means concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.893, + 0.62 + ], + "angle": 0, + "content": "During the testing process, a \\(N\\)-way \\(K\\)-shot support set is randomly sampled in the novel set. For class \\(t\\), we obtain its paraphrased description \\(s_t\\) by Semantic Evolution proposed in Sec. 3.2. Then, support image \\(x\\) and description \\(s_t\\) are encoded by \\(f\\) and \\(g\\), respectively, and fed into \\(h\\) to get the reconstructed prototype:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.637, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\nr _ {t} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} h \\left(f \\left(x _ {i}\\right), g \\left(s _ {t}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.696, + 0.892, + 0.756 + ], + "angle": 0, + "content": "where \\(r_t\\) is the reconstructed prototype. To ensure the reconstructed prototype is further aligned with the groundtruth class prototype, it will be fused with support images in a convex combination manner [57]:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.776, + 0.891, + 0.793 + ], + "angle": 0, + "content": "\\[\np _ {t} = k r _ {t} + (1 - k) u _ {t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\( p_t \\) means the classification prototype of class \\( t \\), \\( u_t = \\frac{1}{K} \\sum_{i=1}^{K} f(x_i) \\) is the mean vector of support samples, \\( r_t \\) is the reconstructed prototype produced by SemAlign, and \\( k \\in [0,1] \\) is a manually controlled fusion factor. Then, Eq. (2) is utilized to assign label for each query sample, where \\( d(\\cdot, \\cdot) \\) is set as cosine distance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28591" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.889, + 0.466 + ], + "angle": 0, + "content": "
MethodVenueBackboneMiniImageNetTieredImageNet
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
Visual-BasedMatchNet [53]NeurIPS'16ResNet-1265.64 ± 0.2078.72 ± 0.1568.50 ± 0.9280.60 ± 0.71
ProtoNet [46]NeurIPS'17ResNet-1262.39 ± 0.2180.53 ± 0.1468.23 ± 0.2384.03 ± 0.16
MAML [13]ICML'17ResNet-1249.24 ± 0.2158.05 ± 0.1067.92 ± 0.1772.41 ± 0.20
TADAM [37]NeurIPS'18ResNet-1258.50 ± 0.3076.70 ± 0.3062.13 ± 0.3181.92 ± 0.30
CAN [20]NeurIPS'19ResNet-1263.85 ± 0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
CTM [29]CVPR'19ResNet-1864.12 ± 0.8280.51 ± 0.1368.41 ± 0.3984.28 ± 1.73
RFS [49]ECCV'20ResNet-1262.02 ± 0.6379.64 ± 0.4469.74 ± 0.7284.41 ± 0.55
FEAT [60]CVPR'20ResNet-1266.78 ± 0.2082.05 ± 0.1470.80 ± 0.2384.79 ± 0.16
Meta-Baseline [6]ICCV'21ResNet-1263.17 ± 0.2379.26 ± 0.1768.62 ± 0.2783.29 ± 0.18
SUN [11]ECCV'22ViT-S67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [19]NeurIPS'22Swin-T72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
FGFL [7]ICCV'23ResNet-1269.14 ± 0.8086.01 ± 0.6273.21 ± 0.8887.21 ± 0.61
Meta-AdaM [45]NeurIPS'23ResNet-1259.89 ± 0.4977.92 ± 0.4365.31 ± 0.4885.24 ± 0.35
Semantic-BasedKTN [38]ICCV'19Conv-12864.42 ± 0.7274.16 ± 0.5674.16 ± 0.56-
AM3 [57]NeurIPS'19ResNet-1265.30 ± 0.4978.10 ± 0.3669.08 ± 0.4782.58 ± 0.31
TRAML[28]CVPR'20ResNet-1267.10 ± 0.5279.54 ± 0.60--
AM3-BERT [59]ICMR'21ResNet-1268.42 ± 0.5181.29 ± 0.5977.03 ± 0.8587.20 ± 0.70
SVAE-Proto [58]CVPR'22ResNet-1274.84 ± 0.2383.28 ± 0.4076.98 ± 0.6585.77 ± 0.50
SP-CLIP [5]CVPR'23Visformer-T72.31 ± 0.4083.42 ± 0.3078.03 ± 0.4688.55 ± 0.32
SemFewOursResNet-1277.63 ± 0.6383.04 ± 0.4878.96 ± 0.8085.88 ± 0.62
SemFew-TransOursSwin-T78.94 ± 0.6686.49 ± 0.5082.37 ± 0.7789.89 ± 0.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.475, + 0.887, + 0.49 + ], + "angle": 0, + "content": "Table 1. Results (%) on MiniImageNet and TieredImageNet. The ± shows 95% confidence intervals. The best results are shown in bold." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.505, + 0.209, + 0.522 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.47, + 0.577 + ], + "angle": 0, + "content": "We exhibit experimental results in this section, including benchmarks introduction, implementation details, and few-shot classification performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.59, + 0.18, + 0.605 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The study evaluates the proposed method on six established FSL datasets: MiniImageNet [53], TieredImageNet [41], CIFAR-FS [27], FC100 [37], Places [63], and CUB [54]. MiniImageNet and TieredImageNet are both subsets of ImageNet [8]. MiniImageNet consists of 64 training classes, 16 validating classes, and 20 testing classes. TieredImageNet encompasses 351 training classes, 97 validating classes, and 160 testing classes. CIFAR-FS and FC100 are derived from the CIFAR-100 [26]. CIFAR-FS employs a random partitioning strategy with 64 training classes, 16 validating classes, and 20 testing classes. FC100 introduces a unique superclass partitioning approach, where the training set comprises 12 subclasses, namely 60 classes, while both the validation and test sets include four subclasses, totaling 20 classes. Places and Caltech-UCSD Birds-200-2011 (CUB) are datasets for cross-domain scenario testing. In a cross-domain evaluation [51], the model is initially trained on MiniImageNet's base split and tested on the corresponding test split." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.719, + 0.521 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.683 + ], + "angle": 0, + "content": "Architecture. We employ a pre-trained ResNet-12 [18] and Swin-T [30] as the vision encoders. Additionally, we apply two types of text encoder: the text encoder from ViT-B/16 CLIP [39] where output dimension is 512, and the BERT-Base text encoder [9] which has 768 output dimension. The network \\( h \\) that combines visual and textual features consists of two fully connected layers and a LeakyReLU [31] activation function. The hidden layer has a dimension of 4,096. In the context of Semantic Evolution, the utilized Large Language Model is GPT-3.5-turbo [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Training Details. A two-step training procedure is performed in experiments. In step one, two kinds of vision encoder are pre-trained, the ResNet-12 [18] follows the same pretraining settings as the FEAT [60], while the Swin-T [30] adopts the identical pretraining strategy as FewTURE [19]. Next, the network \\( h \\) is trained 50 epochs. This training is conducted with a batch size of 128 and employs an Adam optimizer [24] with a learning rate of 1e-4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation protocol. The proposed method is evaluated under 5-way 1-shot and 5-way 5-shot settings. For each setting, 600 classification tasks are uniformly sampled from the novel set. In each task, there are 15 query samples per class. The mean and \\(95\\%\\) confidence interval of the accuracy are reported." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28592" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.872, + 0.316 + ], + "angle": 0, + "content": "
MethodVenueBackboneCIFAR-FSFC100
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
ProtoNet [46]NeurIPS'17ResNet-1272.20 ± 0.7083.50 ± 0.5041.54 ± 0.7657.08 ± 0.76
TADAM [37]NeurIPS'18ResNet-12--40.10 ± 0.4056.10 ± 0.40
MetaOptNet [27]CVPR'19ResNet-1272.80 ± 0.7084.30 ± 0.5047.20 ± 0.6055.50 ± 0.60
MABAS [23]ECCV'20ResNet-1273.51 ± 0.9285.65 ± 0.6542.31 ± 0.7558.16 ± 0.78
RFS [49]ECCV'20ResNet-1271.50 ± 0.8086.00 ± 0.5042.60 ± 0.7059.10 ± 0.60
SUN [11]ECCV'22ViT-S78.37 ± 0.4688.84 ± 0.32--
FewTURE [19]NeurIPS'22Swin-T77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
Meta-AdaM [45]NeurIPS'23ResNet-12--41.12 ± 0.4956.14 ± 0.49
SP-CLIP [5]CVPR'23Visformer-T82.18 ± 0.4088.24 ± 0.3248.53 ± 0.3861.55 ± 0.41
SemFewOursResNet-1283.65 ± 0.7087.66 ± 0.6054.36 ± 0.7162.79 ± 0.74
SemFew-TransOursSwin-T84.34 ± 0.6789.11 ± 0.5454.27 ± 0.7765.02 ± 0.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.325, + 0.848, + 0.339 + ], + "angle": 0, + "content": "Table 2. Results (%) on CIFAR-FS and FC100. The ± shows 95% confidence intervals. The best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.353, + 0.47, + 0.515 + ], + "angle": 0, + "content": "
MethodVenueCUBPlaces
1-shot5-shot1-shot5-shot
GNN [43]ICLR'1845.6962.2553.1070.84
S2M2 [32]WACV'2048.2470.44--
FT [51]ICLR'2047.4766.9855.7773.94
ATA [55]IJCAI'2145.0066.2253.5775.48
AFA [21]ECCV'2246.8668.2554.0476.21
StyleAdv [14]CVPR'2348.4968.7258.5877.73
LDP-net [64]CVPR'2349.8270.3953.8272.90
SemFew-NameOurs57.5872.2663.2274.54
SemFewOurs59.0772.4764.0174.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.524, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Table 3. Average results (\\%) on cross-domain scenarios. SemFew-Name denotes that semantics are class names." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.28, + 0.579 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.586, + 0.295, + 0.601 + ], + "angle": 0, + "content": "4.3.1 Few-Shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.807 + ], + "angle": 0, + "content": "The comparative results of our proposed approach against previous methods across four benchmarks are shown in Tabs. 1 and 2. Specifically, in the 1-shot task, SemFew exhibits notable superiority, surpassing SP-CLIP [5] \\(0.93\\% - 5.83\\%\\) on four benchmarks. Recent studies [5, 11, 19] have explored different Transformer architectures as backbones. Consequently, we conduct experiments with a Transformer backbone, resulting in SemFew-Trans. This variant further elevates performance, outperforming SP-CLIP by \\(2.16\\% - 6.63\\%\\) in the 1-shot setting. In summary, SemFew-Trans achieves state-of-the-art results compared to both previous visual-based and semantic-based approaches across all 1-shot and 5-shot tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.33, + 0.845 + ], + "angle": 0, + "content": "4.3.2 Cross-Domain Classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.9 + ], + "angle": 0, + "content": "To evaluate whether semantics serve as an effective bridge across novel tasks, we test SemFew on completely novel datasets. Such a task is also known as Cross-Domain Few-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.895, + 0.552 + ], + "angle": 0, + "content": "Shot Learning (CDFSL) [51], which involves extending the capacity to recognize novel categories to domains distinct from the training domain. All the results are reported in Tab. 3, where the experimental setup is the same as FT [51]. Tab. 3 indicates that SemFew significantly surpassing the state-of-the-art methods, where it outperforms best compared methods by \\(9.25\\%\\) (1-shot) and \\(1.87\\%\\) (5-shot) on CUB. It also surpasses the lastest method [64] \\(10.19\\%\\) (1-shot) and \\(1.80\\%\\) (5-shot) on Places. Conversely, SemFew-Name, which adopts a template (i.e., The Photo of a bird called {Class Name}) as the semantic, performs worse than SemFew. This observation highlights the inadequacy of using solely names as semantics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.562, + 0.665, + 0.576 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.631 + ], + "angle": 0, + "content": "We conduct ablation experiments on three datasets, i.e., MiniImageNet, CIFAR-FS, and FC100, and all experiments are performed using the ResNet-12 network architecture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.681, + 0.665 + ], + "angle": 0, + "content": "4.4.1 Alignment Source" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given that our framework is primarily responsible for translating visual samples \\( V \\) and semantics descriptions \\( S \\) into ground-truth class prototype (i.e., \\( V + S \\Rightarrow C \\)), we conduct ablation studies by removing \\( S \\) or \\( V \\). The results can be found in Tab. 4. First, we remove semantics and force pure visual samples to reconstruct prototypes (i.e., \\( V \\Rightarrow C \\)). The result shows that it gets the worst performance, indicating that reconstructing a prototype from pure visual features is challenging. Then, we abandon visual features to implement our proposed principle in Sec. 3.3 (i.e., \\( S \\Rightarrow C \\)). As we can see, it achieves substantial performance gains, reflecting the rich class representative features within semantics. However, our method combining visual and semantic acquires the best performance, which indicates the complementarity between visual and semantic modalities does en" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28593" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.09, + 0.865, + 0.187 + ], + "angle": 0, + "content": "
Method5-way 1-shot5-way 5-shot
MiniImageNetCIFAR-FSFC100MiniImageNetCIFAR-FSFC100
V ⇒ C66.27 ± 0.8174.43 ± 0.9044.72 ± 0.7581.70 ± 0.5586.74 ± 0.6160.58 ± 0.76
S ⇒ C76.77 ± 0.6682.27 ± 0.7152.12 ± 0.7382.92 ± 0.4887.42 ± 0.6062.22 ± 0.75
V + S ⇒ C77.63 ± 0.6383.65 ± 0.7054.36 ± 0.7183.04 ± 0.4887.66 ± 0.6062.79 ± 0.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.198, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Table 4. Results (%) on different alignment sources. \\(V\\) means visual features, \\(S\\) represents semantic features, \\(C\\) stands for class prototypes." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.236, + 0.47, + 0.289 + ], + "angle": 0, + "content": "
CenterMiniImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Cluster77.1682.9478.3685.1983.3987.6153.3062.22
Mean77.6383.0478.9685.8883.6587.6654.3662.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.3, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Table 5. Average results (\\%) on different prototypes. Mean represents the mean vector of sample clusters. Cluster means the prototype is determined by Clustering Algorithm." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.356, + 0.47, + 0.418 + ], + "angle": 0, + "content": "
classifierMinImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
LR77.5582.7078.6684.7283.0886.7051.5357.49
EU77.5683.1578.6585.2682.8386.8754.1462.54
CO77.6383.0478.9685.8883.6587.6654.3662.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.428, + 0.47, + 0.47 + ], + "angle": 0, + "content": "Table 6. Average results \\((\\%)\\) on different classifiers. LR: linear logistic regression classifier. EU: euclidean-distance nearest prototype classifier. CO: cosine-distance nearest prototype classifier." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.494, + 0.421, + 0.51 + ], + "angle": 0, + "content": "hance the robustness of the reconstructed prototype." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.429, + 0.546 + ], + "angle": 0, + "content": "4.4.2 Prototype and Classifier Selection Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Tab. 5 displays the performance associated with various learning targets. Results reveal that employing the mean vector as a prototype yields slightly better results compared to prototypes identified by the clustering algorithm [17]. This suggests that the mean vector consolidates representative visual information, making it a highly effective class prototype. As for the classification head design, we present the results in Tab. 6, which indicates the use of cosine distance as the metric to achieve nearest neighbor classification gets the best results, euclidean distance gets similar but lower outcomes, and linear logistic classifier gets the worst performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.267, + 0.77 + ], + "angle": 0, + "content": "4.4.3 Semantic Evolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We use different kinds of semantics and text encoders to evaluate the influence of semantics, where results can be found in Fig. 4. In most scenarios, our model attains optimal results when aided by paraphrased semantics, and the model trained on simple definitions secures the second-best performance. This trend is reasonable since paraphrased descriptions contain more knowledge than simple definitions and class names. Between the comparison of different text" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.236, + 0.695, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.349, + 0.668, + 0.361 + ], + "angle": 0, + "content": "(a) MiniImageNet, 1-shot" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.363, + 0.695, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.475, + 0.658, + 0.486 + ], + "angle": 0, + "content": "(c) CIFAR-FS, 1-shot" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.487, + 0.695, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.601, + 0.648, + 0.611 + ], + "angle": 0, + "content": "(e) FC100, 1-shot" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.236, + 0.89, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.349, + 0.862, + 0.361 + ], + "angle": 0, + "content": "(b) MiniImageNet, 5-shot" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.362, + 0.889, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.475, + 0.852, + 0.486 + ], + "angle": 0, + "content": "(d) CIFAR-FS, 5-shot" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.487, + 0.89, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.6, + 0.841, + 0.611 + ], + "angle": 0, + "content": "(f) FC100, 5-shot" + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.625, + 0.854, + 0.639 + ], + "angle": 0, + "content": "Figure 4. Average results \\((\\%)\\) on different semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.774 + ], + "angle": 0, + "content": "encoders, the model trained on CLIP encoded semantics greatly outperforms the model trained on BERT encoded semantics. This is because CLIP is trained in a multi-modal manner, rendering it better capable of understanding images. Although a multi-modal text encoder has been widely adopted in FSL research, our model obtains competitive results on BERT, demonstrating that uni-modal semantics can also achieve excellent results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.707, + 0.803 + ], + "angle": 0, + "content": "4.5. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We visualize visual space features through the t-SNE algorithm [52], where the results are presented in Fig. 5. We select three novel classes on the MiniImageNet dataset, and then randomly sampled three support samples and one hundred query samples, where the distribution of support and query samples in visual space can be found in Figs. 5a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.273, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.234, + 0.207, + 0.247 + ], + "angle": 0, + "content": "(a) Support" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.088, + 0.48, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.234, + 0.41, + 0.247 + ], + "angle": 0, + "content": "(b) Query" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.088, + 0.688, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.234, + 0.64, + 0.247 + ], + "angle": 0, + "content": "(c) Real Prototype" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.089, + 0.895, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.234, + 0.873, + 0.247 + ], + "angle": 0, + "content": "(d) Reconstructed Prototype" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.894, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Visualization results on the MiniImageNet dataset. Different colors or shapes represent different classes. The \\(\\star\\) represents the class prototypes, and the \\(\\diamond\\) denotes the prototypes reconstructed by our method." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.304, + 0.273, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.417, + 0.245, + 0.43 + ], + "angle": 0, + "content": "(a) MiniImageNet,CLIP" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.304, + 0.47, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.417, + 0.439, + 0.43 + ], + "angle": 0, + "content": "(b) MiniImageNet, BERT" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.43, + 0.274, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.543, + 0.235, + 0.555 + ], + "angle": 0, + "content": "(c) CIFAR-FS,CLIP" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.43, + 0.47, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.543, + 0.429, + 0.555 + ], + "angle": 0, + "content": "(d) CIFAR-FS, BERT" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.556, + 0.274, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.668, + 0.224, + 0.681 + ], + "angle": 0, + "content": "(e) FC100, CLIP" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.556, + 0.47, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.668, + 0.418, + 0.68 + ], + "angle": 0, + "content": "(f) FC100, BERT" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.693, + 0.446, + 0.708 + ], + "angle": 0, + "content": "Figure 6. Average results \\((\\%)\\) on different fusion factor \\(k\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "and 5b. In Fig. 5c, we put support samples, i.e., red dots, mean vector of each class, i.e., blue stars, and query samples in the same picture. As we can see, support samples are close to the cluster edge rather than the mean center, which leads to difficulties in correctly classifying most samples in the query set. However, after being transformed by SemFew, the reconstructed prototypes are closer to the mean center of each cluster compared to the support set samples. This phenomenon demonstrates that our method can reasonably transform visual samples into class prototypes corresponding to their respective categories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.715, + 0.323 + ], + "angle": 0, + "content": "4.6. Fusion Factor Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.527 + ], + "angle": 0, + "content": "The fusion factor, denoted as \\( k \\) in Eq. (6), significantly influences performance. To demonstrate, we systematically vary the value of \\( k \\) between 0 and 1, using a step size of 0.01, and present the accuracy trends across three benchmarks in Fig. 6. Notably, our model, i.e., blue line, gets the highest accuracy when \\( k \\) is large. Given that \\( k \\) represents the percentage of the reconstructed prototype \\( r_t \\) during fusion in Eq. (6), the experiment highlights that our model produces high-quality prototypes. This observation is further supported by the accuracy comparison when \\( k = 1 \\). In such cases, our model consistently outperforms others, providing strong evidence for the effectiveness of our reconstructed prototypes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.893, + 0.68 + ], + "angle": 0, + "content": "Additionally, we add \\( S \\Rightarrow V \\) experiment for comparison, which generates visual samples rather than prototypes through a Conditional Generative Adversarial Network [34]. This experiment allows us to explore whether we can build robust prototypes indirectly by expanding the sample of the support set. As we can see, \\( S \\Rightarrow V \\), i.e., yellow line, gets relatively low performance in all settings. The experiments indicate that producing the prototype by expanding then averaging the support set is not as effective as directly generating class prototypes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.62, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.854 + ], + "angle": 0, + "content": "To unleash the potential of semantics in FSL, we propose the Semantic Evolution to automatically generate high-quality semantics from simple class names. Then, we design a simple Semantic Alignment Network to transform the concatenated visual features and high-quality semantics features into robust class prototypes for classification. The experimental results show that with high-quality semantics, the basic network can easily achieve greater performance compared to other state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by 2023YFF1204901, NSFC-62076172, 2021DQ02-0903, 2023YFG0116." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28595" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.471, + 0.254 + ], + "angle": 0, + "content": "[3] Irving Biederman. Recognition-by-components: a theory of human image understanding. *Psychol Rev*, 94(2):115, 1987. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[5] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. In CVPR, pages 23581-23591, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[6] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In ICCV, pages 9062-9071, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[7] Hao Cheng, Siyuan Yang, Joel Tianyi Zhou, Lanqing Guo, and Bihan Wen. Frequency guidance matters in few-shot learning. In ICCV, pages 11814-11824, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[10] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. In NeurIPS, pages 21981-21993, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[11] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In ECCV, pages 329-347. Springer, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[12] Thomas Elsken, Benedikt Staffler, Jan Hendrik Metzen, and Frank Hutter. Meta-learning of neural architectures for few-shot learning. In CVPR, pages 12365–12375, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Int. Conf. Mach. Learn., pages 1126-1135. PMLR, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[14] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In CVPR, pages 24575-24584, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In ICCV, pages 8059-8068, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[16] Fusheng Hao, Fengxiang He, Liu Liu, Fuxiang Wu, Dacheng Tao, and Jun Cheng. Class-aware patch embedding adaptation for few-shot image classification. In ICCV, pages 18905-18915, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. J R Stat Soc Ser C Appl Stat, 28(1):100-108, 1979. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[19] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. NeurIPS, 35:3582-3595, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[20] Ruibing Hou, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot classification. In NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.266, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[21] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[22] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In CVPR, pages 11719-11727, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[23] Jaekyeom Kim, Hyoungseok Kim, and Gunhee Kim. Model-agnostic boundary-adversarial sampling for test-time generalization in few-shot learning. In ECCV, pages 599–617. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.45 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.454, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[25] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Deep Learn. workshop. Lille, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In CVPR, pages 10657-10665, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[28] Aoxue Li, Weiran Huang, Xu Lan, Jiashi Feng, Zhenguo Li, and Liwei Wang. Boosting few-shot learning with adaptive margin loss. In CVPR, pages 12576-12584, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[29] Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features for few-shot learning by category traversal. In CVPR, pages 1-10, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[31] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, Georgia, USA, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[32] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In IEEE Win. Conf. Appl. Comput. Vis., pages 2218-2227, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.843, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[33] George A Miller. Wordnet: a lexical database for english. Commun. ACM, 38(11):39-41, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28596" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "[35] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. arXiv preprint arXiv:1707.03141, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.162 + ], + "angle": 0, + "content": "[36] OpenAI. Chatgpt. https://chat.openai.com, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[37] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[38] Zhimao Peng, Zechao Li, Junge Zhang, Yan Li, Guo-Jun Qi, and Jinhui Tang. Few-shot image recognition with knowledge transfer. In ICCV, pages 441-449, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763. PMLR, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[40] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[42] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[43] Victor Garcia Satorras and Joan Bruna Estrach. Few-shot learning with graph neural networks. In ICLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[44] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, pages 8247-8255, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[45] Hongyang Gao Siyuan Sun. Meta-adam: A meta-learned adaptive optimizer with momentum for few-shot learning. In NeurIPS, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[46] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[47] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[48] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199–1208, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[49] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, pages 266-282. Springer, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[50] Pavel Tokmakov, Yu-Xiong Wang, and Martial Hebert. Learning compositional representations for few-shot recognition. In ICCV, pages 6372-6381, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[51] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. J Mach Learn Res, 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In NeurIPS, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[55] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In IJCAI, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[56] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, pages 1831-1840, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[57] Chen Xing, Nagar Rostamzadeh, Boris Oreshkin, and Pedro O O Pinheiro. Adaptive cross-modal few-shot learning. In NeurIPS, 2019. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[58] Jingyi Xu and Hieu Le. Generating representative samples for few-shot classification. In CVPR, pages 9003-9013, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[59] Kun Yan, Zied Bouraoui, Ping Wang, Shoaib Jameel, and Steven Schockaert. Aligning visual prototypes with bert embeddings for few-shot learning. In ICMR, pages 367-375, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[60] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, pages 8808-8817, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In CVPR, pages 3754-3762, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[62] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Differentiable earth mover's distance for few-shot learning. IEEE TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.606, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[64] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In CVPR, pages 20061-20070, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[65] Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast context adaptation via meta-learning. In Int. Conf. Mach. Learn., pages 7693-7702. PMLR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28597" + } + ] +] \ No newline at end of file diff --git a/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_origin.pdf b/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e5be3301735e6a3a487e565de7827b2e16b24dda --- /dev/null +++ b/2024/Simple Semantic-Aided Few-Shot Learning/f104bef6-8837-4adc-8ad9-f16490a614a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e949674deb6788922a7d43374d4027514c447f049b9ee728b52ef9b955c59c1e +size 894227 diff --git a/2024/Simple Semantic-Aided Few-Shot Learning/full.md b/2024/Simple Semantic-Aided Few-Shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b974daa4d9a65f775edbf7e5ac87bd0a0d3cf01a --- /dev/null +++ b/2024/Simple Semantic-Aided Few-Shot Learning/full.md @@ -0,0 +1,335 @@ +# Simple Semantic-Aided Few-Shot Learning + +Hai Zhang $^{1*}$ Junzhe Xu $^{1,2*‡}$ Shanlin Jiang $^{3}$ Zhenan He $^{1\dagger}$ \ + $^{1}$ College of Computer Science, Sichuan University\ + $^{2}$ Banma Network Technology, Alibaba Group\ + $^{3}$ Naveen Jindal School of Management, University of Texas at Dallas\ +zhanghi@stu.scu.edu.cn, xujunzhe.xjz@alibaba-inc.com\ +shanlin.jiang@utdallas.edu, zhenan@scu.edu.cn + +# Abstract + +Learning from a limited amount of data, namely Few-Shot Learning, stands out as a challenging computer vision task. Several works exploit semantics and design complicated semantic fusion mechanisms to compensate for rare representative features within restricted data. However, relying on naive semantics such as class names introduces biases due to their brevity, while acquiring extensive semantics from external knowledge takes a huge time and effort. This limitation severely constrains the potential of semantics in Few-Shot Learning. In this paper, we design an automatic way called Semantic Evolution to generate high-quality semantics. The incorporation of high-quality semantics alleviates the need for complex network structures and learning algorithms used in previous works. Hence, we employ a simple two-layer network termed Semantic Alignment Network to transform semantics and visual features into robust class prototypes with rich discriminative features for few-shot classification. The experimental results show our framework outperforms all previous methods on six benchmarks, demonstrating a simple network with high-quality semantics can beat intricate multi-modal modules on few-shot classification tasks. Code is available at https://github.com/zhangdoudou123/SemFew. + +# 1. Introduction + +Deep learning models have witnessed substantial advancements by leveraging extensive annotated data [8]. Nevertheless, in numerous real-world scenarios, the shortage of labeled data restricts the applicability of conventional deep learning approaches. In contrast, humans possess a remarkable cognitive ability to learn new concepts and recognize + +categories from just a few examples [3]. Motivated by this, Few-Shot Learning (FSL) [46] is proposed to mimic the ability of humans to learn from a few labeled samples. + +In a typical setting of FSL [46], the support set, which consists of $N$ novel classes and each class contains $K$ samples, is provided for the model. The model learned on the support set is required to accurately classify test samples, which are labeled as the query set, into one of the $N$ categories. A conventional classification approach involves projecting both support and query sets onto a pre-established metric space. Then, query images are classified by finding the nearest support image. However, few support samples may not contain sufficient discriminative features for recognition. These samples are typically located on the periphery of the sample cluster in the metric space, resulting in unstable distance evaluations and inaccurate classification. + +To address this issue, researchers start from visual-based methods [7, 11, 13, 16, 19, 29, 32, 37, 46, 49, 53, 60], aiming at extracting class-related features from periphery image, reducing the intra-class variation among samples, and thus constructing robust features for classification. Although visual-based methods have undergone extensive investigation and achieved great success, this type of approach still struggles to handle situations when there is only one hard periphery sample with minimal semantic features per class, i.e., the one-shot learning task [25, 53]. Hence, some researchers turned to applying semantics as auxiliary information to help the model better understand periphery images, leveraging the synergies between language and vision modalities. + +From this perspective, semantic-based methods [5, 28, 38, 44, 50, 56-59, 61] dedicated to exploring different types of semantics to enhance periphery samples. One kind of method [57-59] applies the class name as the semantic source. However, class name is not the best way to understand a novel class. For instance, If a person never saw a zebra, it would be easier to identify a zebra by the definition, i.e., a horse with stripes, than by the name zebra. + +The American robin is renowned for its striking appearance, with a rust-red to orange breast and abdomen, a dark grayish-blue upper body, and a white eye ring. This bird plays a pivotal ecological role as a seed disperser, primarily for fruits and berries,... + +![](images/71f35d14f53a0d0f7579004a8d0e8a8c1906588581f2cf88234f41ca965cb527.jpg) +Figure 1. The introduction of how high-quality semantics reconstruct the class prototype through complementarity between modalities. Periphery image is an image with fewer discriminative features and prototype image represents images with concrete and enriched representative features. + +Meanwhile, the class name contains ambiguity in some scenarios. For example, the class `ear` is easily misidentified as the human organ, but images in MiniImageNet [53] belong to this class are actually photos of corn. Another kind of semantic exploits discrete attribute labels as the substitution of mere class names and achieves good performance [61]. Nevertheless, the collection process is time-consuming and expert assistance requiring. Moreover, discrete attribute labels lack robust representation and struggle to benefit from context. Therefore, effectively collecting and leveraging high-quality semantics becomes an urgent research need. + +In this paper, we propose a simple framework called Semantic-Aided Few-Shot Learning (SemFew). To begin with, we identify the vital role of high-quality semantics in FSL and innovatively propose an automatic step-by-step Semantic Evolution (SemEvo) process to acquire detailed and accurate semantics. It first converts the class name into a short description that matches the image content of each class. Next, to further enrich recognizable characteristics, we expand and paraphrase short descriptions. The paraphrased descriptions contain more class-related knowledge than class names and short descriptions, making them better encompass a vast array of visual features. To illustrate, Fig. 1 showcases an instance that the paraphrased semantic contains a large number of detailed descriptions matching visual features e.g. organ characteristics and eating habitat. With this information, the model can better reconstruct an image with periphery feature (left image), into an image with prototype feature (right image), which benefits few-shot classification. Then, to fully exploit the advantage of paraphrased semantics, we design a network called Semantic Alignment Network (SemAlign). Different from previous methods, SemAlign does not adopt complex semantic understanding modules but utilizes a basic two-layer network. After accepting semantics and visual features + +as input, it reconstructs prototype features for subsequent classification. Experimental results on six widely adopted benchmarks demonstrate the effectiveness of our method, providing a simple insight into future FSL research. Our contributions can be summarized as three points: + +- To the best of our knowledge, we are the first to consider the automatic way of collecting high-quality semantics and applying them in FSL. +- We design a simple and efficient way to translate high-quality semantics and visual features into prototypes, without any intricate semantic understanding modules. +- Our approach achieves state-of-the-art performance across six benchmarks in FSL research, underscoring that a basic network can obtain excellent performance when supported by high-quality semantics. + +# 2. Related Work + +The significant issue in FSL is how to effectively and accurately extract class-related features from limited data. Targeting this issue, existing research explored two different types of approaches, i.e., visual-based methods and semantic-based methods. + +Visual-Based Methods focus on extracting class-related features from images for classification. There are two main types of this approach. The approaches with the optimization principle behind [1, 13, 40] concentrated on acquiring a set of initial model parameters capable of rapid adaptation to novel classes. However, updating the entire model with a limited amount of labeled data may result in metaoverfitting [12, 22, 35, 42, 65]. The approaches considered metric design aiming at training a metric space, where inter-class distances are maximized, while intra-class distances are minimized. Within this type of approach, one attracting method focuses on integrating self-supervised loss to pretrain the backbone network [10, 15, 19, 32, 49]. Other methods are dedicated to improving the metric, and previous research spanning from initially manually designed metrics [2, 25, 46, 62] to subsequently model-based metrics [48, 53], all of which have yielded encouraging outcomes. + +Semantic-Based Methods sought to enhance visual recognition performance by fusing complementary information from both visual and textual modalities [5, 28, 57, 58, 61], and related research introducing intricate network frameworks to effectively utilize textual information. Chen et al. [57] introduced an adaptive fusion mechanism to integrate visual prototypes with semantic prototypes obtained through word embeddings of class labels. Zhang et al. [61] designed a framework for prototype completion, leveraging rare attribute-level information from labels to construct representative class prototypes. Xu and Le [58] proposed the use of the Conditional Variational Autoencoder (CVAE) [47] model to generate visual features based on semantic embeddings for further classification. Li et al. [28] pre + +sented a task-relevant adaptive margin loss based on the semantic similarity between class names. Chen et al. [5] devised complementary mechanisms for inserting semantic prompts into the feature extractor, allowing the feature extractor to better focus on class-specific features. + +Previous research emphasized the exploitation of class names and designed a complicated network to integrate semantics with visual information. In contrast, our primary concern lies in the inherent quality of semantics, and how it can relieve the burden of designing intricate network structures and learning algorithms. Remarkably, we exhibit that high-quality semantics information has the ability to yield superior outcomes even within a simple network. + +# 3. Methodology + +This section first revisits the background of FSL, then introduces two components of SemFew: explaining how semantics are gradually improved through SemEvo, and demonstrating the details of our proposed SemAlign. + +# 3.1. Background + +The dataset of FSL includes two components: a base set $\mathcal{D}_{base} = \{(x,y)|x\in \mathcal{X}_{base},y\in \mathcal{C}_{base}\}$ for training a metric space and a novel set $\mathcal{D}_{novel} = \{(x,y)|x\in \mathcal{X}_{novel},y\in \mathcal{C}_{novel}\}$ for testing, where $x$ denotes the image and $y$ represents the label. It is important to note that the label space for both sets is disjoint, i.e., $\mathcal{C}_{base}\cap \mathcal{C}_{novel} = \emptyset$ . During the testing process, the support set $S = \{(x_i,y_i)\}_{i = 0}^{N\times K}$ is randomly sampled from $\mathcal{D}_{novel}$ , which contains $N$ classes with each class consisting of $K$ samples. Then, the model is required to correctly classify images in the query set $\mathcal{Q} = \{(x_i,y_i)\}_{i = 0}^{N\times M}$ into one of $N$ classes in the support set $S$ , where $M$ is the number of query samples in each class. Generally, this type of classification task is named as $N$ -way $K$ -shot task. + +A simple but effective FSL method is to calculate the prototype of each support class, and then classify query samples by finding the nearest prototype in the metric space. To be specific, the prototype of class $t$ , labeled as $p_t$ , is the mean vector of all samples in the support set belonging to class $t$ : + +$$ +p _ {t} = \frac {1}{\left| \mathcal {S} _ {t} \right|} \sum_ {x _ {i} \in \mathcal {S} _ {t}} f \left(x _ {i}\right), \tag {1} +$$ + +where $S_{t}$ is a set of all samples belonging to class $t$ in the support set $S$ and $f$ denotes the backbone network. Then, the probability of a query sample $q$ belonging to class $t$ is related to the distance between $q$ and prototype of class $t$ [46], which is calculated as follows: + +$$ +P (y = t) = \frac {\exp d \left(q , p _ {t}\right)}{\sum_ {i = 1} ^ {N} \exp d \left(q , p _ {i}\right)}, \tag {2} +$$ + +where $d(\cdot, \cdot)$ means distance metric, e.g., cosine distance, and $N$ denotes the number of classes. + +![](images/5b24ad8af9db3e7a6085fff5b1863d6db76c335e90d50d2aa121387db69c6079.jpg) +Figure 2. The illustration of how Semantic Evolution converts the simple name and the definition into the high-quality description. + +# 3.2. Semantic Evolution + +This section explains how we automatically generate high-quality semantics for few-shot classification by the designed process termed Semantic Evolution. This process first converts class names to short descriptions and then paraphrases them into high-quality semantics. + +As we illustrated before, adopting class names as semantics [5, 39, 58] unavoidably introduces ambiguity. Therefore, we first retrieve the definition of each class name from WordNet [33] that matches image content, where definitions are more detailed and concrete than mere class names. However, even though definitions are able to precisely describe a class to some extent, brief descriptions may still overlook key visual information, due to the fact that visual features are highly variable. + +To comprehensively conclude variable visual features, we leverage the extensive knowledge stored in the pretrained Large Language Models (LLMs) [4] to augment definitions. The prompt we used is: $\{\text{WordNet Definition}\}$ is the definition of the $\{\text{Class Name}\}$ . Please rewrite and expand this definition to make it more detailed and consistent with scientific fact. Briefness is required, using only one paragraph. The illustration of the Semantic Evolution process is shown in Fig. 2. + +The paraphrased class description generated by LLMs precisely describes the corresponding class itself, which does not apply to other categories. This characteristic ensures that semantic descriptions do not leak information to new categories, which follows the paradigm of FSL. + +# 3.3. Semantic Alignment Network + +After obtaining paraphrased semantics in the previous section, we devise the Semantic Alignment Network (SemAlign), aiming at translating hard periphery samples into robust class prototypes. + +![](images/ac749bd011964ce9970d2992531a46ccc7543a6aec1a5d9cdfc48ed5d932d7a0.jpg) +Figure 3. The framework of our proposed SemFew. During the training stage, images and paraphrased semantics are encoded and fed into SemAlign $h$ , with the objective of reducing the distance between the output of $h$ and the class prototype in the visual space. During the testing stage, images in the support set are transformed into class prototypes by $h$ , and query images are classified by identifying the nearest prototype. The symbol $\oplus$ denotes a concatenation operation. + +For a cluster of samples, the cluster center best encapsulates the characteristic of a category and has been regarded as a prototype in previous work [46]. Suppose $S$ represents semantic features, and $C$ stands for the center of each class, the simplest implementation of SemAlign is to learn the alignment between semantics $S$ and centers $C$ : + +$$ +\min _ {W _ {1}, W _ {2}} L \left(\sigma \left(S ^ {\top} W _ {1}\right) W _ {2}, C\right), \tag {3} +$$ + +where $W_{1}$ and $W_{2}$ are parameters of network, $\sigma$ is an activation function. + +Considering that each modality views things from a different and complementary perspective, this phenomenon can be reasonably utilized to obtain better prototypes. Therefore, we fuse visual features with high-quality semantics, and then align multi-modal output with the cluster center, the details of the whole pipeline can be found in Fig. 3. Specifically, suppose the $i$ -th image in the dataset is represented as $x_{i}$ , its label and semantic are denoted as $y_{i}$ and $s_i$ , respectively. First, $x_{i}$ is encoded by a vision encoder $f$ , which is pre-trained on the base set. Also, semantic description $s_i$ is encoded by an off-the-shelf text encoder $g$ . Then, the encoded visual and text features are concatenated together and fed into SemAlign $h$ to reconstruct the prototype. The training loss supervises the distance between the reconstructed prototype towards the cluster center: + +$$ +\min _ {W _ {1}, W _ {2}} \mathbb {E} [ \| h (f (x _ {i}), g (s _ {i})) - c _ {y} \| _ {1} ], \tag {4} +$$ + +where $h(f(x_i), g(s_i)) = \sigma([f(x_i) \cdot g(s_i)]^\top W_1)W_2$ is SemAlign with learnable weights $W_1$ and $W_2$ , $[\cdot]$ means concatenation operation. + +During the testing process, a $N$ -way $K$ -shot support set is randomly sampled in the novel set. For class $t$ , we obtain its paraphrased description $s_t$ by Semantic Evolution proposed in Sec. 3.2. Then, support image $x$ and description $s_t$ are encoded by $f$ and $g$ , respectively, and fed into $h$ to get the reconstructed prototype: + +$$ +r _ {t} = \frac {1}{K} \sum_ {i = 1} ^ {K} h \left(f \left(x _ {i}\right), g \left(s _ {t}\right)\right), \tag {5} +$$ + +where $r_t$ is the reconstructed prototype. To ensure the reconstructed prototype is further aligned with the groundtruth class prototype, it will be fused with support images in a convex combination manner [57]: + +$$ +p _ {t} = k r _ {t} + (1 - k) u _ {t}, \tag {6} +$$ + +where $p_t$ means the classification prototype of class $t$ , $u_t = \frac{1}{K} \sum_{i=1}^{K} f(x_i)$ is the mean vector of support samples, $r_t$ is the reconstructed prototype produced by SemAlign, and $k \in [0,1]$ is a manually controlled fusion factor. Then, Eq. (2) is utilized to assign label for each query sample, where $d(\cdot, \cdot)$ is set as cosine distance. + +
MethodVenueBackboneMiniImageNetTieredImageNet
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
Visual-BasedMatchNet [53]NeurIPS'16ResNet-1265.64 ± 0.2078.72 ± 0.1568.50 ± 0.9280.60 ± 0.71
ProtoNet [46]NeurIPS'17ResNet-1262.39 ± 0.2180.53 ± 0.1468.23 ± 0.2384.03 ± 0.16
MAML [13]ICML'17ResNet-1249.24 ± 0.2158.05 ± 0.1067.92 ± 0.1772.41 ± 0.20
TADAM [37]NeurIPS'18ResNet-1258.50 ± 0.3076.70 ± 0.3062.13 ± 0.3181.92 ± 0.30
CAN [20]NeurIPS'19ResNet-1263.85 ± 0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
CTM [29]CVPR'19ResNet-1864.12 ± 0.8280.51 ± 0.1368.41 ± 0.3984.28 ± 1.73
RFS [49]ECCV'20ResNet-1262.02 ± 0.6379.64 ± 0.4469.74 ± 0.7284.41 ± 0.55
FEAT [60]CVPR'20ResNet-1266.78 ± 0.2082.05 ± 0.1470.80 ± 0.2384.79 ± 0.16
Meta-Baseline [6]ICCV'21ResNet-1263.17 ± 0.2379.26 ± 0.1768.62 ± 0.2783.29 ± 0.18
SUN [11]ECCV'22ViT-S67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [19]NeurIPS'22Swin-T72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
FGFL [7]ICCV'23ResNet-1269.14 ± 0.8086.01 ± 0.6273.21 ± 0.8887.21 ± 0.61
Meta-AdaM [45]NeurIPS'23ResNet-1259.89 ± 0.4977.92 ± 0.4365.31 ± 0.4885.24 ± 0.35
Semantic-BasedKTN [38]ICCV'19Conv-12864.42 ± 0.7274.16 ± 0.5674.16 ± 0.56-
AM3 [57]NeurIPS'19ResNet-1265.30 ± 0.4978.10 ± 0.3669.08 ± 0.4782.58 ± 0.31
TRAML[28]CVPR'20ResNet-1267.10 ± 0.5279.54 ± 0.60--
AM3-BERT [59]ICMR'21ResNet-1268.42 ± 0.5181.29 ± 0.5977.03 ± 0.8587.20 ± 0.70
SVAE-Proto [58]CVPR'22ResNet-1274.84 ± 0.2383.28 ± 0.4076.98 ± 0.6585.77 ± 0.50
SP-CLIP [5]CVPR'23Visformer-T72.31 ± 0.4083.42 ± 0.3078.03 ± 0.4688.55 ± 0.32
SemFewOursResNet-1277.63 ± 0.6383.04 ± 0.4878.96 ± 0.8085.88 ± 0.62
SemFew-TransOursSwin-T78.94 ± 0.6686.49 ± 0.5082.37 ± 0.7789.89 ± 0.52
+ +Table 1. Results (%) on MiniImageNet and TieredImageNet. The ± shows 95% confidence intervals. The best results are shown in bold. + +# 4. Experiments + +We exhibit experimental results in this section, including benchmarks introduction, implementation details, and few-shot classification performance. + +# 4.1. Datasets + +The study evaluates the proposed method on six established FSL datasets: MiniImageNet [53], TieredImageNet [41], CIFAR-FS [27], FC100 [37], Places [63], and CUB [54]. MiniImageNet and TieredImageNet are both subsets of ImageNet [8]. MiniImageNet consists of 64 training classes, 16 validating classes, and 20 testing classes. TieredImageNet encompasses 351 training classes, 97 validating classes, and 160 testing classes. CIFAR-FS and FC100 are derived from the CIFAR-100 [26]. CIFAR-FS employs a random partitioning strategy with 64 training classes, 16 validating classes, and 20 testing classes. FC100 introduces a unique superclass partitioning approach, where the training set comprises 12 subclasses, namely 60 classes, while both the validation and test sets include four subclasses, totaling 20 classes. Places and Caltech-UCSD Birds-200-2011 (CUB) are datasets for cross-domain scenario testing. In a cross-domain evaluation [51], the model is initially trained on MiniImageNet's base split and tested on the corresponding test split. + +# 4.2. Implementation Details + +Architecture. We employ a pre-trained ResNet-12 [18] and Swin-T [30] as the vision encoders. Additionally, we apply two types of text encoder: the text encoder from ViT-B/16 CLIP [39] where output dimension is 512, and the BERT-Base text encoder [9] which has 768 output dimension. The network $h$ that combines visual and textual features consists of two fully connected layers and a LeakyReLU [31] activation function. The hidden layer has a dimension of 4,096. In the context of Semantic Evolution, the utilized Large Language Model is GPT-3.5-turbo [36]. + +Training Details. A two-step training procedure is performed in experiments. In step one, two kinds of vision encoder are pre-trained, the ResNet-12 [18] follows the same pretraining settings as the FEAT [60], while the Swin-T [30] adopts the identical pretraining strategy as FewTURE [19]. Next, the network $h$ is trained 50 epochs. This training is conducted with a batch size of 128 and employs an Adam optimizer [24] with a learning rate of 1e-4. + +Evaluation protocol. The proposed method is evaluated under 5-way 1-shot and 5-way 5-shot settings. For each setting, 600 classification tasks are uniformly sampled from the novel set. In each task, there are 15 query samples per class. The mean and $95\%$ confidence interval of the accuracy are reported. + +
MethodVenueBackboneCIFAR-FSFC100
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
ProtoNet [46]NeurIPS'17ResNet-1272.20 ± 0.7083.50 ± 0.5041.54 ± 0.7657.08 ± 0.76
TADAM [37]NeurIPS'18ResNet-12--40.10 ± 0.4056.10 ± 0.40
MetaOptNet [27]CVPR'19ResNet-1272.80 ± 0.7084.30 ± 0.5047.20 ± 0.6055.50 ± 0.60
MABAS [23]ECCV'20ResNet-1273.51 ± 0.9285.65 ± 0.6542.31 ± 0.7558.16 ± 0.78
RFS [49]ECCV'20ResNet-1271.50 ± 0.8086.00 ± 0.5042.60 ± 0.7059.10 ± 0.60
SUN [11]ECCV'22ViT-S78.37 ± 0.4688.84 ± 0.32--
FewTURE [19]NeurIPS'22Swin-T77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
Meta-AdaM [45]NeurIPS'23ResNet-12--41.12 ± 0.4956.14 ± 0.49
SP-CLIP [5]CVPR'23Visformer-T82.18 ± 0.4088.24 ± 0.3248.53 ± 0.3861.55 ± 0.41
SemFewOursResNet-1283.65 ± 0.7087.66 ± 0.6054.36 ± 0.7162.79 ± 0.74
SemFew-TransOursSwin-T84.34 ± 0.6789.11 ± 0.5454.27 ± 0.7765.02 ± 0.72
+ +Table 2. Results (%) on CIFAR-FS and FC100. The ± shows 95% confidence intervals. The best results are shown in bold. + +
MethodVenueCUBPlaces
1-shot5-shot1-shot5-shot
GNN [43]ICLR'1845.6962.2553.1070.84
S2M2 [32]WACV'2048.2470.44--
FT [51]ICLR'2047.4766.9855.7773.94
ATA [55]IJCAI'2145.0066.2253.5775.48
AFA [21]ECCV'2246.8668.2554.0476.21
StyleAdv [14]CVPR'2348.4968.7258.5877.73
LDP-net [64]CVPR'2349.8270.3953.8272.90
SemFew-NameOurs57.5872.2663.2274.54
SemFewOurs59.0772.4764.0174.70
+ +Table 3. Average results (\%) on cross-domain scenarios. SemFew-Name denotes that semantics are class names. + +# 4.3. Experimental Results + +# 4.3.1 Few-Shot Classification + +The comparative results of our proposed approach against previous methods across four benchmarks are shown in Tabs. 1 and 2. Specifically, in the 1-shot task, SemFew exhibits notable superiority, surpassing SP-CLIP [5] $0.93\% - 5.83\%$ on four benchmarks. Recent studies [5, 11, 19] have explored different Transformer architectures as backbones. Consequently, we conduct experiments with a Transformer backbone, resulting in SemFew-Trans. This variant further elevates performance, outperforming SP-CLIP by $2.16\% - 6.63\%$ in the 1-shot setting. In summary, SemFew-Trans achieves state-of-the-art results compared to both previous visual-based and semantic-based approaches across all 1-shot and 5-shot tasks. + +# 4.3.2 Cross-Domain Classification + +To evaluate whether semantics serve as an effective bridge across novel tasks, we test SemFew on completely novel datasets. Such a task is also known as Cross-Domain Few- + +Shot Learning (CDFSL) [51], which involves extending the capacity to recognize novel categories to domains distinct from the training domain. All the results are reported in Tab. 3, where the experimental setup is the same as FT [51]. Tab. 3 indicates that SemFew significantly surpassing the state-of-the-art methods, where it outperforms best compared methods by $9.25\%$ (1-shot) and $1.87\%$ (5-shot) on CUB. It also surpasses the lastest method [64] $10.19\%$ (1-shot) and $1.80\%$ (5-shot) on Places. Conversely, SemFew-Name, which adopts a template (i.e., The Photo of a bird called {Class Name}) as the semantic, performs worse than SemFew. This observation highlights the inadequacy of using solely names as semantics. + +# 4.4. Ablation Studies + +We conduct ablation experiments on three datasets, i.e., MiniImageNet, CIFAR-FS, and FC100, and all experiments are performed using the ResNet-12 network architecture. + +# 4.4.1 Alignment Source + +Given that our framework is primarily responsible for translating visual samples $V$ and semantics descriptions $S$ into ground-truth class prototype (i.e., $V + S \Rightarrow C$ ), we conduct ablation studies by removing $S$ or $V$ . The results can be found in Tab. 4. First, we remove semantics and force pure visual samples to reconstruct prototypes (i.e., $V \Rightarrow C$ ). The result shows that it gets the worst performance, indicating that reconstructing a prototype from pure visual features is challenging. Then, we abandon visual features to implement our proposed principle in Sec. 3.3 (i.e., $S \Rightarrow C$ ). As we can see, it achieves substantial performance gains, reflecting the rich class representative features within semantics. However, our method combining visual and semantic acquires the best performance, which indicates the complementarity between visual and semantic modalities does en + +
Method5-way 1-shot5-way 5-shot
MiniImageNetCIFAR-FSFC100MiniImageNetCIFAR-FSFC100
V ⇒ C66.27 ± 0.8174.43 ± 0.9044.72 ± 0.7581.70 ± 0.5586.74 ± 0.6160.58 ± 0.76
S ⇒ C76.77 ± 0.6682.27 ± 0.7152.12 ± 0.7382.92 ± 0.4887.42 ± 0.6062.22 ± 0.75
V + S ⇒ C77.63 ± 0.6383.65 ± 0.7054.36 ± 0.7183.04 ± 0.4887.66 ± 0.6062.79 ± 0.74
+ +Table 4. Results (%) on different alignment sources. $V$ means visual features, $S$ represents semantic features, $C$ stands for class prototypes. + +
CenterMiniImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Cluster77.1682.9478.3685.1983.3987.6153.3062.22
Mean77.6383.0478.9685.8883.6587.6654.3662.79
+ +Table 5. Average results (\%) on different prototypes. Mean represents the mean vector of sample clusters. Cluster means the prototype is determined by Clustering Algorithm. + +
classifierMinImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
LR77.5582.7078.6684.7283.0886.7051.5357.49
EU77.5683.1578.6585.2682.8386.8754.1462.54
CO77.6383.0478.9685.8883.6587.6654.3662.79
+ +Table 6. Average results $(\%)$ on different classifiers. LR: linear logistic regression classifier. EU: euclidean-distance nearest prototype classifier. CO: cosine-distance nearest prototype classifier. + +hance the robustness of the reconstructed prototype. + +# 4.4.2 Prototype and Classifier Selection Strategy + +Tab. 5 displays the performance associated with various learning targets. Results reveal that employing the mean vector as a prototype yields slightly better results compared to prototypes identified by the clustering algorithm [17]. This suggests that the mean vector consolidates representative visual information, making it a highly effective class prototype. As for the classification head design, we present the results in Tab. 6, which indicates the use of cosine distance as the metric to achieve nearest neighbor classification gets the best results, euclidean distance gets similar but lower outcomes, and linear logistic classifier gets the worst performance. + +# 4.4.3 Semantic Evolution + +We use different kinds of semantics and text encoders to evaluate the influence of semantics, where results can be found in Fig. 4. In most scenarios, our model attains optimal results when aided by paraphrased semantics, and the model trained on simple definitions secures the second-best performance. This trend is reasonable since paraphrased descriptions contain more knowledge than simple definitions and class names. Between the comparison of different text + +![](images/aac258d6026627afec3d05617817eb1fa140f66da681af56a215c01ea3daebc9.jpg) +(a) MiniImageNet, 1-shot + +![](images/b9f8ba4222802d95e8b212a91c57036ac921ec4d66476c9fd68395abdac56354.jpg) +(c) CIFAR-FS, 1-shot + +![](images/0732e6ff299f45084b3b59a268bed241b2d5caa132b551144b24fe98a11cc78a.jpg) +(e) FC100, 1-shot +Figure 4. Average results $(\%)$ on different semantics. + +![](images/0bce1ec77520057a58b9132c2c2eb36e0804e5c0c28afcb7cae2d25a56ac4a42.jpg) +(b) MiniImageNet, 5-shot + +![](images/46af499085228bacf53f45433249ebe175794bac562d87a095b371f055036ae4.jpg) +(d) CIFAR-FS, 5-shot + +![](images/b5178b19cf2cdb301a06e4cb9ea709a7b5101eabd689ba726a4053b139a8c77c.jpg) +(f) FC100, 5-shot + +encoders, the model trained on CLIP encoded semantics greatly outperforms the model trained on BERT encoded semantics. This is because CLIP is trained in a multi-modal manner, rendering it better capable of understanding images. Although a multi-modal text encoder has been widely adopted in FSL research, our model obtains competitive results on BERT, demonstrating that uni-modal semantics can also achieve excellent results. + +# 4.5. Visualization Analysis + +We visualize visual space features through the t-SNE algorithm [52], where the results are presented in Fig. 5. We select three novel classes on the MiniImageNet dataset, and then randomly sampled three support samples and one hundred query samples, where the distribution of support and query samples in visual space can be found in Figs. 5a + +![](images/946115ba043e80bfa1f2164936e653aa653280766b8b4e8f594f2a12b1079258.jpg) +(a) Support + +![](images/260f067000755f8f22ca3e4a24b50ad123eacd792bfe095f3066fab05d9b50c9.jpg) +(b) Query + +![](images/4ba299f6f11c5aab3394802512653b41d24d57df33ef1370aad699f2003e974f.jpg) +(c) Real Prototype + +![](images/28ac5315e67b252f1f1ba15ece9c403d9ce876c550f2031b52a4450b41b94755.jpg) +(d) Reconstructed Prototype + +![](images/d373e0caa9759a29b73689bae818291e4e338c01c941d5b6ab4c682f97c15d3a.jpg) +Figure 5. Visualization results on the MiniImageNet dataset. Different colors or shapes represent different classes. The $\star$ represents the class prototypes, and the $\diamond$ denotes the prototypes reconstructed by our method. + +![](images/2f37cf55549b011c9f554ab973ce627e53bbffc76c1407b9cec4a90bd3b1584a.jpg) + +![](images/ab76e6586d5fdbe335edca0a6d35503d49a6b29cdc6a971f6dbea040ed0e5628.jpg) +(a) MiniImageNet,CLIP + +![](images/c0cfb4a20ced74d64b5cec0bc59a9a404c98f608f3daf84004c038c2b4702128.jpg) +(b) MiniImageNet, BERT + +![](images/fcbbd8bafd95971bf9721b4734ad71b72feaa93bad3fec9fda3a3a24a13bbd05.jpg) +(c) CIFAR-FS,CLIP +(e) FC100, CLIP +Figure 6. Average results $(\%)$ on different fusion factor $k$ + +![](images/374e297665525f1366bed75e7c82c4e4bbf2abb9e9ae9eb526f286bf22771652.jpg) +(d) CIFAR-FS, BERT +(f) FC100, BERT + +and 5b. In Fig. 5c, we put support samples, i.e., red dots, mean vector of each class, i.e., blue stars, and query samples in the same picture. As we can see, support samples are close to the cluster edge rather than the mean center, which leads to difficulties in correctly classifying most samples in the query set. However, after being transformed by SemFew, the reconstructed prototypes are closer to the mean center of each cluster compared to the support set samples. This phenomenon demonstrates that our method can reasonably transform visual samples into class prototypes corresponding to their respective categories. + +# 4.6. Fusion Factor Analysis + +The fusion factor, denoted as $k$ in Eq. (6), significantly influences performance. To demonstrate, we systematically vary the value of $k$ between 0 and 1, using a step size of 0.01, and present the accuracy trends across three benchmarks in Fig. 6. Notably, our model, i.e., blue line, gets the highest accuracy when $k$ is large. Given that $k$ represents the percentage of the reconstructed prototype $r_t$ during fusion in Eq. (6), the experiment highlights that our model produces high-quality prototypes. This observation is further supported by the accuracy comparison when $k = 1$ . In such cases, our model consistently outperforms others, providing strong evidence for the effectiveness of our reconstructed prototypes. + +Additionally, we add $S \Rightarrow V$ experiment for comparison, which generates visual samples rather than prototypes through a Conditional Generative Adversarial Network [34]. This experiment allows us to explore whether we can build robust prototypes indirectly by expanding the sample of the support set. As we can see, $S \Rightarrow V$ , i.e., yellow line, gets relatively low performance in all settings. The experiments indicate that producing the prototype by expanding then averaging the support set is not as effective as directly generating class prototypes. + +# 5. Conclusion + +To unleash the potential of semantics in FSL, we propose the Semantic Evolution to automatically generate high-quality semantics from simple class names. Then, we design a simple Semantic Alignment Network to transform the concatenated visual features and high-quality semantics features into robust class prototypes for classification. The experimental results show that with high-quality semantics, the basic network can easily achieve greater performance compared to other state-of-the-art methods. + +Acknowledgement. This work was supported by 2023YFF1204901, NSFC-62076172, 2021DQ02-0903, 2023YFG0116. + +# References + +[1] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. In NeurIPS, 2016. 2 +[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 2 +[3] Irving Biederman. Recognition-by-components: a theory of human image understanding. *Psychol Rev*, 94(2):115, 1987. 1 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3 +[5] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. In CVPR, pages 23581-23591, 2023. 1, 2, 3, 5, 6 +[6] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In ICCV, pages 9062-9071, 2021. 5 +[7] Hao Cheng, Siyuan Yang, Joel Tianyi Zhou, Lanqing Guo, and Bihan Wen. Frequency guidance matters in few-shot learning. In ICCV, pages 11814-11824, 2023. 1, 5 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 1, 5 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[10] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. In NeurIPS, pages 21981-21993, 2020. 2 +[11] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In ECCV, pages 329-347. Springer, 2022. 1, 5, 6 +[12] Thomas Elsken, Benedikt Staffler, Jan Hendrik Metzen, and Frank Hutter. Meta-learning of neural architectures for few-shot learning. In CVPR, pages 12365–12375, 2020. 2 +[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Int. Conf. Mach. Learn., pages 1126-1135. PMLR, 2017. 1, 2, 5 +[14] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In CVPR, pages 24575-24584, 2023. 6 +[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In ICCV, pages 8059-8068, 2019. 2 +[16] Fusheng Hao, Fengxiang He, Liu Liu, Fuxiang Wu, Dacheng Tao, and Jun Cheng. Class-aware patch embedding adaptation for few-shot image classification. In ICCV, pages 18905-18915, 2023. 1 + +[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. J R Stat Soc Ser C Appl Stat, 28(1):100-108, 1979. 7 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5 +[19] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. NeurIPS, 35:3582-3595, 2022. 1, 2, 5, 6 +[20] Ruibing Hou, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot classification. In NeurIPS, 2019. 5 +[21] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 6 +[22] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In CVPR, pages 11719-11727, 2019. 2 +[23] Jaekyeom Kim, Hyoungseok Kim, and Gunhee Kim. Model-agnostic boundary-adversarial sampling for test-time generalization in few-shot learning. In ECCV, pages 599–617. Springer, 2020. 6 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[25] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Deep Learn. workshop. Lille, 2015. 1, 2 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In CVPR, pages 10657-10665, 2019. 5, 6 +[28] Aoxue Li, Weiran Huang, Xu Lan, Jiashi Feng, Zhenguo Li, and Liwei Wang. Boosting few-shot learning with adaptive margin loss. In CVPR, pages 12576-12584, 2020. 1, 2, 5 +[29] Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features for few-shot learning by category traversal. In CVPR, pages 1-10, 2019. 1, 5 +[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5 +[31] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, Georgia, USA, 2013. 5 +[32] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In IEEE Win. Conf. Appl. Comput. Vis., pages 2218-2227, 2020. 1, 2, 6 +[33] George A Miller. Wordnet: a lexical database for english. Commun. ACM, 38(11):39-41, 1995. 3 +[34] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 8 + +[35] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. arXiv preprint arXiv:1707.03141, 2017. 2 +[36] OpenAI. Chatgpt. https://chat.openai.com, 2023.5 +[37] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 1, 5, 6 +[38] Zhimao Peng, Zechao Li, Junge Zhang, Yan Li, Guo-Jun Qi, and Jinhui Tang. Few-shot image recognition with knowledge transfer. In ICCV, pages 441-449, 2019. 1, 5 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763. PMLR, 2021. 3, 5 +[40] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2 +[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5 +[42] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 2 +[43] Victor Garcia Satorras and Joan Bruna Estrach. Few-shot learning with graph neural networks. In ICLR, 2018. 6 +[44] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, pages 8247-8255, 2019. 1 +[45] Hongyang Gao Siyuan Sun. Meta-adam: A meta-learned adaptive optimizer with momentum for few-shot learning. In NeurIPS, 2023. 5, 6 +[46] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6 +[47] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2 +[48] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199–1208, 2018. 2 +[49] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, pages 266-282. Springer, 2020. 1, 2, 5, 6 +[50] Pavel Tokmakov, Yu-Xiong Wang, and Martial Hebert. Learning compositional representations for few-shot recognition. In ICCV, pages 6372-6381, 2019. 1 +[51] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 5, 6 + +[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. J Mach Learn Res, 9(11), 2008. 7 +[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In NeurIPS, 2016. 1, 2, 5 +[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[55] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In IJCAI, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. 6 +[56] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, pages 1831-1840, 2019. 1 +[57] Chen Xing, Nagar Rostamzadeh, Boris Oreshkin, and Pedro O O Pinheiro. Adaptive cross-modal few-shot learning. In NeurIPS, 2019. 1, 2, 4, 5 +[58] Jingyi Xu and Hieu Le. Generating representative samples for few-shot classification. In CVPR, pages 9003-9013, 2022. 2, 3, 5 +[59] Kun Yan, Zied Bouraoui, Ping Wang, Shoaib Jameel, and Steven Schockaert. Aligning visual prototypes with bert embeddings for few-shot learning. In ICMR, pages 367-375, 2021. 1, 5 +[60] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, pages 8808-8817, 2020. 1, 5 +[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In CVPR, pages 3754-3762, 2021. 1, 2 +[62] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Differentiable earth mover's distance for few-shot learning. IEEE TPAMI, 2022. 2 +[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2017. 5 +[64] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In CVPR, pages 20061-20070, 2023. 6 +[65] Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast context adaptation via meta-learning. In Int. Conf. Mach. Learn., pages 7693-7702. PMLR, 2019. 2 \ No newline at end of file diff --git a/2024/Simple Semantic-Aided Few-Shot Learning/images.zip b/2024/Simple Semantic-Aided Few-Shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e0e47a12aed0d099d3e62a3d871a344f696802e3 --- /dev/null +++ b/2024/Simple Semantic-Aided Few-Shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17681a43728b687a23427c8fe9e77e037381bcbec109d5edb981c529b91403f3 +size 799969 diff --git a/2024/Simple Semantic-Aided Few-Shot Learning/layout.json b/2024/Simple Semantic-Aided Few-Shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c24742d054aa9abb337a7599c97b06128382f52e --- /dev/null +++ b/2024/Simple Semantic-Aided Few-Shot Learning/layout.json @@ -0,0 +1,9843 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 103, + 430, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 103, + 430, + 121 + ], + "spans": [ + { + "bbox": [ + 162, + 103, + 430, + 121 + ], + "type": "text", + "content": "Simple Semantic-Aided Few-Shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": "Hai Zhang" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": " Junzhe Xu" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "inline_equation", + "content": "^{1,2*‡}" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": " Shanlin Jiang" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": " Zhenan He" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": "\\\n" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": "College of Computer Science, Sichuan University\\\n" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": "Banma Network Technology, Alibaba Group\\\n" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 141, + 463, + 225 + ], + "type": "text", + "content": "Naveen Jindal School of Management, University of Texas at Dallas\\\nzhanghi@stu.scu.edu.cn, xujunzhe.xjz@alibaba-inc.com\\\nshanlin.jiang@utdallas.edu, zhenan@scu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 281, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 567 + ], + "type": "text", + "content": "Learning from a limited amount of data, namely Few-Shot Learning, stands out as a challenging computer vision task. Several works exploit semantics and design complicated semantic fusion mechanisms to compensate for rare representative features within restricted data. However, relying on naive semantics such as class names introduces biases due to their brevity, while acquiring extensive semantics from external knowledge takes a huge time and effort. This limitation severely constrains the potential of semantics in Few-Shot Learning. In this paper, we design an automatic way called Semantic Evolution to generate high-quality semantics. The incorporation of high-quality semantics alleviates the need for complex network structures and learning algorithms used in previous works. Hence, we employ a simple two-layer network termed Semantic Alignment Network to transform semantics and visual features into robust class prototypes with rich discriminative features for few-shot classification. The experimental results show our framework outperforms all previous methods on six benchmarks, demonstrating a simple network with high-quality semantics can beat intricate multi-modal modules on few-shot classification tasks. Code is available at https://github.com/zhangdoudou123/SemFew." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 612, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 287, + 684 + ], + "type": "text", + "content": "Deep learning models have witnessed substantial advancements by leveraging extensive annotated data [8]. Nevertheless, in numerous real-world scenarios, the shortage of labeled data restricts the applicability of conventional deep learning approaches. In contrast, humans possess a remarkable cognitive ability to learn new concepts and recognize" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 256, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 545, + 292 + ], + "type": "text", + "content": "categories from just a few examples [3]. Motivated by this, Few-Shot Learning (FSL) [46] is proposed to mimic the ability of humans to learn from a few labeled samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "text", + "content": "In a typical setting of FSL [46], the support set, which consists of " + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "text", + "content": " novel classes and each class contains " + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "text", + "content": " samples, is provided for the model. The model learned on the support set is required to accurately classify test samples, which are labeled as the query set, into one of the " + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 293, + 546, + 448 + ], + "type": "text", + "content": " categories. A conventional classification approach involves projecting both support and query sets onto a pre-established metric space. Then, query images are classified by finding the nearest support image. However, few support samples may not contain sufficient discriminative features for recognition. These samples are typically located on the periphery of the sample cluster in the metric space, resulting in unstable distance evaluations and inaccurate classification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 616 + ], + "type": "text", + "content": "To address this issue, researchers start from visual-based methods [7, 11, 13, 16, 19, 29, 32, 37, 46, 49, 53, 60], aiming at extracting class-related features from periphery image, reducing the intra-class variation among samples, and thus constructing robust features for classification. Although visual-based methods have undergone extensive investigation and achieved great success, this type of approach still struggles to handle situations when there is only one hard periphery sample with minimal semantic features per class, i.e., the one-shot learning task [25, 53]. Hence, some researchers turned to applying semantics as auxiliary information to help the model better understand periphery images, leveraging the synergies between language and vision modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "From this perspective, semantic-based methods [5, 28, 38, 44, 50, 56-59, 61] dedicated to exploring different types of semantics to enhance periphery samples. One kind of method [57-59] applies the class name as the semantic source. However, class name is not the best way to understand a novel class. For instance, If a person never saw a zebra, it would be easier to identify a zebra by the definition, i.e., a horse with stripes, than by the name zebra." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 247, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 247, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 247, + 713 + ], + "type": "text", + "content": "*Equal Contribution. † Corresponding Author. \n‡ Work done during master's degree at Sichuan University." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28588" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 51, + 76, + 284, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 76, + 284, + 107 + ], + "spans": [ + { + "bbox": [ + 51, + 76, + 284, + 107 + ], + "type": "text", + "content": "The American robin is renowned for its striking appearance, with a rust-red to orange breast and abdomen, a dark grayish-blue upper body, and a white eye ring. This bird plays a pivotal ecological role as a seed disperser, primarily for fruits and berries,..." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 122, + 287, + 206 + ], + "blocks": [ + { + "bbox": [ + 48, + 122, + 287, + 206 + ], + "lines": [ + { + "bbox": [ + 48, + 122, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 122, + 287, + 206 + ], + "type": "image", + "image_path": "71f35d14f53a0d0f7579004a8d0e8a8c1906588581f2cf88234f41ca965cb527.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 288, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 288, + 274 + ], + "type": "text", + "content": "Figure 1. The introduction of how high-quality semantics reconstruct the class prototype through complementarity between modalities. Periphery image is an image with fewer discriminative features and prototype image represents images with concrete and enriched representative features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "type": "text", + "content": "Meanwhile, the class name contains ambiguity in some scenarios. For example, the class `ear` is easily misidentified as the human organ, but images in MiniImageNet [53] belong to this class are actually photos of corn. Another kind of semantic exploits discrete attribute labels as the substitution of mere class names and achieves good performance [61]. Nevertheless, the collection process is time-consuming and expert assistance requiring. Moreover, discrete attribute labels lack robust representation and struggle to benefit from context. Therefore, effectively collecting and leveraging high-quality semantics becomes an urgent research need." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 714 + ], + "type": "text", + "content": "In this paper, we propose a simple framework called Semantic-Aided Few-Shot Learning (SemFew). To begin with, we identify the vital role of high-quality semantics in FSL and innovatively propose an automatic step-by-step Semantic Evolution (SemEvo) process to acquire detailed and accurate semantics. It first converts the class name into a short description that matches the image content of each class. Next, to further enrich recognizable characteristics, we expand and paraphrase short descriptions. The paraphrased descriptions contain more class-related knowledge than class names and short descriptions, making them better encompass a vast array of visual features. To illustrate, Fig. 1 showcases an instance that the paraphrased semantic contains a large number of detailed descriptions matching visual features e.g. organ characteristics and eating habitat. With this information, the model can better reconstruct an image with periphery feature (left image), into an image with prototype feature (right image), which benefits few-shot classification. Then, to fully exploit the advantage of paraphrased semantics, we design a network called Semantic Alignment Network (SemAlign). Different from previous methods, SemAlign does not adopt complex semantic understanding modules but utilizes a basic two-layer network. After accepting semantics and visual features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "as input, it reconstructs prototype features for subsequent classification. Experimental results on six widely adopted benchmarks demonstrate the effectiveness of our method, providing a simple insight into future FSL research. Our contributions can be summarized as three points:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 545, + 252 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 304, + 133, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 167 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to consider the automatic way of collecting high-quality semantics and applying them in FSL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 168, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 204 + ], + "type": "text", + "content": "- We design a simple and efficient way to translate high-quality semantics and visual features into prototypes, without any intricate semantic understanding modules." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 204, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 252 + ], + "type": "text", + "content": "- Our approach achieves state-of-the-art performance across six benchmarks in FSL research, underscoring that a basic network can obtain excellent performance when supported by high-quality semantics." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 282, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 342 + ], + "type": "text", + "content": "The significant issue in FSL is how to effectively and accurately extract class-related features from limited data. Targeting this issue, existing research explored two different types of approaches, i.e., visual-based methods and semantic-based methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "content": "Visual-Based Methods focus on extracting class-related features from images for classification. There are two main types of this approach. The approaches with the optimization principle behind [1, 13, 40] concentrated on acquiring a set of initial model parameters capable of rapid adaptation to novel classes. However, updating the entire model with a limited amount of labeled data may result in metaoverfitting [12, 22, 35, 42, 65]. The approaches considered metric design aiming at training a metric space, where inter-class distances are maximized, while intra-class distances are minimized. Within this type of approach, one attracting method focuses on integrating self-supervised loss to pretrain the backbone network [10, 15, 19, 32, 49]. Other methods are dedicated to improving the metric, and previous research spanning from initially manually designed metrics [2, 25, 46, 62] to subsequently model-based metrics [48, 53], all of which have yielded encouraging outcomes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Semantic-Based Methods sought to enhance visual recognition performance by fusing complementary information from both visual and textual modalities [5, 28, 57, 58, 61], and related research introducing intricate network frameworks to effectively utilize textual information. Chen et al. [57] introduced an adaptive fusion mechanism to integrate visual prototypes with semantic prototypes obtained through word embeddings of class labels. Zhang et al. [61] designed a framework for prototype completion, leveraging rare attribute-level information from labels to construct representative class prototypes. Xu and Le [58] proposed the use of the Conditional Variational Autoencoder (CVAE) [47] model to generate visual features based on semantic embeddings for further classification. Li et al. [28] pre" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28589" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "sented a task-relevant adaptive margin loss based on the semantic similarity between class names. Chen et al. [5] devised complementary mechanisms for inserting semantic prompts into the feature extractor, allowing the feature extractor to better focus on class-specific features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": "Previous research emphasized the exploitation of class names and designed a complicated network to integrate semantics with visual information. In contrast, our primary concern lies in the inherent quality of semantics, and how it can relieve the burden of designing intricate network structures and learning algorithms. Remarkably, we exhibit that high-quality semantics information has the ability to yield superior outcomes even within a simple network." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 236, + 129, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 129, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 129, + 250 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 305 + ], + "type": "text", + "content": "This section first revisits the background of FSL, then introduces two components of SemFew: explaining how semantics are gradually improved through SemEvo, and demonstrating the details of our proposed SemAlign." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 310, + 128, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 128, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 128, + 322 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": "The dataset of FSL includes two components: a base set " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base} = \\{(x,y)|x\\in \\mathcal{X}_{base},y\\in \\mathcal{C}_{base}\\}" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " for training a metric space and a novel set " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{novel} = \\{(x,y)|x\\in \\mathcal{X}_{novel},y\\in \\mathcal{C}_{novel}\\}" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " for testing, where " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " denotes the image and " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " represents the label. It is important to note that the label space for both sets is disjoint, i.e., " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{base}\\cap \\mathcal{C}_{novel} = \\emptyset" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": ". During the testing process, the support set " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "S = \\{(x_i,y_i)\\}_{i = 0}^{N\\times K}" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " is randomly sampled from " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{novel}" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": ", which contains " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " classes with each class consisting of " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " samples. Then, the model is required to correctly classify images in the query set " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{(x_i,y_i)\\}_{i = 0}^{N\\times M}" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " into one of " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " classes in the support set " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": " is the number of query samples in each class. Generally, this type of classification task is named as " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 328, + 287, + 495 + ], + "type": "text", + "content": "-shot task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "text", + "content": "A simple but effective FSL method is to calculate the prototype of each support class, and then classify query samples by finding the nearest prototype in the metric space. To be specific, the prototype of class " + }, + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "text", + "content": ", labeled as " + }, + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "text", + "content": ", is the mean vector of all samples in the support set belonging to class " + }, + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 496, + 287, + 566 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 564, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 564, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 119, + 564, + 287, + 594 + ], + "type": "interline_equation", + "content": "p _ {t} = \\frac {1}{\\left| \\mathcal {S} _ {t} \\right|} \\sum_ {x _ {i} \\in \\mathcal {S} _ {t}} f \\left(x _ {i}\\right), \\tag {1}", + "image_path": "e34e61710a2ee74b84ee7f2ee4c44f30c4a02e36b0f95c764934a11f617c4d23.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " is a set of all samples belonging to class " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " in the support set " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " denotes the backbone network. Then, the probability of a query sample " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " belonging to class " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " is related to the distance between " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " and prototype of class " + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": " [46], which is calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 658, + 287, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 658, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 100, + 658, + 287, + 687 + ], + "type": "interline_equation", + "content": "P (y = t) = \\frac {\\exp d \\left(q , p _ {t}\\right)}{\\sum_ {i = 1} ^ {N} \\exp d \\left(q , p _ {i}\\right)}, \\tag {2}", + "image_path": "b33913ce3edbdcc67a54b02ff0e1ddcd93f2537cb3e0f00d517e48a268bd384a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": " means distance metric, e.g., cosine distance, and " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": " denotes the number of classes." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 547, + 234 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 547, + 234 + ], + "type": "image", + "image_path": "5b24ad8af9db3e7a6085fff5b1863d6db76c335e90d50d2aa121387db69c6079.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 243, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 266 + ], + "type": "text", + "content": "Figure 2. The illustration of how Semantic Evolution converts the simple name and the definition into the high-quality description." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 286, + 420, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 420, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 420, + 298 + ], + "type": "text", + "content": "3.2. Semantic Evolution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "text", + "content": "This section explains how we automatically generate high-quality semantics for few-shot classification by the designed process termed Semantic Evolution. This process first converts class names to short descriptions and then paraphrases them into high-quality semantics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 365, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 472 + ], + "type": "text", + "content": "As we illustrated before, adopting class names as semantics [5, 39, 58] unavoidably introduces ambiguity. Therefore, we first retrieve the definition of each class name from WordNet [33] that matches image content, where definitions are more detailed and concrete than mere class names. However, even though definitions are able to precisely describe a class to some extent, brief descriptions may still overlook key visual information, due to the fact that visual features are highly variable." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 472, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 580 + ], + "type": "text", + "content": "To comprehensively conclude variable visual features, we leverage the extensive knowledge stored in the pretrained Large Language Models (LLMs) [4] to augment definitions. The prompt we used is: " + }, + { + "bbox": [ + 304, + 472, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\{\\text{WordNet Definition}\\}" + }, + { + "bbox": [ + 304, + 472, + 545, + 580 + ], + "type": "text", + "content": " is the definition of the " + }, + { + "bbox": [ + 304, + 472, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\{\\text{Class Name}\\}" + }, + { + "bbox": [ + 304, + 472, + 545, + 580 + ], + "type": "text", + "content": ". Please rewrite and expand this definition to make it more detailed and consistent with scientific fact. Briefness is required, using only one paragraph. The illustration of the Semantic Evolution process is shown in Fig. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": "The paraphrased class description generated by LLMs precisely describes the corresponding class itself, which does not apply to other categories. This characteristic ensures that semantic descriptions do not leak information to new categories, which follows the paradigm of FSL." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 647, + 468, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 468, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 468, + 660 + ], + "type": "text", + "content": "3.3. Semantic Alignment Network" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "After obtaining paraphrased semantics in the previous section, we devise the Semantic Alignment Network (SemAlign), aiming at translating hard periphery samples into robust class prototypes." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28590" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 534, + 305 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 534, + 305 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 534, + 305 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 534, + 305 + ], + "type": "image", + "image_path": "ac749bd011964ce9970d2992531a46ccc7543a6aec1a5d9cdfc48ed5d932d7a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "text", + "content": "Figure 3. The framework of our proposed SemFew. During the training stage, images and paraphrased semantics are encoded and fed into SemAlign " + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "text", + "content": ", with the objective of reducing the distance between the output of " + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "text", + "content": " and the class prototype in the visual space. During the testing stage, images in the support set are transformed into class prototypes by " + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "text", + "content": ", and query images are classified by identifying the nearest prototype. The symbol " + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 314, + 546, + 360 + ], + "type": "text", + "content": " denotes a concatenation operation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "content": "For a cluster of samples, the cluster center best encapsulates the characteristic of a category and has been regarded as a prototype in previous work [46]. Suppose " + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "content": " represents semantic features, and " + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "content": " stands for the center of each class, the simplest implementation of SemAlign is to learn the alignment between semantics " + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "content": " and centers " + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 458, + 287, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 458, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 110, + 458, + 287, + 478 + ], + "type": "interline_equation", + "content": "\\min _ {W _ {1}, W _ {2}} L \\left(\\sigma \\left(S ^ {\\top} W _ {1}\\right) W _ {2}, C\\right), \\tag {3}", + "image_path": "2fdf28e0a8b8144fdf919ac1a6703c0baed1547c7d0b801f05ed3173be61468a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "text", + "content": " are parameters of network, " + }, + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 486, + 287, + 508 + ], + "type": "text", + "content": " is an activation function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": "Considering that each modality views things from a different and complementary perspective, this phenomenon can be reasonably utilized to obtain better prototypes. Therefore, we fuse visual features with high-quality semantics, and then align multi-modal output with the cluster center, the details of the whole pipeline can be found in Fig. 3. Specifically, suppose the " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": "-th image in the dataset is represented as " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": ", its label and semantic are denoted as " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": ", respectively. First, " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " is encoded by a vision encoder " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": ", which is pre-trained on the base set. Also, semantic description " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " is encoded by an off-the-shelf text encoder " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": ". Then, the encoded visual and text features are concatenated together and fed into SemAlign " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " to reconstruct the prototype. The training loss supervises the distance between the reconstructed prototype towards the cluster center:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 696, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 696, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 96, + 696, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {W _ {1}, W _ {2}} \\mathbb {E} [ \\| h (f (x _ {i}), g (s _ {i})) - c _ {y} \\| _ {1} ], \\tag {4}", + "image_path": "403c065555c7d01447bcf9dbf97ea0c7b54a46891a6eda1154ad98d90b9ec112.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "inline_equation", + "content": "h(f(x_i), g(s_i)) = \\sigma([f(x_i) \\cdot g(s_i)]^\\top W_1)W_2" + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "text", + "content": " is SemAlign with learnable weights " + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "inline_equation", + "content": "W_1" + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "inline_equation", + "content": "W_2" + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 305, + 379, + 545, + 416 + ], + "type": "text", + "content": " means concatenation operation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": "During the testing process, a " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": "-shot support set is randomly sampled in the novel set. For class " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": ", we obtain its paraphrased description " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": " by Semantic Evolution proposed in Sec. 3.2. Then, support image " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": " and description " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": " are encoded by " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": ", respectively, and fed into " + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 305, + 418, + 546, + 491 + ], + "type": "text", + "content": " to get the reconstructed prototype:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 366, + 504, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 504, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 366, + 504, + 545, + 537 + ], + "type": "interline_equation", + "content": "r _ {t} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} h \\left(f \\left(x _ {i}\\right), g \\left(s _ {t}\\right)\\right), \\tag {5}", + "image_path": "37b270580797160baf73ff02e0a311ceaf4e3c75c7b4687d3f29e427963bc162.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 551, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 551, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 551, + 545, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 551, + 545, + 598 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 305, + 551, + 545, + 598 + ], + "type": "text", + "content": " is the reconstructed prototype. To ensure the reconstructed prototype is further aligned with the groundtruth class prototype, it will be fused with support images in a convex combination manner [57]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 614, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 614, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 378, + 614, + 545, + 628 + ], + "type": "interline_equation", + "content": "p _ {t} = k r _ {t} + (1 - k) u _ {t}, \\tag {6}", + "image_path": "b9d4fc67788ece18cbab5346d6116dd305585ed29c406048ad09bae16dd681d6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "content": " means the classification prototype of class " + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "u_t = \\frac{1}{K} \\sum_{i=1}^{K} f(x_i)" + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "content": " is the mean vector of support samples, " + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "content": " is the reconstructed prototype produced by SemAlign, and " + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "k \\in [0,1]" + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "content": " is a manually controlled fusion factor. Then, Eq. (2) is utilized to assign label for each query sample, where " + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 305, + 642, + 546, + 713 + ], + "type": "text", + "content": " is set as cosine distance." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28591" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 544, + 369 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 544, + 369 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 544, + 369 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 544, + 369 + ], + "type": "table", + "html": "
MethodVenueBackboneMiniImageNetTieredImageNet
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
Visual-BasedMatchNet [53]NeurIPS'16ResNet-1265.64 ± 0.2078.72 ± 0.1568.50 ± 0.9280.60 ± 0.71
ProtoNet [46]NeurIPS'17ResNet-1262.39 ± 0.2180.53 ± 0.1468.23 ± 0.2384.03 ± 0.16
MAML [13]ICML'17ResNet-1249.24 ± 0.2158.05 ± 0.1067.92 ± 0.1772.41 ± 0.20
TADAM [37]NeurIPS'18ResNet-1258.50 ± 0.3076.70 ± 0.3062.13 ± 0.3181.92 ± 0.30
CAN [20]NeurIPS'19ResNet-1263.85 ± 0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
CTM [29]CVPR'19ResNet-1864.12 ± 0.8280.51 ± 0.1368.41 ± 0.3984.28 ± 1.73
RFS [49]ECCV'20ResNet-1262.02 ± 0.6379.64 ± 0.4469.74 ± 0.7284.41 ± 0.55
FEAT [60]CVPR'20ResNet-1266.78 ± 0.2082.05 ± 0.1470.80 ± 0.2384.79 ± 0.16
Meta-Baseline [6]ICCV'21ResNet-1263.17 ± 0.2379.26 ± 0.1768.62 ± 0.2783.29 ± 0.18
SUN [11]ECCV'22ViT-S67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [19]NeurIPS'22Swin-T72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
FGFL [7]ICCV'23ResNet-1269.14 ± 0.8086.01 ± 0.6273.21 ± 0.8887.21 ± 0.61
Meta-AdaM [45]NeurIPS'23ResNet-1259.89 ± 0.4977.92 ± 0.4365.31 ± 0.4885.24 ± 0.35
Semantic-BasedKTN [38]ICCV'19Conv-12864.42 ± 0.7274.16 ± 0.5674.16 ± 0.56-
AM3 [57]NeurIPS'19ResNet-1265.30 ± 0.4978.10 ± 0.3669.08 ± 0.4782.58 ± 0.31
TRAML[28]CVPR'20ResNet-1267.10 ± 0.5279.54 ± 0.60--
AM3-BERT [59]ICMR'21ResNet-1268.42 ± 0.5181.29 ± 0.5977.03 ± 0.8587.20 ± 0.70
SVAE-Proto [58]CVPR'22ResNet-1274.84 ± 0.2383.28 ± 0.4076.98 ± 0.6585.77 ± 0.50
SP-CLIP [5]CVPR'23Visformer-T72.31 ± 0.4083.42 ± 0.3078.03 ± 0.4688.55 ± 0.32
SemFewOursResNet-1277.63 ± 0.6383.04 ± 0.4878.96 ± 0.8085.88 ± 0.62
SemFew-TransOursSwin-T78.94 ± 0.6686.49 ± 0.5082.37 ± 0.7789.89 ± 0.52
", + "image_path": "4a5521e2ab8d1d53d2e45d1e4aede357cec9ecf7d23d7431a3faad569e52fa72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 376, + 542, + 388 + ], + "lines": [ + { + "bbox": [ + 49, + 376, + 542, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 542, + 388 + ], + "type": "text", + "content": "Table 1. Results (%) on MiniImageNet and TieredImageNet. The ± shows 95% confidence intervals. The best results are shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 399, + 127, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 127, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 127, + 413 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 420, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 456 + ], + "type": "text", + "content": "We exhibit experimental results in this section, including benchmarks introduction, implementation details, and few-shot classification performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 467, + 110, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 110, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 110, + 479 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": "The study evaluates the proposed method on six established FSL datasets: MiniImageNet [53], TieredImageNet [41], CIFAR-FS [27], FC100 [37], Places [63], and CUB [54]. MiniImageNet and TieredImageNet are both subsets of ImageNet [8]. MiniImageNet consists of 64 training classes, 16 validating classes, and 20 testing classes. TieredImageNet encompasses 351 training classes, 97 validating classes, and 160 testing classes. CIFAR-FS and FC100 are derived from the CIFAR-100 [26]. CIFAR-FS employs a random partitioning strategy with 64 training classes, 16 validating classes, and 20 testing classes. FC100 introduces a unique superclass partitioning approach, where the training set comprises 12 subclasses, namely 60 classes, while both the validation and test sets include four subclasses, totaling 20 classes. Places and Caltech-UCSD Birds-200-2011 (CUB) are datasets for cross-domain scenario testing. In a cross-domain evaluation [51], the model is initially trained on MiniImageNet's base split and tested on the corresponding test split." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 399, + 440, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 440, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 440, + 412 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 421, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 540 + ], + "type": "text", + "content": "Architecture. We employ a pre-trained ResNet-12 [18] and Swin-T [30] as the vision encoders. Additionally, we apply two types of text encoder: the text encoder from ViT-B/16 CLIP [39] where output dimension is 512, and the BERT-Base text encoder [9] which has 768 output dimension. The network " + }, + { + "bbox": [ + 304, + 421, + 545, + 540 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 421, + 545, + 540 + ], + "type": "text", + "content": " that combines visual and textual features consists of two fully connected layers and a LeakyReLU [31] activation function. The hidden layer has a dimension of 4,096. In the context of Semantic Evolution, the utilized Large Language Model is GPT-3.5-turbo [36]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "content": "Training Details. A two-step training procedure is performed in experiments. In step one, two kinds of vision encoder are pre-trained, the ResNet-12 [18] follows the same pretraining settings as the FEAT [60], while the Swin-T [30] adopts the identical pretraining strategy as FewTURE [19]. Next, the network " + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "content": " is trained 50 epochs. This training is conducted with a batch size of 128 and employs an Adam optimizer [24] with a learning rate of 1e-4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Evaluation protocol. The proposed method is evaluated under 5-way 1-shot and 5-way 5-shot settings. For each setting, 600 classification tasks are uniformly sampled from the novel set. In each task, there are 15 query samples per class. The mean and " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " confidence interval of the accuracy are reported." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28592" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 533, + 250 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 533, + 250 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 533, + 250 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 533, + 250 + ], + "type": "table", + "html": "
MethodVenueBackboneCIFAR-FSFC100
5-way 1-shot5-way 5-shot5-way 1-shot5-way 5-shot
ProtoNet [46]NeurIPS'17ResNet-1272.20 ± 0.7083.50 ± 0.5041.54 ± 0.7657.08 ± 0.76
TADAM [37]NeurIPS'18ResNet-12--40.10 ± 0.4056.10 ± 0.40
MetaOptNet [27]CVPR'19ResNet-1272.80 ± 0.7084.30 ± 0.5047.20 ± 0.6055.50 ± 0.60
MABAS [23]ECCV'20ResNet-1273.51 ± 0.9285.65 ± 0.6542.31 ± 0.7558.16 ± 0.78
RFS [49]ECCV'20ResNet-1271.50 ± 0.8086.00 ± 0.5042.60 ± 0.7059.10 ± 0.60
SUN [11]ECCV'22ViT-S78.37 ± 0.4688.84 ± 0.32--
FewTURE [19]NeurIPS'22Swin-T77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
Meta-AdaM [45]NeurIPS'23ResNet-12--41.12 ± 0.4956.14 ± 0.49
SP-CLIP [5]CVPR'23Visformer-T82.18 ± 0.4088.24 ± 0.3248.53 ± 0.3861.55 ± 0.41
SemFewOursResNet-1283.65 ± 0.7087.66 ± 0.6054.36 ± 0.7162.79 ± 0.74
SemFew-TransOursSwin-T84.34 ± 0.6789.11 ± 0.5454.27 ± 0.7765.02 ± 0.72
", + "image_path": "074e3ca630f83d9ce4f8a842a218a98aa9c012aad9bddb81a73dab69239aad02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 279, + 287, + 407 + ], + "blocks": [ + { + "bbox": [ + 73, + 257, + 518, + 268 + ], + "lines": [ + { + "bbox": [ + 73, + 257, + 518, + 268 + ], + "spans": [ + { + "bbox": [ + 73, + 257, + 518, + 268 + ], + "type": "text", + "content": "Table 2. Results (%) on CIFAR-FS and FC100. The ± shows 95% confidence intervals. The best results are shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 279, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 279, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 279, + 287, + 407 + ], + "type": "table", + "html": "
MethodVenueCUBPlaces
1-shot5-shot1-shot5-shot
GNN [43]ICLR'1845.6962.2553.1070.84
S2M2 [32]WACV'2048.2470.44--
FT [51]ICLR'2047.4766.9855.7773.94
ATA [55]IJCAI'2145.0066.2253.5775.48
AFA [21]ECCV'2246.8668.2554.0476.21
StyleAdv [14]CVPR'2348.4968.7258.5877.73
LDP-net [64]CVPR'2349.8270.3953.8272.90
SemFew-NameOurs57.5872.2663.2274.54
SemFewOurs59.0772.4764.0174.70
", + "image_path": "2f469ac7825f599475b5c87c5788187dc29c329e98d90b951cbcb5955106f080.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 437 + ], + "type": "text", + "content": "Table 3. Average results (\\%) on cross-domain scenarios. SemFew-Name denotes that semantics are class names." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 445, + 171, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 171, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 171, + 458 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 464, + 180, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 180, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 180, + 475 + ], + "type": "text", + "content": "4.3.1 Few-Shot Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "text", + "content": "The comparative results of our proposed approach against previous methods across four benchmarks are shown in Tabs. 1 and 2. Specifically, in the 1-shot task, SemFew exhibits notable superiority, surpassing SP-CLIP [5] " + }, + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "inline_equation", + "content": "0.93\\% - 5.83\\%" + }, + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "text", + "content": " on four benchmarks. Recent studies [5, 11, 19] have explored different Transformer architectures as backbones. Consequently, we conduct experiments with a Transformer backbone, resulting in SemFew-Trans. This variant further elevates performance, outperforming SP-CLIP by " + }, + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "inline_equation", + "content": "2.16\\% - 6.63\\%" + }, + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "text", + "content": " in the 1-shot setting. In summary, SemFew-Trans achieves state-of-the-art results compared to both previous visual-based and semantic-based approaches across all 1-shot and 5-shot tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 657, + 201, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 201, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 201, + 669 + ], + "type": "text", + "content": "4.3.2 Cross-Domain Classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": "To evaluate whether semantics serve as an effective bridge across novel tasks, we test SemFew on completely novel datasets. Such a task is also known as Cross-Domain Few-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "text", + "content": "Shot Learning (CDFSL) [51], which involves extending the capacity to recognize novel categories to domains distinct from the training domain. All the results are reported in Tab. 3, where the experimental setup is the same as FT [51]. Tab. 3 indicates that SemFew significantly surpassing the state-of-the-art methods, where it outperforms best compared methods by " + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "inline_equation", + "content": "9.25\\%" + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "text", + "content": " (1-shot) and " + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "inline_equation", + "content": "1.87\\%" + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "text", + "content": " (5-shot) on CUB. It also surpasses the lastest method [64] " + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "inline_equation", + "content": "10.19\\%" + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "text", + "content": " (1-shot) and " + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "inline_equation", + "content": "1.80\\%" + }, + { + "bbox": [ + 304, + 281, + 547, + 437 + ], + "type": "text", + "content": " (5-shot) on Places. Conversely, SemFew-Name, which adopts a template (i.e., The Photo of a bird called {Class Name}) as the semantic, performs worse than SemFew. This observation highlights the inadequacy of using solely names as semantics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 445, + 406, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 406, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 406, + 456 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 463, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 499 + ], + "type": "text", + "content": "We conduct ablation experiments on three datasets, i.e., MiniImageNet, CIFAR-FS, and FC100, and all experiments are performed using the ResNet-12 network architecture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 515, + 416, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 416, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 416, + 526 + ], + "type": "text", + "content": "4.4.1 Alignment Source" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "Given that our framework is primarily responsible for translating visual samples " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " and semantics descriptions " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " into ground-truth class prototype (i.e., " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "V + S \\Rightarrow C" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "), we conduct ablation studies by removing " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": ". The results can be found in Tab. 4. First, we remove semantics and force pure visual samples to reconstruct prototypes (i.e., " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "V \\Rightarrow C" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "). The result shows that it gets the worst performance, indicating that reconstructing a prototype from pure visual features is challenging. Then, we abandon visual features to implement our proposed principle in Sec. 3.3 (i.e., " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "S \\Rightarrow C" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "). As we can see, it achieves substantial performance gains, reflecting the rich class representative features within semantics. However, our method combining visual and semantic acquires the best performance, which indicates the complementarity between visual and semantic modalities does en" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28593" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 71, + 529, + 148 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 529, + 148 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 529, + 148 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 529, + 148 + ], + "type": "table", + "html": "
Method5-way 1-shot5-way 5-shot
MiniImageNetCIFAR-FSFC100MiniImageNetCIFAR-FSFC100
V ⇒ C66.27 ± 0.8174.43 ± 0.9044.72 ± 0.7581.70 ± 0.5586.74 ± 0.6160.58 ± 0.76
S ⇒ C76.77 ± 0.6682.27 ± 0.7152.12 ± 0.7382.92 ± 0.4887.42 ± 0.6062.22 ± 0.75
V + S ⇒ C77.63 ± 0.6383.65 ± 0.7054.36 ± 0.7183.04 ± 0.4887.66 ± 0.6062.79 ± 0.74
", + "image_path": "44b74f08e7a152ced310721429ff9ea401b4373bafe661549ede1b26148b88a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 186, + 287, + 228 + ], + "blocks": [ + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "type": "text", + "content": "Table 4. Results (%) on different alignment sources. " + }, + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "type": "text", + "content": " means visual features, " + }, + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "type": "text", + "content": " represents semantic features, " + }, + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 156, + 545, + 168 + ], + "type": "text", + "content": " stands for class prototypes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "table", + "html": "
CenterMiniImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Cluster77.1682.9478.3685.1983.3987.6153.3062.22
Mean77.6383.0478.9685.8883.6587.6654.3662.79
", + "image_path": "51d58ea90200c55f05fb7f080edb4863c2a5fe783de914d08bee44412f813f10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 281, + 287, + 331 + ], + "blocks": [ + { + "bbox": [ + 47, + 237, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 237, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 287, + 270 + ], + "type": "text", + "content": "Table 5. Average results (\\%) on different prototypes. Mean represents the mean vector of sample clusters. Cluster means the prototype is determined by Clustering Algorithm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 281, + 287, + 331 + ], + "lines": [ + { + "bbox": [ + 49, + 281, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 281, + 287, + 331 + ], + "type": "table", + "html": "
classifierMinImageNetTieredImageNetCIFAR-FSFC100
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
LR77.5582.7078.6684.7283.0886.7051.5357.49
EU77.5683.1578.6585.2682.8386.8754.1462.54
CO77.6383.0478.9685.8883.6587.6654.3662.79
", + "image_path": "85a6fe121fe6717aae5acfe607292d73266575ba2b2f443325838182059e36be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 338, + 287, + 372 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 372 + ], + "type": "text", + "content": "Table 6. Average results " + }, + { + "bbox": [ + 47, + 338, + 287, + 372 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 338, + 287, + 372 + ], + "type": "text", + "content": " on different classifiers. LR: linear logistic regression classifier. EU: euclidean-distance nearest prototype classifier. CO: cosine-distance nearest prototype classifier." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 391, + 257, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 257, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 257, + 403 + ], + "type": "text", + "content": "hance the robustness of the reconstructed prototype." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 419, + 262, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 262, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 262, + 432 + ], + "type": "text", + "content": "4.4.2 Prototype and Classifier Selection Strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 438, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 581 + ], + "type": "text", + "content": "Tab. 5 displays the performance associated with various learning targets. Results reveal that employing the mean vector as a prototype yields slightly better results compared to prototypes identified by the clustering algorithm [17]. This suggests that the mean vector consolidates representative visual information, making it a highly effective class prototype. As for the classification head design, we present the results in Tab. 6, which indicates the use of cosine distance as the metric to achieve nearest neighbor classification gets the best results, euclidean distance gets similar but lower outcomes, and linear logistic classifier gets the worst performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 598, + 163, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 163, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 163, + 609 + ], + "type": "text", + "content": "4.4.3 Semantic Evolution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "We use different kinds of semantics and text encoders to evaluate the influence of semantics, where results can be found in Fig. 4. In most scenarios, our model attains optimal results when aided by paraphrased semantics, and the model trained on simple definitions secures the second-best performance. This trend is reasonable since paraphrased descriptions contain more knowledge than simple definitions and class names. Between the comparison of different text" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 186, + 425, + 275 + ], + "blocks": [ + { + "bbox": [ + 307, + 186, + 425, + 275 + ], + "lines": [ + { + "bbox": [ + 307, + 186, + 425, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 425, + 275 + ], + "type": "image", + "image_path": "aac258d6026627afec3d05617817eb1fa140f66da681af56a215c01ea3daebc9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 276, + 408, + 285 + ], + "lines": [ + { + "bbox": [ + 326, + 276, + 408, + 285 + ], + "spans": [ + { + "bbox": [ + 326, + 276, + 408, + 285 + ], + "type": "text", + "content": "(a) MiniImageNet, 1-shot" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 287, + 425, + 374 + ], + "blocks": [ + { + "bbox": [ + 307, + 287, + 425, + 374 + ], + "lines": [ + { + "bbox": [ + 307, + 287, + 425, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 425, + 374 + ], + "type": "image", + "image_path": "b9f8ba4222802d95e8b212a91c57036ac921ec4d66476c9fd68395abdac56354.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 376, + 402, + 384 + ], + "lines": [ + { + "bbox": [ + 332, + 376, + 402, + 384 + ], + "spans": [ + { + "bbox": [ + 332, + 376, + 402, + 384 + ], + "type": "text", + "content": "(c) CIFAR-FS, 1-shot" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 385, + 425, + 474 + ], + "blocks": [ + { + "bbox": [ + 307, + 385, + 425, + 474 + ], + "lines": [ + { + "bbox": [ + 307, + 385, + 425, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 385, + 425, + 474 + ], + "type": "image", + "image_path": "0732e6ff299f45084b3b59a268bed241b2d5caa132b551144b24fe98a11cc78a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 475, + 396, + 483 + ], + "lines": [ + { + "bbox": [ + 339, + 475, + 396, + 483 + ], + "spans": [ + { + "bbox": [ + 339, + 475, + 396, + 483 + ], + "type": "text", + "content": "(e) FC100, 1-shot" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 495, + 522, + 506 + ], + "lines": [ + { + "bbox": [ + 329, + 495, + 522, + 506 + ], + "spans": [ + { + "bbox": [ + 329, + 495, + 522, + 506 + ], + "type": "text", + "content": "Figure 4. Average results " + }, + { + "bbox": [ + 329, + 495, + 522, + 506 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 329, + 495, + 522, + 506 + ], + "type": "text", + "content": " on different semantics." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 426, + 186, + 544, + 275 + ], + "blocks": [ + { + "bbox": [ + 426, + 186, + 544, + 275 + ], + "lines": [ + { + "bbox": [ + 426, + 186, + 544, + 275 + ], + "spans": [ + { + "bbox": [ + 426, + 186, + 544, + 275 + ], + "type": "image", + "image_path": "0bce1ec77520057a58b9132c2c2eb36e0804e5c0c28afcb7cae2d25a56ac4a42.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 276, + 527, + 285 + ], + "lines": [ + { + "bbox": [ + 444, + 276, + 527, + 285 + ], + "spans": [ + { + "bbox": [ + 444, + 276, + 527, + 285 + ], + "type": "text", + "content": "(b) MiniImageNet, 5-shot" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 427, + 286, + 544, + 374 + ], + "blocks": [ + { + "bbox": [ + 427, + 286, + 544, + 374 + ], + "lines": [ + { + "bbox": [ + 427, + 286, + 544, + 374 + ], + "spans": [ + { + "bbox": [ + 427, + 286, + 544, + 374 + ], + "type": "image", + "image_path": "46af499085228bacf53f45433249ebe175794bac562d87a095b371f055036ae4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 376, + 521, + 384 + ], + "lines": [ + { + "bbox": [ + 450, + 376, + 521, + 384 + ], + "spans": [ + { + "bbox": [ + 450, + 376, + 521, + 384 + ], + "type": "text", + "content": "(d) CIFAR-FS, 5-shot" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 427, + 385, + 544, + 474 + ], + "blocks": [ + { + "bbox": [ + 427, + 385, + 544, + 474 + ], + "lines": [ + { + "bbox": [ + 427, + 385, + 544, + 474 + ], + "spans": [ + { + "bbox": [ + 427, + 385, + 544, + 474 + ], + "type": "image", + "image_path": "b5178b19cf2cdb301a06e4cb9ea709a7b5101eabd689ba726a4053b139a8c77c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 475, + 514, + 483 + ], + "lines": [ + { + "bbox": [ + 457, + 475, + 514, + 483 + ], + "spans": [ + { + "bbox": [ + 457, + 475, + 514, + 483 + ], + "type": "text", + "content": "(f) FC100, 5-shot" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 517, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 613 + ], + "type": "text", + "content": "encoders, the model trained on CLIP encoded semantics greatly outperforms the model trained on BERT encoded semantics. This is because CLIP is trained in a multi-modal manner, rendering it better capable of understanding images. Although a multi-modal text encoder has been widely adopted in FSL research, our model obtains competitive results on BERT, demonstrating that uni-modal semantics can also achieve excellent results." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 623, + 432, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 432, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 432, + 635 + ], + "type": "text", + "content": "4.5. Visualization Analysis" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "We visualize visual space features through the t-SNE algorithm [52], where the results are presented in Fig. 5. We select three novel classes on the MiniImageNet dataset, and then randomly sampled three support samples and one hundred query samples, where the distribution of support and query samples in visual space can be found in Figs. 5a" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28594" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 167, + 175 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 167, + 175 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 167, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 167, + 175 + ], + "type": "image", + "image_path": "946115ba043e80bfa1f2164936e653aa653280766b8b4e8f594f2a12b1079258.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 185, + 126, + 195 + ], + "lines": [ + { + "bbox": [ + 87, + 185, + 126, + 195 + ], + "spans": [ + { + "bbox": [ + 87, + 185, + 126, + 195 + ], + "type": "text", + "content": "(a) Support" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 69, + 293, + 175 + ], + "blocks": [ + { + "bbox": [ + 173, + 69, + 293, + 175 + ], + "lines": [ + { + "bbox": [ + 173, + 69, + 293, + 175 + ], + "spans": [ + { + "bbox": [ + 173, + 69, + 293, + 175 + ], + "type": "image", + "image_path": "260f067000755f8f22ca3e4a24b50ad123eacd792bfe095f3066fab05d9b50c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 185, + 250, + 195 + ], + "lines": [ + { + "bbox": [ + 216, + 185, + 250, + 195 + ], + "spans": [ + { + "bbox": [ + 216, + 185, + 250, + 195 + ], + "type": "text", + "content": "(b) Query" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 69, + 421, + 175 + ], + "blocks": [ + { + "bbox": [ + 302, + 69, + 421, + 175 + ], + "lines": [ + { + "bbox": [ + 302, + 69, + 421, + 175 + ], + "spans": [ + { + "bbox": [ + 302, + 69, + 421, + 175 + ], + "type": "image", + "image_path": "4ba299f6f11c5aab3394802512653b41d24d57df33ef1370aad699f2003e974f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 185, + 391, + 195 + ], + "lines": [ + { + "bbox": [ + 329, + 185, + 391, + 195 + ], + "spans": [ + { + "bbox": [ + 329, + 185, + 391, + 195 + ], + "type": "text", + "content": "(c) Real Prototype" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 429, + 70, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 429, + 70, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 429, + 70, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 429, + 70, + 547, + 175 + ], + "type": "image", + "image_path": "28ac5315e67b252f1f1ba15ece9c403d9ce876c550f2031b52a4450b41b94755.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 185, + 534, + 195 + ], + "lines": [ + { + "bbox": [ + 441, + 185, + 534, + 195 + ], + "spans": [ + { + "bbox": [ + 441, + 185, + 534, + 195 + ], + "type": "text", + "content": "(d) Reconstructed Prototype" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 240, + 167, + 328 + ], + "blocks": [ + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "text", + "content": "Figure 5. Visualization results on the MiniImageNet dataset. Different colors or shapes represent different classes. The " + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "text", + "content": " represents the class prototypes, and the " + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "text", + "content": " denotes the prototypes reconstructed by our method." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 240, + 167, + 328 + ], + "lines": [ + { + "bbox": [ + 48, + 240, + 167, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 167, + 328 + ], + "type": "image", + "image_path": "d373e0caa9759a29b73689bae818291e4e338c01c941d5b6ab4c682f97c15d3a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 167, + 240, + 287, + 328 + ], + "blocks": [ + { + "bbox": [ + 167, + 240, + 287, + 328 + ], + "lines": [ + { + "bbox": [ + 167, + 240, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 167, + 240, + 287, + 328 + ], + "type": "image", + "image_path": "2f37cf55549b011c9f554ab973ce627e53bbffc76c1407b9cec4a90bd3b1584a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 48, + 340, + 167, + 428 + ], + "blocks": [ + { + "bbox": [ + 67, + 330, + 149, + 340 + ], + "lines": [ + { + "bbox": [ + 67, + 330, + 149, + 340 + ], + "spans": [ + { + "bbox": [ + 67, + 330, + 149, + 340 + ], + "type": "text", + "content": "(a) MiniImageNet,CLIP" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 340, + 167, + 428 + ], + "lines": [ + { + "bbox": [ + 48, + 340, + 167, + 428 + ], + "spans": [ + { + "bbox": [ + 48, + 340, + 167, + 428 + ], + "type": "image", + "image_path": "ab76e6586d5fdbe335edca0a6d35503d49a6b29cdc6a971f6dbea040ed0e5628.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 168, + 340, + 287, + 428 + ], + "blocks": [ + { + "bbox": [ + 185, + 330, + 268, + 340 + ], + "lines": [ + { + "bbox": [ + 185, + 330, + 268, + 340 + ], + "spans": [ + { + "bbox": [ + 185, + 330, + 268, + 340 + ], + "type": "text", + "content": "(b) MiniImageNet, BERT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 340, + 287, + 428 + ], + "lines": [ + { + "bbox": [ + 168, + 340, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 168, + 340, + 287, + 428 + ], + "type": "image", + "image_path": "c0cfb4a20ced74d64b5cec0bc59a9a404c98f608f3daf84004c038c2b4702128.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 48, + 440, + 167, + 526 + ], + "blocks": [ + { + "bbox": [ + 74, + 430, + 143, + 439 + ], + "lines": [ + { + "bbox": [ + 74, + 430, + 143, + 439 + ], + "spans": [ + { + "bbox": [ + 74, + 430, + 143, + 439 + ], + "type": "text", + "content": "(c) CIFAR-FS,CLIP" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 440, + 167, + 526 + ], + "lines": [ + { + "bbox": [ + 48, + 440, + 167, + 526 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 167, + 526 + ], + "type": "image", + "image_path": "fcbbd8bafd95971bf9721b4734ad71b72feaa93bad3fec9fda3a3a24a13bbd05.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 529, + 137, + 539 + ], + "lines": [ + { + "bbox": [ + 80, + 529, + 137, + 539 + ], + "spans": [ + { + "bbox": [ + 80, + 529, + 137, + 539 + ], + "type": "text", + "content": "(e) FC100, CLIP" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 548, + 272, + 560 + ], + "lines": [ + { + "bbox": [ + 61, + 548, + 272, + 560 + ], + "spans": [ + { + "bbox": [ + 61, + 548, + 272, + 560 + ], + "type": "text", + "content": "Figure 6. Average results " + }, + { + "bbox": [ + 61, + 548, + 272, + 560 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 61, + 548, + 272, + 560 + ], + "type": "text", + "content": " on different fusion factor " + }, + { + "bbox": [ + 61, + 548, + 272, + 560 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 168, + 440, + 287, + 526 + ], + "blocks": [ + { + "bbox": [ + 190, + 430, + 262, + 439 + ], + "lines": [ + { + "bbox": [ + 190, + 430, + 262, + 439 + ], + "spans": [ + { + "bbox": [ + 190, + 430, + 262, + 439 + ], + "type": "text", + "content": "(d) CIFAR-FS, BERT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 440, + 287, + 526 + ], + "lines": [ + { + "bbox": [ + 168, + 440, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 168, + 440, + 287, + 526 + ], + "type": "image", + "image_path": "374e297665525f1366bed75e7c82c4e4bbf2abb9e9ae9eb526f286bf22771652.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 529, + 255, + 538 + ], + "lines": [ + { + "bbox": [ + 198, + 529, + 255, + 538 + ], + "spans": [ + { + "bbox": [ + 198, + 529, + 255, + 538 + ], + "type": "text", + "content": "(f) FC100, BERT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "and 5b. In Fig. 5c, we put support samples, i.e., red dots, mean vector of each class, i.e., blue stars, and query samples in the same picture. As we can see, support samples are close to the cluster edge rather than the mean center, which leads to difficulties in correctly classifying most samples in the query set. However, after being transformed by SemFew, the reconstructed prototypes are closer to the mean center of each cluster compared to the support set samples. This phenomenon demonstrates that our method can reasonably transform visual samples into class prototypes corresponding to their respective categories." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 243, + 437, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 437, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 437, + 255 + ], + "type": "text", + "content": "4.6. Fusion Factor Analysis" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "content": "The fusion factor, denoted as " + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "content": " in Eq. (6), significantly influences performance. To demonstrate, we systematically vary the value of " + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "content": " between 0 and 1, using a step size of 0.01, and present the accuracy trends across three benchmarks in Fig. 6. Notably, our model, i.e., blue line, gets the highest accuracy when " + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "content": " is large. Given that " + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "content": " represents the percentage of the reconstructed prototype " + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "content": " during fusion in Eq. (6), the experiment highlights that our model produces high-quality prototypes. This observation is further supported by the accuracy comparison when " + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 304, + 261, + 545, + 417 + ], + "type": "text", + "content": ". In such cases, our model consistently outperforms others, providing strong evidence for the effectiveness of our reconstructed prototypes." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 417, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 546, + 538 + ], + "type": "text", + "content": "Additionally, we add " + }, + { + "bbox": [ + 304, + 417, + 546, + 538 + ], + "type": "inline_equation", + "content": "S \\Rightarrow V" + }, + { + "bbox": [ + 304, + 417, + 546, + 538 + ], + "type": "text", + "content": " experiment for comparison, which generates visual samples rather than prototypes through a Conditional Generative Adversarial Network [34]. This experiment allows us to explore whether we can build robust prototypes indirectly by expanding the sample of the support set. As we can see, " + }, + { + "bbox": [ + 304, + 417, + 546, + 538 + ], + "type": "inline_equation", + "content": "S \\Rightarrow V" + }, + { + "bbox": [ + 304, + 417, + 546, + 538 + ], + "type": "text", + "content": ", i.e., yellow line, gets relatively low performance in all settings. The experiments indicate that producing the prototype by expanding then averaging the support set is not as effective as directly generating class prototypes." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "type": "text", + "content": "To unleash the potential of semantics in FSL, we propose the Semantic Evolution to automatically generate high-quality semantics from simple class names. Then, we design a simple Semantic Alignment Network to transform the concatenated visual features and high-quality semantics features into robust class prototypes for classification. The experimental results show that with high-quality semantics, the basic network can easily achieve greater performance compared to other state-of-the-art methods." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by 2023YFF1204901, NSFC-62076172, 2021DQ02-0903, 2023YFG0116." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28595" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "text", + "content": "[2] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 288, + 201 + ], + "type": "text", + "content": "[3] Irving Biederman. Recognition-by-components: a theory of human image understanding. *Psychol Rev*, 94(2):115, 1987. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "text", + "content": "[5] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. In CVPR, pages 23581-23591, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 345 + ], + "type": "text", + "content": "[6] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In ICCV, pages 9062-9071, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "text", + "content": "[7] Hao Cheng, Siyuan Yang, Joel Tianyi Zhou, Lanqing Guo, and Bihan Wen. Frequency guidance matters in few-shot learning. In ICCV, pages 11814-11824, 2023. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "type": "text", + "content": "[10] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. In NeurIPS, pages 21981-21993, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "text", + "content": "[11] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In ECCV, pages 329-347. Springer, 2022. 1, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "text", + "content": "[12] Thomas Elsken, Benedikt Staffler, Jan Hendrik Metzen, and Frank Hutter. Meta-learning of neural architectures for few-shot learning. In CVPR, pages 12365–12375, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Int. Conf. Mach. Learn., pages 1126-1135. PMLR, 2017. 1, 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[14] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In CVPR, pages 24575-24584, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[15] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In ICCV, pages 8059-8068, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[16] Fusheng Hao, Fengxiang He, Liu Liu, Fuxiang Wu, Dacheng Tao, and Jun Cheng. Class-aware patch embedding adaptation for few-shot image classification. In ICCV, pages 18905-18915, 2023. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. J R Stat Soc Ser C Appl Stat, 28(1):100-108, 1979. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "text", + "content": "[19] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. NeurIPS, 35:3582-3595, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "text", + "content": "[20] Ruibing Hou, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot classification. In NeurIPS, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "text", + "content": "[21] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "text", + "content": "[22] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In CVPR, pages 11719-11727, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "text", + "content": "[23] Jaekyeom Kim, Hyoungseok Kim, and Gunhee Kim. Model-agnostic boundary-adversarial sampling for test-time generalization in few-shot learning. In ECCV, pages 599–617. Springer, 2020. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 325, + 545, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 356 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 359, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 359, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 359, + 545, + 392 + ], + "type": "text", + "content": "[25] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML Deep Learn. workshop. Lille, 2015. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 394, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 416 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "type": "text", + "content": "[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In CVPR, pages 10657-10665, 2019. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 452, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 485 + ], + "type": "text", + "content": "[28] Aoxue Li, Weiran Huang, Xu Lan, Jiashi Feng, Zhenguo Li, and Liwei Wang. Boosting few-shot learning with adaptive margin loss. In CVPR, pages 12576-12584, 2020. 1, 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "text", + "content": "[29] Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features for few-shot learning by category traversal. In CVPR, pages 1-10, 2019. 1, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 531, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 576 + ], + "type": "text", + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "text", + "content": "[31] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, Georgia, USA, 2013. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "text", + "content": "[32] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In IEEE Win. Conf. Appl. Comput. Vis., pages 2218-2227, 2020. 1, 2, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 667, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 667, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 667, + 545, + 689 + ], + "type": "text", + "content": "[33] George A Miller. Wordnet: a lexical database for english. Commun. ACM, 38(11):39-41, 1995. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[34] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 8" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28596" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[35] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. arXiv preprint arXiv:1707.03141, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 128 + ], + "type": "text", + "content": "[36] OpenAI. Chatgpt. https://chat.openai.com, 2023.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "text", + "content": "[37] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. In NeurIPS, 2018. 1, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "text", + "content": "[38] Zhimao Peng, Zechao Li, Junge Zhang, Yan Li, Guo-Jun Qi, and Jinhui Tang. Few-shot image recognition with knowledge transfer. In ICCV, pages 441-449, 2019. 1, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 198, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 262 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn., pages 8748-8763. PMLR, 2021. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "type": "text", + "content": "[40] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "type": "text", + "content": "[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 342, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 387 + ], + "type": "text", + "content": "[42] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 410 + ], + "type": "text", + "content": "[43] Victor Garcia Satorras and Joan Bruna Estrach. Few-shot learning with graph neural networks. In ICLR, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "text", + "content": "[44] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, pages 8247-8255, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "text", + "content": "[45] Hongyang Gao Siyuan Sun. Meta-adam: A meta-learned adaptive optimizer with momentum for few-shot learning. In NeurIPS, 2023. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "type": "text", + "content": "[46] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[47] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In NeurIPS, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "text", + "content": "[48] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199–1208, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "text", + "content": "[49] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In ECCV, pages 266-282. Springer, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[50] Pavel Tokmakov, Yu-Xiong Wang, and Martial Hebert. Learning compositional representations for few-shot recognition. In ICCV, pages 6372-6381, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[51] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 600 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. J Mach Learn Res, 9(11), 2008. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "text", + "content": "[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In NeurIPS, 2016. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "text", + "content": "[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[55] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In IJCAI, pages 1075-1081. International Joint Conferences on Artificial Intelligence Organization, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "type": "text", + "content": "[56] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, pages 1831-1840, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[57] Chen Xing, Nagar Rostamzadeh, Boris Oreshkin, and Pedro O O Pinheiro. Adaptive cross-modal few-shot learning. In NeurIPS, 2019. 1, 2, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "text", + "content": "[58] Jingyi Xu and Hieu Le. Generating representative samples for few-shot classification. In CVPR, pages 9003-9013, 2022. 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "type": "text", + "content": "[59] Kun Yan, Zied Bouraoui, Ping Wang, Shoaib Jameel, and Steven Schockaert. Aligning visual prototypes with bert embeddings for few-shot learning. In ICMR, pages 367-375, 2021. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "text", + "content": "[60] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In CVPR, pages 8808-8817, 2020. 1, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "text", + "content": "[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In CVPR, pages 3754-3762, 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "text", + "content": "[62] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Differentiable earth mover's distance for few-shot learning. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 524, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 556 + ], + "type": "text", + "content": "[64] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In CVPR, pages 20061-20070, 2023. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 558, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 600 + ], + "type": "text", + "content": "[65] Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast context adaptation via meta-learning. In Int. Conf. Mach. Learn., pages 7693-7702. PMLR, 2019. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28597" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_content_list.json b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..874535310060791f208d5a20b8b3c63aa0546fbe --- /dev/null +++ b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_content_list.json @@ -0,0 +1,2268 @@ +[ + { + "type": "text", + "text": "SinSR: Diffusion-Based Image Super-Resolution in a Single Step", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yufei Wang $^{1,2\\dagger}$ , Wenhan Yang $^{3}$ , Xinyuan Chen $^{2*}$ , Yaohui Wang $^{2}$ , Lanqing Guo $^{1}$ , Lap-Pui Chau $^{4}$ , Ziwei Liu $^{1}$ , Yu Qiao $^{2}$ , Alex C. Kot $^{1}$ , Bihan Wen $^{1*}$", + "bbox": [ + 158, + 179, + 803, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanyang Technological University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ PengCheng Laboratory $^{4}$ The Hong Kong Polytechnic University", + "bbox": [ + 158, + 215, + 810, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While super-resolution (SR) methods based on diffusion models exhibit promising results, their practical application is hindered by the substantial number of required inference steps. Recent methods utilize the degraded images in the initial state, thereby shortening the Markov chain. Nevertheless, these solutions either rely on a precise formulation of the degradation process or still necessitate a relatively lengthy generation path (e.g., 15 iterations). To enhance inference speed, we propose a simple yet effective method for achieving single-step SR generation, named SinSR. Specifically, we first derive a deterministic sampling process from the most recent state-of-the-art (SOTA) method for accelerating diffusion-based SR. This allows the mapping between the input random noise and the generated high-resolution image to be obtained in a reduced and acceptable number of inference steps during training. We show that this deterministic mapping can be distilled into a student model that performs SR within only one inference step. Additionally, we propose a novel consistency-preserving loss to simultaneously leverage the ground-truth image during the distillation process, ensuring that the performance of the student model is not solely bound by the feature manifold of the teacher model, resulting in further performance improvement. Extensive experiments conducted on synthetic and real-world datasets demonstrate that the proposed method can achieve comparable or even superior performance compared to both previous SOTA methods and the teacher model, in just one sampling step, resulting in a remarkable up to $\\times 10$ speedup for inference. Our code will be released at https://github.com/wyf0912/SinSR/.", + "bbox": [ + 75, + 311, + 473, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image super-resolution (SR) aims to reconstruct a high-resolution image from a given low-resolution (LR) counterpart [45]. Recently, diffusion models, known for their effectiveness in modeling complex distributions, have gained", + "bbox": [ + 75, + 811, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd8e4b209a435ae13140b3fba7835f61388989ebeea4f7afb718251b2dc8668c.jpg", + "image_caption": [ + "(a) Input LR image (MUSIQ↑: 23.89)" + ], + "image_footnote": [], + "bbox": [ + 503, + 285, + 694, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b5e145289f6883254fe447a846042697e225fbbdda5be8668baa6374196e38a.jpg", + "image_caption": [ + "(b) ResShift-1 (47.43)" + ], + "image_footnote": [], + "bbox": [ + 696, + 285, + 790, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b7b84c675c6744abd1f45ab8ce9104363d32fd706744444fea0c6305a0928b6.jpg", + "image_caption": [ + "(c) ResShift-5 (40.32)" + ], + "image_footnote": [], + "bbox": [ + 792, + 285, + 887, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5888dc01a559b8987a6f5965eb66b701b193ea8eb4647c4ed6555700aae4d96.jpg", + "image_caption": [ + "(d) ResShift-15 (62.83)", + "Figure 1. A comparison between the most recent SOTA method ResShift [46] for the acceleration of diffusion-based SR and the proposed method. We achieve on-par or even superior perceptual quality using only one inference step. (\"-N\" behind the method name represents the number of inference steps, and the value in the bracket is the quantitative result measured by MUSIQ↑ [15].)" + ], + "image_footnote": [], + "bbox": [ + 696, + 364, + 790, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5970d4aec436000b6a7d64ff488b2de2af1798f35523c25af0e440f6b382a26.jpg", + "image_caption": [ + "(e) Ours-1 (68.05)" + ], + "image_footnote": [], + "bbox": [ + 792, + 364, + 887, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "widespread adoption and demonstrated remarkable performance in SR tasks, particularly in terms of perceptual quality.", + "bbox": [ + 498, + 549, + 893, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, current strategies for employing diffusion models can be broadly categorized into two streams: concatenating the LR image to the input of the denoiser in the diffusion models [32, 33], and adjusting the inverse process of a pre-trained diffusion model [4, 5, 14]. Despite achieving promising results, both strategies encounter computational efficiency issues. Notably, the initial state of these conditional diffusion models is a pure Gaussian noise without using the prior knowledge from the LR image. Consequently, a substantial number of inference steps are required to achieve satisfactory performance, significantly hindering the practical applications of diffusion-based SR techniques.", + "bbox": [ + 496, + 580, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efforts have been made to enhance the sampling efficiency of diffusion models, leading to various techniques proposed [22, 28, 37]. However, in the realm of low-level vision where maintaining high fidelity is critical, these techniques often fall short as they achieve acceleration at the cost of performance. More recently, innovative techniques have emerged to reformulate the diffusion process in image restoration tasks, focusing on improving the signal-to-noise ratio of the initial diffusion state and thereby shorten the", + "bbox": [ + 496, + 763, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done as an intern at Shanghai AI Lab. * Corresponding authors.", + "bbox": [ + 76, + 886, + 457, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25796", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72e3a58ba5bb88e91d2ec16f595b397702f31861550a2cfdca2af93af0da45d7.jpg", + "image_caption": [ + "(a) LR image" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 277, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eadaeb709f05d82e1802e314e2b39992e60a48b66b8983994dd2e74ad0b331aa.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 370, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/644fce7cbc9f8f087fbce537cc0a2a959fecaf4d8a0900dda3cf4e7d7373204c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 88, + 459, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b3a4837de073e3d4f7b421e613cae9b3f422881204fcd09513dfb5cd1796250f.jpg", + "image_caption": [ + "(d)", + "Figure 2. An illustration of the generative ability of the proposed method in only one step. Given the same LR image (Fig. (a) and (b)), by using different noise added to the input, HR images (Fig. (c)-(e)) with different details are generated, e.g., eyes of different shapes and colors. Best zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 284, + 167, + 370, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce0a5599d0d6117c09d4c321e7ae24700d926448323de4c6f253999f4d52413b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 167, + 460, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Markov chain. For instance, [43] initiates the denoising diffusion process with the input noisy image, while in the SR task, [46] models the initial step as a combination of the LR image and random noise. Nonetheless, even in these most recent works [43, 46], limitations persist. For instance, while [43] shows promising results within just three inference steps, it requires a clear formulation of the image degradation process. Besides, [46] still necessitates 15 inference steps and exhibits degraded performance with noticeable artifacts if the number of inference steps is further reduced.", + "bbox": [ + 75, + 352, + 470, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce a novel approach that can generate high-resolution (HR) images in only one sampling step, without compromising the diversity and perceptual quality of the diffusion model, as shown in Fig. 1 and Fig. 2. Specifically, we propose to directly learn a well-paired bi-directional deterministic mapping between the input random noise and the generated HR image from a teacher diffusion model. To accelerate the generation of well-matched training data, we first derive a deterministic sampling strategy from the most recent state-of-the-art work [46], designed for accelerating diffusion-based SR, from its original stochastic formulation. Additionally, we propose a novel consistency-preserving loss to leverage ground-truth images, further enhancing the perceptual quality of the generated HR images by minimizing the error between ground-truth (GT) images and those generated from the predicted initial state. Experimental results demonstrate that our method achieves comparable or even better performance compared to SOTA methods and the teacher diffusion model [46], while greatly reducing the number of inference steps from 15 to 1, resulting in up to a $\\times 10$ speedup in inference.", + "bbox": [ + 75, + 505, + 472, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 94, + 823, + 434, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We accelerate the diffusion-based SR model to a single inference step with comparable or even superior performance for the first time. Instead of shortening the Markov chain of the generation process, we propose a simple yet", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "effective approach that directly distills a deterministic generation function into a student network.", + "bbox": [ + 511, + 90, + 893, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To further fasten training, we derive a deterministic sampling strategy from the recent SOTA method [46] on accelerating the SR task, enabling efficient generation of well-matched training pairs.", + "- We propose a novel consistency-preserving loss that can utilize the ground-truth images during training, preventing the student model from only focusing on fitting the deterministic mapping of the teacher diffusion model, therefore leading to better performance.", + "- Extensive experiments on both synthetic and real-world datasets show that our proposed method can achieve comparable or even superior performance compared to SOTA methods and the teacher diffusion model, while greatly reducing the number of inference steps from 15 to 1." + ], + "bbox": [ + 500, + 121, + 893, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 338, + 640, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Super-Resolution", + "text_level": 1, + "bbox": [ + 500, + 362, + 720, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rise of deep learning, deep learning-based techniques gradually become the mainstream of the SR task [8, 45]. One prevalent approach of early works is to train a regression model using paired training data [1, 2, 16, 44]. While the expectation of the posterior distribution can be well modeled, they inevitably suffer from the over-smooth problem [17, 27, 34]. To improve the perceptual quality of the generated HR images, generative-based SR models attract increasing attention, e.g., autoregressive-based models [6, 26, 29, 30]. While significant improvements are achieved, the computational cost of autoregressive models is usually large. Subsequently, normalizing flows [23, 42] are demonstrated to have good perceptual quality under an efficient inference process, while its network design is restricted by the requirements of the invertibility and ease of calculation. Besides, GAN-based methods also achieve great success in terms of perceptual quality [9, 13, 17, 27, 34]. However, the training of GAN-based methods is usually unstable. Recently, diffusion-based models have been widely investigated in SR [4, 5, 14, 32, 33]. The diffusion-based SR methods can be roughly summarized into two categories, concatenating the LR image to the input of the denoiser [32, 33], and modifying the backward process of a pre-trained diffusion model [4, 5, 10, 14]. While promising results are achieved, they rely on a large number of inference steps, which greatly hinders the application of diffusion-based models.", + "bbox": [ + 496, + 385, + 895, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Acceleration of Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 786, + 787, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the acceleration of diffusion models has attracted more and more attention. Several algorithms are proposed for general diffusion models [22, 28, 37, 38] and proved quite effective for image generations. One intuitive strategy among them is to distill the diffusion models to a student model. However, the huge training overhead to solve", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25797", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e1345c93f2d7de5617f9168178235aa1d593c40d6f7cc5808e2225192c1c149f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 90, + 151, + 146 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3d61da1c6348d7e70ca0847167b48334465a92ce0ff65f1e01b22d5a64869b23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 93, + 248, + 145 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41ad4a4cf1e7fc33a6d8a5df4bba512b7f3050650aa25bb9f7bc08cbc221ffed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 95, + 310, + 145 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e86bf3426ed1ad88cb6ea836e95c372bc000a4c581b7501306afcdbc62e7b08c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 93, + 367, + 146 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3f7cc6c4674ecf76bf119a0c35525fef2c8bcc4f237fc8d4b4bdec3458d98883.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 93, + 416, + 146 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c10cb5336b672ab82e8dfd1d4502eec08a3dd4a148f46593d8ead66551dbc63b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 93, + 454, + 146 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/70a4ad8a056d213f9351a3351109791d31035f6f57b4ed5a54f5254fa91fcbc6.jpg", + "image_caption": [ + "(a) The inference of SR3 [33] starts from a pure noise, which requires a large number of inference steps (T=100 after using DDIM [37])." + ], + "image_footnote": [], + "bbox": [ + 81, + 181, + 465, + 252 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09d2e12d0af47b13ca8344dfebb9d6e7fc9114eec488ac78ba2270f8e6d90a95.jpg", + "image_caption": [ + "(b) The recent SOTA method ResShift [46] shortens the Markov chain to speed up the inference process by incorporating the information of the LR image $y$ to the initial state $x_{T}$ $(\\mathrm{T} = 15)$ .", + "(c) A simplified pipeline of the proposed method $\\text{SinSR}$ (distill only). It directly learns the deterministic mapping between $x_{T}$ and $x_{0}$ , therefore the inference process can be further compressed into only one step $(\\mathrm{T} = 1)$ .", + "Figure 3. A comparison between the vanilla diffusion-based SR method [33], a most recent method for acceleration of the diffusion-based SR [46], and the proposed one-step SR. Different from recent works that shorten the Markov chain to speed up the inference process [43, 46], the proposed method directly learns the deterministic generation process and the details can be found in Fig. 4." + ], + "image_footnote": [], + "bbox": [ + 81, + 296, + 467, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the ordinary differential equation (ODE) of the inference process makes this scheme less attractive on a large-scale dataset [24]. To alleviate the training overhead, progressive distillation strategies are usually adopted [25, 35]. Meanwhile, instead of simply simulating the behavior of a teacher diffusion model through distillation, better inference paths are explored in an iterative manner [20, 21]. While progressive distillation effectively decreases the training overhead, the error accumulates at the same time, leading to an obvious performance loss in SR. Most recently, targeting the image restoration task, some works reformulate the diffusion process by either using the knowledge of degradation process [43] or a pre-defined distribution of the initial state [46], yielding a shortened Markov chain of the generation process and better performance than directly applying DDIM [37] in low-level tasks. However, they either require a clear formulation of the degradation or still require a relatively large number of inference steps.", + "bbox": [ + 75, + 529, + 472, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Motivation", + "text_level": 1, + "bbox": [ + 76, + 815, + 192, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. Given an LR image $y$ and its corresponding HR image $x_0$ , existing diffusion-based SR methods aim to model the conditional distribution $q(x_0|y)$ through a Markov chain where a forward process is usually defined", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as $q(x_{t}|x_{t - 1}) = \\mathcal{N}(x_{t};\\sqrt{1 - \\beta_{t}} x_{t - 1},\\beta_{t}I)$ with an initial state $x_{T}\\sim \\mathcal{N}(0,I)$ . The role of the diffusion model can be regarded as transferring the input domain (standard Gaussian noise) to the HR image domain conditioned on the LR image. Since the matching relationship between $x_{T}$ and $x_0$ is unknown, usually a diffusion model [11, 21, 33] through an iterative manner is required to learn/infer from an unknown mapping between $x_{T}$ and $x_0$ . Our method is grounded in the idea that having an SR model that effectively captures the conditional distribution $q(x_0|y)$ and establishes a deterministic mapping between $x_{T}$ and $\\hat{x} 0$ given an LR image $y$ , we can streamline the inference process to a single step by employing another network, denoted as $f_{\\hat{\\theta}}$ , to learn the correspondence between $\\hat{x}_0$ and $x_{T}$ , as illustrated in Fig. 3. Distillation for diffusion SR models: less is more. While the concept of distilling the mapping between $x_{T}$ and $\\hat{x}_0$ to a student network has been previously explored [21], its application to SR introduces several challenges:", + "bbox": [ + 496, + 90, + 893, + 363 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The training overhead becomes substantial for one-step distillation due to a large number of inference steps of previous models, e.g., LDM [32] still need 100 steps after using DDIM [37] for inference to generate high-quality pairs $(\\hat{x}_0, x_T, y)$ as the training data of the student model.", + "- The performance degradation is attributed to the introduction of a more intricate distillation strategy involving iteration. For example, to reduce the training overhead, an iterative distillation strategy [35] is adopted which gradually decreases the number of inference steps during training. However, despite achieving satisfactory results in generation tasks, the cumulative error significantly impacts the fidelity of the SR results, as SR tasks are relatively more sensitive to image quality." + ], + "bbox": [ + 500, + 364, + 893, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the aforementioned two challenges, we propose to distill the diffusion SR process into a single step in a simple but effective way based on the following observations. More details of the observations can be seen in Sec. 5.3", + "bbox": [ + 496, + 575, + 893, + 635 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate that the most recent SOTA method for accelerating the diffusion-based SR [46], which achieves comparable performance in 15 steps as LDM [32] in 100 DDIM steps, has a deterministic mapping between $x_{T}$ and $x_{0}$ . Besides, the greatly reduced number of inference steps and the existence of the deterministic mapping make the training of a single-step distillation possible as shown in Fig. 6 and Table 4.", + "- Learning the mapping between $x_{T}$ and $\\hat{x}_0$ is found to be easier than denoising $x_{t}$ under different noise levels as shown in Table 5. Therefore, it is feasible to directly learn the mapping between $x_{T}$ and $\\hat{x}_0$ so that the accumulated error by the iterative distillation can be avoided.", + "- Due to the accumulated error, a more sophisticated distillation strategy (iterative-based) does not contribute to the improvement in our setting as shown in Table 6." + ], + "bbox": [ + 500, + 638, + 893, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The organization of the following sections is as follows:", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25798", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b04aa19510ca2f5d685799789a32c5c959e3c54abf196e81c92745e97f94e8a9.jpg", + "image_caption": [ + "(a) One-step bi-directional distillation" + ], + "image_footnote": [], + "bbox": [ + 86, + 61, + 566, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e6ce58296fdae0744d7b6df8940dc397882d0084bb09b77366c99ee7facaf1c.jpg", + "image_caption": [ + "(b) Consistency preserving loss", + "Figure 4. The overall framework of the proposed method. By minimizing $\\mathcal{L}_{dist}$ and $\\mathcal{L}_{inverse}$ , the student network $f_{\\hat{\\theta}}$ learns the deterministic bi-directional mapping between $x_{T}$ and $\\hat{x}_0$ obtained from a pre-trained teacher diffusion model in one step. Meanwhile, the proposed consistency preserving loss $\\mathcal{L}_{gt}$ is optimized during training to utilize the information from the GT images to pursue better perceptual quality instead of simply fitting the deterministic mappings from the teacher model. Specifically, the GT image is first converted to its latent code $\\hat{x}_T = f_{\\hat{\\theta}}(x_0,y,0)$ , and then converted back to calculate its reconstruction loss $L_{MSE}(f_{\\hat{\\theta}}(\\hat{x}_T,y,T),x_0)$ ." + ], + "image_footnote": [], + "bbox": [ + 581, + 61, + 885, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we first demonstrate that ResShift [46], in which the inference process is originally stochastic, can be converted to a deterministic model without retraining in Sec 4.1, and then the proposed consistency preserving distillation in Sec 4.2.", + "bbox": [ + 75, + 353, + 470, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 75, + 430, + 210, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Deterministic Sampling", + "text_level": 1, + "bbox": [ + 75, + 455, + 295, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A core difference between ResShift [46] and LDM [32] is the formulation of the initial state $x_{T}$ . Specifically, in ResShift [46], the information from the LR image $y$ is integrated into the diffusion step $x_{t}$ as follows", + "bbox": [ + 75, + 479, + 470, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}, y\\right) = \\mathcal {N} \\left(x _ {t}; x _ {0} + \\eta_ {t} \\left(y - x _ {0}\\right), \\kappa^ {2} \\eta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 554, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta_{t}$ is a serial of hyper-parameters that monotonically increases with the timestep $t$ and obeys $\\eta_T\\rightarrow 1$ and $\\eta_0\\to 0$ . As such, the inverse of the diffusion process starts from an initial state with rich information from the LR image $y$ as follows $x_{T} = y + \\kappa \\sqrt{\\eta_{T}}\\epsilon$ where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . To generate a HR image $x$ from a given image $y$ , the original inverse process of [46] is as follows", + "bbox": [ + 75, + 585, + 468, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, y\\right) = \\mathcal {N} \\left(x _ {t - 1} \\mid \\mu_ {\\theta} \\left(x _ {t}, y, t\\right), \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 703, + 468, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}(x_t,y,t)$ is reparameterized by a deep network. As shown in Eq. 2, given an initial state $x_{T} = y + \\kappa \\sqrt{\\eta_{T}}\\epsilon$ the generated image is stochastic due to the existence of the random noise during the sampling from $p_{\\theta}(x_{t - 1}|x_t,y)$ . Inspired by DDIM sampling [37], we find that a non-Markovian reverse process $q(x_{t - 1}|x_t,x_0,y)$ exists which keeps the marginal distribution $q(x_{t}|x_{0},y)$ unchanged so", + "bbox": [ + 75, + 743, + 468, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that it can be directly adopted to a pre-trained model. The reformulated deterministic reverse process is as follows", + "bbox": [ + 498, + 353, + 888, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}, y\\right) = \\delta \\left(k _ {t} x _ {0} + m _ {t} x _ {t} + j _ {t} y\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 396, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta$ is the unit impulse, and $k_{t},m_{t},j_{t}$ are as follows", + "bbox": [ + 498, + 424, + 870, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} m _ {t} = \\sqrt {\\frac {\\eta_ {t - 1}}{\\eta_ {t}}} \\\\ j _ {t} = \\eta_ {t - 1} - \\sqrt {\\eta_ {t - 1} \\eta_ {t}} \\\\ k _ {t} = 1 - \\eta_ {t - 1} + \\sqrt {\\eta_ {t - 1} \\eta_ {t}} - \\sqrt {\\frac {\\eta_ {t - 1}}{\\eta_ {t}}}. \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 450, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The details of the derivation can be found in the supplementary material. As a consequence, for inference, the reverse process conditioned on $y$ is reformulated as follows", + "bbox": [ + 498, + 525, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t - 1} = k _ {t} \\hat {x} _ {0} + m _ {t} x _ {t} + j _ {t} y \\\\ = k _ {t} f _ {\\theta} \\left(x _ {t}, y, t\\right) + m _ {t} x _ {t} + j _ {t} y, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 583, + 816, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta}(x_t,y,t)$ is the predicted HR image from a pretrained ResShift [46] model. By sampling from the reformulated process in Eq. 5, a deterministic mapping between $x_{T}$ (or $\\epsilon$ ) and $\\hat{x}_0$ can be obtained and is denoted as $F_{\\theta}(x_T,y)$ .", + "bbox": [ + 498, + 628, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Consistency Preserving Distillation", + "text_level": 1, + "bbox": [ + 500, + 699, + 803, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vanilla distillation. We propose utilizing a student network $f_{\\hat{\\theta}}$ to learn the deterministic mapping $F_{\\theta}$ between the random initialized state $x_{T}$ and its deterministic output $F_{\\theta}(x_{T},y)$ from a teacher diffusion model. The vanilla distillation loss is defined as follows", + "bbox": [ + 498, + 723, + 890, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(x _ {T}, y, T\\right), F _ {\\theta} \\left(x _ {T}, y\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 811, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\hat{\\theta}}(x_T,y,T)$ is the student network that directly predicts the HR image in only one step, and $F_{\\theta}$ represents the proposed deterministic inference process of ResShift [46] in Sec. 4.1 through an iterative manner using a pre-trained", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "For ease of presentation, the LR image is $y$ is pre-upsampled to the same spatial resolution with the HR image $x$ . Besides, similar to [32, 46], the diffusion is conducted in the latent space.", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25799", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "network parameterized by $\\theta$ . We observe that the student model trained solely with the distillation loss in Eq. 6 already achieves promising results in just one inference step, as indicated by “(distill only)” in the result tables.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regularization by the ground-truth image. A limitation of the aforementioned vanilla distillation strategy is that the GT image is not utilized during training, thereby restricting the upper performance bound of the student model. To further enhance the student's performance, we propose a novel strategy that incorporates a learned inversion of the HR image to provide additional regularization from the ground-truth images. In addition to the vanilla distillation loss, the student network concurrently learns the inverse mapping during training by minimizing the following loss,", + "bbox": [ + 75, + 152, + 470, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n v e r s e}} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(F _ {\\theta} \\left(x _ {T}, y\\right), y, 0\\right), x _ {T}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 314, + 470, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the last parameter of $f_{\\hat{\\theta}}$ is set from $T$ in Eq. 6 to 0, indicating that the model is predicting the inversion instead of the $\\hat{x}_0$ . Then the GT image $x_0$ can be employed to regularize the output SR image given its predicted inversion $\\hat{x}_T$ as follows", + "bbox": [ + 75, + 342, + 472, + 416 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {T} = \\operatorname {d e t a c h} (f _ {\\hat {\\theta}} (x _ {0}, y, 0))\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 426, + 349, + 444 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g t} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(\\hat {x} _ {T}, y, T\\right), x _ {0}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 444, + 468, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{gt}$ is the proposed consistency preserving loss. By reusing $f_{\\hat{\\theta}}$ to learn both $f_{\\hat{\\theta}}(\\cdot, \\cdot, T)$ and $f_{\\hat{\\theta}}(\\cdot, \\cdot, 0)$ simultaneously, we can initialize the parameter $\\hat{\\theta}$ of the student model from the teacher one $\\theta$ to speed up the training.", + "bbox": [ + 75, + 473, + 468, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall training objective. The student network is trained to minimize the aforementioned three losses at the same time as follows", + "bbox": [ + 75, + 536, + 468, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\theta} = \\arg \\min _ {\\hat {\\theta}} \\mathbb {E} _ {y, x _ {0}, x _ {T}} \\left[ \\mathcal {L} _ {\\text {d i s t i l l}} + \\mathcal {L} _ {\\text {r e v e r s e}} + \\mathcal {L} _ {g t} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 590, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the losses are defined in Eq. 6, 7, and 8 respectively. We assign equal weight to each loss term, and ablation studies are in the supplementary material. The overall of the proposed method is summarized in Algorithm 1 and Fig. 4.", + "bbox": [ + 75, + 627, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 700, + 200, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 726, + 263, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. For a fair comparison, we follow the same experimental setup and backbone design as that in [46]. Specifically, the main difference is that we finetuned the model for 30K iterations instead of training from scratch for 500K in [46]. We find that the student model can converge quickly so that even if for each iteration we need extra time to solve the ODE to get paired training data, the overall training time is still much shorter than retraining a model from scratch following [46]. We train the models on the training set of ImageNet [7] following the same pipeline", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Training" + ], + "code_body": "Require: Pre-trained teacher diffusion model $f_{\\theta}$ \nRequire: Paired training set $(X,Y)$ \n1: Init $f_{\\hat{\\theta}}$ from the pre-trained model, i.e., $\\hat{\\theta}\\gets \\theta$ \n2: while not converged do \n3: sample $x_0,y\\sim (X,Y)$ \n4: sample $\\epsilon \\sim \\mathcal{N}(0,\\kappa^2\\eta_T\\mathbf{I})$ \n5: $x_{T} = y + \\epsilon$ \n6: for $t = T,T - 1,\\dots,1$ do \n7: if $t = 1$ then \n8: $\\hat{x}_0 = f_\\theta (x_1,y,1)$ \n9: else \n10: $x_{t - 1} = k_t f_\\theta (x_t,y,t) + m_tx_t + j_t y$ \n11: end if \n12: end for \n13: $\\mathcal{L}_{distill} = L_{MSE}(f_{\\hat{\\theta}}(x_T,y,T),\\hat{x}_0)$ \n14: $\\mathcal{L}_{inverse} = L_{MSE}(f_{\\hat{\\theta}}(\\hat{x}_0,y,0),x_T)$ \n15: $\\hat{x}_T = f_\\hat{\\theta} (x_0,y,0)$ \n16: $\\mathcal{L}_{gt} = L_{MSE}(f_{\\hat{\\theta}}(\\operatorname {detach}(\\hat{x}_T),y,T),x_0)$ \n17: $\\mathcal{L} = \\mathcal{L}_{distill} + \\mathcal{L}_{inverse} + \\mathcal{L}_{gt}$ \n18: Perform a gradient descent step on $\\nabla_{\\hat{\\theta}}\\mathcal{L}$ \n19: end while \n20: return The student model $f_{\\hat{\\theta}}$", + "bbox": [ + 500, + 107, + 890, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with ResShift [46] where the degradation model is adopted from RealESRGAN [41].", + "bbox": [ + 496, + 472, + 890, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared methods. We compare our method with several representative SR models, including RealSR-JPEG [12], ESRGAN [40], BSRGAN [47], SwinIR [18], RealESRGAN [41], DASR [19], LDM [32], and ResShift [46]. For a comprehensive comparison, we further evaluate the performance of diffusion-based models LDM [32] and ResShift [46] with a reduced number of sampling steps. Besides, we compare the proposed method with Rectified-Flow [21], a SOTA method that can compress the generation process into a single step, in Table 6.", + "bbox": [ + 496, + 503, + 893, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For the evaluation of the proposed method on the synthetic testing dataset with reference images, we utilize PSNR, SSIM, and LPIPS [48] to measure the fidelity performance. Besides, two recent SOTA non-reference metrics are used to justify the realism of all the images, i.e., CLIP-IQA [39] which leverages a CLIP model [31] pre-trained on a large-scale dataset (Laion400M [36]) and MUSIQ [15].", + "bbox": [ + 496, + 655, + 893, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 771, + 700, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation on real-world datasets. RealSR [3] and RealSet65 [46] are adopted to evaluate the generalization ability of the model on unseen real-world data. Specifically, in RealSR [3], there are 100 real images captured by two different cameras in different scenarios. Besides, RealSet65 [46] includes 65 LR images in total, collected from widely used datasets and the internet. The results on these two datasets", + "bbox": [ + 496, + 794, + 893, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25800", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5ed0296b37d4ce3e6598b5894c72043018d6208e02b4d27d55637544ded7f6a.jpg", + "image_caption": [ + "(a) LR input" + ], + "image_footnote": [], + "bbox": [ + 112, + 53, + 331, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/580e65fc391fcad3e31a61acd43bde698b425c15d37943e46d23f6375f59f260.jpg", + "image_caption": [ + "(b) BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 334, + 53, + 437, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f8753f7fcd123ee25997f728b3f8627b1d03615981b4d8f830c0206838dbd7b1.jpg", + "image_caption": [ + "(c) RealESRGAN" + ], + "image_footnote": [], + "bbox": [ + 439, + 53, + 542, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/beec3d4f1e56ce25a7b87889c5d300d25a04514e5cf04cf698802b587256bb58.jpg", + "image_caption": [ + "(d) SwinIR" + ], + "image_footnote": [], + "bbox": [ + 544, + 53, + 648, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3a65906fc0e119f0257d2ffe72d069e5f60498c3671f48a18545266985335b9.jpg", + "image_caption": [ + "(e) DASR" + ], + "image_footnote": [], + "bbox": [ + 650, + 53, + 753, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b6e70bd8733be0840f34790dd9583292f655bbe86df25415d64ca7774606a14.jpg", + "image_caption": [ + "(f) LDM-100" + ], + "image_footnote": [], + "bbox": [ + 754, + 53, + 857, + 136 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17a59811229d322c1aa2787f6eb9355ca004ed286dc3783c0dad5e7ee1a198dc.jpg", + "image_caption": [ + "(a) LR input" + ], + "image_footnote": [], + "bbox": [ + 112, + 239, + 331, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/540c93fec6a816908c0a12d32deee2e754591e16af15f64913f4b892bf37237f.jpg", + "image_caption": [ + "(g) LDM-30" + ], + "image_footnote": [], + "bbox": [ + 334, + 146, + 437, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fff71557f5e5e0d733cf2c268436554a1e973f142c94d3862d8a1e78545e8c5a.jpg", + "image_caption": [ + "(h) LDM-15" + ], + "image_footnote": [], + "bbox": [ + 439, + 146, + 542, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/72eab5a26d1f69e3ec0b3f7228dda200db2344027a9da564ac45c8ae5374d0e6.jpg", + "image_caption": [ + "(i) ResShift-15" + ], + "image_footnote": [], + "bbox": [ + 544, + 146, + 648, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0529e6bbc423f940ba942b949cdffea65308068e762ebf397ce3d772735d2e97.jpg", + "image_caption": [ + "(j) ResShift-5" + ], + "image_footnote": [], + "bbox": [ + 650, + 146, + 753, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cee4c38d8e016b2baaec3d4fca6aac4ecb4040ead2ee2babc2f00750905a1d4.jpg", + "image_caption": [ + "(k) Ours-1" + ], + "image_footnote": [], + "bbox": [ + 754, + 146, + 857, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c9ef5256714e761b8a97dc7c330b509a73468bf04b67d86bf956a888f7526a6.jpg", + "image_caption": [ + "Figure 5. Visual comparison on real-world samples. Please zoom in for more details.", + "(a) LR input" + ], + "image_footnote": [], + "bbox": [ + 109, + 422, + 330, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(a) LR input", + "bbox": [ + 189, + 593, + 251, + 602 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2660b2aa9542303ccafed587fe8deae32408c0e886d2a283326414134f44e89a.jpg", + "image_caption": [ + "(b) BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 333, + 239, + 437, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a7972bb939a3575d336ca78b04f981bcab8c0d37823bb29c2ddb65f4d1501f7e.jpg", + "image_caption": [ + "(c) RealESRGAN" + ], + "image_footnote": [], + "bbox": [ + 439, + 239, + 542, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ac65be59d92ca1ddafee366c833329d32b3638e3b1b66aa6edfd1464b8af59e.jpg", + "image_caption": [ + "(d) SwinIR" + ], + "image_footnote": [], + "bbox": [ + 544, + 239, + 648, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd18df93c60f2d6dd94edbfa37a81a876cb0a8ae93b8924ca1dcb278204f2bcd.jpg", + "image_caption": [ + "(e) DASR" + ], + "image_footnote": [], + "bbox": [ + 650, + 239, + 751, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b3e29b83d86dd71ee339850569df2cb7f617e6f836367af9c8587fdbdfc5e5a8.jpg", + "image_caption": [ + "(f) LDM-100" + ], + "image_footnote": [], + "bbox": [ + 754, + 239, + 857, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea6f0f0495c04ad25223c915695514d36c31ddee6f76c98a7242b4b379835670.jpg", + "image_caption": [ + "(g) LDM-30" + ], + "image_footnote": [], + "bbox": [ + 334, + 329, + 437, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b70570e1df60f4d0b17e93f29f76f2c39575f71fe30053d13a804de910bb39e8.jpg", + "image_caption": [ + "(h) LDM-15" + ], + "image_footnote": [], + "bbox": [ + 439, + 329, + 542, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5bc4aacfd3c67882681afb57010c8ead1bb12961bc883beac500c61fd27c3604.jpg", + "image_caption": [ + "(i) ResShift-15" + ], + "image_footnote": [], + "bbox": [ + 544, + 329, + 648, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/84f519c79d852450507766d1fe7da908fd63073681cd1c8825d4c069e6cd8ad6.jpg", + "image_caption": [ + "(j) ResShift-5" + ], + "image_footnote": [], + "bbox": [ + 650, + 329, + 751, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9737b1b01f883e904c871a88409134aa30bbc0a9ce4aac60ec5961b122b9069.jpg", + "image_caption": [ + "(k) Ours-1" + ], + "image_footnote": [], + "bbox": [ + 754, + 329, + 857, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/975517ad51ff6a562932f41eaef7a27deef68d3bef710d6c4292a9af8ca0753e.jpg", + "image_caption": [ + "(b) BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 333, + 422, + 437, + 502 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/952b2c1d47070c85bed5ff263d44ad1b06437ff246e3a3c8a817c28ff04681b3.jpg", + "image_caption": [ + "(c) RealESRGAN" + ], + "image_footnote": [], + "bbox": [ + 439, + 422, + 540, + 502 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b014e659343c575da89eebe1ed38400ce0b7981401c4b7a51df6d44076772d1a.jpg", + "image_caption": [ + "(d) SwinIR" + ], + "image_footnote": [], + "bbox": [ + 542, + 422, + 647, + 502 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e3cc664c60065748777c6bd82fd4454550005c20f446226e7ef8ef4796744f4.jpg", + "image_caption": [ + "(e) DASR" + ], + "image_footnote": [], + "bbox": [ + 648, + 422, + 751, + 502 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ab80edfb1a73fc481d0741ce9492334e35b24154c4dbd6ddf36eb92ef4b73ee.jpg", + "image_caption": [ + "(f) LDM-100" + ], + "image_footnote": [], + "bbox": [ + 753, + 422, + 856, + 502 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0a103c522c7febed9f61ff5c6fbf8f726e1082e965312f70f7e26c5ec68632b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDatasets
RealSRRealSet65
CLIPIQA↑MUSIQ↑CLIPIQA↑MUSIQ↑
ESRGAN [40]0.236229.0480.373942.369
RealSR-JPEG [12]0.361536.0760.528250.539
BSRGAN [47]0.543963.5860.616365.582
SwinIR [18]0.465459.6360.578263.822
RealESRGAN [41]0.489859.6780.599563.220
DASR [19]0.362945.8250.496555.708
LDM-15 [32]0.383649.3170.427447.488
ResShift-15 [46]0.595859.8730.653761.330
SinSR-1 (distill only)0.611957.1180.682261.267
SinSR-10.688761.5820.715062.169
", + "bbox": [ + 176, + 619, + 794, + 795 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5e2776c1ddea5861af3e55275881e52173f5667897258117444c72c872647e6.jpg", + "image_caption": [ + "(k) Ours-1" + ], + "image_footnote": [], + "bbox": [ + 753, + 512, + 856, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of models on two real-world datasets. The best and second best results are highlighted in bold and underline.", + "bbox": [ + 76, + 797, + 888, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are reported in Table 1. As shown in the table, the proposed method with only one inference step can outperform the teacher model that we used by a large margin. Besides, for the latest metric CLIPIQA, the proposed method archives the best performance among all the competitors. Some visual", + "bbox": [ + 75, + 821, + 468, + 896 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comparisons are shown in Fig. 5, in which the proposed method achieves promising results using only one step.", + "bbox": [ + 500, + 821, + 890, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on synthetic datasets. We further evaluate the performance of different methods on the synthetic dataset ImageNet-Test following the setting in [46]. Specifically,", + "bbox": [ + 500, + 854, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25801", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b861ce15582ffb7c49638f8f2be09309fc67e8d4229cd2c0d1ebb81de1812eff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMetrics
PSNR↑SSIM↑LPIPS↓CLIPQA↑MUSIQ↑
ESRGAN [40]20.670.4480.4850.45143.615
RealSR-JPEG [12]23.110.5910.3260.53746.981
BSRGAN [47]24.420.6590.2590.58154.697
SwinIR [18]23.990.6670.2380.56453.790
RealESRGAN [41]24.040.6650.2540.52352.538
DASR [19]24.750.6750.2500.53648.337
LDM-30 [32]24.490.6510.2480.57250.895
LDM-15 [32]24.890.6700.2690.51246.419
ResShift-15 [46]24.900.6730.2280.60353.897
SinSR-1 (distill only)24.690.6640.2220.60753.316
SinSR-124.560.6570.2210.61153.357
", + "bbox": [ + 158, + 63, + 807, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/136372d08c2c9a99fb8086a69f2ea5edbf28fbcaf5f0802221272d73c74c5bf8.jpg", + "table_caption": [ + "Table 2. Quantitative results of models on ImageNet-Test. The best and second best results are highlighted in bold and underline." + ], + "table_footnote": [], + "table_body": "
MetricsMethods
LDM-15LDM-30LDM-100ResShift-1ResShift-5ResShift-10ResShift-15SinSR-1
LPIPS↓0.2690.2480.2440.3830.3450.2740.2280.221
CLIPQA↑0.5120.5720.6200.3400.4170.5120.6030.611
Runtime (bs=64)0.046s0.080s0.249s0.012s0.021s0.033s0.047s0.012s
Runtime (bs=1)0.408s1.192s3.902s0.058s0.218s0.425s0.633s0.058s
# Parameters (M)113.60118.59118.59
", + "bbox": [ + 98, + 268, + 870, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3000 high-resolution images are first randomly selected from the validation set of ImageNet [7]. The corresponding LR images are obtained by using the provided script in [46]. As shown in Table 2, while reducing the inference step from 15 to only 1 slightly decreases PSNR and SSIM, the proposed method achieves the best perceptual quality measured by LPIPS, a more recent full-reference image quality assessment (IQA) metric than SSIM. Besides, the proposed method also achieves the best performance among all the methods measured on the most recent SOTA metric CLIP-IQA [39], demonstrating that the proposed 1-step model is on par with or even slightly better than the teacher model with 15 inference steps in terms of perceptual qualities.", + "bbox": [ + 75, + 407, + 468, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of the efficiency. We assess the computational efficiency of the proposed method in comparison to SOTA approaches. As shown in Table 3, the proposed method demonstrates superior performance with only one inference step, outperforming ResShift [46]—the adopted teacher model, which had already significantly reduced the inference time compared to LDM [32]. It is worth noting that all methods presented in Table 3 run in latent space, and the computational cost of VQ-VAE is counted.", + "bbox": [ + 75, + 604, + 470, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analysis", + "text_level": 1, + "bbox": [ + 76, + 755, + 179, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How important is the deterministic sampling? We evaluate the performance of the model trained on generated paired samples from the proposed deterministic sampling and the default stochastic sampling strategy $(x_{T},\\hat{F}_{\\theta}(x_{T},y))$ in [46]. Due to the randomness of the generated samples $x\\sim \\hat{F}_{\\theta}(x_T,y)$ , given a random noise $\\epsilon$ , the prediction is an expectation of its conditional distribution. The comparison in Fig. 6 further verifies that the results trained w/o determin", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5eb65a227565809193c27d82fa15becb62ef633389d9f258db5027d558e58f87.jpg", + "table_caption": [ + "Table 3. Efficiency and performance comparisons with SOTA methods on ImageNet-Test. “-N” represents the number of sampling steps the model used. The running time per image is tested on a Tesla A100 GPU on the x4 (64→256) task averaged over the batch size (bs)." + ], + "table_footnote": [], + "table_body": "
MethodsCLIPIQA ↑MUSIQ ↑
w/ default sampling in [46]0.416651.53
SinSR (distill only)0.682261.27
", + "bbox": [ + 524, + 405, + 867, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5 using only distillation loss. We evaluate their performance on the RealSet65 testing set.", + "bbox": [ + 498, + 464, + 890, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5471328b91dbd7fcb239d87f3fcaad6caad89615cde5a7319ccf3e930cce4710.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 526, + 629, + 623 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b16c95a606179683fec11e539dc4aeac961c4085858eefe8bd7dd4db860fa9f.jpg", + "image_caption": [ + "(a) input" + ], + "image_footnote": [], + "bbox": [ + 503, + 625, + 629, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/180ced791f46b8649099b5d537facaaf0e5bff77f971676b52d1e92469b292f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 526, + 756, + 623 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4e33e7b2d85be2fa0879f0cdf408e67da61266aab835cb353e078b3b43e7902.jpg", + "image_caption": [ + "(b) w/ sampling in [46]" + ], + "image_footnote": [], + "bbox": [ + 630, + 625, + 756, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ffca07afc068a78db05ec7f72fb9c118706526c95c709bafa4ed46546ac51a21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 526, + 883, + 623 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0152792ab56f1b16c330dc38c0ad936127d3e60e9a8b52ce3c3291facb37eed7.jpg", + "image_caption": [ + "(c) Ours (deterministic)", + "Figure 6. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5. Best zoom in for more details." + ], + "image_footnote": [], + "bbox": [ + 759, + 625, + 883, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "istic teacher model exhibit blurred details. Besides, as shown in Table 4, there is a significant performance degradation when we replace the proposed deterministic sampling with the default one in [46], demonstrating the effectiveness and necessity of involving the proposed deterministic sampling.", + "bbox": [ + 496, + 762, + 890, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Why does a single-step distillation work? Previous studies suggest that directly learning the mapping between $x_{T}$ and $x_{0}$ is typically challenging due to the non-causal properties of the generation process [20]. However, our empirical find-", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25802", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78a5c85ee849318fbd22fd86edce0a42cf29c45a2b51f7546feb7277b268f5b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCLIPIQA ↑MUSIQ ↑
ResShift [46] (24.32M)0.536552.71
ResShift [46] (118.59M)0.653761.33
SinSR (distill only) (24.32M)0.649958.71
", + "bbox": [ + 96, + 75, + 450, + 147 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65b70a99e046855fda2bbbeed6958d7dc60bff5cbd5d71c6ba70deb07a9f6de7.jpg", + "table_caption": [ + "Table 5. A comparison of the models trained with different strategies on RealSet65. The model trained with the diffusion loss, i.e., ResShift, is more sensitive to the model size than directly learning the deterministic mapping between $x_{T}$ and $\\hat{x}_0$ , indicating that the deterministic mapping is relatively easier to learn." + ], + "table_footnote": [], + "table_body": "
LPIPS↓MUSIQ↑CLIPQA↑
ResShift [46]0.227553.900.6029
w/ Rectified Flow [21]0.232251.050.5753
SinSR (distill only)0.222153.320.6072
", + "bbox": [ + 86, + 227, + 460, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ings indicate that the matching between $x_{T}$ and $x_{0}$ in the SR task is relatively easier to learn than denoising under different noise levels, as diffusion models do. Specifically, the capacity of the student network $f_{\\hat{\\theta}}$ is sufficient to effectively capture the ODE process $F_{\\theta}$ using only one step. To verify our assumption, we evaluate the performance of smaller models trained under different strategies. Specifically, one model is trained following the experimental settings of [46] while the number of parameters decreases from 118.6M to 24.3M. Another model uses the same backbone as the aforementioned small model while directly learning the mapping relationship between $x_{T}$ and $\\hat{x}_{0}$ from the standard-size teacher diffusion model. A comparison between these two small models is reported in Table 5. As demonstrated by the results, the model trained for denoising under different noise levels suffers from a serious performance drop compared with the model that directly learns the deterministic mapping between. This strongly supports our assumption that directly learning the deterministic mapping is relatively easier.", + "bbox": [ + 73, + 369, + 470, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is a more sophisticated distillation strategy necessary? To explore the necessity of more advanced techniques that learn the mapping between $x_{T}$ and $x_{0}$ , we evaluate the performance of Rectified Flow [21], a recent method that learns the mapping to a single step through an iterative manner. Specifically, Reflow operations are conducted to avoid crossing the generation paths, and then followed by distilling the rectified generation process into a single step. However, as shown in Table 6, the involved iterative distillation degrades the performance of the final model due to the accumulated error as discussed by the author [21]. Besides, as verified by the previous section that the deterministic mapping between $x_{T}$ and $x_{0}$ is easy to learn in the SR task, the benefit of a more sophisticated distillation strategy is not obvious.", + "bbox": [ + 75, + 657, + 468, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learned inversion. As the core of the consistency preserving loss, a comparison with the DDIM inversion [37] is", + "bbox": [ + 75, + 869, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e64930e3a6c67fc81e09d1abbc4353883c264697d883d321374c627d276d59f3.jpg", + "image_caption": [ + "LR image", + "Figure 7. A comparison between HR images generated from DDIM inversion and the proposed learned inversion. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 501, + 80, + 596, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b67a75dccd38b54d6fcf65e83a03a57c4fec1d7dc55e6ba27e38ae23d079e283.jpg", + "image_caption": [ + "DDIM inversion" + ], + "image_footnote": [], + "bbox": [ + 598, + 80, + 694, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a6bc83404031f0ae2370710249703e996ab718559869786b2928fcd77cf838f.jpg", + "image_caption": [ + "Learned inversion" + ], + "image_footnote": [], + "bbox": [ + 696, + 80, + 792, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b057f75dbc7c2864c2324fab3e34405de8c005aef79fb3950a79ef0f6e39cd3.jpg", + "image_caption": [ + "GT image" + ], + "image_footnote": [], + "bbox": [ + 794, + 80, + 890, + 157 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in Fig. 7, where the proposed method achieves better fidelity performance. It indicates that the proposed method can obtain a more accurate estimation of $x_{T}$ . Besides, more analyses regarding the consistency preserving loss are in the supplementary material.", + "bbox": [ + 496, + 214, + 890, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training overhead. While the proposed method involves solving ODEs during training, benefiting from a shortened inference process and initializing the student model from the pre-trained teacher model, the training cost of finetuning using the proposed training paradigm is still lower than that of retraining the diffusion model from scratch. Specifically, the training cost is shown in Table 7.", + "bbox": [ + 496, + 289, + 892, + 396 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/45a77d975567617c56e373a8ec8eb0295eea3e2dd2b0357561836bc48ab9ca37.jpg", + "table_caption": [ + "Table 6. A comparison between models accelerated by the proposed method and [46], which includes a reflow and a distillation operation. The models are evaluated on ImageNet-Test [46]." + ], + "table_footnote": [], + "table_body": "
Num of Iterss/IterTraining Time
ResShift [46]500k1.32s~7.64 days
SinSR (Ours)30k7.41s~2.57 days
", + "bbox": [ + 524, + 404, + 867, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. A comparison of the training cost on an NVIDIA A100.", + "bbox": [ + 501, + 464, + 885, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 487, + 617, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel strategy to accelerate the diffusion-based SR models into a single inference step. Specifically, a one-step bi-directional distillation is proposed to learn the deterministic mapping between the input noise and the generated high-resolution image and versa vice from a teacher diffusion model with our derived deterministic sampling. Meanwhile, a novel consistency preserving loss is optimized at the same time during the distillation so that the student model not only uses the information from the pre-trained teacher diffusion model but also directly learns from ground-truth images. Experimental results demonstrate that the proposed method can achieve on-par or even better performance than the teacher model in only one step.", + "bbox": [ + 496, + 512, + 892, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 715, + 666, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The research is supported in part by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation), National Key R&D Program of China under Grand NO.2022ZD0160100, the National Natural Science Foundation of China under Grant No. 62102150, the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800, and the Basic and Frontier Research Project of PCL and the Major Key Project of PCL.", + "bbox": [ + 496, + 733, + 893, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25803", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 791-799, 2018. 2", + "[2] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2", + "[3] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2019. 5", + "[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2", + "[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12413-12422, 2022. 1, 2", + "[6] Ryan Dahl, Mohammad Norouzi, and Jonathon Shlens. Pixel recursive super resolution. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 7", + "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2", + "[9] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1909-1918, 2022. 2", + "[10] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, et al. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. arXiv preprint arXiv:2402.10491, 2024. 2", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3", + "[12] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In CVPR, pages 466-467, 2020. 5, 6, 7", + "[13] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1, 2", + "[15] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 1, 5", + "[16] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2", + "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, pages 1833-1844, 2021. 5, 6, 7", + "[19] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 5, 6, 7", + "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 3, 7", + "[21] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2022. 3, 5, 8", + "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 1, 2", + "[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 715-732. Springer, 2020. 2", + "[24] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 3", + "[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3", + "[26] Jacob Menick and Nal Kalchbrenner. Generating high fidelity images with subscale pixel networks and multidimensional upscaling. International Conference on Learning Representations, International Conference on Learning Representations, 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25804", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2", + "[29] Aaronvanden Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2016. 2", + "[30] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2018. 2", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 5, 6, 7", + "[33] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 1, 2, 3", + "[34] Mehdi S. M. Sajjadi, Bernhard Scholkopf, and Michael Hirsch. Enhancenet: Single image super-resolution through automated texture synthesis. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2021. 3", + "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5", + "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 8", + "[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2", + "[39] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5, 7" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 5, 6, 7", + "[41] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 5, 6, 7", + "[42] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 2", + "[43] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 2, 3", + "[44] Zhaowen Wang, Ding Liu, Jianchao Yang, Wei Han, and ThomasS. Huang. Deep networks for image super-resolution with sparse prior. Cornell University - arXiv, Cornell University - arXiv, 2015. 2", + "[45] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2", + "[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[47] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 5, 6, 7", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + ], + "bbox": [ + 501, + 92, + 893, + 667 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25805", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_model.json b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d87a3bac1d61e54f2c43b276ed845f085395f6c --- /dev/null +++ b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_model.json @@ -0,0 +1,3223 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "SinSR: Diffusion-Based Image Super-Resolution in a Single Step" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.805, + 0.216 + ], + "angle": 0, + "content": "Yufei Wang\\(^{1,2\\dagger}\\), Wenhan Yang\\(^{3}\\), Xinyuan Chen\\(^{2*}\\), Yaohui Wang\\(^{2}\\), Lanqing Guo\\(^{1}\\), Lap-Pui Chau\\(^{4}\\), Ziwei Liu\\(^{1}\\), Yu Qiao\\(^{2}\\), Alex C. Kot\\(^{1}\\), Bihan Wen\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.216, + 0.812, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Nanyang Technological University \\(^{2}\\)Shanghai Artificial Intelligence Laboratory \\(^{3}\\)PengCheng Laboratory \\(^{4}\\)The Hong Kong Polytechnic University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.474, + 0.766 + ], + "angle": 0, + "content": "While super-resolution (SR) methods based on diffusion models exhibit promising results, their practical application is hindered by the substantial number of required inference steps. Recent methods utilize the degraded images in the initial state, thereby shortening the Markov chain. Nevertheless, these solutions either rely on a precise formulation of the degradation process or still necessitate a relatively lengthy generation path (e.g., 15 iterations). To enhance inference speed, we propose a simple yet effective method for achieving single-step SR generation, named SinSR. Specifically, we first derive a deterministic sampling process from the most recent state-of-the-art (SOTA) method for accelerating diffusion-based SR. This allows the mapping between the input random noise and the generated high-resolution image to be obtained in a reduced and acceptable number of inference steps during training. We show that this deterministic mapping can be distilled into a student model that performs SR within only one inference step. Additionally, we propose a novel consistency-preserving loss to simultaneously leverage the ground-truth image during the distillation process, ensuring that the performance of the student model is not solely bound by the feature manifold of the teacher model, resulting in further performance improvement. Extensive experiments conducted on synthetic and real-world datasets demonstrate that the proposed method can achieve comparable or even superior performance compared to both previous SOTA methods and the teacher model, in just one sampling step, resulting in a remarkable up to \\(\\times 10\\) speedup for inference. Our code will be released at https://github.com/wyf0912/SinSR/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Image super-resolution (SR) aims to reconstruct a high-resolution image from a given low-resolution (LR) counterpart [45]. Recently, diffusion models, known for their effectiveness in modeling complex distributions, have gained" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.286, + 0.695, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.435, + 0.668, + 0.445 + ], + "angle": 0, + "content": "(a) Input LR image (MUSIQ↑: 23.89)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.286, + 0.791, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.355, + 0.786, + 0.364 + ], + "angle": 0, + "content": "(b) ResShift-1 (47.43)" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.286, + 0.888, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.355, + 0.882, + 0.364 + ], + "angle": 0, + "content": "(c) ResShift-5 (40.32)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.365, + 0.791, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.435, + 0.788, + 0.445 + ], + "angle": 0, + "content": "(d) ResShift-15 (62.83)" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.365, + 0.888, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.435, + 0.872, + 0.444 + ], + "angle": 0, + "content": "(e) Ours-1 (68.05)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.453, + 0.895, + 0.538 + ], + "angle": 0, + "content": "Figure 1. A comparison between the most recent SOTA method ResShift [46] for the acceleration of diffusion-based SR and the proposed method. We achieve on-par or even superior perceptual quality using only one inference step. (\"-N\" behind the method name represents the number of inference steps, and the value in the bracket is the quantitative result measured by MUSIQ↑ [15].)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.55, + 0.895, + 0.581 + ], + "angle": 0, + "content": "widespread adoption and demonstrated remarkable performance in SR tasks, particularly in terms of perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.895, + 0.764 + ], + "angle": 0, + "content": "Specifically, current strategies for employing diffusion models can be broadly categorized into two streams: concatenating the LR image to the input of the denoiser in the diffusion models [32, 33], and adjusting the inverse process of a pre-trained diffusion model [4, 5, 14]. Despite achieving promising results, both strategies encounter computational efficiency issues. Notably, the initial state of these conditional diffusion models is a pure Gaussian noise without using the prior knowledge from the LR image. Consequently, a substantial number of inference steps are required to achieve satisfactory performance, significantly hindering the practical applications of diffusion-based SR techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Efforts have been made to enhance the sampling efficiency of diffusion models, leading to various techniques proposed [22, 28, 37]. However, in the realm of low-level vision where maintaining high fidelity is critical, these techniques often fall short as they achieve acceleration at the cost of performance. More recently, innovative techniques have emerged to reformulate the diffusion process in image restoration tasks, focusing on improving the signal-to-noise ratio of the initial diffusion state and thereby shorten the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.458, + 0.901 + ], + "angle": 0, + "content": "† Work done as an intern at Shanghai AI Lab. * Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25796" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.278, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.237, + 0.219, + 0.248 + ], + "angle": 0, + "content": "(a) LR image" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.371, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.089, + 0.46, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.157, + 0.338, + 0.167 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.168, + 0.372, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.169, + 0.461, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.237, + 0.337, + 0.247 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.264, + 0.473, + 0.334 + ], + "angle": 0, + "content": "Figure 2. An illustration of the generative ability of the proposed method in only one step. Given the same LR image (Fig. (a) and (b)), by using different noise added to the input, HR images (Fig. (c)-(e)) with different details are generated, e.g., eyes of different shapes and colors. Best zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.471, + 0.504 + ], + "angle": 0, + "content": "Markov chain. For instance, [43] initiates the denoising diffusion process with the input noisy image, while in the SR task, [46] models the initial step as a combination of the LR image and random noise. Nonetheless, even in these most recent works [43, 46], limitations persist. For instance, while [43] shows promising results within just three inference steps, it requires a clear formulation of the image degradation process. Besides, [46] still necessitates 15 inference steps and exhibits degraded performance with noticeable artifacts if the number of inference steps is further reduced." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.822 + ], + "angle": 0, + "content": "To address these challenges, we introduce a novel approach that can generate high-resolution (HR) images in only one sampling step, without compromising the diversity and perceptual quality of the diffusion model, as shown in Fig. 1 and Fig. 2. Specifically, we propose to directly learn a well-paired bi-directional deterministic mapping between the input random noise and the generated HR image from a teacher diffusion model. To accelerate the generation of well-matched training data, we first derive a deterministic sampling strategy from the most recent state-of-the-art work [46], designed for accelerating diffusion-based SR, from its original stochastic formulation. Additionally, we propose a novel consistency-preserving loss to leverage ground-truth images, further enhancing the perceptual quality of the generated HR images by minimizing the error between ground-truth (GT) images and those generated from the predicted initial state. Experimental results demonstrate that our method achieves comparable or even better performance compared to SOTA methods and the teacher diffusion model [46], while greatly reducing the number of inference steps from 15 to 1, resulting in up to a \\(\\times 10\\) speedup in inference." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.824, + 0.436, + 0.838 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- We accelerate the diffusion-based SR model to a single inference step with comparable or even superior performance for the first time. Instead of shortening the Markov chain of the generation process, we propose a simple yet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.894, + 0.121 + ], + "angle": 0, + "content": "effective approach that directly distills a deterministic generation function into a student network." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.895, + 0.182 + ], + "angle": 0, + "content": "- To further fasten training, we derive a deterministic sampling strategy from the recent SOTA method [46] on accelerating the SR task, enabling efficient generation of well-matched training pairs." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.183, + 0.894, + 0.258 + ], + "angle": 0, + "content": "- We propose a novel consistency-preserving loss that can utilize the ground-truth images during training, preventing the student model from only focusing on fitting the deterministic mapping of the teacher diffusion model, therefore leading to better performance." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.258, + 0.894, + 0.333 + ], + "angle": 0, + "content": "- Extensive experiments on both synthetic and real-world datasets show that our proposed method can achieve comparable or even superior performance compared to SOTA methods and the teacher diffusion model, while greatly reducing the number of inference steps from 15 to 1." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.895, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.339, + 0.642, + 0.354 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.363, + 0.722, + 0.379 + ], + "angle": 0, + "content": "2.1. Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.896, + 0.78 + ], + "angle": 0, + "content": "With the rise of deep learning, deep learning-based techniques gradually become the mainstream of the SR task [8, 45]. One prevalent approach of early works is to train a regression model using paired training data [1, 2, 16, 44]. While the expectation of the posterior distribution can be well modeled, they inevitably suffer from the over-smooth problem [17, 27, 34]. To improve the perceptual quality of the generated HR images, generative-based SR models attract increasing attention, e.g., autoregressive-based models [6, 26, 29, 30]. While significant improvements are achieved, the computational cost of autoregressive models is usually large. Subsequently, normalizing flows [23, 42] are demonstrated to have good perceptual quality under an efficient inference process, while its network design is restricted by the requirements of the invertibility and ease of calculation. Besides, GAN-based methods also achieve great success in terms of perceptual quality [9, 13, 17, 27, 34]. However, the training of GAN-based methods is usually unstable. Recently, diffusion-based models have been widely investigated in SR [4, 5, 14, 32, 33]. The diffusion-based SR methods can be roughly summarized into two categories, concatenating the LR image to the input of the denoiser [32, 33], and modifying the backward process of a pre-trained diffusion model [4, 5, 10, 14]. While promising results are achieved, they rely on a large number of inference steps, which greatly hinders the application of diffusion-based models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.788, + 0.802 + ], + "angle": 0, + "content": "2.2. Acceleration of Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Recently, the acceleration of diffusion models has attracted more and more attention. Several algorithms are proposed for general diffusion models [22, 28, 37, 38] and proved quite effective for image generations. One intuitive strategy among them is to distill the diffusion models to a student model. However, the huge training overhead to solve" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.152, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.094, + 0.249, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.097, + 0.311, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.094, + 0.369, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.094, + 0.418, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.094, + 0.455, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.155, + 0.47, + 0.18 + ], + "angle": 0, + "content": "(a) The inference of SR3 [33] starts from a pure noise, which requires a large number of inference steps (T=100 after using DDIM [37])." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.183, + 0.466, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.258, + 0.469, + 0.294 + ], + "angle": 0, + "content": "(b) The recent SOTA method ResShift [46] shortens the Markov chain to speed up the inference process by incorporating the information of the LR image \\(y\\) to the initial state \\(x_{T}\\) \\((\\mathrm{T} = 15)\\)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.297, + 0.468, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.392, + 0.469, + 0.43 + ], + "angle": 0, + "content": "(c) A simplified pipeline of the proposed method \\( \\text{SinSR} \\) (distill only). It directly learns the deterministic mapping between \\( x_{T} \\) and \\( x_{0} \\), therefore the inference process can be further compressed into only one step \\( (\\mathrm{T} = 1) \\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.519 + ], + "angle": 0, + "content": "Figure 3. A comparison between the vanilla diffusion-based SR method [33], a most recent method for acceleration of the diffusion-based SR [46], and the proposed one-step SR. Different from recent works that shorten the Markov chain to speed up the inference process [43, 46], the proposed method directly learns the deterministic generation process and the details can be found in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.803 + ], + "angle": 0, + "content": "the ordinary differential equation (ODE) of the inference process makes this scheme less attractive on a large-scale dataset [24]. To alleviate the training overhead, progressive distillation strategies are usually adopted [25, 35]. Meanwhile, instead of simply simulating the behavior of a teacher diffusion model through distillation, better inference paths are explored in an iterative manner [20, 21]. While progressive distillation effectively decreases the training overhead, the error accumulates at the same time, leading to an obvious performance loss in SR. Most recently, targeting the image restoration task, some works reformulate the diffusion process by either using the knowledge of degradation process [43] or a pre-defined distribution of the initial state [46], yielding a shortened Markov chain of the generation process and better performance than directly applying DDIM [37] in low-level tasks. However, they either require a clear formulation of the degradation or still require a relatively large number of inference steps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.816, + 0.194, + 0.831 + ], + "angle": 0, + "content": "3. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Preliminary. Given an LR image \\(y\\) and its corresponding HR image \\(x_0\\), existing diffusion-based SR methods aim to model the conditional distribution \\(q(x_0|y)\\) through a Markov chain where a forward process is usually defined" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.364 + ], + "angle": 0, + "content": "as \\( q(x_{t}|x_{t - 1}) = \\mathcal{N}(x_{t};\\sqrt{1 - \\beta_{t}} x_{t - 1},\\beta_{t}I) \\) with an initial state \\( x_{T}\\sim \\mathcal{N}(0,I) \\). The role of the diffusion model can be regarded as transferring the input domain (standard Gaussian noise) to the HR image domain conditioned on the LR image. Since the matching relationship between \\( x_{T} \\) and \\( x_0 \\) is unknown, usually a diffusion model [11, 21, 33] through an iterative manner is required to learn/infer from an unknown mapping between \\( x_{T} \\) and \\( x_0 \\). Our method is grounded in the idea that having an SR model that effectively captures the conditional distribution \\( q(x_0|y) \\) and establishes a deterministic mapping between \\( x_{T} \\) and \\( \\hat{x} 0 \\) given an LR image \\( y \\), we can streamline the inference process to a single step by employing another network, denoted as \\( f_{\\hat{\\theta}} \\), to learn the correspondence between \\( \\hat{x}_0 \\) and \\( x_{T} \\), as illustrated in Fig. 3. Distillation for diffusion SR models: less is more. While the concept of distilling the mapping between \\( x_{T} \\) and \\( \\hat{x}_0 \\) to a student network has been previously explored [21], its application to SR introduces several challenges:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.365, + 0.894, + 0.44 + ], + "angle": 0, + "content": "- The training overhead becomes substantial for one-step distillation due to a large number of inference steps of previous models, e.g., LDM [32] still need 100 steps after using DDIM [37] for inference to generate high-quality pairs \\((\\hat{x}_0, x_T, y)\\) as the training data of the student model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.44, + 0.894, + 0.576 + ], + "angle": 0, + "content": "- The performance degradation is attributed to the introduction of a more intricate distillation strategy involving iteration. For example, to reduce the training overhead, an iterative distillation strategy [35] is adopted which gradually decreases the number of inference steps during training. However, despite achieving satisfactory results in generation tasks, the cumulative error significantly impacts the fidelity of the SR results, as SR tasks are relatively more sensitive to image quality." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.365, + 0.894, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.894, + 0.636 + ], + "angle": 0, + "content": "To address the aforementioned two challenges, we propose to distill the diffusion SR process into a single step in a simple but effective way based on the following observations. More details of the observations can be seen in Sec. 5.3" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.64, + 0.894, + 0.761 + ], + "angle": 0, + "content": "- We demonstrate that the most recent SOTA method for accelerating the diffusion-based SR [46], which achieves comparable performance in 15 steps as LDM [32] in 100 DDIM steps, has a deterministic mapping between \\( x_{T} \\) and \\( x_{0} \\). Besides, the greatly reduced number of inference steps and the existence of the deterministic mapping make the training of a single-step distillation possible as shown in Fig. 6 and Table 4." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.762, + 0.892, + 0.836 + ], + "angle": 0, + "content": "- Learning the mapping between \\( x_{T} \\) and \\( \\hat{x}_0 \\) is found to be easier than denoising \\( x_{t} \\) under different noise levels as shown in Table 5. Therefore, it is feasible to directly learn the mapping between \\( x_{T} \\) and \\( \\hat{x}_0 \\) so that the accumulated error by the iterative distillation can be avoided." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.837, + 0.894, + 0.882 + ], + "angle": 0, + "content": "- Due to the accumulated error, a more sophisticated distillation strategy (iterative-based) does not contribute to the improvement in our setting as shown in Table 6." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.64, + 0.894, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The organization of the following sections is as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.062, + 0.568, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.248, + 0.431, + 0.262 + ], + "angle": 0, + "content": "(a) One-step bi-directional distillation" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.062, + 0.887, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.247, + 0.84, + 0.262 + ], + "angle": 0, + "content": "(b) Consistency preserving loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.267, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Figure 4. The overall framework of the proposed method. By minimizing \\(\\mathcal{L}_{dist}\\) and \\(\\mathcal{L}_{inverse}\\), the student network \\(f_{\\hat{\\theta}}\\) learns the deterministic bi-directional mapping between \\(x_{T}\\) and \\(\\hat{x}_0\\) obtained from a pre-trained teacher diffusion model in one step. Meanwhile, the proposed consistency preserving loss \\(\\mathcal{L}_{gt}\\) is optimized during training to utilize the information from the GT images to pursue better perceptual quality instead of simply fitting the deterministic mappings from the teacher model. Specifically, the GT image is first converted to its latent code \\(\\hat{x}_T = f_{\\hat{\\theta}}(x_0,y,0)\\), and then converted back to calculate its reconstruction loss \\(L_{MSE}(f_{\\hat{\\theta}}(\\hat{x}_T,y,T),x_0)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.471, + 0.415 + ], + "angle": 0, + "content": "we first demonstrate that ResShift [46], in which the inference process is originally stochastic, can be converted to a deterministic model without retraining in Sec 4.1, and then the proposed consistency preserving distillation in Sec 4.2." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.431, + 0.212, + 0.449 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.456, + 0.296, + 0.473 + ], + "angle": 0, + "content": "4.1. Deterministic Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.471, + 0.542 + ], + "angle": 0, + "content": "A core difference between ResShift [46] and LDM [32] is the formulation of the initial state \\( x_{T} \\). Specifically, in ResShift [46], the information from the LR image \\( y \\) is integrated into the diffusion step \\( x_{t} \\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.555, + 0.47, + 0.573 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}, y\\right) = \\mathcal {N} \\left(x _ {t}; x _ {0} + \\eta_ {t} \\left(y - x _ {0}\\right), \\kappa^ {2} \\eta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.691 + ], + "angle": 0, + "content": "where \\(\\eta_{t}\\) is a serial of hyper-parameters that monotonically increases with the timestep \\(t\\) and obeys \\(\\eta_T\\rightarrow 1\\) and \\(\\eta_0\\to 0\\). As such, the inverse of the diffusion process starts from an initial state with rich information from the LR image \\(y\\) as follows \\(x_{T} = y + \\kappa \\sqrt{\\eta_{T}}\\epsilon\\) where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). To generate a HR image \\(x\\) from a given image \\(y\\), the original inverse process of [46] is as follows" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.704, + 0.47, + 0.733 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, y\\right) = \\mathcal {N} \\left(x _ {t - 1} \\mid \\mu_ {\\theta} \\left(x _ {t}, y, t\\right), \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.852 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}(x_t,y,t)\\) is reparameterized by a deep network. As shown in Eq. 2, given an initial state \\(x_{T} = y + \\kappa \\sqrt{\\eta_{T}}\\epsilon\\) the generated image is stochastic due to the existence of the random noise during the sampling from \\(p_{\\theta}(x_{t - 1}|x_t,y)\\). Inspired by DDIM sampling [37], we find that a non-Markovian reverse process \\(q(x_{t - 1}|x_t,x_0,y)\\) exists which keeps the marginal distribution \\(q(x_{t}|x_{0},y)\\) unchanged so" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For ease of presentation, the LR image is \\(y\\) is pre-upsampled to the same spatial resolution with the HR image \\(x\\). Besides, similar to [32, 46], the diffusion is conducted in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.89, + 0.384 + ], + "angle": 0, + "content": "that it can be directly adopted to a pre-trained model. The reformulated deterministic reverse process is as follows" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.397, + 0.892, + 0.414 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}, y\\right) = \\delta \\left(k _ {t} x _ {0} + m _ {t} x _ {t} + j _ {t} y\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.425, + 0.871, + 0.44 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the unit impulse, and \\(k_{t},m_{t},j_{t}\\) are as follows" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.451, + 0.892, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} m _ {t} = \\sqrt {\\frac {\\eta_ {t - 1}}{\\eta_ {t}}} \\\\ j _ {t} = \\eta_ {t - 1} - \\sqrt {\\eta_ {t - 1} \\eta_ {t}} \\\\ k _ {t} = 1 - \\eta_ {t - 1} + \\sqrt {\\eta_ {t - 1} \\eta_ {t}} - \\sqrt {\\frac {\\eta_ {t - 1}}{\\eta_ {t}}}. \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.572 + ], + "angle": 0, + "content": "The details of the derivation can be found in the supplementary material. As a consequence, for inference, the reverse process conditioned on \\( y \\) is reformulated as follows" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.584, + 0.817, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t - 1} = k _ {t} \\hat {x} _ {0} + m _ {t} x _ {t} + j _ {t} y \\\\ = k _ {t} f _ {\\theta} \\left(x _ {t}, y, t\\right) + m _ {t} x _ {t} + j _ {t} y, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.63, + 0.892, + 0.692 + ], + "angle": 0, + "content": "where \\(f_{\\theta}(x_t,y,t)\\) is the predicted HR image from a pretrained ResShift [46] model. By sampling from the reformulated process in Eq. 5, a deterministic mapping between \\(x_{T}\\) (or \\(\\epsilon\\)) and \\(\\hat{x}_0\\) can be obtained and is denoted as \\(F_{\\theta}(x_T,y)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.7, + 0.804, + 0.717 + ], + "angle": 0, + "content": "4.2. Consistency Preserving Distillation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.724, + 0.892, + 0.799 + ], + "angle": 0, + "content": "Vanilla distillation. We propose utilizing a student network \\( f_{\\hat{\\theta}} \\) to learn the deterministic mapping \\( F_{\\theta} \\) between the random initialized state \\( x_{T} \\) and its deterministic output \\( F_{\\theta}(x_{T},y) \\) from a teacher diffusion model. The vanilla distillation loss is defined as follows" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.812, + 0.892, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(x _ {T}, y, T\\right), F _ {\\theta} \\left(x _ {T}, y\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( f_{\\hat{\\theta}}(x_T,y,T) \\) is the student network that directly predicts the HR image in only one step, and \\( F_{\\theta} \\) represents the proposed deterministic inference process of ResShift [46] in Sec. 4.1 through an iterative manner using a pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "network parameterized by \\(\\theta\\). We observe that the student model trained solely with the distillation loss in Eq. 6 already achieves promising results in just one inference step, as indicated by “(distill only)” in the result tables." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.304 + ], + "angle": 0, + "content": "Regularization by the ground-truth image. A limitation of the aforementioned vanilla distillation strategy is that the GT image is not utilized during training, thereby restricting the upper performance bound of the student model. To further enhance the student's performance, we propose a novel strategy that incorporates a learned inversion of the HR image to provide additional regularization from the ground-truth images. In addition to the vanilla distillation loss, the student network concurrently learns the inverse mapping during training by minimizing the following loss," + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.315, + 0.471, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n v e r s e}} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(F _ {\\theta} \\left(x _ {T}, y\\right), y, 0\\right), x _ {T}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.473, + 0.417 + ], + "angle": 0, + "content": "where the last parameter of \\( f_{\\hat{\\theta}} \\) is set from \\( T \\) in Eq. 6 to 0, indicating that the model is predicting the inversion instead of the \\( \\hat{x}_0 \\). Then the GT image \\( x_0 \\) can be employed to regularize the output SR image given its predicted inversion \\( \\hat{x}_T \\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.428, + 0.35, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {T} = \\operatorname {d e t a c h} (f _ {\\hat {\\theta}} (x _ {0}, y, 0))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.445, + 0.47, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g t} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(\\hat {x} _ {T}, y, T\\right), x _ {0}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.536 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{gt}\\) is the proposed consistency preserving loss. By reusing \\(f_{\\hat{\\theta}}\\) to learn both \\(f_{\\hat{\\theta}}(\\cdot, \\cdot, T)\\) and \\(f_{\\hat{\\theta}}(\\cdot, \\cdot, 0)\\) simultaneously, we can initialize the parameter \\(\\hat{\\theta}\\) of the student model from the teacher one \\(\\theta\\) to speed up the training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.58 + ], + "angle": 0, + "content": "The overall training objective. The student network is trained to minimize the aforementioned three losses at the same time as follows" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.592, + 0.47, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\hat {\\theta} = \\arg \\min _ {\\hat {\\theta}} \\mathbb {E} _ {y, x _ {0}, x _ {T}} \\left[ \\mathcal {L} _ {\\text {d i s t i l l}} + \\mathcal {L} _ {\\text {r e v e r s e}} + \\mathcal {L} _ {g t} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.689 + ], + "angle": 0, + "content": "where the losses are defined in Eq. 6, 7, and 8 respectively. We assign equal weight to each loss term, and ablation studies are in the supplementary material. The overall of the proposed method is summarized in Algorithm 1 and Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.702, + 0.201, + 0.719 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.264, + 0.743 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training Details. For a fair comparison, we follow the same experimental setup and backbone design as that in [46]. Specifically, the main difference is that we finetuned the model for 30K iterations instead of training from scratch for 500K in [46]. We find that the student model can converge quickly so that even if for each iteration we need extra time to solve the ODE to get paired training data, the overall training time is still much shorter than retraining a model from scratch following [46]. We train the models on the training set of ImageNet [7] following the same pipeline" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.092, + 0.65, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Training" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.108, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Require: Pre-trained teacher diffusion model \\(f_{\\theta}\\) \nRequire: Paired training set \\((X,Y)\\) \n1: Init \\(f_{\\hat{\\theta}}\\) from the pre-trained model, i.e., \\(\\hat{\\theta}\\gets \\theta\\) \n2: while not converged do \n3: sample \\(x_0,y\\sim (X,Y)\\) \n4: sample \\(\\epsilon \\sim \\mathcal{N}(0,\\kappa^2\\eta_T\\mathbf{I})\\) \n5: \\(x_{T} = y + \\epsilon\\) \n6: for \\(t = T,T - 1,\\dots,1\\) do \n7: if \\(t = 1\\) then \n8: \\(\\hat{x}_0 = f_\\theta (x_1,y,1)\\) \n9: else \n10: \\(x_{t - 1} = k_t f_\\theta (x_t,y,t) + m_tx_t + j_t y\\) \n11: end if \n12: end for \n13: \\(\\mathcal{L}_{distill} = L_{MSE}(f_{\\hat{\\theta}}(x_T,y,T),\\hat{x}_0)\\) \n14: \\(\\mathcal{L}_{inverse} = L_{MSE}(f_{\\hat{\\theta}}(\\hat{x}_0,y,0),x_T)\\) \n15: \\(\\hat{x}_T = f_\\hat{\\theta} (x_0,y,0)\\) \n16: \\(\\mathcal{L}_{gt} = L_{MSE}(f_{\\hat{\\theta}}(\\operatorname {detach}(\\hat{x}_T),y,T),x_0)\\) \n17: \\(\\mathcal{L} = \\mathcal{L}_{distill} + \\mathcal{L}_{inverse} + \\mathcal{L}_{gt}\\) \n18: Perform a gradient descent step on \\(\\nabla_{\\hat{\\theta}}\\mathcal{L}\\) \n19: end while \n20: return The student model \\(f_{\\hat{\\theta}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.503 + ], + "angle": 0, + "content": "with ResShift [46] where the degradation model is adopted from RealESRGAN [41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.895, + 0.655 + ], + "angle": 0, + "content": "Compared methods. We compare our method with several representative SR models, including RealSR-JPEG [12], ESRGAN [40], BSRGAN [47], SwinIR [18], RealESRGAN [41], DASR [19], LDM [32], and ResShift [46]. For a comprehensive comparison, we further evaluate the performance of diffusion-based models LDM [32] and ResShift [46] with a reduced number of sampling steps. Besides, we compare the proposed method with Rectified-Flow [21], a SOTA method that can compress the generation process into a single step, in Table 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.894, + 0.761 + ], + "angle": 0, + "content": "Metrics. For the evaluation of the proposed method on the synthetic testing dataset with reference images, we utilize PSNR, SSIM, and LPIPS [48] to measure the fidelity performance. Besides, two recent SOTA non-reference metrics are used to justify the realism of all the images, i.e., CLIP-IQA [39] which leverages a CLIP model [31] pre-trained on a large-scale dataset (Laion400M [36]) and MUSIQ [15]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.702, + 0.787 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Evaluation on real-world datasets. RealSR [3] and RealSet65 [46] are adopted to evaluate the generalization ability of the model on unseen real-world data. Specifically, in RealSR [3], there are 100 real images captured by two different cameras in different scenarios. Besides, RealSet65 [46] includes 65 LR images in total, collected from widely used datasets and the internet. The results on these two datasets" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25800" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.054, + 0.333, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.231, + 0.254, + 0.24 + ], + "angle": 0, + "content": "(a) LR input" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.054, + 0.439, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.138, + 0.42, + 0.146 + ], + "angle": 0, + "content": "(b) BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.054, + 0.544, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.138, + 0.535, + 0.146 + ], + "angle": 0, + "content": "(c) RealESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.054, + 0.649, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.138, + 0.624, + 0.146 + ], + "angle": 0, + "content": "(d) SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.054, + 0.754, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.138, + 0.725, + 0.146 + ], + "angle": 0, + "content": "(e) DASR" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.054, + 0.858, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.138, + 0.838, + 0.146 + ], + "angle": 0, + "content": "(f) LDM-100" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.241, + 0.332, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.411, + 0.254, + 0.42 + ], + "angle": 0, + "content": "(a) LR input" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.147, + 0.439, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.231, + 0.417, + 0.24 + ], + "angle": 0, + "content": "(g) LDM-30" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.147, + 0.544, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.231, + 0.522, + 0.24 + ], + "angle": 0, + "content": "(h) LDM-15" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.147, + 0.649, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.231, + 0.633, + 0.24 + ], + "angle": 0, + "content": "(i) ResShift-15" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.147, + 0.754, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.231, + 0.734, + 0.24 + ], + "angle": 0, + "content": "(j) ResShift-5" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.147, + 0.858, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.231, + 0.831, + 0.239 + ], + "angle": 0, + "content": "(k) Ours-1" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.423, + 0.331, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.594, + 0.252, + 0.603 + ], + "angle": 0, + "content": "(a) LR input" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.241, + 0.439, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.321, + 0.42, + 0.329 + ], + "angle": 0, + "content": "(b) BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.241, + 0.543, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.321, + 0.535, + 0.329 + ], + "angle": 0, + "content": "(c) RealESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.241, + 0.649, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.321, + 0.624, + 0.329 + ], + "angle": 0, + "content": "(d) SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.241, + 0.753, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.321, + 0.726, + 0.329 + ], + "angle": 0, + "content": "(e) DASR" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.241, + 0.858, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.321, + 0.837, + 0.329 + ], + "angle": 0, + "content": "(f) LDM-100" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.605, + 0.735, + 0.617 + ], + "angle": 0, + "content": "Figure 5. Visual comparison on real-world samples. Please zoom in for more details." + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.33, + 0.439, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.411, + 0.416, + 0.42 + ], + "angle": 0, + "content": "(g) LDM-30" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.33, + 0.543, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.411, + 0.522, + 0.42 + ], + "angle": 0, + "content": "(h) LDM-15" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.33, + 0.649, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.411, + 0.633, + 0.42 + ], + "angle": 0, + "content": "(i) ResShift-15" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.33, + 0.753, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.411, + 0.734, + 0.42 + ], + "angle": 0, + "content": "(j) ResShift-5" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.33, + 0.858, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.411, + 0.832, + 0.419 + ], + "angle": 0, + "content": "(k) Ours-1" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.594, + 0.252, + 0.603 + ], + "angle": 0, + "content": "(a) LR input" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.423, + 0.439, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.504, + 0.418, + 0.512 + ], + "angle": 0, + "content": "(b) BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.423, + 0.542, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.504, + 0.532, + 0.513 + ], + "angle": 0, + "content": "(c) RealESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.423, + 0.648, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.504, + 0.623, + 0.512 + ], + "angle": 0, + "content": "(d) SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.423, + 0.752, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.504, + 0.724, + 0.512 + ], + "angle": 0, + "content": "(e) DASR" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.423, + 0.857, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.504, + 0.836, + 0.512 + ], + "angle": 0, + "content": "(f) LDM-100" + }, + { + "type": "table", + "bbox": [ + 0.178, + 0.62, + 0.795, + 0.796 + ], + "angle": 0, + "content": "
MethodsDatasets
RealSRRealSet65
CLIPIQA↑MUSIQ↑CLIPIQA↑MUSIQ↑
ESRGAN [40]0.236229.0480.373942.369
RealSR-JPEG [12]0.361536.0760.528250.539
BSRGAN [47]0.543963.5860.616365.582
SwinIR [18]0.465459.6360.578263.822
RealESRGAN [41]0.489859.6780.599563.220
DASR [19]0.362945.8250.496555.708
LDM-15 [32]0.383649.3170.427447.488
ResShift-15 [46]0.595859.8730.653761.330
SinSR-1 (distill only)0.611957.1180.682261.267
SinSR-10.688761.5820.715062.169
" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.513, + 0.857, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.594, + 0.83, + 0.602 + ], + "angle": 0, + "content": "(k) Ours-1" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.798, + 0.89, + 0.812 + ], + "angle": 0, + "content": "Table 1. Quantitative results of models on two real-world datasets. The best and second best results are highlighted in bold and underline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.47, + 0.897 + ], + "angle": 0, + "content": "are reported in Table 1. As shown in the table, the proposed method with only one inference step can outperform the teacher model that we used by a large margin. Besides, for the latest metric CLIPIQA, the proposed method archives the best performance among all the competitors. Some visual" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.822, + 0.892, + 0.852 + ], + "angle": 0, + "content": "comparisons are shown in Fig. 5, in which the proposed method achieves promising results using only one step." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Evaluation on synthetic datasets. We further evaluate the performance of different methods on the synthetic dataset ImageNet-Test following the setting in [46]. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25801" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.064, + 0.808, + 0.248 + ], + "angle": 0, + "content": "
MethodsMetrics
PSNR↑SSIM↑LPIPS↓CLIPQA↑MUSIQ↑
ESRGAN [40]20.670.4480.4850.45143.615
RealSR-JPEG [12]23.110.5910.3260.53746.981
BSRGAN [47]24.420.6590.2590.58154.697
SwinIR [18]23.990.6670.2380.56453.790
RealESRGAN [41]24.040.6650.2540.52352.538
DASR [19]24.750.6750.2500.53648.337
LDM-30 [32]24.490.6510.2480.57250.895
LDM-15 [32]24.890.6700.2690.51246.419
ResShift-15 [46]24.900.6730.2280.60353.897
SinSR-1 (distill only)24.690.6640.2220.60753.316
SinSR-124.560.6570.2210.61153.357
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.251, + 0.864, + 0.266 + ], + "angle": 0, + "content": "Table 2. Quantitative results of models on ImageNet-Test. The best and second best results are highlighted in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.269, + 0.872, + 0.361 + ], + "angle": 0, + "content": "
MetricsMethods
LDM-15LDM-30LDM-100ResShift-1ResShift-5ResShift-10ResShift-15SinSR-1
LPIPS↓0.2690.2480.2440.3830.3450.2740.2280.221
CLIPQA↑0.5120.5720.6200.3400.4170.5120.6030.611
Runtime (bs=64)0.046s0.080s0.249s0.012s0.021s0.033s0.047s0.012s
Runtime (bs=1)0.408s1.192s3.902s0.058s0.218s0.425s0.633s0.058s
# Parameters (M)113.60118.59118.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.364, + 0.89, + 0.394 + ], + "angle": 0, + "content": "Table 3. Efficiency and performance comparisons with SOTA methods on ImageNet-Test. “-N” represents the number of sampling steps the model used. The running time per image is tested on a Tesla A100 GPU on the x4 (64→256) task averaged over the batch size (bs)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.47, + 0.605 + ], + "angle": 0, + "content": "3000 high-resolution images are first randomly selected from the validation set of ImageNet [7]. The corresponding LR images are obtained by using the provided script in [46]. As shown in Table 2, while reducing the inference step from 15 to only 1 slightly decreases PSNR and SSIM, the proposed method achieves the best perceptual quality measured by LPIPS, a more recent full-reference image quality assessment (IQA) metric than SSIM. Besides, the proposed method also achieves the best performance among all the methods measured on the most recent SOTA metric CLIP-IQA [39], demonstrating that the proposed 1-step model is on par with or even slightly better than the teacher model with 15 inference steps in terms of perceptual qualities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.471, + 0.742 + ], + "angle": 0, + "content": "Evaluation of the efficiency. We assess the computational efficiency of the proposed method in comparison to SOTA approaches. As shown in Table 3, the proposed method demonstrates superior performance with only one inference step, outperforming ResShift [46]—the adopted teacher model, which had already significantly reduced the inference time compared to LDM [32]. It is worth noting that all methods presented in Table 3 run in latent space, and the computational cost of VQ-VAE is counted." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.18, + 0.772 + ], + "angle": 0, + "content": "5.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "How important is the deterministic sampling? We evaluate the performance of the model trained on generated paired samples from the proposed deterministic sampling and the default stochastic sampling strategy \\((x_{T},\\hat{F}_{\\theta}(x_{T},y))\\) in [46]. Due to the randomness of the generated samples \\(x\\sim \\hat{F}_{\\theta}(x_T,y)\\), given a random noise \\(\\epsilon\\), the prediction is an expectation of its conditional distribution. The comparison in Fig. 6 further verifies that the results trained w/o determin" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.406, + 0.868, + 0.461 + ], + "angle": 0, + "content": "
MethodsCLIPIQA ↑MUSIQ ↑
w/ default sampling in [46]0.416651.53
SinSR (distill only)0.682261.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.465, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Table 4. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5 using only distillation loss. We evaluate their performance on the RealSet65 testing set." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.527, + 0.63, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.625, + 0.63, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.695, + 0.591, + 0.706 + ], + "angle": 0, + "content": "(a) input" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.527, + 0.758, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.625, + 0.758, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.695, + 0.757, + 0.706 + ], + "angle": 0, + "content": "(b) w/ sampling in [46]" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.527, + 0.885, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.625, + 0.885, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.695, + 0.885, + 0.706 + ], + "angle": 0, + "content": "(c) Ours (deterministic)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.708, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Figure 6. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5. Best zoom in for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.838 + ], + "angle": 0, + "content": "istic teacher model exhibit blurred details. Besides, as shown in Table 4, there is a significant performance degradation when we replace the proposed deterministic sampling with the default one in [46], demonstrating the effectiveness and necessity of involving the proposed deterministic sampling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Why does a single-step distillation work? Previous studies suggest that directly learning the mapping between \\( x_{T} \\) and \\( x_{0} \\) is typically challenging due to the non-causal properties of the generation process [20]. However, our empirical find-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25802" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.077, + 0.452, + 0.148 + ], + "angle": 0, + "content": "
MethodsCLIPIQA ↑MUSIQ ↑
ResShift [46] (24.32M)0.536552.71
ResShift [46] (118.59M)0.653761.33
SinSR (distill only) (24.32M)0.649958.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.151, + 0.47, + 0.221 + ], + "angle": 0, + "content": "Table 5. A comparison of the models trained with different strategies on RealSet65. The model trained with the diffusion loss, i.e., ResShift, is more sensitive to the model size than directly learning the deterministic mapping between \\( x_{T} \\) and \\( \\hat{x}_0 \\), indicating that the deterministic mapping is relatively easier to learn." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.228, + 0.462, + 0.298 + ], + "angle": 0, + "content": "
LPIPS↓MUSIQ↑CLIPQA↑
ResShift [46]0.227553.900.6029
w/ Rectified Flow [21]0.232251.050.5753
SinSR (distill only)0.222153.320.6072
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.301, + 0.47, + 0.344 + ], + "angle": 0, + "content": "Table 6. A comparison between models accelerated by the proposed method and [46], which includes a reflow and a distillation operation. The models are evaluated on ImageNet-Test [46]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.371, + 0.471, + 0.658 + ], + "angle": 0, + "content": "ings indicate that the matching between \\( x_{T} \\) and \\( x_{0} \\) in the SR task is relatively easier to learn than denoising under different noise levels, as diffusion models do. Specifically, the capacity of the student network \\( f_{\\hat{\\theta}} \\) is sufficient to effectively capture the ODE process \\( F_{\\theta} \\) using only one step. To verify our assumption, we evaluate the performance of smaller models trained under different strategies. Specifically, one model is trained following the experimental settings of [46] while the number of parameters decreases from 118.6M to 24.3M. Another model uses the same backbone as the aforementioned small model while directly learning the mapping relationship between \\( x_{T} \\) and \\( \\hat{x}_{0} \\) from the standard-size teacher diffusion model. A comparison between these two small models is reported in Table 5. As demonstrated by the results, the model trained for denoising under different noise levels suffers from a serious performance drop compared with the model that directly learns the deterministic mapping between. This strongly supports our assumption that directly learning the deterministic mapping is relatively easier." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Is a more sophisticated distillation strategy necessary? To explore the necessity of more advanced techniques that learn the mapping between \\( x_{T} \\) and \\( x_{0} \\), we evaluate the performance of Rectified Flow [21], a recent method that learns the mapping to a single step through an iterative manner. Specifically, Reflow operations are conducted to avoid crossing the generation paths, and then followed by distilling the rectified generation process into a single step. However, as shown in Table 6, the involved iterative distillation degrades the performance of the final model due to the accumulated error as discussed by the author [21]. Besides, as verified by the previous section that the deterministic mapping between \\( x_{T} \\) and \\( x_{0} \\) is easy to learn in the SR task, the benefit of a more sophisticated distillation strategy is not obvious." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Learned inversion. As the core of the consistency preserving loss, a comparison with the DDIM inversion [37] is" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.082, + 0.598, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.159, + 0.578, + 0.171 + ], + "angle": 0, + "content": "LR image" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.082, + 0.696, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.159, + 0.69, + 0.17 + ], + "angle": 0, + "content": "DDIM inversion" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.082, + 0.794, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.159, + 0.793, + 0.17 + ], + "angle": 0, + "content": "Learned inversion" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.082, + 0.892, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.159, + 0.87, + 0.17 + ], + "angle": 0, + "content": "GT image" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.174, + 0.892, + 0.203 + ], + "angle": 0, + "content": "Figure 7. A comparison between HR images generated from DDIM inversion and the proposed learned inversion. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.29 + ], + "angle": 0, + "content": "shown in Fig. 7, where the proposed method achieves better fidelity performance. It indicates that the proposed method can obtain a more accurate estimation of \\( x_{T} \\). Besides, more analyses regarding the consistency preserving loss are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.893, + 0.397 + ], + "angle": 0, + "content": "Training overhead. While the proposed method involves solving ODEs during training, benefiting from a shortened inference process and initializing the student model from the pre-trained teacher model, the training cost of finetuning using the proposed training paradigm is still lower than that of retraining the diffusion model from scratch. Specifically, the training cost is shown in Table 7." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.405, + 0.868, + 0.462 + ], + "angle": 0, + "content": "
Num of Iterss/IterTraining Time
ResShift [46]500k1.32s~7.64 days
SinSR (Ours)30k7.41s~2.57 days
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.465, + 0.887, + 0.479 + ], + "angle": 0, + "content": "Table 7. A comparison of the training cost on an NVIDIA A100." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.488, + 0.618, + 0.503 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.893, + 0.711 + ], + "angle": 0, + "content": "In this work, we propose a novel strategy to accelerate the diffusion-based SR models into a single inference step. Specifically, a one-step bi-directional distillation is proposed to learn the deterministic mapping between the input noise and the generated high-resolution image and versa vice from a teacher diffusion model with our derived deterministic sampling. Meanwhile, a novel consistency preserving loss is optimized at the same time during the distillation so that the student model not only uses the information from the pre-trained teacher diffusion model but also directly learns from ground-truth images. Experimental results demonstrate that the proposed method can achieve on-par or even better performance than the teacher model in only one step." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.668, + 0.733 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.888 + ], + "angle": 0, + "content": "The research is supported in part by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation), National Key R&D Program of China under Grand NO.2022ZD0160100, the National Natural Science Foundation of China under Grant No. 62102150, the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800, and the Basic and Frontier Research Project of PCL and the Major Key Project of PCL." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25803" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 791-799, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[3] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12413-12422, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.472, + 0.471 + ], + "angle": 0, + "content": "[6] Ryan Dahl, Mohammad Norouzi, and Jonathon Shlens. Pixel recursive super resolution. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.472, + 0.529 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[9] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1909-1918, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.728 + ], + "angle": 0, + "content": "[10] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, et al. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. arXiv preprint arXiv:2402.10491, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.472, + 0.785 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.472, + 0.842 + ], + "angle": 0, + "content": "[12] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In CVPR, pages 466-467, 2020. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[13] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[14] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[15] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[16] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.894, + 0.358 + ], + "angle": 0, + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.893, + 0.404 + ], + "angle": 0, + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, pages 1833-1844, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.894, + 0.461 + ], + "angle": 0, + "content": "[19] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.893, + 0.574 + ], + "angle": 0, + "content": "[21] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2022. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 715-732. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.894, + 0.759 + ], + "angle": 0, + "content": "[24] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[26] Jacob Menick and Nal Kalchbrenner. Generating high fidelity images with subscale pixel networks and multidimensional upscaling. International Conference on Learning Representations, International Conference on Learning Representations, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25804" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[27] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.472, + 0.292 + ], + "angle": 0, + "content": "[29] Aaronvanden Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.472, + 0.363 + ], + "angle": 0, + "content": "[30] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.366, + 0.472, + 0.449 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.593 + ], + "angle": 0, + "content": "[33] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.597, + 0.472, + 0.652 + ], + "angle": 0, + "content": "[34] Mehdi S. M. Sajjadi, Bernhard Scholkopf, and Michael Hirsch. Enhancenet: Single image super-resolution through automated texture synthesis. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.655, + 0.472, + 0.696 + ], + "angle": 0, + "content": "[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.472, + 0.768 + ], + "angle": 0, + "content": "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.472, + 0.812 + ], + "angle": 0, + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.815, + 0.472, + 0.842 + ], + "angle": 0, + "content": "[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[39] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[40] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[41] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[42] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.894, + 0.359 + ], + "angle": 0, + "content": "[43] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.894, + 0.415 + ], + "angle": 0, + "content": "[44] Zhaowen Wang, Ding Liu, Jianchao Yang, Wei Han, and ThomasS. Huang. Deep networks for image super-resolution with sparse prior. Cornell University - arXiv, Cornell University - arXiv, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.894, + 0.471 + ], + "angle": 0, + "content": "[45] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.894, + 0.529 + ], + "angle": 0, + "content": "[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.894, + 0.599 + ], + "angle": 0, + "content": "[47] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.894, + 0.669 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25805" + } + ] +] \ No newline at end of file diff --git a/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_origin.pdf b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cfee9fc3db4997bda8e62482369b078c156f4b5e --- /dev/null +++ b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/7e640a14-4213-4e9a-b98f-0e2a0e646830_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9520b06b4fd3d788b06d69d50c2f86551986e0cfeade037a89e32c10d90d6978 +size 1287034 diff --git a/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/full.md b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c606e8f641afb19215d0bd8e1ac68658ed05eaae --- /dev/null +++ b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/full.md @@ -0,0 +1,474 @@ +# SinSR: Diffusion-Based Image Super-Resolution in a Single Step + +Yufei Wang $^{1,2\dagger}$ , Wenhan Yang $^{3}$ , Xinyuan Chen $^{2*}$ , Yaohui Wang $^{2}$ , Lanqing Guo $^{1}$ , Lap-Pui Chau $^{4}$ , Ziwei Liu $^{1}$ , Yu Qiao $^{2}$ , Alex C. Kot $^{1}$ , Bihan Wen $^{1*}$ + +$^{1}$ Nanyang Technological University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ PengCheng Laboratory $^{4}$ The Hong Kong Polytechnic University + +# Abstract + +While super-resolution (SR) methods based on diffusion models exhibit promising results, their practical application is hindered by the substantial number of required inference steps. Recent methods utilize the degraded images in the initial state, thereby shortening the Markov chain. Nevertheless, these solutions either rely on a precise formulation of the degradation process or still necessitate a relatively lengthy generation path (e.g., 15 iterations). To enhance inference speed, we propose a simple yet effective method for achieving single-step SR generation, named SinSR. Specifically, we first derive a deterministic sampling process from the most recent state-of-the-art (SOTA) method for accelerating diffusion-based SR. This allows the mapping between the input random noise and the generated high-resolution image to be obtained in a reduced and acceptable number of inference steps during training. We show that this deterministic mapping can be distilled into a student model that performs SR within only one inference step. Additionally, we propose a novel consistency-preserving loss to simultaneously leverage the ground-truth image during the distillation process, ensuring that the performance of the student model is not solely bound by the feature manifold of the teacher model, resulting in further performance improvement. Extensive experiments conducted on synthetic and real-world datasets demonstrate that the proposed method can achieve comparable or even superior performance compared to both previous SOTA methods and the teacher model, in just one sampling step, resulting in a remarkable up to $\times 10$ speedup for inference. Our code will be released at https://github.com/wyf0912/SinSR/. + +# 1. Introduction + +Image super-resolution (SR) aims to reconstruct a high-resolution image from a given low-resolution (LR) counterpart [45]. Recently, diffusion models, known for their effectiveness in modeling complex distributions, have gained + +![](images/dd8e4b209a435ae13140b3fba7835f61388989ebeea4f7afb718251b2dc8668c.jpg) +(a) Input LR image (MUSIQ↑: 23.89) + +![](images/9b5e145289f6883254fe447a846042697e225fbbdda5be8668baa6374196e38a.jpg) +(b) ResShift-1 (47.43) + +![](images/1b7b84c675c6744abd1f45ab8ce9104363d32fd706744444fea0c6305a0928b6.jpg) +(c) ResShift-5 (40.32) + +![](images/c5888dc01a559b8987a6f5965eb66b701b193ea8eb4647c4ed6555700aae4d96.jpg) +(d) ResShift-15 (62.83) +Figure 1. A comparison between the most recent SOTA method ResShift [46] for the acceleration of diffusion-based SR and the proposed method. We achieve on-par or even superior perceptual quality using only one inference step. ("-N" behind the method name represents the number of inference steps, and the value in the bracket is the quantitative result measured by MUSIQ↑ [15].) + +![](images/a5970d4aec436000b6a7d64ff488b2de2af1798f35523c25af0e440f6b382a26.jpg) +(e) Ours-1 (68.05) + +widespread adoption and demonstrated remarkable performance in SR tasks, particularly in terms of perceptual quality. + +Specifically, current strategies for employing diffusion models can be broadly categorized into two streams: concatenating the LR image to the input of the denoiser in the diffusion models [32, 33], and adjusting the inverse process of a pre-trained diffusion model [4, 5, 14]. Despite achieving promising results, both strategies encounter computational efficiency issues. Notably, the initial state of these conditional diffusion models is a pure Gaussian noise without using the prior knowledge from the LR image. Consequently, a substantial number of inference steps are required to achieve satisfactory performance, significantly hindering the practical applications of diffusion-based SR techniques. + +Efforts have been made to enhance the sampling efficiency of diffusion models, leading to various techniques proposed [22, 28, 37]. However, in the realm of low-level vision where maintaining high fidelity is critical, these techniques often fall short as they achieve acceleration at the cost of performance. More recently, innovative techniques have emerged to reformulate the diffusion process in image restoration tasks, focusing on improving the signal-to-noise ratio of the initial diffusion state and thereby shorten the + +![](images/72e3a58ba5bb88e91d2ec16f595b397702f31861550a2cfdca2af93af0da45d7.jpg) +(a) LR image + +![](images/eadaeb709f05d82e1802e314e2b39992e60a48b66b8983994dd2e74ad0b331aa.jpg) +(b) + +![](images/644fce7cbc9f8f087fbce537cc0a2a959fecaf4d8a0900dda3cf4e7d7373204c.jpg) + +![](images/b3a4837de073e3d4f7b421e613cae9b3f422881204fcd09513dfb5cd1796250f.jpg) +(d) +Figure 2. An illustration of the generative ability of the proposed method in only one step. Given the same LR image (Fig. (a) and (b)), by using different noise added to the input, HR images (Fig. (c)-(e)) with different details are generated, e.g., eyes of different shapes and colors. Best zoom in for details. + +![](images/ce0a5599d0d6117c09d4c321e7ae24700d926448323de4c6f253999f4d52413b.jpg) + +Markov chain. For instance, [43] initiates the denoising diffusion process with the input noisy image, while in the SR task, [46] models the initial step as a combination of the LR image and random noise. Nonetheless, even in these most recent works [43, 46], limitations persist. For instance, while [43] shows promising results within just three inference steps, it requires a clear formulation of the image degradation process. Besides, [46] still necessitates 15 inference steps and exhibits degraded performance with noticeable artifacts if the number of inference steps is further reduced. + +To address these challenges, we introduce a novel approach that can generate high-resolution (HR) images in only one sampling step, without compromising the diversity and perceptual quality of the diffusion model, as shown in Fig. 1 and Fig. 2. Specifically, we propose to directly learn a well-paired bi-directional deterministic mapping between the input random noise and the generated HR image from a teacher diffusion model. To accelerate the generation of well-matched training data, we first derive a deterministic sampling strategy from the most recent state-of-the-art work [46], designed for accelerating diffusion-based SR, from its original stochastic formulation. Additionally, we propose a novel consistency-preserving loss to leverage ground-truth images, further enhancing the perceptual quality of the generated HR images by minimizing the error between ground-truth (GT) images and those generated from the predicted initial state. Experimental results demonstrate that our method achieves comparable or even better performance compared to SOTA methods and the teacher diffusion model [46], while greatly reducing the number of inference steps from 15 to 1, resulting in up to a $\times 10$ speedup in inference. + +Our main contributions are summarized as follows: + +- We accelerate the diffusion-based SR model to a single inference step with comparable or even superior performance for the first time. Instead of shortening the Markov chain of the generation process, we propose a simple yet + +effective approach that directly distills a deterministic generation function into a student network. + +- To further fasten training, we derive a deterministic sampling strategy from the recent SOTA method [46] on accelerating the SR task, enabling efficient generation of well-matched training pairs. +- We propose a novel consistency-preserving loss that can utilize the ground-truth images during training, preventing the student model from only focusing on fitting the deterministic mapping of the teacher diffusion model, therefore leading to better performance. +- Extensive experiments on both synthetic and real-world datasets show that our proposed method can achieve comparable or even superior performance compared to SOTA methods and the teacher diffusion model, while greatly reducing the number of inference steps from 15 to 1. + +# 2. Related Work + +# 2.1. Image Super-Resolution + +With the rise of deep learning, deep learning-based techniques gradually become the mainstream of the SR task [8, 45]. One prevalent approach of early works is to train a regression model using paired training data [1, 2, 16, 44]. While the expectation of the posterior distribution can be well modeled, they inevitably suffer from the over-smooth problem [17, 27, 34]. To improve the perceptual quality of the generated HR images, generative-based SR models attract increasing attention, e.g., autoregressive-based models [6, 26, 29, 30]. While significant improvements are achieved, the computational cost of autoregressive models is usually large. Subsequently, normalizing flows [23, 42] are demonstrated to have good perceptual quality under an efficient inference process, while its network design is restricted by the requirements of the invertibility and ease of calculation. Besides, GAN-based methods also achieve great success in terms of perceptual quality [9, 13, 17, 27, 34]. However, the training of GAN-based methods is usually unstable. Recently, diffusion-based models have been widely investigated in SR [4, 5, 14, 32, 33]. The diffusion-based SR methods can be roughly summarized into two categories, concatenating the LR image to the input of the denoiser [32, 33], and modifying the backward process of a pre-trained diffusion model [4, 5, 10, 14]. While promising results are achieved, they rely on a large number of inference steps, which greatly hinders the application of diffusion-based models. + +# 2.2. Acceleration of Diffusion Models + +Recently, the acceleration of diffusion models has attracted more and more attention. Several algorithms are proposed for general diffusion models [22, 28, 37, 38] and proved quite effective for image generations. One intuitive strategy among them is to distill the diffusion models to a student model. However, the huge training overhead to solve + +![](images/e1345c93f2d7de5617f9168178235aa1d593c40d6f7cc5808e2225192c1c149f.jpg) + +![](images/3d61da1c6348d7e70ca0847167b48334465a92ce0ff65f1e01b22d5a64869b23.jpg) + +![](images/41ad4a4cf1e7fc33a6d8a5df4bba512b7f3050650aa25bb9f7bc08cbc221ffed.jpg) + +![](images/e86bf3426ed1ad88cb6ea836e95c372bc000a4c581b7501306afcdbc62e7b08c.jpg) + +![](images/3f7cc6c4674ecf76bf119a0c35525fef2c8bcc4f237fc8d4b4bdec3458d98883.jpg) + +![](images/c10cb5336b672ab82e8dfd1d4502eec08a3dd4a148f46593d8ead66551dbc63b.jpg) + +![](images/70a4ad8a056d213f9351a3351109791d31035f6f57b4ed5a54f5254fa91fcbc6.jpg) +(a) The inference of SR3 [33] starts from a pure noise, which requires a large number of inference steps (T=100 after using DDIM [37]). + +![](images/09d2e12d0af47b13ca8344dfebb9d6e7fc9114eec488ac78ba2270f8e6d90a95.jpg) +(b) The recent SOTA method ResShift [46] shortens the Markov chain to speed up the inference process by incorporating the information of the LR image $y$ to the initial state $x_{T}$ $(\mathrm{T} = 15)$ . +(c) A simplified pipeline of the proposed method $\text{SinSR}$ (distill only). It directly learns the deterministic mapping between $x_{T}$ and $x_{0}$ , therefore the inference process can be further compressed into only one step $(\mathrm{T} = 1)$ . +Figure 3. A comparison between the vanilla diffusion-based SR method [33], a most recent method for acceleration of the diffusion-based SR [46], and the proposed one-step SR. Different from recent works that shorten the Markov chain to speed up the inference process [43, 46], the proposed method directly learns the deterministic generation process and the details can be found in Fig. 4. + +the ordinary differential equation (ODE) of the inference process makes this scheme less attractive on a large-scale dataset [24]. To alleviate the training overhead, progressive distillation strategies are usually adopted [25, 35]. Meanwhile, instead of simply simulating the behavior of a teacher diffusion model through distillation, better inference paths are explored in an iterative manner [20, 21]. While progressive distillation effectively decreases the training overhead, the error accumulates at the same time, leading to an obvious performance loss in SR. Most recently, targeting the image restoration task, some works reformulate the diffusion process by either using the knowledge of degradation process [43] or a pre-defined distribution of the initial state [46], yielding a shortened Markov chain of the generation process and better performance than directly applying DDIM [37] in low-level tasks. However, they either require a clear formulation of the degradation or still require a relatively large number of inference steps. + +# 3. Motivation + +Preliminary. Given an LR image $y$ and its corresponding HR image $x_0$ , existing diffusion-based SR methods aim to model the conditional distribution $q(x_0|y)$ through a Markov chain where a forward process is usually defined + +as $q(x_{t}|x_{t - 1}) = \mathcal{N}(x_{t};\sqrt{1 - \beta_{t}} x_{t - 1},\beta_{t}I)$ with an initial state $x_{T}\sim \mathcal{N}(0,I)$ . The role of the diffusion model can be regarded as transferring the input domain (standard Gaussian noise) to the HR image domain conditioned on the LR image. Since the matching relationship between $x_{T}$ and $x_0$ is unknown, usually a diffusion model [11, 21, 33] through an iterative manner is required to learn/infer from an unknown mapping between $x_{T}$ and $x_0$ . Our method is grounded in the idea that having an SR model that effectively captures the conditional distribution $q(x_0|y)$ and establishes a deterministic mapping between $x_{T}$ and $\hat{x} 0$ given an LR image $y$ , we can streamline the inference process to a single step by employing another network, denoted as $f_{\hat{\theta}}$ , to learn the correspondence between $\hat{x}_0$ and $x_{T}$ , as illustrated in Fig. 3. Distillation for diffusion SR models: less is more. While the concept of distilling the mapping between $x_{T}$ and $\hat{x}_0$ to a student network has been previously explored [21], its application to SR introduces several challenges: + +- The training overhead becomes substantial for one-step distillation due to a large number of inference steps of previous models, e.g., LDM [32] still need 100 steps after using DDIM [37] for inference to generate high-quality pairs $(\hat{x}_0, x_T, y)$ as the training data of the student model. +- The performance degradation is attributed to the introduction of a more intricate distillation strategy involving iteration. For example, to reduce the training overhead, an iterative distillation strategy [35] is adopted which gradually decreases the number of inference steps during training. However, despite achieving satisfactory results in generation tasks, the cumulative error significantly impacts the fidelity of the SR results, as SR tasks are relatively more sensitive to image quality. + +To address the aforementioned two challenges, we propose to distill the diffusion SR process into a single step in a simple but effective way based on the following observations. More details of the observations can be seen in Sec. 5.3 + +- We demonstrate that the most recent SOTA method for accelerating the diffusion-based SR [46], which achieves comparable performance in 15 steps as LDM [32] in 100 DDIM steps, has a deterministic mapping between $x_{T}$ and $x_{0}$ . Besides, the greatly reduced number of inference steps and the existence of the deterministic mapping make the training of a single-step distillation possible as shown in Fig. 6 and Table 4. +- Learning the mapping between $x_{T}$ and $\hat{x}_0$ is found to be easier than denoising $x_{t}$ under different noise levels as shown in Table 5. Therefore, it is feasible to directly learn the mapping between $x_{T}$ and $\hat{x}_0$ so that the accumulated error by the iterative distillation can be avoided. +- Due to the accumulated error, a more sophisticated distillation strategy (iterative-based) does not contribute to the improvement in our setting as shown in Table 6. + +The organization of the following sections is as follows: + +![](images/b04aa19510ca2f5d685799789a32c5c959e3c54abf196e81c92745e97f94e8a9.jpg) +(a) One-step bi-directional distillation + +![](images/5e6ce58296fdae0744d7b6df8940dc397882d0084bb09b77366c99ee7facaf1c.jpg) +(b) Consistency preserving loss +Figure 4. The overall framework of the proposed method. By minimizing $\mathcal{L}_{dist}$ and $\mathcal{L}_{inverse}$ , the student network $f_{\hat{\theta}}$ learns the deterministic bi-directional mapping between $x_{T}$ and $\hat{x}_0$ obtained from a pre-trained teacher diffusion model in one step. Meanwhile, the proposed consistency preserving loss $\mathcal{L}_{gt}$ is optimized during training to utilize the information from the GT images to pursue better perceptual quality instead of simply fitting the deterministic mappings from the teacher model. Specifically, the GT image is first converted to its latent code $\hat{x}_T = f_{\hat{\theta}}(x_0,y,0)$ , and then converted back to calculate its reconstruction loss $L_{MSE}(f_{\hat{\theta}}(\hat{x}_T,y,T),x_0)$ . + +we first demonstrate that ResShift [46], in which the inference process is originally stochastic, can be converted to a deterministic model without retraining in Sec 4.1, and then the proposed consistency preserving distillation in Sec 4.2. + +# 4. Methodology + +# 4.1. Deterministic Sampling + +A core difference between ResShift [46] and LDM [32] is the formulation of the initial state $x_{T}$ . Specifically, in ResShift [46], the information from the LR image $y$ is integrated into the diffusion step $x_{t}$ as follows + +$$ +q \left(x _ {t} \mid x _ {0}, y\right) = \mathcal {N} \left(x _ {t}; x _ {0} + \eta_ {t} \left(y - x _ {0}\right), \kappa^ {2} \eta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\eta_{t}$ is a serial of hyper-parameters that monotonically increases with the timestep $t$ and obeys $\eta_T\rightarrow 1$ and $\eta_0\to 0$ . As such, the inverse of the diffusion process starts from an initial state with rich information from the LR image $y$ as follows $x_{T} = y + \kappa \sqrt{\eta_{T}}\epsilon$ where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . To generate a HR image $x$ from a given image $y$ , the original inverse process of [46] is as follows + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, y\right) = \mathcal {N} \left(x _ {t - 1} \mid \mu_ {\theta} \left(x _ {t}, y, t\right), \kappa^ {2} \frac {\eta_ {t - 1}}{\eta_ {t}} \alpha_ {t} \mathbf {I}\right), \tag {2} +$$ + +where $\mu_{\theta}(x_t,y,t)$ is reparameterized by a deep network. As shown in Eq. 2, given an initial state $x_{T} = y + \kappa \sqrt{\eta_{T}}\epsilon$ the generated image is stochastic due to the existence of the random noise during the sampling from $p_{\theta}(x_{t - 1}|x_t,y)$ . Inspired by DDIM sampling [37], we find that a non-Markovian reverse process $q(x_{t - 1}|x_t,x_0,y)$ exists which keeps the marginal distribution $q(x_{t}|x_{0},y)$ unchanged so + +that it can be directly adopted to a pre-trained model. The reformulated deterministic reverse process is as follows + +$$ +q \left(x _ {t - 1} \mid x _ {t}, x _ {0}, y\right) = \delta \left(k _ {t} x _ {0} + m _ {t} x _ {t} + j _ {t} y\right), \tag {3} +$$ + +where $\delta$ is the unit impulse, and $k_{t},m_{t},j_{t}$ are as follows + +$$ +\left\{ \begin{array}{c} m _ {t} = \sqrt {\frac {\eta_ {t - 1}}{\eta_ {t}}} \\ j _ {t} = \eta_ {t - 1} - \sqrt {\eta_ {t - 1} \eta_ {t}} \\ k _ {t} = 1 - \eta_ {t - 1} + \sqrt {\eta_ {t - 1} \eta_ {t}} - \sqrt {\frac {\eta_ {t - 1}}{\eta_ {t}}}. \end{array} \right. \tag {4} +$$ + +The details of the derivation can be found in the supplementary material. As a consequence, for inference, the reverse process conditioned on $y$ is reformulated as follows + +$$ +\begin{array}{l} x _ {t - 1} = k _ {t} \hat {x} _ {0} + m _ {t} x _ {t} + j _ {t} y \\ = k _ {t} f _ {\theta} \left(x _ {t}, y, t\right) + m _ {t} x _ {t} + j _ {t} y, \\ \end{array} +$$ + +where $f_{\theta}(x_t,y,t)$ is the predicted HR image from a pretrained ResShift [46] model. By sampling from the reformulated process in Eq. 5, a deterministic mapping between $x_{T}$ (or $\epsilon$ ) and $\hat{x}_0$ can be obtained and is denoted as $F_{\theta}(x_T,y)$ . + +# 4.2. Consistency Preserving Distillation + +Vanilla distillation. We propose utilizing a student network $f_{\hat{\theta}}$ to learn the deterministic mapping $F_{\theta}$ between the random initialized state $x_{T}$ and its deterministic output $F_{\theta}(x_{T},y)$ from a teacher diffusion model. The vanilla distillation loss is defined as follows + +$$ +\mathcal {L} _ {\text {d i s t i l l}} = L _ {M S E} \left(f _ {\hat {\theta}} \left(x _ {T}, y, T\right), F _ {\theta} \left(x _ {T}, y\right)\right), \tag {6} +$$ + +where $f_{\hat{\theta}}(x_T,y,T)$ is the student network that directly predicts the HR image in only one step, and $F_{\theta}$ represents the proposed deterministic inference process of ResShift [46] in Sec. 4.1 through an iterative manner using a pre-trained + +network parameterized by $\theta$ . We observe that the student model trained solely with the distillation loss in Eq. 6 already achieves promising results in just one inference step, as indicated by “(distill only)” in the result tables. + +Regularization by the ground-truth image. A limitation of the aforementioned vanilla distillation strategy is that the GT image is not utilized during training, thereby restricting the upper performance bound of the student model. To further enhance the student's performance, we propose a novel strategy that incorporates a learned inversion of the HR image to provide additional regularization from the ground-truth images. In addition to the vanilla distillation loss, the student network concurrently learns the inverse mapping during training by minimizing the following loss, + +$$ +\mathcal {L} _ {\text {i n v e r s e}} = L _ {M S E} \left(f _ {\hat {\theta}} \left(F _ {\theta} \left(x _ {T}, y\right), y, 0\right), x _ {T}\right), \tag {7} +$$ + +where the last parameter of $f_{\hat{\theta}}$ is set from $T$ in Eq. 6 to 0, indicating that the model is predicting the inversion instead of the $\hat{x}_0$ . Then the GT image $x_0$ can be employed to regularize the output SR image given its predicted inversion $\hat{x}_T$ as follows + +$$ +\hat {x} _ {T} = \operatorname {d e t a c h} (f _ {\hat {\theta}} (x _ {0}, y, 0)) +$$ + +$$ +\mathcal {L} _ {g t} = L _ {M S E} \left(f _ {\hat {\theta}} \left(\hat {x} _ {T}, y, T\right), x _ {0}\right), \tag {8} +$$ + +where $\mathcal{L}_{gt}$ is the proposed consistency preserving loss. By reusing $f_{\hat{\theta}}$ to learn both $f_{\hat{\theta}}(\cdot, \cdot, T)$ and $f_{\hat{\theta}}(\cdot, \cdot, 0)$ simultaneously, we can initialize the parameter $\hat{\theta}$ of the student model from the teacher one $\theta$ to speed up the training. + +The overall training objective. The student network is trained to minimize the aforementioned three losses at the same time as follows + +$$ +\hat {\theta} = \arg \min _ {\hat {\theta}} \mathbb {E} _ {y, x _ {0}, x _ {T}} \left[ \mathcal {L} _ {\text {d i s t i l l}} + \mathcal {L} _ {\text {r e v e r s e}} + \mathcal {L} _ {g t} \right], \tag {9} +$$ + +where the losses are defined in Eq. 6, 7, and 8 respectively. We assign equal weight to each loss term, and ablation studies are in the supplementary material. The overall of the proposed method is summarized in Algorithm 1 and Fig. 4. + +# 5. Experiment + +# 5.1. Experimental setup + +Training Details. For a fair comparison, we follow the same experimental setup and backbone design as that in [46]. Specifically, the main difference is that we finetuned the model for 30K iterations instead of training from scratch for 500K in [46]. We find that the student model can converge quickly so that even if for each iteration we need extra time to solve the ODE to get paired training data, the overall training time is still much shorter than retraining a model from scratch following [46]. We train the models on the training set of ImageNet [7] following the same pipeline + +Algorithm 1 Training +Require: Pre-trained teacher diffusion model $f_{\theta}$ +Require: Paired training set $(X,Y)$ +1: Init $f_{\hat{\theta}}$ from the pre-trained model, i.e., $\hat{\theta}\gets \theta$ +2: while not converged do +3: sample $x_0,y\sim (X,Y)$ +4: sample $\epsilon \sim \mathcal{N}(0,\kappa^2\eta_T\mathbf{I})$ +5: $x_{T} = y + \epsilon$ +6: for $t = T,T - 1,\dots,1$ do +7: if $t = 1$ then +8: $\hat{x}_0 = f_\theta (x_1,y,1)$ +9: else +10: $x_{t - 1} = k_t f_\theta (x_t,y,t) + m_tx_t + j_t y$ +11: end if +12: end for +13: $\mathcal{L}_{distill} = L_{MSE}(f_{\hat{\theta}}(x_T,y,T),\hat{x}_0)$ +14: $\mathcal{L}_{inverse} = L_{MSE}(f_{\hat{\theta}}(\hat{x}_0,y,0),x_T)$ +15: $\hat{x}_T = f_\hat{\theta} (x_0,y,0)$ +16: $\mathcal{L}_{gt} = L_{MSE}(f_{\hat{\theta}}(\operatorname {detach}(\hat{x}_T),y,T),x_0)$ +17: $\mathcal{L} = \mathcal{L}_{distill} + \mathcal{L}_{inverse} + \mathcal{L}_{gt}$ +18: Perform a gradient descent step on $\nabla_{\hat{\theta}}\mathcal{L}$ +19: end while +20: return The student model $f_{\hat{\theta}}$ + +with ResShift [46] where the degradation model is adopted from RealESRGAN [41]. + +Compared methods. We compare our method with several representative SR models, including RealSR-JPEG [12], ESRGAN [40], BSRGAN [47], SwinIR [18], RealESRGAN [41], DASR [19], LDM [32], and ResShift [46]. For a comprehensive comparison, we further evaluate the performance of diffusion-based models LDM [32] and ResShift [46] with a reduced number of sampling steps. Besides, we compare the proposed method with Rectified-Flow [21], a SOTA method that can compress the generation process into a single step, in Table 6. + +Metrics. For the evaluation of the proposed method on the synthetic testing dataset with reference images, we utilize PSNR, SSIM, and LPIPS [48] to measure the fidelity performance. Besides, two recent SOTA non-reference metrics are used to justify the realism of all the images, i.e., CLIP-IQA [39] which leverages a CLIP model [31] pre-trained on a large-scale dataset (Laion400M [36]) and MUSIQ [15]. + +# 5.2. Experimental Results + +Evaluation on real-world datasets. RealSR [3] and RealSet65 [46] are adopted to evaluate the generalization ability of the model on unseen real-world data. Specifically, in RealSR [3], there are 100 real images captured by two different cameras in different scenarios. Besides, RealSet65 [46] includes 65 LR images in total, collected from widely used datasets and the internet. The results on these two datasets + +![](images/c5ed0296b37d4ce3e6598b5894c72043018d6208e02b4d27d55637544ded7f6a.jpg) +(a) LR input + +![](images/580e65fc391fcad3e31a61acd43bde698b425c15d37943e46d23f6375f59f260.jpg) +(b) BSRGAN + +![](images/f8753f7fcd123ee25997f728b3f8627b1d03615981b4d8f830c0206838dbd7b1.jpg) +(c) RealESRGAN + +![](images/beec3d4f1e56ce25a7b87889c5d300d25a04514e5cf04cf698802b587256bb58.jpg) +(d) SwinIR + +![](images/d3a65906fc0e119f0257d2ffe72d069e5f60498c3671f48a18545266985335b9.jpg) +(e) DASR + +![](images/8b6e70bd8733be0840f34790dd9583292f655bbe86df25415d64ca7774606a14.jpg) +(f) LDM-100 + +![](images/17a59811229d322c1aa2787f6eb9355ca004ed286dc3783c0dad5e7ee1a198dc.jpg) +(a) LR input + +![](images/540c93fec6a816908c0a12d32deee2e754591e16af15f64913f4b892bf37237f.jpg) +(g) LDM-30 + +![](images/fff71557f5e5e0d733cf2c268436554a1e973f142c94d3862d8a1e78545e8c5a.jpg) +(h) LDM-15 + +![](images/72eab5a26d1f69e3ec0b3f7228dda200db2344027a9da564ac45c8ae5374d0e6.jpg) +(i) ResShift-15 + +![](images/0529e6bbc423f940ba942b949cdffea65308068e762ebf397ce3d772735d2e97.jpg) +(j) ResShift-5 + +![](images/1cee4c38d8e016b2baaec3d4fca6aac4ecb4040ead2ee2babc2f00750905a1d4.jpg) +(k) Ours-1 + +![](images/2c9ef5256714e761b8a97dc7c330b509a73468bf04b67d86bf956a888f7526a6.jpg) +Figure 5. Visual comparison on real-world samples. Please zoom in for more details. +(a) LR input + +(a) LR input + +![](images/2660b2aa9542303ccafed587fe8deae32408c0e886d2a283326414134f44e89a.jpg) +(b) BSRGAN + +![](images/a7972bb939a3575d336ca78b04f981bcab8c0d37823bb29c2ddb65f4d1501f7e.jpg) +(c) RealESRGAN + +![](images/3ac65be59d92ca1ddafee366c833329d32b3638e3b1b66aa6edfd1464b8af59e.jpg) +(d) SwinIR + +![](images/dd18df93c60f2d6dd94edbfa37a81a876cb0a8ae93b8924ca1dcb278204f2bcd.jpg) +(e) DASR + +![](images/b3e29b83d86dd71ee339850569df2cb7f617e6f836367af9c8587fdbdfc5e5a8.jpg) +(f) LDM-100 + +![](images/ea6f0f0495c04ad25223c915695514d36c31ddee6f76c98a7242b4b379835670.jpg) +(g) LDM-30 + +![](images/b70570e1df60f4d0b17e93f29f76f2c39575f71fe30053d13a804de910bb39e8.jpg) +(h) LDM-15 + +![](images/5bc4aacfd3c67882681afb57010c8ead1bb12961bc883beac500c61fd27c3604.jpg) +(i) ResShift-15 + +![](images/84f519c79d852450507766d1fe7da908fd63073681cd1c8825d4c069e6cd8ad6.jpg) +(j) ResShift-5 + +![](images/c9737b1b01f883e904c871a88409134aa30bbc0a9ce4aac60ec5961b122b9069.jpg) +(k) Ours-1 + +![](images/975517ad51ff6a562932f41eaef7a27deef68d3bef710d6c4292a9af8ca0753e.jpg) +(b) BSRGAN + +![](images/952b2c1d47070c85bed5ff263d44ad1b06437ff246e3a3c8a817c28ff04681b3.jpg) +(c) RealESRGAN + +![](images/b014e659343c575da89eebe1ed38400ce0b7981401c4b7a51df6d44076772d1a.jpg) +(d) SwinIR + +![](images/3e3cc664c60065748777c6bd82fd4454550005c20f446226e7ef8ef4796744f4.jpg) +(e) DASR + +![](images/1ab80edfb1a73fc481d0741ce9492334e35b24154c4dbd6ddf36eb92ef4b73ee.jpg) +(f) LDM-100 + +
MethodsDatasets
RealSRRealSet65
CLIPIQA↑MUSIQ↑CLIPIQA↑MUSIQ↑
ESRGAN [40]0.236229.0480.373942.369
RealSR-JPEG [12]0.361536.0760.528250.539
BSRGAN [47]0.543963.5860.616365.582
SwinIR [18]0.465459.6360.578263.822
RealESRGAN [41]0.489859.6780.599563.220
DASR [19]0.362945.8250.496555.708
LDM-15 [32]0.383649.3170.427447.488
ResShift-15 [46]0.595859.8730.653761.330
SinSR-1 (distill only)0.611957.1180.682261.267
SinSR-10.688761.5820.715062.169
+ +![](images/b5e2776c1ddea5861af3e55275881e52173f5667897258117444c72c872647e6.jpg) +(k) Ours-1 + +Table 1. Quantitative results of models on two real-world datasets. The best and second best results are highlighted in bold and underline. + +are reported in Table 1. As shown in the table, the proposed method with only one inference step can outperform the teacher model that we used by a large margin. Besides, for the latest metric CLIPIQA, the proposed method archives the best performance among all the competitors. Some visual + +comparisons are shown in Fig. 5, in which the proposed method achieves promising results using only one step. + +Evaluation on synthetic datasets. We further evaluate the performance of different methods on the synthetic dataset ImageNet-Test following the setting in [46]. Specifically, + +
MethodsMetrics
PSNR↑SSIM↑LPIPS↓CLIPQA↑MUSIQ↑
ESRGAN [40]20.670.4480.4850.45143.615
RealSR-JPEG [12]23.110.5910.3260.53746.981
BSRGAN [47]24.420.6590.2590.58154.697
SwinIR [18]23.990.6670.2380.56453.790
RealESRGAN [41]24.040.6650.2540.52352.538
DASR [19]24.750.6750.2500.53648.337
LDM-30 [32]24.490.6510.2480.57250.895
LDM-15 [32]24.890.6700.2690.51246.419
ResShift-15 [46]24.900.6730.2280.60353.897
SinSR-1 (distill only)24.690.6640.2220.60753.316
SinSR-124.560.6570.2210.61153.357
+ +Table 2. Quantitative results of models on ImageNet-Test. The best and second best results are highlighted in bold and underline. + +
MetricsMethods
LDM-15LDM-30LDM-100ResShift-1ResShift-5ResShift-10ResShift-15SinSR-1
LPIPS↓0.2690.2480.2440.3830.3450.2740.2280.221
CLIPQA↑0.5120.5720.6200.3400.4170.5120.6030.611
Runtime (bs=64)0.046s0.080s0.249s0.012s0.021s0.033s0.047s0.012s
Runtime (bs=1)0.408s1.192s3.902s0.058s0.218s0.425s0.633s0.058s
# Parameters (M)113.60118.59118.59
+ +3000 high-resolution images are first randomly selected from the validation set of ImageNet [7]. The corresponding LR images are obtained by using the provided script in [46]. As shown in Table 2, while reducing the inference step from 15 to only 1 slightly decreases PSNR and SSIM, the proposed method achieves the best perceptual quality measured by LPIPS, a more recent full-reference image quality assessment (IQA) metric than SSIM. Besides, the proposed method also achieves the best performance among all the methods measured on the most recent SOTA metric CLIP-IQA [39], demonstrating that the proposed 1-step model is on par with or even slightly better than the teacher model with 15 inference steps in terms of perceptual qualities. + +Evaluation of the efficiency. We assess the computational efficiency of the proposed method in comparison to SOTA approaches. As shown in Table 3, the proposed method demonstrates superior performance with only one inference step, outperforming ResShift [46]—the adopted teacher model, which had already significantly reduced the inference time compared to LDM [32]. It is worth noting that all methods presented in Table 3 run in latent space, and the computational cost of VQ-VAE is counted. + +# 5.3. Analysis + +How important is the deterministic sampling? We evaluate the performance of the model trained on generated paired samples from the proposed deterministic sampling and the default stochastic sampling strategy $(x_{T},\hat{F}_{\theta}(x_{T},y))$ in [46]. Due to the randomness of the generated samples $x\sim \hat{F}_{\theta}(x_T,y)$ , given a random noise $\epsilon$ , the prediction is an expectation of its conditional distribution. The comparison in Fig. 6 further verifies that the results trained w/o determin + +Table 3. Efficiency and performance comparisons with SOTA methods on ImageNet-Test. “-N” represents the number of sampling steps the model used. The running time per image is tested on a Tesla A100 GPU on the x4 (64→256) task averaged over the batch size (bs). + +
MethodsCLIPIQA ↑MUSIQ ↑
w/ default sampling in [46]0.416651.53
SinSR (distill only)0.682261.27
+ +Table 4. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5 using only distillation loss. We evaluate their performance on the RealSet65 testing set. + +![](images/5471328b91dbd7fcb239d87f3fcaad6caad89615cde5a7319ccf3e930cce4710.jpg) + +![](images/6b16c95a606179683fec11e539dc4aeac961c4085858eefe8bd7dd4db860fa9f.jpg) +(a) input + +![](images/180ced791f46b8649099b5d537facaaf0e5bff77f971676b52d1e92469b292f7.jpg) + +![](images/c4e33e7b2d85be2fa0879f0cdf408e67da61266aab835cb353e078b3b43e7902.jpg) +(b) w/ sampling in [46] + +![](images/ffca07afc068a78db05ec7f72fb9c118706526c95c709bafa4ed46546ac51a21.jpg) + +![](images/0152792ab56f1b16c330dc38c0ad936127d3e60e9a8b52ce3c3291facb37eed7.jpg) +(c) Ours (deterministic) +Figure 6. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5. Best zoom in for more details. + +istic teacher model exhibit blurred details. Besides, as shown in Table 4, there is a significant performance degradation when we replace the proposed deterministic sampling with the default one in [46], demonstrating the effectiveness and necessity of involving the proposed deterministic sampling. + +Why does a single-step distillation work? Previous studies suggest that directly learning the mapping between $x_{T}$ and $x_{0}$ is typically challenging due to the non-causal properties of the generation process [20]. However, our empirical find- + +
MethodsCLIPIQA ↑MUSIQ ↑
ResShift [46] (24.32M)0.536552.71
ResShift [46] (118.59M)0.653761.33
SinSR (distill only) (24.32M)0.649958.71
+ +Table 5. A comparison of the models trained with different strategies on RealSet65. The model trained with the diffusion loss, i.e., ResShift, is more sensitive to the model size than directly learning the deterministic mapping between $x_{T}$ and $\hat{x}_0$ , indicating that the deterministic mapping is relatively easier to learn. + +
LPIPS↓MUSIQ↑CLIPQA↑
ResShift [46]0.227553.900.6029
w/ Rectified Flow [21]0.232251.050.5753
SinSR (distill only)0.222153.320.6072
+ +ings indicate that the matching between $x_{T}$ and $x_{0}$ in the SR task is relatively easier to learn than denoising under different noise levels, as diffusion models do. Specifically, the capacity of the student network $f_{\hat{\theta}}$ is sufficient to effectively capture the ODE process $F_{\theta}$ using only one step. To verify our assumption, we evaluate the performance of smaller models trained under different strategies. Specifically, one model is trained following the experimental settings of [46] while the number of parameters decreases from 118.6M to 24.3M. Another model uses the same backbone as the aforementioned small model while directly learning the mapping relationship between $x_{T}$ and $\hat{x}_{0}$ from the standard-size teacher diffusion model. A comparison between these two small models is reported in Table 5. As demonstrated by the results, the model trained for denoising under different noise levels suffers from a serious performance drop compared with the model that directly learns the deterministic mapping between. This strongly supports our assumption that directly learning the deterministic mapping is relatively easier. + +Is a more sophisticated distillation strategy necessary? To explore the necessity of more advanced techniques that learn the mapping between $x_{T}$ and $x_{0}$ , we evaluate the performance of Rectified Flow [21], a recent method that learns the mapping to a single step through an iterative manner. Specifically, Reflow operations are conducted to avoid crossing the generation paths, and then followed by distilling the rectified generation process into a single step. However, as shown in Table 6, the involved iterative distillation degrades the performance of the final model due to the accumulated error as discussed by the author [21]. Besides, as verified by the previous section that the deterministic mapping between $x_{T}$ and $x_{0}$ is easy to learn in the SR task, the benefit of a more sophisticated distillation strategy is not obvious. + +Learned inversion. As the core of the consistency preserving loss, a comparison with the DDIM inversion [37] is + +![](images/e64930e3a6c67fc81e09d1abbc4353883c264697d883d321374c627d276d59f3.jpg) +LR image +Figure 7. A comparison between HR images generated from DDIM inversion and the proposed learned inversion. Zoom in for details. + +![](images/b67a75dccd38b54d6fcf65e83a03a57c4fec1d7dc55e6ba27e38ae23d079e283.jpg) +DDIM inversion + +![](images/2a6bc83404031f0ae2370710249703e996ab718559869786b2928fcd77cf838f.jpg) +Learned inversion + +![](images/4b057f75dbc7c2864c2324fab3e34405de8c005aef79fb3950a79ef0f6e39cd3.jpg) +GT image + +shown in Fig. 7, where the proposed method achieves better fidelity performance. It indicates that the proposed method can obtain a more accurate estimation of $x_{T}$ . Besides, more analyses regarding the consistency preserving loss are in the supplementary material. + +Training overhead. While the proposed method involves solving ODEs during training, benefiting from a shortened inference process and initializing the student model from the pre-trained teacher model, the training cost of finetuning using the proposed training paradigm is still lower than that of retraining the diffusion model from scratch. Specifically, the training cost is shown in Table 7. + +Table 6. A comparison between models accelerated by the proposed method and [46], which includes a reflow and a distillation operation. The models are evaluated on ImageNet-Test [46]. + +
Num of Iterss/IterTraining Time
ResShift [46]500k1.32s~7.64 days
SinSR (Ours)30k7.41s~2.57 days
+ +Table 7. A comparison of the training cost on an NVIDIA A100. + +# 6. Conclusion + +In this work, we propose a novel strategy to accelerate the diffusion-based SR models into a single inference step. Specifically, a one-step bi-directional distillation is proposed to learn the deterministic mapping between the input noise and the generated high-resolution image and versa vice from a teacher diffusion model with our derived deterministic sampling. Meanwhile, a novel consistency preserving loss is optimized at the same time during the distillation so that the student model not only uses the information from the pre-trained teacher diffusion model but also directly learns from ground-truth images. Experimental results demonstrate that the proposed method can achieve on-par or even better performance than the teacher model in only one step. + +# Acknowledgements + +The research is supported in part by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation), National Key R&D Program of China under Grand NO.2022ZD0160100, the National Natural Science Foundation of China under Grant No. 62102150, the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800, and the Basic and Frontier Research Project of PCL and the Major Key Project of PCL. + +# References + +[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 791-799, 2018. 2 +[2] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2 +[3] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2019. 5 +[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2 +[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12413-12422, 2022. 1, 2 +[6] Ryan Dahl, Mohammad Norouzi, and Jonathon Shlens. Pixel recursive super resolution. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 7 +[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2 +[9] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1909-1918, 2022. 2 +[10] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, et al. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. arXiv preprint arXiv:2402.10491, 2024. 2 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3 +[12] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In CVPR, pages 466-467, 2020. 5, 6, 7 +[13] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2 + +[14] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1, 2 +[15] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 1, 5 +[16] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2 +[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, pages 1833-1844, 2021. 5, 6, 7 +[19] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 5, 6, 7 +[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 3, 7 +[21] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2022. 3, 5, 8 +[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 1, 2 +[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 715-732. Springer, 2020. 2 +[24] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 3 +[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3 +[26] Jacob Menick and Nal Kalchbrenner. Generating high fidelity images with subscale pixel networks and multidimensional upscaling. International Conference on Learning Representations, International Conference on Learning Representations, 2018. 2 + +[27] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2 +[29] Aaronvanden Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2016. 2 +[30] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2018. 2 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 5, 6, 7 +[33] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 1, 2, 3 +[34] Mehdi S. M. Sajjadi, Bernhard Scholkopf, and Michael Hirsch. Enhancenet: Single image super-resolution through automated texture synthesis. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2021. 3 +[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5 +[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 8 +[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2 +[39] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5, 7 + +[40] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 5, 6, 7 +[41] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 5, 6, 7 +[42] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 2 +[43] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 2, 3 +[44] Zhaowen Wang, Ding Liu, Jianchao Yang, Wei Han, and ThomasS. Huang. Deep networks for image super-resolution with sparse prior. Cornell University - arXiv, Cornell University - arXiv, 2015. 2 +[45] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2 +[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[47] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 5, 6, 7 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 \ No newline at end of file diff --git a/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/images.zip b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..07a1731b70152b2cf98cf9339eaedc4412fe2cbf --- /dev/null +++ b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9574109b17aa76eb276fba4e8437dd10393fa8b9948e46600b56fbba9afedfc +size 790420 diff --git a/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/layout.json b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..16e40ad17364b78b6e83e6259002ca7d7e7ad6ad --- /dev/null +++ b/2024/SinSR_ Diffusion-Based Image Super-Resolution in a Single Step/layout.json @@ -0,0 +1,13128 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "text", + "content": "SinSR: Diffusion-Based Image Super-Resolution in a Single Step" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": "Yufei Wang" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Wenhan Yang" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Xinyuan Chen" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Yaohui Wang" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Lanqing Guo" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Lap-Pui Chau" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Ziwei Liu" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Alex C. Kot" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Bihan Wen" + }, + { + "bbox": [ + 97, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "spans": [ + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "text", + "content": "PengCheng Laboratory " + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 171, + 496, + 200 + ], + "type": "text", + "content": "The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 247, + 290, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 290, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 290, + 606 + ], + "type": "text", + "content": "While super-resolution (SR) methods based on diffusion models exhibit promising results, their practical application is hindered by the substantial number of required inference steps. Recent methods utilize the degraded images in the initial state, thereby shortening the Markov chain. Nevertheless, these solutions either rely on a precise formulation of the degradation process or still necessitate a relatively lengthy generation path (e.g., 15 iterations). To enhance inference speed, we propose a simple yet effective method for achieving single-step SR generation, named SinSR. Specifically, we first derive a deterministic sampling process from the most recent state-of-the-art (SOTA) method for accelerating diffusion-based SR. This allows the mapping between the input random noise and the generated high-resolution image to be obtained in a reduced and acceptable number of inference steps during training. We show that this deterministic mapping can be distilled into a student model that performs SR within only one inference step. Additionally, we propose a novel consistency-preserving loss to simultaneously leverage the ground-truth image during the distillation process, ensuring that the performance of the student model is not solely bound by the feature manifold of the teacher model, resulting in further performance improvement. Extensive experiments conducted on synthetic and real-world datasets demonstrate that the proposed method can achieve comparable or even superior performance compared to both previous SOTA methods and the teacher model, in just one sampling step, resulting in a remarkable up to " + }, + { + "bbox": [ + 46, + 247, + 290, + 606 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 46, + 247, + 290, + 606 + ], + "type": "text", + "content": " speedup for inference. Our code will be released at https://github.com/wyf0912/SinSR/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "type": "text", + "content": "Image super-resolution (SR) aims to reconstruct a high-resolution image from a given low-resolution (LR) counterpart [45]. Recently, diffusion models, known for their effectiveness in modeling complex distributions, have gained" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 226, + 425, + 343 + ], + "blocks": [ + { + "bbox": [ + 308, + 226, + 425, + 343 + ], + "lines": [ + { + "bbox": [ + 308, + 226, + 425, + 343 + ], + "spans": [ + { + "bbox": [ + 308, + 226, + 425, + 343 + ], + "type": "image", + "image_path": "dd8e4b209a435ae13140b3fba7835f61388989ebeea4f7afb718251b2dc8668c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 344, + 408, + 352 + ], + "lines": [ + { + "bbox": [ + 326, + 344, + 408, + 352 + ], + "spans": [ + { + "bbox": [ + 326, + 344, + 408, + 352 + ], + "type": "text", + "content": "(a) Input LR image (MUSIQ↑: 23.89)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 226, + 484, + 280 + ], + "blocks": [ + { + "bbox": [ + 426, + 226, + 484, + 280 + ], + "lines": [ + { + "bbox": [ + 426, + 226, + 484, + 280 + ], + "spans": [ + { + "bbox": [ + 426, + 226, + 484, + 280 + ], + "type": "image", + "image_path": "9b5e145289f6883254fe447a846042697e225fbbdda5be8668baa6374196e38a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 281, + 481, + 288 + ], + "lines": [ + { + "bbox": [ + 429, + 281, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 429, + 281, + 481, + 288 + ], + "type": "text", + "content": "(b) ResShift-1 (47.43)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 485, + 226, + 543, + 280 + ], + "blocks": [ + { + "bbox": [ + 485, + 226, + 543, + 280 + ], + "lines": [ + { + "bbox": [ + 485, + 226, + 543, + 280 + ], + "spans": [ + { + "bbox": [ + 485, + 226, + 543, + 280 + ], + "type": "image", + "image_path": "1b7b84c675c6744abd1f45ab8ce9104363d32fd706744444fea0c6305a0928b6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 281, + 539, + 288 + ], + "lines": [ + { + "bbox": [ + 488, + 281, + 539, + 288 + ], + "spans": [ + { + "bbox": [ + 488, + 281, + 539, + 288 + ], + "type": "text", + "content": "(c) ResShift-5 (40.32)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 426, + 289, + 484, + 343 + ], + "blocks": [ + { + "bbox": [ + 426, + 289, + 484, + 343 + ], + "lines": [ + { + "bbox": [ + 426, + 289, + 484, + 343 + ], + "spans": [ + { + "bbox": [ + 426, + 289, + 484, + 343 + ], + "type": "image", + "image_path": "c5888dc01a559b8987a6f5965eb66b701b193ea8eb4647c4ed6555700aae4d96.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 344, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 427, + 344, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 427, + 344, + 482, + 352 + ], + "type": "text", + "content": "(d) ResShift-15 (62.83)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 358, + 547, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 358, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 547, + 426 + ], + "type": "text", + "content": "Figure 1. A comparison between the most recent SOTA method ResShift [46] for the acceleration of diffusion-based SR and the proposed method. We achieve on-par or even superior perceptual quality using only one inference step. (\"-N\" behind the method name represents the number of inference steps, and the value in the bracket is the quantitative result measured by MUSIQ↑ [15].)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 485, + 289, + 543, + 343 + ], + "blocks": [ + { + "bbox": [ + 485, + 289, + 543, + 343 + ], + "lines": [ + { + "bbox": [ + 485, + 289, + 543, + 343 + ], + "spans": [ + { + "bbox": [ + 485, + 289, + 543, + 343 + ], + "type": "image", + "image_path": "a5970d4aec436000b6a7d64ff488b2de2af1798f35523c25af0e440f6b382a26.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 344, + 533, + 351 + ], + "lines": [ + { + "bbox": [ + 492, + 344, + 533, + 351 + ], + "spans": [ + { + "bbox": [ + 492, + 344, + 533, + 351 + ], + "type": "text", + "content": "(e) Ours-1 (68.05)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 435, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 547, + 460 + ], + "type": "text", + "content": "widespread adoption and demonstrated remarkable performance in SR tasks, particularly in terms of perceptual quality." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 460, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 605 + ], + "type": "text", + "content": "Specifically, current strategies for employing diffusion models can be broadly categorized into two streams: concatenating the LR image to the input of the denoiser in the diffusion models [32, 33], and adjusting the inverse process of a pre-trained diffusion model [4, 5, 14]. Despite achieving promising results, both strategies encounter computational efficiency issues. Notably, the initial state of these conditional diffusion models is a pure Gaussian noise without using the prior knowledge from the LR image. Consequently, a substantial number of inference steps are required to achieve satisfactory performance, significantly hindering the practical applications of diffusion-based SR techniques." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 713 + ], + "type": "text", + "content": "Efforts have been made to enhance the sampling efficiency of diffusion models, leading to various techniques proposed [22, 28, 37]. However, in the realm of low-level vision where maintaining high fidelity is critical, these techniques often fall short as they achieve acceleration at the cost of performance. More recently, innovative techniques have emerged to reformulate the diffusion process in image restoration tasks, focusing on improving the signal-to-noise ratio of the initial diffusion state and thereby shorten the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 280, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 280, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 280, + 713 + ], + "type": "text", + "content": "† Work done as an intern at Shanghai AI Lab. * Corresponding authors." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25796" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 170, + 186 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 170, + 186 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 170, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 170, + 186 + ], + "type": "image", + "image_path": "72e3a58ba5bb88e91d2ec16f595b397702f31861550a2cfdca2af93af0da45d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 187, + 134, + 196 + ], + "lines": [ + { + "bbox": [ + 93, + 187, + 134, + 196 + ], + "spans": [ + { + "bbox": [ + 93, + 187, + 134, + 196 + ], + "type": "text", + "content": "(a) LR image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 227, + 123 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 227, + 123 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 227, + 123 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 227, + 123 + ], + "type": "image", + "image_path": "eadaeb709f05d82e1802e314e2b39992e60a48b66b8983994dd2e74ad0b331aa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 124, + 206, + 132 + ], + "lines": [ + { + "bbox": [ + 196, + 124, + 206, + 132 + ], + "spans": [ + { + "bbox": [ + 196, + 124, + 206, + 132 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 281, + 123 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 281, + 123 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 281, + 123 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 281, + 123 + ], + "type": "image", + "image_path": "644fce7cbc9f8f087fbce537cc0a2a959fecaf4d8a0900dda3cf4e7d7373204c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 174, + 133, + 227, + 186 + ], + "blocks": [ + { + "bbox": [ + 174, + 133, + 227, + 186 + ], + "lines": [ + { + "bbox": [ + 174, + 133, + 227, + 186 + ], + "spans": [ + { + "bbox": [ + 174, + 133, + 227, + 186 + ], + "type": "image", + "image_path": "b3a4837de073e3d4f7b421e613cae9b3f422881204fcd09513dfb5cd1796250f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 187, + 206, + 195 + ], + "lines": [ + { + "bbox": [ + 196, + 187, + 206, + 195 + ], + "spans": [ + { + "bbox": [ + 196, + 187, + 206, + 195 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 209, + 289, + 264 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 289, + 264 + ], + "type": "text", + "content": "Figure 2. An illustration of the generative ability of the proposed method in only one step. Given the same LR image (Fig. (a) and (b)), by using different noise added to the input, HR images (Fig. (c)-(e)) with different details are generated, e.g., eyes of different shapes and colors. Best zoom in for details." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 229, + 133, + 282, + 186 + ], + "blocks": [ + { + "bbox": [ + 229, + 133, + 282, + 186 + ], + "lines": [ + { + "bbox": [ + 229, + 133, + 282, + 186 + ], + "spans": [ + { + "bbox": [ + 229, + 133, + 282, + 186 + ], + "type": "image", + "image_path": "ce0a5599d0d6117c09d4c321e7ae24700d926448323de4c6f253999f4d52413b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 279, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 288, + 399 + ], + "type": "text", + "content": "Markov chain. For instance, [43] initiates the denoising diffusion process with the input noisy image, while in the SR task, [46] models the initial step as a combination of the LR image and random noise. Nonetheless, even in these most recent works [43, 46], limitations persist. For instance, while [43] shows promising results within just three inference steps, it requires a clear formulation of the image degradation process. Besides, [46] still necessitates 15 inference steps and exhibits degraded performance with noticeable artifacts if the number of inference steps is further reduced." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 400, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 651 + ], + "type": "text", + "content": "To address these challenges, we introduce a novel approach that can generate high-resolution (HR) images in only one sampling step, without compromising the diversity and perceptual quality of the diffusion model, as shown in Fig. 1 and Fig. 2. Specifically, we propose to directly learn a well-paired bi-directional deterministic mapping between the input random noise and the generated HR image from a teacher diffusion model. To accelerate the generation of well-matched training data, we first derive a deterministic sampling strategy from the most recent state-of-the-art work [46], designed for accelerating diffusion-based SR, from its original stochastic formulation. Additionally, we propose a novel consistency-preserving loss to leverage ground-truth images, further enhancing the perceptual quality of the generated HR images by minimizing the error between ground-truth (GT) images and those generated from the predicted initial state. Experimental results demonstrate that our method achieves comparable or even better performance compared to SOTA methods and the teacher diffusion model [46], while greatly reducing the number of inference steps from 15 to 1, resulting in up to a " + }, + { + "bbox": [ + 46, + 400, + 289, + 651 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 46, + 400, + 289, + 651 + ], + "type": "text", + "content": " speedup in inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 652, + 266, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 652, + 266, + 663 + ], + "spans": [ + { + "bbox": [ + 58, + 652, + 266, + 663 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "- We accelerate the diffusion-based SR model to a single inference step with comparable or even superior performance for the first time. Instead of shortening the Markov chain of the generation process, we propose a simple yet" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 547, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 547, + 95 + ], + "type": "text", + "content": "effective approach that directly distills a deterministic generation function into a student network." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 96, + 547, + 263 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 306, + 96, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 547, + 144 + ], + "type": "text", + "content": "- To further fasten training, we derive a deterministic sampling strategy from the recent SOTA method [46] on accelerating the SR task, enabling efficient generation of well-matched training pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 144, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 547, + 204 + ], + "type": "text", + "content": "- We propose a novel consistency-preserving loss that can utilize the ground-truth images during training, preventing the student model from only focusing on fitting the deterministic mapping of the teacher diffusion model, therefore leading to better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 204, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 547, + 263 + ], + "type": "text", + "content": "- Extensive experiments on both synthetic and real-world datasets show that our proposed method can achieve comparable or even superior performance compared to SOTA methods and the teacher diffusion model, while greatly reducing the number of inference steps from 15 to 1." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 268, + 392, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 392, + 280 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 392, + 280 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 287, + 441, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 287, + 441, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 441, + 300 + ], + "type": "text", + "content": "2.1. Image Super-Resolution" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 305, + 548, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 548, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 548, + 617 + ], + "type": "text", + "content": "With the rise of deep learning, deep learning-based techniques gradually become the mainstream of the SR task [8, 45]. One prevalent approach of early works is to train a regression model using paired training data [1, 2, 16, 44]. While the expectation of the posterior distribution can be well modeled, they inevitably suffer from the over-smooth problem [17, 27, 34]. To improve the perceptual quality of the generated HR images, generative-based SR models attract increasing attention, e.g., autoregressive-based models [6, 26, 29, 30]. While significant improvements are achieved, the computational cost of autoregressive models is usually large. Subsequently, normalizing flows [23, 42] are demonstrated to have good perceptual quality under an efficient inference process, while its network design is restricted by the requirements of the invertibility and ease of calculation. Besides, GAN-based methods also achieve great success in terms of perceptual quality [9, 13, 17, 27, 34]. However, the training of GAN-based methods is usually unstable. Recently, diffusion-based models have been widely investigated in SR [4, 5, 14, 32, 33]. The diffusion-based SR methods can be roughly summarized into two categories, concatenating the LR image to the input of the denoiser [32, 33], and modifying the backward process of a pre-trained diffusion model [4, 5, 10, 14]. While promising results are achieved, they rely on a large number of inference steps, which greatly hinders the application of diffusion-based models." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 623, + 482, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 482, + 635 + ], + "type": "text", + "content": "2.2. Acceleration of Diffusion Models" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Recently, the acceleration of diffusion models has attracted more and more attention. Several algorithms are proposed for general diffusion models [22, 28, 37, 38] and proved quite effective for image generations. One intuitive strategy among them is to distill the diffusion models to a student model. However, the huge training overhead to solve" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25797" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 93, + 116 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 93, + 116 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 93, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 93, + 116 + ], + "type": "image", + "image_path": "e1345c93f2d7de5617f9168178235aa1d593c40d6f7cc5808e2225192c1c149f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 74, + 152, + 115 + ], + "blocks": [ + { + "bbox": [ + 93, + 74, + 152, + 115 + ], + "lines": [ + { + "bbox": [ + 93, + 74, + 152, + 115 + ], + "spans": [ + { + "bbox": [ + 93, + 74, + 152, + 115 + ], + "type": "image", + "image_path": "3d61da1c6348d7e70ca0847167b48334465a92ce0ff65f1e01b22d5a64869b23.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 155, + 76, + 190, + 115 + ], + "blocks": [ + { + "bbox": [ + 155, + 76, + 190, + 115 + ], + "lines": [ + { + "bbox": [ + 155, + 76, + 190, + 115 + ], + "spans": [ + { + "bbox": [ + 155, + 76, + 190, + 115 + ], + "type": "image", + "image_path": "41ad4a4cf1e7fc33a6d8a5df4bba512b7f3050650aa25bb9f7bc08cbc221ffed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 190, + 74, + 225, + 116 + ], + "blocks": [ + { + "bbox": [ + 190, + 74, + 225, + 116 + ], + "lines": [ + { + "bbox": [ + 190, + 74, + 225, + 116 + ], + "spans": [ + { + "bbox": [ + 190, + 74, + 225, + 116 + ], + "type": "image", + "image_path": "e86bf3426ed1ad88cb6ea836e95c372bc000a4c581b7501306afcdbc62e7b08c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 226, + 74, + 255, + 116 + ], + "blocks": [ + { + "bbox": [ + 226, + 74, + 255, + 116 + ], + "lines": [ + { + "bbox": [ + 226, + 74, + 255, + 116 + ], + "spans": [ + { + "bbox": [ + 226, + 74, + 255, + 116 + ], + "type": "image", + "image_path": "3f7cc6c4674ecf76bf119a0c35525fef2c8bcc4f237fc8d4b4bdec3458d98883.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 74, + 278, + 116 + ], + "blocks": [ + { + "bbox": [ + 257, + 74, + 278, + 116 + ], + "lines": [ + { + "bbox": [ + 257, + 74, + 278, + 116 + ], + "spans": [ + { + "bbox": [ + 257, + 74, + 278, + 116 + ], + "type": "image", + "image_path": "c10cb5336b672ab82e8dfd1d4502eec08a3dd4a148f46593d8ead66551dbc63b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 144, + 285, + 200 + ], + "blocks": [ + { + "bbox": [ + 47, + 122, + 287, + 142 + ], + "lines": [ + { + "bbox": [ + 47, + 122, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 287, + 142 + ], + "type": "text", + "content": "(a) The inference of SR3 [33] starts from a pure noise, which requires a large number of inference steps (T=100 after using DDIM [37])." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 144, + 285, + 200 + ], + "lines": [ + { + "bbox": [ + 50, + 144, + 285, + 200 + ], + "spans": [ + { + "bbox": [ + 50, + 144, + 285, + 200 + ], + "type": "image", + "image_path": "70a4ad8a056d213f9351a3351109791d31035f6f57b4ed5a54f5254fa91fcbc6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 235, + 286, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "type": "text", + "content": "(b) The recent SOTA method ResShift [46] shortens the Markov chain to speed up the inference process by incorporating the information of the LR image " + }, + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "type": "text", + "content": " to the initial state " + }, + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "type": "inline_equation", + "content": "(\\mathrm{T} = 15)" + }, + { + "bbox": [ + 47, + 204, + 287, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 235, + 286, + 308 + ], + "lines": [ + { + "bbox": [ + 50, + 235, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 50, + 235, + 286, + 308 + ], + "type": "image", + "image_path": "09d2e12d0af47b13ca8344dfebb9d6e7fc9114eec488ac78ba2270f8e6d90a95.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "text", + "content": "(c) A simplified pipeline of the proposed method " + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "inline_equation", + "content": "\\text{SinSR}" + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "text", + "content": " (distill only). It directly learns the deterministic mapping between " + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "text", + "content": ", therefore the inference process can be further compressed into only one step " + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "inline_equation", + "content": "(\\mathrm{T} = 1)" + }, + { + "bbox": [ + 47, + 310, + 287, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 344, + 287, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 411 + ], + "type": "text", + "content": "Figure 3. A comparison between the vanilla diffusion-based SR method [33], a most recent method for acceleration of the diffusion-based SR [46], and the proposed one-step SR. Different from recent works that shorten the Markov chain to speed up the inference process [43, 46], the proposed method directly learns the deterministic generation process and the details can be found in Fig. 4." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 419, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 635 + ], + "type": "text", + "content": "the ordinary differential equation (ODE) of the inference process makes this scheme less attractive on a large-scale dataset [24]. To alleviate the training overhead, progressive distillation strategies are usually adopted [25, 35]. Meanwhile, instead of simply simulating the behavior of a teacher diffusion model through distillation, better inference paths are explored in an iterative manner [20, 21]. While progressive distillation effectively decreases the training overhead, the error accumulates at the same time, leading to an obvious performance loss in SR. Most recently, targeting the image restoration task, some works reformulate the diffusion process by either using the knowledge of degradation process [43] or a pre-defined distribution of the initial state [46], yielding a shortened Markov chain of the generation process and better performance than directly applying DDIM [37] in low-level tasks. However, they either require a clear formulation of the degradation or still require a relatively large number of inference steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 646, + 118, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 118, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 118, + 658 + ], + "type": "text", + "content": "3. Motivation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Preliminary. Given an LR image " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " and its corresponding HR image " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ", existing diffusion-based SR methods aim to model the conditional distribution " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "q(x_0|y)" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " through a Markov chain where a forward process is usually defined" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{t - 1}) = \\mathcal{N}(x_{t};\\sqrt{1 - \\beta_{t}} x_{t - 1},\\beta_{t}I)" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " with an initial state " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_{T}\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": ". The role of the diffusion model can be regarded as transferring the input domain (standard Gaussian noise) to the HR image domain conditioned on the LR image. Since the matching relationship between " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " is unknown, usually a diffusion model [11, 21, 33] through an iterative manner is required to learn/infer from an unknown mapping between " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": ". Our method is grounded in the idea that having an SR model that effectively captures the conditional distribution " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "q(x_0|y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " and establishes a deterministic mapping between " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\hat{x} 0" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " given an LR image " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": ", we can streamline the inference process to a single step by employing another network, denoted as " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": ", to learn the correspondence between " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": ", as illustrated in Fig. 3. Distillation for diffusion SR models: less is more. While the concept of distilling the mapping between " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": " to a student network has been previously explored [21], its application to SR introduces several challenges:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 289, + 547, + 456 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 306, + 289, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 547, + 348 + ], + "type": "text", + "content": "- The training overhead becomes substantial for one-step distillation due to a large number of inference steps of previous models, e.g., LDM [32] still need 100 steps after using DDIM [37] for inference to generate high-quality pairs " + }, + { + "bbox": [ + 306, + 289, + 547, + 348 + ], + "type": "inline_equation", + "content": "(\\hat{x}_0, x_T, y)" + }, + { + "bbox": [ + 306, + 289, + 547, + 348 + ], + "type": "text", + "content": " as the training data of the student model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 348, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 547, + 456 + ], + "type": "text", + "content": "- The performance degradation is attributed to the introduction of a more intricate distillation strategy involving iteration. For example, to reduce the training overhead, an iterative distillation strategy [35] is adopted which gradually decreases the number of inference steps during training. However, despite achieving satisfactory results in generation tasks, the cumulative error significantly impacts the fidelity of the SR results, as SR tasks are relatively more sensitive to image quality." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 456, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 503 + ], + "type": "text", + "content": "To address the aforementioned two challenges, we propose to distill the diffusion SR process into a single step in a simple but effective way based on the following observations. More details of the observations can be seen in Sec. 5.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 506, + 547, + 698 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 306, + 506, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 547, + 602 + ], + "type": "text", + "content": "- We demonstrate that the most recent SOTA method for accelerating the diffusion-based SR [46], which achieves comparable performance in 15 steps as LDM [32] in 100 DDIM steps, has a deterministic mapping between " + }, + { + "bbox": [ + 306, + 506, + 547, + 602 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 306, + 506, + 547, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 506, + 547, + 602 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 306, + 506, + 547, + 602 + ], + "type": "text", + "content": ". Besides, the greatly reduced number of inference steps and the existence of the deterministic mapping make the training of a single-step distillation possible as shown in Fig. 6 and Table 4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "text", + "content": "- Learning the mapping between " + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "text", + "content": " is found to be easier than denoising " + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "text", + "content": " under different noise levels as shown in Table 5. Therefore, it is feasible to directly learn the mapping between " + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 306, + 603, + 545, + 662 + ], + "type": "text", + "content": " so that the accumulated error by the iterative distillation can be avoided." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 662, + 547, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 662, + 547, + 698 + ], + "spans": [ + { + "bbox": [ + 306, + 662, + 547, + 698 + ], + "type": "text", + "content": "- Due to the accumulated error, a more sophisticated distillation strategy (iterative-based) does not contribute to the improvement in our setting as shown in Table 6." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "The organization of the following sections is as follows:" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25798" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 49, + 347, + 192 + ], + "blocks": [ + { + "bbox": [ + 53, + 49, + 347, + 192 + ], + "lines": [ + { + "bbox": [ + 53, + 49, + 347, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 49, + 347, + 192 + ], + "type": "image", + "image_path": "b04aa19510ca2f5d685799789a32c5c959e3c54abf196e81c92745e97f94e8a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 196, + 263, + 207 + ], + "lines": [ + { + "bbox": [ + 126, + 196, + 263, + 207 + ], + "spans": [ + { + "bbox": [ + 126, + 196, + 263, + 207 + ], + "type": "text", + "content": "(a) One-step bi-directional distillation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 356, + 49, + 542, + 192 + ], + "blocks": [ + { + "bbox": [ + 356, + 49, + 542, + 192 + ], + "lines": [ + { + "bbox": [ + 356, + 49, + 542, + 192 + ], + "spans": [ + { + "bbox": [ + 356, + 49, + 542, + 192 + ], + "type": "image", + "image_path": "5e6ce58296fdae0744d7b6df8940dc397882d0084bb09b77366c99ee7facaf1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 195, + 514, + 207 + ], + "lines": [ + { + "bbox": [ + 399, + 195, + 514, + 207 + ], + "spans": [ + { + "bbox": [ + 399, + 195, + 514, + 207 + ], + "type": "text", + "content": "(b) Consistency preserving loss" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": "Figure 4. The overall framework of the proposed method. By minimizing " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dist}" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inverse}" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": ", the student network " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": " learns the deterministic bi-directional mapping between " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": " obtained from a pre-trained teacher diffusion model in one step. Meanwhile, the proposed consistency preserving loss " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gt}" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": " is optimized during training to utilize the information from the GT images to pursue better perceptual quality instead of simply fitting the deterministic mappings from the teacher model. Specifically, the GT image is first converted to its latent code " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{x}_T = f_{\\hat{\\theta}}(x_0,y,0)" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": ", and then converted back to calculate its reconstruction loss " + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "inline_equation", + "content": "L_{MSE}(f_{\\hat{\\theta}}(\\hat{x}_T,y,T),x_0)" + }, + { + "bbox": [ + 45, + 211, + 545, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 288, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 288, + 328 + ], + "type": "text", + "content": "we first demonstrate that ResShift [46], in which the inference process is originally stochastic, can be converted to a deterministic model without retraining in Sec 4.1, and then the proposed consistency preserving distillation in Sec 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 341, + 129, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 129, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 129, + 355 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 361, + 181, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 181, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 181, + 374 + ], + "type": "text", + "content": "4.1. Deterministic Sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "text", + "content": "A core difference between ResShift [46] and LDM [32] is the formulation of the initial state " + }, + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "text", + "content": ". Specifically, in ResShift [46], the information from the LR image " + }, + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "text", + "content": " is integrated into the diffusion step " + }, + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 380, + 288, + 429 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 439, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 439, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 72, + 439, + 287, + 453 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}, y\\right) = \\mathcal {N} \\left(x _ {t}; x _ {0} + \\eta_ {t} \\left(y - x _ {0}\\right), \\kappa^ {2} \\eta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "be0a06aeacdcf4226e2d90796805d174851229770d67d0750d0cd79d2400a9dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\eta_{t}" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": " is a serial of hyper-parameters that monotonically increases with the timestep " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": " and obeys " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\eta_T\\rightarrow 1" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\eta_0\\to 0" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": ". As such, the inverse of the diffusion process starts from an initial state with rich information from the LR image " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": " as follows " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "x_{T} = y + \\kappa \\sqrt{\\eta_{T}}\\epsilon" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": ". To generate a HR image " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": " from a given image " + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 464, + 287, + 547 + ], + "type": "text", + "content": ", the original inverse process of [46] is as follows" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 557, + 287, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 287, + 580 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, y\\right) = \\mathcal {N} \\left(x _ {t - 1} \\mid \\mu_ {\\theta} \\left(x _ {t}, y, t\\right), \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\mathbf {I}\\right), \\tag {2}", + "image_path": "560791533b272eeea4330e85b95e016b16ead6b416d7b75c8edce3e1c02d1f15.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,y,t)" + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "text", + "content": " is reparameterized by a deep network. As shown in Eq. 2, given an initial state " + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "inline_equation", + "content": "x_{T} = y + \\kappa \\sqrt{\\eta_{T}}\\epsilon" + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "text", + "content": " the generated image is stochastic due to the existence of the random noise during the sampling from " + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t,y)" + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "text", + "content": ". Inspired by DDIM sampling [37], we find that a non-Markovian reverse process " + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "inline_equation", + "content": "q(x_{t - 1}|x_t,x_0,y)" + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "text", + "content": " exists which keeps the marginal distribution " + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{0},y)" + }, + { + "bbox": [ + 46, + 589, + 287, + 674 + ], + "type": "text", + "content": " unchanged so" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 280, + 544, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 544, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 544, + 304 + ], + "type": "text", + "content": "that it can be directly adopted to a pre-trained model. The reformulated deterministic reverse process is as follows" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 336, + 314, + 545, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 314, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 336, + 314, + 545, + 327 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}, y\\right) = \\delta \\left(k _ {t} x _ {0} + m _ {t} x _ {t} + j _ {t} y\\right), \\tag {3}", + "image_path": "7616530419e8d5b9252623b575ad532b765cec46d3061a7a3614e997785a21a3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 336, + 533, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 533, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 533, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 336, + 533, + 348 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 336, + 533, + 348 + ], + "type": "text", + "content": " is the unit impulse, and " + }, + { + "bbox": [ + 305, + 336, + 533, + 348 + ], + "type": "inline_equation", + "content": "k_{t},m_{t},j_{t}" + }, + { + "bbox": [ + 305, + 336, + 533, + 348 + ], + "type": "text", + "content": " are as follows" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 357, + 545, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 357, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 345, + 357, + 545, + 408 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} m _ {t} = \\sqrt {\\frac {\\eta_ {t - 1}}{\\eta_ {t}}} \\\\ j _ {t} = \\eta_ {t - 1} - \\sqrt {\\eta_ {t - 1} \\eta_ {t}} \\\\ k _ {t} = 1 - \\eta_ {t - 1} + \\sqrt {\\eta_ {t - 1} \\eta_ {t}} - \\sqrt {\\frac {\\eta_ {t - 1}}{\\eta_ {t}}}. \\end{array} \\right. \\tag {4}", + "image_path": "a6355af014abf6b29d585d94d4da98cc14096cd24a0ada64d0407ffa0d8ec11f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "content": "The details of the derivation can be found in the supplementary material. As a consequence, for inference, the reverse process conditioned on " + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "content": " is reformulated as follows" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 349, + 462, + 500, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 462, + 500, + 489 + ], + "spans": [ + { + "bbox": [ + 349, + 462, + 500, + 489 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t - 1} = k _ {t} \\hat {x} _ {0} + m _ {t} x _ {t} + j _ {t} y \\\\ = k _ {t} f _ {\\theta} \\left(x _ {t}, y, t\\right) + m _ {t} x _ {t} + j _ {t} y, \\\\ \\end{array}", + "image_path": "cc76dbf049dade20f3ad1e15c4604dbd56a731e38ffccb7d2f4b30e8228d2b2a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x_t,y,t)" + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "text", + "content": " is the predicted HR image from a pretrained ResShift [46] model. By sampling from the reformulated process in Eq. 5, a deterministic mapping between " + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "text", + "content": " can be obtained and is denoted as " + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "inline_equation", + "content": "F_{\\theta}(x_T,y)" + }, + { + "bbox": [ + 305, + 498, + 545, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 554, + 492, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 492, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 492, + 567 + ], + "type": "text", + "content": "4.2. Consistency Preserving Distillation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "text", + "content": "Vanilla distillation. We propose utilizing a student network " + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "text", + "content": " to learn the deterministic mapping " + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "text", + "content": " between the random initialized state " + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "text", + "content": " and its deterministic output " + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "inline_equation", + "content": "F_{\\theta}(x_{T},y)" + }, + { + "bbox": [ + 305, + 573, + 545, + 632 + ], + "type": "text", + "content": " from a teacher diffusion model. The vanilla distillation loss is defined as follows" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 336, + 643, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 643, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 336, + 643, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(x _ {T}, y, T\\right), F _ {\\theta} \\left(x _ {T}, y\\right)\\right), \\tag {6}", + "image_path": "e4a65158d33aa6c194b1bc31601fd6f8edd9cfa77ae981b3c629fc5d40c10ed8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}(x_T,y,T)" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " is the student network that directly predicts the HR image in only one step, and " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " represents the proposed deterministic inference process of ResShift [46] in Sec. 4.1 through an iterative manner using a pre-trained" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "For ease of presentation, the LR image is " + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": " is pre-upsampled to the same spatial resolution with the HR image " + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": ". Besides, similar to [32, 46], the diffusion is conducted in the latent space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25799" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "network parameterized by " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". We observe that the student model trained solely with the distillation loss in Eq. 6 already achieves promising results in just one inference step, as indicated by “(distill only)” in the result tables." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 240 + ], + "type": "text", + "content": "Regularization by the ground-truth image. A limitation of the aforementioned vanilla distillation strategy is that the GT image is not utilized during training, thereby restricting the upper performance bound of the student model. To further enhance the student's performance, we propose a novel strategy that incorporates a learned inversion of the HR image to provide additional regularization from the ground-truth images. In addition to the vanilla distillation loss, the student network concurrently learns the inverse mapping during training by minimizing the following loss," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 249, + 288, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 249, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 76, + 249, + 288, + 262 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n v e r s e}} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(F _ {\\theta} \\left(x _ {T}, y\\right), y, 0\\right), x _ {T}\\right), \\tag {7}", + "image_path": "91f224eea7e2cefe14ab4bd3ce537fcd5fa166454d4f0abfbe0ea1fdc9adbf22.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "text", + "content": "where the last parameter of " + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "text", + "content": " is set from " + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "text", + "content": " in Eq. 6 to 0, indicating that the model is predicting the inversion instead of the " + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "text", + "content": ". Then the GT image " + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "text", + "content": " can be employed to regularize the output SR image given its predicted inversion " + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "inline_equation", + "content": "\\hat{x}_T" + }, + { + "bbox": [ + 46, + 271, + 289, + 330 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 338, + 214, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 338, + 214, + 352 + ], + "spans": [ + { + "bbox": [ + 101, + 338, + 214, + 352 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {T} = \\operatorname {d e t a c h} (f _ {\\hat {\\theta}} (x _ {0}, y, 0))", + "image_path": "3a378d5ef69c22ff6f74795ddf81ae2c8a5104b3b6884285caf444677cb24bbb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 352, + 287, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 352, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 101, + 352, + 287, + 367 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g t} = L _ {M S E} \\left(f _ {\\hat {\\theta}} \\left(\\hat {x} _ {T}, y, T\\right), x _ {0}\\right), \\tag {8}", + "image_path": "41cc9c85ceafd3783fb1856bddaead577c9145605adc182c2404125cdbbca5ef.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gt}" + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "content": " is the proposed consistency preserving loss. By reusing " + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "content": " to learn both " + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}(\\cdot, \\cdot, T)" + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}(\\cdot, \\cdot, 0)" + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "content": " simultaneously, we can initialize the parameter " + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "content": " of the student model from the teacher one " + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 375, + 287, + 424 + ], + "type": "text", + "content": " to speed up the training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 425, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 459 + ], + "type": "text", + "content": "The overall training objective. The student network is trained to minimize the aforementioned three losses at the same time as follows" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 468, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 468, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 61, + 468, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\hat {\\theta} = \\arg \\min _ {\\hat {\\theta}} \\mathbb {E} _ {y, x _ {0}, x _ {T}} \\left[ \\mathcal {L} _ {\\text {d i s t i l l}} + \\mathcal {L} _ {\\text {r e v e r s e}} + \\mathcal {L} _ {g t} \\right], \\tag {9}", + "image_path": "76c313d7407f501008af2bca294074ec506f764f051ceea22c657980f198be57.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 497, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 545 + ], + "type": "text", + "content": "where the losses are defined in Eq. 6, 7, and 8 respectively. We assign equal weight to each loss term, and ablation studies are in the supplementary material. The overall of the proposed method is summarized in Algorithm 1 and Fig. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 555, + 123, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 123, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 123, + 569 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 575, + 161, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 161, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 161, + 588 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Training Details. For a fair comparison, we follow the same experimental setup and backbone design as that in [46]. Specifically, the main difference is that we finetuned the model for 30K iterations instead of training from scratch for 500K in [46]. We find that the student model can converge quickly so that even if for each iteration we need extra time to solve the ODE to get paired training data, the overall training time is still much shorter than retraining a model from scratch following [46]. We train the models on the training set of ImageNet [7] following the same pipeline" + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 306, + 85, + 545, + 353 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 397, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 397, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 397, + 84 + ], + "type": "text", + "content": "Algorithm 1 Training" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": "Require: Pre-trained teacher diffusion model " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \nRequire: Paired training set " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "(X,Y)" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n1: Init " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " from the pre-trained model, i.e., " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}\\gets \\theta" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n2: while not converged do \n3: sample " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "x_0,y\\sim (X,Y)" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n4: sample " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\kappa^2\\eta_T\\mathbf{I})" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "x_{T} = y + \\epsilon" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n6: for " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "t = T,T - 1,\\dots,1" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " do \n7: if " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " then \n8: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = f_\\theta (x_1,y,1)" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n9: else \n10: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "x_{t - 1} = k_t f_\\theta (x_t,y,t) + m_tx_t + j_t y" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n11: end if \n12: end for \n13: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{distill} = L_{MSE}(f_{\\hat{\\theta}}(x_T,y,T),\\hat{x}_0)" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n14: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inverse} = L_{MSE}(f_{\\hat{\\theta}}(\\hat{x}_0,y,0),x_T)" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n15: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\hat{x}_T = f_\\hat{\\theta} (x_0,y,0)" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n16: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gt} = L_{MSE}(f_{\\hat{\\theta}}(\\operatorname {detach}(\\hat{x}_T),y,T),x_0)" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n17: " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{distill} + \\mathcal{L}_{inverse} + \\mathcal{L}_{gt}" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n18: Perform a gradient descent step on " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\nabla_{\\hat{\\theta}}\\mathcal{L}" + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "text", + "content": " \n19: end while \n20: return The student model " + }, + { + "bbox": [ + 306, + 85, + 545, + 353 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 374, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 398 + ], + "type": "text", + "content": "with ResShift [46] where the degradation model is adopted from RealESRGAN [41]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 399, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 518 + ], + "type": "text", + "content": "Compared methods. We compare our method with several representative SR models, including RealSR-JPEG [12], ESRGAN [40], BSRGAN [47], SwinIR [18], RealESRGAN [41], DASR [19], LDM [32], and ResShift [46]. For a comprehensive comparison, we further evaluate the performance of diffusion-based models LDM [32] and ResShift [46] with a reduced number of sampling steps. Besides, we compare the proposed method with Rectified-Flow [21], a SOTA method that can compress the generation process into a single step, in Table 6." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 519, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 602 + ], + "type": "text", + "content": "Metrics. For the evaluation of the proposed method on the synthetic testing dataset with reference images, we utilize PSNR, SSIM, and LPIPS [48] to measure the fidelity performance. Besides, two recent SOTA non-reference metrics are used to justify the realism of all the images, i.e., CLIP-IQA [39] which leverages a CLIP model [31] pre-trained on a large-scale dataset (Laion400M [36]) and MUSIQ [15]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 611, + 429, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 429, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 429, + 623 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": "Evaluation on real-world datasets. RealSR [3] and RealSet65 [46] are adopted to evaluate the generalization ability of the model on unseen real-world data. Specifically, in RealSR [3], there are 100 real images captured by two different cameras in different scenarios. Besides, RealSet65 [46] includes 65 LR images in total, collected from widely used datasets and the internet. The results on these two datasets" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25800" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 42, + 203, + 182 + ], + "blocks": [ + { + "bbox": [ + 69, + 42, + 203, + 182 + ], + "lines": [ + { + "bbox": [ + 69, + 42, + 203, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 42, + 203, + 182 + ], + "type": "image", + "image_path": "c5ed0296b37d4ce3e6598b5894c72043018d6208e02b4d27d55637544ded7f6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 182, + 155, + 190 + ], + "lines": [ + { + "bbox": [ + 117, + 182, + 155, + 190 + ], + "spans": [ + { + "bbox": [ + 117, + 182, + 155, + 190 + ], + "type": "text", + "content": "(a) LR input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 42, + 268, + 108 + ], + "blocks": [ + { + "bbox": [ + 205, + 42, + 268, + 108 + ], + "lines": [ + { + "bbox": [ + 205, + 42, + 268, + 108 + ], + "spans": [ + { + "bbox": [ + 205, + 42, + 268, + 108 + ], + "type": "image", + "image_path": "580e65fc391fcad3e31a61acd43bde698b425c15d37943e46d23f6375f59f260.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 109, + 257, + 115 + ], + "lines": [ + { + "bbox": [ + 216, + 109, + 257, + 115 + ], + "spans": [ + { + "bbox": [ + 216, + 109, + 257, + 115 + ], + "type": "text", + "content": "(b) BSRGAN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 269, + 42, + 332, + 108 + ], + "blocks": [ + { + "bbox": [ + 269, + 42, + 332, + 108 + ], + "lines": [ + { + "bbox": [ + 269, + 42, + 332, + 108 + ], + "spans": [ + { + "bbox": [ + 269, + 42, + 332, + 108 + ], + "type": "image", + "image_path": "f8753f7fcd123ee25997f728b3f8627b1d03615981b4d8f830c0206838dbd7b1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 109, + 327, + 115 + ], + "lines": [ + { + "bbox": [ + 274, + 109, + 327, + 115 + ], + "spans": [ + { + "bbox": [ + 274, + 109, + 327, + 115 + ], + "type": "text", + "content": "(c) RealESRGAN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 333, + 42, + 397, + 108 + ], + "blocks": [ + { + "bbox": [ + 333, + 42, + 397, + 108 + ], + "lines": [ + { + "bbox": [ + 333, + 42, + 397, + 108 + ], + "spans": [ + { + "bbox": [ + 333, + 42, + 397, + 108 + ], + "type": "image", + "image_path": "beec3d4f1e56ce25a7b87889c5d300d25a04514e5cf04cf698802b587256bb58.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 109, + 381, + 115 + ], + "lines": [ + { + "bbox": [ + 348, + 109, + 381, + 115 + ], + "spans": [ + { + "bbox": [ + 348, + 109, + 381, + 115 + ], + "type": "text", + "content": "(d) SwinIR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 398, + 42, + 461, + 108 + ], + "blocks": [ + { + "bbox": [ + 398, + 42, + 461, + 108 + ], + "lines": [ + { + "bbox": [ + 398, + 42, + 461, + 108 + ], + "spans": [ + { + "bbox": [ + 398, + 42, + 461, + 108 + ], + "type": "image", + "image_path": "d3a65906fc0e119f0257d2ffe72d069e5f60498c3671f48a18545266985335b9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 109, + 443, + 115 + ], + "lines": [ + { + "bbox": [ + 414, + 109, + 443, + 115 + ], + "spans": [ + { + "bbox": [ + 414, + 109, + 443, + 115 + ], + "type": "text", + "content": "(e) DASR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 462, + 42, + 525, + 108 + ], + "blocks": [ + { + "bbox": [ + 462, + 42, + 525, + 108 + ], + "lines": [ + { + "bbox": [ + 462, + 42, + 525, + 108 + ], + "spans": [ + { + "bbox": [ + 462, + 42, + 525, + 108 + ], + "type": "image", + "image_path": "8b6e70bd8733be0840f34790dd9583292f655bbe86df25415d64ca7774606a14.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 109, + 512, + 115 + ], + "lines": [ + { + "bbox": [ + 474, + 109, + 512, + 115 + ], + "spans": [ + { + "bbox": [ + 474, + 109, + 512, + 115 + ], + "type": "text", + "content": "(f) LDM-100" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 69, + 190, + 203, + 323 + ], + "blocks": [ + { + "bbox": [ + 69, + 190, + 203, + 323 + ], + "lines": [ + { + "bbox": [ + 69, + 190, + 203, + 323 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 203, + 323 + ], + "type": "image", + "image_path": "17a59811229d322c1aa2787f6eb9355ca004ed286dc3783c0dad5e7ee1a198dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 325, + 155, + 332 + ], + "lines": [ + { + "bbox": [ + 117, + 325, + 155, + 332 + ], + "spans": [ + { + "bbox": [ + 117, + 325, + 155, + 332 + ], + "type": "text", + "content": "(a) LR input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 205, + 116, + 268, + 182 + ], + "blocks": [ + { + "bbox": [ + 205, + 116, + 268, + 182 + ], + "lines": [ + { + "bbox": [ + 205, + 116, + 268, + 182 + ], + "spans": [ + { + "bbox": [ + 205, + 116, + 268, + 182 + ], + "type": "image", + "image_path": "540c93fec6a816908c0a12d32deee2e754591e16af15f64913f4b892bf37237f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 182, + 255, + 190 + ], + "lines": [ + { + "bbox": [ + 219, + 182, + 255, + 190 + ], + "spans": [ + { + "bbox": [ + 219, + 182, + 255, + 190 + ], + "type": "text", + "content": "(g) LDM-30" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 269, + 116, + 332, + 182 + ], + "blocks": [ + { + "bbox": [ + 269, + 116, + 332, + 182 + ], + "lines": [ + { + "bbox": [ + 269, + 116, + 332, + 182 + ], + "spans": [ + { + "bbox": [ + 269, + 116, + 332, + 182 + ], + "type": "image", + "image_path": "fff71557f5e5e0d733cf2c268436554a1e973f142c94d3862d8a1e78545e8c5a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 182, + 319, + 190 + ], + "lines": [ + { + "bbox": [ + 283, + 182, + 319, + 190 + ], + "spans": [ + { + "bbox": [ + 283, + 182, + 319, + 190 + ], + "type": "text", + "content": "(h) LDM-15" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 333, + 116, + 397, + 182 + ], + "blocks": [ + { + "bbox": [ + 333, + 116, + 397, + 182 + ], + "lines": [ + { + "bbox": [ + 333, + 116, + 397, + 182 + ], + "spans": [ + { + "bbox": [ + 333, + 116, + 397, + 182 + ], + "type": "image", + "image_path": "72eab5a26d1f69e3ec0b3f7228dda200db2344027a9da564ac45c8ae5374d0e6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 182, + 387, + 190 + ], + "lines": [ + { + "bbox": [ + 344, + 182, + 387, + 190 + ], + "spans": [ + { + "bbox": [ + 344, + 182, + 387, + 190 + ], + "type": "text", + "content": "(i) ResShift-15" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 398, + 116, + 461, + 182 + ], + "blocks": [ + { + "bbox": [ + 398, + 116, + 461, + 182 + ], + "lines": [ + { + "bbox": [ + 398, + 116, + 461, + 182 + ], + "spans": [ + { + "bbox": [ + 398, + 116, + 461, + 182 + ], + "type": "image", + "image_path": "0529e6bbc423f940ba942b949cdffea65308068e762ebf397ce3d772735d2e97.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 182, + 449, + 190 + ], + "lines": [ + { + "bbox": [ + 410, + 182, + 449, + 190 + ], + "spans": [ + { + "bbox": [ + 410, + 182, + 449, + 190 + ], + "type": "text", + "content": "(j) ResShift-5" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 462, + 116, + 525, + 182 + ], + "blocks": [ + { + "bbox": [ + 462, + 116, + 525, + 182 + ], + "lines": [ + { + "bbox": [ + 462, + 116, + 525, + 182 + ], + "spans": [ + { + "bbox": [ + 462, + 116, + 525, + 182 + ], + "type": "image", + "image_path": "1cee4c38d8e016b2baaec3d4fca6aac4ecb4040ead2ee2babc2f00750905a1d4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 182, + 508, + 189 + ], + "lines": [ + { + "bbox": [ + 477, + 182, + 508, + 189 + ], + "spans": [ + { + "bbox": [ + 477, + 182, + 508, + 189 + ], + "type": "text", + "content": "(k) Ours-1" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 67, + 335, + 202, + 469 + ], + "blocks": [ + { + "bbox": [ + 67, + 335, + 202, + 469 + ], + "lines": [ + { + "bbox": [ + 67, + 335, + 202, + 469 + ], + "spans": [ + { + "bbox": [ + 67, + 335, + 202, + 469 + ], + "type": "image", + "image_path": "2c9ef5256714e761b8a97dc7c330b509a73468bf04b67d86bf956a888f7526a6.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 479, + 449, + 488 + ], + "lines": [ + { + "bbox": [ + 67, + 479, + 449, + 488 + ], + "spans": [ + { + "bbox": [ + 67, + 479, + 449, + 488 + ], + "type": "text", + "content": "Figure 5. Visual comparison on real-world samples. Please zoom in for more details." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 470, + 154, + 477 + ], + "lines": [ + { + "bbox": [ + 67, + 470, + 154, + 477 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 154, + 477 + ], + "type": "text", + "content": "(a) LR input" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 470, + 154, + 477 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 470, + 154, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 470, + 154, + 477 + ], + "type": "text", + "content": "(a) LR input" + } + ] + } + ], + "index": 25, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 204, + 190, + 268, + 253 + ], + "blocks": [ + { + "bbox": [ + 204, + 190, + 268, + 253 + ], + "lines": [ + { + "bbox": [ + 204, + 190, + 268, + 253 + ], + "spans": [ + { + "bbox": [ + 204, + 190, + 268, + 253 + ], + "type": "image", + "image_path": "2660b2aa9542303ccafed587fe8deae32408c0e886d2a283326414134f44e89a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 254, + 257, + 260 + ], + "lines": [ + { + "bbox": [ + 217, + 254, + 257, + 260 + ], + "spans": [ + { + "bbox": [ + 217, + 254, + 257, + 260 + ], + "type": "text", + "content": "(b) BSRGAN" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 269, + 190, + 332, + 253 + ], + "blocks": [ + { + "bbox": [ + 269, + 190, + 332, + 253 + ], + "lines": [ + { + "bbox": [ + 269, + 190, + 332, + 253 + ], + "spans": [ + { + "bbox": [ + 269, + 190, + 332, + 253 + ], + "type": "image", + "image_path": "a7972bb939a3575d336ca78b04f981bcab8c0d37823bb29c2ddb65f4d1501f7e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 254, + 327, + 260 + ], + "lines": [ + { + "bbox": [ + 275, + 254, + 327, + 260 + ], + "spans": [ + { + "bbox": [ + 275, + 254, + 327, + 260 + ], + "type": "text", + "content": "(c) RealESRGAN" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 333, + 190, + 397, + 253 + ], + "blocks": [ + { + "bbox": [ + 333, + 190, + 397, + 253 + ], + "lines": [ + { + "bbox": [ + 333, + 190, + 397, + 253 + ], + "spans": [ + { + "bbox": [ + 333, + 190, + 397, + 253 + ], + "type": "image", + "image_path": "3ac65be59d92ca1ddafee366c833329d32b3638e3b1b66aa6edfd1464b8af59e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 254, + 381, + 260 + ], + "lines": [ + { + "bbox": [ + 348, + 254, + 381, + 260 + ], + "spans": [ + { + "bbox": [ + 348, + 254, + 381, + 260 + ], + "type": "text", + "content": "(d) SwinIR" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 398, + 190, + 460, + 253 + ], + "blocks": [ + { + "bbox": [ + 398, + 190, + 460, + 253 + ], + "lines": [ + { + "bbox": [ + 398, + 190, + 460, + 253 + ], + "spans": [ + { + "bbox": [ + 398, + 190, + 460, + 253 + ], + "type": "image", + "image_path": "dd18df93c60f2d6dd94edbfa37a81a876cb0a8ae93b8924ca1dcb278204f2bcd.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 254, + 444, + 260 + ], + "lines": [ + { + "bbox": [ + 414, + 254, + 444, + 260 + ], + "spans": [ + { + "bbox": [ + 414, + 254, + 444, + 260 + ], + "type": "text", + "content": "(e) DASR" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 462, + 190, + 525, + 253 + ], + "blocks": [ + { + "bbox": [ + 462, + 190, + 525, + 253 + ], + "lines": [ + { + "bbox": [ + 462, + 190, + 525, + 253 + ], + "spans": [ + { + "bbox": [ + 462, + 190, + 525, + 253 + ], + "type": "image", + "image_path": "b3e29b83d86dd71ee339850569df2cb7f617e6f836367af9c8587fdbdfc5e5a8.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 254, + 512, + 260 + ], + "lines": [ + { + "bbox": [ + 474, + 254, + 512, + 260 + ], + "spans": [ + { + "bbox": [ + 474, + 254, + 512, + 260 + ], + "type": "text", + "content": "(f) LDM-100" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 205, + 261, + 268, + 323 + ], + "blocks": [ + { + "bbox": [ + 205, + 261, + 268, + 323 + ], + "lines": [ + { + "bbox": [ + 205, + 261, + 268, + 323 + ], + "spans": [ + { + "bbox": [ + 205, + 261, + 268, + 323 + ], + "type": "image", + "image_path": "ea6f0f0495c04ad25223c915695514d36c31ddee6f76c98a7242b4b379835670.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 325, + 254, + 332 + ], + "lines": [ + { + "bbox": [ + 219, + 325, + 254, + 332 + ], + "spans": [ + { + "bbox": [ + 219, + 325, + 254, + 332 + ], + "type": "text", + "content": "(g) LDM-30" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 269, + 261, + 332, + 323 + ], + "blocks": [ + { + "bbox": [ + 269, + 261, + 332, + 323 + ], + "lines": [ + { + "bbox": [ + 269, + 261, + 332, + 323 + ], + "spans": [ + { + "bbox": [ + 269, + 261, + 332, + 323 + ], + "type": "image", + "image_path": "b70570e1df60f4d0b17e93f29f76f2c39575f71fe30053d13a804de910bb39e8.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 325, + 319, + 332 + ], + "lines": [ + { + "bbox": [ + 282, + 325, + 319, + 332 + ], + "spans": [ + { + "bbox": [ + 282, + 325, + 319, + 332 + ], + "type": "text", + "content": "(h) LDM-15" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 333, + 261, + 397, + 323 + ], + "blocks": [ + { + "bbox": [ + 333, + 261, + 397, + 323 + ], + "lines": [ + { + "bbox": [ + 333, + 261, + 397, + 323 + ], + "spans": [ + { + "bbox": [ + 333, + 261, + 397, + 323 + ], + "type": "image", + "image_path": "5bc4aacfd3c67882681afb57010c8ead1bb12961bc883beac500c61fd27c3604.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 325, + 387, + 332 + ], + "lines": [ + { + "bbox": [ + 343, + 325, + 387, + 332 + ], + "spans": [ + { + "bbox": [ + 343, + 325, + 387, + 332 + ], + "type": "text", + "content": "(i) ResShift-15" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 398, + 261, + 460, + 323 + ], + "blocks": [ + { + "bbox": [ + 398, + 261, + 460, + 323 + ], + "lines": [ + { + "bbox": [ + 398, + 261, + 460, + 323 + ], + "spans": [ + { + "bbox": [ + 398, + 261, + 460, + 323 + ], + "type": "image", + "image_path": "84f519c79d852450507766d1fe7da908fd63073681cd1c8825d4c069e6cd8ad6.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 325, + 449, + 332 + ], + "lines": [ + { + "bbox": [ + 409, + 325, + 449, + 332 + ], + "spans": [ + { + "bbox": [ + 409, + 325, + 449, + 332 + ], + "type": "text", + "content": "(j) ResShift-5" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 462, + 261, + 525, + 323 + ], + "blocks": [ + { + "bbox": [ + 462, + 261, + 525, + 323 + ], + "lines": [ + { + "bbox": [ + 462, + 261, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 462, + 261, + 525, + 323 + ], + "type": "image", + "image_path": "c9737b1b01f883e904c871a88409134aa30bbc0a9ce4aac60ec5961b122b9069.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 325, + 509, + 331 + ], + "lines": [ + { + "bbox": [ + 477, + 325, + 509, + 331 + ], + "spans": [ + { + "bbox": [ + 477, + 325, + 509, + 331 + ], + "type": "text", + "content": "(k) Ours-1" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 204, + 335, + 268, + 398 + ], + "blocks": [ + { + "bbox": [ + 204, + 335, + 268, + 398 + ], + "lines": [ + { + "bbox": [ + 204, + 335, + 268, + 398 + ], + "spans": [ + { + "bbox": [ + 204, + 335, + 268, + 398 + ], + "type": "image", + "image_path": "975517ad51ff6a562932f41eaef7a27deef68d3bef710d6c4292a9af8ca0753e.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 399, + 255, + 405 + ], + "lines": [ + { + "bbox": [ + 215, + 399, + 255, + 405 + ], + "spans": [ + { + "bbox": [ + 215, + 399, + 255, + 405 + ], + "type": "text", + "content": "(b) BSRGAN" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 269, + 335, + 331, + 398 + ], + "blocks": [ + { + "bbox": [ + 269, + 335, + 331, + 398 + ], + "lines": [ + { + "bbox": [ + 269, + 335, + 331, + 398 + ], + "spans": [ + { + "bbox": [ + 269, + 335, + 331, + 398 + ], + "type": "image", + "image_path": "952b2c1d47070c85bed5ff263d44ad1b06437ff246e3a3c8a817c28ff04681b3.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 399, + 325, + 406 + ], + "lines": [ + { + "bbox": [ + 274, + 399, + 325, + 406 + ], + "spans": [ + { + "bbox": [ + 274, + 399, + 325, + 406 + ], + "type": "text", + "content": "(c) RealESRGAN" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 332, + 335, + 396, + 398 + ], + "blocks": [ + { + "bbox": [ + 332, + 335, + 396, + 398 + ], + "lines": [ + { + "bbox": [ + 332, + 335, + 396, + 398 + ], + "spans": [ + { + "bbox": [ + 332, + 335, + 396, + 398 + ], + "type": "image", + "image_path": "b014e659343c575da89eebe1ed38400ce0b7981401c4b7a51df6d44076772d1a.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 399, + 381, + 405 + ], + "lines": [ + { + "bbox": [ + 347, + 399, + 381, + 405 + ], + "spans": [ + { + "bbox": [ + 347, + 399, + 381, + 405 + ], + "type": "text", + "content": "(d) SwinIR" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 397, + 335, + 460, + 398 + ], + "blocks": [ + { + "bbox": [ + 397, + 335, + 460, + 398 + ], + "lines": [ + { + "bbox": [ + 397, + 335, + 460, + 398 + ], + "spans": [ + { + "bbox": [ + 397, + 335, + 460, + 398 + ], + "type": "image", + "image_path": "3e3cc664c60065748777c6bd82fd4454550005c20f446226e7ef8ef4796744f4.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 399, + 443, + 405 + ], + "lines": [ + { + "bbox": [ + 413, + 399, + 443, + 405 + ], + "spans": [ + { + "bbox": [ + 413, + 399, + 443, + 405 + ], + "type": "text", + "content": "(e) DASR" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 461, + 335, + 524, + 398 + ], + "blocks": [ + { + "bbox": [ + 461, + 335, + 524, + 398 + ], + "lines": [ + { + "bbox": [ + 461, + 335, + 524, + 398 + ], + "spans": [ + { + "bbox": [ + 461, + 335, + 524, + 398 + ], + "type": "image", + "image_path": "1ab80edfb1a73fc481d0741ce9492334e35b24154c4dbd6ddf36eb92ef4b73ee.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 399, + 511, + 405 + ], + "lines": [ + { + "bbox": [ + 473, + 399, + 511, + 405 + ], + "spans": [ + { + "bbox": [ + 473, + 399, + 511, + 405 + ], + "type": "text", + "content": "(f) LDM-100" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "table", + "bbox": [ + 108, + 491, + 486, + 630 + ], + "blocks": [ + { + "bbox": [ + 108, + 491, + 486, + 630 + ], + "lines": [ + { + "bbox": [ + 108, + 491, + 486, + 630 + ], + "spans": [ + { + "bbox": [ + 108, + 491, + 486, + 630 + ], + "type": "table", + "html": "
MethodsDatasets
RealSRRealSet65
CLIPIQA↑MUSIQ↑CLIPIQA↑MUSIQ↑
ESRGAN [40]0.236229.0480.373942.369
RealSR-JPEG [12]0.361536.0760.528250.539
BSRGAN [47]0.543963.5860.616365.582
SwinIR [18]0.465459.6360.578263.822
RealESRGAN [41]0.489859.6780.599563.220
DASR [19]0.362945.8250.496555.708
LDM-15 [32]0.383649.3170.427447.488
ResShift-15 [46]0.595859.8730.653761.330
SinSR-1 (distill only)0.611957.1180.682261.267
SinSR-10.688761.5820.715062.169
", + "image_path": "0a103c522c7febed9f61ff5c6fbf8f726e1082e965312f70f7e26c5ec68632b9.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "table_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 461, + 406, + 524, + 469 + ], + "blocks": [ + { + "bbox": [ + 461, + 406, + 524, + 469 + ], + "lines": [ + { + "bbox": [ + 461, + 406, + 524, + 469 + ], + "spans": [ + { + "bbox": [ + 461, + 406, + 524, + 469 + ], + "type": "image", + "image_path": "b5e2776c1ddea5861af3e55275881e52173f5667897258117444c72c872647e6.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 470, + 507, + 476 + ], + "lines": [ + { + "bbox": [ + 477, + 470, + 507, + 476 + ], + "spans": [ + { + "bbox": [ + 477, + 470, + 507, + 476 + ], + "type": "text", + "content": "(k) Ours-1" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "bbox": [ + 47, + 632, + 544, + 643 + ], + "lines": [ + { + "bbox": [ + 47, + 632, + 544, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 544, + 643 + ], + "type": "text", + "content": "Table 1. Quantitative results of models on two real-world datasets. The best and second best results are highlighted in bold and underline." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 651, + 287, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 710 + ], + "type": "text", + "content": "are reported in Table 1. As shown in the table, the proposed method with only one inference step can outperform the teacher model that we used by a large margin. Besides, for the latest metric CLIPIQA, the proposed method archives the best performance among all the competitors. Some visual" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 306, + 651, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 545, + 674 + ], + "type": "text", + "content": "comparisons are shown in Fig. 5, in which the proposed method achieves promising results using only one step." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 306, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 547, + 714 + ], + "type": "text", + "content": "Evaluation on synthetic datasets. We further evaluate the performance of different methods on the synthetic dataset ImageNet-Test following the setting in [46]. Specifically," + } + ] + } + ], + "index": 64 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25801" + } + ] + } + ], + "index": 65 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 50, + 494, + 196 + ], + "blocks": [ + { + "bbox": [ + 97, + 50, + 494, + 196 + ], + "lines": [ + { + "bbox": [ + 97, + 50, + 494, + 196 + ], + "spans": [ + { + "bbox": [ + 97, + 50, + 494, + 196 + ], + "type": "table", + "html": "
MethodsMetrics
PSNR↑SSIM↑LPIPS↓CLIPQA↑MUSIQ↑
ESRGAN [40]20.670.4480.4850.45143.615
RealSR-JPEG [12]23.110.5910.3260.53746.981
BSRGAN [47]24.420.6590.2590.58154.697
SwinIR [18]23.990.6670.2380.56453.790
RealESRGAN [41]24.040.6650.2540.52352.538
DASR [19]24.750.6750.2500.53648.337
LDM-30 [32]24.490.6510.2480.57250.895
LDM-15 [32]24.890.6700.2690.51246.419
ResShift-15 [46]24.900.6730.2280.60353.897
SinSR-1 (distill only)24.690.6640.2220.60753.316
SinSR-124.560.6570.2210.61153.357
", + "image_path": "b861ce15582ffb7c49638f8f2be09309fc67e8d4229cd2c0d1ebb81de1812eff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 213, + 533, + 285 + ], + "blocks": [ + { + "bbox": [ + 63, + 198, + 528, + 210 + ], + "lines": [ + { + "bbox": [ + 63, + 198, + 528, + 210 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 528, + 210 + ], + "type": "text", + "content": "Table 2. Quantitative results of models on ImageNet-Test. The best and second best results are highlighted in bold and underline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 213, + 533, + 285 + ], + "lines": [ + { + "bbox": [ + 60, + 213, + 533, + 285 + ], + "spans": [ + { + "bbox": [ + 60, + 213, + 533, + 285 + ], + "type": "table", + "html": "
MetricsMethods
LDM-15LDM-30LDM-100ResShift-1ResShift-5ResShift-10ResShift-15SinSR-1
LPIPS↓0.2690.2480.2440.3830.3450.2740.2280.221
CLIPQA↑0.5120.5720.6200.3400.4170.5120.6030.611
Runtime (bs=64)0.046s0.080s0.249s0.012s0.021s0.033s0.047s0.012s
Runtime (bs=1)0.408s1.192s3.902s0.058s0.218s0.425s0.633s0.058s
# Parameters (M)113.60118.59118.59
", + "image_path": "136372d08c2c9a99fb8086a69f2ea5edbf28fbcaf5f0802221272d73c74c5bf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 479 + ], + "type": "text", + "content": "3000 high-resolution images are first randomly selected from the validation set of ImageNet [7]. The corresponding LR images are obtained by using the provided script in [46]. As shown in Table 2, while reducing the inference step from 15 to only 1 slightly decreases PSNR and SSIM, the proposed method achieves the best perceptual quality measured by LPIPS, a more recent full-reference image quality assessment (IQA) metric than SSIM. Besides, the proposed method also achieves the best performance among all the methods measured on the most recent SOTA metric CLIP-IQA [39], demonstrating that the proposed 1-step model is on par with or even slightly better than the teacher model with 15 inference steps in terms of perceptual qualities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "type": "text", + "content": "Evaluation of the efficiency. We assess the computational efficiency of the proposed method in comparison to SOTA approaches. As shown in Table 3, the proposed method demonstrates superior performance with only one inference step, outperforming ResShift [46]—the adopted teacher model, which had already significantly reduced the inference time compared to LDM [32]. It is worth noting that all methods presented in Table 3 run in latent space, and the computational cost of VQ-VAE is counted." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 598, + 110, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 110, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 110, + 611 + ], + "type": "text", + "content": "5.3. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "How important is the deterministic sampling? We evaluate the performance of the model trained on generated paired samples from the proposed deterministic sampling and the default stochastic sampling strategy " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "(x_{T},\\hat{F}_{\\theta}(x_{T},y))" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": " in [46]. Due to the randomness of the generated samples " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "x\\sim \\hat{F}_{\\theta}(x_T,y)" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": ", given a random noise " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": ", the prediction is an expectation of its conditional distribution. The comparison in Fig. 6 further verifies that the results trained w/o determin" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 321, + 531, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 288, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 544, + 312 + ], + "type": "text", + "content": "Table 3. Efficiency and performance comparisons with SOTA methods on ImageNet-Test. “-N” represents the number of sampling steps the model used. The running time per image is tested on a Tesla A100 GPU on the x4 (64→256) task averaged over the batch size (bs)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 321, + 531, + 365 + ], + "lines": [ + { + "bbox": [ + 321, + 321, + 531, + 365 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 531, + 365 + ], + "type": "table", + "html": "
MethodsCLIPIQA ↑MUSIQ ↑
w/ default sampling in [46]0.416651.53
SinSR (distill only)0.682261.27
", + "image_path": "5eb65a227565809193c27d82fa15becb62ef633389d9f258db5027d558e58f87.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 368, + 545, + 412 + ], + "lines": [ + { + "bbox": [ + 305, + 368, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 545, + 412 + ], + "type": "text", + "content": "Table 4. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5 using only distillation loss. We evaluate their performance on the RealSet65 testing set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 417, + 385, + 494 + ], + "blocks": [ + { + "bbox": [ + 308, + 417, + 385, + 494 + ], + "lines": [ + { + "bbox": [ + 308, + 417, + 385, + 494 + ], + "spans": [ + { + "bbox": [ + 308, + 417, + 385, + 494 + ], + "type": "image", + "image_path": "5471328b91dbd7fcb239d87f3fcaad6caad89615cde5a7319ccf3e930cce4710.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 495, + 385, + 548 + ], + "blocks": [ + { + "bbox": [ + 308, + 495, + 385, + 548 + ], + "lines": [ + { + "bbox": [ + 308, + 495, + 385, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 495, + 385, + 548 + ], + "type": "image", + "image_path": "6b16c95a606179683fec11e539dc4aeac961c4085858eefe8bd7dd4db860fa9f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 550, + 361, + 559 + ], + "lines": [ + { + "bbox": [ + 332, + 550, + 361, + 559 + ], + "spans": [ + { + "bbox": [ + 332, + 550, + 361, + 559 + ], + "type": "text", + "content": "(a) input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 386, + 417, + 463, + 494 + ], + "blocks": [ + { + "bbox": [ + 386, + 417, + 463, + 494 + ], + "lines": [ + { + "bbox": [ + 386, + 417, + 463, + 494 + ], + "spans": [ + { + "bbox": [ + 386, + 417, + 463, + 494 + ], + "type": "image", + "image_path": "180ced791f46b8649099b5d537facaaf0e5bff77f971676b52d1e92469b292f7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 386, + 495, + 463, + 548 + ], + "blocks": [ + { + "bbox": [ + 386, + 495, + 463, + 548 + ], + "lines": [ + { + "bbox": [ + 386, + 495, + 463, + 548 + ], + "spans": [ + { + "bbox": [ + 386, + 495, + 463, + 548 + ], + "type": "image", + "image_path": "c4e33e7b2d85be2fa0879f0cdf408e67da61266aab835cb353e078b3b43e7902.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 550, + 463, + 559 + ], + "lines": [ + { + "bbox": [ + 386, + 550, + 463, + 559 + ], + "spans": [ + { + "bbox": [ + 386, + 550, + 463, + 559 + ], + "type": "text", + "content": "(b) w/ sampling in [46]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 465, + 417, + 541, + 494 + ], + "blocks": [ + { + "bbox": [ + 465, + 417, + 541, + 494 + ], + "lines": [ + { + "bbox": [ + 465, + 417, + 541, + 494 + ], + "spans": [ + { + "bbox": [ + 465, + 417, + 541, + 494 + ], + "type": "image", + "image_path": "ffca07afc068a78db05ec7f72fb9c118706526c95c709bafa4ed46546ac51a21.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 465, + 495, + 541, + 548 + ], + "blocks": [ + { + "bbox": [ + 465, + 495, + 541, + 548 + ], + "lines": [ + { + "bbox": [ + 465, + 495, + 541, + 548 + ], + "spans": [ + { + "bbox": [ + 465, + 495, + 541, + 548 + ], + "type": "image", + "image_path": "0152792ab56f1b16c330dc38c0ad936127d3e60e9a8b52ce3c3291facb37eed7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 550, + 541, + 559 + ], + "lines": [ + { + "bbox": [ + 465, + 550, + 541, + 559 + ], + "spans": [ + { + "bbox": [ + 465, + 550, + 541, + 559 + ], + "type": "text", + "content": "(c) Ours (deterministic)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 560, + 545, + 594 + ], + "lines": [ + { + "bbox": [ + 305, + 560, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 560, + 545, + 594 + ], + "type": "text", + "content": "Figure 6. A comparison between the model trained with the default stochastic sampling process in ResShift [46] and the proposed deterministic sampling in Eq. 5. Best zoom in for more details." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "type": "text", + "content": "istic teacher model exhibit blurred details. Besides, as shown in Table 4, there is a significant performance degradation when we replace the proposed deterministic sampling with the default one in [46], demonstrating the effectiveness and necessity of involving the proposed deterministic sampling." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Why does a single-step distillation work? Previous studies suggest that directly learning the mapping between " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " is typically challenging due to the non-causal properties of the generation process [20]. However, our empirical find-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25802" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 60, + 276, + 117 + ], + "blocks": [ + { + "bbox": [ + 59, + 60, + 276, + 117 + ], + "lines": [ + { + "bbox": [ + 59, + 60, + 276, + 117 + ], + "spans": [ + { + "bbox": [ + 59, + 60, + 276, + 117 + ], + "type": "table", + "html": "
MethodsCLIPIQA ↑MUSIQ ↑
ResShift [46] (24.32M)0.536552.71
ResShift [46] (118.59M)0.653761.33
SinSR (distill only) (24.32M)0.649958.71
", + "image_path": "78a5c85ee849318fbd22fd86edce0a42cf29c45a2b51f7546feb7277b268f5b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 180, + 282, + 236 + ], + "blocks": [ + { + "bbox": [ + 46, + 119, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 119, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 119, + 287, + 175 + ], + "type": "text", + "content": "Table 5. A comparison of the models trained with different strategies on RealSet65. The model trained with the diffusion loss, i.e., ResShift, is more sensitive to the model size than directly learning the deterministic mapping between " + }, + { + "bbox": [ + 46, + 119, + 287, + 175 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 119, + 287, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 119, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 46, + 119, + 287, + 175 + ], + "type": "text", + "content": ", indicating that the deterministic mapping is relatively easier to learn." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 180, + 282, + 236 + ], + "lines": [ + { + "bbox": [ + 53, + 180, + 282, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 282, + 236 + ], + "type": "table", + "html": "
LPIPS↓MUSIQ↑CLIPQA↑
ResShift [46]0.227553.900.6029
w/ Rectified Flow [21]0.232251.050.5753
SinSR (distill only)0.222153.320.6072
", + "image_path": "65b70a99e046855fda2bbbeed6958d7dc60bff5cbd5d71c6ba70deb07a9f6de7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "content": "ings indicate that the matching between " + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "content": " in the SR task is relatively easier to learn than denoising under different noise levels, as diffusion models do. Specifically, the capacity of the student network " + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "inline_equation", + "content": "f_{\\hat{\\theta}}" + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "content": " is sufficient to effectively capture the ODE process " + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "content": " using only one step. To verify our assumption, we evaluate the performance of smaller models trained under different strategies. Specifically, one model is trained following the experimental settings of [46] while the number of parameters decreases from 118.6M to 24.3M. Another model uses the same backbone as the aforementioned small model while directly learning the mapping relationship between " + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0}" + }, + { + "bbox": [ + 45, + 293, + 288, + 521 + ], + "type": "text", + "content": " from the standard-size teacher diffusion model. A comparison between these two small models is reported in Table 5. As demonstrated by the results, the model trained for denoising under different noise levels suffers from a serious performance drop compared with the model that directly learns the deterministic mapping between. This strongly supports our assumption that directly learning the deterministic mapping is relatively easier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": "Is a more sophisticated distillation strategy necessary? To explore the necessity of more advanced techniques that learn the mapping between " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": ", we evaluate the performance of Rectified Flow [21], a recent method that learns the mapping to a single step through an iterative manner. Specifically, Reflow operations are conducted to avoid crossing the generation paths, and then followed by distilling the rectified generation process into a single step. However, as shown in Table 6, the involved iterative distillation degrades the performance of the final model due to the accumulated error as discussed by the author [21]. Besides, as verified by the previous section that the deterministic mapping between " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " is easy to learn in the SR task, the benefit of a more sophisticated distillation strategy is not obvious." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "content": "Learned inversion. As the core of the consistency preserving loss, a comparison with the DDIM inversion [37] is" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 64, + 365, + 125 + ], + "blocks": [ + { + "bbox": [ + 307, + 64, + 365, + 125 + ], + "lines": [ + { + "bbox": [ + 307, + 64, + 365, + 125 + ], + "spans": [ + { + "bbox": [ + 307, + 64, + 365, + 125 + ], + "type": "image", + "image_path": "e64930e3a6c67fc81e09d1abbc4353883c264697d883d321374c627d276d59f3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 125, + 353, + 135 + ], + "lines": [ + { + "bbox": [ + 321, + 125, + 353, + 135 + ], + "spans": [ + { + "bbox": [ + 321, + 125, + 353, + 135 + ], + "type": "text", + "content": "LR image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 137, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 160 + ], + "type": "text", + "content": "Figure 7. A comparison between HR images generated from DDIM inversion and the proposed learned inversion. Zoom in for details." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 366, + 64, + 425, + 125 + ], + "blocks": [ + { + "bbox": [ + 366, + 64, + 425, + 125 + ], + "lines": [ + { + "bbox": [ + 366, + 64, + 425, + 125 + ], + "spans": [ + { + "bbox": [ + 366, + 64, + 425, + 125 + ], + "type": "image", + "image_path": "b67a75dccd38b54d6fcf65e83a03a57c4fec1d7dc55e6ba27e38ae23d079e283.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 125, + 422, + 134 + ], + "lines": [ + { + "bbox": [ + 370, + 125, + 422, + 134 + ], + "spans": [ + { + "bbox": [ + 370, + 125, + 422, + 134 + ], + "type": "text", + "content": "DDIM inversion" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 64, + 485, + 125 + ], + "blocks": [ + { + "bbox": [ + 426, + 64, + 485, + 125 + ], + "lines": [ + { + "bbox": [ + 426, + 64, + 485, + 125 + ], + "spans": [ + { + "bbox": [ + 426, + 64, + 485, + 125 + ], + "type": "image", + "image_path": "2a6bc83404031f0ae2370710249703e996ab718559869786b2928fcd77cf838f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 125, + 485, + 134 + ], + "lines": [ + { + "bbox": [ + 428, + 125, + 485, + 134 + ], + "spans": [ + { + "bbox": [ + 428, + 125, + 485, + 134 + ], + "type": "text", + "content": "Learned inversion" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 486, + 64, + 545, + 125 + ], + "blocks": [ + { + "bbox": [ + 486, + 64, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 486, + 64, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 486, + 64, + 545, + 125 + ], + "type": "image", + "image_path": "4b057f75dbc7c2864c2324fab3e34405de8c005aef79fb3950a79ef0f6e39cd3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 125, + 532, + 134 + ], + "lines": [ + { + "bbox": [ + 500, + 125, + 532, + 134 + ], + "spans": [ + { + "bbox": [ + 500, + 125, + 532, + 134 + ], + "type": "text", + "content": "GT image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "type": "text", + "content": "shown in Fig. 7, where the proposed method achieves better fidelity performance. It indicates that the proposed method can obtain a more accurate estimation of " + }, + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 170, + 545, + 229 + ], + "type": "text", + "content": ". Besides, more analyses regarding the consistency preserving loss are in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 229, + 546, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 546, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 546, + 314 + ], + "type": "text", + "content": "Training overhead. While the proposed method involves solving ODEs during training, benefiting from a shortened inference process and initializing the student model from the pre-trained teacher model, the training cost of finetuning using the proposed training paradigm is still lower than that of retraining the diffusion model from scratch. Specifically, the training cost is shown in Table 7." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 321, + 320, + 531, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 238, + 287, + 272 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 272 + ], + "type": "text", + "content": "Table 6. A comparison between models accelerated by the proposed method and [46], which includes a reflow and a distillation operation. The models are evaluated on ImageNet-Test [46]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 320, + 531, + 365 + ], + "lines": [ + { + "bbox": [ + 321, + 320, + 531, + 365 + ], + "spans": [ + { + "bbox": [ + 321, + 320, + 531, + 365 + ], + "type": "table", + "html": "
Num of Iterss/IterTraining Time
ResShift [46]500k1.32s~7.64 days
SinSR (Ours)30k7.41s~2.57 days
", + "image_path": "45a77d975567617c56e373a8ec8eb0295eea3e2dd2b0357561836bc48ab9ca37.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 368, + 542, + 379 + ], + "lines": [ + { + "bbox": [ + 307, + 368, + 542, + 379 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 542, + 379 + ], + "type": "text", + "content": "Table 7. A comparison of the training cost on an NVIDIA A100." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 386, + 378, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 386, + 378, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 386, + 378, + 398 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 406, + 546, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 546, + 563 + ], + "type": "text", + "content": "In this work, we propose a novel strategy to accelerate the diffusion-based SR models into a single inference step. Specifically, a one-step bi-directional distillation is proposed to learn the deterministic mapping between the input noise and the generated high-resolution image and versa vice from a teacher diffusion model with our derived deterministic sampling. Meanwhile, a novel consistency preserving loss is optimized at the same time during the distillation so that the student model not only uses the information from the pre-trained teacher diffusion model but also directly learns from ground-truth images. Experimental results demonstrate that the proposed method can achieve on-par or even better performance than the teacher model in only one step." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 567, + 408, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 408, + 580 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 408, + 580 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 581, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 703 + ], + "type": "text", + "content": "The research is supported in part by the NTU-PKU Joint Research Institute (a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation), National Key R&D Program of China under Grand NO.2022ZD0160100, the National Natural Science Foundation of China under Grant No. 62102150, the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800, No. 22511105800, and the Basic and Frontier Research Project of PCL and the Major Key Project of PCL." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25803" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 791-799, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 158 + ], + "type": "text", + "content": "[2] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 215 + ], + "type": "text", + "content": "[3] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 270 + ], + "type": "text", + "content": "[4] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 337 + ], + "type": "text", + "content": "[5] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12413-12422, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 288, + 373 + ], + "type": "text", + "content": "[6] Ryan Dahl, Mohammad Norouzi, and Jonathon Shlens. Pixel recursive super resolution. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "text", + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "type": "text", + "content": "[9] Baisong Guo, Xiaoyun Zhang, Haoning Wu, Yu Wang, Ya Zhang, and Yan-Feng Wang. Lar-sr: A local autoregressive model for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1909-1918, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 576 + ], + "type": "text", + "content": "[10] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, et al. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. arXiv preprint arXiv:2402.10491, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 666 + ], + "type": "text", + "content": "[12] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In CVPR, pages 466-467, 2020. 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[13] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[14] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "text", + "content": "[15] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 207 + ], + "type": "text", + "content": "[16] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 547, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 283 + ], + "type": "text", + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 546, + 319 + ], + "type": "text", + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, pages 1833-1844, 2021. 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 547, + 365 + ], + "type": "text", + "content": "[19] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 366, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 547, + 409 + ], + "type": "text", + "content": "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 546, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 546, + 454 + ], + "type": "text", + "content": "[21] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2022. 3, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 547, + 510 + ], + "type": "text", + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "type": "text", + "content": "[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 715-732. Springer, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 569, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 547, + 601 + ], + "type": "text", + "content": "[24] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 602, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 547, + 657 + ], + "type": "text", + "content": "[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 712 + ], + "type": "text", + "content": "[26] Jacob Menick and Nal Kalchbrenner. Generating high fidelity images with subscale pixel networks and multidimensional upscaling. International Conference on Learning Representations, International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25804" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[27] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "text", + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 288, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 288, + 231 + ], + "type": "text", + "content": "[29] Aaronvanden Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 233, + 288, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 288, + 287 + ], + "type": "text", + "content": "[30] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. arXiv: Computer Vision and Pattern Recognition, arXiv: Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 289, + 288, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 288, + 355 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 288, + 355 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 358, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 288, + 413 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "text", + "content": "[33] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 472, + 288, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 288, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 288, + 516 + ], + "type": "text", + "content": "[34] Mehdi S. M. Sajjadi, Bernhard Scholkopf, and Michael Hirsch. Enhancenet: Single image super-resolution through automated texture synthesis. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 518, + 288, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 518, + 288, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 518, + 288, + 551 + ], + "type": "text", + "content": "[35] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 553, + 288, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 288, + 608 + ], + "type": "text", + "content": "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 610, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 288, + 643 + ], + "type": "text", + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 645, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 645, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 645, + 288, + 666 + ], + "type": "text", + "content": "[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[39] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 529 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[40] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "text", + "content": "[41] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 228 + ], + "type": "text", + "content": "[42] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 547, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 547, + 284 + ], + "type": "text", + "content": "[43] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 285, + 547, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 547, + 328 + ], + "type": "text", + "content": "[44] Zhaowen Wang, Ding Liu, Jianchao Yang, Wei Han, and ThomasS. Huang. Deep networks for image super-resolution with sparse prior. Cornell University - arXiv, Cornell University - arXiv, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 373 + ], + "type": "text", + "content": "[45] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 375, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 547, + 418 + ], + "type": "text", + "content": "[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 420, + 547, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 547, + 474 + ], + "type": "text", + "content": "[47] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 475, + 547, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 547, + 529 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25805" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_content_list.json b/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a2e00381b732df1e43fdf75eea68e466ebb655e --- /dev/null +++ b/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_content_list.json @@ -0,0 +1,1998 @@ +[ + { + "type": "text", + "text": "Single Domain Generalization for Crowd Counting", + "text_level": 1, + "bbox": [ + 225, + 130, + 743, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuoxuan Peng, S.-H. Gary Chan \nThe Hong Kong University of Science and Technology \n{zpengac, gchan}@cse.ust.hk", + "bbox": [ + 267, + 181, + 700, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to its promising results, density map regression has been widely employed for image-based crowd counting. The approach, however, often suffers from severe performance degradation when tested on data from unseen scenarios, the so-called \"domain shift\" problem. To address the problem, we investigate in this work single domain generalization (SDG) for crowd counting. The existing SDG approaches are mainly for image classification and segmentation, and can hardly be extended to our case due to its regression nature and label ambiguity (i.e., ambiguous pixel-level ground truths). We propose MPCount, a novel effective SDG approach even for narrow source distribution. MPCount stores diverse density values for density map regression and reconstructs domain-invariant features by means of only one memory bank, a content error mask and attention consistency loss. By partitioning the image into grids, it employs patch-wise classification as an auxiliary task to mitigate label ambiguity. Through extensive experiments on different datasets, MPCount is shown to significantly improve counting accuracy compared to the state of the art under diverse scenarios unobserved in the training data characterized by narrow source distribution. Code is available at https://github.com/Shimmer93/MPCount.", + "bbox": [ + 75, + 301, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-based crowd counting is to estimate the number of people (or objects) in an image. The recent mainstream counting approach is density map regression, where a density map with continuous pixel value is predicted from the input image, and the total crowd count is obtained by summing all these pixel values. During training, the ground-truth density map is generated by smoothing point annotations given by the locations of human heads. Density map regression has been demonstrated to achieve encouraging results in crowd counting.", + "bbox": [ + 75, + 703, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing training models predominantly assume that the testing data share the same source distribution, i.e., fully supervised learning. Such assumption, however, often breaks", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29a0596334f283546466ccc996ea121ca039dfdefdffa00c40826bc036a444b6.jpg", + "image_caption": [ + "(a) Source domain" + ], + "image_footnote": [], + "bbox": [ + 506, + 267, + 633, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e1c2b5b41a74ccea9a425f38d9de2b82ffb2bea1af72a4e8560dc10fa6e0207a.jpg", + "image_caption": [ + "gain" + ], + "image_footnote": [], + "bbox": [ + 633, + 268, + 756, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba96caa0eec5e34538c24c910bff308c00da846361b10ea634a546f5f975ea88.jpg", + "image_caption": [ + "(b) Target image" + ], + "image_footnote": [], + "bbox": [ + 759, + 268, + 885, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/640bf12c0e782edc061bb92537af332efa631d01afdf37f90fde7c891cd21219.jpg", + "image_caption": [ + "(c) Ground-truth" + ], + "image_footnote": [], + "bbox": [ + 506, + 339, + 630, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb179390b3f8de5891ec7a74420c5232187dbbe1a59a06b74312c31857fd38c4.jpg", + "image_caption": [ + "(d) DCCUS" + ], + "image_footnote": [], + "bbox": [ + 633, + 339, + 758, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8f1c619400494e6ac391964186cf6a6f5fe7102e55efd18563be023d1d99d02.jpg", + "image_caption": [ + "(e) MPCount (Ours)", + "Figure 1. Single domain generalization for crowd counting. (a) Sample images from a single source domain $S$ with narrow distribution. (b) A challenging target image from an unseen scenario. (c) The ground-truth density map. (d) The predicted density map from the previous work DCCUS [6] trained on $S$ . (e) The predicted density map from our MPCount trained on $S$ . MPCount achieves much lower counting error than DCCUS and makes better predictions in the cropped area against the domain shift." + ], + "image_footnote": [], + "bbox": [ + 759, + 339, + 885, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in practice due to changes in camera position, weather condition, etc. Such deviation from the training data (i.e. source domain) to unseen testing data (i.e. target domain) is known as domain shift, which severely degrades the performance of deep learning models in real-world deployment.", + "bbox": [ + 496, + 561, + 890, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate the domain shift problem, domain adaptation (DA) approaches have been proposed to transfer knowledge from the source domain to the target one, typically by finetuning pre-trained models with target domain data [2, 10, 39, 44]. Despite promising results, the fine-tuning process of DA is usually laborious. Furthermore, target domain data may not be easily available in reality.", + "bbox": [ + 496, + 638, + 890, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to DA, domain generalization (DG) aims to generalize models to any unobserved target domain without the need for target data. As data and model fine-tuning are not necessary at the target domain, DG is more convenient to be deployed in unseen or unpredictable domain. We consider in this work the challenging case of single domain generalization (SDG) for image-based crowd counting, where only one source domain, likely of narrow distribution (Fig. 1a), is used for training.", + "bbox": [ + 496, + 747, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A common general SDG technique is to construct", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28025", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "domain-invariant features from different augmented versions of an image. Such approach has been shown promising for various tasks [5, 13, 34]. Independent of data augmentation, many other SDG approaches have been proposed for classification and segmentation to further boost performance. Unfortunately, most of them cannot be extended to crowd counting, a regression task without category information. Moreover, while labels in classification and segmentation tasks are precise and domain-invariant, the labels for density-based crowd counting are usually ambiguous. Such \"label ambiguity\" stems from the fact that the density values are generated from point annotations, and hence human heads and backgrounds may be assigned to similar value under different scenarios.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent work of [6] achieves DG for crowd counting based on dividing the source into multiple sub-domains. While encouraging, it requires a rather broad source domain distribution. We propose, for the first time, an SDG approach for crowd counting called MPCount without sub-domain division, thereof effective and extensible to narrow source distribution. We highlight its performance as compared with a state-of-the-art scheme in Fig. 1. MPCount achieves its superior performance with the following two novel components:", + "bbox": [ + 75, + 305, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An attention memory bank (AMB) to tackle density regression: MPCount employs only one attention memory bank (AMB) that takes a pair of features encoded from an crowd image and its data-augmented version as input. To cover continuous density values for regression with a finite memory size, AMB reconstructs each feature vector as an attention over memory vectors. These memory vectors learn domain-invariant representations from the pair of features of different styles but similar content (i.e. crowd densities). We present the content error mask (CEM) to eliminate domain-related content from input features, as characterized by a large discrepancy between instance-normalized feature element pairs. Furthermore, the newly employed attention consistency loss (ACL) enforces the similarity of attention scores produced from input features, ensuring the consistency of memory vectors.", + "- Patch-wise classification $(PC)$ to tackle label ambiguity: To address the label ambiguity, MPCount employs a novel auxiliary task called patch-wise classification (PC). In this task, each crowd image is evenly divided into fixed-size square patches, say $16 \\times 16$ , classified into two classes, i.e., containing human heads or not. The density values in the areas classified as without head are filtered out during density regression, so that the pixel-level ambiguity is overcome by coarser but more accurate patchwise binary labels." + ], + "bbox": [ + 76, + 459, + 472, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate MPCount, we conduct extensive experiments on various crowd counting datasets including ShanghaiTech A (SHA) & B (SHB) and JHU-Crowd++. We introduce", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a new challenging setting of narrow source distribution, where only images of the same category, such as \"snow\" (SN) and \"fog/haze\" (FH) in JHU-Crowd++, are used in training. We demonstrate that MPCount achieves excellent performance not only on traditional inter-dataset benchmarks, but also under our newly introduced narrow source setting. Let $S \\rightarrow T$ denote the case where the source domain is $S$ and the target domain is $T$ . As compared to the state of the art, MPCount is shown to reduce significantly counting error by $21.8\\%$ on SN $\\rightarrow$ FH, $18.6\\%$ on FH $\\rightarrow$ SN, $18.2\\%$ on SHB $\\rightarrow$ SHA and $9.5\\%$ on SHA $\\rightarrow$ SHB.", + "bbox": [ + 496, + 90, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 498, + 271, + 650, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Fully Supervised Crowd Counting", + "text_level": 1, + "bbox": [ + 500, + 296, + 799, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The mainstream approach to crowd counting is via density map regression, where each pixel of an image is assigned a count value, and the summation of these values yields an estimate of the total number of people. Most methods focus on improving counting accuracy under the fully supervised setting by employing novel network designs [1, 16, 17, 31, 45] or loss functions [20, 26, 36, 37]. While encouraging results have been shown, their performance usually drop significantly when evaluated on out-of-distribution data, due to their limited generalization ability. Besides, several works [4, 22, 38] have proposed auxiliary tasks in crowd counting for certain purposes. However, these tasks still require pixel-level predictions and contribute little to tackling label ambiguity.", + "bbox": [ + 496, + 320, + 893, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Domain Adaptation (DA) for Crowd Counting", + "text_level": 1, + "bbox": [ + 498, + 541, + 890, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain adaptation (DA) has been widely studied in crowd counting to address domain shift by adapting source domain information to a particular target domain. While a few methods study supervised DA where labeled data from target domain are available [40, 44], most approaches tackle unsupervised DA, utilizing only unlabeled target domain data [2, 7, 18, 39, 41, 47]. Despite the notable advancements, DA methods normally require data from the target domain, which may not be easily available in practice.", + "bbox": [ + 496, + 565, + 893, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Single Domain Generalization (SDG)", + "text_level": 1, + "bbox": [ + 498, + 710, + 821, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Generalization (DG) approaches are designed to train a network with generalization ability solely utilizing source domain data, and single domain generalization (SDG) is the special case when only one source domain is available. There are various techniques for SDG, including 1) adversarial data generation [15, 25, 46, 50], 2) feature normalization/whitening transformation [5, 12, 23, 24, 43] and 3) domain-general network design, such as a convolution layer [35], vision transformer [32] and memory bank [3]. Among various SDG methods, a common general technique is to simulate domain shift with data augmenta", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28026", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion and extract domain-invariant information from different augmented versions of features [5, 13, 43]. While SDG has attracted widespread attention in classification and segmentation, it is still a nascent topic in crowd counting due to its regression nature and label ambiguity.", + "bbox": [ + 75, + 90, + 467, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DCCUS [6] studies SDG for crowd counting. In this method, the source domain is dynamically clustered into sub-domains simulated as meta-training and -test sets in a meta-learning strategy. Two types of memory modules and several losses are proposed to distinguish and record domain-invariant and -related information. DCCUS designs the memory mechanism specifically for regression, but the problem of ambiguous labels is not considered. Moreover, when the source domain follows a narrow distribution, the sub-domain division process may experience reduced effectiveness and affect the generalization ability of the counting model. MPCount, without sub-domain division, employs novel patch-wise classification to address the label ambiguity problem. Additionally, a single memory bank for regression is introduced without the need for sub-domain partitioning, enabling MPCount to generalize well even in source domains of narrow distributions.", + "bbox": [ + 75, + 167, + 468, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. MPCount", + "text_level": 1, + "bbox": [ + 76, + 436, + 184, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides a detailed description of our MPCount scheme. We begin by reviewing the foundational concepts of crowd counting in Sec. 3.1. Next, in Sec. 3.2, we present a comprehensive overview of the entire scheme. We then elaborate the attention memory bank (Sec. 3.3) designed for density regression, followed by the content error mask (Sec. 3.4) and attention consistency loss (Sec. 3.5). After that, we discuss the patch-wise classification (Sec. 3.6) proposed to address the challenge of ambiguous labels. Finally, we detail the overall training loss in Sec. 3.7.", + "bbox": [ + 75, + 460, + 467, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Crowd Counting Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 621, + 349, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we review the preliminary knowledge of deep learning-based crowd counting via density map regression. Our objective is to train a neural network $\\mathcal{N}$ with parameter $\\theta$ that takes in an input image $\\mathcal{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ and outputs a density map $\\hat{\\mathcal{D}} = \\mathcal{N}(\\mathcal{I};\\theta)$ of size $H \\times W$ . The estimated count $\\hat{c}$ is the sum of all pixel-wise density values in $\\hat{\\mathcal{D}}$ , i.e.,", + "bbox": [ + 75, + 645, + 467, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} = \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\hat {\\mathcal {D}} _ {i j}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 757, + 467, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generation of ground-truth density maps usually follows the method in [28]. Given the human head point annotations $\\mathcal{H}$ where each $h\\in \\mathcal{H}$ is the coordinate of a head location, the ground-truth density map $\\mathcal{D}\\in \\mathbb{R}^{H\\times W}$ is calculated as the sum of the 2D Gaussian filter applied to each head coordinate $h$ , i.e.,", + "bbox": [ + 75, + 809, + 467, + 898 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = \\sum_ {h \\in \\mathcal {H}} \\delta (x - h) \\times G _ {\\sigma} (x), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 108, + 890, + 141 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta (\\cdot)$ is the discrete delta function, and $G_{\\sigma}(\\cdot)$ is the Gaussian kernel with a fixed variance $\\sigma$", + "bbox": [ + 498, + 146, + 890, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A common objective function to supervise the network $\\mathcal{N}$ can be written as", + "bbox": [ + 498, + 176, + 890, + 205 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} (\\theta) = | | \\mathcal {D}, \\hat {\\mathcal {D}} | | _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 226, + 890, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Scheme Overview", + "text_level": 1, + "bbox": [ + 500, + 250, + 671, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall structure of our method is illustrated in Fig. 2. During training, photometric transformations are applied to the original image $\\mathcal{I}^{ori}$ to produce an augmented version $\\mathcal{I}^{aug}$ . The feature extractor encodes them into features $\\mathcal{F}^{ori}$ and $\\mathcal{F}^{aug}$ , and a content error mask (CEM) $\\mathcal{M}$ is applied to them. The masked features are reconstructed into $\\tilde{\\mathcal{F}}^{ori}$ and $\\tilde{\\mathcal{F}}^{aug}$ via an attention memory bank (AMB) and passed to the density head for crowd density prediction. Meanwhile, the highest-level features encoded by the feature extractor, $\\mathcal{Z}^{ori}$ and $\\mathcal{Z}^{aug}$ , are fed into the patch-wise classification (PC) head, and the predicted PC maps (PCMs) $\\hat{\\mathcal{C}}^{ori}$ and $\\hat{\\mathcal{C}}^{aug}$ are binarized and resized to serve as a mask to filter out areas without human heads in the estimated density maps $\\hat{\\mathcal{D}}^{ori}$ and $\\hat{\\mathcal{D}}^{aug}$ . The final predictions $\\mathcal{D}'^{ori}$ and $\\mathcal{D}'^{aug}$ are two density maps with certain areas masked out by the PCMs. The entire training process is supervised under a combination of density losses, PC losses and the attention consistency loss (ACL). The inference process solely employs $\\mathcal{I}^{ori}$ as input, and outputs a single density map $\\mathcal{D}'^{ori}$ .", + "bbox": [ + 498, + 272, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Attention Memory Bank (AMB)", + "text_level": 1, + "bbox": [ + 500, + 566, + 782, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attention memory bank (AMB) is designed to automatically learn domain-invariant features for density regression. [14] proposes a memory bank where memory vectors are updated individually, each corresponding to a certain category. However, this design cannot be applied to regression tasks, since a finite number of memory vectors cannot cover continuous density values. Inspired by [6], we reconstruct each feature vector as an attention over memory vectors, so that any density value may be represented as a linear combination of representative values in the memory bank.", + "bbox": [ + 498, + 589, + 890, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The AMB $\\mathcal{V} \\in \\mathbb{R}^{M \\times C}$ consists of $M$ memory vectors of dimension $C$ . Given a flattened input feature map $\\mathcal{F} \\in \\mathbb{R}^{HW \\times C}$ , we first compute the attention scores $\\mathcal{A}$ between the feature $\\mathcal{F}$ (query) and the memory $\\mathcal{V}$ (key). Then the reconstructed feature map $\\tilde{\\mathcal{F}}$ is calculated as a linear combination of $\\mathcal{V}$ (value) with $\\mathcal{A}$ as weights. The reconstruction process can be summarized as follows:", + "bbox": [ + 498, + 739, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (\\mathcal {F}, \\mathcal {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {F V} ^ {T}}{\\sqrt {C}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 863, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28027", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/18074bfdcd490ed255445b4ed43ad1a70629440fb2752c5b0bd7c4fb864b3ea3.jpg", + "image_caption": [ + "Figure 2. The overall training pipeline of our proposed MPCount. All identical modules in this diagram share the same weights. In the encoder-decoder structure, a higher level indicates a deeper feature." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 893, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 76, + 398, + 106, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {F}} = \\mathcal {A} \\mathcal {V}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 425, + 468, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While [6] relies on multiple memory banks corresponding to different sub-domains to distinguish domain-invariant and -related information, we only employ a single AMB and propose the novel CEM (Sec. 3.4) and ACL (Sec. 3.5) to ensure that it stores domain-invariant representations. To this end, our feature reconstruction mechanism can work without sub-domain division, thereof is effective even with a narrow source distribution.", + "bbox": [ + 75, + 457, + 468, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Content Error Mask (CEM)", + "text_level": 1, + "bbox": [ + 76, + 589, + 330, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CEM is proposed to guarantee the resemblance of content contained in the pair of input features by excluding possible domain-related content information. A common technique to disentangle style and content in a feature is instance normalization (IN) [33], where the feature statistics (channel-wise mean and variance) are considered to contain style information, while the instance normalized features retain content information. Based on this idea, we assume that the discrepancy of instance normalized features reflect effects of domain shift on feature content and may mislead AMB to learn domain-related information. Therefore, we filter out elements in the input features with difference between their instance normalized versions above a certain threshold, resulting in a pair of features with similar content information for domain-invariant feature reconstruction.", + "bbox": [ + 75, + 612, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, given the features extracted from the original and augmented images, $\\mathcal{F}^{ori}$ and $\\mathcal{F}^{aug} \\in \\mathbb{R}^{H \\times W \\times C}$ , we define the content error mask $\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times C}$ as follows:", + "bbox": [ + 76, + 839, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {i j k} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | \\mathrm {I N} (\\mathcal {F} ^ {\\text {o r i}}) _ {i j k} - \\mathrm {I N} (\\mathcal {F} ^ {\\text {a u g}}) _ {i j k} | \\leq \\alpha \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 419, + 890, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is the threshold indicating whether the error value reflects possible inconsistency of content information caused by domain shift.", + "bbox": [ + 496, + 465, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $\\mathcal{F}$ to be one of $\\mathcal{F}^{\\text{ori}}$ and $\\mathcal{F}^{\\text{aug}}$ , we filter out the domain-related content information by", + "bbox": [ + 498, + 511, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {\\prime} = \\operatorname {D r o p o u t 2 D} (\\mathcal {F} \\odot \\mathcal {M}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 565, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The random 2D dropout is applied to $\\mathcal{F}'$ to prevent the memory from relying solely on certain channels.", + "bbox": [ + 498, + 589, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Attention Consistency Loss (ACL)", + "text_level": 1, + "bbox": [ + 498, + 627, + 799, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The style information of $\\mathcal{F}^{ori}$ and $\\mathcal{F}^{aug}$ preserved in Sec. 3.4 may produce a large discrepancy on their attention distributions over the memory vectors in AMB even if they contain identical content information. To ensure that each memory vector consistently store specific domain-invariant representations, we propose the attention consistency loss (ACL) to enforce the similarity of attention scores produced from $\\mathcal{F}^{ori}$ and $\\mathcal{F}^{aug}$ .", + "bbox": [ + 496, + 650, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider $\\mathcal{A}$ in Eq. 4 as a distribution, and the ACL $\\mathbb{L}_{con}$ is calculated as the distance between the distribution $\\mathcal{A}^{ori}$ produced by $\\mathcal{F}^{ori}$ and $\\mathcal{A}^{aug}$ produced by $\\mathcal{F}^{aug}$ . Here the simple Euclidean distance is selected as the distance measure due to its computational stability during training:", + "bbox": [ + 496, + 772, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} _ {\\text {c o n}} = \\left\\| \\mathcal {A} ^ {\\text {a u g}}, \\mathcal {A} ^ {\\text {o r i}} \\right\\| _ {2} ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 869, + 890, + 888 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28028", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6d595b4e7315ec7bb19574ecafa9840a473d3ec2e87411a3217473df3fec3771.jpg", + "image_caption": [ + "(a) Density Map" + ], + "image_footnote": [], + "bbox": [ + 153, + 88, + 271, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc336287be96e075878223a874b3f55febe4f4805756b5982f9e8f8f33a487d3.jpg", + "image_caption": [ + "(b) PCM", + "Figure 3. Illustration of label ambiguity and how PCM tackles it. Point A is a typical example which belongs to a human head but is assigned the density value 0 as the background in the density map due to varying head sizes. In the PCM, the patch covering A is correctly classified as containing heads, thanks to the patch-level accuracy of the labels." + ], + "image_footnote": [], + "bbox": [ + 274, + 88, + 393, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Patch-wise Classification (PC)", + "text_level": 1, + "bbox": [ + 76, + 290, + 341, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.1, ground-truth density maps are usually calculated only based on point annotations. In different scenarios, pixels of a human head and the background might be assigned the same density value due to various head sizes (Fig. 3a), violating the common assumption that labels are accurate and domain-invariant, thereof resulting in the ambiguous labels in crowd counting.", + "bbox": [ + 75, + 313, + 468, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To tackle the label ambiguity, we propose a novel auxiliary task named patch-wise classification (PC) and its corresponding supervision signal, patch-wise classification map (PCM), as illustrated in Fig. 3b. A PCM divides an image evenly into $P \\times P$ ( $P = 16$ empirically) patches and classifies each patch into two categories: containing human heads or not. Such patch-level predictions mitigate the uncertainty in pixel-level density maps by providing coarser but more accurate patch-level information.", + "bbox": [ + 75, + 419, + 468, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, a ground truth $\\mathbf{PCM}\\mathcal{C}$ with patch size $P$ can be calculated as:", + "bbox": [ + 75, + 555, + 468, + 584 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i j} = 1 - \\delta \\left(\\sum_ {k = i P} ^ {(i + 1) P} \\sum_ {l = j P} ^ {(j + 1) P} \\mathcal {D} _ {k l}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 589, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is the ground truth density map and $\\delta (\\cdot)$ is the discrete delta function.", + "bbox": [ + 75, + 643, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model predicts a PCM $\\hat{C}$ supervised by the binary cross entropy (BCE) loss:", + "bbox": [ + 75, + 674, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} _ {c l s} = \\operatorname {B C E} (\\mathcal {C}, \\hat {\\mathcal {C}}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 710, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, $\\hat{\\mathcal{C}}$ is binarized and resized to obtain $\\mathcal{C}'$ which matches the dimensions of the predicted density map $\\hat{\\mathcal{D}}$ . $\\mathcal{C}'$ is then used to mask the regions classified without crowds in $\\hat{\\mathcal{D}}$ , resulting in the final density map $\\mathcal{D}'$ :", + "bbox": [ + 76, + 738, + 468, + 800 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} ^ {\\prime} = \\hat {\\mathcal {D}} \\odot \\mathcal {C} ^ {\\prime}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 806, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the density regression head is supervised with the ground truth density map $\\mathcal{D}$ , and the common Euclidean distance in Eq. (3) is used as the objective function:", + "bbox": [ + 75, + 830, + 468, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} _ {d e n} = \\left\\| \\mathcal {D}, \\mathcal {D} ^ {\\prime} \\right\\| _ {2} ^ {2}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.7. Model Training", + "text_level": 1, + "bbox": [ + 500, + 90, + 656, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall training loss is a combination of density losses $\\mathbb{L}_{den}^{ori},\\mathbb{L}_{den}^{aug}$ , PC losses $\\mathbb{L}_{cls}^{ori},\\mathbb{L}_{cls}^{aug}$ and the attention consistency loss $\\mathbb{L}_{con}$ , which can be written as", + "bbox": [ + 500, + 113, + 890, + 159 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} = \\mathbb {L} _ {d e n} ^ {o r i} + \\mathbb {L} _ {d e n} ^ {a u g} + \\lambda_ {c l s} \\left(\\mathbb {L} _ {c l s} ^ {o r i} + \\mathbb {L} _ {c l s} ^ {a u g}\\right) + \\lambda_ {c o n} \\mathbb {L} _ {c o n}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 167, + 890, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{cls}$ and $\\lambda_{con}$ are weighting parameters to balance different loss terms.", + "bbox": [ + 498, + 208, + 890, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Illustrative Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 250, + 802, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the datasets used in experiments in Sec. 4.1. Next, the implementation details are specified in Sec. 4.2, then evaluation metrics are described in Sec. 4.3. We present the results compared with state of the art in Sec. 4.4. Finally, we conduct ablation studies and other additional analysis in Sec. 4.5.", + "bbox": [ + 496, + 276, + 890, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 375, + 601, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on four mainstream crowd counting datasets: ShanghaiTech Part A & B, UCF-QNRF and JHU-Crowd++.", + "bbox": [ + 496, + 398, + 890, + 443 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ShanghaiTech [45] is composed of two parts, SHA (A) and SHB (B). SHA contains 300 training images and 182 testing images, while SHB contains 400 training images and 316 testing images. Images in SHA are collected from the Internet and feature highly crowded scenes. In contrast, SHB is captured from several streets in Shanghai, and its images exhibit crowd densities generally lower than those in SHA.", + "- UCF-QNRF (Q) [11] consists of 1201 training images and 334 testing images. It is a challenging dataset with a large range of crowd densities, scenes, viewpoints and lighting conditions.", + "- JHU-Crowd++ [29] is a large-scale dataset containing 4372 images, with 2722, 500 and 1600 images in the training, validation and testing set, respectively. In addition, image-level labels are also provided in this dataset, including 16 types of scenes and 4 types of weather conditions. For scene annotations, there are 879 images labeled \"stadium\" (SD) and 573 labeled \"street\" (ST), which are utilized as two source domains. For weather annotations, we use 201 images labeled \"snow\" (SN) and 168 images labeled \"fog/haze\" (FH) as source domains. In each domain, $80\\%$ of the images are selected as the training set while the rest $20\\%$ are for testing. Data with the same label belong to a narrower distribution than mainstream datasets, thus are more challenging for SDG. Illustrative samples in these scene-specific and weather-specific domains are available in Fig. 4.", + "Let $S \\to \\mathcal{T}$ denote the case where $S$ is the source domain and $\\mathcal{T}$ is the target domain. The datasets are utilized" + ], + "bbox": [ + 500, + 445, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28029", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ea710ba9842e1382414ebbe96cfe63aa38f448d0c1f3015e982e1a7aef8150b.jpg", + "image_caption": [ + "(a) Stadium" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 176, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02541f207d512fd99e99c6ccf724a287a7b6d932b25ac5d02e65341963169652.jpg", + "image_caption": [ + "(b) Street" + ], + "image_footnote": [], + "bbox": [ + 178, + 88, + 272, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64084dfcd33edd730f2bf68d2bfcb768ba3b9460ca2a401766734fdf0f020e46.jpg", + "image_caption": [ + "(c) Snow", + "Figure 4. Sample images with different instance-level labels in JHU-Crowd++." + ], + "image_footnote": [], + "bbox": [ + 274, + 88, + 370, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88f90215da608759fe740d88269e5fc4c60414cffb34bad4c1f255229c093065.jpg", + "image_caption": [ + "(d) Fog/Haze" + ], + "image_footnote": [], + "bbox": [ + 372, + 88, + 467, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in experiments under two distinct types of settings: 1) An entire dataset is regarded as a single domain: $\\mathrm{A} \\rightarrow \\mathrm{B} / \\mathrm{Q}$ , $\\mathrm{B} \\rightarrow \\mathrm{A} / \\mathrm{Q}$ and $\\mathrm{Q} \\rightarrow \\mathrm{A} / \\mathrm{B}$ ; 2) A dataset is partitioned into subsets according to image-level labels, and one subset constitutes a single domain: $\\mathrm{SD} \\leftrightarrow \\mathrm{SR}$ and $\\mathrm{SN} \\leftrightarrow \\mathrm{FH}$ .", + "bbox": [ + 76, + 292, + 468, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation", + "text_level": 1, + "bbox": [ + 76, + 376, + 235, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt VGG16-BN [27] as the feature extractor in our model. The density regression head is a single convolution layer with $1 \\times 1$ filters, while the classification head consists of a $3 \\times 3$ and a $1 \\times 1$ convolution layers. Batch normalization is employed after all convolution layers except the last ones. For data augmentation, we apply three types of photometric transformations, color jittering, Gaussian blurring and sharpening. We also randomly crop image patches with a size of $320 \\times 320$ and adopt random horizontal flipping to both the original and augmented images simultaneously. We select AdamW [19] as the optimizer and OneCycleLR [30] as the learning rate scheduler with maximum learning rate set to 1e-3 and a maximum epoch of 300. The memory size $M$ is set to 1024 and the dimension $C$ is 256. We use a content error threshold $\\alpha$ of 0.5, and the loss weights $\\lambda_{cls}$ and $\\lambda_{con}$ are both set to 10.", + "bbox": [ + 75, + 398, + 468, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 648, + 261, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method with mean absolute error (MAE) and mean squared error (MSE), defined as follows:", + "bbox": [ + 76, + 672, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM A E = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| c _ {i} - \\hat {c} _ {i} \\right|, M S E = \\sqrt {\\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(c _ {i} - \\hat {c} _ {i}\\right) ^ {2}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 722, + 468, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ is the number of testing images, $c_{i}$ is the ground truth count of the $i$ -th image and $\\hat{c}_i$ is the predicted count. Lower values of both metrics indicate better performance.", + "bbox": [ + 76, + 792, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with State of the Art", + "text_level": 1, + "bbox": [ + 76, + 847, + 369, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we compare our MPCount with state-of-the-art methods on different benchmarks, as illustrated", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1c9c23b03cc636a03c755ed44e872d83be0deeff9cd760032c06f20f1286d61.jpg", + "image_caption": [ + "(a) Image", + "Figure 5. Visualization results of DCCUS [6] and MPCount under different DG settings. First row: A $\\rightarrow$ B; second row: SN $\\rightarrow$ FH; third row: B $\\rightarrow$ A." + ], + "image_footnote": [], + "bbox": [ + 503, + 88, + 598, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b17371f220a1d4f8b2d07c684197dd6e847e7f6446d45b9fc2c615fc9b1ace6a.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 601, + 88, + 694, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0a754df6ec173baa4b9906bab7d18b6614c99bf9bbfc8b548b6554199f1f9df.jpg", + "image_caption": [ + "(c) DCCUS" + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 792, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e05d7b6691c0a3e221adb9d4fb144886576bb6b67ff9683159e44a8e6088e98b.jpg", + "image_caption": [ + "(d) MPCount" + ], + "image_footnote": [], + "bbox": [ + 794, + 88, + 888, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Fig. 5. Tab. 1 shows the results on $\\mathrm{A} \\rightarrow \\mathrm{B} / \\mathrm{Q}$ , $\\mathrm{B} \\rightarrow \\mathrm{A} / \\mathrm{Q}$ and $\\mathrm{Q} \\rightarrow \\mathrm{A} / \\mathrm{B}$ . The selected baselines can be divided into three categories:", + "bbox": [ + 498, + 344, + 890, + 388 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Fully supervised crowd counting: BL [20], DM-Count [37], SASNet [31], ChfL [26] and MAN [17].", + "- Domain adaptation for crowd counting: RBT [18], C²MoT [41], FGFD [47], DAOT [48] and FSIM [49].", + "- Single domain generalization: IBN [23], SW [24], ISW [5], DG-MAN [21] and DCCUS [6]." + ], + "bbox": [ + 500, + 396, + 890, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All results are copied from previous papers except IBN, SW and ISW, which are feature normalization/whitening-based SDG methods originally designed for classification or segmentation and adapted to crowd counting by us. Domain adaptation methods are listed only for reference, as data from target domains are usually visible during adaptation.", + "bbox": [ + 498, + 493, + 890, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our MPCount outperforms all the DG methods under most of these settings, including an significant error reduction by $18.2\\%$ on $\\mathrm{B} \\rightarrow \\mathrm{A}$ . It is notable that MPCount exhibits outstanding performance in some cases even compared to DA methods, providing compelling evidence for the effectiveness of our design.", + "bbox": [ + 498, + 590, + 890, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we further conduct experiments on scene-specific domains $\\mathrm{SD} \\leftrightarrow \\mathrm{SR}$ and weather-specific domains $\\mathrm{SN} \\leftrightarrow \\mathrm{FH}$ . BL [20], MAN [17], DAOT [48] and DCCUS [6] are trained using their released code, while IBN [23], SW [24] and ISW [5] are adapted by ourselves. We observe that the normalization/whitening-based methods cannot achieve satisfactory results, possibly because useful information is eliminated from feature statistics. DCCUS also obtains worse results than fully supervised methods, reflecting that the clustering procedure may fail to produce meaningful sub-domains when the source domain distribution is narrow. In contrast, our MPCount still performs the best among all tested methods, demonstrating its superiority in challenging conditions with narrow domain distributions.", + "bbox": [ + 498, + 689, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28030", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d7755ba285d7ece6d588a693962c40c8fafeba30f5d99f3b4cf41ffe4ee4024.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Source → TargetA → BA → QB → AB → QQ → AQ → B
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]✘✘15.925.8166.7287.6138.1228.1226.4411.0----
DMCount [37]✘✘23.134.9134.4252.1143.9239.6203.0386.173.4135.114.327.5
SASNet [31]✘✘21.333.2211.2418.6132.4225.6273.5481.373.9116.413.022.1
ChfL [26]✘✘18.729.1122.3218.0121.3200.8197.1357.968.7118.514.027.0
MAN [17]✘✘22.132.8138.8266.3133.6255.6209.4378.867.1122.112.522.2
RBT [18]✓✘13.429.3175.0294.8112.2218.2211.3381.9----
C2MoT [41]✓✘12.421.1125.7218.3120.7192.0198.9368.0----
FGFD [47]✓✘12.723.3124.1242.0123.5210.7209.7384.770.2118.412.520.3
DAOT [48]✓✘10.918.8113.9215.6----67.0128.411.319.6
FSIM [49]✓✘11.119.3105.3191.1120.3202.6194.9324.566.8111.511.019.7
IBN [23]✘✓19.130.8280.2561.0125.9202.3183.5317.4105.9174.616.025.5
SW [24]✘✓20.230.4285.5431.0126.7193.8200.7333.2102.4168.819.032.9
ISW [5]✘✓23.937.7215.6399.7156.2291.5263.1442.283.4136.022.134.7
DG-MAN [21]✘✓17.328.7129.1238.2130.7225.1182.4325.8----
DCCUS [6]✘✓12.624.6119.4216.6121.8203.1179.1316.267.4112.812.120.9
MPCount (Ours)✘✓11.419.7115.7199.899.6182.9165.6290.465.5110.112.324.1
", + "bbox": [ + 81, + 89, + 893, + 368 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4574df905a4773786412ac5fd334ae9d01772e1a64ae00356b8058679d394c2f.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art methods on SHA (A), SHB (B) and UCF-QNRF (Q). Domain adaptation methods are only for reference since data in target domain might be used." + ], + "table_footnote": [], + "table_body": "
Source → TargetSD → SRSR → SDSN → FHFH → SN
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]XX42.179.0262.71,063.948.1129.5343.8770.5
MAN [17]XX45.179.0246.1950.838.168.0445.0979.3
DAOT [48]✓X45.388.0278.71,624.342.373.0151.6273.9
IBN [23]X✓92.2178.0318.11,420.4109.7267.7491.81,110.4
SW [24]X✓110.3202.4312.61,072.4131.5306.6381.3825.0
ISW [5]X✓108.1212.4385.91,464.8151.6365.7276.6439.8
DCCUS [6]X✓90.4194.1258.11,005.954.5125.8399.7945.0
MPCount (Ours)X✓37.470.1218.6935.931.355.0216.3421.4
", + "bbox": [ + 156, + 422, + 815, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with the state of the art on data in JHU-Crowd++ with labels \"Stadium\"(SD), \"Street\"(SR), \"Snow\"(SN) and \"Fog/Haze\"(FH). Domain adaptation methods are only for reference since data in target domain might be used.", + "bbox": [ + 75, + 613, + 892, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 667, + 240, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies and other analysis under the setting $\\mathrm{A} \\to \\mathrm{B} / \\mathrm{Q}$ .", + "bbox": [ + 75, + 694, + 468, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of model components. We start from a simple encoder-decoder structured baseline model and add or remove each component individually to validate its effectiveness. The experimental results are displayed in Tab. 3.", + "bbox": [ + 75, + 726, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Attention Memory Bank: After adding AMB, the performance consistently improves on both target domains, as shown in Tab. 3. This validates that AMB can effectively help the model generalize by reconstructing domain-invariant features for density regression.", + "- Content Error Mask & Attention Consistency Loss: Next, we test the effects of CEM and ACL on AMB. The results" + ], + "bbox": [ + 76, + 794, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Tab. 3 show that both components generally improves the performance, especially when PC is present.", + "bbox": [ + 509, + 667, + 888, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Patch-wise Classification: Finally, models with PC outperforms their counterparts without PC, as visible in Tab. 3. This suggests that such an auxiliary task is useful for DG by compensating for ambiguous pixel-level labels with accurate patch-level labels.", + "bbox": [ + 500, + 699, + 890, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of $\\alpha$ . The threshold of CEM $\\alpha$ mentioned in Eq. (6) controls the Portion of Diminished Elements (PDE) in features. According to the results in Tab. 4, $\\alpha = 0.5$ produces the best performance under all settings, with a PDE of $10.5\\%$ . It is also indicated that using a too large value of $\\alpha$ (0.9) may be inadequate to filter out domain-related information, while a too small value (0.1) would let useful information eliminated.", + "bbox": [ + 496, + 779, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28031", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/edc58b2cfe53d6f985f743c689e5655874a9208d7a5cbf3adcc7d5b1671ec5e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ComponentsA → BA → Q
AMBCEMACLPCMAEMSEMAEMSE
21.639.1200.9325.5
✓18.932.0176.3292.8
✓✓15.325.5156.8232.7
✓✓16.928.7172.5281.6
✓✓✓14.124.3170.6286.8
✓14.430.5144.2262.6
✓✓13.425.2133.4224.0
✓✓✓13.323.4126.9222.3
✓✓✓✓11.419.7115.7199.8
", + "bbox": [ + 78, + 89, + 472, + 291 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7783c255780f1add9655b5c17dffab792b40d2c529422979fe817afbc24c7055.jpg", + "table_caption": [ + "Table 3. Ablation studies on the model components." + ], + "table_footnote": [], + "table_body": "
ParameterPropertyA → BA → Q
αPDE(%)MAEMSEMAEMSE
0.128.616.725.9157.3338.5
0.314.612.420.8132.1223.6
0.510.511.419.7115.7199.8
0.75.911.923.8123.8214.1
0.93.613.827.2124.7209.6
", + "bbox": [ + 80, + 332, + 467, + 467 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ff3ac72241aff4eb1f026bf50c0c3d917faa7978d50c2c0e248a6b79b4ef2d3.jpg", + "table_caption": [ + "Table 4. Ablation studies on the error threshold $\\alpha$ in Eq. (6)." + ], + "table_footnote": [], + "table_body": "
ParametersA → BA → Q
MCMAEMSEMAEMSE
25625615.924.0138.4226.1
51225613.728.1124.4217.2
102425611.419.7115.7199.8
204825612.023.3128.1218.1
102412816.328.6127.4223.3
102425611.419.7115.7199.8
102451214.327.7130.1250.1
1024102414.224.8128.4220.2
", + "bbox": [ + 107, + 508, + 439, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on the number of memory vectors $M$ and vector size $C$ .", + "bbox": [ + 75, + 705, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of $M$ and $C$ . The number of memory vectors $M$ and dimension of each vector $C$ are two important parameters of AMB. To demonstrate their effects independently, we adjust one parameter while keeping the other fixed. The results in Tab. 5 shows that among all variations of parameters, $M = 1024$ , $C = 256$ is the optimal choice.", + "bbox": [ + 75, + 763, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations of CEMs. We present visualizations of CEMs in Fig. 6, each summed along the channel dimension. The results indicate that crowd information is generally more", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc23f6061a32fa91dda40a62ab46f563aaba4f86a34f6589342c9313ef33a5a6.jpg", + "image_caption": [ + "Figure 6. Visualized CEMs. Yellower means less filtered." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 568, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/356579510cf69056b7b29fbe086087f3ae2c0861eafab076b2e729904daa0ca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 88, + 630, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f7bba4ed8dfe203fbb76915cacf53ea698797efa04a3c0e367583ad8127845a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 88, + 694, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66f46a03c57008583fe811b9e36cf6b47fb421739408fba5dff2c86dedc7a968.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 88, + 825, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25ca469d8438797335835d5528cd69768902a15075b74d1314a43165ba1681d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 88, + 887, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd211b7d1a61efbb4e8ee5e944203bbe1c6d47e64ebd1f4718a5d747a1bfc1b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 176, + 598, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89e7d6a0a103ceec615581142a89c86886c40de7dae79a55c7841b53da664fb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 233, + 598, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16386179d7e8aeb5adaa9b80042426adf1b30479adc192702f7e03c605e26c7f.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 503, + 287, + 598, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db91b594668215fbe44b9bbd0cf8c688505ee3c3599c6093a3901998b0521265.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 176, + 694, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64cdb54bf3a203785067fa227d11be96cf9524aa6b1754a59da2e934415d2210.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 599, + 233, + 694, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eafc32d6f395f3c3fe5ef9060741fc24ec5bc640404929370bd6893758d04452.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 176, + 792, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1aceeabf9e4476c4590acaaf069a364ba2188429e32b1cbcc2d2d41f001ec97f.jpg", + "image_caption": [ + "(c) Predicted" + ], + "image_footnote": [], + "bbox": [ + 696, + 233, + 792, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51c19f01d875549ace4a2b875b8fd509639a96c7a5c2cef284a2b6d29597a2e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 176, + 888, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/559708f2bde012ed0fe22fca3512ac092e69f27b895a31b4783957e0552ef4ce.jpg", + "image_caption": [ + "(d) Binarized", + "Figure 7. Visualization of (a) sample images from SHB, (b) the ground-truth PCMs, (c) PCMs predicted by MPCount trained on SHA and (d) predicted PCMs after binarization." + ], + "image_footnote": [], + "bbox": [ + 792, + 233, + 888, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sensitive to domain shift than environments.", + "bbox": [ + 500, + 439, + 790, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations of PCMs. We visualize the predicted and binarized PCMs from MPCount for qualitative analysis. As shown in Fig. 7, MPCount predicts accurate PCMs that provide reliable information of crowd locations. In the raw predicted PCMs, there are unconfident predictions with low classification scores, which are filtered out in the binarized PCMs. The results verify that PC can effectively mitigate the label ambiguity problem and contribute to a more robust crowd counting model against domain shift.", + "bbox": [ + 496, + 454, + 890, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 606, + 617, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes MPCount to address single domain generalization for crowd counting with possibly narrow source distribution. MPCount tackles two unique challenges, namely, density regression and label ambiguity. We propose an attention memory bank to reconstruct domain-invariant features for regression, with the content error mask eliminating domain-related content information and the attention consistency loss ensuring the consistency of memory vectors. Patch-wise classification is a novel auxiliary task to mitigate the ambiguous pixel-level labels with accurate information, enhancing the robustness of density predictions. Extensive experiments on well-known datasets show that MPCount achieves significantly the best results on various benchmarks by outperforming the state of the art under different unseen scenarios and narrow source domain. Acknowledgement. We would like to thank the Turing AI Computing Cloud (TACC) [42] and HKUST iSING Lab for providing us computation resources on their platform.", + "bbox": [ + 496, + 632, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28032", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 118, + 174, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haoyue Bai, Hao He, Zhuoxuan Peng, Tianyuan Dai, and S.-H. Gary Chan. CounTr: An End-to-End Transformer Approach for Crowd Counting and Density Estimation. In Proceedings of European Conference on Computer Vision (ECCV) International Workshop on Distributed Smart Cameras, pages 207–222, Tel Aviv, Israel, 2022. Springer Lecture Notes in Computer Science. 2", + "[2] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. Proceedings of the AAAI Conference on Artificial Intelligence, 37:259-267, 2023. 1, 2", + "[3] Yang Chen, Yu Wang, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. A style and semantic memory mechanism for domain generalization*. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9144-9153, 2021. 2", + "[4] Jian Cheng, Haipeng Xiong, Zhiguo Cao, and Hao Lu. Decoupled two-stage crowd counting and beyond. IEEE Transactions on Image Processing, 30:2862-2875, 2021. 2", + "[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 2, 3, 6, 7, 1", + "[6] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 561-570, 2023. 1, 2, 3, 4, 6, 7", + "[7] Junyu Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 2", + "[8] Junyu Gao, Wei Lin, Bin Zhao, Dong Wang, Chenyu Gao, and Jun Wen. $\\mathbf{C}^3$ framework: An open-source pytorch code for crowd counting. arXiv preprint arXiv:1907.02724, 2019. 1", + "[9] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1", + "[10] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and S.-H. Gary Chan. Target-agnostic Source-free Domain Adaptation for Regression Tasks. In Proceedings of 40th IEEE International Conference on Data Engineering (ICDE'24), Utrecht, Netherlands, 13-16 May 2024 (to appear). IEEE. 1", + "[11] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Ali Al-Maadeed, Nasir M. Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In European Conference on Computer Vision, 2018. 5", + "[12] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Stylenormalization and restitution for domain generalization and" + ], + "bbox": [ + 78, + 143, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adaptation. IEEE Transactions on Multimedia, 24:3636-3651, 2022. 2", + "[13] Hyeonseong Kim, Yoonsu Kang, Changgyoon Oh, and Kuk-Jin Yoon. Single domain generalization for lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17587-17598, 2023. 2, 3", + "[14] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4350, 2022. 3", + "[15] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 2", + "[16] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2018. 2", + "[17] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, 2022. 2, 6, 7", + "[18] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 2, 6, 7", + "[19] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 6", + "[20] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6141-6150, 2019. 2, 6, 7", + "[21] Lucas Mansilla, Rodrigo Echeveste, D.H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7", + "[22] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15529-15539, 2021. 2", + "[23] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 2, 6, 7, 1", + "[24] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6, 7, 1", + "[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28033", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2, 6, 7", + "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 6", + "[28] Vishwanath A. Sindagi and Vishal M. Patel. Cnn-based cascaded multi-task learning of high-level prior and density estimation for crowd counting. 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), pages 1-6, 2017. 3", + "[29] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2594-2609, 2020. 5", + "[30] Leslie N. Smith and Nicholay Topin. Super-convergence: very fast training of neural networks using large learning rates. In *Defense + Commercial Sensing*, 2017. 6", + "[31] Qingyu Song, Changan Wang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Jian Wu, and Jiayi Ma. To choose or to fuse? scale selection for crowd counting. The Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21), 2021. 2, 6, 7", + "[32] Maryam Sultana, Muzammal Naseer, Muhammad Haris Khan, Salman Khan, and Fahad Shahbaz Khan. Self-distilled vision transformer for domain generalization. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3068-3085, 2022. 2", + "[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. ArXiv, abs/1607.08022, 2016. 4", + "[34] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2", + "[35] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xiansheng Hua. Meta convolutional neural networks for single domain generalization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4672-4681, 2022. 2", + "[36] Jia Wan, Ziquan Liu, and Antoni B. Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1974-1983, 2021. 2", + "[37] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai. Distribution matching for crowd counting. In Advances in Neural Information Processing Systems, 2020. 2, 6, 7", + "[38] Mingjie Wang, Hao Cai, Xianfeng Han, Jun Zhou, and Minglun Gong. Stnet: Scale tree network with multi-level auxilator for crowd counting. IEEE Transactions on Multimedia, 25:2074-2084, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8198-8207, 2019. 1, 2", + "[40] Qi Wang, Tao Han, Junyu Gao, and Yuan Yuan. Neuron linear transformation: Modeling the domain shift for crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 33(8):3238-3250, 2022. 2", + "[41] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 2, 6, 7", + "[42] Kaiqiang Xu, Xinchen Wan, Hao Wang, Zhenghang Ren, Xudong Liao, Decang Sun, Chaoliang Zeng, and Kai Chen. Tacc: A full-stack cloud computing infrastructure for machine learning tasks. arXiv preprint arXiv:2110.01556, 2021.8", + "[43] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI Conference on Artificial Intelligence, 2022. 2, 3", + "[44] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-841, 2015. 1, 2", + "[45] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2, 5", + "[46] Yuxiang Zhang, Wei Li, Weidong Sun, Ran Tao, and Qian Du. Single-source domain expansion network for cross-scene hyperspectral image classification. IEEE Transactions on Image Processing, 32:1498-1512, 2022. 2", + "[47] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 2, 6, 7", + "[48] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 6, 7", + "[49] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 6, 7", + "[50] Su Zixian, Yao Kai, Yang Xi, Wang Qiufeng, Sun Jie, and Huang Kaizhu. Rethinking data augmentation for single-source domain generalization in medical image segmentation. In AAAI, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28034", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_model.json b/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9727af050eea4eaf38bfcf0cea9d304c8002b2d7 --- /dev/null +++ b/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_model.json @@ -0,0 +1,2783 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.227, + 0.131, + 0.745, + 0.154 + ], + "angle": 0, + "content": "Single Domain Generalization for Crowd Counting" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.182, + 0.701, + 0.234 + ], + "angle": 0, + "content": "Zhuoxuan Peng, S.-H. Gary Chan \nThe Hong Kong University of Science and Technology \n{zpengac, gchan}@cse.ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.65 + ], + "angle": 0, + "content": "Due to its promising results, density map regression has been widely employed for image-based crowd counting. The approach, however, often suffers from severe performance degradation when tested on data from unseen scenarios, the so-called \"domain shift\" problem. To address the problem, we investigate in this work single domain generalization (SDG) for crowd counting. The existing SDG approaches are mainly for image classification and segmentation, and can hardly be extended to our case due to its regression nature and label ambiguity (i.e., ambiguous pixel-level ground truths). We propose MPCount, a novel effective SDG approach even for narrow source distribution. MPCount stores diverse density values for density map regression and reconstructs domain-invariant features by means of only one memory bank, a content error mask and attention consistency loss. By partitioning the image into grids, it employs patch-wise classification as an auxiliary task to mitigate label ambiguity. Through extensive experiments on different datasets, MPCount is shown to significantly improve counting accuracy compared to the state of the art under diverse scenarios unobserved in the training data characterized by narrow source distribution. Code is available at https://github.com/Shimmer93/MPCount." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Image-based crowd counting is to estimate the number of people (or objects) in an image. The recent mainstream counting approach is density map regression, where a density map with continuous pixel value is predicted from the input image, and the total crowd count is obtained by summing all these pixel values. During training, the ground-truth density map is generated by smoothing point annotations given by the locations of human heads. Density map regression has been demonstrated to achieve encouraging results in crowd counting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Existing training models predominantly assume that the testing data share the same source distribution, i.e., fully supervised learning. Such assumption, however, often breaks" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.268, + 0.635, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.327, + 0.683, + 0.337 + ], + "angle": 0, + "content": "(a) Source domain" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.269, + 0.758, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.325, + 0.731, + 0.337 + ], + "angle": 0, + "content": "gain" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.269, + 0.886, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.326, + 0.867, + 0.338 + ], + "angle": 0, + "content": "(b) Target image" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.34, + 0.632, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.397, + 0.614, + 0.408 + ], + "angle": 0, + "content": "(c) Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.34, + 0.759, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.397, + 0.729, + 0.409 + ], + "angle": 0, + "content": "(d) DCCUS" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.34, + 0.887, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.397, + 0.879, + 0.409 + ], + "angle": 0, + "content": "(e) MPCount (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.421, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Figure 1. Single domain generalization for crowd counting. (a) Sample images from a single source domain \\( S \\) with narrow distribution. (b) A challenging target image from an unseen scenario. (c) The ground-truth density map. (d) The predicted density map from the previous work DCCUS [6] trained on \\( S \\). (e) The predicted density map from our MPCount trained on \\( S \\). MPCount achieves much lower counting error than DCCUS and makes better predictions in the cropped area against the domain shift." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.639 + ], + "angle": 0, + "content": "in practice due to changes in camera position, weather condition, etc. Such deviation from the training data (i.e. source domain) to unseen testing data (i.e. target domain) is known as domain shift, which severely degrades the performance of deep learning models in real-world deployment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.746 + ], + "angle": 0, + "content": "To mitigate the domain shift problem, domain adaptation (DA) approaches have been proposed to transfer knowledge from the source domain to the target one, typically by finetuning pre-trained models with target domain data [2, 10, 39, 44]. Despite promising results, the fine-tuning process of DA is usually laborious. Furthermore, target domain data may not be easily available in reality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.884 + ], + "angle": 0, + "content": "In contrast to DA, domain generalization (DG) aims to generalize models to any unobserved target domain without the need for target data. As data and model fine-tuning are not necessary at the target domain, DG is more convenient to be deployed in unseen or unpredictable domain. We consider in this work the challenging case of single domain generalization (SDG) for image-based crowd counting, where only one source domain, likely of narrow distribution (Fig. 1a), is used for training." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A common general SDG technique is to construct" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28025" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "domain-invariant features from different augmented versions of an image. Such approach has been shown promising for various tasks [5, 13, 34]. Independent of data augmentation, many other SDG approaches have been proposed for classification and segmentation to further boost performance. Unfortunately, most of them cannot be extended to crowd counting, a regression task without category information. Moreover, while labels in classification and segmentation tasks are precise and domain-invariant, the labels for density-based crowd counting are usually ambiguous. Such \"label ambiguity\" stems from the fact that the density values are generated from point annotations, and hence human heads and backgrounds may be assigned to similar value under different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.457 + ], + "angle": 0, + "content": "The recent work of [6] achieves DG for crowd counting based on dividing the source into multiple sub-domains. While encouraging, it requires a rather broad source domain distribution. We propose, for the first time, an SDG approach for crowd counting called MPCount without sub-domain division, thereof effective and extensible to narrow source distribution. We highlight its performance as compared with a state-of-the-art scheme in Fig. 1. MPCount achieves its superior performance with the following two novel components:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.46, + 0.473, + 0.702 + ], + "angle": 0, + "content": "- An attention memory bank (AMB) to tackle density regression: MPCount employs only one attention memory bank (AMB) that takes a pair of features encoded from an crowd image and its data-augmented version as input. To cover continuous density values for regression with a finite memory size, AMB reconstructs each feature vector as an attention over memory vectors. These memory vectors learn domain-invariant representations from the pair of features of different styles but similar content (i.e. crowd densities). We present the content error mask (CEM) to eliminate domain-related content from input features, as characterized by a large discrepancy between instance-normalized feature element pairs. Furthermore, the newly employed attention consistency loss (ACL) enforces the similarity of attention scores produced from input features, ensuring the consistency of memory vectors." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.702, + 0.473, + 0.853 + ], + "angle": 0, + "content": "- Patch-wise classification \\((PC)\\) to tackle label ambiguity: To address the label ambiguity, MPCount employs a novel auxiliary task called patch-wise classification (PC). In this task, each crowd image is evenly divided into fixed-size square patches, say \\(16 \\times 16\\), classified into two classes, i.e., containing human heads or not. The density values in the areas classified as without head are filtered out during density regression, so that the pixel-level ambiguity is overcome by coarser but more accurate patchwise binary labels." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.46, + 0.473, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "To validate MPCount, we conduct extensive experiments on various crowd counting datasets including ShanghaiTech A (SHA) & B (SHB) and JHU-Crowd++. We introduce" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.258 + ], + "angle": 0, + "content": "a new challenging setting of narrow source distribution, where only images of the same category, such as \"snow\" (SN) and \"fog/haze\" (FH) in JHU-Crowd++, are used in training. We demonstrate that MPCount achieves excellent performance not only on traditional inter-dataset benchmarks, but also under our newly introduced narrow source setting. Let \\( S \\rightarrow T \\) denote the case where the source domain is \\( S \\) and the target domain is \\( T \\). As compared to the state of the art, MPCount is shown to reduce significantly counting error by \\( 21.8\\% \\) on SN \\( \\rightarrow \\) FH, \\( 18.6\\% \\) on FH \\( \\rightarrow \\) SN, \\( 18.2\\% \\) on SHB \\( \\rightarrow \\) SHA and \\( 9.5\\% \\) on SHA \\( \\rightarrow \\) SHB." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.272, + 0.651, + 0.288 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.297, + 0.8, + 0.314 + ], + "angle": 0, + "content": "2.1. Fully Supervised Crowd Counting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.533 + ], + "angle": 0, + "content": "The mainstream approach to crowd counting is via density map regression, where each pixel of an image is assigned a count value, and the summation of these values yields an estimate of the total number of people. Most methods focus on improving counting accuracy under the fully supervised setting by employing novel network designs [1, 16, 17, 31, 45] or loss functions [20, 26, 36, 37]. While encouraging results have been shown, their performance usually drop significantly when evaluated on out-of-distribution data, due to their limited generalization ability. Besides, several works [4, 22, 38] have proposed auxiliary tasks in crowd counting for certain purposes. However, these tasks still require pixel-level predictions and contribute little to tackling label ambiguity." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.542, + 0.892, + 0.559 + ], + "angle": 0, + "content": "2.2. Domain Adaptation (DA) for Crowd Counting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.895, + 0.703 + ], + "angle": 0, + "content": "Domain adaptation (DA) has been widely studied in crowd counting to address domain shift by adapting source domain information to a particular target domain. While a few methods study supervised DA where labeled data from target domain are available [40, 44], most approaches tackle unsupervised DA, utilizing only unlabeled target domain data [2, 7, 18, 39, 41, 47]. Despite the notable advancements, DA methods normally require data from the target domain, which may not be easily available in practice." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.711, + 0.822, + 0.728 + ], + "angle": 0, + "content": "2.3. Single Domain Generalization (SDG)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Domain Generalization (DG) approaches are designed to train a network with generalization ability solely utilizing source domain data, and single domain generalization (SDG) is the special case when only one source domain is available. There are various techniques for SDG, including 1) adversarial data generation [15, 25, 46, 50], 2) feature normalization/whitening transformation [5, 12, 23, 24, 43] and 3) domain-general network design, such as a convolution layer [35], vision transformer [32] and memory bank [3]. Among various SDG methods, a common general technique is to simulate domain shift with data augmenta" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28026" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.168 + ], + "angle": 0, + "content": "tion and extract domain-invariant information from different augmented versions of features [5, 13, 43]. While SDG has attracted widespread attention in classification and segmentation, it is still a nascent topic in crowd counting due to its regression nature and label ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.424 + ], + "angle": 0, + "content": "DCCUS [6] studies SDG for crowd counting. In this method, the source domain is dynamically clustered into sub-domains simulated as meta-training and -test sets in a meta-learning strategy. Two types of memory modules and several losses are proposed to distinguish and record domain-invariant and -related information. DCCUS designs the memory mechanism specifically for regression, but the problem of ambiguous labels is not considered. Moreover, when the source domain follows a narrow distribution, the sub-domain division process may experience reduced effectiveness and affect the generalization ability of the counting model. MPCount, without sub-domain division, employs novel patch-wise classification to address the label ambiguity problem. Additionally, a single memory bank for regression is introduced without the need for sub-domain partitioning, enabling MPCount to generalize well even in source domains of narrow distributions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.185, + 0.451 + ], + "angle": 0, + "content": "3. MPCount" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.468, + 0.613 + ], + "angle": 0, + "content": "This section provides a detailed description of our MPCount scheme. We begin by reviewing the foundational concepts of crowd counting in Sec. 3.1. Next, in Sec. 3.2, we present a comprehensive overview of the entire scheme. We then elaborate the attention memory bank (Sec. 3.3) designed for density regression, followed by the content error mask (Sec. 3.4) and attention consistency loss (Sec. 3.5). After that, we discuss the patch-wise classification (Sec. 3.6) proposed to address the challenge of ambiguous labels. Finally, we detail the overall training loss in Sec. 3.7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.622, + 0.35, + 0.637 + ], + "angle": 0, + "content": "3.1. Crowd Counting Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.468, + 0.736 + ], + "angle": 0, + "content": "First, we review the preliminary knowledge of deep learning-based crowd counting via density map regression. Our objective is to train a neural network \\(\\mathcal{N}\\) with parameter \\(\\theta\\) that takes in an input image \\(\\mathcal{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) and outputs a density map \\(\\hat{\\mathcal{D}} = \\mathcal{N}(\\mathcal{I};\\theta)\\) of size \\(H \\times W\\). The estimated count \\(\\hat{c}\\) is the sum of all pixel-wise density values in \\(\\hat{\\mathcal{D}}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.758, + 0.468, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\hat {c} = \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\hat {\\mathcal {D}} _ {i j}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Generation of ground-truth density maps usually follows the method in [28]. Given the human head point annotations \\(\\mathcal{H}\\) where each \\(h\\in \\mathcal{H}\\) is the coordinate of a head location, the ground-truth density map \\(\\mathcal{D}\\in \\mathbb{R}^{H\\times W}\\) is calculated as the sum of the 2D Gaussian filter applied to each head coordinate \\(h\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.109, + 0.891, + 0.142 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = \\sum_ {h \\in \\mathcal {H}} \\delta (x - h) \\times G _ {\\sigma} (x), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.147, + 0.891, + 0.176 + ], + "angle": 0, + "content": "where \\(\\delta (\\cdot)\\) is the discrete delta function, and \\(G_{\\sigma}(\\cdot)\\) is the Gaussian kernel with a fixed variance \\(\\sigma\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.178, + 0.891, + 0.207 + ], + "angle": 0, + "content": "A common objective function to supervise the network \\(\\mathcal{N}\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.227, + 0.891, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} (\\theta) = | | \\mathcal {D}, \\hat {\\mathcal {D}} | | _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.25, + 0.673, + 0.264 + ], + "angle": 0, + "content": "3.2. Scheme Overview" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.56 + ], + "angle": 0, + "content": "The overall structure of our method is illustrated in Fig. 2. During training, photometric transformations are applied to the original image \\(\\mathcal{I}^{ori}\\) to produce an augmented version \\(\\mathcal{I}^{aug}\\). The feature extractor encodes them into features \\(\\mathcal{F}^{ori}\\) and \\(\\mathcal{F}^{aug}\\), and a content error mask (CEM) \\(\\mathcal{M}\\) is applied to them. The masked features are reconstructed into \\(\\tilde{\\mathcal{F}}^{ori}\\) and \\(\\tilde{\\mathcal{F}}^{aug}\\) via an attention memory bank (AMB) and passed to the density head for crowd density prediction. Meanwhile, the highest-level features encoded by the feature extractor, \\(\\mathcal{Z}^{ori}\\) and \\(\\mathcal{Z}^{aug}\\), are fed into the patch-wise classification (PC) head, and the predicted PC maps (PCMs) \\(\\hat{\\mathcal{C}}^{ori}\\) and \\(\\hat{\\mathcal{C}}^{aug}\\) are binarized and resized to serve as a mask to filter out areas without human heads in the estimated density maps \\(\\hat{\\mathcal{D}}^{ori}\\) and \\(\\hat{\\mathcal{D}}^{aug}\\). The final predictions \\(\\mathcal{D}'^{ori}\\) and \\(\\mathcal{D}'^{aug}\\) are two density maps with certain areas masked out by the PCMs. The entire training process is supervised under a combination of density losses, PC losses and the attention consistency loss (ACL). The inference process solely employs \\(\\mathcal{I}^{ori}\\) as input, and outputs a single density map \\(\\mathcal{D}'^{ori}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.783, + 0.583 + ], + "angle": 0, + "content": "3.3. Attention Memory Bank (AMB)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.891, + 0.741 + ], + "angle": 0, + "content": "The attention memory bank (AMB) is designed to automatically learn domain-invariant features for density regression. [14] proposes a memory bank where memory vectors are updated individually, each corresponding to a certain category. However, this design cannot be applied to regression tasks, since a finite number of memory vectors cannot cover continuous density values. Inspired by [6], we reconstruct each feature vector as an attention over memory vectors, so that any density value may be represented as a linear combination of representative values in the memory bank." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.741, + 0.892, + 0.847 + ], + "angle": 0, + "content": "The AMB \\(\\mathcal{V} \\in \\mathbb{R}^{M \\times C}\\) consists of \\(M\\) memory vectors of dimension \\(C\\). Given a flattened input feature map \\(\\mathcal{F} \\in \\mathbb{R}^{HW \\times C}\\), we first compute the attention scores \\(\\mathcal{A}\\) between the feature \\(\\mathcal{F}\\) (query) and the memory \\(\\mathcal{V}\\) (key). Then the reconstructed feature map \\(\\tilde{\\mathcal{F}}\\) is calculated as a linear combination of \\(\\mathcal{V}\\) (value) with \\(\\mathcal{A}\\) as weights. The reconstruction process can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.864, + 0.891, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (\\mathcal {F}, \\mathcal {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {F V} ^ {T}}{\\sqrt {C}}\\right), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28027" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.895, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.893, + 0.375 + ], + "angle": 0, + "content": "Figure 2. The overall training pipeline of our proposed MPCount. All identical modules in this diagram share the same weights. In the encoder-decoder structure, a higher level indicates a deeper feature." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.4, + 0.107, + 0.413 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.426, + 0.469, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {F}} = \\mathcal {A} \\mathcal {V}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.579 + ], + "angle": 0, + "content": "While [6] relies on multiple memory banks corresponding to different sub-domains to distinguish domain-invariant and -related information, we only employ a single AMB and propose the novel CEM (Sec. 3.4) and ACL (Sec. 3.5) to ensure that it stores domain-invariant representations. To this end, our feature reconstruction mechanism can work without sub-domain division, thereof is effective even with a narrow source distribution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.331, + 0.606 + ], + "angle": 0, + "content": "3.4. Content Error Mask (CEM)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.839 + ], + "angle": 0, + "content": "CEM is proposed to guarantee the resemblance of content contained in the pair of input features by excluding possible domain-related content information. A common technique to disentangle style and content in a feature is instance normalization (IN) [33], where the feature statistics (channel-wise mean and variance) are considered to contain style information, while the instance normalized features retain content information. Based on this idea, we assume that the discrepancy of instance normalized features reflect effects of domain shift on feature content and may mislead AMB to learn domain-related information. Therefore, we filter out elements in the input features with difference between their instance normalized versions above a certain threshold, resulting in a pair of features with similar content information for domain-invariant feature reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "More specifically, given the features extracted from the original and augmented images, \\(\\mathcal{F}^{ori}\\) and \\(\\mathcal{F}^{aug} \\in \\mathbb{R}^{H \\times W \\times C}\\), we define the content error mask \\(\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times C}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.42, + 0.892, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {i j k} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | \\mathrm {I N} (\\mathcal {F} ^ {\\text {o r i}}) _ {i j k} - \\mathrm {I N} (\\mathcal {F} ^ {\\text {a u g}}) _ {i j k} | \\leq \\alpha \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.512 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the threshold indicating whether the error value reflects possible inconsistency of content information caused by domain shift." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.512, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Given \\(\\mathcal{F}\\) to be one of \\(\\mathcal{F}^{\\text{ori}}\\) and \\(\\mathcal{F}^{\\text{aug}}\\), we filter out the domain-related content information by" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.566, + 0.891, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {\\prime} = \\operatorname {D r o p o u t 2 D} (\\mathcal {F} \\odot \\mathcal {M}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.891, + 0.621 + ], + "angle": 0, + "content": "The random 2D dropout is applied to \\(\\mathcal{F}'\\) to prevent the memory from relying solely on certain channels." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.628, + 0.8, + 0.645 + ], + "angle": 0, + "content": "3.5. Attention Consistency Loss (ACL)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.772 + ], + "angle": 0, + "content": "The style information of \\(\\mathcal{F}^{ori}\\) and \\(\\mathcal{F}^{aug}\\) preserved in Sec. 3.4 may produce a large discrepancy on their attention distributions over the memory vectors in AMB even if they contain identical content information. To ensure that each memory vector consistently store specific domain-invariant representations, we propose the attention consistency loss (ACL) to enforce the similarity of attention scores produced from \\(\\mathcal{F}^{ori}\\) and \\(\\mathcal{F}^{aug}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.849 + ], + "angle": 0, + "content": "We consider \\(\\mathcal{A}\\) in Eq. 4 as a distribution, and the ACL \\(\\mathbb{L}_{con}\\) is calculated as the distance between the distribution \\(\\mathcal{A}^{ori}\\) produced by \\(\\mathcal{F}^{ori}\\) and \\(\\mathcal{A}^{aug}\\) produced by \\(\\mathcal{F}^{aug}\\). Here the simple Euclidean distance is selected as the distance measure due to its computational stability during training:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.871, + 0.891, + 0.889 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} _ {\\text {c o n}} = \\left\\| \\mathcal {A} ^ {\\text {a u g}}, \\mathcal {A} ^ {\\text {o r i}} \\right\\| _ {2} ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28028" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.155, + 0.089, + 0.272, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.161, + 0.259, + 0.173 + ], + "angle": 0, + "content": "(a) Density Map" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.089, + 0.394, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.161, + 0.358, + 0.172 + ], + "angle": 0, + "content": "(b) PCM" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.268 + ], + "angle": 0, + "content": "Figure 3. Illustration of label ambiguity and how PCM tackles it. Point A is a typical example which belongs to a human head but is assigned the density value 0 as the background in the density map due to varying head sizes. In the PCM, the patch covering A is correctly classified as containing heads, thanks to the patch-level accuracy of the labels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.291, + 0.343, + 0.307 + ], + "angle": 0, + "content": "3.6. Patch-wise Classification (PC)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.469, + 0.42 + ], + "angle": 0, + "content": "As discussed in Sec. 3.1, ground-truth density maps are usually calculated only based on point annotations. In different scenarios, pixels of a human head and the background might be assigned the same density value due to various head sizes (Fig. 3a), violating the common assumption that labels are accurate and domain-invariant, thereof resulting in the ambiguous labels in crowd counting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.469, + 0.555 + ], + "angle": 0, + "content": "To tackle the label ambiguity, we propose a novel auxiliary task named patch-wise classification (PC) and its corresponding supervision signal, patch-wise classification map (PCM), as illustrated in Fig. 3b. A PCM divides an image evenly into \\( P \\times P \\) (\\( P = 16 \\) empirically) patches and classifies each patch into two categories: containing human heads or not. Such patch-level predictions mitigate the uncertainty in pixel-level density maps by providing coarser but more accurate patch-level information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.469, + 0.585 + ], + "angle": 0, + "content": "In practice, a ground truth \\(\\mathbf{PCM}\\mathcal{C}\\) with patch size \\(P\\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.59, + 0.469, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {i j} = 1 - \\delta \\left(\\sum_ {k = i P} ^ {(i + 1) P} \\sum_ {l = j P} ^ {(j + 1) P} \\mathcal {D} _ {k l}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.674 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is the ground truth density map and \\(\\delta (\\cdot)\\) is the discrete delta function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.706 + ], + "angle": 0, + "content": "Our model predicts a PCM \\(\\hat{C}\\) supervised by the binary cross entropy (BCE) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.712, + 0.469, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} _ {c l s} = \\operatorname {B C E} (\\mathcal {C}, \\hat {\\mathcal {C}}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.469, + 0.801 + ], + "angle": 0, + "content": "Subsequently, \\(\\hat{\\mathcal{C}}\\) is binarized and resized to obtain \\(\\mathcal{C}'\\) which matches the dimensions of the predicted density map \\(\\hat{\\mathcal{D}}\\). \\(\\mathcal{C}'\\) is then used to mask the regions classified without crowds in \\(\\hat{\\mathcal{D}}\\), resulting in the final density map \\(\\mathcal{D}'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.807, + 0.469, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} ^ {\\prime} = \\hat {\\mathcal {D}} \\odot \\mathcal {C} ^ {\\prime}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.469, + 0.878 + ], + "angle": 0, + "content": "Finally, the density regression head is supervised with the ground truth density map \\(\\mathcal{D}\\), and the common Euclidean distance in Eq. (3) is used as the objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} _ {d e n} = \\left\\| \\mathcal {D}, \\mathcal {D} ^ {\\prime} \\right\\| _ {2} ^ {2}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.658, + 0.108 + ], + "angle": 0, + "content": "3.7. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.114, + 0.892, + 0.16 + ], + "angle": 0, + "content": "The overall training loss is a combination of density losses \\(\\mathbb{L}_{den}^{ori},\\mathbb{L}_{den}^{aug}\\), PC losses \\(\\mathbb{L}_{cls}^{ori},\\mathbb{L}_{cls}^{aug}\\) and the attention consistency loss \\(\\mathbb{L}_{con}\\), which can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.168, + 0.892, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} = \\mathbb {L} _ {d e n} ^ {o r i} + \\mathbb {L} _ {d e n} ^ {a u g} + \\lambda_ {c l s} \\left(\\mathbb {L} _ {c l s} ^ {o r i} + \\mathbb {L} _ {c l s} ^ {a u g}\\right) + \\lambda_ {c o n} \\mathbb {L} _ {c o n}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.209, + 0.892, + 0.238 + ], + "angle": 0, + "content": "where \\(\\lambda_{cls}\\) and \\(\\lambda_{con}\\) are weighting parameters to balance different loss terms." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.251, + 0.803, + 0.269 + ], + "angle": 0, + "content": "4. Illustrative Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.368 + ], + "angle": 0, + "content": "In this section, we first introduce the datasets used in experiments in Sec. 4.1. Next, the implementation details are specified in Sec. 4.2, then evaluation metrics are described in Sec. 4.3. We present the results compared with state of the art in Sec. 4.4. Finally, we conduct ablation studies and other additional analysis in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.602, + 0.39 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.444 + ], + "angle": 0, + "content": "We evaluate our method on four mainstream crowd counting datasets: ShanghaiTech Part A & B, UCF-QNRF and JHU-Crowd++." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.446, + 0.892, + 0.565 + ], + "angle": 0, + "content": "- ShanghaiTech [45] is composed of two parts, SHA (A) and SHB (B). SHA contains 300 training images and 182 testing images, while SHB contains 400 training images and 316 testing images. Images in SHA are collected from the Internet and feature highly crowded scenes. In contrast, SHB is captured from several streets in Shanghai, and its images exhibit crowd densities generally lower than those in SHA." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.567, + 0.892, + 0.627 + ], + "angle": 0, + "content": "- UCF-QNRF (Q) [11] consists of 1201 training images and 334 testing images. It is a challenging dataset with a large range of crowd densities, scenes, viewpoints and lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.628, + 0.892, + 0.869 + ], + "angle": 0, + "content": "- JHU-Crowd++ [29] is a large-scale dataset containing 4372 images, with 2722, 500 and 1600 images in the training, validation and testing set, respectively. In addition, image-level labels are also provided in this dataset, including 16 types of scenes and 4 types of weather conditions. For scene annotations, there are 879 images labeled \"stadium\" (SD) and 573 labeled \"street\" (ST), which are utilized as two source domains. For weather annotations, we use 201 images labeled \"snow\" (SN) and 168 images labeled \"fog/haze\" (FH) as source domains. In each domain, \\(80\\%\\) of the images are selected as the training set while the rest \\(20\\%\\) are for testing. Data with the same label belong to a narrower distribution than mainstream datasets, thus are more challenging for SDG. Illustrative samples in these scene-specific and weather-specific domains are available in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Let \\(S \\to \\mathcal{T}\\) denote the case where \\(S\\) is the source domain and \\(\\mathcal{T}\\) is the target domain. The datasets are utilized" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.446, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.177, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.217, + 0.162, + 0.228 + ], + "angle": 0, + "content": "(a) Stadium" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.273, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.217, + 0.252, + 0.228 + ], + "angle": 0, + "content": "(b) Street" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.089, + 0.371, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.217, + 0.347, + 0.228 + ], + "angle": 0, + "content": "(c) Snow" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.468, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.217, + 0.455, + 0.228 + ], + "angle": 0, + "content": "(d) Fog/Haze" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.241, + 0.47, + 0.268 + ], + "angle": 0, + "content": "Figure 4. Sample images with different instance-level labels in JHU-Crowd++." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.293, + 0.47, + 0.369 + ], + "angle": 0, + "content": "in experiments under two distinct types of settings: 1) An entire dataset is regarded as a single domain: \\(\\mathrm{A} \\rightarrow \\mathrm{B} / \\mathrm{Q}\\), \\(\\mathrm{B} \\rightarrow \\mathrm{A} / \\mathrm{Q}\\) and \\(\\mathrm{Q} \\rightarrow \\mathrm{A} / \\mathrm{B}\\); 2) A dataset is partitioned into subsets according to image-level labels, and one subset constitutes a single domain: \\(\\mathrm{SD} \\leftrightarrow \\mathrm{SR}\\) and \\(\\mathrm{SN} \\leftrightarrow \\mathrm{FH}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.377, + 0.236, + 0.392 + ], + "angle": 0, + "content": "4.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.642 + ], + "angle": 0, + "content": "We adopt VGG16-BN [27] as the feature extractor in our model. The density regression head is a single convolution layer with \\( 1 \\times 1 \\) filters, while the classification head consists of a \\( 3 \\times 3 \\) and a \\( 1 \\times 1 \\) convolution layers. Batch normalization is employed after all convolution layers except the last ones. For data augmentation, we apply three types of photometric transformations, color jittering, Gaussian blurring and sharpening. We also randomly crop image patches with a size of \\( 320 \\times 320 \\) and adopt random horizontal flipping to both the original and augmented images simultaneously. We select AdamW [19] as the optimizer and OneCycleLR [30] as the learning rate scheduler with maximum learning rate set to 1e-3 and a maximum epoch of 300. The memory size \\( M \\) is set to 1024 and the dimension \\( C \\) is 256. We use a content error threshold \\( \\alpha \\) of 0.5, and the loss weights \\( \\lambda_{cls} \\) and \\( \\lambda_{con} \\) are both set to 10." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.262, + 0.663 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.673, + 0.469, + 0.703 + ], + "angle": 0, + "content": "We evaluate our method with mean absolute error (MAE) and mean squared error (MSE), defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.723, + 0.469, + 0.786 + ], + "angle": 0, + "content": "\\[\nM A E = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| c _ {i} - \\hat {c} _ {i} \\right|, M S E = \\sqrt {\\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(c _ {i} - \\hat {c} _ {i}\\right) ^ {2}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.794, + 0.469, + 0.84 + ], + "angle": 0, + "content": "where \\(N\\) is the number of testing images, \\(c_{i}\\) is the ground truth count of the \\(i\\)-th image and \\(\\hat{c}_i\\) is the predicted count. Lower values of both metrics indicate better performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.37, + 0.863 + ], + "angle": 0, + "content": "4.4. Comparison with State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In this subsection, we compare our MPCount with state-of-the-art methods on different benchmarks, as illustrated" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.599, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.239, + 0.578, + 0.25 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.089, + 0.696, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.238, + 0.667, + 0.249 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.793, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.238, + 0.776, + 0.249 + ], + "angle": 0, + "content": "(c) DCCUS" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.089, + 0.89, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.238, + 0.878, + 0.249 + ], + "angle": 0, + "content": "(d) MPCount" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Figure 5. Visualization results of DCCUS [6] and MPCount under different DG settings. First row: A \\(\\rightarrow\\) B; second row: SN \\(\\rightarrow\\) FH; third row: B \\(\\rightarrow\\) A." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.345, + 0.892, + 0.39 + ], + "angle": 0, + "content": "in Fig. 5. Tab. 1 shows the results on \\( \\mathrm{A} \\rightarrow \\mathrm{B} / \\mathrm{Q} \\), \\( \\mathrm{B} \\rightarrow \\mathrm{A} / \\mathrm{Q} \\) and \\( \\mathrm{Q} \\rightarrow \\mathrm{A} / \\mathrm{B} \\). The selected baselines can be divided into three categories:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.397, + 0.891, + 0.427 + ], + "angle": 0, + "content": "- Fully supervised crowd counting: BL [20], DM-Count [37], SASNet [31], ChfL [26] and MAN [17]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.428, + 0.891, + 0.457 + ], + "angle": 0, + "content": "- Domain adaptation for crowd counting: RBT [18], C²MoT [41], FGFD [47], DAOT [48] and FSIM [49]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.458, + 0.891, + 0.488 + ], + "angle": 0, + "content": "- Single domain generalization: IBN [23], SW [24], ISW [5], DG-MAN [21] and DCCUS [6]." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.397, + 0.891, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.494, + 0.892, + 0.585 + ], + "angle": 0, + "content": "All results are copied from previous papers except IBN, SW and ISW, which are feature normalization/whitening-based SDG methods originally designed for classification or segmentation and adapted to crowd counting by us. Domain adaptation methods are listed only for reference, as data from target domains are usually visible during adaptation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.683 + ], + "angle": 0, + "content": "Our MPCount outperforms all the DG methods under most of these settings, including an significant error reduction by \\(18.2\\%\\) on \\(\\mathrm{B} \\rightarrow \\mathrm{A}\\). It is notable that MPCount exhibits outstanding performance in some cases even compared to DA methods, providing compelling evidence for the effectiveness of our design." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Tab. 2, we further conduct experiments on scene-specific domains \\(\\mathrm{SD} \\leftrightarrow \\mathrm{SR}\\) and weather-specific domains \\(\\mathrm{SN} \\leftrightarrow \\mathrm{FH}\\). BL [20], MAN [17], DAOT [48] and DCCUS [6] are trained using their released code, while IBN [23], SW [24] and ISW [5] are adapted by ourselves. We observe that the normalization/whitening-based methods cannot achieve satisfactory results, possibly because useful information is eliminated from feature statistics. DCCUS also obtains worse results than fully supervised methods, reflecting that the clustering procedure may fail to produce meaningful sub-domains when the source domain distribution is narrow. In contrast, our MPCount still performs the best among all tested methods, demonstrating its superiority in challenging conditions with narrow domain distributions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28030" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.894, + 0.369 + ], + "angle": 0, + "content": "
Source → TargetA → BA → QB → AB → QQ → AQ → B
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]✘✘15.925.8166.7287.6138.1228.1226.4411.0----
DMCount [37]✘✘23.134.9134.4252.1143.9239.6203.0386.173.4135.114.327.5
SASNet [31]✘✘21.333.2211.2418.6132.4225.6273.5481.373.9116.413.022.1
ChfL [26]✘✘18.729.1122.3218.0121.3200.8197.1357.968.7118.514.027.0
MAN [17]✘✘22.132.8138.8266.3133.6255.6209.4378.867.1122.112.522.2
RBT [18]✓✘13.429.3175.0294.8112.2218.2211.3381.9----
C2MoT [41]✓✘12.421.1125.7218.3120.7192.0198.9368.0----
FGFD [47]✓✘12.723.3124.1242.0123.5210.7209.7384.770.2118.412.520.3
DAOT [48]✓✘10.918.8113.9215.6----67.0128.411.319.6
FSIM [49]✓✘11.119.3105.3191.1120.3202.6194.9324.566.8111.511.019.7
IBN [23]✘✓19.130.8280.2561.0125.9202.3183.5317.4105.9174.616.025.5
SW [24]✘✓20.230.4285.5431.0126.7193.8200.7333.2102.4168.819.032.9
ISW [5]✘✓23.937.7215.6399.7156.2291.5263.1442.283.4136.022.134.7
DG-MAN [21]✘✓17.328.7129.1238.2130.7225.1182.4325.8----
DCCUS [6]✘✓12.624.6119.4216.6121.8203.1179.1316.267.4112.812.120.9
MPCount (Ours)✘✓11.419.7115.7199.899.6182.9165.6290.465.5110.112.324.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.38, + 0.893, + 0.409 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art methods on SHA (A), SHB (B) and UCF-QNRF (Q). Domain adaptation methods are only for reference since data in target domain might be used." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.424, + 0.816, + 0.603 + ], + "angle": 0, + "content": "
Source → TargetSD → SRSR → SDSN → FHFH → SN
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]XX42.179.0262.71,063.948.1129.5343.8770.5
MAN [17]XX45.179.0246.1950.838.168.0445.0979.3
DAOT [48]✓X45.388.0278.71,624.342.373.0151.6273.9
IBN [23]X✓92.2178.0318.11,420.4109.7267.7491.81,110.4
SW [24]X✓110.3202.4312.61,072.4131.5306.6381.3825.0
ISW [5]X✓108.1212.4385.91,464.8151.6365.7276.6439.8
DCCUS [6]X✓90.4194.1258.11,005.954.5125.8399.7945.0
MPCount (Ours)X✓37.470.1218.6935.931.355.0216.3421.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.614, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Table 2. Comparison with the state of the art on data in JHU-Crowd++ with labels \"Stadium\"(SD), \"Street\"(SR), \"Snow\"(SN) and \"Fog/Haze\"(FH). Domain adaptation methods are only for reference since data in target domain might be used." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.669, + 0.241, + 0.684 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.469, + 0.725 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies and other analysis under the setting \\( \\mathrm{A} \\to \\mathrm{B} / \\mathrm{Q} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Effects of model components. We start from a simple encoder-decoder structured baseline model and add or remove each component individually to validate its effectiveness. The experimental results are displayed in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.87 + ], + "angle": 0, + "content": "- Attention Memory Bank: After adding AMB, the performance consistently improves on both target domains, as shown in Tab. 3. This validates that AMB can effectively help the model generalize by reconstructing domain-invariant features for density regression." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.899 + ], + "angle": 0, + "content": "- Content Error Mask & Attention Consistency Loss: Next, we test the effects of CEM and ACL on AMB. The results" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.669, + 0.89, + 0.699 + ], + "angle": 0, + "content": "in Tab. 3 show that both components generally improves the performance, especially when PC is present." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.7, + 0.892, + 0.775 + ], + "angle": 0, + "content": "- Patch-wise Classification: Finally, models with PC outperforms their counterparts without PC, as visible in Tab. 3. This suggests that such an auxiliary task is useful for DG by compensating for ambiguous pixel-level labels with accurate patch-level labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Effects of \\(\\alpha\\). The threshold of CEM \\(\\alpha\\) mentioned in Eq. (6) controls the Portion of Diminished Elements (PDE) in features. According to the results in Tab. 4, \\(\\alpha = 0.5\\) produces the best performance under all settings, with a PDE of \\(10.5\\%\\). It is also indicated that using a too large value of \\(\\alpha\\) (0.9) may be inadequate to filter out domain-related information, while a too small value (0.1) would let useful information eliminated." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.473, + 0.292 + ], + "angle": 0, + "content": "
ComponentsA → BA → Q
AMBCEMACLPCMAEMSEMAEMSE
21.639.1200.9325.5
✓18.932.0176.3292.8
✓✓15.325.5156.8232.7
✓✓16.928.7172.5281.6
✓✓✓14.124.3170.6286.8
✓14.430.5144.2262.6
✓✓13.425.2133.4224.0
✓✓✓13.323.4126.9222.3
✓✓✓✓11.419.7115.7199.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.302, + 0.428, + 0.316 + ], + "angle": 0, + "content": "Table 3. Ablation studies on the model components." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.333, + 0.468, + 0.468 + ], + "angle": 0, + "content": "
ParameterPropertyA → BA → Q
αPDE(%)MAEMSEMAEMSE
0.128.616.725.9157.3338.5
0.314.612.420.8132.1223.6
0.510.511.419.7115.7199.8
0.75.911.923.8123.8214.1
0.93.613.827.2124.7209.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.479, + 0.452, + 0.493 + ], + "angle": 0, + "content": "Table 4. Ablation studies on the error threshold \\( \\alpha \\) in Eq. (6)." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.51, + 0.441, + 0.696 + ], + "angle": 0, + "content": "
ParametersA → BA → Q
MCMAEMSEMAEMSE
25625615.924.0138.4226.1
51225613.728.1124.4217.2
102425611.419.7115.7199.8
204825612.023.3128.1218.1
102412816.328.6127.4223.3
102425611.419.7115.7199.8
102451214.327.7130.1250.1
1024102414.224.8128.4220.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the number of memory vectors \\( M \\) and vector size \\( C \\) ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Effects of \\( M \\) and \\( C \\). The number of memory vectors \\( M \\) and dimension of each vector \\( C \\) are two important parameters of AMB. To demonstrate their effects independently, we adjust one parameter while keeping the other fixed. The results in Tab. 5 shows that among all variations of parameters, \\( M = 1024 \\), \\( C = 256 \\) is the optimal choice." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Visualizations of CEMs. We present visualizations of CEMs in Fig. 6, each summed along the channel dimension. The results indicate that crowd information is generally more" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.57, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.089, + 0.631, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.089, + 0.695, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.826, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.089, + 0.888, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.149, + 0.866, + 0.162 + ], + "angle": 0, + "content": "Figure 6. Visualized CEMs. Yellower means less filtered." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.178, + 0.599, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.234, + 0.599, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.289, + 0.599, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.346, + 0.578, + 0.358 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.178, + 0.695, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.234, + 0.695, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.346, + 0.668, + 0.358 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.178, + 0.793, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.234, + 0.793, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.346, + 0.78, + 0.358 + ], + "angle": 0, + "content": "(c) Predicted" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.178, + 0.89, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.234, + 0.89, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.346, + 0.878, + 0.358 + ], + "angle": 0, + "content": "(d) Binarized" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.37, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Figure 7. Visualization of (a) sample images from SHB, (b) the ground-truth PCMs, (c) PCMs predicted by MPCount trained on SHA and (d) predicted PCMs after binarization." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.44, + 0.792, + 0.454 + ], + "angle": 0, + "content": "sensitive to domain shift than environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Visualizations of PCMs. We visualize the predicted and binarized PCMs from MPCount for qualitative analysis. As shown in Fig. 7, MPCount predicts accurate PCMs that provide reliable information of crowd locations. In the raw predicted PCMs, there are unconfident predictions with low classification scores, which are filtered out in the binarized PCMs. The results verify that PC can effectively mitigate the label ambiguity problem and contribute to a more robust crowd counting model against domain shift." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.607, + 0.619, + 0.622 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper proposes MPCount to address single domain generalization for crowd counting with possibly narrow source distribution. MPCount tackles two unique challenges, namely, density regression and label ambiguity. We propose an attention memory bank to reconstruct domain-invariant features for regression, with the content error mask eliminating domain-related content information and the attention consistency loss ensuring the consistency of memory vectors. Patch-wise classification is a novel auxiliary task to mitigate the ambiguous pixel-level labels with accurate information, enhancing the robustness of density predictions. Extensive experiments on well-known datasets show that MPCount achieves significantly the best results on various benchmarks by outperforming the state of the art under different unseen scenarios and narrow source domain. Acknowledgement. We would like to thank the Turing AI Computing Cloud (TACC) [42] and HKUST iSING Lab for providing us computation resources on their platform." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28032" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.119, + 0.176, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.47, + 0.241 + ], + "angle": 0, + "content": "[1] Haoyue Bai, Hao He, Zhuoxuan Peng, Tianyuan Dai, and S.-H. Gary Chan. CounTr: An End-to-End Transformer Approach for Crowd Counting and Density Estimation. In Proceedings of European Conference on Computer Vision (ECCV) International Workshop on Distributed Smart Cameras, pages 207–222, Tel Aviv, Israel, 2022. Springer Lecture Notes in Computer Science. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[2] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. Proceedings of the AAAI Conference on Artificial Intelligence, 37:259-267, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[3] Yang Chen, Yu Wang, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. A style and semantic memory mechanism for domain generalization*. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9144-9153, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[4] Jian Cheng, Haipeng Xiong, Zhiguo Cao, and Hao Lu. Decoupled two-stage crowd counting and beyond. IEEE Transactions on Image Processing, 30:2862-2875, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 2, 3, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[6] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 561-570, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[7] Junyu Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[8] Junyu Gao, Wei Lin, Bin Zhao, Dong Wang, Chenyu Gao, and Jun Wen. \\(\\mathbf{C}^3\\) framework: An open-source pytorch code for crowd counting. arXiv preprint arXiv:1907.02724, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.677, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[9] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.734, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[10] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and S.-H. Gary Chan. Target-agnostic Source-free Domain Adaptation for Regression Tasks. In Proceedings of 40th IEEE International Conference on Data Engineering (ICDE'24), Utrecht, Netherlands, 13-16 May 2024 (to appear). IEEE. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[11] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Ali Al-Maadeed, Nasir M. Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In European Conference on Computer Vision, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[12] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Stylenormalization and restitution for domain generalization and" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.145, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "adaptation. IEEE Transactions on Multimedia, 24:3636-3651, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[13] Hyeonseong Kim, Yoonsu Kang, Changgyoon Oh, and Kuk-Jin Yoon. Single domain generalization for lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17587-17598, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[14] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4350, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[15] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[16] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[17] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[18] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[20] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6141-6150, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[21] Lucas Mansilla, Rodrigo Echeveste, D.H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[22] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15529-15539, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 2, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28033" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[28] Vishwanath A. Sindagi and Vishal M. Patel. Cnn-based cascaded multi-task learning of high-level prior and density estimation for crowd counting. 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), pages 1-6, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.47, + 0.336 + ], + "angle": 0, + "content": "[29] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2594-2609, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[30] Leslie N. Smith and Nicholay Topin. Super-convergence: very fast training of neural networks using large learning rates. In *Defense + Commercial Sensing*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[31] Qingyu Song, Changan Wang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Jian Wu, and Jiayi Ma. To choose or to fuse? scale selection for crowd counting. The Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21), 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.525 + ], + "angle": 0, + "content": "[32] Maryam Sultana, Muzammal Naseer, Muhammad Haris Khan, Salman Khan, and Fahad Shahbaz Khan. Self-distilled vision transformer for domain generalization. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3068-3085, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. ArXiv, abs/1607.08022, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.47, + 0.64 + ], + "angle": 0, + "content": "[34] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.47, + 0.725 + ], + "angle": 0, + "content": "[35] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xiansheng Hua. Meta convolutional neural networks for single domain generalization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4672-4681, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[36] Jia Wan, Ziquan Liu, and Antoni B. Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1974-1983, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.842 + ], + "angle": 0, + "content": "[37] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai. Distribution matching for crowd counting. In Advances in Neural Information Processing Systems, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[38] Mingjie Wang, Hao Cai, Xianfeng Han, Jun Zhou, and Minglun Gong. Stnet: Scale tree network with multi-level auxilator for crowd counting. IEEE Transactions on Multimedia, 25:2074-2084, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8198-8207, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[40] Qi Wang, Tao Han, Junyu Gao, and Yuan Yuan. Neuron linear transformation: Modeling the domain shift for crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 33(8):3238-3250, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[41] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[42] Kaiqiang Xu, Xinchen Wan, Hao Wang, Zhenghang Ren, Xudong Liao, Decang Sun, Chaoliang Zeng, and Kai Chen. Tacc: A full-stack cloud computing infrastructure for machine learning tasks. arXiv preprint arXiv:2110.01556, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[43] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI Conference on Artificial Intelligence, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[44] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-841, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[45] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[46] Yuxiang Zhang, Wei Li, Weidong Sun, Ran Tao, and Qian Du. Single-source domain expansion network for cross-scene hyperspectral image classification. IEEE Transactions on Image Processing, 32:1498-1512, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[47] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[48] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[49] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Su Zixian, Yao Kai, Yang Xi, Wang Qiufeng, Sun Jie, and Huang Kaizhu. Rethinking data augmentation for single-source domain generalization in medical image segmentation. In AAAI, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28034" + } + ] +] \ No newline at end of file diff --git a/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_origin.pdf b/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..398c78865fe4db7b46155e6fe6449121737e4ec8 --- /dev/null +++ b/2024/Single Domain Generalization for Crowd Counting/f370ad4c-18eb-4cb3-987b-d53d7152a8e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:909c6046abc977e71526d469d27cd32e63d8c8fc4bc650ecfd14415898d1d067 +size 2468168 diff --git a/2024/Single Domain Generalization for Crowd Counting/full.md b/2024/Single Domain Generalization for Crowd Counting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d81175df8cb81892fc713d4a2893919286b90ab --- /dev/null +++ b/2024/Single Domain Generalization for Crowd Counting/full.md @@ -0,0 +1,410 @@ +# Single Domain Generalization for Crowd Counting + +Zhuoxuan Peng, S.-H. Gary Chan +The Hong Kong University of Science and Technology +{zpengac, gchan}@cse.ust.hk + +# Abstract + +Due to its promising results, density map regression has been widely employed for image-based crowd counting. The approach, however, often suffers from severe performance degradation when tested on data from unseen scenarios, the so-called "domain shift" problem. To address the problem, we investigate in this work single domain generalization (SDG) for crowd counting. The existing SDG approaches are mainly for image classification and segmentation, and can hardly be extended to our case due to its regression nature and label ambiguity (i.e., ambiguous pixel-level ground truths). We propose MPCount, a novel effective SDG approach even for narrow source distribution. MPCount stores diverse density values for density map regression and reconstructs domain-invariant features by means of only one memory bank, a content error mask and attention consistency loss. By partitioning the image into grids, it employs patch-wise classification as an auxiliary task to mitigate label ambiguity. Through extensive experiments on different datasets, MPCount is shown to significantly improve counting accuracy compared to the state of the art under diverse scenarios unobserved in the training data characterized by narrow source distribution. Code is available at https://github.com/Shimmer93/MPCount. + +# 1. Introduction + +Image-based crowd counting is to estimate the number of people (or objects) in an image. The recent mainstream counting approach is density map regression, where a density map with continuous pixel value is predicted from the input image, and the total crowd count is obtained by summing all these pixel values. During training, the ground-truth density map is generated by smoothing point annotations given by the locations of human heads. Density map regression has been demonstrated to achieve encouraging results in crowd counting. + +Existing training models predominantly assume that the testing data share the same source distribution, i.e., fully supervised learning. Such assumption, however, often breaks + +![](images/29a0596334f283546466ccc996ea121ca039dfdefdffa00c40826bc036a444b6.jpg) +(a) Source domain + +![](images/e1c2b5b41a74ccea9a425f38d9de2b82ffb2bea1af72a4e8560dc10fa6e0207a.jpg) +gain + +![](images/ba96caa0eec5e34538c24c910bff308c00da846361b10ea634a546f5f975ea88.jpg) +(b) Target image + +![](images/640bf12c0e782edc061bb92537af332efa631d01afdf37f90fde7c891cd21219.jpg) +(c) Ground-truth + +![](images/cb179390b3f8de5891ec7a74420c5232187dbbe1a59a06b74312c31857fd38c4.jpg) +(d) DCCUS + +![](images/e8f1c619400494e6ac391964186cf6a6f5fe7102e55efd18563be023d1d99d02.jpg) +(e) MPCount (Ours) +Figure 1. Single domain generalization for crowd counting. (a) Sample images from a single source domain $S$ with narrow distribution. (b) A challenging target image from an unseen scenario. (c) The ground-truth density map. (d) The predicted density map from the previous work DCCUS [6] trained on $S$ . (e) The predicted density map from our MPCount trained on $S$ . MPCount achieves much lower counting error than DCCUS and makes better predictions in the cropped area against the domain shift. + +in practice due to changes in camera position, weather condition, etc. Such deviation from the training data (i.e. source domain) to unseen testing data (i.e. target domain) is known as domain shift, which severely degrades the performance of deep learning models in real-world deployment. + +To mitigate the domain shift problem, domain adaptation (DA) approaches have been proposed to transfer knowledge from the source domain to the target one, typically by finetuning pre-trained models with target domain data [2, 10, 39, 44]. Despite promising results, the fine-tuning process of DA is usually laborious. Furthermore, target domain data may not be easily available in reality. + +In contrast to DA, domain generalization (DG) aims to generalize models to any unobserved target domain without the need for target data. As data and model fine-tuning are not necessary at the target domain, DG is more convenient to be deployed in unseen or unpredictable domain. We consider in this work the challenging case of single domain generalization (SDG) for image-based crowd counting, where only one source domain, likely of narrow distribution (Fig. 1a), is used for training. + +A common general SDG technique is to construct + +domain-invariant features from different augmented versions of an image. Such approach has been shown promising for various tasks [5, 13, 34]. Independent of data augmentation, many other SDG approaches have been proposed for classification and segmentation to further boost performance. Unfortunately, most of them cannot be extended to crowd counting, a regression task without category information. Moreover, while labels in classification and segmentation tasks are precise and domain-invariant, the labels for density-based crowd counting are usually ambiguous. Such "label ambiguity" stems from the fact that the density values are generated from point annotations, and hence human heads and backgrounds may be assigned to similar value under different scenarios. + +The recent work of [6] achieves DG for crowd counting based on dividing the source into multiple sub-domains. While encouraging, it requires a rather broad source domain distribution. We propose, for the first time, an SDG approach for crowd counting called MPCount without sub-domain division, thereof effective and extensible to narrow source distribution. We highlight its performance as compared with a state-of-the-art scheme in Fig. 1. MPCount achieves its superior performance with the following two novel components: + +- An attention memory bank (AMB) to tackle density regression: MPCount employs only one attention memory bank (AMB) that takes a pair of features encoded from an crowd image and its data-augmented version as input. To cover continuous density values for regression with a finite memory size, AMB reconstructs each feature vector as an attention over memory vectors. These memory vectors learn domain-invariant representations from the pair of features of different styles but similar content (i.e. crowd densities). We present the content error mask (CEM) to eliminate domain-related content from input features, as characterized by a large discrepancy between instance-normalized feature element pairs. Furthermore, the newly employed attention consistency loss (ACL) enforces the similarity of attention scores produced from input features, ensuring the consistency of memory vectors. +- Patch-wise classification $(PC)$ to tackle label ambiguity: To address the label ambiguity, MPCount employs a novel auxiliary task called patch-wise classification (PC). In this task, each crowd image is evenly divided into fixed-size square patches, say $16 \times 16$ , classified into two classes, i.e., containing human heads or not. The density values in the areas classified as without head are filtered out during density regression, so that the pixel-level ambiguity is overcome by coarser but more accurate patchwise binary labels. + +To validate MPCount, we conduct extensive experiments on various crowd counting datasets including ShanghaiTech A (SHA) & B (SHB) and JHU-Crowd++. We introduce + +a new challenging setting of narrow source distribution, where only images of the same category, such as "snow" (SN) and "fog/haze" (FH) in JHU-Crowd++, are used in training. We demonstrate that MPCount achieves excellent performance not only on traditional inter-dataset benchmarks, but also under our newly introduced narrow source setting. Let $S \rightarrow T$ denote the case where the source domain is $S$ and the target domain is $T$ . As compared to the state of the art, MPCount is shown to reduce significantly counting error by $21.8\%$ on SN $\rightarrow$ FH, $18.6\%$ on FH $\rightarrow$ SN, $18.2\%$ on SHB $\rightarrow$ SHA and $9.5\%$ on SHA $\rightarrow$ SHB. + +# 2. Related Works + +# 2.1. Fully Supervised Crowd Counting + +The mainstream approach to crowd counting is via density map regression, where each pixel of an image is assigned a count value, and the summation of these values yields an estimate of the total number of people. Most methods focus on improving counting accuracy under the fully supervised setting by employing novel network designs [1, 16, 17, 31, 45] or loss functions [20, 26, 36, 37]. While encouraging results have been shown, their performance usually drop significantly when evaluated on out-of-distribution data, due to their limited generalization ability. Besides, several works [4, 22, 38] have proposed auxiliary tasks in crowd counting for certain purposes. However, these tasks still require pixel-level predictions and contribute little to tackling label ambiguity. + +# 2.2. Domain Adaptation (DA) for Crowd Counting + +Domain adaptation (DA) has been widely studied in crowd counting to address domain shift by adapting source domain information to a particular target domain. While a few methods study supervised DA where labeled data from target domain are available [40, 44], most approaches tackle unsupervised DA, utilizing only unlabeled target domain data [2, 7, 18, 39, 41, 47]. Despite the notable advancements, DA methods normally require data from the target domain, which may not be easily available in practice. + +# 2.3. Single Domain Generalization (SDG) + +Domain Generalization (DG) approaches are designed to train a network with generalization ability solely utilizing source domain data, and single domain generalization (SDG) is the special case when only one source domain is available. There are various techniques for SDG, including 1) adversarial data generation [15, 25, 46, 50], 2) feature normalization/whitening transformation [5, 12, 23, 24, 43] and 3) domain-general network design, such as a convolution layer [35], vision transformer [32] and memory bank [3]. Among various SDG methods, a common general technique is to simulate domain shift with data augmenta + +tion and extract domain-invariant information from different augmented versions of features [5, 13, 43]. While SDG has attracted widespread attention in classification and segmentation, it is still a nascent topic in crowd counting due to its regression nature and label ambiguity. + +DCCUS [6] studies SDG for crowd counting. In this method, the source domain is dynamically clustered into sub-domains simulated as meta-training and -test sets in a meta-learning strategy. Two types of memory modules and several losses are proposed to distinguish and record domain-invariant and -related information. DCCUS designs the memory mechanism specifically for regression, but the problem of ambiguous labels is not considered. Moreover, when the source domain follows a narrow distribution, the sub-domain division process may experience reduced effectiveness and affect the generalization ability of the counting model. MPCount, without sub-domain division, employs novel patch-wise classification to address the label ambiguity problem. Additionally, a single memory bank for regression is introduced without the need for sub-domain partitioning, enabling MPCount to generalize well even in source domains of narrow distributions. + +# 3. MPCount + +This section provides a detailed description of our MPCount scheme. We begin by reviewing the foundational concepts of crowd counting in Sec. 3.1. Next, in Sec. 3.2, we present a comprehensive overview of the entire scheme. We then elaborate the attention memory bank (Sec. 3.3) designed for density regression, followed by the content error mask (Sec. 3.4) and attention consistency loss (Sec. 3.5). After that, we discuss the patch-wise classification (Sec. 3.6) proposed to address the challenge of ambiguous labels. Finally, we detail the overall training loss in Sec. 3.7. + +# 3.1. Crowd Counting Preliminaries + +First, we review the preliminary knowledge of deep learning-based crowd counting via density map regression. Our objective is to train a neural network $\mathcal{N}$ with parameter $\theta$ that takes in an input image $\mathcal{I} \in \mathbb{R}^{H \times W \times 3}$ and outputs a density map $\hat{\mathcal{D}} = \mathcal{N}(\mathcal{I};\theta)$ of size $H \times W$ . The estimated count $\hat{c}$ is the sum of all pixel-wise density values in $\hat{\mathcal{D}}$ , i.e., + +$$ +\hat {c} = \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \hat {\mathcal {D}} _ {i j}. \tag {1} +$$ + +Generation of ground-truth density maps usually follows the method in [28]. Given the human head point annotations $\mathcal{H}$ where each $h\in \mathcal{H}$ is the coordinate of a head location, the ground-truth density map $\mathcal{D}\in \mathbb{R}^{H\times W}$ is calculated as the sum of the 2D Gaussian filter applied to each head coordinate $h$ , i.e., + +$$ +\mathcal {D} = \sum_ {h \in \mathcal {H}} \delta (x - h) \times G _ {\sigma} (x), \tag {2} +$$ + +where $\delta (\cdot)$ is the discrete delta function, and $G_{\sigma}(\cdot)$ is the Gaussian kernel with a fixed variance $\sigma$ + +A common objective function to supervise the network $\mathcal{N}$ can be written as + +$$ +\mathbb {L} (\theta) = | | \mathcal {D}, \hat {\mathcal {D}} | | _ {2} ^ {2}. \tag {3} +$$ + +# 3.2. Scheme Overview + +The overall structure of our method is illustrated in Fig. 2. During training, photometric transformations are applied to the original image $\mathcal{I}^{ori}$ to produce an augmented version $\mathcal{I}^{aug}$ . The feature extractor encodes them into features $\mathcal{F}^{ori}$ and $\mathcal{F}^{aug}$ , and a content error mask (CEM) $\mathcal{M}$ is applied to them. The masked features are reconstructed into $\tilde{\mathcal{F}}^{ori}$ and $\tilde{\mathcal{F}}^{aug}$ via an attention memory bank (AMB) and passed to the density head for crowd density prediction. Meanwhile, the highest-level features encoded by the feature extractor, $\mathcal{Z}^{ori}$ and $\mathcal{Z}^{aug}$ , are fed into the patch-wise classification (PC) head, and the predicted PC maps (PCMs) $\hat{\mathcal{C}}^{ori}$ and $\hat{\mathcal{C}}^{aug}$ are binarized and resized to serve as a mask to filter out areas without human heads in the estimated density maps $\hat{\mathcal{D}}^{ori}$ and $\hat{\mathcal{D}}^{aug}$ . The final predictions $\mathcal{D}'^{ori}$ and $\mathcal{D}'^{aug}$ are two density maps with certain areas masked out by the PCMs. The entire training process is supervised under a combination of density losses, PC losses and the attention consistency loss (ACL). The inference process solely employs $\mathcal{I}^{ori}$ as input, and outputs a single density map $\mathcal{D}'^{ori}$ . + +# 3.3. Attention Memory Bank (AMB) + +The attention memory bank (AMB) is designed to automatically learn domain-invariant features for density regression. [14] proposes a memory bank where memory vectors are updated individually, each corresponding to a certain category. However, this design cannot be applied to regression tasks, since a finite number of memory vectors cannot cover continuous density values. Inspired by [6], we reconstruct each feature vector as an attention over memory vectors, so that any density value may be represented as a linear combination of representative values in the memory bank. + +The AMB $\mathcal{V} \in \mathbb{R}^{M \times C}$ consists of $M$ memory vectors of dimension $C$ . Given a flattened input feature map $\mathcal{F} \in \mathbb{R}^{HW \times C}$ , we first compute the attention scores $\mathcal{A}$ between the feature $\mathcal{F}$ (query) and the memory $\mathcal{V}$ (key). Then the reconstructed feature map $\tilde{\mathcal{F}}$ is calculated as a linear combination of $\mathcal{V}$ (value) with $\mathcal{A}$ as weights. The reconstruction process can be summarized as follows: + +$$ +\mathcal {A} (\mathcal {F}, \mathcal {V}) = \operatorname {S o f t m a x} \left(\frac {\mathcal {F V} ^ {T}}{\sqrt {C}}\right), \tag {4} +$$ + +![](images/18074bfdcd490ed255445b4ed43ad1a70629440fb2752c5b0bd7c4fb864b3ea3.jpg) +Figure 2. The overall training pipeline of our proposed MPCount. All identical modules in this diagram share the same weights. In the encoder-decoder structure, a higher level indicates a deeper feature. + +and + +$$ +\tilde {\mathcal {F}} = \mathcal {A} \mathcal {V}. \tag {5} +$$ + +While [6] relies on multiple memory banks corresponding to different sub-domains to distinguish domain-invariant and -related information, we only employ a single AMB and propose the novel CEM (Sec. 3.4) and ACL (Sec. 3.5) to ensure that it stores domain-invariant representations. To this end, our feature reconstruction mechanism can work without sub-domain division, thereof is effective even with a narrow source distribution. + +# 3.4. Content Error Mask (CEM) + +CEM is proposed to guarantee the resemblance of content contained in the pair of input features by excluding possible domain-related content information. A common technique to disentangle style and content in a feature is instance normalization (IN) [33], where the feature statistics (channel-wise mean and variance) are considered to contain style information, while the instance normalized features retain content information. Based on this idea, we assume that the discrepancy of instance normalized features reflect effects of domain shift on feature content and may mislead AMB to learn domain-related information. Therefore, we filter out elements in the input features with difference between their instance normalized versions above a certain threshold, resulting in a pair of features with similar content information for domain-invariant feature reconstruction. + +More specifically, given the features extracted from the original and augmented images, $\mathcal{F}^{ori}$ and $\mathcal{F}^{aug} \in \mathbb{R}^{H \times W \times C}$ , we define the content error mask $\mathcal{M} \in \mathbb{R}^{H \times W \times C}$ as follows: + +$$ +\mathcal {M} _ {i j k} = \left\{ \begin{array}{l l} 1 & \text {i f} | \mathrm {I N} (\mathcal {F} ^ {\text {o r i}}) _ {i j k} - \mathrm {I N} (\mathcal {F} ^ {\text {a u g}}) _ {i j k} | \leq \alpha \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {6} +$$ + +where $\alpha$ is the threshold indicating whether the error value reflects possible inconsistency of content information caused by domain shift. + +Given $\mathcal{F}$ to be one of $\mathcal{F}^{\text{ori}}$ and $\mathcal{F}^{\text{aug}}$ , we filter out the domain-related content information by + +$$ +\mathcal {F} ^ {\prime} = \operatorname {D r o p o u t 2 D} (\mathcal {F} \odot \mathcal {M}). \tag {7} +$$ + +The random 2D dropout is applied to $\mathcal{F}'$ to prevent the memory from relying solely on certain channels. + +# 3.5. Attention Consistency Loss (ACL) + +The style information of $\mathcal{F}^{ori}$ and $\mathcal{F}^{aug}$ preserved in Sec. 3.4 may produce a large discrepancy on their attention distributions over the memory vectors in AMB even if they contain identical content information. To ensure that each memory vector consistently store specific domain-invariant representations, we propose the attention consistency loss (ACL) to enforce the similarity of attention scores produced from $\mathcal{F}^{ori}$ and $\mathcal{F}^{aug}$ . + +We consider $\mathcal{A}$ in Eq. 4 as a distribution, and the ACL $\mathbb{L}_{con}$ is calculated as the distance between the distribution $\mathcal{A}^{ori}$ produced by $\mathcal{F}^{ori}$ and $\mathcal{A}^{aug}$ produced by $\mathcal{F}^{aug}$ . Here the simple Euclidean distance is selected as the distance measure due to its computational stability during training: + +$$ +\mathbb {L} _ {\text {c o n}} = \left\| \mathcal {A} ^ {\text {a u g}}, \mathcal {A} ^ {\text {o r i}} \right\| _ {2} ^ {2}. \tag {8} +$$ + +![](images/6d595b4e7315ec7bb19574ecafa9840a473d3ec2e87411a3217473df3fec3771.jpg) +(a) Density Map + +![](images/cc336287be96e075878223a874b3f55febe4f4805756b5982f9e8f8f33a487d3.jpg) +(b) PCM +Figure 3. Illustration of label ambiguity and how PCM tackles it. Point A is a typical example which belongs to a human head but is assigned the density value 0 as the background in the density map due to varying head sizes. In the PCM, the patch covering A is correctly classified as containing heads, thanks to the patch-level accuracy of the labels. + +# 3.6. Patch-wise Classification (PC) + +As discussed in Sec. 3.1, ground-truth density maps are usually calculated only based on point annotations. In different scenarios, pixels of a human head and the background might be assigned the same density value due to various head sizes (Fig. 3a), violating the common assumption that labels are accurate and domain-invariant, thereof resulting in the ambiguous labels in crowd counting. + +To tackle the label ambiguity, we propose a novel auxiliary task named patch-wise classification (PC) and its corresponding supervision signal, patch-wise classification map (PCM), as illustrated in Fig. 3b. A PCM divides an image evenly into $P \times P$ ( $P = 16$ empirically) patches and classifies each patch into two categories: containing human heads or not. Such patch-level predictions mitigate the uncertainty in pixel-level density maps by providing coarser but more accurate patch-level information. + +In practice, a ground truth $\mathbf{PCM}\mathcal{C}$ with patch size $P$ can be calculated as: + +$$ +\mathcal {C} _ {i j} = 1 - \delta \left(\sum_ {k = i P} ^ {(i + 1) P} \sum_ {l = j P} ^ {(j + 1) P} \mathcal {D} _ {k l}\right), \tag {9} +$$ + +where $\mathcal{D}$ is the ground truth density map and $\delta (\cdot)$ is the discrete delta function. + +Our model predicts a PCM $\hat{C}$ supervised by the binary cross entropy (BCE) loss: + +$$ +\mathbb {L} _ {c l s} = \operatorname {B C E} (\mathcal {C}, \hat {\mathcal {C}}). \tag {10} +$$ + +Subsequently, $\hat{\mathcal{C}}$ is binarized and resized to obtain $\mathcal{C}'$ which matches the dimensions of the predicted density map $\hat{\mathcal{D}}$ . $\mathcal{C}'$ is then used to mask the regions classified without crowds in $\hat{\mathcal{D}}$ , resulting in the final density map $\mathcal{D}'$ : + +$$ +\mathcal {D} ^ {\prime} = \hat {\mathcal {D}} \odot \mathcal {C} ^ {\prime}. \tag {11} +$$ + +Finally, the density regression head is supervised with the ground truth density map $\mathcal{D}$ , and the common Euclidean distance in Eq. (3) is used as the objective function: + +$$ +\mathbb {L} _ {d e n} = \left\| \mathcal {D}, \mathcal {D} ^ {\prime} \right\| _ {2} ^ {2}. \tag {12} +$$ + +# 3.7. Model Training + +The overall training loss is a combination of density losses $\mathbb{L}_{den}^{ori},\mathbb{L}_{den}^{aug}$ , PC losses $\mathbb{L}_{cls}^{ori},\mathbb{L}_{cls}^{aug}$ and the attention consistency loss $\mathbb{L}_{con}$ , which can be written as + +$$ +\mathbb {L} = \mathbb {L} _ {d e n} ^ {o r i} + \mathbb {L} _ {d e n} ^ {a u g} + \lambda_ {c l s} \left(\mathbb {L} _ {c l s} ^ {o r i} + \mathbb {L} _ {c l s} ^ {a u g}\right) + \lambda_ {c o n} \mathbb {L} _ {c o n}, \tag {13} +$$ + +where $\lambda_{cls}$ and $\lambda_{con}$ are weighting parameters to balance different loss terms. + +# 4. Illustrative Experimental Results + +In this section, we first introduce the datasets used in experiments in Sec. 4.1. Next, the implementation details are specified in Sec. 4.2, then evaluation metrics are described in Sec. 4.3. We present the results compared with state of the art in Sec. 4.4. Finally, we conduct ablation studies and other additional analysis in Sec. 4.5. + +# 4.1. Datasets + +We evaluate our method on four mainstream crowd counting datasets: ShanghaiTech Part A & B, UCF-QNRF and JHU-Crowd++. + +- ShanghaiTech [45] is composed of two parts, SHA (A) and SHB (B). SHA contains 300 training images and 182 testing images, while SHB contains 400 training images and 316 testing images. Images in SHA are collected from the Internet and feature highly crowded scenes. In contrast, SHB is captured from several streets in Shanghai, and its images exhibit crowd densities generally lower than those in SHA. +- UCF-QNRF (Q) [11] consists of 1201 training images and 334 testing images. It is a challenging dataset with a large range of crowd densities, scenes, viewpoints and lighting conditions. +- JHU-Crowd++ [29] is a large-scale dataset containing 4372 images, with 2722, 500 and 1600 images in the training, validation and testing set, respectively. In addition, image-level labels are also provided in this dataset, including 16 types of scenes and 4 types of weather conditions. For scene annotations, there are 879 images labeled "stadium" (SD) and 573 labeled "street" (ST), which are utilized as two source domains. For weather annotations, we use 201 images labeled "snow" (SN) and 168 images labeled "fog/haze" (FH) as source domains. In each domain, $80\%$ of the images are selected as the training set while the rest $20\%$ are for testing. Data with the same label belong to a narrower distribution than mainstream datasets, thus are more challenging for SDG. Illustrative samples in these scene-specific and weather-specific domains are available in Fig. 4. +Let $S \to \mathcal{T}$ denote the case where $S$ is the source domain and $\mathcal{T}$ is the target domain. The datasets are utilized + +![](images/6ea710ba9842e1382414ebbe96cfe63aa38f448d0c1f3015e982e1a7aef8150b.jpg) +(a) Stadium + +![](images/02541f207d512fd99e99c6ccf724a287a7b6d932b25ac5d02e65341963169652.jpg) +(b) Street + +![](images/64084dfcd33edd730f2bf68d2bfcb768ba3b9460ca2a401766734fdf0f020e46.jpg) +(c) Snow +Figure 4. Sample images with different instance-level labels in JHU-Crowd++. + +![](images/88f90215da608759fe740d88269e5fc4c60414cffb34bad4c1f255229c093065.jpg) +(d) Fog/Haze + +in experiments under two distinct types of settings: 1) An entire dataset is regarded as a single domain: $\mathrm{A} \rightarrow \mathrm{B} / \mathrm{Q}$ , $\mathrm{B} \rightarrow \mathrm{A} / \mathrm{Q}$ and $\mathrm{Q} \rightarrow \mathrm{A} / \mathrm{B}$ ; 2) A dataset is partitioned into subsets according to image-level labels, and one subset constitutes a single domain: $\mathrm{SD} \leftrightarrow \mathrm{SR}$ and $\mathrm{SN} \leftrightarrow \mathrm{FH}$ . + +# 4.2. Implementation + +We adopt VGG16-BN [27] as the feature extractor in our model. The density regression head is a single convolution layer with $1 \times 1$ filters, while the classification head consists of a $3 \times 3$ and a $1 \times 1$ convolution layers. Batch normalization is employed after all convolution layers except the last ones. For data augmentation, we apply three types of photometric transformations, color jittering, Gaussian blurring and sharpening. We also randomly crop image patches with a size of $320 \times 320$ and adopt random horizontal flipping to both the original and augmented images simultaneously. We select AdamW [19] as the optimizer and OneCycleLR [30] as the learning rate scheduler with maximum learning rate set to 1e-3 and a maximum epoch of 300. The memory size $M$ is set to 1024 and the dimension $C$ is 256. We use a content error threshold $\alpha$ of 0.5, and the loss weights $\lambda_{cls}$ and $\lambda_{con}$ are both set to 10. + +# 4.3. Evaluation Metrics + +We evaluate our method with mean absolute error (MAE) and mean squared error (MSE), defined as follows: + +$$ +M A E = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| c _ {i} - \hat {c} _ {i} \right|, M S E = \sqrt {\frac {1}{N} \sum_ {i = 1} ^ {N} \left(c _ {i} - \hat {c} _ {i}\right) ^ {2}}, \tag {14} +$$ + +where $N$ is the number of testing images, $c_{i}$ is the ground truth count of the $i$ -th image and $\hat{c}_i$ is the predicted count. Lower values of both metrics indicate better performance. + +# 4.4. Comparison with State of the Art + +In this subsection, we compare our MPCount with state-of-the-art methods on different benchmarks, as illustrated + +![](images/c1c9c23b03cc636a03c755ed44e872d83be0deeff9cd760032c06f20f1286d61.jpg) +(a) Image +Figure 5. Visualization results of DCCUS [6] and MPCount under different DG settings. First row: A $\rightarrow$ B; second row: SN $\rightarrow$ FH; third row: B $\rightarrow$ A. + +![](images/b17371f220a1d4f8b2d07c684197dd6e847e7f6446d45b9fc2c615fc9b1ace6a.jpg) +(b) GT + +![](images/b0a754df6ec173baa4b9906bab7d18b6614c99bf9bbfc8b548b6554199f1f9df.jpg) +(c) DCCUS + +![](images/e05d7b6691c0a3e221adb9d4fb144886576bb6b67ff9683159e44a8e6088e98b.jpg) +(d) MPCount + +in Fig. 5. Tab. 1 shows the results on $\mathrm{A} \rightarrow \mathrm{B} / \mathrm{Q}$ , $\mathrm{B} \rightarrow \mathrm{A} / \mathrm{Q}$ and $\mathrm{Q} \rightarrow \mathrm{A} / \mathrm{B}$ . The selected baselines can be divided into three categories: + +- Fully supervised crowd counting: BL [20], DM-Count [37], SASNet [31], ChfL [26] and MAN [17]. +- Domain adaptation for crowd counting: RBT [18], C²MoT [41], FGFD [47], DAOT [48] and FSIM [49]. +- Single domain generalization: IBN [23], SW [24], ISW [5], DG-MAN [21] and DCCUS [6]. + +All results are copied from previous papers except IBN, SW and ISW, which are feature normalization/whitening-based SDG methods originally designed for classification or segmentation and adapted to crowd counting by us. Domain adaptation methods are listed only for reference, as data from target domains are usually visible during adaptation. + +Our MPCount outperforms all the DG methods under most of these settings, including an significant error reduction by $18.2\%$ on $\mathrm{B} \rightarrow \mathrm{A}$ . It is notable that MPCount exhibits outstanding performance in some cases even compared to DA methods, providing compelling evidence for the effectiveness of our design. + +In Tab. 2, we further conduct experiments on scene-specific domains $\mathrm{SD} \leftrightarrow \mathrm{SR}$ and weather-specific domains $\mathrm{SN} \leftrightarrow \mathrm{FH}$ . BL [20], MAN [17], DAOT [48] and DCCUS [6] are trained using their released code, while IBN [23], SW [24] and ISW [5] are adapted by ourselves. We observe that the normalization/whitening-based methods cannot achieve satisfactory results, possibly because useful information is eliminated from feature statistics. DCCUS also obtains worse results than fully supervised methods, reflecting that the clustering procedure may fail to produce meaningful sub-domains when the source domain distribution is narrow. In contrast, our MPCount still performs the best among all tested methods, demonstrating its superiority in challenging conditions with narrow domain distributions. + +
Source → TargetA → BA → QB → AB → QQ → AQ → B
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]✘✘15.925.8166.7287.6138.1228.1226.4411.0----
DMCount [37]✘✘23.134.9134.4252.1143.9239.6203.0386.173.4135.114.327.5
SASNet [31]✘✘21.333.2211.2418.6132.4225.6273.5481.373.9116.413.022.1
ChfL [26]✘✘18.729.1122.3218.0121.3200.8197.1357.968.7118.514.027.0
MAN [17]✘✘22.132.8138.8266.3133.6255.6209.4378.867.1122.112.522.2
RBT [18]✓✘13.429.3175.0294.8112.2218.2211.3381.9----
C2MoT [41]✓✘12.421.1125.7218.3120.7192.0198.9368.0----
FGFD [47]✓✘12.723.3124.1242.0123.5210.7209.7384.770.2118.412.520.3
DAOT [48]✓✘10.918.8113.9215.6----67.0128.411.319.6
FSIM [49]✓✘11.119.3105.3191.1120.3202.6194.9324.566.8111.511.019.7
IBN [23]✘✓19.130.8280.2561.0125.9202.3183.5317.4105.9174.616.025.5
SW [24]✘✓20.230.4285.5431.0126.7193.8200.7333.2102.4168.819.032.9
ISW [5]✘✓23.937.7215.6399.7156.2291.5263.1442.283.4136.022.134.7
DG-MAN [21]✘✓17.328.7129.1238.2130.7225.1182.4325.8----
DCCUS [6]✘✓12.624.6119.4216.6121.8203.1179.1316.267.4112.812.120.9
MPCount (Ours)✘✓11.419.7115.7199.899.6182.9165.6290.465.5110.112.324.1
+ +Table 1. Comparison with the state-of-the-art methods on SHA (A), SHB (B) and UCF-QNRF (Q). Domain adaptation methods are only for reference since data in target domain might be used. + +
Source → TargetSD → SRSR → SDSN → FHFH → SN
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]XX42.179.0262.71,063.948.1129.5343.8770.5
MAN [17]XX45.179.0246.1950.838.168.0445.0979.3
DAOT [48]✓X45.388.0278.71,624.342.373.0151.6273.9
IBN [23]X✓92.2178.0318.11,420.4109.7267.7491.81,110.4
SW [24]X✓110.3202.4312.61,072.4131.5306.6381.3825.0
ISW [5]X✓108.1212.4385.91,464.8151.6365.7276.6439.8
DCCUS [6]X✓90.4194.1258.11,005.954.5125.8399.7945.0
MPCount (Ours)X✓37.470.1218.6935.931.355.0216.3421.4
+ +Table 2. Comparison with the state of the art on data in JHU-Crowd++ with labels "Stadium"(SD), "Street"(SR), "Snow"(SN) and "Fog/Haze"(FH). Domain adaptation methods are only for reference since data in target domain might be used. + +# 4.5. Ablation Studies + +In this section, we conduct ablation studies and other analysis under the setting $\mathrm{A} \to \mathrm{B} / \mathrm{Q}$ . + +Effects of model components. We start from a simple encoder-decoder structured baseline model and add or remove each component individually to validate its effectiveness. The experimental results are displayed in Tab. 3. + +- Attention Memory Bank: After adding AMB, the performance consistently improves on both target domains, as shown in Tab. 3. This validates that AMB can effectively help the model generalize by reconstructing domain-invariant features for density regression. +- Content Error Mask & Attention Consistency Loss: Next, we test the effects of CEM and ACL on AMB. The results + +in Tab. 3 show that both components generally improves the performance, especially when PC is present. + +- Patch-wise Classification: Finally, models with PC outperforms their counterparts without PC, as visible in Tab. 3. This suggests that such an auxiliary task is useful for DG by compensating for ambiguous pixel-level labels with accurate patch-level labels. + +Effects of $\alpha$ . The threshold of CEM $\alpha$ mentioned in Eq. (6) controls the Portion of Diminished Elements (PDE) in features. According to the results in Tab. 4, $\alpha = 0.5$ produces the best performance under all settings, with a PDE of $10.5\%$ . It is also indicated that using a too large value of $\alpha$ (0.9) may be inadequate to filter out domain-related information, while a too small value (0.1) would let useful information eliminated. + +
ComponentsA → BA → Q
AMBCEMACLPCMAEMSEMAEMSE
21.639.1200.9325.5
✓18.932.0176.3292.8
✓✓15.325.5156.8232.7
✓✓16.928.7172.5281.6
✓✓✓14.124.3170.6286.8
✓14.430.5144.2262.6
✓✓13.425.2133.4224.0
✓✓✓13.323.4126.9222.3
✓✓✓✓11.419.7115.7199.8
+ +Table 3. Ablation studies on the model components. + +
ParameterPropertyA → BA → Q
αPDE(%)MAEMSEMAEMSE
0.128.616.725.9157.3338.5
0.314.612.420.8132.1223.6
0.510.511.419.7115.7199.8
0.75.911.923.8123.8214.1
0.93.613.827.2124.7209.6
+ +Table 4. Ablation studies on the error threshold $\alpha$ in Eq. (6). + +
ParametersA → BA → Q
MCMAEMSEMAEMSE
25625615.924.0138.4226.1
51225613.728.1124.4217.2
102425611.419.7115.7199.8
204825612.023.3128.1218.1
102412816.328.6127.4223.3
102425611.419.7115.7199.8
102451214.327.7130.1250.1
1024102414.224.8128.4220.2
+ +Table 5. Ablation studies on the number of memory vectors $M$ and vector size $C$ . + +Effects of $M$ and $C$ . The number of memory vectors $M$ and dimension of each vector $C$ are two important parameters of AMB. To demonstrate their effects independently, we adjust one parameter while keeping the other fixed. The results in Tab. 5 shows that among all variations of parameters, $M = 1024$ , $C = 256$ is the optimal choice. + +Visualizations of CEMs. We present visualizations of CEMs in Fig. 6, each summed along the channel dimension. The results indicate that crowd information is generally more + +![](images/cc23f6061a32fa91dda40a62ab46f563aaba4f86a34f6589342c9313ef33a5a6.jpg) +Figure 6. Visualized CEMs. Yellower means less filtered. + +![](images/356579510cf69056b7b29fbe086087f3ae2c0861eafab076b2e729904daa0ca1.jpg) + +![](images/3f7bba4ed8dfe203fbb76915cacf53ea698797efa04a3c0e367583ad8127845a.jpg) + +![](images/66f46a03c57008583fe811b9e36cf6b47fb421739408fba5dff2c86dedc7a968.jpg) + +![](images/25ca469d8438797335835d5528cd69768902a15075b74d1314a43165ba1681d5.jpg) + +![](images/cd211b7d1a61efbb4e8ee5e944203bbe1c6d47e64ebd1f4718a5d747a1bfc1b4.jpg) + +![](images/89e7d6a0a103ceec615581142a89c86886c40de7dae79a55c7841b53da664fb7.jpg) + +![](images/16386179d7e8aeb5adaa9b80042426adf1b30479adc192702f7e03c605e26c7f.jpg) +(a) Image + +![](images/db91b594668215fbe44b9bbd0cf8c688505ee3c3599c6093a3901998b0521265.jpg) + +![](images/64cdb54bf3a203785067fa227d11be96cf9524aa6b1754a59da2e934415d2210.jpg) +(b) GT + +![](images/eafc32d6f395f3c3fe5ef9060741fc24ec5bc640404929370bd6893758d04452.jpg) + +![](images/1aceeabf9e4476c4590acaaf069a364ba2188429e32b1cbcc2d2d41f001ec97f.jpg) +(c) Predicted + +![](images/51c19f01d875549ace4a2b875b8fd509639a96c7a5c2cef284a2b6d29597a2e4.jpg) + +![](images/559708f2bde012ed0fe22fca3512ac092e69f27b895a31b4783957e0552ef4ce.jpg) +(d) Binarized +Figure 7. Visualization of (a) sample images from SHB, (b) the ground-truth PCMs, (c) PCMs predicted by MPCount trained on SHA and (d) predicted PCMs after binarization. + +sensitive to domain shift than environments. + +Visualizations of PCMs. We visualize the predicted and binarized PCMs from MPCount for qualitative analysis. As shown in Fig. 7, MPCount predicts accurate PCMs that provide reliable information of crowd locations. In the raw predicted PCMs, there are unconfident predictions with low classification scores, which are filtered out in the binarized PCMs. The results verify that PC can effectively mitigate the label ambiguity problem and contribute to a more robust crowd counting model against domain shift. + +# 5. Conclusion + +This paper proposes MPCount to address single domain generalization for crowd counting with possibly narrow source distribution. MPCount tackles two unique challenges, namely, density regression and label ambiguity. We propose an attention memory bank to reconstruct domain-invariant features for regression, with the content error mask eliminating domain-related content information and the attention consistency loss ensuring the consistency of memory vectors. Patch-wise classification is a novel auxiliary task to mitigate the ambiguous pixel-level labels with accurate information, enhancing the robustness of density predictions. Extensive experiments on well-known datasets show that MPCount achieves significantly the best results on various benchmarks by outperforming the state of the art under different unseen scenarios and narrow source domain. Acknowledgement. We would like to thank the Turing AI Computing Cloud (TACC) [42] and HKUST iSING Lab for providing us computation resources on their platform. + +# References + +[1] Haoyue Bai, Hao He, Zhuoxuan Peng, Tianyuan Dai, and S.-H. Gary Chan. CounTr: An End-to-End Transformer Approach for Crowd Counting and Density Estimation. In Proceedings of European Conference on Computer Vision (ECCV) International Workshop on Distributed Smart Cameras, pages 207–222, Tel Aviv, Israel, 2022. Springer Lecture Notes in Computer Science. 2 +[2] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. Proceedings of the AAAI Conference on Artificial Intelligence, 37:259-267, 2023. 1, 2 +[3] Yang Chen, Yu Wang, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. A style and semantic memory mechanism for domain generalization*. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9144-9153, 2021. 2 +[4] Jian Cheng, Haipeng Xiong, Zhiguo Cao, and Hao Lu. Decoupled two-stage crowd counting and beyond. IEEE Transactions on Image Processing, 30:2862-2875, 2021. 2 +[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 2, 3, 6, 7, 1 +[6] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 561-570, 2023. 1, 2, 3, 4, 6, 7 +[7] Junyu Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 2 +[8] Junyu Gao, Wei Lin, Bin Zhao, Dong Wang, Chenyu Gao, and Jun Wen. $\mathbf{C}^3$ framework: An open-source pytorch code for crowd counting. arXiv preprint arXiv:1907.02724, 2019. 1 +[9] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1 +[10] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and S.-H. Gary Chan. Target-agnostic Source-free Domain Adaptation for Regression Tasks. In Proceedings of 40th IEEE International Conference on Data Engineering (ICDE'24), Utrecht, Netherlands, 13-16 May 2024 (to appear). IEEE. 1 +[11] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Ali Al-Maadeed, Nasir M. Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In European Conference on Computer Vision, 2018. 5 +[12] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Stylenormalization and restitution for domain generalization and + +adaptation. IEEE Transactions on Multimedia, 24:3636-3651, 2022. 2 +[13] Hyeonseong Kim, Yoonsu Kang, Changgyoon Oh, and Kuk-Jin Yoon. Single domain generalization for lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17587-17598, 2023. 2, 3 +[14] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4350, 2022. 3 +[15] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 2 +[16] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2018. 2 +[17] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, 2022. 2, 6, 7 +[18] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 2, 6, 7 +[19] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 6 +[20] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6141-6150, 2019. 2, 6, 7 +[21] Lucas Mansilla, Rodrigo Echeveste, D.H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7 +[22] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15529-15539, 2021. 2 +[23] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 2, 6, 7, 1 +[24] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6, 7, 1 +[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 2 + +[26] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2, 6, 7 +[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 6 +[28] Vishwanath A. Sindagi and Vishal M. Patel. Cnn-based cascaded multi-task learning of high-level prior and density estimation for crowd counting. 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), pages 1-6, 2017. 3 +[29] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2594-2609, 2020. 5 +[30] Leslie N. Smith and Nicholay Topin. Super-convergence: very fast training of neural networks using large learning rates. In *Defense + Commercial Sensing*, 2017. 6 +[31] Qingyu Song, Changan Wang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Jian Wu, and Jiayi Ma. To choose or to fuse? scale selection for crowd counting. The Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21), 2021. 2, 6, 7 +[32] Maryam Sultana, Muzammal Naseer, Muhammad Haris Khan, Salman Khan, and Fahad Shahbaz Khan. Self-distilled vision transformer for domain generalization. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3068-3085, 2022. 2 +[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. ArXiv, abs/1607.08022, 2016. 4 +[34] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2 +[35] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xiansheng Hua. Meta convolutional neural networks for single domain generalization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4672-4681, 2022. 2 +[36] Jia Wan, Ziquan Liu, and Antoni B. Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1974-1983, 2021. 2 +[37] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai. Distribution matching for crowd counting. In Advances in Neural Information Processing Systems, 2020. 2, 6, 7 +[38] Mingjie Wang, Hao Cai, Xianfeng Han, Jun Zhou, and Minglun Gong. Stnet: Scale tree network with multi-level auxilator for crowd counting. IEEE Transactions on Multimedia, 25:2074-2084, 2020. 2 + +[39] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8198-8207, 2019. 1, 2 +[40] Qi Wang, Tao Han, Junyu Gao, and Yuan Yuan. Neuron linear transformation: Modeling the domain shift for crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 33(8):3238-3250, 2022. 2 +[41] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 2, 6, 7 +[42] Kaiqiang Xu, Xinchen Wan, Hao Wang, Zhenghang Ren, Xudong Liao, Decang Sun, Chaoliang Zeng, and Kai Chen. Tacc: A full-stack cloud computing infrastructure for machine learning tasks. arXiv preprint arXiv:2110.01556, 2021.8 +[43] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI Conference on Artificial Intelligence, 2022. 2, 3 +[44] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-841, 2015. 1, 2 +[45] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2, 5 +[46] Yuxiang Zhang, Wei Li, Weidong Sun, Ran Tao, and Qian Du. Single-source domain expansion network for cross-scene hyperspectral image classification. IEEE Transactions on Image Processing, 32:1498-1512, 2022. 2 +[47] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 2, 6, 7 +[48] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 6, 7 +[49] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 6, 7 +[50] Su Zixian, Yao Kai, Yang Xi, Wang Qiufeng, Sun Jie, and Huang Kaizhu. Rethinking data augmentation for single-source domain generalization in medical image segmentation. In AAAI, 2023. 2 \ No newline at end of file diff --git a/2024/Single Domain Generalization for Crowd Counting/images.zip b/2024/Single Domain Generalization for Crowd Counting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73dfcf0551e22eb4c552289050777d621e1df6c5 --- /dev/null +++ b/2024/Single Domain Generalization for Crowd Counting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:072e8d437971228277b2da2cdf0373b3ee6e445b5604d0c2741077638bf317b6 +size 715795 diff --git a/2024/Single Domain Generalization for Crowd Counting/layout.json b/2024/Single Domain Generalization for Crowd Counting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a327310963b2d31d73c10e1cbfed4564d82bb68e --- /dev/null +++ b/2024/Single Domain Generalization for Crowd Counting/layout.json @@ -0,0 +1,11568 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 103, + 455, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 455, + 121 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 455, + 121 + ], + "type": "text", + "content": "Single Domain Generalization for Crowd Counting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 144, + 429, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 144, + 429, + 185 + ], + "spans": [ + { + "bbox": [ + 164, + 144, + 429, + 185 + ], + "type": "text", + "content": "Zhuoxuan Peng, S.-H. Gary Chan \nThe Hong Kong University of Science and Technology \n{zpengac, gchan}@cse.ust.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "content": "Due to its promising results, density map regression has been widely employed for image-based crowd counting. The approach, however, often suffers from severe performance degradation when tested on data from unseen scenarios, the so-called \"domain shift\" problem. To address the problem, we investigate in this work single domain generalization (SDG) for crowd counting. The existing SDG approaches are mainly for image classification and segmentation, and can hardly be extended to our case due to its regression nature and label ambiguity (i.e., ambiguous pixel-level ground truths). We propose MPCount, a novel effective SDG approach even for narrow source distribution. MPCount stores diverse density values for density map regression and reconstructs domain-invariant features by means of only one memory bank, a content error mask and attention consistency loss. By partitioning the image into grids, it employs patch-wise classification as an auxiliary task to mitigate label ambiguity. Through extensive experiments on different datasets, MPCount is shown to significantly improve counting accuracy compared to the state of the art under diverse scenarios unobserved in the training data characterized by narrow source distribution. Code is available at https://github.com/Shimmer93/MPCount." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": "Image-based crowd counting is to estimate the number of people (or objects) in an image. The recent mainstream counting approach is density map regression, where a density map with continuous pixel value is predicted from the input image, and the total crowd count is obtained by summing all these pixel values. During training, the ground-truth density map is generated by smoothing point annotations given by the locations of human heads. Density map regression has been demonstrated to achieve encouraging results in crowd counting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Existing training models predominantly assume that the testing data share the same source distribution, i.e., fully supervised learning. Such assumption, however, often breaks" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 212, + 388, + 256 + ], + "blocks": [ + { + "bbox": [ + 310, + 212, + 388, + 256 + ], + "lines": [ + { + "bbox": [ + 310, + 212, + 388, + 256 + ], + "spans": [ + { + "bbox": [ + 310, + 212, + 388, + 256 + ], + "type": "image", + "image_path": "29a0596334f283546466ccc996ea121ca039dfdefdffa00c40826bc036a444b6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 258, + 417, + 266 + ], + "lines": [ + { + "bbox": [ + 357, + 258, + 417, + 266 + ], + "spans": [ + { + "bbox": [ + 357, + 258, + 417, + 266 + ], + "type": "text", + "content": "(a) Source domain" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 388, + 213, + 463, + 256 + ], + "blocks": [ + { + "bbox": [ + 388, + 213, + 463, + 256 + ], + "lines": [ + { + "bbox": [ + 388, + 213, + 463, + 256 + ], + "spans": [ + { + "bbox": [ + 388, + 213, + 463, + 256 + ], + "type": "image", + "image_path": "e1c2b5b41a74ccea9a425f38d9de2b82ffb2bea1af72a4e8560dc10fa6e0207a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 257, + 447, + 266 + ], + "lines": [ + { + "bbox": [ + 406, + 257, + 447, + 266 + ], + "spans": [ + { + "bbox": [ + 406, + 257, + 447, + 266 + ], + "type": "text", + "content": "gain" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 465, + 213, + 542, + 256 + ], + "blocks": [ + { + "bbox": [ + 465, + 213, + 542, + 256 + ], + "lines": [ + { + "bbox": [ + 465, + 213, + 542, + 256 + ], + "spans": [ + { + "bbox": [ + 465, + 213, + 542, + 256 + ], + "type": "image", + "image_path": "ba96caa0eec5e34538c24c910bff308c00da846361b10ea634a546f5f975ea88.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 258, + 530, + 267 + ], + "lines": [ + { + "bbox": [ + 476, + 258, + 530, + 267 + ], + "spans": [ + { + "bbox": [ + 476, + 258, + 530, + 267 + ], + "type": "text", + "content": "(b) Target image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 269, + 386, + 312 + ], + "blocks": [ + { + "bbox": [ + 310, + 269, + 386, + 312 + ], + "lines": [ + { + "bbox": [ + 310, + 269, + 386, + 312 + ], + "spans": [ + { + "bbox": [ + 310, + 269, + 386, + 312 + ], + "type": "image", + "image_path": "640bf12c0e782edc061bb92537af332efa631d01afdf37f90fde7c891cd21219.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 314, + 375, + 323 + ], + "lines": [ + { + "bbox": [ + 321, + 314, + 375, + 323 + ], + "spans": [ + { + "bbox": [ + 321, + 314, + 375, + 323 + ], + "type": "text", + "content": "(c) Ground-truth" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 388, + 269, + 464, + 312 + ], + "blocks": [ + { + "bbox": [ + 388, + 269, + 464, + 312 + ], + "lines": [ + { + "bbox": [ + 388, + 269, + 464, + 312 + ], + "spans": [ + { + "bbox": [ + 388, + 269, + 464, + 312 + ], + "type": "image", + "image_path": "cb179390b3f8de5891ec7a74420c5232187dbbe1a59a06b74312c31857fd38c4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 314, + 446, + 323 + ], + "lines": [ + { + "bbox": [ + 406, + 314, + 446, + 323 + ], + "spans": [ + { + "bbox": [ + 406, + 314, + 446, + 323 + ], + "type": "text", + "content": "(d) DCCUS" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 465, + 269, + 542, + 312 + ], + "blocks": [ + { + "bbox": [ + 465, + 269, + 542, + 312 + ], + "lines": [ + { + "bbox": [ + 465, + 269, + 542, + 312 + ], + "spans": [ + { + "bbox": [ + 465, + 269, + 542, + 312 + ], + "type": "image", + "image_path": "e8f1c619400494e6ac391964186cf6a6f5fe7102e55efd18563be023d1d99d02.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 314, + 537, + 323 + ], + "lines": [ + { + "bbox": [ + 471, + 314, + 537, + 323 + ], + "spans": [ + { + "bbox": [ + 471, + 314, + 537, + 323 + ], + "type": "text", + "content": "(e) MPCount (Ours)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "type": "text", + "content": "Figure 1. Single domain generalization for crowd counting. (a) Sample images from a single source domain " + }, + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "type": "text", + "content": " with narrow distribution. (b) A challenging target image from an unseen scenario. (c) The ground-truth density map. (d) The predicted density map from the previous work DCCUS [6] trained on " + }, + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "type": "text", + "content": ". (e) The predicted density map from our MPCount trained on " + }, + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 333, + 545, + 422 + ], + "type": "text", + "content": ". MPCount achieves much lower counting error than DCCUS and makes better predictions in the cropped area against the domain shift." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 506 + ], + "type": "text", + "content": "in practice due to changes in camera position, weather condition, etc. Such deviation from the training data (i.e. source domain) to unseen testing data (i.e. target domain) is known as domain shift, which severely degrades the performance of deep learning models in real-world deployment." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "type": "text", + "content": "To mitigate the domain shift problem, domain adaptation (DA) approaches have been proposed to transfer knowledge from the source domain to the target one, typically by finetuning pre-trained models with target domain data [2, 10, 39, 44]. Despite promising results, the fine-tuning process of DA is usually laborious. Furthermore, target domain data may not be easily available in reality." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 592, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 700 + ], + "type": "text", + "content": "In contrast to DA, domain generalization (DG) aims to generalize models to any unobserved target domain without the need for target data. As data and model fine-tuning are not necessary at the target domain, DG is more convenient to be deployed in unseen or unpredictable domain. We consider in this work the challenging case of single domain generalization (SDG) for image-based crowd counting, where only one source domain, likely of narrow distribution (Fig. 1a), is used for training." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "A common general SDG technique is to construct" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28025" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "domain-invariant features from different augmented versions of an image. Such approach has been shown promising for various tasks [5, 13, 34]. Independent of data augmentation, many other SDG approaches have been proposed for classification and segmentation to further boost performance. Unfortunately, most of them cannot be extended to crowd counting, a regression task without category information. Moreover, while labels in classification and segmentation tasks are precise and domain-invariant, the labels for density-based crowd counting are usually ambiguous. Such \"label ambiguity\" stems from the fact that the density values are generated from point annotations, and hence human heads and backgrounds may be assigned to similar value under different scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 361 + ], + "type": "text", + "content": "The recent work of [6] achieves DG for crowd counting based on dividing the source into multiple sub-domains. While encouraging, it requires a rather broad source domain distribution. We propose, for the first time, an SDG approach for crowd counting called MPCount without sub-domain division, thereof effective and extensible to narrow source distribution. We highlight its performance as compared with a state-of-the-art scheme in Fig. 1. MPCount achieves its superior performance with the following two novel components:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 364, + 289, + 675 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 364, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 289, + 555 + ], + "type": "text", + "content": "- An attention memory bank (AMB) to tackle density regression: MPCount employs only one attention memory bank (AMB) that takes a pair of features encoded from an crowd image and its data-augmented version as input. To cover continuous density values for regression with a finite memory size, AMB reconstructs each feature vector as an attention over memory vectors. These memory vectors learn domain-invariant representations from the pair of features of different styles but similar content (i.e. crowd densities). We present the content error mask (CEM) to eliminate domain-related content from input features, as characterized by a large discrepancy between instance-normalized feature element pairs. Furthermore, the newly employed attention consistency loss (ACL) enforces the similarity of attention scores produced from input features, ensuring the consistency of memory vectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 555, + 289, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 289, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 289, + 675 + ], + "type": "text", + "content": "- Patch-wise classification " + }, + { + "bbox": [ + 47, + 555, + 289, + 675 + ], + "type": "inline_equation", + "content": "(PC)" + }, + { + "bbox": [ + 47, + 555, + 289, + 675 + ], + "type": "text", + "content": " to tackle label ambiguity: To address the label ambiguity, MPCount employs a novel auxiliary task called patch-wise classification (PC). In this task, each crowd image is evenly divided into fixed-size square patches, say " + }, + { + "bbox": [ + 47, + 555, + 289, + 675 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 47, + 555, + 289, + 675 + ], + "type": "text", + "content": ", classified into two classes, i.e., containing human heads or not. The density values in the areas classified as without head are filtered out during density regression, so that the pixel-level ambiguity is overcome by coarser but more accurate patchwise binary labels." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "To validate MPCount, we conduct extensive experiments on various crowd counting datasets including ShanghaiTech A (SHA) & B (SHB) and JHU-Crowd++. We introduce" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "a new challenging setting of narrow source distribution, where only images of the same category, such as \"snow\" (SN) and \"fog/haze\" (FH) in JHU-Crowd++, are used in training. We demonstrate that MPCount achieves excellent performance not only on traditional inter-dataset benchmarks, but also under our newly introduced narrow source setting. Let " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "S \\rightarrow T" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " denote the case where the source domain is " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " and the target domain is " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": ". As compared to the state of the art, MPCount is shown to reduce significantly counting error by " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " on SN " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " FH, " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " on FH " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " SN, " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "18.2\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " on SHB " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " SHA and " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " on SHA " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " SHB." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 215, + 398, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 215, + 398, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 215, + 398, + 228 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 235, + 489, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 489, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 489, + 248 + ], + "type": "text", + "content": "2.1. Fully Supervised Crowd Counting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 254, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 422 + ], + "type": "text", + "content": "The mainstream approach to crowd counting is via density map regression, where each pixel of an image is assigned a count value, and the summation of these values yields an estimate of the total number of people. Most methods focus on improving counting accuracy under the fully supervised setting by employing novel network designs [1, 16, 17, 31, 45] or loss functions [20, 26, 36, 37]. While encouraging results have been shown, their performance usually drop significantly when evaluated on out-of-distribution data, due to their limited generalization ability. Besides, several works [4, 22, 38] have proposed auxiliary tasks in crowd counting for certain purposes. However, these tasks still require pixel-level predictions and contribute little to tackling label ambiguity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 429, + 545, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 545, + 442 + ], + "type": "text", + "content": "2.2. Domain Adaptation (DA) for Crowd Counting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 448, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 556 + ], + "type": "text", + "content": "Domain adaptation (DA) has been widely studied in crowd counting to address domain shift by adapting source domain information to a particular target domain. While a few methods study supervised DA where labeled data from target domain are available [40, 44], most approaches tackle unsupervised DA, utilizing only unlabeled target domain data [2, 7, 18, 39, 41, 47]. Despite the notable advancements, DA methods normally require data from the target domain, which may not be easily available in practice." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 563, + 503, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 503, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 503, + 576 + ], + "type": "text", + "content": "2.3. Single Domain Generalization (SDG)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Domain Generalization (DG) approaches are designed to train a network with generalization ability solely utilizing source domain data, and single domain generalization (SDG) is the special case when only one source domain is available. There are various techniques for SDG, including 1) adversarial data generation [15, 25, 46, 50], 2) feature normalization/whitening transformation [5, 12, 23, 24, 43] and 3) domain-general network design, such as a convolution layer [35], vision transformer [32] and memory bank [3]. Among various SDG methods, a common general technique is to simulate domain shift with data augmenta" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28026" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "type": "text", + "content": "tion and extract domain-invariant information from different augmented versions of features [5, 13, 43]. While SDG has attracted widespread attention in classification and segmentation, it is still a nascent topic in crowd counting due to its regression nature and label ambiguity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 335 + ], + "type": "text", + "content": "DCCUS [6] studies SDG for crowd counting. In this method, the source domain is dynamically clustered into sub-domains simulated as meta-training and -test sets in a meta-learning strategy. Two types of memory modules and several losses are proposed to distinguish and record domain-invariant and -related information. DCCUS designs the memory mechanism specifically for regression, but the problem of ambiguous labels is not considered. Moreover, when the source domain follows a narrow distribution, the sub-domain division process may experience reduced effectiveness and affect the generalization ability of the counting model. MPCount, without sub-domain division, employs novel patch-wise classification to address the label ambiguity problem. Additionally, a single memory bank for regression is introduced without the need for sub-domain partitioning, enabling MPCount to generalize well even in source domains of narrow distributions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 346, + 113, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 113, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 113, + 357 + ], + "type": "text", + "content": "3. MPCount" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 286, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 485 + ], + "type": "text", + "content": "This section provides a detailed description of our MPCount scheme. We begin by reviewing the foundational concepts of crowd counting in Sec. 3.1. Next, in Sec. 3.2, we present a comprehensive overview of the entire scheme. We then elaborate the attention memory bank (Sec. 3.3) designed for density regression, followed by the content error mask (Sec. 3.4) and attention consistency loss (Sec. 3.5). After that, we discuss the patch-wise classification (Sec. 3.6) proposed to address the challenge of ambiguous labels. Finally, we detail the overall training loss in Sec. 3.7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 492, + 214, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 214, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 214, + 504 + ], + "type": "text", + "content": "3.1. Crowd Counting Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": "First, we review the preliminary knowledge of deep learning-based crowd counting via density map regression. Our objective is to train a neural network " + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": " that takes in an input image " + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": " and outputs a density map " + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}} = \\mathcal{N}(\\mathcal{I};\\theta)" + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": ". The estimated count " + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": " is the sum of all pixel-wise density values in " + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 511, + 286, + 582 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 600, + 286, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 600, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 600, + 286, + 634 + ], + "type": "interline_equation", + "content": "\\hat {c} = \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\hat {\\mathcal {D}} _ {i j}. \\tag {1}", + "image_path": "d67ed5f506affb989e08d4726ad60f89534af426c694fffee00449990d4b97a4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": "Generation of ground-truth density maps usually follows the method in [28]. Given the human head point annotations " + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": " where each " + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "inline_equation", + "content": "h\\in \\mathcal{H}" + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": " is the coordinate of a head location, the ground-truth density map " + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": " is calculated as the sum of the 2D Gaussian filter applied to each head coordinate " + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 365, + 86, + 545, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 86, + 545, + 112 + ], + "spans": [ + { + "bbox": [ + 365, + 86, + 545, + 112 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = \\sum_ {h \\in \\mathcal {H}} \\delta (x - h) \\times G _ {\\sigma} (x), \\tag {2}", + "image_path": "95f9731074021c1c1489dc0979de58eeaa21bd1471e7ca0123d1a72fa3e131e0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "type": "text", + "content": " is the discrete delta function, and " + }, + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "type": "inline_equation", + "content": "G_{\\sigma}(\\cdot)" + }, + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "type": "text", + "content": " is the Gaussian kernel with a fixed variance " + }, + { + "bbox": [ + 305, + 116, + 545, + 139 + ], + "type": "inline_equation", + "content": "\\sigma" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 140, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 163 + ], + "type": "text", + "content": "A common objective function to supervise the network " + }, + { + "bbox": [ + 305, + 140, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 305, + 140, + 545, + 163 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 179, + 545, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 179, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 388, + 179, + 545, + 194 + ], + "type": "interline_equation", + "content": "\\mathbb {L} (\\theta) = | | \\mathcal {D}, \\hat {\\mathcal {D}} | | _ {2} ^ {2}. \\tag {3}", + "image_path": "3ae6ca554c236cdf5c2eaa0f1f68b8f938cac75c6b359b283746de3e170e534c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 198, + 411, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 198, + 411, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 411, + 209 + ], + "type": "text", + "content": "3.2. Scheme Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": "The overall structure of our method is illustrated in Fig. 2. During training, photometric transformations are applied to the original image " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " to produce an augmented version " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{aug}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": ". The feature extractor encodes them into features " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{aug}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": ", and a content error mask (CEM) " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " is applied to them. The masked features are reconstructed into " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}^{aug}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " via an attention memory bank (AMB) and passed to the density head for crowd density prediction. Meanwhile, the highest-level features encoded by the feature extractor, " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{aug}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": ", are fed into the patch-wise classification (PC) head, and the predicted PC maps (PCMs) " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}^{aug}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " are binarized and resized to serve as a mask to filter out areas without human heads in the estimated density maps " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}^{aug}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": ". The final predictions " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'^{aug}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " are two density maps with certain areas masked out by the PCMs. The entire training process is supervised under a combination of density losses, PC losses and the attention consistency loss (ACL). The inference process solely employs " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": " as input, and outputs a single density map " + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'^{ori}" + }, + { + "bbox": [ + 305, + 216, + 545, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 449, + 479, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 479, + 461 + ], + "type": "text", + "content": "3.3. Attention Memory Bank (AMB)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 467, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 545, + 586 + ], + "type": "text", + "content": "The attention memory bank (AMB) is designed to automatically learn domain-invariant features for density regression. [14] proposes a memory bank where memory vectors are updated individually, each corresponding to a certain category. However, this design cannot be applied to regression tasks, since a finite number of memory vectors cannot cover continuous density values. Inspired by [6], we reconstruct each feature vector as an attention over memory vectors, so that any density value may be represented as a linear combination of representative values in the memory bank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": "The AMB " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " memory vectors of dimension " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": ". Given a flattened input feature map " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": ", we first compute the attention scores " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " between the feature " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " (query) and the memory " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " (key). Then the reconstructed feature map " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " is calculated as a linear combination of " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " (value) with " + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 305, + 586, + 545, + 670 + ], + "type": "text", + "content": " as weights. The reconstruction process can be summarized as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 684, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 684, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 360, + 684, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (\\mathcal {F}, \\mathcal {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {F V} ^ {T}}{\\sqrt {C}}\\right), \\tag {4}", + "image_path": "453ffd6fa50a23b97a8a8d223d6fb0e83fcbab8393419ca5df0f1c0149485b17.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28027" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 547, + 266 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 547, + 266 + ], + "type": "image", + "image_path": "18074bfdcd490ed255445b4ed43ad1a70629440fb2752c5b0bd7c4fb864b3ea3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 546, + 297 + ], + "type": "text", + "content": "Figure 2. The overall training pipeline of our proposed MPCount. All identical modules in this diagram share the same weights. In the encoder-decoder structure, a higher level indicates a deeper feature." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 316, + 65, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 65, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 65, + 327 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 337, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 337, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 145, + 337, + 287, + 352 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {F}} = \\mathcal {A} \\mathcal {V}. \\tag {5}", + "image_path": "0e33aa618318b4fbe4d45eb051cbbb2cb8d92ae31f0f3b88c598b9205929e66e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 458 + ], + "type": "text", + "content": "While [6] relies on multiple memory banks corresponding to different sub-domains to distinguish domain-invariant and -related information, we only employ a single AMB and propose the novel CEM (Sec. 3.4) and ACL (Sec. 3.5) to ensure that it stores domain-invariant representations. To this end, our feature reconstruction mechanism can work without sub-domain division, thereof is effective even with a narrow source distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 467, + 202, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 202, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 202, + 479 + ], + "type": "text", + "content": "3.4. Content Error Mask (CEM)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": "CEM is proposed to guarantee the resemblance of content contained in the pair of input features by excluding possible domain-related content information. A common technique to disentangle style and content in a feature is instance normalization (IN) [33], where the feature statistics (channel-wise mean and variance) are considered to contain style information, while the instance normalized features retain content information. Based on this idea, we assume that the discrepancy of instance normalized features reflect effects of domain shift on feature content and may mislead AMB to learn domain-related information. Therefore, we filter out elements in the input features with difference between their instance normalized versions above a certain threshold, resulting in a pair of features with similar content information for domain-invariant feature reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "text", + "content": "More specifically, given the features extracted from the original and augmented images, " + }, + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{ori}" + }, + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{aug} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "text", + "content": ", we define the content error mask " + }, + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 665, + 288, + 712 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 332, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 332, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {i j k} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | \\mathrm {I N} (\\mathcal {F} ^ {\\text {o r i}}) _ {i j k} - \\mathrm {I N} (\\mathcal {F} ^ {\\text {a u g}}) _ {i j k} | \\leq \\alpha \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}", + "image_path": "7d45d87df960697193f82452c22d855656a83e139d1d9f822b725c63dd8fa75a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "text", + "content": " is the threshold indicating whether the error value reflects possible inconsistency of content information caused by domain shift." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "text", + "content": " to be one of " + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\text{ori}}" + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\text{aug}}" + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "text", + "content": ", we filter out the domain-related content information by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 448, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 448, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 367, + 448, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {\\prime} = \\operatorname {D r o p o u t 2 D} (\\mathcal {F} \\odot \\mathcal {M}). \\tag {7}", + "image_path": "334ba80d450f51b22f33a58086fd20864716638b59cbc487fca17d4d51ed8693.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "content": "The random 2D dropout is applied to " + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'" + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "content": " to prevent the memory from relying solely on certain channels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 497, + 489, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 489, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 489, + 510 + ], + "type": "text", + "content": "3.5. Attention Consistency Loss (ACL)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "text", + "content": "The style information of " + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{ori}" + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{aug}" + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "text", + "content": " preserved in Sec. 3.4 may produce a large discrepancy on their attention distributions over the memory vectors in AMB even if they contain identical content information. To ensure that each memory vector consistently store specific domain-invariant representations, we propose the attention consistency loss (ACL) to enforce the similarity of attention scores produced from " + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{ori}" + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{aug}" + }, + { + "bbox": [ + 304, + 515, + 545, + 611 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "content": "We consider " + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "content": " in Eq. 4 as a distribution, and the ACL " + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{con}" + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "content": " is calculated as the distance between the distribution " + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{ori}" + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "content": " produced by " + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{ori}" + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{aug}" + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "content": " produced by " + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{aug}" + }, + { + "bbox": [ + 304, + 612, + 545, + 672 + ], + "type": "text", + "content": ". Here the simple Euclidean distance is selected as the distance measure due to its computational stability during training:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 689, + 545, + 704 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 689, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 375, + 689, + 545, + 704 + ], + "type": "interline_equation", + "content": "\\mathbb {L} _ {\\text {c o n}} = \\left\\| \\mathcal {A} ^ {\\text {a u g}}, \\mathcal {A} ^ {\\text {o r i}} \\right\\| _ {2} ^ {2}. \\tag {8}", + "image_path": "526b4cf55227c132395518baf9be0b4a16417e2bebb48f843968dfe5f9d24191.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28028" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 70, + 166, + 125 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 166, + 125 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 166, + 125 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 166, + 125 + ], + "type": "image", + "image_path": "6d595b4e7315ec7bb19574ecafa9840a473d3ec2e87411a3217473df3fec3771.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 127, + 158, + 137 + ], + "lines": [ + { + "bbox": [ + 104, + 127, + 158, + 137 + ], + "spans": [ + { + "bbox": [ + 104, + 127, + 158, + 137 + ], + "type": "text", + "content": "(a) Density Map" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 241, + 125 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 241, + 125 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 241, + 125 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 241, + 125 + ], + "type": "image", + "image_path": "cc336287be96e075878223a874b3f55febe4f4805756b5982f9e8f8f33a487d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 127, + 219, + 136 + ], + "lines": [ + { + "bbox": [ + 189, + 127, + 219, + 136 + ], + "spans": [ + { + "bbox": [ + 189, + 127, + 219, + 136 + ], + "type": "text", + "content": "(b) PCM" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 146, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 212 + ], + "type": "text", + "content": "Figure 3. Illustration of label ambiguity and how PCM tackles it. Point A is a typical example which belongs to a human head but is assigned the density value 0 as the background in the density map due to varying head sizes. In the PCM, the patch covering A is correctly classified as containing heads, thanks to the patch-level accuracy of the labels." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 230, + 209, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 209, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 209, + 243 + ], + "type": "text", + "content": "3.6. Patch-wise Classification (PC)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 248, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 332 + ], + "type": "text", + "content": "As discussed in Sec. 3.1, ground-truth density maps are usually calculated only based on point annotations. In different scenarios, pixels of a human head and the background might be assigned the same density value due to various head sizes (Fig. 3a), violating the common assumption that labels are accurate and domain-invariant, thereof resulting in the ambiguous labels in crowd counting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 332, + 287, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 287, + 439 + ], + "type": "text", + "content": "To tackle the label ambiguity, we propose a novel auxiliary task named patch-wise classification (PC) and its corresponding supervision signal, patch-wise classification map (PCM), as illustrated in Fig. 3b. A PCM divides an image evenly into " + }, + { + "bbox": [ + 46, + 332, + 287, + 439 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 46, + 332, + 287, + 439 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 332, + 287, + 439 + ], + "type": "inline_equation", + "content": "P = 16" + }, + { + "bbox": [ + 46, + 332, + 287, + 439 + ], + "type": "text", + "content": " empirically) patches and classifies each patch into two categories: containing human heads or not. Such patch-level predictions mitigate the uncertainty in pixel-level density maps by providing coarser but more accurate patch-level information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 440, + 287, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 463 + ], + "type": "text", + "content": "In practice, a ground truth " + }, + { + "bbox": [ + 46, + 440, + 287, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{PCM}\\mathcal{C}" + }, + { + "bbox": [ + 46, + 440, + 287, + 463 + ], + "type": "text", + "content": " with patch size " + }, + { + "bbox": [ + 46, + 440, + 287, + 463 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 440, + 287, + 463 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 467, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 467, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 94, + 467, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {i j} = 1 - \\delta \\left(\\sum_ {k = i P} ^ {(i + 1) P} \\sum_ {l = j P} ^ {(j + 1) P} \\mathcal {D} _ {k l}\\right), \\tag {9}", + "image_path": "514aa6faed2e16227045f6c123dd198f8dc5eeda9e33a9770b826ee24cc45eda.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "content": " is the ground truth density map and " + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "content": " is the discrete delta function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 534, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 559 + ], + "type": "text", + "content": "Our model predicts a PCM " + }, + { + "bbox": [ + 46, + 534, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 46, + 534, + 287, + 559 + ], + "type": "text", + "content": " supervised by the binary cross entropy (BCE) loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 563, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 563, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 127, + 563, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\mathbb {L} _ {c l s} = \\operatorname {B C E} (\\mathcal {C}, \\hat {\\mathcal {C}}). \\tag {10}", + "image_path": "33b121fbbca5f4015e6276638992f68b815ce8ba302fea6b921bd42324753d5e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "content": "Subsequently, " + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}" + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "content": " is binarized and resized to obtain " + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{C}'" + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "content": " which matches the dimensions of the predicted density map " + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{C}'" + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "content": " is then used to mask the regions classified without crowds in " + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "content": ", resulting in the final density map " + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'" + }, + { + "bbox": [ + 47, + 585, + 287, + 634 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 639, + 287, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 639, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 137, + 639, + 287, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {D} ^ {\\prime} = \\hat {\\mathcal {D}} \\odot \\mathcal {C} ^ {\\prime}. \\tag {11}", + "image_path": "a029c7d0624e170abb509dbec04a7e2ae1ba974c3eaef7c18e7be181af23c4c6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "type": "text", + "content": "Finally, the density regression head is supervised with the ground truth density map " + }, + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 658, + 287, + 695 + ], + "type": "text", + "content": ", and the common Euclidean distance in Eq. (3) is used as the objective function:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 127, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbb {L} _ {d e n} = \\left\\| \\mathcal {D}, \\mathcal {D} ^ {\\prime} \\right\\| _ {2} ^ {2}. \\tag {12}", + "image_path": "32e2ab3dfa23c64bf55205e75076835303ab385b73f01afa7fa5180c8d032297.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 72, + 402, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 402, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 402, + 85 + ], + "type": "text", + "content": "3.7. Model Training" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "text", + "content": "The overall training loss is a combination of density losses " + }, + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{den}^{ori},\\mathbb{L}_{den}^{aug}" + }, + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "text", + "content": ", PC losses " + }, + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{cls}^{ori},\\mathbb{L}_{cls}^{aug}" + }, + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "text", + "content": " and the attention consistency loss " + }, + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{con}" + }, + { + "bbox": [ + 306, + 90, + 545, + 126 + ], + "type": "text", + "content": ", which can be written as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 133, + 545, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 133, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 315, + 133, + 545, + 158 + ], + "type": "interline_equation", + "content": "\\mathbb {L} = \\mathbb {L} _ {d e n} ^ {o r i} + \\mathbb {L} _ {d e n} ^ {a u g} + \\lambda_ {c l s} \\left(\\mathbb {L} _ {c l s} ^ {o r i} + \\mathbb {L} _ {c l s} ^ {a u g}\\right) + \\lambda_ {c o n} \\mathbb {L} _ {c o n}, \\tag {13}", + "image_path": "3d4044ebd6836e69130632cbeb04d663e2f8bc42429340dca71827a58d847d57.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 165, + 545, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 165, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 545, + 188 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 165, + 545, + 188 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 305, + 165, + 545, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 165, + 545, + 188 + ], + "type": "inline_equation", + "content": "\\lambda_{con}" + }, + { + "bbox": [ + 305, + 165, + 545, + 188 + ], + "type": "text", + "content": " are weighting parameters to balance different loss terms." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 198, + 491, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 491, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 491, + 213 + ], + "type": "text", + "content": "4. Illustrative Experimental Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 219, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 291 + ], + "type": "text", + "content": "In this section, we first introduce the datasets used in experiments in Sec. 4.1. Next, the implementation details are specified in Sec. 4.2, then evaluation metrics are described in Sec. 4.3. We present the results compared with state of the art in Sec. 4.4. Finally, we conduct ablation studies and other additional analysis in Sec. 4.5." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 297, + 368, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 368, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 368, + 308 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 316, + 545, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 351 + ], + "type": "text", + "content": "We evaluate our method on four mainstream crowd counting datasets: ShanghaiTech Part A & B, UCF-QNRF and JHU-Crowd++." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 353, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 306, + 353, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 353, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 545, + 447 + ], + "type": "text", + "content": "- ShanghaiTech [45] is composed of two parts, SHA (A) and SHB (B). SHA contains 300 training images and 182 testing images, while SHB contains 400 training images and 316 testing images. Images in SHA are collected from the Internet and feature highly crowded scenes. In contrast, SHB is captured from several streets in Shanghai, and its images exhibit crowd densities generally lower than those in SHA." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 449, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 545, + 496 + ], + "type": "text", + "content": "- UCF-QNRF (Q) [11] consists of 1201 training images and 334 testing images. It is a challenging dataset with a large range of crowd densities, scenes, viewpoints and lighting conditions." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 497, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 545, + 688 + ], + "type": "text", + "content": "- JHU-Crowd++ [29] is a large-scale dataset containing 4372 images, with 2722, 500 and 1600 images in the training, validation and testing set, respectively. In addition, image-level labels are also provided in this dataset, including 16 types of scenes and 4 types of weather conditions. For scene annotations, there are 879 images labeled \"stadium\" (SD) and 573 labeled \"street\" (ST), which are utilized as two source domains. For weather annotations, we use 201 images labeled \"snow\" (SN) and 168 images labeled \"fog/haze\" (FH) as source domains. In each domain, " + }, + { + "bbox": [ + 306, + 497, + 545, + 688 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 306, + 497, + 545, + 688 + ], + "type": "text", + "content": " of the images are selected as the training set while the rest " + }, + { + "bbox": [ + 306, + 497, + 545, + 688 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 306, + 497, + 545, + 688 + ], + "type": "text", + "content": " are for testing. Data with the same label belong to a narrower distribution than mainstream datasets, thus are more challenging for SDG. Illustrative samples in these scene-specific and weather-specific domains are available in Fig. 4." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "S \\to \\mathcal{T}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " denote the case where " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the source domain and " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the target domain. The datasets are utilized" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28029" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 108, + 169 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 108, + 169 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 108, + 169 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 108, + 169 + ], + "type": "image", + "image_path": "6ea710ba9842e1382414ebbe96cfe63aa38f448d0c1f3015e982e1a7aef8150b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 171, + 99, + 180 + ], + "lines": [ + { + "bbox": [ + 59, + 171, + 99, + 180 + ], + "spans": [ + { + "bbox": [ + 59, + 171, + 99, + 180 + ], + "type": "text", + "content": "(a) Stadium" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 70, + 167, + 169 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 167, + 169 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 167, + 169 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 167, + 169 + ], + "type": "image", + "image_path": "02541f207d512fd99e99c6ccf724a287a7b6d932b25ac5d02e65341963169652.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 171, + 154, + 180 + ], + "lines": [ + { + "bbox": [ + 122, + 171, + 154, + 180 + ], + "spans": [ + { + "bbox": [ + 122, + 171, + 154, + 180 + ], + "type": "text", + "content": "(b) Street" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 227, + 170 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 227, + 170 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 227, + 170 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 227, + 170 + ], + "type": "image", + "image_path": "64084dfcd33edd730f2bf68d2bfcb768ba3b9460ca2a401766734fdf0f020e46.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 171, + 212, + 180 + ], + "lines": [ + { + "bbox": [ + 182, + 171, + 212, + 180 + ], + "spans": [ + { + "bbox": [ + 182, + 171, + 212, + 180 + ], + "type": "text", + "content": "(c) Snow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 190, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 212 + ], + "type": "text", + "content": "Figure 4. Sample images with different instance-level labels in JHU-Crowd++." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 286, + 170 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 286, + 170 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 286, + 170 + ], + "type": "image", + "image_path": "88f90215da608759fe740d88269e5fc4c60414cffb34bad4c1f255229c093065.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 171, + 278, + 180 + ], + "lines": [ + { + "bbox": [ + 234, + 171, + 278, + 180 + ], + "spans": [ + { + "bbox": [ + 234, + 171, + 278, + 180 + ], + "type": "text", + "content": "(d) Fog/Haze" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "text", + "content": "in experiments under two distinct types of settings: 1) An entire dataset is regarded as a single domain: " + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\rightarrow \\mathrm{B} / \\mathrm{Q}" + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{B} \\rightarrow \\mathrm{A} / \\mathrm{Q}" + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{Q} \\rightarrow \\mathrm{A} / \\mathrm{B}" + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "text", + "content": "; 2) A dataset is partitioned into subsets according to image-level labels, and one subset constitutes a single domain: " + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{SD} \\leftrightarrow \\mathrm{SR}" + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{SN} \\leftrightarrow \\mathrm{FH}" + }, + { + "bbox": [ + 47, + 232, + 287, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 298, + 144, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 144, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 144, + 310 + ], + "type": "text", + "content": "4.2. Implementation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": "We adopt VGG16-BN [27] as the feature extractor in our model. The density regression head is a single convolution layer with " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " filters, while the classification head consists of a " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " convolution layers. Batch normalization is employed after all convolution layers except the last ones. For data augmentation, we apply three types of photometric transformations, color jittering, Gaussian blurring and sharpening. We also randomly crop image patches with a size of " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " and adopt random horizontal flipping to both the original and augmented images simultaneously. We select AdamW [19] as the optimizer and OneCycleLR [30] as the learning rate scheduler with maximum learning rate set to 1e-3 and a maximum epoch of 300. The memory size " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " is set to 1024 and the dimension " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " is 256. We use a content error threshold " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " of 0.5, and the loss weights " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\lambda_{con}" + }, + { + "bbox": [ + 46, + 316, + 287, + 508 + ], + "type": "text", + "content": " are both set to 10." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 514, + 160, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 160, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 160, + 525 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "type": "text", + "content": "We evaluate our method with mean absolute error (MAE) and mean squared error (MSE), defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 572, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 572, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 572, + 287, + 622 + ], + "type": "interline_equation", + "content": "M A E = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| c _ {i} - \\hat {c} _ {i} \\right|, M S E = \\sqrt {\\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(c _ {i} - \\hat {c} _ {i}\\right) ^ {2}}, \\tag {14}", + "image_path": "813469ab89f2579b35bd384f3283d5c99bf5763695d511aacfb890da04640869.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "text", + "content": " is the number of testing images, " + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "text", + "content": " is the ground truth count of the " + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "text", + "content": "-th image and " + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 47, + 628, + 287, + 665 + ], + "type": "text", + "content": " is the predicted count. Lower values of both metrics indicate better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 671, + 226, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 226, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 226, + 683 + ], + "type": "text", + "content": "4.4. Comparison with State of the Art" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "In this subsection, we compare our MPCount with state-of-the-art methods on different benchmarks, as illustrated" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 366, + 186 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 366, + 186 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 366, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 366, + 186 + ], + "type": "image", + "image_path": "c1c9c23b03cc636a03c755ed44e872d83be0deeff9cd760032c06f20f1286d61.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 189, + 353, + 198 + ], + "lines": [ + { + "bbox": [ + 321, + 189, + 353, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 189, + 353, + 198 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "text", + "content": "Figure 5. Visualization results of DCCUS [6] and MPCount under different DG settings. First row: A " + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "text", + "content": " B; second row: SN " + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "text", + "content": " FH; third row: B " + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 207, + 545, + 239 + ], + "type": "text", + "content": " A." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 368, + 70, + 425, + 186 + ], + "blocks": [ + { + "bbox": [ + 368, + 70, + 425, + 186 + ], + "lines": [ + { + "bbox": [ + 368, + 70, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 368, + 70, + 425, + 186 + ], + "type": "image", + "image_path": "b17371f220a1d4f8b2d07c684197dd6e847e7f6446d45b9fc2c615fc9b1ace6a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 188, + 408, + 197 + ], + "lines": [ + { + "bbox": [ + 385, + 188, + 408, + 197 + ], + "spans": [ + { + "bbox": [ + 385, + 188, + 408, + 197 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 485, + 186 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 485, + 186 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 485, + 186 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 485, + 186 + ], + "type": "image", + "image_path": "b0a754df6ec173baa4b9906bab7d18b6614c99bf9bbfc8b548b6554199f1f9df.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 188, + 474, + 197 + ], + "lines": [ + { + "bbox": [ + 436, + 188, + 474, + 197 + ], + "spans": [ + { + "bbox": [ + 436, + 188, + 474, + 197 + ], + "type": "text", + "content": "(c) DCCUS" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 486, + 70, + 544, + 186 + ], + "blocks": [ + { + "bbox": [ + 486, + 70, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 486, + 70, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 486, + 70, + 544, + 186 + ], + "type": "image", + "image_path": "e05d7b6691c0a3e221adb9d4fb144886576bb6b67ff9683159e44a8e6088e98b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 188, + 537, + 197 + ], + "lines": [ + { + "bbox": [ + 493, + 188, + 537, + 197 + ], + "spans": [ + { + "bbox": [ + 493, + 188, + 537, + 197 + ], + "type": "text", + "content": "(d) MPCount" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "text", + "content": "in Fig. 5. Tab. 1 shows the results on " + }, + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\rightarrow \\mathrm{B} / \\mathrm{Q}" + }, + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{B} \\rightarrow \\mathrm{A} / \\mathrm{Q}" + }, + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{Q} \\rightarrow \\mathrm{A} / \\mathrm{B}" + }, + { + "bbox": [ + 305, + 273, + 545, + 308 + ], + "type": "text", + "content": ". The selected baselines can be divided into three categories:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 314, + 545, + 386 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 306, + 314, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 545, + 338 + ], + "type": "text", + "content": "- Fully supervised crowd counting: BL [20], DM-Count [37], SASNet [31], ChfL [26] and MAN [17]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 338, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 545, + 361 + ], + "type": "text", + "content": "- Domain adaptation for crowd counting: RBT [18], C²MoT [41], FGFD [47], DAOT [48] and FSIM [49]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 362, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 362, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 362, + 545, + 386 + ], + "type": "text", + "content": "- Single domain generalization: IBN [23], SW [24], ISW [5], DG-MAN [21] and DCCUS [6]." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 391, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 463 + ], + "type": "text", + "content": "All results are copied from previous papers except IBN, SW and ISW, which are feature normalization/whitening-based SDG methods originally designed for classification or segmentation and adapted to crowd counting by us. Domain adaptation methods are listed only for reference, as data from target domains are usually visible during adaptation." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 468, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 540 + ], + "type": "text", + "content": "Our MPCount outperforms all the DG methods under most of these settings, including an significant error reduction by " + }, + { + "bbox": [ + 305, + 468, + 545, + 540 + ], + "type": "inline_equation", + "content": "18.2\\%" + }, + { + "bbox": [ + 305, + 468, + 545, + 540 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 468, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{B} \\rightarrow \\mathrm{A}" + }, + { + "bbox": [ + 305, + 468, + 545, + 540 + ], + "type": "text", + "content": ". It is notable that MPCount exhibits outstanding performance in some cases even compared to DA methods, providing compelling evidence for the effectiveness of our design." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": "In Tab. 2, we further conduct experiments on scene-specific domains " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SD} \\leftrightarrow \\mathrm{SR}" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": " and weather-specific domains " + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SN} \\leftrightarrow \\mathrm{FH}" + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": ". BL [20], MAN [17], DAOT [48] and DCCUS [6] are trained using their released code, while IBN [23], SW [24] and ISW [5] are adapted by ourselves. We observe that the normalization/whitening-based methods cannot achieve satisfactory results, possibly because useful information is eliminated from feature statistics. DCCUS also obtains worse results than fully supervised methods, reflecting that the clustering procedure may fail to produce meaningful sub-domains when the source domain distribution is narrow. In contrast, our MPCount still performs the best among all tested methods, demonstrating its superiority in challenging conditions with narrow domain distributions." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28030" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 547, + 292 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 292 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 292 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 292 + ], + "type": "table", + "html": "
Source → TargetA → BA → QB → AB → QQ → AQ → B
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]✘✘15.925.8166.7287.6138.1228.1226.4411.0----
DMCount [37]✘✘23.134.9134.4252.1143.9239.6203.0386.173.4135.114.327.5
SASNet [31]✘✘21.333.2211.2418.6132.4225.6273.5481.373.9116.413.022.1
ChfL [26]✘✘18.729.1122.3218.0121.3200.8197.1357.968.7118.514.027.0
MAN [17]✘✘22.132.8138.8266.3133.6255.6209.4378.867.1122.112.522.2
RBT [18]✓✘13.429.3175.0294.8112.2218.2211.3381.9----
C2MoT [41]✓✘12.421.1125.7218.3120.7192.0198.9368.0----
FGFD [47]✓✘12.723.3124.1242.0123.5210.7209.7384.770.2118.412.520.3
DAOT [48]✓✘10.918.8113.9215.6----67.0128.411.319.6
FSIM [49]✓✘11.119.3105.3191.1120.3202.6194.9324.566.8111.511.019.7
IBN [23]✘✓19.130.8280.2561.0125.9202.3183.5317.4105.9174.616.025.5
SW [24]✘✓20.230.4285.5431.0126.7193.8200.7333.2102.4168.819.032.9
ISW [5]✘✓23.937.7215.6399.7156.2291.5263.1442.283.4136.022.134.7
DG-MAN [21]✘✓17.328.7129.1238.2130.7225.1182.4325.8----
DCCUS [6]✘✓12.624.6119.4216.6121.8203.1179.1316.267.4112.812.120.9
MPCount (Ours)✘✓11.419.7115.7199.899.6182.9165.6290.465.5110.112.324.1
", + "image_path": "7d7755ba285d7ece6d588a693962c40c8fafeba30f5d99f3b4cf41ffe4ee4024.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 96, + 335, + 499, + 477 + ], + "blocks": [ + { + "bbox": [ + 46, + 300, + 546, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 546, + 323 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art methods on SHA (A), SHB (B) and UCF-QNRF (Q). Domain adaptation methods are only for reference since data in target domain might be used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 335, + 499, + 477 + ], + "lines": [ + { + "bbox": [ + 96, + 335, + 499, + 477 + ], + "spans": [ + { + "bbox": [ + 96, + 335, + 499, + 477 + ], + "type": "table", + "html": "
Source → TargetSD → SRSR → SDSN → FHFH → SN
MethodDADGMAEMSEMAEMSEMAEMSEMAEMSE
BL [20]XX42.179.0262.71,063.948.1129.5343.8770.5
MAN [17]XX45.179.0246.1950.838.168.0445.0979.3
DAOT [48]✓X45.388.0278.71,624.342.373.0151.6273.9
IBN [23]X✓92.2178.0318.11,420.4109.7267.7491.81,110.4
SW [24]X✓110.3202.4312.61,072.4131.5306.6381.3825.0
ISW [5]X✓108.1212.4385.91,464.8151.6365.7276.6439.8
DCCUS [6]X✓90.4194.1258.11,005.954.5125.8399.7945.0
MPCount (Ours)X✓37.470.1218.6935.931.355.0216.3421.4
", + "image_path": "4574df905a4773786412ac5fd334ae9d01772e1a64ae00356b8058679d394c2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 486, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 546, + 510 + ], + "type": "text", + "content": "Table 2. Comparison with the state of the art on data in JHU-Crowd++ with labels \"Stadium\"(SD), \"Street\"(SR), \"Snow\"(SN) and \"Fog/Haze\"(FH). Domain adaptation methods are only for reference since data in target domain might be used." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 529, + 147, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 147, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 147, + 541 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 550, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 574 + ], + "type": "text", + "content": "In this section, we conduct ablation studies and other analysis under the setting " + }, + { + "bbox": [ + 46, + 550, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\to \\mathrm{B} / \\mathrm{Q}" + }, + { + "bbox": [ + 46, + 550, + 287, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 575, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 624 + ], + "type": "text", + "content": "Effects of model components. We start from a simple encoder-decoder structured baseline model and add or remove each component individually to validate its effectiveness. The experimental results are displayed in Tab. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 629, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 629, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 689 + ], + "type": "text", + "content": "- Attention Memory Bank: After adding AMB, the performance consistently improves on both target domains, as shown in Tab. 3. This validates that AMB can effectively help the model generalize by reconstructing domain-invariant features for density regression." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "- Content Error Mask & Attention Consistency Loss: Next, we test the effects of CEM and ACL on AMB. The results" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 529, + 544, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 529, + 544, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 529, + 544, + 553 + ], + "type": "text", + "content": "in Tab. 3 show that both components generally improves the performance, especially when PC is present." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 554, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 545, + 613 + ], + "type": "text", + "content": "- Patch-wise Classification: Finally, models with PC outperforms their counterparts without PC, as visible in Tab. 3. This suggests that such an auxiliary task is useful for DG by compensating for ambiguous pixel-level labels with accurate patch-level labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "text", + "content": "Effects of " + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "text", + "content": ". The threshold of CEM " + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "text", + "content": " mentioned in Eq. (6) controls the Portion of Diminished Elements (PDE) in features. According to the results in Tab. 4, " + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "text", + "content": " produces the best performance under all settings, with a PDE of " + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "text", + "content": ". It is also indicated that using a too large value of " + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 617, + 546, + 712 + ], + "type": "text", + "content": " (0.9) may be inadequate to filter out domain-related information, while a too small value (0.1) would let useful information eliminated." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28031" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 289, + 231 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 289, + 231 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 289, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 289, + 231 + ], + "type": "table", + "html": "
ComponentsA → BA → Q
AMBCEMACLPCMAEMSEMAEMSE
21.639.1200.9325.5
✓18.932.0176.3292.8
✓✓15.325.5156.8232.7
✓✓16.928.7172.5281.6
✓✓✓14.124.3170.6286.8
✓14.430.5144.2262.6
✓✓13.425.2133.4224.0
✓✓✓13.323.4126.9222.3
✓✓✓✓11.419.7115.7199.8
", + "image_path": "edc58b2cfe53d6f985f743c689e5655874a9208d7a5cbf3adcc7d5b1671ec5e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 263, + 286, + 370 + ], + "blocks": [ + { + "bbox": [ + 72, + 239, + 261, + 250 + ], + "lines": [ + { + "bbox": [ + 72, + 239, + 261, + 250 + ], + "spans": [ + { + "bbox": [ + 72, + 239, + 261, + 250 + ], + "type": "text", + "content": "Table 3. Ablation studies on the model components." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 263, + 286, + 370 + ], + "lines": [ + { + "bbox": [ + 49, + 263, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 286, + 370 + ], + "type": "table", + "html": "
ParameterPropertyA → BA → Q
αPDE(%)MAEMSEMAEMSE
0.128.616.725.9157.3338.5
0.314.612.420.8132.1223.6
0.510.511.419.7115.7199.8
0.75.911.923.8123.8214.1
0.93.613.827.2124.7209.6
", + "image_path": "7783c255780f1add9655b5c17dffab792b40d2c529422979fe817afbc24c7055.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 66, + 403, + 269, + 551 + ], + "blocks": [ + { + "bbox": [ + 57, + 379, + 276, + 390 + ], + "lines": [ + { + "bbox": [ + 57, + 379, + 276, + 390 + ], + "spans": [ + { + "bbox": [ + 57, + 379, + 276, + 390 + ], + "type": "text", + "content": "Table 4. Ablation studies on the error threshold " + }, + { + "bbox": [ + 57, + 379, + 276, + 390 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 57, + 379, + 276, + 390 + ], + "type": "text", + "content": " in Eq. (6)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 403, + 269, + 551 + ], + "lines": [ + { + "bbox": [ + 66, + 403, + 269, + 551 + ], + "spans": [ + { + "bbox": [ + 66, + 403, + 269, + 551 + ], + "type": "table", + "html": "
ParametersA → BA → Q
MCMAEMSEMAEMSE
25625615.924.0138.4226.1
51225613.728.1124.4217.2
102425611.419.7115.7199.8
204825612.023.3128.1218.1
102412816.328.6127.4223.3
102425611.419.7115.7199.8
102451214.327.7130.1250.1
1024102414.224.8128.4220.2
", + "image_path": "6ff3ac72241aff4eb1f026bf50c0c3d917faa7978d50c2c0e248a6b79b4ef2d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 559, + 287, + 581 + ], + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 581 + ], + "type": "text", + "content": "Table 5. Ablation studies on the number of memory vectors " + }, + { + "bbox": [ + 46, + 559, + 287, + 581 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 559, + 287, + 581 + ], + "type": "text", + "content": " and vector size " + }, + { + "bbox": [ + 46, + 559, + 287, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 559, + 287, + 581 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": "Effects of " + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": ". The number of memory vectors " + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": " and dimension of each vector " + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": " are two important parameters of AMB. To demonstrate their effects independently, we adjust one parameter while keeping the other fixed. The results in Tab. 5 shows that among all variations of parameters, " + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "inline_equation", + "content": "M = 1024" + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": " is the optimal choice." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Visualizations of CEMs. We present visualizations of CEMs in Fig. 6, each summed along the channel dimension. The results indicate that crowd information is generally more" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 348, + 109 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 348, + 109 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 348, + 109 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 348, + 109 + ], + "type": "image", + "image_path": "cc23f6061a32fa91dda40a62ab46f563aaba4f86a34f6589342c9313ef33a5a6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 118, + 529, + 128 + ], + "lines": [ + { + "bbox": [ + 321, + 118, + 529, + 128 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 529, + 128 + ], + "type": "text", + "content": "Figure 6. Visualized CEMs. Yellower means less filtered." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 349, + 70, + 386, + 109 + ], + "blocks": [ + { + "bbox": [ + 349, + 70, + 386, + 109 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 386, + 109 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 386, + 109 + ], + "type": "image", + "image_path": "356579510cf69056b7b29fbe086087f3ae2c0861eafab076b2e729904daa0ca1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 386, + 70, + 425, + 109 + ], + "blocks": [ + { + "bbox": [ + 386, + 70, + 425, + 109 + ], + "lines": [ + { + "bbox": [ + 386, + 70, + 425, + 109 + ], + "spans": [ + { + "bbox": [ + 386, + 70, + 425, + 109 + ], + "type": "image", + "image_path": "3f7bba4ed8dfe203fbb76915cacf53ea698797efa04a3c0e367583ad8127845a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 505, + 109 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 505, + 109 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 505, + 109 + ], + "type": "image", + "image_path": "66f46a03c57008583fe811b9e36cf6b47fb421739408fba5dff2c86dedc7a968.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 506, + 70, + 543, + 109 + ], + "blocks": [ + { + "bbox": [ + 506, + 70, + 543, + 109 + ], + "lines": [ + { + "bbox": [ + 506, + 70, + 543, + 109 + ], + "spans": [ + { + "bbox": [ + 506, + 70, + 543, + 109 + ], + "type": "image", + "image_path": "25ca469d8438797335835d5528cd69768902a15075b74d1314a43165ba1681d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 140, + 366, + 185 + ], + "blocks": [ + { + "bbox": [ + 308, + 140, + 366, + 185 + ], + "lines": [ + { + "bbox": [ + 308, + 140, + 366, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 366, + 185 + ], + "type": "image", + "image_path": "cd211b7d1a61efbb4e8ee5e944203bbe1c6d47e64ebd1f4718a5d747a1bfc1b4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 185, + 366, + 228 + ], + "blocks": [ + { + "bbox": [ + 308, + 185, + 366, + 228 + ], + "lines": [ + { + "bbox": [ + 308, + 185, + 366, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 366, + 228 + ], + "type": "image", + "image_path": "89e7d6a0a103ceec615581142a89c86886c40de7dae79a55c7841b53da664fb7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 228, + 366, + 272 + ], + "blocks": [ + { + "bbox": [ + 308, + 228, + 366, + 272 + ], + "lines": [ + { + "bbox": [ + 308, + 228, + 366, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 366, + 272 + ], + "type": "image", + "image_path": "16386179d7e8aeb5adaa9b80042426adf1b30479adc192702f7e03c605e26c7f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 274, + 353, + 283 + ], + "lines": [ + { + "bbox": [ + 321, + 274, + 353, + 283 + ], + "spans": [ + { + "bbox": [ + 321, + 274, + 353, + 283 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 367, + 140, + 425, + 185 + ], + "blocks": [ + { + "bbox": [ + 367, + 140, + 425, + 185 + ], + "lines": [ + { + "bbox": [ + 367, + 140, + 425, + 185 + ], + "spans": [ + { + "bbox": [ + 367, + 140, + 425, + 185 + ], + "type": "image", + "image_path": "db91b594668215fbe44b9bbd0cf8c688505ee3c3599c6093a3901998b0521265.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 367, + 185, + 425, + 272 + ], + "blocks": [ + { + "bbox": [ + 367, + 185, + 425, + 272 + ], + "lines": [ + { + "bbox": [ + 367, + 185, + 425, + 272 + ], + "spans": [ + { + "bbox": [ + 367, + 185, + 425, + 272 + ], + "type": "image", + "image_path": "64cdb54bf3a203785067fa227d11be96cf9524aa6b1754a59da2e934415d2210.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 274, + 408, + 283 + ], + "lines": [ + { + "bbox": [ + 385, + 274, + 408, + 283 + ], + "spans": [ + { + "bbox": [ + 385, + 274, + 408, + 283 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 426, + 140, + 485, + 185 + ], + "blocks": [ + { + "bbox": [ + 426, + 140, + 485, + 185 + ], + "lines": [ + { + "bbox": [ + 426, + 140, + 485, + 185 + ], + "spans": [ + { + "bbox": [ + 426, + 140, + 485, + 185 + ], + "type": "image", + "image_path": "eafc32d6f395f3c3fe5ef9060741fc24ec5bc640404929370bd6893758d04452.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 426, + 185, + 485, + 272 + ], + "blocks": [ + { + "bbox": [ + 426, + 185, + 485, + 272 + ], + "lines": [ + { + "bbox": [ + 426, + 185, + 485, + 272 + ], + "spans": [ + { + "bbox": [ + 426, + 185, + 485, + 272 + ], + "type": "image", + "image_path": "1aceeabf9e4476c4590acaaf069a364ba2188429e32b1cbcc2d2d41f001ec97f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 274, + 477, + 283 + ], + "lines": [ + { + "bbox": [ + 435, + 274, + 477, + 283 + ], + "spans": [ + { + "bbox": [ + 435, + 274, + 477, + 283 + ], + "type": "text", + "content": "(c) Predicted" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 485, + 140, + 544, + 185 + ], + "blocks": [ + { + "bbox": [ + 485, + 140, + 544, + 185 + ], + "lines": [ + { + "bbox": [ + 485, + 140, + 544, + 185 + ], + "spans": [ + { + "bbox": [ + 485, + 140, + 544, + 185 + ], + "type": "image", + "image_path": "51c19f01d875549ace4a2b875b8fd509639a96c7a5c2cef284a2b6d29597a2e4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 485, + 185, + 544, + 272 + ], + "blocks": [ + { + "bbox": [ + 485, + 185, + 544, + 272 + ], + "lines": [ + { + "bbox": [ + 485, + 185, + 544, + 272 + ], + "spans": [ + { + "bbox": [ + 485, + 185, + 544, + 272 + ], + "type": "image", + "image_path": "559708f2bde012ed0fe22fca3512ac092e69f27b895a31b4783957e0552ef4ce.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 274, + 537, + 283 + ], + "lines": [ + { + "bbox": [ + 493, + 274, + 537, + 283 + ], + "spans": [ + { + "bbox": [ + 493, + 274, + 537, + 283 + ], + "type": "text", + "content": "(d) Binarized" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 293, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 306, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 545, + 326 + ], + "type": "text", + "content": "Figure 7. Visualization of (a) sample images from SHB, (b) the ground-truth PCMs, (c) PCMs predicted by MPCount trained on SHA and (d) predicted PCMs after binarization." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 348, + 484, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 484, + 359 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 484, + 359 + ], + "type": "text", + "content": "sensitive to domain shift than environments." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 360, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 468 + ], + "type": "text", + "content": "Visualizations of PCMs. We visualize the predicted and binarized PCMs from MPCount for qualitative analysis. As shown in Fig. 7, MPCount predicts accurate PCMs that provide reliable information of crowd locations. In the raw predicted PCMs, there are unconfident predictions with low classification scores, which are filtered out in the binarized PCMs. The results verify that PC can effectively mitigate the label ambiguity problem and contribute to a more robust crowd counting model against domain shift." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 480, + 378, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 378, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 378, + 492 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 501, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 713 + ], + "type": "text", + "content": "This paper proposes MPCount to address single domain generalization for crowd counting with possibly narrow source distribution. MPCount tackles two unique challenges, namely, density regression and label ambiguity. We propose an attention memory bank to reconstruct domain-invariant features for regression, with the content error mask eliminating domain-related content information and the attention consistency loss ensuring the consistency of memory vectors. Patch-wise classification is a novel auxiliary task to mitigate the ambiguous pixel-level labels with accurate information, enhancing the robustness of density predictions. Extensive experiments on well-known datasets show that MPCount achieves significantly the best results on various benchmarks by outperforming the state of the art under different unseen scenarios and narrow source domain. Acknowledgement. We would like to thank the Turing AI Computing Cloud (TACC) [42] and HKUST iSING Lab for providing us computation resources on their platform." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28032" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 114, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 190 + ], + "type": "text", + "content": "[1] Haoyue Bai, Hao He, Zhuoxuan Peng, Tianyuan Dai, and S.-H. Gary Chan. CounTr: An End-to-End Transformer Approach for Crowd Counting and Density Estimation. In Proceedings of European Conference on Computer Vision (ECCV) International Workshop on Distributed Smart Cameras, pages 207–222, Tel Aviv, Israel, 2022. Springer Lecture Notes in Computer Science. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 192, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 246 + ], + "type": "text", + "content": "[2] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. Proceedings of the AAAI Conference on Artificial Intelligence, 37:259-267, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[3] Yang Chen, Yu Wang, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. A style and semantic memory mechanism for domain generalization*. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9144-9153, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[4] Jian Cheng, Haipeng Xiong, Zhiguo Cao, and Hao Lu. Decoupled two-stage crowd counting and beyond. IEEE Transactions on Image Processing, 30:2862-2875, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 336, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 402 + ], + "type": "text", + "content": "[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 2, 3, 6, 7, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[6] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 561-570, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[7] Junyu Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "text", + "content": "[8] Junyu Gao, Wei Lin, Bin Zhao, Dong Wang, Chenyu Gao, and Jun Wen. " + }, + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^3" + }, + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "text", + "content": " framework: An open-source pytorch code for crowd counting. arXiv preprint arXiv:1907.02724, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 580 + ], + "type": "text", + "content": "[9] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[10] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and S.-H. Gary Chan. Target-agnostic Source-free Domain Adaptation for Regression Tasks. In Proceedings of 40th IEEE International Conference on Data Engineering (ICDE'24), Utrecht, Netherlands, 13-16 May 2024 (to appear). IEEE. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "text", + "content": "[11] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Ali Al-Maadeed, Nasir M. Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In European Conference on Computer Vision, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[12] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Stylenormalization and restitution for domain generalization and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "adaptation. IEEE Transactions on Multimedia, 24:3636-3651, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "text", + "content": "[13] Hyeonseong Kim, Yoonsu Kang, Changgyoon Oh, and Kuk-Jin Yoon. Single domain generalization for lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17587-17598, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "text", + "content": "[14] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4350, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 205, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 261 + ], + "type": "text", + "content": "[15] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "text", + "content": "[16] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[17] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 349, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 415 + ], + "type": "text", + "content": "[18] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 415, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 437 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 438, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 482 + ], + "type": "text", + "content": "[20] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6141-6150, 2019. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[21] Lucas Mansilla, Rodrigo Echeveste, D.H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "text", + "content": "[22] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15529-15539, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[23] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 2, 6, 7, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[24] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 6, 7, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28033" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "text", + "content": "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "type": "text", + "content": "[28] Vishwanath A. Sindagi and Vishal M. Patel. Cnn-based cascaded multi-task learning of high-level prior and density estimation for crowd counting. 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), pages 1-6, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "type": "text", + "content": "[29] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2594-2609, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 268, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 287, + 301 + ], + "type": "text", + "content": "[30] Leslie N. Smith and Nicholay Topin. Super-convergence: very fast training of neural networks using large learning rates. In *Defense + Commercial Sensing*, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 303, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 357 + ], + "type": "text", + "content": "[31] Qingyu Song, Changan Wang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Jian Wu, and Jiayi Ma. To choose or to fuse? scale selection for crowd counting. The Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21), 2021. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "text", + "content": "[32] Maryam Sultana, Muzammal Naseer, Muhammad Haris Khan, Salman Khan, and Fahad Shahbaz Khan. Self-distilled vision transformer for domain generalization. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3068-3085, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "text", + "content": "[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. ArXiv, abs/1607.08022, 2016. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 452, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 506 + ], + "type": "text", + "content": "[34] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 509, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 574 + ], + "type": "text", + "content": "[35] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xiansheng Hua. Meta convolutional neural networks for single domain generalization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4672-4681, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 577, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 621 + ], + "type": "text", + "content": "[36] Jia Wan, Ziquan Liu, and Antoni B. Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1974-1983, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "text", + "content": "[37] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai. Distribution matching for crowd counting. In Advances in Neural Information Processing Systems, 2020. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[38] Mingjie Wang, Hao Cai, Xianfeng Han, Jun Zhou, and Minglun Gong. Stnet: Scale tree network with multi-level auxilator for crowd counting. IEEE Transactions on Multimedia, 25:2074-2084, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[39] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8198-8207, 2019. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[40] Qi Wang, Tao Han, Junyu Gao, and Yuan Yuan. Neuron linear transformation: Modeling the domain shift for crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 33(8):3238-3250, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "text", + "content": "[41] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "text", + "content": "[42] Kaiqiang Xu, Xinchen Wan, Hao Wang, Zhenghang Ren, Xudong Liao, Decang Sun, Chaoliang Zeng, and Kai Chen. Tacc: A full-stack cloud computing infrastructure for machine learning tasks. arXiv preprint arXiv:2110.01556, 2021.8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "text", + "content": "[43] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI Conference on Artificial Intelligence, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[44] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-841, 2015. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[45] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "text", + "content": "[46] Yuxiang Zhang, Wei Li, Weidong Sun, Ran Tao, and Qian Du. Single-source domain expansion network for cross-scene hyperspectral image classification. IEEE Transactions on Image Processing, 32:1498-1512, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "text", + "content": "[47] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "type": "text", + "content": "[48] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[49] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[50] Su Zixian, Yao Kai, Yang Xi, Wang Qiufeng, Sun Jie, and Huang Kaizhu. Rethinking data augmentation for single-source domain generalization in medical image segmentation. In AAAI, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28034" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_content_list.json b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..475df51cc836cae753ff964382982eb1d5183beb --- /dev/null +++ b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_content_list.json @@ -0,0 +1,2164 @@ +[ + { + "type": "text", + "text": "Single Mesh Diffusion Models with Field Latents for Texture Generation", + "text_level": 1, + "bbox": [ + 119, + 130, + 849, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thomas W. Mitchel1,2† Carlos Esteves1 Ameesh Makadia1 1Google Research 2PlayStation", + "bbox": [ + 233, + 179, + 730, + 223 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a00a360104b82d0ae3fe81b413533e9f05a385623c4ba7b5cd73c16127e41b35.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 78, + 258, + 220, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c9f448fcd5320c1a5bdec708ab2a363a80c47d43271add1701150559da97947.jpg", + "image_caption": [ + "Figure 1. Our latent diffusion models operate directly on the surfaces of 3D shapes, synthesizing new high-quality textures (center) after training on a single example (left). Both our novel latent representation and diffusion models are isometry-equivariant, facilitating a notion of generative texture transfer by sampling pre-trained models on new geometries (right). Above our models are conditioned on coarse semantic labels reflecting a subjective distribution of content, which delineate the sole and interior of the shoes, the eyes and mouths of the skulls, and the decals on the bottles." + ], + "image_footnote": [], + "bbox": [ + 241, + 273, + 354, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cced20cab01424c70df95a6caa97ce3cdf1b93275382754214b82dbcf2c7c992.jpg", + "image_caption": [ + "Generated textures" + ], + "image_footnote": [], + "bbox": [ + 364, + 273, + 475, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4ebd282a16e9549f0749cfa4eb455fd69342bf55cc0b3f72aa1a2659af62431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 273, + 596, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d27906b94c9926129db751a83f30e561506b97bf8be3e715321946d534d20306.jpg", + "image_caption": [ + "Generative transfer to new geometry" + ], + "image_footnote": [], + "bbox": [ + 622, + 261, + 767, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a90d4f2212b730fe890a0efbb1323d3bfd6e12568f0b5684a9689b2865fdddf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 276, + 887, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 607, + 313, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a framework for intrinsic latent diffusion models operating directly on the surfaces of 3D shapes, with the goal of synthesizing high-quality textures. Our approach is underpinned by two contributions: Field Latents, a latent representation encoding textures as discrete vector fields on the mesh vertices, and Field Latent Diffusion Models, which learn to denoise a diffusion process in the learned latent space on the surface. We consider a single-textured-mesh paradigm, where our models are trained to generate variations of a given texture on a mesh. We show the synthesized textures are of superior fidelity compared those from existing single-textured-mesh generative models. Our models can also be adapted for user-controlled editing tasks such as inpainting and label-guided generation. The efficacy of our approach is due in part to the equivariance", + "bbox": [ + 75, + 641, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of our proposed framework under isometries, allowing our models to seamlessly reproduce details across locally similar regions and opening the door to a notion of generative texture transfer. Code and visualizations are available at https://single-mesh-diffusion.github.io/.", + "bbox": [ + 500, + 609, + 892, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 734, + 632, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of latent diffusion models (LDMs) [44] as powerful tools for 2D content creation has motivated efforts to replicate their success across different modalities. A particularly attractive direction is the synthesis of textured 3D assets, due to the high cost of building photorealistic objects which can require intricately configured scanning solutions [13] or expertise in 3D modeling.", + "bbox": [ + 496, + 762, + 892, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The majority of 3D synthesis methods employing LDMs operate in an image-based setting where the 3D repre", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done while at Google Research.", + "bbox": [ + 101, + 886, + 308, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sentation is optimized through differentiably rendered images, often in conjunction with score distillation sampling (SDS) [3, 4, 32, 41-43, 52, 53]. This design allows the models to leverage the full capabilities of pre-trained 2D LDMs. A second, alternative method is to rasterize both geometry and texture onto a 3D grid which are compressed into triplane latent features over which the DM operates [58]. However, in applications where texture synthesis takes precedence over 3D geometry, both approaches sacrifice fidelity in favor of a workable representation — models relying on planar renderings suffer from view inconsistencies when mapping back to the 3D shape while rasterization inherently aliases fine textural details.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we observe that the discretization of a surface as a triangle mesh is itself a tractable and effective representational space. To this end, we present an original framework for intrinsic LDMs operating directly on the surfaces of shapes, with the goal of generating high-quality textures.", + "bbox": [ + 76, + 289, + 468, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method consists of two distinct novel components, serving as the main contributions of our work. The first is a latent representation called Field Latents (FLs). Textures are mapped to tangent vector features at the vertices which are responsible for characterizing the local texture. The choice of tangent vector features over scalars allows for the capture of directional information related to the local texture which we show enables superior quality reconstructions. More generally, FLs offer an effective form of perceptual compression in which a high-resolution texture, analogous to a continuous signal defined over a surface, is mapped to a collection of discrete vector fields taking values at the vertices of a lower-resolution mesh. The second component is a Field Latent Diffusion Model (FLDM), which learns a denoising diffusion process in field latent space. Our FLDMs are built upon field convolutions [35], surface convolution operators designed specifically to process tangent vector features.", + "bbox": [ + 75, + 367, + 472, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although we outline a general framework for constructing diffusion models for surfaces, inspired by the success of single-image generative models [28, 38, 46, 54], we tailor and deploy our FLDM architecture in the single-textured-mesh setting. We address the problem of generating variations of a given texture on a mesh (Figure 1, left and center), as well as tasks supporting user control such as label-guided generation and inpainting. The single-textured-mesh setting also lets us circumvent the data scarcity issue. Large-scale 3D datasets with high-quality textures have only recently become available [8, 9], with only a fraction of samples across disparate categories possessing complex non-uniform textures. Furthermore, many textured 3D assets that are handcrafted or derived from real-world scans are geometrically or stylistically unique, upon which large-scale models can be challenging to condition.", + "bbox": [ + 75, + 641, + 472, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments reveal our models enable flexible synthe", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sis of high-resolution textures of qualitatively superior fidelity compared to existing single-textured-asset diffusion models, while simultaneously sidestepping the challenges of using 2D-to-3D approaches. Furthermore, both FLs and FLDMs are isometry-equivariant — each commute with distance-preserving shape deformations. As a result, we find that our approach can be used for generative texture transfer, wherein a FLDM trained on a single textured mesh can be sampled on a second similar mesh to texture it in the style of the first (Figure 1, right).", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing generative models that synthesize textured 3D assets typically map content to an internal 2D representation which is computationally preferable to working directly over 3D Euclidean space. Perhaps the most popular approach involves iteratively rendering objects from different viewpoints, enabling the use of pre-trained 2D LDMs (e.g. Stable Diffusion [44]) either indirectly as optimization priors [30, 32, 41, 42] or to directly apply a full denoising process [3, 4, 43, 53]. The former case, pioneered by DreamFusion [41], uses score distillation sampling (SDS), in which a neural radiance field is optimized such that the rendered images appear to be reasonable samples from a pre-trained LDM; in the latter, LDMs with depth conditioning are used to iteratively denoise the rendered images which are aggregatively projected to the texture map. However, SDS-based methods are computationally feasible only with low-resolution LDMs and small NeRFs, and are unable to synthesize fine details. Additionally, approaches that directly denoise and project must contend with artifacts arising from view inconsistencies and synthesized textures may contain unnatural tones or lighting effects as residuals from their image-trained LDM backbone.", + "bbox": [ + 496, + 280, + 893, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternative representational spaces to drive textured 3D synthesis have also been explored. Texture Fields [17, 39] encode textures as 3D neural fields and triplane feature maps have emerged recently as an efficient domain for 3D geometry [56, 58, 63]. In particular, Sin3DM [58] encodes geometry and texture into an implicit triplane latent representation, then trains and samples a 2D DM on the resulting feature maps. However, geometry and texture must first be rasterized to a 3D grid before triplane convolutions can be employed, which can lead to aliasing of high-fidelity textures and geometric details due to memory constraints on the maximal grid resolution. To disentangle geometry from texture, several methods map textures to domains such as the image plane [6, 40, 51, 60] or sphere [7]. Point-UV diffusion [61] trains and samples 2D DMs directly in the UV-atlas, though this approach suffers on meshes with disconnected or fragmented UV mappings.", + "bbox": [ + 496, + 613, + 893, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similar to our approach, a third class of models encode textures as features at the simplices of a mesh. GAN-based", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods Texturify [48] and Mesh2Tex [2] compress textures to attributes at the mesh facets, and differentiable rendering enables adversarial supervision in the image space.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, Intrinsic Neural Fields (INFs) [27] introduces a latent representation wherein textures are encoded as eigenfunctions of the Laplacian taking values at the mesh vertices. However, features at points on faces are recovered through barycentric interpolation, which we will show leads to a breakdown in reconstruction fidelity at higher compression ratios. Recently, Manifold Diffusion Fields [14, 64] introduce the first fully intrinsic DMs defined over 2D Riemannian manifolds, and are based on a global attention mechanism which aggregates features at the vertices, limiting their applicability to coarse, low-resolution meshes due to complexity constraints. Here, our proposed FLDMs are also fully intrinsic but are built upon locally-supported convolutions defined over the manifold [35], allowing them to scale to higher-resolution meshes.", + "bbox": [ + 75, + 136, + 468, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our single-textured-mesh FLDMs are inspired by recent work in which DMs are trained to capture internal patch distributions from a single image with the goal of generating diverse samples with similar visual content [28, 38, 54]; In turn, these models can trace their lineage to the influential SinGAN architecture [46] and its derivatives [19, 21]. Here, we mirror the approach proposed by SinDiffusion [54] and denoise with a shallow convolutional UNet to control the receptive field and prevent the DM from overfitting to the texture. Similar to our setting, Sin3DM [58] was the first to train LDMs in a single-textured-mesh paradigm, however they take an extrinsic approach to generating geometry and texture, whereas our approach focuses only on synthesizing textures which we show qualitatively are of higher-fidelity than those produced by Sin3DM.", + "bbox": [ + 75, + 363, + 468, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method Overview", + "text_level": 1, + "bbox": [ + 76, + 603, + 251, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a framework for intrinsic LDMs operating directly on surfaces. The first component is a latent representation of mesh textures comprised of tangent vector features. At each vertex, the tangent features characterize the texture of a local neighborhood on the surface, and the collection of these features constitute a stack of vector fields we call Field Latents (FLs). The FL space is learned with a locally-supported variational autoencoder [25] (FL-VAE) which is described in Section 4.", + "bbox": [ + 75, + 628, + 467, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second key component is the Field Latent Diffusion Model (FLDM), which learns to denoise a diffusion process in the tangent space of a surface. Denoising networks are constructed with field convolutions (FCs), surface convolution operators acting on tangent vector fields [35]. We extend FCs to interleave scalar embeddings with the tangent vector features, enabling the injection of diffusion timesteps and optional conditioning, such as user-specified labels, into the denoising model.", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, FL-VAEs and FLDMs are applied to synthesize new textures from a single textured triangle mesh. We pre-train a single FL-VAE by superimposing planar meshes over a large high-quality image dataset, obtaining a general-purpose latent space which can be used to train FLDMs on arbitrary textured meshes. To do so, a texture is first mapped to distributions in the tangent space at the vertices with the FL-VAE encoder, and an FLDM is trained to iteratively denoise tangent vector latent features. Afterwards, samples from the FLDM can be decoded with the FL-VAE decoder to synthesize new textures.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Importance of Isometry-Equivariance Our FL-VAEs and FLDMs are designed to be equivariant under isometries, i.e. they commute with distance-preserving shape deformations. Unlike images, points on a surface have no canonical orientation. The choice of local coordinates is ambiguous up to an arbitrary rotation, making it impossible to simply adapt standard image-based VAEs and diffusion models to surfaces. However, isometries manifest locally as rotations. Thus, designing our framework to be isometry-equivariant inherently solves the orientation ambiguity problem, enabling consistent, repeatable results. Isometry-equivariance also results in tangible benefits — our models are able to seamlessly reproduce textural details across locally similar areas of meshes. Furthermore, while existing models transfer textures with pointwise maps [11] or generatively via conditioning on a learned token [5, 43], we can simply sample our pre-trained FLDMs on new, similar meshes to texture them in the learned style.", + "bbox": [ + 496, + 276, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Field Latents", + "text_level": 1, + "bbox": [ + 500, + 561, + 635, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As in Knoppel et al. [26] we associate tangent vectors with complex numbers. For a surface $M$ and point $p \\in M$ , we assign to the tangent space an arbitrary orthonormal basis such that for any $\\mathbf{v} \\in T_pM$ we have $\\mathbf{v} \\equiv r e^{i\\theta}$ , with $r = |\\mathbf{v}|$ and $\\theta$ the direction of $\\mathbf{v}$ expressed in the frame. We make this convention explicit by denoting $\\mathbb{C}_p \\equiv T_pM$ .", + "bbox": [ + 496, + 587, + 890, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To describe the FL-VAE, we require a notion of multidimensional Gaussian noise in the tangent bundle of a surface $M$ . Specifically, we denote the Gaussian distribution over $d$ copies of the tangent bundle $TM^{d}$ by $\\mathcal{T}\\mathcal{N}_M(0,I_d)$ . Samples $\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)$ are vector fields in $TM^{d}$ such that at each point $p\\in M$ , the coefficients of $\\epsilon (p)\\in \\mathbb{C}_p^d$ expressed in the orthonormal basis are themselves samples from the $d$ -dimensional standard normal distribution", + "bbox": [ + 496, + 678, + 890, + 797 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon (p) = \\epsilon_ {1} + i \\epsilon_ {2}, \\epsilon_ {1}, \\epsilon_ {2} \\sim \\mathcal {N} (0, I _ {d}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 811, + 890, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work we are concerned with isometries $\\gamma : M \\to N$ , which enjoy the property that their push-forwards $d\\gamma : TM \\to TN$ , which define a smooth map between tangent spaces, manifest pointwise as rotations. From Equa-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7955", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion (1) it is easy to see that $\\mathcal{T}\\mathcal{N}_M(0,I_d)$ is symmetric under such transformations. Thus, it follows that sampling from $\\mathcal{T}\\mathcal{N}_N(0,I_d)$ is equivalent to pushing forward samples from $\\mathcal{T}\\mathcal{N}_M(0,I_d)$ ,", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {\\prime} = \\gamma \\epsilon \\quad \\text {f o r} \\quad \\begin{array}{l} \\epsilon^ {\\prime} \\sim \\mathcal {T N} _ {N} (0, I _ {d}) \\\\ \\epsilon \\sim \\mathcal {T N} _ {M} (0, I _ {d}) \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 161, + 468, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\gamma \\epsilon = [d\\gamma \\cdot \\epsilon] \\circ \\gamma^{-1}$ denoting the standard action of diffeomorphisms on vector fields via left-shifts.", + "bbox": [ + 76, + 209, + 468, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FL-VAE Encoder The encoder $\\mathcal{E}$ is a network that, for any point $p\\in M$ , takes $n$ -dimensional scalar functions $\\psi \\in L^2 (M,\\mathbb{R}^n)$ (e.g. texture RGB values with $n = 3$ restricted to the surrounding geodesic neighborhood $B_{p}\\subset$ $M$ to the parameters of $d$ independent normal distributions in the tangent space at $p$ ..", + "bbox": [ + 76, + 258, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {E} _ {p}: L ^ {2} \\left(B _ {p}, \\mathbb {R} ^ {n}\\right)\\rightarrow \\mathbb {C} _ {p} ^ {d} \\times \\mathbb {R} _ {\\geq 0} ^ {d} \\tag {3} \\\\ \\psi \\mapsto \\left(\\mu_ {p} ^ {\\psi}, \\sigma_ {p} ^ {\\psi}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 358, + 468, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the means of the distributions $\\mu_p^\\psi$ are themselves tangent vectors while the standard deviations $\\sigma_p^\\psi$ are scalars. Latent codes characterizing the local restriction of $\\psi$ are collections of tangent vectors $z_{p}^{\\psi}\\in \\mathbb{C}_{p}^{d}$ drawn from the multivariate normal distribution parameterized by $(\\mu_p^\\psi ,\\sigma_p^\\psi)$ . Using the reparameterization trick [25] latent codes can be generated as $z_{p}^{\\psi} = \\mu_{p}^{\\psi} + \\sigma_{p}^{\\psi}\\odot \\epsilon (p)$ with $\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)$ .", + "bbox": [ + 76, + 410, + 468, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FL-VAE Decoder The principal advantage in a tangent vector latent representation is that it naturally admits a descriptive coordinate function which can be queried to associate features with neighboring points. The decoder $\\mathcal{D}$ is designed as a neural field operating over the local parameterization of the surface induced by the logarithm map about a point $p$ , $\\log_p:B_p\\to \\mathbb{C}_p$ , with $\\log_p q$ encoding the position of $q\\in B_{p}$ in the tangent space of $p$ . Formally,", + "bbox": [ + 76, + 537, + 468, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} _ {p}: \\log_ {p} \\left(B _ {p}\\right) \\times \\mathbb {C} _ {p} ^ {d} \\rightarrow \\mathbb {R} ^ {n} \\tag {4} \\\\ \\left(\\log_ {p} q, z _ {p} ^ {\\psi}\\right) \\mapsto \\widehat {\\psi} _ {p} (q), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 669, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\widehat{\\psi}_p$ denoting the prediction of $\\psi$ made from the perspective of $p$ . Decoding follows a similar process as proposed in [18]. Specifically, given a tuple $(\\log_p q, z_p^\\psi)$ , the decoder constructs two features. First, an invariant scalar feature is obtained by vectorizing the upper-triangular part of the Hermitian outer product of the latent code with itself, $\\mathrm{vec}_{j \\geq i} \\left(z_p^\\psi \\left[z_p^\\psi\\right]^*\\right) \\in \\mathbb{C}^{\\frac{d(d+1)}{2}}$ . Second, a positionally-aware scalar feature $c_{pq}^\\psi \\in \\mathbb{C}^d$ is formed by the natural coordinate function", + "bbox": [ + 76, + 723, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {p q} ^ {\\psi} \\equiv \\log_ {p} q \\cdot \\overline {{z _ {p} ^ {\\psi}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 876, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which corresponds to storing the inner product and determinant of each tangent vector with the position of $q$ relative to $p$ given by $\\log_p q$ . The neural field receives the concatenation of these two features and returns the prediction $\\widehat{\\psi}_p(q)$ .", + "bbox": [ + 498, + 90, + 890, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equivalence Recall we are concerned with constructing a latent representation equivariant under isometries — that for any $\\psi \\in L^2(M, \\mathbb{R}^n)$ and isometry $\\gamma: M \\to N$ we have", + "bbox": [ + 498, + 172, + 890, + 218 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\widehat {\\gamma \\psi} \\right] _ {\\gamma (p)} = \\gamma \\widehat {\\psi} _ {p}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 227, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma \\psi = \\psi \\circ \\gamma^{-1}$ denotes the standard action of diffeomorphisms by left-shifts. Thus, it follows from Equation (2) that a sufficient condition for Equation (6) to hold is if for all $\\psi \\in L^2(M, \\mathbb{R}^n)$ , isometries $\\gamma : M \\to N$ , and $p \\in M$ , the encoded mean $\\mu_p^\\psi$ and standard deviation $\\sigma_p^\\psi$ satisfy", + "bbox": [ + 498, + 258, + 890, + 337 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. d \\gamma \\right| _ {p} \\cdot \\mu_ {p} ^ {\\psi} = \\mu_ {\\gamma (p)} ^ {\\gamma \\psi} \\quad \\text {a n d} \\quad \\sigma_ {p} ^ {\\psi} = \\sigma_ {\\gamma (p)} ^ {\\gamma \\psi}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 345, + 890, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $d\\gamma|_p$ denoting the push-forward of $\\gamma$ in the tangent space at $p$ , which rotates the local coordinate system. A detailed proof appears in the supplement, sec. 8.1.", + "bbox": [ + 498, + 376, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture In practice, we discretize a surface $M$ by a triangle mesh with vertices $V$ and faces $F$ . Textures are scalar functions $\\psi$ defined continuously over the faces and vertices, taking values in $\\mathbb{R}^3$ . For each $p \\in V$ , we consider its neighborhood to be the surrounding one-ring, the collection of faces in $F$ that share $p$ as a vertex. The encoder architecture is chosen so as to satisfy the conditions in Equation (7). A fixed number of points $\\{q_i\\}$ are uniformly sampled in the one-ring neighborhood at which the texture function is evaluated. The resulting scalar features $\\{\\psi(q_i)\\}$ are interleaved with the logarithms $\\{\\log_p q_i\\}$ to construct tangent vector features which are concatenated with a token following [12]. The features are passed to eight successive VN-Transformer layers [1], modified to process tangent vector features. Afterwards, the token is extracted and used to predict the mean and standard deviation.", + "bbox": [ + 496, + 439, + 890, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The neural field comprising the decoder consists of a five-layer real-valued MLP, to which the real and imaginary parts of the complex feature vector are passed. For any point in a triangle, the latent codes at the adjacent vertices provide three distinct predictions which are linearly blended with barycentric interpolation only at inference. This is in contrast to INF [27], wherein linearly blended features are passed to the neural field to make predictions. A comprehensive discussion of the FL-VAE architecture and training objective is provided in section 9.1 in the supplement.", + "bbox": [ + 496, + 681, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Field Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 844, + 774, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FLDMs define a noising process on surface vector fields, and learn a denoising diffusion probabilistic model", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7956", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c94607f7ffe0c85fc2d05ef3dd95e2c4ea4287e6166cb6507c43602592e69443.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FL-VAE22.38 (20.59)0.51 (0.83)1.02 (1.81)
FL-VAE (Bary.)21.33 (19.61)0.66 (1.01)1.31 (2.03)
INFs [27]18.86 (16.45)1.16 (1.64)2.15 (2.73)
", + "bbox": [ + 109, + 101, + 431, + 140 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Texture reconstruction quality on meshes from the Google Scanned Objects dataset [13]. Reconstructions are evaluated on high-res (30K vertices) and low-res (5K vertices, in parenthesis) meshings of the objects. Metrics are computed in the texture atlases, with DSSIM and LPIPS scaled by a factor of 10.", + "bbox": [ + 75, + 142, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(DDPM) [22] using an equivariant surface network. In principle they can be applied to any vector data on surfaces, and here we use them to learn feature distributions in FL space.", + "bbox": [ + 75, + 217, + 468, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent vector fields $Z \\in TM^{d}$ can be produced by compressing a texture $\\psi \\in L^{2}(M, \\mathbb{R}^{3})$ to a collection of tangent vector features at each point with the FL-VAE encoder as in Equation (3) and sampling from the distributions. The forward process progressively adds noise in the tangent space, dictated by a monotonically decreasing schedule $\\{\\alpha_{t}\\}_{0 \\leq t \\leq T}$ with $\\alpha_{0} = 1$ and $\\alpha_{T} = 0$ . The noised vector fields can be expressed at an arbitrary time $t$ in the discretized forward process with", + "bbox": [ + 75, + 263, + 468, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {t} = \\sqrt {\\alpha_ {t}} Z + \\sqrt {1 - \\alpha_ {t}} \\epsilon \\in T M ^ {d}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 405, + 468, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)$ as in Equation (1).", + "bbox": [ + 76, + 431, + 367, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The denoising function $\\varepsilon$ is a surface network trained to reverse the forward process. Specifically, it is a learnable map on the space of latent vector fields, conditioned on timesteps $t$ and optional scalar embeddings $\\rho \\in L^{2}(M,\\mathbb{R}^{m})$ , with the goal of predicting the added noise,", + "bbox": [ + 76, + 446, + 468, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\varepsilon : T M ^ {d} \\times \\mathbb {R} _ {\\geq 0} \\times L ^ {2} (M, \\mathbb {R} ^ {m}) \\rightarrow T M ^ {d}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 529, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "trained subject to the loss", + "bbox": [ + 76, + 556, + 248, + 570 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {F L D M}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {T N} _ {M} (0, I _ {d}), 0 \\leq t \\leq T} \\| \\epsilon - \\varepsilon \\left(Z _ {t}, t, \\rho\\right) \\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 577, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At inference, new latent vector fields are generated following the iterative DDPM denoising process. Beginning with representing the fully-noised vector fields at step $T$ by a sample $\\tilde{Z} \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)$ , $\\tilde{Z}_T = \\tilde{Z}$ , the trained denoising function $\\varepsilon$ is used to estimate the previous step in the forward process via the relation", + "bbox": [ + 76, + 604, + 468, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widetilde {Z} _ {t - 1} = C _ {1} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\widetilde {Z} _ {t} + C _ {2} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\varepsilon \\left(\\widetilde {Z} _ {t}, t, \\rho\\right) \\tag {11} \\\\ + C _ {3} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\epsilon , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 702, + 467, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\epsilon = 0$ for $t = 1$ and $\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)$ otherwise, and $C_i(\\alpha_t,\\alpha_{t - 1})\\in \\mathbb{R}$ as defined in 9.1 of the supplement.", + "bbox": [ + 76, + 747, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equivalence From the equivalence of the distributions $\\mathcal{T}\\mathcal{N}_M(0,I_d)$ and $\\mathcal{T}\\mathcal{N}_N(0,I_d)$ under isometries as in Equation (2), it follows that for any isometry $\\gamma :M\\to N$ a sufficient condition for both the forward and reverse processes in Equations (8) and (11) to satisfy", + "bbox": [ + 76, + 795, + 468, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma Z _ {t} = \\left[ \\gamma Z \\right] _ {t} \\quad \\text {a n d} \\quad \\gamma \\widetilde {Z} _ {t} = \\left[ \\gamma \\widetilde {Z} \\right] _ {t}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 877, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63aa0937b3f29c8b94195d24c6d5cee7638a21bdca27d7c809e816a3e9a4036f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 109, + 596, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae336f6b2f726b5de02550b5ebf7e68b2210b329ccbafdf06f88ee2138c31873.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 157, + 596, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70388add48c7ab22cf2ca238ca263be278195e2ada3d3339a86d2668c00135a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 111, + 691, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44f08460a5951f34d3f96f17cfa280c9cd3ad7031638c69d9321225cddf6df43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 157, + 694, + 226 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6a688597a5ff870651a05b2f0e38e1fb2ad063e274e5cbfaf8c3d580c4dbc04f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 111, + 787, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a342393567a96786c96eb29cfafa83b3faf22ef63c91d4598a030a12ddf3a8b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 111, + 885, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eaebf572d728ea734cb7aa96bc0801ca31832af8099aab4cb30dabc4899bbd83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 157, + 790, + 226 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02579870a66ff82c2480b1755db40022a86861c59411cf9a2fde4ad379bc3cdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 157, + 887, + 226 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1deb7074a634d3fddfa7522b18a356eb8e7a04524c3b68c5d37aa735cd2c0fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 233, + 593, + 270 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a7640465f10645669e6687715d83009fc5e35b4a45aa39ebbbd7649fab780506.jpg", + "image_caption": [ + "Figure 2. Visual comparison of textures compressed and reconstructed with the FL-VAE and INF [27] on 30K vertex meshes. Compared to barycentric coordinates, our proposed logarithmic coordinate function more richly extends latent features across the mesh, enabling the reconstruction of finer details. Zoom in to view." + ], + "image_footnote": [], + "bbox": [ + 504, + 271, + 591, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aeb2230304ddfa3e2618256dcb1c45c272da83db316f064b7b9062e28cdbd379.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 233, + 692, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b13595d074786f40764e31b3c1710f887a75c01b7084d7abce7a0db15faf0a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 272, + 691, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0aa349d9b15df4819afdad900a8967ad32e95a4917cbcee429b09deca8e68c7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 233, + 792, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c110f4efea2af05364ededd69d3518732e2bc890f4274d8895ff389d6e89053b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 272, + 790, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a53dd0f6ade80852fd2ef99fe245a55752d305902d47d930b7ecacd64354e073.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 233, + 890, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f0859b85236abd84b97d9f4baf2e643916bb89e4fdabe9356946babec1031dc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 272, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $0 \\leq t \\leq T$ is if the denoising network $\\varepsilon$ in Equation (9) has the property that for any vector field $Z \\in TM^{d}$ ,", + "bbox": [ + 498, + 422, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma \\varepsilon (Z, t, \\rho) = \\varepsilon (\\gamma Z, t, \\gamma \\rho). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 467, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A detailed proof is provided in sec. 8.2 of the supplement.", + "bbox": [ + 500, + 497, + 882, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denoising with Field Convolutions To satisfy the condition in Equation (12), our denoising network is constructed with field convolutions (FC) [35], which belong to a class of models that can operate on tangent vector fields [45, 57].", + "bbox": [ + 498, + 536, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FCs convolve vector fields $Z \\in TM^{C}$ with $C' \\times C$ filter banks $\\mathbf{f}_{c'c} \\in L^{2}(\\mathbb{C},\\mathbb{C})$ , returning vector fields $Z * \\mathbf{f} \\in TM^{C'}$ . In conventional diffusion models, features inside the denoising network are conditioned on timesteps $t$ via an additive embedding applied inside the convolution blocks. Unfortunately, no direct analogue exists for vector field features on surfaces — additive embeddings break equivariance and we find that multiplicative embeddings destabilize training. Thus, to enable conditioning of our denoising network $\\varepsilon$ on both timesteps $t$ and optional user-input features $\\rho \\in L^{2}(M,\\mathbb{R}^{m})$ , we inject embeddings directly into convolutions by extending FCs to convolve vector fields with filters $\\mathbf{f}_{c'c} \\in L^{2}(\\mathbb{C} \\times \\mathbb{R}^{e},\\mathbb{C})$ , with $e$ the embedding dimension. Critically, this formulation preserves equivariance without sacrificing stability.", + "bbox": [ + 498, + 597, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [54], the denoising network $\\varepsilon$ takes the form of a shallow two-level U-Net but with field convolutions; the relatively small size of the network's receptive field allows distributions of latent features to be learned without overfitting to the single textural example [28, 54, 58].", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Method", + "bbox": [ + 112, + 89, + 158, + 99 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PSNR↑", + "bbox": [ + 225, + 89, + 269, + 99 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "DSSIM", + "bbox": [ + 300, + 89, + 351, + 99 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LPIPS", + "bbox": [ + 372, + 89, + 419, + 99 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Ground truth", + "bbox": [ + 504, + 92, + 584, + 104 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "FL-VAE", + "bbox": [ + 614, + 92, + 671, + 104 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "FL-VAE (Bary.)", + "bbox": [ + 694, + 92, + 790, + 104 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "INFs [27]", + "bbox": [ + 812, + 92, + 870, + 104 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7957", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/89286772ec9ca37e6ef226617e7584df1bb16c2c37964c67f7070548d8051ac5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 79, + 156, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6613f4a9698ceb5e75e2b08d786bff7cd1003b115a6eea1e36dbdfcadfa49c92.jpg", + "image_caption": [ + "Figure 3. Visual comparison of unconditionally (label-free) generated textures from FLDMs and Sin3DM [58]. Our FLDM's isometry-equivariant construction allows for replication of textural details across locally similar regions. In contrast, Sin3DM's extrinsic approach associates textural features with 3D space; Synthesized details appear as repetitions or extrusions of previous patterns along the major axes of the mesh, and we observe that novel textures cannot be created without modifying geometry. Zoom in to compare." + ], + "image_footnote": [], + "bbox": [ + 88, + 171, + 158, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5283b246c8ecd84f70136fc56a568a6d95fa51b11153d0059994c15615374ed9.jpg", + "image_caption": [ + "FLDM" + ], + "image_footnote": [], + "bbox": [ + 169, + 98, + 230, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8604ae5206e9c1fe87abc7c2440a3084246b6286888b675292ebced34502433f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 101, + 307, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e4923766cf947fc7d7d9745ed3081f6f0fed49b82fe75ee42704ca32893a132.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 170, + 235, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16e6d47d14b2e30f986e94b85867a4ab77ce80380a652009699f5462c6f6922c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 171, + 310, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3fd7fbbca6319ae00f4572a50b323e97db54259e3d629ba5e944245fcbd1612.jpg", + "image_caption": [ + "Sin3DM [58]" + ], + "image_footnote": [], + "bbox": [ + 326, + 99, + 385, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/165e4738813c0d860ab3141560a148bee157967e3c1f00a5d0772defbcb31934.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 101, + 460, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/547c9e9021a7d0de0267627142322c2941bfcfe4a73e758f2447ae8ca61bc0af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 171, + 393, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0bbb7cf194d616ba200e2b29d659952f2b9cc4f5dcc2242b83c05ffd6b8a39d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 172, + 464, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8ac167370b7d8a63d65bd087caad30909aca91346ff76aeb9be7f969fec8df7.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 480, + 101, + 542, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5b1aa814fca902e7c7f692954e09aa5dc48a6d306295fc1a047a061e0a8ed15.jpg", + "image_caption": [ + "FLDM" + ], + "image_footnote": [], + "bbox": [ + 566, + 101, + 627, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c07ecad3d77d567a7c0e67950921e1fc3a4f1dd8be453f61c48b9f5c0a47181b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 101, + 710, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55d5c396fe30615aafcf0e4e07a9e8635292d28063d61ed73ca0ffa841d556ca.jpg", + "image_caption": [ + "Sin3DM [58]" + ], + "image_footnote": [], + "bbox": [ + 733, + 101, + 795, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/721c0cbd2fff4f86060934de71bd4e97803189172b4f2ca664fc7d87bbbf0c67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 101, + 880, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44cd2fd89375eba06d65280e4f46d1101ce39261a2f01efcaf59966169f69eae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 172, + 720, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95944ef82cf4e16994f6b8b162a6909666b1428bc680c4a8d2b217b98057a67d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 172, + 803, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fbcbf17b062b7b2ccfcfa262582c79117500b87a4e788a928742996988618dcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 172, + 888, + 236 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e5f800e10dbcab0c897c7e353e75dd359740064b87a61e753a85faf9317a5ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSIFID ↓LPIPS ↑
FLDM3.271.15
Sample (left)0.710.94
Sin3DM [58]6.582.20
Sample (right)2.422.91
", + "bbox": [ + 86, + 330, + 294, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6ed50c83dd3bed48f94ef1102ecc92854b7ffa4cf2b3dcd284cd2798ac4d0a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 321, + 375, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a17bb953d8d301a11e5657a74f90b11831f0b50e0dbac5324bdd3d56a0ec4363.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 321, + 450, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Fidelity and diversity of unconditionally generated textures trained on selected Objaverse [8] and Scanned Objects [13] meshes. SIFID and LPIPS are scaled by factors of $10^{5}$ and 10, respectively. Unlike our FLDMs, Sin3DM synthesizes new geometry which can inflate diversity scores. As an example, the individual metrics are shown for the FLDM and Sin3DM samples on the right. The latter attains a high LPIPS score, but exhibits poor textural diversity and quality.", + "bbox": [ + 75, + 417, + 468, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 539, + 209, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training the FL-VAE It is important to note that the FL-VAE proposed in Section 4 only interfaces with the mesh geometry through its local flattening in the image of the logarithm map. Thus, the FL-VAE operates independently of the local curvature, and a model pre-trained' on flat textured meshes can be used to encode and decode textures on arbitrary 3D shapes.", + "bbox": [ + 75, + 566, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In all of our experiments we use the same FL-VAE with a $d = 8$ dimensional latent space, pre-trained on the OpenImagesV4 dataset [29]. Specifically, we generate 1K different planar meshes containing 10K vertices each. Each sample $512 \\times 512$ image is overlaid by a randomly chosen triangulation, with the image serving as the \"texture\" defined over the mesh. The FL-VAE is then trained with the sum of reconstruction loss and KL divergence using the Adam [24] optimizer. We deliberately train the FL-VAE on coarser triangulations than we expect during inference on textured meshes. This forces the model to reconstruct high frequencies during training, which we find increases robustness to sampling density and triangulation quality, in addition to improving the fidelity of reconstructed FLDM samples.", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training single-textured-mesh FLDMs All diffusion experiments share the same training regime. During preprocessing, we create 500 copies of a given mesh each with 30K vertices and different triangulations to prevent the denoising network from learning the connectivity. Each copy shares the same UV-atlas as the original and the texture is mapped to vector field latent features at the vertices using the pre-trained FL-VAE encoder. FLDMs have a max timestep of $T = 1000$ and noise is added on a linear schedule. The denoising network is trained subject to the loss in Equation (10) using the Adam [24] optimizer.", + "bbox": [ + 496, + 333, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Texture Compression and Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 512, + 854, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First we quantitatively evaluate the capacity of our proposed FL-VAE for compressing and reconstructing textures. We compare FL-VAEs against two other approaches: the state-of-the-art INF [27] and a modified version of FL-VAEs serving as an ablation. INFs train a neural field per texture and predicts the texture values using the values of the Laplacian eigenfunctions extended over triangles through barycentric interpolation. Similarly, the modified FL-VAE (denoted FL-VAE Barycentric) omits our proposed logarithmic coordinate function in Equation (5) in favor of linearly interpolating the values of the invariant outer product features across triangles. It is otherwise identical to our proposed approach, including pre-training methodology. Thus, this also serves to compare the efficacy of our proposed coordinate function against barycentric interpolation in extending vertex features continuously over the mesh.", + "bbox": [ + 496, + 537, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate, we select 16 objects exhibiting a range of complex textures with high-frequency details from the Google Scanned Objects dataset [13]. Two versions of the dataset are used, with objects remeshed to have $30\\mathrm{K}$ (high-res) and $5\\mathrm{K}$ (low-res) vertices, respectively. The single pretrained FL-VAE is used to map the textures to distributions over the tangent space, and the mean is used to sample the reconstructed texture at the pixel indices in the texture atlas.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7958", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ccdf64a5f1b102d6d0808145bdd38b6605576cf305c631992e128d922208115.jpg", + "image_caption": [ + "Figure 4. Label-guided generation. FLDMs can be conditioned on a subjective user-input labeling which generated textures will reflect. See Figure 1 for more examples. Zoom in to view." + ], + "image_footnote": [], + "bbox": [ + 78, + 78, + 468, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A separate INF neural field is trained for each mesh, using the specified regime for high-res reconstruction [27].", + "bbox": [ + 75, + 234, + 468, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following INFs, we evaluate the quality of the reconstructed textures using the average PSNR, DSSIM [55], and LPIPS [62] metrics computed in texture atlases. The results are reported in Table 1, with several examples of reconstructions shown in Figure 2. Our proposed FL-VAE achieves the best results across all metrics at both resolutions, followed by the barycentric FL-VAE and INFs. The significant gap in performance between the barycentric FL-VAE and INFs is likely due to the reliance of the latter the on the Laplacian eigenfunctions as input features. Areas of fine textural detail may not necessarily overlap with regions where the eigenfunctions exhibit high variability, potentially contributing to the \"raggedness\" observed in INF reconstructions. Furthermore, the superior performance of our proposed FL-VAE relative to its barycentric counterpart suggests that our logarithmic coordinate function in Equation (5) provides a richer interpolant over the mesh faces. This directly increases the representational capacity of the latent features, enabling the reconstruction of finer textural details with the same number of features.", + "bbox": [ + 75, + 265, + 468, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Unconditional (Label-Free) Generation", + "text_level": 1, + "bbox": [ + 76, + 575, + 413, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Label-free generation can be performed by training and sampling FLDMs conditioned only on the diffusion timescale (i.e. $\\rho = 0$ ). We compare against Sin3DM [58] which to our knowledge is the only prior generative model designed to operate in a single-textured-mesh paradigm. We follow their proposed evaluation protocol, and train FLDMs and Sin3DM on 10 assets from the Objaverse [8] and Scanned Objects [13] datasets selected for texture complexity and diversity. For each asset, 64 textures are sampled from the trained models, which are rendered on the mesh from 24 different views. Fidelity is measured with the mean SIFID (Single Image Fréchet Inception Distance) [46] pairwise between the renderings and those of the input mesh from the same view; the mean LPIPS [62] between renders of samples from the same viewpoint measures diversity. We note that unlike FLDMs, which only synthesize new textures, Sin3DM synthesizes both textures and geometries, with the latter potentially inflating LPIPS scores.", + "bbox": [ + 75, + 598, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are shown in Table 2, Figure 3, and in the supplement (Figure 8). Compared to Sin3DM, our FLDMs", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b61e246cc5e841824e8907e64d220456dc4a5b6718c26610993c0e1f1b1a0ac.jpg", + "image_caption": [ + "Figure 5. Inpainting with FLDMs. The input texture is preserved in the masked regions and new content is synthesized elsewhere with agreement on the boundaries. Zoom in to view." + ], + "image_footnote": [], + "bbox": [ + 504, + 78, + 887, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are able to synthesize higher-fidelity textures which is likely due in part to the former's rasterization of textured geometry to a 3D grid before encoding to the latent space, potentially aliasing fine details. More generally, we observe that by operating over an extrinsic representation of textured geometry, Sin3DM intertwines texture with a mesh's 3D embedding. Thus, new textural details appear as repetitions or extrusions of earlier patterns along the major axes of the models, as seen in the vase, snake, and sculpture samples in Figure 3. Furthermore, we observe that texture and position are so strongly linked that Sin3DM is unable to produce significant textural alterations without correspondingly modifying the geometry, as seen in the vase, snake, and lantern samples. Here, the advantage of our isometry-equivariant formulation becomes clear as our FLDMs can effectively replicate textural details across areas of the mesh that are locally similar, regardless of the relative configuration of these regions in the 3D embedding.", + "bbox": [ + 496, + 233, + 892, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Label-Guided Generation", + "text_level": 1, + "bbox": [ + 500, + 513, + 733, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Textured objects often exhibit distinct content specific to certain regions of the mesh which may not be geometrically unique; for example, in Figure 1 regions on the sole, body, and inside of the shoe lack distinguishing geometric features, as is the case with the eyes and mouth of the skull and the decal on the bottle. In such cases, a user may find it desirable for synthesized textures to reflect a subjective distribution of content on the input mesh, which we facilitate by extending our denoising model to incorporate optional conditioning from user-designated mesh features $\\rho \\in L^{2}(M,\\mathbb{R}^{m})$ . In the simplest case, the user can paint a coarse, semantic labeling to subjectively delineate salient regions, such as the flippers, head, and rock base of the seal sculpture in Figure 4. After training an FLDM conditioned on the labels, generated textures reflect the specified segmentation. Further examples are shown in Figure 1. Labeled areas contain different textural features, and conditioning ensures the FLDM samples respect these distributions, e.g. that teeth are synthesized only on the mouth.", + "bbox": [ + 496, + 536, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Inpainting", + "text_level": 1, + "bbox": [ + 500, + 832, + 617, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In certain cases a user may wish to preserve the input texture in certain regions while synthesizing new content elsewhere. Texture inpainting is a well studied topic", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7959", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f72422ff655744fd6b1f7f490df08dc519195e489d602626c39463dbd4a07826.jpg", + "image_caption": [ + "FLDM samples across geometries", + "Figure 6. Generative texture transfer. Since our FL-VAE and FLDMs commute with isometries, we can encode a texture and train an FLDM on an input mesh (left), then sample the pre-trained FLDM on a new, similar mesh and decode the latents to texture it (center to right). Conditioning labels segmenting the shell apertures are not visible. Zoom in to view." + ], + "image_footnote": [], + "bbox": [ + 86, + 102, + 460, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[15, 20, 31], and here we propose a straight-forward approach with FLDMs. Given a pre-trained FLDM conditioned on a user-input binary mask $m: M \\to \\{0, 1\\}$ specifying the regions to preserve (Figure 5), generative inpainting can be performed by modifying the iterative sampling process. Specifically, after estimating previous the latent features $\\widetilde{Z}_{t-1}$ as in Equation (11), the features in the mask region are replaced with the appropriately noised input latents from the forward process $Z_{t-1}$ such that", + "bbox": [ + 75, + 503, + 468, + 640 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {Z} _ {t - 1} \\mapsto m \\cdot Z _ {t - 1} + (1 - m) \\cdot \\tilde {Z} _ {t - 1}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 647, + 468, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This approach encourages the inpainted texture to agree with the original at the mask boundaries due to convolutional structure of the denoising network, with this effect observed on the hind leg of sampled textures in Figure 5.", + "bbox": [ + 75, + 672, + 468, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Generative Texture Transfer", + "text_level": 1, + "bbox": [ + 76, + 741, + 333, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perhaps the most dramatic consequence of isometry-equivariance is that it naturally facilitates a notion of generative texture transfer. Since the FL-VAE and FLDMs commute with isometries $\\gamma : M \\to N$ as in Equations (6) and (12), our models see $TM^{d}$ and $TN^{d}$ as functionally the same space. In practice, we find that long as surfaces $M$ and $N$ are approximately isometric, we can encode a texture and train an FLDM on $M$ , then sample the model on $N$ and decode the synthesized latents to texture it.", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9370a8f36dc0dadec68557824cf9f71cfdebb2208ddbf2569c4dc7773ae9b9c2.jpg", + "image_caption": [ + "FLDM samples on increasingly coarse remeshings", + "Figure 7. Controlling the scale of synthesized textures. Here, the FLDM trained on the textured skull in the center-left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry, dilating the size of the tesserae. Zoom in to view details." + ], + "image_footnote": [], + "bbox": [ + 506, + 102, + 880, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Several examples are shown in Figure 6. An FLDM is pre-trained with label conditioning on each of the textured shells in the first column and sampled with corresponding labels on the others to texture them in the same style. Additional examples are shown on the right side of Figure 1. Notably, our model successfully transfers between different topologies, as the skull meshes are genus zero and three. More generally, we are able to transfer textures between highly dissimilar shapes (see supplement for examples).", + "bbox": [ + 496, + 248, + 890, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.6. Controlling Textural Scale", + "text_level": 1, + "bbox": [ + 500, + 396, + 740, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "An additional feature of our FLDMs is that they offer user-control over the scale of sampled textures. Internally, FLDMs normalize the filter support by dividing the distances between adjacent vertices by the radius of the mean vertex area element $\\widetilde{r} = \\sqrt{A / (\\pi|V|)}$ , where $A$ is the surface area of the mesh and $|V|$ is the number of vertices. Thus, the scale of the synthesized textural details can be controlled by increasing or decreasing the resolution of the mesh upon which the pre-trained FLDM is sampled. An example shown in Figure 7, where the FLDM trained on the textured skull on the center left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry. As the number of vertices decreases, the average one-ring occupies an increasingly larger fraction of the surface area, dilating the size of the synthesized textural details.", + "bbox": [ + 496, + 420, + 890, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 500, + 662, + 756, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present an original framework for latent diffusion models operating directly on the surfaces to generate high-quality textures. Our approach consists of two main contributions: a novel latent representation mapping textures to vector fields called Field Latents and Field Latent Diffusion Models which learn to denoise a diffusion process in the latent space on the surface. We apply our method in a single-textured-mesh paradigm, and generate variations of textures with state-of-the-art fidelity. Limitations of our approach include the relatively long training time of FLDMs and an inability to reflect directional information in synthesized textures. The latter could potentially be addressed by conditioning FLDMs on user-specified vector fields, with embeddings formed via dot products with network features.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7960", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Serge Assaad, Carlton Downey, Rami Al-Rfou, Nigamaa Nayakanti, and Ben Sapp. Vn-transformer: Rotation-equivariant attention for vector neurons. arXiv preprint arXiv:2206.04176, 2022. 4, 3", + "[2] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. arXiv preprint arXiv:2304.05868, 2023. 3", + "[3] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, pages 4169-4181, 2023. 2", + "[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2", + "[5] Qimin Chen, Zhiqin Chen, Hang Zhou, and Hao Zhang. Shaddr: Real-time example-based geometry and texture generation via 3d shape detailization and differentiable rendering. arXiv preprint arXiv:2306.04889, 2023. 3", + "[6] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2", + "[7] An-Chieh Cheng, Xueting Li, Sifei Liu, and Xiaolong Wang. Tuvf: Learning generalizable texture uv radiance fields. arXiv preprint arXiv:2305.03040, 2023. 2", + "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 2, 6, 7", + "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects, 2023. 2", + "[10] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In ICCV, pages 12200-12209, 2021. 2", + "[11] Nicolas Donati, Etienne Corman, Simone Melzi, and Maks Ovsjanikov. Complex functional maps: A conformal link between tangent bundles. In Computer Graphics Forum, pages 317-334. Wiley Online Library, 2022. 3, 7", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4", + "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items, 2022. 1, 5, 6, 7", + "[14] Ahmed A Elhag, Joshua M Susskind, and Miguel An" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gel Bautista. Manifold diffusion fields. arXiv preprint arXiv:2305.15586, 2023. 3", + "[15] Julien Fayer, Bastien Durix, Simone Gasparini, and Géraldine Morin. Texturing and inpainting a complete tubular 3d object reconstructed from partial views. Computers & Graphics, 74:126-136, 2018. 8", + "[16] Marc Finzi, Samuel Stanton, Pavel Izmailov, and Andrew Gordon Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In ICML, pages 3165-3176. PMLR, 2020. 4", + "[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 2", + "[18] James Gardner, Bernhard Egger, and William Smith. Rotation-equivariant conditional spherical neural fields for learning a natural illumination prior. NeurIPS, 35:26309-26323, 2022. 4", + "[19] Niv Granot, Ben Feinstein, Assaf Shocher, Shai Bagon, and Michal Irani. Drop the gan: In defense of patches nearest neighbors as single image generative models. In CVPR, pages 13460-13469, 2022. 3", + "[20] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, pages 12135-12144, 2019. 8", + "[21] Tobias Hinz, Matthew Fisher, Oliver Wang, and Stefan Wermter. Improved techniques for training single-image gans. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1300–1309, 2021. 3", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5, 3", + "[23] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 4", + "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 6, 5", + "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3, 4", + "[26] Felix Knöppel, Keenan Crane, Ulrich Pinkall, and Peter Schröder. Globally optimal direction fields. ACM TOG, 32 (4):1-10, 2013. 3", + "[27] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, pages 622-639. Springer, 2022. 3, 4, 5, 6, 7", + "[28] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930. PMLR, 2023. 2, 3, 5, 4", + "[29] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7961", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6, 3", + "[30] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2", + "[31] A Maggiordomo, P Cignoni, and M Tarini. Texture inpainting for photogrammetric models. In Computer Graphics Forum. Wiley Online Library, 2023. 8", + "[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 2", + "[33] Thomas W. Mitchel. Extending Convolution Through Spatially Adaptive Alignment. PhD thesis, Johns Hopkins University, 2022. 4", + "[34] Thomas W. Mitchel, Benedict Brown, David Koller, Tim Weyrich, Szymon Rusinkiewicz, and Michael Kazhdan. Efficient spatially adaptive convolution and correlation, 2020. 4", + "[35] Thomas W Mitchel, Vladimir G Kim, and Michael Kazhdan. Field convolutions for surface cnns. In ICCV, pages 10001-10011, 2021. 2, 3, 5, 4", + "[36] Thomas W Mitchel, Szymon Rusinkiewicz, Gregory S Chirikjian, and Michael Kazhdan. Echo: Extended convolution histogram of orientations for local surface description. In Computer Graphics Forum, pages 180-194. Wiley Online Library, 2021. 4", + "[37] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 4", + "[38] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2, 3", + "[39] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, pages 4531-4540, 2019. 2", + "[40] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In ICCV, pages 13879-13889, 2021. 2", + "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2", + "[43] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 3", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image syn" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2", + "[45] Klaus Hildebrandt Ruben Wiersma, Elmar Eisemann. Cnns on surfaces using rotation-equivariant features. ACM TOG, 39(4), 2020. 5, 4", + "[46] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019. 2, 3, 7", + "[47] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM TOG, XX(X), 2022. 3, 4", + "[48] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, pages 72-88. Springer, 2022. 3", + "[49] Saurabh Singh and Shankar Krishnan. Filter response normalization layer: Eliminating batch dependence in the training of deep neural networks. In CVPR, pages 11237-11246, 2020. 2", + "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[51] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In ICCV, pages 7062-7072, 2023. 2", + "[52] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model, 2023. 2", + "[53] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting, 2023. 2", + "[54] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 2, 3, 5, 4", + "[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[56] Jiacheng Wei, Hao Wang, Jiashi Feng, Guosheng Lin, and Kim-Hui Yap. Taps3d: Text-guided 3d textured shape generation from pseudo supervision. In CVPR, pages 16805-16815, 2023. 2", + "[57] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. Deltaconv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 5, 4", + "[58] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2, 3, 5, 6, 7", + "[59] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 4", + "[60] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7962", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In ICCV, pages 4206–4216, 2023. 2", + "[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7", + "[63] Xuanmeng Zhang, Jianfeng Zhang, Rohan Chacko, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, pages 2273-2282, 2023. 2", + "[64] Peiye Zhuang, Samira Abnar, Jiatao Gu, Alex Schwing, Joshua M Susskind, and Miguel Ángel Bautista. Diffusion probabilistic fields. In ICLR, 2022. 3" + ], + "bbox": [ + 78, + 90, + 467, + 289 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7963", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_model.json b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4ef8027ab2c60364e3b96f7e15a28655aa20d0cd --- /dev/null +++ b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_model.json @@ -0,0 +1,2873 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.851, + 0.152 + ], + "angle": 0, + "content": "Single Mesh Diffusion Models with Field Latents for Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.18, + 0.731, + 0.224 + ], + "angle": 0, + "content": "Thomas W. Mitchel1,2† Carlos Esteves1 Ameesh Makadia1 1Google Research 2PlayStation" + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.241, + 0.169, + 0.255 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.259, + 0.221, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.241, + 0.474, + 0.254 + ], + "angle": 0, + "content": "Generated textures" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.274, + 0.355, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.275, + 0.476, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.275, + 0.597, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.241, + 0.866, + 0.255 + ], + "angle": 0, + "content": "Generative transfer to new geometry" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.262, + 0.768, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.277, + 0.888, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.526, + 0.894, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Our latent diffusion models operate directly on the surfaces of 3D shapes, synthesizing new high-quality textures (center) after training on a single example (left). Both our novel latent representation and diffusion models are isometry-equivariant, facilitating a notion of generative texture transfer by sampling pre-trained models on new geometries (right). Above our models are conditioned on coarse semantic labels reflecting a subjective distribution of content, which delineate the sole and interior of the shoes, the eyes and mouths of the skulls, and the decals on the bottles." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.608, + 0.314, + 0.625 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.474, + 0.871 + ], + "angle": 0, + "content": "We introduce a framework for intrinsic latent diffusion models operating directly on the surfaces of 3D shapes, with the goal of synthesizing high-quality textures. Our approach is underpinned by two contributions: Field Latents, a latent representation encoding textures as discrete vector fields on the mesh vertices, and Field Latent Diffusion Models, which learn to denoise a diffusion process in the learned latent space on the surface. We consider a single-textured-mesh paradigm, where our models are trained to generate variations of a given texture on a mesh. We show the synthesized textures are of superior fidelity compared those from existing single-textured-mesh generative models. Our models can also be adapted for user-controlled editing tasks such as inpainting and label-guided generation. The efficacy of our approach is due in part to the equivariance" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.61, + 0.893, + 0.7 + ], + "angle": 0, + "content": "of our proposed framework under isometries, allowing our models to seamlessly reproduce details across locally similar regions and opening the door to a notion of generative texture transfer. Code and visualizations are available at https://single-mesh-diffusion.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.735, + 0.633, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.893, + 0.869 + ], + "angle": 0, + "content": "The emergence of latent diffusion models (LDMs) [44] as powerful tools for 2D content creation has motivated efforts to replicate their success across different modalities. A particularly attractive direction is the synthesis of textured 3D assets, due to the high cost of building photorealistic objects which can require intricately configured scanning solutions [13] or expertise in 3D modeling." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The majority of 3D synthesis methods employing LDMs operate in an image-based setting where the 3D repre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.31, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done while at Google Research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7953" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "sentation is optimized through differentiably rendered images, often in conjunction with score distillation sampling (SDS) [3, 4, 32, 41-43, 52, 53]. This design allows the models to leverage the full capabilities of pre-trained 2D LDMs. A second, alternative method is to rasterize both geometry and texture onto a 3D grid which are compressed into triplane latent features over which the DM operates [58]. However, in applications where texture synthesis takes precedence over 3D geometry, both approaches sacrifice fidelity in favor of a workable representation — models relying on planar renderings suffer from view inconsistencies when mapping back to the 3D shape while rasterization inherently aliases fine textural details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.29, + 0.47, + 0.367 + ], + "angle": 0, + "content": "Here, we observe that the discretization of a surface as a triangle mesh is itself a tractable and effective representational space. To this end, we present an original framework for intrinsic LDMs operating directly on the surfaces of shapes, with the goal of generating high-quality textures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.64 + ], + "angle": 0, + "content": "Our method consists of two distinct novel components, serving as the main contributions of our work. The first is a latent representation called Field Latents (FLs). Textures are mapped to tangent vector features at the vertices which are responsible for characterizing the local texture. The choice of tangent vector features over scalars allows for the capture of directional information related to the local texture which we show enables superior quality reconstructions. More generally, FLs offer an effective form of perceptual compression in which a high-resolution texture, analogous to a continuous signal defined over a surface, is mapped to a collection of discrete vector fields taking values at the vertices of a lower-resolution mesh. The second component is a Field Latent Diffusion Model (FLDM), which learns a denoising diffusion process in field latent space. Our FLDMs are built upon field convolutions [35], surface convolution operators designed specifically to process tangent vector features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Although we outline a general framework for constructing diffusion models for surfaces, inspired by the success of single-image generative models [28, 38, 46, 54], we tailor and deploy our FLDM architecture in the single-textured-mesh setting. We address the problem of generating variations of a given texture on a mesh (Figure 1, left and center), as well as tasks supporting user control such as label-guided generation and inpainting. The single-textured-mesh setting also lets us circumvent the data scarcity issue. Large-scale 3D datasets with high-quality textures have only recently become available [8, 9], with only a fraction of samples across disparate categories possessing complex non-uniform textures. Furthermore, many textured 3D assets that are handcrafted or derived from real-world scans are geometrically or stylistically unique, upon which large-scale models can be challenging to condition." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Experiments reveal our models enable flexible synthe" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "sis of high-resolution textures of qualitatively superior fidelity compared to existing single-textured-asset diffusion models, while simultaneously sidestepping the challenges of using 2D-to-3D approaches. Furthermore, both FLs and FLDMs are isometry-equivariant — each commute with distance-preserving shape deformations. As a result, we find that our approach can be used for generative texture transfer, wherein a FLDM trained on a single textured mesh can be sampled on a second similar mesh to texture it in the style of the first (Figure 1, right)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Existing generative models that synthesize textured 3D assets typically map content to an internal 2D representation which is computationally preferable to working directly over 3D Euclidean space. Perhaps the most popular approach involves iteratively rendering objects from different viewpoints, enabling the use of pre-trained 2D LDMs (e.g. Stable Diffusion [44]) either indirectly as optimization priors [30, 32, 41, 42] or to directly apply a full denoising process [3, 4, 43, 53]. The former case, pioneered by DreamFusion [41], uses score distillation sampling (SDS), in which a neural radiance field is optimized such that the rendered images appear to be reasonable samples from a pre-trained LDM; in the latter, LDMs with depth conditioning are used to iteratively denoise the rendered images which are aggregatively projected to the texture map. However, SDS-based methods are computationally feasible only with low-resolution LDMs and small NeRFs, and are unable to synthesize fine details. Additionally, approaches that directly denoise and project must contend with artifacts arising from view inconsistencies and synthesized textures may contain unnatural tones or lighting effects as residuals from their image-trained LDM backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Alternative representational spaces to drive textured 3D synthesis have also been explored. Texture Fields [17, 39] encode textures as 3D neural fields and triplane feature maps have emerged recently as an efficient domain for 3D geometry [56, 58, 63]. In particular, Sin3DM [58] encodes geometry and texture into an implicit triplane latent representation, then trains and samples a 2D DM on the resulting feature maps. However, geometry and texture must first be rasterized to a 3D grid before triplane convolutions can be employed, which can lead to aliasing of high-fidelity textures and geometric details due to memory constraints on the maximal grid resolution. To disentangle geometry from texture, several methods map textures to domains such as the image plane [6, 40, 51, 60] or sphere [7]. Point-UV diffusion [61] trains and samples 2D DMs directly in the UV-atlas, though this approach suffers on meshes with disconnected or fragmented UV mappings." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Similar to our approach, a third class of models encode textures as features at the simplices of a mesh. GAN-based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "methods Texturify [48] and Mesh2Tex [2] compress textures to attributes at the mesh facets, and differentiable rendering enables adversarial supervision in the image space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.362 + ], + "angle": 0, + "content": "Alternatively, Intrinsic Neural Fields (INFs) [27] introduces a latent representation wherein textures are encoded as eigenfunctions of the Laplacian taking values at the mesh vertices. However, features at points on faces are recovered through barycentric interpolation, which we will show leads to a breakdown in reconstruction fidelity at higher compression ratios. Recently, Manifold Diffusion Fields [14, 64] introduce the first fully intrinsic DMs defined over 2D Riemannian manifolds, and are based on a global attention mechanism which aggregates features at the vertices, limiting their applicability to coarse, low-resolution meshes due to complexity constraints. Here, our proposed FLDMs are also fully intrinsic but are built upon locally-supported convolutions defined over the manifold [35], allowing them to scale to higher-resolution meshes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Our single-textured-mesh FLDMs are inspired by recent work in which DMs are trained to capture internal patch distributions from a single image with the goal of generating diverse samples with similar visual content [28, 38, 54]; In turn, these models can trace their lineage to the influential SinGAN architecture [46] and its derivatives [19, 21]. Here, we mirror the approach proposed by SinDiffusion [54] and denoise with a shallow convolutional UNet to control the receptive field and prevent the DM from overfitting to the texture. Similar to our setting, Sin3DM [58] was the first to train LDMs in a single-textured-mesh paradigm, however they take an extrinsic approach to generating geometry and texture, whereas our approach focuses only on synthesizing textures which we show qualitatively are of higher-fidelity than those produced by Sin3DM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.253, + 0.62 + ], + "angle": 0, + "content": "3. Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.763 + ], + "angle": 0, + "content": "We introduce a framework for intrinsic LDMs operating directly on surfaces. The first component is a latent representation of mesh textures comprised of tangent vector features. At each vertex, the tangent features characterize the texture of a local neighborhood on the surface, and the collection of these features constitute a stack of vector fields we call Field Latents (FLs). The FL space is learned with a locally-supported variational autoencoder [25] (FL-VAE) which is described in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The second key component is the Field Latent Diffusion Model (FLDM), which learns to denoise a diffusion process in the tangent space of a surface. Denoising networks are constructed with field convolutions (FCs), surface convolution operators acting on tangent vector fields [35]. We extend FCs to interleave scalar embeddings with the tangent vector features, enabling the injection of diffusion timesteps and optional conditioning, such as user-specified labels, into the denoising model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "In practice, FL-VAEs and FLDMs are applied to synthesize new textures from a single textured triangle mesh. We pre-train a single FL-VAE by superimposing planar meshes over a large high-quality image dataset, obtaining a general-purpose latent space which can be used to train FLDMs on arbitrary textured meshes. To do so, a texture is first mapped to distributions in the tangent space at the vertices with the FL-VAE encoder, and an FLDM is trained to iteratively denoise tangent vector latent features. Afterwards, samples from the FLDM can be decoded with the FL-VAE decoder to synthesize new textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Importance of Isometry-Equivariance Our FL-VAEs and FLDMs are designed to be equivariant under isometries, i.e. they commute with distance-preserving shape deformations. Unlike images, points on a surface have no canonical orientation. The choice of local coordinates is ambiguous up to an arbitrary rotation, making it impossible to simply adapt standard image-based VAEs and diffusion models to surfaces. However, isometries manifest locally as rotations. Thus, designing our framework to be isometry-equivariant inherently solves the orientation ambiguity problem, enabling consistent, repeatable results. Isometry-equivariance also results in tangible benefits — our models are able to seamlessly reproduce textural details across locally similar areas of meshes. Furthermore, while existing models transfer textures with pointwise maps [11] or generatively via conditioning on a learned token [5, 43], we can simply sample our pre-trained FLDMs on new, similar meshes to texture them in the learned style." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.563, + 0.637, + 0.578 + ], + "angle": 0, + "content": "4. Field Latents" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.679 + ], + "angle": 0, + "content": "As in Knoppel et al. [26] we associate tangent vectors with complex numbers. For a surface \\(M\\) and point \\(p \\in M\\), we assign to the tangent space an arbitrary orthonormal basis such that for any \\(\\mathbf{v} \\in T_pM\\) we have \\(\\mathbf{v} \\equiv r e^{i\\theta}\\), with \\(r = |\\mathbf{v}|\\) and \\(\\theta\\) the direction of \\(\\mathbf{v}\\) expressed in the frame. We make this convention explicit by denoting \\(\\mathbb{C}_p \\equiv T_pM\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.799 + ], + "angle": 0, + "content": "To describe the FL-VAE, we require a notion of multidimensional Gaussian noise in the tangent bundle of a surface \\(M\\). Specifically, we denote the Gaussian distribution over \\(d\\) copies of the tangent bundle \\(TM^{d}\\) by \\(\\mathcal{T}\\mathcal{N}_M(0,I_d)\\). Samples \\(\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)\\) are vector fields in \\(TM^{d}\\) such that at each point \\(p\\in M\\), the coefficients of \\(\\epsilon (p)\\in \\mathbb{C}_p^d\\) expressed in the orthonormal basis are themselves samples from the \\(d\\)-dimensional standard normal distribution" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.812, + 0.892, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\epsilon (p) = \\epsilon_ {1} + i \\epsilon_ {2}, \\epsilon_ {1}, \\epsilon_ {2} \\sim \\mathcal {N} (0, I _ {d}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work we are concerned with isometries \\(\\gamma : M \\to N\\), which enjoy the property that their push-forwards \\(d\\gamma : TM \\to TN\\), which define a smooth map between tangent spaces, manifest pointwise as rotations. From Equa-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7955" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "tion (1) it is easy to see that \\(\\mathcal{T}\\mathcal{N}_M(0,I_d)\\) is symmetric under such transformations. Thus, it follows that sampling from \\(\\mathcal{T}\\mathcal{N}_N(0,I_d)\\) is equivalent to pushing forward samples from \\(\\mathcal{T}\\mathcal{N}_M(0,I_d)\\)," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.162, + 0.469, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {\\prime} = \\gamma \\epsilon \\quad \\text {f o r} \\quad \\begin{array}{l} \\epsilon^ {\\prime} \\sim \\mathcal {T N} _ {N} (0, I _ {d}) \\\\ \\epsilon \\sim \\mathcal {T N} _ {M} (0, I _ {d}) \\end{array} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.21, + 0.47, + 0.24 + ], + "angle": 0, + "content": "with \\(\\gamma \\epsilon = [d\\gamma \\cdot \\epsilon] \\circ \\gamma^{-1}\\) denoting the standard action of diffeomorphisms on vector fields via left-shifts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.259, + 0.469, + 0.35 + ], + "angle": 0, + "content": "FL-VAE Encoder The encoder \\(\\mathcal{E}\\) is a network that, for any point \\(p\\in M\\) , takes \\(n\\) -dimensional scalar functions \\(\\psi \\in L^2 (M,\\mathbb{R}^n)\\) (e.g. texture RGB values with \\(n = 3\\) restricted to the surrounding geodesic neighborhood \\(B_{p}\\subset\\) \\(M\\) to the parameters of \\(d\\) independent normal distributions in the tangent space at \\(p\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.359, + 0.469, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {E} _ {p}: L ^ {2} \\left(B _ {p}, \\mathbb {R} ^ {n}\\right)\\rightarrow \\mathbb {C} _ {p} ^ {d} \\times \\mathbb {R} _ {\\geq 0} ^ {d} \\tag {3} \\\\ \\psi \\mapsto \\left(\\mu_ {p} ^ {\\psi}, \\sigma_ {p} ^ {\\psi}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.411, + 0.469, + 0.523 + ], + "angle": 0, + "content": "Here, the means of the distributions \\(\\mu_p^\\psi\\) are themselves tangent vectors while the standard deviations \\(\\sigma_p^\\psi\\) are scalars. Latent codes characterizing the local restriction of \\(\\psi\\) are collections of tangent vectors \\(z_{p}^{\\psi}\\in \\mathbb{C}_{p}^{d}\\) drawn from the multivariate normal distribution parameterized by \\((\\mu_p^\\psi ,\\sigma_p^\\psi)\\). Using the reparameterization trick [25] latent codes can be generated as \\(z_{p}^{\\psi} = \\mu_{p}^{\\psi} + \\sigma_{p}^{\\psi}\\odot \\epsilon (p)\\) with \\(\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.469, + 0.66 + ], + "angle": 0, + "content": "FL-VAE Decoder The principal advantage in a tangent vector latent representation is that it naturally admits a descriptive coordinate function which can be queried to associate features with neighboring points. The decoder \\(\\mathcal{D}\\) is designed as a neural field operating over the local parameterization of the surface induced by the logarithm map about a point \\(p\\), \\(\\log_p:B_p\\to \\mathbb{C}_p\\), with \\(\\log_p q\\) encoding the position of \\(q\\in B_{p}\\) in the tangent space of \\(p\\). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.67, + 0.469, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} _ {p}: \\log_ {p} \\left(B _ {p}\\right) \\times \\mathbb {C} _ {p} ^ {d} \\rightarrow \\mathbb {R} ^ {n} \\tag {4} \\\\ \\left(\\log_ {p} q, z _ {p} ^ {\\psi}\\right) \\mapsto \\widehat {\\psi} _ {p} (q), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.724, + 0.469, + 0.867 + ], + "angle": 0, + "content": "with \\(\\widehat{\\psi}_p\\) denoting the prediction of \\(\\psi\\) made from the perspective of \\(p\\). Decoding follows a similar process as proposed in [18]. Specifically, given a tuple \\((\\log_p q, z_p^\\psi)\\), the decoder constructs two features. First, an invariant scalar feature is obtained by vectorizing the upper-triangular part of the Hermitian outer product of the latent code with itself, \\(\\mathrm{vec}_{j \\geq i} \\left(z_p^\\psi \\left[z_p^\\psi\\right]^*\\right) \\in \\mathbb{C}^{\\frac{d(d+1)}{2}}\\). Second, a positionally-aware scalar feature \\(c_{pq}^\\psi \\in \\mathbb{C}^d\\) is formed by the natural coordinate function" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.877, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\nc _ {p q} ^ {\\psi} \\equiv \\log_ {p} q \\cdot \\overline {{z _ {p} ^ {\\psi}}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.157 + ], + "angle": 0, + "content": "which corresponds to storing the inner product and determinant of each tangent vector with the position of \\(q\\) relative to \\(p\\) given by \\(\\log_p q\\). The neural field receives the concatenation of these two features and returns the prediction \\(\\widehat{\\psi}_p(q)\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.173, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Equivalence Recall we are concerned with constructing a latent representation equivariant under isometries — that for any \\(\\psi \\in L^2(M, \\mathbb{R}^n)\\) and isometry \\(\\gamma: M \\to N\\) we have" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.228, + 0.891, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\left[ \\widehat {\\gamma \\psi} \\right] _ {\\gamma (p)} = \\gamma \\widehat {\\psi} _ {p}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.338 + ], + "angle": 0, + "content": "where \\(\\gamma \\psi = \\psi \\circ \\gamma^{-1}\\) denotes the standard action of diffeomorphisms by left-shifts. Thus, it follows from Equation (2) that a sufficient condition for Equation (6) to hold is if for all \\(\\psi \\in L^2(M, \\mathbb{R}^n)\\), isometries \\(\\gamma : M \\to N\\), and \\(p \\in M\\), the encoded mean \\(\\mu_p^\\psi\\) and standard deviation \\(\\sigma_p^\\psi\\) satisfy" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.347, + 0.891, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\left. d \\gamma \\right| _ {p} \\cdot \\mu_ {p} ^ {\\psi} = \\mu_ {\\gamma (p)} ^ {\\gamma \\psi} \\quad \\text {a n d} \\quad \\sigma_ {p} ^ {\\psi} = \\sigma_ {\\gamma (p)} ^ {\\gamma \\psi}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.423 + ], + "angle": 0, + "content": "with \\(d\\gamma|_p\\) denoting the push-forward of \\(\\gamma\\) in the tangent space at \\(p\\), which rotates the local coordinate system. A detailed proof appears in the supplement, sec. 8.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.68 + ], + "angle": 0, + "content": "Architecture In practice, we discretize a surface \\(M\\) by a triangle mesh with vertices \\(V\\) and faces \\(F\\). Textures are scalar functions \\(\\psi\\) defined continuously over the faces and vertices, taking values in \\(\\mathbb{R}^3\\). For each \\(p \\in V\\), we consider its neighborhood to be the surrounding one-ring, the collection of faces in \\(F\\) that share \\(p\\) as a vertex. The encoder architecture is chosen so as to satisfy the conditions in Equation (7). A fixed number of points \\(\\{q_i\\}\\) are uniformly sampled in the one-ring neighborhood at which the texture function is evaluated. The resulting scalar features \\(\\{\\psi(q_i)\\}\\) are interleaved with the logarithms \\(\\{\\log_p q_i\\}\\) to construct tangent vector features which are concatenated with a token following [12]. The features are passed to eight successive VN-Transformer layers [1], modified to process tangent vector features. Afterwards, the token is extracted and used to predict the mean and standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.833 + ], + "angle": 0, + "content": "The neural field comprising the decoder consists of a five-layer real-valued MLP, to which the real and imaginary parts of the complex feature vector are passed. For any point in a triangle, the latent codes at the adjacent vertices provide three distinct predictions which are linearly blended with barycentric interpolation only at inference. This is in contrast to INF [27], wherein linearly blended features are passed to the neural field to make predictions. A comprehensive discussion of the FL-VAE architecture and training objective is provided in section 9.1 in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.776, + 0.861 + ], + "angle": 0, + "content": "5. Field Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "FLDMs define a noising process on surface vector fields, and learn a denoising diffusion probabilistic model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7956" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.114, + 0.09, + 0.16, + 0.1 + ], + "angle": 0, + "content": "Method" + }, + { + "type": "header", + "bbox": [ + 0.227, + 0.09, + 0.271, + 0.1 + ], + "angle": 0, + "content": "PSNR↑" + }, + { + "type": "header", + "bbox": [ + 0.301, + 0.09, + 0.352, + 0.1 + ], + "angle": 0, + "content": "DSSIM" + }, + { + "type": "header", + "bbox": [ + 0.374, + 0.09, + 0.42, + 0.1 + ], + "angle": 0, + "content": "LPIPS" + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.102, + 0.433, + 0.141 + ], + "angle": 0, + "content": "
FL-VAE22.38 (20.59)0.51 (0.83)1.02 (1.81)
FL-VAE (Bary.)21.33 (19.61)0.66 (1.01)1.31 (2.03)
INFs [27]18.86 (16.45)1.16 (1.64)2.15 (2.73)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.143, + 0.47, + 0.213 + ], + "angle": 0, + "content": "Table 1. Texture reconstruction quality on meshes from the Google Scanned Objects dataset [13]. Reconstructions are evaluated on high-res (30K vertices) and low-res (5K vertices, in parenthesis) meshings of the objects. Metrics are computed in the texture atlases, with DSSIM and LPIPS scaled by a factor of 10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.469, + 0.263 + ], + "angle": 0, + "content": "(DDPM) [22] using an equivariant surface network. In principle they can be applied to any vector data on surfaces, and here we use them to learn feature distributions in FL space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.47, + 0.399 + ], + "angle": 0, + "content": "Latent vector fields \\( Z \\in TM^{d} \\) can be produced by compressing a texture \\( \\psi \\in L^{2}(M, \\mathbb{R}^{3}) \\) to a collection of tangent vector features at each point with the FL-VAE encoder as in Equation (3) and sampling from the distributions. The forward process progressively adds noise in the tangent space, dictated by a monotonically decreasing schedule \\( \\{\\alpha_{t}\\}_{0 \\leq t \\leq T} \\) with \\( \\alpha_{0} = 1 \\) and \\( \\alpha_{T} = 0 \\). The noised vector fields can be expressed at an arbitrary time \\( t \\) in the discretized forward process with" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.406, + 0.469, + 0.425 + ], + "angle": 0, + "content": "\\[\nZ _ {t} = \\sqrt {\\alpha_ {t}} Z + \\sqrt {1 - \\alpha_ {t}} \\epsilon \\in T M ^ {d}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.432, + 0.369, + 0.447 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)\\) as in Equation (1)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.448, + 0.469, + 0.523 + ], + "angle": 0, + "content": "The denoising function \\(\\varepsilon\\) is a surface network trained to reverse the forward process. Specifically, it is a learnable map on the space of latent vector fields, conditioned on timesteps \\(t\\) and optional scalar embeddings \\(\\rho \\in L^{2}(M,\\mathbb{R}^{m})\\), with the goal of predicting the added noise," + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.53, + 0.469, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\varepsilon : T M ^ {d} \\times \\mathbb {R} _ {\\geq 0} \\times L ^ {2} (M, \\mathbb {R} ^ {m}) \\rightarrow T M ^ {d}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.557, + 0.249, + 0.571 + ], + "angle": 0, + "content": "trained subject to the loss" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.578, + 0.469, + 0.598 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {F L D M}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {T N} _ {M} (0, I _ {d}), 0 \\leq t \\leq T} \\| \\epsilon - \\varepsilon \\left(Z _ {t}, t, \\rho\\right) \\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.605, + 0.469, + 0.695 + ], + "angle": 0, + "content": "At inference, new latent vector fields are generated following the iterative DDPM denoising process. Beginning with representing the fully-noised vector fields at step \\(T\\) by a sample \\(\\tilde{Z} \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)\\), \\(\\tilde{Z}_T = \\tilde{Z}\\), the trained denoising function \\(\\varepsilon\\) is used to estimate the previous step in the forward process via the relation" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.703, + 0.468, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widetilde {Z} _ {t - 1} = C _ {1} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\widetilde {Z} _ {t} + C _ {2} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\varepsilon \\left(\\widetilde {Z} _ {t}, t, \\rho\\right) \\tag {11} \\\\ + C _ {3} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\epsilon , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.748, + 0.469, + 0.78 + ], + "angle": 0, + "content": "with \\(\\epsilon = 0\\) for \\(t = 1\\) and \\(\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)\\) otherwise, and \\(C_i(\\alpha_t,\\alpha_{t - 1})\\in \\mathbb{R}\\) as defined in 9.1 of the supplement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.469, + 0.872 + ], + "angle": 0, + "content": "Equivalence From the equivalence of the distributions \\(\\mathcal{T}\\mathcal{N}_M(0,I_d)\\) and \\(\\mathcal{T}\\mathcal{N}_N(0,I_d)\\) under isometries as in Equation (2), it follows that for any isometry \\(\\gamma :M\\to N\\) a sufficient condition for both the forward and reverse processes in Equations (8) and (11) to satisfy" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.878, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\gamma Z _ {t} = \\left[ \\gamma Z \\right] _ {t} \\quad \\text {a n d} \\quad \\gamma \\widetilde {Z} _ {t} = \\left[ \\gamma \\widetilde {Z} \\right] _ {t}, \\tag {12}\n\\]" + }, + { + "type": "header", + "bbox": [ + 0.505, + 0.093, + 0.586, + 0.105 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "header", + "bbox": [ + 0.616, + 0.093, + 0.672, + 0.105 + ], + "angle": 0, + "content": "FL-VAE" + }, + { + "type": "header", + "bbox": [ + 0.695, + 0.093, + 0.791, + 0.106 + ], + "angle": 0, + "content": "FL-VAE (Bary.)" + }, + { + "type": "header", + "bbox": [ + 0.813, + 0.093, + 0.872, + 0.106 + ], + "angle": 0, + "content": "INFs [27]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.111, + 0.597, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.158, + 0.598, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.112, + 0.692, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.158, + 0.695, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.112, + 0.788, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.112, + 0.886, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.158, + 0.792, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.159, + 0.888, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.234, + 0.594, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.272, + 0.592, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.234, + 0.693, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.273, + 0.692, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.234, + 0.793, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.273, + 0.791, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.234, + 0.892, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.273, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Figure 2. Visual comparison of textures compressed and reconstructed with the FL-VAE and INF [27] on 30K vertex meshes. Compared to barycentric coordinates, our proposed logarithmic coordinate function more richly extends latent features across the mesh, enabling the reconstruction of finer details. Zoom in to view." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.892, + 0.455 + ], + "angle": 0, + "content": "for \\(0 \\leq t \\leq T\\) is if the denoising network \\(\\varepsilon\\) in Equation (9) has the property that for any vector field \\(Z \\in TM^{d}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.468, + 0.891, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\gamma \\varepsilon (Z, t, \\rho) = \\varepsilon (\\gamma Z, t, \\gamma \\rho). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.498, + 0.883, + 0.514 + ], + "angle": 0, + "content": "A detailed proof is provided in sec. 8.2 of the supplement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Denoising with Field Convolutions To satisfy the condition in Equation (12), our denoising network is constructed with field convolutions (FC) [35], which belong to a class of models that can operate on tangent vector fields [45, 57]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.825 + ], + "angle": 0, + "content": "FCs convolve vector fields \\( Z \\in TM^{C} \\) with \\( C' \\times C \\) filter banks \\( \\mathbf{f}_{c'c} \\in L^{2}(\\mathbb{C},\\mathbb{C}) \\), returning vector fields \\( Z * \\mathbf{f} \\in TM^{C'} \\). In conventional diffusion models, features inside the denoising network are conditioned on timesteps \\( t \\) via an additive embedding applied inside the convolution blocks. Unfortunately, no direct analogue exists for vector field features on surfaces — additive embeddings break equivariance and we find that multiplicative embeddings destabilize training. Thus, to enable conditioning of our denoising network \\( \\varepsilon \\) on both timesteps \\( t \\) and optional user-input features \\( \\rho \\in L^{2}(M,\\mathbb{R}^{m}) \\), we inject embeddings directly into convolutions by extending FCs to convolve vector fields with filters \\( \\mathbf{f}_{c'c} \\in L^{2}(\\mathbb{C} \\times \\mathbb{R}^{e},\\mathbb{C}) \\), with \\( e \\) the embedding dimension. Critically, this formulation preserves equivariance without sacrificing stability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Following [54], the denoising network \\(\\varepsilon\\) takes the form of a shallow two-level U-Net but with field convolutions; the relatively small size of the network's receptive field allows distributions of latent features to be learned without overfitting to the single textural example [28, 54, 58]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7957" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.08, + 0.158, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.172, + 0.159, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.081, + 0.262, + 0.094 + ], + "angle": 0, + "content": "FLDM" + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.099, + 0.232, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.102, + 0.308, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.171, + 0.236, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.172, + 0.311, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.082, + 0.436, + 0.094 + ], + "angle": 0, + "content": "Sin3DM [58]" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.101, + 0.387, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.102, + 0.462, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.172, + 0.394, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.173, + 0.465, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.082, + 0.528, + 0.094 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.102, + 0.543, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.082, + 0.66, + 0.093 + ], + "angle": 0, + "content": "FLDM" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.102, + 0.628, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.102, + 0.712, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.082, + 0.846, + 0.094 + ], + "angle": 0, + "content": "Sin3DM [58]" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.102, + 0.796, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.102, + 0.882, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.173, + 0.721, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.174, + 0.805, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.174, + 0.89, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.893, + 0.308 + ], + "angle": 0, + "content": "Figure 3. Visual comparison of unconditionally (label-free) generated textures from FLDMs and Sin3DM [58]. Our FLDM's isometry-equivariant construction allows for replication of textural details across locally similar regions. In contrast, Sin3DM's extrinsic approach associates textural features with 3D space; Synthesized details appear as repetitions or extrusions of previous patterns along the major axes of the mesh, and we observe that novel textures cannot be created without modifying geometry. Zoom in to compare." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.331, + 0.295, + 0.395 + ], + "angle": 0, + "content": "
MethodSIFID ↓LPIPS ↑
FLDM3.271.15
Sample (left)0.710.94
Sin3DM [58]6.582.20
Sample (right)2.422.91
" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.322, + 0.376, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.322, + 0.452, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.53 + ], + "angle": 0, + "content": "Table 2. Fidelity and diversity of unconditionally generated textures trained on selected Objaverse [8] and Scanned Objects [13] meshes. SIFID and LPIPS are scaled by factors of \\(10^{5}\\) and 10, respectively. Unlike our FLDMs, Sin3DM synthesizes new geometry which can inflate diversity scores. As an example, the individual metrics are shown for the FLDM and Sin3DM samples on the right. The latter attains a high LPIPS score, but exhibits poor textural diversity and quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.21, + 0.557 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Pre-training the FL-VAE It is important to note that the FL-VAE proposed in Section 4 only interfaces with the mesh geometry through its local flattening in the image of the logarithm map. Thus, the FL-VAE operates independently of the local curvature, and a model pre-trained' on flat textured meshes can be used to encode and decode textures on arbitrary 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In all of our experiments we use the same FL-VAE with a \\(d = 8\\) dimensional latent space, pre-trained on the OpenImagesV4 dataset [29]. Specifically, we generate 1K different planar meshes containing 10K vertices each. Each sample \\(512 \\times 512\\) image is overlaid by a randomly chosen triangulation, with the image serving as the \"texture\" defined over the mesh. The FL-VAE is then trained with the sum of reconstruction loss and KL divergence using the Adam [24] optimizer. We deliberately train the FL-VAE on coarser triangulations than we expect during inference on textured meshes. This forces the model to reconstruct high frequencies during training, which we find increases robustness to sampling density and triangulation quality, in addition to improving the fidelity of reconstructed FLDM samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Training single-textured-mesh FLDMs All diffusion experiments share the same training regime. During preprocessing, we create 500 copies of a given mesh each with 30K vertices and different triangulations to prevent the denoising network from learning the connectivity. Each copy shares the same UV-atlas as the original and the texture is mapped to vector field latent features at the vertices using the pre-trained FL-VAE encoder. FLDMs have a max timestep of \\( T = 1000 \\) and noise is added on a linear schedule. The denoising network is trained subject to the loss in Equation (10) using the Adam [24] optimizer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.855, + 0.529 + ], + "angle": 0, + "content": "6.1. Texture Compression and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.779 + ], + "angle": 0, + "content": "First we quantitatively evaluate the capacity of our proposed FL-VAE for compressing and reconstructing textures. We compare FL-VAEs against two other approaches: the state-of-the-art INF [27] and a modified version of FL-VAEs serving as an ablation. INFs train a neural field per texture and predicts the texture values using the values of the Laplacian eigenfunctions extended over triangles through barycentric interpolation. Similarly, the modified FL-VAE (denoted FL-VAE Barycentric) omits our proposed logarithmic coordinate function in Equation (5) in favor of linearly interpolating the values of the invariant outer product features across triangles. It is otherwise identical to our proposed approach, including pre-training methodology. Thus, this also serves to compare the efficacy of our proposed coordinate function against barycentric interpolation in extending vertex features continuously over the mesh." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate, we select 16 objects exhibiting a range of complex textures with high-frequency details from the Google Scanned Objects dataset [13]. Two versions of the dataset are used, with objects remeshed to have \\(30\\mathrm{K}\\) (high-res) and \\(5\\mathrm{K}\\) (low-res) vertices, respectively. The single pretrained FL-VAE is used to map the textures to distributions over the tangent space, and the mean is used to sample the reconstructed texture at the pixel indices in the texture atlas." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7958" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.079, + 0.47, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.186, + 0.471, + 0.228 + ], + "angle": 0, + "content": "Figure 4. Label-guided generation. FLDMs can be conditioned on a subjective user-input labeling which generated textures will reflect. See Figure 1 for more examples. Zoom in to view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.469, + 0.265 + ], + "angle": 0, + "content": "A separate INF neural field is trained for each mesh, using the specified regime for high-res reconstruction [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Following INFs, we evaluate the quality of the reconstructed textures using the average PSNR, DSSIM [55], and LPIPS [62] metrics computed in texture atlases. The results are reported in Table 1, with several examples of reconstructions shown in Figure 2. Our proposed FL-VAE achieves the best results across all metrics at both resolutions, followed by the barycentric FL-VAE and INFs. The significant gap in performance between the barycentric FL-VAE and INFs is likely due to the reliance of the latter the on the Laplacian eigenfunctions as input features. Areas of fine textural detail may not necessarily overlap with regions where the eigenfunctions exhibit high variability, potentially contributing to the \"raggedness\" observed in INF reconstructions. Furthermore, the superior performance of our proposed FL-VAE relative to its barycentric counterpart suggests that our logarithmic coordinate function in Equation (5) provides a richer interpolant over the mesh faces. This directly increases the representational capacity of the latent features, enabling the reconstruction of finer textural details with the same number of features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.415, + 0.591 + ], + "angle": 0, + "content": "6.2. Unconditional (Label-Free) Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Label-free generation can be performed by training and sampling FLDMs conditioned only on the diffusion timescale (i.e. \\(\\rho = 0\\)). We compare against Sin3DM [58] which to our knowledge is the only prior generative model designed to operate in a single-textured-mesh paradigm. We follow their proposed evaluation protocol, and train FLDMs and Sin3DM on 10 assets from the Objaverse [8] and Scanned Objects [13] datasets selected for texture complexity and diversity. For each asset, 64 textures are sampled from the trained models, which are rendered on the mesh from 24 different views. Fidelity is measured with the mean SIFID (Single Image Fréchet Inception Distance) [46] pairwise between the renderings and those of the input mesh from the same view; the mean LPIPS [62] between renders of samples from the same viewpoint measures diversity. We note that unlike FLDMs, which only synthesize new textures, Sin3DM synthesizes both textures and geometries, with the latter potentially inflating LPIPS scores." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The results are shown in Table 2, Figure 3, and in the supplement (Figure 8). Compared to Sin3DM, our FLDMs" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.079, + 0.888, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.186, + 0.892, + 0.228 + ], + "angle": 0, + "content": "Figure 5. Inpainting with FLDMs. The input texture is preserved in the masked regions and new content is synthesized elsewhere with agreement on the boundaries. Zoom in to view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.893, + 0.506 + ], + "angle": 0, + "content": "are able to synthesize higher-fidelity textures which is likely due in part to the former's rasterization of textured geometry to a 3D grid before encoding to the latent space, potentially aliasing fine details. More generally, we observe that by operating over an extrinsic representation of textured geometry, Sin3DM intertwines texture with a mesh's 3D embedding. Thus, new textural details appear as repetitions or extrusions of earlier patterns along the major axes of the models, as seen in the vase, snake, and sculpture samples in Figure 3. Furthermore, we observe that texture and position are so strongly linked that Sin3DM is unable to produce significant textural alterations without correspondingly modifying the geometry, as seen in the vase, snake, and lantern samples. Here, the advantage of our isometry-equivariant formulation becomes clear as our FLDMs can effectively replicate textural details across areas of the mesh that are locally similar, regardless of the relative configuration of these regions in the 3D embedding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.514, + 0.734, + 0.529 + ], + "angle": 0, + "content": "6.3. Label-Guided Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Textured objects often exhibit distinct content specific to certain regions of the mesh which may not be geometrically unique; for example, in Figure 1 regions on the sole, body, and inside of the shoe lack distinguishing geometric features, as is the case with the eyes and mouth of the skull and the decal on the bottle. In such cases, a user may find it desirable for synthesized textures to reflect a subjective distribution of content on the input mesh, which we facilitate by extending our denoising model to incorporate optional conditioning from user-designated mesh features \\(\\rho \\in L^{2}(M,\\mathbb{R}^{m})\\). In the simplest case, the user can paint a coarse, semantic labeling to subjectively delineate salient regions, such as the flippers, head, and rock base of the seal sculpture in Figure 4. After training an FLDM conditioned on the labels, generated textures reflect the specified segmentation. Further examples are shown in Figure 1. Labeled areas contain different textural features, and conditioning ensures the FLDM samples respect these distributions, e.g. that teeth are synthesized only on the mouth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.618, + 0.849 + ], + "angle": 0, + "content": "6.4. Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In certain cases a user may wish to preserve the input texture in certain regions while synthesizing new content elsewhere. Texture inpainting is a well studied topic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7959" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.088, + 0.427, + 0.102 + ], + "angle": 0, + "content": "FLDM samples across geometries" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.103, + 0.462, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.499 + ], + "angle": 0, + "content": "Figure 6. Generative texture transfer. Since our FL-VAE and FLDMs commute with isometries, we can encode a texture and train an FLDM on an input mesh (left), then sample the pre-trained FLDM on a new, similar mesh and decode the latents to texture it (center to right). Conditioning labels segmenting the shell apertures are not visible. Zoom in to view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.641 + ], + "angle": 0, + "content": "[15, 20, 31], and here we propose a straight-forward approach with FLDMs. Given a pre-trained FLDM conditioned on a user-input binary mask \\( m: M \\to \\{0, 1\\} \\) specifying the regions to preserve (Figure 5), generative inpainting can be performed by modifying the iterative sampling process. Specifically, after estimating previous the latent features \\( \\widetilde{Z}_{t-1} \\) as in Equation (11), the features in the mask region are replaced with the appropriately noised input latents from the forward process \\( Z_{t-1} \\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.648, + 0.469, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\tilde {Z} _ {t - 1} \\mapsto m \\cdot Z _ {t - 1} + (1 - m) \\cdot \\tilde {Z} _ {t - 1}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.735 + ], + "angle": 0, + "content": "This approach encourages the inpainted texture to agree with the original at the mask boundaries due to convolutional structure of the denoising network, with this effect observed on the hind leg of sampled textures in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.334, + 0.758 + ], + "angle": 0, + "content": "6.5. Generative Texture Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Perhaps the most dramatic consequence of isometry-equivariance is that it naturally facilitates a notion of generative texture transfer. Since the FL-VAE and FLDMs commute with isometries \\(\\gamma : M \\to N\\) as in Equations (6) and (12), our models see \\(TM^{d}\\) and \\(TN^{d}\\) as functionally the same space. In practice, we find that long as surfaces \\(M\\) and \\(N\\) are approximately isometric, we can encode a texture and train an FLDM on \\(M\\), then sample the model on \\(N\\) and decode the synthesized latents to texture it." + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.085, + 0.843, + 0.099 + ], + "angle": 0, + "content": "FLDM samples on increasingly coarse remeshings" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.103, + 0.882, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.239 + ], + "angle": 0, + "content": "Figure 7. Controlling the scale of synthesized textures. Here, the FLDM trained on the textured skull in the center-left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry, dilating the size of the tesserae. Zoom in to view details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Several examples are shown in Figure 6. An FLDM is pre-trained with label conditioning on each of the textured shells in the first column and sampled with corresponding labels on the others to texture them in the same style. Additional examples are shown on the right side of Figure 1. Notably, our model successfully transfers between different topologies, as the skull meshes are genus zero and three. More generally, we are able to transfer textures between highly dissimilar shapes (see supplement for examples)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.397, + 0.741, + 0.413 + ], + "angle": 0, + "content": "6.6. Controlling Textural Scale" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.648 + ], + "angle": 0, + "content": "An additional feature of our FLDMs is that they offer user-control over the scale of sampled textures. Internally, FLDMs normalize the filter support by dividing the distances between adjacent vertices by the radius of the mean vertex area element \\(\\widetilde{r} = \\sqrt{A / (\\pi|V|)}\\), where \\(A\\) is the surface area of the mesh and \\(|V|\\) is the number of vertices. Thus, the scale of the synthesized textural details can be controlled by increasing or decreasing the resolution of the mesh upon which the pre-trained FLDM is sampled. An example shown in Figure 7, where the FLDM trained on the textured skull on the center left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry. As the number of vertices decreases, the average one-ring occupies an increasingly larger fraction of the surface area, dilating the size of the synthesized textural details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.757, + 0.68 + ], + "angle": 0, + "content": "7. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present an original framework for latent diffusion models operating directly on the surfaces to generate high-quality textures. Our approach consists of two main contributions: a novel latent representation mapping textures to vector fields called Field Latents and Field Latent Diffusion Models which learn to denoise a diffusion process in the latent space on the surface. We apply our method in a single-textured-mesh paradigm, and generate variations of textures with state-of-the-art fidelity. Limitations of our approach include the relatively long training time of FLDMs and an inability to reflect directional information in synthesized textures. The latter could potentially be addressed by conditioning FLDMs on user-specified vector fields, with embeddings formed via dot products with network features." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7960" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Serge Assaad, Carlton Downey, Rami Al-Rfou, Nigamaa Nayakanti, and Ben Sapp. Vn-transformer: Rotation-equivariant attention for vector neurons. arXiv preprint arXiv:2206.04176, 2022. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[2] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. arXiv preprint arXiv:2304.05868, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.468, + 0.27 + ], + "angle": 0, + "content": "[3] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, pages 4169-4181, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.468, + 0.327 + ], + "angle": 0, + "content": "[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.468, + 0.385 + ], + "angle": 0, + "content": "[5] Qimin Chen, Zhiqin Chen, Hang Zhou, and Hao Zhang. Shaddr: Real-time example-based geometry and texture generation via 3d shape detailization and differentiable rendering. arXiv preprint arXiv:2306.04889, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.468, + 0.429 + ], + "angle": 0, + "content": "[6] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.468, + 0.472 + ], + "angle": 0, + "content": "[7] An-Chieh Cheng, Xueting Li, Sifei Liu, and Xiaolong Wang. Tuvf: Learning generalizable texture uv radiance fields. arXiv preprint arXiv:2305.03040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.468, + 0.53 + ], + "angle": 0, + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.468, + 0.614 + ], + "angle": 0, + "content": "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[10] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In ICCV, pages 12200-12209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.468, + 0.729 + ], + "angle": 0, + "content": "[11] Nicolas Donati, Etienne Corman, Simone Melzi, and Maks Ovsjanikov. Complex functional maps: A conformal link between tangent bundles. In Computer Graphics Forum, pages 317-334. Wiley Online Library, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.468, + 0.813 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.882 + ], + "angle": 0, + "content": "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items, 2022. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Ahmed A Elhag, Joshua M Susskind, and Miguel An" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "gel Bautista. Manifold diffusion fields. arXiv preprint arXiv:2305.15586, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[15] Julien Fayer, Bastien Durix, Simone Gasparini, and Géraldine Morin. Texturing and inpainting a complete tubular 3d object reconstructed from partial views. Computers & Graphics, 74:126-136, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Marc Finzi, Samuel Stanton, Pavel Izmailov, and Andrew Gordon Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In ICML, pages 3165-3176. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[18] James Gardner, Bernhard Egger, and William Smith. Rotation-equivariant conditional spherical neural fields for learning a natural illumination prior. NeurIPS, 35:26309-26323, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[19] Niv Granot, Ben Feinstein, Assaf Shocher, Shai Bagon, and Michal Irani. Drop the gan: In defense of patches nearest neighbors as single image generative models. In CVPR, pages 13460-13469, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[20] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, pages 12135-12144, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[21] Tobias Hinz, Matthew Fisher, Oliver Wang, and Stefan Wermter. Improved techniques for training single-image gans. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1300–1309, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[23] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[26] Felix Knöppel, Keenan Crane, Ulrich Pinkall, and Peter Schröder. Globally optimal direction fields. ACM TOG, 32 (4):1-10, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[27] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, pages 622-639. Springer, 2022. 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[28] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930. PMLR, 2023. 2, 3, 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[30] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[31] A Maggiordomo, P Cignoni, and M Tarini. Texture inpainting for photogrammetric models. In Computer Graphics Forum. Wiley Online Library, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[33] Thomas W. Mitchel. Extending Convolution Through Spatially Adaptive Alignment. PhD thesis, Johns Hopkins University, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[34] Thomas W. Mitchel, Benedict Brown, David Koller, Tim Weyrich, Szymon Rusinkiewicz, and Michael Kazhdan. Efficient spatially adaptive convolution and correlation, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[35] Thomas W Mitchel, Vladimir G Kim, and Michael Kazhdan. Field convolutions for surface cnns. In ICCV, pages 10001-10011, 2021. 2, 3, 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[36] Thomas W Mitchel, Szymon Rusinkiewicz, Gregory S Chirikjian, and Michael Kazhdan. Echo: Extended convolution histogram of orientations for local surface description. In Computer Graphics Forum, pages 180-194. Wiley Online Library, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[37] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[38] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[39] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, pages 4531-4540, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[40] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In ICCV, pages 13879-13889, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[43] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image syn" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "thesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[45] Klaus Hildebrandt Ruben Wiersma, Elmar Eisemann. Cnns on surfaces using rotation-equivariant features. ACM TOG, 39(4), 2020. 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[46] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[47] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM TOG, XX(X), 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[48] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, pages 72-88. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[49] Saurabh Singh and Shankar Krishnan. Filter response normalization layer: Eliminating batch dependence in the training of deep neural networks. In CVPR, pages 11237-11246, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[51] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In ICCV, pages 7062-7072, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[52] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[53] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[54] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 2, 3, 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[56] Jiacheng Wei, Hao Wang, Jiashi Feng, Guosheng Lin, and Kim-Hui Yap. Taps3d: Text-guided 3d textured shape generation from pseudo supervision. In CVPR, pages 16805-16815, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[57] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. Deltaconv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[58] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[59] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[60] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7962" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[61] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In ICCV, pages 4206–4216, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[63] Xuanmeng Zhang, Jianfeng Zhang, Rohan Chacko, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, pages 2273-2282, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[64] Peiye Zhuang, Samira Abnar, Jiatao Gu, Alex Schwing, Joshua M Susskind, and Miguel Ángel Bautista. Diffusion probabilistic fields. In ICLR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7963" + } + ] +] \ No newline at end of file diff --git a/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_origin.pdf b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed809c9728c24b235f4ceec78090db2515723815 --- /dev/null +++ b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/bf087a44-a85c-4958-a01e-c35dc0ca2939_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faeb2bdd96bc745711eaad63331388222e9c1d39d1a5aef8f5ee467f28a3df76 +size 2106756 diff --git a/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/full.md b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a30e5a9fd0da45a9dd378d1125e38a424647d222 --- /dev/null +++ b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/full.md @@ -0,0 +1,412 @@ +# Single Mesh Diffusion Models with Field Latents for Texture Generation + +Thomas W. Mitchel1,2† Carlos Esteves1 Ameesh Makadia1 1Google Research 2PlayStation + +![](images/a00a360104b82d0ae3fe81b413533e9f05a385623c4ba7b5cd73c16127e41b35.jpg) +Input + +![](images/6c9f448fcd5320c1a5bdec708ab2a363a80c47d43271add1701150559da97947.jpg) +Figure 1. Our latent diffusion models operate directly on the surfaces of 3D shapes, synthesizing new high-quality textures (center) after training on a single example (left). Both our novel latent representation and diffusion models are isometry-equivariant, facilitating a notion of generative texture transfer by sampling pre-trained models on new geometries (right). Above our models are conditioned on coarse semantic labels reflecting a subjective distribution of content, which delineate the sole and interior of the shoes, the eyes and mouths of the skulls, and the decals on the bottles. + +![](images/cced20cab01424c70df95a6caa97ce3cdf1b93275382754214b82dbcf2c7c992.jpg) +Generated textures + +![](images/e4ebd282a16e9549f0749cfa4eb455fd69342bf55cc0b3f72aa1a2659af62431.jpg) + +![](images/d27906b94c9926129db751a83f30e561506b97bf8be3e715321946d534d20306.jpg) +Generative transfer to new geometry + +![](images/a90d4f2212b730fe890a0efbb1323d3bfd6e12568f0b5684a9689b2865fdddf8.jpg) + +# Abstract + +We introduce a framework for intrinsic latent diffusion models operating directly on the surfaces of 3D shapes, with the goal of synthesizing high-quality textures. Our approach is underpinned by two contributions: Field Latents, a latent representation encoding textures as discrete vector fields on the mesh vertices, and Field Latent Diffusion Models, which learn to denoise a diffusion process in the learned latent space on the surface. We consider a single-textured-mesh paradigm, where our models are trained to generate variations of a given texture on a mesh. We show the synthesized textures are of superior fidelity compared those from existing single-textured-mesh generative models. Our models can also be adapted for user-controlled editing tasks such as inpainting and label-guided generation. The efficacy of our approach is due in part to the equivariance + +of our proposed framework under isometries, allowing our models to seamlessly reproduce details across locally similar regions and opening the door to a notion of generative texture transfer. Code and visualizations are available at https://single-mesh-diffusion.github.io/. + +# 1. Introduction + +The emergence of latent diffusion models (LDMs) [44] as powerful tools for 2D content creation has motivated efforts to replicate their success across different modalities. A particularly attractive direction is the synthesis of textured 3D assets, due to the high cost of building photorealistic objects which can require intricately configured scanning solutions [13] or expertise in 3D modeling. + +The majority of 3D synthesis methods employing LDMs operate in an image-based setting where the 3D repre + +sentation is optimized through differentiably rendered images, often in conjunction with score distillation sampling (SDS) [3, 4, 32, 41-43, 52, 53]. This design allows the models to leverage the full capabilities of pre-trained 2D LDMs. A second, alternative method is to rasterize both geometry and texture onto a 3D grid which are compressed into triplane latent features over which the DM operates [58]. However, in applications where texture synthesis takes precedence over 3D geometry, both approaches sacrifice fidelity in favor of a workable representation — models relying on planar renderings suffer from view inconsistencies when mapping back to the 3D shape while rasterization inherently aliases fine textural details. + +Here, we observe that the discretization of a surface as a triangle mesh is itself a tractable and effective representational space. To this end, we present an original framework for intrinsic LDMs operating directly on the surfaces of shapes, with the goal of generating high-quality textures. + +Our method consists of two distinct novel components, serving as the main contributions of our work. The first is a latent representation called Field Latents (FLs). Textures are mapped to tangent vector features at the vertices which are responsible for characterizing the local texture. The choice of tangent vector features over scalars allows for the capture of directional information related to the local texture which we show enables superior quality reconstructions. More generally, FLs offer an effective form of perceptual compression in which a high-resolution texture, analogous to a continuous signal defined over a surface, is mapped to a collection of discrete vector fields taking values at the vertices of a lower-resolution mesh. The second component is a Field Latent Diffusion Model (FLDM), which learns a denoising diffusion process in field latent space. Our FLDMs are built upon field convolutions [35], surface convolution operators designed specifically to process tangent vector features. + +Although we outline a general framework for constructing diffusion models for surfaces, inspired by the success of single-image generative models [28, 38, 46, 54], we tailor and deploy our FLDM architecture in the single-textured-mesh setting. We address the problem of generating variations of a given texture on a mesh (Figure 1, left and center), as well as tasks supporting user control such as label-guided generation and inpainting. The single-textured-mesh setting also lets us circumvent the data scarcity issue. Large-scale 3D datasets with high-quality textures have only recently become available [8, 9], with only a fraction of samples across disparate categories possessing complex non-uniform textures. Furthermore, many textured 3D assets that are handcrafted or derived from real-world scans are geometrically or stylistically unique, upon which large-scale models can be challenging to condition. + +Experiments reveal our models enable flexible synthe + +sis of high-resolution textures of qualitatively superior fidelity compared to existing single-textured-asset diffusion models, while simultaneously sidestepping the challenges of using 2D-to-3D approaches. Furthermore, both FLs and FLDMs are isometry-equivariant — each commute with distance-preserving shape deformations. As a result, we find that our approach can be used for generative texture transfer, wherein a FLDM trained on a single textured mesh can be sampled on a second similar mesh to texture it in the style of the first (Figure 1, right). + +# 2. Related Work + +Existing generative models that synthesize textured 3D assets typically map content to an internal 2D representation which is computationally preferable to working directly over 3D Euclidean space. Perhaps the most popular approach involves iteratively rendering objects from different viewpoints, enabling the use of pre-trained 2D LDMs (e.g. Stable Diffusion [44]) either indirectly as optimization priors [30, 32, 41, 42] or to directly apply a full denoising process [3, 4, 43, 53]. The former case, pioneered by DreamFusion [41], uses score distillation sampling (SDS), in which a neural radiance field is optimized such that the rendered images appear to be reasonable samples from a pre-trained LDM; in the latter, LDMs with depth conditioning are used to iteratively denoise the rendered images which are aggregatively projected to the texture map. However, SDS-based methods are computationally feasible only with low-resolution LDMs and small NeRFs, and are unable to synthesize fine details. Additionally, approaches that directly denoise and project must contend with artifacts arising from view inconsistencies and synthesized textures may contain unnatural tones or lighting effects as residuals from their image-trained LDM backbone. + +Alternative representational spaces to drive textured 3D synthesis have also been explored. Texture Fields [17, 39] encode textures as 3D neural fields and triplane feature maps have emerged recently as an efficient domain for 3D geometry [56, 58, 63]. In particular, Sin3DM [58] encodes geometry and texture into an implicit triplane latent representation, then trains and samples a 2D DM on the resulting feature maps. However, geometry and texture must first be rasterized to a 3D grid before triplane convolutions can be employed, which can lead to aliasing of high-fidelity textures and geometric details due to memory constraints on the maximal grid resolution. To disentangle geometry from texture, several methods map textures to domains such as the image plane [6, 40, 51, 60] or sphere [7]. Point-UV diffusion [61] trains and samples 2D DMs directly in the UV-atlas, though this approach suffers on meshes with disconnected or fragmented UV mappings. + +Similar to our approach, a third class of models encode textures as features at the simplices of a mesh. GAN-based + +methods Texturify [48] and Mesh2Tex [2] compress textures to attributes at the mesh facets, and differentiable rendering enables adversarial supervision in the image space. + +Alternatively, Intrinsic Neural Fields (INFs) [27] introduces a latent representation wherein textures are encoded as eigenfunctions of the Laplacian taking values at the mesh vertices. However, features at points on faces are recovered through barycentric interpolation, which we will show leads to a breakdown in reconstruction fidelity at higher compression ratios. Recently, Manifold Diffusion Fields [14, 64] introduce the first fully intrinsic DMs defined over 2D Riemannian manifolds, and are based on a global attention mechanism which aggregates features at the vertices, limiting their applicability to coarse, low-resolution meshes due to complexity constraints. Here, our proposed FLDMs are also fully intrinsic but are built upon locally-supported convolutions defined over the manifold [35], allowing them to scale to higher-resolution meshes. + +Our single-textured-mesh FLDMs are inspired by recent work in which DMs are trained to capture internal patch distributions from a single image with the goal of generating diverse samples with similar visual content [28, 38, 54]; In turn, these models can trace their lineage to the influential SinGAN architecture [46] and its derivatives [19, 21]. Here, we mirror the approach proposed by SinDiffusion [54] and denoise with a shallow convolutional UNet to control the receptive field and prevent the DM from overfitting to the texture. Similar to our setting, Sin3DM [58] was the first to train LDMs in a single-textured-mesh paradigm, however they take an extrinsic approach to generating geometry and texture, whereas our approach focuses only on synthesizing textures which we show qualitatively are of higher-fidelity than those produced by Sin3DM. + +# 3. Method Overview + +We introduce a framework for intrinsic LDMs operating directly on surfaces. The first component is a latent representation of mesh textures comprised of tangent vector features. At each vertex, the tangent features characterize the texture of a local neighborhood on the surface, and the collection of these features constitute a stack of vector fields we call Field Latents (FLs). The FL space is learned with a locally-supported variational autoencoder [25] (FL-VAE) which is described in Section 4. + +The second key component is the Field Latent Diffusion Model (FLDM), which learns to denoise a diffusion process in the tangent space of a surface. Denoising networks are constructed with field convolutions (FCs), surface convolution operators acting on tangent vector fields [35]. We extend FCs to interleave scalar embeddings with the tangent vector features, enabling the injection of diffusion timesteps and optional conditioning, such as user-specified labels, into the denoising model. + +In practice, FL-VAEs and FLDMs are applied to synthesize new textures from a single textured triangle mesh. We pre-train a single FL-VAE by superimposing planar meshes over a large high-quality image dataset, obtaining a general-purpose latent space which can be used to train FLDMs on arbitrary textured meshes. To do so, a texture is first mapped to distributions in the tangent space at the vertices with the FL-VAE encoder, and an FLDM is trained to iteratively denoise tangent vector latent features. Afterwards, samples from the FLDM can be decoded with the FL-VAE decoder to synthesize new textures. + +Importance of Isometry-Equivariance Our FL-VAEs and FLDMs are designed to be equivariant under isometries, i.e. they commute with distance-preserving shape deformations. Unlike images, points on a surface have no canonical orientation. The choice of local coordinates is ambiguous up to an arbitrary rotation, making it impossible to simply adapt standard image-based VAEs and diffusion models to surfaces. However, isometries manifest locally as rotations. Thus, designing our framework to be isometry-equivariant inherently solves the orientation ambiguity problem, enabling consistent, repeatable results. Isometry-equivariance also results in tangible benefits — our models are able to seamlessly reproduce textural details across locally similar areas of meshes. Furthermore, while existing models transfer textures with pointwise maps [11] or generatively via conditioning on a learned token [5, 43], we can simply sample our pre-trained FLDMs on new, similar meshes to texture them in the learned style. + +# 4. Field Latents + +As in Knoppel et al. [26] we associate tangent vectors with complex numbers. For a surface $M$ and point $p \in M$ , we assign to the tangent space an arbitrary orthonormal basis such that for any $\mathbf{v} \in T_pM$ we have $\mathbf{v} \equiv r e^{i\theta}$ , with $r = |\mathbf{v}|$ and $\theta$ the direction of $\mathbf{v}$ expressed in the frame. We make this convention explicit by denoting $\mathbb{C}_p \equiv T_pM$ . + +To describe the FL-VAE, we require a notion of multidimensional Gaussian noise in the tangent bundle of a surface $M$ . Specifically, we denote the Gaussian distribution over $d$ copies of the tangent bundle $TM^{d}$ by $\mathcal{T}\mathcal{N}_M(0,I_d)$ . Samples $\epsilon \sim \mathcal{T}\mathcal{N}_M(0,I_d)$ are vector fields in $TM^{d}$ such that at each point $p\in M$ , the coefficients of $\epsilon (p)\in \mathbb{C}_p^d$ expressed in the orthonormal basis are themselves samples from the $d$ -dimensional standard normal distribution + +$$ +\epsilon (p) = \epsilon_ {1} + i \epsilon_ {2}, \epsilon_ {1}, \epsilon_ {2} \sim \mathcal {N} (0, I _ {d}). \tag {1} +$$ + +In this work we are concerned with isometries $\gamma : M \to N$ , which enjoy the property that their push-forwards $d\gamma : TM \to TN$ , which define a smooth map between tangent spaces, manifest pointwise as rotations. From Equa- + +tion (1) it is easy to see that $\mathcal{T}\mathcal{N}_M(0,I_d)$ is symmetric under such transformations. Thus, it follows that sampling from $\mathcal{T}\mathcal{N}_N(0,I_d)$ is equivalent to pushing forward samples from $\mathcal{T}\mathcal{N}_M(0,I_d)$ , + +$$ +\epsilon^ {\prime} = \gamma \epsilon \quad \text {f o r} \quad \begin{array}{l} \epsilon^ {\prime} \sim \mathcal {T N} _ {N} (0, I _ {d}) \\ \epsilon \sim \mathcal {T N} _ {M} (0, I _ {d}) \end{array} \tag {2} +$$ + +with $\gamma \epsilon = [d\gamma \cdot \epsilon] \circ \gamma^{-1}$ denoting the standard action of diffeomorphisms on vector fields via left-shifts. + +FL-VAE Encoder The encoder $\mathcal{E}$ is a network that, for any point $p\in M$ , takes $n$ -dimensional scalar functions $\psi \in L^2 (M,\mathbb{R}^n)$ (e.g. texture RGB values with $n = 3$ restricted to the surrounding geodesic neighborhood $B_{p}\subset$ $M$ to the parameters of $d$ independent normal distributions in the tangent space at $p$ .. + +$$ +\begin{array}{l} \mathcal {E} _ {p}: L ^ {2} \left(B _ {p}, \mathbb {R} ^ {n}\right)\rightarrow \mathbb {C} _ {p} ^ {d} \times \mathbb {R} _ {\geq 0} ^ {d} \tag {3} \\ \psi \mapsto \left(\mu_ {p} ^ {\psi}, \sigma_ {p} ^ {\psi}\right) \\ \end{array} +$$ + +Here, the means of the distributions $\mu_p^\psi$ are themselves tangent vectors while the standard deviations $\sigma_p^\psi$ are scalars. Latent codes characterizing the local restriction of $\psi$ are collections of tangent vectors $z_{p}^{\psi}\in \mathbb{C}_{p}^{d}$ drawn from the multivariate normal distribution parameterized by $(\mu_p^\psi ,\sigma_p^\psi)$ . Using the reparameterization trick [25] latent codes can be generated as $z_{p}^{\psi} = \mu_{p}^{\psi} + \sigma_{p}^{\psi}\odot \epsilon (p)$ with $\epsilon \sim \mathcal{T}\mathcal{N}_M(0,I_d)$ . + +FL-VAE Decoder The principal advantage in a tangent vector latent representation is that it naturally admits a descriptive coordinate function which can be queried to associate features with neighboring points. The decoder $\mathcal{D}$ is designed as a neural field operating over the local parameterization of the surface induced by the logarithm map about a point $p$ , $\log_p:B_p\to \mathbb{C}_p$ , with $\log_p q$ encoding the position of $q\in B_{p}$ in the tangent space of $p$ . Formally, + +$$ +\begin{array}{l} \mathcal {D} _ {p}: \log_ {p} \left(B _ {p}\right) \times \mathbb {C} _ {p} ^ {d} \rightarrow \mathbb {R} ^ {n} \tag {4} \\ \left(\log_ {p} q, z _ {p} ^ {\psi}\right) \mapsto \widehat {\psi} _ {p} (q), \\ \end{array} +$$ + +with $\widehat{\psi}_p$ denoting the prediction of $\psi$ made from the perspective of $p$ . Decoding follows a similar process as proposed in [18]. Specifically, given a tuple $(\log_p q, z_p^\psi)$ , the decoder constructs two features. First, an invariant scalar feature is obtained by vectorizing the upper-triangular part of the Hermitian outer product of the latent code with itself, $\mathrm{vec}_{j \geq i} \left(z_p^\psi \left[z_p^\psi\right]^*\right) \in \mathbb{C}^{\frac{d(d+1)}{2}}$ . Second, a positionally-aware scalar feature $c_{pq}^\psi \in \mathbb{C}^d$ is formed by the natural coordinate function + +$$ +c _ {p q} ^ {\psi} \equiv \log_ {p} q \cdot \overline {{z _ {p} ^ {\psi}}}, \tag {5} +$$ + +which corresponds to storing the inner product and determinant of each tangent vector with the position of $q$ relative to $p$ given by $\log_p q$ . The neural field receives the concatenation of these two features and returns the prediction $\widehat{\psi}_p(q)$ . + +Equivalence Recall we are concerned with constructing a latent representation equivariant under isometries — that for any $\psi \in L^2(M, \mathbb{R}^n)$ and isometry $\gamma: M \to N$ we have + +$$ +\left[ \widehat {\gamma \psi} \right] _ {\gamma (p)} = \gamma \widehat {\psi} _ {p}, \tag {6} +$$ + +where $\gamma \psi = \psi \circ \gamma^{-1}$ denotes the standard action of diffeomorphisms by left-shifts. Thus, it follows from Equation (2) that a sufficient condition for Equation (6) to hold is if for all $\psi \in L^2(M, \mathbb{R}^n)$ , isometries $\gamma : M \to N$ , and $p \in M$ , the encoded mean $\mu_p^\psi$ and standard deviation $\sigma_p^\psi$ satisfy + +$$ +\left. d \gamma \right| _ {p} \cdot \mu_ {p} ^ {\psi} = \mu_ {\gamma (p)} ^ {\gamma \psi} \quad \text {a n d} \quad \sigma_ {p} ^ {\psi} = \sigma_ {\gamma (p)} ^ {\gamma \psi}, \tag {7} +$$ + +with $d\gamma|_p$ denoting the push-forward of $\gamma$ in the tangent space at $p$ , which rotates the local coordinate system. A detailed proof appears in the supplement, sec. 8.1. + +Architecture In practice, we discretize a surface $M$ by a triangle mesh with vertices $V$ and faces $F$ . Textures are scalar functions $\psi$ defined continuously over the faces and vertices, taking values in $\mathbb{R}^3$ . For each $p \in V$ , we consider its neighborhood to be the surrounding one-ring, the collection of faces in $F$ that share $p$ as a vertex. The encoder architecture is chosen so as to satisfy the conditions in Equation (7). A fixed number of points $\{q_i\}$ are uniformly sampled in the one-ring neighborhood at which the texture function is evaluated. The resulting scalar features $\{\psi(q_i)\}$ are interleaved with the logarithms $\{\log_p q_i\}$ to construct tangent vector features which are concatenated with a token following [12]. The features are passed to eight successive VN-Transformer layers [1], modified to process tangent vector features. Afterwards, the token is extracted and used to predict the mean and standard deviation. + +The neural field comprising the decoder consists of a five-layer real-valued MLP, to which the real and imaginary parts of the complex feature vector are passed. For any point in a triangle, the latent codes at the adjacent vertices provide three distinct predictions which are linearly blended with barycentric interpolation only at inference. This is in contrast to INF [27], wherein linearly blended features are passed to the neural field to make predictions. A comprehensive discussion of the FL-VAE architecture and training objective is provided in section 9.1 in the supplement. + +# 5. Field Latent Diffusion Models + +FLDMs define a noising process on surface vector fields, and learn a denoising diffusion probabilistic model + +
FL-VAE22.38 (20.59)0.51 (0.83)1.02 (1.81)
FL-VAE (Bary.)21.33 (19.61)0.66 (1.01)1.31 (2.03)
INFs [27]18.86 (16.45)1.16 (1.64)2.15 (2.73)
+ +Table 1. Texture reconstruction quality on meshes from the Google Scanned Objects dataset [13]. Reconstructions are evaluated on high-res (30K vertices) and low-res (5K vertices, in parenthesis) meshings of the objects. Metrics are computed in the texture atlases, with DSSIM and LPIPS scaled by a factor of 10. + +(DDPM) [22] using an equivariant surface network. In principle they can be applied to any vector data on surfaces, and here we use them to learn feature distributions in FL space. + +Latent vector fields $Z \in TM^{d}$ can be produced by compressing a texture $\psi \in L^{2}(M, \mathbb{R}^{3})$ to a collection of tangent vector features at each point with the FL-VAE encoder as in Equation (3) and sampling from the distributions. The forward process progressively adds noise in the tangent space, dictated by a monotonically decreasing schedule $\{\alpha_{t}\}_{0 \leq t \leq T}$ with $\alpha_{0} = 1$ and $\alpha_{T} = 0$ . The noised vector fields can be expressed at an arbitrary time $t$ in the discretized forward process with + +$$ +Z _ {t} = \sqrt {\alpha_ {t}} Z + \sqrt {1 - \alpha_ {t}} \epsilon \in T M ^ {d}, \tag {8} +$$ + +where $\epsilon \sim \mathcal{T}\mathcal{N}_M(0,I_d)$ as in Equation (1). + +The denoising function $\varepsilon$ is a surface network trained to reverse the forward process. Specifically, it is a learnable map on the space of latent vector fields, conditioned on timesteps $t$ and optional scalar embeddings $\rho \in L^{2}(M,\mathbb{R}^{m})$ , with the goal of predicting the added noise, + +$$ +\varepsilon : T M ^ {d} \times \mathbb {R} _ {\geq 0} \times L ^ {2} (M, \mathbb {R} ^ {m}) \rightarrow T M ^ {d}, \tag {9} +$$ + +trained subject to the loss + +$$ +L _ {\mathrm {F L D M}} = \mathbb {E} _ {\epsilon \sim \mathcal {T N} _ {M} (0, I _ {d}), 0 \leq t \leq T} \| \epsilon - \varepsilon \left(Z _ {t}, t, \rho\right) \| _ {2} ^ {2}. \tag {10} +$$ + +At inference, new latent vector fields are generated following the iterative DDPM denoising process. Beginning with representing the fully-noised vector fields at step $T$ by a sample $\tilde{Z} \sim \mathcal{T}\mathcal{N}_M(0,I_d)$ , $\tilde{Z}_T = \tilde{Z}$ , the trained denoising function $\varepsilon$ is used to estimate the previous step in the forward process via the relation + +$$ +\begin{array}{l} \widetilde {Z} _ {t - 1} = C _ {1} \left(\alpha_ {t}, \alpha_ {t - 1}\right) \widetilde {Z} _ {t} + C _ {2} \left(\alpha_ {t}, \alpha_ {t - 1}\right) \varepsilon \left(\widetilde {Z} _ {t}, t, \rho\right) \tag {11} \\ + C _ {3} \left(\alpha_ {t}, \alpha_ {t - 1}\right) \epsilon , \\ \end{array} +$$ + +with $\epsilon = 0$ for $t = 1$ and $\epsilon \sim \mathcal{T}\mathcal{N}_M(0,I_d)$ otherwise, and $C_i(\alpha_t,\alpha_{t - 1})\in \mathbb{R}$ as defined in 9.1 of the supplement. + +Equivalence From the equivalence of the distributions $\mathcal{T}\mathcal{N}_M(0,I_d)$ and $\mathcal{T}\mathcal{N}_N(0,I_d)$ under isometries as in Equation (2), it follows that for any isometry $\gamma :M\to N$ a sufficient condition for both the forward and reverse processes in Equations (8) and (11) to satisfy + +$$ +\gamma Z _ {t} = \left[ \gamma Z \right] _ {t} \quad \text {a n d} \quad \gamma \widetilde {Z} _ {t} = \left[ \gamma \widetilde {Z} \right] _ {t}, \tag {12} +$$ + +![](images/63aa0937b3f29c8b94195d24c6d5cee7638a21bdca27d7c809e816a3e9a4036f.jpg) + +![](images/ae336f6b2f726b5de02550b5ebf7e68b2210b329ccbafdf06f88ee2138c31873.jpg) + +![](images/70388add48c7ab22cf2ca238ca263be278195e2ada3d3339a86d2668c00135a9.jpg) + +![](images/44f08460a5951f34d3f96f17cfa280c9cd3ad7031638c69d9321225cddf6df43.jpg) + +![](images/6a688597a5ff870651a05b2f0e38e1fb2ad063e274e5cbfaf8c3d580c4dbc04f.jpg) + +![](images/a342393567a96786c96eb29cfafa83b3faf22ef63c91d4598a030a12ddf3a8b2.jpg) + +![](images/eaebf572d728ea734cb7aa96bc0801ca31832af8099aab4cb30dabc4899bbd83.jpg) + +![](images/02579870a66ff82c2480b1755db40022a86861c59411cf9a2fde4ad379bc3cdd.jpg) + +![](images/d1deb7074a634d3fddfa7522b18a356eb8e7a04524c3b68c5d37aa735cd2c0fd.jpg) + +![](images/a7640465f10645669e6687715d83009fc5e35b4a45aa39ebbbd7649fab780506.jpg) +Figure 2. Visual comparison of textures compressed and reconstructed with the FL-VAE and INF [27] on 30K vertex meshes. Compared to barycentric coordinates, our proposed logarithmic coordinate function more richly extends latent features across the mesh, enabling the reconstruction of finer details. Zoom in to view. + +![](images/aeb2230304ddfa3e2618256dcb1c45c272da83db316f064b7b9062e28cdbd379.jpg) + +![](images/7b13595d074786f40764e31b3c1710f887a75c01b7084d7abce7a0db15faf0a4.jpg) + +![](images/0aa349d9b15df4819afdad900a8967ad32e95a4917cbcee429b09deca8e68c7f.jpg) + +![](images/c110f4efea2af05364ededd69d3518732e2bc890f4274d8895ff389d6e89053b.jpg) + +![](images/a53dd0f6ade80852fd2ef99fe245a55752d305902d47d930b7ecacd64354e073.jpg) + +![](images/f0859b85236abd84b97d9f4baf2e643916bb89e4fdabe9356946babec1031dc9.jpg) + +for $0 \leq t \leq T$ is if the denoising network $\varepsilon$ in Equation (9) has the property that for any vector field $Z \in TM^{d}$ , + +$$ +\gamma \varepsilon (Z, t, \rho) = \varepsilon (\gamma Z, t, \gamma \rho). \tag {13} +$$ + +A detailed proof is provided in sec. 8.2 of the supplement. + +Denoising with Field Convolutions To satisfy the condition in Equation (12), our denoising network is constructed with field convolutions (FC) [35], which belong to a class of models that can operate on tangent vector fields [45, 57]. + +FCs convolve vector fields $Z \in TM^{C}$ with $C' \times C$ filter banks $\mathbf{f}_{c'c} \in L^{2}(\mathbb{C},\mathbb{C})$ , returning vector fields $Z * \mathbf{f} \in TM^{C'}$ . In conventional diffusion models, features inside the denoising network are conditioned on timesteps $t$ via an additive embedding applied inside the convolution blocks. Unfortunately, no direct analogue exists for vector field features on surfaces — additive embeddings break equivariance and we find that multiplicative embeddings destabilize training. Thus, to enable conditioning of our denoising network $\varepsilon$ on both timesteps $t$ and optional user-input features $\rho \in L^{2}(M,\mathbb{R}^{m})$ , we inject embeddings directly into convolutions by extending FCs to convolve vector fields with filters $\mathbf{f}_{c'c} \in L^{2}(\mathbb{C} \times \mathbb{R}^{e},\mathbb{C})$ , with $e$ the embedding dimension. Critically, this formulation preserves equivariance without sacrificing stability. + +Following [54], the denoising network $\varepsilon$ takes the form of a shallow two-level U-Net but with field convolutions; the relatively small size of the network's receptive field allows distributions of latent features to be learned without overfitting to the single textural example [28, 54, 58]. + +![](images/89286772ec9ca37e6ef226617e7584df1bb16c2c37964c67f7070548d8051ac5.jpg) + +![](images/6613f4a9698ceb5e75e2b08d786bff7cd1003b115a6eea1e36dbdfcadfa49c92.jpg) +Figure 3. Visual comparison of unconditionally (label-free) generated textures from FLDMs and Sin3DM [58]. Our FLDM's isometry-equivariant construction allows for replication of textural details across locally similar regions. In contrast, Sin3DM's extrinsic approach associates textural features with 3D space; Synthesized details appear as repetitions or extrusions of previous patterns along the major axes of the mesh, and we observe that novel textures cannot be created without modifying geometry. Zoom in to compare. + +![](images/5283b246c8ecd84f70136fc56a568a6d95fa51b11153d0059994c15615374ed9.jpg) +FLDM + +![](images/8604ae5206e9c1fe87abc7c2440a3084246b6286888b675292ebced34502433f.jpg) + +![](images/5e4923766cf947fc7d7d9745ed3081f6f0fed49b82fe75ee42704ca32893a132.jpg) + +![](images/16e6d47d14b2e30f986e94b85867a4ab77ce80380a652009699f5462c6f6922c.jpg) + +![](images/a3fd7fbbca6319ae00f4572a50b323e97db54259e3d629ba5e944245fcbd1612.jpg) +Sin3DM [58] + +![](images/165e4738813c0d860ab3141560a148bee157967e3c1f00a5d0772defbcb31934.jpg) + +![](images/547c9e9021a7d0de0267627142322c2941bfcfe4a73e758f2447ae8ca61bc0af.jpg) + +![](images/0bbb7cf194d616ba200e2b29d659952f2b9cc4f5dcc2242b83c05ffd6b8a39d1.jpg) + +![](images/e8ac167370b7d8a63d65bd087caad30909aca91346ff76aeb9be7f969fec8df7.jpg) +Input + +![](images/f5b1aa814fca902e7c7f692954e09aa5dc48a6d306295fc1a047a061e0a8ed15.jpg) +FLDM + +![](images/c07ecad3d77d567a7c0e67950921e1fc3a4f1dd8be453f61c48b9f5c0a47181b.jpg) + +![](images/55d5c396fe30615aafcf0e4e07a9e8635292d28063d61ed73ca0ffa841d556ca.jpg) +Sin3DM [58] + +![](images/721c0cbd2fff4f86060934de71bd4e97803189172b4f2ca664fc7d87bbbf0c67.jpg) + +![](images/44cd2fd89375eba06d65280e4f46d1101ce39261a2f01efcaf59966169f69eae.jpg) + +![](images/95944ef82cf4e16994f6b8b162a6909666b1428bc680c4a8d2b217b98057a67d.jpg) + +![](images/fbcbf17b062b7b2ccfcfa262582c79117500b87a4e788a928742996988618dcc.jpg) + +
MethodSIFID ↓LPIPS ↑
FLDM3.271.15
Sample (left)0.710.94
Sin3DM [58]6.582.20
Sample (right)2.422.91
+ +![](images/a6ed50c83dd3bed48f94ef1102ecc92854b7ffa4cf2b3dcd284cd2798ac4d0a6.jpg) + +![](images/a17bb953d8d301a11e5657a74f90b11831f0b50e0dbac5324bdd3d56a0ec4363.jpg) + +Table 2. Fidelity and diversity of unconditionally generated textures trained on selected Objaverse [8] and Scanned Objects [13] meshes. SIFID and LPIPS are scaled by factors of $10^{5}$ and 10, respectively. Unlike our FLDMs, Sin3DM synthesizes new geometry which can inflate diversity scores. As an example, the individual metrics are shown for the FLDM and Sin3DM samples on the right. The latter attains a high LPIPS score, but exhibits poor textural diversity and quality. + +# 6. Experiments + +Pre-training the FL-VAE It is important to note that the FL-VAE proposed in Section 4 only interfaces with the mesh geometry through its local flattening in the image of the logarithm map. Thus, the FL-VAE operates independently of the local curvature, and a model pre-trained' on flat textured meshes can be used to encode and decode textures on arbitrary 3D shapes. + +In all of our experiments we use the same FL-VAE with a $d = 8$ dimensional latent space, pre-trained on the OpenImagesV4 dataset [29]. Specifically, we generate 1K different planar meshes containing 10K vertices each. Each sample $512 \times 512$ image is overlaid by a randomly chosen triangulation, with the image serving as the "texture" defined over the mesh. The FL-VAE is then trained with the sum of reconstruction loss and KL divergence using the Adam [24] optimizer. We deliberately train the FL-VAE on coarser triangulations than we expect during inference on textured meshes. This forces the model to reconstruct high frequencies during training, which we find increases robustness to sampling density and triangulation quality, in addition to improving the fidelity of reconstructed FLDM samples. + +Training single-textured-mesh FLDMs All diffusion experiments share the same training regime. During preprocessing, we create 500 copies of a given mesh each with 30K vertices and different triangulations to prevent the denoising network from learning the connectivity. Each copy shares the same UV-atlas as the original and the texture is mapped to vector field latent features at the vertices using the pre-trained FL-VAE encoder. FLDMs have a max timestep of $T = 1000$ and noise is added on a linear schedule. The denoising network is trained subject to the loss in Equation (10) using the Adam [24] optimizer. + +# 6.1. Texture Compression and Reconstruction + +First we quantitatively evaluate the capacity of our proposed FL-VAE for compressing and reconstructing textures. We compare FL-VAEs against two other approaches: the state-of-the-art INF [27] and a modified version of FL-VAEs serving as an ablation. INFs train a neural field per texture and predicts the texture values using the values of the Laplacian eigenfunctions extended over triangles through barycentric interpolation. Similarly, the modified FL-VAE (denoted FL-VAE Barycentric) omits our proposed logarithmic coordinate function in Equation (5) in favor of linearly interpolating the values of the invariant outer product features across triangles. It is otherwise identical to our proposed approach, including pre-training methodology. Thus, this also serves to compare the efficacy of our proposed coordinate function against barycentric interpolation in extending vertex features continuously over the mesh. + +To evaluate, we select 16 objects exhibiting a range of complex textures with high-frequency details from the Google Scanned Objects dataset [13]. Two versions of the dataset are used, with objects remeshed to have $30\mathrm{K}$ (high-res) and $5\mathrm{K}$ (low-res) vertices, respectively. The single pretrained FL-VAE is used to map the textures to distributions over the tangent space, and the mean is used to sample the reconstructed texture at the pixel indices in the texture atlas. + +![](images/3ccdf64a5f1b102d6d0808145bdd38b6605576cf305c631992e128d922208115.jpg) +Figure 4. Label-guided generation. FLDMs can be conditioned on a subjective user-input labeling which generated textures will reflect. See Figure 1 for more examples. Zoom in to view. + +A separate INF neural field is trained for each mesh, using the specified regime for high-res reconstruction [27]. + +Following INFs, we evaluate the quality of the reconstructed textures using the average PSNR, DSSIM [55], and LPIPS [62] metrics computed in texture atlases. The results are reported in Table 1, with several examples of reconstructions shown in Figure 2. Our proposed FL-VAE achieves the best results across all metrics at both resolutions, followed by the barycentric FL-VAE and INFs. The significant gap in performance between the barycentric FL-VAE and INFs is likely due to the reliance of the latter the on the Laplacian eigenfunctions as input features. Areas of fine textural detail may not necessarily overlap with regions where the eigenfunctions exhibit high variability, potentially contributing to the "raggedness" observed in INF reconstructions. Furthermore, the superior performance of our proposed FL-VAE relative to its barycentric counterpart suggests that our logarithmic coordinate function in Equation (5) provides a richer interpolant over the mesh faces. This directly increases the representational capacity of the latent features, enabling the reconstruction of finer textural details with the same number of features. + +# 6.2. Unconditional (Label-Free) Generation + +Label-free generation can be performed by training and sampling FLDMs conditioned only on the diffusion timescale (i.e. $\rho = 0$ ). We compare against Sin3DM [58] which to our knowledge is the only prior generative model designed to operate in a single-textured-mesh paradigm. We follow their proposed evaluation protocol, and train FLDMs and Sin3DM on 10 assets from the Objaverse [8] and Scanned Objects [13] datasets selected for texture complexity and diversity. For each asset, 64 textures are sampled from the trained models, which are rendered on the mesh from 24 different views. Fidelity is measured with the mean SIFID (Single Image Fréchet Inception Distance) [46] pairwise between the renderings and those of the input mesh from the same view; the mean LPIPS [62] between renders of samples from the same viewpoint measures diversity. We note that unlike FLDMs, which only synthesize new textures, Sin3DM synthesizes both textures and geometries, with the latter potentially inflating LPIPS scores. + +The results are shown in Table 2, Figure 3, and in the supplement (Figure 8). Compared to Sin3DM, our FLDMs + +![](images/5b61e246cc5e841824e8907e64d220456dc4a5b6718c26610993c0e1f1b1a0ac.jpg) +Figure 5. Inpainting with FLDMs. The input texture is preserved in the masked regions and new content is synthesized elsewhere with agreement on the boundaries. Zoom in to view. + +are able to synthesize higher-fidelity textures which is likely due in part to the former's rasterization of textured geometry to a 3D grid before encoding to the latent space, potentially aliasing fine details. More generally, we observe that by operating over an extrinsic representation of textured geometry, Sin3DM intertwines texture with a mesh's 3D embedding. Thus, new textural details appear as repetitions or extrusions of earlier patterns along the major axes of the models, as seen in the vase, snake, and sculpture samples in Figure 3. Furthermore, we observe that texture and position are so strongly linked that Sin3DM is unable to produce significant textural alterations without correspondingly modifying the geometry, as seen in the vase, snake, and lantern samples. Here, the advantage of our isometry-equivariant formulation becomes clear as our FLDMs can effectively replicate textural details across areas of the mesh that are locally similar, regardless of the relative configuration of these regions in the 3D embedding. + +# 6.3. Label-Guided Generation + +Textured objects often exhibit distinct content specific to certain regions of the mesh which may not be geometrically unique; for example, in Figure 1 regions on the sole, body, and inside of the shoe lack distinguishing geometric features, as is the case with the eyes and mouth of the skull and the decal on the bottle. In such cases, a user may find it desirable for synthesized textures to reflect a subjective distribution of content on the input mesh, which we facilitate by extending our denoising model to incorporate optional conditioning from user-designated mesh features $\rho \in L^{2}(M,\mathbb{R}^{m})$ . In the simplest case, the user can paint a coarse, semantic labeling to subjectively delineate salient regions, such as the flippers, head, and rock base of the seal sculpture in Figure 4. After training an FLDM conditioned on the labels, generated textures reflect the specified segmentation. Further examples are shown in Figure 1. Labeled areas contain different textural features, and conditioning ensures the FLDM samples respect these distributions, e.g. that teeth are synthesized only on the mouth. + +# 6.4. Inpainting + +In certain cases a user may wish to preserve the input texture in certain regions while synthesizing new content elsewhere. Texture inpainting is a well studied topic + +![](images/f72422ff655744fd6b1f7f490df08dc519195e489d602626c39463dbd4a07826.jpg) +FLDM samples across geometries +Figure 6. Generative texture transfer. Since our FL-VAE and FLDMs commute with isometries, we can encode a texture and train an FLDM on an input mesh (left), then sample the pre-trained FLDM on a new, similar mesh and decode the latents to texture it (center to right). Conditioning labels segmenting the shell apertures are not visible. Zoom in to view. + +[15, 20, 31], and here we propose a straight-forward approach with FLDMs. Given a pre-trained FLDM conditioned on a user-input binary mask $m: M \to \{0, 1\}$ specifying the regions to preserve (Figure 5), generative inpainting can be performed by modifying the iterative sampling process. Specifically, after estimating previous the latent features $\widetilde{Z}_{t-1}$ as in Equation (11), the features in the mask region are replaced with the appropriately noised input latents from the forward process $Z_{t-1}$ such that + +$$ +\tilde {Z} _ {t - 1} \mapsto m \cdot Z _ {t - 1} + (1 - m) \cdot \tilde {Z} _ {t - 1}. \tag {14} +$$ + +This approach encourages the inpainted texture to agree with the original at the mask boundaries due to convolutional structure of the denoising network, with this effect observed on the hind leg of sampled textures in Figure 5. + +# 6.5. Generative Texture Transfer + +Perhaps the most dramatic consequence of isometry-equivariance is that it naturally facilitates a notion of generative texture transfer. Since the FL-VAE and FLDMs commute with isometries $\gamma : M \to N$ as in Equations (6) and (12), our models see $TM^{d}$ and $TN^{d}$ as functionally the same space. In practice, we find that long as surfaces $M$ and $N$ are approximately isometric, we can encode a texture and train an FLDM on $M$ , then sample the model on $N$ and decode the synthesized latents to texture it. + +![](images/9370a8f36dc0dadec68557824cf9f71cfdebb2208ddbf2569c4dc7773ae9b9c2.jpg) +FLDM samples on increasingly coarse remeshings +Figure 7. Controlling the scale of synthesized textures. Here, the FLDM trained on the textured skull in the center-left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry, dilating the size of the tesserae. Zoom in to view details. + +Several examples are shown in Figure 6. An FLDM is pre-trained with label conditioning on each of the textured shells in the first column and sampled with corresponding labels on the others to texture them in the same style. Additional examples are shown on the right side of Figure 1. Notably, our model successfully transfers between different topologies, as the skull meshes are genus zero and three. More generally, we are able to transfer textures between highly dissimilar shapes (see supplement for examples). + +# 6.6. Controlling Textural Scale + +An additional feature of our FLDMs is that they offer user-control over the scale of sampled textures. Internally, FLDMs normalize the filter support by dividing the distances between adjacent vertices by the radius of the mean vertex area element $\widetilde{r} = \sqrt{A / (\pi|V|)}$ , where $A$ is the surface area of the mesh and $|V|$ is the number of vertices. Thus, the scale of the synthesized textural details can be controlled by increasing or decreasing the resolution of the mesh upon which the pre-trained FLDM is sampled. An example shown in Figure 7, where the FLDM trained on the textured skull on the center left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry. As the number of vertices decreases, the average one-ring occupies an increasingly larger fraction of the surface area, dilating the size of the synthesized textural details. + +# 7. Conclusion and Limitations + +We present an original framework for latent diffusion models operating directly on the surfaces to generate high-quality textures. Our approach consists of two main contributions: a novel latent representation mapping textures to vector fields called Field Latents and Field Latent Diffusion Models which learn to denoise a diffusion process in the latent space on the surface. We apply our method in a single-textured-mesh paradigm, and generate variations of textures with state-of-the-art fidelity. Limitations of our approach include the relatively long training time of FLDMs and an inability to reflect directional information in synthesized textures. The latter could potentially be addressed by conditioning FLDMs on user-specified vector fields, with embeddings formed via dot products with network features. + +# References + +[1] Serge Assaad, Carlton Downey, Rami Al-Rfou, Nigamaa Nayakanti, and Ben Sapp. Vn-transformer: Rotation-equivariant attention for vector neurons. arXiv preprint arXiv:2206.04176, 2022. 4, 3 +[2] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. arXiv preprint arXiv:2304.05868, 2023. 3 +[3] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, pages 4169-4181, 2023. 2 +[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2 +[5] Qimin Chen, Zhiqin Chen, Hang Zhou, and Hao Zhang. Shaddr: Real-time example-based geometry and texture generation via 3d shape detailization and differentiable rendering. arXiv preprint arXiv:2306.04889, 2023. 3 +[6] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2 +[7] An-Chieh Cheng, Xueting Li, Sifei Liu, and Xiaolong Wang. Tuvf: Learning generalizable texture uv radiance fields. arXiv preprint arXiv:2305.03040, 2023. 2 +[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 2, 6, 7 +[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects, 2023. 2 +[10] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In ICCV, pages 12200-12209, 2021. 2 +[11] Nicolas Donati, Etienne Corman, Simone Melzi, and Maks Ovsjanikov. Complex functional maps: A conformal link between tangent bundles. In Computer Graphics Forum, pages 317-334. Wiley Online Library, 2022. 3, 7 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4 +[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items, 2022. 1, 5, 6, 7 +[14] Ahmed A Elhag, Joshua M Susskind, and Miguel An + +gel Bautista. Manifold diffusion fields. arXiv preprint arXiv:2305.15586, 2023. 3 +[15] Julien Fayer, Bastien Durix, Simone Gasparini, and Géraldine Morin. Texturing and inpainting a complete tubular 3d object reconstructed from partial views. Computers & Graphics, 74:126-136, 2018. 8 +[16] Marc Finzi, Samuel Stanton, Pavel Izmailov, and Andrew Gordon Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In ICML, pages 3165-3176. PMLR, 2020. 4 +[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 2 +[18] James Gardner, Bernhard Egger, and William Smith. Rotation-equivariant conditional spherical neural fields for learning a natural illumination prior. NeurIPS, 35:26309-26323, 2022. 4 +[19] Niv Granot, Ben Feinstein, Assaf Shocher, Shai Bagon, and Michal Irani. Drop the gan: In defense of patches nearest neighbors as single image generative models. In CVPR, pages 13460-13469, 2022. 3 +[20] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, pages 12135-12144, 2019. 8 +[21] Tobias Hinz, Matthew Fisher, Oliver Wang, and Stefan Wermter. Improved techniques for training single-image gans. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1300–1309, 2021. 3 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5, 3 +[23] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 4 +[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 6, 5 +[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3, 4 +[26] Felix Knöppel, Keenan Crane, Ulrich Pinkall, and Peter Schröder. Globally optimal direction fields. ACM TOG, 32 (4):1-10, 2013. 3 +[27] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, pages 622-639. Springer, 2022. 3, 4, 5, 6, 7 +[28] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930. PMLR, 2023. 2, 3, 5, 4 +[29] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object + +detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6, 3 +[30] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2 +[31] A Maggiordomo, P Cignoni, and M Tarini. Texture inpainting for photogrammetric models. In Computer Graphics Forum. Wiley Online Library, 2023. 8 +[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 2 +[33] Thomas W. Mitchel. Extending Convolution Through Spatially Adaptive Alignment. PhD thesis, Johns Hopkins University, 2022. 4 +[34] Thomas W. Mitchel, Benedict Brown, David Koller, Tim Weyrich, Szymon Rusinkiewicz, and Michael Kazhdan. Efficient spatially adaptive convolution and correlation, 2020. 4 +[35] Thomas W Mitchel, Vladimir G Kim, and Michael Kazhdan. Field convolutions for surface cnns. In ICCV, pages 10001-10011, 2021. 2, 3, 5, 4 +[36] Thomas W Mitchel, Szymon Rusinkiewicz, Gregory S Chirikjian, and Michael Kazhdan. Echo: Extended convolution histogram of orientations for local surface description. In Computer Graphics Forum, pages 180-194. Wiley Online Library, 2021. 4 +[37] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 4 +[38] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2, 3 +[39] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, pages 4531-4540, 2019. 2 +[40] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In ICCV, pages 13879-13889, 2021. 2 +[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2 +[43] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 3 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image syn + +thesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2 +[45] Klaus Hildebrandt Ruben Wiersma, Elmar Eisemann. Cnns on surfaces using rotation-equivariant features. ACM TOG, 39(4), 2020. 5, 4 +[46] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019. 2, 3, 7 +[47] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM TOG, XX(X), 2022. 3, 4 +[48] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, pages 72-88. Springer, 2022. 3 +[49] Saurabh Singh and Shankar Krishnan. Filter response normalization layer: Eliminating batch dependence in the training of deep neural networks. In CVPR, pages 11237-11246, 2020. 2 +[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[51] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In ICCV, pages 7062-7072, 2023. 2 +[52] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model, 2023. 2 +[53] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting, 2023. 2 +[54] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 2, 3, 5, 4 +[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[56] Jiacheng Wei, Hao Wang, Jiashi Feng, Guosheng Lin, and Kim-Hui Yap. Taps3d: Text-guided 3d textured shape generation from pseudo supervision. In CVPR, pages 16805-16815, 2023. 2 +[57] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. Deltaconv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 5, 4 +[58] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2, 3, 5, 6, 7 +[59] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 4 +[60] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2 + +[61] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In ICCV, pages 4206–4216, 2023. 2 +[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7 +[63] Xuanmeng Zhang, Jianfeng Zhang, Rohan Chacko, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, pages 2273-2282, 2023. 2 +[64] Peiye Zhuang, Samira Abnar, Jiatao Gu, Alex Schwing, Joshua M Susskind, and Miguel Ángel Bautista. Diffusion probabilistic fields. In ICLR, 2022. 3 \ No newline at end of file diff --git a/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/images.zip b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3c149942a631197fed31822b8aa8504cfcd26792 --- /dev/null +++ b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b2120d0b8aee38fd974887b284caa2ec5686305ac038990be8781cf6febaad7 +size 544807 diff --git a/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/layout.json b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..324498ef28ce1d6c8005b0d66ffa76826f7758b1 --- /dev/null +++ b/2024/Single Mesh Diffusion Models with Field Latents for Texture Generation/layout.json @@ -0,0 +1,11851 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 520, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 520, + 120 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 520, + 120 + ], + "type": "text", + "content": "Single Mesh Diffusion Models with Field Latents for Texture Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 142, + 447, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 142, + 447, + 177 + ], + "spans": [ + { + "bbox": [ + 143, + 142, + 447, + 177 + ], + "type": "text", + "content": "Thomas W. Mitchel1,2† Carlos Esteves1 Ameesh Makadia1 1Google Research 2PlayStation" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 205, + 135, + 411 + ], + "blocks": [ + { + "bbox": [ + 81, + 190, + 103, + 201 + ], + "lines": [ + { + "bbox": [ + 81, + 190, + 103, + 201 + ], + "spans": [ + { + "bbox": [ + 81, + 190, + 103, + 201 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 205, + 135, + 411 + ], + "lines": [ + { + "bbox": [ + 48, + 205, + 135, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 135, + 411 + ], + "type": "image", + "image_path": "a00a360104b82d0ae3fe81b413533e9f05a385623c4ba7b5cd73c16127e41b35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 148, + 217, + 217, + 411 + ], + "blocks": [ + { + "bbox": [ + 148, + 217, + 217, + 411 + ], + "lines": [ + { + "bbox": [ + 148, + 217, + 217, + 411 + ], + "spans": [ + { + "bbox": [ + 148, + 217, + 217, + 411 + ], + "type": "image", + "image_path": "6c9f448fcd5320c1a5bdec708ab2a363a80c47d43271add1701150559da97947.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 416, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 547, + 472 + ], + "type": "text", + "content": "Figure 1. Our latent diffusion models operate directly on the surfaces of 3D shapes, synthesizing new high-quality textures (center) after training on a single example (left). Both our novel latent representation and diffusion models are isometry-equivariant, facilitating a notion of generative texture transfer by sampling pre-trained models on new geometries (right). Above our models are conditioned on coarse semantic labels reflecting a subjective distribution of content, which delineate the sole and interior of the shoes, the eyes and mouths of the skulls, and the decals on the bottles." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 223, + 217, + 291, + 411 + ], + "blocks": [ + { + "bbox": [ + 220, + 190, + 290, + 201 + ], + "lines": [ + { + "bbox": [ + 220, + 190, + 290, + 201 + ], + "spans": [ + { + "bbox": [ + 220, + 190, + 290, + 201 + ], + "type": "text", + "content": "Generated textures" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 217, + 291, + 411 + ], + "lines": [ + { + "bbox": [ + 223, + 217, + 291, + 411 + ], + "spans": [ + { + "bbox": [ + 223, + 217, + 291, + 411 + ], + "type": "image", + "image_path": "cced20cab01424c70df95a6caa97ce3cdf1b93275382754214b82dbcf2c7c992.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 297, + 217, + 365, + 411 + ], + "blocks": [ + { + "bbox": [ + 297, + 217, + 365, + 411 + ], + "lines": [ + { + "bbox": [ + 297, + 217, + 365, + 411 + ], + "spans": [ + { + "bbox": [ + 297, + 217, + 365, + 411 + ], + "type": "image", + "image_path": "e4ebd282a16e9549f0749cfa4eb455fd69342bf55cc0b3f72aa1a2659af62431.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 381, + 207, + 470, + 411 + ], + "blocks": [ + { + "bbox": [ + 396, + 190, + 529, + 201 + ], + "lines": [ + { + "bbox": [ + 396, + 190, + 529, + 201 + ], + "spans": [ + { + "bbox": [ + 396, + 190, + 529, + 201 + ], + "type": "text", + "content": "Generative transfer to new geometry" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 207, + 470, + 411 + ], + "lines": [ + { + "bbox": [ + 381, + 207, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 381, + 207, + 470, + 411 + ], + "type": "image", + "image_path": "d27906b94c9926129db751a83f30e561506b97bf8be3e715321946d534d20306.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 473, + 219, + 543, + 411 + ], + "blocks": [ + { + "bbox": [ + 473, + 219, + 543, + 411 + ], + "lines": [ + { + "bbox": [ + 473, + 219, + 543, + 411 + ], + "spans": [ + { + "bbox": [ + 473, + 219, + 543, + 411 + ], + "type": "image", + "image_path": "a90d4f2212b730fe890a0efbb1323d3bfd6e12568f0b5684a9689b2865fdddf8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 481, + 192, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 481, + 192, + 495 + ], + "spans": [ + { + "bbox": [ + 143, + 481, + 192, + 495 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 508, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 290, + 689 + ], + "type": "text", + "content": "We introduce a framework for intrinsic latent diffusion models operating directly on the surfaces of 3D shapes, with the goal of synthesizing high-quality textures. Our approach is underpinned by two contributions: Field Latents, a latent representation encoding textures as discrete vector fields on the mesh vertices, and Field Latent Diffusion Models, which learn to denoise a diffusion process in the learned latent space on the surface. We consider a single-textured-mesh paradigm, where our models are trained to generate variations of a given texture on a mesh. We show the synthesized textures are of superior fidelity compared those from existing single-textured-mesh generative models. Our models can also be adapted for user-controlled editing tasks such as inpainting and label-guided generation. The efficacy of our approach is due in part to the equivariance" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 483, + 546, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 546, + 554 + ], + "type": "text", + "content": "of our proposed framework under isometries, allowing our models to seamlessly reproduce details across locally similar regions and opening the door to a notion of generative texture transfer. Code and visualizations are available at https://single-mesh-diffusion.github.io/." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 582, + 387, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 387, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 387, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 604, + 546, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 546, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 546, + 688 + ], + "type": "text", + "content": "The emergence of latent diffusion models (LDMs) [44] as powerful tools for 2D content creation has motivated efforts to replicate their success across different modalities. A particularly attractive direction is the synthesis of textured 3D assets, due to the high cost of building photorealistic objects which can require intricately configured scanning solutions [13] or expertise in 3D modeling." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "The majority of 3D synthesis methods employing LDMs operate in an image-based setting where the 3D repre" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 189, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 189, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 189, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 62, + 702, + 189, + 712 + ], + "type": "text", + "content": " Work done while at Google Research." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7953" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "sentation is optimized through differentiably rendered images, often in conjunction with score distillation sampling (SDS) [3, 4, 32, 41-43, 52, 53]. This design allows the models to leverage the full capabilities of pre-trained 2D LDMs. A second, alternative method is to rasterize both geometry and texture onto a 3D grid which are compressed into triplane latent features over which the DM operates [58]. However, in applications where texture synthesis takes precedence over 3D geometry, both approaches sacrifice fidelity in favor of a workable representation — models relying on planar renderings suffer from view inconsistencies when mapping back to the 3D shape while rasterization inherently aliases fine textural details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 229, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 287, + 290 + ], + "type": "text", + "content": "Here, we observe that the discretization of a surface as a triangle mesh is itself a tractable and effective representational space. To this end, we present an original framework for intrinsic LDMs operating directly on the surfaces of shapes, with the goal of generating high-quality textures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 506 + ], + "type": "text", + "content": "Our method consists of two distinct novel components, serving as the main contributions of our work. The first is a latent representation called Field Latents (FLs). Textures are mapped to tangent vector features at the vertices which are responsible for characterizing the local texture. The choice of tangent vector features over scalars allows for the capture of directional information related to the local texture which we show enables superior quality reconstructions. More generally, FLs offer an effective form of perceptual compression in which a high-resolution texture, analogous to a continuous signal defined over a surface, is mapped to a collection of discrete vector fields taking values at the vertices of a lower-resolution mesh. The second component is a Field Latent Diffusion Model (FLDM), which learns a denoising diffusion process in field latent space. Our FLDMs are built upon field convolutions [35], surface convolution operators designed specifically to process tangent vector features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 289, + 700 + ], + "type": "text", + "content": "Although we outline a general framework for constructing diffusion models for surfaces, inspired by the success of single-image generative models [28, 38, 46, 54], we tailor and deploy our FLDM architecture in the single-textured-mesh setting. We address the problem of generating variations of a given texture on a mesh (Figure 1, left and center), as well as tasks supporting user control such as label-guided generation and inpainting. The single-textured-mesh setting also lets us circumvent the data scarcity issue. Large-scale 3D datasets with high-quality textures have only recently become available [8, 9], with only a fraction of samples across disparate categories possessing complex non-uniform textures. Furthermore, many textured 3D assets that are handcrafted or derived from real-world scans are geometrically or stylistically unique, upon which large-scale models can be challenging to condition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Experiments reveal our models enable flexible synthe" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "sis of high-resolution textures of qualitatively superior fidelity compared to existing single-textured-asset diffusion models, while simultaneously sidestepping the challenges of using 2D-to-3D approaches. Furthermore, both FLs and FLDMs are isometry-equivariant — each commute with distance-preserving shape deformations. As a result, we find that our approach can be used for generative texture transfer, wherein a FLDM trained on a single textured mesh can be sampled on a second similar mesh to texture it in the style of the first (Figure 1, right)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 222, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 547, + 486 + ], + "type": "text", + "content": "Existing generative models that synthesize textured 3D assets typically map content to an internal 2D representation which is computationally preferable to working directly over 3D Euclidean space. Perhaps the most popular approach involves iteratively rendering objects from different viewpoints, enabling the use of pre-trained 2D LDMs (e.g. Stable Diffusion [44]) either indirectly as optimization priors [30, 32, 41, 42] or to directly apply a full denoising process [3, 4, 43, 53]. The former case, pioneered by DreamFusion [41], uses score distillation sampling (SDS), in which a neural radiance field is optimized such that the rendered images appear to be reasonable samples from a pre-trained LDM; in the latter, LDMs with depth conditioning are used to iteratively denoise the rendered images which are aggregatively projected to the texture map. However, SDS-based methods are computationally feasible only with low-resolution LDMs and small NeRFs, and are unable to synthesize fine details. Additionally, approaches that directly denoise and project must contend with artifacts arising from view inconsistencies and synthesized textures may contain unnatural tones or lighting effects as residuals from their image-trained LDM backbone." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 486, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 689 + ], + "type": "text", + "content": "Alternative representational spaces to drive textured 3D synthesis have also been explored. Texture Fields [17, 39] encode textures as 3D neural fields and triplane feature maps have emerged recently as an efficient domain for 3D geometry [56, 58, 63]. In particular, Sin3DM [58] encodes geometry and texture into an implicit triplane latent representation, then trains and samples a 2D DM on the resulting feature maps. However, geometry and texture must first be rasterized to a 3D grid before triplane convolutions can be employed, which can lead to aliasing of high-fidelity textures and geometric details due to memory constraints on the maximal grid resolution. To disentangle geometry from texture, several methods map textures to domains such as the image plane [6, 40, 51, 60] or sphere [7]. Point-UV diffusion [61] trains and samples 2D DMs directly in the UV-atlas, though this approach suffers on meshes with disconnected or fragmented UV mappings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Similar to our approach, a third class of models encode textures as features at the simplices of a mesh. GAN-based" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7954" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "methods Texturify [48] and Mesh2Tex [2] compress textures to attributes at the mesh facets, and differentiable rendering enables adversarial supervision in the image space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 286 + ], + "type": "text", + "content": "Alternatively, Intrinsic Neural Fields (INFs) [27] introduces a latent representation wherein textures are encoded as eigenfunctions of the Laplacian taking values at the mesh vertices. However, features at points on faces are recovered through barycentric interpolation, which we will show leads to a breakdown in reconstruction fidelity at higher compression ratios. Recently, Manifold Diffusion Fields [14, 64] introduce the first fully intrinsic DMs defined over 2D Riemannian manifolds, and are based on a global attention mechanism which aggregates features at the vertices, limiting their applicability to coarse, low-resolution meshes due to complexity constraints. Here, our proposed FLDMs are also fully intrinsic but are built upon locally-supported convolutions defined over the manifold [35], allowing them to scale to higher-resolution meshes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 467 + ], + "type": "text", + "content": "Our single-textured-mesh FLDMs are inspired by recent work in which DMs are trained to capture internal patch distributions from a single image with the goal of generating diverse samples with similar visual content [28, 38, 54]; In turn, these models can trace their lineage to the influential SinGAN architecture [46] and its derivatives [19, 21]. Here, we mirror the approach proposed by SinDiffusion [54] and denoise with a shallow convolutional UNet to control the receptive field and prevent the DM from overfitting to the texture. Similar to our setting, Sin3DM [58] was the first to train LDMs in a single-textured-mesh paradigm, however they take an extrinsic approach to generating geometry and texture, whereas our approach focuses only on synthesizing textures which we show qualitatively are of higher-fidelity than those produced by Sin3DM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 478, + 154, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 154, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 154, + 491 + ], + "type": "text", + "content": "3. Method Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 604 + ], + "type": "text", + "content": "We introduce a framework for intrinsic LDMs operating directly on surfaces. The first component is a latent representation of mesh textures comprised of tangent vector features. At each vertex, the tangent features characterize the texture of a local neighborhood on the surface, and the collection of these features constitute a stack of vector fields we call Field Latents (FLs). The FL space is learned with a locally-supported variational autoencoder [25] (FL-VAE) which is described in Section 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "The second key component is the Field Latent Diffusion Model (FLDM), which learns to denoise a diffusion process in the tangent space of a surface. Denoising networks are constructed with field convolutions (FCs), surface convolution operators acting on tangent vector fields [35]. We extend FCs to interleave scalar embeddings with the tangent vector features, enabling the injection of diffusion timesteps and optional conditioning, such as user-specified labels, into the denoising model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "In practice, FL-VAEs and FLDMs are applied to synthesize new textures from a single textured triangle mesh. We pre-train a single FL-VAE by superimposing planar meshes over a large high-quality image dataset, obtaining a general-purpose latent space which can be used to train FLDMs on arbitrary textured meshes. To do so, a texture is first mapped to distributions in the tangent space at the vertices with the FL-VAE encoder, and an FLDM is trained to iteratively denoise tangent vector latent features. Afterwards, samples from the FLDM can be decoded with the FL-VAE decoder to synthesize new textures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 219, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 434 + ], + "type": "text", + "content": "Importance of Isometry-Equivariance Our FL-VAEs and FLDMs are designed to be equivariant under isometries, i.e. they commute with distance-preserving shape deformations. Unlike images, points on a surface have no canonical orientation. The choice of local coordinates is ambiguous up to an arbitrary rotation, making it impossible to simply adapt standard image-based VAEs and diffusion models to surfaces. However, isometries manifest locally as rotations. Thus, designing our framework to be isometry-equivariant inherently solves the orientation ambiguity problem, enabling consistent, repeatable results. Isometry-equivariance also results in tangible benefits — our models are able to seamlessly reproduce textural details across locally similar areas of meshes. Furthermore, while existing models transfer textures with pointwise maps [11] or generatively via conditioning on a learned token [5, 43], we can simply sample our pre-trained FLDMs on new, similar meshes to texture them in the learned style." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 445, + 389, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 389, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 389, + 457 + ], + "type": "text", + "content": "4. Field Latents" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": "As in Knoppel et al. [26] we associate tangent vectors with complex numbers. For a surface " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": " and point " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "p \\in M" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": ", we assign to the tangent space an arbitrary orthonormal basis such that for any " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in T_pM" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": " we have " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\equiv r e^{i\\theta}" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "r = |\\mathbf{v}|" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": " the direction of " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": " expressed in the frame. We make this convention explicit by denoting " + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mathbb{C}_p \\equiv T_pM" + }, + { + "bbox": [ + 304, + 465, + 545, + 537 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": "To describe the FL-VAE, we require a notion of multidimensional Gaussian noise in the tangent bundle of a surface " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": ". Specifically, we denote the Gaussian distribution over " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": " copies of the tangent bundle " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "TM^{d}" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": ". Samples " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": " are vector fields in " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "TM^{d}" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": " such that at each point " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "p\\in M" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": ", the coefficients of " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\epsilon (p)\\in \\mathbb{C}_p^d" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": " expressed in the orthonormal basis are themselves samples from the " + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 537, + 545, + 632 + ], + "type": "text", + "content": "-dimensional standard normal distribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 352, + 643, + 545, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 643, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 352, + 643, + 545, + 656 + ], + "type": "interline_equation", + "content": "\\epsilon (p) = \\epsilon_ {1} + i \\epsilon_ {2}, \\epsilon_ {1}, \\epsilon_ {2} \\sim \\mathcal {N} (0, I _ {d}). \\tag {1}", + "image_path": "1d56a0e64336e397b5498d514bf940d3452e38a5c21ceab2775db266e152c417.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "In this work we are concerned with isometries " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma : M \\to N" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", which enjoy the property that their push-forwards " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "d\\gamma : TM \\to TN" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", which define a smooth map between tangent spaces, manifest pointwise as rotations. From Equa-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7955" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "tion (1) it is easy to see that " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " is symmetric under such transformations. Thus, it follows that sampling from " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{N}_N(0,I_d)" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " is equivalent to pushing forward samples from " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 128, + 287, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 128, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 87, + 128, + 287, + 158 + ], + "type": "interline_equation", + "content": "\\epsilon^ {\\prime} = \\gamma \\epsilon \\quad \\text {f o r} \\quad \\begin{array}{l} \\epsilon^ {\\prime} \\sim \\mathcal {T N} _ {N} (0, I _ {d}) \\\\ \\epsilon \\sim \\mathcal {T N} _ {M} (0, I _ {d}) \\end{array} \\tag {2}", + "image_path": "e9263ac44053037001b6f53904f69c4d8b29bd6dddad6264dd5039caa94b5db6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 166, + 287, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 166, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 287, + 190 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 166, + 287, + 190 + ], + "type": "inline_equation", + "content": "\\gamma \\epsilon = [d\\gamma \\cdot \\epsilon] \\circ \\gamma^{-1}" + }, + { + "bbox": [ + 47, + 166, + 287, + 190 + ], + "type": "text", + "content": " denoting the standard action of diffeomorphisms on vector fields via left-shifts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": "FL-VAE Encoder The encoder " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " is a network that, for any point " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "p\\in M" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " , takes " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " -dimensional scalar functions " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\psi \\in L^2 (M,\\mathbb{R}^n)" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " (e.g. texture RGB values with " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " restricted to the surrounding geodesic neighborhood " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "B_{p}\\subset" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " to the parameters of " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " independent normal distributions in the tangent space at " + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 205, + 287, + 277 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 284, + 287, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 284, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 102, + 284, + 287, + 317 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {E} _ {p}: L ^ {2} \\left(B _ {p}, \\mathbb {R} ^ {n}\\right)\\rightarrow \\mathbb {C} _ {p} ^ {d} \\times \\mathbb {R} _ {\\geq 0} ^ {d} \\tag {3} \\\\ \\psi \\mapsto \\left(\\mu_ {p} ^ {\\psi}, \\sigma_ {p} ^ {\\psi}\\right) \\\\ \\end{array}", + "image_path": "d5092c1e24fdaf8d35254e2ad9ae4eafe247694ade6571a3fbe41ffccb33fdfe.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": "Here, the means of the distributions " + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\mu_p^\\psi" + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": " are themselves tangent vectors while the standard deviations " + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\sigma_p^\\psi" + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": " are scalars. Latent codes characterizing the local restriction of " + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": " are collections of tangent vectors " + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "inline_equation", + "content": "z_{p}^{\\psi}\\in \\mathbb{C}_{p}^{d}" + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": " drawn from the multivariate normal distribution parameterized by " + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "inline_equation", + "content": "(\\mu_p^\\psi ,\\sigma_p^\\psi)" + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": ". Using the reparameterization trick [25] latent codes can be generated as " + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "inline_equation", + "content": "z_{p}^{\\psi} = \\mu_{p}^{\\psi} + \\sigma_{p}^{\\psi}\\odot \\epsilon (p)" + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 47, + 325, + 287, + 414 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "content": "FL-VAE Decoder The principal advantage in a tangent vector latent representation is that it naturally admits a descriptive coordinate function which can be queried to associate features with neighboring points. The decoder " + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "content": " is designed as a neural field operating over the local parameterization of the surface induced by the logarithm map about a point " + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\log_p:B_p\\to \\mathbb{C}_p" + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\log_p q" + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "content": " encoding the position of " + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "inline_equation", + "content": "q\\in B_{p}" + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "content": " in the tangent space of " + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 426, + 287, + 522 + ], + "type": "text", + "content": ". Formally," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 530, + 287, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 530, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 102, + 530, + 287, + 564 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} _ {p}: \\log_ {p} \\left(B _ {p}\\right) \\times \\mathbb {C} _ {p} ^ {d} \\rightarrow \\mathbb {R} ^ {n} \\tag {4} \\\\ \\left(\\log_ {p} q, z _ {p} ^ {\\psi}\\right) \\mapsto \\widehat {\\psi} _ {p} (q), \\\\ \\end{array}", + "image_path": "587ab42cfbda29bce4f2abd9f63481d878af2638739d970e25f36c512d1c6b9f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\widehat{\\psi}_p" + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "content": " denoting the prediction of " + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "content": " made from the perspective of " + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "content": ". Decoding follows a similar process as proposed in [18]. Specifically, given a tuple " + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "inline_equation", + "content": "(\\log_p q, z_p^\\psi)" + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "content": ", the decoder constructs two features. First, an invariant scalar feature is obtained by vectorizing the upper-triangular part of the Hermitian outer product of the latent code with itself, " + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{vec}_{j \\geq i} \\left(z_p^\\psi \\left[z_p^\\psi\\right]^*\\right) \\in \\mathbb{C}^{\\frac{d(d+1)}{2}}" + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "content": ". Second, a positionally-aware scalar feature " + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "inline_equation", + "content": "c_{pq}^\\psi \\in \\mathbb{C}^d" + }, + { + "bbox": [ + 47, + 573, + 287, + 686 + ], + "type": "text", + "content": " is formed by the natural coordinate function" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 694, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 694, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 129, + 694, + 287, + 712 + ], + "type": "interline_equation", + "content": "c _ {p q} ^ {\\psi} \\equiv \\log_ {p} q \\cdot \\overline {{z _ {p} ^ {\\psi}}}, \\tag {5}", + "image_path": "a7fe8b1087904a39733e71174dfa22590d1419aedc7771d9cfc9caaad4d93467.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "text", + "content": "which corresponds to storing the inner product and determinant of each tangent vector with the position of " + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "text", + "content": " relative to " + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "text", + "content": " given by " + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\log_p q" + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "text", + "content": ". The neural field receives the concatenation of these two features and returns the prediction " + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\widehat{\\psi}_p(q)" + }, + { + "bbox": [ + 305, + 72, + 545, + 124 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "content": "Equivalence Recall we are concerned with constructing a latent representation equivariant under isometries — that for any " + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "inline_equation", + "content": "\\psi \\in L^2(M, \\mathbb{R}^n)" + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "content": " and isometry " + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "inline_equation", + "content": "\\gamma: M \\to N" + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "content": " we have" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 180, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 180, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 391, + 180, + 545, + 198 + ], + "type": "interline_equation", + "content": "\\left[ \\widehat {\\gamma \\psi} \\right] _ {\\gamma (p)} = \\gamma \\widehat {\\psi} _ {p}, \\tag {6}", + "image_path": "b2d4cd578091ec8e460e58af8dce3c40e8ba49fd949193f38a5b9d51881acf71.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\gamma \\psi = \\psi \\circ \\gamma^{-1}" + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "content": " denotes the standard action of diffeomorphisms by left-shifts. Thus, it follows from Equation (2) that a sufficient condition for Equation (6) to hold is if for all " + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\psi \\in L^2(M, \\mathbb{R}^n)" + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "content": ", isometries " + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\gamma : M \\to N" + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "inline_equation", + "content": "p \\in M" + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "content": ", the encoded mean " + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\mu_p^\\psi" + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\sigma_p^\\psi" + }, + { + "bbox": [ + 305, + 205, + 545, + 267 + ], + "type": "text", + "content": " satisfy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 336, + 274, + 545, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 274, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 336, + 274, + 545, + 292 + ], + "type": "interline_equation", + "content": "\\left. d \\gamma \\right| _ {p} \\cdot \\mu_ {p} ^ {\\psi} = \\mu_ {\\gamma (p)} ^ {\\gamma \\psi} \\quad \\text {a n d} \\quad \\sigma_ {p} ^ {\\psi} = \\sigma_ {\\gamma (p)} ^ {\\gamma \\psi}, \\tag {7}", + "image_path": "3b5f78f75ec4b89a661d833b9c56b54ca7253b21ab654817e89633d79992b418.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "inline_equation", + "content": "d\\gamma|_p" + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "text", + "content": " denoting the push-forward of " + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "text", + "content": " in the tangent space at " + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "text", + "content": ", which rotates the local coordinate system. A detailed proof appears in the supplement, sec. 8.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": "Architecture In practice, we discretize a surface " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " by a triangle mesh with vertices " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " and faces " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": ". Textures are scalar functions " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " defined continuously over the faces and vertices, taking values in " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "p \\in V" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": ", we consider its neighborhood to be the surrounding one-ring, the collection of faces in " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " that share " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " as a vertex. The encoder architecture is chosen so as to satisfy the conditions in Equation (7). A fixed number of points " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\{q_i\\}" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " are uniformly sampled in the one-ring neighborhood at which the texture function is evaluated. The resulting scalar features " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\{\\psi(q_i)\\}" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " are interleaved with the logarithms " + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\{\\log_p q_i\\}" + }, + { + "bbox": [ + 304, + 348, + 545, + 538 + ], + "type": "text", + "content": " to construct tangent vector features which are concatenated with a token following [12]. The features are passed to eight successive VN-Transformer layers [1], modified to process tangent vector features. Afterwards, the token is extracted and used to predict the mean and standard deviation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 540, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 659 + ], + "type": "text", + "content": "The neural field comprising the decoder consists of a five-layer real-valued MLP, to which the real and imaginary parts of the complex feature vector are passed. For any point in a triangle, the latent codes at the adjacent vertices provide three distinct predictions which are linearly blended with barycentric interpolation only at inference. This is in contrast to INF [27], wherein linearly blended features are passed to the neural field to make predictions. A comprehensive discussion of the FL-VAE architecture and training objective is provided in section 9.1 in the supplement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 669, + 474, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 474, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 474, + 681 + ], + "type": "text", + "content": "5. Field Latent Diffusion Models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "FLDMs define a noising process on surface vector fields, and learn a denoising diffusion probabilistic model" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7956" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 80, + 264, + 111 + ], + "blocks": [ + { + "bbox": [ + 67, + 80, + 264, + 111 + ], + "lines": [ + { + "bbox": [ + 67, + 80, + 264, + 111 + ], + "spans": [ + { + "bbox": [ + 67, + 80, + 264, + 111 + ], + "type": "table", + "html": "
FL-VAE22.38 (20.59)0.51 (0.83)1.02 (1.81)
FL-VAE (Bary.)21.33 (19.61)0.66 (1.01)1.31 (2.03)
INFs [27]18.86 (16.45)1.16 (1.64)2.15 (2.73)
", + "image_path": "c94607f7ffe0c85fc2d05ef3dd95e2c4ea4287e6166cb6507c43602592e69443.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 113, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 113, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 113, + 287, + 168 + ], + "type": "text", + "content": "Table 1. Texture reconstruction quality on meshes from the Google Scanned Objects dataset [13]. Reconstructions are evaluated on high-res (30K vertices) and low-res (5K vertices, in parenthesis) meshings of the objects. Metrics are computed in the texture atlases, with DSSIM and LPIPS scaled by a factor of 10." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 172, + 287, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 208 + ], + "type": "text", + "content": "(DDPM) [22] using an equivariant surface network. In principle they can be applied to any vector data on surfaces, and here we use them to learn feature distributions in FL space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": "Latent vector fields " + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "inline_equation", + "content": "Z \\in TM^{d}" + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": " can be produced by compressing a texture " + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\psi \\in L^{2}(M, \\mathbb{R}^{3})" + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": " to a collection of tangent vector features at each point with the FL-VAE encoder as in Equation (3) and sampling from the distributions. The forward process progressively adds noise in the tangent space, dictated by a monotonically decreasing schedule " + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{t}\\}_{0 \\leq t \\leq T}" + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\alpha_{0} = 1" + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\alpha_{T} = 0" + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": ". The noised vector fields can be expressed at an arbitrary time " + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 209, + 287, + 316 + ], + "type": "text", + "content": " in the discretized forward process with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 321, + 287, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 321, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 94, + 321, + 287, + 336 + ], + "type": "interline_equation", + "content": "Z _ {t} = \\sqrt {\\alpha_ {t}} Z + \\sqrt {1 - \\alpha_ {t}} \\epsilon \\in T M ^ {d}, \\tag {8}", + "image_path": "fc2633d1f988e835d194fa02467dda5664b0c77cf67b4e34236ad31516d1e6b4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 342, + 225, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 225, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 225, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 342, + 225, + 354 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 47, + 342, + 225, + 354 + ], + "type": "text", + "content": " as in Equation (1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "text", + "content": "The denoising function " + }, + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "text", + "content": " is a surface network trained to reverse the forward process. Specifically, it is a learnable map on the space of latent vector fields, conditioned on timesteps " + }, + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "text", + "content": " and optional scalar embeddings " + }, + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\rho \\in L^{2}(M,\\mathbb{R}^{m})" + }, + { + "bbox": [ + 47, + 354, + 287, + 414 + ], + "type": "text", + "content": ", with the goal of predicting the added noise," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 81, + 419, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 419, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 81, + 419, + 287, + 434 + ], + "type": "interline_equation", + "content": "\\varepsilon : T M ^ {d} \\times \\mathbb {R} _ {\\geq 0} \\times L ^ {2} (M, \\mathbb {R} ^ {m}) \\rightarrow T M ^ {d}, \\tag {9}", + "image_path": "de0c6eaade1a5d2a5e0f6aa51acc63219742d01b2be71d24f7b8fb6296326ff8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 441, + 152, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 152, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 152, + 452 + ], + "type": "text", + "content": "trained subject to the loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 457, + 287, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 287, + 473 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {F L D M}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {T N} _ {M} (0, I _ {d}), 0 \\leq t \\leq T} \\| \\epsilon - \\varepsilon \\left(Z _ {t}, t, \\rho\\right) \\| _ {2} ^ {2}. \\tag {10}", + "image_path": "7fad25d389ff4f17e726f1a8f2d60cae13f5323bf28d164a210c08803b9e653f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "text", + "content": "At inference, new latent vector fields are generated following the iterative DDPM denoising process. Beginning with representing the fully-noised vector fields at step " + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "text", + "content": " by a sample " + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\tilde{Z} \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\tilde{Z}_T = \\tilde{Z}" + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "text", + "content": ", the trained denoising function " + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 47, + 479, + 287, + 550 + ], + "type": "text", + "content": " is used to estimate the previous step in the forward process via the relation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 556, + 286, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 286, + 586 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widetilde {Z} _ {t - 1} = C _ {1} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\widetilde {Z} _ {t} + C _ {2} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\varepsilon \\left(\\widetilde {Z} _ {t}, t, \\rho\\right) \\tag {11} \\\\ + C _ {3} \\left(\\alpha_ {t}, \\alpha_ {t - 1}\\right) \\epsilon , \\\\ \\end{array}", + "image_path": "7e3ef75e199d480bcab0527c192051ab4a88b89207009a86bb26544b88250b41.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon = 0" + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "text", + "content": " otherwise, and " + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "inline_equation", + "content": "C_i(\\alpha_t,\\alpha_{t - 1})\\in \\mathbb{R}" + }, + { + "bbox": [ + 47, + 592, + 287, + 617 + ], + "type": "text", + "content": " as defined in 9.1 of the supplement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "text", + "content": "Equivalence From the equivalence of the distributions " + }, + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{N}_M(0,I_d)" + }, + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{N}_N(0,I_d)" + }, + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "text", + "content": " under isometries as in Equation (2), it follows that for any isometry " + }, + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\gamma :M\\to N" + }, + { + "bbox": [ + 47, + 630, + 287, + 690 + ], + "type": "text", + "content": " a sufficient condition for both the forward and reverse processes in Equations (8) and (11) to satisfy" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 82, + 695, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 695, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 82, + 695, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\gamma Z _ {t} = \\left[ \\gamma Z \\right] _ {t} \\quad \\text {a n d} \\quad \\gamma \\widetilde {Z} _ {t} = \\left[ \\gamma \\widetilde {Z} \\right] _ {t}, \\tag {12}", + "image_path": "56841e09bf42f19a401225ee57ba4d593c0347fe315689cd253c53118d359d35.jpg" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 87, + 365, + 123 + ], + "blocks": [ + { + "bbox": [ + 307, + 87, + 365, + 123 + ], + "lines": [ + { + "bbox": [ + 307, + 87, + 365, + 123 + ], + "spans": [ + { + "bbox": [ + 307, + 87, + 365, + 123 + ], + "type": "image", + "image_path": "63aa0937b3f29c8b94195d24c6d5cee7638a21bdca27d7c809e816a3e9a4036f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 313, + 125, + 365, + 178 + ], + "blocks": [ + { + "bbox": [ + 313, + 125, + 365, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 125, + 365, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 365, + 178 + ], + "type": "image", + "image_path": "ae336f6b2f726b5de02550b5ebf7e68b2210b329ccbafdf06f88ee2138c31873.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 367, + 88, + 423, + 123 + ], + "blocks": [ + { + "bbox": [ + 367, + 88, + 423, + 123 + ], + "lines": [ + { + "bbox": [ + 367, + 88, + 423, + 123 + ], + "spans": [ + { + "bbox": [ + 367, + 88, + 423, + 123 + ], + "type": "image", + "image_path": "70388add48c7ab22cf2ca238ca263be278195e2ada3d3339a86d2668c00135a9.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 371, + 125, + 425, + 179 + ], + "blocks": [ + { + "bbox": [ + 371, + 125, + 425, + 179 + ], + "lines": [ + { + "bbox": [ + 371, + 125, + 425, + 179 + ], + "spans": [ + { + "bbox": [ + 371, + 125, + 425, + 179 + ], + "type": "image", + "image_path": "44f08460a5951f34d3f96f17cfa280c9cd3ad7031638c69d9321225cddf6df43.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 426, + 88, + 482, + 123 + ], + "blocks": [ + { + "bbox": [ + 426, + 88, + 482, + 123 + ], + "lines": [ + { + "bbox": [ + 426, + 88, + 482, + 123 + ], + "spans": [ + { + "bbox": [ + 426, + 88, + 482, + 123 + ], + "type": "image", + "image_path": "6a688597a5ff870651a05b2f0e38e1fb2ad063e274e5cbfaf8c3d580c4dbc04f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 485, + 88, + 542, + 123 + ], + "blocks": [ + { + "bbox": [ + 485, + 88, + 542, + 123 + ], + "lines": [ + { + "bbox": [ + 485, + 88, + 542, + 123 + ], + "spans": [ + { + "bbox": [ + 485, + 88, + 542, + 123 + ], + "type": "image", + "image_path": "a342393567a96786c96eb29cfafa83b3faf22ef63c91d4598a030a12ddf3a8b2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 427, + 125, + 484, + 179 + ], + "blocks": [ + { + "bbox": [ + 427, + 125, + 484, + 179 + ], + "lines": [ + { + "bbox": [ + 427, + 125, + 484, + 179 + ], + "spans": [ + { + "bbox": [ + 427, + 125, + 484, + 179 + ], + "type": "image", + "image_path": "eaebf572d728ea734cb7aa96bc0801ca31832af8099aab4cb30dabc4899bbd83.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 490, + 125, + 543, + 179 + ], + "blocks": [ + { + "bbox": [ + 490, + 125, + 543, + 179 + ], + "lines": [ + { + "bbox": [ + 490, + 125, + 543, + 179 + ], + "spans": [ + { + "bbox": [ + 490, + 125, + 543, + 179 + ], + "type": "image", + "image_path": "02579870a66ff82c2480b1755db40022a86861c59411cf9a2fde4ad379bc3cdd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 306, + 185, + 363, + 214 + ], + "blocks": [ + { + "bbox": [ + 306, + 185, + 363, + 214 + ], + "lines": [ + { + "bbox": [ + 306, + 185, + 363, + 214 + ], + "spans": [ + { + "bbox": [ + 306, + 185, + 363, + 214 + ], + "type": "image", + "image_path": "d1deb7074a634d3fddfa7522b18a356eb8e7a04524c3b68c5d37aa735cd2c0fd.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 309, + 215, + 362, + 266 + ], + "blocks": [ + { + "bbox": [ + 309, + 215, + 362, + 266 + ], + "lines": [ + { + "bbox": [ + 309, + 215, + 362, + 266 + ], + "spans": [ + { + "bbox": [ + 309, + 215, + 362, + 266 + ], + "type": "image", + "image_path": "a7640465f10645669e6687715d83009fc5e35b4a45aa39ebbbd7649fab780506.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 272, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 305, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 545, + 327 + ], + "type": "text", + "content": "Figure 2. Visual comparison of textures compressed and reconstructed with the FL-VAE and INF [27] on 30K vertex meshes. Compared to barycentric coordinates, our proposed logarithmic coordinate function more richly extends latent features across the mesh, enabling the reconstruction of finer details. Zoom in to view." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 367, + 185, + 424, + 215 + ], + "blocks": [ + { + "bbox": [ + 367, + 185, + 424, + 215 + ], + "lines": [ + { + "bbox": [ + 367, + 185, + 424, + 215 + ], + "spans": [ + { + "bbox": [ + 367, + 185, + 424, + 215 + ], + "type": "image", + "image_path": "aeb2230304ddfa3e2618256dcb1c45c272da83db316f064b7b9062e28cdbd379.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 370, + 216, + 423, + 266 + ], + "blocks": [ + { + "bbox": [ + 370, + 216, + 423, + 266 + ], + "lines": [ + { + "bbox": [ + 370, + 216, + 423, + 266 + ], + "spans": [ + { + "bbox": [ + 370, + 216, + 423, + 266 + ], + "type": "image", + "image_path": "7b13595d074786f40764e31b3c1710f887a75c01b7084d7abce7a0db15faf0a4.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 428, + 185, + 485, + 215 + ], + "blocks": [ + { + "bbox": [ + 428, + 185, + 485, + 215 + ], + "lines": [ + { + "bbox": [ + 428, + 185, + 485, + 215 + ], + "spans": [ + { + "bbox": [ + 428, + 185, + 485, + 215 + ], + "type": "image", + "image_path": "0aa349d9b15df4819afdad900a8967ad32e95a4917cbcee429b09deca8e68c7f.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 430, + 216, + 484, + 266 + ], + "blocks": [ + { + "bbox": [ + 430, + 216, + 484, + 266 + ], + "lines": [ + { + "bbox": [ + 430, + 216, + 484, + 266 + ], + "spans": [ + { + "bbox": [ + 430, + 216, + 484, + 266 + ], + "type": "image", + "image_path": "c110f4efea2af05364ededd69d3518732e2bc890f4274d8895ff389d6e89053b.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 488, + 185, + 545, + 215 + ], + "blocks": [ + { + "bbox": [ + 488, + 185, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 488, + 185, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 488, + 185, + 545, + 215 + ], + "type": "image", + "image_path": "a53dd0f6ade80852fd2ef99fe245a55752d305902d47d930b7ecacd64354e073.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 488, + 216, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 488, + 216, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 488, + 216, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 488, + 216, + 545, + 266 + ], + "type": "image", + "image_path": "f0859b85236abd84b97d9f4baf2e643916bb89e4fdabe9356946babec1031dc9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "inline_equation", + "content": "0 \\leq t \\leq T" + }, + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "text", + "content": " is if the denoising network " + }, + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "text", + "content": " in Equation (9) has the property that for any vector field " + }, + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "inline_equation", + "content": "Z \\in TM^{d}" + }, + { + "bbox": [ + 305, + 335, + 545, + 360 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 370, + 370, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 370, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 370, + 370, + 545, + 384 + ], + "type": "interline_equation", + "content": "\\gamma \\varepsilon (Z, t, \\rho) = \\varepsilon (\\gamma Z, t, \\gamma \\rho). \\tag {13}", + "image_path": "f59992869a305645192ea2c2e1e5645d4cf640d06682f2c59cf297185f9dc99d.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 394, + 540, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 540, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 540, + 407 + ], + "type": "text", + "content": "A detailed proof is provided in sec. 8.2 of the supplement." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 305, + 425, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 472 + ], + "type": "text", + "content": "Denoising with Field Convolutions To satisfy the condition in Equation (12), our denoising network is constructed with field convolutions (FC) [35], which belong to a class of models that can operate on tangent vector fields [45, 57]." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": "FCs convolve vector fields " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "Z \\in TM^{C}" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "C' \\times C" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": " filter banks " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{c'c} \\in L^{2}(\\mathbb{C},\\mathbb{C})" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": ", returning vector fields " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "Z * \\mathbf{f} \\in TM^{C'}" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": ". In conventional diffusion models, features inside the denoising network are conditioned on timesteps " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": " via an additive embedding applied inside the convolution blocks. Unfortunately, no direct analogue exists for vector field features on surfaces — additive embeddings break equivariance and we find that multiplicative embeddings destabilize training. Thus, to enable conditioning of our denoising network " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": " on both timesteps " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": " and optional user-input features " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\rho \\in L^{2}(M,\\mathbb{R}^{m})" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": ", we inject embeddings directly into convolutions by extending FCs to convolve vector fields with filters " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{c'c} \\in L^{2}(\\mathbb{C} \\times \\mathbb{R}^{e},\\mathbb{C})" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 305, + 473, + 545, + 653 + ], + "type": "text", + "content": " the embedding dimension. Critically, this formulation preserves equivariance without sacrificing stability." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Following [54], the denoising network " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " takes the form of a shallow two-level U-Net but with field convolutions; the relatively small size of the network's receptive field allows distributions of latent features to be learned without overfitting to the single textural example [28, 54, 58]." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 69, + 71, + 97, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 97, + 79 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 97, + 79 + ], + "type": "text", + "content": "Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 71, + 165, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 71, + 165, + 79 + ], + "spans": [ + { + "bbox": [ + 138, + 71, + 165, + 79 + ], + "type": "text", + "content": "PSNR↑" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 71, + 215, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 71, + 215, + 79 + ], + "spans": [ + { + "bbox": [ + 184, + 71, + 215, + 79 + ], + "type": "text", + "content": "DSSIM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 71, + 257, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 71, + 257, + 79 + ], + "spans": [ + { + "bbox": [ + 228, + 71, + 257, + 79 + ], + "type": "text", + "content": "LPIPS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 309, + 73, + 358, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 73, + 358, + 83 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 358, + 83 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 376, + 73, + 411, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 73, + 411, + 83 + ], + "spans": [ + { + "bbox": [ + 376, + 73, + 411, + 83 + ], + "type": "text", + "content": "FL-VAE" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 425, + 73, + 484, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 73, + 484, + 83 + ], + "spans": [ + { + "bbox": [ + 425, + 73, + 484, + 83 + ], + "type": "text", + "content": "FL-VAE (Bary.)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 497, + 73, + 533, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 497, + 73, + 533, + 83 + ], + "spans": [ + { + "bbox": [ + 497, + 73, + 533, + 83 + ], + "type": "text", + "content": "INFs [27]" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7957" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 63, + 96, + 133 + ], + "blocks": [ + { + "bbox": [ + 55, + 63, + 96, + 133 + ], + "lines": [ + { + "bbox": [ + 55, + 63, + 96, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 63, + 96, + 133 + ], + "type": "image", + "image_path": "89286772ec9ca37e6ef226617e7584df1bb16c2c37964c67f7070548d8051ac5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 136, + 97, + 189 + ], + "blocks": [ + { + "bbox": [ + 54, + 136, + 97, + 189 + ], + "lines": [ + { + "bbox": [ + 54, + 136, + 97, + 189 + ], + "spans": [ + { + "bbox": [ + 54, + 136, + 97, + 189 + ], + "type": "image", + "image_path": "6613f4a9698ceb5e75e2b08d786bff7cd1003b115a6eea1e36dbdfcadfa49c92.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 546, + 243 + ], + "type": "text", + "content": "Figure 3. Visual comparison of unconditionally (label-free) generated textures from FLDMs and Sin3DM [58]. Our FLDM's isometry-equivariant construction allows for replication of textural details across locally similar regions. In contrast, Sin3DM's extrinsic approach associates textural features with 3D space; Synthesized details appear as repetitions or extrusions of previous patterns along the major axes of the mesh, and we observe that novel textures cannot be created without modifying geometry. Zoom in to compare." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 104, + 78, + 141, + 133 + ], + "blocks": [ + { + "bbox": [ + 131, + 64, + 160, + 74 + ], + "lines": [ + { + "bbox": [ + 131, + 64, + 160, + 74 + ], + "spans": [ + { + "bbox": [ + 131, + 64, + 160, + 74 + ], + "type": "text", + "content": "FLDM" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 104, + 78, + 141, + 133 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 141, + 133 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 141, + 133 + ], + "type": "image", + "image_path": "5283b246c8ecd84f70136fc56a568a6d95fa51b11153d0059994c15615374ed9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 80, + 188, + 133 + ], + "blocks": [ + { + "bbox": [ + 151, + 80, + 188, + 133 + ], + "lines": [ + { + "bbox": [ + 151, + 80, + 188, + 133 + ], + "spans": [ + { + "bbox": [ + 151, + 80, + 188, + 133 + ], + "type": "image", + "image_path": "8604ae5206e9c1fe87abc7c2440a3084246b6286888b675292ebced34502433f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 102, + 135, + 144, + 189 + ], + "blocks": [ + { + "bbox": [ + 102, + 135, + 144, + 189 + ], + "lines": [ + { + "bbox": [ + 102, + 135, + 144, + 189 + ], + "spans": [ + { + "bbox": [ + 102, + 135, + 144, + 189 + ], + "type": "image", + "image_path": "5e4923766cf947fc7d7d9745ed3081f6f0fed49b82fe75ee42704ca32893a132.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 150, + 136, + 190, + 190 + ], + "blocks": [ + { + "bbox": [ + 150, + 136, + 190, + 190 + ], + "lines": [ + { + "bbox": [ + 150, + 136, + 190, + 190 + ], + "spans": [ + { + "bbox": [ + 150, + 136, + 190, + 190 + ], + "type": "image", + "image_path": "16e6d47d14b2e30f986e94b85867a4ab77ce80380a652009699f5462c6f6922c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 200, + 79, + 236, + 133 + ], + "blocks": [ + { + "bbox": [ + 216, + 64, + 266, + 74 + ], + "lines": [ + { + "bbox": [ + 216, + 64, + 266, + 74 + ], + "spans": [ + { + "bbox": [ + 216, + 64, + 266, + 74 + ], + "type": "text", + "content": "Sin3DM [58]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 79, + 236, + 133 + ], + "lines": [ + { + "bbox": [ + 200, + 79, + 236, + 133 + ], + "spans": [ + { + "bbox": [ + 200, + 79, + 236, + 133 + ], + "type": "image", + "image_path": "a3fd7fbbca6319ae00f4572a50b323e97db54259e3d629ba5e944245fcbd1612.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 246, + 80, + 282, + 133 + ], + "blocks": [ + { + "bbox": [ + 246, + 80, + 282, + 133 + ], + "lines": [ + { + "bbox": [ + 246, + 80, + 282, + 133 + ], + "spans": [ + { + "bbox": [ + 246, + 80, + 282, + 133 + ], + "type": "image", + "image_path": "165e4738813c0d860ab3141560a148bee157967e3c1f00a5d0772defbcb31934.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 195, + 136, + 241, + 189 + ], + "blocks": [ + { + "bbox": [ + 195, + 136, + 241, + 189 + ], + "lines": [ + { + "bbox": [ + 195, + 136, + 241, + 189 + ], + "spans": [ + { + "bbox": [ + 195, + 136, + 241, + 189 + ], + "type": "image", + "image_path": "547c9e9021a7d0de0267627142322c2941bfcfe4a73e758f2447ae8ca61bc0af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 244, + 137, + 284, + 188 + ], + "blocks": [ + { + "bbox": [ + 244, + 137, + 284, + 188 + ], + "lines": [ + { + "bbox": [ + 244, + 137, + 284, + 188 + ], + "spans": [ + { + "bbox": [ + 244, + 137, + 284, + 188 + ], + "type": "image", + "image_path": "0bbb7cf194d616ba200e2b29d659952f2b9cc4f5dcc2242b83c05ffd6b8a39d1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 294, + 80, + 332, + 133 + ], + "blocks": [ + { + "bbox": [ + 302, + 64, + 323, + 74 + ], + "lines": [ + { + "bbox": [ + 302, + 64, + 323, + 74 + ], + "spans": [ + { + "bbox": [ + 302, + 64, + 323, + 74 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 80, + 332, + 133 + ], + "lines": [ + { + "bbox": [ + 294, + 80, + 332, + 133 + ], + "spans": [ + { + "bbox": [ + 294, + 80, + 332, + 133 + ], + "type": "image", + "image_path": "e8ac167370b7d8a63d65bd087caad30909aca91346ff76aeb9be7f969fec8df7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 347, + 80, + 384, + 133 + ], + "blocks": [ + { + "bbox": [ + 375, + 64, + 403, + 73 + ], + "lines": [ + { + "bbox": [ + 375, + 64, + 403, + 73 + ], + "spans": [ + { + "bbox": [ + 375, + 64, + 403, + 73 + ], + "type": "text", + "content": "FLDM" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 80, + 384, + 133 + ], + "lines": [ + { + "bbox": [ + 347, + 80, + 384, + 133 + ], + "spans": [ + { + "bbox": [ + 347, + 80, + 384, + 133 + ], + "type": "image", + "image_path": "f5b1aa814fca902e7c7f692954e09aa5dc48a6d306295fc1a047a061e0a8ed15.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 399, + 80, + 435, + 133 + ], + "blocks": [ + { + "bbox": [ + 399, + 80, + 435, + 133 + ], + "lines": [ + { + "bbox": [ + 399, + 80, + 435, + 133 + ], + "spans": [ + { + "bbox": [ + 399, + 80, + 435, + 133 + ], + "type": "image", + "image_path": "c07ecad3d77d567a7c0e67950921e1fc3a4f1dd8be453f61c48b9f5c0a47181b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 449, + 80, + 487, + 133 + ], + "blocks": [ + { + "bbox": [ + 468, + 64, + 517, + 74 + ], + "lines": [ + { + "bbox": [ + 468, + 64, + 517, + 74 + ], + "spans": [ + { + "bbox": [ + 468, + 64, + 517, + 74 + ], + "type": "text", + "content": "Sin3DM [58]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 80, + 487, + 133 + ], + "lines": [ + { + "bbox": [ + 449, + 80, + 487, + 133 + ], + "spans": [ + { + "bbox": [ + 449, + 80, + 487, + 133 + ], + "type": "image", + "image_path": "55d5c396fe30615aafcf0e4e07a9e8635292d28063d61ed73ca0ffa841d556ca.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 501, + 80, + 539, + 133 + ], + "blocks": [ + { + "bbox": [ + 501, + 80, + 539, + 133 + ], + "lines": [ + { + "bbox": [ + 501, + 80, + 539, + 133 + ], + "spans": [ + { + "bbox": [ + 501, + 80, + 539, + 133 + ], + "type": "image", + "image_path": "721c0cbd2fff4f86060934de71bd4e97803189172b4f2ca664fc7d87bbbf0c67.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 286, + 137, + 441, + 188 + ], + "blocks": [ + { + "bbox": [ + 286, + 137, + 441, + 188 + ], + "lines": [ + { + "bbox": [ + 286, + 137, + 441, + 188 + ], + "spans": [ + { + "bbox": [ + 286, + 137, + 441, + 188 + ], + "type": "image", + "image_path": "44cd2fd89375eba06d65280e4f46d1101ce39261a2f01efcaf59966169f69eae.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 442, + 137, + 492, + 187 + ], + "blocks": [ + { + "bbox": [ + 442, + 137, + 492, + 187 + ], + "lines": [ + { + "bbox": [ + 442, + 137, + 492, + 187 + ], + "spans": [ + { + "bbox": [ + 442, + 137, + 492, + 187 + ], + "type": "image", + "image_path": "95944ef82cf4e16994f6b8b162a6909666b1428bc680c4a8d2b217b98057a67d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 493, + 137, + 544, + 187 + ], + "blocks": [ + { + "bbox": [ + 493, + 137, + 544, + 187 + ], + "lines": [ + { + "bbox": [ + 493, + 137, + 544, + 187 + ], + "spans": [ + { + "bbox": [ + 493, + 137, + 544, + 187 + ], + "type": "image", + "image_path": "fbcbf17b062b7b2ccfcfa262582c79117500b87a4e788a928742996988618dcc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 53, + 262, + 180, + 312 + ], + "blocks": [ + { + "bbox": [ + 53, + 262, + 180, + 312 + ], + "lines": [ + { + "bbox": [ + 53, + 262, + 180, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 180, + 312 + ], + "type": "table", + "html": "
MethodSIFID ↓LPIPS ↑
FLDM3.271.15
Sample (left)0.710.94
Sin3DM [58]6.582.20
Sample (right)2.422.91
", + "image_path": "6e5f800e10dbcab0c897c7e353e75dd359740064b87a61e753a85faf9317a5ea.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 195, + 255, + 230, + 323 + ], + "blocks": [ + { + "bbox": [ + 195, + 255, + 230, + 323 + ], + "lines": [ + { + "bbox": [ + 195, + 255, + 230, + 323 + ], + "spans": [ + { + "bbox": [ + 195, + 255, + 230, + 323 + ], + "type": "image", + "image_path": "a6ed50c83dd3bed48f94ef1102ecc92854b7ffa4cf2b3dcd284cd2798ac4d0a6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 240, + 255, + 276, + 322 + ], + "blocks": [ + { + "bbox": [ + 240, + 255, + 276, + 322 + ], + "lines": [ + { + "bbox": [ + 240, + 255, + 276, + 322 + ], + "spans": [ + { + "bbox": [ + 240, + 255, + 276, + 322 + ], + "type": "image", + "image_path": "a17bb953d8d301a11e5657a74f90b11831f0b50e0dbac5324bdd3d56a0ec4363.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 331, + 287, + 419 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 419 + ], + "type": "text", + "content": "Table 2. Fidelity and diversity of unconditionally generated textures trained on selected Objaverse [8] and Scanned Objects [13] meshes. SIFID and LPIPS are scaled by factors of " + }, + { + "bbox": [ + 46, + 331, + 287, + 419 + ], + "type": "inline_equation", + "content": "10^{5}" + }, + { + "bbox": [ + 46, + 331, + 287, + 419 + ], + "type": "text", + "content": " and 10, respectively. Unlike our FLDMs, Sin3DM synthesizes new geometry which can inflate diversity scores. As an example, the individual metrics are shown for the FLDM and Sin3DM samples on the right. The latter attains a high LPIPS score, but exhibits poor textural diversity and quality." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 427, + 128, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 128, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 128, + 441 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "content": "Pre-training the FL-VAE It is important to note that the FL-VAE proposed in Section 4 only interfaces with the mesh geometry through its local flattening in the image of the logarithm map. Thus, the FL-VAE operates independently of the local curvature, and a model pre-trained' on flat textured meshes can be used to encode and decode textures on arbitrary 3D shapes." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "In all of our experiments we use the same FL-VAE with a " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "d = 8" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " dimensional latent space, pre-trained on the OpenImagesV4 dataset [29]. Specifically, we generate 1K different planar meshes containing 10K vertices each. Each sample " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " image is overlaid by a randomly chosen triangulation, with the image serving as the \"texture\" defined over the mesh. The FL-VAE is then trained with the sum of reconstruction loss and KL divergence using the Adam [24] optimizer. We deliberately train the FL-VAE on coarser triangulations than we expect during inference on textured meshes. This forces the model to reconstruct high frequencies during training, which we find increases robustness to sampling density and triangulation quality, in addition to improving the fidelity of reconstructed FLDM samples." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "text", + "content": "Training single-textured-mesh FLDMs All diffusion experiments share the same training regime. During preprocessing, we create 500 copies of a given mesh each with 30K vertices and different triangulations to prevent the denoising network from learning the connectivity. Each copy shares the same UV-atlas as the original and the texture is mapped to vector field latent features at the vertices using the pre-trained FL-VAE encoder. FLDMs have a max timestep of " + }, + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 304, + 264, + 545, + 396 + ], + "type": "text", + "content": " and noise is added on a linear schedule. The denoising network is trained subject to the loss in Equation (10) using the Adam [24] optimizer." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 406, + 523, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 523, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 523, + 418 + ], + "type": "text", + "content": "6.1. Texture Compression and Reconstruction" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "text", + "content": "First we quantitatively evaluate the capacity of our proposed FL-VAE for compressing and reconstructing textures. We compare FL-VAEs against two other approaches: the state-of-the-art INF [27] and a modified version of FL-VAEs serving as an ablation. INFs train a neural field per texture and predicts the texture values using the values of the Laplacian eigenfunctions extended over triangles through barycentric interpolation. Similarly, the modified FL-VAE (denoted FL-VAE Barycentric) omits our proposed logarithmic coordinate function in Equation (5) in favor of linearly interpolating the values of the invariant outer product features across triangles. It is otherwise identical to our proposed approach, including pre-training methodology. Thus, this also serves to compare the efficacy of our proposed coordinate function against barycentric interpolation in extending vertex features continuously over the mesh." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "To evaluate, we select 16 objects exhibiting a range of complex textures with high-frequency details from the Google Scanned Objects dataset [13]. Two versions of the dataset are used, with objects remeshed to have " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "30\\mathrm{K}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " (high-res) and " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "5\\mathrm{K}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " (low-res) vertices, respectively. The single pretrained FL-VAE is used to map the textures to distributions over the tangent space, and the mean is used to sample the reconstructed texture at the pixel indices in the texture atlas." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7958" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 62, + 287, + 137 + ], + "blocks": [ + { + "bbox": [ + 48, + 62, + 287, + 137 + ], + "lines": [ + { + "bbox": [ + 48, + 62, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 62, + 287, + 137 + ], + "type": "image", + "image_path": "3ccdf64a5f1b102d6d0808145bdd38b6605576cf305c631992e128d922208115.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 147, + 288, + 180 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 288, + 180 + ], + "type": "text", + "content": "Figure 4. Label-guided generation. FLDMs can be conditioned on a subjective user-input labeling which generated textures will reflect. See Figure 1 for more examples. Zoom in to view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 209 + ], + "type": "text", + "content": "A separate INF neural field is trained for each mesh, using the specified regime for high-res reconstruction [27]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 210, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 448 + ], + "type": "text", + "content": "Following INFs, we evaluate the quality of the reconstructed textures using the average PSNR, DSSIM [55], and LPIPS [62] metrics computed in texture atlases. The results are reported in Table 1, with several examples of reconstructions shown in Figure 2. Our proposed FL-VAE achieves the best results across all metrics at both resolutions, followed by the barycentric FL-VAE and INFs. The significant gap in performance between the barycentric FL-VAE and INFs is likely due to the reliance of the latter the on the Laplacian eigenfunctions as input features. Areas of fine textural detail may not necessarily overlap with regions where the eigenfunctions exhibit high variability, potentially contributing to the \"raggedness\" observed in INF reconstructions. Furthermore, the superior performance of our proposed FL-VAE relative to its barycentric counterpart suggests that our logarithmic coordinate function in Equation (5) provides a richer interpolant over the mesh faces. This directly increases the representational capacity of the latent features, enabling the reconstruction of finer textural details with the same number of features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 456, + 253, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 253, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 253, + 468 + ], + "type": "text", + "content": "6.2. Unconditional (Label-Free) Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 688 + ], + "type": "text", + "content": "Label-free generation can be performed by training and sampling FLDMs conditioned only on the diffusion timescale (i.e. " + }, + { + "bbox": [ + 46, + 474, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\rho = 0" + }, + { + "bbox": [ + 46, + 474, + 287, + 688 + ], + "type": "text", + "content": "). We compare against Sin3DM [58] which to our knowledge is the only prior generative model designed to operate in a single-textured-mesh paradigm. We follow their proposed evaluation protocol, and train FLDMs and Sin3DM on 10 assets from the Objaverse [8] and Scanned Objects [13] datasets selected for texture complexity and diversity. For each asset, 64 textures are sampled from the trained models, which are rendered on the mesh from 24 different views. Fidelity is measured with the mean SIFID (Single Image Fréchet Inception Distance) [46] pairwise between the renderings and those of the input mesh from the same view; the mean LPIPS [62] between renders of samples from the same viewpoint measures diversity. We note that unlike FLDMs, which only synthesize new textures, Sin3DM synthesizes both textures and geometries, with the latter potentially inflating LPIPS scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The results are shown in Table 2, Figure 3, and in the supplement (Figure 8). Compared to Sin3DM, our FLDMs" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 62, + 543, + 137 + ], + "blocks": [ + { + "bbox": [ + 309, + 62, + 543, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 62, + 543, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 62, + 543, + 137 + ], + "type": "image", + "image_path": "5b61e246cc5e841824e8907e64d220456dc4a5b6718c26610993c0e1f1b1a0ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 147, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 305, + 147, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 147, + 545, + 180 + ], + "type": "text", + "content": "Figure 5. Inpainting with FLDMs. The input texture is preserved in the masked regions and new content is synthesized elsewhere with agreement on the boundaries. Zoom in to view." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "content": "are able to synthesize higher-fidelity textures which is likely due in part to the former's rasterization of textured geometry to a 3D grid before encoding to the latent space, potentially aliasing fine details. More generally, we observe that by operating over an extrinsic representation of textured geometry, Sin3DM intertwines texture with a mesh's 3D embedding. Thus, new textural details appear as repetitions or extrusions of earlier patterns along the major axes of the models, as seen in the vase, snake, and sculpture samples in Figure 3. Furthermore, we observe that texture and position are so strongly linked that Sin3DM is unable to produce significant textural alterations without correspondingly modifying the geometry, as seen in the vase, snake, and lantern samples. Here, the advantage of our isometry-equivariant formulation becomes clear as our FLDMs can effectively replicate textural details across areas of the mesh that are locally similar, regardless of the relative configuration of these regions in the 3D embedding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 407, + 449, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 449, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 449, + 418 + ], + "type": "text", + "content": "6.3. Label-Guided Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 653 + ], + "type": "text", + "content": "Textured objects often exhibit distinct content specific to certain regions of the mesh which may not be geometrically unique; for example, in Figure 1 regions on the sole, body, and inside of the shoe lack distinguishing geometric features, as is the case with the eyes and mouth of the skull and the decal on the bottle. In such cases, a user may find it desirable for synthesized textures to reflect a subjective distribution of content on the input mesh, which we facilitate by extending our denoising model to incorporate optional conditioning from user-designated mesh features " + }, + { + "bbox": [ + 304, + 425, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\rho \\in L^{2}(M,\\mathbb{R}^{m})" + }, + { + "bbox": [ + 304, + 425, + 545, + 653 + ], + "type": "text", + "content": ". In the simplest case, the user can paint a coarse, semantic labeling to subjectively delineate salient regions, such as the flippers, head, and rock base of the seal sculpture in Figure 4. After training an FLDM conditioned on the labels, generated textures reflect the specified segmentation. Further examples are shown in Figure 1. Labeled areas contain different textural features, and conditioning ensures the FLDM samples respect these distributions, e.g. that teeth are synthesized only on the mouth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 659, + 378, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 378, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 378, + 672 + ], + "type": "text", + "content": "6.4. Inpainting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In certain cases a user may wish to preserve the input texture in certain regions while synthesizing new content elsewhere. Texture inpainting is a well studied topic" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7959" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 81, + 282, + 316 + ], + "blocks": [ + { + "bbox": [ + 135, + 69, + 261, + 80 + ], + "lines": [ + { + "bbox": [ + 135, + 69, + 261, + 80 + ], + "spans": [ + { + "bbox": [ + 135, + 69, + 261, + 80 + ], + "type": "text", + "content": "FLDM samples across geometries" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 81, + 282, + 316 + ], + "lines": [ + { + "bbox": [ + 53, + 81, + 282, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 81, + 282, + 316 + ], + "type": "image", + "image_path": "f72422ff655744fd6b1f7f490df08dc519195e489d602626c39463dbd4a07826.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 328, + 287, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 395 + ], + "type": "text", + "content": "Figure 6. Generative texture transfer. Since our FL-VAE and FLDMs commute with isometries, we can encode a texture and train an FLDM on an input mesh (left), then sample the pre-trained FLDM on a new, similar mesh and decode the latents to texture it (center to right). Conditioning labels segmenting the shell apertures are not visible. Zoom in to view." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "text", + "content": "[15, 20, 31], and here we propose a straight-forward approach with FLDMs. Given a pre-trained FLDM conditioned on a user-input binary mask " + }, + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "inline_equation", + "content": "m: M \\to \\{0, 1\\}" + }, + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "text", + "content": " specifying the regions to preserve (Figure 5), generative inpainting can be performed by modifying the iterative sampling process. Specifically, after estimating previous the latent features " + }, + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\widetilde{Z}_{t-1}" + }, + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "text", + "content": " as in Equation (11), the features in the mask region are replaced with the appropriately noised input latents from the forward process " + }, + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "inline_equation", + "content": "Z_{t-1}" + }, + { + "bbox": [ + 46, + 399, + 287, + 507 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 513, + 287, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 513, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 91, + 513, + 287, + 528 + ], + "type": "interline_equation", + "content": "\\tilde {Z} _ {t - 1} \\mapsto m \\cdot Z _ {t - 1} + (1 - m) \\cdot \\tilde {Z} _ {t - 1}. \\tag {14}", + "image_path": "0e3a9bdc033b669c51aa11dad8666973e112b8e179b62c8e3fcff680fd1c19fd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "type": "text", + "content": "This approach encourages the inpainted texture to agree with the original at the mask boundaries due to convolutional structure of the denoising network, with this effect observed on the hind leg of sampled textures in Figure 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 587, + 204, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 204, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 204, + 600 + ], + "type": "text", + "content": "6.5. Generative Texture Transfer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Perhaps the most dramatic consequence of isometry-equivariance is that it naturally facilitates a notion of generative texture transfer. Since the FL-VAE and FLDMs commute with isometries " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\gamma : M \\to N" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " as in Equations (6) and (12), our models see " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "TM^{d}" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "TN^{d}" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " as functionally the same space. In practice, we find that long as surfaces " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " are approximately isometric, we can encode a texture and train an FLDM on " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": ", then sample the model on " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " and decode the synthesized latents to texture it." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 81, + 539, + 136 + ], + "blocks": [ + { + "bbox": [ + 331, + 67, + 515, + 78 + ], + "lines": [ + { + "bbox": [ + 331, + 67, + 515, + 78 + ], + "spans": [ + { + "bbox": [ + 331, + 67, + 515, + 78 + ], + "type": "text", + "content": "FLDM samples on increasingly coarse remeshings" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 81, + 539, + 136 + ], + "lines": [ + { + "bbox": [ + 310, + 81, + 539, + 136 + ], + "spans": [ + { + "bbox": [ + 310, + 81, + 539, + 136 + ], + "type": "image", + "image_path": "9370a8f36dc0dadec68557824cf9f71cfdebb2208ddbf2569c4dc7773ae9b9c2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 144, + 545, + 189 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 189 + ], + "type": "text", + "content": "Figure 7. Controlling the scale of synthesized textures. Here, the FLDM trained on the textured skull in the center-left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry, dilating the size of the tesserae. Zoom in to view details." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 197, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 305 + ], + "type": "text", + "content": "Several examples are shown in Figure 6. An FLDM is pre-trained with label conditioning on each of the textured shells in the first column and sampled with corresponding labels on the others to texture them in the same style. Additional examples are shown on the right side of Figure 1. Notably, our model successfully transfers between different topologies, as the skull meshes are genus zero and three. More generally, we are able to transfer textures between highly dissimilar shapes (see supplement for examples)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 314, + 453, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 453, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 453, + 327 + ], + "type": "text", + "content": "6.6. Controlling Textural Scale" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "text", + "content": "An additional feature of our FLDMs is that they offer user-control over the scale of sampled textures. Internally, FLDMs normalize the filter support by dividing the distances between adjacent vertices by the radius of the mean vertex area element " + }, + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\widetilde{r} = \\sqrt{A / (\\pi|V|)}" + }, + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "text", + "content": " is the surface area of the mesh and " + }, + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 304, + 333, + 545, + 513 + ], + "type": "text", + "content": " is the number of vertices. Thus, the scale of the synthesized textural details can be controlled by increasing or decreasing the resolution of the mesh upon which the pre-trained FLDM is sampled. An example shown in Figure 7, where the FLDM trained on the textured skull on the center left of Figure 1 is sampled on progressively coarser remeshings of a new skull geometry. As the number of vertices decreases, the average one-ring occupies an increasingly larger fraction of the surface area, dilating the size of the synthesized textural details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 525, + 463, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 463, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 463, + 538 + ], + "type": "text", + "content": "7. Conclusion and Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "We present an original framework for latent diffusion models operating directly on the surfaces to generate high-quality textures. Our approach consists of two main contributions: a novel latent representation mapping textures to vector fields called Field Latents and Field Latent Diffusion Models which learn to denoise a diffusion process in the latent space on the surface. We apply our method in a single-textured-mesh paradigm, and generate variations of textures with state-of-the-art fidelity. Limitations of our approach include the relatively long training time of FLDMs and an inability to reflect directional information in synthesized textures. The latter could potentially be addressed by conditioning FLDMs on user-specified vector fields, with embeddings formed via dot products with network features." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7960" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Serge Assaad, Carlton Downey, Rami Al-Rfou, Nigamaa Nayakanti, and Ben Sapp. Vn-transformer: Rotation-equivariant attention for vector neurons. arXiv preprint arXiv:2206.04176, 2022. 4, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 169 + ], + "type": "text", + "content": "[2] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. arXiv preprint arXiv:2304.05868, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 213 + ], + "type": "text", + "content": "[3] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In ICCV, pages 4169-4181, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 286, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 286, + 258 + ], + "type": "text", + "content": "[4] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 286, + 304 + ], + "type": "text", + "content": "[5] Qimin Chen, Zhiqin Chen, Hang Zhou, and Hao Zhang. Shaddr: Real-time example-based geometry and texture generation via 3d shape detailization and differentiable rendering. arXiv preprint arXiv:2306.04889, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 286, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 286, + 339 + ], + "type": "text", + "content": "[6] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 286, + 373 + ], + "type": "text", + "content": "[7] An-Chieh Cheng, Xueting Li, Sifei Liu, and Xiaolong Wang. Tuvf: Learning generalizable texture uv radiance fields. arXiv preprint arXiv:2305.03040, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 375, + 286, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 286, + 419 + ], + "type": "text", + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 286, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 286, + 486 + ], + "type": "text", + "content": "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 53, + 420, + 286, + 486 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 53, + 420, + 286, + 486 + ], + "type": "text", + "content": " objects, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 286, + 531 + ], + "type": "text", + "content": "[10] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In ICCV, pages 12200-12209, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 286, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 286, + 577 + ], + "type": "text", + "content": "[11] Nicolas Donati, Etienne Corman, Simone Melzi, and Maks Ovsjanikov. Complex functional maps: A conformal link between tangent bundles. In Computer Graphics Forum, pages 317-334. Wiley Online Library, 2022. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 643 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 286, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 698 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 698 + ], + "type": "text", + "content": "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items, 2022. 1, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "text", + "content": "[14] Ahmed A Elhag, Joshua M Susskind, and Miguel An" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "gel Bautista. Manifold diffusion fields. arXiv preprint arXiv:2305.15586, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[15] Julien Fayer, Bastien Durix, Simone Gasparini, and Géraldine Morin. Texturing and inpainting a complete tubular 3d object reconstructed from partial views. Computers & Graphics, 74:126-136, 2018. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[16] Marc Finzi, Samuel Stanton, Pavel Izmailov, and Andrew Gordon Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In ICML, pages 3165-3176. PMLR, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS, 35:31841-31854, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "type": "text", + "content": "[18] James Gardner, Bernhard Egger, and William Smith. Rotation-equivariant conditional spherical neural fields for learning a natural illumination prior. NeurIPS, 35:26309-26323, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[19] Niv Granot, Ben Feinstein, Assaf Shocher, Shai Bagon, and Michal Irani. Drop the gan: In defense of patches nearest neighbors as single image generative models. In CVPR, pages 13460-13469, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "text", + "content": "[20] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, pages 12135-12144, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 375, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 420 + ], + "type": "text", + "content": "[21] Tobias Hinz, Matthew Fisher, Oliver Wang, and Stefan Wermter. Improved techniques for training single-image gans. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1300–1309, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 421, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 452 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 5, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 454, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 498 + ], + "type": "text", + "content": "[23] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 499, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 545, + 521 + ], + "type": "text", + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 6, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 522, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 544 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[26] Felix Knöppel, Keenan Crane, Ulrich Pinkall, and Peter Schröder. Globally optimal direction fields. ACM TOG, 32 (4):1-10, 2013. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "text", + "content": "[27] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, pages 622-639. Springer, 2022. 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 666 + ], + "type": "text", + "content": "[28] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In ICML, pages 17920-17930. PMLR, 2023. 2, 3, 5, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[29] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7961" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[30] Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[31] A Maggiordomo, P Cignoni, and M Tarini. Texture inpainting for photogrammetric models. In Computer Graphics Forum. Wiley Online Library, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "text", + "content": "[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "type": "text", + "content": "[33] Thomas W. Mitchel. Extending Convolution Through Spatially Adaptive Alignment. PhD thesis, Johns Hopkins University, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "text", + "content": "[34] Thomas W. Mitchel, Benedict Brown, David Koller, Tim Weyrich, Szymon Rusinkiewicz, and Michael Kazhdan. Efficient spatially adaptive convolution and correlation, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "type": "text", + "content": "[35] Thomas W Mitchel, Vladimir G Kim, and Michael Kazhdan. Field convolutions for surface cnns. In ICCV, pages 10001-10011, 2021. 2, 3, 5, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "text", + "content": "[36] Thomas W Mitchel, Szymon Rusinkiewicz, Gregory S Chirikjian, and Michael Kazhdan. Echo: Extended convolution histogram of orientations for local surface description. In Computer Graphics Forum, pages 180-194. Wiley Online Library, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "text", + "content": "[37] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "text", + "content": "[38] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "text", + "content": "[39] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, pages 4531-4540, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[40] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In ICCV, pages 13879-13889, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "text", + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 591, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 655 + ], + "type": "text", + "content": "[42] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[43] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image syn" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "thesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[45] Klaus Hildebrandt Ruben Wiersma, Elmar Eisemann. Cnns on surfaces using rotation-equivariant features. ACM TOG, 39(4), 2020. 5, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[46] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "text", + "content": "[47] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM TOG, XX(X), 2022. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "text", + "content": "[48] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, pages 72-88. Springer, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[49] Saurabh Singh and Shankar Krishnan. Filter response normalization layer: Eliminating batch dependence in the training of deep neural networks. In CVPR, pages 11237-11246, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "text", + "content": "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 321, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 363 + ], + "type": "text", + "content": "[51] David Svitov, Dmitrii Gudkov, Renat Bashirov, and Victor Lempitsky. Dinar: Diffusion inpainting of neural textures for one-shot human avatars. In ICCV, pages 7062-7072, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 365, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 388 + ], + "type": "text", + "content": "[52] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 422 + ], + "type": "text", + "content": "[53] Tianfu Wang, Menelaos Kanakis, Konrad Schindler, Luc Van Gool, and Anton Obukhov. Breathing new life into 3d assets with generative repainting, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 423, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 466 + ], + "type": "text", + "content": "[54] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 2, 3, 5, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "text", + "content": "[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[56] Jiacheng Wei, Hao Wang, Jiashi Feng, Guosheng Lin, and Kim-Hui Yap. Taps3d: Text-guided 3d textured shape generation from pseudo supervision. In CVPR, pages 16805-16815, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[57] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. Deltaconv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 5, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "type": "text", + "content": "[58] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape. arXiv preprint arXiv:2305.15399, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[59] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, pages 9621-9630, 2019. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[60] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7962" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 229 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[61] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with point-uv diffusion. In ICCV, pages 4206–4216, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "text", + "content": "[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "text", + "content": "[63] Xuanmeng Zhang, Jianfeng Zhang, Rohan Chacko, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, pages 2273-2282, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 229 + ], + "type": "text", + "content": "[64] Peiye Zhuang, Samira Abnar, Jiatao Gu, Alex Schwing, Joshua M Susskind, and Miguel Ángel Bautista. Diffusion probabilistic fields. In ICLR, 2022. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7963" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_content_list.json b/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..648d531b3b9559a5b6d91086fd6de84024fdb365 --- /dev/null +++ b/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_content_list.json @@ -0,0 +1,1550 @@ +[ + { + "type": "text", + "text": "Single View Refractive Index Tomography with Neural Fields", + "text_level": 1, + "bbox": [ + 174, + 130, + 795, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brandon Zhao1* Aviad Levis1 Liam Connor2 Pratul P. Srinivasan3 Katherine L. Bouman1,2", + "bbox": [ + 93, + 179, + 875, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computing and Mathematical Sciences, California Institute of Technology $^{2}$ Cahill Center for Astronomy and Astrophysics, California Institute of Technology $^{3}$ Google Research", + "bbox": [ + 127, + 214, + 841, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{*}$ byzhao@caltech.edu", + "bbox": [ + 398, + 271, + 566, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Refractive Index Tomography is the inverse problem of reconstructing the continuously-varying 3D refractive index in a scene using 2D projected image measurements. Although a purely refractive field is not directly visible, it bends light rays as they travel through space, thus providing a signal for reconstruction. The effects of such fields appear in many scientific computer vision settings, ranging from refraction due to transparent cells in microscopy to the lensing of distant galaxies caused by dark matter in astrophysics. Reconstructing these fields is particularly difficult due to the complex nonlinear effects of the refractive field on observed images. Furthermore, while standard 3D reconstruction and tomography settings typically have access to observations of the scene from many viewpoints, many refractive index tomography problem settings only have access to images observed from a single viewpoint. We introduce a method that leverages prior knowledge of light sources scattered throughout the refractive medium to help disambiguate the single-view refractive index tomography problem. We differentiably trace curved rays through a neural field representation of the refractive field, and optimize its parameters to best reproduce the observed image. We demonstrate the efficacy of our approach by reconstructing simulated refractive fields, analyze the effects of light source distribution on the recovered field, and test our method on a simulated dark matter mapping problem where we successfully recover the 3D refractive field caused by a realistic dark matter distribution.", + "bbox": [ + 73, + 353, + 473, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 784, + 209, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When light passes from one medium to another, such as from air to water, it encounters a change in refractive index (RI). As a result, light rays are bent, or refracted, according to the magnitude of this change, causing effects like the shimmering (caustics) at the bottom of a pool of water. Media such as hot gas or translucent cell samples have RIs", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that continuously vary in space, causing light to take paths curved by the gradient of the underlying RI field. Although the RI field is not directly visible, its effect on the scene's illumination provides clues to its structure.", + "bbox": [ + 496, + 323, + 893, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatially varying RI fields are present in a diverse range of scientific applications. For example, certain fiber optic cables are designed to smoothly refract light away from the cable's outside padding [31]. In seismology, pressure waves from earthquakes are refracted continuously as they travel through ground structures of varying density [23]. In astrophysics, gravitational effects from massive dark matter structures lens the light from faraway galaxies resulting in observable effects in telescope images. When the degree of these gravitational lensing effects is relatively weak, they can be modeled as refraction [15].", + "bbox": [ + 496, + 383, + 893, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we seek to tackle the problem of single-view RI tomography: 3D reconstruction of RI fields where image measurements are available only from a single viewpoint. This problem setting occurs when using telescope images or a single camera on scenes in the wild, and poses several challenges. First, image measurements are produced by rays traveling along continuously curving trajectories through the scene, where curvature is due to the structure of the underlying RI field. Unlike in linear tomography problems, where the projection geometry is independent of the medium, in RI tomography the ray paths are dictated by the unknown medium properties themselves. Thus the measurements are a non-linear function of the targeted refractive field. Second, limiting image measurements to a single viewing angle makes the problem severely ill-posed. That is, there are many possible refractive fields which could potentially produce the same image measurements. However, by leveraging the location of light sources distributed throughout the scene we can better constrain the recovery.", + "bbox": [ + 496, + 551, + 895, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel approach using neural fields with continuous ray tracing that leverages the physics of light propagation to recover an underlying refractive field. In particular, we represent the spatially varying RI field with", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25358", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a90c7c540a37ff2c8c3a97cb5384bfce87b5d87ca128443be0b2d777579b2f4.jpg", + "image_caption": [ + "Figure 1. We formulate and present an approach for single-view refractive field reconstruction that explicitly models curved ray paths through a neural field. By leveraging spatial information and the emission profile of light sources located throughout the volume, we are able to localize refractive structures in 3D with just a single viewpoint. Here we demonstrate our approach on a simulation involving a realistic refractive field induced by simulated dark matter halos. Despite the large patches of space that are void of galaxy emitters, our method is still able to recover the rough structure of the 3D dark matter distribution from a single simulated telescope image of refracted galaxies. This reconstruction is also especially challenging because of the extremely weak degree of refraction, as seen in the relative scale of the intensity difference plot; the refractive effects of dark matter result in an intensity change of at most $\\pm 0.3\\%$ of the maximum intensity of the original image. This experiment uses a realistic dark matter distribution derived from IllustrisTNG [27], but we leave the inclusion of realistic measurement noise to future work." + ], + "image_footnote": [], + "bbox": [ + 99, + 99, + 872, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a coordinate-based neural network, which we optimize through a differentiable rendering procedure that traces paths of light through the refractive field to best reproduce measurements. Coordinate-based neural networks are advantageous for this problem because they provide a continuous, easily-differentiable representation of the refractive field as a function of space. We also find empirically that coordinate-based neural networks lead to better optimization behavior out-of-the-box than other representations such as linearly interpolated grids.", + "bbox": [ + 75, + 402, + 468, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach expands upon recent works in neural scene representation [21, 25]. While in these works ray paths through the scene are assumed to be known, our approach explicitly models and solves for unknown refracted ray paths. We first study the conditions under which single-view reconstruction is possible, then test our approach by analyzing its ability to recover weak refractive fields. Next, we analyze the sensitivity of our approach to the density of light sources in the scene. Finally, as shown in Fig. 1, we include a case study motivated by the application of weak lensing where our method recovers the refractive field generated by physically realistic simulations of dark matter.", + "bbox": [ + 75, + 554, + 468, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 742, + 218, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-Based Refractive Field Recovery The reconstruction of refractive objects and phenomena is an ongoing field of study, both in applications that make use of standard consumer cameras as well as scientific imaging applications. In Background-oriented Schlieren (BoS) imaging, images taken of a background with and without refractive effects are correlated to estimate an optical flow describing a projected measurement of the refractive field [4]. Combining these projected measurements from multiple viewpoints", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "through tomography gives an estimate of the 3D gradients, from which the 3D field can be estimated from a Poisson solver. A similar type of tomography can be done to passively model 3D turbulence statistics by combining time resolved deflection measurements from multiple viewpoints [1]. Optical flow estimates such as those in BoS can also be used to detect refractive \"wiggles\" in fluid flow that can then be used to track the projected motion of the refractive media; by observing these wiggles with stereo cameras the fluid's depth can also be recovered [40]. Fluid refraction from water can also be estimated by submerged cameras viewing the sun as a guiding point [2].", + "bbox": [ + 496, + 402, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the aforementioned methods utilize the physics behind ray optics, they don't explicitly model the continuously curved paths that rays take through these refractive media during 3D reconstruction. A method using the refractive curvature of traced rays to supervise an interpolated grid model has been shown to be effective for reconstructing and designing refractive media [37]. However, as with the previous methods, this method still relies on access to multiple viewpoints to reconstruct the refractive volume. In this work, we instead seek to tackle the challenge of accurately reconstructing 3D refractive fields from a single viewpoint.", + "bbox": [ + 496, + 584, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Applications Recent works in optics have used differentiable rendering techniques to design refractive surfaces for various applications. The materials making up the lens typically have a constant RI, meaning light is refracted a discrete number of times at the interface between the lens and the surrounding air. Thus, these works solve for the optimal shape of the material rather than an RI field varying within the lens. Fast differentiable ray tracing techniques have been used for the design of a single lens to produce a", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25359", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "target image [22], as well as compound lenses for camera design [38]. These types of lens designs are aimed at reproducing a target image or favorable image effects. While we also aim to optimize a refractive field to match a set of observed image measurements, we are interested in recovering the properties of the refractive medium itself.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In biomedical imaging, Optical diffraction tomography (ODT) uses the intrinsic optical variation in a sample to reconstruct its RI field [11, 34]. Holographic measurements are taken from multiple illumination angles with coherent light modeled by wave optics. For example, in [12], the authors present a technique to recover RI by using intensity-only measurements from multiple viewpoints. DeCAF [24] uses a neural field representation to solve this inverse problem. Our problem setup differs from these biomedical applications in that we attempt to solve this problem without access to active illumination and coherent light sources. Additionally, we use ray optics rather than wave optics in our characterization of light transport.", + "bbox": [ + 75, + 181, + 467, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are motivated in particular by the application of gravitational lensing, where light coming from small elliptical sources (galaxies) curves on its path to Earth due to an invisible dark matter distribution. In the weak lensing regime, the small ray deflections caused by dark matter can be well-approximated as a refractive effect [15], allowing us to model the large-scale structure of the dark matter distribution as an RI field. Existing works use a linear tomography approach similar to that of BoS [4] by first computing a 2D projected map of mass density, then backprojecting into 3D to obtain a reconstruction. Because image measurements are limited to a single viewpoint, filtered backprojection-style methods [33, 39] are susceptible to heavy smearing along the line of sight, even in the absence of noise. In [19, 20] an approach is used that assumes the density field can be sparsely represented by a dictionary of dark matter halos. However, this assumption is only approximately true in practice; our method can instead reconstruct 3D RI fields induced by a general matter distribution.", + "bbox": [ + 75, + 378, + 467, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coordinate-based Neural Representations Recently coordinate-based neural networks have gained traction for applications in vision and graphics. These models typically parameterize a three-dimensional field, such as a scene's radiance, with the weights of a multi-layer perceptron (MLP). Coordinate-based neural representations have been used in graphics to model shapes with occupancy fields [10], as well as three-dimensional realistic scenes with neural radiance fields [25]. Both of these works rely on a fixed straight ray assumption to render scenes for training and inference. Coordinate-based MLPs have also been used in black hole imaging [21] to reconstruct emission fields with curved ray paths dictated by general relativity; however, the paths taken by rays of light are calculated then fixed beforehand according to a specific model of gravitational lens. In contrast, our", + "bbox": [ + 75, + 674, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work seeks to optimize ray paths that are unknown and directly estimated during model training.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eikonal ray-tracing with neural fields has been used for refractive tomography in the context of novel-view synthesis [6]. However, this work uses images of the scene from multiple viewpoints for supervision. Additionally, it focuses on recovering scenes which have a piecewise constant RI (e.g. a glass block with constant RI surrounded by air), where no light emission occurs within the refractive medium. While it achieves impressive results for rendering images from novel viewpoints, the reconstructed refractive indices of these volumes fail to match the ground truth. In contrast, our method leverages visible light sources from within the refractive medium to enable accurate reconstruction of a refractive volume from only a single viewpoint.", + "bbox": [ + 496, + 121, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Forward: Image Formation Model", + "text_level": 1, + "bbox": [ + 498, + 329, + 813, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section details the components of the image formation (forward) model. First, we introduce the equations governing the refraction and intensity measurements of individual rays. Next, we detail the numerical integration used to model image intensity measurements at a sensor.", + "bbox": [ + 496, + 354, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Refractive Ray Tracing", + "text_level": 1, + "bbox": [ + 498, + 436, + 715, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Purely refractive media are translucent and affect light rays by curving their paths rather than modifying their intensity. Light travels more slowly through more optically dense material, causing it to bend towards an increase in optical density. The RI $\\eta$ is a scalar used to quantify the optical density of a material relative to a vacuum; in natural settings, $\\eta \\geq 1$ as light cannot exceed its speed in a vacuum. In this work, we are interested in volumes with a continuously varying refractive index. The effect of a continuously varying RI $\\eta$ on a light ray's position $x$ and direction $v$ is described by Hamilton's equations for ray tracing [16]:", + "bbox": [ + 496, + 460, + 890, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\frac {d x}{d s} = \\frac {v (s)}{\\eta (s)} \\\\ \\frac {d v}{d s} = \\nabla \\eta (s) \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 637, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here the derivatives of the 3D ray position $x$ and direction $v$ are taken with respect to the ray path length $s$ , and are described in terms of the RI $\\eta$ and its gradient $\\nabla \\eta$ . The directional derivative of Eq. (2) can be understood as an attractive force towards local increases in RI.", + "bbox": [ + 496, + 704, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an initial position and direction $x_0, v_0 \\in \\mathbb{R}^3$ , equations (1) and (2) describe a first-order initial value problem whose solution $\\Gamma(s) = (x, v)$ is a curve describing the trajectory of a ray as a function of its path length $s$ :", + "bbox": [ + 496, + 777, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma (s) = \\left(x _ {0}, v _ {0}\\right) + \\int_ {0} ^ {s} \\left(x ^ {\\prime} (\\tilde {s}), v ^ {\\prime} (\\tilde {s})\\right) d \\tilde {s} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 849, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An illustration of refractive ray tracing is given in Figure 2.", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25360", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca87681480d606e18ed1d8bcbadad1d0920025621061f76f133622d069721748.jpg", + "image_caption": [ + "Figure 2. Illustration of ray tracing in a refractive field. Ray paths are calculated iteratively by solving the Hamiltonian equations (1), (2) parameterized by the refractive field $\\eta$ . Rays are curved locally towards areas of high RI in the direction of $\\nabla \\eta$ ." + ], + "image_footnote": [], + "bbox": [ + 112, + 93, + 433, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sensor and Light Emission Integration", + "text_level": 1, + "bbox": [ + 76, + 359, + 410, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sensor is modeled as a single-wavelength pinhole camera positioned outside of the scene. The intensity image $I$ of a given scene is represented as a discrete function indexed by pixel coordinates $i, j$ . The specific intensity of a pixel $I(i, j)$ is given by integration of light emission over a ray path solution $\\Gamma_{i,j}$ , to Eq. (3). The initial conditions $\\Gamma_{i,j}(0)$ are defined by the position and direction of a ray traveling backwards from the sensor through the pixel center.", + "bbox": [ + 75, + 385, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within the scene, light emission $e(\\mathbf{x})$ is represented as a continuous scalar function of a 3D coordinate $\\mathbf{x} = (x,y,z)$ , and is assumed to be known. This assumption is commonly used for refractive modeling problems in the forms of calibration images in BoS imaging [4] or galaxy catalogs in dark matter mapping [15]. Given the ray path $\\Gamma_{i,j}$ starting at the pixel in location $(i,j)$ , we model the specific intensity $I(i,j)$ by taking the line integral of emission coefficients $e$ along the path $\\Gamma_{i,j}$ :", + "bbox": [ + 75, + 508, + 468, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI (i, j) = \\int_ {\\Gamma_ {i, j}} e (\\mathbf {x} (s)) d s \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 670, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. (4) corresponds to a simplified model of light transport in which the specific intensity is conserved along a ray's path. In gravitational lens theory [32], the emission term in the integrand is decreased for sources in areas of high refractive index due to gravitational redshift. However, for gravitational RI fields with $\\eta$ close to 1, which we focus on in this work, this effect is negligible and we thus exclude it from our forward model [21]. In other applications of refractive tomography [6, 28], the refractive radiative transfer equation (RRTE) [3] may be more appropriate for the rendering forward model. Further details, including experiments with the RRTE, can be found in the supplement.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Numerical Integration", + "text_level": 1, + "bbox": [ + 500, + 90, + 709, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A ray path solution $\\Gamma = (x, v)$ to the Hamiltonian initial value problem given by equations (1) and (2) is fully determined by the initial position and direction $\\Gamma_{i,j}(0)$ of the ray traced through a pixel $(i, j)$ , the refractive field $\\eta$ , and its gradient $\\nabla \\eta$ . Thus, we rewrite the integral in Eq. (3) as:", + "bbox": [ + 498, + 113, + 890, + 190 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {i, j} = \\operatorname {s o l v e H a m} \\left(\\Gamma_ {i, j} (0), \\eta , \\nabla \\eta\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 205, + 890, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Integrating the scene's emission function $e$ along the ray path $\\Gamma$ as in Eq. (4) allows us to calculate image intensities:", + "bbox": [ + 498, + 234, + 890, + 266 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI (i, j) = \\operatorname {r e n d e r} \\operatorname {I m} \\left(\\Gamma_ {i, j}, e\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 281, + 890, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, closed-form solutions for integrals (5) and (6) do not exist, so they are solved numerically. We use an adaptive step-size Dormand-Prince integrator [13] to solve Eqs. (5) and (6) in our experiments. Backpropagating through the integrator using autograd requires memory in the order of the number of solver steps. Instead, we follow previous works [6, 9, 37] in using the adjoint formulation [30] which has constant memory requirements for backpropagation.", + "bbox": [ + 496, + 310, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Inverse: Estimating a Refractive Field", + "text_level": 1, + "bbox": [ + 498, + 445, + 843, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we focus on weakly refractive fields where the deflections of a single source are on the order of fractions of a pixel. Although this may seem small, we rely on the combined signal across an ensemble of lensed sources.", + "bbox": [ + 496, + 470, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve the inverse problem of refractive tomography in this setting, we use an analysis-by-synthesis approach. We model the continuous refractive field with a coordinate-based MLP, which is optimized to fit a true observed image. In the following section we describe the neural representation and optimization scheme.", + "bbox": [ + 496, + 532, + 890, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Continuous Refractive Field Representation", + "text_level": 1, + "bbox": [ + 498, + 632, + 872, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To form a candidate continuously-varying 3D refractive field, we use a coordinate-based MLP inspired by the NeRF [25] model. The MLP, parameterized by weights $\\theta$ , takes as input spatial coordinates $\\mathbf{x}$ and outputs a scalar index of refraction. To ensure that our reconstruction is physically plausible, we apply the Softplus function $\\sigma$ to restrict the output of our network to be greater than 1:", + "bbox": [ + 496, + 656, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\eta} _ {\\theta} (\\mathbf {x}) = \\sigma (\\operatorname {M L P} (\\gamma (\\mathbf {x})); \\theta). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 779, + 890, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\gamma (\\mathbf{x})$ is a positional encoding layer mapping an input coordinate to a set of Fourier basis functions [36] with exponentially-increasing frequencies:", + "bbox": [ + 496, + 806, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma (\\mathbf {x}) = \\left[ \\sin (\\mathbf {x}), \\cos (\\mathbf {x}), \\dots , \\sin \\left(2 ^ {L - 1} \\mathbf {x}\\right), \\cos \\left(2 ^ {L - 1} \\mathbf {x}\\right) \\right] ^ {T}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 864, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25361", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50beebf9300f58d88c684c04463ab757658491b6e7ce04c716119e72fb8ca110.jpg", + "image_caption": [ + "Figure 3. Refractive Tomography Pipeline. We model a 3D RI field $\\hat{\\eta}(\\mathbf{x})$ as a continuous function using a neural network parameterized by $\\theta$ . This refractive field, along with the known visible emission field $e(x)$ from point-like light sources, is fed through a ray tracing simulator to produce a predicted 2D image measurement. An image loss is then taken between the model and true image measurements to solve for $\\theta$ . We differentiably trace rays through this refractive field and accumulate the light emitted along these curved ray paths to produce a 2D image. This allows us to optimize MLP parameters $\\theta$ to minimize the loss between the rendered and true image measurements." + ], + "image_footnote": [], + "bbox": [ + 120, + 97, + 831, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adding this positional encoding allows the MLP to fit continuous fields with higher spatial frequencies as the degree $L$ is increased [36]. We find empirically that a degree of $L = 4$ is suitable for naturally smooth refractive fields.", + "bbox": [ + 76, + 334, + 470, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ray tracing through our model field requires integrating equation (5) during each training iteration. This requires evaluating the model $\\hat{\\eta}_{\\theta}$ and its spatial gradient $\\nabla \\hat{\\eta}_{\\theta}$ at continuous spatial locations over the course of optimization, and for both of these functions to be differentiable with respect to the model parameters $\\theta$ for gradient descent. In other words, we need to be able to evaluate $\\nabla_{\\theta} \\nabla_{x} \\hat{\\eta}_{\\theta}$ at arbitrary points throughout the modeled refractive field. We use the eLU activation function rather than a ReLU so that this gradient is smooth during optimization, and can be easily evaluated with existing autograd frameworks.", + "bbox": [ + 75, + 396, + 472, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Optimization", + "text_level": 1, + "bbox": [ + 76, + 574, + 217, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our neural field model defines a continuous refractive field $\\hat{\\eta}_{\\theta}$ and its gradient $\\nabla_{x}\\hat{\\eta}_{\\theta}$ parameterized by its weights $\\theta$ . Thus, a rendered image $\\hat{I}$ obtained by ray-tracing through a neural field $\\hat{\\eta}_{\\theta}$ with Eqs. (5) and (6) is a function of $\\theta$ .", + "bbox": [ + 75, + 598, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of our pipeline is shown in Fig. 3. To train the neural network weights $\\theta$ , we solve a minimization problem whose objective is a mean-squared error between the observed image $I(i,j)$ , and the rendered image $\\hat{I}(i,j;\\theta)$ , plus a regularization term $\\mathcal{R}(\\theta)$ to enforce a boundary condition of $\\eta = 1$ on the cube's faces:", + "bbox": [ + 75, + 659, + 470, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} (\\theta) = \\sum_ {i, j} | | I (i, j) - \\hat {I} (i, j, \\theta) | | _ {2} ^ {2} + \\lambda \\mathcal {R} (\\theta). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 777, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This loss is fully differentiable with respect to model parameters $\\theta$ , allowing us to use gradient descent to solve the optimization. The term $\\mathcal{R}(\\theta)$ acts only as a constraint on the boundary conditions of the volume; unlike explicit regularizers, such as total variation, we rely on the implicit regu", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "larization from the neural field for our reconstruction. More details on the $\\mathcal{R}(\\theta)$ term can be found in the supplement.", + "bbox": [ + 498, + 334, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 375, + 718, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we use an MLP with 4 layers, where each layer is 256 units wide. The neural network is implemented in JAX [7] and trained with the Adam optimizer [17] with exponential learning rate decay from $10^{-4}$ to $5 \\times 10^{-6}$ over 10K iterations. The weights of the neural network were randomly initialized with the He uniform variance scaling initializer [14]. Code implementation will be publicly available on the project page.", + "bbox": [ + 496, + 401, + 890, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unmodeled Sources of Noise Because our method leverages information about the location and shape of light sources to enable reconstruction from a single viewpoint, it is sensitive to noise in both our knowledge of light source locations and in image measurements. An analysis of our method's sensitivity to these noise sources can be found in the supplement. We plan to investigate ways to overcome this sensitivity in future work.", + "bbox": [ + 496, + 534, + 892, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 666, + 586, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we present results of our method on general refractive tomography problems. We generate ground truth refractive fields inside of a three-dimensional volume, where the RI takes a constant value of 1 outside the region of interest. Each ground truth refractive field is defined by evaluating a continuous function at regularly spaced grid points. The continuous field is then evaluated by linearly interpolating between these grid points. Image measurements are rendered using a pinhole camera with a $64^2$ resolution. Throughout this section, we quantify the accuracy of our reconstructions by sampling the ground truth and reconstructed fields on the same uniformly spaced grid of sample points as the ground truth and report the PSNR. When calculating the PSNR, we first subtract 1 from the RI.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25362", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20ce296905d1ca91806fdc5dad9b468905fbf9bce39e343dc3348d8d705214ca.jpg", + "image_caption": [ + "Light Source Distribution" + ], + "image_footnote": [], + "bbox": [ + 99, + 125, + 215, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dbdb63f900c319f49d2b40ccc2ccd404564244c86ab36be8e1a290eed4e3b753.jpg", + "image_caption": [ + "Measurement Type" + ], + "image_footnote": [], + "bbox": [ + 251, + 133, + 341, + 198 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a8284e3d2cdbdbb56f9e1e820434df2f897182218ff72b2d3ed9b00277cdfd28.jpg", + "image_caption": [ + "Neural Field" + ], + "image_footnote": [], + "bbox": [ + 375, + 119, + 485, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79521e40691946946a6ac961187f6ac4c7ed709540ef65c35aef0c85d5abe52a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 213, + 477, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/398a3e1052a5fa415585c90cbe88b7ccf87279cadcceb27f940c36a826172a61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 301, + 478, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/465d011bdbeeb9880c9f6591f8c676885ced2e23818011520ca6194a62adf6c3.jpg", + "image_caption": [ + "Voxel Grid" + ], + "image_footnote": [], + "bbox": [ + 519, + 126, + 612, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb0591dd5f1a88e38f5fc0cf7fe8f137997402a1dfce0af0a85b041f0be1d7fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 214, + 612, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8da236723c30692d2b5fb3e27267a58e42003fe9fac7fb2a0755494ea9e12d85.jpg", + "image_caption": [ + "Figure 4. Measurement and Reconstruction Conditions. We compare our neural field reconstruction model to a linear voxel grid with and without $\\mathrm{TV}^2$ regularization for gas flow reconstruction under three different measurement conditions. While single-view recovery is difficult with measurements of ray displacement, we find that image measurements of light sources spread throughout the refractive volume makes recovery possible with our method. In addition, a neural field's implicit regularization results in better-constrained reconstructions than the linear voxel grid, even when an explicit regularizer is used. In each recovery we super-impose the ground-truth gas plume in blue over the reconstruction in yellow for comparison. We note that all reconstructions shown fit the measurements almost perfectly; differences in their structures highlight the ill-posedness of the problem." + ], + "image_footnote": [], + "bbox": [ + 522, + 303, + 612, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89b6cb89eff35bd796598340c44f980e73a4633a94ee243014c0ec5f5f9e406f.jpg", + "image_caption": [ + "Voxel Grid (TV2)" + ], + "image_footnote": [], + "bbox": [ + 650, + 127, + 741, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b573e2257e1ebe3181cb417f9e9175be8f56f316ad97861398d0e30d6a5c90c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 215, + 740, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5f0e342e78680e8c1515c04631d65bbb238c62f54eb31fcf970b3e4790cf912.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 303, + 740, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0cce2ddefadb8624086684af89f3fed9f814f0d3d93b961ce48019920844a4f0.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 769, + 191, + 862, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bed70aff96c6933919682dbca83f519d60d2246e06b626a4a19fdb39fd46e4f4.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 772, + 284, + 792, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be6835291efc219663b9d319a1460a0d2494ffa0fe9b9a553b2bfe3c5059617b.jpg", + "image_caption": [ + "Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 772, + 301, + 792, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Measurement and Reconstruction Conditions", + "text_level": 1, + "bbox": [ + 76, + 506, + 460, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we compare refractive reconstructions under nine conditions: three sets of measurement conditions paired with three different reconstruction strategies. We show that single-view RI tomography is made possible by a combination of the measurement conditions of our problem setting and the implicit regularization given by our neural field model. Results are shown in Fig. 4.", + "bbox": [ + 75, + 532, + 468, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Measurement Conditions: The first measurement conditions mirror the astronomical domain that motivates our work, in which image measurements are taken of light sources distributed throughout the refractive volume (Fig. 4 top row). Secondly, we compare to image measurements taken of light sources behind the volume (middle row). Third, we investigate using ray displacement measurements rather than image measurements; in this case, the displacements correspond to a plane of light sources behind the volume (bottom row). To simulate the displacement measurements, we ray trace through the ground truth refractive field to recover a perfect optical flow map at the same resolution as our sensor. Although this would be infeasible in practice, we use these measurements to represent the ideal limit of measurements from a single-view BoS setup [4].", + "bbox": [ + 75, + 641, + 468, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstruction Conditions: We also compare reconstructions from three models: our neural field representa", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion, a linear voxel grid, and a linear voxel grid with $\\mathrm{TV}^2$ regularization, with regularizer weight chosen to maximize PSNR with the ground truth volume. These correspond to the last three columns of Fig. 4, respectively. The goal of the regularized voxel grid is to show that our method can outperform even the highest-performing $\\mathrm{TV}^2$ model; in practice, it would be infeasible to choose the regularizer weight in this way as we don't have access to the ground truth.", + "bbox": [ + 496, + 507, + 890, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data & Results In Fig. 4 we demonstrate the effect of each pair of conditions in reconstructing the fuel injection dataset (a $64^{3}$ RI grid) from SFB 382 of the German Research Council (DFG). Both a method using an unregularized voxel grid as well as BoS have been shown to reconstruct this volume given access to multiple viewpoints [4, 37]. We report the PSNR for each reconstruction, but note that it isn't fully representative of reconstruction quality, which should also be evaluated visually.", + "bbox": [ + 496, + 628, + 892, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Measurement Conditions: When optimizing our neural field model on image measurements of light sources located throughout the refractive volume, accurate 3D recovery of the gas flow is possible with minimal artifacts (A). However, in (B), where the light sources are instead located behind the volume, the neural field is able to recover the rough structure of the gas flow but not its absolute location. When using displacement measurements, none of the reconstructions (C, F, I) are able to correctly recover the shape", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25363", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b320a93dbd033148251f51bd6286a96e51f43d17f4c788bdef89a029235626fd.jpg", + "image_caption": [ + "Figure 5. Recovery of Smooth Refractive Fields. Each row corresponds to a refractive tomography problem with a different random refractive field against a random distribution of 290 light sources. Although refractive recovery from a single viewpoint is highly ill-posed, our method can leverage source location information to detect and localize 3D refractive structure. However, as can be seen in the third example, due to the ill-posed and highly non-convex nature of this tomography problem, our model is still be susceptible to error. The yellow cones indicate sensor direction." + ], + "image_footnote": [], + "bbox": [ + 106, + 99, + 433, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the gas flow from a single viewpoint, even with a perfect optical flow map. While it has been shown that voxel grid optimization with displacement measurements can accurately reconstruct this volume [37], having access to only a single viewpoint causes the reconstruction to be smeared along the optical axis. An analysis of the voxel grid recovery with access to additional viewpoints can be found in the supplement. We conclude that using image measurements of light sources spread throughout the volume is a key ingredient enabling 3D reconstruction from a single viewpoint.", + "bbox": [ + 75, + 593, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Reconstruction Conditions: An image-based loss on point-like light sources is highly nonconvex, and thus models such as an unregularized voxel grid can be prone to converging to bad minima. Although all reconstructions in Fig. 4 fit the measurements almost perfectly, each converges to a different refractive field. The unregularized grid reconstructions (D, E) optimize locally along ray paths to fit the image measurements, resulting in streaklike artifacts and a highly non-smooth recovery. Even when an explicit $\\mathrm{TV}^2$ regularizer is used on the voxel grid, as in", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "recoveries (G) and (H), it still fails to recover the structure of the gas flow along the optical axis. The implicit regularization of our neural field model appears to favor compact reconstructions, which matches the gas flow volume well. In the following section, we show that this implicit regularization also allows us to recover more general smooth fields.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Recovery of Smooth Refractive Fields", + "text_level": 1, + "bbox": [ + 498, + 195, + 825, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate our approach by recovering randomly generated smooth refractive fields in a dense field of light sources from a single viewpoint. We use Poisson Disk sampling [8] to generate an emission field of 290 light source locations. Light sources are amplitude scaled 3D Gaussians with independent random elliptical orientations. Ground truth refractive fields are generated from a random Gaussian Process with a smooth covariance kernel described in the supplement. These true fields were then scaled to an RI range of 1 to 1.003. To put this RI magnitude into context, the refractive effect of these fields causes a median deviation of 0.03 to 0.07 pixels on a $64^2$ pixel image for rays shot parallel through the volume from the front to the back.", + "bbox": [ + 496, + 220, + 890, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 shows that our method has the potential to accurately reconstruct complex, smooth fields. However, due to the ill-posedness and highly non-convex nature of this problem, our model is still susceptible to reconstruction artifacts, as seen in the last row. Note that no explicit regularization besides a boundary condition term was used in solving for these refractive fields. Thus, our results suggest that our neural network architecture combined with a low degree of positional encoding act as an implicit prior on the smoothness of our reconstruction, allowing it to effectively recover smooth refractive fields.", + "bbox": [ + 496, + 417, + 892, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Sensitivity to Light Source Density", + "text_level": 1, + "bbox": [ + 498, + 592, + 802, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From a single viewpoint, refractive fields are detected when they lie between light sources and the sensor. Background refractive structures become difficult to detect with a sparse array of light sources since fewer light sources lie behind them. Because we supervise our model with curved ray paths throughout the scene, our model uses the deflections between adjacent light sources to localize the refractive field. Therefore, we expect that as available light sources in the scene become sparser, our model will struggle to detect and localize the structure of a refractive field.", + "bbox": [ + 496, + 612, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform a refractive field reconstruction on a scene containing five elliptical refractive objects in light source emission fields of varying densities. We sample 250 light source locations uniformly, and create emission fields of decreasing density by taking random subsets of this initial sample. We find that reconstruction quality of our approach degrades with sparser light source fields, especially in the background portion of the field. However, even with only 50 light sources in the volume, our method is able to accu", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25364", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9855240e3cee6a5daa938b590fabd4a90715435fcf1df972c2e7fb46c14a63b5.jpg", + "image_caption": [ + "Figure 6. Sensitivity to Light Source Density. We analyze our approach's reconstruction of a refractive field given light source fields with a range of densities. As the number of light sources decreases, it becomes harder to detect and localize refractive objects, especially towards the back of the volume. However, even with sparse light sources, our method is able to accurately resolve refractive objects in the foreground. The yellow cones indicate sensor direction." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 450, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rately resolve the structure of refractive objects in the foreground. Results are shown in Fig. 6.", + "bbox": [ + 75, + 494, + 470, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Case Study: Dark Matter Mass Mapping", + "text_level": 1, + "bbox": [ + 75, + 531, + 444, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section we show an application to astronomical imaging of dark matter from simulations. In outer space, clumps of invisible mass called dark matter permeate the cosmic web. These clumps, called halos, are so massive that their gravitational pull warps the surrounding space, which lenses nearby passing rays of light. In the weak lensing regime, the effects of gravitational lensing are well-described by refraction from an \"effective\" RI field [26]. In fact, the degree of curvature in weak lensing is so small that its effect on galaxy images is often modeled by a rotation (rather than a translation) of the galaxy's appearance on the sky [5]. By combining telescope images with photometric redshift measurements of the distances of galaxies from Earth, scientists are able to determine their approximate spatial locations. The inverse problem of dark matter mass mapping is to reconstruct an invisible distribution of dark matter given refracted images from telescope surveys and photometric measurements of galaxy distances [15].", + "bbox": [ + 73, + 547, + 472, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Simulated Dark Matter Halo Recovery", + "text_level": 1, + "bbox": [ + 76, + 830, + 413, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now use our method to recover a realistic dark matter refractive field. To generate this refractive field, we use the cosmological hydrodynamical simulation IllustrisTNG,", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which includes both dark and baryonic matter [27, 29]. We use the catalog of dark matter halos produced by the TNG300-1 simulation, focusing on a $20^{3}$ $\\mathrm{Mpc}^3$ region containing 1747 massive halos. As galaxy placement is known to trace the dark matter distribution, we placed Gaussian emission functions emulating galaxies at the center of the 300 most massive halos. We aim to reconstruct a realistic dark matter distribution, but leave the inclusion of realistic measurement noise to future work.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RI tomography in the weak lensing regime requires sensors with sufficient resolution to detect the small shearing effects from gravitational lensing. In our experiments, we use an increased sensor resolution of $512 \\times 512$ ; this resolution in comparison with the scale of our scene is consistent with that of current and upcoming weak lensing surveys [18, 35]. We train our model using SGD, sampling batches of 4096 rays per iteration due to memory constraints.", + "bbox": [ + 496, + 227, + 890, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our reconstruction of the invisible dark matter field is shown in Fig. 1. Our method is able to accurately recover the structure of the dark matter field with slight damping on the largest peaks of halo centers. This recovery is especially challenging because of the distribution of galaxies throughout the scene; large patches of the image are dark as galaxies are not uniformly distributed throughout the volume. In addition, the amplitude of the refractive field generated by these halos is three orders of magnitude lower than in our previous results, resulting in a weak refractive signal; the median deflection caused by the refractive field is $6\\mathrm{e}-3$ pixels. We find that despite these challenges, our method is able to recover a refractive field that correlates strongly with true simulated dark matter halos, and shows promise for application to 3D dark matter recovery in the future.", + "bbox": [ + 496, + 347, + 892, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 584, + 619, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper we present a novel method for the tomography of continuous refractive fields from single-view image measurements with neural fields. Our method succeeds at reconstructing simulated refractive fields with varying light source density. By using the known location of light emitters, our method reconstructs refractive volumes from a single viewpoint that rival those of multiview methods. Finally, our results on physically realistic simulations of dark matter show promise for the application of our method to real weak lensing survey data for dark matter mapping. Beyond dark matter fields, our method opens doors to reconstructing other refractive phenomena, underscoring the computer vision community's potential to drive significant advances in scientific discovery with computational imaging.", + "bbox": [ + 496, + 608, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements BZ was partially supported by the Gupta Fellowship. This work was supported by NSF award 2048237, Google, BP, and AWS Credits from the AI4Science Partnership. We also thank Arjun Teh, Adithya Pediredla and Andrew Chael for their helpful discussions.", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25365", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marina Alterman, Yoav Y Schechner, Minh Vo, and Srinivasa G Narasimhan. Passive tomography of turbulence strength. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 47-60. Springer, 2014. 2", + "[2] Marina Alterman, Yohay Swirski, and Yoav Y Schechner. Stella maris: Stellar marine refractive imaging sensor. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2014. 2", + "[3] Marco Ament, Christoph Bergmann, and Daniel Weiskopf. Refractive radiative transfer equation. ACM Transactions on Graphics (TOG), 33(2):1-22, 2014. 4", + "[4] Bradley Atcheson, Ivo Ihrke, Wolfgang Heidrich, Art Tevs, Derek Bradley, Marcus Magnor, and Hans-Peter Seidel. Time-resolved 3d capture of non-stationary gas flows. ACM transactions on graphics (TOG), 27(5):1-9, 2008. 2, 3, 4, 6", + "[5] Matthias Bartelmann and Peter Schneider. Weak gravitational lensing. Physics Reports, 340(4-5):291-472, 2001. 8", + "[6] Mojtaba Bemana, Karol Myszkowski, Jeppe Revall Frisvad, Hans-Peter Seidel, and Tobias Ritschel. Eikonal fields for refractive novel-view synthesis. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3, 4", + "[7] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 5", + "[8] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 10(1):1, 2007. 7", + "[9] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 4", + "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3", + "[11] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007. 3", + "[12] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019. 3", + "[13] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 6(1):19-26, 1980. 4", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Henk Hoekstra and Bhuvnesh Jain. Weak gravitational lensing and its cosmological applications. Annual Review of Nuclear and Particle Science, 58:99-123, 2008. 1, 3, 4, 8", + "[16] Ivo Ihrke, Gernot Ziegler, Art Tevs, Christian Theobalt, Marcus Magnor, and Hans-Peter Seidel. Eikonal rendering: Efficient light transport in refractive objects. ACM Transactions on Graphics (TOG), 26(3):59–es, 2007. 3", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[18] Rene Laureejs, J Amiaux, S Arduini, J-L Augueres, J Brinchmann, R Cole, M Cropper, C Dabin, L Duvet, A Ealet, et al. Euclid definition study report. arXiv preprint arXiv:1110.3193, 2011. 8", + "[19] Adrienne Leonard, F-X Dupé, and J-L Starck. A compressed sensing approach to 3d weak lensing. *Astronomy & Astrophysics*, 539:A85, 2012. 3", + "[20] Adrienne Leonard, François Lanusse, and Jean-Luc Starck. Glimpse: accurate 3d weak lensing reconstructions using sparsity. Monthly Notices of the Royal Astronomical Society, 440(2):1281-1294, 2014. 3", + "[21] Aviad Levis, Pratul P Srinivasan, Andrew A Chael, Ren Ng, and Katherine L Bouman. Gravitationally lensed black hole emission tomography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19841-19850, 2022. 2, 3, 4", + "[22] Zongling Li, Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Jin Liu, and Wei Zhang. End-to-end learned single lens design using fast differentiable ray tracing. Optics Letters, 46(21): 5453-5456, 2021. 3", + "[23] Fan-Chi Lin, Michael H Ritzwoller, and Roel Snieder. Eikonal tomography: surface wave tomography by phase front tracking across a regional broad-band seismic array. Geophysical Journal International, 177(3):1091-1110, 2009. 1", + "[24] Renhao Liu, Yu Sun, Jiabei Zhu, Lei Tian, and Ulugbek S Kamilov. Recovery of continuous 3d refractive index maps from discrete intensity-only measurements using neural fields. Nature Machine Intelligence, 4(9):781-791, 2022. 3", + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4", + "[26] Ramesh Narayan and Matthias Bartelmann. Lectures on gravitational lensing. arXiv preprint astro-ph/9606001, 1996. 8", + "[27] Dylan Nelson, Annalisa Pillepich, Shy Genel, Mark Vogelsberger, Volker Springel, Paul Torrey, Vicente Rodriguez-Gomez, Debora Sijacki, Gregory F Snyder, Brendan Griffen, et al. The illustris simulation: Public data release. *Astronomy and Computing*, 13:12–37, 2015. 2, 8", + "[28] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25366", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Annalisa Pillepich, Volker Springel, Dylan Nelson, Shy Genel, Jill Naiman, Rüdiger Pakmor, Lars Hernquist, Paul Torrey, Mark Vogelsberger, Rainer Weinberger, et al. Simulating galaxy formation with the illustrstng model. Monthly Notices of the Royal Astronomical Society, 473(3):4077-4106, 2018. 8", + "[30] Lev Semenovich Pontryagin. Mathematical theory of optimal processes. CRC press, 1987. 4", + "[31] William A Reed, Man F Yan, and Mark J Schnitzer. Gradient-index fiber-optic microprobes for minimally invasive in vivo low-coherence interferometry. Optics letters, 27 (20):1794-1796, 2002. 1", + "[32] Peter Schneider, Jürgen Ehlers, Emilio E Falco, et al. Gravitational lenses [electronic resource]. 4", + "[33] Patrick Simon, AN Taylor, and Jan Hartlap. Unfolding the matter distribution using three-dimensional weak gravitational lensing. Monthly Notices of the Royal Astronomical Society, 399(1):48-68, 2009. 3", + "[34] Yongjin Sung, Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Ramachandra R Dasari, and Michael S Feld. Optical diffraction tomography for high resolution live cell imaging. Optics express, 17(1):266-277, 2009. 3", + "[35] Masahiro Takada. Subaru hyper supprime-cam project. In AIP Conference Proceedings, pages 120-127. American Institute of Physics, 2010. 8", + "[36] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4, 5", + "[37] Arjun Teh, Matthew O'Toole, and Ioannis Gkioulekas. Adjoint nonlinear ray tracing. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4, 6, 7", + "[38] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 3", + "[39] JT VanderPlas, AJ Connolly, Bhuvnesh Jain, and Mike Jarvis. Three-dimensional reconstruction of the density field: An svd approach to weak-lensing tomography. The Astrophysical Journal, 727(2):118, 2011. 3", + "[40] Tianfan Xue, Michael Rubinstein, Neal Wadhwa, Anat Levin, Fredo Durand, and William T Freeman. Refraction wiggles for measuring fluid depth and velocity from video. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 767-782. Springer, 2014. 2" + ], + "bbox": [ + 78, + 90, + 468, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25367", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_model.json b/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c69dc6791470de7e09169d3c26ad36a16e7dc20 --- /dev/null +++ b/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.797, + 0.154 + ], + "angle": 0, + "content": "Single View Refractive Index Tomography with Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.18, + 0.877, + 0.199 + ], + "angle": 0, + "content": "Brandon Zhao1* Aviad Levis1 Liam Connor2 Pratul P. Srinivasan3 Katherine L. Bouman1,2" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.215, + 0.842, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Computing and Mathematical Sciences, California Institute of Technology \\(^{2}\\)Cahill Center for Astronomy and Astrophysics, California Institute of Technology \\(^{3}\\)Google Research" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.272, + 0.568, + 0.287 + ], + "angle": 0, + "content": "\\(^{*}\\)byzhao@caltech.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.354, + 0.474, + 0.779 + ], + "angle": 0, + "content": "Refractive Index Tomography is the inverse problem of reconstructing the continuously-varying 3D refractive index in a scene using 2D projected image measurements. Although a purely refractive field is not directly visible, it bends light rays as they travel through space, thus providing a signal for reconstruction. The effects of such fields appear in many scientific computer vision settings, ranging from refraction due to transparent cells in microscopy to the lensing of distant galaxies caused by dark matter in astrophysics. Reconstructing these fields is particularly difficult due to the complex nonlinear effects of the refractive field on observed images. Furthermore, while standard 3D reconstruction and tomography settings typically have access to observations of the scene from many viewpoints, many refractive index tomography problem settings only have access to images observed from a single viewpoint. We introduce a method that leverages prior knowledge of light sources scattered throughout the refractive medium to help disambiguate the single-view refractive index tomography problem. We differentiably trace curved rays through a neural field representation of the refractive field, and optimize its parameters to best reproduce the observed image. We demonstrate the efficacy of our approach by reconstructing simulated refractive fields, analyze the effects of light source distribution on the recovered field, and test our method on a simulated dark matter mapping problem where we successfully recover the 3D refractive field caused by a realistic dark matter distribution." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.21, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "When light passes from one medium to another, such as from air to water, it encounters a change in refractive index (RI). As a result, light rays are bent, or refracted, according to the magnitude of this change, causing effects like the shimmering (caustics) at the bottom of a pool of water. Media such as hot gas or translucent cell samples have RIs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.895, + 0.383 + ], + "angle": 0, + "content": "that continuously vary in space, causing light to take paths curved by the gradient of the underlying RI field. Although the RI field is not directly visible, its effect on the scene's illumination provides clues to its structure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.895, + 0.552 + ], + "angle": 0, + "content": "Spatially varying RI fields are present in a diverse range of scientific applications. For example, certain fiber optic cables are designed to smoothly refract light away from the cable's outside padding [31]. In seismology, pressure waves from earthquakes are refracted continuously as they travel through ground structures of varying density [23]. In astrophysics, gravitational effects from massive dark matter structures lens the light from faraway galaxies resulting in observable effects in telescope images. When the degree of these gravitational lensing effects is relatively weak, they can be modeled as refraction [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.896, + 0.84 + ], + "angle": 0, + "content": "In this work, we seek to tackle the problem of single-view RI tomography: 3D reconstruction of RI fields where image measurements are available only from a single viewpoint. This problem setting occurs when using telescope images or a single camera on scenes in the wild, and poses several challenges. First, image measurements are produced by rays traveling along continuously curving trajectories through the scene, where curvature is due to the structure of the underlying RI field. Unlike in linear tomography problems, where the projection geometry is independent of the medium, in RI tomography the ray paths are dictated by the unknown medium properties themselves. Thus the measurements are a non-linear function of the targeted refractive field. Second, limiting image measurements to a single viewing angle makes the problem severely ill-posed. That is, there are many possible refractive fields which could potentially produce the same image measurements. However, by leveraging the location of light sources distributed throughout the scene we can better constrain the recovery." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We propose a novel approach using neural fields with continuous ray tracing that leverages the physics of light propagation to recover an underlying refractive field. In particular, we represent the spatially varying RI field with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25358" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.1, + 0.874, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Figure 1. We formulate and present an approach for single-view refractive field reconstruction that explicitly models curved ray paths through a neural field. By leveraging spatial information and the emission profile of light sources located throughout the volume, we are able to localize refractive structures in 3D with just a single viewpoint. Here we demonstrate our approach on a simulation involving a realistic refractive field induced by simulated dark matter halos. Despite the large patches of space that are void of galaxy emitters, our method is still able to recover the rough structure of the 3D dark matter distribution from a single simulated telescope image of refracted galaxies. This reconstruction is also especially challenging because of the extremely weak degree of refraction, as seen in the relative scale of the intensity difference plot; the refractive effects of dark matter result in an intensity change of at most \\(\\pm 0.3\\%\\) of the maximum intensity of the original image. This experiment uses a realistic dark matter distribution derived from IllustrisTNG [27], but we leave the inclusion of realistic measurement noise to future work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.553 + ], + "angle": 0, + "content": "a coordinate-based neural network, which we optimize through a differentiable rendering procedure that traces paths of light through the refractive field to best reproduce measurements. Coordinate-based neural networks are advantageous for this problem because they provide a continuous, easily-differentiable representation of the refractive field as a function of space. We also find empirically that coordinate-based neural networks lead to better optimization behavior out-of-the-box than other representations such as linearly interpolated grids." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.47, + 0.736 + ], + "angle": 0, + "content": "Our approach expands upon recent works in neural scene representation [21, 25]. While in these works ray paths through the scene are assumed to be known, our approach explicitly models and solves for unknown refracted ray paths. We first study the conditions under which single-view reconstruction is possible, then test our approach by analyzing its ability to recover weak refractive fields. Next, we analyze the sensitivity of our approach to the density of light sources in the scene. Finally, as shown in Fig. 1, we include a case study motivated by the application of weak lensing where our method recovers the refractive field generated by physically realistic simulations of dark matter." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.743, + 0.219, + 0.758 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Image-Based Refractive Field Recovery The reconstruction of refractive objects and phenomena is an ongoing field of study, both in applications that make use of standard consumer cameras as well as scientific imaging applications. In Background-oriented Schlieren (BoS) imaging, images taken of a background with and without refractive effects are correlated to estimate an optical flow describing a projected measurement of the refractive field [4]. Combining these projected measurements from multiple viewpoints" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.584 + ], + "angle": 0, + "content": "through tomography gives an estimate of the 3D gradients, from which the 3D field can be estimated from a Poisson solver. A similar type of tomography can be done to passively model 3D turbulence statistics by combining time resolved deflection measurements from multiple viewpoints [1]. Optical flow estimates such as those in BoS can also be used to detect refractive \"wiggles\" in fluid flow that can then be used to track the projected motion of the refractive media; by observing these wiggles with stereo cameras the fluid's depth can also be recovered [40]. Fluid refraction from water can also be estimated by submerged cameras viewing the sun as a guiding point [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.751 + ], + "angle": 0, + "content": "While the aforementioned methods utilize the physics behind ray optics, they don't explicitly model the continuously curved paths that rays take through these refractive media during 3D reconstruction. A method using the refractive curvature of traced rays to supervise an interpolated grid model has been shown to be effective for reconstructing and designing refractive media [37]. However, as with the previous methods, this method still relies on access to multiple viewpoints to reconstruct the refractive volume. In this work, we instead seek to tackle the challenge of accurately reconstructing 3D refractive fields from a single viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Applications Recent works in optics have used differentiable rendering techniques to design refractive surfaces for various applications. The materials making up the lens typically have a constant RI, meaning light is refracted a discrete number of times at the interface between the lens and the surrounding air. Thus, these works solve for the optimal shape of the material rather than an RI field varying within the lens. Fast differentiable ray tracing techniques have been used for the design of a single lens to produce a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25359" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "target image [22], as well as compound lenses for camera design [38]. These types of lens designs are aimed at reproducing a target image or favorable image effects. While we also aim to optimize a refractive field to match a set of observed image measurements, we are interested in recovering the properties of the refractive medium itself." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.378 + ], + "angle": 0, + "content": "In biomedical imaging, Optical diffraction tomography (ODT) uses the intrinsic optical variation in a sample to reconstruct its RI field [11, 34]. Holographic measurements are taken from multiple illumination angles with coherent light modeled by wave optics. For example, in [12], the authors present a technique to recover RI by using intensity-only measurements from multiple viewpoints. DeCAF [24] uses a neural field representation to solve this inverse problem. Our problem setup differs from these biomedical applications in that we attempt to solve this problem without access to active illumination and coherent light sources. Additionally, we use ray optics rather than wave optics in our characterization of light transport." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.468, + 0.664 + ], + "angle": 0, + "content": "We are motivated in particular by the application of gravitational lensing, where light coming from small elliptical sources (galaxies) curves on its path to Earth due to an invisible dark matter distribution. In the weak lensing regime, the small ray deflections caused by dark matter can be well-approximated as a refractive effect [15], allowing us to model the large-scale structure of the dark matter distribution as an RI field. Existing works use a linear tomography approach similar to that of BoS [4] by first computing a 2D projected map of mass density, then backprojecting into 3D to obtain a reconstruction. Because image measurements are limited to a single viewpoint, filtered backprojection-style methods [33, 39] are susceptible to heavy smearing along the line of sight, even in the absence of noise. In [19, 20] an approach is used that assumes the density field can be sparsely represented by a dictionary of dark matter halos. However, this assumption is only approximately true in practice; our method can instead reconstruct 3D RI fields induced by a general matter distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Coordinate-based Neural Representations Recently coordinate-based neural networks have gained traction for applications in vision and graphics. These models typically parameterize a three-dimensional field, such as a scene's radiance, with the weights of a multi-layer perceptron (MLP). Coordinate-based neural representations have been used in graphics to model shapes with occupancy fields [10], as well as three-dimensional realistic scenes with neural radiance fields [25]. Both of these works rely on a fixed straight ray assumption to render scenes for training and inference. Coordinate-based MLPs have also been used in black hole imaging [21] to reconstruct emission fields with curved ray paths dictated by general relativity; however, the paths taken by rays of light are calculated then fixed beforehand according to a specific model of gravitational lens. In contrast, our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "work seeks to optimize ray paths that are unknown and directly estimated during model training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Eikonal ray-tracing with neural fields has been used for refractive tomography in the context of novel-view synthesis [6]. However, this work uses images of the scene from multiple viewpoints for supervision. Additionally, it focuses on recovering scenes which have a piecewise constant RI (e.g. a glass block with constant RI surrounded by air), where no light emission occurs within the refractive medium. While it achieves impressive results for rendering images from novel viewpoints, the reconstructed refractive indices of these volumes fail to match the ground truth. In contrast, our method leverages visible light sources from within the refractive medium to enable accurate reconstruction of a refractive volume from only a single viewpoint." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.33, + 0.815, + 0.347 + ], + "angle": 0, + "content": "3. Forward: Image Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.43 + ], + "angle": 0, + "content": "This section details the components of the image formation (forward) model. First, we introduce the equations governing the refraction and intensity measurements of individual rays. Next, we detail the numerical integration used to model image intensity measurements at a sensor." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.438, + 0.716, + 0.453 + ], + "angle": 0, + "content": "3.1. Refractive Ray Tracing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Purely refractive media are translucent and affect light rays by curving their paths rather than modifying their intensity. Light travels more slowly through more optically dense material, causing it to bend towards an increase in optical density. The RI \\(\\eta\\) is a scalar used to quantify the optical density of a material relative to a vacuum; in natural settings, \\(\\eta \\geq 1\\) as light cannot exceed its speed in a vacuum. In this work, we are interested in volumes with a continuously varying refractive index. The effect of a continuously varying RI \\(\\eta\\) on a light ray's position \\(x\\) and direction \\(v\\) is described by Hamilton's equations for ray tracing [16]:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.638, + 0.892, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\frac {d x}{d s} = \\frac {v (s)}{\\eta (s)} \\\\ \\frac {d v}{d s} = \\nabla \\eta (s) \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.778 + ], + "angle": 0, + "content": "Here the derivatives of the 3D ray position \\( x \\) and direction \\( v \\) are taken with respect to the ray path length \\( s \\), and are described in terms of the RI \\( \\eta \\) and its gradient \\( \\nabla \\eta \\). The directional derivative of Eq. (2) can be understood as an attractive force towards local increases in RI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Given an initial position and direction \\( x_0, v_0 \\in \\mathbb{R}^3 \\), equations (1) and (2) describe a first-order initial value problem whose solution \\( \\Gamma(s) = (x, v) \\) is a curve describing the trajectory of a ray as a function of its path length \\( s \\):" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.851, + 0.892, + 0.884 + ], + "angle": 0, + "content": "\\[\n\\Gamma (s) = \\left(x _ {0}, v _ {0}\\right) + \\int_ {0} ^ {s} \\left(x ^ {\\prime} (\\tilde {s}), v ^ {\\prime} (\\tilde {s})\\right) d \\tilde {s} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "An illustration of refractive ray tracing is given in Figure 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25360" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.094, + 0.434, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.471, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Illustration of ray tracing in a refractive field. Ray paths are calculated iteratively by solving the Hamiltonian equations (1), (2) parameterized by the refractive field \\(\\eta\\). Rays are curved locally towards areas of high RI in the direction of \\(\\nabla \\eta\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.36, + 0.411, + 0.377 + ], + "angle": 0, + "content": "3.2. Sensor and Light Emission Integration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.508 + ], + "angle": 0, + "content": "The sensor is modeled as a single-wavelength pinhole camera positioned outside of the scene. The intensity image \\(I\\) of a given scene is represented as a discrete function indexed by pixel coordinates \\(i, j\\). The specific intensity of a pixel \\(I(i, j)\\) is given by integration of light emission over a ray path solution \\(\\Gamma_{i,j}\\), to Eq. (3). The initial conditions \\(\\Gamma_{i,j}(0)\\) are defined by the position and direction of a ray traveling backwards from the sensor through the pixel center." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Within the scene, light emission \\( e(\\mathbf{x}) \\) is represented as a continuous scalar function of a 3D coordinate \\( \\mathbf{x} = (x,y,z) \\), and is assumed to be known. This assumption is commonly used for refractive modeling problems in the forms of calibration images in BoS imaging [4] or galaxy catalogs in dark matter mapping [15]. Given the ray path \\( \\Gamma_{i,j} \\) starting at the pixel in location \\( (i,j) \\), we model the specific intensity \\( I(i,j) \\) by taking the line integral of emission coefficients \\( e \\) along the path \\( \\Gamma_{i,j} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.671, + 0.469, + 0.707 + ], + "angle": 0, + "content": "\\[\nI (i, j) = \\int_ {\\Gamma_ {i, j}} e (\\mathbf {x} (s)) d s \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Eq. (4) corresponds to a simplified model of light transport in which the specific intensity is conserved along a ray's path. In gravitational lens theory [32], the emission term in the integrand is decreased for sources in areas of high refractive index due to gravitational redshift. However, for gravitational RI fields with \\(\\eta\\) close to 1, which we focus on in this work, this effect is negligible and we thus exclude it from our forward model [21]. In other applications of refractive tomography [6, 28], the refractive radiative transfer equation (RRTE) [3] may be more appropriate for the rendering forward model. Further details, including experiments with the RRTE, can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.71, + 0.108 + ], + "angle": 0, + "content": "3.3. Numerical Integration" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.114, + 0.892, + 0.191 + ], + "angle": 0, + "content": "A ray path solution \\(\\Gamma = (x, v)\\) to the Hamiltonian initial value problem given by equations (1) and (2) is fully determined by the initial position and direction \\(\\Gamma_{i,j}(0)\\) of the ray traced through a pixel \\((i, j)\\), the refractive field \\(\\eta\\), and its gradient \\(\\nabla \\eta\\). Thus, we rewrite the integral in Eq. (3) as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.207, + 0.892, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {i, j} = \\operatorname {s o l v e H a m} \\left(\\Gamma_ {i, j} (0), \\eta , \\nabla \\eta\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.235, + 0.892, + 0.267 + ], + "angle": 0, + "content": "Integrating the scene's emission function \\( e \\) along the ray path \\( \\Gamma \\) as in Eq. (4) allows us to calculate image intensities:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.282, + 0.892, + 0.3 + ], + "angle": 0, + "content": "\\[\nI (i, j) = \\operatorname {r e n d e r} \\operatorname {I m} \\left(\\Gamma_ {i, j}, e\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.432 + ], + "angle": 0, + "content": "In general, closed-form solutions for integrals (5) and (6) do not exist, so they are solved numerically. We use an adaptive step-size Dormand-Prince integrator [13] to solve Eqs. (5) and (6) in our experiments. Backpropagating through the integrator using autograd requires memory in the order of the number of solver steps. Instead, we follow previous works [6, 9, 37] in using the adjoint formulation [30] which has constant memory requirements for backpropagation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.446, + 0.844, + 0.464 + ], + "angle": 0, + "content": "4. Inverse: Estimating a Refractive Field" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.532 + ], + "angle": 0, + "content": "In this work, we focus on weakly refractive fields where the deflections of a single source are on the order of fractions of a pixel. Although this may seem small, we rely on the combined signal across an ensemble of lensed sources." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.624 + ], + "angle": 0, + "content": "To solve the inverse problem of refractive tomography in this setting, we use an analysis-by-synthesis approach. We model the continuous refractive field with a coordinate-based MLP, which is optimized to fit a true observed image. In the following section we describe the neural representation and optimization scheme." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.633, + 0.874, + 0.65 + ], + "angle": 0, + "content": "4.1. Continuous Refractive Field Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.764 + ], + "angle": 0, + "content": "To form a candidate continuously-varying 3D refractive field, we use a coordinate-based MLP inspired by the NeRF [25] model. The MLP, parameterized by weights \\(\\theta\\), takes as input spatial coordinates \\(\\mathbf{x}\\) and outputs a scalar index of refraction. To ensure that our reconstruction is physically plausible, we apply the Softplus function \\(\\sigma\\) to restrict the output of our network to be greater than 1:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.78, + 0.891, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\hat {\\eta} _ {\\theta} (\\mathbf {x}) = \\sigma (\\operatorname {M L P} (\\gamma (\\mathbf {x})); \\theta). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Here, \\(\\gamma (\\mathbf{x})\\) is a positional encoding layer mapping an input coordinate to a set of Fourier basis functions [36] with exponentially-increasing frequencies:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.865, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\gamma (\\mathbf {x}) = \\left[ \\sin (\\mathbf {x}), \\cos (\\mathbf {x}), \\dots , \\sin \\left(2 ^ {L - 1} \\mathbf {x}\\right), \\cos \\left(2 ^ {L - 1} \\mathbf {x}\\right) \\right] ^ {T}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25361" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.098, + 0.833, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.895, + 0.328 + ], + "angle": 0, + "content": "Figure 3. Refractive Tomography Pipeline. We model a 3D RI field \\(\\hat{\\eta}(\\mathbf{x})\\) as a continuous function using a neural network parameterized by \\(\\theta\\). This refractive field, along with the known visible emission field \\(e(x)\\) from point-like light sources, is fed through a ray tracing simulator to produce a predicted 2D image measurement. An image loss is then taken between the model and true image measurements to solve for \\(\\theta\\). We differentiably trace rays through this refractive field and accumulate the light emitted along these curved ray paths to produce a 2D image. This allows us to optimize MLP parameters \\(\\theta\\) to minimize the loss between the rendered and true image measurements." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.335, + 0.471, + 0.396 + ], + "angle": 0, + "content": "Adding this positional encoding allows the MLP to fit continuous fields with higher spatial frequencies as the degree \\( L \\) is increased [36]. We find empirically that a degree of \\( L = 4 \\) is suitable for naturally smooth refractive fields." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.473, + 0.564 + ], + "angle": 0, + "content": "Ray tracing through our model field requires integrating equation (5) during each training iteration. This requires evaluating the model \\(\\hat{\\eta}_{\\theta}\\) and its spatial gradient \\(\\nabla \\hat{\\eta}_{\\theta}\\) at continuous spatial locations over the course of optimization, and for both of these functions to be differentiable with respect to the model parameters \\(\\theta\\) for gradient descent. In other words, we need to be able to evaluate \\(\\nabla_{\\theta} \\nabla_{x} \\hat{\\eta}_{\\theta}\\) at arbitrary points throughout the modeled refractive field. We use the eLU activation function rather than a ReLU so that this gradient is smooth during optimization, and can be easily evaluated with existing autograd frameworks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.218, + 0.59 + ], + "angle": 0, + "content": "4.2. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Our neural field model defines a continuous refractive field \\(\\hat{\\eta}_{\\theta}\\) and its gradient \\(\\nabla_{x}\\hat{\\eta}_{\\theta}\\) parameterized by its weights \\(\\theta\\). Thus, a rendered image \\(\\hat{I}\\) obtained by ray-tracing through a neural field \\(\\hat{\\eta}_{\\theta}\\) with Eqs. (5) and (6) is a function of \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.751 + ], + "angle": 0, + "content": "An overview of our pipeline is shown in Fig. 3. To train the neural network weights \\(\\theta\\), we solve a minimization problem whose objective is a mean-squared error between the observed image \\(I(i,j)\\), and the rendered image \\(\\hat{I}(i,j;\\theta)\\), plus a regularization term \\(\\mathcal{R}(\\theta)\\) to enforce a boundary condition of \\(\\eta = 1\\) on the cube's faces:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.779, + 0.47, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} (\\theta) = \\sum_ {i, j} | | I (i, j) - \\hat {I} (i, j, \\theta) | | _ {2} ^ {2} + \\lambda \\mathcal {R} (\\theta). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This loss is fully differentiable with respect to model parameters \\(\\theta\\), allowing us to use gradient descent to solve the optimization. The term \\(\\mathcal{R}(\\theta)\\) acts only as a constraint on the boundary conditions of the volume; unlike explicit regularizers, such as total variation, we rely on the implicit regu" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.366 + ], + "angle": 0, + "content": "larization from the neural field for our reconstruction. More details on the \\(\\mathcal{R}(\\theta)\\) term can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.376, + 0.719, + 0.393 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.525 + ], + "angle": 0, + "content": "In our experiments, we use an MLP with 4 layers, where each layer is 256 units wide. The neural network is implemented in JAX [7] and trained with the Adam optimizer [17] with exponential learning rate decay from \\(10^{-4}\\) to \\(5 \\times 10^{-6}\\) over 10K iterations. The weights of the neural network were randomly initialized with the He uniform variance scaling initializer [14]. Code implementation will be publicly available on the project page." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.656 + ], + "angle": 0, + "content": "Unmodeled Sources of Noise Because our method leverages information about the location and shape of light sources to enable reconstruction from a single viewpoint, it is sensitive to noise in both our knowledge of light source locations and in image measurements. An analysis of our method's sensitivity to these noise sources can be found in the supplement. We plan to investigate ways to overcome this sensitivity in future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.587, + 0.683 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this section we present results of our method on general refractive tomography problems. We generate ground truth refractive fields inside of a three-dimensional volume, where the RI takes a constant value of 1 outside the region of interest. Each ground truth refractive field is defined by evaluating a continuous function at regularly spaced grid points. The continuous field is then evaluated by linearly interpolating between these grid points. Image measurements are rendered using a pinhole camera with a \\(64^2\\) resolution. Throughout this section, we quantify the accuracy of our reconstructions by sampling the ground truth and reconstructed fields on the same uniformly spaced grid of sample points as the ground truth and report the PSNR. When calculating the PSNR, we first subtract 1 from the RI." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25362" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.096, + 0.193, + 0.12 + ], + "angle": 0, + "content": "Light Source Distribution" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.125, + 0.216, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.096, + 0.338, + 0.121 + ], + "angle": 0, + "content": "Measurement Type" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.134, + 0.343, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.102, + 0.466, + 0.114 + ], + "angle": 0, + "content": "Neural Field" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.121, + 0.486, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.214, + 0.478, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.303, + 0.48, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.102, + 0.597, + 0.113 + ], + "angle": 0, + "content": "Voxel Grid" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.127, + 0.614, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.215, + 0.614, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.304, + 0.614, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.102, + 0.741, + 0.114 + ], + "angle": 0, + "content": "Voxel Grid (TV2)" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.128, + 0.742, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.216, + 0.741, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.304, + 0.741, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.177, + 0.857, + 0.188 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.192, + 0.864, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.285, + 0.794, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.286, + 0.859, + 0.295 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.302, + 0.794, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.302, + 0.869, + 0.312 + ], + "angle": 0, + "content": "Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Measurement and Reconstruction Conditions. We compare our neural field reconstruction model to a linear voxel grid with and without \\(\\mathrm{TV}^2\\) regularization for gas flow reconstruction under three different measurement conditions. While single-view recovery is difficult with measurements of ray displacement, we find that image measurements of light sources spread throughout the refractive volume makes recovery possible with our method. In addition, a neural field's implicit regularization results in better-constrained reconstructions than the linear voxel grid, even when an explicit regularizer is used. In each recovery we super-impose the ground-truth gas plume in blue over the reconstruction in yellow for comparison. We note that all reconstructions shown fit the measurements almost perfectly; differences in their structures highlight the ill-posedness of the problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.507, + 0.462, + 0.523 + ], + "angle": 0, + "content": "5.1. Measurement and Reconstruction Conditions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.469, + 0.639 + ], + "angle": 0, + "content": "In this section we compare refractive reconstructions under nine conditions: three sets of measurement conditions paired with three different reconstruction strategies. We show that single-view RI tomography is made possible by a combination of the measurement conditions of our problem setting and the implicit regularization given by our neural field model. Results are shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Measurement Conditions: The first measurement conditions mirror the astronomical domain that motivates our work, in which image measurements are taken of light sources distributed throughout the refractive volume (Fig. 4 top row). Secondly, we compare to image measurements taken of light sources behind the volume (middle row). Third, we investigate using ray displacement measurements rather than image measurements; in this case, the displacements correspond to a plane of light sources behind the volume (bottom row). To simulate the displacement measurements, we ray trace through the ground truth refractive field to recover a perfect optical flow map at the same resolution as our sensor. Although this would be infeasible in practice, we use these measurements to represent the ideal limit of measurements from a single-view BoS setup [4]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Reconstruction Conditions: We also compare reconstructions from three models: our neural field representa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.629 + ], + "angle": 0, + "content": "tion, a linear voxel grid, and a linear voxel grid with \\(\\mathrm{TV}^2\\) regularization, with regularizer weight chosen to maximize PSNR with the ground truth volume. These correspond to the last three columns of Fig. 4, respectively. The goal of the regularized voxel grid is to show that our method can outperform even the highest-performing \\(\\mathrm{TV}^2\\) model; in practice, it would be infeasible to choose the regularizer weight in this way as we don't have access to the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.765 + ], + "angle": 0, + "content": "Data & Results In Fig. 4 we demonstrate the effect of each pair of conditions in reconstructing the fuel injection dataset (a \\(64^{3}\\) RI grid) from SFB 382 of the German Research Council (DFG). Both a method using an unregularized voxel grid as well as BoS have been shown to reconstruct this volume given access to multiple viewpoints [4, 37]. We report the PSNR for each reconstruction, but note that it isn't fully representative of reconstruction quality, which should also be evaluated visually." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of Measurement Conditions: When optimizing our neural field model on image measurements of light sources located throughout the refractive volume, accurate 3D recovery of the gas flow is possible with minimal artifacts (A). However, in (B), where the light sources are instead located behind the volume, the neural field is able to recover the rough structure of the gas flow but not its absolute location. When using displacement measurements, none of the reconstructions (C, F, I) are able to correctly recover the shape" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25363" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.101, + 0.434, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.46, + 0.473, + 0.587 + ], + "angle": 0, + "content": "Figure 5. Recovery of Smooth Refractive Fields. Each row corresponds to a refractive tomography problem with a different random refractive field against a random distribution of 290 light sources. Although refractive recovery from a single viewpoint is highly ill-posed, our method can leverage source location information to detect and localize 3D refractive structure. However, as can be seen in the third example, due to the ill-posed and highly non-convex nature of this tomography problem, our model is still be susceptible to error. The yellow cones indicate sensor direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.746 + ], + "angle": 0, + "content": "of the gas flow from a single viewpoint, even with a perfect optical flow map. While it has been shown that voxel grid optimization with displacement measurements can accurately reconstruct this volume [37], having access to only a single viewpoint causes the reconstruction to be smeared along the optical axis. An analysis of the voxel grid recovery with access to additional viewpoints can be found in the supplement. We conclude that using image measurements of light sources spread throughout the volume is a key ingredient enabling 3D reconstruction from a single viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of Reconstruction Conditions: An image-based loss on point-like light sources is highly nonconvex, and thus models such as an unregularized voxel grid can be prone to converging to bad minima. Although all reconstructions in Fig. 4 fit the measurements almost perfectly, each converges to a different refractive field. The unregularized grid reconstructions (D, E) optimize locally along ray paths to fit the image measurements, resulting in streaklike artifacts and a highly non-smooth recovery. Even when an explicit \\(\\mathrm{TV}^2\\) regularizer is used on the voxel grid, as in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "recoveries (G) and (H), it still fails to recover the structure of the gas flow along the optical axis. The implicit regularization of our neural field model appears to favor compact reconstructions, which matches the gas flow volume well. In the following section, we show that this implicit regularization also allows us to recover more general smooth fields." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.196, + 0.826, + 0.212 + ], + "angle": 0, + "content": "5.2. Recovery of Smooth Refractive Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.221, + 0.892, + 0.417 + ], + "angle": 0, + "content": "We demonstrate our approach by recovering randomly generated smooth refractive fields in a dense field of light sources from a single viewpoint. We use Poisson Disk sampling [8] to generate an emission field of 290 light source locations. Light sources are amplitude scaled 3D Gaussians with independent random elliptical orientations. Ground truth refractive fields are generated from a random Gaussian Process with a smooth covariance kernel described in the supplement. These true fields were then scaled to an RI range of 1 to 1.003. To put this RI magnitude into context, the refractive effect of these fields causes a median deviation of 0.03 to 0.07 pixels on a \\(64^2\\) pixel image for rays shot parallel through the volume from the front to the back." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.583 + ], + "angle": 0, + "content": "Fig. 5 shows that our method has the potential to accurately reconstruct complex, smooth fields. However, due to the ill-posedness and highly non-convex nature of this problem, our model is still susceptible to reconstruction artifacts, as seen in the last row. Note that no explicit regularization besides a boundary condition term was used in solving for these refractive fields. Thus, our results suggest that our neural network architecture combined with a low degree of positional encoding act as an implicit prior on the smoothness of our reconstruction, allowing it to effectively recover smooth refractive fields." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.593, + 0.803, + 0.609 + ], + "angle": 0, + "content": "5.3. Sensitivity to Light Source Density" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.763 + ], + "angle": 0, + "content": "From a single viewpoint, refractive fields are detected when they lie between light sources and the sensor. Background refractive structures become difficult to detect with a sparse array of light sources since fewer light sources lie behind them. Because we supervise our model with curved ray paths throughout the scene, our model uses the deflections between adjacent light sources to localize the refractive field. Therefore, we expect that as available light sources in the scene become sparser, our model will struggle to detect and localize the structure of a refractive field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We perform a refractive field reconstruction on a scene containing five elliptical refractive objects in light source emission fields of varying densities. We sample 250 light source locations uniformly, and create emission fields of decreasing density by taking random subsets of this initial sample. We find that reconstruction quality of our approach degrades with sparser light source fields, especially in the background portion of the field. However, even with only 50 light sources in the volume, our method is able to accu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25364" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.452, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.473, + 0.493 + ], + "angle": 0, + "content": "Figure 6. Sensitivity to Light Source Density. We analyze our approach's reconstruction of a refractive field given light source fields with a range of densities. As the number of light sources decreases, it becomes harder to detect and localize refractive objects, especially towards the back of the volume. However, even with sparse light sources, our method is able to accurately resolve refractive objects in the foreground. The yellow cones indicate sensor direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.471, + 0.527 + ], + "angle": 0, + "content": "rately resolve the structure of refractive objects in the foreground. Results are shown in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.532, + 0.446, + 0.549 + ], + "angle": 0, + "content": "6. Case Study: Dark Matter Mass Mapping" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.549, + 0.473, + 0.822 + ], + "angle": 0, + "content": "In this section we show an application to astronomical imaging of dark matter from simulations. In outer space, clumps of invisible mass called dark matter permeate the cosmic web. These clumps, called halos, are so massive that their gravitational pull warps the surrounding space, which lenses nearby passing rays of light. In the weak lensing regime, the effects of gravitational lensing are well-described by refraction from an \"effective\" RI field [26]. In fact, the degree of curvature in weak lensing is so small that its effect on galaxy images is often modeled by a rotation (rather than a translation) of the galaxy's appearance on the sky [5]. By combining telescope images with photometric redshift measurements of the distances of galaxies from Earth, scientists are able to determine their approximate spatial locations. The inverse problem of dark matter mass mapping is to reconstruct an invisible distribution of dark matter given refracted images from telescope surveys and photometric measurements of galaxy distances [15]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.414, + 0.849 + ], + "angle": 0, + "content": "6.1. Simulated Dark Matter Halo Recovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We now use our method to recover a realistic dark matter refractive field. To generate this refractive field, we use the cosmological hydrodynamical simulation IllustrisTNG," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "which includes both dark and baryonic matter [27, 29]. We use the catalog of dark matter halos produced by the TNG300-1 simulation, focusing on a \\(20^{3}\\) \\(\\mathrm{Mpc}^3\\) region containing 1747 massive halos. As galaxy placement is known to trace the dark matter distribution, we placed Gaussian emission functions emulating galaxies at the center of the 300 most massive halos. We aim to reconstruct a realistic dark matter distribution, but leave the inclusion of realistic measurement noise to future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.348 + ], + "angle": 0, + "content": "RI tomography in the weak lensing regime requires sensors with sufficient resolution to detect the small shearing effects from gravitational lensing. In our experiments, we use an increased sensor resolution of \\(512 \\times 512\\); this resolution in comparison with the scale of our scene is consistent with that of current and upcoming weak lensing surveys [18, 35]. We train our model using SGD, sampling batches of 4096 rays per iteration due to memory constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.893, + 0.575 + ], + "angle": 0, + "content": "Our reconstruction of the invisible dark matter field is shown in Fig. 1. Our method is able to accurately recover the structure of the dark matter field with slight damping on the largest peaks of halo centers. This recovery is especially challenging because of the distribution of galaxies throughout the scene; large patches of the image are dark as galaxies are not uniformly distributed throughout the volume. In addition, the amplitude of the refractive field generated by these halos is three orders of magnitude lower than in our previous results, resulting in a weak refractive signal; the median deflection caused by the refractive field is \\(6\\mathrm{e}-3\\) pixels. We find that despite these challenges, our method is able to recover a refractive field that correlates strongly with true simulated dark matter halos, and shows promise for application to 3D dark matter recovery in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.585, + 0.62, + 0.6 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.821 + ], + "angle": 0, + "content": "In this paper we present a novel method for the tomography of continuous refractive fields from single-view image measurements with neural fields. Our method succeeds at reconstructing simulated refractive fields with varying light source density. By using the known location of light emitters, our method reconstructs refractive volumes from a single viewpoint that rival those of multiview methods. Finally, our results on physically realistic simulations of dark matter show promise for the application of our method to real weak lensing survey data for dark matter mapping. Beyond dark matter fields, our method opens doors to reconstructing other refractive phenomena, underscoring the computer vision community's potential to drive significant advances in scientific discovery with computational imaging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements BZ was partially supported by the Gupta Fellowship. This work was supported by NSF award 2048237, Google, BP, and AWS Credits from the AI4Science Partnership. We also thank Arjun Teh, Adithya Pediredla and Andrew Chael for their helpful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25365" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Marina Alterman, Yoav Y Schechner, Minh Vo, and Srinivasa G Narasimhan. Passive tomography of turbulence strength. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 47-60. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Marina Alterman, Yohay Swirski, and Yoav Y Schechner. Stella maris: Stellar marine refractive imaging sensor. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[3] Marco Ament, Christoph Bergmann, and Daniel Weiskopf. Refractive radiative transfer equation. ACM Transactions on Graphics (TOG), 33(2):1-22, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.29, + 0.472, + 0.346 + ], + "angle": 0, + "content": "[4] Bradley Atcheson, Ivo Ihrke, Wolfgang Heidrich, Art Tevs, Derek Bradley, Marcus Magnor, and Hans-Peter Seidel. Time-resolved 3d capture of non-stationary gas flows. ACM transactions on graphics (TOG), 27(5):1-9, 2008. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.471, + 0.377 + ], + "angle": 0, + "content": "[5] Matthias Bartelmann and Peter Schneider. Weak gravitational lensing. Physics Reports, 340(4-5):291-472, 2001. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.471, + 0.435 + ], + "angle": 0, + "content": "[6] Mojtaba Bemana, Karol Myszkowski, Jeppe Revall Frisvad, Hans-Peter Seidel, and Tobias Ritschel. Eikonal fields for refractive novel-view synthesis. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[8] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 10(1):1, 2007. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.54, + 0.471, + 0.595 + ], + "angle": 0, + "content": "[9] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.598, + 0.47, + 0.654 + ], + "angle": 0, + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[11] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[12] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[13] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 6(1):19-26, 1980. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Henk Hoekstra and Bhuvnesh Jain. Weak gravitational lensing and its cosmological applications. Annual Review of Nuclear and Particle Science, 58:99-123, 2008. 1, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Ivo Ihrke, Gernot Ziegler, Art Tevs, Christian Theobalt, Marcus Magnor, and Hans-Peter Seidel. Eikonal rendering: Efficient light transport in refractive objects. ACM Transactions on Graphics (TOG), 26(3):59–es, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[18] Rene Laureejs, J Amiaux, S Arduini, J-L Augueres, J Brinchmann, R Cole, M Cropper, C Dabin, L Duvet, A Ealet, et al. Euclid definition study report. arXiv preprint arXiv:1110.3193, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[19] Adrienne Leonard, F-X Dupé, and J-L Starck. A compressed sensing approach to 3d weak lensing. *Astronomy & Astrophysics*, 539:A85, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.331, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[20] Adrienne Leonard, François Lanusse, and Jean-Luc Starck. Glimpse: accurate 3d weak lensing reconstructions using sparsity. Monthly Notices of the Royal Astronomical Society, 440(2):1281-1294, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[21] Aviad Levis, Pratul P Srinivasan, Andrew A Chael, Ren Ng, and Katherine L Bouman. Gravitationally lensed black hole emission tomography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19841-19850, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[22] Zongling Li, Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Jin Liu, and Wei Zhang. End-to-end learned single lens design using fast differentiable ray tracing. Optics Letters, 46(21): 5453-5456, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[23] Fan-Chi Lin, Michael H Ritzwoller, and Roel Snieder. Eikonal tomography: surface wave tomography by phase front tracking across a regional broad-band seismic array. Geophysical Journal International, 177(3):1091-1110, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.582, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[24] Renhao Liu, Yu Sun, Jiabei Zhu, Lei Tian, and Ulugbek S Kamilov. Recovery of continuous 3d refractive index maps from discrete intensity-only measurements using neural fields. Nature Machine Intelligence, 4(9):781-791, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[26] Ramesh Narayan and Matthias Bartelmann. Lectures on gravitational lensing. arXiv preprint astro-ph/9606001, 1996. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[27] Dylan Nelson, Annalisa Pillepich, Shy Genel, Mark Vogelsberger, Volker Springel, Paul Torrey, Vicente Rodriguez-Gomez, Debora Sijacki, Gregory F Snyder, Brendan Griffen, et al. The illustris simulation: Public data release. *Astronomy and Computing*, 13:12–37, 2015. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.833, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25366" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[29] Annalisa Pillepich, Volker Springel, Dylan Nelson, Shy Genel, Jill Naiman, Rüdiger Pakmor, Lars Hernquist, Paul Torrey, Mark Vogelsberger, Rainer Weinberger, et al. Simulating galaxy formation with the illustrstng model. Monthly Notices of the Royal Astronomical Society, 473(3):4077-4106, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[30] Lev Semenovich Pontryagin. Mathematical theory of optimal processes. CRC press, 1987. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[31] William A Reed, Man F Yan, and Mark J Schnitzer. Gradient-index fiber-optic microprobes for minimally invasive in vivo low-coherence interferometry. Optics letters, 27 (20):1794-1796, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[32] Peter Schneider, Jürgen Ehlers, Emilio E Falco, et al. Gravitational lenses [electronic resource]. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[33] Patrick Simon, AN Taylor, and Jan Hartlap. Unfolding the matter distribution using three-dimensional weak gravitational lensing. Monthly Notices of the Royal Astronomical Society, 399(1):48-68, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[34] Yongjin Sung, Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Ramachandra R Dasari, and Michael S Feld. Optical diffraction tomography for high resolution live cell imaging. Optics express, 17(1):266-277, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[35] Masahiro Takada. Subaru hyper supprime-cam project. In AIP Conference Proceedings, pages 120-127. American Institute of Physics, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.468, + 0.529 + ], + "angle": 0, + "content": "[36] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.532, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[37] Arjun Teh, Matthew O'Toole, and Ioannis Gkioulekas. Adjoint nonlinear ray tracing. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[38] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.66, + 0.468, + 0.713 + ], + "angle": 0, + "content": "[39] JT VanderPlas, AJ Connolly, Bhuvnesh Jain, and Mike Jarvis. Three-dimensional reconstruction of the density field: An svd approach to weak-lensing tomography. The Astrophysical Journal, 727(2):118, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.716, + 0.468, + 0.798 + ], + "angle": 0, + "content": "[40] Tianfan Xue, Michael Rubinstein, Neal Wadhwa, Anat Levin, Fredo Durand, and William T Freeman. Refraction wiggles for measuring fluid depth and velocity from video. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 767-782. Springer, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25367" + } + ] +] \ No newline at end of file diff --git a/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_origin.pdf b/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e882086ea414fc1efb670dc39de5a7b1a401c073 --- /dev/null +++ b/2024/Single View Refractive Index Tomography with Neural Fields/887b4199-788b-44cf-9fe5-91d50d30037c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bc10a20f434ae504bc370712fc3f3e6cadfacf686c1eb28cdccc384d7e4266b +size 1093680 diff --git a/2024/Single View Refractive Index Tomography with Neural Fields/full.md b/2024/Single View Refractive Index Tomography with Neural Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a09e322d660e1cc4fb101cacc639f6f6b6e9476c --- /dev/null +++ b/2024/Single View Refractive Index Tomography with Neural Fields/full.md @@ -0,0 +1,298 @@ +# Single View Refractive Index Tomography with Neural Fields + +Brandon Zhao1* Aviad Levis1 Liam Connor2 Pratul P. Srinivasan3 Katherine L. Bouman1,2 + +$^{1}$ Department of Computing and Mathematical Sciences, California Institute of Technology $^{2}$ Cahill Center for Astronomy and Astrophysics, California Institute of Technology $^{3}$ Google Research + +$^{*}$ byzhao@caltech.edu + +# Abstract + +Refractive Index Tomography is the inverse problem of reconstructing the continuously-varying 3D refractive index in a scene using 2D projected image measurements. Although a purely refractive field is not directly visible, it bends light rays as they travel through space, thus providing a signal for reconstruction. The effects of such fields appear in many scientific computer vision settings, ranging from refraction due to transparent cells in microscopy to the lensing of distant galaxies caused by dark matter in astrophysics. Reconstructing these fields is particularly difficult due to the complex nonlinear effects of the refractive field on observed images. Furthermore, while standard 3D reconstruction and tomography settings typically have access to observations of the scene from many viewpoints, many refractive index tomography problem settings only have access to images observed from a single viewpoint. We introduce a method that leverages prior knowledge of light sources scattered throughout the refractive medium to help disambiguate the single-view refractive index tomography problem. We differentiably trace curved rays through a neural field representation of the refractive field, and optimize its parameters to best reproduce the observed image. We demonstrate the efficacy of our approach by reconstructing simulated refractive fields, analyze the effects of light source distribution on the recovered field, and test our method on a simulated dark matter mapping problem where we successfully recover the 3D refractive field caused by a realistic dark matter distribution. + +# 1. Introduction + +When light passes from one medium to another, such as from air to water, it encounters a change in refractive index (RI). As a result, light rays are bent, or refracted, according to the magnitude of this change, causing effects like the shimmering (caustics) at the bottom of a pool of water. Media such as hot gas or translucent cell samples have RIs + +that continuously vary in space, causing light to take paths curved by the gradient of the underlying RI field. Although the RI field is not directly visible, its effect on the scene's illumination provides clues to its structure. + +Spatially varying RI fields are present in a diverse range of scientific applications. For example, certain fiber optic cables are designed to smoothly refract light away from the cable's outside padding [31]. In seismology, pressure waves from earthquakes are refracted continuously as they travel through ground structures of varying density [23]. In astrophysics, gravitational effects from massive dark matter structures lens the light from faraway galaxies resulting in observable effects in telescope images. When the degree of these gravitational lensing effects is relatively weak, they can be modeled as refraction [15]. + +In this work, we seek to tackle the problem of single-view RI tomography: 3D reconstruction of RI fields where image measurements are available only from a single viewpoint. This problem setting occurs when using telescope images or a single camera on scenes in the wild, and poses several challenges. First, image measurements are produced by rays traveling along continuously curving trajectories through the scene, where curvature is due to the structure of the underlying RI field. Unlike in linear tomography problems, where the projection geometry is independent of the medium, in RI tomography the ray paths are dictated by the unknown medium properties themselves. Thus the measurements are a non-linear function of the targeted refractive field. Second, limiting image measurements to a single viewing angle makes the problem severely ill-posed. That is, there are many possible refractive fields which could potentially produce the same image measurements. However, by leveraging the location of light sources distributed throughout the scene we can better constrain the recovery. + +We propose a novel approach using neural fields with continuous ray tracing that leverages the physics of light propagation to recover an underlying refractive field. In particular, we represent the spatially varying RI field with + +![](images/6a90c7c540a37ff2c8c3a97cb5384bfce87b5d87ca128443be0b2d777579b2f4.jpg) +Figure 1. We formulate and present an approach for single-view refractive field reconstruction that explicitly models curved ray paths through a neural field. By leveraging spatial information and the emission profile of light sources located throughout the volume, we are able to localize refractive structures in 3D with just a single viewpoint. Here we demonstrate our approach on a simulation involving a realistic refractive field induced by simulated dark matter halos. Despite the large patches of space that are void of galaxy emitters, our method is still able to recover the rough structure of the 3D dark matter distribution from a single simulated telescope image of refracted galaxies. This reconstruction is also especially challenging because of the extremely weak degree of refraction, as seen in the relative scale of the intensity difference plot; the refractive effects of dark matter result in an intensity change of at most $\pm 0.3\%$ of the maximum intensity of the original image. This experiment uses a realistic dark matter distribution derived from IllustrisTNG [27], but we leave the inclusion of realistic measurement noise to future work. + +a coordinate-based neural network, which we optimize through a differentiable rendering procedure that traces paths of light through the refractive field to best reproduce measurements. Coordinate-based neural networks are advantageous for this problem because they provide a continuous, easily-differentiable representation of the refractive field as a function of space. We also find empirically that coordinate-based neural networks lead to better optimization behavior out-of-the-box than other representations such as linearly interpolated grids. + +Our approach expands upon recent works in neural scene representation [21, 25]. While in these works ray paths through the scene are assumed to be known, our approach explicitly models and solves for unknown refracted ray paths. We first study the conditions under which single-view reconstruction is possible, then test our approach by analyzing its ability to recover weak refractive fields. Next, we analyze the sensitivity of our approach to the density of light sources in the scene. Finally, as shown in Fig. 1, we include a case study motivated by the application of weak lensing where our method recovers the refractive field generated by physically realistic simulations of dark matter. + +# 2. Related Work + +Image-Based Refractive Field Recovery The reconstruction of refractive objects and phenomena is an ongoing field of study, both in applications that make use of standard consumer cameras as well as scientific imaging applications. In Background-oriented Schlieren (BoS) imaging, images taken of a background with and without refractive effects are correlated to estimate an optical flow describing a projected measurement of the refractive field [4]. Combining these projected measurements from multiple viewpoints + +through tomography gives an estimate of the 3D gradients, from which the 3D field can be estimated from a Poisson solver. A similar type of tomography can be done to passively model 3D turbulence statistics by combining time resolved deflection measurements from multiple viewpoints [1]. Optical flow estimates such as those in BoS can also be used to detect refractive "wiggles" in fluid flow that can then be used to track the projected motion of the refractive media; by observing these wiggles with stereo cameras the fluid's depth can also be recovered [40]. Fluid refraction from water can also be estimated by submerged cameras viewing the sun as a guiding point [2]. + +While the aforementioned methods utilize the physics behind ray optics, they don't explicitly model the continuously curved paths that rays take through these refractive media during 3D reconstruction. A method using the refractive curvature of traced rays to supervise an interpolated grid model has been shown to be effective for reconstructing and designing refractive media [37]. However, as with the previous methods, this method still relies on access to multiple viewpoints to reconstruct the refractive volume. In this work, we instead seek to tackle the challenge of accurately reconstructing 3D refractive fields from a single viewpoint. + +Applications Recent works in optics have used differentiable rendering techniques to design refractive surfaces for various applications. The materials making up the lens typically have a constant RI, meaning light is refracted a discrete number of times at the interface between the lens and the surrounding air. Thus, these works solve for the optimal shape of the material rather than an RI field varying within the lens. Fast differentiable ray tracing techniques have been used for the design of a single lens to produce a + +target image [22], as well as compound lenses for camera design [38]. These types of lens designs are aimed at reproducing a target image or favorable image effects. While we also aim to optimize a refractive field to match a set of observed image measurements, we are interested in recovering the properties of the refractive medium itself. + +In biomedical imaging, Optical diffraction tomography (ODT) uses the intrinsic optical variation in a sample to reconstruct its RI field [11, 34]. Holographic measurements are taken from multiple illumination angles with coherent light modeled by wave optics. For example, in [12], the authors present a technique to recover RI by using intensity-only measurements from multiple viewpoints. DeCAF [24] uses a neural field representation to solve this inverse problem. Our problem setup differs from these biomedical applications in that we attempt to solve this problem without access to active illumination and coherent light sources. Additionally, we use ray optics rather than wave optics in our characterization of light transport. + +We are motivated in particular by the application of gravitational lensing, where light coming from small elliptical sources (galaxies) curves on its path to Earth due to an invisible dark matter distribution. In the weak lensing regime, the small ray deflections caused by dark matter can be well-approximated as a refractive effect [15], allowing us to model the large-scale structure of the dark matter distribution as an RI field. Existing works use a linear tomography approach similar to that of BoS [4] by first computing a 2D projected map of mass density, then backprojecting into 3D to obtain a reconstruction. Because image measurements are limited to a single viewpoint, filtered backprojection-style methods [33, 39] are susceptible to heavy smearing along the line of sight, even in the absence of noise. In [19, 20] an approach is used that assumes the density field can be sparsely represented by a dictionary of dark matter halos. However, this assumption is only approximately true in practice; our method can instead reconstruct 3D RI fields induced by a general matter distribution. + +Coordinate-based Neural Representations Recently coordinate-based neural networks have gained traction for applications in vision and graphics. These models typically parameterize a three-dimensional field, such as a scene's radiance, with the weights of a multi-layer perceptron (MLP). Coordinate-based neural representations have been used in graphics to model shapes with occupancy fields [10], as well as three-dimensional realistic scenes with neural radiance fields [25]. Both of these works rely on a fixed straight ray assumption to render scenes for training and inference. Coordinate-based MLPs have also been used in black hole imaging [21] to reconstruct emission fields with curved ray paths dictated by general relativity; however, the paths taken by rays of light are calculated then fixed beforehand according to a specific model of gravitational lens. In contrast, our + +work seeks to optimize ray paths that are unknown and directly estimated during model training. + +Eikonal ray-tracing with neural fields has been used for refractive tomography in the context of novel-view synthesis [6]. However, this work uses images of the scene from multiple viewpoints for supervision. Additionally, it focuses on recovering scenes which have a piecewise constant RI (e.g. a glass block with constant RI surrounded by air), where no light emission occurs within the refractive medium. While it achieves impressive results for rendering images from novel viewpoints, the reconstructed refractive indices of these volumes fail to match the ground truth. In contrast, our method leverages visible light sources from within the refractive medium to enable accurate reconstruction of a refractive volume from only a single viewpoint. + +# 3. Forward: Image Formation Model + +This section details the components of the image formation (forward) model. First, we introduce the equations governing the refraction and intensity measurements of individual rays. Next, we detail the numerical integration used to model image intensity measurements at a sensor. + +# 3.1. Refractive Ray Tracing + +Purely refractive media are translucent and affect light rays by curving their paths rather than modifying their intensity. Light travels more slowly through more optically dense material, causing it to bend towards an increase in optical density. The RI $\eta$ is a scalar used to quantify the optical density of a material relative to a vacuum; in natural settings, $\eta \geq 1$ as light cannot exceed its speed in a vacuum. In this work, we are interested in volumes with a continuously varying refractive index. The effect of a continuously varying RI $\eta$ on a light ray's position $x$ and direction $v$ is described by Hamilton's equations for ray tracing [16]: + +$$ +\left\{ \begin{array}{l} \frac {d x}{d s} = \frac {v (s)}{\eta (s)} \\ \frac {d v}{d s} = \nabla \eta (s) \end{array} \right. \tag {1} +$$ + +Here the derivatives of the 3D ray position $x$ and direction $v$ are taken with respect to the ray path length $s$ , and are described in terms of the RI $\eta$ and its gradient $\nabla \eta$ . The directional derivative of Eq. (2) can be understood as an attractive force towards local increases in RI. + +Given an initial position and direction $x_0, v_0 \in \mathbb{R}^3$ , equations (1) and (2) describe a first-order initial value problem whose solution $\Gamma(s) = (x, v)$ is a curve describing the trajectory of a ray as a function of its path length $s$ : + +$$ +\Gamma (s) = \left(x _ {0}, v _ {0}\right) + \int_ {0} ^ {s} \left(x ^ {\prime} (\tilde {s}), v ^ {\prime} (\tilde {s})\right) d \tilde {s} \tag {3} +$$ + +An illustration of refractive ray tracing is given in Figure 2. + +![](images/ca87681480d606e18ed1d8bcbadad1d0920025621061f76f133622d069721748.jpg) +Figure 2. Illustration of ray tracing in a refractive field. Ray paths are calculated iteratively by solving the Hamiltonian equations (1), (2) parameterized by the refractive field $\eta$ . Rays are curved locally towards areas of high RI in the direction of $\nabla \eta$ . + +# 3.2. Sensor and Light Emission Integration + +The sensor is modeled as a single-wavelength pinhole camera positioned outside of the scene. The intensity image $I$ of a given scene is represented as a discrete function indexed by pixel coordinates $i, j$ . The specific intensity of a pixel $I(i, j)$ is given by integration of light emission over a ray path solution $\Gamma_{i,j}$ , to Eq. (3). The initial conditions $\Gamma_{i,j}(0)$ are defined by the position and direction of a ray traveling backwards from the sensor through the pixel center. + +Within the scene, light emission $e(\mathbf{x})$ is represented as a continuous scalar function of a 3D coordinate $\mathbf{x} = (x,y,z)$ , and is assumed to be known. This assumption is commonly used for refractive modeling problems in the forms of calibration images in BoS imaging [4] or galaxy catalogs in dark matter mapping [15]. Given the ray path $\Gamma_{i,j}$ starting at the pixel in location $(i,j)$ , we model the specific intensity $I(i,j)$ by taking the line integral of emission coefficients $e$ along the path $\Gamma_{i,j}$ : + +$$ +I (i, j) = \int_ {\Gamma_ {i, j}} e (\mathbf {x} (s)) d s \tag {4} +$$ + +Eq. (4) corresponds to a simplified model of light transport in which the specific intensity is conserved along a ray's path. In gravitational lens theory [32], the emission term in the integrand is decreased for sources in areas of high refractive index due to gravitational redshift. However, for gravitational RI fields with $\eta$ close to 1, which we focus on in this work, this effect is negligible and we thus exclude it from our forward model [21]. In other applications of refractive tomography [6, 28], the refractive radiative transfer equation (RRTE) [3] may be more appropriate for the rendering forward model. Further details, including experiments with the RRTE, can be found in the supplement. + +# 3.3. Numerical Integration + +A ray path solution $\Gamma = (x, v)$ to the Hamiltonian initial value problem given by equations (1) and (2) is fully determined by the initial position and direction $\Gamma_{i,j}(0)$ of the ray traced through a pixel $(i, j)$ , the refractive field $\eta$ , and its gradient $\nabla \eta$ . Thus, we rewrite the integral in Eq. (3) as: + +$$ +\Gamma_ {i, j} = \operatorname {s o l v e H a m} \left(\Gamma_ {i, j} (0), \eta , \nabla \eta\right). \tag {5} +$$ + +Integrating the scene's emission function $e$ along the ray path $\Gamma$ as in Eq. (4) allows us to calculate image intensities: + +$$ +I (i, j) = \operatorname {r e n d e r} \operatorname {I m} \left(\Gamma_ {i, j}, e\right). \tag {6} +$$ + +In general, closed-form solutions for integrals (5) and (6) do not exist, so they are solved numerically. We use an adaptive step-size Dormand-Prince integrator [13] to solve Eqs. (5) and (6) in our experiments. Backpropagating through the integrator using autograd requires memory in the order of the number of solver steps. Instead, we follow previous works [6, 9, 37] in using the adjoint formulation [30] which has constant memory requirements for backpropagation. + +# 4. Inverse: Estimating a Refractive Field + +In this work, we focus on weakly refractive fields where the deflections of a single source are on the order of fractions of a pixel. Although this may seem small, we rely on the combined signal across an ensemble of lensed sources. + +To solve the inverse problem of refractive tomography in this setting, we use an analysis-by-synthesis approach. We model the continuous refractive field with a coordinate-based MLP, which is optimized to fit a true observed image. In the following section we describe the neural representation and optimization scheme. + +# 4.1. Continuous Refractive Field Representation + +To form a candidate continuously-varying 3D refractive field, we use a coordinate-based MLP inspired by the NeRF [25] model. The MLP, parameterized by weights $\theta$ , takes as input spatial coordinates $\mathbf{x}$ and outputs a scalar index of refraction. To ensure that our reconstruction is physically plausible, we apply the Softplus function $\sigma$ to restrict the output of our network to be greater than 1: + +$$ +\hat {\eta} _ {\theta} (\mathbf {x}) = \sigma (\operatorname {M L P} (\gamma (\mathbf {x})); \theta). \tag {7} +$$ + +Here, $\gamma (\mathbf{x})$ is a positional encoding layer mapping an input coordinate to a set of Fourier basis functions [36] with exponentially-increasing frequencies: + +$$ +\gamma (\mathbf {x}) = \left[ \sin (\mathbf {x}), \cos (\mathbf {x}), \dots , \sin \left(2 ^ {L - 1} \mathbf {x}\right), \cos \left(2 ^ {L - 1} \mathbf {x}\right) \right] ^ {T}. \tag {8} +$$ + +![](images/50beebf9300f58d88c684c04463ab757658491b6e7ce04c716119e72fb8ca110.jpg) +Figure 3. Refractive Tomography Pipeline. We model a 3D RI field $\hat{\eta}(\mathbf{x})$ as a continuous function using a neural network parameterized by $\theta$ . This refractive field, along with the known visible emission field $e(x)$ from point-like light sources, is fed through a ray tracing simulator to produce a predicted 2D image measurement. An image loss is then taken between the model and true image measurements to solve for $\theta$ . We differentiably trace rays through this refractive field and accumulate the light emitted along these curved ray paths to produce a 2D image. This allows us to optimize MLP parameters $\theta$ to minimize the loss between the rendered and true image measurements. + +Adding this positional encoding allows the MLP to fit continuous fields with higher spatial frequencies as the degree $L$ is increased [36]. We find empirically that a degree of $L = 4$ is suitable for naturally smooth refractive fields. + +Ray tracing through our model field requires integrating equation (5) during each training iteration. This requires evaluating the model $\hat{\eta}_{\theta}$ and its spatial gradient $\nabla \hat{\eta}_{\theta}$ at continuous spatial locations over the course of optimization, and for both of these functions to be differentiable with respect to the model parameters $\theta$ for gradient descent. In other words, we need to be able to evaluate $\nabla_{\theta} \nabla_{x} \hat{\eta}_{\theta}$ at arbitrary points throughout the modeled refractive field. We use the eLU activation function rather than a ReLU so that this gradient is smooth during optimization, and can be easily evaluated with existing autograd frameworks. + +# 4.2. Optimization + +Our neural field model defines a continuous refractive field $\hat{\eta}_{\theta}$ and its gradient $\nabla_{x}\hat{\eta}_{\theta}$ parameterized by its weights $\theta$ . Thus, a rendered image $\hat{I}$ obtained by ray-tracing through a neural field $\hat{\eta}_{\theta}$ with Eqs. (5) and (6) is a function of $\theta$ . + +An overview of our pipeline is shown in Fig. 3. To train the neural network weights $\theta$ , we solve a minimization problem whose objective is a mean-squared error between the observed image $I(i,j)$ , and the rendered image $\hat{I}(i,j;\theta)$ , plus a regularization term $\mathcal{R}(\theta)$ to enforce a boundary condition of $\eta = 1$ on the cube's faces: + +$$ +\min _ {\theta} \mathcal {L} (\theta) = \sum_ {i, j} | | I (i, j) - \hat {I} (i, j, \theta) | | _ {2} ^ {2} + \lambda \mathcal {R} (\theta). \tag {9} +$$ + +This loss is fully differentiable with respect to model parameters $\theta$ , allowing us to use gradient descent to solve the optimization. The term $\mathcal{R}(\theta)$ acts only as a constraint on the boundary conditions of the volume; unlike explicit regularizers, such as total variation, we rely on the implicit regu + +larization from the neural field for our reconstruction. More details on the $\mathcal{R}(\theta)$ term can be found in the supplement. + +# 4.3. Implementation Details + +In our experiments, we use an MLP with 4 layers, where each layer is 256 units wide. The neural network is implemented in JAX [7] and trained with the Adam optimizer [17] with exponential learning rate decay from $10^{-4}$ to $5 \times 10^{-6}$ over 10K iterations. The weights of the neural network were randomly initialized with the He uniform variance scaling initializer [14]. Code implementation will be publicly available on the project page. + +Unmodeled Sources of Noise Because our method leverages information about the location and shape of light sources to enable reconstruction from a single viewpoint, it is sensitive to noise in both our knowledge of light source locations and in image measurements. An analysis of our method's sensitivity to these noise sources can be found in the supplement. We plan to investigate ways to overcome this sensitivity in future work. + +# 5. Results + +In this section we present results of our method on general refractive tomography problems. We generate ground truth refractive fields inside of a three-dimensional volume, where the RI takes a constant value of 1 outside the region of interest. Each ground truth refractive field is defined by evaluating a continuous function at regularly spaced grid points. The continuous field is then evaluated by linearly interpolating between these grid points. Image measurements are rendered using a pinhole camera with a $64^2$ resolution. Throughout this section, we quantify the accuracy of our reconstructions by sampling the ground truth and reconstructed fields on the same uniformly spaced grid of sample points as the ground truth and report the PSNR. When calculating the PSNR, we first subtract 1 from the RI. + +![](images/20ce296905d1ca91806fdc5dad9b468905fbf9bce39e343dc3348d8d705214ca.jpg) +Light Source Distribution + +![](images/dbdb63f900c319f49d2b40ccc2ccd404564244c86ab36be8e1a290eed4e3b753.jpg) +Measurement Type + +![](images/a8284e3d2cdbdbb56f9e1e820434df2f897182218ff72b2d3ed9b00277cdfd28.jpg) +Neural Field + +![](images/79521e40691946946a6ac961187f6ac4c7ed709540ef65c35aef0c85d5abe52a.jpg) + +![](images/398a3e1052a5fa415585c90cbe88b7ccf87279cadcceb27f940c36a826172a61.jpg) + +![](images/465d011bdbeeb9880c9f6591f8c676885ced2e23818011520ca6194a62adf6c3.jpg) +Voxel Grid + +![](images/eb0591dd5f1a88e38f5fc0cf7fe8f137997402a1dfce0af0a85b041f0be1d7fb.jpg) + +![](images/8da236723c30692d2b5fb3e27267a58e42003fe9fac7fb2a0755494ea9e12d85.jpg) +Figure 4. Measurement and Reconstruction Conditions. We compare our neural field reconstruction model to a linear voxel grid with and without $\mathrm{TV}^2$ regularization for gas flow reconstruction under three different measurement conditions. While single-view recovery is difficult with measurements of ray displacement, we find that image measurements of light sources spread throughout the refractive volume makes recovery possible with our method. In addition, a neural field's implicit regularization results in better-constrained reconstructions than the linear voxel grid, even when an explicit regularizer is used. In each recovery we super-impose the ground-truth gas plume in blue over the reconstruction in yellow for comparison. We note that all reconstructions shown fit the measurements almost perfectly; differences in their structures highlight the ill-posedness of the problem. + +![](images/89b6cb89eff35bd796598340c44f980e73a4633a94ee243014c0ec5f5f9e406f.jpg) +Voxel Grid (TV2) + +![](images/b573e2257e1ebe3181cb417f9e9175be8f56f316ad97861398d0e30d6a5c90c7.jpg) + +![](images/a5f0e342e78680e8c1515c04631d65bbb238c62f54eb31fcf970b3e4790cf912.jpg) + +![](images/0cce2ddefadb8624086684af89f3fed9f814f0d3d93b961ce48019920844a4f0.jpg) +Ground Truth + +![](images/bed70aff96c6933919682dbca83f519d60d2246e06b626a4a19fdb39fd46e4f4.jpg) +Ground Truth + +![](images/be6835291efc219663b9d319a1460a0d2494ffa0fe9b9a553b2bfe3c5059617b.jpg) +Reconstruction + +# 5.1. Measurement and Reconstruction Conditions + +In this section we compare refractive reconstructions under nine conditions: three sets of measurement conditions paired with three different reconstruction strategies. We show that single-view RI tomography is made possible by a combination of the measurement conditions of our problem setting and the implicit regularization given by our neural field model. Results are shown in Fig. 4. + +Measurement Conditions: The first measurement conditions mirror the astronomical domain that motivates our work, in which image measurements are taken of light sources distributed throughout the refractive volume (Fig. 4 top row). Secondly, we compare to image measurements taken of light sources behind the volume (middle row). Third, we investigate using ray displacement measurements rather than image measurements; in this case, the displacements correspond to a plane of light sources behind the volume (bottom row). To simulate the displacement measurements, we ray trace through the ground truth refractive field to recover a perfect optical flow map at the same resolution as our sensor. Although this would be infeasible in practice, we use these measurements to represent the ideal limit of measurements from a single-view BoS setup [4]. + +Reconstruction Conditions: We also compare reconstructions from three models: our neural field representa + +tion, a linear voxel grid, and a linear voxel grid with $\mathrm{TV}^2$ regularization, with regularizer weight chosen to maximize PSNR with the ground truth volume. These correspond to the last three columns of Fig. 4, respectively. The goal of the regularized voxel grid is to show that our method can outperform even the highest-performing $\mathrm{TV}^2$ model; in practice, it would be infeasible to choose the regularizer weight in this way as we don't have access to the ground truth. + +Data & Results In Fig. 4 we demonstrate the effect of each pair of conditions in reconstructing the fuel injection dataset (a $64^{3}$ RI grid) from SFB 382 of the German Research Council (DFG). Both a method using an unregularized voxel grid as well as BoS have been shown to reconstruct this volume given access to multiple viewpoints [4, 37]. We report the PSNR for each reconstruction, but note that it isn't fully representative of reconstruction quality, which should also be evaluated visually. + +Effect of Measurement Conditions: When optimizing our neural field model on image measurements of light sources located throughout the refractive volume, accurate 3D recovery of the gas flow is possible with minimal artifacts (A). However, in (B), where the light sources are instead located behind the volume, the neural field is able to recover the rough structure of the gas flow but not its absolute location. When using displacement measurements, none of the reconstructions (C, F, I) are able to correctly recover the shape + +![](images/b320a93dbd033148251f51bd6286a96e51f43d17f4c788bdef89a029235626fd.jpg) +Figure 5. Recovery of Smooth Refractive Fields. Each row corresponds to a refractive tomography problem with a different random refractive field against a random distribution of 290 light sources. Although refractive recovery from a single viewpoint is highly ill-posed, our method can leverage source location information to detect and localize 3D refractive structure. However, as can be seen in the third example, due to the ill-posed and highly non-convex nature of this tomography problem, our model is still be susceptible to error. The yellow cones indicate sensor direction. + +of the gas flow from a single viewpoint, even with a perfect optical flow map. While it has been shown that voxel grid optimization with displacement measurements can accurately reconstruct this volume [37], having access to only a single viewpoint causes the reconstruction to be smeared along the optical axis. An analysis of the voxel grid recovery with access to additional viewpoints can be found in the supplement. We conclude that using image measurements of light sources spread throughout the volume is a key ingredient enabling 3D reconstruction from a single viewpoint. + +Effect of Reconstruction Conditions: An image-based loss on point-like light sources is highly nonconvex, and thus models such as an unregularized voxel grid can be prone to converging to bad minima. Although all reconstructions in Fig. 4 fit the measurements almost perfectly, each converges to a different refractive field. The unregularized grid reconstructions (D, E) optimize locally along ray paths to fit the image measurements, resulting in streaklike artifacts and a highly non-smooth recovery. Even when an explicit $\mathrm{TV}^2$ regularizer is used on the voxel grid, as in + +recoveries (G) and (H), it still fails to recover the structure of the gas flow along the optical axis. The implicit regularization of our neural field model appears to favor compact reconstructions, which matches the gas flow volume well. In the following section, we show that this implicit regularization also allows us to recover more general smooth fields. + +# 5.2. Recovery of Smooth Refractive Fields + +We demonstrate our approach by recovering randomly generated smooth refractive fields in a dense field of light sources from a single viewpoint. We use Poisson Disk sampling [8] to generate an emission field of 290 light source locations. Light sources are amplitude scaled 3D Gaussians with independent random elliptical orientations. Ground truth refractive fields are generated from a random Gaussian Process with a smooth covariance kernel described in the supplement. These true fields were then scaled to an RI range of 1 to 1.003. To put this RI magnitude into context, the refractive effect of these fields causes a median deviation of 0.03 to 0.07 pixels on a $64^2$ pixel image for rays shot parallel through the volume from the front to the back. + +Fig. 5 shows that our method has the potential to accurately reconstruct complex, smooth fields. However, due to the ill-posedness and highly non-convex nature of this problem, our model is still susceptible to reconstruction artifacts, as seen in the last row. Note that no explicit regularization besides a boundary condition term was used in solving for these refractive fields. Thus, our results suggest that our neural network architecture combined with a low degree of positional encoding act as an implicit prior on the smoothness of our reconstruction, allowing it to effectively recover smooth refractive fields. + +# 5.3. Sensitivity to Light Source Density + +From a single viewpoint, refractive fields are detected when they lie between light sources and the sensor. Background refractive structures become difficult to detect with a sparse array of light sources since fewer light sources lie behind them. Because we supervise our model with curved ray paths throughout the scene, our model uses the deflections between adjacent light sources to localize the refractive field. Therefore, we expect that as available light sources in the scene become sparser, our model will struggle to detect and localize the structure of a refractive field. + +We perform a refractive field reconstruction on a scene containing five elliptical refractive objects in light source emission fields of varying densities. We sample 250 light source locations uniformly, and create emission fields of decreasing density by taking random subsets of this initial sample. We find that reconstruction quality of our approach degrades with sparser light source fields, especially in the background portion of the field. However, even with only 50 light sources in the volume, our method is able to accu + +![](images/9855240e3cee6a5daa938b590fabd4a90715435fcf1df972c2e7fb46c14a63b5.jpg) +Figure 6. Sensitivity to Light Source Density. We analyze our approach's reconstruction of a refractive field given light source fields with a range of densities. As the number of light sources decreases, it becomes harder to detect and localize refractive objects, especially towards the back of the volume. However, even with sparse light sources, our method is able to accurately resolve refractive objects in the foreground. The yellow cones indicate sensor direction. + +rately resolve the structure of refractive objects in the foreground. Results are shown in Fig. 6. + +# 6. Case Study: Dark Matter Mass Mapping + +In this section we show an application to astronomical imaging of dark matter from simulations. In outer space, clumps of invisible mass called dark matter permeate the cosmic web. These clumps, called halos, are so massive that their gravitational pull warps the surrounding space, which lenses nearby passing rays of light. In the weak lensing regime, the effects of gravitational lensing are well-described by refraction from an "effective" RI field [26]. In fact, the degree of curvature in weak lensing is so small that its effect on galaxy images is often modeled by a rotation (rather than a translation) of the galaxy's appearance on the sky [5]. By combining telescope images with photometric redshift measurements of the distances of galaxies from Earth, scientists are able to determine their approximate spatial locations. The inverse problem of dark matter mass mapping is to reconstruct an invisible distribution of dark matter given refracted images from telescope surveys and photometric measurements of galaxy distances [15]. + +# 6.1. Simulated Dark Matter Halo Recovery + +We now use our method to recover a realistic dark matter refractive field. To generate this refractive field, we use the cosmological hydrodynamical simulation IllustrisTNG, + +which includes both dark and baryonic matter [27, 29]. We use the catalog of dark matter halos produced by the TNG300-1 simulation, focusing on a $20^{3}$ $\mathrm{Mpc}^3$ region containing 1747 massive halos. As galaxy placement is known to trace the dark matter distribution, we placed Gaussian emission functions emulating galaxies at the center of the 300 most massive halos. We aim to reconstruct a realistic dark matter distribution, but leave the inclusion of realistic measurement noise to future work. + +RI tomography in the weak lensing regime requires sensors with sufficient resolution to detect the small shearing effects from gravitational lensing. In our experiments, we use an increased sensor resolution of $512 \times 512$ ; this resolution in comparison with the scale of our scene is consistent with that of current and upcoming weak lensing surveys [18, 35]. We train our model using SGD, sampling batches of 4096 rays per iteration due to memory constraints. + +Our reconstruction of the invisible dark matter field is shown in Fig. 1. Our method is able to accurately recover the structure of the dark matter field with slight damping on the largest peaks of halo centers. This recovery is especially challenging because of the distribution of galaxies throughout the scene; large patches of the image are dark as galaxies are not uniformly distributed throughout the volume. In addition, the amplitude of the refractive field generated by these halos is three orders of magnitude lower than in our previous results, resulting in a weak refractive signal; the median deflection caused by the refractive field is $6\mathrm{e}-3$ pixels. We find that despite these challenges, our method is able to recover a refractive field that correlates strongly with true simulated dark matter halos, and shows promise for application to 3D dark matter recovery in the future. + +# 7. Conclusion + +In this paper we present a novel method for the tomography of continuous refractive fields from single-view image measurements with neural fields. Our method succeeds at reconstructing simulated refractive fields with varying light source density. By using the known location of light emitters, our method reconstructs refractive volumes from a single viewpoint that rival those of multiview methods. Finally, our results on physically realistic simulations of dark matter show promise for the application of our method to real weak lensing survey data for dark matter mapping. Beyond dark matter fields, our method opens doors to reconstructing other refractive phenomena, underscoring the computer vision community's potential to drive significant advances in scientific discovery with computational imaging. + +Acknowledgements BZ was partially supported by the Gupta Fellowship. This work was supported by NSF award 2048237, Google, BP, and AWS Credits from the AI4Science Partnership. We also thank Arjun Teh, Adithya Pediredla and Andrew Chael for their helpful discussions. + +# References + +[1] Marina Alterman, Yoav Y Schechner, Minh Vo, and Srinivasa G Narasimhan. Passive tomography of turbulence strength. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 47-60. Springer, 2014. 2 +[2] Marina Alterman, Yohay Swirski, and Yoav Y Schechner. Stella maris: Stellar marine refractive imaging sensor. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2014. 2 +[3] Marco Ament, Christoph Bergmann, and Daniel Weiskopf. Refractive radiative transfer equation. ACM Transactions on Graphics (TOG), 33(2):1-22, 2014. 4 +[4] Bradley Atcheson, Ivo Ihrke, Wolfgang Heidrich, Art Tevs, Derek Bradley, Marcus Magnor, and Hans-Peter Seidel. Time-resolved 3d capture of non-stationary gas flows. ACM transactions on graphics (TOG), 27(5):1-9, 2008. 2, 3, 4, 6 +[5] Matthias Bartelmann and Peter Schneider. Weak gravitational lensing. Physics Reports, 340(4-5):291-472, 2001. 8 +[6] Mojtaba Bemana, Karol Myszkowski, Jeppe Revall Frisvad, Hans-Peter Seidel, and Tobias Ritschel. Eikonal fields for refractive novel-view synthesis. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3, 4 +[7] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 5 +[8] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 10(1):1, 2007. 7 +[9] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 4 +[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3 +[11] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007. 3 +[12] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019. 3 +[13] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 6(1):19-26, 1980. 4 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5 + +[15] Henk Hoekstra and Bhuvnesh Jain. Weak gravitational lensing and its cosmological applications. Annual Review of Nuclear and Particle Science, 58:99-123, 2008. 1, 3, 4, 8 +[16] Ivo Ihrke, Gernot Ziegler, Art Tevs, Christian Theobalt, Marcus Magnor, and Hans-Peter Seidel. Eikonal rendering: Efficient light transport in refractive objects. ACM Transactions on Graphics (TOG), 26(3):59–es, 2007. 3 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[18] Rene Laureejs, J Amiaux, S Arduini, J-L Augueres, J Brinchmann, R Cole, M Cropper, C Dabin, L Duvet, A Ealet, et al. Euclid definition study report. arXiv preprint arXiv:1110.3193, 2011. 8 +[19] Adrienne Leonard, F-X Dupé, and J-L Starck. A compressed sensing approach to 3d weak lensing. *Astronomy & Astrophysics*, 539:A85, 2012. 3 +[20] Adrienne Leonard, François Lanusse, and Jean-Luc Starck. Glimpse: accurate 3d weak lensing reconstructions using sparsity. Monthly Notices of the Royal Astronomical Society, 440(2):1281-1294, 2014. 3 +[21] Aviad Levis, Pratul P Srinivasan, Andrew A Chael, Ren Ng, and Katherine L Bouman. Gravitationally lensed black hole emission tomography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19841-19850, 2022. 2, 3, 4 +[22] Zongling Li, Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Jin Liu, and Wei Zhang. End-to-end learned single lens design using fast differentiable ray tracing. Optics Letters, 46(21): 5453-5456, 2021. 3 +[23] Fan-Chi Lin, Michael H Ritzwoller, and Roel Snieder. Eikonal tomography: surface wave tomography by phase front tracking across a regional broad-band seismic array. Geophysical Journal International, 177(3):1091-1110, 2009. 1 +[24] Renhao Liu, Yu Sun, Jiabei Zhu, Lei Tian, and Ulugbek S Kamilov. Recovery of continuous 3d refractive index maps from discrete intensity-only measurements using neural fields. Nature Machine Intelligence, 4(9):781-791, 2022. 3 +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4 +[26] Ramesh Narayan and Matthias Bartelmann. Lectures on gravitational lensing. arXiv preprint astro-ph/9606001, 1996. 8 +[27] Dylan Nelson, Annalisa Pillepich, Shy Genel, Mark Vogelsberger, Volker Springel, Paul Torrey, Vicente Rodriguez-Gomez, Debora Sijacki, Gregory F Snyder, Brendan Griffen, et al. The illustris simulation: Public data release. *Astronomy and Computing*, 13:12–37, 2015. 2, 8 +[28] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 4 + +[29] Annalisa Pillepich, Volker Springel, Dylan Nelson, Shy Genel, Jill Naiman, Rüdiger Pakmor, Lars Hernquist, Paul Torrey, Mark Vogelsberger, Rainer Weinberger, et al. Simulating galaxy formation with the illustrstng model. Monthly Notices of the Royal Astronomical Society, 473(3):4077-4106, 2018. 8 +[30] Lev Semenovich Pontryagin. Mathematical theory of optimal processes. CRC press, 1987. 4 +[31] William A Reed, Man F Yan, and Mark J Schnitzer. Gradient-index fiber-optic microprobes for minimally invasive in vivo low-coherence interferometry. Optics letters, 27 (20):1794-1796, 2002. 1 +[32] Peter Schneider, Jürgen Ehlers, Emilio E Falco, et al. Gravitational lenses [electronic resource]. 4 +[33] Patrick Simon, AN Taylor, and Jan Hartlap. Unfolding the matter distribution using three-dimensional weak gravitational lensing. Monthly Notices of the Royal Astronomical Society, 399(1):48-68, 2009. 3 +[34] Yongjin Sung, Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Ramachandra R Dasari, and Michael S Feld. Optical diffraction tomography for high resolution live cell imaging. Optics express, 17(1):266-277, 2009. 3 +[35] Masahiro Takada. Subaru hyper supprime-cam project. In AIP Conference Proceedings, pages 120-127. American Institute of Physics, 2010. 8 +[36] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4, 5 +[37] Arjun Teh, Matthew O'Toole, and Ioannis Gkioulekas. Adjoint nonlinear ray tracing. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4, 6, 7 +[38] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 3 +[39] JT VanderPlas, AJ Connolly, Bhuvnesh Jain, and Mike Jarvis. Three-dimensional reconstruction of the density field: An svd approach to weak-lensing tomography. The Astrophysical Journal, 727(2):118, 2011. 3 +[40] Tianfan Xue, Michael Rubinstein, Neal Wadhwa, Anat Levin, Fredo Durand, and William T Freeman. Refraction wiggles for measuring fluid depth and velocity from video. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 767-782. Springer, 2014. 2 \ No newline at end of file diff --git a/2024/Single View Refractive Index Tomography with Neural Fields/images.zip b/2024/Single View Refractive Index Tomography with Neural Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec80a29ceadb44407051abb88acfee465216ddc1 --- /dev/null +++ b/2024/Single View Refractive Index Tomography with Neural Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:569aa26a27609177136b29b5d66fddfb5cc3308718f58bbed673e72539ca2572 +size 264327 diff --git a/2024/Single View Refractive Index Tomography with Neural Fields/layout.json b/2024/Single View Refractive Index Tomography with Neural Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f37b1e344fdf6d6d1147afd23eb4dd8e02445b2f --- /dev/null +++ b/2024/Single View Refractive Index Tomography with Neural Fields/layout.json @@ -0,0 +1,7911 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 487, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 487, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 487, + 121 + ], + "type": "text", + "content": "Single View Refractive Index Tomography with Neural Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 142, + 536, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 142, + 536, + 157 + ], + "spans": [ + { + "bbox": [ + 57, + 142, + 536, + 157 + ], + "type": "text", + "content": "Brandon Zhao1* Aviad Levis1 Liam Connor2 Pratul P. Srinivasan3 Katherine L. Bouman1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "spans": [ + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "type": "text", + "content": "Department of Computing and Mathematical Sciences, California Institute of Technology " + }, + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "type": "text", + "content": "Cahill Center for Astronomy and Astrophysics, California Institute of Technology " + }, + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 170, + 515, + 213 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 215, + 347, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 215, + 347, + 227 + ], + "spans": [ + { + "bbox": [ + 244, + 215, + 347, + 227 + ], + "type": "inline_equation", + "content": "^{*}" + }, + { + "bbox": [ + 244, + 215, + 347, + 227 + ], + "type": "text", + "content": "byzhao@caltech.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 280, + 290, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 280, + 290, + 616 + ], + "spans": [ + { + "bbox": [ + 45, + 280, + 290, + 616 + ], + "type": "text", + "content": "Refractive Index Tomography is the inverse problem of reconstructing the continuously-varying 3D refractive index in a scene using 2D projected image measurements. Although a purely refractive field is not directly visible, it bends light rays as they travel through space, thus providing a signal for reconstruction. The effects of such fields appear in many scientific computer vision settings, ranging from refraction due to transparent cells in microscopy to the lensing of distant galaxies caused by dark matter in astrophysics. Reconstructing these fields is particularly difficult due to the complex nonlinear effects of the refractive field on observed images. Furthermore, while standard 3D reconstruction and tomography settings typically have access to observations of the scene from many viewpoints, many refractive index tomography problem settings only have access to images observed from a single viewpoint. We introduce a method that leverages prior knowledge of light sources scattered throughout the refractive medium to help disambiguate the single-view refractive index tomography problem. We differentiably trace curved rays through a neural field representation of the refractive field, and optimize its parameters to best reproduce the observed image. We demonstrate the efficacy of our approach by reconstructing simulated refractive fields, analyze the effects of light source distribution on the recovered field, and test our method on a simulated dark matter mapping problem where we successfully recover the 3D refractive field caused by a realistic dark matter distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "When light passes from one medium to another, such as from air to water, it encounters a change in refractive index (RI). As a result, light rays are bent, or refracted, according to the magnitude of this change, causing effects like the shimmering (caustics) at the bottom of a pool of water. Media such as hot gas or translucent cell samples have RIs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 256, + 547, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 547, + 303 + ], + "type": "text", + "content": "that continuously vary in space, causing light to take paths curved by the gradient of the underlying RI field. Although the RI field is not directly visible, its effect on the scene's illumination provides clues to its structure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 437 + ], + "type": "text", + "content": "Spatially varying RI fields are present in a diverse range of scientific applications. For example, certain fiber optic cables are designed to smoothly refract light away from the cable's outside padding [31]. In seismology, pressure waves from earthquakes are refracted continuously as they travel through ground structures of varying density [23]. In astrophysics, gravitational effects from massive dark matter structures lens the light from faraway galaxies resulting in observable effects in telescope images. When the degree of these gravitational lensing effects is relatively weak, they can be modeled as refraction [15]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 548, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 548, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 548, + 665 + ], + "type": "text", + "content": "In this work, we seek to tackle the problem of single-view RI tomography: 3D reconstruction of RI fields where image measurements are available only from a single viewpoint. This problem setting occurs when using telescope images or a single camera on scenes in the wild, and poses several challenges. First, image measurements are produced by rays traveling along continuously curving trajectories through the scene, where curvature is due to the structure of the underlying RI field. Unlike in linear tomography problems, where the projection geometry is independent of the medium, in RI tomography the ray paths are dictated by the unknown medium properties themselves. Thus the measurements are a non-linear function of the targeted refractive field. Second, limiting image measurements to a single viewing angle makes the problem severely ill-posed. That is, there are many possible refractive fields which could potentially produce the same image measurements. However, by leveraging the location of light sources distributed throughout the scene we can better constrain the recovery." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "We propose a novel approach using neural fields with continuous ray tracing that leverages the physics of light propagation to recover an underlying refractive field. In particular, we represent the spatially varying RI field with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25358" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 79, + 534, + 198 + ], + "blocks": [ + { + "bbox": [ + 61, + 79, + 534, + 198 + ], + "lines": [ + { + "bbox": [ + 61, + 79, + 534, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 79, + 534, + 198 + ], + "type": "image", + "image_path": "6a90c7c540a37ff2c8c3a97cb5384bfce87b5d87ca128443be0b2d777579b2f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 545, + 312 + ], + "type": "text", + "content": "Figure 1. We formulate and present an approach for single-view refractive field reconstruction that explicitly models curved ray paths through a neural field. By leveraging spatial information and the emission profile of light sources located throughout the volume, we are able to localize refractive structures in 3D with just a single viewpoint. Here we demonstrate our approach on a simulation involving a realistic refractive field induced by simulated dark matter halos. Despite the large patches of space that are void of galaxy emitters, our method is still able to recover the rough structure of the 3D dark matter distribution from a single simulated telescope image of refracted galaxies. This reconstruction is also especially challenging because of the extremely weak degree of refraction, as seen in the relative scale of the intensity difference plot; the refractive effects of dark matter result in an intensity change of at most " + }, + { + "bbox": [ + 46, + 213, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\pm 0.3\\%" + }, + { + "bbox": [ + 46, + 213, + 545, + 312 + ], + "type": "text", + "content": " of the maximum intensity of the original image. This experiment uses a realistic dark matter distribution derived from IllustrisTNG [27], but we leave the inclusion of realistic measurement noise to future work." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 319, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 437 + ], + "type": "text", + "content": "a coordinate-based neural network, which we optimize through a differentiable rendering procedure that traces paths of light through the refractive field to best reproduce measurements. Coordinate-based neural networks are advantageous for this problem because they provide a continuous, easily-differentiable representation of the refractive field as a function of space. We also find empirically that coordinate-based neural networks lead to better optimization behavior out-of-the-box than other representations such as linearly interpolated grids." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 439, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 582 + ], + "type": "text", + "content": "Our approach expands upon recent works in neural scene representation [21, 25]. While in these works ray paths through the scene are assumed to be known, our approach explicitly models and solves for unknown refracted ray paths. We first study the conditions under which single-view reconstruction is possible, then test our approach by analyzing its ability to recover weak refractive fields. Next, we analyze the sensitivity of our approach to the density of light sources in the scene. Finally, as shown in Fig. 1, we include a case study motivated by the application of weak lensing where our method recovers the refractive field generated by physically realistic simulations of dark matter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 588, + 134, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 134, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 134, + 600 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Image-Based Refractive Field Recovery The reconstruction of refractive objects and phenomena is an ongoing field of study, both in applications that make use of standard consumer cameras as well as scientific imaging applications. In Background-oriented Schlieren (BoS) imaging, images taken of a background with and without refractive effects are correlated to estimate an optical flow describing a projected measurement of the refractive field [4]. Combining these projected measurements from multiple viewpoints" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "type": "text", + "content": "through tomography gives an estimate of the 3D gradients, from which the 3D field can be estimated from a Poisson solver. A similar type of tomography can be done to passively model 3D turbulence statistics by combining time resolved deflection measurements from multiple viewpoints [1]. Optical flow estimates such as those in BoS can also be used to detect refractive \"wiggles\" in fluid flow that can then be used to track the projected motion of the refractive media; by observing these wiggles with stereo cameras the fluid's depth can also be recovered [40]. Fluid refraction from water can also be estimated by submerged cameras viewing the sun as a guiding point [2]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 463, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 594 + ], + "type": "text", + "content": "While the aforementioned methods utilize the physics behind ray optics, they don't explicitly model the continuously curved paths that rays take through these refractive media during 3D reconstruction. A method using the refractive curvature of traced rays to supervise an interpolated grid model has been shown to be effective for reconstructing and designing refractive media [37]. However, as with the previous methods, this method still relies on access to multiple viewpoints to reconstruct the refractive volume. In this work, we instead seek to tackle the challenge of accurately reconstructing 3D refractive fields from a single viewpoint." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Applications Recent works in optics have used differentiable rendering techniques to design refractive surfaces for various applications. The materials making up the lens typically have a constant RI, meaning light is refracted a discrete number of times at the interface between the lens and the surrounding air. Thus, these works solve for the optimal shape of the material rather than an RI field varying within the lens. Fast differentiable ray tracing techniques have been used for the design of a single lens to produce a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25359" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "target image [22], as well as compound lenses for camera design [38]. These types of lens designs are aimed at reproducing a target image or favorable image effects. While we also aim to optimize a refractive field to match a set of observed image measurements, we are interested in recovering the properties of the refractive medium itself." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 299 + ], + "type": "text", + "content": "In biomedical imaging, Optical diffraction tomography (ODT) uses the intrinsic optical variation in a sample to reconstruct its RI field [11, 34]. Holographic measurements are taken from multiple illumination angles with coherent light modeled by wave optics. For example, in [12], the authors present a technique to recover RI by using intensity-only measurements from multiple viewpoints. DeCAF [24] uses a neural field representation to solve this inverse problem. Our problem setup differs from these biomedical applications in that we attempt to solve this problem without access to active illumination and coherent light sources. Additionally, we use ray optics rather than wave optics in our characterization of light transport." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 286, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 525 + ], + "type": "text", + "content": "We are motivated in particular by the application of gravitational lensing, where light coming from small elliptical sources (galaxies) curves on its path to Earth due to an invisible dark matter distribution. In the weak lensing regime, the small ray deflections caused by dark matter can be well-approximated as a refractive effect [15], allowing us to model the large-scale structure of the dark matter distribution as an RI field. Existing works use a linear tomography approach similar to that of BoS [4] by first computing a 2D projected map of mass density, then backprojecting into 3D to obtain a reconstruction. Because image measurements are limited to a single viewpoint, filtered backprojection-style methods [33, 39] are susceptible to heavy smearing along the line of sight, even in the absence of noise. In [19, 20] an approach is used that assumes the density field can be sparsely represented by a dictionary of dark matter halos. However, this assumption is only approximately true in practice; our method can instead reconstruct 3D RI fields induced by a general matter distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "type": "text", + "content": "Coordinate-based Neural Representations Recently coordinate-based neural networks have gained traction for applications in vision and graphics. These models typically parameterize a three-dimensional field, such as a scene's radiance, with the weights of a multi-layer perceptron (MLP). Coordinate-based neural representations have been used in graphics to model shapes with occupancy fields [10], as well as three-dimensional realistic scenes with neural radiance fields [25]. Both of these works rely on a fixed straight ray assumption to render scenes for training and inference. Coordinate-based MLPs have also been used in black hole imaging [21] to reconstruct emission fields with curved ray paths dictated by general relativity; however, the paths taken by rays of light are calculated then fixed beforehand according to a specific model of gravitational lens. In contrast, our" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "work seeks to optimize ray paths that are unknown and directly estimated during model training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 96, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 251 + ], + "type": "text", + "content": "Eikonal ray-tracing with neural fields has been used for refractive tomography in the context of novel-view synthesis [6]. However, this work uses images of the scene from multiple viewpoints for supervision. Additionally, it focuses on recovering scenes which have a piecewise constant RI (e.g. a glass block with constant RI surrounded by air), where no light emission occurs within the refractive medium. While it achieves impressive results for rendering images from novel viewpoints, the reconstructed refractive indices of these volumes fail to match the ground truth. In contrast, our method leverages visible light sources from within the refractive medium to enable accurate reconstruction of a refractive volume from only a single viewpoint." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 261, + 498, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 498, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 498, + 274 + ], + "type": "text", + "content": "3. Forward: Image Formation Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 281, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 340 + ], + "type": "text", + "content": "This section details the components of the image formation (forward) model. First, we introduce the equations governing the refraction and intensity measurements of individual rays. Next, we detail the numerical integration used to model image intensity measurements at a sensor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 346, + 438, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 438, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 438, + 358 + ], + "type": "text", + "content": "3.1. Refractive Ray Tracing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": "Purely refractive media are translucent and affect light rays by curving their paths rather than modifying their intensity. Light travels more slowly through more optically dense material, causing it to bend towards an increase in optical density. The RI " + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": " is a scalar used to quantify the optical density of a material relative to a vacuum; in natural settings, " + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\eta \\geq 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": " as light cannot exceed its speed in a vacuum. In this work, we are interested in volumes with a continuously varying refractive index. The effect of a continuously varying RI " + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": " on a light ray's position " + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": " is described by Hamilton's equations for ray tracing [16]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 394, + 505, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 505, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 394, + 505, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\frac {d x}{d s} = \\frac {v (s)}{\\eta (s)} \\\\ \\frac {d v}{d s} = \\nabla \\eta (s) \\end{array} \\right. \\tag {1}", + "image_path": "fd0b46fd3058f2d0594b5be14ca889c14b62867930e94fa4d79a50fa409adb7b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "text", + "content": "Here the derivatives of the 3D ray position " + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "text", + "content": " are taken with respect to the ray path length " + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "text", + "content": ", and are described in terms of the RI " + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "text", + "content": " and its gradient " + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\nabla \\eta" + }, + { + "bbox": [ + 304, + 558, + 545, + 616 + ], + "type": "text", + "content": ". The directional derivative of Eq. (2) can be understood as an attractive force towards local increases in RI." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": "Given an initial position and direction " + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "inline_equation", + "content": "x_0, v_0 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": ", equations (1) and (2) describe a first-order initial value problem whose solution " + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\Gamma(s) = (x, v)" + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": " is a curve describing the trajectory of a ray as a function of its path length " + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 616, + 545, + 665 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 346, + 673, + 545, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 673, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 346, + 673, + 545, + 700 + ], + "type": "interline_equation", + "content": "\\Gamma (s) = \\left(x _ {0}, v _ {0}\\right) + \\int_ {0} ^ {s} \\left(x ^ {\\prime} (\\tilde {s}), v ^ {\\prime} (\\tilde {s})\\right) d \\tilde {s} \\tag {3}", + "image_path": "8defb6dde7fbf9565f76709349051bad7a9c1da04df6ee4db05041a7ac970795.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "An illustration of refractive ray tracing is given in Figure 2." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25360" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 74, + 265, + 217 + ], + "blocks": [ + { + "bbox": [ + 69, + 74, + 265, + 217 + ], + "lines": [ + { + "bbox": [ + 69, + 74, + 265, + 217 + ], + "spans": [ + { + "bbox": [ + 69, + 74, + 265, + 217 + ], + "type": "image", + "image_path": "ca87681480d606e18ed1d8bcbadad1d0920025621061f76f133622d069721748.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "type": "text", + "content": "Figure 2. Illustration of ray tracing in a refractive field. Ray paths are calculated iteratively by solving the Hamiltonian equations (1), (2) parameterized by the refractive field " + }, + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "type": "text", + "content": ". Rays are curved locally towards areas of high RI in the direction of " + }, + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\nabla \\eta" + }, + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 285, + 251, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 251, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 251, + 298 + ], + "type": "text", + "content": "3.2. Sensor and Light Emission Integration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "content": "The sensor is modeled as a single-wavelength pinhole camera positioned outside of the scene. The intensity image " + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "content": " of a given scene is represented as a discrete function indexed by pixel coordinates " + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "content": ". The specific intensity of a pixel " + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "inline_equation", + "content": "I(i, j)" + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "content": " is given by integration of light emission over a ray path solution " + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\Gamma_{i,j}" + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "content": ", to Eq. (3). The initial conditions " + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\Gamma_{i,j}(0)" + }, + { + "bbox": [ + 46, + 305, + 287, + 402 + ], + "type": "text", + "content": " are defined by the position and direction of a ray traveling backwards from the sensor through the pixel center." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": "Within the scene, light emission " + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "inline_equation", + "content": "e(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": " is represented as a continuous scalar function of a 3D coordinate " + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y,z)" + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": ", and is assumed to be known. This assumption is commonly used for refractive modeling problems in the forms of calibration images in BoS imaging [4] or galaxy catalogs in dark matter mapping [15]. Given the ray path " + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\Gamma_{i,j}" + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": " starting at the pixel in location " + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": ", we model the specific intensity " + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "inline_equation", + "content": "I(i,j)" + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": " by taking the line integral of emission coefficients " + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": " along the path " + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\Gamma_{i,j}" + }, + { + "bbox": [ + 46, + 403, + 287, + 513 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 531, + 287, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 531, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 114, + 531, + 287, + 559 + ], + "type": "interline_equation", + "content": "I (i, j) = \\int_ {\\Gamma_ {i, j}} e (\\mathbf {x} (s)) d s \\tag {4}", + "image_path": "1e95037763d1ad100883892c7142cca1422e19e216bfef0718fedd4dbbe82204.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Eq. (4) corresponds to a simplified model of light transport in which the specific intensity is conserved along a ray's path. In gravitational lens theory [32], the emission term in the integrand is decreased for sources in areas of high refractive index due to gravitational redshift. However, for gravitational RI fields with " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " close to 1, which we focus on in this work, this effect is negligible and we thus exclude it from our forward model [21]. In other applications of refractive tomography [6, 28], the refractive radiative transfer equation (RRTE) [3] may be more appropriate for the rendering forward model. Further details, including experiments with the RRTE, can be found in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 434, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 434, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 434, + 85 + ], + "type": "text", + "content": "3.3. Numerical Integration" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "text", + "content": "A ray path solution " + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "inline_equation", + "content": "\\Gamma = (x, v)" + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "text", + "content": " to the Hamiltonian initial value problem given by equations (1) and (2) is fully determined by the initial position and direction " + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "inline_equation", + "content": "\\Gamma_{i,j}(0)" + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "text", + "content": " of the ray traced through a pixel " + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "inline_equation", + "content": "(i, j)" + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "text", + "content": ", the refractive field " + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "text", + "content": ", and its gradient " + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "inline_equation", + "content": "\\nabla \\eta" + }, + { + "bbox": [ + 305, + 90, + 545, + 151 + ], + "type": "text", + "content": ". Thus, we rewrite the integral in Eq. (3) as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 163, + 545, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 163, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 351, + 163, + 545, + 177 + ], + "type": "interline_equation", + "content": "\\Gamma_ {i, j} = \\operatorname {s o l v e H a m} \\left(\\Gamma_ {i, j} (0), \\eta , \\nabla \\eta\\right). \\tag {5}", + "image_path": "174c78207ed3b5a6a01928e39456efa535327c7757f8d6987935e459918702b4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 186, + 545, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 211 + ], + "type": "text", + "content": "Integrating the scene's emission function " + }, + { + "bbox": [ + 305, + 186, + 545, + 211 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 305, + 186, + 545, + 211 + ], + "type": "text", + "content": " along the ray path " + }, + { + "bbox": [ + 305, + 186, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 305, + 186, + 545, + 211 + ], + "type": "text", + "content": " as in Eq. (4) allows us to calculate image intensities:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 363, + 223, + 545, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 223, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 363, + 223, + 545, + 237 + ], + "type": "interline_equation", + "content": "I (i, j) = \\operatorname {r e n d e r} \\operatorname {I m} \\left(\\Gamma_ {i, j}, e\\right). \\tag {6}", + "image_path": "f7ab7b51684d89168c3262be61276ec0a039ccb239d9a31168028cef00661a1b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 342 + ], + "type": "text", + "content": "In general, closed-form solutions for integrals (5) and (6) do not exist, so they are solved numerically. We use an adaptive step-size Dormand-Prince integrator [13] to solve Eqs. (5) and (6) in our experiments. Backpropagating through the integrator using autograd requires memory in the order of the number of solver steps. Instead, we follow previous works [6, 9, 37] in using the adjoint formulation [30] which has constant memory requirements for backpropagation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 353, + 516, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 516, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 516, + 367 + ], + "type": "text", + "content": "4. Inverse: Estimating a Refractive Field" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": "In this work, we focus on weakly refractive fields where the deflections of a single source are on the order of fractions of a pixel. Although this may seem small, we rely on the combined signal across an ensemble of lensed sources." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 422, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 494 + ], + "type": "text", + "content": "To solve the inverse problem of refractive tomography in this setting, we use an analysis-by-synthesis approach. We model the continuous refractive field with a coordinate-based MLP, which is optimized to fit a true observed image. In the following section we describe the neural representation and optimization scheme." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 501, + 534, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 534, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 534, + 514 + ], + "type": "text", + "content": "4.1. Continuous Refractive Field Representation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "text", + "content": "To form a candidate continuously-varying 3D refractive field, we use a coordinate-based MLP inspired by the NeRF [25] model. The MLP, parameterized by weights " + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "text", + "content": ", takes as input spatial coordinates " + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "text", + "content": " and outputs a scalar index of refraction. To ensure that our reconstruction is physically plausible, we apply the Softplus function " + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "text", + "content": " to restrict the output of our network to be greater than 1:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 368, + 617, + 545, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 617, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 368, + 617, + 545, + 630 + ], + "type": "interline_equation", + "content": "\\hat {\\eta} _ {\\theta} (\\mathbf {x}) = \\sigma (\\operatorname {M L P} (\\gamma (\\mathbf {x})); \\theta). \\tag {7}", + "image_path": "156700fe6fd686b7829300205ad8027fa3dad170a8afde132ce41a97ce475e8b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " is a positional encoding layer mapping an input coordinate to a set of Fourier basis functions [36] with exponentially-increasing frequencies:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 685, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 685, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 310, + 685, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\gamma (\\mathbf {x}) = \\left[ \\sin (\\mathbf {x}), \\cos (\\mathbf {x}), \\dots , \\sin \\left(2 ^ {L - 1} \\mathbf {x}\\right), \\cos \\left(2 ^ {L - 1} \\mathbf {x}\\right) \\right] ^ {T}. \\tag {8}", + "image_path": "aa7ed729e7b9f4d092e294af6d3b59cf67f835cc468a978aca241601a45ec0b2.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25361" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 77, + 509, + 191 + ], + "blocks": [ + { + "bbox": [ + 74, + 77, + 509, + 191 + ], + "lines": [ + { + "bbox": [ + 74, + 77, + 509, + 191 + ], + "spans": [ + { + "bbox": [ + 74, + 77, + 509, + 191 + ], + "type": "image", + "image_path": "50beebf9300f58d88c684c04463ab757658491b6e7ce04c716119e72fb8ca110.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "text", + "content": "Figure 3. Refractive Tomography Pipeline. We model a 3D RI field " + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "inline_equation", + "content": "\\hat{\\eta}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "text", + "content": " as a continuous function using a neural network parameterized by " + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "text", + "content": ". This refractive field, along with the known visible emission field " + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "inline_equation", + "content": "e(x)" + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "text", + "content": " from point-like light sources, is fed through a ray tracing simulator to produce a predicted 2D image measurement. An image loss is then taken between the model and true image measurements to solve for " + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "text", + "content": ". We differentiably trace rays through this refractive field and accumulate the light emitted along these curved ray paths to produce a 2D image. This allows us to optimize MLP parameters " + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 202, + 547, + 259 + ], + "type": "text", + "content": " to minimize the loss between the rendered and true image measurements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 265, + 288, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 288, + 313 + ], + "type": "text", + "content": "Adding this positional encoding allows the MLP to fit continuous fields with higher spatial frequencies as the degree " + }, + { + "bbox": [ + 47, + 265, + 288, + 313 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 265, + 288, + 313 + ], + "type": "text", + "content": " is increased [36]. We find empirically that a degree of " + }, + { + "bbox": [ + 47, + 265, + 288, + 313 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 47, + 265, + 288, + 313 + ], + "type": "text", + "content": " is suitable for naturally smooth refractive fields." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "text", + "content": "Ray tracing through our model field requires integrating equation (5) during each training iteration. This requires evaluating the model " + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\hat{\\eta}_{\\theta}" + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "text", + "content": " and its spatial gradient " + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\nabla \\hat{\\eta}_{\\theta}" + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "text", + "content": " at continuous spatial locations over the course of optimization, and for both of these functions to be differentiable with respect to the model parameters " + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "text", + "content": " for gradient descent. In other words, we need to be able to evaluate " + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta} \\nabla_{x} \\hat{\\eta}_{\\theta}" + }, + { + "bbox": [ + 46, + 314, + 289, + 446 + ], + "type": "text", + "content": " at arbitrary points throughout the modeled refractive field. We use the eLU activation function rather than a ReLU so that this gradient is smooth during optimization, and can be easily evaluated with existing autograd frameworks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 455, + 133, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 133, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 133, + 467 + ], + "type": "text", + "content": "4.2. Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": "Our neural field model defines a continuous refractive field " + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\eta}_{\\theta}" + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": " and its gradient " + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\hat{\\eta}_{\\theta}" + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": " parameterized by its weights " + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": ". Thus, a rendered image " + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": " obtained by ray-tracing through a neural field " + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\eta}_{\\theta}" + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": " with Eqs. (5) and (6) is a function of " + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 474, + 287, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "text", + "content": "An overview of our pipeline is shown in Fig. 3. To train the neural network weights " + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "text", + "content": ", we solve a minimization problem whose objective is a mean-squared error between the observed image " + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "inline_equation", + "content": "I(i,j)" + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "text", + "content": ", and the rendered image " + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\hat{I}(i,j;\\theta)" + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "text", + "content": ", plus a regularization term " + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\theta)" + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "text", + "content": " to enforce a boundary condition of " + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "inline_equation", + "content": "\\eta = 1" + }, + { + "bbox": [ + 46, + 522, + 288, + 594 + ], + "type": "text", + "content": " on the cube's faces:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 616, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 616, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 61, + 616, + 287, + 643 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} (\\theta) = \\sum_ {i, j} | | I (i, j) - \\hat {I} (i, j, \\theta) | | _ {2} ^ {2} + \\lambda \\mathcal {R} (\\theta). \\tag {9}", + "image_path": "44877d479697cccf5ee0e6e90e758921c6c3e53a98cd60527f0ed108042e6646.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "This loss is fully differentiable with respect to model parameters " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", allowing us to use gradient descent to solve the optimization. The term " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\theta)" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " acts only as a constraint on the boundary conditions of the volume; unlike explicit regularizers, such as total variation, we rely on the implicit regu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 265, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 289 + ], + "type": "text", + "content": "larization from the neural field for our reconstruction. More details on the " + }, + { + "bbox": [ + 305, + 265, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\theta)" + }, + { + "bbox": [ + 305, + 265, + 545, + 289 + ], + "type": "text", + "content": " term can be found in the supplement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 297, + 440, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 440, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 440, + 311 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "text", + "content": "In our experiments, we use an MLP with 4 layers, where each layer is 256 units wide. The neural network is implemented in JAX [7] and trained with the Adam optimizer [17] with exponential learning rate decay from " + }, + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-6}" + }, + { + "bbox": [ + 304, + 318, + 545, + 415 + ], + "type": "text", + "content": " over 10K iterations. The weights of the neural network were randomly initialized with the He uniform variance scaling initializer [14]. Code implementation will be publicly available on the project page." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 423, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 519 + ], + "type": "text", + "content": "Unmodeled Sources of Noise Because our method leverages information about the location and shape of light sources to enable reconstruction from a single viewpoint, it is sensitive to noise in both our knowledge of light source locations and in image measurements. An analysis of our method's sensitivity to these noise sources can be found in the supplement. We plan to investigate ways to overcome this sensitivity in future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 528, + 359, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 359, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 359, + 540 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "In this section we present results of our method on general refractive tomography problems. We generate ground truth refractive fields inside of a three-dimensional volume, where the RI takes a constant value of 1 outside the region of interest. Each ground truth refractive field is defined by evaluating a continuous function at regularly spaced grid points. The continuous field is then evaluated by linearly interpolating between these grid points. Image measurements are rendered using a pinhole camera with a " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "64^2" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " resolution. Throughout this section, we quantify the accuracy of our reconstructions by sampling the ground truth and reconstructed fields on the same uniformly spaced grid of sample points as the ground truth and report the PSNR. When calculating the PSNR, we first subtract 1 from the RI." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25362" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 99, + 132, + 167 + ], + "blocks": [ + { + "bbox": [ + 75, + 76, + 118, + 95 + ], + "lines": [ + { + "bbox": [ + 75, + 76, + 118, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 76, + 118, + 95 + ], + "type": "text", + "content": "Light Source Distribution" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 99, + 132, + 167 + ], + "lines": [ + { + "bbox": [ + 61, + 99, + 132, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 99, + 132, + 167 + ], + "type": "image", + "image_path": "20ce296905d1ca91806fdc5dad9b468905fbf9bce39e343dc3348d8d705214ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 106, + 209, + 157 + ], + "blocks": [ + { + "bbox": [ + 157, + 76, + 206, + 95 + ], + "lines": [ + { + "bbox": [ + 157, + 76, + 206, + 95 + ], + "spans": [ + { + "bbox": [ + 157, + 76, + 206, + 95 + ], + "type": "text", + "content": "Measurement Type" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 106, + 209, + 157 + ], + "lines": [ + { + "bbox": [ + 154, + 106, + 209, + 157 + ], + "spans": [ + { + "bbox": [ + 154, + 106, + 209, + 157 + ], + "type": "image", + "image_path": "dbdb63f900c319f49d2b40ccc2ccd404564244c86ab36be8e1a290eed4e3b753.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 230, + 95, + 297, + 167 + ], + "blocks": [ + { + "bbox": [ + 242, + 80, + 285, + 90 + ], + "lines": [ + { + "bbox": [ + 242, + 80, + 285, + 90 + ], + "spans": [ + { + "bbox": [ + 242, + 80, + 285, + 90 + ], + "type": "text", + "content": "Neural Field" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 95, + 297, + 167 + ], + "lines": [ + { + "bbox": [ + 230, + 95, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 230, + 95, + 297, + 167 + ], + "type": "image", + "image_path": "a8284e3d2cdbdbb56f9e1e820434df2f897182218ff72b2d3ed9b00277cdfd28.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 236, + 169, + 292, + 235 + ], + "blocks": [ + { + "bbox": [ + 236, + 169, + 292, + 235 + ], + "lines": [ + { + "bbox": [ + 236, + 169, + 292, + 235 + ], + "spans": [ + { + "bbox": [ + 236, + 169, + 292, + 235 + ], + "type": "image", + "image_path": "79521e40691946946a6ac961187f6ac4c7ed709540ef65c35aef0c85d5abe52a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 236, + 239, + 293, + 303 + ], + "blocks": [ + { + "bbox": [ + 236, + 239, + 293, + 303 + ], + "lines": [ + { + "bbox": [ + 236, + 239, + 293, + 303 + ], + "spans": [ + { + "bbox": [ + 236, + 239, + 293, + 303 + ], + "type": "image", + "image_path": "398a3e1052a5fa415585c90cbe88b7ccf87279cadcceb27f940c36a826172a61.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 100, + 375, + 163 + ], + "blocks": [ + { + "bbox": [ + 329, + 80, + 365, + 89 + ], + "lines": [ + { + "bbox": [ + 329, + 80, + 365, + 89 + ], + "spans": [ + { + "bbox": [ + 329, + 80, + 365, + 89 + ], + "type": "text", + "content": "Voxel Grid" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 100, + 375, + 163 + ], + "lines": [ + { + "bbox": [ + 318, + 100, + 375, + 163 + ], + "spans": [ + { + "bbox": [ + 318, + 100, + 375, + 163 + ], + "type": "image", + "image_path": "465d011bdbeeb9880c9f6591f8c676885ced2e23818011520ca6194a62adf6c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 170, + 375, + 235 + ], + "blocks": [ + { + "bbox": [ + 319, + 170, + 375, + 235 + ], + "lines": [ + { + "bbox": [ + 319, + 170, + 375, + 235 + ], + "spans": [ + { + "bbox": [ + 319, + 170, + 375, + 235 + ], + "type": "image", + "image_path": "eb0591dd5f1a88e38f5fc0cf7fe8f137997402a1dfce0af0a85b041f0be1d7fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 240, + 375, + 304 + ], + "blocks": [ + { + "bbox": [ + 320, + 240, + 375, + 304 + ], + "lines": [ + { + "bbox": [ + 320, + 240, + 375, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 240, + 375, + 304 + ], + "type": "image", + "image_path": "8da236723c30692d2b5fb3e27267a58e42003fe9fac7fb2a0755494ea9e12d85.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 322, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 545, + 399 + ], + "type": "text", + "content": "Figure 4. Measurement and Reconstruction Conditions. We compare our neural field reconstruction model to a linear voxel grid with and without " + }, + { + "bbox": [ + 46, + 322, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{TV}^2" + }, + { + "bbox": [ + 46, + 322, + 545, + 399 + ], + "type": "text", + "content": " regularization for gas flow reconstruction under three different measurement conditions. While single-view recovery is difficult with measurements of ray displacement, we find that image measurements of light sources spread throughout the refractive volume makes recovery possible with our method. In addition, a neural field's implicit regularization results in better-constrained reconstructions than the linear voxel grid, even when an explicit regularizer is used. In each recovery we super-impose the ground-truth gas plume in blue over the reconstruction in yellow for comparison. We note that all reconstructions shown fit the measurements almost perfectly; differences in their structures highlight the ill-posedness of the problem." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 398, + 101, + 454, + 163 + ], + "blocks": [ + { + "bbox": [ + 399, + 80, + 453, + 90 + ], + "lines": [ + { + "bbox": [ + 399, + 80, + 453, + 90 + ], + "spans": [ + { + "bbox": [ + 399, + 80, + 453, + 90 + ], + "type": "text", + "content": "Voxel Grid (TV2)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 398, + 101, + 454, + 163 + ], + "lines": [ + { + "bbox": [ + 398, + 101, + 454, + 163 + ], + "spans": [ + { + "bbox": [ + 398, + 101, + 454, + 163 + ], + "type": "image", + "image_path": "89b6cb89eff35bd796598340c44f980e73a4633a94ee243014c0ec5f5f9e406f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 398, + 171, + 453, + 234 + ], + "blocks": [ + { + "bbox": [ + 398, + 171, + 453, + 234 + ], + "lines": [ + { + "bbox": [ + 398, + 171, + 453, + 234 + ], + "spans": [ + { + "bbox": [ + 398, + 171, + 453, + 234 + ], + "type": "image", + "image_path": "b573e2257e1ebe3181cb417f9e9175be8f56f316ad97861398d0e30d6a5c90c7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 398, + 240, + 453, + 303 + ], + "blocks": [ + { + "bbox": [ + 398, + 240, + 453, + 303 + ], + "lines": [ + { + "bbox": [ + 398, + 240, + 453, + 303 + ], + "spans": [ + { + "bbox": [ + 398, + 240, + 453, + 303 + ], + "type": "image", + "image_path": "a5f0e342e78680e8c1515c04631d65bbb238c62f54eb31fcf970b3e4790cf912.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 471, + 152, + 528, + 209 + ], + "blocks": [ + { + "bbox": [ + 477, + 140, + 524, + 148 + ], + "lines": [ + { + "bbox": [ + 477, + 140, + 524, + 148 + ], + "spans": [ + { + "bbox": [ + 477, + 140, + 524, + 148 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 152, + 528, + 209 + ], + "lines": [ + { + "bbox": [ + 471, + 152, + 528, + 209 + ], + "spans": [ + { + "bbox": [ + 471, + 152, + 528, + 209 + ], + "type": "image", + "image_path": "0cce2ddefadb8624086684af89f3fed9f814f0d3d93b961ce48019920844a4f0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 473, + 225, + 485, + 233 + ], + "blocks": [ + { + "bbox": [ + 473, + 225, + 485, + 233 + ], + "lines": [ + { + "bbox": [ + 473, + 225, + 485, + 233 + ], + "spans": [ + { + "bbox": [ + 473, + 225, + 485, + 233 + ], + "type": "image", + "image_path": "bed70aff96c6933919682dbca83f519d60d2246e06b626a4a19fdb39fd46e4f4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 226, + 525, + 233 + ], + "lines": [ + { + "bbox": [ + 486, + 226, + 525, + 233 + ], + "spans": [ + { + "bbox": [ + 486, + 226, + 525, + 233 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 473, + 239, + 485, + 247 + ], + "blocks": [ + { + "bbox": [ + 473, + 239, + 485, + 247 + ], + "lines": [ + { + "bbox": [ + 473, + 239, + 485, + 247 + ], + "spans": [ + { + "bbox": [ + 473, + 239, + 485, + 247 + ], + "type": "image", + "image_path": "be6835291efc219663b9d319a1460a0d2494ffa0fe9b9a553b2bfe3c5059617b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 239, + 531, + 247 + ], + "lines": [ + { + "bbox": [ + 486, + 239, + 531, + 247 + ], + "spans": [ + { + "bbox": [ + 486, + 239, + 531, + 247 + ], + "type": "text", + "content": "Reconstruction" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 401, + 282, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 282, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 282, + 414 + ], + "type": "text", + "content": "5.1. Measurement and Reconstruction Conditions" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 422, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 506 + ], + "type": "text", + "content": "In this section we compare refractive reconstructions under nine conditions: three sets of measurement conditions paired with three different reconstruction strategies. We show that single-view RI tomography is made possible by a combination of the measurement conditions of our problem setting and the implicit regularization given by our neural field model. Results are shown in Fig. 4." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 508, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 687 + ], + "type": "text", + "content": "Measurement Conditions: The first measurement conditions mirror the astronomical domain that motivates our work, in which image measurements are taken of light sources distributed throughout the refractive volume (Fig. 4 top row). Secondly, we compare to image measurements taken of light sources behind the volume (middle row). Third, we investigate using ray displacement measurements rather than image measurements; in this case, the displacements correspond to a plane of light sources behind the volume (bottom row). To simulate the displacement measurements, we ray trace through the ground truth refractive field to recover a perfect optical flow map at the same resolution as our sensor. Although this would be infeasible in practice, we use these measurements to represent the ideal limit of measurements from a single-view BoS setup [4]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Reconstruction Conditions: We also compare reconstructions from three models: our neural field representa" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 402, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 498 + ], + "type": "text", + "content": "tion, a linear voxel grid, and a linear voxel grid with " + }, + { + "bbox": [ + 304, + 402, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{TV}^2" + }, + { + "bbox": [ + 304, + 402, + 545, + 498 + ], + "type": "text", + "content": " regularization, with regularizer weight chosen to maximize PSNR with the ground truth volume. These correspond to the last three columns of Fig. 4, respectively. The goal of the regularized voxel grid is to show that our method can outperform even the highest-performing " + }, + { + "bbox": [ + 304, + 402, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{TV}^2" + }, + { + "bbox": [ + 304, + 402, + 545, + 498 + ], + "type": "text", + "content": " model; in practice, it would be infeasible to choose the regularizer weight in this way as we don't have access to the ground truth." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 498, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 605 + ], + "type": "text", + "content": "Data & Results In Fig. 4 we demonstrate the effect of each pair of conditions in reconstructing the fuel injection dataset (a " + }, + { + "bbox": [ + 304, + 498, + 546, + 605 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 304, + 498, + 546, + 605 + ], + "type": "text", + "content": " RI grid) from SFB 382 of the German Research Council (DFG). Both a method using an unregularized voxel grid as well as BoS have been shown to reconstruct this volume given access to multiple viewpoints [4, 37]. We report the PSNR for each reconstruction, but note that it isn't fully representative of reconstruction quality, which should also be evaluated visually." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "Effect of Measurement Conditions: When optimizing our neural field model on image measurements of light sources located throughout the refractive volume, accurate 3D recovery of the gas flow is possible with minimal artifacts (A). However, in (B), where the light sources are instead located behind the volume, the neural field is able to recover the rough structure of the gas flow but not its absolute location. When using displacement measurements, none of the reconstructions (C, F, I) are able to correctly recover the shape" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25363" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 79, + 265, + 348 + ], + "blocks": [ + { + "bbox": [ + 65, + 79, + 265, + 348 + ], + "lines": [ + { + "bbox": [ + 65, + 79, + 265, + 348 + ], + "spans": [ + { + "bbox": [ + 65, + 79, + 265, + 348 + ], + "type": "image", + "image_path": "b320a93dbd033148251f51bd6286a96e51f43d17f4c788bdef89a029235626fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 364, + 289, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 364, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 289, + 464 + ], + "type": "text", + "content": "Figure 5. Recovery of Smooth Refractive Fields. Each row corresponds to a refractive tomography problem with a different random refractive field against a random distribution of 290 light sources. Although refractive recovery from a single viewpoint is highly ill-posed, our method can leverage source location information to detect and localize 3D refractive structure. However, as can be seen in the third example, due to the ill-posed and highly non-convex nature of this tomography problem, our model is still be susceptible to error. The yellow cones indicate sensor direction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 470, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 590 + ], + "type": "text", + "content": "of the gas flow from a single viewpoint, even with a perfect optical flow map. While it has been shown that voxel grid optimization with displacement measurements can accurately reconstruct this volume [37], having access to only a single viewpoint causes the reconstruction to be smeared along the optical axis. An analysis of the voxel grid recovery with access to additional viewpoints can be found in the supplement. We conclude that using image measurements of light sources spread throughout the volume is a key ingredient enabling 3D reconstruction from a single viewpoint." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Effect of Reconstruction Conditions: An image-based loss on point-like light sources is highly nonconvex, and thus models such as an unregularized voxel grid can be prone to converging to bad minima. Although all reconstructions in Fig. 4 fit the measurements almost perfectly, each converges to a different refractive field. The unregularized grid reconstructions (D, E) optimize locally along ray paths to fit the image measurements, resulting in streaklike artifacts and a highly non-smooth recovery. Even when an explicit " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TV}^2" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " regularizer is used on the voxel grid, as in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "recoveries (G) and (H), it still fails to recover the structure of the gas flow along the optical axis. The implicit regularization of our neural field model appears to favor compact reconstructions, which matches the gas flow volume well. In the following section, we show that this implicit regularization also allows us to recover more general smooth fields." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 155, + 505, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 505, + 167 + ], + "type": "text", + "content": "5.2. Recovery of Smooth Refractive Fields" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 175, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 330 + ], + "type": "text", + "content": "We demonstrate our approach by recovering randomly generated smooth refractive fields in a dense field of light sources from a single viewpoint. We use Poisson Disk sampling [8] to generate an emission field of 290 light source locations. Light sources are amplitude scaled 3D Gaussians with independent random elliptical orientations. Ground truth refractive fields are generated from a random Gaussian Process with a smooth covariance kernel described in the supplement. These true fields were then scaled to an RI range of 1 to 1.003. To put this RI magnitude into context, the refractive effect of these fields causes a median deviation of 0.03 to 0.07 pixels on a " + }, + { + "bbox": [ + 304, + 175, + 545, + 330 + ], + "type": "inline_equation", + "content": "64^2" + }, + { + "bbox": [ + 304, + 175, + 545, + 330 + ], + "type": "text", + "content": " pixel image for rays shot parallel through the volume from the front to the back." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 331, + 546, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 461 + ], + "type": "text", + "content": "Fig. 5 shows that our method has the potential to accurately reconstruct complex, smooth fields. However, due to the ill-posedness and highly non-convex nature of this problem, our model is still susceptible to reconstruction artifacts, as seen in the last row. Note that no explicit regularization besides a boundary condition term was used in solving for these refractive fields. Thus, our results suggest that our neural network architecture combined with a low degree of positional encoding act as an implicit prior on the smoothness of our reconstruction, allowing it to effectively recover smooth refractive fields." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 469, + 491, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 469, + 491, + 482 + ], + "spans": [ + { + "bbox": [ + 305, + 469, + 491, + 482 + ], + "type": "text", + "content": "5.3. Sensitivity to Light Source Density" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "content": "From a single viewpoint, refractive fields are detected when they lie between light sources and the sensor. Background refractive structures become difficult to detect with a sparse array of light sources since fewer light sources lie behind them. Because we supervise our model with curved ray paths throughout the scene, our model uses the deflections between adjacent light sources to localize the refractive field. Therefore, we expect that as available light sources in the scene become sparser, our model will struggle to detect and localize the structure of a refractive field." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We perform a refractive field reconstruction on a scene containing five elliptical refractive objects in light source emission fields of varying densities. We sample 250 light source locations uniformly, and create emission fields of decreasing density by taking random subsets of this initial sample. We find that reconstruction quality of our approach degrades with sparser light source fields, especially in the background portion of the field. However, even with only 50 light sources in the volume, our method is able to accu" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25364" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 276, + 291 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 276, + 291 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 276, + 291 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 276, + 291 + ], + "type": "image", + "image_path": "9855240e3cee6a5daa938b590fabd4a90715435fcf1df972c2e7fb46c14a63b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 289, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 289, + 390 + ], + "type": "text", + "content": "Figure 6. Sensitivity to Light Source Density. We analyze our approach's reconstruction of a refractive field given light source fields with a range of densities. As the number of light sources decreases, it becomes harder to detect and localize refractive objects, especially towards the back of the volume. However, even with sparse light sources, our method is able to accurately resolve refractive objects in the foreground. The yellow cones indicate sensor direction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 392, + 288, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 417 + ], + "type": "text", + "content": "rately resolve the structure of refractive objects in the foreground. Results are shown in Fig. 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 421, + 272, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 272, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 272, + 434 + ], + "type": "text", + "content": "6. Case Study: Dark Matter Mass Mapping" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 434, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 434, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 45, + 434, + 289, + 651 + ], + "type": "text", + "content": "In this section we show an application to astronomical imaging of dark matter from simulations. In outer space, clumps of invisible mass called dark matter permeate the cosmic web. These clumps, called halos, are so massive that their gravitational pull warps the surrounding space, which lenses nearby passing rays of light. In the weak lensing regime, the effects of gravitational lensing are well-described by refraction from an \"effective\" RI field [26]. In fact, the degree of curvature in weak lensing is so small that its effect on galaxy images is often modeled by a rotation (rather than a translation) of the galaxy's appearance on the sky [5]. By combining telescope images with photometric redshift measurements of the distances of galaxies from Earth, scientists are able to determine their approximate spatial locations. The inverse problem of dark matter mass mapping is to reconstruct an invisible distribution of dark matter given refracted images from telescope surveys and photometric measurements of galaxy distances [15]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 253, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 253, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 253, + 672 + ], + "type": "text", + "content": "6.1. Simulated Dark Matter Halo Recovery" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We now use our method to recover a realistic dark matter refractive field. To generate this refractive field, we use the cosmological hydrodynamical simulation IllustrisTNG," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "which includes both dark and baryonic matter [27, 29]. We use the catalog of dark matter halos produced by the TNG300-1 simulation, focusing on a " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "20^{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{Mpc}^3" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " region containing 1747 massive halos. As galaxy placement is known to trace the dark matter distribution, we placed Gaussian emission functions emulating galaxies at the center of the 300 most massive halos. We aim to reconstruct a realistic dark matter distribution, but leave the inclusion of realistic measurement noise to future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "type": "text", + "content": "RI tomography in the weak lensing regime requires sensors with sufficient resolution to detect the small shearing effects from gravitational lensing. In our experiments, we use an increased sensor resolution of " + }, + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 180, + 545, + 275 + ], + "type": "text", + "content": "; this resolution in comparison with the scale of our scene is consistent with that of current and upcoming weak lensing surveys [18, 35]. We train our model using SGD, sampling batches of 4096 rays per iteration due to memory constraints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 275, + 546, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 546, + 455 + ], + "type": "text", + "content": "Our reconstruction of the invisible dark matter field is shown in Fig. 1. Our method is able to accurately recover the structure of the dark matter field with slight damping on the largest peaks of halo centers. This recovery is especially challenging because of the distribution of galaxies throughout the scene; large patches of the image are dark as galaxies are not uniformly distributed throughout the volume. In addition, the amplitude of the refractive field generated by these halos is three orders of magnitude lower than in our previous results, resulting in a weak refractive signal; the median deflection caused by the refractive field is " + }, + { + "bbox": [ + 304, + 275, + 546, + 455 + ], + "type": "inline_equation", + "content": "6\\mathrm{e}-3" + }, + { + "bbox": [ + 304, + 275, + 546, + 455 + ], + "type": "text", + "content": " pixels. We find that despite these challenges, our method is able to recover a refractive field that correlates strongly with true simulated dark matter halos, and shows promise for application to 3D dark matter recovery in the future." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 463, + 379, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 379, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 379, + 475 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 482, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 650 + ], + "type": "text", + "content": "In this paper we present a novel method for the tomography of continuous refractive fields from single-view image measurements with neural fields. Our method succeeds at reconstructing simulated refractive fields with varying light source density. By using the known location of light emitters, our method reconstructs refractive volumes from a single viewpoint that rival those of multiview methods. Finally, our results on physically realistic simulations of dark matter show promise for the application of our method to real weak lensing survey data for dark matter mapping. Beyond dark matter fields, our method opens doors to reconstructing other refractive phenomena, underscoring the computer vision community's potential to drive significant advances in scientific discovery with computational imaging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements BZ was partially supported by the Gupta Fellowship. This work was supported by NSF award 2048237, Google, BP, and AWS Credits from the AI4Science Partnership. We also thank Arjun Teh, Adithya Pediredla and Andrew Chael for their helpful discussions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25365" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Marina Alterman, Yoav Y Schechner, Minh Vo, and Srinivasa G Narasimhan. Passive tomography of turbulence strength. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 47-60. Springer, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Marina Alterman, Yohay Swirski, and Yoav Y Schechner. Stella maris: Stellar marine refractive imaging sensor. In 2014 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 227 + ], + "type": "text", + "content": "[3] Marco Ament, Christoph Bergmann, and Daniel Weiskopf. Refractive radiative transfer equation. ACM Transactions on Graphics (TOG), 33(2):1-22, 2014. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 229, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 229, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 229, + 288, + 274 + ], + "type": "text", + "content": "[4] Bradley Atcheson, Ivo Ihrke, Wolfgang Heidrich, Art Tevs, Derek Bradley, Marcus Magnor, and Hans-Peter Seidel. Time-resolved 3d capture of non-stationary gas flows. ACM transactions on graphics (TOG), 27(5):1-9, 2008. 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 298 + ], + "type": "text", + "content": "[5] Matthias Bartelmann and Peter Schneider. Weak gravitational lensing. Physics Reports, 340(4-5):291-472, 2001. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 300, + 288, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 288, + 344 + ], + "type": "text", + "content": "[6] Mojtaba Bemana, Karol Myszkowski, Jeppe Revall Frisvad, Hans-Peter Seidel, and Tobias Ritschel. Eikonal fields for refractive novel-view synthesis. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "text", + "content": "[7] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 425 + ], + "type": "text", + "content": "[8] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 10(1):1, 2007. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 427, + 288, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 427, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 53, + 427, + 288, + 471 + ], + "type": "text", + "content": "[9] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 473, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 473, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 473, + 287, + 517 + ], + "type": "text", + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[11] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "type": "text", + "content": "[12] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 623, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 288, + 655 + ], + "type": "text", + "content": "[13] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 6(1):19-26, 1980. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Henk Hoekstra and Bhuvnesh Jain. Weak gravitational lensing and its cosmological applications. Annual Review of Nuclear and Particle Science, 58:99-123, 2008. 1, 3, 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[16] Ivo Ihrke, Gernot Ziegler, Art Tevs, Christian Theobalt, Marcus Magnor, and Hans-Peter Seidel. Eikonal rendering: Efficient light transport in refractive objects. ACM Transactions on Graphics (TOG), 26(3):59–es, 2007. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 185, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 227 + ], + "type": "text", + "content": "[18] Rene Laureejs, J Amiaux, S Arduini, J-L Augueres, J Brinchmann, R Cole, M Cropper, C Dabin, L Duvet, A Ealet, et al. Euclid definition study report. arXiv preprint arXiv:1110.3193, 2011. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": "[19] Adrienne Leonard, F-X Dupé, and J-L Starck. A compressed sensing approach to 3d weak lensing. *Astronomy & Astrophysics*, 539:A85, 2012. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 262, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 304 + ], + "type": "text", + "content": "[20] Adrienne Leonard, François Lanusse, and Jean-Luc Starck. Glimpse: accurate 3d weak lensing reconstructions using sparsity. Monthly Notices of the Royal Astronomical Society, 440(2):1281-1294, 2014. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 306, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 306, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 545, + 360 + ], + "type": "text", + "content": "[21] Aviad Levis, Pratul P Srinivasan, Andrew A Chael, Ren Ng, and Katherine L Bouman. Gravitationally lensed black hole emission tomography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19841-19850, 2022. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "type": "text", + "content": "[22] Zongling Li, Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Jin Liu, and Wei Zhang. End-to-end learned single lens design using fast differentiable ray tracing. Optics Letters, 46(21): 5453-5456, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 405, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 458 + ], + "type": "text", + "content": "[23] Fan-Chi Lin, Michael H Ritzwoller, and Roel Snieder. Eikonal tomography: surface wave tomography by phase front tracking across a regional broad-band seismic array. Geophysical Journal International, 177(3):1091-1110, 2009. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 460, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 513 + ], + "type": "text", + "content": "[24] Renhao Liu, Yu Sun, Jiabei Zhu, Lei Tian, and Ulugbek S Kamilov. Recovery of continuous 3d refractive index maps from discrete intensity-only measurements using neural fields. Nature Machine Intelligence, 4(9):781-791, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 571, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 602 + ], + "type": "text", + "content": "[26] Ramesh Narayan and Matthias Bartelmann. Lectures on gravitational lensing. arXiv preprint astro-ph/9606001, 1996. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 658 + ], + "type": "text", + "content": "[27] Dylan Nelson, Annalisa Pillepich, Shy Genel, Mark Vogelsberger, Volker Springel, Paul Torrey, Vicente Rodriguez-Gomez, Debora Sijacki, Gregory F Snyder, Brendan Griffen, et al. The illustris simulation: Public data release. *Astronomy and Computing*, 13:12–37, 2015. 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 659, + 545, + 712 + ], + "type": "text", + "content": "[28] Adithya Pediredla, Yasin Karimi Chalmiani, Matteo Giuseppe Scopelliti, Maysamreza Chamanzar, Srinivasa Narasimhan, and Ioannis Gkioulekas. Path tracing estimators for refractive radiative transfer. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25366" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 632 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[29] Annalisa Pillepich, Volker Springel, Dylan Nelson, Shy Genel, Jill Naiman, Rüdiger Pakmor, Lars Hernquist, Paul Torrey, Mark Vogelsberger, Rainer Weinberger, et al. Simulating galaxy formation with the illustrstng model. Monthly Notices of the Royal Astronomical Society, 473(3):4077-4106, 2018. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 140, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 286, + 161 + ], + "type": "text", + "content": "[30] Lev Semenovich Pontryagin. Mathematical theory of optimal processes. CRC press, 1987. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "type": "text", + "content": "[31] William A Reed, Man F Yan, and Mark J Schnitzer. Gradient-index fiber-optic microprobes for minimally invasive in vivo low-coherence interferometry. Optics letters, 27 (20):1794-1796, 2002. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 286, + 229 + ], + "type": "text", + "content": "[32] Peter Schneider, Jürgen Ehlers, Emilio E Falco, et al. Gravitational lenses [electronic resource]. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "text", + "content": "[33] Patrick Simon, AN Taylor, and Jan Hartlap. Unfolding the matter distribution using three-dimensional weak gravitational lensing. Monthly Notices of the Royal Astronomical Society, 399(1):48-68, 2009. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 319 + ], + "type": "text", + "content": "[34] Yongjin Sung, Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Ramachandra R Dasari, and Michael S Feld. Optical diffraction tomography for high resolution live cell imaging. Optics express, 17(1):266-277, 2009. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 321, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 286, + 352 + ], + "type": "text", + "content": "[35] Masahiro Takada. Subaru hyper supprime-cam project. In AIP Conference Proceedings, pages 120-127. American Institute of Physics, 2010. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 418 + ], + "type": "text", + "content": "[36] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 421, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 421, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 421, + 286, + 453 + ], + "type": "text", + "content": "[37] Arjun Teh, Matthew O'Toole, and Ioannis Gkioulekas. Adjoint nonlinear ray tracing. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 455, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 286, + 520 + ], + "type": "text", + "content": "[38] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 522, + 286, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 522, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 49, + 522, + 286, + 564 + ], + "type": "text", + "content": "[39] JT VanderPlas, AJ Connolly, Bhuvnesh Jain, and Mike Jarvis. Three-dimensional reconstruction of the density field: An svd approach to weak-lensing tomography. The Astrophysical Journal, 727(2):118, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 567, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 286, + 632 + ], + "type": "text", + "content": "[40] Tianfan Xue, Michael Rubinstein, Neal Wadhwa, Anat Levin, Fredo Durand, and William T Freeman. Refraction wiggles for measuring fluid depth and velocity from video. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 767-782. Springer, 2014. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25367" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_content_list.json b/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..110f03ad7a7504fbb6a002aad196e9f218c8938a --- /dev/null +++ b/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_content_list.json @@ -0,0 +1,1756 @@ +[ + { + "type": "text", + "text": "Single-Model and Any-Modality for Video Object Tracking", + "text_level": 1, + "bbox": [ + 184, + 130, + 784, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zongwei Wu $^{1}$ Jilai Zheng $^{2}$ Xiangxuan Ren $^{2}$ Florin-Alexandru Vasluianu $^{1}$ Chao Ma $^{2*}$ Danda Pani Paudel $^{3}$ Luc Van Gool $^{3,4}$ Radu Timofte $^{1}$", + "bbox": [ + 163, + 178, + 802, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Computer Vision Lab, CAIDAS & IFI, University of Wurzburg 2 AI Institute, Shanghai Jiao Tong University 3 INSAIT, Sofia University 4 CVL, ETH Zurich", + "bbox": [ + 148, + 217, + 820, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the realm of video object tracking, auxiliary modalities such as depth, thermal, or event data have emerged as valuable assets to complement the RGB trackers. In practice, most existing RGB trackers learn a single set of parameters to use them across datasets and applications. However, a similar single-model unification for multimodality tracking presents several challenges. These challenges stem from the inherent heterogeneity of inputs - each with modality-specific representations, the scarcity of multimodal datasets, and the absence of all the modalities at all times. In this work, we introduce Un-Track, a Unified Tracker of a single set of parameters for any modality. To handle any modality, our method learns their common latent space through low-rank factorization and reconstruction techniques. More importantly, we use only the RGB-X pairs to learn the common latent space. This unique shared representation seamlessly binds all modalities together, enabling effective unification and accommodating any missing modality, all within a single transformer-based architecture. Our Un-Track achieves +8.1 absolute F-score gain, on the DepthTrack dataset, by introducing only +2.14 (over 21.50) GFLOPs with +6.6M (over 93M) parameters, through a simple yet efficient prompting strategy. Extensive comparisons on five benchmark datasets with different modalities show that Un-Track surpasses both SOTA unified trackers and modality-specific counterparts, validating our effectiveness and practicality. The source code is publicly available at https://github.com/Zongwei97/UnTrack.", + "bbox": [ + 76, + 319, + 473, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 787, + 209, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video object tracking [9, 12, 86, 90] is a fundamental task in computer vision with wide-ranging applications spanning from surveillance [89] to augmented reality [23], where accuracy and robustness are paramount. While traditional", + "bbox": [ + 75, + 814, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9333d4bb350e45c134026b34eab75cd3425c27e62ff5aff14b298ff7dfb44f1.jpg", + "image_caption": [ + "Figure 1. Un-Track is a unified tracker with a single parameter set that seamlessly integrates any modality (of RGB-X)." + ], + "image_footnote": [], + "bbox": [ + 509, + 287, + 888, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RGB trackers have shown promise in general settings, they often struggle with challenging scenarios such as occlusions [40, 47], low visibility [83, 92], or fast-moving objects [10, 48]. For more reliable tracking under such challenging conditions, the integration of auxiliary modalities (X) like depth [20, 68], thermal [83, 87], and event [79, 98] have proven effective in multimodal tracking.", + "bbox": [ + 496, + 578, + 890, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the idea of fusing RGB with other modalities holds promise [82, 84], the main challenge is the discrepancy in the representation of information across different modalities. Despite the success of previous fusion works [15, 60] tailored for each specific scenario to improve RGB trackers, their reliance on modality-specific designs limits adaptability. Recent initiatives [69, 91] towards a uniform architecture for various modalities show promise but necessitate modality-specific fine-tuning. This approach leads to multiple parameter sets, as shown in Fig. 1(a), thereby compromising practicality in diverse real-world applications.", + "bbox": [ + 496, + 686, + 892, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we aim to avoid such modality-specific fine-tuning to keep only one model-parameter set at all times. Another practical constraint arises from the differ", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 94, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19156", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ences in available auxiliary modalities across settings. Unifying modalities by a common representation can handle any modality at its disposal, addressing the mentioned problems by its virtue. However, two additional multifaceted challenges emerge from the scarcity of multimodal datasets and the absence of all paired data combinations. The former makes cross-modal mappings through a large-scale data prior unfeasible [19, 39, 76], while the latter leads to missing modalities and renders joint learning using all possible combinations of paired data unfeasible [41, 80].", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve such a unification, in this paper, we present Un-Track, denoting \"one\" in French, which learns a cohesive embedding across diverse input modalities. Unlike previous approaches [51, 58, 73, 80], Un-Track relies solely on RGB-X pairs for training, with X representing auxiliary modalities, without the need for all modalities to co-occur. Our objective is to discover a shared embedding seamlessly binding all auxiliary modalities (as depicted in Fig. 1(b)). More specifically, we leverage the factorization prior, allowing reasoning about a common embedding directly from the low-rank latent space. Factorization is a simple composition prior with the assumption that the global approximation can be constructed from a set of subset vectors. The factorization prior, successfully utilized in previous studies [1, 4, 24], is employed in our work to reconstruct a shared embedding. This process transforms the heterogeneous modal representation into a uniform one, thereby facilitating the emergent cross-modal alignment.", + "bbox": [ + 75, + 242, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, to harness the full potential of auxiliary inputs while maintaining efficiency, we leverage cross-modal features as prompts to enable RGB-X interaction. Different from previous works [25, 27], our goal is to enhance less reliable tokens, defined by a learnable score function, using multimodal cues. We approach this as a token recovery problem and leverage low-rank factorization to achieve the goal, which is first suggested in multimodal fusion, to the best of our knowledge. With its unified model architecture and prompting blocks, Un-Track is the first to offer support for cross-modality alignment under a single architecture with uniform parameters. In comparison to our RGB baseline with 21.50 GFLOPs and 92M parameters, Un-Track introduces only +2.14 GFLOPs with +6.6M parameters, resulting in a significant +8.1 absolute F-score gain demonstrated on the DepthTrack dataset. Extensive comparisons across five datasets with different modalities validate UnTrack's superiority over specialized SOTA models, surpassing both unified trackers and modality-specific fine-tuned counterparts by a substantial margin.", + "bbox": [ + 75, + 513, + 472, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 830, + 225, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal tracking: Video object tracking [7, 71] aims to detect objects in a video sequence based on their initial positions. Early approaches treated tracking as a per-frame", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "target-matching problem, with Siamese networks [32, 33, 50, 61, 74, 85] being a notable example. More recently, transformer-based methods [2, 5, 6, 11, 54, 63] have gained popularity for feature extraction and per-frame correlation in tracking. Large-scale training datasets [14, 22, 42, 45] have empowered RGB trackers to uniformly apply parameters across various applications. While RGB trackers deliver promising results, challenges such as occlusion, low illumination, and fast-moving scenes have led to the exploration of additional modalities. Several works have investigated the role of depth [70, 95], thermal [37, 38], and event modalities [49, 55] in enhancing tracking performance. Specifically, depth cues [16, 67] contribute to handling objects with different camera distances; thermal cameras [56, 60] address challenges such as low illumination; event cameras improve the temporal awareness [57, 78, 97].", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the plausible achievements, many rely on modality-specific blocks designed for individual modalities [88, 93], limiting their adaptability. Recent efforts have focused on achieving architectural unification [69, 91], yet they still necessitate modality-specific fine-tuning, resulting in distinct parameter sets for different modalities. The ideal scenario would involve a large-scale dataset encompassing all possible modal combinations, but current tracking datasets predominantly feature a single modality — depth [66, 94], thermal [35, 36], or event [55] — posing challenges for a unified model with a single parameter set.", + "bbox": [ + 496, + 334, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning with Missing modalities: Recent research has addressed real-world scenarios where models must cope with missing modalities [41, 44]. One common strategy involves estimating missing values by learning joint multimodal representations [31, 75], feasible when complete samples are available during training. However, tracking datasets typically exhibit only one modality at a time, complicating the learning of such joint representations. Other works [19, 39] implicitly learn cross-modal alignment end-to-end using large-scale datasets and deep networks, demanding substantial computational resources. Extending such approaches to tracking is challenging due to limited downstream datasets and real-world applicability constraints [26, 64]. In contrast to existing methods, our approach investigates cross-modal relationships by leveraging edge priors to learn a joint representation that unifies all modalities. Our method does not require the simultaneous occurrence of all modalities, offering a unique perspective on dealing with diverse and individual modalities.", + "bbox": [ + 496, + 501, + 890, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 805, + 598, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 500, + 830, + 689, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, our primary focus is on multimodal tracking, with the constraint that only one modality is available at a time, as shown in Fig. 2. We define our multimodal", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9251a0b0f8c112a9e6b5721515276e5d325ad8135ce281de200d47396b7fc90d.jpg", + "image_caption": [ + "Figure 2. Our proposed framework, termed Un-Track, is composed of a shared embedding, a modal prompting, and a LoRA-finetuned pretrained RGB tracker. The shared embedding learns a joint representation that unifies all modalities (Sec. 3.2). The modal prompting block enhances feature modeling with modal awareness at each scale (Sec. 3.3). To track the target object, we finetune a pretrained foundation model [72] using the LoRA technique (Sec. 3.4). Our model achieves a unified model applicable across different modalities under a single parameter set. During inference, Un-Track seamlessly integrates any image-paired data, thanks to the emergent alignment." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 844, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tracking dataset as $M = \\{M^D, M^T, M^E\\}$ , where $M^X$ represents the subset dataset with only a single modality $X$ available. Conventional methods tackling missing modalities often use dummy inputs for the absent pixels [31, 53], simulating complete datasets with all possible combinations of paired datasets. In contrast, our method transforms any auxiliary input into a shared embedding, seamlessly binding all modalities together and creating a complete and paired representation with the master RGB feature.", + "bbox": [ + 75, + 314, + 470, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate overfitting on sparse downstream multimodal datasets, we adopt a transformer-based RGB tracker with frozen parameters and fine-tune it for multimodal tracking. Leveraging a lightweight outer prompting method, we identify uncertain tokens at each scale and enhance them with cross-modal awareness. Simultaneously, an inner fine-tuning process is implemented using the LoRA technique [21].", + "bbox": [ + 75, + 453, + 468, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, our model learns the shared embedding from samples in the mixed dataset $M$ , effectively binding all modalities together. As for inference, our model can accommodate any modal input $X$ , thanks to the emergent alignment. Our trainable parameters only include cross-modal binding, outer prompting parameters, and inner LoRA parameters, ensuring a training-friendly pipeline that can be efficiently employed end-to-end on a single 24G GPU.", + "bbox": [ + 75, + 575, + 468, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Shared Embedding", + "text_level": 1, + "bbox": [ + 76, + 709, + 261, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Explicit Edge Awareness: We observe that, as illustrated in Fig. 2(a), depth data introduce 3D distance information, effectively delineating objects with varying granularity and enhancing the sharpness of 3D boundaries; thermal images generate a scene heat map, highlighting objects based on their temperatures and providing clearer contours; event data capture intensity changes, particularly around an object's outbound region. Notably, a consistent feature emerges across these modalities: the representation of the \"true\" objects' shape, often manifested through edges.", + "bbox": [ + 75, + 732, + 468, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by this observation, our objective is to harness", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "edge embedding to unify all modalities. To achieve this, as shown in Fig. 3, we generate gradient maps from auxiliary modalities by computing differences between neighboring pixels along both the x- and y-axes. Simultaneously, without compromising texture edge, we also generate RGB gradient maps. Subsequently, all gradient maps are integrated with the visual feature, forming the gradient feature $G$ .", + "bbox": [ + 496, + 316, + 890, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Low-Rank Reconstruction: While edges present a shared feature across different modalities, exclusively transforming all modalities into edges may risk overlooking modality-specific clues. Therefore, we introduce an implicit learning pipeline to complement this by discovering the shared embedding, guided by the previously generated explicit edge awareness. This combined approach allows for a more effective identification of the shared embedding, leveraging both data-driven learning and edge priors.", + "bbox": [ + 496, + 422, + 892, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In technical terms, we redefine the challenge of learning the shared embedding as a quest for the shared low-rank approximation. Both objectives share the essence of distilling common features across all modalities. However, direct estimation of the low-rank approximation becomes impractical due to the distinct data domains and modal representations. In response, we propose a pragmatic strategy: decomposing the shared low-rank vector into the low-rank of each subset component. This alternative, more manageable and feasible within a single domain, lays the groundwork for approximating the global shared low-rank from these individual low-rank components.", + "bbox": [ + 496, + 558, + 892, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall algorithm can be found in Algorithm 1. Specifically, let $M$ be the input feature with mixed auxiliary modalities, decomposed into subset features $D,T,E$ representing depth, thermal, and event samples from subset datasets $M^D,M^T,M^E$ . Their respective $k_{th}$ low-rank matrices $D_{k},T_{k},E_{k}$ , are approximated by:", + "bbox": [ + 496, + 739, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {k} = \\sigma_ {d} (D), \\quad T _ {k} = \\sigma_ {t} (T), \\quad E _ {k} = \\sigma_ {e} (E), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 842, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma_{x}$ is the modality-specific learning through a simple MLP projecting features from input channel $c$ to the low", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19158", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Implicit Shared Embedding", + "text_level": 1, + "bbox": [ + 78, + 90, + 352, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: $M$ - Mix Modal Feature, $G$ - Explicit Gradient Binding. \nOutput: $F$ - Reconstructed feature", + "bbox": [ + 89, + 109, + 464, + 136 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Initialize: Separate $M$ into subsets, each representing a modality-specific feature: $D, T, E$ .", + "2. In-Domain Approximation: Train individual approximators $\\sigma_{x}$ for each modality $x$ to derive domain-specific low-rank matrices $D_{k}, T_{k}, E_{k}$ . Idem for explicit low-rank gradient $G_{k}$ .", + "3. Fuse and Guide: Merge $D_{k}, T_{k}, E_{k}$ through a fusion function $\\varphi_{1}$ . Incorporate the explicit gradient $G_{k}$ using addition, after projecting it into a compatible latent space through $\\varphi_{2}$ .", + "4. Reconstruct: Train a reconstruction function $\\Phi_R$ to construct the shared embedding $F$ , considering the explicit guidance $G$ in the original feature space." + ], + "bbox": [ + 91, + 137, + 468, + 289 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de2111bf26edd166c00a6777b458a86d3d1ebac711b3b88db515cb9ed22cb92c.jpg", + "image_caption": [ + "RGB: Image M: Mixed Modalities G: Gradient Feature F: Shared Embedding", + "Figure 3. Shared Embedding. We derive a joint representation through low-rank factorization and reconstruction. Such an implicit learning is additionally integrated with explicit edge awareness to enhance the embedding." + ], + "image_footnote": [], + "bbox": [ + 81, + 310, + 439, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rank space $k$ ( $k < c$ ). Simultaneously, we compute the low-rank matrix $G_{k}$ from the gradient feature $G$ .", + "bbox": [ + 75, + 561, + 468, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The global shared low-rank matrix $M_{k}$ is then approximated by fusing the subset low-rank matrices $D_{k}, T_{k}, E_{k}$ , along with the gradient guidance. Technically, we concatenate $D_{k}, T_{k}, E_{k}$ and learn the joint low-rank approximation, incorporating it with the gradient guidance. This pipeline can be expressed as follows:", + "bbox": [ + 75, + 593, + 468, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {k} = \\varphi_ {R _ {1}} \\left(\\left[ D _ {k}, T _ {k}, E _ {k} \\right]\\right) + \\varphi_ {R _ {2}} (G _ {k}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 699, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[.]$ is the channel concatenation and $\\varphi_{R_i}$ are the MLP projections to the low-rank latent space. Finally, we reconstruct the shared embedding $F$ through:", + "bbox": [ + 75, + 731, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF = \\Phi_ {R} (M _ {k}) + G, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 792, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi_R$ is another MLP that projects the jointly-learned low-rank matrix back to the departing feature space. Our ablation studies validate that this subset low-rank approximation and regrouping efficiently unify all input modalities, despite their heterogeneous representations.", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/97127b3f803f23f10955139fe90f88451f56c935043afdfdbc41bae3793fb447.jpg", + "image_caption": [ + "Figure 4. Modal Prompting. For the visual feature $I$ , we employ a score function to categorize tokens into negative, uncertain, and positive segments. Using a token exchange policy, we discard negative tokens, enhance uncertain ones with corresponding tokens from $F$ , and retain positive ones. Then, we transform the feature fusion task into a token recovery problem, addressed by low-rank factorization. Similarly, we extract the most informative low-rank matrix from $F$ to fuse and reconstruct the shared output." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Outer Modal Prompting", + "text_level": 1, + "bbox": [ + 500, + 407, + 723, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RGB-tracker may fail to perform accurately in corner cases where auxiliary clues can contribute. Drawing inspiration from the success of adapting large pre-trained models to specific downstream tasks [21, 25], we introduce a modal prompting method devised to enhance RGB token $I$ with modality-awareness $F$ , as shown in Fig. 4.", + "bbox": [ + 496, + 431, + 890, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, our approach employs a shrinkage token fusion strategy. Taking $I$ as an example, we categorize tokens into three groups—negative, uncertain, and positive—based on a dynamic scoring function $s$ . These regions are defined with mask form, expressed as $m_{n}, m_{u}, m_{p} = s(I)$ . To harness multimodal clues effectively, we replace negative tokens with those from the other modality, omit uncertain ones with dummy values, and retain the positive tokens. Subsequently, these modified tokens undergo projection into the low-rank space using the approximation function $\\sigma_{c}$ . Our objective is to enhance robustness by completing uncertain tokens with information from reliable neighboring tokens. Mathematically, we obtain the first low-rank matrix $I_{l_1}$ by:", + "bbox": [ + 496, + 521, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l _ {1}} = \\sigma_ {c} \\left(m _ {n} \\cdot F + m _ {p} \\cdot I\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 729, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we target the uncertain tokens by merging them with the paired tokens from the other modality and approximate the low-rank matrix similarly. Here, we aim to throw out the possible noise, resulting in a matrix that is more informative than the original. Let $\\sigma_{n}$ be another approximation function, we obtain the second low-rank matrix $I_{l_n}$ by:", + "bbox": [ + 496, + 757, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l _ {2}} = \\sigma_ {n} \\left(m _ {u} \\cdot F + m _ {u} \\cdot I\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 863, + 890, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we fuse these two low-rank matrices in low-rank", + "bbox": [ + 519, + 885, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19159", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/da94fd0625f618c03c162d51e4e64b71734ded5ffca2677059b3ec86c52e8a3c.jpg", + "table_caption": [ + "Depth-Specific Parameters", + "Uni-model with a Single Set of Parameters" + ], + "table_footnote": [], + "table_body": "
ATCAIS [28]DDiMP [28]DeT [66]SPT [94]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]UniNext [65]SeqTrack [8]ViPT [91]Un-Track (ours)
F-score(↑)0.4760.4850.5320.5380.5780.5940.6120.3970.5150.5690.4220.5900.5610.610
Re(↑)0.4550.4690.5060.5490.5730.5960.6100.4060.5260.5820.4320.6000.5620.610
Pr(↑)0.5000.5030.5600.5270.5830.5920.6130.3880.5050.5570.4130.5800.5600.610
", + "bbox": [ + 80, + 104, + 888, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "space, forming the shared low-rank matrix $I_{l}$ for input $I$ . In such a manner, we improve the image token modeling by fully benefiting from the auxiliary clues. Mathematically, the whole process can be formulated as:", + "bbox": [ + 75, + 217, + 467, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l} = \\varphi_ {P} ([ I _ {l _ {1}}, I _ {l _ {2}} ]), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 291, + 468, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi_{P}$ is learnable fusion. Similarly, from the input $F$ , we follow the same process to obtain the low-rank matrix $F_{l}$ . For the cross-modal fusion, we add these two low-rank matrices and then project back to the original space. We can obtain the fused output $O$ by:", + "bbox": [ + 75, + 320, + 468, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO = \\Phi_ {P} \\left(I _ {l} + F _ {l}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 409, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_P$ is another learnable MLP.", + "bbox": [ + 76, + 439, + 318, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method can be regarded as a mixer of token exchange (for negative tokens) and token fusion (for uncertain tokens), while retaining the most informative modality-specific clues (for positive tokens). As the majority of fusion operations occur in the low-rank feature space, our progressive cross-modal shrinkage avoids imposing a significant additional computational burden, while being able to excavate and accumulate the rich cues from each modality for effective modal prompting.", + "bbox": [ + 75, + 455, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inner Finetuning", + "text_level": 1, + "bbox": [ + 76, + 602, + 245, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the outer modal prompting, we incorporate the LoRA technique [21] for more efficient finetuning. For each transformer attention module with the weight matrix $W_0 \\in \\mathbb{R}^{d \\times k}$ , we introduce two learnable matrices: $B \\in \\mathbb{R}^{d \\times r}$ and $A \\in \\mathbb{R}^{r \\times k}$ . This leads to the replacement of the frozen attention mechanism $h = W_0 x$ with the new LoRA attention:", + "bbox": [ + 75, + 626, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh = W _ {0} \\boldsymbol {x} + B A \\boldsymbol {x}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 734, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train our network, we adopt the same loss functions as our baseline tracker [72] for end-to-end learning.", + "bbox": [ + 75, + 760, + 468, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 805, + 207, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training Data", + "text_level": 1, + "bbox": [ + 76, + 830, + 220, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the absence of a comprehensive multi-modal tracking dataset encompassing all possible combinations (RGB-D-T-E), we have only one $\\mathrm{RGB + X}$ (i.e. RGB+depth, or", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RGB+thermal, or RGB+event) at a time. The conventional modality-specific adapted settings are referred as \"X-Specific\", whereas the main target of this paper with one model trained on all modality pairs is called \"Uni-model\" (or a single set of parameters). The training and evaluation settings are summarized as follows:", + "bbox": [ + 496, + 217, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4268f811f0d7a505cec64750def463f96ecdc85f5624c25568c7e412e3dd144b.jpg", + "table_caption": [ + "Table 1. Overall performance on DepthTrack test set [66]. Red/Green/Blue stands for the best/second/third performance. Our depth-specific model sets the SOTA record, while our uni-model with single parameters set also outperforms previous depth-specific SOTA." + ], + "table_footnote": [], + "table_body": "
ModelTrained onEvaluated on
X-SpecificRGBi+XiRGBi+Xi
Uni-model∪iRGBi+XiRGBi+Xi
", + "bbox": [ + 501, + 309, + 870, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our RGB-D samples are sourced from DepthTrack [66], a pioneering RGB-D tracking benchmark with 150 training long-term sequences. RGB-T samples are extracted from the extensive LasHeR [36] dataset, featuring 979 diverse training sequences. RGB-E samples are obtained from VisEvent [55], which boasts 500 real-world sequences. Each D/T/E input is transformed into an RGB-like form.", + "bbox": [ + 496, + 368, + 890, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Within distribution Evaluation", + "text_level": 1, + "bbox": [ + 500, + 484, + 772, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that DepthTrack [66], Lasher [36], and VisEvent [55] provide domain-specific testing sequences, our initial evaluation focuses on these within distribution sequences. For each dataset, we adhere to the metrics specified in the original papers and prior standards [69, 91] for evaluation.", + "bbox": [ + 496, + 507, + 890, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison on DepthTrack [66]: For evaluation, we use metrics such as precision (Pr) and recall (Re), as well as the F-score, which are the primary metrics. As shown in Tab. 1 When exclusively trained and fine-tuned on DepthTrack, Un-Track achieves a $+2.1\\%$ absolute precision improvement over the current depth-specific SOTA ViPT [91]. Notably, even when trained with mixed data using a single parameter set, Un-Track outperforms the depth-specific ViPT. Furthermore, when the current ViPT is jointly trained on all datasets with a single set of parameters, its performance significantly deteriorates.", + "bbox": [ + 496, + 583, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we observe that trackers excelling in other tracking datasets [14, 22, 42] might struggle in RGB-D downstream settings. UniNext [65], a leading tracker trained on various large-scale tracking datasets and related image/video tasks, exhibits poor performance. In contrast, our model achieves cross-modal unification within a single set of parameters, surpassing all depth-specific counterparts and performing closely to our specialized version. This underscores the efficacy of our shared embedding in achieving global alignment across diverse modalities.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19160", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f425a4c874b72aa69f9efe5f746961c03f4cd6ecdda6ed4b6e3f66d141816375.jpg", + "table_caption": [ + "Thermal-Specific Parameters", + "Uni-model with a Single Set of Parameters" + ], + "table_footnote": [], + "table_body": "
SGT [34]FANet [96]mfDiMP [81]DAFNet [18]MaCNet [77]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
PR(↑)0.3270.4410.4470.4480.4820.5380.6510.6670.4180.4630.5300.5820.6080.646
SR(↑)0.2320.3090.3430.3110.3500.4200.5250.5360.3330.3650.4220.4410.4900.513
", + "bbox": [ + 81, + 106, + 888, + 171 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/faa06b262871b53165fe5ba4a8262fce5334d38de755e5085ff4f89728e2a8b6.jpg", + "table_caption": [ + "Table 2. Overall performance on the Lasher thermal testing set [36]. Our thermal-specific model sets a new SOTA record. Our uni-model variant competes strongly with the previous SOTA thermal-specific models and significantly surpasses its unimodel version.", + "Event-Specific Parameters", + "Uni-model with a Single Set of Parameters" + ], + "table_footnote": [], + "table_body": "
SiamCar [62]Stark_E [62]VITAL_E [46]PrDiMP_E [13]TransT_E [5]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
Precision(↑)0.5990.6120.6490.6440.6500.6320.7580.7630.6160.6260.6910.6650.7400.755
Success(↑)0.4200.4460.4150.4530.4740.4710.5920.5970.4480.4440.5250.5040.5790.589
", + "bbox": [ + 81, + 229, + 888, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Overall performance on VisEvent dataset [55]. Our event-specific model sets a new SOTA record. Our uni-variant, with the same parameter set as in Depth and Thermal, consistently achieves competitive performance across various modalities, leading to a significant margin over the uni-variant of the SOTA modality-specific model [91].", + "bbox": [ + 75, + 297, + 893, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on LasHer [36]: Similarly, we conduct evaluations on the LasHer testing set for RGB-T tracker assessment as shown in Tab. 2. Precision (PR) and success rates (SR) are reported following conventional standards [36, 69, 91]. Initial comparisons under domain-specific settings reveal the challenges of achieving an architecture-unified model across RGB-D and RGB-T domains, with ProTrack [69] and ViPT [91] being the only works consistently leading in both settings. Our Un-Track, following domain-specific finetuning, outperforms the leading ViPT by a significant margin and sets a new SOTA record. In the cross-domain joint learning with a single set of parameters, our uni-model achieves a $+3.8\\%$ absolute gain over ViPT. Remarkably, our model with a single set of parameters already achieves very competitive performance compared to the thermal-specific ViPT version.", + "bbox": [ + 75, + 358, + 472, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on VisEvent [55]: We also evaluate tracker performance with RGB-Event input. Event data, being inherently sparse compared to depth and thermal information, presents challenges in extending existing RGB-D or RGB-T fusion designs for effective integration, hence necessitating specific fusion designs [3, 43, 98]. In contrast, we adopt a unified cross-modal prompting method based on shrinkage fusion. Our approach, with gradual token exchanges between RGB and event modalities, effectively preserves crucial modality-specific clues to enhance feature modeling. Performance-wise, under the event-specific setting, our Un-Track outperforms all counterparts, as shown in Tab. 3 and in Fig. 5. In the single set of parameters setting, our uni-model achieves a $+1.1\\%$ absolute gain in precision over the current SOTA. This underscores the effectiveness of our approach in handling the unique challenges posed by event data integration, which can be mainly contributed to our shared binding that learns the global RGB+X alignment.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29f2be60cfcc7d917d98781312dfdc23fd658d477dd5208a20bec076dbaadb10.jpg", + "image_caption": [ + "Figure 5. More precision/success comparisons on VisEvent dataset [55]. \"Uni\" stands for models with a single parameter set. \"E\" stands for the extension of RGB trackers with event fusion." + ], + "image_footnote": [], + "bbox": [ + 501, + 354, + 694, + 505 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d89a35b2ccda4e1346510bdcf6a832c3983146410eb10179321964ca153ee48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 354, + 880, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Generalization Across Datasets", + "text_level": 1, + "bbox": [ + 500, + 571, + 774, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we assess the versatility by evaluating performance on datasets that differ from the training ones, aligning with the goal of achieving a universal model checkpoint applicable to diverse scenarios.", + "bbox": [ + 496, + 595, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VOT-RGBD2022 [30]: We first perform inference on the VOT-RGBD2022 dataset. Notably, our uni-model achieves superior performance compared to the depth-specific ViPT with a $+0.5\\%$ absolute gain in accuracy.", + "bbox": [ + 496, + 657, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RGBT234 [35]: We also test our model on the other thermal dataset RGBT234, which encompasses sequences with different distributions. Our uni-model surpasses the thermal-specific ViPT with a notable $+0.7\\%$ absolute precision gain, as shown in Tab. 5.", + "bbox": [ + 496, + 718, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we present a detailed per-attribute comparison with the fine-tuned ViPT in Fig. 6. We are particularly interested in sequences with motion blur-fast motion-camera moving, as well as sequences with heavy occlusion-scale variants-background clutter. The former motion-related challenges can typically benefit from event clues, renowned for asynchronous computing, while the latter geometry", + "bbox": [ + 496, + 795, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19161", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3e110bd08cca8f49878bd634c8c900a314e44c2cd9867fdd9ad31525f1f7dcdf.jpg", + "table_caption": [ + "Depth-Specific Parameters", + "Uni-model with a Single Set of Parameters" + ], + "table_footnote": [], + "table_body": "
DRefine [29]Stark_D [29]DMTracker [30]DeT [66]SBT_D [30]SPT [94]ProTrack [69]ViPT [91]Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
EAO(↑)0.5920.6470.6580.6570.7080.6510.6510.7210.4450.6410.6660.6790.718
Accuracy(↑)0.7750.8030.7580.7600.8090.7980.8010.8150.7140.7690.8080.8020.820
Robustness(↑)0.7600.7980.8510.8450.8640.8510.8020.8710.5980.8320.8140.8460.864
", + "bbox": [ + 81, + 106, + 890, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b67334091280628925072b216e457eac80cee8b83f512d3e12f7f9dba8423e9f.jpg", + "table_caption": [ + "Table 4. Overall performance on VOT-RGBD2022 [30]. Our uni-model, trained on a mix of all modalities, shows robust generalization and outperforms all depth-specific models and other uni-model counterparts." + ], + "table_footnote": [], + "table_body": "
Thermal-Specific ParametersUni-model with a Single Set of Parameters
mfDiMP [81]SGT [34]DAFNet [18]FANet [96]MaCNet [77]CMPP [52]APFNet [59]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
MPR(↑)0.6460.7200.7960.7870.7900.8230.8270.7950.8350.8370.6770.7110.7550.8060.842
MSR(↑)0.4280.4720.5440.5530.5540.5750.5790.5990.6170.6180.4960.5080.5690.5990.625
", + "bbox": [ + 81, + 218, + 890, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98c0d70b043dfc08881f66c6ef0e3a941242ab9ff552177207e05de66533baea.jpg", + "image_caption": [ + "Figure 6. Per-attribute analysis on the thermal dataset RGBT234 [35]. Challenges related to motion and geometry are generally better addressed by event and depth cameras, respectively. Nevertheless, when inferring only with RGB-T data, our Un-Track surpasses both the SOTA thermal-specific method and the current leading uni-tracker. This success underscores our ability to learn emergent alignment across diverse modalities." + ], + "image_footnote": [], + "bbox": [ + 81, + 319, + 465, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "related challenges can typically benefit from depth cameras. However, these event/depth clues are not directly available in the RGB-T setting. Nevertheless, our Un-Track, trained on all modalities, outperforms both thermal fine-tuned ViPT [91] and the current leading uni-tracker [8] with large margins. This underscores our capability in learning event and depth priors through shared binding, without the need for the presence of these modalities during inference.", + "bbox": [ + 75, + 625, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RGB-only: In practical scenarios, challenges arise when there are no modal clues available, a typical case is when the auxiliary sensor fails to work properly. We address this demanding case in our study by substituting the modal input with dummy values. Under a such challenging setting, as shown in Tab. 6, our uni-method consistently outperforms both our RGB baseline fine-tuned counterparts significantly. Notably, such an improvement is achieved with a very limited increase in learning parameters, i.e., $+6.65\\mathrm{M}$ with $+2.14$ GFLOPs.", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b5321f98ba82d82641771286325a20cdd714dc96a545a43c7742e366e0150fcd.jpg", + "table_caption": [ + "Table 5. Overall performance on RGBT234 dataset [35]. Our uni-model sets new SOTA records without specific thermal finetuning." + ], + "table_footnote": [], + "table_body": "
RGB Baseline [72]ViPT [91]Un-Track (ours)
GFLOPs21.5021.8023.64
Params (M)92.0892.9698.73
F-score(↑)0.5290.5420.558
Re(↑)0.5220.5380.557
Pr(↑)0.5360.5460.560
", + "bbox": [ + 503, + 316, + 888, + 426 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a207d800b14a42e71a9d560f7e02b2bc720917342eafa65fc09cf1d23c8208b1.jpg", + "table_caption": [ + "Table 6. Overall performance on DepthTrack test set [66] with dummy depth input." + ], + "table_footnote": [], + "table_body": "
w/o Shared Embed[91] as Modal Promptw/o LoRA Finetune
F-score(↑)0.5990.5790.594
Re(↑)0.6020.5750.598
Pr(↑)0.5970.5840.596
", + "bbox": [ + 503, + 460, + 888, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Key component analysis.", + "bbox": [ + 593, + 540, + 797, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 570, + 663, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform all experiments on the DepthTrack testing set [66] under a single parameter set setting. Key Component Analysis: We begin by studying the effectiveness of key components, including the shared embedding, modal prompting, and LoRA finetuning, as summarized in Tab. 7. We initially remove the shared embedding by directly feeding mixed modalities into the learning diagram. It can be seen that this equal treatment of all modalities harms network performance due to the heterogeneous representation across modal domains. Secondly, we replace our prompting block with a recent counterpart that computes fovea attention from additional input [91]. Our designed gradual shrinkage fusion, allowing token-wise interaction, proves to be more effective. We also report performance when we remove the inner fine-tuning with Lora. It can be seen that the performance deteriorates.", + "bbox": [ + 496, + 595, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Low-Rank: Low-rank approximation plays a vital role in our model, influencing our shared embedding, modal fusion, and LoRA-finetuning. Hence, the choice of rank is crucial for both the efficiency and effectiveness of our ap", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19162", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Low Rank Approximation Our plain version is highlighted in gray.", + "bbox": [ + 250, + 88, + 717, + 104 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa9164f814325db0da629b7328c678ba15a8702b456c9402efeb6cd9dfd5ac69.jpg", + "table_caption": [ + "(a) $\\text{Rank}_k(\\text{Sec. 3.2})$" + ], + "table_footnote": [], + "table_body": "
248
F-score(↑)0.6070.6100.602
Re(↑)0.6060.6080.601
Pr(↑)0.6080.6110.604
", + "bbox": [ + 86, + 125, + 330, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6be6d9835993e1435d69451f6279fb2ce79e6f7ce0535fe2037d6db17be3e1d4.jpg", + "table_caption": [ + "(b) $\\text{Rank}_l$ (Sec. 3.3)" + ], + "table_footnote": [], + "table_body": "
4816
0.5960.6100.606
0.5930.6080.609
0.5990.6110.604
", + "bbox": [ + 344, + 125, + 506, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8682634c645e4a1cd0cbd2f9a02e34dc53b70515923ec0891bd517d6ea97ad36.jpg", + "table_caption": [ + "(c) LoRA (Sec. 3.4)" + ], + "table_footnote": [], + "table_body": "
248
0.6010.6100.600
0.5990.6080.598
0.6020.6110.602
", + "bbox": [ + 519, + 125, + 702, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df4e4b713fdbbac621291e722557f5ad088c5d38a5c2d748d8e97e7208e61cad.jpg", + "table_caption": [ + "(d) Percentile (Sec. 3.3)" + ], + "table_footnote": [], + "table_body": "
1/81/41/3
0.6040.6100.595
0.6060.6080.593
0.6020.6110.596
", + "bbox": [ + 715, + 125, + 880, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proach. In Tab. 8, we systematically explore the impact of low-rank choices within each component.", + "bbox": [ + 75, + 220, + 468, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shared Embedding: Our objective is to identify an optimal low-rank latent space for merging different modalities effectively. Tab. 8(a) presents our ablation study, where we explore ranks of 2, 4, and 8. Lower ranks result in poorer performance due to sparse representations. Conversely, higher ranks capture too many modality-specific details, complicating the search for a shared embedding.", + "bbox": [ + 75, + 251, + 468, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modal Prompting: As shown in Tab. 8(b), similar trends are observed when investigating the low-rank choices for modal prompting with rank values of 4, 8, and 16, as low ranks struggle to capture essential information, while higher ranks introduce an overload of modality-specific details.", + "bbox": [ + 75, + 357, + 468, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LoRA-finetuning: We also vary the ranks between 2, 4, and 8 for the LoRA finetuning technique, as shown in Tab. 8(c). Lower ranks exhibited consistently poor performance, while higher ranks in this case tend to result in poorer performance, likely due to overfitting.", + "bbox": [ + 75, + 431, + 468, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remarks: These experiments emphasize the importance of selecting the best low-rank representations. Nevertheless, our model shows great resilience to the choice of LoRA, showcasing consistent performance across different low-rank configurations. Notably, all our low-rank variants outperform the current SOTA ViPT under the uni-setting, validating our robustness and effectiveness.", + "bbox": [ + 75, + 508, + 468, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modal Prompting: During the prompting fusion, a learnable score function is used to categorize tokens into three groups based on their confidence scores. Here, we explore different percentiles for the number of positive, which is the same as the number of negative tokens, leaving the rest as uncertain tokens. As shown in Tab. 8(d), the choice of percentile can influence overall performance. Lower percentiles result in poorer performance, as recovering uncertain tokens from very few neighboring tokens is challenging. Higher percentiles also lead to performance degradation since the focus is on token exchange rather than token fusion. The choice of $1/4$ leads to the best balance between exchanger and fuser, leading to the best performance.", + "bbox": [ + 75, + 613, + 468, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shared Embedding: Here, we perform ablation studies on our shared embedding, a foundational component of our uni-model. The quantitative results are presented in Tab. 9. We begin by exploring a scenario where our shared embedding is replaced with a variant lacking explicit edge guidance (w/o Explicit Edge). In this configuration, the network", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ef2d7dae8f8a3448ddee9089031402f23a0e4a87ac1fb759ffd1449d19eee6f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
w/o Explicit Edgew/o Implicit Learningw/o In-domain Approx.
F-score(↑)0.6000.6040.581
Re(↑)0.6020.6090.583
Pr(↑)0.5980.5990.579
", + "bbox": [ + 504, + 218, + 888, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation on shared embedding", + "bbox": [ + 576, + 304, + 813, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learns the shared embedding solely without any edge prior. The results highlight a substantial performance drop, underscoring the pivotal role of explicit edge guidance — a natural and static embedding that binds all modalities together — in facilitating this implicit learning process.", + "bbox": [ + 496, + 327, + 890, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conduct experiments using only the explicit edge as the shared embedding, excluding any additional learning modules (w/o Implicit Learning). This approach, too, yields suboptimal performance, primarily due to the neglect of modality-specific clues within each domain.", + "bbox": [ + 496, + 404, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we directly compute the low-rank approximation from mixed modalities, bypassing the initial in-domain approximation and subsequent fusion steps (w/o In-domain Approx.). We observe that the network struggles in learning a shared embedding in this direct low-rank approximation setup, primarily due to the intricately mixed representation caused by the domain gap between different modalities.", + "bbox": [ + 496, + 479, + 890, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 601, + 767, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a successful case of a single-model and any-modality tracker for video object tracking. The proposed method achieves a shared embedding that binds all modalities together, overcoming their heterogeneous representations. This unification is facilitated by lightweight modal prompting and inner finetuning, inheriting benefits from large-scale pre-trained trackers without introducing a substantial computational burden. Exhaustive experiments showcase our improved tracking performance and robust generalization, with any modality input.", + "bbox": [ + 496, + 628, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The authors thank the reviewers and ACs for their tremendous efforts and helpful comments. The event icon is credited to Zuowen Wang. This research is financed in part by the Alexander von Humboldt Foundation, in part by NSFC (62376156, 62322113), and in part by the Ministry of Education and Science of Bulgaria (support for INSAIT, part of the Bulgarian National Roadmap for Research Infrastructure).", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19163", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Simon Arridge, Pascal Fernsel, and Andreas Hauptmann. Joint reconstruction and low-rank decomposition for dynamic inverse problems. Inverse Problems and Imaging, 16 (3):483-523, 2022. 2", + "[2] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. In ECCV, 2022. 2", + "[3] Qinyu Chen, Zuowen Wang, Shih-Chii Liu, and Chang Gao. 3et: Efficient event-based eye tracking using a change-based convlstm network. arXiv preprint arXiv:2308.11771, 2023. 6", + "[4] Wanli Chen, Xinge Zhu, Ruoqi Sun, Junjun He, Ruiyu Li, Xiaoyong Shen, and Bei Yu. Tensor low-rank reconstruction for semantic segmentation. In ECCV, 2020. 2", + "[5] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2, 6", + "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2", + "[7] Xin Chen, Bin Yan, Jiawen Zhu, Huchuan Lu, Xiang Ruan, and Dong Wang. High-performance transformer tracking. TPAMI, 2022. 2", + "[8] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 5, 6, 7", + "[9] Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, and Rongrong Ji. Siamese box adaptive network for visual tracking. In CVPR, 2020. 1", + "[10] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In CVPR, 2022. 1", + "[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 2", + "[12] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 1", + "[13] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6", + "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 5", + "[15] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. TIP, 2023. 1", + "[16] Shang Gao, Jinyu Yang, Zhe Li, Feng Zheng, Alex Leonardis, and Jingkuan Song. Learning dual-fused modality-aware representations for rgbd tracking. In ECCV, 2022. 2", + "[17] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In ECCV, 2022. 5, 6, 7" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Yuan Gao, Chenglong Li, Yabin Zhu, Jin Tang, Tao He, and Futian Wang. Deep adaptive fusion network for high performance RGBT tracking. In ICCVW, 2019. 6, 7", + "[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 2", + "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In IROS, 2021. 1", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4, 5", + "[22] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5", + "[23] Sajid Javed, Martin Danelljan, Fahad Shahbaz Khan, Muhammad Haris Khan, Michael Felsberg, and Jiri Matas. Visual object tracking with discriminative filters and siamese networks: a survey and outlook. TPAMI, 45(5):6552-6574, 2022. 1", + "[24] I-Hong Jhuo, Dong Liu, DT Lee, and Shih-Fu Chang. Robust visual domain adaptation with low-rank reconstruction. In CVPR, 2012. 2", + "[25] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 4", + "[26] Ben Kang, Xin Chen, Dong Wang, Houwen Peng, and Huchuan Lu. Exploring lightweight hierarchical vision transformers for efficient visual tracking. In ICCV, 2023. 2", + "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2", + "[28] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, Ondrej Drbohlav, et al. The eighth visual object tracking vot2020 challenge results. In ECCVW, 2020. 5", + "[29] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In ICCVW, 2021. 7", + "[30] Matej Kristan, Aleš Leonardis, Jiri Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, et al. The tenth visual object tracking vot2022 challenge results. In ECCVW, 2023. 6, 7", + "[31] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 2, 3", + "[32] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19164", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2", + "[34] Chenglong Li, Nan Zhao, Yijuan Lu, Chengli Zhu, and Jin Tang. Weighted sparse representation regularized graph learning for RGB-T object tracking. In ACM MM, 2017. 6, 7", + "[35] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. RGB-T object tracking: Benchmark and baseline. PR, 96:106977, 2019. 2, 6, 7", + "[36] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for RGBT tracking. TIP, 31:392-404, 2021. 2, 5, 6", + "[37] Qiao Liu, Xin Li, Zhenyu He, Nana Fan, Di Yuan, and Hong-peng Wang. Learning deep multi-level similarity for thermal infrared object tracking. TMM, 23:2114-2126, 2020. 2", + "[38] Cheng Long Li, Andong Lu, Ai Hua Zheng, Zhengzheng Tu, and Jin Tang. Multi-adapter rgbt tracking. In ICCVW, 2019. 2", + "[39] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2023. 2", + "[40] Alan Lukezic, Ugur Kart, Jani Kapyla, Ahmed Durmush, Joni-Kristian Kamarainen, Jiri Matas, and Matej Kristan. Cdtb: A color and depth visual object tracking dataset and benchmark. In ICCV, 2019. 1", + "[41] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 2", + "[42] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2, 5", + "[43] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. In ICCV, 2023. 6", + "[44] Yansheng Qiu, Ziyuan Zhao, Hongdou Yao, Delin Chen, and Zheng Wang. Modal-aware visual prompting for incomplete multi-modal brain tumor segmentation. In ACM MM, 2023. 2", + "[45] Esteban Real, Jonathon Shlens, Stefano Mazzocchi, Xin Pan, and Vincent Vanhoucke. Youtube-boundingboxes: A large high-precision human-annotated data set for object detection in video. In CVPR, 2017. 2", + "[46] Yibing Song, Chao Ma, Xiaohe Wu, Lijun Gong, Linchao Bao, Wangmeng Zuo, Chunhua Shen, Rynson WH Lau, and Ming-Hsuan Yang. Vital: Visual tracking via adversarial learning. In CVPR, 2018. 6", + "[47] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In CVPR, 2021. 1", + "[48] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 2", + "[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: Visual tracking by redetection. In CVPR, 2020. 2", + "[51] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In ICCV, 2023. 2", + "[52] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for RGB-T tracking. In CVPR, 2020. 7", + "[53] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In CVPR, 2023. 3", + "[54] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2", + "[55] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. arXiv preprint arXiv:2108.05015, 2021. 2, 5, 6", + "[56] Xiao Wang, Xiujun Shu, Shilliang Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Mfgnet: Dynamic modality-aware filter generation for rgb-t tracking. TMM, 2022. 2", + "[57] Zuowen Wang, Yuhuang Hu, and Shih-Chii Liu. Exploiting spatial sparsity for event cameras with visual transformers. In ICIP, 2022. 2", + "[58] David Wisth, Marco Camurri, Sandipan Das, and Maurice Fallon. Unified multi-modal landmark tracking for tightly coupled lidar-visual-inertial odometry. RA-L, 6(2):1004-1011, 2021. 2", + "[59] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for RGBT tracking. In AAAI, 2022. 7", + "[60] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for rgbt tracking. In AAAI, 2022. 1, 2", + "[61] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. SiamFC++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, 2020. 2", + "[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 5, 6, 7", + "[63] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 2", + "[64] Bin Yan, Houwen Peng, Kan Wu, Dong Wang, Jianlong Fu, and Huchuan Lu. Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search. In CVPR, 2021. 2", + "[65] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19165", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Song Yan, Jinyu Yang, Jani Käpylä, Feng Zheng, Ales Leonardis, and Joni-Kristian Kämäräinen. Depthtrack: Unveiling the power of RGBD tracking. In ICCV, 2021. 2, 5, 7", + "[67] Song Yan, Jinyu Yang, Ales Leonardis, and Joni-Kristian Kamarainen. Depth-only object tracking. In BMVC, 2021. 2", + "[68] Jinyu Yang, Zhe Li, Song Yan, Feng Zheng, Aleš Leonardis, Joni-Kristian Kāmäräinen, and Ling Shao. Rgbd object tracking: An in-depth review. arXiv preprint arXiv:2203.14134, 2022. 1", + "[69] Jinyu Yang, Zhe Li, Feng Zheng, Ales Leonardis, and Jingkuan Song. Prompting for multi-modal tracking. In ACMM, 2022. 1, 2, 5, 6, 7", + "[70] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Aleš Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, 2023. 2", + "[71] Rui Yao, Guosheng Lin, Shixiong Xia, Jiaqi Zhao, and Yong Zhou. Video object segmentation and tracking: A survey. ACM TIST, 11(4):1-47, 2020. 2", + "[72] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In ECCV, 2022. 3, 5, 6, 7", + "[73] Jie Yin, Ang Li, Tao Li, Wenxian Yu, and Danping Zou. M2dgr: A multi-sensor and multi-scenario slam dataset for ground robots. RA-L, 7(2):2266-2273, 2021. 2", + "[74] Yuechen Yu, Yilei Xiong, Weilin Huang, and Matthew R Scott. Deformable siamese attention networks for visual object tracking. In CVPR, 2020. 2", + "[75] Jiandian Zeng, Tianyi Liu, and Jiantao Zhou. Tag-assisted multimodal sentiment analysis under uncertain missing modalities. In ACM SIGIR, 2022. 2", + "[76] Chunhui Zhang, Xin Sun, Yiqian Yang, Li Liu, Qiong Liu, Xi Zhou, and Yanfeng Wang. All in one: Exploring unified vision-language tracking with multi-modal alignment. In ACM MM, 2023. 2", + "[77] Hui Zhang, Lei Zhang, Li Zhuo, and Jing Zhang. Object tracking in RGB-T videos using modal-aware attention network and competitive learning. Sensors, 20(2):393, 2020. 6, 7", + "[78] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Bao-cai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, 2021. 2", + "[79] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 1", + "[80] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In CVPR, 2023. 2", + "[81] Lichao Zhang, Martin Danelljan, Abel Gonzalez-Garcia, Joost van de Weijer, and Fahad Shahbaz Khan. Multi-modal fusion for end-to-end RGB-T tracking. In ICCVW, 2019. 6, 7", + "[82] Pengyu Zhang, Dong Wang, and Hutchuan Lu. Multi-modal visual tracking: Review and experimental comparison. arXiv preprint arXiv:2012.04176, 2020. 1" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, 2022. 1", + "[84] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In ICCV, 2019. 1", + "[85] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 2", + "[86] Haojie Zhao, Dong Wang, and Hutchuan Lu. Representation learning for visual object tracking by masked appearance transfer. In CVPR, 2023. 1", + "[87] Jinjian Zhao, Xiaohan Zhang, and Pengyu Zhang. A unified approach for tracking uavs in infrared. In ICCV, 2021. 1", + "[88] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. PR, 111:107679, 2021. 2", + "[89] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, 2021. 1", + "[90] Jiawen Zhu, Zhenyu Chen, Zeqi Hao, Shijie Chang, Lu Zhang, Dong Wang, Huchuan Lu, Bin Luo, Jun-Yan He, Jin-Peng Lan, et al. Tracking anything in high quality. arXiv preprint arXiv:2307.13974, 2023. 1", + "[91] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, 2023. 1, 2, 5, 6, 7", + "[92] Jiawen Zhu, Huayi Tang, Zhi-Qi Cheng, Jun-Yan He, Bin Luo, Shihao Qiu, Shengming Li, and Huchuan Lu. Dcpt: Darkness clue-prompted tracking in nighttime uavs. arXiv preprint arXiv:2309.10491, 2023. 1", + "[93] Xue-Feng Zhu, Xiao-Jun Wu, Tianyang Xu, Zhen-Hua Feng, and Josef Kittler. Robust visual object tracking via adaptive attribute-aware discriminative correlation filters. TMM, 24: 301-312, 2021. 2", + "[94] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. RGBD1K: A large-scale dataset and benchmark for RGB-D object tracking. AAAI, 2023. 2, 5, 7", + "[95] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking. In AAAI, 2023. 2", + "[96] Yabin Zhu, Chenglong Li, Jin Tang, and Bin Luo. Quality-aware feature aggregation network for robust RGBT tracking. IEEE TIV, 6(1):121-130, 2020. 6, 7", + "[97] Zhiyu Zhu, Junhui Hou, and Xianqiang Lyu. Learning graph-embedded key-event back-tracing for object tracking in event clouds. NeurIPS, 2022. 2", + "[98] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event transformer-trackers. In ICCV, 2023. 1, 6" + ], + "bbox": [ + 501, + 92, + 890, + 829 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19166", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_model.json b/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3f68f818db1193ef48096061bdab86ccd511a09f --- /dev/null +++ b/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_model.json @@ -0,0 +1,2928 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.131, + 0.785, + 0.154 + ], + "angle": 0, + "content": "Single-Model and Any-Modality for Video Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.179, + 0.803, + 0.216 + ], + "angle": 0, + "content": "Zongwei Wu\\(^{1}\\) Jilai Zheng\\(^{2}\\) Xiangxuan Ren\\(^{2}\\) Florin-Alexandru Vasluianu\\(^{1}\\) Chao Ma\\(^{2*}\\) Danda Pani Paudel\\(^{3}\\) Luc Van Gool\\(^{3,4}\\) Radu Timofte\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.218, + 0.821, + 0.253 + ], + "angle": 0, + "content": "1 Computer Vision Lab, CAIDAS & IFI, University of Wurzburg 2 AI Institute, Shanghai Jiao Tong University 3 INSAIT, Sofia University 4 CVL, ETH Zurich" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.474, + 0.759 + ], + "angle": 0, + "content": "In the realm of video object tracking, auxiliary modalities such as depth, thermal, or event data have emerged as valuable assets to complement the RGB trackers. In practice, most existing RGB trackers learn a single set of parameters to use them across datasets and applications. However, a similar single-model unification for multimodality tracking presents several challenges. These challenges stem from the inherent heterogeneity of inputs - each with modality-specific representations, the scarcity of multimodal datasets, and the absence of all the modalities at all times. In this work, we introduce Un-Track, a Unified Tracker of a single set of parameters for any modality. To handle any modality, our method learns their common latent space through low-rank factorization and reconstruction techniques. More importantly, we use only the RGB-X pairs to learn the common latent space. This unique shared representation seamlessly binds all modalities together, enabling effective unification and accommodating any missing modality, all within a single transformer-based architecture. Our Un-Track achieves +8.1 absolute F-score gain, on the DepthTrack dataset, by introducing only +2.14 (over 21.50) GFLOPs with +6.6M (over 93M) parameters, through a simple yet efficient prompting strategy. Extensive comparisons on five benchmark datasets with different modalities show that Un-Track surpasses both SOTA unified trackers and modality-specific counterparts, validating our effectiveness and practicality. The source code is publicly available at https://github.com/Zongwei97/UnTrack." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.789, + 0.21, + 0.805 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.815, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Video object tracking [9, 12, 86, 90] is a fundamental task in computer vision with wide-ranging applications spanning from surveillance [89] to augmented reality [23], where accuracy and robustness are paramount. While traditional" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.289, + 0.889, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.533, + 0.895, + 0.561 + ], + "angle": 0, + "content": "Figure 1. Un-Track is a unified tracker with a single parameter set that seamlessly integrates any modality (of RGB-X)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.685 + ], + "angle": 0, + "content": "RGB trackers have shown promise in general settings, they often struggle with challenging scenarios such as occlusions [40, 47], low visibility [83, 92], or fast-moving objects [10, 48]. For more reliable tracking under such challenging conditions, the integration of auxiliary modalities (X) like depth [20, 68], thermal [83, 87], and event [79, 98] have proven effective in multimodal tracking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.893, + 0.853 + ], + "angle": 0, + "content": "While the idea of fusing RGB with other modalities holds promise [82, 84], the main challenge is the discrepancy in the representation of information across different modalities. Despite the success of previous fusion works [15, 60] tailored for each specific scenario to improve RGB trackers, their reliance on modality-specific designs limits adaptability. Recent initiatives [69, 91] towards a uniform architecture for various modalities show promise but necessitate modality-specific fine-tuning. This approach leads to multiple parameter sets, as shown in Fig. 1(a), thereby compromising practicality in diverse real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this work, we aim to avoid such modality-specific fine-tuning to keep only one model-parameter set at all times. Another practical constraint arises from the differ" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "ences in available auxiliary modalities across settings. Unifying modalities by a common representation can handle any modality at its disposal, addressing the mentioned problems by its virtue. However, two additional multifaceted challenges emerge from the scarcity of multimodal datasets and the absence of all paired data combinations. The former makes cross-modal mappings through a large-scale data prior unfeasible [19, 39, 76], while the latter leads to missing modalities and renders joint learning using all possible combinations of paired data unfeasible [41, 80]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.514 + ], + "angle": 0, + "content": "To achieve such a unification, in this paper, we present Un-Track, denoting \"one\" in French, which learns a cohesive embedding across diverse input modalities. Unlike previous approaches [51, 58, 73, 80], Un-Track relies solely on RGB-X pairs for training, with X representing auxiliary modalities, without the need for all modalities to co-occur. Our objective is to discover a shared embedding seamlessly binding all auxiliary modalities (as depicted in Fig. 1(b)). More specifically, we leverage the factorization prior, allowing reasoning about a common embedding directly from the low-rank latent space. Factorization is a simple composition prior with the assumption that the global approximation can be constructed from a set of subset vectors. The factorization prior, successfully utilized in previous studies [1, 4, 24], is employed in our work to reconstruct a shared embedding. This process transforms the heterogeneous modal representation into a uniform one, thereby facilitating the emergent cross-modal alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.473, + 0.818 + ], + "angle": 0, + "content": "Moreover, to harness the full potential of auxiliary inputs while maintaining efficiency, we leverage cross-modal features as prompts to enable RGB-X interaction. Different from previous works [25, 27], our goal is to enhance less reliable tokens, defined by a learnable score function, using multimodal cues. We approach this as a token recovery problem and leverage low-rank factorization to achieve the goal, which is first suggested in multimodal fusion, to the best of our knowledge. With its unified model architecture and prompting blocks, Un-Track is the first to offer support for cross-modality alignment under a single architecture with uniform parameters. In comparison to our RGB baseline with 21.50 GFLOPs and 92M parameters, Un-Track introduces only +2.14 GFLOPs with +6.6M parameters, resulting in a significant +8.1 absolute F-score gain demonstrated on the DepthTrack dataset. Extensive comparisons across five datasets with different modalities validate UnTrack's superiority over specialized SOTA models, surpassing both unified trackers and modality-specific fine-tuned counterparts by a substantial margin." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.227, + 0.845 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Multimodal tracking: Video object tracking [7, 71] aims to detect objects in a video sequence based on their initial positions. Early approaches treated tracking as a per-frame" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "target-matching problem, with Siamese networks [32, 33, 50, 61, 74, 85] being a notable example. More recently, transformer-based methods [2, 5, 6, 11, 54, 63] have gained popularity for feature extraction and per-frame correlation in tracking. Large-scale training datasets [14, 22, 42, 45] have empowered RGB trackers to uniformly apply parameters across various applications. While RGB trackers deliver promising results, challenges such as occlusion, low illumination, and fast-moving scenes have led to the exploration of additional modalities. Several works have investigated the role of depth [70, 95], thermal [37, 38], and event modalities [49, 55] in enhancing tracking performance. Specifically, depth cues [16, 67] contribute to handling objects with different camera distances; thermal cameras [56, 60] address challenges such as low illumination; event cameras improve the temporal awareness [57, 78, 97]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Despite the plausible achievements, many rely on modality-specific blocks designed for individual modalities [88, 93], limiting their adaptability. Recent efforts have focused on achieving architectural unification [69, 91], yet they still necessitate modality-specific fine-tuning, resulting in distinct parameter sets for different modalities. The ideal scenario would involve a large-scale dataset encompassing all possible modal combinations, but current tracking datasets predominantly feature a single modality — depth [66, 94], thermal [35, 36], or event [55] — posing challenges for a unified model with a single parameter set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Learning with Missing modalities: Recent research has addressed real-world scenarios where models must cope with missing modalities [41, 44]. One common strategy involves estimating missing values by learning joint multimodal representations [31, 75], feasible when complete samples are available during training. However, tracking datasets typically exhibit only one modality at a time, complicating the learning of such joint representations. Other works [19, 39] implicitly learn cross-modal alignment end-to-end using large-scale datasets and deep networks, demanding substantial computational resources. Extending such approaches to tracking is challenging due to limited downstream datasets and real-world applicability constraints [26, 64]. In contrast to existing methods, our approach investigates cross-modal relationships by leveraging edge priors to learn a joint representation that unifies all modalities. Our method does not require the simultaneous occurrence of all modalities, offering a unique perspective on dealing with diverse and individual modalities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.806, + 0.599, + 0.821 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.691, + 0.846 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, our primary focus is on multimodal tracking, with the constraint that only one modality is available at a time, as shown in Fig. 2. We define our multimodal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19157" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.845, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.894, + 0.302 + ], + "angle": 0, + "content": "Figure 2. Our proposed framework, termed Un-Track, is composed of a shared embedding, a modal prompting, and a LoRA-finetuned pretrained RGB tracker. The shared embedding learns a joint representation that unifies all modalities (Sec. 3.2). The modal prompting block enhances feature modeling with modal awareness at each scale (Sec. 3.3). To track the target object, we finetune a pretrained foundation model [72] using the LoRA technique (Sec. 3.4). Our model achieves a unified model applicable across different modalities under a single parameter set. During inference, Un-Track seamlessly integrates any image-paired data, thanks to the emergent alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.471, + 0.452 + ], + "angle": 0, + "content": "tracking dataset as \\( M = \\{M^D, M^T, M^E\\} \\), where \\( M^X \\) represents the subset dataset with only a single modality \\( X \\) available. Conventional methods tackling missing modalities often use dummy inputs for the absent pixels [31, 53], simulating complete datasets with all possible combinations of paired datasets. In contrast, our method transforms any auxiliary input into a shared embedding, seamlessly binding all modalities together and creating a complete and paired representation with the master RGB feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.575 + ], + "angle": 0, + "content": "To mitigate overfitting on sparse downstream multimodal datasets, we adopt a transformer-based RGB tracker with frozen parameters and fine-tune it for multimodal tracking. Leveraging a lightweight outer prompting method, we identify uncertain tokens at each scale and enhance them with cross-modal awareness. Simultaneously, an inner fine-tuning process is implemented using the LoRA technique [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.697 + ], + "angle": 0, + "content": "During training, our model learns the shared embedding from samples in the mixed dataset \\(M\\), effectively binding all modalities together. As for inference, our model can accommodate any modal input \\(X\\), thanks to the emergent alignment. Our trainable parameters only include cross-modal binding, outer prompting parameters, and inner LoRA parameters, ensuring a training-friendly pipeline that can be efficiently employed end-to-end on a single 24G GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.263, + 0.726 + ], + "angle": 0, + "content": "3.2. Shared Embedding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Explicit Edge Awareness: We observe that, as illustrated in Fig. 2(a), depth data introduce 3D distance information, effectively delineating objects with varying granularity and enhancing the sharpness of 3D boundaries; thermal images generate a scene heat map, highlighting objects based on their temperatures and providing clearer contours; event data capture intensity changes, particularly around an object's outbound region. Notably, a consistent feature emerges across these modalities: the representation of the \"true\" objects' shape, often manifested through edges." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Motivated by this observation, our objective is to harness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.422 + ], + "angle": 0, + "content": "edge embedding to unify all modalities. To achieve this, as shown in Fig. 3, we generate gradient maps from auxiliary modalities by computing differences between neighboring pixels along both the x- and y-axes. Simultaneously, without compromising texture edge, we also generate RGB gradient maps. Subsequently, all gradient maps are integrated with the visual feature, forming the gradient feature \\( G \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.893, + 0.559 + ], + "angle": 0, + "content": "Implicit Low-Rank Reconstruction: While edges present a shared feature across different modalities, exclusively transforming all modalities into edges may risk overlooking modality-specific clues. Therefore, we introduce an implicit learning pipeline to complement this by discovering the shared embedding, guided by the previously generated explicit edge awareness. This combined approach allows for a more effective identification of the shared embedding, leveraging both data-driven learning and edge priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.893, + 0.739 + ], + "angle": 0, + "content": "In technical terms, we redefine the challenge of learning the shared embedding as a quest for the shared low-rank approximation. Both objectives share the essence of distilling common features across all modalities. However, direct estimation of the low-rank approximation becomes impractical due to the distinct data domains and modal representations. In response, we propose a pragmatic strategy: decomposing the shared low-rank vector into the low-rank of each subset component. This alternative, more manageable and feasible within a single domain, lays the groundwork for approximating the global shared low-rank from these individual low-rank components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.831 + ], + "angle": 0, + "content": "The overall algorithm can be found in Algorithm 1. Specifically, let \\(M\\) be the input feature with mixed auxiliary modalities, decomposed into subset features \\(D,T,E\\) representing depth, thermal, and event samples from subset datasets \\(M^D,M^T,M^E\\). Their respective \\(k_{th}\\) low-rank matrices \\(D_{k},T_{k},E_{k}\\), are approximated by:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.843, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\nD _ {k} = \\sigma_ {d} (D), \\quad T _ {k} = \\sigma_ {t} (T), \\quad E _ {k} = \\sigma_ {e} (E), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\sigma_{x}\\) is the modality-specific learning through a simple MLP projecting features from input channel \\(c\\) to the low" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19158" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.092, + 0.353, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Implicit Shared Embedding" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.11, + 0.465, + 0.137 + ], + "angle": 0, + "content": "Input: \\(M\\) - Mix Modal Feature, \\(G\\) - Explicit Gradient Binding. \nOutput: \\(F\\) - Reconstructed feature" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.138, + 0.469, + 0.164 + ], + "angle": 0, + "content": "1. Initialize: Separate \\(M\\) into subsets, each representing a modality-specific feature: \\(D, T, E\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.166, + 0.469, + 0.207 + ], + "angle": 0, + "content": "2. In-Domain Approximation: Train individual approximators \\(\\sigma_{x}\\) for each modality \\(x\\) to derive domain-specific low-rank matrices \\(D_{k}, T_{k}, E_{k}\\). Idem for explicit low-rank gradient \\(G_{k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.208, + 0.469, + 0.248 + ], + "angle": 0, + "content": "3. Fuse and Guide: Merge \\(D_{k}, T_{k}, E_{k}\\) through a fusion function \\(\\varphi_{1}\\). Incorporate the explicit gradient \\(G_{k}\\) using addition, after projecting it into a compatible latent space through \\(\\varphi_{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.249, + 0.469, + 0.29 + ], + "angle": 0, + "content": "4. Reconstruct: Train a reconstruction function \\(\\Phi_R\\) to construct the shared embedding \\(F\\), considering the explicit guidance \\(G\\) in the original feature space." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.138, + 0.469, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.311, + 0.441, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.473, + 0.469, + 0.485 + ], + "angle": 0, + "content": "RGB: Image M: Mixed Modalities G: Gradient Feature F: Shared Embedding" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.49, + 0.469, + 0.545 + ], + "angle": 0, + "content": "Figure 3. Shared Embedding. We derive a joint representation through low-rank factorization and reconstruction. Such an implicit learning is additionally integrated with explicit edge awareness to enhance the embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.469, + 0.592 + ], + "angle": 0, + "content": "rank space \\(k\\) (\\(k < c\\)). Simultaneously, we compute the low-rank matrix \\(G_{k}\\) from the gradient feature \\(G\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.469, + 0.684 + ], + "angle": 0, + "content": "The global shared low-rank matrix \\( M_{k} \\) is then approximated by fusing the subset low-rank matrices \\( D_{k}, T_{k}, E_{k} \\), along with the gradient guidance. Technically, we concatenate \\( D_{k}, T_{k}, E_{k} \\) and learn the joint low-rank approximation, incorporating it with the gradient guidance. This pipeline can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.7, + 0.469, + 0.718 + ], + "angle": 0, + "content": "\\[\nM _ {k} = \\varphi_ {R _ {1}} \\left(\\left[ D _ {k}, T _ {k}, E _ {k} \\right]\\right) + \\varphi_ {R _ {2}} (G _ {k}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.469, + 0.778 + ], + "angle": 0, + "content": "where \\([.]\\) is the channel concatenation and \\(\\varphi_{R_i}\\) are the MLP projections to the low-rank latent space. Finally, we reconstruct the shared embedding \\(F\\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.794, + 0.469, + 0.81 + ], + "angle": 0, + "content": "\\[\nF = \\Phi_ {R} (M _ {k}) + G, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "where \\(\\Phi_R\\) is another MLP that projects the jointly-learned low-rank matrix back to the departing feature space. Our ablation studies validate that this subset low-rank approximation and regrouping efficiently unify all input modalities, despite their heterogeneous representations." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.891, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.39 + ], + "angle": 0, + "content": "Figure 4. Modal Prompting. For the visual feature \\( I \\), we employ a score function to categorize tokens into negative, uncertain, and positive segments. Using a token exchange policy, we discard negative tokens, enhance uncertain ones with corresponding tokens from \\( F \\), and retain positive ones. Then, we transform the feature fusion task into a token recovery problem, addressed by low-rank factorization. Similarly, we extract the most informative low-rank matrix from \\( F \\) to fuse and reconstruct the shared output." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.724, + 0.425 + ], + "angle": 0, + "content": "3.3. Outer Modal Prompting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.522 + ], + "angle": 0, + "content": "RGB-tracker may fail to perform accurately in corner cases where auxiliary clues can contribute. Drawing inspiration from the success of adapting large pre-trained models to specific downstream tasks [21, 25], we introduce a modal prompting method devised to enhance RGB token \\( I \\) with modality-awareness \\( F \\), as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Specifically, our approach employs a shrinkage token fusion strategy. Taking \\( I \\) as an example, we categorize tokens into three groups—negative, uncertain, and positive—based on a dynamic scoring function \\( s \\). These regions are defined with mask form, expressed as \\( m_{n}, m_{u}, m_{p} = s(I) \\). To harness multimodal clues effectively, we replace negative tokens with those from the other modality, omit uncertain ones with dummy values, and retain the positive tokens. Subsequently, these modified tokens undergo projection into the low-rank space using the approximation function \\( \\sigma_{c} \\). Our objective is to enhance robustness by completing uncertain tokens with information from reliable neighboring tokens. Mathematically, we obtain the first low-rank matrix \\( I_{l_1} \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.731, + 0.891, + 0.747 + ], + "angle": 0, + "content": "\\[\nI _ {l _ {1}} = \\sigma_ {c} \\left(m _ {n} \\cdot F + m _ {p} \\cdot I\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.849 + ], + "angle": 0, + "content": "Next, we target the uncertain tokens by merging them with the paired tokens from the other modality and approximate the low-rank matrix similarly. Here, we aim to throw out the possible noise, resulting in a matrix that is more informative than the original. Let \\(\\sigma_{n}\\) be another approximation function, we obtain the second low-rank matrix \\(I_{l_n}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.864, + 0.891, + 0.88 + ], + "angle": 0, + "content": "\\[\nI _ {l _ {2}} = \\sigma_ {n} \\left(m _ {u} \\cdot F + m _ {u} \\cdot I\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Then, we fuse these two low-rank matrices in low-rank" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19159" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.269, + 0.094, + 0.383, + 0.104 + ], + "angle": 0, + "content": "Depth-Specific Parameters" + }, + { + "type": "table_caption", + "bbox": [ + 0.606, + 0.094, + 0.788, + 0.104 + ], + "angle": 0, + "content": "Uni-model with a Single Set of Parameters" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.105, + 0.89, + 0.173 + ], + "angle": 0, + "content": "
ATCAIS [28]DDiMP [28]DeT [66]SPT [94]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]UniNext [65]SeqTrack [8]ViPT [91]Un-Track (ours)
F-score(↑)0.4760.4850.5320.5380.5780.5940.6120.3970.5150.5690.4220.5900.5610.610
Re(↑)0.4550.4690.5060.5490.5730.5960.6100.4060.5260.5820.4320.6000.5620.610
Pr(↑)0.5000.5030.5600.5270.5830.5920.6130.3880.5050.5570.4130.5800.5600.610
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.174, + 0.89, + 0.203 + ], + "angle": 0, + "content": "Table 1. Overall performance on DepthTrack test set [66]. Red/Green/Blue stands for the best/second/third performance. Our depth-specific model sets the SOTA record, while our uni-model with single parameters set also outperforms previous depth-specific SOTA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.218, + 0.468, + 0.279 + ], + "angle": 0, + "content": "space, forming the shared low-rank matrix \\( I_{l} \\) for input \\( I \\). In such a manner, we improve the image token modeling by fully benefiting from the auxiliary clues. Mathematically, the whole process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.292, + 0.469, + 0.309 + ], + "angle": 0, + "content": "\\[\nI _ {l} = \\varphi_ {P} ([ I _ {l _ {1}}, I _ {l _ {2}} ]), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.469, + 0.397 + ], + "angle": 0, + "content": "where \\(\\varphi_{P}\\) is learnable fusion. Similarly, from the input \\(F\\), we follow the same process to obtain the low-rank matrix \\(F_{l}\\). For the cross-modal fusion, we add these two low-rank matrices and then project back to the original space. We can obtain the fused output \\(O\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.41, + 0.469, + 0.427 + ], + "angle": 0, + "content": "\\[\nO = \\Phi_ {P} \\left(I _ {l} + F _ {l}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.319, + 0.454 + ], + "angle": 0, + "content": "where \\(\\Phi_P\\) is another learnable MLP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.469, + 0.591 + ], + "angle": 0, + "content": "Our method can be regarded as a mixer of token exchange (for negative tokens) and token fusion (for uncertain tokens), while retaining the most informative modality-specific clues (for positive tokens). As the majority of fusion operations occur in the low-rank feature space, our progressive cross-modal shrinkage avoids imposing a significant additional computational burden, while being able to excavate and accumulate the rich cues from each modality for effective modal prompting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.246, + 0.619 + ], + "angle": 0, + "content": "3.4. Inner Finetuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.469, + 0.716 + ], + "angle": 0, + "content": "In addition to the outer modal prompting, we incorporate the LoRA technique [21] for more efficient finetuning. For each transformer attention module with the weight matrix \\( W_0 \\in \\mathbb{R}^{d \\times k} \\), we introduce two learnable matrices: \\( B \\in \\mathbb{R}^{d \\times r} \\) and \\( A \\in \\mathbb{R}^{r \\times k} \\). This leads to the replacement of the frozen attention mechanism \\( h = W_0 x \\) with the new LoRA attention:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.736, + 0.469, + 0.75 + ], + "angle": 0, + "content": "\\[\nh = W _ {0} \\boldsymbol {x} + B A \\boldsymbol {x}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.469, + 0.79 + ], + "angle": 0, + "content": "To train our network, we adopt the same loss functions as our baseline tracker [72] for end-to-end learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.209, + 0.823 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.221, + 0.847 + ], + "angle": 0, + "content": "4.1. Training Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In the absence of a comprehensive multi-modal tracking dataset encompassing all possible combinations (RGB-D-T-E), we have only one \\(\\mathrm{RGB + X}\\) (i.e. RGB+depth, or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.308 + ], + "angle": 0, + "content": "RGB+thermal, or RGB+event) at a time. The conventional modality-specific adapted settings are referred as \"X-Specific\", whereas the main target of this paper with one model trained on all modality pairs is called \"Uni-model\" (or a single set of parameters). The training and evaluation settings are summarized as follows:" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.31, + 0.871, + 0.364 + ], + "angle": 0, + "content": "
ModelTrained onEvaluated on
X-SpecificRGBi+XiRGBi+Xi
Uni-model∪iRGBi+XiRGBi+Xi
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Our RGB-D samples are sourced from DepthTrack [66], a pioneering RGB-D tracking benchmark with 150 training long-term sequences. RGB-T samples are extracted from the extensive LasHeR [36] dataset, featuring 979 diverse training sequences. RGB-E samples are obtained from VisEvent [55], which boasts 500 real-world sequences. Each D/T/E input is transformed into an RGB-like form." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.773, + 0.499 + ], + "angle": 0, + "content": "4.2. Within distribution Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.891, + 0.583 + ], + "angle": 0, + "content": "Given that DepthTrack [66], Lasher [36], and VisEvent [55] provide domain-specific testing sequences, our initial evaluation focuses on these within distribution sequences. For each dataset, we adhere to the metrics specified in the original papers and prior standards [69, 91] for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Comparison on DepthTrack [66]: For evaluation, we use metrics such as precision (Pr) and recall (Re), as well as the F-score, which are the primary metrics. As shown in Tab. 1 When exclusively trained and fine-tuned on DepthTrack, Un-Track achieves a \\(+2.1\\%\\) absolute precision improvement over the current depth-specific SOTA ViPT [91]. Notably, even when trained with mixed data using a single parameter set, Un-Track outperforms the depth-specific ViPT. Furthermore, when the current ViPT is jointly trained on all datasets with a single set of parameters, its performance significantly deteriorates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Additionally, we observe that trackers excelling in other tracking datasets [14, 22, 42] might struggle in RGB-D downstream settings. UniNext [65], a leading tracker trained on various large-scale tracking datasets and related image/video tasks, exhibits poor performance. In contrast, our model achieves cross-modal unification within a single set of parameters, surpassing all depth-specific counterparts and performing closely to our specialized version. This underscores the efficacy of our shared embedding in achieving global alignment across diverse modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19160" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.094, + 0.413, + 0.106 + ], + "angle": 0, + "content": "Thermal-Specific Parameters" + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.094, + 0.825, + 0.106 + ], + "angle": 0, + "content": "Uni-model with a Single Set of Parameters" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.107, + 0.89, + 0.172 + ], + "angle": 0, + "content": "
SGT [34]FANet [96]mfDiMP [81]DAFNet [18]MaCNet [77]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
PR(↑)0.3270.4410.4470.4480.4820.5380.6510.6670.4180.4630.5300.5820.6080.646
SR(↑)0.2320.3090.3430.3110.3500.4200.5250.5360.3330.3650.4220.4410.4900.513
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.174, + 0.893, + 0.203 + ], + "angle": 0, + "content": "Table 2. Overall performance on the Lasher thermal testing set [36]. Our thermal-specific model sets a new SOTA record. Our uni-model variant competes strongly with the previous SOTA thermal-specific models and significantly surpasses its unimodel version." + }, + { + "type": "table_caption", + "bbox": [ + 0.304, + 0.218, + 0.434, + 0.229 + ], + "angle": 0, + "content": "Event-Specific Parameters" + }, + { + "type": "table_caption", + "bbox": [ + 0.632, + 0.218, + 0.84, + 0.229 + ], + "angle": 0, + "content": "Uni-model with a Single Set of Parameters" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.23, + 0.89, + 0.296 + ], + "angle": 0, + "content": "
SiamCar [62]Stark_E [62]VITAL_E [46]PrDiMP_E [13]TransT_E [5]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
Precision(↑)0.5990.6120.6490.6440.6500.6320.7580.7630.6160.6260.6910.6650.7400.755
Success(↑)0.4200.4460.4150.4530.4740.4710.5920.5970.4480.4440.5250.5040.5790.589
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.298, + 0.894, + 0.34 + ], + "angle": 0, + "content": "Table 3. Overall performance on VisEvent dataset [55]. Our event-specific model sets a new SOTA record. Our uni-variant, with the same parameter set as in Depth and Thermal, consistently achieves competitive performance across various modalities, leading to a significant margin over the uni-variant of the SOTA modality-specific model [91]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.601 + ], + "angle": 0, + "content": "Comparison on LasHer [36]: Similarly, we conduct evaluations on the LasHer testing set for RGB-T tracker assessment as shown in Tab. 2. Precision (PR) and success rates (SR) are reported following conventional standards [36, 69, 91]. Initial comparisons under domain-specific settings reveal the challenges of achieving an architecture-unified model across RGB-D and RGB-T domains, with ProTrack [69] and ViPT [91] being the only works consistently leading in both settings. Our Un-Track, following domain-specific finetuning, outperforms the leading ViPT by a significant margin and sets a new SOTA record. In the cross-domain joint learning with a single set of parameters, our uni-model achieves a \\(+3.8\\%\\) absolute gain over ViPT. Remarkably, our model with a single set of parameters already achieves very competitive performance compared to the thermal-specific ViPT version." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Comparison on VisEvent [55]: We also evaluate tracker performance with RGB-Event input. Event data, being inherently sparse compared to depth and thermal information, presents challenges in extending existing RGB-D or RGB-T fusion designs for effective integration, hence necessitating specific fusion designs [3, 43, 98]. In contrast, we adopt a unified cross-modal prompting method based on shrinkage fusion. Our approach, with gradual token exchanges between RGB and event modalities, effectively preserves crucial modality-specific clues to enhance feature modeling. Performance-wise, under the event-specific setting, our Un-Track outperforms all counterparts, as shown in Tab. 3 and in Fig. 5. In the single set of parameters setting, our uni-model achieves a \\(+1.1\\%\\) absolute gain in precision over the current SOTA. This underscores the effectiveness of our approach in handling the unique challenges posed by event data integration, which can be mainly contributed to our shared binding that learns the global RGB+X alignment." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.356, + 0.695, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.356, + 0.882, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.507, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Figure 5. More precision/success comparisons on VisEvent dataset [55]. \"Uni\" stands for models with a single parameter set. \"E\" stands for the extension of RGB trackers with event fusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.776, + 0.587 + ], + "angle": 0, + "content": "4.3. Generalization Across Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.656 + ], + "angle": 0, + "content": "In this section, we assess the versatility by evaluating performance on datasets that differ from the training ones, aligning with the goal of achieving a universal model checkpoint applicable to diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.718 + ], + "angle": 0, + "content": "VOT-RGBD2022 [30]: We first perform inference on the VOT-RGBD2022 dataset. Notably, our uni-model achieves superior performance compared to the depth-specific ViPT with a \\(+0.5\\%\\) absolute gain in accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.794 + ], + "angle": 0, + "content": "RGBT234 [35]: We also test our model on the other thermal dataset RGBT234, which encompasses sequences with different distributions. Our uni-model surpasses the thermal-specific ViPT with a notable \\(+0.7\\%\\) absolute precision gain, as shown in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Moreover, we present a detailed per-attribute comparison with the fine-tuned ViPT in Fig. 6. We are particularly interested in sequences with motion blur-fast motion-camera moving, as well as sequences with heavy occlusion-scale variants-background clutter. The former motion-related challenges can typically benefit from event clues, renowned for asynchronous computing, while the latter geometry" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19161" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.318, + 0.094, + 0.443, + 0.105 + ], + "angle": 0, + "content": "Depth-Specific Parameters" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.094, + 0.84, + 0.105 + ], + "angle": 0, + "content": "Uni-model with a Single Set of Parameters" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.107, + 0.891, + 0.182 + ], + "angle": 0, + "content": "
DRefine [29]Stark_D [29]DMTracker [30]DeT [66]SBT_D [30]SPT [94]ProTrack [69]ViPT [91]Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
EAO(↑)0.5920.6470.6580.6570.7080.6510.6510.7210.4450.6410.6660.6790.718
Accuracy(↑)0.7750.8030.7580.7600.8090.7980.8010.8150.7140.7690.8080.8020.820
Robustness(↑)0.7600.7980.8510.8450.8640.8510.8020.8710.5980.8320.8140.8460.864
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.183, + 0.893, + 0.213 + ], + "angle": 0, + "content": "Table 4. Overall performance on VOT-RGBD2022 [30]. Our uni-model, trained on a mix of all modalities, shows robust generalization and outperforms all depth-specific models and other uni-model counterparts." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.219, + 0.891, + 0.286 + ], + "angle": 0, + "content": "
Thermal-Specific ParametersUni-model with a Single Set of Parameters
mfDiMP [81]SGT [34]DAFNet [18]FANet [96]MaCNet [77]CMPP [52]APFNet [59]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
MPR(↑)0.6460.7200.7960.7870.7900.8230.8270.7950.8350.8370.6770.7110.7550.8060.842
MSR(↑)0.4280.4720.5440.5530.5540.5750.5790.5990.6170.6180.4960.5080.5690.5990.625
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.287, + 0.875, + 0.302 + ], + "angle": 0, + "content": "Table 5. Overall performance on RGBT234 dataset [35]. Our uni-model sets new SOTA records without specific thermal finetuning." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.32, + 0.466, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.511, + 0.471, + 0.609 + ], + "angle": 0, + "content": "Figure 6. Per-attribute analysis on the thermal dataset RGBT234 [35]. Challenges related to motion and geometry are generally better addressed by event and depth cameras, respectively. Nevertheless, when inferring only with RGB-T data, our Un-Track surpasses both the SOTA thermal-specific method and the current leading uni-tracker. This success underscores our ability to learn emergent alignment across diverse modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.748 + ], + "angle": 0, + "content": "related challenges can typically benefit from depth cameras. However, these event/depth clues are not directly available in the RGB-T setting. Nevertheless, our Un-Track, trained on all modalities, outperforms both thermal fine-tuned ViPT [91] and the current leading uni-tracker [8] with large margins. This underscores our capability in learning event and depth priors through shared binding, without the need for the presence of these modalities during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "RGB-only: In practical scenarios, challenges arise when there are no modal clues available, a typical case is when the auxiliary sensor fails to work properly. We address this demanding case in our study by substituting the modal input with dummy values. Under a such challenging setting, as shown in Tab. 6, our uni-method consistently outperforms both our RGB baseline fine-tuned counterparts significantly. Notably, such an improvement is achieved with a very limited increase in learning parameters, i.e., \\(+6.65\\mathrm{M}\\) with \\(+2.14\\) GFLOPs." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.318, + 0.89, + 0.427 + ], + "angle": 0, + "content": "
RGB Baseline [72]ViPT [91]Un-Track (ours)
GFLOPs21.5021.8023.64
Params (M)92.0892.9698.73
F-score(↑)0.5290.5420.558
Re(↑)0.5220.5380.557
Pr(↑)0.5360.5460.560
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.428, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Table 6. Overall performance on DepthTrack test set [66] with dummy depth input." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.461, + 0.89, + 0.54 + ], + "angle": 0, + "content": "
w/o Shared Embed[91] as Modal Promptw/o LoRA Finetune
F-score(↑)0.5990.5790.594
Re(↑)0.6020.5750.598
Pr(↑)0.5970.5840.596
" + }, + { + "type": "table_caption", + "bbox": [ + 0.594, + 0.541, + 0.798, + 0.555 + ], + "angle": 0, + "content": "Table 7. Key component analysis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.665, + 0.587 + ], + "angle": 0, + "content": "5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In this section, we perform all experiments on the DepthTrack testing set [66] under a single parameter set setting. Key Component Analysis: We begin by studying the effectiveness of key components, including the shared embedding, modal prompting, and LoRA finetuning, as summarized in Tab. 7. We initially remove the shared embedding by directly feeding mixed modalities into the learning diagram. It can be seen that this equal treatment of all modalities harms network performance due to the heterogeneous representation across modal domains. Secondly, we replace our prompting block with a recent counterpart that computes fovea attention from additional input [91]. Our designed gradual shrinkage fusion, allowing token-wise interaction, proves to be more effective. We also report performance when we remove the inner fine-tuning with Lora. It can be seen that the performance deteriorates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Low-Rank: Low-rank approximation plays a vital role in our model, influencing our shared embedding, modal fusion, and LoRA-finetuning. Hence, the choice of rank is crucial for both the efficiency and effectiveness of our ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19162" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.25, + 0.089, + 0.718, + 0.105 + ], + "angle": 0, + "content": "Table 8. Low Rank Approximation Our plain version is highlighted in gray." + }, + { + "type": "table_caption", + "bbox": [ + 0.151, + 0.111, + 0.264, + 0.123 + ], + "angle": 0, + "content": "(a) \\( \\text{Rank}_k(\\text{Sec. 3.2}) \\)" + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.125, + 0.331, + 0.201 + ], + "angle": 0, + "content": "
248
F-score(↑)0.6070.6100.602
Re(↑)0.6060.6080.601
Pr(↑)0.6080.6110.604
" + }, + { + "type": "table_caption", + "bbox": [ + 0.368, + 0.111, + 0.479, + 0.123 + ], + "angle": 0, + "content": "(b) \\( \\text{Rank}_l \\) (Sec. 3.3)" + }, + { + "type": "table", + "bbox": [ + 0.345, + 0.125, + 0.508, + 0.201 + ], + "angle": 0, + "content": "
4816
0.5960.6100.606
0.5930.6080.609
0.5990.6110.604
" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.111, + 0.665, + 0.123 + ], + "angle": 0, + "content": "(c) LoRA (Sec. 3.4)" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.125, + 0.704, + 0.201 + ], + "angle": 0, + "content": "
248
0.6010.6100.600
0.5990.6080.598
0.6020.6110.602
" + }, + { + "type": "table_caption", + "bbox": [ + 0.727, + 0.111, + 0.861, + 0.123 + ], + "angle": 0, + "content": "(d) Percentile (Sec. 3.3)" + }, + { + "type": "table", + "bbox": [ + 0.716, + 0.125, + 0.882, + 0.201 + ], + "angle": 0, + "content": "
1/81/41/3
0.6040.6100.595
0.6060.6080.593
0.6020.6110.596
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.47, + 0.251 + ], + "angle": 0, + "content": "proach. In Tab. 8, we systematically explore the impact of low-rank choices within each component." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.469, + 0.357 + ], + "angle": 0, + "content": "Shared Embedding: Our objective is to identify an optimal low-rank latent space for merging different modalities effectively. Tab. 8(a) presents our ablation study, where we explore ranks of 2, 4, and 8. Lower ranks result in poorer performance due to sparse representations. Conversely, higher ranks capture too many modality-specific details, complicating the search for a shared embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.469, + 0.432 + ], + "angle": 0, + "content": "Modal Prompting: As shown in Tab. 8(b), similar trends are observed when investigating the low-rank choices for modal prompting with rank values of 4, 8, and 16, as low ranks struggle to capture essential information, while higher ranks introduce an overload of modality-specific details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.469, + 0.508 + ], + "angle": 0, + "content": "LoRA-finetuning: We also vary the ranks between 2, 4, and 8 for the LoRA finetuning technique, as shown in Tab. 8(c). Lower ranks exhibited consistently poor performance, while higher ranks in this case tend to result in poorer performance, likely due to overfitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.469, + 0.613 + ], + "angle": 0, + "content": "Remarks: These experiments emphasize the importance of selecting the best low-rank representations. Nevertheless, our model shows great resilience to the choice of LoRA, showcasing consistent performance across different low-rank configurations. Notably, all our low-rank variants outperform the current SOTA ViPT under the uni-setting, validating our robustness and effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Modal Prompting: During the prompting fusion, a learnable score function is used to categorize tokens into three groups based on their confidence scores. Here, we explore different percentiles for the number of positive, which is the same as the number of negative tokens, leaving the rest as uncertain tokens. As shown in Tab. 8(d), the choice of percentile can influence overall performance. Lower percentiles result in poorer performance, as recovering uncertain tokens from very few neighboring tokens is challenging. Higher percentiles also lead to performance degradation since the focus is on token exchange rather than token fusion. The choice of \\(1/4\\) leads to the best balance between exchanger and fuser, leading to the best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Shared Embedding: Here, we perform ablation studies on our shared embedding, a foundational component of our uni-model. The quantitative results are presented in Tab. 9. We begin by exploring a scenario where our shared embedding is replaced with a variant lacking explicit edge guidance (w/o Explicit Edge). In this configuration, the network" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.219, + 0.89, + 0.304 + ], + "angle": 0, + "content": "
w/o Explicit Edgew/o Implicit Learningw/o In-domain Approx.
F-score(↑)0.6000.6040.581
Re(↑)0.6020.6090.583
Pr(↑)0.5980.5990.579
" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.305, + 0.815, + 0.318 + ], + "angle": 0, + "content": "Table 9. Ablation on shared embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.404 + ], + "angle": 0, + "content": "learns the shared embedding solely without any edge prior. The results highlight a substantial performance drop, underscoring the pivotal role of explicit edge guidance — a natural and static embedding that binds all modalities together — in facilitating this implicit learning process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.48 + ], + "angle": 0, + "content": "We also conduct experiments using only the explicit edge as the shared embedding, excluding any additional learning modules (w/o Implicit Learning). This approach, too, yields suboptimal performance, primarily due to the neglect of modality-specific clues within each domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Finally, we directly compute the low-rank approximation from mixed modalities, bypassing the initial in-domain approximation and subsequent fusion steps (w/o In-domain Approx.). We observe that the network struggles in learning a shared embedding in this direct low-rank approximation setup, primarily due to the intricately mixed representation caused by the domain gap between different modalities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.769, + 0.618 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We present a successful case of a single-model and any-modality tracker for video object tracking. The proposed method achieves a shared embedding that binds all modalities together, overcoming their heterogeneous representations. This unification is facilitated by lightweight modal prompting and inner finetuning, inheriting benefits from large-scale pre-trained trackers without introducing a substantial computational burden. Exhaustive experiments showcase our improved tracking performance and robust generalization, with any modality input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. The authors thank the reviewers and ACs for their tremendous efforts and helpful comments. The event icon is credited to Zuowen Wang. This research is financed in part by the Alexander von Humboldt Foundation, in part by NSFC (62376156, 62322113), and in part by the Ministry of Education and Science of Bulgaria (support for INSAIT, part of the Bulgarian National Roadmap for Research Infrastructure)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19163" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Simon Arridge, Pascal Fernsel, and Andreas Hauptmann. Joint reconstruction and low-rank decomposition for dynamic inverse problems. Inverse Problems and Imaging, 16 (3):483-523, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.225 + ], + "angle": 0, + "content": "[2] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.468, + 0.28 + ], + "angle": 0, + "content": "[3] Qinyu Chen, Zuowen Wang, Shih-Chii Liu, and Chang Gao. 3et: Efficient event-based eye tracking using a change-based convlstm network. arXiv preprint arXiv:2308.11771, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.468, + 0.324 + ], + "angle": 0, + "content": "[4] Wanli Chen, Xinge Zhu, Ruoqi Sun, Junjun He, Ruiyu Li, Xiaoyong Shen, and Bei Yu. Tensor low-rank reconstruction for semantic segmentation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.468, + 0.366 + ], + "angle": 0, + "content": "[5] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[7] Xin Chen, Bin Yan, Jiawen Zhu, Huchuan Lu, Xiang Ruan, and Dong Wang. High-performance transformer tracking. TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.521 + ], + "angle": 0, + "content": "[9] Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, and Rongrong Ji. Siamese box adaptive network for visual tracking. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.523, + 0.468, + 0.589 + ], + "angle": 0, + "content": "[10] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[12] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[13] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[15] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. TIP, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[16] Shang Gao, Jinyu Yang, Zhe Li, Feng Zheng, Alex Leonardis, and Jingkuan Song. Learning dual-fused modality-aware representations for rgbd tracking. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In ECCV, 2022. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Yuan Gao, Chenglong Li, Yabin Zhu, Jin Tang, Tao He, and Futian Wang. Deep adaptive fusion network for high performance RGBT tracking. In ICCVW, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In IROS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[22] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[23] Sajid Javed, Martin Danelljan, Fahad Shahbaz Khan, Muhammad Haris Khan, Michael Felsberg, and Jiri Matas. Visual object tracking with discriminative filters and siamese networks: a survey and outlook. TPAMI, 45(5):6552-6574, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[24] I-Hong Jhuo, Dong Liu, DT Lee, and Shih-Fu Chang. Robust visual domain adaptation with low-rank reconstruction. In CVPR, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[25] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[26] Ben Kang, Xin Chen, Dong Wang, Houwen Peng, and Huchuan Lu. Exploring lightweight hierarchical vision transformers for efficient visual tracking. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[28] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, Ondrej Drbohlav, et al. The eighth visual object tracking vot2020 challenge results. In ECCVW, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[29] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In ICCVW, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[30] Matej Kristan, Aleš Leonardis, Jiri Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, et al. The tenth visual object tracking vot2022 challenge results. In ECCVW, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[31] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19164" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[33] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[34] Chenglong Li, Nan Zhao, Yijuan Lu, Chengli Zhu, and Jin Tang. Weighted sparse representation regularized graph learning for RGB-T object tracking. In ACM MM, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[35] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. RGB-T object tracking: Benchmark and baseline. PR, 96:106977, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[36] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for RGBT tracking. TIP, 31:392-404, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[37] Qiao Liu, Xin Li, Zhenyu He, Nana Fan, Di Yuan, and Hong-peng Wang. Learning deep multi-level similarity for thermal infrared object tracking. TMM, 23:2114-2126, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[38] Cheng Long Li, Andong Lu, Ai Hua Zheng, Zhengzheng Tu, and Jin Tang. Multi-adapter rgbt tracking. In ICCVW, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[39] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[40] Alan Lukezic, Ugur Kart, Jani Kapyla, Ahmed Durmush, Joni-Kristian Kamarainen, Jiri Matas, and Matej Kristan. Cdtb: A color and depth visual object tracking dataset and benchmark. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[41] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.534, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[42] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[43] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[44] Yansheng Qiu, Ziyuan Zhao, Hongdou Yao, Delin Chen, and Zheng Wang. Modal-aware visual prompting for incomplete multi-modal brain tumor segmentation. In ACM MM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[45] Esteban Real, Jonathon Shlens, Stefano Mazzocchi, Xin Pan, and Vincent Vanhoucke. Youtube-boundingboxes: A large high-precision human-annotated data set for object detection in video. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[46] Yibing Song, Chao Ma, Xiaohe Wu, Lijun Gong, Linchao Bao, Wangmeng Zuo, Chunhua Shen, Rynson WH Lau, and Ming-Hsuan Yang. Vital: Visual tracking via adversarial learning. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[47] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[48] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[49] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: Visual tracking by redetection. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[51] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[52] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for RGB-T tracking. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[53] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.343, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[54] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[55] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. arXiv preprint arXiv:2108.05015, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[56] Xiao Wang, Xiujun Shu, Shilliang Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Mfgnet: Dynamic modality-aware filter generation for rgb-t tracking. TMM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[57] Zuowen Wang, Yuhuang Hu, and Shih-Chii Liu. Exploiting spatial sparsity for event cameras with visual transformers. In ICIP, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[58] David Wisth, Marco Camurri, Sandipan Das, and Maurice Fallon. Unified multi-modal landmark tracking for tightly coupled lidar-visual-inertial odometry. RA-L, 6(2):1004-1011, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[59] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for RGBT tracking. In AAAI, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[60] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for rgbt tracking. In AAAI, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[61] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. SiamFC++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[63] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[64] Bin Yan, Houwen Peng, Kan Wu, Dong Wang, Jianlong Fu, and Huchuan Lu. Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[65] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19165" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[66] Song Yan, Jinyu Yang, Jani Käpylä, Feng Zheng, Ales Leonardis, and Joni-Kristian Kämäräinen. Depthtrack: Unveiling the power of RGBD tracking. In ICCV, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.177 + ], + "angle": 0, + "content": "[67] Song Yan, Jinyu Yang, Ales Leonardis, and Joni-Kristian Kamarainen. Depth-only object tracking. In BMVC, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[68] Jinyu Yang, Zhe Li, Song Yan, Feng Zheng, Aleš Leonardis, Joni-Kristian Kāmäräinen, and Ling Shao. Rgbd object tracking: An in-depth review. arXiv preprint arXiv:2203.14134, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[69] Jinyu Yang, Zhe Li, Feng Zheng, Ales Leonardis, and Jingkuan Song. Prompting for multi-modal tracking. In ACMM, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[70] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Aleš Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[71] Rui Yao, Guosheng Lin, Shixiong Xia, Jiaqi Zhao, and Yong Zhou. Video object segmentation and tracking: A survey. ACM TIST, 11(4):1-47, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[72] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In ECCV, 2022. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[73] Jie Yin, Ang Li, Tao Li, Wenxian Yu, and Danping Zou. M2dgr: A multi-sensor and multi-scenario slam dataset for ground robots. RA-L, 7(2):2266-2273, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[74] Yuechen Yu, Yilei Xiong, Weilin Huang, and Matthew R Scott. Deformable siamese attention networks for visual object tracking. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[75] Jiandian Zeng, Tianyi Liu, and Jiantao Zhou. Tag-assisted multimodal sentiment analysis under uncertain missing modalities. In ACM SIGIR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[76] Chunhui Zhang, Xin Sun, Yiqian Yang, Li Liu, Qiong Liu, Xi Zhou, and Yanfeng Wang. All in one: Exploring unified vision-language tracking with multi-modal alignment. In ACM MM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[77] Hui Zhang, Lei Zhang, Li Zhuo, and Jing Zhang. Object tracking in RGB-T videos using modal-aware attention network and competitive learning. Sensors, 20(2):393, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[78] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Bao-cai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[79] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.747, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[80] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.471, + 0.856 + ], + "angle": 0, + "content": "[81] Lichao Zhang, Martin Danelljan, Abel Gonzalez-Garcia, Joost van de Weijer, and Fahad Shahbaz Khan. Multi-modal fusion for end-to-end RGB-T tracking. In ICCVW, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[82] Pengyu Zhang, Dong Wang, and Hutchuan Lu. Multi-modal visual tracking: Review and experimental comparison. arXiv preprint arXiv:2012.04176, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[83] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[84] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[85] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[86] Haojie Zhao, Dong Wang, and Hutchuan Lu. Representation learning for visual object tracking by masked appearance transfer. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[87] Jinjian Zhao, Xiaohan Zhang, and Pengyu Zhang. A unified approach for tracking uavs in infrared. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.294, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[88] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. PR, 111:107679, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.336, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[89] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.434 + ], + "angle": 0, + "content": "[90] Jiawen Zhu, Zhenyu Chen, Zeqi Hao, Shijie Chang, Lu Zhang, Dong Wang, Huchuan Lu, Bin Luo, Jun-Yan He, Jin-Peng Lan, et al. Tracking anything in high quality. arXiv preprint arXiv:2307.13974, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.436, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[91] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[92] Jiawen Zhu, Huayi Tang, Zhi-Qi Cheng, Jun-Yan He, Bin Luo, Shihao Qiu, Shengming Li, and Huchuan Lu. Dcpt: Darkness clue-prompted tracking in nighttime uavs. arXiv preprint arXiv:2309.10491, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.535, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[93] Xue-Feng Zhu, Xiao-Jun Wu, Tianyang Xu, Zhen-Hua Feng, and Josef Kittler. Robust visual object tracking via adaptive attribute-aware discriminative correlation filters. TMM, 24: 301-312, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[94] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. RGBD1K: A large-scale dataset and benchmark for RGB-D object tracking. AAAI, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[95] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking. In AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[96] Yabin Zhu, Chenglong Li, Jin Tang, and Bin Luo. Quality-aware feature aggregation network for robust RGBT tracking. IEEE TIV, 6(1):121-130, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[97] Zhiyu Zhu, Junhui Hou, and Xianqiang Lyu. Learning graph-embedded key-event back-tracing for object tracking in event clouds. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[98] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event transformer-trackers. In ICCV, 2023. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19166" + } + ] +] \ No newline at end of file diff --git a/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_origin.pdf b/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79e2be8f34b51cdff5a0c779cde00572bb2f9d86 --- /dev/null +++ b/2024/Single-Model and Any-Modality for Video Object Tracking/06dd6563-2018-4142-8391-b2e6da00243a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01a6836c3c9c99f425dbfc1a46f3719aae2be7cc280302faddaece6681293ca7 +size 6723732 diff --git a/2024/Single-Model and Any-Modality for Video Object Tracking/full.md b/2024/Single-Model and Any-Modality for Video Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c67c69fe106f89195a340b025982a902fe3a2f3 --- /dev/null +++ b/2024/Single-Model and Any-Modality for Video Object Tracking/full.md @@ -0,0 +1,395 @@ +# Single-Model and Any-Modality for Video Object Tracking + +Zongwei Wu $^{1}$ Jilai Zheng $^{2}$ Xiangxuan Ren $^{2}$ Florin-Alexandru Vasluianu $^{1}$ Chao Ma $^{2*}$ Danda Pani Paudel $^{3}$ Luc Van Gool $^{3,4}$ Radu Timofte $^{1}$ + +1 Computer Vision Lab, CAIDAS & IFI, University of Wurzburg 2 AI Institute, Shanghai Jiao Tong University 3 INSAIT, Sofia University 4 CVL, ETH Zurich + +# Abstract + +In the realm of video object tracking, auxiliary modalities such as depth, thermal, or event data have emerged as valuable assets to complement the RGB trackers. In practice, most existing RGB trackers learn a single set of parameters to use them across datasets and applications. However, a similar single-model unification for multimodality tracking presents several challenges. These challenges stem from the inherent heterogeneity of inputs - each with modality-specific representations, the scarcity of multimodal datasets, and the absence of all the modalities at all times. In this work, we introduce Un-Track, a Unified Tracker of a single set of parameters for any modality. To handle any modality, our method learns their common latent space through low-rank factorization and reconstruction techniques. More importantly, we use only the RGB-X pairs to learn the common latent space. This unique shared representation seamlessly binds all modalities together, enabling effective unification and accommodating any missing modality, all within a single transformer-based architecture. Our Un-Track achieves +8.1 absolute F-score gain, on the DepthTrack dataset, by introducing only +2.14 (over 21.50) GFLOPs with +6.6M (over 93M) parameters, through a simple yet efficient prompting strategy. Extensive comparisons on five benchmark datasets with different modalities show that Un-Track surpasses both SOTA unified trackers and modality-specific counterparts, validating our effectiveness and practicality. The source code is publicly available at https://github.com/Zongwei97/UnTrack. + +# 1. Introduction + +Video object tracking [9, 12, 86, 90] is a fundamental task in computer vision with wide-ranging applications spanning from surveillance [89] to augmented reality [23], where accuracy and robustness are paramount. While traditional + +![](images/a9333d4bb350e45c134026b34eab75cd3425c27e62ff5aff14b298ff7dfb44f1.jpg) +Figure 1. Un-Track is a unified tracker with a single parameter set that seamlessly integrates any modality (of RGB-X). + +RGB trackers have shown promise in general settings, they often struggle with challenging scenarios such as occlusions [40, 47], low visibility [83, 92], or fast-moving objects [10, 48]. For more reliable tracking under such challenging conditions, the integration of auxiliary modalities (X) like depth [20, 68], thermal [83, 87], and event [79, 98] have proven effective in multimodal tracking. + +While the idea of fusing RGB with other modalities holds promise [82, 84], the main challenge is the discrepancy in the representation of information across different modalities. Despite the success of previous fusion works [15, 60] tailored for each specific scenario to improve RGB trackers, their reliance on modality-specific designs limits adaptability. Recent initiatives [69, 91] towards a uniform architecture for various modalities show promise but necessitate modality-specific fine-tuning. This approach leads to multiple parameter sets, as shown in Fig. 1(a), thereby compromising practicality in diverse real-world applications. + +In this work, we aim to avoid such modality-specific fine-tuning to keep only one model-parameter set at all times. Another practical constraint arises from the differ + +ences in available auxiliary modalities across settings. Unifying modalities by a common representation can handle any modality at its disposal, addressing the mentioned problems by its virtue. However, two additional multifaceted challenges emerge from the scarcity of multimodal datasets and the absence of all paired data combinations. The former makes cross-modal mappings through a large-scale data prior unfeasible [19, 39, 76], while the latter leads to missing modalities and renders joint learning using all possible combinations of paired data unfeasible [41, 80]. + +To achieve such a unification, in this paper, we present Un-Track, denoting "one" in French, which learns a cohesive embedding across diverse input modalities. Unlike previous approaches [51, 58, 73, 80], Un-Track relies solely on RGB-X pairs for training, with X representing auxiliary modalities, without the need for all modalities to co-occur. Our objective is to discover a shared embedding seamlessly binding all auxiliary modalities (as depicted in Fig. 1(b)). More specifically, we leverage the factorization prior, allowing reasoning about a common embedding directly from the low-rank latent space. Factorization is a simple composition prior with the assumption that the global approximation can be constructed from a set of subset vectors. The factorization prior, successfully utilized in previous studies [1, 4, 24], is employed in our work to reconstruct a shared embedding. This process transforms the heterogeneous modal representation into a uniform one, thereby facilitating the emergent cross-modal alignment. + +Moreover, to harness the full potential of auxiliary inputs while maintaining efficiency, we leverage cross-modal features as prompts to enable RGB-X interaction. Different from previous works [25, 27], our goal is to enhance less reliable tokens, defined by a learnable score function, using multimodal cues. We approach this as a token recovery problem and leverage low-rank factorization to achieve the goal, which is first suggested in multimodal fusion, to the best of our knowledge. With its unified model architecture and prompting blocks, Un-Track is the first to offer support for cross-modality alignment under a single architecture with uniform parameters. In comparison to our RGB baseline with 21.50 GFLOPs and 92M parameters, Un-Track introduces only +2.14 GFLOPs with +6.6M parameters, resulting in a significant +8.1 absolute F-score gain demonstrated on the DepthTrack dataset. Extensive comparisons across five datasets with different modalities validate UnTrack's superiority over specialized SOTA models, surpassing both unified trackers and modality-specific fine-tuned counterparts by a substantial margin. + +# 2. Related Works + +Multimodal tracking: Video object tracking [7, 71] aims to detect objects in a video sequence based on their initial positions. Early approaches treated tracking as a per-frame + +target-matching problem, with Siamese networks [32, 33, 50, 61, 74, 85] being a notable example. More recently, transformer-based methods [2, 5, 6, 11, 54, 63] have gained popularity for feature extraction and per-frame correlation in tracking. Large-scale training datasets [14, 22, 42, 45] have empowered RGB trackers to uniformly apply parameters across various applications. While RGB trackers deliver promising results, challenges such as occlusion, low illumination, and fast-moving scenes have led to the exploration of additional modalities. Several works have investigated the role of depth [70, 95], thermal [37, 38], and event modalities [49, 55] in enhancing tracking performance. Specifically, depth cues [16, 67] contribute to handling objects with different camera distances; thermal cameras [56, 60] address challenges such as low illumination; event cameras improve the temporal awareness [57, 78, 97]. + +Despite the plausible achievements, many rely on modality-specific blocks designed for individual modalities [88, 93], limiting their adaptability. Recent efforts have focused on achieving architectural unification [69, 91], yet they still necessitate modality-specific fine-tuning, resulting in distinct parameter sets for different modalities. The ideal scenario would involve a large-scale dataset encompassing all possible modal combinations, but current tracking datasets predominantly feature a single modality — depth [66, 94], thermal [35, 36], or event [55] — posing challenges for a unified model with a single parameter set. + +Learning with Missing modalities: Recent research has addressed real-world scenarios where models must cope with missing modalities [41, 44]. One common strategy involves estimating missing values by learning joint multimodal representations [31, 75], feasible when complete samples are available during training. However, tracking datasets typically exhibit only one modality at a time, complicating the learning of such joint representations. Other works [19, 39] implicitly learn cross-modal alignment end-to-end using large-scale datasets and deep networks, demanding substantial computational resources. Extending such approaches to tracking is challenging due to limited downstream datasets and real-world applicability constraints [26, 64]. In contrast to existing methods, our approach investigates cross-modal relationships by leveraging edge priors to learn a joint representation that unifies all modalities. Our method does not require the simultaneous occurrence of all modalities, offering a unique perspective on dealing with diverse and individual modalities. + +# 3. Methods + +# 3.1. Overall Framework + +In this paper, our primary focus is on multimodal tracking, with the constraint that only one modality is available at a time, as shown in Fig. 2. We define our multimodal + +![](images/9251a0b0f8c112a9e6b5721515276e5d325ad8135ce281de200d47396b7fc90d.jpg) +Figure 2. Our proposed framework, termed Un-Track, is composed of a shared embedding, a modal prompting, and a LoRA-finetuned pretrained RGB tracker. The shared embedding learns a joint representation that unifies all modalities (Sec. 3.2). The modal prompting block enhances feature modeling with modal awareness at each scale (Sec. 3.3). To track the target object, we finetune a pretrained foundation model [72] using the LoRA technique (Sec. 3.4). Our model achieves a unified model applicable across different modalities under a single parameter set. During inference, Un-Track seamlessly integrates any image-paired data, thanks to the emergent alignment. + +tracking dataset as $M = \{M^D, M^T, M^E\}$ , where $M^X$ represents the subset dataset with only a single modality $X$ available. Conventional methods tackling missing modalities often use dummy inputs for the absent pixels [31, 53], simulating complete datasets with all possible combinations of paired datasets. In contrast, our method transforms any auxiliary input into a shared embedding, seamlessly binding all modalities together and creating a complete and paired representation with the master RGB feature. + +To mitigate overfitting on sparse downstream multimodal datasets, we adopt a transformer-based RGB tracker with frozen parameters and fine-tune it for multimodal tracking. Leveraging a lightweight outer prompting method, we identify uncertain tokens at each scale and enhance them with cross-modal awareness. Simultaneously, an inner fine-tuning process is implemented using the LoRA technique [21]. + +During training, our model learns the shared embedding from samples in the mixed dataset $M$ , effectively binding all modalities together. As for inference, our model can accommodate any modal input $X$ , thanks to the emergent alignment. Our trainable parameters only include cross-modal binding, outer prompting parameters, and inner LoRA parameters, ensuring a training-friendly pipeline that can be efficiently employed end-to-end on a single 24G GPU. + +# 3.2. Shared Embedding + +Explicit Edge Awareness: We observe that, as illustrated in Fig. 2(a), depth data introduce 3D distance information, effectively delineating objects with varying granularity and enhancing the sharpness of 3D boundaries; thermal images generate a scene heat map, highlighting objects based on their temperatures and providing clearer contours; event data capture intensity changes, particularly around an object's outbound region. Notably, a consistent feature emerges across these modalities: the representation of the "true" objects' shape, often manifested through edges. + +Motivated by this observation, our objective is to harness + +edge embedding to unify all modalities. To achieve this, as shown in Fig. 3, we generate gradient maps from auxiliary modalities by computing differences between neighboring pixels along both the x- and y-axes. Simultaneously, without compromising texture edge, we also generate RGB gradient maps. Subsequently, all gradient maps are integrated with the visual feature, forming the gradient feature $G$ . + +Implicit Low-Rank Reconstruction: While edges present a shared feature across different modalities, exclusively transforming all modalities into edges may risk overlooking modality-specific clues. Therefore, we introduce an implicit learning pipeline to complement this by discovering the shared embedding, guided by the previously generated explicit edge awareness. This combined approach allows for a more effective identification of the shared embedding, leveraging both data-driven learning and edge priors. + +In technical terms, we redefine the challenge of learning the shared embedding as a quest for the shared low-rank approximation. Both objectives share the essence of distilling common features across all modalities. However, direct estimation of the low-rank approximation becomes impractical due to the distinct data domains and modal representations. In response, we propose a pragmatic strategy: decomposing the shared low-rank vector into the low-rank of each subset component. This alternative, more manageable and feasible within a single domain, lays the groundwork for approximating the global shared low-rank from these individual low-rank components. + +The overall algorithm can be found in Algorithm 1. Specifically, let $M$ be the input feature with mixed auxiliary modalities, decomposed into subset features $D,T,E$ representing depth, thermal, and event samples from subset datasets $M^D,M^T,M^E$ . Their respective $k_{th}$ low-rank matrices $D_{k},T_{k},E_{k}$ , are approximated by: + +$$ +D _ {k} = \sigma_ {d} (D), \quad T _ {k} = \sigma_ {t} (T), \quad E _ {k} = \sigma_ {e} (E), \tag {1} +$$ + +where $\sigma_{x}$ is the modality-specific learning through a simple MLP projecting features from input channel $c$ to the low + +# Algorithm 1 Implicit Shared Embedding + +Input: $M$ - Mix Modal Feature, $G$ - Explicit Gradient Binding. +Output: $F$ - Reconstructed feature + +1. Initialize: Separate $M$ into subsets, each representing a modality-specific feature: $D, T, E$ . +2. In-Domain Approximation: Train individual approximators $\sigma_{x}$ for each modality $x$ to derive domain-specific low-rank matrices $D_{k}, T_{k}, E_{k}$ . Idem for explicit low-rank gradient $G_{k}$ . +3. Fuse and Guide: Merge $D_{k}, T_{k}, E_{k}$ through a fusion function $\varphi_{1}$ . Incorporate the explicit gradient $G_{k}$ using addition, after projecting it into a compatible latent space through $\varphi_{2}$ . +4. Reconstruct: Train a reconstruction function $\Phi_R$ to construct the shared embedding $F$ , considering the explicit guidance $G$ in the original feature space. + +![](images/de2111bf26edd166c00a6777b458a86d3d1ebac711b3b88db515cb9ed22cb92c.jpg) +RGB: Image M: Mixed Modalities G: Gradient Feature F: Shared Embedding +Figure 3. Shared Embedding. We derive a joint representation through low-rank factorization and reconstruction. Such an implicit learning is additionally integrated with explicit edge awareness to enhance the embedding. + +rank space $k$ ( $k < c$ ). Simultaneously, we compute the low-rank matrix $G_{k}$ from the gradient feature $G$ . + +The global shared low-rank matrix $M_{k}$ is then approximated by fusing the subset low-rank matrices $D_{k}, T_{k}, E_{k}$ , along with the gradient guidance. Technically, we concatenate $D_{k}, T_{k}, E_{k}$ and learn the joint low-rank approximation, incorporating it with the gradient guidance. This pipeline can be expressed as follows: + +$$ +M _ {k} = \varphi_ {R _ {1}} \left(\left[ D _ {k}, T _ {k}, E _ {k} \right]\right) + \varphi_ {R _ {2}} (G _ {k}), \tag {2} +$$ + +where $[.]$ is the channel concatenation and $\varphi_{R_i}$ are the MLP projections to the low-rank latent space. Finally, we reconstruct the shared embedding $F$ through: + +$$ +F = \Phi_ {R} (M _ {k}) + G, \tag {3} +$$ + +where $\Phi_R$ is another MLP that projects the jointly-learned low-rank matrix back to the departing feature space. Our ablation studies validate that this subset low-rank approximation and regrouping efficiently unify all input modalities, despite their heterogeneous representations. + +![](images/97127b3f803f23f10955139fe90f88451f56c935043afdfdbc41bae3793fb447.jpg) +Figure 4. Modal Prompting. For the visual feature $I$ , we employ a score function to categorize tokens into negative, uncertain, and positive segments. Using a token exchange policy, we discard negative tokens, enhance uncertain ones with corresponding tokens from $F$ , and retain positive ones. Then, we transform the feature fusion task into a token recovery problem, addressed by low-rank factorization. Similarly, we extract the most informative low-rank matrix from $F$ to fuse and reconstruct the shared output. + +# 3.3. Outer Modal Prompting + +RGB-tracker may fail to perform accurately in corner cases where auxiliary clues can contribute. Drawing inspiration from the success of adapting large pre-trained models to specific downstream tasks [21, 25], we introduce a modal prompting method devised to enhance RGB token $I$ with modality-awareness $F$ , as shown in Fig. 4. + +Specifically, our approach employs a shrinkage token fusion strategy. Taking $I$ as an example, we categorize tokens into three groups—negative, uncertain, and positive—based on a dynamic scoring function $s$ . These regions are defined with mask form, expressed as $m_{n}, m_{u}, m_{p} = s(I)$ . To harness multimodal clues effectively, we replace negative tokens with those from the other modality, omit uncertain ones with dummy values, and retain the positive tokens. Subsequently, these modified tokens undergo projection into the low-rank space using the approximation function $\sigma_{c}$ . Our objective is to enhance robustness by completing uncertain tokens with information from reliable neighboring tokens. Mathematically, we obtain the first low-rank matrix $I_{l_1}$ by: + +$$ +I _ {l _ {1}} = \sigma_ {c} \left(m _ {n} \cdot F + m _ {p} \cdot I\right). \tag {4} +$$ + +Next, we target the uncertain tokens by merging them with the paired tokens from the other modality and approximate the low-rank matrix similarly. Here, we aim to throw out the possible noise, resulting in a matrix that is more informative than the original. Let $\sigma_{n}$ be another approximation function, we obtain the second low-rank matrix $I_{l_n}$ by: + +$$ +I _ {l _ {2}} = \sigma_ {n} \left(m _ {u} \cdot F + m _ {u} \cdot I\right). \tag {5} +$$ + +Then, we fuse these two low-rank matrices in low-rank + +Depth-Specific Parameters +Uni-model with a Single Set of Parameters + +
ATCAIS [28]DDiMP [28]DeT [66]SPT [94]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]UniNext [65]SeqTrack [8]ViPT [91]Un-Track (ours)
F-score(↑)0.4760.4850.5320.5380.5780.5940.6120.3970.5150.5690.4220.5900.5610.610
Re(↑)0.4550.4690.5060.5490.5730.5960.6100.4060.5260.5820.4320.6000.5620.610
Pr(↑)0.5000.5030.5600.5270.5830.5920.6130.3880.5050.5570.4130.5800.5600.610
+ +space, forming the shared low-rank matrix $I_{l}$ for input $I$ . In such a manner, we improve the image token modeling by fully benefiting from the auxiliary clues. Mathematically, the whole process can be formulated as: + +$$ +I _ {l} = \varphi_ {P} ([ I _ {l _ {1}}, I _ {l _ {2}} ]), \tag {6} +$$ + +where $\varphi_{P}$ is learnable fusion. Similarly, from the input $F$ , we follow the same process to obtain the low-rank matrix $F_{l}$ . For the cross-modal fusion, we add these two low-rank matrices and then project back to the original space. We can obtain the fused output $O$ by: + +$$ +O = \Phi_ {P} \left(I _ {l} + F _ {l}\right), \tag {7} +$$ + +where $\Phi_P$ is another learnable MLP. + +Our method can be regarded as a mixer of token exchange (for negative tokens) and token fusion (for uncertain tokens), while retaining the most informative modality-specific clues (for positive tokens). As the majority of fusion operations occur in the low-rank feature space, our progressive cross-modal shrinkage avoids imposing a significant additional computational burden, while being able to excavate and accumulate the rich cues from each modality for effective modal prompting. + +# 3.4. Inner Finetuning + +In addition to the outer modal prompting, we incorporate the LoRA technique [21] for more efficient finetuning. For each transformer attention module with the weight matrix $W_0 \in \mathbb{R}^{d \times k}$ , we introduce two learnable matrices: $B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$ . This leads to the replacement of the frozen attention mechanism $h = W_0 x$ with the new LoRA attention: + +$$ +h = W _ {0} \boldsymbol {x} + B A \boldsymbol {x}. \tag {8} +$$ + +To train our network, we adopt the same loss functions as our baseline tracker [72] for end-to-end learning. + +# 4. Experiments + +# 4.1. Training Data + +In the absence of a comprehensive multi-modal tracking dataset encompassing all possible combinations (RGB-D-T-E), we have only one $\mathrm{RGB + X}$ (i.e. RGB+depth, or + +RGB+thermal, or RGB+event) at a time. The conventional modality-specific adapted settings are referred as "X-Specific", whereas the main target of this paper with one model trained on all modality pairs is called "Uni-model" (or a single set of parameters). The training and evaluation settings are summarized as follows: + +Table 1. Overall performance on DepthTrack test set [66]. Red/Green/Blue stands for the best/second/third performance. Our depth-specific model sets the SOTA record, while our uni-model with single parameters set also outperforms previous depth-specific SOTA. + +
ModelTrained onEvaluated on
X-SpecificRGBi+XiRGBi+Xi
Uni-model∪iRGBi+XiRGBi+Xi
+ +Our RGB-D samples are sourced from DepthTrack [66], a pioneering RGB-D tracking benchmark with 150 training long-term sequences. RGB-T samples are extracted from the extensive LasHeR [36] dataset, featuring 979 diverse training sequences. RGB-E samples are obtained from VisEvent [55], which boasts 500 real-world sequences. Each D/T/E input is transformed into an RGB-like form. + +# 4.2. Within distribution Evaluation + +Given that DepthTrack [66], Lasher [36], and VisEvent [55] provide domain-specific testing sequences, our initial evaluation focuses on these within distribution sequences. For each dataset, we adhere to the metrics specified in the original papers and prior standards [69, 91] for evaluation. + +Comparison on DepthTrack [66]: For evaluation, we use metrics such as precision (Pr) and recall (Re), as well as the F-score, which are the primary metrics. As shown in Tab. 1 When exclusively trained and fine-tuned on DepthTrack, Un-Track achieves a $+2.1\%$ absolute precision improvement over the current depth-specific SOTA ViPT [91]. Notably, even when trained with mixed data using a single parameter set, Un-Track outperforms the depth-specific ViPT. Furthermore, when the current ViPT is jointly trained on all datasets with a single set of parameters, its performance significantly deteriorates. + +Additionally, we observe that trackers excelling in other tracking datasets [14, 22, 42] might struggle in RGB-D downstream settings. UniNext [65], a leading tracker trained on various large-scale tracking datasets and related image/video tasks, exhibits poor performance. In contrast, our model achieves cross-modal unification within a single set of parameters, surpassing all depth-specific counterparts and performing closely to our specialized version. This underscores the efficacy of our shared embedding in achieving global alignment across diverse modalities. + +Thermal-Specific Parameters +Uni-model with a Single Set of Parameters + +
SGT [34]FANet [96]mfDiMP [81]DAFNet [18]MaCNet [77]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
PR(↑)0.3270.4410.4470.4480.4820.5380.6510.6670.4180.4630.5300.5820.6080.646
SR(↑)0.2320.3090.3430.3110.3500.4200.5250.5360.3330.3650.4220.4410.4900.513
+ +Table 2. Overall performance on the Lasher thermal testing set [36]. Our thermal-specific model sets a new SOTA record. Our uni-model variant competes strongly with the previous SOTA thermal-specific models and significantly surpasses its unimodel version. +Event-Specific Parameters +Uni-model with a Single Set of Parameters + +
SiamCar [62]Stark_E [62]VITAL_E [46]PrDiMP_E [13]TransT_E [5]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
Precision(↑)0.5990.6120.6490.6440.6500.6320.7580.7630.6160.6260.6910.6650.7400.755
Success(↑)0.4200.4460.4150.4530.4740.4710.5920.5970.4480.4440.5250.5040.5790.589
+ +Table 3. Overall performance on VisEvent dataset [55]. Our event-specific model sets a new SOTA record. Our uni-variant, with the same parameter set as in Depth and Thermal, consistently achieves competitive performance across various modalities, leading to a significant margin over the uni-variant of the SOTA modality-specific model [91]. + +Comparison on LasHer [36]: Similarly, we conduct evaluations on the LasHer testing set for RGB-T tracker assessment as shown in Tab. 2. Precision (PR) and success rates (SR) are reported following conventional standards [36, 69, 91]. Initial comparisons under domain-specific settings reveal the challenges of achieving an architecture-unified model across RGB-D and RGB-T domains, with ProTrack [69] and ViPT [91] being the only works consistently leading in both settings. Our Un-Track, following domain-specific finetuning, outperforms the leading ViPT by a significant margin and sets a new SOTA record. In the cross-domain joint learning with a single set of parameters, our uni-model achieves a $+3.8\%$ absolute gain over ViPT. Remarkably, our model with a single set of parameters already achieves very competitive performance compared to the thermal-specific ViPT version. + +Comparison on VisEvent [55]: We also evaluate tracker performance with RGB-Event input. Event data, being inherently sparse compared to depth and thermal information, presents challenges in extending existing RGB-D or RGB-T fusion designs for effective integration, hence necessitating specific fusion designs [3, 43, 98]. In contrast, we adopt a unified cross-modal prompting method based on shrinkage fusion. Our approach, with gradual token exchanges between RGB and event modalities, effectively preserves crucial modality-specific clues to enhance feature modeling. Performance-wise, under the event-specific setting, our Un-Track outperforms all counterparts, as shown in Tab. 3 and in Fig. 5. In the single set of parameters setting, our uni-model achieves a $+1.1\%$ absolute gain in precision over the current SOTA. This underscores the effectiveness of our approach in handling the unique challenges posed by event data integration, which can be mainly contributed to our shared binding that learns the global RGB+X alignment. + +![](images/29f2be60cfcc7d917d98781312dfdc23fd658d477dd5208a20bec076dbaadb10.jpg) +Figure 5. More precision/success comparisons on VisEvent dataset [55]. "Uni" stands for models with a single parameter set. "E" stands for the extension of RGB trackers with event fusion. + +![](images/3d89a35b2ccda4e1346510bdcf6a832c3983146410eb10179321964ca153ee48.jpg) + +# 4.3. Generalization Across Datasets + +In this section, we assess the versatility by evaluating performance on datasets that differ from the training ones, aligning with the goal of achieving a universal model checkpoint applicable to diverse scenarios. + +VOT-RGBD2022 [30]: We first perform inference on the VOT-RGBD2022 dataset. Notably, our uni-model achieves superior performance compared to the depth-specific ViPT with a $+0.5\%$ absolute gain in accuracy. + +RGBT234 [35]: We also test our model on the other thermal dataset RGBT234, which encompasses sequences with different distributions. Our uni-model surpasses the thermal-specific ViPT with a notable $+0.7\%$ absolute precision gain, as shown in Tab. 5. + +Moreover, we present a detailed per-attribute comparison with the fine-tuned ViPT in Fig. 6. We are particularly interested in sequences with motion blur-fast motion-camera moving, as well as sequences with heavy occlusion-scale variants-background clutter. The former motion-related challenges can typically benefit from event clues, renowned for asynchronous computing, while the latter geometry + +Depth-Specific Parameters +Uni-model with a Single Set of Parameters + +
DRefine [29]Stark_D [29]DMTracker [30]DeT [66]SBT_D [30]SPT [94]ProTrack [69]ViPT [91]Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
EAO(↑)0.5920.6470.6580.6570.7080.6510.6510.7210.4450.6410.6660.6790.718
Accuracy(↑)0.7750.8030.7580.7600.8090.7980.8010.8150.7140.7690.8080.8020.820
Robustness(↑)0.7600.7980.8510.8450.8640.8510.8020.8710.5980.8320.8140.8460.864
+ +Table 4. Overall performance on VOT-RGBD2022 [30]. Our uni-model, trained on a mix of all modalities, shows robust generalization and outperforms all depth-specific models and other uni-model counterparts. + +
Thermal-Specific ParametersUni-model with a Single Set of Parameters
mfDiMP [81]SGT [34]DAFNet [18]FANet [96]MaCNet [77]CMPP [52]APFNet [59]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
MPR(↑)0.6460.7200.7960.7870.7900.8230.8270.7950.8350.8370.6770.7110.7550.8060.842
MSR(↑)0.4280.4720.5440.5530.5540.5750.5790.5990.6170.6180.4960.5080.5690.5990.625
+ +![](images/98c0d70b043dfc08881f66c6ef0e3a941242ab9ff552177207e05de66533baea.jpg) +Figure 6. Per-attribute analysis on the thermal dataset RGBT234 [35]. Challenges related to motion and geometry are generally better addressed by event and depth cameras, respectively. Nevertheless, when inferring only with RGB-T data, our Un-Track surpasses both the SOTA thermal-specific method and the current leading uni-tracker. This success underscores our ability to learn emergent alignment across diverse modalities. + +related challenges can typically benefit from depth cameras. However, these event/depth clues are not directly available in the RGB-T setting. Nevertheless, our Un-Track, trained on all modalities, outperforms both thermal fine-tuned ViPT [91] and the current leading uni-tracker [8] with large margins. This underscores our capability in learning event and depth priors through shared binding, without the need for the presence of these modalities during inference. + +RGB-only: In practical scenarios, challenges arise when there are no modal clues available, a typical case is when the auxiliary sensor fails to work properly. We address this demanding case in our study by substituting the modal input with dummy values. Under a such challenging setting, as shown in Tab. 6, our uni-method consistently outperforms both our RGB baseline fine-tuned counterparts significantly. Notably, such an improvement is achieved with a very limited increase in learning parameters, i.e., $+6.65\mathrm{M}$ with $+2.14$ GFLOPs. + +Table 5. Overall performance on RGBT234 dataset [35]. Our uni-model sets new SOTA records without specific thermal finetuning. + +
RGB Baseline [72]ViPT [91]Un-Track (ours)
GFLOPs21.5021.8023.64
Params (M)92.0892.9698.73
F-score(↑)0.5290.5420.558
Re(↑)0.5220.5380.557
Pr(↑)0.5360.5460.560
+ +Table 6. Overall performance on DepthTrack test set [66] with dummy depth input. + +
w/o Shared Embed[91] as Modal Promptw/o LoRA Finetune
F-score(↑)0.5990.5790.594
Re(↑)0.6020.5750.598
Pr(↑)0.5970.5840.596
+ +Table 7. Key component analysis. + +# 5. Ablation Studies + +In this section, we perform all experiments on the DepthTrack testing set [66] under a single parameter set setting. Key Component Analysis: We begin by studying the effectiveness of key components, including the shared embedding, modal prompting, and LoRA finetuning, as summarized in Tab. 7. We initially remove the shared embedding by directly feeding mixed modalities into the learning diagram. It can be seen that this equal treatment of all modalities harms network performance due to the heterogeneous representation across modal domains. Secondly, we replace our prompting block with a recent counterpart that computes fovea attention from additional input [91]. Our designed gradual shrinkage fusion, allowing token-wise interaction, proves to be more effective. We also report performance when we remove the inner fine-tuning with Lora. It can be seen that the performance deteriorates. + +Low-Rank: Low-rank approximation plays a vital role in our model, influencing our shared embedding, modal fusion, and LoRA-finetuning. Hence, the choice of rank is crucial for both the efficiency and effectiveness of our ap + +Table 8. Low Rank Approximation Our plain version is highlighted in gray. + +(a) $\text{Rank}_k(\text{Sec. 3.2})$ + +
248
F-score(↑)0.6070.6100.602
Re(↑)0.6060.6080.601
Pr(↑)0.6080.6110.604
+ +(b) $\text{Rank}_l$ (Sec. 3.3) + +
4816
0.5960.6100.606
0.5930.6080.609
0.5990.6110.604
+ +(c) LoRA (Sec. 3.4) + +
248
0.6010.6100.600
0.5990.6080.598
0.6020.6110.602
+ +(d) Percentile (Sec. 3.3) + +
1/81/41/3
0.6040.6100.595
0.6060.6080.593
0.6020.6110.596
+ +proach. In Tab. 8, we systematically explore the impact of low-rank choices within each component. + +Shared Embedding: Our objective is to identify an optimal low-rank latent space for merging different modalities effectively. Tab. 8(a) presents our ablation study, where we explore ranks of 2, 4, and 8. Lower ranks result in poorer performance due to sparse representations. Conversely, higher ranks capture too many modality-specific details, complicating the search for a shared embedding. + +Modal Prompting: As shown in Tab. 8(b), similar trends are observed when investigating the low-rank choices for modal prompting with rank values of 4, 8, and 16, as low ranks struggle to capture essential information, while higher ranks introduce an overload of modality-specific details. + +LoRA-finetuning: We also vary the ranks between 2, 4, and 8 for the LoRA finetuning technique, as shown in Tab. 8(c). Lower ranks exhibited consistently poor performance, while higher ranks in this case tend to result in poorer performance, likely due to overfitting. + +Remarks: These experiments emphasize the importance of selecting the best low-rank representations. Nevertheless, our model shows great resilience to the choice of LoRA, showcasing consistent performance across different low-rank configurations. Notably, all our low-rank variants outperform the current SOTA ViPT under the uni-setting, validating our robustness and effectiveness. + +Modal Prompting: During the prompting fusion, a learnable score function is used to categorize tokens into three groups based on their confidence scores. Here, we explore different percentiles for the number of positive, which is the same as the number of negative tokens, leaving the rest as uncertain tokens. As shown in Tab. 8(d), the choice of percentile can influence overall performance. Lower percentiles result in poorer performance, as recovering uncertain tokens from very few neighboring tokens is challenging. Higher percentiles also lead to performance degradation since the focus is on token exchange rather than token fusion. The choice of $1/4$ leads to the best balance between exchanger and fuser, leading to the best performance. + +Shared Embedding: Here, we perform ablation studies on our shared embedding, a foundational component of our uni-model. The quantitative results are presented in Tab. 9. We begin by exploring a scenario where our shared embedding is replaced with a variant lacking explicit edge guidance (w/o Explicit Edge). In this configuration, the network + +
w/o Explicit Edgew/o Implicit Learningw/o In-domain Approx.
F-score(↑)0.6000.6040.581
Re(↑)0.6020.6090.583
Pr(↑)0.5980.5990.579
+ +Table 9. Ablation on shared embedding + +learns the shared embedding solely without any edge prior. The results highlight a substantial performance drop, underscoring the pivotal role of explicit edge guidance — a natural and static embedding that binds all modalities together — in facilitating this implicit learning process. + +We also conduct experiments using only the explicit edge as the shared embedding, excluding any additional learning modules (w/o Implicit Learning). This approach, too, yields suboptimal performance, primarily due to the neglect of modality-specific clues within each domain. + +Finally, we directly compute the low-rank approximation from mixed modalities, bypassing the initial in-domain approximation and subsequent fusion steps (w/o In-domain Approx.). We observe that the network struggles in learning a shared embedding in this direct low-rank approximation setup, primarily due to the intricately mixed representation caused by the domain gap between different modalities. + +# 6. Conclusion and Future Work + +We present a successful case of a single-model and any-modality tracker for video object tracking. The proposed method achieves a shared embedding that binds all modalities together, overcoming their heterogeneous representations. This unification is facilitated by lightweight modal prompting and inner finetuning, inheriting benefits from large-scale pre-trained trackers without introducing a substantial computational burden. Exhaustive experiments showcase our improved tracking performance and robust generalization, with any modality input. + +Acknowledgements. The authors thank the reviewers and ACs for their tremendous efforts and helpful comments. The event icon is credited to Zuowen Wang. This research is financed in part by the Alexander von Humboldt Foundation, in part by NSFC (62376156, 62322113), and in part by the Ministry of Education and Science of Bulgaria (support for INSAIT, part of the Bulgarian National Roadmap for Research Infrastructure). + +# References + +[1] Simon Arridge, Pascal Fernsel, and Andreas Hauptmann. Joint reconstruction and low-rank decomposition for dynamic inverse problems. Inverse Problems and Imaging, 16 (3):483-523, 2022. 2 +[2] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. In ECCV, 2022. 2 +[3] Qinyu Chen, Zuowen Wang, Shih-Chii Liu, and Chang Gao. 3et: Efficient event-based eye tracking using a change-based convlstm network. arXiv preprint arXiv:2308.11771, 2023. 6 +[4] Wanli Chen, Xinge Zhu, Ruoqi Sun, Junjun He, Ruiyu Li, Xiaoyong Shen, and Bei Yu. Tensor low-rank reconstruction for semantic segmentation. In ECCV, 2020. 2 +[5] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2, 6 +[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2 +[7] Xin Chen, Bin Yan, Jiawen Zhu, Huchuan Lu, Xiang Ruan, and Dong Wang. High-performance transformer tracking. TPAMI, 2022. 2 +[8] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 5, 6, 7 +[9] Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, and Rongrong Ji. Siamese box adaptive network for visual tracking. In CVPR, 2020. 1 +[10] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In CVPR, 2022. 1 +[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 2 +[12] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 1 +[13] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6 +[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 5 +[15] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. TIP, 2023. 1 +[16] Shang Gao, Jinyu Yang, Zhe Li, Feng Zheng, Alex Leonardis, and Jingkuan Song. Learning dual-fused modality-aware representations for rgbd tracking. In ECCV, 2022. 2 +[17] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In ECCV, 2022. 5, 6, 7 + +[18] Yuan Gao, Chenglong Li, Yabin Zhu, Jin Tang, Tao He, and Futian Wang. Deep adaptive fusion network for high performance RGBT tracking. In ICCVW, 2019. 6, 7 +[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 2 +[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In IROS, 2021. 1 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4, 5 +[22] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5 +[23] Sajid Javed, Martin Danelljan, Fahad Shahbaz Khan, Muhammad Haris Khan, Michael Felsberg, and Jiri Matas. Visual object tracking with discriminative filters and siamese networks: a survey and outlook. TPAMI, 45(5):6552-6574, 2022. 1 +[24] I-Hong Jhuo, Dong Liu, DT Lee, and Shih-Fu Chang. Robust visual domain adaptation with low-rank reconstruction. In CVPR, 2012. 2 +[25] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 4 +[26] Ben Kang, Xin Chen, Dong Wang, Houwen Peng, and Huchuan Lu. Exploring lightweight hierarchical vision transformers for efficient visual tracking. In ICCV, 2023. 2 +[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2 +[28] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, Ondrej Drbohlav, et al. The eighth visual object tracking vot2020 challenge results. In ECCVW, 2020. 5 +[29] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In ICCVW, 2021. 7 +[30] Matej Kristan, Aleš Leonardis, Jiri Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, et al. The tenth visual object tracking vot2022 challenge results. In ECCVW, 2023. 6, 7 +[31] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 2, 3 +[32] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2 + +[33] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2 +[34] Chenglong Li, Nan Zhao, Yijuan Lu, Chengli Zhu, and Jin Tang. Weighted sparse representation regularized graph learning for RGB-T object tracking. In ACM MM, 2017. 6, 7 +[35] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. RGB-T object tracking: Benchmark and baseline. PR, 96:106977, 2019. 2, 6, 7 +[36] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for RGBT tracking. TIP, 31:392-404, 2021. 2, 5, 6 +[37] Qiao Liu, Xin Li, Zhenyu He, Nana Fan, Di Yuan, and Hong-peng Wang. Learning deep multi-level similarity for thermal infrared object tracking. TMM, 23:2114-2126, 2020. 2 +[38] Cheng Long Li, Andong Lu, Ai Hua Zheng, Zhengzheng Tu, and Jin Tang. Multi-adapter rgbt tracking. In ICCVW, 2019. 2 +[39] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2023. 2 +[40] Alan Lukezic, Ugur Kart, Jani Kapyla, Ahmed Durmush, Joni-Kristian Kamarainen, Jiri Matas, and Matej Kristan. Cdtb: A color and depth visual object tracking dataset and benchmark. In ICCV, 2019. 1 +[41] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 2 +[42] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2, 5 +[43] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. In ICCV, 2023. 6 +[44] Yansheng Qiu, Ziyuan Zhao, Hongdou Yao, Delin Chen, and Zheng Wang. Modal-aware visual prompting for incomplete multi-modal brain tumor segmentation. In ACM MM, 2023. 2 +[45] Esteban Real, Jonathon Shlens, Stefano Mazzocchi, Xin Pan, and Vincent Vanhoucke. Youtube-boundingboxes: A large high-precision human-annotated data set for object detection in video. In CVPR, 2017. 2 +[46] Yibing Song, Chao Ma, Xiaohe Wu, Lijun Gong, Linchao Bao, Wangmeng Zuo, Chunhua Shen, Rynson WH Lau, and Ming-Hsuan Yang. Vital: Visual tracking via adversarial learning. In CVPR, 2018. 6 +[47] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In CVPR, 2021. 1 +[48] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1 + +[49] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 2 +[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: Visual tracking by redetection. In CVPR, 2020. 2 +[51] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In ICCV, 2023. 2 +[52] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for RGB-T tracking. In CVPR, 2020. 7 +[53] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In CVPR, 2023. 3 +[54] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2 +[55] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. arXiv preprint arXiv:2108.05015, 2021. 2, 5, 6 +[56] Xiao Wang, Xiujun Shu, Shilliang Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Mfgnet: Dynamic modality-aware filter generation for rgb-t tracking. TMM, 2022. 2 +[57] Zuowen Wang, Yuhuang Hu, and Shih-Chii Liu. Exploiting spatial sparsity for event cameras with visual transformers. In ICIP, 2022. 2 +[58] David Wisth, Marco Camurri, Sandipan Das, and Maurice Fallon. Unified multi-modal landmark tracking for tightly coupled lidar-visual-inertial odometry. RA-L, 6(2):1004-1011, 2021. 2 +[59] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for RGBT tracking. In AAAI, 2022. 7 +[60] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for rgbt tracking. In AAAI, 2022. 1, 2 +[61] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. SiamFC++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, 2020. 2 +[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 5, 6, 7 +[63] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 2 +[64] Bin Yan, Houwen Peng, Kan Wu, Dong Wang, Jianlong Fu, and Huchuan Lu. Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search. In CVPR, 2021. 2 +[65] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 5 + +[66] Song Yan, Jinyu Yang, Jani Käpylä, Feng Zheng, Ales Leonardis, and Joni-Kristian Kämäräinen. Depthtrack: Unveiling the power of RGBD tracking. In ICCV, 2021. 2, 5, 7 +[67] Song Yan, Jinyu Yang, Ales Leonardis, and Joni-Kristian Kamarainen. Depth-only object tracking. In BMVC, 2021. 2 +[68] Jinyu Yang, Zhe Li, Song Yan, Feng Zheng, Aleš Leonardis, Joni-Kristian Kāmäräinen, and Ling Shao. Rgbd object tracking: An in-depth review. arXiv preprint arXiv:2203.14134, 2022. 1 +[69] Jinyu Yang, Zhe Li, Feng Zheng, Ales Leonardis, and Jingkuan Song. Prompting for multi-modal tracking. In ACMM, 2022. 1, 2, 5, 6, 7 +[70] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Aleš Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, 2023. 2 +[71] Rui Yao, Guosheng Lin, Shixiong Xia, Jiaqi Zhao, and Yong Zhou. Video object segmentation and tracking: A survey. ACM TIST, 11(4):1-47, 2020. 2 +[72] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In ECCV, 2022. 3, 5, 6, 7 +[73] Jie Yin, Ang Li, Tao Li, Wenxian Yu, and Danping Zou. M2dgr: A multi-sensor and multi-scenario slam dataset for ground robots. RA-L, 7(2):2266-2273, 2021. 2 +[74] Yuechen Yu, Yilei Xiong, Weilin Huang, and Matthew R Scott. Deformable siamese attention networks for visual object tracking. In CVPR, 2020. 2 +[75] Jiandian Zeng, Tianyi Liu, and Jiantao Zhou. Tag-assisted multimodal sentiment analysis under uncertain missing modalities. In ACM SIGIR, 2022. 2 +[76] Chunhui Zhang, Xin Sun, Yiqian Yang, Li Liu, Qiong Liu, Xi Zhou, and Yanfeng Wang. All in one: Exploring unified vision-language tracking with multi-modal alignment. In ACM MM, 2023. 2 +[77] Hui Zhang, Lei Zhang, Li Zhuo, and Jing Zhang. Object tracking in RGB-T videos using modal-aware attention network and competitive learning. Sensors, 20(2):393, 2020. 6, 7 +[78] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Bao-cai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, 2021. 2 +[79] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 1 +[80] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In CVPR, 2023. 2 +[81] Lichao Zhang, Martin Danelljan, Abel Gonzalez-Garcia, Joost van de Weijer, and Fahad Shahbaz Khan. Multi-modal fusion for end-to-end RGB-T tracking. In ICCVW, 2019. 6, 7 +[82] Pengyu Zhang, Dong Wang, and Hutchuan Lu. Multi-modal visual tracking: Review and experimental comparison. arXiv preprint arXiv:2012.04176, 2020. 1 + +[83] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, 2022. 1 +[84] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In ICCV, 2019. 1 +[85] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 2 +[86] Haojie Zhao, Dong Wang, and Hutchuan Lu. Representation learning for visual object tracking by masked appearance transfer. In CVPR, 2023. 1 +[87] Jinjian Zhao, Xiaohan Zhang, and Pengyu Zhang. A unified approach for tracking uavs in infrared. In ICCV, 2021. 1 +[88] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. PR, 111:107679, 2021. 2 +[89] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, 2021. 1 +[90] Jiawen Zhu, Zhenyu Chen, Zeqi Hao, Shijie Chang, Lu Zhang, Dong Wang, Huchuan Lu, Bin Luo, Jun-Yan He, Jin-Peng Lan, et al. Tracking anything in high quality. arXiv preprint arXiv:2307.13974, 2023. 1 +[91] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, 2023. 1, 2, 5, 6, 7 +[92] Jiawen Zhu, Huayi Tang, Zhi-Qi Cheng, Jun-Yan He, Bin Luo, Shihao Qiu, Shengming Li, and Huchuan Lu. Dcpt: Darkness clue-prompted tracking in nighttime uavs. arXiv preprint arXiv:2309.10491, 2023. 1 +[93] Xue-Feng Zhu, Xiao-Jun Wu, Tianyang Xu, Zhen-Hua Feng, and Josef Kittler. Robust visual object tracking via adaptive attribute-aware discriminative correlation filters. TMM, 24: 301-312, 2021. 2 +[94] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. RGBD1K: A large-scale dataset and benchmark for RGB-D object tracking. AAAI, 2023. 2, 5, 7 +[95] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking. In AAAI, 2023. 2 +[96] Yabin Zhu, Chenglong Li, Jin Tang, and Bin Luo. Quality-aware feature aggregation network for robust RGBT tracking. IEEE TIV, 6(1):121-130, 2020. 6, 7 +[97] Zhiyu Zhu, Junhui Hou, and Xianqiang Lyu. Learning graph-embedded key-event back-tracing for object tracking in event clouds. NeurIPS, 2022. 2 +[98] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event transformer-trackers. In ICCV, 2023. 1, 6 \ No newline at end of file diff --git a/2024/Single-Model and Any-Modality for Video Object Tracking/images.zip b/2024/Single-Model and Any-Modality for Video Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..97bae9994da5c4314f51dd89052972c1ab1ce04e --- /dev/null +++ b/2024/Single-Model and Any-Modality for Video Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a2d762fc942df717f8a8a7516ed31510673f89d3c726bdc0c7988db7d8de0e7 +size 530185 diff --git a/2024/Single-Model and Any-Modality for Video Object Tracking/layout.json b/2024/Single-Model and Any-Modality for Video Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f00ec176d0aca75580ac4d56a1a9de81a6de0d20 --- /dev/null +++ b/2024/Single-Model and Any-Modality for Video Object Tracking/layout.json @@ -0,0 +1,10587 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 103, + 480, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 103, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 113, + 103, + 480, + 121 + ], + "type": "text", + "content": "Single-Model and Any-Modality for Video Object Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "spans": [ + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": "Zongwei Wu" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": " Jilai Zheng" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": " Xiangxuan Ren" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": " Florin-Alexandru Vasluianu" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": " Chao Ma" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": " Danda Pani Paudel" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": " Luc Van Gool" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "text", + "content": " Radu Timofte" + }, + { + "bbox": [ + 100, + 141, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 172, + 502, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 172, + 502, + 200 + ], + "spans": [ + { + "bbox": [ + 91, + 172, + 502, + 200 + ], + "type": "text", + "content": "1 Computer Vision Lab, CAIDAS & IFI, University of Wurzburg 2 AI Institute, Shanghai Jiao Tong University 3 INSAIT, Sofia University 4 CVL, ETH Zurich" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 253, + 290, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 290, + 601 + ], + "type": "text", + "content": "In the realm of video object tracking, auxiliary modalities such as depth, thermal, or event data have emerged as valuable assets to complement the RGB trackers. In practice, most existing RGB trackers learn a single set of parameters to use them across datasets and applications. However, a similar single-model unification for multimodality tracking presents several challenges. These challenges stem from the inherent heterogeneity of inputs - each with modality-specific representations, the scarcity of multimodal datasets, and the absence of all the modalities at all times. In this work, we introduce Un-Track, a Unified Tracker of a single set of parameters for any modality. To handle any modality, our method learns their common latent space through low-rank factorization and reconstruction techniques. More importantly, we use only the RGB-X pairs to learn the common latent space. This unique shared representation seamlessly binds all modalities together, enabling effective unification and accommodating any missing modality, all within a single transformer-based architecture. Our Un-Track achieves +8.1 absolute F-score gain, on the DepthTrack dataset, by introducing only +2.14 (over 21.50) GFLOPs with +6.6M (over 93M) parameters, through a simple yet efficient prompting strategy. Extensive comparisons on five benchmark datasets with different modalities show that Un-Track surpasses both SOTA unified trackers and modality-specific counterparts, validating our effectiveness and practicality. The source code is publicly available at https://github.com/Zongwei97/UnTrack." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 624, + 128, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 128, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 128, + 637 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 645, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 645, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 645, + 287, + 694 + ], + "type": "text", + "content": "Video object tracking [9, 12, 86, 90] is a fundamental task in computer vision with wide-ranging applications spanning from surveillance [89] to augmented reality [23], where accuracy and robustness are paramount. While traditional" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 228, + 544, + 419 + ], + "blocks": [ + { + "bbox": [ + 312, + 228, + 544, + 419 + ], + "lines": [ + { + "bbox": [ + 312, + 228, + 544, + 419 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 544, + 419 + ], + "type": "image", + "image_path": "a9333d4bb350e45c134026b34eab75cd3425c27e62ff5aff14b298ff7dfb44f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 422, + 547, + 444 + ], + "lines": [ + { + "bbox": [ + 305, + 422, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 547, + 444 + ], + "type": "text", + "content": "Figure 1. Un-Track is a unified tracker with a single parameter set that seamlessly integrates any modality (of RGB-X)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "type": "text", + "content": "RGB trackers have shown promise in general settings, they often struggle with challenging scenarios such as occlusions [40, 47], low visibility [83, 92], or fast-moving objects [10, 48]. For more reliable tracking under such challenging conditions, the integration of auxiliary modalities (X) like depth [20, 68], thermal [83, 87], and event [79, 98] have proven effective in multimodal tracking." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 546, + 675 + ], + "type": "text", + "content": "While the idea of fusing RGB with other modalities holds promise [82, 84], the main challenge is the discrepancy in the representation of information across different modalities. Despite the success of previous fusion works [15, 60] tailored for each specific scenario to improve RGB trackers, their reliance on modality-specific designs limits adaptability. Recent initiatives [69, 91] towards a uniform architecture for various modalities show promise but necessitate modality-specific fine-tuning. This approach leads to multiple parameter sets, as shown in Fig. 1(a), thereby compromising practicality in diverse real-world applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "In this work, we aim to avoid such modality-specific fine-tuning to keep only one model-parameter set at all times. Another practical constraint arises from the differ" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 138, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19156" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "ences in available auxiliary modalities across settings. Unifying modalities by a common representation can handle any modality at its disposal, addressing the mentioned problems by its virtue. However, two additional multifaceted challenges emerge from the scarcity of multimodal datasets and the absence of all paired data combinations. The former makes cross-modal mappings through a large-scale data prior unfeasible [19, 39, 76], while the latter leads to missing modalities and renders joint learning using all possible combinations of paired data unfeasible [41, 80]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "type": "text", + "content": "To achieve such a unification, in this paper, we present Un-Track, denoting \"one\" in French, which learns a cohesive embedding across diverse input modalities. Unlike previous approaches [51, 58, 73, 80], Un-Track relies solely on RGB-X pairs for training, with X representing auxiliary modalities, without the need for all modalities to co-occur. Our objective is to discover a shared embedding seamlessly binding all auxiliary modalities (as depicted in Fig. 1(b)). More specifically, we leverage the factorization prior, allowing reasoning about a common embedding directly from the low-rank latent space. Factorization is a simple composition prior with the assumption that the global approximation can be constructed from a set of subset vectors. The factorization prior, successfully utilized in previous studies [1, 4, 24], is employed in our work to reconstruct a shared embedding. This process transforms the heterogeneous modal representation into a uniform one, thereby facilitating the emergent cross-modal alignment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 407, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 289, + 647 + ], + "type": "text", + "content": "Moreover, to harness the full potential of auxiliary inputs while maintaining efficiency, we leverage cross-modal features as prompts to enable RGB-X interaction. Different from previous works [25, 27], our goal is to enhance less reliable tokens, defined by a learnable score function, using multimodal cues. We approach this as a token recovery problem and leverage low-rank factorization to achieve the goal, which is first suggested in multimodal fusion, to the best of our knowledge. With its unified model architecture and prompting blocks, Un-Track is the first to offer support for cross-modality alignment under a single architecture with uniform parameters. In comparison to our RGB baseline with 21.50 GFLOPs and 92M parameters, Un-Track introduces only +2.14 GFLOPs with +6.6M parameters, resulting in a significant +8.1 absolute F-score gain demonstrated on the DepthTrack dataset. Extensive comparisons across five datasets with different modalities validate UnTrack's superiority over specialized SOTA models, surpassing both unified trackers and modality-specific fine-tuned counterparts by a substantial margin." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 138, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 138, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 138, + 669 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Multimodal tracking: Video object tracking [7, 71] aims to detect objects in a video sequence based on their initial positions. Early approaches treated tracking as a per-frame" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "target-matching problem, with Siamese networks [32, 33, 50, 61, 74, 85] being a notable example. More recently, transformer-based methods [2, 5, 6, 11, 54, 63] have gained popularity for feature extraction and per-frame correlation in tracking. Large-scale training datasets [14, 22, 42, 45] have empowered RGB trackers to uniformly apply parameters across various applications. While RGB trackers deliver promising results, challenges such as occlusion, low illumination, and fast-moving scenes have led to the exploration of additional modalities. Several works have investigated the role of depth [70, 95], thermal [37, 38], and event modalities [49, 55] in enhancing tracking performance. Specifically, depth cues [16, 67] contribute to handling objects with different camera distances; thermal cameras [56, 60] address challenges such as low illumination; event cameras improve the temporal awareness [57, 78, 97]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 396 + ], + "type": "text", + "content": "Despite the plausible achievements, many rely on modality-specific blocks designed for individual modalities [88, 93], limiting their adaptability. Recent efforts have focused on achieving architectural unification [69, 91], yet they still necessitate modality-specific fine-tuning, resulting in distinct parameter sets for different modalities. The ideal scenario would involve a large-scale dataset encompassing all possible modal combinations, but current tracking datasets predominantly feature a single modality — depth [66, 94], thermal [35, 36], or event [55] — posing challenges for a unified model with a single parameter set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 397, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 624 + ], + "type": "text", + "content": "Learning with Missing modalities: Recent research has addressed real-world scenarios where models must cope with missing modalities [41, 44]. One common strategy involves estimating missing values by learning joint multimodal representations [31, 75], feasible when complete samples are available during training. However, tracking datasets typically exhibit only one modality at a time, complicating the learning of such joint representations. Other works [19, 39] implicitly learn cross-modal alignment end-to-end using large-scale datasets and deep networks, demanding substantial computational resources. Extending such approaches to tracking is challenging due to limited downstream datasets and real-world applicability constraints [26, 64]. In contrast to existing methods, our approach investigates cross-modal relationships by leveraging edge priors to learn a joint representation that unifies all modalities. Our method does not require the simultaneous occurrence of all modalities, offering a unique perspective on dealing with diverse and individual modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 638, + 366, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 366, + 650 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 366, + 650 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 658, + 422, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 422, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 422, + 670 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In this paper, our primary focus is on multimodal tracking, with the constraint that only one modality is available at a time, as shown in Fig. 2. We define our multimodal" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19157" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 517, + 181 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 517, + 181 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 517, + 181 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 517, + 181 + ], + "type": "image", + "image_path": "9251a0b0f8c112a9e6b5721515276e5d325ad8135ce281de200d47396b7fc90d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 547, + 239 + ], + "type": "text", + "content": "Figure 2. Our proposed framework, termed Un-Track, is composed of a shared embedding, a modal prompting, and a LoRA-finetuned pretrained RGB tracker. The shared embedding learns a joint representation that unifies all modalities (Sec. 3.2). The modal prompting block enhances feature modeling with modal awareness at each scale (Sec. 3.3). To track the target object, we finetune a pretrained foundation model [72] using the LoRA technique (Sec. 3.4). Our model achieves a unified model applicable across different modalities under a single parameter set. During inference, Un-Track seamlessly integrates any image-paired data, thanks to the emergent alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "text", + "content": "tracking dataset as " + }, + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "inline_equation", + "content": "M = \\{M^D, M^T, M^E\\}" + }, + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "inline_equation", + "content": "M^X" + }, + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "text", + "content": " represents the subset dataset with only a single modality " + }, + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 249, + 288, + 357 + ], + "type": "text", + "content": " available. Conventional methods tackling missing modalities often use dummy inputs for the absent pixels [31, 53], simulating complete datasets with all possible combinations of paired datasets. In contrast, our method transforms any auxiliary input into a shared embedding, seamlessly binding all modalities together and creating a complete and paired representation with the master RGB feature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 359, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 455 + ], + "type": "text", + "content": "To mitigate overfitting on sparse downstream multimodal datasets, we adopt a transformer-based RGB tracker with frozen parameters and fine-tune it for multimodal tracking. Leveraging a lightweight outer prompting method, we identify uncertain tokens at each scale and enhance them with cross-modal awareness. Simultaneously, an inner fine-tuning process is implemented using the LoRA technique [21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "content": "During training, our model learns the shared embedding from samples in the mixed dataset " + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "content": ", effectively binding all modalities together. As for inference, our model can accommodate any modal input " + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "content": ", thanks to the emergent alignment. Our trainable parameters only include cross-modal binding, outer prompting parameters, and inner LoRA parameters, ensuring a training-friendly pipeline that can be efficiently employed end-to-end on a single 24G GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 562, + 160, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 160, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 160, + 574 + ], + "type": "text", + "content": "3.2. Shared Embedding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "Explicit Edge Awareness: We observe that, as illustrated in Fig. 2(a), depth data introduce 3D distance information, effectively delineating objects with varying granularity and enhancing the sharpness of 3D boundaries; thermal images generate a scene heat map, highlighting objects based on their temperatures and providing clearer contours; event data capture intensity changes, particularly around an object's outbound region. Notably, a consistent feature emerges across these modalities: the representation of the \"true\" objects' shape, often manifested through edges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Motivated by this observation, our objective is to harness" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 251, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 334 + ], + "type": "text", + "content": "edge embedding to unify all modalities. To achieve this, as shown in Fig. 3, we generate gradient maps from auxiliary modalities by computing differences between neighboring pixels along both the x- and y-axes. Simultaneously, without compromising texture edge, we also generate RGB gradient maps. Subsequently, all gradient maps are integrated with the visual feature, forming the gradient feature " + }, + { + "bbox": [ + 304, + 251, + 545, + 334 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 251, + 545, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 546, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 442 + ], + "type": "text", + "content": "Implicit Low-Rank Reconstruction: While edges present a shared feature across different modalities, exclusively transforming all modalities into edges may risk overlooking modality-specific clues. Therefore, we introduce an implicit learning pipeline to complement this by discovering the shared embedding, guided by the previously generated explicit edge awareness. This combined approach allows for a more effective identification of the shared embedding, leveraging both data-driven learning and edge priors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 442, + 546, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 546, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 546, + 585 + ], + "type": "text", + "content": "In technical terms, we redefine the challenge of learning the shared embedding as a quest for the shared low-rank approximation. Both objectives share the essence of distilling common features across all modalities. However, direct estimation of the low-rank approximation becomes impractical due to the distinct data domains and modal representations. In response, we propose a pragmatic strategy: decomposing the shared low-rank vector into the low-rank of each subset component. This alternative, more manageable and feasible within a single domain, lays the groundwork for approximating the global shared low-rank from these individual low-rank components." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "content": "The overall algorithm can be found in Algorithm 1. Specifically, let " + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "content": " be the input feature with mixed auxiliary modalities, decomposed into subset features " + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "inline_equation", + "content": "D,T,E" + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "content": " representing depth, thermal, and event samples from subset datasets " + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "inline_equation", + "content": "M^D,M^T,M^E" + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "content": ". Their respective " + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "inline_equation", + "content": "k_{th}" + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "content": " low-rank matrices " + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "inline_equation", + "content": "D_{k},T_{k},E_{k}" + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "content": ", are approximated by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 667, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 667, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 339, + 667, + 545, + 681 + ], + "type": "interline_equation", + "content": "D _ {k} = \\sigma_ {d} (D), \\quad T _ {k} = \\sigma_ {t} (T), \\quad E _ {k} = \\sigma_ {e} (E), \\tag {1}", + "image_path": "c2c77ceca392ce2dec3deacd577da815580c8b0468e5d3d472f7f6262d0f6af0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{x}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the modality-specific learning through a simple MLP projecting features from input channel " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " to the low" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19158" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 216, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 216, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 216, + 84 + ], + "type": "text", + "content": "Algorithm 1 Implicit Shared Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "text", + "content": " - Mix Modal Feature, " + }, + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "text", + "content": " - Explicit Gradient Binding. \nOutput: " + }, + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 87, + 284, + 108 + ], + "type": "text", + "content": " - Reconstructed feature" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 109, + 287, + 229 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 109, + 287, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 287, + 129 + ], + "type": "text", + "content": "1. Initialize: Separate " + }, + { + "bbox": [ + 56, + 109, + 287, + 129 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 56, + 109, + 287, + 129 + ], + "type": "text", + "content": " into subsets, each representing a modality-specific feature: " + }, + { + "bbox": [ + 56, + 109, + 287, + 129 + ], + "type": "inline_equation", + "content": "D, T, E" + }, + { + "bbox": [ + 56, + 109, + 287, + 129 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "text", + "content": "2. In-Domain Approximation: Train individual approximators " + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "inline_equation", + "content": "\\sigma_{x}" + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "text", + "content": " for each modality " + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "text", + "content": " to derive domain-specific low-rank matrices " + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "inline_equation", + "content": "D_{k}, T_{k}, E_{k}" + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "text", + "content": ". Idem for explicit low-rank gradient " + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "inline_equation", + "content": "G_{k}" + }, + { + "bbox": [ + 56, + 131, + 287, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "text", + "content": "3. Fuse and Guide: Merge " + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "inline_equation", + "content": "D_{k}, T_{k}, E_{k}" + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "text", + "content": " through a fusion function " + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "inline_equation", + "content": "\\varphi_{1}" + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "text", + "content": ". Incorporate the explicit gradient " + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "inline_equation", + "content": "G_{k}" + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "text", + "content": " using addition, after projecting it into a compatible latent space through " + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "inline_equation", + "content": "\\varphi_{2}" + }, + { + "bbox": [ + 56, + 164, + 287, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "text", + "content": "4. Reconstruct: Train a reconstruction function " + }, + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\Phi_R" + }, + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "text", + "content": " to construct the shared embedding " + }, + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "text", + "content": ", considering the explicit guidance " + }, + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 56, + 197, + 287, + 229 + ], + "type": "text", + "content": " in the original feature space." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 246, + 269, + 369 + ], + "blocks": [ + { + "bbox": [ + 50, + 246, + 269, + 369 + ], + "lines": [ + { + "bbox": [ + 50, + 246, + 269, + 369 + ], + "spans": [ + { + "bbox": [ + 50, + 246, + 269, + 369 + ], + "type": "image", + "image_path": "de2111bf26edd166c00a6777b458a86d3d1ebac711b3b88db515cb9ed22cb92c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 374, + 287, + 384 + ], + "lines": [ + { + "bbox": [ + 50, + 374, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 50, + 374, + 287, + 384 + ], + "type": "text", + "content": "RGB: Image M: Mixed Modalities G: Gradient Feature F: Shared Embedding" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 388, + 287, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 431 + ], + "type": "text", + "content": "Figure 3. Shared Embedding. We derive a joint representation through low-rank factorization and reconstruction. Such an implicit learning is additionally integrated with explicit edge awareness to enhance the embedding." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "text", + "content": "rank space " + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "inline_equation", + "content": "k < c" + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "text", + "content": "). Simultaneously, we compute the low-rank matrix " + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "inline_equation", + "content": "G_{k}" + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "text", + "content": " from the gradient feature " + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 445, + 287, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "text", + "content": "The global shared low-rank matrix " + }, + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "inline_equation", + "content": "M_{k}" + }, + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "text", + "content": " is then approximated by fusing the subset low-rank matrices " + }, + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "inline_equation", + "content": "D_{k}, T_{k}, E_{k}" + }, + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "text", + "content": ", along with the gradient guidance. Technically, we concatenate " + }, + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "inline_equation", + "content": "D_{k}, T_{k}, E_{k}" + }, + { + "bbox": [ + 46, + 470, + 287, + 541 + ], + "type": "text", + "content": " and learn the joint low-rank approximation, incorporating it with the gradient guidance. This pipeline can be expressed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 96, + 554, + 287, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 554, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 96, + 554, + 287, + 568 + ], + "type": "interline_equation", + "content": "M _ {k} = \\varphi_ {R _ {1}} \\left(\\left[ D _ {k}, T _ {k}, E _ {k} \\right]\\right) + \\varphi_ {R _ {2}} (G _ {k}), \\tag {2}", + "image_path": "00be2b63eb5fadf1aab7589a45e82510e0a05469aeb878da74b88e6d1aaf4050.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "inline_equation", + "content": "[.]" + }, + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "text", + "content": " is the channel concatenation and " + }, + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\varphi_{R_i}" + }, + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "text", + "content": " are the MLP projections to the low-rank latent space. Finally, we reconstruct the shared embedding " + }, + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 579, + 287, + 616 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 129, + 628, + 287, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 628, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 129, + 628, + 287, + 641 + ], + "type": "interline_equation", + "content": "F = \\Phi_ {R} (M _ {k}) + G, \\tag {3}", + "image_path": "493a9414d72c2dc866d0d23a16bf055c4b94281e212f9bf551ce5cb65969ff8d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\Phi_R" + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": " is another MLP that projects the jointly-learned low-rank matrix back to the departing feature space. Our ablation studies validate that this subset low-rank approximation and regrouping efficiently unify all input modalities, despite their heterogeneous representations." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 215 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 215 + ], + "type": "image", + "image_path": "97127b3f803f23f10955139fe90f88451f56c935043afdfdbc41bae3793fb447.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "text", + "content": "Figure 4. Modal Prompting. For the visual feature " + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "text", + "content": ", we employ a score function to categorize tokens into negative, uncertain, and positive segments. Using a token exchange policy, we discard negative tokens, enhance uncertain ones with corresponding tokens from " + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "text", + "content": ", and retain positive ones. Then, we transform the feature fusion task into a token recovery problem, addressed by low-rank factorization. Similarly, we extract the most informative low-rank matrix from " + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 220, + 545, + 308 + ], + "type": "text", + "content": " to fuse and reconstruct the shared output." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 323, + 443, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 443, + 336 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 443, + 336 + ], + "type": "text", + "content": "3.3. Outer Modal Prompting" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": "RGB-tracker may fail to perform accurately in corner cases where auxiliary clues can contribute. Drawing inspiration from the success of adapting large pre-trained models to specific downstream tasks [21, 25], we introduce a modal prompting method devised to enhance RGB token " + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": " with modality-awareness " + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": ", as shown in Fig. 4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "text", + "content": "Specifically, our approach employs a shrinkage token fusion strategy. Taking " + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "text", + "content": " as an example, we categorize tokens into three groups—negative, uncertain, and positive—based on a dynamic scoring function " + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "text", + "content": ". These regions are defined with mask form, expressed as " + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "inline_equation", + "content": "m_{n}, m_{u}, m_{p} = s(I)" + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "text", + "content": ". To harness multimodal clues effectively, we replace negative tokens with those from the other modality, omit uncertain ones with dummy values, and retain the positive tokens. Subsequently, these modified tokens undergo projection into the low-rank space using the approximation function " + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\sigma_{c}" + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "text", + "content": ". Our objective is to enhance robustness by completing uncertain tokens with information from reliable neighboring tokens. Mathematically, we obtain the first low-rank matrix " + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "inline_equation", + "content": "I_{l_1}" + }, + { + "bbox": [ + 304, + 413, + 545, + 569 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 375, + 578, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 578, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 375, + 578, + 545, + 591 + ], + "type": "interline_equation", + "content": "I _ {l _ {1}} = \\sigma_ {c} \\left(m _ {n} \\cdot F + m _ {p} \\cdot I\\right). \\tag {4}", + "image_path": "4b571fad1a77fcd37dddf92171d671780bcbd003d465f0d434cbb5219a57ccb4.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 600, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 672 + ], + "type": "text", + "content": "Next, we target the uncertain tokens by merging them with the paired tokens from the other modality and approximate the low-rank matrix similarly. Here, we aim to throw out the possible noise, resulting in a matrix that is more informative than the original. Let " + }, + { + "bbox": [ + 304, + 600, + 545, + 672 + ], + "type": "inline_equation", + "content": "\\sigma_{n}" + }, + { + "bbox": [ + 304, + 600, + 545, + 672 + ], + "type": "text", + "content": " be another approximation function, we obtain the second low-rank matrix " + }, + { + "bbox": [ + 304, + 600, + 545, + 672 + ], + "type": "inline_equation", + "content": "I_{l_n}" + }, + { + "bbox": [ + 304, + 600, + 545, + 672 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 375, + 684, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 684, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 375, + 684, + 545, + 696 + ], + "type": "interline_equation", + "content": "I _ {l _ {2}} = \\sigma_ {n} \\left(m _ {u} \\cdot F + m _ {u} \\cdot I\\right). \\tag {5}", + "image_path": "c7c82ed9b9f2f4ce14860b654ecf7c235f3f75c4e60aa11a7128aabf4e27c5c9.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "content": "Then, we fuse these two low-rank matrices in low-rank" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19159" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 83, + 544, + 137 + ], + "blocks": [ + { + "bbox": [ + 164, + 74, + 234, + 82 + ], + "lines": [ + { + "bbox": [ + 164, + 74, + 234, + 82 + ], + "spans": [ + { + "bbox": [ + 164, + 74, + 234, + 82 + ], + "type": "text", + "content": "Depth-Specific Parameters" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 370, + 74, + 482, + 82 + ], + "lines": [ + { + "bbox": [ + 370, + 74, + 482, + 82 + ], + "spans": [ + { + "bbox": [ + 370, + 74, + 482, + 82 + ], + "type": "text", + "content": "Uni-model with a Single Set of Parameters" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 83, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 49, + 83, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 83, + 544, + 137 + ], + "type": "table", + "html": "
ATCAIS [28]DDiMP [28]DeT [66]SPT [94]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]UniNext [65]SeqTrack [8]ViPT [91]Un-Track (ours)
F-score(↑)0.4760.4850.5320.5380.5780.5940.6120.3970.5150.5690.4220.5900.5610.610
Re(↑)0.4550.4690.5060.5490.5730.5960.6100.4060.5260.5820.4320.6000.5620.610
Pr(↑)0.5000.5030.5600.5270.5830.5920.6130.3880.5050.5570.4130.5800.5600.610
", + "image_path": "da94fd0625f618c03c162d51e4e64b71734ded5ffca2677059b3ec86c52e8a3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 172, + 286, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 172, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 286, + 220 + ], + "type": "text", + "content": "space, forming the shared low-rank matrix " + }, + { + "bbox": [ + 46, + 172, + 286, + 220 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 46, + 172, + 286, + 220 + ], + "type": "text", + "content": " for input " + }, + { + "bbox": [ + 46, + 172, + 286, + 220 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 172, + 286, + 220 + ], + "type": "text", + "content": ". In such a manner, we improve the image token modeling by fully benefiting from the auxiliary clues. Mathematically, the whole process can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 231, + 287, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 231, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 287, + 244 + ], + "type": "interline_equation", + "content": "I _ {l} = \\varphi_ {P} ([ I _ {l _ {1}}, I _ {l _ {2}} ]), \\tag {6}", + "image_path": "284520640fd101376b4dbf0ee236f9cd2d59bdc525ffca214b3ccfe9c6e0d4c9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\varphi_{P}" + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "content": " is learnable fusion. Similarly, from the input " + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "content": ", we follow the same process to obtain the low-rank matrix " + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "content": ". For the cross-modal fusion, we add these two low-rank matrices and then project back to the original space. We can obtain the fused output " + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 324, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 287, + 338 + ], + "type": "interline_equation", + "content": "O = \\Phi_ {P} \\left(I _ {l} + F _ {l}\\right), \\tag {7}", + "image_path": "311cfccaf0e1f9e9524fcb16e52196aa41f8a101214f8981ed7c03640164bb0d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 348, + 195, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 195, + 359 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 195, + 359 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 348, + 195, + 359 + ], + "type": "inline_equation", + "content": "\\Phi_P" + }, + { + "bbox": [ + 47, + 348, + 195, + 359 + ], + "type": "text", + "content": " is another learnable MLP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "content": "Our method can be regarded as a mixer of token exchange (for negative tokens) and token fusion (for uncertain tokens), while retaining the most informative modality-specific clues (for positive tokens). As the majority of fusion operations occur in the low-rank feature space, our progressive cross-modal shrinkage avoids imposing a significant additional computational burden, while being able to excavate and accumulate the rich cues from each modality for effective modal prompting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 477, + 150, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 150, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 150, + 490 + ], + "type": "text", + "content": "3.4. Inner Finetuning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "text", + "content": "In addition to the outer modal prompting, we incorporate the LoRA technique [21] for more efficient finetuning. For each transformer attention module with the weight matrix " + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "text", + "content": ", we introduce two learnable matrices: " + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "text", + "content": ". This leads to the replacement of the frozen attention mechanism " + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "inline_equation", + "content": "h = W_0 x" + }, + { + "bbox": [ + 46, + 496, + 287, + 567 + ], + "type": "text", + "content": " with the new LoRA attention:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 582, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 287, + 594 + ], + "type": "interline_equation", + "content": "h = W _ {0} \\boldsymbol {x} + B A \\boldsymbol {x}. \\tag {8}", + "image_path": "332a435280b6d79ba914343cd3d32d102f78018bba1cab92e80d5659f68e1d24.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 602, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 625 + ], + "type": "text", + "content": "To train our network, we adopt the same loss functions as our baseline tracker [72] for end-to-end learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 638, + 127, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 127, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 127, + 651 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 658, + 135, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 135, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 135, + 670 + ], + "type": "text", + "content": "4.1. Training Data" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In the absence of a comprehensive multi-modal tracking dataset encompassing all possible combinations (RGB-D-T-E), we have only one " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{RGB + X}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " (i.e. RGB+depth, or" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 172, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 243 + ], + "type": "text", + "content": "RGB+thermal, or RGB+event) at a time. The conventional modality-specific adapted settings are referred as \"X-Specific\", whereas the main target of this paper with one model trained on all modality pairs is called \"Uni-model\" (or a single set of parameters). The training and evaluation settings are summarized as follows:" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 307, + 245, + 533, + 288 + ], + "blocks": [ + { + "bbox": [ + 46, + 137, + 544, + 160 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 544, + 160 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 544, + 160 + ], + "type": "text", + "content": "Table 1. Overall performance on DepthTrack test set [66]. Red/Green/Blue stands for the best/second/third performance. Our depth-specific model sets the SOTA record, while our uni-model with single parameters set also outperforms previous depth-specific SOTA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 245, + 533, + 288 + ], + "lines": [ + { + "bbox": [ + 307, + 245, + 533, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 533, + 288 + ], + "type": "table", + "html": "
ModelTrained onEvaluated on
X-SpecificRGBi+XiRGBi+Xi
Uni-model∪iRGBi+XiRGBi+Xi
", + "image_path": "4268f811f0d7a505cec64750def463f96ecdc85f5624c25568c7e412e3dd144b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 292, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 375 + ], + "type": "text", + "content": "Our RGB-D samples are sourced from DepthTrack [66], a pioneering RGB-D tracking benchmark with 150 training long-term sequences. RGB-T samples are extracted from the extensive LasHeR [36] dataset, featuring 979 diverse training sequences. RGB-E samples are obtained from VisEvent [55], which boasts 500 real-world sequences. Each D/T/E input is transformed into an RGB-like form." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 384, + 473, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 473, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 473, + 395 + ], + "type": "text", + "content": "4.2. Within distribution Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 402, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 461 + ], + "type": "text", + "content": "Given that DepthTrack [66], Lasher [36], and VisEvent [55] provide domain-specific testing sequences, our initial evaluation focuses on these within distribution sequences. For each dataset, we adhere to the metrics specified in the original papers and prior standards [69, 91] for evaluation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 462, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 593 + ], + "type": "text", + "content": "Comparison on DepthTrack [66]: For evaluation, we use metrics such as precision (Pr) and recall (Re), as well as the F-score, which are the primary metrics. As shown in Tab. 1 When exclusively trained and fine-tuned on DepthTrack, Un-Track achieves a " + }, + { + "bbox": [ + 304, + 462, + 545, + 593 + ], + "type": "inline_equation", + "content": "+2.1\\%" + }, + { + "bbox": [ + 304, + 462, + 545, + 593 + ], + "type": "text", + "content": " absolute precision improvement over the current depth-specific SOTA ViPT [91]. Notably, even when trained with mixed data using a single parameter set, Un-Track outperforms the depth-specific ViPT. Furthermore, when the current ViPT is jointly trained on all datasets with a single set of parameters, its performance significantly deteriorates." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Additionally, we observe that trackers excelling in other tracking datasets [14, 22, 42] might struggle in RGB-D downstream settings. UniNext [65], a leading tracker trained on various large-scale tracking datasets and related image/video tasks, exhibits poor performance. In contrast, our model achieves cross-modal unification within a single set of parameters, surpassing all depth-specific counterparts and performing closely to our specialized version. This underscores the efficacy of our shared embedding in achieving global alignment across diverse modalities." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19160" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 84, + 544, + 136 + ], + "blocks": [ + { + "bbox": [ + 166, + 74, + 252, + 83 + ], + "lines": [ + { + "bbox": [ + 166, + 74, + 252, + 83 + ], + "spans": [ + { + "bbox": [ + 166, + 74, + 252, + 83 + ], + "type": "text", + "content": "Thermal-Specific Parameters" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 378, + 74, + 504, + 83 + ], + "lines": [ + { + "bbox": [ + 378, + 74, + 504, + 83 + ], + "spans": [ + { + "bbox": [ + 378, + 74, + 504, + 83 + ], + "type": "text", + "content": "Uni-model with a Single Set of Parameters" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 84, + 544, + 136 + ], + "lines": [ + { + "bbox": [ + 50, + 84, + 544, + 136 + ], + "spans": [ + { + "bbox": [ + 50, + 84, + 544, + 136 + ], + "type": "table", + "html": "
SGT [34]FANet [96]mfDiMP [81]DAFNet [18]MaCNet [77]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
PR(↑)0.3270.4410.4470.4480.4820.5380.6510.6670.4180.4630.5300.5820.6080.646
SR(↑)0.2320.3090.3430.3110.3500.4200.5250.5360.3330.3650.4220.4410.4900.513
", + "image_path": "f425a4c874b72aa69f9efe5f746961c03f4cd6ecdda6ed4b6e3f66d141816375.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 182, + 544, + 234 + ], + "blocks": [ + { + "bbox": [ + 47, + 137, + 546, + 160 + ], + "lines": [ + { + "bbox": [ + 47, + 137, + 546, + 160 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 546, + 160 + ], + "type": "text", + "content": "Table 2. Overall performance on the Lasher thermal testing set [36]. Our thermal-specific model sets a new SOTA record. Our uni-model variant competes strongly with the previous SOTA thermal-specific models and significantly surpasses its unimodel version." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 172, + 265, + 181 + ], + "lines": [ + { + "bbox": [ + 186, + 172, + 265, + 181 + ], + "spans": [ + { + "bbox": [ + 186, + 172, + 265, + 181 + ], + "type": "text", + "content": "Event-Specific Parameters" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 386, + 172, + 514, + 181 + ], + "lines": [ + { + "bbox": [ + 386, + 172, + 514, + 181 + ], + "spans": [ + { + "bbox": [ + 386, + 172, + 514, + 181 + ], + "type": "text", + "content": "Uni-model with a Single Set of Parameters" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 182, + 544, + 234 + ], + "lines": [ + { + "bbox": [ + 50, + 182, + 544, + 234 + ], + "spans": [ + { + "bbox": [ + 50, + 182, + 544, + 234 + ], + "type": "table", + "html": "
SiamCar [62]Stark_E [62]VITAL_E [46]PrDiMP_E [13]TransT_E [5]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]ViPT [91]Un-Track (ours)
Precision(↑)0.5990.6120.6490.6440.6500.6320.7580.7630.6160.6260.6910.6650.7400.755
Success(↑)0.4200.4460.4150.4530.4740.4710.5920.5970.4480.4440.5250.5040.5790.589
", + "image_path": "faa06b262871b53165fe5ba4a8262fce5334d38de755e5085ff4f89728e2a8b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 236, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 269 + ], + "type": "text", + "content": "Table 3. Overall performance on VisEvent dataset [55]. Our event-specific model sets a new SOTA record. Our uni-variant, with the same parameter set as in Depth and Thermal, consistently achieves competitive performance across various modalities, leading to a significant margin over the uni-variant of the SOTA modality-specific model [91]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 284, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 475 + ], + "type": "text", + "content": "Comparison on LasHer [36]: Similarly, we conduct evaluations on the LasHer testing set for RGB-T tracker assessment as shown in Tab. 2. Precision (PR) and success rates (SR) are reported following conventional standards [36, 69, 91]. Initial comparisons under domain-specific settings reveal the challenges of achieving an architecture-unified model across RGB-D and RGB-T domains, with ProTrack [69] and ViPT [91] being the only works consistently leading in both settings. Our Un-Track, following domain-specific finetuning, outperforms the leading ViPT by a significant margin and sets a new SOTA record. In the cross-domain joint learning with a single set of parameters, our uni-model achieves a " + }, + { + "bbox": [ + 46, + 284, + 289, + 475 + ], + "type": "inline_equation", + "content": "+3.8\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 475 + ], + "type": "text", + "content": " absolute gain over ViPT. Remarkably, our model with a single set of parameters already achieves very competitive performance compared to the thermal-specific ViPT version." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "Comparison on VisEvent [55]: We also evaluate tracker performance with RGB-Event input. Event data, being inherently sparse compared to depth and thermal information, presents challenges in extending existing RGB-D or RGB-T fusion designs for effective integration, hence necessitating specific fusion designs [3, 43, 98]. In contrast, we adopt a unified cross-modal prompting method based on shrinkage fusion. Our approach, with gradual token exchanges between RGB and event modalities, effectively preserves crucial modality-specific clues to enhance feature modeling. Performance-wise, under the event-specific setting, our Un-Track outperforms all counterparts, as shown in Tab. 3 and in Fig. 5. In the single set of parameters setting, our uni-model achieves a " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " absolute gain in precision over the current SOTA. This underscores the effectiveness of our approach in handling the unique challenges posed by event data integration, which can be mainly contributed to our shared binding that learns the global RGB+X alignment." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 281, + 425, + 400 + ], + "blocks": [ + { + "bbox": [ + 307, + 281, + 425, + 400 + ], + "lines": [ + { + "bbox": [ + 307, + 281, + 425, + 400 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 425, + 400 + ], + "type": "image", + "image_path": "29f2be60cfcc7d917d98781312dfdc23fd658d477dd5208a20bec076dbaadb10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 401, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 306, + 401, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 545, + 434 + ], + "type": "text", + "content": "Figure 5. More precision/success comparisons on VisEvent dataset [55]. \"Uni\" stands for models with a single parameter set. \"E\" stands for the extension of RGB trackers with event fusion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 281, + 539, + 400 + ], + "blocks": [ + { + "bbox": [ + 425, + 281, + 539, + 400 + ], + "lines": [ + { + "bbox": [ + 425, + 281, + 539, + 400 + ], + "spans": [ + { + "bbox": [ + 425, + 281, + 539, + 400 + ], + "type": "image", + "image_path": "3d89a35b2ccda4e1346510bdcf6a832c3983146410eb10179321964ca153ee48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 453, + 474, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 474, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 474, + 464 + ], + "type": "text", + "content": "4.3. Generalization Across Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 519 + ], + "type": "text", + "content": "In this section, we assess the versatility by evaluating performance on datasets that differ from the training ones, aligning with the goal of achieving a universal model checkpoint applicable to diverse scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": "VOT-RGBD2022 [30]: We first perform inference on the VOT-RGBD2022 dataset. Notably, our uni-model achieves superior performance compared to the depth-specific ViPT with a " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "+0.5\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": " absolute gain in accuracy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "content": "RGBT234 [35]: We also test our model on the other thermal dataset RGBT234, which encompasses sequences with different distributions. Our uni-model surpasses the thermal-specific ViPT with a notable " + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "content": " absolute precision gain, as shown in Tab. 5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": "Moreover, we present a detailed per-attribute comparison with the fine-tuned ViPT in Fig. 6. We are particularly interested in sequences with motion blur-fast motion-camera moving, as well as sequences with heavy occlusion-scale variants-background clutter. The former motion-related challenges can typically benefit from event clues, renowned for asynchronous computing, while the latter geometry" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19161" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 84, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 194, + 74, + 271, + 83 + ], + "lines": [ + { + "bbox": [ + 194, + 74, + 271, + 83 + ], + "spans": [ + { + "bbox": [ + 194, + 74, + 271, + 83 + ], + "type": "text", + "content": "Depth-Specific Parameters" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 391, + 74, + 514, + 83 + ], + "lines": [ + { + "bbox": [ + 391, + 74, + 514, + 83 + ], + "spans": [ + { + "bbox": [ + 391, + 74, + 514, + 83 + ], + "type": "text", + "content": "Uni-model with a Single Set of Parameters" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 84, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 50, + 84, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 50, + 84, + 545, + 144 + ], + "type": "table", + "html": "
DRefine [29]Stark_D [29]DMTracker [30]DeT [66]SBT_D [30]SPT [94]ProTrack [69]ViPT [91]Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
EAO(↑)0.5920.6470.6580.6570.7080.6510.6510.7210.4450.6410.6660.6790.718
Accuracy(↑)0.7750.8030.7580.7600.8090.7980.8010.8150.7140.7690.8080.8020.820
Robustness(↑)0.7600.7980.8510.8450.8640.8510.8020.8710.5980.8320.8140.8460.864
", + "image_path": "3e110bd08cca8f49878bd634c8c900a314e44c2cd9867fdd9ad31525f1f7dcdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 173, + 545, + 226 + ], + "blocks": [ + { + "bbox": [ + 47, + 144, + 546, + 168 + ], + "lines": [ + { + "bbox": [ + 47, + 144, + 546, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 546, + 168 + ], + "type": "text", + "content": "Table 4. Overall performance on VOT-RGBD2022 [30]. Our uni-model, trained on a mix of all modalities, shows robust generalization and outperforms all depth-specific models and other uni-model counterparts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 173, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 50, + 173, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 50, + 173, + 545, + 226 + ], + "type": "table", + "html": "
Thermal-Specific ParametersUni-model with a Single Set of Parameters
mfDiMP [81]SGT [34]DAFNet [18]FANet [96]MaCNet [77]CMPP [52]APFNet [59]ProTrack [69]ViPT [91]Un-Track (ours)Stark [62]AiATrack [17]OSTrack [72]SeqTrack [8]Un-Track (ours)
MPR(↑)0.6460.7200.7960.7870.7900.8230.8270.7950.8350.8370.6770.7110.7550.8060.842
MSR(↑)0.4280.4720.5440.5530.5540.5750.5790.5990.6170.6180.4960.5080.5690.5990.625
", + "image_path": "b67334091280628925072b216e457eac80cee8b83f512d3e12f7f9dba8423e9f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 253, + 285, + 400 + ], + "blocks": [ + { + "bbox": [ + 50, + 253, + 285, + 400 + ], + "lines": [ + { + "bbox": [ + 50, + 253, + 285, + 400 + ], + "spans": [ + { + "bbox": [ + 50, + 253, + 285, + 400 + ], + "type": "image", + "image_path": "98c0d70b043dfc08881f66c6ef0e3a941242ab9ff552177207e05de66533baea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 404, + 288, + 482 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 482 + ], + "type": "text", + "content": "Figure 6. Per-attribute analysis on the thermal dataset RGBT234 [35]. Challenges related to motion and geometry are generally better addressed by event and depth cameras, respectively. Nevertheless, when inferring only with RGB-T data, our Un-Track surpasses both the SOTA thermal-specific method and the current leading uni-tracker. This success underscores our ability to learn emergent alignment across diverse modalities." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 592 + ], + "type": "text", + "content": "related challenges can typically benefit from depth cameras. However, these event/depth clues are not directly available in the RGB-T setting. Nevertheless, our Un-Track, trained on all modalities, outperforms both thermal fine-tuned ViPT [91] and the current leading uni-tracker [8] with large margins. This underscores our capability in learning event and depth priors through shared binding, without the need for the presence of these modalities during inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "RGB-only: In practical scenarios, challenges arise when there are no modal clues available, a typical case is when the auxiliary sensor fails to work properly. We address this demanding case in our study by substituting the modal input with dummy values. Under a such challenging setting, as shown in Tab. 6, our uni-method consistently outperforms both our RGB baseline fine-tuned counterparts significantly. Notably, such an improvement is achieved with a very limited increase in learning parameters, i.e., " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "+6.65\\mathrm{M}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "+2.14" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " GFLOPs." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 251, + 544, + 338 + ], + "blocks": [ + { + "bbox": [ + 56, + 227, + 535, + 239 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 535, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 535, + 239 + ], + "type": "text", + "content": "Table 5. Overall performance on RGBT234 dataset [35]. Our uni-model sets new SOTA records without specific thermal finetuning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 251, + 544, + 338 + ], + "lines": [ + { + "bbox": [ + 308, + 251, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 544, + 338 + ], + "type": "table", + "html": "
RGB Baseline [72]ViPT [91]Un-Track (ours)
GFLOPs21.5021.8023.64
Params (M)92.0892.9698.73
F-score(↑)0.5290.5420.558
Re(↑)0.5220.5380.557
Pr(↑)0.5360.5460.560
", + "image_path": "b5321f98ba82d82641771286325a20cdd714dc96a545a43c7742e366e0150fcd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 365, + 544, + 427 + ], + "blocks": [ + { + "bbox": [ + 306, + 338, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 306, + 338, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 545, + 361 + ], + "type": "text", + "content": "Table 6. Overall performance on DepthTrack test set [66] with dummy depth input." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 365, + 544, + 427 + ], + "lines": [ + { + "bbox": [ + 308, + 365, + 544, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 544, + 427 + ], + "type": "table", + "html": "
w/o Shared Embed[91] as Modal Promptw/o LoRA Finetune
F-score(↑)0.5990.5790.594
Re(↑)0.6020.5750.598
Pr(↑)0.5970.5840.596
", + "image_path": "a207d800b14a42e71a9d560f7e02b2bc720917342eafa65fc09cf1d23c8208b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 363, + 428, + 488, + 439 + ], + "lines": [ + { + "bbox": [ + 363, + 428, + 488, + 439 + ], + "spans": [ + { + "bbox": [ + 363, + 428, + 488, + 439 + ], + "type": "text", + "content": "Table 7. Key component analysis." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 452, + 406, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 406, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 406, + 464 + ], + "type": "text", + "content": "5. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 472, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 665 + ], + "type": "text", + "content": "In this section, we perform all experiments on the DepthTrack testing set [66] under a single parameter set setting. Key Component Analysis: We begin by studying the effectiveness of key components, including the shared embedding, modal prompting, and LoRA finetuning, as summarized in Tab. 7. We initially remove the shared embedding by directly feeding mixed modalities into the learning diagram. It can be seen that this equal treatment of all modalities harms network performance due to the heterogeneous representation across modal domains. Secondly, we replace our prompting block with a recent counterpart that computes fovea attention from additional input [91]. Our designed gradual shrinkage fusion, allowing token-wise interaction, proves to be more effective. We also report performance when we remove the inner fine-tuning with Lora. It can be seen that the performance deteriorates." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Low-Rank: Low-rank approximation plays a vital role in our model, influencing our shared embedding, modal fusion, and LoRA-finetuning. Hence, the choice of rank is crucial for both the efficiency and effectiveness of our ap" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19162" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 153, + 70, + 439, + 83 + ], + "lines": [ + { + "bbox": [ + 153, + 70, + 439, + 83 + ], + "spans": [ + { + "bbox": [ + 153, + 70, + 439, + 83 + ], + "type": "text", + "content": "Table 8. Low Rank Approximation Our plain version is highlighted in gray." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 53, + 99, + 202, + 159 + ], + "blocks": [ + { + "bbox": [ + 92, + 87, + 161, + 97 + ], + "lines": [ + { + "bbox": [ + 92, + 87, + 161, + 97 + ], + "spans": [ + { + "bbox": [ + 92, + 87, + 161, + 97 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 92, + 87, + 161, + 97 + ], + "type": "inline_equation", + "content": "\\text{Rank}_k(\\text{Sec. 3.2})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 99, + 202, + 159 + ], + "lines": [ + { + "bbox": [ + 53, + 99, + 202, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 99, + 202, + 159 + ], + "type": "table", + "html": "
248
F-score(↑)0.6070.6100.602
Re(↑)0.6060.6080.601
Pr(↑)0.6080.6110.604
", + "image_path": "aa9164f814325db0da629b7328c678ba15a8702b456c9402efeb6cd9dfd5ac69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 211, + 99, + 310, + 159 + ], + "blocks": [ + { + "bbox": [ + 225, + 87, + 293, + 97 + ], + "lines": [ + { + "bbox": [ + 225, + 87, + 293, + 97 + ], + "spans": [ + { + "bbox": [ + 225, + 87, + 293, + 97 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 225, + 87, + 293, + 97 + ], + "type": "inline_equation", + "content": "\\text{Rank}_l" + }, + { + "bbox": [ + 225, + 87, + 293, + 97 + ], + "type": "text", + "content": " (Sec. 3.3)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 211, + 99, + 310, + 159 + ], + "lines": [ + { + "bbox": [ + 211, + 99, + 310, + 159 + ], + "spans": [ + { + "bbox": [ + 211, + 99, + 310, + 159 + ], + "type": "table", + "html": "
4816
0.5960.6100.606
0.5930.6080.609
0.5990.6110.604
", + "image_path": "6be6d9835993e1435d69451f6279fb2ce79e6f7ce0535fe2037d6db17be3e1d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 318, + 99, + 430, + 159 + ], + "blocks": [ + { + "bbox": [ + 338, + 87, + 406, + 97 + ], + "lines": [ + { + "bbox": [ + 338, + 87, + 406, + 97 + ], + "spans": [ + { + "bbox": [ + 338, + 87, + 406, + 97 + ], + "type": "text", + "content": "(c) LoRA (Sec. 3.4)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 99, + 430, + 159 + ], + "lines": [ + { + "bbox": [ + 318, + 99, + 430, + 159 + ], + "spans": [ + { + "bbox": [ + 318, + 99, + 430, + 159 + ], + "type": "table", + "html": "
248
0.6010.6100.600
0.5990.6080.598
0.6020.6110.602
", + "image_path": "8682634c645e4a1cd0cbd2f9a02e34dc53b70515923ec0891bd517d6ea97ad36.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 438, + 99, + 539, + 159 + ], + "blocks": [ + { + "bbox": [ + 444, + 87, + 526, + 97 + ], + "lines": [ + { + "bbox": [ + 444, + 87, + 526, + 97 + ], + "spans": [ + { + "bbox": [ + 444, + 87, + 526, + 97 + ], + "type": "text", + "content": "(d) Percentile (Sec. 3.3)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 438, + 99, + 539, + 159 + ], + "lines": [ + { + "bbox": [ + 438, + 99, + 539, + 159 + ], + "spans": [ + { + "bbox": [ + 438, + 99, + 539, + 159 + ], + "type": "table", + "html": "
1/81/41/3
0.6040.6100.595
0.6060.6080.593
0.6020.6110.596
", + "image_path": "df4e4b713fdbbac621291e722557f5ad088c5d38a5c2d748d8e97e7208e61cad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 175, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 198 + ], + "type": "text", + "content": "proach. In Tab. 8, we systematically explore the impact of low-rank choices within each component." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 199, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 282 + ], + "type": "text", + "content": "Shared Embedding: Our objective is to identify an optimal low-rank latent space for merging different modalities effectively. Tab. 8(a) presents our ablation study, where we explore ranks of 2, 4, and 8. Lower ranks result in poorer performance due to sparse representations. Conversely, higher ranks capture too many modality-specific details, complicating the search for a shared embedding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "content": "Modal Prompting: As shown in Tab. 8(b), similar trends are observed when investigating the low-rank choices for modal prompting with rank values of 4, 8, and 16, as low ranks struggle to capture essential information, while higher ranks introduce an overload of modality-specific details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 342, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 402 + ], + "type": "text", + "content": "LoRA-finetuning: We also vary the ranks between 2, 4, and 8 for the LoRA finetuning technique, as shown in Tab. 8(c). Lower ranks exhibited consistently poor performance, while higher ranks in this case tend to result in poorer performance, likely due to overfitting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 403, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 485 + ], + "type": "text", + "content": "Remarks: These experiments emphasize the importance of selecting the best low-rank representations. Nevertheless, our model shows great resilience to the choice of LoRA, showcasing consistent performance across different low-rank configurations. Notably, all our low-rank variants outperform the current SOTA ViPT under the uni-setting, validating our robustness and effectiveness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "type": "text", + "content": "Modal Prompting: During the prompting fusion, a learnable score function is used to categorize tokens into three groups based on their confidence scores. Here, we explore different percentiles for the number of positive, which is the same as the number of negative tokens, leaving the rest as uncertain tokens. As shown in Tab. 8(d), the choice of percentile can influence overall performance. Lower percentiles result in poorer performance, as recovering uncertain tokens from very few neighboring tokens is challenging. Higher percentiles also lead to performance degradation since the focus is on token exchange rather than token fusion. The choice of " + }, + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "type": "text", + "content": " leads to the best balance between exchanger and fuser, leading to the best performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Shared Embedding: Here, we perform ablation studies on our shared embedding, a foundational component of our uni-model. The quantitative results are presented in Tab. 9. We begin by exploring a scenario where our shared embedding is replaced with a variant lacking explicit edge guidance (w/o Explicit Edge). In this configuration, the network" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 309, + 173, + 544, + 240 + ], + "blocks": [ + { + "bbox": [ + 309, + 173, + 544, + 240 + ], + "lines": [ + { + "bbox": [ + 309, + 173, + 544, + 240 + ], + "spans": [ + { + "bbox": [ + 309, + 173, + 544, + 240 + ], + "type": "table", + "html": "
w/o Explicit Edgew/o Implicit Learningw/o In-domain Approx.
F-score(↑)0.6000.6040.581
Re(↑)0.6020.6090.583
Pr(↑)0.5980.5990.579
", + "image_path": "ef2d7dae8f8a3448ddee9089031402f23a0e4a87ac1fb759ffd1449d19eee6f1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 241, + 498, + 251 + ], + "lines": [ + { + "bbox": [ + 353, + 241, + 498, + 251 + ], + "spans": [ + { + "bbox": [ + 353, + 241, + 498, + 251 + ], + "type": "text", + "content": "Table 9. Ablation on shared embedding" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 319 + ], + "type": "text", + "content": "learns the shared embedding solely without any edge prior. The results highlight a substantial performance drop, underscoring the pivotal role of explicit edge guidance — a natural and static embedding that binds all modalities together — in facilitating this implicit learning process." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 320, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 380 + ], + "type": "text", + "content": "We also conduct experiments using only the explicit edge as the shared embedding, excluding any additional learning modules (w/o Implicit Learning). This approach, too, yields suboptimal performance, primarily due to the neglect of modality-specific clues within each domain." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 380, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 464 + ], + "type": "text", + "content": "Finally, we directly compute the low-rank approximation from mixed modalities, bypassing the initial in-domain approximation and subsequent fusion steps (w/o In-domain Approx.). We observe that the network struggles in learning a shared embedding in this direct low-rank approximation setup, primarily due to the intricately mixed representation caused by the domain gap between different modalities." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 476, + 470, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 470, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 470, + 489 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": "We present a successful case of a single-model and any-modality tracker for video object tracking. The proposed method achieves a shared embedding that binds all modalities together, overcoming their heterogeneous representations. This unification is facilitated by lightweight modal prompting and inner finetuning, inheriting benefits from large-scale pre-trained trackers without introducing a substantial computational burden. Exhaustive experiments showcase our improved tracking performance and robust generalization, with any modality input." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. The authors thank the reviewers and ACs for their tremendous efforts and helpful comments. The event icon is credited to Zuowen Wang. This research is financed in part by the Alexander von Humboldt Foundation, in part by NSFC (62376156, 62322113), and in part by the Ministry of Education and Science of Bulgaria (support for INSAIT, part of the Bulgarian National Roadmap for Research Infrastructure)." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19163" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Simon Arridge, Pascal Fernsel, and Andreas Hauptmann. Joint reconstruction and low-rank decomposition for dynamic inverse problems. Inverse Problems and Imaging, 16 (3):483-523, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 178 + ], + "type": "text", + "content": "[2] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 286, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 286, + 221 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 286, + 221 + ], + "type": "text", + "content": "[3] Qinyu Chen, Zuowen Wang, Shih-Chii Liu, and Chang Gao. 3et: Efficient event-based eye tracking using a change-based convlstm network. arXiv preprint arXiv:2308.11771, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 286, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 286, + 256 + ], + "type": "text", + "content": "[4] Wanli Chen, Xinge Zhu, Ruoqi Sun, Junjun He, Ruiyu Li, Xiaoyong Shen, and Bei Yu. Tensor low-rank reconstruction for semantic segmentation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 286, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 289 + ], + "type": "text", + "content": "[5] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 292, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 312 + ], + "type": "text", + "content": "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 345 + ], + "type": "text", + "content": "[7] Xin Chen, Bin Yan, Jiawen Zhu, Huchuan Lu, Xiang Ruan, and Dong Wang. High-performance transformer tracking. TPAMI, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "type": "text", + "content": "[8] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "type": "text", + "content": "[9] Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, and Rongrong Ji. Siamese box adaptive network for visual tracking. In CVPR, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 414, + 286, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 414, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 414, + 286, + 466 + ], + "type": "text", + "content": "[10] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In CVPR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 286, + 501 + ], + "type": "text", + "content": "[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 502, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 534 + ], + "type": "text", + "content": "[12] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. ATOM: Accurate tracking by overlap maximization. In CVPR, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "type": "text", + "content": "[13] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "text", + "content": "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 603, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 635 + ], + "type": "text", + "content": "[15] Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking. TIP, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 636, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 677 + ], + "type": "text", + "content": "[16] Shang Gao, Jinyu Yang, Zhe Li, Feng Zheng, Alex Leonardis, and Jingkuan Song. Learning dual-fused modality-aware representations for rgbd tracking. In ECCV, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[17] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In ECCV, 2022. 5, 6, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Yuan Gao, Chenglong Li, Yabin Zhu, Jin Tang, Tao He, and Futian Wang. Deep adaptive fusion network for high performance RGBT tracking. In ICCVW, 2019. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In IROS, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "text", + "content": "[22] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "type": "text", + "content": "[23] Sajid Javed, Martin Danelljan, Fahad Shahbaz Khan, Muhammad Haris Khan, Michael Felsberg, and Jiri Matas. Visual object tracking with discriminative filters and siamese networks: a survey and outlook. TPAMI, 45(5):6552-6574, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 332, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 364 + ], + "type": "text", + "content": "[24] I-Hong Jhuo, Dong Liu, DT Lee, and Shih-Fu Chang. Robust visual domain adaptation with low-rank reconstruction. In CVPR, 2012. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 366, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 398 + ], + "type": "text", + "content": "[25] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 400, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 441 + ], + "type": "text", + "content": "[26] Ben Kang, Xin Chen, Dong Wang, Houwen Peng, and Huchuan Lu. Exploring lightweight hierarchical vision transformers for efficient visual tracking. In ICCV, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "text", + "content": "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "text", + "content": "[28] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, Ondrej Drbohlav, et al. The eighth visual object tracking vot2020 challenge results. In ECCVW, 2020. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 535, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 588 + ], + "type": "text", + "content": "[29] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In ICCVW, 2021. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "text", + "content": "[30] Matej Kristan, Aleš Leonardis, Jiri Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, et al. The tenth visual object tracking vot2022 challenge results. In ECCVW, 2023. 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[31] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[32] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19164" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[33] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[34] Chenglong Li, Nan Zhao, Yijuan Lu, Chengli Zhu, and Jin Tang. Weighted sparse representation regularized graph learning for RGB-T object tracking. In ACM MM, 2017. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "text", + "content": "[35] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. RGB-T object tracking: Benchmark and baseline. PR, 96:106977, 2019. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "text", + "content": "[36] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for RGBT tracking. TIP, 31:392-404, 2021. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 230, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 287, + 263 + ], + "type": "text", + "content": "[37] Qiao Liu, Xin Li, Zhenyu He, Nana Fan, Di Yuan, and Hong-peng Wang. Learning deep multi-level similarity for thermal infrared object tracking. TMM, 23:2114-2126, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "text", + "content": "[38] Cheng Long Li, Andong Lu, Ai Hua Zheng, Zhengzheng Tu, and Jin Tang. Multi-adapter rgbt tracking. In ICCVW, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "type": "text", + "content": "[39] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 387 + ], + "type": "text", + "content": "[40] Alan Lukezic, Ugur Kart, Jani Kapyla, Ahmed Durmush, Joni-Kristian Kamarainen, Jiri Matas, and Matej Kristan. Cdtb: A color and depth visual object tracking dataset and benchmark. In ICCV, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 388, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 422 + ], + "type": "text", + "content": "[41] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 287, + 465 + ], + "type": "text", + "content": "[42] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. TrackingNet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 499 + ], + "type": "text", + "content": "[43] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. In ICCV, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "text", + "content": "[44] Yansheng Qiu, Ziyuan Zhao, Hongdou Yao, Delin Chen, and Zheng Wang. Modal-aware visual prompting for incomplete multi-modal brain tumor segmentation. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 287, + 589 + ], + "type": "text", + "content": "[45] Esteban Real, Jonathon Shlens, Stefano Mazzocchi, Xin Pan, and Vincent Vanhoucke. Youtube-boundingboxes: A large high-precision human-annotated data set for object detection in video. In CVPR, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 634 + ], + "type": "text", + "content": "[46] Yibing Song, Chao Ma, Xiaohe Wu, Lijun Gong, Linchao Bao, Wangmeng Zuo, Chunhua Shen, Rynson WH Lau, and Ming-Hsuan Yang. Vital: Visual tracking via adversarial learning. In CVPR, 2018. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 287, + 668 + ], + "type": "text", + "content": "[47] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In CVPR, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "text", + "content": "[48] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[49] Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric. arXiv preprint arXiv:2211.11010, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 149 + ], + "type": "text", + "content": "[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: Visual tracking by redetection. In CVPR, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 151, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 193 + ], + "type": "text", + "content": "[51] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. Rpeflow: Multimodal fusion of rgb-pointcloud-event for joint optical flow and scene flow estimation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 194, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 228 + ], + "type": "text", + "content": "[52] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for RGB-T tracking. In CVPR, 2020. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[53] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In CVPR, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 271, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 271, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 545, + 304 + ], + "type": "text", + "content": "[54] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[55] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. arXiv preprint arXiv:2108.05015, 2021. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[56] Xiao Wang, Xiujun Shu, Shilliang Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Mfgnet: Dynamic modality-aware filter generation for rgb-t tracking. TMM, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 393, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 425 + ], + "type": "text", + "content": "[57] Zuowen Wang, Yuhuang Hu, and Shih-Chii Liu. Exploiting spatial sparsity for event cameras with visual transformers. In ICIP, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[58] David Wisth, Marco Camurri, Sandipan Das, and Maurice Fallon. Unified multi-modal landmark tracking for tightly coupled lidar-visual-inertial odometry. RA-L, 6(2):1004-1011, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "text", + "content": "[59] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for RGBT tracking. In AAAI, 2022. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[60] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for rgbt tracking. In AAAI, 2022. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "text", + "content": "[61] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. SiamFC++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "text", + "content": "[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 635 + ], + "type": "text", + "content": "[63] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "text", + "content": "[64] Bin Yan, Houwen Peng, Kan Wu, Dong Wang, Jianlong Fu, and Huchuan Lu. Lighttrack: Finding lightweight neural networks for object tracking via one-shot architecture search. In CVPR, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[65] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In CVPR, 2023. 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19165" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[66] Song Yan, Jinyu Yang, Jani Käpylä, Feng Zheng, Ales Leonardis, and Joni-Kristian Kämäräinen. Depthtrack: Unveiling the power of RGBD tracking. In ICCV, 2021. 2, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 140 + ], + "type": "text", + "content": "[67] Song Yan, Jinyu Yang, Ales Leonardis, and Joni-Kristian Kamarainen. Depth-only object tracking. In BMVC, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "type": "text", + "content": "[68] Jinyu Yang, Zhe Li, Song Yan, Feng Zheng, Aleš Leonardis, Joni-Kristian Kāmäräinen, and Ling Shao. Rgbd object tracking: An in-depth review. arXiv preprint arXiv:2203.14134, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 288, + 217 + ], + "type": "text", + "content": "[69] Jinyu Yang, Zhe Li, Feng Zheng, Ales Leonardis, and Jingkuan Song. Prompting for multi-modal tracking. In ACMM, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 288, + 251 + ], + "type": "text", + "content": "[70] Jinyu Yang, Shang Gao, Zhe Li, Feng Zheng, and Aleš Leonardis. Resource-efficient rgbd aerial tracking. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 253, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 288, + 285 + ], + "type": "text", + "content": "[71] Rui Yao, Guosheng Lin, Shixiong Xia, Jiaqi Zhao, and Yong Zhou. Video object segmentation and tracking: A survey. ACM TIST, 11(4):1-47, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 287, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 288, + 330 + ], + "type": "text", + "content": "[72] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In ECCV, 2022. 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 332, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 288, + 364 + ], + "type": "text", + "content": "[73] Jie Yin, Ang Li, Tao Li, Wenxian Yu, and Danping Zou. M2dgr: A multi-sensor and multi-scenario slam dataset for ground robots. RA-L, 7(2):2266-2273, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 365, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 288, + 398 + ], + "type": "text", + "content": "[74] Yuechen Yu, Yilei Xiong, Weilin Huang, and Matthew R Scott. Deformable siamese attention networks for visual object tracking. In CVPR, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 399, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 288, + 431 + ], + "type": "text", + "content": "[75] Jiandian Zeng, Tianyi Liu, and Jiantao Zhou. Tag-assisted multimodal sentiment analysis under uncertain missing modalities. In ACM SIGIR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 434, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 288, + 476 + ], + "type": "text", + "content": "[76] Chunhui Zhang, Xin Sun, Yiqian Yang, Li Liu, Qiong Liu, Xi Zhou, and Yanfeng Wang. All in one: Exploring unified vision-language tracking with multi-modal alignment. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 478, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 288, + 520 + ], + "type": "text", + "content": "[77] Hui Zhang, Lei Zhang, Li Zhuo, and Jing Zhang. Object tracking in RGB-T videos using modal-aware attention network and competitive learning. Sensors, 20(2):393, 2020. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 523, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 288, + 555 + ], + "type": "text", + "content": "[78] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Bao-cai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 557, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 288, + 590 + ], + "type": "text", + "content": "[79] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 591, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 591, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 591, + 288, + 634 + ], + "type": "text", + "content": "[80] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 636, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 288, + 677 + ], + "type": "text", + "content": "[81] Lichao Zhang, Martin Danelljan, Abel Gonzalez-Garcia, Joost van de Weijer, and Fahad Shahbaz Khan. Multi-modal fusion for end-to-end RGB-T tracking. In ICCVW, 2019. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 288, + 713 + ], + "type": "text", + "content": "[82] Pengyu Zhang, Dong Wang, and Hutchuan Lu. Multi-modal visual tracking: Review and experimental comparison. arXiv preprint arXiv:2012.04176, 2020. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 657 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[83] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[84] Wenwei Zhang, Hui Zhou, Shuyang Sun, Zhe Wang, Jianping Shi, and Chen Change Loy. Robust multi-modality multi-object tracking. In ICCV, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 172 + ], + "type": "text", + "content": "[85] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In CVPR, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "text", + "content": "[86] Haojie Zhao, Dong Wang, and Hutchuan Lu. Representation learning for visual object tracking by masked appearance transfer. In CVPR, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 209, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 231 + ], + "type": "text", + "content": "[87] Jinjian Zhao, Xiaohan Zhang, and Pengyu Zhang. A unified approach for tracking uavs in infrared. In ICCV, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 232, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 263 + ], + "type": "text", + "content": "[88] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. PR, 111:107679, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 266, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 297 + ], + "type": "text", + "content": "[89] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 300, + 545, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 343 + ], + "type": "text", + "content": "[90] Jiawen Zhu, Zhenyu Chen, Zeqi Hao, Shijie Chang, Lu Zhang, Dong Wang, Huchuan Lu, Bin Luo, Jun-Yan He, Jin-Peng Lan, et al. Tracking anything in high quality. arXiv preprint arXiv:2307.13974, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 345, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 345, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 345, + 545, + 376 + ], + "type": "text", + "content": "[91] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 378, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 422 + ], + "type": "text", + "content": "[92] Jiawen Zhu, Huayi Tang, Zhi-Qi Cheng, Jun-Yan He, Bin Luo, Shihao Qiu, Shengming Li, and Huchuan Lu. Dcpt: Darkness clue-prompted tracking in nighttime uavs. arXiv preprint arXiv:2309.10491, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 423, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 423, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 423, + 545, + 465 + ], + "type": "text", + "content": "[93] Xue-Feng Zhu, Xiao-Jun Wu, Tianyang Xu, Zhen-Hua Feng, and Josef Kittler. Robust visual object tracking via adaptive attribute-aware discriminative correlation filters. TMM, 24: 301-312, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[94] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. RGBD1K: A large-scale dataset and benchmark for RGB-D object tracking. AAAI, 2023. 2, 5, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 514, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 555 + ], + "type": "text", + "content": "[95] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking. In AAAI, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "type": "text", + "content": "[96] Yabin Zhu, Chenglong Li, Jin Tang, and Bin Luo. Quality-aware feature aggregation network for robust RGBT tracking. IEEE TIV, 6(1):121-130, 2020. 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 591, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 624 + ], + "type": "text", + "content": "[97] Zhiyu Zhu, Junhui Hou, and Xianqiang Lyu. Learning graph-embedded key-event back-tracing for object tracking in event clouds. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "text", + "content": "[98] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event transformer-trackers. In ICCV, 2023. 1, 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19166" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_content_list.json b/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18aa4545a1ef4a5ba186f92ca031990b74101f4d --- /dev/null +++ b/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "Single-View Scene Point Cloud Human Grasp Generation", + "text_level": 1, + "bbox": [ + 192, + 130, + 776, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yan-Kang Wang1 Chengyi Xing2 Yi-Lin Wei1 Xiao-Ming Wu1 Wei-Shi Zheng1,3* \n1School of Computer Science and Engineering, Sun Yat-sen University, China \n2Stanford University, Stanford, CA", + "bbox": [ + 155, + 179, + 818, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 96, + 233, + 872, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangyk65, weiylin5, wuxm65}@mail2.sysu.edu.cn chengyix@stanford.edu wszheng@ieee.org", + "bbox": [ + 101, + 253, + 859, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fcaeed68e163bb55fa933cc7f4dec2172dc5722bcef0cddc46aac51d90a76c6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 297, + 364, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c671fe8c73dcb014f03abdbc967077ad82c22a49d00357d97f96d1dcc2fe283.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 297, + 890, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/428381b2947458e925e9fe3c5ce57a437ba11b19d7565b6c31b7fd70872f0ed3.jpg", + "image_caption": [ + "Dataset Scene" + ], + "image_footnote": [], + "bbox": [ + 80, + 414, + 364, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55f399e5732a9ab3693b12cc556f3f53f033005e758797426a3e6e2900ad752c.jpg", + "image_caption": [ + "Red hands (second row): GraspCVAE Yellow hands (third row): Ours", + "Figure 1. The scene of our dataset and comparison between our model and GraspCVAE [25]. The top-left image depicts the scene in our dataset, comprising a table and an object. Below it are images from four random viewpoints out of 36. The three rows of images on the right side represent, respectively, the single-view images, the generation results of GraspCVAE and our method. The green box indicates the area where the hand-object penetration occurs due to the lack of global perception ability of GraspCVAE." + ], + "image_footnote": [], + "bbox": [ + 364, + 414, + 890, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 559, + 313, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we explore a novel task of generating human grasps based on single-view scene point clouds, which more accurately mirrors the typical real-world situation of observing objects from a single viewpoint. Due to the incompleteness of object point clouds and the presence of numerous scene points, the generated hand is prone to penetrating into the invisible parts of the object and the model is easily affected by scene points. Thus, we introduce S2HGrasp, a framework composed of two key modules: the Global Perception module that globally perceives partial object point clouds, and the DiffuGrasp module designed to generate high-quality human grasps based on complex inputs that include scene points. Additionally, we introduce S2HGD dataset, which comprises approximately 99,000 single-object single-view scene point clouds of 1,668 unique objects, each annotated with one human grasp. Our extensive experiments demonstrate that S2HGrasp can not only generate natural human grasps regardless of scene points, but also effectively prevent penetration between the", + "bbox": [ + 75, + 590, + 473, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hand and invisible parts of the object. Moreover, our model showcases strong generalization capability when applied to unseen objects. Our code and dataset are available at https://github.com/iSEE-Laboratory/S2HGrasp.", + "bbox": [ + 500, + 559, + 892, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 648, + 632, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning and capturing hand-object interactions stands as a foundation for understanding human behaviors and has many practical applications. From augmented and virtual reality [24] to the domains of robotic grasps, human-computer interaction [35, 41] and learning from human demonstrations [13, 15, 37, 45], the comprehension and replication of these interactions are of significance.", + "bbox": [ + 496, + 674, + 890, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we focus on understanding the interactions involved in generating 3D human grasps for objects. We recognize most existing grasp generation methods depend on full 3D object models [25, 26], yet in real-world scenarios, objects and scenes are commonly perceived from a single viewpoint, leading to incomplete object point clouds with potential scene noise. Consequently, existing methods find it challenging to capture the complete geometric fea", + "bbox": [ + 496, + 781, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 888, + 220, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "831", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tures of a partial object and generate human grasps based on such single-view point clouds, limiting their practical applicability. Clearly, a major challenge of this task is the incompleteness of the object. If the model cannot perceive the object's overall shape, the generated grasp is likely to penetrate into the invisible parts of the object. Furthermore, due to interference from scene points, there is a risk that the generated hand will collide with the scene elements.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges, we introduce S2HGrasp, a novel framework designed for generating human grasps on single-view scene point clouds. S2HGrasp incorporates a Global Perception module that facilitates the global understanding of partial object point clouds, and a DiffuGrasp module aiming at generating high-quality human grasps from complex inputs that contain numerous scene points.", + "bbox": [ + 75, + 214, + 468, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, the Global Perception module employs a multi-task learning paradigm, including single-view point cloud completion and classification tasks. This module helps to capture the object's global geometric features, and effectively prevents the generated hands from penetrating into the object's invisible parts. During testing, we can directly extract global features from the input single-view point clouds without the need of point cloud completion and classification tasks. Through feature alignment in latent space, our module can capture global features from the partial object in a single-stage process. Concurrently, the DiffuGrasp module is designed to generate human grasps and it adds noise to the normalized hand parameters when training. Then it conditions on the scene features and predicts the original hand parameters. In testing, the DiffuGrasp module progressively denoises a random noise with the condition of scene features, and predicts the final reasonable hand parameters. Leveraging its powerful conditional generation capabilities, the DiffuGrasp module is capable of generating high-quality grasps that are close to the object, even when the input includes scene points.", + "bbox": [ + 75, + 321, + 470, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we introduce a novel dataset named S2HGD, specifically for human grasp generation on single-view scene point clouds. Constructed using the BlenderProc [11] simulator and derived from OakInk dataset [51], S2HGD has a collection of 1,668 unique objects. Each object is individually placed on a tabletop setting. RGB-D images are captured across 36 distinct viewpoints, and corresponding single-view point clouds are derived from depth information. Overall, S2HGD comprises roughly 99,000 point clouds, each paired with a reasonable human grasp.", + "bbox": [ + 75, + 641, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experimental results indicate that our end-to-end approach S2HGrasp outperforms existing generation methods and two-stage methods without data preprocessing and TTA (test-time adaptation, which fails to work well in our task due to the incompleteness of the objects). Additionally, our model not only generates natural and plausible grasps but also exhibits the ability to generalize to unseen objects.", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows: 1) We explore a new task of human grasp generation on single-view scene point clouds and achieve favorable results. 2) We propose a novel model named S2HGrasp, which employs a Global Perception module to endow the model with the ability to globally perceive single-view point clouds, and a DiffuGrasp module to generate high-quality grasps despite the complex input that contains the object and scene points. 3) We construct S2HGD, a new dataset for human grasp generation based on single-object single-view scene point clouds. Our code and dataset will both be available.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 273, + 640, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hand-object interaction and grasp. Hand-object interaction is currently receiving increasing attention and involves two main tasks: 1) reconstructing 3D models of hands and objects based on images [6, 12, 16-18, 27, 32, 33, 43, 50] and 2) predicting how to grasp an object based on its visual representation (images or point clouds) [4, 8, 10, 25, 26, 29, 42, 52]. GraspTTA [25] can generate human grasps based on the consistency between hand contact points and object contact regions given complete objects. Affordance Diffusion [52] employs diffusion model to generate hand-object interacting images given object images. When it comes to dexterous hand grasp, [49] and [44] work well using generative models and reinforcement learning, and [29] use a single-shot network to predict dexterous hand grasp in clutter. Existing works mainly focus on generating grasps for complete objects, however, in real world, objects are often observed from a single viewpoint. Thus we study human grasp generation on single-view scene point clouds, employing two modules to tackle the challenges of incompleteness of the object and impact of scene elements.", + "bbox": [ + 496, + 297, + 890, + 601 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D point cloud completion. 3D point cloud completion enables the model to have global perception capability and has various methods, including point-based [30, 47, 54], view-based [21, 55], convolution-based [48], graph-based [34], generative model-based [23], transformer-based [53] approaches, etc. Point Completion Network (PCN) [54] is a learning-based shape completion method, and its decoder designed for point cloud features does not require any assumptions about the shape or category of the point clouds. It can generate both coarse and fine-grained completions with a relatively small number of parameters. Motivated by PCN [54], we design a Global Shape Perception (GSP) to complete partial point clouds. Compared to other SOTA methods for point cloud completion, this approach allows for the presence of scene points and achieves a satisfactory completion result, giving the model global awareness.", + "bbox": [ + 496, + 606, + 890, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion model. Diffusion models have emerged as the new deep generative models that progressively destruct data by injecting noise, then learn to reverse this process for sam", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "832", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cb99ead72145e7cc8fb6ffdc8dac14fdabbdb9ee8038ffa5935736a164d2273c.jpg", + "image_caption": [ + "Figure 2. S2HGrasp framework. The scene encoder takes single-view scene point clouds as input and extracts their features through PointNet++ [36] and a transformer block. The features are then used in point cloud completion (GSP), classification (GCP), and grasp generation (DiffuGrasp). GSP and GCP won't be used in testing. In the DiffuGrasp Training, the model adds noise to the normalized hand parameters and extracts hand features after passing the parameters into the MANO layer [39]. Then the object features and hand features will be fed into a transformer decoder to predict the original hand parameters. When testing, the DiffuGrasp Sampling starts from a random noise and iteratively denoises it, resulting in final hand parameters." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ple generation. They have been used in a variety of domains, like computer vision [3, 5, 20, 57], natural language processing [1, 28], temporal data modeling [9] and multimodal modeling [2]. Notably, [22] introduces a SceneDiffuser for 3D scene understanding, and it also generates dexterous hands by adding noise to hand parameters and denoising them. Differently, we convert noised hand parameters into point clouds, extract their features, and then decode them to get final grasps. This improves the naturalness of generated hands, as demonstrated later in our experiments.", + "bbox": [ + 75, + 448, + 472, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. S2HGrasp", + "text_level": 1, + "bbox": [ + 76, + 609, + 191, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement and Method Overview", + "text_level": 1, + "bbox": [ + 76, + 635, + 436, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We explore a novel task: generating physically plausible human grasps for objects based on single-view scene point clouds. We enable the model to globally perceive partial objects and generate high-quality grasps despite interference from scene points. The input is point clouds $\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}$ (where $N$ is the number of points including tabletop and partial object), captured from a single viewpoint. The output is human grasp parameters $\\mathcal{H} \\in \\mathbb{R}^{61}$ modeled via differentiable MANO layer [39], with shape parameter $\\beta \\in \\mathbb{R}^{10}$ for size of the hand and pose parameter $\\theta \\in \\mathbb{R}^{45}$ for joint angle of 15 joints, as well as $\\mathcal{R} \\in \\mathbb{R}^3$ and $\\mathcal{T} \\in \\mathbb{R}^3$ for rotation and translation of wrist joint respectively. Given the hand parameters, MANO layer will output the hand mesh with $\\hat{\\mathcal{M}} = (\\hat{\\mathcal{V}} \\in \\mathcal{R}^{778 \\times 3}, \\hat{\\mathcal{F}})$ , where $\\hat{\\mathcal{V}}$ and $\\hat{\\mathcal{F}}$ denote hand vertices and faces. This task is more aligned with real-world scenarios and is beneficial for practical applications.", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The generated grasps should be not only natural and plausible, but also able to firmly hold objects in physics-based simulators. Due to the incompleteness of object point clouds, the model must possess the capability to globally perceive the object to prevent hands from penetrating into the invisible parts of the object. Moreover, the model needs to generate high-quality human grasps despite interference from scene points. Therefore, we propose a model named S2HGrasp, as summarized in Fig. 2, which consists of two modules: Global Perception module and DiffuGrasp module, to generate human grasps on single-view point clouds.", + "bbox": [ + 496, + 448, + 890, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "S2HGrasp is an end-to-end framework. The Global Perception module equips the model with the ability to perceive global geometric features of partial objects, while the DiffuGrasp module is responsible for resisting interference from scene points and generating high-quality human grasps. In the training process, our model takes single-view point clouds and corresponding hand parameters as input and uses an encoder to extract scene features. The Global Perception module endows these features with the global perception of the object. The DiffuGrasp module conditions on these features, adds noise to hand parameters, and trains a decoder to recover the original hand parameters. In testing, DiffuGrasp starts by sampling a random noise and progressively denoises it to derive the final hand parameters.", + "bbox": [ + 496, + 616, + 892, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Global Perception", + "text_level": 1, + "bbox": [ + 500, + 844, + 676, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Global Perception module aims at enabling the model to perceive the global geometric characteristics of the ob", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "833", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ject from single-view point clouds. Thanks to the multitask learning paradigm in training, our model can directly extract the scene features that contain global perception of the partial object in a single-stage process when testing. The resulting scene features will be used in the DiffuGrasp module, making the generated human grasps more natural, avoiding penetration with invisible parts of the object. Our scene point cloud features $\\mathcal{F}_s$ are primarily extracted using PointNet++ [36], and we utilize a transformer block similar to [56] to add attention mechanisms, making the model focus more on the local regions suitable for grasping.", + "bbox": [ + 75, + 90, + 468, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, we design a Global Shape Perception (GSP) to perform point cloud completion task on the extracted feature, motivated by PCN [54]. After getting the feature $\\mathcal{F}_s$ of single-view scene point clouds, we pass it through our GSP and result in $\\mathcal{N}$ points of the completed object. Subsequently, containing the global perception of the object, $\\mathcal{F}_s$ can be better utilized in the human grasp generation.", + "bbox": [ + 75, + 257, + 468, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, we design a Global Category Perception (GCP) to perform point cloud classification task. Specifically, we feed the obtained scene feature $\\mathcal{F}_s$ into a classification head to output the object category. Since the object is incomplete and the input contains both the object and scene points, this classification task is challenging. Nevertheless, we calculate the F1 score for the classification results of S2HGrasp trained on S2HGD, achieving a score of 0.96. It indicates that GCP performs well, enabling the model to focus more on the object. By predicting the object's category, it can better perceive the object's overall shape.", + "bbox": [ + 75, + 363, + 468, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Perception Loss. In the GSP of the Global Perception module, we use a Chamfer Distance loss $L_{GSP}$ between the completed point clouds and the corresponding complete object point clouds for supervision, enabling it to have the capability to perceive the global feature of the object. In the GCP, we use Cross Entropy Loss $L_{GCP}$ . The total loss of the Global Perception module is defined as,", + "bbox": [ + 75, + 529, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {G l o} = \\lambda_ {G S P} \\cdot L _ {G S P} + \\lambda_ {G C P} \\cdot L _ {G C P}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 638, + 468, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. DiffuGrasp", + "text_level": 1, + "bbox": [ + 76, + 665, + 202, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the input for our task is single-view scene point clouds including the tabletop and the object, the presence of numerous scene points can affect the quality of the generated hand and may lead to collision with the scene points. Therefore, we design the DiffuGrasp module with two innovations, to generate high-quality and diverse human grasps from complex input. Besides, we also design a plane loss to avoid collision between the hand and the tabletop.", + "bbox": [ + 75, + 688, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we add noise to hand parameters, but direct noise injection is not feasible due to the differences in the ranges and meanings of the parameters. Therefore, we perform specific normalization for each parameter based on its meaning and range. Second, hand parameters include hand shape, pose, and position information, and directly decod", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing parameters does not effectively model these information. Thus, we pass them through MANO layer [39] to convert them into point clouds and use hand encoder to obtain their features before feeding them into the decoder.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the practices of conditional diffusion model [20, 38], in our work, the DiffuGrasp module is with condition $\\mathcal{F}_s$ , i.e., the feature of single-view point clouds. The DiffuGrasp module can progressively generate human grasps that are close to the object and avoid penetration with the object and the scene during the denoising process.", + "bbox": [ + 496, + 154, + 890, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Training Process. Conditioning on the single-view scene point cloud feature $\\mathcal{F}_s$ , the DiffuGrasp module adds noise to the normalized hand parameters $\\mathcal{H} \\in \\mathbb{R}^{61}$ . After getting the hand feature $\\mathcal{F}_h$ , we feed it together with $\\mathcal{F}_s$ into a transformer decoder and train the decoder at various noise levels to predict the original hand parameters, i.e., denoise $\\mathcal{H}_0$ for different $\\mathcal{H}_t$ given $t$ , shown as below:", + "bbox": [ + 496, + 248, + 890, + 352 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {t} = \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathcal {H} _ {0} + \\sqrt {\\bar {\\alpha} _ {t}} \\cdot \\epsilon , \\mathcal {H} _ {0} ^ {\\prime} = \\operatorname {M o d e l} \\left(\\mathcal {H} _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 356, + 890, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is noise, $\\alpha$ is noise coefficient and $\\overline{\\alpha}_t = \\prod_{i=1}^t \\alpha_i$ .", + "bbox": [ + 496, + 376, + 880, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- DiffuGrasp Loss. Referring to the codes in DDPM [19], calculating the loss on the input and calculating the loss on the noise are equivalent, so we use Reconstruction Loss along with Cmap Loss, Penetration Loss and Plane Loss to train our transformer decoder. Specifically, 1) Reconstruction Loss of hand parameters and hand point clouds $\\mathcal{P}^h\\in \\mathcal{R}^{778\\times 3}$ is used to minimize reconstruction error,", + "bbox": [ + 496, + 393, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p a r a m} = \\left| \\mathcal {H} _ {0} ^ {\\prime} - \\mathcal {H} _ {0} \\right|, L _ {\\mathcal {V}} = \\left| \\left| \\hat {\\mathcal {V}} - \\mathcal {P} ^ {h} \\right| \\right| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 502, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2) Cmap Loss is used to ensure the generated hand is as close to the object as possible. We compute the distance from the hand's contact points $\\mathcal{V}^p$ to the object's point clouds. For each point $\\mathcal{P}_i^o$ , the distance $\\mathbf{D}(\\mathcal{P}_i^o) = \\min_j ||\\mathcal{V}_j^p - \\mathcal{P}_i^o||_2^2$ . If the distance is less than a threshold, we consider the point as a potentially contactable point on the object, i.e., contact map. One of our training objectives is to make the contact points on the hand as close as possible to the contact map on the object as,", + "bbox": [ + 496, + 525, + 890, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c m a p} = \\sum_ {i} \\mathbf {D} \\left(\\mathcal {P} _ {i} ^ {o}\\right), f o r a l l \\mathbf {D} \\left(\\mathcal {P} _ {i} ^ {o}\\right) \\leq \\mathcal {T}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 665, + 890, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3) Penetration Loss is used to penalize cases where the hand penetrates into the object. Specifically, we define the subset of object points that penetrate into the hand's mesh as $\\mathcal{P}_{in}^{o}$ , and the penetration loss is defined as minimizing the distances from these points to their nearest points on the hand,", + "bbox": [ + 496, + 689, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p e n e t r} = \\frac {1}{\\left| \\mathcal {P} _ {i n} ^ {o} \\right|} \\sum_ {p \\in \\mathcal {P} _ {i n} ^ {o}} \\min _ {i} \\| p - \\hat {\\mathcal {V}} _ {i} \\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 766, + 890, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that we don't use the complete object to calculate the Cmap loss and the Penetration loss, and it is only used in the GSP module. Instead, we combine the input single-view point clouds $\\mathcal{P}^s$ with the corresponding completion result $\\mathcal{P}^c$ from the GSP for loss calculations. Specifically, we first calculate the minimum distance $\\mathbf{D}(\\mathcal{P}_i^c) = \\min_j||\\mathcal{P}_j^s -$", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "834", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{P}_i^c ||_2^2$ , and keep the points $\\hat{\\mathcal{P}}^c = \\{\\mathcal{P}_i^c\\in \\mathcal{P}^c\\mid \\mathbf{D}(\\mathcal{P}_i^c)\\geq \\theta \\}$ We get the final completion result $\\mathcal{P}^C = \\hat{\\mathcal{P}}^c\\cup \\mathcal{P}^s$ , which is used for loss calculation. This approach prevents large errors from directly using rough completion results to compute loss functions. It also allows the model to have a global perception of the object's shape and avoids generating hands that penetrate into the unseen parts of the object. 4) Plane Loss. As for generating human grasps based on scene point clouds, one of the challenges is to avoid contact between the hand and the tabletop point clouds. Thus, we design a plane loss to penalize cases where there is penetration between the tabletop point clouds and the hand. To avoid the enormous computational cost of calculating distances between tabletop point clouds and hand vertices, we regress the parameters of the tabletop plane. We define the hand points which are on the backside of the plane as $\\hat{\\mathcal{V}}_{back}$ and calculate the distance $\\mathbf{D}(\\hat{\\mathcal{V}}_i,P)$ from each point $\\hat{\\mathcal{V}}_i$ in $\\hat{\\mathcal{V}}_{back}$ to the plane $P$ , and the plane loss is defined as,", + "bbox": [ + 76, + 89, + 472, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p l a n e} = \\sum_ {i} \\mathbf {D} (\\hat {\\mathcal {V}} _ {i}, P), f o r a l l \\hat {\\mathcal {V}} _ {i} \\in \\hat {\\mathcal {V}} _ {b a c k}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 367, + 468, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, the loss for the DiffuGrasp is:", + "bbox": [ + 76, + 391, + 316, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D i f f} = \\lambda_ {p a r a m} \\cdot L _ {p a r a m} + \\lambda_ {\\mathcal {V}} \\cdot L _ {\\mathcal {V}} +\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 409, + 411, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {c} \\cdot L _ {c m a p} + \\lambda_ {p e} \\cdot L _ {p e n e t r} + \\lambda_ {p l} \\cdot L _ {p l a n e}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 426, + 411, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Sampling Process. The Sampling Process begins by randomly sampling a noise from standard Gaussian distribution. Subsequently, it undergoes a denoising process for $T$ timesteps with the feature of single-view point clouds serving as a condition, and ultimately yields the final hand parameters. The normalization follows the same procedure as described in the Training Process. We add noise calculated from $\\mathcal{H}_t$ and $t$ to hand parameters $\\mathcal{H}_t$ , and then feed them into the Model, which includes the MANO layer, Hand Encoder, and the Transformer Decoder as shown in Fig. 2, to get $\\mathcal{H}_{t - 1}$ . After $T$ iterations of the denoising process, the model will get final $\\mathcal{H}_0'$ . The distribution $p_{\\theta}\\left(\\mathcal{H}_0|\\mathcal{F}_s\\right)$ represents our sampling process and can be shown as,", + "bbox": [ + 76, + 446, + 472, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathcal {H} _ {0} \\mid \\mathcal {F} _ {s}\\right) = p \\left(\\mathcal {H} _ {T}\\right) \\prod_ {t = 1} ^ {T} p \\left(\\mathcal {H} _ {t - 1} \\mid \\mathcal {H} _ {t}, \\mathcal {F} _ {s}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 652, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathcal {H} _ {t - 1} \\mid \\mathcal {H} _ {t}, \\mathcal {F} _ {s}\\right) = \\mathcal {N} \\left(\\mathcal {H} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathcal {H} _ {t}, t, \\mathcal {F} _ {s}\\right), \\sum_ {\\theta} \\left(\\mathcal {H} _ {t}, t, \\mathcal {F} _ {s}\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 696, + 472, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\mathcal {H} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\epsilon_ {\\theta} ^ {(t)} (\\mathcal {H} _ {t}) + \\sigma_ {t} \\epsilon_ {t}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 742, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_{t} = \\eta \\sqrt{\\frac{1 - \\overline{\\alpha}_{t - 1}}{1 - \\overline{\\alpha}_{t}}\\cdot\\left(1 - \\frac{\\overline{\\alpha}_{t}}{\\overline{\\alpha}_{t - 1}}\\right)}$ $\\eta$ is DDIM [40] sampling coefficient and $\\epsilon_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ is standard Gaussian noise. For more details, please refer to supplementary file.", + "bbox": [ + 76, + 779, + 468, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. S2HGD Dataset", + "text_level": 1, + "bbox": [ + 76, + 844, + 235, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better align with real-world scenarios where objects are typically observed from a single viewpoint and to fur", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ther explore generating human grasps based on single-view scene point clouds, we construct the S2HGD dataset. Here we show how we build our S2HGD with BlenderProc [11]:", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Source. Our raw data (including objects and human grasp annotations) all come from the OakInk dataset [51]. We select a total of 1,667 objects from 35 different categories (e.g. bottle, knife, mug) as the dataset's objects.", + "bbox": [ + 496, + 142, + 890, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Simulation Environment Setting. In the simulation environment, we select a table from ShapeNet [7] and position 36 cameras above the table, arranged in three concentric circular arrays at heights of 0.25 meters, 0.45 meters, and 0.65 meters from the table respectively. Each circular array contains 12 cameras at the same height, with a 30-degree spacing between each camera, as shown in Fig. 1.", + "bbox": [ + 496, + 209, + 890, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Processing. We drop an object from $20~cm$ above the table, allowing it to fall onto the tabletop freely. This is repeated ten times to get ten different poses. For each pose, we capture an RGBD image from each camera and convert it into single-view point clouds. After collecting all point clouds, we test all grasps of each object, discard any that contact the table and remove scenes lacking collision-free grasp annotations. From the remaining grasps, we choose the one closest to the object as the final annotation.", + "bbox": [ + 496, + 321, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable our model to generate high-quality grasps for both varying viewpoints of seen objects and entirely unseen objects, we partition our dataset in two unique ways: ViewS2HGD focuses on varying viewpoints, whereas ObjectS2HGD is based on object diversity. Specifically, ViewS2HGD includes different single-view point clouds of the same objects in training and testing sets. Object-S2HGD ensures that the objects in the testing set are entirely absent from the training set, allowing us to assess the model's capacity for generalization to unseen objects.", + "bbox": [ + 496, + 457, + 890, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In total, S2HGD contains 1,667 objects and about 99,000 single-view scene point clouds with one annotation each.", + "bbox": [ + 496, + 609, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 654, + 633, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 680, + 777, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The single-view point clouds in our dataset consist of 2,000 points, with 1,000 points belonging to the object and 1,000 points to the tabletop. In training, we use Adam optimizer and $\\mathrm{LR} = 1e - 4$ , where the LR is reduced by half when the model trained 100, 160 and 180 epochs. We totally train for 200 epochs and the batch size is 60. Before training the entire model with View-S2HGD and Object-S2HGD, we pre-train the GSP module using single-view point clouds and their corresponding complete objects from each dataset respectively for 200 epochs. The loss weights we use are $\\lambda_{GSP} = 5$ , $\\lambda_{GCP} = 2$ , $\\lambda_{param} = 300$ , $\\lambda_V = 15$ , $\\lambda_{cmap} = 300$ , $\\lambda_{pe} = 15$ , $\\lambda_{pl} = 1$ for View-S2HGD. For Object-S2HGD, only $\\lambda_{pe}$", + "bbox": [ + 496, + 703, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "835", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b3ef22b32e59e048b5413b6a3393056369174ad3707a76fc0b249f1e62470246.jpg", + "image_caption": [ + "Figure 3. The visualization of generated grasps of our S2HGrasp on our two datasets. The blue objects on the left represent the results of View-S2HGD, while the red objects on the right represent the results of Object-S2HGD." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 885, + 310 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b2bc911b7c131fceecec1386072e8842b857be3ecd35effaefa8ac667fe42e96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
View-S2HGDObject-S2HGDHO3D-objectObman-object
GTGC [25]OursGTGC [25]OursGC [25]OursGC [25]Ours
PenetrationDepth(cm)↓0.270.230.210.260.250.212.111.380.320.21
Volume(cm3)↓6.6013.546.587.2214.485.588.664.879.844.62
Grasp DisplaceMean(cm)↓2.013.102.732.183.693.263.613.304.183.25
Variance(cm)↓±2.73±3.49±3.16±2.69±3.94±3.38±3.93±3.73±4.31±3.75
Perceptual Score{1,...,5}↑4.542.603.254.522.543.142.112.892.043.13
ContactRatio(%)↑10098.0799.4110097.8798.6797.6299.0685.8296.06
Diversity(σ2)Axis(10-2)↑2.150.851.032.290.520.710.500.790.560.70
Angle(10-2)↑8.863.294.519.682.042.751.321.952.282.67
", + "bbox": [ + 96, + 345, + 874, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results compared with GraspCVAE [25] on S2HGD. The objects in HO3D and Obman datasets are used to demonstrate the generalization capability. The Penetration Volume metric and the Grasp displace metric are our main indicators. The results show that our method outperforms GraspCVAE on all indicators (in bold) and is significantly effective in reducing hand-object penetration caused by incomplete point clouds. The Penetration results of our method are even better than GT (with a blue background).", + "bbox": [ + 75, + 494, + 892, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $\\lambda_{\\text{param}}$ differ, being set to 20 and 250, respectively.", + "bbox": [ + 75, + 566, + 444, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use S2HGD, Obman [16] and HO3D [14] for our experiments. S2HGD is the dataset we mainly conduct our experiments on. We train and test our model on both View-S2HGD and Object-S2HGD to demonstrate the overall performance. We also conduct ablation experiments on View-S2HGD. Obman and HO3D are used for studying hand-object interaction and Obman is synthetic, while HO3D is real. To validate generalization capability of our method on unseen objects, we utilize the Object-S2HGD for training and testing. We also select 30 objects from Obman and 10 objects from HO3D to construct two testing sets using the same methodology employed in the construction of S2HGD, to further assess generalization capability.", + "bbox": [ + 75, + 580, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The primary goal of our model is to generate plausible and natural human grasps for objects based on single-view scene point clouds. Therefore, our main evaluation metrics involve calculating the penetration depth and volume between the object and the hand. We also use a physics-based simulation environment to calculate the grasp displace to assess the stability of the hand grasping the object. Perceptual score is used to evaluate the natural", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ness of the generated grasps and we invite 50 participants to rate the naturalness of generated grasps, using a scale from 1 to 5, where higher scores indicate higher naturalness. What's more, to demonstrate the contact between the generated hand and the object, we also calculate contact ratio between the hand and the object. Besides, to demonstrate diversity of generated grasp poses, we introduce a diversity metric. Specifically, we calculate the variance of the rotation axes $\\sigma_{axis}^{2}$ and angles $\\sigma_{angle}^{2}$ for 15 joints, excluding the wrist joint, across all grasp samples. Details about the evaluation metrics can be found in the supplementary file.", + "bbox": [ + 496, + 566, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Grasp Generation Performance", + "text_level": 1, + "bbox": [ + 498, + 739, + 779, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualization results. We visualize the generated grasps on both the View-S2HGD and Object-S2HGD datasets, as shown in Fig. 3. The results demonstrate that our model can generate natural and plausible human grasps for different single-view point clouds of seen objects and shows a notable level of generalization for unseen objects. We also visualize the performance of our GSP module, i.e. the partial object point cloud completion results, as shown in the left of Fig. 4. The first row is the input single-view point", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "836", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b8cf197749997a808823859b50c18474e1d99bdf57a32b0ca22791d73da4f9e.jpg", + "image_caption": [ + "Figure 4. Visualization of point cloud completion and our failure cases. Left: Point cloud completion results, with the input single-view point clouds in the top row and completion results below (gray points for the tabletop, red for the object). Right: Failure cases, with single-view images at the top and corresponding failure cases of our method below." + ], + "image_footnote": [], + "bbox": [ + 88, + 89, + 341, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d5ef548b3afac0cebfb3af2e2d9c14097ba91ecca5c8b978baaac7e32d2a7ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 90, + 426, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edd4a9e1e06fad87ed0e661bf8d403d0ebdb8ad038083057c9321a0f407a9fc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 90, + 511, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9aa40d49eea5b1591ef1577eb63665668aa426c3be0d763cfb27e950ee92397.jpg", + "image_caption": [ + "Figure 5. The balance between penetration volume (x-axis) and grasp stability (y-axis). Method whose result point is closer to the origin is more effective." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 596, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3cd44a49d12c5ada88e333c53e1690b6d31d332b6cd488fb4025dc3314ad932.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 97, + 880, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b87a2fbe7bde0d4605ef21061ac41ed44e56f2f119a8b9e69b33a414609e29bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
w/o encoder [22]0.216.703.13±3.4098.49
w/ encoder(ours)0.216.582.73±3.1699.41
", + "bbox": [ + 78, + 297, + 467, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "clouds and the second row is the completion results. The output of our GSP module consists of only the red points. We also visualize some failure cases, shown in the right part of Fig. 4. The first example fails due to the complex shape of the object and it's difficult to perceive its global shape from only partial point clouds, leading to the generated hand penetrating into the object. The last two examples fail because only a small part of the object is visible from the viewpoint. There are too few points to generate a reasonable grasp. More visualization results are in the supplementary file, including the HO3D and Obman datasets.", + "bbox": [ + 75, + 397, + 467, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing with GraspCVAE. As shown in Table 1, we conduct experiments on our View-S2HGD and Object-S2HGD datasets and the testing sets we construct using the objects selected from HO3D [14] and Obman [16] datasets. We compare our model with the SOTA method GraspTTA [25] for the task of generating human grasps based on complete object point clouds. GraspTTA employs ContactNet for Test-Time Adaptation (TTA), which predicts contact areas with the condition of complete object point clouds. Since our task involves single-view scene point clouds as input, with partial object point clouds and interference from scene point clouds, TTA is challenging to apply. Thus, in the comparative experiments, we did not include the TTA module, only utilizing the GraspCVAE network (GC).", + "bbox": [ + 75, + 565, + 467, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results indicate that our model outperforms GraspCVAE across all datasets. From the \"Penetration\" metric, it can be seen that GraspCVAE, due to its lack of global perception of single-view point clouds, produces grasps that are significantly penetrate into the object, while ours are much less, even better than GT. Additionally, our model also demonstrates superior generalization capability on unseen objects compared to GraspCVAE. Regarding the", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/435c738b6e34eae6fd7c123ad1ccfc9230bbe1cdec19774c0aee89dc89b95640.jpg", + "table_caption": [ + "Table 2. Comparison with method [22] that without hand encoder." + ], + "table_footnote": [], + "table_body": "
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
PCN+GC0.2211.346.00±4.8385.66
PCN+DG0.207.686.50±4.5172.85
Ours0.215.583.26±3.3898.67
", + "bbox": [ + 517, + 297, + 870, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with two-stage methods. We use our model trained on Object-S2HGD to compare with others. The two-stage methods initially complete the single-view point cloud (by PCN [54]), and generate human hand grasps on the completed point cloud using GraspCVAE (GC) [25] and our DiffuGrasp(DG).", + "bbox": [ + 498, + 382, + 890, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"Grasp Displace\" metric, a larger penetration volume may result in better grasping stability (the larger the penetration, the greater the force on the fingers in the simulator), which could make the results of GraspCVAE seem favorable. However, an ideal grasp should have both smaller penetration volume and good stability. Therefore, we present the balance on these two metrics of both methods in Fig. 5. It is evident that the points representing our method are closer to the origin point, which also indicates that our method better unifies penetration with grasp stability, meaning the generated grasps are more natural and reasonable.", + "bbox": [ + 496, + 460, + 890, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To generate more natural and stable grasps, instead of directly passing hand parameters into the decoder as done in [22], we first convert the hand parameters into point clouds and extract their features. We then feed the features into the decoder to predict the hand parameters. Our method incorporates more shape and positional information about the hand than direct decoding the hand parameters. To verify the effectiveness of our method, we conduct comparative experiments with the approach without the hand encoder used in [22], as shown in Table 2. The results show that, with a similar penetration volume, our method achieves higher grasp stability (smaller grasp displace), thereby effectively facilitating the generation of higher quality grasps and demonstrating the advantages of our approach.", + "bbox": [ + 496, + 628, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing with two-stage methods. For the task of generating human grasps based on single-view scene point clouds, one intuitive approach is to first utilize point cloud completion methods to complete the single-view point", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "837", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bb2ce2da8431cf2fccae514f9f47b4c8639f8c3f5817123a4aeda6adee603fb8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCVAEDiffuGraspGPpenetrationGrasp Displace Mean ± Variance(cm)↓Contact Ratio(%)↑
Depth(cm)↓Volume(cm3)↓
CVAE✓0.2313.543.10 ± 3.4998.07
DiffuGrasp✓0.2011.543.06 ± 3.3398.80
DiffuGrasp + GP✓✓0.216.582.73 ± 3.1699.41
", + "bbox": [ + 109, + 88, + 857, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation results of the DiffuGrasp module and the Global Perception module(GP) on our View-S2HGD dataset. The role of DiffuGrasp is demonstrated through a comparison with CVAE.", + "bbox": [ + 75, + 172, + 893, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "clouds. Subsequently, a generation method like GraspTTA [25] is employed to generate grasps on the completed point clouds. [46] and [31] utilize such two-stage methods to generate dexterous hand grasps in single-view scenes.", + "bbox": [ + 75, + 209, + 470, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first train a Point Completion Network (PCN) [54] to complete the input single-view point clouds into full objects with tabletop points. Then, we train GraspCVAE (GC) [25] and our DiffuGrasp (DG) to generate human grasps based on these completed results. The testing results are shown in Table 3. The results demonstrate that our method outperforms both two-stage methods by a large margin. The reason is that the two-stage method is not end-to-end, which leads to accumulated errors in the process, impacting the quality of generated grasps. In addition, most works of point cloud completion are not suitable for such a situation with scene points, which will lead to a decrease in the accuracy of point cloud completion. However, the two-stage method has very high requirements for the completion results of single-view point clouds. If the result is poor, the generated grasp can easily penetrate into the object or not make contact with it. In contrast, our end-to-end model effectively avoids such problems. Last, the overall models of two-stage methods tend to be larger, require longer and more tedious training processes, and are less straightforward to train compared to our end-to-end method.", + "bbox": [ + 75, + 270, + 473, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 601, + 232, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study on our S2HGrasp using our View-S2HGD dataset to highlight the role of the Global Perception module and demonstrate the advantages of our DiffuGrasp module over CVAE, as shown in Table 4.", + "bbox": [ + 75, + 626, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Global Perception. The comparative results, as well as the visualizations of point cloud completion, highlight that this module helps the object encoder capture global object characteristics. This, in turn, assists the generation models in producing more natural, collision-free human grasps.", + "bbox": [ + 75, + 686, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, by comparing the results of \"DiffuGrasp\" and \"DiffuGrasp+GP\", we can find that with the Global Perception module, our model can effectively reduce the volume of hand-object penetration (from 11.54 to 6.58) and enhance the stability of the grasp. This also demonstrates that, with the aid of the Global Perception module, our model has a good perception of the overall shape of local objects, and can effectively prevent the generated hand from penetrating into the invisible areas of the object.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DiffuGrasp. The previous work GraspTTA [25] leverages CVAE to generate human grasps based on the full object model, but suffers from severe model collapse and can't generate diverse grasps poses. We believe that the generative capabilities of CVAE are not sufficient, especially for the more challenging task of generating human grasps based on single-view scene point clouds. Therefore, we conduct comparative experiments on our View-S2HGD dataset using our DiffuGrasp and CVAE (without TTA, as TTA cannot be applied to incomplete point clouds) to demonstrate the advantages of our DiffuGrasp in this task.", + "bbox": [ + 496, + 209, + 893, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results show that even without the Global Perception module, the penetration volume of DiffuGrasp is smaller than that of CVAE (11.54 to 13.54). Moreover, DiffuGrasp achieves higher grasp stability than CVAE with smaller penetration, which also shows that the grasps generated by DiffuGrasp are more reasonable and natural. In short, despite the incompleteness of the object and the interference from numerous scene points, DiffuGrasp outperforms CVAE and can generate stable human grasps with minimal penetration into the object. Additional ablation studies and experiments are shown in the supplementary file.", + "bbox": [ + 496, + 376, + 893, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 559, + 619, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we explore a new task of generating human grasps based on single-view scene point clouds rather than full object models. We propose S2HGrasp, along with a new synthetic dataset S2HGD, to address the problem of hand-object penetration caused by the incompleteness of object point clouds. We design a Global Perception module to globally perceive partial objects and a DiffuGrasp module to generate plausible and natural human grasps despite numerous scene points. Our S2HGrasp effectively overcomes the limitation of existing approaches and extends human grasp generation beyond full object models, benefiting the study of hand-object interactions. The experimental results demonstrate that S2HGrasp outperforms other methods and achieves satisfying performance and generalization capabilities across different datasets and unseen objects.", + "bbox": [ + 496, + 585, + 893, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported partially by the National Key Research and Development Program of China(2023YFA1008503),NSFC(U21A20471,U1911401), Guangdong NSF Project (No. 2023B1515040025, 2020B1 515120085).", + "bbox": [ + 496, + 811, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "838", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021. 3", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3", + "[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2", + "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, 2020. 3", + "[6] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing hand-object interactions in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5", + "[8] Jiayi Chen, Mi Yan, Jiazhao Zhang, Yinzhen Xu, Xiaolong Li, Yijia Weng, Li Yi, Shuran Song, and He Wang. Tracking and reconstructing hand object interactions from point cloud sequences in the wild. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2", + "[9] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020. 3", + "[10] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2", + "[11] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. arXiv preprint arXiv:1911.01911, 2019. 2, 5", + "[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2", + "[13] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning." + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1", + "[14] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 6, 7", + "[15] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), 2020. 1", + "[16] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2, 6, 7", + "[17] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020.", + "[18] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), 2021. 2", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 2020. 4", + "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 3, 4", + "[21] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2", + "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7", + "[23] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2", + "[24] Wolfgang Hürst and Casper Van Wezel. Gesture-based interaction via finger tracking for mobile augmented reality. Multimedia Tools and Applications, 2013. 1", + "[25] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2, 6, 7, 8" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "839", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), 2020. 1, 2", + "[27] Mia Kokic, Danica Kragic, and Jeannette Bohg. Learning to estimate pose and shape of hand-held objects from rgb images. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2", + "[28] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 2022. 3", + "[29] Yiming Li, Wei Wei, Daheng Li, Peng Wang, Wanyi Li, and Jun Zhong. Hgc-net: Deep anthropomorphic hand grasping in clutter. In 2022 International Conference on Robotics and Automation (ICRA), 2022. 2", + "[30] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In Proceedings of the AAAI conference on artificial intelligence, 2020. 2", + "[31] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 2021. 8", + "[32] Meysam Madadi, Sergio Escalera, Xavier Baró, and Jordi Gonzalez. End-to-end global to local cnn learning for hand pose recovery in depth data. arXiv preprint arXiv:1705.09606, 2017. 2", + "[33] Markus Oberweger, Paul Wohlhart, and Vincent Lepetit. Hands deep in deep learning for hand pose estimation. arXiv preprint arXiv:1502.06807, 2015. 2", + "[34] Liang Pan. Ecg: Edge-aware point cloud completion with graph convolution. IEEE Robotics and Automation Letters, 2020. 2", + "[35] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems. 2013. 1", + "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 2017. 3, 4", + "[37] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, 2022. 1", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 4", + "[39] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 3, 4", + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Srinath Sridhar, Anna Maria Feit, Christian Theobalt, and Antti Oulasvirta. Investigating the dexterity of multi-finger input for mid-air text entry. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 2015. 1", + "[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, 2020. 2", + "[43] Bugra Tekin, Federica Bogo, and Marc Pollefeys. H+ o: Unified egocentric recognition of 3d hand-object poses and interactions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2", + "[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. arXiv preprint arXiv:2304.00464, 2023. 2", + "[45] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1", + "[46] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 2022. 8", + "[47] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2", + "[48] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In European Conference on Computer Vision, 2020. 2", + "[49] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[50] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[51] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 5", + "[52] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "840", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[53] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Point: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2", + "[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV), 2018. 2, 4, 7, 8", + "[55] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[56] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4", + "[57] Dian Zheng, Xiao-Ming Wu, Zuhao Liu, Jingke Meng, and Wei-shi Zheng. Diffuvolume: Diffusion model for volume based stereo matching. arXiv preprint arXiv:2308.15989, 2023. 3" + ], + "bbox": [ + 78, + 90, + 465, + 441 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "841", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_model.json b/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8c3e9be5cd572322b663c3fb71c324fd191ad5f6 --- /dev/null +++ b/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.131, + 0.777, + 0.154 + ], + "angle": 0, + "content": "Single-View Scene Point Cloud Human Grasp Generation" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.18, + 0.819, + 0.233 + ], + "angle": 0, + "content": "Yan-Kang Wang1 Chengyi Xing2 Yi-Lin Wei1 Xiao-Ming Wu1 Wei-Shi Zheng1,3* \n1School of Computer Science and Engineering, Sun Yat-sen University, China \n2Stanford University, Stanford, CA" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.234, + 0.873, + 0.253 + ], + "angle": 0, + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.255, + 0.86, + 0.27 + ], + "angle": 0, + "content": "{wangyk65, weiylin5, wuxm65}@mail2.sysu.edu.cn chengyix@stanford.edu wszheng@ieee.org" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.299, + 0.365, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.299, + 0.891, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.415, + 0.365, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.415, + 0.891, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.475, + 0.268, + 0.487 + ], + "angle": 0, + "content": "Dataset Scene" + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.475, + 0.865, + 0.489 + ], + "angle": 0, + "content": "Red hands (second row): GraspCVAE Yellow hands (third row): Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.491, + 0.893, + 0.547 + ], + "angle": 0, + "content": "Figure 1. The scene of our dataset and comparison between our model and GraspCVAE [25]. The top-left image depicts the scene in our dataset, comprising a table and an object. Below it are images from four random viewpoints out of 36. The three rows of images on the right side represent, respectively, the single-view images, the generation results of GraspCVAE and our method. The green box indicates the area where the hand-object penetration occurs due to the lack of global perception ability of GraspCVAE." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.56, + 0.314, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.474, + 0.879 + ], + "angle": 0, + "content": "In this work, we explore a novel task of generating human grasps based on single-view scene point clouds, which more accurately mirrors the typical real-world situation of observing objects from a single viewpoint. Due to the incompleteness of object point clouds and the presence of numerous scene points, the generated hand is prone to penetrating into the invisible parts of the object and the model is easily affected by scene points. Thus, we introduce S2HGrasp, a framework composed of two key modules: the Global Perception module that globally perceives partial object point clouds, and the DiffuGrasp module designed to generate high-quality human grasps based on complex inputs that include scene points. Additionally, we introduce S2HGD dataset, which comprises approximately 99,000 single-object single-view scene point clouds of 1,668 unique objects, each annotated with one human grasp. Our extensive experiments demonstrate that S2HGrasp can not only generate natural human grasps regardless of scene points, but also effectively prevent penetration between the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.56, + 0.893, + 0.621 + ], + "angle": 0, + "content": "hand and invisible parts of the object. Moreover, our model showcases strong generalization capability when applied to unseen objects. Our code and dataset are available at https://github.com/iSEE-Laboratory/S2HGrasp." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.65, + 0.633, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.781 + ], + "angle": 0, + "content": "Learning and capturing hand-object interactions stands as a foundation for understanding human behaviors and has many practical applications. From augmented and virtual reality [24] to the domains of robotic grasps, human-computer interaction [35, 41] and learning from human demonstrations [13, 15, 37, 45], the comprehension and replication of these interactions are of significance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.893, + 0.903 + ], + "angle": 0, + "content": "In this work, we focus on understanding the interactions involved in generating 3D human grasps for objects. We recognize most existing grasp generation methods depend on full 3D object models [25, 26], yet in real-world scenarios, objects and scenes are commonly perceived from a single viewpoint, leading to incomplete object point clouds with potential scene noise. Consequently, existing methods find it challenging to capture the complete geometric fea" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.889, + 0.221, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "831" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "tures of a partial object and generate human grasps based on such single-view point clouds, limiting their practical applicability. Clearly, a major challenge of this task is the incompleteness of the object. If the model cannot perceive the object's overall shape, the generated grasp is likely to penetrate into the invisible parts of the object. Furthermore, due to interference from scene points, there is a risk that the generated hand will collide with the scene elements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.47, + 0.32 + ], + "angle": 0, + "content": "To address the challenges, we introduce S2HGrasp, a novel framework designed for generating human grasps on single-view scene point clouds. S2HGrasp incorporates a Global Perception module that facilitates the global understanding of partial object point clouds, and a DiffuGrasp module aiming at generating high-quality human grasps from complex inputs that contain numerous scene points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.471, + 0.64 + ], + "angle": 0, + "content": "Specifically, the Global Perception module employs a multi-task learning paradigm, including single-view point cloud completion and classification tasks. This module helps to capture the object's global geometric features, and effectively prevents the generated hands from penetrating into the object's invisible parts. During testing, we can directly extract global features from the input single-view point clouds without the need of point cloud completion and classification tasks. Through feature alignment in latent space, our module can capture global features from the partial object in a single-stage process. Concurrently, the DiffuGrasp module is designed to generate human grasps and it adds noise to the normalized hand parameters when training. Then it conditions on the scene features and predicts the original hand parameters. In testing, the DiffuGrasp module progressively denoises a random noise with the condition of scene features, and predicts the final reasonable hand parameters. Leveraging its powerful conditional generation capabilities, the DiffuGrasp module is capable of generating high-quality grasps that are close to the object, even when the input includes scene points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.793 + ], + "angle": 0, + "content": "In addition, we introduce a novel dataset named S2HGD, specifically for human grasp generation on single-view scene point clouds. Constructed using the BlenderProc [11] simulator and derived from OakInk dataset [51], S2HGD has a collection of 1,668 unique objects. Each object is individually placed on a tabletop setting. RGB-D images are captured across 36 distinct viewpoints, and corresponding single-view point clouds are derived from depth information. Overall, S2HGD comprises roughly 99,000 point clouds, each paired with a reasonable human grasp." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Our experimental results indicate that our end-to-end approach S2HGrasp outperforms existing generation methods and two-stage methods without data preprocessing and TTA (test-time adaptation, which fails to work well in our task due to the incompleteness of the objects). Additionally, our model not only generates natural and plausible grasps but also exhibits the ability to generalize to unseen objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows: 1) We explore a new task of human grasp generation on single-view scene point clouds and achieve favorable results. 2) We propose a novel model named S2HGrasp, which employs a Global Perception module to endow the model with the ability to globally perceive single-view point clouds, and a DiffuGrasp module to generate high-quality grasps despite the complex input that contains the object and scene points. 3) We construct S2HGD, a new dataset for human grasp generation based on single-object single-view scene point clouds. Our code and dataset will both be available." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.274, + 0.642, + 0.289 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Hand-object interaction and grasp. Hand-object interaction is currently receiving increasing attention and involves two main tasks: 1) reconstructing 3D models of hands and objects based on images [6, 12, 16-18, 27, 32, 33, 43, 50] and 2) predicting how to grasp an object based on its visual representation (images or point clouds) [4, 8, 10, 25, 26, 29, 42, 52]. GraspTTA [25] can generate human grasps based on the consistency between hand contact points and object contact regions given complete objects. Affordance Diffusion [52] employs diffusion model to generate hand-object interacting images given object images. When it comes to dexterous hand grasp, [49] and [44] work well using generative models and reinforcement learning, and [29] use a single-shot network to predict dexterous hand grasp in clutter. Existing works mainly focus on generating grasps for complete objects, however, in real world, objects are often observed from a single viewpoint. Thus we study human grasp generation on single-view scene point clouds, employing two modules to tackle the challenges of incompleteness of the object and impact of scene elements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.85 + ], + "angle": 0, + "content": "3D point cloud completion. 3D point cloud completion enables the model to have global perception capability and has various methods, including point-based [30, 47, 54], view-based [21, 55], convolution-based [48], graph-based [34], generative model-based [23], transformer-based [53] approaches, etc. Point Completion Network (PCN) [54] is a learning-based shape completion method, and its decoder designed for point cloud features does not require any assumptions about the shape or category of the point clouds. It can generate both coarse and fine-grained completions with a relatively small number of parameters. Motivated by PCN [54], we design a Global Shape Perception (GSP) to complete partial point clouds. Compared to other SOTA methods for point cloud completion, this approach allows for the presence of scene points and achieves a satisfactory completion result, giving the model global awareness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Diffusion model. Diffusion models have emerged as the new deep generative models that progressively destruct data by injecting noise, then learn to reverse this process for sam" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Figure 2. S2HGrasp framework. The scene encoder takes single-view scene point clouds as input and extracts their features through PointNet++ [36] and a transformer block. The features are then used in point cloud completion (GSP), classification (GCP), and grasp generation (DiffuGrasp). GSP and GCP won't be used in testing. In the DiffuGrasp Training, the model adds noise to the normalized hand parameters and extracts hand features after passing the parameters into the MANO layer [39]. Then the object features and hand features will be fed into a transformer decoder to predict the original hand parameters. When testing, the DiffuGrasp Sampling starts from a random noise and iteratively denoises it, resulting in final hand parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.473, + 0.603 + ], + "angle": 0, + "content": "ple generation. They have been used in a variety of domains, like computer vision [3, 5, 20, 57], natural language processing [1, 28], temporal data modeling [9] and multimodal modeling [2]. Notably, [22] introduces a SceneDiffuser for 3D scene understanding, and it also generates dexterous hands by adding noise to hand parameters and denoising them. Differently, we convert noised hand parameters into point clouds, extract their features, and then decode them to get final grasps. This improves the naturalness of generated hands, as demonstrated later in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.611, + 0.192, + 0.629 + ], + "angle": 0, + "content": "3. S2HGrasp" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.437, + 0.652 + ], + "angle": 0, + "content": "3.1. Problem Statement and Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We explore a novel task: generating physically plausible human grasps for objects based on single-view scene point clouds. We enable the model to globally perceive partial objects and generate high-quality grasps despite interference from scene points. The input is point clouds \\(\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}\\) (where \\(N\\) is the number of points including tabletop and partial object), captured from a single viewpoint. The output is human grasp parameters \\(\\mathcal{H} \\in \\mathbb{R}^{61}\\) modeled via differentiable MANO layer [39], with shape parameter \\(\\beta \\in \\mathbb{R}^{10}\\) for size of the hand and pose parameter \\(\\theta \\in \\mathbb{R}^{45}\\) for joint angle of 15 joints, as well as \\(\\mathcal{R} \\in \\mathbb{R}^3\\) and \\(\\mathcal{T} \\in \\mathbb{R}^3\\) for rotation and translation of wrist joint respectively. Given the hand parameters, MANO layer will output the hand mesh with \\(\\hat{\\mathcal{M}} = (\\hat{\\mathcal{V}} \\in \\mathcal{R}^{778 \\times 3}, \\hat{\\mathcal{F}})\\), where \\(\\hat{\\mathcal{V}}\\) and \\(\\hat{\\mathcal{F}}\\) denote hand vertices and faces. This task is more aligned with real-world scenarios and is beneficial for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.617 + ], + "angle": 0, + "content": "The generated grasps should be not only natural and plausible, but also able to firmly hold objects in physics-based simulators. Due to the incompleteness of object point clouds, the model must possess the capability to globally perceive the object to prevent hands from penetrating into the invisible parts of the object. Moreover, the model needs to generate high-quality human grasps despite interference from scene points. Therefore, we propose a model named S2HGrasp, as summarized in Fig. 2, which consists of two modules: Global Perception module and DiffuGrasp module, to generate human grasps on single-view point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.893, + 0.83 + ], + "angle": 0, + "content": "S2HGrasp is an end-to-end framework. The Global Perception module equips the model with the ability to perceive global geometric features of partial objects, while the DiffuGrasp module is responsible for resisting interference from scene points and generating high-quality human grasps. In the training process, our model takes single-view point clouds and corresponding hand parameters as input and uses an encoder to extract scene features. The Global Perception module endows these features with the global perception of the object. The DiffuGrasp module conditions on these features, adds noise to hand parameters, and trains a decoder to recover the original hand parameters. In testing, DiffuGrasp starts by sampling a random noise and progressively denoises it to derive the final hand parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.677, + 0.862 + ], + "angle": 0, + "content": "3.2. Global Perception" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The Global Perception module aims at enabling the model to perceive the global geometric characteristics of the ob" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "833" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.257 + ], + "angle": 0, + "content": "ject from single-view point clouds. Thanks to the multitask learning paradigm in training, our model can directly extract the scene features that contain global perception of the partial object in a single-stage process when testing. The resulting scene features will be used in the DiffuGrasp module, making the generated human grasps more natural, avoiding penetration with invisible parts of the object. Our scene point cloud features \\(\\mathcal{F}_s\\) are primarily extracted using PointNet++ [36], and we utilize a transformer block similar to [56] to add attention mechanisms, making the model focus more on the local regions suitable for grasping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.469, + 0.363 + ], + "angle": 0, + "content": "Firstly, we design a Global Shape Perception (GSP) to perform point cloud completion task on the extracted feature, motivated by PCN [54]. After getting the feature \\(\\mathcal{F}_s\\) of single-view scene point clouds, we pass it through our GSP and result in \\(\\mathcal{N}\\) points of the completed object. Subsequently, containing the global perception of the object, \\(\\mathcal{F}_s\\) can be better utilized in the human grasp generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.469, + 0.529 + ], + "angle": 0, + "content": "Secondly, we design a Global Category Perception (GCP) to perform point cloud classification task. Specifically, we feed the obtained scene feature \\(\\mathcal{F}_s\\) into a classification head to output the object category. Since the object is incomplete and the input contains both the object and scene points, this classification task is challenging. Nevertheless, we calculate the F1 score for the classification results of S2HGrasp trained on S2HGD, achieving a score of 0.96. It indicates that GCP performs well, enabling the model to focus more on the object. By predicting the object's category, it can better perceive the object's overall shape." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.469, + 0.635 + ], + "angle": 0, + "content": "Global Perception Loss. In the GSP of the Global Perception module, we use a Chamfer Distance loss \\( L_{GSP} \\) between the completed point clouds and the corresponding complete object point clouds for supervision, enabling it to have the capability to perceive the global feature of the object. In the GCP, we use Cross Entropy Loss \\( L_{GCP} \\). The total loss of the Global Perception module is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.64, + 0.469, + 0.655 + ], + "angle": 0, + "content": "\\[\nL _ {G l o} = \\lambda_ {G S P} \\cdot L _ {G S P} + \\lambda_ {G C P} \\cdot L _ {G C P}. \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.666, + 0.203, + 0.682 + ], + "angle": 0, + "content": "3.3. DiffuGrasp" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Since the input for our task is single-view scene point clouds including the tabletop and the object, the presence of numerous scene points can affect the quality of the generated hand and may lead to collision with the scene points. Therefore, we design the DiffuGrasp module with two innovations, to generate high-quality and diverse human grasps from complex input. Besides, we also design a plane loss to avoid collision between the hand and the tabletop." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "First, we add noise to hand parameters, but direct noise injection is not feasible due to the differences in the ranges and meanings of the parameters. Therefore, we perform specific normalization for each parameter based on its meaning and range. Second, hand parameters include hand shape, pose, and position information, and directly decod" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "ing parameters does not effectively model these information. Thus, we pass them through MANO layer [39] to convert them into point clouds and use hand encoder to obtain their features before feeding them into the decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.892, + 0.245 + ], + "angle": 0, + "content": "Following the practices of conditional diffusion model [20, 38], in our work, the DiffuGrasp module is with condition \\(\\mathcal{F}_s\\), i.e., the feature of single-view point clouds. The DiffuGrasp module can progressively generate human grasps that are close to the object and avoid penetration with the object and the scene during the denoising process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.353 + ], + "angle": 0, + "content": "- Training Process. Conditioning on the single-view scene point cloud feature \\(\\mathcal{F}_s\\), the DiffuGrasp module adds noise to the normalized hand parameters \\(\\mathcal{H} \\in \\mathbb{R}^{61}\\). After getting the hand feature \\(\\mathcal{F}_h\\), we feed it together with \\(\\mathcal{F}_s\\) into a transformer decoder and train the decoder at various noise levels to predict the original hand parameters, i.e., denoise \\(\\mathcal{H}_0\\) for different \\(\\mathcal{H}_t\\) given \\(t\\), shown as below:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.357, + 0.892, + 0.374 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {t} = \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathcal {H} _ {0} + \\sqrt {\\bar {\\alpha} _ {t}} \\cdot \\epsilon , \\mathcal {H} _ {0} ^ {\\prime} = \\operatorname {M o d e l} \\left(\\mathcal {H} _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.882, + 0.394 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is noise, \\(\\alpha\\) is noise coefficient and \\(\\overline{\\alpha}_t = \\prod_{i=1}^t \\alpha_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.5 + ], + "angle": 0, + "content": "- DiffuGrasp Loss. Referring to the codes in DDPM [19], calculating the loss on the input and calculating the loss on the noise are equivalent, so we use Reconstruction Loss along with Cmap Loss, Penetration Loss and Plane Loss to train our transformer decoder. Specifically, 1) Reconstruction Loss of hand parameters and hand point clouds \\(\\mathcal{P}^h\\in \\mathcal{R}^{778\\times 3}\\) is used to minimize reconstruction error," + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.503, + 0.891, + 0.521 + ], + "angle": 0, + "content": "\\[\nL _ {p a r a m} = \\left| \\mathcal {H} _ {0} ^ {\\prime} - \\mathcal {H} _ {0} \\right|, L _ {\\mathcal {V}} = \\left| \\left| \\hat {\\mathcal {V}} - \\mathcal {P} ^ {h} \\right| \\right| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.662 + ], + "angle": 0, + "content": "2) Cmap Loss is used to ensure the generated hand is as close to the object as possible. We compute the distance from the hand's contact points \\(\\mathcal{V}^p\\) to the object's point clouds. For each point \\(\\mathcal{P}_i^o\\), the distance \\(\\mathbf{D}(\\mathcal{P}_i^o) = \\min_j ||\\mathcal{V}_j^p - \\mathcal{P}_i^o||_2^2\\). If the distance is less than a threshold, we consider the point as a potentially contactable point on the object, i.e., contact map. One of our training objectives is to make the contact points on the hand as close as possible to the contact map on the object as," + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.666, + 0.891, + 0.689 + ], + "angle": 0, + "content": "\\[\nL _ {c m a p} = \\sum_ {i} \\mathbf {D} \\left(\\mathcal {P} _ {i} ^ {o}\\right), f o r a l l \\mathbf {D} \\left(\\mathcal {P} _ {i} ^ {o}\\right) \\leq \\mathcal {T}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.764 + ], + "angle": 0, + "content": "3) Penetration Loss is used to penalize cases where the hand penetrates into the object. Specifically, we define the subset of object points that penetrate into the hand's mesh as \\(\\mathcal{P}_{in}^{o}\\), and the penetration loss is defined as minimizing the distances from these points to their nearest points on the hand," + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.767, + 0.891, + 0.805 + ], + "angle": 0, + "content": "\\[\nL _ {p e n e t r} = \\frac {1}{\\left| \\mathcal {P} _ {i n} ^ {o} \\right|} \\sum_ {p \\in \\mathcal {P} _ {i n} ^ {o}} \\min _ {i} \\| p - \\hat {\\mathcal {V}} _ {i} \\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Note that we don't use the complete object to calculate the Cmap loss and the Penetration loss, and it is only used in the GSP module. Instead, we combine the input single-view point clouds \\(\\mathcal{P}^s\\) with the corresponding completion result \\(\\mathcal{P}^c\\) from the GSP for loss calculations. Specifically, we first calculate the minimum distance \\(\\mathbf{D}(\\mathcal{P}_i^c) = \\min_j||\\mathcal{P}_j^s -\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "834" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.473, + 0.365 + ], + "angle": 0, + "content": "\\(\\mathcal{P}_i^c ||_2^2\\) , and keep the points \\(\\hat{\\mathcal{P}}^c = \\{\\mathcal{P}_i^c\\in \\mathcal{P}^c\\mid \\mathbf{D}(\\mathcal{P}_i^c)\\geq \\theta \\}\\) We get the final completion result \\(\\mathcal{P}^C = \\hat{\\mathcal{P}}^c\\cup \\mathcal{P}^s\\) , which is used for loss calculation. This approach prevents large errors from directly using rough completion results to compute loss functions. It also allows the model to have a global perception of the object's shape and avoids generating hands that penetrate into the unseen parts of the object. 4) Plane Loss. As for generating human grasps based on scene point clouds, one of the challenges is to avoid contact between the hand and the tabletop point clouds. Thus, we design a plane loss to penalize cases where there is penetration between the tabletop point clouds and the hand. To avoid the enormous computational cost of calculating distances between tabletop point clouds and hand vertices, we regress the parameters of the tabletop plane. We define the hand points which are on the backside of the plane as \\(\\hat{\\mathcal{V}}_{back}\\) and calculate the distance \\(\\mathbf{D}(\\hat{\\mathcal{V}}_i,P)\\) from each point \\(\\hat{\\mathcal{V}}_i\\) in \\(\\hat{\\mathcal{V}}_{back}\\) to the plane \\(P\\) , and the plane loss is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.368, + 0.47, + 0.391 + ], + "angle": 0, + "content": "\\[\nL _ {p l a n e} = \\sum_ {i} \\mathbf {D} (\\hat {\\mathcal {V}} _ {i}, P), f o r a l l \\hat {\\mathcal {V}} _ {i} \\in \\hat {\\mathcal {V}} _ {b a c k}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.392, + 0.318, + 0.407 + ], + "angle": 0, + "content": "Thus, the loss for the DiffuGrasp is:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.41, + 0.412, + 0.428 + ], + "angle": 0, + "content": "\\[\nL _ {D i f f} = \\lambda_ {p a r a m} \\cdot L _ {p a r a m} + \\lambda_ {\\mathcal {V}} \\cdot L _ {\\mathcal {V}} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.428, + 0.412, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {c} \\cdot L _ {c m a p} + \\lambda_ {p e} \\cdot L _ {p e n e t r} + \\lambda_ {p l} \\cdot L _ {p l a n e}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.448, + 0.473, + 0.645 + ], + "angle": 0, + "content": "- Sampling Process. The Sampling Process begins by randomly sampling a noise from standard Gaussian distribution. Subsequently, it undergoes a denoising process for \\( T \\) timesteps with the feature of single-view point clouds serving as a condition, and ultimately yields the final hand parameters. The normalization follows the same procedure as described in the Training Process. We add noise calculated from \\( \\mathcal{H}_t \\) and \\( t \\) to hand parameters \\( \\mathcal{H}_t \\), and then feed them into the Model, which includes the MANO layer, Hand Encoder, and the Transformer Decoder as shown in Fig. 2, to get \\( \\mathcal{H}_{t - 1} \\). After \\( T \\) iterations of the denoising process, the model will get final \\( \\mathcal{H}_0' \\). The distribution \\( p_{\\theta}\\left(\\mathcal{H}_0|\\mathcal{F}_s\\right) \\) represents our sampling process and can be shown as," + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.653, + 0.47, + 0.693 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathcal {H} _ {0} \\mid \\mathcal {F} _ {s}\\right) = p \\left(\\mathcal {H} _ {T}\\right) \\prod_ {t = 1} ^ {T} p \\left(\\mathcal {H} _ {t - 1} \\mid \\mathcal {H} _ {t}, \\mathcal {F} _ {s}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.697, + 0.473, + 0.742 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathcal {H} _ {t - 1} \\mid \\mathcal {H} _ {t}, \\mathcal {F} _ {s}\\right) = \\mathcal {N} \\left(\\mathcal {H} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathcal {H} _ {t}, t, \\mathcal {F} _ {s}\\right), \\sum_ {\\theta} \\left(\\mathcal {H} _ {t}, t, \\mathcal {F} _ {s}\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\mathcal {H} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\epsilon_ {\\theta} ^ {(t)} (\\mathcal {H} _ {t}) + \\sigma_ {t} \\epsilon_ {t}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.47, + 0.834 + ], + "angle": 0, + "content": "where \\(\\sigma_{t} = \\eta \\sqrt{\\frac{1 - \\overline{\\alpha}_{t - 1}}{1 - \\overline{\\alpha}_{t}}\\cdot\\left(1 - \\frac{\\overline{\\alpha}_{t}}{\\overline{\\alpha}_{t - 1}}\\right)}\\) \\(\\eta\\) is DDIM [40] sampling coefficient and \\(\\epsilon_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) is standard Gaussian noise. For more details, please refer to supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.236, + 0.861 + ], + "angle": 0, + "content": "4. S2HGD Dataset" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To better align with real-world scenarios where objects are typically observed from a single viewpoint and to fur" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "ther explore generating human grasps based on single-view scene point clouds, we construct the S2HGD dataset. Here we show how we build our S2HGD with BlenderProc [11]:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.143, + 0.892, + 0.204 + ], + "angle": 0, + "content": "Data Source. Our raw data (including objects and human grasp annotations) all come from the OakInk dataset [51]. We select a total of 1,667 objects from 35 different categories (e.g. bottle, knife, mug) as the dataset's objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.21, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Simulation Environment Setting. In the simulation environment, we select a table from ShapeNet [7] and position 36 cameras above the table, arranged in three concentric circular arrays at heights of 0.25 meters, 0.45 meters, and 0.65 meters from the table respectively. Each circular array contains 12 cameras at the same height, with a 30-degree spacing between each camera, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Data Processing. We drop an object from \\(20~cm\\) above the table, allowing it to fall onto the tabletop freely. This is repeated ten times to get ten different poses. For each pose, we capture an RGBD image from each camera and convert it into single-view point clouds. After collecting all point clouds, we test all grasps of each object, discard any that contact the table and remove scenes lacking collision-free grasp annotations. From the remaining grasps, we choose the one closest to the object as the final annotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.609 + ], + "angle": 0, + "content": "To enable our model to generate high-quality grasps for both varying viewpoints of seen objects and entirely unseen objects, we partition our dataset in two unique ways: ViewS2HGD focuses on varying viewpoints, whereas ObjectS2HGD is based on object diversity. Specifically, ViewS2HGD includes different single-view point clouds of the same objects in training and testing sets. Object-S2HGD ensures that the objects in the testing set are entirely absent from the training set, allowing us to assess the model's capacity for generalization to unseen objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.641 + ], + "angle": 0, + "content": "In total, S2HGD contains 1,667 objects and about 99,000 single-view scene point clouds with one annotation each." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.634, + 0.672 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.779, + 0.698 + ], + "angle": 0, + "content": "5.1. Datasets and Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. The single-view point clouds in our dataset consist of 2,000 points, with 1,000 points belonging to the object and 1,000 points to the tabletop. In training, we use Adam optimizer and \\(\\mathrm{LR} = 1e - 4\\), where the LR is reduced by half when the model trained 100, 160 and 180 epochs. We totally train for 200 epochs and the batch size is 60. Before training the entire model with View-S2HGD and Object-S2HGD, we pre-train the GSP module using single-view point clouds and their corresponding complete objects from each dataset respectively for 200 epochs. The loss weights we use are \\(\\lambda_{GSP} = 5\\), \\(\\lambda_{GCP} = 2\\), \\(\\lambda_{param} = 300\\), \\(\\lambda_V = 15\\), \\(\\lambda_{cmap} = 300\\), \\(\\lambda_{pe} = 15\\), \\(\\lambda_{pl} = 1\\) for View-S2HGD. For Object-S2HGD, only \\(\\lambda_{pe}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "835" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.887, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 3. The visualization of generated grasps of our S2HGrasp on our two datasets. The blue objects on the left represent the results of View-S2HGD, while the red objects on the right represent the results of Object-S2HGD." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.346, + 0.875, + 0.486 + ], + "angle": 0, + "content": "
View-S2HGDObject-S2HGDHO3D-objectObman-object
GTGC [25]OursGTGC [25]OursGC [25]OursGC [25]Ours
PenetrationDepth(cm)↓0.270.230.210.260.250.212.111.380.320.21
Volume(cm3)↓6.6013.546.587.2214.485.588.664.879.844.62
Grasp DisplaceMean(cm)↓2.013.102.732.183.693.263.613.304.183.25
Variance(cm)↓±2.73±3.49±3.16±2.69±3.94±3.38±3.93±3.73±4.31±3.75
Perceptual Score{1,...,5}↑4.542.603.254.522.543.142.112.892.043.13
ContactRatio(%)↑10098.0799.4110097.8798.6797.6299.0685.8296.06
Diversity(σ2)Axis(10-2)↑2.150.851.032.290.520.710.500.790.560.70
Angle(10-2)↑8.863.294.519.682.042.751.321.952.282.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.496, + 0.893, + 0.555 + ], + "angle": 0, + "content": "Table 1. Quantitative results compared with GraspCVAE [25] on S2HGD. The objects in HO3D and Obman datasets are used to demonstrate the generalization capability. The Penetration Volume metric and the Grasp displace metric are our main indicators. The results show that our method outperforms GraspCVAE on all indicators (in bold) and is significantly effective in reducing hand-object penetration caused by incomplete point clouds. The Penetration results of our method are even better than GT (with a blue background)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.446, + 0.582 + ], + "angle": 0, + "content": "and \\(\\lambda_{\\text{param}}\\) differ, being set to 20 and 250, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Datasets. We use S2HGD, Obman [16] and HO3D [14] for our experiments. S2HGD is the dataset we mainly conduct our experiments on. We train and test our model on both View-S2HGD and Object-S2HGD to demonstrate the overall performance. We also conduct ablation experiments on View-S2HGD. Obman and HO3D are used for studying hand-object interaction and Obman is synthetic, while HO3D is real. To validate generalization capability of our method on unseen objects, we utilize the Object-S2HGD for training and testing. We also select 30 objects from Obman and 10 objects from HO3D to construct two testing sets using the same methodology employed in the construction of S2HGD, to further assess generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. The primary goal of our model is to generate plausible and natural human grasps for objects based on single-view scene point clouds. Therefore, our main evaluation metrics involve calculating the penetration depth and volume between the object and the hand. We also use a physics-based simulation environment to calculate the grasp displace to assess the stability of the hand grasping the object. Perceptual score is used to evaluate the natural" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.734 + ], + "angle": 0, + "content": "ness of the generated grasps and we invite 50 participants to rate the naturalness of generated grasps, using a scale from 1 to 5, where higher scores indicate higher naturalness. What's more, to demonstrate the contact between the generated hand and the object, we also calculate contact ratio between the hand and the object. Besides, to demonstrate diversity of generated grasp poses, we introduce a diversity metric. Specifically, we calculate the variance of the rotation axes \\(\\sigma_{axis}^{2}\\) and angles \\(\\sigma_{angle}^{2}\\) for 15 joints, excluding the wrist joint, across all grasp samples. Details about the evaluation metrics can be found in the supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.78, + 0.758 + ], + "angle": 0, + "content": "5.2. Grasp Generation Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Visualization results. We visualize the generated grasps on both the View-S2HGD and Object-S2HGD datasets, as shown in Fig. 3. The results demonstrate that our model can generate natural and plausible human grasps for different single-view point clouds of seen objects and shows a notable level of generalization for unseen objects. We also visualize the performance of our GSP module, i.e. the partial object point cloud completion results, as shown in the left of Fig. 4. The first row is the input single-view point" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "836" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.09, + 0.342, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.091, + 0.427, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.091, + 0.512, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.598, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.231, + 0.611, + 0.286 + ], + "angle": 0, + "content": "Figure 4. Visualization of point cloud completion and our failure cases. Left: Point cloud completion results, with the input single-view point clouds in the top row and completion results below (gray points for the tabletop, red for the object). Right: Failure cases, with single-view images at the top and corresponding failure cases of our method below." + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.098, + 0.882, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.231, + 0.887, + 0.286 + ], + "angle": 0, + "content": "Figure 5. The balance between penetration volume (x-axis) and grasp stability (y-axis). Method whose result point is closer to the origin is more effective." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.298, + 0.468, + 0.365 + ], + "angle": 0, + "content": "
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
w/o encoder [22]0.216.703.13±3.4098.49
w/ encoder(ours)0.216.582.73±3.1699.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.371, + 0.468, + 0.384 + ], + "angle": 0, + "content": "Table 2. Comparison with method [22] that without hand encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.468, + 0.563 + ], + "angle": 0, + "content": "clouds and the second row is the completion results. The output of our GSP module consists of only the red points. We also visualize some failure cases, shown in the right part of Fig. 4. The first example fails due to the complex shape of the object and it's difficult to perceive its global shape from only partial point clouds, leading to the generated hand penetrating into the object. The last two examples fail because only a small part of the object is visible from the viewpoint. There are too few points to generate a reasonable grasp. More visualization results are in the supplementary file, including the HO3D and Obman datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.468, + 0.777 + ], + "angle": 0, + "content": "Comparing with GraspCVAE. As shown in Table 1, we conduct experiments on our View-S2HGD and Object-S2HGD datasets and the testing sets we construct using the objects selected from HO3D [14] and Obman [16] datasets. We compare our model with the SOTA method GraspTTA [25] for the task of generating human grasps based on complete object point clouds. GraspTTA employs ContactNet for Test-Time Adaptation (TTA), which predicts contact areas with the condition of complete object point clouds. Since our task involves single-view scene point clouds as input, with partial object point clouds and interference from scene point clouds, TTA is challenging to apply. Thus, in the comparative experiments, we did not include the TTA module, only utilizing the GraspCVAE network (GC)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The results indicate that our model outperforms GraspCVAE across all datasets. From the \"Penetration\" metric, it can be seen that GraspCVAE, due to its lack of global perception of single-view point clouds, produces grasps that are significantly penetrate into the object, while ours are much less, even better than GT. Additionally, our model also demonstrates superior generalization capability on unseen objects compared to GraspCVAE. Regarding the" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.298, + 0.871, + 0.378 + ], + "angle": 0, + "content": "
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
PCN+GC0.2211.346.00±4.8385.66
PCN+DG0.207.686.50±4.5172.85
Ours0.215.583.26±3.3898.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Table 3. Comparison with two-stage methods. We use our model trained on Object-S2HGD to compare with others. The two-stage methods initially complete the single-view point cloud (by PCN [54]), and generate human hand grasps on the completed point cloud using GraspCVAE (GC) [25] and our DiffuGrasp(DG)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.628 + ], + "angle": 0, + "content": "\"Grasp Displace\" metric, a larger penetration volume may result in better grasping stability (the larger the penetration, the greater the force on the fingers in the simulator), which could make the results of GraspCVAE seem favorable. However, an ideal grasp should have both smaller penetration volume and good stability. Therefore, we present the balance on these two metrics of both methods in Fig. 5. It is evident that the points representing our method are closer to the origin point, which also indicates that our method better unifies penetration with grasp stability, meaning the generated grasps are more natural and reasonable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.84 + ], + "angle": 0, + "content": "To generate more natural and stable grasps, instead of directly passing hand parameters into the decoder as done in [22], we first convert the hand parameters into point clouds and extract their features. We then feed the features into the decoder to predict the hand parameters. Our method incorporates more shape and positional information about the hand than direct decoding the hand parameters. To verify the effectiveness of our method, we conduct comparative experiments with the approach without the hand encoder used in [22], as shown in Table 2. The results show that, with a similar penetration volume, our method achieves higher grasp stability (smaller grasp displace), thereby effectively facilitating the generation of higher quality grasps and demonstrating the advantages of our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparing with two-stage methods. For the task of generating human grasps based on single-view scene point clouds, one intuitive approach is to first utilize point cloud completion methods to complete the single-view point" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "837" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.859, + 0.169 + ], + "angle": 0, + "content": "
MethodCVAEDiffuGraspGPpenetrationGrasp Displace Mean ± Variance(cm)↓Contact Ratio(%)↑
Depth(cm)↓Volume(cm3)↓
CVAE✓0.2313.543.10 ± 3.4998.07
DiffuGrasp✓0.2011.543.06 ± 3.3398.80
DiffuGrasp + GP✓✓0.216.582.73 ± 3.1699.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.174, + 0.894, + 0.203 + ], + "angle": 0, + "content": "Table 4. Ablation results of the DiffuGrasp module and the Global Perception module(GP) on our View-S2HGD dataset. The role of DiffuGrasp is demonstrated through a comparison with CVAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.471, + 0.27 + ], + "angle": 0, + "content": "clouds. Subsequently, a generation method like GraspTTA [25] is employed to generate grasps on the completed point clouds. [46] and [31] utilize such two-stage methods to generate dexterous hand grasps in single-view scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.474, + 0.591 + ], + "angle": 0, + "content": "We first train a Point Completion Network (PCN) [54] to complete the input single-view point clouds into full objects with tabletop points. Then, we train GraspCVAE (GC) [25] and our DiffuGrasp (DG) to generate human grasps based on these completed results. The testing results are shown in Table 3. The results demonstrate that our method outperforms both two-stage methods by a large margin. The reason is that the two-stage method is not end-to-end, which leads to accumulated errors in the process, impacting the quality of generated grasps. In addition, most works of point cloud completion are not suitable for such a situation with scene points, which will lead to a decrease in the accuracy of point cloud completion. However, the two-stage method has very high requirements for the completion results of single-view point clouds. If the result is poor, the generated grasp can easily penetrate into the object or not make contact with it. In contrast, our end-to-end model effectively avoids such problems. Last, the overall models of two-stage methods tend to be larger, require longer and more tedious training processes, and are less straightforward to train compared to our end-to-end method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.602, + 0.233, + 0.619 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.687 + ], + "angle": 0, + "content": "We conduct an ablation study on our S2HGrasp using our View-S2HGD dataset to highlight the role of the Global Perception module and demonstrate the advantages of our DiffuGrasp module over CVAE, as shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Global Perception. The comparative results, as well as the visualizations of point cloud completion, highlight that this module helps the object encoder capture global object characteristics. This, in turn, assists the generation models in producing more natural, collision-free human grasps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Specifically, by comparing the results of \"DiffuGrasp\" and \"DiffuGrasp+GP\", we can find that with the Global Perception module, our model can effectively reduce the volume of hand-object penetration (from 11.54 to 6.58) and enhance the stability of the grasp. This also demonstrates that, with the aid of the Global Perception module, our model has a good perception of the overall shape of local objects, and can effectively prevent the generated hand from penetrating into the invisible areas of the object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.21, + 0.895, + 0.376 + ], + "angle": 0, + "content": "DiffuGrasp. The previous work GraspTTA [25] leverages CVAE to generate human grasps based on the full object model, but suffers from severe model collapse and can't generate diverse grasps poses. We believe that the generative capabilities of CVAE are not sufficient, especially for the more challenging task of generating human grasps based on single-view scene point clouds. Therefore, we conduct comparative experiments on our View-S2HGD dataset using our DiffuGrasp and CVAE (without TTA, as TTA cannot be applied to incomplete point clouds) to demonstrate the advantages of our DiffuGrasp in this task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Results show that even without the Global Perception module, the penetration volume of DiffuGrasp is smaller than that of CVAE (11.54 to 13.54). Moreover, DiffuGrasp achieves higher grasp stability than CVAE with smaller penetration, which also shows that the grasps generated by DiffuGrasp are more reasonable and natural. In short, despite the incompleteness of the object and the interference from numerous scene points, DiffuGrasp outperforms CVAE and can generate stable human grasps with minimal penetration into the object. Additional ablation studies and experiments are shown in the supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.62, + 0.575 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.894, + 0.812 + ], + "angle": 0, + "content": "In this work, we explore a new task of generating human grasps based on single-view scene point clouds rather than full object models. We propose S2HGrasp, along with a new synthetic dataset S2HGD, to address the problem of hand-object penetration caused by the incompleteness of object point clouds. We design a Global Perception module to globally perceive partial objects and a DiffuGrasp module to generate plausible and natural human grasps despite numerous scene points. Our S2HGrasp effectively overcomes the limitation of existing approaches and extends human grasp generation beyond full object models, benefiting the study of hand-object interactions. The experimental results demonstrate that S2HGrasp outperforms other methods and achieves satisfying performance and generalization capabilities across different datasets and unseen objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.813, + 0.892, + 0.888 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported partially by the National Key Research and Development Program of China(2023YFA1008503),NSFC(U21A20471,U1911401), Guangdong NSF Project (No. 2023B1515040025, 2020B1 515120085)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.513, + 0.956 + ], + "angle": 0, + "content": "838" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.446, + 0.17 + ], + "angle": 0, + "content": "[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.465, + 0.226 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.443, + 0.282 + ], + "angle": 0, + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.468, + 0.408 + ], + "angle": 0, + "content": "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.469, + 0.465 + ], + "angle": 0, + "content": "[6] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing hand-object interactions in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.466, + 0.534 + ], + "angle": 0, + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[8] Jiayi Chen, Mi Yan, Jiazhao Zhang, Yinzhen Xu, Xiaolong Li, Yijia Weng, Li Yi, Shuran Song, and He Wang. Tracking and reconstructing hand object interactions from point cloud sequences in the wild. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.426, + 0.661 + ], + "angle": 0, + "content": "[9] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.462, + 0.732 + ], + "angle": 0, + "content": "[10] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.458, + 0.788 + ], + "angle": 0, + "content": "[11] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. arXiv preprint arXiv:1911.01911, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.46, + 0.901 + ], + "angle": 0, + "content": "[13] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.88, + 0.12 + ], + "angle": 0, + "content": "In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.887, + 0.19 + ], + "angle": 0, + "content": "[14] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.872, + 0.274 + ], + "angle": 0, + "content": "[15] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.866, + 0.358 + ], + "angle": 0, + "content": "[16] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.879, + 0.444 + ], + "angle": 0, + "content": "[17] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.889, + 0.501 + ], + "angle": 0, + "content": "[18] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.856, + 0.545 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.875, + 0.601 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.868, + 0.658 + ], + "angle": 0, + "content": "[21] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.885, + 0.729 + ], + "angle": 0, + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.89, + 0.786 + ], + "angle": 0, + "content": "[23] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.888, + 0.83 + ], + "angle": 0, + "content": "[24] Wolfgang Hürst and Casper Van Wezel. Gesture-based interaction via finger tracking for mobile augmented reality. Multimedia Tools and Applications, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.891, + 0.899 + ], + "angle": 0, + "content": "[25] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.891, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "839" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.466, + 0.161 + ], + "angle": 0, + "content": "[26] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.468, + 0.218 + ], + "angle": 0, + "content": "[27] Mia Kokic, Danica Kragic, and Jeannette Bohg. Learning to estimate pose and shape of hand-held objects from rgb images. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.465, + 0.273 + ], + "angle": 0, + "content": "[28] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.468, + 0.328 + ], + "angle": 0, + "content": "[29] Yiming Li, Wei Wei, Daheng Li, Peng Wang, Wanyi Li, and Jun Zhong. Hgc-net: Deep anthropomorphic hand grasping in clutter. In 2022 International Conference on Robotics and Automation (ICRA), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.445, + 0.385 + ], + "angle": 0, + "content": "[30] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In Proceedings of the AAAI conference on artificial intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.449, + 0.426 + ], + "angle": 0, + "content": "[31] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.463, + 0.482 + ], + "angle": 0, + "content": "[32] Meysam Madadi, Sergio Escalera, Xavier Baró, and Jordi Gonzalez. End-to-end global to local cnn learning for hand pose recovery in depth data. arXiv preprint arXiv:1705.09606, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.445, + 0.525 + ], + "angle": 0, + "content": "[33] Markus Oberweger, Paul Wohlhart, and Vincent Lepetit. Hands deep in deep learning for hand pose estimation. arXiv preprint arXiv:1502.06807, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.463, + 0.566 + ], + "angle": 0, + "content": "[34] Liang Pan. Ecg: Edge-aware point cloud completion with graph convolution. IEEE Robotics and Automation Letters, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.468, + 0.622 + ], + "angle": 0, + "content": "[35] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems. 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.678 + ], + "angle": 0, + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.468, + 0.747 + ], + "angle": 0, + "content": "[37] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.455, + 0.817 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.859 + ], + "angle": 0, + "content": "[39] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.421, + 0.9 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.885, + 0.161 + ], + "angle": 0, + "content": "[41] Srinath Sridhar, Anna Maria Feit, Christian Theobalt, and Antti Oulasvirta. Investigating the dexterity of multi-finger input for mid-air text entry. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.885, + 0.233 + ], + "angle": 0, + "content": "[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.885, + 0.29 + ], + "angle": 0, + "content": "[43] Bugra Tekin, Federica Bogo, and Marc Pollefeys. H+ o: Unified egocentric recognition of 3d hand-object poses and interactions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.853, + 0.374 + ], + "angle": 0, + "content": "[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. arXiv preprint arXiv:2304.00464, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[45] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.887, + 0.503 + ], + "angle": 0, + "content": "[46] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.871, + 0.572 + ], + "angle": 0, + "content": "[47] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.888, + 0.631 + ], + "angle": 0, + "content": "[48] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.887, + 0.73 + ], + "angle": 0, + "content": "[49] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.885, + 0.799 + ], + "angle": 0, + "content": "[50] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.885, + 0.871 + ], + "angle": 0, + "content": "[51] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.887, + 0.901 + ], + "angle": 0, + "content": "[52] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "840" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.455, + 0.134 + ], + "angle": 0, + "content": "Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.452, + 0.203 + ], + "angle": 0, + "content": "[53] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Point: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.206, + 0.467, + 0.26 + ], + "angle": 0, + "content": "[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV), 2018. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.462, + 0.331 + ], + "angle": 0, + "content": "[55] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.456, + 0.386 + ], + "angle": 0, + "content": "[56] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.461, + 0.443 + ], + "angle": 0, + "content": "[57] Dian Zheng, Xiao-Ming Wu, Zuhao Liu, Jingke Meng, and Wei-shi Zheng. Diffuvolume: Diffusion model for volume based stereo matching. arXiv preprint arXiv:2308.15989, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.467, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.956 + ], + "angle": 0, + "content": "841" + } + ] +] \ No newline at end of file diff --git a/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_origin.pdf b/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bde5760fed0f0c1290c6868558c29900cf171002 --- /dev/null +++ b/2024/Single-View Scene Point Cloud Human Grasp Generation/4153e0a8-7264-4ade-83b0-cfe42df66a8b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df3b053c3811d97f294eaa23432de2a7a16f20c6a67e5b04c0f6522fc675c2b0 +size 1476006 diff --git a/2024/Single-View Scene Point Cloud Human Grasp Generation/full.md b/2024/Single-View Scene Point Cloud Human Grasp Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bf106e4595167ee117353ee1485aef5b6e6e49cd --- /dev/null +++ b/2024/Single-View Scene Point Cloud Human Grasp Generation/full.md @@ -0,0 +1,318 @@ +# Single-View Scene Point Cloud Human Grasp Generation + +Yan-Kang Wang1 Chengyi Xing2 Yi-Lin Wei1 Xiao-Ming Wu1 Wei-Shi Zheng1,3* +1School of Computer Science and Engineering, Sun Yat-sen University, China +2Stanford University, Stanford, CA + +3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +{wangyk65, weiylin5, wuxm65}@mail2.sysu.edu.cn chengyix@stanford.edu wszheng@ieee.org + +![](images/fcaeed68e163bb55fa933cc7f4dec2172dc5722bcef0cddc46aac51d90a76c6c.jpg) + +![](images/3c671fe8c73dcb014f03abdbc967077ad82c22a49d00357d97f96d1dcc2fe283.jpg) + +![](images/428381b2947458e925e9fe3c5ce57a437ba11b19d7565b6c31b7fd70872f0ed3.jpg) +Dataset Scene + +![](images/55f399e5732a9ab3693b12cc556f3f53f033005e758797426a3e6e2900ad752c.jpg) +Red hands (second row): GraspCVAE Yellow hands (third row): Ours +Figure 1. The scene of our dataset and comparison between our model and GraspCVAE [25]. The top-left image depicts the scene in our dataset, comprising a table and an object. Below it are images from four random viewpoints out of 36. The three rows of images on the right side represent, respectively, the single-view images, the generation results of GraspCVAE and our method. The green box indicates the area where the hand-object penetration occurs due to the lack of global perception ability of GraspCVAE. + +# Abstract + +In this work, we explore a novel task of generating human grasps based on single-view scene point clouds, which more accurately mirrors the typical real-world situation of observing objects from a single viewpoint. Due to the incompleteness of object point clouds and the presence of numerous scene points, the generated hand is prone to penetrating into the invisible parts of the object and the model is easily affected by scene points. Thus, we introduce S2HGrasp, a framework composed of two key modules: the Global Perception module that globally perceives partial object point clouds, and the DiffuGrasp module designed to generate high-quality human grasps based on complex inputs that include scene points. Additionally, we introduce S2HGD dataset, which comprises approximately 99,000 single-object single-view scene point clouds of 1,668 unique objects, each annotated with one human grasp. Our extensive experiments demonstrate that S2HGrasp can not only generate natural human grasps regardless of scene points, but also effectively prevent penetration between the + +hand and invisible parts of the object. Moreover, our model showcases strong generalization capability when applied to unseen objects. Our code and dataset are available at https://github.com/iSEE-Laboratory/S2HGrasp. + +# 1. Introduction + +Learning and capturing hand-object interactions stands as a foundation for understanding human behaviors and has many practical applications. From augmented and virtual reality [24] to the domains of robotic grasps, human-computer interaction [35, 41] and learning from human demonstrations [13, 15, 37, 45], the comprehension and replication of these interactions are of significance. + +In this work, we focus on understanding the interactions involved in generating 3D human grasps for objects. We recognize most existing grasp generation methods depend on full 3D object models [25, 26], yet in real-world scenarios, objects and scenes are commonly perceived from a single viewpoint, leading to incomplete object point clouds with potential scene noise. Consequently, existing methods find it challenging to capture the complete geometric fea + +tures of a partial object and generate human grasps based on such single-view point clouds, limiting their practical applicability. Clearly, a major challenge of this task is the incompleteness of the object. If the model cannot perceive the object's overall shape, the generated grasp is likely to penetrate into the invisible parts of the object. Furthermore, due to interference from scene points, there is a risk that the generated hand will collide with the scene elements. + +To address the challenges, we introduce S2HGrasp, a novel framework designed for generating human grasps on single-view scene point clouds. S2HGrasp incorporates a Global Perception module that facilitates the global understanding of partial object point clouds, and a DiffuGrasp module aiming at generating high-quality human grasps from complex inputs that contain numerous scene points. + +Specifically, the Global Perception module employs a multi-task learning paradigm, including single-view point cloud completion and classification tasks. This module helps to capture the object's global geometric features, and effectively prevents the generated hands from penetrating into the object's invisible parts. During testing, we can directly extract global features from the input single-view point clouds without the need of point cloud completion and classification tasks. Through feature alignment in latent space, our module can capture global features from the partial object in a single-stage process. Concurrently, the DiffuGrasp module is designed to generate human grasps and it adds noise to the normalized hand parameters when training. Then it conditions on the scene features and predicts the original hand parameters. In testing, the DiffuGrasp module progressively denoises a random noise with the condition of scene features, and predicts the final reasonable hand parameters. Leveraging its powerful conditional generation capabilities, the DiffuGrasp module is capable of generating high-quality grasps that are close to the object, even when the input includes scene points. + +In addition, we introduce a novel dataset named S2HGD, specifically for human grasp generation on single-view scene point clouds. Constructed using the BlenderProc [11] simulator and derived from OakInk dataset [51], S2HGD has a collection of 1,668 unique objects. Each object is individually placed on a tabletop setting. RGB-D images are captured across 36 distinct viewpoints, and corresponding single-view point clouds are derived from depth information. Overall, S2HGD comprises roughly 99,000 point clouds, each paired with a reasonable human grasp. + +Our experimental results indicate that our end-to-end approach S2HGrasp outperforms existing generation methods and two-stage methods without data preprocessing and TTA (test-time adaptation, which fails to work well in our task due to the incompleteness of the objects). Additionally, our model not only generates natural and plausible grasps but also exhibits the ability to generalize to unseen objects. + +In summary, our main contributions are as follows: 1) We explore a new task of human grasp generation on single-view scene point clouds and achieve favorable results. 2) We propose a novel model named S2HGrasp, which employs a Global Perception module to endow the model with the ability to globally perceive single-view point clouds, and a DiffuGrasp module to generate high-quality grasps despite the complex input that contains the object and scene points. 3) We construct S2HGD, a new dataset for human grasp generation based on single-object single-view scene point clouds. Our code and dataset will both be available. + +# 2. Related Work + +Hand-object interaction and grasp. Hand-object interaction is currently receiving increasing attention and involves two main tasks: 1) reconstructing 3D models of hands and objects based on images [6, 12, 16-18, 27, 32, 33, 43, 50] and 2) predicting how to grasp an object based on its visual representation (images or point clouds) [4, 8, 10, 25, 26, 29, 42, 52]. GraspTTA [25] can generate human grasps based on the consistency between hand contact points and object contact regions given complete objects. Affordance Diffusion [52] employs diffusion model to generate hand-object interacting images given object images. When it comes to dexterous hand grasp, [49] and [44] work well using generative models and reinforcement learning, and [29] use a single-shot network to predict dexterous hand grasp in clutter. Existing works mainly focus on generating grasps for complete objects, however, in real world, objects are often observed from a single viewpoint. Thus we study human grasp generation on single-view scene point clouds, employing two modules to tackle the challenges of incompleteness of the object and impact of scene elements. + +3D point cloud completion. 3D point cloud completion enables the model to have global perception capability and has various methods, including point-based [30, 47, 54], view-based [21, 55], convolution-based [48], graph-based [34], generative model-based [23], transformer-based [53] approaches, etc. Point Completion Network (PCN) [54] is a learning-based shape completion method, and its decoder designed for point cloud features does not require any assumptions about the shape or category of the point clouds. It can generate both coarse and fine-grained completions with a relatively small number of parameters. Motivated by PCN [54], we design a Global Shape Perception (GSP) to complete partial point clouds. Compared to other SOTA methods for point cloud completion, this approach allows for the presence of scene points and achieves a satisfactory completion result, giving the model global awareness. + +Diffusion model. Diffusion models have emerged as the new deep generative models that progressively destruct data by injecting noise, then learn to reverse this process for sam + +![](images/cb99ead72145e7cc8fb6ffdc8dac14fdabbdb9ee8038ffa5935736a164d2273c.jpg) +Figure 2. S2HGrasp framework. The scene encoder takes single-view scene point clouds as input and extracts their features through PointNet++ [36] and a transformer block. The features are then used in point cloud completion (GSP), classification (GCP), and grasp generation (DiffuGrasp). GSP and GCP won't be used in testing. In the DiffuGrasp Training, the model adds noise to the normalized hand parameters and extracts hand features after passing the parameters into the MANO layer [39]. Then the object features and hand features will be fed into a transformer decoder to predict the original hand parameters. When testing, the DiffuGrasp Sampling starts from a random noise and iteratively denoises it, resulting in final hand parameters. + +ple generation. They have been used in a variety of domains, like computer vision [3, 5, 20, 57], natural language processing [1, 28], temporal data modeling [9] and multimodal modeling [2]. Notably, [22] introduces a SceneDiffuser for 3D scene understanding, and it also generates dexterous hands by adding noise to hand parameters and denoising them. Differently, we convert noised hand parameters into point clouds, extract their features, and then decode them to get final grasps. This improves the naturalness of generated hands, as demonstrated later in our experiments. + +# 3. S2HGrasp + +# 3.1. Problem Statement and Method Overview + +We explore a novel task: generating physically plausible human grasps for objects based on single-view scene point clouds. We enable the model to globally perceive partial objects and generate high-quality grasps despite interference from scene points. The input is point clouds $\mathcal{P} \in \mathbb{R}^{N \times 3}$ (where $N$ is the number of points including tabletop and partial object), captured from a single viewpoint. The output is human grasp parameters $\mathcal{H} \in \mathbb{R}^{61}$ modeled via differentiable MANO layer [39], with shape parameter $\beta \in \mathbb{R}^{10}$ for size of the hand and pose parameter $\theta \in \mathbb{R}^{45}$ for joint angle of 15 joints, as well as $\mathcal{R} \in \mathbb{R}^3$ and $\mathcal{T} \in \mathbb{R}^3$ for rotation and translation of wrist joint respectively. Given the hand parameters, MANO layer will output the hand mesh with $\hat{\mathcal{M}} = (\hat{\mathcal{V}} \in \mathcal{R}^{778 \times 3}, \hat{\mathcal{F}})$ , where $\hat{\mathcal{V}}$ and $\hat{\mathcal{F}}$ denote hand vertices and faces. This task is more aligned with real-world scenarios and is beneficial for practical applications. + +The generated grasps should be not only natural and plausible, but also able to firmly hold objects in physics-based simulators. Due to the incompleteness of object point clouds, the model must possess the capability to globally perceive the object to prevent hands from penetrating into the invisible parts of the object. Moreover, the model needs to generate high-quality human grasps despite interference from scene points. Therefore, we propose a model named S2HGrasp, as summarized in Fig. 2, which consists of two modules: Global Perception module and DiffuGrasp module, to generate human grasps on single-view point clouds. + +S2HGrasp is an end-to-end framework. The Global Perception module equips the model with the ability to perceive global geometric features of partial objects, while the DiffuGrasp module is responsible for resisting interference from scene points and generating high-quality human grasps. In the training process, our model takes single-view point clouds and corresponding hand parameters as input and uses an encoder to extract scene features. The Global Perception module endows these features with the global perception of the object. The DiffuGrasp module conditions on these features, adds noise to hand parameters, and trains a decoder to recover the original hand parameters. In testing, DiffuGrasp starts by sampling a random noise and progressively denoises it to derive the final hand parameters. + +# 3.2. Global Perception + +The Global Perception module aims at enabling the model to perceive the global geometric characteristics of the ob + +ject from single-view point clouds. Thanks to the multitask learning paradigm in training, our model can directly extract the scene features that contain global perception of the partial object in a single-stage process when testing. The resulting scene features will be used in the DiffuGrasp module, making the generated human grasps more natural, avoiding penetration with invisible parts of the object. Our scene point cloud features $\mathcal{F}_s$ are primarily extracted using PointNet++ [36], and we utilize a transformer block similar to [56] to add attention mechanisms, making the model focus more on the local regions suitable for grasping. + +Firstly, we design a Global Shape Perception (GSP) to perform point cloud completion task on the extracted feature, motivated by PCN [54]. After getting the feature $\mathcal{F}_s$ of single-view scene point clouds, we pass it through our GSP and result in $\mathcal{N}$ points of the completed object. Subsequently, containing the global perception of the object, $\mathcal{F}_s$ can be better utilized in the human grasp generation. + +Secondly, we design a Global Category Perception (GCP) to perform point cloud classification task. Specifically, we feed the obtained scene feature $\mathcal{F}_s$ into a classification head to output the object category. Since the object is incomplete and the input contains both the object and scene points, this classification task is challenging. Nevertheless, we calculate the F1 score for the classification results of S2HGrasp trained on S2HGD, achieving a score of 0.96. It indicates that GCP performs well, enabling the model to focus more on the object. By predicting the object's category, it can better perceive the object's overall shape. + +Global Perception Loss. In the GSP of the Global Perception module, we use a Chamfer Distance loss $L_{GSP}$ between the completed point clouds and the corresponding complete object point clouds for supervision, enabling it to have the capability to perceive the global feature of the object. In the GCP, we use Cross Entropy Loss $L_{GCP}$ . The total loss of the Global Perception module is defined as, + +$$ +L _ {G l o} = \lambda_ {G S P} \cdot L _ {G S P} + \lambda_ {G C P} \cdot L _ {G C P}. \tag {1} +$$ + +# 3.3. DiffuGrasp + +Since the input for our task is single-view scene point clouds including the tabletop and the object, the presence of numerous scene points can affect the quality of the generated hand and may lead to collision with the scene points. Therefore, we design the DiffuGrasp module with two innovations, to generate high-quality and diverse human grasps from complex input. Besides, we also design a plane loss to avoid collision between the hand and the tabletop. + +First, we add noise to hand parameters, but direct noise injection is not feasible due to the differences in the ranges and meanings of the parameters. Therefore, we perform specific normalization for each parameter based on its meaning and range. Second, hand parameters include hand shape, pose, and position information, and directly decod + +ing parameters does not effectively model these information. Thus, we pass them through MANO layer [39] to convert them into point clouds and use hand encoder to obtain their features before feeding them into the decoder. + +Following the practices of conditional diffusion model [20, 38], in our work, the DiffuGrasp module is with condition $\mathcal{F}_s$ , i.e., the feature of single-view point clouds. The DiffuGrasp module can progressively generate human grasps that are close to the object and avoid penetration with the object and the scene during the denoising process. + +- Training Process. Conditioning on the single-view scene point cloud feature $\mathcal{F}_s$ , the DiffuGrasp module adds noise to the normalized hand parameters $\mathcal{H} \in \mathbb{R}^{61}$ . After getting the hand feature $\mathcal{F}_h$ , we feed it together with $\mathcal{F}_s$ into a transformer decoder and train the decoder at various noise levels to predict the original hand parameters, i.e., denoise $\mathcal{H}_0$ for different $\mathcal{H}_t$ given $t$ , shown as below: + +$$ +\mathcal {H} _ {t} = \sqrt {1 - \bar {\alpha} _ {t}} \mathcal {H} _ {0} + \sqrt {\bar {\alpha} _ {t}} \cdot \epsilon , \mathcal {H} _ {0} ^ {\prime} = \operatorname {M o d e l} \left(\mathcal {H} _ {t}\right), \tag {2} +$$ + +where $\epsilon$ is noise, $\alpha$ is noise coefficient and $\overline{\alpha}_t = \prod_{i=1}^t \alpha_i$ . + +- DiffuGrasp Loss. Referring to the codes in DDPM [19], calculating the loss on the input and calculating the loss on the noise are equivalent, so we use Reconstruction Loss along with Cmap Loss, Penetration Loss and Plane Loss to train our transformer decoder. Specifically, 1) Reconstruction Loss of hand parameters and hand point clouds $\mathcal{P}^h\in \mathcal{R}^{778\times 3}$ is used to minimize reconstruction error, + +$$ +L _ {p a r a m} = \left| \mathcal {H} _ {0} ^ {\prime} - \mathcal {H} _ {0} \right|, L _ {\mathcal {V}} = \left| \left| \hat {\mathcal {V}} - \mathcal {P} ^ {h} \right| \right| _ {2} ^ {2}. \tag {3} +$$ + +2) Cmap Loss is used to ensure the generated hand is as close to the object as possible. We compute the distance from the hand's contact points $\mathcal{V}^p$ to the object's point clouds. For each point $\mathcal{P}_i^o$ , the distance $\mathbf{D}(\mathcal{P}_i^o) = \min_j ||\mathcal{V}_j^p - \mathcal{P}_i^o||_2^2$ . If the distance is less than a threshold, we consider the point as a potentially contactable point on the object, i.e., contact map. One of our training objectives is to make the contact points on the hand as close as possible to the contact map on the object as, + +$$ +L _ {c m a p} = \sum_ {i} \mathbf {D} \left(\mathcal {P} _ {i} ^ {o}\right), f o r a l l \mathbf {D} \left(\mathcal {P} _ {i} ^ {o}\right) \leq \mathcal {T}. \tag {4} +$$ + +3) Penetration Loss is used to penalize cases where the hand penetrates into the object. Specifically, we define the subset of object points that penetrate into the hand's mesh as $\mathcal{P}_{in}^{o}$ , and the penetration loss is defined as minimizing the distances from these points to their nearest points on the hand, + +$$ +L _ {p e n e t r} = \frac {1}{\left| \mathcal {P} _ {i n} ^ {o} \right|} \sum_ {p \in \mathcal {P} _ {i n} ^ {o}} \min _ {i} \| p - \hat {\mathcal {V}} _ {i} \| _ {2} ^ {2}. \tag {5} +$$ + +Note that we don't use the complete object to calculate the Cmap loss and the Penetration loss, and it is only used in the GSP module. Instead, we combine the input single-view point clouds $\mathcal{P}^s$ with the corresponding completion result $\mathcal{P}^c$ from the GSP for loss calculations. Specifically, we first calculate the minimum distance $\mathbf{D}(\mathcal{P}_i^c) = \min_j||\mathcal{P}_j^s -$ + +$\mathcal{P}_i^c ||_2^2$ , and keep the points $\hat{\mathcal{P}}^c = \{\mathcal{P}_i^c\in \mathcal{P}^c\mid \mathbf{D}(\mathcal{P}_i^c)\geq \theta \}$ We get the final completion result $\mathcal{P}^C = \hat{\mathcal{P}}^c\cup \mathcal{P}^s$ , which is used for loss calculation. This approach prevents large errors from directly using rough completion results to compute loss functions. It also allows the model to have a global perception of the object's shape and avoids generating hands that penetrate into the unseen parts of the object. 4) Plane Loss. As for generating human grasps based on scene point clouds, one of the challenges is to avoid contact between the hand and the tabletop point clouds. Thus, we design a plane loss to penalize cases where there is penetration between the tabletop point clouds and the hand. To avoid the enormous computational cost of calculating distances between tabletop point clouds and hand vertices, we regress the parameters of the tabletop plane. We define the hand points which are on the backside of the plane as $\hat{\mathcal{V}}_{back}$ and calculate the distance $\mathbf{D}(\hat{\mathcal{V}}_i,P)$ from each point $\hat{\mathcal{V}}_i$ in $\hat{\mathcal{V}}_{back}$ to the plane $P$ , and the plane loss is defined as, + +$$ +L _ {p l a n e} = \sum_ {i} \mathbf {D} (\hat {\mathcal {V}} _ {i}, P), f o r a l l \hat {\mathcal {V}} _ {i} \in \hat {\mathcal {V}} _ {b a c k}. \tag {6} +$$ + +Thus, the loss for the DiffuGrasp is: + +$$ +L _ {D i f f} = \lambda_ {p a r a m} \cdot L _ {p a r a m} + \lambda_ {\mathcal {V}} \cdot L _ {\mathcal {V}} + +$$ + +$$ +\lambda_ {c} \cdot L _ {c m a p} + \lambda_ {p e} \cdot L _ {p e n e t r} + \lambda_ {p l} \cdot L _ {p l a n e} +$$ + +- Sampling Process. The Sampling Process begins by randomly sampling a noise from standard Gaussian distribution. Subsequently, it undergoes a denoising process for $T$ timesteps with the feature of single-view point clouds serving as a condition, and ultimately yields the final hand parameters. The normalization follows the same procedure as described in the Training Process. We add noise calculated from $\mathcal{H}_t$ and $t$ to hand parameters $\mathcal{H}_t$ , and then feed them into the Model, which includes the MANO layer, Hand Encoder, and the Transformer Decoder as shown in Fig. 2, to get $\mathcal{H}_{t - 1}$ . After $T$ iterations of the denoising process, the model will get final $\mathcal{H}_0'$ . The distribution $p_{\theta}\left(\mathcal{H}_0|\mathcal{F}_s\right)$ represents our sampling process and can be shown as, + +$$ +p _ {\theta} \left(\mathcal {H} _ {0} \mid \mathcal {F} _ {s}\right) = p \left(\mathcal {H} _ {T}\right) \prod_ {t = 1} ^ {T} p \left(\mathcal {H} _ {t - 1} \mid \mathcal {H} _ {t}, \mathcal {F} _ {s}\right), \tag {8} +$$ + +$$ +p \left(\mathcal {H} _ {t - 1} \mid \mathcal {H} _ {t}, \mathcal {F} _ {s}\right) = \mathcal {N} \left(\mathcal {H} _ {t - 1}; \mu_ {\theta} \left(\mathcal {H} _ {t}, t, \mathcal {F} _ {s}\right), \sum_ {\theta} \left(\mathcal {H} _ {t}, t, \mathcal {F} _ {s}\right)\right) \tag {9} +$$ + +$$ +\mathcal {H} _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \mathcal {H} _ {0} + \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \cdot \epsilon_ {\theta} ^ {(t)} (\mathcal {H} _ {t}) + \sigma_ {t} \epsilon_ {t}, \tag {10} +$$ + +where $\sigma_{t} = \eta \sqrt{\frac{1 - \overline{\alpha}_{t - 1}}{1 - \overline{\alpha}_{t}}\cdot\left(1 - \frac{\overline{\alpha}_{t}}{\overline{\alpha}_{t - 1}}\right)}$ $\eta$ is DDIM [40] sampling coefficient and $\epsilon_t\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ is standard Gaussian noise. For more details, please refer to supplementary file. + +# 4. S2HGD Dataset + +To better align with real-world scenarios where objects are typically observed from a single viewpoint and to fur + +ther explore generating human grasps based on single-view scene point clouds, we construct the S2HGD dataset. Here we show how we build our S2HGD with BlenderProc [11]: + +Data Source. Our raw data (including objects and human grasp annotations) all come from the OakInk dataset [51]. We select a total of 1,667 objects from 35 different categories (e.g. bottle, knife, mug) as the dataset's objects. + +Simulation Environment Setting. In the simulation environment, we select a table from ShapeNet [7] and position 36 cameras above the table, arranged in three concentric circular arrays at heights of 0.25 meters, 0.45 meters, and 0.65 meters from the table respectively. Each circular array contains 12 cameras at the same height, with a 30-degree spacing between each camera, as shown in Fig. 1. + +Data Processing. We drop an object from $20~cm$ above the table, allowing it to fall onto the tabletop freely. This is repeated ten times to get ten different poses. For each pose, we capture an RGBD image from each camera and convert it into single-view point clouds. After collecting all point clouds, we test all grasps of each object, discard any that contact the table and remove scenes lacking collision-free grasp annotations. From the remaining grasps, we choose the one closest to the object as the final annotation. + +To enable our model to generate high-quality grasps for both varying viewpoints of seen objects and entirely unseen objects, we partition our dataset in two unique ways: ViewS2HGD focuses on varying viewpoints, whereas ObjectS2HGD is based on object diversity. Specifically, ViewS2HGD includes different single-view point clouds of the same objects in training and testing sets. Object-S2HGD ensures that the objects in the testing set are entirely absent from the training set, allowing us to assess the model's capacity for generalization to unseen objects. + +In total, S2HGD contains 1,667 objects and about 99,000 single-view scene point clouds with one annotation each. + +# 5. Experiments + +# 5.1. Datasets and Experiment Setup + +Implementation Details. The single-view point clouds in our dataset consist of 2,000 points, with 1,000 points belonging to the object and 1,000 points to the tabletop. In training, we use Adam optimizer and $\mathrm{LR} = 1e - 4$ , where the LR is reduced by half when the model trained 100, 160 and 180 epochs. We totally train for 200 epochs and the batch size is 60. Before training the entire model with View-S2HGD and Object-S2HGD, we pre-train the GSP module using single-view point clouds and their corresponding complete objects from each dataset respectively for 200 epochs. The loss weights we use are $\lambda_{GSP} = 5$ , $\lambda_{GCP} = 2$ , $\lambda_{param} = 300$ , $\lambda_V = 15$ , $\lambda_{cmap} = 300$ , $\lambda_{pe} = 15$ , $\lambda_{pl} = 1$ for View-S2HGD. For Object-S2HGD, only $\lambda_{pe}$ + +![](images/b3ef22b32e59e048b5413b6a3393056369174ad3707a76fc0b249f1e62470246.jpg) +Figure 3. The visualization of generated grasps of our S2HGrasp on our two datasets. The blue objects on the left represent the results of View-S2HGD, while the red objects on the right represent the results of Object-S2HGD. + +
View-S2HGDObject-S2HGDHO3D-objectObman-object
GTGC [25]OursGTGC [25]OursGC [25]OursGC [25]Ours
PenetrationDepth(cm)↓0.270.230.210.260.250.212.111.380.320.21
Volume(cm3)↓6.6013.546.587.2214.485.588.664.879.844.62
Grasp DisplaceMean(cm)↓2.013.102.732.183.693.263.613.304.183.25
Variance(cm)↓±2.73±3.49±3.16±2.69±3.94±3.38±3.93±3.73±4.31±3.75
Perceptual Score{1,...,5}↑4.542.603.254.522.543.142.112.892.043.13
ContactRatio(%)↑10098.0799.4110097.8798.6797.6299.0685.8296.06
Diversity(σ2)Axis(10-2)↑2.150.851.032.290.520.710.500.790.560.70
Angle(10-2)↑8.863.294.519.682.042.751.321.952.282.67
+ +Table 1. Quantitative results compared with GraspCVAE [25] on S2HGD. The objects in HO3D and Obman datasets are used to demonstrate the generalization capability. The Penetration Volume metric and the Grasp displace metric are our main indicators. The results show that our method outperforms GraspCVAE on all indicators (in bold) and is significantly effective in reducing hand-object penetration caused by incomplete point clouds. The Penetration results of our method are even better than GT (with a blue background). + +and $\lambda_{\text{param}}$ differ, being set to 20 and 250, respectively. + +Datasets. We use S2HGD, Obman [16] and HO3D [14] for our experiments. S2HGD is the dataset we mainly conduct our experiments on. We train and test our model on both View-S2HGD and Object-S2HGD to demonstrate the overall performance. We also conduct ablation experiments on View-S2HGD. Obman and HO3D are used for studying hand-object interaction and Obman is synthetic, while HO3D is real. To validate generalization capability of our method on unseen objects, we utilize the Object-S2HGD for training and testing. We also select 30 objects from Obman and 10 objects from HO3D to construct two testing sets using the same methodology employed in the construction of S2HGD, to further assess generalization capability. + +Evaluation Metrics. The primary goal of our model is to generate plausible and natural human grasps for objects based on single-view scene point clouds. Therefore, our main evaluation metrics involve calculating the penetration depth and volume between the object and the hand. We also use a physics-based simulation environment to calculate the grasp displace to assess the stability of the hand grasping the object. Perceptual score is used to evaluate the natural + +ness of the generated grasps and we invite 50 participants to rate the naturalness of generated grasps, using a scale from 1 to 5, where higher scores indicate higher naturalness. What's more, to demonstrate the contact between the generated hand and the object, we also calculate contact ratio between the hand and the object. Besides, to demonstrate diversity of generated grasp poses, we introduce a diversity metric. Specifically, we calculate the variance of the rotation axes $\sigma_{axis}^{2}$ and angles $\sigma_{angle}^{2}$ for 15 joints, excluding the wrist joint, across all grasp samples. Details about the evaluation metrics can be found in the supplementary file. + +# 5.2. Grasp Generation Performance + +Visualization results. We visualize the generated grasps on both the View-S2HGD and Object-S2HGD datasets, as shown in Fig. 3. The results demonstrate that our model can generate natural and plausible human grasps for different single-view point clouds of seen objects and shows a notable level of generalization for unseen objects. We also visualize the performance of our GSP module, i.e. the partial object point cloud completion results, as shown in the left of Fig. 4. The first row is the input single-view point + +![](images/2b8cf197749997a808823859b50c18474e1d99bdf57a32b0ca22791d73da4f9e.jpg) +Figure 4. Visualization of point cloud completion and our failure cases. Left: Point cloud completion results, with the input single-view point clouds in the top row and completion results below (gray points for the tabletop, red for the object). Right: Failure cases, with single-view images at the top and corresponding failure cases of our method below. + +![](images/3d5ef548b3afac0cebfb3af2e2d9c14097ba91ecca5c8b978baaac7e32d2a7ee.jpg) + +![](images/edd4a9e1e06fad87ed0e661bf8d403d0ebdb8ad038083057c9321a0f407a9fc2.jpg) + +![](images/b9aa40d49eea5b1591ef1577eb63665668aa426c3be0d763cfb27e950ee92397.jpg) +Figure 5. The balance between penetration volume (x-axis) and grasp stability (y-axis). Method whose result point is closer to the origin is more effective. + +![](images/d3cd44a49d12c5ada88e333c53e1690b6d31d332b6cd488fb4025dc3314ad932.jpg) + +
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
w/o encoder [22]0.216.703.13±3.4098.49
w/ encoder(ours)0.216.582.73±3.1699.41
+ +clouds and the second row is the completion results. The output of our GSP module consists of only the red points. We also visualize some failure cases, shown in the right part of Fig. 4. The first example fails due to the complex shape of the object and it's difficult to perceive its global shape from only partial point clouds, leading to the generated hand penetrating into the object. The last two examples fail because only a small part of the object is visible from the viewpoint. There are too few points to generate a reasonable grasp. More visualization results are in the supplementary file, including the HO3D and Obman datasets. + +Comparing with GraspCVAE. As shown in Table 1, we conduct experiments on our View-S2HGD and Object-S2HGD datasets and the testing sets we construct using the objects selected from HO3D [14] and Obman [16] datasets. We compare our model with the SOTA method GraspTTA [25] for the task of generating human grasps based on complete object point clouds. GraspTTA employs ContactNet for Test-Time Adaptation (TTA), which predicts contact areas with the condition of complete object point clouds. Since our task involves single-view scene point clouds as input, with partial object point clouds and interference from scene point clouds, TTA is challenging to apply. Thus, in the comparative experiments, we did not include the TTA module, only utilizing the GraspCVAE network (GC). + +The results indicate that our model outperforms GraspCVAE across all datasets. From the "Penetration" metric, it can be seen that GraspCVAE, due to its lack of global perception of single-view point clouds, produces grasps that are significantly penetrate into the object, while ours are much less, even better than GT. Additionally, our model also demonstrates superior generalization capability on unseen objects compared to GraspCVAE. Regarding the + +Table 2. Comparison with method [22] that without hand encoder. + +
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
PCN+GC0.2211.346.00±4.8385.66
PCN+DG0.207.686.50±4.5172.85
Ours0.215.583.26±3.3898.67
+ +Table 3. Comparison with two-stage methods. We use our model trained on Object-S2HGD to compare with others. The two-stage methods initially complete the single-view point cloud (by PCN [54]), and generate human hand grasps on the completed point cloud using GraspCVAE (GC) [25] and our DiffuGrasp(DG). + +"Grasp Displace" metric, a larger penetration volume may result in better grasping stability (the larger the penetration, the greater the force on the fingers in the simulator), which could make the results of GraspCVAE seem favorable. However, an ideal grasp should have both smaller penetration volume and good stability. Therefore, we present the balance on these two metrics of both methods in Fig. 5. It is evident that the points representing our method are closer to the origin point, which also indicates that our method better unifies penetration with grasp stability, meaning the generated grasps are more natural and reasonable. + +To generate more natural and stable grasps, instead of directly passing hand parameters into the decoder as done in [22], we first convert the hand parameters into point clouds and extract their features. We then feed the features into the decoder to predict the hand parameters. Our method incorporates more shape and positional information about the hand than direct decoding the hand parameters. To verify the effectiveness of our method, we conduct comparative experiments with the approach without the hand encoder used in [22], as shown in Table 2. The results show that, with a similar penetration volume, our method achieves higher grasp stability (smaller grasp displace), thereby effectively facilitating the generation of higher quality grasps and demonstrating the advantages of our approach. + +Comparing with two-stage methods. For the task of generating human grasps based on single-view scene point clouds, one intuitive approach is to first utilize point cloud completion methods to complete the single-view point + +
MethodCVAEDiffuGraspGPpenetrationGrasp Displace Mean ± Variance(cm)↓Contact Ratio(%)↑
Depth(cm)↓Volume(cm3)↓
CVAE✓0.2313.543.10 ± 3.4998.07
DiffuGrasp✓0.2011.543.06 ± 3.3398.80
DiffuGrasp + GP✓✓0.216.582.73 ± 3.1699.41
+ +Table 4. Ablation results of the DiffuGrasp module and the Global Perception module(GP) on our View-S2HGD dataset. The role of DiffuGrasp is demonstrated through a comparison with CVAE. + +clouds. Subsequently, a generation method like GraspTTA [25] is employed to generate grasps on the completed point clouds. [46] and [31] utilize such two-stage methods to generate dexterous hand grasps in single-view scenes. + +We first train a Point Completion Network (PCN) [54] to complete the input single-view point clouds into full objects with tabletop points. Then, we train GraspCVAE (GC) [25] and our DiffuGrasp (DG) to generate human grasps based on these completed results. The testing results are shown in Table 3. The results demonstrate that our method outperforms both two-stage methods by a large margin. The reason is that the two-stage method is not end-to-end, which leads to accumulated errors in the process, impacting the quality of generated grasps. In addition, most works of point cloud completion are not suitable for such a situation with scene points, which will lead to a decrease in the accuracy of point cloud completion. However, the two-stage method has very high requirements for the completion results of single-view point clouds. If the result is poor, the generated grasp can easily penetrate into the object or not make contact with it. In contrast, our end-to-end model effectively avoids such problems. Last, the overall models of two-stage methods tend to be larger, require longer and more tedious training processes, and are less straightforward to train compared to our end-to-end method. + +# 5.3. Ablation Study + +We conduct an ablation study on our S2HGrasp using our View-S2HGD dataset to highlight the role of the Global Perception module and demonstrate the advantages of our DiffuGrasp module over CVAE, as shown in Table 4. + +Global Perception. The comparative results, as well as the visualizations of point cloud completion, highlight that this module helps the object encoder capture global object characteristics. This, in turn, assists the generation models in producing more natural, collision-free human grasps. + +Specifically, by comparing the results of "DiffuGrasp" and "DiffuGrasp+GP", we can find that with the Global Perception module, our model can effectively reduce the volume of hand-object penetration (from 11.54 to 6.58) and enhance the stability of the grasp. This also demonstrates that, with the aid of the Global Perception module, our model has a good perception of the overall shape of local objects, and can effectively prevent the generated hand from penetrating into the invisible areas of the object. + +DiffuGrasp. The previous work GraspTTA [25] leverages CVAE to generate human grasps based on the full object model, but suffers from severe model collapse and can't generate diverse grasps poses. We believe that the generative capabilities of CVAE are not sufficient, especially for the more challenging task of generating human grasps based on single-view scene point clouds. Therefore, we conduct comparative experiments on our View-S2HGD dataset using our DiffuGrasp and CVAE (without TTA, as TTA cannot be applied to incomplete point clouds) to demonstrate the advantages of our DiffuGrasp in this task. + +Results show that even without the Global Perception module, the penetration volume of DiffuGrasp is smaller than that of CVAE (11.54 to 13.54). Moreover, DiffuGrasp achieves higher grasp stability than CVAE with smaller penetration, which also shows that the grasps generated by DiffuGrasp are more reasonable and natural. In short, despite the incompleteness of the object and the interference from numerous scene points, DiffuGrasp outperforms CVAE and can generate stable human grasps with minimal penetration into the object. Additional ablation studies and experiments are shown in the supplementary file. + +# 6. Conclusion + +In this work, we explore a new task of generating human grasps based on single-view scene point clouds rather than full object models. We propose S2HGrasp, along with a new synthetic dataset S2HGD, to address the problem of hand-object penetration caused by the incompleteness of object point clouds. We design a Global Perception module to globally perceive partial objects and a DiffuGrasp module to generate plausible and natural human grasps despite numerous scene points. Our S2HGrasp effectively overcomes the limitation of existing approaches and extends human grasp generation beyond full object models, benefiting the study of hand-object interactions. The experimental results demonstrate that S2HGrasp outperforms other methods and achieves satisfying performance and generalization capabilities across different datasets and unseen objects. + +Acknowledgments. This work was supported partially by the National Key Research and Development Program of China(2023YFA1008503),NSFC(U21A20471,U1911401), Guangdong NSF Project (No. 2023B1515040025, 2020B1 515120085). + +# References + +[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021. 3 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3 +[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2 +[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, 2020. 3 +[6] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing hand-object interactions in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5 +[8] Jiayi Chen, Mi Yan, Jiazhao Zhang, Yinzhen Xu, Xiaolong Li, Yijia Weng, Li Yi, Shuran Song, and He Wang. Tracking and reconstructing hand object interactions from point cloud sequences in the wild. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2 +[9] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020. 3 +[10] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2 +[11] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. arXiv preprint arXiv:1911.01911, 2019. 2, 5 +[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2 +[13] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning. + +In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1 +[14] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 6, 7 +[15] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), 2020. 1 +[16] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2, 6, 7 +[17] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. +[18] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), 2021. 2 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 2020. 4 +[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 3, 4 +[21] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2 +[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7 +[23] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2 +[24] Wolfgang Hürst and Casper Van Wezel. Gesture-based interaction via finger tracking for mobile augmented reality. Multimedia Tools and Applications, 2013. 1 +[25] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2, 6, 7, 8 + +[26] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), 2020. 1, 2 +[27] Mia Kokic, Danica Kragic, and Jeannette Bohg. Learning to estimate pose and shape of hand-held objects from rgb images. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2 +[28] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 2022. 3 +[29] Yiming Li, Wei Wei, Daheng Li, Peng Wang, Wanyi Li, and Jun Zhong. Hgc-net: Deep anthropomorphic hand grasping in clutter. In 2022 International Conference on Robotics and Automation (ICRA), 2022. 2 +[30] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In Proceedings of the AAAI conference on artificial intelligence, 2020. 2 +[31] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 2021. 8 +[32] Meysam Madadi, Sergio Escalera, Xavier Baró, and Jordi Gonzalez. End-to-end global to local cnn learning for hand pose recovery in depth data. arXiv preprint arXiv:1705.09606, 2017. 2 +[33] Markus Oberweger, Paul Wohlhart, and Vincent Lepetit. Hands deep in deep learning for hand pose estimation. arXiv preprint arXiv:1502.06807, 2015. 2 +[34] Liang Pan. Ecg: Edge-aware point cloud completion with graph convolution. IEEE Robotics and Automation Letters, 2020. 2 +[35] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems. 2013. 1 +[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 2017. 3, 4 +[37] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, 2022. 1 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 4 +[39] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 3, 4 +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 + +[41] Srinath Sridhar, Anna Maria Feit, Christian Theobalt, and Antti Oulasvirta. Investigating the dexterity of multi-finger input for mid-air text entry. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 2015. 1 +[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, 2020. 2 +[43] Bugra Tekin, Federica Bogo, and Marc Pollefeys. H+ o: Unified egocentric recognition of 3d hand-object poses and interactions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2 +[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. arXiv preprint arXiv:2304.00464, 2023. 2 +[45] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1 +[46] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 2022. 8 +[47] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2 +[48] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In European Conference on Computer Vision, 2020. 2 +[49] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[50] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[51] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 5 +[52] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei + +Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[53] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Point: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2 +[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV), 2018. 2, 4, 7, 8 +[55] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[56] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4 +[57] Dian Zheng, Xiao-Ming Wu, Zuhao Liu, Jingke Meng, and Wei-shi Zheng. Diffuvolume: Diffusion model for volume based stereo matching. arXiv preprint arXiv:2308.15989, 2023. 3 \ No newline at end of file diff --git a/2024/Single-View Scene Point Cloud Human Grasp Generation/images.zip b/2024/Single-View Scene Point Cloud Human Grasp Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..756433e9eb48ab7a1ee66df53e4c6cecdda60e30 --- /dev/null +++ b/2024/Single-View Scene Point Cloud Human Grasp Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2c9d4fd2dc280d19027987246d607cbf7d10ab0f79803ebeb2e93162ac51398 +size 480912 diff --git a/2024/Single-View Scene Point Cloud Human Grasp Generation/layout.json b/2024/Single-View Scene Point Cloud Human Grasp Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..61a550d5c251d079c5056eb20e2ce8ca696e1a01 --- /dev/null +++ b/2024/Single-View Scene Point Cloud Human Grasp Generation/layout.json @@ -0,0 +1,8048 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 475, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 475, + 121 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 475, + 121 + ], + "type": "text", + "content": "Single-View Scene Point Cloud Human Grasp Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 142, + 501, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 142, + 501, + 184 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 501, + 184 + ], + "type": "text", + "content": "Yan-Kang Wang1 Chengyi Xing2 Yi-Lin Wei1 Xiao-Ming Wu1 Wei-Shi Zheng1,3* \n1School of Computer Science and Engineering, Sun Yat-sen University, China \n2Stanford University, Stanford, CA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 185, + 534, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 185, + 534, + 200 + ], + "spans": [ + { + "bbox": [ + 59, + 185, + 534, + 200 + ], + "type": "text", + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 201, + 526, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 201, + 526, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 201, + 526, + 213 + ], + "type": "text", + "content": "{wangyk65, weiylin5, wuxm65}@mail2.sysu.edu.cn chengyix@stanford.edu wszheng@ieee.org" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 236, + 223, + 327 + ], + "blocks": [ + { + "bbox": [ + 50, + 236, + 223, + 327 + ], + "lines": [ + { + "bbox": [ + 50, + 236, + 223, + 327 + ], + "spans": [ + { + "bbox": [ + 50, + 236, + 223, + 327 + ], + "type": "image", + "image_path": "fcaeed68e163bb55fa933cc7f4dec2172dc5722bcef0cddc46aac51d90a76c6c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 221, + 236, + 545, + 327 + ], + "blocks": [ + { + "bbox": [ + 221, + 236, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 221, + 236, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 221, + 236, + 545, + 327 + ], + "type": "image", + "image_path": "3c671fe8c73dcb014f03abdbc967077ad82c22a49d00357d97f96d1dcc2fe283.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 49, + 328, + 223, + 373 + ], + "blocks": [ + { + "bbox": [ + 49, + 328, + 223, + 373 + ], + "lines": [ + { + "bbox": [ + 49, + 328, + 223, + 373 + ], + "spans": [ + { + "bbox": [ + 49, + 328, + 223, + 373 + ], + "type": "image", + "image_path": "428381b2947458e925e9fe3c5ce57a437ba11b19d7565b6c31b7fd70872f0ed3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 376, + 164, + 385 + ], + "lines": [ + { + "bbox": [ + 108, + 376, + 164, + 385 + ], + "spans": [ + { + "bbox": [ + 108, + 376, + 164, + 385 + ], + "type": "text", + "content": "Dataset Scene" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 328, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 223, + 328, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 223, + 328, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 223, + 328, + 545, + 373 + ], + "type": "image", + "image_path": "55f399e5732a9ab3693b12cc556f3f53f033005e758797426a3e6e2900ad752c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 376, + 529, + 387 + ], + "lines": [ + { + "bbox": [ + 247, + 376, + 529, + 387 + ], + "spans": [ + { + "bbox": [ + 247, + 376, + 529, + 387 + ], + "type": "text", + "content": "Red hands (second row): GraspCVAE Yellow hands (third row): Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 388, + 546, + 433 + ], + "lines": [ + { + "bbox": [ + 45, + 388, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 45, + 388, + 546, + 433 + ], + "type": "text", + "content": "Figure 1. The scene of our dataset and comparison between our model and GraspCVAE [25]. The top-left image depicts the scene in our dataset, comprising a table and an object. Below it are images from four random viewpoints out of 36. The three rows of images on the right side represent, respectively, the single-view images, the generation results of GraspCVAE and our method. The green box indicates the area where the hand-object penetration occurs due to the lack of global perception ability of GraspCVAE." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 443, + 192, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 443, + 192, + 455 + ], + "spans": [ + { + "bbox": [ + 143, + 443, + 192, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 468, + 290, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 290, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 290, + 696 + ], + "type": "text", + "content": "In this work, we explore a novel task of generating human grasps based on single-view scene point clouds, which more accurately mirrors the typical real-world situation of observing objects from a single viewpoint. Due to the incompleteness of object point clouds and the presence of numerous scene points, the generated hand is prone to penetrating into the invisible parts of the object and the model is easily affected by scene points. Thus, we introduce S2HGrasp, a framework composed of two key modules: the Global Perception module that globally perceives partial object point clouds, and the DiffuGrasp module designed to generate high-quality human grasps based on complex inputs that include scene points. Additionally, we introduce S2HGD dataset, which comprises approximately 99,000 single-object single-view scene point clouds of 1,668 unique objects, each annotated with one human grasp. Our extensive experiments demonstrate that S2HGrasp can not only generate natural human grasps regardless of scene points, but also effectively prevent penetration between the" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 443, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 546, + 491 + ], + "type": "text", + "content": "hand and invisible parts of the object. Moreover, our model showcases strong generalization capability when applied to unseen objects. Our code and dataset are available at https://github.com/iSEE-Laboratory/S2HGrasp." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 514, + 387, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 387, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 387, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 618 + ], + "type": "text", + "content": "Learning and capturing hand-object interactions stands as a foundation for understanding human behaviors and has many practical applications. From augmented and virtual reality [24] to the domains of robotic grasps, human-computer interaction [35, 41] and learning from human demonstrations [13, 15, 37, 45], the comprehension and replication of these interactions are of significance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 619, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 546, + 715 + ], + "type": "text", + "content": "In this work, we focus on understanding the interactions involved in generating 3D human grasps for objects. We recognize most existing grasp generation methods depend on full 3D object models [25, 26], yet in real-world scenarios, objects and scenes are commonly perceived from a single viewpoint, leading to incomplete object point clouds with potential scene noise. Consequently, existing methods find it challenging to capture the complete geometric fea" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "831" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "tures of a partial object and generate human grasps based on such single-view point clouds, limiting their practical applicability. Clearly, a major challenge of this task is the incompleteness of the object. If the model cannot perceive the object's overall shape, the generated grasp is likely to penetrate into the invisible parts of the object. Furthermore, due to interference from scene points, there is a risk that the generated hand will collide with the scene elements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 287, + 253 + ], + "type": "text", + "content": "To address the challenges, we introduce S2HGrasp, a novel framework designed for generating human grasps on single-view scene point clouds. S2HGrasp incorporates a Global Perception module that facilitates the global understanding of partial object point clouds, and a DiffuGrasp module aiming at generating high-quality human grasps from complex inputs that contain numerous scene points." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 288, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 506 + ], + "type": "text", + "content": "Specifically, the Global Perception module employs a multi-task learning paradigm, including single-view point cloud completion and classification tasks. This module helps to capture the object's global geometric features, and effectively prevents the generated hands from penetrating into the object's invisible parts. During testing, we can directly extract global features from the input single-view point clouds without the need of point cloud completion and classification tasks. Through feature alignment in latent space, our module can capture global features from the partial object in a single-stage process. Concurrently, the DiffuGrasp module is designed to generate human grasps and it adds noise to the normalized hand parameters when training. Then it conditions on the scene features and predicts the original hand parameters. In testing, the DiffuGrasp module progressively denoises a random noise with the condition of scene features, and predicts the final reasonable hand parameters. Leveraging its powerful conditional generation capabilities, the DiffuGrasp module is capable of generating high-quality grasps that are close to the object, even when the input includes scene points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "text", + "content": "In addition, we introduce a novel dataset named S2HGD, specifically for human grasp generation on single-view scene point clouds. Constructed using the BlenderProc [11] simulator and derived from OakInk dataset [51], S2HGD has a collection of 1,668 unique objects. Each object is individually placed on a tabletop setting. RGB-D images are captured across 36 distinct viewpoints, and corresponding single-view point clouds are derived from depth information. Overall, S2HGD comprises roughly 99,000 point clouds, each paired with a reasonable human grasp." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Our experimental results indicate that our end-to-end approach S2HGrasp outperforms existing generation methods and two-stage methods without data preprocessing and TTA (test-time adaptation, which fails to work well in our task due to the incompleteness of the objects). Additionally, our model not only generates natural and plausible grasps but also exhibits the ability to generalize to unseen objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "In summary, our main contributions are as follows: 1) We explore a new task of human grasp generation on single-view scene point clouds and achieve favorable results. 2) We propose a novel model named S2HGrasp, which employs a Global Perception module to endow the model with the ability to globally perceive single-view point clouds, and a DiffuGrasp module to generate high-quality grasps despite the complex input that contains the object and scene points. 3) We construct S2HGD, a new dataset for human grasp generation based on single-object single-view scene point clouds. Our code and dataset will both be available." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 217, + 392, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 392, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 392, + 228 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 236, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 476 + ], + "type": "text", + "content": "Hand-object interaction and grasp. Hand-object interaction is currently receiving increasing attention and involves two main tasks: 1) reconstructing 3D models of hands and objects based on images [6, 12, 16-18, 27, 32, 33, 43, 50] and 2) predicting how to grasp an object based on its visual representation (images or point clouds) [4, 8, 10, 25, 26, 29, 42, 52]. GraspTTA [25] can generate human grasps based on the consistency between hand contact points and object contact regions given complete objects. Affordance Diffusion [52] employs diffusion model to generate hand-object interacting images given object images. When it comes to dexterous hand grasp, [49] and [44] work well using generative models and reinforcement learning, and [29] use a single-shot network to predict dexterous hand grasp in clutter. Existing works mainly focus on generating grasps for complete objects, however, in real world, objects are often observed from a single viewpoint. Thus we study human grasp generation on single-view scene point clouds, employing two modules to tackle the challenges of incompleteness of the object and impact of scene elements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 480, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 673 + ], + "type": "text", + "content": "3D point cloud completion. 3D point cloud completion enables the model to have global perception capability and has various methods, including point-based [30, 47, 54], view-based [21, 55], convolution-based [48], graph-based [34], generative model-based [23], transformer-based [53] approaches, etc. Point Completion Network (PCN) [54] is a learning-based shape completion method, and its decoder designed for point cloud features does not require any assumptions about the shape or category of the point clouds. It can generate both coarse and fine-grained completions with a relatively small number of parameters. Motivated by PCN [54], we design a Global Shape Perception (GSP) to complete partial point clouds. Compared to other SOTA methods for point cloud completion, this approach allows for the presence of scene points and achieves a satisfactory completion result, giving the model global awareness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Diffusion model. Diffusion models have emerged as the new deep generative models that progressively destruct data by injecting noise, then learn to reverse this process for sam" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "832" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 273 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 273 + ], + "type": "image", + "image_path": "cb99ead72145e7cc8fb6ffdc8dac14fdabbdb9ee8038ffa5935736a164d2273c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 546, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 546, + 344 + ], + "type": "text", + "content": "Figure 2. S2HGrasp framework. The scene encoder takes single-view scene point clouds as input and extracts their features through PointNet++ [36] and a transformer block. The features are then used in point cloud completion (GSP), classification (GCP), and grasp generation (DiffuGrasp). GSP and GCP won't be used in testing. In the DiffuGrasp Training, the model adds noise to the normalized hand parameters and extracts hand features after passing the parameters into the MANO layer [39]. Then the object features and hand features will be fed into a transformer decoder to predict the original hand parameters. When testing, the DiffuGrasp Sampling starts from a random noise and iteratively denoises it, resulting in final hand parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 355, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 289, + 477 + ], + "type": "text", + "content": "ple generation. They have been used in a variety of domains, like computer vision [3, 5, 20, 57], natural language processing [1, 28], temporal data modeling [9] and multimodal modeling [2]. Notably, [22] introduces a SceneDiffuser for 3D scene understanding, and it also generates dexterous hands by adding noise to hand parameters and denoising them. Differently, we convert noised hand parameters into point clouds, extract their features, and then decode them to get final grasps. This improves the naturalness of generated hands, as demonstrated later in our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 483, + 117, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 117, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 117, + 498 + ], + "type": "text", + "content": "3. S2HGrasp" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 503, + 267, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 267, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 267, + 516 + ], + "type": "text", + "content": "3.1. Problem Statement and Method Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "We explore a novel task: generating physically plausible human grasps for objects based on single-view scene point clouds. We enable the model to globally perceive partial objects and generate high-quality grasps despite interference from scene points. The input is point clouds " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " is the number of points including tabletop and partial object), captured from a single viewpoint. The output is human grasp parameters " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\in \\mathbb{R}^{61}" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " modeled via differentiable MANO layer [39], with shape parameter " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " for size of the hand and pose parameter " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{45}" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " for joint angle of 15 joints, as well as " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " for rotation and translation of wrist joint respectively. Given the hand parameters, MANO layer will output the hand mesh with " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}} = (\\hat{\\mathcal{V}} \\in \\mathcal{R}^{778 \\times 3}, \\hat{\\mathcal{F}})" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " denote hand vertices and faces. This task is more aligned with real-world scenarios and is beneficial for practical applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 355, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 488 + ], + "type": "text", + "content": "The generated grasps should be not only natural and plausible, but also able to firmly hold objects in physics-based simulators. Due to the incompleteness of object point clouds, the model must possess the capability to globally perceive the object to prevent hands from penetrating into the invisible parts of the object. Moreover, the model needs to generate high-quality human grasps despite interference from scene points. Therefore, we propose a model named S2HGrasp, as summarized in Fig. 2, which consists of two modules: Global Perception module and DiffuGrasp module, to generate human grasps on single-view point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 488, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 546, + 657 + ], + "type": "text", + "content": "S2HGrasp is an end-to-end framework. The Global Perception module equips the model with the ability to perceive global geometric features of partial objects, while the DiffuGrasp module is responsible for resisting interference from scene points and generating high-quality human grasps. In the training process, our model takes single-view point clouds and corresponding hand parameters as input and uses an encoder to extract scene features. The Global Perception module endows these features with the global perception of the object. The DiffuGrasp module conditions on these features, adds noise to hand parameters, and trains a decoder to recover the original hand parameters. In testing, DiffuGrasp starts by sampling a random noise and progressively denoises it to derive the final hand parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 669, + 414, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 414, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 414, + 682 + ], + "type": "text", + "content": "3.2. Global Perception" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The Global Perception module aims at enabling the model to perceive the global geometric characteristics of the ob" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "833" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "text", + "content": "ject from single-view point clouds. Thanks to the multitask learning paradigm in training, our model can directly extract the scene features that contain global perception of the partial object in a single-stage process when testing. The resulting scene features will be used in the DiffuGrasp module, making the generated human grasps more natural, avoiding penetration with invisible parts of the object. Our scene point cloud features " + }, + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "text", + "content": " are primarily extracted using PointNet++ [36], and we utilize a transformer block similar to [56] to add attention mechanisms, making the model focus more on the local regions suitable for grasping." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "text", + "content": "Firstly, we design a Global Shape Perception (GSP) to perform point cloud completion task on the extracted feature, motivated by PCN [54]. After getting the feature " + }, + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "text", + "content": " of single-view scene point clouds, we pass it through our GSP and result in " + }, + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "text", + "content": " points of the completed object. Subsequently, containing the global perception of the object, " + }, + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 46, + 204, + 287, + 287 + ], + "type": "text", + "content": " can be better utilized in the human grasp generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 418 + ], + "type": "text", + "content": "Secondly, we design a Global Category Perception (GCP) to perform point cloud classification task. Specifically, we feed the obtained scene feature " + }, + { + "bbox": [ + 46, + 288, + 287, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 46, + 288, + 287, + 418 + ], + "type": "text", + "content": " into a classification head to output the object category. Since the object is incomplete and the input contains both the object and scene points, this classification task is challenging. Nevertheless, we calculate the F1 score for the classification results of S2HGrasp trained on S2HGD, achieving a score of 0.96. It indicates that GCP performs well, enabling the model to focus more on the object. By predicting the object's category, it can better perceive the object's overall shape." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "text", + "content": "Global Perception Loss. In the GSP of the Global Perception module, we use a Chamfer Distance loss " + }, + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "inline_equation", + "content": "L_{GSP}" + }, + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "text", + "content": " between the completed point clouds and the corresponding complete object point clouds for supervision, enabling it to have the capability to perceive the global feature of the object. In the GCP, we use Cross Entropy Loss " + }, + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "inline_equation", + "content": "L_{GCP}" + }, + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "text", + "content": ". The total loss of the Global Perception module is defined as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 506, + 287, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 506, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 84, + 506, + 287, + 518 + ], + "type": "interline_equation", + "content": "L _ {G l o} = \\lambda_ {G S P} \\cdot L _ {G S P} + \\lambda_ {G C P} \\cdot L _ {G C P}. \\tag {1}", + "image_path": "92d13b92ed445e1bc33fb2cf4281dadf2551c3e3265f8e31d2419efd426b0796.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 527, + 124, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 124, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 124, + 540 + ], + "type": "text", + "content": "3.3. DiffuGrasp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 640 + ], + "type": "text", + "content": "Since the input for our task is single-view scene point clouds including the tabletop and the object, the presence of numerous scene points can affect the quality of the generated hand and may lead to collision with the scene points. Therefore, we design the DiffuGrasp module with two innovations, to generate high-quality and diverse human grasps from complex input. Besides, we also design a plane loss to avoid collision between the hand and the tabletop." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "First, we add noise to hand parameters, but direct noise injection is not feasible due to the differences in the ranges and meanings of the parameters. Therefore, we perform specific normalization for each parameter based on its meaning and range. Second, hand parameters include hand shape, pose, and position information, and directly decod" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "ing parameters does not effectively model these information. Thus, we pass them through MANO layer [39] to convert them into point clouds and use hand encoder to obtain their features before feeding them into the decoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 122, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 545, + 194 + ], + "type": "text", + "content": "Following the practices of conditional diffusion model [20, 38], in our work, the DiffuGrasp module is with condition " + }, + { + "bbox": [ + 304, + 122, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 304, + 122, + 545, + 194 + ], + "type": "text", + "content": ", i.e., the feature of single-view point clouds. The DiffuGrasp module can progressively generate human grasps that are close to the object and avoid penetration with the object and the scene during the denoising process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": "- Training Process. Conditioning on the single-view scene point cloud feature " + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": ", the DiffuGrasp module adds noise to the normalized hand parameters " + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\in \\mathbb{R}^{61}" + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": ". After getting the hand feature " + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h" + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": ", we feed it together with " + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": " into a transformer decoder and train the decoder at various noise levels to predict the original hand parameters, i.e., denoise " + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_0" + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": " for different " + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 197, + 545, + 279 + ], + "type": "text", + "content": ", shown as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 282, + 545, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 282, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 320, + 282, + 545, + 296 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {t} = \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathcal {H} _ {0} + \\sqrt {\\bar {\\alpha} _ {t}} \\cdot \\epsilon , \\mathcal {H} _ {0} ^ {\\prime} = \\operatorname {M o d e l} \\left(\\mathcal {H} _ {t}\\right), \\tag {2}", + "image_path": "6c3b0db640c4293180a79483d12a98389e78900301ded75c2ab7aae7d8d65179.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "text", + "content": " is noise, " + }, + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "text", + "content": " is noise coefficient and " + }, + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_t = \\prod_{i=1}^t \\alpha_i" + }, + { + "bbox": [ + 304, + 298, + 539, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "type": "text", + "content": "- DiffuGrasp Loss. Referring to the codes in DDPM [19], calculating the loss on the input and calculating the loss on the noise are equivalent, so we use Reconstruction Loss along with Cmap Loss, Penetration Loss and Plane Loss to train our transformer decoder. Specifically, 1) Reconstruction Loss of hand parameters and hand point clouds " + }, + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^h\\in \\mathcal{R}^{778\\times 3}" + }, + { + "bbox": [ + 304, + 312, + 545, + 396 + ], + "type": "text", + "content": " is used to minimize reconstruction error," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 398, + 545, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 398, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 339, + 398, + 545, + 412 + ], + "type": "interline_equation", + "content": "L _ {p a r a m} = \\left| \\mathcal {H} _ {0} ^ {\\prime} - \\mathcal {H} _ {0} \\right|, L _ {\\mathcal {V}} = \\left| \\left| \\hat {\\mathcal {V}} - \\mathcal {P} ^ {h} \\right| \\right| _ {2} ^ {2}. \\tag {3}", + "image_path": "01622b134973e8010ccd071f3d0dc3b52e481068f03889123a58a105f8882760.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "text", + "content": "2) Cmap Loss is used to ensure the generated hand is as close to the object as possible. We compute the distance from the hand's contact points " + }, + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^p" + }, + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "text", + "content": " to the object's point clouds. For each point " + }, + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^o" + }, + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "text", + "content": ", the distance " + }, + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\mathcal{P}_i^o) = \\min_j ||\\mathcal{V}_j^p - \\mathcal{P}_i^o||_2^2" + }, + { + "bbox": [ + 304, + 416, + 545, + 524 + ], + "type": "text", + "content": ". If the distance is less than a threshold, we consider the point as a potentially contactable point on the object, i.e., contact map. One of our training objectives is to make the contact points on the hand as close as possible to the contact map on the object as," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 338, + 527, + 545, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 527, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 338, + 527, + 545, + 545 + ], + "type": "interline_equation", + "content": "L _ {c m a p} = \\sum_ {i} \\mathbf {D} \\left(\\mathcal {P} _ {i} ^ {o}\\right), f o r a l l \\mathbf {D} \\left(\\mathcal {P} _ {i} ^ {o}\\right) \\leq \\mathcal {T}. \\tag {4}", + "image_path": "1989f4e840c0a2f082640ca1ca140b1fd06671c55345ac0d2a60789a65854914.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "type": "text", + "content": "3) Penetration Loss is used to penalize cases where the hand penetrates into the object. Specifically, we define the subset of object points that penetrate into the hand's mesh as " + }, + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}^{o}" + }, + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "type": "text", + "content": ", and the penetration loss is defined as minimizing the distances from these points to their nearest points on the hand," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 344, + 607, + 545, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 607, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 344, + 607, + 545, + 637 + ], + "type": "interline_equation", + "content": "L _ {p e n e t r} = \\frac {1}{\\left| \\mathcal {P} _ {i n} ^ {o} \\right|} \\sum_ {p \\in \\mathcal {P} _ {i n} ^ {o}} \\min _ {i} \\| p - \\hat {\\mathcal {V}} _ {i} \\| _ {2} ^ {2}. \\tag {5}", + "image_path": "3049f00f24e30b9c6a92a85187e1a5f07c6566c06f2f12c87282651c09c10839.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "Note that we don't use the complete object to calculate the Cmap loss and the Penetration loss, and it is only used in the GSP module. Instead, we combine the input single-view point clouds " + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s" + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": " with the corresponding completion result " + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^c" + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": " from the GSP for loss calculations. Specifically, we first calculate the minimum distance " + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\mathcal{P}_i^c) = \\min_j||\\mathcal{P}_j^s -" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "834" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^c ||_2^2" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " , and keep the points " + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^c = \\{\\mathcal{P}_i^c\\in \\mathcal{P}^c\\mid \\mathbf{D}(\\mathcal{P}_i^c)\\geq \\theta \\}" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " We get the final completion result " + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^C = \\hat{\\mathcal{P}}^c\\cup \\mathcal{P}^s" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " , which is used for loss calculation. This approach prevents large errors from directly using rough completion results to compute loss functions. It also allows the model to have a global perception of the object's shape and avoids generating hands that penetrate into the unseen parts of the object. 4) Plane Loss. As for generating human grasps based on scene point clouds, one of the challenges is to avoid contact between the hand and the tabletop point clouds. Thus, we design a plane loss to penalize cases where there is penetration between the tabletop point clouds and the hand. To avoid the enormous computational cost of calculating distances between tabletop point clouds and hand vertices, we regress the parameters of the tabletop plane. We define the hand points which are on the backside of the plane as " + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_{back}" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " and calculate the distance " + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\hat{\\mathcal{V}}_i,P)" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " from each point " + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_i" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_{back}" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " to the plane " + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 71, + 289, + 289 + ], + "type": "text", + "content": " , and the plane loss is defined as," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 291, + 287, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 287, + 309 + ], + "type": "interline_equation", + "content": "L _ {p l a n e} = \\sum_ {i} \\mathbf {D} (\\hat {\\mathcal {V}} _ {i}, P), f o r a l l \\hat {\\mathcal {V}} _ {i} \\in \\hat {\\mathcal {V}} _ {b a c k}. \\tag {6}", + "image_path": "c527f962b050c757ef44f817825f9bb6104af3bc42bbe694d41de674f97823db.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 310, + 194, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 194, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 194, + 322 + ], + "type": "text", + "content": "Thus, the loss for the DiffuGrasp is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 324, + 252, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 324, + 252, + 338 + ], + "spans": [ + { + "bbox": [ + 89, + 324, + 252, + 338 + ], + "type": "interline_equation", + "content": "L _ {D i f f} = \\lambda_ {p a r a m} \\cdot L _ {p a r a m} + \\lambda_ {\\mathcal {V}} \\cdot L _ {\\mathcal {V}} +", + "image_path": "a57f7b08419fea76b6dc19bd3d5ebee8f60d14c0129fac00dfd37192933afddb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 338, + 252, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 338, + 252, + 353 + ], + "spans": [ + { + "bbox": [ + 83, + 338, + 252, + 353 + ], + "type": "interline_equation", + "content": "\\lambda_ {c} \\cdot L _ {c m a p} + \\lambda_ {p e} \\cdot L _ {p e n e t r} + \\lambda_ {p l} \\cdot L _ {p l a n e}", + "image_path": "9646cf095c8a60ed241b94e0ee10c7ca187ac2f1364d7eea11f90e57f56a94e0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": "- Sampling Process. The Sampling Process begins by randomly sampling a noise from standard Gaussian distribution. Subsequently, it undergoes a denoising process for " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": " timesteps with the feature of single-view point clouds serving as a condition, and ultimately yields the final hand parameters. The normalization follows the same procedure as described in the Training Process. We add noise calculated from " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": " to hand parameters " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": ", and then feed them into the Model, which includes the MANO layer, Hand Encoder, and the Transformer Decoder as shown in Fig. 2, to get " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{t - 1}" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": ". After " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": " iterations of the denoising process, the model will get final " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_0'" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": ". The distribution " + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "inline_equation", + "content": "p_{\\theta}\\left(\\mathcal{H}_0|\\mathcal{F}_s\\right)" + }, + { + "bbox": [ + 47, + 354, + 289, + 510 + ], + "type": "text", + "content": " represents our sampling process and can be shown as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 517, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 517, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 66, + 517, + 287, + 548 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathcal {H} _ {0} \\mid \\mathcal {F} _ {s}\\right) = p \\left(\\mathcal {H} _ {T}\\right) \\prod_ {t = 1} ^ {T} p \\left(\\mathcal {H} _ {t - 1} \\mid \\mathcal {H} _ {t}, \\mathcal {F} _ {s}\\right), \\tag {8}", + "image_path": "8f249e55ae3bc1f4575643204ff6830d18405a95f546c3581e169b407b48ead8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 552, + 289, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 289, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 289, + 587 + ], + "type": "interline_equation", + "content": "p \\left(\\mathcal {H} _ {t - 1} \\mid \\mathcal {H} _ {t}, \\mathcal {F} _ {s}\\right) = \\mathcal {N} \\left(\\mathcal {H} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathcal {H} _ {t}, t, \\mathcal {F} _ {s}\\right), \\sum_ {\\theta} \\left(\\mathcal {H} _ {t}, t, \\mathcal {F} _ {s}\\right)\\right) \\tag {9}", + "image_path": "f9843c55272a360e2d5492f5357764c50aaca5cfbcefc6fc71fb6b16379321eb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 588, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\mathcal {H} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\epsilon_ {\\theta} ^ {(t)} (\\mathcal {H} _ {t}) + \\sigma_ {t} \\epsilon_ {t}, \\tag {10}", + "image_path": "96974759d64456c30a9fbb62551d4cb359a3e64a92250bbce4ec234510bcffcf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\sigma_{t} = \\eta \\sqrt{\\frac{1 - \\overline{\\alpha}_{t - 1}}{1 - \\overline{\\alpha}_{t}}\\cdot\\left(1 - \\frac{\\overline{\\alpha}_{t}}{\\overline{\\alpha}_{t - 1}}\\right)}" + }, + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "type": "text", + "content": " is DDIM [40] sampling coefficient and " + }, + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\epsilon_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 617, + 287, + 660 + ], + "type": "text", + "content": " is standard Gaussian noise. For more details, please refer to supplementary file." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 669, + 144, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 144, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 144, + 681 + ], + "type": "text", + "content": "4. S2HGD Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "To better align with real-world scenarios where objects are typically observed from a single viewpoint and to fur" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "ther explore generating human grasps based on single-view scene point clouds, we construct the S2HGD dataset. Here we show how we build our S2HGD with BlenderProc [11]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 113, + 545, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 113, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 545, + 161 + ], + "type": "text", + "content": "Data Source. Our raw data (including objects and human grasp annotations) all come from the OakInk dataset [51]. We select a total of 1,667 objects from 35 different categories (e.g. bottle, knife, mug) as the dataset's objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 166, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 545, + 250 + ], + "type": "text", + "content": "Simulation Environment Setting. In the simulation environment, we select a table from ShapeNet [7] and position 36 cameras above the table, arranged in three concentric circular arrays at heights of 0.25 meters, 0.45 meters, and 0.65 meters from the table respectively. Each circular array contains 12 cameras at the same height, with a 30-degree spacing between each camera, as shown in Fig. 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 255, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 361 + ], + "type": "text", + "content": "Data Processing. We drop an object from " + }, + { + "bbox": [ + 304, + 255, + 545, + 361 + ], + "type": "inline_equation", + "content": "20~cm" + }, + { + "bbox": [ + 304, + 255, + 545, + 361 + ], + "type": "text", + "content": " above the table, allowing it to fall onto the tabletop freely. This is repeated ten times to get ten different poses. For each pose, we capture an RGBD image from each camera and convert it into single-view point clouds. After collecting all point clouds, we test all grasps of each object, discard any that contact the table and remove scenes lacking collision-free grasp annotations. From the remaining grasps, we choose the one closest to the object as the final annotation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 362, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 482 + ], + "type": "text", + "content": "To enable our model to generate high-quality grasps for both varying viewpoints of seen objects and entirely unseen objects, we partition our dataset in two unique ways: ViewS2HGD focuses on varying viewpoints, whereas ObjectS2HGD is based on object diversity. Specifically, ViewS2HGD includes different single-view point clouds of the same objects in training and testing sets. Object-S2HGD ensures that the objects in the testing set are entirely absent from the training set, allowing us to assess the model's capacity for generalization to unseen objects." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 483, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 507 + ], + "type": "text", + "content": "In total, S2HGD contains 1,667 objects and about 99,000 single-view scene point clouds with one annotation each." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 518, + 388, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 388, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 388, + 532 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 539, + 476, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 476, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 476, + 552 + ], + "type": "text", + "content": "5.1. Datasets and Experiment Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": "Implementation Details. The single-view point clouds in our dataset consist of 2,000 points, with 1,000 points belonging to the object and 1,000 points to the tabletop. In training, we use Adam optimizer and " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LR} = 1e - 4" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": ", where the LR is reduced by half when the model trained 100, 160 and 180 epochs. We totally train for 200 epochs and the batch size is 60. Before training the entire model with View-S2HGD and Object-S2HGD, we pre-train the GSP module using single-view point clouds and their corresponding complete objects from each dataset respectively for 200 epochs. The loss weights we use are " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{GSP} = 5" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{GCP} = 2" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{param} = 300" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_V = 15" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{cmap} = 300" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pe} = 15" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pl} = 1" + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "text", + "content": " for View-S2HGD. For Object-S2HGD, only " + }, + { + "bbox": [ + 304, + 557, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pe}" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "835" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 542, + 246 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 542, + 246 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 542, + 246 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 542, + 246 + ], + "type": "image", + "image_path": "b3ef22b32e59e048b5413b6a3393056369174ad3707a76fc0b249f1e62470246.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 273 + ], + "type": "text", + "content": "Figure 3. The visualization of generated grasps of our S2HGrasp on our two datasets. The blue objects on the left represent the results of View-S2HGD, while the red objects on the right represent the results of Object-S2HGD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 274, + 535, + 384 + ], + "blocks": [ + { + "bbox": [ + 59, + 274, + 535, + 384 + ], + "lines": [ + { + "bbox": [ + 59, + 274, + 535, + 384 + ], + "spans": [ + { + "bbox": [ + 59, + 274, + 535, + 384 + ], + "type": "table", + "html": "
View-S2HGDObject-S2HGDHO3D-objectObman-object
GTGC [25]OursGTGC [25]OursGC [25]OursGC [25]Ours
PenetrationDepth(cm)↓0.270.230.210.260.250.212.111.380.320.21
Volume(cm3)↓6.6013.546.587.2214.485.588.664.879.844.62
Grasp DisplaceMean(cm)↓2.013.102.732.183.693.263.613.304.183.25
Variance(cm)↓±2.73±3.49±3.16±2.69±3.94±3.38±3.93±3.73±4.31±3.75
Perceptual Score{1,...,5}↑4.542.603.254.522.543.142.112.892.043.13
ContactRatio(%)↑10098.0799.4110097.8798.6797.6299.0685.8296.06
Diversity(σ2)Axis(10-2)↑2.150.851.032.290.520.710.500.790.560.70
Angle(10-2)↑8.863.294.519.682.042.751.321.952.282.67
", + "image_path": "b2bc911b7c131fceecec1386072e8842b857be3ecd35effaefa8ac667fe42e96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 546, + 439 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 546, + 439 + ], + "type": "text", + "content": "Table 1. Quantitative results compared with GraspCVAE [25] on S2HGD. The objects in HO3D and Obman datasets are used to demonstrate the generalization capability. The Penetration Volume metric and the Grasp displace metric are our main indicators. The results show that our method outperforms GraspCVAE on all indicators (in bold) and is significantly effective in reducing hand-object penetration caused by incomplete point clouds. The Penetration results of our method are even better than GT (with a blue background)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 449, + 272, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 272, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 272, + 460 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 449, + 272, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{param}}" + }, + { + "bbox": [ + 46, + 449, + 272, + 460 + ], + "type": "text", + "content": " differ, being set to 20 and 250, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": "Datasets. We use S2HGD, Obman [16] and HO3D [14] for our experiments. S2HGD is the dataset we mainly conduct our experiments on. We train and test our model on both View-S2HGD and Object-S2HGD to demonstrate the overall performance. We also conduct ablation experiments on View-S2HGD. Obman and HO3D are used for studying hand-object interaction and Obman is synthetic, while HO3D is real. To validate generalization capability of our method on unseen objects, we utilize the Object-S2HGD for training and testing. We also select 30 objects from Obman and 10 objects from HO3D to construct two testing sets using the same methodology employed in the construction of S2HGD, to further assess generalization capability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. The primary goal of our model is to generate plausible and natural human grasps for objects based on single-view scene point clouds. Therefore, our main evaluation metrics involve calculating the penetration depth and volume between the object and the hand. We also use a physics-based simulation environment to calculate the grasp displace to assess the stability of the hand grasping the object. Perceptual score is used to evaluate the natural" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "content": "ness of the generated grasps and we invite 50 participants to rate the naturalness of generated grasps, using a scale from 1 to 5, where higher scores indicate higher naturalness. What's more, to demonstrate the contact between the generated hand and the object, we also calculate contact ratio between the hand and the object. Besides, to demonstrate diversity of generated grasp poses, we introduce a diversity metric. Specifically, we calculate the variance of the rotation axes " + }, + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\sigma_{axis}^{2}" + }, + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "content": " and angles " + }, + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\sigma_{angle}^{2}" + }, + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "content": " for 15 joints, excluding the wrist joint, across all grasp samples. Details about the evaluation metrics can be found in the supplementary file." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 586, + 477, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 477, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 477, + 600 + ], + "type": "text", + "content": "5.2. Grasp Generation Performance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Visualization results. We visualize the generated grasps on both the View-S2HGD and Object-S2HGD datasets, as shown in Fig. 3. The results demonstrate that our model can generate natural and plausible human grasps for different single-view point clouds of seen objects and shows a notable level of generalization for unseen objects. We also visualize the performance of our GSP module, i.e. the partial object point cloud completion results, as shown in the left of Fig. 4. The first row is the input single-view point" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "836" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 71, + 209, + 173 + ], + "blocks": [ + { + "bbox": [ + 54, + 71, + 209, + 173 + ], + "lines": [ + { + "bbox": [ + 54, + 71, + 209, + 173 + ], + "spans": [ + { + "bbox": [ + 54, + 71, + 209, + 173 + ], + "type": "image", + "image_path": "2b8cf197749997a808823859b50c18474e1d99bdf57a32b0ca22791d73da4f9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 182, + 373, + 226 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 373, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 373, + 226 + ], + "type": "text", + "content": "Figure 4. Visualization of point cloud completion and our failure cases. Left: Point cloud completion results, with the input single-view point clouds in the top row and completion results below (gray points for the tabletop, red for the object). Right: Failure cases, with single-view images at the top and corresponding failure cases of our method below." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 72, + 261, + 171 + ], + "blocks": [ + { + "bbox": [ + 211, + 72, + 261, + 171 + ], + "lines": [ + { + "bbox": [ + 211, + 72, + 261, + 171 + ], + "spans": [ + { + "bbox": [ + 211, + 72, + 261, + 171 + ], + "type": "image", + "image_path": "3d5ef548b3afac0cebfb3af2e2d9c14097ba91ecca5c8b978baaac7e32d2a7ee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 263, + 72, + 313, + 171 + ], + "blocks": [ + { + "bbox": [ + 263, + 72, + 313, + 171 + ], + "lines": [ + { + "bbox": [ + 263, + 72, + 313, + 171 + ], + "spans": [ + { + "bbox": [ + 263, + 72, + 313, + 171 + ], + "type": "image", + "image_path": "edd4a9e1e06fad87ed0e661bf8d403d0ebdb8ad038083057c9321a0f407a9fc2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 365, + 171 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 365, + 171 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 365, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 365, + 171 + ], + "type": "image", + "image_path": "b9aa40d49eea5b1591ef1577eb63665668aa426c3be0d763cfb27e950ee92397.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 182, + 542, + 226 + ], + "lines": [ + { + "bbox": [ + 380, + 182, + 542, + 226 + ], + "spans": [ + { + "bbox": [ + 380, + 182, + 542, + 226 + ], + "type": "text", + "content": "Figure 5. The balance between penetration volume (x-axis) and grasp stability (y-axis). Method whose result point is closer to the origin is more effective." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 384, + 77, + 539, + 171 + ], + "blocks": [ + { + "bbox": [ + 384, + 77, + 539, + 171 + ], + "lines": [ + { + "bbox": [ + 384, + 77, + 539, + 171 + ], + "spans": [ + { + "bbox": [ + 384, + 77, + 539, + 171 + ], + "type": "image", + "image_path": "d3cd44a49d12c5ada88e333c53e1690b6d31d332b6cd488fb4025dc3314ad932.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 236, + 286, + 289 + ], + "blocks": [ + { + "bbox": [ + 48, + 236, + 286, + 289 + ], + "lines": [ + { + "bbox": [ + 48, + 236, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 236, + 286, + 289 + ], + "type": "table", + "html": "
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
w/o encoder [22]0.216.703.13±3.4098.49
w/ encoder(ours)0.216.582.73±3.1699.41
", + "image_path": "b87a2fbe7bde0d4605ef21061ac41ed44e56f2f119a8b9e69b33a414609e29bb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 315, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 286, + 445 + ], + "type": "text", + "content": "clouds and the second row is the completion results. The output of our GSP module consists of only the red points. We also visualize some failure cases, shown in the right part of Fig. 4. The first example fails due to the complex shape of the object and it's difficult to perceive its global shape from only partial point clouds, leading to the generated hand penetrating into the object. The last two examples fail because only a small part of the object is visible from the viewpoint. There are too few points to generate a reasonable grasp. More visualization results are in the supplementary file, including the HO3D and Obman datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 448, + 286, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 286, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 286, + 615 + ], + "type": "text", + "content": "Comparing with GraspCVAE. As shown in Table 1, we conduct experiments on our View-S2HGD and Object-S2HGD datasets and the testing sets we construct using the objects selected from HO3D [14] and Obman [16] datasets. We compare our model with the SOTA method GraspTTA [25] for the task of generating human grasps based on complete object point clouds. GraspTTA employs ContactNet for Test-Time Adaptation (TTA), which predicts contact areas with the condition of complete object point clouds. Since our task involves single-view scene point clouds as input, with partial object point clouds and interference from scene point clouds, TTA is challenging to apply. Thus, in the comparative experiments, we did not include the TTA module, only utilizing the GraspCVAE network (GC)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "The results indicate that our model outperforms GraspCVAE across all datasets. From the \"Penetration\" metric, it can be seen that GraspCVAE, due to its lack of global perception of single-view point clouds, produces grasps that are significantly penetrate into the object, while ours are much less, even better than GT. Additionally, our model also demonstrates superior generalization capability on unseen objects compared to GraspCVAE. Regarding the" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 236, + 533, + 299 + ], + "blocks": [ + { + "bbox": [ + 47, + 293, + 286, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 293, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 286, + 304 + ], + "type": "text", + "content": "Table 2. Comparison with method [22] that without hand encoder." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 236, + 533, + 299 + ], + "lines": [ + { + "bbox": [ + 317, + 236, + 533, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 533, + 299 + ], + "type": "table", + "html": "
MethodPenetrationGrasp DisplaceContact
Depth↓Volume↓Mean±Variance↓Ratio↑
PCN+GC0.2211.346.00±4.8385.66
PCN+DG0.207.686.50±4.5172.85
Ours0.215.583.26±3.3898.67
", + "image_path": "435c738b6e34eae6fd7c123ad1ccfc9230bbe1cdec19774c0aee89dc89b95640.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "type": "text", + "content": "Table 3. Comparison with two-stage methods. We use our model trained on Object-S2HGD to compare with others. The two-stage methods initially complete the single-view point cloud (by PCN [54]), and generate human hand grasps on the completed point cloud using GraspCVAE (GC) [25] and our DiffuGrasp(DG)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 365, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 497 + ], + "type": "text", + "content": "\"Grasp Displace\" metric, a larger penetration volume may result in better grasping stability (the larger the penetration, the greater the force on the fingers in the simulator), which could make the results of GraspCVAE seem favorable. However, an ideal grasp should have both smaller penetration volume and good stability. Therefore, we present the balance on these two metrics of both methods in Fig. 5. It is evident that the points representing our method are closer to the origin point, which also indicates that our method better unifies penetration with grasp stability, meaning the generated grasps are more natural and reasonable." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "type": "text", + "content": "To generate more natural and stable grasps, instead of directly passing hand parameters into the decoder as done in [22], we first convert the hand parameters into point clouds and extract their features. We then feed the features into the decoder to predict the hand parameters. Our method incorporates more shape and positional information about the hand than direct decoding the hand parameters. To verify the effectiveness of our method, we conduct comparative experiments with the approach without the hand encoder used in [22], as shown in Table 2. The results show that, with a similar penetration volume, our method achieves higher grasp stability (smaller grasp displace), thereby effectively facilitating the generation of higher quality grasps and demonstrating the advantages of our approach." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Comparing with two-stage methods. For the task of generating human grasps based on single-view scene point clouds, one intuitive approach is to first utilize point cloud completion methods to complete the single-view point" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "837" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 525, + 133 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 525, + 133 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 525, + 133 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 525, + 133 + ], + "type": "table", + "html": "
MethodCVAEDiffuGraspGPpenetrationGrasp Displace Mean ± Variance(cm)↓Contact Ratio(%)↑
Depth(cm)↓Volume(cm3)↓
CVAE✓0.2313.543.10 ± 3.4998.07
DiffuGrasp✓0.2011.543.06 ± 3.3398.80
DiffuGrasp + GP✓✓0.216.582.73 ± 3.1699.41
", + "image_path": "bb2ce2da8431cf2fccae514f9f47b4c8639f8c3f5817123a4aeda6adee603fb8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 137, + 547, + 160 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 547, + 160 + ], + "type": "text", + "content": "Table 4. Ablation results of the DiffuGrasp module and the Global Perception module(GP) on our View-S2HGD dataset. The role of DiffuGrasp is demonstrated through a comparison with CVAE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 166, + 288, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 288, + 213 + ], + "type": "text", + "content": "clouds. Subsequently, a generation method like GraspTTA [25] is employed to generate grasps on the completed point clouds. [46] and [31] utilize such two-stage methods to generate dexterous hand grasps in single-view scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 214, + 290, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 290, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 290, + 468 + ], + "type": "text", + "content": "We first train a Point Completion Network (PCN) [54] to complete the input single-view point clouds into full objects with tabletop points. Then, we train GraspCVAE (GC) [25] and our DiffuGrasp (DG) to generate human grasps based on these completed results. The testing results are shown in Table 3. The results demonstrate that our method outperforms both two-stage methods by a large margin. The reason is that the two-stage method is not end-to-end, which leads to accumulated errors in the process, impacting the quality of generated grasps. In addition, most works of point cloud completion are not suitable for such a situation with scene points, which will lead to a decrease in the accuracy of point cloud completion. However, the two-stage method has very high requirements for the completion results of single-view point clouds. If the result is poor, the generated grasp can easily penetrate into the object or not make contact with it. In contrast, our end-to-end model effectively avoids such problems. Last, the overall models of two-stage methods tend to be larger, require longer and more tedious training processes, and are less straightforward to train compared to our end-to-end method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 476, + 142, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 142, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 142, + 490 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 544 + ], + "type": "text", + "content": "We conduct an ablation study on our S2HGrasp using our View-S2HGD dataset to highlight the role of the Global Perception module and demonstrate the advantages of our DiffuGrasp module over CVAE, as shown in Table 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "content": "Global Perception. The comparative results, as well as the visualizations of point cloud completion, highlight that this module helps the object encoder capture global object characteristics. This, in turn, assists the generation models in producing more natural, collision-free human grasps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Specifically, by comparing the results of \"DiffuGrasp\" and \"DiffuGrasp+GP\", we can find that with the Global Perception module, our model can effectively reduce the volume of hand-object penetration (from 11.54 to 6.58) and enhance the stability of the grasp. This also demonstrates that, with the aid of the Global Perception module, our model has a good perception of the overall shape of local objects, and can effectively prevent the generated hand from penetrating into the invisible areas of the object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 166, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 547, + 297 + ], + "type": "text", + "content": "DiffuGrasp. The previous work GraspTTA [25] leverages CVAE to generate human grasps based on the full object model, but suffers from severe model collapse and can't generate diverse grasps poses. We believe that the generative capabilities of CVAE are not sufficient, especially for the more challenging task of generating human grasps based on single-view scene point clouds. Therefore, we conduct comparative experiments on our View-S2HGD dataset using our DiffuGrasp and CVAE (without TTA, as TTA cannot be applied to incomplete point clouds) to demonstrate the advantages of our DiffuGrasp in this task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 298, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 547, + 431 + ], + "type": "text", + "content": "Results show that even without the Global Perception module, the penetration volume of DiffuGrasp is smaller than that of CVAE (11.54 to 13.54). Moreover, DiffuGrasp achieves higher grasp stability than CVAE with smaller penetration, which also shows that the grasps generated by DiffuGrasp are more reasonable and natural. In short, despite the incompleteness of the object and the interference from numerous scene points, DiffuGrasp outperforms CVAE and can generate stable human grasps with minimal penetration into the object. Additional ablation studies and experiments are shown in the supplementary file." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 443, + 379, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 379, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 379, + 455 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 464, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 643 + ], + "type": "text", + "content": "In this work, we explore a new task of generating human grasps based on single-view scene point clouds rather than full object models. We propose S2HGrasp, along with a new synthetic dataset S2HGD, to address the problem of hand-object penetration caused by the incompleteness of object point clouds. We design a Global Perception module to globally perceive partial objects and a DiffuGrasp module to generate plausible and natural human grasps despite numerous scene points. Our S2HGrasp effectively overcomes the limitation of existing approaches and extends human grasp generation beyond full object models, benefiting the study of hand-object interactions. The experimental results demonstrate that S2HGrasp outperforms other methods and achieves satisfying performance and generalization capabilities across different datasets and unseen objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 643, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 643, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 545, + 703 + ], + "type": "text", + "content": "Acknowledgments. This work was supported partially by the National Key Research and Development Program of China(2023YFA1008503),NSFC(U21A20471,U1911401), Guangdong NSF Project (No. 2023B1515040025, 2020B1 515120085)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "838" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 272, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 272, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 272, + 134 + ], + "type": "text", + "content": "[1] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 284, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 284, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 284, + 178 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 271, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 271, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 271, + 223 + ], + "type": "text", + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "text", + "content": "[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 286, + 323 + ], + "type": "text", + "content": "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "type": "text", + "content": "[6] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing hand-object interactions in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 285, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 285, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 285, + 422 + ], + "type": "text", + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "type": "text", + "content": "[8] Jiayi Chen, Mi Yan, Jiazhao Zhang, Yinzhen Xu, Xiaolong Li, Yijia Weng, Li Yi, Shuran Song, and He Wang. Tracking and reconstructing hand object interactions from point cloud sequences in the wild. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 260, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 260, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 260, + 523 + ], + "type": "text", + "content": "[9] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 282, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 282, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 282, + 579 + ], + "type": "text", + "content": "[10] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 280, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 280, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 280, + 624 + ], + "type": "text", + "content": "[11] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. arXiv preprint arXiv:1911.01911, 2019. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 677 + ], + "type": "text", + "content": "[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 281, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 281, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 281, + 713 + ], + "type": "text", + "content": "[13] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 538, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 538, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 538, + 95 + ], + "type": "text", + "content": "In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 542, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 542, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 542, + 150 + ], + "type": "text", + "content": "[14] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 533, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 533, + 217 + ], + "type": "text", + "content": "[15] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 219, + 529, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 529, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 529, + 283 + ], + "type": "text", + "content": "[16] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 537, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 537, + 351 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 537, + 351 + ], + "type": "text", + "content": "[17] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 544, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 544, + 396 + ], + "type": "text", + "content": "[18] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 523, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 523, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 523, + 431 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 434, + 535, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 535, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 535, + 475 + ], + "type": "text", + "content": "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 2022. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 478, + 531, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 531, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 531, + 521 + ], + "type": "text", + "content": "[21] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 541, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 541, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 541, + 577 + ], + "type": "text", + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 579, + 544, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 544, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 544, + 622 + ], + "type": "text", + "content": "[23] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 543, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 543, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 543, + 657 + ], + "type": "text", + "content": "[24] Wolfgang Hürst and Casper Van Wezel. Gesture-based interaction via finger tracking for mobile augmented reality. Multimedia Tools and Applications, 2013. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "839" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 285, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 285, + 127 + ], + "type": "text", + "content": "[26] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 286, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 286, + 172 + ], + "type": "text", + "content": "[27] Mia Kokic, Danica Kragic, and Jeannette Bohg. Learning to estimate pose and shape of hand-held objects from rgb images. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 284, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 284, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 284, + 216 + ], + "type": "text", + "content": "[28] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 216, + 286, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 286, + 259 + ], + "type": "text", + "content": "[29] Yiming Li, Wei Wei, Daheng Li, Peng Wang, Wanyi Li, and Jun Zhong. Hgc-net: Deep anthropomorphic hand grasping in clutter. In 2022 International Conference on Robotics and Automation (ICRA), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 272, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 272, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 272, + 304 + ], + "type": "text", + "content": "[30] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In Proceedings of the AAAI conference on artificial intelligence, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 274, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 274, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 274, + 337 + ], + "type": "text", + "content": "[31] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 283, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 283, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 283, + 381 + ], + "type": "text", + "content": "[32] Meysam Madadi, Sergio Escalera, Xavier Baró, and Jordi Gonzalez. End-to-end global to local cnn learning for hand pose recovery in depth data. arXiv preprint arXiv:1705.09606, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 381, + 272, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 272, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 272, + 415 + ], + "type": "text", + "content": "[33] Markus Oberweger, Paul Wohlhart, and Vincent Lepetit. Hands deep in deep learning for hand pose estimation. arXiv preprint arXiv:1502.06807, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 283, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 283, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 283, + 448 + ], + "type": "text", + "content": "[34] Liang Pan. Ecg: Edge-aware point cloud completion with graph convolution. IEEE Robotics and Automation Letters, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 286, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 286, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 286, + 492 + ], + "type": "text", + "content": "[35] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems. 2013. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "type": "text", + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 2017. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 286, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 286, + 591 + ], + "type": "text", + "content": "[37] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 278, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 278, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 278, + 647 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "type": "text", + "content": "[39] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 257, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 257, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 257, + 712 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 541, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 541, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 541, + 127 + ], + "type": "text", + "content": "[41] Srinath Sridhar, Anna Maria Feit, Christian Theobalt, and Antti Oulasvirta. Investigating the dexterity of multi-finger input for mid-air text entry. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 541, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 541, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 541, + 184 + ], + "type": "text", + "content": "[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 541, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 541, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 541, + 229 + ], + "type": "text", + "content": "[43] Bugra Tekin, Federica Bogo, and Marc Pollefeys. H+ o: Unified egocentric recognition of 3d hand-object poses and interactions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 522, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 522, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 522, + 296 + ], + "type": "text", + "content": "[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. arXiv preprint arXiv:2304.00464, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "type": "text", + "content": "[45] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 353, + 542, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 542, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 542, + 398 + ], + "type": "text", + "content": "[46] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 533, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 533, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 533, + 453 + ], + "type": "text", + "content": "[47] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 454, + 543, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 543, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 543, + 499 + ], + "type": "text", + "content": "[48] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 542, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 542, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 542, + 578 + ], + "type": "text", + "content": "[49] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 578, + 541, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 541, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 541, + 632 + ], + "type": "text", + "content": "[50] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 634, + 541, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 541, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 541, + 689 + ], + "type": "text", + "content": "[51] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 690, + 542, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 542, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 542, + 713 + ], + "type": "text", + "content": "[52] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "840" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 350 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 278, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 278, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 278, + 106 + ], + "type": "text", + "content": "Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 276, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 276, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 276, + 160 + ], + "type": "text", + "content": "[53] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Point: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 285, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 285, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 285, + 205 + ], + "type": "text", + "content": "[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV), 2018. 2, 4, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 282, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 282, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 282, + 262 + ], + "type": "text", + "content": "[55] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 279, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 279, + 305 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 279, + 305 + ], + "type": "text", + "content": "[56] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 282, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 282, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 282, + 350 + ], + "type": "text", + "content": "[57] Dian Zheng, Xiao-Ming Wu, Zuhao Liu, Jingke Meng, and Wei-shi Zheng. Diffuvolume: Diffusion model for volume based stereo matching. arXiv preprint arXiv:2308.15989, 2023. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "841" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_content_list.json b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2811602e39a606c4457c3e8e1a16a7555b4586da --- /dev/null +++ b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_content_list.json @@ -0,0 +1,1699 @@ +[ + { + "type": "text", + "text": "Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruicong Liu", + "bbox": [ + 184, + 180, + 289, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Takehiko Ohkawa", + "bbox": [ + 328, + 181, + 473, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingfang Zhang", + "bbox": [ + 514, + 181, + 648, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yoichi Sato", + "bbox": [ + 689, + 181, + 784, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Tokyo, Japan", + "bbox": [ + 359, + 199, + 609, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lruicong, ohkawa-t, mfzhang, ysato}@iis.u-tokyo.ac.jp", + "bbox": [ + 243, + 219, + 720, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The pursuit of accurate 3D hand pose estimation stands as a keystone for understanding human activity in the realm of egocentric vision. The majority of existing estimation methods still rely on single-view images as input, leading to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, adding another camera to better capture the shape of hands is a practical direction. However, existing multi-view hand pose estimation methods suffer from two main drawbacks: 1) Requiring multi-view annotations for training, which are expensive. 2) During testing, the model becomes inapplicable if camera parameters/layout are not the same as those used in training. In this paper, we propose a novel Single-to-Dual-view adaptation (S2DHand) solution that adapts a pre-trained single-view estimator to dual views. Compared with existing multi-view training methods, 1) our adaptation process is unsupervised, eliminating the need for multi-view annotation. 2) Moreover, our method can handle arbitrary dual-view pairs with unknown camera parameters, making the model applicable to diverse camera settings. Specifically, S2DHand is built on certain stereo constraints, including pair-wise cross-view consensus and invariance of transformation between both views. These two stereo constraints are used in a complementary manner to generate pseudo-labels, allowing reliable adaptation. Evaluation results reveal that S2DHand achieves significant improvements on arbitrary camera pairs under both in-dataset and cross-dataset settings, and outperforms existing adaptation methods with leading performance. Project page: https://github.com/ut-vision/S2DHand.", + "bbox": [ + 76, + 300, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 767, + 209, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Delving into the realm of egocentric vision (first-person view), the pursuit of refining 3D hand pose estimation stands as a keystone for understanding human activity. This quest not only forges new paths in human-computer interaction [31, 34, 38], but also empowers imitation learning", + "bbox": [ + 75, + 792, + 468, + 869 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da105d94f95ccd3b3b84db61e86194932ff7d10116dbfbd49601f145da7eac0d.jpg", + "image_caption": [ + "Figure 1. From (a) to (b), our single-to-dual-view adaptation method adapts a traditional single-view hand pose estimator to arbitrary dual views and achieves better accuracy. (a) Traditional single-view hand pose estimation. (b) Inference process of the adapted model under a dual-view setting." + ], + "image_footnote": [], + "bbox": [ + 522, + 272, + 870, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[8, 13, 37]. Moreover, it enhances the immersive experience in augmented/virtual reality (AR/VR) to new heights [15, 33]. Recently, with the advancements of AR/VR headsets, egocentric data has become increasingly prevalent [5, 10], leading to an increasing demand for estimating 3D hand poses from egocentric viewpoints.", + "bbox": [ + 496, + 537, + 893, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve better 3D hand pose estimation performance, recent years have witnessed many networks with various structures [7, 39, 46]. However, the majority of existing hand pose estimation methods are still under a single-view setting, which is convenient but leads to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, a potential solution is to add another camera to expand the field-of-view and reduce depth ambiguity by capturing the hand shape from an additional view angle. Furthermore, the use of multiple cameras also aligns with industry trends, as demonstrated by the latest AR/VR headsets such as the Apple Vision Pro and Meta Quest, which feature multiple egocentric cameras. Overall, an unavoidable trend towards multi-view settings in hand pose estimation is emerging, driven by its technological advantages and the direction of industrial development.", + "bbox": [ + 496, + 628, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, several existing studies [4, 12, 19] have paid attention to hand pose estimation under multi-view settings.", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by JSPS KAKENHI Grant Number JP20H04205 and JST ACT-X Grant Number JPMJAX2007.", + "bbox": [ + 75, + 875, + 470, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "677", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/13778c7f7c503fa63a11134c453c5599a58bb3baa399a0e69cf39bfe9193e023.jpg", + "table_caption": [ + "Table 1. Differences among traditional single-view methods, multi-view training methods, and our proposed single-to-dual-view adaptation method. Green indicates lower training requirements, enhanced testing results, and reduced camera parameter requirements, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsPre-training dataset requiredTestCamera parameters
Traditional single-view methods [7, 39, 46]Single-view (common single-camera setting)Single-viewNot required
Multi-view training methods [4, 12, 19]Dual-view (need multi-camera setting)Dual-view & same camera poses with trainingRequired & same with training
Single-to-dual-view adaptation (Ours)Single-view (common single-camera setting)Dual-view & arbitrary camera posesNot required
", + "bbox": [ + 78, + 119, + 890, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These methods typically process input images from multiple views simultaneously, utilizing a feature fusion module to arrive at a final prediction [22, 41]. However, all these methods have two significant drawbacks that limit their applicability. 1) The training, especially for the feature fusion module, necessitates multi-view labels, which are costly to annotate. 2) During testing, the same camera parameters as in training must be used. An estimator trained under a specific multi-camera setup becomes inapplicable if there are any changes to the camera layout or parameters.", + "bbox": [ + 75, + 181, + 468, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike existing multi-view training methods, we propose a new solution that adapts an estimator from single-view to dual-view without needing multi-view labels or camera parameters. As shown in Fig. 1, given a pre-trained estimator, our method adapts it to an arbitrary dual-view setting (from (a) to (b)), where two cameras are placed in any layout without knowing their parameters. Here, all we need is a pre-trained estimator and a sufficient number of unlabeled dual-view inputs from the two cameras. As compared in Tab. 1 (row 2-3), in contrast to multi-view training, our method only needs common and cheaper single-view data for training. During testing, unlike existing methods, our method is compatible with arbitrary dual-view pairs, making the model applicable to flexible and changeable camera settings. Specifically, when camera settings change, it is easy and swift to repeat our method's adaptation process to re-adapt the pre-trained estimator to work well with new camera parameters. For camera parameters, existing methods not only need them for training, but also require them the same with testing. Conversely, our method is clearly more practical since no camera parameters are required.", + "bbox": [ + 75, + 335, + 470, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these advancements, we present a novel unsupervised Single-to-Dual-view adaptation framework (S2DHand) for egocentric 3D hand pose estimation. It uses certain stereo constraints for adaptation, including cross-view consensus (pair-wise) and invariance of transformation between both camera coordinate systems (to all input pairs). These two stereo constraints are used in a complementary manner to refine the accuracy of pseudo-labels, allowing the model to better fit to the dual views. Specifically, the cross-view consensus is leveraged through an attention-based merging module, and the invariance of transformation is utilized via a rotation-guided refinement module.", + "bbox": [ + 75, + 656, + 468, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method by adapting a pre-trained estimator to several dual-camera pairs placed in arbitrary poses [30]. Our evaluation encompasses both in-dataset and cross-dataset scenarios. Experimental results reveal that our", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "technique not only realizes notable improvements across all pairs but also surpasses state-of-the-art adaptation methods. The primary contributions of this paper are summarized as:", + "bbox": [ + 496, + 181, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel unsupervised single-to-dual-view adaptation (S2DHand) solution for egocentric 3D hand pose estimation. Our method can adapt a traditional single-view estimator for arbitrary dual views without requiring annotations or camera parameters.", + "- We build a pseudo-label-based strategy for adaptation. It leverages cross-view consensus and invariance of transformation between both camera coordinate systems for reliable pseudo-labeling. This leads to two key modules: attention-based merging and rotation-guided refinement.", + "- Evaluation results demonstrate the benefits of our approach for arbitrarily placed camera pairs. Our method achieves significant improvements for all pairs both under in-dataset and cross-dataset settings." + ], + "bbox": [ + 500, + 228, + 890, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 454, + 640, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-view hand pose estimation", + "text_level": 1, + "bbox": [ + 500, + 479, + 785, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view hand pose estimation accepts multi-view images as input and outputs a final 3D hand pose, which remains a relatively unexplored research area. Previous works [4, 12, 19] design various network structures to predict 3D hand poses through a multi-view fusion module. Similarly, multiple studies have been done [44, 45] through feature fusion for body pose estimation under multi-view settings.", + "bbox": [ + 496, + 503, + 890, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All these studies have two limitations: 1) they require costly multi-view images and annotations for training, 2) during testing, camera poses are assumed to be known and covered by training data, thereby limiting their applicability. In contrast, our method eliminates the need for multi-view annotations and is adaptable to arbitrary dual views.", + "bbox": [ + 496, + 609, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Adaptation in hand pose estimation", + "text_level": 1, + "bbox": [ + 500, + 710, + 808, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptation aims at tailoring a model for specific application scenarios [20, 24, 29]. Existing adaptation methods in hand pose estimation mainly focus on adaptation across different domains (datasets), e.g., entropy minimization [17, 32], consistency regularization [3, 21, 29], and pseudo-labeling methods [16, 35, 42]. Prior works only use constraints in single-view settings for adaptation, e.g., bio-mechanical constraint [21, 42], and are thus limited to single-view inference. Unlike these methods, we propose stereo constraints from dual views for adaptation and supports dual-view inference, which extends the application scenarios.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "678", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac57a26fcd87a67b09f50e22e0d29616418a35bb90272dd234bb52ef3a7cab2b.jpg", + "image_caption": [ + "Figure 2. Problem setting of single-to-dual-view adaptation for hand pose estimation. (a) The input and output of adaptation. (b) The dual-view testing scheme after adaptation." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 467, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Setting", + "text_level": 1, + "bbox": [ + 76, + 358, + 238, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the task setting of single-to-dual-view adaptation for hand pose estimation. We denote unlabeled dual-view data as $\\mathcal{D} = \\{\\mathbf{x}_i^{v1},\\mathbf{x}_i^{v2}|_{i = 1}^N\\}$ , where $\\mathbf{x}_i^{v1}$ and $\\mathbf{x}_i^{v2}$ denote the $i$ -th image from view1 and view2, respectively, $N$ is the number of image pairs. The dual-view data $\\mathcal{D}$ contains no ground-truth hand poses or camera parameters.", + "bbox": [ + 75, + 383, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (a), suppose we have a baseline hand pose estimator $\\mathcal{H}(\\cdot|\\theta)$ with parameters $\\theta$ pre-trained from common single-view data. Leveraging $\\mathcal{D}$ can enhance its performance, as $\\mathcal{D}$ provides additional information from a dual-view setup. Our objective is to adapt this pre-trained estimator, $\\mathcal{H}(\\cdot|\\theta)$ , to an arbitrary yet fixed dual-view setting (with unknown camera poses) without needing groundtruths or camera parameters. By inputting $\\mathcal{H}(\\cdot|\\theta)$ and $\\mathcal{D}$ into our method, it outputs an adapted estimator $\\mathcal{H}(\\cdot|\\theta')$ with parameters $\\theta'$ tailored for the dual-view scenario.", + "bbox": [ + 75, + 476, + 468, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Upon adapting the estimator, its inference mechanism is correspondingly tailored for dual-view scenarios (Fig. 1 (b)). During testing, the adapted estimator $\\mathcal{H}(\\cdot |\\theta^{\\prime})$ processes a dual-view input pair $(\\mathbf{x}^{v1},\\mathbf{x}^{v2})$ and produces two predictions $(\\mathbf{J}^{v1},\\mathbf{J}^{v2})$ , where each $\\mathbf{J}^v\\in \\mathbb{R}^{21\\times 3}$ represents the 21 3D joints of the hand. These predictions denote the 3D hand joints for each view and can be combined together to generate a final output, e.g., through a simple average.", + "bbox": [ + 75, + 628, + 468, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera layout for a multi-view headset. Fig. 3 illustrates an example of headset-mounted camera setups for multiview egocentric data capture, with four cameras at each corner for different views. The top-right of Fig. 3 displays images from these cameras. Six distinct dual-view pairs can be created from these four views. As a supplement, the bottom of Fig. 3 shows the synthetic training data, highlighting variations in style and lighting. Such data helps to explore the performance of our method under cross-dataset or simulate-to-real settings. See Sec. 5.1 for dataset details.", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d98d7ba4e1360a685602effc4b571cc2a9a2ec620bd9659c736e40017cc7990.jpg", + "image_caption": [ + "Figure 3. Top: Headset and its camera layout to collect multi-view data, and samples from the four views. Bottom: Samples of synthetic data. Image samples are from AssemblyHands [30], GANerated Hands [28], and Rendered Handpose [47], respectively." + ], + "image_footnote": [], + "bbox": [ + 532, + 92, + 859, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 459, + 671, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel unsupervised single-to-dual-view adaptation framework (S2DHand). Before adaptation, an initialization step is performed to initialize the rotation matrix between both views (Sec. 4.1). The rotation matrix is essential to establish the transformation between two camera coordinate systems. The architecture overview, as illustrated in Fig. 4, comprises two branches, an estimator $\\mathcal{H}$ and a its momentum version $\\overline{\\mathcal{H}}$ . The adaptation process is designed from two stereo constraints, pair-wise cross-view consensus and invariant rotation transformation between both camera coordinate systems. This leads to two key pseudo-labeling modules: attention-based merging and rotation-guided refinement (Secs. 4.3 and 4.4). Notably, these two modules function in a complementary manner, depending on the prediction accuracy, ensuring reliable pseudo-labeling.", + "bbox": [ + 496, + 484, + 892, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Initialization", + "text_level": 1, + "bbox": [ + 500, + 724, + 635, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initialization step aims to estimate a relatively accurate rotation matrix $R$ , since $R$ is necessary to link the two camera coordinate systems [23, 43]. It should be noted that translation vector between the two cameras is not necessary, as the predicted hand poses are usually aligned by the wrist during testing [30, 46]. Assuming that the initial pre-trained estimator is sufficiently accurate to generate reasonable predictions, we estimate the $R$ using the predictions of unlabeled dual-view data $\\mathcal{D} = \\{\\mathbf{x}_i^{v1}, \\mathbf{x}_i^{v2}|_{i=1}^N\\}$ . Given $\\mathcal{D}$ , the estimator $\\mathcal{H}$ can output $N$ pairs of predictions $\\{\\mathbf{J}_i^{v1}, \\mathbf{J}_i^{v2}|_{i=1}^N\\}$ ,", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "679", + "bbox": [ + 486, + 944, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e0d20658f0cf2f945f1885462c7ddc2b354f0b5fc21b484ad0b8802e033f319.jpg", + "image_caption": [ + "Figure 4. Overview of the proposed S2DHand, image pairs captured from arbitrarily placed dual cameras are input for adaptation. The architecture of S2DHand is illustrated in the dark dashed box, which contains a dynamically updated estimator and a momentum estimator. The momentum estimator's predictions are used to generate pseudo-labels, which are then processed by our pseudo-labeling module (Secs. 4.3 and 4.4). Using the pseudo-labels, a loss function is computed to update the estimator. The rotation matrix $R$ from the initialization step (Sec. 4.1) is required for the pseudo-labeling." + ], + "image_footnote": [], + "bbox": [ + 148, + 88, + 823, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{J}_i^v\\in \\mathbb{R}^{21\\times 3}$ (21 is the number of 3D joints). Then, the rotation matrix $R$ is estimated by:", + "bbox": [ + 75, + 446, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {(0)} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} r o t \\left(\\mathbf {J} _ {i} ^ {v 1}, \\mathbf {J} _ {i} ^ {v 2}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 491, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the superscript of $R$ denotes the iteration number, (0) indicates that it is before the first iteration. The rot function [18] generates a $3 \\times 3$ rotation matrix from two $21 \\times 3$ joint predictions. Note that the average in Eq. (1) is not elementwise, but an average of rotation matrices.", + "bbox": [ + 76, + 544, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Single-to-dual-view adaptation", + "text_level": 1, + "bbox": [ + 76, + 630, + 349, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the initialized $R$ , the adaptation process begins. The S2DHand framework comprises two branches, an estimator $\\mathcal{H}(\\cdot |\\theta)$ with dynamically updating parameters $\\theta$ , and its momentum version $\\overline{\\mathcal{H}} (\\cdot |\\overline{\\theta})$ , which updates its parameters $\\bar{\\theta}$ using temporal moving average. Temporal moving average has been proved by many works [6, 14, 25, 27] that can help to stabilize the training process. The $\\bar{\\theta}$ is updated as:", + "bbox": [ + 75, + 652, + 468, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\theta} ^ {(T)} = \\eta_ {\\theta} \\bar {\\theta} ^ {(T - 1)} + (1 - \\eta_ {\\theta}) \\theta , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 770, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\overline{\\theta}^{(T - 1)}$ indicates the temporal averaged parameters in the previous iteration $T - 1$ , and $\\eta_{\\theta}$ represents the ensembling momentum, which is set as 0.99 [25, 27].", + "bbox": [ + 75, + 804, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 4, during the single-to-dual-view adaptation, the role of the momentum model $\\overline{\\mathcal{H}}$ is to generate pseudo-labels, which are then utilized to supervise the", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model $\\mathcal{H}$ . The pseudo-labeling module (Secs. 4.3 and 4.4) outputs pseudo-labels $\\hat{\\mathbf{y}}^{v1},\\hat{\\mathbf{y}}^{v2}$ based on the predictions $\\overline{\\mathbf{J}}^{v1},\\overline{\\mathbf{J}}^{v2}$ from $\\overline{\\mathcal{H}}$ . These pseudo-labels are then used to supervise the predictions $\\mathbf{J}^{v1},\\mathbf{J}^{v2}$ of $\\mathcal{H}$ . The loss function is computed as:", + "bbox": [ + 496, + 448, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| \\mathbf {J} ^ {v 1} - \\hat {\\mathbf {y}} ^ {v 1} \\right\\| _ {2} + \\left\\| \\mathbf {J} ^ {v 2} - \\hat {\\mathbf {y}} ^ {v 2} \\right\\| _ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 534, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The estimator follows the implementation of DetNet [46], where $\\mathcal{H}$ directly outputs heatmaps, and $\\mathbf{J}$ is calculated from the heatmaps. Therefore, the loss function is actually computed from corresponding heatmaps, here we write these 3D-joint variables only for better understanding.", + "bbox": [ + 496, + 560, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Pseudo-labeling: attention-based merging", + "text_level": 1, + "bbox": [ + 498, + 643, + 857, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention-based merging (ABM) module, which constitutes the first part of pseudo-labeling, is derived from cross-view consensus. Cross-view consensus refers to the concept of achieving agreement or consistency between different views of the same data [40]. Theoretically, when transformed into the same coordinate system, the two predictions $\\overline{\\mathbf{J}}^{v1}$ and $\\overline{\\mathbf{J}}^{v2}$ from different views should be identical, i.e., $R\\overline{\\mathbf{J}}^{v1} = \\overline{\\mathbf{J}}^{v2}$ , with $\\overline{\\mathbf{J}}^{v1}\\overline{\\mathbf{J}}^{v2}$ being aligned with wrist joint.", + "bbox": [ + 496, + 667, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This stereo constraint is the foundation for this module to generate accurate pseudo-labels. Prior works utilize a simple average [26, 29] (e.g., $(R\\overline{\\mathbf{J}}^{v1} + \\overline{\\mathbf{J}}^{v2}) / 2$ ) or sample-wise confidence [2, 29] to improve the quality of pseudolabels. However, these approaches overlook the varying confidence in joints that is caused by differences in image capture across views. For instance, a joint that is occluded", + "bbox": [ + 496, + 792, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "680", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dd6e00e991ca4caf4f4760ba95568fd0f3c456761820a11e9bc45f0e7d102862.jpg", + "image_caption": [ + "Figure 5. Top: illustration of the first part of pseudo-labeling: attention-based merging module. The generating process of $\\hat{y}_{abm}^{v2}$ in view2 is shown as an example, the process of view1 is the same. Bottom: visualizations of heatmaps with different accuracy." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 444, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in one view but fully visible in another could lead to reliable predictions being hindered by unreliable ones.", + "bbox": [ + 75, + 391, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we propose joint-wise attention $\\mathbf{w} \\in \\mathbb{R}^{21 \\times 1}$ to represent each joint's prediction confidence. It is derived from the 2D heatmap $\\overline{\\mathbf{h}} \\in \\mathbb{R}^{21 \\times 32 \\times 32}$ output from $\\overline{\\mathcal{H}}$ . The $\\overline{\\mathbf{h}}$ indicates the probability of each joint's presence at every pixel in the 2D image space. This approach is from an observation (bottom of Fig. 5): as the error of prediction increases, the intensity of the heatmap's hotspot decreases, i.e., darker indicates low accuracy. Inspired by this, we propose an attention-generating function:", + "bbox": [ + 75, + 421, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {j} ^ {v} = \\frac {\\beta^ {\\max \\left(\\overline {{\\mathbf {h}}} _ {j} ^ {v}\\right)}}{\\sum_ {v \\in \\{v 1 , v 2 \\}} \\beta^ {\\max \\left(\\overline {{\\mathbf {h}}} _ {j} ^ {v}\\right)}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 565, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the subscript $j$ indicate the index of joint, i.e., $j = 1,2,\\ldots ,21$ . We introduce a hyper-parameter $\\beta$ here to adjust softness. Please refer to Sec. 5.8 for parameter choosing.", + "bbox": [ + 76, + 618, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The workflow of this attention-based merging module is illustrated at the top of Fig. 5. First, we transform both predictions into the same camera coordinate system. Then, a joint-wise multiplication is performed for each of them using the attention $\\mathbf{w}^{v1},\\mathbf{w}^{v2}$ . Finally, the pseudo-label $\\hat{\\mathbf{y}}_{abm}$ is calculated through a summation operation. In summary:", + "bbox": [ + 75, + 662, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {a b m} ^ {v 1} = \\mathbf {w} ^ {v 1} \\overline {{\\mathbf {J}}} ^ {v 1} + \\mathbf {w} ^ {v 2} R ^ {T} \\overline {{\\mathbf {J}}} ^ {v 2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 763, + 468, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {a b m} ^ {v 2} = \\mathbf {w} ^ {v 1} R \\overline {{\\mathbf {J}}} ^ {v 1} + \\mathbf {w} ^ {v 2} \\overline {{\\mathbf {J}}} ^ {v 2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 787, + 367, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Pseudo-labeling: rotation-guided refinement", + "text_level": 1, + "bbox": [ + 76, + 816, + 452, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The rotation-guided refinement (RGR) module is based on another stereo constraint: invariance of rotation transformation between both views. This implies that the estimated rotation matrix should remain unchanged since the cameras", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e99f4a925ed0416018d1aa30067e0f58f892d493ca4801a04510408bb4912ef6.jpg", + "image_caption": [ + "Figure 6. Illustration of the second part of pseudo-labeling: rotation-guided refinement module." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 879, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are fixed, i.e., theoretically $rot(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2}) = C$ . In light of this, this module aims to refine the predictions $\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2}$ such that $rot(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2})$ becomes invariant across all input data. The refinement result becomes the pseudo-label $\\hat{\\mathbf{y}}_{rgr}$ .", + "bbox": [ + 496, + 224, + 890, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The workflow is shown in Fig. 6. Given a pair of predictions, our method estimates a new rotation matrix $R' = \\text{rot}(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2})$ . Subsequently, the more accurately estimated $R$ from the initialization step is set as the target for refinement, aiming to minimize $\\| R - R' \\|_F$ . The refinement process can be expressed as:", + "bbox": [ + 496, + 294, + 890, + 385 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {r g r} ^ {v 1}, \\hat {\\mathbf {y}} _ {r g r} ^ {v 2} = \\underset {\\overline {{\\mathbf {J}}} ^ {v 1}, \\overline {{\\mathbf {J}}} ^ {v 2}} {\\arg \\min } \\| R - r o t (\\overline {{\\mathbf {J}}} ^ {v 1}, \\overline {{\\mathbf {J}}} ^ {v 2}) \\| _ {F}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 392, + 890, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In detail, we employ BFGS [1] algorithm for minimizing.", + "bbox": [ + 496, + 431, + 879, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Final pseudo-label. Finally, the pseudo-label is calculated from a weighted average of $\\hat{\\mathbf{y}}_{abm}$ and $\\hat{\\mathbf{y}}_{rgr}$ :", + "bbox": [ + 496, + 446, + 890, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} = \\alpha \\hat {\\mathbf {y}} _ {a b m} + (1 - \\alpha) \\hat {\\mathbf {y}} _ {r g r}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 484, + 890, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we introduce another pre-fixed hyper-parameter $\\alpha$ , which adjusts the weight of the two parts of the pseudolabel. Empirically, we set $\\alpha = 0.7$ (see Sec. 5.8).", + "bbox": [ + 496, + 508, + 890, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complement between two pseudo-labels. When the predictions $\\overline{\\mathbf{J}}^{v1}$ and $\\overline{\\mathbf{J}}^{v2}$ are accurate, $R^{\\prime}$ closely approximates $R$ , making $\\hat{\\mathbf{y}}_{rgr}$ redundant. In such cases, $\\hat{\\mathbf{y}}_{abm}$ is beneficial as it merges the two accurate predictions. Conversely, if $\\overline{\\mathbf{J}}^{v1}$ and $\\overline{\\mathbf{J}}^{v2}$ are unreliable, $\\hat{\\mathbf{y}}_{abm}$ will consequently also lack accuracy. In such instances, $\\hat{\\mathbf{y}}_{rgr}$ steps in as a complementary tool for refining pseudo-labels. By minimizing $\\| R - R^{\\prime}\\|_{F}$ , it optimizes the predictions towards an alignment with real-world condition.", + "bbox": [ + 496, + 554, + 890, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update rotation matrix R. Clearly, the rotation matrix $R$ from the initialization step plays an important role in bridging the two camera coordinate systems. As can be expected, its accuracy significantly affects the final performance. To enhance its accuracy with each iteration, we also employ a temporal moving average for its updates. Given an input batch containing $B$ image pairs, the $R$ is updated as:", + "bbox": [ + 496, + 695, + 890, + 800 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {(T)} = \\eta_ {R} R ^ {(T - 1)} + (1 - \\eta_ {R}) \\cdot \\frac {1}{B} \\sum_ {i = 1} ^ {B} r o t \\left(\\overline {{\\mathbf {J}}} _ {i} ^ {v 1}, \\overline {{\\mathbf {J}}} _ {i} ^ {v 2}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 808, + 890, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the ensembling momentum $\\eta_{R}$ is set as 0.999 [14]. This way, the rotation matrix $R$ can be updated slowly and provide a more accurate rotation function with iteration.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "681", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5e548ac51d891c9d5c91b67a51928e54281acea8507ce3b0f207431bdee2a0b6.jpg", + "table_caption": [ + "Table 2. Adaptation results for all dual-camera pairs. The camera pairs are from $\\mathcal{D}_{ah}$ dataset [30], which is divided into two parts according to the collecting headset. In \"In-dataset\" and \"Cross-dataset\" settings, the baseline model is pre-trained on $\\mathcal{D}_{ah}$ and $\\mathcal{D}_{syn}$ [28, 47], respectively. The adaptation yields a uniquely adapted model for each pair, and \"Overall\" averages the results across all 6 pairs." + ], + "table_footnote": [], + "table_body": "
Camera pairMethodIn-dataset (Dah → Dah)Cross-dataset (Dsyn → Dah)
Dah - Headset1Dah - Headset2Dah - Headset1Dah - Headset2
Mono-MDual-MMono-MDual-MMono-MDual-MMono-MDual-M
cam 1,2Baseline43.0039.2054.7152.3867.9360.4870.3262.26
S2DHand31.01 ▼ 27.9%31.36 ▼ 20.0%45.52 ▼ 16.8%45.14 ▼ 13.8%63.46 ▼ 6.6%59.32 ▼ 1.9%70.09 ▼ 0.3%60.97 ▼ 2.1%
cam 1,3Baseline25.0023.2922.5921.0857.7951.4264.0060.25
S2DHand19.73 ▼ 21.1%19.92 ▼ 14.5%17.90 ▼ 20.8%17.68 ▼ 16.1%50.84 ▼ 12.0%47.55 ▼ 7.5%61.81 ▼ 3.4%58.34 ▼ 3.2%
cam 1,4Baseline24.9022.7016.7314.9152.7146.5554.3250.57
S2DHand20.88 ▼ 16.1%20.87 ▼ 8.1%14.64 ▼ 12.5%14.29 ▼ 4.2%46.05 ▼ 12.6%42.50 ▼ 8.7%46.59 ▼ 14.2%45.66 ▼ 9.7%
cam 2,3Baseline17.9615.2317.1015.0853.3648.4252.8448.84
S2DHand14.97 ▼ 16.6%14.44 ▼ 5.2%14.42 ▼ 15.7%14.20 ▼ 5.8%40.26 ▼ 24.6%39.32 ▼ 18.8%43.61 ▼ 17.5%42.88 ▼ 12.2%
cam 2,4Baseline22.0919.8423.2420.9659.4454.3261.1357.41
S2DHand17.98 ▼ 18.6%17.75 ▼ 10.5%18.31 ▼ 21.2%18.41 ▼ 12.2%50.59 ▼ 14.9%49.41 ▼ 9.0%52.45 ▼ 14.2%51.48 ▼ 10.3%
cam 3,4Baseline16.8315.7719.9318.0845.8242.3449.8448.99
S2DHand16.36 ▼ 2.8%15.55 ▼ 1.4%19.25 ▼ 3.4%17.80 ▼ 1.5%39.46 ▼ 13.9%37.43 ▼ 11.6%44.04 ▼ 11.6%42.88 ▼ 12.5%
OverallBaseline24.9622.6725.7223.7556.1850.5958.7454.72
S2DHand20.16 ▼ 19.2%19.98 ▼ 11.9%21.67 ▼ 15.7%21.25 ▼ 10.5%48.44 ▼ 13.8%45.92 ▼ 9.2%53.11 ▼ 9.6%50.37 ▼ 7.9%
", + "bbox": [ + 86, + 133, + 885, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 369, + 202, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 393, + 174, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ AssemblyHands [30] $(\\mathcal{D}_{ah})$ as the evaluation set, as it is the newest large-scale benchmark dataset with high-quality multi-view 3D hand pose annotations. As for the training set, we set two adaptation scenarios: 1) in-dataset setting where the training set is drawn from the same dataset $\\mathcal{D}_{ah}$ and 2) cross-dataset setting where we use synthetic dataset $(\\mathcal{D}_{syn})$ as the training set, consisting of Rendered Handpose [47] and GANerated Hands [28]. The details of the datasets are as below:", + "bbox": [ + 75, + 416, + 470, + 551 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- **AssemblyHands** [30] is a large-scale benchmark dataset featuring accurate 3D hand pose annotations. Collected using two AR headsets, it comprises images captured from four synchronized egocentric cameras. The dataset includes 412K training samples and 62K testing samples.", + "- GANerated Hands [28] includes over 330K color images of hands. The images are synthetically generated and then fed to a GAN [9] to make the features closer to real hands.", + "- Rendered Handpose [47] contains about 44K samples. The images are rendered with freely available characters. In detail, AssemblyHands is collected by two VR headsets, as shown in Fig. 3, each headset has four egocentric cameras at four corners. Following the collecting devices, we separate $\\mathcal{D}_{ah}$ into two parts, each part is collected using one headset, namely $\\mathcal{D}_{ah} - Headset1/2$ ." + ], + "bbox": [ + 76, + 553, + 470, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 786, + 264, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric. We compare the predictions from our model with the ground-truth labels in root-relative coordinates, and use the common mean per joint position error (MPJPE) in millimeters as the evaluation metric. However, since our focus is on the single-to-dual-view adaptation task, the adapted estimator is expected to perform in dual-", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "view settings. This implies that traditional MPJPE computed from single-view (monocular MPJPE, Mono-M) cannot be sufficient. As a result, we propose a new dual-view MPJPE metric Dual-M in addition to monocular MPJPE. The metrics are defined as follows:", + "bbox": [ + 496, + 369, + 890, + 445 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Mono-M: the traditional monocular MPJPE, which collects all the single-view errors from both views and calculates their average.", + "- Dual-M: the proposed metric under dual views. To calculate it, first the predictions from both views are averaged using a rotation matrix $R$ . Then, we calculate the MPJPE of the averaged predictions as the Dual-M. Usually, the $R$ is from the initialization step of our method." + ], + "bbox": [ + 500, + 446, + 893, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation detail. We employ PyTorch for implementation. All experiments run on a single NVIDIA A100 GPU. DetNet from [46] is adopted as the backbone of our hand pose estimator. Adam optimizer is employed with a learning rate of $1 \\times 10^{-3}$ to pre-train the 3D hand pose estimation network. For the single-to-dual-view adaptation, we use the Adam optimizer with a learning rate of $5 \\times 10^{-4}$ .", + "bbox": [ + 496, + 566, + 890, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Adaptation results for all camera pairs", + "text_level": 1, + "bbox": [ + 498, + 681, + 834, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we use our method to adapt the same pretrained single-view hand pose estimator to all dual-view pairs from the evaluation set $\\mathcal{D}_{ah}$ independently, yielding one adapted model for each pair. Experiments are conducted under both in-dataset and cross-dataset settings. Under the in-dataset setting (Tab. 2, $\\mathcal{D}_{ah} \\to \\mathcal{D}_{ah}$ ), the baseline model is pre-trained on $\\mathcal{D}_{ah}$ , while under the cross-dataset setting $(\\mathcal{D}_{syn} \\to \\mathcal{D}_{ah})$ , the baseline model is pre-trained on $\\mathcal{D}_{syn}$ before being adapted to the camera pairs from $\\mathcal{D}_{ah}$ .", + "bbox": [ + 496, + 704, + 890, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, compared with the pre-trained model (Baseline), our S2DHand offers significant accuracy gains under both settings among all camera pairs. This indicates that our method can adapt well to arbitrary dual views re", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "682", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gardless of the camera positions or pre-training datasets.", + "bbox": [ + 76, + 90, + 447, + 106 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results demonstrate that the S2DHand offers substantial improvements. On average, the improvement in both monocular (Mono-M) and dual-view (Dual-M) metrics exceeds $10\\%$ , with the maximum improvement exceeding $20\\%$ . Interestingly, we can see that the improvement for cam 1, 2 under the cross-dataset setting is relatively small. This indicates that low initial accuracy limits the performance.", + "bbox": [ + 75, + 107, + 467, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Comparison under cross-dataset settings", + "text_level": 1, + "bbox": [ + 76, + 242, + 424, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method is compared with state-of-the-art adaptation techniques in cross-dataset settings. Considering the prevalence and significance of cross-dataset scenarios in real-world applications, this experiment evaluates the capability of S2DHand in comparison to leading domain adaptation methods. Specifically, adaptation methods included in the comparison are: SFDAHPE [32], RegDA [17], DAGEN [11], and ADDA [36].", + "bbox": [ + 75, + 267, + 467, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For fairness, we do not include existing multi-view methods [4, 12, 19] in this comparison. This is because these methods require 1) multi-view labels and 2) camera parameters, whereas our approach is unsupervised and does not require such parameters. In contrast, all the comparison methods are unsupervised, leading to a fair comparison.", + "bbox": [ + 75, + 388, + 467, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In detail, SFDAHPE [32], RegDA [17] are developed for pose estimation. DAGEN [11] and ADDA [36] are originally proposed for gaze estimation and classification, respectively. We include these two methods here to show potential of these state-of-the-art methods in enhancing the cross-dataset performance of hand pose estimation. To make a fair comparison, their original networks are replaced with the same DetNet [46] as our baseline.", + "bbox": [ + 75, + 479, + 467, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results of different methods are shown in Tab. 3. Our method not only significantly outperforms the state-of-the-art methods, but also shows an advantage of being source-free. The superior performance verifies the effectiveness of the proposed S2DHand for single-to-dual-view adaptation under cross-dataset settings. For reference, we also provide the result of fine-tuning as the upper bound of this cross-dataset task.", + "bbox": [ + 75, + 603, + 467, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 738, + 225, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the $\\mathcal{D}_{ah} \\to \\mathcal{D}_{ah}$ task for clearer observation. The components are shown below:", + "bbox": [ + 75, + 762, + 467, + 823 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ABM: Attention-based merging module, which generates pseudo-labels based on the cross-view consensus.", + "- RGR: Rotation-guided refinement module, which generates pseudo-labels based on the invariance of rotation transformation between both camera coordinate systems." + ], + "bbox": [ + 76, + 825, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b78d688e896a52c3b26bcc377abcb9c28f047435337c2b6d438eb23378c689de.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art adaptation methods under cross-dataset settings. \"SF\" indicates if the method is source-free (requiring no data from source dataset, $\\mathcal{D}_{syn}$ ). * denotes that labels from the target dataset ( $\\mathcal{D}_{ah}$ ) are needed." + ], + "table_footnote": [], + "table_body": "
\\(D_{syn}\\rightarrow\\mathcal{D}_{ah}\\)\\(\\begin{array}{c}\\mathsf{SF}\\\\ \\end{array}\\)\\(\\frac{\\mathcal{D}_{ah}-\\text{Headset1}}{\\text{Mono-M}}\\)\\(\\frac{\\mathcal{D}_{ah}-\\text{Headset2}}{\\text{Mono-M}}\\)
Dual-MDual-M
Source Only56.1850.5958.74
Fine-tune*45.0338.1147.75
ADDA [36]X56.9048.4857.87
DAGEN [11]X55.3749.7257.62
RegDA [17]X51.4147.8554.75
SFDAHPE [32]✓54.0649.1157.22
S2DHand (Ours)✓48.4445.9253.11
50.37
", + "bbox": [ + 504, + 148, + 888, + 289 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4bf5b0cb1ab244b35d04aee82f0d9fe6950cc24d2e99f126a1874a1e5a7635b4.jpg", + "table_caption": [ + "Table 4. Ablation study of our method on $\\mathcal{D}_{ah} \\rightarrow \\mathcal{D}_{ah}$ task. ABM and RGR stand for the two pseudo-labeling modules, respectively." + ], + "table_footnote": [], + "table_body": "
ABMRGR\\( \\mathcal{D}_{ah} - Headset1 \\)\\( \\mathcal{D}_{ah} - Headset2 \\)
Mono-MDual-MMono-MDual-M
XX24.9622.6725.7223.75
✓X20.8120.5422.2421.71
X✓21.8921.3323.5422.75
✓✓20.1619.9821.6721.25
", + "bbox": [ + 514, + 329, + 877, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 shows the hand pose estimation errors under different combinations. We observe that both ABM and RGR can significantly improve the hand pose estimation performance over the pre-trained baseline (first row). Our final version achieves the best results for all metrics, confirming the optimality of our method.", + "bbox": [ + 498, + 438, + 890, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Number of input image pairs", + "text_level": 1, + "bbox": [ + 500, + 540, + 756, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To find the optimal number of input image pairs (i.e., $N$ ) for our method, we evaluate the S2DHand's performance under different numbers of input image pairs. Specifically, the experiments are conducted on the cam 2,3 pair in the $\\mathcal{D}_{ah} \\to \\mathcal{D}_{ah}$ task. The results are illustrated in Fig. 7. It indicates that the performance of S2DHand constantly improves as the number of input image pairs increase. Our method's performance converges when $N \\geq 1000$ . Consequently, we choose $N = 1000$ for our S2DHand.", + "bbox": [ + 498, + 564, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.7. Complement between two pseudo-labels", + "text_level": 1, + "bbox": [ + 500, + 709, + 841, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8 demonstrates the complementary nature, as stated in Sec. 4.4, of the pseudo-labeling. Using camera pair $cam1,2 - Headset1$ under in-dataset setting, we analyze the error of pseudo-labels with and without the refinement term $\\hat{\\mathbf{y}}_{rgr}$ , in relation to the prediction error of $\\overline{\\mathbf{J}}$ . Note that $\\overline{\\mathbf{J}}$ is the prediction used to compute these pseudolabels. For better observation, the predictions are first divided into seven equal intervals according to their errors, [9.4, 34.0), [34.0, 58.6), ..., [157.1, 181.7]. Then, the average pseudo-label error for each interval is calculated.", + "bbox": [ + 498, + 734, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 8, the bars are placed in the middle of each inter", + "bbox": [ + 517, + 885, + 888, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "683", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de0074811df19a878036390d2f064e64b1058f0cef65e5aeb73c10debbe6c2ec.jpg", + "image_caption": [ + "Figure 7. Evaluation of S2DHand's performance with gradually increasing the number of input image pairs." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 467, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d5e1e6eaef800f5f16f7b50d2195b5a21f77babf9cdc010bb201ab04554bc9a.jpg", + "image_caption": [ + "Figure 8. The error of pseudo-labels with and without $\\hat{\\mathbf{y}}_{rgr}$ , in relation to the error of prediction $\\overline{\\mathbf{J}}$ , where $\\overline{\\mathbf{J}}$ is what we use to compute the pseudo-labels. MPJPE in millimeter is the metric." + ], + "image_footnote": [], + "bbox": [ + 89, + 244, + 454, + 354 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f8c398a591518bc145c4cbba015b8830104007820e8a3697afd2b211a1bc54ac.jpg", + "table_caption": [ + "Table 5. Performance of our method with different hyperparameters $\\alpha$ (Eq. (7)) and $\\beta$ (Eq. (4))." + ], + "table_footnote": [], + "table_body": "
\\( {\\mathcal{D}}_{ah} \\rightarrow {\\mathcal{D}}_{ah} \\)\\( {\\mathcal{D}}_{ah} - Headset1 \\)\\( {\\mathcal{D}}_{ah} - Headset2 \\)
Mono-MDual-MMono-MDual-M
\\( \\alpha = {0.3} \\)20.8320.5422.3921.82
\\( \\alpha = {0.5} \\)20.3420.1721.9021.49
\\( \\alpha = {0.7} \\)20.1619.9821.6721.25
\\( \\alpha = {0.9} \\)20.4620.1822.1721.53
\\( \\beta = 1 \\)21.0220.9422.7722.56
\\( \\beta = e \\)20.6620.4722.5022.10
\\( \\beta = \\infty \\)20.1619.9821.6721.25
", + "bbox": [ + 89, + 439, + 457, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "val, with y-axis representing the pseudo-label errors. The result indicates that the $\\hat{\\mathbf{y}}_{rgr}$ term is redundant for accurate predictions ( $< 60mm$ ) but significantly reduces pseudolabel errors for larger prediction errors ( $\\geq 60mm$ ). This finding supports the statement in Sec. 4.4 about the importance of $\\hat{\\mathbf{y}}_{rgr}$ for complementing pseudo-labels in cases of inaccurate predictions.", + "bbox": [ + 75, + 580, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.8. Hyper-parameters", + "text_level": 1, + "bbox": [ + 76, + 696, + 256, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate how the S2DHand's performance varies with the change of weight parameter $\\alpha$ (Eq. (7)). $\\alpha$ controls the weights for averaging the two parts of pseudo-labeling. We test four values of $\\alpha$ , 0.3, 0.5, 0.7, and 0.9. The results are shown in the row 1-4 of Tab. 5, where our method achieves the best performance when $\\alpha = 0.7$ .", + "bbox": [ + 75, + 719, + 468, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also test the performance with varying $\\beta$ (Eq. (4)). $\\beta$ is the parameter in the attention-generating function, which generates the joint-wise attention in attention-based merging module. In fact, the $\\beta$ acts the role of $e$ in the softmax function. When $\\beta = 1, e, \\infty$ , the merging becomes a simple average (where attention becomes invalidated), soft-", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30eab19f249ea218839214664e88d7a687d80e8efed123cd6ee6435b0e4b1727.jpg", + "image_caption": [ + "Figure 9. Visual examples of estimated 3D hand poses under both views. The joints are portrayed by projecting the final 3D predictions to the image plane." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 874, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "max function, and maximum function, respectively. We can see that the S2DHand achieves the best performance when $\\beta = \\infty$ . This suggests that selecting the prediction with higher confidence as a pseudo-label in a joint-wise manner is the most effective strategy. Consequently, we set $\\alpha = 0.7$ and $\\beta = \\infty$ for all the experiments.", + "bbox": [ + 498, + 356, + 893, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.9. Qualitative result", + "text_level": 1, + "bbox": [ + 500, + 458, + 669, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To understand how our method improves the performance of hand pose estimation under dual-view settings, we visually present typical cases by portraying 3D hand joints onto the input image pairs. In detail, the 3D joints are projected to the image plane, with the visualization depicted in Fig. 9.", + "bbox": [ + 496, + 481, + 890, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Notably, when confronted with extreme view angles (see the left pair), the predictions of baseline model tend to be unreliable. Conversely, our technique gives a prediction much closer to the actual hand shape after adaptation. In the 3rd column, even when the hand is partially out of field-of-view, leading to a truncated hand, our S2DHand continues to deliver trustworthy predictions. These results indicate that S2DHand can utilize additional information from dual views to provide significant improvements even under extreme challenging cases.", + "bbox": [ + 496, + 558, + 892, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 723, + 617, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a novel single-to-dual-view adaptation framework (S2DHand), designed to adapt a single-view hand pose estimator to dual-view settings. The S2DHand is unsupervised, eliminating the need for multi-view labels. Our method also requires no camera parameters, enabling compatibility with arbitrary dual views. Two stereo constraints are employed as two pseudo-labeling modules in an complementary manner. Our method achieves significant performance gains across all dual-view pairs under both in-dataset and cross-dataset settings.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "684", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Charles George Broyden. The convergence of a class of double-rank minimization algorithms 1. general considerations. IMA Journal of Applied Mathematics, 6(1):76-90, 1970. 5", + "[2] Minjie Cai, Feng Lu, and Yoichi Sato. Generalizing hand segmentation in egocentric videos with uncertainty-guided model adaptation. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 14392-14401, 2020. 4", + "[3] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, Rohith Mv, Stefan Stojanov, and James M Rehg. Unsupervised 3d pose estimation with geometric self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5714-5724, 2019. 2", + "[4] Liangjian Chen, Shih-Yao Lin, Yusheng Xie, Yen-Yu Lin, and Xiaohui Xie. Mvhm: A large-scale multi-view hand mesh benchmark for accurate 3d hand pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 836-845, 2021. 1, 2, 7", + "[5] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 1", + "[6] Fabian Dubourvieux, Romaric Audigier, Angelique Loesch, Samia Ainouz, and Stephane Canu. Unsupervised domain adaptation for person re-identification through source-guided pseudo-labeling. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4957–4964. IEEE, 2021. 4", + "[7] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M. Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23600-23611, 2023. 1, 2", + "[8] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9561-9568. IEEE, 2020. 1", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 6", + "[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 1", + "[11] Zidong Guo, Zejian Yuan, Chong Zhang, Wanchao Chi, Yonggen Ling, and Shenghao Zhang. Domain adaptation gaze estimation by embedding with prediction consistency." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the Asian Conference on Computer Vision, 2020. 7", + "[12] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. Umetrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 7", + "[13] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9164–9170. IEEE, 2020. 1", + "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5", + "[15] Markus Höll, Markus Oberweger, Clemens Arth, and Vincent Lepetit. Efficient physics-based implementation for realistic hand-object interaction in virtual reality. In 2018 IEEE conference on virtual reality and 3D user interfaces (VR), pages 175-182. IEEE, 2018. 1", + "[16] Linzhi Huang, Yulong Li, Hongbo Tian, Yue Yang, Xianggang Li, Weihong Deng, and Jieping Ye. Semi-supervised 2d human pose estimation driven by position inconsistency pseudo label correction module. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 693-703, 2023. 2", + "[17] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021. 2, 7", + "[18] Wolfgang Kabsch. A discussion of the solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 34(5):827-828, 1978. 4", + "[19] Leyla Khaleghi, Alireza Sepas-Moghaddam, Joshua Marshall, and Ali Etemad. Multi-view video-based 3d hand pose estimation. IEEE Transactions on Artificial Intelligence, 2022. 1, 2, 7", + "[20] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 2", + "[21] Qiuxia Lin, Linlin Yang, and Angela Yao. Cross-domain 3d hand pose estimation with dual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17184-17193, 2023. 2", + "[22] Qunming Liu, Xiaodong Li, Xiaofei Zhang, Xiaojun Tan, and Bodong Shi. Multi-view joint learning and bev feature-fusion network for 3d object detection. Applied Sciences, 13 (9):5274, 2023. 2", + "[23] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. arXiv preprint arXiv:2312.15644, 2023. 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "685", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Ruicong Liu, Yiwei Bao, Mingjie Xu, Haofei Wang, Yunfei Liu, and Feng Lu. Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082, 2022. 2", + "[25] Ruicong Liu, Yunfei Liu, Haofei Wang, and Feng Lu. Pnpga+: Plug-and-play domain adaptation for gaze estimation using model variants. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4", + "[26] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 4", + "[27] Yunfei Liu, Ruicong Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with outlier-guided collaborative adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3835-3844, 2021. 4", + "[28] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Generated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6", + "[29] Takehiko Ohkawa, Yu-Jhe Li, Qichen Fu, Ryosuke Furuta, Kris M Kitani, and Yoichi Sato. Domain adaptive hand keypoint and pixel localization in the wild. In European Conference on Computer Vision, pages 68-87. Springer, 2022. 2, 4", + "[30] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 2, 3, 6", + "[31] Vladimir I Pavlovic, Rajeev Sharma, and Thomas S. Huang. Visual interpretation of hand gestures for human-computer interaction: A review. IEEE Transactions on pattern analysis and machine intelligence, 19(7):677-695, 1997. 1", + "[32] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4826-4836, 2023. 2, 7", + "[33] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems, pages 955-960. 2013. 1", + "[34] Siddharth S Rautaray and Anupam Agrawal. Vision based hand gesture recognition for human computer interaction: a survey. Artificial intelligence review, 43:1-54, 2015. 1", + "[35] Dripta S Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14996-15006, 2023. 2", + "[36] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Stefan Vogt, Giovanni Buccino, Afra M Wohlschlager, Nicola Canessa, N Jon Shah, Karl Zilles, Simon B Eickhoff, Hans-Joachim Freund, Giacomo Rizzolatti, and Gereon R Fink. Prefrontal involvement in imitation learning of hand actions: effects of practice and expertise. Neuroimage, 37 (4):1371-1383, 2007. 1", + "[38] Christian Von Hardenberg and François Bérard. Bare-hand human-computer interaction. In Proceedings of the 2001 workshop on Perceptive user interfaces, pages 1-8, 2001. 1", + "[39] Yilin Wen, Hao Pan, Lei Yang, Jia Pan, Taku Komura, and Wenping Wang. Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21243-21253, 2023. 1, 2", + "[40] Wenhao Wu, Hau San Wong, and Si Wu. Semi-supervised stereo-based 3d object detection via cross-view consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17471-17481, 2023. 4", + "[41] Yinsong Xu, Zhuqing Jiang, Aidong Men, Haiying Wang, and Haiyong Luo. Multi-view feature fusion for person re-identification. Knowledge-Based Systems, 229:107344, 2021. 2", + "[42] Linlin Yang, Shicheng Chen, and Angela Yao. Semi-hand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2", + "[43] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 3", + "[44] Yuxiang Zhang, Zhe Li, Liang An, Mengcheng Li, Tao Yu, and Yebin Liu. Lightweight multi-person total motion capture using sparse multi-view cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5560-5569, 2021. 2", + "[45] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2", + "[46] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 3, 4, 6, 7", + "[47] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. Technical report, arXiv:1705.01389, 2017. https://arxiv.org/abs/1705.01389. 3,6" + ], + "bbox": [ + 501, + 92, + 890, + 849 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "686", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_model.json b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f658c35060d9ba841f8214cffe65e215f48072e5 --- /dev/null +++ b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.181, + 0.29, + 0.199 + ], + "angle": 0, + "content": "Ruicong Liu" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.182, + 0.474, + 0.198 + ], + "angle": 0, + "content": "Takehiko Ohkawa" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.182, + 0.65, + 0.199 + ], + "angle": 0, + "content": "Mingfang Zhang" + }, + { + "type": "text", + "bbox": [ + 0.691, + 0.182, + 0.785, + 0.198 + ], + "angle": 0, + "content": "Yoichi Sato" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.2, + 0.611, + 0.218 + ], + "angle": 0, + "content": "The University of Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.22, + 0.721, + 0.235 + ], + "angle": 0, + "content": "{lruicong, ohkawa-t, mfzhang, ysato}@iis.u-tokyo.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.755 + ], + "angle": 0, + "content": "The pursuit of accurate 3D hand pose estimation stands as a keystone for understanding human activity in the realm of egocentric vision. The majority of existing estimation methods still rely on single-view images as input, leading to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, adding another camera to better capture the shape of hands is a practical direction. However, existing multi-view hand pose estimation methods suffer from two main drawbacks: 1) Requiring multi-view annotations for training, which are expensive. 2) During testing, the model becomes inapplicable if camera parameters/layout are not the same as those used in training. In this paper, we propose a novel Single-to-Dual-view adaptation (S2DHand) solution that adapts a pre-trained single-view estimator to dual views. Compared with existing multi-view training methods, 1) our adaptation process is unsupervised, eliminating the need for multi-view annotation. 2) Moreover, our method can handle arbitrary dual-view pairs with unknown camera parameters, making the model applicable to diverse camera settings. Specifically, S2DHand is built on certain stereo constraints, including pair-wise cross-view consensus and invariance of transformation between both views. These two stereo constraints are used in a complementary manner to generate pseudo-labels, allowing reliable adaptation. Evaluation results reveal that S2DHand achieves significant improvements on arbitrary camera pairs under both in-dataset and cross-dataset settings, and outperforms existing adaptation methods with leading performance. Project page: https://github.com/ut-vision/S2DHand." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.768, + 0.21, + 0.783 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Delving into the realm of egocentric vision (first-person view), the pursuit of refining 3D hand pose estimation stands as a keystone for understanding human activity. This quest not only forges new paths in human-computer interaction [31, 34, 38], but also empowers imitation learning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.876, + 0.471, + 0.9 + ], + "angle": 0, + "content": "This work was supported by JSPS KAKENHI Grant Number JP20H04205 and JST ACT-X Grant Number JPMJAX2007." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.273, + 0.871, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.533 + ], + "angle": 0, + "content": "Figure 1. From (a) to (b), our single-to-dual-view adaptation method adapts a traditional single-view hand pose estimator to arbitrary dual views and achieves better accuracy. (a) Traditional single-view hand pose estimation. (b) Inference process of the adapted model under a dual-view setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.894, + 0.629 + ], + "angle": 0, + "content": "[8, 13, 37]. Moreover, it enhances the immersive experience in augmented/virtual reality (AR/VR) to new heights [15, 33]. Recently, with the advancements of AR/VR headsets, egocentric data has become increasingly prevalent [5, 10], leading to an increasing demand for estimating 3D hand poses from egocentric viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.87 + ], + "angle": 0, + "content": "To achieve better 3D hand pose estimation performance, recent years have witnessed many networks with various structures [7, 39, 46]. However, the majority of existing hand pose estimation methods are still under a single-view setting, which is convenient but leads to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, a potential solution is to add another camera to expand the field-of-view and reduce depth ambiguity by capturing the hand shape from an additional view angle. Furthermore, the use of multiple cameras also aligns with industry trends, as demonstrated by the latest AR/VR headsets such as the Apple Vision Pro and Meta Quest, which feature multiple egocentric cameras. Overall, an unavoidable trend towards multi-view settings in hand pose estimation is emerging, driven by its technological advantages and the direction of industrial development." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Currently, several existing studies [4, 12, 19] have paid attention to hand pose estimation under multi-view settings." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "677" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Differences among traditional single-view methods, multi-view training methods, and our proposed single-to-dual-view adaptation method. Green indicates lower training requirements, enhanced testing results, and reduced camera parameter requirements, respectively." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.121, + 0.891, + 0.177 + ], + "angle": 0, + "content": "
MethodsPre-training dataset requiredTestCamera parameters
Traditional single-view methods [7, 39, 46]Single-view (common single-camera setting)Single-viewNot required
Multi-view training methods [4, 12, 19]Dual-view (need multi-camera setting)Dual-view & same camera poses with trainingRequired & same with training
Single-to-dual-view adaptation (Ours)Single-view (common single-camera setting)Dual-view & arbitrary camera posesNot required
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.47, + 0.333 + ], + "angle": 0, + "content": "These methods typically process input images from multiple views simultaneously, utilizing a feature fusion module to arrive at a final prediction [22, 41]. However, all these methods have two significant drawbacks that limit their applicability. 1) The training, especially for the feature fusion module, necessitates multi-view labels, which are costly to annotate. 2) During testing, the same camera parameters as in training must be used. An estimator trained under a specific multi-camera setup becomes inapplicable if there are any changes to the camera layout or parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.471, + 0.654 + ], + "angle": 0, + "content": "Unlike existing multi-view training methods, we propose a new solution that adapts an estimator from single-view to dual-view without needing multi-view labels or camera parameters. As shown in Fig. 1, given a pre-trained estimator, our method adapts it to an arbitrary dual-view setting (from (a) to (b)), where two cameras are placed in any layout without knowing their parameters. Here, all we need is a pre-trained estimator and a sufficient number of unlabeled dual-view inputs from the two cameras. As compared in Tab. 1 (row 2-3), in contrast to multi-view training, our method only needs common and cheaper single-view data for training. During testing, unlike existing methods, our method is compatible with arbitrary dual-view pairs, making the model applicable to flexible and changeable camera settings. Specifically, when camera settings change, it is easy and swift to repeat our method's adaptation process to re-adapt the pre-trained estimator to work well with new camera parameters. For camera parameters, existing methods not only need them for training, but also require them the same with testing. Conversely, our method is clearly more practical since no camera parameters are required." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.838 + ], + "angle": 0, + "content": "Building on these advancements, we present a novel unsupervised Single-to-Dual-view adaptation framework (S2DHand) for egocentric 3D hand pose estimation. It uses certain stereo constraints for adaptation, including cross-view consensus (pair-wise) and invariance of transformation between both camera coordinate systems (to all input pairs). These two stereo constraints are used in a complementary manner to refine the accuracy of pseudo-labels, allowing the model to better fit to the dual views. Specifically, the cross-view consensus is leveraged through an attention-based merging module, and the invariance of transformation is utilized via a rotation-guided refinement module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We evaluate our method by adapting a pre-trained estimator to several dual-camera pairs placed in arbitrary poses [30]. Our evaluation encompasses both in-dataset and cross-dataset scenarios. Experimental results reveal that our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.228 + ], + "angle": 0, + "content": "technique not only realizes notable improvements across all pairs but also surpasses state-of-the-art adaptation methods. The primary contributions of this paper are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.229, + 0.892, + 0.304 + ], + "angle": 0, + "content": "- We propose a novel unsupervised single-to-dual-view adaptation (S2DHand) solution for egocentric 3D hand pose estimation. Our method can adapt a traditional single-view estimator for arbitrary dual views without requiring annotations or camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.892, + 0.379 + ], + "angle": 0, + "content": "- We build a pseudo-label-based strategy for adaptation. It leverages cross-view consensus and invariance of transformation between both camera coordinate systems for reliable pseudo-labeling. This leads to two key modules: attention-based merging and rotation-guided refinement." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.38, + 0.892, + 0.44 + ], + "angle": 0, + "content": "- Evaluation results demonstrate the benefits of our approach for arbitrarily placed camera pairs. Our method achieves significant improvements for all pairs both under in-dataset and cross-dataset settings." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.229, + 0.892, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.642, + 0.47 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.48, + 0.787, + 0.496 + ], + "angle": 0, + "content": "2.1. Multi-view hand pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.609 + ], + "angle": 0, + "content": "Multi-view hand pose estimation accepts multi-view images as input and outputs a final 3D hand pose, which remains a relatively unexplored research area. Previous works [4, 12, 19] design various network structures to predict 3D hand poses through a multi-view fusion module. Similarly, multiple studies have been done [44, 45] through feature fusion for body pose estimation under multi-view settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.701 + ], + "angle": 0, + "content": "All these studies have two limitations: 1) they require costly multi-view images and annotations for training, 2) during testing, camera poses are assumed to be known and covered by training data, thereby limiting their applicability. In contrast, our method eliminates the need for multi-view annotations and is adaptable to arbitrary dual views." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.809, + 0.727 + ], + "angle": 0, + "content": "2.2. Adaptation in hand pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Adaptation aims at tailoring a model for specific application scenarios [20, 24, 29]. Existing adaptation methods in hand pose estimation mainly focus on adaptation across different domains (datasets), e.g., entropy minimization [17, 32], consistency regularization [3, 21, 29], and pseudo-labeling methods [16, 35, 42]. Prior works only use constraints in single-view settings for adaptation, e.g., bio-mechanical constraint [21, 42], and are thus limited to single-view inference. Unlike these methods, we propose stereo constraints from dual views for adaptation and supports dual-view inference, which extends the application scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.468, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.356 + ], + "angle": 0, + "content": "Figure 2. Problem setting of single-to-dual-view adaptation for hand pose estimation. (a) The input and output of adaptation. (b) The dual-view testing scheme after adaptation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.24, + 0.377 + ], + "angle": 0, + "content": "3. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Fig. 2 illustrates the task setting of single-to-dual-view adaptation for hand pose estimation. We denote unlabeled dual-view data as \\(\\mathcal{D} = \\{\\mathbf{x}_i^{v1},\\mathbf{x}_i^{v2}|_{i = 1}^N\\}\\), where \\(\\mathbf{x}_i^{v1}\\) and \\(\\mathbf{x}_i^{v2}\\) denote the \\(i\\)-th image from view1 and view2, respectively, \\(N\\) is the number of image pairs. The dual-view data \\(\\mathcal{D}\\) contains no ground-truth hand poses or camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.627 + ], + "angle": 0, + "content": "As shown in Fig. 2 (a), suppose we have a baseline hand pose estimator \\(\\mathcal{H}(\\cdot|\\theta)\\) with parameters \\(\\theta\\) pre-trained from common single-view data. Leveraging \\(\\mathcal{D}\\) can enhance its performance, as \\(\\mathcal{D}\\) provides additional information from a dual-view setup. Our objective is to adapt this pre-trained estimator, \\(\\mathcal{H}(\\cdot|\\theta)\\), to an arbitrary yet fixed dual-view setting (with unknown camera poses) without needing groundtruths or camera parameters. By inputting \\(\\mathcal{H}(\\cdot|\\theta)\\) and \\(\\mathcal{D}\\) into our method, it outputs an adapted estimator \\(\\mathcal{H}(\\cdot|\\theta')\\) with parameters \\(\\theta'\\) tailored for the dual-view scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Upon adapting the estimator, its inference mechanism is correspondingly tailored for dual-view scenarios (Fig. 1 (b)). During testing, the adapted estimator \\(\\mathcal{H}(\\cdot |\\theta^{\\prime})\\) processes a dual-view input pair \\((\\mathbf{x}^{v1},\\mathbf{x}^{v2})\\) and produces two predictions \\((\\mathbf{J}^{v1},\\mathbf{J}^{v2})\\), where each \\(\\mathbf{J}^v\\in \\mathbb{R}^{21\\times 3}\\) represents the 21 3D joints of the hand. These predictions denote the 3D hand joints for each view and can be combined together to generate a final output, e.g., through a simple average." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Camera layout for a multi-view headset. Fig. 3 illustrates an example of headset-mounted camera setups for multiview egocentric data capture, with four cameras at each corner for different views. The top-right of Fig. 3 displays images from these cameras. Six distinct dual-view pairs can be created from these four views. As a supplement, the bottom of Fig. 3 shows the synthetic training data, highlighting variations in style and lighting. Such data helps to explore the performance of our method under cross-dataset or simulate-to-real settings. See Sec. 5.1 for dataset details." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.093, + 0.86, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.398, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 3. Top: Headset and its camera layout to collect multi-view data, and samples from the four views. Bottom: Samples of synthetic data. Image samples are from AssemblyHands [30], GANerated Hands [28], and Rendered Handpose [47], respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.673, + 0.478 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.715 + ], + "angle": 0, + "content": "We propose a novel unsupervised single-to-dual-view adaptation framework (S2DHand). Before adaptation, an initialization step is performed to initialize the rotation matrix between both views (Sec. 4.1). The rotation matrix is essential to establish the transformation between two camera coordinate systems. The architecture overview, as illustrated in Fig. 4, comprises two branches, an estimator \\(\\mathcal{H}\\) and a its momentum version \\(\\overline{\\mathcal{H}}\\). The adaptation process is designed from two stereo constraints, pair-wise cross-view consensus and invariant rotation transformation between both camera coordinate systems. This leads to two key pseudo-labeling modules: attention-based merging and rotation-guided refinement (Secs. 4.3 and 4.4). Notably, these two modules function in a complementary manner, depending on the prediction accuracy, ensuring reliable pseudo-labeling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.637, + 0.74 + ], + "angle": 0, + "content": "4.1. Initialization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The initialization step aims to estimate a relatively accurate rotation matrix \\(R\\), since \\(R\\) is necessary to link the two camera coordinate systems [23, 43]. It should be noted that translation vector between the two cameras is not necessary, as the predicted hand poses are usually aligned by the wrist during testing [30, 46]. Assuming that the initial pre-trained estimator is sufficiently accurate to generate reasonable predictions, we estimate the \\(R\\) using the predictions of unlabeled dual-view data \\(\\mathcal{D} = \\{\\mathbf{x}_i^{v1}, \\mathbf{x}_i^{v2}|_{i=1}^N\\}\\). Given \\(\\mathcal{D}\\), the estimator \\(\\mathcal{H}\\) can output \\(N\\) pairs of predictions \\(\\{\\mathbf{J}_i^{v1}, \\mathbf{J}_i^{v2}|_{i=1}^N\\}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.089, + 0.825, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Figure 4. Overview of the proposed S2DHand, image pairs captured from arbitrarily placed dual cameras are input for adaptation. The architecture of S2DHand is illustrated in the dark dashed box, which contains a dynamically updated estimator and a momentum estimator. The momentum estimator's predictions are used to generate pseudo-labels, which are then processed by our pseudo-labeling module (Secs. 4.3 and 4.4). Using the pseudo-labels, a loss function is computed to update the estimator. The rotation matrix \\( R \\) from the initialization step (Sec. 4.1) is required for the pseudo-labeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.48 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}_i^v\\in \\mathbb{R}^{21\\times 3}\\) (21 is the number of 3D joints). Then, the rotation matrix \\(R\\) is estimated by:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.492, + 0.47, + 0.533 + ], + "angle": 0, + "content": "\\[\nR ^ {(0)} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} r o t \\left(\\mathbf {J} _ {i} ^ {v 1}, \\mathbf {J} _ {i} ^ {v 2}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.47, + 0.621 + ], + "angle": 0, + "content": "where the superscript of \\( R \\) denotes the iteration number, (0) indicates that it is before the first iteration. The rot function [18] generates a \\( 3 \\times 3 \\) rotation matrix from two \\( 21 \\times 3 \\) joint predictions. Note that the average in Eq. (1) is not elementwise, but an average of rotation matrices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.35, + 0.648 + ], + "angle": 0, + "content": "4.2. Single-to-dual-view adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.761 + ], + "angle": 0, + "content": "With the initialized \\(R\\), the adaptation process begins. The S2DHand framework comprises two branches, an estimator \\(\\mathcal{H}(\\cdot |\\theta)\\) with dynamically updating parameters \\(\\theta\\), and its momentum version \\(\\overline{\\mathcal{H}} (\\cdot |\\overline{\\theta})\\), which updates its parameters \\(\\bar{\\theta}\\) using temporal moving average. Temporal moving average has been proved by many works [6, 14, 25, 27] that can help to stabilize the training process. The \\(\\bar{\\theta}\\) is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.771, + 0.469, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\bar {\\theta} ^ {(T)} = \\eta_ {\\theta} \\bar {\\theta} ^ {(T - 1)} + (1 - \\eta_ {\\theta}) \\theta , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.805, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\(\\overline{\\theta}^{(T - 1)}\\) indicates the temporal averaged parameters in the previous iteration \\(T - 1\\), and \\(\\eta_{\\theta}\\) represents the ensembling momentum, which is set as 0.99 [25, 27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 4, during the single-to-dual-view adaptation, the role of the momentum model \\(\\overline{\\mathcal{H}}\\) is to generate pseudo-labels, which are then utilized to supervise the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.528 + ], + "angle": 0, + "content": "model \\(\\mathcal{H}\\). The pseudo-labeling module (Secs. 4.3 and 4.4) outputs pseudo-labels \\(\\hat{\\mathbf{y}}^{v1},\\hat{\\mathbf{y}}^{v2}\\) based on the predictions \\(\\overline{\\mathbf{J}}^{v1},\\overline{\\mathbf{J}}^{v2}\\) from \\(\\overline{\\mathcal{H}}\\). These pseudo-labels are then used to supervise the predictions \\(\\mathbf{J}^{v1},\\mathbf{J}^{v2}\\) of \\(\\mathcal{H}\\). The loss function is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.535, + 0.892, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| \\mathbf {J} ^ {v 1} - \\hat {\\mathbf {y}} ^ {v 1} \\right\\| _ {2} + \\left\\| \\mathbf {J} ^ {v 2} - \\hat {\\mathbf {y}} ^ {v 2} \\right\\| _ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.638 + ], + "angle": 0, + "content": "The estimator follows the implementation of DetNet [46], where \\(\\mathcal{H}\\) directly outputs heatmaps, and \\(\\mathbf{J}\\) is calculated from the heatmaps. Therefore, the loss function is actually computed from corresponding heatmaps, here we write these 3D-joint variables only for better understanding." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.644, + 0.858, + 0.661 + ], + "angle": 0, + "content": "4.3. Pseudo-labeling: attention-based merging" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.892, + 0.793 + ], + "angle": 0, + "content": "The attention-based merging (ABM) module, which constitutes the first part of pseudo-labeling, is derived from cross-view consensus. Cross-view consensus refers to the concept of achieving agreement or consistency between different views of the same data [40]. Theoretically, when transformed into the same coordinate system, the two predictions \\(\\overline{\\mathbf{J}}^{v1}\\) and \\(\\overline{\\mathbf{J}}^{v2}\\) from different views should be identical, i.e., \\(R\\overline{\\mathbf{J}}^{v1} = \\overline{\\mathbf{J}}^{v2}\\), with \\(\\overline{\\mathbf{J}}^{v1}\\overline{\\mathbf{J}}^{v2}\\) being aligned with wrist joint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This stereo constraint is the foundation for this module to generate accurate pseudo-labels. Prior works utilize a simple average [26, 29] (e.g., \\((R\\overline{\\mathbf{J}}^{v1} + \\overline{\\mathbf{J}}^{v2}) / 2\\)) or sample-wise confidence [2, 29] to improve the quality of pseudolabels. However, these approaches overlook the varying confidence in joints that is caused by differences in image capture across views. For instance, a joint that is occluded" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "680" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.445, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Figure 5. Top: illustration of the first part of pseudo-labeling: attention-based merging module. The generating process of \\(\\hat{y}_{abm}^{v2}\\) in view2 is shown as an example, the process of view1 is the same. Bottom: visualizations of heatmaps with different accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.469, + 0.422 + ], + "angle": 0, + "content": "in one view but fully visible in another could lead to reliable predictions being hindered by unreliable ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.558 + ], + "angle": 0, + "content": "To address this, we propose joint-wise attention \\(\\mathbf{w} \\in \\mathbb{R}^{21 \\times 1}\\) to represent each joint's prediction confidence. It is derived from the 2D heatmap \\(\\overline{\\mathbf{h}} \\in \\mathbb{R}^{21 \\times 32 \\times 32}\\) output from \\(\\overline{\\mathcal{H}}\\). The \\(\\overline{\\mathbf{h}}\\) indicates the probability of each joint's presence at every pixel in the 2D image space. This approach is from an observation (bottom of Fig. 5): as the error of prediction increases, the intensity of the heatmap's hotspot decreases, i.e., darker indicates low accuracy. Inspired by this, we propose an attention-generating function:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.566, + 0.469, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {j} ^ {v} = \\frac {\\beta^ {\\max \\left(\\overline {{\\mathbf {h}}} _ {j} ^ {v}\\right)}}{\\sum_ {v \\in \\{v 1 , v 2 \\}} \\beta^ {\\max \\left(\\overline {{\\mathbf {h}}} _ {j} ^ {v}\\right)}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.619, + 0.469, + 0.663 + ], + "angle": 0, + "content": "where the subscript \\( j \\) indicate the index of joint, i.e., \\( j = 1,2,\\ldots ,21 \\). We introduce a hyper-parameter \\( \\beta \\) here to adjust softness. Please refer to Sec. 5.8 for parameter choosing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.469, + 0.755 + ], + "angle": 0, + "content": "The workflow of this attention-based merging module is illustrated at the top of Fig. 5. First, we transform both predictions into the same camera coordinate system. Then, a joint-wise multiplication is performed for each of them using the attention \\(\\mathbf{w}^{v1},\\mathbf{w}^{v2}\\). Finally, the pseudo-label \\(\\hat{\\mathbf{y}}_{abm}\\) is calculated through a summation operation. In summary:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.764, + 0.469, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {a b m} ^ {v 1} = \\mathbf {w} ^ {v 1} \\overline {{\\mathbf {J}}} ^ {v 1} + \\mathbf {w} ^ {v 2} R ^ {T} \\overline {{\\mathbf {J}}} ^ {v 2}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.789, + 0.369, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {a b m} ^ {v 2} = \\mathbf {w} ^ {v 1} R \\overline {{\\mathbf {J}}} ^ {v 1} + \\mathbf {w} ^ {v 2} \\overline {{\\mathbf {J}}} ^ {v 2}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.454, + 0.833 + ], + "angle": 0, + "content": "4.4. Pseudo-labeling: rotation-guided refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The rotation-guided refinement (RGR) module is based on another stereo constraint: invariance of rotation transformation between both views. This implies that the estimated rotation matrix should remain unchanged since the cameras" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.88, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.197, + 0.892, + 0.224 + ], + "angle": 0, + "content": "Figure 6. Illustration of the second part of pseudo-labeling: rotation-guided refinement module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.225, + 0.892, + 0.295 + ], + "angle": 0, + "content": "are fixed, i.e., theoretically \\( rot(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2}) = C \\). In light of this, this module aims to refine the predictions \\( \\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2} \\) such that \\( rot(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2}) \\) becomes invariant across all input data. The refinement result becomes the pseudo-label \\( \\hat{\\mathbf{y}}_{rgr} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.386 + ], + "angle": 0, + "content": "The workflow is shown in Fig. 6. Given a pair of predictions, our method estimates a new rotation matrix \\( R' = \\text{rot}(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2}) \\). Subsequently, the more accurately estimated \\( R \\) from the initialization step is set as the target for refinement, aiming to minimize \\( \\| R - R' \\|_F \\). The refinement process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.393, + 0.891, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {r g r} ^ {v 1}, \\hat {\\mathbf {y}} _ {r g r} ^ {v 2} = \\underset {\\overline {{\\mathbf {J}}} ^ {v 1}, \\overline {{\\mathbf {J}}} ^ {v 2}} {\\arg \\min } \\| R - r o t (\\overline {{\\mathbf {J}}} ^ {v 1}, \\overline {{\\mathbf {J}}} ^ {v 2}) \\| _ {F}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.88, + 0.447 + ], + "angle": 0, + "content": "In detail, we employ BFGS [1] algorithm for minimizing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Final pseudo-label. Finally, the pseudo-label is calculated from a weighted average of \\(\\hat{\\mathbf{y}}_{abm}\\) and \\(\\hat{\\mathbf{y}}_{rgr}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.486, + 0.891, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} = \\alpha \\hat {\\mathbf {y}} _ {a b m} + (1 - \\alpha) \\hat {\\mathbf {y}} _ {r g r}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.891, + 0.554 + ], + "angle": 0, + "content": "Here, we introduce another pre-fixed hyper-parameter \\(\\alpha\\), which adjusts the weight of the two parts of the pseudolabel. Empirically, we set \\(\\alpha = 0.7\\) (see Sec. 5.8)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.892, + 0.694 + ], + "angle": 0, + "content": "Complement between two pseudo-labels. When the predictions \\(\\overline{\\mathbf{J}}^{v1}\\) and \\(\\overline{\\mathbf{J}}^{v2}\\) are accurate, \\(R^{\\prime}\\) closely approximates \\(R\\), making \\(\\hat{\\mathbf{y}}_{rgr}\\) redundant. In such cases, \\(\\hat{\\mathbf{y}}_{abm}\\) is beneficial as it merges the two accurate predictions. Conversely, if \\(\\overline{\\mathbf{J}}^{v1}\\) and \\(\\overline{\\mathbf{J}}^{v2}\\) are unreliable, \\(\\hat{\\mathbf{y}}_{abm}\\) will consequently also lack accuracy. In such instances, \\(\\hat{\\mathbf{y}}_{rgr}\\) steps in as a complementary tool for refining pseudo-labels. By minimizing \\(\\| R - R^{\\prime}\\|_{F}\\), it optimizes the predictions towards an alignment with real-world condition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.892, + 0.801 + ], + "angle": 0, + "content": "Update rotation matrix R. Clearly, the rotation matrix \\(R\\) from the initialization step plays an important role in bridging the two camera coordinate systems. As can be expected, its accuracy significantly affects the final performance. To enhance its accuracy with each iteration, we also employ a temporal moving average for its updates. Given an input batch containing \\(B\\) image pairs, the \\(R\\) is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.809, + 0.891, + 0.849 + ], + "angle": 0, + "content": "\\[\nR ^ {(T)} = \\eta_ {R} R ^ {(T - 1)} + (1 - \\eta_ {R}) \\cdot \\frac {1}{B} \\sum_ {i = 1} ^ {B} r o t \\left(\\overline {{\\mathbf {J}}} _ {i} ^ {v 1}, \\overline {{\\mathbf {J}}} _ {i} ^ {v 2}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where the ensembling momentum \\(\\eta_{R}\\) is set as 0.999 [14]. This way, the rotation matrix \\(R\\) can be updated slowly and provide a more accurate rotation function with iteration." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "681" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 2. Adaptation results for all dual-camera pairs. The camera pairs are from \\( \\mathcal{D}_{ah} \\) dataset [30], which is divided into two parts according to the collecting headset. In \"In-dataset\" and \"Cross-dataset\" settings, the baseline model is pre-trained on \\( \\mathcal{D}_{ah} \\) and \\( \\mathcal{D}_{syn} \\) [28, 47], respectively. The adaptation yields a uniquely adapted model for each pair, and \"Overall\" averages the results across all 6 pairs." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.135, + 0.887, + 0.361 + ], + "angle": 0, + "content": "
Camera pairMethodIn-dataset (Dah → Dah)Cross-dataset (Dsyn → Dah)
Dah - Headset1Dah - Headset2Dah - Headset1Dah - Headset2
Mono-MDual-MMono-MDual-MMono-MDual-MMono-MDual-M
cam 1,2Baseline43.0039.2054.7152.3867.9360.4870.3262.26
S2DHand31.01 ▼ 27.9%31.36 ▼ 20.0%45.52 ▼ 16.8%45.14 ▼ 13.8%63.46 ▼ 6.6%59.32 ▼ 1.9%70.09 ▼ 0.3%60.97 ▼ 2.1%
cam 1,3Baseline25.0023.2922.5921.0857.7951.4264.0060.25
S2DHand19.73 ▼ 21.1%19.92 ▼ 14.5%17.90 ▼ 20.8%17.68 ▼ 16.1%50.84 ▼ 12.0%47.55 ▼ 7.5%61.81 ▼ 3.4%58.34 ▼ 3.2%
cam 1,4Baseline24.9022.7016.7314.9152.7146.5554.3250.57
S2DHand20.88 ▼ 16.1%20.87 ▼ 8.1%14.64 ▼ 12.5%14.29 ▼ 4.2%46.05 ▼ 12.6%42.50 ▼ 8.7%46.59 ▼ 14.2%45.66 ▼ 9.7%
cam 2,3Baseline17.9615.2317.1015.0853.3648.4252.8448.84
S2DHand14.97 ▼ 16.6%14.44 ▼ 5.2%14.42 ▼ 15.7%14.20 ▼ 5.8%40.26 ▼ 24.6%39.32 ▼ 18.8%43.61 ▼ 17.5%42.88 ▼ 12.2%
cam 2,4Baseline22.0919.8423.2420.9659.4454.3261.1357.41
S2DHand17.98 ▼ 18.6%17.75 ▼ 10.5%18.31 ▼ 21.2%18.41 ▼ 12.2%50.59 ▼ 14.9%49.41 ▼ 9.0%52.45 ▼ 14.2%51.48 ▼ 10.3%
cam 3,4Baseline16.8315.7719.9318.0845.8242.3449.8448.99
S2DHand16.36 ▼ 2.8%15.55 ▼ 1.4%19.25 ▼ 3.4%17.80 ▼ 1.5%39.46 ▼ 13.9%37.43 ▼ 11.6%44.04 ▼ 11.6%42.88 ▼ 12.5%
OverallBaseline24.9622.6725.7223.7556.1850.5958.7454.72
S2DHand20.16 ▼ 19.2%19.98 ▼ 11.9%21.67 ▼ 15.7%21.25 ▼ 10.5%48.44 ▼ 13.8%45.92 ▼ 9.2%53.11 ▼ 9.6%50.37 ▼ 7.9%
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.204, + 0.387 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.394, + 0.175, + 0.41 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.471, + 0.553 + ], + "angle": 0, + "content": "We employ AssemblyHands [30] \\((\\mathcal{D}_{ah})\\) as the evaluation set, as it is the newest large-scale benchmark dataset with high-quality multi-view 3D hand pose annotations. As for the training set, we set two adaptation scenarios: 1) in-dataset setting where the training set is drawn from the same dataset \\(\\mathcal{D}_{ah}\\) and 2) cross-dataset setting where we use synthetic dataset \\((\\mathcal{D}_{syn})\\) as the training set, consisting of Rendered Handpose [47] and GANerated Hands [28]. The details of the datasets are as below:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.554, + 0.472, + 0.63 + ], + "angle": 0, + "content": "- **AssemblyHands** [30] is a large-scale benchmark dataset featuring accurate 3D hand pose annotations. Collected using two AR headsets, it comprises images captured from four synchronized egocentric cameras. The dataset includes 412K training samples and 62K testing samples." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.63, + 0.471, + 0.675 + ], + "angle": 0, + "content": "- GANerated Hands [28] includes over 330K color images of hands. The images are synthetically generated and then fed to a GAN [9] to make the features closer to real hands." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.472, + 0.782 + ], + "angle": 0, + "content": "- Rendered Handpose [47] contains about 44K samples. The images are rendered with freely available characters. In detail, AssemblyHands is collected by two VR headsets, as shown in Fig. 3, each headset has four egocentric cameras at four corners. Following the collecting devices, we separate \\( \\mathcal{D}_{ah} \\) into two parts, each part is collected using one headset, namely \\( \\mathcal{D}_{ah} - Headset1/2 \\)." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.554, + 0.472, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.265, + 0.804 + ], + "angle": 0, + "content": "5.2. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Evaluation metric. We compare the predictions from our model with the ground-truth labels in root-relative coordinates, and use the common mean per joint position error (MPJPE) in millimeters as the evaluation metric. However, since our focus is on the single-to-dual-view adaptation task, the adapted estimator is expected to perform in dual-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.446 + ], + "angle": 0, + "content": "view settings. This implies that traditional MPJPE computed from single-view (monocular MPJPE, Mono-M) cannot be sufficient. As a result, we propose a new dual-view MPJPE metric Dual-M in addition to monocular MPJPE. The metrics are defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.447, + 0.892, + 0.492 + ], + "angle": 0, + "content": "- Mono-M: the traditional monocular MPJPE, which collects all the single-view errors from both views and calculates their average." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.492, + 0.894, + 0.567 + ], + "angle": 0, + "content": "- Dual-M: the proposed metric under dual views. To calculate it, first the predictions from both views are averaged using a rotation matrix \\( R \\). Then, we calculate the MPJPE of the averaged predictions as the Dual-M. Usually, the \\( R \\) is from the initialization step of our method." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.447, + 0.894, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Implementation detail. We employ PyTorch for implementation. All experiments run on a single NVIDIA A100 GPU. DetNet from [46] is adopted as the backbone of our hand pose estimator. Adam optimizer is employed with a learning rate of \\( 1 \\times 10^{-3} \\) to pre-train the 3D hand pose estimation network. For the single-to-dual-view adaptation, we use the Adam optimizer with a learning rate of \\( 5 \\times 10^{-4} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.682, + 0.836, + 0.699 + ], + "angle": 0, + "content": "5.3. Adaptation results for all camera pairs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.841 + ], + "angle": 0, + "content": "In this section, we use our method to adapt the same pretrained single-view hand pose estimator to all dual-view pairs from the evaluation set \\( \\mathcal{D}_{ah} \\) independently, yielding one adapted model for each pair. Experiments are conducted under both in-dataset and cross-dataset settings. Under the in-dataset setting (Tab. 2, \\( \\mathcal{D}_{ah} \\to \\mathcal{D}_{ah} \\)), the baseline model is pre-trained on \\( \\mathcal{D}_{ah} \\), while under the cross-dataset setting \\( (\\mathcal{D}_{syn} \\to \\mathcal{D}_{ah}) \\), the baseline model is pre-trained on \\( \\mathcal{D}_{syn} \\) before being adapted to the camera pairs from \\( \\mathcal{D}_{ah} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 2, compared with the pre-trained model (Baseline), our S2DHand offers significant accuracy gains under both settings among all camera pairs. This indicates that our method can adapt well to arbitrary dual views re" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "682" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.448, + 0.107 + ], + "angle": 0, + "content": "gardless of the camera positions or pre-training datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Quantitative results demonstrate that the S2DHand offers substantial improvements. On average, the improvement in both monocular (Mono-M) and dual-view (Dual-M) metrics exceeds \\(10\\%\\), with the maximum improvement exceeding \\(20\\%\\). Interestingly, we can see that the improvement for cam 1, 2 under the cross-dataset setting is relatively small. This indicates that low initial accuracy limits the performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.243, + 0.426, + 0.26 + ], + "angle": 0, + "content": "5.4. Comparison under cross-dataset settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.468, + 0.388 + ], + "angle": 0, + "content": "Our method is compared with state-of-the-art adaptation techniques in cross-dataset settings. Considering the prevalence and significance of cross-dataset scenarios in real-world applications, this experiment evaluates the capability of S2DHand in comparison to leading domain adaptation methods. Specifically, adaptation methods included in the comparison are: SFDAHPE [32], RegDA [17], DAGEN [11], and ADDA [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.468, + 0.48 + ], + "angle": 0, + "content": "For fairness, we do not include existing multi-view methods [4, 12, 19] in this comparison. This is because these methods require 1) multi-view labels and 2) camera parameters, whereas our approach is unsupervised and does not require such parameters. In contrast, all the comparison methods are unsupervised, leading to a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.468, + 0.602 + ], + "angle": 0, + "content": "In detail, SFDAHPE [32], RegDA [17] are developed for pose estimation. DAGEN [11] and ADDA [36] are originally proposed for gaze estimation and classification, respectively. We include these two methods here to show potential of these state-of-the-art methods in enhancing the cross-dataset performance of hand pose estimation. To make a fair comparison, their original networks are replaced with the same DetNet [46] as our baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.468, + 0.724 + ], + "angle": 0, + "content": "Quantitative results of different methods are shown in Tab. 3. Our method not only significantly outperforms the state-of-the-art methods, but also shows an advantage of being source-free. The superior performance verifies the effectiveness of the proposed S2DHand for single-to-dual-view adaptation under cross-dataset settings. For reference, we also provide the result of fine-tuning as the upper bound of this cross-dataset task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.227, + 0.755 + ], + "angle": 0, + "content": "5.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.468, + 0.824 + ], + "angle": 0, + "content": "We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the \\( \\mathcal{D}_{ah} \\to \\mathcal{D}_{ah} \\) task for clearer observation. The components are shown below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.468, + 0.854 + ], + "angle": 0, + "content": "- ABM: Attention-based merging module, which generates pseudo-labels based on the cross-view consensus." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- RGR: Rotation-guided refinement module, which generates pseudo-labels based on the invariance of rotation transformation between both camera coordinate systems." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.826, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art adaptation methods under cross-dataset settings. \"SF\" indicates if the method is source-free (requiring no data from source dataset, \\( \\mathcal{D}_{syn} \\)). * denotes that labels from the target dataset (\\( \\mathcal{D}_{ah} \\)) are needed." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.149, + 0.89, + 0.29 + ], + "angle": 0, + "content": "
\\(D_{syn}\\rightarrow\\mathcal{D}_{ah}\\)\\(\\begin{array}{c}\\mathsf{SF}\\\\ \\end{array}\\)\\(\\frac{\\mathcal{D}_{ah}-\\text{Headset1}}{\\text{Mono-M}}\\)\\(\\frac{\\mathcal{D}_{ah}-\\text{Headset2}}{\\text{Mono-M}}\\)
Dual-MDual-M
Source Only56.1850.5958.74
Fine-tune*45.0338.1147.75
ADDA [36]X56.9048.4857.87
DAGEN [11]X55.3749.7257.62
RegDA [17]X51.4147.8554.75
SFDAHPE [32]✓54.0649.1157.22
S2DHand (Ours)✓48.4445.9253.11
50.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.298, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Table 4. Ablation study of our method on \\( \\mathcal{D}_{ah} \\rightarrow \\mathcal{D}_{ah} \\) task. ABM and RGR stand for the two pseudo-labeling modules, respectively." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.33, + 0.879, + 0.426 + ], + "angle": 0, + "content": "
ABMRGR\\( \\mathcal{D}_{ah} - Headset1 \\)\\( \\mathcal{D}_{ah} - Headset2 \\)
Mono-MDual-MMono-MDual-M
XX24.9622.6725.7223.75
✓X20.8120.5422.2421.71
X✓21.8921.3323.5422.75
✓✓20.1619.9821.6721.25
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Tab. 4 shows the hand pose estimation errors under different combinations. We observe that both ABM and RGR can significantly improve the hand pose estimation performance over the pre-trained baseline (first row). Our final version achieves the best results for all metrics, confirming the optimality of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.758, + 0.557 + ], + "angle": 0, + "content": "5.6. Number of input image pairs" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.565, + 0.892, + 0.701 + ], + "angle": 0, + "content": "To find the optimal number of input image pairs (i.e., \\(N\\)) for our method, we evaluate the S2DHand's performance under different numbers of input image pairs. Specifically, the experiments are conducted on the cam 2,3 pair in the \\(\\mathcal{D}_{ah} \\to \\mathcal{D}_{ah}\\) task. The results are illustrated in Fig. 7. It indicates that the performance of S2DHand constantly improves as the number of input image pairs increase. Our method's performance converges when \\(N \\geq 1000\\). Consequently, we choose \\(N = 1000\\) for our S2DHand." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.843, + 0.727 + ], + "angle": 0, + "content": "5.7. Complement between two pseudo-labels" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Fig. 8 demonstrates the complementary nature, as stated in Sec. 4.4, of the pseudo-labeling. Using camera pair \\(cam1,2 - Headset1\\) under in-dataset setting, we analyze the error of pseudo-labels with and without the refinement term \\(\\hat{\\mathbf{y}}_{rgr}\\), in relation to the prediction error of \\(\\overline{\\mathbf{J}}\\). Note that \\(\\overline{\\mathbf{J}}\\) is the prediction used to compute these pseudolabels. For better observation, the predictions are first divided into seven equal intervals according to their errors, [9.4, 34.0), [34.0, 58.6), ..., [157.1, 181.7]. Then, the average pseudo-label error for each interval is calculated." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.89, + 0.901 + ], + "angle": 0, + "content": "In Fig. 8, the bars are placed in the middle of each inter" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "683" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.468, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.209, + 0.47, + 0.239 + ], + "angle": 0, + "content": "Figure 7. Evaluation of S2DHand's performance with gradually increasing the number of input image pairs." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.245, + 0.455, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.399 + ], + "angle": 0, + "content": "Figure 8. The error of pseudo-labels with and without \\(\\hat{\\mathbf{y}}_{rgr}\\), in relation to the error of prediction \\(\\overline{\\mathbf{J}}\\), where \\(\\overline{\\mathbf{J}}\\) is what we use to compute the pseudo-labels. MPJPE in millimeter is the metric." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.406, + 0.47, + 0.435 + ], + "angle": 0, + "content": "Table 5. Performance of our method with different hyperparameters \\( \\alpha \\) (Eq. (7)) and \\( \\beta \\) (Eq. (4))." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.44, + 0.458, + 0.574 + ], + "angle": 0, + "content": "
\\( {\\mathcal{D}}_{ah} \\rightarrow {\\mathcal{D}}_{ah} \\)\\( {\\mathcal{D}}_{ah} - Headset1 \\)\\( {\\mathcal{D}}_{ah} - Headset2 \\)
Mono-MDual-MMono-MDual-M
\\( \\alpha = {0.3} \\)20.8320.5422.3921.82
\\( \\alpha = {0.5} \\)20.3420.1721.9021.49
\\( \\alpha = {0.7} \\)20.1619.9821.6721.25
\\( \\alpha = {0.9} \\)20.4620.1822.1721.53
\\( \\beta = 1 \\)21.0220.9422.7722.56
\\( \\beta = e \\)20.6620.4722.5022.10
\\( \\beta = \\infty \\)20.1619.9821.6721.25
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.688 + ], + "angle": 0, + "content": "val, with y-axis representing the pseudo-label errors. The result indicates that the \\(\\hat{\\mathbf{y}}_{rgr}\\) term is redundant for accurate predictions (\\(< 60mm\\)) but significantly reduces pseudolabel errors for larger prediction errors (\\(\\geq 60mm\\)). This finding supports the statement in Sec. 4.4 about the importance of \\(\\hat{\\mathbf{y}}_{rgr}\\) for complementing pseudo-labels in cases of inaccurate predictions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.258, + 0.714 + ], + "angle": 0, + "content": "5.8. Hyper-parameters" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.81 + ], + "angle": 0, + "content": "We evaluate how the S2DHand's performance varies with the change of weight parameter \\(\\alpha\\) (Eq. (7)). \\(\\alpha\\) controls the weights for averaging the two parts of pseudo-labeling. We test four values of \\(\\alpha\\), 0.3, 0.5, 0.7, and 0.9. The results are shown in the row 1-4 of Tab. 5, where our method achieves the best performance when \\(\\alpha = 0.7\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We also test the performance with varying \\(\\beta\\) (Eq. (4)). \\(\\beta\\) is the parameter in the attention-generating function, which generates the joint-wise attention in attention-based merging module. In fact, the \\(\\beta\\) acts the role of \\(e\\) in the softmax function. When \\(\\beta = 1, e, \\infty\\), the merging becomes a simple average (where attention becomes invalidated), soft-" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.875, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Figure 9. Visual examples of estimated 3D hand poses under both views. The joints are portrayed by projecting the final 3D predictions to the image plane." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.357, + 0.894, + 0.449 + ], + "angle": 0, + "content": "max function, and maximum function, respectively. We can see that the S2DHand achieves the best performance when \\(\\beta = \\infty\\). This suggests that selecting the prediction with higher confidence as a pseudo-label in a joint-wise manner is the most effective strategy. Consequently, we set \\(\\alpha = 0.7\\) and \\(\\beta = \\infty\\) for all the experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.459, + 0.671, + 0.475 + ], + "angle": 0, + "content": "5.9. Qualitative result" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.558 + ], + "angle": 0, + "content": "To understand how our method improves the performance of hand pose estimation under dual-view settings, we visually present typical cases by portraying 3D hand joints onto the input image pairs. In detail, the 3D joints are projected to the image plane, with the visualization depicted in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.893, + 0.71 + ], + "angle": 0, + "content": "Notably, when confronted with extreme view angles (see the left pair), the predictions of baseline model tend to be unreliable. Conversely, our technique gives a prediction much closer to the actual hand shape after adaptation. In the 3rd column, even when the hand is partially out of field-of-view, leading to a truncated hand, our S2DHand continues to deliver trustworthy predictions. These results indicate that S2DHand can utilize additional information from dual views to provide significant improvements even under extreme challenging cases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.619, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present a novel single-to-dual-view adaptation framework (S2DHand), designed to adapt a single-view hand pose estimator to dual-view settings. The S2DHand is unsupervised, eliminating the need for multi-view labels. Our method also requires no camera parameters, enabling compatibility with arbitrary dual views. Two stereo constraints are employed as two pseudo-labeling modules in an complementary manner. Our method achieves significant performance gains across all dual-view pairs under both in-dataset and cross-dataset settings." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "684" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Charles George Broyden. The convergence of a class of double-rank minimization algorithms 1. general considerations. IMA Journal of Applied Mathematics, 6(1):76-90, 1970. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.471, + 0.241 + ], + "angle": 0, + "content": "[2] Minjie Cai, Feng Lu, and Yoichi Sato. Generalizing hand segmentation in egocentric videos with uncertainty-guided model adaptation. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 14392-14401, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[3] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, Rohith Mv, Stefan Stojanov, and James M Rehg. Unsupervised 3d pose estimation with geometric self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5714-5724, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[4] Liangjian Chen, Shih-Yao Lin, Yusheng Xie, Yen-Yu Lin, and Xiaohui Xie. Mvhm: A large-scale multi-view hand mesh benchmark for accurate 3d hand pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 836-845, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[5] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.569 + ], + "angle": 0, + "content": "[6] Fabian Dubourvieux, Romaric Audigier, Angelique Loesch, Samia Ainouz, and Stephane Canu. Unsupervised domain adaptation for person re-identification through source-guided pseudo-labeling. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4957–4964. IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.471, + 0.643 + ], + "angle": 0, + "content": "[7] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M. Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23600-23611, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.471, + 0.713 + ], + "angle": 0, + "content": "[8] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9561-9568. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.716, + 0.471, + 0.771 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[11] Zidong Guo, Zejian Yuan, Chong Zhang, Wanchao Chi, Yonggen Ling, and Shenghao Zhang. Domain adaptation gaze estimation by embedding with prediction consistency." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "In Proceedings of the Asian Conference on Computer Vision, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[12] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. Umetrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[13] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9164–9170. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[15] Markus Höll, Markus Oberweger, Clemens Arth, and Vincent Lepetit. Efficient physics-based implementation for realistic hand-object interaction in virtual reality. In 2018 IEEE conference on virtual reality and 3D user interfaces (VR), pages 175-182. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[16] Linzhi Huang, Yulong Li, Hongbo Tian, Yue Yang, Xianggang Li, Weihong Deng, and Jieping Ye. Semi-supervised 2d human pose estimation driven by position inconsistency pseudo label correction module. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 693-703, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[17] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[18] Wolfgang Kabsch. A discussion of the solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 34(5):827-828, 1978. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[19] Leyla Khaleghi, Alireza Sepas-Moghaddam, Joshua Marshall, and Ali Etemad. Multi-view video-based 3d hand pose estimation. IEEE Transactions on Artificial Intelligence, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.678, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[20] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[21] Qiuxia Lin, Linlin Yang, and Angela Yao. Cross-domain 3d hand pose estimation with dual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17184-17193, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[22] Qunming Liu, Xiaodong Li, Xiaofei Zhang, Xiaojun Tan, and Bodong Shi. Multi-view joint learning and bev feature-fusion network for 3d object detection. Applied Sciences, 13 (9):5274, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[23] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. arXiv preprint arXiv:2312.15644, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "685" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[24] Ruicong Liu, Yiwei Bao, Mingjie Xu, Haofei Wang, Yunfei Liu, and Feng Lu. Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[25] Ruicong Liu, Yunfei Liu, Haofei Wang, and Feng Lu. Pnpga+: Plug-and-play domain adaptation for gaze estimation using model variants. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[26] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[27] Yunfei Liu, Ruicong Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with outlier-guided collaborative adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3835-3844, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[28] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Generated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[29] Takehiko Ohkawa, Yu-Jhe Li, Qichen Fu, Ryosuke Furuta, Kris M Kitani, and Yoichi Sato. Domain adaptive hand keypoint and pixel localization in the wild. In European Conference on Computer Vision, pages 68-87. Springer, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[30] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.561, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[31] Vladimir I Pavlovic, Rajeev Sharma, and Thomas S. Huang. Visual interpretation of hand gestures for human-computer interaction: A review. IEEE Transactions on pattern analysis and machine intelligence, 19(7):677-695, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[32] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4826-4836, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.676, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[33] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems, pages 955-960. 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.746, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[34] Siddharth S Rautaray and Anupam Agrawal. Vision based hand gesture recognition for human computer interaction: a survey. Artificial intelligence review, 43:1-54, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[35] Dripta S Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14996-15006, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[36] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[37] Stefan Vogt, Giovanni Buccino, Afra M Wohlschlager, Nicola Canessa, N Jon Shah, Karl Zilles, Simon B Eickhoff, Hans-Joachim Freund, Giacomo Rizzolatti, and Gereon R Fink. Prefrontal involvement in imitation learning of hand actions: effects of practice and expertise. Neuroimage, 37 (4):1371-1383, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[38] Christian Von Hardenberg and François Bérard. Bare-hand human-computer interaction. In Proceedings of the 2001 workshop on Perceptive user interfaces, pages 1-8, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[39] Yilin Wen, Hao Pan, Lei Yang, Jia Pan, Taku Komura, and Wenping Wang. Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21243-21253, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[40] Wenhao Wu, Hau San Wong, and Si Wu. Semi-supervised stereo-based 3d object detection via cross-view consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17471-17481, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[41] Yinsong Xu, Zhuqing Jiang, Aidong Men, Haiying Wang, and Haiyong Luo. Multi-view feature fusion for person re-identification. Knowledge-Based Systems, 229:107344, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[42] Linlin Yang, Shicheng Chen, and Angela Yao. Semi-hand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[43] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[44] Yuxiang Zhang, Zhe Li, Liang An, Mengcheng Li, Tao Yu, and Yebin Liu. Lightweight multi-person total motion capture using sparse multi-view cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5560-5569, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.71 + ], + "angle": 0, + "content": "[45] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.794 + ], + "angle": 0, + "content": "[46] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.798, + 0.892, + 0.851 + ], + "angle": 0, + "content": "[47] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. Technical report, arXiv:1705.01389, 2017. https://arxiv.org/abs/1705.01389. 3,6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "686" + } + ] +] \ No newline at end of file diff --git a/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_origin.pdf b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..386a6ba58d5464fbbaee15956608d07bdcb33239 --- /dev/null +++ b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/899acf81-bf1d-4090-8b2e-d789914e0b58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a09239486b878949768d3a6b1a3553dcac25e2a3d45f61a2bf15868b16537edf +size 2865042 diff --git a/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/full.md b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..098a68ceb128a944080749aba841b7b1cfd2d2cb --- /dev/null +++ b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/full.md @@ -0,0 +1,338 @@ +# Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation + +Ruicong Liu + +Takehiko Ohkawa + +Mingfang Zhang + +Yoichi Sato + +The University of Tokyo, Japan + +{lruicong, ohkawa-t, mfzhang, ysato}@iis.u-tokyo.ac.jp + +# Abstract + +The pursuit of accurate 3D hand pose estimation stands as a keystone for understanding human activity in the realm of egocentric vision. The majority of existing estimation methods still rely on single-view images as input, leading to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, adding another camera to better capture the shape of hands is a practical direction. However, existing multi-view hand pose estimation methods suffer from two main drawbacks: 1) Requiring multi-view annotations for training, which are expensive. 2) During testing, the model becomes inapplicable if camera parameters/layout are not the same as those used in training. In this paper, we propose a novel Single-to-Dual-view adaptation (S2DHand) solution that adapts a pre-trained single-view estimator to dual views. Compared with existing multi-view training methods, 1) our adaptation process is unsupervised, eliminating the need for multi-view annotation. 2) Moreover, our method can handle arbitrary dual-view pairs with unknown camera parameters, making the model applicable to diverse camera settings. Specifically, S2DHand is built on certain stereo constraints, including pair-wise cross-view consensus and invariance of transformation between both views. These two stereo constraints are used in a complementary manner to generate pseudo-labels, allowing reliable adaptation. Evaluation results reveal that S2DHand achieves significant improvements on arbitrary camera pairs under both in-dataset and cross-dataset settings, and outperforms existing adaptation methods with leading performance. Project page: https://github.com/ut-vision/S2DHand. + +# 1. Introduction + +Delving into the realm of egocentric vision (first-person view), the pursuit of refining 3D hand pose estimation stands as a keystone for understanding human activity. This quest not only forges new paths in human-computer interaction [31, 34, 38], but also empowers imitation learning + +![](images/da105d94f95ccd3b3b84db61e86194932ff7d10116dbfbd49601f145da7eac0d.jpg) +Figure 1. From (a) to (b), our single-to-dual-view adaptation method adapts a traditional single-view hand pose estimator to arbitrary dual views and achieves better accuracy. (a) Traditional single-view hand pose estimation. (b) Inference process of the adapted model under a dual-view setting. + +[8, 13, 37]. Moreover, it enhances the immersive experience in augmented/virtual reality (AR/VR) to new heights [15, 33]. Recently, with the advancements of AR/VR headsets, egocentric data has become increasingly prevalent [5, 10], leading to an increasing demand for estimating 3D hand poses from egocentric viewpoints. + +To achieve better 3D hand pose estimation performance, recent years have witnessed many networks with various structures [7, 39, 46]. However, the majority of existing hand pose estimation methods are still under a single-view setting, which is convenient but leads to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, a potential solution is to add another camera to expand the field-of-view and reduce depth ambiguity by capturing the hand shape from an additional view angle. Furthermore, the use of multiple cameras also aligns with industry trends, as demonstrated by the latest AR/VR headsets such as the Apple Vision Pro and Meta Quest, which feature multiple egocentric cameras. Overall, an unavoidable trend towards multi-view settings in hand pose estimation is emerging, driven by its technological advantages and the direction of industrial development. + +Currently, several existing studies [4, 12, 19] have paid attention to hand pose estimation under multi-view settings. + +Table 1. Differences among traditional single-view methods, multi-view training methods, and our proposed single-to-dual-view adaptation method. Green indicates lower training requirements, enhanced testing results, and reduced camera parameter requirements, respectively. + +
MethodsPre-training dataset requiredTestCamera parameters
Traditional single-view methods [7, 39, 46]Single-view (common single-camera setting)Single-viewNot required
Multi-view training methods [4, 12, 19]Dual-view (need multi-camera setting)Dual-view & same camera poses with trainingRequired & same with training
Single-to-dual-view adaptation (Ours)Single-view (common single-camera setting)Dual-view & arbitrary camera posesNot required
+ +These methods typically process input images from multiple views simultaneously, utilizing a feature fusion module to arrive at a final prediction [22, 41]. However, all these methods have two significant drawbacks that limit their applicability. 1) The training, especially for the feature fusion module, necessitates multi-view labels, which are costly to annotate. 2) During testing, the same camera parameters as in training must be used. An estimator trained under a specific multi-camera setup becomes inapplicable if there are any changes to the camera layout or parameters. + +Unlike existing multi-view training methods, we propose a new solution that adapts an estimator from single-view to dual-view without needing multi-view labels or camera parameters. As shown in Fig. 1, given a pre-trained estimator, our method adapts it to an arbitrary dual-view setting (from (a) to (b)), where two cameras are placed in any layout without knowing their parameters. Here, all we need is a pre-trained estimator and a sufficient number of unlabeled dual-view inputs from the two cameras. As compared in Tab. 1 (row 2-3), in contrast to multi-view training, our method only needs common and cheaper single-view data for training. During testing, unlike existing methods, our method is compatible with arbitrary dual-view pairs, making the model applicable to flexible and changeable camera settings. Specifically, when camera settings change, it is easy and swift to repeat our method's adaptation process to re-adapt the pre-trained estimator to work well with new camera parameters. For camera parameters, existing methods not only need them for training, but also require them the same with testing. Conversely, our method is clearly more practical since no camera parameters are required. + +Building on these advancements, we present a novel unsupervised Single-to-Dual-view adaptation framework (S2DHand) for egocentric 3D hand pose estimation. It uses certain stereo constraints for adaptation, including cross-view consensus (pair-wise) and invariance of transformation between both camera coordinate systems (to all input pairs). These two stereo constraints are used in a complementary manner to refine the accuracy of pseudo-labels, allowing the model to better fit to the dual views. Specifically, the cross-view consensus is leveraged through an attention-based merging module, and the invariance of transformation is utilized via a rotation-guided refinement module. + +We evaluate our method by adapting a pre-trained estimator to several dual-camera pairs placed in arbitrary poses [30]. Our evaluation encompasses both in-dataset and cross-dataset scenarios. Experimental results reveal that our + +technique not only realizes notable improvements across all pairs but also surpasses state-of-the-art adaptation methods. The primary contributions of this paper are summarized as: + +- We propose a novel unsupervised single-to-dual-view adaptation (S2DHand) solution for egocentric 3D hand pose estimation. Our method can adapt a traditional single-view estimator for arbitrary dual views without requiring annotations or camera parameters. +- We build a pseudo-label-based strategy for adaptation. It leverages cross-view consensus and invariance of transformation between both camera coordinate systems for reliable pseudo-labeling. This leads to two key modules: attention-based merging and rotation-guided refinement. +- Evaluation results demonstrate the benefits of our approach for arbitrarily placed camera pairs. Our method achieves significant improvements for all pairs both under in-dataset and cross-dataset settings. + +# 2. Related Work + +# 2.1. Multi-view hand pose estimation + +Multi-view hand pose estimation accepts multi-view images as input and outputs a final 3D hand pose, which remains a relatively unexplored research area. Previous works [4, 12, 19] design various network structures to predict 3D hand poses through a multi-view fusion module. Similarly, multiple studies have been done [44, 45] through feature fusion for body pose estimation under multi-view settings. + +All these studies have two limitations: 1) they require costly multi-view images and annotations for training, 2) during testing, camera poses are assumed to be known and covered by training data, thereby limiting their applicability. In contrast, our method eliminates the need for multi-view annotations and is adaptable to arbitrary dual views. + +# 2.2. Adaptation in hand pose estimation + +Adaptation aims at tailoring a model for specific application scenarios [20, 24, 29]. Existing adaptation methods in hand pose estimation mainly focus on adaptation across different domains (datasets), e.g., entropy minimization [17, 32], consistency regularization [3, 21, 29], and pseudo-labeling methods [16, 35, 42]. Prior works only use constraints in single-view settings for adaptation, e.g., bio-mechanical constraint [21, 42], and are thus limited to single-view inference. Unlike these methods, we propose stereo constraints from dual views for adaptation and supports dual-view inference, which extends the application scenarios. + +![](images/ac57a26fcd87a67b09f50e22e0d29616418a35bb90272dd234bb52ef3a7cab2b.jpg) +Figure 2. Problem setting of single-to-dual-view adaptation for hand pose estimation. (a) The input and output of adaptation. (b) The dual-view testing scheme after adaptation. + +# 3. Problem Setting + +Fig. 2 illustrates the task setting of single-to-dual-view adaptation for hand pose estimation. We denote unlabeled dual-view data as $\mathcal{D} = \{\mathbf{x}_i^{v1},\mathbf{x}_i^{v2}|_{i = 1}^N\}$ , where $\mathbf{x}_i^{v1}$ and $\mathbf{x}_i^{v2}$ denote the $i$ -th image from view1 and view2, respectively, $N$ is the number of image pairs. The dual-view data $\mathcal{D}$ contains no ground-truth hand poses or camera parameters. + +As shown in Fig. 2 (a), suppose we have a baseline hand pose estimator $\mathcal{H}(\cdot|\theta)$ with parameters $\theta$ pre-trained from common single-view data. Leveraging $\mathcal{D}$ can enhance its performance, as $\mathcal{D}$ provides additional information from a dual-view setup. Our objective is to adapt this pre-trained estimator, $\mathcal{H}(\cdot|\theta)$ , to an arbitrary yet fixed dual-view setting (with unknown camera poses) without needing groundtruths or camera parameters. By inputting $\mathcal{H}(\cdot|\theta)$ and $\mathcal{D}$ into our method, it outputs an adapted estimator $\mathcal{H}(\cdot|\theta')$ with parameters $\theta'$ tailored for the dual-view scenario. + +Upon adapting the estimator, its inference mechanism is correspondingly tailored for dual-view scenarios (Fig. 1 (b)). During testing, the adapted estimator $\mathcal{H}(\cdot |\theta^{\prime})$ processes a dual-view input pair $(\mathbf{x}^{v1},\mathbf{x}^{v2})$ and produces two predictions $(\mathbf{J}^{v1},\mathbf{J}^{v2})$ , where each $\mathbf{J}^v\in \mathbb{R}^{21\times 3}$ represents the 21 3D joints of the hand. These predictions denote the 3D hand joints for each view and can be combined together to generate a final output, e.g., through a simple average. + +Camera layout for a multi-view headset. Fig. 3 illustrates an example of headset-mounted camera setups for multiview egocentric data capture, with four cameras at each corner for different views. The top-right of Fig. 3 displays images from these cameras. Six distinct dual-view pairs can be created from these four views. As a supplement, the bottom of Fig. 3 shows the synthetic training data, highlighting variations in style and lighting. Such data helps to explore the performance of our method under cross-dataset or simulate-to-real settings. See Sec. 5.1 for dataset details. + +![](images/4d98d7ba4e1360a685602effc4b571cc2a9a2ec620bd9659c736e40017cc7990.jpg) +Figure 3. Top: Headset and its camera layout to collect multi-view data, and samples from the four views. Bottom: Samples of synthetic data. Image samples are from AssemblyHands [30], GANerated Hands [28], and Rendered Handpose [47], respectively. + +# 4. Proposed Method + +We propose a novel unsupervised single-to-dual-view adaptation framework (S2DHand). Before adaptation, an initialization step is performed to initialize the rotation matrix between both views (Sec. 4.1). The rotation matrix is essential to establish the transformation between two camera coordinate systems. The architecture overview, as illustrated in Fig. 4, comprises two branches, an estimator $\mathcal{H}$ and a its momentum version $\overline{\mathcal{H}}$ . The adaptation process is designed from two stereo constraints, pair-wise cross-view consensus and invariant rotation transformation between both camera coordinate systems. This leads to two key pseudo-labeling modules: attention-based merging and rotation-guided refinement (Secs. 4.3 and 4.4). Notably, these two modules function in a complementary manner, depending on the prediction accuracy, ensuring reliable pseudo-labeling. + +# 4.1. Initialization + +The initialization step aims to estimate a relatively accurate rotation matrix $R$ , since $R$ is necessary to link the two camera coordinate systems [23, 43]. It should be noted that translation vector between the two cameras is not necessary, as the predicted hand poses are usually aligned by the wrist during testing [30, 46]. Assuming that the initial pre-trained estimator is sufficiently accurate to generate reasonable predictions, we estimate the $R$ using the predictions of unlabeled dual-view data $\mathcal{D} = \{\mathbf{x}_i^{v1}, \mathbf{x}_i^{v2}|_{i=1}^N\}$ . Given $\mathcal{D}$ , the estimator $\mathcal{H}$ can output $N$ pairs of predictions $\{\mathbf{J}_i^{v1}, \mathbf{J}_i^{v2}|_{i=1}^N\}$ , + +![](images/7e0d20658f0cf2f945f1885462c7ddc2b354f0b5fc21b484ad0b8802e033f319.jpg) +Figure 4. Overview of the proposed S2DHand, image pairs captured from arbitrarily placed dual cameras are input for adaptation. The architecture of S2DHand is illustrated in the dark dashed box, which contains a dynamically updated estimator and a momentum estimator. The momentum estimator's predictions are used to generate pseudo-labels, which are then processed by our pseudo-labeling module (Secs. 4.3 and 4.4). Using the pseudo-labels, a loss function is computed to update the estimator. The rotation matrix $R$ from the initialization step (Sec. 4.1) is required for the pseudo-labeling. + +where $\mathbf{J}_i^v\in \mathbb{R}^{21\times 3}$ (21 is the number of 3D joints). Then, the rotation matrix $R$ is estimated by: + +$$ +R ^ {(0)} = \frac {1}{N} \sum_ {i = 1} ^ {N} r o t \left(\mathbf {J} _ {i} ^ {v 1}, \mathbf {J} _ {i} ^ {v 2}\right), \tag {1} +$$ + +where the superscript of $R$ denotes the iteration number, (0) indicates that it is before the first iteration. The rot function [18] generates a $3 \times 3$ rotation matrix from two $21 \times 3$ joint predictions. Note that the average in Eq. (1) is not elementwise, but an average of rotation matrices. + +# 4.2. Single-to-dual-view adaptation + +With the initialized $R$ , the adaptation process begins. The S2DHand framework comprises two branches, an estimator $\mathcal{H}(\cdot |\theta)$ with dynamically updating parameters $\theta$ , and its momentum version $\overline{\mathcal{H}} (\cdot |\overline{\theta})$ , which updates its parameters $\bar{\theta}$ using temporal moving average. Temporal moving average has been proved by many works [6, 14, 25, 27] that can help to stabilize the training process. The $\bar{\theta}$ is updated as: + +$$ +\bar {\theta} ^ {(T)} = \eta_ {\theta} \bar {\theta} ^ {(T - 1)} + (1 - \eta_ {\theta}) \theta , \tag {2} +$$ + +where $\overline{\theta}^{(T - 1)}$ indicates the temporal averaged parameters in the previous iteration $T - 1$ , and $\eta_{\theta}$ represents the ensembling momentum, which is set as 0.99 [25, 27]. + +As shown in Fig. 4, during the single-to-dual-view adaptation, the role of the momentum model $\overline{\mathcal{H}}$ is to generate pseudo-labels, which are then utilized to supervise the + +model $\mathcal{H}$ . The pseudo-labeling module (Secs. 4.3 and 4.4) outputs pseudo-labels $\hat{\mathbf{y}}^{v1},\hat{\mathbf{y}}^{v2}$ based on the predictions $\overline{\mathbf{J}}^{v1},\overline{\mathbf{J}}^{v2}$ from $\overline{\mathcal{H}}$ . These pseudo-labels are then used to supervise the predictions $\mathbf{J}^{v1},\mathbf{J}^{v2}$ of $\mathcal{H}$ . The loss function is computed as: + +$$ +\mathcal {L} = \left\| \mathbf {J} ^ {v 1} - \hat {\mathbf {y}} ^ {v 1} \right\| _ {2} + \left\| \mathbf {J} ^ {v 2} - \hat {\mathbf {y}} ^ {v 2} \right\| _ {2}. \tag {3} +$$ + +The estimator follows the implementation of DetNet [46], where $\mathcal{H}$ directly outputs heatmaps, and $\mathbf{J}$ is calculated from the heatmaps. Therefore, the loss function is actually computed from corresponding heatmaps, here we write these 3D-joint variables only for better understanding. + +# 4.3. Pseudo-labeling: attention-based merging + +The attention-based merging (ABM) module, which constitutes the first part of pseudo-labeling, is derived from cross-view consensus. Cross-view consensus refers to the concept of achieving agreement or consistency between different views of the same data [40]. Theoretically, when transformed into the same coordinate system, the two predictions $\overline{\mathbf{J}}^{v1}$ and $\overline{\mathbf{J}}^{v2}$ from different views should be identical, i.e., $R\overline{\mathbf{J}}^{v1} = \overline{\mathbf{J}}^{v2}$ , with $\overline{\mathbf{J}}^{v1}\overline{\mathbf{J}}^{v2}$ being aligned with wrist joint. + +This stereo constraint is the foundation for this module to generate accurate pseudo-labels. Prior works utilize a simple average [26, 29] (e.g., $(R\overline{\mathbf{J}}^{v1} + \overline{\mathbf{J}}^{v2}) / 2$ ) or sample-wise confidence [2, 29] to improve the quality of pseudolabels. However, these approaches overlook the varying confidence in joints that is caused by differences in image capture across views. For instance, a joint that is occluded + +![](images/dd6e00e991ca4caf4f4760ba95568fd0f3c456761820a11e9bc45f0e7d102862.jpg) +Figure 5. Top: illustration of the first part of pseudo-labeling: attention-based merging module. The generating process of $\hat{y}_{abm}^{v2}$ in view2 is shown as an example, the process of view1 is the same. Bottom: visualizations of heatmaps with different accuracy. + +in one view but fully visible in another could lead to reliable predictions being hindered by unreliable ones. + +To address this, we propose joint-wise attention $\mathbf{w} \in \mathbb{R}^{21 \times 1}$ to represent each joint's prediction confidence. It is derived from the 2D heatmap $\overline{\mathbf{h}} \in \mathbb{R}^{21 \times 32 \times 32}$ output from $\overline{\mathcal{H}}$ . The $\overline{\mathbf{h}}$ indicates the probability of each joint's presence at every pixel in the 2D image space. This approach is from an observation (bottom of Fig. 5): as the error of prediction increases, the intensity of the heatmap's hotspot decreases, i.e., darker indicates low accuracy. Inspired by this, we propose an attention-generating function: + +$$ +\mathbf {w} _ {j} ^ {v} = \frac {\beta^ {\max \left(\overline {{\mathbf {h}}} _ {j} ^ {v}\right)}}{\sum_ {v \in \{v 1 , v 2 \}} \beta^ {\max \left(\overline {{\mathbf {h}}} _ {j} ^ {v}\right)}}, \tag {4} +$$ + +where the subscript $j$ indicate the index of joint, i.e., $j = 1,2,\ldots ,21$ . We introduce a hyper-parameter $\beta$ here to adjust softness. Please refer to Sec. 5.8 for parameter choosing. + +The workflow of this attention-based merging module is illustrated at the top of Fig. 5. First, we transform both predictions into the same camera coordinate system. Then, a joint-wise multiplication is performed for each of them using the attention $\mathbf{w}^{v1},\mathbf{w}^{v2}$ . Finally, the pseudo-label $\hat{\mathbf{y}}_{abm}$ is calculated through a summation operation. In summary: + +$$ +\hat {\mathbf {y}} _ {a b m} ^ {v 1} = \mathbf {w} ^ {v 1} \overline {{\mathbf {J}}} ^ {v 1} + \mathbf {w} ^ {v 2} R ^ {T} \overline {{\mathbf {J}}} ^ {v 2}, \tag {5} +$$ + +$$ +\hat {\mathbf {y}} _ {a b m} ^ {v 2} = \mathbf {w} ^ {v 1} R \overline {{\mathbf {J}}} ^ {v 1} + \mathbf {w} ^ {v 2} \overline {{\mathbf {J}}} ^ {v 2}. +$$ + +# 4.4. Pseudo-labeling: rotation-guided refinement + +The rotation-guided refinement (RGR) module is based on another stereo constraint: invariance of rotation transformation between both views. This implies that the estimated rotation matrix should remain unchanged since the cameras + +![](images/e99f4a925ed0416018d1aa30067e0f58f892d493ca4801a04510408bb4912ef6.jpg) +Figure 6. Illustration of the second part of pseudo-labeling: rotation-guided refinement module. + +are fixed, i.e., theoretically $rot(\overline{\mathbf{J}}^{v1}, \overline{\mathbf{J}}^{v2}) = C$ . In light of this, this module aims to refine the predictions $\overline{\mathbf{J}}^{v1}, \overline{\mathbf{J}}^{v2}$ such that $rot(\overline{\mathbf{J}}^{v1}, \overline{\mathbf{J}}^{v2})$ becomes invariant across all input data. The refinement result becomes the pseudo-label $\hat{\mathbf{y}}_{rgr}$ . + +The workflow is shown in Fig. 6. Given a pair of predictions, our method estimates a new rotation matrix $R' = \text{rot}(\overline{\mathbf{J}}^{v1}, \overline{\mathbf{J}}^{v2})$ . Subsequently, the more accurately estimated $R$ from the initialization step is set as the target for refinement, aiming to minimize $\| R - R' \|_F$ . The refinement process can be expressed as: + +$$ +\hat {\mathbf {y}} _ {r g r} ^ {v 1}, \hat {\mathbf {y}} _ {r g r} ^ {v 2} = \underset {\overline {{\mathbf {J}}} ^ {v 1}, \overline {{\mathbf {J}}} ^ {v 2}} {\arg \min } \| R - r o t (\overline {{\mathbf {J}}} ^ {v 1}, \overline {{\mathbf {J}}} ^ {v 2}) \| _ {F}. \tag {6} +$$ + +In detail, we employ BFGS [1] algorithm for minimizing. + +Final pseudo-label. Finally, the pseudo-label is calculated from a weighted average of $\hat{\mathbf{y}}_{abm}$ and $\hat{\mathbf{y}}_{rgr}$ : + +$$ +\hat {\mathbf {y}} = \alpha \hat {\mathbf {y}} _ {a b m} + (1 - \alpha) \hat {\mathbf {y}} _ {r g r}. \tag {7} +$$ + +Here, we introduce another pre-fixed hyper-parameter $\alpha$ , which adjusts the weight of the two parts of the pseudolabel. Empirically, we set $\alpha = 0.7$ (see Sec. 5.8). + +Complement between two pseudo-labels. When the predictions $\overline{\mathbf{J}}^{v1}$ and $\overline{\mathbf{J}}^{v2}$ are accurate, $R^{\prime}$ closely approximates $R$ , making $\hat{\mathbf{y}}_{rgr}$ redundant. In such cases, $\hat{\mathbf{y}}_{abm}$ is beneficial as it merges the two accurate predictions. Conversely, if $\overline{\mathbf{J}}^{v1}$ and $\overline{\mathbf{J}}^{v2}$ are unreliable, $\hat{\mathbf{y}}_{abm}$ will consequently also lack accuracy. In such instances, $\hat{\mathbf{y}}_{rgr}$ steps in as a complementary tool for refining pseudo-labels. By minimizing $\| R - R^{\prime}\|_{F}$ , it optimizes the predictions towards an alignment with real-world condition. + +Update rotation matrix R. Clearly, the rotation matrix $R$ from the initialization step plays an important role in bridging the two camera coordinate systems. As can be expected, its accuracy significantly affects the final performance. To enhance its accuracy with each iteration, we also employ a temporal moving average for its updates. Given an input batch containing $B$ image pairs, the $R$ is updated as: + +$$ +R ^ {(T)} = \eta_ {R} R ^ {(T - 1)} + (1 - \eta_ {R}) \cdot \frac {1}{B} \sum_ {i = 1} ^ {B} r o t \left(\overline {{\mathbf {J}}} _ {i} ^ {v 1}, \overline {{\mathbf {J}}} _ {i} ^ {v 2}\right), \tag {8} +$$ + +where the ensembling momentum $\eta_{R}$ is set as 0.999 [14]. This way, the rotation matrix $R$ can be updated slowly and provide a more accurate rotation function with iteration. + +Table 2. Adaptation results for all dual-camera pairs. The camera pairs are from $\mathcal{D}_{ah}$ dataset [30], which is divided into two parts according to the collecting headset. In "In-dataset" and "Cross-dataset" settings, the baseline model is pre-trained on $\mathcal{D}_{ah}$ and $\mathcal{D}_{syn}$ [28, 47], respectively. The adaptation yields a uniquely adapted model for each pair, and "Overall" averages the results across all 6 pairs. + +
Camera pairMethodIn-dataset (Dah → Dah)Cross-dataset (Dsyn → Dah)
Dah - Headset1Dah - Headset2Dah - Headset1Dah - Headset2
Mono-MDual-MMono-MDual-MMono-MDual-MMono-MDual-M
cam 1,2Baseline43.0039.2054.7152.3867.9360.4870.3262.26
S2DHand31.01 ▼ 27.9%31.36 ▼ 20.0%45.52 ▼ 16.8%45.14 ▼ 13.8%63.46 ▼ 6.6%59.32 ▼ 1.9%70.09 ▼ 0.3%60.97 ▼ 2.1%
cam 1,3Baseline25.0023.2922.5921.0857.7951.4264.0060.25
S2DHand19.73 ▼ 21.1%19.92 ▼ 14.5%17.90 ▼ 20.8%17.68 ▼ 16.1%50.84 ▼ 12.0%47.55 ▼ 7.5%61.81 ▼ 3.4%58.34 ▼ 3.2%
cam 1,4Baseline24.9022.7016.7314.9152.7146.5554.3250.57
S2DHand20.88 ▼ 16.1%20.87 ▼ 8.1%14.64 ▼ 12.5%14.29 ▼ 4.2%46.05 ▼ 12.6%42.50 ▼ 8.7%46.59 ▼ 14.2%45.66 ▼ 9.7%
cam 2,3Baseline17.9615.2317.1015.0853.3648.4252.8448.84
S2DHand14.97 ▼ 16.6%14.44 ▼ 5.2%14.42 ▼ 15.7%14.20 ▼ 5.8%40.26 ▼ 24.6%39.32 ▼ 18.8%43.61 ▼ 17.5%42.88 ▼ 12.2%
cam 2,4Baseline22.0919.8423.2420.9659.4454.3261.1357.41
S2DHand17.98 ▼ 18.6%17.75 ▼ 10.5%18.31 ▼ 21.2%18.41 ▼ 12.2%50.59 ▼ 14.9%49.41 ▼ 9.0%52.45 ▼ 14.2%51.48 ▼ 10.3%
cam 3,4Baseline16.8315.7719.9318.0845.8242.3449.8448.99
S2DHand16.36 ▼ 2.8%15.55 ▼ 1.4%19.25 ▼ 3.4%17.80 ▼ 1.5%39.46 ▼ 13.9%37.43 ▼ 11.6%44.04 ▼ 11.6%42.88 ▼ 12.5%
OverallBaseline24.9622.6725.7223.7556.1850.5958.7454.72
S2DHand20.16 ▼ 19.2%19.98 ▼ 11.9%21.67 ▼ 15.7%21.25 ▼ 10.5%48.44 ▼ 13.8%45.92 ▼ 9.2%53.11 ▼ 9.6%50.37 ▼ 7.9%
+ +# 5. Experiment + +# 5.1. Dataset + +We employ AssemblyHands [30] $(\mathcal{D}_{ah})$ as the evaluation set, as it is the newest large-scale benchmark dataset with high-quality multi-view 3D hand pose annotations. As for the training set, we set two adaptation scenarios: 1) in-dataset setting where the training set is drawn from the same dataset $\mathcal{D}_{ah}$ and 2) cross-dataset setting where we use synthetic dataset $(\mathcal{D}_{syn})$ as the training set, consisting of Rendered Handpose [47] and GANerated Hands [28]. The details of the datasets are as below: + +- **AssemblyHands** [30] is a large-scale benchmark dataset featuring accurate 3D hand pose annotations. Collected using two AR headsets, it comprises images captured from four synchronized egocentric cameras. The dataset includes 412K training samples and 62K testing samples. +- GANerated Hands [28] includes over 330K color images of hands. The images are synthetically generated and then fed to a GAN [9] to make the features closer to real hands. +- Rendered Handpose [47] contains about 44K samples. The images are rendered with freely available characters. In detail, AssemblyHands is collected by two VR headsets, as shown in Fig. 3, each headset has four egocentric cameras at four corners. Following the collecting devices, we separate $\mathcal{D}_{ah}$ into two parts, each part is collected using one headset, namely $\mathcal{D}_{ah} - Headset1/2$ . + +# 5.2. Experimental setup + +Evaluation metric. We compare the predictions from our model with the ground-truth labels in root-relative coordinates, and use the common mean per joint position error (MPJPE) in millimeters as the evaluation metric. However, since our focus is on the single-to-dual-view adaptation task, the adapted estimator is expected to perform in dual- + +view settings. This implies that traditional MPJPE computed from single-view (monocular MPJPE, Mono-M) cannot be sufficient. As a result, we propose a new dual-view MPJPE metric Dual-M in addition to monocular MPJPE. The metrics are defined as follows: + +- Mono-M: the traditional monocular MPJPE, which collects all the single-view errors from both views and calculates their average. +- Dual-M: the proposed metric under dual views. To calculate it, first the predictions from both views are averaged using a rotation matrix $R$ . Then, we calculate the MPJPE of the averaged predictions as the Dual-M. Usually, the $R$ is from the initialization step of our method. + +Implementation detail. We employ PyTorch for implementation. All experiments run on a single NVIDIA A100 GPU. DetNet from [46] is adopted as the backbone of our hand pose estimator. Adam optimizer is employed with a learning rate of $1 \times 10^{-3}$ to pre-train the 3D hand pose estimation network. For the single-to-dual-view adaptation, we use the Adam optimizer with a learning rate of $5 \times 10^{-4}$ . + +# 5.3. Adaptation results for all camera pairs + +In this section, we use our method to adapt the same pretrained single-view hand pose estimator to all dual-view pairs from the evaluation set $\mathcal{D}_{ah}$ independently, yielding one adapted model for each pair. Experiments are conducted under both in-dataset and cross-dataset settings. Under the in-dataset setting (Tab. 2, $\mathcal{D}_{ah} \to \mathcal{D}_{ah}$ ), the baseline model is pre-trained on $\mathcal{D}_{ah}$ , while under the cross-dataset setting $(\mathcal{D}_{syn} \to \mathcal{D}_{ah})$ , the baseline model is pre-trained on $\mathcal{D}_{syn}$ before being adapted to the camera pairs from $\mathcal{D}_{ah}$ . + +As shown in Tab. 2, compared with the pre-trained model (Baseline), our S2DHand offers significant accuracy gains under both settings among all camera pairs. This indicates that our method can adapt well to arbitrary dual views re + +gardless of the camera positions or pre-training datasets. + +Quantitative results demonstrate that the S2DHand offers substantial improvements. On average, the improvement in both monocular (Mono-M) and dual-view (Dual-M) metrics exceeds $10\%$ , with the maximum improvement exceeding $20\%$ . Interestingly, we can see that the improvement for cam 1, 2 under the cross-dataset setting is relatively small. This indicates that low initial accuracy limits the performance. + +# 5.4. Comparison under cross-dataset settings + +Our method is compared with state-of-the-art adaptation techniques in cross-dataset settings. Considering the prevalence and significance of cross-dataset scenarios in real-world applications, this experiment evaluates the capability of S2DHand in comparison to leading domain adaptation methods. Specifically, adaptation methods included in the comparison are: SFDAHPE [32], RegDA [17], DAGEN [11], and ADDA [36]. + +For fairness, we do not include existing multi-view methods [4, 12, 19] in this comparison. This is because these methods require 1) multi-view labels and 2) camera parameters, whereas our approach is unsupervised and does not require such parameters. In contrast, all the comparison methods are unsupervised, leading to a fair comparison. + +In detail, SFDAHPE [32], RegDA [17] are developed for pose estimation. DAGEN [11] and ADDA [36] are originally proposed for gaze estimation and classification, respectively. We include these two methods here to show potential of these state-of-the-art methods in enhancing the cross-dataset performance of hand pose estimation. To make a fair comparison, their original networks are replaced with the same DetNet [46] as our baseline. + +Quantitative results of different methods are shown in Tab. 3. Our method not only significantly outperforms the state-of-the-art methods, but also shows an advantage of being source-free. The superior performance verifies the effectiveness of the proposed S2DHand for single-to-dual-view adaptation under cross-dataset settings. For reference, we also provide the result of fine-tuning as the upper bound of this cross-dataset task. + +# 5.5. Ablation study + +We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the $\mathcal{D}_{ah} \to \mathcal{D}_{ah}$ task for clearer observation. The components are shown below: + +- ABM: Attention-based merging module, which generates pseudo-labels based on the cross-view consensus. +- RGR: Rotation-guided refinement module, which generates pseudo-labels based on the invariance of rotation transformation between both camera coordinate systems. + +Table 3. Comparison with state-of-the-art adaptation methods under cross-dataset settings. "SF" indicates if the method is source-free (requiring no data from source dataset, $\mathcal{D}_{syn}$ ). * denotes that labels from the target dataset ( $\mathcal{D}_{ah}$ ) are needed. + +
\(D_{syn}\rightarrow\mathcal{D}_{ah}\)\(\begin{array}{c}\mathsf{SF}\\ \end{array}\)\(\frac{\mathcal{D}_{ah}-\text{Headset1}}{\text{Mono-M}}\)\(\frac{\mathcal{D}_{ah}-\text{Headset2}}{\text{Mono-M}}\)
Dual-MDual-M
Source Only56.1850.5958.74
Fine-tune*45.0338.1147.75
ADDA [36]X56.9048.4857.87
DAGEN [11]X55.3749.7257.62
RegDA [17]X51.4147.8554.75
SFDAHPE [32]✓54.0649.1157.22
S2DHand (Ours)✓48.4445.9253.11
50.37
+ +Table 4. Ablation study of our method on $\mathcal{D}_{ah} \rightarrow \mathcal{D}_{ah}$ task. ABM and RGR stand for the two pseudo-labeling modules, respectively. + +
ABMRGR\( \mathcal{D}_{ah} - Headset1 \)\( \mathcal{D}_{ah} - Headset2 \)
Mono-MDual-MMono-MDual-M
XX24.9622.6725.7223.75
✓X20.8120.5422.2421.71
X✓21.8921.3323.5422.75
✓✓20.1619.9821.6721.25
+ +Tab. 4 shows the hand pose estimation errors under different combinations. We observe that both ABM and RGR can significantly improve the hand pose estimation performance over the pre-trained baseline (first row). Our final version achieves the best results for all metrics, confirming the optimality of our method. + +# 5.6. Number of input image pairs + +To find the optimal number of input image pairs (i.e., $N$ ) for our method, we evaluate the S2DHand's performance under different numbers of input image pairs. Specifically, the experiments are conducted on the cam 2,3 pair in the $\mathcal{D}_{ah} \to \mathcal{D}_{ah}$ task. The results are illustrated in Fig. 7. It indicates that the performance of S2DHand constantly improves as the number of input image pairs increase. Our method's performance converges when $N \geq 1000$ . Consequently, we choose $N = 1000$ for our S2DHand. + +# 5.7. Complement between two pseudo-labels + +Fig. 8 demonstrates the complementary nature, as stated in Sec. 4.4, of the pseudo-labeling. Using camera pair $cam1,2 - Headset1$ under in-dataset setting, we analyze the error of pseudo-labels with and without the refinement term $\hat{\mathbf{y}}_{rgr}$ , in relation to the prediction error of $\overline{\mathbf{J}}$ . Note that $\overline{\mathbf{J}}$ is the prediction used to compute these pseudolabels. For better observation, the predictions are first divided into seven equal intervals according to their errors, [9.4, 34.0), [34.0, 58.6), ..., [157.1, 181.7]. Then, the average pseudo-label error for each interval is calculated. + +In Fig. 8, the bars are placed in the middle of each inter + +![](images/de0074811df19a878036390d2f064e64b1058f0cef65e5aeb73c10debbe6c2ec.jpg) +Figure 7. Evaluation of S2DHand's performance with gradually increasing the number of input image pairs. + +![](images/8d5e1e6eaef800f5f16f7b50d2195b5a21f77babf9cdc010bb201ab04554bc9a.jpg) +Figure 8. The error of pseudo-labels with and without $\hat{\mathbf{y}}_{rgr}$ , in relation to the error of prediction $\overline{\mathbf{J}}$ , where $\overline{\mathbf{J}}$ is what we use to compute the pseudo-labels. MPJPE in millimeter is the metric. + +Table 5. Performance of our method with different hyperparameters $\alpha$ (Eq. (7)) and $\beta$ (Eq. (4)). + +
\( {\mathcal{D}}_{ah} \rightarrow {\mathcal{D}}_{ah} \)\( {\mathcal{D}}_{ah} - Headset1 \)\( {\mathcal{D}}_{ah} - Headset2 \)
Mono-MDual-MMono-MDual-M
\( \alpha = {0.3} \)20.8320.5422.3921.82
\( \alpha = {0.5} \)20.3420.1721.9021.49
\( \alpha = {0.7} \)20.1619.9821.6721.25
\( \alpha = {0.9} \)20.4620.1822.1721.53
\( \beta = 1 \)21.0220.9422.7722.56
\( \beta = e \)20.6620.4722.5022.10
\( \beta = \infty \)20.1619.9821.6721.25
+ +val, with y-axis representing the pseudo-label errors. The result indicates that the $\hat{\mathbf{y}}_{rgr}$ term is redundant for accurate predictions ( $< 60mm$ ) but significantly reduces pseudolabel errors for larger prediction errors ( $\geq 60mm$ ). This finding supports the statement in Sec. 4.4 about the importance of $\hat{\mathbf{y}}_{rgr}$ for complementing pseudo-labels in cases of inaccurate predictions. + +# 5.8. Hyper-parameters + +We evaluate how the S2DHand's performance varies with the change of weight parameter $\alpha$ (Eq. (7)). $\alpha$ controls the weights for averaging the two parts of pseudo-labeling. We test four values of $\alpha$ , 0.3, 0.5, 0.7, and 0.9. The results are shown in the row 1-4 of Tab. 5, where our method achieves the best performance when $\alpha = 0.7$ . + +We also test the performance with varying $\beta$ (Eq. (4)). $\beta$ is the parameter in the attention-generating function, which generates the joint-wise attention in attention-based merging module. In fact, the $\beta$ acts the role of $e$ in the softmax function. When $\beta = 1, e, \infty$ , the merging becomes a simple average (where attention becomes invalidated), soft- + +![](images/30eab19f249ea218839214664e88d7a687d80e8efed123cd6ee6435b0e4b1727.jpg) +Figure 9. Visual examples of estimated 3D hand poses under both views. The joints are portrayed by projecting the final 3D predictions to the image plane. + +max function, and maximum function, respectively. We can see that the S2DHand achieves the best performance when $\beta = \infty$ . This suggests that selecting the prediction with higher confidence as a pseudo-label in a joint-wise manner is the most effective strategy. Consequently, we set $\alpha = 0.7$ and $\beta = \infty$ for all the experiments. + +# 5.9. Qualitative result + +To understand how our method improves the performance of hand pose estimation under dual-view settings, we visually present typical cases by portraying 3D hand joints onto the input image pairs. In detail, the 3D joints are projected to the image plane, with the visualization depicted in Fig. 9. + +Notably, when confronted with extreme view angles (see the left pair), the predictions of baseline model tend to be unreliable. Conversely, our technique gives a prediction much closer to the actual hand shape after adaptation. In the 3rd column, even when the hand is partially out of field-of-view, leading to a truncated hand, our S2DHand continues to deliver trustworthy predictions. These results indicate that S2DHand can utilize additional information from dual views to provide significant improvements even under extreme challenging cases. + +# 6. Conclusion + +In this paper, we present a novel single-to-dual-view adaptation framework (S2DHand), designed to adapt a single-view hand pose estimator to dual-view settings. The S2DHand is unsupervised, eliminating the need for multi-view labels. Our method also requires no camera parameters, enabling compatibility with arbitrary dual views. Two stereo constraints are employed as two pseudo-labeling modules in an complementary manner. Our method achieves significant performance gains across all dual-view pairs under both in-dataset and cross-dataset settings. + +# References + +[1] Charles George Broyden. The convergence of a class of double-rank minimization algorithms 1. general considerations. IMA Journal of Applied Mathematics, 6(1):76-90, 1970. 5 +[2] Minjie Cai, Feng Lu, and Yoichi Sato. Generalizing hand segmentation in egocentric videos with uncertainty-guided model adaptation. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 14392-14401, 2020. 4 +[3] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, Rohith Mv, Stefan Stojanov, and James M Rehg. Unsupervised 3d pose estimation with geometric self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5714-5724, 2019. 2 +[4] Liangjian Chen, Shih-Yao Lin, Yusheng Xie, Yen-Yu Lin, and Xiaohui Xie. Mvhm: A large-scale multi-view hand mesh benchmark for accurate 3d hand pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 836-845, 2021. 1, 2, 7 +[5] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 1 +[6] Fabian Dubourvieux, Romaric Audigier, Angelique Loesch, Samia Ainouz, and Stephane Canu. Unsupervised domain adaptation for person re-identification through source-guided pseudo-labeling. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4957–4964. IEEE, 2021. 4 +[7] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M. Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23600-23611, 2023. 1, 2 +[8] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9561-9568. IEEE, 2020. 1 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 6 +[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 1 +[11] Zidong Guo, Zejian Yuan, Chong Zhang, Wanchao Chi, Yonggen Ling, and Shenghao Zhang. Domain adaptation gaze estimation by embedding with prediction consistency. + +In Proceedings of the Asian Conference on Computer Vision, 2020. 7 +[12] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. Umetrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 7 +[13] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9164–9170. IEEE, 2020. 1 +[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5 +[15] Markus Höll, Markus Oberweger, Clemens Arth, and Vincent Lepetit. Efficient physics-based implementation for realistic hand-object interaction in virtual reality. In 2018 IEEE conference on virtual reality and 3D user interfaces (VR), pages 175-182. IEEE, 2018. 1 +[16] Linzhi Huang, Yulong Li, Hongbo Tian, Yue Yang, Xianggang Li, Weihong Deng, and Jieping Ye. Semi-supervised 2d human pose estimation driven by position inconsistency pseudo label correction module. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 693-703, 2023. 2 +[17] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021. 2, 7 +[18] Wolfgang Kabsch. A discussion of the solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 34(5):827-828, 1978. 4 +[19] Leyla Khaleghi, Alireza Sepas-Moghaddam, Joshua Marshall, and Ali Etemad. Multi-view video-based 3d hand pose estimation. IEEE Transactions on Artificial Intelligence, 2022. 1, 2, 7 +[20] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 2 +[21] Qiuxia Lin, Linlin Yang, and Angela Yao. Cross-domain 3d hand pose estimation with dual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17184-17193, 2023. 2 +[22] Qunming Liu, Xiaodong Li, Xiaofei Zhang, Xiaojun Tan, and Bodong Shi. Multi-view joint learning and bev feature-fusion network for 3d object detection. Applied Sciences, 13 (9):5274, 2023. 2 +[23] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. arXiv preprint arXiv:2312.15644, 2023. 3 + +[24] Ruicong Liu, Yiwei Bao, Mingjie Xu, Haofei Wang, Yunfei Liu, and Feng Lu. Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082, 2022. 2 +[25] Ruicong Liu, Yunfei Liu, Haofei Wang, and Feng Lu. Pnpga+: Plug-and-play domain adaptation for gaze estimation using model variants. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4 +[26] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 4 +[27] Yunfei Liu, Ruicong Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with outlier-guided collaborative adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3835-3844, 2021. 4 +[28] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Generated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6 +[29] Takehiko Ohkawa, Yu-Jhe Li, Qichen Fu, Ryosuke Furuta, Kris M Kitani, and Yoichi Sato. Domain adaptive hand keypoint and pixel localization in the wild. In European Conference on Computer Vision, pages 68-87. Springer, 2022. 2, 4 +[30] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 2, 3, 6 +[31] Vladimir I Pavlovic, Rajeev Sharma, and Thomas S. Huang. Visual interpretation of hand gestures for human-computer interaction: A review. IEEE Transactions on pattern analysis and machine intelligence, 19(7):677-695, 1997. 1 +[32] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4826-4836, 2023. 2, 7 +[33] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems, pages 955-960. 2013. 1 +[34] Siddharth S Rautaray and Anupam Agrawal. Vision based hand gesture recognition for human computer interaction: a survey. Artificial intelligence review, 43:1-54, 2015. 1 +[35] Dripta S Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14996-15006, 2023. 2 +[36] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 7 + +[37] Stefan Vogt, Giovanni Buccino, Afra M Wohlschlager, Nicola Canessa, N Jon Shah, Karl Zilles, Simon B Eickhoff, Hans-Joachim Freund, Giacomo Rizzolatti, and Gereon R Fink. Prefrontal involvement in imitation learning of hand actions: effects of practice and expertise. Neuroimage, 37 (4):1371-1383, 2007. 1 +[38] Christian Von Hardenberg and François Bérard. Bare-hand human-computer interaction. In Proceedings of the 2001 workshop on Perceptive user interfaces, pages 1-8, 2001. 1 +[39] Yilin Wen, Hao Pan, Lei Yang, Jia Pan, Taku Komura, and Wenping Wang. Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21243-21253, 2023. 1, 2 +[40] Wenhao Wu, Hau San Wong, and Si Wu. Semi-supervised stereo-based 3d object detection via cross-view consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17471-17481, 2023. 4 +[41] Yinsong Xu, Zhuqing Jiang, Aidong Men, Haiying Wang, and Haiyong Luo. Multi-view feature fusion for person re-identification. Knowledge-Based Systems, 229:107344, 2021. 2 +[42] Linlin Yang, Shicheng Chen, and Angela Yao. Semi-hand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2 +[43] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 3 +[44] Yuxiang Zhang, Zhe Li, Liang An, Mengcheng Li, Tao Yu, and Yebin Liu. Lightweight multi-person total motion capture using sparse multi-view cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5560-5569, 2021. 2 +[45] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2 +[46] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 3, 4, 6, 7 +[47] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. Technical report, arXiv:1705.01389, 2017. https://arxiv.org/abs/1705.01389. 3,6 \ No newline at end of file diff --git a/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/images.zip b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..35562779847f33b790d3d19c6d7b9b1e8302cc14 --- /dev/null +++ b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83bf1921d8f13d6ddf641d4e0c453c296344a5449654b95896921c17966f908d +size 677454 diff --git a/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/layout.json b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a8f4ccd2177cdb864593e9ec1cbca6cb5d132fd2 --- /dev/null +++ b/2024/Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation/layout.json @@ -0,0 +1,10069 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "text", + "content": "Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 143, + 177, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 143, + 177, + 157 + ], + "spans": [ + { + "bbox": [ + 113, + 143, + 177, + 157 + ], + "type": "text", + "content": "Ruicong Liu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 144, + 290, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 144, + 290, + 156 + ], + "spans": [ + { + "bbox": [ + 201, + 144, + 290, + 156 + ], + "type": "text", + "content": "Takehiko Ohkawa" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 144, + 397, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 144, + 397, + 157 + ], + "spans": [ + { + "bbox": [ + 315, + 144, + 397, + 157 + ], + "type": "text", + "content": "Mingfang Zhang" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 422, + 144, + 480, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 144, + 480, + 156 + ], + "spans": [ + { + "bbox": [ + 422, + 144, + 480, + 156 + ], + "type": "text", + "content": "Yoichi Sato" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 220, + 158, + 373, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 158, + 373, + 172 + ], + "spans": [ + { + "bbox": [ + 220, + 158, + 373, + 172 + ], + "type": "text", + "content": "The University of Tokyo, Japan" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 149, + 174, + 441, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 174, + 441, + 186 + ], + "spans": [ + { + "bbox": [ + 149, + 174, + 441, + 186 + ], + "type": "text", + "content": "{lruicong, ohkawa-t, mfzhang, ysato}@iis.u-tokyo.ac.jp" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 238, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 597 + ], + "type": "text", + "content": "The pursuit of accurate 3D hand pose estimation stands as a keystone for understanding human activity in the realm of egocentric vision. The majority of existing estimation methods still rely on single-view images as input, leading to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, adding another camera to better capture the shape of hands is a practical direction. However, existing multi-view hand pose estimation methods suffer from two main drawbacks: 1) Requiring multi-view annotations for training, which are expensive. 2) During testing, the model becomes inapplicable if camera parameters/layout are not the same as those used in training. In this paper, we propose a novel Single-to-Dual-view adaptation (S2DHand) solution that adapts a pre-trained single-view estimator to dual views. Compared with existing multi-view training methods, 1) our adaptation process is unsupervised, eliminating the need for multi-view annotation. 2) Moreover, our method can handle arbitrary dual-view pairs with unknown camera parameters, making the model applicable to diverse camera settings. Specifically, S2DHand is built on certain stereo constraints, including pair-wise cross-view consensus and invariance of transformation between both views. These two stereo constraints are used in a complementary manner to generate pseudo-labels, allowing reliable adaptation. Evaluation results reveal that S2DHand achieves significant improvements on arbitrary camera pairs under both in-dataset and cross-dataset settings, and outperforms existing adaptation methods with leading performance. Project page: https://github.com/ut-vision/S2DHand." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 608, + 128, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 128, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 128, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "type": "text", + "content": "Delving into the realm of egocentric vision (first-person view), the pursuit of refining 3D hand pose estimation stands as a keystone for understanding human activity. This quest not only forges new paths in human-computer interaction [31, 34, 38], but also empowers imitation learning" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 216, + 533, + 353 + ], + "blocks": [ + { + "bbox": [ + 320, + 216, + 533, + 353 + ], + "lines": [ + { + "bbox": [ + 320, + 216, + 533, + 353 + ], + "spans": [ + { + "bbox": [ + 320, + 216, + 533, + 353 + ], + "type": "image", + "image_path": "da105d94f95ccd3b3b84db61e86194932ff7d10116dbfbd49601f145da7eac0d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 366, + 547, + 422 + ], + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 422 + ], + "type": "text", + "content": "Figure 1. From (a) to (b), our single-to-dual-view adaptation method adapts a traditional single-view hand pose estimator to arbitrary dual views and achieves better accuracy. (a) Traditional single-view hand pose estimation. (b) Inference process of the adapted model under a dual-view setting." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 498 + ], + "type": "text", + "content": "[8, 13, 37]. Moreover, it enhances the immersive experience in augmented/virtual reality (AR/VR) to new heights [15, 33]. Recently, with the advancements of AR/VR headsets, egocentric data has become increasingly prevalent [5, 10], leading to an increasing demand for estimating 3D hand poses from egocentric viewpoints." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "content": "To achieve better 3D hand pose estimation performance, recent years have witnessed many networks with various structures [7, 39, 46]. However, the majority of existing hand pose estimation methods are still under a single-view setting, which is convenient but leads to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, a potential solution is to add another camera to expand the field-of-view and reduce depth ambiguity by capturing the hand shape from an additional view angle. Furthermore, the use of multiple cameras also aligns with industry trends, as demonstrated by the latest AR/VR headsets such as the Apple Vision Pro and Meta Quest, which feature multiple egocentric cameras. Overall, an unavoidable trend towards multi-view settings in hand pose estimation is emerging, driven by its technological advantages and the direction of industrial development." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Currently, several existing studies [4, 12, 19] have paid attention to hand pose estimation under multi-view settings." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 712 + ], + "type": "text", + "content": "This work was supported by JSPS KAKENHI Grant Number JP20H04205 and JST ACT-X Grant Number JPMJAX2007." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "677" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 95, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Differences among traditional single-view methods, multi-view training methods, and our proposed single-to-dual-view adaptation method. Green indicates lower training requirements, enhanced testing results, and reduced camera parameter requirements, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 95, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 48, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 545, + 140 + ], + "type": "table", + "html": "
MethodsPre-training dataset requiredTestCamera parameters
Traditional single-view methods [7, 39, 46]Single-view (common single-camera setting)Single-viewNot required
Multi-view training methods [4, 12, 19]Dual-view (need multi-camera setting)Dual-view & same camera poses with trainingRequired & same with training
Single-to-dual-view adaptation (Ours)Single-view (common single-camera setting)Dual-view & arbitrary camera posesNot required
", + "image_path": "13778c7f7c503fa63a11134c453c5599a58bb3baa399a0e69cf39bfe9193e023.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 263 + ], + "type": "text", + "content": "These methods typically process input images from multiple views simultaneously, utilizing a feature fusion module to arrive at a final prediction [22, 41]. However, all these methods have two significant drawbacks that limit their applicability. 1) The training, especially for the feature fusion module, necessitates multi-view labels, which are costly to annotate. 2) During testing, the same camera parameters as in training must be used. An estimator trained under a specific multi-camera setup becomes inapplicable if there are any changes to the camera layout or parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 517 + ], + "type": "text", + "content": "Unlike existing multi-view training methods, we propose a new solution that adapts an estimator from single-view to dual-view without needing multi-view labels or camera parameters. As shown in Fig. 1, given a pre-trained estimator, our method adapts it to an arbitrary dual-view setting (from (a) to (b)), where two cameras are placed in any layout without knowing their parameters. Here, all we need is a pre-trained estimator and a sufficient number of unlabeled dual-view inputs from the two cameras. As compared in Tab. 1 (row 2-3), in contrast to multi-view training, our method only needs common and cheaper single-view data for training. During testing, unlike existing methods, our method is compatible with arbitrary dual-view pairs, making the model applicable to flexible and changeable camera settings. Specifically, when camera settings change, it is easy and swift to repeat our method's adaptation process to re-adapt the pre-trained estimator to work well with new camera parameters. For camera parameters, existing methods not only need them for training, but also require them the same with testing. Conversely, our method is clearly more practical since no camera parameters are required." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 520, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 663 + ], + "type": "text", + "content": "Building on these advancements, we present a novel unsupervised Single-to-Dual-view adaptation framework (S2DHand) for egocentric 3D hand pose estimation. It uses certain stereo constraints for adaptation, including cross-view consensus (pair-wise) and invariance of transformation between both camera coordinate systems (to all input pairs). These two stereo constraints are used in a complementary manner to refine the accuracy of pseudo-labels, allowing the model to better fit to the dual views. Specifically, the cross-view consensus is leveraged through an attention-based merging module, and the invariance of transformation is utilized via a rotation-guided refinement module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "We evaluate our method by adapting a pre-trained estimator to several dual-camera pairs placed in arbitrary poses [30]. Our evaluation encompasses both in-dataset and cross-dataset scenarios. Experimental results reveal that our" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 144, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 180 + ], + "type": "text", + "content": "technique not only realizes notable improvements across all pairs but also surpasses state-of-the-art adaptation methods. The primary contributions of this paper are summarized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 181, + 545, + 348 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 181, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 181, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 181, + 545, + 240 + ], + "type": "text", + "content": "- We propose a novel unsupervised single-to-dual-view adaptation (S2DHand) solution for egocentric 3D hand pose estimation. Our method can adapt a traditional single-view estimator for arbitrary dual views without requiring annotations or camera parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 240, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 300 + ], + "type": "text", + "content": "- We build a pseudo-label-based strategy for adaptation. It leverages cross-view consensus and invariance of transformation between both camera coordinate systems for reliable pseudo-labeling. This leads to two key modules: attention-based merging and rotation-guided refinement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 300, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 348 + ], + "type": "text", + "content": "- Evaluation results demonstrate the benefits of our approach for arbitrarily placed camera pairs. Our method achieves significant improvements for all pairs both under in-dataset and cross-dataset settings." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 360, + 392, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 392, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 392, + 372 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 380, + 481, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 481, + 392 + ], + "type": "text", + "content": "2.1. Multi-view hand pose estimation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 399, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 482 + ], + "type": "text", + "content": "Multi-view hand pose estimation accepts multi-view images as input and outputs a final 3D hand pose, which remains a relatively unexplored research area. Previous works [4, 12, 19] design various network structures to predict 3D hand poses through a multi-view fusion module. Similarly, multiple studies have been done [44, 45] through feature fusion for body pose estimation under multi-view settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": "All these studies have two limitations: 1) they require costly multi-view images and annotations for training, 2) during testing, camera poses are assumed to be known and covered by training data, thereby limiting their applicability. In contrast, our method eliminates the need for multi-view annotations and is adaptable to arbitrary dual views." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 563, + 495, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 495, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 495, + 575 + ], + "type": "text", + "content": "2.2. Adaptation in hand pose estimation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Adaptation aims at tailoring a model for specific application scenarios [20, 24, 29]. Existing adaptation methods in hand pose estimation mainly focus on adaptation across different domains (datasets), e.g., entropy minimization [17, 32], consistency regularization [3, 21, 29], and pseudo-labeling methods [16, 35, 42]. Prior works only use constraints in single-view settings for adaptation, e.g., bio-mechanical constraint [21, 42], and are thus limited to single-view inference. Unlike these methods, we propose stereo constraints from dual views for adaptation and supports dual-view inference, which extends the application scenarios." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "678" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 286, + 243 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 286, + 243 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 286, + 243 + ], + "type": "image", + "image_path": "ac57a26fcd87a67b09f50e22e0d29616418a35bb90272dd234bb52ef3a7cab2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 288, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 281 + ], + "type": "text", + "content": "Figure 2. Problem setting of single-to-dual-view adaptation for hand pose estimation. (a) The input and output of adaptation. (b) The dual-view testing scheme after adaptation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 284, + 146, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 146, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 146, + 298 + ], + "type": "text", + "content": "3. Problem Setting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": "Fig. 2 illustrates the task setting of single-to-dual-view adaptation for hand pose estimation. We denote unlabeled dual-view data as " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{x}_i^{v1},\\mathbf{x}_i^{v2}|_{i = 1}^N\\}" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{v1}" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{v2}" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": "-th image from view1 and view2, respectively, " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " is the number of image pairs. The dual-view data " + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": " contains no ground-truth hand poses or camera parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": "As shown in Fig. 2 (a), suppose we have a baseline hand pose estimator " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot|\\theta)" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " pre-trained from common single-view data. Leveraging " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " can enhance its performance, as " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " provides additional information from a dual-view setup. Our objective is to adapt this pre-trained estimator, " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot|\\theta)" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": ", to an arbitrary yet fixed dual-view setting (with unknown camera poses) without needing groundtruths or camera parameters. By inputting " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot|\\theta)" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " into our method, it outputs an adapted estimator " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot|\\theta')" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " tailored for the dual-view scenario." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "content": "Upon adapting the estimator, its inference mechanism is correspondingly tailored for dual-view scenarios (Fig. 1 (b)). During testing, the adapted estimator " + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot |\\theta^{\\prime})" + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "content": " processes a dual-view input pair " + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}^{v1},\\mathbf{x}^{v2})" + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "content": " and produces two predictions " + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "inline_equation", + "content": "(\\mathbf{J}^{v1},\\mathbf{J}^{v2})" + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{J}^v\\in \\mathbb{R}^{21\\times 3}" + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "content": " represents the 21 3D joints of the hand. These predictions denote the 3D hand joints for each view and can be combined together to generate a final output, e.g., through a simple average." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Camera layout for a multi-view headset. Fig. 3 illustrates an example of headset-mounted camera setups for multiview egocentric data capture, with four cameras at each corner for different views. The top-right of Fig. 3 displays images from these cameras. Six distinct dual-view pairs can be created from these four views. As a supplement, the bottom of Fig. 3 shows the synthetic training data, highlighting variations in style and lighting. Such data helps to explore the performance of our method under cross-dataset or simulate-to-real settings. See Sec. 5.1 for dataset details." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 73, + 526, + 310 + ], + "blocks": [ + { + "bbox": [ + 326, + 73, + 526, + 310 + ], + "lines": [ + { + "bbox": [ + 326, + 73, + 526, + 310 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 526, + 310 + ], + "type": "image", + "image_path": "4d98d7ba4e1360a685602effc4b571cc2a9a2ec620bd9659c736e40017cc7990.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 315, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 304, + 315, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 546, + 361 + ], + "type": "text", + "content": "Figure 3. Top: Headset and its camera layout to collect multi-view data, and samples from the four views. Bottom: Samples of synthetic data. Image samples are from AssemblyHands [30], GANerated Hands [28], and Rendered Handpose [47], respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 364, + 411, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 411, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 411, + 378 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 384, + 546, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 566 + ], + "type": "text", + "content": "We propose a novel unsupervised single-to-dual-view adaptation framework (S2DHand). Before adaptation, an initialization step is performed to initialize the rotation matrix between both views (Sec. 4.1). The rotation matrix is essential to establish the transformation between two camera coordinate systems. The architecture overview, as illustrated in Fig. 4, comprises two branches, an estimator " + }, + { + "bbox": [ + 304, + 384, + 546, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 384, + 546, + 566 + ], + "type": "text", + "content": " and a its momentum version " + }, + { + "bbox": [ + 304, + 384, + 546, + 566 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{H}}" + }, + { + "bbox": [ + 304, + 384, + 546, + 566 + ], + "type": "text", + "content": ". The adaptation process is designed from two stereo constraints, pair-wise cross-view consensus and invariant rotation transformation between both camera coordinate systems. This leads to two key pseudo-labeling modules: attention-based merging and rotation-guided refinement (Secs. 4.3 and 4.4). Notably, these two modules function in a complementary manner, depending on the prediction accuracy, ensuring reliable pseudo-labeling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 574, + 389, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 389, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 389, + 586 + ], + "type": "text", + "content": "4.1. Initialization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "The initialization step aims to estimate a relatively accurate rotation matrix " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", since " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is necessary to link the two camera coordinate systems [23, 43]. It should be noted that translation vector between the two cameras is not necessary, as the predicted hand poses are usually aligned by the wrist during testing [30, 46]. Assuming that the initial pre-trained estimator is sufficiently accurate to generate reasonable predictions, we estimate the " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " using the predictions of unlabeled dual-view data " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{x}_i^{v1}, \\mathbf{x}_i^{v2}|_{i=1}^N\\}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", the estimator " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " can output " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " pairs of predictions " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{J}_i^{v1}, \\mathbf{J}_i^{v2}|_{i=1}^N\\}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "text", + "content": "679" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 70, + 504, + 289 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 504, + 289 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 504, + 289 + ], + "type": "image", + "image_path": "7e0d20658f0cf2f945f1885462c7ddc2b354f0b5fc21b484ad0b8802e033f319.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 545, + 350 + ], + "type": "text", + "content": "Figure 4. Overview of the proposed S2DHand, image pairs captured from arbitrarily placed dual cameras are input for adaptation. The architecture of S2DHand is illustrated in the dark dashed box, which contains a dynamically updated estimator and a momentum estimator. The momentum estimator's predictions are used to generate pseudo-labels, which are then processed by our pseudo-labeling module (Secs. 4.3 and 4.4). Using the pseudo-labels, a loss function is computed to update the estimator. The rotation matrix " + }, + { + "bbox": [ + 46, + 293, + 545, + 350 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 293, + 545, + 350 + ], + "type": "text", + "content": " from the initialization step (Sec. 4.1) is required for the pseudo-labeling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 354, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_i^v\\in \\mathbb{R}^{21\\times 3}" + }, + { + "bbox": [ + 46, + 354, + 287, + 380 + ], + "type": "text", + "content": " (21 is the number of 3D joints). Then, the rotation matrix " + }, + { + "bbox": [ + 46, + 354, + 287, + 380 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 354, + 287, + 380 + ], + "type": "text", + "content": " is estimated by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 389, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 389, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 287, + 422 + ], + "type": "interline_equation", + "content": "R ^ {(0)} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} r o t \\left(\\mathbf {J} _ {i} ^ {v 1}, \\mathbf {J} _ {i} ^ {v 2}\\right), \\tag {1}", + "image_path": "cbe0f5c7f7c063f8c85740aff8cc95d6ebedf8954195c2831e53e882799b349b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "text", + "content": "where the superscript of " + }, + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "text", + "content": " denotes the iteration number, (0) indicates that it is before the first iteration. The rot function [18] generates a " + }, + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "text", + "content": " rotation matrix from two " + }, + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "inline_equation", + "content": "21 \\times 3" + }, + { + "bbox": [ + 47, + 431, + 287, + 491 + ], + "type": "text", + "content": " joint predictions. Note that the average in Eq. (1) is not elementwise, but an average of rotation matrices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 499, + 214, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 214, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 214, + 513 + ], + "type": "text", + "content": "4.2. Single-to-dual-view adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "content": "With the initialized " + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "content": ", the adaptation process begins. The S2DHand framework comprises two branches, an estimator " + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot |\\theta)" + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "content": " with dynamically updating parameters " + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "content": ", and its momentum version " + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{H}} (\\cdot |\\overline{\\theta})" + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "content": ", which updates its parameters " + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}" + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "content": " using temporal moving average. Temporal moving average has been proved by many works [6, 14, 25, 27] that can help to stabilize the training process. The " + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}" + }, + { + "bbox": [ + 46, + 517, + 287, + 602 + ], + "type": "text", + "content": " is updated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 610, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 610, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\bar {\\theta} ^ {(T)} = \\eta_ {\\theta} \\bar {\\theta} ^ {(T - 1)} + (1 - \\eta_ {\\theta}) \\theta , \\tag {2}", + "image_path": "3c4c7bbe655e52f6f212b90b829faa528207b61ac4b85673893c7aa2cdfed1ef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\overline{\\theta}^{(T - 1)}" + }, + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "text", + "content": " indicates the temporal averaged parameters in the previous iteration " + }, + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\eta_{\\theta}" + }, + { + "bbox": [ + 46, + 637, + 287, + 677 + ], + "type": "text", + "content": " represents the ensembling momentum, which is set as 0.99 [25, 27]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "As shown in Fig. 4, during the single-to-dual-view adaptation, the role of the momentum model " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{H}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " is to generate pseudo-labels, which are then utilized to supervise the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "content": "model " + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "content": ". The pseudo-labeling module (Secs. 4.3 and 4.4) outputs pseudo-labels " + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{v1},\\hat{\\mathbf{y}}^{v2}" + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "content": " based on the predictions " + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v1},\\overline{\\mathbf{J}}^{v2}" + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{H}}" + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "content": ". These pseudo-labels are then used to supervise the predictions " + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{J}^{v1},\\mathbf{J}^{v2}" + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 355, + 545, + 418 + ], + "type": "text", + "content": ". The loss function is computed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 423, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 423, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 350, + 423, + 545, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| \\mathbf {J} ^ {v 1} - \\hat {\\mathbf {y}} ^ {v 1} \\right\\| _ {2} + \\left\\| \\mathbf {J} ^ {v 2} - \\hat {\\mathbf {y}} ^ {v 2} \\right\\| _ {2}. \\tag {3}", + "image_path": "80197bac8e627487e3b7ceaec420748b12485797073cf82c1f8e936c45e4f6c2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 444, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 505 + ], + "type": "text", + "content": "The estimator follows the implementation of DetNet [46], where " + }, + { + "bbox": [ + 304, + 444, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 444, + 545, + 505 + ], + "type": "text", + "content": " directly outputs heatmaps, and " + }, + { + "bbox": [ + 304, + 444, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 304, + 444, + 545, + 505 + ], + "type": "text", + "content": " is calculated from the heatmaps. Therefore, the loss function is actually computed from corresponding heatmaps, here we write these 3D-joint variables only for better understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 510, + 525, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 525, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 525, + 523 + ], + "type": "text", + "content": "4.3. Pseudo-labeling: attention-based merging" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": "The attention-based merging (ABM) module, which constitutes the first part of pseudo-labeling, is derived from cross-view consensus. Cross-view consensus refers to the concept of achieving agreement or consistency between different views of the same data [40]. Theoretically, when transformed into the same coordinate system, the two predictions " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v1}" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v2}" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": " from different views should be identical, i.e., " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "R\\overline{\\mathbf{J}}^{v1} = \\overline{\\mathbf{J}}^{v2}" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v1}\\overline{\\mathbf{J}}^{v2}" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": " being aligned with wrist joint." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": "This stereo constraint is the foundation for this module to generate accurate pseudo-labels. Prior works utilize a simple average [26, 29] (e.g., " + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "inline_equation", + "content": "(R\\overline{\\mathbf{J}}^{v1} + \\overline{\\mathbf{J}}^{v2}) / 2" + }, + { + "bbox": [ + 304, + 628, + 545, + 713 + ], + "type": "text", + "content": ") or sample-wise confidence [2, 29] to improve the quality of pseudolabels. However, these approaches overlook the varying confidence in joints that is caused by differences in image capture across views. For instance, a joint that is occluded" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "680" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 272, + 258 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 272, + 258 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 272, + 258 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 272, + 258 + ], + "type": "image", + "image_path": "dd6e00e991ca4caf4f4760ba95568fd0f3c456761820a11e9bc45f0e7d102862.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 307 + ], + "type": "text", + "content": "Figure 5. Top: illustration of the first part of pseudo-labeling: attention-based merging module. The generating process of " + }, + { + "bbox": [ + 46, + 262, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\hat{y}_{abm}^{v2}" + }, + { + "bbox": [ + 46, + 262, + 287, + 307 + ], + "type": "text", + "content": " in view2 is shown as an example, the process of view1 is the same. Bottom: visualizations of heatmaps with different accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 310, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 334 + ], + "type": "text", + "content": "in one view but fully visible in another could lead to reliable predictions being hindered by unreliable ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "content": "To address this, we propose joint-wise attention " + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^{21 \\times 1}" + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "content": " to represent each joint's prediction confidence. It is derived from the 2D heatmap " + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{h}} \\in \\mathbb{R}^{21 \\times 32 \\times 32}" + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "content": " output from " + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{H}}" + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{h}}" + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "content": " indicates the probability of each joint's presence at every pixel in the 2D image space. This approach is from an observation (bottom of Fig. 5): as the error of prediction increases, the intensity of the heatmap's hotspot decreases, i.e., darker indicates low accuracy. Inspired by this, we propose an attention-generating function:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 448, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 448, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {j} ^ {v} = \\frac {\\beta^ {\\max \\left(\\overline {{\\mathbf {h}}} _ {j} ^ {v}\\right)}}{\\sum_ {v \\in \\{v 1 , v 2 \\}} \\beta^ {\\max \\left(\\overline {{\\mathbf {h}}} _ {j} ^ {v}\\right)}}, \\tag {4}", + "image_path": "f33359fe5a319ebdac4c2ef59ebfe13afc72e5fe442c947919f171b0dd6d640b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "text", + "content": "where the subscript " + }, + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "text", + "content": " indicate the index of joint, i.e., " + }, + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "inline_equation", + "content": "j = 1,2,\\ldots ,21" + }, + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "text", + "content": ". We introduce a hyper-parameter " + }, + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 490, + 287, + 525 + ], + "type": "text", + "content": " here to adjust softness. Please refer to Sec. 5.8 for parameter choosing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": "The workflow of this attention-based merging module is illustrated at the top of Fig. 5. First, we transform both predictions into the same camera coordinate system. Then, a joint-wise multiplication is performed for each of them using the attention " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{v1},\\mathbf{w}^{v2}" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": ". Finally, the pseudo-label " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{abm}" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " is calculated through a summation operation. In summary:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 605, + 287, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 605, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 100, + 605, + 287, + 627 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {a b m} ^ {v 1} = \\mathbf {w} ^ {v 1} \\overline {{\\mathbf {J}}} ^ {v 1} + \\mathbf {w} ^ {v 2} R ^ {T} \\overline {{\\mathbf {J}}} ^ {v 2}, \\tag {5}", + "image_path": "a9c2827ae3ee9705b355b912d422c65c06f6dbbc8e6719905662c7d9df0b1748.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 624, + 225, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 624, + 225, + 639 + ], + "spans": [ + { + "bbox": [ + 102, + 624, + 225, + 639 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {a b m} ^ {v 2} = \\mathbf {w} ^ {v 1} R \\overline {{\\mathbf {J}}} ^ {v 1} + \\mathbf {w} ^ {v 2} \\overline {{\\mathbf {J}}} ^ {v 2}.", + "image_path": "0ac2ee2ba535b91c3a62053b9e72900655c9c062617c74b2ad11e02da0158fbd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 647, + 277, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 277, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 277, + 659 + ], + "type": "text", + "content": "4.4. Pseudo-labeling: rotation-guided refinement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "The rotation-guided refinement (RGR) module is based on another stereo constraint: invariance of rotation transformation between both views. This implies that the estimated rotation matrix should remain unchanged since the cameras" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 538, + 151 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 538, + 151 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 538, + 151 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 538, + 151 + ], + "type": "image", + "image_path": "e99f4a925ed0416018d1aa30067e0f58f892d493ca4801a04510408bb4912ef6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 156, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 177 + ], + "type": "text", + "content": "Figure 6. Illustration of the second part of pseudo-labeling: rotation-guided refinement module." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "text", + "content": "are fixed, i.e., theoretically " + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "inline_equation", + "content": "rot(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2}) = C" + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "text", + "content": ". In light of this, this module aims to refine the predictions " + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2}" + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "inline_equation", + "content": "rot(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2})" + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "text", + "content": " becomes invariant across all input data. The refinement result becomes the pseudo-label " + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 304, + 178, + 545, + 233 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "text", + "content": "The workflow is shown in Fig. 6. Given a pair of predictions, our method estimates a new rotation matrix " + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "inline_equation", + "content": "R' = \\text{rot}(\\overline{\\mathbf{J}}^{v1}, \\overline{\\mathbf{J}}^{v2})" + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "text", + "content": ". Subsequently, the more accurately estimated " + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "text", + "content": " from the initialization step is set as the target for refinement, aiming to minimize " + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\| R - R' \\|_F" + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "text", + "content": ". The refinement process can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 311, + 545, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 311, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 332, + 311, + 545, + 336 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {r g r} ^ {v 1}, \\hat {\\mathbf {y}} _ {r g r} ^ {v 2} = \\underset {\\overline {{\\mathbf {J}}} ^ {v 1}, \\overline {{\\mathbf {J}}} ^ {v 2}} {\\arg \\min } \\| R - r o t (\\overline {{\\mathbf {J}}} ^ {v 1}, \\overline {{\\mathbf {J}}} ^ {v 2}) \\| _ {F}. \\tag {6}", + "image_path": "72a6cb0f20ecbefd9ab6a0010893af95a889245dd28b5d4e8990f4eb0d6bbfc7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 342, + 538, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 538, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 538, + 354 + ], + "type": "text", + "content": "In detail, we employ BFGS [1] algorithm for minimizing." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 354, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 378 + ], + "type": "text", + "content": "Final pseudo-label. Finally, the pseudo-label is calculated from a weighted average of " + }, + { + "bbox": [ + 304, + 354, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{abm}" + }, + { + "bbox": [ + 304, + 354, + 545, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 304, + 354, + 545, + 378 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 367, + 384, + 545, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 384, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 367, + 384, + 545, + 398 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} = \\alpha \\hat {\\mathbf {y}} _ {a b m} + (1 - \\alpha) \\hat {\\mathbf {y}} _ {r g r}. \\tag {7}", + "image_path": "9863e99b5d7c9071021aa7152b7d08f172c3480b6f87f250d15dc7a7b1b4403c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "content": "Here, we introduce another pre-fixed hyper-parameter " + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "content": ", which adjusts the weight of the two parts of the pseudolabel. Empirically, we set " + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\alpha = 0.7" + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "content": " (see Sec. 5.8)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": "Complement between two pseudo-labels. When the predictions " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v1}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v2}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " are accurate, " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "R^{\\prime}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " closely approximates " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": ", making " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " redundant. In such cases, " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{abm}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " is beneficial as it merges the two accurate predictions. Conversely, if " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v1}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}^{v2}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " are unreliable, " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{abm}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " will consequently also lack accuracy. In such instances, " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": " steps in as a complementary tool for refining pseudo-labels. By minimizing " + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\| R - R^{\\prime}\\|_{F}" + }, + { + "bbox": [ + 304, + 439, + 545, + 549 + ], + "type": "text", + "content": ", it optimizes the predictions towards an alignment with real-world condition." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "text", + "content": "Update rotation matrix R. Clearly, the rotation matrix " + }, + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "text", + "content": " from the initialization step plays an important role in bridging the two camera coordinate systems. As can be expected, its accuracy significantly affects the final performance. To enhance its accuracy with each iteration, we also employ a temporal moving average for its updates. Given an input batch containing " + }, + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "text", + "content": " image pairs, the " + }, + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 551, + 545, + 634 + ], + "type": "text", + "content": " is updated as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 640, + 545, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 640, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 311, + 640, + 545, + 672 + ], + "type": "interline_equation", + "content": "R ^ {(T)} = \\eta_ {R} R ^ {(T - 1)} + (1 - \\eta_ {R}) \\cdot \\frac {1}{B} \\sum_ {i = 1} ^ {B} r o t \\left(\\overline {{\\mathbf {J}}} _ {i} ^ {v 1}, \\overline {{\\mathbf {J}}} _ {i} ^ {v 2}\\right), \\tag {8}", + "image_path": "f4e97e85382fd8ed6a6028be1d110c181b4715d1968a08dd700e0f4532a82bec.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where the ensembling momentum " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\eta_{R}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is set as 0.999 [14]. This way, the rotation matrix " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " can be updated slowly and provide a more accurate rotation function with iteration." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "681" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 106, + 542, + 285 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Adaptation results for all dual-camera pairs. The camera pairs are from " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " dataset [30], which is divided into two parts according to the collecting headset. In \"In-dataset\" and \"Cross-dataset\" settings, the baseline model is pre-trained on " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{syn}" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " [28, 47], respectively. The adaptation yields a uniquely adapted model for each pair, and \"Overall\" averages the results across all 6 pairs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 106, + 542, + 285 + ], + "lines": [ + { + "bbox": [ + 53, + 106, + 542, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 542, + 285 + ], + "type": "table", + "html": "
Camera pairMethodIn-dataset (Dah → Dah)Cross-dataset (Dsyn → Dah)
Dah - Headset1Dah - Headset2Dah - Headset1Dah - Headset2
Mono-MDual-MMono-MDual-MMono-MDual-MMono-MDual-M
cam 1,2Baseline43.0039.2054.7152.3867.9360.4870.3262.26
S2DHand31.01 ▼ 27.9%31.36 ▼ 20.0%45.52 ▼ 16.8%45.14 ▼ 13.8%63.46 ▼ 6.6%59.32 ▼ 1.9%70.09 ▼ 0.3%60.97 ▼ 2.1%
cam 1,3Baseline25.0023.2922.5921.0857.7951.4264.0060.25
S2DHand19.73 ▼ 21.1%19.92 ▼ 14.5%17.90 ▼ 20.8%17.68 ▼ 16.1%50.84 ▼ 12.0%47.55 ▼ 7.5%61.81 ▼ 3.4%58.34 ▼ 3.2%
cam 1,4Baseline24.9022.7016.7314.9152.7146.5554.3250.57
S2DHand20.88 ▼ 16.1%20.87 ▼ 8.1%14.64 ▼ 12.5%14.29 ▼ 4.2%46.05 ▼ 12.6%42.50 ▼ 8.7%46.59 ▼ 14.2%45.66 ▼ 9.7%
cam 2,3Baseline17.9615.2317.1015.0853.3648.4252.8448.84
S2DHand14.97 ▼ 16.6%14.44 ▼ 5.2%14.42 ▼ 15.7%14.20 ▼ 5.8%40.26 ▼ 24.6%39.32 ▼ 18.8%43.61 ▼ 17.5%42.88 ▼ 12.2%
cam 2,4Baseline22.0919.8423.2420.9659.4454.3261.1357.41
S2DHand17.98 ▼ 18.6%17.75 ▼ 10.5%18.31 ▼ 21.2%18.41 ▼ 12.2%50.59 ▼ 14.9%49.41 ▼ 9.0%52.45 ▼ 14.2%51.48 ▼ 10.3%
cam 3,4Baseline16.8315.7719.9318.0845.8242.3449.8448.99
S2DHand16.36 ▼ 2.8%15.55 ▼ 1.4%19.25 ▼ 3.4%17.80 ▼ 1.5%39.46 ▼ 13.9%37.43 ▼ 11.6%44.04 ▼ 11.6%42.88 ▼ 12.5%
OverallBaseline24.9622.6725.7223.7556.1850.5958.7454.72
S2DHand20.16 ▼ 19.2%19.98 ▼ 11.9%21.67 ▼ 15.7%21.25 ▼ 10.5%48.44 ▼ 13.8%45.92 ▼ 9.2%53.11 ▼ 9.6%50.37 ▼ 7.9%
", + "image_path": "5e548ac51d891c9d5c91b67a51928e54281acea8507ce3b0f207431bdee2a0b6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 293, + 124, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 124, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 124, + 306 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 107, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 107, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 107, + 324 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "text", + "content": "We employ AssemblyHands [30] " + }, + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{ah})" + }, + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "text", + "content": " as the evaluation set, as it is the newest large-scale benchmark dataset with high-quality multi-view 3D hand pose annotations. As for the training set, we set two adaptation scenarios: 1) in-dataset setting where the training set is drawn from the same dataset " + }, + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "text", + "content": " and 2) cross-dataset setting where we use synthetic dataset " + }, + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{syn})" + }, + { + "bbox": [ + 46, + 330, + 288, + 437 + ], + "type": "text", + "content": " as the training set, consisting of Rendered Handpose [47] and GANerated Hands [28]. The details of the datasets are as below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 438, + 288, + 619 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 438, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 288, + 498 + ], + "type": "text", + "content": "- **AssemblyHands** [30] is a large-scale benchmark dataset featuring accurate 3D hand pose annotations. Collected using two AR headsets, it comprises images captured from four synchronized egocentric cameras. The dataset includes 412K training samples and 62K testing samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 498, + 288, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 288, + 534 + ], + "type": "text", + "content": "- GANerated Hands [28] includes over 330K color images of hands. The images are synthetically generated and then fed to a GAN [9] to make the features closer to real hands." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 534, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 288, + 619 + ], + "type": "text", + "content": "- Rendered Handpose [47] contains about 44K samples. The images are rendered with freely available characters. In detail, AssemblyHands is collected by two VR headsets, as shown in Fig. 3, each headset has four egocentric cameras at four corners. Following the collecting devices, we separate " + }, + { + "bbox": [ + 47, + 534, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 47, + 534, + 288, + 619 + ], + "type": "text", + "content": " into two parts, each part is collected using one headset, namely " + }, + { + "bbox": [ + 47, + 534, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah} - Headset1/2" + }, + { + "bbox": [ + 47, + 534, + 288, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 623, + 162, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 162, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 162, + 636 + ], + "type": "text", + "content": "5.2. Experimental setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Evaluation metric. We compare the predictions from our model with the ground-truth labels in root-relative coordinates, and use the common mean per joint position error (MPJPE) in millimeters as the evaluation metric. However, since our focus is on the single-to-dual-view adaptation task, the adapted estimator is expected to perform in dual-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 293, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 353 + ], + "type": "text", + "content": "view settings. This implies that traditional MPJPE computed from single-view (monocular MPJPE, Mono-M) cannot be sufficient. As a result, we propose a new dual-view MPJPE metric Dual-M in addition to monocular MPJPE. The metrics are defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 354, + 547, + 449 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 306, + 354, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 545, + 389 + ], + "type": "text", + "content": "- Mono-M: the traditional monocular MPJPE, which collects all the single-view errors from both views and calculates their average." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 389, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 547, + 449 + ], + "type": "text", + "content": "- Dual-M: the proposed metric under dual views. To calculate it, first the predictions from both views are averaged using a rotation matrix " + }, + { + "bbox": [ + 306, + 389, + 547, + 449 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 389, + 547, + 449 + ], + "type": "text", + "content": ". Then, we calculate the MPJPE of the averaged predictions as the Dual-M. Usually, the " + }, + { + "bbox": [ + 306, + 389, + 547, + 449 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 389, + 547, + 449 + ], + "type": "text", + "content": " is from the initialization step of our method." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 449, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 534 + ], + "type": "text", + "content": "Implementation detail. We employ PyTorch for implementation. All experiments run on a single NVIDIA A100 GPU. DetNet from [46] is adopted as the backbone of our hand pose estimator. Adam optimizer is employed with a learning rate of " + }, + { + "bbox": [ + 304, + 449, + 545, + 534 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 449, + 545, + 534 + ], + "type": "text", + "content": " to pre-train the 3D hand pose estimation network. For the single-to-dual-view adaptation, we use the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 449, + 545, + 534 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 449, + 545, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 540, + 511, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 511, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 511, + 553 + ], + "type": "text", + "content": "5.3. Adaptation results for all camera pairs" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": "In this section, we use our method to adapt the same pretrained single-view hand pose estimator to all dual-view pairs from the evaluation set " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": " independently, yielding one adapted model for each pair. Experiments are conducted under both in-dataset and cross-dataset settings. Under the in-dataset setting (Tab. 2, " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah} \\to \\mathcal{D}_{ah}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": "), the baseline model is pre-trained on " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": ", while under the cross-dataset setting " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_{syn} \\to \\mathcal{D}_{ah})" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": ", the baseline model is pre-trained on " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{syn}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": " before being adapted to the camera pairs from " + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 304, + 558, + 545, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "As shown in Tab. 2, compared with the pre-trained model (Baseline), our S2DHand offers significant accuracy gains under both settings among all camera pairs. This indicates that our method can adapt well to arbitrary dual views re" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "682" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 274, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 274, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 274, + 84 + ], + "type": "text", + "content": "gardless of the camera positions or pre-training datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 180 + ], + "type": "text", + "content": "Quantitative results demonstrate that the S2DHand offers substantial improvements. On average, the improvement in both monocular (Mono-M) and dual-view (Dual-M) metrics exceeds " + }, + { + "bbox": [ + 46, + 85, + 286, + 180 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 85, + 286, + 180 + ], + "type": "text", + "content": ", with the maximum improvement exceeding " + }, + { + "bbox": [ + 46, + 85, + 286, + 180 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 85, + 286, + 180 + ], + "type": "text", + "content": ". Interestingly, we can see that the improvement for cam 1, 2 under the cross-dataset setting is relatively small. This indicates that low initial accuracy limits the performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 192, + 260, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 260, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 260, + 205 + ], + "type": "text", + "content": "5.4. Comparison under cross-dataset settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 212, + 286, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 286, + 307 + ], + "type": "text", + "content": "Our method is compared with state-of-the-art adaptation techniques in cross-dataset settings. Considering the prevalence and significance of cross-dataset scenarios in real-world applications, this experiment evaluates the capability of S2DHand in comparison to leading domain adaptation methods. Specifically, adaptation methods included in the comparison are: SFDAHPE [32], RegDA [17], DAGEN [11], and ADDA [36]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 286, + 380 + ], + "type": "text", + "content": "For fairness, we do not include existing multi-view methods [4, 12, 19] in this comparison. This is because these methods require 1) multi-view labels and 2) camera parameters, whereas our approach is unsupervised and does not require such parameters. In contrast, all the comparison methods are unsupervised, leading to a fair comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 286, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 286, + 476 + ], + "type": "text", + "content": "In detail, SFDAHPE [32], RegDA [17] are developed for pose estimation. DAGEN [11] and ADDA [36] are originally proposed for gaze estimation and classification, respectively. We include these two methods here to show potential of these state-of-the-art methods in enhancing the cross-dataset performance of hand pose estimation. To make a fair comparison, their original networks are replaced with the same DetNet [46] as our baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 478, + 286, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 286, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 286, + 573 + ], + "type": "text", + "content": "Quantitative results of different methods are shown in Tab. 3. Our method not only significantly outperforms the state-of-the-art methods, but also shows an advantage of being source-free. The superior performance verifies the effectiveness of the proposed S2DHand for single-to-dual-view adaptation under cross-dataset settings. For reference, we also provide the result of fine-tuning as the upper bound of this cross-dataset task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 585, + 138, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 138, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 138, + 597 + ], + "type": "text", + "content": "5.5. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 604, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 286, + 652 + ], + "type": "text", + "content": "We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the " + }, + { + "bbox": [ + 46, + 604, + 286, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah} \\to \\mathcal{D}_{ah}" + }, + { + "bbox": [ + 46, + 604, + 286, + 652 + ], + "type": "text", + "content": " task for clearer observation. The components are shown below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 654, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 47, + 654, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 676 + ], + "type": "text", + "content": "- ABM: Attention-based merging module, which generates pseudo-labels based on the cross-view consensus." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "- RGR: Rotation-guided refinement module, which generates pseudo-labels based on the invariance of rotation transformation between both camera coordinate systems." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 309, + 118, + 544, + 229 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art adaptation methods under cross-dataset settings. \"SF\" indicates if the method is source-free (requiring no data from source dataset, " + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{syn}" + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": "). * denotes that labels from the target dataset (" + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah}" + }, + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": ") are needed." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 118, + 544, + 229 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 544, + 229 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 544, + 229 + ], + "type": "table", + "html": "
\\(D_{syn}\\rightarrow\\mathcal{D}_{ah}\\)\\(\\begin{array}{c}\\mathsf{SF}\\\\ \\end{array}\\)\\(\\frac{\\mathcal{D}_{ah}-\\text{Headset1}}{\\text{Mono-M}}\\)\\(\\frac{\\mathcal{D}_{ah}-\\text{Headset2}}{\\text{Mono-M}}\\)
Dual-MDual-M
Source Only56.1850.5958.74
Fine-tune*45.0338.1147.75
ADDA [36]X56.9048.4857.87
DAGEN [11]X55.3749.7257.62
RegDA [17]X51.4147.8554.75
SFDAHPE [32]✓54.0649.1157.22
S2DHand (Ours)✓48.4445.9253.11
50.37
", + "image_path": "b78d688e896a52c3b26bcc377abcb9c28f047435337c2b6d438eb23378c689de.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 261, + 537, + 337 + ], + "blocks": [ + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "type": "text", + "content": "Table 4. Ablation study of our method on " + }, + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah} \\rightarrow \\mathcal{D}_{ah}" + }, + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "type": "text", + "content": " task. ABM and RGR stand for the two pseudo-labeling modules, respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 261, + 537, + 337 + ], + "lines": [ + { + "bbox": [ + 315, + 261, + 537, + 337 + ], + "spans": [ + { + "bbox": [ + 315, + 261, + 537, + 337 + ], + "type": "table", + "html": "
ABMRGR\\( \\mathcal{D}_{ah} - Headset1 \\)\\( \\mathcal{D}_{ah} - Headset2 \\)
Mono-MDual-MMono-MDual-M
XX24.9622.6725.7223.75
✓X20.8120.5422.2421.71
X✓21.8921.3323.5422.75
✓✓20.1619.9821.6721.25
", + "image_path": "4bf5b0cb1ab244b35d04aee82f0d9fe6950cc24d2e99f126a1874a1e5a7635b4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 347, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 545, + 419 + ], + "type": "text", + "content": "Tab. 4 shows the hand pose estimation errors under different combinations. We observe that both ABM and RGR can significantly improve the hand pose estimation performance over the pre-trained baseline (first row). Our final version achieves the best results for all metrics, confirming the optimality of our method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 428, + 463, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 463, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 463, + 441 + ], + "type": "text", + "content": "5.6. Number of input image pairs" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "text", + "content": "To find the optimal number of input image pairs (i.e., " + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "text", + "content": ") for our method, we evaluate the S2DHand's performance under different numbers of input image pairs. Specifically, the experiments are conducted on the cam 2,3 pair in the " + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ah} \\to \\mathcal{D}_{ah}" + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "text", + "content": " task. The results are illustrated in Fig. 7. It indicates that the performance of S2DHand constantly improves as the number of input image pairs increase. Our method's performance converges when " + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "inline_equation", + "content": "N \\geq 1000" + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "text", + "content": ". Consequently, we choose " + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "inline_equation", + "content": "N = 1000" + }, + { + "bbox": [ + 305, + 447, + 545, + 555 + ], + "type": "text", + "content": " for our S2DHand." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 562, + 515, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 515, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 515, + 575 + ], + "type": "text", + "content": "5.7. Complement between two pseudo-labels" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "text", + "content": "Fig. 8 demonstrates the complementary nature, as stated in Sec. 4.4, of the pseudo-labeling. Using camera pair " + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "cam1,2 - Headset1" + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "text", + "content": " under in-dataset setting, we analyze the error of pseudo-labels with and without the refinement term " + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "text", + "content": ", in relation to the prediction error of " + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}" + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}" + }, + { + "bbox": [ + 305, + 582, + 545, + 700 + ], + "type": "text", + "content": " is the prediction used to compute these pseudolabels. For better observation, the predictions are first divided into seven equal intervals according to their errors, [9.4, 34.0), [34.0, 58.6), ..., [157.1, 181.7]. Then, the average pseudo-label error for each interval is calculated." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 701, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 544, + 713 + ], + "type": "text", + "content": "In Fig. 8, the bars are placed in the middle of each inter" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "683" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 286, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 163 + ], + "type": "image", + "image_path": "de0074811df19a878036390d2f064e64b1058f0cef65e5aeb73c10debbe6c2ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 165, + 287, + 189 + ], + "lines": [ + { + "bbox": [ + 47, + 165, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 287, + 189 + ], + "type": "text", + "content": "Figure 7. Evaluation of S2DHand's performance with gradually increasing the number of input image pairs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 194, + 278, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 194, + 278, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 278, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 278, + 281 + ], + "type": "image", + "image_path": "8d5e1e6eaef800f5f16f7b50d2195b5a21f77babf9cdc010bb201ab04554bc9a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "type": "text", + "content": "Figure 8. The error of pseudo-labels with and without " + }, + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "type": "text", + "content": ", in relation to the error of prediction " + }, + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}" + }, + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{J}}" + }, + { + "bbox": [ + 46, + 282, + 287, + 316 + ], + "type": "text", + "content": " is what we use to compute the pseudo-labels. MPJPE in millimeter is the metric." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 348, + 280, + 454 + ], + "blocks": [ + { + "bbox": [ + 47, + 321, + 287, + 344 + ], + "lines": [ + { + "bbox": [ + 47, + 321, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 287, + 344 + ], + "type": "text", + "content": "Table 5. Performance of our method with different hyperparameters " + }, + { + "bbox": [ + 47, + 321, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 321, + 287, + 344 + ], + "type": "text", + "content": " (Eq. (7)) and " + }, + { + "bbox": [ + 47, + 321, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 321, + 287, + 344 + ], + "type": "text", + "content": " (Eq. (4))." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 348, + 280, + 454 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 280, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 280, + 454 + ], + "type": "table", + "html": "
\\( {\\mathcal{D}}_{ah} \\rightarrow {\\mathcal{D}}_{ah} \\)\\( {\\mathcal{D}}_{ah} - Headset1 \\)\\( {\\mathcal{D}}_{ah} - Headset2 \\)
Mono-MDual-MMono-MDual-M
\\( \\alpha = {0.3} \\)20.8320.5422.3921.82
\\( \\alpha = {0.5} \\)20.3420.1721.9021.49
\\( \\alpha = {0.7} \\)20.1619.9821.6721.25
\\( \\alpha = {0.9} \\)20.4620.1822.1721.53
\\( \\beta = 1 \\)21.0220.9422.7722.56
\\( \\beta = e \\)20.6620.4722.5022.10
\\( \\beta = \\infty \\)20.1619.9821.6721.25
", + "image_path": "f8c398a591518bc145c4cbba015b8830104007820e8a3697afd2b211a1bc54ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": "val, with y-axis representing the pseudo-label errors. The result indicates that the " + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": " term is redundant for accurate predictions (" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "inline_equation", + "content": "< 60mm" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": ") but significantly reduces pseudolabel errors for larger prediction errors (" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\geq 60mm" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": "). This finding supports the statement in Sec. 4.4 about the importance of " + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{rgr}" + }, + { + "bbox": [ + 46, + 460, + 287, + 544 + ], + "type": "text", + "content": " for complementing pseudo-labels in cases of inaccurate predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 552, + 157, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 157, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 157, + 565 + ], + "type": "text", + "content": "5.8. Hyper-parameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "content": "We evaluate how the S2DHand's performance varies with the change of weight parameter " + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "content": " (Eq. (7)). " + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "content": " controls the weights for averaging the two parts of pseudo-labeling. We test four values of " + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "content": ", 0.3, 0.5, 0.7, and 0.9. The results are shown in the row 1-4 of Tab. 5, where our method achieves the best performance when " + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\alpha = 0.7" + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "We also test the performance with varying " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " (Eq. (4)). " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " is the parameter in the attention-generating function, which generates the joint-wise attention in attention-based merging module. In fact, the " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " acts the role of " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " in the softmax function. When " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\beta = 1, e, \\infty" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": ", the merging becomes a simple average (where attention becomes invalidated), soft-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 535, + 243 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 535, + 243 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 535, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 535, + 243 + ], + "type": "image", + "image_path": "30eab19f249ea218839214664e88d7a687d80e8efed123cd6ee6435b0e4b1727.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 244, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 278 + ], + "type": "text", + "content": "Figure 9. Visual examples of estimated 3D hand poses under both views. The joints are portrayed by projecting the final 3D predictions to the image plane." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "text", + "content": "max function, and maximum function, respectively. We can see that the S2DHand achieves the best performance when " + }, + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "inline_equation", + "content": "\\beta = \\infty" + }, + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "text", + "content": ". This suggests that selecting the prediction with higher confidence as a pseudo-label in a joint-wise manner is the most effective strategy. Consequently, we set " + }, + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "inline_equation", + "content": "\\alpha = 0.7" + }, + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "inline_equation", + "content": "\\beta = \\infty" + }, + { + "bbox": [ + 305, + 282, + 547, + 355 + ], + "type": "text", + "content": " for all the experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 363, + 410, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 363, + 410, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 363, + 410, + 376 + ], + "type": "text", + "content": "5.9. Qualitative result" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "content": "To understand how our method improves the performance of hand pose estimation under dual-view settings, we visually present typical cases by portraying 3D hand joints onto the input image pairs. In detail, the 3D joints are projected to the image plane, with the visualization depicted in Fig. 9." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 442, + 546, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 546, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 546, + 562 + ], + "type": "text", + "content": "Notably, when confronted with extreme view angles (see the left pair), the predictions of baseline model tend to be unreliable. Conversely, our technique gives a prediction much closer to the actual hand shape after adaptation. In the 3rd column, even when the hand is partially out of field-of-view, leading to a truncated hand, our S2DHand continues to deliver trustworthy predictions. These results indicate that S2DHand can utilize additional information from dual views to provide significant improvements even under extreme challenging cases." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 573, + 378, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 378, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 378, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "In this paper, we present a novel single-to-dual-view adaptation framework (S2DHand), designed to adapt a single-view hand pose estimator to dual-view settings. The S2DHand is unsupervised, eliminating the need for multi-view labels. Our method also requires no camera parameters, enabling compatibility with arbitrary dual views. Two stereo constraints are employed as two pseudo-labeling modules in an complementary manner. Our method achieves significant performance gains across all dual-view pairs under both in-dataset and cross-dataset settings." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "684" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Charles George Broyden. The convergence of a class of double-rank minimization algorithms 1. general considerations. IMA Journal of Applied Mathematics, 6(1):76-90, 1970. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "text", + "content": "[2] Minjie Cai, Feng Lu, and Yoichi Sato. Generalizing hand segmentation in egocentric videos with uncertainty-guided model adaptation. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 14392-14401, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "text", + "content": "[3] Ching-Hang Chen, Ambrish Tyagi, Amit Agrawal, Dylan Drover, Rohith Mv, Stefan Stojanov, and James M Rehg. Unsupervised 3d pose estimation with geometric self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5714-5724, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 316 + ], + "type": "text", + "content": "[4] Liangjian Chen, Shih-Yao Lin, Yusheng Xie, Yen-Yu Lin, and Xiaohui Xie. Mvhm: A large-scale multi-view hand mesh benchmark for accurate 3d hand pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 836-845, 2021. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 384 + ], + "type": "text", + "content": "[5] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 386, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 450 + ], + "type": "text", + "content": "[6] Fabian Dubourvieux, Romaric Audigier, Angelique Loesch, Samia Ainouz, and Stephane Canu. Unsupervised domain adaptation for person re-identification through source-guided pseudo-labeling. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4957–4964. IEEE, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 288, + 509 + ], + "type": "text", + "content": "[7] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M. Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23600-23611, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "type": "text", + "content": "[8] Guillermo Garcia-Hernando, Edward Johns, and Tae-Kyun Kim. Physics-based dexterous manipulations with estimated hand poses and residual reinforcement learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9561-9568. IEEE, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 288, + 610 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[11] Zidong Guo, Zejian Yuan, Chong Zhang, Wanchao Chi, Yonggen Ling, and Shenghao Zhang. Domain adaptation gaze estimation by embedding with prediction consistency." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "In Proceedings of the Asian Conference on Computer Vision, 2020. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[12] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. Umetrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 152, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 216 + ], + "type": "text", + "content": "[13] Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9164–9170. IEEE, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "text", + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "text", + "content": "[15] Markus Höll, Markus Oberweger, Clemens Arth, and Vincent Lepetit. Efficient physics-based implementation for realistic hand-object interaction in virtual reality. In 2018 IEEE conference on virtual reality and 3D user interfaces (VR), pages 175-182. IEEE, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "type": "text", + "content": "[16] Linzhi Huang, Yulong Li, Hongbo Tian, Yue Yang, Xianggang Li, Weihong Deng, and Jieping Ye. Semi-supervised 2d human pose estimation driven by position inconsistency pseudo label correction module. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 693-703, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "text", + "content": "[17] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[18] Wolfgang Kabsch. A discussion of the solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 34(5):827-828, 1978. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "text", + "content": "[19] Leyla Khaleghi, Alireza Sepas-Moghaddam, Joshua Marshall, and Ali Etemad. Multi-view video-based 3d hand pose estimation. IEEE Transactions on Artificial Intelligence, 2022. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 536, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 536, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 536, + 545, + 591 + ], + "type": "text", + "content": "[20] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[21] Qiuxia Lin, Linlin Yang, and Angela Yao. Cross-domain 3d hand pose estimation with dual modalities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17184-17193, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[22] Qunming Liu, Xiaodong Li, Xiaofei Zhang, Xiaojun Tan, and Bodong Shi. Multi-view joint learning and bev feature-fusion network for 3d object detection. Applied Sciences, 13 (9):5274, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[23] Ruicong Liu and Feng Lu. Uvagaze: Unsupervised 1-to-2 views adaptation for gaze estimation. arXiv preprint arXiv:2312.15644, 2023. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "685" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[24] Ruicong Liu, Yiwei Bao, Mingjie Xu, Haofei Wang, Yunfei Liu, and Feng Lu. Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[25] Ruicong Liu, Yunfei Liu, Haofei Wang, and Feng Lu. Pnpga+: Plug-and-play domain adaptation for gaze estimation using model variants. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "text", + "content": "[26] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "text", + "content": "[27] Yunfei Liu, Ruicong Liu, Haofei Wang, and Feng Lu. Generalizing gaze estimation with outlier-guided collaborative adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3835-3844, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "type": "text", + "content": "[28] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Generated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 321, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 287, + 374 + ], + "type": "text", + "content": "[29] Takehiko Ohkawa, Yu-Jhe Li, Qichen Fu, Ryosuke Furuta, Kris M Kitani, and Yoichi Sato. Domain adaptive hand keypoint and pixel localization in the wild. In European Conference on Computer Vision, pages 68-87. Springer, 2022. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 376, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 441 + ], + "type": "text", + "content": "[30] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12999-13008, 2023. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 444, + 287, + 487 + ], + "type": "text", + "content": "[31] Vladimir I Pavlovic, Rajeev Sharma, and Thomas S. Huang. Visual interpretation of hand gestures for human-computer interaction: A review. IEEE Transactions on pattern analysis and machine intelligence, 19(7):677-695, 1997. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 489, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 287, + 533 + ], + "type": "text", + "content": "[32] Qucheng Peng, Ce Zheng, and Chen Chen. Source-free domain adaptive human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4826-4836, 2023. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 535, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 535, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 287, + 587 + ], + "type": "text", + "content": "[33] Thammathip Piumsomboon, Adrian Clark, Mark Billinghurst, and Andy Cockburn. User-defined gestures for augmented reality. In CHI'13 Extended Abstracts on Human Factors in Computing Systems, pages 955-960. 2013. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 623 + ], + "type": "text", + "content": "[34] Siddharth S Rautaray and Anupam Agrawal. Vision based hand gesture recognition for human computer interaction: a survey. Artificial intelligence review, 43:1-54, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "text", + "content": "[35] Dripta S Raychaudhuri, Calvin-Khang Ta, Arindam Dutta, Rohit Lal, and Amit K Roy-Chowdhury. Prior-guided source-free domain adaptation for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14996-15006, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[36] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 673 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[37] Stefan Vogt, Giovanni Buccino, Afra M Wohlschlager, Nicola Canessa, N Jon Shah, Karl Zilles, Simon B Eickhoff, Hans-Joachim Freund, Giacomo Rizzolatti, and Gereon R Fink. Prefrontal involvement in imitation learning of hand actions: effects of practice and expertise. Neuroimage, 37 (4):1371-1383, 2007. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "text", + "content": "[38] Christian Von Hardenberg and François Bérard. Bare-hand human-computer interaction. In Proceedings of the 2001 workshop on Perceptive user interfaces, pages 1-8, 2001. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 238 + ], + "type": "text", + "content": "[39] Yilin Wen, Hao Pan, Lei Yang, Jia Pan, Taku Komura, and Wenping Wang. Hierarchical temporal transformer for 3d hand pose estimation and action recognition from egocentric rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21243-21253, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 293 + ], + "type": "text", + "content": "[40] Wenhao Wu, Hau San Wong, and Si Wu. Semi-supervised stereo-based 3d object detection via cross-view consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17471-17481, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 339 + ], + "type": "text", + "content": "[41] Yinsong Xu, Zhuqing Jiang, Aidong Men, Haiying Wang, and Haiyong Luo. Multi-view feature fusion for person re-identification. Knowledge-Based Systems, 229:107344, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[42] Linlin Yang, Shicheng Chen, and Angela Yao. Semi-hand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "type": "text", + "content": "[43] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 442, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 545, + 496 + ], + "type": "text", + "content": "[44] Yuxiang Zhang, Zhe Li, Liang An, Mengcheng Li, Tao Yu, and Yebin Liu. Lightweight multi-person total motion capture using sparse multi-view cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5560-5569, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 498, + 545, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 562 + ], + "type": "text", + "content": "[45] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 565, + 545, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 628 + ], + "type": "text", + "content": "[46] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 632, + 545, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 632, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 307, + 632, + 545, + 673 + ], + "type": "text", + "content": "[47] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. Technical report, arXiv:1705.01389, 2017. https://arxiv.org/abs/1705.01389. 3,6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "686" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_content_list.json b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3a879a87cbea4e59413e076c1597dcd073ed73cc --- /dev/null +++ b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_content_list.json @@ -0,0 +1,1802 @@ +[ + { + "type": "text", + "text": "SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model", + "text_level": 1, + "bbox": [ + 107, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inhwan Bae, Young-Jae Park and Hae-Gon Jeon* \nAI Graduate School, GIST, South Korea", + "bbox": [ + 292, + 181, + 681, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{inhwanbae, youngjae.park}@gm.gist.ac.kr, haegonj@gist.ac.kr", + "bbox": [ + 217, + 219, + 745, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are five types of trajectory prediction tasks: deterministic, stochastic, domain adaptation, momentary observation, and few-shot. These associated tasks are defined by various factors, such as the length of input paths, data split and pre-processing methods. Interestingly, even though they commonly take sequential coordinates of observations as input and infer future paths in the same coordinates as output, designing specialized architectures for each task is still necessary. For the other task, generality issues can lead to sub-optimal performances. In this paper, we propose SingularTrajectory, a diffusion-based universal trajectory prediction framework to reduce the performance gap across the five tasks. The core of SingularTrajectory is to unify a variety of human dynamics representations on the associated tasks. To do this, we first build a Singular space to project all types of motion patterns from each task into one embedding space. We next propose an adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods that sometimes yield unacceptable paths, our adaptive anchor enables correct anchors, which are put into a wrong location, based on a traversability map. Finally, we adopt a diffusion-based predictor to further enhance the prototype paths using a cascaded denoising process. Our unified framework ensures the generality across various benchmark settings such as input modality, and trajectory lengths. Extensive experiments on five public benchmarks demonstrate that SingularTrajectory substantially outperforms existing models, highlighting its effectiveness in estimating general dynamics of human movements. Code is publicly available at https://github.com/inhwanbae/SingularTrajectory.", + "bbox": [ + 75, + 301, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extensive studies of trajectory prediction methods have been conducted in the computer vision field for several decades [20, 52]. They have demonstrated its importance in various applications, including crowd simulation, social robot navi", + "bbox": [ + 75, + 811, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6cf9b4473ac034a222fb40a9a888610f4aecf95e6ae1053e9f1b7693c80a41a.jpg", + "image_caption": [ + "Figure 1. An overview of our SingularTrajectory framework. All relevant human trajectory prediction tasks can be represented in our Singular space, a unified feature embedding space for human dynamics. Using the embedding features, our diffusion-based model for a universal trajectory prediction makes prediction for all the tasks in this same space." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 890, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gation, obstacle avoidance and security and surveillance systems, etc. Trajectory prediction takes sequential coordinate values of agents as input and infers their future pathways in common [1, 19, 83]. Such tasks vary depending on the application. The tasks are determined with respect to the number of input/output, data processing, and the use of geological information.", + "bbox": [ + 496, + 491, + 893, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The tasks for trajectory prediction can be categorized into five groups: (1) Stochastic prediction is to predict 20 multi-modal future trajectories from one observation with 8 frames [19]. Each future trajectory consists of 12 frames. (2) Deterministic prediction takes one observation with 8 frames, but infers only one future trajectory with 12 frames [1]. (3) The momentary observation uses only two frames to predict 20 multi-modal future paths with 12 frames [77]. (4) Domain adaptation splits training data with respect to places in datasets, trains a model on one place, and then checks the transferability to other places [93]. (5) The few-shot task only uses partial data to build a dataset-efficient model [53], as illustrated in Fig. 1. Until now, each specialized architecture for a task type has provided performance gains.", + "bbox": [ + 496, + 598, + 895, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, two questions arise. First, why do state-of-the-art models for one task undergo significant performance drops when applied to other trajectory prediction tasks? Second, is it feasible to design a general predictor that works across the five tasks?", + "bbox": [ + 496, + 810, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As answers to these questions, we present a universal tra", + "bbox": [ + 517, + 885, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17890", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jectionary predictor to achieve generality in predictions, named SingularTrajectory. The main idea is to unify the modalities of human dynamic representations across the five tasks. To do this, we first introduce a Singular space, an embedding space consisting of representative motion patterns for each task. The motion patterns play a role in the basis function for pedestrian movements, and are extracted using Singular Value Decomposition (SVD). They are then projected onto Singular space.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We next propose an environment-adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods [6, 29] that sometimes fail to handle different target data distribution, our adaptive anchor is able to correct prototype paths from the adaptive anchor in Singular space if they are put into the wrong locations, based on an input traversability map. Lastly, we generate socially-acceptable future trajectories for all agents in scenes with a diffusion-based predictor which denoises residuals of perturbed prototype paths. Thanks to the cascaded denoising process of diffusion models, we refine the prototype paths. Here, historical pathways, agent interactions, and environmental information are provided as conditions to guide them in the Markov chain of the denoising diffusion processes.", + "bbox": [ + 75, + 227, + 472, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results demonstrate that our SingularTrajectory can successfully represent pedestrian motion dynamics, and significantly improve prediction accuracy for the five tasks on challenging public benchmark datasets.", + "bbox": [ + 76, + 438, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 511, + 227, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review previous studies on trajectory prediction that have attempted to address various benchmark scenarios.", + "bbox": [ + 76, + 537, + 468, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Pedestrian Trajectory Prediction", + "text_level": 1, + "bbox": [ + 76, + 575, + 364, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ways of predicting pedestrian future trajectories have been studied for a long time in the computer vision field. Pioneering works [20, 52, 60, 95] model an invisible social norm based on motion dynamics as an energy minimization problem. Introducing recurrent neural networks [1, 19, 64, 97] achieves significant improvements by providing highly sequential representations of high-level shape of paths. These methods determine the most probable path, called deterministic trajectory prediction. The following works model mutual influences among agents using attention mechanisms [17, 23, 64, 83], graph convolutional networks [2, 9, 27, 41, 53, 54, 75], graph attention networks [5, 22, 28, 38, 39, 66, 82], and transformers [4, 18, 55, 68, 86, 87, 97, 98]. Additional visual information allows us to leverage environmental constraints from traversability maps [13-15, 28, 37, 46, 47, 49, 50, 63, 65, 74, 75, 79, 81, 85, 94, 99, 102]. Depending on the constraints, predictors take either recurrent [1, 8, 10, 11, 18, 19, 30, 43, 44, 51, 56, 61, 64, 92, 100, 101] or simultaneous approaches [2, 3, 35, 53, 66, 67, 89] to extrapolate the future pathways.", + "bbox": [ + 75, + 598, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, with the success of generative models, the importance of multimodality has begun to emerge, called stochastic trajectory prediction. Stochastic prediction enables us to consider all of an agent's possible future pathways. For example, an agent at a crossroads may either walk straight or turn left/right. Here, the stochastic prediction infers all potential future modes. This approach has become mainstream in this field. Starting from Social-GAN [19], a bivariate Gaussian distribution [1, 2, 34, 53, 66, 69, 70, 93, 96, 97], generative adversarial network [14, 19, 22, 28, 36, 40, 63, 74, 78, 102], and conditional variational autoEncoder [7, 10, 23, 30, 31, 33, 45, 64, 76, 84, 88, 91] have been adopted for stochastic trajectory prediction. Anchorm-conditioned methods can explicitly represent different modalities by prototyping possible paths [6, 29]. Most recently, diffusion-based models have revealed their tremendous representation capacities in numerous tasks [16, 21, 57, 71-73], proving its potential for stochastic trajectory prediction [18, 25, 48, 62]. In this study, we take full advantage of both the anchor-conditioned approach and the diffusion-based model to achieve the explainability and generalizability of the trajectory prediction tasks.", + "bbox": [ + 496, + 90, + 895, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Various Trajectory Prediction Tasks", + "text_level": 1, + "bbox": [ + 500, + 431, + 815, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond the standard benchmark protocol of stochastic prediction, there are three other variants of this task: momentary observation, domain adaptation, and few-shot learning. Works in [12, 55, 77] only take two frames as input for the momentary trajectory prediction. Multi-task learning, self-supervised learning, and knowledge distillation techniques have been used to extract rich features from the limited data. Another works [24, 93, 103] focuses on domain adaptation across trajectory domains, captured from different surveillance views. A transferable graph neural network is introduced to adaptively learn domain-invariant knowledge. The others [53, 101] adopt few-shot learning for better training efficiency.", + "bbox": [ + 496, + 454, + 893, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these works take and infer the sequential coordinates of agent trajectories in common, there is no unified model for all the tasks. Despite the tremendous efforts to design a specialized architecture for one task, they cannot be applied to the other tasks without suffering significant performance drops. In the next section, we will describe how to design a unified architecture, which consistently produces promising results on the five associated tasks.", + "bbox": [ + 496, + 651, + 893, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 784, + 635, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We describe how to learn a general representation of human motions. We first define a general trajectory prediction problem in Sec. 3.1 and provide preliminaries of explicit formulations on the SVD and diffusion process in Sec. 3.2. We then introduce a motion feature extraction to build our Singular space, and a projection of any trajectory from each task onto", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17891", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "it in Sec. 3.3. Next, we propose an environment-adaptive anchor using motion vectors and input image in Sec. 3.4. Finally, we present the SingularTrajectory predictor based on the diffusion model in Sec. 3.5.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 159, + 261, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Trajectory prediction aims to predict the future paths of agents based on their historical path and surrounding contexts. Suppose that at each timestamp $t$ , there are $N$ pedestrians in a scene with the 2D spatial coordinate position $\\{\\pmb{p}_t^n \\in \\mathbb{R}^2 | n \\in [1, \\dots, N]\\}$ . A pedestrian historical trajectory $\\pmb{X}_n$ over $T_{hist}$ timesteps can be represented as the cumulative coordinates $\\pmb{X}_n = \\{\\pmb{p}_t^n | t \\in [1, \\dots, T_{hist}]\\}$ . Similarly, future trajectories $\\pmb{Y}_n$ for the time duration $T_{fut}$ to be predicted can be written as $Y_n = \\{\\pmb{p}_t^n | t \\in [T_{hist} + 1, \\dots, T_{hist} + T_{fut}]\\}$ . The prediction system takes both the historical trajectories for all $N$ people $X = \\{X_n | n \\in [1, \\dots, N]\\}$ and the scene image map $I$ for environmental information as input. The deterministic prediction system infers one sequence of the most reliable future trajectory $\\widehat{Y} = \\{\\widehat{Y}_n | n \\in [1, \\dots, N]\\}$ . For the stochastic prediction, because of the indeterminacy of the future movements, $S$ multiple pathways for all the $N$ pedestrians $\\widehat{Y} = \\{\\widehat{Y}_n^s | n \\in [1, \\dots, N], s \\in [1, \\dots, S]\\}$ are generated so that at least one sample is close to the ground-truth trajectory.", + "bbox": [ + 75, + 181, + 472, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 463, + 217, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Singular Value Decomposition. Singular Value Decomposition (SVD) decomposes a matrix into three resultant matrices. Given a matrix $\\mathbf{A}$ , its SVD is represented as:", + "bbox": [ + 75, + 486, + 470, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} = \\boldsymbol {U} \\boldsymbol {\\Sigma} \\boldsymbol {V} ^ {\\top}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 539, + 468, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{U}$ is an orthogonal left singular vector matrix, whose columns are eigenvectors of $A\\times A^{\\top}$ . $\\pmb{\\Sigma}$ is a diagonal matrix with the singular values of $A$ , consisting of $K$ non-negative values in descending order. $\\pmb{V}$ is a right singular vector matrix, which is also orthogonal and its columns are the eigenvectors of $A^{\\top}\\times A$ .", + "bbox": [ + 75, + 566, + 468, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To remove the redundant part of the raw data, the truncation technique is often applied to the results after the decomposition. The idea behind the truncated SVD is to approximate the original matrix $\\mathbf{A}$ with the lower rank. With the $K$ to determine the number of singular values to retain, $\\pmb{\\Sigma}$ can be simplified to $\\pmb{\\Sigma}_{K}$ which contains only the $K$ largest singular values. Similarly, $\\pmb{U}$ and $V^{\\top}$ are reduced to $U_{K}$ and $V_{K}^{\\top}$ by keeping the first $K$ columns and rows, respectively. This process eliminates the smallest singular values, which are not needed to express the original data and often correspond to noise or redundant information. This is useful for practical scenarios dealing with large and potentially sparse matrices because we can reconstruct a close approximation of the original data with significantly less storage space.", + "bbox": [ + 75, + 657, + 470, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. The diffusion model operates by transforming a noisy distribution, represented by the noise vector", + "bbox": [ + 75, + 869, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{y}_M$ , into the desired data $\\mathbf{y}_0$ through a series of $M$ diffusion steps. These steps involve intermediate latent variables $\\{\\mathbf{y}_m|m\\in [1,\\dots ,M]\\}$ , and encompass both the diffusion and denoising processes. The diffusion process adds a small amount of noise to data in order to obtain the standard normal distribution $q(\\mathbf{y}_M)$ from the distribution $q(\\mathbf{y}_0)$ using the Markov chain as:", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {y} _ {1: M} \\mid \\mathbf {y} _ {0}\\right) := \\prod_ {m = 1} ^ {M} q \\left(\\mathbf {y} _ {m} \\mid \\mathbf {y} _ {m - 1}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 204, + 890, + 244 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {y} _ {m} \\mid \\mathbf {y} _ {m - 1}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {m}; \\sqrt {1 - \\beta_ {m}} \\mathbf {y} _ {m - 1}, \\beta_ {m} \\mathbf {I}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 246, + 852, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ is a small positive constant and a variance schedule for adding noise. The denoising process uses the $\\mathbf{y}_m$ to recover $\\mathbf{y}_0$ with a learnable network as follows:", + "bbox": [ + 496, + 273, + 890, + 319 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {y} _ {0: M}\\right) := p \\left(\\mathbf {y} _ {M}\\right) \\prod_ {m = 1} ^ {M} p _ {\\theta} \\left(\\mathbf {y} _ {m - 1} \\mid \\mathbf {y} _ {m}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 329, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\mathbf {y} _ {m - 1} | \\mathbf {y} _ {m}) := \\mathcal {N} (\\mathbf {y} _ {m - 1}; \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {y} _ {m}, m), \\beta_ {m} \\mathbf {I}).\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 371, + 851, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{y}_M\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ is an initial noise sampled from the Gaussian distribution $p(\\mathbf{y}_M)$ , and $\\theta$ denotes the learnable parameter of the diffusion model. $\\epsilon_{\\theta}$ is a learnable denoising model of a clean data $\\mathbf{y}_0$ , and a corrupted data $\\mathbf{y}_m$ at a step $m$ . The objective is to train the neural network so that the denoising process predicts the true data-generating distribution well. This is often done by maximizing the evidence lower bound, ensuring that the samples generated by the diffusion model are indistinguishable from the real data.", + "bbox": [ + 496, + 398, + 893, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Unifying the Motion Space", + "text_level": 1, + "bbox": [ + 500, + 542, + 741, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The trajectory prediction model uses a learnable network to capture the relationship in consideration of the input coordinates of pedestrians, input images and output coordinates for each pedestrian. Since expectations for input and output trajectories are different for each associated task (e.g., length and multimodality), they should be viewed as different data spaces even if they use the same coordinate systems. We introduce a method to merge raw data in each space into a Singular space for human motion dynamics.", + "bbox": [ + 496, + 566, + 893, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Singular space construction. To discover motion dynamics from the raw data, we first extract primitive motion features. Inspired by the successful low-rank approximation of raw trajectory data using eigenvectors in [6], we also employ a similar strategy using singular vectors from the truncated SVD to extract principal motion components from the entire training dataset.", + "bbox": [ + 496, + 704, + 893, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we cut-off paths of all pedestrians in the dataset into $T_{win}$ lengths through a sliding window to create a total of $L$ gist of trajectory set $A \\in \\mathbb{R}^{L \\times (2 \\times T_{win})}$ . Here, $A$ is a temporary matrix to extract a motion vector from a set of trajectory in the initialization phase. Next, we decompose $A$ to obtain truncated matrices $U_K \\in \\mathbb{R}^{L \\times K}$ , $\\pmb{\\Sigma}_K \\in \\mathbb{R}^{K \\times K}$", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17892", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab4c686765e04400c9a24dc333bb76be181683d3ab945f60b722742b7b8d2dc7.jpg", + "image_caption": [ + "Figure 2. Visualization of the trajectories in Singular space. (a) The circle and triangle markers indicate the history and future trajectory coordinates, respectively. Each color also refers to each associated task. (b-d) Each marker in the Singular space corresponds to each trajectory in raw data. And, the slicing planes, representing a set of arrows, mean human dynamics of straight forward motions and turning motions." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 184, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a07dc740514b4b0cc3fe84a0295dacfd257db2a5019d448d04a1fd5208c1175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 90, + 277, + 171 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/897df13df0786f68e036ad5758762e54972a06bc1b36f6c231928b4a098606df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 90, + 372, + 171 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f800c2596ad3f2d2c80ab498aa9a7be8a1395a9e5b21235f181adb944a62389.jpg", + "image_caption": [ + "Figure 3. An example of the adaptive anchor generation. (a) The initial prototype anchor $\\mathcal{P}$ is placed on the last observation coordinate of a person. In this instance, four prototype paths (highlighted in red) are incorrectly placed at the non-traversable locations. (b) Vector field $\\vec{F}_I(x,y)$ is computed to guide toward in the nearest traversable areas. (c) The initial prototype paths are then tailored to the environment using the vector field." + ], + "image_footnote": [], + "bbox": [ + 380, + 90, + 468, + 171 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $V_{K}^{\\top}\\in \\mathbb{R}^{K\\times (2\\times T_{win})}$ . Here, $V_{K}^{\\top}$ is a set of $K$ spatiotemporal motion vectors $\\mathrm{v}_k\\in \\mathbb{R}^{2\\times T_{win}}$ representing the most dominant motion dynamics of pedestrians. Since $\\mathrm{v}_k$ are orthogonal to each other, we define a Singular space coordinate system using the $K$ singular vectors as basis vectors for each axis. In Singular space, trajectories $\\pmb{A}$ can be a coordinate of $\\mathcal{A} = U_K\\cdot \\pmb {\\Sigma}_K\\in \\mathbb{R}^{L\\times K}$ , indicating that each motion vector has an influence on reconstructing the $L$ trajectories. In other words, we can project the $\\pmb{A}$ into the coordinate in Singular space $\\mathcal{A}$ as follows:", + "bbox": [ + 75, + 295, + 472, + 449 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} = \\boldsymbol {A} \\times \\left(\\boldsymbol {V} _ {K} ^ {\\top}\\right) ^ {- 1} = \\boldsymbol {A} \\times \\boldsymbol {V} _ {K}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 457, + 468, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\mathbf{V}_K^\\top)^{-1}$ can be simplified into $\\mathbf{V}_K$ with the property of an orthogonal matrix1.", + "bbox": [ + 75, + 484, + 468, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the Singular space, we can now concentrate on the motion flow by using the coefficients of motion patterns, instead of considering every position over time. Note that the results from the SVD depend highly on various factors such as the window size for inputs/outputs, data processing, and the differences in the raw data caused by geological variations. Due to the issue, works using SVD only convert the output space [26, 59, 80] or conduct separate decompositions for each of input and output data spaces [6]. However, in this case, they cause inconsistent representations of motion dynamics, even for the same pedestrian, which leads to a lack of generality. In the next step, we introduce a new method to address this issue.", + "bbox": [ + 75, + 517, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Projection of any trajectory into Singular space. We aim to express the input/output trajectories $\\mathbf{X}$ and $\\mathbf{Y}$ of the five associated tasks in Singular space all at once. Our core idea stems from the notion that all pedestrians share the same human motion dynamics and thus will likely show a similar pattern. Starting with the projection function Eq. (4), which is a projection matrix for the fixed trajectory length $T_{win}$ , we extend it to any length $T_{hist}$ and $T_{fut}$ , which varies from task to task. To handle the motion patterns regardless of their lengths, we interpolate $\\mathrm{v}_k \\in \\mathbb{R}^{2 \\times T_{win}}$ to $\\mathrm{v}_{\\mathrm{x},k} \\in \\mathbb{R}^{2 \\times T_{hist}}$ .", + "bbox": [ + 75, + 714, + 470, + 866 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/daba69e12be8e95b117f5a30ce2d5f08c28f51143f2fa1557b9c7bdca9a2c8ec.jpg", + "image_caption": [ + "(a) Initial prototype paths" + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 629, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/70ab6f266e750f341fc8ab0e1381dafd4e118302daabb102f0be48a0ca6f9bb8.jpg", + "image_caption": [ + "(b) Vector field" + ], + "image_footnote": [], + "bbox": [ + 632, + 88, + 759, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d2215bdab00f8ff960015d0cfb9cfc00536b11e95ed457dd3130f3343b5f196b.jpg", + "image_caption": [ + "(c) Final prototype paths" + ], + "image_footnote": [], + "bbox": [ + 763, + 88, + 892, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the motion pattern can be seen as a 2-dimensional curve, we use Cardinal B-splines to make a transformation matrix $C_{T_{hist}} \\in \\mathbb{R}^{(2 \\times T_{hist}) \\times (2 \\times T_{win})}$ using the Irwin-Hall distribution. The constant value $C$ depends only on the length, so $\\mathrm{v}_{\\mathrm{x},k}$ can be approximated. The trajectory $\\mathbf{X} \\in \\mathbb{R}^{N \\times (2 \\times T_{hist})}$ is projected to the coordinate $\\mathcal{X} \\in \\mathbb{R}^{N \\times (2 \\times T_{hist})}$ in Singular space as follows:", + "bbox": [ + 496, + 305, + 893, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {x, k} = C _ {T _ {\\text {h i s t}}} \\times v _ {k}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 424, + 890, + 440 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} = \\boldsymbol {X} \\times \\boldsymbol {C} _ {T _ {\\text {h i s t}}} \\times \\boldsymbol {V} _ {K}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 449, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the same way as Eq. (6), the trajectory $\\mathbf{Y} \\in \\mathbb{R}^{N \\times (2 \\times T_{\\text{fit}})}$ also can be projected to $\\mathcal{V} \\in \\mathbb{R}^{N \\times K}$ using $\\mathbf{V}_K$ and $C_{T_{\\text{fit}}}$ .", + "bbox": [ + 496, + 470, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through this process, Singular space can represent a variety of trajectories even with different lengths, including motion vectors for the input/output as well as the task-specific data, as visualized in Fig. 2. Establishing a common ground for the trajectory representation is crucial for our model's adaptability and robust performances across different trajectory prediction benchmarks. Moreover, the strategy of focusing on the overall motion flow, rather than frame-by-frame coordinates, further improves the model's capacity to understand and predict socially compliant trajectories in diverse real-world scenarios.", + "bbox": [ + 496, + 503, + 893, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Adaptive Anchor", + "text_level": 1, + "bbox": [ + 500, + 679, + 668, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Beyond the trajectory data integration, we also introduce how to incorporate environmental contexts into trajectory predictors. To do this, we propose an adaptive anchor which is expressed as a set of motion vectors from input images and consists of prototype paths in Singular space.", + "bbox": [ + 496, + 702, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prototype anchor formation. Previous anchor-based human trajectory prediction approaches [6, 29] use a fixed anchor for any pedestrian, and the anchor is refined for output trajectories. Although these methods can explicitly model the multimodality, they sometimes fail to handle the case where the prototype path is put into wrong locations, or blocked by static obstacles. This is because the wrong prototypes are treated as a hard constraint. To avoid this problem,", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1For better understanding, we display the coordinate variable in the Singular space using the calligraphic font", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17893", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we use the input image as a traversability map to correct the wrong prototype paths.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start with the converted future trajectory $\\mathcal{V}$ in the whole dataset $Y$ using Sec. 3.3. Following [6, 59], the coordinates $\\mathcal{V}$ are then clustered into $S$ centroids, which can be viewed as groups of components representing different multimodal futures. In addition, each centroid in Singular space is used as an initial prototype path $\\mathcal{P}_s$ for constructing an anchor $\\mathcal{P} \\in \\mathbb{R}^{S \\times K}$ . But, these initial prototype paths are still fixed, so they still have a limitation which is unlikely to consider environmental information.", + "bbox": [ + 75, + 121, + 468, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive anchor generation. To make use of environmental information, we introduce a module to deform the anchor. Using an off-the-shelf semantic segmentation model from [46], we can obtain binary traversable maps $\\mathbf{I}_{\\text{map}}$ from an input image. We then derive a vector field $\\vec{F}_I(x,y)$ to fix the wrongly located prototype path by directing it to a nearest traversable regions. If the initial prototype path $\\mathcal{P}_s$ is in the wrong place, we deform it by adding the vector fields into the initial anchor until they reach equilibrium states to obtain the final prototype paths $\\mathcal{P}_s'$ as follows:", + "bbox": [ + 75, + 260, + 470, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {s} ^ {\\prime} = \\mathcal {P} _ {s} + \\vec {F} _ {I} \\left(\\mathcal {P} _ {s} \\boldsymbol {V} _ {K} ^ {\\top} \\boldsymbol {C} _ {T _ {f u t}} ^ {- 1}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 425, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this process, the prototype path is re-located toward the nearest walkway, as demonstrated in Fig. 3. In other words, the anchor plays a role in the environment-adaptive prototype paths, unlike the existing fixed anchor methods.", + "bbox": [ + 75, + 455, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because the prototype paths are in Singular space, the scene image can be implicitly represented as the adaptive anchor. In addition, the scene image can be approximated with the set of motions. By projecting the scene images into the Singular space, our model understands the surrounding environment. This holistic approach, integrating both coordinate and environmental cues, sets the stage for more realistic and reliable trajectory predictions.", + "bbox": [ + 75, + 517, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Diffusion-Based SingularTrajectory Model", + "text_level": 1, + "bbox": [ + 76, + 648, + 441, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the final step, we develop a framework called Singular-Trajectory, a unified model that works well across the five tasks. Leveraging the Singular space and the adaptive anchor, our diffusion-based SingularTrajectory model can precisely forecast potential future paths.", + "bbox": [ + 75, + 671, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denoising a perturbed trajectory anchor. Unlike previous diffusion-based predictors, which directly forecast the future path from Gaussian noise [18, 25], we devise a stepwise refinement from the adaptive anchor for a realistic trajectory. Here, historical pathways $\\mathcal{X}$ , environmental information $\\mathcal{P}$ , and agent interactions $\\mathcal{G}(\\mathcal{X},\\mathcal{P})$ are encoded and used as conditions to guide the denoising processes. To better capture agents interactions $\\mathcal{G}$ in the denoising process, we adopt the transformer model. Similar to other transformer-based trajectory predictors [18, 48], our SingularTrajectory", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encodes spatio-temporal information to account for agent-agent and agent-environment interactions using $\\mathcal{X}$ and $\\mathcal{P}$ , respectively. These conditions are then concatenated into one feature vector representation and fed into the diffusion model $\\epsilon_{\\theta}(\\mathbf{y}_m, m, \\mathcal{X}, \\mathcal{P}, \\mathcal{G})$ to learn, by contrasting the motion patterns from previous diffusion steps to distinguish the added noise, and to close the reality gap by generating socially-acceptable future paths $\\widehat{\\mathbf{Y}}$ .", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This refinement process works in a cascading manner by $\\{\\mathcal{P},\\dots,\\widehat{\\mathcal{V}}\\}_{m = 1}^{M}$ as follows:", + "bbox": [ + 500, + 213, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {Y}} = \\mathcal {P} + p \\left(\\mathbf {y} _ {M}\\right) \\prod_ {m = 1} ^ {M} p _ {\\theta} \\left(\\mathbf {y} _ {m - 1} \\mid \\mathbf {y} _ {m}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 246, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By predicting only the residuals $\\mathbf{y}_m\\in \\mathbb{R}^{S\\times K}$ to adjust the anchors, the problem is simplified, in that the model is able to use a prior knowledge on the initial state. With this process, we thus ensure more precise and reliable trajectory generations, achieving generality in various environments and applications.", + "bbox": [ + 496, + 287, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. To construct Singular space, we empirically set $K$ to 4. We set $T_{win} = T_{fut} = 12$ in order to prevent information loss due to the approximation of the motion vector during prediction. For an anchor diffusion model, we devise a one-layer transformer for encoding motion and context information, where the dimension is set to 256 with 4-head attention. We set $M = 10$ and schedule the diffusion timesteps following DDIM [72]. To train the SingularTrajectory in an end-to-end manner, we use a mean square error (MSE) as a loss function between the output and a random Gaussian noise for the current iteration. The training is performed with AdamW optimizer [42], with a learning rate of 0.001 and batch size of 512 for 256 epochs. All the experiments are conducted on a single NVIDIA A6000 GPU, which usually takes about an hour to train each scene.", + "bbox": [ + 496, + 382, + 893, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 626, + 633, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive experiments to verify the generality of our SingularTrajectory model for the trajectory prediction tasks. We first describe our experimental setup in Sec. 4.1. We then provide comparison results with state-of-the-art models on public benchmark datasets in Sec. 4.2. Finally, we perform an extensive ablation study demonstrating the effect of each component of our method in Sec. 4.3.", + "bbox": [ + 496, + 652, + 893, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 785, + 689, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To compare our SingularTrajectory with state-of-the-art baselines, we conduct quantitative evaluations on two common datasets, ETH [60] and UCY [32], for all the associated tasks. The ETH and UCY datasets consist of various motions of 1,536 pedestrians across five unique scenes: ETH, Hotel, Univ, Zara1 and Zara2. They are recorded in", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17894", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bf7eb52475bd8cc9e1048e212b59dde79bc1e7613522c0ae027bb6ae0da8a734.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Domain Adaptation (ADE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAVG
STGCNN [53]1.831.581.301.311.513.021.382.631.582.151.160.700.820.540.811.041.050.730.470.820.981.090.740.500.831.22
PECNet [45]1.971.681.241.351.563.111.352.691.622.191.390.820.930.570.931.101.170.920.520.931.011.250.830.610.931.31
AgentFormer [98]1.731.641.871.571.701.491.762.131.731.781.730.192.011.791.431.712.051.921.321.751.961.911.891.911.921.72
MID [18]1.261.431.651.431.441.230.870.980.770.961.610.630.810.580.911.420.750.810.740.931.470.680.880.910.981.04
EigenTrajectory [6]0.390.710.730.450.570.950.680.510.420.640.940.320.490.370.531.101.031.360.430.980.920.510.620.520.640.67
T-GNN [93]1.131.250.941.031.092.541.082.251.411.820.970.540.610.230.590.880.780.590.320.640.870.720.650.340.650.96
K0 [24]0.450.780.680.590.630.870.650.490.360.591.080.460.440.360.591.130.560.590.530.701.130.390.590.480.650.63
Adaptive [24]0.360.640.530.450.500.900.620.500.360.601.110.460.480.420.621.110.520.540.470.661.050.340.560.440.600.59
SingularF trajectory0.290.590.510.420.450.660.550.450.350.500.790.300.480.360.480.730.290.550.360.490.640.230.550.450.470.48
Domain Adaptation (FDE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAvgV
STGCNN [53]3.242.862.532.432.775.162.514.862.883.852.301.341.741.101.622.211.991.410.881.622.102.051.471.011.662.30
PECNet [45]3.332.832.532.452.795.232.484.902.863.872.221.321.681.121.592.202.051.520.881.662.101.841.450.981.592.29
AgentFormer [98]3.603.544.523.953.903.763.675.173.424.013.773.494.123.763.793.883.754.273.613.883.823.553.744.323.863.89
MID [18]2.422.773.222.772.802.341.661.871.441.833.131.281.741.241.852.731.491.671.541.862.871.411.881.932.022.07
EigenTrajectory [6]0.761.431.560.961.182.101.381.070.861.352.010.641.100.821.142.421.031.360.971.442.040.511.291.111.241.27
T-GNN [93]2.182.251.781.842.014.151.824.042.533.141.911.121.300.871.301.921.461.250.651.321.861.451.280.721.331.82
K0 [24]0.941.721.511.291.371.811.331.020.751.232.170.910.940.811.212.180.981.201.031.352.120.681.221.031.261.28
Adaptive [24]0.691.341.120.911.021.851.271.050.761.232.260.921.030.941.292.231.021.171.001.362.070.631.190.941.211.22
SingularF trajectory0.571.191.080.810.911.161.170.970.731.011.360.621.110.790.971.220.521.180.730.921.070.431.150.940.900.94
", + "bbox": [ + 80, + 88, + 888, + 343 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/88a58d0d1d2c39bcb78dd3cf5d0724d68cd52ec4ba9a7428664403b183df1baf.jpg", + "table_caption": [ + "Table 1. Comparison of our SingularTrajectory with other state-of-the-art methods in domain adaptation task. Following [24, 93], we deterministically predict one future path for all dataset pairs (ADE/FDE). Bold: Best, Underline: Second best." + ], + "table_footnote": [ + "Table 2. Evaluation on the stochastic trajectory prediction task." + ], + "table_body": "
StochasticETHHOTELUNIVZARA1ZARA2AVG
Social-GAN [19]0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
STGCNN [53]0.64/1.110.49/0.850.44/0.790.34/0.530.30/0.480.44/0.75
STAR [97]0.57/1.110.19/0.370.35/0.750.26/0.570.25/0.580.33/0.68
PECNet [45]0.61/1.070.22/0.390.34/0.560.25/0.450.19/0.330.32/0.56
MID [18]0.57/0.930.21/0.330.29/0.550.28/0.500.20/0.370.31/0.54
LBEBM [58]0.60/1.060.21/0.380.28/0.540.21/0.390.15/0.300.29/0.53
Trajectory++[64]0.61/1.030.20/0.280.30/0.550.24/0.410.18/0.320.31/0.52
GroupNet [88]0.46/0.730.15/0.250.26/0.490.21/0.390.17/0.330.25/0.44
AgentFormer [98]0.46/0.800.14/0.220.25/0.450.18/0.300.14/0.240.23/0.40
GP-Graph [4]0.43/0.630.18/0.300.24/0.420.17/0.310.15/0.290.23/0.39
NPSN [5]0.36/0.590.16/0.250.23/0.390.18/0.320.14/0.250.21/0.36
EqMotion [90]0.40/0.610.12/0.180.23/0.430.18/0.320.13/0.230.21/0.35
EigenIProjection [6]0.36/0.530.12/0.190.24/0.430.19/0.330.14/0.240.21/0.34
SocialVAE [91]0.41/0.580.13/0.190.21/0.360.17/0.290.13/0.220.21/0.33
LED [48]0.39/0.580.11/0.170.26/0.430.18/0.260.13/0.220.21/0.33
SingularProjection0.35/0.420.13/0.190.25/0.440.19/0.320.15/0.250.21/0.32
", + "bbox": [ + 78, + 388, + 465, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "surveillance views, and the trajectories are labeled in world coordinates.", + "bbox": [ + 76, + 623, + 468, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarks. We evaluate our model for the five trajectory prediction tasks: (1) stochastic prediction, (2) deterministic prediction, (3) momentary observation, (4) domain adaptation, and (5) few-shot learning. The stochastic trajectory prediction employs a leave-one-out strategy [5, 19, 22, 53, 66] for training and inference across the five ETH-UCY scenes. In the scenes, the initial 3.2 seconds (equivalent to $T_{hist} = 8$ frames) of all paths are used as input, and the subsequent 4.8 seconds (corresponding to $T_{fut} = 12$ frames) are predicted. The model generates $S = 20$ paths, and one with the lowest error is chosen as the final prediction. In the deterministic prediction, the trajectory prediction methods generate only $S = 1$ path. For the momentary observation, the observation length is $T_{hist} = 2$ whereas still $T_{fut} = 12$ . For the domain adaptation, models are trained on each of the five scenes individually, and then evaluated on the other scenes, instead", + "bbox": [ + 76, + 659, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/284d9aa57509c06e041a206f4e55b9583e94d28863f9924419dd5b12658fe71c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DeterministicETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.04/1.930.88/1.630.71/1.340.66/1.220.57/1.060.77/1.44
PECNet [45]1.20/2.730.68/1.510.78/1.710.82/1.850.62/1.460.82/1.85
AgentFormer [98]1.67/3.651.61/3.591.62/3.531.85/4.131.68/3.741.69/3.73
MID [18]1.42/2.940.64/1.300.76/1.620.74/1.590.60/1.310.83/1.75
EigenTrajectory [6]0.93/2.050.33/0.640.58/1.230.45/0.990.34/0.750.53/1.13
Social-LSTM [1]1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social-GAN [19]1.13/2.211.01/2.180.60/1.280.42/0.910.52/1.110.67/1.41
SR-LSTM [100]1.01/1.930.35/0.720.66/1.380.56/1.230.44/0.900.60/1.23
STAR-D [97]0.97/2.000.32/0.730.56/1.250.44/0.960.35/0.770.53/1.14
Trajectory++[64]1.02/2.090.33/0.630.52/1.160.42/0.940.32/0.710.52/1.11
SingularTrajectory0.72/1.230.27/0.500.57/1.120.44/0.930.35/0.730.47/0.90
", + "bbox": [ + 501, + 388, + 888, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Evaluation on the deterministic trajectory prediction task.", + "bbox": [ + 500, + 526, + 890, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of using the leave-one-out strategy. In the few-shot learning, they are trained using only $10\\%$ of the training set, and then assessed on the entire test set. To measure the prediction performances, two common metrics are used for all the tasks, Average Displacement Error (ADE) and Final Displacement Error (FDE), measuring the average and destination error between prediction and its ground-truth path, respectively.", + "bbox": [ + 496, + 570, + 893, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Results", + "text_level": 1, + "bbox": [ + 500, + 690, + 679, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stochastic prediction task. In Tab. 2, since the ETH/UCY dataset release, many state-of-the-art models have made new records. Among them, our model also achieves performance competitive with the comparison methods. Our model regards historical, social, and environmental conditions of observations and images as the most important factors for successfully generating multimodal trajectories.", + "bbox": [ + 496, + 715, + 893, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Deterministic prediction task. Next, we evaluate our model in the deterministic trajectory prediction task in Tab. 3. Our SingularTrajectory model has the best performance over both these latest state-of-the-art models, and the models optimized for deterministic prediction. In particular, our model", + "bbox": [ + 496, + 824, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17895", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/41c74b1faede7cefd483bb93b6f6c2953df2b98c4bf7a1fa1443ac22303b1f93.jpg", + "image_caption": [ + "Figure 4. Visualization of prediction results on (a) momentary observation task and (b) few-shot task. To aid visualization, the best trajectory among $S = 20$ samples are reported." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 893, + 231 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d2fe4c99c3b4464a9a031de5da62bb4c5b629fc36bd4beb0d43fc6207b520ad5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MomentaryETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.24/2.230.77/1.440.45/0.810.38/0.570.35/0.580.64/1.13
PECNet [45]0.63/1.040.28/0.530.28/0.490.25/0.440.19/0.340.33/0.57
AgentFormer [98]1.10/2.110.50/1.020.52/1.100.56/1.180.43/0.890.62/1.26
MID [18]0.63/1.050.29/0.490.30/0.560.30/0.560.22/0.400.35/0.61
EigenF trajectory [6]0.46/0.760.17/0.280.25/0.440.19/0.350.15/0.270.25/0.42
STT [55]0.72/1.450.48/0.480.53/1.090.64/1.210.44/0.880.57/0.93
STT+DTO [55]0.62/1.220.29/0.560.58/1.140.45/0.980.34/0.740.46/0.93
MOE-Next [77]0.71/1.570.30/0.580.52/1.120.38/0.810.33/0.730.45/0.96
MOE-Traj++ [77]0.64/1.120.20/0.330.33/0.620.22/0.420.17/0.320.31/0.56
SingularF trajectory0.45/0.670.18/0.290.24/0.430.19/0.330.17/0.280.25/0.40
", + "bbox": [ + 78, + 284, + 467, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exhibits significant performance improvements in the ETH and HOTEL scenes. These scenes have noisy observation paths, which have a negative impact on the prediction. We are able to achieve accurate predictions by leveraging the overall motion flow, which acts as a low-pass filter over the noisy sequence, in Singular space.", + "bbox": [ + 75, + 455, + 468, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Momentary observation task. In Tab. 4, the state-of-the-art predictors for momentary observation mainly focus on the coordinates themselves, leading to a performance drop, given only two frames. Some models for this task try to bridge these gaps, but have not been sufficient. Fortunately, with the benefit of our Singular Space, even with only a two-frame input, our model can successfully represent the overall long-term flow. Note that our SingularTrajectory with two-frame observation demonstrates the closest performance to state-of-the-art stochastic prediction models with an entire history frame; even this is achieved without any masked trajectory complement or knowledge distillation. Consequently, this allows it to accurately pinpoint the future locations of pedestrians, whose examples are displayed in Fig. 4.", + "bbox": [ + 75, + 549, + 472, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain adaptation task. We next evaluate the performance of SingularTrajectory in a domain adaptation task. For simplicity, the ETH, HOTEL, UNIV, ZARA1, and ZARA2 scenes are denoted as A, B, C, D, and E, respectively. For example, 'A2B' means that a model is trained on the ETH scene and tested on the HOTEL scene. As demonstrated in Tab. 1, our SingularTrajectory model shows the performance nearly equivalent to those of models specifically designed for deterministic prediction. Particularly,", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/17c24b334870d57838e72b03acd3b3d2cd90451a9b7c7e057d8beaa4ab5d7f6f.jpg", + "table_caption": [ + "Table 4. Evaluation on the momentary observation task." + ], + "table_footnote": [], + "table_body": "
Few-ShotETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]0.89/1.631.22/2.480.90/1.610.68/1.251.36/2.121.01/1.82
PECNet [45]0.72/1.460.29/0.530.58/0.930.27/0.440.22/0.380.41/0.75
AgentFormer [98]1.60/2.651.02/1.641.13/1.901.19/2.011.08/1.591.20/1.96
MID [18]0.57/0.920.21/0.330.32/0.600.27/0.490.24/0.420.32/0.55
EigenF trajectory [6]0.39/0.640.13/0.210.25/0.430.21/0.390.15/0.270.23/0.39
SingularF trajectory0.35/0.460.14/0.210.26/0.440.21/0.360.18/0.310.23/0.35
", + "bbox": [ + 501, + 284, + 890, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Evaluation on the few-shot trajectory prediction task.", + "bbox": [ + 509, + 369, + 879, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our model produces impressive results in the challenging B2A, C2A, D2A and E2A scenarios where even models specialized in domain adaptation fail. Our SingularTrajectory framework specializes in learning general human motions, and is not limited to a specific domain, enabling accurate predictions even in extreme cases.", + "bbox": [ + 496, + 411, + 893, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot task. Finally, the results for the few-shot task are reported in Tab. 5, whose examples are in Fig. 4. As expected, our SingularTrajectory significantly outperforms the existing models. With limited data, the existing works, except the diffusion-based model, tend to easily overfit. In contrast, our methods make it possible to explicitly designate future prototype paths, while taking full advantage of the expression ability of the diffusion model. In particular, even with only $10\\%$ of the data, our model achieves substantial improvements with respect to both data efficiency and performance. Figure 5 illustrates several cases where there are differences between the predictions of SingularTrajectory and other comparison methods. Previous works often show significant performance drops in other associated tasks compared to stochastic prediction tasks. In contrast, our model consistently predicts the best trajectories across multiple tasks.", + "bbox": [ + 496, + 503, + 895, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 771, + 663, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The number of motion vectors $K$ . First, we conduct a component study by varying the dimension $K$ of Singular space in Tab. 6. To find the best number of singular vectors in general, we carry out experiments across all five tasks using ZARA1 scene where there are both human-environmental and human-human interactions, following [93]. As the number of motion vectors $K$ increases, more detailed movements", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17896", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1b2d777706d42baa698fe7f459ebe68b89215dc214120d48937d2a40bc2fc2d.jpg", + "image_caption": [ + "Figure 5. Visualization of prediction consistency across five tasks. The more consistent the prediction is the better." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 468, + 228 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/26ac6bfe0b4624b886696339905f3fdf4df1ee68c4647510fcce868abd3db792.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KDeterministicStochasticMomentaryDomainFew-shotAverage
10.52 / 1.030.30 / 0.550.31 / 0.580.51 / 1.070.34 / 0.610.40 / 0.77
20.51 / 1.010.20 / 0.320.20 / 0.330.50 / 1.060.23 / 0.370.33 / 0.62
30.45 / 0.930.19 / 0.330.20 / 0.330.47 / 1.030.22 / 0.380.31 / 0.60
40.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
50.43 / 0.940.19 / 0.330.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
60.43 / 0.940.19 / 0.320.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
", + "bbox": [ + 78, + 273, + 465, + 356 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dc173bb1db23101557da66271c6af104dc8ee042a0cfefd84b9926310e34d688.jpg", + "table_caption": [ + "Table 6. Ablation study on the Singular space dimension $K$ ." + ], + "table_footnote": [], + "table_body": "
AdoptionDeterministicStochasticMomentaryDomainFew-shotAverage
Direct0.75 / 1.470.27 / 0.480.31 / 0.540.78 / 1.600.27 / 0.480.48 / 0.91
Initial0.47 / 0.950.22 / 0.400.22 / 0.390.50 / 1.170.23 / 0.400.33 / 0.66
Residual0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
Independent0.46 / 0.940.21 / 0.390.21 / 0.390.49 / 1.160.21 / 0.390.32 / 0.65
Jointly0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
", + "bbox": [ + 78, + 388, + 465, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are captured. In contrast, when using the smaller $K$ , it compresses the space, mainly to represent the overall motion flow. The performance tends to plateau when $K$ is larger than 3, as long as $K$ is not too small to cover most movements. We set $K = 4$ as the dimension for the Singular space because it shows the most effective prediction results across all tasks.", + "bbox": [ + 75, + 505, + 470, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trajectory denoising methods. Next, we evaluate three types of diffusion models for the trajectory prediction tasks, as shown in Tab. 7. First, we use a basic model similar to MID [18], which directly denoises from a Gaussian noise to a trajectory. This method fails to achieve good performances. The use of an anchor as an intermediate state, similar to LED [48], which reduces the denoising steps by skipping the initial denoising steps, seems to validate its generality. This demonstrates that our adaptive anchor can function as a good initializer for the diffusion model, particularly showing nearly identical outcomes in the stochastic, momentary observation, and few-shot tasks when predicting a multimodal path. However, we confirm that our scheme, which denoises only a residual by adding perturbation to a prototype path in Fig. 6, showcases the best performance in all the tasks. Additionally, compared to refining the prototype paths independently, regarding them as a batch dimension, our model can accurately predict the future when prototype paths are jointly refined.", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df7e34619e76f5d29e9b9a15a1ddd4bea2c99f89453fd42e28d4eed41c4ec12f.jpg", + "image_caption": [ + "Figure 6. Visualization of anchor refinement. The denoising process progressively refines the prototype paths at each diffusion step $m$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4cec812089f2c595af831e2b9e59e0cb10279017e8cab547a028f660e6ebe88b.jpg", + "table_caption": [ + "Table 7. Ablation study on the adoption of diffusion model." + ], + "table_footnote": [], + "table_body": "
MDeterministicStochasticMomentaryDomainFew-shotAverage
10.45 / 0.950.20 / 0.340.20 / 0.350.48 / 1.050.21 / 0.370.31 / 0.61
20.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
50.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
100.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
250.44 / 0.930.19 / 0.320.20 / 0.340.47 / 1.030.20 / 0.360.30 / 0.60
", + "bbox": [ + 501, + 242, + 890, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study on the diffusion steps $M$", + "bbox": [ + 545, + 315, + 841, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffusion steps $M$ . Lastly, we check how many steps in the diffusion model are needed for the cascaded refinement of the adaptive anchor. In Tab. 8, we confirm that the best performance comes from $M = 10$ . This is because the DDIM scheduler accelerates its convergence and the prototype path provides a rough initial trajectory, and so fewer denoising steps are sufficient. However, as the number of denoising steps increases, the information in the initial prototype path becomes attenuated due to the noise. As a result, we observe a slight decrease in performance.", + "bbox": [ + 496, + 357, + 890, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 522, + 617, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we introduce SingularTrajectory, a universal trajectory predictor model for all related trajectory prediction tasks. By unifying trajectory modalities into one Singular space, our approach standardizes trajectory data with shared motion dynamics, which eliminates the need for task-specific adjustments. The incorporation of an adaptive anchor system further personalizes the prototype paths, allowing them to interpret and adapt to environmental contexts and enhancing the reliability of the trajectory prediction. By successfully incorporating Singular space into the diffusion model, our SingularTrajectory framework successfully achieves the state-of-the-art results across five different benchmarks. This establishes SingularTrajectory itself as a general solution that covers a multitude of scenarios.", + "bbox": [ + 496, + 547, + 893, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), Vehicles AI Convergence Research & Development Program through the National IT Industry Promotion Agency of Korea (NIPA) funded by the Ministry of Science and ICT (No.S1602-20-1001), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST).", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17897", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 6", + "[2] Inhwan Bae and Hae-Gon Jeon. Disentangled multi-relational graph convolutional network for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2", + "[3] Inhwan Bae and Hae-Gon Jeon. A set of control points conditioned pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 2", + "[4] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Learning pedestrian group representations for multi-modal trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6", + "[5] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6", + "[6] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. EigenTrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7", + "[7] Apratim Bhattacharyya, Michael Hanselmann, Mario Fritz, Bernt Schiele, and Christoph-Nikolas Straehle. Conditional flow variational autoencoders for structured sequence prediction. arXiv preprint arXiv:1908.09008, 2020. 2", + "[8] Niccolò Bisagno, Bo Zhang, and Nicola Concì. Group lstm: Group trajectory prediction in crowded scenarios. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 2", + "[9] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[10] Guangyi Chen, Junlong Li, Nuoxing Zhou, Liangliang Ren, and Jiwen Lu. Personalized trajectory prediction via distribution discrimination. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[11] Guangyi Chen, Zhenhao Chen, Shunxing Fan, and Kun Zhang. Unsupervised sampling promoting for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[12] Sourav Das, Guglielmo Camporese, and Lamberto Ballan. Distilling knowledge for short-to-long term trajectory prediction. arXiv preprint arXiv:2305.08553, 2023. 2", + "[13] Patrick Dendorfer, Aljosa Osep, and Laura Leal-Taixe. Goalgan: Multimodal trajectory prediction based on goal position estimation. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Patrick Dendorfer, Sven Elflein, and Laura Leal-Taixe. M. Gagan: A multi-generator model preventing out-of-distribution samples in pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[15] Nachiket Deo and Mohan M. Trivedi. Trajectory forecasts in unknown environments conditioned on grid-based plans. arXiv preprint arXiv:2001.00735, 2020. 2", + "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Proceedings of the Neural Information Processing Systems (NeurIPS), 2021. 2", + "[17] Tharindu Fernando, Simon Denman, Sridha Sridharan, and Clinton Fookes. Soft+ hardwired attention: An lstm framework for human trajectory prediction and abnormal event detection. Neural Networks, 108:466-478, 2018. 2", + "[18] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6, 7, 8", + "[19] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 6", + "[20] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 1, 2", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2", + "[22] Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. Stgat: Modeling spatial-temporal interactions for human trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6", + "[23] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.", + "[24] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 2, 6", + "[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5", + "[26] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4", + "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proceedings of" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17898", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the International Conference on Learning Representations (ICLR), 2017. 2", + "[28] Vineet Kosaraju, Amir Sadeghian, Roberto Martín-Martín, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2", + "[29] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4", + "[30] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Muse-vae: Multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[31] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B. Choy, Philip H. S. Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[32] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Computer Graphics Forum, 26(3): 655-664, 2007. 5", + "[33] Jiachen Li, Hengbo Ma, and Masayoshi Tomizuka. Conditional generative neural system for probabilistic trajectory prediction. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2", + "[34] Jiachen Li, Fan Yang, Masayoshi Tomizuka, and Chiho Choi. Evolvegraph: Multi-agent trajectory prediction with dynamic relational reasoning. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2", + "[35] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[36] Yuke Li. Which way are you going? imitative decision learning for path forecasting in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[37] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[38] Junwei Liang, Lu Jiang, and Alexander Hauptmann. Simaug: Learning robust representations from simulation for trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[39] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[40] Rongqin Liang, Yuanman Li, Xia Li, Yi Tang, Jiantao Zhou, and Wenbin Zou. Temporal pyramid network for pedestrian" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "trajectory prediction with multi-supervision. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2", + "[41] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5", + "[43] Yuexin Ma, Xinge Zhu, Xinjing Cheng, Ruigang Yang, Jiming Liu, and Dinesh Manocha. Autotrajectory: Label-free trajectory extraction and prediction from videos using dynamic points. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[44] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[45] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6, 7", + "[46] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 5", + "[47] Huynh Manh and Gita Alaghband. Scene-lstm: A model for human trajectory prediction. arXiv preprint arXiv:1808.04018, 2018. 2", + "[48] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 8", + "[49] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), 2020. 2", + "[50] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Multiple trajectory prediction of moving agents with memory augmented networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2", + "[51] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Smemo: Social memory for trajectory forecasting. arXiv preprint arXiv:2203.12446, 2022. 2", + "[52] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 1, 2", + "[53] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgcnn: A social spatio-temporal" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17899", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 6, 7", + "[54] Abduallah Mohamed, Deyao Zhu, Warren Vu, Mohamed Elhoseiny, and Christian Claudel. Social-implicit: Rethinking trajectory prediction evaluation and the effectiveness of implicit maximum likelihood estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[55] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 7", + "[56] Ingrid Navarro and Jean Oh. Social-patternn: Socially-aware trajectory prediction guided by motion patterns. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 2", + "[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2", + "[58] Bo Pang, Tianyang Zhao, Xu Xie, and Ying Nian Wu. Trajectory prediction with latent belief energy-based model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6", + "[59] Wonhui Park, Dongkwon Jin, and Chang-Su Kim. Eigencontours: Novel contour descriptors based on low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5", + "[60] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2, 5", + "[61] Mark Pfeiffer, Giuseppe Paolo, Hannes Sommer, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. A data-driven model for interaction-aware pedestrian motion prediction in object cluttered environments. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 2", + "[62] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[63] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[64] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajec" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tory forecasting with heterogeneous data. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6", + "[65] Nasim Shafiee, Taskin Padir, and Ehsan Elhamifar. Introvert: Human trajectory prediction via conditional 3d attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[66] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6", + "[67] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 2", + "[68] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[69] Xiaodan Shi, Xiaowei Shao, Zipei Fan, Renhe Jiang, Haoran Zhang, Zhiling Guo, Guangming Wu, Wei Yuan, and Ryosuke Shibasaki. Multimodal interaction-aware trajectory prediction in crowded space. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2", + "[70] Xiaodan Shi, Xiaowei Shao, Guangming Wu, Haoran Zhang, Zhiling Guo, Renhe Jiang, and Ryosuke Shibasaki. Socialdpf: Socially acceptable distribution prediction of futures. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2", + "[71] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2", + "[72] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[73] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2", + "[74] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[75] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[76] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[77] Jianhua Sun, Yuxuan Li, Liang Chai, Hao-Shu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17900", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "momentary observation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7", + "[78] Jianhua Sun, Yuxuan Li, Liang Chai, and Cewu Lu. Stimulus verification is a universal and effective sampler in multi-modal human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[79] Chaofan Tao, Qinhong Jiang, and Lixin Duan. Dynamic and static context-aware LSTM for multi-agent motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[80] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1991. 4", + "[81] Daksh Varshneya and G. Srinivasaraghavan. Human trajectory prediction using spatially aware deep attention models. arXiv preprint arXiv:1705.09436, 2017. 2", + "[82] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2", + "[83] Anirudh Vemula, Katharina Muelling, and Jean Oh. Social attention: Modeling attention in human crowds. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2", + "[84] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters (RA-L), 2022. 2", + "[85] Yuning Wang, Pu Zhang, Lei Bai, and Jianru Xue. Fend: A future enhanced distribution-aware contrastive learning framework for long-tail trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[86] Song Wen, Hao Wang, and Dimitris Metaxas. Social ode: Multi-agent trajectory forecasting with neural ordinary differential equations. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[87] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[88] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6", + "[89] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[90] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[91] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6", + "[92] Yi Xu, Jing Yang, and Shaoyi Du. Cf-lstm: Cascaded feature-based long short-term networks for predicting pedestrian trajectory. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2", + "[93] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7", + "[94] Hao Xue, Du Q Huynh, and Mark Reynolds. Ss-lstm: A hierarchical LSTM model for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2018. 2", + "[95] Kota Yamaguchi, Alexander C Berg, Luis E Ortiz, and Tamara L Berg. Who are you with and where are you going? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2", + "[96] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters (RA-L), 2021. 2", + "[97] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6", + "[98] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6, 7", + "[99] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[100] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6", + "[101] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[102] Tianyang Zhao, Yifei Xu, Mathew Monfort, Wongun Choi, Chris Baker, Yibiao Zhao, Yizhou Wang, and Ying Nian Wu. Multi-agent tensor fusion for contextual trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[103] ChunYu Zhi, HuaiJiang Sun, and Tian Xu. Adaptive trajectory prediction without catastrophic forgetting. The Journal of Supercomputing, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 883 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17901", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_model.json b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cbc878320dba225b00666ba09f8560c1d09a3be1 --- /dev/null +++ b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_model.json @@ -0,0 +1,2952 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.861, + 0.153 + ], + "angle": 0, + "content": "SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.182, + 0.682, + 0.216 + ], + "angle": 0, + "content": "Inhwan Bae, Young-Jae Park and Hae-Gon Jeon* \nAI Graduate School, GIST, South Korea" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.22, + 0.746, + 0.234 + ], + "angle": 0, + "content": "{inhwanbae, youngjae.park}@gm.gist.ac.kr, haegonj@gist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.755 + ], + "angle": 0, + "content": "There are five types of trajectory prediction tasks: deterministic, stochastic, domain adaptation, momentary observation, and few-shot. These associated tasks are defined by various factors, such as the length of input paths, data split and pre-processing methods. Interestingly, even though they commonly take sequential coordinates of observations as input and infer future paths in the same coordinates as output, designing specialized architectures for each task is still necessary. For the other task, generality issues can lead to sub-optimal performances. In this paper, we propose SingularTrajectory, a diffusion-based universal trajectory prediction framework to reduce the performance gap across the five tasks. The core of SingularTrajectory is to unify a variety of human dynamics representations on the associated tasks. To do this, we first build a Singular space to project all types of motion patterns from each task into one embedding space. We next propose an adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods that sometimes yield unacceptable paths, our adaptive anchor enables correct anchors, which are put into a wrong location, based on a traversability map. Finally, we adopt a diffusion-based predictor to further enhance the prototype paths using a cascaded denoising process. Our unified framework ensures the generality across various benchmark settings such as input modality, and trajectory lengths. Extensive experiments on five public benchmarks demonstrate that SingularTrajectory substantially outperforms existing models, highlighting its effectiveness in estimating general dynamics of human movements. Code is publicly available at https://github.com/inhwanbae/SingularTrajectory." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Extensive studies of trajectory prediction methods have been conducted in the computer vision field for several decades [20, 52]. They have demonstrated its importance in various applications, including crowd simulation, social robot navi" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.268, + 0.891, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.383, + 0.895, + 0.467 + ], + "angle": 0, + "content": "Figure 1. An overview of our SingularTrajectory framework. All relevant human trajectory prediction tasks can be represented in our Singular space, a unified feature embedding space for human dynamics. Using the embedding features, our diffusion-based model for a universal trajectory prediction makes prediction for all the tasks in this same space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.895, + 0.597 + ], + "angle": 0, + "content": "gation, obstacle avoidance and security and surveillance systems, etc. Trajectory prediction takes sequential coordinate values of agents as input and infers their future pathways in common [1, 19, 83]. Such tasks vary depending on the application. The tasks are determined with respect to the number of input/output, data processing, and the use of geological information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.896, + 0.81 + ], + "angle": 0, + "content": "The tasks for trajectory prediction can be categorized into five groups: (1) Stochastic prediction is to predict 20 multi-modal future trajectories from one observation with 8 frames [19]. Each future trajectory consists of 12 frames. (2) Deterministic prediction takes one observation with 8 frames, but infers only one future trajectory with 12 frames [1]. (3) The momentary observation uses only two frames to predict 20 multi-modal future paths with 12 frames [77]. (4) Domain adaptation splits training data with respect to places in datasets, trains a model on one place, and then checks the transferability to other places [93]. (5) The few-shot task only uses partial data to build a dataset-efficient model [53], as illustrated in Fig. 1. Until now, each specialized architecture for a task type has provided performance gains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.884 + ], + "angle": 0, + "content": "However, two questions arise. First, why do state-of-the-art models for one task undergo significant performance drops when applied to other trajectory prediction tasks? Second, is it feasible to design a general predictor that works across the five tasks?" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As answers to these questions, we present a universal tra" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17890" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "jectionary predictor to achieve generality in predictions, named SingularTrajectory. The main idea is to unify the modalities of human dynamic representations across the five tasks. To do this, we first introduce a Singular space, an embedding space consisting of representative motion patterns for each task. The motion patterns play a role in the basis function for pedestrian movements, and are extracted using Singular Value Decomposition (SVD). They are then projected onto Singular space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.439 + ], + "angle": 0, + "content": "We next propose an environment-adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods [6, 29] that sometimes fail to handle different target data distribution, our adaptive anchor is able to correct prototype paths from the adaptive anchor in Singular space if they are put into the wrong locations, based on an input traversability map. Lastly, we generate socially-acceptable future trajectories for all agents in scenes with a diffusion-based predictor which denoises residuals of perturbed prototype paths. Thanks to the cascaded denoising process of diffusion models, we refine the prototype paths. Here, historical pathways, agent interactions, and environmental information are provided as conditions to guide them in the Markov chain of the denoising diffusion processes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.439, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Experimental results demonstrate that our SingularTrajectory can successfully represent pedestrian motion dynamics, and significantly improve prediction accuracy for the five tasks on challenging public benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.512, + 0.228, + 0.527 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.47, + 0.567 + ], + "angle": 0, + "content": "We review previous studies on trajectory prediction that have attempted to address various benchmark scenarios." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.576, + 0.365, + 0.592 + ], + "angle": 0, + "content": "2.1. Pedestrian Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Ways of predicting pedestrian future trajectories have been studied for a long time in the computer vision field. Pioneering works [20, 52, 60, 95] model an invisible social norm based on motion dynamics as an energy minimization problem. Introducing recurrent neural networks [1, 19, 64, 97] achieves significant improvements by providing highly sequential representations of high-level shape of paths. These methods determine the most probable path, called deterministic trajectory prediction. The following works model mutual influences among agents using attention mechanisms [17, 23, 64, 83], graph convolutional networks [2, 9, 27, 41, 53, 54, 75], graph attention networks [5, 22, 28, 38, 39, 66, 82], and transformers [4, 18, 55, 68, 86, 87, 97, 98]. Additional visual information allows us to leverage environmental constraints from traversability maps [13-15, 28, 37, 46, 47, 49, 50, 63, 65, 74, 75, 79, 81, 85, 94, 99, 102]. Depending on the constraints, predictors take either recurrent [1, 8, 10, 11, 18, 19, 30, 43, 44, 51, 56, 61, 64, 92, 100, 101] or simultaneous approaches [2, 3, 35, 53, 66, 67, 89] to extrapolate the future pathways." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.896, + 0.426 + ], + "angle": 0, + "content": "Meanwhile, with the success of generative models, the importance of multimodality has begun to emerge, called stochastic trajectory prediction. Stochastic prediction enables us to consider all of an agent's possible future pathways. For example, an agent at a crossroads may either walk straight or turn left/right. Here, the stochastic prediction infers all potential future modes. This approach has become mainstream in this field. Starting from Social-GAN [19], a bivariate Gaussian distribution [1, 2, 34, 53, 66, 69, 70, 93, 96, 97], generative adversarial network [14, 19, 22, 28, 36, 40, 63, 74, 78, 102], and conditional variational autoEncoder [7, 10, 23, 30, 31, 33, 45, 64, 76, 84, 88, 91] have been adopted for stochastic trajectory prediction. Anchorm-conditioned methods can explicitly represent different modalities by prototyping possible paths [6, 29]. Most recently, diffusion-based models have revealed their tremendous representation capacities in numerous tasks [16, 21, 57, 71-73], proving its potential for stochastic trajectory prediction [18, 25, 48, 62]. In this study, we take full advantage of both the anchor-conditioned approach and the diffusion-based model to achieve the explainability and generalizability of the trajectory prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.432, + 0.816, + 0.449 + ], + "angle": 0, + "content": "2.2. Various Trajectory Prediction Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.652 + ], + "angle": 0, + "content": "Beyond the standard benchmark protocol of stochastic prediction, there are three other variants of this task: momentary observation, domain adaptation, and few-shot learning. Works in [12, 55, 77] only take two frames as input for the momentary trajectory prediction. Multi-task learning, self-supervised learning, and knowledge distillation techniques have been used to extract rich features from the limited data. Another works [24, 93, 103] focuses on domain adaptation across trajectory domains, captured from different surveillance views. A transferable graph neural network is introduced to adaptively learn domain-invariant knowledge. The others [53, 101] adopt few-shot learning for better training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.895, + 0.773 + ], + "angle": 0, + "content": "Although these works take and infer the sequential coordinates of agent trajectories in common, there is no unified model for all the tasks. Despite the tremendous efforts to design a specialized architecture for one task, they cannot be applied to the other tasks without suffering significant performance drops. In the next section, we will describe how to design a unified architecture, which consistently produces promising results on the five associated tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.637, + 0.803 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We describe how to learn a general representation of human motions. We first define a general trajectory prediction problem in Sec. 3.1 and provide preliminaries of explicit formulations on the SVD and diffusion process in Sec. 3.2. We then introduce a motion feature extraction to build our Singular space, and a projection of any trajectory from each task onto" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17891" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "it in Sec. 3.3. Next, we propose an environment-adaptive anchor using motion vectors and input image in Sec. 3.4. Finally, we present the SingularTrajectory predictor based on the diffusion model in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.16, + 0.262, + 0.175 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Trajectory prediction aims to predict the future paths of agents based on their historical path and surrounding contexts. Suppose that at each timestamp \\( t \\), there are \\( N \\) pedestrians in a scene with the 2D spatial coordinate position \\( \\{\\pmb{p}_t^n \\in \\mathbb{R}^2 | n \\in [1, \\dots, N]\\} \\). A pedestrian historical trajectory \\( \\pmb{X}_n \\) over \\( T_{hist} \\) timesteps can be represented as the cumulative coordinates \\( \\pmb{X}_n = \\{\\pmb{p}_t^n | t \\in [1, \\dots, T_{hist}]\\} \\). Similarly, future trajectories \\( \\pmb{Y}_n \\) for the time duration \\( T_{fut} \\) to be predicted can be written as \\( Y_n = \\{\\pmb{p}_t^n | t \\in [T_{hist} + 1, \\dots, T_{hist} + T_{fut}]\\} \\). The prediction system takes both the historical trajectories for all \\( N \\) people \\( X = \\{X_n | n \\in [1, \\dots, N]\\} \\) and the scene image map \\( I \\) for environmental information as input. The deterministic prediction system infers one sequence of the most reliable future trajectory \\( \\widehat{Y} = \\{\\widehat{Y}_n | n \\in [1, \\dots, N]\\} \\). For the stochastic prediction, because of the indeterminacy of the future movements, \\( S \\) multiple pathways for all the \\( N \\) pedestrians \\( \\widehat{Y} = \\{\\widehat{Y}_n^s | n \\in [1, \\dots, N], s \\in [1, \\dots, S]\\} \\) are generated so that at least one sample is close to the ground-truth trajectory." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.464, + 0.218, + 0.479 + ], + "angle": 0, + "content": "3.2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.533 + ], + "angle": 0, + "content": "Singular Value Decomposition. Singular Value Decomposition (SVD) decomposes a matrix into three resultant matrices. Given a matrix \\( \\mathbf{A} \\), its SVD is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.54, + 0.47, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} = \\boldsymbol {U} \\boldsymbol {\\Sigma} \\boldsymbol {V} ^ {\\top}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.657 + ], + "angle": 0, + "content": "where \\(\\pmb{U}\\) is an orthogonal left singular vector matrix, whose columns are eigenvectors of \\(A\\times A^{\\top}\\). \\(\\pmb{\\Sigma}\\) is a diagonal matrix with the singular values of \\(A\\), consisting of \\(K\\) non-negative values in descending order. \\(\\pmb{V}\\) is a right singular vector matrix, which is also orthogonal and its columns are the eigenvectors of \\(A^{\\top}\\times A\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.472, + 0.869 + ], + "angle": 0, + "content": "To remove the redundant part of the raw data, the truncation technique is often applied to the results after the decomposition. The idea behind the truncated SVD is to approximate the original matrix \\( \\mathbf{A} \\) with the lower rank. With the \\( K \\) to determine the number of singular values to retain, \\( \\pmb{\\Sigma} \\) can be simplified to \\( \\pmb{\\Sigma}_{K} \\) which contains only the \\( K \\) largest singular values. Similarly, \\( \\pmb{U} \\) and \\( V^{\\top} \\) are reduced to \\( U_{K} \\) and \\( V_{K}^{\\top} \\) by keeping the first \\( K \\) columns and rows, respectively. This process eliminates the smallest singular values, which are not needed to express the original data and often correspond to noise or redundant information. This is useful for practical scenarios dealing with large and potentially sparse matrices because we can reconstruct a close approximation of the original data with significantly less storage space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Diffusion models. The diffusion model operates by transforming a noisy distribution, represented by the noise vector" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "\\(\\mathbf{y}_M\\), into the desired data \\(\\mathbf{y}_0\\) through a series of \\(M\\) diffusion steps. These steps involve intermediate latent variables \\(\\{\\mathbf{y}_m|m\\in [1,\\dots ,M]\\}\\), and encompass both the diffusion and denoising processes. The diffusion process adds a small amount of noise to data in order to obtain the standard normal distribution \\(q(\\mathbf{y}_M)\\) from the distribution \\(q(\\mathbf{y}_0)\\) using the Markov chain as:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {y} _ {1: M} \\mid \\mathbf {y} _ {0}\\right) := \\prod_ {m = 1} ^ {M} q \\left(\\mathbf {y} _ {m} \\mid \\mathbf {y} _ {m - 1}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.247, + 0.854, + 0.267 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {y} _ {m} \\mid \\mathbf {y} _ {m - 1}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {m}; \\sqrt {1 - \\beta_ {m}} \\mathbf {y} _ {m - 1}, \\beta_ {m} \\mathbf {I}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.32 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) is a small positive constant and a variance schedule for adding noise. The denoising process uses the \\(\\mathbf{y}_m\\) to recover \\(\\mathbf{y}_0\\) with a learnable network as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.33, + 0.892, + 0.371 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {y} _ {0: M}\\right) := p \\left(\\mathbf {y} _ {M}\\right) \\prod_ {m = 1} ^ {M} p _ {\\theta} \\left(\\mathbf {y} _ {m - 1} \\mid \\mathbf {y} _ {m}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.372, + 0.852, + 0.39 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\mathbf {y} _ {m - 1} | \\mathbf {y} _ {m}) := \\mathcal {N} (\\mathbf {y} _ {m - 1}; \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {y} _ {m}, m), \\beta_ {m} \\mathbf {I}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.536 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}_M\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) is an initial noise sampled from the Gaussian distribution \\(p(\\mathbf{y}_M)\\), and \\(\\theta\\) denotes the learnable parameter of the diffusion model. \\(\\epsilon_{\\theta}\\) is a learnable denoising model of a clean data \\(\\mathbf{y}_0\\), and a corrupted data \\(\\mathbf{y}_m\\) at a step \\(m\\). The objective is to train the neural network so that the denoising process predicts the true data-generating distribution well. This is often done by maximizing the evidence lower bound, ensuring that the samples generated by the diffusion model are indistinguishable from the real data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.742, + 0.56 + ], + "angle": 0, + "content": "3.3. Unifying the Motion Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.703 + ], + "angle": 0, + "content": "The trajectory prediction model uses a learnable network to capture the relationship in consideration of the input coordinates of pedestrians, input images and output coordinates for each pedestrian. Since expectations for input and output trajectories are different for each associated task (e.g., length and multimodality), they should be viewed as different data spaces even if they use the same coordinate systems. We introduce a method to merge raw data in each space into a Singular space for human motion dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Singular space construction. To discover motion dynamics from the raw data, we first extract primitive motion features. Inspired by the successful low-rank approximation of raw trajectory data using eigenvectors in [6], we also employ a similar strategy using singular vectors from the truncated SVD to extract principal motion components from the entire training dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "First, we cut-off paths of all pedestrians in the dataset into \\( T_{win} \\) lengths through a sliding window to create a total of \\( L \\) gist of trajectory set \\( A \\in \\mathbb{R}^{L \\times (2 \\times T_{win})} \\). Here, \\( A \\) is a temporary matrix to extract a motion vector from a set of trajectory in the initialization phase. Next, we decompose \\( A \\) to obtain truncated matrices \\( U_K \\in \\mathbb{R}^{L \\times K} \\), \\( \\pmb{\\Sigma}_K \\in \\mathbb{R}^{K \\times K} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17892" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.186, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.092, + 0.278, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.092, + 0.373, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.092, + 0.469, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.175, + 0.47, + 0.272 + ], + "angle": 0, + "content": "Figure 2. Visualization of the trajectories in Singular space. (a) The circle and triangle markers indicate the history and future trajectory coordinates, respectively. Each color also refers to each associated task. (b-d) Each marker in the Singular space corresponds to each trajectory in raw data. And, the slicing planes, representing a set of arrows, mean human dynamics of straight forward motions and turning motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.473, + 0.45 + ], + "angle": 0, + "content": "and \\(V_{K}^{\\top}\\in \\mathbb{R}^{K\\times (2\\times T_{win})}\\). Here, \\(V_{K}^{\\top}\\) is a set of \\(K\\) spatiotemporal motion vectors \\(\\mathrm{v}_k\\in \\mathbb{R}^{2\\times T_{win}}\\) representing the most dominant motion dynamics of pedestrians. Since \\(\\mathrm{v}_k\\) are orthogonal to each other, we define a Singular space coordinate system using the \\(K\\) singular vectors as basis vectors for each axis. In Singular space, trajectories \\(\\pmb{A}\\) can be a coordinate of \\(\\mathcal{A} = U_K\\cdot \\pmb {\\Sigma}_K\\in \\mathbb{R}^{L\\times K}\\), indicating that each motion vector has an influence on reconstructing the \\(L\\) trajectories. In other words, we can project the \\(\\pmb{A}\\) into the coordinate in Singular space \\(\\mathcal{A}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.458, + 0.47, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} = \\boldsymbol {A} \\times \\left(\\boldsymbol {V} _ {K} ^ {\\top}\\right) ^ {- 1} = \\boldsymbol {A} \\times \\boldsymbol {V} _ {K}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.469, + 0.517 + ], + "angle": 0, + "content": "where \\((\\mathbf{V}_K^\\top)^{-1}\\) can be simplified into \\(\\mathbf{V}_K\\) with the property of an orthogonal matrix1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.713 + ], + "angle": 0, + "content": "In the Singular space, we can now concentrate on the motion flow by using the coefficients of motion patterns, instead of considering every position over time. Note that the results from the SVD depend highly on various factors such as the window size for inputs/outputs, data processing, and the differences in the raw data caused by geological variations. Due to the issue, works using SVD only convert the output space [26, 59, 80] or conduct separate decompositions for each of input and output data spaces [6]. However, in this case, they cause inconsistent representations of motion dynamics, even for the same pedestrian, which leads to a lack of generality. In the next step, we introduce a new method to address this issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.471, + 0.867 + ], + "angle": 0, + "content": "Projection of any trajectory into Singular space. We aim to express the input/output trajectories \\( \\mathbf{X} \\) and \\( \\mathbf{Y} \\) of the five associated tasks in Singular space all at once. Our core idea stems from the notion that all pedestrians share the same human motion dynamics and thus will likely show a similar pattern. Starting with the projection function Eq. (4), which is a projection matrix for the fixed trajectory length \\( T_{win} \\), we extend it to any length \\( T_{hist} \\) and \\( T_{fut} \\), which varies from task to task. To handle the motion patterns regardless of their lengths, we interpolate \\( \\mathrm{v}_k \\in \\mathbb{R}^{2 \\times T_{win}} \\) to \\( \\mathrm{v}_{\\mathrm{x},k} \\in \\mathbb{R}^{2 \\times T_{hist}} \\)." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.63, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.169, + 0.625, + 0.179 + ], + "angle": 0, + "content": "(a) Initial prototype paths" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.089, + 0.761, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.169, + 0.73, + 0.178 + ], + "angle": 0, + "content": "(b) Vector field" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.089, + 0.893, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.169, + 0.884, + 0.179 + ], + "angle": 0, + "content": "(c) Final prototype paths" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.182, + 0.894, + 0.28 + ], + "angle": 0, + "content": "Figure 3. An example of the adaptive anchor generation. (a) The initial prototype anchor \\(\\mathcal{P}\\) is placed on the last observation coordinate of a person. In this instance, four prototype paths (highlighted in red) are incorrectly placed at the non-traversable locations. (b) Vector field \\(\\vec{F}_I(x,y)\\) is computed to guide toward in the nearest traversable areas. (c) The initial prototype paths are then tailored to the environment using the vector field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.894, + 0.412 + ], + "angle": 0, + "content": "Since the motion pattern can be seen as a 2-dimensional curve, we use Cardinal B-splines to make a transformation matrix \\(C_{T_{hist}} \\in \\mathbb{R}^{(2 \\times T_{hist}) \\times (2 \\times T_{win})}\\) using the Irwin-Hall distribution. The constant value \\(C\\) depends only on the length, so \\(\\mathrm{v}_{\\mathrm{x},k}\\) can be approximated. The trajectory \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times (2 \\times T_{hist})}\\) is projected to the coordinate \\(\\mathcal{X} \\in \\mathbb{R}^{N \\times (2 \\times T_{hist})}\\) in Singular space as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.425, + 0.892, + 0.441 + ], + "angle": 0, + "content": "\\[\nv _ {x, k} = C _ {T _ {\\text {h i s t}}} \\times v _ {k}. \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.45, + 0.892, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} = \\boldsymbol {X} \\times \\boldsymbol {C} _ {T _ {\\text {h i s t}}} \\times \\boldsymbol {V} _ {K}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.504 + ], + "angle": 0, + "content": "In the same way as Eq. (6), the trajectory \\(\\mathbf{Y} \\in \\mathbb{R}^{N \\times (2 \\times T_{\\text{fit}})}\\) also can be projected to \\(\\mathcal{V} \\in \\mathbb{R}^{N \\times K}\\) using \\(\\mathbf{V}_K\\) and \\(C_{T_{\\text{fit}}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.895, + 0.669 + ], + "angle": 0, + "content": "Through this process, Singular space can represent a variety of trajectories even with different lengths, including motion vectors for the input/output as well as the task-specific data, as visualized in Fig. 2. Establishing a common ground for the trajectory representation is crucial for our model's adaptability and robust performances across different trajectory prediction benchmarks. Moreover, the strategy of focusing on the overall motion flow, rather than frame-by-frame coordinates, further improves the model's capacity to understand and predict socially compliant trajectories in diverse real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.669, + 0.696 + ], + "angle": 0, + "content": "3.4. Adaptive Anchor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Beyond the trajectory data integration, we also introduce how to incorporate environmental contexts into trajectory predictors. To do this, we propose an adaptive anchor which is expressed as a set of motion vectors from input images and consists of prototype paths in Singular space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Prototype anchor formation. Previous anchor-based human trajectory prediction approaches [6, 29] use a fixed anchor for any pedestrian, and the anchor is refined for output trajectories. Although these methods can explicitly model the multimodality, they sometimes fail to handle the case where the prototype path is put into wrong locations, or blocked by static obstacles. This is because the wrong prototypes are treated as a hard constraint. To avoid this problem," + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1For better understanding, we display the coordinate variable in the Singular space using the calligraphic font" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17893" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "we use the input image as a traversability map to correct the wrong prototype paths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.257 + ], + "angle": 0, + "content": "We start with the converted future trajectory \\(\\mathcal{V}\\) in the whole dataset \\(Y\\) using Sec. 3.3. Following [6, 59], the coordinates \\(\\mathcal{V}\\) are then clustered into \\(S\\) centroids, which can be viewed as groups of components representing different multimodal futures. In addition, each centroid in Singular space is used as an initial prototype path \\(\\mathcal{P}_s\\) for constructing an anchor \\(\\mathcal{P} \\in \\mathbb{R}^{S \\times K}\\). But, these initial prototype paths are still fixed, so they still have a limitation which is unlikely to consider environmental information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.471, + 0.414 + ], + "angle": 0, + "content": "Adaptive anchor generation. To make use of environmental information, we introduce a module to deform the anchor. Using an off-the-shelf semantic segmentation model from [46], we can obtain binary traversable maps \\( \\mathbf{I}_{\\text{map}} \\) from an input image. We then derive a vector field \\( \\vec{F}_I(x,y) \\) to fix the wrongly located prototype path by directing it to a nearest traversable regions. If the initial prototype path \\( \\mathcal{P}_s \\) is in the wrong place, we deform it by adding the vector fields into the initial anchor until they reach equilibrium states to obtain the final prototype paths \\( \\mathcal{P}_s' \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.426, + 0.47, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {s} ^ {\\prime} = \\mathcal {P} _ {s} + \\vec {F} _ {I} \\left(\\mathcal {P} _ {s} \\boldsymbol {V} _ {K} ^ {\\top} \\boldsymbol {C} _ {T _ {f u t}} ^ {- 1}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.517 + ], + "angle": 0, + "content": "Through this process, the prototype path is re-located toward the nearest walkway, as demonstrated in Fig. 3. In other words, the anchor plays a role in the environment-adaptive prototype paths, unlike the existing fixed anchor methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Because the prototype paths are in Singular space, the scene image can be implicitly represented as the adaptive anchor. In addition, the scene image can be approximated with the set of motions. By projecting the scene images into the Singular space, our model understands the surrounding environment. This holistic approach, integrating both coordinate and environmental cues, sets the stage for more realistic and reliable trajectory predictions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.442, + 0.666 + ], + "angle": 0, + "content": "3.5. Diffusion-Based SingularTrajectory Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.748 + ], + "angle": 0, + "content": "As the final step, we develop a framework called Singular-Trajectory, a unified model that works well across the five tasks. Leveraging the Singular space and the adaptive anchor, our diffusion-based SingularTrajectory model can precisely forecast potential future paths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Denoising a perturbed trajectory anchor. Unlike previous diffusion-based predictors, which directly forecast the future path from Gaussian noise [18, 25], we devise a stepwise refinement from the adaptive anchor for a realistic trajectory. Here, historical pathways \\(\\mathcal{X}\\), environmental information \\(\\mathcal{P}\\), and agent interactions \\(\\mathcal{G}(\\mathcal{X},\\mathcal{P})\\) are encoded and used as conditions to guide the denoising processes. To better capture agents interactions \\(\\mathcal{G}\\) in the denoising process, we adopt the transformer model. Similar to other transformer-based trajectory predictors [18, 48], our SingularTrajectory" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "encodes spatio-temporal information to account for agent-agent and agent-environment interactions using \\(\\mathcal{X}\\) and \\(\\mathcal{P}\\), respectively. These conditions are then concatenated into one feature vector representation and fed into the diffusion model \\(\\epsilon_{\\theta}(\\mathbf{y}_m, m, \\mathcal{X}, \\mathcal{P}, \\mathcal{G})\\) to learn, by contrasting the motion patterns from previous diffusion steps to distinguish the added noise, and to close the reality gap by generating socially-acceptable future paths \\(\\widehat{\\mathbf{Y}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.214, + 0.892, + 0.245 + ], + "angle": 0, + "content": "This refinement process works in a cascading manner by \\(\\{\\mathcal{P},\\dots,\\widehat{\\mathcal{V}}\\}_{m = 1}^{M}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {Y}} = \\mathcal {P} + p \\left(\\mathbf {y} _ {M}\\right) \\prod_ {m = 1} ^ {M} p _ {\\theta} \\left(\\mathbf {y} _ {m - 1} \\mid \\mathbf {y} _ {m}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.38 + ], + "angle": 0, + "content": "By predicting only the residuals \\(\\mathbf{y}_m\\in \\mathbb{R}^{S\\times K}\\) to adjust the anchors, the problem is simplified, in that the model is able to use a prior knowledge on the initial state. With this process, we thus ensure more precise and reliable trajectory generations, achieving generality in various environments and applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Implementation details. To construct Singular space, we empirically set \\( K \\) to 4. We set \\( T_{win} = T_{fut} = 12 \\) in order to prevent information loss due to the approximation of the motion vector during prediction. For an anchor diffusion model, we devise a one-layer transformer for encoding motion and context information, where the dimension is set to 256 with 4-head attention. We set \\( M = 10 \\) and schedule the diffusion timesteps following DDIM [72]. To train the SingularTrajectory in an end-to-end manner, we use a mean square error (MSE) as a loss function between the output and a random Gaussian noise for the current iteration. The training is performed with AdamW optimizer [42], with a learning rate of 0.001 and batch size of 512 for 256 epochs. All the experiments are conducted on a single NVIDIA A6000 GPU, which usually takes about an hour to train each scene." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.634, + 0.643 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.894, + 0.773 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive experiments to verify the generality of our SingularTrajectory model for the trajectory prediction tasks. We first describe our experimental setup in Sec. 4.1. We then provide comparison results with state-of-the-art models on public benchmark datasets in Sec. 4.2. Finally, we perform an extensive ablation study demonstrating the effect of each component of our method in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.691, + 0.803 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Datasets. To compare our SingularTrajectory with state-of-the-art baselines, we conduct quantitative evaluations on two common datasets, ETH [60] and UCY [32], for all the associated tasks. The ETH and UCY datasets consist of various motions of 1,536 pedestrians across five unique scenes: ETH, Hotel, Univ, Zara1 and Zara2. They are recorded in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17894" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.89, + 0.344 + ], + "angle": 0, + "content": "
Domain Adaptation (ADE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAVG
STGCNN [53]1.831.581.301.311.513.021.382.631.582.151.160.700.820.540.811.041.050.730.470.820.981.090.740.500.831.22
PECNet [45]1.971.681.241.351.563.111.352.691.622.191.390.820.930.570.931.101.170.920.520.931.011.250.830.610.931.31
AgentFormer [98]1.731.641.871.571.701.491.762.131.731.781.730.192.011.791.431.712.051.921.321.751.961.911.891.911.921.72
MID [18]1.261.431.651.431.441.230.870.980.770.961.610.630.810.580.911.420.750.810.740.931.470.680.880.910.981.04
EigenTrajectory [6]0.390.710.730.450.570.950.680.510.420.640.940.320.490.370.531.101.031.360.430.980.920.510.620.520.640.67
T-GNN [93]1.131.250.941.031.092.541.082.251.411.820.970.540.610.230.590.880.780.590.320.640.870.720.650.340.650.96
K0 [24]0.450.780.680.590.630.870.650.490.360.591.080.460.440.360.591.130.560.590.530.701.130.390.590.480.650.63
Adaptive [24]0.360.640.530.450.500.900.620.500.360.601.110.460.480.420.621.110.520.540.470.661.050.340.560.440.600.59
SingularF trajectory0.290.590.510.420.450.660.550.450.350.500.790.300.480.360.480.730.290.550.360.490.640.230.550.450.470.48
Domain Adaptation (FDE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAvgV
STGCNN [53]3.242.862.532.432.775.162.514.862.883.852.301.341.741.101.622.211.991.410.881.622.102.051.471.011.662.30
PECNet [45]3.332.832.532.452.795.232.484.902.863.872.221.321.681.121.592.202.051.520.881.662.101.841.450.981.592.29
AgentFormer [98]3.603.544.523.953.903.763.675.173.424.013.773.494.123.763.793.883.754.273.613.883.823.553.744.323.863.89
MID [18]2.422.773.222.772.802.341.661.871.441.833.131.281.741.241.852.731.491.671.541.862.871.411.881.932.022.07
EigenTrajectory [6]0.761.431.560.961.182.101.381.070.861.352.010.641.100.821.142.421.031.360.971.442.040.511.291.111.241.27
T-GNN [93]2.182.251.781.842.014.151.824.042.533.141.911.121.300.871.301.921.461.250.651.321.861.451.280.721.331.82
K0 [24]0.941.721.511.291.371.811.331.020.751.232.170.910.940.811.212.180.981.201.031.352.120.681.221.031.261.28
Adaptive [24]0.691.341.120.911.021.851.271.050.761.232.260.921.030.941.292.231.021.171.001.362.070.631.190.941.211.22
SingularF trajectory0.571.191.080.810.911.161.170.970.731.011.360.621.110.790.971.220.521.180.730.921.070.431.150.940.900.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.345, + 0.893, + 0.375 + ], + "angle": 0, + "content": "Table 1. Comparison of our SingularTrajectory with other state-of-the-art methods in domain adaptation task. Following [24, 93], we deterministically predict one future path for all dataset pairs (ADE/FDE). Bold: Best, Underline: Second best." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.389, + 0.466, + 0.572 + ], + "angle": 0, + "content": "
StochasticETHHOTELUNIVZARA1ZARA2AVG
Social-GAN [19]0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
STGCNN [53]0.64/1.110.49/0.850.44/0.790.34/0.530.30/0.480.44/0.75
STAR [97]0.57/1.110.19/0.370.35/0.750.26/0.570.25/0.580.33/0.68
PECNet [45]0.61/1.070.22/0.390.34/0.560.25/0.450.19/0.330.32/0.56
MID [18]0.57/0.930.21/0.330.29/0.550.28/0.500.20/0.370.31/0.54
LBEBM [58]0.60/1.060.21/0.380.28/0.540.21/0.390.15/0.300.29/0.53
Trajectory++[64]0.61/1.030.20/0.280.30/0.550.24/0.410.18/0.320.31/0.52
GroupNet [88]0.46/0.730.15/0.250.26/0.490.21/0.390.17/0.330.25/0.44
AgentFormer [98]0.46/0.800.14/0.220.25/0.450.18/0.300.14/0.240.23/0.40
GP-Graph [4]0.43/0.630.18/0.300.24/0.420.17/0.310.15/0.290.23/0.39
NPSN [5]0.36/0.590.16/0.250.23/0.390.18/0.320.14/0.250.21/0.36
EqMotion [90]0.40/0.610.12/0.180.23/0.430.18/0.320.13/0.230.21/0.35
EigenIProjection [6]0.36/0.530.12/0.190.24/0.430.19/0.330.14/0.240.21/0.34
SocialVAE [91]0.41/0.580.13/0.190.21/0.360.17/0.290.13/0.220.21/0.33
LED [48]0.39/0.580.11/0.170.26/0.430.18/0.260.13/0.220.21/0.33
SingularProjection0.35/0.420.13/0.190.25/0.440.19/0.320.15/0.250.21/0.32
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.086, + 0.575, + 0.46, + 0.588 + ], + "angle": 0, + "content": "Table 2. Evaluation on the stochastic trajectory prediction task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.624, + 0.47, + 0.653 + ], + "angle": 0, + "content": "surveillance views, and the trajectories are labeled in world coordinates." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Benchmarks. We evaluate our model for the five trajectory prediction tasks: (1) stochastic prediction, (2) deterministic prediction, (3) momentary observation, (4) domain adaptation, and (5) few-shot learning. The stochastic trajectory prediction employs a leave-one-out strategy [5, 19, 22, 53, 66] for training and inference across the five ETH-UCY scenes. In the scenes, the initial 3.2 seconds (equivalent to \\( T_{hist} = 8 \\) frames) of all paths are used as input, and the subsequent 4.8 seconds (corresponding to \\( T_{fut} = 12 \\) frames) are predicted. The model generates \\( S = 20 \\) paths, and one with the lowest error is chosen as the final prediction. In the deterministic prediction, the trajectory prediction methods generate only \\( S = 1 \\) path. For the momentary observation, the observation length is \\( T_{hist} = 2 \\) whereas still \\( T_{fut} = 12 \\). For the domain adaptation, models are trained on each of the five scenes individually, and then evaluated on the other scenes, instead" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.389, + 0.89, + 0.524 + ], + "angle": 0, + "content": "
DeterministicETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.04/1.930.88/1.630.71/1.340.66/1.220.57/1.060.77/1.44
PECNet [45]1.20/2.730.68/1.510.78/1.710.82/1.850.62/1.460.82/1.85
AgentFormer [98]1.67/3.651.61/3.591.62/3.531.85/4.131.68/3.741.69/3.73
MID [18]1.42/2.940.64/1.300.76/1.620.74/1.590.60/1.310.83/1.75
EigenTrajectory [6]0.93/2.050.33/0.640.58/1.230.45/0.990.34/0.750.53/1.13
Social-LSTM [1]1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social-GAN [19]1.13/2.211.01/2.180.60/1.280.42/0.910.52/1.110.67/1.41
SR-LSTM [100]1.01/1.930.35/0.720.66/1.380.56/1.230.44/0.900.60/1.23
STAR-D [97]0.97/2.000.32/0.730.56/1.250.44/0.960.35/0.770.53/1.14
Trajectory++[64]1.02/2.090.33/0.630.52/1.160.42/0.940.32/0.710.52/1.11
SingularTrajectory0.72/1.230.27/0.500.57/1.120.44/0.930.35/0.730.47/0.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.527, + 0.892, + 0.541 + ], + "angle": 0, + "content": "Table 3. Evaluation on the deterministic trajectory prediction task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.894, + 0.678 + ], + "angle": 0, + "content": "of using the leave-one-out strategy. In the few-shot learning, they are trained using only \\(10\\%\\) of the training set, and then assessed on the entire test set. To measure the prediction performances, two common metrics are used for all the tasks, Average Displacement Error (ADE) and Final Displacement Error (FDE), measuring the average and destination error between prediction and its ground-truth path, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.691, + 0.681, + 0.706 + ], + "angle": 0, + "content": "4.2. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Stochastic prediction task. In Tab. 2, since the ETH/UCY dataset release, many state-of-the-art models have made new records. Among them, our model also achieves performance competitive with the comparison methods. Our model regards historical, social, and environmental conditions of observations and images as the most important factors for successfully generating multimodal trajectories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Deterministic prediction task. Next, we evaluate our model in the deterministic trajectory prediction task in Tab. 3. Our SingularTrajectory model has the best performance over both these latest state-of-the-art models, and the models optimized for deterministic prediction. In particular, our model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17895" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.894, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.893, + 0.262 + ], + "angle": 0, + "content": "Figure 4. Visualization of prediction results on (a) momentary observation task and (b) few-shot task. To aid visualization, the best trajectory among \\( S = 20 \\) samples are reported." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.285, + 0.468, + 0.412 + ], + "angle": 0, + "content": "
MomentaryETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.24/2.230.77/1.440.45/0.810.38/0.570.35/0.580.64/1.13
PECNet [45]0.63/1.040.28/0.530.28/0.490.25/0.440.19/0.340.33/0.57
AgentFormer [98]1.10/2.110.50/1.020.52/1.100.56/1.180.43/0.890.62/1.26
MID [18]0.63/1.050.29/0.490.30/0.560.30/0.560.22/0.400.35/0.61
EigenF trajectory [6]0.46/0.760.17/0.280.25/0.440.19/0.350.15/0.270.25/0.42
STT [55]0.72/1.450.48/0.480.53/1.090.64/1.210.44/0.880.57/0.93
STT+DTO [55]0.62/1.220.29/0.560.58/1.140.45/0.980.34/0.740.46/0.93
MOE-Next [77]0.71/1.570.30/0.580.52/1.120.38/0.810.33/0.730.45/0.96
MOE-Traj++ [77]0.64/1.120.20/0.330.33/0.620.22/0.420.17/0.320.31/0.56
SingularF trajectory0.45/0.670.18/0.290.24/0.430.19/0.330.17/0.280.25/0.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.413, + 0.438, + 0.426 + ], + "angle": 0, + "content": "Table 4. Evaluation on the momentary observation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.547 + ], + "angle": 0, + "content": "exhibits significant performance improvements in the ETH and HOTEL scenes. These scenes have noisy observation paths, which have a negative impact on the prediction. We are able to achieve accurate predictions by leveraging the overall motion flow, which acts as a low-pass filter over the noisy sequence, in Singular space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.473, + 0.763 + ], + "angle": 0, + "content": "Momentary observation task. In Tab. 4, the state-of-the-art predictors for momentary observation mainly focus on the coordinates themselves, leading to a performance drop, given only two frames. Some models for this task try to bridge these gaps, but have not been sufficient. Fortunately, with the benefit of our Singular Space, even with only a two-frame input, our model can successfully represent the overall long-term flow. Note that our SingularTrajectory with two-frame observation demonstrates the closest performance to state-of-the-art stochastic prediction models with an entire history frame; even this is achieved without any masked trajectory complement or knowledge distillation. Consequently, this allows it to accurately pinpoint the future locations of pedestrians, whose examples are displayed in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Domain adaptation task. We next evaluate the performance of SingularTrajectory in a domain adaptation task. For simplicity, the ETH, HOTEL, UNIV, ZARA1, and ZARA2 scenes are denoted as A, B, C, D, and E, respectively. For example, 'A2B' means that a model is trained on the ETH scene and tested on the HOTEL scene. As demonstrated in Tab. 1, our SingularTrajectory model shows the performance nearly equivalent to those of models specifically designed for deterministic prediction. Particularly," + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.285, + 0.891, + 0.368 + ], + "angle": 0, + "content": "
Few-ShotETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]0.89/1.631.22/2.480.90/1.610.68/1.251.36/2.121.01/1.82
PECNet [45]0.72/1.460.29/0.530.58/0.930.27/0.440.22/0.380.41/0.75
AgentFormer [98]1.60/2.651.02/1.641.13/1.901.19/2.011.08/1.591.20/1.96
MID [18]0.57/0.920.21/0.330.32/0.600.27/0.490.24/0.420.32/0.55
EigenF trajectory [6]0.39/0.640.13/0.210.25/0.430.21/0.390.15/0.270.23/0.39
SingularF trajectory0.35/0.460.14/0.210.26/0.440.21/0.360.18/0.310.23/0.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.37, + 0.88, + 0.384 + ], + "angle": 0, + "content": "Table 5. Evaluation on the few-shot trajectory prediction task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.501 + ], + "angle": 0, + "content": "our model produces impressive results in the challenging B2A, C2A, D2A and E2A scenarios where even models specialized in domain adaptation fail. Our SingularTrajectory framework specializes in learning general human motions, and is not limited to a specific domain, enabling accurate predictions even in extreme cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.896, + 0.761 + ], + "angle": 0, + "content": "Few-shot task. Finally, the results for the few-shot task are reported in Tab. 5, whose examples are in Fig. 4. As expected, our SingularTrajectory significantly outperforms the existing models. With limited data, the existing works, except the diffusion-based model, tend to easily overfit. In contrast, our methods make it possible to explicitly designate future prototype paths, while taking full advantage of the expression ability of the diffusion model. In particular, even with only \\(10\\%\\) of the data, our model achieves substantial improvements with respect to both data efficiency and performance. Figure 5 illustrates several cases where there are differences between the predictions of SingularTrajectory and other comparison methods. Previous works often show significant performance drops in other associated tasks compared to stochastic prediction tasks. In contrast, our model consistently predicts the best trajectories across multiple tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.665, + 0.786 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The number of motion vectors \\( K \\). First, we conduct a component study by varying the dimension \\( K \\) of Singular space in Tab. 6. To find the best number of singular vectors in general, we carry out experiments across all five tasks using ZARA1 scene where there are both human-environmental and human-human interactions, following [93]. As the number of motion vectors \\( K \\) increases, more detailed movements" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17896" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.47, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.232, + 0.47, + 0.26 + ], + "angle": 0, + "content": "Figure 5. Visualization of prediction consistency across five tasks. The more consistent the prediction is the better." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.275, + 0.466, + 0.357 + ], + "angle": 0, + "content": "
KDeterministicStochasticMomentaryDomainFew-shotAverage
10.52 / 1.030.30 / 0.550.31 / 0.580.51 / 1.070.34 / 0.610.40 / 0.77
20.51 / 1.010.20 / 0.320.20 / 0.330.50 / 1.060.23 / 0.370.33 / 0.62
30.45 / 0.930.19 / 0.330.20 / 0.330.47 / 1.030.22 / 0.380.31 / 0.60
40.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
50.43 / 0.940.19 / 0.330.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
60.43 / 0.940.19 / 0.320.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.36, + 0.452, + 0.373 + ], + "angle": 0, + "content": "Table 6. Ablation study on the Singular space dimension \\(K\\)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.389, + 0.466, + 0.461 + ], + "angle": 0, + "content": "
AdoptionDeterministicStochasticMomentaryDomainFew-shotAverage
Direct0.75 / 1.470.27 / 0.480.31 / 0.540.78 / 1.600.27 / 0.480.48 / 0.91
Initial0.47 / 0.950.22 / 0.400.22 / 0.390.50 / 1.170.23 / 0.400.33 / 0.66
Residual0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
Independent0.46 / 0.940.21 / 0.390.21 / 0.390.49 / 1.160.21 / 0.390.32 / 0.65
Jointly0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.463, + 0.45, + 0.477 + ], + "angle": 0, + "content": "Table 7. Ablation study on the adoption of diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.61 + ], + "angle": 0, + "content": "are captured. In contrast, when using the smaller \\( K \\), it compresses the space, mainly to represent the overall motion flow. The performance tends to plateau when \\( K \\) is larger than 3, as long as \\( K \\) is not too small to cover most movements. We set \\( K = 4 \\) as the dimension for the Singular space because it shows the most effective prediction results across all tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Trajectory denoising methods. Next, we evaluate three types of diffusion models for the trajectory prediction tasks, as shown in Tab. 7. First, we use a basic model similar to MID [18], which directly denoises from a Gaussian noise to a trajectory. This method fails to achieve good performances. The use of an anchor as an intermediate state, similar to LED [48], which reduces the denoising steps by skipping the initial denoising steps, seems to validate its generality. This demonstrates that our adaptive anchor can function as a good initializer for the diffusion model, particularly showing nearly identical outcomes in the stochastic, momentary observation, and few-shot tasks when predicting a multimodal path. However, we confirm that our scheme, which denoises only a residual by adding perturbation to a prototype path in Fig. 6, showcases the best performance in all the tasks. Additionally, compared to refining the prototype paths independently, regarding them as a batch dimension, our model can accurately predict the future when prototype paths are jointly refined." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.894, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.201, + 0.892, + 0.229 + ], + "angle": 0, + "content": "Figure 6. Visualization of anchor refinement. The denoising process progressively refines the prototype paths at each diffusion step \\( m \\)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.243, + 0.891, + 0.315 + ], + "angle": 0, + "content": "
MDeterministicStochasticMomentaryDomainFew-shotAverage
10.45 / 0.950.20 / 0.340.20 / 0.350.48 / 1.050.21 / 0.370.31 / 0.61
20.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
50.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
100.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
250.44 / 0.930.19 / 0.320.20 / 0.340.47 / 1.030.20 / 0.360.30 / 0.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.316, + 0.843, + 0.33 + ], + "angle": 0, + "content": "Table 8. Ablation study on the diffusion steps \\(M\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.509 + ], + "angle": 0, + "content": "Diffusion steps \\( M \\). Lastly, we check how many steps in the diffusion model are needed for the cascaded refinement of the adaptive anchor. In Tab. 8, we confirm that the best performance comes from \\( M = 10 \\). This is because the DDIM scheduler accelerates its convergence and the prototype path provides a rough initial trajectory, and so fewer denoising steps are sufficient. However, as the number of denoising steps increases, the information in the initial prototype path becomes attenuated due to the noise. As a result, we observe a slight decrease in performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.619, + 0.538 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.759 + ], + "angle": 0, + "content": "In this study, we introduce SingularTrajectory, a universal trajectory predictor model for all related trajectory prediction tasks. By unifying trajectory modalities into one Singular space, our approach standardizes trajectory data with shared motion dynamics, which eliminates the need for task-specific adjustments. The incorporation of an adaptive anchor system further personalizes the prototype paths, allowing them to interpret and adapt to environmental contexts and enhancing the reliability of the trajectory prediction. By successfully incorporating Singular space into the diffusion model, our SingularTrajectory framework successfully achieves the state-of-the-art results across five different benchmarks. This establishes SingularTrajectory itself as a general solution that covers a multitude of scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), Vehicles AI Convergence Research & Development Program through the National IT Industry Promotion Agency of Korea (NIPA) funded by the Ministry of Science and ICT (No.S1602-20-1001), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17897" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.471, + 0.184 + ], + "angle": 0, + "content": "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Inhwan Bae and Hae-Gon Jeon. Disentangled multi-relational graph convolutional network for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.244, + 0.472, + 0.298 + ], + "angle": 0, + "content": "[3] Inhwan Bae and Hae-Gon Jeon. A set of control points conditioned pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.301, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Learning pedestrian group representations for multi-modal trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.359, + 0.472, + 0.428 + ], + "angle": 0, + "content": "[5] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.43, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[6] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. EigenTrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.488, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Apratim Bhattacharyya, Michael Hanselmann, Mario Fritz, Bernt Schiele, and Christoph-Nikolas Straehle. Conditional flow variational autoencoders for structured sequence prediction. arXiv preprint arXiv:1908.09008, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.545, + 0.471, + 0.6 + ], + "angle": 0, + "content": "[8] Niccolò Bisagno, Bo Zhang, and Nicola Concì. Group lstm: Group trajectory prediction in crowded scenarios. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.66, + 0.471, + 0.728 + ], + "angle": 0, + "content": "[10] Guangyi Chen, Junlong Li, Nuoxing Zhou, Liangliang Ren, and Jiwen Lu. Personalized trajectory prediction via distribution discrimination. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.731, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[11] Guangyi Chen, Zhenhao Chen, Shunxing Fan, and Kun Zhang. Unsupervised sampling promoting for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.802, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[12] Sourav Das, Guglielmo Camporese, and Lamberto Ballan. Distilling knowledge for short-to-long term trajectory prediction. arXiv preprint arXiv:2305.08553, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Patrick Dendorfer, Aljosa Osep, and Laura Leal-Taixe. Goalgan: Multimodal trajectory prediction based on goal position estimation. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[14] Patrick Dendorfer, Sven Elflein, and Laura Leal-Taixe. M. Gagan: A multi-generator model preventing out-of-distribution samples in pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.163, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[15] Nachiket Deo and Mohan M. Trivedi. Trajectory forecasts in unknown environments conditioned on grid-based plans. arXiv preprint arXiv:2001.00735, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.205, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Proceedings of the Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.247, + 0.894, + 0.301 + ], + "angle": 0, + "content": "[17] Tharindu Fernando, Simon Denman, Sridha Sridharan, and Clinton Fookes. Soft+ hardwired attention: An lstm framework for human trajectory prediction and abnormal event detection. Neural Networks, 108:466-478, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.302, + 0.894, + 0.37 + ], + "angle": 0, + "content": "[18] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.372, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[19] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.442, + 0.894, + 0.481 + ], + "angle": 0, + "content": "[20] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.484, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[22] Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. Stgat: Modeling spatial-temporal interactions for human trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.596, + 0.894, + 0.662 + ], + "angle": 0, + "content": "[23] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.665, + 0.894, + 0.733 + ], + "angle": 0, + "content": "[24] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.735, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[26] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17898" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the International Conference on Learning Representations (ICLR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[28] Vineet Kosaraju, Amir Sadeghian, Roberto Martín-Martín, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.192, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[29] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.263, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[30] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Muse-vae: Multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[31] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B. Choy, Philip H. S. Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.405, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[32] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Computer Graphics Forum, 26(3): 655-664, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.447, + 0.472, + 0.504 + ], + "angle": 0, + "content": "[33] Jiachen Li, Hengbo Ma, and Masayoshi Tomizuka. Conditional generative neural system for probabilistic trajectory prediction. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.505, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[34] Jiachen Li, Fan Yang, Masayoshi Tomizuka, and Chiho Choi. Evolvegraph: Multi-agent trajectory prediction with dynamic relational reasoning. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.561, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[35] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.618, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[36] Yuke Li. Which way are you going? imitative decision learning for path forecasting in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.675, + 0.472, + 0.744 + ], + "angle": 0, + "content": "[37] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[38] Junwei Liang, Lu Jiang, and Alexander Hauptmann. Simaug: Learning robust representations from simulation for trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.802, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[39] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[40] Rongqin Liang, Yuanman Li, Xia Li, Yi Tang, Jiantao Zhou, and Wenbin Zou. Temporal pyramid network for pedestrian" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.895, + 0.133 + ], + "angle": 0, + "content": "trajectory prediction with multi-supervision. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.135, + 0.895, + 0.19 + ], + "angle": 0, + "content": "[41] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.191, + 0.895, + 0.232 + ], + "angle": 0, + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.233, + 0.895, + 0.301 + ], + "angle": 0, + "content": "[43] Yuexin Ma, Xinge Zhu, Xinjing Cheng, Ruigang Yang, Jiming Liu, and Dinesh Manocha. Autotrajectory: Label-free trajectory extraction and prediction from videos using dynamic points. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.302, + 0.895, + 0.357 + ], + "angle": 0, + "content": "[44] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.358, + 0.895, + 0.44 + ], + "angle": 0, + "content": "[45] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.441, + 0.895, + 0.509 + ], + "angle": 0, + "content": "[46] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.511, + 0.895, + 0.551 + ], + "angle": 0, + "content": "[47] Huynh Manh and Gita Alaghband. Scene-lstm: A model for human trajectory prediction. arXiv preprint arXiv:1808.04018, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.553, + 0.895, + 0.621 + ], + "angle": 0, + "content": "[48] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.622, + 0.895, + 0.692 + ], + "angle": 0, + "content": "[49] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.692, + 0.895, + 0.761 + ], + "angle": 0, + "content": "[50] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Multiple trajectory prediction of moving agents with memory augmented networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.761, + 0.895, + 0.815 + ], + "angle": 0, + "content": "[51] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Smemo: Social memory for trajectory forecasting. arXiv preprint arXiv:2203.12446, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.818, + 0.895, + 0.872 + ], + "angle": 0, + "content": "[52] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[53] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgcnn: A social spatio-temporal" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17899" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.471, + 0.147 + ], + "angle": 0, + "content": "graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.149, + 0.472, + 0.231 + ], + "angle": 0, + "content": "[54] Abduallah Mohamed, Deyao Zhu, Warren Vu, Mohamed Elhoseiny, and Christian Claudel. Social-implicit: Rethinking trajectory prediction evaluation and the effectiveness of implicit maximum likelihood estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.233, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[55] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.317, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[56] Ingrid Navarro and Jean Oh. Social-patternn: Socially-aware trajectory prediction guided by motion patterns. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.373, + 0.471, + 0.425 + ], + "angle": 0, + "content": "[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.428, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[58] Bo Pang, Tianyang Zhao, Xu Xie, and Ying Nian Wu. Trajectory prediction with latent belief energy-based model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.485, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[59] Wonhui Park, Dongkwon Jin, and Chang-Su Kim. Eigencontours: Novel contour descriptors based on low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.553, + 0.471, + 0.621 + ], + "angle": 0, + "content": "[60] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.624, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[61] Mark Pfeiffer, Giuseppe Paolo, Hannes Sommer, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. A data-driven model for interaction-aware pedestrian motion prediction in object cluttered environments. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.707, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[62] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.79, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[63] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[64] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajec" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "tory forecasting with heterogeneous data. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[65] Nasim Shafiee, Taskin Padir, and Ehsan Elhamifar. Introvert: Human trajectory prediction via conditional 3d attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.192, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[66] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.261, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[67] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.331, + 0.894, + 0.384 + ], + "angle": 0, + "content": "[68] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.387, + 0.894, + 0.454 + ], + "angle": 0, + "content": "[69] Xiaodan Shi, Xiaowei Shao, Zipei Fan, Renhe Jiang, Haoran Zhang, Zhiling Guo, Guangming Wu, Wei Yuan, and Ryosuke Shibasaki. Multimodal interaction-aware trajectory prediction in crowded space. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.457, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[70] Xiaodan Shi, Xiaowei Shao, Guangming Wu, Haoran Zhang, Zhiling Guo, Renhe Jiang, and Ryosuke Shibasaki. Socialdpf: Socially acceptable distribution prediction of futures. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.894, + 0.592 + ], + "angle": 0, + "content": "[71] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.596, + 0.894, + 0.635 + ], + "angle": 0, + "content": "[72] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.637, + 0.894, + 0.69 + ], + "angle": 0, + "content": "[73] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.693, + 0.894, + 0.747 + ], + "angle": 0, + "content": "[74] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.749, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[75] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[76] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[77] Jianhua Sun, Yuxuan Li, Liang Chai, Hao-Shu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17900" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "momentary observation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.135, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[78] Jianhua Sun, Yuxuan Li, Liang Chai, and Cewu Lu. Stimulus verification is a universal and effective sampler in multi-modal human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.205, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[79] Chaofan Tao, Qinhong Jiang, and Lixin Duan. Dynamic and static context-aware LSTM for multi-agent motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.261, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[80] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1991. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.316, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[81] Daksh Varshneya and G. Srinivasaraghavan. Human trajectory prediction using spatially aware deep attention models. arXiv preprint arXiv:1705.09436, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.359, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[82] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.414, + 0.472, + 0.469 + ], + "angle": 0, + "content": "[83] Anirudh Vemula, Katharina Muelling, and Jean Oh. Social attention: Modeling attention in human crowds. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.47, + 0.472, + 0.511 + ], + "angle": 0, + "content": "[84] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters (RA-L), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.511, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[85] Yuning Wang, Pu Zhang, Lei Bai, and Jianru Xue. Fend: A future enhanced distribution-aware contrastive learning framework for long-tail trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.581, + 0.472, + 0.636 + ], + "angle": 0, + "content": "[86] Song Wen, Hao Wang, and Dimitris Metaxas. Social ode: Multi-agent trajectory forecasting with neural ordinary differential equations. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.637, + 0.472, + 0.705 + ], + "angle": 0, + "content": "[87] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.706, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[88] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.776, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[89] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.845, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[90] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[91] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[92] Yi Xu, Jing Yang, and Shaoyi Du. Cf-lstm: Cascaded feature-based long short-term networks for predicting pedestrian trajectory. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.234, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[93] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.291, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[94] Hao Xue, Du Q Huynh, and Mark Reynolds. Ss-lstm: A hierarchical LSTM model for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.348, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[95] Kota Yamaguchi, Alexander C Berg, Luis E Ortiz, and Tamara L Berg. Who are you with and where are you going? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.405, + 0.894, + 0.46 + ], + "angle": 0, + "content": "[96] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters (RA-L), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.461, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[97] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.518, + 0.894, + 0.586 + ], + "angle": 0, + "content": "[98] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.588, + 0.894, + 0.642 + ], + "angle": 0, + "content": "[99] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.894, + 0.713 + ], + "angle": 0, + "content": "[100] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.894, + 0.77 + ], + "angle": 0, + "content": "[101] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.894, + 0.841 + ], + "angle": 0, + "content": "[102] Tianyang Zhao, Yifei Xu, Mathew Monfort, Wongun Choi, Chris Baker, Yibiao Zhao, Yizhou Wang, and Ying Nian Wu. Multi-agent tensor fusion for contextual trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.842, + 0.894, + 0.884 + ], + "angle": 0, + "content": "[103] ChunYu Zhi, HuaiJiang Sun, and Tian Xu. Adaptive trajectory prediction without catastrophic forgetting. The Journal of Supercomputing, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17901" + } + ] +] \ No newline at end of file diff --git a/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_origin.pdf b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb77e70086b43d9593a79e8abbfa72be7a9976a7 --- /dev/null +++ b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/2a853686-8806-4e80-a078-f0cdecbd3caf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cebb55f923b9c740fee041a3828eaba76ad7a60f5796dd450b7a3a32c926fa5 +size 2163163 diff --git a/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/full.md b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ce2abcc789f211487b62c93c8d6bdfc180fc9c47 --- /dev/null +++ b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/full.md @@ -0,0 +1,390 @@ +# SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model + +Inhwan Bae, Young-Jae Park and Hae-Gon Jeon* +AI Graduate School, GIST, South Korea + +{inhwanbae, youngjae.park}@gm.gist.ac.kr, haegonj@gist.ac.kr + +# Abstract + +There are five types of trajectory prediction tasks: deterministic, stochastic, domain adaptation, momentary observation, and few-shot. These associated tasks are defined by various factors, such as the length of input paths, data split and pre-processing methods. Interestingly, even though they commonly take sequential coordinates of observations as input and infer future paths in the same coordinates as output, designing specialized architectures for each task is still necessary. For the other task, generality issues can lead to sub-optimal performances. In this paper, we propose SingularTrajectory, a diffusion-based universal trajectory prediction framework to reduce the performance gap across the five tasks. The core of SingularTrajectory is to unify a variety of human dynamics representations on the associated tasks. To do this, we first build a Singular space to project all types of motion patterns from each task into one embedding space. We next propose an adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods that sometimes yield unacceptable paths, our adaptive anchor enables correct anchors, which are put into a wrong location, based on a traversability map. Finally, we adopt a diffusion-based predictor to further enhance the prototype paths using a cascaded denoising process. Our unified framework ensures the generality across various benchmark settings such as input modality, and trajectory lengths. Extensive experiments on five public benchmarks demonstrate that SingularTrajectory substantially outperforms existing models, highlighting its effectiveness in estimating general dynamics of human movements. Code is publicly available at https://github.com/inhwanbae/SingularTrajectory. + +# 1. Introduction + +Extensive studies of trajectory prediction methods have been conducted in the computer vision field for several decades [20, 52]. They have demonstrated its importance in various applications, including crowd simulation, social robot navi + +![](images/c6cf9b4473ac034a222fb40a9a888610f4aecf95e6ae1053e9f1b7693c80a41a.jpg) +Figure 1. An overview of our SingularTrajectory framework. All relevant human trajectory prediction tasks can be represented in our Singular space, a unified feature embedding space for human dynamics. Using the embedding features, our diffusion-based model for a universal trajectory prediction makes prediction for all the tasks in this same space. + +gation, obstacle avoidance and security and surveillance systems, etc. Trajectory prediction takes sequential coordinate values of agents as input and infers their future pathways in common [1, 19, 83]. Such tasks vary depending on the application. The tasks are determined with respect to the number of input/output, data processing, and the use of geological information. + +The tasks for trajectory prediction can be categorized into five groups: (1) Stochastic prediction is to predict 20 multi-modal future trajectories from one observation with 8 frames [19]. Each future trajectory consists of 12 frames. (2) Deterministic prediction takes one observation with 8 frames, but infers only one future trajectory with 12 frames [1]. (3) The momentary observation uses only two frames to predict 20 multi-modal future paths with 12 frames [77]. (4) Domain adaptation splits training data with respect to places in datasets, trains a model on one place, and then checks the transferability to other places [93]. (5) The few-shot task only uses partial data to build a dataset-efficient model [53], as illustrated in Fig. 1. Until now, each specialized architecture for a task type has provided performance gains. + +However, two questions arise. First, why do state-of-the-art models for one task undergo significant performance drops when applied to other trajectory prediction tasks? Second, is it feasible to design a general predictor that works across the five tasks? + +As answers to these questions, we present a universal tra + +jectionary predictor to achieve generality in predictions, named SingularTrajectory. The main idea is to unify the modalities of human dynamic representations across the five tasks. To do this, we first introduce a Singular space, an embedding space consisting of representative motion patterns for each task. The motion patterns play a role in the basis function for pedestrian movements, and are extracted using Singular Value Decomposition (SVD). They are then projected onto Singular space. + +We next propose an environment-adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods [6, 29] that sometimes fail to handle different target data distribution, our adaptive anchor is able to correct prototype paths from the adaptive anchor in Singular space if they are put into the wrong locations, based on an input traversability map. Lastly, we generate socially-acceptable future trajectories for all agents in scenes with a diffusion-based predictor which denoises residuals of perturbed prototype paths. Thanks to the cascaded denoising process of diffusion models, we refine the prototype paths. Here, historical pathways, agent interactions, and environmental information are provided as conditions to guide them in the Markov chain of the denoising diffusion processes. + +Experimental results demonstrate that our SingularTrajectory can successfully represent pedestrian motion dynamics, and significantly improve prediction accuracy for the five tasks on challenging public benchmark datasets. + +# 2. Related Works + +We review previous studies on trajectory prediction that have attempted to address various benchmark scenarios. + +# 2.1. Pedestrian Trajectory Prediction + +Ways of predicting pedestrian future trajectories have been studied for a long time in the computer vision field. Pioneering works [20, 52, 60, 95] model an invisible social norm based on motion dynamics as an energy minimization problem. Introducing recurrent neural networks [1, 19, 64, 97] achieves significant improvements by providing highly sequential representations of high-level shape of paths. These methods determine the most probable path, called deterministic trajectory prediction. The following works model mutual influences among agents using attention mechanisms [17, 23, 64, 83], graph convolutional networks [2, 9, 27, 41, 53, 54, 75], graph attention networks [5, 22, 28, 38, 39, 66, 82], and transformers [4, 18, 55, 68, 86, 87, 97, 98]. Additional visual information allows us to leverage environmental constraints from traversability maps [13-15, 28, 37, 46, 47, 49, 50, 63, 65, 74, 75, 79, 81, 85, 94, 99, 102]. Depending on the constraints, predictors take either recurrent [1, 8, 10, 11, 18, 19, 30, 43, 44, 51, 56, 61, 64, 92, 100, 101] or simultaneous approaches [2, 3, 35, 53, 66, 67, 89] to extrapolate the future pathways. + +Meanwhile, with the success of generative models, the importance of multimodality has begun to emerge, called stochastic trajectory prediction. Stochastic prediction enables us to consider all of an agent's possible future pathways. For example, an agent at a crossroads may either walk straight or turn left/right. Here, the stochastic prediction infers all potential future modes. This approach has become mainstream in this field. Starting from Social-GAN [19], a bivariate Gaussian distribution [1, 2, 34, 53, 66, 69, 70, 93, 96, 97], generative adversarial network [14, 19, 22, 28, 36, 40, 63, 74, 78, 102], and conditional variational autoEncoder [7, 10, 23, 30, 31, 33, 45, 64, 76, 84, 88, 91] have been adopted for stochastic trajectory prediction. Anchorm-conditioned methods can explicitly represent different modalities by prototyping possible paths [6, 29]. Most recently, diffusion-based models have revealed their tremendous representation capacities in numerous tasks [16, 21, 57, 71-73], proving its potential for stochastic trajectory prediction [18, 25, 48, 62]. In this study, we take full advantage of both the anchor-conditioned approach and the diffusion-based model to achieve the explainability and generalizability of the trajectory prediction tasks. + +# 2.2. Various Trajectory Prediction Tasks + +Beyond the standard benchmark protocol of stochastic prediction, there are three other variants of this task: momentary observation, domain adaptation, and few-shot learning. Works in [12, 55, 77] only take two frames as input for the momentary trajectory prediction. Multi-task learning, self-supervised learning, and knowledge distillation techniques have been used to extract rich features from the limited data. Another works [24, 93, 103] focuses on domain adaptation across trajectory domains, captured from different surveillance views. A transferable graph neural network is introduced to adaptively learn domain-invariant knowledge. The others [53, 101] adopt few-shot learning for better training efficiency. + +Although these works take and infer the sequential coordinates of agent trajectories in common, there is no unified model for all the tasks. Despite the tremendous efforts to design a specialized architecture for one task, they cannot be applied to the other tasks without suffering significant performance drops. In the next section, we will describe how to design a unified architecture, which consistently produces promising results on the five associated tasks. + +# 3. Methodology + +We describe how to learn a general representation of human motions. We first define a general trajectory prediction problem in Sec. 3.1 and provide preliminaries of explicit formulations on the SVD and diffusion process in Sec. 3.2. We then introduce a motion feature extraction to build our Singular space, and a projection of any trajectory from each task onto + +it in Sec. 3.3. Next, we propose an environment-adaptive anchor using motion vectors and input image in Sec. 3.4. Finally, we present the SingularTrajectory predictor based on the diffusion model in Sec. 3.5. + +# 3.1. Problem Definition + +Trajectory prediction aims to predict the future paths of agents based on their historical path and surrounding contexts. Suppose that at each timestamp $t$ , there are $N$ pedestrians in a scene with the 2D spatial coordinate position $\{\pmb{p}_t^n \in \mathbb{R}^2 | n \in [1, \dots, N]\}$ . A pedestrian historical trajectory $\pmb{X}_n$ over $T_{hist}$ timesteps can be represented as the cumulative coordinates $\pmb{X}_n = \{\pmb{p}_t^n | t \in [1, \dots, T_{hist}]\}$ . Similarly, future trajectories $\pmb{Y}_n$ for the time duration $T_{fut}$ to be predicted can be written as $Y_n = \{\pmb{p}_t^n | t \in [T_{hist} + 1, \dots, T_{hist} + T_{fut}]\}$ . The prediction system takes both the historical trajectories for all $N$ people $X = \{X_n | n \in [1, \dots, N]\}$ and the scene image map $I$ for environmental information as input. The deterministic prediction system infers one sequence of the most reliable future trajectory $\widehat{Y} = \{\widehat{Y}_n | n \in [1, \dots, N]\}$ . For the stochastic prediction, because of the indeterminacy of the future movements, $S$ multiple pathways for all the $N$ pedestrians $\widehat{Y} = \{\widehat{Y}_n^s | n \in [1, \dots, N], s \in [1, \dots, S]\}$ are generated so that at least one sample is close to the ground-truth trajectory. + +# 3.2. Preliminaries + +Singular Value Decomposition. Singular Value Decomposition (SVD) decomposes a matrix into three resultant matrices. Given a matrix $\mathbf{A}$ , its SVD is represented as: + +$$ +\boldsymbol {A} = \boldsymbol {U} \boldsymbol {\Sigma} \boldsymbol {V} ^ {\top}, \tag {1} +$$ + +where $\pmb{U}$ is an orthogonal left singular vector matrix, whose columns are eigenvectors of $A\times A^{\top}$ . $\pmb{\Sigma}$ is a diagonal matrix with the singular values of $A$ , consisting of $K$ non-negative values in descending order. $\pmb{V}$ is a right singular vector matrix, which is also orthogonal and its columns are the eigenvectors of $A^{\top}\times A$ . + +To remove the redundant part of the raw data, the truncation technique is often applied to the results after the decomposition. The idea behind the truncated SVD is to approximate the original matrix $\mathbf{A}$ with the lower rank. With the $K$ to determine the number of singular values to retain, $\pmb{\Sigma}$ can be simplified to $\pmb{\Sigma}_{K}$ which contains only the $K$ largest singular values. Similarly, $\pmb{U}$ and $V^{\top}$ are reduced to $U_{K}$ and $V_{K}^{\top}$ by keeping the first $K$ columns and rows, respectively. This process eliminates the smallest singular values, which are not needed to express the original data and often correspond to noise or redundant information. This is useful for practical scenarios dealing with large and potentially sparse matrices because we can reconstruct a close approximation of the original data with significantly less storage space. + +Diffusion models. The diffusion model operates by transforming a noisy distribution, represented by the noise vector + +$\mathbf{y}_M$ , into the desired data $\mathbf{y}_0$ through a series of $M$ diffusion steps. These steps involve intermediate latent variables $\{\mathbf{y}_m|m\in [1,\dots ,M]\}$ , and encompass both the diffusion and denoising processes. The diffusion process adds a small amount of noise to data in order to obtain the standard normal distribution $q(\mathbf{y}_M)$ from the distribution $q(\mathbf{y}_0)$ using the Markov chain as: + +$$ +q \left(\mathbf {y} _ {1: M} \mid \mathbf {y} _ {0}\right) := \prod_ {m = 1} ^ {M} q \left(\mathbf {y} _ {m} \mid \mathbf {y} _ {m - 1}\right) \tag {2} +$$ + +$$ +q \left(\mathbf {y} _ {m} \mid \mathbf {y} _ {m - 1}\right) := \mathcal {N} \left(\mathbf {y} _ {m}; \sqrt {1 - \beta_ {m}} \mathbf {y} _ {m - 1}, \beta_ {m} \mathbf {I}\right) +$$ + +where $\beta_{t}$ is a small positive constant and a variance schedule for adding noise. The denoising process uses the $\mathbf{y}_m$ to recover $\mathbf{y}_0$ with a learnable network as follows: + +$$ +p _ {\theta} \left(\mathbf {y} _ {0: M}\right) := p \left(\mathbf {y} _ {M}\right) \prod_ {m = 1} ^ {M} p _ {\theta} \left(\mathbf {y} _ {m - 1} \mid \mathbf {y} _ {m}\right), \tag {3} +$$ + +$$ +p _ {\theta} (\mathbf {y} _ {m - 1} | \mathbf {y} _ {m}) := \mathcal {N} (\mathbf {y} _ {m - 1}; \boldsymbol {\epsilon} _ {\theta} (\mathbf {y} _ {m}, m), \beta_ {m} \mathbf {I}). +$$ + +where $\mathbf{y}_M\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ is an initial noise sampled from the Gaussian distribution $p(\mathbf{y}_M)$ , and $\theta$ denotes the learnable parameter of the diffusion model. $\epsilon_{\theta}$ is a learnable denoising model of a clean data $\mathbf{y}_0$ , and a corrupted data $\mathbf{y}_m$ at a step $m$ . The objective is to train the neural network so that the denoising process predicts the true data-generating distribution well. This is often done by maximizing the evidence lower bound, ensuring that the samples generated by the diffusion model are indistinguishable from the real data. + +# 3.3. Unifying the Motion Space + +The trajectory prediction model uses a learnable network to capture the relationship in consideration of the input coordinates of pedestrians, input images and output coordinates for each pedestrian. Since expectations for input and output trajectories are different for each associated task (e.g., length and multimodality), they should be viewed as different data spaces even if they use the same coordinate systems. We introduce a method to merge raw data in each space into a Singular space for human motion dynamics. + +Singular space construction. To discover motion dynamics from the raw data, we first extract primitive motion features. Inspired by the successful low-rank approximation of raw trajectory data using eigenvectors in [6], we also employ a similar strategy using singular vectors from the truncated SVD to extract principal motion components from the entire training dataset. + +First, we cut-off paths of all pedestrians in the dataset into $T_{win}$ lengths through a sliding window to create a total of $L$ gist of trajectory set $A \in \mathbb{R}^{L \times (2 \times T_{win})}$ . Here, $A$ is a temporary matrix to extract a motion vector from a set of trajectory in the initialization phase. Next, we decompose $A$ to obtain truncated matrices $U_K \in \mathbb{R}^{L \times K}$ , $\pmb{\Sigma}_K \in \mathbb{R}^{K \times K}$ + +![](images/ab4c686765e04400c9a24dc333bb76be181683d3ab945f60b722742b7b8d2dc7.jpg) +Figure 2. Visualization of the trajectories in Singular space. (a) The circle and triangle markers indicate the history and future trajectory coordinates, respectively. Each color also refers to each associated task. (b-d) Each marker in the Singular space corresponds to each trajectory in raw data. And, the slicing planes, representing a set of arrows, mean human dynamics of straight forward motions and turning motions. + +![](images/6a07dc740514b4b0cc3fe84a0295dacfd257db2a5019d448d04a1fd5208c1175.jpg) + +![](images/897df13df0786f68e036ad5758762e54972a06bc1b36f6c231928b4a098606df.jpg) + +![](images/9f800c2596ad3f2d2c80ab498aa9a7be8a1395a9e5b21235f181adb944a62389.jpg) +Figure 3. An example of the adaptive anchor generation. (a) The initial prototype anchor $\mathcal{P}$ is placed on the last observation coordinate of a person. In this instance, four prototype paths (highlighted in red) are incorrectly placed at the non-traversable locations. (b) Vector field $\vec{F}_I(x,y)$ is computed to guide toward in the nearest traversable areas. (c) The initial prototype paths are then tailored to the environment using the vector field. + +and $V_{K}^{\top}\in \mathbb{R}^{K\times (2\times T_{win})}$ . Here, $V_{K}^{\top}$ is a set of $K$ spatiotemporal motion vectors $\mathrm{v}_k\in \mathbb{R}^{2\times T_{win}}$ representing the most dominant motion dynamics of pedestrians. Since $\mathrm{v}_k$ are orthogonal to each other, we define a Singular space coordinate system using the $K$ singular vectors as basis vectors for each axis. In Singular space, trajectories $\pmb{A}$ can be a coordinate of $\mathcal{A} = U_K\cdot \pmb {\Sigma}_K\in \mathbb{R}^{L\times K}$ , indicating that each motion vector has an influence on reconstructing the $L$ trajectories. In other words, we can project the $\pmb{A}$ into the coordinate in Singular space $\mathcal{A}$ as follows: + +$$ +\mathcal {A} = \boldsymbol {A} \times \left(\boldsymbol {V} _ {K} ^ {\top}\right) ^ {- 1} = \boldsymbol {A} \times \boldsymbol {V} _ {K}, \tag {4} +$$ + +where $(\mathbf{V}_K^\top)^{-1}$ can be simplified into $\mathbf{V}_K$ with the property of an orthogonal matrix1. + +In the Singular space, we can now concentrate on the motion flow by using the coefficients of motion patterns, instead of considering every position over time. Note that the results from the SVD depend highly on various factors such as the window size for inputs/outputs, data processing, and the differences in the raw data caused by geological variations. Due to the issue, works using SVD only convert the output space [26, 59, 80] or conduct separate decompositions for each of input and output data spaces [6]. However, in this case, they cause inconsistent representations of motion dynamics, even for the same pedestrian, which leads to a lack of generality. In the next step, we introduce a new method to address this issue. + +Projection of any trajectory into Singular space. We aim to express the input/output trajectories $\mathbf{X}$ and $\mathbf{Y}$ of the five associated tasks in Singular space all at once. Our core idea stems from the notion that all pedestrians share the same human motion dynamics and thus will likely show a similar pattern. Starting with the projection function Eq. (4), which is a projection matrix for the fixed trajectory length $T_{win}$ , we extend it to any length $T_{hist}$ and $T_{fut}$ , which varies from task to task. To handle the motion patterns regardless of their lengths, we interpolate $\mathrm{v}_k \in \mathbb{R}^{2 \times T_{win}}$ to $\mathrm{v}_{\mathrm{x},k} \in \mathbb{R}^{2 \times T_{hist}}$ . + +![](images/daba69e12be8e95b117f5a30ce2d5f08c28f51143f2fa1557b9c7bdca9a2c8ec.jpg) +(a) Initial prototype paths + +![](images/70ab6f266e750f341fc8ab0e1381dafd4e118302daabb102f0be48a0ca6f9bb8.jpg) +(b) Vector field + +![](images/d2215bdab00f8ff960015d0cfb9cfc00536b11e95ed457dd3130f3343b5f196b.jpg) +(c) Final prototype paths + +Since the motion pattern can be seen as a 2-dimensional curve, we use Cardinal B-splines to make a transformation matrix $C_{T_{hist}} \in \mathbb{R}^{(2 \times T_{hist}) \times (2 \times T_{win})}$ using the Irwin-Hall distribution. The constant value $C$ depends only on the length, so $\mathrm{v}_{\mathrm{x},k}$ can be approximated. The trajectory $\mathbf{X} \in \mathbb{R}^{N \times (2 \times T_{hist})}$ is projected to the coordinate $\mathcal{X} \in \mathbb{R}^{N \times (2 \times T_{hist})}$ in Singular space as follows: + +$$ +v _ {x, k} = C _ {T _ {\text {h i s t}}} \times v _ {k}. \tag {5} +$$ + +$$ +\mathcal {X} = \boldsymbol {X} \times \boldsymbol {C} _ {T _ {\text {h i s t}}} \times \boldsymbol {V} _ {K}. \tag {6} +$$ + +In the same way as Eq. (6), the trajectory $\mathbf{Y} \in \mathbb{R}^{N \times (2 \times T_{\text{fit}})}$ also can be projected to $\mathcal{V} \in \mathbb{R}^{N \times K}$ using $\mathbf{V}_K$ and $C_{T_{\text{fit}}}$ . + +Through this process, Singular space can represent a variety of trajectories even with different lengths, including motion vectors for the input/output as well as the task-specific data, as visualized in Fig. 2. Establishing a common ground for the trajectory representation is crucial for our model's adaptability and robust performances across different trajectory prediction benchmarks. Moreover, the strategy of focusing on the overall motion flow, rather than frame-by-frame coordinates, further improves the model's capacity to understand and predict socially compliant trajectories in diverse real-world scenarios. + +# 3.4. Adaptive Anchor + +Beyond the trajectory data integration, we also introduce how to incorporate environmental contexts into trajectory predictors. To do this, we propose an adaptive anchor which is expressed as a set of motion vectors from input images and consists of prototype paths in Singular space. + +Prototype anchor formation. Previous anchor-based human trajectory prediction approaches [6, 29] use a fixed anchor for any pedestrian, and the anchor is refined for output trajectories. Although these methods can explicitly model the multimodality, they sometimes fail to handle the case where the prototype path is put into wrong locations, or blocked by static obstacles. This is because the wrong prototypes are treated as a hard constraint. To avoid this problem, + +we use the input image as a traversability map to correct the wrong prototype paths. + +We start with the converted future trajectory $\mathcal{V}$ in the whole dataset $Y$ using Sec. 3.3. Following [6, 59], the coordinates $\mathcal{V}$ are then clustered into $S$ centroids, which can be viewed as groups of components representing different multimodal futures. In addition, each centroid in Singular space is used as an initial prototype path $\mathcal{P}_s$ for constructing an anchor $\mathcal{P} \in \mathbb{R}^{S \times K}$ . But, these initial prototype paths are still fixed, so they still have a limitation which is unlikely to consider environmental information. + +Adaptive anchor generation. To make use of environmental information, we introduce a module to deform the anchor. Using an off-the-shelf semantic segmentation model from [46], we can obtain binary traversable maps $\mathbf{I}_{\text{map}}$ from an input image. We then derive a vector field $\vec{F}_I(x,y)$ to fix the wrongly located prototype path by directing it to a nearest traversable regions. If the initial prototype path $\mathcal{P}_s$ is in the wrong place, we deform it by adding the vector fields into the initial anchor until they reach equilibrium states to obtain the final prototype paths $\mathcal{P}_s'$ as follows: + +$$ +\mathcal {P} _ {s} ^ {\prime} = \mathcal {P} _ {s} + \vec {F} _ {I} \left(\mathcal {P} _ {s} \boldsymbol {V} _ {K} ^ {\top} \boldsymbol {C} _ {T _ {f u t}} ^ {- 1}\right). \tag {7} +$$ + +Through this process, the prototype path is re-located toward the nearest walkway, as demonstrated in Fig. 3. In other words, the anchor plays a role in the environment-adaptive prototype paths, unlike the existing fixed anchor methods. + +Because the prototype paths are in Singular space, the scene image can be implicitly represented as the adaptive anchor. In addition, the scene image can be approximated with the set of motions. By projecting the scene images into the Singular space, our model understands the surrounding environment. This holistic approach, integrating both coordinate and environmental cues, sets the stage for more realistic and reliable trajectory predictions. + +# 3.5. Diffusion-Based SingularTrajectory Model + +As the final step, we develop a framework called Singular-Trajectory, a unified model that works well across the five tasks. Leveraging the Singular space and the adaptive anchor, our diffusion-based SingularTrajectory model can precisely forecast potential future paths. + +Denoising a perturbed trajectory anchor. Unlike previous diffusion-based predictors, which directly forecast the future path from Gaussian noise [18, 25], we devise a stepwise refinement from the adaptive anchor for a realistic trajectory. Here, historical pathways $\mathcal{X}$ , environmental information $\mathcal{P}$ , and agent interactions $\mathcal{G}(\mathcal{X},\mathcal{P})$ are encoded and used as conditions to guide the denoising processes. To better capture agents interactions $\mathcal{G}$ in the denoising process, we adopt the transformer model. Similar to other transformer-based trajectory predictors [18, 48], our SingularTrajectory + +encodes spatio-temporal information to account for agent-agent and agent-environment interactions using $\mathcal{X}$ and $\mathcal{P}$ , respectively. These conditions are then concatenated into one feature vector representation and fed into the diffusion model $\epsilon_{\theta}(\mathbf{y}_m, m, \mathcal{X}, \mathcal{P}, \mathcal{G})$ to learn, by contrasting the motion patterns from previous diffusion steps to distinguish the added noise, and to close the reality gap by generating socially-acceptable future paths $\widehat{\mathbf{Y}}$ . + +This refinement process works in a cascading manner by $\{\mathcal{P},\dots,\widehat{\mathcal{V}}\}_{m = 1}^{M}$ as follows: + +$$ +\widehat {\mathcal {Y}} = \mathcal {P} + p \left(\mathbf {y} _ {M}\right) \prod_ {m = 1} ^ {M} p _ {\theta} \left(\mathbf {y} _ {m - 1} \mid \mathbf {y} _ {m}\right). \tag {8} +$$ + +By predicting only the residuals $\mathbf{y}_m\in \mathbb{R}^{S\times K}$ to adjust the anchors, the problem is simplified, in that the model is able to use a prior knowledge on the initial state. With this process, we thus ensure more precise and reliable trajectory generations, achieving generality in various environments and applications. + +Implementation details. To construct Singular space, we empirically set $K$ to 4. We set $T_{win} = T_{fut} = 12$ in order to prevent information loss due to the approximation of the motion vector during prediction. For an anchor diffusion model, we devise a one-layer transformer for encoding motion and context information, where the dimension is set to 256 with 4-head attention. We set $M = 10$ and schedule the diffusion timesteps following DDIM [72]. To train the SingularTrajectory in an end-to-end manner, we use a mean square error (MSE) as a loss function between the output and a random Gaussian noise for the current iteration. The training is performed with AdamW optimizer [42], with a learning rate of 0.001 and batch size of 512 for 256 epochs. All the experiments are conducted on a single NVIDIA A6000 GPU, which usually takes about an hour to train each scene. + +# 4. Experiments + +In this section, we conduct comprehensive experiments to verify the generality of our SingularTrajectory model for the trajectory prediction tasks. We first describe our experimental setup in Sec. 4.1. We then provide comparison results with state-of-the-art models on public benchmark datasets in Sec. 4.2. Finally, we perform an extensive ablation study demonstrating the effect of each component of our method in Sec. 4.3. + +# 4.1. Experimental Setup + +Datasets. To compare our SingularTrajectory with state-of-the-art baselines, we conduct quantitative evaluations on two common datasets, ETH [60] and UCY [32], for all the associated tasks. The ETH and UCY datasets consist of various motions of 1,536 pedestrians across five unique scenes: ETH, Hotel, Univ, Zara1 and Zara2. They are recorded in + +
Domain Adaptation (ADE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAVG
STGCNN [53]1.831.581.301.311.513.021.382.631.582.151.160.700.820.540.811.041.050.730.470.820.981.090.740.500.831.22
PECNet [45]1.971.681.241.351.563.111.352.691.622.191.390.820.930.570.931.101.170.920.520.931.011.250.830.610.931.31
AgentFormer [98]1.731.641.871.571.701.491.762.131.731.781.730.192.011.791.431.712.051.921.321.751.961.911.891.911.921.72
MID [18]1.261.431.651.431.441.230.870.980.770.961.610.630.810.580.911.420.750.810.740.931.470.680.880.910.981.04
EigenTrajectory [6]0.390.710.730.450.570.950.680.510.420.640.940.320.490.370.531.101.031.360.430.980.920.510.620.520.640.67
T-GNN [93]1.131.250.941.031.092.541.082.251.411.820.970.540.610.230.590.880.780.590.320.640.870.720.650.340.650.96
K0 [24]0.450.780.680.590.630.870.650.490.360.591.080.460.440.360.591.130.560.590.530.701.130.390.590.480.650.63
Adaptive [24]0.360.640.530.450.500.900.620.500.360.601.110.460.480.420.621.110.520.540.470.661.050.340.560.440.600.59
SingularF trajectory0.290.590.510.420.450.660.550.450.350.500.790.300.480.360.480.730.290.550.360.490.640.230.550.450.470.48
Domain Adaptation (FDE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAvgV
STGCNN [53]3.242.862.532.432.775.162.514.862.883.852.301.341.741.101.622.211.991.410.881.622.102.051.471.011.662.30
PECNet [45]3.332.832.532.452.795.232.484.902.863.872.221.321.681.121.592.202.051.520.881.662.101.841.450.981.592.29
AgentFormer [98]3.603.544.523.953.903.763.675.173.424.013.773.494.123.763.793.883.754.273.613.883.823.553.744.323.863.89
MID [18]2.422.773.222.772.802.341.661.871.441.833.131.281.741.241.852.731.491.671.541.862.871.411.881.932.022.07
EigenTrajectory [6]0.761.431.560.961.182.101.381.070.861.352.010.641.100.821.142.421.031.360.971.442.040.511.291.111.241.27
T-GNN [93]2.182.251.781.842.014.151.824.042.533.141.911.121.300.871.301.921.461.250.651.321.861.451.280.721.331.82
K0 [24]0.941.721.511.291.371.811.331.020.751.232.170.910.940.811.212.180.981.201.031.352.120.681.221.031.261.28
Adaptive [24]0.691.341.120.911.021.851.271.050.761.232.260.921.030.941.292.231.021.171.001.362.070.631.190.941.211.22
SingularF trajectory0.571.191.080.810.911.161.170.970.731.011.360.621.110.790.971.220.521.180.730.921.070.431.150.940.900.94
+ +Table 1. Comparison of our SingularTrajectory with other state-of-the-art methods in domain adaptation task. Following [24, 93], we deterministically predict one future path for all dataset pairs (ADE/FDE). Bold: Best, Underline: Second best. + +
StochasticETHHOTELUNIVZARA1ZARA2AVG
Social-GAN [19]0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
STGCNN [53]0.64/1.110.49/0.850.44/0.790.34/0.530.30/0.480.44/0.75
STAR [97]0.57/1.110.19/0.370.35/0.750.26/0.570.25/0.580.33/0.68
PECNet [45]0.61/1.070.22/0.390.34/0.560.25/0.450.19/0.330.32/0.56
MID [18]0.57/0.930.21/0.330.29/0.550.28/0.500.20/0.370.31/0.54
LBEBM [58]0.60/1.060.21/0.380.28/0.540.21/0.390.15/0.300.29/0.53
Trajectory++[64]0.61/1.030.20/0.280.30/0.550.24/0.410.18/0.320.31/0.52
GroupNet [88]0.46/0.730.15/0.250.26/0.490.21/0.390.17/0.330.25/0.44
AgentFormer [98]0.46/0.800.14/0.220.25/0.450.18/0.300.14/0.240.23/0.40
GP-Graph [4]0.43/0.630.18/0.300.24/0.420.17/0.310.15/0.290.23/0.39
NPSN [5]0.36/0.590.16/0.250.23/0.390.18/0.320.14/0.250.21/0.36
EqMotion [90]0.40/0.610.12/0.180.23/0.430.18/0.320.13/0.230.21/0.35
EigenIProjection [6]0.36/0.530.12/0.190.24/0.430.19/0.330.14/0.240.21/0.34
SocialVAE [91]0.41/0.580.13/0.190.21/0.360.17/0.290.13/0.220.21/0.33
LED [48]0.39/0.580.11/0.170.26/0.430.18/0.260.13/0.220.21/0.33
SingularProjection0.35/0.420.13/0.190.25/0.440.19/0.320.15/0.250.21/0.32
+ +Table 2. Evaluation on the stochastic trajectory prediction task. + +surveillance views, and the trajectories are labeled in world coordinates. + +Benchmarks. We evaluate our model for the five trajectory prediction tasks: (1) stochastic prediction, (2) deterministic prediction, (3) momentary observation, (4) domain adaptation, and (5) few-shot learning. The stochastic trajectory prediction employs a leave-one-out strategy [5, 19, 22, 53, 66] for training and inference across the five ETH-UCY scenes. In the scenes, the initial 3.2 seconds (equivalent to $T_{hist} = 8$ frames) of all paths are used as input, and the subsequent 4.8 seconds (corresponding to $T_{fut} = 12$ frames) are predicted. The model generates $S = 20$ paths, and one with the lowest error is chosen as the final prediction. In the deterministic prediction, the trajectory prediction methods generate only $S = 1$ path. For the momentary observation, the observation length is $T_{hist} = 2$ whereas still $T_{fut} = 12$ . For the domain adaptation, models are trained on each of the five scenes individually, and then evaluated on the other scenes, instead + +
DeterministicETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.04/1.930.88/1.630.71/1.340.66/1.220.57/1.060.77/1.44
PECNet [45]1.20/2.730.68/1.510.78/1.710.82/1.850.62/1.460.82/1.85
AgentFormer [98]1.67/3.651.61/3.591.62/3.531.85/4.131.68/3.741.69/3.73
MID [18]1.42/2.940.64/1.300.76/1.620.74/1.590.60/1.310.83/1.75
EigenTrajectory [6]0.93/2.050.33/0.640.58/1.230.45/0.990.34/0.750.53/1.13
Social-LSTM [1]1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social-GAN [19]1.13/2.211.01/2.180.60/1.280.42/0.910.52/1.110.67/1.41
SR-LSTM [100]1.01/1.930.35/0.720.66/1.380.56/1.230.44/0.900.60/1.23
STAR-D [97]0.97/2.000.32/0.730.56/1.250.44/0.960.35/0.770.53/1.14
Trajectory++[64]1.02/2.090.33/0.630.52/1.160.42/0.940.32/0.710.52/1.11
SingularTrajectory0.72/1.230.27/0.500.57/1.120.44/0.930.35/0.730.47/0.90
+ +Table 3. Evaluation on the deterministic trajectory prediction task. + +of using the leave-one-out strategy. In the few-shot learning, they are trained using only $10\%$ of the training set, and then assessed on the entire test set. To measure the prediction performances, two common metrics are used for all the tasks, Average Displacement Error (ADE) and Final Displacement Error (FDE), measuring the average and destination error between prediction and its ground-truth path, respectively. + +# 4.2. Evaluation Results + +Stochastic prediction task. In Tab. 2, since the ETH/UCY dataset release, many state-of-the-art models have made new records. Among them, our model also achieves performance competitive with the comparison methods. Our model regards historical, social, and environmental conditions of observations and images as the most important factors for successfully generating multimodal trajectories. + +Deterministic prediction task. Next, we evaluate our model in the deterministic trajectory prediction task in Tab. 3. Our SingularTrajectory model has the best performance over both these latest state-of-the-art models, and the models optimized for deterministic prediction. In particular, our model + +![](images/41c74b1faede7cefd483bb93b6f6c2953df2b98c4bf7a1fa1443ac22303b1f93.jpg) +Figure 4. Visualization of prediction results on (a) momentary observation task and (b) few-shot task. To aid visualization, the best trajectory among $S = 20$ samples are reported. + +
MomentaryETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.24/2.230.77/1.440.45/0.810.38/0.570.35/0.580.64/1.13
PECNet [45]0.63/1.040.28/0.530.28/0.490.25/0.440.19/0.340.33/0.57
AgentFormer [98]1.10/2.110.50/1.020.52/1.100.56/1.180.43/0.890.62/1.26
MID [18]0.63/1.050.29/0.490.30/0.560.30/0.560.22/0.400.35/0.61
EigenF trajectory [6]0.46/0.760.17/0.280.25/0.440.19/0.350.15/0.270.25/0.42
STT [55]0.72/1.450.48/0.480.53/1.090.64/1.210.44/0.880.57/0.93
STT+DTO [55]0.62/1.220.29/0.560.58/1.140.45/0.980.34/0.740.46/0.93
MOE-Next [77]0.71/1.570.30/0.580.52/1.120.38/0.810.33/0.730.45/0.96
MOE-Traj++ [77]0.64/1.120.20/0.330.33/0.620.22/0.420.17/0.320.31/0.56
SingularF trajectory0.45/0.670.18/0.290.24/0.430.19/0.330.17/0.280.25/0.40
+ +exhibits significant performance improvements in the ETH and HOTEL scenes. These scenes have noisy observation paths, which have a negative impact on the prediction. We are able to achieve accurate predictions by leveraging the overall motion flow, which acts as a low-pass filter over the noisy sequence, in Singular space. + +Momentary observation task. In Tab. 4, the state-of-the-art predictors for momentary observation mainly focus on the coordinates themselves, leading to a performance drop, given only two frames. Some models for this task try to bridge these gaps, but have not been sufficient. Fortunately, with the benefit of our Singular Space, even with only a two-frame input, our model can successfully represent the overall long-term flow. Note that our SingularTrajectory with two-frame observation demonstrates the closest performance to state-of-the-art stochastic prediction models with an entire history frame; even this is achieved without any masked trajectory complement or knowledge distillation. Consequently, this allows it to accurately pinpoint the future locations of pedestrians, whose examples are displayed in Fig. 4. + +Domain adaptation task. We next evaluate the performance of SingularTrajectory in a domain adaptation task. For simplicity, the ETH, HOTEL, UNIV, ZARA1, and ZARA2 scenes are denoted as A, B, C, D, and E, respectively. For example, 'A2B' means that a model is trained on the ETH scene and tested on the HOTEL scene. As demonstrated in Tab. 1, our SingularTrajectory model shows the performance nearly equivalent to those of models specifically designed for deterministic prediction. Particularly, + +Table 4. Evaluation on the momentary observation task. + +
Few-ShotETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]0.89/1.631.22/2.480.90/1.610.68/1.251.36/2.121.01/1.82
PECNet [45]0.72/1.460.29/0.530.58/0.930.27/0.440.22/0.380.41/0.75
AgentFormer [98]1.60/2.651.02/1.641.13/1.901.19/2.011.08/1.591.20/1.96
MID [18]0.57/0.920.21/0.330.32/0.600.27/0.490.24/0.420.32/0.55
EigenF trajectory [6]0.39/0.640.13/0.210.25/0.430.21/0.390.15/0.270.23/0.39
SingularF trajectory0.35/0.460.14/0.210.26/0.440.21/0.360.18/0.310.23/0.35
+ +Table 5. Evaluation on the few-shot trajectory prediction task. + +our model produces impressive results in the challenging B2A, C2A, D2A and E2A scenarios where even models specialized in domain adaptation fail. Our SingularTrajectory framework specializes in learning general human motions, and is not limited to a specific domain, enabling accurate predictions even in extreme cases. + +Few-shot task. Finally, the results for the few-shot task are reported in Tab. 5, whose examples are in Fig. 4. As expected, our SingularTrajectory significantly outperforms the existing models. With limited data, the existing works, except the diffusion-based model, tend to easily overfit. In contrast, our methods make it possible to explicitly designate future prototype paths, while taking full advantage of the expression ability of the diffusion model. In particular, even with only $10\%$ of the data, our model achieves substantial improvements with respect to both data efficiency and performance. Figure 5 illustrates several cases where there are differences between the predictions of SingularTrajectory and other comparison methods. Previous works often show significant performance drops in other associated tasks compared to stochastic prediction tasks. In contrast, our model consistently predicts the best trajectories across multiple tasks. + +# 4.3. Ablation Studies + +The number of motion vectors $K$ . First, we conduct a component study by varying the dimension $K$ of Singular space in Tab. 6. To find the best number of singular vectors in general, we carry out experiments across all five tasks using ZARA1 scene where there are both human-environmental and human-human interactions, following [93]. As the number of motion vectors $K$ increases, more detailed movements + +![](images/a1b2d777706d42baa698fe7f459ebe68b89215dc214120d48937d2a40bc2fc2d.jpg) +Figure 5. Visualization of prediction consistency across five tasks. The more consistent the prediction is the better. + +
KDeterministicStochasticMomentaryDomainFew-shotAverage
10.52 / 1.030.30 / 0.550.31 / 0.580.51 / 1.070.34 / 0.610.40 / 0.77
20.51 / 1.010.20 / 0.320.20 / 0.330.50 / 1.060.23 / 0.370.33 / 0.62
30.45 / 0.930.19 / 0.330.20 / 0.330.47 / 1.030.22 / 0.380.31 / 0.60
40.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
50.43 / 0.940.19 / 0.330.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
60.43 / 0.940.19 / 0.320.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
+ +Table 6. Ablation study on the Singular space dimension $K$ . + +
AdoptionDeterministicStochasticMomentaryDomainFew-shotAverage
Direct0.75 / 1.470.27 / 0.480.31 / 0.540.78 / 1.600.27 / 0.480.48 / 0.91
Initial0.47 / 0.950.22 / 0.400.22 / 0.390.50 / 1.170.23 / 0.400.33 / 0.66
Residual0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
Independent0.46 / 0.940.21 / 0.390.21 / 0.390.49 / 1.160.21 / 0.390.32 / 0.65
Jointly0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
+ +are captured. In contrast, when using the smaller $K$ , it compresses the space, mainly to represent the overall motion flow. The performance tends to plateau when $K$ is larger than 3, as long as $K$ is not too small to cover most movements. We set $K = 4$ as the dimension for the Singular space because it shows the most effective prediction results across all tasks. + +Trajectory denoising methods. Next, we evaluate three types of diffusion models for the trajectory prediction tasks, as shown in Tab. 7. First, we use a basic model similar to MID [18], which directly denoises from a Gaussian noise to a trajectory. This method fails to achieve good performances. The use of an anchor as an intermediate state, similar to LED [48], which reduces the denoising steps by skipping the initial denoising steps, seems to validate its generality. This demonstrates that our adaptive anchor can function as a good initializer for the diffusion model, particularly showing nearly identical outcomes in the stochastic, momentary observation, and few-shot tasks when predicting a multimodal path. However, we confirm that our scheme, which denoises only a residual by adding perturbation to a prototype path in Fig. 6, showcases the best performance in all the tasks. Additionally, compared to refining the prototype paths independently, regarding them as a batch dimension, our model can accurately predict the future when prototype paths are jointly refined. + +![](images/df7e34619e76f5d29e9b9a15a1ddd4bea2c99f89453fd42e28d4eed41c4ec12f.jpg) +Figure 6. Visualization of anchor refinement. The denoising process progressively refines the prototype paths at each diffusion step $m$ . + +Table 7. Ablation study on the adoption of diffusion model. + +
MDeterministicStochasticMomentaryDomainFew-shotAverage
10.45 / 0.950.20 / 0.340.20 / 0.350.48 / 1.050.21 / 0.370.31 / 0.61
20.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
50.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
100.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
250.44 / 0.930.19 / 0.320.20 / 0.340.47 / 1.030.20 / 0.360.30 / 0.60
+ +Table 8. Ablation study on the diffusion steps $M$ + +Diffusion steps $M$ . Lastly, we check how many steps in the diffusion model are needed for the cascaded refinement of the adaptive anchor. In Tab. 8, we confirm that the best performance comes from $M = 10$ . This is because the DDIM scheduler accelerates its convergence and the prototype path provides a rough initial trajectory, and so fewer denoising steps are sufficient. However, as the number of denoising steps increases, the information in the initial prototype path becomes attenuated due to the noise. As a result, we observe a slight decrease in performance. + +# 5. Conclusion + +In this study, we introduce SingularTrajectory, a universal trajectory predictor model for all related trajectory prediction tasks. By unifying trajectory modalities into one Singular space, our approach standardizes trajectory data with shared motion dynamics, which eliminates the need for task-specific adjustments. The incorporation of an adaptive anchor system further personalizes the prototype paths, allowing them to interpret and adapt to environmental contexts and enhancing the reliability of the trajectory prediction. By successfully incorporating Singular space into the diffusion model, our SingularTrajectory framework successfully achieves the state-of-the-art results across five different benchmarks. This establishes SingularTrajectory itself as a general solution that covers a multitude of scenarios. + +Acknowledgement This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), Vehicles AI Convergence Research & Development Program through the National IT Industry Promotion Agency of Korea (NIPA) funded by the Ministry of Science and ICT (No.S1602-20-1001), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST). + +# References + +[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 6 +[2] Inhwan Bae and Hae-Gon Jeon. Disentangled multi-relational graph convolutional network for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2 +[3] Inhwan Bae and Hae-Gon Jeon. A set of control points conditioned pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 2 +[4] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Learning pedestrian group representations for multi-modal trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6 +[5] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6 +[6] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. EigenTrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7 +[7] Apratim Bhattacharyya, Michael Hanselmann, Mario Fritz, Bernt Schiele, and Christoph-Nikolas Straehle. Conditional flow variational autoencoders for structured sequence prediction. arXiv preprint arXiv:1908.09008, 2020. 2 +[8] Niccolò Bisagno, Bo Zhang, and Nicola Concì. Group lstm: Group trajectory prediction in crowded scenarios. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 2 +[9] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[10] Guangyi Chen, Junlong Li, Nuoxing Zhou, Liangliang Ren, and Jiwen Lu. Personalized trajectory prediction via distribution discrimination. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[11] Guangyi Chen, Zhenhao Chen, Shunxing Fan, and Kun Zhang. Unsupervised sampling promoting for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[12] Sourav Das, Guglielmo Camporese, and Lamberto Ballan. Distilling knowledge for short-to-long term trajectory prediction. arXiv preprint arXiv:2305.08553, 2023. 2 +[13] Patrick Dendorfer, Aljosa Osep, and Laura Leal-Taixe. Goalgan: Multimodal trajectory prediction based on goal position estimation. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2 + +[14] Patrick Dendorfer, Sven Elflein, and Laura Leal-Taixe. M. Gagan: A multi-generator model preventing out-of-distribution samples in pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[15] Nachiket Deo and Mohan M. Trivedi. Trajectory forecasts in unknown environments conditioned on grid-based plans. arXiv preprint arXiv:2001.00735, 2020. 2 +[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Proceedings of the Neural Information Processing Systems (NeurIPS), 2021. 2 +[17] Tharindu Fernando, Simon Denman, Sridha Sridharan, and Clinton Fookes. Soft+ hardwired attention: An lstm framework for human trajectory prediction and abnormal event detection. Neural Networks, 108:466-478, 2018. 2 +[18] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6, 7, 8 +[19] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 6 +[20] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 1, 2 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2 +[22] Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. Stgat: Modeling spatial-temporal interactions for human trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6 +[23] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. +[24] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 2, 6 +[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5 +[26] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4 +[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proceedings of + +the International Conference on Learning Representations (ICLR), 2017. 2 +[28] Vineet Kosaraju, Amir Sadeghian, Roberto Martín-Martín, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2 +[29] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4 +[30] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Muse-vae: Multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[31] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B. Choy, Philip H. S. Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[32] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Computer Graphics Forum, 26(3): 655-664, 2007. 5 +[33] Jiachen Li, Hengbo Ma, and Masayoshi Tomizuka. Conditional generative neural system for probabilistic trajectory prediction. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2 +[34] Jiachen Li, Fan Yang, Masayoshi Tomizuka, and Chiho Choi. Evolvegraph: Multi-agent trajectory prediction with dynamic relational reasoning. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2 +[35] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[36] Yuke Li. Which way are you going? imitative decision learning for path forecasting in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[37] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[38] Junwei Liang, Lu Jiang, and Alexander Hauptmann. Simaug: Learning robust representations from simulation for trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[39] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[40] Rongqin Liang, Yuanman Li, Xia Li, Yi Tang, Jiantao Zhou, and Wenbin Zou. Temporal pyramid network for pedestrian + +trajectory prediction with multi-supervision. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2 +[41] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5 +[43] Yuexin Ma, Xinge Zhu, Xinjing Cheng, Ruigang Yang, Jiming Liu, and Dinesh Manocha. Autotrajectory: Label-free trajectory extraction and prediction from videos using dynamic points. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[44] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[45] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6, 7 +[46] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 5 +[47] Huynh Manh and Gita Alaghband. Scene-lstm: A model for human trajectory prediction. arXiv preprint arXiv:1808.04018, 2018. 2 +[48] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 8 +[49] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), 2020. 2 +[50] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Multiple trajectory prediction of moving agents with memory augmented networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2 +[51] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Smemo: Social memory for trajectory forecasting. arXiv preprint arXiv:2203.12446, 2022. 2 +[52] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 1, 2 +[53] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgcnn: A social spatio-temporal + +graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 6, 7 +[54] Abduallah Mohamed, Deyao Zhu, Warren Vu, Mohamed Elhoseiny, and Christian Claudel. Social-implicit: Rethinking trajectory prediction evaluation and the effectiveness of implicit maximum likelihood estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[55] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 7 +[56] Ingrid Navarro and Jean Oh. Social-patternn: Socially-aware trajectory prediction guided by motion patterns. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 2 +[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2 +[58] Bo Pang, Tianyang Zhao, Xu Xie, and Ying Nian Wu. Trajectory prediction with latent belief energy-based model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6 +[59] Wonhui Park, Dongkwon Jin, and Chang-Su Kim. Eigencontours: Novel contour descriptors based on low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5 +[60] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2, 5 +[61] Mark Pfeiffer, Giuseppe Paolo, Hannes Sommer, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. A data-driven model for interaction-aware pedestrian motion prediction in object cluttered environments. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 2 +[62] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[63] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[64] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajec + +tory forecasting with heterogeneous data. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6 +[65] Nasim Shafiee, Taskin Padir, and Ehsan Elhamifar. Introvert: Human trajectory prediction via conditional 3d attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[66] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6 +[67] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 2 +[68] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[69] Xiaodan Shi, Xiaowei Shao, Zipei Fan, Renhe Jiang, Haoran Zhang, Zhiling Guo, Guangming Wu, Wei Yuan, and Ryosuke Shibasaki. Multimodal interaction-aware trajectory prediction in crowded space. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2 +[70] Xiaodan Shi, Xiaowei Shao, Guangming Wu, Haoran Zhang, Zhiling Guo, Renhe Jiang, and Ryosuke Shibasaki. Socialdpf: Socially acceptable distribution prediction of futures. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2 +[71] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2 +[72] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[73] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2 +[74] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[75] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[76] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[77] Jianhua Sun, Yuxuan Li, Liang Chai, Hao-Shu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with + +momentary observation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7 +[78] Jianhua Sun, Yuxuan Li, Liang Chai, and Cewu Lu. Stimulus verification is a universal and effective sampler in multi-modal human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[79] Chaofan Tao, Qinhong Jiang, and Lixin Duan. Dynamic and static context-aware LSTM for multi-agent motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[80] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1991. 4 +[81] Daksh Varshneya and G. Srinivasaraghavan. Human trajectory prediction using spatially aware deep attention models. arXiv preprint arXiv:1705.09436, 2017. 2 +[82] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2 +[83] Anirudh Vemula, Katharina Muelling, and Jean Oh. Social attention: Modeling attention in human crowds. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2 +[84] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters (RA-L), 2022. 2 +[85] Yuning Wang, Pu Zhang, Lei Bai, and Jianru Xue. Fend: A future enhanced distribution-aware contrastive learning framework for long-tail trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[86] Song Wen, Hao Wang, and Dimitris Metaxas. Social ode: Multi-agent trajectory forecasting with neural ordinary differential equations. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[87] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[88] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6 +[89] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[90] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[91] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6 +[92] Yi Xu, Jing Yang, and Shaoyi Du. Cf-lstm: Cascaded feature-based long short-term networks for predicting pedestrian trajectory. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2 +[93] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7 +[94] Hao Xue, Du Q Huynh, and Mark Reynolds. Ss-lstm: A hierarchical LSTM model for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2018. 2 +[95] Kota Yamaguchi, Alexander C Berg, Luis E Ortiz, and Tamara L Berg. Who are you with and where are you going? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2 +[96] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters (RA-L), 2021. 2 +[97] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6 +[98] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6, 7 +[99] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[100] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6 +[101] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[102] Tianyang Zhao, Yifei Xu, Mathew Monfort, Wongun Choi, Chris Baker, Yibiao Zhao, Yizhou Wang, and Ying Nian Wu. Multi-agent tensor fusion for contextual trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[103] ChunYu Zhi, HuaiJiang Sun, and Tian Xu. Adaptive trajectory prediction without catastrophic forgetting. The Journal of Supercomputing, 2023. 2 \ No newline at end of file diff --git a/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/images.zip b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aede2cd93905f94b3a7219e5884c237bb657c38c --- /dev/null +++ b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:194bff8dde81a63ab93831d98b79cb0d181a9051426f9e6894e50f420b5b6efb +size 868840 diff --git a/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/layout.json b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bf84f5c7993c3eeafb0fd130a02d125bc3d7d4e1 --- /dev/null +++ b/2024/SingularTrajectory_ Universal Trajectory Predictor Using Diffusion Model/layout.json @@ -0,0 +1,11620 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 526, + 121 + ], + "type": "text", + "content": "SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 144, + 417, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 144, + 417, + 171 + ], + "spans": [ + { + "bbox": [ + 179, + 144, + 417, + 171 + ], + "type": "text", + "content": "Inhwan Bae, Young-Jae Park and Hae-Gon Jeon* \nAI Graduate School, GIST, South Korea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 174, + 456, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 456, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 456, + 185 + ], + "type": "text", + "content": "{inhwanbae, youngjae.park}@gm.gist.ac.kr, haegonj@gist.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 239, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 597 + ], + "type": "text", + "content": "There are five types of trajectory prediction tasks: deterministic, stochastic, domain adaptation, momentary observation, and few-shot. These associated tasks are defined by various factors, such as the length of input paths, data split and pre-processing methods. Interestingly, even though they commonly take sequential coordinates of observations as input and infer future paths in the same coordinates as output, designing specialized architectures for each task is still necessary. For the other task, generality issues can lead to sub-optimal performances. In this paper, we propose SingularTrajectory, a diffusion-based universal trajectory prediction framework to reduce the performance gap across the five tasks. The core of SingularTrajectory is to unify a variety of human dynamics representations on the associated tasks. To do this, we first build a Singular space to project all types of motion patterns from each task into one embedding space. We next propose an adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods that sometimes yield unacceptable paths, our adaptive anchor enables correct anchors, which are put into a wrong location, based on a traversability map. Finally, we adopt a diffusion-based predictor to further enhance the prototype paths using a cascaded denoising process. Our unified framework ensures the generality across various benchmark settings such as input modality, and trajectory lengths. Extensive experiments on five public benchmarks demonstrate that SingularTrajectory substantially outperforms existing models, highlighting its effectiveness in estimating general dynamics of human movements. Code is publicly available at https://github.com/inhwanbae/SingularTrajectory." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 288, + 693 + ], + "type": "text", + "content": "Extensive studies of trajectory prediction methods have been conducted in the computer vision field for several decades [20, 52]. They have demonstrated its importance in various applications, including crowd simulation, social robot navi" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 545, + 300 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 300 + ], + "type": "image", + "image_path": "c6cf9b4473ac034a222fb40a9a888610f4aecf95e6ae1053e9f1b7693c80a41a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 303, + 547, + 369 + ], + "lines": [ + { + "bbox": [ + 305, + 303, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 547, + 369 + ], + "type": "text", + "content": "Figure 1. An overview of our SingularTrajectory framework. All relevant human trajectory prediction tasks can be represented in our Singular space, a unified feature embedding space for human dynamics. Using the embedding features, our diffusion-based model for a universal trajectory prediction makes prediction for all the tasks in this same space." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 389, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 547, + 472 + ], + "type": "text", + "content": "gation, obstacle avoidance and security and surveillance systems, etc. Trajectory prediction takes sequential coordinate values of agents as input and infers their future pathways in common [1, 19, 83]. Such tasks vary depending on the application. The tasks are determined with respect to the number of input/output, data processing, and the use of geological information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 548, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 548, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 548, + 641 + ], + "type": "text", + "content": "The tasks for trajectory prediction can be categorized into five groups: (1) Stochastic prediction is to predict 20 multi-modal future trajectories from one observation with 8 frames [19]. Each future trajectory consists of 12 frames. (2) Deterministic prediction takes one observation with 8 frames, but infers only one future trajectory with 12 frames [1]. (3) The momentary observation uses only two frames to predict 20 multi-modal future paths with 12 frames [77]. (4) Domain adaptation splits training data with respect to places in datasets, trains a model on one place, and then checks the transferability to other places [93]. (5) The few-shot task only uses partial data to build a dataset-efficient model [53], as illustrated in Fig. 1. Until now, each specialized architecture for a task type has provided performance gains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 700 + ], + "type": "text", + "content": "However, two questions arise. First, why do state-of-the-art models for one task undergo significant performance drops when applied to other trajectory prediction tasks? Second, is it feasible to design a general predictor that works across the five tasks?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 547, + 713 + ], + "type": "text", + "content": "As answers to these questions, we present a universal tra" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17890" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "jectionary predictor to achieve generality in predictions, named SingularTrajectory. The main idea is to unify the modalities of human dynamic representations across the five tasks. To do this, we first introduce a Singular space, an embedding space consisting of representative motion patterns for each task. The motion patterns play a role in the basis function for pedestrian movements, and are extracted using Singular Value Decomposition (SVD). They are then projected onto Singular space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 347 + ], + "type": "text", + "content": "We next propose an environment-adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods [6, 29] that sometimes fail to handle different target data distribution, our adaptive anchor is able to correct prototype paths from the adaptive anchor in Singular space if they are put into the wrong locations, based on an input traversability map. Lastly, we generate socially-acceptable future trajectories for all agents in scenes with a diffusion-based predictor which denoises residuals of perturbed prototype paths. Thanks to the cascaded denoising process of diffusion models, we refine the prototype paths. Here, historical pathways, agent interactions, and environmental information are provided as conditions to guide them in the Markov chain of the denoising diffusion processes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 347, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 289, + 396 + ], + "type": "text", + "content": "Experimental results demonstrate that our SingularTrajectory can successfully represent pedestrian motion dynamics, and significantly improve prediction accuracy for the five tasks on challenging public benchmark datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 405, + 139, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 139, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 139, + 417 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 426, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 449 + ], + "type": "text", + "content": "We review previous studies on trajectory prediction that have attempted to address various benchmark scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 456, + 223, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 223, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 223, + 468 + ], + "type": "text", + "content": "2.1. Pedestrian Trajectory Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": "Ways of predicting pedestrian future trajectories have been studied for a long time in the computer vision field. Pioneering works [20, 52, 60, 95] model an invisible social norm based on motion dynamics as an energy minimization problem. Introducing recurrent neural networks [1, 19, 64, 97] achieves significant improvements by providing highly sequential representations of high-level shape of paths. These methods determine the most probable path, called deterministic trajectory prediction. The following works model mutual influences among agents using attention mechanisms [17, 23, 64, 83], graph convolutional networks [2, 9, 27, 41, 53, 54, 75], graph attention networks [5, 22, 28, 38, 39, 66, 82], and transformers [4, 18, 55, 68, 86, 87, 97, 98]. Additional visual information allows us to leverage environmental constraints from traversability maps [13-15, 28, 37, 46, 47, 49, 50, 63, 65, 74, 75, 79, 81, 85, 94, 99, 102]. Depending on the constraints, predictors take either recurrent [1, 8, 10, 11, 18, 19, 30, 43, 44, 51, 56, 61, 64, 92, 100, 101] or simultaneous approaches [2, 3, 35, 53, 66, 67, 89] to extrapolate the future pathways." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 548, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 548, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 548, + 337 + ], + "type": "text", + "content": "Meanwhile, with the success of generative models, the importance of multimodality has begun to emerge, called stochastic trajectory prediction. Stochastic prediction enables us to consider all of an agent's possible future pathways. For example, an agent at a crossroads may either walk straight or turn left/right. Here, the stochastic prediction infers all potential future modes. This approach has become mainstream in this field. Starting from Social-GAN [19], a bivariate Gaussian distribution [1, 2, 34, 53, 66, 69, 70, 93, 96, 97], generative adversarial network [14, 19, 22, 28, 36, 40, 63, 74, 78, 102], and conditional variational autoEncoder [7, 10, 23, 30, 31, 33, 45, 64, 76, 84, 88, 91] have been adopted for stochastic trajectory prediction. Anchorm-conditioned methods can explicitly represent different modalities by prototyping possible paths [6, 29]. Most recently, diffusion-based models have revealed their tremendous representation capacities in numerous tasks [16, 21, 57, 71-73], proving its potential for stochastic trajectory prediction [18, 25, 48, 62]. In this study, we take full advantage of both the anchor-conditioned approach and the diffusion-based model to achieve the explainability and generalizability of the trajectory prediction tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 342, + 499, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 499, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 499, + 355 + ], + "type": "text", + "content": "2.2. Various Trajectory Prediction Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 547, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 516 + ], + "type": "text", + "content": "Beyond the standard benchmark protocol of stochastic prediction, there are three other variants of this task: momentary observation, domain adaptation, and few-shot learning. Works in [12, 55, 77] only take two frames as input for the momentary trajectory prediction. Multi-task learning, self-supervised learning, and knowledge distillation techniques have been used to extract rich features from the limited data. Another works [24, 93, 103] focuses on domain adaptation across trajectory domains, captured from different surveillance views. A transferable graph neural network is introduced to adaptively learn domain-invariant knowledge. The others [53, 101] adopt few-shot learning for better training efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 516, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 612 + ], + "type": "text", + "content": "Although these works take and infer the sequential coordinates of agent trajectories in common, there is no unified model for all the tasks. Despite the tremendous efforts to design a specialized architecture for one task, they cannot be applied to the other tasks without suffering significant performance drops. In the next section, we will describe how to design a unified architecture, which consistently produces promising results on the five associated tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 621, + 389, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 389, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 389, + 635 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "We describe how to learn a general representation of human motions. We first define a general trajectory prediction problem in Sec. 3.1 and provide preliminaries of explicit formulations on the SVD and diffusion process in Sec. 3.2. We then introduce a motion feature extraction to build our Singular space, and a projection of any trajectory from each task onto" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17891" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "it in Sec. 3.3. Next, we propose an environment-adaptive anchor using motion vectors and input image in Sec. 3.4. Finally, we present the SingularTrajectory predictor based on the diffusion model in Sec. 3.5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 126, + 160, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 126, + 160, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 126, + 160, + 138 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": "Trajectory prediction aims to predict the future paths of agents based on their historical path and surrounding contexts. Suppose that at each timestamp " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": ", there are " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " pedestrians in a scene with the 2D spatial coordinate position " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_t^n \\in \\mathbb{R}^2 | n \\in [1, \\dots, N]\\}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": ". A pedestrian historical trajectory " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{X}_n" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "T_{hist}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " timesteps can be represented as the cumulative coordinates " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{X}_n = \\{\\pmb{p}_t^n | t \\in [1, \\dots, T_{hist}]\\}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": ". Similarly, future trajectories " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{Y}_n" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " for the time duration " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "T_{fut}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " to be predicted can be written as " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "Y_n = \\{\\pmb{p}_t^n | t \\in [T_{hist} + 1, \\dots, T_{hist} + T_{fut}]\\}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": ". The prediction system takes both the historical trajectories for all " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " people " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "X = \\{X_n | n \\in [1, \\dots, N]\\}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " and the scene image map " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " for environmental information as input. The deterministic prediction system infers one sequence of the most reliable future trajectory " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\widehat{Y} = \\{\\widehat{Y}_n | n \\in [1, \\dots, N]\\}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": ". For the stochastic prediction, because of the indeterminacy of the future movements, " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " multiple pathways for all the " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " pedestrians " + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\widehat{Y} = \\{\\widehat{Y}_n^s | n \\in [1, \\dots, N], s \\in [1, \\dots, S]\\}" + }, + { + "bbox": [ + 46, + 144, + 289, + 361 + ], + "type": "text", + "content": " are generated so that at least one sample is close to the ground-truth trajectory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 367, + 133, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 133, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 133, + 379 + ], + "type": "text", + "content": "3.2. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": "Singular Value Decomposition. Singular Value Decomposition (SVD) decomposes a matrix into three resultant matrices. Given a matrix " + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 385, + 288, + 422 + ], + "type": "text", + "content": ", its SVD is represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 427, + 287, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 427, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 135, + 427, + 287, + 442 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} = \\boldsymbol {U} \\boldsymbol {\\Sigma} \\boldsymbol {V} ^ {\\top}, \\tag {1}", + "image_path": "9929f6ebb670fa42c962984b7684b551e90fdd1621f04c2f1be472a678cb4e79.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": " is an orthogonal left singular vector matrix, whose columns are eigenvectors of " + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "inline_equation", + "content": "A\\times A^{\\top}" + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": " is a diagonal matrix with the singular values of " + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": ", consisting of " + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": " non-negative values in descending order. " + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": " is a right singular vector matrix, which is also orthogonal and its columns are the eigenvectors of " + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "inline_equation", + "content": "A^{\\top}\\times A" + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": "To remove the redundant part of the raw data, the truncation technique is often applied to the results after the decomposition. The idea behind the truncated SVD is to approximate the original matrix " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " with the lower rank. With the " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " to determine the number of singular values to retain, " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " can be simplified to " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{K}" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " which contains only the " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " largest singular values. Similarly, " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "V^{\\top}" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " are reduced to " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "U_{K}" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "V_{K}^{\\top}" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " by keeping the first " + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 521, + 288, + 688 + ], + "type": "text", + "content": " columns and rows, respectively. This process eliminates the smallest singular values, which are not needed to express the original data and often correspond to noise or redundant information. This is useful for practical scenarios dealing with large and potentially sparse matrices because we can reconstruct a close approximation of the original data with significantly less storage space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "type": "text", + "content": "Diffusion models. The diffusion model operates by transforming a noisy distribution, represented by the noise vector" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_M" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ", into the desired data " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " through a series of " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " diffusion steps. These steps involve intermediate latent variables " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_m|m\\in [1,\\dots ,M]\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ", and encompass both the diffusion and denoising processes. The diffusion process adds a small amount of noise to data in order to obtain the standard normal distribution " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_M)" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " from the distribution " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_0)" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " using the Markov chain as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 162, + 545, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 351, + 162, + 545, + 194 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {y} _ {1: M} \\mid \\mathbf {y} _ {0}\\right) := \\prod_ {m = 1} ^ {M} q \\left(\\mathbf {y} _ {m} \\mid \\mathbf {y} _ {m - 1}\\right) \\tag {2}", + "image_path": "4f896439298ad4cc309057240437e9ef594b9194bac0448ddb782e2c7a7db657.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 327, + 195, + 522, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 195, + 522, + 211 + ], + "spans": [ + { + "bbox": [ + 327, + 195, + 522, + 211 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {y} _ {m} \\mid \\mathbf {y} _ {m - 1}\\right) := \\mathcal {N} \\left(\\mathbf {y} _ {m}; \\sqrt {1 - \\beta_ {m}} \\mathbf {y} _ {m - 1}, \\beta_ {m} \\mathbf {I}\\right)", + "image_path": "744b664159ae2085665f1c95d4d89b69b63e31e4f6c4205c73e91a5196a470fd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "text", + "content": " is a small positive constant and a variance schedule for adding noise. The denoising process uses the " + }, + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_m" + }, + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "text", + "content": " to recover " + }, + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 304, + 217, + 545, + 253 + ], + "type": "text", + "content": " with a learnable network as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 261, + 545, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 341, + 261, + 545, + 293 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {y} _ {0: M}\\right) := p \\left(\\mathbf {y} _ {M}\\right) \\prod_ {m = 1} ^ {M} p _ {\\theta} \\left(\\mathbf {y} _ {m - 1} \\mid \\mathbf {y} _ {m}\\right), \\tag {3}", + "image_path": "052091aa5a8c3f1a750c108d360056384550114d95883882fb38d6557bf2f8fd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 294, + 521, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 294, + 521, + 308 + ], + "spans": [ + { + "bbox": [ + 330, + 294, + 521, + 308 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\mathbf {y} _ {m - 1} | \\mathbf {y} _ {m}) := \\mathcal {N} (\\mathbf {y} _ {m - 1}; \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {y} _ {m}, m), \\beta_ {m} \\mathbf {I}).", + "image_path": "dd50c639c53a0fddddce06102b19ffb374e962b66cd0bef59ca32a1adfbdebd9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_M\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": " is an initial noise sampled from the Gaussian distribution " + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}_M)" + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": " denotes the learnable parameter of the diffusion model. " + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": " is a learnable denoising model of a clean data " + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": ", and a corrupted data " + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_m" + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": " at a step " + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": ". The objective is to train the neural network so that the denoising process predicts the true data-generating distribution well. This is often done by maximizing the evidence lower bound, ensuring that the samples generated by the diffusion model are indistinguishable from the real data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 430, + 454, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 454, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 454, + 443 + ], + "type": "text", + "content": "3.3. Unifying the Motion Space" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 449, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 556 + ], + "type": "text", + "content": "The trajectory prediction model uses a learnable network to capture the relationship in consideration of the input coordinates of pedestrians, input images and output coordinates for each pedestrian. Since expectations for input and output trajectories are different for each associated task (e.g., length and multimodality), they should be viewed as different data spaces even if they use the same coordinate systems. We introduce a method to merge raw data in each space into a Singular space for human motion dynamics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 641 + ], + "type": "text", + "content": "Singular space construction. To discover motion dynamics from the raw data, we first extract primitive motion features. Inspired by the successful low-rank approximation of raw trajectory data using eigenvectors in [6], we also employ a similar strategy using singular vectors from the truncated SVD to extract principal motion components from the entire training dataset." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "First, we cut-off paths of all pedestrians in the dataset into " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "T_{win}" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " lengths through a sliding window to create a total of " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " gist of trajectory set " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{L \\times (2 \\times T_{win})}" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " is a temporary matrix to extract a motion vector from a set of trajectory in the initialization phase. Next, we decompose " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " to obtain truncated matrices " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "U_K \\in \\mathbb{R}^{L \\times K}" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_K \\in \\mathbb{R}^{K \\times K}" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17892" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 113, + 135 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 113, + 135 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 113, + 135 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 113, + 135 + ], + "type": "image", + "image_path": "ab4c686765e04400c9a24dc333bb76be181683d3ab945f60b722742b7b8d2dc7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 138, + 287, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 138, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 287, + 215 + ], + "type": "text", + "content": "Figure 2. Visualization of the trajectories in Singular space. (a) The circle and triangle markers indicate the history and future trajectory coordinates, respectively. Each color also refers to each associated task. (b-d) Each marker in the Singular space corresponds to each trajectory in raw data. And, the slicing planes, representing a set of arrows, mean human dynamics of straight forward motions and turning motions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 116, + 72, + 170, + 136 + ], + "blocks": [ + { + "bbox": [ + 116, + 72, + 170, + 136 + ], + "lines": [ + { + "bbox": [ + 116, + 72, + 170, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 72, + 170, + 136 + ], + "type": "image", + "image_path": "6a07dc740514b4b0cc3fe84a0295dacfd257db2a5019d448d04a1fd5208c1175.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 228, + 136 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 228, + 136 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 228, + 136 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 228, + 136 + ], + "type": "image", + "image_path": "897df13df0786f68e036ad5758762e54972a06bc1b36f6c231928b4a098606df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 72, + 287, + 136 + ], + "blocks": [ + { + "bbox": [ + 233, + 72, + 287, + 136 + ], + "lines": [ + { + "bbox": [ + 233, + 72, + 287, + 136 + ], + "spans": [ + { + "bbox": [ + 233, + 72, + 287, + 136 + ], + "type": "image", + "image_path": "9f800c2596ad3f2d2c80ab498aa9a7be8a1395a9e5b21235f181adb944a62389.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 144, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 547, + 221 + ], + "type": "text", + "content": "Figure 3. An example of the adaptive anchor generation. (a) The initial prototype anchor " + }, + { + "bbox": [ + 305, + 144, + 547, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 305, + 144, + 547, + 221 + ], + "type": "text", + "content": " is placed on the last observation coordinate of a person. In this instance, four prototype paths (highlighted in red) are incorrectly placed at the non-traversable locations. (b) Vector field " + }, + { + "bbox": [ + 305, + 144, + 547, + 221 + ], + "type": "inline_equation", + "content": "\\vec{F}_I(x,y)" + }, + { + "bbox": [ + 305, + 144, + 547, + 221 + ], + "type": "text", + "content": " is computed to guide toward in the nearest traversable areas. (c) The initial prototype paths are then tailored to the environment using the vector field." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "V_{K}^{\\top}\\in \\mathbb{R}^{K\\times (2\\times T_{win})}" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "V_{K}^{\\top}" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " is a set of " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " spatiotemporal motion vectors " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_k\\in \\mathbb{R}^{2\\times T_{win}}" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " representing the most dominant motion dynamics of pedestrians. Since " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_k" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " are orthogonal to each other, we define a Singular space coordinate system using the " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " singular vectors as basis vectors for each axis. In Singular space, trajectories " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\pmb{A}" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " can be a coordinate of " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = U_K\\cdot \\pmb {\\Sigma}_K\\in \\mathbb{R}^{L\\times K}" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": ", indicating that each motion vector has an influence on reconstructing the " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " trajectories. In other words, we can project the " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\pmb{A}" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " into the coordinate in Singular space " + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 234, + 289, + 356 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 362, + 287, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 362, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 101, + 362, + 287, + 377 + ], + "type": "interline_equation", + "content": "\\mathcal {A} = \\boldsymbol {A} \\times \\left(\\boldsymbol {V} _ {K} ^ {\\top}\\right) ^ {- 1} = \\boldsymbol {A} \\times \\boldsymbol {V} _ {K}, \\tag {4}", + "image_path": "d57e6d125e2b89741962d58decced0ad83ccfd550765c8e50ff3ab77a046c391.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 384, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 384, + 287, + 409 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_K^\\top)^{-1}" + }, + { + "bbox": [ + 46, + 384, + 287, + 409 + ], + "type": "text", + "content": " can be simplified into " + }, + { + "bbox": [ + 46, + 384, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_K" + }, + { + "bbox": [ + 46, + 384, + 287, + 409 + ], + "type": "text", + "content": " with the property of an orthogonal matrix1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 410, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 564 + ], + "type": "text", + "content": "In the Singular space, we can now concentrate on the motion flow by using the coefficients of motion patterns, instead of considering every position over time. Note that the results from the SVD depend highly on various factors such as the window size for inputs/outputs, data processing, and the differences in the raw data caused by geological variations. Due to the issue, works using SVD only convert the output space [26, 59, 80] or conduct separate decompositions for each of input and output data spaces [6]. However, in this case, they cause inconsistent representations of motion dynamics, even for the same pedestrian, which leads to a lack of generality. In the next step, we introduce a new method to address this issue." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": "Projection of any trajectory into Singular space. We aim to express the input/output trajectories " + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": " of the five associated tasks in Singular space all at once. Our core idea stems from the notion that all pedestrians share the same human motion dynamics and thus will likely show a similar pattern. Starting with the projection function Eq. (4), which is a projection matrix for the fixed trajectory length " + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "inline_equation", + "content": "T_{win}" + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": ", we extend it to any length " + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "inline_equation", + "content": "T_{hist}" + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "inline_equation", + "content": "T_{fut}" + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": ", which varies from task to task. To handle the motion patterns regardless of their lengths, we interpolate " + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_k \\in \\mathbb{R}^{2 \\times T_{win}}" + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_{\\mathrm{x},k} \\in \\mathbb{R}^{2 \\times T_{hist}}" + }, + { + "bbox": [ + 46, + 566, + 288, + 686 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 385, + 133 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 385, + 133 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 385, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 385, + 133 + ], + "type": "image", + "image_path": "daba69e12be8e95b117f5a30ce2d5f08c28f51143f2fa1557b9c7bdca9a2c8ec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 133, + 382, + 141 + ], + "lines": [ + { + "bbox": [ + 310, + 133, + 382, + 141 + ], + "spans": [ + { + "bbox": [ + 310, + 133, + 382, + 141 + ], + "type": "text", + "content": "(a) Initial prototype paths" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 387, + 70, + 465, + 133 + ], + "blocks": [ + { + "bbox": [ + 387, + 70, + 465, + 133 + ], + "lines": [ + { + "bbox": [ + 387, + 70, + 465, + 133 + ], + "spans": [ + { + "bbox": [ + 387, + 70, + 465, + 133 + ], + "type": "image", + "image_path": "70ab6f266e750f341fc8ab0e1381dafd4e118302daabb102f0be48a0ca6f9bb8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 133, + 446, + 140 + ], + "lines": [ + { + "bbox": [ + 403, + 133, + 446, + 140 + ], + "spans": [ + { + "bbox": [ + 403, + 133, + 446, + 140 + ], + "type": "text", + "content": "(b) Vector field" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 467, + 70, + 546, + 133 + ], + "blocks": [ + { + "bbox": [ + 467, + 70, + 546, + 133 + ], + "lines": [ + { + "bbox": [ + 467, + 70, + 546, + 133 + ], + "spans": [ + { + "bbox": [ + 467, + 70, + 546, + 133 + ], + "type": "image", + "image_path": "d2215bdab00f8ff960015d0cfb9cfc00536b11e95ed457dd3130f3343b5f196b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 133, + 541, + 141 + ], + "lines": [ + { + "bbox": [ + 471, + 133, + 541, + 141 + ], + "spans": [ + { + "bbox": [ + 471, + 133, + 541, + 141 + ], + "type": "text", + "content": "(c) Final prototype paths" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "content": "Since the motion pattern can be seen as a 2-dimensional curve, we use Cardinal B-splines to make a transformation matrix " + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "inline_equation", + "content": "C_{T_{hist}} \\in \\mathbb{R}^{(2 \\times T_{hist}) \\times (2 \\times T_{win})}" + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "content": " using the Irwin-Hall distribution. The constant value " + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "content": " depends only on the length, so " + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_{\\mathrm{x},k}" + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "content": " can be approximated. The trajectory " + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times (2 \\times T_{hist})}" + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "content": " is projected to the coordinate " + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^{N \\times (2 \\times T_{hist})}" + }, + { + "bbox": [ + 304, + 242, + 547, + 326 + ], + "type": "text", + "content": " in Singular space as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 336, + 545, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 336, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 386, + 336, + 545, + 349 + ], + "type": "interline_equation", + "content": "v _ {x, k} = C _ {T _ {\\text {h i s t}}} \\times v _ {k}. \\tag {5}", + "image_path": "68f4836b26391c3cd8fafd435b7a771dbb4001db03edcc00bf1a45a83680b000.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 376, + 356, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 356, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 376, + 356, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {X} = \\boldsymbol {X} \\times \\boldsymbol {C} _ {T _ {\\text {h i s t}}} \\times \\boldsymbol {V} _ {K}. \\tag {6}", + "image_path": "6c5bef99387782ff67e293a19877e527a8082a57e8ca16fd7fe6810c1269510c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "text", + "content": "In the same way as Eq. (6), the trajectory " + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{N \\times (2 \\times T_{\\text{fit}})}" + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "text", + "content": " also can be projected to " + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{N \\times K}" + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_K" + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "inline_equation", + "content": "C_{T_{\\text{fit}}}" + }, + { + "bbox": [ + 304, + 373, + 545, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 399, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 529 + ], + "type": "text", + "content": "Through this process, Singular space can represent a variety of trajectories even with different lengths, including motion vectors for the input/output as well as the task-specific data, as visualized in Fig. 2. Establishing a common ground for the trajectory representation is crucial for our model's adaptability and robust performances across different trajectory prediction benchmarks. Moreover, the strategy of focusing on the overall motion flow, rather than frame-by-frame coordinates, further improves the model's capacity to understand and predict socially compliant trajectories in diverse real-world scenarios." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 538, + 409, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 409, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 409, + 551 + ], + "type": "text", + "content": "3.4. Adaptive Anchor" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "type": "text", + "content": "Beyond the trajectory data integration, we also introduce how to incorporate environmental contexts into trajectory predictors. To do this, we propose an adaptive anchor which is expressed as a set of motion vectors from input images and consists of prototype paths in Singular space." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Prototype anchor formation. Previous anchor-based human trajectory prediction approaches [6, 29] use a fixed anchor for any pedestrian, and the anchor is refined for output trajectories. Although these methods can explicitly model the multimodality, they sometimes fail to handle the case where the prototype path is put into wrong locations, or blocked by static obstacles. This is because the wrong prototypes are treated as a hard constraint. To avoid this problem," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "1For better understanding, we display the coordinate variable in the Singular space using the calligraphic font" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17893" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "we use the input image as a traversability map to correct the wrong prototype paths." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "content": "We start with the converted future trajectory " + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "content": " in the whole dataset " + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "content": " using Sec. 3.3. Following [6, 59], the coordinates " + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "content": " are then clustered into " + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "content": " centroids, which can be viewed as groups of components representing different multimodal futures. In addition, each centroid in Singular space is used as an initial prototype path " + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s" + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "content": " for constructing an anchor " + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{S \\times K}" + }, + { + "bbox": [ + 46, + 96, + 287, + 203 + ], + "type": "text", + "content": ". But, these initial prototype paths are still fixed, so they still have a limitation which is unlikely to consider environmental information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "text", + "content": "Adaptive anchor generation. To make use of environmental information, we introduce a module to deform the anchor. Using an off-the-shelf semantic segmentation model from [46], we can obtain binary traversable maps " + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text{map}}" + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "text", + "content": " from an input image. We then derive a vector field " + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "inline_equation", + "content": "\\vec{F}_I(x,y)" + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "text", + "content": " to fix the wrongly located prototype path by directing it to a nearest traversable regions. If the initial prototype path " + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s" + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "text", + "content": " is in the wrong place, we deform it by adding the vector fields into the initial anchor until they reach equilibrium states to obtain the final prototype paths " + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s'" + }, + { + "bbox": [ + 46, + 206, + 288, + 327 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 337, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 337, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 107, + 337, + 287, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {s} ^ {\\prime} = \\mathcal {P} _ {s} + \\vec {F} _ {I} \\left(\\mathcal {P} _ {s} \\boldsymbol {V} _ {K} ^ {\\top} \\boldsymbol {C} _ {T _ {f u t}} ^ {- 1}\\right). \\tag {7}", + "image_path": "59a8926449b996f5205a090c23fad64bd56a6d6afe6f1a53276324a3c61986e2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 409 + ], + "type": "text", + "content": "Through this process, the prototype path is re-located toward the nearest walkway, as demonstrated in Fig. 3. In other words, the anchor plays a role in the environment-adaptive prototype paths, unlike the existing fixed anchor methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 410, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 506 + ], + "type": "text", + "content": "Because the prototype paths are in Singular space, the scene image can be implicitly represented as the adaptive anchor. In addition, the scene image can be approximated with the set of motions. By projecting the scene images into the Singular space, our model understands the surrounding environment. This holistic approach, integrating both coordinate and environmental cues, sets the stage for more realistic and reliable trajectory predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 514, + 270, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 270, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 270, + 527 + ], + "type": "text", + "content": "3.5. Diffusion-Based SingularTrajectory Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": "As the final step, we develop a framework called Singular-Trajectory, a unified model that works well across the five tasks. Leveraging the Singular space and the adaptive anchor, our diffusion-based SingularTrajectory model can precisely forecast potential future paths." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Denoising a perturbed trajectory anchor. Unlike previous diffusion-based predictors, which directly forecast the future path from Gaussian noise [18, 25], we devise a stepwise refinement from the adaptive anchor for a realistic trajectory. Here, historical pathways " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", environmental information " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", and agent interactions " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathcal{X},\\mathcal{P})" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " are encoded and used as conditions to guide the denoising processes. To better capture agents interactions " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " in the denoising process, we adopt the transformer model. Similar to other transformer-based trajectory predictors [18, 48], our SingularTrajectory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "encodes spatio-temporal information to account for agent-agent and agent-environment interactions using " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": ", respectively. These conditions are then concatenated into one feature vector representation and fed into the diffusion model " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{y}_m, m, \\mathcal{X}, \\mathcal{P}, \\mathcal{G})" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": " to learn, by contrasting the motion patterns from previous diffusion steps to distinguish the added noise, and to close the reality gap by generating socially-acceptable future paths " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 169, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 169, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 169, + 545, + 194 + ], + "type": "text", + "content": "This refinement process works in a cascading manner by " + }, + { + "bbox": [ + 306, + 169, + 545, + 194 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P},\\dots,\\widehat{\\mathcal{V}}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 306, + 169, + 545, + 194 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 195, + 545, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 350, + 195, + 545, + 227 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {Y}} = \\mathcal {P} + p \\left(\\mathbf {y} _ {M}\\right) \\prod_ {m = 1} ^ {M} p _ {\\theta} \\left(\\mathbf {y} _ {m - 1} \\mid \\mathbf {y} _ {m}\\right). \\tag {8}", + "image_path": "624bcdf6ea34c430b8d07c156045f7100abc1478c2261fcdfc7481dbcb5a73f0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "content": "By predicting only the residuals " + }, + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_m\\in \\mathbb{R}^{S\\times K}" + }, + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "content": " to adjust the anchors, the problem is simplified, in that the model is able to use a prior knowledge on the initial state. With this process, we thus ensure more precise and reliable trajectory generations, achieving generality in various environments and applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "text", + "content": "Implementation details. To construct Singular space, we empirically set " + }, + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "text", + "content": " to 4. We set " + }, + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "inline_equation", + "content": "T_{win} = T_{fut} = 12" + }, + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "text", + "content": " in order to prevent information loss due to the approximation of the motion vector during prediction. For an anchor diffusion model, we devise a one-layer transformer for encoding motion and context information, where the dimension is set to 256 with 4-head attention. We set " + }, + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 304, + 303, + 547, + 483 + ], + "type": "text", + "content": " and schedule the diffusion timesteps following DDIM [72]. To train the SingularTrajectory in an end-to-end manner, we use a mean square error (MSE) as a loss function between the output and a random Gaussian noise for the current iteration. The training is performed with AdamW optimizer [42], with a learning rate of 0.001 and batch size of 512 for 256 epochs. All the experiments are conducted on a single NVIDIA A6000 GPU, which usually takes about an hour to train each scene." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 496, + 388, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 388, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 388, + 509 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 612 + ], + "type": "text", + "content": "In this section, we conduct comprehensive experiments to verify the generality of our SingularTrajectory model for the trajectory prediction tasks. We first describe our experimental setup in Sec. 4.1. We then provide comparison results with state-of-the-art models on public benchmark datasets in Sec. 4.2. Finally, we perform an extensive ablation study demonstrating the effect of each component of our method in Sec. 4.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 622, + 422, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 422, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 422, + 635 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Datasets. To compare our SingularTrajectory with state-of-the-art baselines, we conduct quantitative evaluations on two common datasets, ETH [60] and UCY [32], for all the associated tasks. The ETH and UCY datasets consist of various motions of 1,536 pedestrians across five unique scenes: ETH, Hotel, Univ, Zara1 and Zara2. They are recorded in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17894" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 544, + 272 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 544, + 272 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 544, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 544, + 272 + ], + "type": "table", + "html": "
Domain Adaptation (ADE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAVG
STGCNN [53]1.831.581.301.311.513.021.382.631.582.151.160.700.820.540.811.041.050.730.470.820.981.090.740.500.831.22
PECNet [45]1.971.681.241.351.563.111.352.691.622.191.390.820.930.570.931.101.170.920.520.931.011.250.830.610.931.31
AgentFormer [98]1.731.641.871.571.701.491.762.131.731.781.730.192.011.791.431.712.051.921.321.751.961.911.891.911.921.72
MID [18]1.261.431.651.431.441.230.870.980.770.961.610.630.810.580.911.420.750.810.740.931.470.680.880.910.981.04
EigenTrajectory [6]0.390.710.730.450.570.950.680.510.420.640.940.320.490.370.531.101.031.360.430.980.920.510.620.520.640.67
T-GNN [93]1.131.250.941.031.092.541.082.251.411.820.970.540.610.230.590.880.780.590.320.640.870.720.650.340.650.96
K0 [24]0.450.780.680.590.630.870.650.490.360.591.080.460.440.360.591.130.560.590.530.701.130.390.590.480.650.63
Adaptive [24]0.360.640.530.450.500.900.620.500.360.601.110.460.480.420.621.110.520.540.470.661.050.340.560.440.600.59
SingularF trajectory0.290.590.510.420.450.660.550.450.350.500.790.300.480.360.480.730.290.550.360.490.640.230.550.450.470.48
Domain Adaptation (FDE)A2*B2*C2*D2*E2**2*
A2BA2CA2DA2EAVGB2AB2CB2DB2EAVGC2AC2BC2DC2EAVGD2AD2BD2CD2EAVGE2AE2BE2CE2DAVGAvgV
STGCNN [53]3.242.862.532.432.775.162.514.862.883.852.301.341.741.101.622.211.991.410.881.622.102.051.471.011.662.30
PECNet [45]3.332.832.532.452.795.232.484.902.863.872.221.321.681.121.592.202.051.520.881.662.101.841.450.981.592.29
AgentFormer [98]3.603.544.523.953.903.763.675.173.424.013.773.494.123.763.793.883.754.273.613.883.823.553.744.323.863.89
MID [18]2.422.773.222.772.802.341.661.871.441.833.131.281.741.241.852.731.491.671.541.862.871.411.881.932.022.07
EigenTrajectory [6]0.761.431.560.961.182.101.381.070.861.352.010.641.100.821.142.421.031.360.971.442.040.511.291.111.241.27
T-GNN [93]2.182.251.781.842.014.151.824.042.533.141.911.121.300.871.301.921.461.250.651.321.861.451.280.721.331.82
K0 [24]0.941.721.511.291.371.811.331.020.751.232.170.910.940.811.212.180.981.201.031.352.120.681.221.031.261.28
Adaptive [24]0.691.341.120.911.021.851.271.050.761.232.260.921.030.941.292.231.021.171.001.362.070.631.190.941.211.22
SingularF trajectory0.571.191.080.810.911.161.170.970.731.011.360.621.110.790.971.220.521.180.730.921.070.431.150.940.900.94
", + "image_path": "bf7eb52475bd8cc9e1048e212b59dde79bc1e7613522c0ae027bb6ae0da8a734.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 308, + 285, + 453 + ], + "blocks": [ + { + "bbox": [ + 47, + 273, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 47, + 273, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 546, + 297 + ], + "type": "text", + "content": "Table 1. Comparison of our SingularTrajectory with other state-of-the-art methods in domain adaptation task. Following [24, 93], we deterministically predict one future path for all dataset pairs (ADE/FDE). Bold: Best, Underline: Second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 308, + 285, + 453 + ], + "lines": [ + { + "bbox": [ + 48, + 308, + 285, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 285, + 453 + ], + "type": "table", + "html": "
StochasticETHHOTELUNIVZARA1ZARA2AVG
Social-GAN [19]0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
STGCNN [53]0.64/1.110.49/0.850.44/0.790.34/0.530.30/0.480.44/0.75
STAR [97]0.57/1.110.19/0.370.35/0.750.26/0.570.25/0.580.33/0.68
PECNet [45]0.61/1.070.22/0.390.34/0.560.25/0.450.19/0.330.32/0.56
MID [18]0.57/0.930.21/0.330.29/0.550.28/0.500.20/0.370.31/0.54
LBEBM [58]0.60/1.060.21/0.380.28/0.540.21/0.390.15/0.300.29/0.53
Trajectory++[64]0.61/1.030.20/0.280.30/0.550.24/0.410.18/0.320.31/0.52
GroupNet [88]0.46/0.730.15/0.250.26/0.490.21/0.390.17/0.330.25/0.44
AgentFormer [98]0.46/0.800.14/0.220.25/0.450.18/0.300.14/0.240.23/0.40
GP-Graph [4]0.43/0.630.18/0.300.24/0.420.17/0.310.15/0.290.23/0.39
NPSN [5]0.36/0.590.16/0.250.23/0.390.18/0.320.14/0.250.21/0.36
EqMotion [90]0.40/0.610.12/0.180.23/0.430.18/0.320.13/0.230.21/0.35
EigenIProjection [6]0.36/0.530.12/0.190.24/0.430.19/0.330.14/0.240.21/0.34
SocialVAE [91]0.41/0.580.13/0.190.21/0.360.17/0.290.13/0.220.21/0.33
LED [48]0.39/0.580.11/0.170.26/0.430.18/0.260.13/0.220.21/0.33
SingularProjection0.35/0.420.13/0.190.25/0.440.19/0.320.15/0.250.21/0.32
", + "image_path": "88a58d0d1d2c39bcb78dd3cf5d0724d68cd52ec4ba9a7428664403b183df1baf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 52, + 455, + 281, + 465 + ], + "lines": [ + { + "bbox": [ + 52, + 455, + 281, + 465 + ], + "spans": [ + { + "bbox": [ + 52, + 455, + 281, + 465 + ], + "type": "text", + "content": "Table 2. Evaluation on the stochastic trajectory prediction task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 494, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 287, + 517 + ], + "type": "text", + "content": "surveillance views, and the trajectories are labeled in world coordinates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "content": "Benchmarks. We evaluate our model for the five trajectory prediction tasks: (1) stochastic prediction, (2) deterministic prediction, (3) momentary observation, (4) domain adaptation, and (5) few-shot learning. The stochastic trajectory prediction employs a leave-one-out strategy [5, 19, 22, 53, 66] for training and inference across the five ETH-UCY scenes. In the scenes, the initial 3.2 seconds (equivalent to " + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "T_{hist} = 8" + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "content": " frames) of all paths are used as input, and the subsequent 4.8 seconds (corresponding to " + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "T_{fut} = 12" + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "content": " frames) are predicted. The model generates " + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "S = 20" + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "content": " paths, and one with the lowest error is chosen as the final prediction. In the deterministic prediction, the trajectory prediction methods generate only " + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "S = 1" + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "content": " path. For the momentary observation, the observation length is " + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "T_{hist} = 2" + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "content": " whereas still " + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "T_{fut} = 12" + }, + { + "bbox": [ + 47, + 522, + 289, + 713 + ], + "type": "text", + "content": ". For the domain adaptation, models are trained on each of the five scenes individually, and then evaluated on the other scenes, instead" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 308, + 544, + 415 + ], + "blocks": [ + { + "bbox": [ + 307, + 308, + 544, + 415 + ], + "lines": [ + { + "bbox": [ + 307, + 308, + 544, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 544, + 415 + ], + "type": "table", + "html": "
DeterministicETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.04/1.930.88/1.630.71/1.340.66/1.220.57/1.060.77/1.44
PECNet [45]1.20/2.730.68/1.510.78/1.710.82/1.850.62/1.460.82/1.85
AgentFormer [98]1.67/3.651.61/3.591.62/3.531.85/4.131.68/3.741.69/3.73
MID [18]1.42/2.940.64/1.300.76/1.620.74/1.590.60/1.310.83/1.75
EigenTrajectory [6]0.93/2.050.33/0.640.58/1.230.45/0.990.34/0.750.53/1.13
Social-LSTM [1]1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social-GAN [19]1.13/2.211.01/2.180.60/1.280.42/0.910.52/1.110.67/1.41
SR-LSTM [100]1.01/1.930.35/0.720.66/1.380.56/1.230.44/0.900.60/1.23
STAR-D [97]0.97/2.000.32/0.730.56/1.250.44/0.960.35/0.770.53/1.14
Trajectory++[64]1.02/2.090.33/0.630.52/1.160.42/0.940.32/0.710.52/1.11
SingularTrajectory0.72/1.230.27/0.500.57/1.120.44/0.930.35/0.730.47/0.90
", + "image_path": "284d9aa57509c06e041a206f4e55b9583e94d28863f9924419dd5b12658fe71c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 417, + 545, + 428 + ], + "lines": [ + { + "bbox": [ + 306, + 417, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 545, + 428 + ], + "type": "text", + "content": "Table 3. Evaluation on the deterministic trajectory prediction task." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "type": "text", + "content": "of using the leave-one-out strategy. In the few-shot learning, they are trained using only " + }, + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "type": "text", + "content": " of the training set, and then assessed on the entire test set. To measure the prediction performances, two common metrics are used for all the tasks, Average Displacement Error (ADE) and Final Displacement Error (FDE), measuring the average and destination error between prediction and its ground-truth path, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 547, + 416, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 547, + 416, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 547, + 416, + 559 + ], + "type": "text", + "content": "4.2. Evaluation Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 567, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 651 + ], + "type": "text", + "content": "Stochastic prediction task. In Tab. 2, since the ETH/UCY dataset release, many state-of-the-art models have made new records. Among them, our model also achieves performance competitive with the comparison methods. Our model regards historical, social, and environmental conditions of observations and images as the most important factors for successfully generating multimodal trajectories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "content": "Deterministic prediction task. Next, we evaluate our model in the deterministic trajectory prediction task in Tab. 3. Our SingularTrajectory model has the best performance over both these latest state-of-the-art models, and the models optimized for deterministic prediction. In particular, our model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17895" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 547, + 183 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 547, + 183 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 547, + 183 + ], + "type": "image", + "image_path": "41c74b1faede7cefd483bb93b6f6c2953df2b98c4bf7a1fa1443ac22303b1f93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "type": "text", + "content": "Figure 4. Visualization of prediction results on (a) momentary observation task and (b) few-shot task. To aid visualization, the best trajectory among " + }, + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "type": "inline_equation", + "content": "S = 20" + }, + { + "bbox": [ + 46, + 185, + 546, + 207 + ], + "type": "text", + "content": " samples are reported." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 225, + 286, + 326 + ], + "blocks": [ + { + "bbox": [ + 48, + 225, + 286, + 326 + ], + "lines": [ + { + "bbox": [ + 48, + 225, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 225, + 286, + 326 + ], + "type": "table", + "html": "
MomentaryETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]1.24/2.230.77/1.440.45/0.810.38/0.570.35/0.580.64/1.13
PECNet [45]0.63/1.040.28/0.530.28/0.490.25/0.440.19/0.340.33/0.57
AgentFormer [98]1.10/2.110.50/1.020.52/1.100.56/1.180.43/0.890.62/1.26
MID [18]0.63/1.050.29/0.490.30/0.560.30/0.560.22/0.400.35/0.61
EigenF trajectory [6]0.46/0.760.17/0.280.25/0.440.19/0.350.15/0.270.25/0.42
STT [55]0.72/1.450.48/0.480.53/1.090.64/1.210.44/0.880.57/0.93
STT+DTO [55]0.62/1.220.29/0.560.58/1.140.45/0.980.34/0.740.46/0.93
MOE-Next [77]0.71/1.570.30/0.580.52/1.120.38/0.810.33/0.730.45/0.96
MOE-Traj++ [77]0.64/1.120.20/0.330.33/0.620.22/0.420.17/0.320.31/0.56
SingularF trajectory0.45/0.670.18/0.290.24/0.430.19/0.330.17/0.280.25/0.40
", + "image_path": "d2fe4c99c3b4464a9a031de5da62bb4c5b629fc36bd4beb0d43fc6207b520ad5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 433 + ], + "type": "text", + "content": "exhibits significant performance improvements in the ETH and HOTEL scenes. These scenes have noisy observation paths, which have a negative impact on the prediction. We are able to achieve accurate predictions by leveraging the overall motion flow, which acts as a low-pass filter over the noisy sequence, in Singular space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 435, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 289, + 604 + ], + "type": "text", + "content": "Momentary observation task. In Tab. 4, the state-of-the-art predictors for momentary observation mainly focus on the coordinates themselves, leading to a performance drop, given only two frames. Some models for this task try to bridge these gaps, but have not been sufficient. Fortunately, with the benefit of our Singular Space, even with only a two-frame input, our model can successfully represent the overall long-term flow. Note that our SingularTrajectory with two-frame observation demonstrates the closest performance to state-of-the-art stochastic prediction models with an entire history frame; even this is achieved without any masked trajectory complement or knowledge distillation. Consequently, this allows it to accurately pinpoint the future locations of pedestrians, whose examples are displayed in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Domain adaptation task. We next evaluate the performance of SingularTrajectory in a domain adaptation task. For simplicity, the ETH, HOTEL, UNIV, ZARA1, and ZARA2 scenes are denoted as A, B, C, D, and E, respectively. For example, 'A2B' means that a model is trained on the ETH scene and tested on the HOTEL scene. As demonstrated in Tab. 1, our SingularTrajectory model shows the performance nearly equivalent to those of models specifically designed for deterministic prediction. Particularly," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 225, + 545, + 291 + ], + "blocks": [ + { + "bbox": [ + 65, + 327, + 268, + 337 + ], + "lines": [ + { + "bbox": [ + 65, + 327, + 268, + 337 + ], + "spans": [ + { + "bbox": [ + 65, + 327, + 268, + 337 + ], + "type": "text", + "content": "Table 4. Evaluation on the momentary observation task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 225, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 307, + 225, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 225, + 545, + 291 + ], + "type": "table", + "html": "
Few-ShotETHHOTELUNIVZARA1ZARA2AVG
STGCNN [53]0.89/1.631.22/2.480.90/1.610.68/1.251.36/2.121.01/1.82
PECNet [45]0.72/1.460.29/0.530.58/0.930.27/0.440.22/0.380.41/0.75
AgentFormer [98]1.60/2.651.02/1.641.13/1.901.19/2.011.08/1.591.20/1.96
MID [18]0.57/0.920.21/0.330.32/0.600.27/0.490.24/0.420.32/0.55
EigenF trajectory [6]0.39/0.640.13/0.210.25/0.430.21/0.390.15/0.270.23/0.39
SingularF trajectory0.35/0.460.14/0.210.26/0.440.21/0.360.18/0.310.23/0.35
", + "image_path": "17c24b334870d57838e72b03acd3b3d2cd90451a9b7c7e057d8beaa4ab5d7f6f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 293, + 538, + 304 + ], + "lines": [ + { + "bbox": [ + 312, + 293, + 538, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 538, + 304 + ], + "type": "text", + "content": "Table 5. Evaluation on the few-shot trajectory prediction task." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 326, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 396 + ], + "type": "text", + "content": "our model produces impressive results in the challenging B2A, C2A, D2A and E2A scenarios where even models specialized in domain adaptation fail. Our SingularTrajectory framework specializes in learning general human motions, and is not limited to a specific domain, enabling accurate predictions even in extreme cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 399, + 548, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 548, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 548, + 602 + ], + "type": "text", + "content": "Few-shot task. Finally, the results for the few-shot task are reported in Tab. 5, whose examples are in Fig. 4. As expected, our SingularTrajectory significantly outperforms the existing models. With limited data, the existing works, except the diffusion-based model, tend to easily overfit. In contrast, our methods make it possible to explicitly designate future prototype paths, while taking full advantage of the expression ability of the diffusion model. In particular, even with only " + }, + { + "bbox": [ + 304, + 399, + 548, + 602 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 399, + 548, + 602 + ], + "type": "text", + "content": " of the data, our model achieves substantial improvements with respect to both data efficiency and performance. Figure 5 illustrates several cases where there are differences between the predictions of SingularTrajectory and other comparison methods. Previous works often show significant performance drops in other associated tasks compared to stochastic prediction tasks. In contrast, our model consistently predicts the best trajectories across multiple tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 611, + 406, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 406, + 622 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 406, + 622 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "The number of motion vectors " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": ". First, we conduct a component study by varying the dimension " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " of Singular space in Tab. 6. To find the best number of singular vectors in general, we carry out experiments across all five tasks using ZARA1 scene where there are both human-environmental and human-human interactions, following [93]. As the number of motion vectors " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " increases, more detailed movements" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17896" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 287, + 181 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 181 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 181 + ], + "type": "image", + "image_path": "a1b2d777706d42baa698fe7f459ebe68b89215dc214120d48937d2a40bc2fc2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 183, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 183, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 287, + 205 + ], + "type": "text", + "content": "Figure 5. Visualization of prediction consistency across five tasks. The more consistent the prediction is the better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 217, + 285, + 282 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 285, + 282 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 285, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 285, + 282 + ], + "type": "table", + "html": "
KDeterministicStochasticMomentaryDomainFew-shotAverage
10.52 / 1.030.30 / 0.550.31 / 0.580.51 / 1.070.34 / 0.610.40 / 0.77
20.51 / 1.010.20 / 0.320.20 / 0.330.50 / 1.060.23 / 0.370.33 / 0.62
30.45 / 0.930.19 / 0.330.20 / 0.330.47 / 1.030.22 / 0.380.31 / 0.60
40.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
50.43 / 0.940.19 / 0.330.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
60.43 / 0.940.19 / 0.320.20 / 0.340.46 / 1.040.22 / 0.370.30 / 0.60
", + "image_path": "26ac6bfe0b4624b886696339905f3fdf4df1ee68c4647510fcce868abd3db792.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 308, + 285, + 365 + ], + "blocks": [ + { + "bbox": [ + 57, + 285, + 276, + 295 + ], + "lines": [ + { + "bbox": [ + 57, + 285, + 276, + 295 + ], + "spans": [ + { + "bbox": [ + 57, + 285, + 276, + 295 + ], + "type": "text", + "content": "Table 6. Ablation study on the Singular space dimension " + }, + { + "bbox": [ + 57, + 285, + 276, + 295 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 57, + 285, + 276, + 295 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 308, + 285, + 365 + ], + "lines": [ + { + "bbox": [ + 48, + 308, + 285, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 285, + 365 + ], + "type": "table", + "html": "
AdoptionDeterministicStochasticMomentaryDomainFew-shotAverage
Direct0.75 / 1.470.27 / 0.480.31 / 0.540.78 / 1.600.27 / 0.480.48 / 0.91
Initial0.47 / 0.950.22 / 0.400.22 / 0.390.50 / 1.170.23 / 0.400.33 / 0.66
Residual0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
Independent0.46 / 0.940.21 / 0.390.21 / 0.390.49 / 1.160.21 / 0.390.32 / 0.65
Jointly0.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
", + "image_path": "dc173bb1db23101557da66271c6af104dc8ee042a0cfefd84b9926310e34d688.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "text", + "content": "are captured. In contrast, when using the smaller " + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "text", + "content": ", it compresses the space, mainly to represent the overall motion flow. The performance tends to plateau when " + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "text", + "content": " is larger than 3, as long as " + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "text", + "content": " is not too small to cover most movements. We set " + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 46, + 400, + 288, + 483 + ], + "type": "text", + "content": " as the dimension for the Singular space because it shows the most effective prediction results across all tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "Trajectory denoising methods. Next, we evaluate three types of diffusion models for the trajectory prediction tasks, as shown in Tab. 7. First, we use a basic model similar to MID [18], which directly denoises from a Gaussian noise to a trajectory. This method fails to achieve good performances. The use of an anchor as an intermediate state, similar to LED [48], which reduces the denoising steps by skipping the initial denoising steps, seems to validate its generality. This demonstrates that our adaptive anchor can function as a good initializer for the diffusion model, particularly showing nearly identical outcomes in the stochastic, momentary observation, and few-shot tasks when predicting a multimodal path. However, we confirm that our scheme, which denoises only a residual by adding perturbation to a prototype path in Fig. 6, showcases the best performance in all the tasks. Additionally, compared to refining the prototype paths independently, regarding them as a batch dimension, our model can accurately predict the future when prototype paths are jointly refined." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 159 + ], + "type": "image", + "image_path": "df7e34619e76f5d29e9b9a15a1ddd4bea2c99f89453fd42e28d4eed41c4ec12f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 159, + 545, + 181 + ], + "lines": [ + { + "bbox": [ + 306, + 159, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 545, + 181 + ], + "type": "text", + "content": "Figure 6. Visualization of anchor refinement. The denoising process progressively refines the prototype paths at each diffusion step " + }, + { + "bbox": [ + 306, + 159, + 545, + 181 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 159, + 545, + 181 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 192, + 545, + 249 + ], + "blocks": [ + { + "bbox": [ + 58, + 366, + 275, + 377 + ], + "lines": [ + { + "bbox": [ + 58, + 366, + 275, + 377 + ], + "spans": [ + { + "bbox": [ + 58, + 366, + 275, + 377 + ], + "type": "text", + "content": "Table 7. Ablation study on the adoption of diffusion model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 192, + 545, + 249 + ], + "lines": [ + { + "bbox": [ + 307, + 192, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 192, + 545, + 249 + ], + "type": "table", + "html": "
MDeterministicStochasticMomentaryDomainFew-shotAverage
10.45 / 0.950.20 / 0.340.20 / 0.350.48 / 1.050.21 / 0.370.31 / 0.61
20.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
50.44 / 0.930.19 / 0.330.20 / 0.340.47 / 1.040.20 / 0.360.30 / 0.60
100.44 / 0.930.19 / 0.320.19 / 0.330.47 / 1.030.21 / 0.360.30 / 0.59
250.44 / 0.930.19 / 0.320.20 / 0.340.47 / 1.030.20 / 0.360.30 / 0.60
", + "image_path": "4cec812089f2c595af831e2b9e59e0cb10279017e8cab547a028f660e6ebe88b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 334, + 250, + 515, + 261 + ], + "lines": [ + { + "bbox": [ + 334, + 250, + 515, + 261 + ], + "spans": [ + { + "bbox": [ + 334, + 250, + 515, + 261 + ], + "type": "text", + "content": "Table 8. Ablation study on the diffusion steps " + }, + { + "bbox": [ + 334, + 250, + 515, + 261 + ], + "type": "inline_equation", + "content": "M" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": "Diffusion steps " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": ". Lastly, we check how many steps in the diffusion model are needed for the cascaded refinement of the adaptive anchor. In Tab. 8, we confirm that the best performance comes from " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": ". This is because the DDIM scheduler accelerates its convergence and the prototype path provides a rough initial trajectory, and so fewer denoising steps are sufficient. However, as the number of denoising steps increases, the information in the initial prototype path becomes attenuated due to the noise. As a result, we observe a slight decrease in performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 414, + 378, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 378, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 378, + 426 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 434, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 601 + ], + "type": "text", + "content": "In this study, we introduce SingularTrajectory, a universal trajectory predictor model for all related trajectory prediction tasks. By unifying trajectory modalities into one Singular space, our approach standardizes trajectory data with shared motion dynamics, which eliminates the need for task-specific adjustments. The incorporation of an adaptive anchor system further personalizes the prototype paths, allowing them to interpret and adapt to environmental contexts and enhancing the reliability of the trajectory prediction. By successfully incorporating Singular space into the diffusion model, our SingularTrajectory framework successfully achieves the state-of-the-art results across five different benchmarks. This establishes SingularTrajectory itself as a general solution that covers a multitude of scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), Vehicles AI Convergence Research & Development Program through the National IT Industry Promotion Agency of Korea (NIPA) funded by the Ministry of Science and ICT (No.S1602-20-1001), and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17897" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 58, + 91, + 288, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 288, + 145 + ], + "type": "text", + "content": "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 58, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Inhwan Bae and Hae-Gon Jeon. Disentangled multi-relational graph convolutional network for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 193, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 57, + 193, + 288, + 236 + ], + "type": "text", + "content": "[3] Inhwan Bae and Hae-Gon Jeon. A set of control points conditioned pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 57, + 238, + 288, + 282 + ], + "type": "text", + "content": "[4] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Learning pedestrian group representations for multi-modal trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 284, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 284, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 57, + 284, + 288, + 338 + ], + "type": "text", + "content": "[5] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 340, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 340, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 57, + 340, + 288, + 384 + ], + "type": "text", + "content": "[6] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. EigenTrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 386, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 386, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 386, + 288, + 430 + ], + "type": "text", + "content": "[7] Apratim Bhattacharyya, Michael Hanselmann, Mario Fritz, Bernt Schiele, and Christoph-Nikolas Straehle. Conditional flow variational autoencoders for structured sequence prediction. arXiv preprint arXiv:1908.09008, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 431, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 431, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 431, + 288, + 475 + ], + "type": "text", + "content": "[8] Niccolò Bisagno, Bo Zhang, and Nicola Concì. Group lstm: Group trajectory prediction in crowded scenarios. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 476, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 288, + 520 + ], + "type": "text", + "content": "[9] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 522, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 288, + 576 + ], + "type": "text", + "content": "[10] Guangyi Chen, Junlong Li, Nuoxing Zhou, Liangliang Ren, and Jiwen Lu. Personalized trajectory prediction via distribution discrimination. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 578, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 288, + 633 + ], + "type": "text", + "content": "[11] Guangyi Chen, Zhenhao Chen, Shunxing Fan, and Kun Zhang. Unsupervised sampling promoting for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 635, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 668 + ], + "type": "text", + "content": "[12] Sourav Das, Guglielmo Camporese, and Lamberto Ballan. Distilling knowledge for short-to-long term trajectory prediction. arXiv preprint arXiv:2305.08553, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "text", + "content": "[13] Patrick Dendorfer, Aljosa Osep, and Laura Leal-Taixe. Goalgan: Multimodal trajectory prediction based on goal position estimation. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "type": "text", + "content": "[14] Patrick Dendorfer, Sven Elflein, and Laura Leal-Taixe. M. Gagan: A multi-generator model preventing out-of-distribution samples in pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 129, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 547, + 161 + ], + "type": "text", + "content": "[15] Nachiket Deo and Mohan M. Trivedi. Trajectory forecasts in unknown environments conditioned on grid-based plans. arXiv preprint arXiv:2001.00735, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 162, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 547, + 194 + ], + "type": "text", + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Proceedings of the Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 195, + 547, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 547, + 238 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 547, + 238 + ], + "type": "text", + "content": "[17] Tharindu Fernando, Simon Denman, Sridha Sridharan, and Clinton Fookes. Soft+ hardwired attention: An lstm framework for human trajectory prediction and abnormal event detection. Neural Networks, 108:466-478, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 239, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 547, + 293 + ], + "type": "text", + "content": "[18] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 294, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 294, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 547, + 348 + ], + "type": "text", + "content": "[19] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 350, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 547, + 380 + ], + "type": "text", + "content": "[20] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 383, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 547, + 415 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 416, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 547, + 470 + ], + "type": "text", + "content": "[22] Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. Stgat: Modeling spatial-temporal interactions for human trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 472, + 547, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 547, + 524 + ], + "type": "text", + "content": "[23] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 526, + 547, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 526, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 526, + 547, + 580 + ], + "type": "text", + "content": "[24] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 582, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 547, + 635 + ], + "type": "text", + "content": "[25] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 636, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 547, + 690 + ], + "type": "text", + "content": "[26] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "type": "text", + "content": "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17898" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "text", + "content": "the International Conference on Learning Representations (ICLR), 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 151 + ], + "type": "text", + "content": "[28] Vineet Kosaraju, Amir Sadeghian, Roberto Martín-Martín, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 206 + ], + "type": "text", + "content": "[29] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "text", + "content": "[30] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Muse-vae: Multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "text", + "content": "[31] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B. Choy, Philip H. S. Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 320, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 288, + 352 + ], + "type": "text", + "content": "[32] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Computer Graphics Forum, 26(3): 655-664, 2007. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 288, + 399 + ], + "type": "text", + "content": "[33] Jiachen Li, Hengbo Ma, and Masayoshi Tomizuka. Conditional generative neural system for probabilistic trajectory prediction. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 443 + ], + "type": "text", + "content": "[34] Jiachen Li, Fan Yang, Masayoshi Tomizuka, and Chiho Choi. Evolvegraph: Multi-agent trajectory prediction with dynamic relational reasoning. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 444, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 288, + 488 + ], + "type": "text", + "content": "[35] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 288, + 533 + ], + "type": "text", + "content": "[36] Yuke Li. Which way are you going? imitative decision learning for path forecasting in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 534, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 288, + 589 + ], + "type": "text", + "content": "[37] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 288, + 634 + ], + "type": "text", + "content": "[38] Junwei Liang, Lu Jiang, and Alexander Hauptmann. Simaug: Learning robust representations from simulation for trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 689 + ], + "type": "text", + "content": "[39] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 713 + ], + "type": "text", + "content": "[40] Rongqin Liang, Yuanman Li, Xia Li, Yi Tang, Jiantao Zhou, and Wenbin Zou. Temporal pyramid network for pedestrian" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "type": "text", + "content": "trajectory prediction with multi-supervision. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 106, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 547, + 150 + ], + "type": "text", + "content": "[41] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 151, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 151, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 151, + 547, + 183 + ], + "type": "text", + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 184, + 547, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 547, + 238 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 547, + 238 + ], + "type": "text", + "content": "[43] Yuexin Ma, Xinge Zhu, Xinjing Cheng, Ruigang Yang, Jiming Liu, and Dinesh Manocha. Autotrajectory: Label-free trajectory extraction and prediction from videos using dynamic points. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 239, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 547, + 282 + ], + "type": "text", + "content": "[44] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 283, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 547, + 348 + ], + "type": "text", + "content": "[45] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 349, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 349, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 349, + 547, + 403 + ], + "type": "text", + "content": "[46] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 404, + 547, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 547, + 436 + ], + "type": "text", + "content": "[47] Huynh Manh and Gita Alaghband. Scene-lstm: A model for human trajectory prediction. arXiv preprint arXiv:1808.04018, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 437, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 547, + 491 + ], + "type": "text", + "content": "[48] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 492, + 547, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 492, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 312, + 492, + 547, + 548 + ], + "type": "text", + "content": "[49] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 548, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 547, + 602 + ], + "type": "text", + "content": "[50] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Multiple trajectory prediction of moving agents with memory augmented networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 602, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 602, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 312, + 602, + 547, + 645 + ], + "type": "text", + "content": "[51] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Smemo: Social memory for trajectory forecasting. arXiv preprint arXiv:2203.12446, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 547, + 690 + ], + "type": "text", + "content": "[52] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "type": "text", + "content": "[53] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgcnn: A social spatio-temporal" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17899" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 72, + 73, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 288, + 116 + ], + "type": "text", + "content": "graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 182 + ], + "type": "text", + "content": "[54] Abduallah Mohamed, Deyao Zhu, Warren Vu, Mohamed Elhoseiny, and Christian Claudel. Social-implicit: Rethinking trajectory prediction evaluation and the effectiveness of implicit maximum likelihood estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 184, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 184, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 54, + 184, + 288, + 248 + ], + "type": "text", + "content": "[55] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 251, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 288, + 293 + ], + "type": "text", + "content": "[56] Ingrid Navarro and Jean Oh. Social-patternn: Socially-aware trajectory prediction guided by motion patterns. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 295, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 288, + 336 + ], + "type": "text", + "content": "[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 288, + 381 + ], + "type": "text", + "content": "[58] Bo Pang, Tianyang Zhao, Xu Xie, and Ying Nian Wu. Trajectory prediction with latent belief energy-based model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 384, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 288, + 435 + ], + "type": "text", + "content": "[59] Wonhui Park, Dongkwon Jin, and Chang-Su Kim. Eigencontours: Novel contour descriptors based on low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 437, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 491 + ], + "type": "text", + "content": "[60] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 494, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 494, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 494, + 288, + 557 + ], + "type": "text", + "content": "[61] Mark Pfeiffer, Giuseppe Paolo, Hannes Sommer, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. A data-driven model for interaction-aware pedestrian motion prediction in object cluttered environments. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 559, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 288, + 623 + ], + "type": "text", + "content": "[62] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 625, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 288, + 689 + ], + "type": "text", + "content": "[63] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 288, + 713 + ], + "type": "text", + "content": "[64] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajec" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 547, + 105 + ], + "type": "text", + "content": "tory forecasting with heterogeneous data. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 107, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 547, + 149 + ], + "type": "text", + "content": "[65] Nasim Shafiee, Taskin Padir, and Ehsan Elhamifar. Introvert: Human trajectory prediction via conditional 3d attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 152, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 205 + ], + "type": "text", + "content": "[66] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 206, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 547, + 259 + ], + "type": "text", + "content": "[67] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 262, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 547, + 304 + ], + "type": "text", + "content": "[68] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 306, + 547, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 306, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 312, + 306, + 547, + 359 + ], + "type": "text", + "content": "[69] Xiaodan Shi, Xiaowei Shao, Zipei Fan, Renhe Jiang, Haoran Zhang, Zhiling Guo, Guangming Wu, Wei Yuan, and Ryosuke Shibasaki. Multimodal interaction-aware trajectory prediction in crowded space. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 361, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 547, + 415 + ], + "type": "text", + "content": "[70] Xiaodan Shi, Xiaowei Shao, Guangming Wu, Haoran Zhang, Zhiling Guo, Renhe Jiang, and Ryosuke Shibasaki. Socialdpf: Socially acceptable distribution prediction of futures. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 416, + 547, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 547, + 468 + ], + "type": "text", + "content": "[71] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the International Conference on Machine Learning (ICML), 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 472, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 547, + 502 + ], + "type": "text", + "content": "[72] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 504, + 547, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 547, + 546 + ], + "type": "text", + "content": "[73] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Proceedings of the Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 548, + 547, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 547, + 591 + ], + "type": "text", + "content": "[74] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 593, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 547, + 635 + ], + "type": "text", + "content": "[75] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 636, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 547, + 689 + ], + "type": "text", + "content": "[76] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 547, + 713 + ], + "type": "text", + "content": "[77] Jianhua Sun, Yuxuan Li, Liang Chai, Hao-Shu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17900" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "text", + "content": "momentary observation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 162 + ], + "type": "text", + "content": "[78] Jianhua Sun, Yuxuan Li, Liang Chai, and Cewu Lu. Stimulus verification is a universal and effective sampler in multi-modal human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 205 + ], + "type": "text", + "content": "[79] Chaofan Tao, Qinhong Jiang, and Lixin Duan. Dynamic and static context-aware LSTM for multi-agent motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 206, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 288, + 248 + ], + "type": "text", + "content": "[80] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1991. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "type": "text", + "content": "[81] Daksh Varshneya and G. Srinivasaraghavan. Human trajectory prediction using spatially aware deep attention models. arXiv preprint arXiv:1705.09436, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "text", + "content": "[82] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 327, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 371 + ], + "type": "text", + "content": "[83] Anirudh Vemula, Katharina Muelling, and Jean Oh. Social attention: Modeling attention in human crowds. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 372, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 288, + 404 + ], + "type": "text", + "content": "[84] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters (RA-L), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 459 + ], + "type": "text", + "content": "[85] Yuning Wang, Pu Zhang, Lei Bai, and Jianru Xue. Fend: A future enhanced distribution-aware contrastive learning framework for long-tail trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 288, + 503 + ], + "type": "text", + "content": "[86] Song Wen, Hao Wang, and Dimitris Metaxas. Social ode: Multi-agent trajectory forecasting with neural ordinary differential equations. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 504, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 288, + 558 + ], + "type": "text", + "content": "[87] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 288, + 613 + ], + "type": "text", + "content": "[88] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 614, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 288, + 667 + ], + "type": "text", + "content": "[89] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "type": "text", + "content": "[90] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 700 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 140 + ], + "type": "text", + "content": "[91] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 547, + 184 + ], + "type": "text", + "content": "[92] Yi Xu, Jing Yang, and Shaoyi Du. Cf-lstm: Cascaded feature-based long short-term networks for predicting pedestrian trajectory. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 185, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 547, + 229 + ], + "type": "text", + "content": "[93] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 230, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 230, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 230, + 547, + 274 + ], + "type": "text", + "content": "[94] Hao Xue, Du Q Huynh, and Mark Reynolds. Ss-lstm: A hierarchical LSTM model for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 275, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 275, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 275, + 547, + 319 + ], + "type": "text", + "content": "[95] Kota Yamaguchi, Alexander C Berg, Luis E Ortiz, and Tamara L Berg. Who are you with and where are you going? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 320, + 547, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 320, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 320, + 547, + 364 + ], + "type": "text", + "content": "[96] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters (RA-L), 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 365, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 365, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 312, + 365, + 547, + 409 + ], + "type": "text", + "content": "[97] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 410, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 410, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 312, + 410, + 547, + 464 + ], + "type": "text", + "content": "[98] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 465, + 547, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 465, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 312, + 465, + 547, + 508 + ], + "type": "text", + "content": "[99] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 510, + 547, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 547, + 564 + ], + "type": "text", + "content": "[100] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 566, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 547, + 609 + ], + "type": "text", + "content": "[101] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 611, + 547, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 547, + 666 + ], + "type": "text", + "content": "[102] Tianyang Zhao, Yifei Xu, Mathew Monfort, Wongun Choi, Chris Baker, Yibiao Zhao, Yizhou Wang, and Ying Nian Wu. Multi-agent tensor fusion for contextual trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 666, + 547, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 547, + 700 + ], + "type": "text", + "content": "[103] ChunYu Zhi, HuaiJiang Sun, and Tian Xu. Adaptive trajectory prediction without catastrophic forgetting. The Journal of Supercomputing, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17901" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_content_list.json b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..979ef30ed2f5298a1877528e811d5a6c1fd8f168 --- /dev/null +++ b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_content_list.json @@ -0,0 +1,1354 @@ +[ + { + "type": "text", + "text": "Situational Awareness Matters in 3D Vision Language Reasoning", + "text_level": 1, + "bbox": [ + 156, + 130, + 816, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunze Man Liang-Yan Gui Yu-Xiong Wang University of Illinois Urbana-Champaign", + "bbox": [ + 269, + 180, + 700, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yunzem2,lgui,yxw}@illinois.edu", + "bbox": [ + 344, + 219, + 620, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Being able to carry out complicated vision language reasoning tasks in 3D space represents a significant milestone in developing household robots and human-centered embodied AI. In this work, we demonstrate that a critical and distinct challenge in 3D vision language reasoning is the situational awareness, which incorporates two key components: (1) The autonomous agent grounds its self-location based on a language prompt. (2) The agent answers open-ended questions from the perspective of its calculated position. To address this challenge, we introduce SIG3D, an end-to-end Situation-Grounded model for 3D vision language reasoning. We tokenize the 3D scene into sparse voxel representation, and propose a language-grounded situation estimator, followed by a situated question answering module. Experiments on the SQA3D and ScanQA datasets show that SIG3D outperforms state-of-the-art models in situational estimation and question answering by a large margin (e.g., an enhancement of over $30\\%$ on situation accuracy). Subsequent analysis corroborates our architectural design choices, explores the distinct functions of visual and textual tokens, and highlights the importance of situational awareness in the domain of 3D question-answering. Project page is available at https://yunzeman.github.io/situation3d.", + "bbox": [ + 75, + 301, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans learn knowledge efficiently through the interactions with the 3D world and the integration of multi-modal information, such as verbal guidance or instructions. Similarly, introducing language guidance into the visual comprehension task can greatly enhance models' learning efficiency [3, 42]. Nonetheless, despite considerable advancements in linguistic understanding [7, 11, 31, 58] and vision-language integration [3, 37, 54, 61], current methodologies remain deficient in accurately perceiving and rationalizing within real-world 3D environments, which is largely attributed to the lack of 3D situational reasoning capabilities.", + "bbox": [ + 75, + 718, + 468, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared with machine learning models, humans put", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3743d41e5d3a84bf69adb681218bac04137bc2965fe7e2159dbb3884a53b9f95.jpg", + "image_caption": [ + "Figure 1. Previous methods perform direct 3D vision language reasoning without modeling the situation of an embodied agent in the 3D environment. Our method, SIG3D, grounds the situational description in the 3D space, and then re-encodes the visual tokens from the agent's intended perspective before vision-language fusion, resulting in a more comprehensive and generalized 3D vision language (3DVL) representation and reasoning pipeline. Q, K, V stands for query, key, and value, respectively." + ], + "image_footnote": [], + "bbox": [ + 501, + 268, + 890, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "themselves inside the 3D world and then perceive and interact with the surrounding environment from their egoperspective (Figure 1). Such situational awareness is a crucial difference between 2D and 3D visual understanding, and a key to achieve seamless understanding of spatial concepts in more complex real-world environments. Several existing methods recognize the lack of positional understanding in 3D and propose new benchmarks and joint optimization functions [44], or positional embedding methods [26] to enhance the overall reasoning performance.", + "bbox": [ + 496, + 594, + 893, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the lack of an explicit situation modeling and situation-grounded 3D reasoning method restricts them from obtaining a generalizable and consistent 3D vision-language (VL) representation. As shown in Figure 2, the situation prediction of the state-of-the-art method [44] (in blue) diverges significantly from the ground truth vectors (in red) in almost all scenes in the dataset [13]. Moreover, our pilot study in Section 3 also reveals that situational understanding, despite being very crucial in comprehending the context of questions, only plays a minor role in the final", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13678", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2652a4744e4f1b9b629ac3b59c6a5e5c71bcbe328d03926159a7f561e502346.jpg", + "image_caption": [ + "I am facing a trash can, while there is a door on my right." + ], + "image_footnote": [], + "bbox": [ + 86, + 94, + 259, + 193 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/976466d9324826ee50a29bec14a5c7f88056478dd6b771190e13afc97ae6a01c.jpg", + "image_caption": [ + "I am facing the paper towel dispenser with a chair in my six o'clock direction." + ], + "image_footnote": [], + "bbox": [ + 271, + 94, + 447, + 194 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2165f6d32d3e15bd7ee574ddfc128ad3bb03b33c2d22d0851737bc3dddd20906.jpg", + "image_caption": [ + "I am facing a pool table, taking out the white ball from the side pocket. I am also facing a painting on the wall." + ], + "image_footnote": [], + "bbox": [ + 83, + 227, + 256, + 303 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c2658c3e9812b52434ac8e0ee2d0b95a7787c20d17889903217dccc87ce8cc0b.jpg", + "image_caption": [ + "I am sitting on a couch with a pillow facing another couch and the pillow is on my right.", + "Figure 2. Situational estimation in existing methods [44] fails in most scenarios, indicating the missing registration between the situational description and 3D embeddings. Red: Ground truth (GT) vector. Blue: Estimated vector." + ], + "image_footnote": [], + "bbox": [ + 272, + 224, + 437, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "question-answering performance of existing methods.", + "bbox": [ + 75, + 424, + 431, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose SIG3D, a novel approach designed to precisely model and estimate an embodied agent's ego-location and orientation from a textual description, before performing multi-modal QA tasks from the agent's egocentric perspective, as shown in Figure 1. Specifically, we leverage large-scale pretrained language and visual encoders to process the input text and 3D data, and fuse the tokens with attention modules to predict a situational vector. Previous attempts to directly predict the ego-situation are hindered by the expansive search space inherent in 3D environments. To address this challenge, we re-conceptualize the task as an anchor-based classification, where visual tokens are regarded as anchor points, and a likelihood of position together with a set of rotation parameters are concurrently regressed for each visual token. After obtaining the situational estimation, we propose a situational alignment and a situation-guided token re-encoding strategy, to perceive the environment from the agent's intended perspective. These strategies enhance the visual token with more accurate situational awareness for subsequent QA tasks.", + "bbox": [ + 73, + 441, + 468, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on two challenging 3D visual question-answering (VQA) datasets [5, 44] demonstrate the significant improvement on situational estimation and QA tasks of our model. In particular, we improve the situational estimation accuracy by over $30\\%$ , and the subsequent QA performance by up to $3\\%$ . Further qualitative and quantitative analysis verifies our design choices and highlights the significance of situational awareness in 3D reasoning tasks.", + "bbox": [ + 75, + 747, + 467, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our paper has the following contributions: (1) We recognize the lack of situational awareness as a sig-", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nificant oversight in existing research. To address this, we introduce SIG3D, a situation-grounded 3D VL reasoning architecture, specifically designed to fill this void. (2) We propose an anchor-based approach to situational estimation, which effectively narrows the extensive search space in 3D environments for precise grounding of 3D positions and orientations with textual descriptions. Additionally, we investigate situational alignment and visual re-encoding mechanisms to leverage situational awareness for enhanced QA performance. (3) Our model demonstrates superior performance on two challenging datasets, SQA3D and ScanQA, surpassing the state of the art in both situational estimation and QA metrics. Ablation studies highlight the importance of situation-guided encoding, revealing its beneficial impact on general QA tasks.", + "bbox": [ + 496, + 90, + 890, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 335, + 640, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Language Models (VLMs). Early transformer-driven [59] textual and visual encoders [17, 31] have facilitated great progress in recent vision language learning. Text-image contrastive models [30, 54] propose to align the feature space of two modalities with large-scale pretraining, fueling numerous downstream tasks from generalized open-vocabulary visual perception [21, 34, 36] to text-to-image generation [56]. Concurrently, some work uses text and vision encoders on separate modalities followed by feature fusion [18, 33] for multi-modal reasoning tasks. Since the emergence of Large Language Models (LLMs) [7, 58, 69], VLMs have experienced huge improvement with the help of LLMs as building blocks for multi-modal learning architectures. Specifically, recent work directly projects visual embeddings into language-space tokens as input to LLM [23, 40, 47, 70], or use the latent bottleneck structure for cross-modal visual decoding [3, 26, 37, 38], or treat LLM layers as encoder blocks for various visual tasks [49].", + "bbox": [ + 496, + 363, + 890, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the domain of visual question-answering (VQA) [4, 73], recent work has pushed the frontier towards video understanding [14, 28, 29, 40, 62], knowledge-based understanding [20, 22, 23, 41, 46, 57, 63], and commonsense reasoning [68]. Despite the outstanding performance in 2D image interpretation, most existing methods lack the capability to generalize to 3D scenarios. In contrast, our work studies the representation of visual information and its fusion with language embeddings in the 3D domain by targeting on the 3D situation-guided visual language interpretation.", + "bbox": [ + 496, + 637, + 890, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Grounding Language in 3D Space. Compared with 2D images, knowledge such as spatial relationships, interactive exploration, and topological analysis, which only exist in the 3D world provides additional challenges and opportunities to develop better language models with stronger commonsense reasoning capability grounded in the real-world 3D scenarios. In this direction, early work seeks", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd7ff2f87d095cdce14564a7bb9feee5ce4a817302e0629fecdf932f2ab0596a.jpg", + "image_caption": [ + "Figure 3. Results on variants of the representative SQA3D baseline method [44] demonstrate that situational understanding, despite being indispensable in perceiving the context of questions, makes negligible contribution in existing methods. This motivates a situation-guided 3D encoding mechanism in our model." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to ground isolated objects [1, 9] or objects within more complex scenes [2, 8, 19, 27] using natural language descriptions. Recently, with more collected 3D vision language datasets, several work starts to explore language-guided 3D visual interpretation and reasoning on a diverse set of datasets, including 3D scene captioning [10], open-vocabulary segmentation [16, 32, 51], and question-answering [5, 15, 25, 65, 74].", + "bbox": [ + 75, + 271, + 467, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The success of LLMs also elicits the usage of them in 3D vision language reasoning for task decomposition [64], data generation, and multi-modal feature fusion [26]. Motivated by ScanQA [5], SQA3D [44] takes the first step into exploring the challenging 3D situational reasoning problem by developing a situated question-answering benchmark, and proposing the first joint learning baseline on the benchmark. Our work highlights the uniqueness and significance of situational awareness in the 3D vision language learning paradigm, leading to notably better 3D situational grounding and question-answering performance.", + "bbox": [ + 75, + 392, + 467, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Pilot Study on Situated Reasoning", + "text_level": 1, + "bbox": [ + 76, + 571, + 387, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite pointing out the significance of situated understanding and reasoning, existing methods [44] fall short in providing effective situational estimation, as illustrated in Figure 2. This section delves into a pilot study examining the impact of situational understanding on downstream reasoning tasks. The SQA3D baseline [44] incorporates situational descriptions and uses ground truth (GT) situational vectors for supervision in a direct regression task. We investigate three variants of this baseline to assess the effect of situational understanding. In the first variant, we remove the situational description and supervision from the model, by passing in empty situational tokens. In another variant, we corrupt the situation supervision by introducing very large Gaussian noise to the GT vectors to effectively randomizing them. Finally, we try to encode the GT situational vector in the input with learnable multi-layer perceptron layers, which are then integrated with the visual and textual tokens.", + "bbox": [ + 75, + 598, + 467, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3 demonstrates the results of this study, revealing negligible changes in performance across these variants. Notably, corrupting the GT situational information or di", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rectly incorporating it results in only marginal alterations in the QA outcomes. Omitting the situational description entirely from the input results in a mere $2\\%$ decline in accuracy. However, in the absence of this information, the model resorts to random guessing when determining the correct answer, as all responses depend on situational context. The findings from Figures 2 and 3 collectively indicate a deficiency in existing methods regarding situation vector estimation and the application of situational understanding in subsequent reasoning tasks. These unresolved challenges motivate the development of our proposed method.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 498, + 273, + 589, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our approach SIG3D is illustrated in Figure 4. Our method begins with a set of points that represent a 3D scene, accompanied by a situational description and a question, which defines the overall context of the problem. We tokenize them into separate token embeddings (Section 4.1), and ground the textual description in the 3D scene with a vector comprising of location and orientation. We find direct single vector estimation to be challenging due to the vast and complex nature of the 3D search space, so we propose an anchor-based situational estimation strategy (Section 4.2). Subsequently, we re-encode the visual tokens from the perspective of situational vectors, enhancing the situational awareness for downstream reasoning tasks (Section 4.3). The finalized visual and textual tokens are fused by a transformer decoder to generate the final response.", + "bbox": [ + 496, + 299, + 890, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Visual and Textual Tokenization", + "text_level": 1, + "bbox": [ + 498, + 537, + 782, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging input scene point clouds and textual prompts, our objective is to generate three distinct types of tokens: 3D visual tokens $z^{3\\mathrm{D}} \\in \\mathbb{R}^{N_v \\times C_v}$ , situational tokens $z^{\\mathrm{S}} \\in \\mathbb{R}^{N_s \\times C_s}$ , and question tokens $z^{\\mathrm{Q}} \\in \\mathbb{R}^{N_q \\times C_q}$ . Each type of token is characterized by two primary components: $N$ , representing the number of tokens, and $C$ , encapsulating the feature embeddings. Tockenize and capture the feature embeddings for situational and question inputs, we employ a shared text tokenizer $\\mathrm{E}^{\\mathrm{TXT}}$ following prior methods [5, 44]. We assume that situational and question prompts are separated in the input data. If not, LLMs [7] can be used to parse the textual input without changing the semantic meaning of the sentences. However, there is a lack of consensus on a standard 3D visual tokenization method $\\mathrm{E}^{3\\mathrm{D}}$ that is apt for the 3DVL reasoning task, prompting a more detailed exploration in the subsequent paragraphs.", + "bbox": [ + 496, + 561, + 890, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Tokenization. Given an input point cloud $\\mathbf{p} \\in \\mathbb{R}^{N \\times 3}$ , most prior methods [5, 15, 44] adopt a VoteNet [53] detector to acquire object-level tokens $z^{3\\mathrm{D}} \\in \\mathbb{R}^{N_{\\mathrm{obj}} \\times C_{\\mathrm{obj}}}$ as the visual representation, where $N_{\\mathrm{obj}}$ is the number of object proposals, and $C_{\\mathrm{obj}}$ is the object-level feature embeddings. However, we point out several problems with", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13680", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7b27bdd4f2816e51261ad8540905c741dd7b3c0866dab967ea8a81792e51acf8.jpg", + "image_caption": [ + "Figure 4. Overview of our SIG3D model, which includes 3D scene and text encoding, anchor-based situational estimation, situation-guided visual re-encoding, and multi-modal decoder modules. We tokenize the 3D scene into voxels, treat each token as an anchor point, and query the text tokens to predict a token-level position likelihood and rotation matrix to locate the situational vector associated with the text descriptions. Then we update the scene tokens with situational position encoding (PE), and finally perform the 3DVL reasoning task with a large transformer decoder." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 897, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this abstraction strategy: (1) A detection-based tokenization method tends to ignore the non-object regions in the scene, which can be indispensable in some reasoning scenarios (e.g., carpets on the ground, ceiling, walls.) (2) After object-level abstraction, the visual representation loses the high-level information of the scene (e.g., the shape of the living room, the corner of the kitchen.) (3) A supervised detector trained from scratch can only recognize objects within the training set (e.g., only 20 categories for ScanNet [13]), meaning that the method does not have zero-shot capability to reason about novel unseen objects that are inevitably common in real-world scenarios.", + "bbox": [ + 75, + 444, + 472, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In light of this, we adopt a pretrained open-vocabulary voxel-based tokenization method from OpenScene [51]. The scene is first discretized into regular small 3D voxels and fed into a visual encoder for feature extraction:", + "bbox": [ + 75, + 632, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\mathrm {3 D}} = \\operatorname {E} ^ {\\mathrm {3 D}} (\\mathcal {V} (\\mathbf {p})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 695, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{V}$ represents the voxelization process, and $\\mathrm{E}^{3\\mathrm{D}}$ is a Minkowski sparse 3D convolutional network [12]. The sparse network is pretrained by distillation from CLIP [54] embeddings of rendered multi-view 2D images, resulting in a feature map with better language alignment and 3D awareness. We take the upsampled bottleneck-layer feature embeddings from the encoder network, and compute the mean average over the $z$ -axis (vertical) to project the voxels onto the $x - y$ plane and treat the feature grids in the resulting 2D feature map as our $N_{v}$ visual tokens. We find that this bird's-eye-view projection results in a more compact representation and improves the final performance.", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Situational Estimation", + "text_level": 1, + "bbox": [ + 498, + 443, + 709, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given 3D visual tokens $z^{3\\mathrm{D}}$ and situational tokens $z^{\\mathrm{S}}$ , our objective is to estimate the situational vector $\\vec{s}$ referred by the situational description, which comprises of a position component $s^{\\mathrm{pos}}$ represented by $(x,y,z)$ coordinates, and a rotation component $s^{\\mathrm{rot}}$ represented by $(\\theta ,\\psi ,\\phi)$ Euler angles, where pitch angles $\\psi$ are always defined as 0, meaning that situational vectors are defined to be parallel with the ground plane. The prior method [44] utilizes a transformer block to calculate the cross-attention feature between visual and language tokens, and directly regress a final situational vector out of the averaged attention map. We find such strategy producing very inaccurate estimation, as shown in Figure 2, due to the large search space in the entire 3D volume. Inspired by the recent 3D object detection methods [45, 66, 71], we reduce the search space by turning the localization problem into a classification problem.", + "bbox": [ + 496, + 465, + 890, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positional Embedding and Feature Fusion. After the voxelization and 3D encoding process, each 3D token associates with a 3D position $(x,y,z)$ representing the center of its voxel. We first provide positional information to the model by generating learnable positional embeddings (PE) using a two-layer perceptron for each of the $N_v$ visual tokens, and add learnable positional embeddings to the token features $z^{3\\mathrm{D}}$ . We use a situation interpreter [55] to extract situational information, and ask the updated visual tokens to attend to these situational tokens with several transformer layers to produce the joint feature embeddings.", + "bbox": [ + 496, + 713, + 890, + 880 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anchor-based Situational Estimation. We treat each", + "bbox": [ + 500, + 885, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "output token of the feature fusion module as an anchor point, and use it to predict a position likelihood $p \\in [0,1]$ and a rotation estimation. Since each token has an associated 3D position $(x,y,z)$ , the position likelihood $p$ indicates how likely the situational vector locates at the center of this token (voxel). We define a soft ground truth for this classification task with a Gaussian kernel, meaning that the closer a token is to the actual situational vector $s^{\\mathrm{pos}}$ , a higher ground truth probability $p$ will be assigned to that token. In order to counteract the sparse supervisory signal and increase the positive supervision around the vector position, we adopt the peak enlarging technique in CenterPoint [66], where the size of the Gaussian kernel is increased (meaning that the $\\sigma$ is increased) to allow denser supervision around the vector position. Furthermore, we explore different rotation representation and find that compared with quaternion and $(\\sin \\theta, \\cos \\theta)$ representations, 6D vector proposed by [72] achieves the best performance. Hence, we adopt a situational estimation head with MLP layers to output 7-dimensional vector for each of the tokens, where the first channel represents the position likelihood and the other six channels represent the 6D rotation matrix. We take the center of the token with the peak position likelihood as our estimated $s^{\\mathrm{pos}}$ , and convert its corresponding 6D rotation vector as our estimated $s^{\\mathrm{rot}}$ . The estimation can be equivalently represented as a rotation matrix $R$ and a translation matrix $T$ . More discussion about the architecture and design choices is in Section 5.3.", + "bbox": [ + 76, + 90, + 472, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Situation-guided Visual Encoding", + "text_level": 1, + "bbox": [ + 76, + 527, + 372, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the situational estimation, we investigate a better approach to enhancing the downstream response generation, inspired by human cognitive processes. Intuitively, humans typically comprehend their immediate 3D environment by first interpreting their own situation in space, and then discerning their surroundings from an appropriate viewpoint. Our model is designed to emulate this natural strategy. Utilizing the situational vector $\\vec{s}$ , we adjust the coordinate system by repositioning the origin at $s^{\\mathrm{pos}}$ , and reorienting the axes according to $s^{\\mathrm{rot}}$ such that the new y-axis is aligned with the indicated direction. We keep the $z$ -axis vertically oriented, and project the situational vectors onto the $x - y$ plane. This is in line with the format of the dataset [44], where situational vectors are assumed to be parallel with the ground plane. Subsequently, we compute a new situation-guided PE for each of the $N_{v}$ visual tokens, akin to the learnable 3D PE outlined in Section 4.2. They allow the model to grasp positional interrelations from the perspective of the current situation. These situational embeddings are added to the output embeddings of the situational estimation module, which consists of blocks featuring self-attention layers for visual tokens, succeeded by cross-attention layers that bridge visual", + "bbox": [ + 76, + 553, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and situational information. This structure allows for the re-encoding of visual tokens under the influence of situational and question context, guiding the model to assign higher weights to situation-related and question-related visual tokens. The output, termed situation-guided visual tokens, embodies this re-contextualized understanding.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Question Answering Head", + "text_level": 1, + "bbox": [ + 498, + 191, + 741, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow existing methods [26] to use a large vision-language decoder to fuse the final visual and textual tokens and generate textual response to the input question. We explore both auto-regressive response generation, and classification-based answer prediction [5, 44]. For classification, we predict a vector $v^{\\mathrm{ans}} \\in \\mathbb{R}^{n_a}$ for the $n_a$ answer candidates in the training set following [5].", + "bbox": [ + 496, + 215, + 893, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Analysis in 3D VQA Task", + "text_level": 1, + "bbox": [ + 498, + 335, + 740, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SIG3D for 3DVL reasoning on two challenging benchmarks, addressing both visually-oriented situational estimation and textually-focused QA tasks. We present a detailed examination of the implementation strategies adopted, the datasets employed, and the metrics applied in our research. For an exhaustive understanding, implementation and training details, and other additional information are available in the supplementary materials.", + "bbox": [ + 496, + 362, + 893, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on SQA3D [44] and ScanQA [5], two challenging indoor 3D VQA datasets. Both datasets are derived from the ScanNet dataset [13], serving as the foundational source for their 3D scenes. SQA3D features over 33K question-answer pairs for the 3D VQA task and 26K unique situational descriptions for the situational estimation task. Each entry in this dataset includes a 3D scene point cloud, a situational description, a question, and pertinent annotations. ScanQA consists of over 41K question-answer pairs, without situational descriptions and situational annotations. We use it to demonstrate the generalizability of our method on general QA tasks. We use the splits provided by these datasets.", + "bbox": [ + 496, + 488, + 893, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For SQA3D, in order to compare with baseline methods [44, 49, 74], we use a shallow transformer decoder task head to perform answer classification task, and evaluate the performance with exact matches (EM@1), which is equivalent to Top-1 answer accuracy. We also provide EM@1 on a breakdown of question types, including \"What,\" \"Is,\" \"How,\" \"Can,\" \"Which,\" and \"Other,\" based on the first word in the question sentence. Additionally, we evaluate situational estimation performance with localization accuracy and orientation accuracy. In both tasks, we use accuracy within different distance or angle thresholds as our metrics. For example, \"Acc@0.5m\" means accuracy of location estimation when positive threshold is set to 0.5 meter. For ScanQA,", + "bbox": [ + 496, + 688, + 895, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13682", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e968f98110c94f87dcee2bf5fef4d0fa03c6446195059bd2fd2941924ceeadcc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelQuestion BreakdownOverall
WhatIsHowCanWhichOther
GPT-3 [7]39.746.040.545.636.138.441.0
ClipBERT [35]30.260.138.763.342.542.743.3
MCAN [67]28.959.744.168.340.740.543.4
ScanQA [5]28.665.047.366.343.942.945.3
SQA3D [44]33.566.142.469.543.046.447.2
Multi-CLIP [15]------48.0
LM4Vision [49]34.367.148.268.348.945.648.1
3D-LLM [26]36.565.647.268.848.046.348.1
3D-VisTA [74]34.863.345.469.847.248.148.5
SIG3D (Ours)35.667.248.571.449.145.852.6
", + "bbox": [ + 202, + 88, + 772, + 294 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6a88f7ab52cc259c88a72dad97ce0220ef440063a5be4815b758bc87acac0df4.jpg", + "table_caption": [ + "Table 1. Our proposed method SIG3D achieves state-of-the-art performance on SQA3D benchmark [44]. We perform the best on \"Is\", \"How\", and \"Can\" breakdown types of questions, as well as the average accuracy. The results are reported on test set." + ], + "table_footnote": [], + "table_body": "
ModelLocalizationOrientation
Acc@0.5mAcc@1.0mAcc@15°Acc@30°
Random7.225.88.416.9
SQA3D [44]9.529.68.716.5
SQA3D (separate)10.331.417.122.8
3D-VisTA [74]11.734.516.924.2
SIG3D (Ours)27.459.128.742.5
", + "bbox": [ + 81, + 339, + 468, + 450 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff9c6ba01265f79c16e9ea6408657073157a0715824c21bfb36c4f71d480d2ae.jpg", + "table_caption": [ + "Table 2. Our proposed method SIG3D performs significantly better than prior method [44] in situational estimation task. \"Acc@0.5m\" stands for localization accuracy with $0.5\\mathrm{m}$ threshold. \"Acc@15°\" represents orientation accuracy with $15^{\\circ}$ threshold. separate means disabling other tasks to let the model focus on situational estimation only." + ], + "table_footnote": [], + "table_body": "
ModelBLEU-1BLEU-4ROUGEMETEORCIDEr
BLIP2 [38]29.75.926.611.345.7
Flamingo [3]25.68.431.111.355.0
VN+MCAN [67]28.06.229.811.454.7
SR+MCAN [67]26.97.930.011.555.4
ScanQA [5]30.210.133.313.164.9
3D-LLM [26]39.312.035.714.569.4
SIG3D39.512.435.913.468.8
", + "bbox": [ + 81, + 547, + 468, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Performance of SIG3D on ScanQA dataset [5] is on-par with the state-of-the-art with large-scale text-3D pre-training. VN and SR stand for VoteNet and ScanRefer, respectively. 3D-LLM [74] leverages pretrained 2DVL foundation models and LLM models [3, 7, 37, 38], and is pretrained on a large-scale held-in 3D-text dataset before the finetuning on ScanQA.", + "bbox": [ + 75, + 669, + 468, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we perform auto-regressive answer generation with large transformer decoder [26], and evaluate with BLEU [50], ROUGE [39], METEOR [6], and CIDEr [60] metrics.", + "bbox": [ + 75, + 773, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Situated Question Answering", + "text_level": 1, + "bbox": [ + 76, + 830, + 339, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Our study involves a comparative analysis with a range of representative baselines on the SQA3D dataset. In particular, we evaluate against GPT-3 [7], Clip", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BERT [35], and MCAN [67], which are, as reported in prior work [44], baselines focused on language-only, 2D video, and 2D image QA, respectively. For GPT-3, we follow SQA3D [44] to convert the visual input into a caption using Scan2Cap [10] for LLMs to process. ScanQA [5] represents a 3D QA baseline that ignores the situational input. Both SQA3D [44] and Multi-CLIP [15] employ situational descriptions and annotations for direct regression tasks. LM4Vision [49] utilizes LLMs as visual and textual encoders. Additionally, 3D-VisTA [74] undergoes a pretraining procedure on their large-scale 3D scene-text dataset, ScanScribe, prior to the finetuning on this dataset.", + "bbox": [ + 496, + 342, + 892, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Situational Estimation. As shown in Table 2, our work performs significantly better than the state of the art [44, 74] in both localization and orientation estimation tasks. For 3D-VisTA [74], we use a pretrained model and finetune a new situation head with the SQA3D dataset following [44]. We also report a random baseline, in which we randomly sample position and orientation from a uniform distribution as a lower-bound performance. Note that original SQA3D performs only marginally better than the random baseline, meaning that it does not acquire any situational awareness, despite having the situational estimation loss. Disabling the QA task and asking the model to exclusively focus on the situational estimation task results in a slight better performance. Our method, with the anchor-based position likelihood estimation, results in a much better understanding of the 3D situational relationship. Our method also outperforms 3D-VisTA, which is pretrained on a large-scale 3D-text dataset, indicating that large pretraining alone is not enough to address the situational awareness problem. Note that we do not include the random baseline performance reported in [44], because each value is obtained by generating three random values and taking the closest one to the ground true, and thus it does not reflect a true \"random\" baseline.", + "bbox": [ + 496, + 530, + 893, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Situated Question Answering. SIG3D outperforms prior", + "bbox": [ + 500, + 885, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/521f1b5e69b3c801641ff85268a47c78db0e7b2a8c435984687d245a2c50d981.jpg", + "table_caption": [ + "(a) Number of Visual Tokens" + ], + "table_footnote": [], + "table_body": "
Acc@1.0mAcc@30°EM@1
12848.938.249.2
25659.142.550.9
51257.842.150.7
", + "bbox": [ + 96, + 112, + 359, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2800257079b5c6dfeadd5dae2322d175306ff92ab071c2da7090af59beb629e.jpg", + "table_caption": [ + "(b) Voxel size (in meters)" + ], + "table_footnote": [], + "table_body": "
Acc@1.0mEM@1
0.0154.149.5
0.0259.150.9
0.0547.348.8
", + "bbox": [ + 406, + 112, + 599, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e92edbac765b99b3df55a336252871ba6062f618745bcb82cb54048c5b19d24.jpg", + "table_caption": [ + "(c) Rotation representation" + ], + "table_footnote": [], + "table_body": "
Acc@30°EM@1
Quaternion31.450.0
6D vector42.550.9
sinθ, cosθ42.650.6
", + "bbox": [ + 640, + 112, + 874, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fddfa3960749c3c2ed36d9a1bc4cdf7f1e2569f74a268f73ffeb0238998da4a1.jpg", + "table_caption": [ + "Table 4. Ablation study validates that our various design choices improves the performance. \"Acc@1.0m\", \"Acc@30°\", and \"EM@1\" are accuracy (%) for localization estimation, orientation estimation, and QA tasks. Our settings are marked in gray." + ], + "table_footnote": [ + "Table 5. Performance of SIG3D improves with stronger visual and language encoders. We find that open-vocabulary point encoder and MPNet-based sentence BERT leads to best performance. \"Acc@1.0m\" and \"Acc@30°\" stands for localization and orientation accuracy in situational estimation task. \"EM@1\" demonstrate exact match metric in QA task." + ], + "table_body": "
Acc@1.0mAcc@30°EM@1
3D Vision Encoders
Text-only (no vision input)--47.5
VoteNet [53]37.428.249.1
3DETR [48]47.229.149.4
OpenScene - OpenSeg [51]57.541.650.2
OpenScene - LSeg [51]59.142.550.9
Language Tokenizer / Encoders
GloVe + LSTM [24, 52]44.330.948.7
SBERT - MiniLM [55]56.138.649.4
SBERT - MPNet [55]55.940.649.7
SBERT - MPNet (finetune)59.142.550.9
", + "bbox": [ + 81, + 234, + 468, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods in most question breakdown categories and overall accuracy, as shown in Table 1. Our work achieves leading results without large-scale pretraining (compared with 3D-VisTA) and LLM (compared with GPT-3), indicating its superiority in situational awareness. Note that the LLM baseline GPT-3 achieves the best performance on the \"What\" category, suggesting the potential of stronger language encoder in interpreting the complicated question prompt.", + "bbox": [ + 75, + 515, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. General Question Answering on ScanQA", + "text_level": 1, + "bbox": [ + 76, + 655, + 426, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare with 2D image VQA MCAN-based baselines [67], ScanQA [5], 3D-LLM [26] which leverages large-scale pretrained 2D VLMs and LLMs as backbone models, and 3D-VisTA [74] pretrained on their proposed large-scale 3D-text dataset.", + "bbox": [ + 75, + 680, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Question Answering. As shown in Table 3, despite that the questions do not explicitly require situational understanding to answer in ScanQA, SIG3D achieves comparable results with state-of-the-art methods without the large-scale 3D-text pretraining and powerful 2D VLM and LLM backbone models. Our work pretrained on SQA3D [44] leads to higher performance on BLEU-1, BLEU-4, and ROUGE metrics, showing its generalizability on general 3D QA scenarios.", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bd148491d7663bdf3fe91fe0dcac67f9e9289ddfe1b3b17671d20a50a2213b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Acc@1.0mAcc@30°EM@1
Baseline (joint optimization)29.523.147.7
How to achieve better situational estimation
+ 3D PE38.823.647.8
+ 6D Representation38.527.447.7
+ Anchor-based Estimation58.841.948.2
How to utilize situational estimation for better QA
+ 3D Situational PE58.941.850.0
+ Visual Token Re-encoding59.142.550.9
Oracle Models (Ground Truth Situation Information)
Situation as direct input10010047.7
Situation as intermediate input10010053.9
", + "bbox": [ + 504, + 234, + 890, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study verifies that our proposed modules leads to better situational estimation and better QA performance.", + "bbox": [ + 500, + 398, + 890, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study & Analysis", + "text_level": 1, + "bbox": [ + 500, + 448, + 741, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vision & Language Encoders. We study the impacts of different visual and textual tokenizers in Table 5. It is observed that the open-vocabulary visual encoder (OpenScene) outperforms detection-based encoders (such as VoteNet and 3DETR) across all metrics. This superior performance of OpenScene is attributed to the limitations of 3D detectors, which are typically trained on a limited set of object categories, rendering them less effective in recognizing novel objects mentioned in textual prompts. Regarding language encoders, our findings indicate that a stronger backbone correlates with better performance, primarily due to its improved capability to interpret complex textual inputs. This leads to the suggestion of integrating LLM with our method to potentially further enhance performance, an avenue we intend to explore in future research.", + "bbox": [ + 496, + 470, + 890, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Situated Awareness. In Table 6 we verify the crucial role of situational awareness in 3DVL task. Firstly, we show that 3D PE, 6D rotation estimation, and anchor-based position estimation all leads to much better position and orientation estimation performance. We further establish that situational PE and visual token re-encoding modules lead to better utilization of the predicted situational vector for QA task. Additionally, We design two oracle models under the assumption of having access to the ground truth situational vector as an input. The outcomes from these models reveal a critical insight: the model fails to effectively interpret situational information when it is directly incorporated into the input visual embeddings. This underlines the necessity", + "bbox": [ + 496, + 703, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13684", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d3791acd66cf36a696504a6298f47bf9b90e9459e4b89094fb7fa4bcc5a5786.jpg", + "image_caption": [ + "Figure 5. Qualitative results demonstrate significant improvement of SIG3D over prior method. The first row is results from SQA3D [44], and the second row is results from our methods. In the 3D scene, red: Ground truth (GT) vector, and blue: Estimated vector." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the intermediate representation and encoding mechanism we have proposed, affirming its importance in achieving optimal 3DVL task performance.", + "bbox": [ + 75, + 479, + 468, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architectural Design. We explore different architectural design choices of our model in Table 4. We find that the number of visual tokens sampled from the visual feature embeddings affects the performance of both situational estimation and QA tasks. Sampling fewer visual tokens increases the risk of missing the region of significance, while sampling more does not lead to a better performance as well. We study the size of voxels and find $0.02\\mathrm{m}$ to be the most effective choice, as OpenScene [51] backbone is pretrained with the same voxel size. We also find $\\sin \\theta$ , $\\cos \\theta$ and 6D vector representation performs a lot better than quaternion in rotation estimation task. This is consistent with the finding reported in [72].", + "bbox": [ + 75, + 530, + 470, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 739, + 269, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we demonstrate some qualitative results of our SIG3D in Figure 5. We show the ground truth and estimated situational vectors in red and blue colors, respectively, in their corresponding 3D scenes. We also print the answers with red cross or green checkmark indicating the correctness. It is clear that our method performs significantly better in situational estimation task, resulting in vectors very close to the ground truth in both position and orientation perspectives. The better situational awareness", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also aids the complicated embodied navigation and common sense QA activities. This also demonstrate great potential of our method in the development of indoor robotics and/or conversational agents.", + "bbox": [ + 496, + 479, + 890, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Supplementary Material. The supplementary section offers an extensive analysis, encompassing a detailed examination of the 3D visual token activation changes pre and post situational re-encoding. Additionally, it includes a comprehensive collection of positive and negative samples, an insightful failure case analysis, and a forward-looking discussion on limitations and future work.", + "bbox": [ + 496, + 544, + 892, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SIG3D, a situation-aware vision language model for 3D reasoning tasks. We propose to represent 3D scene as feature tokens, treat tokens as anchors points to estimate a situational vector from a text description, and use the estimated situation as guidance to align and re-encode the visual tokens to enhance the features for reasoning tasks. We observe consistent and significant performance gain on both situational estimation and question answering tasks.", + "bbox": [ + 496, + 688, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13685", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. ShapeGlot: Learning language for shape differentiation. In ICCV, 2019. 3", + "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In ECCV, 2020. 3", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurlPS, 2022. 1, 2, 6", + "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 2", + "[5] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In CVPR, 2022. 2, 3, 5, 6, 7", + "[6] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005. 6", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1, 2, 3, 6", + "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3D object localization in rgb-d scans using natural language. In ECCV, 2020. 3", + "[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In ACCV, 2019. 3", + "[10] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, 2021. 3, 6", + "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1", + "[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In CVPR, 2019. 4, 1", + "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 4, 5", + "[14] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. In CVPR, 2022. 2", + "[15] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann," + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Thomas Hofmann. Multi-CLIP: Contrastive Vision-Language Pre-training for Question Answering tasks in 3D Scenes. In BMVC, 2023. 3, 6", + "[16] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In CVPR, 2023. 3", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2", + "[18] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In CVPR, 2022. 2", + "[19] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3D visual graph network for object grounding in point cloud. In ICCV, 2021. 3", + "[20] François Gardères, Maryam Ziaeefard, Baptiste Abeloos, and Freddy Lecue. Conceptbert: Concept-aware representation for visual question answering. In EMNLP, 2020. 2", + "[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2", + "[22] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, and Jianfeng Gao. Kat: A knowledge augmented transformer for vision-and-language. In NAACL, 2022. 2", + "[23] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From Images to Textual Prompts: Zero-shot Visual Question Answering with Frozen Large Language Models. In CVPR, 2023. 2", + "[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 7", + "[25] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3D Concept Learning and Reasoning from Multi-View Images. In CVPR, 2023. 3", + "[26] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3D World into Large Language Models. NeurIPS, 2023. 1, 2, 3, 5, 6, 7", + "[27] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3D instance segmentation. In AAAI, 2021. 3", + "[28] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. LEMMA: A Multi-view Dataset for LE arning Multi-agent M multi-task A civities. In ECCV, 2020. 2", + "[29] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In NeurIPS, 2022. 2", + "[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13686", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2", + "[31] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL, 2019. 1, 2", + "[32] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In iccv, 2023. 3", + "[33] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2", + "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2", + "[35] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021. 6", + "[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven Semantic Segmentation. In ICLR, 2022. 2", + "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6", + "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ICML, 2023. 2, 6, 1", + "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 6", + "[40] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang. Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In CVPR, 2023. 2, 1", + "[41] Yuanze Lin, Yujia Xie, Dongdong Chen, Yichong Xu, Chenguang Zhu, and Lu Yuan. Revive: Regional visual representation matters in knowledge-based visual question answering. In NeurIPS, 2022. 2", + "[42] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In CVPR, 2023. 1", + "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1", + "[44] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: Situated question answering in 3D scenes. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[45] Yunze Man, Liang-Yan Gui, and Yu-Xiong Wang. BEV-Guided Multi-Modality Fusion for Driving Perception. In CVPR, 2023. 4", + "[46] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Jack Merullo, Louis Castricato, Carsten Eickhoff, and Ellie Pavlick. Linearly mapping from image to text space. *ICLR*, 2023. 2", + "[48] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, 2021. 7", + "[49] Ziqi Pang, Ziyang Xie, Yunze Man, and Yu-Xiong Wang. Frozen Transformers in Language Models Are Effective Visual Encoder Layers. arXiv preprint arXiv:2310.12973, 2023. 2, 5, 6", + "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6", + "[51] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3, 4, 7, 8, 1", + "[52] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 7", + "[53] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, 2019. 3, 7", + "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 4", + "[55] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In EMNLP, 2019. 4, 7, 1", + "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[57] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, 2022. 2", + "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2", + "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2, 1", + "[60] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015. 6", + "[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 1", + "[62] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In NeurIPS (Datasets and Benchmarks Track), 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13687", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Jialin Wu, Jiasen Lu, Ashish Sabharwal, and Roozbeh Mottaghi. Multi-modal answer validation for knowledge-based vqa. In AAAI, 2022. 2", + "[64] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. LLM-Grounder: Open-Vocabulary 3D Visual Grounding with Large Language Model as an Agent. arXiv preprint arXiv:2309.12311, 2023. 3", + "[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3D question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3", + "[66] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In CVPR, 2021. 4, 5", + "[67] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, 2019. 6, 7, 1", + "[68] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In CVPR, 2019. 2", + "[69] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2", + "[70] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. MotionGPT: Finetuned LLMs are General-Purpose Motion Generators. arXiv preprint arXiv:2306.10900, 2023. 2", + "[71] Brady Zhou and Philipp Krahenbuhl. Cross-view Transformers for real-time Map-view Semantic Segmentation. In CVPR, 2022. 4", + "[72] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 5, 8", + "[73] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In CVPR, 2016. 2", + "[74] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment. In ICCV, 2023. 3, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 674 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13688", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_model.json b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e206de03e5df57281a79b9461a8616ec5a63324e --- /dev/null +++ b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.817, + 0.154 + ], + "angle": 0, + "content": "Situational Awareness Matters in 3D Vision Language Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.181, + 0.701, + 0.218 + ], + "angle": 0, + "content": "Yunze Man Liang-Yan Gui Yu-Xiong Wang University of Illinois Urbana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.22, + 0.621, + 0.234 + ], + "angle": 0, + "content": "{yunzem2,lgui,yxw}@illinois.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Being able to carry out complicated vision language reasoning tasks in 3D space represents a significant milestone in developing household robots and human-centered embodied AI. In this work, we demonstrate that a critical and distinct challenge in 3D vision language reasoning is the situational awareness, which incorporates two key components: (1) The autonomous agent grounds its self-location based on a language prompt. (2) The agent answers open-ended questions from the perspective of its calculated position. To address this challenge, we introduce SIG3D, an end-to-end Situation-Grounded model for 3D vision language reasoning. We tokenize the 3D scene into sparse voxel representation, and propose a language-grounded situation estimator, followed by a situated question answering module. Experiments on the SQA3D and ScanQA datasets show that SIG3D outperforms state-of-the-art models in situational estimation and question answering by a large margin (e.g., an enhancement of over \\(30\\%\\) on situation accuracy). Subsequent analysis corroborates our architectural design choices, explores the distinct functions of visual and textual tokens, and highlights the importance of situational awareness in the domain of 3D question-answering. Project page is available at https://yunzeman.github.io/situation3d." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.885 + ], + "angle": 0, + "content": "Humans learn knowledge efficiently through the interactions with the 3D world and the integration of multi-modal information, such as verbal guidance or instructions. Similarly, introducing language guidance into the visual comprehension task can greatly enhance models' learning efficiency [3, 42]. Nonetheless, despite considerable advancements in linguistic understanding [7, 11, 31, 58] and vision-language integration [3, 37, 54, 61], current methodologies remain deficient in accurately perceiving and rationalizing within real-world 3D environments, which is largely attributed to the lack of 3D situational reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Compared with machine learning models, humans put" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.269, + 0.891, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.576 + ], + "angle": 0, + "content": "Figure 1. Previous methods perform direct 3D vision language reasoning without modeling the situation of an embodied agent in the 3D environment. Our method, SIG3D, grounds the situational description in the 3D space, and then re-encodes the visual tokens from the agent's intended perspective before vision-language fusion, resulting in a more comprehensive and generalized 3D vision language (3DVL) representation and reasoning pipeline. Q, K, V stands for query, key, and value, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.746 + ], + "angle": 0, + "content": "themselves inside the 3D world and then perceive and interact with the surrounding environment from their egoperspective (Figure 1). Such situational awareness is a crucial difference between 2D and 3D visual understanding, and a key to achieve seamless understanding of spatial concepts in more complex real-world environments. Several existing methods recognize the lack of positional understanding in 3D and propose new benchmarks and joint optimization functions [44], or positional embedding methods [26] to enhance the overall reasoning performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "However, the lack of an explicit situation modeling and situation-grounded 3D reasoning method restricts them from obtaining a generalizable and consistent 3D vision-language (VL) representation. As shown in Figure 2, the situation prediction of the state-of-the-art method [44] (in blue) diverges significantly from the ground truth vectors (in red) in almost all scenes in the dataset [13]. Moreover, our pilot study in Section 3 also reveals that situational understanding, despite being very crucial in comprehending the context of questions, only plays a minor role in the final" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.095, + 0.26, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.196, + 0.258, + 0.216 + ], + "angle": 0, + "content": "I am facing a trash can, while there is a door on my right." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.095, + 0.449, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.196, + 0.453, + 0.215 + ], + "angle": 0, + "content": "I am facing the paper towel dispenser with a chair in my six o'clock direction." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.228, + 0.258, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.308, + 0.255, + 0.337 + ], + "angle": 0, + "content": "I am facing a pool table, taking out the white ball from the side pocket. I am also facing a painting on the wall." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.225, + 0.439, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.316, + 0.459, + 0.336 + ], + "angle": 0, + "content": "I am sitting on a couch with a pillow facing another couch and the pillow is on my right." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.406 + ], + "angle": 0, + "content": "Figure 2. Situational estimation in existing methods [44] fails in most scenarios, indicating the missing registration between the situational description and 3D embeddings. Red: Ground truth (GT) vector. Blue: Estimated vector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.433, + 0.44 + ], + "angle": 0, + "content": "question-answering performance of existing methods." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.442, + 0.47, + 0.745 + ], + "angle": 0, + "content": "In this work, we propose SIG3D, a novel approach designed to precisely model and estimate an embodied agent's ego-location and orientation from a textual description, before performing multi-modal QA tasks from the agent's egocentric perspective, as shown in Figure 1. Specifically, we leverage large-scale pretrained language and visual encoders to process the input text and 3D data, and fuse the tokens with attention modules to predict a situational vector. Previous attempts to directly predict the ego-situation are hindered by the expansive search space inherent in 3D environments. To address this challenge, we re-conceptualize the task as an anchor-based classification, where visual tokens are regarded as anchor points, and a likelihood of position together with a set of rotation parameters are concurrently regressed for each visual token. After obtaining the situational estimation, we propose a situational alignment and a situation-guided token re-encoding strategy, to perceive the environment from the agent's intended perspective. These strategies enhance the visual token with more accurate situational awareness for subsequent QA tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.468, + 0.868 + ], + "angle": 0, + "content": "Experiments on two challenging 3D visual question-answering (VQA) datasets [5, 44] demonstrate the significant improvement on situational estimation and QA tasks of our model. In particular, we improve the situational estimation accuracy by over \\(30\\%\\), and the subsequent QA performance by up to \\(3\\%\\). Further qualitative and quantitative analysis verifies our design choices and highlights the significance of situational awareness in 3D reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "To sum up, our paper has the following contributions: (1) We recognize the lack of situational awareness as a sig-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.319 + ], + "angle": 0, + "content": "nificant oversight in existing research. To address this, we introduce SIG3D, a situation-grounded 3D VL reasoning architecture, specifically designed to fill this void. (2) We propose an anchor-based approach to situational estimation, which effectively narrows the extensive search space in 3D environments for precise grounding of 3D positions and orientations with textual descriptions. Additionally, we investigate situational alignment and visual re-encoding mechanisms to leverage situational awareness for enhanced QA performance. (3) Our model demonstrates superior performance on two challenging datasets, SQA3D and ScanQA, surpassing the state of the art in both situational estimation and QA metrics. Ablation studies highlight the importance of situation-guided encoding, revealing its beneficial impact on general QA tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.337, + 0.642, + 0.354 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Vision Language Models (VLMs). Early transformer-driven [59] textual and visual encoders [17, 31] have facilitated great progress in recent vision language learning. Text-image contrastive models [30, 54] propose to align the feature space of two modalities with large-scale pretraining, fueling numerous downstream tasks from generalized open-vocabulary visual perception [21, 34, 36] to text-to-image generation [56]. Concurrently, some work uses text and vision encoders on separate modalities followed by feature fusion [18, 33] for multi-modal reasoning tasks. Since the emergence of Large Language Models (LLMs) [7, 58, 69], VLMs have experienced huge improvement with the help of LLMs as building blocks for multi-modal learning architectures. Specifically, recent work directly projects visual embeddings into language-space tokens as input to LLM [23, 40, 47, 70], or use the latent bottleneck structure for cross-modal visual decoding [3, 26, 37, 38], or treat LLM layers as encoder blocks for various visual tasks [49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.789 + ], + "angle": 0, + "content": "In the domain of visual question-answering (VQA) [4, 73], recent work has pushed the frontier towards video understanding [14, 28, 29, 40, 62], knowledge-based understanding [20, 22, 23, 41, 46, 57, 63], and commonsense reasoning [68]. Despite the outstanding performance in 2D image interpretation, most existing methods lack the capability to generalize to 3D scenarios. In contrast, our work studies the representation of visual information and its fusion with language embeddings in the 3D domain by targeting on the 3D situation-guided visual language interpretation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Grounding Language in 3D Space. Compared with 2D images, knowledge such as spatial relationships, interactive exploration, and topological analysis, which only exist in the 3D world provides additional challenges and opportunities to develop better language models with stronger commonsense reasoning capability grounded in the real-world 3D scenarios. In this direction, early work seeks" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.186, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Figure 3. Results on variants of the representative SQA3D baseline method [44] demonstrate that situational understanding, despite being indispensable in perceiving the context of questions, makes negligible contribution in existing methods. This motivates a situation-guided 3D encoding mechanism in our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.468, + 0.392 + ], + "angle": 0, + "content": "to ground isolated objects [1, 9] or objects within more complex scenes [2, 8, 19, 27] using natural language descriptions. Recently, with more collected 3D vision language datasets, several work starts to explore language-guided 3D visual interpretation and reasoning on a diverse set of datasets, including 3D scene captioning [10], open-vocabulary segmentation [16, 32, 51], and question-answering [5, 15, 25, 65, 74]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.468, + 0.56 + ], + "angle": 0, + "content": "The success of LLMs also elicits the usage of them in 3D vision language reasoning for task decomposition [64], data generation, and multi-modal feature fusion [26]. Motivated by ScanQA [5], SQA3D [44] takes the first step into exploring the challenging 3D situational reasoning problem by developing a situated question-answering benchmark, and proposing the first joint learning baseline on the benchmark. Our work highlights the uniqueness and significance of situational awareness in the 3D vision language learning paradigm, leading to notably better 3D situational grounding and question-answering performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.388, + 0.59 + ], + "angle": 0, + "content": "3. Pilot Study on Situated Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Despite pointing out the significance of situated understanding and reasoning, existing methods [44] fall short in providing effective situational estimation, as illustrated in Figure 2. This section delves into a pilot study examining the impact of situational understanding on downstream reasoning tasks. The SQA3D baseline [44] incorporates situational descriptions and uses ground truth (GT) situational vectors for supervision in a direct regression task. We investigate three variants of this baseline to assess the effect of situational understanding. In the first variant, we remove the situational description and supervision from the model, by passing in empty situational tokens. In another variant, we corrupt the situation supervision by introducing very large Gaussian noise to the GT vectors to effectively randomizing them. Finally, we try to encode the GT situational vector in the input with learnable multi-layer perceptron layers, which are then integrated with the visual and textual tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Figure 3 demonstrates the results of this study, revealing negligible changes in performance across these variants. Notably, corrupting the GT situational information or di" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "rectly incorporating it results in only marginal alterations in the QA outcomes. Omitting the situational description entirely from the input results in a mere \\(2\\%\\) decline in accuracy. However, in the absence of this information, the model resorts to random guessing when determining the correct answer, as all responses depend on situational context. The findings from Figures 2 and 3 collectively indicate a deficiency in existing methods regarding situation vector estimation and the application of situational understanding in subsequent reasoning tasks. These unresolved challenges motivate the development of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.274, + 0.591, + 0.289 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.527 + ], + "angle": 0, + "content": "An overview of our approach SIG3D is illustrated in Figure 4. Our method begins with a set of points that represent a 3D scene, accompanied by a situational description and a question, which defines the overall context of the problem. We tokenize them into separate token embeddings (Section 4.1), and ground the textual description in the 3D scene with a vector comprising of location and orientation. We find direct single vector estimation to be challenging due to the vast and complex nature of the 3D search space, so we propose an anchor-based situational estimation strategy (Section 4.2). Subsequently, we re-encode the visual tokens from the perspective of situational vectors, enhancing the situational awareness for downstream reasoning tasks (Section 4.3). The finalized visual and textual tokens are fused by a transformer decoder to generate the final response." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.539, + 0.783, + 0.554 + ], + "angle": 0, + "content": "4.1. Visual and Textual Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.805 + ], + "angle": 0, + "content": "Leveraging input scene point clouds and textual prompts, our objective is to generate three distinct types of tokens: 3D visual tokens \\( z^{3\\mathrm{D}} \\in \\mathbb{R}^{N_v \\times C_v} \\), situational tokens \\( z^{\\mathrm{S}} \\in \\mathbb{R}^{N_s \\times C_s} \\), and question tokens \\( z^{\\mathrm{Q}} \\in \\mathbb{R}^{N_q \\times C_q} \\). Each type of token is characterized by two primary components: \\( N \\), representing the number of tokens, and \\( C \\), encapsulating the feature embeddings. Tockenize and capture the feature embeddings for situational and question inputs, we employ a shared text tokenizer \\( \\mathrm{E}^{\\mathrm{TXT}} \\) following prior methods [5, 44]. We assume that situational and question prompts are separated in the input data. If not, LLMs [7] can be used to parse the textual input without changing the semantic meaning of the sentences. However, there is a lack of consensus on a standard 3D visual tokenization method \\( \\mathrm{E}^{3\\mathrm{D}} \\) that is apt for the 3DVL reasoning task, prompting a more detailed exploration in the subsequent paragraphs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visual Tokenization. Given an input point cloud \\(\\mathbf{p} \\in \\mathbb{R}^{N \\times 3}\\), most prior methods [5, 15, 44] adopt a VoteNet [53] detector to acquire object-level tokens \\(z^{3\\mathrm{D}} \\in \\mathbb{R}^{N_{\\mathrm{obj}} \\times C_{\\mathrm{obj}}}\\) as the visual representation, where \\(N_{\\mathrm{obj}}\\) is the number of object proposals, and \\(C_{\\mathrm{obj}}\\) is the object-level feature embeddings. However, we point out several problems with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13680" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.077, + 0.089, + 0.898, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.895, + 0.422 + ], + "angle": 0, + "content": "Figure 4. Overview of our SIG3D model, which includes 3D scene and text encoding, anchor-based situational estimation, situation-guided visual re-encoding, and multi-modal decoder modules. We tokenize the 3D scene into voxels, treat each token as an anchor point, and query the text tokens to predict a token-level position likelihood and rotation matrix to locate the situational vector associated with the text descriptions. Then we update the scene tokens with situational position encoding (PE), and finally perform the 3DVL reasoning task with a large transformer decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.626 + ], + "angle": 0, + "content": "this abstraction strategy: (1) A detection-based tokenization method tends to ignore the non-object regions in the scene, which can be indispensable in some reasoning scenarios (e.g., carpets on the ground, ceiling, walls.) (2) After object-level abstraction, the visual representation loses the high-level information of the scene (e.g., the shape of the living room, the corner of the kitchen.) (3) A supervised detector trained from scratch can only recognize objects within the training set (e.g., only 20 categories for ScanNet [13]), meaning that the method does not have zero-shot capability to reason about novel unseen objects that are inevitably common in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.692 + ], + "angle": 0, + "content": "In light of this, we adopt a pretrained open-vocabulary voxel-based tokenization method from OpenScene [51]. The scene is first discretized into regular small 3D voxels and fed into a visual encoder for feature extraction:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.696, + 0.47, + 0.715 + ], + "angle": 0, + "content": "\\[\nx ^ {\\mathrm {3 D}} = \\operatorname {E} ^ {\\mathrm {3 D}} (\\mathcal {V} (\\mathbf {p})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}\\) represents the voxelization process, and \\(\\mathrm{E}^{3\\mathrm{D}}\\) is a Minkowski sparse 3D convolutional network [12]. The sparse network is pretrained by distillation from CLIP [54] embeddings of rendered multi-view 2D images, resulting in a feature map with better language alignment and 3D awareness. We take the upsampled bottleneck-layer feature embeddings from the encoder network, and compute the mean average over the \\(z\\)-axis (vertical) to project the voxels onto the \\(x - y\\) plane and treat the feature grids in the resulting 2D feature map as our \\(N_{v}\\) visual tokens. We find that this bird's-eye-view projection results in a more compact representation and improves the final performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.444, + 0.71, + 0.459 + ], + "angle": 0, + "content": "4.2. Situational Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.71 + ], + "angle": 0, + "content": "Given 3D visual tokens \\( z^{3\\mathrm{D}} \\) and situational tokens \\( z^{\\mathrm{S}} \\), our objective is to estimate the situational vector \\( \\vec{s} \\) referred by the situational description, which comprises of a position component \\( s^{\\mathrm{pos}} \\) represented by \\( (x,y,z) \\) coordinates, and a rotation component \\( s^{\\mathrm{rot}} \\) represented by \\( (\\theta ,\\psi ,\\phi) \\) Euler angles, where pitch angles \\( \\psi \\) are always defined as 0, meaning that situational vectors are defined to be parallel with the ground plane. The prior method [44] utilizes a transformer block to calculate the cross-attention feature between visual and language tokens, and directly regress a final situational vector out of the averaged attention map. We find such strategy producing very inaccurate estimation, as shown in Figure 2, due to the large search space in the entire 3D volume. Inspired by the recent 3D object detection methods [45, 66, 71], we reduce the search space by turning the localization problem into a classification problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.881 + ], + "angle": 0, + "content": "Positional Embedding and Feature Fusion. After the voxelization and 3D encoding process, each 3D token associates with a 3D position \\((x,y,z)\\) representing the center of its voxel. We first provide positional information to the model by generating learnable positional embeddings (PE) using a two-layer perceptron for each of the \\(N_v\\) visual tokens, and add learnable positional embeddings to the token features \\(z^{3\\mathrm{D}}\\). We use a situation interpreter [55] to extract situational information, and ask the updated visual tokens to attend to these situational tokens with several transformer layers to produce the joint feature embeddings." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Anchor-based Situational Estimation. We treat each" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13681" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.514 + ], + "angle": 0, + "content": "output token of the feature fusion module as an anchor point, and use it to predict a position likelihood \\( p \\in [0,1] \\) and a rotation estimation. Since each token has an associated 3D position \\( (x,y,z) \\), the position likelihood \\( p \\) indicates how likely the situational vector locates at the center of this token (voxel). We define a soft ground truth for this classification task with a Gaussian kernel, meaning that the closer a token is to the actual situational vector \\( s^{\\mathrm{pos}} \\), a higher ground truth probability \\( p \\) will be assigned to that token. In order to counteract the sparse supervisory signal and increase the positive supervision around the vector position, we adopt the peak enlarging technique in CenterPoint [66], where the size of the Gaussian kernel is increased (meaning that the \\( \\sigma \\) is increased) to allow denser supervision around the vector position. Furthermore, we explore different rotation representation and find that compared with quaternion and \\( (\\sin \\theta, \\cos \\theta) \\) representations, 6D vector proposed by [72] achieves the best performance. Hence, we adopt a situational estimation head with MLP layers to output 7-dimensional vector for each of the tokens, where the first channel represents the position likelihood and the other six channels represent the 6D rotation matrix. We take the center of the token with the peak position likelihood as our estimated \\( s^{\\mathrm{pos}} \\), and convert its corresponding 6D rotation vector as our estimated \\( s^{\\mathrm{rot}} \\). The estimation can be equivalently represented as a rotation matrix \\( R \\) and a translation matrix \\( T \\). More discussion about the architecture and design choices is in Section 5.3." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.529, + 0.373, + 0.545 + ], + "angle": 0, + "content": "4.3. Situation-guided Visual Encoding" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.554, + 0.473, + 0.903 + ], + "angle": 0, + "content": "After obtaining the situational estimation, we investigate a better approach to enhancing the downstream response generation, inspired by human cognitive processes. Intuitively, humans typically comprehend their immediate 3D environment by first interpreting their own situation in space, and then discerning their surroundings from an appropriate viewpoint. Our model is designed to emulate this natural strategy. Utilizing the situational vector \\(\\vec{s}\\), we adjust the coordinate system by repositioning the origin at \\(s^{\\mathrm{pos}}\\), and reorienting the axes according to \\(s^{\\mathrm{rot}}\\) such that the new y-axis is aligned with the indicated direction. We keep the \\(z\\)-axis vertically oriented, and project the situational vectors onto the \\(x - y\\) plane. This is in line with the format of the dataset [44], where situational vectors are assumed to be parallel with the ground plane. Subsequently, we compute a new situation-guided PE for each of the \\(N_{v}\\) visual tokens, akin to the learnable 3D PE outlined in Section 4.2. They allow the model to grasp positional interrelations from the perspective of the current situation. These situational embeddings are added to the output embeddings of the situational estimation module, which consists of blocks featuring self-attention layers for visual tokens, succeeded by cross-attention layers that bridge visual" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "and situational information. This structure allows for the re-encoding of visual tokens under the influence of situational and question context, guiding the model to assign higher weights to situation-related and question-related visual tokens. The output, termed situation-guided visual tokens, embodies this re-contextualized understanding." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.193, + 0.742, + 0.21 + ], + "angle": 0, + "content": "4.4. Question Answering Head" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.216, + 0.895, + 0.324 + ], + "angle": 0, + "content": "We follow existing methods [26] to use a large vision-language decoder to fuse the final visual and textual tokens and generate textual response to the input question. We explore both auto-regressive response generation, and classification-based answer prediction [5, 44]. For classification, we predict a vector \\( v^{\\mathrm{ans}} \\in \\mathbb{R}^{n_a} \\) for the \\( n_a \\) answer candidates in the training set following [5]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.741, + 0.354 + ], + "angle": 0, + "content": "5. Analysis in 3D VQA Task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.484 + ], + "angle": 0, + "content": "We evaluate SIG3D for 3DVL reasoning on two challenging benchmarks, addressing both visually-oriented situational estimation and textually-focused QA tasks. We present a detailed examination of the implementation strategies adopted, the datasets employed, and the metrics applied in our research. For an exhaustive understanding, implementation and training details, and other additional information are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.895, + 0.685 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on SQA3D [44] and ScanQA [5], two challenging indoor 3D VQA datasets. Both datasets are derived from the ScanNet dataset [13], serving as the foundational source for their 3D scenes. SQA3D features over 33K question-answer pairs for the 3D VQA task and 26K unique situational descriptions for the situational estimation task. Each entry in this dataset includes a 3D scene point cloud, a situational description, a question, and pertinent annotations. ScanQA consists of over 41K question-answer pairs, without situational descriptions and situational annotations. We use it to demonstrate the generalizability of our method on general QA tasks. We use the splits provided by these datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. For SQA3D, in order to compare with baseline methods [44, 49, 74], we use a shallow transformer decoder task head to perform answer classification task, and evaluate the performance with exact matches (EM@1), which is equivalent to Top-1 answer accuracy. We also provide EM@1 on a breakdown of question types, including \"What,\" \"Is,\" \"How,\" \"Can,\" \"Which,\" and \"Other,\" based on the first word in the question sentence. Additionally, we evaluate situational estimation performance with localization accuracy and orientation accuracy. In both tasks, we use accuracy within different distance or angle thresholds as our metrics. For example, \"Acc@0.5m\" means accuracy of location estimation when positive threshold is set to 0.5 meter. For ScanQA," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13682" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.204, + 0.089, + 0.773, + 0.295 + ], + "angle": 0, + "content": "
ModelQuestion BreakdownOverall
WhatIsHowCanWhichOther
GPT-3 [7]39.746.040.545.636.138.441.0
ClipBERT [35]30.260.138.763.342.542.743.3
MCAN [67]28.959.744.168.340.740.543.4
ScanQA [5]28.665.047.366.343.942.945.3
SQA3D [44]33.566.142.469.543.046.447.2
Multi-CLIP [15]------48.0
LM4Vision [49]34.367.148.268.348.945.648.1
3D-LLM [26]36.565.647.268.848.046.348.1
3D-VisTA [74]34.863.345.469.847.248.148.5
SIG3D (Ours)35.667.248.571.449.145.852.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.299, + 0.893, + 0.328 + ], + "angle": 0, + "content": "Table 1. Our proposed method SIG3D achieves state-of-the-art performance on SQA3D benchmark [44]. We perform the best on \"Is\", \"How\", and \"Can\" breakdown types of questions, as well as the average accuracy. The results are reported on test set." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.34, + 0.47, + 0.452 + ], + "angle": 0, + "content": "
ModelLocalizationOrientation
Acc@0.5mAcc@1.0mAcc@15°Acc@30°
Random7.225.88.416.9
SQA3D [44]9.529.68.716.5
SQA3D (separate)10.331.417.122.8
3D-VisTA [74]11.734.516.924.2
SIG3D (Ours)27.459.128.742.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.542 + ], + "angle": 0, + "content": "Table 2. Our proposed method SIG3D performs significantly better than prior method [44] in situational estimation task. \"Acc@0.5m\" stands for localization accuracy with \\(0.5\\mathrm{m}\\) threshold. \"Acc@15°\" represents orientation accuracy with \\(15^{\\circ}\\) threshold. separate means disabling other tasks to let the model focus on situational estimation only." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.549, + 0.47, + 0.663 + ], + "angle": 0, + "content": "
ModelBLEU-1BLEU-4ROUGEMETEORCIDEr
BLIP2 [38]29.75.926.611.345.7
Flamingo [3]25.68.431.111.355.0
VN+MCAN [67]28.06.229.811.454.7
SR+MCAN [67]26.97.930.011.555.4
ScanQA [5]30.210.133.313.164.9
3D-LLM [26]39.312.035.714.569.4
SIG3D39.512.435.913.468.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.754 + ], + "angle": 0, + "content": "Table 3. Performance of SIG3D on ScanQA dataset [5] is on-par with the state-of-the-art with large-scale text-3D pre-training. VN and SR stand for VoteNet and ScanRefer, respectively. 3D-LLM [74] leverages pretrained 2DVL foundation models and LLM models [3, 7, 37, 38], and is pretrained on a large-scale held-in 3D-text dataset before the finetuning on ScanQA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.469, + 0.822 + ], + "angle": 0, + "content": "we perform auto-regressive answer generation with large transformer decoder [26], and evaluate with BLEU [50], ROUGE [39], METEOR [6], and CIDEr [60] metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.34, + 0.849 + ], + "angle": 0, + "content": "5.1. Situated Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Baselines. Our study involves a comparative analysis with a range of representative baselines on the SQA3D dataset. In particular, we evaluate against GPT-3 [7], Clip" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.893, + 0.525 + ], + "angle": 0, + "content": "BERT [35], and MCAN [67], which are, as reported in prior work [44], baselines focused on language-only, 2D video, and 2D image QA, respectively. For GPT-3, we follow SQA3D [44] to convert the visual input into a caption using Scan2Cap [10] for LLMs to process. ScanQA [5] represents a 3D QA baseline that ignores the situational input. Both SQA3D [44] and Multi-CLIP [15] employ situational descriptions and annotations for direct regression tasks. LM4Vision [49] utilizes LLMs as visual and textual encoders. Additionally, 3D-VisTA [74] undergoes a pretraining procedure on their large-scale 3D scene-text dataset, ScanScribe, prior to the finetuning on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.879 + ], + "angle": 0, + "content": "Situational Estimation. As shown in Table 2, our work performs significantly better than the state of the art [44, 74] in both localization and orientation estimation tasks. For 3D-VisTA [74], we use a pretrained model and finetune a new situation head with the SQA3D dataset following [44]. We also report a random baseline, in which we randomly sample position and orientation from a uniform distribution as a lower-bound performance. Note that original SQA3D performs only marginally better than the random baseline, meaning that it does not acquire any situational awareness, despite having the situational estimation loss. Disabling the QA task and asking the model to exclusively focus on the situational estimation task results in a slight better performance. Our method, with the anchor-based position likelihood estimation, results in a much better understanding of the 3D situational relationship. Our method also outperforms 3D-VisTA, which is pretrained on a large-scale 3D-text dataset, indicating that large pretraining alone is not enough to address the situational awareness problem. Note that we do not include the random baseline performance reported in [44], because each value is obtained by generating three random values and taking the closest one to the ground true, and thus it does not reflect a true \"random\" baseline." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Situated Question Answering. SIG3D outperforms prior" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13683" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.101, + 0.308, + 0.112 + ], + "angle": 0, + "content": "(a) Number of Visual Tokens" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.113, + 0.361, + 0.182 + ], + "angle": 0, + "content": "
Acc@1.0mAcc@30°EM@1
12848.938.249.2
25659.142.550.9
51257.842.150.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.439, + 0.101, + 0.573, + 0.112 + ], + "angle": 0, + "content": "(b) Voxel size (in meters)" + }, + { + "type": "table", + "bbox": [ + 0.407, + 0.113, + 0.601, + 0.182 + ], + "angle": 0, + "content": "
Acc@1.0mEM@1
0.0154.149.5
0.0259.150.9
0.0547.348.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.69, + 0.101, + 0.832, + 0.113 + ], + "angle": 0, + "content": "(c) Rotation representation" + }, + { + "type": "table", + "bbox": [ + 0.642, + 0.113, + 0.875, + 0.182 + ], + "angle": 0, + "content": "
Acc@30°EM@1
Quaternion31.450.0
6D vector42.550.9
sinθ, cosθ42.650.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.192, + 0.894, + 0.222 + ], + "angle": 0, + "content": "Table 4. Ablation study validates that our various design choices improves the performance. \"Acc@1.0m\", \"Acc@30°\", and \"EM@1\" are accuracy (%) for localization estimation, orientation estimation, and QA tasks. Our settings are marked in gray." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.235, + 0.47, + 0.402 + ], + "angle": 0, + "content": "
Acc@1.0mAcc@30°EM@1
3D Vision Encoders
Text-only (no vision input)--47.5
VoteNet [53]37.428.249.1
3DETR [48]47.229.149.4
OpenScene - OpenSeg [51]57.541.650.2
OpenScene - LSeg [51]59.142.550.9
Language Tokenizer / Encoders
GloVe + LSTM [24, 52]44.330.948.7
SBERT - MiniLM [55]56.138.649.4
SBERT - MPNet [55]55.940.649.7
SBERT - MPNet (finetune)59.142.550.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.406, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Table 5. Performance of SIG3D improves with stronger visual and language encoders. We find that open-vocabulary point encoder and MPNet-based sentence BERT leads to best performance. \"Acc@1.0m\" and \"Acc@30°\" stands for localization and orientation accuracy in situational estimation task. \"EM@1\" demonstrate exact match metric in QA task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.639 + ], + "angle": 0, + "content": "methods in most question breakdown categories and overall accuracy, as shown in Table 1. Our work achieves leading results without large-scale pretraining (compared with 3D-VisTA) and LLM (compared with GPT-3), indicating its superiority in situational awareness. Note that the LLM baseline GPT-3 achieves the best performance on the \"What\" category, suggesting the potential of stronger language encoder in interpreting the complicated question prompt." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.656, + 0.427, + 0.673 + ], + "angle": 0, + "content": "5.2. General Question Answering on ScanQA" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Baselines. We compare with 2D image VQA MCAN-based baselines [67], ScanQA [5], 3D-LLM [26] which leverages large-scale pretrained 2D VLMs and LLMs as backbone models, and 3D-VisTA [74] pretrained on their proposed large-scale 3D-text dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Question Answering. As shown in Table 3, despite that the questions do not explicitly require situational understanding to answer in ScanQA, SIG3D achieves comparable results with state-of-the-art methods without the large-scale 3D-text pretraining and powerful 2D VLM and LLM backbone models. Our work pretrained on SQA3D [44] leads to higher performance on BLEU-1, BLEU-4, and ROUGE metrics, showing its generalizability on general 3D QA scenarios." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.235, + 0.892, + 0.396 + ], + "angle": 0, + "content": "
Acc@1.0mAcc@30°EM@1
Baseline (joint optimization)29.523.147.7
How to achieve better situational estimation
+ 3D PE38.823.647.8
+ 6D Representation38.527.447.7
+ Anchor-based Estimation58.841.948.2
How to utilize situational estimation for better QA
+ 3D Situational PE58.941.850.0
+ Visual Token Re-encoding59.142.550.9
Oracle Models (Ground Truth Situation Information)
Situation as direct input10010047.7
Situation as intermediate input10010053.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.4, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Table 6. Ablation study verifies that our proposed modules leads to better situational estimation and better QA performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.449, + 0.742, + 0.466 + ], + "angle": 0, + "content": "5.3. Ablation Study & Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Vision & Language Encoders. We study the impacts of different visual and textual tokenizers in Table 5. It is observed that the open-vocabulary visual encoder (OpenScene) outperforms detection-based encoders (such as VoteNet and 3DETR) across all metrics. This superior performance of OpenScene is attributed to the limitations of 3D detectors, which are typically trained on a limited set of object categories, rendering them less effective in recognizing novel objects mentioned in textual prompts. Regarding language encoders, our findings indicate that a stronger backbone correlates with better performance, primarily due to its improved capability to interpret complex textual inputs. This leads to the suggestion of integrating LLM with our method to potentially further enhance performance, an avenue we intend to explore in future research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Situated Awareness. In Table 6 we verify the crucial role of situational awareness in 3DVL task. Firstly, we show that 3D PE, 6D rotation estimation, and anchor-based position estimation all leads to much better position and orientation estimation performance. We further establish that situational PE and visual token re-encoding modules lead to better utilization of the predicted situational vector for QA task. Additionally, We design two oracle models under the assumption of having access to the ground truth situational vector as an input. The outcomes from these models reveal a critical insight: the model fails to effectively interpret situational information when it is directly incorporated into the input visual embeddings. This underlines the necessity" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.892, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.893, + 0.465 + ], + "angle": 0, + "content": "Figure 5. Qualitative results demonstrate significant improvement of SIG3D over prior method. The first row is results from SQA3D [44], and the second row is results from our methods. In the 3D scene, red: Ground truth (GT) vector, and blue: Estimated vector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.526 + ], + "angle": 0, + "content": "of the intermediate representation and encoding mechanism we have proposed, affirming its importance in achieving optimal 3DVL task performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.471, + 0.729 + ], + "angle": 0, + "content": "Architectural Design. We explore different architectural design choices of our model in Table 4. We find that the number of visual tokens sampled from the visual feature embeddings affects the performance of both situational estimation and QA tasks. Sampling fewer visual tokens increases the risk of missing the region of significance, while sampling more does not lead to a better performance as well. We study the size of voxels and find \\(0.02\\mathrm{m}\\) to be the most effective choice, as OpenScene [51] backbone is pretrained with the same voxel size. We also find \\(\\sin \\theta\\), \\(\\cos \\theta\\) and 6D vector representation performs a lot better than quaternion in rotation estimation task. This is consistent with the finding reported in [72]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.27, + 0.758 + ], + "angle": 0, + "content": "5.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Finally, we demonstrate some qualitative results of our SIG3D in Figure 5. We show the ground truth and estimated situational vectors in red and blue colors, respectively, in their corresponding 3D scenes. We also print the answers with red cross or green checkmark indicating the correctness. It is clear that our method performs significantly better in situational estimation task, resulting in vectors very close to the ground truth in both position and orientation perspectives. The better situational awareness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.542 + ], + "angle": 0, + "content": "also aids the complicated embodied navigation and common sense QA activities. This also demonstrate great potential of our method in the development of indoor robotics and/or conversational agents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.893, + 0.651 + ], + "angle": 0, + "content": "Supplementary Material. The supplementary section offers an extensive analysis, encompassing a detailed examination of the 3D visual token activation changes pre and post situational re-encoding. Additionally, it includes a comprehensive collection of positive and negative samples, an insightful failure case analysis, and a forward-looking discussion on limitations and future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.619, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In this paper, we introduce SIG3D, a situation-aware vision language model for 3D reasoning tasks. We propose to represent 3D scene as feature tokens, treat tokens as anchors points to estimate a situational vector from a text description, and use the estimated situation as guidance to align and re-encode the visual tokens to enhance the features for reasoning tasks. We observe consistent and significant performance gain on both situational estimation and question answering tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13685" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. ShapeGlot: Learning language for shape differentiation. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurlPS, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[5] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In CVPR, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[6] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.512 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3D object localization in rgb-d scans using natural language. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In ACCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.615, + 0.47, + 0.656 + ], + "angle": 0, + "content": "[10] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.659, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.73, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In CVPR, 2019. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.773, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.83, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[14] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[15] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann," + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "and Thomas Hofmann. Multi-CLIP: Contrastive Vision-Language Pre-training for Question Answering tasks in 3D Scenes. In BMVC, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[16] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[18] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[19] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3D visual graph network for object grounding in point cloud. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] François Gardères, Maryam Ziaeefard, Baptiste Abeloos, and Freddy Lecue. Conceptbert: Concept-aware representation for visual question answering. In EMNLP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[22] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, and Jianfeng Gao. Kat: A knowledge augmented transformer for vision-and-language. In NAACL, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[23] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From Images to Textual Prompts: Zero-shot Visual Question Answering with Frozen Large Language Models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[25] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3D Concept Learning and Reasoning from Multi-View Images. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[26] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3D World into Large Language Models. NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[27] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3D instance segmentation. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[28] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. LEMMA: A Multi-view Dataset for LE arning Multi-agent M multi-task A civities. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13686" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[31] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[32] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In iccv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[33] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[35] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven Semantic Segmentation. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ICML, 2023. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[40] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang. Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In CVPR, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[41] Yuanze Lin, Yujia Xie, Dongdong Chen, Yichong Xu, Chenguang Zhu, and Lu Yuan. Revive: Regional visual representation matters in knowledge-based visual question answering. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[42] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[44] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: Situated question answering in 3D scenes. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[45] Yunze Man, Liang-Yan Gui, and Yu-Xiong Wang. BEV-Guided Multi-Modality Fusion for Driving Perception. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[46] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[47] Jack Merullo, Louis Castricato, Carsten Eickhoff, and Ellie Pavlick. Linearly mapping from image to text space. *ICLR*, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[48] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[49] Ziqi Pang, Ziyang Xie, Yunze Man, and Yu-Xiong Wang. Frozen Transformers in Language Models Are Effective Visual Encoder Layers. arXiv preprint arXiv:2310.12973, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[51] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3, 4, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[52] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[53] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[55] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In EMNLP, 2019. 4, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[57] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[60] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[62] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In NeurIPS (Datasets and Benchmarks Track), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13687" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[63] Jialin Wu, Jiasen Lu, Ashish Sabharwal, and Roozbeh Mottaghi. Multi-modal answer validation for knowledge-based vqa. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[64] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. LLM-Grounder: Open-Vocabulary 3D Visual Grounding with Large Language Model as an Agent. arXiv preprint arXiv:2309.12311, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3D question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[66] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In CVPR, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[67] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, 2019. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.376 + ], + "angle": 0, + "content": "[68] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[69] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[70] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. MotionGPT: Finetuned LLMs are General-Purpose Motion Generators. arXiv preprint arXiv:2306.10900, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[71] Brady Zhou and Philipp Krahenbuhl. Cross-view Transformers for real-time Map-view Semantic Segmentation. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[72] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[73] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[74] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment. In ICCV, 2023. 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13688" + } + ] +] \ No newline at end of file diff --git a/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_origin.pdf b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c9ffb383995d2a8c2633be061c5fe706262ac10 --- /dev/null +++ b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/1ff0a4cb-e0c0-41a2-9620-42d78c2488cd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b04bef49ec5ea21cd4049c657e4f5f4b1f07b0f12282d640b74306a62d59d123 +size 1107669 diff --git a/2024/Situational Awareness Matters in 3D Vision Language Reasoning/full.md b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3fcdd3fa05b7baab5f89806abb37f115b1241ed6 --- /dev/null +++ b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/full.md @@ -0,0 +1,283 @@ +# Situational Awareness Matters in 3D Vision Language Reasoning + +Yunze Man Liang-Yan Gui Yu-Xiong Wang University of Illinois Urbana-Champaign + +{yunzem2,lgui,yxw}@illinois.edu + +# Abstract + +Being able to carry out complicated vision language reasoning tasks in 3D space represents a significant milestone in developing household robots and human-centered embodied AI. In this work, we demonstrate that a critical and distinct challenge in 3D vision language reasoning is the situational awareness, which incorporates two key components: (1) The autonomous agent grounds its self-location based on a language prompt. (2) The agent answers open-ended questions from the perspective of its calculated position. To address this challenge, we introduce SIG3D, an end-to-end Situation-Grounded model for 3D vision language reasoning. We tokenize the 3D scene into sparse voxel representation, and propose a language-grounded situation estimator, followed by a situated question answering module. Experiments on the SQA3D and ScanQA datasets show that SIG3D outperforms state-of-the-art models in situational estimation and question answering by a large margin (e.g., an enhancement of over $30\%$ on situation accuracy). Subsequent analysis corroborates our architectural design choices, explores the distinct functions of visual and textual tokens, and highlights the importance of situational awareness in the domain of 3D question-answering. Project page is available at https://yunzeman.github.io/situation3d. + +# 1. Introduction + +Humans learn knowledge efficiently through the interactions with the 3D world and the integration of multi-modal information, such as verbal guidance or instructions. Similarly, introducing language guidance into the visual comprehension task can greatly enhance models' learning efficiency [3, 42]. Nonetheless, despite considerable advancements in linguistic understanding [7, 11, 31, 58] and vision-language integration [3, 37, 54, 61], current methodologies remain deficient in accurately perceiving and rationalizing within real-world 3D environments, which is largely attributed to the lack of 3D situational reasoning capabilities. + +Compared with machine learning models, humans put + +![](images/3743d41e5d3a84bf69adb681218bac04137bc2965fe7e2159dbb3884a53b9f95.jpg) +Figure 1. Previous methods perform direct 3D vision language reasoning without modeling the situation of an embodied agent in the 3D environment. Our method, SIG3D, grounds the situational description in the 3D space, and then re-encodes the visual tokens from the agent's intended perspective before vision-language fusion, resulting in a more comprehensive and generalized 3D vision language (3DVL) representation and reasoning pipeline. Q, K, V stands for query, key, and value, respectively. + +themselves inside the 3D world and then perceive and interact with the surrounding environment from their egoperspective (Figure 1). Such situational awareness is a crucial difference between 2D and 3D visual understanding, and a key to achieve seamless understanding of spatial concepts in more complex real-world environments. Several existing methods recognize the lack of positional understanding in 3D and propose new benchmarks and joint optimization functions [44], or positional embedding methods [26] to enhance the overall reasoning performance. + +However, the lack of an explicit situation modeling and situation-grounded 3D reasoning method restricts them from obtaining a generalizable and consistent 3D vision-language (VL) representation. As shown in Figure 2, the situation prediction of the state-of-the-art method [44] (in blue) diverges significantly from the ground truth vectors (in red) in almost all scenes in the dataset [13]. Moreover, our pilot study in Section 3 also reveals that situational understanding, despite being very crucial in comprehending the context of questions, only plays a minor role in the final + +![](images/f2652a4744e4f1b9b629ac3b59c6a5e5c71bcbe328d03926159a7f561e502346.jpg) +I am facing a trash can, while there is a door on my right. + +![](images/976466d9324826ee50a29bec14a5c7f88056478dd6b771190e13afc97ae6a01c.jpg) +I am facing the paper towel dispenser with a chair in my six o'clock direction. + +![](images/2165f6d32d3e15bd7ee574ddfc128ad3bb03b33c2d22d0851737bc3dddd20906.jpg) +I am facing a pool table, taking out the white ball from the side pocket. I am also facing a painting on the wall. + +![](images/c2658c3e9812b52434ac8e0ee2d0b95a7787c20d17889903217dccc87ce8cc0b.jpg) +I am sitting on a couch with a pillow facing another couch and the pillow is on my right. +Figure 2. Situational estimation in existing methods [44] fails in most scenarios, indicating the missing registration between the situational description and 3D embeddings. Red: Ground truth (GT) vector. Blue: Estimated vector. + +question-answering performance of existing methods. + +In this work, we propose SIG3D, a novel approach designed to precisely model and estimate an embodied agent's ego-location and orientation from a textual description, before performing multi-modal QA tasks from the agent's egocentric perspective, as shown in Figure 1. Specifically, we leverage large-scale pretrained language and visual encoders to process the input text and 3D data, and fuse the tokens with attention modules to predict a situational vector. Previous attempts to directly predict the ego-situation are hindered by the expansive search space inherent in 3D environments. To address this challenge, we re-conceptualize the task as an anchor-based classification, where visual tokens are regarded as anchor points, and a likelihood of position together with a set of rotation parameters are concurrently regressed for each visual token. After obtaining the situational estimation, we propose a situational alignment and a situation-guided token re-encoding strategy, to perceive the environment from the agent's intended perspective. These strategies enhance the visual token with more accurate situational awareness for subsequent QA tasks. + +Experiments on two challenging 3D visual question-answering (VQA) datasets [5, 44] demonstrate the significant improvement on situational estimation and QA tasks of our model. In particular, we improve the situational estimation accuracy by over $30\%$ , and the subsequent QA performance by up to $3\%$ . Further qualitative and quantitative analysis verifies our design choices and highlights the significance of situational awareness in 3D reasoning tasks. + +To sum up, our paper has the following contributions: (1) We recognize the lack of situational awareness as a sig- + +nificant oversight in existing research. To address this, we introduce SIG3D, a situation-grounded 3D VL reasoning architecture, specifically designed to fill this void. (2) We propose an anchor-based approach to situational estimation, which effectively narrows the extensive search space in 3D environments for precise grounding of 3D positions and orientations with textual descriptions. Additionally, we investigate situational alignment and visual re-encoding mechanisms to leverage situational awareness for enhanced QA performance. (3) Our model demonstrates superior performance on two challenging datasets, SQA3D and ScanQA, surpassing the state of the art in both situational estimation and QA metrics. Ablation studies highlight the importance of situation-guided encoding, revealing its beneficial impact on general QA tasks. + +# 2. Related Work + +Vision Language Models (VLMs). Early transformer-driven [59] textual and visual encoders [17, 31] have facilitated great progress in recent vision language learning. Text-image contrastive models [30, 54] propose to align the feature space of two modalities with large-scale pretraining, fueling numerous downstream tasks from generalized open-vocabulary visual perception [21, 34, 36] to text-to-image generation [56]. Concurrently, some work uses text and vision encoders on separate modalities followed by feature fusion [18, 33] for multi-modal reasoning tasks. Since the emergence of Large Language Models (LLMs) [7, 58, 69], VLMs have experienced huge improvement with the help of LLMs as building blocks for multi-modal learning architectures. Specifically, recent work directly projects visual embeddings into language-space tokens as input to LLM [23, 40, 47, 70], or use the latent bottleneck structure for cross-modal visual decoding [3, 26, 37, 38], or treat LLM layers as encoder blocks for various visual tasks [49]. + +In the domain of visual question-answering (VQA) [4, 73], recent work has pushed the frontier towards video understanding [14, 28, 29, 40, 62], knowledge-based understanding [20, 22, 23, 41, 46, 57, 63], and commonsense reasoning [68]. Despite the outstanding performance in 2D image interpretation, most existing methods lack the capability to generalize to 3D scenarios. In contrast, our work studies the representation of visual information and its fusion with language embeddings in the 3D domain by targeting on the 3D situation-guided visual language interpretation. + +Grounding Language in 3D Space. Compared with 2D images, knowledge such as spatial relationships, interactive exploration, and topological analysis, which only exist in the 3D world provides additional challenges and opportunities to develop better language models with stronger commonsense reasoning capability grounded in the real-world 3D scenarios. In this direction, early work seeks + +![](images/dd7ff2f87d095cdce14564a7bb9feee5ce4a817302e0629fecdf932f2ab0596a.jpg) +Figure 3. Results on variants of the representative SQA3D baseline method [44] demonstrate that situational understanding, despite being indispensable in perceiving the context of questions, makes negligible contribution in existing methods. This motivates a situation-guided 3D encoding mechanism in our model. + +to ground isolated objects [1, 9] or objects within more complex scenes [2, 8, 19, 27] using natural language descriptions. Recently, with more collected 3D vision language datasets, several work starts to explore language-guided 3D visual interpretation and reasoning on a diverse set of datasets, including 3D scene captioning [10], open-vocabulary segmentation [16, 32, 51], and question-answering [5, 15, 25, 65, 74]. + +The success of LLMs also elicits the usage of them in 3D vision language reasoning for task decomposition [64], data generation, and multi-modal feature fusion [26]. Motivated by ScanQA [5], SQA3D [44] takes the first step into exploring the challenging 3D situational reasoning problem by developing a situated question-answering benchmark, and proposing the first joint learning baseline on the benchmark. Our work highlights the uniqueness and significance of situational awareness in the 3D vision language learning paradigm, leading to notably better 3D situational grounding and question-answering performance. + +# 3. Pilot Study on Situated Reasoning + +Despite pointing out the significance of situated understanding and reasoning, existing methods [44] fall short in providing effective situational estimation, as illustrated in Figure 2. This section delves into a pilot study examining the impact of situational understanding on downstream reasoning tasks. The SQA3D baseline [44] incorporates situational descriptions and uses ground truth (GT) situational vectors for supervision in a direct regression task. We investigate three variants of this baseline to assess the effect of situational understanding. In the first variant, we remove the situational description and supervision from the model, by passing in empty situational tokens. In another variant, we corrupt the situation supervision by introducing very large Gaussian noise to the GT vectors to effectively randomizing them. Finally, we try to encode the GT situational vector in the input with learnable multi-layer perceptron layers, which are then integrated with the visual and textual tokens. + +Figure 3 demonstrates the results of this study, revealing negligible changes in performance across these variants. Notably, corrupting the GT situational information or di + +rectly incorporating it results in only marginal alterations in the QA outcomes. Omitting the situational description entirely from the input results in a mere $2\%$ decline in accuracy. However, in the absence of this information, the model resorts to random guessing when determining the correct answer, as all responses depend on situational context. The findings from Figures 2 and 3 collectively indicate a deficiency in existing methods regarding situation vector estimation and the application of situational understanding in subsequent reasoning tasks. These unresolved challenges motivate the development of our proposed method. + +# 4. Method + +An overview of our approach SIG3D is illustrated in Figure 4. Our method begins with a set of points that represent a 3D scene, accompanied by a situational description and a question, which defines the overall context of the problem. We tokenize them into separate token embeddings (Section 4.1), and ground the textual description in the 3D scene with a vector comprising of location and orientation. We find direct single vector estimation to be challenging due to the vast and complex nature of the 3D search space, so we propose an anchor-based situational estimation strategy (Section 4.2). Subsequently, we re-encode the visual tokens from the perspective of situational vectors, enhancing the situational awareness for downstream reasoning tasks (Section 4.3). The finalized visual and textual tokens are fused by a transformer decoder to generate the final response. + +# 4.1. Visual and Textual Tokenization + +Leveraging input scene point clouds and textual prompts, our objective is to generate three distinct types of tokens: 3D visual tokens $z^{3\mathrm{D}} \in \mathbb{R}^{N_v \times C_v}$ , situational tokens $z^{\mathrm{S}} \in \mathbb{R}^{N_s \times C_s}$ , and question tokens $z^{\mathrm{Q}} \in \mathbb{R}^{N_q \times C_q}$ . Each type of token is characterized by two primary components: $N$ , representing the number of tokens, and $C$ , encapsulating the feature embeddings. Tockenize and capture the feature embeddings for situational and question inputs, we employ a shared text tokenizer $\mathrm{E}^{\mathrm{TXT}}$ following prior methods [5, 44]. We assume that situational and question prompts are separated in the input data. If not, LLMs [7] can be used to parse the textual input without changing the semantic meaning of the sentences. However, there is a lack of consensus on a standard 3D visual tokenization method $\mathrm{E}^{3\mathrm{D}}$ that is apt for the 3DVL reasoning task, prompting a more detailed exploration in the subsequent paragraphs. + +Visual Tokenization. Given an input point cloud $\mathbf{p} \in \mathbb{R}^{N \times 3}$ , most prior methods [5, 15, 44] adopt a VoteNet [53] detector to acquire object-level tokens $z^{3\mathrm{D}} \in \mathbb{R}^{N_{\mathrm{obj}} \times C_{\mathrm{obj}}}$ as the visual representation, where $N_{\mathrm{obj}}$ is the number of object proposals, and $C_{\mathrm{obj}}$ is the object-level feature embeddings. However, we point out several problems with + +![](images/7b27bdd4f2816e51261ad8540905c741dd7b3c0866dab967ea8a81792e51acf8.jpg) +Figure 4. Overview of our SIG3D model, which includes 3D scene and text encoding, anchor-based situational estimation, situation-guided visual re-encoding, and multi-modal decoder modules. We tokenize the 3D scene into voxels, treat each token as an anchor point, and query the text tokens to predict a token-level position likelihood and rotation matrix to locate the situational vector associated with the text descriptions. Then we update the scene tokens with situational position encoding (PE), and finally perform the 3DVL reasoning task with a large transformer decoder. + +this abstraction strategy: (1) A detection-based tokenization method tends to ignore the non-object regions in the scene, which can be indispensable in some reasoning scenarios (e.g., carpets on the ground, ceiling, walls.) (2) After object-level abstraction, the visual representation loses the high-level information of the scene (e.g., the shape of the living room, the corner of the kitchen.) (3) A supervised detector trained from scratch can only recognize objects within the training set (e.g., only 20 categories for ScanNet [13]), meaning that the method does not have zero-shot capability to reason about novel unseen objects that are inevitably common in real-world scenarios. + +In light of this, we adopt a pretrained open-vocabulary voxel-based tokenization method from OpenScene [51]. The scene is first discretized into regular small 3D voxels and fed into a visual encoder for feature extraction: + +$$ +x ^ {\mathrm {3 D}} = \operatorname {E} ^ {\mathrm {3 D}} (\mathcal {V} (\mathbf {p})), \tag {1} +$$ + +where $\mathcal{V}$ represents the voxelization process, and $\mathrm{E}^{3\mathrm{D}}$ is a Minkowski sparse 3D convolutional network [12]. The sparse network is pretrained by distillation from CLIP [54] embeddings of rendered multi-view 2D images, resulting in a feature map with better language alignment and 3D awareness. We take the upsampled bottleneck-layer feature embeddings from the encoder network, and compute the mean average over the $z$ -axis (vertical) to project the voxels onto the $x - y$ plane and treat the feature grids in the resulting 2D feature map as our $N_{v}$ visual tokens. We find that this bird's-eye-view projection results in a more compact representation and improves the final performance. + +# 4.2. Situational Estimation + +Given 3D visual tokens $z^{3\mathrm{D}}$ and situational tokens $z^{\mathrm{S}}$ , our objective is to estimate the situational vector $\vec{s}$ referred by the situational description, which comprises of a position component $s^{\mathrm{pos}}$ represented by $(x,y,z)$ coordinates, and a rotation component $s^{\mathrm{rot}}$ represented by $(\theta ,\psi ,\phi)$ Euler angles, where pitch angles $\psi$ are always defined as 0, meaning that situational vectors are defined to be parallel with the ground plane. The prior method [44] utilizes a transformer block to calculate the cross-attention feature between visual and language tokens, and directly regress a final situational vector out of the averaged attention map. We find such strategy producing very inaccurate estimation, as shown in Figure 2, due to the large search space in the entire 3D volume. Inspired by the recent 3D object detection methods [45, 66, 71], we reduce the search space by turning the localization problem into a classification problem. + +Positional Embedding and Feature Fusion. After the voxelization and 3D encoding process, each 3D token associates with a 3D position $(x,y,z)$ representing the center of its voxel. We first provide positional information to the model by generating learnable positional embeddings (PE) using a two-layer perceptron for each of the $N_v$ visual tokens, and add learnable positional embeddings to the token features $z^{3\mathrm{D}}$ . We use a situation interpreter [55] to extract situational information, and ask the updated visual tokens to attend to these situational tokens with several transformer layers to produce the joint feature embeddings. + +Anchor-based Situational Estimation. We treat each + +output token of the feature fusion module as an anchor point, and use it to predict a position likelihood $p \in [0,1]$ and a rotation estimation. Since each token has an associated 3D position $(x,y,z)$ , the position likelihood $p$ indicates how likely the situational vector locates at the center of this token (voxel). We define a soft ground truth for this classification task with a Gaussian kernel, meaning that the closer a token is to the actual situational vector $s^{\mathrm{pos}}$ , a higher ground truth probability $p$ will be assigned to that token. In order to counteract the sparse supervisory signal and increase the positive supervision around the vector position, we adopt the peak enlarging technique in CenterPoint [66], where the size of the Gaussian kernel is increased (meaning that the $\sigma$ is increased) to allow denser supervision around the vector position. Furthermore, we explore different rotation representation and find that compared with quaternion and $(\sin \theta, \cos \theta)$ representations, 6D vector proposed by [72] achieves the best performance. Hence, we adopt a situational estimation head with MLP layers to output 7-dimensional vector for each of the tokens, where the first channel represents the position likelihood and the other six channels represent the 6D rotation matrix. We take the center of the token with the peak position likelihood as our estimated $s^{\mathrm{pos}}$ , and convert its corresponding 6D rotation vector as our estimated $s^{\mathrm{rot}}$ . The estimation can be equivalently represented as a rotation matrix $R$ and a translation matrix $T$ . More discussion about the architecture and design choices is in Section 5.3. + +# 4.3. Situation-guided Visual Encoding + +After obtaining the situational estimation, we investigate a better approach to enhancing the downstream response generation, inspired by human cognitive processes. Intuitively, humans typically comprehend their immediate 3D environment by first interpreting their own situation in space, and then discerning their surroundings from an appropriate viewpoint. Our model is designed to emulate this natural strategy. Utilizing the situational vector $\vec{s}$ , we adjust the coordinate system by repositioning the origin at $s^{\mathrm{pos}}$ , and reorienting the axes according to $s^{\mathrm{rot}}$ such that the new y-axis is aligned with the indicated direction. We keep the $z$ -axis vertically oriented, and project the situational vectors onto the $x - y$ plane. This is in line with the format of the dataset [44], where situational vectors are assumed to be parallel with the ground plane. Subsequently, we compute a new situation-guided PE for each of the $N_{v}$ visual tokens, akin to the learnable 3D PE outlined in Section 4.2. They allow the model to grasp positional interrelations from the perspective of the current situation. These situational embeddings are added to the output embeddings of the situational estimation module, which consists of blocks featuring self-attention layers for visual tokens, succeeded by cross-attention layers that bridge visual + +and situational information. This structure allows for the re-encoding of visual tokens under the influence of situational and question context, guiding the model to assign higher weights to situation-related and question-related visual tokens. The output, termed situation-guided visual tokens, embodies this re-contextualized understanding. + +# 4.4. Question Answering Head + +We follow existing methods [26] to use a large vision-language decoder to fuse the final visual and textual tokens and generate textual response to the input question. We explore both auto-regressive response generation, and classification-based answer prediction [5, 44]. For classification, we predict a vector $v^{\mathrm{ans}} \in \mathbb{R}^{n_a}$ for the $n_a$ answer candidates in the training set following [5]. + +# 5. Analysis in 3D VQA Task + +We evaluate SIG3D for 3DVL reasoning on two challenging benchmarks, addressing both visually-oriented situational estimation and textually-focused QA tasks. We present a detailed examination of the implementation strategies adopted, the datasets employed, and the metrics applied in our research. For an exhaustive understanding, implementation and training details, and other additional information are available in the supplementary materials. + +Datasets. We evaluate our method on SQA3D [44] and ScanQA [5], two challenging indoor 3D VQA datasets. Both datasets are derived from the ScanNet dataset [13], serving as the foundational source for their 3D scenes. SQA3D features over 33K question-answer pairs for the 3D VQA task and 26K unique situational descriptions for the situational estimation task. Each entry in this dataset includes a 3D scene point cloud, a situational description, a question, and pertinent annotations. ScanQA consists of over 41K question-answer pairs, without situational descriptions and situational annotations. We use it to demonstrate the generalizability of our method on general QA tasks. We use the splits provided by these datasets. + +Evaluation Metrics. For SQA3D, in order to compare with baseline methods [44, 49, 74], we use a shallow transformer decoder task head to perform answer classification task, and evaluate the performance with exact matches (EM@1), which is equivalent to Top-1 answer accuracy. We also provide EM@1 on a breakdown of question types, including "What," "Is," "How," "Can," "Which," and "Other," based on the first word in the question sentence. Additionally, we evaluate situational estimation performance with localization accuracy and orientation accuracy. In both tasks, we use accuracy within different distance or angle thresholds as our metrics. For example, "Acc@0.5m" means accuracy of location estimation when positive threshold is set to 0.5 meter. For ScanQA, + +
ModelQuestion BreakdownOverall
WhatIsHowCanWhichOther
GPT-3 [7]39.746.040.545.636.138.441.0
ClipBERT [35]30.260.138.763.342.542.743.3
MCAN [67]28.959.744.168.340.740.543.4
ScanQA [5]28.665.047.366.343.942.945.3
SQA3D [44]33.566.142.469.543.046.447.2
Multi-CLIP [15]------48.0
LM4Vision [49]34.367.148.268.348.945.648.1
3D-LLM [26]36.565.647.268.848.046.348.1
3D-VisTA [74]34.863.345.469.847.248.148.5
SIG3D (Ours)35.667.248.571.449.145.852.6
+ +Table 1. Our proposed method SIG3D achieves state-of-the-art performance on SQA3D benchmark [44]. We perform the best on "Is", "How", and "Can" breakdown types of questions, as well as the average accuracy. The results are reported on test set. + +
ModelLocalizationOrientation
Acc@0.5mAcc@1.0mAcc@15°Acc@30°
Random7.225.88.416.9
SQA3D [44]9.529.68.716.5
SQA3D (separate)10.331.417.122.8
3D-VisTA [74]11.734.516.924.2
SIG3D (Ours)27.459.128.742.5
+ +Table 2. Our proposed method SIG3D performs significantly better than prior method [44] in situational estimation task. "Acc@0.5m" stands for localization accuracy with $0.5\mathrm{m}$ threshold. "Acc@15°" represents orientation accuracy with $15^{\circ}$ threshold. separate means disabling other tasks to let the model focus on situational estimation only. + +
ModelBLEU-1BLEU-4ROUGEMETEORCIDEr
BLIP2 [38]29.75.926.611.345.7
Flamingo [3]25.68.431.111.355.0
VN+MCAN [67]28.06.229.811.454.7
SR+MCAN [67]26.97.930.011.555.4
ScanQA [5]30.210.133.313.164.9
3D-LLM [26]39.312.035.714.569.4
SIG3D39.512.435.913.468.8
+ +Table 3. Performance of SIG3D on ScanQA dataset [5] is on-par with the state-of-the-art with large-scale text-3D pre-training. VN and SR stand for VoteNet and ScanRefer, respectively. 3D-LLM [74] leverages pretrained 2DVL foundation models and LLM models [3, 7, 37, 38], and is pretrained on a large-scale held-in 3D-text dataset before the finetuning on ScanQA. + +we perform auto-regressive answer generation with large transformer decoder [26], and evaluate with BLEU [50], ROUGE [39], METEOR [6], and CIDEr [60] metrics. + +# 5.1. Situated Question Answering + +Baselines. Our study involves a comparative analysis with a range of representative baselines on the SQA3D dataset. In particular, we evaluate against GPT-3 [7], Clip + +BERT [35], and MCAN [67], which are, as reported in prior work [44], baselines focused on language-only, 2D video, and 2D image QA, respectively. For GPT-3, we follow SQA3D [44] to convert the visual input into a caption using Scan2Cap [10] for LLMs to process. ScanQA [5] represents a 3D QA baseline that ignores the situational input. Both SQA3D [44] and Multi-CLIP [15] employ situational descriptions and annotations for direct regression tasks. LM4Vision [49] utilizes LLMs as visual and textual encoders. Additionally, 3D-VisTA [74] undergoes a pretraining procedure on their large-scale 3D scene-text dataset, ScanScribe, prior to the finetuning on this dataset. + +Situational Estimation. As shown in Table 2, our work performs significantly better than the state of the art [44, 74] in both localization and orientation estimation tasks. For 3D-VisTA [74], we use a pretrained model and finetune a new situation head with the SQA3D dataset following [44]. We also report a random baseline, in which we randomly sample position and orientation from a uniform distribution as a lower-bound performance. Note that original SQA3D performs only marginally better than the random baseline, meaning that it does not acquire any situational awareness, despite having the situational estimation loss. Disabling the QA task and asking the model to exclusively focus on the situational estimation task results in a slight better performance. Our method, with the anchor-based position likelihood estimation, results in a much better understanding of the 3D situational relationship. Our method also outperforms 3D-VisTA, which is pretrained on a large-scale 3D-text dataset, indicating that large pretraining alone is not enough to address the situational awareness problem. Note that we do not include the random baseline performance reported in [44], because each value is obtained by generating three random values and taking the closest one to the ground true, and thus it does not reflect a true "random" baseline. + +Situated Question Answering. SIG3D outperforms prior + +(a) Number of Visual Tokens + +
Acc@1.0mAcc@30°EM@1
12848.938.249.2
25659.142.550.9
51257.842.150.7
+ +(b) Voxel size (in meters) + +
Acc@1.0mEM@1
0.0154.149.5
0.0259.150.9
0.0547.348.8
+ +(c) Rotation representation + +
Acc@30°EM@1
Quaternion31.450.0
6D vector42.550.9
sinθ, cosθ42.650.6
+ +Table 4. Ablation study validates that our various design choices improves the performance. "Acc@1.0m", "Acc@30°", and "EM@1" are accuracy (%) for localization estimation, orientation estimation, and QA tasks. Our settings are marked in gray. + +
Acc@1.0mAcc@30°EM@1
3D Vision Encoders
Text-only (no vision input)--47.5
VoteNet [53]37.428.249.1
3DETR [48]47.229.149.4
OpenScene - OpenSeg [51]57.541.650.2
OpenScene - LSeg [51]59.142.550.9
Language Tokenizer / Encoders
GloVe + LSTM [24, 52]44.330.948.7
SBERT - MiniLM [55]56.138.649.4
SBERT - MPNet [55]55.940.649.7
SBERT - MPNet (finetune)59.142.550.9
+ +Table 5. Performance of SIG3D improves with stronger visual and language encoders. We find that open-vocabulary point encoder and MPNet-based sentence BERT leads to best performance. "Acc@1.0m" and "Acc@30°" stands for localization and orientation accuracy in situational estimation task. "EM@1" demonstrate exact match metric in QA task. + +methods in most question breakdown categories and overall accuracy, as shown in Table 1. Our work achieves leading results without large-scale pretraining (compared with 3D-VisTA) and LLM (compared with GPT-3), indicating its superiority in situational awareness. Note that the LLM baseline GPT-3 achieves the best performance on the "What" category, suggesting the potential of stronger language encoder in interpreting the complicated question prompt. + +# 5.2. General Question Answering on ScanQA + +Baselines. We compare with 2D image VQA MCAN-based baselines [67], ScanQA [5], 3D-LLM [26] which leverages large-scale pretrained 2D VLMs and LLMs as backbone models, and 3D-VisTA [74] pretrained on their proposed large-scale 3D-text dataset. + +Question Answering. As shown in Table 3, despite that the questions do not explicitly require situational understanding to answer in ScanQA, SIG3D achieves comparable results with state-of-the-art methods without the large-scale 3D-text pretraining and powerful 2D VLM and LLM backbone models. Our work pretrained on SQA3D [44] leads to higher performance on BLEU-1, BLEU-4, and ROUGE metrics, showing its generalizability on general 3D QA scenarios. + +
Acc@1.0mAcc@30°EM@1
Baseline (joint optimization)29.523.147.7
How to achieve better situational estimation
+ 3D PE38.823.647.8
+ 6D Representation38.527.447.7
+ Anchor-based Estimation58.841.948.2
How to utilize situational estimation for better QA
+ 3D Situational PE58.941.850.0
+ Visual Token Re-encoding59.142.550.9
Oracle Models (Ground Truth Situation Information)
Situation as direct input10010047.7
Situation as intermediate input10010053.9
+ +Table 6. Ablation study verifies that our proposed modules leads to better situational estimation and better QA performance. + +# 5.3. Ablation Study & Analysis + +Vision & Language Encoders. We study the impacts of different visual and textual tokenizers in Table 5. It is observed that the open-vocabulary visual encoder (OpenScene) outperforms detection-based encoders (such as VoteNet and 3DETR) across all metrics. This superior performance of OpenScene is attributed to the limitations of 3D detectors, which are typically trained on a limited set of object categories, rendering them less effective in recognizing novel objects mentioned in textual prompts. Regarding language encoders, our findings indicate that a stronger backbone correlates with better performance, primarily due to its improved capability to interpret complex textual inputs. This leads to the suggestion of integrating LLM with our method to potentially further enhance performance, an avenue we intend to explore in future research. + +Situated Awareness. In Table 6 we verify the crucial role of situational awareness in 3DVL task. Firstly, we show that 3D PE, 6D rotation estimation, and anchor-based position estimation all leads to much better position and orientation estimation performance. We further establish that situational PE and visual token re-encoding modules lead to better utilization of the predicted situational vector for QA task. Additionally, We design two oracle models under the assumption of having access to the ground truth situational vector as an input. The outcomes from these models reveal a critical insight: the model fails to effectively interpret situational information when it is directly incorporated into the input visual embeddings. This underlines the necessity + +![](images/3d3791acd66cf36a696504a6298f47bf9b90e9459e4b89094fb7fa4bcc5a5786.jpg) +Figure 5. Qualitative results demonstrate significant improvement of SIG3D over prior method. The first row is results from SQA3D [44], and the second row is results from our methods. In the 3D scene, red: Ground truth (GT) vector, and blue: Estimated vector. + +of the intermediate representation and encoding mechanism we have proposed, affirming its importance in achieving optimal 3DVL task performance. + +Architectural Design. We explore different architectural design choices of our model in Table 4. We find that the number of visual tokens sampled from the visual feature embeddings affects the performance of both situational estimation and QA tasks. Sampling fewer visual tokens increases the risk of missing the region of significance, while sampling more does not lead to a better performance as well. We study the size of voxels and find $0.02\mathrm{m}$ to be the most effective choice, as OpenScene [51] backbone is pretrained with the same voxel size. We also find $\sin \theta$ , $\cos \theta$ and 6D vector representation performs a lot better than quaternion in rotation estimation task. This is consistent with the finding reported in [72]. + +# 5.4. Qualitative Analysis + +Finally, we demonstrate some qualitative results of our SIG3D in Figure 5. We show the ground truth and estimated situational vectors in red and blue colors, respectively, in their corresponding 3D scenes. We also print the answers with red cross or green checkmark indicating the correctness. It is clear that our method performs significantly better in situational estimation task, resulting in vectors very close to the ground truth in both position and orientation perspectives. The better situational awareness + +also aids the complicated embodied navigation and common sense QA activities. This also demonstrate great potential of our method in the development of indoor robotics and/or conversational agents. + +Supplementary Material. The supplementary section offers an extensive analysis, encompassing a detailed examination of the 3D visual token activation changes pre and post situational re-encoding. Additionally, it includes a comprehensive collection of positive and negative samples, an insightful failure case analysis, and a forward-looking discussion on limitations and future work. + +# 6. Conclusion + +In this paper, we introduce SIG3D, a situation-aware vision language model for 3D reasoning tasks. We propose to represent 3D scene as feature tokens, treat tokens as anchors points to estimate a situational vector from a text description, and use the estimated situation as guidance to align and re-encode the visual tokens to enhance the features for reasoning tasks. We observe consistent and significant performance gain on both situational estimation and question answering tasks. + +Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program. + +# References + +[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. ShapeGlot: Learning language for shape differentiation. In ICCV, 2019. 3 +[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In ECCV, 2020. 3 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurlPS, 2022. 1, 2, 6 +[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 2 +[5] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In CVPR, 2022. 2, 3, 5, 6, 7 +[6] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005. 6 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1, 2, 3, 6 +[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3D object localization in rgb-d scans using natural language. In ECCV, 2020. 3 +[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In ACCV, 2019. 3 +[10] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, 2021. 3, 6 +[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1 +[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In CVPR, 2019. 4, 1 +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 4, 5 +[14] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. In CVPR, 2022. 2 +[15] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, + +and Thomas Hofmann. Multi-CLIP: Contrastive Vision-Language Pre-training for Question Answering tasks in 3D Scenes. In BMVC, 2023. 3, 6 +[16] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In CVPR, 2023. 3 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2 +[18] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In CVPR, 2022. 2 +[19] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3D visual graph network for object grounding in point cloud. In ICCV, 2021. 3 +[20] François Gardères, Maryam Ziaeefard, Baptiste Abeloos, and Freddy Lecue. Conceptbert: Concept-aware representation for visual question answering. In EMNLP, 2020. 2 +[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2 +[22] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, and Jianfeng Gao. Kat: A knowledge augmented transformer for vision-and-language. In NAACL, 2022. 2 +[23] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From Images to Textual Prompts: Zero-shot Visual Question Answering with Frozen Large Language Models. In CVPR, 2023. 2 +[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 7 +[25] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3D Concept Learning and Reasoning from Multi-View Images. In CVPR, 2023. 3 +[26] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3D World into Large Language Models. NeurIPS, 2023. 1, 2, 3, 5, 6, 7 +[27] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3D instance segmentation. In AAAI, 2021. 3 +[28] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. LEMMA: A Multi-view Dataset for LE arning Multi-agent M multi-task A civities. In ECCV, 2020. 2 +[29] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In NeurIPS, 2022. 2 +[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom + +Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2 +[31] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL, 2019. 1, 2 +[32] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In iccv, 2023. 3 +[33] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2 +[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2 +[35] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021. 6 +[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven Semantic Segmentation. In ICLR, 2022. 2 +[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6 +[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ICML, 2023. 2, 6, 1 +[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 6 +[40] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang. Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In CVPR, 2023. 2, 1 +[41] Yuanze Lin, Yujia Xie, Dongdong Chen, Yichong Xu, Chenguang Zhu, and Lu Yuan. Revive: Regional visual representation matters in knowledge-based visual question answering. In NeurIPS, 2022. 2 +[42] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In CVPR, 2023. 1 +[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1 +[44] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: Situated question answering in 3D scenes. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[45] Yunze Man, Liang-Yan Gui, and Yu-Xiong Wang. BEV-Guided Multi-Modality Fusion for Driving Perception. In CVPR, 2023. 4 +[46] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 2 + +[47] Jack Merullo, Louis Castricato, Carsten Eickhoff, and Ellie Pavlick. Linearly mapping from image to text space. *ICLR*, 2023. 2 +[48] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, 2021. 7 +[49] Ziqi Pang, Ziyang Xie, Yunze Man, and Yu-Xiong Wang. Frozen Transformers in Language Models Are Effective Visual Encoder Layers. arXiv preprint arXiv:2310.12973, 2023. 2, 5, 6 +[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6 +[51] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3, 4, 7, 8, 1 +[52] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 7 +[53] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, 2019. 3, 7 +[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 4 +[55] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In EMNLP, 2019. 4, 7, 1 +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[57] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, 2022. 2 +[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2 +[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2, 1 +[60] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015. 6 +[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 1 +[62] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In NeurIPS (Datasets and Benchmarks Track), 2021. 2 + +[63] Jialin Wu, Jiasen Lu, Ashish Sabharwal, and Roozbeh Mottaghi. Multi-modal answer validation for knowledge-based vqa. In AAAI, 2022. 2 +[64] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. LLM-Grounder: Open-Vocabulary 3D Visual Grounding with Large Language Model as an Agent. arXiv preprint arXiv:2309.12311, 2023. 3 +[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3D question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3 +[66] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In CVPR, 2021. 4, 5 +[67] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, 2019. 6, 7, 1 +[68] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In CVPR, 2019. 2 +[69] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2 +[70] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. MotionGPT: Finetuned LLMs are General-Purpose Motion Generators. arXiv preprint arXiv:2306.10900, 2023. 2 +[71] Brady Zhou and Philipp Krahenbuhl. Cross-view Transformers for real-time Map-view Semantic Segmentation. In CVPR, 2022. 4 +[72] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 5, 8 +[73] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In CVPR, 2016. 2 +[74] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment. In ICCV, 2023. 3, 5, 6, 7 \ No newline at end of file diff --git a/2024/Situational Awareness Matters in 3D Vision Language Reasoning/images.zip b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f57e473abc42721eb0d7110f2f6b751e56ecfee3 --- /dev/null +++ b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fde71ae6797e03d8913ace3431c80c9ffb497e520b557f920b695982f31cfe2b +size 590973 diff --git a/2024/Situational Awareness Matters in 3D Vision Language Reasoning/layout.json b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63f0f4ca824e45b3658a9003924e4e4052f04ed3 --- /dev/null +++ b/2024/Situational Awareness Matters in 3D Vision Language Reasoning/layout.json @@ -0,0 +1,7900 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 500, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 500, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 500, + 121 + ], + "type": "text", + "content": "Situational Awareness Matters in 3D Vision Language Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 143, + 429, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 143, + 429, + 172 + ], + "spans": [ + { + "bbox": [ + 165, + 143, + 429, + 172 + ], + "type": "text", + "content": "Yunze Man Liang-Yan Gui Yu-Xiong Wang University of Illinois Urbana-Champaign" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 174, + 380, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 174, + 380, + 185 + ], + "spans": [ + { + "bbox": [ + 211, + 174, + 380, + 185 + ], + "type": "text", + "content": "{yunzem2,lgui,yxw}@illinois.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "type": "text", + "content": "Being able to carry out complicated vision language reasoning tasks in 3D space represents a significant milestone in developing household robots and human-centered embodied AI. In this work, we demonstrate that a critical and distinct challenge in 3D vision language reasoning is the situational awareness, which incorporates two key components: (1) The autonomous agent grounds its self-location based on a language prompt. (2) The agent answers open-ended questions from the perspective of its calculated position. To address this challenge, we introduce SIG3D, an end-to-end Situation-Grounded model for 3D vision language reasoning. We tokenize the 3D scene into sparse voxel representation, and propose a language-grounded situation estimator, followed by a situated question answering module. Experiments on the SQA3D and ScanQA datasets show that SIG3D outperforms state-of-the-art models in situational estimation and question answering by a large margin (e.g., an enhancement of over " + }, + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 525 + ], + "type": "text", + "content": " on situation accuracy). Subsequent analysis corroborates our architectural design choices, explores the distinct functions of visual and textual tokens, and highlights the importance of situational awareness in the domain of 3D question-answering. Project page is available at https://yunzeman.github.io/situation3d." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": "Humans learn knowledge efficiently through the interactions with the 3D world and the integration of multi-modal information, such as verbal guidance or instructions. Similarly, introducing language guidance into the visual comprehension task can greatly enhance models' learning efficiency [3, 42]. Nonetheless, despite considerable advancements in linguistic understanding [7, 11, 31, 58] and vision-language integration [3, 37, 54, 61], current methodologies remain deficient in accurately perceiving and rationalizing within real-world 3D environments, which is largely attributed to the lack of 3D situational reasoning capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Compared with machine learning models, humans put" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 213, + 545, + 363 + ], + "blocks": [ + { + "bbox": [ + 307, + 213, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 363 + ], + "type": "image", + "image_path": "3743d41e5d3a84bf69adb681218bac04137bc2965fe7e2159dbb3884a53b9f95.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 366, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 456 + ], + "type": "text", + "content": "Figure 1. Previous methods perform direct 3D vision language reasoning without modeling the situation of an embodied agent in the 3D environment. Our method, SIG3D, grounds the situational description in the 3D space, and then re-encodes the visual tokens from the agent's intended perspective before vision-language fusion, resulting in a more comprehensive and generalized 3D vision language (3DVL) representation and reasoning pipeline. Q, K, V stands for query, key, and value, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 471, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 590 + ], + "type": "text", + "content": "themselves inside the 3D world and then perceive and interact with the surrounding environment from their egoperspective (Figure 1). Such situational awareness is a crucial difference between 2D and 3D visual understanding, and a key to achieve seamless understanding of spatial concepts in more complex real-world environments. Several existing methods recognize the lack of positional understanding in 3D and propose new benchmarks and joint optimization functions [44], or positional embedding methods [26] to enhance the overall reasoning performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "However, the lack of an explicit situation modeling and situation-grounded 3D reasoning method restricts them from obtaining a generalizable and consistent 3D vision-language (VL) representation. As shown in Figure 2, the situation prediction of the state-of-the-art method [44] (in blue) diverges significantly from the ground truth vectors (in red) in almost all scenes in the dataset [13]. Moreover, our pilot study in Section 3 also reveals that situational understanding, despite being very crucial in comprehending the context of questions, only plays a minor role in the final" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13678" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 75, + 159, + 153 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 159, + 153 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 159, + 153 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 159, + 153 + ], + "type": "image", + "image_path": "f2652a4744e4f1b9b629ac3b59c6a5e5c71bcbe328d03926159a7f561e502346.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 155, + 157, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 155, + 157, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 157, + 171 + ], + "type": "text", + "content": "I am facing a trash can, while there is a door on my right." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 166, + 75, + 274, + 154 + ], + "blocks": [ + { + "bbox": [ + 166, + 75, + 274, + 154 + ], + "lines": [ + { + "bbox": [ + 166, + 75, + 274, + 154 + ], + "spans": [ + { + "bbox": [ + 166, + 75, + 274, + 154 + ], + "type": "image", + "image_path": "976466d9324826ee50a29bec14a5c7f88056478dd6b771190e13afc97ae6a01c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 155, + 277, + 170 + ], + "lines": [ + { + "bbox": [ + 165, + 155, + 277, + 170 + ], + "spans": [ + { + "bbox": [ + 165, + 155, + 277, + 170 + ], + "type": "text", + "content": "I am facing the paper towel dispenser with a chair in my six o'clock direction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 180, + 157, + 240 + ], + "blocks": [ + { + "bbox": [ + 51, + 180, + 157, + 240 + ], + "lines": [ + { + "bbox": [ + 51, + 180, + 157, + 240 + ], + "spans": [ + { + "bbox": [ + 51, + 180, + 157, + 240 + ], + "type": "image", + "image_path": "2165f6d32d3e15bd7ee574ddfc128ad3bb03b33c2d22d0851737bc3dddd20906.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 243, + 156, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 156, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 156, + 266 + ], + "type": "text", + "content": "I am facing a pool table, taking out the white ball from the side pocket. I am also facing a painting on the wall." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 167, + 178, + 268, + 245 + ], + "blocks": [ + { + "bbox": [ + 167, + 178, + 268, + 245 + ], + "lines": [ + { + "bbox": [ + 167, + 178, + 268, + 245 + ], + "spans": [ + { + "bbox": [ + 167, + 178, + 268, + 245 + ], + "type": "image", + "image_path": "c2658c3e9812b52434ac8e0ee2d0b95a7787c20d17889903217dccc87ce8cc0b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 250, + 280, + 266 + ], + "lines": [ + { + "bbox": [ + 165, + 250, + 280, + 266 + ], + "spans": [ + { + "bbox": [ + 165, + 250, + 280, + 266 + ], + "type": "text", + "content": "I am sitting on a couch with a pillow facing another couch and the pillow is on my right." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 277, + 287, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 321 + ], + "type": "text", + "content": "Figure 2. Situational estimation in existing methods [44] fails in most scenarios, indicating the missing registration between the situational description and 3D embeddings. Red: Ground truth (GT) vector. Blue: Estimated vector." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 336, + 264, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 264, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 264, + 348 + ], + "type": "text", + "content": "question-answering performance of existing methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 350, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 350, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 45, + 350, + 287, + 590 + ], + "type": "text", + "content": "In this work, we propose SIG3D, a novel approach designed to precisely model and estimate an embodied agent's ego-location and orientation from a textual description, before performing multi-modal QA tasks from the agent's egocentric perspective, as shown in Figure 1. Specifically, we leverage large-scale pretrained language and visual encoders to process the input text and 3D data, and fuse the tokens with attention modules to predict a situational vector. Previous attempts to directly predict the ego-situation are hindered by the expansive search space inherent in 3D environments. To address this challenge, we re-conceptualize the task as an anchor-based classification, where visual tokens are regarded as anchor points, and a likelihood of position together with a set of rotation parameters are concurrently regressed for each visual token. After obtaining the situational estimation, we propose a situational alignment and a situation-guided token re-encoding strategy, to perceive the environment from the agent's intended perspective. These strategies enhance the visual token with more accurate situational awareness for subsequent QA tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "text", + "content": "Experiments on two challenging 3D visual question-answering (VQA) datasets [5, 44] demonstrate the significant improvement on situational estimation and QA tasks of our model. In particular, we improve the situational estimation accuracy by over " + }, + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "text", + "content": ", and the subsequent QA performance by up to " + }, + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 592, + 286, + 687 + ], + "type": "text", + "content": ". Further qualitative and quantitative analysis verifies our design choices and highlights the significance of situational awareness in 3D reasoning tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "To sum up, our paper has the following contributions: (1) We recognize the lack of situational awareness as a sig-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "content": "nificant oversight in existing research. To address this, we introduce SIG3D, a situation-grounded 3D VL reasoning architecture, specifically designed to fill this void. (2) We propose an anchor-based approach to situational estimation, which effectively narrows the extensive search space in 3D environments for precise grounding of 3D positions and orientations with textual descriptions. Additionally, we investigate situational alignment and visual re-encoding mechanisms to leverage situational awareness for enhanced QA performance. (3) Our model demonstrates superior performance on two challenging datasets, SQA3D and ScanQA, surpassing the state of the art in both situational estimation and QA metrics. Ablation studies highlight the importance of situation-guided encoding, revealing its beneficial impact on general QA tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 266, + 392, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 266, + 392, + 280 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 392, + 280 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 288, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 503 + ], + "type": "text", + "content": "Vision Language Models (VLMs). Early transformer-driven [59] textual and visual encoders [17, 31] have facilitated great progress in recent vision language learning. Text-image contrastive models [30, 54] propose to align the feature space of two modalities with large-scale pretraining, fueling numerous downstream tasks from generalized open-vocabulary visual perception [21, 34, 36] to text-to-image generation [56]. Concurrently, some work uses text and vision encoders on separate modalities followed by feature fusion [18, 33] for multi-modal reasoning tasks. Since the emergence of Large Language Models (LLMs) [7, 58, 69], VLMs have experienced huge improvement with the help of LLMs as building blocks for multi-modal learning architectures. Specifically, recent work directly projects visual embeddings into language-space tokens as input to LLM [23, 40, 47, 70], or use the latent bottleneck structure for cross-modal visual decoding [3, 26, 37, 38], or treat LLM layers as encoder blocks for various visual tasks [49]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "content": "In the domain of visual question-answering (VQA) [4, 73], recent work has pushed the frontier towards video understanding [14, 28, 29, 40, 62], knowledge-based understanding [20, 22, 23, 41, 46, 57, 63], and commonsense reasoning [68]. Despite the outstanding performance in 2D image interpretation, most existing methods lack the capability to generalize to 3D scenarios. In contrast, our work studies the representation of visual information and its fusion with language embeddings in the 3D domain by targeting on the 3D situation-guided visual language interpretation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Grounding Language in 3D Space. Compared with 2D images, knowledge such as spatial relationships, interactive exploration, and topological analysis, which only exist in the 3D world provides additional challenges and opportunities to develop better language models with stronger commonsense reasoning capability grounded in the real-world 3D scenarios. In this direction, early work seeks" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13679" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 144 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 144 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 144 + ], + "type": "image", + "image_path": "dd7ff2f87d095cdce14564a7bb9feee5ce4a817302e0629fecdf932f2ab0596a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 147, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 202 + ], + "type": "text", + "content": "Figure 3. Results on variants of the representative SQA3D baseline method [44] demonstrate that situational understanding, despite being indispensable in perceiving the context of questions, makes negligible contribution in existing methods. This motivates a situation-guided 3D encoding mechanism in our model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 286, + 310 + ], + "type": "text", + "content": "to ground isolated objects [1, 9] or objects within more complex scenes [2, 8, 19, 27] using natural language descriptions. Recently, with more collected 3D vision language datasets, several work starts to explore language-guided 3D visual interpretation and reasoning on a diverse set of datasets, including 3D scene captioning [10], open-vocabulary segmentation [16, 32, 51], and question-answering [5, 15, 25, 65, 74]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 286, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 286, + 443 + ], + "type": "text", + "content": "The success of LLMs also elicits the usage of them in 3D vision language reasoning for task decomposition [64], data generation, and multi-modal feature fusion [26]. Motivated by ScanQA [5], SQA3D [44] takes the first step into exploring the challenging 3D situational reasoning problem by developing a situated question-answering benchmark, and proposing the first joint learning baseline on the benchmark. Our work highlights the uniqueness and significance of situational awareness in the 3D vision language learning paradigm, leading to notably better 3D situational grounding and question-answering performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 237, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 237, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 237, + 467 + ], + "type": "text", + "content": "3. Pilot Study on Situated Reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 676 + ], + "type": "text", + "content": "Despite pointing out the significance of situated understanding and reasoning, existing methods [44] fall short in providing effective situational estimation, as illustrated in Figure 2. This section delves into a pilot study examining the impact of situational understanding on downstream reasoning tasks. The SQA3D baseline [44] incorporates situational descriptions and uses ground truth (GT) situational vectors for supervision in a direct regression task. We investigate three variants of this baseline to assess the effect of situational understanding. In the first variant, we remove the situational description and supervision from the model, by passing in empty situational tokens. In another variant, we corrupt the situation supervision by introducing very large Gaussian noise to the GT vectors to effectively randomizing them. Finally, we try to encode the GT situational vector in the input with learnable multi-layer perceptron layers, which are then integrated with the visual and textual tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Figure 3 demonstrates the results of this study, revealing negligible changes in performance across these variants. Notably, corrupting the GT situational information or di" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "rectly incorporating it results in only marginal alterations in the QA outcomes. Omitting the situational description entirely from the input results in a mere " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " decline in accuracy. However, in the absence of this information, the model resorts to random guessing when determining the correct answer, as all responses depend on situational context. The findings from Figures 2 and 3 collectively indicate a deficiency in existing methods regarding situation vector estimation and the application of situational understanding in subsequent reasoning tasks. These unresolved challenges motivate the development of our proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 217, + 361, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 361, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 361, + 228 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 237, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 417 + ], + "type": "text", + "content": "An overview of our approach SIG3D is illustrated in Figure 4. Our method begins with a set of points that represent a 3D scene, accompanied by a situational description and a question, which defines the overall context of the problem. We tokenize them into separate token embeddings (Section 4.1), and ground the textual description in the 3D scene with a vector comprising of location and orientation. We find direct single vector estimation to be challenging due to the vast and complex nature of the 3D search space, so we propose an anchor-based situational estimation strategy (Section 4.2). Subsequently, we re-encode the visual tokens from the perspective of situational vectors, enhancing the situational awareness for downstream reasoning tasks (Section 4.3). The finalized visual and textual tokens are fused by a transformer decoder to generate the final response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 426, + 479, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 479, + 438 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 479, + 438 + ], + "type": "text", + "content": "4.1. Visual and Textual Tokenization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": "Leveraging input scene point clouds and textual prompts, our objective is to generate three distinct types of tokens: 3D visual tokens " + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "inline_equation", + "content": "z^{3\\mathrm{D}} \\in \\mathbb{R}^{N_v \\times C_v}" + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": ", situational tokens " + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{S}} \\in \\mathbb{R}^{N_s \\times C_s}" + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": ", and question tokens " + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{Q}} \\in \\mathbb{R}^{N_q \\times C_q}" + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": ". Each type of token is characterized by two primary components: " + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": ", representing the number of tokens, and " + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": ", encapsulating the feature embeddings. Tockenize and capture the feature embeddings for situational and question inputs, we employ a shared text tokenizer " + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^{\\mathrm{TXT}}" + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": " following prior methods [5, 44]. We assume that situational and question prompts are separated in the input data. If not, LLMs [7] can be used to parse the textual input without changing the semantic meaning of the sentences. However, there is a lack of consensus on a standard 3D visual tokenization method " + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 445, + 545, + 637 + ], + "type": "text", + "content": " that is apt for the 3DVL reasoning task, prompting a more detailed exploration in the subsequent paragraphs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Visual Tokenization. Given an input point cloud " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ", most prior methods [5, 15, 44] adopt a VoteNet [53] detector to acquire object-level tokens " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "z^{3\\mathrm{D}} \\in \\mathbb{R}^{N_{\\mathrm{obj}} \\times C_{\\mathrm{obj}}}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " as the visual representation, where " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " is the number of object proposals, and " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " is the object-level feature embeddings. However, we point out several problems with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13680" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 549, + 274 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 549, + 274 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 549, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 549, + 274 + ], + "type": "image", + "image_path": "7b27bdd4f2816e51261ad8540905c741dd7b3c0866dab967ea8a81792e51acf8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 334 + ], + "type": "text", + "content": "Figure 4. Overview of our SIG3D model, which includes 3D scene and text encoding, anchor-based situational estimation, situation-guided visual re-encoding, and multi-modal decoder modules. We tokenize the 3D scene into voxels, treat each token as an anchor point, and query the text tokens to predict a token-level position likelihood and rotation matrix to locate the situational vector associated with the text descriptions. Then we update the scene tokens with situational position encoding (PE), and finally perform the 3DVL reasoning task with a large transformer decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "content": "this abstraction strategy: (1) A detection-based tokenization method tends to ignore the non-object regions in the scene, which can be indispensable in some reasoning scenarios (e.g., carpets on the ground, ceiling, walls.) (2) After object-level abstraction, the visual representation loses the high-level information of the scene (e.g., the shape of the living room, the corner of the kitchen.) (3) A supervised detector trained from scratch can only recognize objects within the training set (e.g., only 20 categories for ScanNet [13]), meaning that the method does not have zero-shot capability to reason about novel unseen objects that are inevitably common in real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 501, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 548 + ], + "type": "text", + "content": "In light of this, we adopt a pretrained open-vocabulary voxel-based tokenization method from OpenScene [51]. The scene is first discretized into regular small 3D voxels and fed into a visual encoder for feature extraction:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 551, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 551, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 126, + 551, + 287, + 566 + ], + "type": "interline_equation", + "content": "x ^ {\\mathrm {3 D}} = \\operatorname {E} ^ {\\mathrm {3 D}} (\\mathcal {V} (\\mathbf {p})), \\tag {1}", + "image_path": "f64a2026848891f65eca49ab68495e04fbf85577d7d52d7c5db593250f0e5f38.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " represents the voxelization process, and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " is a Minkowski sparse 3D convolutional network [12]. The sparse network is pretrained by distillation from CLIP [54] embeddings of rendered multi-view 2D images, resulting in a feature map with better language alignment and 3D awareness. We take the upsampled bottleneck-layer feature embeddings from the encoder network, and compute the mean average over the " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "-axis (vertical) to project the voxels onto the " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " plane and treat the feature grids in the resulting 2D feature map as our " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " visual tokens. We find that this bird's-eye-view projection results in a more compact representation and improves the final performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 351, + 434, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 434, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 434, + 363 + ], + "type": "text", + "content": "4.2. Situational Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": "Given 3D visual tokens " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "z^{3\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": " and situational tokens " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{S}}" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": ", our objective is to estimate the situational vector " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\vec{s}" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": " referred by the situational description, which comprises of a position component " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{pos}}" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": " represented by " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": " coordinates, and a rotation component " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{rot}}" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": " represented by " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "(\\theta ,\\psi ,\\phi)" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": " Euler angles, where pitch angles " + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 369, + 545, + 562 + ], + "type": "text", + "content": " are always defined as 0, meaning that situational vectors are defined to be parallel with the ground plane. The prior method [44] utilizes a transformer block to calculate the cross-attention feature between visual and language tokens, and directly regress a final situational vector out of the averaged attention map. We find such strategy producing very inaccurate estimation, as shown in Figure 2, due to the large search space in the entire 3D volume. Inspired by the recent 3D object detection methods [45, 66, 71], we reduce the search space by turning the localization problem into a classification problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "text", + "content": "Positional Embedding and Feature Fusion. After the voxelization and 3D encoding process, each 3D token associates with a 3D position " + }, + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "text", + "content": " representing the center of its voxel. We first provide positional information to the model by generating learnable positional embeddings (PE) using a two-layer perceptron for each of the " + }, + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "text", + "content": " visual tokens, and add learnable positional embeddings to the token features " + }, + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "inline_equation", + "content": "z^{3\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 565, + 545, + 697 + ], + "type": "text", + "content": ". We use a situation interpreter [55] to extract situational information, and ask the updated visual tokens to attend to these situational tokens with several transformer layers to produce the joint feature embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "content": "Anchor-based Situational Estimation. We treat each" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13681" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "output token of the feature fusion module as an anchor point, and use it to predict a position likelihood " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "p \\in [0,1]" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " and a rotation estimation. Since each token has an associated 3D position " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": ", the position likelihood " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " indicates how likely the situational vector locates at the center of this token (voxel). We define a soft ground truth for this classification task with a Gaussian kernel, meaning that the closer a token is to the actual situational vector " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{pos}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": ", a higher ground truth probability " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " will be assigned to that token. In order to counteract the sparse supervisory signal and increase the positive supervision around the vector position, we adopt the peak enlarging technique in CenterPoint [66], where the size of the Gaussian kernel is increased (meaning that the " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " is increased) to allow denser supervision around the vector position. Furthermore, we explore different rotation representation and find that compared with quaternion and " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "(\\sin \\theta, \\cos \\theta)" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " representations, 6D vector proposed by [72] achieves the best performance. Hence, we adopt a situational estimation head with MLP layers to output 7-dimensional vector for each of the tokens, where the first channel represents the position likelihood and the other six channels represent the 6D rotation matrix. We take the center of the token with the peak position likelihood as our estimated " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{pos}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": ", and convert its corresponding 6D rotation vector as our estimated " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{rot}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": ". The estimation can be equivalently represented as a rotation matrix " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " and a translation matrix " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": ". More discussion about the architecture and design choices is in Section 5.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 418, + 228, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 228, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 228, + 431 + ], + "type": "text", + "content": "4.3. Situation-guided Visual Encoding" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "content": "After obtaining the situational estimation, we investigate a better approach to enhancing the downstream response generation, inspired by human cognitive processes. Intuitively, humans typically comprehend their immediate 3D environment by first interpreting their own situation in space, and then discerning their surroundings from an appropriate viewpoint. Our model is designed to emulate this natural strategy. Utilizing the situational vector " + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\vec{s}" + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "content": ", we adjust the coordinate system by repositioning the origin at " + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{pos}}" + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "content": ", and reorienting the axes according to " + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{rot}}" + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "content": " such that the new y-axis is aligned with the indicated direction. We keep the " + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "content": "-axis vertically oriented, and project the situational vectors onto the " + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "content": " plane. This is in line with the format of the dataset [44], where situational vectors are assumed to be parallel with the ground plane. Subsequently, we compute a new situation-guided PE for each of the " + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 47, + 438, + 289, + 715 + ], + "type": "text", + "content": " visual tokens, akin to the learnable 3D PE outlined in Section 4.2. They allow the model to grasp positional interrelations from the perspective of the current situation. These situational embeddings are added to the output embeddings of the situational estimation module, which consists of blocks featuring self-attention layers for visual tokens, succeeded by cross-attention layers that bridge visual" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "and situational information. This structure allows for the re-encoding of visual tokens under the influence of situational and question context, guiding the model to assign higher weights to situation-related and question-related visual tokens. The output, termed situation-guided visual tokens, embodies this re-contextualized understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 152, + 454, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 454, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 454, + 166 + ], + "type": "text", + "content": "4.4. Question Answering Head" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": "We follow existing methods [26] to use a large vision-language decoder to fuse the final visual and textual tokens and generate textual response to the input question. We explore both auto-regressive response generation, and classification-based answer prediction [5, 44]. For classification, we predict a vector " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "v^{\\mathrm{ans}} \\in \\mathbb{R}^{n_a}" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "inline_equation", + "content": "n_a" + }, + { + "bbox": [ + 304, + 171, + 547, + 256 + ], + "type": "text", + "content": " answer candidates in the training set following [5]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 266, + 453, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 453, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 453, + 280 + ], + "type": "text", + "content": "5. Analysis in 3D VQA Task" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 287, + 547, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 383 + ], + "type": "text", + "content": "We evaluate SIG3D for 3DVL reasoning on two challenging benchmarks, addressing both visually-oriented situational estimation and textually-focused QA tasks. We present a detailed examination of the implementation strategies adopted, the datasets employed, and the metrics applied in our research. For an exhaustive understanding, implementation and training details, and other additional information are available in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 387, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 547, + 542 + ], + "type": "text", + "content": "Datasets. We evaluate our method on SQA3D [44] and ScanQA [5], two challenging indoor 3D VQA datasets. Both datasets are derived from the ScanNet dataset [13], serving as the foundational source for their 3D scenes. SQA3D features over 33K question-answer pairs for the 3D VQA task and 26K unique situational descriptions for the situational estimation task. Each entry in this dataset includes a 3D scene point cloud, a situational description, a question, and pertinent annotations. ScanQA consists of over 41K question-answer pairs, without situational descriptions and situational annotations. We use it to demonstrate the generalizability of our method on general QA tasks. We use the splits provided by these datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. For SQA3D, in order to compare with baseline methods [44, 49, 74], we use a shallow transformer decoder task head to perform answer classification task, and evaluate the performance with exact matches (EM@1), which is equivalent to Top-1 answer accuracy. We also provide EM@1 on a breakdown of question types, including \"What,\" \"Is,\" \"How,\" \"Can,\" \"Which,\" and \"Other,\" based on the first word in the question sentence. Additionally, we evaluate situational estimation performance with localization accuracy and orientation accuracy. In both tasks, we use accuracy within different distance or angle thresholds as our metrics. For example, \"Acc@0.5m\" means accuracy of location estimation when positive threshold is set to 0.5 meter. For ScanQA," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13682" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 70, + 473, + 233 + ], + "blocks": [ + { + "bbox": [ + 124, + 70, + 473, + 233 + ], + "lines": [ + { + "bbox": [ + 124, + 70, + 473, + 233 + ], + "spans": [ + { + "bbox": [ + 124, + 70, + 473, + 233 + ], + "type": "table", + "html": "
ModelQuestion BreakdownOverall
WhatIsHowCanWhichOther
GPT-3 [7]39.746.040.545.636.138.441.0
ClipBERT [35]30.260.138.763.342.542.743.3
MCAN [67]28.959.744.168.340.740.543.4
ScanQA [5]28.665.047.366.343.942.945.3
SQA3D [44]33.566.142.469.543.046.447.2
Multi-CLIP [15]------48.0
LM4Vision [49]34.367.148.268.348.945.648.1
3D-LLM [26]36.565.647.268.848.046.348.1
3D-VisTA [74]34.863.345.469.847.248.148.5
SIG3D (Ours)35.667.248.571.449.145.852.6
", + "image_path": "e968f98110c94f87dcee2bf5fef4d0fa03c6446195059bd2fd2941924ceeadcc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 269, + 287, + 357 + ], + "blocks": [ + { + "bbox": [ + 46, + 236, + 546, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 546, + 259 + ], + "type": "text", + "content": "Table 1. Our proposed method SIG3D achieves state-of-the-art performance on SQA3D benchmark [44]. We perform the best on \"Is\", \"How\", and \"Can\" breakdown types of questions, as well as the average accuracy. The results are reported on test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 269, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 50, + 269, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 50, + 269, + 287, + 357 + ], + "type": "table", + "html": "
ModelLocalizationOrientation
Acc@0.5mAcc@1.0mAcc@15°Acc@30°
Random7.225.88.416.9
SQA3D [44]9.529.68.716.5
SQA3D (separate)10.331.417.122.8
3D-VisTA [74]11.734.516.924.2
SIG3D (Ours)27.459.128.742.5
", + "image_path": "6a88f7ab52cc259c88a72dad97ce0220ef440063a5be4815b758bc87acac0df4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 434, + 287, + 525 + ], + "blocks": [ + { + "bbox": [ + 46, + 361, + 287, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 429 + ], + "type": "text", + "content": "Table 2. Our proposed method SIG3D performs significantly better than prior method [44] in situational estimation task. \"Acc@0.5m\" stands for localization accuracy with " + }, + { + "bbox": [ + 46, + 361, + 287, + 429 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 361, + 287, + 429 + ], + "type": "text", + "content": " threshold. \"Acc@15°\" represents orientation accuracy with " + }, + { + "bbox": [ + 46, + 361, + 287, + 429 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 46, + 361, + 287, + 429 + ], + "type": "text", + "content": " threshold. separate means disabling other tasks to let the model focus on situational estimation only." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 434, + 287, + 525 + ], + "lines": [ + { + "bbox": [ + 50, + 434, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 50, + 434, + 287, + 525 + ], + "type": "table", + "html": "
ModelBLEU-1BLEU-4ROUGEMETEORCIDEr
BLIP2 [38]29.75.926.611.345.7
Flamingo [3]25.68.431.111.355.0
VN+MCAN [67]28.06.229.811.454.7
SR+MCAN [67]26.97.930.011.555.4
ScanQA [5]30.210.133.313.164.9
3D-LLM [26]39.312.035.714.569.4
SIG3D39.512.435.913.468.8
", + "image_path": "ff9c6ba01265f79c16e9ea6408657073157a0715824c21bfb36c4f71d480d2ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 530, + 287, + 597 + ], + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 597 + ], + "type": "text", + "content": "Table 3. Performance of SIG3D on ScanQA dataset [5] is on-par with the state-of-the-art with large-scale text-3D pre-training. VN and SR stand for VoteNet and ScanRefer, respectively. 3D-LLM [74] leverages pretrained 2DVL foundation models and LLM models [3, 7, 37, 38], and is pretrained on a large-scale held-in 3D-text dataset before the finetuning on ScanQA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 613, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 651 + ], + "type": "text", + "content": "we perform auto-regressive answer generation with large transformer decoder [26], and evaluate with BLEU [50], ROUGE [39], METEOR [6], and CIDEr [60] metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 208, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 208, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 208, + 672 + ], + "type": "text", + "content": "5.1. Situated Question Answering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Baselines. Our study involves a comparative analysis with a range of representative baselines on the SQA3D dataset. In particular, we evaluate against GPT-3 [7], Clip" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 271, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 546, + 415 + ], + "type": "text", + "content": "BERT [35], and MCAN [67], which are, as reported in prior work [44], baselines focused on language-only, 2D video, and 2D image QA, respectively. For GPT-3, we follow SQA3D [44] to convert the visual input into a caption using Scan2Cap [10] for LLMs to process. ScanQA [5] represents a 3D QA baseline that ignores the situational input. Both SQA3D [44] and Multi-CLIP [15] employ situational descriptions and annotations for direct regression tasks. LM4Vision [49] utilizes LLMs as visual and textual encoders. Additionally, 3D-VisTA [74] undergoes a pretraining procedure on their large-scale 3D scene-text dataset, ScanScribe, prior to the finetuning on this dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 420, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 696 + ], + "type": "text", + "content": "Situational Estimation. As shown in Table 2, our work performs significantly better than the state of the art [44, 74] in both localization and orientation estimation tasks. For 3D-VisTA [74], we use a pretrained model and finetune a new situation head with the SQA3D dataset following [44]. We also report a random baseline, in which we randomly sample position and orientation from a uniform distribution as a lower-bound performance. Note that original SQA3D performs only marginally better than the random baseline, meaning that it does not acquire any situational awareness, despite having the situational estimation loss. Disabling the QA task and asking the model to exclusively focus on the situational estimation task results in a slight better performance. Our method, with the anchor-based position likelihood estimation, results in a much better understanding of the 3D situational relationship. Our method also outperforms 3D-VisTA, which is pretrained on a large-scale 3D-text dataset, indicating that large pretraining alone is not enough to address the situational awareness problem. Note that we do not include the random baseline performance reported in [44], because each value is obtained by generating three random values and taking the closest one to the ground true, and thus it does not reflect a true \"random\" baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "text", + "content": "Situated Question Answering. SIG3D outperforms prior" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13683" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 89, + 220, + 144 + ], + "blocks": [ + { + "bbox": [ + 94, + 79, + 188, + 88 + ], + "lines": [ + { + "bbox": [ + 94, + 79, + 188, + 88 + ], + "spans": [ + { + "bbox": [ + 94, + 79, + 188, + 88 + ], + "type": "text", + "content": "(a) Number of Visual Tokens" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 89, + 220, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 89, + 220, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 89, + 220, + 144 + ], + "type": "table", + "html": "
Acc@1.0mAcc@30°EM@1
12848.938.249.2
25659.142.550.9
51257.842.150.7
", + "image_path": "521f1b5e69b3c801641ff85268a47c78db0e7b2a8c435984687d245a2c50d981.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 249, + 89, + 367, + 144 + ], + "blocks": [ + { + "bbox": [ + 268, + 79, + 350, + 88 + ], + "lines": [ + { + "bbox": [ + 268, + 79, + 350, + 88 + ], + "spans": [ + { + "bbox": [ + 268, + 79, + 350, + 88 + ], + "type": "text", + "content": "(b) Voxel size (in meters)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 249, + 89, + 367, + 144 + ], + "lines": [ + { + "bbox": [ + 249, + 89, + 367, + 144 + ], + "spans": [ + { + "bbox": [ + 249, + 89, + 367, + 144 + ], + "type": "table", + "html": "
Acc@1.0mEM@1
0.0154.149.5
0.0259.150.9
0.0547.348.8
", + "image_path": "c2800257079b5c6dfeadd5dae2322d175306ff92ab071c2da7090af59beb629e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 392, + 89, + 535, + 144 + ], + "blocks": [ + { + "bbox": [ + 422, + 79, + 509, + 89 + ], + "lines": [ + { + "bbox": [ + 422, + 79, + 509, + 89 + ], + "spans": [ + { + "bbox": [ + 422, + 79, + 509, + 89 + ], + "type": "text", + "content": "(c) Rotation representation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 392, + 89, + 535, + 144 + ], + "lines": [ + { + "bbox": [ + 392, + 89, + 535, + 144 + ], + "spans": [ + { + "bbox": [ + 392, + 89, + 535, + 144 + ], + "type": "table", + "html": "
Acc@30°EM@1
Quaternion31.450.0
6D vector42.550.9
sinθ, cosθ42.650.6
", + "image_path": "8e92edbac765b99b3df55a336252871ba6062f618745bcb82cb54048c5b19d24.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 186, + 287, + 318 + ], + "blocks": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 547, + 175 + ], + "type": "text", + "content": "Table 4. Ablation study validates that our various design choices improves the performance. \"Acc@1.0m\", \"Acc@30°\", and \"EM@1\" are accuracy (%) for localization estimation, orientation estimation, and QA tasks. Our settings are marked in gray." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 186, + 287, + 318 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 287, + 318 + ], + "type": "table", + "html": "
Acc@1.0mAcc@30°EM@1
3D Vision Encoders
Text-only (no vision input)--47.5
VoteNet [53]37.428.249.1
3DETR [48]47.229.149.4
OpenScene - OpenSeg [51]57.541.650.2
OpenScene - LSeg [51]59.142.550.9
Language Tokenizer / Encoders
GloVe + LSTM [24, 52]44.330.948.7
SBERT - MiniLM [55]56.138.649.4
SBERT - MPNet [55]55.940.649.7
SBERT - MPNet (finetune)59.142.550.9
", + "image_path": "fddfa3960749c3c2ed36d9a1bc4cdf7f1e2569f74a268f73ffeb0238998da4a1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 321, + 287, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 287, + 389 + ], + "type": "text", + "content": "Table 5. Performance of SIG3D improves with stronger visual and language encoders. We find that open-vocabulary point encoder and MPNet-based sentence BERT leads to best performance. \"Acc@1.0m\" and \"Acc@30°\" stands for localization and orientation accuracy in situational estimation task. \"EM@1\" demonstrate exact match metric in QA task." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 408, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 506 + ], + "type": "text", + "content": "methods in most question breakdown categories and overall accuracy, as shown in Table 1. Our work achieves leading results without large-scale pretraining (compared with 3D-VisTA) and LLM (compared with GPT-3), indicating its superiority in situational awareness. Note that the LLM baseline GPT-3 achieves the best performance on the \"What\" category, suggesting the potential of stronger language encoder in interpreting the complicated question prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 519, + 261, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 261, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 261, + 533 + ], + "type": "text", + "content": "5.2. General Question Answering on ScanQA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 539, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 601 + ], + "type": "text", + "content": "Baselines. We compare with 2D image VQA MCAN-based baselines [67], ScanQA [5], 3D-LLM [26] which leverages large-scale pretrained 2D VLMs and LLMs as backbone models, and 3D-VisTA [74] pretrained on their proposed large-scale 3D-text dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Question Answering. As shown in Table 3, despite that the questions do not explicitly require situational understanding to answer in ScanQA, SIG3D achieves comparable results with state-of-the-art methods without the large-scale 3D-text pretraining and powerful 2D VLM and LLM backbone models. Our work pretrained on SQA3D [44] leads to higher performance on BLEU-1, BLEU-4, and ROUGE metrics, showing its generalizability on general 3D QA scenarios." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 186, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 309, + 186, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 309, + 186, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 309, + 186, + 545, + 313 + ], + "type": "table", + "html": "
Acc@1.0mAcc@30°EM@1
Baseline (joint optimization)29.523.147.7
How to achieve better situational estimation
+ 3D PE38.823.647.8
+ 6D Representation38.527.447.7
+ Anchor-based Estimation58.841.948.2
How to utilize situational estimation for better QA
+ 3D Situational PE58.941.850.0
+ Visual Token Re-encoding59.142.550.9
Oracle Models (Ground Truth Situation Information)
Situation as direct input10010047.7
Situation as intermediate input10010053.9
", + "image_path": "3bd148491d7663bdf3fe91fe0dcac67f9e9289ddfe1b3b17671d20a50a2213b4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 316, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 306, + 316, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 545, + 340 + ], + "type": "text", + "content": "Table 6. Ablation study verifies that our proposed modules leads to better situational estimation and better QA performance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 355, + 454, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 355, + 454, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 355, + 454, + 369 + ], + "type": "text", + "content": "5.3. Ablation Study & Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": "Vision & Language Encoders. We study the impacts of different visual and textual tokenizers in Table 5. It is observed that the open-vocabulary visual encoder (OpenScene) outperforms detection-based encoders (such as VoteNet and 3DETR) across all metrics. This superior performance of OpenScene is attributed to the limitations of 3D detectors, which are typically trained on a limited set of object categories, rendering them less effective in recognizing novel objects mentioned in textual prompts. Regarding language encoders, our findings indicate that a stronger backbone correlates with better performance, primarily due to its improved capability to interpret complex textual inputs. This leads to the suggestion of integrating LLM with our method to potentially further enhance performance, an avenue we intend to explore in future research." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "content": "Situated Awareness. In Table 6 we verify the crucial role of situational awareness in 3DVL task. Firstly, we show that 3D PE, 6D rotation estimation, and anchor-based position estimation all leads to much better position and orientation estimation performance. We further establish that situational PE and visual token re-encoding modules lead to better utilization of the predicted situational vector for QA task. Additionally, We design two oracle models under the assumption of having access to the ground truth situational vector as an input. The outcomes from these models reveal a critical insight: the model fails to effectively interpret situational information when it is directly incorporated into the input visual embeddings. This underlines the necessity" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13684" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 340 + ], + "type": "image", + "image_path": "3d3791acd66cf36a696504a6298f47bf9b90e9459e4b89094fb7fa4bcc5a5786.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 345, + 546, + 368 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 546, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 546, + 368 + ], + "type": "text", + "content": "Figure 5. Qualitative results demonstrate significant improvement of SIG3D over prior method. The first row is results from SQA3D [44], and the second row is results from our methods. In the 3D scene, red: Ground truth (GT) vector, and blue: Estimated vector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 416 + ], + "type": "text", + "content": "of the intermediate representation and encoding mechanism we have proposed, affirming its importance in achieving optimal 3DVL task performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "text", + "content": "Architectural Design. We explore different architectural design choices of our model in Table 4. We find that the number of visual tokens sampled from the visual feature embeddings affects the performance of both situational estimation and QA tasks. Sampling fewer visual tokens increases the risk of missing the region of significance, while sampling more does not lead to a better performance as well. We study the size of voxels and find " + }, + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "inline_equation", + "content": "0.02\\mathrm{m}" + }, + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "text", + "content": " to be the most effective choice, as OpenScene [51] backbone is pretrained with the same voxel size. We also find " + }, + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "inline_equation", + "content": "\\sin \\theta" + }, + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 420, + 288, + 577 + ], + "type": "text", + "content": " and 6D vector representation performs a lot better than quaternion in rotation estimation task. This is consistent with the finding reported in [72]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 165, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 165, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 165, + 600 + ], + "type": "text", + "content": "5.4. Qualitative Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Finally, we demonstrate some qualitative results of our SIG3D in Figure 5. We show the ground truth and estimated situational vectors in red and blue colors, respectively, in their corresponding 3D scenes. We also print the answers with red cross or green checkmark indicating the correctness. It is clear that our method performs significantly better in situational estimation task, resulting in vectors very close to the ground truth in both position and orientation perspectives. The better situational awareness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 380, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 429 + ], + "type": "text", + "content": "also aids the complicated embodied navigation and common sense QA activities. This also demonstrate great potential of our method in the development of indoor robotics and/or conversational agents." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 431, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 546, + 515 + ], + "type": "text", + "content": "Supplementary Material. The supplementary section offers an extensive analysis, encompassing a detailed examination of the 3D visual token activation changes pre and post situational re-encoding. Additionally, it includes a comprehensive collection of positive and negative samples, an insightful failure case analysis, and a forward-looking discussion on limitations and future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": "In this paper, we introduce SIG3D, a situation-aware vision language model for 3D reasoning tasks. We propose to represent 3D scene as feature tokens, treat tokens as anchors points to estimate a situational vector from a text description, and use the estimated situation as guidance to align and re-encode the visual tokens to enhance the features for reasoning tasks. We observe consistent and significant performance gain on both situational estimation and question answering tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13685" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. ShapeGlot: Learning language for shape differentiation. In ICCV, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3D: Neural listeners for fine-grained 3D object identification in real-world scenes. In ECCV, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 224 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurlPS, 2022. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 259 + ], + "type": "text", + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual question answering. In ICCV, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 294 + ], + "type": "text", + "content": "[5] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In CVPR, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "text", + "content": "[6] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 287, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 405 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 407, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 441 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3D object localization in rgb-d scans using natural language. In ECCV, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "text", + "content": "[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In ACCV, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "type": "text", + "content": "[10] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, 2021. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "type": "text", + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In CVPR, 2019. 4, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "type": "text", + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 1, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[14] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[15] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "and Thomas Hofmann. Multi-CLIP: Contrastive Vision-Language Pre-training for Question Answering tasks in 3D Scenes. In BMVC, 2023. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[16] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. PLA: Language-Driven Open-Vocabulary 3D Scene Understanding. In CVPR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "text", + "content": "[18] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[19] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3D visual graph network for object grounding in point cloud. In ICCV, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "text", + "content": "[20] François Gardères, Maryam Ziaeefard, Baptiste Abeloos, and Freddy Lecue. Conceptbert: Concept-aware representation for visual question answering. In EMNLP, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "text", + "content": "[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[22] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, and Jianfeng Gao. Kat: A knowledge augmented transformer for vision-and-language. In NAACL, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "text", + "content": "[23] Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From Images to Textual Prompts: Zero-shot Visual Question Answering with Frozen Large Language Models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "type": "text", + "content": "[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "text", + "content": "[25] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3D Concept Learning and Reasoning from Multi-View Images. In CVPR, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "type": "text", + "content": "[26] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3D World into Large Language Models. NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[27] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3D instance segmentation. In AAAI, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "text", + "content": "[28] Baoxiong Jia, Yixin Chen, Siyuan Huang, Yixin Zhu, and Song-Chun Zhu. LEMMA: A Multi-view Dataset for LE arning Multi-agent M multi-task A civities. In ECCV, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "text", + "content": "[29] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13686" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[31] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[32] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In iccv, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[33] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "text", + "content": "[35] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "text", + "content": "[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven Semantic Segmentation. In ICLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "type": "text", + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 377, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 419 + ], + "type": "text", + "content": "[38] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ICML, 2023. 2, 6, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 443 + ], + "type": "text", + "content": "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[40] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang. Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In CVPR, 2023. 2, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[41] Yuanze Lin, Yujia Xie, Dongdong Chen, Yichong Xu, Chenguang Zhu, and Lu Yuan. Revive: Regional visual representation matters in knowledge-based visual question answering. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "type": "text", + "content": "[42] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In CVPR, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 601 + ], + "type": "text", + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "type": "text", + "content": "[44] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: Situated question answering in 3D scenes. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[45] Yunze Man, Liang-Yan Gui, and Yu-Xiong Wang. BEV-Guided Multi-Modality Fusion for Driving Perception. In CVPR, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[46] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[47] Jack Merullo, Louis Castricato, Carsten Eickhoff, and Ellie Pavlick. Linearly mapping from image to text space. *ICLR*, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 137 + ], + "type": "text", + "content": "[48] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[49] Ziqi Pang, Ziyang Xie, Yunze Man, and Yu-Xiong Wang. Frozen Transformers in Language Models Are Effective Visual Encoder Layers. arXiv preprint arXiv:2310.12973, 2023. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 185, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 215 + ], + "type": "text", + "content": "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[51] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3, 4, 7, 8, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 262, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 293 + ], + "type": "text", + "content": "[52] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 295, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 326 + ], + "type": "text", + "content": "[53] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, 2019. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 328, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 380 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "text", + "content": "[55] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In EMNLP, 2019. 4, 7, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "text", + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[57] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 494, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 547 + ], + "type": "text", + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "text", + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2, 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "text", + "content": "[60] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "type": "text", + "content": "[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[62] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In NeurIPS (Datasets and Benchmarks Track), 2021. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13687" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 534 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[63] Jialin Wu, Jiasen Lu, Ashish Sabharwal, and Roozbeh Mottaghi. Multi-modal answer validation for knowledge-based vqa. In AAAI, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[64] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. LLM-Grounder: Open-Vocabulary 3D Visual Grounding with Large Language Model as an Agent. arXiv preprint arXiv:2309.12311, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "text", + "content": "[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3D question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "text", + "content": "[66] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In CVPR, 2021. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[67] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, 2019. 6, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[68] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In CVPR, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "text", + "content": "[69] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[70] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. MotionGPT: Finetuned LLMs are General-Purpose Motion Generators. arXiv preprint arXiv:2306.10900, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "text", + "content": "[71] Brady Zhou and Philipp Krahenbuhl. Cross-view Transformers for real-time Map-view Semantic Segmentation. In CVPR, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "text", + "content": "[72] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "text", + "content": "[73] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In CVPR, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 534 + ], + "type": "text", + "content": "[74] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment. In ICCV, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13688" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_content_list.json b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d29f21ee00601ac1df5d728da16d2fb7409fe289 --- /dev/null +++ b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_content_list.json @@ -0,0 +1,1539 @@ +[ + { + "type": "text", + "text": "Skeleton-in-Context: Unified Skeleton Sequence Modeling with In-Context Learning", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinshun Wang $^{1,2,*}$ , Zhongbin Fang $^{1,*}$ , Xia Li $^{3}$ , Xiangtai Li $^{4}$ , Chen Chen $^{5}$ , Mengyuan Liu $^{2, \\boxtimes}$ , Sun Yat-sen University", + "bbox": [ + 116, + 188, + 852, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Key Laboratory of General Artificial Intelligence, Peking University, Shenzhen Graduate School", + "bbox": [ + 166, + 227, + 799, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science, ETH Zurich $^{4}$ S-Lab, Nanyang Technological University", + "bbox": [ + 210, + 244, + 756, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Center for Research in Computer Vision, University of Central Florida", + "bbox": [ + 271, + 263, + 696, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangxsh36,fangzhb5}@mail2.sysu.edu.cn，xia.li@inf.ethz.ch，xiangtai94@gmail.com", + "bbox": [ + 127, + 281, + 841, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chen.chen@crcv.ucf.edu, liumengyuan@pku.edu.cn", + "bbox": [ + 277, + 300, + 684, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/fanglaosi/Skeleton-in-Context", + "bbox": [ + 267, + 319, + 694, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 359, + 313, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In-context learning provides a new perspective for multi-task modeling for vision and NLP. Under this setting, the model can perceive tasks from prompts and accomplish them without any extra task-specific head predictions or model fine-tuning. However, skeleton sequence modeling via in-context learning remains unexplored. Directly applying existing in-context models from other areas onto skeleton sequences fails due to the similarity between inter-frame and cross-task poses, which makes it exceptionally hard to perceive the task correctly from a subtle context. To address this challenge, we propose Skeleton-in-Context (SiC), an effective framework for in-context skeleton sequence modeling. Our SiC is able to handle multiple skeleton-based tasks simultaneously after a single training process and accomplish each task from context according to the given prompt. It can further generalize to new, unseen tasks according to customized prompts. To facilitate context perception, we additionally propose a task-unified prompt, which adaptively learns tasks of different natures, such as partial joint-level generation, sequence-level prediction, or 2D-to-3D motion prediction. We conduct extensive experiments to evaluate the effectiveness of our SiC on multiple tasks, including motion prediction, pose estimation, joint completion, and future pose estimation. We also evaluate its generalization capability on unseen tasks such as motion-in-between. These experiments show that our model achieves state-of-the-art multi-task performance and even outperforms single-task methods on certain tasks.", + "bbox": [ + 75, + 378, + 473, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The idea of multi-task models has attracted much research interest over the years in a wide range of areas [6, 16, 17, 58], such as computer vision and natural language processing", + "bbox": [ + 75, + 830, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50b983fa9eab11369c47a0eaa679044708a3e891ab5e57c82a5c2dc72a9227ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 358, + 883, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2de66d6fc38ae2c4f4e17186707ca7dd9c91d6dff9c04e373f646dc75e83679.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 446, + 885, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f8b91c68bb67cb337a07a1263d4fa0a79c3d915a703f9d361b17a941f59c566.jpg", + "image_caption": [ + "Figure 1. In-context learning in a) image modeling [58], b) point cloud modeling [16], and c) skeleton sequence modeling (ours)." + ], + "image_footnote": [], + "bbox": [ + 503, + 540, + 883, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(NLP). To build such models, most previous works focus on deploying task-specific heads on a shared backbone and learning them through pretraining and fine-tuning. Models with such task-specific architectures are not learned in an end-to-end manner and are hard to generalize to new, unseen tasks and data. Most existing methods for skeleton-based human-centric tasks rely on task-specific designs and/or non-end-to-end frameworks to accomplish multiple tasks. Several works [9, 36] learn unified human motion representations. In particular, MotionBERT [77] adopts two-stage training, including pre-training on the 2D-3D lifting task", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ☑ Corresponding author is Mengyuan Liu.", + "bbox": [ + 99, + 886, + 442, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2436", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and fine-tuning on each task. UniHCP [12] connects a well-designed shared backbone network to specific task heads according to different tasks. Meanwhile, UPS [17] unifies action labels and motion sequences into text-based language sequences and fine-tunes them on each task for better performance. These models can have a limited scope of tasks they are capable of handling and lack flexibility. This limits their capability to generalize to unseen tasks and data. To overcome these challenges, in-context learning represents a new trend to build generalist models, which was originally proposed in NLP [6, 45, 46] and then introduced into other areas such as images and point clouds. Some generalist models [5, 58, 59, 72] utilize the Masked Image Modeling (MIM) framework to exhibit multi-task capability via in-context learning in image processing, as shown in Fig. 1(a). Similarly, in Fig. 1(b), Point-In-Context [16] proposes a key joint sampling module and introduces in-context learning into point cloud processing. Specifically, in-context models learn patterns of provided input-output pairs, called in-context examples or prompts, and then perform the desired task according to the prompt. An in-context model can accomplish diverse tasks end-to-end without any task-specific heads or fine-tuning. It also has excellent generalization capability. Directly adapting existing in-context models from other areas onto skeleton sequences fails because they largely focus on static objects (images or point clouds) and thus cannot capture spatial-temporal dependencies in skeleton sequences. To the best of our knowledge, in-context learning remains unexplored in 3D skeleton sequence modeling.", + "bbox": [ + 76, + 90, + 472, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Skeleton-in-Context (SiC), a novel framework for skeleton sequence modeling, which is able to process multiple tasks simultaneously after a single training process without any task-specific designs, as Fig. 1(c) shows. Specifically, SiC perceives tasks from context, i.e., the prompt from a desired task provided as additional input, and then accomplishes the task for the query. SiC employs two types of prompts to facilitate context perception: the task-guided prompt (TGP) and the task-unified prompt (TUP). TGP provides a task template for the model, giving the model the ability to make analogies. This is the key to the model's capability to generalize to new tasks, as we can guide the model to do unseen tasks by customizing specific prompts. TUP can be implemented in two ways: static and dynamic. The static TUP leverages prior knowledge shared among tasks, while the dynamic TUP adaptively learns to perceive task context. Skeleton-in-Context provides a new perspective for training a skeleton-based multi-task in-context model. Since there is no existing benchmark for measuring such a model, we establish a new in-context benchmark based on four tasks, including three skeleton-based tasks: motion prediction, pose estimation, and joint completion, and a new task: future pose estimation.", + "bbox": [ + 76, + 532, + 472, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows: 1) We introduce", + "bbox": [ + 96, + 885, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a novel framework, Skeleton-in-Context (SiC), designed to handle multiple skeleton-based tasks simultaneously. Our approach learns the potential pattern of the given prompt and performs it on the query sample. 2) We establish a new 3D human in-context learning benchmark comprised of multiple skeleton-based tasks, such as motion prediction, pose estimation, joint completion, and future motion estimation. We further benchmark several representative in-context learning baselines for reference. 3) We conduct extensive experiments to evaluate our SiC. SiC achieves state-of-the-art results in multi-tasking and even outperforms task-specific models on certain tasks. Moreover, SiC also generalizes well to unseen tasks such as motion in-between.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 301, + 640, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skeleton-Based Motion Analysis. Many of the skeleton-based human-centric tasks have shared similar development patterns in terms of the methods used to address them. On human motion prediction, early successes are achieved with RNNs and CNNs [28, 40]. Nowadays, the task is dominated by Graph Convolutional Networks (GCNs). LTD [38] considers each joint coordinate as a node in a graph on which to employ learnable spatial graph convolutions. Follow-up works [13, 39, 60, 61] focus on designing better graph convolution networks. Different from motion prediction, the goal of motion synthesis is to complete the human motion sequence. To complete the missing parts of human motion, many works adopt convolutional models [22, 26, 68] or generative adversarial networks [7, 20, 23, 76]. One category of 3D pose estimation [18, 32, 34] is to lift the estimated 2D pose to 3D, which often employs spatial-temporal convolution [8, 10, 11, 41, 44, 57, 67] or transformer-based [14, 49, 71, 73-75] methods. However, these approaches focus on designing customized architectures specific to a task, which are limited in range of applications. To this end, we propose to unify the input and output formats of multiple tasks and utilize a single model to process multiple tasks simultaneously via in-context learning.", + "bbox": [ + 496, + 325, + 895, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In-Context Learning. The emergence of in-context learning provides a new paradigm for achieving multitasking, which originated from Natural Language Processing (NLP) [1, 15, 45-47]. It endows models with the capability to handle multiple tasks and even unseen tasks by incorporating domain-specific input-target pairs, known as prompts, into the query sample [6]. Following NLP, in-context learning is extended to computer vision in some work, such as 2D images [51, 56, 58, 59, 62, 72] and 3D point clouds [3, 16, 64]. Specifically, Visual Prompt [5], Painter [58], and Point-In-Context [16] are pioneers of in-context methods in computer vision. They verify the effectiveness and flexibility of in-context learning in vision-based tasks, and most of them adopt pure-transformer-based approaches [4, 21, 29, 63]. Meanwhile, several approaches", + "bbox": [ + 496, + 674, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2437", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "deeply explore Prompt Engineering [3, 51, 72] and awe-inspiring generalization [16, 59, 64] in in-context learning. Recently, LVM [2] unifies most vision tasks and data via joint co-training on one visual in-context model. Compared to the works focused on static objects (images and point clouds), our work mainly delves into modeling in-context dynamic skeleton sequences.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified Skeleton-Centric Models. A generalist model that can excel in handling multiple tasks is a profound step toward general artificial intelligence [45, 58]. Previous works [25, 27, 31, 35, 50, 53, 66, 69] give the effort to explore co-training several human-centric tasks. Meanwhile, some works [30, 42] show great progress in simultaneously predicting parsing and pose. Recent research [9, 36] demonstrates the effectiveness of modeling unified human motion sequences. In particular, UniHCP [12] designs specific task heads according to different 2D human-centric tasks. Motion-BERT [77] is pre-trained in the 2D-3D lifting task and fine-tuned on each task. UPS [17] unifies heterogeneous human behavior understanding tasks via language sequences and cycle training for each task in order. Nevertheless, these works are limited by the dilemma of complex training stages or task heads, which introduce extra costs. We integrate the output format of different tasks in skeleton sequence format and train the model just once.", + "bbox": [ + 75, + 198, + 473, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. In-Context Skeleton Sequence Modeling", + "text_level": 1, + "bbox": [ + 76, + 481, + 437, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Formulation", + "text_level": 1, + "bbox": [ + 76, + 506, + 210, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the skeleton sequence data in skeleton-based tasks are sequential by nature, we treat the skeleton at each frame as a token, which can be processed by transformer [54]. Following previous works [5, 16, 58], we curate an in-context skeleton sequence paradigm. Given an input-target skeleton sequence pair as a task-guided prompt $G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\}$ and a query input $Q_{j}^{k}$ , where $I_{i}^{k}, T_{i}^{k}, Q_{j}^{k} \\in \\mathbb{R}^{F \\times J \\times 3}$ each have $F$ frames and $J$ joints, the model is expected to perceive the context from the prompt, i.e., which task it represents, and perform the corresponding operations on the query in analogy to the prompt. This process can be formulated as:", + "bbox": [ + 75, + 529, + 472, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = f \\left(G _ {i} ^ {k}, Q _ {j} ^ {k}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 705, + 470, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $k$ represents the task index, while $i$ and $j$ indicate different example indexes. Thus, we can determine what task will be performed on the query sample by selecting the task-specific prompt from the corresponding training set.", + "bbox": [ + 75, + 732, + 470, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task Definitions", + "text_level": 1, + "bbox": [ + 76, + 801, + 236, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To measure multitasking capability, we establish a new in-context benchmark from common skeleton-based datasets, including H3.6M [24], AMASS [37], and 3DPW [55]. In this benchmark, we select three common tasks: motion prediction, pose estimation, and joint completion, and a new", + "bbox": [ + 75, + 824, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c3f6a43a5f06aa2a32c8a96539d22a1ca563999491ee6f59866ccf78f60081f0.jpg", + "image_caption": [ + "Figure 2. Top: Training with MIM-style framework [21, 43] used in previous works [5, 16, 58]. The model is able to reconstruct the masked frames well. Bottom: During inference, the reconstructed sequence gradually shrinks the skeleton as time goes by because the model only learns frame interpolation during training. Once the model loses subsequent reference frames when interpolating frames, the generated sequence will tend to shrink." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task: future pose estimation. Consequently, we conduct extensive experiments on a large and diverse dataset collection, which consists of 176K skeleton sequence samples.", + "bbox": [ + 496, + 377, + 893, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion Prediction. Motion prediction [38] is to predict future motion based on historical motion. There is continuity between the input and output sequences, which constitute a complete motion. Thus, the input $I$ and target $T$ are complete skeleton sequences with shapes of $(F, J, 3)$ .", + "bbox": [ + 496, + 429, + 893, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose Estimation. Pose estimation [18, 32, 34] is to lift 2D poses to 3D. The model takes the detected 2D skeleton sequence $I \\in \\mathbb{R}^{F \\times J \\times 2}$ , which contain noise, as input, and estimates the accurate 3D skeleton sequence $T \\in \\mathbb{R}^{F \\times J \\times 3}$ . Note that we expand the 2D input to 3D by filling the 3rd dimension with 0, in order to unify the data format with other tasks. Therefore, the input skeleton sequence becomes the same shape $(F, J, 3)$ as the target sequence.", + "bbox": [ + 496, + 510, + 893, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Future Pose Estimation. Based on prediction and estimation, we propose a new task: future pose estimation, which is to predict future 3D motion based on history 2D poses. As 2D images are easier to obtain than 3D data, this task unlocks a host of new real-world applications. Given a historical 2D skeleton sequence $I \\in \\mathbb{R}^{F \\times J \\times 3}$ (one dimension with 0), future pose estimation expects the model to output its future 3D skeleton sequence $T \\in \\mathbb{R}^{F \\times J \\times 3}$ . To our knowledge, no work has ever proposed future pose estimation.", + "bbox": [ + 496, + 637, + 893, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Joint Completion. Joint completion is to recover skeleton sequences that have missing joints. This task can reflect cases where joints may be missing or displaced when collected with Lidar cameras [33, 48]. We set two levels to verify the ability to complete joints, which are $40\\%$ and $60\\%$ of the joints missing in each frame. Also, the coordinates of missing joints are replaced with 0, with the goal of unifying the input skeleton sequence format with $(F, J, 3)$ .", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2438", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0ec1efdc4867b9d4bb6239732dfa04df9a7098f76d6928d5cc6b1799ab26928c.jpg", + "image_caption": [ + "Figure 3. Overall framework of our Skeleton-in-Context. Specifically, we establish a skeleton bank by integrating training sets under different tasks, which contain a large amount of input-target pairs performing different tasks. Next, we randomly select a sample pair as the task-guided prompt (TGP) and a query input from the skeleton bank, undergo encoding and concatenating, respectively, and then input them into the transformer in parallel. In particular, during this process, the query input and task-unified prompt (TUP) are combined to form a new query. After iterating $n_1$ times, the TGP and query are aggregated through $\\mathcal{A}(\\cdot)$ and then input into the transformer for $n_2$ iterations. Lastly, the second half of the model output is used as our prediction." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 872, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Skeleton-in-Context", + "text_level": 1, + "bbox": [ + 76, + 388, + 272, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In-context learning on other data modalities such as images and point clouds [5, 16, 58] largely employ Mask Image Modeling (MIM) [21] to reconstruct the randomly masked elements. During inference, only the target elements are masked. On skeleton sequences, MIM fails due to spatial-temporal complexity and inter-pose similarity, which makes it outstandingly hard to perceive the task correctly from a subtle context. In our Skeleton-in Context, we employ a new framework to address this challenge.", + "bbox": [ + 75, + 414, + 470, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Skeleton Sequence Prompt", + "text_level": 1, + "bbox": [ + 76, + 559, + 318, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-Guided Prompt. In in-context learning, a prompt is needed for the model to learn the underlying task pattern from context. As mentioned in Sec. 3.1, we propose the Task-Guided Prompt (TGP) $G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\}$ , which comprises of an example skeleton pair performing a specific task $k$ . In this way, our proposed model is able to perform a variety of tasks according to the provided prompts that could be seen or even unseen in the training set. It is worth noting that our model can perform unseen tasks and transfer to datasets that are different styles from the training dataset. It will be elaborated in Sec. 5.2.", + "bbox": [ + 75, + 583, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Avoiding Over-Fitting. Previous works have made great progress in unifying different tasks via the MIM-style framework [5, 16, 58]. Compared with static objects in 2D and 3D, such as images and point clouds, the obstacle to in-context learning of skeleton sequences lies in the wondrous similarity of each frame, so directly using the random masking strategy [5, 16, 58] will cause the model to fall into overfitting, leading it to only learn simple interpolation frames rather than capturing meaningful motion sequence representations during the training stage. As shown in Fig. 2, when", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we directly utilize the MPM-style [21, 43] framework to train our model, although it can perfectly reconstruct the masked frame during training, the experimental results reveal that it outputs a devastating result during inference. It is obvious that the output skeleton sequence gradually shrinks as the length increases because the next frame of the current frame is 0, and interpolating with 0 will lead to shrinking. Such a phenomenon underscores the challenges of learning contextual information about dynamic skeleton sequences.", + "bbox": [ + 496, + 388, + 893, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-Unified Prompt. With the Task-Guided Prompt, which exhibits unique capabilities specific to in-domain or out-of-domain tasks, mitigating the gap between different tasks presents a new challenge. Meanwhile, as discussed above, a straight extension of previous works will lead to a simple model that can only perform the interpolation task based on the given skeleton sequence. To tackle these challenges, we introduce the Task-Unified Prompt (TUP), a well-designed and task-agnostic module, which can adaptively learn to incorporate multiple tasks into the framework of in-context learning, i.e., a one-off end-to-end framework for all tasks without having to deploy task-specific heads or fine-tuning. We introduce two versions of TUP:", + "bbox": [ + 496, + 526, + 893, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Version 1: Static-TUP with static pseudo pose:", + "bbox": [ + 517, + 722, + 839, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU ^ {S} = \\mathcal {E} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\sum_ {k = 1} ^ {K} \\left(T _ {m} ^ {k}\\right) \\in \\mathbb {R} ^ {F \\times J \\times 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 747, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}(\\cdot)$ is a embedding layer. $M$ and $K$ indicate the total number of training samples and tasks, respectively.", + "bbox": [ + 496, + 799, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Version 2: Dynamic-TUP with dynamic pseudo pose:", + "bbox": [ + 517, + 830, + 880, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU ^ {D} = W _ {1} W _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 856, + 890, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{1}\\in \\mathbb{R}^{F\\times J\\times 1}$ and $W_{2}\\in \\mathbb{R}^{1\\times 1\\times C}$ .", + "bbox": [ + 500, + 883, + 763, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2439", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Comparison of task-specific models, multi-task models, and in-context models on four human skeleton tasks. For Motion Prediction (MP.), Joint Completion (JC.), and Future Pose Estimation (FPE.), we report Mean Per Joint Position Error (mm) [24]. For Pose Estimation (PE.), we additionally report another indicator, Normalized Mean Per Joint Position Error (N-MPJPE) [77]. For joint completion, we report two settings: $40\\%$ and $60\\%$ of the joints missing in each frame. $\\dagger$ represents the multi-task-capable models re-structured from task-specific models. $\\ddagger$ represents the multi-task-capable models re-structured from multi-stage models. Static-SiC and Dynamic-SiC represent the static and dynamic versions of TUP, respectively.", + "bbox": [ + 75, + 88, + 893, + 172 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b4e362122d937656b351b2e1d94fc126870d52633c0a303b6e8fbd52855833ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsVenuesMP. (AMASS)PE. (H3.6M)JC. (3DPW)FPE. (H3.6M)
80ms160ms200ms320ms400msAvg.MPJPE↓ Avg.N-MPJPE↓ Avg.MPJPE↓ 40%60% Avg.MPJPE↓ 200ms300ms Avg.
Task-Specific Model: one architecture for one task
siMLPe [19]WACV'2312.121.625.434.439.926.758.949.748.458.553.575.4
EqMotion [65]CVPR'2314.426.631.744.050.033.363.259.340.449.745.169.5
STCFormaler [52]CVPR'2319.525.128.235.841.528.754.645.231.246.738.960.9
GLA-GCN [70]ICCV'2319.525.128.235.342.030.049.640.462.272.567.449.5
MotionBERT [77]ICCV'2312.921.425.536.342.326.749.441.129.344.336.875.5
Multi-Task Model: one architecture (w. multiple task-specific heads) for multiple tasks
siMLPe†[41]WACV'2319.030.035.346.852.436.774.660.059.067.463.264.4
EqMotion†[65]CVPR'2314.025.330.240.546.631.3163.3110.138.956.047.588.1
STCFormaler†[52]CVPR'2317.428.031.543.040.433.955.645.633.650.642.162.2
GLA-GCN†[70]ICCV'2335.339.842.349.256.144.578.662.974.181.978.060.6
MotionBERT‡[77]ICCV'2314.322.826.736.441.728.455.845.731.346.739.075.5
In-Context Model: one task-agnostic architecture for multiple tasks
Copy [5, 16]-134.7133.7133.3132.1131.4133.0351.0171.6196.5195.5196.0316.8
PointMAE [43]ECCV'2219.131.236.654.967.741.9144.396.4167.679.073.3139.9
PIC [16]NeurIPS'2325.040.246.867.279.451.7181.6118.885.796.491.1173.1
Static-SiC (ours)-11.019.022.732.637.824.651.641.829.844.737.259.8
Dynamic-SiC (ours)-11.018.922.632.237.424.451.842.529.544.036.859.2
", + "bbox": [ + 112, + 178, + 854, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the SiC-based on Static-/Dynamic-TUP is referred to as Static-/Dynamic-SiC, respectively. The proposed TUP facilitates the modeling of distribution differences among tasks and datasets, so the model focuses more on learning unified representations shared across tasks.", + "bbox": [ + 75, + 457, + 473, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Model Architecture", + "text_level": 1, + "bbox": [ + 76, + 541, + 264, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we build a skeleton bank, a collection of publicly available datasets including H3.6M, AMASS and 3DPW, covering common skeleton analysis tasks, containing input-target pairs formed as sequences of equal length, where prompts are randomly selected. Given a prompt pair $G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\}$ and a query sample $Q_{j}^{k}$ , we first map them to the C-dimensional feature space, and then splice $I$ and $T$ , $Q$ and $U$ respectively. The former represents our proposed Task-Guided Prompt, and the latter is the objective sequence of our focused analysis. The process can be formulated as:", + "bbox": [ + 75, + 564, + 473, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {G} = \\mathcal {E} \\left(G _ {i} ^ {k}\\right) \\in \\mathbb {R} ^ {2 F \\times J \\times C}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 724, + 470, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {Q} = \\left(\\mathcal {E} \\left(Q _ {j} ^ {k}\\right) \\| U\\right) \\in \\mathbb {R} ^ {2 F \\times J \\times C}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 765, + 468, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}(\\cdot)$ is a linear transformation and $\\|\\cdot\\|$ is concatenating operator. In essence, $U$ represents the prior knowledge of the target skeleton sequence and is also an adapter that adapts to different tasks and different distributions of data for learning.", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Drawing inspirations from [77], we connect spatial attention blocks and temporal attention blocks together in an alternating combination manner to form a two-stream", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "transformer $\\mathcal{M}(\\cdot)$ dedicated to capturing the dynamic characteristics of motion, which is implemented as:", + "bbox": [ + 498, + 457, + 893, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (\\mathrm {X}) = \\alpha^ {i} \\mathcal {T} _ {1} ^ {i} \\left(\\mathcal {S} _ {1} ^ {i} \\left(\\mathrm {X} ^ {i - 1}\\right)\\right) + \\beta^ {i} \\mathcal {S} _ {2} ^ {i} \\left(\\mathcal {T} _ {2} ^ {i} \\left(\\mathrm {X} ^ {i - 1}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 497, + 892, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are balanced parameters determined adaptively. $S_{1}^{i}$ and $\\mathcal{T}_1^i$ are implemented using spatial selfattention mechanism and temporal self-attention mechanism, respectively. Additionally, $i = 1,\\dots,N$ .", + "bbox": [ + 496, + 532, + 893, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, $\\mathcal{M}(\\cdot)$ takes $\\widetilde{G}$ and $\\widetilde{Q}$ as input and parallel iterates over them $n_1$ times respectively. Then, $\\widetilde{G}^{n1}$ and $\\widetilde{Q}^{n1}$ are integrated via aggregation function $\\mathcal{A}(\\cdot)$ :", + "bbox": [ + 496, + 592, + 893, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {n 1} = \\mathcal {A} \\left(\\widetilde {G} ^ {n 1}, \\widetilde {Q} ^ {n 1}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 648, + 892, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we consider $\\mathcal{A}(\\cdot)$ to be a simple weighted summation function. The final output is obtained after iterating $W$ $n_2$ times in $\\mathcal{M}(\\cdot)$ , where $n_1 + n_2 = N$ . Note that we take the second half of the output as the prediction.", + "bbox": [ + 496, + 681, + 893, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training and Inference", + "text_level": 1, + "bbox": [ + 500, + 751, + 714, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We train our model in our proposed skeleton-centric in-context dataset mentioned in Sec. 3. In this training way, our Skeleton-in-Context can perform multiple tasks after one training. Since our output format is skeleton-like, following previous work [77], we adopt the 3D joint-level reconstruction loss:", + "bbox": [ + 496, + 773, + 893, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL o s s = \\sum_ {f = 1} ^ {F} \\sum_ {j = 1} ^ {J} \\left\\| P _ {f, j} - G _ {f, j} \\right\\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 863, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2440", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e88b61c3309694e53b4db3a976de1f3633df3f7143bed116c9f205a0e50dbafc.jpg", + "image_caption": [ + "Figure 4. The comparison of visual results between our Skeleton-in-Context and recent SoTA method MotionBERT [77] on four tasks. Our method generates more accurate poses than MotionBERT, as can be seen from the spots with red arrows $(\\searrow)$ pointed to them." + ], + "image_footnote": [], + "bbox": [ + 120, + 89, + 848, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. After a training phase involving all tasks, we get an in-context model that focuses on dynamic skeleton sequences. It can perform multiple tasks at the same time according to the provided task-dependent prompt without any further parameter updates.", + "bbox": [ + 76, + 688, + 470, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 782, + 209, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement the proposed Skeleton-in-Context model with the number of layers $N = 5$ , number of attention heads $H = 8$ , and hidden feature dimension $C = 256$ . For each prompt input/target and query input/target, the sequence length is $F = 16$ . All tasks are unified into a one-off, end-to-end training process without", + "bbox": [ + 76, + 809, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "any task-specific designs. The training takes about 6 hours for 120 epochs on 4 NVIDIA GeForce RTX 4090 GPUs.", + "bbox": [ + 496, + 685, + 890, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Metrics. H3.6M (Human3.6M) is used for the pose estimation and future pose estimation tasks. AMASS is used for the motion prediction task. 3DPW (3D Pose in the Wild) is used for the joint completion task. All of the tasks can be evaluated using Mean Per Joint Position Error.", + "bbox": [ + 496, + 715, + 890, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Baseline Methods", + "text_level": 1, + "bbox": [ + 500, + 801, + 671, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to make a qualitative comparison with other methods in a relatively fair situation, we reimplement the task-specific SoTA methods in our proposed benchmark. Additionally, we transform recent methods into multi-task models and compare them with our model.", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f956ce6c2b907a14017fa3b0ec1a6addbd8daa1a3247810dc230e948760c4cb5.jpg", + "table_caption": [ + "Table 2. Comparison results between task-specific models and our proposed models on generalization testing experiments on different datasets. We report Mean Per Joint Position Error (mm) for all tasks." + ], + "table_footnote": [], + "table_body": "
MethodsMP. (3DPW)PE.JC. (H3.6M)
40ms80ms120ms160ms200msAvg.AMASS3DPW40%60%Avg.
Task-Specific Model: one architecture for one task
EqMotion [65]15.635.652.167.979.550.151.8156.6193.6194.0193.8
STCFformer [52]18.339.358.975.088.656.045.0141.2177.2178.4177.8
siMLPe [41]14.926.244.056.668.842.148.8150.4236.3257.8247.1
In-Context Model: one task-agnostic architecture for multiple tasks
Static-SiC (ours)7.921.537.754.069.138.122.076.4162.5171.9167.2
Dynamic-SiC (ours)7.121.337.453.568.037.521.776.0150.1151.7151.0
", + "bbox": [ + 101, + 109, + 864, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e4bd665dde05fc600a9f35b9eaf5524175ba4a79c5d194fd2f5df0034412678.jpg", + "image_caption": [ + "Figure 5. Comparison of generalization. Our SiC completes the missing skeletons according to the customized prompt for an unseen task." + ], + "image_footnote": [], + "bbox": [ + 117, + 247, + 851, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Copy Method. Following [5, 16], we copy the prompt target $T_{i}^{k}$ as prediction results for comparison.", + "bbox": [ + 75, + 518, + 468, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Task-Capable Models. For task-specific models [41, 52, 65] and multi-stage models [77], they are not able to handle multiple tasks at the same time. Therefore, we remodel them into end-to-end multi-task-capable models to achieve fair comparison.", + "bbox": [ + 75, + 550, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In-Context Models. Since there are no in-context models excelling in processing multiple human perception tasks, we select the most relative in-context models with our benchmark, Point-In-Context [16] and PointMAE [43], which is proposed in 3D point cloud understanding. Specifically, we take each frame as a token in their framework. For PointMAE, we use the transformed version in [16].", + "bbox": [ + 75, + 626, + 468, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Compared with SoTA Methods", + "text_level": 1, + "bbox": [ + 76, + 741, + 352, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. We report experimental results on four tasks in Tab. 1. We term Skeleton-in-Context implemented with static/dynamic-TUP as Static/Dynamic-SiC, respectively. For other models, we reproduce them and report the results. As can be seen from Tab. 1, our proposed model outperforms all multi-task models, achieving state-of-the-art results on the in-context multi-task benchmark. Notably, the results of task-specific models on a single task reflect the difficulty of each task. However, we also exhibit impressive", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance compared to task-specific models, demonstrating that our proposed model is an expert in handling skeleton sequence-based multi-tasking.", + "bbox": [ + 496, + 518, + 892, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We visualize and compare the results of our SiC and the most recent SoTA model, MotionBERT [77], which is re-structured as an end-to-end multi-task model for fair comparison. As highlighted in Fig. 4, our SiC can generate more accurate poses than MotionBERT according to the provided task-guided prompt.", + "bbox": [ + 496, + 565, + 893, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Generalization Capability.", + "text_level": 1, + "bbox": [ + 500, + 666, + 738, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalize to New Datasets. To verify the generalization ability of our model, we conduct cross-dataset experiments on a specific task. For instance, our proposed model learns motion prediction on the AMASS dataset under the default benchmark, and then it is used to predict future motion sequences on 3DPW during testing. Similarly, we set up four experiments: motion prediction (AMASS $\\rightarrow$ 3DPW), pose estimation (H3.6M $\\rightarrow$ AMASS, 3DPW), and joint completion (3DPW $\\rightarrow$ H3.6M). As shown in Tab. 2, our model is able to generalize the learned task well to other datasets, while other task-specific models cannot accommodate the large bridge introduced across datasets even if they are specifically trained on this single task. Additionally, benefiting from the dynamic pseudo pose, the Dynamic-SiC exhibits", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2442", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/999a9f81ba85f32bac6384abbf49ba50af00f70e73725ae8a46b18a0d63bcbc1.jpg", + "table_caption": [ + "Table 3. Effectiveness of the Skeleton Prompt." + ], + "table_footnote": [], + "table_body": "
#TGPTUPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1XX111.7159.2543.7160.2
2✓X25.052.537.765.2
3X✓26.052.437.878.4
4✓✓24.451.836.862.9
", + "bbox": [ + 86, + 108, + 455, + 175 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f4deef0573e672eb4c942a89b2a43b435ad4d84208285a8c1a7d6f2f95fedb65.jpg", + "table_caption": [ + "Table 4. Different Prompting (TGP)." + ], + "table_footnote": [], + "table_body": "
#TGPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1Feature Sim.25.252.137.764.1
2Motion Sim.25.252.137.864.1
3Random24.451.836.862.9
", + "bbox": [ + 83, + 196, + 460, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more excellent generalization ability than the Static-SiC.", + "bbox": [ + 75, + 258, + 447, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalize to New Task. Furthermore, We construct a prompt that has not appeared in the training set, which performs motion in-between. As Fig. 5 shows, our trained model can complete the missing skeletons according to the customized task-guided prompt. Our SiC generalizes well to a new task, which is unexplored in other works.", + "bbox": [ + 75, + 273, + 468, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 375, + 230, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Skeleton Prompt. We conduct ablation experiments on our proposed skeleton prompts. As shown in Tab. 3, the skeleton prompts solve the problem of directly extending previous work to skeleton sequences, which is falling into over-fitting. Meanwhile, applying both TGP and TUP to the model can achieve the best results. Our proposed skeleton prompts replace the random masking training strategy used in previous work. In particular, TUP creates a training environment that does not require the query target, which ensures that the model will not see the ground truth of the query input in advance.", + "bbox": [ + 75, + 398, + 468, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Engineering. Since previous work has proven that different prompts in context learning affect the performance of the model, we conduct ablation experiments on the choice of prompts. As Tab. 4 shows, we select the prompt by the feature similarity and MPJPE minimum between the input skeleton sequence and the prompt, respectively. Experimental results show that the random selection method achieves the best results. Unlike 2D images and 3D point clouds, skeleton-based task data is highly similar, resulting in different prompts playing similar roles in the same task. While the random selection method expands the model's receptive field to the data and achieves better results.", + "bbox": [ + 75, + 565, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture. We conduct ablation experiments on the architecture of the model. As shown in Tab. 5, after exploring different hidden feature dimensions and transformer depths, we find that the model performs best when the feature dimension is 256, and the depth is 5.", + "bbox": [ + 75, + 746, + 468, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. In-Context Initial Pose", + "text_level": 1, + "bbox": [ + 76, + 830, + 287, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an in-depth exploration of Dynamic-TUP. We extract the trained Dynamic-TUP weights, then randomly initialize a $(F,J,3)$ tensor and use the trained encoder to in", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ae595ac376388337924cb89960655ffcfe911d91a4df2754e246bbdd29df0e6.jpg", + "table_caption": [ + "Table 5. Ablation on model architecture." + ], + "table_footnote": [], + "table_body": "
Arch.Num.MP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
Dim.6429.357.145.880.5
12826.453.839.266.1
25624.451.836.862.9
51224.750.737.663.0
Dep.325.153.739.265.6
425.152.238.664.7
524.451.836.862.9
624.752.938.163.7
(a) Task-Unified Prompt (Static)
(b) Task-Unified Prompt (Dynamic)
", + "bbox": [ + 506, + 108, + 888, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6. (a) The visualization of Static-TUP. (b) The concrete visual skeleton sequence of Dynamic-TUP.", + "bbox": [ + 498, + 340, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "crease its feature dimension to C-dimension. Finally, an optimizer is used to optimize the tensor and minimize its feature similarity with the trained Dynamic-TUP. The top of Fig. 6 represents the visualization of Static-TUP, while the bottom is the Dynamic-TUP. The concrete skeleton of dynamic-TUP is similar to the mean pose Static-TUP of the training set, which demonstrates that the Dynamic-TUP learned by our model is actually the prior knowledge whose receptive field is all ground truth in the training set. This prior knowledge helps our model learn more smoothly on multiple tasks, and adding learnable modules helps our model generalize to other datasets or new tasks (Sec. 5.3).", + "bbox": [ + 496, + 377, + 893, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 573, + 617, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose the Skeleton-in-Context, designed to process multiple skeleton-base tasks simultaneously after just one training time. Specifically, we build a novel skeleton-based in-context benchmark covering various tasks. In particular, we propose skeleton prompts composed of TGP and TUP, which solve the overfitting problem of skeleton sequence data trained under the training framework commonly applied in previous 2D and 3D in-context models. Besides, we demonstrate that our model can generalize to different datasets and new tasks, such as motion completion. We hope our research builds the first step in the exploration of in-context learning for skeleton-based sequences, which paves the way for further research in this area.", + "bbox": [ + 496, + 598, + 893, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work is supported by National Natural Science Foundation of China (No. 62203476), Natural Science Foundation of Shenzhen (No. JCYJ20230807120801002). This project is also supported by the interdisciplinary doctoral grants (iDoc 2021-360) from the Personalized Health and Related Technologies (PHRT) of the ETH domain, Switzerland.", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 2", + "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models, 2023. 3", + "[3] Ivana Balažević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Hénaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 2, 3", + "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2", + "[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 2022. 2, 3, 4, 5, 7", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 1, 2", + "[7] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In ECCV, 2018. 2", + "[8] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 2", + "[9] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In ICCV, 2021. 1, 3", + "[10] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, 2020. 2", + "[11] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In ICCV, 2019. 2", + "[12] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. Unihcp: A unified model for human-centric perceptions. In CVPR, 2023. 2, 3", + "[13] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In CVPR, 2020. 2", + "[14] Zhichao Deng, Xiangtai Li, Xia Li, Yunhai Tong, Shen Zhao, and Mengyuan Liu. Vg4d: Vision-language model goes 4d video recognition. In ICRA, 2024. 2", + "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[16] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning for 3d point cloud understanding. arXiv preprint arXiv:2306.08659, 2023. 1, 2, 3, 4, 5, 7", + "[17] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In CVPR, 2023. 1, 2, 3", + "[18] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 2, 3", + "[19] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 5", + "[20] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. TOG, 2020. 2", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3, 4", + "[22] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. TOG, 2020. 2", + "[23] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In ICCV, 2019. 2", + "[24] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. PAMI, 2013. 3, 5", + "[25] Mahdi M Kalayeh, Emrah Basaran, Muhittin Gokmen, Mustafa E Kamasak, and Mubarak Shah. Human semantic parsing for person re-identification. In CVPR, 2018. 3", + "[26] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 3DV, 2020. 2", + "[27] Sameh Khamis, Cheng-Hao Kuo, Vivek K Singh, Vinay D Shet, and Larry S Davis. Joint learning for attribute-consistent person re-identification. In ECCV, 2015. 3", + "[28] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In CVPR), 2018. 2", + "[29] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Guangliang Cheng, Pang Jiangmiao, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv pre-print, 2023. 2", + "[30] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into person: Joint body parsing & pose estimation network and a new benchmark. PAMI, 2018. 3", + "[31] Chunze Lin, Jiwen Lu, and Jie Zhou. Multi-grained deep feature learning for pedestrian detection. In ICME, 2018. 3", + "[32] Jun Liu, Henghui Ding, Amir Shahroudy, Ling-Yu Duan, Xudong Jiang, Gang Wang, and Alex C Kot. Feature boosting network for 3d pose estimation. PAMI, 2019. 2, 3", + "[33] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. PAMI, 2019. 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2444", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In ECCV, 2020. 2, 3", + "[35] Diogo C Luvizon, David Picard, and Hedi Tabia. 2d/3d pose estimation and action recognition using multitask deep learning. In CVPR, 2018. 3", + "[36] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 1, 3", + "[37] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3", + "[38] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2, 3", + "[39] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2", + "[40] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2", + "[41] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 5, 7", + "[42] Xuecheng Nie, Jiashi Feng, and Shuicheng Yan. Mutual learning to adapt for joint human parsing and pose estimation. In ECCV, 2018. 3", + "[43] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3, 4, 5, 7", + "[44] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3", + "[46] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633, 2021. 2", + "[47] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 2021. 2", + "[48] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu $\\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 3", + "[49] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In ECCV, 2022. 2", + "[50] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry Steven Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for multi-camera person re-identification. PAMI, 2017. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023. 2, 3", + "[52] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatio-temporal criss-cross attention. In CVPR, 2023. 5, 7", + "[53] Yonglong Tian, Ping Luo, Xiaogang Wang, and Xiaou Tang. Pedestrian detection aided by deep learning semantic tasks. In CVPR, 2015. 3", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3", + "[55] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3", + "[56] Chaoyang Wang, Xiangtai Li, Henghui Ding, Lu Qi, Jiangning Zhang, Yunhai Tong, Chen Change Loy, and Shuicheng Yan. Explore in-context segmentation via latent diffusion models. arXiv, 2024. 2", + "[57] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 2", + "[58] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4", + "[59] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, 2023. 2, 3", + "[60] Xinshun Wang, Qiongjie Cui, Chen Chen, and Mengyuan Liu. Gcnext: Towards the unity of graph convolutions for human motion prediction. In AAAI, 2024. 2", + "[61] Xinshun Wang, Wanying Zhang, Can Wang, Yuan Gao, and Mengyuan Liu. Dynamic dense graph convolutional network for skeleton-based human motion prediction. TIP, 2024. 2", + "[62] Zhendong Wang, Yifan Jiang, Yadong Lu, Yelong Shen, Pengcheng He, Weizhu Chen, Zhangyang Wang, and Mingyuan Zhou. In-context learning unlocked for diffusion models. arXiv preprint arXiv:2305.01115, 2023. 2", + "[63] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 2", + "[64] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023. 2, 3", + "[65] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 5, 7", + "[66] Shilin Xu, Xiangtai Li, Jingbo Wang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Fashionformer: A simple, effective and unified baseline for human fashion segmentation and recognition. In ECCV, 2022. 3", + "[67] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In CVPR, 2021. 2" + ], + "bbox": [ + 501, + 90, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2445", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In ICCV, 2019. 2", + "[69] Angela Yao, Juergen Gall, and Luc Van Gool. Coupled action recognition and pose estimation from multiple views. IJCV, 2012. 3", + "[70] Bruce XB Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In ICCV, 2023. 5", + "[71] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2", + "[72] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670, 2023. 2, 3", + "[73] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In CVPR, 2022. 2", + "[74] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021.", + "[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 2", + "[76] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, et al. Generative tweening: Long-term inbetweening of 3d human motions. arXiv preprint arXiv:2005.08891, 2020.2", + "[77] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In ICCV, 2023. 1, 3, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 559 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2446", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_model.json b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d1ecd8c3c76602fdf87d559fb05616cd8f7ba25 --- /dev/null +++ b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.804, + 0.176 + ], + "angle": 0, + "content": "Skeleton-in-Context: Unified Skeleton Sequence Modeling with In-Context Learning" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.189, + 0.854, + 0.226 + ], + "angle": 0, + "content": "Xinshun Wang\\(^{1,2,*}\\), Zhongbin Fang\\(^{1,*}\\), Xia Li\\(^{3}\\), Xiangtai Li\\(^{4}\\), Chen Chen\\(^{5}\\), Mengyuan Liu\\(^{2, \\boxtimes}\\), Sun Yat-sen University" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.228, + 0.8, + 0.244 + ], + "angle": 0, + "content": "\\(^{2}\\)National Key Laboratory of General Artificial Intelligence, Peking University, Shenzhen Graduate School" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.246, + 0.758, + 0.262 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Computer Science, ETH Zurich \\(^{4}\\)S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.264, + 0.697, + 0.279 + ], + "angle": 0, + "content": "\\(^{5}\\)Center for Research in Computer Vision, University of Central Florida" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.282, + 0.842, + 0.298 + ], + "angle": 0, + "content": "{wangxsh36,fangzhb5}@mail2.sysu.edu.cn，xia.li@inf.ethz.ch，xiangtai94@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.301, + 0.686, + 0.314 + ], + "angle": 0, + "content": "chen.chen@crcv.ucf.edu, liumengyuan@pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.32, + 0.696, + 0.334 + ], + "angle": 0, + "content": "https://github.com/fanglaosi/Skeleton-in-Context" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.36, + 0.314, + 0.375 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.474, + 0.802 + ], + "angle": 0, + "content": "In-context learning provides a new perspective for multi-task modeling for vision and NLP. Under this setting, the model can perceive tasks from prompts and accomplish them without any extra task-specific head predictions or model fine-tuning. However, skeleton sequence modeling via in-context learning remains unexplored. Directly applying existing in-context models from other areas onto skeleton sequences fails due to the similarity between inter-frame and cross-task poses, which makes it exceptionally hard to perceive the task correctly from a subtle context. To address this challenge, we propose Skeleton-in-Context (SiC), an effective framework for in-context skeleton sequence modeling. Our SiC is able to handle multiple skeleton-based tasks simultaneously after a single training process and accomplish each task from context according to the given prompt. It can further generalize to new, unseen tasks according to customized prompts. To facilitate context perception, we additionally propose a task-unified prompt, which adaptively learns tasks of different natures, such as partial joint-level generation, sequence-level prediction, or 2D-to-3D motion prediction. We conduct extensive experiments to evaluate the effectiveness of our SiC on multiple tasks, including motion prediction, pose estimation, joint completion, and future pose estimation. We also evaluate its generalization capability on unseen tasks such as motion-in-between. These experiments show that our model achieves state-of-the-art multi-task performance and even outperforms single-task methods on certain tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.807, + 0.21, + 0.822 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.471, + 0.879 + ], + "angle": 0, + "content": "The idea of multi-task models has attracted much research interest over the years in a wide range of areas [6, 16, 17, 58], such as computer vision and natural language processing" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.359, + 0.885, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.447, + 0.887, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.541, + 0.885, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.7, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Figure 1. In-context learning in a) image modeling [58], b) point cloud modeling [16], and c) skeleton sequence modeling (ours)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "(NLP). To build such models, most previous works focus on deploying task-specific heads on a shared backbone and learning them through pretraining and fine-tuning. Models with such task-specific architectures are not learned in an end-to-end manner and are hard to generalize to new, unseen tasks and data. Most existing methods for skeleton-based human-centric tasks rely on task-specific designs and/or non-end-to-end frameworks to accomplish multiple tasks. Several works [9, 36] learn unified human motion representations. In particular, MotionBERT [77] adopts two-stage training, including pre-training on the 2D-3D lifting task" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.444, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. ☑ Corresponding author is Mengyuan Liu." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2436" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.531 + ], + "angle": 0, + "content": "and fine-tuning on each task. UniHCP [12] connects a well-designed shared backbone network to specific task heads according to different tasks. Meanwhile, UPS [17] unifies action labels and motion sequences into text-based language sequences and fine-tunes them on each task for better performance. These models can have a limited scope of tasks they are capable of handling and lack flexibility. This limits their capability to generalize to unseen tasks and data. To overcome these challenges, in-context learning represents a new trend to build generalist models, which was originally proposed in NLP [6, 45, 46] and then introduced into other areas such as images and point clouds. Some generalist models [5, 58, 59, 72] utilize the Masked Image Modeling (MIM) framework to exhibit multi-task capability via in-context learning in image processing, as shown in Fig. 1(a). Similarly, in Fig. 1(b), Point-In-Context [16] proposes a key joint sampling module and introduces in-context learning into point cloud processing. Specifically, in-context models learn patterns of provided input-output pairs, called in-context examples or prompts, and then perform the desired task according to the prompt. An in-context model can accomplish diverse tasks end-to-end without any task-specific heads or fine-tuning. It also has excellent generalization capability. Directly adapting existing in-context models from other areas onto skeleton sequences fails because they largely focus on static objects (images or point clouds) and thus cannot capture spatial-temporal dependencies in skeleton sequences. To the best of our knowledge, in-context learning remains unexplored in 3D skeleton sequence modeling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.534, + 0.473, + 0.882 + ], + "angle": 0, + "content": "In this paper, we propose Skeleton-in-Context (SiC), a novel framework for skeleton sequence modeling, which is able to process multiple tasks simultaneously after a single training process without any task-specific designs, as Fig. 1(c) shows. Specifically, SiC perceives tasks from context, i.e., the prompt from a desired task provided as additional input, and then accomplishes the task for the query. SiC employs two types of prompts to facilitate context perception: the task-guided prompt (TGP) and the task-unified prompt (TUP). TGP provides a task template for the model, giving the model the ability to make analogies. This is the key to the model's capability to generalize to new tasks, as we can guide the model to do unseen tasks by customizing specific prompts. TUP can be implemented in two ways: static and dynamic. The static TUP leverages prior knowledge shared among tasks, while the dynamic TUP adaptively learns to perceive task context. Skeleton-in-Context provides a new perspective for training a skeleton-based multi-task in-context model. Since there is no existing benchmark for measuring such a model, we establish a new in-context benchmark based on four tasks, including three skeleton-based tasks: motion prediction, pose estimation, and joint completion, and a new task: future pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Our main contributions are as follows: 1) We introduce" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "a novel framework, Skeleton-in-Context (SiC), designed to handle multiple skeleton-based tasks simultaneously. Our approach learns the potential pattern of the given prompt and performs it on the query sample. 2) We establish a new 3D human in-context learning benchmark comprised of multiple skeleton-based tasks, such as motion prediction, pose estimation, joint completion, and future motion estimation. We further benchmark several representative in-context learning baselines for reference. 3) We conduct extensive experiments to evaluate our SiC. SiC achieves state-of-the-art results in multi-tasking and even outperforms task-specific models on certain tasks. Moreover, SiC also generalizes well to unseen tasks such as motion in-between." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.302, + 0.642, + 0.317 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.897, + 0.674 + ], + "angle": 0, + "content": "Skeleton-Based Motion Analysis. Many of the skeleton-based human-centric tasks have shared similar development patterns in terms of the methods used to address them. On human motion prediction, early successes are achieved with RNNs and CNNs [28, 40]. Nowadays, the task is dominated by Graph Convolutional Networks (GCNs). LTD [38] considers each joint coordinate as a node in a graph on which to employ learnable spatial graph convolutions. Follow-up works [13, 39, 60, 61] focus on designing better graph convolution networks. Different from motion prediction, the goal of motion synthesis is to complete the human motion sequence. To complete the missing parts of human motion, many works adopt convolutional models [22, 26, 68] or generative adversarial networks [7, 20, 23, 76]. One category of 3D pose estimation [18, 32, 34] is to lift the estimated 2D pose to 3D, which often employs spatial-temporal convolution [8, 10, 11, 41, 44, 57, 67] or transformer-based [14, 49, 71, 73-75] methods. However, these approaches focus on designing customized architectures specific to a task, which are limited in range of applications. To this end, we propose to unify the input and output formats of multiple tasks and utilize a single model to process multiple tasks simultaneously via in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.897, + 0.903 + ], + "angle": 0, + "content": "In-Context Learning. The emergence of in-context learning provides a new paradigm for achieving multitasking, which originated from Natural Language Processing (NLP) [1, 15, 45-47]. It endows models with the capability to handle multiple tasks and even unseen tasks by incorporating domain-specific input-target pairs, known as prompts, into the query sample [6]. Following NLP, in-context learning is extended to computer vision in some work, such as 2D images [51, 56, 58, 59, 62, 72] and 3D point clouds [3, 16, 64]. Specifically, Visual Prompt [5], Painter [58], and Point-In-Context [16] are pioneers of in-context methods in computer vision. They verify the effectiveness and flexibility of in-context learning in vision-based tasks, and most of them adopt pure-transformer-based approaches [4, 21, 29, 63]. Meanwhile, several approaches" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2437" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "deeply explore Prompt Engineering [3, 51, 72] and awe-inspiring generalization [16, 59, 64] in in-context learning. Recently, LVM [2] unifies most vision tasks and data via joint co-training on one visual in-context model. Compared to the works focused on static objects (images and point clouds), our work mainly delves into modeling in-context dynamic skeleton sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.474, + 0.471 + ], + "angle": 0, + "content": "Unified Skeleton-Centric Models. A generalist model that can excel in handling multiple tasks is a profound step toward general artificial intelligence [45, 58]. Previous works [25, 27, 31, 35, 50, 53, 66, 69] give the effort to explore co-training several human-centric tasks. Meanwhile, some works [30, 42] show great progress in simultaneously predicting parsing and pose. Recent research [9, 36] demonstrates the effectiveness of modeling unified human motion sequences. In particular, UniHCP [12] designs specific task heads according to different 2D human-centric tasks. Motion-BERT [77] is pre-trained in the 2D-3D lifting task and fine-tuned on each task. UPS [17] unifies heterogeneous human behavior understanding tasks via language sequences and cycle training for each task in order. Nevertheless, these works are limited by the dilemma of complex training stages or task heads, which introduce extra costs. We integrate the output format of different tasks in skeleton sequence format and train the model just once." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.439, + 0.5 + ], + "angle": 0, + "content": "3. In-Context Skeleton Sequence Modeling" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.507, + 0.212, + 0.521 + ], + "angle": 0, + "content": "3.1. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.698 + ], + "angle": 0, + "content": "Since the skeleton sequence data in skeleton-based tasks are sequential by nature, we treat the skeleton at each frame as a token, which can be processed by transformer [54]. Following previous works [5, 16, 58], we curate an in-context skeleton sequence paradigm. Given an input-target skeleton sequence pair as a task-guided prompt \\( G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\} \\) and a query input \\( Q_{j}^{k} \\), where \\( I_{i}^{k}, T_{i}^{k}, Q_{j}^{k} \\in \\mathbb{R}^{F \\times J \\times 3} \\) each have \\( F \\) frames and \\( J \\) joints, the model is expected to perceive the context from the prompt, i.e., which task it represents, and perform the corresponding operations on the query in analogy to the prompt. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.706, + 0.471, + 0.726 + ], + "angle": 0, + "content": "\\[\ny = f \\left(G _ {i} ^ {k}, Q _ {j} ^ {k}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.471, + 0.795 + ], + "angle": 0, + "content": "where \\( k \\) represents the task index, while \\( i \\) and \\( j \\) indicate different example indexes. Thus, we can determine what task will be performed on the query sample by selecting the task-specific prompt from the corresponding training set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.238, + 0.817 + ], + "angle": 0, + "content": "3.2. Task Definitions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To measure multitasking capability, we establish a new in-context benchmark from common skeleton-based datasets, including H3.6M [24], AMASS [37], and 3DPW [55]. In this benchmark, we select three common tasks: motion prediction, pose estimation, and joint completion, and a new" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.891, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.26, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Top: Training with MIM-style framework [21, 43] used in previous works [5, 16, 58]. The model is able to reconstruct the masked frames well. Bottom: During inference, the reconstructed sequence gradually shrinks the skeleton as time goes by because the model only learns frame interpolation during training. Once the model loses subsequent reference frames when interpolating frames, the generated sequence will tend to shrink." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.895, + 0.424 + ], + "angle": 0, + "content": "task: future pose estimation. Consequently, we conduct extensive experiments on a large and diverse dataset collection, which consists of 176K skeleton sequence samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.894, + 0.505 + ], + "angle": 0, + "content": "Motion Prediction. Motion prediction [38] is to predict future motion based on historical motion. There is continuity between the input and output sequences, which constitute a complete motion. Thus, the input \\( I \\) and target \\( T \\) are complete skeleton sequences with shapes of \\( (F, J, 3) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.894, + 0.633 + ], + "angle": 0, + "content": "Pose Estimation. Pose estimation [18, 32, 34] is to lift 2D poses to 3D. The model takes the detected 2D skeleton sequence \\( I \\in \\mathbb{R}^{F \\times J \\times 2} \\), which contain noise, as input, and estimates the accurate 3D skeleton sequence \\( T \\in \\mathbb{R}^{F \\times J \\times 3} \\). Note that we expand the 2D input to 3D by filling the 3rd dimension with 0, in order to unify the data format with other tasks. Therefore, the input skeleton sequence becomes the same shape \\( (F, J, 3) \\) as the target sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.895, + 0.775 + ], + "angle": 0, + "content": "Future Pose Estimation. Based on prediction and estimation, we propose a new task: future pose estimation, which is to predict future 3D motion based on history 2D poses. As 2D images are easier to obtain than 3D data, this task unlocks a host of new real-world applications. Given a historical 2D skeleton sequence \\( I \\in \\mathbb{R}^{F \\times J \\times 3} \\) (one dimension with 0), future pose estimation expects the model to output its future 3D skeleton sequence \\( T \\in \\mathbb{R}^{F \\times J \\times 3} \\). To our knowledge, no work has ever proposed future pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Joint Completion. Joint completion is to recover skeleton sequences that have missing joints. This task can reflect cases where joints may be missing or displaced when collected with Lidar cameras [33, 48]. We set two levels to verify the ability to complete joints, which are \\(40\\%\\) and \\(60\\%\\) of the joints missing in each frame. Also, the coordinates of missing joints are replaced with 0, with the goal of unifying the input skeleton sequence format with \\((F, J, 3)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2438" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.087, + 0.873, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.896, + 0.373 + ], + "angle": 0, + "content": "Figure 3. Overall framework of our Skeleton-in-Context. Specifically, we establish a skeleton bank by integrating training sets under different tasks, which contain a large amount of input-target pairs performing different tasks. Next, we randomly select a sample pair as the task-guided prompt (TGP) and a query input from the skeleton bank, undergo encoding and concatenating, respectively, and then input them into the transformer in parallel. In particular, during this process, the query input and task-unified prompt (TUP) are combined to form a new query. After iterating \\( n_1 \\) times, the TGP and query are aggregated through \\( \\mathcal{A}(\\cdot) \\) and then input into the transformer for \\( n_2 \\) iterations. Lastly, the second half of the model output is used as our prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.389, + 0.273, + 0.405 + ], + "angle": 0, + "content": "4. Skeleton-in-Context" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.471, + 0.552 + ], + "angle": 0, + "content": "In-context learning on other data modalities such as images and point clouds [5, 16, 58] largely employ Mask Image Modeling (MIM) [21] to reconstruct the randomly masked elements. During inference, only the target elements are masked. On skeleton sequences, MIM fails due to spatial-temporal complexity and inter-pose similarity, which makes it outstandingly hard to perceive the task correctly from a subtle context. In our Skeleton-in Context, we employ a new framework to address this challenge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.32, + 0.577 + ], + "angle": 0, + "content": "4.1. Skeleton Sequence Prompt" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Task-Guided Prompt. In in-context learning, a prompt is needed for the model to learn the underlying task pattern from context. As mentioned in Sec. 3.1, we propose the Task-Guided Prompt (TGP) \\( G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\} \\), which comprises of an example skeleton pair performing a specific task \\( k \\). In this way, our proposed model is able to perform a variety of tasks according to the provided prompts that could be seen or even unseen in the training set. It is worth noting that our model can perform unseen tasks and transfer to datasets that are different styles from the training dataset. It will be elaborated in Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Avoiding Over-Fitting. Previous works have made great progress in unifying different tasks via the MIM-style framework [5, 16, 58]. Compared with static objects in 2D and 3D, such as images and point clouds, the obstacle to in-context learning of skeleton sequences lies in the wondrous similarity of each frame, so directly using the random masking strategy [5, 16, 58] will cause the model to fall into overfitting, leading it to only learn simple interpolation frames rather than capturing meaningful motion sequence representations during the training stage. As shown in Fig. 2, when" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.526 + ], + "angle": 0, + "content": "we directly utilize the MPM-style [21, 43] framework to train our model, although it can perfectly reconstruct the masked frame during training, the experimental results reveal that it outputs a devastating result during inference. It is obvious that the output skeleton sequence gradually shrinks as the length increases because the next frame of the current frame is 0, and interpolating with 0 will lead to shrinking. Such a phenomenon underscores the challenges of learning contextual information about dynamic skeleton sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.895, + 0.721 + ], + "angle": 0, + "content": "Task-Unified Prompt. With the Task-Guided Prompt, which exhibits unique capabilities specific to in-domain or out-of-domain tasks, mitigating the gap between different tasks presents a new challenge. Meanwhile, as discussed above, a straight extension of previous works will lead to a simple model that can only perform the interpolation task based on the given skeleton sequence. To tackle these challenges, we introduce the Task-Unified Prompt (TUP), a well-designed and task-agnostic module, which can adaptively learn to incorporate multiple tasks into the framework of in-context learning, i.e., a one-off end-to-end framework for all tasks without having to deploy task-specific heads or fine-tuning. We introduce two versions of TUP:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.723, + 0.84, + 0.738 + ], + "angle": 0, + "content": "Version 1: Static-TUP with static pseudo pose:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.748, + 0.892, + 0.79 + ], + "angle": 0, + "content": "\\[\nU ^ {S} = \\mathcal {E} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\sum_ {k = 1} ^ {K} \\left(T _ {m} ^ {k}\\right) \\in \\mathbb {R} ^ {F \\times J \\times 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.891, + 0.831 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}(\\cdot)\\) is a embedding layer. \\(M\\) and \\(K\\) indicate the total number of training samples and tasks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.832, + 0.882, + 0.846 + ], + "angle": 0, + "content": "Version 2: Dynamic-TUP with dynamic pseudo pose:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.857, + 0.892, + 0.874 + ], + "angle": 0, + "content": "\\[\nU ^ {D} = W _ {1} W _ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.884, + 0.764, + 0.901 + ], + "angle": 0, + "content": "where \\(W_{1}\\in \\mathbb{R}^{F\\times J\\times 1}\\) and \\(W_{2}\\in \\mathbb{R}^{1\\times 1\\times C}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.174 + ], + "angle": 0, + "content": "Table 1. Comparison of task-specific models, multi-task models, and in-context models on four human skeleton tasks. For Motion Prediction (MP.), Joint Completion (JC.), and Future Pose Estimation (FPE.), we report Mean Per Joint Position Error (mm) [24]. For Pose Estimation (PE.), we additionally report another indicator, Normalized Mean Per Joint Position Error (N-MPJPE) [77]. For joint completion, we report two settings: \\(40\\%\\) and \\(60\\%\\) of the joints missing in each frame. \\(\\dagger\\) represents the multi-task-capable models re-structured from task-specific models. \\(\\ddagger\\) represents the multi-task-capable models re-structured from multi-stage models. Static-SiC and Dynamic-SiC represent the static and dynamic versions of TUP, respectively." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.179, + 0.855, + 0.452 + ], + "angle": 0, + "content": "
MethodsVenuesMP. (AMASS)PE. (H3.6M)JC. (3DPW)FPE. (H3.6M)
80ms160ms200ms320ms400msAvg.MPJPE↓ Avg.N-MPJPE↓ Avg.MPJPE↓ 40%60% Avg.MPJPE↓ 200ms300ms Avg.
Task-Specific Model: one architecture for one task
siMLPe [19]WACV'2312.121.625.434.439.926.758.949.748.458.553.575.4
EqMotion [65]CVPR'2314.426.631.744.050.033.363.259.340.449.745.169.5
STCFormaler [52]CVPR'2319.525.128.235.841.528.754.645.231.246.738.960.9
GLA-GCN [70]ICCV'2319.525.128.235.342.030.049.640.462.272.567.449.5
MotionBERT [77]ICCV'2312.921.425.536.342.326.749.441.129.344.336.875.5
Multi-Task Model: one architecture (w. multiple task-specific heads) for multiple tasks
siMLPe†[41]WACV'2319.030.035.346.852.436.774.660.059.067.463.264.4
EqMotion†[65]CVPR'2314.025.330.240.546.631.3163.3110.138.956.047.588.1
STCFormaler†[52]CVPR'2317.428.031.543.040.433.955.645.633.650.642.162.2
GLA-GCN†[70]ICCV'2335.339.842.349.256.144.578.662.974.181.978.060.6
MotionBERT‡[77]ICCV'2314.322.826.736.441.728.455.845.731.346.739.075.5
In-Context Model: one task-agnostic architecture for multiple tasks
Copy [5, 16]-134.7133.7133.3132.1131.4133.0351.0171.6196.5195.5196.0316.8
PointMAE [43]ECCV'2219.131.236.654.967.741.9144.396.4167.679.073.3139.9
PIC [16]NeurIPS'2325.040.246.867.279.451.7181.6118.885.796.491.1173.1
Static-SiC (ours)-11.019.022.732.637.824.651.641.829.844.737.259.8
Dynamic-SiC (ours)-11.018.922.632.237.424.451.842.529.544.036.859.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.475, + 0.534 + ], + "angle": 0, + "content": "We denote the SiC-based on Static-/Dynamic-TUP is referred to as Static-/Dynamic-SiC, respectively. The proposed TUP facilitates the modeling of distribution differences among tasks and datasets, so the model focuses more on learning unified representations shared across tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.266, + 0.558 + ], + "angle": 0, + "content": "4.2. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.474, + 0.718 + ], + "angle": 0, + "content": "As shown in Fig. 3, we build a skeleton bank, a collection of publicly available datasets including H3.6M, AMASS and 3DPW, covering common skeleton analysis tasks, containing input-target pairs formed as sequences of equal length, where prompts are randomly selected. Given a prompt pair \\( G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\} \\) and a query sample \\( Q_{j}^{k} \\), we first map them to the C-dimensional feature space, and then splice \\( I \\) and \\( T \\), \\( Q \\) and \\( U \\) respectively. The former represents our proposed Task-Guided Prompt, and the latter is the objective sequence of our focused analysis. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.726, + 0.471, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\widetilde {G} = \\mathcal {E} \\left(G _ {i} ^ {k}\\right) \\in \\mathbb {R} ^ {2 F \\times J \\times C}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.766, + 0.47, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\widetilde {Q} = \\left(\\mathcal {E} \\left(Q _ {j} ^ {k}\\right) \\| U\\right) \\in \\mathbb {R} ^ {2 F \\times J \\times C}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}(\\cdot)\\) is a linear transformation and \\(\\|\\cdot\\|\\) is concatenating operator. In essence, \\(U\\) represents the prior knowledge of the target skeleton sequence and is also an adapter that adapts to different tasks and different distributions of data for learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Drawing inspirations from [77], we connect spatial attention blocks and temporal attention blocks together in an alternating combination manner to form a two-stream" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.458, + 0.895, + 0.489 + ], + "angle": 0, + "content": "transformer \\(\\mathcal{M}(\\cdot)\\) dedicated to capturing the dynamic characteristics of motion, which is implemented as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.498, + 0.893, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (\\mathrm {X}) = \\alpha^ {i} \\mathcal {T} _ {1} ^ {i} \\left(\\mathcal {S} _ {1} ^ {i} \\left(\\mathrm {X} ^ {i - 1}\\right)\\right) + \\beta^ {i} \\mathcal {S} _ {2} ^ {i} \\left(\\mathcal {T} _ {2} ^ {i} \\left(\\mathrm {X} ^ {i - 1}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.593 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are balanced parameters determined adaptively. \\(S_{1}^{i}\\) and \\(\\mathcal{T}_1^i\\) are implemented using spatial selfattention mechanism and temporal self-attention mechanism, respectively. Additionally, \\(i = 1,\\dots,N\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.895, + 0.64 + ], + "angle": 0, + "content": "Then, \\(\\mathcal{M}(\\cdot)\\) takes \\(\\widetilde{G}\\) and \\(\\widetilde{Q}\\) as input and parallel iterates over them \\(n_1\\) times respectively. Then, \\(\\widetilde{G}^{n1}\\) and \\(\\widetilde{Q}^{n1}\\) are integrated via aggregation function \\(\\mathcal{A}(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.649, + 0.893, + 0.673 + ], + "angle": 0, + "content": "\\[\nW ^ {n 1} = \\mathcal {A} \\left(\\widetilde {G} ^ {n 1}, \\widetilde {Q} ^ {n 1}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.895, + 0.745 + ], + "angle": 0, + "content": "we consider \\(\\mathcal{A}(\\cdot)\\) to be a simple weighted summation function. The final output is obtained after iterating \\(W\\) \\(n_2\\) times in \\(\\mathcal{M}(\\cdot)\\), where \\(n_1 + n_2 = N\\). Note that we take the second half of the output as the prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.715, + 0.769 + ], + "angle": 0, + "content": "4.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.895, + 0.864 + ], + "angle": 0, + "content": "Training. We train our model in our proposed skeleton-centric in-context dataset mentioned in Sec. 3. In this training way, our Skeleton-in-Context can perform multiple tasks after one training. Since our output format is skeleton-like, following previous work [77], we adopt the 3D joint-level reconstruction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.864, + 0.893, + 0.904 + ], + "angle": 0, + "content": "\\[\nL o s s = \\sum_ {f = 1} ^ {F} \\sum_ {j = 1} ^ {J} \\left\\| P _ {f, j} - G _ {f, j} \\right\\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2440" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.09, + 0.849, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.649, + 0.895, + 0.679 + ], + "angle": 0, + "content": "Figure 4. The comparison of visual results between our Skeleton-in-Context and recent SoTA method MotionBERT [77] on four tasks. Our method generates more accurate poses than MotionBERT, as can be seen from the spots with red arrows \\((\\searrow)\\) pointed to them." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.471, + 0.766 + ], + "angle": 0, + "content": "Inference. After a training phase involving all tasks, we get an in-context model that focuses on dynamic skeleton sequences. It can perform multiple tasks at the same time according to the provided task-dependent prompt without any further parameter updates." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.784, + 0.21, + 0.801 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We implement the proposed Skeleton-in-Context model with the number of layers \\( N = 5 \\), number of attention heads \\( H = 8 \\), and hidden feature dimension \\( C = 256 \\). For each prompt input/target and query input/target, the sequence length is \\( F = 16 \\). All tasks are unified into a one-off, end-to-end training process without" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.715 + ], + "angle": 0, + "content": "any task-specific designs. The training takes about 6 hours for 120 epochs on 4 NVIDIA GeForce RTX 4090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Datasets and Metrics. H3.6M (Human3.6M) is used for the pose estimation and future pose estimation tasks. AMASS is used for the motion prediction task. 3DPW (3D Pose in the Wild) is used for the joint completion task. All of the tasks can be evaluated using Mean Per Joint Position Error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.673, + 0.817 + ], + "angle": 0, + "content": "5.1. Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In order to make a qualitative comparison with other methods in a relatively fair situation, we reimplement the task-specific SoTA methods in our proposed benchmark. Additionally, we transform recent methods into multi-task models and compare them with our model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2441" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.079, + 0.892, + 0.107 + ], + "angle": 0, + "content": "Table 2. Comparison results between task-specific models and our proposed models on generalization testing experiments on different datasets. We report Mean Per Joint Position Error (mm) for all tasks." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.111, + 0.865, + 0.237 + ], + "angle": 0, + "content": "
MethodsMP. (3DPW)PE.JC. (H3.6M)
40ms80ms120ms160ms200msAvg.AMASS3DPW40%60%Avg.
Task-Specific Model: one architecture for one task
EqMotion [65]15.635.652.167.979.550.151.8156.6193.6194.0193.8
STCFformer [52]18.339.358.975.088.656.045.0141.2177.2178.4177.8
siMLPe [41]14.926.244.056.668.842.148.8150.4236.3257.8247.1
In-Context Model: one task-agnostic architecture for multiple tasks
Static-SiC (ours)7.921.537.754.069.138.122.076.4162.5171.9167.2
Dynamic-SiC (ours)7.121.337.453.568.037.521.776.0150.1151.7151.0
" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.248, + 0.852, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.498, + 0.891, + 0.513 + ], + "angle": 0, + "content": "Figure 5. Comparison of generalization. Our SiC completes the missing skeletons according to the customized prompt for an unseen task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.469, + 0.551 + ], + "angle": 0, + "content": "Copy Method. Following [5, 16], we copy the prompt target \\( T_{i}^{k} \\) as prediction results for comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Multi-Task-Capable Models. For task-specific models [41, 52, 65] and multi-stage models [77], they are not able to handle multiple tasks at the same time. Therefore, we remodel them into end-to-end multi-task-capable models to achieve fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.732 + ], + "angle": 0, + "content": "In-Context Models. Since there are no in-context models excelling in processing multiple human perception tasks, we select the most relative in-context models with our benchmark, Point-In-Context [16] and PointMAE [43], which is proposed in 3D point cloud understanding. Specifically, we take each frame as a token in their framework. For PointMAE, we use the transformed version in [16]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.353, + 0.758 + ], + "angle": 0, + "content": "5.2. Compared with SoTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We report experimental results on four tasks in Tab. 1. We term Skeleton-in-Context implemented with static/dynamic-TUP as Static/Dynamic-SiC, respectively. For other models, we reproduce them and report the results. As can be seen from Tab. 1, our proposed model outperforms all multi-task models, achieving state-of-the-art results on the in-context multi-task benchmark. Notably, the results of task-specific models on a single task reflect the difficulty of each task. However, we also exhibit impressive" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.893, + 0.565 + ], + "angle": 0, + "content": "performance compared to task-specific models, demonstrating that our proposed model is an expert in handling skeleton sequence-based multi-tasking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.894, + 0.657 + ], + "angle": 0, + "content": "Qualitative Results. We visualize and compare the results of our SiC and the most recent SoTA model, MotionBERT [77], which is re-structured as an end-to-end multi-task model for fair comparison. As highlighted in Fig. 4, our SiC can generate more accurate poses than MotionBERT according to the provided task-guided prompt." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.74, + 0.683 + ], + "angle": 0, + "content": "5.3. Generalization Capability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Generalize to New Datasets. To verify the generalization ability of our model, we conduct cross-dataset experiments on a specific task. For instance, our proposed model learns motion prediction on the AMASS dataset under the default benchmark, and then it is used to predict future motion sequences on 3DPW during testing. Similarly, we set up four experiments: motion prediction (AMASS \\(\\rightarrow\\) 3DPW), pose estimation (H3.6M \\(\\rightarrow\\) AMASS, 3DPW), and joint completion (3DPW \\(\\rightarrow\\) H3.6M). As shown in Tab. 2, our model is able to generalize the learned task well to other datasets, while other task-specific models cannot accommodate the large bridge introduced across datasets even if they are specifically trained on this single task. Additionally, benefiting from the dynamic pseudo pose, the Dynamic-SiC exhibits" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2442" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.09, + 0.411, + 0.104 + ], + "angle": 0, + "content": "Table 3. Effectiveness of the Skeleton Prompt." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.109, + 0.457, + 0.176 + ], + "angle": 0, + "content": "
#TGPTUPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1XX111.7159.2543.7160.2
2✓X25.052.537.765.2
3X✓26.052.437.878.4
4✓✓24.451.836.862.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.164, + 0.178, + 0.382, + 0.191 + ], + "angle": 0, + "content": "Table 4. Different Prompting (TGP)." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.197, + 0.461, + 0.247 + ], + "angle": 0, + "content": "
#TGPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1Feature Sim.25.252.137.764.1
2Motion Sim.25.252.137.864.1
3Random24.451.836.862.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.449, + 0.274 + ], + "angle": 0, + "content": "more excellent generalization ability than the Static-SiC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Generalize to New Task. Furthermore, We construct a prompt that has not appeared in the training set, which performs motion in-between. As Fig. 5 shows, our trained model can complete the missing skeletons according to the customized task-guided prompt. Our SiC generalizes well to a new task, which is unexplored in other works." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.231, + 0.391 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.565 + ], + "angle": 0, + "content": "Effectiveness of the Skeleton Prompt. We conduct ablation experiments on our proposed skeleton prompts. As shown in Tab. 3, the skeleton prompts solve the problem of directly extending previous work to skeleton sequences, which is falling into over-fitting. Meanwhile, applying both TGP and TUP to the model can achieve the best results. Our proposed skeleton prompts replace the random masking training strategy used in previous work. In particular, TUP creates a training environment that does not require the query target, which ensures that the model will not see the ground truth of the query input in advance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.745 + ], + "angle": 0, + "content": "Prompt Engineering. Since previous work has proven that different prompts in context learning affect the performance of the model, we conduct ablation experiments on the choice of prompts. As Tab. 4 shows, we select the prompt by the feature similarity and MPJPE minimum between the input skeleton sequence and the prompt, respectively. Experimental results show that the random selection method achieves the best results. Unlike 2D images and 3D point clouds, skeleton-based task data is highly similar, resulting in different prompts playing similar roles in the same task. While the random selection method expands the model's receptive field to the data and achieves better results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Architecture. We conduct ablation experiments on the architecture of the model. As shown in Tab. 5, after exploring different hidden feature dimensions and transformer depths, we find that the model performs best when the feature dimension is 256, and the depth is 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.288, + 0.847 + ], + "angle": 0, + "content": "5.5. In-Context Initial Pose" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We conduct an in-depth exploration of Dynamic-TUP. We extract the trained Dynamic-TUP weights, then randomly initialize a \\((F,J,3)\\) tensor and use the trained encoder to in" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.09, + 0.816, + 0.103 + ], + "angle": 0, + "content": "Table 5. Ablation on model architecture." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.109, + 0.889, + 0.34 + ], + "angle": 0, + "content": "
Arch.Num.MP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
Dim.6429.357.145.880.5
12826.453.839.266.1
25624.451.836.862.9
51224.750.737.663.0
Dep.325.153.739.265.6
425.152.238.664.7
524.451.836.862.9
624.752.938.163.7
(a) Task-Unified Prompt (Static)
(b) Task-Unified Prompt (Dynamic)
" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 6. (a) The visualization of Static-TUP. (b) The concrete visual skeleton sequence of Dynamic-TUP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.895, + 0.56 + ], + "angle": 0, + "content": "crease its feature dimension to C-dimension. Finally, an optimizer is used to optimize the tensor and minimize its feature similarity with the trained Dynamic-TUP. The top of Fig. 6 represents the visualization of Static-TUP, while the bottom is the Dynamic-TUP. The concrete skeleton of dynamic-TUP is similar to the mean pose Static-TUP of the training set, which demonstrates that the Dynamic-TUP learned by our model is actually the prior knowledge whose receptive field is all ground truth in the training set. This prior knowledge helps our model learn more smoothly on multiple tasks, and adding learnable modules helps our model generalize to other datasets or new tasks (Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.619, + 0.589 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.794 + ], + "angle": 0, + "content": "In this work, we propose the Skeleton-in-Context, designed to process multiple skeleton-base tasks simultaneously after just one training time. Specifically, we build a novel skeleton-based in-context benchmark covering various tasks. In particular, we propose skeleton prompts composed of TGP and TUP, which solve the overfitting problem of skeleton sequence data trained under the training framework commonly applied in previous 2D and 3D in-context models. Besides, we demonstrate that our model can generalize to different datasets and new tasks, such as motion completion. We hope our research builds the first step in the exploration of in-context learning for skeleton-based sequences, which paves the way for further research in this area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work is supported by National Natural Science Foundation of China (No. 62203476), Natural Science Foundation of Shenzhen (No. JCYJ20230807120801002). This project is also supported by the interdisciplinary doctoral grants (iDoc 2021-360) from the Personalized Health and Related Technologies (PHRT) of the ETH domain, Switzerland." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2443" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.183 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.472, + 0.282 + ], + "angle": 0, + "content": "[3] Ivana Balažević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Hénaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.472, + 0.324 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.367 + ], + "angle": 0, + "content": "[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.472, + 0.422 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.472, + 0.465 + ], + "angle": 0, + "content": "[7] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.472, + 0.521 + ], + "angle": 0, + "content": "[8] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.472, + 0.59 + ], + "angle": 0, + "content": "[9] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.592, + 0.472, + 0.633 + ], + "angle": 0, + "content": "[10] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.635, + 0.472, + 0.674 + ], + "angle": 0, + "content": "[11] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.676, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[12] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. Unihcp: A unified model for human-centric perceptions. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[13] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.775, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[14] Zhichao Deng, Xiangtai Li, Xia Li, Yunhai Tong, Shen Zhao, and Mengyuan Liu. Vg4d: Vision-language model goes 4d video recognition. In ICRA, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[16] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "learning for 3d point cloud understanding. arXiv preprint arXiv:2306.08659, 2023. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.163 + ], + "angle": 0, + "content": "[17] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In CVPR, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[18] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.894, + 0.263 + ], + "angle": 0, + "content": "[19] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.266, + 0.894, + 0.306 + ], + "angle": 0, + "content": "[20] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.353, + 0.894, + 0.395 + ], + "angle": 0, + "content": "[22] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.397, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[23] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.894, + 0.495 + ], + "angle": 0, + "content": "[24] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. PAMI, 2013. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[25] Mahdi M Kalayeh, Emrah Basaran, Muhittin Gokmen, Mustafa E Kamasak, and Mubarak Shah. Human semantic parsing for person re-identification. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.894, + 0.582 + ], + "angle": 0, + "content": "[26] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 3DV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[27] Sameh Khamis, Cheng-Hao Kuo, Vivek K Singh, Vinay D Shet, and Larry S Davis. Joint learning for attribute-consistent person re-identification. In ECCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[28] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[29] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Guangliang Cheng, Pang Jiangmiao, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv pre-print, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[30] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into person: Joint body parsing & pose estimation network and a new benchmark. PAMI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[31] Chunze Lin, Jiwen Lu, and Jie Zhou. Multi-grained deep feature learning for pedestrian detection. In ICME, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[32] Jun Liu, Henghui Ding, Amir Shahroudy, Ling-Yu Duan, Xudong Jiang, Gang Wang, and Alex C Kot. Feature boosting network for 3d pose estimation. PAMI, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[33] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. PAMI, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2444" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[34] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[35] Diogo C Luvizon, David Picard, and Hedi Tabia. 2d/3d pose estimation and action recognition using multitask deep learning. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[36] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[37] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[38] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[39] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[40] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[41] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[42] Xuecheng Nie, Jiashi Feng, and Shuicheng Yan. Mutual learning to adapt for joint human parsing and pose estimation. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.511 + ], + "angle": 0, + "content": "[43] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[44] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[46] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[47] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[48] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu \\(\\mathrm{rgb + }\\) d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[49] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[50] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry Steven Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for multi-camera person re-identification. PAMI, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[51] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.893, + 0.189 + ], + "angle": 0, + "content": "[52] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatio-temporal criss-cross attention. In CVPR, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.189, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[53] Yonglong Tian, Ping Luo, Xiaogang Wang, and Xiaou Tang. Pedestrian detection aided by deep learning semantic tasks. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.231, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.328 + ], + "angle": 0, + "content": "[55] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.328, + 0.893, + 0.384 + ], + "angle": 0, + "content": "[56] Chaoyang Wang, Xiangtai Li, Henghui Ding, Lu Qi, Jiangning Zhang, Yunhai Tong, Chen Change Loy, and Shuicheng Yan. Explore in-context segmentation via latent diffusion models. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.893, + 0.426 + ], + "angle": 0, + "content": "[57] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.893, + 0.469 + ], + "angle": 0, + "content": "[58] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.893, + 0.511 + ], + "angle": 0, + "content": "[59] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.553 + ], + "angle": 0, + "content": "[60] Xinshun Wang, Qiongjie Cui, Chen Chen, and Mengyuan Liu. Gcnext: Towards the unity of graph convolutions for human motion prediction. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.893, + 0.595 + ], + "angle": 0, + "content": "[61] Xinshun Wang, Wanying Zhang, Can Wang, Yuan Gao, and Mengyuan Liu. Dynamic dense graph convolutional network for skeleton-based human motion prediction. TIP, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.893, + 0.65 + ], + "angle": 0, + "content": "[62] Zhendong Wang, Yifan Jiang, Yadong Lu, Yelong Shen, Pengcheng He, Weizhu Chen, Zhangyang Wang, and Mingyuan Zhou. In-context learning unlocked for diffusion models. arXiv preprint arXiv:2305.01115, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.893, + 0.707 + ], + "angle": 0, + "content": "[63] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.893, + 0.762 + ], + "angle": 0, + "content": "[64] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[65] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[66] Shilin Xu, Xiangtai Li, Jingbo Wang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Fashionformer: A simple, effective and unified baseline for human fashion segmentation and recognition. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[67] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2445" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[68] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.471, + 0.176 + ], + "angle": 0, + "content": "[69] Angela Yao, Juergen Gall, and Luc Van Gool. Coupled action recognition and pose estimation from multiple views. IJCV, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[70] Bruce XB Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[71] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[72] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[73] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[74] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.407, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.449, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[76] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, et al. Generative tweening: Long-term inbetweening of 3d human motions. arXiv preprint arXiv:2005.08891, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[77] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In ICCV, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2446" + } + ] +] \ No newline at end of file diff --git a/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_origin.pdf b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..67a26cf845c5ea935fc3390a8e250dbb62806ca0 --- /dev/null +++ b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/d6e51c0b-3655-41d0-b00a-3888d1dd8906_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6ba5b599198c4488266d31da3c1b851f7ca9f75a3b325a67fd3d13d7fa065e4 +size 4686531 diff --git a/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/full.md b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb3684ab6a7aae57f7d0622d5961d60814c714ce --- /dev/null +++ b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/full.md @@ -0,0 +1,323 @@ +# Skeleton-in-Context: Unified Skeleton Sequence Modeling with In-Context Learning + +Xinshun Wang $^{1,2,*}$ , Zhongbin Fang $^{1,*}$ , Xia Li $^{3}$ , Xiangtai Li $^{4}$ , Chen Chen $^{5}$ , Mengyuan Liu $^{2, \boxtimes}$ , Sun Yat-sen University + +$^{2}$ National Key Laboratory of General Artificial Intelligence, Peking University, Shenzhen Graduate School + +$^{3}$ Department of Computer Science, ETH Zurich $^{4}$ S-Lab, Nanyang Technological University + +$^{5}$ Center for Research in Computer Vision, University of Central Florida + +{wangxsh36,fangzhb5}@mail2.sysu.edu.cn，xia.li@inf.ethz.ch，xiangtai94@gmail.com + +chen.chen@crcv.ucf.edu, liumengyuan@pku.edu.cn + +https://github.com/fanglaosi/Skeleton-in-Context + +# Abstract + +In-context learning provides a new perspective for multi-task modeling for vision and NLP. Under this setting, the model can perceive tasks from prompts and accomplish them without any extra task-specific head predictions or model fine-tuning. However, skeleton sequence modeling via in-context learning remains unexplored. Directly applying existing in-context models from other areas onto skeleton sequences fails due to the similarity between inter-frame and cross-task poses, which makes it exceptionally hard to perceive the task correctly from a subtle context. To address this challenge, we propose Skeleton-in-Context (SiC), an effective framework for in-context skeleton sequence modeling. Our SiC is able to handle multiple skeleton-based tasks simultaneously after a single training process and accomplish each task from context according to the given prompt. It can further generalize to new, unseen tasks according to customized prompts. To facilitate context perception, we additionally propose a task-unified prompt, which adaptively learns tasks of different natures, such as partial joint-level generation, sequence-level prediction, or 2D-to-3D motion prediction. We conduct extensive experiments to evaluate the effectiveness of our SiC on multiple tasks, including motion prediction, pose estimation, joint completion, and future pose estimation. We also evaluate its generalization capability on unseen tasks such as motion-in-between. These experiments show that our model achieves state-of-the-art multi-task performance and even outperforms single-task methods on certain tasks. + +# 1. Introduction + +The idea of multi-task models has attracted much research interest over the years in a wide range of areas [6, 16, 17, 58], such as computer vision and natural language processing + +![](images/50b983fa9eab11369c47a0eaa679044708a3e891ab5e57c82a5c2dc72a9227ed.jpg) + +![](images/b2de66d6fc38ae2c4f4e17186707ca7dd9c91d6dff9c04e373f646dc75e83679.jpg) + +![](images/5f8b91c68bb67cb337a07a1263d4fa0a79c3d915a703f9d361b17a941f59c566.jpg) +Figure 1. In-context learning in a) image modeling [58], b) point cloud modeling [16], and c) skeleton sequence modeling (ours). + +(NLP). To build such models, most previous works focus on deploying task-specific heads on a shared backbone and learning them through pretraining and fine-tuning. Models with such task-specific architectures are not learned in an end-to-end manner and are hard to generalize to new, unseen tasks and data. Most existing methods for skeleton-based human-centric tasks rely on task-specific designs and/or non-end-to-end frameworks to accomplish multiple tasks. Several works [9, 36] learn unified human motion representations. In particular, MotionBERT [77] adopts two-stage training, including pre-training on the 2D-3D lifting task + +and fine-tuning on each task. UniHCP [12] connects a well-designed shared backbone network to specific task heads according to different tasks. Meanwhile, UPS [17] unifies action labels and motion sequences into text-based language sequences and fine-tunes them on each task for better performance. These models can have a limited scope of tasks they are capable of handling and lack flexibility. This limits their capability to generalize to unseen tasks and data. To overcome these challenges, in-context learning represents a new trend to build generalist models, which was originally proposed in NLP [6, 45, 46] and then introduced into other areas such as images and point clouds. Some generalist models [5, 58, 59, 72] utilize the Masked Image Modeling (MIM) framework to exhibit multi-task capability via in-context learning in image processing, as shown in Fig. 1(a). Similarly, in Fig. 1(b), Point-In-Context [16] proposes a key joint sampling module and introduces in-context learning into point cloud processing. Specifically, in-context models learn patterns of provided input-output pairs, called in-context examples or prompts, and then perform the desired task according to the prompt. An in-context model can accomplish diverse tasks end-to-end without any task-specific heads or fine-tuning. It also has excellent generalization capability. Directly adapting existing in-context models from other areas onto skeleton sequences fails because they largely focus on static objects (images or point clouds) and thus cannot capture spatial-temporal dependencies in skeleton sequences. To the best of our knowledge, in-context learning remains unexplored in 3D skeleton sequence modeling. + +In this paper, we propose Skeleton-in-Context (SiC), a novel framework for skeleton sequence modeling, which is able to process multiple tasks simultaneously after a single training process without any task-specific designs, as Fig. 1(c) shows. Specifically, SiC perceives tasks from context, i.e., the prompt from a desired task provided as additional input, and then accomplishes the task for the query. SiC employs two types of prompts to facilitate context perception: the task-guided prompt (TGP) and the task-unified prompt (TUP). TGP provides a task template for the model, giving the model the ability to make analogies. This is the key to the model's capability to generalize to new tasks, as we can guide the model to do unseen tasks by customizing specific prompts. TUP can be implemented in two ways: static and dynamic. The static TUP leverages prior knowledge shared among tasks, while the dynamic TUP adaptively learns to perceive task context. Skeleton-in-Context provides a new perspective for training a skeleton-based multi-task in-context model. Since there is no existing benchmark for measuring such a model, we establish a new in-context benchmark based on four tasks, including three skeleton-based tasks: motion prediction, pose estimation, and joint completion, and a new task: future pose estimation. + +Our main contributions are as follows: 1) We introduce + +a novel framework, Skeleton-in-Context (SiC), designed to handle multiple skeleton-based tasks simultaneously. Our approach learns the potential pattern of the given prompt and performs it on the query sample. 2) We establish a new 3D human in-context learning benchmark comprised of multiple skeleton-based tasks, such as motion prediction, pose estimation, joint completion, and future motion estimation. We further benchmark several representative in-context learning baselines for reference. 3) We conduct extensive experiments to evaluate our SiC. SiC achieves state-of-the-art results in multi-tasking and even outperforms task-specific models on certain tasks. Moreover, SiC also generalizes well to unseen tasks such as motion in-between. + +# 2. Related Work + +Skeleton-Based Motion Analysis. Many of the skeleton-based human-centric tasks have shared similar development patterns in terms of the methods used to address them. On human motion prediction, early successes are achieved with RNNs and CNNs [28, 40]. Nowadays, the task is dominated by Graph Convolutional Networks (GCNs). LTD [38] considers each joint coordinate as a node in a graph on which to employ learnable spatial graph convolutions. Follow-up works [13, 39, 60, 61] focus on designing better graph convolution networks. Different from motion prediction, the goal of motion synthesis is to complete the human motion sequence. To complete the missing parts of human motion, many works adopt convolutional models [22, 26, 68] or generative adversarial networks [7, 20, 23, 76]. One category of 3D pose estimation [18, 32, 34] is to lift the estimated 2D pose to 3D, which often employs spatial-temporal convolution [8, 10, 11, 41, 44, 57, 67] or transformer-based [14, 49, 71, 73-75] methods. However, these approaches focus on designing customized architectures specific to a task, which are limited in range of applications. To this end, we propose to unify the input and output formats of multiple tasks and utilize a single model to process multiple tasks simultaneously via in-context learning. + +In-Context Learning. The emergence of in-context learning provides a new paradigm for achieving multitasking, which originated from Natural Language Processing (NLP) [1, 15, 45-47]. It endows models with the capability to handle multiple tasks and even unseen tasks by incorporating domain-specific input-target pairs, known as prompts, into the query sample [6]. Following NLP, in-context learning is extended to computer vision in some work, such as 2D images [51, 56, 58, 59, 62, 72] and 3D point clouds [3, 16, 64]. Specifically, Visual Prompt [5], Painter [58], and Point-In-Context [16] are pioneers of in-context methods in computer vision. They verify the effectiveness and flexibility of in-context learning in vision-based tasks, and most of them adopt pure-transformer-based approaches [4, 21, 29, 63]. Meanwhile, several approaches + +deeply explore Prompt Engineering [3, 51, 72] and awe-inspiring generalization [16, 59, 64] in in-context learning. Recently, LVM [2] unifies most vision tasks and data via joint co-training on one visual in-context model. Compared to the works focused on static objects (images and point clouds), our work mainly delves into modeling in-context dynamic skeleton sequences. + +Unified Skeleton-Centric Models. A generalist model that can excel in handling multiple tasks is a profound step toward general artificial intelligence [45, 58]. Previous works [25, 27, 31, 35, 50, 53, 66, 69] give the effort to explore co-training several human-centric tasks. Meanwhile, some works [30, 42] show great progress in simultaneously predicting parsing and pose. Recent research [9, 36] demonstrates the effectiveness of modeling unified human motion sequences. In particular, UniHCP [12] designs specific task heads according to different 2D human-centric tasks. Motion-BERT [77] is pre-trained in the 2D-3D lifting task and fine-tuned on each task. UPS [17] unifies heterogeneous human behavior understanding tasks via language sequences and cycle training for each task in order. Nevertheless, these works are limited by the dilemma of complex training stages or task heads, which introduce extra costs. We integrate the output format of different tasks in skeleton sequence format and train the model just once. + +# 3. In-Context Skeleton Sequence Modeling + +# 3.1. Formulation + +Since the skeleton sequence data in skeleton-based tasks are sequential by nature, we treat the skeleton at each frame as a token, which can be processed by transformer [54]. Following previous works [5, 16, 58], we curate an in-context skeleton sequence paradigm. Given an input-target skeleton sequence pair as a task-guided prompt $G_{i}^{k} = \{I_{i}^{k}, T_{i}^{k}\}$ and a query input $Q_{j}^{k}$ , where $I_{i}^{k}, T_{i}^{k}, Q_{j}^{k} \in \mathbb{R}^{F \times J \times 3}$ each have $F$ frames and $J$ joints, the model is expected to perceive the context from the prompt, i.e., which task it represents, and perform the corresponding operations on the query in analogy to the prompt. This process can be formulated as: + +$$ +y = f \left(G _ {i} ^ {k}, Q _ {j} ^ {k}\right), \tag {1} +$$ + +where $k$ represents the task index, while $i$ and $j$ indicate different example indexes. Thus, we can determine what task will be performed on the query sample by selecting the task-specific prompt from the corresponding training set. + +# 3.2. Task Definitions + +To measure multitasking capability, we establish a new in-context benchmark from common skeleton-based datasets, including H3.6M [24], AMASS [37], and 3DPW [55]. In this benchmark, we select three common tasks: motion prediction, pose estimation, and joint completion, and a new + +![](images/c3f6a43a5f06aa2a32c8a96539d22a1ca563999491ee6f59866ccf78f60081f0.jpg) +Figure 2. Top: Training with MIM-style framework [21, 43] used in previous works [5, 16, 58]. The model is able to reconstruct the masked frames well. Bottom: During inference, the reconstructed sequence gradually shrinks the skeleton as time goes by because the model only learns frame interpolation during training. Once the model loses subsequent reference frames when interpolating frames, the generated sequence will tend to shrink. + +task: future pose estimation. Consequently, we conduct extensive experiments on a large and diverse dataset collection, which consists of 176K skeleton sequence samples. + +Motion Prediction. Motion prediction [38] is to predict future motion based on historical motion. There is continuity between the input and output sequences, which constitute a complete motion. Thus, the input $I$ and target $T$ are complete skeleton sequences with shapes of $(F, J, 3)$ . + +Pose Estimation. Pose estimation [18, 32, 34] is to lift 2D poses to 3D. The model takes the detected 2D skeleton sequence $I \in \mathbb{R}^{F \times J \times 2}$ , which contain noise, as input, and estimates the accurate 3D skeleton sequence $T \in \mathbb{R}^{F \times J \times 3}$ . Note that we expand the 2D input to 3D by filling the 3rd dimension with 0, in order to unify the data format with other tasks. Therefore, the input skeleton sequence becomes the same shape $(F, J, 3)$ as the target sequence. + +Future Pose Estimation. Based on prediction and estimation, we propose a new task: future pose estimation, which is to predict future 3D motion based on history 2D poses. As 2D images are easier to obtain than 3D data, this task unlocks a host of new real-world applications. Given a historical 2D skeleton sequence $I \in \mathbb{R}^{F \times J \times 3}$ (one dimension with 0), future pose estimation expects the model to output its future 3D skeleton sequence $T \in \mathbb{R}^{F \times J \times 3}$ . To our knowledge, no work has ever proposed future pose estimation. + +Joint Completion. Joint completion is to recover skeleton sequences that have missing joints. This task can reflect cases where joints may be missing or displaced when collected with Lidar cameras [33, 48]. We set two levels to verify the ability to complete joints, which are $40\%$ and $60\%$ of the joints missing in each frame. Also, the coordinates of missing joints are replaced with 0, with the goal of unifying the input skeleton sequence format with $(F, J, 3)$ . + +![](images/0ec1efdc4867b9d4bb6239732dfa04df9a7098f76d6928d5cc6b1799ab26928c.jpg) +Figure 3. Overall framework of our Skeleton-in-Context. Specifically, we establish a skeleton bank by integrating training sets under different tasks, which contain a large amount of input-target pairs performing different tasks. Next, we randomly select a sample pair as the task-guided prompt (TGP) and a query input from the skeleton bank, undergo encoding and concatenating, respectively, and then input them into the transformer in parallel. In particular, during this process, the query input and task-unified prompt (TUP) are combined to form a new query. After iterating $n_1$ times, the TGP and query are aggregated through $\mathcal{A}(\cdot)$ and then input into the transformer for $n_2$ iterations. Lastly, the second half of the model output is used as our prediction. + +# 4. Skeleton-in-Context + +In-context learning on other data modalities such as images and point clouds [5, 16, 58] largely employ Mask Image Modeling (MIM) [21] to reconstruct the randomly masked elements. During inference, only the target elements are masked. On skeleton sequences, MIM fails due to spatial-temporal complexity and inter-pose similarity, which makes it outstandingly hard to perceive the task correctly from a subtle context. In our Skeleton-in Context, we employ a new framework to address this challenge. + +# 4.1. Skeleton Sequence Prompt + +Task-Guided Prompt. In in-context learning, a prompt is needed for the model to learn the underlying task pattern from context. As mentioned in Sec. 3.1, we propose the Task-Guided Prompt (TGP) $G_{i}^{k} = \{I_{i}^{k}, T_{i}^{k}\}$ , which comprises of an example skeleton pair performing a specific task $k$ . In this way, our proposed model is able to perform a variety of tasks according to the provided prompts that could be seen or even unseen in the training set. It is worth noting that our model can perform unseen tasks and transfer to datasets that are different styles from the training dataset. It will be elaborated in Sec. 5.2. + +Avoiding Over-Fitting. Previous works have made great progress in unifying different tasks via the MIM-style framework [5, 16, 58]. Compared with static objects in 2D and 3D, such as images and point clouds, the obstacle to in-context learning of skeleton sequences lies in the wondrous similarity of each frame, so directly using the random masking strategy [5, 16, 58] will cause the model to fall into overfitting, leading it to only learn simple interpolation frames rather than capturing meaningful motion sequence representations during the training stage. As shown in Fig. 2, when + +we directly utilize the MPM-style [21, 43] framework to train our model, although it can perfectly reconstruct the masked frame during training, the experimental results reveal that it outputs a devastating result during inference. It is obvious that the output skeleton sequence gradually shrinks as the length increases because the next frame of the current frame is 0, and interpolating with 0 will lead to shrinking. Such a phenomenon underscores the challenges of learning contextual information about dynamic skeleton sequences. + +Task-Unified Prompt. With the Task-Guided Prompt, which exhibits unique capabilities specific to in-domain or out-of-domain tasks, mitigating the gap between different tasks presents a new challenge. Meanwhile, as discussed above, a straight extension of previous works will lead to a simple model that can only perform the interpolation task based on the given skeleton sequence. To tackle these challenges, we introduce the Task-Unified Prompt (TUP), a well-designed and task-agnostic module, which can adaptively learn to incorporate multiple tasks into the framework of in-context learning, i.e., a one-off end-to-end framework for all tasks without having to deploy task-specific heads or fine-tuning. We introduce two versions of TUP: + +Version 1: Static-TUP with static pseudo pose: + +$$ +U ^ {S} = \mathcal {E} \left(\frac {1}{M} \sum_ {m = 1} ^ {M} \sum_ {k = 1} ^ {K} \left(T _ {m} ^ {k}\right) \in \mathbb {R} ^ {F \times J \times 1}\right), \tag {2} +$$ + +where $\mathcal{E}(\cdot)$ is a embedding layer. $M$ and $K$ indicate the total number of training samples and tasks, respectively. + +Version 2: Dynamic-TUP with dynamic pseudo pose: + +$$ +U ^ {D} = W _ {1} W _ {2}, \tag {3} +$$ + +where $W_{1}\in \mathbb{R}^{F\times J\times 1}$ and $W_{2}\in \mathbb{R}^{1\times 1\times C}$ . + +Table 1. Comparison of task-specific models, multi-task models, and in-context models on four human skeleton tasks. For Motion Prediction (MP.), Joint Completion (JC.), and Future Pose Estimation (FPE.), we report Mean Per Joint Position Error (mm) [24]. For Pose Estimation (PE.), we additionally report another indicator, Normalized Mean Per Joint Position Error (N-MPJPE) [77]. For joint completion, we report two settings: $40\%$ and $60\%$ of the joints missing in each frame. $\dagger$ represents the multi-task-capable models re-structured from task-specific models. $\ddagger$ represents the multi-task-capable models re-structured from multi-stage models. Static-SiC and Dynamic-SiC represent the static and dynamic versions of TUP, respectively. + +
MethodsVenuesMP. (AMASS)PE. (H3.6M)JC. (3DPW)FPE. (H3.6M)
80ms160ms200ms320ms400msAvg.MPJPE↓ Avg.N-MPJPE↓ Avg.MPJPE↓ 40%60% Avg.MPJPE↓ 200ms300ms Avg.
Task-Specific Model: one architecture for one task
siMLPe [19]WACV'2312.121.625.434.439.926.758.949.748.458.553.575.4
EqMotion [65]CVPR'2314.426.631.744.050.033.363.259.340.449.745.169.5
STCFormaler [52]CVPR'2319.525.128.235.841.528.754.645.231.246.738.960.9
GLA-GCN [70]ICCV'2319.525.128.235.342.030.049.640.462.272.567.449.5
MotionBERT [77]ICCV'2312.921.425.536.342.326.749.441.129.344.336.875.5
Multi-Task Model: one architecture (w. multiple task-specific heads) for multiple tasks
siMLPe†[41]WACV'2319.030.035.346.852.436.774.660.059.067.463.264.4
EqMotion†[65]CVPR'2314.025.330.240.546.631.3163.3110.138.956.047.588.1
STCFormaler†[52]CVPR'2317.428.031.543.040.433.955.645.633.650.642.162.2
GLA-GCN†[70]ICCV'2335.339.842.349.256.144.578.662.974.181.978.060.6
MotionBERT‡[77]ICCV'2314.322.826.736.441.728.455.845.731.346.739.075.5
In-Context Model: one task-agnostic architecture for multiple tasks
Copy [5, 16]-134.7133.7133.3132.1131.4133.0351.0171.6196.5195.5196.0316.8
PointMAE [43]ECCV'2219.131.236.654.967.741.9144.396.4167.679.073.3139.9
PIC [16]NeurIPS'2325.040.246.867.279.451.7181.6118.885.796.491.1173.1
Static-SiC (ours)-11.019.022.732.637.824.651.641.829.844.737.259.8
Dynamic-SiC (ours)-11.018.922.632.237.424.451.842.529.544.036.859.2
+ +We denote the SiC-based on Static-/Dynamic-TUP is referred to as Static-/Dynamic-SiC, respectively. The proposed TUP facilitates the modeling of distribution differences among tasks and datasets, so the model focuses more on learning unified representations shared across tasks. + +# 4.2. Model Architecture + +As shown in Fig. 3, we build a skeleton bank, a collection of publicly available datasets including H3.6M, AMASS and 3DPW, covering common skeleton analysis tasks, containing input-target pairs formed as sequences of equal length, where prompts are randomly selected. Given a prompt pair $G_{i}^{k} = \{I_{i}^{k}, T_{i}^{k}\}$ and a query sample $Q_{j}^{k}$ , we first map them to the C-dimensional feature space, and then splice $I$ and $T$ , $Q$ and $U$ respectively. The former represents our proposed Task-Guided Prompt, and the latter is the objective sequence of our focused analysis. The process can be formulated as: + +$$ +\widetilde {G} = \mathcal {E} \left(G _ {i} ^ {k}\right) \in \mathbb {R} ^ {2 F \times J \times C}, \tag {4} +$$ + +$$ +\widetilde {Q} = \left(\mathcal {E} \left(Q _ {j} ^ {k}\right) \| U\right) \in \mathbb {R} ^ {2 F \times J \times C}, \tag {5} +$$ + +where $\mathcal{E}(\cdot)$ is a linear transformation and $\|\cdot\|$ is concatenating operator. In essence, $U$ represents the prior knowledge of the target skeleton sequence and is also an adapter that adapts to different tasks and different distributions of data for learning. + +Drawing inspirations from [77], we connect spatial attention blocks and temporal attention blocks together in an alternating combination manner to form a two-stream + +transformer $\mathcal{M}(\cdot)$ dedicated to capturing the dynamic characteristics of motion, which is implemented as: + +$$ +\mathcal {M} (\mathrm {X}) = \alpha^ {i} \mathcal {T} _ {1} ^ {i} \left(\mathcal {S} _ {1} ^ {i} \left(\mathrm {X} ^ {i - 1}\right)\right) + \beta^ {i} \mathcal {S} _ {2} ^ {i} \left(\mathcal {T} _ {2} ^ {i} \left(\mathrm {X} ^ {i - 1}\right)\right), \tag {6} +$$ + +where $\alpha$ and $\beta$ are balanced parameters determined adaptively. $S_{1}^{i}$ and $\mathcal{T}_1^i$ are implemented using spatial selfattention mechanism and temporal self-attention mechanism, respectively. Additionally, $i = 1,\dots,N$ . + +Then, $\mathcal{M}(\cdot)$ takes $\widetilde{G}$ and $\widetilde{Q}$ as input and parallel iterates over them $n_1$ times respectively. Then, $\widetilde{G}^{n1}$ and $\widetilde{Q}^{n1}$ are integrated via aggregation function $\mathcal{A}(\cdot)$ : + +$$ +W ^ {n 1} = \mathcal {A} \left(\widetilde {G} ^ {n 1}, \widetilde {Q} ^ {n 1}\right), \tag {7} +$$ + +we consider $\mathcal{A}(\cdot)$ to be a simple weighted summation function. The final output is obtained after iterating $W$ $n_2$ times in $\mathcal{M}(\cdot)$ , where $n_1 + n_2 = N$ . Note that we take the second half of the output as the prediction. + +# 4.3. Training and Inference + +Training. We train our model in our proposed skeleton-centric in-context dataset mentioned in Sec. 3. In this training way, our Skeleton-in-Context can perform multiple tasks after one training. Since our output format is skeleton-like, following previous work [77], we adopt the 3D joint-level reconstruction loss: + +$$ +L o s s = \sum_ {f = 1} ^ {F} \sum_ {j = 1} ^ {J} \left\| P _ {f, j} - G _ {f, j} \right\| _ {2}, \tag {8} +$$ + +![](images/e88b61c3309694e53b4db3a976de1f3633df3f7143bed116c9f205a0e50dbafc.jpg) +Figure 4. The comparison of visual results between our Skeleton-in-Context and recent SoTA method MotionBERT [77] on four tasks. Our method generates more accurate poses than MotionBERT, as can be seen from the spots with red arrows $(\searrow)$ pointed to them. + +Inference. After a training phase involving all tasks, we get an in-context model that focuses on dynamic skeleton sequences. It can perform multiple tasks at the same time according to the provided task-dependent prompt without any further parameter updates. + +# 5. Experiments + +Implementation Details. We implement the proposed Skeleton-in-Context model with the number of layers $N = 5$ , number of attention heads $H = 8$ , and hidden feature dimension $C = 256$ . For each prompt input/target and query input/target, the sequence length is $F = 16$ . All tasks are unified into a one-off, end-to-end training process without + +any task-specific designs. The training takes about 6 hours for 120 epochs on 4 NVIDIA GeForce RTX 4090 GPUs. + +Datasets and Metrics. H3.6M (Human3.6M) is used for the pose estimation and future pose estimation tasks. AMASS is used for the motion prediction task. 3DPW (3D Pose in the Wild) is used for the joint completion task. All of the tasks can be evaluated using Mean Per Joint Position Error. + +# 5.1. Baseline Methods + +In order to make a qualitative comparison with other methods in a relatively fair situation, we reimplement the task-specific SoTA methods in our proposed benchmark. Additionally, we transform recent methods into multi-task models and compare them with our model. + +Table 2. Comparison results between task-specific models and our proposed models on generalization testing experiments on different datasets. We report Mean Per Joint Position Error (mm) for all tasks. + +
MethodsMP. (3DPW)PE.JC. (H3.6M)
40ms80ms120ms160ms200msAvg.AMASS3DPW40%60%Avg.
Task-Specific Model: one architecture for one task
EqMotion [65]15.635.652.167.979.550.151.8156.6193.6194.0193.8
STCFformer [52]18.339.358.975.088.656.045.0141.2177.2178.4177.8
siMLPe [41]14.926.244.056.668.842.148.8150.4236.3257.8247.1
In-Context Model: one task-agnostic architecture for multiple tasks
Static-SiC (ours)7.921.537.754.069.138.122.076.4162.5171.9167.2
Dynamic-SiC (ours)7.121.337.453.568.037.521.776.0150.1151.7151.0
+ +![](images/0e4bd665dde05fc600a9f35b9eaf5524175ba4a79c5d194fd2f5df0034412678.jpg) +Figure 5. Comparison of generalization. Our SiC completes the missing skeletons according to the customized prompt for an unseen task. + +Copy Method. Following [5, 16], we copy the prompt target $T_{i}^{k}$ as prediction results for comparison. + +Multi-Task-Capable Models. For task-specific models [41, 52, 65] and multi-stage models [77], they are not able to handle multiple tasks at the same time. Therefore, we remodel them into end-to-end multi-task-capable models to achieve fair comparison. + +In-Context Models. Since there are no in-context models excelling in processing multiple human perception tasks, we select the most relative in-context models with our benchmark, Point-In-Context [16] and PointMAE [43], which is proposed in 3D point cloud understanding. Specifically, we take each frame as a token in their framework. For PointMAE, we use the transformed version in [16]. + +# 5.2. Compared with SoTA Methods + +Quantitative Results. We report experimental results on four tasks in Tab. 1. We term Skeleton-in-Context implemented with static/dynamic-TUP as Static/Dynamic-SiC, respectively. For other models, we reproduce them and report the results. As can be seen from Tab. 1, our proposed model outperforms all multi-task models, achieving state-of-the-art results on the in-context multi-task benchmark. Notably, the results of task-specific models on a single task reflect the difficulty of each task. However, we also exhibit impressive + +performance compared to task-specific models, demonstrating that our proposed model is an expert in handling skeleton sequence-based multi-tasking. + +Qualitative Results. We visualize and compare the results of our SiC and the most recent SoTA model, MotionBERT [77], which is re-structured as an end-to-end multi-task model for fair comparison. As highlighted in Fig. 4, our SiC can generate more accurate poses than MotionBERT according to the provided task-guided prompt. + +# 5.3. Generalization Capability. + +Generalize to New Datasets. To verify the generalization ability of our model, we conduct cross-dataset experiments on a specific task. For instance, our proposed model learns motion prediction on the AMASS dataset under the default benchmark, and then it is used to predict future motion sequences on 3DPW during testing. Similarly, we set up four experiments: motion prediction (AMASS $\rightarrow$ 3DPW), pose estimation (H3.6M $\rightarrow$ AMASS, 3DPW), and joint completion (3DPW $\rightarrow$ H3.6M). As shown in Tab. 2, our model is able to generalize the learned task well to other datasets, while other task-specific models cannot accommodate the large bridge introduced across datasets even if they are specifically trained on this single task. Additionally, benefiting from the dynamic pseudo pose, the Dynamic-SiC exhibits + +Table 3. Effectiveness of the Skeleton Prompt. + +
#TGPTUPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1XX111.7159.2543.7160.2
2✓X25.052.537.765.2
3X✓26.052.437.878.4
4✓✓24.451.836.862.9
+ +Table 4. Different Prompting (TGP). + +
#TGPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1Feature Sim.25.252.137.764.1
2Motion Sim.25.252.137.864.1
3Random24.451.836.862.9
+ +more excellent generalization ability than the Static-SiC. + +Generalize to New Task. Furthermore, We construct a prompt that has not appeared in the training set, which performs motion in-between. As Fig. 5 shows, our trained model can complete the missing skeletons according to the customized task-guided prompt. Our SiC generalizes well to a new task, which is unexplored in other works. + +# 5.4. Ablation Study + +Effectiveness of the Skeleton Prompt. We conduct ablation experiments on our proposed skeleton prompts. As shown in Tab. 3, the skeleton prompts solve the problem of directly extending previous work to skeleton sequences, which is falling into over-fitting. Meanwhile, applying both TGP and TUP to the model can achieve the best results. Our proposed skeleton prompts replace the random masking training strategy used in previous work. In particular, TUP creates a training environment that does not require the query target, which ensures that the model will not see the ground truth of the query input in advance. + +Prompt Engineering. Since previous work has proven that different prompts in context learning affect the performance of the model, we conduct ablation experiments on the choice of prompts. As Tab. 4 shows, we select the prompt by the feature similarity and MPJPE minimum between the input skeleton sequence and the prompt, respectively. Experimental results show that the random selection method achieves the best results. Unlike 2D images and 3D point clouds, skeleton-based task data is highly similar, resulting in different prompts playing similar roles in the same task. While the random selection method expands the model's receptive field to the data and achieves better results. + +Architecture. We conduct ablation experiments on the architecture of the model. As shown in Tab. 5, after exploring different hidden feature dimensions and transformer depths, we find that the model performs best when the feature dimension is 256, and the depth is 5. + +# 5.5. In-Context Initial Pose + +We conduct an in-depth exploration of Dynamic-TUP. We extract the trained Dynamic-TUP weights, then randomly initialize a $(F,J,3)$ tensor and use the trained encoder to in + +Table 5. Ablation on model architecture. + +
Arch.Num.MP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
Dim.6429.357.145.880.5
12826.453.839.266.1
25624.451.836.862.9
51224.750.737.663.0
Dep.325.153.739.265.6
425.152.238.664.7
524.451.836.862.9
624.752.938.163.7
(a) Task-Unified Prompt (Static)
(b) Task-Unified Prompt (Dynamic)
+ +Figure 6. (a) The visualization of Static-TUP. (b) The concrete visual skeleton sequence of Dynamic-TUP. + +crease its feature dimension to C-dimension. Finally, an optimizer is used to optimize the tensor and minimize its feature similarity with the trained Dynamic-TUP. The top of Fig. 6 represents the visualization of Static-TUP, while the bottom is the Dynamic-TUP. The concrete skeleton of dynamic-TUP is similar to the mean pose Static-TUP of the training set, which demonstrates that the Dynamic-TUP learned by our model is actually the prior knowledge whose receptive field is all ground truth in the training set. This prior knowledge helps our model learn more smoothly on multiple tasks, and adding learnable modules helps our model generalize to other datasets or new tasks (Sec. 5.3). + +# 6. Conclusion + +In this work, we propose the Skeleton-in-Context, designed to process multiple skeleton-base tasks simultaneously after just one training time. Specifically, we build a novel skeleton-based in-context benchmark covering various tasks. In particular, we propose skeleton prompts composed of TGP and TUP, which solve the overfitting problem of skeleton sequence data trained under the training framework commonly applied in previous 2D and 3D in-context models. Besides, we demonstrate that our model can generalize to different datasets and new tasks, such as motion completion. We hope our research builds the first step in the exploration of in-context learning for skeleton-based sequences, which paves the way for further research in this area. + +Acknowledgement This work is supported by National Natural Science Foundation of China (No. 62203476), Natural Science Foundation of Shenzhen (No. JCYJ20230807120801002). This project is also supported by the interdisciplinary doctoral grants (iDoc 2021-360) from the Personalized Health and Related Technologies (PHRT) of the ETH domain, Switzerland. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 2 +[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models, 2023. 3 +[3] Ivana Balažević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Hénaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 2, 3 +[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2 +[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 2022. 2, 3, 4, 5, 7 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 1, 2 +[7] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In ECCV, 2018. 2 +[8] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 2 +[9] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In ICCV, 2021. 1, 3 +[10] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, 2020. 2 +[11] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In ICCV, 2019. 2 +[12] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. Unihcp: A unified model for human-centric perceptions. In CVPR, 2023. 2, 3 +[13] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In CVPR, 2020. 2 +[14] Zhichao Deng, Xiangtai Li, Xia Li, Yunhai Tong, Shen Zhao, and Mengyuan Liu. Vg4d: Vision-language model goes 4d video recognition. In ICRA, 2024. 2 +[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[16] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context + +learning for 3d point cloud understanding. arXiv preprint arXiv:2306.08659, 2023. 1, 2, 3, 4, 5, 7 +[17] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In CVPR, 2023. 1, 2, 3 +[18] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 2, 3 +[19] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 5 +[20] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. TOG, 2020. 2 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3, 4 +[22] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. TOG, 2020. 2 +[23] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In ICCV, 2019. 2 +[24] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. PAMI, 2013. 3, 5 +[25] Mahdi M Kalayeh, Emrah Basaran, Muhittin Gokmen, Mustafa E Kamasak, and Mubarak Shah. Human semantic parsing for person re-identification. In CVPR, 2018. 3 +[26] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 3DV, 2020. 2 +[27] Sameh Khamis, Cheng-Hao Kuo, Vivek K Singh, Vinay D Shet, and Larry S Davis. Joint learning for attribute-consistent person re-identification. In ECCV, 2015. 3 +[28] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In CVPR), 2018. 2 +[29] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Guangliang Cheng, Pang Jiangmiao, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv pre-print, 2023. 2 +[30] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into person: Joint body parsing & pose estimation network and a new benchmark. PAMI, 2018. 3 +[31] Chunze Lin, Jiwen Lu, and Jie Zhou. Multi-grained deep feature learning for pedestrian detection. In ICME, 2018. 3 +[32] Jun Liu, Henghui Ding, Amir Shahroudy, Ling-Yu Duan, Xudong Jiang, Gang Wang, and Alex C Kot. Feature boosting network for 3d pose estimation. PAMI, 2019. 2, 3 +[33] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. PAMI, 2019. 3 + +[34] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In ECCV, 2020. 2, 3 +[35] Diogo C Luvizon, David Picard, and Hedi Tabia. 2d/3d pose estimation and action recognition using multitask deep learning. In CVPR, 2018. 3 +[36] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 1, 3 +[37] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3 +[38] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2, 3 +[39] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2 +[40] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2 +[41] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 5, 7 +[42] Xuecheng Nie, Jiashi Feng, and Shuicheng Yan. Mutual learning to adapt for joint human parsing and pose estimation. In ECCV, 2018. 3 +[43] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3, 4, 5, 7 +[44] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3 +[46] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633, 2021. 2 +[47] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 2021. 2 +[48] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu $\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 3 +[49] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In ECCV, 2022. 2 +[50] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry Steven Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for multi-camera person re-identification. PAMI, 2017. 3 + +[51] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023. 2, 3 +[52] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatio-temporal criss-cross attention. In CVPR, 2023. 5, 7 +[53] Yonglong Tian, Ping Luo, Xiaogang Wang, and Xiaou Tang. Pedestrian detection aided by deep learning semantic tasks. In CVPR, 2015. 3 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3 +[55] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3 +[56] Chaoyang Wang, Xiangtai Li, Henghui Ding, Lu Qi, Jiangning Zhang, Yunhai Tong, Chen Change Loy, and Shuicheng Yan. Explore in-context segmentation via latent diffusion models. arXiv, 2024. 2 +[57] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 2 +[58] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4 +[59] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, 2023. 2, 3 +[60] Xinshun Wang, Qiongjie Cui, Chen Chen, and Mengyuan Liu. Gcnext: Towards the unity of graph convolutions for human motion prediction. In AAAI, 2024. 2 +[61] Xinshun Wang, Wanying Zhang, Can Wang, Yuan Gao, and Mengyuan Liu. Dynamic dense graph convolutional network for skeleton-based human motion prediction. TIP, 2024. 2 +[62] Zhendong Wang, Yifan Jiang, Yadong Lu, Yelong Shen, Pengcheng He, Weizhu Chen, Zhangyang Wang, and Mingyuan Zhou. In-context learning unlocked for diffusion models. arXiv preprint arXiv:2305.01115, 2023. 2 +[63] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 2 +[64] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023. 2, 3 +[65] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 5, 7 +[66] Shilin Xu, Xiangtai Li, Jingbo Wang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Fashionformer: A simple, effective and unified baseline for human fashion segmentation and recognition. In ECCV, 2022. 3 +[67] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In CVPR, 2021. 2 + +[68] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In ICCV, 2019. 2 +[69] Angela Yao, Juergen Gall, and Luc Van Gool. Coupled action recognition and pose estimation from multiple views. IJCV, 2012. 3 +[70] Bruce XB Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In ICCV, 2023. 5 +[71] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2 +[72] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670, 2023. 2, 3 +[73] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In CVPR, 2022. 2 +[74] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. +[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 2 +[76] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, et al. Generative tweening: Long-term inbetweening of 3d human motions. arXiv preprint arXiv:2005.08891, 2020.2 +[77] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In ICCV, 2023. 1, 3, 5, 6, 7 \ No newline at end of file diff --git a/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/images.zip b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..133956801854a39feecd22e55263c9e74f069c16 --- /dev/null +++ b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ae7e5ebadf017623ecf121dd38904be7273337b720ab99a4452fdc3d4d314f6 +size 799914 diff --git a/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/layout.json b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05b20c663786d4406dfe4075faf595cf20e04379 --- /dev/null +++ b/2024/Skeleton-in-Context_ Unified Skeleton Sequence Modeling with In-Context Learning/layout.json @@ -0,0 +1,8765 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 139 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 139 + ], + "type": "text", + "content": "Skeleton-in-Context: Unified Skeleton Sequence Modeling with In-Context Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "spans": [ + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "content": "Xinshun Wang" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "content": ", Zhongbin Fang" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "content": ", Xia Li" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "content": ", Xiangtai Li" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "content": ", Chen Chen" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "content": ", Mengyuan Liu" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "inline_equation", + "content": "^{2, \\boxtimes}" + }, + { + "bbox": [ + 71, + 149, + 522, + 178 + ], + "type": "text", + "content": ", Sun Yat-sen University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 180, + 489, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 180, + 489, + 193 + ], + "spans": [ + { + "bbox": [ + 102, + 180, + 489, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 180, + 489, + 193 + ], + "type": "text", + "content": "National Key Laboratory of General Artificial Intelligence, Peking University, Shenzhen Graduate School" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 194, + 463, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 194, + 463, + 207 + ], + "spans": [ + { + "bbox": [ + 129, + 194, + 463, + 207 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 129, + 194, + 463, + 207 + ], + "type": "text", + "content": "Department of Computer Science, ETH Zurich " + }, + { + "bbox": [ + 129, + 194, + 463, + 207 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 129, + 194, + 463, + 207 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 166, + 209, + 426, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 209, + 426, + 220 + ], + "spans": [ + { + "bbox": [ + 166, + 209, + 426, + 220 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 166, + 209, + 426, + 220 + ], + "type": "text", + "content": "Center for Research in Computer Vision, University of Central Florida" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 223, + 515, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 223, + 515, + 236 + ], + "spans": [ + { + "bbox": [ + 78, + 223, + 515, + 236 + ], + "type": "text", + "content": "{wangxsh36,fangzhb5}@mail2.sysu.edu.cn，xia.li@inf.ethz.ch，xiangtai94@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 170, + 238, + 419, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 238, + 419, + 248 + ], + "spans": [ + { + "bbox": [ + 170, + 238, + 419, + 248 + ], + "type": "text", + "content": "chen.chen@crcv.ucf.edu, liumengyuan@pku.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 164, + 253, + 425, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 253, + 425, + 264 + ], + "spans": [ + { + "bbox": [ + 164, + 253, + 425, + 264 + ], + "type": "text", + "content": "https://github.com/fanglaosi/Skeleton-in-Context" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 285, + 192, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 285, + 192, + 297 + ], + "spans": [ + { + "bbox": [ + 143, + 285, + 192, + 297 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 300, + 290, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 290, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 290, + 635 + ], + "type": "text", + "content": "In-context learning provides a new perspective for multi-task modeling for vision and NLP. Under this setting, the model can perceive tasks from prompts and accomplish them without any extra task-specific head predictions or model fine-tuning. However, skeleton sequence modeling via in-context learning remains unexplored. Directly applying existing in-context models from other areas onto skeleton sequences fails due to the similarity between inter-frame and cross-task poses, which makes it exceptionally hard to perceive the task correctly from a subtle context. To address this challenge, we propose Skeleton-in-Context (SiC), an effective framework for in-context skeleton sequence modeling. Our SiC is able to handle multiple skeleton-based tasks simultaneously after a single training process and accomplish each task from context according to the given prompt. It can further generalize to new, unseen tasks according to customized prompts. To facilitate context perception, we additionally propose a task-unified prompt, which adaptively learns tasks of different natures, such as partial joint-level generation, sequence-level prediction, or 2D-to-3D motion prediction. We conduct extensive experiments to evaluate the effectiveness of our SiC on multiple tasks, including motion prediction, pose estimation, joint completion, and future pose estimation. We also evaluate its generalization capability on unseen tasks such as motion-in-between. These experiments show that our model achieves state-of-the-art multi-task performance and even outperforms single-task methods on certain tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 658, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 288, + 696 + ], + "type": "text", + "content": "The idea of multi-task models has attracted much research interest over the years in a wide range of areas [6, 16, 17, 58], such as computer vision and natural language processing" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 284, + 541, + 351 + ], + "blocks": [ + { + "bbox": [ + 309, + 284, + 541, + 351 + ], + "lines": [ + { + "bbox": [ + 309, + 284, + 541, + 351 + ], + "spans": [ + { + "bbox": [ + 309, + 284, + 541, + 351 + ], + "type": "image", + "image_path": "50b983fa9eab11369c47a0eaa679044708a3e891ab5e57c82a5c2dc72a9227ed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 354, + 542, + 426 + ], + "blocks": [ + { + "bbox": [ + 308, + 354, + 542, + 426 + ], + "lines": [ + { + "bbox": [ + 308, + 354, + 542, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 542, + 426 + ], + "type": "image", + "image_path": "b2de66d6fc38ae2c4f4e17186707ca7dd9c91d6dff9c04e373f646dc75e83679.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 428, + 541, + 550 + ], + "blocks": [ + { + "bbox": [ + 308, + 428, + 541, + 550 + ], + "lines": [ + { + "bbox": [ + 308, + 428, + 541, + 550 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 541, + 550 + ], + "type": "image", + "image_path": "5f8b91c68bb67cb337a07a1263d4fa0a79c3d915a703f9d361b17a941f59c566.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 554, + 545, + 576 + ], + "lines": [ + { + "bbox": [ + 305, + 554, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 545, + 576 + ], + "type": "text", + "content": "Figure 1. In-context learning in a) image modeling [58], b) point cloud modeling [16], and c) skeleton sequence modeling (ours)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "(NLP). To build such models, most previous works focus on deploying task-specific heads on a shared backbone and learning them through pretraining and fine-tuning. Models with such task-specific architectures are not learned in an end-to-end manner and are hard to generalize to new, unseen tasks and data. Most existing methods for skeleton-based human-centric tasks rely on task-specific designs and/or non-end-to-end frameworks to accomplish multiple tasks. Several works [9, 36] learn unified human motion representations. In particular, MotionBERT [77] adopts two-stage training, including pre-training on the 2D-3D lifting task" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 271, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 271, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 271, + 713 + ], + "type": "text", + "content": "* Equal contribution. ☑ Corresponding author is Mengyuan Liu." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2436" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "content": "and fine-tuning on each task. UniHCP [12] connects a well-designed shared backbone network to specific task heads according to different tasks. Meanwhile, UPS [17] unifies action labels and motion sequences into text-based language sequences and fine-tunes them on each task for better performance. These models can have a limited scope of tasks they are capable of handling and lack flexibility. This limits their capability to generalize to unseen tasks and data. To overcome these challenges, in-context learning represents a new trend to build generalist models, which was originally proposed in NLP [6, 45, 46] and then introduced into other areas such as images and point clouds. Some generalist models [5, 58, 59, 72] utilize the Masked Image Modeling (MIM) framework to exhibit multi-task capability via in-context learning in image processing, as shown in Fig. 1(a). Similarly, in Fig. 1(b), Point-In-Context [16] proposes a key joint sampling module and introduces in-context learning into point cloud processing. Specifically, in-context models learn patterns of provided input-output pairs, called in-context examples or prompts, and then perform the desired task according to the prompt. An in-context model can accomplish diverse tasks end-to-end without any task-specific heads or fine-tuning. It also has excellent generalization capability. Directly adapting existing in-context models from other areas onto skeleton sequences fails because they largely focus on static objects (images or point clouds) and thus cannot capture spatial-temporal dependencies in skeleton sequences. To the best of our knowledge, in-context learning remains unexplored in 3D skeleton sequence modeling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 422, + 289, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 289, + 698 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 289, + 698 + ], + "type": "text", + "content": "In this paper, we propose Skeleton-in-Context (SiC), a novel framework for skeleton sequence modeling, which is able to process multiple tasks simultaneously after a single training process without any task-specific designs, as Fig. 1(c) shows. Specifically, SiC perceives tasks from context, i.e., the prompt from a desired task provided as additional input, and then accomplishes the task for the query. SiC employs two types of prompts to facilitate context perception: the task-guided prompt (TGP) and the task-unified prompt (TUP). TGP provides a task template for the model, giving the model the ability to make analogies. This is the key to the model's capability to generalize to new tasks, as we can guide the model to do unseen tasks by customizing specific prompts. TUP can be implemented in two ways: static and dynamic. The static TUP leverages prior knowledge shared among tasks, while the dynamic TUP adaptively learns to perceive task context. Skeleton-in-Context provides a new perspective for training a skeleton-based multi-task in-context model. Since there is no existing benchmark for measuring such a model, we establish a new in-context benchmark based on four tasks, including three skeleton-based tasks: motion prediction, pose estimation, and joint completion, and a new task: future pose estimation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "content": "Our main contributions are as follows: 1) We introduce" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "a novel framework, Skeleton-in-Context (SiC), designed to handle multiple skeleton-based tasks simultaneously. Our approach learns the potential pattern of the given prompt and performs it on the query sample. 2) We establish a new 3D human in-context learning benchmark comprised of multiple skeleton-based tasks, such as motion prediction, pose estimation, joint completion, and future motion estimation. We further benchmark several representative in-context learning baselines for reference. 3) We conduct extensive experiments to evaluate our SiC. SiC achieves state-of-the-art results in multi-tasking and even outperforms task-specific models on certain tasks. Moreover, SiC also generalizes well to unseen tasks such as motion in-between." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 239, + 392, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 392, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 392, + 251 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 258, + 548, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 548, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 548, + 533 + ], + "type": "text", + "content": "Skeleton-Based Motion Analysis. Many of the skeleton-based human-centric tasks have shared similar development patterns in terms of the methods used to address them. On human motion prediction, early successes are achieved with RNNs and CNNs [28, 40]. Nowadays, the task is dominated by Graph Convolutional Networks (GCNs). LTD [38] considers each joint coordinate as a node in a graph on which to employ learnable spatial graph convolutions. Follow-up works [13, 39, 60, 61] focus on designing better graph convolution networks. Different from motion prediction, the goal of motion synthesis is to complete the human motion sequence. To complete the missing parts of human motion, many works adopt convolutional models [22, 26, 68] or generative adversarial networks [7, 20, 23, 76]. One category of 3D pose estimation [18, 32, 34] is to lift the estimated 2D pose to 3D, which often employs spatial-temporal convolution [8, 10, 11, 41, 44, 57, 67] or transformer-based [14, 49, 71, 73-75] methods. However, these approaches focus on designing customized architectures specific to a task, which are limited in range of applications. To this end, we propose to unify the input and output formats of multiple tasks and utilize a single model to process multiple tasks simultaneously via in-context learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "content": "In-Context Learning. The emergence of in-context learning provides a new paradigm for achieving multitasking, which originated from Natural Language Processing (NLP) [1, 15, 45-47]. It endows models with the capability to handle multiple tasks and even unseen tasks by incorporating domain-specific input-target pairs, known as prompts, into the query sample [6]. Following NLP, in-context learning is extended to computer vision in some work, such as 2D images [51, 56, 58, 59, 62, 72] and 3D point clouds [3, 16, 64]. Specifically, Visual Prompt [5], Painter [58], and Point-In-Context [16] are pioneers of in-context methods in computer vision. They verify the effectiveness and flexibility of in-context learning in vision-based tasks, and most of them adopt pure-transformer-based approaches [4, 21, 29, 63]. Meanwhile, several approaches" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2437" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "deeply explore Prompt Engineering [3, 51, 72] and awe-inspiring generalization [16, 59, 64] in in-context learning. Recently, LVM [2] unifies most vision tasks and data via joint co-training on one visual in-context model. Compared to the works focused on static objects (images and point clouds), our work mainly delves into modeling in-context dynamic skeleton sequences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 290, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 290, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 290, + 373 + ], + "type": "text", + "content": "Unified Skeleton-Centric Models. A generalist model that can excel in handling multiple tasks is a profound step toward general artificial intelligence [45, 58]. Previous works [25, 27, 31, 35, 50, 53, 66, 69] give the effort to explore co-training several human-centric tasks. Meanwhile, some works [30, 42] show great progress in simultaneously predicting parsing and pose. Recent research [9, 36] demonstrates the effectiveness of modeling unified human motion sequences. In particular, UniHCP [12] designs specific task heads according to different 2D human-centric tasks. Motion-BERT [77] is pre-trained in the 2D-3D lifting task and fine-tuned on each task. UPS [17] unifies heterogeneous human behavior understanding tasks via language sequences and cycle training for each task in order. Nevertheless, these works are limited by the dilemma of complex training stages or task heads, which introduce extra costs. We integrate the output format of different tasks in skeleton sequence format and train the model just once." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 381, + 268, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 268, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 268, + 396 + ], + "type": "text", + "content": "3. In-Context Skeleton Sequence Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 401, + 129, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 129, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 129, + 412 + ], + "type": "text", + "content": "3.1. Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "text", + "content": "Since the skeleton sequence data in skeleton-based tasks are sequential by nature, we treat the skeleton at each frame as a token, which can be processed by transformer [54]. Following previous works [5, 16, 58], we curate an in-context skeleton sequence paradigm. Given an input-target skeleton sequence pair as a task-guided prompt " + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "inline_equation", + "content": "G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\}" + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "text", + "content": " and a query input " + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "inline_equation", + "content": "Q_{j}^{k}" + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "inline_equation", + "content": "I_{i}^{k}, T_{i}^{k}, Q_{j}^{k} \\in \\mathbb{R}^{F \\times J \\times 3}" + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "text", + "content": " each have " + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 419, + 289, + 552 + ], + "type": "text", + "content": " joints, the model is expected to perceive the context from the prompt, i.e., which task it represents, and perform the corresponding operations on the query in analogy to the prompt. This process can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 559, + 288, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 559, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 133, + 559, + 288, + 574 + ], + "type": "interline_equation", + "content": "y = f \\left(G _ {i} ^ {k}, Q _ {j} ^ {k}\\right), \\tag {1}", + "image_path": "0d750e76db3a14202a7be967a1ddf11707f76204efff7fb06d5edebf88ca00bc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "text", + "content": " represents the task index, while " + }, + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 580, + 288, + 629 + ], + "type": "text", + "content": " indicate different example indexes. Thus, we can determine what task will be performed on the query sample by selecting the task-specific prompt from the corresponding training set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 145, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 145, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 145, + 647 + ], + "type": "text", + "content": "3.2. Task Definitions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "content": "To measure multitasking capability, we establish a new in-context benchmark from common skeleton-based datasets, including H3.6M [24], AMASS [37], and 3DPW [55]. In this benchmark, we select three common tasks: motion prediction, pose estimation, and joint completion, and a new" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 203 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 203 + ], + "type": "image", + "image_path": "c3f6a43a5f06aa2a32c8a96539d22a1ca563999491ee6f59866ccf78f60081f0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 205, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 282 + ], + "type": "text", + "content": "Figure 2. Top: Training with MIM-style framework [21, 43] used in previous works [5, 16, 58]. The model is able to reconstruct the masked frames well. Bottom: During inference, the reconstructed sequence gradually shrinks the skeleton as time goes by because the model only learns frame interpolation during training. Once the model loses subsequent reference frames when interpolating frames, the generated sequence will tend to shrink." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 299, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 547, + 335 + ], + "type": "text", + "content": "task: future pose estimation. Consequently, we conduct extensive experiments on a large and diverse dataset collection, which consists of 176K skeleton sequence samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "text", + "content": "Motion Prediction. Motion prediction [38] is to predict future motion based on historical motion. There is continuity between the input and output sequences, which constitute a complete motion. Thus, the input " + }, + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "text", + "content": " are complete skeleton sequences with shapes of " + }, + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "inline_equation", + "content": "(F, J, 3)" + }, + { + "bbox": [ + 304, + 340, + 547, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "text", + "content": "Pose Estimation. Pose estimation [18, 32, 34] is to lift 2D poses to 3D. The model takes the detected 2D skeleton sequence " + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{F \\times J \\times 2}" + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "text", + "content": ", which contain noise, as input, and estimates the accurate 3D skeleton sequence " + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{F \\times J \\times 3}" + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "text", + "content": ". Note that we expand the 2D input to 3D by filling the 3rd dimension with 0, in order to unify the data format with other tasks. Therefore, the input skeleton sequence becomes the same shape " + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "inline_equation", + "content": "(F, J, 3)" + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "text", + "content": " as the target sequence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "text", + "content": "Future Pose Estimation. Based on prediction and estimation, we propose a new task: future pose estimation, which is to predict future 3D motion based on history 2D poses. As 2D images are easier to obtain than 3D data, this task unlocks a host of new real-world applications. Given a historical 2D skeleton sequence " + }, + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{F \\times J \\times 3}" + }, + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "text", + "content": " (one dimension with 0), future pose estimation expects the model to output its future 3D skeleton sequence " + }, + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{F \\times J \\times 3}" + }, + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "text", + "content": ". To our knowledge, no work has ever proposed future pose estimation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Joint Completion. Joint completion is to recover skeleton sequences that have missing joints. This task can reflect cases where joints may be missing or displaced when collected with Lidar cameras [33, 48]. We set two levels to verify the ability to complete joints, which are " + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": " of the joints missing in each frame. Also, the coordinates of missing joints are replaced with 0, with the goal of unifying the input skeleton sequence format with " + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "inline_equation", + "content": "(F, J, 3)" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2438" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 534, + 222 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 534, + 222 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 534, + 222 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 534, + 222 + ], + "type": "image", + "image_path": "0ec1efdc4867b9d4bb6239732dfa04df9a7098f76d6928d5cc6b1799ab26928c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "type": "text", + "content": "Figure 3. Overall framework of our Skeleton-in-Context. Specifically, we establish a skeleton bank by integrating training sets under different tasks, which contain a large amount of input-target pairs performing different tasks. Next, we randomly select a sample pair as the task-guided prompt (TGP) and a query input from the skeleton bank, undergo encoding and concatenating, respectively, and then input them into the transformer in parallel. In particular, during this process, the query input and task-unified prompt (TUP) are combined to form a new query. After iterating " + }, + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "type": "text", + "content": " times, the TGP and query are aggregated through " + }, + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "type": "text", + "content": " and then input into the transformer for " + }, + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 46, + 226, + 548, + 295 + ], + "type": "text", + "content": " iterations. Lastly, the second half of the model output is used as our prediction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 308, + 167, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 167, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 167, + 320 + ], + "type": "text", + "content": "4. Skeleton-in-Context" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 328, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 288, + 437 + ], + "type": "text", + "content": "In-context learning on other data modalities such as images and point clouds [5, 16, 58] largely employ Mask Image Modeling (MIM) [21] to reconstruct the randomly masked elements. During inference, only the target elements are masked. On skeleton sequences, MIM fails due to spatial-temporal complexity and inter-pose similarity, which makes it outstandingly hard to perceive the task correctly from a subtle context. In our Skeleton-in Context, we employ a new framework to address this challenge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 443, + 195, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 195, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 195, + 456 + ], + "type": "text", + "content": "4.1. Skeleton Sequence Prompt" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 593 + ], + "type": "text", + "content": "Task-Guided Prompt. In in-context learning, a prompt is needed for the model to learn the underlying task pattern from context. As mentioned in Sec. 3.1, we propose the Task-Guided Prompt (TGP) " + }, + { + "bbox": [ + 46, + 462, + 287, + 593 + ], + "type": "inline_equation", + "content": "G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\}" + }, + { + "bbox": [ + 46, + 462, + 287, + 593 + ], + "type": "text", + "content": ", which comprises of an example skeleton pair performing a specific task " + }, + { + "bbox": [ + 46, + 462, + 287, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 462, + 287, + 593 + ], + "type": "text", + "content": ". In this way, our proposed model is able to perform a variety of tasks according to the provided prompts that could be seen or even unseen in the training set. It is worth noting that our model can perform unseen tasks and transfer to datasets that are different styles from the training dataset. It will be elaborated in Sec. 5.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Avoiding Over-Fitting. Previous works have made great progress in unifying different tasks via the MIM-style framework [5, 16, 58]. Compared with static objects in 2D and 3D, such as images and point clouds, the obstacle to in-context learning of skeleton sequences lies in the wondrous similarity of each frame, so directly using the random masking strategy [5, 16, 58] will cause the model to fall into overfitting, leading it to only learn simple interpolation frames rather than capturing meaningful motion sequence representations during the training stage. As shown in Fig. 2, when" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 308, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 416 + ], + "type": "text", + "content": "we directly utilize the MPM-style [21, 43] framework to train our model, although it can perfectly reconstruct the masked frame during training, the experimental results reveal that it outputs a devastating result during inference. It is obvious that the output skeleton sequence gradually shrinks as the length increases because the next frame of the current frame is 0, and interpolating with 0 will lead to shrinking. Such a phenomenon underscores the challenges of learning contextual information about dynamic skeleton sequences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 417, + 547, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 547, + 571 + ], + "type": "text", + "content": "Task-Unified Prompt. With the Task-Guided Prompt, which exhibits unique capabilities specific to in-domain or out-of-domain tasks, mitigating the gap between different tasks presents a new challenge. Meanwhile, as discussed above, a straight extension of previous works will lead to a simple model that can only perform the interpolation task based on the given skeleton sequence. To tackle these challenges, we introduce the Task-Unified Prompt (TUP), a well-designed and task-agnostic module, which can adaptively learn to incorporate multiple tasks into the framework of in-context learning, i.e., a one-off end-to-end framework for all tasks without having to deploy task-specific heads or fine-tuning. We introduce two versions of TUP:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 572, + 514, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 572, + 514, + 584 + ], + "spans": [ + { + "bbox": [ + 317, + 572, + 514, + 584 + ], + "type": "text", + "content": "Version 1: Static-TUP with static pseudo pose:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 592, + 545, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 592, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 347, + 592, + 545, + 625 + ], + "type": "interline_equation", + "content": "U ^ {S} = \\mathcal {E} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\sum_ {k = 1} ^ {K} \\left(T _ {m} ^ {k}\\right) \\in \\mathbb {R} ^ {F \\times J \\times 1}\\right), \\tag {2}", + "image_path": "97e2476b90751392d609663504b87b8cf99cc56dc454de62a92e01cd8e19ed6e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "text", + "content": " is a embedding layer. " + }, + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 633, + 545, + 658 + ], + "type": "text", + "content": " indicate the total number of training samples and tasks, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 658, + 539, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 539, + 670 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 539, + 670 + ], + "type": "text", + "content": "Version 2: Dynamic-TUP with dynamic pseudo pose:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 398, + 678, + 545, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 678, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 398, + 678, + 545, + 692 + ], + "type": "interline_equation", + "content": "U ^ {D} = W _ {1} W _ {2}, \\tag {3}", + "image_path": "466b87e915d83af9b5e8cde98e5e96ab8a2047f5a45df5c7367a6c761c1d225e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 700, + 467, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 467, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 467, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 700, + 467, + 713 + ], + "type": "inline_equation", + "content": "W_{1}\\in \\mathbb{R}^{F\\times J\\times 1}" + }, + { + "bbox": [ + 306, + 700, + 467, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 700, + 467, + 713 + ], + "type": "inline_equation", + "content": "W_{2}\\in \\mathbb{R}^{1\\times 1\\times C}" + }, + { + "bbox": [ + 306, + 700, + 467, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2439" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "content": "Table 1. Comparison of task-specific models, multi-task models, and in-context models on four human skeleton tasks. For Motion Prediction (MP.), Joint Completion (JC.), and Future Pose Estimation (FPE.), we report Mean Per Joint Position Error (mm) [24]. For Pose Estimation (PE.), we additionally report another indicator, Normalized Mean Per Joint Position Error (N-MPJPE) [77]. For joint completion, we report two settings: " + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "content": " of the joints missing in each frame. " + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "content": " represents the multi-task-capable models re-structured from task-specific models. " + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "content": " represents the multi-task-capable models re-structured from multi-stage models. Static-SiC and Dynamic-SiC represent the static and dynamic versions of TUP, respectively." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 141, + 523, + 357 + ], + "blocks": [ + { + "bbox": [ + 69, + 141, + 523, + 357 + ], + "lines": [ + { + "bbox": [ + 69, + 141, + 523, + 357 + ], + "spans": [ + { + "bbox": [ + 69, + 141, + 523, + 357 + ], + "type": "table", + "html": "
MethodsVenuesMP. (AMASS)PE. (H3.6M)JC. (3DPW)FPE. (H3.6M)
80ms160ms200ms320ms400msAvg.MPJPE↓ Avg.N-MPJPE↓ Avg.MPJPE↓ 40%60% Avg.MPJPE↓ 200ms300ms Avg.
Task-Specific Model: one architecture for one task
siMLPe [19]WACV'2312.121.625.434.439.926.758.949.748.458.553.575.4
EqMotion [65]CVPR'2314.426.631.744.050.033.363.259.340.449.745.169.5
STCFormaler [52]CVPR'2319.525.128.235.841.528.754.645.231.246.738.960.9
GLA-GCN [70]ICCV'2319.525.128.235.342.030.049.640.462.272.567.449.5
MotionBERT [77]ICCV'2312.921.425.536.342.326.749.441.129.344.336.875.5
Multi-Task Model: one architecture (w. multiple task-specific heads) for multiple tasks
siMLPe†[41]WACV'2319.030.035.346.852.436.774.660.059.067.463.264.4
EqMotion†[65]CVPR'2314.025.330.240.546.631.3163.3110.138.956.047.588.1
STCFormaler†[52]CVPR'2317.428.031.543.040.433.955.645.633.650.642.162.2
GLA-GCN†[70]ICCV'2335.339.842.349.256.144.578.662.974.181.978.060.6
MotionBERT‡[77]ICCV'2314.322.826.736.441.728.455.845.731.346.739.075.5
In-Context Model: one task-agnostic architecture for multiple tasks
Copy [5, 16]-134.7133.7133.3132.1131.4133.0351.0171.6196.5195.5196.0316.8
PointMAE [43]ECCV'2219.131.236.654.967.741.9144.396.4167.679.073.3139.9
PIC [16]NeurIPS'2325.040.246.867.279.451.7181.6118.885.796.491.1173.1
Static-SiC (ours)-11.019.022.732.637.824.651.641.829.844.737.259.8
Dynamic-SiC (ours)-11.018.922.632.237.424.451.842.529.544.036.859.2
", + "image_path": "b4e362122d937656b351b2e1d94fc126870d52633c0a303b6e8fbd52855833ed.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 362, + 290, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 290, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 290, + 422 + ], + "type": "text", + "content": "We denote the SiC-based on Static-/Dynamic-TUP is referred to as Static-/Dynamic-SiC, respectively. The proposed TUP facilitates the modeling of distribution differences among tasks and datasets, so the model focuses more on learning unified representations shared across tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 429, + 162, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 162, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 162, + 441 + ], + "type": "text", + "content": "4.2. Model Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "content": "As shown in Fig. 3, we build a skeleton bank, a collection of publicly available datasets including H3.6M, AMASS and 3DPW, covering common skeleton analysis tasks, containing input-target pairs formed as sequences of equal length, where prompts are randomly selected. Given a prompt pair " + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "inline_equation", + "content": "G_{i}^{k} = \\{I_{i}^{k}, T_{i}^{k}\\}" + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "content": " and a query sample " + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "inline_equation", + "content": "Q_{j}^{k}" + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "content": ", we first map them to the C-dimensional feature space, and then splice " + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 447, + 290, + 568 + ], + "type": "text", + "content": " respectively. The former represents our proposed Task-Guided Prompt, and the latter is the objective sequence of our focused analysis. The process can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 574, + 288, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 574, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 114, + 574, + 288, + 594 + ], + "type": "interline_equation", + "content": "\\widetilde {G} = \\mathcal {E} \\left(G _ {i} ^ {k}\\right) \\in \\mathbb {R} ^ {2 F \\times J \\times C}, \\tag {4}", + "image_path": "45200d898ef8196ec1b17327b834c56f3636f0b2f6a35f163cec65fe64e752b0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 606, + 287, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 606, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 101, + 606, + 287, + 626 + ], + "type": "interline_equation", + "content": "\\widetilde {Q} = \\left(\\mathcal {E} \\left(Q _ {j} ^ {k}\\right) \\| U\\right) \\in \\mathbb {R} ^ {2 F \\times J \\times C}, \\tag {5}", + "image_path": "a1a1bef17d9030e0e3e18f5bf6f6aa47e4f60e4c569ed63e3c32ee741731187c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " is a linear transformation and " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " is concatenating operator. In essence, " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " represents the prior knowledge of the target skeleton sequence and is also an adapter that adapts to different tasks and different distributions of data for learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Drawing inspirations from [77], we connect spatial attention blocks and temporal attention blocks together in an alternating combination manner to form a two-stream" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 362, + 547, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 547, + 387 + ], + "type": "text", + "content": "transformer " + }, + { + "bbox": [ + 305, + 362, + 547, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot)" + }, + { + "bbox": [ + 305, + 362, + 547, + 387 + ], + "type": "text", + "content": " dedicated to capturing the dynamic characteristics of motion, which is implemented as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 394, + 546, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 394, + 546, + 416 + ], + "spans": [ + { + "bbox": [ + 319, + 394, + 546, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (\\mathrm {X}) = \\alpha^ {i} \\mathcal {T} _ {1} ^ {i} \\left(\\mathcal {S} _ {1} ^ {i} \\left(\\mathrm {X} ^ {i - 1}\\right)\\right) + \\beta^ {i} \\mathcal {S} _ {2} ^ {i} \\left(\\mathcal {T} _ {2} ^ {i} \\left(\\mathrm {X} ^ {i - 1}\\right)\\right), \\tag {6}", + "image_path": "cba770052b39e27cccd2582563262b949434156c6541e38981d2098523a7ed32.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "text", + "content": " are balanced parameters determined adaptively. " + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "inline_equation", + "content": "S_{1}^{i}" + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1^i" + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "text", + "content": " are implemented using spatial selfattention mechanism and temporal self-attention mechanism, respectively. Additionally, " + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "inline_equation", + "content": "i = 1,\\dots,N" + }, + { + "bbox": [ + 304, + 422, + 547, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot)" + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{G}" + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{Q}" + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": " as input and parallel iterates over them " + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": " times respectively. Then, " + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{G}^{n1}" + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{Q}^{n1}" + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": " are integrated via aggregation function " + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 304, + 469, + 547, + 506 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 379, + 514, + 546, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 514, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 379, + 514, + 546, + 533 + ], + "type": "interline_equation", + "content": "W ^ {n 1} = \\mathcal {A} \\left(\\widetilde {G} ^ {n 1}, \\widetilde {Q} ^ {n 1}\\right), \\tag {7}", + "image_path": "032b76ff19f180edfaad69737c39db533e281d90dd3cd75ad9ba707c7026c061.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "text", + "content": "we consider " + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "text", + "content": " to be a simple weighted summation function. The final output is obtained after iterating " + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "text", + "content": " times in " + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot)" + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "inline_equation", + "content": "n_1 + n_2 = N" + }, + { + "bbox": [ + 304, + 540, + 547, + 590 + ], + "type": "text", + "content": ". Note that we take the second half of the output as the prediction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 595, + 437, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 437, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 437, + 609 + ], + "type": "text", + "content": "4.3. Training and Inference" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 613, + 547, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 547, + 684 + ], + "type": "text", + "content": "Training. We train our model in our proposed skeleton-centric in-context dataset mentioned in Sec. 3. In this training way, our Skeleton-in-Context can perform multiple tasks after one training. Since our output format is skeleton-like, following previous work [77], we adopt the 3D joint-level reconstruction loss:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 684, + 546, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 684, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 360, + 684, + 546, + 715 + ], + "type": "interline_equation", + "content": "L o s s = \\sum_ {f = 1} ^ {F} \\sum_ {j = 1} ^ {J} \\left\\| P _ {f, j} - G _ {f, j} \\right\\| _ {2}, \\tag {8}", + "image_path": "d07c340cfc3fe23b8a4a9b52557303de45c1aac997cbea38ee0575ee227f274c.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2440" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 71, + 519, + 507 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 519, + 507 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 519, + 507 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 519, + 507 + ], + "type": "image", + "image_path": "e88b61c3309694e53b4db3a976de1f3633df3f7143bed116c9f205a0e50dbafc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 514, + 547, + 537 + ], + "lines": [ + { + "bbox": [ + 47, + 514, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 547, + 537 + ], + "type": "text", + "content": "Figure 4. The comparison of visual results between our Skeleton-in-Context and recent SoTA method MotionBERT [77] on four tasks. Our method generates more accurate poses than MotionBERT, as can be seen from the spots with red arrows " + }, + { + "bbox": [ + 47, + 514, + 547, + 537 + ], + "type": "inline_equation", + "content": "(\\searrow)" + }, + { + "bbox": [ + 47, + 514, + 547, + 537 + ], + "type": "text", + "content": " pointed to them." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 545, + 288, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 288, + 606 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 288, + 606 + ], + "type": "text", + "content": "Inference. After a training phase involving all tasks, we get an in-context model that focuses on dynamic skeleton sequences. It can perform multiple tasks at the same time according to the provided task-dependent prompt without any further parameter updates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": "Implementation Details. We implement the proposed Skeleton-in-Context model with the number of layers " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": ", number of attention heads " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "H = 8" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": ", and hidden feature dimension " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": ". For each prompt input/target and query input/target, the sequence length is " + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "F = 16" + }, + { + "bbox": [ + 47, + 641, + 288, + 715 + ], + "type": "text", + "content": ". All tasks are unified into a one-off, end-to-end training process without" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 543, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 566 + ], + "type": "text", + "content": "any task-specific designs. The training takes about 6 hours for 120 epochs on 4 NVIDIA GeForce RTX 4090 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 567, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 626 + ], + "type": "text", + "content": "Datasets and Metrics. H3.6M (Human3.6M) is used for the pose estimation and future pose estimation tasks. AMASS is used for the motion prediction task. 3DPW (3D Pose in the Wild) is used for the joint completion task. All of the tasks can be evaluated using Mean Per Joint Position Error." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 635, + 411, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 411, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 411, + 647 + ], + "type": "text", + "content": "5.1. Baseline Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "In order to make a qualitative comparison with other methods in a relatively fair situation, we reimplement the task-specific SoTA methods in our proposed benchmark. Additionally, we transform recent methods into multi-task models and compare them with our model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2441" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 87, + 529, + 187 + ], + "blocks": [ + { + "bbox": [ + 46, + 62, + 545, + 84 + ], + "lines": [ + { + "bbox": [ + 46, + 62, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 46, + 62, + 545, + 84 + ], + "type": "text", + "content": "Table 2. Comparison results between task-specific models and our proposed models on generalization testing experiments on different datasets. We report Mean Per Joint Position Error (mm) for all tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 87, + 529, + 187 + ], + "lines": [ + { + "bbox": [ + 62, + 87, + 529, + 187 + ], + "spans": [ + { + "bbox": [ + 62, + 87, + 529, + 187 + ], + "type": "table", + "html": "
MethodsMP. (3DPW)PE.JC. (H3.6M)
40ms80ms120ms160ms200msAvg.AMASS3DPW40%60%Avg.
Task-Specific Model: one architecture for one task
EqMotion [65]15.635.652.167.979.550.151.8156.6193.6194.0193.8
STCFformer [52]18.339.358.975.088.656.045.0141.2177.2178.4177.8
siMLPe [41]14.926.244.056.668.842.148.8150.4236.3257.8247.1
In-Context Model: one task-agnostic architecture for multiple tasks
Static-SiC (ours)7.921.537.754.069.138.122.076.4162.5171.9167.2
Dynamic-SiC (ours)7.121.337.453.568.037.521.776.0150.1151.7151.0
", + "image_path": "f956ce6c2b907a14017fa3b0ec1a6addbd8daa1a3247810dc230e948760c4cb5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 196, + 521, + 387 + ], + "blocks": [ + { + "bbox": [ + 72, + 196, + 521, + 387 + ], + "lines": [ + { + "bbox": [ + 72, + 196, + 521, + 387 + ], + "spans": [ + { + "bbox": [ + 72, + 196, + 521, + 387 + ], + "type": "image", + "image_path": "0e4bd665dde05fc600a9f35b9eaf5524175ba4a79c5d194fd2f5df0034412678.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 394, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 46, + 394, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 545, + 406 + ], + "type": "text", + "content": "Figure 5. Comparison of generalization. Our SiC completes the missing skeletons according to the customized prompt for an unseen task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": "Copy Method. Following [5, 16], we copy the prompt target " + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "inline_equation", + "content": "T_{i}^{k}" + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": " as prediction results for comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "type": "text", + "content": "Multi-Task-Capable Models. For task-specific models [41, 52, 65] and multi-stage models [77], they are not able to handle multiple tasks at the same time. Therefore, we remodel them into end-to-end multi-task-capable models to achieve fair comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 496, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 579 + ], + "type": "text", + "content": "In-Context Models. Since there are no in-context models excelling in processing multiple human perception tasks, we select the most relative in-context models with our benchmark, Point-In-Context [16] and PointMAE [43], which is proposed in 3D point cloud understanding. Specifically, we take each frame as a token in their framework. For PointMAE, we use the transformed version in [16]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 587, + 216, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 216, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 216, + 600 + ], + "type": "text", + "content": "5.2. Compared with SoTA Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Quantitative Results. We report experimental results on four tasks in Tab. 1. We term Skeleton-in-Context implemented with static/dynamic-TUP as Static/Dynamic-SiC, respectively. For other models, we reproduce them and report the results. As can be seen from Tab. 1, our proposed model outperforms all multi-task models, achieving state-of-the-art results on the in-context multi-task benchmark. Notably, the results of task-specific models on a single task reflect the difficulty of each task. However, we also exhibit impressive" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 411, + 546, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 447 + ], + "type": "text", + "content": "performance compared to task-specific models, demonstrating that our proposed model is an expert in handling skeleton sequence-based multi-tasking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 448, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 520 + ], + "type": "text", + "content": "Qualitative Results. We visualize and compare the results of our SiC and the most recent SoTA model, MotionBERT [77], which is re-structured as an end-to-end multi-task model for fair comparison. As highlighted in Fig. 4, our SiC can generate more accurate poses than MotionBERT according to the provided task-guided prompt." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 528, + 452, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 452, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 452, + 540 + ], + "type": "text", + "content": "5.3. Generalization Capability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "Generalize to New Datasets. To verify the generalization ability of our model, we conduct cross-dataset experiments on a specific task. For instance, our proposed model learns motion prediction on the AMASS dataset under the default benchmark, and then it is used to predict future motion sequences on 3DPW during testing. Similarly, we set up four experiments: motion prediction (AMASS " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": " 3DPW), pose estimation (H3.6M " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": " AMASS, 3DPW), and joint completion (3DPW " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": " H3.6M). As shown in Tab. 2, our model is able to generalize the learned task well to other datasets, while other task-specific models cannot accommodate the large bridge introduced across datasets even if they are specifically trained on this single task. Additionally, benefiting from the dynamic pseudo pose, the Dynamic-SiC exhibits" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2442" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 86, + 279, + 139 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 251, + 82 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 251, + 82 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 251, + 82 + ], + "type": "text", + "content": "Table 3. Effectiveness of the Skeleton Prompt." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 86, + 279, + 139 + ], + "lines": [ + { + "bbox": [ + 53, + 86, + 279, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 86, + 279, + 139 + ], + "type": "table", + "html": "
#TGPTUPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1XX111.7159.2543.7160.2
2✓X25.052.537.765.2
3X✓26.052.437.878.4
4✓✓24.451.836.862.9
", + "image_path": "999a9f81ba85f32bac6384abbf49ba50af00f70e73725ae8a46b18a0d63bcbc1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 156, + 282, + 195 + ], + "blocks": [ + { + "bbox": [ + 100, + 140, + 233, + 151 + ], + "lines": [ + { + "bbox": [ + 100, + 140, + 233, + 151 + ], + "spans": [ + { + "bbox": [ + 100, + 140, + 233, + 151 + ], + "type": "text", + "content": "Table 4. Different Prompting (TGP)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 156, + 282, + 195 + ], + "lines": [ + { + "bbox": [ + 51, + 156, + 282, + 195 + ], + "spans": [ + { + "bbox": [ + 51, + 156, + 282, + 195 + ], + "type": "table", + "html": "
#TGPMP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
1Feature Sim.25.252.137.764.1
2Motion Sim.25.252.137.864.1
3Random24.451.836.862.9
", + "image_path": "f4deef0573e672eb4c942a89b2a43b435ad4d84208285a8c1a7d6f2f95fedb65.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 205, + 274, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 274, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 274, + 217 + ], + "type": "text", + "content": "more excellent generalization ability than the Static-SiC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 217, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 289 + ], + "type": "text", + "content": "Generalize to New Task. Furthermore, We construct a prompt that has not appeared in the training set, which performs motion in-between. As Fig. 5 shows, our trained model can complete the missing skeletons according to the customized task-guided prompt. Our SiC generalizes well to a new task, which is unexplored in other works." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 297, + 141, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 141, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 141, + 309 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 316, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 447 + ], + "type": "text", + "content": "Effectiveness of the Skeleton Prompt. We conduct ablation experiments on our proposed skeleton prompts. As shown in Tab. 3, the skeleton prompts solve the problem of directly extending previous work to skeleton sequences, which is falling into over-fitting. Meanwhile, applying both TGP and TUP to the model can achieve the best results. Our proposed skeleton prompts replace the random masking training strategy used in previous work. In particular, TUP creates a training environment that does not require the query target, which ensures that the model will not see the ground truth of the query input in advance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 448, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 590 + ], + "type": "text", + "content": "Prompt Engineering. Since previous work has proven that different prompts in context learning affect the performance of the model, we conduct ablation experiments on the choice of prompts. As Tab. 4 shows, we select the prompt by the feature similarity and MPJPE minimum between the input skeleton sequence and the prompt, respectively. Experimental results show that the random selection method achieves the best results. Unlike 2D images and 3D point clouds, skeleton-based task data is highly similar, resulting in different prompts playing similar roles in the same task. While the random selection method expands the model's receptive field to the data and achieves better results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "content": "Architecture. We conduct ablation experiments on the architecture of the model. As shown in Tab. 5, after exploring different hidden feature dimensions and transformer depths, we find that the model performs best when the feature dimension is 256, and the depth is 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 176, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 176, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 176, + 670 + ], + "type": "text", + "content": "5.5. In-Context Initial Pose" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We conduct an in-depth exploration of Dynamic-TUP. We extract the trained Dynamic-TUP weights, then randomly initialize a " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "(F,J,3)" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " tensor and use the trained encoder to in" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 86, + 544, + 269 + ], + "blocks": [ + { + "bbox": [ + 351, + 71, + 499, + 81 + ], + "lines": [ + { + "bbox": [ + 351, + 71, + 499, + 81 + ], + "spans": [ + { + "bbox": [ + 351, + 71, + 499, + 81 + ], + "type": "text", + "content": "Table 5. Ablation on model architecture." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 86, + 544, + 269 + ], + "lines": [ + { + "bbox": [ + 310, + 86, + 544, + 269 + ], + "spans": [ + { + "bbox": [ + 310, + 86, + 544, + 269 + ], + "type": "table", + "html": "
Arch.Num.MP. MPJPE↓PE. MPJPE↓JC. MPJPE↓FPE. MPJPE↓
Dim.6429.357.145.880.5
12826.453.839.266.1
25624.451.836.862.9
51224.750.737.663.0
Dep.325.153.739.265.6
425.152.238.664.7
524.451.836.862.9
624.752.938.163.7
(a) Task-Unified Prompt (Static)
(b) Task-Unified Prompt (Dynamic)
", + "image_path": "7ae595ac376388337924cb89960655ffcfe911d91a4df2754e246bbdd29df0e6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "type": "text", + "content": "Figure 6. (a) The visualization of Static-TUP. (b) The concrete visual skeleton sequence of Dynamic-TUP." + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 304, + 299, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 547, + 443 + ], + "type": "text", + "content": "crease its feature dimension to C-dimension. Finally, an optimizer is used to optimize the tensor and minimize its feature similarity with the trained Dynamic-TUP. The top of Fig. 6 represents the visualization of Static-TUP, while the bottom is the Dynamic-TUP. The concrete skeleton of dynamic-TUP is similar to the mean pose Static-TUP of the training set, which demonstrates that the Dynamic-TUP learned by our model is actually the prior knowledge whose receptive field is all ground truth in the training set. This prior knowledge helps our model learn more smoothly on multiple tasks, and adding learnable modules helps our model generalize to other datasets or new tasks (Sec. 5.3)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 628 + ], + "type": "text", + "content": "In this work, we propose the Skeleton-in-Context, designed to process multiple skeleton-base tasks simultaneously after just one training time. Specifically, we build a novel skeleton-based in-context benchmark covering various tasks. In particular, we propose skeleton prompts composed of TGP and TUP, which solve the overfitting problem of skeleton sequence data trained under the training framework commonly applied in previous 2D and 3D in-context models. Besides, we demonstrate that our model can generalize to different datasets and new tasks, such as motion completion. We hope our research builds the first step in the exploration of in-context learning for skeleton-based sequences, which paves the way for further research in this area." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement This work is supported by National Natural Science Foundation of China (No. 62203476), Natural Science Foundation of Shenzhen (No. JCYJ20230807120801002). This project is also supported by the interdisciplinary doctoral grants (iDoc 2021-360) from the Personalized Health and Related Technologies (PHRT) of the ETH domain, Switzerland." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2443" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 144 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "text", + "content": "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "type": "text", + "content": "[3] Ivana Balažević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Hénaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 290 + ], + "type": "text", + "content": "[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[7] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In ECCV, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 412 + ], + "type": "text", + "content": "[8] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 467 + ], + "type": "text", + "content": "[9] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 288, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 501 + ], + "type": "text", + "content": "[10] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 288, + 533 + ], + "type": "text", + "content": "[11] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 579 + ], + "type": "text", + "content": "[12] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. Unihcp: A unified model for human-centric perceptions. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 612 + ], + "type": "text", + "content": "[13] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In CVPR, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 645 + ], + "type": "text", + "content": "[14] Zhichao Deng, Xiangtai Li, Xia Li, Yunhai Tong, Shen Zhao, and Mengyuan Liu. Vg4d: Vision-language model goes 4d video recognition. In ICRA, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[16] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "learning for 3d point cloud understanding. arXiv preprint arXiv:2306.08659, 2023. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 129 + ], + "type": "text", + "content": "[17] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In CVPR, 2023. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[18] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 165, + 547, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 547, + 208 + ], + "type": "text", + "content": "[19] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 210, + 547, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 210, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 547, + 242 + ], + "type": "text", + "content": "[20] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening. TOG, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 279, + 547, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 279, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 547, + 312 + ], + "type": "text", + "content": "[22] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. TOG, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 314, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 314, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 314, + 545, + 346 + ], + "type": "text", + "content": "[23] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In ICCV, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 349, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 547, + 392 + ], + "type": "text", + "content": "[24] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. PAMI, 2013. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "text", + "content": "[25] Mahdi M Kalayeh, Emrah Basaran, Muhittin Gokmen, Mustafa E Kamasak, and Mubarak Shah. Human semantic parsing for person re-identification. In CVPR, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 428, + 547, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 547, + 460 + ], + "type": "text", + "content": "[26] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 3DV, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "type": "text", + "content": "[27] Sameh Khamis, Cheng-Hao Kuo, Vivek K Singh, Vinay D Shet, and Larry S Davis. Joint learning for attribute-consistent person re-identification. In ECCV, 2015. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "type": "text", + "content": "[28] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In CVPR), 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 576 + ], + "type": "text", + "content": "[29] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Guangliang Cheng, Pang Jiangmiao, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv pre-print, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 609 + ], + "type": "text", + "content": "[30] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into person: Joint body parsing & pose estimation network and a new benchmark. PAMI, 2018. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 611, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 633 + ], + "type": "text", + "content": "[31] Chunze Lin, Jiwen Lu, and Jie Zhou. Multi-grained deep feature learning for pedestrian detection. In ICME, 2018. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[32] Jun Liu, Henghui Ding, Amir Shahroudy, Ling-Yu Duan, Xudong Jiang, Gang Wang, and Alex C Kot. Feature boosting network for 3d pose estimation. PAMI, 2019. 2, 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "text", + "content": "[33] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. PAMI, 2019. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2444" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[34] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[35] Diogo C Luvizon, David Picard, and Hedi Tabia. 2d/3d pose estimation and action recognition using multitask deep learning. In CVPR, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[36] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "text", + "content": "[37] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "text", + "content": "[38] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 239, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 271 + ], + "type": "text", + "content": "[39] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 271, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 304 + ], + "type": "text", + "content": "[40] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 304, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 338 + ], + "type": "text", + "content": "[41] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 370 + ], + "type": "text", + "content": "[42] Xuecheng Nie, Jiashi Feng, and Shuicheng Yan. Mutual learning to adapt for joint human parsing and pose estimation. In ECCV, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "type": "text", + "content": "[43] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3, 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[44] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 448, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 288, + 503 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "text", + "content": "[46] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "type": "text", + "content": "[47] Justyna Sarzynska-Wawer, Aleksander Wawer, Aleksandra Pawlak, Julia Szymanowska, Izabela Stefaniak, Michal Jarkiewicz, and Lukasz Okruszek. Detecting formal thought disorder by deep contextualized word representations. *Psychiatry Research*, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "text", + "content": "[48] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu " + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{rgb + }" + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "text", + "content": " d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "type": "text", + "content": "[49] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[50] Chi Su, Fan Yang, Shiliang Zhang, Qi Tian, Larry Steven Davis, and Wen Gao. Multi-task learning with low rank attribute embedding for multi-camera person re-identification. PAMI, 2017. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[51] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 116, + 546, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 546, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 546, + 149 + ], + "type": "text", + "content": "[52] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatio-temporal criss-cross attention. In CVPR, 2023. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 149, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 149, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 149, + 547, + 182 + ], + "type": "text", + "content": "[53] Yonglong Tian, Ping Luo, Xiaogang Wang, and Xiaou Tang. Pedestrian detection aided by deep learning semantic tasks. In CVPR, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 182, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 182, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 547, + 217 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "text", + "content": "[55] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 259, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 259, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 546, + 304 + ], + "type": "text", + "content": "[56] Chaoyang Wang, Xiangtai Li, Henghui Ding, Lu Qi, Jiangning Zhang, Yunhai Tong, Chen Change Loy, and Shuicheng Yan. Explore in-context segmentation via latent diffusion models. arXiv, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 304, + 546, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 337 + ], + "type": "text", + "content": "[57] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 337, + 546, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 546, + 371 + ], + "type": "text", + "content": "[58] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 371, + 546, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 546, + 404 + ], + "type": "text", + "content": "[59] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 404, + 546, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 437 + ], + "type": "text", + "content": "[60] Xinshun Wang, Qiongjie Cui, Chen Chen, and Mengyuan Liu. Gcnext: Towards the unity of graph convolutions for human motion prediction. In AAAI, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 437, + 546, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 546, + 471 + ], + "type": "text", + "content": "[61] Xinshun Wang, Wanying Zhang, Can Wang, Yuan Gao, and Mengyuan Liu. Dynamic dense graph convolutional network for skeleton-based human motion prediction. TIP, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "type": "text", + "content": "[62] Zhendong Wang, Yifan Jiang, Yadong Lu, Yelong Shen, Pengcheng He, Weizhu Chen, Zhangyang Wang, and Mingyuan Zhou. In-context learning unlocked for diffusion models. arXiv preprint arXiv:2305.01115, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 514, + 546, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 559 + ], + "type": "text", + "content": "[63] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 559, + 546, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 546, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 546, + 603 + ], + "type": "text", + "content": "[64] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. arXiv preprint arXiv:2308.09718, 2023. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 603, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 647 + ], + "type": "text", + "content": "[65] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 5, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "type": "text", + "content": "[66] Shilin Xu, Xiangtai Li, Jingbo Wang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Fashionformer: A simple, effective and unified baseline for human fashion segmentation and recognition. In ECCV, 2022. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 690, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 546, + 714 + ], + "type": "text", + "content": "[67] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2445" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 443 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[68] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In ICCV, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 139 + ], + "type": "text", + "content": "[69] Angela Yao, Juergen Gall, and Luc Van Gool. Coupled action recognition and pose estimation from multiple views. IJCV, 2012. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 184 + ], + "type": "text", + "content": "[70] Bruce XB Yu, Zhi Zhang, Yongxu Liu, Sheng-hua Zhong, Yan Liu, and Chang Wen Chen. Gla-gcn: Global-local adaptive graph convolutional network for 3d human pose estimation from monocular video. In ICCV, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 288, + 218 + ], + "type": "text", + "content": "[71] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "type": "text", + "content": "[72] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 254, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 288, + 285 + ], + "type": "text", + "content": "[73] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 287, + 319 + ], + "type": "text", + "content": "[74] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 322, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 322, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 322, + 287, + 354 + ], + "type": "text", + "content": "[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 355, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 355, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 355, + 287, + 397 + ], + "type": "text", + "content": "[76] Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, et al. Generative tweening: Long-term inbetweening of 3d human motions. arXiv preprint arXiv:2005.08891, 2020.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 443 + ], + "type": "text", + "content": "[77] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In ICCV, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2446" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_content_list.json b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c1bb08d76838fdbf636af8473447600ae74a8e00 --- /dev/null +++ b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_content_list.json @@ -0,0 +1,1404 @@ +[ + { + "type": "text", + "text": "SketchINR: A First Look into Sketches as Implicit Neural Representations", + "text_level": 1, + "bbox": [ + 109, + 108, + 856, + 128 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hmrishav Bandyopadhyay1 Ayan Kumar Bhunia1 Pinaki Nath Chowdhury1 Aneeshan Sain1 Tao Xiang1,2 Timothy Hospedales3 Yi-Zhe Song1,2 \n1SketchX, CVSSP, University of Surrey, United Kingdom. \n2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence. \n3University of Edinburgh, United Kingdom \n{h.bandyopadhyay, a.bhunia, p.chowdhury, a.sain, t.xiang, y-song}@surrey.ac.uk t.hospedales@ed.ac.uk", + "bbox": [ + 101, + 136, + 867, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0326cb215bf5f979385ad4ca53dc3f4c8d7fb3a7bd4bed2acdb998828504ebc5.jpg", + "image_caption": [ + "Figure 1. SketchINR is an implicit neural representation for sequential vector sketches. It is the first neural representation with sufficient fidelity to be a drop-in replacement for the raw sketch data (a). The representation is significantly higher fidelity than existing learned representations such as SketchRNN [28], especially for more complex sketches (b, left). SketchINR also provides state-of-the-art sketch compression, with a substantially more compact representation than either vector or raster sketches (b, right)." + ], + "image_footnote": [], + "bbox": [ + 81, + 268, + 584, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b2dc68e7f26fc593d940f54567ddc9b80140fadfe77ac3e58b593325520dec8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 268, + 890, + 416 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 486, + 313, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose SketchINR, to advance the representation of vector sketches with implicit neural models. A variable length vector sketch is compressed into a latent space of fixed dimension that implicitly encodes the underlying shape as a function of time and strokes. The learned function predicts the $xy$ point coordinates in a sketch at each time and stroke. Despite its simplicity, SketchINR outperforms existing representations at multiple tasks: (i) Encoding an entire sketch dataset into a fixed size latent vector, SketchINR gives $60 \\times$ and $10 \\times$ data compression over raster and vector sketches, respectively. (ii) SketchINR's auto-decoder provides a much higher-fidelity representation than other learned vector sketch representations, and is uniquely able to scale to complex vector sketches such as FS-COCO. (iii) SketchINR supports parallelisation that can decode/render $\\sim 100 \\times$ faster than other learned vector representations such as SketchRNN. (iv) SketchINR, for the first time, emulates the human ability to reproduce a sketch with varying abstraction in terms of number and complexity of strokes. As a first look at implicit sketches, SketchINR's compact high-fidelity representation will support future work in modelling long and complex sketches.", + "bbox": [ + 73, + 507, + 473, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 848, + 209, + 862 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The prevalence of touch-screen devices has triggered significant research progress on sketches [4, 6, 8, 51, 57]. Mod-", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "elling digital sketches has become an important challenge for learning systems that aim to stimulate human creativity [17, 27]. Large-scale datasets [24, 28, 53] motivated several downstream applications like image retrieval [15, 61], image generation [37, 65], image editing [1, 41], and 3D content creation [5, 8, 38], among others [58, 60].", + "bbox": [ + 496, + 487, + 892, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These applications are underpinned by sketches captured in raw form either as raster or vector images, and usually encoded into derived representations by ConvNets [15], splines, and so on. For over a decade, discourse around what is a \"good\" representation of human sketches has persisted [31]. A now substantial body of work has focused on representations for sequential vector sketches that model both explicit strokes, and their drawing over time — most famously by the auto-regressive SketchRNN [28], but also using representations such as parametric curves [16], Gaussians [2], and Neural ODEs [17]. We introduce a novel continuous-time representation of sequential vector sketches by taking an implicit neural representation perspective on sketches for the first time.", + "bbox": [ + 496, + 580, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, we review the limitations of current sketch representations. The two most popular raw sketch representations are (i) raster sketch - a black and white image of a line drawing, and (ii) vector sketch - a temporal sequence of points and strokes. Raster sketch has a large but fixed storage (e.g., $256 \\times 256$ ). Its compatibility with ConvNets [64] made raster sketches popular but they lack the tempo-", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ral information (i.e., stroke order) necessary for generative modelling [17]. Vector sketches efficiently store the $xy$ -coordinates of points and strokes ( $N \\times 2$ ). However, the storage increases with the length of sketch $N$ (Fig. 1b). (iii) Raw vector sketches can be encoded by popular learned sketch representations such as SketchRNN [28] and others [16, 17, 39]. However, for longer ( $N \\geq 1000$ ) scene sketches, the quadratic computational cost for transformers becomes intractable [39] and recurrent nets become unstable [28]. Nevertheless, most of these still suffer from slow sequential autoregressive decoding, and they mostly cannot accurately reconstruct complex sketches with many strokes. To summarise, current sketch representations either discard temporal information (raster), scale poorly in storage size (vector), or become slow and inaccurate with complex sketches (learned autoregressive vector).", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the above issues, we propose an implicit neural representation for sketches - SketchINR. We piggyback on the well-established [8, 30, 34, 48] potential of implicit functions [22] to encode an underlying shape as a function of time/location. Different to standard implicits [43, 48], the input of sketches as parametric functions is a hierarchy of stroke sequence and point sequence. We develop SketchINR to generate points (xy-coordinates) given both time and stroke inputs. To represent multiple sketches, SketchINR learns a fixed size latent space and a conditional implicit function that generates xy-coordinates of any sketch instance given its corresponding latent code.", + "bbox": [ + 75, + 335, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its simplicity, SketchINR is competitive with existing sketch representations. (i) It can encode an entire sketch dataset into a compact latent space $(e.g., \\mathbb{R}^{N \\times 512})$ and decode using a fixed size function (e.g., an 8 layer MLP). Compared to raster sketches $(256 \\times 256)$ , SketchINR provides $128 \\times$ storage compression. While simple object-level vector sketches (length $\\leq 300$ ) have comparable storage with SketchINR ( $\\mathbb{R}^{512}$ ), for practical scene-level [13] vector sketches (length $\\geq 1000$ ), SketchINR provides $10 \\times$ better compression (Fig. 1b). (ii) For decoding/generation, SketchINR can parallelly predict the $xy$ -coordinates for all time and strokes. In practice, this is $100 \\times$ faster than an autoregressive vector sketch generator with length $\\geq 300$ . (iii) SketchINR can accurately represent complex sketches (Fig. 1a,b) unlike competitors that suffer from complexity limitations [16, 17] or the length limitations of auto-regression [28]. The encoding fidelity is sufficiently high that SketchINR provides a drop-in replacement for raw sketch data, while being more compact. (iv) SketchINR supports diverse applications spanning smooth latent space interpolation, and sketch generation, sketch completion. Uniquely, it also supports sketch abstraction - the human-like ability to replicate the essence of a sketch with a variable number and complexity of strokes [31, 54].", + "bbox": [ + 75, + 518, + 472, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (i) We present the", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "first implicit neural representation approach to vector sketch modelling, extending INRs as a function of time and strokes, and defining a training objective for learning them. (ii) SketchINR provides the first learned representation capable of representing complex sketches compactly and with high-fidelity (Fig. 1a). Based on our compact high-fidelity representation we introduce the task of sketch compression, and demonstrate excellent compression results (Fig. 1b)). (iii) We further demonstrate SketchINR's applicability to a variety of sketch related tasks including generation, interpolation, completion and abstraction.", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 271, + 640, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sketch Representations: Digital sketches are predominantly captured as a function of 'time' [28], in a sequence of coordinates traced by an artist on a canvas. Recent works [24, 28] emphasise this temporal order of coordinates as an indicator of their relative importance in depicting a sketch-concept, as humans draw in a coarse-to-fine fashion. As such, they [2, 7, 18, 28] exploit temporal information in sketches for downstream tasks like sketch-assisted retrieval [7], generation [2, 18, 28] and modelling [38]. Vector sequences are further parameterised, for representation with parametric curves like Béziers [16, 57] and B-Splines [50, 66] for early-handwritten digit representation [50] to recent representations of complex structures like real-world objects [16, 57] and scenes [58]. Discarding temporal significance of strokes, vector-sketches can be converted to raster images by rendering coordinate strokes [9] on a 2D canvas. While raster sketches are less informative than their vector counterparts [28], they offer enhanced applicability by expanding beyond digital sketches to paper-sketches [18] where stroke order is already lost. Raster representations are encoded with translation-invariant networks like CNNs [63] and Vision-Transformers [8], making them extremely useful to represent object-level information.", + "bbox": [ + 496, + 296, + 893, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-Parametric Modelling: Both vector sequences and raster renderings of these sequences are non-parametric representations [16] of sketches. To preserve temporal stroke order, vector sketches are handled as sequences [28] of coordinates rather than unordered sets. Hence, they are encoded with position-aware networks like RNNs [28], LSTMs [15, 61], and Transformers [10, 51]. Without positional cues, raster sketches are processed by spatially-aware CNNs [63] and Vision Transformers [52]. Semantic encodings of non-parametric representations capture the expressive nature of sketch, allowing its use as a creative input for downstream segmentation [33], retrieval [7], and editing [41], as well as for interactive tasks like object-detection [14] and image-inpainting [62]. These encodings are further used in auto-regressive [28, 59] and VAE-like [10] modelling of vector coordinate sequences for the generation of textual-characters [1] and object sketches [28].", + "bbox": [ + 496, + 643, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12566", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parametric Representation learning: Parametric Splines [20], such as Beizer curves [42], approximate vector sketches by per-sample fitting on individual sketch samples. Amortised frameworks bypass this per-sample optimisation by inferring curve control points [16] and degree [18] from individual stroke features. Recent works exploit the representative power of parametric curves for generation of simple characters [26] to complex object [16, 18, 57] and scene level sketches[58]. Both non-parametric and curve-fitted Bézier parametric representations capture the explicit form of a sketch as spatial and sequential forms.", + "bbox": [ + 76, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Beyond explicit representations of sketches, we learn visual implicits to capture spatial dynamics in vectorsketches. Learned implicits can be used for sketch reconstruction with particular control over abstraction through number of reconstructed strokes. Somewhat similar to implicit representations for vector sketches, CoSE [2] parameterises Gaussian Mixtures on individual strokes, learning one implicit for one stroke only. New implicits are generated auto-regressively from previous predictions for auto-regressive generation and completion of sketches. Unlike CoSE, we learn a single implicit for the entire sketch, thus having one implicit parameter to sample the sketch in one pass. While SketchODE [17] similarly parameterises Neural Ordinary Differential Equations [12] to capture vector-sketch dynamics, it's optimisation is extremely slow (as much as $120 \\times$ slower) because of high time complexity in solving higher order differentials. Three dimensional implicits like Signed-Distance [48] and Occupancy [43] functions learned from 3D point-clouds are particularly analogous to our implicit representation.", + "bbox": [ + 76, + 261, + 472, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch Abstraction: The continuity of human cognition in visual perception [36] is directly reflected in how we sketch. This enables versatility [64] in sketch, allowing us to express a huge range of thoughts and visions directly on paper in the form of coarse-grained 'ideas' [32] and fine-grained 'objects' [64]. Sketch abstraction, as a direct inverse of this granularity, has been studied in depth as a function of (i) drawing time: humans draw most representative strokes first [28], and (ii) compactness: in a constrained stroke setting, salient strokes are prioritised [46]. As abstraction gives sketches a human touch, generative modelling of sketches is focused on abstraction towards making generated drawings \"more humane\" [19] and augmenting limited sketch datasets with simulated sketches from photos [57, 58]. Importantly, recent works [57] model abstraction as a function of strokes to explicitly control abstraction levels by restricting the number of generated strokes. Here, we represent sketches as visual implicits which allows us to reconstruct them at varying levels of abstraction and detail by controlling the number of reconstructed strokes. We demonstrate several downstream applications of this representation, particularly sketch compression, generation and completion.", + "bbox": [ + 76, + 568, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 89, + 635, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview: Vector sketches are captured as pen movements on a digital canvas, represented by coordinates $(x,y) \\in [0,1]$ with binary pen-states $\\delta \\in \\{0:$ pen-down, $1:$ pen-up $\\}$ at those coordinates. A sketch $p = \\{(x_i,y_i,\\delta_i)\\}_{i=1}^N$ comprises $N$ vector way-points (coordinates), and is portrayed (rendered) by retracing these points on a 2D canvas. This process, termed as rasterisation produces a pixelated raster sketch $p^s$ on the canvas. In this work, we leverage scalability [48] in implicit functions to learn a flexible representation for vector sketches with raster guidance.", + "bbox": [ + 496, + 114, + 893, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learning Visual Implicits for Sketches", + "text_level": 1, + "bbox": [ + 498, + 273, + 831, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vector sketches can be modelled as implicit, where a sketch is represented as function of time $f_{\\theta}(t_j) : \\mathbb{R} \\to \\mathbb{R}^{2+1}$ parameterised by $\\theta$ . Each time-stamp $t_j \\in [0,1)$ is mapped to a pen position in the form of $xy$ -coordinates $(x,y) \\in \\mathbb{R}^2$ and binary pen-states $\\delta \\in \\mathbb{R}^1$ . The entire sketch can be reconstructed as a vector sequence by sampling $t$ from 0 to 1 with a learned $\\theta$ as $\\mathbf{p} = \\{f_{\\theta}(0),\\dots,f_{\\theta}(t_j),\\dots,f_{\\theta}(\\frac{J-1}{J})\\}$ for a total of $J$ timestamps. This in itself is a flexible representation for a vector sketch, as it allows for sampling with arbitrary timestamps, thus potentially increasing or decreasing resolution of the resulting sketch by modulating number of strokes with sampling resolution $(t : \\{0,\\frac{1}{3},\\frac{2}{3}\\} \\mathrm{v/s} t : \\{0,\\frac{1}{10},\\dots,\\frac{9}{10}\\})$ . For increased control over reconstruction granularity in the form of abstraction, we model $f_{\\theta}$ on an additional variable in the form of strokestamps $\\{s_k\\}_{k=1}^K$ , representing the sketch as $\\hat{p} = \\{f_{\\theta}(0,0),\\dots,f_{\\theta}(t_j,s_k),\\dots,f_{\\theta}(\\frac{J-1}{J},\\frac{K-1}{K})\\}$ for $K$ strokes (Fig. 2). These strokestamps provide explicit control over pen-up and pen-down states, allowing us to reconstruct the sketch with re-defined granularity during inference (e.g. 7 or 9 strokes). Specifically, strokestamps represent the current stroke number as a fraction of the total number of strokes $K$ which allows us to change strokes by changing $s_k$ . Pen-states $\\delta$ are thus explicitly determined from strokestamps only by toggling when $s_k$ changes value. Our practical implementation of sketch implicits follows:", + "bbox": [ + 496, + 297, + 893, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(t _ {j}, \\Delta t, s _ {k}, \\Delta s\\right) = p _ {j} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 712, + 890, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where, time instances are computed as $t_j = t_{j-1} + \\Delta t$ , $t_0 = 0$ , and $\\Delta t = \\frac{1}{J}$ . Similarly, $K$ -stroke instances are computed as $s_k = s_{k-1} + \\Delta s$ , $s_0 = 0$ , and $\\Delta s = \\frac{1}{K}$ . To exploit bias in neural networks towards learning low frequency functions [44, 49], we smoothen out high frequency variations $(0 \\text{to} \\frac{1}{K}, 0 \\text{to} \\frac{1}{S})$ for timestamps and strokestamps by mapping them from $\\mathbb{R}$ to a higher dimension $\\mathbb{R}^{2L}$ using positional embeddings $\\mathrm{PE}(\\cdot)$ as", + "bbox": [ + 496, + 736, + 893, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {P E} (x) = \\left[ \\sin \\left(1 0 ^ {- 4 / L} x\\right), \\cos \\left(1 0 ^ {- 4 / L} x\\right), \\right. \\tag {2} \\\\ \\ldots , \\sin (1 0 ^ {- 4 L / L} x), \\cos (1 0 ^ {- 4 L / L} x) ] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 863, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12567", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for $x \\in \\{t_n, \\Delta t, s_k, \\Delta s\\}$ . We uniformly distribute $J$ times-stamps across $K$ strokes, where each stroke has $J / K$ points. For data augmentation, we train $f_{\\theta}$ with varying $J$ and $K$ ( $\\pm 50\\%$ of ground truth value). This helps $f_{\\theta}$ to scale to different levels of abstraction, and number of points and strokes during inference.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1302058cfe7037f4cc43014ed0e7dce1a0670bd315cb43ae2fedc80c87f5fdc3.jpg", + "image_caption": [ + "Figure 2. We model sketch as a function $f_{\\theta}(t_j, s_k)$ of $J$ timestamps and $K$ strokes. Timestamps $\\{t_a, \\ldots, t_b\\}$ where $t_j \\in [0,1)$ correspond to way-points for a specific stroke $s_k$ . Finally, $f_{\\theta}$ with learned weights $\\theta$ can be sampled with arbitrary number of strokes as $t_j \\in \\{0, \\frac{1}{J}, \\dots, \\frac{J - 1}{J}\\}$ and $s_k \\in \\{0, \\frac{1}{5}, \\dots, \\frac{4}{5}\\}$ or $s_k \\in \\{0, \\frac{1}{10}, \\dots, \\frac{9}{10}\\}$ , leading to increasing or decreasing abstraction for $K = 5$ or $K = 10$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 83, + 191, + 464, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Loss functions", + "text_level": 1, + "bbox": [ + 76, + 497, + 225, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To fit our implicit function $f_{\\theta}$ to a particular sketch instance $p = \\{(x_0, y_0, \\delta_0), \\ldots, (x_{M-1}, y_{M-1}, \\delta_{M-1})\\}$ with $M$ vector way-points $\\{p_m\\}_{m=0}^{M-1}$ , a naive approach would be to optimise $\\theta$ with mean squared error against the ground truth vectors as:", + "bbox": [ + 75, + 521, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} = \\sum_ {m = 0} ^ {M} \\left| \\left| f _ {\\theta} \\left(\\frac {m}{M}, \\frac {k}{K}\\right) - p _ {m} \\right| \\right| _ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 602, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each predicted coordinate and pen-state is optimised directly against corresponding vector way-points from the ground truth. However, this rigid point-to-point matching enforces memorisation of particular points and strokes without any spatial understanding. As such, sampling with a different resolution at this stage yields noisy output that does not correspond to the raster sketch. To explicitly optimise for sampling flexibility, we introduce a visual loss, penalising sketch reconstructions that do not match visually. For this, we sample a template grid $G$ of 2D coordinates and compute the region of influence of each vector-waypoint $p_m$ of stroke $s_k$ on coordinate $g$ as an intensity map [3, 59]:", + "bbox": [ + 75, + 648, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I} _ {k} (g, s _ {k}) = \\max _ {p _ {m} \\in s _ {k}, r \\in [ 0, 1 ]} \\exp (\\gamma | | g - r p _ {m} - (1 - r) p _ {m + 1} | | _ {2}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 838, + 468, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where intensity $\\mathrm{I}_k$ is smoothed with $\\gamma$ for points $p_m$ in stroke $s_k$ where $m \\in \\{0, \\dots, \\frac{K - 1}{K}\\}$ . The intensity map", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the entire sketch is formulated as a summation of maps from all strokes for $k \\in (0,1]$ for both the ground truth sketch vector $p$ with $K$ strokes and the arbitrarily sampled sketch $\\hat{p}$ with $\\hat{K}$ strokes as $\\hat{\\mathrm{I}}_k$ , composing the loss as:", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V} = \\left| \\left| \\sum_ {k = 0} ^ {\\frac {K - 1}{K}} \\mathrm {I} _ {k} - \\sum_ {k = 0} ^ {\\frac {\\hat {K} - 1}{\\hat {K}}} \\hat {\\mathrm {I}} _ {k} \\right| \\right| _ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 162, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intensity Smoothing with $\\gamma$ : The sketch intensity map $I_k$ helps us to ground vector sketches to raster definiteness (Eq. (4)) by a visual rendering of individual strokes where intensity in coordinates $g \\in G$ drops exponentially with distance from $p_m \\in s_k$ . This exponential drop is controlled by a smoothing factor $\\gamma$ as a hyperparameter, where (i) a low $\\gamma$ yields thicker lines on $I_k$ (low-quality sketch reconstructions) but offers better optimisation, while (ii) a high $\\gamma$ yields more refined lines (higher quality reconstructions), making optimisation significantly slower (Fig. 3).", + "bbox": [ + 496, + 222, + 890, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For faster optimisation, we assist the visual penalty with a weighted MSE loss from Eq. (3) and a lower gamma, empirically set to $\\gamma = 150$ . The loss function weighted with $\\lambda_{\\mathrm{MSE}}$ , looks can be written as:", + "bbox": [ + 496, + 373, + 890, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i m p l i c i t}} = \\mathcal {L} _ {V} ^ {\\gamma = 1 5 0} + \\lambda_ {\\mathrm {M S E}} \\cdot \\mathcal {L} _ {\\mathrm {M S E}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 443, + 890, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we set $\\lambda_{\\mathrm{MSE}}$ to 0.7. This, however, inherently introduces bias in our optimisation to match the order of strokes with the ground truth order to some extent. Future works could ideally set $\\lambda_{\\mathrm{MSE}} = 0$ if they obtain faster convergence from a better engineered implicit function $f_{\\theta}$ . Fig. 4a summarises our implicit sketch learning framework.", + "bbox": [ + 496, + 473, + 890, + 565 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f25d662a62327603735692a8289c6b17afa2685b24422537af4b8eb38e066618.jpg", + "image_caption": [ + "Figure 3. Effect of smoothing factor $\\gamma$ on training: Reducing $\\gamma$ leads to higher intensity in the surrounding region near stroke $s_k$ . This is similar to stroke dilation of intensity map $\\mathrm{I}_k$ . A lower $\\gamma$ leads to stable training (plot on right) but lacks fine-grained details. A higher $\\gamma$ gives a fine-grained sketch but is harder to train." + ], + "image_footnote": [], + "bbox": [ + 501, + 588, + 908, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Generalising to Multiple Sketches", + "text_level": 1, + "bbox": [ + 500, + 786, + 794, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To efficiently represent multiple sketch implicits with a single global and general function, we modify Eq. (1) to include a sketch descriptor $\\nu_{i}\\in \\mathbb{R}^{d}$ for each sketch $i$ . The function $f_{\\theta}(\\nu_{i},t,\\Delta t,s_{k},\\Delta s)$ is approximated with a fully connected auto-decoder network, similar to DeepSDF [48]. Thus, feature descriptors $\\nu_{i}$ are hidden vectors, learned", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe7980f604d4fcbf9b3db3f358080564a901f7311acd41ed00da39b329de927c.jpg", + "image_caption": [ + "(a): Single sketch Implicit Learning" + ], + "image_footnote": [], + "bbox": [ + 71, + 107, + 356, + 256 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6dcc40403b97c8c12ee7a18cf1a6536edc27118b5e77f87db5a294e3fb4d7828.jpg", + "image_caption": [ + "(b): Multi-sketch auto-decoder", + "Figure 4. SketchINR model diagram. (a) Embedding a single vector sketch as an implicit model $F$ mapping stroke and time $(\\mathbf{s},\\mathbf{t})$ to ink coordinate $\\mathbf{p}$ . (b) Embedding a vector sketch dataset as a shared decoder and set of latent vectors $\\nu$ . (c) Training a generative model for implicit sketches by generating latent codes $\\nu$ using an encoder $E$ that inputs raster sketches." + ], + "image_footnote": [], + "bbox": [ + 361, + 112, + 576, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/46cd639eaf6fb2e4abe776e82e3e5bd93b2bf9776689786754d0c428ddc7488b.jpg", + "image_caption": [ + "(c): Mapping to $\\mathcal{N}(0,\\mathbf{I})$ for generative modelling" + ], + "image_footnote": [], + "bbox": [ + 586, + 112, + 906, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "jointly with the function parameters $\\theta$ on a training dataset $(i\\in D)$ . Similar to single sketch optimisation, we train on multiple samples with a combination of MSE and visual loss (Eq. (6)). Fig. 4b summarises our multi-sketch autodecoder through block diagrams.", + "bbox": [ + 73, + 319, + 468, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generative Modelling: The generalised auto-decoder, enables simple generative modelling of sketches by generating their feature descriptors $\\nu_{i}\\in \\mathcal{V}$ , which allows us to create novel sketch implicits. Towards this, we train a Variational Auto-Encoder [35] as our generator, where we replace the decoder with our pre-trained implicit decoder. For variational inference, our encoder $E(\\cdot)$ represents each raster sketch instance $i$ as a set of $d$ Gaussians with mean $\\mu_i\\in \\mathbb{R}^d$ and variance $\\sigma_{i}\\in \\mathbb{R}^{d}$ which are re-parameterised [35] to a latent encoding as $z_{i} = \\mu_{i} + \\sigma_{i}\\cdot \\epsilon$ where $\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . This latent is then projected to the $d$ -dimensional implicit latent space $\\nu_{i}\\in \\mathcal{V}$ . We train the encoder as a ResNet-18 network with (i) reconstruction loss from both our visual and vector spaces (Eq. (6)) and (ii) KL-divergence loss to bring the distribution closer to a unit normal $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The loss function for the encoder can be written as:", + "bbox": [ + 75, + 395, + 472, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e n c}} = \\mathcal {L} _ {\\text {i m p l i c i t}} - \\beta \\cdot D _ {K L} \\left(q \\left(z _ {i} \\mid i\\right), \\mathcal {N} (\\mathbf {0}, \\mathbf {I})\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 650, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a weighing factor, set to 0.7 and $q(z_i|i)$ represents the underlying probability distribution modelled with encoder $E$ that infers $z_{i}$ for a given sketch $i$ . Fig. 4c summarises our implicit sketch generative model. The decoder can be used for sketch generation, while the encoder can encode raster sketches for vectorisation.", + "bbox": [ + 75, + 678, + 470, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Applications", + "text_level": 1, + "bbox": [ + 76, + 784, + 207, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We learn implicits for sketches of various complexities and abstraction levels, ranging from highly complex scene sketches that correspond to a photo scenery [13] to abstract doodles [28] drawn from memory. Based off complexity in downstream applications, we use scene sketches to demonstrate self-reconstruction and sketch", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "compression, while focusing on simpler object level sketches for more complex tasks like sketch mixing (interpolation) and generation for faster optimisation. Specifically, we use the FS-COCO [13], Sketchy [53], and QuickDraw! [28] datasets. (i) FS-COCO [13] contains 10,000 hand-drawn sketches of complex MS-COCO [40] scenes, drawn by amateurs without time constraints. Sketches in FS-COCO have on a median of 64 strokes, resulting in the sketch having $\\sim 3000$ waypoints. (ii) Comparatively simpler than scene sketches, Sketchy [53] consists of $\\sim 75\\mathrm{K}$ sketches of $12.5\\mathrm{K}$ object photographs from 125 categories. These sketches similar to FS-COCO are drawn from reference photographs and have photo-sketch pairs. (iii) QuickDraw! [28] sketches are drawn from memory and in a constrained time settings ( $< 20\\mathrm{s}$ ). Hence, these sketches (doodles) are abstract in nature, lacking any pre-defined configuration or orientation from reference photos. In addition to these three datasets, we use a vectorised form of the MNIST dataset, Vector-MNIST [17], consisting of $\\sim 10\\mathrm{K}$ samples as a simpler dataset to demonstrate complex downstream applications like generation and interpolation.", + "bbox": [ + 496, + 319, + 893, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. SketchINR: A Compact High-Fidelity Codec", + "text_level": 1, + "bbox": [ + 498, + 648, + 879, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by demonstrating that SketchINR provides a high fidelity neural representation for sketches. Specifically, we train SketchINR on complex variable length vector sketch datasets such as FS-COCO and Sketchy. We represent each sketch in these datasets as a fixed size vector with a dataset-specific decoder $f_{\\theta}$ . We then render each implicit sketch $\\nu_{i} \\in \\mathcal{V}$ with the same number of strokes as the ground-truth sketch $\\mathbf{p}^i$ to evaluate reconstruction quality. Following [17, 28], we measure reconstruction quality using (i) Chamfer Distance (CD), and (ii) retrieval accuracy (top-10) with a naive sketch retrieval network (ResNet-18) trained on ground-truth sketch vectors $\\mathbf{p}$ and evaluated on rendered $\\hat{\\mathbf{p}}$ . The quantitative results in Tab. 1, show that SketchINR provides substantially higher fidelity encoding than alternative representations like RNN [28], Bézier strokes [16],", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12569", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c436081b815089d448db45b554119c70d0a96d8a63d997a3ee8f3a09cf909e51.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FS-COCO [13]Sketchy [53]
CDR@10Time (s)CDR@10Time(s)
SketchRNN [28]---0.1641.650.11
CoSE [2]---0.1445.820.10
SketchODE [17]0.415.171.680.1737.200.13
BézierSketch [16]0.572.961.220.2325.720.08
Ours0.01115.610.30.00857.290.0007
", + "bbox": [ + 78, + 88, + 467, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Evaluating neural sketch representations' reconstruction quality (Chamfer distance $\\downarrow$ ; recall@ $10\\uparrow$ ) and decoding speed.", + "bbox": [ + 76, + 188, + 468, + 217 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/adea143865d18c7223b4c716cfdbb2e36a60b7d26bb8b33d56dea1c1de6b1c1c.jpg", + "image_caption": [ + "Figure 5. Qualitative reconstruction results for neural sketch representations. SketchINR is uniquely able to scale to sketches of Sketchy and FS-COCO complexity." + ], + "image_footnote": [], + "bbox": [ + 81, + 232, + 467, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ODE dynamics [17], and stroke embeddings [2] particularly for extremely complex sketches in FS-COCO, where SketchRNN and CoSE fail entirely. Qualitative results in Fig. 5 show SketchINR to reconstruct simple and complex sketches with high fidelity while SketchRNN performs poorly for Sketchy and fails completely on FS-COCO.", + "bbox": [ + 75, + 579, + 468, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sketch Compression: We next show that SketchINR's high-fidelity encoding is extremely compact, thus providing an efficient sketch codec. INRs store information implicitly [55] in the network weights $(\\theta)$ and latent codes $(\\nu_{i})$ . To store or transmit a sketch dataset, we only stream weights $\\theta$ once (fixed cost) and a compressed latent vector $\\nu_{i} \\in \\mathbb{R}^{D}$ for each sketch (variable cost). Hence, the space complexity for $N_{D}$ sketches with SketchINR is $|\\theta| + N_{D} \\times |\\nu_{i}|$ bytes, where $\\nu_{i}$ is 64 dimensional latent. For streaming a large number of sketches, we discount $\\theta$ as negligible streaming overhead. Comparing to alternative representations, streaming $N_{D}$ (i) binary raster (BR) sketches of size $(N_{x} \\times N_{y})$ take $N_{D} \\times N_{x} \\times N_{y}$ bits, and (ii) vector scene sketches with $N$ coordinates and pen states $(\\mathbb{R}^{2+1})$ take $N_{D} \\times N \\times 2 \\times 3$ bytes in 16-bit floating precision ( $N \\approx 3000$ in FS-COCO", + "bbox": [ + 75, + 672, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[13]). Raster representations can be further stored by only storing the coordinates $(\\mathbb{R}^2)$ of pixels containing the sketch (i.e., inked pixels) as sparse binary rasters (SBR). Fig. 6 shows rate-distortion for storing 10,000 scene sketches [13] with $\\nu_{i}$ varying from 128 to 8 dimensional latent, and quality for raster and vector forms dropped with downsampling and RDP simplification [23] respectively. Quality metrics like PSNR and SSIM are more suited to photos. We use Chamfer Distance as a quality metric for vector sketch and observe $\\sim 10\\times$ and $\\sim 60\\times$ more efficiency than vector and raster sketches, respectively. Future works can further optimise our models and thus reduce storage requirements via model compression [21, 29, 56] and quantisation [11].", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1c6198caa03f00604f631ac1001a30e0a4ff7eaecd8419e117a6b8a4cb5f0f4.jpg", + "image_caption": [ + "Figure 6. Rate Distortion: SketchINR can encode complex scene sketches from FS-COCO [13] in highly compact $(\\mathbb{R}^{64})$ latent codes. Specifically, despite nearly identical sketch quality (low CD represents higher fidelity), SketchINR has $\\sim 60\\times$ lower BPP than PNG raster sketches and $\\sim 10\\times$ lower than vectors sketches." + ], + "image_footnote": [], + "bbox": [ + 501, + 301, + 890, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intra-Sketch Variation: Unlike prior representations (RNN [28], ODE dynamics [17]), humans do not draw exact stroke-level replica of a raster sketch. Instead, we choose the level of detail, the number of points, and the strokes used to convey the underlying shape. Given a latent code $\\nu_{i}$ , SketchINR allows users to explicitly choose $N$ and $K$ , where the number of points $p = \\{p_1,\\dots ,p_N\\}$ and strokes $s = \\{s_1,\\ldots ,s_K\\}$ . Choosing a low $N$ and $K$ leads to abstract self-reconstructions with fewer points and strokes. Fig. 7(top to bottom) shows that we can control $K$ to decrease or increase the level detail in the sketch.", + "bbox": [ + 496, + 505, + 893, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Latent Space Interpolation (Creative Mixing)", + "text_level": 1, + "bbox": [ + 498, + 681, + 883, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interpolation of learned latents allows us to explore latent space continuity. Auto-regressive modelling of sketches builds poor latents as explored by Das et. al. [19] due to the lack of a full visibility of the sketch at any decoding timestep. Specifically, auto-regressive models use the partially predicted sketch along with the sketch's latent representation to predict new way-points with consistency. This, however, introduces noise in the latent $\\rightarrow$ sketch decoding so that small changes in these latents lead to big changes in the final sketch. Deterministic latent to sketch decoding solves this problem as noise is not added at any point, leading to meaningful sketch interpolation from corresponding latent space interpolations. We visualise in", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12570", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b46ea493f27fe1a381abda23a10a47263c0590d32d66c3d01d509d97e455da19.jpg", + "image_caption": [ + "Figure 7. Given a latent code $\\nu_{i}$ , we can reconstruct a sketch using learned $f_{\\theta}$ at multiple resolutions by controlling the number of strokes $K$ during sampling." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 464, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Sketch Interpolation with auto-regressive networks - SketchRNN", + "bbox": [ + 84, + 313, + 431, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5533c1a13f69b71f81b46f45ee080bf8a47161f049f4c5215d8834eaaf7872ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 330, + 455, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Sketch interpolation with implicits - SketchINR", + "bbox": [ + 86, + 412, + 352, + 425 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0d842a57d73d2ac8488ffdc6888a567f9ffff78778527e2102e93cfbbc39f74.jpg", + "image_caption": [ + "Figure 8. Latent space interpolation shows continuity and translation from one sketch concept to another." + ], + "image_footnote": [], + "bbox": [ + 86, + 431, + 455, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8 how one sketch morphs into another by reconstructing from $f_{\\theta}(\\nu', t_n, \\Delta t, s_k, \\Delta s)$ , where $\\nu'$ is sampled along a latent walk: $\\nu' = (1 - \\delta)\\nu_1 + \\delta \\nu_2$ , for $\\delta \\in [0,1]$ . Interpolating between two sketches in the continuous latent space gives creative mixing – modifying the sketch of airplane-1 (facing right) into airplane-2 (facing left) by varying $\\delta$ . Fig. 8 shows that small changes in the latent (columns) leads to smoother changes in sketch space for SketchINR. Meanwhile SketchRNN can exhibit big jumps that can zig-zag over successive latent increments, as indicated by an example tracked point in the figure.", + "bbox": [ + 75, + 559, + 468, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. INR inversion for sketch completion", + "text_level": 1, + "bbox": [ + 76, + 739, + 390, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Neural networks used to approximate implicit representations have been inverted in parallel literature [30] to appropriate samples not seen during training. This inversion, to arrive at a latent $\\nu$ that describes the new sample, is usually made possible by learned priors in $f_{\\theta}$ . Specifically, an implicit decoder $f_{\\theta}$ is frozen with latent $\\nu$ optimised for $f_{\\theta}(\\nu, \\ldots)$ against the ground truth. This behaviour is easily replicated in SketchINR, where we optimise the latent $\\nu$ to describe sketch samples (Fig. 5). To implement", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a) Input Partial Stroke Vector-MNIST (Left) and Quick-Draw! (Right)", + "bbox": [ + 509, + 84, + 862, + 97 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7898dce9327a887394bedb35dcbbd9accd32630f7ecad25846834534df9aa74a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 97, + 883, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f1ca3a3ef16062faf1340b788350408e9f7df30f0043f1bf990f966bb43514d.jpg", + "image_caption": [ + "Sketch Completion using SketchRNN", + "Sketch Completion using SketchINR" + ], + "image_footnote": [], + "bbox": [ + 506, + 152, + 883, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8b8505fc47496acf52cf3d713808f03d5d485f34ffeea1fc5dc17528d3aa569.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 210, + 880, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2cd8f870418982026c14026f1fa834ad6392f262952f41dcaab0e641a2e2e863.jpg", + "image_caption": [ + "Figure 9. We perform sketch completion by inverting learned implicit representations and decoding for occluded strokes or waypoints. (a) Temporal completion (b) Unlike auto-regressive methods like SketchRNN [28], we can perform a-temporal (or inverse-temporal) completion, where given the second half of a partial sketch, we can successfully reconstruct the first part." + ], + "image_footnote": [], + "bbox": [ + 503, + 263, + 870, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sketch completion, we obtain a sketch descriptor $\\nu$ from a partial sketch by just optimising for specific time-steps (e.g. 1/4th of a sketch; $t \\in [0, \\frac{1}{4})$ ) available in a partial doodle. Then, by sampling for rest of the timesteps we can complete the incomplete sketch. We summarise some qualitative sketch completion results in Fig. 9(a). Completed sketches are varied because of randomness in optimisation, as partial sketches are often ambiguous. INR inversion thus provides a novel approach to sketch completion, previously dominated by auto-regressive and models. Unlike auto-regressive models SketchINR can also perform a-temporal completion, inferring the first half of a sketch given the second half of the sketch (Fig. 9b). We remark that popular INR applications such as super-resolution [47] and point-cloud densification [48] perform interpolation after inversion on sparse inputs (e.g.: $t \\in [0, 1)$ given $t \\in \\{0, \\frac{1}{4}, \\frac{1}{2}, \\frac{3}{4}\\}$ ), while sketch completion requires extrapolation (e.g.: $t \\in [0, 1)$ given $t \\in [0, \\frac{1}{4})$ ), making it more analogous to visual outpainting.", + "bbox": [ + 496, + 406, + 892, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Sketch Generation", + "text_level": 1, + "bbox": [ + 500, + 708, + 679, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can perform conditional and unconditional sequential vector sketch generation after learning a generative model for sketch latents as described in Sec. 3.3.", + "bbox": [ + 496, + 732, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unconditional Generation: We unconditionally generate sketches by random sampling $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ with the learnt VAE latent space, obtaining novel sketch descriptors $\\nu_{i}\\in \\mathcal{V}$ . We further sample $f_{\\theta}(\\nu_{i},\\dots)$ with varying timestamps $T\\in [100,300]$ and strokestamps $K\\in [10,30]$ obtaining sketches at varying levels of abstraction. Qualitative results in Fig. 10 show diverse and plausible samples from Vector-MNIST [17] and Quick-Draw! [28] datasets.", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c071b249dd46dd7892a71ac96dca3d64f907bb483cce2bf2664dbe6d8249f73b.jpg", + "image_caption": [ + "Figure 10. Unconditional sketch generation. Sampling sketch latents $\\nu$ as well as stroke and point complexity $(K, T)$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 468, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conditional Generation: To condition sketch generation, we use our pre-trained VAE encoder $E$ to embed raster sketches, which can then be vectorised by the decoder $f_{\\theta}$ . Fig. 11 shows qualitative results for sequential vector sketch generation from raster images, where we obtain comparable results with auto-regressive generative modelling [28]. We observe that treating vectorisation as a generative problem rather than a deterministic raster-to-vector mapping [7] leads us to have variations in generated vectors particularly in the form of stroke order resembling real copies of sketches made by humans.", + "bbox": [ + 75, + 273, + 472, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ddf0fc6e578f25e0212718635520cfa79c9b40491e03db5197833cfb8968344b.jpg", + "image_caption": [ + "Figure 11. Conditional sketch generation (vectorisation) on Vector-MNIST. Raster sketches (top left) can have multiple vector forms, e.g., sketched with different numbers of strokes (bottom)." + ], + "image_footnote": [], + "bbox": [ + 78, + 448, + 468, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablations and Discussion", + "text_level": 1, + "bbox": [ + 76, + 619, + 307, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Global vs Local Time Modelling: We model $J$ timestamps across $K$ strokes, by distributing timestamps uniformly over each stroke such that each stroke has $\\frac{J}{K}$ timestamps. In other words, time $(t,\\Delta t)$ is modelled globally for the entire sketch $\\mathbf{p} = \\{\\mathrm{p}_1,\\dots ,\\mathrm{p}_N\\}$ . An alternative is to model $(t,\\Delta t)$ as 0to1 for each stroke separately (local time-modelling) instead of the entire sketch. Particularly, each of the $K$ strokes are modelled using $N$ points, representing the entire sketch using $N\\cdot K$ points as $\\mathbf{p} = \\{\\mathrm{p}_1,\\dots ,\\mathrm{p}_{N\\cdot K}\\}$ . Visually inspecting global vs local time embedding shows that local modelling leads to jittery sketches (see Fig. 12), whereas global modelling gives visually smooth sketches. We hypothesise that time is a global property of a sketch and not an independent property localised to each stroke. Hence, modelling time locally puts an additional optimisation burden on our implicit decoder to ensure smoothness and continuity, resulting in poorly reconstructed sketches.", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f8e4975916570af285cfd13ce622394fb52e77f569e562e5f0f99ade64f6219.jpg", + "image_caption": [ + "Figure 12. Global vs Local time embedding: Global embedding gives smooth sketches compared to local, which gives jittering." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 883, + 157 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fixed vs Variable Smoothing Factor $(\\gamma)$ : The intensity map $\\mathrm{I}_k$ in Eq. (4) guides our raster-based visual loss $\\mathcal{L}_V$ . A key component in $\\mathrm{I}_k$ is the smoothing parameter $\\gamma -$ low $\\gamma$ (thicker $\\mathrm{I}_k$ ) gives less accurate but smoother values on the entire visual region (easier training); high $\\gamma$ (thinner $\\mathrm{I}_k$ ) gives more accurate but sharper (steeper) intensity-gradients near ground-truth strokes (harder training). While we use a fixed $\\gamma$ with MSE Loss in Eq. (6) to balance out the optimisation complexity, we note that an alternative is to vary $\\gamma$ every fixed number of steps with a scheduler to allow faster optimisation. Specifically, we vary $\\gamma$ as 20to200 on a linear scale by incremental $\\Delta \\gamma$ every few iterations. Despite its theoretical advantage [25, 45], our initial experiments suggest only minor improvement that reduces training time by $6.1\\%$ . A detailed study of bundle-adjustment [45] for visual loss is an interesting future work.", + "bbox": [ + 496, + 193, + 893, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work Despite it's efficiency as a representation for hand-drawn sketches, SketchINR suffers from a number of limitations. We naturally share limitations with other implicit representations including optimisation-based encoding and poor cross category generalisation. While unlike Sketch-RNN our generation network can be trained on 2-3 categories at once, it fails to generalise on more. In addition, a core limitation lies in slow convergence due to pixel space optimisation. While vector-space loss in the form of mean squared error mitigates this to some extent (Fig. 3), convergence is still slow and can possibly be improved by better engineering the implicit function. Strokes could also be smoothed by explicit inclusion of stroke gradient (slope) based regularisation in future work. Finally, vectorisation for conditional generation performs poorly on high complexity sketches.", + "bbox": [ + 496, + 438, + 893, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 686, + 619, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce visual implicits to represent vector sketches with compressed latent descriptors. This neural representation provides a high-fidelity and compact representation that raises the possibility of a sketch-specific codec for compactly representing large sketch datasets. Our SketchINR can decode sketches at varying levels of detail with controllable number of strokes, and provides superior cross-sketch interpolation between implicits, demonstrating a smoother latent space than auto-regressive models. Applications such as sketch-completion are also supported, including a-temporal completion not available with auto-regressive models. Decoding is inherently parallel and can be over $100 \\times$ faster than autoregressive models in practice.", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12572", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emre Aksan, Fabrizio Pece, and Otmar Hilliges. Deepwriting: Making digital ink editable via deep generative modeling. In CHI, 2018. 1, 2", + "[2] Emre Aksan, Thomas Deselaers, Andrea Tagliasacchi, and Otmar Hilliges. Cose: Compositional stroke embeddings. In NeurIPS, 2020. 1, 2, 3, 6", + "[3] Stephan Alaniz, Massimiliano Mancini, Anjan Dutta, Diego Marcos, and Zeynep Akata. Abstracting sketches through simple primitives. In ECCV, 2022. 4", + "[4] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks. In CVPR, 2024. 1", + "[5] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 1", + "[6] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can beat me? In SIGGRAPH Asia, 2020. 1", + "[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 2, 8", + "[8] Alexandre Binninger, Amir Hertz, Olga Sorkine-Hornung, Daniel Cohen-Or, and Raja Giryes. Sens: Part-aware sketch-based implicit neural shape modeling. arXiv preprint arXiv:2306.06088, 2024. 1, 2", + "[9] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 1965. 2", + "[10] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation. In NeurIPS, 2020. 2", + "[11] Qualcomm Innovation Center. Ai model efficiency toolkit (aimet). https://github.com/quic/aimet, 2023.6", + "[12] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. In NeurIPS, 2018. 3", + "[13] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2, 5, 6", + "[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aaneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What can human sketches do for object detection? In CVPR, 2023. 2", + "[15] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2", + "[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 1, 2, 3, 5, 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Sketchode: Learning neural sketch representation in continuous time. In ICLR, 2021. 1, 2, 3, 5, 6, 7", + "[18] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Cloud2curve: Generation and vectorization of parametric sketches. In CVPR, 2021. 2, 3", + "[19] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In ICLR, 2023. 3, 6", + "[20] Carl De Boor and Carl De Boor. A practical guide to splines. Applied Mathematical Sciences, 1978. 3", + "[21] Zhen Dong, Zhewei Yao, Amir Gholami, Michael Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In ICCV, 2019. 6", + "[22] Asen L. Dontchev and R. Tyrrell Rockafellar. Implicit Functions and Solution Mappings. Springer-Verlag, 2014. 2", + "[23] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 1973. 6", + "[24] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 2", + "[25] F. Dan Foresee and Martin T. Hagan. Gauss-newton approximation to bayesian learning. In ICNN, 1997. 8", + "[26] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, SM Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In ICML, 2018. 3", + "[27] Songwei Ge, Vedanuj Goswami, C. Lawrence Zitnick, and Devi Parikh. Creative sketch generation. In ICLR, 2021. 1", + "[28] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 2, 3, 5, 6, 7, 8", + "[29] Hai Victor Habi, Roy H. Jennings, and Arnon Netzer. Hmq: Hardware friendly mixed precision quantization block for cnns. In ECCV, 2020. 6", + "[30] Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. Spaghetti: Editing implicit shapes through part aware generation. ACM TOG, 2022. 2, 7", + "[31] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2", + "[32] Conghui Hu, Da Li, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Sketch-a-classifier: Sketch-based photo classifier generation. In CVPR, 2018. 3", + "[33] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2", + "[34] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2", + "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5", + "[36] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1", + "[38] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Free2cad: Parsing freehand drawings into cad commands. ACM TOG, 2022. 1, 2", + "[39] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 2", + "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5", + "[41] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Deflocnet: Deep image editing via flexible low-level controls. In CVPR, 2021. 1, 2", + "[42] Asif Masood and Sidra Ejaz. An efficient algorithm for robust curve fitting using cubicBezier curves. In ICIC, 2010. 3", + "[43] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 2, 3", + "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3", + "[45] Jorge J. Mofe. The levenberg-marquardt algorithm. In Numerical Analysis, 1976. 8", + "[46] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3", + "[47] Quan H Nguyen and William J Beksi. Single image superresolution via a dual interactive implicit neural network. In WACV, 2023. 7", + "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3, 4, 7", + "[49] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML, 2019. 3", + "[50] Michael Revow, Christopher KI Williams, and Geoffrey E Hinton. Using generative models for handwritten digit recognition. IEEE TPAMI, 1996. 2", + "[51] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In CVPR, 2020. 1, 2", + "[52] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 2" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 1, 5, 6", + "[54] Heping Sheng, John Wilder, and Dirk B. Walther. Where to draw the line? PLOS One, 2021. 2", + "[55] Yannick Strumpler, Janis Postels, Ren Yang, Luc van Gool, and Federico Tombari. Implicit neural representations for image compression. In ECCV, 2022. 6", + "[56] Stefan Uhlich, Lukas Mauch, Fabien Cardinaux, Kazuki Yoshiyama, Javier Alonso Garcia, Stephen Tiedemann, Thomas Kemp, and Akira Nakamura. Mixed precision dnns: All you need is a good parametrization. In ICLR, 2020. 6", + "[57] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM TOG, 2022. 1, 2, 3", + "[58] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In ICCV, 2023. 1, 2, 3", + "[59] Alexander Wang, Mengye Ren, and Richard Zemel. Sketchembednet: Learning novel concepts by imitating drawings. In ICML, 2021. 2, 4", + "[60] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In NeurIPS, 2023. 1", + "[61] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 1, 2", + "[62] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2", + "[63] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 2", + "[64] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 3", + "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2021. 1", + "[66] Wenni Zheng, Pengbo Bo, Yang Liu, and Wenping Wang. Fast b-spline curve fitting by 1-bfgs. Computer Aided Geometric Design, 2012. 2" + ], + "bbox": [ + 501, + 92, + 890, + 744 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12574", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_model.json b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d612dfb2699faa355625c4b99b8391f3219c7aa2 --- /dev/null +++ b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.109, + 0.857, + 0.13 + ], + "angle": 0, + "content": "SketchINR: A First Look into Sketches as Implicit Neural Representations" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.137, + 0.868, + 0.26 + ], + "angle": 0, + "content": "Hmrishav Bandyopadhyay1 Ayan Kumar Bhunia1 Pinaki Nath Chowdhury1 Aneeshan Sain1 Tao Xiang1,2 Timothy Hospedales3 Yi-Zhe Song1,2 \n1SketchX, CVSSP, University of Surrey, United Kingdom. \n2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence. \n3University of Edinburgh, United Kingdom \n{h.bandyopadhyay, a.bhunia, p.chowdhury, a.sain, t.xiang, y-song}@surrey.ac.uk t.hospedales@ed.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.269, + 0.585, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.27, + 0.892, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 1. SketchINR is an implicit neural representation for sequential vector sketches. It is the first neural representation with sufficient fidelity to be a drop-in replacement for the raw sketch data (a). The representation is significantly higher fidelity than existing learned representations such as SketchRNN [28], especially for more complex sketches (b, left). SketchINR also provides state-of-the-art sketch compression, with a substantially more compact representation than either vector or raster sketches (b, right)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.487, + 0.314, + 0.503 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.508, + 0.474, + 0.841 + ], + "angle": 0, + "content": "We propose SketchINR, to advance the representation of vector sketches with implicit neural models. A variable length vector sketch is compressed into a latent space of fixed dimension that implicitly encodes the underlying shape as a function of time and strokes. The learned function predicts the \\( xy \\) point coordinates in a sketch at each time and stroke. Despite its simplicity, SketchINR outperforms existing representations at multiple tasks: (i) Encoding an entire sketch dataset into a fixed size latent vector, SketchINR gives \\( 60 \\times \\) and \\( 10 \\times \\) data compression over raster and vector sketches, respectively. (ii) SketchINR's auto-decoder provides a much higher-fidelity representation than other learned vector sketch representations, and is uniquely able to scale to complex vector sketches such as FS-COCO. (iii) SketchINR supports parallelisation that can decode/render \\( \\sim 100 \\times \\) faster than other learned vector representations such as SketchRNN. (iv) SketchINR, for the first time, emulates the human ability to reproduce a sketch with varying abstraction in terms of number and complexity of strokes. As a first look at implicit sketches, SketchINR's compact high-fidelity representation will support future work in modelling long and complex sketches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.849, + 0.21, + 0.863 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The prevalence of touch-screen devices has triggered significant research progress on sketches [4, 6, 8, 51, 57]. Mod-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.893, + 0.579 + ], + "angle": 0, + "content": "elling digital sketches has become an important challenge for learning systems that aim to stimulate human creativity [17, 27]. Large-scale datasets [24, 28, 53] motivated several downstream applications like image retrieval [15, 61], image generation [37, 65], image editing [1, 41], and 3D content creation [5, 8, 38], among others [58, 60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.793 + ], + "angle": 0, + "content": "These applications are underpinned by sketches captured in raw form either as raster or vector images, and usually encoded into derived representations by ConvNets [15], splines, and so on. For over a decade, discourse around what is a \"good\" representation of human sketches has persisted [31]. A now substantial body of work has focused on representations for sequential vector sketches that model both explicit strokes, and their drawing over time — most famously by the auto-regressive SketchRNN [28], but also using representations such as parametric curves [16], Gaussians [2], and Neural ODEs [17]. We introduce a novel continuous-time representation of sequential vector sketches by taking an implicit neural representation perspective on sketches for the first time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "First, we review the limitations of current sketch representations. The two most popular raw sketch representations are (i) raster sketch - a black and white image of a line drawing, and (ii) vector sketch - a temporal sequence of points and strokes. Raster sketch has a large but fixed storage (e.g., \\(256 \\times 256\\)). Its compatibility with ConvNets [64] made raster sketches popular but they lack the tempo-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12565" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "ral information (i.e., stroke order) necessary for generative modelling [17]. Vector sketches efficiently store the \\( xy \\)-coordinates of points and strokes (\\( N \\times 2 \\)). However, the storage increases with the length of sketch \\( N \\) (Fig. 1b). (iii) Raw vector sketches can be encoded by popular learned sketch representations such as SketchRNN [28] and others [16, 17, 39]. However, for longer (\\( N \\geq 1000 \\)) scene sketches, the quadratic computational cost for transformers becomes intractable [39] and recurrent nets become unstable [28]. Nevertheless, most of these still suffer from slow sequential autoregressive decoding, and they mostly cannot accurately reconstruct complex sketches with many strokes. To summarise, current sketch representations either discard temporal information (raster), scale poorly in storage size (vector), or become slow and inaccurate with complex sketches (learned autoregressive vector)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.518 + ], + "angle": 0, + "content": "To overcome the above issues, we propose an implicit neural representation for sketches - SketchINR. We piggyback on the well-established [8, 30, 34, 48] potential of implicit functions [22] to encode an underlying shape as a function of time/location. Different to standard implicits [43, 48], the input of sketches as parametric functions is a hierarchy of stroke sequence and point sequence. We develop SketchINR to generate points (xy-coordinates) given both time and stroke inputs. To represent multiple sketches, SketchINR learns a fixed size latent space and a conditional implicit function that generates xy-coordinates of any sketch instance given its corresponding latent code." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Despite its simplicity, SketchINR is competitive with existing sketch representations. (i) It can encode an entire sketch dataset into a compact latent space \\((e.g., \\mathbb{R}^{N \\times 512})\\) and decode using a fixed size function (e.g., an 8 layer MLP). Compared to raster sketches \\((256 \\times 256)\\), SketchINR provides \\(128 \\times\\) storage compression. While simple object-level vector sketches (length \\(\\leq 300\\)) have comparable storage with SketchINR (\\(\\mathbb{R}^{512}\\)), for practical scene-level [13] vector sketches (length \\(\\geq 1000\\)), SketchINR provides \\(10 \\times\\) better compression (Fig. 1b). (ii) For decoding/generation, SketchINR can parallelly predict the \\(xy\\)-coordinates for all time and strokes. In practice, this is \\(100 \\times\\) faster than an autoregressive vector sketch generator with length \\(\\geq 300\\). (iii) SketchINR can accurately represent complex sketches (Fig. 1a,b) unlike competitors that suffer from complexity limitations [16, 17] or the length limitations of auto-regression [28]. The encoding fidelity is sufficiently high that SketchINR provides a drop-in replacement for raw sketch data, while being more compact. (iv) SketchINR supports diverse applications spanning smooth latent space interpolation, and sketch generation, sketch completion. Uniquely, it also supports sketch abstraction - the human-like ability to replicate the essence of a sketch with a variable number and complexity of strokes [31, 54]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In summary, our contributions are: (i) We present the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.26 + ], + "angle": 0, + "content": "first implicit neural representation approach to vector sketch modelling, extending INRs as a function of time and strokes, and defining a training objective for learning them. (ii) SketchINR provides the first learned representation capable of representing complex sketches compactly and with high-fidelity (Fig. 1a). Based on our compact high-fidelity representation we introduce the task of sketch compression, and demonstrate excellent compression results (Fig. 1b)). (iii) We further demonstrate SketchINR's applicability to a variety of sketch related tasks including generation, interpolation, completion and abstraction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.272, + 0.642, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Sketch Representations: Digital sketches are predominantly captured as a function of 'time' [28], in a sequence of coordinates traced by an artist on a canvas. Recent works [24, 28] emphasise this temporal order of coordinates as an indicator of their relative importance in depicting a sketch-concept, as humans draw in a coarse-to-fine fashion. As such, they [2, 7, 18, 28] exploit temporal information in sketches for downstream tasks like sketch-assisted retrieval [7], generation [2, 18, 28] and modelling [38]. Vector sequences are further parameterised, for representation with parametric curves like Béziers [16, 57] and B-Splines [50, 66] for early-handwritten digit representation [50] to recent representations of complex structures like real-world objects [16, 57] and scenes [58]. Discarding temporal significance of strokes, vector-sketches can be converted to raster images by rendering coordinate strokes [9] on a 2D canvas. While raster sketches are less informative than their vector counterparts [28], they offer enhanced applicability by expanding beyond digital sketches to paper-sketches [18] where stroke order is already lost. Raster representations are encoded with translation-invariant networks like CNNs [63] and Vision-Transformers [8], making them extremely useful to represent object-level information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Non-Parametric Modelling: Both vector sequences and raster renderings of these sequences are non-parametric representations [16] of sketches. To preserve temporal stroke order, vector sketches are handled as sequences [28] of coordinates rather than unordered sets. Hence, they are encoded with position-aware networks like RNNs [28], LSTMs [15, 61], and Transformers [10, 51]. Without positional cues, raster sketches are processed by spatially-aware CNNs [63] and Vision Transformers [52]. Semantic encodings of non-parametric representations capture the expressive nature of sketch, allowing its use as a creative input for downstream segmentation [33], retrieval [7], and editing [41], as well as for interactive tasks like object-detection [14] and image-inpainting [62]. These encodings are further used in auto-regressive [28, 59] and VAE-like [10] modelling of vector coordinate sequences for the generation of textual-characters [1] and object sketches [28]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "12566" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Parametric Representation learning: Parametric Splines [20], such as Beizer curves [42], approximate vector sketches by per-sample fitting on individual sketch samples. Amortised frameworks bypass this per-sample optimisation by inferring curve control points [16] and degree [18] from individual stroke features. Recent works exploit the representative power of parametric curves for generation of simple characters [26] to complex object [16, 18, 57] and scene level sketches[58]. Both non-parametric and curve-fitted Bézier parametric representations capture the explicit form of a sketch as spatial and sequential forms." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.262, + 0.473, + 0.565 + ], + "angle": 0, + "content": "Beyond explicit representations of sketches, we learn visual implicits to capture spatial dynamics in vectorsketches. Learned implicits can be used for sketch reconstruction with particular control over abstraction through number of reconstructed strokes. Somewhat similar to implicit representations for vector sketches, CoSE [2] parameterises Gaussian Mixtures on individual strokes, learning one implicit for one stroke only. New implicits are generated auto-regressively from previous predictions for auto-regressive generation and completion of sketches. Unlike CoSE, we learn a single implicit for the entire sketch, thus having one implicit parameter to sample the sketch in one pass. While SketchODE [17] similarly parameterises Neural Ordinary Differential Equations [12] to capture vector-sketch dynamics, it's optimisation is extremely slow (as much as \\(120 \\times\\) slower) because of high time complexity in solving higher order differentials. Three dimensional implicits like Signed-Distance [48] and Occupancy [43] functions learned from 3D point-clouds are particularly analogous to our implicit representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Sketch Abstraction: The continuity of human cognition in visual perception [36] is directly reflected in how we sketch. This enables versatility [64] in sketch, allowing us to express a huge range of thoughts and visions directly on paper in the form of coarse-grained 'ideas' [32] and fine-grained 'objects' [64]. Sketch abstraction, as a direct inverse of this granularity, has been studied in depth as a function of (i) drawing time: humans draw most representative strokes first [28], and (ii) compactness: in a constrained stroke setting, salient strokes are prioritised [46]. As abstraction gives sketches a human touch, generative modelling of sketches is focused on abstraction towards making generated drawings \"more humane\" [19] and augmenting limited sketch datasets with simulated sketches from photos [57, 58]. Importantly, recent works [57] model abstraction as a function of strokes to explicitly control abstraction levels by restricting the number of generated strokes. Here, we represent sketches as visual implicits which allows us to reconstruct them at varying levels of abstraction and detail by controlling the number of reconstructed strokes. We demonstrate several downstream applications of this representation, particularly sketch compression, generation and completion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.636, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Overview: Vector sketches are captured as pen movements on a digital canvas, represented by coordinates \\((x,y) \\in [0,1]\\) with binary pen-states \\(\\delta \\in \\{0:\\) pen-down, \\(1:\\) pen-up\\(\\}\\) at those coordinates. A sketch \\(p = \\{(x_i,y_i,\\delta_i)\\}_{i=1}^N\\) comprises \\(N\\) vector way-points (coordinates), and is portrayed (rendered) by retracing these points on a 2D canvas. This process, termed as rasterisation produces a pixelated raster sketch \\(p^s\\) on the canvas. In this work, we leverage scalability [48] in implicit functions to learn a flexible representation for vector sketches with raster guidance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.274, + 0.832, + 0.291 + ], + "angle": 0, + "content": "3.1. Learning Visual Implicits for Sketches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.895, + 0.706 + ], + "angle": 0, + "content": "Vector sketches can be modelled as implicit, where a sketch is represented as function of time \\( f_{\\theta}(t_j) : \\mathbb{R} \\to \\mathbb{R}^{2+1} \\) parameterised by \\( \\theta \\). Each time-stamp \\( t_j \\in [0,1) \\) is mapped to a pen position in the form of \\( xy \\)-coordinates \\( (x,y) \\in \\mathbb{R}^2 \\) and binary pen-states \\( \\delta \\in \\mathbb{R}^1 \\). The entire sketch can be reconstructed as a vector sequence by sampling \\( t \\) from 0 to 1 with a learned \\( \\theta \\) as \\( \\mathbf{p} = \\{f_{\\theta}(0),\\dots,f_{\\theta}(t_j),\\dots,f_{\\theta}(\\frac{J-1}{J})\\} \\) for a total of \\( J \\) timestamps. This in itself is a flexible representation for a vector sketch, as it allows for sampling with arbitrary timestamps, thus potentially increasing or decreasing resolution of the resulting sketch by modulating number of strokes with sampling resolution \\( (t : \\{0,\\frac{1}{3},\\frac{2}{3}\\} \\mathrm{v/s} t : \\{0,\\frac{1}{10},\\dots,\\frac{9}{10}\\}) \\). For increased control over reconstruction granularity in the form of abstraction, we model \\( f_{\\theta} \\) on an additional variable in the form of strokestamps \\( \\{s_k\\}_{k=1}^K \\), representing the sketch as \\( \\hat{p} = \\{f_{\\theta}(0,0),\\dots,f_{\\theta}(t_j,s_k),\\dots,f_{\\theta}(\\frac{J-1}{J},\\frac{K-1}{K})\\} \\) for \\( K \\) strokes (Fig. 2). These strokestamps provide explicit control over pen-up and pen-down states, allowing us to reconstruct the sketch with re-defined granularity during inference (e.g. 7 or 9 strokes). Specifically, strokestamps represent the current stroke number as a fraction of the total number of strokes \\( K \\) which allows us to change strokes by changing \\( s_k \\). Pen-states \\( \\delta \\) are thus explicitly determined from strokestamps only by toggling when \\( s_k \\) changes value. Our practical implementation of sketch implicits follows:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.713, + 0.892, + 0.731 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(t _ {j}, \\Delta t, s _ {k}, \\Delta s\\right) = p _ {j} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.737, + 0.895, + 0.858 + ], + "angle": 0, + "content": "where, time instances are computed as \\( t_j = t_{j-1} + \\Delta t \\), \\( t_0 = 0 \\), and \\( \\Delta t = \\frac{1}{J} \\). Similarly, \\( K \\)-stroke instances are computed as \\( s_k = s_{k-1} + \\Delta s \\), \\( s_0 = 0 \\), and \\( \\Delta s = \\frac{1}{K} \\). To exploit bias in neural networks towards learning low frequency functions [44, 49], we smoothen out high frequency variations \\( (0 \\text{to} \\frac{1}{K}, 0 \\text{to} \\frac{1}{S}) \\) for timestamps and strokestamps by mapping them from \\( \\mathbb{R} \\) to a higher dimension \\( \\mathbb{R}^{2L} \\) using positional embeddings \\( \\mathrm{PE}(\\cdot) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.864, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {P E} (x) = \\left[ \\sin \\left(1 0 ^ {- 4 / L} x\\right), \\cos \\left(1 0 ^ {- 4 / L} x\\right), \\right. \\tag {2} \\\\ \\ldots , \\sin (1 0 ^ {- 4 L / L} x), \\cos (1 0 ^ {- 4 L / L} x) ] \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12567" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.182 + ], + "angle": 0, + "content": "for \\( x \\in \\{t_n, \\Delta t, s_k, \\Delta s\\} \\). We uniformly distribute \\( J \\) times-stamps across \\( K \\) strokes, where each stroke has \\( J / K \\) points. For data augmentation, we train \\( f_{\\theta} \\) with varying \\( J \\) and \\( K \\) (\\( \\pm 50\\% \\) of ground truth value). This helps \\( f_{\\theta} \\) to scale to different levels of abstraction, and number of points and strokes during inference." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.192, + 0.465, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Figure 2. We model sketch as a function \\( f_{\\theta}(t_j, s_k) \\) of \\( J \\) timestamps and \\( K \\) strokes. Timestamps \\( \\{t_a, \\ldots, t_b\\} \\) where \\( t_j \\in [0,1) \\) correspond to way-points for a specific stroke \\( s_k \\). Finally, \\( f_{\\theta} \\) with learned weights \\( \\theta \\) can be sampled with arbitrary number of strokes as \\( t_j \\in \\{0, \\frac{1}{J}, \\dots, \\frac{J - 1}{J}\\} \\) and \\( s_k \\in \\{0, \\frac{1}{5}, \\dots, \\frac{4}{5}\\} \\) or \\( s_k \\in \\{0, \\frac{1}{10}, \\dots, \\frac{9}{10}\\} \\), leading to increasing or decreasing abstraction for \\( K = 5 \\) or \\( K = 10 \\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.226, + 0.513 + ], + "angle": 0, + "content": "3.2. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.469, + 0.596 + ], + "angle": 0, + "content": "To fit our implicit function \\( f_{\\theta} \\) to a particular sketch instance \\( p = \\{(x_0, y_0, \\delta_0), \\ldots, (x_{M-1}, y_{M-1}, \\delta_{M-1})\\} \\) with \\( M \\) vector way-points \\( \\{p_m\\}_{m=0}^{M-1} \\), a naive approach would be to optimise \\( \\theta \\) with mean squared error against the ground truth vectors as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.603, + 0.47, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} = \\sum_ {m = 0} ^ {M} \\left| \\left| f _ {\\theta} \\left(\\frac {m}{M}, \\frac {k}{K}\\right) - p _ {m} \\right| \\right| _ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.833 + ], + "angle": 0, + "content": "where each predicted coordinate and pen-state is optimised directly against corresponding vector way-points from the ground truth. However, this rigid point-to-point matching enforces memorisation of particular points and strokes without any spatial understanding. As such, sampling with a different resolution at this stage yields noisy output that does not correspond to the raster sketch. To explicitly optimise for sampling flexibility, we introduce a visual loss, penalising sketch reconstructions that do not match visually. For this, we sample a template grid \\( G \\) of 2D coordinates and compute the region of influence of each vector-waypoint \\( p_m \\) of stroke \\( s_k \\) on coordinate \\( g \\) as an intensity map [3, 59]:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.839, + 0.469, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I} _ {k} (g, s _ {k}) = \\max _ {p _ {m} \\in s _ {k}, r \\in [ 0, 1 ]} \\exp (\\gamma | | g - r p _ {m} - (1 - r) p _ {m + 1} | | _ {2}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.904 + ], + "angle": 0, + "content": "where intensity \\( \\mathrm{I}_k \\) is smoothed with \\( \\gamma \\) for points \\( p_m \\) in stroke \\( s_k \\) where \\( m \\in \\{0, \\dots, \\frac{K - 1}{K}\\} \\). The intensity map" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "for the entire sketch is formulated as a summation of maps from all strokes for \\( k \\in (0,1] \\) for both the ground truth sketch vector \\( p \\) with \\( K \\) strokes and the arbitrarily sampled sketch \\( \\hat{p} \\) with \\( \\hat{K} \\) strokes as \\( \\hat{\\mathrm{I}}_k \\), composing the loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.164, + 0.892, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V} = \\left| \\left| \\sum_ {k = 0} ^ {\\frac {K - 1}{K}} \\mathrm {I} _ {k} - \\sum_ {k = 0} ^ {\\frac {\\hat {K} - 1}{\\hat {K}}} \\hat {\\mathrm {I}} _ {k} \\right| \\right| _ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Intensity Smoothing with \\(\\gamma\\): The sketch intensity map \\(I_k\\) helps us to ground vector sketches to raster definiteness (Eq. (4)) by a visual rendering of individual strokes where intensity in coordinates \\(g \\in G\\) drops exponentially with distance from \\(p_m \\in s_k\\). This exponential drop is controlled by a smoothing factor \\(\\gamma\\) as a hyperparameter, where (i) a low \\(\\gamma\\) yields thicker lines on \\(I_k\\) (low-quality sketch reconstructions) but offers better optimisation, while (ii) a high \\(\\gamma\\) yields more refined lines (higher quality reconstructions), making optimisation significantly slower (Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.435 + ], + "angle": 0, + "content": "For faster optimisation, we assist the visual penalty with a weighted MSE loss from Eq. (3) and a lower gamma, empirically set to \\(\\gamma = 150\\). The loss function weighted with \\(\\lambda_{\\mathrm{MSE}}\\), looks can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.444, + 0.892, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i m p l i c i t}} = \\mathcal {L} _ {V} ^ {\\gamma = 1 5 0} + \\lambda_ {\\mathrm {M S E}} \\cdot \\mathcal {L} _ {\\mathrm {M S E}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.566 + ], + "angle": 0, + "content": "where we set \\(\\lambda_{\\mathrm{MSE}}\\) to 0.7. This, however, inherently introduces bias in our optimisation to match the order of strokes with the ground truth order to some extent. Future works could ideally set \\(\\lambda_{\\mathrm{MSE}} = 0\\) if they obtain faster convergence from a better engineered implicit function \\(f_{\\theta}\\). Fig. 4a summarises our implicit sketch learning framework." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.589, + 0.909, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Figure 3. Effect of smoothing factor \\(\\gamma\\) on training: Reducing \\(\\gamma\\) leads to higher intensity in the surrounding region near stroke \\(s_k\\). This is similar to stroke dilation of intensity map \\(\\mathrm{I}_k\\). A lower \\(\\gamma\\) leads to stable training (plot on right) but lacks fine-grained details. A higher \\(\\gamma\\) gives a fine-grained sketch but is harder to train." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.795, + 0.803 + ], + "angle": 0, + "content": "3.3. Generalising to Multiple Sketches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To efficiently represent multiple sketch implicits with a single global and general function, we modify Eq. (1) to include a sketch descriptor \\(\\nu_{i}\\in \\mathbb{R}^{d}\\) for each sketch \\(i\\). The function \\(f_{\\theta}(\\nu_{i},t,\\Delta t,s_{k},\\Delta s)\\) is approximated with a fully connected auto-decoder network, similar to DeepSDF [48]. Thus, feature descriptors \\(\\nu_{i}\\) are hidden vectors, learned" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12568" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.07, + 0.092, + 0.288, + 0.106 + ], + "angle": 0, + "content": "(a): Single sketch Implicit Learning" + }, + { + "type": "image", + "bbox": [ + 0.073, + 0.108, + 0.357, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.092, + 0.558, + 0.106 + ], + "angle": 0, + "content": "(b): Multi-sketch auto-decoder" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.113, + 0.577, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.092, + 0.896, + 0.107 + ], + "angle": 0, + "content": "(c): Mapping to \\(\\mathcal{N}(0,\\mathbf{I})\\) for generative modelling" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.113, + 0.907, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.895, + 0.306 + ], + "angle": 0, + "content": "Figure 4. SketchINR model diagram. (a) Embedding a single vector sketch as an implicit model \\(F\\) mapping stroke and time \\((\\mathbf{s},\\mathbf{t})\\) to ink coordinate \\(\\mathbf{p}\\). (b) Embedding a vector sketch dataset as a shared decoder and set of latent vectors \\(\\nu\\). (c) Training a generative model for implicit sketches by generating latent codes \\(\\nu\\) using an encoder \\(E\\) that inputs raster sketches." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.32, + 0.47, + 0.396 + ], + "angle": 0, + "content": "jointly with the function parameters \\(\\theta\\) on a training dataset \\((i\\in D)\\). Similar to single sketch optimisation, we train on multiple samples with a combination of MSE and visual loss (Eq. (6)). Fig. 4b summarises our multi-sketch autodecoder through block diagrams." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.473, + 0.637 + ], + "angle": 0, + "content": "Generative Modelling: The generalised auto-decoder, enables simple generative modelling of sketches by generating their feature descriptors \\(\\nu_{i}\\in \\mathcal{V}\\), which allows us to create novel sketch implicits. Towards this, we train a Variational Auto-Encoder [35] as our generator, where we replace the decoder with our pre-trained implicit decoder. For variational inference, our encoder \\(E(\\cdot)\\) represents each raster sketch instance \\(i\\) as a set of \\(d\\) Gaussians with mean \\(\\mu_i\\in \\mathbb{R}^d\\) and variance \\(\\sigma_{i}\\in \\mathbb{R}^{d}\\) which are re-parameterised [35] to a latent encoding as \\(z_{i} = \\mu_{i} + \\sigma_{i}\\cdot \\epsilon\\) where \\(\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). This latent is then projected to the \\(d\\)-dimensional implicit latent space \\(\\nu_{i}\\in \\mathcal{V}\\). We train the encoder as a ResNet-18 network with (i) reconstruction loss from both our visual and vector spaces (Eq. (6)) and (ii) KL-divergence loss to bring the distribution closer to a unit normal \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The loss function for the encoder can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.651, + 0.47, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e n c}} = \\mathcal {L} _ {\\text {i m p l i c i t}} - \\beta \\cdot D _ {K L} \\left(q \\left(z _ {i} \\mid i\\right), \\mathcal {N} (\\mathbf {0}, \\mathbf {I})\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.471, + 0.771 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a weighing factor, set to 0.7 and \\(q(z_i|i)\\) represents the underlying probability distribution modelled with encoder \\(E\\) that infers \\(z_{i}\\) for a given sketch \\(i\\). Fig. 4c summarises our implicit sketch generative model. The decoder can be used for sketch generation, while the encoder can encode raster sketches for vectorisation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.208, + 0.803 + ], + "angle": 0, + "content": "4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Datasets: We learn implicits for sketches of various complexities and abstraction levels, ranging from highly complex scene sketches that correspond to a photo scenery [13] to abstract doodles [28] drawn from memory. Based off complexity in downstream applications, we use scene sketches to demonstrate self-reconstruction and sketch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.895, + 0.638 + ], + "angle": 0, + "content": "compression, while focusing on simpler object level sketches for more complex tasks like sketch mixing (interpolation) and generation for faster optimisation. Specifically, we use the FS-COCO [13], Sketchy [53], and QuickDraw! [28] datasets. (i) FS-COCO [13] contains 10,000 hand-drawn sketches of complex MS-COCO [40] scenes, drawn by amateurs without time constraints. Sketches in FS-COCO have on a median of 64 strokes, resulting in the sketch having \\(\\sim 3000\\) waypoints. (ii) Comparatively simpler than scene sketches, Sketchy [53] consists of \\(\\sim 75\\mathrm{K}\\) sketches of \\(12.5\\mathrm{K}\\) object photographs from 125 categories. These sketches similar to FS-COCO are drawn from reference photographs and have photo-sketch pairs. (iii) QuickDraw! [28] sketches are drawn from memory and in a constrained time settings (\\(< 20\\mathrm{s}\\)). Hence, these sketches (doodles) are abstract in nature, lacking any pre-defined configuration or orientation from reference photos. In addition to these three datasets, we use a vectorised form of the MNIST dataset, Vector-MNIST [17], consisting of \\(\\sim 10\\mathrm{K}\\) samples as a simpler dataset to demonstrate complex downstream applications like generation and interpolation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.65, + 0.88, + 0.667 + ], + "angle": 0, + "content": "4.1. SketchINR: A Compact High-Fidelity Codec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We begin by demonstrating that SketchINR provides a high fidelity neural representation for sketches. Specifically, we train SketchINR on complex variable length vector sketch datasets such as FS-COCO and Sketchy. We represent each sketch in these datasets as a fixed size vector with a dataset-specific decoder \\(f_{\\theta}\\). We then render each implicit sketch \\(\\nu_{i} \\in \\mathcal{V}\\) with the same number of strokes as the ground-truth sketch \\(\\mathbf{p}^i\\) to evaluate reconstruction quality. Following [17, 28], we measure reconstruction quality using (i) Chamfer Distance (CD), and (ii) retrieval accuracy (top-10) with a naive sketch retrieval network (ResNet-18) trained on ground-truth sketch vectors \\(\\mathbf{p}\\) and evaluated on rendered \\(\\hat{\\mathbf{p}}\\). The quantitative results in Tab. 1, show that SketchINR provides substantially higher fidelity encoding than alternative representations like RNN [28], Bézier strokes [16]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12569" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.183 + ], + "angle": 0, + "content": "
FS-COCO [13]Sketchy [53]
CDR@10Time (s)CDR@10Time(s)
SketchRNN [28]---0.1641.650.11
CoSE [2]---0.1445.820.10
SketchODE [17]0.415.171.680.1737.200.13
BézierSketch [16]0.572.961.220.2325.720.08
Ours0.01115.610.30.00857.290.0007
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.189, + 0.47, + 0.218 + ], + "angle": 0, + "content": "Table 1. Evaluating neural sketch representations' reconstruction quality (Chamfer distance \\(\\downarrow\\); recall@ \\(10\\uparrow\\)) and decoding speed." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.233, + 0.468, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.56 + ], + "angle": 0, + "content": "Figure 5. Qualitative reconstruction results for neural sketch representations. SketchINR is uniquely able to scale to sketches of Sketchy and FS-COCO complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.672 + ], + "angle": 0, + "content": "ODE dynamics [17], and stroke embeddings [2] particularly for extremely complex sketches in FS-COCO, where SketchRNN and CoSE fail entirely. Qualitative results in Fig. 5 show SketchINR to reconstruct simple and complex sketches with high fidelity while SketchRNN performs poorly for Sketchy and fails completely on FS-COCO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Sketch Compression: We next show that SketchINR's high-fidelity encoding is extremely compact, thus providing an efficient sketch codec. INRs store information implicitly [55] in the network weights \\((\\theta)\\) and latent codes \\((\\nu_{i})\\). To store or transmit a sketch dataset, we only stream weights \\(\\theta\\) once (fixed cost) and a compressed latent vector \\(\\nu_{i} \\in \\mathbb{R}^{D}\\) for each sketch (variable cost). Hence, the space complexity for \\(N_{D}\\) sketches with SketchINR is \\(|\\theta| + N_{D} \\times |\\nu_{i}|\\) bytes, where \\(\\nu_{i}\\) is 64 dimensional latent. For streaming a large number of sketches, we discount \\(\\theta\\) as negligible streaming overhead. Comparing to alternative representations, streaming \\(N_{D}\\) (i) binary raster (BR) sketches of size \\((N_{x} \\times N_{y})\\) take \\(N_{D} \\times N_{x} \\times N_{y}\\) bits, and (ii) vector scene sketches with \\(N\\) coordinates and pen states \\((\\mathbb{R}^{2+1})\\) take \\(N_{D} \\times N \\times 2 \\times 3\\) bytes in 16-bit floating precision (\\(N \\approx 3000\\) in FS-COCO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "[13]). Raster representations can be further stored by only storing the coordinates \\((\\mathbb{R}^2)\\) of pixels containing the sketch (i.e., inked pixels) as sparse binary rasters (SBR). Fig. 6 shows rate-distortion for storing 10,000 scene sketches [13] with \\(\\nu_{i}\\) varying from 128 to 8 dimensional latent, and quality for raster and vector forms dropped with downsampling and RDP simplification [23] respectively. Quality metrics like PSNR and SSIM are more suited to photos. We use Chamfer Distance as a quality metric for vector sketch and observe \\(\\sim 10\\times\\) and \\(\\sim 60\\times\\) more efficiency than vector and raster sketches, respectively. Future works can further optimise our models and thus reduce storage requirements via model compression [21, 29, 56] and quantisation [11]." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.302, + 0.891, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.487 + ], + "angle": 0, + "content": "Figure 6. Rate Distortion: SketchINR can encode complex scene sketches from FS-COCO [13] in highly compact \\((\\mathbb{R}^{64})\\) latent codes. Specifically, despite nearly identical sketch quality (low CD represents higher fidelity), SketchINR has \\(\\sim 60\\times\\) lower BPP than PNG raster sketches and \\(\\sim 10\\times\\) lower than vectors sketches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Intra-Sketch Variation: Unlike prior representations (RNN [28], ODE dynamics [17]), humans do not draw exact stroke-level replica of a raster sketch. Instead, we choose the level of detail, the number of points, and the strokes used to convey the underlying shape. Given a latent code \\(\\nu_{i}\\), SketchINR allows users to explicitly choose \\(N\\) and \\(K\\), where the number of points \\(p = \\{p_1,\\dots ,p_N\\}\\) and strokes \\(s = \\{s_1,\\ldots ,s_K\\}\\). Choosing a low \\(N\\) and \\(K\\) leads to abstract self-reconstructions with fewer points and strokes. Fig. 7(top to bottom) shows that we can control \\(K\\) to decrease or increase the level detail in the sketch." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.682, + 0.885, + 0.699 + ], + "angle": 0, + "content": "4.2. Latent Space Interpolation (Creative Mixing)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Interpolation of learned latents allows us to explore latent space continuity. Auto-regressive modelling of sketches builds poor latents as explored by Das et. al. [19] due to the lack of a full visibility of the sketch at any decoding timestep. Specifically, auto-regressive models use the partially predicted sketch along with the sketch's latent representation to predict new way-points with consistency. This, however, introduces noise in the latent \\(\\rightarrow\\) sketch decoding so that small changes in these latents lead to big changes in the final sketch. Deterministic latent to sketch decoding solves this problem as noise is not added at any point, leading to meaningful sketch interpolation from corresponding latent space interpolations. We visualise in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12570" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.465, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.264, + 0.471, + 0.306 + ], + "angle": 0, + "content": "Figure 7. Given a latent code \\(\\nu_{i}\\), we can reconstruct a sketch using learned \\(f_{\\theta}\\) at multiple resolutions by controlling the number of strokes \\(K\\) during sampling." + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.314, + 0.433, + 0.326 + ], + "angle": 0, + "content": "Cross-Sketch Interpolation with auto-regressive networks - SketchRNN" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.331, + 0.457, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.414, + 0.353, + 0.426 + ], + "angle": 0, + "content": "Cross-Sketch interpolation with implicits - SketchINR" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.433, + 0.457, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.542 + ], + "angle": 0, + "content": "Figure 8. Latent space interpolation shows continuity and translation from one sketch concept to another." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Fig. 8 how one sketch morphs into another by reconstructing from \\( f_{\\theta}(\\nu', t_n, \\Delta t, s_k, \\Delta s) \\), where \\( \\nu' \\) is sampled along a latent walk: \\( \\nu' = (1 - \\delta)\\nu_1 + \\delta \\nu_2 \\), for \\( \\delta \\in [0,1] \\). Interpolating between two sketches in the continuous latent space gives creative mixing – modifying the sketch of airplane-1 (facing right) into airplane-2 (facing left) by varying \\( \\delta \\). Fig. 8 shows that small changes in the latent (columns) leads to smoother changes in sketch space for SketchINR. Meanwhile SketchRNN can exhibit big jumps that can zig-zag over successive latent increments, as indicated by an example tracked point in the figure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.391, + 0.757 + ], + "angle": 0, + "content": "4.3. INR inversion for sketch completion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Neural networks used to approximate implicit representations have been inverted in parallel literature [30] to appropriate samples not seen during training. This inversion, to arrive at a latent \\(\\nu\\) that describes the new sample, is usually made possible by learned priors in \\(f_{\\theta}\\). Specifically, an implicit decoder \\(f_{\\theta}\\) is frozen with latent \\(\\nu\\) optimised for \\(f_{\\theta}(\\nu, \\ldots)\\) against the ground truth. This behaviour is easily replicated in SketchINR, where we optimise the latent \\(\\nu\\) to describe sketch samples (Fig. 5). To implement" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.085, + 0.864, + 0.098 + ], + "angle": 0, + "content": "(a) Input Partial Stroke Vector-MNIST (Left) and Quick-Draw! (Right)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.098, + 0.885, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.144, + 0.783, + 0.155 + ], + "angle": 0, + "content": "Sketch Completion using SketchRNN" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.154, + 0.885, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.2, + 0.78, + 0.212 + ], + "angle": 0, + "content": "Sketch Completion using SketchINR" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.212, + 0.882, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.265, + 0.872, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.309, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Figure 9. We perform sketch completion by inverting learned implicit representations and decoding for occluded strokes or waypoints. (a) Temporal completion (b) Unlike auto-regressive methods like SketchRNN [28], we can perform a-temporal (or inverse-temporal) completion, where given the second half of a partial sketch, we can successfully reconstruct the first part." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.893, + 0.695 + ], + "angle": 0, + "content": "sketch completion, we obtain a sketch descriptor \\(\\nu\\) from a partial sketch by just optimising for specific time-steps (e.g. 1/4th of a sketch; \\(t \\in [0, \\frac{1}{4})\\)) available in a partial doodle. Then, by sampling for rest of the timesteps we can complete the incomplete sketch. We summarise some qualitative sketch completion results in Fig. 9(a). Completed sketches are varied because of randomness in optimisation, as partial sketches are often ambiguous. INR inversion thus provides a novel approach to sketch completion, previously dominated by auto-regressive and models. Unlike auto-regressive models SketchINR can also perform a-temporal completion, inferring the first half of a sketch given the second half of the sketch (Fig. 9b). We remark that popular INR applications such as super-resolution [47] and point-cloud densification [48] perform interpolation after inversion on sparse inputs (e.g.: \\(t \\in [0, 1)\\) given \\(t \\in \\{0, \\frac{1}{4}, \\frac{1}{2}, \\frac{3}{4}\\}\\)), while sketch completion requires extrapolation (e.g.: \\(t \\in [0, 1)\\) given \\(t \\in [0, \\frac{1}{4})\\)), making it more analogous to visual outpainting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.681, + 0.723 + ], + "angle": 0, + "content": "4.4. Sketch Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.778 + ], + "angle": 0, + "content": "We can perform conditional and unconditional sequential vector sketch generation after learning a generative model for sketch latents as described in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Unconditional Generation: We unconditionally generate sketches by random sampling \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) with the learnt VAE latent space, obtaining novel sketch descriptors \\(\\nu_{i}\\in \\mathcal{V}\\). We further sample \\(f_{\\theta}(\\nu_{i},\\dots)\\) with varying timestamps \\(T\\in [100,300]\\) and strokestamps \\(K\\in [10,30]\\) obtaining sketches at varying levels of abstraction. Qualitative results in Fig. 10 show diverse and plausible samples from Vector-MNIST [17] and Quick-Draw! [28] datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.47, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.471, + 0.263 + ], + "angle": 0, + "content": "Figure 10. Unconditional sketch generation. Sampling sketch latents \\(\\nu\\) as well as stroke and point complexity \\((K, T)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Conditional Generation: To condition sketch generation, we use our pre-trained VAE encoder \\( E \\) to embed raster sketches, which can then be vectorised by the decoder \\( f_{\\theta} \\). Fig. 11 shows qualitative results for sequential vector sketch generation from raster images, where we obtain comparable results with auto-regressive generative modelling [28]. We observe that treating vectorisation as a generative problem rather than a deterministic raster-to-vector mapping [7] leads us to have variations in generated vectors particularly in the form of stroke order resembling real copies of sketches made by humans." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.608 + ], + "angle": 0, + "content": "Figure 11. Conditional sketch generation (vectorisation) on Vector-MNIST. Raster sketches (top left) can have multiple vector forms, e.g., sketched with different numbers of strokes (bottom)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.308, + 0.636 + ], + "angle": 0, + "content": "4.5. Ablations and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Global vs Local Time Modelling: We model \\(J\\) timestamps across \\(K\\) strokes, by distributing timestamps uniformly over each stroke such that each stroke has \\(\\frac{J}{K}\\) timestamps. In other words, time \\((t,\\Delta t)\\) is modelled globally for the entire sketch \\(\\mathbf{p} = \\{\\mathrm{p}_1,\\dots ,\\mathrm{p}_N\\}\\). An alternative is to model \\((t,\\Delta t)\\) as 0to1 for each stroke separately (local time-modelling) instead of the entire sketch. Particularly, each of the \\(K\\) strokes are modelled using \\(N\\) points, representing the entire sketch using \\(N\\cdot K\\) points as \\(\\mathbf{p} = \\{\\mathrm{p}_1,\\dots ,\\mathrm{p}_{N\\cdot K}\\}\\). Visually inspecting global vs local time embedding shows that local modelling leads to jittery sketches (see Fig. 12), whereas global modelling gives visually smooth sketches. We hypothesise that time is a global property of a sketch and not an independent property localised to each stroke. Hence, modelling time locally puts an additional optimisation burden on our implicit decoder to ensure smoothness and continuity, resulting in poorly reconstructed sketches." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.885, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.16, + 0.892, + 0.188 + ], + "angle": 0, + "content": "Figure 12. Global vs Local time embedding: Global embedding gives smooth sketches compared to local, which gives jittering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.895, + 0.437 + ], + "angle": 0, + "content": "Fixed vs Variable Smoothing Factor \\((\\gamma)\\): The intensity map \\(\\mathrm{I}_k\\) in Eq. (4) guides our raster-based visual loss \\(\\mathcal{L}_V\\). A key component in \\(\\mathrm{I}_k\\) is the smoothing parameter \\(\\gamma -\\) low \\(\\gamma\\) (thicker \\(\\mathrm{I}_k\\)) gives less accurate but smoother values on the entire visual region (easier training); high \\(\\gamma\\) (thinner \\(\\mathrm{I}_k\\)) gives more accurate but sharper (steeper) intensity-gradients near ground-truth strokes (harder training). While we use a fixed \\(\\gamma\\) with MSE Loss in Eq. (6) to balance out the optimisation complexity, we note that an alternative is to vary \\(\\gamma\\) every fixed number of steps with a scheduler to allow faster optimisation. Specifically, we vary \\(\\gamma\\) as 20to200 on a linear scale by incremental \\(\\Delta \\gamma\\) every few iterations. Despite its theoretical advantage [25, 45], our initial experiments suggest only minor improvement that reduces training time by \\(6.1\\%\\). A detailed study of bundle-adjustment [45] for visual loss is an interesting future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.682 + ], + "angle": 0, + "content": "Limitations and Future Work Despite it's efficiency as a representation for hand-drawn sketches, SketchINR suffers from a number of limitations. We naturally share limitations with other implicit representations including optimisation-based encoding and poor cross category generalisation. While unlike Sketch-RNN our generation network can be trained on 2-3 categories at once, it fails to generalise on more. In addition, a core limitation lies in slow convergence due to pixel space optimisation. While vector-space loss in the form of mean squared error mitigates this to some extent (Fig. 3), convergence is still slow and can possibly be improved by better engineering the implicit function. Strokes could also be smoothed by explicit inclusion of stroke gradient (slope) based regularisation in future work. Finally, vectorisation for conditional generation performs poorly on high complexity sketches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.687, + 0.62, + 0.702 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We introduce visual implicits to represent vector sketches with compressed latent descriptors. This neural representation provides a high-fidelity and compact representation that raises the possibility of a sketch-specific codec for compactly representing large sketch datasets. Our SketchINR can decode sketches at varying levels of detail with controllable number of strokes, and provides superior cross-sketch interpolation between implicits, demonstrating a smoother latent space than auto-regressive models. Applications such as sketch-completion are also supported, including a-temporal completion not available with auto-regressive models. Decoding is inherently parallel and can be over \\(100 \\times\\) faster than autoregressive models in practice." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12572" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Emre Aksan, Fabrizio Pece, and Otmar Hilliges. Deepwriting: Making digital ink editable via deep generative modeling. In CHI, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.198 + ], + "angle": 0, + "content": "[2] Emre Aksan, Thomas Deselaers, Andrea Tagliasacchi, and Otmar Hilliges. Cose: Compositional stroke embeddings. In NeurIPS, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.242 + ], + "angle": 0, + "content": "[3] Stephan Alaniz, Massimiliano Mancini, Anjan Dutta, Diego Marcos, and Zeynep Akata. Abstracting sketches through simple primitives. In ECCV, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[4] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[5] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[6] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can beat me? In SIGGRAPH Asia, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[8] Alexandre Binninger, Amir Hertz, Olga Sorkine-Hornung, Daniel Cohen-Or, and Raja Giryes. Sens: Part-aware sketch-based implicit neural shape modeling. arXiv preprint arXiv:2306.06088, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[9] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 1965. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[10] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[11] Qualcomm Innovation Center. Ai model efficiency toolkit (aimet). https://github.com/quic/aimet, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[12] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. In NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[13] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aaneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What can human sketches do for object detection? In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[15] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Sketchode: Learning neural sketch representation in continuous time. In ICLR, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[18] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Cloud2curve: Generation and vectorization of parametric sketches. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[19] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In ICLR, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.267 + ], + "angle": 0, + "content": "[20] Carl De Boor and Carl De Boor. A practical guide to splines. Applied Mathematical Sciences, 1978. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.311 + ], + "angle": 0, + "content": "[21] Zhen Dong, Zhewei Yao, Amir Gholami, Michael Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In ICCV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.312, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[22] Asen L. Dontchev and R. Tyrrell Rockafellar. Implicit Functions and Solution Mappings. Springer-Verlag, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[23] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 1973. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[24] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[25] F. Dan Foresee and Martin T. Hagan. Gauss-newton approximation to bayesian learning. In ICNN, 1997. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[26] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, SM Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[27] Songwei Ge, Vedanuj Goswami, C. Lawrence Zitnick, and Devi Parikh. Creative sketch generation. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.591 + ], + "angle": 0, + "content": "[28] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.593, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[29] Hai Victor Habi, Roy H. Jennings, and Arnon Netzer. Hmq: Hardware friendly mixed precision quantization block for cnns. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.68 + ], + "angle": 0, + "content": "[30] Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. Spaghetti: Editing implicit shapes through part aware generation. ACM TOG, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.681, + 0.892, + 0.709 + ], + "angle": 0, + "content": "[31] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.711, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[32] Conghui Hu, Da Li, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Sketch-a-classifier: Sketch-based photo classifier generation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.755, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[33] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.799, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[34] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.842, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[36] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12573" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[37] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[38] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Free2cad: Parsing freehand drawings into cad commands. ACM TOG, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[39] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.469, + 0.363 + ], + "angle": 0, + "content": "[41] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Deflocnet: Deep image editing via flexible low-level controls. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[42] Asif Masood and Sidra Ejaz. An efficient algorithm for robust curve fitting using cubicBezier curves. In ICIC, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.411, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[43] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[45] Jorge J. Mofe. The levenberg-marquardt algorithm. In Numerical Analysis, 1976. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[46] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.6, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[47] Quan H Nguyen and William J Beksi. Single image superresolution via a dual interactive implicit neural network. In WACV, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[49] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[50] Michael Revow, Christopher KI Williams, and Geoffrey E Hinton. Using generative models for handwritten digit recognition. IEEE TPAMI, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[51] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[52] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[53] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[54] Heping Sheng, John Wilder, and Dirk B. Walther. Where to draw the line? PLOS One, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[55] Yannick Strumpler, Janis Postels, Ren Yang, Luc van Gool, and Federico Tombari. Implicit neural representations for image compression. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[56] Stefan Uhlich, Lukas Mauch, Fabien Cardinaux, Kazuki Yoshiyama, Javier Alonso Garcia, Stephen Tiedemann, Thomas Kemp, and Akira Nakamura. Mixed precision dnns: All you need is a good parametrization. In ICLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[57] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM TOG, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[58] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In ICCV, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[59] Alexander Wang, Mengye Ren, and Richard Zemel. Sketchembednet: Learning novel concepts by imitating drawings. In ICML, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[60] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[61] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[62] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[63] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[64] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[66] Wenni Zheng, Pengbo Bo, Yang Liu, and Wenping Wang. Fast b-spline curve fitting by 1-bfgs. Computer Aided Geometric Design, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12574" + } + ] +] \ No newline at end of file diff --git a/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_origin.pdf b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac1118c94edf124b6aa364e7b2b743ad127a4c10 --- /dev/null +++ b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/ab26a978-e483-4292-89b7-9fe0134cf389_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c90dae50411ea7a52429bae0b1b2c0ad47999851e61dbf8c025627c1ac999de +size 6887947 diff --git a/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/full.md b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1a983618b31c26308498f520daea12b6b74fe5a8 --- /dev/null +++ b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/full.md @@ -0,0 +1,296 @@ +# SketchINR: A First Look into Sketches as Implicit Neural Representations + +Hmrishav Bandyopadhyay1 Ayan Kumar Bhunia1 Pinaki Nath Chowdhury1 Aneeshan Sain1 Tao Xiang1,2 Timothy Hospedales3 Yi-Zhe Song1,2 +1SketchX, CVSSP, University of Surrey, United Kingdom. +2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence. +3University of Edinburgh, United Kingdom +{h.bandyopadhyay, a.bhunia, p.chowdhury, a.sain, t.xiang, y-song}@surrey.ac.uk t.hospedales@ed.ac.uk + +![](images/0326cb215bf5f979385ad4ca53dc3f4c8d7fb3a7bd4bed2acdb998828504ebc5.jpg) +Figure 1. SketchINR is an implicit neural representation for sequential vector sketches. It is the first neural representation with sufficient fidelity to be a drop-in replacement for the raw sketch data (a). The representation is significantly higher fidelity than existing learned representations such as SketchRNN [28], especially for more complex sketches (b, left). SketchINR also provides state-of-the-art sketch compression, with a substantially more compact representation than either vector or raster sketches (b, right). + +![](images/0b2dc68e7f26fc593d940f54567ddc9b80140fadfe77ac3e58b593325520dec8.jpg) + +# Abstract + +We propose SketchINR, to advance the representation of vector sketches with implicit neural models. A variable length vector sketch is compressed into a latent space of fixed dimension that implicitly encodes the underlying shape as a function of time and strokes. The learned function predicts the $xy$ point coordinates in a sketch at each time and stroke. Despite its simplicity, SketchINR outperforms existing representations at multiple tasks: (i) Encoding an entire sketch dataset into a fixed size latent vector, SketchINR gives $60 \times$ and $10 \times$ data compression over raster and vector sketches, respectively. (ii) SketchINR's auto-decoder provides a much higher-fidelity representation than other learned vector sketch representations, and is uniquely able to scale to complex vector sketches such as FS-COCO. (iii) SketchINR supports parallelisation that can decode/render $\sim 100 \times$ faster than other learned vector representations such as SketchRNN. (iv) SketchINR, for the first time, emulates the human ability to reproduce a sketch with varying abstraction in terms of number and complexity of strokes. As a first look at implicit sketches, SketchINR's compact high-fidelity representation will support future work in modelling long and complex sketches. + +# 1. Introduction + +The prevalence of touch-screen devices has triggered significant research progress on sketches [4, 6, 8, 51, 57]. Mod- + +elling digital sketches has become an important challenge for learning systems that aim to stimulate human creativity [17, 27]. Large-scale datasets [24, 28, 53] motivated several downstream applications like image retrieval [15, 61], image generation [37, 65], image editing [1, 41], and 3D content creation [5, 8, 38], among others [58, 60]. + +These applications are underpinned by sketches captured in raw form either as raster or vector images, and usually encoded into derived representations by ConvNets [15], splines, and so on. For over a decade, discourse around what is a "good" representation of human sketches has persisted [31]. A now substantial body of work has focused on representations for sequential vector sketches that model both explicit strokes, and their drawing over time — most famously by the auto-regressive SketchRNN [28], but also using representations such as parametric curves [16], Gaussians [2], and Neural ODEs [17]. We introduce a novel continuous-time representation of sequential vector sketches by taking an implicit neural representation perspective on sketches for the first time. + +First, we review the limitations of current sketch representations. The two most popular raw sketch representations are (i) raster sketch - a black and white image of a line drawing, and (ii) vector sketch - a temporal sequence of points and strokes. Raster sketch has a large but fixed storage (e.g., $256 \times 256$ ). Its compatibility with ConvNets [64] made raster sketches popular but they lack the tempo- + +ral information (i.e., stroke order) necessary for generative modelling [17]. Vector sketches efficiently store the $xy$ -coordinates of points and strokes ( $N \times 2$ ). However, the storage increases with the length of sketch $N$ (Fig. 1b). (iii) Raw vector sketches can be encoded by popular learned sketch representations such as SketchRNN [28] and others [16, 17, 39]. However, for longer ( $N \geq 1000$ ) scene sketches, the quadratic computational cost for transformers becomes intractable [39] and recurrent nets become unstable [28]. Nevertheless, most of these still suffer from slow sequential autoregressive decoding, and they mostly cannot accurately reconstruct complex sketches with many strokes. To summarise, current sketch representations either discard temporal information (raster), scale poorly in storage size (vector), or become slow and inaccurate with complex sketches (learned autoregressive vector). + +To overcome the above issues, we propose an implicit neural representation for sketches - SketchINR. We piggyback on the well-established [8, 30, 34, 48] potential of implicit functions [22] to encode an underlying shape as a function of time/location. Different to standard implicits [43, 48], the input of sketches as parametric functions is a hierarchy of stroke sequence and point sequence. We develop SketchINR to generate points (xy-coordinates) given both time and stroke inputs. To represent multiple sketches, SketchINR learns a fixed size latent space and a conditional implicit function that generates xy-coordinates of any sketch instance given its corresponding latent code. + +Despite its simplicity, SketchINR is competitive with existing sketch representations. (i) It can encode an entire sketch dataset into a compact latent space $(e.g., \mathbb{R}^{N \times 512})$ and decode using a fixed size function (e.g., an 8 layer MLP). Compared to raster sketches $(256 \times 256)$ , SketchINR provides $128 \times$ storage compression. While simple object-level vector sketches (length $\leq 300$ ) have comparable storage with SketchINR ( $\mathbb{R}^{512}$ ), for practical scene-level [13] vector sketches (length $\geq 1000$ ), SketchINR provides $10 \times$ better compression (Fig. 1b). (ii) For decoding/generation, SketchINR can parallelly predict the $xy$ -coordinates for all time and strokes. In practice, this is $100 \times$ faster than an autoregressive vector sketch generator with length $\geq 300$ . (iii) SketchINR can accurately represent complex sketches (Fig. 1a,b) unlike competitors that suffer from complexity limitations [16, 17] or the length limitations of auto-regression [28]. The encoding fidelity is sufficiently high that SketchINR provides a drop-in replacement for raw sketch data, while being more compact. (iv) SketchINR supports diverse applications spanning smooth latent space interpolation, and sketch generation, sketch completion. Uniquely, it also supports sketch abstraction - the human-like ability to replicate the essence of a sketch with a variable number and complexity of strokes [31, 54]. + +In summary, our contributions are: (i) We present the + +first implicit neural representation approach to vector sketch modelling, extending INRs as a function of time and strokes, and defining a training objective for learning them. (ii) SketchINR provides the first learned representation capable of representing complex sketches compactly and with high-fidelity (Fig. 1a). Based on our compact high-fidelity representation we introduce the task of sketch compression, and demonstrate excellent compression results (Fig. 1b)). (iii) We further demonstrate SketchINR's applicability to a variety of sketch related tasks including generation, interpolation, completion and abstraction. + +# 2. Related Work + +Sketch Representations: Digital sketches are predominantly captured as a function of 'time' [28], in a sequence of coordinates traced by an artist on a canvas. Recent works [24, 28] emphasise this temporal order of coordinates as an indicator of their relative importance in depicting a sketch-concept, as humans draw in a coarse-to-fine fashion. As such, they [2, 7, 18, 28] exploit temporal information in sketches for downstream tasks like sketch-assisted retrieval [7], generation [2, 18, 28] and modelling [38]. Vector sequences are further parameterised, for representation with parametric curves like Béziers [16, 57] and B-Splines [50, 66] for early-handwritten digit representation [50] to recent representations of complex structures like real-world objects [16, 57] and scenes [58]. Discarding temporal significance of strokes, vector-sketches can be converted to raster images by rendering coordinate strokes [9] on a 2D canvas. While raster sketches are less informative than their vector counterparts [28], they offer enhanced applicability by expanding beyond digital sketches to paper-sketches [18] where stroke order is already lost. Raster representations are encoded with translation-invariant networks like CNNs [63] and Vision-Transformers [8], making them extremely useful to represent object-level information. + +Non-Parametric Modelling: Both vector sequences and raster renderings of these sequences are non-parametric representations [16] of sketches. To preserve temporal stroke order, vector sketches are handled as sequences [28] of coordinates rather than unordered sets. Hence, they are encoded with position-aware networks like RNNs [28], LSTMs [15, 61], and Transformers [10, 51]. Without positional cues, raster sketches are processed by spatially-aware CNNs [63] and Vision Transformers [52]. Semantic encodings of non-parametric representations capture the expressive nature of sketch, allowing its use as a creative input for downstream segmentation [33], retrieval [7], and editing [41], as well as for interactive tasks like object-detection [14] and image-inpainting [62]. These encodings are further used in auto-regressive [28, 59] and VAE-like [10] modelling of vector coordinate sequences for the generation of textual-characters [1] and object sketches [28]. + +Parametric Representation learning: Parametric Splines [20], such as Beizer curves [42], approximate vector sketches by per-sample fitting on individual sketch samples. Amortised frameworks bypass this per-sample optimisation by inferring curve control points [16] and degree [18] from individual stroke features. Recent works exploit the representative power of parametric curves for generation of simple characters [26] to complex object [16, 18, 57] and scene level sketches[58]. Both non-parametric and curve-fitted Bézier parametric representations capture the explicit form of a sketch as spatial and sequential forms. + +Beyond explicit representations of sketches, we learn visual implicits to capture spatial dynamics in vectorsketches. Learned implicits can be used for sketch reconstruction with particular control over abstraction through number of reconstructed strokes. Somewhat similar to implicit representations for vector sketches, CoSE [2] parameterises Gaussian Mixtures on individual strokes, learning one implicit for one stroke only. New implicits are generated auto-regressively from previous predictions for auto-regressive generation and completion of sketches. Unlike CoSE, we learn a single implicit for the entire sketch, thus having one implicit parameter to sample the sketch in one pass. While SketchODE [17] similarly parameterises Neural Ordinary Differential Equations [12] to capture vector-sketch dynamics, it's optimisation is extremely slow (as much as $120 \times$ slower) because of high time complexity in solving higher order differentials. Three dimensional implicits like Signed-Distance [48] and Occupancy [43] functions learned from 3D point-clouds are particularly analogous to our implicit representation. + +Sketch Abstraction: The continuity of human cognition in visual perception [36] is directly reflected in how we sketch. This enables versatility [64] in sketch, allowing us to express a huge range of thoughts and visions directly on paper in the form of coarse-grained 'ideas' [32] and fine-grained 'objects' [64]. Sketch abstraction, as a direct inverse of this granularity, has been studied in depth as a function of (i) drawing time: humans draw most representative strokes first [28], and (ii) compactness: in a constrained stroke setting, salient strokes are prioritised [46]. As abstraction gives sketches a human touch, generative modelling of sketches is focused on abstraction towards making generated drawings "more humane" [19] and augmenting limited sketch datasets with simulated sketches from photos [57, 58]. Importantly, recent works [57] model abstraction as a function of strokes to explicitly control abstraction levels by restricting the number of generated strokes. Here, we represent sketches as visual implicits which allows us to reconstruct them at varying levels of abstraction and detail by controlling the number of reconstructed strokes. We demonstrate several downstream applications of this representation, particularly sketch compression, generation and completion. + +# 3. Methodology + +Overview: Vector sketches are captured as pen movements on a digital canvas, represented by coordinates $(x,y) \in [0,1]$ with binary pen-states $\delta \in \{0:$ pen-down, $1:$ pen-up $\}$ at those coordinates. A sketch $p = \{(x_i,y_i,\delta_i)\}_{i=1}^N$ comprises $N$ vector way-points (coordinates), and is portrayed (rendered) by retracing these points on a 2D canvas. This process, termed as rasterisation produces a pixelated raster sketch $p^s$ on the canvas. In this work, we leverage scalability [48] in implicit functions to learn a flexible representation for vector sketches with raster guidance. + +# 3.1. Learning Visual Implicits for Sketches + +Vector sketches can be modelled as implicit, where a sketch is represented as function of time $f_{\theta}(t_j) : \mathbb{R} \to \mathbb{R}^{2+1}$ parameterised by $\theta$ . Each time-stamp $t_j \in [0,1)$ is mapped to a pen position in the form of $xy$ -coordinates $(x,y) \in \mathbb{R}^2$ and binary pen-states $\delta \in \mathbb{R}^1$ . The entire sketch can be reconstructed as a vector sequence by sampling $t$ from 0 to 1 with a learned $\theta$ as $\mathbf{p} = \{f_{\theta}(0),\dots,f_{\theta}(t_j),\dots,f_{\theta}(\frac{J-1}{J})\}$ for a total of $J$ timestamps. This in itself is a flexible representation for a vector sketch, as it allows for sampling with arbitrary timestamps, thus potentially increasing or decreasing resolution of the resulting sketch by modulating number of strokes with sampling resolution $(t : \{0,\frac{1}{3},\frac{2}{3}\} \mathrm{v/s} t : \{0,\frac{1}{10},\dots,\frac{9}{10}\})$ . For increased control over reconstruction granularity in the form of abstraction, we model $f_{\theta}$ on an additional variable in the form of strokestamps $\{s_k\}_{k=1}^K$ , representing the sketch as $\hat{p} = \{f_{\theta}(0,0),\dots,f_{\theta}(t_j,s_k),\dots,f_{\theta}(\frac{J-1}{J},\frac{K-1}{K})\}$ for $K$ strokes (Fig. 2). These strokestamps provide explicit control over pen-up and pen-down states, allowing us to reconstruct the sketch with re-defined granularity during inference (e.g. 7 or 9 strokes). Specifically, strokestamps represent the current stroke number as a fraction of the total number of strokes $K$ which allows us to change strokes by changing $s_k$ . Pen-states $\delta$ are thus explicitly determined from strokestamps only by toggling when $s_k$ changes value. Our practical implementation of sketch implicits follows: + +$$ +f _ {\theta} \left(t _ {j}, \Delta t, s _ {k}, \Delta s\right) = p _ {j} \tag {1} +$$ + +where, time instances are computed as $t_j = t_{j-1} + \Delta t$ , $t_0 = 0$ , and $\Delta t = \frac{1}{J}$ . Similarly, $K$ -stroke instances are computed as $s_k = s_{k-1} + \Delta s$ , $s_0 = 0$ , and $\Delta s = \frac{1}{K}$ . To exploit bias in neural networks towards learning low frequency functions [44, 49], we smoothen out high frequency variations $(0 \text{to} \frac{1}{K}, 0 \text{to} \frac{1}{S})$ for timestamps and strokestamps by mapping them from $\mathbb{R}$ to a higher dimension $\mathbb{R}^{2L}$ using positional embeddings $\mathrm{PE}(\cdot)$ as + +$$ +\begin{array}{l} \mathrm {P E} (x) = \left[ \sin \left(1 0 ^ {- 4 / L} x\right), \cos \left(1 0 ^ {- 4 / L} x\right), \right. \tag {2} \\ \ldots , \sin (1 0 ^ {- 4 L / L} x), \cos (1 0 ^ {- 4 L / L} x) ] \\ \end{array} +$$ + +for $x \in \{t_n, \Delta t, s_k, \Delta s\}$ . We uniformly distribute $J$ times-stamps across $K$ strokes, where each stroke has $J / K$ points. For data augmentation, we train $f_{\theta}$ with varying $J$ and $K$ ( $\pm 50\%$ of ground truth value). This helps $f_{\theta}$ to scale to different levels of abstraction, and number of points and strokes during inference. + +![](images/1302058cfe7037f4cc43014ed0e7dce1a0670bd315cb43ae2fedc80c87f5fdc3.jpg) +Figure 2. We model sketch as a function $f_{\theta}(t_j, s_k)$ of $J$ timestamps and $K$ strokes. Timestamps $\{t_a, \ldots, t_b\}$ where $t_j \in [0,1)$ correspond to way-points for a specific stroke $s_k$ . Finally, $f_{\theta}$ with learned weights $\theta$ can be sampled with arbitrary number of strokes as $t_j \in \{0, \frac{1}{J}, \dots, \frac{J - 1}{J}\}$ and $s_k \in \{0, \frac{1}{5}, \dots, \frac{4}{5}\}$ or $s_k \in \{0, \frac{1}{10}, \dots, \frac{9}{10}\}$ , leading to increasing or decreasing abstraction for $K = 5$ or $K = 10$ , respectively. + +# 3.2. Loss functions + +To fit our implicit function $f_{\theta}$ to a particular sketch instance $p = \{(x_0, y_0, \delta_0), \ldots, (x_{M-1}, y_{M-1}, \delta_{M-1})\}$ with $M$ vector way-points $\{p_m\}_{m=0}^{M-1}$ , a naive approach would be to optimise $\theta$ with mean squared error against the ground truth vectors as: + +$$ +\mathcal {L} _ {\mathrm {M S E}} = \sum_ {m = 0} ^ {M} \left| \left| f _ {\theta} \left(\frac {m}{M}, \frac {k}{K}\right) - p _ {m} \right| \right| _ {2} \tag {3} +$$ + +where each predicted coordinate and pen-state is optimised directly against corresponding vector way-points from the ground truth. However, this rigid point-to-point matching enforces memorisation of particular points and strokes without any spatial understanding. As such, sampling with a different resolution at this stage yields noisy output that does not correspond to the raster sketch. To explicitly optimise for sampling flexibility, we introduce a visual loss, penalising sketch reconstructions that do not match visually. For this, we sample a template grid $G$ of 2D coordinates and compute the region of influence of each vector-waypoint $p_m$ of stroke $s_k$ on coordinate $g$ as an intensity map [3, 59]: + +$$ +\mathrm {I} _ {k} (g, s _ {k}) = \max _ {p _ {m} \in s _ {k}, r \in [ 0, 1 ]} \exp (\gamma | | g - r p _ {m} - (1 - r) p _ {m + 1} | | _ {2}) \tag {4} +$$ + +where intensity $\mathrm{I}_k$ is smoothed with $\gamma$ for points $p_m$ in stroke $s_k$ where $m \in \{0, \dots, \frac{K - 1}{K}\}$ . The intensity map + +for the entire sketch is formulated as a summation of maps from all strokes for $k \in (0,1]$ for both the ground truth sketch vector $p$ with $K$ strokes and the arbitrarily sampled sketch $\hat{p}$ with $\hat{K}$ strokes as $\hat{\mathrm{I}}_k$ , composing the loss as: + +$$ +\mathcal {L} _ {V} = \left| \left| \sum_ {k = 0} ^ {\frac {K - 1}{K}} \mathrm {I} _ {k} - \sum_ {k = 0} ^ {\frac {\hat {K} - 1}{\hat {K}}} \hat {\mathrm {I}} _ {k} \right| \right| _ {2} \tag {5} +$$ + +Intensity Smoothing with $\gamma$ : The sketch intensity map $I_k$ helps us to ground vector sketches to raster definiteness (Eq. (4)) by a visual rendering of individual strokes where intensity in coordinates $g \in G$ drops exponentially with distance from $p_m \in s_k$ . This exponential drop is controlled by a smoothing factor $\gamma$ as a hyperparameter, where (i) a low $\gamma$ yields thicker lines on $I_k$ (low-quality sketch reconstructions) but offers better optimisation, while (ii) a high $\gamma$ yields more refined lines (higher quality reconstructions), making optimisation significantly slower (Fig. 3). + +For faster optimisation, we assist the visual penalty with a weighted MSE loss from Eq. (3) and a lower gamma, empirically set to $\gamma = 150$ . The loss function weighted with $\lambda_{\mathrm{MSE}}$ , looks can be written as: + +$$ +\mathcal {L} _ {\text {i m p l i c i t}} = \mathcal {L} _ {V} ^ {\gamma = 1 5 0} + \lambda_ {\mathrm {M S E}} \cdot \mathcal {L} _ {\mathrm {M S E}} \tag {6} +$$ + +where we set $\lambda_{\mathrm{MSE}}$ to 0.7. This, however, inherently introduces bias in our optimisation to match the order of strokes with the ground truth order to some extent. Future works could ideally set $\lambda_{\mathrm{MSE}} = 0$ if they obtain faster convergence from a better engineered implicit function $f_{\theta}$ . Fig. 4a summarises our implicit sketch learning framework. + +![](images/f25d662a62327603735692a8289c6b17afa2685b24422537af4b8eb38e066618.jpg) +Figure 3. Effect of smoothing factor $\gamma$ on training: Reducing $\gamma$ leads to higher intensity in the surrounding region near stroke $s_k$ . This is similar to stroke dilation of intensity map $\mathrm{I}_k$ . A lower $\gamma$ leads to stable training (plot on right) but lacks fine-grained details. A higher $\gamma$ gives a fine-grained sketch but is harder to train. + +# 3.3. Generalising to Multiple Sketches + +To efficiently represent multiple sketch implicits with a single global and general function, we modify Eq. (1) to include a sketch descriptor $\nu_{i}\in \mathbb{R}^{d}$ for each sketch $i$ . The function $f_{\theta}(\nu_{i},t,\Delta t,s_{k},\Delta s)$ is approximated with a fully connected auto-decoder network, similar to DeepSDF [48]. Thus, feature descriptors $\nu_{i}$ are hidden vectors, learned + +![](images/fe7980f604d4fcbf9b3db3f358080564a901f7311acd41ed00da39b329de927c.jpg) +(a): Single sketch Implicit Learning + +![](images/6dcc40403b97c8c12ee7a18cf1a6536edc27118b5e77f87db5a294e3fb4d7828.jpg) +(b): Multi-sketch auto-decoder +Figure 4. SketchINR model diagram. (a) Embedding a single vector sketch as an implicit model $F$ mapping stroke and time $(\mathbf{s},\mathbf{t})$ to ink coordinate $\mathbf{p}$ . (b) Embedding a vector sketch dataset as a shared decoder and set of latent vectors $\nu$ . (c) Training a generative model for implicit sketches by generating latent codes $\nu$ using an encoder $E$ that inputs raster sketches. + +![](images/46cd639eaf6fb2e4abe776e82e3e5bd93b2bf9776689786754d0c428ddc7488b.jpg) +(c): Mapping to $\mathcal{N}(0,\mathbf{I})$ for generative modelling + +jointly with the function parameters $\theta$ on a training dataset $(i\in D)$ . Similar to single sketch optimisation, we train on multiple samples with a combination of MSE and visual loss (Eq. (6)). Fig. 4b summarises our multi-sketch autodecoder through block diagrams. + +Generative Modelling: The generalised auto-decoder, enables simple generative modelling of sketches by generating their feature descriptors $\nu_{i}\in \mathcal{V}$ , which allows us to create novel sketch implicits. Towards this, we train a Variational Auto-Encoder [35] as our generator, where we replace the decoder with our pre-trained implicit decoder. For variational inference, our encoder $E(\cdot)$ represents each raster sketch instance $i$ as a set of $d$ Gaussians with mean $\mu_i\in \mathbb{R}^d$ and variance $\sigma_{i}\in \mathbb{R}^{d}$ which are re-parameterised [35] to a latent encoding as $z_{i} = \mu_{i} + \sigma_{i}\cdot \epsilon$ where $\epsilon \in \mathcal{N}(\mathbf{0},\mathbf{I})$ . This latent is then projected to the $d$ -dimensional implicit latent space $\nu_{i}\in \mathcal{V}$ . We train the encoder as a ResNet-18 network with (i) reconstruction loss from both our visual and vector spaces (Eq. (6)) and (ii) KL-divergence loss to bring the distribution closer to a unit normal $\mathcal{N}(\mathbf{0},\mathbf{I})$ . The loss function for the encoder can be written as: + +$$ +\mathcal {L} _ {\mathrm {e n c}} = \mathcal {L} _ {\text {i m p l i c i t}} - \beta \cdot D _ {K L} \left(q \left(z _ {i} \mid i\right), \mathcal {N} (\mathbf {0}, \mathbf {I})\right) \tag {7} +$$ + +where $\beta$ is a weighing factor, set to 0.7 and $q(z_i|i)$ represents the underlying probability distribution modelled with encoder $E$ that infers $z_{i}$ for a given sketch $i$ . Fig. 4c summarises our implicit sketch generative model. The decoder can be used for sketch generation, while the encoder can encode raster sketches for vectorisation. + +# 4. Applications + +Datasets: We learn implicits for sketches of various complexities and abstraction levels, ranging from highly complex scene sketches that correspond to a photo scenery [13] to abstract doodles [28] drawn from memory. Based off complexity in downstream applications, we use scene sketches to demonstrate self-reconstruction and sketch + +compression, while focusing on simpler object level sketches for more complex tasks like sketch mixing (interpolation) and generation for faster optimisation. Specifically, we use the FS-COCO [13], Sketchy [53], and QuickDraw! [28] datasets. (i) FS-COCO [13] contains 10,000 hand-drawn sketches of complex MS-COCO [40] scenes, drawn by amateurs without time constraints. Sketches in FS-COCO have on a median of 64 strokes, resulting in the sketch having $\sim 3000$ waypoints. (ii) Comparatively simpler than scene sketches, Sketchy [53] consists of $\sim 75\mathrm{K}$ sketches of $12.5\mathrm{K}$ object photographs from 125 categories. These sketches similar to FS-COCO are drawn from reference photographs and have photo-sketch pairs. (iii) QuickDraw! [28] sketches are drawn from memory and in a constrained time settings ( $< 20\mathrm{s}$ ). Hence, these sketches (doodles) are abstract in nature, lacking any pre-defined configuration or orientation from reference photos. In addition to these three datasets, we use a vectorised form of the MNIST dataset, Vector-MNIST [17], consisting of $\sim 10\mathrm{K}$ samples as a simpler dataset to demonstrate complex downstream applications like generation and interpolation. + +# 4.1. SketchINR: A Compact High-Fidelity Codec + +We begin by demonstrating that SketchINR provides a high fidelity neural representation for sketches. Specifically, we train SketchINR on complex variable length vector sketch datasets such as FS-COCO and Sketchy. We represent each sketch in these datasets as a fixed size vector with a dataset-specific decoder $f_{\theta}$ . We then render each implicit sketch $\nu_{i} \in \mathcal{V}$ with the same number of strokes as the ground-truth sketch $\mathbf{p}^i$ to evaluate reconstruction quality. Following [17, 28], we measure reconstruction quality using (i) Chamfer Distance (CD), and (ii) retrieval accuracy (top-10) with a naive sketch retrieval network (ResNet-18) trained on ground-truth sketch vectors $\mathbf{p}$ and evaluated on rendered $\hat{\mathbf{p}}$ . The quantitative results in Tab. 1, show that SketchINR provides substantially higher fidelity encoding than alternative representations like RNN [28], Bézier strokes [16], + +
FS-COCO [13]Sketchy [53]
CDR@10Time (s)CDR@10Time(s)
SketchRNN [28]---0.1641.650.11
CoSE [2]---0.1445.820.10
SketchODE [17]0.415.171.680.1737.200.13
BézierSketch [16]0.572.961.220.2325.720.08
Ours0.01115.610.30.00857.290.0007
+ +Table 1. Evaluating neural sketch representations' reconstruction quality (Chamfer distance $\downarrow$ ; recall@ $10\uparrow$ ) and decoding speed. + +![](images/adea143865d18c7223b4c716cfdbb2e36a60b7d26bb8b33d56dea1c1de6b1c1c.jpg) +Figure 5. Qualitative reconstruction results for neural sketch representations. SketchINR is uniquely able to scale to sketches of Sketchy and FS-COCO complexity. + +ODE dynamics [17], and stroke embeddings [2] particularly for extremely complex sketches in FS-COCO, where SketchRNN and CoSE fail entirely. Qualitative results in Fig. 5 show SketchINR to reconstruct simple and complex sketches with high fidelity while SketchRNN performs poorly for Sketchy and fails completely on FS-COCO. + +Sketch Compression: We next show that SketchINR's high-fidelity encoding is extremely compact, thus providing an efficient sketch codec. INRs store information implicitly [55] in the network weights $(\theta)$ and latent codes $(\nu_{i})$ . To store or transmit a sketch dataset, we only stream weights $\theta$ once (fixed cost) and a compressed latent vector $\nu_{i} \in \mathbb{R}^{D}$ for each sketch (variable cost). Hence, the space complexity for $N_{D}$ sketches with SketchINR is $|\theta| + N_{D} \times |\nu_{i}|$ bytes, where $\nu_{i}$ is 64 dimensional latent. For streaming a large number of sketches, we discount $\theta$ as negligible streaming overhead. Comparing to alternative representations, streaming $N_{D}$ (i) binary raster (BR) sketches of size $(N_{x} \times N_{y})$ take $N_{D} \times N_{x} \times N_{y}$ bits, and (ii) vector scene sketches with $N$ coordinates and pen states $(\mathbb{R}^{2+1})$ take $N_{D} \times N \times 2 \times 3$ bytes in 16-bit floating precision ( $N \approx 3000$ in FS-COCO + +[13]). Raster representations can be further stored by only storing the coordinates $(\mathbb{R}^2)$ of pixels containing the sketch (i.e., inked pixels) as sparse binary rasters (SBR). Fig. 6 shows rate-distortion for storing 10,000 scene sketches [13] with $\nu_{i}$ varying from 128 to 8 dimensional latent, and quality for raster and vector forms dropped with downsampling and RDP simplification [23] respectively. Quality metrics like PSNR and SSIM are more suited to photos. We use Chamfer Distance as a quality metric for vector sketch and observe $\sim 10\times$ and $\sim 60\times$ more efficiency than vector and raster sketches, respectively. Future works can further optimise our models and thus reduce storage requirements via model compression [21, 29, 56] and quantisation [11]. + +![](images/d1c6198caa03f00604f631ac1001a30e0a4ff7eaecd8419e117a6b8a4cb5f0f4.jpg) +Figure 6. Rate Distortion: SketchINR can encode complex scene sketches from FS-COCO [13] in highly compact $(\mathbb{R}^{64})$ latent codes. Specifically, despite nearly identical sketch quality (low CD represents higher fidelity), SketchINR has $\sim 60\times$ lower BPP than PNG raster sketches and $\sim 10\times$ lower than vectors sketches. + +Intra-Sketch Variation: Unlike prior representations (RNN [28], ODE dynamics [17]), humans do not draw exact stroke-level replica of a raster sketch. Instead, we choose the level of detail, the number of points, and the strokes used to convey the underlying shape. Given a latent code $\nu_{i}$ , SketchINR allows users to explicitly choose $N$ and $K$ , where the number of points $p = \{p_1,\dots ,p_N\}$ and strokes $s = \{s_1,\ldots ,s_K\}$ . Choosing a low $N$ and $K$ leads to abstract self-reconstructions with fewer points and strokes. Fig. 7(top to bottom) shows that we can control $K$ to decrease or increase the level detail in the sketch. + +# 4.2. Latent Space Interpolation (Creative Mixing) + +Interpolation of learned latents allows us to explore latent space continuity. Auto-regressive modelling of sketches builds poor latents as explored by Das et. al. [19] due to the lack of a full visibility of the sketch at any decoding timestep. Specifically, auto-regressive models use the partially predicted sketch along with the sketch's latent representation to predict new way-points with consistency. This, however, introduces noise in the latent $\rightarrow$ sketch decoding so that small changes in these latents lead to big changes in the final sketch. Deterministic latent to sketch decoding solves this problem as noise is not added at any point, leading to meaningful sketch interpolation from corresponding latent space interpolations. We visualise in + +![](images/b46ea493f27fe1a381abda23a10a47263c0590d32d66c3d01d509d97e455da19.jpg) +Figure 7. Given a latent code $\nu_{i}$ , we can reconstruct a sketch using learned $f_{\theta}$ at multiple resolutions by controlling the number of strokes $K$ during sampling. + +Cross-Sketch Interpolation with auto-regressive networks - SketchRNN + +![](images/5533c1a13f69b71f81b46f45ee080bf8a47161f049f4c5215d8834eaaf7872ca.jpg) + +Cross-Sketch interpolation with implicits - SketchINR + +![](images/a0d842a57d73d2ac8488ffdc6888a567f9ffff78778527e2102e93cfbbc39f74.jpg) +Figure 8. Latent space interpolation shows continuity and translation from one sketch concept to another. + +Fig. 8 how one sketch morphs into another by reconstructing from $f_{\theta}(\nu', t_n, \Delta t, s_k, \Delta s)$ , where $\nu'$ is sampled along a latent walk: $\nu' = (1 - \delta)\nu_1 + \delta \nu_2$ , for $\delta \in [0,1]$ . Interpolating between two sketches in the continuous latent space gives creative mixing – modifying the sketch of airplane-1 (facing right) into airplane-2 (facing left) by varying $\delta$ . Fig. 8 shows that small changes in the latent (columns) leads to smoother changes in sketch space for SketchINR. Meanwhile SketchRNN can exhibit big jumps that can zig-zag over successive latent increments, as indicated by an example tracked point in the figure. + +# 4.3. INR inversion for sketch completion + +Neural networks used to approximate implicit representations have been inverted in parallel literature [30] to appropriate samples not seen during training. This inversion, to arrive at a latent $\nu$ that describes the new sample, is usually made possible by learned priors in $f_{\theta}$ . Specifically, an implicit decoder $f_{\theta}$ is frozen with latent $\nu$ optimised for $f_{\theta}(\nu, \ldots)$ against the ground truth. This behaviour is easily replicated in SketchINR, where we optimise the latent $\nu$ to describe sketch samples (Fig. 5). To implement + +(a) Input Partial Stroke Vector-MNIST (Left) and Quick-Draw! (Right) + +![](images/7898dce9327a887394bedb35dcbbd9accd32630f7ecad25846834534df9aa74a.jpg) + +![](images/9f1ca3a3ef16062faf1340b788350408e9f7df30f0043f1bf990f966bb43514d.jpg) +Sketch Completion using SketchRNN +Sketch Completion using SketchINR + +![](images/a8b8505fc47496acf52cf3d713808f03d5d485f34ffeea1fc5dc17528d3aa569.jpg) + +![](images/2cd8f870418982026c14026f1fa834ad6392f262952f41dcaab0e641a2e2e863.jpg) +Figure 9. We perform sketch completion by inverting learned implicit representations and decoding for occluded strokes or waypoints. (a) Temporal completion (b) Unlike auto-regressive methods like SketchRNN [28], we can perform a-temporal (or inverse-temporal) completion, where given the second half of a partial sketch, we can successfully reconstruct the first part. + +sketch completion, we obtain a sketch descriptor $\nu$ from a partial sketch by just optimising for specific time-steps (e.g. 1/4th of a sketch; $t \in [0, \frac{1}{4})$ ) available in a partial doodle. Then, by sampling for rest of the timesteps we can complete the incomplete sketch. We summarise some qualitative sketch completion results in Fig. 9(a). Completed sketches are varied because of randomness in optimisation, as partial sketches are often ambiguous. INR inversion thus provides a novel approach to sketch completion, previously dominated by auto-regressive and models. Unlike auto-regressive models SketchINR can also perform a-temporal completion, inferring the first half of a sketch given the second half of the sketch (Fig. 9b). We remark that popular INR applications such as super-resolution [47] and point-cloud densification [48] perform interpolation after inversion on sparse inputs (e.g.: $t \in [0, 1)$ given $t \in \{0, \frac{1}{4}, \frac{1}{2}, \frac{3}{4}\}$ ), while sketch completion requires extrapolation (e.g.: $t \in [0, 1)$ given $t \in [0, \frac{1}{4})$ ), making it more analogous to visual outpainting. + +# 4.4. Sketch Generation + +We can perform conditional and unconditional sequential vector sketch generation after learning a generative model for sketch latents as described in Sec. 3.3. + +Unconditional Generation: We unconditionally generate sketches by random sampling $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ with the learnt VAE latent space, obtaining novel sketch descriptors $\nu_{i}\in \mathcal{V}$ . We further sample $f_{\theta}(\nu_{i},\dots)$ with varying timestamps $T\in [100,300]$ and strokestamps $K\in [10,30]$ obtaining sketches at varying levels of abstraction. Qualitative results in Fig. 10 show diverse and plausible samples from Vector-MNIST [17] and Quick-Draw! [28] datasets. + +![](images/c071b249dd46dd7892a71ac96dca3d64f907bb483cce2bf2664dbe6d8249f73b.jpg) +Figure 10. Unconditional sketch generation. Sampling sketch latents $\nu$ as well as stroke and point complexity $(K, T)$ . + +Conditional Generation: To condition sketch generation, we use our pre-trained VAE encoder $E$ to embed raster sketches, which can then be vectorised by the decoder $f_{\theta}$ . Fig. 11 shows qualitative results for sequential vector sketch generation from raster images, where we obtain comparable results with auto-regressive generative modelling [28]. We observe that treating vectorisation as a generative problem rather than a deterministic raster-to-vector mapping [7] leads us to have variations in generated vectors particularly in the form of stroke order resembling real copies of sketches made by humans. + +![](images/ddf0fc6e578f25e0212718635520cfa79c9b40491e03db5197833cfb8968344b.jpg) +Figure 11. Conditional sketch generation (vectorisation) on Vector-MNIST. Raster sketches (top left) can have multiple vector forms, e.g., sketched with different numbers of strokes (bottom). + +# 4.5. Ablations and Discussion + +Global vs Local Time Modelling: We model $J$ timestamps across $K$ strokes, by distributing timestamps uniformly over each stroke such that each stroke has $\frac{J}{K}$ timestamps. In other words, time $(t,\Delta t)$ is modelled globally for the entire sketch $\mathbf{p} = \{\mathrm{p}_1,\dots ,\mathrm{p}_N\}$ . An alternative is to model $(t,\Delta t)$ as 0to1 for each stroke separately (local time-modelling) instead of the entire sketch. Particularly, each of the $K$ strokes are modelled using $N$ points, representing the entire sketch using $N\cdot K$ points as $\mathbf{p} = \{\mathrm{p}_1,\dots ,\mathrm{p}_{N\cdot K}\}$ . Visually inspecting global vs local time embedding shows that local modelling leads to jittery sketches (see Fig. 12), whereas global modelling gives visually smooth sketches. We hypothesise that time is a global property of a sketch and not an independent property localised to each stroke. Hence, modelling time locally puts an additional optimisation burden on our implicit decoder to ensure smoothness and continuity, resulting in poorly reconstructed sketches. + +![](images/3f8e4975916570af285cfd13ce622394fb52e77f569e562e5f0f99ade64f6219.jpg) +Figure 12. Global vs Local time embedding: Global embedding gives smooth sketches compared to local, which gives jittering. + +Fixed vs Variable Smoothing Factor $(\gamma)$ : The intensity map $\mathrm{I}_k$ in Eq. (4) guides our raster-based visual loss $\mathcal{L}_V$ . A key component in $\mathrm{I}_k$ is the smoothing parameter $\gamma -$ low $\gamma$ (thicker $\mathrm{I}_k$ ) gives less accurate but smoother values on the entire visual region (easier training); high $\gamma$ (thinner $\mathrm{I}_k$ ) gives more accurate but sharper (steeper) intensity-gradients near ground-truth strokes (harder training). While we use a fixed $\gamma$ with MSE Loss in Eq. (6) to balance out the optimisation complexity, we note that an alternative is to vary $\gamma$ every fixed number of steps with a scheduler to allow faster optimisation. Specifically, we vary $\gamma$ as 20to200 on a linear scale by incremental $\Delta \gamma$ every few iterations. Despite its theoretical advantage [25, 45], our initial experiments suggest only minor improvement that reduces training time by $6.1\%$ . A detailed study of bundle-adjustment [45] for visual loss is an interesting future work. + +Limitations and Future Work Despite it's efficiency as a representation for hand-drawn sketches, SketchINR suffers from a number of limitations. We naturally share limitations with other implicit representations including optimisation-based encoding and poor cross category generalisation. While unlike Sketch-RNN our generation network can be trained on 2-3 categories at once, it fails to generalise on more. In addition, a core limitation lies in slow convergence due to pixel space optimisation. While vector-space loss in the form of mean squared error mitigates this to some extent (Fig. 3), convergence is still slow and can possibly be improved by better engineering the implicit function. Strokes could also be smoothed by explicit inclusion of stroke gradient (slope) based regularisation in future work. Finally, vectorisation for conditional generation performs poorly on high complexity sketches. + +# 5. Conclusion + +We introduce visual implicits to represent vector sketches with compressed latent descriptors. This neural representation provides a high-fidelity and compact representation that raises the possibility of a sketch-specific codec for compactly representing large sketch datasets. Our SketchINR can decode sketches at varying levels of detail with controllable number of strokes, and provides superior cross-sketch interpolation between implicits, demonstrating a smoother latent space than auto-regressive models. Applications such as sketch-completion are also supported, including a-temporal completion not available with auto-regressive models. Decoding is inherently parallel and can be over $100 \times$ faster than autoregressive models in practice. + +# References + +[1] Emre Aksan, Fabrizio Pece, and Otmar Hilliges. Deepwriting: Making digital ink editable via deep generative modeling. In CHI, 2018. 1, 2 +[2] Emre Aksan, Thomas Deselaers, Andrea Tagliasacchi, and Otmar Hilliges. Cose: Compositional stroke embeddings. In NeurIPS, 2020. 1, 2, 3, 6 +[3] Stephan Alaniz, Massimiliano Mancini, Anjan Dutta, Diego Marcos, and Zeynep Akata. Abstracting sketches through simple primitives. In ECCV, 2022. 4 +[4] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks. In CVPR, 2024. 1 +[5] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 1 +[6] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can beat me? In SIGGRAPH Asia, 2020. 1 +[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 2, 8 +[8] Alexandre Binninger, Amir Hertz, Olga Sorkine-Hornung, Daniel Cohen-Or, and Raja Giryes. Sens: Part-aware sketch-based implicit neural shape modeling. arXiv preprint arXiv:2306.06088, 2024. 1, 2 +[9] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 1965. 2 +[10] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation. In NeurIPS, 2020. 2 +[11] Qualcomm Innovation Center. Ai model efficiency toolkit (aimet). https://github.com/quic/aimet, 2023.6 +[12] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. In NeurIPS, 2018. 3 +[13] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2, 5, 6 +[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aaneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What can human sketches do for object detection? In CVPR, 2023. 2 +[15] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2 +[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 1, 2, 3, 5, 6 + +[17] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Sketchode: Learning neural sketch representation in continuous time. In ICLR, 2021. 1, 2, 3, 5, 6, 7 +[18] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Cloud2curve: Generation and vectorization of parametric sketches. In CVPR, 2021. 2, 3 +[19] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In ICLR, 2023. 3, 6 +[20] Carl De Boor and Carl De Boor. A practical guide to splines. Applied Mathematical Sciences, 1978. 3 +[21] Zhen Dong, Zhewei Yao, Amir Gholami, Michael Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In ICCV, 2019. 6 +[22] Asen L. Dontchev and R. Tyrrell Rockafellar. Implicit Functions and Solution Mappings. Springer-Verlag, 2014. 2 +[23] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 1973. 6 +[24] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 2 +[25] F. Dan Foresee and Martin T. Hagan. Gauss-newton approximation to bayesian learning. In ICNN, 1997. 8 +[26] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, SM Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In ICML, 2018. 3 +[27] Songwei Ge, Vedanuj Goswami, C. Lawrence Zitnick, and Devi Parikh. Creative sketch generation. In ICLR, 2021. 1 +[28] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 2, 3, 5, 6, 7, 8 +[29] Hai Victor Habi, Roy H. Jennings, and Arnon Netzer. Hmq: Hardware friendly mixed precision quantization block for cnns. In ECCV, 2020. 6 +[30] Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. Spaghetti: Editing implicit shapes through part aware generation. ACM TOG, 2022. 2, 7 +[31] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2 +[32] Conghui Hu, Da Li, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Sketch-a-classifier: Sketch-based photo classifier generation. In CVPR, 2018. 3 +[33] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2 +[34] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2 +[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5 +[36] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 3 + +[37] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1 +[38] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Free2cad: Parsing freehand drawings into cad commands. ACM TOG, 2022. 1, 2 +[39] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 2 +[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5 +[41] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Deflocnet: Deep image editing via flexible low-level controls. In CVPR, 2021. 1, 2 +[42] Asif Masood and Sidra Ejaz. An efficient algorithm for robust curve fitting using cubicBezier curves. In ICIC, 2010. 3 +[43] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 2, 3 +[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3 +[45] Jorge J. Mofe. The levenberg-marquardt algorithm. In Numerical Analysis, 1976. 8 +[46] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3 +[47] Quan H Nguyen and William J Beksi. Single image superresolution via a dual interactive implicit neural network. In WACV, 2023. 7 +[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3, 4, 7 +[49] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML, 2019. 3 +[50] Michael Revow, Christopher KI Williams, and Geoffrey E Hinton. Using generative models for handwritten digit recognition. IEEE TPAMI, 1996. 2 +[51] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In CVPR, 2020. 1, 2 +[52] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 2 + +[53] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 1, 5, 6 +[54] Heping Sheng, John Wilder, and Dirk B. Walther. Where to draw the line? PLOS One, 2021. 2 +[55] Yannick Strumpler, Janis Postels, Ren Yang, Luc van Gool, and Federico Tombari. Implicit neural representations for image compression. In ECCV, 2022. 6 +[56] Stefan Uhlich, Lukas Mauch, Fabien Cardinaux, Kazuki Yoshiyama, Javier Alonso Garcia, Stephen Tiedemann, Thomas Kemp, and Akira Nakamura. Mixed precision dnns: All you need is a good parametrization. In ICLR, 2020. 6 +[57] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM TOG, 2022. 1, 2, 3 +[58] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In ICCV, 2023. 1, 2, 3 +[59] Alexander Wang, Mengye Ren, and Richard Zemel. Sketchembednet: Learning novel concepts by imitating drawings. In ICML, 2021. 2, 4 +[60] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In NeurIPS, 2023. 1 +[61] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 1, 2 +[62] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2 +[63] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 2 +[64] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 3 +[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2021. 1 +[66] Wenni Zheng, Pengbo Bo, Yang Liu, and Wenping Wang. Fast b-spline curve fitting by 1-bfgs. Computer Aided Geometric Design, 2012. 2 \ No newline at end of file diff --git a/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/images.zip b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16a555e907e21ebdfa264a22465070a76c045864 --- /dev/null +++ b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c50bcd21b313d0ba65f42dab1c6612e8fba38466d564dc9596f85795aeff011 +size 458410 diff --git a/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/layout.json b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..39ebcb9ce8e96338c771e83f19e6fffcf299a7e2 --- /dev/null +++ b/2024/SketchINR_ A First Look into Sketches as Implicit Neural Representations/layout.json @@ -0,0 +1,11603 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 86, + 524, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 86, + 524, + 102 + ], + "spans": [ + { + "bbox": [ + 67, + 86, + 524, + 102 + ], + "type": "text", + "content": "SketchINR: A First Look into Sketches as Implicit Neural Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 108, + 531, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 108, + 531, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 108, + 531, + 205 + ], + "type": "text", + "content": "Hmrishav Bandyopadhyay1 Ayan Kumar Bhunia1 Pinaki Nath Chowdhury1 Aneeshan Sain1 Tao Xiang1,2 Timothy Hospedales3 Yi-Zhe Song1,2 \n1SketchX, CVSSP, University of Surrey, United Kingdom. \n2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence. \n3University of Edinburgh, United Kingdom \n{h.bandyopadhyay, a.bhunia, p.chowdhury, a.sain, t.xiang, y-song}@surrey.ac.uk t.hospedales@ed.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 213, + 358, + 330 + ], + "blocks": [ + { + "bbox": [ + 50, + 213, + 358, + 330 + ], + "lines": [ + { + "bbox": [ + 50, + 213, + 358, + 330 + ], + "spans": [ + { + "bbox": [ + 50, + 213, + 358, + 330 + ], + "type": "image", + "image_path": "0326cb215bf5f979385ad4ca53dc3f4c8d7fb3a7bd4bed2acdb998828504ebc5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 377 + ], + "type": "text", + "content": "Figure 1. SketchINR is an implicit neural representation for sequential vector sketches. It is the first neural representation with sufficient fidelity to be a drop-in replacement for the raw sketch data (a). The representation is significantly higher fidelity than existing learned representations such as SketchRNN [28], especially for more complex sketches (b, left). SketchINR also provides state-of-the-art sketch compression, with a substantially more compact representation than either vector or raster sketches (b, right)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 363, + 213, + 545, + 330 + ], + "blocks": [ + { + "bbox": [ + 363, + 213, + 545, + 330 + ], + "lines": [ + { + "bbox": [ + 363, + 213, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 363, + 213, + 545, + 330 + ], + "type": "image", + "image_path": "0b2dc68e7f26fc593d940f54567ddc9b80140fadfe77ac3e58b593325520dec8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 385, + 192, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 385, + 192, + 398 + ], + "spans": [ + { + "bbox": [ + 143, + 385, + 192, + 398 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "spans": [ + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "text", + "content": "We propose SketchINR, to advance the representation of vector sketches with implicit neural models. A variable length vector sketch is compressed into a latent space of fixed dimension that implicitly encodes the underlying shape as a function of time and strokes. The learned function predicts the " + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "text", + "content": " point coordinates in a sketch at each time and stroke. Despite its simplicity, SketchINR outperforms existing representations at multiple tasks: (i) Encoding an entire sketch dataset into a fixed size latent vector, SketchINR gives " + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "inline_equation", + "content": "60 \\times" + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "text", + "content": " data compression over raster and vector sketches, respectively. (ii) SketchINR's auto-decoder provides a much higher-fidelity representation than other learned vector sketch representations, and is uniquely able to scale to complex vector sketches such as FS-COCO. (iii) SketchINR supports parallelisation that can decode/render " + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\sim 100 \\times" + }, + { + "bbox": [ + 45, + 402, + 290, + 666 + ], + "type": "text", + "content": " faster than other learned vector representations such as SketchRNN. (iv) SketchINR, for the first time, emulates the human ability to reproduce a sketch with varying abstraction in terms of number and complexity of strokes. As a first look at implicit sketches, SketchINR's compact high-fidelity representation will support future work in modelling long and complex sketches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 672, + 128, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 128, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 128, + 683 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "The prevalence of touch-screen devices has triggered significant research progress on sketches [4, 6, 8, 51, 57]. Mod-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 386, + 546, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 546, + 458 + ], + "type": "text", + "content": "elling digital sketches has become an important challenge for learning systems that aim to stimulate human creativity [17, 27]. Large-scale datasets [24, 28, 53] motivated several downstream applications like image retrieval [15, 61], image generation [37, 65], image editing [1, 41], and 3D content creation [5, 8, 38], among others [58, 60]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": "These applications are underpinned by sketches captured in raw form either as raster or vector images, and usually encoded into derived representations by ConvNets [15], splines, and so on. For over a decade, discourse around what is a \"good\" representation of human sketches has persisted [31]. A now substantial body of work has focused on representations for sequential vector sketches that model both explicit strokes, and their drawing over time — most famously by the auto-regressive SketchRNN [28], but also using representations such as parametric curves [16], Gaussians [2], and Neural ODEs [17]. We introduce a novel continuous-time representation of sequential vector sketches by taking an implicit neural representation perspective on sketches for the first time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "First, we review the limitations of current sketch representations. The two most popular raw sketch representations are (i) raster sketch - a black and white image of a line drawing, and (ii) vector sketch - a temporal sequence of points and strokes. Raster sketch has a large but fixed storage (e.g., " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "). Its compatibility with ConvNets [64] made raster sketches popular but they lack the tempo-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12565" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "ral information (i.e., stroke order) necessary for generative modelling [17]. Vector sketches efficiently store the " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "-coordinates of points and strokes (" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "N \\times 2" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "). However, the storage increases with the length of sketch " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " (Fig. 1b). (iii) Raw vector sketches can be encoded by popular learned sketch representations such as SketchRNN [28] and others [16, 17, 39]. However, for longer (" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "N \\geq 1000" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": ") scene sketches, the quadratic computational cost for transformers becomes intractable [39] and recurrent nets become unstable [28]. Nevertheless, most of these still suffer from slow sequential autoregressive decoding, and they mostly cannot accurately reconstruct complex sketches with many strokes. To summarise, current sketch representations either discard temporal information (raster), scale poorly in storage size (vector), or become slow and inaccurate with complex sketches (learned autoregressive vector)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 410 + ], + "type": "text", + "content": "To overcome the above issues, we propose an implicit neural representation for sketches - SketchINR. We piggyback on the well-established [8, 30, 34, 48] potential of implicit functions [22] to encode an underlying shape as a function of time/location. Different to standard implicits [43, 48], the input of sketches as parametric functions is a hierarchy of stroke sequence and point sequence. We develop SketchINR to generate points (xy-coordinates) given both time and stroke inputs. To represent multiple sketches, SketchINR learns a fixed size latent space and a conditional implicit function that generates xy-coordinates of any sketch instance given its corresponding latent code." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": "Despite its simplicity, SketchINR is competitive with existing sketch representations. (i) It can encode an entire sketch dataset into a compact latent space " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "(e.g., \\mathbb{R}^{N \\times 512})" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": " and decode using a fixed size function (e.g., an 8 layer MLP). Compared to raster sketches " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": ", SketchINR provides " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "128 \\times" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": " storage compression. While simple object-level vector sketches (length " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "\\leq 300" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": ") have comparable storage with SketchINR (" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{512}" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": "), for practical scene-level [13] vector sketches (length " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "\\geq 1000" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": "), SketchINR provides " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": " better compression (Fig. 1b). (ii) For decoding/generation, SketchINR can parallelly predict the " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": "-coordinates for all time and strokes. In practice, this is " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": " faster than an autoregressive vector sketch generator with length " + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "inline_equation", + "content": "\\geq 300" + }, + { + "bbox": [ + 46, + 411, + 289, + 700 + ], + "type": "text", + "content": ". (iii) SketchINR can accurately represent complex sketches (Fig. 1a,b) unlike competitors that suffer from complexity limitations [16, 17] or the length limitations of auto-regression [28]. The encoding fidelity is sufficiently high that SketchINR provides a drop-in replacement for raw sketch data, while being more compact. (iv) SketchINR supports diverse applications spanning smooth latent space interpolation, and sketch generation, sketch completion. Uniquely, it also supports sketch abstraction - the human-like ability to replicate the essence of a sketch with a variable number and complexity of strokes [31, 54]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "In summary, our contributions are: (i) We present the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "first implicit neural representation approach to vector sketch modelling, extending INRs as a function of time and strokes, and defining a training objective for learning them. (ii) SketchINR provides the first learned representation capable of representing complex sketches compactly and with high-fidelity (Fig. 1a). Based on our compact high-fidelity representation we introduce the task of sketch compression, and demonstrate excellent compression results (Fig. 1b)). (iii) We further demonstrate SketchINR's applicability to a variety of sketch related tasks including generation, interpolation, completion and abstraction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 235, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 547, + 510 + ], + "type": "text", + "content": "Sketch Representations: Digital sketches are predominantly captured as a function of 'time' [28], in a sequence of coordinates traced by an artist on a canvas. Recent works [24, 28] emphasise this temporal order of coordinates as an indicator of their relative importance in depicting a sketch-concept, as humans draw in a coarse-to-fine fashion. As such, they [2, 7, 18, 28] exploit temporal information in sketches for downstream tasks like sketch-assisted retrieval [7], generation [2, 18, 28] and modelling [38]. Vector sequences are further parameterised, for representation with parametric curves like Béziers [16, 57] and B-Splines [50, 66] for early-handwritten digit representation [50] to recent representations of complex structures like real-world objects [16, 57] and scenes [58]. Discarding temporal significance of strokes, vector-sketches can be converted to raster images by rendering coordinate strokes [9] on a 2D canvas. While raster sketches are less informative than their vector counterparts [28], they offer enhanced applicability by expanding beyond digital sketches to paper-sketches [18] where stroke order is already lost. Raster representations are encoded with translation-invariant networks like CNNs [63] and Vision-Transformers [8], making them extremely useful to represent object-level information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "content": "Non-Parametric Modelling: Both vector sequences and raster renderings of these sequences are non-parametric representations [16] of sketches. To preserve temporal stroke order, vector sketches are handled as sequences [28] of coordinates rather than unordered sets. Hence, they are encoded with position-aware networks like RNNs [28], LSTMs [15, 61], and Transformers [10, 51]. Without positional cues, raster sketches are processed by spatially-aware CNNs [63] and Vision Transformers [52]. Semantic encodings of non-parametric representations capture the expressive nature of sketch, allowing its use as a creative input for downstream segmentation [33], retrieval [7], and editing [41], as well as for interactive tasks like object-detection [14] and image-inpainting [62]. These encodings are further used in auto-regressive [28, 59] and VAE-like [10] modelling of vector coordinate sequences for the generation of textual-characters [1] and object sketches [28]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12566" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": "Parametric Representation learning: Parametric Splines [20], such as Beizer curves [42], approximate vector sketches by per-sample fitting on individual sketch samples. Amortised frameworks bypass this per-sample optimisation by inferring curve control points [16] and degree [18] from individual stroke features. Recent works exploit the representative power of parametric curves for generation of simple characters [26] to complex object [16, 18, 57] and scene level sketches[58]. Both non-parametric and curve-fitted Bézier parametric representations capture the explicit form of a sketch as spatial and sequential forms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 207, + 289, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 289, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 289, + 447 + ], + "type": "text", + "content": "Beyond explicit representations of sketches, we learn visual implicits to capture spatial dynamics in vectorsketches. Learned implicits can be used for sketch reconstruction with particular control over abstraction through number of reconstructed strokes. Somewhat similar to implicit representations for vector sketches, CoSE [2] parameterises Gaussian Mixtures on individual strokes, learning one implicit for one stroke only. New implicits are generated auto-regressively from previous predictions for auto-regressive generation and completion of sketches. Unlike CoSE, we learn a single implicit for the entire sketch, thus having one implicit parameter to sample the sketch in one pass. While SketchODE [17] similarly parameterises Neural Ordinary Differential Equations [12] to capture vector-sketch dynamics, it's optimisation is extremely slow (as much as " + }, + { + "bbox": [ + 47, + 207, + 289, + 447 + ], + "type": "inline_equation", + "content": "120 \\times" + }, + { + "bbox": [ + 47, + 207, + 289, + 447 + ], + "type": "text", + "content": " slower) because of high time complexity in solving higher order differentials. Three dimensional implicits like Signed-Distance [48] and Occupancy [43] functions learned from 3D point-clouds are particularly analogous to our implicit representation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": "Sketch Abstraction: The continuity of human cognition in visual perception [36] is directly reflected in how we sketch. This enables versatility [64] in sketch, allowing us to express a huge range of thoughts and visions directly on paper in the form of coarse-grained 'ideas' [32] and fine-grained 'objects' [64]. Sketch abstraction, as a direct inverse of this granularity, has been studied in depth as a function of (i) drawing time: humans draw most representative strokes first [28], and (ii) compactness: in a constrained stroke setting, salient strokes are prioritised [46]. As abstraction gives sketches a human touch, generative modelling of sketches is focused on abstraction towards making generated drawings \"more humane\" [19] and augmenting limited sketch datasets with simulated sketches from photos [57, 58]. Importantly, recent works [57] model abstraction as a function of strokes to explicitly control abstraction levels by restricting the number of generated strokes. Here, we represent sketches as visual implicits which allows us to reconstruct them at varying levels of abstraction and detail by controlling the number of reconstructed strokes. We demonstrate several downstream applications of this representation, particularly sketch compression, generation and completion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 389, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": "Overview: Vector sketches are captured as pen movements on a digital canvas, represented by coordinates " + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "inline_equation", + "content": "(x,y) \\in [0,1]" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": " with binary pen-states " + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "inline_equation", + "content": "\\delta \\in \\{0:" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": " pen-down, " + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "inline_equation", + "content": "1:" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": " pen-up" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "inline_equation", + "content": "\\}" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": " at those coordinates. A sketch " + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "inline_equation", + "content": "p = \\{(x_i,y_i,\\delta_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": " vector way-points (coordinates), and is portrayed (rendered) by retracing these points on a 2D canvas. This process, termed as rasterisation produces a pixelated raster sketch " + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "inline_equation", + "content": "p^s" + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": " on the canvas. In this work, we leverage scalability [48] in implicit functions to learn a flexible representation for vector sketches with raster guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 217, + 509, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 509, + 230 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 509, + 230 + ], + "type": "text", + "content": "3.1. Learning Visual Implicits for Sketches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": "Vector sketches can be modelled as implicit, where a sketch is represented as function of time " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "f_{\\theta}(t_j) : \\mathbb{R} \\to \\mathbb{R}^{2+1}" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " parameterised by " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": ". Each time-stamp " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "t_j \\in [0,1)" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " is mapped to a pen position in the form of " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": "-coordinates " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "(x,y) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " and binary pen-states " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^1" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": ". The entire sketch can be reconstructed as a vector sequence by sampling " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " from 0 to 1 with a learned " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\{f_{\\theta}(0),\\dots,f_{\\theta}(t_j),\\dots,f_{\\theta}(\\frac{J-1}{J})\\}" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " for a total of " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " timestamps. This in itself is a flexible representation for a vector sketch, as it allows for sampling with arbitrary timestamps, thus potentially increasing or decreasing resolution of the resulting sketch by modulating number of strokes with sampling resolution " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "(t : \\{0,\\frac{1}{3},\\frac{2}{3}\\} \\mathrm{v/s} t : \\{0,\\frac{1}{10},\\dots,\\frac{9}{10}\\})" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": ". For increased control over reconstruction granularity in the form of abstraction, we model " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " on an additional variable in the form of strokestamps " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\{s_k\\}_{k=1}^K" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": ", representing the sketch as " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\hat{p} = \\{f_{\\theta}(0,0),\\dots,f_{\\theta}(t_j,s_k),\\dots,f_{\\theta}(\\frac{J-1}{J},\\frac{K-1}{K})\\}" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " strokes (Fig. 2). These strokestamps provide explicit control over pen-up and pen-down states, allowing us to reconstruct the sketch with re-defined granularity during inference (e.g. 7 or 9 strokes). Specifically, strokestamps represent the current stroke number as a fraction of the total number of strokes " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " which allows us to change strokes by changing " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": ". Pen-states " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " are thus explicitly determined from strokestamps only by toggling when " + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 304, + 236, + 547, + 559 + ], + "type": "text", + "content": " changes value. Our practical implementation of sketch implicits follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 376, + 564, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 564, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 376, + 564, + 545, + 578 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(t _ {j}, \\Delta t, s _ {k}, \\Delta s\\right) = p _ {j} \\tag {1}", + "image_path": "d7058bb9bad0f271f1392081d9ea2dbc982f580b8e80b886fc355bf83dbdbb24.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": "where, time instances are computed as " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "t_j = t_{j-1} + \\Delta t" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "t_0 = 0" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "\\Delta t = \\frac{1}{J}" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": "-stroke instances are computed as " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "s_k = s_{k-1} + \\Delta s" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "s_0 = 0" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "\\Delta s = \\frac{1}{K}" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": ". To exploit bias in neural networks towards learning low frequency functions [44, 49], we smoothen out high frequency variations " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "(0 \\text{to} \\frac{1}{K}, 0 \\text{to} \\frac{1}{S})" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": " for timestamps and strokestamps by mapping them from " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": " to a higher dimension " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2L}" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": " using positional embeddings " + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "inline_equation", + "content": "\\mathrm{PE}(\\cdot)" + }, + { + "bbox": [ + 304, + 583, + 547, + 679 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 330, + 684, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 684, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 330, + 684, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {P E} (x) = \\left[ \\sin \\left(1 0 ^ {- 4 / L} x\\right), \\cos \\left(1 0 ^ {- 4 / L} x\\right), \\right. \\tag {2} \\\\ \\ldots , \\sin (1 0 ^ {- 4 L / L} x), \\cos (1 0 ^ {- 4 L / L} x) ] \\\\ \\end{array}", + "image_path": "bb7212d25a3d2b836a039ab8c07a40e81c2203504cb36a1dcce5e3de5a8077ce.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "12567" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "x \\in \\{t_n, \\Delta t, s_k, \\Delta s\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ". We uniformly distribute " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " times-stamps across " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " strokes, where each stroke has " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "J / K" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " points. For data augmentation, we train " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " with varying " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\pm 50\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " of ground truth value). This helps " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " to scale to different levels of abstraction, and number of points and strokes during inference." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 152, + 284, + 291 + ], + "blocks": [ + { + "bbox": [ + 51, + 152, + 284, + 291 + ], + "lines": [ + { + "bbox": [ + 51, + 152, + 284, + 291 + ], + "spans": [ + { + "bbox": [ + 51, + 152, + 284, + 291 + ], + "type": "image", + "image_path": "1302058cfe7037f4cc43014ed0e7dce1a0670bd315cb43ae2fedc80c87f5fdc3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": "Figure 2. We model sketch as a function " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "f_{\\theta}(t_j, s_k)" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " timestamps and " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " strokes. Timestamps " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\{t_a, \\ldots, t_b\\}" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "t_j \\in [0,1)" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " correspond to way-points for a specific stroke " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " with learned weights " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " can be sampled with arbitrary number of strokes as " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "t_j \\in \\{0, \\frac{1}{J}, \\dots, \\frac{J - 1}{J}\\}" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "s_k \\in \\{0, \\frac{1}{5}, \\dots, \\frac{4}{5}\\}" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "s_k \\in \\{0, \\frac{1}{10}, \\dots, \\frac{9}{10}\\}" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": ", leading to increasing or decreasing abstraction for " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 46, + 299, + 287, + 376 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 394, + 138, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 138, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 138, + 406 + ], + "type": "text", + "content": "3.2. Loss functions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "content": "To fit our implicit function " + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "content": " to a particular sketch instance " + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "inline_equation", + "content": "p = \\{(x_0, y_0, \\delta_0), \\ldots, (x_{M-1}, y_{M-1}, \\delta_{M-1})\\}" + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "content": " vector way-points " + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\{p_m\\}_{m=0}^{M-1}" + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "content": ", a naive approach would be to optimise " + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "content": " with mean squared error against the ground truth vectors as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 477, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 91, + 477, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} = \\sum_ {m = 0} ^ {M} \\left| \\left| f _ {\\theta} \\left(\\frac {m}{M}, \\frac {k}{K}\\right) - p _ {m} \\right| \\right| _ {2} \\tag {3}", + "image_path": "43274ee88d5dbfae7d587df59c5bf0845245f8bda8286a5b3b0c756cc851985d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "text", + "content": "where each predicted coordinate and pen-state is optimised directly against corresponding vector way-points from the ground truth. However, this rigid point-to-point matching enforces memorisation of particular points and strokes without any spatial understanding. As such, sampling with a different resolution at this stage yields noisy output that does not correspond to the raster sketch. To explicitly optimise for sampling flexibility, we introduce a visual loss, penalising sketch reconstructions that do not match visually. For this, we sample a template grid " + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "text", + "content": " of 2D coordinates and compute the region of influence of each vector-waypoint " + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "text", + "content": " of stroke " + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "text", + "content": " on coordinate " + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 514, + 287, + 659 + ], + "type": "text", + "content": " as an intensity map [3, 59]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 664, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 664, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 52, + 664, + 287, + 684 + ], + "type": "interline_equation", + "content": "\\mathrm {I} _ {k} (g, s _ {k}) = \\max _ {p _ {m} \\in s _ {k}, r \\in [ 0, 1 ]} \\exp (\\gamma | | g - r p _ {m} - (1 - r) p _ {m + 1} | | _ {2}) \\tag {4}", + "image_path": "399ff7aba386bf7c46f58d0b60ede87e68b5e0c889b4e7298b55038d7d932501.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "where intensity " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_k" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " is smoothed with " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " for points " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " in stroke " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "m \\in \\{0, \\dots, \\frac{K - 1}{K}\\}" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": ". The intensity map" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "for the entire sketch is formulated as a summation of maps from all strokes for " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "k \\in (0,1]" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " for both the ground truth sketch vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " strokes and the arbitrarily sampled sketch " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\hat{K}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " strokes as " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{I}}_k" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", composing the loss as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 129, + 545, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 129, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 367, + 129, + 545, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V} = \\left| \\left| \\sum_ {k = 0} ^ {\\frac {K - 1}{K}} \\mathrm {I} _ {k} - \\sum_ {k = 0} ^ {\\frac {\\hat {K} - 1}{\\hat {K}}} \\hat {\\mathrm {I}} _ {k} \\right| \\right| _ {2} \\tag {5}", + "image_path": "67856442945dbc8b1f7c6f6d2bafa706db79e33829dd97b565d34fcf3a307b56.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": "Intensity Smoothing with " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": ": The sketch intensity map " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " helps us to ground vector sketches to raster definiteness (Eq. (4)) by a visual rendering of individual strokes where intensity in coordinates " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "g \\in G" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " drops exponentially with distance from " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "p_m \\in s_k" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": ". This exponential drop is controlled by a smoothing factor " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " as a hyperparameter, where (i) a low " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " yields thicker lines on " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " (low-quality sketch reconstructions) but offers better optimisation, while (ii) a high " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " yields more refined lines (higher quality reconstructions), making optimisation significantly slower (Fig. 3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "content": "For faster optimisation, we assist the visual penalty with a weighted MSE loss from Eq. (3) and a lower gamma, empirically set to " + }, + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\gamma = 150" + }, + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "content": ". The loss function weighted with " + }, + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "content": ", looks can be written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 351, + 545, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 351, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 357, + 351, + 545, + 368 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i m p l i c i t}} = \\mathcal {L} _ {V} ^ {\\gamma = 1 5 0} + \\lambda_ {\\mathrm {M S E}} \\cdot \\mathcal {L} _ {\\mathrm {M S E}} \\tag {6}", + "image_path": "3b5f36b64f84d1ac6d6269d34da2cf20278e1d3f0af9589822446d5c56dd7356.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": " to 0.7. This, however, inherently introduces bias in our optimisation to match the order of strokes with the ground truth order to some extent. Future works could ideally set " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MSE}} = 0" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": " if they obtain faster convergence from a better engineered implicit function " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": ". Fig. 4a summarises our implicit sketch learning framework." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 466, + 556, + 552 + ], + "blocks": [ + { + "bbox": [ + 307, + 466, + 556, + 552 + ], + "lines": [ + { + "bbox": [ + 307, + 466, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 556, + 552 + ], + "type": "image", + "image_path": "f25d662a62327603735692a8289c6b17afa2685b24422537af4b8eb38e066618.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "text", + "content": "Figure 3. Effect of smoothing factor " + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "text", + "content": " on training: Reducing " + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "text", + "content": " leads to higher intensity in the surrounding region near stroke " + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "text", + "content": ". This is similar to stroke dilation of intensity map " + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_k" + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "text", + "content": ". A lower " + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "text", + "content": " leads to stable training (plot on right) but lacks fine-grained details. A higher " + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 556, + 545, + 611 + ], + "type": "text", + "content": " gives a fine-grained sketch but is harder to train." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 623, + 486, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 486, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 486, + 635 + ], + "type": "text", + "content": "3.3. Generalising to Multiple Sketches" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "To efficiently represent multiple sketch implicits with a single global and general function, we modify Eq. (1) to include a sketch descriptor " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\nu_{i}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " for each sketch " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\nu_{i},t,\\Delta t,s_{k},\\Delta s)" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " is approximated with a fully connected auto-decoder network, similar to DeepSDF [48]. Thus, feature descriptors " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " are hidden vectors, learned" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12568" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 44, + 85, + 218, + 203 + ], + "blocks": [ + { + "bbox": [ + 42, + 72, + 176, + 83 + ], + "lines": [ + { + "bbox": [ + 42, + 72, + 176, + 83 + ], + "spans": [ + { + "bbox": [ + 42, + 72, + 176, + 83 + ], + "type": "text", + "content": "(a): Single sketch Implicit Learning" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 44, + 85, + 218, + 203 + ], + "lines": [ + { + "bbox": [ + 44, + 85, + 218, + 203 + ], + "spans": [ + { + "bbox": [ + 44, + 85, + 218, + 203 + ], + "type": "image", + "image_path": "fe7980f604d4fcbf9b3db3f358080564a901f7311acd41ed00da39b329de927c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 221, + 89, + 353, + 200 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 341, + 83 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 341, + 83 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 341, + 83 + ], + "type": "text", + "content": "(b): Multi-sketch auto-decoder" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 89, + 353, + 200 + ], + "lines": [ + { + "bbox": [ + 221, + 89, + 353, + 200 + ], + "spans": [ + { + "bbox": [ + 221, + 89, + 353, + 200 + ], + "type": "image", + "image_path": "6dcc40403b97c8c12ee7a18cf1a6536edc27118b5e77f87db5a294e3fb4d7828.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "text", + "content": "Figure 4. SketchINR model diagram. (a) Embedding a single vector sketch as an implicit model " + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "text", + "content": " mapping stroke and time " + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "inline_equation", + "content": "(\\mathbf{s},\\mathbf{t})" + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "text", + "content": " to ink coordinate " + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "text", + "content": ". (b) Embedding a vector sketch dataset as a shared decoder and set of latent vectors " + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "text", + "content": ". (c) Training a generative model for implicit sketches by generating latent codes " + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "text", + "content": " using an encoder " + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 207, + 547, + 242 + ], + "type": "text", + "content": " that inputs raster sketches." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 359, + 89, + 555, + 196 + ], + "blocks": [ + { + "bbox": [ + 362, + 72, + 548, + 84 + ], + "lines": [ + { + "bbox": [ + 362, + 72, + 548, + 84 + ], + "spans": [ + { + "bbox": [ + 362, + 72, + 548, + 84 + ], + "type": "text", + "content": "(c): Mapping to " + }, + { + "bbox": [ + 362, + 72, + 548, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 362, + 72, + 548, + 84 + ], + "type": "text", + "content": " for generative modelling" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 89, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 359, + 89, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 359, + 89, + 555, + 196 + ], + "type": "image", + "image_path": "46cd639eaf6fb2e4abe776e82e3e5bd93b2bf9776689786754d0c428ddc7488b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 253, + 287, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 253, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 45, + 253, + 287, + 313 + ], + "type": "text", + "content": "jointly with the function parameters " + }, + { + "bbox": [ + 45, + 253, + 287, + 313 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 253, + 287, + 313 + ], + "type": "text", + "content": " on a training dataset " + }, + { + "bbox": [ + 45, + 253, + 287, + 313 + ], + "type": "inline_equation", + "content": "(i\\in D)" + }, + { + "bbox": [ + 45, + 253, + 287, + 313 + ], + "type": "text", + "content": ". Similar to single sketch optimisation, we train on multiple samples with a combination of MSE and visual loss (Eq. (6)). Fig. 4b summarises our multi-sketch autodecoder through block diagrams." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": "Generative Modelling: The generalised auto-decoder, enables simple generative modelling of sketches by generating their feature descriptors " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "\\nu_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": ", which allows us to create novel sketch implicits. Towards this, we train a Variational Auto-Encoder [35] as our generator, where we replace the decoder with our pre-trained implicit decoder. For variational inference, our encoder " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": " represents each raster sketch instance " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": " as a set of " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": " Gaussians with mean " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "\\mu_i\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "\\sigma_{i}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": " which are re-parameterised [35] to a latent encoding as " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "z_{i} = \\mu_{i} + \\sigma_{i}\\cdot \\epsilon" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": ". This latent is then projected to the " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": "-dimensional implicit latent space " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "\\nu_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": ". We train the encoder as a ResNet-18 network with (i) reconstruction loss from both our visual and vector spaces (Eq. (6)) and (ii) KL-divergence loss to bring the distribution closer to a unit normal " + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 313, + 289, + 504 + ], + "type": "text", + "content": ". The loss function for the encoder can be written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 515, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 515, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 77, + 515, + 287, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e n c}} = \\mathcal {L} _ {\\text {i m p l i c i t}} - \\beta \\cdot D _ {K L} \\left(q \\left(z _ {i} \\mid i\\right), \\mathcal {N} (\\mathbf {0}, \\mathbf {I})\\right) \\tag {7}", + "image_path": "1ee4939279b6aebcd197ad10866e2033b6dfdd4469b0c0576e9158fcfd725d36.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "text", + "content": " is a weighing factor, set to 0.7 and " + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "inline_equation", + "content": "q(z_i|i)" + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "text", + "content": " represents the underlying probability distribution modelled with encoder " + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "text", + "content": " that infers " + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "text", + "content": " for a given sketch " + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 537, + 288, + 610 + ], + "type": "text", + "content": ". Fig. 4c summarises our implicit sketch generative model. The decoder can be used for sketch generation, while the encoder can encode raster sketches for vectorisation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 621, + 127, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 127, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 127, + 635 + ], + "type": "text", + "content": "4. Applications" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "Datasets: We learn implicits for sketches of various complexities and abstraction levels, ranging from highly complex scene sketches that correspond to a photo scenery [13] to abstract doodles [28] drawn from memory. Based off complexity in downstream applications, we use scene sketches to demonstrate self-reconstruction and sketch" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "text", + "content": "compression, while focusing on simpler object level sketches for more complex tasks like sketch mixing (interpolation) and generation for faster optimisation. Specifically, we use the FS-COCO [13], Sketchy [53], and QuickDraw! [28] datasets. (i) FS-COCO [13] contains 10,000 hand-drawn sketches of complex MS-COCO [40] scenes, drawn by amateurs without time constraints. Sketches in FS-COCO have on a median of 64 strokes, resulting in the sketch having " + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "inline_equation", + "content": "\\sim 3000" + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "text", + "content": " waypoints. (ii) Comparatively simpler than scene sketches, Sketchy [53] consists of " + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "inline_equation", + "content": "\\sim 75\\mathrm{K}" + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "text", + "content": " sketches of " + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "inline_equation", + "content": "12.5\\mathrm{K}" + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "text", + "content": " object photographs from 125 categories. These sketches similar to FS-COCO are drawn from reference photographs and have photo-sketch pairs. (iii) QuickDraw! [28] sketches are drawn from memory and in a constrained time settings (" + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "inline_equation", + "content": "< 20\\mathrm{s}" + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "text", + "content": "). Hence, these sketches (doodles) are abstract in nature, lacking any pre-defined configuration or orientation from reference photos. In addition to these three datasets, we use a vectorised form of the MNIST dataset, Vector-MNIST [17], consisting of " + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "inline_equation", + "content": "\\sim 10\\mathrm{K}" + }, + { + "bbox": [ + 304, + 253, + 547, + 505 + ], + "type": "text", + "content": " samples as a simpler dataset to demonstrate complex downstream applications like generation and interpolation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 514, + 538, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 538, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 538, + 528 + ], + "type": "text", + "content": "4.1. SketchINR: A Compact High-Fidelity Codec" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "We begin by demonstrating that SketchINR provides a high fidelity neural representation for sketches. Specifically, we train SketchINR on complex variable length vector sketch datasets such as FS-COCO and Sketchy. We represent each sketch in these datasets as a fixed size vector with a dataset-specific decoder " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": ". We then render each implicit sketch " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\nu_{i} \\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " with the same number of strokes as the ground-truth sketch " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^i" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " to evaluate reconstruction quality. Following [17, 28], we measure reconstruction quality using (i) Chamfer Distance (CD), and (ii) retrieval accuracy (top-10) with a naive sketch retrieval network (ResNet-18) trained on ground-truth sketch vectors " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " and evaluated on rendered " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": ". The quantitative results in Tab. 1, show that SketchINR provides substantially higher fidelity encoding than alternative representations like RNN [28], Bézier strokes [16]," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12569" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 144 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 144 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 144 + ], + "type": "table", + "html": "
FS-COCO [13]Sketchy [53]
CDR@10Time (s)CDR@10Time(s)
SketchRNN [28]---0.1641.650.11
CoSE [2]---0.1445.820.10
SketchODE [17]0.415.171.680.1737.200.13
BézierSketch [16]0.572.961.220.2325.720.08
Ours0.01115.610.30.00857.290.0007
", + "image_path": "c436081b815089d448db45b554119c70d0a96d8a63d997a3ee8f3a09cf909e51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "lines": [ + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "type": "text", + "content": "Table 1. Evaluating neural sketch representations' reconstruction quality (Chamfer distance " + }, + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "type": "text", + "content": "; recall@ " + }, + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "type": "inline_equation", + "content": "10\\uparrow" + }, + { + "bbox": [ + 47, + 149, + 287, + 172 + ], + "type": "text", + "content": ") and decoding speed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 184, + 286, + 403 + ], + "blocks": [ + { + "bbox": [ + 50, + 184, + 286, + 403 + ], + "lines": [ + { + "bbox": [ + 50, + 184, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 50, + 184, + 286, + 403 + ], + "type": "image", + "image_path": "adea143865d18c7223b4c716cfdbb2e36a60b7d26bb8b33d56dea1c1de6b1c1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 410, + 288, + 443 + ], + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 443 + ], + "type": "text", + "content": "Figure 5. Qualitative reconstruction results for neural sketch representations. SketchINR is uniquely able to scale to sketches of Sketchy and FS-COCO complexity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 459, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 532 + ], + "type": "text", + "content": "ODE dynamics [17], and stroke embeddings [2] particularly for extremely complex sketches in FS-COCO, where SketchRNN and CoSE fail entirely. Qualitative results in Fig. 5 show SketchINR to reconstruct simple and complex sketches with high fidelity while SketchRNN performs poorly for Sketchy and fails completely on FS-COCO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": "Sketch Compression: We next show that SketchINR's high-fidelity encoding is extremely compact, thus providing an efficient sketch codec. INRs store information implicitly [55] in the network weights " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " and latent codes " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\nu_{i})" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": ". To store or transmit a sketch dataset, we only stream weights " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " once (fixed cost) and a compressed latent vector " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\nu_{i} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " for each sketch (variable cost). Hence, the space complexity for " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{D}" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " sketches with SketchINR is " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "|\\theta| + N_{D} \\times |\\nu_{i}|" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " bytes, where " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " is 64 dimensional latent. For streaming a large number of sketches, we discount " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " as negligible streaming overhead. Comparing to alternative representations, streaming " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{D}" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " (i) binary raster (BR) sketches of size " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "(N_{x} \\times N_{y})" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " take " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{D} \\times N_{x} \\times N_{y}" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " bits, and (ii) vector scene sketches with " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " coordinates and pen states " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{2+1})" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " take " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{D} \\times N \\times 2 \\times 3" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " bytes in 16-bit floating precision (" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "N \\approx 3000" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " in FS-COCO" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "[13]). Raster representations can be further stored by only storing the coordinates " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^2)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " of pixels containing the sketch (i.e., inked pixels) as sparse binary rasters (SBR). Fig. 6 shows rate-distortion for storing 10,000 scene sketches [13] with " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " varying from 128 to 8 dimensional latent, and quality for raster and vector forms dropped with downsampling and RDP simplification [23] respectively. Quality metrics like PSNR and SSIM are more suited to photos. We use Chamfer Distance as a quality metric for vector sketch and observe " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\sim 10\\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\sim 60\\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " more efficiency than vector and raster sketches, respectively. Future works can further optimise our models and thus reduce storage requirements via model compression [21, 29, 56] and quantisation [11]." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 239, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 307, + 239, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 324 + ], + "type": "image", + "image_path": "d1c6198caa03f00604f631ac1001a30e0a4ff7eaecd8419e117a6b8a4cb5f0f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "type": "text", + "content": "Figure 6. Rate Distortion: SketchINR can encode complex scene sketches from FS-COCO [13] in highly compact " + }, + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{64})" + }, + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "type": "text", + "content": " latent codes. Specifically, despite nearly identical sketch quality (low CD represents higher fidelity), SketchINR has " + }, + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "type": "inline_equation", + "content": "\\sim 60\\times" + }, + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "type": "text", + "content": " lower BPP than PNG raster sketches and " + }, + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "type": "inline_equation", + "content": "\\sim 10\\times" + }, + { + "bbox": [ + 304, + 330, + 547, + 385 + ], + "type": "text", + "content": " lower than vectors sketches." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": "Intra-Sketch Variation: Unlike prior representations (RNN [28], ODE dynamics [17]), humans do not draw exact stroke-level replica of a raster sketch. Instead, we choose the level of detail, the number of points, and the strokes used to convey the underlying shape. Given a latent code " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": ", SketchINR allows users to explicitly choose " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": ", where the number of points " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "p = \\{p_1,\\dots ,p_N\\}" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": " and strokes " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "s = \\{s_1,\\ldots ,s_K\\}" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": ". Choosing a low " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": " leads to abstract self-reconstructions with fewer points and strokes. Fig. 7(top to bottom) shows that we can control " + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 400, + 547, + 533 + ], + "type": "text", + "content": " to decrease or increase the level detail in the sketch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 540, + 541, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 541, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 541, + 553 + ], + "type": "text", + "content": "4.2. Latent Space Interpolation (Creative Mixing)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "Interpolation of learned latents allows us to explore latent space continuity. Auto-regressive modelling of sketches builds poor latents as explored by Das et. al. [19] due to the lack of a full visibility of the sketch at any decoding timestep. Specifically, auto-regressive models use the partially predicted sketch along with the sketch's latent representation to predict new way-points with consistency. This, however, introduces noise in the latent " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " sketch decoding so that small changes in these latents lead to big changes in the final sketch. Deterministic latent to sketch decoding solves this problem as noise is not added at any point, leading to meaningful sketch interpolation from corresponding latent space interpolations. We visualise in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12570" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 284, + 200 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 284, + 200 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 284, + 200 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 284, + 200 + ], + "type": "image", + "image_path": "b46ea493f27fe1a381abda23a10a47263c0590d32d66c3d01d509d97e455da19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "type": "text", + "content": "Figure 7. Given a latent code " + }, + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "type": "text", + "content": ", we can reconstruct a sketch using learned " + }, + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "type": "text", + "content": " at multiple resolutions by controlling the number of strokes " + }, + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 209, + 288, + 242 + ], + "type": "text", + "content": " during sampling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 248, + 264, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 248, + 264, + 258 + ], + "spans": [ + { + "bbox": [ + 52, + 248, + 264, + 258 + ], + "type": "text", + "content": "Cross-Sketch Interpolation with auto-regressive networks - SketchRNN" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 262, + 279, + 322 + ], + "blocks": [ + { + "bbox": [ + 53, + 262, + 279, + 322 + ], + "lines": [ + { + "bbox": [ + 53, + 262, + 279, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 279, + 322 + ], + "type": "image", + "image_path": "5533c1a13f69b71f81b46f45ee080bf8a47161f049f4c5215d8834eaaf7872ca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 327, + 216, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 216, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 216, + 337 + ], + "type": "text", + "content": "Cross-Sketch interpolation with implicits - SketchINR" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 342, + 279, + 397 + ], + "blocks": [ + { + "bbox": [ + 53, + 342, + 279, + 397 + ], + "lines": [ + { + "bbox": [ + 53, + 342, + 279, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 279, + 397 + ], + "type": "image", + "image_path": "a0d842a57d73d2ac8488ffdc6888a567f9ffff78778527e2102e93cfbbc39f74.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 406, + 287, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 429 + ], + "type": "text", + "content": "Figure 8. Latent space interpolation shows continuity and translation from one sketch concept to another." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "content": "Fig. 8 how one sketch morphs into another by reconstructing from " + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\nu', t_n, \\Delta t, s_k, \\Delta s)" + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\nu'" + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "content": " is sampled along a latent walk: " + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\nu' = (1 - \\delta)\\nu_1 + \\delta \\nu_2" + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\delta \\in [0,1]" + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "content": ". Interpolating between two sketches in the continuous latent space gives creative mixing – modifying the sketch of airplane-1 (facing right) into airplane-2 (facing left) by varying " + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 443, + 287, + 575 + ], + "type": "text", + "content": ". Fig. 8 shows that small changes in the latent (columns) leads to smoother changes in sketch space for SketchINR. Meanwhile SketchRNN can exhibit big jumps that can zig-zag over successive latent increments, as indicated by an example tracked point in the figure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 239, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 239, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 239, + 599 + ], + "type": "text", + "content": "4.3. INR inversion for sketch completion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Neural networks used to approximate implicit representations have been inverted in parallel literature [30] to appropriate samples not seen during training. This inversion, to arrive at a latent " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " that describes the new sample, is usually made possible by learned priors in " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". Specifically, an implicit decoder " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " is frozen with latent " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " optimised for " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\nu, \\ldots)" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " against the ground truth. This behaviour is easily replicated in SketchINR, where we optimise the latent " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " to describe sketch samples (Fig. 5). To implement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 67, + 528, + 77 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 67, + 528, + 77 + ], + "spans": [ + { + "bbox": [ + 312, + 67, + 528, + 77 + ], + "type": "text", + "content": "(a) Input Partial Stroke Vector-MNIST (Left) and Quick-Draw! (Right)" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 77, + 541, + 112 + ], + "blocks": [ + { + "bbox": [ + 310, + 77, + 541, + 112 + ], + "lines": [ + { + "bbox": [ + 310, + 77, + 541, + 112 + ], + "spans": [ + { + "bbox": [ + 310, + 77, + 541, + 112 + ], + "type": "image", + "image_path": "7898dce9327a887394bedb35dcbbd9accd32630f7ecad25846834534df9aa74a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 121, + 541, + 158 + ], + "blocks": [ + { + "bbox": [ + 361, + 114, + 479, + 122 + ], + "lines": [ + { + "bbox": [ + 361, + 114, + 479, + 122 + ], + "spans": [ + { + "bbox": [ + 361, + 114, + 479, + 122 + ], + "type": "text", + "content": "Sketch Completion using SketchRNN" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 121, + 541, + 158 + ], + "lines": [ + { + "bbox": [ + 310, + 121, + 541, + 158 + ], + "spans": [ + { + "bbox": [ + 310, + 121, + 541, + 158 + ], + "type": "image", + "image_path": "9f1ca3a3ef16062faf1340b788350408e9f7df30f0043f1bf990f966bb43514d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 158, + 477, + 167 + ], + "lines": [ + { + "bbox": [ + 364, + 158, + 477, + 167 + ], + "spans": [ + { + "bbox": [ + 364, + 158, + 477, + 167 + ], + "type": "text", + "content": "Sketch Completion using SketchINR" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 167, + 539, + 204 + ], + "blocks": [ + { + "bbox": [ + 309, + 167, + 539, + 204 + ], + "lines": [ + { + "bbox": [ + 309, + 167, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 309, + 167, + 539, + 204 + ], + "type": "image", + "image_path": "a8b8505fc47496acf52cf3d713808f03d5d485f34ffeea1fc5dc17528d3aa569.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 209, + 533, + 241 + ], + "blocks": [ + { + "bbox": [ + 308, + 209, + 533, + 241 + ], + "lines": [ + { + "bbox": [ + 308, + 209, + 533, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 533, + 241 + ], + "type": "image", + "image_path": "2cd8f870418982026c14026f1fa834ad6392f262952f41dcaab0e641a2e2e863.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 244, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 304, + 244, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 545, + 310 + ], + "type": "text", + "content": "Figure 9. We perform sketch completion by inverting learned implicit representations and decoding for occluded strokes or waypoints. (a) Temporal completion (b) Unlike auto-regressive methods like SketchRNN [28], we can perform a-temporal (or inverse-temporal) completion, where given the second half of a partial sketch, we can successfully reconstruct the first part." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "content": "sketch completion, we obtain a sketch descriptor " + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "content": " from a partial sketch by just optimising for specific time-steps (e.g. 1/4th of a sketch; " + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "inline_equation", + "content": "t \\in [0, \\frac{1}{4})" + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "content": ") available in a partial doodle. Then, by sampling for rest of the timesteps we can complete the incomplete sketch. We summarise some qualitative sketch completion results in Fig. 9(a). Completed sketches are varied because of randomness in optimisation, as partial sketches are often ambiguous. INR inversion thus provides a novel approach to sketch completion, previously dominated by auto-regressive and models. Unlike auto-regressive models SketchINR can also perform a-temporal completion, inferring the first half of a sketch given the second half of the sketch (Fig. 9b). We remark that popular INR applications such as super-resolution [47] and point-cloud densification [48] perform interpolation after inversion on sparse inputs (e.g.: " + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "inline_equation", + "content": "t \\in [0, 1)" + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "inline_equation", + "content": "t \\in \\{0, \\frac{1}{4}, \\frac{1}{2}, \\frac{3}{4}\\}" + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "content": "), while sketch completion requires extrapolation (e.g.: " + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "inline_equation", + "content": "t \\in [0, 1)" + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "inline_equation", + "content": "t \\in [0, \\frac{1}{4})" + }, + { + "bbox": [ + 304, + 322, + 546, + 550 + ], + "type": "text", + "content": "), making it more analogous to visual outpainting." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 561, + 416, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 416, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 416, + 572 + ], + "type": "text", + "content": "4.4. Sketch Generation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "type": "text", + "content": "We can perform conditional and unconditional sequential vector sketch generation after learning a generative model for sketch latents as described in Sec. 3.3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Unconditional Generation: We unconditionally generate sketches by random sampling " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " with the learnt VAE latent space, obtaining novel sketch descriptors " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\nu_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ". We further sample " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\nu_{i},\\dots)" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " with varying timestamps " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "T\\in [100,300]" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " and strokestamps " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "K\\in [10,30]" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " obtaining sketches at varying levels of abstraction. Qualitative results in Fig. 10 show diverse and plausible samples from Vector-MNIST [17] and Quick-Draw! [28] datasets." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12571" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 287, + 176 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 287, + 176 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 287, + 176 + ], + "type": "image", + "image_path": "c071b249dd46dd7892a71ac96dca3d64f907bb483cce2bf2664dbe6d8249f73b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 288, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 288, + 208 + ], + "type": "text", + "content": "Figure 10. Unconditional sketch generation. Sampling sketch latents " + }, + { + "bbox": [ + 46, + 185, + 288, + 208 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 185, + 288, + 208 + ], + "type": "text", + "content": " as well as stroke and point complexity " + }, + { + "bbox": [ + 46, + 185, + 288, + 208 + ], + "type": "inline_equation", + "content": "(K, T)" + }, + { + "bbox": [ + 46, + 185, + 288, + 208 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 350 + ], + "type": "text", + "content": "Conditional Generation: To condition sketch generation, we use our pre-trained VAE encoder " + }, + { + "bbox": [ + 46, + 217, + 289, + 350 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 217, + 289, + 350 + ], + "type": "text", + "content": " to embed raster sketches, which can then be vectorised by the decoder " + }, + { + "bbox": [ + 46, + 217, + 289, + 350 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 217, + 289, + 350 + ], + "type": "text", + "content": ". Fig. 11 shows qualitative results for sequential vector sketch generation from raster images, where we obtain comparable results with auto-regressive generative modelling [28]. We observe that treating vectorisation as a generative problem rather than a deterministic raster-to-vector mapping [7] leads us to have variations in generated vectors particularly in the form of stroke order resembling real copies of sketches made by humans." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 355, + 287, + 428 + ], + "blocks": [ + { + "bbox": [ + 48, + 355, + 287, + 428 + ], + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 428 + ], + "type": "image", + "image_path": "ddf0fc6e578f25e0212718635520cfa79c9b40491e03db5197833cfb8968344b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 449, + 287, + 481 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 481 + ], + "type": "text", + "content": "Figure 11. Conditional sketch generation (vectorisation) on Vector-MNIST. Raster sketches (top left) can have multiple vector forms, e.g., sketched with different numbers of strokes (bottom)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 188, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 188, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 188, + 503 + ], + "type": "text", + "content": "4.5. Ablations and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Global vs Local Time Modelling: We model " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " timestamps across " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " strokes, by distributing timestamps uniformly over each stroke such that each stroke has " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{J}{K}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " timestamps. In other words, time " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "(t,\\Delta t)" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " is modelled globally for the entire sketch " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\{\\mathrm{p}_1,\\dots ,\\mathrm{p}_N\\}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ". An alternative is to model " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "(t,\\Delta t)" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " as 0to1 for each stroke separately (local time-modelling) instead of the entire sketch. Particularly, each of the " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " strokes are modelled using " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " points, representing the entire sketch using " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "N\\cdot K" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " points as " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\{\\mathrm{p}_1,\\dots ,\\mathrm{p}_{N\\cdot K}\\}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ". Visually inspecting global vs local time embedding shows that local modelling leads to jittery sketches (see Fig. 12), whereas global modelling gives visually smooth sketches. We hypothesise that time is a global property of a sketch and not an independent property localised to each stroke. Hence, modelling time locally puts an additional optimisation burden on our implicit decoder to ensure smoothness and continuity, resulting in poorly reconstructed sketches." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 541, + 125 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 541, + 125 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 541, + 125 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 541, + 125 + ], + "type": "image", + "image_path": "3f8e4975916570af285cfd13ce622394fb52e77f569e562e5f0f99ade64f6219.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 126, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 305, + 126, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 545, + 148 + ], + "type": "text", + "content": "Figure 12. Global vs Local time embedding: Global embedding gives smooth sketches compared to local, which gives jittering." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": "Fixed vs Variable Smoothing Factor " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": ": The intensity map " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_k" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " in Eq. (4) guides our raster-based visual loss " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_V" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": ". A key component in " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_k" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " is the smoothing parameter " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\gamma -" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " low " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " (thicker " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_k" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": ") gives less accurate but smoother values on the entire visual region (easier training); high " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " (thinner " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_k" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": ") gives more accurate but sharper (steeper) intensity-gradients near ground-truth strokes (harder training). While we use a fixed " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " with MSE Loss in Eq. (6) to balance out the optimisation complexity, we note that an alternative is to vary " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " every fixed number of steps with a scheduler to allow faster optimisation. Specifically, we vary " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " as 20to200 on a linear scale by incremental " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\Delta \\gamma" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": " every few iterations. Despite its theoretical advantage [25, 45], our initial experiments suggest only minor improvement that reduces training time by " + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 304, + 153, + 547, + 346 + ], + "type": "text", + "content": ". A detailed study of bundle-adjustment [45] for visual loss is an interesting future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "type": "text", + "content": "Limitations and Future Work Despite it's efficiency as a representation for hand-drawn sketches, SketchINR suffers from a number of limitations. We naturally share limitations with other implicit representations including optimisation-based encoding and poor cross category generalisation. While unlike Sketch-RNN our generation network can be trained on 2-3 categories at once, it fails to generalise on more. In addition, a core limitation lies in slow convergence due to pixel space optimisation. While vector-space loss in the form of mean squared error mitigates this to some extent (Fig. 3), convergence is still slow and can possibly be improved by better engineering the implicit function. Strokes could also be smoothed by explicit inclusion of stroke gradient (slope) based regularisation in future work. Finally, vectorisation for conditional generation performs poorly on high complexity sketches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 544, + 379, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 379, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 379, + 555 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "We introduce visual implicits to represent vector sketches with compressed latent descriptors. This neural representation provides a high-fidelity and compact representation that raises the possibility of a sketch-specific codec for compactly representing large sketch datasets. Our SketchINR can decode sketches at varying levels of detail with controllable number of strokes, and provides superior cross-sketch interpolation between implicits, demonstrating a smoother latent space than auto-regressive models. Applications such as sketch-completion are also supported, including a-temporal completion not available with auto-regressive models. Decoding is inherently parallel and can be over " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " faster than autoregressive models in practice." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12572" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Emre Aksan, Fabrizio Pece, and Otmar Hilliges. Deepwriting: Making digital ink editable via deep generative modeling. In CHI, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "type": "text", + "content": "[2] Emre Aksan, Thomas Deselaers, Andrea Tagliasacchi, and Otmar Hilliges. Cose: Compositional stroke embeddings. In NeurIPS, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "text", + "content": "[3] Stephan Alaniz, Massimiliano Mancini, Anjan Dutta, Diego Marcos, and Zeynep Akata. Abstracting sketches through simple primitives. In ECCV, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 288, + 235 + ], + "type": "text", + "content": "[4] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks. In CVPR, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[5] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle your 3d: From abstract freehand sketches to precise 3d shapes. In CVPR, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 345 + ], + "type": "text", + "content": "[6] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can beat me? In SIGGRAPH Asia, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "text", + "content": "[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "type": "text", + "content": "[8] Alexandre Binninger, Amir Hertz, Olga Sorkine-Hornung, Daniel Cohen-Or, and Raja Giryes. Sens: Part-aware sketch-based implicit neural shape modeling. arXiv preprint arXiv:2306.06088, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 457 + ], + "type": "text", + "content": "[9] Jack E Bresenham. Algorithm for computer control of a digital plotter. IBM Systems journal, 1965. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 288, + 491 + ], + "type": "text", + "content": "[10] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 523 + ], + "type": "text", + "content": "[11] Qualcomm Innovation Center. Ai model efficiency toolkit (aimet). https://github.com/quic/aimet, 2023.6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 557 + ], + "type": "text", + "content": "[12] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. In NeurIPS, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "text", + "content": "[13] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "text", + "content": "[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aaneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What can human sketches do for object detection? In CVPR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "text", + "content": "[15] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In ECCV, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Sketchode: Learning neural sketch representation in continuous time. In ICLR, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "text", + "content": "[18] Ayan Das, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Cloud2curve: Generation and vectorization of parametric sketches. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "text", + "content": "[19] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Chirodiff: Modelling chirographic data with diffusion models. In ICLR, 2023. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 188, + 545, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 211 + ], + "type": "text", + "content": "[20] Carl De Boor and Carl De Boor. A practical guide to splines. Applied Mathematical Sciences, 1978. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 212, + 545, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 246 + ], + "type": "text", + "content": "[21] Zhen Dong, Zhewei Yao, Amir Gholami, Michael Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In ICCV, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 247, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 247, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 545, + 270 + ], + "type": "text", + "content": "[22] Asen L. Dontchev and R. Tyrrell Rockafellar. Implicit Functions and Solution Mappings. Springer-Verlag, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 271, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 326 + ], + "type": "text", + "content": "[23] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 1973. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 327, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 350 + ], + "type": "text", + "content": "[24] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 352, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 545, + 374 + ], + "type": "text", + "content": "[25] F. Dan Foresee and Martin T. Hagan. Gauss-newton approximation to bayesian learning. In ICNN, 1997. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "type": "text", + "content": "[26] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, SM Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In ICML, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 421, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 444 + ], + "type": "text", + "content": "[27] Songwei Ge, Vedanuj Goswami, C. Lawrence Zitnick, and Devi Parikh. Creative sketch generation. In ICLR, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 445, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 468 + ], + "type": "text", + "content": "[28] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 469, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 469, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 469, + 545, + 502 + ], + "type": "text", + "content": "[29] Hai Victor Habi, Roy H. Jennings, and Arnon Netzer. Hmq: Hardware friendly mixed precision quantization block for cnns. In ECCV, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 504, + 545, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 538 + ], + "type": "text", + "content": "[30] Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. Spaghetti: Editing implicit shapes through part aware generation. ACM TOG, 2022. 2, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 539, + 545, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 539, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 307, + 539, + 545, + 561 + ], + "type": "text", + "content": "[31] Aaron Hertzmann. Why do line drawings work? a realism hypothesis. Perception, 2020. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 563, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 545, + 596 + ], + "type": "text", + "content": "[32] Conghui Hu, Da Li, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Sketch-a-classifier: Sketch-based photo classifier generation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 597, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 597, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 307, + 597, + 545, + 631 + ], + "type": "text", + "content": "[33] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-segmenter: Sketch-based photo segmenter generation. IEEE TIP, 2020. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 632, + 545, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 632, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 632, + 545, + 665 + ], + "type": "text", + "content": "[34] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 666, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 545, + 689 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "text", + "content": "[36] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 3" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12573" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "type": "text", + "content": "[37] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In CVPR, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 286, + 150 + ], + "type": "text", + "content": "[38] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Free2cad: Parsing freehand drawings into cad commands. ACM TOG, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "type": "text", + "content": "[39] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 241 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 244, + 287, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 287 + ], + "type": "text", + "content": "[41] Hongyu Liu, Ziyu Wan, Wei Huang, Yibing Song, Xintong Han, Jing Liao, Bin Jiang, and Wei Liu. Deflocnet: Deep image editing via flexible low-level controls. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 290, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 322 + ], + "type": "text", + "content": "[42] Asif Masood and Sidra Ejaz. An efficient algorithm for robust curve fitting using cubicBezier curves. In ICIC, 2010. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 325, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 325, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 325, + 287, + 367 + ], + "type": "text", + "content": "[43] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "type": "text", + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "type": "text", + "content": "[45] Jorge J. Mofe. The levenberg-marquardt algorithm. In Numerical Analysis, 1976. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 440, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 472 + ], + "type": "text", + "content": "[46] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 506 + ], + "type": "text", + "content": "[47] Quan H Nguyen and William J Beksi. Single image superresolution via a dual interactive implicit neural network. In WACV, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 509, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 552 + ], + "type": "text", + "content": "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "text", + "content": "[49] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In ICML, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[50] Michael Revow, Christopher KI Williams, and Geoffrey E Hinton. Using generative models for handwritten digit recognition. IEEE TPAMI, 1996. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[51] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[52] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 590 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[53] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 1, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[54] Heping Sheng, John Wilder, and Dirk B. Walther. Where to draw the line? PLOS One, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "text", + "content": "[55] Yannick Strumpler, Janis Postels, Ren Yang, Luc van Gool, and Federico Tombari. Implicit neural representations for image compression. In ECCV, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[56] Stefan Uhlich, Lukas Mauch, Fabien Cardinaux, Kazuki Yoshiyama, Javier Alonso Garcia, Stephen Tiedemann, Thomas Kemp, and Akira Nakamura. Mixed precision dnns: All you need is a good parametrization. In ICLR, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "text", + "content": "[57] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM TOG, 2022. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[58] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In ICCV, 2023. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "type": "text", + "content": "[59] Alexander Wang, Mengye Ren, and Richard Zemel. Sketchembednet: Learning novel concepts by imitating drawings. In ICML, 2021. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "text", + "content": "[60] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "text", + "content": "[61] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. Sketchmate: Deep hashing for million-scale human sketch retrieval. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "text", + "content": "[62] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "type": "text", + "content": "[63] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[64] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 524, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 555 + ], + "type": "text", + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "type": "text", + "content": "[66] Wenni Zheng, Pengbo Bo, Yang Liu, and Wenping Wang. Fast b-spline curve fitting by 1-bfgs. Computer Aided Geometric Design, 2012. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12574" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_content_list.json b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..19e3671b5f81e29dcca9d1ddac754d94a1a0cdf5 --- /dev/null +++ b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_content_list.json @@ -0,0 +1,1798 @@ +[ + { + "type": "text", + "text": "SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution", + "text_level": 1, + "bbox": [ + 107, + 128, + 864, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhixuan Liang $^{1,3}$ Yao Mu $^{1,3}$ Hengbo Ma $^{2}$", + "bbox": [ + 295, + 202, + 668, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masayoshi Tomizuka2 Mingyu Ding $^{2\\dagger}$ Ping Luo $^{1,3\\dagger}$", + "bbox": [ + 267, + 220, + 707, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2University of California, Berkeley 3Shanghai AI Laboratory", + "bbox": [ + 98, + 243, + 867, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zxliang, ymu, pluo}@cs.hku.hk {hengbo.ma, tomizuka, myding}@berkeley.edu", + "bbox": [ + 156, + 263, + 810, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://skilldiffuser.github.io/", + "bbox": [ + 323, + 281, + 640, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 332, + 313, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have demonstrated strong potential for robotic trajectory planning. However, generating coherent trajectories from high-level instructions remains challenging, especially for long-range composition tasks requiring multiple sequential skills. We propose SkillDiffuser, an end-to-end hierarchical planning framework integrating interpretable skill learning with conditional diffusion planning to address this problem. At the higher level, the skill abstraction module learns discrete, human-understandable skill representations from visual observations and language instructions. These learned skill embeddings are then used to condition the diffusion model to generate customized latent trajectories aligned with the skills. This allows generating diverse state trajectories that adhere to the learnable skills. By integrating skill learning with conditional trajectory generation, SkillDiffuser produces coherent behavior following abstract instructions across diverse tasks. Experiments on multi-task robotic manipulation benchmarks like Meta-World and LOReL demonstrate state-of-the-art performance and human-interpretable skill representations from SkillDiffuser. More visualization results and information could be found on our website.", + "bbox": [ + 75, + 364, + 473, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 727, + 209, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent research [6, 7, 18, 19] has demonstrated diffusion models' superior generative capabilities compared to previous models that help enhance reinforcement learning across various dimensions, including the generation of action trajectories [2, 19], policy representation [5, 50], and data synthesis [14, 23]. However, their ability to generate coherent trajectories for intricate tasks still poses challenges in terms of performance and generalizability, as these tasks often re", + "bbox": [ + 75, + 753, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6656ca45fad5bf13f40a16020660d2f13b0ac8d09b01a8723a896f2f326f6ab.jpg", + "image_caption": [ + "Non-transferable" + ], + "image_footnote": [], + "bbox": [ + 524, + 345, + 616, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b27cdbb0feada732e1dc54ba9986bc764a77f02e17d3698481b46050f8b9fdb.jpg", + "image_caption": [ + "1" + ], + "image_footnote": [], + "bbox": [ + 545, + 419, + 606, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccfb012514736eb469a2ddd9e299e76cda4863ccc61704245f82aeded07579d8.jpg", + "image_caption": [ + "Denoise U-Net", + "#" + ], + "image_footnote": [], + "bbox": [ + 545, + 446, + 606, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b836bbfc981a16ec7aac44594fcdeab0b722f76c3cfc8861e78c355557d99ea.jpg", + "image_caption": [ + "Instruction Input" + ], + "image_footnote": [], + "bbox": [ + 516, + 513, + 616, + 531 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3672c9593dbb4b3c420c081dda63ded0befab90f551ec280017ee17bb160f784.jpg", + "image_caption": [ + "(a) Language-conditioned Diffuser", + "Transferable and Interpretable", + "Visual Input", + "Instruction Input", + "(b) Our SkillDiffuser", + "Figure 1. Comparison of SkillDiffuser and previous language conditioned diffusers. SkillDiffuser utilizes high-level abstraction to translate visual observations and language instructions into human understandable skills with language grounding. It then enables the low-level diffusion model condition on these skills, not only improving the execution performance of multi-step composition tasks but greatly enhancing the generalization and adaptability of the framework." + ], + "image_footnote": [], + "bbox": [ + 638, + 345, + 880, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "require the fulfillment of abstract instructions that consist of numerous coordination-intensive sequential steps.", + "bbox": [ + 496, + 703, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous approaches [2, 14], such as Decision Diffuser, aim to tackle this challenge by decomposing complex tasks into simpler sub-skills, organized within a predefined skill library. These methods rely on conditioning the diffusion model with one-hot skill representations to generate trajectories for each of these sub-skills. Nevertheless, these techniques encounter difficulties in attempting to autonomously learn end-to-end from a wide range of datasets, which hinders their scalability and ability to achieve end-to-end learning [1]. In addition, without explicitly learning reusable skills, models cannot capture intricate inter-step dependen", + "bbox": [ + 496, + 734, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16467", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cies and constraints, yielding fragmented and illogical trajectories. Decomposing ambiguous instructions into learnable sub-goals and skills can better enable models to follow logical step sequences, respect task structures, and transfer common procedures (e.g. reusable skill abstraction and adaptable skill-based diffusion) between different tasks. This skill-centric paradigm paves the way for diffusion models that can interpret and execute elaborate, abstract instructions necessitating numerous sequential steps.", + "bbox": [ + 75, + 90, + 468, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the above observations, we propose SkillDiffuser, a hierarchical planning framework unifying high-level skill learning with low-level conditional diffusion-based execution. As shown in Fig. 1, compared with previous language-conditioned diffusion policies, SkillDiffuser is able to interpret and execute complex instructions end-to-end with higher transferability. SkillDiffuser induces interpretable sub-latent goals by learning reusable skills tailored to the task instructions. The framework conditions a diffusion model on these learned skills to generate customized, coherent trajectories aligned with the overall objectives. By integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser achieves consistent, skill-oriented behavior without relying on a predefined skill library. Moreover, SkillDiffuser is designed to operate solely on visual observations, eliminating the need for robot proprioception (i.e. fully observed states). This end-to-end methodology, featuring learnable skills, enables SkillDiffuser to execute abstract instructions across a variety of tasks efficiently.", + "bbox": [ + 75, + 229, + 470, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SkillDiffuser works as follows. 1) A skill predictor with vector quantization [45] is used for high-level skill learning to distill tasks into discrete and interpretable skills. Rather than forecasting skill durations, we adopt a fixed prediction horizon - predicting skills at regular intervals. This horizon-based discretization process seamlessly integrates visual and linguistic inputs into a cohesive skill set guiding the low-level diffusion model. 2) For skill-based trajectory generation, we utilize a classifier-free diffusion model as policy, with skills directly embedded as guidance. This setting allows for generating multi-modal state trajectories aligned with skill specifications while avoiding overfitting to a closed dataset. 3) To enable action inference from predicted states, we train an inverse dynamics network to decode motions between two consecutive frames generated. By separating state prediction from motion decoding, SkillDiffuser yields a fully adaptive framework for directing diverse embodiments via transferable state-space plans.", + "bbox": [ + 75, + 534, + 468, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the model's performance in both skill learning and multi-task planning on the LOReL [29] Sawyer and Meta-World [51] datasets, with crucial experimental settings by considering robotic agents in real-world scenarios must operate with limited state information, chiefly visual observations. Furthermore, we present success rates on un", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "seen compositional tasks, the reusability and visualizations of learned skills to illustrate the model does have the ability to abstract high-level skills that are not only effective but also understandable to humans, bringing us closer to intelligent agents that acquire skills in a direct manner.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are three-fold: 1) We propose an end-to-end hierarchical planning framework via skill learning for sub-goal abstraction; 2) We adopt a classifier-free diffusion model conditioned on learned skills to generate skill-oriented transferable state trajectories; 3) We demonstrate state-of-the-art performance on complex benchmarks and provide interpretable visualizations of human-understandable skill representations.", + "bbox": [ + 496, + 167, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 303, + 648, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Imitation Learning and Multi-task Learning", + "text_level": 1, + "bbox": [ + 500, + 327, + 877, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imitation learning (IL) has evolved from foundational behavioral cloning to sophisticated multi-task learning frameworks. With traditional approaches relying on supervised learning from expert demonstrations [33, 37, 38], recent advancements have shifted towards learning the reward [16] or Q-function [12] from expert data, enhancing the ability to mimic complex behaviors. A new challenge lies in multi-task IL [41], where imitators are trained across varied tasks, aiming for generalization to new scenarios with task specifications ranging from vector states [28] to vision and language descriptions [8, 11, 13, 48].", + "bbox": [ + 496, + 351, + 890, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-task learning approaches often leverage shared representations to learn a spectrum of tasks simultaneously, enhancing the flexibility and efficiency of the learning process. The Meta-World benchmark [51] assesses multi-task and meta reinforcement learning, highlighting the need for algorithms capable of rapid adaptation. Building on this, the Prompting Decision Transformer [49] showcases few-shot policy generalization using task-specific prompts. And diffusion policy has also been explored in multi-task settings [14], which shows proficiency in generating diverse behaviors across tasks. However, unlike methods that leverage state inputs [14, 49] or access robot proprioception [30], SkillDiffuser uses raw visual inputs only.", + "bbox": [ + 496, + 518, + 890, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Skill Discovery and Hierarchical Learning", + "text_level": 1, + "bbox": [ + 500, + 724, + 861, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skill learning, the process by which robots acquire new abilities or refine existing ones, is gaining increasing attention due to its pivotal role in enabling autonomous systems to adapt to new tasks, improve the performance over time, and interact naturally with humans and complex environments. Traditionally, this domain was influenced by handcrafted features and expert demonstrations [32].", + "bbox": [ + 496, + 750, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the development of deep learning, Eysenbach et al. [9] and A. Sharma et al. [40] investigated skill discovery in learning methods, achieving policies conditioned", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16468", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on learned latent variables and maintaining consistent skill codes throughout trajectories. In the domain of skill learning through language instructions, LISA [13] stands out by sampling multiple skills per trajectory, integrating language conditioning in a unique manner.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our SkillDiffuser follows this way to extract sub-skills of each instruction at the higher-level. But differently, SkillDiffuser employs an adaptable diffusion policy at the lower-level to condition on different sub-skills to generate different actions, which formulates a creative hierarchical planning framework also advancing research on hierarchical techniques of reinforcement learning [22, 27, 52].", + "bbox": [ + 75, + 167, + 470, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Planning with Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 282, + 349, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [18] make great breakthroughs in image synthesis [18] recent years and has shown promising results in various generative applications. A seminal work that performs planning with diffusion directly is Diffuser [19], which laid the groundwork for using diffusion models in behavior synthesis. Then, a branch of this kind planning methods achieved state-of-the-art performance in a variety of decision tasks [3, 5, 7, 23, 31]. Among them, the work done by Chi et al. in [5] introduces the concept of learning the gradient of the action-distribution score function and iteratively optimizing it, demonstrating its significant potential in visuomotor policy learning. These works have further extended this direction and strengthened the versatility and generalization of diffusion-based planners.", + "bbox": [ + 75, + 306, + 468, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach is inspired by classifier-free diffusion guidance [17], which offers a significant advantage over classifier-guided models [6]. By adopting the classifier-free approach, we can circumvent the challenges associated with training a reward model and $Q$ -function, which are particularly cumbersome in many real-world planning scenarios of which the complexity is very high. Recent studies have also extended this direction, which use conditional diffusion models to generate customized trajectories. Decision Diffuser [2] is an example which is designed to generate trajectories for a predefined skill library. However, unlike our method, it can't autonomously learn skill abstractions in an end-to-end fashion, which makes it difficult to scale to more tasks. This highlights the necessity for diffusion models with dynamic, learnable skill abstractions, facilitating complex instruction execution.", + "bbox": [ + 75, + 518, + 468, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 776, + 202, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Planning with Diffusion over States", + "text_level": 1, + "bbox": [ + 76, + 801, + 383, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As introduced in previous works [2, 19], diffusion model is a promising tool to address the problem of planning in reinforcement learning which is cast as a Markov Decision Process (MDP) [34]. Within the MDP framework $\\mathcal{M} = (\\mathcal{S},\\mathcal{A},\\mathcal{T},\\mathcal{R},\\gamma)$ , the planning policy aims to iden", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tify an optimal action sequence $\\pmb{a}_{0:T}^{*}$ that maximizes the expected cumulative rewards over a finite time horizon $T$ , governed by the state transition dynamics $\\mathcal{T}$ and reward function $\\mathcal{R}$ . $S$ is the state space and $\\mathcal{A}$ is the action space.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By treating the state trajectory as sequence data $\\pmb{\\tau}$ , with sequence modeling, diffusion probabilistic models conceptualize planning as an iterative denoising process. The model progressively refines trajectories by reversing a forward diffusion process that is modeled as a Gaussian process, whereby noise is incrementally added to the data, denoted as $p_{\\theta}(\\pmb{\\tau}^{i-1} \\mid \\pmb{\\tau}^i)$ . Training involves minimizing the ELBO of the data's negative log-likelihood, similar to variational Bayesian inference, with the optimization objective:", + "bbox": [ + 496, + 151, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\arg \\min _ {\\theta} - \\mathbb {E} _ {\\boldsymbol {\\tau} ^ {0}} \\left[ \\log p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {0}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 290, + 890, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(\\boldsymbol{\\tau}^N)$ is a standard normal distribution and $\\boldsymbol{\\tau}^0$ denotes noiseless sequence data.", + "bbox": [ + 496, + 313, + 890, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For practical implementation, a simplified surrogate loss function is proposed in [18], focusing on predicting the Gaussian mean of the reverse diffusion step:", + "bbox": [ + 496, + 344, + 890, + 388 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e n o i s e}} (\\theta) = \\mathbb {E} _ {i, \\boldsymbol {\\tau} ^ {0} \\sim q, \\epsilon \\sim \\mathcal {N}} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {\\tau} ^ {i}, i) \\| ^ {2} ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 395, + 890, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Classifier-free Diffusion Guidance", + "text_level": 1, + "bbox": [ + 500, + 422, + 795, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the basis of unconditioned diffusion-based method, in the realm of offline reinforcement learning, a critical endeavor is to generate trajectories with the highest reward-to-go. With the flourishing development of conditional diffusion models [6], classifier-guided approaches embark on this by infusing specific trajectory information (encoded in $y(\\tau)$ ), such as the return $\\mathcal{J}(\\tau^0)$ or designated constraints, into the diffusion process:", + "bbox": [ + 496, + 445, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {\\tau} ^ {i} \\mid \\boldsymbol {\\tau} ^ {i - 1}\\right), \\quad p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i - 1} \\mid \\boldsymbol {\\tau} ^ {i}, \\boldsymbol {y} (\\boldsymbol {\\tau})\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 568, + 890, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With assumptions specified in [10], we have", + "bbox": [ + 500, + 592, + 792, + 607 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\tau} ^ {i - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} + \\alpha \\Sigma \\nabla_ {\\boldsymbol {\\tau}} \\log p (\\boldsymbol {y} (\\boldsymbol {\\tau} ^ {i})), \\Sigma\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 613, + 890, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter that adjusts the guidance strength, $\\Sigma$ is the specified covariance of the noise and $\\mu_{\\theta}$ is the learned mean value of noise in unconditional diffusion.", + "bbox": [ + 496, + 633, + 890, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the classifier-guided diffusion model requires an accurate estimation of guidance gradient based on the trajectory classifier $\\pmb{y}(\\pmb{\\tau})$ , which may not be feasible and need to introduce a separate dynamic programming procedure to estimate a $Q$ -function in the training process.", + "bbox": [ + 496, + 680, + 890, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, classifier-free guidance offers an alternative, which augments the trajectory generation process with a guidance signal that amplifies the features of high-reward or optimal characteristics, i.e. $y(\\tau)$ , that are implicitly present in the data. Mathematically, the noise to add during the reverse denoising process is:", + "bbox": [ + 496, + 755, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} = \\epsilon_ {\\theta} (\\boldsymbol {\\tau} ^ {i}, \\varnothing , i) + \\omega \\left(\\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, \\boldsymbol {y}, i\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, \\varnothing , i\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 849, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\omega$ is the guidance scale, and $\\varnothing$ represents the absence of guidance. Setting $\\omega = 0$ removes the classifier-free", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16469", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guidance towards an unconditional diffusion model, while a large value of $\\omega$ strengthens the influence of the conditional information during trajectory generation.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Also, the loss function to minimize can be rewritten as,", + "bbox": [ + 96, + 137, + 460, + 150 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} (\\theta) = \\mathbb {E} _ {i, \\boldsymbol {\\tau}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, (1 - \\beta) \\boldsymbol {y} \\left(\\boldsymbol {\\tau} ^ {i}\\right) + \\beta \\emptyset , i\\right) \\right\\| ^ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 151, + 473, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyperparameter that controls the probability of dropping specific condition $\\pmb{y}(\\pmb{\\tau})$ , enhancing sample diversity while maintaining context relevance.", + "bbox": [ + 76, + 190, + 468, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After training the noise prediction model $\\epsilon_{\\theta}$ with the above $\\mathcal{L}(\\theta)$ , the trajectory is sampled starting from Gaussian noise and progressively denoised with modified $\\hat{\\epsilon}$ through Eq. 5, employing re-parameterization technique.", + "bbox": [ + 76, + 234, + 468, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, through classifier-free guidance, we can modulate the trajectory sampling process to require the generated trajectories more aligned with the desired features represented by $\\pmb{y}(\\pmb{\\tau})$ . This process iteratively applies the conditioned noise model to refine the target trajectories that contain future states satisfying the constraints.", + "bbox": [ + 75, + 296, + 468, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 398, + 210, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 422, + 187, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building upon the motivations discussed above, we present SkillDiffuser, an advanced methodological framework for robust multi-task learning across various robots. This dynamic approach leverages the cooperation of skill learning at the higher level and a conditional diffusion model at the lower level. An overall framework is illustrated in Fig. 2. Notably, we leverage language-grounded representations to abstract skills, thereby rendering the execution of tasks through our diffusion policy both interpretable and comprehensible to humans.", + "bbox": [ + 75, + 446, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. High Level Interpretable Skill Abstractions", + "text_level": 1, + "bbox": [ + 76, + 604, + 444, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our SkillDiffuser framework, the high level interpretable skill abstraction module plays a crucial role in understanding and executing complex tasks. However, given the multitask environments we consider, each task with only a single language instruction may be broken down into a sequence of sub-tasks or skills, which are not explicitly delineated within the instruction itself. Furthermore, suppose the offline training dataset is denoted as $\\mathcal{D}$ , it consists of trajectories derived from a sub-optimal policy for various tasks. A trajectory $\\boldsymbol{\\tau} = (l, \\{\\boldsymbol{i}_t, \\boldsymbol{a}_t\\}_{t=1}^T)$ includes language description $l$ and a sequence of image observations and actions $(\\boldsymbol{i}_t, \\boldsymbol{a}_t)$ over time steps $T$ , with no reward labels attached. But the trajectories do not indicate the boundaries between sub-tasks as well, which thus requires the proposed methods capable of segmenting and interpreting the tasks into sub-goals in an unsupervised manner.", + "bbox": [ + 75, + 628, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we build a skill abstraction component upon a horizon-based skill predictor adapted", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db18d1540526ccb8eb5383397f98c047494aadddf21f202a52fdfb7eac6784fb.jpg", + "image_caption": [ + "Figure 2. Overall framework of SkillDiffuser. It's a hierarchical planning model that leverages the cooperation of interpretable skill abstractions at the higher level and a skill conditioned diffusion model at the lower level for task execution in a multi-task learning environment. The high-level skill abstraction is achieved through a skill predictor and a vector quantization operation, generating sub-goals (skill set) that the diffusion model employs to determine the appropriate future states. Future states are converted to actions using an inverse dynamics model. This unique fusion enables a consistent underlying planner across different tasks, with the variation only in the inverse dynamics model." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 857, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from GPT-2 [35]. It is designed to parse and decompose tasks by fusing visual input and natural language instructions, alongside a Vector Quantization (VQ) sub-module that discretizes the learned skills into a skill set. The specifics of this component are illustrated below.", + "bbox": [ + 496, + 551, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, as we utilize only images as robot state information, we transform the images into latent space features with a fixed image encoder (e.g. R3M [30]). For convenience, we denote the image encoder as $\\Phi_{im}:\\mathcal{I}\\to \\mathcal{R}^I$ , where $\\mathcal{I}$ represents the space of input images and $\\mathcal{R}^I$ , the resultant feature space, serves to condense the visual information into a form that is conducive for high-level semantics. Simultaneously, we use a language encoder to pre-process the natural language instructions, which is formalized as $\\Phi_{lang}:L\\rightarrow \\mathcal{R}^{L}$ , with $L$ being the space of language instructions and $\\mathcal{R}^L$ the language feature space. The outputs of both encoders are then fed into the skill predictor, which operates to integrate these two modalities.", + "bbox": [ + 496, + 628, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The skill predicting process is as follows: An image $\\boldsymbol{i}_t \\in \\mathcal{I}$ at time step $t$ is encoded into a visual embedding $s_t = \\Phi_{im}(\\boldsymbol{i}_t)$ . This embedding $s_t$ is then input to the skill predictor, along with the language instruction $l \\in L$ through the language encoder's output $l_t = \\Phi_{lang}(l)$ . The", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16470", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b8b722df06785f69ad27546734123b48ad4ba01e9688a0b22029a98de30a00fc.jpg", + "image_caption": [ + "Figure 3. SkillDiffuser's low level skill-conditioned diffusion planning model. Notably, while the schematic here employs images to represent visual features for illustrative purposes, in actual implementation, both the input to and the sampling output of the diffusion model are state embeddings. The current observation is also the feature embedding of current visual observation." + ], + "image_footnote": [], + "bbox": [ + 140, + 87, + 831, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "skill predictor, represented as $f: \\mathcal{R}^I \\times \\mathcal{R}^L \\to \\mathcal{C}$ , generates a skill code $\\tilde{z}$ by $\\tilde{z} = f(\\boldsymbol{s}_t, \\boldsymbol{l}_t)$ that encapsulates task's requirements interpreted from the visual and language inputs.", + "bbox": [ + 75, + 440, + 468, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After that, Vector Quantization [45] operation $\\mathbf{q}(\\cdot)$ is taken to transform these predicted skills into a low-dimensional discrete set $\\mathcal{C}$ . The discrete skill set contains $K$ skill embeddings $\\{z^1, z^2, \\ldots, z^K\\}$ which represent different potential skills. VQ operation is achieved by mapping the latent $\\tilde{z}$ to its closest entry of the skill set with skill vectors updated to be the moving average of the embeddings $z$ closest to them, same as [45]. This process is as follows,", + "bbox": [ + 75, + 486, + 468, + 608 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} = f \\left(\\Phi_ {i m} (\\boldsymbol {i} _ {t}), \\Phi_ {l a n g} (l)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 611, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz = \\mathbf {q} (\\tilde {z}) = \\underset {z ^ {k} \\in \\mathcal {C}} {\\arg \\min } \\| \\tilde {z} - z ^ {k} \\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 632, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VQ enforces each learnt skill $z$ to lie in $\\mathcal{C}$ , which is equal to learning $K$ prototypes for the language embeddings utilizing $k$ -means [24] algorithm. This acts as a bottleneck, restricting the flow of language information and aiding in the learning of discrete skill codes. The back-propagation through the non-differentiable quantization is achieved by a straight-through gradient estimator, which simply copies the gradients and enables model to be trained end-to-end.", + "bbox": [ + 75, + 659, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our approach, we apply a consistent skill code $z$ across a defined number of time steps, termed the horizon. This consistent application across multiple horizons adeptly addresses the challenge of varying sub-task durations without altering the horizon itself. Consequently, this strategy not only preserves the flexibility required for diverse task execution but also maintains the model's architectural stability by avoiding horizon-induced structural changes.", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, the discrete nature of the learned skill codes enhances the interpretability and controllability of the system's behavior, as each skill code is associated with some human-understandable language phrases. An example is depicted in Fig 6. Through this method, SkillDiffuser can not only learn to perform tasks based on language instructions but also achieve them in a human-interpretable way, allowing for a deeper understanding and control of the decision-making process of specific embodied agents.", + "bbox": [ + 496, + 440, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Low Level Skill-conditioned Diffusion Planning", + "text_level": 1, + "bbox": [ + 498, + 588, + 890, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As highlighted in Section 1, while existing approaches like Decision Diffuser [2] have introduced conditional diffusion models constrained by skills, their capability is limited to generating trajectories that meet only predefined skills. Consequently, these models fall short of realizing a diffusion model capable of conditioning on an arbitrary learned skill. To overcome these constraints and enable diffusion models to plan over a learned continuous spectrum of skills, we propose an approach that leverages the classifier-free guided diffusion model with the skills embedded.", + "bbox": [ + 496, + 612, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SkillDiffuser begins by employing a diffusion model to operate over the continuous skill space $\\mathcal{Z}$ learned during high-level skill abstraction. We employ a U-Net to serve as the noise prediction model $\\epsilon_{\\theta}(\\cdot)$ and guide it by in-context conditioning. More specifically, we firstly utilize a skill MLP (similar to point-wise feed-forward network [46]), denoted as $\\Lambda$ , to align skill features with denoising model. After that, we fuse the skill embeddings $\\Lambda(z)$ into the state features throughout the residual blocks of U-Net. Details", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are depicted in Fig. 3. In this way, we make the diffusion model contextually modulate the influence of skill embeddings at each step. And following Eq. 5, we can synthesize future states attending to these given skills. This is a significant shift from previous static conditioning framework to a more dynamic and adaptive trajectory generation process.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, following previous works [2, 7, 23], we adopt a state-only diffusion model, which eschews the direct generation of actions in diffusion model. And instead, we utilize another MLP, denoted as $\\Psi(\\cdot)$ , to perform inverse dynamics after state generation to infer feasible actions that can achieve transitions between two continuous states. Additionally, we integrate observations from the current frame to provide more detailed information for motion prediction and achieve closed-loop control. Mathematically, it is:", + "bbox": [ + 75, + 181, + 467, + 316 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {t} = \\Psi ([ \\tilde {\\boldsymbol {s}} _ {t}, \\tilde {\\boldsymbol {s}} _ {t + 1} ], \\boldsymbol {i} _ {t}) \\quad \\text {f o r} \\quad t = 0, \\dots , T - 1, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 321, + 467, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\boldsymbol{s}}_t$ and $\\tilde{\\boldsymbol{s}}_{t + 1}$ are consecutive observation embeddings within $\\tilde{\\tau}$ generated by diffusion model, $i_t$ is the current observation and $a_{t}$ is the inferred action.", + "bbox": [ + 76, + 343, + 467, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the resulting action sequence $\\{a_0, a_1, \\ldots, a_{T-1}\\}$ derived from the generated states, it encapsulates the skills to execute the tasks, which allows for remarkable adaptability across multiple tasks. And when faced with a new task, we are only required to change the inverse dynamics model $\\Psi(\\cdot)$ specific to the new task's kinematics, with the architecture and parameters of diffusion model unchanged. Such modularity ensures the generative capabilities of SkillDiffuser are not task-specific but can be leveraged across a diverse range of tasks with varying dynamics. A schematic diagram of the low level module is shown in Fig. 3.", + "bbox": [ + 75, + 388, + 467, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Training the SkillDiffuser", + "text_level": 1, + "bbox": [ + 76, + 564, + 310, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We engineered a threefold loss function for SkillDiffuser. Firstly, for the inverse dynamics model which is task-specific, we employ a behavior cloning loss [44] to train our inverse MLP emulating expert actions from observed state transitions. This loss, denoted as $\\mathcal{L}_{inv}$ , is formulated as:", + "bbox": [ + 75, + 585, + 467, + 661 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n v} = \\mathbb {E} _ {\\boldsymbol {\\tau} \\sim \\mathcal {D}} \\left[ \\left\\| \\boldsymbol {a} - \\Psi (\\tilde {\\boldsymbol {s}}, \\tilde {\\boldsymbol {s}} ^ {\\prime}, \\boldsymbol {i}) \\right\\| _ {2} ^ {2} \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 665, + 467, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the notations are similar to Eq. 9.", + "bbox": [ + 76, + 694, + 341, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Correspondingly, the other parts including both skill abstraction and low level execution are task-agnostic. For the high-level skill abstraction module, we apply a vector quantization (VQ) loss to refine the skill predictor. This VQ loss, $\\mathcal{L}_{\\mathrm{VQ}}$ , ensures the embeddings produced by the skill predictor closely match the skill set vectors, thereby improving the interpretability and consistency of the skill representations. Inspired by VQ-VAE [45], we formulate it as:", + "bbox": [ + 75, + 709, + 467, + 828 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V Q}} = \\mathbb {E} _ {\\boldsymbol {\\tau}} \\left[ \\| \\mathbf {q} (\\tilde {z}) - \\tilde {z} \\| _ {2} ^ {2} \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 833, + 467, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{z}$ follows Eq. 7.", + "bbox": [ + 76, + 854, + 230, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lastly, for the low-level state-only skill-conditioned diffusion execution, we incorporate a diffusion loss $\\mathcal{L}_{diff}$ as", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "per Eq. 6, ensuring our model's state predictions are in line with both the skill guidance and temporal dynamics observed in expert demonstrations. Here, we take $y(\\tau) = \\Lambda(z)$ with $z$ derived from Eq. 7 and Eq. 8.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To be noted, we train our SkillDiffuser with two optimizers, one for inverse dynamics model with $\\mathcal{L}_{inv}$ and the other for overall parameters of high-level skill abstraction and low-level planning with $\\mathcal{L}_{\\mathrm{VQ}} + \\lambda \\mathcal{L}_{diff}$ , where $\\lambda$ is a loss weight. This carefully constructed loss architecture enables SkillDiffuser to excel in a multi-task environment, generalizing across tasks by simply substituting the inverse dynamics model $\\Psi(\\cdot)$ specific to each new task's requirements.", + "bbox": [ + 496, + 152, + 890, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we utilize a pre-trained distilBERT [39] as our language encoder, adopting the configuration consistent with LOReL [29] and LISA [13], while freezing its parameters to guarantee stability in language understanding. And we employ diverse settings to serve as the visual encoder to ensure fair comparison in different datasets. We elaborate the details in corresponding parts of Sec. 5. More training details are shown in Appendix G and we also provide some pseudo-code of our algorithm in Appendix B.", + "bbox": [ + 496, + 273, + 890, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 426, + 630, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first present a comprehensive evaluation on the LOReL Sawyer dataset, and then perform the ablation study compared on Meta-World benchmark to illustrate the efficiency of our method. Finally, we visualize the learned skills of our method both on LOReL and Meta-World MT10.", + "bbox": [ + 496, + 452, + 890, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 540, + 601, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LOReL Sawyer Dataset [29] which is abbreviated from Language-conditioned Offline Reward learning, is composed of pseudo-expert trajectories or play data gathered from an arbitrary reinforcement learning policy, annotated with post-hoc crowd-sourced language directives. The LOReL Sawyer dataset encompasses 50k trajectories, each with 20 steps, in a simulated Sawyer robot environment. We assess our approach using the same six tasks as the original paper [29] with paraphrased instructions under five different situations (i.e. seen, unseen verb, unseen none, unseen verb + noun and human provided). This comes to a total of 77 instructions for all 6 tasks combined. More details about the dataset can be found in Appendix D.1.", + "bbox": [ + 496, + 564, + 890, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meta-World Dataset [51]. It is a comprehensive benchmark designed for evaluating multi-task and meta reinforcement learning algorithms. It introduces a comprehensive suite of 50 distinct robotic manipulation tasks, all located within a unified tabletop environment controlled by a simulated Sawyer arm. The Multi-Task 10 (MT10) within MetaWorld is a subset comprising ten carefully selected tasks, balanced in terms of diversity and complexity. Details of the ten tasks can be found in Appendix D.3.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16472", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2bc08da22ecf64a0db2483c0534ac4485cc9b901e71d00695ffc6d65a82df2ad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Task InstructionRandomLCBC [42]LCRL [20]Lang DT [13]LISA [13]SkillDiffuser
close drawer52%50%58%10%100%95 ± 3.2%
open drawer14%0%8%60%20%55 ± 13.3%
turn faucet left24%12%13%0%0%55 ± 9.3%
turn faucet right15%31%0%0%30%25 ± 4.4%
move black mug right12%73%0%20%60%18 ± 3.9%
move white mug down5%6%0%0%30%10 ± 1.7%
Average over tasks20%29%13%15%40%43 ± 1.1%
", + "bbox": [ + 135, + 88, + 834, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7573d4234c80256f828ed4f76bfca29aa3779ab027a0c838fd454959c5e99be0.jpg", + "table_caption": [ + "Table 1. Task-wise success rates on LOReL Sawyer Dataset. We show our success rates compared to random policy, language-conditioned imitation learning (LCBC), language-conditioned Q-learning (LCRL), a flat non-hierarchical Decision Transformer (Lang-DT), and LISA. The results on each dataset are calculated over 3 seeds. SkillDiffuser outperforms all other methods in terms of average performance over all tasks. Best methods and those within $6\\%$ of the best are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Rephrasal TypeLang DTLISA [13]SkillDiffuser
seen154043.65 ±4.7
unseen noun13.3333.3336.01 ±6.3
unseen verb28.333036.70 ±9.5
unseen noun+verb6.72042.02 ±3.8
human provided26.9827.3540.16 ±2.1
Average18.0730.1439.71
", + "bbox": [ + 78, + 287, + 468, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Evaluation Results on LOReL Sawyer Dataset", + "text_level": 1, + "bbox": [ + 76, + 460, + 468, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We employ random policy, language conditioned imitation learning (LCBC) [42], language conditioned Q-learning (LCRL) [20], Lang-DT (also known as Flat Baseline in [13]) and state-of-the-art skill-learning method LISA [13] as our baselines. We follow the same settings as LOReL [29] for the first three algorithms, and follow LISA [13] for the last two. The random policy serves as a baseline. And LCBC mimics offline dataset behaviors based on instructions, aligning with previous works focusing on imitation learning to achieve language conditioned behavior. In contrast, LCRL employs reinforcement learning, labeling each episode's final state with language-instructed rewards. Lang-DT plays as a non-hierarchical benchmark with a Causal Transformer [25], while LISA incorporates a dual-transformer structure, with one for skill prediction and the other for action planning. We do not compare with LOReL planner [29] as it uses MPC on a learned reward function relying on human annotations, while LISA and ours learn with trajectory data only.", + "bbox": [ + 75, + 484, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. To ensure fair comparison, our method, SkillDiffuser, is designed to maintain a parameter count similar to baseline models. It employs the same visual and language encoder architecture as used in LISA [13] with meticulously matching of embedding dimensions and the number of heads across the layers of SkillDiffuser.", + "bbox": [ + 75, + 777, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are present in Tables 1 and 2, showing taskwise and rephrasal-wise success rates for LOReL, averaged", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb563b0ec7a1c637853d1ec3fb4c4ba0bd6de1811e2bb656b096a92499d9946e.jpg", + "table_caption": [ + "Table 2. Rephrasal-wise success rates (in %) on LOReL Sawyer. Results of Lang DT, LISA and our SkillDiffuser are shown here. The standard error is calculated over 3 random seeds." + ], + "table_footnote": [], + "table_body": "
MethodLang DTLOReL [29]LISA [13]SkillDiffuser
Success Rate13.33 ±1.318.18 ±1.820.89 ±0.625.21 ±2.7
", + "bbox": [ + 503, + 287, + 890, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance on LOReL multi-step composition tasks.", + "bbox": [ + 500, + 325, + 890, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "over 10 runs with a 20-step time horizon. SkillDiffuser, our approach, achieves the highest average performance in six different tasks, indicating its superior cross-task adaptability particularly when compared to similar approaches which yet are based on Decision Transformer [4], such as LISA [13]. Moreover, SkillDiffuser consistently excels in all rephrased types for LOReL test tasks, outperforming LISA by $9.6\\%$ on average. This demonstrates the model's robustness in handling varied skill representations, marking a notable advancement in skill-conditioned diffusion model.", + "bbox": [ + 496, + 345, + 890, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Performance on LOReL Compositional Tasks", + "text_level": 1, + "bbox": [ + 498, + 506, + 885, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. We conduct experiments following the same settings of unseen composition tasks of LISA [13] with 12 composition instructions in Tab. 3. Detailed instructions are listed in Appendix D.2 with such an example that \"open drawer and move black mug right\". We extend the max number of episode steps from customary 20 to 40, as LISA.", + "bbox": [ + 496, + 529, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We observe SkillDiffuser achieves 2x the performance of non-hierarchical baseline (i.e. w/o skill abstraction) and also improves about $25\\%$ over LISA, highlighting its effectiveness. MPC-based LOReL planners are unable to perform as well in open scenarios like composition tasks.", + "bbox": [ + 496, + 622, + 890, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study on Meta-World Dataset", + "text_level": 1, + "bbox": [ + 498, + 707, + 836, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. We conduct experiments on Meta-World MT10 benchmark with finely annotated instructions. We also use visual observations only. Details are in Appendix D.3.", + "bbox": [ + 496, + 729, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We evaluate our method against three baselines, all modified from existing models. The first, Flat R3M, is adapted from R3M [30] paper's planner. As the original one utilizes the first four terms of robot proprioception, we eliminate them and make the planner focus exclusively on visual observations. The second baseline is a variant of our SkillDiffuser, lacking the high-level skill abstraction module but retaining the low-level conditional diffusion-based planner,", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5586fe43ee72ae2c1eec5b91c2b3393c9cc200069e1eb6f2deb570fbc0b91f1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLangSkill SetDiffuserPerformance
Flat R3M [30]XXX13.3%
LISA [13]✓✓X13.8%
Lang Diffuser✓X✓16.7%
SkillDiffuser✓✓✓23.3%
", + "bbox": [ + 83, + 88, + 460, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to assess the impact of skill abstraction. This version integrates a two-layer MLP to predict options from visual and language inputs, functioning as a language-conditioned diffusion planner. The last baseline is our re-implemented LISA [13] for Meta-World Dataset to validate the efficiency of diffusion model. To ensure fairness, we use R3M as the visual encoder for all of these methods on Meta-World.", + "bbox": [ + 75, + 220, + 467, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Language Skill Conditioning. Table 4 indicates our Meta-World MT10 task is quite different from and much more challenging than previous tasks that use states as observations [49] or take into robot proprioception [30]. We only utilize single-frame visual input and instructions. The Flat R3M method, lacking language conditioning and skill sets can only succeed on tasks like drawer-close and reach through behavior cloning. Lang-conditioned Diffuser and modified LISA both outperform Flat R3M, suggesting the value of each corresponding module. Our SkillDiffuser, discretizing skills into a skill set, achieves a $6\\%$ higher performance than language-conditioned diffuser and a $9.5\\%$ higher than LISA, demonstrating the effectiveness of this combinational architecture.", + "bbox": [ + 75, + 325, + 467, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study on Reusability of Learned Skills", + "text_level": 1, + "bbox": [ + 76, + 544, + 467, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the reusability of our learned skills, we calculate the average number of different skills used for a single instruction and the total number of instructions using each skill of LOReL Sawyer Dataset in Table 5. (With max episode step being 20, we experiment with skill horizon 10.) We observe each single instruction uses 1.55 sub-skills on average and each skill is called multiple times than the number of instructions (75 with 5 eval episodes), verifying the transferability of learnt skills. As suggested in Tab.10 of [13], except a very small horizon will hurt the performance, learning sub-skills to get refined semantics helps perform different actions at different stages. Besides, we also visualize resulting images from applying discrete skills in Appendix E.1 to further validate skills' interpretability.", + "bbox": [ + 75, + 566, + 467, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Visualization Results of Learned Skill Set", + "text_level": 1, + "bbox": [ + 76, + 786, + 429, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the visualization of skill set on LOReL compositional tasks here in Fig. 4 and results on original LOReL dataset in Fig. 5 and Meta-World dataset in Fig. 6 in Appendix C. The visual analysis of our SkillDiffuser's learned skills on LOReL compositional tasks reveals that out of the 20-size skill set, our method learned 11 skills (e.g. pull han", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b108d10deaa10a7a73af4a59cf357403da887acbd1a1e9cd7da0e9574a8b7f67.jpg", + "table_caption": [ + "Table 4. Ablation study of language skill conditioning on Meta-World. Results of Flat R3M, Language-condition diffuser, LISA and our SkillDiffuser are shown here. All are averaged over 3 runs." + ], + "table_footnote": [], + "table_body": "
# of learnt skills# of inst# of successUse 1 skillUse 2 skillsAverage
1737514464801.55
Freq of 17 skills30, 8, 14, 10, 5, 19, 4, 7, 20, 2, 9, 20, 25, 6, 10, 3, 18
", + "bbox": [ + 504, + 89, + 908, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c68b2ccb7870580917e050779f5398c452eb66f7b0b96b5fca037d98b5d5b06a.jpg", + "image_caption": [ + "Table 5. Average number of different skills used for a single instruction and total number of instructions used for each skill.", + "Figure 4. Visualization of skill heat map on LOReL Sawyer compositional tasks. We display the word frequency associated with a skill set of size 20 in LOReL, normalized by column. The data's sparsity and distinct highlights indicate certain language tokens are uniquely linked to specific skills. There are eleven skills learned by our method. (zoom in for best view)" + ], + "image_footnote": [], + "bbox": [ + 503, + 170, + 888, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dle [skill 0], open counter [skill 14], etc.) notably distinguished by their unique word highlights. These bright spots across eleven columns (changing from only one column at initial which corresponds to default BC) in the heatmap underscore the model's ability to identify and isolate distinct skills from visual inputs, without an explicitly defined skill library. This indicates not only a significant interpretative advancement over previous diffusion-based planning but a successful abstraction of high-level skill representations.", + "bbox": [ + 496, + 429, + 890, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 571, + 617, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents SkillDiffuser, an integrated framework that enables robots to perform tasks from natural language instructions by enabling interpretable skill learning and conditional diffusion planning. It employs vector quantization to learn discrete and comprehensible skill representations directly from visual and linguistic demonstrations. Subsequently, these skills condition a diffusion model to generate state trajectories adhering to the learned skills. Through integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser can comprehend and execute abstract instructions for various manipulation tasks. Extensive experiments on manipulation benchmarks demonstrate state-of-the-art performance, highlighting its effectiveness for multi-step composition tasks and ability to automatically learn interpretable skills.", + "bbox": [ + 496, + 595, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 830, + 666, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16474", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. 2022. 1", + "[2] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 5, 6", + "[3] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 3", + "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 7", + "[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 3", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 3", + "[7] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 2023. 1, 3, 6", + "[8] Yan Duan, Marcin Andrychowicz, Bradly Stadie, OpenAI Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel, and Wojciech Zaremba. One-shot imitation learning. Advances in neural information processing systems, 30, 2017. 2", + "[9] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018.2", + "[10] William Feller. On the theory of stochastic processes, with particular reference to applications. In Selected Papers I, pages 769-798. Springer, 2015. 3", + "[11] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 2", + "[12] Divyansh Garg, Shuvam Chakraborty, Chris Cundy, Jiaming Song, and Stefano Ermon. Iq-learn: Inverse soft-q learning for imitation. Advances in Neural Information Processing Systems, 34:4028–4039, 2021. 2", + "[13] Divyansh Garg, Skanda Vaidyanath, Kuno Kim, Jiaming Song, and Stefano Ermon. Lisa: Learning interpretable skill abstractions from language. Advances in Neural Information Processing Systems, 35:21711-21724, 2022. 2, 3, 6, 7, 8, 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Haoran He, Chenjia Bai, Kang Xu, Zhuoran Yang, Weinan Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Diffusion model is an effective planner and data synthesizer for multitask reinforcement learning. Advances in neural information processing systems, 2023. 1, 2", + "[15] Bernd Heidergott, editor. *Taylor Series Expansions*, pages 179–263. Springer US, Boston, MA, 2007. 1", + "[16] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2", + "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 2", + "[19] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 1, 3", + "[20] Yiding Jiang, Shixiang Shane Gu, Kevin P Murphy, and Chelsea Finn. Language as an abstraction for hierarchical deep reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 7", + "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7", + "[22] Alexander C Li, Carlos Florensa, Ignasi Clavera, and Pieter Abbeel. Sub-policy adaptation for hierarchical reinforcement learning. arXiv preprint arXiv:1906.05862, 2019. 3", + "[23] Zhixuan Liang, Yao Mu, Mingyu Ding, Fei Ni, Masayoshi Tomizuka, and Ping Luo. Adaptdiffuser: Diffusion models as adaptive self-evolving planners. In International Conference on Machine Learning, 2023. 1, 3, 6", + "[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 5", + "[25] Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Causal transformer for estimating counterfactual outcomes. In International Conference on Machine Learning, pages 15293-15329. PMLR, 2022. 7", + "[26] Diganta Misra. Mish: A self regularized non-monotonic activation function. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020, 2020. 6", + "[27] Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical reinforcement learning. Advances in neural information processing systems, 31, 2018. 3", + "[28] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2", + "[29] Suraj Nair, Eric Mitchell, Kevin Chen, Silvio Savarese, Chelsea Finn, et al. Learning language-conditioned robot behavior from offline data and crowd-sourced annotation. In" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16475", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Robot Learning, pages 1303-1315. PMLR, 2022. 2, 6, 7, 4", + "[30] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In Conference on Robot Learning, pages 892-909. PMLR, 2023. 2, 4, 7, 8, 6", + "[31] Fei Ni, Jianye Hao, Yao Mu, Yifu Yuan, Yan Zheng, Bin Wang, and Zhixuan Liang. Metadiffuser: Diffusion model as conditional planner for offline meta-rl. In International Conference on Machine Learning, 2023. 3", + "[32] Scott Niekum, Sarah Osentoski, George Konidaris, Sachin Chitta, Bhaskara Marthi, and Andrew G Barto. Learning grounded finite-state representations from unstructured demonstrations. The International Journal of Robotics Research, 34(2):131-157, 2015. 2", + "[33] Dean A Pomerleau. Efficient training of artificial neural networks for autonomous navigation. Neural computation, 3 (1):88-97, 1991. 2", + "[34] Martin L Puterman. Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons, 1994. 3", + "[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6", + "[37] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 661-668. JMLR Workshop and Conference Proceedings, 2010. 2", + "[38] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 2", + "[39] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In Proceedings of Thirty-third Conference on Neural Information Processing Systems, 2019. 6", + "[40] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019. 2", + "[41] Avi Singh, Eric Jang, Alexander Irpan, Daniel Kappler, Murtaza Dalal, Sergey Levinev, Mohi Khansari, and Chelsea Finn. Scalable multi-task imitation learning with autonomous improvement. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2167-2173. IEEE, 2020. 2", + "[42] Simon Stepputtis, Joseph Campbell, Mariano Phielipp, Stefan Lee, Chitta Baral, and Henri Ben Amor. Language-conditioned imitation learning for robot manipulation tasks. Advances in Neural Information Processing Systems, 33: 13139-13150, 2020. 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Alan Stuart and J Keith Ord. Kendall's advanced theory of statistics. vol. 1: Distribution theory. Kendall's advanced theory of statistics. Vol. 1: Distribution theory, 1994. 2", + "[44] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4950-4957, 2018. 6", + "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5, 6", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[47] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 6", + "[48] Danfei Xu, Suraj Nair, Yuke Zhu, Julian Gao, Animesh Garg, Li Fei-Fei, and Silvio Savarese. Neural task programming: Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 3795–3802. IEEE, 2018. 2", + "[49] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In international conference on machine learning, pages 24631-24645. PMLR, 2022. 2, 8", + "[50] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 1", + "[51] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 2, 6", + "[52] Jesse Zhang, Haonan Yu, and Wei Xu. Hierarchical reinforcement learning by discovering intrinsic options. arXiv preprint arXiv:2101.06521, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 657 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16476", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_model.json b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..873403070b90a1878de1386ed652334d308edcca --- /dev/null +++ b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.865, + 0.175 + ], + "angle": 0, + "content": "SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.203, + 0.669, + 0.221 + ], + "angle": 0, + "content": "Zhixuan Liang\\(^{1,3}\\) Yao Mu\\(^{1,3}\\) Hengbo Ma\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.221, + 0.708, + 0.24 + ], + "angle": 0, + "content": "Masayoshi Tomizuka2 Mingyu Ding\\(^{2\\dagger}\\) Ping Luo\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.244, + 0.868, + 0.264 + ], + "angle": 0, + "content": "1The University of Hong Kong 2University of California, Berkeley 3Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.265, + 0.812, + 0.28 + ], + "angle": 0, + "content": "{zxliang, ymu, pluo}@cs.hku.hk {hengbo.ma, tomizuka, myding}@berkeley.edu" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.282, + 0.642, + 0.297 + ], + "angle": 0, + "content": "https://skilldiffuser.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.333, + 0.314, + 0.349 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.474, + 0.698 + ], + "angle": 0, + "content": "Diffusion models have demonstrated strong potential for robotic trajectory planning. However, generating coherent trajectories from high-level instructions remains challenging, especially for long-range composition tasks requiring multiple sequential skills. We propose SkillDiffuser, an end-to-end hierarchical planning framework integrating interpretable skill learning with conditional diffusion planning to address this problem. At the higher level, the skill abstraction module learns discrete, human-understandable skill representations from visual observations and language instructions. These learned skill embeddings are then used to condition the diffusion model to generate customized latent trajectories aligned with the skills. This allows generating diverse state trajectories that adhere to the learnable skills. By integrating skill learning with conditional trajectory generation, SkillDiffuser produces coherent behavior following abstract instructions across diverse tasks. Experiments on multi-task robotic manipulation benchmarks like Meta-World and LOReL demonstrate state-of-the-art performance and human-interpretable skill representations from SkillDiffuser. More visualization results and information could be found on our website." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.728, + 0.21, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Recent research [6, 7, 18, 19] has demonstrated diffusion models' superior generative capabilities compared to previous models that help enhance reinforcement learning across various dimensions, including the generation of action trajectories [2, 19], policy representation [5, 50], and data synthesis [14, 23]. However, their ability to generate coherent trajectories for intricate tasks still poses challenges in terms of performance and generalizability, as these tasks often re" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.335, + 0.605, + 0.345 + ], + "angle": 0, + "content": "Non-transferable" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.346, + 0.617, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.411, + 0.559, + 0.419 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.42, + 0.607, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.438, + 0.607, + 0.447 + ], + "angle": 0, + "content": "Denoise U-Net" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.448, + 0.607, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.462, + 0.56, + 0.472 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.514, + 0.617, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.544, + 0.597, + 0.555 + ], + "angle": 0, + "content": "Instruction Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.558, + 0.625, + 0.577 + ], + "angle": 0, + "content": "(a) Language-conditioned Diffuser" + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.335, + 0.825, + 0.345 + ], + "angle": 0, + "content": "Transferable and Interpretable" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.346, + 0.882, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.544, + 0.713, + 0.555 + ], + "angle": 0, + "content": "Visual Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.544, + 0.852, + 0.554 + ], + "angle": 0, + "content": "Instruction Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.562, + 0.808, + 0.572 + ], + "angle": 0, + "content": "(b) Our SkillDiffuser" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.694 + ], + "angle": 0, + "content": "Figure 1. Comparison of SkillDiffuser and previous language conditioned diffusers. SkillDiffuser utilizes high-level abstraction to translate visual observations and language instructions into human understandable skills with language grounding. It then enables the low-level diffusion model condition on these skills, not only improving the execution performance of multi-step composition tasks but greatly enhancing the generalization and adaptability of the framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.734 + ], + "angle": 0, + "content": "require the fulfillment of abstract instructions that consist of numerous coordination-intensive sequential steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Previous approaches [2, 14], such as Decision Diffuser, aim to tackle this challenge by decomposing complex tasks into simpler sub-skills, organized within a predefined skill library. These methods rely on conditioning the diffusion model with one-hot skill representations to generate trajectories for each of these sub-skills. Nevertheless, these techniques encounter difficulties in attempting to autonomously learn end-to-end from a wide range of datasets, which hinders their scalability and ability to achieve end-to-end learning [1]. In addition, without explicitly learning reusable skills, models cannot capture intricate inter-step dependen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16467" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.228 + ], + "angle": 0, + "content": "cies and constraints, yielding fragmented and illogical trajectories. Decomposing ambiguous instructions into learnable sub-goals and skills can better enable models to follow logical step sequences, respect task structures, and transfer common procedures (e.g. reusable skill abstraction and adaptable skill-based diffusion) between different tasks. This skill-centric paradigm paves the way for diffusion models that can interpret and execute elaborate, abstract instructions necessitating numerous sequential steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.471, + 0.533 + ], + "angle": 0, + "content": "Inspired by the above observations, we propose SkillDiffuser, a hierarchical planning framework unifying high-level skill learning with low-level conditional diffusion-based execution. As shown in Fig. 1, compared with previous language-conditioned diffusion policies, SkillDiffuser is able to interpret and execute complex instructions end-to-end with higher transferability. SkillDiffuser induces interpretable sub-latent goals by learning reusable skills tailored to the task instructions. The framework conditions a diffusion model on these learned skills to generate customized, coherent trajectories aligned with the overall objectives. By integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser achieves consistent, skill-oriented behavior without relying on a predefined skill library. Moreover, SkillDiffuser is designed to operate solely on visual observations, eliminating the need for robot proprioception (i.e. fully observed states). This end-to-end methodology, featuring learnable skills, enables SkillDiffuser to execute abstract instructions across a variety of tasks efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.808 + ], + "angle": 0, + "content": "SkillDiffuser works as follows. 1) A skill predictor with vector quantization [45] is used for high-level skill learning to distill tasks into discrete and interpretable skills. Rather than forecasting skill durations, we adopt a fixed prediction horizon - predicting skills at regular intervals. This horizon-based discretization process seamlessly integrates visual and linguistic inputs into a cohesive skill set guiding the low-level diffusion model. 2) For skill-based trajectory generation, we utilize a classifier-free diffusion model as policy, with skills directly embedded as guidance. This setting allows for generating multi-modal state trajectories aligned with skill specifications while avoiding overfitting to a closed dataset. 3) To enable action inference from predicted states, we train an inverse dynamics network to decode motions between two consecutive frames generated. By separating state prediction from motion decoding, SkillDiffuser yields a fully adaptive framework for directing diverse embodiments via transferable state-space plans." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We evaluate the model's performance in both skill learning and multi-task planning on the LOReL [29] Sawyer and Meta-World [51] datasets, with crucial experimental settings by considering robotic agents in real-world scenarios must operate with limited state information, chiefly visual observations. Furthermore, we present success rates on un" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "seen compositional tasks, the reusability and visualizations of learned skills to illustrate the model does have the ability to abstract high-level skills that are not only effective but also understandable to humans, bringing us closer to intelligent agents that acquire skills in a direct manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Our contributions are three-fold: 1) We propose an end-to-end hierarchical planning framework via skill learning for sub-goal abstraction; 2) We adopt a classifier-free diffusion model conditioned on learned skills to generate skill-oriented transferable state trajectories; 3) We demonstrate state-of-the-art performance on complex benchmarks and provide interpretable visualizations of human-understandable skill representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.304, + 0.65, + 0.319 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.328, + 0.878, + 0.344 + ], + "angle": 0, + "content": "2.1. Imitation Learning and Multi-task Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.518 + ], + "angle": 0, + "content": "Imitation learning (IL) has evolved from foundational behavioral cloning to sophisticated multi-task learning frameworks. With traditional approaches relying on supervised learning from expert demonstrations [33, 37, 38], recent advancements have shifted towards learning the reward [16] or Q-function [12] from expert data, enhancing the ability to mimic complex behaviors. A new challenge lies in multi-task IL [41], where imitators are trained across varied tasks, aiming for generalization to new scenarios with task specifications ranging from vector states [28] to vision and language descriptions [8, 11, 13, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Multi-task learning approaches often leverage shared representations to learn a spectrum of tasks simultaneously, enhancing the flexibility and efficiency of the learning process. The Meta-World benchmark [51] assesses multi-task and meta reinforcement learning, highlighting the need for algorithms capable of rapid adaptation. Building on this, the Prompting Decision Transformer [49] showcases few-shot policy generalization using task-specific prompts. And diffusion policy has also been explored in multi-task settings [14], which shows proficiency in generating diverse behaviors across tasks. However, unlike methods that leverage state inputs [14, 49] or access robot proprioception [30], SkillDiffuser uses raw visual inputs only." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.862, + 0.742 + ], + "angle": 0, + "content": "2.2. Skill Discovery and Hierarchical Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Skill learning, the process by which robots acquire new abilities or refine existing ones, is gaining increasing attention due to its pivotal role in enabling autonomous systems to adapt to new tasks, improve the performance over time, and interact naturally with humans and complex environments. Traditionally, this domain was influenced by handcrafted features and expert demonstrations [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With the development of deep learning, Eysenbach et al. [9] and A. Sharma et al. [40] investigated skill discovery in learning methods, achieving policies conditioned" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "on learned latent variables and maintaining consistent skill codes throughout trajectories. In the domain of skill learning through language instructions, LISA [13] stands out by sampling multiple skills per trajectory, integrating language conditioning in a unique manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.274 + ], + "angle": 0, + "content": "Our SkillDiffuser follows this way to extract sub-skills of each instruction at the higher-level. But differently, SkillDiffuser employs an adaptable diffusion policy at the lower-level to condition on different sub-skills to generate different actions, which formulates a creative hierarchical planning framework also advancing research on hierarchical techniques of reinforcement learning [22, 27, 52]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.284, + 0.35, + 0.3 + ], + "angle": 0, + "content": "2.3. Planning with Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.519 + ], + "angle": 0, + "content": "Diffusion models [18] make great breakthroughs in image synthesis [18] recent years and has shown promising results in various generative applications. A seminal work that performs planning with diffusion directly is Diffuser [19], which laid the groundwork for using diffusion models in behavior synthesis. Then, a branch of this kind planning methods achieved state-of-the-art performance in a variety of decision tasks [3, 5, 7, 23, 31]. Among them, the work done by Chi et al. in [5] introduces the concept of learning the gradient of the action-distribution score function and iteratively optimizing it, demonstrating its significant potential in visuomotor policy learning. These works have further extended this direction and strengthened the versatility and generalization of diffusion-based planners." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Our approach is inspired by classifier-free diffusion guidance [17], which offers a significant advantage over classifier-guided models [6]. By adopting the classifier-free approach, we can circumvent the challenges associated with training a reward model and \\(Q\\)-function, which are particularly cumbersome in many real-world planning scenarios of which the complexity is very high. Recent studies have also extended this direction, which use conditional diffusion models to generate customized trajectories. Decision Diffuser [2] is an example which is designed to generate trajectories for a predefined skill library. However, unlike our method, it can't autonomously learn skill abstractions in an end-to-end fashion, which makes it difficult to scale to more tasks. This highlights the necessity for diffusion models with dynamic, learnable skill abstractions, facilitating complex instruction execution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.777, + 0.204, + 0.794 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.385, + 0.817 + ], + "angle": 0, + "content": "3.1. Planning with Diffusion over States" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "As introduced in previous works [2, 19], diffusion model is a promising tool to address the problem of planning in reinforcement learning which is cast as a Markov Decision Process (MDP) [34]. Within the MDP framework \\(\\mathcal{M} = (\\mathcal{S},\\mathcal{A},\\mathcal{T},\\mathcal{R},\\gamma)\\), the planning policy aims to iden" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "tify an optimal action sequence \\( \\pmb{a}_{0:T}^{*} \\) that maximizes the expected cumulative rewards over a finite time horizon \\( T \\), governed by the state transition dynamics \\( \\mathcal{T} \\) and reward function \\( \\mathcal{R} \\). \\( S \\) is the state space and \\( \\mathcal{A} \\) is the action space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.288 + ], + "angle": 0, + "content": "By treating the state trajectory as sequence data \\(\\pmb{\\tau}\\), with sequence modeling, diffusion probabilistic models conceptualize planning as an iterative denoising process. The model progressively refines trajectories by reversing a forward diffusion process that is modeled as a Gaussian process, whereby noise is incrementally added to the data, denoted as \\(p_{\\theta}(\\pmb{\\tau}^{i-1} \\mid \\pmb{\\tau}^i)\\). Training involves minimizing the ELBO of the data's negative log-likelihood, similar to variational Bayesian inference, with the optimization objective:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.291, + 0.892, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\arg \\min _ {\\theta} - \\mathbb {E} _ {\\boldsymbol {\\tau} ^ {0}} \\left[ \\log p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {0}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.344 + ], + "angle": 0, + "content": "where \\(p(\\boldsymbol{\\tau}^N)\\) is a standard normal distribution and \\(\\boldsymbol{\\tau}^0\\) denotes noiseless sequence data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.39 + ], + "angle": 0, + "content": "For practical implementation, a simplified surrogate loss function is proposed in [18], focusing on predicting the Gaussian mean of the reverse diffusion step:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.396, + 0.892, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e n o i s e}} (\\theta) = \\mathbb {E} _ {i, \\boldsymbol {\\tau} ^ {0} \\sim q, \\epsilon \\sim \\mathcal {N}} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {\\tau} ^ {i}, i) \\| ^ {2} ]. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.797, + 0.437 + ], + "angle": 0, + "content": "3.2. Classifier-free Diffusion Guidance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.566 + ], + "angle": 0, + "content": "On the basis of unconditioned diffusion-based method, in the realm of offline reinforcement learning, a critical endeavor is to generate trajectories with the highest reward-to-go. With the flourishing development of conditional diffusion models [6], classifier-guided approaches embark on this by infusing specific trajectory information (encoded in \\( y(\\tau) \\)), such as the return \\( \\mathcal{J}(\\tau^0) \\) or designated constraints, into the diffusion process:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.569, + 0.892, + 0.588 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {\\tau} ^ {i} \\mid \\boldsymbol {\\tau} ^ {i - 1}\\right), \\quad p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i - 1} \\mid \\boldsymbol {\\tau} ^ {i}, \\boldsymbol {y} (\\boldsymbol {\\tau})\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.593, + 0.794, + 0.608 + ], + "angle": 0, + "content": "With assumptions specified in [10], we have" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.614, + 0.892, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\tau} ^ {i - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} + \\alpha \\Sigma \\nabla_ {\\boldsymbol {\\tau}} \\log p (\\boldsymbol {y} (\\boldsymbol {\\tau} ^ {i})), \\Sigma\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.679 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter that adjusts the guidance strength, \\(\\Sigma\\) is the specified covariance of the noise and \\(\\mu_{\\theta}\\) is the learned mean value of noise in unconditional diffusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.756 + ], + "angle": 0, + "content": "However, the classifier-guided diffusion model requires an accurate estimation of guidance gradient based on the trajectory classifier \\(\\pmb{y}(\\pmb{\\tau})\\), which may not be feasible and need to introduce a separate dynamic programming procedure to estimate a \\(Q\\)-function in the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.846 + ], + "angle": 0, + "content": "Thus, classifier-free guidance offers an alternative, which augments the trajectory generation process with a guidance signal that amplifies the features of high-reward or optimal characteristics, i.e. \\( y(\\tau) \\), that are implicitly present in the data. Mathematically, the noise to add during the reverse denoising process is:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.85, + 0.892, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} = \\epsilon_ {\\theta} (\\boldsymbol {\\tau} ^ {i}, \\varnothing , i) + \\omega \\left(\\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, \\boldsymbol {y}, i\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, \\varnothing , i\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\omega\\) is the guidance scale, and \\(\\varnothing\\) represents the absence of guidance. Setting \\(\\omega = 0\\) removes the classifier-free" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "guidance towards an unconditional diffusion model, while a large value of \\(\\omega\\) strengthens the influence of the conditional information during trajectory generation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.138, + 0.462, + 0.151 + ], + "angle": 0, + "content": "Also, the loss function to minimize can be rewritten as," + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.152, + 0.474, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} (\\theta) = \\mathbb {E} _ {i, \\boldsymbol {\\tau}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, (1 - \\beta) \\boldsymbol {y} \\left(\\boldsymbol {\\tau} ^ {i}\\right) + \\beta \\emptyset , i\\right) \\right\\| ^ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.191, + 0.47, + 0.235 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyperparameter that controls the probability of dropping specific condition \\(\\pmb{y}(\\pmb{\\tau})\\), enhancing sample diversity while maintaining context relevance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.236, + 0.469, + 0.296 + ], + "angle": 0, + "content": "After training the noise prediction model \\(\\epsilon_{\\theta}\\) with the above \\(\\mathcal{L}(\\theta)\\), the trajectory is sampled starting from Gaussian noise and progressively denoised with modified \\(\\hat{\\epsilon}\\) through Eq. 5, employing re-parameterization technique." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.469, + 0.387 + ], + "angle": 0, + "content": "In summary, through classifier-free guidance, we can modulate the trajectory sampling process to require the generated trajectories more aligned with the desired features represented by \\( \\pmb{y}(\\pmb{\\tau}) \\). This process iteratively applies the conditioned noise model to refine the target trajectories that contain future states satisfying the constraints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.212, + 0.416 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.188, + 0.438 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.598 + ], + "angle": 0, + "content": "Building upon the motivations discussed above, we present SkillDiffuser, an advanced methodological framework for robust multi-task learning across various robots. This dynamic approach leverages the cooperation of skill learning at the higher level and a conditional diffusion model at the lower level. An overall framework is illustrated in Fig. 2. Notably, we leverage language-grounded representations to abstract skills, thereby rendering the execution of tasks through our diffusion policy both interpretable and comprehensible to humans." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.446, + 0.623 + ], + "angle": 0, + "content": "4.2. High Level Interpretable Skill Abstractions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.87 + ], + "angle": 0, + "content": "In our SkillDiffuser framework, the high level interpretable skill abstraction module plays a crucial role in understanding and executing complex tasks. However, given the multitask environments we consider, each task with only a single language instruction may be broken down into a sequence of sub-tasks or skills, which are not explicitly delineated within the instruction itself. Furthermore, suppose the offline training dataset is denoted as \\(\\mathcal{D}\\), it consists of trajectories derived from a sub-optimal policy for various tasks. A trajectory \\(\\boldsymbol{\\tau} = (l, \\{\\boldsymbol{i}_t, \\boldsymbol{a}_t\\}_{t=1}^T)\\) includes language description \\(l\\) and a sequence of image observations and actions \\((\\boldsymbol{i}_t, \\boldsymbol{a}_t)\\) over time steps \\(T\\), with no reward labels attached. But the trajectories do not indicate the boundaries between sub-tasks as well, which thus requires the proposed methods capable of segmenting and interpreting the tasks into sub-goals in an unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To address this challenge, we build a skill abstraction component upon a horizon-based skill predictor adapted" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.091, + 0.858, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Figure 2. Overall framework of SkillDiffuser. It's a hierarchical planning model that leverages the cooperation of interpretable skill abstractions at the higher level and a skill conditioned diffusion model at the lower level for task execution in a multi-task learning environment. The high-level skill abstraction is achieved through a skill predictor and a vector quantization operation, generating sub-goals (skill set) that the diffusion model employs to determine the appropriate future states. Future states are converted to actions using an inverse dynamics model. This unique fusion enables a consistent underlying planner across different tasks, with the variation only in the inverse dynamics model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.629 + ], + "angle": 0, + "content": "from GPT-2 [35]. It is designed to parse and decompose tasks by fusing visual input and natural language instructions, alongside a Vector Quantization (VQ) sub-module that discretizes the learned skills into a skill set. The specifics of this component are illustrated below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Firstly, as we utilize only images as robot state information, we transform the images into latent space features with a fixed image encoder (e.g. R3M [30]). For convenience, we denote the image encoder as \\(\\Phi_{im}:\\mathcal{I}\\to \\mathcal{R}^I\\), where \\(\\mathcal{I}\\) represents the space of input images and \\(\\mathcal{R}^I\\), the resultant feature space, serves to condense the visual information into a form that is conducive for high-level semantics. Simultaneously, we use a language encoder to pre-process the natural language instructions, which is formalized as \\(\\Phi_{lang}:L\\rightarrow \\mathcal{R}^{L}\\), with \\(L\\) being the space of language instructions and \\(\\mathcal{R}^L\\) the language feature space. The outputs of both encoders are then fed into the skill predictor, which operates to integrate these two modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The skill predicting process is as follows: An image \\( \\boldsymbol{i}_t \\in \\mathcal{I} \\) at time step \\( t \\) is encoded into a visual embedding \\( s_t = \\Phi_{im}(\\boldsymbol{i}_t) \\). This embedding \\( s_t \\) is then input to the skill predictor, along with the language instruction \\( l \\in L \\) through the language encoder's output \\( l_t = \\Phi_{lang}(l) \\). The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.088, + 0.833, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.396, + 0.895, + 0.438 + ], + "angle": 0, + "content": "Figure 3. SkillDiffuser's low level skill-conditioned diffusion planning model. Notably, while the schematic here employs images to represent visual features for illustrative purposes, in actual implementation, both the input to and the sampling output of the diffusion model are state embeddings. The current observation is also the feature embedding of current visual observation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.488 + ], + "angle": 0, + "content": "skill predictor, represented as \\( f: \\mathcal{R}^I \\times \\mathcal{R}^L \\to \\mathcal{C} \\), generates a skill code \\( \\tilde{z} \\) by \\( \\tilde{z} = f(\\boldsymbol{s}_t, \\boldsymbol{l}_t) \\) that encapsulates task's requirements interpreted from the visual and language inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.609 + ], + "angle": 0, + "content": "After that, Vector Quantization [45] operation \\(\\mathbf{q}(\\cdot)\\) is taken to transform these predicted skills into a low-dimensional discrete set \\(\\mathcal{C}\\). The discrete skill set contains \\(K\\) skill embeddings \\(\\{z^1, z^2, \\ldots, z^K\\}\\) which represent different potential skills. VQ operation is achieved by mapping the latent \\(\\tilde{z}\\) to its closest entry of the skill set with skill vectors updated to be the moving average of the embeddings \\(z\\) closest to them, same as [45]. This process is as follows," + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.612, + 0.469, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} = f \\left(\\Phi_ {i m} (\\boldsymbol {i} _ {t}), \\Phi_ {l a n g} (l)\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.633, + 0.469, + 0.658 + ], + "angle": 0, + "content": "\\[\nz = \\mathbf {q} (\\tilde {z}) = \\underset {z ^ {k} \\in \\mathcal {C}} {\\arg \\min } \\| \\tilde {z} - z ^ {k} \\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.78 + ], + "angle": 0, + "content": "VQ enforces each learnt skill \\( z \\) to lie in \\( \\mathcal{C} \\), which is equal to learning \\( K \\) prototypes for the language embeddings utilizing \\( k \\)-means [24] algorithm. This acts as a bottleneck, restricting the flow of language information and aiding in the learning of discrete skill codes. The back-propagation through the non-differentiable quantization is achieved by a straight-through gradient estimator, which simply copies the gradients and enables model to be trained end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In our approach, we apply a consistent skill code \\( z \\) across a defined number of time steps, termed the horizon. This consistent application across multiple horizons adeptly addresses the challenge of varying sub-task durations without altering the horizon itself. Consequently, this strategy not only preserves the flexibility required for diverse task execution but also maintains the model's architectural stability by avoiding horizon-induced structural changes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Importantly, the discrete nature of the learned skill codes enhances the interpretability and controllability of the system's behavior, as each skill code is associated with some human-understandable language phrases. An example is depicted in Fig 6. Through this method, SkillDiffuser can not only learn to perform tasks based on language instructions but also achieve them in a human-interpretable way, allowing for a deeper understanding and control of the decision-making process of specific embodied agents." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.589, + 0.892, + 0.607 + ], + "angle": 0, + "content": "4.3. Low Level Skill-conditioned Diffusion Planning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.764 + ], + "angle": 0, + "content": "As highlighted in Section 1, while existing approaches like Decision Diffuser [2] have introduced conditional diffusion models constrained by skills, their capability is limited to generating trajectories that meet only predefined skills. Consequently, these models fall short of realizing a diffusion model capable of conditioning on an arbitrary learned skill. To overcome these constraints and enable diffusion models to plan over a learned continuous spectrum of skills, we propose an approach that leverages the classifier-free guided diffusion model with the skills embedded." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SkillDiffuser begins by employing a diffusion model to operate over the continuous skill space \\(\\mathcal{Z}\\) learned during high-level skill abstraction. We employ a U-Net to serve as the noise prediction model \\(\\epsilon_{\\theta}(\\cdot)\\) and guide it by in-context conditioning. More specifically, we firstly utilize a skill MLP (similar to point-wise feed-forward network [46]), denoted as \\(\\Lambda\\), to align skill features with denoising model. After that, we fuse the skill embeddings \\(\\Lambda(z)\\) into the state features throughout the residual blocks of U-Net. Details" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "are depicted in Fig. 3. In this way, we make the diffusion model contextually modulate the influence of skill embeddings at each step. And following Eq. 5, we can synthesize future states attending to these given skills. This is a significant shift from previous static conditioning framework to a more dynamic and adaptive trajectory generation process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.318 + ], + "angle": 0, + "content": "Moreover, following previous works [2, 7, 23], we adopt a state-only diffusion model, which eschews the direct generation of actions in diffusion model. And instead, we utilize another MLP, denoted as \\(\\Psi(\\cdot)\\), to perform inverse dynamics after state generation to infer feasible actions that can achieve transitions between two continuous states. Additionally, we integrate observations from the current frame to provide more detailed information for motion prediction and achieve closed-loop control. Mathematically, it is:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.323, + 0.468, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {t} = \\Psi ([ \\tilde {\\boldsymbol {s}} _ {t}, \\tilde {\\boldsymbol {s}} _ {t + 1} ], \\boldsymbol {i} _ {t}) \\quad \\text {f o r} \\quad t = 0, \\dots , T - 1, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.468, + 0.388 + ], + "angle": 0, + "content": "where \\(\\tilde{\\boldsymbol{s}}_t\\) and \\(\\tilde{\\boldsymbol{s}}_{t + 1}\\) are consecutive observation embeddings within \\(\\tilde{\\tau}\\) generated by diffusion model, \\(i_t\\) is the current observation and \\(a_{t}\\) is the inferred action." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.468, + 0.555 + ], + "angle": 0, + "content": "As the resulting action sequence \\(\\{a_0, a_1, \\ldots, a_{T-1}\\}\\) derived from the generated states, it encapsulates the skills to execute the tasks, which allows for remarkable adaptability across multiple tasks. And when faced with a new task, we are only required to change the inverse dynamics model \\(\\Psi(\\cdot)\\) specific to the new task's kinematics, with the architecture and parameters of diffusion model unchanged. Such modularity ensures the generative capabilities of SkillDiffuser are not task-specific but can be leveraged across a diverse range of tasks with varying dynamics. A schematic diagram of the low level module is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.312, + 0.58 + ], + "angle": 0, + "content": "4.4. Training the SkillDiffuser" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.468, + 0.662 + ], + "angle": 0, + "content": "We engineered a threefold loss function for SkillDiffuser. Firstly, for the inverse dynamics model which is task-specific, we employ a behavior cloning loss [44] to train our inverse MLP emulating expert actions from observed state transitions. This loss, denoted as \\(\\mathcal{L}_{inv}\\), is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.666, + 0.468, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n v} = \\mathbb {E} _ {\\boldsymbol {\\tau} \\sim \\mathcal {D}} \\left[ \\left\\| \\boldsymbol {a} - \\Psi (\\tilde {\\boldsymbol {s}}, \\tilde {\\boldsymbol {s}} ^ {\\prime}, \\boldsymbol {i}) \\right\\| _ {2} ^ {2} \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.695, + 0.342, + 0.709 + ], + "angle": 0, + "content": "where the notations are similar to Eq. 9." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.468, + 0.829 + ], + "angle": 0, + "content": "Correspondingly, the other parts including both skill abstraction and low level execution are task-agnostic. For the high-level skill abstraction module, we apply a vector quantization (VQ) loss to refine the skill predictor. This VQ loss, \\(\\mathcal{L}_{\\mathrm{VQ}}\\), ensures the embeddings produced by the skill predictor closely match the skill set vectors, thereby improving the interpretability and consistency of the skill representations. Inspired by VQ-VAE [45], we formulate it as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.834, + 0.468, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V Q}} = \\mathbb {E} _ {\\boldsymbol {\\tau}} \\left[ \\| \\mathbf {q} (\\tilde {z}) - \\tilde {z} \\| _ {2} ^ {2} \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.231, + 0.87 + ], + "angle": 0, + "content": "where \\(\\tilde{z}\\) follows Eq. 7." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Lastly, for the low-level state-only skill-conditioned diffusion execution, we incorporate a diffusion loss \\(\\mathcal{L}_{diff}\\) as" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "per Eq. 6, ensuring our model's state predictions are in line with both the skill guidance and temporal dynamics observed in expert demonstrations. Here, we take \\( y(\\tau) = \\Lambda(z) \\) with \\( z \\) derived from Eq. 7 and Eq. 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.273 + ], + "angle": 0, + "content": "To be noted, we train our SkillDiffuser with two optimizers, one for inverse dynamics model with \\(\\mathcal{L}_{inv}\\) and the other for overall parameters of high-level skill abstraction and low-level planning with \\(\\mathcal{L}_{\\mathrm{VQ}} + \\lambda \\mathcal{L}_{diff}\\), where \\(\\lambda\\) is a loss weight. This carefully constructed loss architecture enables SkillDiffuser to excel in a multi-task environment, generalizing across tasks by simply substituting the inverse dynamics model \\(\\Psi(\\cdot)\\) specific to each new task's requirements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.411 + ], + "angle": 0, + "content": "Additionally, we utilize a pre-trained distilBERT [39] as our language encoder, adopting the configuration consistent with LOReL [29] and LISA [13], while freezing its parameters to guarantee stability in language understanding. And we employ diverse settings to serve as the visual encoder to ensure fair comparison in different datasets. We elaborate the details in corresponding parts of Sec. 5. More training details are shown in Appendix G and we also provide some pseudo-code of our algorithm in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.632, + 0.443 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.527 + ], + "angle": 0, + "content": "We first present a comprehensive evaluation on the LOReL Sawyer dataset, and then perform the ablation study compared on Meta-World benchmark to illustrate the efficiency of our method. Finally, we visualize the learned skills of our method both on LOReL and Meta-World MT10." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.602, + 0.555 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.761 + ], + "angle": 0, + "content": "LOReL Sawyer Dataset [29] which is abbreviated from Language-conditioned Offline Reward learning, is composed of pseudo-expert trajectories or play data gathered from an arbitrary reinforcement learning policy, annotated with post-hoc crowd-sourced language directives. The LOReL Sawyer dataset encompasses 50k trajectories, each with 20 steps, in a simulated Sawyer robot environment. We assess our approach using the same six tasks as the original paper [29] with paraphrased instructions under five different situations (i.e. seen, unseen verb, unseen none, unseen verb + noun and human provided). This comes to a total of 77 instructions for all 6 tasks combined. More details about the dataset can be found in Appendix D.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Meta-World Dataset [51]. It is a comprehensive benchmark designed for evaluating multi-task and meta reinforcement learning algorithms. It introduces a comprehensive suite of 50 distinct robotic manipulation tasks, all located within a unified tabletop environment controlled by a simulated Sawyer arm. The Multi-Task 10 (MT10) within MetaWorld is a subset comprising ten carefully selected tasks, balanced in terms of diversity and complexity. Details of the ten tasks can be found in Appendix D.3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16472" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.835, + 0.221 + ], + "angle": 0, + "content": "
Task InstructionRandomLCBC [42]LCRL [20]Lang DT [13]LISA [13]SkillDiffuser
close drawer52%50%58%10%100%95 ± 3.2%
open drawer14%0%8%60%20%55 ± 13.3%
turn faucet left24%12%13%0%0%55 ± 9.3%
turn faucet right15%31%0%0%30%25 ± 4.4%
move black mug right12%73%0%20%60%18 ± 3.9%
move white mug down5%6%0%0%30%10 ± 1.7%
Average over tasks20%29%13%15%40%43 ± 1.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.226, + 0.895, + 0.284 + ], + "angle": 0, + "content": "Table 1. Task-wise success rates on LOReL Sawyer Dataset. We show our success rates compared to random policy, language-conditioned imitation learning (LCBC), language-conditioned Q-learning (LCRL), a flat non-hierarchical Decision Transformer (Lang-DT), and LISA. The results on each dataset are calculated over 3 seeds. SkillDiffuser outperforms all other methods in terms of average performance over all tasks. Best methods and those within \\(6\\%\\) of the best are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.407 + ], + "angle": 0, + "content": "
Rephrasal TypeLang DTLISA [13]SkillDiffuser
seen154043.65 ±4.7
unseen noun13.3333.3336.01 ±6.3
unseen verb28.333036.70 ±9.5
unseen noun+verb6.72042.02 ±3.8
human provided26.9827.3540.16 ±2.1
Average18.0730.1439.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.408, + 0.472, + 0.451 + ], + "angle": 0, + "content": "Table 2. Rephrasal-wise success rates (in %) on LOReL Sawyer. Results of Lang DT, LISA and our SkillDiffuser are shown here. The standard error is calculated over 3 random seeds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.469, + 0.478 + ], + "angle": 0, + "content": "5.2. Evaluation Results on LOReL Sawyer Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Baselines. We employ random policy, language conditioned imitation learning (LCBC) [42], language conditioned Q-learning (LCRL) [20], Lang-DT (also known as Flat Baseline in [13]) and state-of-the-art skill-learning method LISA [13] as our baselines. We follow the same settings as LOReL [29] for the first three algorithms, and follow LISA [13] for the last two. The random policy serves as a baseline. And LCBC mimics offline dataset behaviors based on instructions, aligning with previous works focusing on imitation learning to achieve language conditioned behavior. In contrast, LCRL employs reinforcement learning, labeling each episode's final state with language-instructed rewards. Lang-DT plays as a non-hierarchical benchmark with a Causal Transformer [25], while LISA incorporates a dual-transformer structure, with one for skill prediction and the other for action planning. We do not compare with LOReL planner [29] as it uses MPC on a learned reward function relying on human annotations, while LISA and ours learn with trajectory data only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Results. To ensure fair comparison, our method, SkillDiffuser, is designed to maintain a parameter count similar to baseline models. It employs the same visual and language encoder architecture as used in LISA [13] with meticulously matching of embedding dimensions and the number of heads across the layers of SkillDiffuser." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The results are present in Tables 1 and 2, showing taskwise and rephrasal-wise success rates for LOReL, averaged" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.325 + ], + "angle": 0, + "content": "
MethodLang DTLOReL [29]LISA [13]SkillDiffuser
Success Rate13.33 ±1.318.18 ±1.820.89 ±0.625.21 ±2.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.326, + 0.892, + 0.341 + ], + "angle": 0, + "content": "Table 3. Performance on LOReL multi-step composition tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.892, + 0.499 + ], + "angle": 0, + "content": "over 10 runs with a 20-step time horizon. SkillDiffuser, our approach, achieves the highest average performance in six different tasks, indicating its superior cross-task adaptability particularly when compared to similar approaches which yet are based on Decision Transformer [4], such as LISA [13]. Moreover, SkillDiffuser consistently excels in all rephrased types for LOReL test tasks, outperforming LISA by \\(9.6\\%\\) on average. This demonstrates the model's robustness in handling varied skill representations, marking a notable advancement in skill-conditioned diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.507, + 0.887, + 0.524 + ], + "angle": 0, + "content": "5.3. Performance on LOReL Compositional Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Settings. We conduct experiments following the same settings of unseen composition tasks of LISA [13] with 12 composition instructions in Tab. 3. Detailed instructions are listed in Appendix D.2 with such an example that \"open drawer and move black mug right\". We extend the max number of episode steps from customary 20 to 40, as LISA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Results. We observe SkillDiffuser achieves 2x the performance of non-hierarchical baseline (i.e. w/o skill abstraction) and also improves about \\(25\\%\\) over LISA, highlighting its effectiveness. MPC-based LOReL planners are unable to perform as well in open scenarios like composition tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.708, + 0.838, + 0.724 + ], + "angle": 0, + "content": "5.4. Ablation Study on Meta-World Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.778 + ], + "angle": 0, + "content": "Settings. We conduct experiments on Meta-World MT10 benchmark with finely annotated instructions. We also use visual observations only. Details are in Appendix D.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Baselines. We evaluate our method against three baselines, all modified from existing models. The first, Flat R3M, is adapted from R3M [30] paper's planner. As the original one utilizes the first four terms of robot proprioception, we eliminate them and make the planner focus exclusively on visual observations. The second baseline is a variant of our SkillDiffuser, lacking the high-level skill abstraction module but retaining the low-level conditional diffusion-based planner," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16473" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.169 + ], + "angle": 0, + "content": "
MethodLangSkill SetDiffuserPerformance
Flat R3M [30]XXX13.3%
LISA [13]✓✓X13.8%
Lang Diffuser✓X✓16.7%
SkillDiffuser✓✓✓23.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.172, + 0.468, + 0.214 + ], + "angle": 0, + "content": "Table 4. Ablation study of language skill conditioning on Meta-World. Results of Flat R3M, Language-condition diffuser, LISA and our SkillDiffuser are shown here. All are averaged over 3 runs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.468, + 0.324 + ], + "angle": 0, + "content": "to assess the impact of skill abstraction. This version integrates a two-layer MLP to predict options from visual and language inputs, functioning as a language-conditioned diffusion planner. The last baseline is our re-implemented LISA [13] for Meta-World Dataset to validate the efficiency of diffusion model. To ensure fairness, we use R3M as the visual encoder for all of these methods on Meta-World." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.468, + 0.536 + ], + "angle": 0, + "content": "Ablation on Language Skill Conditioning. Table 4 indicates our Meta-World MT10 task is quite different from and much more challenging than previous tasks that use states as observations [49] or take into robot proprioception [30]. We only utilize single-frame visual input and instructions. The Flat R3M method, lacking language conditioning and skill sets can only succeed on tasks like drawer-close and reach through behavior cloning. Lang-conditioned Diffuser and modified LISA both outperform Flat R3M, suggesting the value of each corresponding module. Our SkillDiffuser, discretizing skills into a skill set, achieves a \\(6\\%\\) higher performance than language-conditioned diffuser and a \\(9.5\\%\\) higher than LISA, demonstrating the effectiveness of this combinational architecture." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.468, + 0.56 + ], + "angle": 0, + "content": "5.5. Ablation Study on Reusability of Learned Skills" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.468, + 0.78 + ], + "angle": 0, + "content": "To evaluate the reusability of our learned skills, we calculate the average number of different skills used for a single instruction and the total number of instructions using each skill of LOReL Sawyer Dataset in Table 5. (With max episode step being 20, we experiment with skill horizon 10.) We observe each single instruction uses 1.55 sub-skills on average and each skill is called multiple times than the number of instructions (75 with 5 eval episodes), verifying the transferability of learnt skills. As suggested in Tab.10 of [13], except a very small horizon will hurt the performance, learning sub-skills to get refined semantics helps perform different actions at different stages. Besides, we also visualize resulting images from applying discrete skills in Appendix E.1 to further validate skills' interpretability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.43, + 0.802 + ], + "angle": 0, + "content": "5.6. Visualization Results of Learned Skill Set" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We show the visualization of skill set on LOReL compositional tasks here in Fig. 4 and results on original LOReL dataset in Fig. 5 and Meta-World dataset in Fig. 6 in Appendix C. The visual analysis of our SkillDiffuser's learned skills on LOReL compositional tasks reveals that out of the 20-size skill set, our method learned 11 skills (e.g. pull han" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.91, + 0.138 + ], + "angle": 0, + "content": "
# of learnt skills# of inst# of successUse 1 skillUse 2 skillsAverage
1737514464801.55
Freq of 17 skills30, 8, 14, 10, 5, 19, 4, 7, 20, 2, 9, 20, 25, 6, 10, 3, 18
" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.141, + 0.892, + 0.167 + ], + "angle": 0, + "content": "Table 5. Average number of different skills used for a single instruction and total number of instructions used for each skill." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.171, + 0.89, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Figure 4. Visualization of skill heat map on LOReL Sawyer compositional tasks. We display the word frequency associated with a skill set of size 20 in LOReL, normalized by column. The data's sparsity and distinct highlights indicate certain language tokens are uniquely linked to specific skills. There are eleven skills learned by our method. (zoom in for best view)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.566 + ], + "angle": 0, + "content": "dle [skill 0], open counter [skill 14], etc.) notably distinguished by their unique word highlights. These bright spots across eleven columns (changing from only one column at initial which corresponds to default BC) in the heatmap underscore the model's ability to identify and isolate distinct skills from visual inputs, without an explicitly defined skill library. This indicates not only a significant interpretative advancement over previous diffusion-based planning but a successful abstraction of high-level skill representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.619, + 0.587 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.824 + ], + "angle": 0, + "content": "This paper presents SkillDiffuser, an integrated framework that enables robots to perform tasks from natural language instructions by enabling interpretable skill learning and conditional diffusion planning. It employs vector quantization to learn discrete and comprehensible skill representations directly from visual and linguistic demonstrations. Subsequently, these skills condition a diffusion model to generate state trajectories adhering to the learned skills. Through integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser can comprehend and execute abstract instructions for various manipulation tasks. Extensive experiments on manipulation benchmarks demonstrate state-of-the-art performance, highlighting its effectiveness for multi-step composition tasks and ability to automatically learn interpretable skills." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.667, + 0.847 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.891, + 0.901 + ], + "angle": 0, + "content": "This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16474" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.632 + ], + "angle": 0, + "content": "[8] Yan Duan, Marcin Andrychowicz, Bradly Stadie, OpenAI Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel, and Wojciech Zaremba. One-shot imitation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[9] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.69, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[10] William Feller. On the theory of stochastic processes, with particular reference to applications. In Selected Papers I, pages 769-798. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.734, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[11] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Divyansh Garg, Shuvam Chakraborty, Chris Cundy, Jiaming Song, and Stefano Ermon. Iq-learn: Inverse soft-q learning for imitation. Advances in Neural Information Processing Systems, 34:4028–4039, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Divyansh Garg, Skanda Vaidyanath, Kuno Kim, Jiaming Song, and Stefano Ermon. Lisa: Learning interpretable skill abstractions from language. Advances in Neural Information Processing Systems, 35:21711-21724, 2022. 2, 3, 6, 7, 8, 5" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[14] Haoran He, Chenjia Bai, Kang Xu, Zhuoran Yang, Weinan Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Diffusion model is an effective planner and data synthesizer for multitask reinforcement learning. Advances in neural information processing systems, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Bernd Heidergott, editor. *Taylor Series Expansions*, pages 179–263. Springer US, Boston, MA, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[19] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[20] Yiding Jiang, Shixiang Shane Gu, Kevin P Murphy, and Chelsea Finn. Language as an abstraction for hierarchical deep reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[22] Alexander C Li, Carlos Florensa, Ignasi Clavera, and Pieter Abbeel. Sub-policy adaptation for hierarchical reinforcement learning. arXiv preprint arXiv:1906.05862, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[23] Zhixuan Liang, Yao Mu, Mingyu Ding, Fei Ni, Masayoshi Tomizuka, and Ping Luo. Adaptdiffuser: Diffusion models as adaptive self-evolving planners. In International Conference on Machine Learning, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[25] Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Causal transformer for estimating counterfactual outcomes. In International Conference on Machine Learning, pages 15293-15329. PMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[26] Diganta Misra. Mish: A self regularized non-monotonic activation function. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[27] Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical reinforcement learning. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[28] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Suraj Nair, Eric Mitchell, Kevin Chen, Silvio Savarese, Chelsea Finn, et al. Learning language-conditioned robot behavior from offline data and crowd-sourced annotation. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Conference on Robot Learning, pages 1303-1315. PMLR, 2022. 2, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[30] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In Conference on Robot Learning, pages 892-909. PMLR, 2023. 2, 4, 7, 8, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[31] Fei Ni, Jianye Hao, Yao Mu, Yifu Yuan, Yan Zheng, Bin Wang, and Zhixuan Liang. Metadiffuser: Diffusion model as conditional planner for offline meta-rl. In International Conference on Machine Learning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.301 + ], + "angle": 0, + "content": "[32] Scott Niekum, Sarah Osentoski, George Konidaris, Sachin Chitta, Bhaskara Marthi, and Andrew G Barto. Learning grounded finite-state representations from unstructured demonstrations. The International Journal of Robotics Research, 34(2):131-157, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[33] Dean A Pomerleau. Efficient training of artificial neural networks for autonomous navigation. Neural computation, 3 (1):88-97, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[34] Martin L Puterman. Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.428 + ], + "angle": 0, + "content": "[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.429, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[37] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 661-668. JMLR Workshop and Conference Proceedings, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[38] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[39] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In Proceedings of Thirty-third Conference on Neural Information Processing Systems, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[40] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[41] Avi Singh, Eric Jang, Alexander Irpan, Daniel Kappler, Murtaza Dalal, Sergey Levinev, Mohi Khansari, and Chelsea Finn. Scalable multi-task imitation learning with autonomous improvement. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2167-2173. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[42] Simon Stepputtis, Joseph Campbell, Mariano Phielipp, Stefan Lee, Chitta Baral, and Henri Ben Amor. Language-conditioned imitation learning for robot manipulation tasks. Advances in Neural Information Processing Systems, 33: 13139-13150, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[43] Alan Stuart and J Keith Ord. Kendall's advanced theory of statistics. vol. 1: Distribution theory. Kendall's advanced theory of statistics. Vol. 1: Distribution theory, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[44] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4950-4957, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[47] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[48] Danfei Xu, Suraj Nair, Yuke Zhu, Julian Gao, Animesh Garg, Li Fei-Fei, and Silvio Savarese. Neural task programming: Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 3795–3802. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[49] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In international conference on machine learning, pages 24631-24645. PMLR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[50] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[51] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[52] Jesse Zhang, Haonan Yu, and Wei Xu. Hierarchical reinforcement learning by discovering intrinsic options. arXiv preprint arXiv:2101.06521, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16476" + } + ] +] \ No newline at end of file diff --git a/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_origin.pdf b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..af1722948d243da7c9045d4ed0c55530b8689810 --- /dev/null +++ b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/7ae363f2-9817-4f7b-ab1b-3e5ad622cb0d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:324be2efbe8d5e146960e5037509486b5cb594339c68f9c37571ef071e8f40a9 +size 867871 diff --git a/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/full.md b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7cb41bef1170a7c5b5cee4e8de6e0511f151e7a --- /dev/null +++ b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/full.md @@ -0,0 +1,363 @@ +# SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution + +Zhixuan Liang $^{1,3}$ Yao Mu $^{1,3}$ Hengbo Ma $^{2}$ + +Masayoshi Tomizuka2 Mingyu Ding $^{2\dagger}$ Ping Luo $^{1,3\dagger}$ + +1The University of Hong Kong 2University of California, Berkeley 3Shanghai AI Laboratory + +{zxliang, ymu, pluo}@cs.hku.hk {hengbo.ma, tomizuka, myding}@berkeley.edu + +https://skilldiffuser.github.io/ + +# Abstract + +Diffusion models have demonstrated strong potential for robotic trajectory planning. However, generating coherent trajectories from high-level instructions remains challenging, especially for long-range composition tasks requiring multiple sequential skills. We propose SkillDiffuser, an end-to-end hierarchical planning framework integrating interpretable skill learning with conditional diffusion planning to address this problem. At the higher level, the skill abstraction module learns discrete, human-understandable skill representations from visual observations and language instructions. These learned skill embeddings are then used to condition the diffusion model to generate customized latent trajectories aligned with the skills. This allows generating diverse state trajectories that adhere to the learnable skills. By integrating skill learning with conditional trajectory generation, SkillDiffuser produces coherent behavior following abstract instructions across diverse tasks. Experiments on multi-task robotic manipulation benchmarks like Meta-World and LOReL demonstrate state-of-the-art performance and human-interpretable skill representations from SkillDiffuser. More visualization results and information could be found on our website. + +# 1. Introduction + +Recent research [6, 7, 18, 19] has demonstrated diffusion models' superior generative capabilities compared to previous models that help enhance reinforcement learning across various dimensions, including the generation of action trajectories [2, 19], policy representation [5, 50], and data synthesis [14, 23]. However, their ability to generate coherent trajectories for intricate tasks still poses challenges in terms of performance and generalizability, as these tasks often re + +![](images/b6656ca45fad5bf13f40a16020660d2f13b0ac8d09b01a8723a896f2f326f6ab.jpg) +Non-transferable + +![](images/7b27cdbb0feada732e1dc54ba9986bc764a77f02e17d3698481b46050f8b9fdb.jpg) +1 + +![](images/ccfb012514736eb469a2ddd9e299e76cda4863ccc61704245f82aeded07579d8.jpg) +Denoise U-Net +# + +![](images/2b836bbfc981a16ec7aac44594fcdeab0b722f76c3cfc8861e78c355557d99ea.jpg) +Instruction Input + +![](images/3672c9593dbb4b3c420c081dda63ded0befab90f551ec280017ee17bb160f784.jpg) +(a) Language-conditioned Diffuser +Transferable and Interpretable +Visual Input +Instruction Input +(b) Our SkillDiffuser +Figure 1. Comparison of SkillDiffuser and previous language conditioned diffusers. SkillDiffuser utilizes high-level abstraction to translate visual observations and language instructions into human understandable skills with language grounding. It then enables the low-level diffusion model condition on these skills, not only improving the execution performance of multi-step composition tasks but greatly enhancing the generalization and adaptability of the framework. + +require the fulfillment of abstract instructions that consist of numerous coordination-intensive sequential steps. + +Previous approaches [2, 14], such as Decision Diffuser, aim to tackle this challenge by decomposing complex tasks into simpler sub-skills, organized within a predefined skill library. These methods rely on conditioning the diffusion model with one-hot skill representations to generate trajectories for each of these sub-skills. Nevertheless, these techniques encounter difficulties in attempting to autonomously learn end-to-end from a wide range of datasets, which hinders their scalability and ability to achieve end-to-end learning [1]. In addition, without explicitly learning reusable skills, models cannot capture intricate inter-step dependen + +cies and constraints, yielding fragmented and illogical trajectories. Decomposing ambiguous instructions into learnable sub-goals and skills can better enable models to follow logical step sequences, respect task structures, and transfer common procedures (e.g. reusable skill abstraction and adaptable skill-based diffusion) between different tasks. This skill-centric paradigm paves the way for diffusion models that can interpret and execute elaborate, abstract instructions necessitating numerous sequential steps. + +Inspired by the above observations, we propose SkillDiffuser, a hierarchical planning framework unifying high-level skill learning with low-level conditional diffusion-based execution. As shown in Fig. 1, compared with previous language-conditioned diffusion policies, SkillDiffuser is able to interpret and execute complex instructions end-to-end with higher transferability. SkillDiffuser induces interpretable sub-latent goals by learning reusable skills tailored to the task instructions. The framework conditions a diffusion model on these learned skills to generate customized, coherent trajectories aligned with the overall objectives. By integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser achieves consistent, skill-oriented behavior without relying on a predefined skill library. Moreover, SkillDiffuser is designed to operate solely on visual observations, eliminating the need for robot proprioception (i.e. fully observed states). This end-to-end methodology, featuring learnable skills, enables SkillDiffuser to execute abstract instructions across a variety of tasks efficiently. + +SkillDiffuser works as follows. 1) A skill predictor with vector quantization [45] is used for high-level skill learning to distill tasks into discrete and interpretable skills. Rather than forecasting skill durations, we adopt a fixed prediction horizon - predicting skills at regular intervals. This horizon-based discretization process seamlessly integrates visual and linguistic inputs into a cohesive skill set guiding the low-level diffusion model. 2) For skill-based trajectory generation, we utilize a classifier-free diffusion model as policy, with skills directly embedded as guidance. This setting allows for generating multi-modal state trajectories aligned with skill specifications while avoiding overfitting to a closed dataset. 3) To enable action inference from predicted states, we train an inverse dynamics network to decode motions between two consecutive frames generated. By separating state prediction from motion decoding, SkillDiffuser yields a fully adaptive framework for directing diverse embodiments via transferable state-space plans. + +We evaluate the model's performance in both skill learning and multi-task planning on the LOReL [29] Sawyer and Meta-World [51] datasets, with crucial experimental settings by considering robotic agents in real-world scenarios must operate with limited state information, chiefly visual observations. Furthermore, we present success rates on un + +seen compositional tasks, the reusability and visualizations of learned skills to illustrate the model does have the ability to abstract high-level skills that are not only effective but also understandable to humans, bringing us closer to intelligent agents that acquire skills in a direct manner. + +Our contributions are three-fold: 1) We propose an end-to-end hierarchical planning framework via skill learning for sub-goal abstraction; 2) We adopt a classifier-free diffusion model conditioned on learned skills to generate skill-oriented transferable state trajectories; 3) We demonstrate state-of-the-art performance on complex benchmarks and provide interpretable visualizations of human-understandable skill representations. + +# 2. Related Works + +# 2.1. Imitation Learning and Multi-task Learning + +Imitation learning (IL) has evolved from foundational behavioral cloning to sophisticated multi-task learning frameworks. With traditional approaches relying on supervised learning from expert demonstrations [33, 37, 38], recent advancements have shifted towards learning the reward [16] or Q-function [12] from expert data, enhancing the ability to mimic complex behaviors. A new challenge lies in multi-task IL [41], where imitators are trained across varied tasks, aiming for generalization to new scenarios with task specifications ranging from vector states [28] to vision and language descriptions [8, 11, 13, 48]. + +Multi-task learning approaches often leverage shared representations to learn a spectrum of tasks simultaneously, enhancing the flexibility and efficiency of the learning process. The Meta-World benchmark [51] assesses multi-task and meta reinforcement learning, highlighting the need for algorithms capable of rapid adaptation. Building on this, the Prompting Decision Transformer [49] showcases few-shot policy generalization using task-specific prompts. And diffusion policy has also been explored in multi-task settings [14], which shows proficiency in generating diverse behaviors across tasks. However, unlike methods that leverage state inputs [14, 49] or access robot proprioception [30], SkillDiffuser uses raw visual inputs only. + +# 2.2. Skill Discovery and Hierarchical Learning + +Skill learning, the process by which robots acquire new abilities or refine existing ones, is gaining increasing attention due to its pivotal role in enabling autonomous systems to adapt to new tasks, improve the performance over time, and interact naturally with humans and complex environments. Traditionally, this domain was influenced by handcrafted features and expert demonstrations [32]. + +With the development of deep learning, Eysenbach et al. [9] and A. Sharma et al. [40] investigated skill discovery in learning methods, achieving policies conditioned + +on learned latent variables and maintaining consistent skill codes throughout trajectories. In the domain of skill learning through language instructions, LISA [13] stands out by sampling multiple skills per trajectory, integrating language conditioning in a unique manner. + +Our SkillDiffuser follows this way to extract sub-skills of each instruction at the higher-level. But differently, SkillDiffuser employs an adaptable diffusion policy at the lower-level to condition on different sub-skills to generate different actions, which formulates a creative hierarchical planning framework also advancing research on hierarchical techniques of reinforcement learning [22, 27, 52]. + +# 2.3. Planning with Diffusion Model + +Diffusion models [18] make great breakthroughs in image synthesis [18] recent years and has shown promising results in various generative applications. A seminal work that performs planning with diffusion directly is Diffuser [19], which laid the groundwork for using diffusion models in behavior synthesis. Then, a branch of this kind planning methods achieved state-of-the-art performance in a variety of decision tasks [3, 5, 7, 23, 31]. Among them, the work done by Chi et al. in [5] introduces the concept of learning the gradient of the action-distribution score function and iteratively optimizing it, demonstrating its significant potential in visuomotor policy learning. These works have further extended this direction and strengthened the versatility and generalization of diffusion-based planners. + +Our approach is inspired by classifier-free diffusion guidance [17], which offers a significant advantage over classifier-guided models [6]. By adopting the classifier-free approach, we can circumvent the challenges associated with training a reward model and $Q$ -function, which are particularly cumbersome in many real-world planning scenarios of which the complexity is very high. Recent studies have also extended this direction, which use conditional diffusion models to generate customized trajectories. Decision Diffuser [2] is an example which is designed to generate trajectories for a predefined skill library. However, unlike our method, it can't autonomously learn skill abstractions in an end-to-end fashion, which makes it difficult to scale to more tasks. This highlights the necessity for diffusion models with dynamic, learnable skill abstractions, facilitating complex instruction execution. + +# 3. Preliminary + +# 3.1. Planning with Diffusion over States + +As introduced in previous works [2, 19], diffusion model is a promising tool to address the problem of planning in reinforcement learning which is cast as a Markov Decision Process (MDP) [34]. Within the MDP framework $\mathcal{M} = (\mathcal{S},\mathcal{A},\mathcal{T},\mathcal{R},\gamma)$ , the planning policy aims to iden + +tify an optimal action sequence $\pmb{a}_{0:T}^{*}$ that maximizes the expected cumulative rewards over a finite time horizon $T$ , governed by the state transition dynamics $\mathcal{T}$ and reward function $\mathcal{R}$ . $S$ is the state space and $\mathcal{A}$ is the action space. + +By treating the state trajectory as sequence data $\pmb{\tau}$ , with sequence modeling, diffusion probabilistic models conceptualize planning as an iterative denoising process. The model progressively refines trajectories by reversing a forward diffusion process that is modeled as a Gaussian process, whereby noise is incrementally added to the data, denoted as $p_{\theta}(\pmb{\tau}^{i-1} \mid \pmb{\tau}^i)$ . Training involves minimizing the ELBO of the data's negative log-likelihood, similar to variational Bayesian inference, with the optimization objective: + +$$ +\theta^ {*} = \arg \min _ {\theta} - \mathbb {E} _ {\boldsymbol {\tau} ^ {0}} \left[ \log p _ {\theta} \left(\boldsymbol {\tau} ^ {0}\right) \right], \tag {1} +$$ + +where $p(\boldsymbol{\tau}^N)$ is a standard normal distribution and $\boldsymbol{\tau}^0$ denotes noiseless sequence data. + +For practical implementation, a simplified surrogate loss function is proposed in [18], focusing on predicting the Gaussian mean of the reverse diffusion step: + +$$ +\mathcal {L} _ {\text {d e n o i s e}} (\theta) = \mathbb {E} _ {i, \boldsymbol {\tau} ^ {0} \sim q, \epsilon \sim \mathcal {N}} [ \| \epsilon - \epsilon_ {\theta} (\boldsymbol {\tau} ^ {i}, i) \| ^ {2} ]. \tag {2} +$$ + +# 3.2. Classifier-free Diffusion Guidance + +On the basis of unconditioned diffusion-based method, in the realm of offline reinforcement learning, a critical endeavor is to generate trajectories with the highest reward-to-go. With the flourishing development of conditional diffusion models [6], classifier-guided approaches embark on this by infusing specific trajectory information (encoded in $y(\tau)$ ), such as the return $\mathcal{J}(\tau^0)$ or designated constraints, into the diffusion process: + +$$ +q \left(\boldsymbol {\tau} ^ {i} \mid \boldsymbol {\tau} ^ {i - 1}\right), \quad p _ {\theta} \left(\boldsymbol {\tau} ^ {i - 1} \mid \boldsymbol {\tau} ^ {i}, \boldsymbol {y} (\boldsymbol {\tau})\right). \tag {3} +$$ + +With assumptions specified in [10], we have + +$$ +\boldsymbol {\tau} ^ {i - 1} \sim \mathcal {N} \left(\mu_ {\theta} + \alpha \Sigma \nabla_ {\boldsymbol {\tau}} \log p (\boldsymbol {y} (\boldsymbol {\tau} ^ {i})), \Sigma\right), \tag {4} +$$ + +where $\alpha$ is a hyperparameter that adjusts the guidance strength, $\Sigma$ is the specified covariance of the noise and $\mu_{\theta}$ is the learned mean value of noise in unconditional diffusion. + +However, the classifier-guided diffusion model requires an accurate estimation of guidance gradient based on the trajectory classifier $\pmb{y}(\pmb{\tau})$ , which may not be feasible and need to introduce a separate dynamic programming procedure to estimate a $Q$ -function in the training process. + +Thus, classifier-free guidance offers an alternative, which augments the trajectory generation process with a guidance signal that amplifies the features of high-reward or optimal characteristics, i.e. $y(\tau)$ , that are implicitly present in the data. Mathematically, the noise to add during the reverse denoising process is: + +$$ +\hat {\epsilon} = \epsilon_ {\theta} (\boldsymbol {\tau} ^ {i}, \varnothing , i) + \omega \left(\epsilon_ {\theta} \left(\boldsymbol {\tau} ^ {i}, \boldsymbol {y}, i\right) - \epsilon_ {\theta} \left(\boldsymbol {\tau} ^ {i}, \varnothing , i\right)\right), \tag {5} +$$ + +where $\omega$ is the guidance scale, and $\varnothing$ represents the absence of guidance. Setting $\omega = 0$ removes the classifier-free + +guidance towards an unconditional diffusion model, while a large value of $\omega$ strengthens the influence of the conditional information during trajectory generation. + +Also, the loss function to minimize can be rewritten as, + +$$ +\mathcal {L} _ {\text {d i f f}} (\theta) = \mathbb {E} _ {i, \boldsymbol {\tau}, \epsilon} \left[ \left\| \epsilon - \epsilon_ {\theta} \left(\boldsymbol {\tau} ^ {i}, (1 - \beta) \boldsymbol {y} \left(\boldsymbol {\tau} ^ {i}\right) + \beta \emptyset , i\right) \right\| ^ {2} \right], \tag {6} +$$ + +where $\beta$ is a hyperparameter that controls the probability of dropping specific condition $\pmb{y}(\pmb{\tau})$ , enhancing sample diversity while maintaining context relevance. + +After training the noise prediction model $\epsilon_{\theta}$ with the above $\mathcal{L}(\theta)$ , the trajectory is sampled starting from Gaussian noise and progressively denoised with modified $\hat{\epsilon}$ through Eq. 5, employing re-parameterization technique. + +In summary, through classifier-free guidance, we can modulate the trajectory sampling process to require the generated trajectories more aligned with the desired features represented by $\pmb{y}(\pmb{\tau})$ . This process iteratively applies the conditioned noise model to refine the target trajectories that contain future states satisfying the constraints. + +# 4. Methodology + +# 4.1. Overview + +Building upon the motivations discussed above, we present SkillDiffuser, an advanced methodological framework for robust multi-task learning across various robots. This dynamic approach leverages the cooperation of skill learning at the higher level and a conditional diffusion model at the lower level. An overall framework is illustrated in Fig. 2. Notably, we leverage language-grounded representations to abstract skills, thereby rendering the execution of tasks through our diffusion policy both interpretable and comprehensible to humans. + +# 4.2. High Level Interpretable Skill Abstractions + +In our SkillDiffuser framework, the high level interpretable skill abstraction module plays a crucial role in understanding and executing complex tasks. However, given the multitask environments we consider, each task with only a single language instruction may be broken down into a sequence of sub-tasks or skills, which are not explicitly delineated within the instruction itself. Furthermore, suppose the offline training dataset is denoted as $\mathcal{D}$ , it consists of trajectories derived from a sub-optimal policy for various tasks. A trajectory $\boldsymbol{\tau} = (l, \{\boldsymbol{i}_t, \boldsymbol{a}_t\}_{t=1}^T)$ includes language description $l$ and a sequence of image observations and actions $(\boldsymbol{i}_t, \boldsymbol{a}_t)$ over time steps $T$ , with no reward labels attached. But the trajectories do not indicate the boundaries between sub-tasks as well, which thus requires the proposed methods capable of segmenting and interpreting the tasks into sub-goals in an unsupervised manner. + +To address this challenge, we build a skill abstraction component upon a horizon-based skill predictor adapted + +![](images/db18d1540526ccb8eb5383397f98c047494aadddf21f202a52fdfb7eac6784fb.jpg) +Figure 2. Overall framework of SkillDiffuser. It's a hierarchical planning model that leverages the cooperation of interpretable skill abstractions at the higher level and a skill conditioned diffusion model at the lower level for task execution in a multi-task learning environment. The high-level skill abstraction is achieved through a skill predictor and a vector quantization operation, generating sub-goals (skill set) that the diffusion model employs to determine the appropriate future states. Future states are converted to actions using an inverse dynamics model. This unique fusion enables a consistent underlying planner across different tasks, with the variation only in the inverse dynamics model. + +from GPT-2 [35]. It is designed to parse and decompose tasks by fusing visual input and natural language instructions, alongside a Vector Quantization (VQ) sub-module that discretizes the learned skills into a skill set. The specifics of this component are illustrated below. + +Firstly, as we utilize only images as robot state information, we transform the images into latent space features with a fixed image encoder (e.g. R3M [30]). For convenience, we denote the image encoder as $\Phi_{im}:\mathcal{I}\to \mathcal{R}^I$ , where $\mathcal{I}$ represents the space of input images and $\mathcal{R}^I$ , the resultant feature space, serves to condense the visual information into a form that is conducive for high-level semantics. Simultaneously, we use a language encoder to pre-process the natural language instructions, which is formalized as $\Phi_{lang}:L\rightarrow \mathcal{R}^{L}$ , with $L$ being the space of language instructions and $\mathcal{R}^L$ the language feature space. The outputs of both encoders are then fed into the skill predictor, which operates to integrate these two modalities. + +The skill predicting process is as follows: An image $\boldsymbol{i}_t \in \mathcal{I}$ at time step $t$ is encoded into a visual embedding $s_t = \Phi_{im}(\boldsymbol{i}_t)$ . This embedding $s_t$ is then input to the skill predictor, along with the language instruction $l \in L$ through the language encoder's output $l_t = \Phi_{lang}(l)$ . The + +![](images/b8b722df06785f69ad27546734123b48ad4ba01e9688a0b22029a98de30a00fc.jpg) +Figure 3. SkillDiffuser's low level skill-conditioned diffusion planning model. Notably, while the schematic here employs images to represent visual features for illustrative purposes, in actual implementation, both the input to and the sampling output of the diffusion model are state embeddings. The current observation is also the feature embedding of current visual observation. + +skill predictor, represented as $f: \mathcal{R}^I \times \mathcal{R}^L \to \mathcal{C}$ , generates a skill code $\tilde{z}$ by $\tilde{z} = f(\boldsymbol{s}_t, \boldsymbol{l}_t)$ that encapsulates task's requirements interpreted from the visual and language inputs. + +After that, Vector Quantization [45] operation $\mathbf{q}(\cdot)$ is taken to transform these predicted skills into a low-dimensional discrete set $\mathcal{C}$ . The discrete skill set contains $K$ skill embeddings $\{z^1, z^2, \ldots, z^K\}$ which represent different potential skills. VQ operation is achieved by mapping the latent $\tilde{z}$ to its closest entry of the skill set with skill vectors updated to be the moving average of the embeddings $z$ closest to them, same as [45]. This process is as follows, + +$$ +\tilde {z} = f \left(\Phi_ {i m} (\boldsymbol {i} _ {t}), \Phi_ {l a n g} (l)\right), \tag {7} +$$ + +$$ +z = \mathbf {q} (\tilde {z}) = \underset {z ^ {k} \in \mathcal {C}} {\arg \min } \| \tilde {z} - z ^ {k} \| _ {2}. \tag {8} +$$ + +VQ enforces each learnt skill $z$ to lie in $\mathcal{C}$ , which is equal to learning $K$ prototypes for the language embeddings utilizing $k$ -means [24] algorithm. This acts as a bottleneck, restricting the flow of language information and aiding in the learning of discrete skill codes. The back-propagation through the non-differentiable quantization is achieved by a straight-through gradient estimator, which simply copies the gradients and enables model to be trained end-to-end. + +In our approach, we apply a consistent skill code $z$ across a defined number of time steps, termed the horizon. This consistent application across multiple horizons adeptly addresses the challenge of varying sub-task durations without altering the horizon itself. Consequently, this strategy not only preserves the flexibility required for diverse task execution but also maintains the model's architectural stability by avoiding horizon-induced structural changes. + +Importantly, the discrete nature of the learned skill codes enhances the interpretability and controllability of the system's behavior, as each skill code is associated with some human-understandable language phrases. An example is depicted in Fig 6. Through this method, SkillDiffuser can not only learn to perform tasks based on language instructions but also achieve them in a human-interpretable way, allowing for a deeper understanding and control of the decision-making process of specific embodied agents. + +# 4.3. Low Level Skill-conditioned Diffusion Planning + +As highlighted in Section 1, while existing approaches like Decision Diffuser [2] have introduced conditional diffusion models constrained by skills, their capability is limited to generating trajectories that meet only predefined skills. Consequently, these models fall short of realizing a diffusion model capable of conditioning on an arbitrary learned skill. To overcome these constraints and enable diffusion models to plan over a learned continuous spectrum of skills, we propose an approach that leverages the classifier-free guided diffusion model with the skills embedded. + +SkillDiffuser begins by employing a diffusion model to operate over the continuous skill space $\mathcal{Z}$ learned during high-level skill abstraction. We employ a U-Net to serve as the noise prediction model $\epsilon_{\theta}(\cdot)$ and guide it by in-context conditioning. More specifically, we firstly utilize a skill MLP (similar to point-wise feed-forward network [46]), denoted as $\Lambda$ , to align skill features with denoising model. After that, we fuse the skill embeddings $\Lambda(z)$ into the state features throughout the residual blocks of U-Net. Details + +are depicted in Fig. 3. In this way, we make the diffusion model contextually modulate the influence of skill embeddings at each step. And following Eq. 5, we can synthesize future states attending to these given skills. This is a significant shift from previous static conditioning framework to a more dynamic and adaptive trajectory generation process. + +Moreover, following previous works [2, 7, 23], we adopt a state-only diffusion model, which eschews the direct generation of actions in diffusion model. And instead, we utilize another MLP, denoted as $\Psi(\cdot)$ , to perform inverse dynamics after state generation to infer feasible actions that can achieve transitions between two continuous states. Additionally, we integrate observations from the current frame to provide more detailed information for motion prediction and achieve closed-loop control. Mathematically, it is: + +$$ +\boldsymbol {a} _ {t} = \Psi ([ \tilde {\boldsymbol {s}} _ {t}, \tilde {\boldsymbol {s}} _ {t + 1} ], \boldsymbol {i} _ {t}) \quad \text {f o r} \quad t = 0, \dots , T - 1, \tag {9} +$$ + +where $\tilde{\boldsymbol{s}}_t$ and $\tilde{\boldsymbol{s}}_{t + 1}$ are consecutive observation embeddings within $\tilde{\tau}$ generated by diffusion model, $i_t$ is the current observation and $a_{t}$ is the inferred action. + +As the resulting action sequence $\{a_0, a_1, \ldots, a_{T-1}\}$ derived from the generated states, it encapsulates the skills to execute the tasks, which allows for remarkable adaptability across multiple tasks. And when faced with a new task, we are only required to change the inverse dynamics model $\Psi(\cdot)$ specific to the new task's kinematics, with the architecture and parameters of diffusion model unchanged. Such modularity ensures the generative capabilities of SkillDiffuser are not task-specific but can be leveraged across a diverse range of tasks with varying dynamics. A schematic diagram of the low level module is shown in Fig. 3. + +# 4.4. Training the SkillDiffuser + +We engineered a threefold loss function for SkillDiffuser. Firstly, for the inverse dynamics model which is task-specific, we employ a behavior cloning loss [44] to train our inverse MLP emulating expert actions from observed state transitions. This loss, denoted as $\mathcal{L}_{inv}$ , is formulated as: + +$$ +\mathcal {L} _ {i n v} = \mathbb {E} _ {\boldsymbol {\tau} \sim \mathcal {D}} \left[ \left\| \boldsymbol {a} - \Psi (\tilde {\boldsymbol {s}}, \tilde {\boldsymbol {s}} ^ {\prime}, \boldsymbol {i}) \right\| _ {2} ^ {2} \right], \tag {10} +$$ + +where the notations are similar to Eq. 9. + +Correspondingly, the other parts including both skill abstraction and low level execution are task-agnostic. For the high-level skill abstraction module, we apply a vector quantization (VQ) loss to refine the skill predictor. This VQ loss, $\mathcal{L}_{\mathrm{VQ}}$ , ensures the embeddings produced by the skill predictor closely match the skill set vectors, thereby improving the interpretability and consistency of the skill representations. Inspired by VQ-VAE [45], we formulate it as: + +$$ +\mathcal {L} _ {\mathrm {V Q}} = \mathbb {E} _ {\boldsymbol {\tau}} \left[ \| \mathbf {q} (\tilde {z}) - \tilde {z} \| _ {2} ^ {2} \right], \tag {11} +$$ + +where $\tilde{z}$ follows Eq. 7. + +Lastly, for the low-level state-only skill-conditioned diffusion execution, we incorporate a diffusion loss $\mathcal{L}_{diff}$ as + +per Eq. 6, ensuring our model's state predictions are in line with both the skill guidance and temporal dynamics observed in expert demonstrations. Here, we take $y(\tau) = \Lambda(z)$ with $z$ derived from Eq. 7 and Eq. 8. + +To be noted, we train our SkillDiffuser with two optimizers, one for inverse dynamics model with $\mathcal{L}_{inv}$ and the other for overall parameters of high-level skill abstraction and low-level planning with $\mathcal{L}_{\mathrm{VQ}} + \lambda \mathcal{L}_{diff}$ , where $\lambda$ is a loss weight. This carefully constructed loss architecture enables SkillDiffuser to excel in a multi-task environment, generalizing across tasks by simply substituting the inverse dynamics model $\Psi(\cdot)$ specific to each new task's requirements. + +Additionally, we utilize a pre-trained distilBERT [39] as our language encoder, adopting the configuration consistent with LOReL [29] and LISA [13], while freezing its parameters to guarantee stability in language understanding. And we employ diverse settings to serve as the visual encoder to ensure fair comparison in different datasets. We elaborate the details in corresponding parts of Sec. 5. More training details are shown in Appendix G and we also provide some pseudo-code of our algorithm in Appendix B. + +# 5. Experiments + +We first present a comprehensive evaluation on the LOReL Sawyer dataset, and then perform the ablation study compared on Meta-World benchmark to illustrate the efficiency of our method. Finally, we visualize the learned skills of our method both on LOReL and Meta-World MT10. + +# 5.1. Datasets + +LOReL Sawyer Dataset [29] which is abbreviated from Language-conditioned Offline Reward learning, is composed of pseudo-expert trajectories or play data gathered from an arbitrary reinforcement learning policy, annotated with post-hoc crowd-sourced language directives. The LOReL Sawyer dataset encompasses 50k trajectories, each with 20 steps, in a simulated Sawyer robot environment. We assess our approach using the same six tasks as the original paper [29] with paraphrased instructions under five different situations (i.e. seen, unseen verb, unseen none, unseen verb + noun and human provided). This comes to a total of 77 instructions for all 6 tasks combined. More details about the dataset can be found in Appendix D.1. + +Meta-World Dataset [51]. It is a comprehensive benchmark designed for evaluating multi-task and meta reinforcement learning algorithms. It introduces a comprehensive suite of 50 distinct robotic manipulation tasks, all located within a unified tabletop environment controlled by a simulated Sawyer arm. The Multi-Task 10 (MT10) within MetaWorld is a subset comprising ten carefully selected tasks, balanced in terms of diversity and complexity. Details of the ten tasks can be found in Appendix D.3. + +
Task InstructionRandomLCBC [42]LCRL [20]Lang DT [13]LISA [13]SkillDiffuser
close drawer52%50%58%10%100%95 ± 3.2%
open drawer14%0%8%60%20%55 ± 13.3%
turn faucet left24%12%13%0%0%55 ± 9.3%
turn faucet right15%31%0%0%30%25 ± 4.4%
move black mug right12%73%0%20%60%18 ± 3.9%
move white mug down5%6%0%0%30%10 ± 1.7%
Average over tasks20%29%13%15%40%43 ± 1.1%
+ +Table 1. Task-wise success rates on LOReL Sawyer Dataset. We show our success rates compared to random policy, language-conditioned imitation learning (LCBC), language-conditioned Q-learning (LCRL), a flat non-hierarchical Decision Transformer (Lang-DT), and LISA. The results on each dataset are calculated over 3 seeds. SkillDiffuser outperforms all other methods in terms of average performance over all tasks. Best methods and those within $6\%$ of the best are highlighted in bold. + +
Rephrasal TypeLang DTLISA [13]SkillDiffuser
seen154043.65 ±4.7
unseen noun13.3333.3336.01 ±6.3
unseen verb28.333036.70 ±9.5
unseen noun+verb6.72042.02 ±3.8
human provided26.9827.3540.16 ±2.1
Average18.0730.1439.71
+ +# 5.2. Evaluation Results on LOReL Sawyer Dataset + +Baselines. We employ random policy, language conditioned imitation learning (LCBC) [42], language conditioned Q-learning (LCRL) [20], Lang-DT (also known as Flat Baseline in [13]) and state-of-the-art skill-learning method LISA [13] as our baselines. We follow the same settings as LOReL [29] for the first three algorithms, and follow LISA [13] for the last two. The random policy serves as a baseline. And LCBC mimics offline dataset behaviors based on instructions, aligning with previous works focusing on imitation learning to achieve language conditioned behavior. In contrast, LCRL employs reinforcement learning, labeling each episode's final state with language-instructed rewards. Lang-DT plays as a non-hierarchical benchmark with a Causal Transformer [25], while LISA incorporates a dual-transformer structure, with one for skill prediction and the other for action planning. We do not compare with LOReL planner [29] as it uses MPC on a learned reward function relying on human annotations, while LISA and ours learn with trajectory data only. + +Results. To ensure fair comparison, our method, SkillDiffuser, is designed to maintain a parameter count similar to baseline models. It employs the same visual and language encoder architecture as used in LISA [13] with meticulously matching of embedding dimensions and the number of heads across the layers of SkillDiffuser. + +The results are present in Tables 1 and 2, showing taskwise and rephrasal-wise success rates for LOReL, averaged + +Table 2. Rephrasal-wise success rates (in %) on LOReL Sawyer. Results of Lang DT, LISA and our SkillDiffuser are shown here. The standard error is calculated over 3 random seeds. + +
MethodLang DTLOReL [29]LISA [13]SkillDiffuser
Success Rate13.33 ±1.318.18 ±1.820.89 ±0.625.21 ±2.7
+ +Table 3. Performance on LOReL multi-step composition tasks. + +over 10 runs with a 20-step time horizon. SkillDiffuser, our approach, achieves the highest average performance in six different tasks, indicating its superior cross-task adaptability particularly when compared to similar approaches which yet are based on Decision Transformer [4], such as LISA [13]. Moreover, SkillDiffuser consistently excels in all rephrased types for LOReL test tasks, outperforming LISA by $9.6\%$ on average. This demonstrates the model's robustness in handling varied skill representations, marking a notable advancement in skill-conditioned diffusion model. + +# 5.3. Performance on LOReL Compositional Tasks + +Settings. We conduct experiments following the same settings of unseen composition tasks of LISA [13] with 12 composition instructions in Tab. 3. Detailed instructions are listed in Appendix D.2 with such an example that "open drawer and move black mug right". We extend the max number of episode steps from customary 20 to 40, as LISA. + +Results. We observe SkillDiffuser achieves 2x the performance of non-hierarchical baseline (i.e. w/o skill abstraction) and also improves about $25\%$ over LISA, highlighting its effectiveness. MPC-based LOReL planners are unable to perform as well in open scenarios like composition tasks. + +# 5.4. Ablation Study on Meta-World Dataset + +Settings. We conduct experiments on Meta-World MT10 benchmark with finely annotated instructions. We also use visual observations only. Details are in Appendix D.3. + +Baselines. We evaluate our method against three baselines, all modified from existing models. The first, Flat R3M, is adapted from R3M [30] paper's planner. As the original one utilizes the first four terms of robot proprioception, we eliminate them and make the planner focus exclusively on visual observations. The second baseline is a variant of our SkillDiffuser, lacking the high-level skill abstraction module but retaining the low-level conditional diffusion-based planner, + +
MethodLangSkill SetDiffuserPerformance
Flat R3M [30]XXX13.3%
LISA [13]✓✓X13.8%
Lang Diffuser✓X✓16.7%
SkillDiffuser✓✓✓23.3%
+ +to assess the impact of skill abstraction. This version integrates a two-layer MLP to predict options from visual and language inputs, functioning as a language-conditioned diffusion planner. The last baseline is our re-implemented LISA [13] for Meta-World Dataset to validate the efficiency of diffusion model. To ensure fairness, we use R3M as the visual encoder for all of these methods on Meta-World. + +Ablation on Language Skill Conditioning. Table 4 indicates our Meta-World MT10 task is quite different from and much more challenging than previous tasks that use states as observations [49] or take into robot proprioception [30]. We only utilize single-frame visual input and instructions. The Flat R3M method, lacking language conditioning and skill sets can only succeed on tasks like drawer-close and reach through behavior cloning. Lang-conditioned Diffuser and modified LISA both outperform Flat R3M, suggesting the value of each corresponding module. Our SkillDiffuser, discretizing skills into a skill set, achieves a $6\%$ higher performance than language-conditioned diffuser and a $9.5\%$ higher than LISA, demonstrating the effectiveness of this combinational architecture. + +# 5.5. Ablation Study on Reusability of Learned Skills + +To evaluate the reusability of our learned skills, we calculate the average number of different skills used for a single instruction and the total number of instructions using each skill of LOReL Sawyer Dataset in Table 5. (With max episode step being 20, we experiment with skill horizon 10.) We observe each single instruction uses 1.55 sub-skills on average and each skill is called multiple times than the number of instructions (75 with 5 eval episodes), verifying the transferability of learnt skills. As suggested in Tab.10 of [13], except a very small horizon will hurt the performance, learning sub-skills to get refined semantics helps perform different actions at different stages. Besides, we also visualize resulting images from applying discrete skills in Appendix E.1 to further validate skills' interpretability. + +# 5.6. Visualization Results of Learned Skill Set + +We show the visualization of skill set on LOReL compositional tasks here in Fig. 4 and results on original LOReL dataset in Fig. 5 and Meta-World dataset in Fig. 6 in Appendix C. The visual analysis of our SkillDiffuser's learned skills on LOReL compositional tasks reveals that out of the 20-size skill set, our method learned 11 skills (e.g. pull han + +Table 4. Ablation study of language skill conditioning on Meta-World. Results of Flat R3M, Language-condition diffuser, LISA and our SkillDiffuser are shown here. All are averaged over 3 runs. + +
# of learnt skills# of inst# of successUse 1 skillUse 2 skillsAverage
1737514464801.55
Freq of 17 skills30, 8, 14, 10, 5, 19, 4, 7, 20, 2, 9, 20, 25, 6, 10, 3, 18
+ +![](images/c68b2ccb7870580917e050779f5398c452eb66f7b0b96b5fca037d98b5d5b06a.jpg) +Table 5. Average number of different skills used for a single instruction and total number of instructions used for each skill. +Figure 4. Visualization of skill heat map on LOReL Sawyer compositional tasks. We display the word frequency associated with a skill set of size 20 in LOReL, normalized by column. The data's sparsity and distinct highlights indicate certain language tokens are uniquely linked to specific skills. There are eleven skills learned by our method. (zoom in for best view) + +dle [skill 0], open counter [skill 14], etc.) notably distinguished by their unique word highlights. These bright spots across eleven columns (changing from only one column at initial which corresponds to default BC) in the heatmap underscore the model's ability to identify and isolate distinct skills from visual inputs, without an explicitly defined skill library. This indicates not only a significant interpretative advancement over previous diffusion-based planning but a successful abstraction of high-level skill representations. + +# 6. Conclusion + +This paper presents SkillDiffuser, an integrated framework that enables robots to perform tasks from natural language instructions by enabling interpretable skill learning and conditional diffusion planning. It employs vector quantization to learn discrete and comprehensible skill representations directly from visual and linguistic demonstrations. Subsequently, these skills condition a diffusion model to generate state trajectories adhering to the learned skills. Through integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser can comprehend and execute abstract instructions for various manipulation tasks. Extensive experiments on manipulation benchmarks demonstrate state-of-the-art performance, highlighting its effectiveness for multi-step composition tasks and ability to automatically learn interpretable skills. + +# Acknowledgements + +This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622. + +# References + +[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. 2022. 1 +[2] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 5, 6 +[3] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 3 +[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 7 +[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 3 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 3 +[7] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 2023. 1, 3, 6 +[8] Yan Duan, Marcin Andrychowicz, Bradly Stadie, OpenAI Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel, and Wojciech Zaremba. One-shot imitation learning. Advances in neural information processing systems, 30, 2017. 2 +[9] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018.2 +[10] William Feller. On the theory of stochastic processes, with particular reference to applications. In Selected Papers I, pages 769-798. Springer, 2015. 3 +[11] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 2 +[12] Divyansh Garg, Shuvam Chakraborty, Chris Cundy, Jiaming Song, and Stefano Ermon. Iq-learn: Inverse soft-q learning for imitation. Advances in Neural Information Processing Systems, 34:4028–4039, 2021. 2 +[13] Divyansh Garg, Skanda Vaidyanath, Kuno Kim, Jiaming Song, and Stefano Ermon. Lisa: Learning interpretable skill abstractions from language. Advances in Neural Information Processing Systems, 35:21711-21724, 2022. 2, 3, 6, 7, 8, 5 + +[14] Haoran He, Chenjia Bai, Kang Xu, Zhuoran Yang, Weinan Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Diffusion model is an effective planner and data synthesizer for multitask reinforcement learning. Advances in neural information processing systems, 2023. 1, 2 +[15] Bernd Heidergott, editor. *Taylor Series Expansions*, pages 179–263. Springer US, Boston, MA, 2007. 1 +[16] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2 +[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 2 +[19] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 1, 3 +[20] Yiding Jiang, Shixiang Shane Gu, Kevin P Murphy, and Chelsea Finn. Language as an abstraction for hierarchical deep reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 7 +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7 +[22] Alexander C Li, Carlos Florensa, Ignasi Clavera, and Pieter Abbeel. Sub-policy adaptation for hierarchical reinforcement learning. arXiv preprint arXiv:1906.05862, 2019. 3 +[23] Zhixuan Liang, Yao Mu, Mingyu Ding, Fei Ni, Masayoshi Tomizuka, and Ping Luo. Adaptdiffuser: Diffusion models as adaptive self-evolving planners. In International Conference on Machine Learning, 2023. 1, 3, 6 +[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 5 +[25] Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Causal transformer for estimating counterfactual outcomes. In International Conference on Machine Learning, pages 15293-15329. PMLR, 2022. 7 +[26] Diganta Misra. Mish: A self regularized non-monotonic activation function. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020, 2020. 6 +[27] Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical reinforcement learning. Advances in neural information processing systems, 31, 2018. 3 +[28] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2 +[29] Suraj Nair, Eric Mitchell, Kevin Chen, Silvio Savarese, Chelsea Finn, et al. Learning language-conditioned robot behavior from offline data and crowd-sourced annotation. In + +Conference on Robot Learning, pages 1303-1315. PMLR, 2022. 2, 6, 7, 4 +[30] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In Conference on Robot Learning, pages 892-909. PMLR, 2023. 2, 4, 7, 8, 6 +[31] Fei Ni, Jianye Hao, Yao Mu, Yifu Yuan, Yan Zheng, Bin Wang, and Zhixuan Liang. Metadiffuser: Diffusion model as conditional planner for offline meta-rl. In International Conference on Machine Learning, 2023. 3 +[32] Scott Niekum, Sarah Osentoski, George Konidaris, Sachin Chitta, Bhaskara Marthi, and Andrew G Barto. Learning grounded finite-state representations from unstructured demonstrations. The International Journal of Robotics Research, 34(2):131-157, 2015. 2 +[33] Dean A Pomerleau. Efficient training of artificial neural networks for autonomous navigation. Neural computation, 3 (1):88-97, 1991. 2 +[34] Martin L Puterman. Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons, 1994. 3 +[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6 +[37] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 661-668. JMLR Workshop and Conference Proceedings, 2010. 2 +[38] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 2 +[39] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In Proceedings of Thirty-third Conference on Neural Information Processing Systems, 2019. 6 +[40] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019. 2 +[41] Avi Singh, Eric Jang, Alexander Irpan, Daniel Kappler, Murtaza Dalal, Sergey Levinev, Mohi Khansari, and Chelsea Finn. Scalable multi-task imitation learning with autonomous improvement. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2167-2173. IEEE, 2020. 2 +[42] Simon Stepputtis, Joseph Campbell, Mariano Phielipp, Stefan Lee, Chitta Baral, and Henri Ben Amor. Language-conditioned imitation learning for robot manipulation tasks. Advances in Neural Information Processing Systems, 33: 13139-13150, 2020. 7 + +[43] Alan Stuart and J Keith Ord. Kendall's advanced theory of statistics. vol. 1: Distribution theory. Kendall's advanced theory of statistics. Vol. 1: Distribution theory, 1994. 2 +[44] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4950-4957, 2018. 6 +[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5, 6 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[47] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 6 +[48] Danfei Xu, Suraj Nair, Yuke Zhu, Julian Gao, Animesh Garg, Li Fei-Fei, and Silvio Savarese. Neural task programming: Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 3795–3802. IEEE, 2018. 2 +[49] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In international conference on machine learning, pages 24631-24645. PMLR, 2022. 2, 8 +[50] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 1 +[51] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 2, 6 +[52] Jesse Zhang, Haonan Yu, and Wei Xu. Hierarchical reinforcement learning by discovering intrinsic options. arXiv preprint arXiv:2101.06521, 2021. 3 \ No newline at end of file diff --git a/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/images.zip b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bfeec103d7b1423854d7cd63ba25b7a2c99d40b9 --- /dev/null +++ b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f790836399e3491545579e574697881ae6eca477f607936c5c0059b7312f18b +size 362089 diff --git a/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/layout.json b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2f271bfe75a55a099761da4c36ff8a7744cf7402 --- /dev/null +++ b/2024/SkillDiffuser_ Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution/layout.json @@ -0,0 +1,9312 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 529, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 529, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 529, + 138 + ], + "type": "text", + "content": "SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "spans": [ + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "type": "text", + "content": "Zhixuan Liang" + }, + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "type": "text", + "content": " Yao Mu" + }, + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "type": "text", + "content": " Hengbo Ma" + }, + { + "bbox": [ + 181, + 160, + 409, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 175, + 433, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 175, + 433, + 190 + ], + "spans": [ + { + "bbox": [ + 164, + 175, + 433, + 190 + ], + "type": "text", + "content": "Masayoshi Tomizuka2 Mingyu Ding" + }, + { + "bbox": [ + 164, + 175, + 433, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 164, + 175, + 433, + 190 + ], + "type": "text", + "content": " Ping Luo" + }, + { + "bbox": [ + 164, + 175, + 433, + 190 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 193, + 531, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 193, + 531, + 209 + ], + "spans": [ + { + "bbox": [ + 60, + 193, + 531, + 209 + ], + "type": "text", + "content": "1The University of Hong Kong 2University of California, Berkeley 3Shanghai AI Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 209, + 496, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 209, + 496, + 221 + ], + "spans": [ + { + "bbox": [ + 96, + 209, + 496, + 221 + ], + "type": "text", + "content": "{zxliang, ymu, pluo}@cs.hku.hk {hengbo.ma, tomizuka, myding}@berkeley.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 198, + 223, + 392, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 223, + 392, + 235 + ], + "spans": [ + { + "bbox": [ + 198, + 223, + 392, + 235 + ], + "type": "text", + "content": "https://skilldiffuser.github.io/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 263, + 192, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 263, + 192, + 276 + ], + "spans": [ + { + "bbox": [ + 143, + 263, + 192, + 276 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 289, + 290, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 290, + 552 + ], + "type": "text", + "content": "Diffusion models have demonstrated strong potential for robotic trajectory planning. However, generating coherent trajectories from high-level instructions remains challenging, especially for long-range composition tasks requiring multiple sequential skills. We propose SkillDiffuser, an end-to-end hierarchical planning framework integrating interpretable skill learning with conditional diffusion planning to address this problem. At the higher level, the skill abstraction module learns discrete, human-understandable skill representations from visual observations and language instructions. These learned skill embeddings are then used to condition the diffusion model to generate customized latent trajectories aligned with the skills. This allows generating diverse state trajectories that adhere to the learnable skills. By integrating skill learning with conditional trajectory generation, SkillDiffuser produces coherent behavior following abstract instructions across diverse tasks. Experiments on multi-task robotic manipulation benchmarks like Meta-World and LOReL demonstrate state-of-the-art performance and human-interpretable skill representations from SkillDiffuser. More visualization results and information could be found on our website." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 576, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 128, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "type": "text", + "content": "Recent research [6, 7, 18, 19] has demonstrated diffusion models' superior generative capabilities compared to previous models that help enhance reinforcement learning across various dimensions, including the generation of action trajectories [2, 19], policy representation [5, 50], and data synthesis [14, 23]. However, their ability to generate coherent trajectories for intricate tasks still poses challenges in terms of performance and generalizability, as these tasks often re" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 274, + 377, + 321 + ], + "blocks": [ + { + "bbox": [ + 321, + 265, + 370, + 273 + ], + "lines": [ + { + "bbox": [ + 321, + 265, + 370, + 273 + ], + "spans": [ + { + "bbox": [ + 321, + 265, + 370, + 273 + ], + "type": "text", + "content": "Non-transferable" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 274, + 377, + 321 + ], + "lines": [ + { + "bbox": [ + 321, + 274, + 377, + 321 + ], + "spans": [ + { + "bbox": [ + 321, + 274, + 377, + 321 + ], + "type": "image", + "image_path": "b6656ca45fad5bf13f40a16020660d2f13b0ac8d09b01a8723a896f2f326f6ab.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 334, + 332, + 371, + 340 + ], + "blocks": [ + { + "bbox": [ + 333, + 325, + 342, + 331 + ], + "lines": [ + { + "bbox": [ + 333, + 325, + 342, + 331 + ], + "spans": [ + { + "bbox": [ + 333, + 325, + 342, + 331 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 332, + 371, + 340 + ], + "lines": [ + { + "bbox": [ + 334, + 332, + 371, + 340 + ], + "spans": [ + { + "bbox": [ + 334, + 332, + 371, + 340 + ], + "type": "image", + "image_path": "7b27cdbb0feada732e1dc54ba9986bc764a77f02e17d3698481b46050f8b9fdb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 334, + 354, + 371, + 361 + ], + "blocks": [ + { + "bbox": [ + 334, + 346, + 371, + 354 + ], + "lines": [ + { + "bbox": [ + 334, + 346, + 371, + 354 + ], + "spans": [ + { + "bbox": [ + 334, + 346, + 371, + 354 + ], + "type": "text", + "content": "Denoise U-Net" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 354, + 371, + 361 + ], + "lines": [ + { + "bbox": [ + 334, + 354, + 371, + 361 + ], + "spans": [ + { + "bbox": [ + 334, + 354, + 371, + 361 + ], + "type": "image", + "image_path": "ccfb012514736eb469a2ddd9e299e76cda4863ccc61704245f82aeded07579d8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 365, + 342, + 373 + ], + "lines": [ + { + "bbox": [ + 334, + 365, + 342, + 373 + ], + "spans": [ + { + "bbox": [ + 334, + 365, + 342, + 373 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 407, + 377, + 421 + ], + "blocks": [ + { + "bbox": [ + 316, + 407, + 377, + 421 + ], + "lines": [ + { + "bbox": [ + 316, + 407, + 377, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 377, + 421 + ], + "type": "image", + "image_path": "2b836bbfc981a16ec7aac44594fcdeab0b722f76c3cfc8861e78c355557d99ea.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 430, + 365, + 439 + ], + "lines": [ + { + "bbox": [ + 324, + 430, + 365, + 439 + ], + "spans": [ + { + "bbox": [ + 324, + 430, + 365, + 439 + ], + "type": "text", + "content": "Instruction Input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 391, + 274, + 539, + 430 + ], + "blocks": [ + { + "bbox": [ + 311, + 441, + 382, + 456 + ], + "lines": [ + { + "bbox": [ + 311, + 441, + 382, + 456 + ], + "spans": [ + { + "bbox": [ + 311, + 441, + 382, + 456 + ], + "type": "text", + "content": "(a) Language-conditioned Diffuser" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 265, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 420, + 265, + 504, + 273 + ], + "spans": [ + { + "bbox": [ + 420, + 265, + 504, + 273 + ], + "type": "text", + "content": "Transferable and Interpretable" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 274, + 539, + 430 + ], + "lines": [ + { + "bbox": [ + 391, + 274, + 539, + 430 + ], + "spans": [ + { + "bbox": [ + 391, + 274, + 539, + 430 + ], + "type": "image", + "image_path": "3672c9593dbb4b3c420c081dda63ded0befab90f551ec280017ee17bb160f784.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 430, + 436, + 439 + ], + "lines": [ + { + "bbox": [ + 406, + 430, + 436, + 439 + ], + "spans": [ + { + "bbox": [ + 406, + 430, + 436, + 439 + ], + "type": "text", + "content": "Visual Input" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 480, + 430, + 521, + 438 + ], + "lines": [ + { + "bbox": [ + 480, + 430, + 521, + 438 + ], + "spans": [ + { + "bbox": [ + 480, + 430, + 521, + 438 + ], + "type": "text", + "content": "Instruction Input" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 445, + 494, + 453 + ], + "lines": [ + { + "bbox": [ + 436, + 445, + 494, + 453 + ], + "spans": [ + { + "bbox": [ + 436, + 445, + 494, + 453 + ], + "type": "text", + "content": "(b) Our SkillDiffuser" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 461, + 545, + 549 + ], + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 549 + ], + "type": "text", + "content": "Figure 1. Comparison of SkillDiffuser and previous language conditioned diffusers. SkillDiffuser utilizes high-level abstraction to translate visual observations and language instructions into human understandable skills with language grounding. It then enables the low-level diffusion model condition on these skills, not only improving the execution performance of multi-step composition tasks but greatly enhancing the generalization and adaptability of the framework." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 557, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 581 + ], + "type": "text", + "content": "require the fulfillment of abstract instructions that consist of numerous coordination-intensive sequential steps." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": "Previous approaches [2, 14], such as Decision Diffuser, aim to tackle this challenge by decomposing complex tasks into simpler sub-skills, organized within a predefined skill library. These methods rely on conditioning the diffusion model with one-hot skill representations to generate trajectories for each of these sub-skills. Nevertheless, these techniques encounter difficulties in attempting to autonomously learn end-to-end from a wide range of datasets, which hinders their scalability and ability to achieve end-to-end learning [1]. In addition, without explicitly learning reusable skills, models cannot capture intricate inter-step dependen" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16467" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": "cies and constraints, yielding fragmented and illogical trajectories. Decomposing ambiguous instructions into learnable sub-goals and skills can better enable models to follow logical step sequences, respect task structures, and transfer common procedures (e.g. reusable skill abstraction and adaptable skill-based diffusion) between different tasks. This skill-centric paradigm paves the way for diffusion models that can interpret and execute elaborate, abstract instructions necessitating numerous sequential steps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 422 + ], + "type": "text", + "content": "Inspired by the above observations, we propose SkillDiffuser, a hierarchical planning framework unifying high-level skill learning with low-level conditional diffusion-based execution. As shown in Fig. 1, compared with previous language-conditioned diffusion policies, SkillDiffuser is able to interpret and execute complex instructions end-to-end with higher transferability. SkillDiffuser induces interpretable sub-latent goals by learning reusable skills tailored to the task instructions. The framework conditions a diffusion model on these learned skills to generate customized, coherent trajectories aligned with the overall objectives. By integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser achieves consistent, skill-oriented behavior without relying on a predefined skill library. Moreover, SkillDiffuser is designed to operate solely on visual observations, eliminating the need for robot proprioception (i.e. fully observed states). This end-to-end methodology, featuring learnable skills, enables SkillDiffuser to execute abstract instructions across a variety of tasks efficiently." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 423, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 639 + ], + "type": "text", + "content": "SkillDiffuser works as follows. 1) A skill predictor with vector quantization [45] is used for high-level skill learning to distill tasks into discrete and interpretable skills. Rather than forecasting skill durations, we adopt a fixed prediction horizon - predicting skills at regular intervals. This horizon-based discretization process seamlessly integrates visual and linguistic inputs into a cohesive skill set guiding the low-level diffusion model. 2) For skill-based trajectory generation, we utilize a classifier-free diffusion model as policy, with skills directly embedded as guidance. This setting allows for generating multi-modal state trajectories aligned with skill specifications while avoiding overfitting to a closed dataset. 3) To enable action inference from predicted states, we train an inverse dynamics network to decode motions between two consecutive frames generated. By separating state prediction from motion decoding, SkillDiffuser yields a fully adaptive framework for directing diverse embodiments via transferable state-space plans." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "We evaluate the model's performance in both skill learning and multi-task planning on the LOReL [29] Sawyer and Meta-World [51] datasets, with crucial experimental settings by considering robotic agents in real-world scenarios must operate with limited state information, chiefly visual observations. Furthermore, we present success rates on un" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "seen compositional tasks, the reusability and visualizations of learned skills to illustrate the model does have the ability to abstract high-level skills that are not only effective but also understandable to humans, bringing us closer to intelligent agents that acquire skills in a direct manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": "Our contributions are three-fold: 1) We propose an end-to-end hierarchical planning framework via skill learning for sub-goal abstraction; 2) We adopt a classifier-free diffusion model conditioned on learned skills to generate skill-oriented transferable state trajectories; 3) We demonstrate state-of-the-art performance on complex benchmarks and provide interpretable visualizations of human-understandable skill representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 240, + 397, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 397, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 397, + 252 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 259, + 537, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 537, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 537, + 272 + ], + "type": "text", + "content": "2.1. Imitation Learning and Multi-task Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 278, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 410 + ], + "type": "text", + "content": "Imitation learning (IL) has evolved from foundational behavioral cloning to sophisticated multi-task learning frameworks. With traditional approaches relying on supervised learning from expert demonstrations [33, 37, 38], recent advancements have shifted towards learning the reward [16] or Q-function [12] from expert data, enhancing the ability to mimic complex behaviors. A new challenge lies in multi-task IL [41], where imitators are trained across varied tasks, aiming for generalization to new scenarios with task specifications ranging from vector states [28] to vision and language descriptions [8, 11, 13, 48]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "content": "Multi-task learning approaches often leverage shared representations to learn a spectrum of tasks simultaneously, enhancing the flexibility and efficiency of the learning process. The Meta-World benchmark [51] assesses multi-task and meta reinforcement learning, highlighting the need for algorithms capable of rapid adaptation. Building on this, the Prompting Decision Transformer [49] showcases few-shot policy generalization using task-specific prompts. And diffusion policy has also been explored in multi-task settings [14], which shows proficiency in generating diverse behaviors across tasks. However, unlike methods that leverage state inputs [14, 49] or access robot proprioception [30], SkillDiffuser uses raw visual inputs only." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 574, + 527, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 527, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 527, + 587 + ], + "type": "text", + "content": "2.2. Skill Discovery and Hierarchical Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": "Skill learning, the process by which robots acquire new abilities or refine existing ones, is gaining increasing attention due to its pivotal role in enabling autonomous systems to adapt to new tasks, improve the performance over time, and interact naturally with humans and complex environments. Traditionally, this domain was influenced by handcrafted features and expert demonstrations [32]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "With the development of deep learning, Eysenbach et al. [9] and A. Sharma et al. [40] investigated skill discovery in learning methods, achieving policies conditioned" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16468" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "on learned latent variables and maintaining consistent skill codes throughout trajectories. In the domain of skill learning through language instructions, LISA [13] stands out by sampling multiple skills per trajectory, integrating language conditioning in a unique manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 217 + ], + "type": "text", + "content": "Our SkillDiffuser follows this way to extract sub-skills of each instruction at the higher-level. But differently, SkillDiffuser employs an adaptable diffusion policy at the lower-level to condition on different sub-skills to generate different actions, which formulates a creative hierarchical planning framework also advancing research on hierarchical techniques of reinforcement learning [22, 27, 52]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 224, + 214, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 214, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 214, + 237 + ], + "type": "text", + "content": "2.3. Planning with Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 411 + ], + "type": "text", + "content": "Diffusion models [18] make great breakthroughs in image synthesis [18] recent years and has shown promising results in various generative applications. A seminal work that performs planning with diffusion directly is Diffuser [19], which laid the groundwork for using diffusion models in behavior synthesis. Then, a branch of this kind planning methods achieved state-of-the-art performance in a variety of decision tasks [3, 5, 7, 23, 31]. Among them, the work done by Chi et al. in [5] introduces the concept of learning the gradient of the action-distribution score function and iteratively optimizing it, demonstrating its significant potential in visuomotor policy learning. These works have further extended this direction and strengthened the versatility and generalization of diffusion-based planners." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "type": "text", + "content": "Our approach is inspired by classifier-free diffusion guidance [17], which offers a significant advantage over classifier-guided models [6]. By adopting the classifier-free approach, we can circumvent the challenges associated with training a reward model and " + }, + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 411, + 287, + 603 + ], + "type": "text", + "content": "-function, which are particularly cumbersome in many real-world planning scenarios of which the complexity is very high. Recent studies have also extended this direction, which use conditional diffusion models to generate customized trajectories. Decision Diffuser [2] is an example which is designed to generate trajectories for a predefined skill library. However, unlike our method, it can't autonomously learn skill abstractions in an end-to-end fashion, which makes it difficult to scale to more tasks. This highlights the necessity for diffusion models with dynamic, learnable skill abstractions, facilitating complex instruction execution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 615, + 124, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 124, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 124, + 628 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 235, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 235, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 235, + 647 + ], + "type": "text", + "content": "3.1. Planning with Diffusion over States" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "As introduced in previous works [2, 19], diffusion model is a promising tool to address the problem of planning in reinforcement learning which is cast as a Markov Decision Process (MDP) [34]. Within the MDP framework " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathcal{S},\\mathcal{A},\\mathcal{T},\\mathcal{R},\\gamma)" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": ", the planning policy aims to iden" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "tify an optimal action sequence " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{0:T}^{*}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " that maximizes the expected cumulative rewards over a finite time horizon " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", governed by the state transition dynamics " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " and reward function " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the state space and " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the action space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 120, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 228 + ], + "type": "text", + "content": "By treating the state trajectory as sequence data " + }, + { + "bbox": [ + 304, + 120, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 304, + 120, + 545, + 228 + ], + "type": "text", + "content": ", with sequence modeling, diffusion probabilistic models conceptualize planning as an iterative denoising process. The model progressively refines trajectories by reversing a forward diffusion process that is modeled as a Gaussian process, whereby noise is incrementally added to the data, denoted as " + }, + { + "bbox": [ + 304, + 120, + 545, + 228 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{\\tau}^{i-1} \\mid \\pmb{\\tau}^i)" + }, + { + "bbox": [ + 304, + 120, + 545, + 228 + ], + "type": "text", + "content": ". Training involves minimizing the ELBO of the data's negative log-likelihood, similar to variational Bayesian inference, with the optimization objective:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 230, + 545, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 230, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 353, + 230, + 545, + 248 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\arg \\min _ {\\theta} - \\mathbb {E} _ {\\boldsymbol {\\tau} ^ {0}} \\left[ \\log p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {0}\\right) \\right], \\tag {1}", + "image_path": "6256dcafb43b5aecdb993bd3b6e6487d08697e51a23b07207654eca1660e72ff.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 248, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 248, + 545, + 272 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{\\tau}^N)" + }, + { + "bbox": [ + 304, + 248, + 545, + 272 + ], + "type": "text", + "content": " is a standard normal distribution and " + }, + { + "bbox": [ + 304, + 248, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\tau}^0" + }, + { + "bbox": [ + 304, + 248, + 545, + 272 + ], + "type": "text", + "content": " denotes noiseless sequence data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 273, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 308 + ], + "type": "text", + "content": "For practical implementation, a simplified surrogate loss function is proposed in [18], focusing on predicting the Gaussian mean of the reverse diffusion step:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 334, + 313, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 313, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 334, + 313, + 545, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e n o i s e}} (\\theta) = \\mathbb {E} _ {i, \\boldsymbol {\\tau} ^ {0} \\sim q, \\epsilon \\sim \\mathcal {N}} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {\\tau} ^ {i}, i) \\| ^ {2} ]. \\tag {2}", + "image_path": "db46e02619f4432c2b8532dbb28e52939c8cfeca6e17423872e0eacf9224b609.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 335, + 487, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 487, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 487, + 346 + ], + "type": "text", + "content": "3.2. Classifier-free Diffusion Guidance" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "text", + "content": "On the basis of unconditioned diffusion-based method, in the realm of offline reinforcement learning, a critical endeavor is to generate trajectories with the highest reward-to-go. With the flourishing development of conditional diffusion models [6], classifier-guided approaches embark on this by infusing specific trajectory information (encoded in " + }, + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "inline_equation", + "content": "y(\\tau)" + }, + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "text", + "content": "), such as the return " + }, + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{J}(\\tau^0)" + }, + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "text", + "content": " or designated constraints, into the diffusion process:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 450, + 545, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 450, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 354, + 450, + 545, + 465 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {\\tau} ^ {i} \\mid \\boldsymbol {\\tau} ^ {i - 1}\\right), \\quad p _ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i - 1} \\mid \\boldsymbol {\\tau} ^ {i}, \\boldsymbol {y} (\\boldsymbol {\\tau})\\right). \\tag {3}", + "image_path": "aa8d3fad701096159991fc984f92fdc88c939f12c40020a007f0a4d3bd29f4f5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 469, + 485, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 485, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 485, + 481 + ], + "type": "text", + "content": "With assumptions specified in [10], we have" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 337, + 486, + 545, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 486, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 337, + 486, + 545, + 500 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\tau} ^ {i - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} + \\alpha \\Sigma \\nabla_ {\\boldsymbol {\\tau}} \\log p (\\boldsymbol {y} (\\boldsymbol {\\tau} ^ {i})), \\Sigma\\right), \\tag {4}", + "image_path": "296ef595d58afd59487cbf03715d941642ac9e6fe647039b219f4ef62e5fcc77.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "text", + "content": " is a hyperparameter that adjusts the guidance strength, " + }, + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "text", + "content": " is the specified covariance of the noise and " + }, + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 304, + 502, + 545, + 537 + ], + "type": "text", + "content": " is the learned mean value of noise in unconditional diffusion." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 539, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 598 + ], + "type": "text", + "content": "However, the classifier-guided diffusion model requires an accurate estimation of guidance gradient based on the trajectory classifier " + }, + { + "bbox": [ + 304, + 539, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{y}(\\pmb{\\tau})" + }, + { + "bbox": [ + 304, + 539, + 545, + 598 + ], + "type": "text", + "content": ", which may not be feasible and need to introduce a separate dynamic programming procedure to estimate a " + }, + { + "bbox": [ + 304, + 539, + 545, + 598 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 539, + 545, + 598 + ], + "type": "text", + "content": "-function in the training process." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 598, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 545, + 670 + ], + "type": "text", + "content": "Thus, classifier-free guidance offers an alternative, which augments the trajectory generation process with a guidance signal that amplifies the features of high-reward or optimal characteristics, i.e. " + }, + { + "bbox": [ + 304, + 598, + 545, + 670 + ], + "type": "inline_equation", + "content": "y(\\tau)" + }, + { + "bbox": [ + 304, + 598, + 545, + 670 + ], + "type": "text", + "content": ", that are implicitly present in the data. Mathematically, the noise to add during the reverse denoising process is:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 319, + 673, + 545, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 673, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 319, + 673, + 545, + 687 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} = \\epsilon_ {\\theta} (\\boldsymbol {\\tau} ^ {i}, \\varnothing , i) + \\omega \\left(\\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, \\boldsymbol {y}, i\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, \\varnothing , i\\right)\\right), \\tag {5}", + "image_path": "d028f87e293f4c9a3611def2605654d9767611da81c8bffa8de3541a116d5f51.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the guidance scale, and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " represents the absence of guidance. Setting " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\omega = 0" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " removes the classifier-free" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16469" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "guidance towards an unconditional diffusion model, while a large value of " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " strengthens the influence of the conditional information during trajectory generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 109, + 282, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 109, + 282, + 119 + ], + "spans": [ + { + "bbox": [ + 59, + 109, + 282, + 119 + ], + "type": "text", + "content": "Also, the loss function to minimize can be rewritten as," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 120, + 290, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 290, + 149 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 290, + 149 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} (\\theta) = \\mathbb {E} _ {i, \\boldsymbol {\\tau}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {\\tau} ^ {i}, (1 - \\beta) \\boldsymbol {y} \\left(\\boldsymbol {\\tau} ^ {i}\\right) + \\beta \\emptyset , i\\right) \\right\\| ^ {2} \\right], \\tag {6}", + "image_path": "2395b002abace6b6b6eac017a242293bc3aed8fda9376d8814d6bad8128e3ae9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 151, + 287, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 151, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 151, + 287, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 151, + 287, + 186 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 151, + 287, + 186 + ], + "type": "text", + "content": " is a hyperparameter that controls the probability of dropping specific condition " + }, + { + "bbox": [ + 47, + 151, + 287, + 186 + ], + "type": "inline_equation", + "content": "\\pmb{y}(\\pmb{\\tau})" + }, + { + "bbox": [ + 47, + 151, + 287, + 186 + ], + "type": "text", + "content": ", enhancing sample diversity while maintaining context relevance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "text", + "content": "After training the noise prediction model " + }, + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "text", + "content": " with the above " + }, + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta)" + }, + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "text", + "content": ", the trajectory is sampled starting from Gaussian noise and progressively denoised with modified " + }, + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 47, + 186, + 287, + 234 + ], + "type": "text", + "content": " through Eq. 5, employing re-parameterization technique." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 235, + 287, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 306 + ], + "type": "text", + "content": "In summary, through classifier-free guidance, we can modulate the trajectory sampling process to require the generated trajectories more aligned with the desired features represented by " + }, + { + "bbox": [ + 46, + 235, + 287, + 306 + ], + "type": "inline_equation", + "content": "\\pmb{y}(\\pmb{\\tau})" + }, + { + "bbox": [ + 46, + 235, + 287, + 306 + ], + "type": "text", + "content": ". This process iteratively applies the conditioned noise model to refine the target trajectories that contain future states satisfying the constraints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 316, + 129, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 129, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 129, + 329 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 335, + 115, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 115, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 115, + 346 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 354, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 473 + ], + "type": "text", + "content": "Building upon the motivations discussed above, we present SkillDiffuser, an advanced methodological framework for robust multi-task learning across various robots. This dynamic approach leverages the cooperation of skill learning at the higher level and a conditional diffusion model at the lower level. An overall framework is illustrated in Fig. 2. Notably, we leverage language-grounded representations to abstract skills, thereby rendering the execution of tasks through our diffusion policy both interpretable and comprehensible to humans." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 479, + 272, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 272, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 272, + 493 + ], + "type": "text", + "content": "4.2. High Level Interpretable Skill Abstractions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "content": "In our SkillDiffuser framework, the high level interpretable skill abstraction module plays a crucial role in understanding and executing complex tasks. However, given the multitask environments we consider, each task with only a single language instruction may be broken down into a sequence of sub-tasks or skills, which are not explicitly delineated within the instruction itself. Furthermore, suppose the offline training dataset is denoted as " + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "content": ", it consists of trajectories derived from a sub-optimal policy for various tasks. A trajectory " + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\tau} = (l, \\{\\boldsymbol{i}_t, \\boldsymbol{a}_t\\}_{t=1}^T)" + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "content": " includes language description " + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "content": " and a sequence of image observations and actions " + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{i}_t, \\boldsymbol{a}_t)" + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "content": " over time steps " + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "content": ", with no reward labels attached. But the trajectories do not indicate the boundaries between sub-tasks as well, which thus requires the proposed methods capable of segmenting and interpreting the tasks into sub-goals in an unsupervised manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "To address this challenge, we build a skill abstraction component upon a horizon-based skill predictor adapted" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 525, + 309 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 525, + 309 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 525, + 309 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 525, + 309 + ], + "type": "image", + "image_path": "db18d1540526ccb8eb5383397f98c047494aadddf21f202a52fdfb7eac6784fb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 314, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 434 + ], + "type": "text", + "content": "Figure 2. Overall framework of SkillDiffuser. It's a hierarchical planning model that leverages the cooperation of interpretable skill abstractions at the higher level and a skill conditioned diffusion model at the lower level for task execution in a multi-task learning environment. The high-level skill abstraction is achieved through a skill predictor and a vector quantization operation, generating sub-goals (skill set) that the diffusion model employs to determine the appropriate future states. Future states are converted to actions using an inverse dynamics model. This unique fusion enables a consistent underlying planner across different tasks, with the variation only in the inverse dynamics model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": "from GPT-2 [35]. It is designed to parse and decompose tasks by fusing visual input and natural language instructions, alongside a Vector Quantization (VQ) sub-module that discretizes the learned skills into a skill set. The specifics of this component are illustrated below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": "Firstly, as we utilize only images as robot state information, we transform the images into latent space features with a fixed image encoder (e.g. R3M [30]). For convenience, we denote the image encoder as " + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\Phi_{im}:\\mathcal{I}\\to \\mathcal{R}^I" + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": " represents the space of input images and " + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^I" + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": ", the resultant feature space, serves to condense the visual information into a form that is conducive for high-level semantics. Simultaneously, we use a language encoder to pre-process the natural language instructions, which is formalized as " + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\Phi_{lang}:L\\rightarrow \\mathcal{R}^{L}" + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": " being the space of language instructions and " + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^L" + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": " the language feature space. The outputs of both encoders are then fed into the skill predictor, which operates to integrate these two modalities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "The skill predicting process is as follows: An image " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{i}_t \\in \\mathcal{I}" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " is encoded into a visual embedding " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "s_t = \\Phi_{im}(\\boldsymbol{i}_t)" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": ". This embedding " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " is then input to the skill predictor, along with the language instruction " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "l \\in L" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " through the language encoder's output " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "l_t = \\Phi_{lang}(l)" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16470" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 69, + 509, + 309 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 509, + 309 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 509, + 309 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 509, + 309 + ], + "type": "image", + "image_path": "b8b722df06785f69ad27546734123b48ad4ba01e9688a0b22029a98de30a00fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 313, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 313, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 547, + 346 + ], + "type": "text", + "content": "Figure 3. SkillDiffuser's low level skill-conditioned diffusion planning model. Notably, while the schematic here employs images to represent visual features for illustrative purposes, in actual implementation, both the input to and the sampling output of the diffusion model are state embeddings. The current observation is also the feature embedding of current visual observation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "text", + "content": "skill predictor, represented as " + }, + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "inline_equation", + "content": "f: \\mathcal{R}^I \\times \\mathcal{R}^L \\to \\mathcal{C}" + }, + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "text", + "content": ", generates a skill code " + }, + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\tilde{z} = f(\\boldsymbol{s}_t, \\boldsymbol{l}_t)" + }, + { + "bbox": [ + 46, + 349, + 287, + 386 + ], + "type": "text", + "content": " that encapsulates task's requirements interpreted from the visual and language inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "content": "After that, Vector Quantization [45] operation " + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(\\cdot)" + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "content": " is taken to transform these predicted skills into a low-dimensional discrete set " + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "content": ". The discrete skill set contains " + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "content": " skill embeddings " + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\{z^1, z^2, \\ldots, z^K\\}" + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "content": " which represent different potential skills. VQ operation is achieved by mapping the latent " + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "content": " to its closest entry of the skill set with skill vectors updated to be the moving average of the embeddings " + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 385, + 287, + 482 + ], + "type": "text", + "content": " closest to them, same as [45]. This process is as follows," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 484, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 484, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 110, + 484, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\tilde {z} = f \\left(\\Phi_ {i m} (\\boldsymbol {i} _ {t}), \\Phi_ {l a n g} (l)\\right), \\tag {7}", + "image_path": "b681a0f7787a60a1fd71c0d9221d80fea3cd3ac4fac3b3d335a2f03a4eba9323.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 501, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 501, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 99, + 501, + 287, + 521 + ], + "type": "interline_equation", + "content": "z = \\mathbf {q} (\\tilde {z}) = \\underset {z ^ {k} \\in \\mathcal {C}} {\\arg \\min } \\| \\tilde {z} - z ^ {k} \\| _ {2}. \\tag {8}", + "image_path": "530ef642b006b8d9bbbd740814d8a6615fa4baa461b622459bc098eb64d5fbbf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": "VQ enforces each learnt skill " + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": " to lie in " + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": ", which is equal to learning " + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": " prototypes for the language embeddings utilizing " + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": "-means [24] algorithm. This acts as a bottleneck, restricting the flow of language information and aiding in the learning of discrete skill codes. The back-propagation through the non-differentiable quantization is achieved by a straight-through gradient estimator, which simply copies the gradients and enables model to be trained end-to-end." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "In our approach, we apply a consistent skill code " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " across a defined number of time steps, termed the horizon. This consistent application across multiple horizons adeptly addresses the challenge of varying sub-task durations without altering the horizon itself. Consequently, this strategy not only preserves the flexibility required for diverse task execution but also maintains the model's architectural stability by avoiding horizon-induced structural changes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 349, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 458 + ], + "type": "text", + "content": "Importantly, the discrete nature of the learned skill codes enhances the interpretability and controllability of the system's behavior, as each skill code is associated with some human-understandable language phrases. An example is depicted in Fig 6. Through this method, SkillDiffuser can not only learn to perform tasks based on language instructions but also achieve them in a human-interpretable way, allowing for a deeper understanding and control of the decision-making process of specific embodied agents." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 466, + 545, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 466, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 305, + 466, + 545, + 480 + ], + "type": "text", + "content": "4.3. Low Level Skill-conditioned Diffusion Planning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 485, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 605 + ], + "type": "text", + "content": "As highlighted in Section 1, while existing approaches like Decision Diffuser [2] have introduced conditional diffusion models constrained by skills, their capability is limited to generating trajectories that meet only predefined skills. Consequently, these models fall short of realizing a diffusion model capable of conditioning on an arbitrary learned skill. To overcome these constraints and enable diffusion models to plan over a learned continuous spectrum of skills, we propose an approach that leverages the classifier-free guided diffusion model with the skills embedded." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "SkillDiffuser begins by employing a diffusion model to operate over the continuous skill space " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " learned during high-level skill abstraction. We employ a U-Net to serve as the noise prediction model " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and guide it by in-context conditioning. More specifically, we firstly utilize a skill MLP (similar to point-wise feed-forward network [46]), denoted as " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", to align skill features with denoising model. After that, we fuse the skill embeddings " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Lambda(z)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " into the state features throughout the residual blocks of U-Net. Details" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16471" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "are depicted in Fig. 3. In this way, we make the diffusion model contextually modulate the influence of skill embeddings at each step. And following Eq. 5, we can synthesize future states attending to these given skills. This is a significant shift from previous static conditioning framework to a more dynamic and adaptive trajectory generation process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 251 + ], + "type": "text", + "content": "Moreover, following previous works [2, 7, 23], we adopt a state-only diffusion model, which eschews the direct generation of actions in diffusion model. And instead, we utilize another MLP, denoted as " + }, + { + "bbox": [ + 46, + 144, + 286, + 251 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 46, + 144, + 286, + 251 + ], + "type": "text", + "content": ", to perform inverse dynamics after state generation to infer feasible actions that can achieve transitions between two continuous states. Additionally, we integrate observations from the current frame to provide more detailed information for motion prediction and achieve closed-loop control. Mathematically, it is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 255, + 286, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 255, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 64, + 255, + 286, + 269 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {t} = \\Psi ([ \\tilde {\\boldsymbol {s}} _ {t}, \\tilde {\\boldsymbol {s}} _ {t + 1} ], \\boldsymbol {i} _ {t}) \\quad \\text {f o r} \\quad t = 0, \\dots , T - 1, \\tag {9}", + "image_path": "311b3e5a66f5fe0ea11ce72edd6a1dc9d71ba78c1b6a6279463df5ba4ceabfe7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{s}}_t" + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{s}}_{t + 1}" + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "text", + "content": " are consecutive observation embeddings within " + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "inline_equation", + "content": "\\tilde{\\tau}" + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "text", + "content": " generated by diffusion model, " + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "inline_equation", + "content": "i_t" + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "text", + "content": " is the current observation and " + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 47, + 272, + 286, + 307 + ], + "type": "text", + "content": " is the inferred action." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 286, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 286, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 286, + 439 + ], + "type": "text", + "content": "As the resulting action sequence " + }, + { + "bbox": [ + 46, + 308, + 286, + 439 + ], + "type": "inline_equation", + "content": "\\{a_0, a_1, \\ldots, a_{T-1}\\}" + }, + { + "bbox": [ + 46, + 308, + 286, + 439 + ], + "type": "text", + "content": " derived from the generated states, it encapsulates the skills to execute the tasks, which allows for remarkable adaptability across multiple tasks. And when faced with a new task, we are only required to change the inverse dynamics model " + }, + { + "bbox": [ + 46, + 308, + 286, + 439 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 46, + 308, + 286, + 439 + ], + "type": "text", + "content": " specific to the new task's kinematics, with the architecture and parameters of diffusion model unchanged. Such modularity ensures the generative capabilities of SkillDiffuser are not task-specific but can be leveraged across a diverse range of tasks with varying dynamics. A schematic diagram of the low level module is shown in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 447, + 190, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 190, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 190, + 459 + ], + "type": "text", + "content": "4.4. Training the SkillDiffuser" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 464, + 286, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 286, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 286, + 524 + ], + "type": "text", + "content": "We engineered a threefold loss function for SkillDiffuser. Firstly, for the inverse dynamics model which is task-specific, we employ a behavior cloning loss [44] to train our inverse MLP emulating expert actions from observed state transitions. This loss, denoted as " + }, + { + "bbox": [ + 46, + 464, + 286, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inv}" + }, + { + "bbox": [ + 46, + 464, + 286, + 524 + ], + "type": "text", + "content": ", is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 527, + 286, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 527, + 286, + 547 + ], + "spans": [ + { + "bbox": [ + 91, + 527, + 286, + 547 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n v} = \\mathbb {E} _ {\\boldsymbol {\\tau} \\sim \\mathcal {D}} \\left[ \\left\\| \\boldsymbol {a} - \\Psi (\\tilde {\\boldsymbol {s}}, \\tilde {\\boldsymbol {s}} ^ {\\prime}, \\boldsymbol {i}) \\right\\| _ {2} ^ {2} \\right], \\tag {10}", + "image_path": "5789287a252fb41fb7c91cae9c8a4692449cee03d932151f4fb6d3a805401f6f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 550, + 209, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 209, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 209, + 561 + ], + "type": "text", + "content": "where the notations are similar to Eq. 9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 562, + 286, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 286, + 656 + ], + "type": "text", + "content": "Correspondingly, the other parts including both skill abstraction and low level execution are task-agnostic. For the high-level skill abstraction module, we apply a vector quantization (VQ) loss to refine the skill predictor. This VQ loss, " + }, + { + "bbox": [ + 46, + 562, + 286, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VQ}}" + }, + { + "bbox": [ + 46, + 562, + 286, + 656 + ], + "type": "text", + "content": ", ensures the embeddings produced by the skill predictor closely match the skill set vectors, thereby improving the interpretability and consistency of the skill representations. Inspired by VQ-VAE [45], we formulate it as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 660, + 286, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 660, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 110, + 660, + 286, + 674 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V Q}} = \\mathbb {E} _ {\\boldsymbol {\\tau}} \\left[ \\| \\mathbf {q} (\\tilde {z}) - \\tilde {z} \\| _ {2} ^ {2} \\right], \\tag {11}", + "image_path": "62f65b040c95c5c0d1efe345b69b587fc999cdc550cb8fdb8069fbfc88998ec1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 141, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 141, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 141, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 141, + 689 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 47, + 677, + 141, + 689 + ], + "type": "text", + "content": " follows Eq. 7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "Lastly, for the low-level state-only skill-conditioned diffusion execution, we incorporate a diffusion loss " + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{diff}" + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "per Eq. 6, ensuring our model's state predictions are in line with both the skill guidance and temporal dynamics observed in expert demonstrations. Here, we take " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "y(\\tau) = \\Lambda(z)" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " derived from Eq. 7 and Eq. 8." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "text", + "content": "To be noted, we train our SkillDiffuser with two optimizers, one for inverse dynamics model with " + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inv}" + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "text", + "content": " and the other for overall parameters of high-level skill abstraction and low-level planning with " + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VQ}} + \\lambda \\mathcal{L}_{diff}" + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "text", + "content": " is a loss weight. This carefully constructed loss architecture enables SkillDiffuser to excel in a multi-task environment, generalizing across tasks by simply substituting the inverse dynamics model " + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 304, + 121, + 545, + 216 + ], + "type": "text", + "content": " specific to each new task's requirements." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 217, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 325 + ], + "type": "text", + "content": "Additionally, we utilize a pre-trained distilBERT [39] as our language encoder, adopting the configuration consistent with LOReL [29] and LISA [13], while freezing its parameters to guarantee stability in language understanding. And we employ diverse settings to serve as the visual encoder to ensure fair comparison in different datasets. We elaborate the details in corresponding parts of Sec. 5. More training details are shown in Appendix G and we also provide some pseudo-code of our algorithm in Appendix B." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 338, + 386, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 386, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 386, + 350 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 358, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 417 + ], + "type": "text", + "content": "We first present a comprehensive evaluation on the LOReL Sawyer dataset, and then perform the ablation study compared on Meta-World benchmark to illustrate the efficiency of our method. Finally, we visualize the learned skills of our method both on LOReL and Meta-World MT10." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 428, + 368, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 368, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 368, + 439 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 447, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 602 + ], + "type": "text", + "content": "LOReL Sawyer Dataset [29] which is abbreviated from Language-conditioned Offline Reward learning, is composed of pseudo-expert trajectories or play data gathered from an arbitrary reinforcement learning policy, annotated with post-hoc crowd-sourced language directives. The LOReL Sawyer dataset encompasses 50k trajectories, each with 20 steps, in a simulated Sawyer robot environment. We assess our approach using the same six tasks as the original paper [29] with paraphrased instructions under five different situations (i.e. seen, unseen verb, unseen none, unseen verb + noun and human provided). This comes to a total of 77 instructions for all 6 tasks combined. More details about the dataset can be found in Appendix D.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Meta-World Dataset [51]. It is a comprehensive benchmark designed for evaluating multi-task and meta reinforcement learning algorithms. It introduces a comprehensive suite of 50 distinct robotic manipulation tasks, all located within a unified tabletop environment controlled by a simulated Sawyer arm. The Multi-Task 10 (MT10) within MetaWorld is a subset comprising ten carefully selected tasks, balanced in terms of diversity and complexity. Details of the ten tasks can be found in Appendix D.3." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16472" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 511, + 175 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 511, + 175 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 511, + 175 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 511, + 175 + ], + "type": "table", + "html": "
Task InstructionRandomLCBC [42]LCRL [20]Lang DT [13]LISA [13]SkillDiffuser
close drawer52%50%58%10%100%95 ± 3.2%
open drawer14%0%8%60%20%55 ± 13.3%
turn faucet left24%12%13%0%0%55 ± 9.3%
turn faucet right15%31%0%0%30%25 ± 4.4%
move black mug right12%73%0%20%60%18 ± 3.9%
move white mug down5%6%0%0%30%10 ± 1.7%
Average over tasks20%29%13%15%40%43 ± 1.1%
", + "image_path": "2bc08da22ecf64a0db2483c0534ac4485cc9b901e71d00695ffc6d65a82df2ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 228, + 287, + 322 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 224 + ], + "type": "text", + "content": "Table 1. Task-wise success rates on LOReL Sawyer Dataset. We show our success rates compared to random policy, language-conditioned imitation learning (LCBC), language-conditioned Q-learning (LCRL), a flat non-hierarchical Decision Transformer (Lang-DT), and LISA. The results on each dataset are calculated over 3 seeds. SkillDiffuser outperforms all other methods in terms of average performance over all tasks. Best methods and those within " + }, + { + "bbox": [ + 46, + 178, + 547, + 224 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 178, + 547, + 224 + ], + "type": "text", + "content": " of the best are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 228, + 287, + 322 + ], + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 322 + ], + "type": "table", + "html": "
Rephrasal TypeLang DTLISA [13]SkillDiffuser
seen154043.65 ±4.7
unseen noun13.3333.3336.01 ±6.3
unseen verb28.333036.70 ±9.5
unseen noun+verb6.72042.02 ±3.8
human provided26.9827.3540.16 ±2.1
Average18.0730.1439.71
", + "image_path": "7573d4234c80256f828ed4f76bfca29aa3779ab027a0c838fd454959c5e99be0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 365, + 287, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 378 + ], + "type": "text", + "content": "5.2. Evaluation Results on LOReL Sawyer Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 613 + ], + "type": "text", + "content": "Baselines. We employ random policy, language conditioned imitation learning (LCBC) [42], language conditioned Q-learning (LCRL) [20], Lang-DT (also known as Flat Baseline in [13]) and state-of-the-art skill-learning method LISA [13] as our baselines. We follow the same settings as LOReL [29] for the first three algorithms, and follow LISA [13] for the last two. The random policy serves as a baseline. And LCBC mimics offline dataset behaviors based on instructions, aligning with previous works focusing on imitation learning to achieve language conditioned behavior. In contrast, LCRL employs reinforcement learning, labeling each episode's final state with language-instructed rewards. Lang-DT plays as a non-hierarchical benchmark with a Causal Transformer [25], while LISA incorporates a dual-transformer structure, with one for skill prediction and the other for action planning. We do not compare with LOReL planner [29] as it uses MPC on a learned reward function relying on human annotations, while LISA and ours learn with trajectory data only." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "content": "Results. To ensure fair comparison, our method, SkillDiffuser, is designed to maintain a parameter count similar to baseline models. It employs the same visual and language encoder architecture as used in LISA [13] with meticulously matching of embedding dimensions and the number of heads across the layers of SkillDiffuser." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "The results are present in Tables 1 and 2, showing taskwise and rephrasal-wise success rates for LOReL, averaged" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 228, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 46, + 323, + 288, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 357 + ], + "type": "text", + "content": "Table 2. Rephrasal-wise success rates (in %) on LOReL Sawyer. Results of Lang DT, LISA and our SkillDiffuser are shown here. The standard error is calculated over 3 random seeds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 228, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 257 + ], + "type": "table", + "html": "
MethodLang DTLOReL [29]LISA [13]SkillDiffuser
Success Rate13.33 ±1.318.18 ±1.820.89 ±0.625.21 ±2.7
", + "image_path": "fb563b0ec7a1c637853d1ec3fb4c4ba0bd6de1811e2bb656b096a92499d9946e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 258, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 306, + 258, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 545, + 270 + ], + "type": "text", + "content": "Table 3. Performance on LOReL multi-step composition tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "text", + "content": "over 10 runs with a 20-step time horizon. SkillDiffuser, our approach, achieves the highest average performance in six different tasks, indicating its superior cross-task adaptability particularly when compared to similar approaches which yet are based on Decision Transformer [4], such as LISA [13]. Moreover, SkillDiffuser consistently excels in all rephrased types for LOReL test tasks, outperforming LISA by " + }, + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 304, + 274, + 545, + 395 + ], + "type": "text", + "content": " on average. This demonstrates the model's robustness in handling varied skill representations, marking a notable advancement in skill-conditioned diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 401, + 542, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 542, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 542, + 415 + ], + "type": "text", + "content": "5.3. Performance on LOReL Compositional Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 419, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 491 + ], + "type": "text", + "content": "Settings. We conduct experiments following the same settings of unseen composition tasks of LISA [13] with 12 composition instructions in Tab. 3. Detailed instructions are listed in Appendix D.2 with such an example that \"open drawer and move black mug right\". We extend the max number of episode steps from customary 20 to 40, as LISA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 493, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 554 + ], + "type": "text", + "content": "Results. We observe SkillDiffuser achieves 2x the performance of non-hierarchical baseline (i.e. w/o skill abstraction) and also improves about " + }, + { + "bbox": [ + 304, + 493, + 545, + 554 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 493, + 545, + 554 + ], + "type": "text", + "content": " over LISA, highlighting its effectiveness. MPC-based LOReL planners are unable to perform as well in open scenarios like composition tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 560, + 512, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 560, + 512, + 573 + ], + "spans": [ + { + "bbox": [ + 305, + 560, + 512, + 573 + ], + "type": "text", + "content": "5.4. Ablation Study on Meta-World Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 578, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 616 + ], + "type": "text", + "content": "Settings. We conduct experiments on Meta-World MT10 benchmark with finely annotated instructions. We also use visual observations only. Details are in Appendix D.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Baselines. We evaluate our method against three baselines, all modified from existing models. The first, Flat R3M, is adapted from R3M [30] paper's planner. As the original one utilizes the first four terms of robot proprioception, we eliminate them and make the planner focus exclusively on visual observations. The second baseline is a variant of our SkillDiffuser, lacking the high-level skill abstraction module but retaining the low-level conditional diffusion-based planner," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16473" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 282, + 133 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 133 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 133 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 133 + ], + "type": "table", + "html": "
MethodLangSkill SetDiffuserPerformance
Flat R3M [30]XXX13.3%
LISA [13]✓✓X13.8%
Lang Diffuser✓X✓16.7%
SkillDiffuser✓✓✓23.3%
", + "image_path": "5586fe43ee72ae2c1eec5b91c2b3393c9cc200069e1eb6f2deb570fbc0b91f1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 175, + 286, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 286, + 256 + ], + "type": "text", + "content": "to assess the impact of skill abstraction. This version integrates a two-layer MLP to predict options from visual and language inputs, functioning as a language-conditioned diffusion planner. The last baseline is our re-implemented LISA [13] for Meta-World Dataset to validate the efficiency of diffusion model. To ensure fairness, we use R3M as the visual encoder for all of these methods on Meta-World." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 286, + 424 + ], + "type": "text", + "content": "Ablation on Language Skill Conditioning. Table 4 indicates our Meta-World MT10 task is quite different from and much more challenging than previous tasks that use states as observations [49] or take into robot proprioception [30]. We only utilize single-frame visual input and instructions. The Flat R3M method, lacking language conditioning and skill sets can only succeed on tasks like drawer-close and reach through behavior cloning. Lang-conditioned Diffuser and modified LISA both outperform Flat R3M, suggesting the value of each corresponding module. Our SkillDiffuser, discretizing skills into a skill set, achieves a " + }, + { + "bbox": [ + 46, + 258, + 286, + 424 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 258, + 286, + 424 + ], + "type": "text", + "content": " higher performance than language-conditioned diffuser and a " + }, + { + "bbox": [ + 46, + 258, + 286, + 424 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 46, + 258, + 286, + 424 + ], + "type": "text", + "content": " higher than LISA, demonstrating the effectiveness of this combinational architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 286, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 286, + 443 + ], + "type": "text", + "content": "5.5. Ablation Study on Reusability of Learned Skills" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 617 + ], + "type": "text", + "content": "To evaluate the reusability of our learned skills, we calculate the average number of different skills used for a single instruction and the total number of instructions using each skill of LOReL Sawyer Dataset in Table 5. (With max episode step being 20, we experiment with skill horizon 10.) We observe each single instruction uses 1.55 sub-skills on average and each skill is called multiple times than the number of instructions (75 with 5 eval episodes), verifying the transferability of learnt skills. As suggested in Tab.10 of [13], except a very small horizon will hurt the performance, learning sub-skills to get refined semantics helps perform different actions at different stages. Besides, we also visualize resulting images from applying discrete skills in Appendix E.1 to further validate skills' interpretability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 263, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 263, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 263, + 635 + ], + "type": "text", + "content": "5.6. Visualization Results of Learned Skill Set" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "We show the visualization of skill set on LOReL compositional tasks here in Fig. 4 and results on original LOReL dataset in Fig. 5 and Meta-World dataset in Fig. 6 in Appendix C. The visual analysis of our SkillDiffuser's learned skills on LOReL compositional tasks reveals that out of the 20-size skill set, our method learned 11 skills (e.g. pull han" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 556, + 109 + ], + "blocks": [ + { + "bbox": [ + 47, + 136, + 286, + 169 + ], + "lines": [ + { + "bbox": [ + 47, + 136, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 286, + 169 + ], + "type": "text", + "content": "Table 4. Ablation study of language skill conditioning on Meta-World. Results of Flat R3M, Language-condition diffuser, LISA and our SkillDiffuser are shown here. All are averaged over 3 runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 556, + 109 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 556, + 109 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 556, + 109 + ], + "type": "table", + "html": "
# of learnt skills# of inst# of successUse 1 skillUse 2 skillsAverage
1737514464801.55
Freq of 17 skills30, 8, 14, 10, 5, 19, 4, 7, 20, 2, 9, 20, 25, 6, 10, 3, 18
", + "image_path": "b108d10deaa10a7a73af4a59cf357403da887acbd1a1e9cd7da0e9574a8b7f67.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 135, + 544, + 268 + ], + "blocks": [ + { + "bbox": [ + 306, + 111, + 545, + 132 + ], + "lines": [ + { + "bbox": [ + 306, + 111, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 111, + 545, + 132 + ], + "type": "text", + "content": "Table 5. Average number of different skills used for a single instruction and total number of instructions used for each skill." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 135, + 544, + 268 + ], + "lines": [ + { + "bbox": [ + 308, + 135, + 544, + 268 + ], + "spans": [ + { + "bbox": [ + 308, + 135, + 544, + 268 + ], + "type": "image", + "image_path": "c68b2ccb7870580917e050779f5398c452eb66f7b0b96b5fca037d98b5d5b06a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 269, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 334 + ], + "type": "text", + "content": "Figure 4. Visualization of skill heat map on LOReL Sawyer compositional tasks. We display the word frequency associated with a skill set of size 20 in LOReL, normalized by column. The data's sparsity and distinct highlights indicate certain language tokens are uniquely linked to specific skills. There are eleven skills learned by our method. (zoom in for best view)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 448 + ], + "type": "text", + "content": "dle [skill 0], open counter [skill 14], etc.) notably distinguished by their unique word highlights. These bright spots across eleven columns (changing from only one column at initial which corresponds to default BC) in the heatmap underscore the model's ability to identify and isolate distinct skills from visual inputs, without an explicitly defined skill library. This indicates not only a significant interpretative advancement over previous diffusion-based planning but a successful abstraction of high-level skill representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 453, + 378, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 378, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 378, + 464 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 652 + ], + "type": "text", + "content": "This paper presents SkillDiffuser, an integrated framework that enables robots to perform tasks from natural language instructions by enabling interpretable skill learning and conditional diffusion planning. It employs vector quantization to learn discrete and comprehensible skill representations directly from visual and linguistic demonstrations. Subsequently, these skills condition a diffusion model to generate state trajectories adhering to the learned skills. Through integrating hierarchical skill decomposition with conditional trajectory generation, SkillDiffuser can comprehend and execute abstract instructions for various manipulation tasks. Extensive experiments on manipulation benchmarks demonstrate state-of-the-art performance, highlighting its effectiveness for multi-step composition tasks and ability to automatically learn interpretable skills." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 658, + 408, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 408, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 408, + 670 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "This paper is partially supported by the National Key R&D Program of China No.2022ZD0161000 and the General Research Fund of Hong Kong No.17200622." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16474" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "text", + "content": "[2] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision making? In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "text", + "content": "[3] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 302 + ], + "type": "text", + "content": "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 356 + ], + "type": "text", + "content": "[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "type": "text", + "content": "[7] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 2023. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 500 + ], + "type": "text", + "content": "[8] Yan Duan, Marcin Andrychowicz, Bradly Stadie, OpenAI Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel, and Wojciech Zaremba. One-shot imitation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "text", + "content": "[9] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018.2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 581 + ], + "type": "text", + "content": "[10] William Feller. On the theory of stochastic processes, with particular reference to applications. In Selected Papers I, pages 769-798. Springer, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[11] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357-368. PMLR, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[12] Divyansh Garg, Shuvam Chakraborty, Chris Cundy, Jiaming Song, and Stefano Ermon. Iq-learn: Inverse soft-q learning for imitation. Advances in Neural Information Processing Systems, 34:4028–4039, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[13] Divyansh Garg, Skanda Vaidyanath, Kuno Kim, Jiaming Song, and Stefano Ermon. Lisa: Learning interpretable skill abstractions from language. Advances in Neural Information Processing Systems, 35:21711-21724, 2022. 2, 3, 6, 7, 8, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "type": "text", + "content": "[14] Haoran He, Chenjia Bai, Kang Xu, Zhuoran Yang, Weinan Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Diffusion model is an effective planner and data synthesizer for multitask reinforcement learning. Advances in neural information processing systems, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "text", + "content": "[15] Bernd Heidergott, editor. *Taylor Series Expansions*, pages 179–263. Springer US, Boston, MA, 2007. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[16] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "text", + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[19] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, pages 9902-9915. PMLR, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[20] Yiding Jiang, Shixiang Shane Gu, Kevin P Murphy, and Chelsea Finn. Language as an abstraction for hierarchical deep reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 377, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 410 + ], + "type": "text", + "content": "[22] Alexander C Li, Carlos Florensa, Ignasi Clavera, and Pieter Abbeel. Sub-policy adaptation for hierarchical reinforcement learning. arXiv preprint arXiv:1906.05862, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "type": "text", + "content": "[23] Zhixuan Liang, Yao Mu, Mingyu Ding, Fei Ni, Masayoshi Tomizuka, and Ping Luo. Adaptdiffuser: Diffusion models as adaptive self-evolving planners. In International Conference on Machine Learning, 2023. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 456, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 500 + ], + "type": "text", + "content": "[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297. Oakland, CA, USA, 1967. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 501, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 545 + ], + "type": "text", + "content": "[25] Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Causal transformer for estimating counterfactual outcomes. In International Conference on Machine Learning, pages 15293-15329. PMLR, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "type": "text", + "content": "[26] Diganta Misra. Mish: A self regularized non-monotonic activation function. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020, 2020. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[27] Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical reinforcement learning. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[28] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[29] Suraj Nair, Eric Mitchell, Kevin Chen, Silvio Savarese, Chelsea Finn, et al. Learning language-conditioned robot behavior from offline data and crowd-sourced annotation. In" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16475" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "Conference on Robot Learning, pages 1303-1315. PMLR, 2022. 2, 6, 7, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[30] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In Conference on Robot Learning, pages 892-909. PMLR, 2023. 2, 4, 7, 8, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[31] Fei Ni, Jianye Hao, Yao Mu, Yifu Yuan, Yan Zheng, Bin Wang, and Zhixuan Liang. Metadiffuser: Diffusion model as conditional planner for offline meta-rl. In International Conference on Machine Learning, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 238 + ], + "type": "text", + "content": "[32] Scott Niekum, Sarah Osentoski, George Konidaris, Sachin Chitta, Bhaskara Marthi, and Andrew G Barto. Learning grounded finite-state representations from unstructured demonstrations. The International Journal of Robotics Research, 34(2):131-157, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "text", + "content": "[33] Dean A Pomerleau. Efficient training of artificial neural networks for autonomous navigation. Neural computation, 3 (1):88-97, 1991. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 303 + ], + "type": "text", + "content": "[34] Martin L Puterman. Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons, 1994. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 338 + ], + "type": "text", + "content": "[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 339, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 339, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 287, + 404 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[37] Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 661-668. JMLR Workshop and Conference Proceedings, 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 525 + ], + "type": "text", + "content": "[38] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "type": "text", + "content": "[39] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In Proceedings of Thirty-third Conference on Neural Information Processing Systems, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "type": "text", + "content": "[40] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 593, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 657 + ], + "type": "text", + "content": "[41] Avi Singh, Eric Jang, Alexander Irpan, Daniel Kappler, Murtaza Dalal, Sergey Levinev, Mohi Khansari, and Chelsea Finn. Scalable multi-task imitation learning with autonomous improvement. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2167-2173. IEEE, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[42] Simon Stepputtis, Joseph Campbell, Mariano Phielipp, Stefan Lee, Chitta Baral, and Henri Ben Amor. Language-conditioned imitation learning for robot manipulation tasks. Advances in Neural Information Processing Systems, 33: 13139-13150, 2020. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 521 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[43] Alan Stuart and J Keith Ord. Kendall's advanced theory of statistics. vol. 1: Distribution theory. Kendall's advanced theory of statistics. Vol. 1: Distribution theory, 1994. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[44] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4950-4957, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "text", + "content": "[47] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[48] Danfei Xu, Suraj Nair, Yuke Zhu, Julian Gao, Animesh Garg, Li Fei-Fei, and Silvio Savarese. Neural task programming: Learning to generalize across hierarchical tasks. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 3795–3802. IEEE, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "text", + "content": "[49] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In international conference on machine learning, pages 24631-24645. PMLR, 2022. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "text", + "content": "[50] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 432, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 487 + ], + "type": "text", + "content": "[51] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 487, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 521 + ], + "type": "text", + "content": "[52] Jesse Zhang, Haonan Yu, and Wei Xu. Hierarchical reinforcement learning by discovering intrinsic options. arXiv preprint arXiv:2101.06521, 2021. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16476" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_content_list.json b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..084b031e3252994cc226e57c68abc74e4068bfdb --- /dev/null +++ b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_content_list.json @@ -0,0 +1,1712 @@ +[ + { + "type": "text", + "text": "SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery", + "text_level": 1, + "bbox": [ + 76, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Guo $^{1*}$ , Jiangwei Lao $^{1*}$ , Bo Dang $^{\\S 2*}$ , Yingying Zhang $^{1}$ , Lei Yu $^{1}$ , Lixiang Ru $^{1}$ , Liheng Zhong $^{1}$ , Ziyuan Huang $^{1}$ , Kang Wu $^{\\S 2}$ , Dingxiang Hu $^{3,1}$ , Huimei He $^{3,1}$ , Jian Wang $^{1}$ , Jingdong Chen $^{1}$ , Ming Yang $^{1\\dagger}$ , Yongjun Zhang $^{2}$ , Yansheng Li $^{2\\dagger}$ $^{1}$ Ant Group $^{2}$ Wuhan University $^{3}$ MYBank", + "bbox": [ + 86, + 202, + 901, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bangzhu.gx, wenshuo.ljw}@antgroup.com, bodang@whu.edu.cn", + "bbox": [ + 246, + 277, + 743, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior studies on Remote Sensing Foundation Model (RSFM) reveal immense potential towards a generic model for Earth Observation. Nevertheless, these works primarily focus on a single modality without temporal and geocontext modeling, hampering their capabilities for diverse tasks. In this study, we present SkySense, a generic billion-scale model, pre-trained on a curated multi-modal Remote Sensing Imagery (RSI) dataset with 21.5 million temporal sequences. SkySense incorporates a factorized multi-modal spatiotemporal encoder taking temporal sequences of optical and Synthetic Aperture Radar (SAR) data as input. This encoder is pre-trained by our proposed Multi-Granularity Contrastive Learning to learn representations across different modal and spatial granularities. To further enhance the RSI representations by the geo-context clue, we introduce Geo-Context Prototype Learning to learn region-aware prototypes upon RSI's multi-modal spatiotemporal features. To our best knowledge, SkySense is the largest Multi-Modal RSFM to date, whose modules can be flexibly combined or used individually to accommodate various tasks. It demonstrates remarkable generalization capabilities on a thorough evaluation encompassing 16 datasets over 7 tasks, from single- to multi-modal, static to temporal, and classification to localization. SkySense surpasses 18 recent RSFMs in all test scenarios. Specifically, it outperforms the latest models such as GFM, SatLas and Scale-MAE by a large margin, i.e., $2.76\\%$ , $3.67\\%$ and $3.61\\%$ on average respectively. We will release the pre-trained weights to facilitate future research and Earth Observation applications.", + "bbox": [ + 73, + 358, + 473, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 809, + 209, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remote Sensing Imagery (RSI) interpretation is crucial in understanding our common home, the Earth [16, 60], via", + "bbox": [ + 76, + 835, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7a5a415ce6e68b2561b5f704c42d5811180eb157b6de1703152cc7f3a8a8d9d.jpg", + "image_caption": [ + "Figure 1. SkySense has achieved superior performance on 16 datasets over 7 distinct tasks compared with 18 state-of-the-art RSFMs and supports a board range of EO imagery interpretations." + ], + "image_footnote": [], + "bbox": [ + 503, + 325, + 897, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quite diverse tasks [5, 13, 43, 72], e.g. crop monitoring, natural disaster management, etc. Every task may require significant dedicated efforts and resources to build a task-specific model. Recently, Foundation Model emerges as a pre-trained generic model that excels in a wide range of downstream tasks [70, 76]. Hence, there is a soaring interest in exploring a comprehensive Remote Sensing Foundation Model (RSFM) for many Earth Observation (EO) tasks.", + "bbox": [ + 496, + 642, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key question naturally arises: What is essential for a RSFM? First of all, an ideal RSFM should possess the ability to perceive multi-modal temporal RSI. EO heavily relies on multi-modal time series of remote sensing data, including temporal optical and Synthetic Aperture Radar (SAR) data. Individual modality offers unique advantages and complements to each other. For example, optical images provide rich spectral bands and texture details but are susceptible to weather [77]. In contrast, SAR sensors cap", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equally contributing first authors. † Corresponding authors. §Work done during the internship of the author at Ant Group.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27672", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/b4d8a304f15428aa642fb9359edecc34d09fb75f29cbd7f46eed4467e8560e01.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDifferent EO Interpretation Input Types
Single-Modal O(RGB)Single-Modal O(Ms)Multi-Modal Static O & SARMulti-Modal Temporal O & SAR
SkySense✓✓✓✓
SatLas[4]✓✓
GFM[47]✓
Scale-MAE[50]✓
", + "bbox": [ + 78, + 88, + 470, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. SkySense supports various input types. O(RGB): Optical RGB images; O(Ms): Optical multispectral images.", + "bbox": [ + 76, + 198, + 468, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ture clear imagery in all weather conditions [31, 38]. Moreover, the time series of such data provide the crucial temporal clue to various tasks [5, 22, 73] like change prediction. Second, a RSFM should be easy to tailor when being deployed for EO tasks using different modalities (i.e., single- and multi-modal) at different spatial (i.e., pixel-, object-, and image-level) granularities. Last but not the least, remote sensing data is inherently contingent on their spacetime coordinates, which provide rich regional and seasonal geo-context that benefits RSI interpretation a lot, as indicated in [11, 24, 32, 39, 40]. Therefore, a RSFM shall bear the vital capability of effective geo-context learning and utilization.", + "bbox": [ + 75, + 247, + 468, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works on RSFM [1, 2, 4, 8, 17, 33, 45-48, 50, 55, 57, 62, 63, 65, 66] have demonstrated their preliminary success on several specific datasets. However, these RSFMs, while proficient in certain areas, are limited in their applications to EO tasks, due to factors such as single-modal pre-training and the neglect of geo-context.", + "bbox": [ + 75, + 445, + 468, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose SkySense, a billion-scale Multi-Modal Remote Sensing Foundation Model (MM-RSFM). SkySense incorporates 2.06 billion parameters and is pre-trained on a large-scale multi-modal dataset which comprises 21.5 million RSI temporal sequences extracted from high-spatial-resolution optical images (HSROIs), medium-resolution temporal multispectral imagery (TMsI) and temporal SAR imagery (TSARI). To handle the multimodal temporal RSI sequences, SkySense employs a factorized multi-modal spatiotemporal encoder to perform spatial feature extraction and multi-modal temporal fusion independently, since RSI sequence are spatially-aligned in nature. It leads to a modular design allowing flexible use of its modules, i.e., the spatial encoder can be either used alone or in combination of the fusion module to support tasks from static single-modal to temporal multi-modal. This design delivers strong modeling of RSI sequences while using substantially less parameters compared to common 3D structures [44, 75]. The factorized encoder is pre-trained by Multi-Granularity Contrastive Learning to construct features from different modal and spatial granularities. Furthermore, we propose Geo-Context Prototype Learning to generate regional prototypes from RSI features given geolocations. This approach enhances multi-modal spatiotem", + "bbox": [ + 75, + 537, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "poral representation learning by leveraging the regional context clue hidden in numerous unlabeled RSI.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SkySense has achieved the state-of-the-art (SOTA) performance across a variety of modalities and EO tasks, as shown in Fig. 1. We evaluate SkySense on a diverse set of 16 datasets [9, 14, 15, 17, 18, 22, 36, 53, 56, 59, 67, 68], where the selection covers different task types, modalities and spatial scales. The results demonstrate that Sky-Sense outperforms 18 advanced RSFMs [1, 2, 4, 8, 17, 45-48, 50, 55, 57, 62, 63, 65, 66] in all test scenarios, validating its competitive edge for a broad range of EO interpretation tasks. Tab. 1 compares our work with latest representative studies w.r.t. various input types of EO interpretation.", + "bbox": [ + 496, + 121, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our technical contributions are:", + "bbox": [ + 517, + 289, + 812, + 304 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SkySense, the largest MM-RSFM to date with a modular design, which is capable of handling diverse tasks, from single- to multi-modal, static to temporal, and classification to localization.", + "- The design of SkySense involves three novel technical components: a) A factorized multi-modal spatiotemporal encoder to effectively process multi-modal temporal RSI; b) Multi-Granularity Contrastive Learning that learns features at various levels of granularities to facilitate different tasks; c) Geo-Context Prototype Learning to extract region-aware geo-context clue to enable implicit geo-knowledge integration.", + "- We extensively compare SkySense with 18 recently published RSFMs. Our model has achieved the SOTA performance, supparssing the latest models like GFM, SatLas and Scale-MAE by over $2.5\\%$ on average. We hope the release of pre-trained weights will contribute to the Remote Sensing community and facilitate future research." + ], + "bbox": [ + 500, + 305, + 890, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 593, + 640, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Remote Sensing Foundation Model", + "text_level": 1, + "bbox": [ + 500, + 619, + 803, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent Remote Sensing Foundation Models draw their primary inspiration from the research on Vision Foundation Model [3, 7, 10, 19, 23, 25-27, 41, 49, 61]. Remote sensing data inherently integrates space-time coordinates and has diverse spatial scales. The mainstream RSFMs extend the foundation model techniques to space-time RS data, such as Contrastive Learning. For instance, GASSL [2] utilized geo-location prediction as an additional pre-text task in the MoCo-v2 framework [12]. Multiple views with different sizes were utilized by DINO-MC [66] for self-supervised learning within the DINO framework [7]. SeCo [46] and CACo [45] both proposed Contrastive Learning to perceive short-term and long-term changes by using the spatiotemporal structure of temporal RSI sequences. Besides, there are works either improving the MIM-based framework [50, 55, 62] or exploring the model scale-up [8]. For example, RingMo [55] modified MAE to adapt to the dense", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27673", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b4c7d33ef7197a5a1250fcd5e501c8dee2f25ff825e11ba32886d897ac087221.jpg", + "image_caption": [ + "Figure 2. The overview of our SkySense model architecture." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 477, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "objects in RSI. SatMAE [17] employed TMsi to enhance the performance on temporal sequences. Scale-MAE [50] built a framework with scale-aware encoder. Recent efforts such as CMID [48] and GFM [47] have commenced to explore amalgamation of CL and MIM strategies. Concurrently, CROMA [21] and DeCUR [64] investigated multimodal pre-training for single- and multi-modal tasks using static imagery. In this study, we propose a comprehensive MM-RSFM, SkySense, to fill the gap in existing RSFMs, i.e., single modality of RingMo, CACo, etc., static input of Scale-MAE, CROMA, etc., and the neglect of geo-context of SatLas, RVSA, etc.", + "bbox": [ + 75, + 369, + 472, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SkySense", + "text_level": 1, + "bbox": [ + 76, + 566, + 181, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the pre-training dataset and the design choices for individual module respectively.", + "bbox": [ + 75, + 593, + 468, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pre-training Dataset", + "text_level": 1, + "bbox": [ + 76, + 635, + 272, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We curate an extensive multi-modal remote sensing dataset with temporal sequences, containing RSI from various sources: HSROIs from WorldView-3, 4, etc. (RGB band), TMsi from Sentinel-2 (B2-8, B8A, B11-12 band) and TSARI from Sentinel-1 (VV, VH Polarization). All data is geo-spatially aligned. Strictly speaking, HSROIs and TMsi shall be categorized to the optical modality, while TSARI falls to the SAR modality. However, due to HSROIs and TMsi's significant difference in spectral band and ground sample distance, we regard HSROIs and TMsi as two distinct modalities for simplicity in this paper. The dataset comprises 21.5 million training samples, each consisting of a static HSROI with rich texture details, a TMsi containing temporal and multispectral data, a TSARI providing backscatter polarization under cloud coverage, and the metadata like geo-location and acquisition date for geo", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "context modeling. This dataset covers a great variety of scenarios across resolution, spectrum, and imaging mechanism. More details of the data are included in the supplementary materials. We construct the input for SkySense as $\\{x_{HR}, x_{Ms}, x_{SAR}\\}$ , where $x_{HR}$ represents a static HSROI; $x_{Ms}$ is a Sentinel-2 TMsi after filtering cloudy images, where we randomly select 20 images to form the sequence; and $x_{SAR}$ stands for a standard-calibrated TSARI, from which we randomly select 10 images for training.", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Model Architecture", + "text_level": 1, + "bbox": [ + 498, + 237, + 689, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Factorized Multi-Modal Spatiotemporal Encoder. The overall architecture of our method is illustrated in Fig. 2. In a multi-modal input $\\{x_{HR}, x_{Ms}, x_{SAR}\\}$ , the pixels within each RSI naturally align with the others given the same geo-location. Upon this, we propose a factorized encoder that initially extracts spatial features from each RSI independently and then fuses them to capture a multi-modal spatiotemporal representation. The design separates spatial feature extraction from the feature fusion, enabling the integration of the clues from modality, time and geo-context.", + "bbox": [ + 496, + 260, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Feature Extraction. To handle the spatially aligned sequence input $\\{x_{HR}, x_{Ms}, x_{SAR}\\}$ , we utilize the spatial encoder $g_{HR}, g_{Ms}$ and $g_{SAR}$ for each individual RSI from HSROI, TMsI and TSARI respectively. As shown in Eq. (1), the obtained feature $F_i \\in \\mathbb{R}^{h \\times w \\times T_i \\times d}$ , $i \\in \\{HR, Ms, SAR\\}$ are of the same size in spatial dimension, where $h$ and $w$ are the height and width of $F_i$ , $T_{HR}, T_{Ms}, T_{SAR}$ represent the sequence lengths of HSROI, TMsI, and TSARI respectively, and $d$ is the feature dimension. The initial multi-modal temporal feature representation $F_T \\in \\mathbb{R}^{N_S \\times N_T \\times d}$ is generated by concatenating all $F_i$ along the time dimension, where $N_S = h \\times w$ represents the feature size in the spatial dimension, and $N_T = \\sum_{i \\in \\{HR, Ms, SAR\\}} T_i$ represents the total sequence length across all modalities,", + "bbox": [ + 496, + 411, + 892, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} F _ {i} = g _ {i} \\left(x _ {i}\\right), i \\in \\{H R, M s, S A R \\}, \\\\ F _ {i} = G _ {i} + \\left[ F _ {i - 1}, F _ {i - 2}, F _ {i - 3} \\right] \\end{array} \\tag {1} \\\\ F _ {T} = \\operatorname {C o n c a t} \\left[ F _ {H R}, F _ {M s}, F _ {S A R} \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 646, + 890, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal Temporal Fusion. Next, we incorporate the date-specific temporal positional encoding $P_{DTPE}[:,\\mathbf{t},:]\\in \\mathbb{R}^{1\\times N_T\\times d}$ to $F_{T}$ through broadcasting, creating $F_{T}^{date}$ for date-aware modeling. $F_{T}^{date}$ is then concatenated with an extra token $F_{\\mathbf{e}}\\in \\mathbb{R}^{N_S\\times 1\\times d}$ [20] (see Eq. (2)),", + "bbox": [ + 496, + 693, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {T} ^ {\\text {d a t e}} = F _ {T} + P _ {D T P E} [:, \\mathbf {t}, : ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 773, + 890, + 799 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {T} ^ {c a t} = \\operatorname {C o n c a t} \\left[ F _ {\\mathbf {e}}, F _ {T} ^ {d a t e} \\right] \\in \\mathbb {R} ^ {N _ {S} \\times (1 + N _ {T}) \\times d},\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 792, + 844, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{t} \\in \\mathbb{R}^{N_T}$ is a vector containing the acquisition dates of all RSI in the current batch. $P_{DTPE} \\in \\mathbb{R}^{1 \\times 365 \\times d}$ is a learnable parameter representing different dates of a year, which is essential for tasks affected by seasons (e.g., crop recognition). $F_T^{cat}$ is then fed into the Multi-modal", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27674", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Fusion Transformer, composed of multiple Naive Transformer encoder layers. This module employs selfattention to integrate multi-modal temporal data, generating the multi-modal spatiotemporal feature $F_{\\mathbf{fus}}^{\\mathrm{mm}} \\in \\mathbb{R}^{N_S \\times 1 \\times d}$ . Attentional Geo-Context Integration. Each RSI's geographical location may reveal rich region-specific geocontext. It is valuable for RSI interpretation as indicated by [11, 24, 32, 40]. To utilize this contextual clue to enhance $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ , we employ a region-specific prototype set $\\mathcal{P} \\in \\mathbb{R}^{N_R \\times N_p \\times d}$ (shown on the right side of Fig. 2), where $N_R$ is the number of regions, $N_p$ represents the number of prototypes for each region and $d$ denotes the feature dimension. The learning procedure of $\\mathcal{P}$ will be elaborated in Sec. 3.3. Specifically, a regional prototype subset $\\mathcal{P}_r \\in \\mathbb{R}^{N_p \\times d}$ is chosen from $\\mathcal{P}$ based on the geo-location embedded with $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ . $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ is then attended to the prototypes of $\\mathcal{P}_r$ through the attention mechanism, as shown in Eq. (3). The weights, computed from Softmax $\\left(\\frac{QK^T}{\\sqrt{d}}\\right)$ , facilitate a soft selection of prototypes in accordance with their similarity to $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ . The final representation $F_{\\mathbf{fus}} \\in \\mathbb{R}^{N_S \\times 2d}$ is generated by concatenation of $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ and weighted sum of prototypes from $\\mathcal{P}_r$ along the feature dimension. The prototypes represent a set of discriminative features linked to certain semantics like water body, cropland, etc. By finding the similar ones to $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ , we provide the standard representations of certain semantics to complement $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ ,", + "bbox": [ + 75, + 90, + 468, + 493 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {\\mathbf {f u s}} = \\operatorname {C o n c a t} \\left[ F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}, \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\right], \\tag {3} \\\\ Q = F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}, K = V = \\mathcal {P} _ {r}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 500, + 468, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pre-training", + "text_level": 1, + "bbox": [ + 76, + 579, + 209, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An overview of our pre-training procedure is illustrated in Fig. 3. We build the pre-training framework on a common teacher-student structure [7], since it conducts self-supervised learning using only positive pairs, which is easily accessible given spatially aligned RSI, avoiding complicated design of negative pairs. Teacher's parameter set $\\theta^{\\prime}$ is updated through exponential moving average (EMA) [26] from student's parameter set $\\theta$ .", + "bbox": [ + 75, + 603, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Granularity Contrastive Learning. We propose Multi-Granularity Contrastive Learning for self-supervised learning on different modal and spatial granularities for diverse tasks. Given the input $\\{x_{HR}, x_{Ms}, x_{SAR}\\}$ , two sets of random augmentations are employed, generating two groups of views $\\{u_i\\}$ and $\\{v_i\\}$ , where $i \\in \\{HR, Ms, SAR\\}$ . $u_i$ and $v_i$ are subsequently fed into the spatial encoders from the student and teacher branches respectively. $g_i$ is the student's spatial encoder and $g_i'$ is the teacher's. The features are generated as in Eq. (4),", + "bbox": [ + 75, + 724, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} = g _ {i} \\left(u _ {i}\\right), F _ {i} ^ {\\prime} = g _ {i} ^ {\\prime} \\left(v _ {i}\\right) \\quad i \\in \\{H R, M s, S A R \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After applying the multi-modal temporal fusion and geocontext integration on $F_{i}$ and $F_{i}^{\\prime}$ , the final feature $F_{\\mathrm{fus}}$ and $F_{\\mathrm{fus}}^{\\prime}$ are obtained. Initially, we establish pixel-, object- and image-level contrastive learning to progressively learn coarse-to-fine spatial features for various tasks.", + "bbox": [ + 498, + 90, + 890, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each temporal slice of $F_{i}$ can be viewed as a pixel-level feature $F_{i}^{\\mathbf{pix}} \\in \\mathbb{R}^{N_{S} \\times d}$ . Pixel-level contrastive learning loss $\\mathcal{L}_{\\mathbf{pix}}$ is obtained by averaging all $\\mathcal{L}_{CL}$ across the spatial $(s)$ and temporal $(t)$ dimensions, as shown in Eq. (5). $f_{i}^{\\mathbf{pix}} \\in \\mathbb{R}^{d}$ represents a feature vector from $F_{i}^{\\mathbf{pix}}$ and $f_{i}^{\\mathbf{pix'}}$ is its correspondence at the same geo-location. $\\mathcal{L}_{CL}$ denotes the learning loss [7] between $f_{i}^{\\mathbf{pix}}$ and $f_{i}^{\\mathbf{pix'}}$ , and", + "bbox": [ + 496, + 167, + 890, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {p i x}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{N _ {S} T _ {i}} \\sum_ {s} \\sum_ {t} \\mathcal {L} _ {C L} \\left(f _ {i} ^ {\\mathbf {p i x}}, f _ {i} ^ {\\mathbf {p i x} ^ {\\prime}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 277, + 890, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$F_{i}^{\\mathbf{obj}} \\in \\mathbb{R}^{N_{C} \\times d}$ denotes object-level feature generated from unsupervised clustering on pixel-level feature vectors $f_{i}^{\\mathbf{pix}}$ in a single RSI, where $N_{C}$ is the number of clusters. The clustering employs the same Sinkhorn-Knopp algorithm [6] we apply for Geo-Context Prototype Learning, as shown later. $f_{i}^{\\mathbf{obj}} \\in \\mathbb{R}^{d}$ is the vector representing the cluster centers in $F_{i}^{\\mathbf{obj}}$ , which can be viewed as a general representation for a set of collected $f_{i}^{\\mathbf{pix}}$ . It usually corresponds to a certain ground object or semantics. The object-level contrastive learning loss is computed as Eq. (6),", + "bbox": [ + 496, + 321, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {o b j}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{N _ {C} T _ {i}} \\sum_ {s} \\sum_ {t} \\mathcal {L} _ {C L} \\left(f _ {i} ^ {\\mathbf {o b j}}, f _ {i} ^ {\\mathbf {o b j} \\prime}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 482, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$F_{i}^{\\mathrm{img}}\\in \\mathbb{R}^{d}$ corresponds to the image-level feature, which is an average pooling result from $F_{i}^{\\mathrm{pix}}$ . Image-level contrastive learning loss is illustrated by Eq. (7),", + "bbox": [ + 496, + 527, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i m g}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{T _ {i}} \\sum_ {t} \\mathcal {L} _ {C L} \\left(F _ {i} ^ {\\mathrm {i m g}}, F _ {i} ^ {\\mathrm {i m g} ^ {\\prime}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 580, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The fine-grained contrastive learning loss $\\mathcal{L}_{FGCL}$ is the sum of pixel-, object- and image-level contrastive learning losses as Eq. (8). Finally we form the Multi-Granularity Contrastive Learning loss $\\mathcal{L}_{MGCL}$ in Eq. (9). The concept of multi-granularity is reflected in two aspects: space and modality. In terms of space, contrastive learning is performed at the pixel-, object-, and image-level, facilitating representation learning that encapsulates diverse spatial dimensions. Regarding modality, we conduct contrastive learning on the feature of each single modality, i.e., $F_{i}$ , and the multi-modal feature after fusion, i.e., $F_{\\mathrm{fus}}$ ,", + "bbox": [ + 496, + 623, + 890, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F G C L} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\sum_ {n \\in \\{\\mathbf {p i x}, \\mathbf {o b j}, \\mathbf {i m g} \\}} \\mathcal {L} _ {n} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 806, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {M G C L} = \\sum_ {i \\in \\{H R, M s, S A R \\}} \\mathcal {L} _ {F G C L} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) \\tag {9} \\\\ + \\mathcal {L} _ {F G C L} \\left(F _ {\\mathbf {f u s}}, F _ {\\mathbf {f u s}} ^ {\\prime}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 848, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27675", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c1f1a9749b4a465bc9af20b8080df47baee31044b61d218e87a5f095f7627bc5.jpg", + "image_caption": [ + "SkySense Pre-training", + "Figure 3. Overview of SkySense pre-training and downstream usage. SkySense employs data augmentations on the input and then feeds the augmented data into the student and teacher networks respectively. Multi-Granularity Contrastive Learning and Cross-Modal Alignment are proposed to pre-train the overall network. The region-specific prototype set $\\mathcal{P}$ is learned on the student branch and it is frozen for downstream usage. Enhancing feature with $\\mathcal{P}$ is optional. After pre-training, we adopt the parameters of the teacher branch for downstream tasks. Each pre-trained module can be used alone or combined with the others, with the chosen ones either frozen or fine-tuned." + ], + "image_footnote": [], + "bbox": [ + 91, + 101, + 686, + 318 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a7375c35a02e1959e57cc9aef9624442c7d87c33e4a1254179990279299f272.jpg", + "image_caption": [ + "Downstream Usage" + ], + "image_footnote": [], + "bbox": [ + 694, + 103, + 885, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Modal Alignment. The heterogeneity of multi-modal data poses a challenge for effective multi-modal feature fusion. We address this issue by adopting multi-modal contrastive loss $\\mathcal{L}_{MMCL}$ [35] to form the alignment loss $\\mathcal{L}_{\\text {align }}$ , as shown in Eq. (10),", + "bbox": [ + 76, + 402, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {i \\neq j} \\mathcal {L} _ {\\text {M M C L}} \\left(F _ {i}, F _ {j}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 488, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ni, j \\in \\{H R, M s, S A R \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 525, + 377, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{MMCL}$ maximizes the similarity of cross-modal features from the same geo-location, while minimizing it otherwise. Cross-modal alignment is performed on the student branch. Unsupervised Geo-Context Prototype Learning. Different regions characterize distinct geographic landscapes and seasonal dynamics [30, 32] due to disparities in topography and climate. Prior arts have shown that the enlarged context can benefit the RSI interpretations [11, 24, 32, 40]. In this work, $F_{\\mathrm{fus}}^{\\mathrm{mm}}$ captures rich spatiotemporal clues for a small area. By clustering on numerous $F_{\\mathrm{fus}}^{\\mathrm{mm}}$ , higher-level regional semantics are obtained as implicit geo-knowledge for a vast geo-spatial scope (see Fig. 5). Thus, we propose Geo-Context Prototype Learning to unsupervisedly extract regional geo-context from $F_{\\mathrm{fus}}^{\\mathrm{mm}}$ during pre-training.", + "bbox": [ + 75, + 553, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We divide the globe into $N_{R}$ regions and initialize a region-specific prototype set $\\mathcal{P} \\in \\mathbb{R}^{N_R \\times N_p \\times d}$ . Each prototype is learned from $F_{\\mathrm{fus}}^{\\mathrm{mm}}$ . We leverage the geo-location of the RSI to retrieve the regional subset $\\mathcal{P}_r \\in \\mathbb{R}^{N_p \\times d}$ from $\\mathcal{P}$ . Then, we calculate the cosine similarity matrix $\\mathbf{M} \\in \\mathbb{R}^{N_S \\times N_p}$ between $F_{\\mathrm{fus}}^{\\mathrm{mm}}$ and $\\mathcal{P}_r$ as in Eq. (11),", + "bbox": [ + 75, + 766, + 468, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\frac {F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}} \\cdot \\mathcal {P} _ {r} ^ {\\mathrm {T}}}{\\| F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}} \\| \\| \\mathcal {P} _ {r} \\|}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 863, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the Sinkhorn-Knopp algorithm [6] on $\\mathbf{M}$ to find the optimal assignment matrix $\\mathbf{S} \\in \\mathbb{R}^{N_S \\times N_P}$ between $F_{\\mathbf{fus}}^{\\mathrm{mm}}$ and the prototypes. This algorithm introduces the uniform distribution constraint to avoid trivial solution while achieving maximal similarity possible. We then use $\\mathbf{S}$ to generate an update value for current sample's corresponding $\\mathcal{P}_r$ , denoted as $\\overline{\\mathcal{P}_r}$ , as shown in Eq. (12),", + "bbox": [ + 498, + 402, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathcal {P} _ {r}}} = \\mathbf {S} ^ {\\mathrm {T}} F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 530, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, we update $\\mathcal{P}_r$ through EMA [26] as in Eq. (13), where $m\\in [0,1)$ is a momentum coefficient,", + "bbox": [ + 498, + 560, + 888, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {r} \\leftarrow m \\mathcal {P} _ {r} + (1 - m) \\overline {{\\mathcal {P} _ {r}}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 597, + 890, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each $\\mathcal{P}_r$ is updated during pre-training and used as the fixed geo-context for downstream tasks. Geo-Context Prototype Learning is only conducted on the student branch. It extracts generalized region-aware representations from numerous RSI within a consistent region, offering a complementary clue to enhance the feature of a single RSI.", + "bbox": [ + 496, + 625, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As Geo-Context Prototype Learning is incorporated without an explicit loss term, our pre-training objective is shown in Eq. (14), where $\\alpha$ and $\\beta$ are trade-off weights,", + "bbox": [ + 496, + 715, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha \\mathcal {L} _ {M G C L} + \\beta \\mathcal {L} _ {\\text {a l i g n}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 770, + 890, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 799, + 632, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 1 demonstrates SkySense's superior performance in all test scenarios. We conduct experiments on 16 datasets, covering different modalities and tasks, to ensure a comprehensive assessment. The right side of Fig. 3 shows how to apply SkySense to different tasks. Each pre-trained module is", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27676", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5cef1cb98a993408b3e4ff4b41bd58744e10a5f756aa6e8c28492220df32e88c.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Semantic segmentation results." + ], + "table_body": "
ModelPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [2]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [46]ICCV'21-57.2089.0329.4/39.8
SatMAE [17]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo† [55]TGRS'22-67.2091.27-
RVSA [62]TGRS'2234.3/44.464.49--
BFM† [8]Arxiv'23--92.12-
TOV [57]JSTARS'2332.1/37.866.2492.03-
SSL4EO [65]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [48]TGRS'2336.4/43.566.2191.86-
CACo [45]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS† [63]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [47]ICCV'2336.7/45.666.6291.85-
Scale-MAE [50]ICCV'2334.0/41.765.7791.54-
SkySense-39.7/46.570.9193.9933.1/46.2
", + "bbox": [ + 78, + 88, + 408, + 291 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff9d77ad6b0a06f60000593f1cd572a8299a8b168c358463624154a25480ef6c.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Object detection results." + ], + "table_body": "
ModelHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [2]67.4065.6548.15
SatMAE [17]70.8965.6646.55
RingMo†[55]75.90-46.21
RVSA [62]73.2271.0547.04
BFM†[8]-73.62-
TOV [57]70.1666.3349.62
SSL4EO [65]64.8261.2349.37
CMID [48]75.1166.3750.58
CACo [45]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [47]72.8467.6749.69
Scale-MAE [50]73.8166.4748.31
SkySense78.7374.2754.57
", + "bbox": [ + 416, + 90, + 648, + 290 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fd0478a527524239d8c8d83f5bf391dacec752e046f9ad0f3d3c9387f98ffb2a.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Change detection results." + ], + "table_body": "
ModelLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [2]78.1946.2613.6/16.7
SeCo [46]90.1447.6713.9/16.0
SatMAE [17]87.6552.7614.8/16.2
RingMo† [55]91.86--
RVSA [62]90.86--
SpectralGPT† [28]-54.29-
MATTER† [1]-59.37-
DINO-MC [66]-52.7014.5/15.6
SSL4EO [65]89.0535.0812.3/17.5
CMID [48]91.72--
CACo [45]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [47]91.7359.82-
Scale-MAE [50]92.07--
SkySense92.5860.0615.4/18.0
", + "bbox": [ + 653, + 88, + 897, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0907e007c7dd1305360c51222426b4b9ef6858cac758507f1467ffb222309db8.jpg", + "table_caption": [ + "Table 2. Results of semantic segmentation, object detection and change detection. $\\dagger$ means the code and weights are not released until November 11th, 2023, thus we report the metrics from the paper. - means the task is not supported or the value is unavailable in the paper." + ], + "table_footnote": [], + "table_body": "
ModelSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [2]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [46]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [17]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo†[55]96.90/98.3494.25/95.67--
RVSA [62]97.03/98.5093.93/95.69--
DINO-MC [66]--84.20/88.7560.16/83.49
TOV [57]95.16/97.0990.97/93.79--
SSL4EO [65]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [48]96.11/97.7994.05/95.53--
CACo [45]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMAT[21]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [47]95.47/97.0992.73/94.6486.30/--
Scale-MAE [50]96.44/97.5892.63/95.04--
SkySense97.68/98.6094.85/96.3288.67/92.0964.38/87.27
", + "bbox": [ + 89, + 354, + 457, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Scene classification results.", + "bbox": [ + 163, + 594, + 380, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "designed to allow for combined or individual use, with the flexibility to be either frozen or fine-tuned as needed. More details are included in the supplementary materials.", + "bbox": [ + 75, + 625, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Pre-training Implementation", + "text_level": 1, + "bbox": [ + 76, + 680, + 336, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The model is pre-trained with a batch size of 240 samples, distributed over 80 A100-80GB GPUs. For HSROIs, we apply data augmentations including multi-crop [6], Gaussian blur, solarization [23], etc. As for TMsiI and TSARI, we randomly select a fixed-sized sequence from the original one and perform random disturbances on the RSI acquisition date. We employ the huge version of the Swin Transformer (Swin-H) [41] as the spatial encoder of HSROIs, for its design efficiency in minimizing computational costs for high-resolution imagery [74]. RSI from TMsiI or TSARI is processed with corresponding ViT-L [20]. For Geo-Context Prototype Learning, we divide the globe into 4096 regions, each containing 100 prototypes.", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance on Single-Modal Tasks", + "text_level": 1, + "bbox": [ + 498, + 357, + 815, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate SkySense on 4 representative single-modal tasks. All experiments are conducted using consistent finetuning settings for fairness. The supplementary materials include implementation details, visualization results and additional experiments on frozen backbone tuning.", + "bbox": [ + 496, + 380, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Segmentation. We adopt Dyna.-Pla. [59], iSAID [67], Potsdam [52] and Dyna.-S2 [59] for the segmentation experiment. They are chosen considering factors such as spatial resolution, spectrum and category type. UperNet [69] serves as the segmentation head. For Dyna.-Pla. and Dyna.-S2 datasets, we report mIoU results on official validation and test sets. For iSAID and Potsdam, we follow the settings of [55]. As depicted in Tab. 2a, SkySense has achieved the SOTA performance on all four segmentation datasets. On average, it surpasses the previous SOTA by an impressive improvement of $1.86\\%$ .", + "bbox": [ + 496, + 455, + 892, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Horizontal & Oriented Object Detection. We employ the widely recognized DIOR [36] dataset for Horizontal Object Detection and its enhanced version DIOR-R [15], along with FAIR1M [56], for Oriented Object Detection. All datasets consist of optical RGB images. Faster RCNN [51] and Oriented RCNN [37] are used for the experiment, following the setup of [55, 62]. SkySense excels on all three datasets (Tab. 2b). Notably, we surpass the second best CMID by $3.99\\%$ mAP and have achieved the best performance on the FAIR1M v2.01 leaderboard. More importantly, our results are accomplished without using any sophisticated Oriented Detection designs [29, 37, 71].", + "bbox": [ + 496, + 622, + 893, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Change Detection. We assess SkySense's Change Detection performance on LEVIR-CD [9], OSCD [18], and Dyna.-S2 [59] datasets. For LEVIR-CD and OSCD, we follow the frameworks of [55, 66] and report the F1 metric. For Dyna.-S2, we utilize the UperNet head since the re", + "bbox": [ + 496, + 803, + 893, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://www.gaofen-challenge.com/benchmark (2023.11.17)", + "bbox": [ + 517, + 886, + 844, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27677", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82f381d2e9a4d11418ca391e0203c6e75fabb22acabb644c7fb6163893446ec1.jpg", + "image_caption": [ + "Figure 4. (a) Experiment on fine-tuning using different percentages of training data on the AID dataset. (b) The impact of S1-Ts data under varying cloud coverage conditions." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 285, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7621b30499eb0385330ac21cdabd3fe9dbc32b8c7235d68f1b767ee5e32ac6ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 88, + 464, + 209 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d69bdae501a8d19f6fe9bacc472b26057948978c28fde02d4908062ce7a9e58f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Task & DatasetData Source & Geo-ContextPrevious SOTASkySense
(a) Multi-Modal Seg: Dyna.-MM(i) Planet.45.6 [47]46.5 ↑ 0.9
(ii) Planet. with GCP-47.0
(iii) S243.5 [4]46.2 ↑ 2.7
(iv) Planet. + S2-47.3
(v) Planet. + S2 + S1-47.7
(vi) Planet. + S2 + S1 with GCP-48.2
(b) Multi-Modal Seg: PASTIS-MM(i) S2-static-73.5
(ii) S2-Ts83.4 [58]84.6 ↑ 1.2
(iii) S2-Ts + S1-Ts84.2 [22]84.8 ↑ 0.6
(iv) S2-Ts + GEP-85.8
(v) S2-Ts + S1-Ts + GEP-85.9
(c) Multi-Modal Cls: BEN-MM(i) S183.70 [64]86.25 ↑ 2.55
(ii) S2 + S189.70 [64]92.21 ↑ 2.51
", + "bbox": [ + 78, + 263, + 478, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Fine-tuning results on multi-modal tasks.", + "bbox": [ + 124, + 454, + 421, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ported semantic change segmentation (SCS) score is calculated from the segmentation results [59]. Both results from the validation and test sets of Dyna.-S2 are presented. The remarkable generalization ability of SkySense is evident in the consistent improvements shown in Tab. 2c. Unlike CACo [45], which shows proficiency mainly on the Dyna.-S2 validation set, our model excels across all datasets.", + "bbox": [ + 75, + 487, + 468, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Classification. We utilize four scene classification datasets: AID [68] and RESISC-45 [14] with static RGB images, BEN-S2 [53] with static multispectral images, and fMoW-S2 [17] with temporal multispectral images. The training ratio (TR) follows [17, 46, 55]. We use a linear classifier head for experiment. For AID and RESISC-45, we report Overall Accuracy (OA), for BEN-S2 we report mAP, and for fMoW-S2 we report both Top-1 and Top-5 Accuracy. SkySense overly outperforms competitive baselines and achieves the best results on all datasets (Tab. 3). Additionally, with limited labeled data on the AID dataset, SkySense consistently outperforms CMID, ScaleMAE, and random initialization, with a $4.17\\%$ higher OA than the second best Scale-MAE using only $1\\%$ training data (Fig. 4a). These results highlight the robustness and generalization ability of SkySense's pre-trained features.", + "bbox": [ + 75, + 594, + 470, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Performance on Multi-Modal Tasks", + "text_level": 1, + "bbox": [ + 76, + 845, + 387, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Modal Segmentation: Time-insensitive Land Cover Mapping. We employ the Dyna.-MM dataset [59]", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for fine-tuning and report mIoU on the official test set. The dataset comprises HSROIs from PlanetFusion (Planet.), multispectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). We use a simple UperNet head. As shown in Tab. 4a, SkySense achieves the best results in single-modal scenarios (i) and (iii), clearly outperforming the previous SOTA by roughly $1\\%$ mIoU. Moreover, combining all three modalities as (v) further improves the mIoU by $1.2\\%$ compared to (i). Notably, without bells and whistles, SkySense ranks No.1 on the challenging DynamicEarthNet leaderboard2.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Modal Segmentation: Time-sensitive Crop Mapping. We evaluate SkySense's fine-tuning result on the PASTIS-MM dataset, an enhanced version of PASTIS-R [22]. PASTIS-MM includes HSROIs from Google Earth Pro (GEP), TMsi from Sentinel-2 (S2-Ts), and TSARI from Sentinel-1 (S1-Ts). We use a naive FCN head [42] and report the OA from the official five-fold validation on PASTIS-MM dataset. In Tab. 4b, comparing S2-Ts and static multispectral data (S2-static), we observe a significant $11.1\\%$ OA increase, highlighting the importance of incorporating temporal clue for crop mapping.", + "bbox": [ + 496, + 261, + 892, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, both (ii) and (iii) exceed the performance of the previous SOTA, affirming the superior capabilities of SkySense. When more modalities are added as (ii), (iv), and (v), the OA increases accordingly. However, integrating Sentinel-1 data yields no substantial improvement, presumably because of the cloud-free imagery from PASTIS-MM dataset. To further investigate, we compare OA using S2-Ts data at different cloud ratios with and without Sentinel-1. Fig. 4b illustrates that the performance difference between utilizing and foregoing Sentinel-1 data becomes more pronounced with an increasing cloud ratio. Specifically, when the cloud ratio exceeds $50\\%$ , the result of using Sentinel-1 outperforms its counterpart by $13\\%$ . This highlights the importance of SAR data in situations with cloud coverage and rainfall.", + "bbox": [ + 496, + 431, + 892, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Modal Scene Classification. We utilize the BENMM [54] dataset for evaluating the multi-modal scene classification task. This dataset includes both Sentinel-1 (S1) and Sentinel-2 (S2) imagery. The evaluation protocol from DeCUR [64] is followed, and we report the mAP metric of fine-tuning with $100\\%$ training data. In Tab. 4c, both (i) and (ii) significantly outperform the previous SOTA by more than $2.5\\%$ mAP. Furthermore, the inclusion of Sentinel-2 imagery greatly enhances performance compared to using Sentinel-1 imagery alone.", + "bbox": [ + 496, + 661, + 890, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All these results show a notable gain for the tasks using multi-modal data, affirming the necessity of SkySense's multi-modal pre-training from one perspective.", + "bbox": [ + 496, + 816, + 890, + 862 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2https://codalab.lisn.upsaclay.fr/competitions/2882#results(2023.11.17)", + "bbox": [ + 514, + 886, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27678", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8cdc2207c7c369f46a059ceedb7944ba5887bc078d74902033f866e183a71449.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-trainingiSAIDfMoW-S2
mIoUTop-5 Acc
Simple Ver.68.9885.69
SkySense70.91 ↑ 1.9387.27 ↑ 1.58
", + "bbox": [ + 78, + 88, + 279, + 166 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eef9906bdb89fa3294db1ed00e928e37e87d57415790a4850c0d6ab5f8b25de1.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Pre-trainingDyna.-MM
mIoU
Baseline42.2
+ MGCL44.4 ↑ 2.2
+ MM47.0 ↑ 2.6
+ CMA47.7 ↑ 0.7
+ GCPL48.2 ↑ 0.5
", + "bbox": [ + 290, + 88, + 450, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 361, + 184, + 380, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. (a) Discussion on multi-modal pre-training effectiveness. (b) Ablation study on the pre-training design.", + "bbox": [ + 76, + 207, + 467, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussions & Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 250, + 361, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-modal Pre-training Effectiveness. In addition to confirming the effectiveness of using multi-modal data in downstream tasks, we investigate the impact of multi-modal pre-training on single-modal tasks, compared with pretraining on fewer modalities. We conduct experiments on iSAID for static HSROI segmentation and fMoW-S2 for temporal multispectral classification. Two versions of the pre-trained model are tested: a simple version pre-trained only with optical imagery (HSROIs, TMsI), and SkySense, which includes HSROIs, TMsI, and TSARI for pre-training. The rest of the settings remain consistent. The results in Tab. 5a show that SkySense consistently outperforms the simple version, suggesting that the introduction of SAR data benefits representation learning of other modalities. This may attribute to the implicit clue brought by SAR data through Cross-Modal Alignment. It provides another perspective on the necessity of SkySense's multi-modal pretraining.", + "bbox": [ + 75, + 276, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What does Geo-Context Prototype (GCP) Learn? We utilize the Dyna.-MM dataset for experiment as it contains diverse geo-locations worldwide. For the segmentation task in Tab. 4a, adding GCP in downstream tasks leads to a further gain of $0.5\\%$ mIoU compared to the strong multi-modal baseline (v). Moreover, a comparison between (i) and (ii) shows a $0.5\\%$ mIoU improvement using GCP for the single-modal task. It demonstrates GCP's consistent performance gain in single- and multi-modal scenarios.", + "bbox": [ + 76, + 549, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we visualize the learned prototypes on the Map by calculating the pre-trained feature of each pixel and assigning the most similar prototype to it. A comparison with the ESRI LandCover Map [34] reveals GCP's promising results in segmenting different areas. Moreover, GCP exhibit fine-grained advantage, as shown in the middle of Fig. 5. The prototypes learned from unsupervised clustering segment cropland within the town, which is overlooked by the LandCover Map. Notably, the visualization shares the same spatial resolution with ESRI LandCover Map.", + "bbox": [ + 75, + 686, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design of Pre-training. Tab. 5b presents the ablation study to assess our pre-training design, namely Multi-Granularity Contrastive Learning (MGCL), Multi-Modal (MM) integration, Cross-Modal Alignment (CMA) and Geo-Context", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cafeb5289967c67afa0371759e6485cae4f30adfd0b88321acbfeae07de0c87c.jpg", + "image_caption": [ + "Figure 5. Comparison between (a) ESRI LandCover Map and (b) Geo-Context Prototype. The visualization process of Geo-Context Prototype is illustrated in the upper part of this figure." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 893, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prototype Learning (GCPL). We utilize the Dyna.-MM dataset for the experiment and report mIoU metric on the official test set.", + "bbox": [ + 496, + 313, + 890, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initially, we utilize a single-modal version, using $g_{HR}$ spatial encoder and HSROIs for pre-training. The training settings are kept the same as described in Sec. 4.1. The results show that MGCL leads to a notable improvement compared to a simple baseline [7]. Then we integrate further multi-modal data (i.e., TMsiI and TSARI) into the pretraining and downstream evaluation. This effectively improves the performance on the test set to $47.0\\%$ mIoU, validating the necessity of multi-modal pre-training.", + "bbox": [ + 496, + 358, + 890, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CMA is another necessary design for SkySense's pretraining, which explicitly pulls features from different modalities together, encouraging cross-modal interactions. The results show that the incorporation of modal alignment leads to $0.7\\%$ mIoU improvement. Finally, we introduce GCPL, which learns complementary regional context clues to facilitate downstream tasks and further pushes the very strong performance to $48.2\\%$ mIoU.", + "bbox": [ + 496, + 494, + 890, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion & Future Work", + "text_level": 1, + "bbox": [ + 500, + 628, + 754, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present SkySense, a large-scale MM-RSFM for interpretation of EO imagery. SkySense allows using its modules flexibly to accommodate different scenarios and consistently outperforms other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope SkySense will inspire further research on MM-RSFM and its release may contribute to sustainable innovations thriving in the Remote Sensing community. As part of our future work, we plan to incorporate the language modality, thereby extending SkySense's applications to more EO tasks.", + "bbox": [ + 496, + 654, + 890, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 833, + 669, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was in part supported by the National Natural Science Foundation of China under Grants 42030102 and 42371321.", + "bbox": [ + 498, + 857, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27679", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 2, 6", + "[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 2, 6", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 2", + "[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinand, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 2, 6, 7", + "[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1, 2", + "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4, 5, 6", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 8", + "[8] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 2, 6", + "[9] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[11] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 2, 4, 5", + "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Mocov2: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2", + "[13] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1" + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 7", + "[15] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6", + "[16] Mingmin Chi, Antonio Plaza, Jon Atli Benediktsson, Zhongyi Sun, Jinsheng Shen, and Yangyong Zhu. Big data for remote sensing: Challenges and opportunities. Proceedings of the IEEE, 104(11):2207-2219, 2016. 1", + "[17] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 3, 6, 7", + "[18] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6", + "[21] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 3, 6", + "[22] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7", + "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 6", + "[24] Shaohua Guo, Liang Liu, Zhenye Gan, Yabiao Wang, Wuhao Zhang, Chengjie Wang, Guannan Jiang, Wei Zhang, Ran Yi, Lizhuang Ma, et al. Isdnet: Integrating shallow and deep networks for efficient ultra-high resolution segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4361-4370, 2022. 2, 4, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27680", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2", + "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5", + "[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2", + "[28] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Gamba Paolo, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral foundation model. arXiv preprint arXiv:2311.07113, 2023. 6", + "[29] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 923-932, 2022. 6", + "[30] Jingliang Hu, Lichao Mou, and Xiao Xiang Zhu. Unsupervised domain adaptation using a teacher-student network for cross-city classification of sentinel-2 images. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 43:1569–1574, 2020. 5", + "[31] Lanqing Huang, Bin Liu, Boying Li, Weiwei Guo, Henhao Yu, Zenghui Zhang, and Wenxian Yu. Opersarship: A dataset dedicated to sentinel-1 ship interpretation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(1):195–208, 2017. 2", + "[32] Xin Huang, Yihong Song, Jie Yang, Wenrui Wang, Huiqun Ren, Mengjie Dong, Yujin Feng, Haidan Yin, and Jiayi Li. Toward accurate mapping of 30-m time-series global impervious surface area (gisa). International Journal of Applied Earth Observation and Geoinformation, 109:102787, 2022. 2, 4, 5", + "[33] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Nyirjesy Gabby, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mukkavilli, Devyani Lambate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence. arXiv preprint arXiv:2310.18660, 2023. 2", + "[34] Krishna Karra, Caitlin Kontgis, Zoe Statman-Weil, Joseph C Mazzariello, Mark Mathis, and Steven P Brumby. Global land use/land cover with sentinel 2 and deep learning. In 2021 IEEE international geoscience and remote sensing symposium IGARSS, pages 4704-4707. IEEE, 2021. 8", + "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5", + "[36] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6", + "[37] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6", + "[38] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. Mcanet: A joint semantic segmentation framework of optical and sar images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106: 102638, 2022. 2", + "[39] Yansheng Li, Wei Chen, Xin Huang, Gao Zhi, Siwei Li, Tao He, and Yongjun Zhang. Mfvnet: a deep adaptive fusion network with multiple field-of-views for remote sensing image semantic segmentation. Science China Information Sciences, 2023. 2", + "[40] Yinhe Liu, Sunan Shi, Junjue Wang, and Yanfei Zhong. Seeing beyond the patch: Scale-adaptive semantic segmentation of high-resolution remote sensing imagery based on reinforcement learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16868-16878, 2023. 2, 4, 5", + "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6", + "[42] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 7", + "[43] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1", + "[44] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 75-82, 2019. 2", + "[45] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7", + "[46] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27681", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 3, 6, 7", + "[48] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 3, 6", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[50] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 2, 3, 6", + "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6", + "[52] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6", + "[53] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 7", + "[54] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7", + "[55] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7", + "[56] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6", + "[57] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learn" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 2, 6", + "[58] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10418-10428, 2023. 7", + "[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7", + "[60] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N van Rijn, Holger H Hoos, Fabio Del Frate, et al. Artificial intelligence to advance earth observation: a perspective. arXiv preprint arXiv:2305.08413, 2023. 1", + "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[62] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 2, 6", + "[63] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 2, 6", + "[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Graham, Chenying Liu, Zhitong Xiong, and Xiao Xiang Zhu. Decur: decoupling common & unique representations for multimodal self-supervision. arXiv preprint arXiv:2309.05300, 2023. 3, 7", + "[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 2, 6", + "[66] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6", + "[67] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6", + "[68] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 7" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27682", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6", + "[70] Binbin Yang, Xinchi Deng, Han Shi, Changlin Li, Gengwei Zhang, Hang Xu, Shen Zhao, Liang Lin, and Xiaodan Liang. Continual object detection via prototypical task correlation guided gating mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9255-9264, 2022. 1", + "[71] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13354-13363, 2023. 6", + "[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1", + "[73] Hankui K Zhang, David P Roy, and Dong Luo. Demonstration of large area land cover classification with a one-dimensional convolutional neural network applied to single pixel temporal metric percentiles. Remote Sensing of Environment, 295:113653, 2023. 2", + "[74] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 6", + "[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 2", + "[76] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 1", + "[77] Stefano Zorzi, Shabab Bazrafkan, Stefan Habenschuss, and Friedrich Fraundorfer. *Polyworld: Polygonal building extraction with graph neural networks in satellite images*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1848-1857, 2022. 1" + ], + "bbox": [ + 78, + 90, + 470, + 696 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "27683", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_model.json b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6c22c5c40a066c393b9886ee0b95d572b6984ed5 --- /dev/null +++ b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_model.json @@ -0,0 +1,2578 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.203, + 0.903, + 0.277 + ], + "angle": 0, + "content": "Xin Guo\\(^{1*}\\), Jiangwei Lao\\(^{1*}\\), Bo Dang\\(^{\\S 2*}\\), Yingying Zhang\\(^{1}\\), Lei Yu\\(^{1}\\), Lixiang Ru\\(^{1}\\), Liheng Zhong\\(^{1}\\), Ziyuan Huang\\(^{1}\\), Kang Wu\\(^{\\S 2}\\), Dingxiang Hu\\(^{3,1}\\), Huimei He\\(^{3,1}\\), Jian Wang\\(^{1}\\), Jingdong Chen\\(^{1}\\), Ming Yang\\(^{1\\dagger}\\), Yongjun Zhang\\(^{2}\\), Yansheng Li\\(^{2\\dagger}\\) \n\\(^{1}\\)Ant Group \\(^{2}\\)Wuhan University \\(^{3}\\)MYBank" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.278, + 0.744, + 0.293 + ], + "angle": 0, + "content": "{bangzhu.gx, wenshuo.ljw}@antgroup.com, bodang@whu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.359, + 0.474, + 0.798 + ], + "angle": 0, + "content": "Prior studies on Remote Sensing Foundation Model (RSFM) reveal immense potential towards a generic model for Earth Observation. Nevertheless, these works primarily focus on a single modality without temporal and geocontext modeling, hampering their capabilities for diverse tasks. In this study, we present SkySense, a generic billion-scale model, pre-trained on a curated multi-modal Remote Sensing Imagery (RSI) dataset with 21.5 million temporal sequences. SkySense incorporates a factorized multi-modal spatiotemporal encoder taking temporal sequences of optical and Synthetic Aperture Radar (SAR) data as input. This encoder is pre-trained by our proposed Multi-Granularity Contrastive Learning to learn representations across different modal and spatial granularities. To further enhance the RSI representations by the geo-context clue, we introduce Geo-Context Prototype Learning to learn region-aware prototypes upon RSI's multi-modal spatiotemporal features. To our best knowledge, SkySense is the largest Multi-Modal RSFM to date, whose modules can be flexibly combined or used individually to accommodate various tasks. It demonstrates remarkable generalization capabilities on a thorough evaluation encompassing 16 datasets over 7 tasks, from single- to multi-modal, static to temporal, and classification to localization. SkySense surpasses 18 recent RSFMs in all test scenarios. Specifically, it outperforms the latest models such as GFM, SatLas and Scale-MAE by a large margin, i.e., \\(2.76\\%\\), \\(3.67\\%\\) and \\(3.61\\%\\) on average respectively. We will release the pre-trained weights to facilitate future research and Earth Observation applications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.81, + 0.21, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.836, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Remote Sensing Imagery (RSI) interpretation is crucial in understanding our common home, the Earth [16, 60], via" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.327, + 0.898, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.588, + 0.895, + 0.631 + ], + "angle": 0, + "content": "Figure 1. SkySense has achieved superior performance on 16 datasets over 7 distinct tasks compared with 18 state-of-the-art RSFMs and supports a board range of EO imagery interpretations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.764 + ], + "angle": 0, + "content": "quite diverse tasks [5, 13, 43, 72], e.g. crop monitoring, natural disaster management, etc. Every task may require significant dedicated efforts and resources to build a task-specific model. Recently, Foundation Model emerges as a pre-trained generic model that excels in a wide range of downstream tasks [70, 76]. Hence, there is a soaring interest in exploring a comprehensive Remote Sensing Foundation Model (RSFM) for many Earth Observation (EO) tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "The key question naturally arises: What is essential for a RSFM? First of all, an ideal RSFM should possess the ability to perceive multi-modal temporal RSI. EO heavily relies on multi-modal time series of remote sensing data, including temporal optical and Synthetic Aperture Radar (SAR) data. Individual modality offers unique advantages and complements to each other. For example, optical images provide rich spectral bands and texture details but are susceptible to weather [77]. In contrast, SAR sensors cap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Equally contributing first authors. † Corresponding authors. §Work done during the internship of the author at Ant Group." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27672" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.472, + 0.188 + ], + "angle": 0, + "content": "
ModelDifferent EO Interpretation Input Types
Single-Modal O(RGB)Single-Modal O(Ms)Multi-Modal Static O & SARMulti-Modal Temporal O & SAR
SkySense✓✓✓✓
SatLas[4]✓✓
GFM[47]✓
Scale-MAE[50]✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.199, + 0.47, + 0.227 + ], + "angle": 0, + "content": "Table 1. SkySense supports various input types. O(RGB): Optical RGB images; O(Ms): Optical multispectral images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.469, + 0.442 + ], + "angle": 0, + "content": "ture clear imagery in all weather conditions [31, 38]. Moreover, the time series of such data provide the crucial temporal clue to various tasks [5, 22, 73] like change prediction. Second, a RSFM should be easy to tailor when being deployed for EO tasks using different modalities (i.e., single- and multi-modal) at different spatial (i.e., pixel-, object-, and image-level) granularities. Last but not the least, remote sensing data is inherently contingent on their spacetime coordinates, which provide rich regional and seasonal geo-context that benefits RSI interpretation a lot, as indicated in [11, 24, 32, 39, 40]. Therefore, a RSFM shall bear the vital capability of effective geo-context learning and utilization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Previous works on RSFM [1, 2, 4, 8, 17, 33, 45-48, 50, 55, 57, 62, 63, 65, 66] have demonstrated their preliminary success on several specific datasets. However, these RSFMs, while proficient in certain areas, are limited in their applications to EO tasks, due to factors such as single-modal pre-training and the neglect of geo-context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose SkySense, a billion-scale Multi-Modal Remote Sensing Foundation Model (MM-RSFM). SkySense incorporates 2.06 billion parameters and is pre-trained on a large-scale multi-modal dataset which comprises 21.5 million RSI temporal sequences extracted from high-spatial-resolution optical images (HSROIs), medium-resolution temporal multispectral imagery (TMsI) and temporal SAR imagery (TSARI). To handle the multimodal temporal RSI sequences, SkySense employs a factorized multi-modal spatiotemporal encoder to perform spatial feature extraction and multi-modal temporal fusion independently, since RSI sequence are spatially-aligned in nature. It leads to a modular design allowing flexible use of its modules, i.e., the spatial encoder can be either used alone or in combination of the fusion module to support tasks from static single-modal to temporal multi-modal. This design delivers strong modeling of RSI sequences while using substantially less parameters compared to common 3D structures [44, 75]. The factorized encoder is pre-trained by Multi-Granularity Contrastive Learning to construct features from different modal and spatial granularities. Furthermore, we propose Geo-Context Prototype Learning to generate regional prototypes from RSI features given geolocations. This approach enhances multi-modal spatiotem" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "poral representation learning by leveraging the regional context clue hidden in numerous unlabeled RSI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.289 + ], + "angle": 0, + "content": "SkySense has achieved the state-of-the-art (SOTA) performance across a variety of modalities and EO tasks, as shown in Fig. 1. We evaluate SkySense on a diverse set of 16 datasets [9, 14, 15, 17, 18, 22, 36, 53, 56, 59, 67, 68], where the selection covers different task types, modalities and spatial scales. The results demonstrate that Sky-Sense outperforms 18 advanced RSFMs [1, 2, 4, 8, 17, 45-48, 50, 55, 57, 62, 63, 65, 66] in all test scenarios, validating its competitive edge for a broad range of EO interpretation tasks. Tab. 1 compares our work with latest representative studies w.r.t. various input types of EO interpretation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.29, + 0.813, + 0.305 + ], + "angle": 0, + "content": "In summary, our technical contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.891, + 0.365 + ], + "angle": 0, + "content": "- We propose SkySense, the largest MM-RSFM to date with a modular design, which is capable of handling diverse tasks, from single- to multi-modal, static to temporal, and classification to localization." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.367, + 0.892, + 0.487 + ], + "angle": 0, + "content": "- The design of SkySense involves three novel technical components: a) A factorized multi-modal spatiotemporal encoder to effectively process multi-modal temporal RSI; b) Multi-Granularity Contrastive Learning that learns features at various levels of granularities to facilitate different tasks; c) Geo-Context Prototype Learning to extract region-aware geo-context clue to enable implicit geo-knowledge integration." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.488, + 0.891, + 0.577 + ], + "angle": 0, + "content": "- We extensively compare SkySense with 18 recently published RSFMs. Our model has achieved the SOTA performance, supparssing the latest models like GFM, SatLas and Scale-MAE by over \\(2.5\\%\\) on average. We hope the release of pre-trained weights will contribute to the Remote Sensing community and facilitate future research." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.306, + 0.892, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.642, + 0.61 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.805, + 0.636 + ], + "angle": 0, + "content": "2.1. Remote Sensing Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recent Remote Sensing Foundation Models draw their primary inspiration from the research on Vision Foundation Model [3, 7, 10, 19, 23, 25-27, 41, 49, 61]. Remote sensing data inherently integrates space-time coordinates and has diverse spatial scales. The mainstream RSFMs extend the foundation model techniques to space-time RS data, such as Contrastive Learning. For instance, GASSL [2] utilized geo-location prediction as an additional pre-text task in the MoCo-v2 framework [12]. Multiple views with different sizes were utilized by DINO-MC [66] for self-supervised learning within the DINO framework [7]. SeCo [46] and CACo [45] both proposed Contrastive Learning to perceive short-term and long-term changes by using the spatiotemporal structure of temporal RSI sequences. Besides, there are works either improving the MIM-based framework [50, 55, 62] or exploring the model scale-up [8]. For example, RingMo [55] modified MAE to adapt to the dense" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.478, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.343, + 0.454, + 0.358 + ], + "angle": 0, + "content": "Figure 2. The overview of our SkySense model architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.552 + ], + "angle": 0, + "content": "objects in RSI. SatMAE [17] employed TMsi to enhance the performance on temporal sequences. Scale-MAE [50] built a framework with scale-aware encoder. Recent efforts such as CMID [48] and GFM [47] have commenced to explore amalgamation of CL and MIM strategies. Concurrently, CROMA [21] and DeCUR [64] investigated multimodal pre-training for single- and multi-modal tasks using static imagery. In this study, we propose a comprehensive MM-RSFM, SkySense, to fill the gap in existing RSFMs, i.e., single modality of RingMo, CACo, etc., static input of Scale-MAE, CROMA, etc., and the neglect of geo-context of SatLas, RVSA, etc." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.182, + 0.584 + ], + "angle": 0, + "content": "3. SkySense" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.625 + ], + "angle": 0, + "content": "In this section, we introduce the pre-training dataset and the design choices for individual module respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.273, + 0.651 + ], + "angle": 0, + "content": "3.1. Pre-training Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We curate an extensive multi-modal remote sensing dataset with temporal sequences, containing RSI from various sources: HSROIs from WorldView-3, 4, etc. (RGB band), TMsi from Sentinel-2 (B2-8, B8A, B11-12 band) and TSARI from Sentinel-1 (VV, VH Polarization). All data is geo-spatially aligned. Strictly speaking, HSROIs and TMsi shall be categorized to the optical modality, while TSARI falls to the SAR modality. However, due to HSROIs and TMsi's significant difference in spectral band and ground sample distance, we regard HSROIs and TMsi as two distinct modalities for simplicity in this paper. The dataset comprises 21.5 million training samples, each consisting of a static HSROI with rich texture details, a TMsi containing temporal and multispectral data, a TSARI providing backscatter polarization under cloud coverage, and the metadata like geo-location and acquisition date for geo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "context modeling. This dataset covers a great variety of scenarios across resolution, spectrum, and imaging mechanism. More details of the data are included in the supplementary materials. We construct the input for SkySense as \\(\\{x_{HR}, x_{Ms}, x_{SAR}\\}\\), where \\(x_{HR}\\) represents a static HSROI; \\(x_{Ms}\\) is a Sentinel-2 TMsi after filtering cloudy images, where we randomly select 20 images to form the sequence; and \\(x_{SAR}\\) stands for a standard-calibrated TSARI, from which we randomly select 10 images for training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.238, + 0.69, + 0.253 + ], + "angle": 0, + "content": "3.2. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Factorized Multi-Modal Spatiotemporal Encoder. The overall architecture of our method is illustrated in Fig. 2. In a multi-modal input \\(\\{x_{HR}, x_{Ms}, x_{SAR}\\}\\), the pixels within each RSI naturally align with the others given the same geo-location. Upon this, we propose a factorized encoder that initially extracts spatial features from each RSI independently and then fuses them to capture a multi-modal spatiotemporal representation. The design separates spatial feature extraction from the feature fusion, enabling the integration of the clues from modality, time and geo-context." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.893, + 0.639 + ], + "angle": 0, + "content": "Spatial Feature Extraction. To handle the spatially aligned sequence input \\(\\{x_{HR}, x_{Ms}, x_{SAR}\\}\\), we utilize the spatial encoder \\(g_{HR}, g_{Ms}\\) and \\(g_{SAR}\\) for each individual RSI from HSROI, TMsI and TSARI respectively. As shown in Eq. (1), the obtained feature \\(F_i \\in \\mathbb{R}^{h \\times w \\times T_i \\times d}\\), \\(i \\in \\{HR, Ms, SAR\\}\\) are of the same size in spatial dimension, where \\(h\\) and \\(w\\) are the height and width of \\(F_i\\), \\(T_{HR}, T_{Ms}, T_{SAR}\\) represent the sequence lengths of HSROI, TMsI, and TSARI respectively, and \\(d\\) is the feature dimension. The initial multi-modal temporal feature representation \\(F_T \\in \\mathbb{R}^{N_S \\times N_T \\times d}\\) is generated by concatenating all \\(F_i\\) along the time dimension, where \\(N_S = h \\times w\\) represents the feature size in the spatial dimension, and \\(N_T = \\sum_{i \\in \\{HR, Ms, SAR\\}} T_i\\) represents the total sequence length across all modalities," + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.647, + 0.892, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} F _ {i} = g _ {i} \\left(x _ {i}\\right), i \\in \\{H R, M s, S A R \\}, \\\\ F _ {i} = G _ {i} + \\left[ F _ {i - 1}, F _ {i - 2}, F _ {i - 3} \\right] \\end{array} \\tag {1} \\\\ F _ {T} = \\operatorname {C o n c a t} \\left[ F _ {H R}, F _ {M s}, F _ {S A R} \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Multi-modal Temporal Fusion. Next, we incorporate the date-specific temporal positional encoding \\(P_{DTPE}[:,\\mathbf{t},:]\\in \\mathbb{R}^{1\\times N_T\\times d}\\) to \\(F_{T}\\) through broadcasting, creating \\(F_{T}^{date}\\) for date-aware modeling. \\(F_{T}^{date}\\) is then concatenated with an extra token \\(F_{\\mathbf{e}}\\in \\mathbb{R}^{N_S\\times 1\\times d}\\) [20] (see Eq. (2))," + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.774, + 0.891, + 0.8 + ], + "angle": 0, + "content": "\\[\nF _ {T} ^ {\\text {d a t e}} = F _ {T} + P _ {D T P E} [:, \\mathbf {t}, : ], \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.794, + 0.845, + 0.813 + ], + "angle": 0, + "content": "\\[\nF _ {T} ^ {c a t} = \\operatorname {C o n c a t} \\left[ F _ {\\mathbf {e}}, F _ {T} ^ {d a t e} \\right] \\in \\mathbb {R} ^ {N _ {S} \\times (1 + N _ {T}) \\times d},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{t} \\in \\mathbb{R}^{N_T}\\) is a vector containing the acquisition dates of all RSI in the current batch. \\(P_{DTPE} \\in \\mathbb{R}^{1 \\times 365 \\times d}\\) is a learnable parameter representing different dates of a year, which is essential for tasks affected by seasons (e.g., crop recognition). \\(F_T^{cat}\\) is then fed into the Multi-modal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27674" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.494 + ], + "angle": 0, + "content": "Temporal Fusion Transformer, composed of multiple Naive Transformer encoder layers. This module employs selfattention to integrate multi-modal temporal data, generating the multi-modal spatiotemporal feature \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\in \\mathbb{R}^{N_S \\times 1 \\times d} \\). Attentional Geo-Context Integration. Each RSI's geographical location may reveal rich region-specific geocontext. It is valuable for RSI interpretation as indicated by [11, 24, 32, 40]. To utilize this contextual clue to enhance \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\), we employ a region-specific prototype set \\( \\mathcal{P} \\in \\mathbb{R}^{N_R \\times N_p \\times d} \\) (shown on the right side of Fig. 2), where \\( N_R \\) is the number of regions, \\( N_p \\) represents the number of prototypes for each region and \\( d \\) denotes the feature dimension. The learning procedure of \\( \\mathcal{P} \\) will be elaborated in Sec. 3.3. Specifically, a regional prototype subset \\( \\mathcal{P}_r \\in \\mathbb{R}^{N_p \\times d} \\) is chosen from \\( \\mathcal{P} \\) based on the geo-location embedded with \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\). \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\) is then attended to the prototypes of \\( \\mathcal{P}_r \\) through the attention mechanism, as shown in Eq. (3). The weights, computed from Softmax \\( \\left(\\frac{QK^T}{\\sqrt{d}}\\right) \\), facilitate a soft selection of prototypes in accordance with their similarity to \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\). The final representation \\( F_{\\mathbf{fus}} \\in \\mathbb{R}^{N_S \\times 2d} \\) is generated by concatenation of \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\) and weighted sum of prototypes from \\( \\mathcal{P}_r \\) along the feature dimension. The prototypes represent a set of discriminative features linked to certain semantics like water body, cropland, etc. By finding the similar ones to \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\), we provide the standard representations of certain semantics to complement \\( F_{\\mathbf{fus}}^{\\mathrm{mm}} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.501, + 0.469, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {\\mathbf {f u s}} = \\operatorname {C o n c a t} \\left[ F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}, \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\right], \\tag {3} \\\\ Q = F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}, K = V = \\mathcal {P} _ {r}. \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.21, + 0.595 + ], + "angle": 0, + "content": "3.3. Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.469, + 0.724 + ], + "angle": 0, + "content": "An overview of our pre-training procedure is illustrated in Fig. 3. We build the pre-training framework on a common teacher-student structure [7], since it conducts self-supervised learning using only positive pairs, which is easily accessible given spatially aligned RSI, avoiding complicated design of negative pairs. Teacher's parameter set \\(\\theta^{\\prime}\\) is updated through exponential moving average (EMA) [26] from student's parameter set \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Multi-Granularity Contrastive Learning. We propose Multi-Granularity Contrastive Learning for self-supervised learning on different modal and spatial granularities for diverse tasks. Given the input \\(\\{x_{HR}, x_{Ms}, x_{SAR}\\}\\), two sets of random augmentations are employed, generating two groups of views \\(\\{u_i\\}\\) and \\(\\{v_i\\}\\), where \\(i \\in \\{HR, Ms, SAR\\}\\). \\(u_i\\) and \\(v_i\\) are subsequently fed into the spatial encoders from the student and teacher branches respectively. \\(g_i\\) is the student's spatial encoder and \\(g_i'\\) is the teacher's. The features are generated as in Eq. (4)," + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {i} = g _ {i} \\left(u _ {i}\\right), F _ {i} ^ {\\prime} = g _ {i} ^ {\\prime} \\left(v _ {i}\\right) \\quad i \\in \\{H R, M s, S A R \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "After applying the multi-modal temporal fusion and geocontext integration on \\( F_{i} \\) and \\( F_{i}^{\\prime} \\), the final feature \\( F_{\\mathrm{fus}} \\) and \\( F_{\\mathrm{fus}}^{\\prime} \\) are obtained. Initially, we establish pixel-, object- and image-level contrastive learning to progressively learn coarse-to-fine spatial features for various tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Each temporal slice of \\(F_{i}\\) can be viewed as a pixel-level feature \\(F_{i}^{\\mathbf{pix}} \\in \\mathbb{R}^{N_{S} \\times d}\\). Pixel-level contrastive learning loss \\(\\mathcal{L}_{\\mathbf{pix}}\\) is obtained by averaging all \\(\\mathcal{L}_{CL}\\) across the spatial \\((s)\\) and temporal \\((t)\\) dimensions, as shown in Eq. (5). \\(f_{i}^{\\mathbf{pix}} \\in \\mathbb{R}^{d}\\) represents a feature vector from \\(F_{i}^{\\mathbf{pix}}\\) and \\(f_{i}^{\\mathbf{pix'}}\\) is its correspondence at the same geo-location. \\(\\mathcal{L}_{CL}\\) denotes the learning loss [7] between \\(f_{i}^{\\mathbf{pix}}\\) and \\(f_{i}^{\\mathbf{pix'}}\\), and" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.278, + 0.892, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {p i x}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{N _ {S} T _ {i}} \\sum_ {s} \\sum_ {t} \\mathcal {L} _ {C L} \\left(f _ {i} ^ {\\mathbf {p i x}}, f _ {i} ^ {\\mathbf {p i x} ^ {\\prime}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.48 + ], + "angle": 0, + "content": "\\(F_{i}^{\\mathbf{obj}} \\in \\mathbb{R}^{N_{C} \\times d}\\) denotes object-level feature generated from unsupervised clustering on pixel-level feature vectors \\(f_{i}^{\\mathbf{pix}}\\) in a single RSI, where \\(N_{C}\\) is the number of clusters. The clustering employs the same Sinkhorn-Knopp algorithm [6] we apply for Geo-Context Prototype Learning, as shown later. \\(f_{i}^{\\mathbf{obj}} \\in \\mathbb{R}^{d}\\) is the vector representing the cluster centers in \\(F_{i}^{\\mathbf{obj}}\\), which can be viewed as a general representation for a set of collected \\(f_{i}^{\\mathbf{pix}}\\). It usually corresponds to a certain ground object or semantics. The object-level contrastive learning loss is computed as Eq. (6)," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.483, + 0.892, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {o b j}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{N _ {C} T _ {i}} \\sum_ {s} \\sum_ {t} \\mathcal {L} _ {C L} \\left(f _ {i} ^ {\\mathbf {o b j}}, f _ {i} ^ {\\mathbf {o b j} \\prime}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.577 + ], + "angle": 0, + "content": "\\(F_{i}^{\\mathrm{img}}\\in \\mathbb{R}^{d}\\) corresponds to the image-level feature, which is an average pooling result from \\(F_{i}^{\\mathrm{pix}}\\) . Image-level contrastive learning loss is illustrated by Eq. (7)," + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.582, + 0.892, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i m g}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{T _ {i}} \\sum_ {t} \\mathcal {L} _ {C L} \\left(F _ {i} ^ {\\mathrm {i m g}}, F _ {i} ^ {\\mathrm {i m g} ^ {\\prime}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.791 + ], + "angle": 0, + "content": "The fine-grained contrastive learning loss \\(\\mathcal{L}_{FGCL}\\) is the sum of pixel-, object- and image-level contrastive learning losses as Eq. (8). Finally we form the Multi-Granularity Contrastive Learning loss \\(\\mathcal{L}_{MGCL}\\) in Eq. (9). The concept of multi-granularity is reflected in two aspects: space and modality. In terms of space, contrastive learning is performed at the pixel-, object-, and image-level, facilitating representation learning that encapsulates diverse spatial dimensions. Regarding modality, we conduct contrastive learning on the feature of each single modality, i.e., \\(F_{i}\\), and the multi-modal feature after fusion, i.e., \\(F_{\\mathrm{fus}}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.808, + 0.891, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F G C L} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\sum_ {n \\in \\{\\mathbf {p i x}, \\mathbf {o b j}, \\mathbf {i m g} \\}} \\mathcal {L} _ {n} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.849, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {M G C L} = \\sum_ {i \\in \\{H R, M s, S A R \\}} \\mathcal {L} _ {F G C L} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) \\tag {9} \\\\ + \\mathcal {L} _ {F G C L} \\left(F _ {\\mathbf {f u s}}, F _ {\\mathbf {f u s}} ^ {\\prime}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27675" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.09, + 0.447, + 0.102 + ], + "angle": 0, + "content": "SkySense Pre-training" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.102, + 0.687, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.091, + 0.843, + 0.102 + ], + "angle": 0, + "content": "Downstream Usage" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.104, + 0.887, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.894, + 0.392 + ], + "angle": 0, + "content": "Figure 3. Overview of SkySense pre-training and downstream usage. SkySense employs data augmentations on the input and then feeds the augmented data into the student and teacher networks respectively. Multi-Granularity Contrastive Learning and Cross-Modal Alignment are proposed to pre-train the overall network. The region-specific prototype set \\(\\mathcal{P}\\) is learned on the student branch and it is frozen for downstream usage. Enhancing feature with \\(\\mathcal{P}\\) is optional. After pre-training, we adopt the parameters of the teacher branch for downstream tasks. Each pre-trained module can be used alone or combined with the others, with the chosen ones either frozen or fine-tuned." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.404, + 0.47, + 0.481 + ], + "angle": 0, + "content": "Cross-Modal Alignment. The heterogeneity of multi-modal data poses a challenge for effective multi-modal feature fusion. We address this issue by adopting multi-modal contrastive loss \\(\\mathcal{L}_{MMCL}\\) [35] to form the alignment loss \\(\\mathcal{L}_{\\text {align }}\\), as shown in Eq. (10)," + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.489, + 0.469, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {i \\neq j} \\mathcal {L} _ {\\text {M M C L}} \\left(F _ {i}, F _ {j}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.526, + 0.378, + 0.542 + ], + "angle": 0, + "content": "\\[\ni, j \\in \\{H R, M s, S A R \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.469, + 0.766 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{MMCL}\\) maximizes the similarity of cross-modal features from the same geo-location, while minimizing it otherwise. Cross-modal alignment is performed on the student branch. Unsupervised Geo-Context Prototype Learning. Different regions characterize distinct geographic landscapes and seasonal dynamics [30, 32] due to disparities in topography and climate. Prior arts have shown that the enlarged context can benefit the RSI interpretations [11, 24, 32, 40]. In this work, \\(F_{\\mathrm{fus}}^{\\mathrm{mm}}\\) captures rich spatiotemporal clues for a small area. By clustering on numerous \\(F_{\\mathrm{fus}}^{\\mathrm{mm}}\\), higher-level regional semantics are obtained as implicit geo-knowledge for a vast geo-spatial scope (see Fig. 5). Thus, we propose Geo-Context Prototype Learning to unsupervisedly extract regional geo-context from \\(F_{\\mathrm{fus}}^{\\mathrm{mm}}\\) during pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.469, + 0.857 + ], + "angle": 0, + "content": "We divide the globe into \\(N_{R}\\) regions and initialize a region-specific prototype set \\(\\mathcal{P} \\in \\mathbb{R}^{N_R \\times N_p \\times d}\\). Each prototype is learned from \\(F_{\\mathrm{fus}}^{\\mathrm{mm}}\\). We leverage the geo-location of the RSI to retrieve the regional subset \\(\\mathcal{P}_r \\in \\mathbb{R}^{N_p \\times d}\\) from \\(\\mathcal{P}\\). Then, we calculate the cosine similarity matrix \\(\\mathbf{M} \\in \\mathbb{R}^{N_S \\times N_p}\\) between \\(F_{\\mathrm{fus}}^{\\mathrm{mm}}\\) and \\(\\mathcal{P}_r\\) as in Eq. (11)," + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.864, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\frac {F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}} \\cdot \\mathcal {P} _ {r} ^ {\\mathrm {T}}}{\\| F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}} \\| \\| \\mathcal {P} _ {r} \\|}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.892, + 0.51 + ], + "angle": 0, + "content": "We utilize the Sinkhorn-Knopp algorithm [6] on \\(\\mathbf{M}\\) to find the optimal assignment matrix \\(\\mathbf{S} \\in \\mathbb{R}^{N_S \\times N_P}\\) between \\(F_{\\mathbf{fus}}^{\\mathrm{mm}}\\) and the prototypes. This algorithm introduces the uniform distribution constraint to avoid trivial solution while achieving maximal similarity possible. We then use \\(\\mathbf{S}\\) to generate an update value for current sample's corresponding \\(\\mathcal{P}_r\\), denoted as \\(\\overline{\\mathcal{P}_r}\\), as shown in Eq. (12)," + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.531, + 0.891, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathcal {P} _ {r}}} = \\mathbf {S} ^ {\\mathrm {T}} F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.561, + 0.89, + 0.591 + ], + "angle": 0, + "content": "Afterwards, we update \\(\\mathcal{P}_r\\) through EMA [26] as in Eq. (13), where \\(m\\in [0,1)\\) is a momentum coefficient," + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.598, + 0.891, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {r} \\leftarrow m \\mathcal {P} _ {r} + (1 - m) \\overline {{\\mathcal {P} _ {r}}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Each \\(\\mathcal{P}_r\\) is updated during pre-training and used as the fixed geo-context for downstream tasks. Geo-Context Prototype Learning is only conducted on the student branch. It extracts generalized region-aware representations from numerous RSI within a consistent region, offering a complementary clue to enhance the feature of a single RSI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.763 + ], + "angle": 0, + "content": "As Geo-Context Prototype Learning is incorporated without an explicit loss term, our pre-training objective is shown in Eq. (14), where \\(\\alpha\\) and \\(\\beta\\) are trade-off weights," + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.771, + 0.891, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha \\mathcal {L} _ {M G C L} + \\beta \\mathcal {L} _ {\\text {a l i g n}}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.633, + 0.817 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 1 demonstrates SkySense's superior performance in all test scenarios. We conduct experiments on 16 datasets, covering different modalities and tasks, to ensure a comprehensive assessment. The right side of Fig. 3 shows how to apply SkySense to different tasks. Each pre-trained module is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27676" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.41, + 0.292 + ], + "angle": 0, + "content": "
ModelPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [2]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [46]ICCV'21-57.2089.0329.4/39.8
SatMAE [17]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo† [55]TGRS'22-67.2091.27-
RVSA [62]TGRS'2234.3/44.464.49--
BFM† [8]Arxiv'23--92.12-
TOV [57]JSTARS'2332.1/37.866.2492.03-
SSL4EO [65]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [48]TGRS'2336.4/43.566.2191.86-
CACo [45]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS† [63]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [47]ICCV'2336.7/45.666.6291.85-
Scale-MAE [50]ICCV'2334.0/41.765.7791.54-
SkySense-39.7/46.570.9193.9933.1/46.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.151, + 0.295, + 0.334, + 0.307 + ], + "angle": 0, + "content": "(a) Semantic segmentation results." + }, + { + "type": "table", + "bbox": [ + 0.417, + 0.091, + 0.649, + 0.291 + ], + "angle": 0, + "content": "
ModelHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [2]67.4065.6548.15
SatMAE [17]70.8965.6646.55
RingMo†[55]75.90-46.21
RVSA [62]73.2271.0547.04
BFM†[8]-73.62-
TOV [57]70.1666.3349.62
SSL4EO [65]64.8261.2349.37
CMID [48]75.1166.3750.58
CACo [45]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [47]72.8467.6749.69
Scale-MAE [50]73.8166.4748.31
SkySense78.7374.2754.57
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.457, + 0.292, + 0.604, + 0.304 + ], + "angle": 0, + "content": "(b) Object detection results." + }, + { + "type": "table", + "bbox": [ + 0.654, + 0.089, + 0.898, + 0.293 + ], + "angle": 0, + "content": "
ModelLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [2]78.1946.2613.6/16.7
SeCo [46]90.1447.6713.9/16.0
SatMAE [17]87.6552.7614.8/16.2
RingMo† [55]91.86--
RVSA [62]90.86--
SpectralGPT† [28]-54.29-
MATTER† [1]-59.37-
DINO-MC [66]-52.7014.5/15.6
SSL4EO [65]89.0535.0812.3/17.5
CMID [48]91.72--
CACo [45]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [47]91.7359.82-
Scale-MAE [50]92.07--
SkySense92.5860.0615.4/18.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.695, + 0.295, + 0.846, + 0.307 + ], + "angle": 0, + "content": "(c) Change detection results." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.319, + 0.895, + 0.349 + ], + "angle": 0, + "content": "Table 2. Results of semantic segmentation, object detection and change detection. \\(\\dagger\\) means the code and weights are not released until November 11th, 2023, thus we report the metrics from the paper. - means the task is not supported or the value is unavailable in the paper." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.356, + 0.459, + 0.585 + ], + "angle": 0, + "content": "
ModelSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [2]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [46]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [17]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo†[55]96.90/98.3494.25/95.67--
RVSA [62]97.03/98.5093.93/95.69--
DINO-MC [66]--84.20/88.7560.16/83.49
TOV [57]95.16/97.0990.97/93.79--
SSL4EO [65]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [48]96.11/97.7994.05/95.53--
CACo [45]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMAT[21]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [47]95.47/97.0992.73/94.6486.30/--
Scale-MAE [50]96.44/97.5892.63/95.04--
SkySense97.68/98.6094.85/96.3288.67/92.0964.38/87.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.165, + 0.595, + 0.382, + 0.608 + ], + "angle": 0, + "content": "Table 3. Scene classification results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.673 + ], + "angle": 0, + "content": "designed to allow for combined or individual use, with the flexibility to be either frozen or fine-tuned as needed. More details are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.337, + 0.698 + ], + "angle": 0, + "content": "4.1. Pre-training Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The model is pre-trained with a batch size of 240 samples, distributed over 80 A100-80GB GPUs. For HSROIs, we apply data augmentations including multi-crop [6], Gaussian blur, solarization [23], etc. As for TMsiI and TSARI, we randomly select a fixed-sized sequence from the original one and perform random disturbances on the RSI acquisition date. We employ the huge version of the Swin Transformer (Swin-H) [41] as the spatial encoder of HSROIs, for its design efficiency in minimizing computational costs for high-resolution imagery [74]. RSI from TMsiI or TSARI is processed with corresponding ViT-L [20]. For Geo-Context Prototype Learning, we divide the globe into 4096 regions, each containing 100 prototypes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.358, + 0.816, + 0.374 + ], + "angle": 0, + "content": "4.2. Performance on Single-Modal Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.456 + ], + "angle": 0, + "content": "We evaluate SkySense on 4 representative single-modal tasks. All experiments are conducted using consistent finetuning settings for fairness. The supplementary materials include implementation details, visualization results and additional experiments on frozen backbone tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.622 + ], + "angle": 0, + "content": "Semantic Segmentation. We adopt Dyna.-Pla. [59], iSAID [67], Potsdam [52] and Dyna.-S2 [59] for the segmentation experiment. They are chosen considering factors such as spatial resolution, spectrum and category type. UperNet [69] serves as the segmentation head. For Dyna.-Pla. and Dyna.-S2 datasets, we report mIoU results on official validation and test sets. For iSAID and Potsdam, we follow the settings of [55]. As depicted in Tab. 2a, SkySense has achieved the SOTA performance on all four segmentation datasets. On average, it surpasses the previous SOTA by an impressive improvement of \\(1.86\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.895, + 0.803 + ], + "angle": 0, + "content": "Horizontal & Oriented Object Detection. We employ the widely recognized DIOR [36] dataset for Horizontal Object Detection and its enhanced version DIOR-R [15], along with FAIR1M [56], for Oriented Object Detection. All datasets consist of optical RGB images. Faster RCNN [51] and Oriented RCNN [37] are used for the experiment, following the setup of [55, 62]. SkySense excels on all three datasets (Tab. 2b). Notably, we surpass the second best CMID by \\(3.99\\%\\) mAP and have achieved the best performance on the FAIR1M v2.01 leaderboard. More importantly, our results are accomplished without using any sophisticated Oriented Detection designs [29, 37, 71]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.895, + 0.88 + ], + "angle": 0, + "content": "Change Detection. We assess SkySense's Change Detection performance on LEVIR-CD [9], OSCD [18], and Dyna.-S2 [59] datasets. For LEVIR-CD and OSCD, we follow the frameworks of [55, 66] and report the F1 metric. For Dyna.-S2, we utilize the UperNet head since the re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.846, + 0.902 + ], + "angle": 0, + "content": "1https://www.gaofen-challenge.com/benchmark (2023.11.17)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27677" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.287, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.465, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.214, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Figure 4. (a) Experiment on fine-tuning using different percentages of training data on the AID dataset. (b) The impact of S1-Ts data under varying cloud coverage conditions." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.264, + 0.48, + 0.446 + ], + "angle": 0, + "content": "
Task & DatasetData Source & Geo-ContextPrevious SOTASkySense
(a) Multi-Modal Seg: Dyna.-MM(i) Planet.45.6 [47]46.5 ↑ 0.9
(ii) Planet. with GCP-47.0
(iii) S243.5 [4]46.2 ↑ 2.7
(iv) Planet. + S2-47.3
(v) Planet. + S2 + S1-47.7
(vi) Planet. + S2 + S1 with GCP-48.2
(b) Multi-Modal Seg: PASTIS-MM(i) S2-static-73.5
(ii) S2-Ts83.4 [58]84.6 ↑ 1.2
(iii) S2-Ts + S1-Ts84.2 [22]84.8 ↑ 0.6
(iv) S2-Ts + GEP-85.8
(v) S2-Ts + S1-Ts + GEP-85.9
(c) Multi-Modal Cls: BEN-MM(i) S183.70 [64]86.25 ↑ 2.55
(ii) S2 + S189.70 [64]92.21 ↑ 2.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.455, + 0.422, + 0.469 + ], + "angle": 0, + "content": "Table 4. Fine-tuning results on multi-modal tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.594 + ], + "angle": 0, + "content": "ported semantic change segmentation (SCS) score is calculated from the segmentation results [59]. Both results from the validation and test sets of Dyna.-S2 are presented. The remarkable generalization ability of SkySense is evident in the consistent improvements shown in Tab. 2c. Unlike CACo [45], which shows proficiency mainly on the Dyna.-S2 validation set, our model excels across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.471, + 0.837 + ], + "angle": 0, + "content": "Scene Classification. We utilize four scene classification datasets: AID [68] and RESISC-45 [14] with static RGB images, BEN-S2 [53] with static multispectral images, and fMoW-S2 [17] with temporal multispectral images. The training ratio (TR) follows [17, 46, 55]. We use a linear classifier head for experiment. For AID and RESISC-45, we report Overall Accuracy (OA), for BEN-S2 we report mAP, and for fMoW-S2 we report both Top-1 and Top-5 Accuracy. SkySense overly outperforms competitive baselines and achieves the best results on all datasets (Tab. 3). Additionally, with limited labeled data on the AID dataset, SkySense consistently outperforms CMID, ScaleMAE, and random initialization, with a \\(4.17\\%\\) higher OA than the second best Scale-MAE using only \\(1\\%\\) training data (Fig. 4a). These results highlight the robustness and generalization ability of SkySense's pre-trained features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.388, + 0.862 + ], + "angle": 0, + "content": "4.3. Performance on Multi-Modal Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Multi-Modal Segmentation: Time-insensitive Land Cover Mapping. We employ the Dyna.-MM dataset [59]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "for fine-tuning and report mIoU on the official test set. The dataset comprises HSROIs from PlanetFusion (Planet.), multispectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). We use a simple UperNet head. As shown in Tab. 4a, SkySense achieves the best results in single-modal scenarios (i) and (iii), clearly outperforming the previous SOTA by roughly \\(1\\%\\) mIoU. Moreover, combining all three modalities as (v) further improves the mIoU by \\(1.2\\%\\) compared to (i). Notably, without bells and whistles, SkySense ranks No.1 on the challenging DynamicEarthNet leaderboard2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.893, + 0.429 + ], + "angle": 0, + "content": "Multi-Modal Segmentation: Time-sensitive Crop Mapping. We evaluate SkySense's fine-tuning result on the PASTIS-MM dataset, an enhanced version of PASTIS-R [22]. PASTIS-MM includes HSROIs from Google Earth Pro (GEP), TMsi from Sentinel-2 (S2-Ts), and TSARI from Sentinel-1 (S1-Ts). We use a naive FCN head [42] and report the OA from the official five-fold validation on PASTIS-MM dataset. In Tab. 4b, comparing S2-Ts and static multispectral data (S2-static), we observe a significant \\(11.1\\%\\) OA increase, highlighting the importance of incorporating temporal clue for crop mapping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.658 + ], + "angle": 0, + "content": "Furthermore, both (ii) and (iii) exceed the performance of the previous SOTA, affirming the superior capabilities of SkySense. When more modalities are added as (ii), (iv), and (v), the OA increases accordingly. However, integrating Sentinel-1 data yields no substantial improvement, presumably because of the cloud-free imagery from PASTIS-MM dataset. To further investigate, we compare OA using S2-Ts data at different cloud ratios with and without Sentinel-1. Fig. 4b illustrates that the performance difference between utilizing and foregoing Sentinel-1 data becomes more pronounced with an increasing cloud ratio. Specifically, when the cloud ratio exceeds \\(50\\%\\), the result of using Sentinel-1 outperforms its counterpart by \\(13\\%\\). This highlights the importance of SAR data in situations with cloud coverage and rainfall." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Multi-Modal Scene Classification. We utilize the BENMM [54] dataset for evaluating the multi-modal scene classification task. This dataset includes both Sentinel-1 (S1) and Sentinel-2 (S2) imagery. The evaluation protocol from DeCUR [64] is followed, and we report the mAP metric of fine-tuning with \\(100\\%\\) training data. In Tab. 4c, both (i) and (ii) significantly outperform the previous SOTA by more than \\(2.5\\%\\) mAP. Furthermore, the inclusion of Sentinel-2 imagery greatly enhances performance compared to using Sentinel-1 imagery alone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.863 + ], + "angle": 0, + "content": "All these results show a notable gain for the tasks using multi-modal data, affirming the necessity of SkySense's multi-modal pre-training from one perspective." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.897, + 0.902 + ], + "angle": 0, + "content": "2https://codalab.lisn.upsaclay.fr/competitions/2882#results(2023.11.17)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27678" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.28, + 0.167 + ], + "angle": 0, + "content": "
Pre-trainingiSAIDfMoW-S2
mIoUTop-5 Acc
Simple Ver.68.9885.69
SkySense70.91 ↑ 1.9387.27 ↑ 1.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.171, + 0.185, + 0.18 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.291, + 0.089, + 0.451, + 0.183 + ], + "angle": 0, + "content": "
Pre-trainingDyna.-MM
mIoU
Baseline42.2
+ MGCL44.4 ↑ 2.2
+ MM47.0 ↑ 2.6
+ CMA47.7 ↑ 0.7
+ GCPL48.2 ↑ 0.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.362, + 0.185, + 0.381, + 0.196 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.208, + 0.468, + 0.237 + ], + "angle": 0, + "content": "Table 5. (a) Discussion on multi-modal pre-training effectiveness. (b) Ablation study on the pre-training design." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.25, + 0.362, + 0.266 + ], + "angle": 0, + "content": "5. Discussions & Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.548 + ], + "angle": 0, + "content": "Multi-modal Pre-training Effectiveness. In addition to confirming the effectiveness of using multi-modal data in downstream tasks, we investigate the impact of multi-modal pre-training on single-modal tasks, compared with pretraining on fewer modalities. We conduct experiments on iSAID for static HSROI segmentation and fMoW-S2 for temporal multispectral classification. Two versions of the pre-trained model are tested: a simple version pre-trained only with optical imagery (HSROIs, TMsI), and SkySense, which includes HSROIs, TMsI, and TSARI for pre-training. The rest of the settings remain consistent. The results in Tab. 5a show that SkySense consistently outperforms the simple version, suggesting that the introduction of SAR data benefits representation learning of other modalities. This may attribute to the implicit clue brought by SAR data through Cross-Modal Alignment. It provides another perspective on the necessity of SkySense's multi-modal pretraining." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.55, + 0.469, + 0.687 + ], + "angle": 0, + "content": "What does Geo-Context Prototype (GCP) Learn? We utilize the Dyna.-MM dataset for experiment as it contains diverse geo-locations worldwide. For the segmentation task in Tab. 4a, adding GCP in downstream tasks leads to a further gain of \\(0.5\\%\\) mIoU compared to the strong multi-modal baseline (v). Moreover, a comparison between (i) and (ii) shows a \\(0.5\\%\\) mIoU improvement using GCP for the single-modal task. It demonstrates GCP's consistent performance gain in single- and multi-modal scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.839 + ], + "angle": 0, + "content": "In Fig. 5, we visualize the learned prototypes on the Map by calculating the pre-trained feature of each pixel and assigning the most similar prototype to it. A comparison with the ESRI LandCover Map [34] reveals GCP's promising results in segmenting different areas. Moreover, GCP exhibit fine-grained advantage, as shown in the middle of Fig. 5. The prototypes learned from unsupervised clustering segment cropland within the town, which is overlooked by the LandCover Map. Notably, the visualization shares the same spatial resolution with ESRI LandCover Map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Design of Pre-training. Tab. 5b presents the ablation study to assess our pre-training design, namely Multi-Granularity Contrastive Learning (MGCL), Multi-Modal (MM) integration, Cross-Modal Alignment (CMA) and Geo-Context" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.894, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.258, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Figure 5. Comparison between (a) ESRI LandCover Map and (b) Geo-Context Prototype. The visualization process of Geo-Context Prototype is illustrated in the upper part of this figure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Prototype Learning (GCPL). We utilize the Dyna.-MM dataset for the experiment and report mIoU metric on the official test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.495 + ], + "angle": 0, + "content": "Initially, we utilize a single-modal version, using \\( g_{HR} \\) spatial encoder and HSROIs for pre-training. The training settings are kept the same as described in Sec. 4.1. The results show that MGCL leads to a notable improvement compared to a simple baseline [7]. Then we integrate further multi-modal data (i.e., TMsiI and TSARI) into the pretraining and downstream evaluation. This effectively improves the performance on the test set to \\( 47.0\\% \\) mIoU, validating the necessity of multi-modal pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.616 + ], + "angle": 0, + "content": "CMA is another necessary design for SkySense's pretraining, which explicitly pulls features from different modalities together, encouraging cross-modal interactions. The results show that the incorporation of modal alignment leads to \\(0.7\\%\\) mIoU improvement. Finally, we introduce GCPL, which learns complementary regional context clues to facilitate downstream tasks and further pushes the very strong performance to \\(48.2\\%\\) mIoU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.755, + 0.645 + ], + "angle": 0, + "content": "6. Conclusion & Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.82 + ], + "angle": 0, + "content": "In this paper, we present SkySense, a large-scale MM-RSFM for interpretation of EO imagery. SkySense allows using its modules flexibly to accommodate different scenarios and consistently outperforms other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope SkySense will inspire further research on MM-RSFM and its release may contribute to sustainable innovations thriving in the Remote Sensing community. As part of our future work, we plan to incorporate the language modality, thereby extending SkySense's applications to more EO tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.834, + 0.671, + 0.851 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.858, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was in part supported by the National Natural Science Foundation of China under Grants 42030102 and 42371321." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27679" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinand, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.471, + 0.441 + ], + "angle": 0, + "content": "[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.443, + 0.471, + 0.513 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.471, + 0.584 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.471, + 0.628 + ], + "angle": 0, + "content": "[8] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[9] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.673, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.73, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[11] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.814, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Mocov2: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[14] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[15] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Mingmin Chi, Antonio Plaza, Jon Atli Benediktsson, Zhongyi Sun, Jinsheng Shen, and Yangyong Zhu. Big data for remote sensing: Challenges and opportunities. Proceedings of the IEEE, 104(11):2207-2219, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[17] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[18] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[21] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.649, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[22] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[24] Shaohua Guo, Liang Liu, Zhenye Gan, Yabiao Wang, Wuhao Zhang, Chengjie Wang, Guannan Jiang, Wei Zhang, Ran Yi, Lizhuang Ma, et al. Isdnet: Integrating shallow and deep networks for efficient ultra-high resolution segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4361-4370, 2022. 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27680" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[28] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Gamba Paolo, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral foundation model. arXiv preprint arXiv:2311.07113, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[29] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 923-932, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.483 + ], + "angle": 0, + "content": "[30] Jingliang Hu, Lichao Mou, and Xiao Xiang Zhu. Unsupervised domain adaptation using a teacher-student network for cross-city classification of sentinel-2 images. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 43:1569–1574, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[31] Lanqing Huang, Bin Liu, Boying Li, Weiwei Guo, Henhao Yu, Zenghui Zhang, and Wenxian Yu. Opersarship: A dataset dedicated to sentinel-1 ship interpretation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(1):195–208, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[32] Xin Huang, Yihong Song, Jie Yang, Wenrui Wang, Huiqun Ren, Mengjie Dong, Yujin Feng, Haidan Yin, and Jiayi Li. Toward accurate mapping of 30-m time-series global impervious surface area (gisa). International Journal of Applied Earth Observation and Geoinformation, 109:102787, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[33] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Nyirjesy Gabby, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mukkavilli, Devyani Lambate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence. arXiv preprint arXiv:2310.18660, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[34] Krishna Karra, Caitlin Kontgis, Zoe Statman-Weil, Joseph C Mazzariello, Mark Mathis, and Steven P Brumby. Global land use/land cover with sentinel 2 and deep learning. In 2021 IEEE international geoscience and remote sensing symposium IGARSS, pages 4704-4707. IEEE, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[36] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[37] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[38] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. Mcanet: A joint semantic segmentation framework of optical and sar images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106: 102638, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[39] Yansheng Li, Wei Chen, Xin Huang, Gao Zhi, Siwei Li, Tao He, and Yongjun Zhang. Mfvnet: a deep adaptive fusion network with multiple field-of-views for remote sensing image semantic segmentation. Science China Information Sciences, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[40] Yinhe Liu, Sunan Shi, Junjue Wang, and Yanfei Zhong. Seeing beyond the patch: Scale-adaptive semantic segmentation of high-resolution remote sensing imagery based on reinforcement learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16868-16878, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[42] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[43] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[44] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 75-82, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[45] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27681" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[47] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[48] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.471, + 0.416 + ], + "angle": 0, + "content": "[50] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[52] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[53] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[54] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[55] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[56] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[57] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learn" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ing. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[58] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10418-10428, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[60] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N van Rijn, Holger H Hoos, Fabio Del Frate, et al. Artificial intelligence to advance earth observation: a perspective. arXiv preprint arXiv:2305.08413, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[62] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[63] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Graham, Chenying Liu, Zhitong Xiong, and Xiao Xiang Zhu. Decur: decoupling common & unique representations for multimodal self-supervision. arXiv preprint arXiv:2309.05300, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[66] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[67] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[68] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27682" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[69] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[70] Binbin Yang, Xinchi Deng, Han Shi, Changlin Li, Gengwei Zhang, Hang Xu, Shen Zhao, Liang Lin, and Xiaodan Liang. Continual object detection via prototypical task correlation guided gating mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9255-9264, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[71] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13354-13363, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[73] Hankui K Zhang, David P Roy, and Dong Luo. Demonstration of large area land cover classification with a one-dimensional convolutional neural network applied to single pixel temporal metric percentiles. Remote Sensing of Environment, 295:113653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.47, + 0.486 + ], + "angle": 0, + "content": "[74] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.47, + 0.556 + ], + "angle": 0, + "content": "[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.47, + 0.626 + ], + "angle": 0, + "content": "[76] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.47, + 0.697 + ], + "angle": 0, + "content": "[77] Stefano Zorzi, Shabab Bazrafkan, Stefan Habenschuss, and Friedrich Fraundorfer. *Polyworld: Polygonal building extraction with graph neural networks in satellite images*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1848-1857, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27683" + } + ] +] \ No newline at end of file diff --git a/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_origin.pdf b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b55966ace91b09c8f0c75bec1ed778f51ec6058f --- /dev/null +++ b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/e566e7dc-1647-497b-a48b-fadeb9fef24f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95425fc9b0f10ae3c9b8ad9d9f6b43fce11c8c033d8b888c7f7398ffe7a01c16 +size 9042185 diff --git a/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/full.md b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dddb8801c1577c16aed071d3cdcb9a5b01973688 --- /dev/null +++ b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/full.md @@ -0,0 +1,364 @@ +# SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery + +Xin Guo $^{1*}$ , Jiangwei Lao $^{1*}$ , Bo Dang $^{\S 2*}$ , Yingying Zhang $^{1}$ , Lei Yu $^{1}$ , Lixiang Ru $^{1}$ , Liheng Zhong $^{1}$ , Ziyuan Huang $^{1}$ , Kang Wu $^{\S 2}$ , Dingxiang Hu $^{3,1}$ , Huimei He $^{3,1}$ , Jian Wang $^{1}$ , Jingdong Chen $^{1}$ , Ming Yang $^{1\dagger}$ , Yongjun Zhang $^{2}$ , Yansheng Li $^{2\dagger}$ $^{1}$ Ant Group $^{2}$ Wuhan University $^{3}$ MYBank + +{bangzhu.gx, wenshuo.ljw}@antgroup.com, bodang@whu.edu.cn + +# Abstract + +Prior studies on Remote Sensing Foundation Model (RSFM) reveal immense potential towards a generic model for Earth Observation. Nevertheless, these works primarily focus on a single modality without temporal and geocontext modeling, hampering their capabilities for diverse tasks. In this study, we present SkySense, a generic billion-scale model, pre-trained on a curated multi-modal Remote Sensing Imagery (RSI) dataset with 21.5 million temporal sequences. SkySense incorporates a factorized multi-modal spatiotemporal encoder taking temporal sequences of optical and Synthetic Aperture Radar (SAR) data as input. This encoder is pre-trained by our proposed Multi-Granularity Contrastive Learning to learn representations across different modal and spatial granularities. To further enhance the RSI representations by the geo-context clue, we introduce Geo-Context Prototype Learning to learn region-aware prototypes upon RSI's multi-modal spatiotemporal features. To our best knowledge, SkySense is the largest Multi-Modal RSFM to date, whose modules can be flexibly combined or used individually to accommodate various tasks. It demonstrates remarkable generalization capabilities on a thorough evaluation encompassing 16 datasets over 7 tasks, from single- to multi-modal, static to temporal, and classification to localization. SkySense surpasses 18 recent RSFMs in all test scenarios. Specifically, it outperforms the latest models such as GFM, SatLas and Scale-MAE by a large margin, i.e., $2.76\%$ , $3.67\%$ and $3.61\%$ on average respectively. We will release the pre-trained weights to facilitate future research and Earth Observation applications. + +# 1. Introduction + +Remote Sensing Imagery (RSI) interpretation is crucial in understanding our common home, the Earth [16, 60], via + +![](images/b7a5a415ce6e68b2561b5f704c42d5811180eb157b6de1703152cc7f3a8a8d9d.jpg) +Figure 1. SkySense has achieved superior performance on 16 datasets over 7 distinct tasks compared with 18 state-of-the-art RSFMs and supports a board range of EO imagery interpretations. + +quite diverse tasks [5, 13, 43, 72], e.g. crop monitoring, natural disaster management, etc. Every task may require significant dedicated efforts and resources to build a task-specific model. Recently, Foundation Model emerges as a pre-trained generic model that excels in a wide range of downstream tasks [70, 76]. Hence, there is a soaring interest in exploring a comprehensive Remote Sensing Foundation Model (RSFM) for many Earth Observation (EO) tasks. + +The key question naturally arises: What is essential for a RSFM? First of all, an ideal RSFM should possess the ability to perceive multi-modal temporal RSI. EO heavily relies on multi-modal time series of remote sensing data, including temporal optical and Synthetic Aperture Radar (SAR) data. Individual modality offers unique advantages and complements to each other. For example, optical images provide rich spectral bands and texture details but are susceptible to weather [77]. In contrast, SAR sensors cap + +
ModelDifferent EO Interpretation Input Types
Single-Modal O(RGB)Single-Modal O(Ms)Multi-Modal Static O & SARMulti-Modal Temporal O & SAR
SkySense✓✓✓✓
SatLas[4]✓✓
GFM[47]✓
Scale-MAE[50]✓
+ +Table 1. SkySense supports various input types. O(RGB): Optical RGB images; O(Ms): Optical multispectral images. + +ture clear imagery in all weather conditions [31, 38]. Moreover, the time series of such data provide the crucial temporal clue to various tasks [5, 22, 73] like change prediction. Second, a RSFM should be easy to tailor when being deployed for EO tasks using different modalities (i.e., single- and multi-modal) at different spatial (i.e., pixel-, object-, and image-level) granularities. Last but not the least, remote sensing data is inherently contingent on their spacetime coordinates, which provide rich regional and seasonal geo-context that benefits RSI interpretation a lot, as indicated in [11, 24, 32, 39, 40]. Therefore, a RSFM shall bear the vital capability of effective geo-context learning and utilization. + +Previous works on RSFM [1, 2, 4, 8, 17, 33, 45-48, 50, 55, 57, 62, 63, 65, 66] have demonstrated their preliminary success on several specific datasets. However, these RSFMs, while proficient in certain areas, are limited in their applications to EO tasks, due to factors such as single-modal pre-training and the neglect of geo-context. + +In this paper, we propose SkySense, a billion-scale Multi-Modal Remote Sensing Foundation Model (MM-RSFM). SkySense incorporates 2.06 billion parameters and is pre-trained on a large-scale multi-modal dataset which comprises 21.5 million RSI temporal sequences extracted from high-spatial-resolution optical images (HSROIs), medium-resolution temporal multispectral imagery (TMsI) and temporal SAR imagery (TSARI). To handle the multimodal temporal RSI sequences, SkySense employs a factorized multi-modal spatiotemporal encoder to perform spatial feature extraction and multi-modal temporal fusion independently, since RSI sequence are spatially-aligned in nature. It leads to a modular design allowing flexible use of its modules, i.e., the spatial encoder can be either used alone or in combination of the fusion module to support tasks from static single-modal to temporal multi-modal. This design delivers strong modeling of RSI sequences while using substantially less parameters compared to common 3D structures [44, 75]. The factorized encoder is pre-trained by Multi-Granularity Contrastive Learning to construct features from different modal and spatial granularities. Furthermore, we propose Geo-Context Prototype Learning to generate regional prototypes from RSI features given geolocations. This approach enhances multi-modal spatiotem + +poral representation learning by leveraging the regional context clue hidden in numerous unlabeled RSI. + +SkySense has achieved the state-of-the-art (SOTA) performance across a variety of modalities and EO tasks, as shown in Fig. 1. We evaluate SkySense on a diverse set of 16 datasets [9, 14, 15, 17, 18, 22, 36, 53, 56, 59, 67, 68], where the selection covers different task types, modalities and spatial scales. The results demonstrate that Sky-Sense outperforms 18 advanced RSFMs [1, 2, 4, 8, 17, 45-48, 50, 55, 57, 62, 63, 65, 66] in all test scenarios, validating its competitive edge for a broad range of EO interpretation tasks. Tab. 1 compares our work with latest representative studies w.r.t. various input types of EO interpretation. + +In summary, our technical contributions are: + +- We propose SkySense, the largest MM-RSFM to date with a modular design, which is capable of handling diverse tasks, from single- to multi-modal, static to temporal, and classification to localization. +- The design of SkySense involves three novel technical components: a) A factorized multi-modal spatiotemporal encoder to effectively process multi-modal temporal RSI; b) Multi-Granularity Contrastive Learning that learns features at various levels of granularities to facilitate different tasks; c) Geo-Context Prototype Learning to extract region-aware geo-context clue to enable implicit geo-knowledge integration. +- We extensively compare SkySense with 18 recently published RSFMs. Our model has achieved the SOTA performance, supparssing the latest models like GFM, SatLas and Scale-MAE by over $2.5\%$ on average. We hope the release of pre-trained weights will contribute to the Remote Sensing community and facilitate future research. + +# 2. Related Work + +# 2.1. Remote Sensing Foundation Model + +Recent Remote Sensing Foundation Models draw their primary inspiration from the research on Vision Foundation Model [3, 7, 10, 19, 23, 25-27, 41, 49, 61]. Remote sensing data inherently integrates space-time coordinates and has diverse spatial scales. The mainstream RSFMs extend the foundation model techniques to space-time RS data, such as Contrastive Learning. For instance, GASSL [2] utilized geo-location prediction as an additional pre-text task in the MoCo-v2 framework [12]. Multiple views with different sizes were utilized by DINO-MC [66] for self-supervised learning within the DINO framework [7]. SeCo [46] and CACo [45] both proposed Contrastive Learning to perceive short-term and long-term changes by using the spatiotemporal structure of temporal RSI sequences. Besides, there are works either improving the MIM-based framework [50, 55, 62] or exploring the model scale-up [8]. For example, RingMo [55] modified MAE to adapt to the dense + +![](images/b4c7d33ef7197a5a1250fcd5e501c8dee2f25ff825e11ba32886d897ac087221.jpg) +Figure 2. The overview of our SkySense model architecture. + +objects in RSI. SatMAE [17] employed TMsi to enhance the performance on temporal sequences. Scale-MAE [50] built a framework with scale-aware encoder. Recent efforts such as CMID [48] and GFM [47] have commenced to explore amalgamation of CL and MIM strategies. Concurrently, CROMA [21] and DeCUR [64] investigated multimodal pre-training for single- and multi-modal tasks using static imagery. In this study, we propose a comprehensive MM-RSFM, SkySense, to fill the gap in existing RSFMs, i.e., single modality of RingMo, CACo, etc., static input of Scale-MAE, CROMA, etc., and the neglect of geo-context of SatLas, RVSA, etc. + +# 3. SkySense + +In this section, we introduce the pre-training dataset and the design choices for individual module respectively. + +# 3.1. Pre-training Dataset + +We curate an extensive multi-modal remote sensing dataset with temporal sequences, containing RSI from various sources: HSROIs from WorldView-3, 4, etc. (RGB band), TMsi from Sentinel-2 (B2-8, B8A, B11-12 band) and TSARI from Sentinel-1 (VV, VH Polarization). All data is geo-spatially aligned. Strictly speaking, HSROIs and TMsi shall be categorized to the optical modality, while TSARI falls to the SAR modality. However, due to HSROIs and TMsi's significant difference in spectral band and ground sample distance, we regard HSROIs and TMsi as two distinct modalities for simplicity in this paper. The dataset comprises 21.5 million training samples, each consisting of a static HSROI with rich texture details, a TMsi containing temporal and multispectral data, a TSARI providing backscatter polarization under cloud coverage, and the metadata like geo-location and acquisition date for geo + +context modeling. This dataset covers a great variety of scenarios across resolution, spectrum, and imaging mechanism. More details of the data are included in the supplementary materials. We construct the input for SkySense as $\{x_{HR}, x_{Ms}, x_{SAR}\}$ , where $x_{HR}$ represents a static HSROI; $x_{Ms}$ is a Sentinel-2 TMsi after filtering cloudy images, where we randomly select 20 images to form the sequence; and $x_{SAR}$ stands for a standard-calibrated TSARI, from which we randomly select 10 images for training. + +# 3.2. Model Architecture + +Factorized Multi-Modal Spatiotemporal Encoder. The overall architecture of our method is illustrated in Fig. 2. In a multi-modal input $\{x_{HR}, x_{Ms}, x_{SAR}\}$ , the pixels within each RSI naturally align with the others given the same geo-location. Upon this, we propose a factorized encoder that initially extracts spatial features from each RSI independently and then fuses them to capture a multi-modal spatiotemporal representation. The design separates spatial feature extraction from the feature fusion, enabling the integration of the clues from modality, time and geo-context. + +Spatial Feature Extraction. To handle the spatially aligned sequence input $\{x_{HR}, x_{Ms}, x_{SAR}\}$ , we utilize the spatial encoder $g_{HR}, g_{Ms}$ and $g_{SAR}$ for each individual RSI from HSROI, TMsI and TSARI respectively. As shown in Eq. (1), the obtained feature $F_i \in \mathbb{R}^{h \times w \times T_i \times d}$ , $i \in \{HR, Ms, SAR\}$ are of the same size in spatial dimension, where $h$ and $w$ are the height and width of $F_i$ , $T_{HR}, T_{Ms}, T_{SAR}$ represent the sequence lengths of HSROI, TMsI, and TSARI respectively, and $d$ is the feature dimension. The initial multi-modal temporal feature representation $F_T \in \mathbb{R}^{N_S \times N_T \times d}$ is generated by concatenating all $F_i$ along the time dimension, where $N_S = h \times w$ represents the feature size in the spatial dimension, and $N_T = \sum_{i \in \{HR, Ms, SAR\}} T_i$ represents the total sequence length across all modalities, + +$$ +\begin{array}{l} \begin{array}{l} F _ {i} = g _ {i} \left(x _ {i}\right), i \in \{H R, M s, S A R \}, \\ F _ {i} = G _ {i} + \left[ F _ {i - 1}, F _ {i - 2}, F _ {i - 3} \right] \end{array} \tag {1} \\ F _ {T} = \operatorname {C o n c a t} \left[ F _ {H R}, F _ {M s}, F _ {S A R} \right]. \\ \end{array} +$$ + +Multi-modal Temporal Fusion. Next, we incorporate the date-specific temporal positional encoding $P_{DTPE}[:,\mathbf{t},:]\in \mathbb{R}^{1\times N_T\times d}$ to $F_{T}$ through broadcasting, creating $F_{T}^{date}$ for date-aware modeling. $F_{T}^{date}$ is then concatenated with an extra token $F_{\mathbf{e}}\in \mathbb{R}^{N_S\times 1\times d}$ [20] (see Eq. (2)), + +$$ +F _ {T} ^ {\text {d a t e}} = F _ {T} + P _ {D T P E} [:, \mathbf {t}, : ], \tag {2} +$$ + +$$ +F _ {T} ^ {c a t} = \operatorname {C o n c a t} \left[ F _ {\mathbf {e}}, F _ {T} ^ {d a t e} \right] \in \mathbb {R} ^ {N _ {S} \times (1 + N _ {T}) \times d}, +$$ + +where $\mathbf{t} \in \mathbb{R}^{N_T}$ is a vector containing the acquisition dates of all RSI in the current batch. $P_{DTPE} \in \mathbb{R}^{1 \times 365 \times d}$ is a learnable parameter representing different dates of a year, which is essential for tasks affected by seasons (e.g., crop recognition). $F_T^{cat}$ is then fed into the Multi-modal + +Temporal Fusion Transformer, composed of multiple Naive Transformer encoder layers. This module employs selfattention to integrate multi-modal temporal data, generating the multi-modal spatiotemporal feature $F_{\mathbf{fus}}^{\mathrm{mm}} \in \mathbb{R}^{N_S \times 1 \times d}$ . Attentional Geo-Context Integration. Each RSI's geographical location may reveal rich region-specific geocontext. It is valuable for RSI interpretation as indicated by [11, 24, 32, 40]. To utilize this contextual clue to enhance $F_{\mathbf{fus}}^{\mathrm{mm}}$ , we employ a region-specific prototype set $\mathcal{P} \in \mathbb{R}^{N_R \times N_p \times d}$ (shown on the right side of Fig. 2), where $N_R$ is the number of regions, $N_p$ represents the number of prototypes for each region and $d$ denotes the feature dimension. The learning procedure of $\mathcal{P}$ will be elaborated in Sec. 3.3. Specifically, a regional prototype subset $\mathcal{P}_r \in \mathbb{R}^{N_p \times d}$ is chosen from $\mathcal{P}$ based on the geo-location embedded with $F_{\mathbf{fus}}^{\mathrm{mm}}$ . $F_{\mathbf{fus}}^{\mathrm{mm}}$ is then attended to the prototypes of $\mathcal{P}_r$ through the attention mechanism, as shown in Eq. (3). The weights, computed from Softmax $\left(\frac{QK^T}{\sqrt{d}}\right)$ , facilitate a soft selection of prototypes in accordance with their similarity to $F_{\mathbf{fus}}^{\mathrm{mm}}$ . The final representation $F_{\mathbf{fus}} \in \mathbb{R}^{N_S \times 2d}$ is generated by concatenation of $F_{\mathbf{fus}}^{\mathrm{mm}}$ and weighted sum of prototypes from $\mathcal{P}_r$ along the feature dimension. The prototypes represent a set of discriminative features linked to certain semantics like water body, cropland, etc. By finding the similar ones to $F_{\mathbf{fus}}^{\mathrm{mm}}$ , we provide the standard representations of certain semantics to complement $F_{\mathbf{fus}}^{\mathrm{mm}}$ , + +$$ +\begin{array}{l} F _ {\mathbf {f u s}} = \operatorname {C o n c a t} \left[ F _ {\mathbf {f u s}} ^ {\mathrm {m m}}, \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V \right], \tag {3} \\ Q = F _ {\mathbf {f u s}} ^ {\mathrm {m m}}, K = V = \mathcal {P} _ {r}. \\ \end{array} +$$ + +# 3.3. Pre-training + +An overview of our pre-training procedure is illustrated in Fig. 3. We build the pre-training framework on a common teacher-student structure [7], since it conducts self-supervised learning using only positive pairs, which is easily accessible given spatially aligned RSI, avoiding complicated design of negative pairs. Teacher's parameter set $\theta^{\prime}$ is updated through exponential moving average (EMA) [26] from student's parameter set $\theta$ . + +Multi-Granularity Contrastive Learning. We propose Multi-Granularity Contrastive Learning for self-supervised learning on different modal and spatial granularities for diverse tasks. Given the input $\{x_{HR}, x_{Ms}, x_{SAR}\}$ , two sets of random augmentations are employed, generating two groups of views $\{u_i\}$ and $\{v_i\}$ , where $i \in \{HR, Ms, SAR\}$ . $u_i$ and $v_i$ are subsequently fed into the spatial encoders from the student and teacher branches respectively. $g_i$ is the student's spatial encoder and $g_i'$ is the teacher's. The features are generated as in Eq. (4), + +$$ +F _ {i} = g _ {i} \left(u _ {i}\right), F _ {i} ^ {\prime} = g _ {i} ^ {\prime} \left(v _ {i}\right) \quad i \in \{H R, M s, S A R \}. \tag {4} +$$ + +After applying the multi-modal temporal fusion and geocontext integration on $F_{i}$ and $F_{i}^{\prime}$ , the final feature $F_{\mathrm{fus}}$ and $F_{\mathrm{fus}}^{\prime}$ are obtained. Initially, we establish pixel-, object- and image-level contrastive learning to progressively learn coarse-to-fine spatial features for various tasks. + +Each temporal slice of $F_{i}$ can be viewed as a pixel-level feature $F_{i}^{\mathbf{pix}} \in \mathbb{R}^{N_{S} \times d}$ . Pixel-level contrastive learning loss $\mathcal{L}_{\mathbf{pix}}$ is obtained by averaging all $\mathcal{L}_{CL}$ across the spatial $(s)$ and temporal $(t)$ dimensions, as shown in Eq. (5). $f_{i}^{\mathbf{pix}} \in \mathbb{R}^{d}$ represents a feature vector from $F_{i}^{\mathbf{pix}}$ and $f_{i}^{\mathbf{pix'}}$ is its correspondence at the same geo-location. $\mathcal{L}_{CL}$ denotes the learning loss [7] between $f_{i}^{\mathbf{pix}}$ and $f_{i}^{\mathbf{pix'}}$ , and + +$$ +\mathcal {L} _ {\mathbf {p i x}} \left(F _ {i}, F _ {i} ^ {\prime}\right) = \frac {1}{N _ {S} T _ {i}} \sum_ {s} \sum_ {t} \mathcal {L} _ {C L} \left(f _ {i} ^ {\mathbf {p i x}}, f _ {i} ^ {\mathbf {p i x} ^ {\prime}}\right). \tag {5} +$$ + +$F_{i}^{\mathbf{obj}} \in \mathbb{R}^{N_{C} \times d}$ denotes object-level feature generated from unsupervised clustering on pixel-level feature vectors $f_{i}^{\mathbf{pix}}$ in a single RSI, where $N_{C}$ is the number of clusters. The clustering employs the same Sinkhorn-Knopp algorithm [6] we apply for Geo-Context Prototype Learning, as shown later. $f_{i}^{\mathbf{obj}} \in \mathbb{R}^{d}$ is the vector representing the cluster centers in $F_{i}^{\mathbf{obj}}$ , which can be viewed as a general representation for a set of collected $f_{i}^{\mathbf{pix}}$ . It usually corresponds to a certain ground object or semantics. The object-level contrastive learning loss is computed as Eq. (6), + +$$ +\mathcal {L} _ {\mathbf {o b j}} \left(F _ {i}, F _ {i} ^ {\prime}\right) = \frac {1}{N _ {C} T _ {i}} \sum_ {s} \sum_ {t} \mathcal {L} _ {C L} \left(f _ {i} ^ {\mathbf {o b j}}, f _ {i} ^ {\mathbf {o b j} \prime}\right). \tag {6} +$$ + +$F_{i}^{\mathrm{img}}\in \mathbb{R}^{d}$ corresponds to the image-level feature, which is an average pooling result from $F_{i}^{\mathrm{pix}}$ . Image-level contrastive learning loss is illustrated by Eq. (7), + +$$ +\mathcal {L} _ {\mathrm {i m g}} \left(F _ {i}, F _ {i} ^ {\prime}\right) = \frac {1}{T _ {i}} \sum_ {t} \mathcal {L} _ {C L} \left(F _ {i} ^ {\mathrm {i m g}}, F _ {i} ^ {\mathrm {i m g} ^ {\prime}}\right). \tag {7} +$$ + +The fine-grained contrastive learning loss $\mathcal{L}_{FGCL}$ is the sum of pixel-, object- and image-level contrastive learning losses as Eq. (8). Finally we form the Multi-Granularity Contrastive Learning loss $\mathcal{L}_{MGCL}$ in Eq. (9). The concept of multi-granularity is reflected in two aspects: space and modality. In terms of space, contrastive learning is performed at the pixel-, object-, and image-level, facilitating representation learning that encapsulates diverse spatial dimensions. Regarding modality, we conduct contrastive learning on the feature of each single modality, i.e., $F_{i}$ , and the multi-modal feature after fusion, i.e., $F_{\mathrm{fus}}$ , + +$$ +\mathcal {L} _ {F G C L} \left(F _ {i}, F _ {i} ^ {\prime}\right) = \sum_ {n \in \{\mathbf {p i x}, \mathbf {o b j}, \mathbf {i m g} \}} \mathcal {L} _ {n} \left(F _ {i}, F _ {i} ^ {\prime}\right), \tag {8} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {M G C L} = \sum_ {i \in \{H R, M s, S A R \}} \mathcal {L} _ {F G C L} \left(F _ {i}, F _ {i} ^ {\prime}\right) \tag {9} \\ + \mathcal {L} _ {F G C L} \left(F _ {\mathbf {f u s}}, F _ {\mathbf {f u s}} ^ {\prime}\right). \\ \end{array} +$$ + +![](images/c1f1a9749b4a465bc9af20b8080df47baee31044b61d218e87a5f095f7627bc5.jpg) +SkySense Pre-training +Figure 3. Overview of SkySense pre-training and downstream usage. SkySense employs data augmentations on the input and then feeds the augmented data into the student and teacher networks respectively. Multi-Granularity Contrastive Learning and Cross-Modal Alignment are proposed to pre-train the overall network. The region-specific prototype set $\mathcal{P}$ is learned on the student branch and it is frozen for downstream usage. Enhancing feature with $\mathcal{P}$ is optional. After pre-training, we adopt the parameters of the teacher branch for downstream tasks. Each pre-trained module can be used alone or combined with the others, with the chosen ones either frozen or fine-tuned. + +![](images/7a7375c35a02e1959e57cc9aef9624442c7d87c33e4a1254179990279299f272.jpg) +Downstream Usage + +Cross-Modal Alignment. The heterogeneity of multi-modal data poses a challenge for effective multi-modal feature fusion. We address this issue by adopting multi-modal contrastive loss $\mathcal{L}_{MMCL}$ [35] to form the alignment loss $\mathcal{L}_{\text {align }}$ , as shown in Eq. (10), + +$$ +\mathcal {L} _ {\text {a l i g n}} = \sum_ {i \neq j} \mathcal {L} _ {\text {M M C L}} \left(F _ {i}, F _ {j}\right), \tag {10} +$$ + +$$ +i, j \in \{H R, M s, S A R \}. +$$ + +$\mathcal{L}_{MMCL}$ maximizes the similarity of cross-modal features from the same geo-location, while minimizing it otherwise. Cross-modal alignment is performed on the student branch. Unsupervised Geo-Context Prototype Learning. Different regions characterize distinct geographic landscapes and seasonal dynamics [30, 32] due to disparities in topography and climate. Prior arts have shown that the enlarged context can benefit the RSI interpretations [11, 24, 32, 40]. In this work, $F_{\mathrm{fus}}^{\mathrm{mm}}$ captures rich spatiotemporal clues for a small area. By clustering on numerous $F_{\mathrm{fus}}^{\mathrm{mm}}$ , higher-level regional semantics are obtained as implicit geo-knowledge for a vast geo-spatial scope (see Fig. 5). Thus, we propose Geo-Context Prototype Learning to unsupervisedly extract regional geo-context from $F_{\mathrm{fus}}^{\mathrm{mm}}$ during pre-training. + +We divide the globe into $N_{R}$ regions and initialize a region-specific prototype set $\mathcal{P} \in \mathbb{R}^{N_R \times N_p \times d}$ . Each prototype is learned from $F_{\mathrm{fus}}^{\mathrm{mm}}$ . We leverage the geo-location of the RSI to retrieve the regional subset $\mathcal{P}_r \in \mathbb{R}^{N_p \times d}$ from $\mathcal{P}$ . Then, we calculate the cosine similarity matrix $\mathbf{M} \in \mathbb{R}^{N_S \times N_p}$ between $F_{\mathrm{fus}}^{\mathrm{mm}}$ and $\mathcal{P}_r$ as in Eq. (11), + +$$ +\mathbf {M} = \frac {F _ {\mathbf {f u s}} ^ {\mathrm {m m}} \cdot \mathcal {P} _ {r} ^ {\mathrm {T}}}{\| F _ {\mathbf {f u s}} ^ {\mathrm {m m}} \| \| \mathcal {P} _ {r} \|}, \tag {11} +$$ + +We utilize the Sinkhorn-Knopp algorithm [6] on $\mathbf{M}$ to find the optimal assignment matrix $\mathbf{S} \in \mathbb{R}^{N_S \times N_P}$ between $F_{\mathbf{fus}}^{\mathrm{mm}}$ and the prototypes. This algorithm introduces the uniform distribution constraint to avoid trivial solution while achieving maximal similarity possible. We then use $\mathbf{S}$ to generate an update value for current sample's corresponding $\mathcal{P}_r$ , denoted as $\overline{\mathcal{P}_r}$ , as shown in Eq. (12), + +$$ +\overline {{\mathcal {P} _ {r}}} = \mathbf {S} ^ {\mathrm {T}} F _ {\mathbf {f u s}} ^ {\mathrm {m m}}. \tag {12} +$$ + +Afterwards, we update $\mathcal{P}_r$ through EMA [26] as in Eq. (13), where $m\in [0,1)$ is a momentum coefficient, + +$$ +\mathcal {P} _ {r} \leftarrow m \mathcal {P} _ {r} + (1 - m) \overline {{\mathcal {P} _ {r}}}. \tag {13} +$$ + +Each $\mathcal{P}_r$ is updated during pre-training and used as the fixed geo-context for downstream tasks. Geo-Context Prototype Learning is only conducted on the student branch. It extracts generalized region-aware representations from numerous RSI within a consistent region, offering a complementary clue to enhance the feature of a single RSI. + +As Geo-Context Prototype Learning is incorporated without an explicit loss term, our pre-training objective is shown in Eq. (14), where $\alpha$ and $\beta$ are trade-off weights, + +$$ +\mathcal {L} = \alpha \mathcal {L} _ {M G C L} + \beta \mathcal {L} _ {\text {a l i g n}}. \tag {14} +$$ + +# 4. Experiments + +Fig. 1 demonstrates SkySense's superior performance in all test scenarios. We conduct experiments on 16 datasets, covering different modalities and tasks, to ensure a comprehensive assessment. The right side of Fig. 3 shows how to apply SkySense to different tasks. Each pre-trained module is + +
ModelPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [2]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [46]ICCV'21-57.2089.0329.4/39.8
SatMAE [17]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo† [55]TGRS'22-67.2091.27-
RVSA [62]TGRS'2234.3/44.464.49--
BFM† [8]Arxiv'23--92.12-
TOV [57]JSTARS'2332.1/37.866.2492.03-
SSL4EO [65]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [48]TGRS'2336.4/43.566.2191.86-
CACo [45]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS† [63]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [47]ICCV'2336.7/45.666.6291.85-
Scale-MAE [50]ICCV'2334.0/41.765.7791.54-
SkySense-39.7/46.570.9193.9933.1/46.2
+ +(a) Semantic segmentation results. + +
ModelHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [2]67.4065.6548.15
SatMAE [17]70.8965.6646.55
RingMo†[55]75.90-46.21
RVSA [62]73.2271.0547.04
BFM†[8]-73.62-
TOV [57]70.1666.3349.62
SSL4EO [65]64.8261.2349.37
CMID [48]75.1166.3750.58
CACo [45]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [47]72.8467.6749.69
Scale-MAE [50]73.8166.4748.31
SkySense78.7374.2754.57
+ +(b) Object detection results. + +
ModelLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [2]78.1946.2613.6/16.7
SeCo [46]90.1447.6713.9/16.0
SatMAE [17]87.6552.7614.8/16.2
RingMo† [55]91.86--
RVSA [62]90.86--
SpectralGPT† [28]-54.29-
MATTER† [1]-59.37-
DINO-MC [66]-52.7014.5/15.6
SSL4EO [65]89.0535.0812.3/17.5
CMID [48]91.72--
CACo [45]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [47]91.7359.82-
Scale-MAE [50]92.07--
SkySense92.5860.0615.4/18.0
+ +(c) Change detection results. + +Table 2. Results of semantic segmentation, object detection and change detection. $\dagger$ means the code and weights are not released until November 11th, 2023, thus we report the metrics from the paper. - means the task is not supported or the value is unavailable in the paper. + +
ModelSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [2]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [46]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [17]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo†[55]96.90/98.3494.25/95.67--
RVSA [62]97.03/98.5093.93/95.69--
DINO-MC [66]--84.20/88.7560.16/83.49
TOV [57]95.16/97.0990.97/93.79--
SSL4EO [65]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [48]96.11/97.7994.05/95.53--
CACo [45]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMAT[21]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [47]95.47/97.0992.73/94.6486.30/--
Scale-MAE [50]96.44/97.5892.63/95.04--
SkySense97.68/98.6094.85/96.3288.67/92.0964.38/87.27
+ +Table 3. Scene classification results. + +designed to allow for combined or individual use, with the flexibility to be either frozen or fine-tuned as needed. More details are included in the supplementary materials. + +# 4.1. Pre-training Implementation + +The model is pre-trained with a batch size of 240 samples, distributed over 80 A100-80GB GPUs. For HSROIs, we apply data augmentations including multi-crop [6], Gaussian blur, solarization [23], etc. As for TMsiI and TSARI, we randomly select a fixed-sized sequence from the original one and perform random disturbances on the RSI acquisition date. We employ the huge version of the Swin Transformer (Swin-H) [41] as the spatial encoder of HSROIs, for its design efficiency in minimizing computational costs for high-resolution imagery [74]. RSI from TMsiI or TSARI is processed with corresponding ViT-L [20]. For Geo-Context Prototype Learning, we divide the globe into 4096 regions, each containing 100 prototypes. + +# 4.2. Performance on Single-Modal Tasks + +We evaluate SkySense on 4 representative single-modal tasks. All experiments are conducted using consistent finetuning settings for fairness. The supplementary materials include implementation details, visualization results and additional experiments on frozen backbone tuning. + +Semantic Segmentation. We adopt Dyna.-Pla. [59], iSAID [67], Potsdam [52] and Dyna.-S2 [59] for the segmentation experiment. They are chosen considering factors such as spatial resolution, spectrum and category type. UperNet [69] serves as the segmentation head. For Dyna.-Pla. and Dyna.-S2 datasets, we report mIoU results on official validation and test sets. For iSAID and Potsdam, we follow the settings of [55]. As depicted in Tab. 2a, SkySense has achieved the SOTA performance on all four segmentation datasets. On average, it surpasses the previous SOTA by an impressive improvement of $1.86\%$ . + +Horizontal & Oriented Object Detection. We employ the widely recognized DIOR [36] dataset for Horizontal Object Detection and its enhanced version DIOR-R [15], along with FAIR1M [56], for Oriented Object Detection. All datasets consist of optical RGB images. Faster RCNN [51] and Oriented RCNN [37] are used for the experiment, following the setup of [55, 62]. SkySense excels on all three datasets (Tab. 2b). Notably, we surpass the second best CMID by $3.99\%$ mAP and have achieved the best performance on the FAIR1M v2.01 leaderboard. More importantly, our results are accomplished without using any sophisticated Oriented Detection designs [29, 37, 71]. + +Change Detection. We assess SkySense's Change Detection performance on LEVIR-CD [9], OSCD [18], and Dyna.-S2 [59] datasets. For LEVIR-CD and OSCD, we follow the frameworks of [55, 66] and report the F1 metric. For Dyna.-S2, we utilize the UperNet head since the re + +![](images/82f381d2e9a4d11418ca391e0203c6e75fabb22acabb644c7fb6163893446ec1.jpg) +Figure 4. (a) Experiment on fine-tuning using different percentages of training data on the AID dataset. (b) The impact of S1-Ts data under varying cloud coverage conditions. + +![](images/7621b30499eb0385330ac21cdabd3fe9dbc32b8c7235d68f1b767ee5e32ac6ac.jpg) + +
Task & DatasetData Source & Geo-ContextPrevious SOTASkySense
(a) Multi-Modal Seg: Dyna.-MM(i) Planet.45.6 [47]46.5 ↑ 0.9
(ii) Planet. with GCP-47.0
(iii) S243.5 [4]46.2 ↑ 2.7
(iv) Planet. + S2-47.3
(v) Planet. + S2 + S1-47.7
(vi) Planet. + S2 + S1 with GCP-48.2
(b) Multi-Modal Seg: PASTIS-MM(i) S2-static-73.5
(ii) S2-Ts83.4 [58]84.6 ↑ 1.2
(iii) S2-Ts + S1-Ts84.2 [22]84.8 ↑ 0.6
(iv) S2-Ts + GEP-85.8
(v) S2-Ts + S1-Ts + GEP-85.9
(c) Multi-Modal Cls: BEN-MM(i) S183.70 [64]86.25 ↑ 2.55
(ii) S2 + S189.70 [64]92.21 ↑ 2.51
+ +Table 4. Fine-tuning results on multi-modal tasks. + +ported semantic change segmentation (SCS) score is calculated from the segmentation results [59]. Both results from the validation and test sets of Dyna.-S2 are presented. The remarkable generalization ability of SkySense is evident in the consistent improvements shown in Tab. 2c. Unlike CACo [45], which shows proficiency mainly on the Dyna.-S2 validation set, our model excels across all datasets. + +Scene Classification. We utilize four scene classification datasets: AID [68] and RESISC-45 [14] with static RGB images, BEN-S2 [53] with static multispectral images, and fMoW-S2 [17] with temporal multispectral images. The training ratio (TR) follows [17, 46, 55]. We use a linear classifier head for experiment. For AID and RESISC-45, we report Overall Accuracy (OA), for BEN-S2 we report mAP, and for fMoW-S2 we report both Top-1 and Top-5 Accuracy. SkySense overly outperforms competitive baselines and achieves the best results on all datasets (Tab. 3). Additionally, with limited labeled data on the AID dataset, SkySense consistently outperforms CMID, ScaleMAE, and random initialization, with a $4.17\%$ higher OA than the second best Scale-MAE using only $1\%$ training data (Fig. 4a). These results highlight the robustness and generalization ability of SkySense's pre-trained features. + +# 4.3. Performance on Multi-Modal Tasks + +Multi-Modal Segmentation: Time-insensitive Land Cover Mapping. We employ the Dyna.-MM dataset [59] + +for fine-tuning and report mIoU on the official test set. The dataset comprises HSROIs from PlanetFusion (Planet.), multispectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). We use a simple UperNet head. As shown in Tab. 4a, SkySense achieves the best results in single-modal scenarios (i) and (iii), clearly outperforming the previous SOTA by roughly $1\%$ mIoU. Moreover, combining all three modalities as (v) further improves the mIoU by $1.2\%$ compared to (i). Notably, without bells and whistles, SkySense ranks No.1 on the challenging DynamicEarthNet leaderboard2. + +Multi-Modal Segmentation: Time-sensitive Crop Mapping. We evaluate SkySense's fine-tuning result on the PASTIS-MM dataset, an enhanced version of PASTIS-R [22]. PASTIS-MM includes HSROIs from Google Earth Pro (GEP), TMsi from Sentinel-2 (S2-Ts), and TSARI from Sentinel-1 (S1-Ts). We use a naive FCN head [42] and report the OA from the official five-fold validation on PASTIS-MM dataset. In Tab. 4b, comparing S2-Ts and static multispectral data (S2-static), we observe a significant $11.1\%$ OA increase, highlighting the importance of incorporating temporal clue for crop mapping. + +Furthermore, both (ii) and (iii) exceed the performance of the previous SOTA, affirming the superior capabilities of SkySense. When more modalities are added as (ii), (iv), and (v), the OA increases accordingly. However, integrating Sentinel-1 data yields no substantial improvement, presumably because of the cloud-free imagery from PASTIS-MM dataset. To further investigate, we compare OA using S2-Ts data at different cloud ratios with and without Sentinel-1. Fig. 4b illustrates that the performance difference between utilizing and foregoing Sentinel-1 data becomes more pronounced with an increasing cloud ratio. Specifically, when the cloud ratio exceeds $50\%$ , the result of using Sentinel-1 outperforms its counterpart by $13\%$ . This highlights the importance of SAR data in situations with cloud coverage and rainfall. + +Multi-Modal Scene Classification. We utilize the BENMM [54] dataset for evaluating the multi-modal scene classification task. This dataset includes both Sentinel-1 (S1) and Sentinel-2 (S2) imagery. The evaluation protocol from DeCUR [64] is followed, and we report the mAP metric of fine-tuning with $100\%$ training data. In Tab. 4c, both (i) and (ii) significantly outperform the previous SOTA by more than $2.5\%$ mAP. Furthermore, the inclusion of Sentinel-2 imagery greatly enhances performance compared to using Sentinel-1 imagery alone. + +All these results show a notable gain for the tasks using multi-modal data, affirming the necessity of SkySense's multi-modal pre-training from one perspective. + +
Pre-trainingiSAIDfMoW-S2
mIoUTop-5 Acc
Simple Ver.68.9885.69
SkySense70.91 ↑ 1.9387.27 ↑ 1.58
+ +(a) + +
Pre-trainingDyna.-MM
mIoU
Baseline42.2
+ MGCL44.4 ↑ 2.2
+ MM47.0 ↑ 2.6
+ CMA47.7 ↑ 0.7
+ GCPL48.2 ↑ 0.5
+ +(b) + +Table 5. (a) Discussion on multi-modal pre-training effectiveness. (b) Ablation study on the pre-training design. + +# 5. Discussions & Ablation Studies + +Multi-modal Pre-training Effectiveness. In addition to confirming the effectiveness of using multi-modal data in downstream tasks, we investigate the impact of multi-modal pre-training on single-modal tasks, compared with pretraining on fewer modalities. We conduct experiments on iSAID for static HSROI segmentation and fMoW-S2 for temporal multispectral classification. Two versions of the pre-trained model are tested: a simple version pre-trained only with optical imagery (HSROIs, TMsI), and SkySense, which includes HSROIs, TMsI, and TSARI for pre-training. The rest of the settings remain consistent. The results in Tab. 5a show that SkySense consistently outperforms the simple version, suggesting that the introduction of SAR data benefits representation learning of other modalities. This may attribute to the implicit clue brought by SAR data through Cross-Modal Alignment. It provides another perspective on the necessity of SkySense's multi-modal pretraining. + +What does Geo-Context Prototype (GCP) Learn? We utilize the Dyna.-MM dataset for experiment as it contains diverse geo-locations worldwide. For the segmentation task in Tab. 4a, adding GCP in downstream tasks leads to a further gain of $0.5\%$ mIoU compared to the strong multi-modal baseline (v). Moreover, a comparison between (i) and (ii) shows a $0.5\%$ mIoU improvement using GCP for the single-modal task. It demonstrates GCP's consistent performance gain in single- and multi-modal scenarios. + +In Fig. 5, we visualize the learned prototypes on the Map by calculating the pre-trained feature of each pixel and assigning the most similar prototype to it. A comparison with the ESRI LandCover Map [34] reveals GCP's promising results in segmenting different areas. Moreover, GCP exhibit fine-grained advantage, as shown in the middle of Fig. 5. The prototypes learned from unsupervised clustering segment cropland within the town, which is overlooked by the LandCover Map. Notably, the visualization shares the same spatial resolution with ESRI LandCover Map. + +Design of Pre-training. Tab. 5b presents the ablation study to assess our pre-training design, namely Multi-Granularity Contrastive Learning (MGCL), Multi-Modal (MM) integration, Cross-Modal Alignment (CMA) and Geo-Context + +![](images/cafeb5289967c67afa0371759e6485cae4f30adfd0b88321acbfeae07de0c87c.jpg) +Figure 5. Comparison between (a) ESRI LandCover Map and (b) Geo-Context Prototype. The visualization process of Geo-Context Prototype is illustrated in the upper part of this figure. + +Prototype Learning (GCPL). We utilize the Dyna.-MM dataset for the experiment and report mIoU metric on the official test set. + +Initially, we utilize a single-modal version, using $g_{HR}$ spatial encoder and HSROIs for pre-training. The training settings are kept the same as described in Sec. 4.1. The results show that MGCL leads to a notable improvement compared to a simple baseline [7]. Then we integrate further multi-modal data (i.e., TMsiI and TSARI) into the pretraining and downstream evaluation. This effectively improves the performance on the test set to $47.0\%$ mIoU, validating the necessity of multi-modal pre-training. + +CMA is another necessary design for SkySense's pretraining, which explicitly pulls features from different modalities together, encouraging cross-modal interactions. The results show that the incorporation of modal alignment leads to $0.7\%$ mIoU improvement. Finally, we introduce GCPL, which learns complementary regional context clues to facilitate downstream tasks and further pushes the very strong performance to $48.2\%$ mIoU. + +# 6. Conclusion & Future Work + +In this paper, we present SkySense, a large-scale MM-RSFM for interpretation of EO imagery. SkySense allows using its modules flexibly to accommodate different scenarios and consistently outperforms other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope SkySense will inspire further research on MM-RSFM and its release may contribute to sustainable innovations thriving in the Remote Sensing community. As part of our future work, we plan to incorporate the language modality, thereby extending SkySense's applications to more EO tasks. + +# 7. Acknowledgment + +This work was in part supported by the National Natural Science Foundation of China under Grants 42030102 and 42371321. + +# References + +[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 2, 6 +[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 2, 6 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 2 +[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinand, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 2, 6, 7 +[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1, 2 +[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4, 5, 6 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 8 +[8] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 2, 6 +[9] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[11] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 2, 4, 5 +[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Mocov2: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2 +[13] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1 + +[14] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 7 +[15] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6 +[16] Mingmin Chi, Antonio Plaza, Jon Atli Benediktsson, Zhongyi Sun, Jinsheng Shen, and Yangyong Zhu. Big data for remote sensing: Challenges and opportunities. Proceedings of the IEEE, 104(11):2207-2219, 2016. 1 +[17] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 3, 6, 7 +[18] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6 +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6 +[21] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 3, 6 +[22] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7 +[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 6 +[24] Shaohua Guo, Liang Liu, Zhenye Gan, Yabiao Wang, Wuhao Zhang, Chengjie Wang, Guannan Jiang, Wei Zhang, Ran Yi, Lizhuang Ma, et al. Isdnet: Integrating shallow and deep networks for efficient ultra-high resolution segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4361-4370, 2022. 2, 4, 5 + +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2 +[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5 +[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 +[28] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Gamba Paolo, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral foundation model. arXiv preprint arXiv:2311.07113, 2023. 6 +[29] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 923-932, 2022. 6 +[30] Jingliang Hu, Lichao Mou, and Xiao Xiang Zhu. Unsupervised domain adaptation using a teacher-student network for cross-city classification of sentinel-2 images. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 43:1569–1574, 2020. 5 +[31] Lanqing Huang, Bin Liu, Boying Li, Weiwei Guo, Henhao Yu, Zenghui Zhang, and Wenxian Yu. Opersarship: A dataset dedicated to sentinel-1 ship interpretation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(1):195–208, 2017. 2 +[32] Xin Huang, Yihong Song, Jie Yang, Wenrui Wang, Huiqun Ren, Mengjie Dong, Yujin Feng, Haidan Yin, and Jiayi Li. Toward accurate mapping of 30-m time-series global impervious surface area (gisa). International Journal of Applied Earth Observation and Geoinformation, 109:102787, 2022. 2, 4, 5 +[33] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Nyirjesy Gabby, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mukkavilli, Devyani Lambate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence. arXiv preprint arXiv:2310.18660, 2023. 2 +[34] Krishna Karra, Caitlin Kontgis, Zoe Statman-Weil, Joseph C Mazzariello, Mark Mathis, and Steven P Brumby. Global land use/land cover with sentinel 2 and deep learning. In 2021 IEEE international geoscience and remote sensing symposium IGARSS, pages 4704-4707. IEEE, 2021. 8 +[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. + +Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5 +[36] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6 +[37] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6 +[38] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. Mcanet: A joint semantic segmentation framework of optical and sar images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106: 102638, 2022. 2 +[39] Yansheng Li, Wei Chen, Xin Huang, Gao Zhi, Siwei Li, Tao He, and Yongjun Zhang. Mfvnet: a deep adaptive fusion network with multiple field-of-views for remote sensing image semantic segmentation. Science China Information Sciences, 2023. 2 +[40] Yinhe Liu, Sunan Shi, Junjue Wang, and Yanfei Zhong. Seeing beyond the patch: Scale-adaptive semantic segmentation of high-resolution remote sensing imagery based on reinforcement learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16868-16878, 2023. 2, 4, 5 +[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6 +[42] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 7 +[43] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1 +[44] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 75-82, 2019. 2 +[45] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7 +[46] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7 + +[47] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 3, 6, 7 +[48] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 3, 6 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[50] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 2, 3, 6 +[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6 +[52] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6 +[53] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 7 +[54] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7 +[55] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7 +[56] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6 +[57] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learn + +ing. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 2, 6 +[58] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10418-10428, 2023. 7 +[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7 +[60] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N van Rijn, Holger H Hoos, Fabio Del Frate, et al. Artificial intelligence to advance earth observation: a perspective. arXiv preprint arXiv:2305.08413, 2023. 1 +[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[62] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 2, 6 +[63] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 2, 6 +[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Graham, Chenying Liu, Zhitong Xiong, and Xiao Xiang Zhu. Decur: decoupling common & unique representations for multimodal self-supervision. arXiv preprint arXiv:2309.05300, 2023. 3, 7 +[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 2, 6 +[66] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6 +[67] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6 +[68] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 7 + +[69] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6 +[70] Binbin Yang, Xinchi Deng, Han Shi, Changlin Li, Gengwei Zhang, Hang Xu, Shen Zhao, Liang Lin, and Xiaodan Liang. Continual object detection via prototypical task correlation guided gating mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9255-9264, 2022. 1 +[71] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13354-13363, 2023. 6 +[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1 +[73] Hankui K Zhang, David P Roy, and Dong Luo. Demonstration of large area land cover classification with a one-dimensional convolutional neural network applied to single pixel temporal metric percentiles. Remote Sensing of Environment, 295:113653, 2023. 2 +[74] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 6 +[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 2 +[76] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 1 +[77] Stefano Zorzi, Shabab Bazrafkan, Stefan Habenschuss, and Friedrich Fraundorfer. *Polyworld: Polygonal building extraction with graph neural networks in satellite images*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1848-1857, 2022. 1 \ No newline at end of file diff --git a/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/images.zip b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..905a07f8dfaa729172ba71ae46ff92620708d6db --- /dev/null +++ b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b0b6c566596d33ac276ac3d7ed32b1fbd75291bb5a2c6b2936f0006de41cf7d +size 683712 diff --git a/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/layout.json b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4424dd722b712061874011f0a595689cb3f45254 --- /dev/null +++ b/2024/SkySense_ A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery/layout.json @@ -0,0 +1,10846 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 547, + 140 + ], + "type": "text", + "content": "SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": "Xin Guo" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Jiangwei Lao" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Bo Dang" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{\\S 2*}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Yingying Zhang" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Lei Yu" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Lixiang Ru" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Liheng Zhong" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Ziyuan Huang" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Kang Wu" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{\\S 2}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Dingxiang Hu" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{3,1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Huimei He" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{3,1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Jian Wang" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Jingdong Chen" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Ming Yang" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Yongjun Zhang" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": ", Yansheng Li" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": "Ant Group " + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": "Wuhan University " + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 160, + 552, + 219 + ], + "type": "text", + "content": "MYBank" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 220, + 455, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 220, + 455, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 220, + 455, + 232 + ], + "type": "text", + "content": "{bangzhu.gx, wenshuo.ljw}@antgroup.com, bodang@whu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "spans": [ + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "text", + "content": "Prior studies on Remote Sensing Foundation Model (RSFM) reveal immense potential towards a generic model for Earth Observation. Nevertheless, these works primarily focus on a single modality without temporal and geocontext modeling, hampering their capabilities for diverse tasks. In this study, we present SkySense, a generic billion-scale model, pre-trained on a curated multi-modal Remote Sensing Imagery (RSI) dataset with 21.5 million temporal sequences. SkySense incorporates a factorized multi-modal spatiotemporal encoder taking temporal sequences of optical and Synthetic Aperture Radar (SAR) data as input. This encoder is pre-trained by our proposed Multi-Granularity Contrastive Learning to learn representations across different modal and spatial granularities. To further enhance the RSI representations by the geo-context clue, we introduce Geo-Context Prototype Learning to learn region-aware prototypes upon RSI's multi-modal spatiotemporal features. To our best knowledge, SkySense is the largest Multi-Modal RSFM to date, whose modules can be flexibly combined or used individually to accommodate various tasks. It demonstrates remarkable generalization capabilities on a thorough evaluation encompassing 16 datasets over 7 tasks, from single- to multi-modal, static to temporal, and classification to localization. SkySense surpasses 18 recent RSFMs in all test scenarios. Specifically, it outperforms the latest models such as GFM, SatLas and Scale-MAE by a large margin, i.e., " + }, + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "inline_equation", + "content": "2.76\\%" + }, + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "inline_equation", + "content": "3.67\\%" + }, + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "inline_equation", + "content": "3.61\\%" + }, + { + "bbox": [ + 45, + 284, + 290, + 632 + ], + "type": "text", + "content": " on average respectively. We will release the pre-trained weights to facilitate future research and Earth Observation applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 641, + 128, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 128, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 128, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "type": "text", + "content": "Remote Sensing Imagery (RSI) interpretation is crucial in understanding our common home, the Earth [16, 60], via" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 258, + 549, + 458 + ], + "blocks": [ + { + "bbox": [ + 308, + 258, + 549, + 458 + ], + "lines": [ + { + "bbox": [ + 308, + 258, + 549, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 549, + 458 + ], + "type": "image", + "image_path": "b7a5a415ce6e68b2561b5f704c42d5811180eb157b6de1703152cc7f3a8a8d9d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 465, + 547, + 499 + ], + "lines": [ + { + "bbox": [ + 305, + 465, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 547, + 499 + ], + "type": "text", + "content": "Figure 1. SkySense has achieved superior performance on 16 datasets over 7 distinct tasks compared with 18 state-of-the-art RSFMs and supports a board range of EO imagery interpretations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 605 + ], + "type": "text", + "content": "quite diverse tasks [5, 13, 43, 72], e.g. crop monitoring, natural disaster management, etc. Every task may require significant dedicated efforts and resources to build a task-specific model. Recently, Foundation Model emerges as a pre-trained generic model that excels in a wide range of downstream tasks [70, 76]. Hence, there is a soaring interest in exploring a comprehensive Remote Sensing Foundation Model (RSFM) for many Earth Observation (EO) tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "The key question naturally arises: What is essential for a RSFM? First of all, an ideal RSFM should possess the ability to perceive multi-modal temporal RSI. EO heavily relies on multi-modal time series of remote sensing data, including temporal optical and Synthetic Aperture Radar (SAR) data. Individual modality offers unique advantages and complements to each other. For example, optical images provide rich spectral bands and texture details but are susceptible to weather [77]. In contrast, SAR sensors cap" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "*Equally contributing first authors. † Corresponding authors. §Work done during the internship of the author at Ant Group." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27672" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 148 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 148 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 148 + ], + "type": "table", + "html": "
ModelDifferent EO Interpretation Input Types
Single-Modal O(RGB)Single-Modal O(Ms)Multi-Modal Static O & SARMulti-Modal Temporal O & SAR
SkySense✓✓✓✓
SatLas[4]✓✓
GFM[47]✓
Scale-MAE[50]✓
", + "image_path": "b4d8a304f15428aa642fb9359edecc34d09fb75f29cbd7f46eed4467e8560e01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 157, + 287, + 179 + ], + "lines": [ + { + "bbox": [ + 47, + 157, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 287, + 179 + ], + "type": "text", + "content": "Table 1. SkySense supports various input types. O(RGB): Optical RGB images; O(Ms): Optical multispectral images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 196, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 350 + ], + "type": "text", + "content": "ture clear imagery in all weather conditions [31, 38]. Moreover, the time series of such data provide the crucial temporal clue to various tasks [5, 22, 73] like change prediction. Second, a RSFM should be easy to tailor when being deployed for EO tasks using different modalities (i.e., single- and multi-modal) at different spatial (i.e., pixel-, object-, and image-level) granularities. Last but not the least, remote sensing data is inherently contingent on their spacetime coordinates, which provide rich regional and seasonal geo-context that benefits RSI interpretation a lot, as indicated in [11, 24, 32, 39, 40]. Therefore, a RSFM shall bear the vital capability of effective geo-context learning and utilization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 425 + ], + "type": "text", + "content": "Previous works on RSFM [1, 2, 4, 8, 17, 33, 45-48, 50, 55, 57, 62, 63, 65, 66] have demonstrated their preliminary success on several specific datasets. However, these RSFMs, while proficient in certain areas, are limited in their applications to EO tasks, due to factors such as single-modal pre-training and the neglect of geo-context." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 714 + ], + "type": "text", + "content": "In this paper, we propose SkySense, a billion-scale Multi-Modal Remote Sensing Foundation Model (MM-RSFM). SkySense incorporates 2.06 billion parameters and is pre-trained on a large-scale multi-modal dataset which comprises 21.5 million RSI temporal sequences extracted from high-spatial-resolution optical images (HSROIs), medium-resolution temporal multispectral imagery (TMsI) and temporal SAR imagery (TSARI). To handle the multimodal temporal RSI sequences, SkySense employs a factorized multi-modal spatiotemporal encoder to perform spatial feature extraction and multi-modal temporal fusion independently, since RSI sequence are spatially-aligned in nature. It leads to a modular design allowing flexible use of its modules, i.e., the spatial encoder can be either used alone or in combination of the fusion module to support tasks from static single-modal to temporal multi-modal. This design delivers strong modeling of RSI sequences while using substantially less parameters compared to common 3D structures [44, 75]. The factorized encoder is pre-trained by Multi-Granularity Contrastive Learning to construct features from different modal and spatial granularities. Furthermore, we propose Geo-Context Prototype Learning to generate regional prototypes from RSI features given geolocations. This approach enhances multi-modal spatiotem" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "poral representation learning by leveraging the regional context clue hidden in numerous unlabeled RSI." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "text", + "content": "SkySense has achieved the state-of-the-art (SOTA) performance across a variety of modalities and EO tasks, as shown in Fig. 1. We evaluate SkySense on a diverse set of 16 datasets [9, 14, 15, 17, 18, 22, 36, 53, 56, 59, 67, 68], where the selection covers different task types, modalities and spatial scales. The results demonstrate that Sky-Sense outperforms 18 advanced RSFMs [1, 2, 4, 8, 17, 45-48, 50, 55, 57, 62, 63, 65, 66] in all test scenarios, validating its competitive edge for a broad range of EO interpretation tasks. Tab. 1 compares our work with latest representative studies w.r.t. various input types of EO interpretation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 229, + 497, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 229, + 497, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 229, + 497, + 241 + ], + "type": "text", + "content": "In summary, our technical contributions are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 242, + 545, + 456 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 242, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 545, + 289 + ], + "type": "text", + "content": "- We propose SkySense, the largest MM-RSFM to date with a modular design, which is capable of handling diverse tasks, from single- to multi-modal, static to temporal, and classification to localization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 290, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 545, + 385 + ], + "type": "text", + "content": "- The design of SkySense involves three novel technical components: a) A factorized multi-modal spatiotemporal encoder to effectively process multi-modal temporal RSI; b) Multi-Granularity Contrastive Learning that learns features at various levels of granularities to facilitate different tasks; c) Geo-Context Prototype Learning to extract region-aware geo-context clue to enable implicit geo-knowledge integration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 386, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 386, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 386, + 545, + 456 + ], + "type": "text", + "content": "- We extensively compare SkySense with 18 recently published RSFMs. Our model has achieved the SOTA performance, supparssing the latest models like GFM, SatLas and Scale-MAE by over " + }, + { + "bbox": [ + 306, + 386, + 545, + 456 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 306, + 386, + 545, + 456 + ], + "type": "text", + "content": " on average. We hope the release of pre-trained weights will contribute to the Remote Sensing community and facilitate future research." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 470, + 392, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 392, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 392, + 483 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 492, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 492, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 492, + 503 + ], + "type": "text", + "content": "2.1. Remote Sensing Foundation Model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Recent Remote Sensing Foundation Models draw their primary inspiration from the research on Vision Foundation Model [3, 7, 10, 19, 23, 25-27, 41, 49, 61]. Remote sensing data inherently integrates space-time coordinates and has diverse spatial scales. The mainstream RSFMs extend the foundation model techniques to space-time RS data, such as Contrastive Learning. For instance, GASSL [2] utilized geo-location prediction as an additional pre-text task in the MoCo-v2 framework [12]. Multiple views with different sizes were utilized by DINO-MC [66] for self-supervised learning within the DINO framework [7]. SeCo [46] and CACo [45] both proposed Contrastive Learning to perceive short-term and long-term changes by using the spatiotemporal structure of temporal RSI sequences. Besides, there are works either improving the MIM-based framework [50, 55, 62] or exploring the model scale-up [8]. For example, RingMo [55] modified MAE to adapt to the dense" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27673" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 292, + 262 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 292, + 262 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 292, + 262 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 292, + 262 + ], + "type": "image", + "image_path": "b4c7d33ef7197a5a1250fcd5e501c8dee2f25ff825e11ba32886d897ac087221.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 271, + 277, + 283 + ], + "lines": [ + { + "bbox": [ + 56, + 271, + 277, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 277, + 283 + ], + "type": "text", + "content": "Figure 2. The overview of our SkySense model architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 437 + ], + "type": "text", + "content": "objects in RSI. SatMAE [17] employed TMsi to enhance the performance on temporal sequences. Scale-MAE [50] built a framework with scale-aware encoder. Recent efforts such as CMID [48] and GFM [47] have commenced to explore amalgamation of CL and MIM strategies. Concurrently, CROMA [21] and DeCUR [64] investigated multimodal pre-training for single- and multi-modal tasks using static imagery. In this study, we propose a comprehensive MM-RSFM, SkySense, to fill the gap in existing RSFMs, i.e., single modality of RingMo, CACo, etc., static input of Scale-MAE, CROMA, etc., and the neglect of geo-context of SatLas, RVSA, etc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 111, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 111, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 111, + 462 + ], + "type": "text", + "content": "3. SkySense" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 470, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 495 + ], + "type": "text", + "content": "In this section, we introduce the pre-training dataset and the design choices for individual module respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 503, + 167, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 167, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 167, + 515 + ], + "type": "text", + "content": "3.1. Pre-training Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "We curate an extensive multi-modal remote sensing dataset with temporal sequences, containing RSI from various sources: HSROIs from WorldView-3, 4, etc. (RGB band), TMsi from Sentinel-2 (B2-8, B8A, B11-12 band) and TSARI from Sentinel-1 (VV, VH Polarization). All data is geo-spatially aligned. Strictly speaking, HSROIs and TMsi shall be categorized to the optical modality, while TSARI falls to the SAR modality. However, due to HSROIs and TMsi's significant difference in spectral band and ground sample distance, we regard HSROIs and TMsi as two distinct modalities for simplicity in this paper. The dataset comprises 21.5 million training samples, each consisting of a static HSROI with rich texture details, a TMsi containing temporal and multispectral data, a TSARI providing backscatter polarization under cloud coverage, and the metadata like geo-location and acquisition date for geo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "context modeling. This dataset covers a great variety of scenarios across resolution, spectrum, and imaging mechanism. More details of the data are included in the supplementary materials. We construct the input for SkySense as " + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "inline_equation", + "content": "\\{x_{HR}, x_{Ms}, x_{SAR}\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "inline_equation", + "content": "x_{HR}" + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": " represents a static HSROI; " + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "inline_equation", + "content": "x_{Ms}" + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": " is a Sentinel-2 TMsi after filtering cloudy images, where we randomly select 20 images to form the sequence; and " + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "inline_equation", + "content": "x_{SAR}" + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": " stands for a standard-calibrated TSARI, from which we randomly select 10 images for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 188, + 422, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 188, + 422, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 422, + 200 + ], + "type": "text", + "content": "3.2. Model Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 206, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 326 + ], + "type": "text", + "content": "Factorized Multi-Modal Spatiotemporal Encoder. The overall architecture of our method is illustrated in Fig. 2. In a multi-modal input " + }, + { + "bbox": [ + 304, + 206, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\{x_{HR}, x_{Ms}, x_{SAR}\\}" + }, + { + "bbox": [ + 304, + 206, + 545, + 326 + ], + "type": "text", + "content": ", the pixels within each RSI naturally align with the others given the same geo-location. Upon this, we propose a factorized encoder that initially extracts spatial features from each RSI independently and then fuses them to capture a multi-modal spatiotemporal representation. The design separates spatial feature extraction from the feature fusion, enabling the integration of the clues from modality, time and geo-context." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": "Spatial Feature Extraction. To handle the spatially aligned sequence input " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\{x_{HR}, x_{Ms}, x_{SAR}\\}" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": ", we utilize the spatial encoder " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "g_{HR}, g_{Ms}" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "g_{SAR}" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " for each individual RSI from HSROI, TMsI and TSARI respectively. As shown in Eq. (1), the obtained feature " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "F_i \\in \\mathbb{R}^{h \\times w \\times T_i \\times d}" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "i \\in \\{HR, Ms, SAR\\}" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " are of the same size in spatial dimension, where " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " are the height and width of " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "T_{HR}, T_{Ms}, T_{SAR}" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " represent the sequence lengths of HSROI, TMsI, and TSARI respectively, and " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " is the feature dimension. The initial multi-modal temporal feature representation " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "F_T \\in \\mathbb{R}^{N_S \\times N_T \\times d}" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " is generated by concatenating all " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " along the time dimension, where " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "N_S = h \\times w" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " represents the feature size in the spatial dimension, and " + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "inline_equation", + "content": "N_T = \\sum_{i \\in \\{HR, Ms, SAR\\}} T_i" + }, + { + "bbox": [ + 304, + 326, + 546, + 506 + ], + "type": "text", + "content": " represents the total sequence length across all modalities," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 512, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 512, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 351, + 512, + 545, + 540 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} F _ {i} = g _ {i} \\left(x _ {i}\\right), i \\in \\{H R, M s, S A R \\}, \\\\ F _ {i} = G _ {i} + \\left[ F _ {i - 1}, F _ {i - 2}, F _ {i - 3} \\right] \\end{array} \\tag {1} \\\\ F _ {T} = \\operatorname {C o n c a t} \\left[ F _ {H R}, F _ {M s}, F _ {S A R} \\right]. \\\\ \\end{array}", + "image_path": "24d23b2ee81a14f0408e08c751d87b7361b3c1b39c245b7caa70769274c80756.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": "Multi-modal Temporal Fusion. Next, we incorporate the date-specific temporal positional encoding " + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "inline_equation", + "content": "P_{DTPE}[:,\\mathbf{t},:]\\in \\mathbb{R}^{1\\times N_T\\times d}" + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": " through broadcasting, creating " + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "inline_equation", + "content": "F_{T}^{date}" + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": " for date-aware modeling. " + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "inline_equation", + "content": "F_{T}^{date}" + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": " is then concatenated with an extra token " + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{e}}\\in \\mathbb{R}^{N_S\\times 1\\times d}" + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": " [20] (see Eq. (2))," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 613, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 613, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 358, + 613, + 545, + 633 + ], + "type": "interline_equation", + "content": "F _ {T} ^ {\\text {d a t e}} = F _ {T} + P _ {D T P E} [:, \\mathbf {t}, : ], \\tag {2}", + "image_path": "a72c1d467dd317dd2a42e1a15353bdb963c37c15379d289a35232c275f08e765.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 628, + 517, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 628, + 517, + 643 + ], + "spans": [ + { + "bbox": [ + 323, + 628, + 517, + 643 + ], + "type": "interline_equation", + "content": "F _ {T} ^ {c a t} = \\operatorname {C o n c a t} \\left[ F _ {\\mathbf {e}}, F _ {T} ^ {d a t e} \\right] \\in \\mathbb {R} ^ {N _ {S} \\times (1 + N _ {T}) \\times d},", + "image_path": "b0b5f17f473a9833eab0d2c7b442b2d007c1736cbb38cf135866f1637c7cad46.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^{N_T}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " is a vector containing the acquisition dates of all RSI in the current batch. " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "P_{DTPE} \\in \\mathbb{R}^{1 \\times 365 \\times d}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " is a learnable parameter representing different dates of a year, which is essential for tasks affected by seasons (e.g., crop recognition). " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "F_T^{cat}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " is then fed into the Multi-modal" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27674" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": "Temporal Fusion Transformer, composed of multiple Naive Transformer encoder layers. This module employs selfattention to integrate multi-modal temporal data, generating the multi-modal spatiotemporal feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}} \\in \\mathbb{R}^{N_S \\times 1 \\times d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": ". Attentional Geo-Context Integration. Each RSI's geographical location may reveal rich region-specific geocontext. It is valuable for RSI interpretation as indicated by [11, 24, 32, 40]. To utilize this contextual clue to enhance " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": ", we employ a region-specific prototype set " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N_R \\times N_p \\times d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " (shown on the right side of Fig. 2), where " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "N_R" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " is the number of regions, " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " represents the number of prototypes for each region and " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " denotes the feature dimension. The learning procedure of " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " will be elaborated in Sec. 3.3. Specifically, a regional prototype subset " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r \\in \\mathbb{R}^{N_p \\times d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " is chosen from " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " based on the geo-location embedded with " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " is then attended to the prototypes of " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " through the attention mechanism, as shown in Eq. (3). The weights, computed from Softmax " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "\\left(\\frac{QK^T}{\\sqrt{d}}\\right)" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": ", facilitate a soft selection of prototypes in accordance with their similarity to " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": ". The final representation " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}} \\in \\mathbb{R}^{N_S \\times 2d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " is generated by concatenation of " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " and weighted sum of prototypes from " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": " along the feature dimension. The prototypes represent a set of discriminative features linked to certain semantics like water body, cropland, etc. By finding the similar ones to " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": ", we provide the standard representations of certain semantics to complement " + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 391 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 396, + 287, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 396, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 64, + 396, + 287, + 439 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {\\mathbf {f u s}} = \\operatorname {C o n c a t} \\left[ F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}, \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\right], \\tag {3} \\\\ Q = F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}, K = V = \\mathcal {P} _ {r}. \\\\ \\end{array}", + "image_path": "5126a040ab56e555c343634fcd3a418530e8852b47c38c47fc24cafb9136616c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 459, + 128, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 128, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 128, + 471 + ], + "type": "text", + "content": "3.3. Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 478, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 573 + ], + "type": "text", + "content": "An overview of our pre-training procedure is illustrated in Fig. 3. We build the pre-training framework on a common teacher-student structure [7], since it conducts self-supervised learning using only positive pairs, which is easily accessible given spatially aligned RSI, avoiding complicated design of negative pairs. Teacher's parameter set " + }, + { + "bbox": [ + 46, + 478, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + }, + { + "bbox": [ + 46, + 478, + 287, + 573 + ], + "type": "text", + "content": " is updated through exponential moving average (EMA) [26] from student's parameter set " + }, + { + "bbox": [ + 46, + 478, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 478, + 287, + 573 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": "Multi-Granularity Contrastive Learning. We propose Multi-Granularity Contrastive Learning for self-supervised learning on different modal and spatial granularities for diverse tasks. Given the input " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\{x_{HR}, x_{Ms}, x_{SAR}\\}" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": ", two sets of random augmentations are employed, generating two groups of views " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\{u_i\\}" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\{v_i\\}" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "i \\in \\{HR, Ms, SAR\\}" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": " are subsequently fed into the spatial encoders from the student and teacher branches respectively. " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": " is the student's spatial encoder and " + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "inline_equation", + "content": "g_i'" + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": " is the teacher's. The features are generated as in Eq. (4)," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "F _ {i} = g _ {i} \\left(u _ {i}\\right), F _ {i} ^ {\\prime} = g _ {i} ^ {\\prime} \\left(v _ {i}\\right) \\quad i \\in \\{H R, M s, S A R \\}. \\tag {4}", + "image_path": "d4082f2c51c23b6e5168bcf8cecbbb8f25b057d23e92319989923f4406289dd5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": "After applying the multi-modal temporal fusion and geocontext integration on " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "F_{i}^{\\prime}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": ", the final feature " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}^{\\prime}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " are obtained. Initially, we establish pixel-, object- and image-level contrastive learning to progressively learn coarse-to-fine spatial features for various tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": "Each temporal slice of " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " can be viewed as a pixel-level feature " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathbf{pix}} \\in \\mathbb{R}^{N_{S} \\times d}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": ". Pixel-level contrastive learning loss " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathbf{pix}}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " is obtained by averaging all " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CL}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " across the spatial " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " and temporal " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " dimensions, as shown in Eq. (5). " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathbf{pix}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " represents a feature vector from " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathbf{pix}}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathbf{pix'}}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " is its correspondence at the same geo-location. " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CL}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " denotes the learning loss [7] between " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathbf{pix}}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathbf{pix'}}" + }, + { + "bbox": [ + 304, + 133, + 545, + 217 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 220, + 545, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 220, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 318, + 220, + 545, + 248 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {p i x}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{N _ {S} T _ {i}} \\sum_ {s} \\sum_ {t} \\mathcal {L} _ {C L} \\left(f _ {i} ^ {\\mathbf {p i x}}, f _ {i} ^ {\\mathbf {p i x} ^ {\\prime}}\\right). \\tag {5}", + "image_path": "bbbbdc9b9a234fd14070d8ef57f56699f74ee46f4cb1cf886f811f3837bbaa4a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathbf{obj}} \\in \\mathbb{R}^{N_{C} \\times d}" + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "text", + "content": " denotes object-level feature generated from unsupervised clustering on pixel-level feature vectors " + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathbf{pix}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "text", + "content": " in a single RSI, where " + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "inline_equation", + "content": "N_{C}" + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "text", + "content": " is the number of clusters. The clustering employs the same Sinkhorn-Knopp algorithm [6] we apply for Geo-Context Prototype Learning, as shown later. " + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathbf{obj}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "text", + "content": " is the vector representing the cluster centers in " + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathbf{obj}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "text", + "content": ", which can be viewed as a general representation for a set of collected " + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathbf{pix}}" + }, + { + "bbox": [ + 304, + 255, + 545, + 380 + ], + "type": "text", + "content": ". It usually corresponds to a certain ground object or semantics. The object-level contrastive learning loss is computed as Eq. (6)," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 382, + 545, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 382, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 318, + 382, + 545, + 410 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {o b j}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{N _ {C} T _ {i}} \\sum_ {s} \\sum_ {t} \\mathcal {L} _ {C L} \\left(f _ {i} ^ {\\mathbf {o b j}}, f _ {i} ^ {\\mathbf {o b j} \\prime}\\right). \\tag {6}", + "image_path": "25d8bd7345bdf28182bf62852a11b0ba8d35b3741b4acbb88b6fd85b9478783d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 418, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 456 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathrm{img}}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 418, + 545, + 456 + ], + "type": "text", + "content": " corresponds to the image-level feature, which is an average pooling result from " + }, + { + "bbox": [ + 304, + 418, + 545, + 456 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathrm{pix}}" + }, + { + "bbox": [ + 304, + 418, + 545, + 456 + ], + "type": "text", + "content": " . Image-level contrastive learning loss is illustrated by Eq. (7)," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 460, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 460, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 334, + 460, + 545, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i m g}} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\frac {1}{T _ {i}} \\sum_ {t} \\mathcal {L} _ {C L} \\left(F _ {i} ^ {\\mathrm {i m g}}, F _ {i} ^ {\\mathrm {i m g} ^ {\\prime}}\\right). \\tag {7}", + "image_path": "c489b6514491b4e106ea48c91f79b6a90b6fec720080294970c14472f9c424ca.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "text", + "content": "The fine-grained contrastive learning loss " + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FGCL}" + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "text", + "content": " is the sum of pixel-, object- and image-level contrastive learning losses as Eq. (8). Finally we form the Multi-Granularity Contrastive Learning loss " + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MGCL}" + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "text", + "content": " in Eq. (9). The concept of multi-granularity is reflected in two aspects: space and modality. In terms of space, contrastive learning is performed at the pixel-, object-, and image-level, facilitating representation learning that encapsulates diverse spatial dimensions. Regarding modality, we conduct contrastive learning on the feature of each single modality, i.e., " + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "text", + "content": ", and the multi-modal feature after fusion, i.e., " + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}" + }, + { + "bbox": [ + 304, + 494, + 545, + 626 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 639, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 639, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 329, + 639, + 545, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F G C L} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) = \\sum_ {n \\in \\{\\mathbf {p i x}, \\mathbf {o b j}, \\mathbf {i m g} \\}} \\mathcal {L} _ {n} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right), \\tag {8}", + "image_path": "148b93ae8fd2c8b14f2725919ea3153bc635d70b99f39f81fa8ed4d817cb996c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 672, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 672, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 335, + 672, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {M G C L} = \\sum_ {i \\in \\{H R, M s, S A R \\}} \\mathcal {L} _ {F G C L} \\left(F _ {i}, F _ {i} ^ {\\prime}\\right) \\tag {9} \\\\ + \\mathcal {L} _ {F G C L} \\left(F _ {\\mathbf {f u s}}, F _ {\\mathbf {f u s}} ^ {\\prime}\\right). \\\\ \\end{array}", + "image_path": "1e92ce500fc47b32fd7a2f42b44015f0fc5d5b134781e8ccc05afabe8319e67d.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27675" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 80, + 420, + 252 + ], + "blocks": [ + { + "bbox": [ + 198, + 71, + 273, + 80 + ], + "lines": [ + { + "bbox": [ + 198, + 71, + 273, + 80 + ], + "spans": [ + { + "bbox": [ + 198, + 71, + 273, + 80 + ], + "type": "text", + "content": "SkySense Pre-training" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 80, + 420, + 252 + ], + "lines": [ + { + "bbox": [ + 56, + 80, + 420, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 80, + 420, + 252 + ], + "type": "image", + "image_path": "c1f1a9749b4a465bc9af20b8080df47baee31044b61d218e87a5f095f7627bc5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 310 + ], + "type": "text", + "content": "Figure 3. Overview of SkySense pre-training and downstream usage. SkySense employs data augmentations on the input and then feeds the augmented data into the student and teacher networks respectively. Multi-Granularity Contrastive Learning and Cross-Modal Alignment are proposed to pre-train the overall network. The region-specific prototype set " + }, + { + "bbox": [ + 46, + 255, + 547, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 255, + 547, + 310 + ], + "type": "text", + "content": " is learned on the student branch and it is frozen for downstream usage. Enhancing feature with " + }, + { + "bbox": [ + 46, + 255, + 547, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 255, + 547, + 310 + ], + "type": "text", + "content": " is optional. After pre-training, we adopt the parameters of the teacher branch for downstream tasks. Each pre-trained module can be used alone or combined with the others, with the chosen ones either frozen or fine-tuned." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 425, + 82, + 542, + 250 + ], + "blocks": [ + { + "bbox": [ + 451, + 72, + 515, + 80 + ], + "lines": [ + { + "bbox": [ + 451, + 72, + 515, + 80 + ], + "spans": [ + { + "bbox": [ + 451, + 72, + 515, + 80 + ], + "type": "text", + "content": "Downstream Usage" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 82, + 542, + 250 + ], + "lines": [ + { + "bbox": [ + 425, + 82, + 542, + 250 + ], + "spans": [ + { + "bbox": [ + 425, + 82, + 542, + 250 + ], + "type": "image", + "image_path": "7a7375c35a02e1959e57cc9aef9624442c7d87c33e4a1254179990279299f272.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": "Cross-Modal Alignment. The heterogeneity of multi-modal data poses a challenge for effective multi-modal feature fusion. We address this issue by adopting multi-modal contrastive loss " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MMCL}" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": " [35] to form the alignment loss " + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {align }}" + }, + { + "bbox": [ + 47, + 319, + 287, + 380 + ], + "type": "text", + "content": ", as shown in Eq. (10)," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 387, + 287, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 387, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 100, + 387, + 287, + 413 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {i \\neq j} \\mathcal {L} _ {\\text {M M C L}} \\left(F _ {i}, F _ {j}\\right), \\tag {10}", + "image_path": "2873e4f75370905522e292395d97cdee081caaa919a3a592aaa62319323bc750.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 416, + 231, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 416, + 231, + 429 + ], + "spans": [ + { + "bbox": [ + 127, + 416, + 231, + 429 + ], + "type": "interline_equation", + "content": "i, j \\in \\{H R, M s, S A R \\}.", + "image_path": "5ebde691acc41869cab090d02a213271407aeecd4ad3de03cb10c9fb58cb66b2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MMCL}" + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "text", + "content": " maximizes the similarity of cross-modal features from the same geo-location, while minimizing it otherwise. Cross-modal alignment is performed on the student branch. Unsupervised Geo-Context Prototype Learning. Different regions characterize distinct geographic landscapes and seasonal dynamics [30, 32] due to disparities in topography and climate. Prior arts have shown that the enlarged context can benefit the RSI interpretations [11, 24, 32, 40]. In this work, " + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "text", + "content": " captures rich spatiotemporal clues for a small area. By clustering on numerous " + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "text", + "content": ", higher-level regional semantics are obtained as implicit geo-knowledge for a vast geo-spatial scope (see Fig. 5). Thus, we propose Geo-Context Prototype Learning to unsupervisedly extract regional geo-context from " + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 438, + 287, + 606 + ], + "type": "text", + "content": " during pre-training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": "We divide the globe into " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "N_{R}" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": " regions and initialize a region-specific prototype set " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N_R \\times N_p \\times d}" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": ". Each prototype is learned from " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": ". We leverage the geo-location of the RSI to retrieve the regional subset " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r \\in \\mathbb{R}^{N_p \\times d}" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": ". Then, we calculate the cosine similarity matrix " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{N_S \\times N_p}" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r" + }, + { + "bbox": [ + 46, + 607, + 287, + 678 + ], + "type": "text", + "content": " as in Eq. (11)," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 684, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 684, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 125, + 684, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\frac {F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}} \\cdot \\mathcal {P} _ {r} ^ {\\mathrm {T}}}{\\| F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}} \\| \\| \\mathcal {P} _ {r} \\|}, \\tag {11}", + "image_path": "3344ed02417656ef2dcb6d72eae7f9a81fdbd28847607dfaab3797a1a6b362bb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "content": "We utilize the Sinkhorn-Knopp algorithm [6] on " + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "content": " to find the optimal assignment matrix " + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N_S \\times N_P}" + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "inline_equation", + "content": "F_{\\mathbf{fus}}^{\\mathrm{mm}}" + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "content": " and the prototypes. This algorithm introduces the uniform distribution constraint to avoid trivial solution while achieving maximal similarity possible. We then use " + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "content": " to generate an update value for current sample's corresponding " + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r" + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{P}_r}" + }, + { + "bbox": [ + 305, + 319, + 545, + 403 + ], + "type": "text", + "content": ", as shown in Eq. (12)," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 420, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 420, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 394, + 420, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathcal {P} _ {r}}} = \\mathbf {S} ^ {\\mathrm {T}} F _ {\\mathbf {f u s}} ^ {\\mathrm {m m}}. \\tag {12}", + "image_path": "00b3ba8eaf71c95453af55273a19f57afed1d46c45a790804a0ace8910f863a7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 444, + 544, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 544, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 544, + 468 + ], + "type": "text", + "content": "Afterwards, we update " + }, + { + "bbox": [ + 305, + 444, + 544, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r" + }, + { + "bbox": [ + 305, + 444, + 544, + 468 + ], + "type": "text", + "content": " through EMA [26] as in Eq. (13), where " + }, + { + "bbox": [ + 305, + 444, + 544, + 468 + ], + "type": "inline_equation", + "content": "m\\in [0,1)" + }, + { + "bbox": [ + 305, + 444, + 544, + 468 + ], + "type": "text", + "content": " is a momentum coefficient," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 473, + 545, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 473, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 370, + 473, + 545, + 486 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {r} \\leftarrow m \\mathcal {P} _ {r} + (1 - m) \\overline {{\\mathcal {P} _ {r}}}. \\tag {13}", + "image_path": "17705aa62fcf8f7be37e7f9d5243a9510f318b0bd7cd3e9464965e6c625cd1f9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": "Each " + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_r" + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": " is updated during pre-training and used as the fixed geo-context for downstream tasks. Geo-Context Prototype Learning is only conducted on the student branch. It extracts generalized region-aware representations from numerous RSI within a consistent region, offering a complementary clue to enhance the feature of a single RSI." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "content": "As Geo-Context Prototype Learning is incorporated without an explicit loss term, our pre-training objective is shown in Eq. (14), where " + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "content": " are trade-off weights," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 370, + 610, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 610, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 370, + 610, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha \\mathcal {L} _ {M G C L} + \\beta \\mathcal {L} _ {\\text {a l i g n}}. \\tag {14}", + "image_path": "904be58a5093bc817df25c7ef603ce0448d25b511bb7881faa898a04eb01abac.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 633, + 387, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 387, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 387, + 647 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Fig. 1 demonstrates SkySense's superior performance in all test scenarios. We conduct experiments on 16 datasets, covering different modalities and tasks, to ensure a comprehensive assessment. The right side of Fig. 3 shows how to apply SkySense to different tasks. Each pre-trained module is" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27676" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 250, + 231 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 250, + 231 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 250, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 250, + 231 + ], + "type": "table", + "html": "
ModelPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [2]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [46]ICCV'21-57.2089.0329.4/39.8
SatMAE [17]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo† [55]TGRS'22-67.2091.27-
RVSA [62]TGRS'2234.3/44.464.49--
BFM† [8]Arxiv'23--92.12-
TOV [57]JSTARS'2332.1/37.866.2492.03-
SSL4EO [65]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [48]TGRS'2336.4/43.566.2191.86-
CACo [45]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS† [63]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [47]ICCV'2336.7/45.666.6291.85-
Scale-MAE [50]ICCV'2334.0/41.765.7791.54-
SkySense-39.7/46.570.9193.9933.1/46.2
", + "image_path": "5cef1cb98a993408b3e4ff4b41bd58744e10a5f756aa6e8c28492220df32e88c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 92, + 233, + 204, + 243 + ], + "lines": [ + { + "bbox": [ + 92, + 233, + 204, + 243 + ], + "spans": [ + { + "bbox": [ + 92, + 233, + 204, + 243 + ], + "type": "text", + "content": "(a) Semantic segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 255, + 72, + 397, + 230 + ], + "blocks": [ + { + "bbox": [ + 255, + 72, + 397, + 230 + ], + "lines": [ + { + "bbox": [ + 255, + 72, + 397, + 230 + ], + "spans": [ + { + "bbox": [ + 255, + 72, + 397, + 230 + ], + "type": "table", + "html": "
ModelHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [2]67.4065.6548.15
SatMAE [17]70.8965.6646.55
RingMo†[55]75.90-46.21
RVSA [62]73.2271.0547.04
BFM†[8]-73.62-
TOV [57]70.1666.3349.62
SSL4EO [65]64.8261.2349.37
CMID [48]75.1166.3750.58
CACo [45]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [47]72.8467.6749.69
Scale-MAE [50]73.8166.4748.31
SkySense78.7374.2754.57
", + "image_path": "ff9d77ad6b0a06f60000593f1cd572a8299a8b168c358463624154a25480ef6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 279, + 231, + 369, + 240 + ], + "lines": [ + { + "bbox": [ + 279, + 231, + 369, + 240 + ], + "spans": [ + { + "bbox": [ + 279, + 231, + 369, + 240 + ], + "type": "text", + "content": "(b) Object detection results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 400, + 70, + 549, + 232 + ], + "blocks": [ + { + "bbox": [ + 400, + 70, + 549, + 232 + ], + "lines": [ + { + "bbox": [ + 400, + 70, + 549, + 232 + ], + "spans": [ + { + "bbox": [ + 400, + 70, + 549, + 232 + ], + "type": "table", + "html": "
ModelLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [2]78.1946.2613.6/16.7
SeCo [46]90.1447.6713.9/16.0
SatMAE [17]87.6552.7614.8/16.2
RingMo† [55]91.86--
RVSA [62]90.86--
SpectralGPT† [28]-54.29-
MATTER† [1]-59.37-
DINO-MC [66]-52.7014.5/15.6
SSL4EO [65]89.0535.0812.3/17.5
CMID [48]91.72--
CACo [45]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [47]91.7359.82-
Scale-MAE [50]92.07--
SkySense92.5860.0615.4/18.0
", + "image_path": "fd0478a527524239d8c8d83f5bf391dacec752e046f9ad0f3d3c9387f98ffb2a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 425, + 233, + 517, + 243 + ], + "lines": [ + { + "bbox": [ + 425, + 233, + 517, + 243 + ], + "spans": [ + { + "bbox": [ + 425, + 233, + 517, + 243 + ], + "type": "text", + "content": "(c) Change detection results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 55, + 281, + 280, + 463 + ], + "blocks": [ + { + "bbox": [ + 46, + 252, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 276 + ], + "type": "text", + "content": "Table 2. Results of semantic segmentation, object detection and change detection. " + }, + { + "bbox": [ + 46, + 252, + 547, + 276 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 252, + 547, + 276 + ], + "type": "text", + "content": " means the code and weights are not released until November 11th, 2023, thus we report the metrics from the paper. - means the task is not supported or the value is unavailable in the paper." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 281, + 280, + 463 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 280, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 280, + 463 + ], + "type": "table", + "html": "
ModelSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [2]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [46]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [17]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo†[55]96.90/98.3494.25/95.67--
RVSA [62]97.03/98.5093.93/95.69--
DINO-MC [66]--84.20/88.7560.16/83.49
TOV [57]95.16/97.0990.97/93.79--
SSL4EO [65]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [48]96.11/97.7994.05/95.53--
CACo [45]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMAT[21]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [47]95.47/97.0992.73/94.6486.30/--
Scale-MAE [50]96.44/97.5892.63/95.04--
SkySense97.68/98.6094.85/96.3288.67/92.0964.38/87.27
", + "image_path": "0907e007c7dd1305360c51222426b4b9ef6858cac758507f1467ffb222309db8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 471, + 233, + 481 + ], + "lines": [ + { + "bbox": [ + 100, + 471, + 233, + 481 + ], + "spans": [ + { + "bbox": [ + 100, + 471, + 233, + 481 + ], + "type": "text", + "content": "Table 3. Scene classification results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 495, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 533 + ], + "type": "text", + "content": "designed to allow for combined or individual use, with the flexibility to be either frozen or fine-tuned as needed. More details are included in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 539, + 206, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 206, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 206, + 552 + ], + "type": "text", + "content": "4.1. Pre-training Implementation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "The model is pre-trained with a batch size of 240 samples, distributed over 80 A100-80GB GPUs. For HSROIs, we apply data augmentations including multi-crop [6], Gaussian blur, solarization [23], etc. As for TMsiI and TSARI, we randomly select a fixed-sized sequence from the original one and perform random disturbances on the RSI acquisition date. We employ the huge version of the Swin Transformer (Swin-H) [41] as the spatial encoder of HSROIs, for its design efficiency in minimizing computational costs for high-resolution imagery [74]. RSI from TMsiI or TSARI is processed with corresponding ViT-L [20]. For Geo-Context Prototype Learning, we divide the globe into 4096 regions, each containing 100 prototypes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 283, + 499, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 499, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 499, + 296 + ], + "type": "text", + "content": "4.2. Performance on Single-Modal Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "content": "We evaluate SkySense on 4 representative single-modal tasks. All experiments are conducted using consistent finetuning settings for fairness. The supplementary materials include implementation details, visualization results and additional experiments on frozen backbone tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "text", + "content": "Semantic Segmentation. We adopt Dyna.-Pla. [59], iSAID [67], Potsdam [52] and Dyna.-S2 [59] for the segmentation experiment. They are chosen considering factors such as spatial resolution, spectrum and category type. UperNet [69] serves as the segmentation head. For Dyna.-Pla. and Dyna.-S2 datasets, we report mIoU results on official validation and test sets. For iSAID and Potsdam, we follow the settings of [55]. As depicted in Tab. 2a, SkySense has achieved the SOTA performance on all four segmentation datasets. On average, it surpasses the previous SOTA by an impressive improvement of " + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "inline_equation", + "content": "1.86\\%" + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 493, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 547, + 635 + ], + "type": "text", + "content": "Horizontal & Oriented Object Detection. We employ the widely recognized DIOR [36] dataset for Horizontal Object Detection and its enhanced version DIOR-R [15], along with FAIR1M [56], for Oriented Object Detection. All datasets consist of optical RGB images. Faster RCNN [51] and Oriented RCNN [37] are used for the experiment, following the setup of [55, 62]. SkySense excels on all three datasets (Tab. 2b). Notably, we surpass the second best CMID by " + }, + { + "bbox": [ + 304, + 493, + 547, + 635 + ], + "type": "inline_equation", + "content": "3.99\\%" + }, + { + "bbox": [ + 304, + 493, + 547, + 635 + ], + "type": "text", + "content": " mAP and have achieved the best performance on the FAIR1M v2.01 leaderboard. More importantly, our results are accomplished without using any sophisticated Oriented Detection designs [29, 37, 71]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 636, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 547, + 696 + ], + "type": "text", + "content": "Change Detection. We assess SkySense's Change Detection performance on LEVIR-CD [9], OSCD [18], and Dyna.-S2 [59] datasets. For LEVIR-CD and OSCD, we follow the frameworks of [55, 66] and report the F1 metric. For Dyna.-S2, we utilize the UperNet head since the re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 517, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 517, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 517, + 714 + ], + "type": "text", + "content": "1https://www.gaofen-challenge.com/benchmark (2023.11.17)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27677" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 175, + 165 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 175, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 175, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 175, + 165 + ], + "type": "image", + "image_path": "82f381d2e9a4d11418ca391e0203c6e75fabb22acabb644c7fb6163893446ec1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 169, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 287, + 202 + ], + "type": "text", + "content": "Figure 4. (a) Experiment on fine-tuning using different percentages of training data on the AID dataset. (b) The impact of S1-Ts data under varying cloud coverage conditions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 284, + 166 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 284, + 166 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 284, + 166 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 284, + 166 + ], + "type": "image", + "image_path": "7621b30499eb0385330ac21cdabd3fe9dbc32b8c7235d68f1b767ee5e32ac6ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 209, + 293, + 353 + ], + "blocks": [ + { + "bbox": [ + 48, + 209, + 293, + 353 + ], + "lines": [ + { + "bbox": [ + 48, + 209, + 293, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 293, + 353 + ], + "type": "table", + "html": "
Task & DatasetData Source & Geo-ContextPrevious SOTASkySense
(a) Multi-Modal Seg: Dyna.-MM(i) Planet.45.6 [47]46.5 ↑ 0.9
(ii) Planet. with GCP-47.0
(iii) S243.5 [4]46.2 ↑ 2.7
(iv) Planet. + S2-47.3
(v) Planet. + S2 + S1-47.7
(vi) Planet. + S2 + S1 with GCP-48.2
(b) Multi-Modal Seg: PASTIS-MM(i) S2-static-73.5
(ii) S2-Ts83.4 [58]84.6 ↑ 1.2
(iii) S2-Ts + S1-Ts84.2 [22]84.8 ↑ 0.6
(iv) S2-Ts + GEP-85.8
(v) S2-Ts + S1-Ts + GEP-85.9
(c) Multi-Modal Cls: BEN-MM(i) S183.70 [64]86.25 ↑ 2.55
(ii) S2 + S189.70 [64]92.21 ↑ 2.51
", + "image_path": "d69bdae501a8d19f6fe9bacc472b26057948978c28fde02d4908062ce7a9e58f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 360, + 258, + 371 + ], + "lines": [ + { + "bbox": [ + 76, + 360, + 258, + 371 + ], + "spans": [ + { + "bbox": [ + 76, + 360, + 258, + 371 + ], + "type": "text", + "content": "Table 4. Fine-tuning results on multi-modal tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 386, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 470 + ], + "type": "text", + "content": "ported semantic change segmentation (SCS) score is calculated from the segmentation results [59]. Both results from the validation and test sets of Dyna.-S2 are presented. The remarkable generalization ability of SkySense is evident in the consistent improvements shown in Tab. 2c. Unlike CACo [45], which shows proficiency mainly on the Dyna.-S2 validation set, our model excels across all datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 471, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 288, + 662 + ], + "type": "text", + "content": "Scene Classification. We utilize four scene classification datasets: AID [68] and RESISC-45 [14] with static RGB images, BEN-S2 [53] with static multispectral images, and fMoW-S2 [17] with temporal multispectral images. The training ratio (TR) follows [17, 46, 55]. We use a linear classifier head for experiment. For AID and RESISC-45, we report Overall Accuracy (OA), for BEN-S2 we report mAP, and for fMoW-S2 we report both Top-1 and Top-5 Accuracy. SkySense overly outperforms competitive baselines and achieves the best results on all datasets (Tab. 3). Additionally, with limited labeled data on the AID dataset, SkySense consistently outperforms CMID, ScaleMAE, and random initialization, with a " + }, + { + "bbox": [ + 46, + 471, + 288, + 662 + ], + "type": "inline_equation", + "content": "4.17\\%" + }, + { + "bbox": [ + 46, + 471, + 288, + 662 + ], + "type": "text", + "content": " higher OA than the second best Scale-MAE using only " + }, + { + "bbox": [ + 46, + 471, + 288, + 662 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 471, + 288, + 662 + ], + "type": "text", + "content": " training data (Fig. 4a). These results highlight the robustness and generalization ability of SkySense's pre-trained features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 237, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 237, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 237, + 682 + ], + "type": "text", + "content": "4.3. Performance on Multi-Modal Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Multi-Modal Segmentation: Time-insensitive Land Cover Mapping. We employ the Dyna.-MM dataset [59]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "for fine-tuning and report mIoU on the official test set. The dataset comprises HSROIs from PlanetFusion (Planet.), multispectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). We use a simple UperNet head. As shown in Tab. 4a, SkySense achieves the best results in single-modal scenarios (i) and (iii), clearly outperforming the previous SOTA by roughly " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " mIoU. Moreover, combining all three modalities as (v) further improves the mIoU by " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " compared to (i). Notably, without bells and whistles, SkySense ranks No.1 on the challenging DynamicEarthNet leaderboard2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 207, + 546, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 546, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 546, + 339 + ], + "type": "text", + "content": "Multi-Modal Segmentation: Time-sensitive Crop Mapping. We evaluate SkySense's fine-tuning result on the PASTIS-MM dataset, an enhanced version of PASTIS-R [22]. PASTIS-MM includes HSROIs from Google Earth Pro (GEP), TMsi from Sentinel-2 (S2-Ts), and TSARI from Sentinel-1 (S1-Ts). We use a naive FCN head [42] and report the OA from the official five-fold validation on PASTIS-MM dataset. In Tab. 4b, comparing S2-Ts and static multispectral data (S2-static), we observe a significant " + }, + { + "bbox": [ + 304, + 207, + 546, + 339 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 304, + 207, + 546, + 339 + ], + "type": "text", + "content": " OA increase, highlighting the importance of incorporating temporal clue for crop mapping." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "content": "Furthermore, both (ii) and (iii) exceed the performance of the previous SOTA, affirming the superior capabilities of SkySense. When more modalities are added as (ii), (iv), and (v), the OA increases accordingly. However, integrating Sentinel-1 data yields no substantial improvement, presumably because of the cloud-free imagery from PASTIS-MM dataset. To further investigate, we compare OA using S2-Ts data at different cloud ratios with and without Sentinel-1. Fig. 4b illustrates that the performance difference between utilizing and foregoing Sentinel-1 data becomes more pronounced with an increasing cloud ratio. Specifically, when the cloud ratio exceeds " + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "content": ", the result of using Sentinel-1 outperforms its counterpart by " + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "content": ". This highlights the importance of SAR data in situations with cloud coverage and rainfall." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "content": "Multi-Modal Scene Classification. We utilize the BENMM [54] dataset for evaluating the multi-modal scene classification task. This dataset includes both Sentinel-1 (S1) and Sentinel-2 (S2) imagery. The evaluation protocol from DeCUR [64] is followed, and we report the mAP metric of fine-tuning with " + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "content": " training data. In Tab. 4c, both (i) and (ii) significantly outperform the previous SOTA by more than " + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "content": " mAP. Furthermore, the inclusion of Sentinel-2 imagery greatly enhances performance compared to using Sentinel-1 imagery alone." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "content": "All these results show a notable gain for the tasks using multi-modal data, affirming the necessity of SkySense's multi-modal pre-training from one perspective." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 548, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 548, + 714 + ], + "type": "text", + "content": "2https://codalab.lisn.upsaclay.fr/competitions/2882#results(2023.11.17)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27678" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 171, + 132 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 171, + 132 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 171, + 132 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 171, + 132 + ], + "type": "table", + "html": "
Pre-trainingiSAIDfMoW-S2
mIoUTop-5 Acc
Simple Ver.68.9885.69
SkySense70.91 ↑ 1.9387.27 ↑ 1.58
", + "image_path": "8cdc2207c7c369f46a059ceedb7944ba5887bc078d74902033f866e183a71449.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 178, + 70, + 276, + 144 + ], + "blocks": [ + { + "bbox": [ + 102, + 135, + 113, + 142 + ], + "lines": [ + { + "bbox": [ + 102, + 135, + 113, + 142 + ], + "spans": [ + { + "bbox": [ + 102, + 135, + 113, + 142 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 70, + 276, + 144 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 276, + 144 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 276, + 144 + ], + "type": "table", + "html": "
Pre-trainingDyna.-MM
mIoU
Baseline42.2
+ MGCL44.4 ↑ 2.2
+ MM47.0 ↑ 2.6
+ CMA47.7 ↑ 0.7
+ GCPL48.2 ↑ 0.5
", + "image_path": "eef9906bdb89fa3294db1ed00e928e37e87d57415790a4850c0d6ab5f8b25de1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 221, + 146, + 233, + 155 + ], + "lines": [ + { + "bbox": [ + 221, + 146, + 233, + 155 + ], + "spans": [ + { + "bbox": [ + 221, + 146, + 233, + 155 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 164, + 286, + 187 + ], + "lines": [ + { + "bbox": [ + 47, + 164, + 286, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 286, + 187 + ], + "type": "text", + "content": "Table 5. (a) Discussion on multi-modal pre-training effectiveness. (b) Ablation study on the pre-training design." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 198, + 221, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 221, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 221, + 210 + ], + "type": "text", + "content": "5. Discussions & Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 219, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 434 + ], + "type": "text", + "content": "Multi-modal Pre-training Effectiveness. In addition to confirming the effectiveness of using multi-modal data in downstream tasks, we investigate the impact of multi-modal pre-training on single-modal tasks, compared with pretraining on fewer modalities. We conduct experiments on iSAID for static HSROI segmentation and fMoW-S2 for temporal multispectral classification. Two versions of the pre-trained model are tested: a simple version pre-trained only with optical imagery (HSROIs, TMsI), and SkySense, which includes HSROIs, TMsI, and TSARI for pre-training. The rest of the settings remain consistent. The results in Tab. 5a show that SkySense consistently outperforms the simple version, suggesting that the introduction of SAR data benefits representation learning of other modalities. This may attribute to the implicit clue brought by SAR data through Cross-Modal Alignment. It provides another perspective on the necessity of SkySense's multi-modal pretraining." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 435, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 287, + 544 + ], + "type": "text", + "content": "What does Geo-Context Prototype (GCP) Learn? We utilize the Dyna.-MM dataset for experiment as it contains diverse geo-locations worldwide. For the segmentation task in Tab. 4a, adding GCP in downstream tasks leads to a further gain of " + }, + { + "bbox": [ + 47, + 435, + 287, + 544 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 47, + 435, + 287, + 544 + ], + "type": "text", + "content": " mIoU compared to the strong multi-modal baseline (v). Moreover, a comparison between (i) and (ii) shows a " + }, + { + "bbox": [ + 47, + 435, + 287, + 544 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 47, + 435, + 287, + 544 + ], + "type": "text", + "content": " mIoU improvement using GCP for the single-modal task. It demonstrates GCP's consistent performance gain in single- and multi-modal scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": "In Fig. 5, we visualize the learned prototypes on the Map by calculating the pre-trained feature of each pixel and assigning the most similar prototype to it. A comparison with the ESRI LandCover Map [34] reveals GCP's promising results in segmenting different areas. Moreover, GCP exhibit fine-grained advantage, as shown in the middle of Fig. 5. The prototypes learned from unsupervised clustering segment cropland within the town, which is overlooked by the LandCover Map. Notably, the visualization shares the same spatial resolution with ESRI LandCover Map." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Design of Pre-training. Tab. 5b presents the ablation study to assess our pre-training design, namely Multi-Granularity Contrastive Learning (MGCL), Multi-Modal (MM) integration, Cross-Modal Alignment (CMA) and Geo-Context" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 547, + 200 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 547, + 200 + ], + "type": "image", + "image_path": "cafeb5289967c67afa0371759e6485cae4f30adfd0b88321acbfeae07de0c87c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 204, + 545, + 238 + ], + "lines": [ + { + "bbox": [ + 305, + 204, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 545, + 238 + ], + "type": "text", + "content": "Figure 5. Comparison between (a) ESRI LandCover Map and (b) Geo-Context Prototype. The visualization process of Geo-Context Prototype is illustrated in the upper part of this figure." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 248, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 283 + ], + "type": "text", + "content": "Prototype Learning (GCPL). We utilize the Dyna.-MM dataset for the experiment and report mIoU metric on the official test set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "text", + "content": "Initially, we utilize a single-modal version, using " + }, + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "inline_equation", + "content": "g_{HR}" + }, + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "text", + "content": " spatial encoder and HSROIs for pre-training. The training settings are kept the same as described in Sec. 4.1. The results show that MGCL leads to a notable improvement compared to a simple baseline [7]. Then we integrate further multi-modal data (i.e., TMsiI and TSARI) into the pretraining and downstream evaluation. This effectively improves the performance on the test set to " + }, + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "inline_equation", + "content": "47.0\\%" + }, + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "text", + "content": " mIoU, validating the necessity of multi-modal pre-training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "text", + "content": "CMA is another necessary design for SkySense's pretraining, which explicitly pulls features from different modalities together, encouraging cross-modal interactions. The results show that the incorporation of modal alignment leads to " + }, + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "text", + "content": " mIoU improvement. Finally, we introduce GCPL, which learns complementary regional context clues to facilitate downstream tasks and further pushes the very strong performance to " + }, + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "inline_equation", + "content": "48.2\\%" + }, + { + "bbox": [ + 304, + 392, + 545, + 487 + ], + "type": "text", + "content": " mIoU." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 498, + 462, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 462, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 462, + 510 + ], + "type": "text", + "content": "6. Conclusion & Future Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 518, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 649 + ], + "type": "text", + "content": "In this paper, we present SkySense, a large-scale MM-RSFM for interpretation of EO imagery. SkySense allows using its modules flexibly to accommodate different scenarios and consistently outperforms other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope SkySense will inspire further research on MM-RSFM and its release may contribute to sustainable innovations thriving in the Remote Sensing community. As part of our future work, we plan to incorporate the language modality, thereby extending SkySense's applications to more EO tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 660, + 410, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 660, + 410, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 660, + 410, + 673 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 679, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 679, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 679, + 545, + 712 + ], + "type": "text", + "content": "This work was in part supported by the National Natural Science Foundation of China under Grants 42030102 and 42371321." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27679" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "type": "text", + "content": "[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 237 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 288, + 293 + ], + "type": "text", + "content": "[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinand, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 349 + ], + "type": "text", + "content": "[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 350, + 288, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 406 + ], + "type": "text", + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 497 + ], + "type": "text", + "content": "[8] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 532 + ], + "type": "text", + "content": "[9] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 578, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 288, + 643 + ], + "type": "text", + "content": "[11] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 644, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 288, + 679 + ], + "type": "text", + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Mocov2: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "text", + "content": "[13] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[14] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "text", + "content": "[15] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[16] Mingmin Chi, Antonio Plaza, Jon Atli Benediktsson, Zhongyi Sun, Jinsheng Shen, and Yangyong Zhu. Big data for remote sensing: Challenges and opportunities. Proceedings of the IEEE, 104(11):2207-2219, 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "text", + "content": "[17] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 277, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 342 + ], + "type": "text", + "content": "[18] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 344, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 398 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 400, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 465 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "text", + "content": "[21] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 514, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 567 + ], + "type": "text", + "content": "[22] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 569, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 635 + ], + "type": "text", + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 636, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 712 + ], + "type": "text", + "content": "[24] Shaohua Guo, Liang Liu, Zhenye Gan, Yabiao Wang, Wuhao Zhang, Chengjie Wang, Guannan Jiang, Wei Zhang, Ran Yi, Lizhuang Ma, et al. Isdnet: Integrating shallow and deep networks for efficient ultra-high resolution segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4361-4370, 2022. 2, 4, 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27680" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "text", + "content": "[28] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Gamba Paolo, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral foundation model. arXiv preprint arXiv:2311.07113, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "text", + "content": "[29] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 923-932, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "type": "text", + "content": "[30] Jingliang Hu, Lichao Mou, and Xiao Xiang Zhu. Unsupervised domain adaptation using a teacher-student network for cross-city classification of sentinel-2 images. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 43:1569–1574, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "text", + "content": "[31] Lanqing Huang, Bin Liu, Boying Li, Weiwei Guo, Henhao Yu, Zenghui Zhang, and Wenxian Yu. Opersarship: A dataset dedicated to sentinel-1 ship interpretation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(1):195–208, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "text", + "content": "[32] Xin Huang, Yihong Song, Jie Yang, Wenrui Wang, Huiqun Ren, Mengjie Dong, Yujin Feng, Haidan Yin, and Jiayi Li. Toward accurate mapping of 30-m time-series global impervious surface area (gisa). International Journal of Applied Earth Observation and Geoinformation, 109:102787, 2022. 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 504, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 635 + ], + "type": "text", + "content": "[33] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Nyirjesy Gabby, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mukkavilli, Devyani Lambate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence. arXiv preprint arXiv:2310.18660, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "text", + "content": "[34] Krishna Karra, Caitlin Kontgis, Zoe Statman-Weil, Joseph C Mazzariello, Mark Mathis, and Steven P Brumby. Global land use/land cover with sentinel 2 and deep learning. In 2021 IEEE international geoscience and remote sensing symposium IGARSS, pages 4704-4707. IEEE, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[35] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "text", + "content": "[36] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "text", + "content": "[37] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "text", + "content": "[38] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. Mcanet: A joint semantic segmentation framework of optical and sar images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106: 102638, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "text", + "content": "[39] Yansheng Li, Wei Chen, Xin Huang, Gao Zhi, Siwei Li, Tao He, and Yongjun Zhang. Mfvnet: a deep adaptive fusion network with multiple field-of-views for remote sensing image semantic segmentation. Science China Information Sciences, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "text", + "content": "[40] Yinhe Liu, Sunan Shi, Junjue Wang, and Yanfei Zhong. Seeing beyond the patch: Scale-adaptive semantic segmentation of high-resolution remote sensing imagery based on reinforcement learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16868-16878, 2023. 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[42] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[43] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "text", + "content": "[44] Rose M Rustowicz, Robin Cheong, Lijing Wang, Stefano Ermon, Marshall Burke, and David Lobell. Semantic segmentation of crop type in africa: A novel dataset and analysis of deep learning methods. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 75-82, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "text", + "content": "[45] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[46] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27681" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[47] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[48] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 251 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 253, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 329 + ], + "type": "text", + "content": "[50] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "text", + "content": "[52] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 411, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 477 + ], + "type": "text", + "content": "[53] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 479, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 553 + ], + "type": "text", + "content": "[54] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "text", + "content": "[55] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[56] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[57] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learn" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ing. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "text", + "content": "[58] Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10418-10428, 2023. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 150, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 547, + 227 + ], + "type": "text", + "content": "[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "text", + "content": "[60] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N van Rijn, Holger H Hoos, Fabio Del Frate, et al. Artificial intelligence to advance earth observation: a perspective. arXiv preprint arXiv:2305.08413, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "text", + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "text", + "content": "[62] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[63] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Graham, Chenying Liu, Zhitong Xiong, and Xiao Xiang Zhu. Decur: decoupling common & unique representations for multimodal self-supervision. arXiv preprint arXiv:2309.05300, 2023. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[66] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 582, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 657 + ], + "type": "text", + "content": "[67] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[68] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27682" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 552 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[69] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "text", + "content": "[70] Binbin Yang, Xinchi Deng, Han Shi, Changlin Li, Gengwei Zhang, Hang Xu, Shen Zhao, Liang Lin, and Xiaodan Liang. Continual object detection via prototypical task correlation guided gating mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9255-9264, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[71] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13354-13363, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 272 + ], + "type": "text", + "content": "[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "text", + "content": "[73] Hankui K Zhang, David P Roy, and Dong Luo. Demonstration of large area land cover classification with a one-dimensional convolutional neural network applied to single pixel temporal metric percentiles. Remote Sensing of Environment, 295:113653, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "type": "text", + "content": "[74] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998-3008, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 287, + 440 + ], + "type": "text", + "content": "[75] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 442, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 495 + ], + "type": "text", + "content": "[76] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "type": "text", + "content": "[77] Stefano Zorzi, Shabab Bazrafkan, Stefan Habenschuss, and Friedrich Fraundorfer. *Polyworld: Polygonal building extraction with graph neural networks in satellite images*. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1848-1857, 2022. 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27683" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_content_list.json b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f61f686ecafcc175fb9f39d471945362bf88ecbb --- /dev/null +++ b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_content_list.json @@ -0,0 +1,1643 @@ +[ + { + "type": "text", + "text": "SleepVST: Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers", + "text_level": 1, + "bbox": [ + 187, + 128, + 782, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonathan F. Carter $^{1,2}$ João Jorge $^{2}$ Oliver Gibson $^{2}$ Lionel Tarassenko $^{1}$", + "bbox": [ + 207, + 214, + 759, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Biomedical Engineering, University of Oxford $^{2}$ Oxehealth Ltd., Oxford", + "bbox": [ + 251, + 244, + 720, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advances in camera-based physiological monitoring have enabled the robust, non-contact measurement of respiration and the cardiac pulse, which are known to be indicative of the sleep stage. This has led to research into camera-based sleep monitoring as a promising alternative to \"gold-standard\" polysomnography, which is cumbersome, expensive to administer, and hence unsuitable for longer-term clinical studies. In this paper, we introduce SleepVST, a transformer model which enables state-of-the-art performance in camera-based sleep stage classification (sleep staging). After pre-training on contact sensor data, SleepVST outperforms existing methods for cardiorespiratory sleep staging on the SHHS and MESA datasets, achieving total Cohen's kappa scores of 0.75 and 0.77 respectively. We then show that SleepVST can be successfully transferred to cardio-respiratory waveforms extracted from video, enabling fully contact-free sleep staging. Using a video dataset of 50 nights, we achieve a total accuracy of $78.8\\%$ and a Cohen's $\\kappa$ of 0.71 in four-class video-based sleep staging, setting a new state-of-the-art in the domain.", + "bbox": [ + 75, + 358, + 472, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 693, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate sleep monitoring is critical for the diagnosis of sleep disorders and the discovery of novel treatments and biomarkers. Poor sleep is broadly linked with a number of health conditions, including cardiovascular diseases such as diabetes and hypertension [15]. Additionally, there are links between specific sleep stages and neurodegenerative conditions, such as a decrease in non-rapid-eye-movement stage 3 (N3) sleep with Alzheimer's disease [23].", + "bbox": [ + 75, + 719, + 468, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Overnight video polysomnography (vPSG), the \"gold-standard\" [52] technique for sleep monitoring, requires the", + "bbox": [ + 76, + 842, + 468, + 873 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "use of a large number of contact sensors including electrodes placed on the scalp (EEG), near the eyes (EOG), and under the chin (EMG) of the patient. Human experts (polysomnographers) must then review the recorded data, classifying the patient's sleep into five stages at 30-second intervals (epochs) and annotating other important events such as leg movements and apnoeas. This manual process often takes multiple hours to complete.", + "bbox": [ + 496, + 327, + 893, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Experts primarily rely on characteristic patterns in brain activity measured from the EEG to classify sleep stages [20]. However, sleep stage information is also encoded in measures of autonomic activity, including cardiac [48] and respiratory signals [19], and body movements [68]. This has led to the development of methods for automatically classifying sleep stages from sensors that measure these parameters, such as smartwatches [63].", + "bbox": [ + 496, + 452, + 893, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior work has shown that these physiological parameters can also be measured using video cameras [18, 65, 66], leading to the development of methods for classifying sleep stages entirely from video input [6, 35, 60]. Camera-based methods have particular promise as part of a multipurpose sleep monitoring system. For example, within elderly care settings, in addition to classifying sleep stages, they could also be used to detect specific sleep movement disorders [18], which are linked with conditions such as Parkinson's, and which are typically distinguished via manual video review [51]. They could also be used to detect falls [12], one of the leading causes of injury and mortality amongst elderly individuals.", + "bbox": [ + 496, + 578, + 893, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art video-based sleep staging [6] has used the heart rate (HR) and breathing rate (BR) derived from video to perform sleep staging. However, these rates are typically measured as averages over time, losing higher frequency inter-pulse and inter-breath interval information. EEG- [40] and wearable-based [22] sleep staging methods have both been improved by using raw waveforms as inputs, rather than derived quantities e.g. signal features [37, 45].", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*jcarter@robots.ox.ac.uk", + "bbox": [ + 94, + 887, + 285, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12479", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a915d98d330c93a316f9c2cff731f64bbeacff570a48d1addd7b205a04b1e161.jpg", + "image_caption": [ + "Figure 1. Sleep staging from near-infrared video signals using SleepVST. (a) We first pre-train the model end-to-end on cardiac (heart) and respiratory (breathing) waveforms $x_{HW}(t)$ and $x_{BW}(t)$ derived from the electrocardiogram (ECG) and a thoracic respiratory belt (THX) respectively. (b) After pre-training, we use the model as a frozen feature extractor, applying it to cardio-respiratory waveforms derived from near-infrared (NIR) video to generate sequences of features. When transferring to video data, we additionally use a set of motion features $f_{i}(t)$ derived from an optical flow field $\\pmb{u}(x,y,t)$ as inputs to the classifier C. This approach allows us to utilise much larger contact-sensor datasets to train SleepVST, whilst also enabling the incorporation of motion information when transferring to video data. (c) We evaluate SleepVST using overnight video polysomnography (vPSG) studies, comparing expert-labelled sleep stage sequences $\\hat{y}_{1:T}$ against those generated entirely from near-infrared video $y_{1:T}$ using our method." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 872, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce SleepVST1, a transformer model for sleep stage classification from cardio-respiratory waveforms. First, we pre-train the model on waveforms derived from contact sensors (Fig. 1a), after which it outperforms existing methods for cardio-respiratory sleep staging on two widely used public datasets, highlighting the effectiveness of our transformer architecture.", + "bbox": [ + 75, + 397, + 467, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After pre-training, we then show that SleepVST can be transferred to cardiac and respiratory waveforms measured entirely from near-infrared video, where our method additionally enables us to incorporate motion features derived from the camera data (Fig. 1b). Using this novel combination of inputs available from a near-infrared camera, we obtain state-of-the-art results for video-based sleep staging. Through a series of ablation studies we demonstrate the effectiveness of our approach. Our work further highlights the potential of camera-based sleep monitoring for important clinical applications.", + "bbox": [ + 75, + 503, + 467, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Work", + "text_level": 1, + "bbox": [ + 76, + 684, + 361, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automated sleep staging. To reduce the manual effort of sleep assessment, deep learning methods have been developed for the automatic classification of sleep stages from polysomnography (PSG) recordings, achieving near-human accuracy [39]. Accuracy is commonly measured using Cohen's kappa statistic $(\\kappa, [9])$ , which measures inter-rater agreement between two sources. Expert annotation of sleep stages is a subjective process, making $100\\%$ agreement unachievable [59]. Prior work has shown that sleep scorers agree around $80\\%$ of the time ( $\\kappa = 0.71 - 0.81$ , [11, 24]) when classifying sleep stages according to American Academy of", + "bbox": [ + 75, + 709, + 467, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sleep Medicine (AASM, [20]) scoring rules. These rules divide sleep into five classes: Wake, N1 (light), N2 (intermediate), N3 (deep) and rapid-eye-movement (REM) sleep. However, due to low inter-scorer agreement in identifying N1 [11], prior work has often treated sleep classification as a four-class problem, by merging N1 and N2 into a single class [22, 50].", + "bbox": [ + 504, + 397, + 888, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the drawbacks of conventional polysomnography, prior work has investigated the potential of automatic sleep staging from a reduced number of contact sensors, including the electrocardiogram ([50], $\\kappa = 0.66$ ), and the photoplethysmogram ([44], $\\kappa = 0.65$ ).", + "bbox": [ + 496, + 503, + 888, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera-based physiological measurement. Respiratory measurement is possible from video by measuring the induced movements of the chest and abdomen [65]. This has been achieved through methods such as frame-differencing [2, 54], using dense optical flow magnitudes [32, 33] and by tracking the movement of sparse optical flow vectors [25] across video frames.", + "bbox": [ + 496, + 580, + 888, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera-based measurement of the pulse is possible using methods which measure specific cardiac phenomena [30]. These include remote photoplethysmography (rPPG, [42, 66]), which measures skin colour changes caused by variations in blood volume over the cardiac cycle, and ballistocardiography [4], which measures head movements induced as blood is pumped through the carotid arteries. More recently, deep learning methods have been developed which learn to extract a pulse signal directly from input video frames [7, 26, 70] rather than targeting specific underlying cardiac phenomena, including unsupervised approaches [49, 69].", + "bbox": [ + 496, + 688, + 888, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera-based vital-sign measurement has been validated within clinical settings [55, 62], during overnight", + "bbox": [ + 500, + 869, + 888, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Sleep Video-Signal Transformer.", + "bbox": [ + 94, + 886, + 274, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12480", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "polysomnography recordings [58, 64], and has led to products with regulatory clearance for camera-based measurement of the heart and breathing rate, including the Gili Pro Biosensor [57] and Oxehealth Vital Signs software [56].", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera-based sleep monitoring. Prior work has investigated the binary problem of distinguishing sleep from wake using video cameras, using measures of activity derived from the input frames [10, 33, 46]. Whilst activity information alone can be effective in distinguishing sleep from wake, cardio-respiratory signals carry important additional information for classifying specific sleep stages [19, 48]. The current state-of-the-art approach for video-based sleep staging ([6], $\\kappa = 0.61$ ) combines multiple modalities derived from video: the heart rate, breathing rate and activity information, to perform sleep stage classification. This approach outperforms prior methods which had used only a single video-derived modality: motion [21, 35] or the cardiac pulse [60].", + "bbox": [ + 75, + 152, + 472, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Datasets and Preprocessing", + "text_level": 1, + "bbox": [ + 76, + 378, + 331, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Contact Sensor Data", + "text_level": 1, + "bbox": [ + 76, + 402, + 272, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the SleepVST model, we used the SHHS [43, 71] and MESA [8] contact sensor datasets. These datasets contain PSG recordings and sleep stages annotated according to AASM scoring rules. The cardio-respiratory signals available in each recording include the electrocardiogram (ECG), which measures electrical activity of the heart, and respiratory signals, measured using respiratory bands placed around the abdomen and thorax.", + "bbox": [ + 75, + 426, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the available signals, we seek to produce generic cardiac and respiratory waveforms to train the SleepVST model, preventing the model from learning to use information which may be present in signals derived from contact sensors but not present in those derived from video. This is to mitigate the problem of domain shift [28, 47], where the performance of a model is significantly reduced when applied to a different data distribution.", + "bbox": [ + 75, + 547, + 468, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To produce a cardiac (heart) waveform, we first applied the same filtering steps used by the Pan-Tompkins algorithm [36] to the ECG signal. We resampled the output from the moving-window integration step of the algorithm to $10\\mathrm{Hz}$ , then applied a Butterworth bandpass filter with a passband of $0.66 - 2.8\\mathrm{Hz}$ (40-168 beats per minute or BPM) and Gaussian smoothing. To produce a respiratory (breathing) waveform, we used the thoracic respiratory signal as the input source, as previously used in [3]. This was downsampled to $5\\mathrm{Hz}$ and smoothed using a median filter with a kernel of length 5, to remove artefacts.", + "bbox": [ + 75, + 669, + 468, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Video Data", + "text_level": 1, + "bbox": [ + 76, + 845, + 199, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We used the dataset previously introduced in [6] for our video-based sleep staging experiments. This dataset con", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sists of overnight vPSG recordings from 50 volunteers, with sleep stages annotated according to the AASM guidelines at 30-second intervals (epochs) by a sleep physiologist. Throughout this work, we refer to this as the Oxford Sleep Volunteers (OSV) dataset. Each recording includes $850\\mathrm{nm}$ near-infrared (NIR) video data captured at 20 FPS. An example camera view can be seen in Figure 5. More information on the dataset is provided in $\\S B$ .", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To derive cardiac and respiratory waveforms from the video data, we used components of FDA-cleared Oxhealth Vital Signs software [56]. For each overnight video recording, we used the software to obtain cardiac and respiratory waveforms sampled at $10\\mathrm{Hz}$ and $5\\mathrm{Hz}$ respectively. We then applied the same filtering steps applied to the contact sensor waveforms: Butterworth filtering and Gaussian smoothing to the heart waveforms, and median filtering to the breathing waveforms.", + "bbox": [ + 496, + 212, + 892, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows examples of processed cardio-respiratory waveforms measured from contact sensors and video for a short section of a recording from the OSV dataset. These have additionally been normalised to zero mean and unit variance.", + "bbox": [ + 496, + 348, + 893, + 421 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/13a2e06f84183abdde81392f75867ab555309f312b8bc93e075e587686f53b08.jpg", + "image_caption": [ + "Figure 2. Example (normalized) cardiac and respiratory waveforms, $x_{HW}(t)$ and $x_{BW}(t)$ , derived from contact sensors (blue) and video (orange) from the OSV dataset." + ], + "image_footnote": [], + "bbox": [ + 506, + 438, + 883, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 500, + 671, + 598, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. SleepVST Architecture", + "text_level": 1, + "bbox": [ + 500, + 695, + 714, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3 illustrates the SleepVST model architecture, which turns cardiac and respiratory waveforms into sequences of feature vectors using a transformer encoder. These feature vectors are then used to classify each 30-second sleep epoch. During pre-training, we used a linear layer for classification. When transferring to video data, we trained a new classifier that additionally used motion features derived from video as inputs, discussed further in Section 4.2.", + "bbox": [ + 496, + 718, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patchification. The continuous heart and breathing waveforms (HW and BW) are divided into N windows i.e. patches, drawing on ideas from PatchTST [34]. We used non-overlapping 30-second windows since this is the inter", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12481", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8d940f7e4bb5edf9b1e04902ea003124a51496629d71b795a736b64fcad25298.jpg", + "image_caption": [ + "Figure 3. SleepVST architecture. Each 30-second window of heart $(x_{HW})$ and breathing waveforms $(x_{BW})$ is passed to a patch encoder, which turns them into patch-level features. These features are concatenated and passed to a transformer encoder. During pre-training, a linear layer turns the output feature sequence $z_{o}$ of length N from SleepVST into sleep stage classifications." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 444, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "val at which sleep stages are classified according to AASM guidelines. We denote the patchified inputs as $\\boldsymbol{x}_{HW} \\in \\mathbb{R}^{300 \\times N}$ and $\\boldsymbol{x}_{BW} \\in \\mathbb{R}^{150 \\times N}$ respectively, where $N$ is the sequence length.", + "bbox": [ + 76, + 426, + 468, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encodings. Each waveform patch is normalised to zero mean and unit variance before being passed to encoding layers $E_{HW}$ and $E_{BW}$ , which transform them into feature vectors $\\pmb{z}_{HW} \\in \\mathbb{R}^{D_{HW} \\times N}$ and $\\pmb{z}_{BW} \\in \\mathbb{R}^{D_{BW} \\times N}$ . The feature vectors from each patch are then concatenated, producing a sequence of feature vectors $\\pmb{z}_i \\in \\mathbb{R}^{(D_{HW} + D_{BW}) \\times N}$ which are passed to a transformer encoder.", + "bbox": [ + 76, + 489, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We used identical, shallow 1D ResNet-style [17] models for $E_{HW}$ and $E_{BW}$ , illustrated in Figure 4. This convolutional encoder design [1] compresses the waveform patches into lower-dimensional feature vectors. This allowed us to keep $D_{HW}$ and $D_{BW}$ small, reducing the computational and memory complexity of the downstream transformer and the likelihood of over-fitting, particularly when transferring the model to the smaller video dataset.", + "bbox": [ + 75, + 597, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This encoder design additionally provides a high degree of flexibility at the input level, such as the ability to use input signals of differing frequencies. In later ablation studies, this enables us to compare the performance of the SleepVST model with simpler variants which use 30-second patches of $1\\mathrm{Hz}$ heart rate $\\pmb{x}_{HR} \\in \\mathbb{R}^{30 \\times N}$ and/or breathing rate $\\pmb{x}_{BR} \\in \\mathbb{R}^{30 \\times N}$ time-series as inputs, rather than underlying waveforms. When using either of these as inputs, we used linear projections [13, 34] for the input encoders, because of the lower dimensionality of $\\pmb{x}_{HR}$ and $\\pmb{x}_{BR}$ . Throughout our experiments we use the same feature dimension i.e. $D_{HW} = D_{BW} = D_{HR} = D_{BR} = 64$ .", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd35a6b7e6e1cf582dc432a7f0dc609018fcec645f9926444d7cf0e0429209a1.jpg", + "image_caption": [ + "Figure 4. Waveform encoder design. Using a series of convolutional layers, the encoder turns sequences of signal patches into sequences of lower-dimensional feature vectors." + ], + "image_footnote": [], + "bbox": [ + 519, + 97, + 874, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer layer. We used a Transformer encoder [61] with sinusoidal position encodings to turn the sequence of input features $\\mathbf{z}_i$ into an output sequence of features $\\mathbf{z}_o$ with an identical feature dimension. Normal sleep exhibits long-range structures, known as sleep cycles [38], which typically last around 90-120 minutes and typically occur four to five times per night. From its superior performance in many sequence modelling tasks, including EEG-based sleep staging [41], we expect that the inductive biases of the transformer model (i.e. the attention mechanism) can enable it to learn to use relative position within a sleep cycle to classify sleep stages.", + "bbox": [ + 496, + 368, + 890, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classification. When training and testing on contact sensor data, we used a linear classification layer $\\mathbf{W}_C \\in \\mathbb{R}^{C \\times (D_{HW} + D_{BW})}$ to transform the sequence of output feature vectors $\\mathbf{z}_o$ into a sequence of sleep stage probabilities $\\mathbf{y} \\in \\mathbb{R}^{C \\times N}$ , where the number of classes $C$ depends on the sleep classification strategy used. In line with prior work, we classified sleep stages into four classes during pretraining: Wake-N1/N2-N3-REM, i.e. $C = 4$ .", + "bbox": [ + 496, + 550, + 890, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model training and inference. In all experiments, we used a sequence length of $N = 240$ windows i.e. two hours. During training, we created batches by sampling sequences of length $N$ with a step of 5 minutes from within each recording. The model was trained to minimise the multiclass cross-entropy loss using the AdamW [27] optimiser with a learning rate of 0.0005 and weight decay of 0.001. Other hyper-parameters, such as the no. encoder layers, are detailed in §D. To apply the model to arbitrary length input sequences during evaluation, we re-applied it at regular intervals. More details on this process are provided in §E.", + "bbox": [ + 496, + 670, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Motion features from Video", + "text_level": 1, + "bbox": [ + 500, + 847, + 748, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following prior work [35], we seek to produce motion features from video which encode for information which is pre", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12482", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dictive of the sleep stage, such as the rate and distribution of body movements [68]. To do so, we define a set of parameterised time-series features that collectively quantify motion over a range of spatial and temporal scales.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optical flow estimation. We first estimated a 2D optical flow field $\\pmb{u}(t,x,y) \\in \\mathbb{R}^2$ over the bed region. Inspired by sleep pose detection work [31], we applied a homography transformation [16] to the individual video frames before estimating the flow field, to achieve camera viewpoint invariance (shown in Figure 5).", + "bbox": [ + 75, + 151, + 467, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d62adc07c5c05ad174f74c1c78cd6c8a43f496c74b8ac3715d3a5d004de9300b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 88, + 255, + 236, + 362 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f43ab59badbe965d2739e0e1735184249d2d76bb15410cb12d4ee8051b20c94.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 246, + 255, + 348, + 363 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35d473d2c27d3b528714d70946bd483a1a99b46eea75f67d05e4d225ce19bb26.jpg", + "image_caption": [ + "(c)", + "Figure 5. Example processing of an NIR video frame from the OSV dataset. (a) Real viewpoint. (b) Virtual viewpoint. (c) Head (H), body (B), and outer (O) bed regions. The distance from the camera to the head region is around $1.5\\mathrm{m}$ ." + ], + "image_footnote": [], + "bbox": [ + 357, + 255, + 459, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature definitions. Given an arbitrary region $\\mathbf{R}$ of the optical flow field, we define the following parametrised signals:", + "bbox": [ + 75, + 468, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv (t; R) = \\max _ {(x, y) \\in R} | \\boldsymbol {u} (t, x, y) | \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 526, + 468, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns (t; R) = \\frac {1}{| R |} \\sum_ {(x, y) \\in R} | \\boldsymbol {u} (t, x, y) | \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 553, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "i.e. the maximum (1) and the average (2) of the optical flow magnitude in the region R at time $t$ . Using these signals, we define two time-series features that quantify motion within a time window $\\Delta$ up to time $t$ :", + "bbox": [ + 75, + 603, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {1} (t; R, \\Delta) = \\sum_ {t - \\Delta} ^ {t} v (t; R) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 676, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {2} (t; R, \\Delta) = \\sum_ {t - \\Delta} ^ {t} s (t; R) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 720, + 468, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Plus two additional features which measure the time elapsed since either signal went above a threshold $\\delta$ :", + "bbox": [ + 75, + 773, + 468, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {3} (t; R, \\delta) = \\sum_ {t - \\tau} ^ {t} 1\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 816, + 325, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad \\tau = \\max _ {t ^ {\\prime} \\leq t} t ^ {\\prime} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 853, + 468, + 882 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c c} \\text {s . t .} & v (t ^ {\\prime}; R) > \\delta \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 886, + 351, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {4} (t; R, \\delta) = \\sum_ {t - \\tau} ^ {t} 1\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 103, + 748, + 142 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad \\tau = \\max _ {t ^ {\\prime} \\leq t} t ^ {\\prime} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 140, + 890, + 170 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l l} \\text {s . t .} & s (t ^ {\\prime}; R) > \\delta \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 172, + 772, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We calculated each feature for all possible combinations of parameters $\\Delta \\in \\{30,300\\}$ seconds, $\\delta \\in \\{0.01,0.1,1\\}$ and $R \\in \\{H,B,O\\}$ , where $\\{H,B,O\\}$ are approximate head, body and outer bed regions as illustrated in Figure 5. This resulted in $(2 + 2 + 3 + 3)$ time-series features per region, i.e. 30 unique time-series features. Additionally, we provided the video-based classifier (see Section 4.3) with time-shifted features $f_{i}(t + T)$ for $T \\in \\{-90,0, + 90\\}$ seconds. This gave a total of 90 motion features for each sleep epoch.", + "bbox": [ + 496, + 194, + 890, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Transferring SleepVST to Video", + "text_level": 1, + "bbox": [ + 498, + 339, + 782, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To transfer the SleepVST model to video using the OSV dataset, we removed the linear classification layer learnt during pre-training and used the model as a frozen feature extractor, applying it to sequences of cardiac and respiratory waveforms measured from video to produce sequences of output features $\\mathbf{z}_o$ . We then trained a new classifier that used these features, plus the motion features derived in the previous section, to classify individual sleep epochs. This allowed us to leverage the representational capacity of the transformer model, by pre-training on the much larger contact-sensor datasets, whilst also enabling us to incorporate motion information from the video dataset. We followed the design choice of [6] and used a Random Forest [5] for our classifier due to its low computational cost and robustness to overfitting.", + "bbox": [ + 496, + 362, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 602, + 705, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sleep staging performance has commonly been reported in terms of classification accuracy $(\\mathrm{Acc}_T)$ and total Cohen's kappa $(\\kappa_T)$ over all sleep epochs in the dataset, and/or using per-night mean statistics, $\\mathrm{Acc}_{\\mu}$ and $\\kappa_{\\mu}$ , i.e.:", + "bbox": [ + 496, + 627, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa_ {T} = K \\left(\\sum_ {i = 1} ^ {N} c \\left(\\boldsymbol {y} _ {i}, \\hat {\\boldsymbol {y}} _ {i}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 698, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa_ {\\mu} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} K \\left(c \\left(\\boldsymbol {y} _ {i}, \\hat {\\boldsymbol {y}} _ {i}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 742, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{y}_i, \\hat{\\pmb{y}}_i \\in \\mathbb{R}^{C \\times L_i}$ are the one-hot encoded model and scorer labels for recording $i$ of length $L_i$ and $c(\\pmb{y}_i, \\hat{\\pmb{y}}_i) \\in \\mathbb{R}^{C \\times C}$ is the confusion matrix between them. $K(\\cdot)$ is the function which calculates the Cohen's kappa statistic [9] from a confusion matrix and is defined in §G. We report all of these statistics, to make fairer comparisons with prior work, and to aid future comparisons.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12483", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77d3913bfd21dd10d4b8964c9f7266bc0c1d88709d6c0c2dfa38dfc559816add.jpg", + "image_caption": [ + "Figure 6. Example four-class sleep hypnograms from the OSV dataset. (top) Scored by an expert using signals from the vPSG recording. (bottom) Automatically generated from near-infrared video using SleepVST. (Cohen's $\\kappa = 0.71$ between model and expert sleep stages.)" + ], + "image_footnote": [], + "bbox": [ + 96, + 98, + 877, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Pre-training on Contact Sensors", + "text_level": 1, + "bbox": [ + 76, + 402, + 361, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From the SHHS dataset, we randomly selected 500 participants who participated in both visits, to form a test set of 1000 nights of data. This ensured that no participant appeared in both the training/validation and test partitions. Our test set is listed in $\\S H$ . From the MESA dataset, we used the same test set of 204 nights as Kotzen et al. [22]. The remaining nights from both the SHHS and MESA datasets were pooled together and randomly split into training and validation sets using a 75:25 split.", + "bbox": [ + 75, + 426, + 468, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares the overall performance of SleepVST with prior methods for cardio-respiratory sleep staging from contact sensors. These have used input modalities such as the photoplethysmogram (PPG) waveform and the ECG-derived heart rate. Using the ECG and thoracic respiratory signals, SleepVST outperforms prior methods for sleep staging from cardio-respiratory signals on both datasets.", + "bbox": [ + 75, + 563, + 468, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Video-based Sleep Staging", + "text_level": 1, + "bbox": [ + 76, + 679, + 316, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we report the performance of the SleepVST model after transfer to the OSV video dataset. We follow the evaluation procedure of [6], training and evaluating on the 50 nights of data using 10-fold cross-validation with non-overlapping folds each containing five nights of data.", + "bbox": [ + 75, + 703, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we see that our method achieves significantly better performance in terms of both accuracy and Cohen's $\\kappa$ statistic when compared with prior video-based methods evaluated on similar study populations.", + "bbox": [ + 75, + 779, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 7 shows the total test confusion matrix between model and human expert (polysomnographer) classifications on the OSV dataset. We observe particularly high accuracy in distinguishing Wake and REM sleep stages.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9d84e28653074c427a8cd06f230fe54cc96db6dbead15932cabe78b06086e8f.jpg", + "image_caption": [ + "Figure 7. Confusion matrix between SleepVST and expert labels on four-class video based sleep staging." + ], + "image_footnote": [], + "bbox": [ + 511, + 407, + 880, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Population performance. Figure 8 shows the distribution of Cohen's $\\kappa$ values across participants from the OSV dataset plotted against age, sex and Fitzpatrick skin type[2] [14]. We observe a decrease in performance with age, as observed in prior sleep staging work [6, 53]. In older adults, measures of cardio-respiratory activity (e.g. heart rate variability) are known to decline with age [72]. This likely reduces the distance between sleep stages in cardiorespiratory input space and thus makes them harder to distinguish. Additional video data from older subjects would likely help to improve this, which is discussed further in §C.", + "bbox": [ + 496, + 693, + 890, + 859 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2 Plotted using representative skin tones from [67].", + "bbox": [ + 514, + 886, + 785, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12484", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa56b9fb2d63635d8828f99834ae8ca1250d4c34a24c702b992f8780027026ef.jpg", + "table_caption": [ + "Table 1. Comparison of four-class cardio-respiratory sleep staging methods using contact sensors." + ], + "table_footnote": [ + "*Thoracic respiratory effort. †Nasal airflow. ‡From reported confusion matrices." + ], + "table_body": "
DatasetNMethodModalitiesCohen's κAccuracy / %
κμκTAccμAccT
SHHS [43]296Bakker et al. [3]HR+Thor* +Air†-0.64-76.7
800Sridhar et al. [50]HR (ECG)0.650.6677.377.0
1000SleepVSTECG+Thor*0.730.7582.883.0
MESA [8]296Bakker et al. [3]HR+Thor* +Air†-0.68-79.8
194Sridhar et al. [50]HR (ECG)-0.69-80.0
204Kotzen et al. [22]PPG-0.73‡-82.6‡
204SleepVSTECG+Thor*0.760.7785.185.2
", + "bbox": [ + 200, + 114, + 769, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/481ef70e474959d9f37e10f314b3ac5466a0e8ffa72418b317c30d4ffabcac6d.jpg", + "table_caption": [ + "Table 2. Comparison of four-class sleep staging methods using video cameras." + ], + "table_footnote": [ + "**Camera-derived. *Cardiac pulse (heart) waveforms.† Respiration (breathing) waveforms.‡ From reported confusion matrices." + ], + "table_body": "
DatasetNMethodModalities**Cohen's κAccuracy / %
κμκTAccμAccT
Healthy infants [21]8Kamon et al. [21]Motion0.26-48.0-
Healthy adults [35]6Nochino et al. [35]Motion0.19-40.5-
HealthBed Study [60]46van Meulen et al. [60]HW*0.490.49‡67.967.9‡
Oxford Sleep Volunteers50Carter et al. [6]HR+BR+Motion0.610.64‡73.474.3‡
50SleepVSTHW* + BW† + Motion0.680.7177.778.8
", + "bbox": [ + 145, + 334, + 823, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f4a18a5fbd1f584f58d720232e940c2dc5bdf04f9e91452845d9df4589524aa.jpg", + "image_caption": [ + "Figure 8. Scatter and box plot of Cohen's $\\kappa$ against age, sex and Fitzpatrick skin type $^2$ for individuals from the OSV dataset." + ], + "image_footnote": [], + "bbox": [ + 89, + 512, + 457, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Example SleepVST output. Figure 6 shows an example hypnogram generated entirely from near-infrared video using our method. This example corresponds to the median Cohen's $\\kappa$ between model and expert from the OSV dataset. The model correctly identifies all three REM cycles, periods of deep sleep, and brief awakenings throughout the night. Additional hypnogram examples can be found in §C.", + "bbox": [ + 76, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 506, + 665, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video transfer strategy. In Table 3, we compare the performance of three methods for transferring the SleepVST model to video data:", + "bbox": [ + 498, + 530, + 890, + 573 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Directly applying the pre-trained SleepVST model to cardiac and respiratory waveforms from video data.", + "2. Using SleepVST as a feature extractor and:" + ], + "bbox": [ + 500, + 575, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) training a new classification layer without using motion features.", + "(b) training a new classification layer that also uses motion features i.e. our best-performing approach." + ], + "bbox": [ + 524, + 621, + 890, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we see that training a new classifier for video data is essential to the effectiveness of our approach. This step likely allows the classifier to ignore parts of the learnt feature space that do not transfer between contact-sensor and video-derived waveforms i.e. sensor-specific features.", + "bbox": [ + 498, + 681, + 890, + 756 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc3c9db71f33499f3c73106eca00049cb66b6a354a60964391dd5214309dc893.jpg", + "table_caption": [ + "Table 3. SleepVST video-based sleep staging performance for different model transfer strategies." + ], + "table_footnote": [], + "table_body": "
κμκTAccμ / %AccT / %
Direct apply0.5100.53666.767.3
Transfer w/o motion0.6600.68976.477.4
Transfer w/ motion0.6770.70877.778.8
", + "bbox": [ + 503, + 809, + 885, + 883 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12485", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c35a369e0fc4a80ccef27bd57c870129abc084cefc0256345880af757353eb8c.jpg", + "image_caption": [ + "Figure 9. Example four-class hypnograms during fragmented sleep from the OSV dataset. The use of motion features improves accuracy during short awakenings between 22.00 and 00.00." + ], + "image_footnote": [], + "bbox": [ + 86, + 94, + 467, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion features. Table 4 reports the effectiveness of the two main components of our motion feature set: the set of feature definitions $f_{i}$ and the set of regions used $R$ . Additional ablations can be found in §C. Here we define and report an additional metric $\\kappa_{F}$ which is the Cohen's $\\kappa$ statistic computed over epochs near transitions to/from Wake (W) i.e. where $\\hat{y}_i^t\\neq \\hat{y}_i^{t + \\Delta} = W$ or $\\hat{y}_i^t = W\\neq \\hat{y}_i^{t + \\Delta}$ for any $\\Delta \\in [-2,2]$ epochs. This metric better quantifies accuracy during fragmented sleep, such as in Figure 9.", + "bbox": [ + 75, + 472, + 468, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As measured by $\\kappa_{F}$ , we see that motion features aid classification during fragmented sleep. The ability to accurately measure sleep fragmentation is particularly important since it is linked with daytime sleepiness and decreased cognitive performance [29].", + "bbox": [ + 75, + 611, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e3cf9649ad6d12d46f9ca2d4c69b5999fe1075eaacf3595a011c6ec1a0f2e4f8.jpg", + "table_caption": [ + "Table 4. Ablation study of motion feature parameters." + ], + "table_footnote": [ + "\\*No. features." + ], + "table_body": "
AblationParameter Setnf*Cohen's κ
κTκF
Feature setfi ∈ {}00.6890.397
fi ∈ {f1, f2}360.6970.467
fi ∈ {f3, f4}540.7040.465
fi ∈ {f1, f2, f3, f4}900.7080.491
Region(s)R ∈ {H ∪ B}300.7020.461
R ∈ {H ∪ B, O}600.7040.468
R ∈ {H, B, O}900.7080.491
", + "bbox": [ + 93, + 729, + 450, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cardio-respiratory inputs. In Table 5, we compare the performance of the SleepVST model with variants which use heart rate $\\boldsymbol{x}_{HR}$ or breathing rate $\\boldsymbol{x}_{BR}$ as inputs instead of the underlying waveforms, as described in Section 4.1. We see that the use of waveforms, rather than derived heart and breathing rates, leads to improvements in sleep staging accuracy across all datasets.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cc4d5190bc6a3c064218d64c9ac02a83829fa96857a817a6a916fc3f62d75680.jpg", + "table_caption": [ + "Table 5. Comparison of four-class sleep staging performance for different cardio-respiratory input combinations to SleepVST." + ], + "table_footnote": [], + "table_body": "
Input signalsCohen's κT
SHHSMESAOSV
xHR, xBR0.6830.7220.681
xHW, xBR0.7180.7360.684
xHR, xBW0.7230.7580.702
xHW, xBW0.7490.7650.708
", + "bbox": [ + 566, + 256, + 823, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 409, + 627, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced SleepVST, a transformer model for sleep stage classification from cardio-respiratory waveforms. Using a transfer learning approach, we have shown how SleepVST can be used to perform sleep stage classification entirely from near-infrared video.", + "bbox": [ + 496, + 438, + 893, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our results advance the state-of-the-art in video-based sleep staging, narrowing the gap to expert-level performance and taking a significant step towards important clinical applications. More broadly, our work shows that a model pre-trained on a task using cardio-respiratory waveforms from large contact-sensor datasets can be effectively transferred to waveforms measured from video. This approach can benefit wider applications such as the detection of arrhythmias and sleep apnoes, from cardiac and respiratory waveforms respectively.", + "bbox": [ + 496, + 517, + 893, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we believe there is important future work in learning more expressive representations of motion during sleep. In addition to improving the performance of video-based sleep staging, this could also enable the detection of sleep movement disorders, such as REM behaviour disorder, removing the existing need for time-consuming, manual video review [51].", + "bbox": [ + 496, + 671, + 893, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 801, + 668, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the EPSRC Centre for Doctoral Training in Autonomous Intelligent Machines and Systems [EP/S024050/1] and funded by Oxehealth Ltd. Figure 1 was created with BioRender.com. We kindly thank the National Sleep Research Resource for granting access to SHHS and MESA.", + "bbox": [ + 496, + 830, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12486", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 4", + "[2] Ying-Wen Bai, Wen-Tai Li, and Cheng-Hsiang Yeh. Design and implementation of an embedded monitor system for body breath detection by using image processing methods. In 2010 Digest of Technical Papers International Conference on Consumer Electronics (ICCE), pages 193-194, 2010. 2", + "[3] Jessie P. Bakker, Marco Ross, Ray Vasko, Andreas Cerny, Pedro Fonseca, Jeff Jasko, Edmund Shaw, David P. White, and Peter Anderer. Estimating sleep stages using cardiorespiratory signals: validation of a novel algorithm across a wide range of sleep-disordered breathing severity. Journal of Clinical Sleep Medicine, 17(7):1343-1354, 2021. 3, 7", + "[4] Guha Balakrishnan, Fredo Durand, and John Guttag. Detecting Pulse from Head Motions in Video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, Portland, OR, USA, 2013. IEEE. 2", + "[5] Leo Breiman. Random Forests. Machine Learning, 45(1): 5-32, 2001. 5", + "[6] Jonathan Carter, João Jorge, Bindia Venugopal, Oliver Gibson, and Lionel Tarassenko. Deep Learning-Enabled Sleep Staging From Vital Signs and Activity Measured Using a Near-Infrared Video Camera. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 5940-5949, Vancouver, BC, Canada, 2023. IEEE. 1, 3, 5, 6, 7", + "[7] Weixuan Chen and Daniel McDuff. Deepphys: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 349-365, 2018. 2", + "[8] Xiaoli Chen, Rui Wang, Phyllis Zee, Pamela L. Lutsey, Sogol Javaheri, Carmela Alcantara, Chandra L. Jackson, Michelle A. Williams, and Susan Redline. Racial/Ethnic Differences in Sleep Disturbances: The Multi-Ethnic Study of Atherosclerosis (MESA). Sleep, 38(6):877-888, 2015. 3, 7", + "[9] Jacob Cohen. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1): 37-46, 1960. 2, 5", + "[10] Kris Cuppens, Lieven Lagae, Berten Ceulemans, Sabine Van Huffel, and Bart Vanrumste. Automatic video detection of body movement during sleep based on optical flow in pediatric patients with epilepsy. Medical & Biological Engineering & Computing, 48(9):923-931, 2010. 3", + "[11] Heidi Danker-Hopfe, Peter Anderer, Josef Zeitlhofer, Marion Boeck, Hans Dorn, Georg Gruber, Esther Heller, Erna Loretz, Doris Moser, Silvia Parapatics, Bernd Saletu, Andrea Schmidt, and Georg Dorffner. _Interrater reliability for sleep scoring according to the Rechtschaffen & Kales and the new AASM standard_. Journal of Sleep Research, 18(1):74-84, 2009. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1365-2869.2008.00700.x. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Koldo De Miguel, Alberto Brunete, Miguel Hernando, and Ernesto Gambao. Home Camera-Based Fall Detection System for the Elderly. Sensors, 17(12):2864, 2017. 1", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 4", + "[14] T. B. Fitzpatrick. The validity and practicality of sun-reactive skin types I through VI. Archives of Dermatology, 124(6): 869-871, 1988. 6", + "[15] Marco Hafner, Martin Stepanek, Jirka Taylor, Wendy M. Troxel, and Christian van Stolk. Why Sleep Matters—The Economic Costs of Insufficient Sleep. *Rand Health Quarterly*, 6(4):11, 2017. 1", + "[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 5", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4", + "[18] Adrienne Heinrich, Xavier Aubert, and Gerard de Haan. Body movement analysis during sleep based on video motion estimation. In 2013 IEEE 15th International Conference on e-Health Networking, Applications and Services (Healthcom 2013), pages 539–543, 2013. 1", + "[19] D. W. Hudgel, R. J. Martin, B. Johnson, and P. Hill. Mechanics of the respiratory system and breathing pattern during sleep in normal humans. Journal of Applied Physiology: Respiratory, Environmental and Exercise Physiology, 56(1): 133-137, 1984. 1, 3", + "[20] C. Iber. The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology, and Technical Specification. 2007. 1, 2", + "[21] Masamitsu Kamon, Shima Okada, Masafumi Furuta, and Koki Yoshida. Development of a non-contact sleep monitoring system for children. Frontiers in Digital Health, 4, 2022. 3, 7", + "[22] Kevin Kotzen, Peter H. Charlton, Sharon Salabi, Lea Amar, Amir Landesberg, and Joachim A. Behar. SleepPPG-Net: A Deep Learning Algorithm for Robust Sleep Staging From Continuous Photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 27(2):924–932, 2023. 1, 2, 6, 7", + "[23] Yee Fun Lee, Dmitry Gerashchenko, Igor Timofeev, Brian J. Bacskai, and Ksenia V. Kastanenka. Slow Wave Sleep Is a Promising Intervention Target for Alzheimer's Disease. Frontiers in Neuroscience, 14, 2020. 1", + "[24] Yun Ji Lee, Jae Yong Lee, Jae Hoon Cho, and Ji Ho Choi. Interrater reliability of sleep stage scoring: a meta-analysis. Journal of Clinical Sleep Medicine, 18(1):193-202, 2022. 2", + "[25] Michael H. Li, Azadeh Yadollahi, and Babak Taati. A noncontact vision-based system for respiratory rate estimation. In 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pages 2119-2122, 2014. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12487", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xin Liu, Brian Hill, Ziheng Jiang, Shwetak Patel, and Daniel McDuff. EfficientPhys: Enabling Simple, Fast and Accurate Camera-Based Cardiac Measurement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5008-5017, 2023. 2", + "[27] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, 2019. 4", + "[28] Yawei Luo, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Taking a Closer Look at Domain Shift: Category-Level Adversaries for Semantics Consistent Domain Adaptation. pages 2507-2516, 2019. 3", + "[29] S E Martin, H M Engleman, I J Deary, and N J Douglas. The effect of sleep fragmentation on daytime function. American Journal of Respiratory and Critical Care Medicine, 153(4): 1328-1332, 1996. Publisher: American Thoracic Society - AJRCCM. 8", + "[30] Daniel McDuff. Camera Measurement of Physiological Vital Signs. ACM Computing Surveys, 55(9):1-40, 2023. 2", + "[31] Sara Mahvash Mohammadi, Shirin Enshaeifar, Adrian Hilton, Derk-Jan Dijk, and Kevin Wells. Transfer Learning for Clinical Sleep Pose Detection Using a Single 2D IR Camera. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 29:290-299, 2021. 5", + "[32] K. Nakajima, A. Osa, and H. Miike. A method for measuring respiration and physical activity in bed by optical flow analysis. In Proceedings of the 19th Annual International Conference of the IEEE Engineering in Medicine and Biology Society., pages 2054-2057 vol.5, 1997. 2", + "[33] Kazuki Nakajima, Yoshiaki Matsumoto, and Toshiyo Tamura. Development of real-time image sequence analysis for evaluating posture change and respiratory rate of a subject in bed. Physiological Measurement, 22(3):N21-N28, 2001. 2, 3", + "[34] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations, 2022. 3, 4", + "[35] Teruaki Nochino, Yuko Ohno, Takafumi Kato, Masako Tani-ike, and Shima Okada. Sleep stage estimation method using a camera for home use. Biomedical Engineering Letters, 9 (2):257-265, 2019. 1, 3, 4, 7", + "[36] Jiapu Pan and Willis J. Tompkins. A Real-Time QRS Detection Algorithm. IEEE Transactions on Biomedical Engineering, BME-32(3):230-236, 1985. 3", + "[37] James Pardey, Stephen Roberts, Lionel Tarassenko, and John Stradling. A new approach to the analysis of the human sleep/wakefulness continuum. Journal of sleep research, 5 (4):201-210, 1996. 1", + "[38] Aakash K. Patel, Vamsi Reddy, and John F. Araujo. Physiology, Sleep Stages. StatPearls Publishing, 2022. 4", + "[39] Huy Phan and Kaare Mikkelsen. Automatic sleep staging of EEG signals: recent development, challenges, and future directions. Physiological Measurement, 43(4):04TR01, 2022. 2", + "[40] Huy Phan, Oliver Y. Chén, Minh C. Tran, Philipp Koch, Alfred Mertins, and Maarten De Vos. XSleepNet: Multi-View" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sequential Model for Automatic Sleep Staging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 5903-5915, 2022. 1", + "[41] Huy Phan, Kaare Mikkelsen, Oliver Y. Chen, Philipp Koch, Alfred Mertins, and Maarten De Vos. SleepTransformer: Automatic Sleep Staging With Interpretability and Uncertainty Quantification. IEEE Transactions on Biomedical Engineering, 69(8):2456-2467, 2022. 4", + "[42] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Advancements in Noncontact, Multiparameter Physiological Measurements Using a Webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. Conference Name: IEEE Transactions on Biomedical Engineering. 2", + "[43] S. F. Quan, B. V. Howard, C. Iber, J. P. Kiley, F. J. Nieto, G. T. O'Connor, D. M. Rapoport, S. Redline, J. Robbins, J. M. Samet, and P. W. Wahl. The Sleep Heart Health Study: design, rationale, and methods. Sleep, 20(12):1077-1085, 1997. 3, 7", + "[44] Mustafa Radha, Pedro Fonseca, Arnaud Moreau, Marco Ross, Andreas Cerny, Peter Anderer, Xi Long, and Ronald M. Aarts. A deep transfer learning approach for wearable sleep stage classification with photoplethysmography. npj Digital Medicine, 4(1):1-11, 2021. 2", + "[45] S.J. Redmond and C. Heneghan. Cardiorespiratory-based sleep staging in subjects with obstructive sleep apnea. IEEE Transactions on Biomedical Engineering, 53(3):485-496, 2006. 1", + "[46] A. J. Schwichtenberg, Jeehyun Choe, Ashleigh Kellerman, Emily A. Abel, and Edward J. Delp. Pediatric Videosomnography: Can Signal/Video Processing Distinguish Sleep and Wake States? Frontiers in Pediatrics, 6:158, 2018. 3", + "[47] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90(2):227-244, 2000. 3", + "[48] Z. Shinar, A. Baharav, Y. Dagan, and S. Akselrod. Automatic detection of slow-wave-sleep using heart rate variability. In Computers in Cardiology 2001. Vol.28, pages 593-596, 2001. 1, 3", + "[49] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 2", + "[50] Niranjan Sridhar, Ali Shoeb, Philip Stephens, Alaa Kharbouch, David Ben Shimol, Joshua Burkart, Atiyeh Ghoreyshi, and Lance Myers. Deep learning for automated sleep staging using instantaneous heart rate. npj Digital Medicine, 3(1):1-10, 2020. 2, 7", + "[51] Ambra Stefani and Birgit Hogl. Sleep in Parkinson's disease. Neuropsychopharmacology, 45(1):121-128, 2020. 1, 8", + "[52] Ambra Stefani, David Gabelaia, Thomas Mitterling, Werner Poewe, Birgit Hogl, and Birgit Frauscher. A Prospective Video-Polysomnographic Analysis of Movements during Physiological Sleep in 100 Healthy Sleepers. Sleep, 38 (9):1479-1487, 2015. 1", + "[53] Haoqi Sun, Wolfgang Ganglberger, Ezhil Panneerselvam, Michael J Leone, Syed A Quadri, Balaji Goparaju, Ryan A" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12488", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tesh, Oluwaseun Akeju, Robert J Thomas, and M Brandon Westover. Sleep staging from electrocardiography and respiration with deep learning. *Sleep*, 43(7):zsz306, 2020. 6", + "[54] K. S. Tan, R. Saatchi, H. Elphick, and D. Burke. Real-time vision based respiration monitoring system. In 2010 7th International Symposium on Communication Systems, Networks & Digital Signal Processing (CSNDSP 2010), pages 770-774, 2010. 2", + "[55] Alexander Trumpp, Johannes Lohr, Daniel Wedekind, Martin Schmidt, Matthias Burghardt, Axel R. Heller, Hagen Malberg, and Sebastian Zaunseder. Camera-based photoplethysmography in an intraoperative setting. *BioMedical Engineering OnLine*, 17(1):33, 2018. 2", + "[56] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Oxhealth Vital Signs, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200019.pdf.3", + "[57] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Gili Pro Biosensor, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200038.pdf.3", + "[58] Mark van Gastel, Sander Stuijk, Sebastiaan Overeem, Johannes P. van Dijk, Merel M. van Gilst, and Gerard de Haan. Camera-Based Vital Signs Monitoring During Sleep - A Proof of Concept Study. IEEE Journal of Biomedical and Health Informatics, 25(5):1409-1418, 2021. 3", + "[59] Hans van Gorp, Iris A M Huijben, Pedro Fonseca, Ruud J G van Sloun, Sebastiaan Overeem, and Merel M van Gilst. Certainty about uncertainty in sleep staging: a theoretical framework. Sleep, 45(8), 2022. 2", + "[60] Fokke B. van Meulen, Angela Grassi, Leonie van den Heuvel, Sebastiaan Overeem, Merel M. van Gilst, Johannes P. van Dijk, Henning Maass, Mark J. H. van Gastel, and Pedro Fonseca. Contactless Camera-Based Sleep Staging: The HealthBed Study. Bioengineering, 10(1):109, 2023. 1, 3, 7", + "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. arXiv:1706.03762 [cs], 2017. arXiv: 1706.03762. 4", + "[62] Mauricio Villarroel, Sitthichok Chaichulee, João Jorge, Sara Davis, Gabrielle Green, Carlos Arteta, Andrew Zisserman, Kenny McCormick, Peter Watkinson, and Lionel Tarassenko. Non-contact physiological monitoring of preterm infants in the Neonatal Intensive Care Unit. npj Digital Medicine, 2(1):1-18, 2019. 2", + "[63] Olivia Walch, Yitong Huang, Daniel Forger, and Cathy Goldstein. Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device. *Sleep*, 42(12), 2019. 1", + "[64] Qiongyan Wang, Hanrong Cheng, and Wenjin Wang. Feasibility of Exploiting Physiological and Motion Features for Camera-based Sleep Staging: A Clinical Study. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5, 2023. ISSN: 2694-0604. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Wenjin Wang and Albertus C. den Brinker. Algorithmic insights of camera-based respiratory motion extraction. Physiological measurement, 43(7):075004, 2022. 1, 2", + "[66] Wenjin Wang, Albertus C. den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic Principles of Remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 1, 2", + "[67] William H. Ward, Fernando Lambreton, Neha Goel, Jian Q. Yu, and Jeffrey M. Farma. Clinical Presentation and Staging of Melanoma. In Cutaneous Melanoma: Etiology and Therapy. Codon Publications, Brisbane (AU), 2017. 6", + "[68] Johanna Wilde-Frenz and Hartmut Schulz. Rate and Distribution of Body Movements during Sleep in Humans. Perceptual and Motor Skills, 56(1):275-283, 1983. 1, 5", + "[69] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. SimPer: Simple Self-Supervised Learning of Periodic Targets, 2023. 2", + "[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. PhysFormer: Facial Video-based Physiological Measurement with Temporal Difference Transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4186, New Orleans, LA, USA, 2022. IEEE. 2", + "[71] Guo-Qiang Zhang, Licong Cui, Remo Mueller, Shiqiang Tao, Matthew Kim, Michael Rueschman, Sara Mariani, Daniel Mobley, and Susan Redline. The National Sleep Research Resource: towards a sleep data commons. Journal of the American Medical Informatics Association: JAMIA, 25 (10):1351-1358, 2018. 3", + "[72] Dan Ziegler, G. Laux, K. Dannehl, M. Spüler, H. Mühlen, P. Mayer, and F.a. Gries. Assessment of Cardiovascular Autonomic Function: Age-related Normal Ranges and Reproducibility of Spectral Analysis, Vector Analysis, and Standard Tests of Heart Rate Variation and Blood Pressure Responses. Diabetic Medicine, 9(2):166-175, 1992. 6" + ], + "bbox": [ + 501, + 92, + 890, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12489", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_model.json b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1ae86b341ea57a1f384db56b0627ad164729bc87 --- /dev/null +++ b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.13, + 0.783, + 0.175 + ], + "angle": 0, + "content": "SleepVST: Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.215, + 0.761, + 0.235 + ], + "angle": 0, + "content": "Jonathan F. Carter\\(^{1,2}\\) João Jorge\\(^{2}\\) Oliver Gibson\\(^{2}\\) Lionel Tarassenko\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.246, + 0.721, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Biomedical Engineering, University of Oxford \\(^{2}\\)Oxehealth Ltd., Oxford" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.326, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.662 + ], + "angle": 0, + "content": "Advances in camera-based physiological monitoring have enabled the robust, non-contact measurement of respiration and the cardiac pulse, which are known to be indicative of the sleep stage. This has led to research into camera-based sleep monitoring as a promising alternative to \"gold-standard\" polysomnography, which is cumbersome, expensive to administer, and hence unsuitable for longer-term clinical studies. In this paper, we introduce SleepVST, a transformer model which enables state-of-the-art performance in camera-based sleep stage classification (sleep staging). After pre-training on contact sensor data, SleepVST outperforms existing methods for cardiorespiratory sleep staging on the SHHS and MESA datasets, achieving total Cohen's kappa scores of 0.75 and 0.77 respectively. We then show that SleepVST can be successfully transferred to cardio-respiratory waveforms extracted from video, enabling fully contact-free sleep staging. Using a video dataset of 50 nights, we achieve a total accuracy of \\(78.8\\%\\) and a Cohen's \\(\\kappa\\) of 0.71 in four-class video-based sleep staging, setting a new state-of-the-art in the domain." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.842 + ], + "angle": 0, + "content": "Accurate sleep monitoring is critical for the diagnosis of sleep disorders and the discovery of novel treatments and biomarkers. Poor sleep is broadly linked with a number of health conditions, including cardiovascular diseases such as diabetes and hypertension [15]. Additionally, there are links between specific sleep stages and neurodegenerative conditions, such as a decrease in non-rapid-eye-movement stage 3 (N3) sleep with Alzheimer's disease [23]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.843, + 0.47, + 0.874 + ], + "angle": 0, + "content": "Overnight video polysomnography (vPSG), the \"gold-standard\" [52] technique for sleep monitoring, requires the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.895, + 0.45 + ], + "angle": 0, + "content": "use of a large number of contact sensors including electrodes placed on the scalp (EEG), near the eyes (EOG), and under the chin (EMG) of the patient. Human experts (polysomnographers) must then review the recorded data, classifying the patient's sleep into five stages at 30-second intervals (epochs) and annotating other important events such as leg movements and apnoeas. This manual process often takes multiple hours to complete." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.895, + 0.575 + ], + "angle": 0, + "content": "Experts primarily rely on characteristic patterns in brain activity measured from the EEG to classify sleep stages [20]. However, sleep stage information is also encoded in measures of autonomic activity, including cardiac [48] and respiratory signals [19], and body movements [68]. This has led to the development of methods for automatically classifying sleep stages from sensors that measure these parameters, such as smartwatches [63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.777 + ], + "angle": 0, + "content": "Prior work has shown that these physiological parameters can also be measured using video cameras [18, 65, 66], leading to the development of methods for classifying sleep stages entirely from video input [6, 35, 60]. Camera-based methods have particular promise as part of a multipurpose sleep monitoring system. For example, within elderly care settings, in addition to classifying sleep stages, they could also be used to detect specific sleep movement disorders [18], which are linked with conditions such as Parkinson's, and which are typically distinguished via manual video review [51]. They could also be used to detect falls [12], one of the leading causes of injury and mortality amongst elderly individuals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "State-of-the-art video-based sleep staging [6] has used the heart rate (HR) and breathing rate (BR) derived from video to perform sleep staging. However, these rates are typically measured as averages over time, losing higher frequency inter-pulse and inter-breath interval information. EEG- [40] and wearable-based [22] sleep staging methods have both been improved by using raw waveforms as inputs, rather than derived quantities e.g. signal features [37, 45]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.286, + 0.9 + ], + "angle": 0, + "content": "*jcarter@robots.ox.ac.uk" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12479" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.874, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 1. Sleep staging from near-infrared video signals using SleepVST. (a) We first pre-train the model end-to-end on cardiac (heart) and respiratory (breathing) waveforms \\( x_{HW}(t) \\) and \\( x_{BW}(t) \\) derived from the electrocardiogram (ECG) and a thoracic respiratory belt (THX) respectively. (b) After pre-training, we use the model as a frozen feature extractor, applying it to cardio-respiratory waveforms derived from near-infrared (NIR) video to generate sequences of features. When transferring to video data, we additionally use a set of motion features \\( f_{i}(t) \\) derived from an optical flow field \\( \\pmb{u}(x,y,t) \\) as inputs to the classifier C. This approach allows us to utilise much larger contact-sensor datasets to train SleepVST, whilst also enabling the incorporation of motion information when transferring to video data. (c) We evaluate SleepVST using overnight video polysomnography (vPSG) studies, comparing expert-labelled sleep stage sequences \\( \\hat{y}_{1:T} \\) against those generated entirely from near-infrared video \\( y_{1:T} \\) using our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.468, + 0.502 + ], + "angle": 0, + "content": "In this paper, we introduce SleepVST1, a transformer model for sleep stage classification from cardio-respiratory waveforms. First, we pre-train the model on waveforms derived from contact sensors (Fig. 1a), after which it outperforms existing methods for cardio-respiratory sleep staging on two widely used public datasets, highlighting the effectiveness of our transformer architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.468, + 0.67 + ], + "angle": 0, + "content": "After pre-training, we then show that SleepVST can be transferred to cardiac and respiratory waveforms measured entirely from near-infrared video, where our method additionally enables us to incorporate motion features derived from the camera data (Fig. 1b). Using this novel combination of inputs available from a near-infrared camera, we obtain state-of-the-art results for video-based sleep staging. Through a series of ablation studies we demonstrate the effectiveness of our approach. Our work further highlights the potential of camera-based sleep monitoring for important clinical applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.362, + 0.701 + ], + "angle": 0, + "content": "2. Background and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.468, + 0.876 + ], + "angle": 0, + "content": "Automated sleep staging. To reduce the manual effort of sleep assessment, deep learning methods have been developed for the automatic classification of sleep stages from polysomnography (PSG) recordings, achieving near-human accuracy [39]. Accuracy is commonly measured using Cohen's kappa statistic \\((\\kappa, [9])\\), which measures inter-rater agreement between two sources. Expert annotation of sleep stages is a subjective process, making \\(100\\%\\) agreement unachievable [59]. Prior work has shown that sleep scorers agree around \\(80\\%\\) of the time (\\(\\kappa = 0.71 - 0.81\\), [11, 24]) when classifying sleep stages according to American Academy of" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.398, + 0.89, + 0.503 + ], + "angle": 0, + "content": "Sleep Medicine (AASM, [20]) scoring rules. These rules divide sleep into five classes: Wake, N1 (light), N2 (intermediate), N3 (deep) and rapid-eye-movement (REM) sleep. However, due to low inter-scorer agreement in identifying N1 [11], prior work has often treated sleep classification as a four-class problem, by merging N1 and N2 into a single class [22, 50]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.89, + 0.58 + ], + "angle": 0, + "content": "Given the drawbacks of conventional polysomnography, prior work has investigated the potential of automatic sleep staging from a reduced number of contact sensors, including the electrocardiogram ([50], \\(\\kappa = 0.66\\)), and the photoplethysmogram ([44], \\(\\kappa = 0.65\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.89, + 0.687 + ], + "angle": 0, + "content": "Camera-based physiological measurement. Respiratory measurement is possible from video by measuring the induced movements of the chest and abdomen [65]. This has been achieved through methods such as frame-differencing [2, 54], using dense optical flow magnitudes [32, 33] and by tracking the movement of sparse optical flow vectors [25] across video frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.89, + 0.869 + ], + "angle": 0, + "content": "Camera-based measurement of the pulse is possible using methods which measure specific cardiac phenomena [30]. These include remote photoplethysmography (rPPG, [42, 66]), which measures skin colour changes caused by variations in blood volume over the cardiac cycle, and ballistocardiography [4], which measures head movements induced as blood is pumped through the carotid arteries. More recently, deep learning methods have been developed which learn to extract a pulse signal directly from input video frames [7, 26, 70] rather than targeting specific underlying cardiac phenomena, including unsupervised approaches [49, 69]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.89, + 0.901 + ], + "angle": 0, + "content": "Camera-based vital-sign measurement has been validated within clinical settings [55, 62], during overnight" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.276, + 0.9 + ], + "angle": 0, + "content": "Sleep Video-Signal Transformer." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "12480" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "polysomnography recordings [58, 64], and has led to products with regulatory clearance for camera-based measurement of the heart and breathing rate, including the Gili Pro Biosensor [57] and Oxehealth Vital Signs software [56]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.473, + 0.365 + ], + "angle": 0, + "content": "Camera-based sleep monitoring. Prior work has investigated the binary problem of distinguishing sleep from wake using video cameras, using measures of activity derived from the input frames [10, 33, 46]. Whilst activity information alone can be effective in distinguishing sleep from wake, cardio-respiratory signals carry important additional information for classifying specific sleep stages [19, 48]. The current state-of-the-art approach for video-based sleep staging ([6], \\(\\kappa = 0.61\\)) combines multiple modalities derived from video: the heart rate, breathing rate and activity information, to perform sleep stage classification. This approach outperforms prior methods which had used only a single video-derived modality: motion [21, 35] or the cardiac pulse [60]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.379, + 0.332, + 0.397 + ], + "angle": 0, + "content": "3. Datasets and Preprocessing" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.273, + 0.419 + ], + "angle": 0, + "content": "3.1. Contact Sensor Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.469, + 0.548 + ], + "angle": 0, + "content": "To train the SleepVST model, we used the SHHS [43, 71] and MESA [8] contact sensor datasets. These datasets contain PSG recordings and sleep stages annotated according to AASM scoring rules. The cardio-respiratory signals available in each recording include the electrocardiogram (ECG), which measures electrical activity of the heart, and respiratory signals, measured using respiratory bands placed around the abdomen and thorax." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.469, + 0.669 + ], + "angle": 0, + "content": "From the available signals, we seek to produce generic cardiac and respiratory waveforms to train the SleepVST model, preventing the model from learning to use information which may be present in signals derived from contact sensors but not present in those derived from video. This is to mitigate the problem of domain shift [28, 47], where the performance of a model is significantly reduced when applied to a different data distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.837 + ], + "angle": 0, + "content": "To produce a cardiac (heart) waveform, we first applied the same filtering steps used by the Pan-Tompkins algorithm [36] to the ECG signal. We resampled the output from the moving-window integration step of the algorithm to \\(10\\mathrm{Hz}\\), then applied a Butterworth bandpass filter with a passband of \\(0.66 - 2.8\\mathrm{Hz}\\) (40-168 beats per minute or BPM) and Gaussian smoothing. To produce a respiratory (breathing) waveform, we used the thoracic respiratory signal as the input source, as previously used in [3]. This was downsampled to \\(5\\mathrm{Hz}\\) and smoothed using a median filter with a kernel of length 5, to remove artefacts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.2, + 0.862 + ], + "angle": 0, + "content": "3.2. Video Data" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We used the dataset previously introduced in [6] for our video-based sleep staging experiments. This dataset con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "sists of overnight vPSG recordings from 50 volunteers, with sleep stages annotated according to the AASM guidelines at 30-second intervals (epochs) by a sleep physiologist. Throughout this work, we refer to this as the Oxford Sleep Volunteers (OSV) dataset. Each recording includes \\(850\\mathrm{nm}\\) near-infrared (NIR) video data captured at 20 FPS. An example camera view can be seen in Figure 5. More information on the dataset is provided in \\(\\S B\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.893, + 0.348 + ], + "angle": 0, + "content": "To derive cardiac and respiratory waveforms from the video data, we used components of FDA-cleared Oxhealth Vital Signs software [56]. For each overnight video recording, we used the software to obtain cardiac and respiratory waveforms sampled at \\(10\\mathrm{Hz}\\) and \\(5\\mathrm{Hz}\\) respectively. We then applied the same filtering steps applied to the contact sensor waveforms: Butterworth filtering and Gaussian smoothing to the heart waveforms, and median filtering to the breathing waveforms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.894, + 0.422 + ], + "angle": 0, + "content": "Figure 2 shows examples of processed cardio-respiratory waveforms measured from contact sensors and video for a short section of a recording from the OSV dataset. These have additionally been normalised to zero mean and unit variance." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.439, + 0.885, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Figure 2. Example (normalized) cardiac and respiratory waveforms, \\( x_{HW}(t) \\) and \\( x_{BW}(t) \\), derived from contact sensors (blue) and video (orange) from the OSV dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.599, + 0.687 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.715, + 0.712 + ], + "angle": 0, + "content": "4.1. SleepVST Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Figure 3 illustrates the SleepVST model architecture, which turns cardiac and respiratory waveforms into sequences of feature vectors using a transformer encoder. These feature vectors are then used to classify each 30-second sleep epoch. During pre-training, we used a linear layer for classification. When transferring to video data, we trained a new classifier that additionally used motion features derived from video as inputs, discussed further in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Patchification. The continuous heart and breathing waveforms (HW and BW) are divided into N windows i.e. patches, drawing on ideas from PatchTST [34]. We used non-overlapping 30-second windows since this is the inter" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12481" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.445, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.471, + 0.399 + ], + "angle": 0, + "content": "Figure 3. SleepVST architecture. Each 30-second window of heart \\((x_{HW})\\) and breathing waveforms \\((x_{BW})\\) is passed to a patch encoder, which turns them into patch-level features. These features are concatenated and passed to a transformer encoder. During pre-training, a linear layer turns the output feature sequence \\(z_{o}\\) of length N from SleepVST into sleep stage classifications." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.469, + 0.489 + ], + "angle": 0, + "content": "val at which sleep stages are classified according to AASM guidelines. We denote the patchified inputs as \\( \\boldsymbol{x}_{HW} \\in \\mathbb{R}^{300 \\times N} \\) and \\( \\boldsymbol{x}_{BW} \\in \\mathbb{R}^{150 \\times N} \\) respectively, where \\( N \\) is the sequence length." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.49, + 0.469, + 0.595 + ], + "angle": 0, + "content": "Encodings. Each waveform patch is normalised to zero mean and unit variance before being passed to encoding layers \\( E_{HW} \\) and \\( E_{BW} \\), which transform them into feature vectors \\( \\pmb{z}_{HW} \\in \\mathbb{R}^{D_{HW} \\times N} \\) and \\( \\pmb{z}_{BW} \\in \\mathbb{R}^{D_{BW} \\times N} \\). The feature vectors from each patch are then concatenated, producing a sequence of feature vectors \\( \\pmb{z}_i \\in \\mathbb{R}^{(D_{HW} + D_{BW}) \\times N} \\) which are passed to a transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.717 + ], + "angle": 0, + "content": "We used identical, shallow 1D ResNet-style [17] models for \\( E_{HW} \\) and \\( E_{BW} \\), illustrated in Figure 4. This convolutional encoder design [1] compresses the waveform patches into lower-dimensional feature vectors. This allowed us to keep \\( D_{HW} \\) and \\( D_{BW} \\) small, reducing the computational and memory complexity of the downstream transformer and the likelihood of over-fitting, particularly when transferring the model to the smaller video dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This encoder design additionally provides a high degree of flexibility at the input level, such as the ability to use input signals of differing frequencies. In later ablation studies, this enables us to compare the performance of the SleepVST model with simpler variants which use 30-second patches of \\(1\\mathrm{Hz}\\) heart rate \\(\\pmb{x}_{HR} \\in \\mathbb{R}^{30 \\times N}\\) and/or breathing rate \\(\\pmb{x}_{BR} \\in \\mathbb{R}^{30 \\times N}\\) time-series as inputs, rather than underlying waveforms. When using either of these as inputs, we used linear projections [13, 34] for the input encoders, because of the lower dimensionality of \\(\\pmb{x}_{HR}\\) and \\(\\pmb{x}_{BR}\\). Throughout our experiments we use the same feature dimension i.e. \\(D_{HW} = D_{BW} = D_{HR} = D_{BR} = 64\\)." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.098, + 0.875, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Waveform encoder design. Using a series of convolutional layers, the encoder turns sequences of signal patches into sequences of lower-dimensional feature vectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Transformer layer. We used a Transformer encoder [61] with sinusoidal position encodings to turn the sequence of input features \\( \\mathbf{z}_i \\) into an output sequence of features \\( \\mathbf{z}_o \\) with an identical feature dimension. Normal sleep exhibits long-range structures, known as sleep cycles [38], which typically last around 90-120 minutes and typically occur four to five times per night. From its superior performance in many sequence modelling tasks, including EEG-based sleep staging [41], we expect that the inductive biases of the transformer model (i.e. the attention mechanism) can enable it to learn to use relative position within a sleep cycle to classify sleep stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.671 + ], + "angle": 0, + "content": "Classification. When training and testing on contact sensor data, we used a linear classification layer \\( \\mathbf{W}_C \\in \\mathbb{R}^{C \\times (D_{HW} + D_{BW})} \\) to transform the sequence of output feature vectors \\( \\mathbf{z}_o \\) into a sequence of sleep stage probabilities \\( \\mathbf{y} \\in \\mathbb{R}^{C \\times N} \\), where the number of classes \\( C \\) depends on the sleep classification strategy used. In line with prior work, we classified sleep stages into four classes during pretraining: Wake-N1/N2-N3-REM, i.e. \\( C = 4 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Model training and inference. In all experiments, we used a sequence length of \\( N = 240 \\) windows i.e. two hours. During training, we created batches by sampling sequences of length \\( N \\) with a step of 5 minutes from within each recording. The model was trained to minimise the multiclass cross-entropy loss using the AdamW [27] optimiser with a learning rate of 0.0005 and weight decay of 0.001. Other hyper-parameters, such as the no. encoder layers, are detailed in §D. To apply the model to arbitrary length input sequences during evaluation, we re-applied it at regular intervals. More details on this process are provided in §E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.749, + 0.862 + ], + "angle": 0, + "content": "4.2. Motion features from Video" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Following prior work [35], we seek to produce motion features from video which encode for information which is pre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12482" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "dictive of the sleep stage, such as the rate and distribution of body movements [68]. To do so, we define a set of parameterised time-series features that collectively quantify motion over a range of spatial and temporal scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.468, + 0.243 + ], + "angle": 0, + "content": "Optical flow estimation. We first estimated a 2D optical flow field \\(\\pmb{u}(t,x,y) \\in \\mathbb{R}^2\\) over the bed region. Inspired by sleep pose detection work [31], we applied a homography transformation [16] to the individual video frames before estimating the flow field, to achieve camera viewpoint invariance (shown in Figure 5)." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.256, + 0.238, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.369, + 0.174, + 0.381 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.256, + 0.349, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.369, + 0.308, + 0.381 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.256, + 0.46, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.369, + 0.419, + 0.381 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.453 + ], + "angle": 0, + "content": "Figure 5. Example processing of an NIR video frame from the OSV dataset. (a) Real viewpoint. (b) Virtual viewpoint. (c) Head (H), body (B), and outer (O) bed regions. The distance from the camera to the head region is around \\(1.5\\mathrm{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.469, + 0.514 + ], + "angle": 0, + "content": "Feature definitions. Given an arbitrary region \\( \\mathbf{R} \\) of the optical flow field, we define the following parametrised signals:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.527, + 0.469, + 0.551 + ], + "angle": 0, + "content": "\\[\nv (t; R) = \\max _ {(x, y) \\in R} | \\boldsymbol {u} (t, x, y) | \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.554, + 0.469, + 0.592 + ], + "angle": 0, + "content": "\\[\ns (t; R) = \\frac {1}{| R |} \\sum_ {(x, y) \\in R} | \\boldsymbol {u} (t, x, y) | \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.469, + 0.664 + ], + "angle": 0, + "content": "i.e. the maximum (1) and the average (2) of the optical flow magnitude in the region R at time \\( t \\). Using these signals, we define two time-series features that quantify motion within a time window \\( \\Delta \\) up to time \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.677, + 0.469, + 0.717 + ], + "angle": 0, + "content": "\\[\nf _ {1} (t; R, \\Delta) = \\sum_ {t - \\Delta} ^ {t} v (t; R) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.722, + 0.469, + 0.762 + ], + "angle": 0, + "content": "\\[\nf _ {2} (t; R, \\Delta) = \\sum_ {t - \\Delta} ^ {t} s (t; R) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.469, + 0.804 + ], + "angle": 0, + "content": "Plus two additional features which measure the time elapsed since either signal went above a threshold \\(\\delta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.817, + 0.326, + 0.856 + ], + "angle": 0, + "content": "\\[\nf _ {3} (t; R, \\delta) = \\sum_ {t - \\tau} ^ {t} 1\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.854, + 0.469, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad \\tau = \\max _ {t ^ {\\prime} \\leq t} t ^ {\\prime} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.887, + 0.352, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c c} \\text {s . t .} & v (t ^ {\\prime}; R) > \\delta \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.104, + 0.749, + 0.143 + ], + "angle": 0, + "content": "\\[\nf _ {4} (t; R, \\delta) = \\sum_ {t - \\tau} ^ {t} 1\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.141, + 0.891, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad \\tau = \\max _ {t ^ {\\prime} \\leq t} t ^ {\\prime} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.174, + 0.773, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l l} \\text {s . t .} & s (t ^ {\\prime}; R) > \\delta \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.195, + 0.892, + 0.332 + ], + "angle": 0, + "content": "We calculated each feature for all possible combinations of parameters \\(\\Delta \\in \\{30,300\\}\\) seconds, \\(\\delta \\in \\{0.01,0.1,1\\}\\) and \\(R \\in \\{H,B,O\\}\\), where \\(\\{H,B,O\\}\\) are approximate head, body and outer bed regions as illustrated in Figure 5. This resulted in \\((2 + 2 + 3 + 3)\\) time-series features per region, i.e. 30 unique time-series features. Additionally, we provided the video-based classifier (see Section 4.3) with time-shifted features \\(f_{i}(t + T)\\) for \\(T \\in \\{-90,0, + 90\\}\\) seconds. This gave a total of 90 motion features for each sleep epoch." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.34, + 0.783, + 0.356 + ], + "angle": 0, + "content": "4.3. Transferring SleepVST to Video" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.59 + ], + "angle": 0, + "content": "To transfer the SleepVST model to video using the OSV dataset, we removed the linear classification layer learnt during pre-training and used the model as a frozen feature extractor, applying it to sequences of cardiac and respiratory waveforms measured from video to produce sequences of output features \\( \\mathbf{z}_o \\). We then trained a new classifier that used these features, plus the motion features derived in the previous section, to classify individual sleep epochs. This allowed us to leverage the representational capacity of the transformer model, by pre-training on the much larger contact-sensor datasets, whilst also enabling us to incorporate motion information from the video dataset. We followed the design choice of [6] and used a Random Forest [5] for our classifier due to its low computational cost and robustness to overfitting." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.603, + 0.706, + 0.62 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Sleep staging performance has commonly been reported in terms of classification accuracy \\((\\mathrm{Acc}_T)\\) and total Cohen's kappa \\((\\kappa_T)\\) over all sleep epochs in the dataset, and/or using per-night mean statistics, \\(\\mathrm{Acc}_{\\mu}\\) and \\(\\kappa_{\\mu}\\), i.e.:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.699, + 0.891, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\kappa_ {T} = K \\left(\\sum_ {i = 1} ^ {N} c \\left(\\boldsymbol {y} _ {i}, \\hat {\\boldsymbol {y}} _ {i}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.743, + 0.891, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\kappa_ {\\mu} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} K \\left(c \\left(\\boldsymbol {y} _ {i}, \\hat {\\boldsymbol {y}} _ {i}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\pmb{y}_i, \\hat{\\pmb{y}}_i \\in \\mathbb{R}^{C \\times L_i}\\) are the one-hot encoded model and scorer labels for recording \\(i\\) of length \\(L_i\\) and \\(c(\\pmb{y}_i, \\hat{\\pmb{y}}_i) \\in \\mathbb{R}^{C \\times C}\\) is the confusion matrix between them. \\(K(\\cdot)\\) is the function which calculates the Cohen's kappa statistic [9] from a confusion matrix and is defined in §G. We report all of these statistics, to make fairer comparisons with prior work, and to aid future comparisons." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.099, + 0.878, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Figure 6. Example four-class sleep hypnograms from the OSV dataset. (top) Scored by an expert using signals from the vPSG recording. (bottom) Automatically generated from near-infrared video using SleepVST. (Cohen's \\(\\kappa = 0.71\\) between model and expert sleep stages.)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.362, + 0.421 + ], + "angle": 0, + "content": "5.1. Pre-training on Contact Sensors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.563 + ], + "angle": 0, + "content": "From the SHHS dataset, we randomly selected 500 participants who participated in both visits, to form a test set of 1000 nights of data. This ensured that no participant appeared in both the training/validation and test partitions. Our test set is listed in \\(\\S H\\). From the MESA dataset, we used the same test set of 204 nights as Kotzen et al. [22]. The remaining nights from both the SHHS and MESA datasets were pooled together and randomly split into training and validation sets using a 75:25 split." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Table 1 compares the overall performance of SleepVST with prior methods for cardio-respiratory sleep staging from contact sensors. These have used input modalities such as the photoplethysmogram (PPG) waveform and the ECG-derived heart rate. Using the ECG and thoracic respiratory signals, SleepVST outperforms prior methods for sleep staging from cardio-respiratory signals on both datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.317, + 0.697 + ], + "angle": 0, + "content": "5.2. Video-based Sleep Staging" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.78 + ], + "angle": 0, + "content": "In this section, we report the performance of the SleepVST model after transfer to the OSV video dataset. We follow the evaluation procedure of [6], training and evaluating on the 50 nights of data using 10-fold cross-validation with non-overlapping folds each containing five nights of data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.841 + ], + "angle": 0, + "content": "In Table 2, we see that our method achieves significantly better performance in terms of both accuracy and Cohen's \\(\\kappa\\) statistic when compared with prior video-based methods evaluated on similar study populations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Figure 7 shows the total test confusion matrix between model and human expert (polysomnographer) classifications on the OSV dataset. We observe particularly high accuracy in distinguishing Wake and REM sleep stages." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.409, + 0.882, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.637, + 0.892, + 0.667 + ], + "angle": 0, + "content": "Figure 7. Confusion matrix between SleepVST and expert labels on four-class video based sleep staging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.861 + ], + "angle": 0, + "content": "Population performance. Figure 8 shows the distribution of Cohen's \\(\\kappa\\) values across participants from the OSV dataset plotted against age, sex and Fitzpatrick skin type[2] [14]. We observe a decrease in performance with age, as observed in prior sleep staging work [6, 53]. In older adults, measures of cardio-respiratory activity (e.g. heart rate variability) are known to decline with age [72]. This likely reduces the distance between sleep stages in cardiorespiratory input space and thus makes them harder to distinguish. Additional video data from older subjects would likely help to improve this, which is discussed further in §C." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.787, + 0.902 + ], + "angle": 0, + "content": "2 Plotted using representative skin tones from [67]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12484" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.089, + 0.775, + 0.105 + ], + "angle": 0, + "content": "Table 1. Comparison of four-class cardio-respiratory sleep staging methods using contact sensors." + }, + { + "type": "table", + "bbox": [ + 0.202, + 0.115, + 0.77, + 0.27 + ], + "angle": 0, + "content": "
DatasetNMethodModalitiesCohen's κAccuracy / %
κμκTAccμAccT
SHHS [43]296Bakker et al. [3]HR+Thor* +Air†-0.64-76.7
800Sridhar et al. [50]HR (ECG)0.650.6677.377.0
1000SleepVSTECG+Thor*0.730.7582.883.0
MESA [8]296Bakker et al. [3]HR+Thor* +Air†-0.68-79.8
194Sridhar et al. [50]HR (ECG)-0.69-80.0
204Kotzen et al. [22]PPG-0.73‡-82.6‡
204SleepVSTECG+Thor*0.760.7785.185.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.212, + 0.27, + 0.591, + 0.282 + ], + "angle": 0, + "content": "*Thoracic respiratory effort. †Nasal airflow. ‡From reported confusion matrices." + }, + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.31, + 0.719, + 0.325 + ], + "angle": 0, + "content": "Table 2. Comparison of four-class sleep staging methods using video cameras." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.335, + 0.825, + 0.456 + ], + "angle": 0, + "content": "
DatasetNMethodModalities**Cohen's κAccuracy / %
κμκTAccμAccT
Healthy infants [21]8Kamon et al. [21]Motion0.26-48.0-
Healthy adults [35]6Nochino et al. [35]Motion0.19-40.5-
HealthBed Study [60]46van Meulen et al. [60]HW*0.490.49‡67.967.9‡
Oxford Sleep Volunteers50Carter et al. [6]HR+BR+Motion0.610.64‡73.474.3‡
50SleepVSTHW* + BW† + Motion0.680.7177.778.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.155, + 0.457, + 0.751, + 0.468 + ], + "angle": 0, + "content": "**Camera-derived. *Cardiac pulse (heart) waveforms.† Respiration (breathing) waveforms.‡ From reported confusion matrices." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.513, + 0.459, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.739, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Figure 8. Scatter and box plot of Cohen's \\(\\kappa\\) against age, sex and Fitzpatrick skin type\\(^2\\) for individuals from the OSV dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Example SleepVST output. Figure 6 shows an example hypnogram generated entirely from near-infrared video using our method. This example corresponds to the median Cohen's \\(\\kappa\\) between model and expert from the OSV dataset. The model correctly identifies all three REM cycles, periods of deep sleep, and brief awakenings throughout the night. Additional hypnogram examples can be found in §C." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.666, + 0.523 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.574 + ], + "angle": 0, + "content": "Video transfer strategy. In Table 3, we compare the performance of three methods for transferring the SleepVST model to video data:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.576, + 0.892, + 0.606 + ], + "angle": 0, + "content": "1. Directly applying the pre-trained SleepVST model to cardiac and respiratory waveforms from video data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.607, + 0.809, + 0.621 + ], + "angle": 0, + "content": "2. Using SleepVST as a feature extractor and:" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.576, + 0.892, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.622, + 0.891, + 0.65 + ], + "angle": 0, + "content": "(a) training a new classification layer without using motion features." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.652, + 0.892, + 0.681 + ], + "angle": 0, + "content": "(b) training a new classification layer that also uses motion features i.e. our best-performing approach." + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.622, + 0.892, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.682, + 0.892, + 0.757 + ], + "angle": 0, + "content": "Here we see that training a new classifier for video data is essential to the effectiveness of our approach. This step likely allows the classifier to ignore parts of the learnt feature space that do not transfer between contact-sensor and video-derived waveforms i.e. sensor-specific features." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.771, + 0.892, + 0.8 + ], + "angle": 0, + "content": "Table 3. SleepVST video-based sleep staging performance for different model transfer strategies." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.81, + 0.887, + 0.884 + ], + "angle": 0, + "content": "
κμκTAccμ / %AccT / %
Direct apply0.5100.53666.767.3
Transfer w/o motion0.6600.68976.477.4
Transfer w/ motion0.6770.70877.778.8
" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12485" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.095, + 0.468, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.472, + 0.445 + ], + "angle": 0, + "content": "Figure 9. Example four-class hypnograms during fragmented sleep from the OSV dataset. The use of motion features improves accuracy during short awakenings between 22.00 and 00.00." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Motion features. Table 4 reports the effectiveness of the two main components of our motion feature set: the set of feature definitions \\( f_{i} \\) and the set of regions used \\( R \\). Additional ablations can be found in §C. Here we define and report an additional metric \\( \\kappa_{F} \\) which is the Cohen's \\( \\kappa \\) statistic computed over epochs near transitions to/from Wake (W) i.e. where \\( \\hat{y}_i^t\\neq \\hat{y}_i^{t + \\Delta} = W \\) or \\( \\hat{y}_i^t = W\\neq \\hat{y}_i^{t + \\Delta} \\) for any \\( \\Delta \\in [-2,2] \\) epochs. This metric better quantifies accuracy during fragmented sleep, such as in Figure 9." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.687 + ], + "angle": 0, + "content": "As measured by \\(\\kappa_{F}\\), we see that motion features aid classification during fragmented sleep. The ability to accurately measure sleep fragmentation is particularly important since it is linked with daytime sleepiness and decreased cognitive performance [29]." + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.706, + 0.434, + 0.721 + ], + "angle": 0, + "content": "Table 4. Ablation study of motion feature parameters." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.731, + 0.451, + 0.884 + ], + "angle": 0, + "content": "
AblationParameter Setnf*Cohen's κ
κTκF
Feature setfi ∈ {}00.6890.397
fi ∈ {f1, f2}360.6970.467
fi ∈ {f3, f4}540.7040.465
fi ∈ {f1, f2, f3, f4}900.7080.491
Region(s)R ∈ {H ∪ B}300.7020.461
R ∈ {H ∪ B, O}600.7040.468
R ∈ {H, B, O}900.7080.491
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.097, + 0.885, + 0.165, + 0.896 + ], + "angle": 0, + "content": "\\*No. features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "Cardio-respiratory inputs. In Table 5, we compare the performance of the SleepVST model with variants which use heart rate \\( \\boldsymbol{x}_{HR} \\) or breathing rate \\( \\boldsymbol{x}_{BR} \\) as inputs instead of the underlying waveforms, as described in Section 4.1. We see that the use of waveforms, rather than derived heart and breathing rates, leads to improvements in sleep staging accuracy across all datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.218, + 0.895, + 0.248 + ], + "angle": 0, + "content": "Table 5. Comparison of four-class sleep staging performance for different cardio-respiratory input combinations to SleepVST." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.257, + 0.825, + 0.365 + ], + "angle": 0, + "content": "
Input signalsCohen's κT
SHHSMESAOSV
xHR, xBR0.6830.7220.681
xHW, xBR0.7180.7360.684
xHR, xBW0.7230.7580.702
xHW, xBW0.7490.7650.708
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.628, + 0.426 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.894, + 0.514 + ], + "angle": 0, + "content": "We have introduced SleepVST, a transformer model for sleep stage classification from cardio-respiratory waveforms. Using a transfer learning approach, we have shown how SleepVST can be used to perform sleep stage classification entirely from near-infrared video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.895, + 0.67 + ], + "angle": 0, + "content": "Our results advance the state-of-the-art in video-based sleep staging, narrowing the gap to expert-level performance and taking a significant step towards important clinical applications. More broadly, our work shows that a model pre-trained on a task using cardio-respiratory waveforms from large contact-sensor datasets can be effectively transferred to waveforms measured from video. This approach can benefit wider applications such as the detection of arrhythmias and sleep apnoes, from cardiac and respiratory waveforms respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Finally, we believe there is important future work in learning more expressive representations of motion during sleep. In addition to improving the performance of video-based sleep staging, this could also enable the detection of sleep movement disorders, such as REM behaviour disorder, removing the existing need for time-consuming, manual video review [51]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.803, + 0.669, + 0.82 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.831, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This work was supported by the EPSRC Centre for Doctoral Training in Autonomous Intelligent Machines and Systems [EP/S024050/1] and funded by Oxehealth Ltd. Figure 1 was created with BioRender.com. We kindly thank the National Sleep Research Resource for granting access to SHHS and MESA." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12486" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Ying-Wen Bai, Wen-Tai Li, and Cheng-Hsiang Yeh. Design and implementation of an embedded monitor system for body breath detection by using image processing methods. In 2010 Digest of Technical Papers International Conference on Consumer Electronics (ICCE), pages 193-194, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[3] Jessie P. Bakker, Marco Ross, Ray Vasko, Andreas Cerny, Pedro Fonseca, Jeff Jasko, Edmund Shaw, David P. White, and Peter Anderer. Estimating sleep stages using cardiorespiratory signals: validation of a novel algorithm across a wide range of sleep-disordered breathing severity. Journal of Clinical Sleep Medicine, 17(7):1343-1354, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[4] Guha Balakrishnan, Fredo Durand, and John Guttag. Detecting Pulse from Head Motions in Video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, Portland, OR, USA, 2013. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.468, + 0.429 + ], + "angle": 0, + "content": "[5] Leo Breiman. Random Forests. Machine Learning, 45(1): 5-32, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[6] Jonathan Carter, João Jorge, Bindia Venugopal, Oliver Gibson, and Lionel Tarassenko. Deep Learning-Enabled Sleep Staging From Vital Signs and Activity Measured Using a Near-Infrared Video Camera. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 5940-5949, Vancouver, BC, Canada, 2023. IEEE. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[7] Weixuan Chen and Daniel McDuff. Deepphys: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 349-365, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[8] Xiaoli Chen, Rui Wang, Phyllis Zee, Pamela L. Lutsey, Sogol Javaheri, Carmela Alcantara, Chandra L. Jackson, Michelle A. Williams, and Susan Redline. Racial/Ethnic Differences in Sleep Disturbances: The Multi-Ethnic Study of Atherosclerosis (MESA). Sleep, 38(6):877-888, 2015. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[9] Jacob Cohen. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1): 37-46, 1960. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[10] Kris Cuppens, Lieven Lagae, Berten Ceulemans, Sabine Van Huffel, and Bart Vanrumste. Automatic video detection of body movement during sleep based on optical flow in pediatric patients with epilepsy. Medical & Biological Engineering & Computing, 48(9):923-931, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[11] Heidi Danker-Hopfe, Peter Anderer, Josef Zeitlhofer, Marion Boeck, Hans Dorn, Georg Gruber, Esther Heller, Erna Loretz, Doris Moser, Silvia Parapatics, Bernd Saletu, Andrea Schmidt, and Georg Dorffner. _Interrater reliability for sleep scoring according to the Rechtschaffen & Kales and the new AASM standard_. Journal of Sleep Research, 18(1):74-84, 2009. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1365-2869.2008.00700.x. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[12] Koldo De Miguel, Alberto Brunete, Miguel Hernando, and Ernesto Gambao. Home Camera-Based Fall Detection System for the Elderly. Sensors, 17(12):2864, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[14] T. B. Fitzpatrick. The validity and practicality of sun-reactive skin types I through VI. Archives of Dermatology, 124(6): 869-871, 1988. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Marco Hafner, Martin Stepanek, Jirka Taylor, Wendy M. Troxel, and Christian van Stolk. Why Sleep Matters—The Economic Costs of Insufficient Sleep. *Rand Health Quarterly*, 6(4):11, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[18] Adrienne Heinrich, Xavier Aubert, and Gerard de Haan. Body movement analysis during sleep based on video motion estimation. In 2013 IEEE 15th International Conference on e-Health Networking, Applications and Services (Healthcom 2013), pages 539–543, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[19] D. W. Hudgel, R. J. Martin, B. Johnson, and P. Hill. Mechanics of the respiratory system and breathing pattern during sleep in normal humans. Journal of Applied Physiology: Respiratory, Environmental and Exercise Physiology, 56(1): 133-137, 1984. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[20] C. Iber. The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology, and Technical Specification. 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[21] Masamitsu Kamon, Shima Okada, Masafumi Furuta, and Koki Yoshida. Development of a non-contact sleep monitoring system for children. Frontiers in Digital Health, 4, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.894, + 0.733 + ], + "angle": 0, + "content": "[22] Kevin Kotzen, Peter H. Charlton, Sharon Salabi, Lea Amar, Amir Landesberg, and Joachim A. Behar. SleepPPG-Net: A Deep Learning Algorithm for Robust Sleep Staging From Continuous Photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 27(2):924–932, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Yee Fun Lee, Dmitry Gerashchenko, Igor Timofeev, Brian J. Bacskai, and Ksenia V. Kastanenka. Slow Wave Sleep Is a Promising Intervention Target for Alzheimer's Disease. Frontiers in Neuroscience, 14, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Yun Ji Lee, Jae Yong Lee, Jae Hoon Cho, and Ji Ho Choi. Interrater reliability of sleep stage scoring: a meta-analysis. Journal of Clinical Sleep Medicine, 18(1):193-202, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Michael H. Li, Azadeh Yadollahi, and Babak Taati. A noncontact vision-based system for respiratory rate estimation. In 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pages 2119-2122, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12487" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Xin Liu, Brian Hill, Ziheng Jiang, Shwetak Patel, and Daniel McDuff. EfficientPhys: Enabling Simple, Fast and Accurate Camera-Based Cardiac Measurement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5008-5017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[28] Yawei Luo, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Taking a Closer Look at Domain Shift: Category-Level Adversaries for Semantics Consistent Domain Adaptation. pages 2507-2516, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[29] S E Martin, H M Engleman, I J Deary, and N J Douglas. The effect of sleep fragmentation on daytime function. American Journal of Respiratory and Critical Care Medicine, 153(4): 1328-1332, 1996. Publisher: American Thoracic Society - AJRCCM. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[30] Daniel McDuff. Camera Measurement of Physiological Vital Signs. ACM Computing Surveys, 55(9):1-40, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[31] Sara Mahvash Mohammadi, Shirin Enshaeifar, Adrian Hilton, Derk-Jan Dijk, and Kevin Wells. Transfer Learning for Clinical Sleep Pose Detection Using a Single 2D IR Camera. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 29:290-299, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[32] K. Nakajima, A. Osa, and H. Miike. A method for measuring respiration and physical activity in bed by optical flow analysis. In Proceedings of the 19th Annual International Conference of the IEEE Engineering in Medicine and Biology Society., pages 2054-2057 vol.5, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[33] Kazuki Nakajima, Yoshiaki Matsumoto, and Toshiyo Tamura. Development of real-time image sequence analysis for evaluating posture change and respiratory rate of a subject in bed. Physiological Measurement, 22(3):N21-N28, 2001. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[34] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[35] Teruaki Nochino, Yuko Ohno, Takafumi Kato, Masako Tani-ike, and Shima Okada. Sleep stage estimation method using a camera for home use. Biomedical Engineering Letters, 9 (2):257-265, 2019. 1, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[36] Jiapu Pan and Willis J. Tompkins. A Real-Time QRS Detection Algorithm. IEEE Transactions on Biomedical Engineering, BME-32(3):230-236, 1985. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[37] James Pardey, Stephen Roberts, Lionel Tarassenko, and John Stradling. A new approach to the analysis of the human sleep/wakefulness continuum. Journal of sleep research, 5 (4):201-210, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.817 + ], + "angle": 0, + "content": "[38] Aakash K. Patel, Vamsi Reddy, and John F. Araujo. Physiology, Sleep Stages. StatPearls Publishing, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[39] Huy Phan and Kaare Mikkelsen. Automatic sleep staging of EEG signals: recent development, challenges, and future directions. Physiological Measurement, 43(4):04TR01, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Huy Phan, Oliver Y. Chén, Minh C. Tran, Philipp Koch, Alfred Mertins, and Maarten De Vos. XSleepNet: Multi-View" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Sequential Model for Automatic Sleep Staging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 5903-5915, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Huy Phan, Kaare Mikkelsen, Oliver Y. Chen, Philipp Koch, Alfred Mertins, and Maarten De Vos. SleepTransformer: Automatic Sleep Staging With Interpretability and Uncertainty Quantification. IEEE Transactions on Biomedical Engineering, 69(8):2456-2467, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[42] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Advancements in Noncontact, Multiparameter Physiological Measurements Using a Webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. Conference Name: IEEE Transactions on Biomedical Engineering. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[43] S. F. Quan, B. V. Howard, C. Iber, J. P. Kiley, F. J. Nieto, G. T. O'Connor, D. M. Rapoport, S. Redline, J. Robbins, J. M. Samet, and P. W. Wahl. The Sleep Heart Health Study: design, rationale, and methods. Sleep, 20(12):1077-1085, 1997. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[44] Mustafa Radha, Pedro Fonseca, Arnaud Moreau, Marco Ross, Andreas Cerny, Peter Anderer, Xi Long, and Ronald M. Aarts. A deep transfer learning approach for wearable sleep stage classification with photoplethysmography. npj Digital Medicine, 4(1):1-11, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[45] S.J. Redmond and C. Heneghan. Cardiorespiratory-based sleep staging in subjects with obstructive sleep apnea. IEEE Transactions on Biomedical Engineering, 53(3):485-496, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[46] A. J. Schwichtenberg, Jeehyun Choe, Ashleigh Kellerman, Emily A. Abel, and Edward J. Delp. Pediatric Videosomnography: Can Signal/Video Processing Distinguish Sleep and Wake States? Frontiers in Pediatrics, 6:158, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[47] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90(2):227-244, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[48] Z. Shinar, A. Baharav, Y. Dagan, and S. Akselrod. Automatic detection of slow-wave-sleep using heart rate variability. In Computers in Cardiology 2001. Vol.28, pages 593-596, 2001. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[49] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[50] Niranjan Sridhar, Ali Shoeb, Philip Stephens, Alaa Kharbouch, David Ben Shimol, Joshua Burkart, Atiyeh Ghoreyshi, and Lance Myers. Deep learning for automated sleep staging using instantaneous heart rate. npj Digital Medicine, 3(1):1-10, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[51] Ambra Stefani and Birgit Hogl. Sleep in Parkinson's disease. Neuropsychopharmacology, 45(1):121-128, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[52] Ambra Stefani, David Gabelaia, Thomas Mitterling, Werner Poewe, Birgit Hogl, and Birgit Frauscher. A Prospective Video-Polysomnographic Analysis of Movements during Physiological Sleep in 100 Healthy Sleepers. Sleep, 38 (9):1479-1487, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Haoqi Sun, Wolfgang Ganglberger, Ezhil Panneerselvam, Michael J Leone, Syed A Quadri, Balaji Goparaju, Ryan A" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12488" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "Tesh, Oluwaseun Akeju, Robert J Thomas, and M Brandon Westover. Sleep staging from electrocardiography and respiration with deep learning. *Sleep*, 43(7):zsz306, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[54] K. S. Tan, R. Saatchi, H. Elphick, and D. Burke. Real-time vision based respiration monitoring system. In 2010 7th International Symposium on Communication Systems, Networks & Digital Signal Processing (CSNDSP 2010), pages 770-774, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.47, + 0.279 + ], + "angle": 0, + "content": "[55] Alexander Trumpp, Johannes Lohr, Daniel Wedekind, Martin Schmidt, Matthias Burghardt, Axel R. Heller, Hagen Malberg, and Sebastian Zaunseder. Camera-based photoplethysmography in an intraoperative setting. *BioMedical Engineering OnLine*, 17(1):33, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[56] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Oxhealth Vital Signs, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200019.pdf.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[57] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Gili Pro Biosensor, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200038.pdf.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.399, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[58] Mark van Gastel, Sander Stuijk, Sebastiaan Overeem, Johannes P. van Dijk, Merel M. van Gilst, and Gerard de Haan. Camera-Based Vital Signs Monitoring During Sleep - A Proof of Concept Study. IEEE Journal of Biomedical and Health Informatics, 25(5):1409-1418, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.47, + 0.526 + ], + "angle": 0, + "content": "[59] Hans van Gorp, Iris A M Huijben, Pedro Fonseca, Ruud J G van Sloun, Sebastiaan Overeem, and Merel M van Gilst. Certainty about uncertainty in sleep staging: a theoretical framework. Sleep, 45(8), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.47, + 0.611 + ], + "angle": 0, + "content": "[60] Fokke B. van Meulen, Angela Grassi, Leonie van den Heuvel, Sebastiaan Overeem, Merel M. van Gilst, Johannes P. van Dijk, Henning Maass, Mark J. H. van Gastel, and Pedro Fonseca. Contactless Camera-Based Sleep Staging: The HealthBed Study. Bioengineering, 10(1):109, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. arXiv:1706.03762 [cs], 2017. arXiv: 1706.03762. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[62] Mauricio Villarroel, Sitthichok Chaichulee, João Jorge, Sara Davis, Gabrielle Green, Carlos Arteta, Andrew Zisserman, Kenny McCormick, Peter Watkinson, and Lionel Tarassenko. Non-contact physiological monitoring of preterm infants in the Neonatal Intensive Care Unit. npj Digital Medicine, 2(1):1-18, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[63] Olivia Walch, Yitong Huang, Daniel Forger, and Cathy Goldstein. Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device. *Sleep*, 42(12), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[64] Qiongyan Wang, Hanrong Cheng, and Wenjin Wang. Feasibility of Exploiting Physiological and Motion Features for Camera-based Sleep Staging: A Clinical Study. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5, 2023. ISSN: 2694-0604. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[65] Wenjin Wang and Albertus C. den Brinker. Algorithmic insights of camera-based respiratory motion extraction. Physiological measurement, 43(7):075004, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[66] Wenjin Wang, Albertus C. den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic Principles of Remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[67] William H. Ward, Fernando Lambreton, Neha Goel, Jian Q. Yu, and Jeffrey M. Farma. Clinical Presentation and Staging of Melanoma. In Cutaneous Melanoma: Etiology and Therapy. Codon Publications, Brisbane (AU), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[68] Johanna Wilde-Frenz and Hartmut Schulz. Rate and Distribution of Body Movements during Sleep in Humans. Perceptual and Motor Skills, 56(1):275-283, 1983. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[69] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. SimPer: Simple Self-Supervised Learning of Periodic Targets, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. PhysFormer: Facial Video-based Physiological Measurement with Temporal Difference Transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4186, New Orleans, LA, USA, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[71] Guo-Qiang Zhang, Licong Cui, Remo Mueller, Shiqiang Tao, Matthew Kim, Michael Rueschman, Sara Mariani, Daniel Mobley, and Susan Redline. The National Sleep Research Resource: towards a sleep data commons. Journal of the American Medical Informatics Association: JAMIA, 25 (10):1351-1358, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[72] Dan Ziegler, G. Laux, K. Dannehl, M. Spüler, H. Mühlen, P. Mayer, and F.a. Gries. Assessment of Cardiovascular Autonomic Function: Age-related Normal Ranges and Reproducibility of Spectral Analysis, Vector Analysis, and Standard Tests of Heart Rate Variation and Blood Pressure Responses. Diabetic Medicine, 9(2):166-175, 1992. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12489" + } + ] +] \ No newline at end of file diff --git a/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_origin.pdf b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..58552bbfaec5551923d1e9c5cf973270c6b0029e --- /dev/null +++ b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/fb8d4880-9157-4c04-a223-f42e3fd034fb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b091b263e32f0a9b1390b449ff76f6a8d28b0e27ea8884dcfcb4ec2fe14b48ea +size 2931077 diff --git a/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/full.md b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ecffdf52e1a34066b860d101e39efde69c7a276 --- /dev/null +++ b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/full.md @@ -0,0 +1,351 @@ +# SleepVST: Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers + +Jonathan F. Carter $^{1,2}$ João Jorge $^{2}$ Oliver Gibson $^{2}$ Lionel Tarassenko $^{1}$ + +$^{1}$ Institute of Biomedical Engineering, University of Oxford $^{2}$ Oxehealth Ltd., Oxford + +# Abstract + +Advances in camera-based physiological monitoring have enabled the robust, non-contact measurement of respiration and the cardiac pulse, which are known to be indicative of the sleep stage. This has led to research into camera-based sleep monitoring as a promising alternative to "gold-standard" polysomnography, which is cumbersome, expensive to administer, and hence unsuitable for longer-term clinical studies. In this paper, we introduce SleepVST, a transformer model which enables state-of-the-art performance in camera-based sleep stage classification (sleep staging). After pre-training on contact sensor data, SleepVST outperforms existing methods for cardiorespiratory sleep staging on the SHHS and MESA datasets, achieving total Cohen's kappa scores of 0.75 and 0.77 respectively. We then show that SleepVST can be successfully transferred to cardio-respiratory waveforms extracted from video, enabling fully contact-free sleep staging. Using a video dataset of 50 nights, we achieve a total accuracy of $78.8\%$ and a Cohen's $\kappa$ of 0.71 in four-class video-based sleep staging, setting a new state-of-the-art in the domain. + +# 1. Introduction + +Accurate sleep monitoring is critical for the diagnosis of sleep disorders and the discovery of novel treatments and biomarkers. Poor sleep is broadly linked with a number of health conditions, including cardiovascular diseases such as diabetes and hypertension [15]. Additionally, there are links between specific sleep stages and neurodegenerative conditions, such as a decrease in non-rapid-eye-movement stage 3 (N3) sleep with Alzheimer's disease [23]. + +Overnight video polysomnography (vPSG), the "gold-standard" [52] technique for sleep monitoring, requires the + +use of a large number of contact sensors including electrodes placed on the scalp (EEG), near the eyes (EOG), and under the chin (EMG) of the patient. Human experts (polysomnographers) must then review the recorded data, classifying the patient's sleep into five stages at 30-second intervals (epochs) and annotating other important events such as leg movements and apnoeas. This manual process often takes multiple hours to complete. + +Experts primarily rely on characteristic patterns in brain activity measured from the EEG to classify sleep stages [20]. However, sleep stage information is also encoded in measures of autonomic activity, including cardiac [48] and respiratory signals [19], and body movements [68]. This has led to the development of methods for automatically classifying sleep stages from sensors that measure these parameters, such as smartwatches [63]. + +Prior work has shown that these physiological parameters can also be measured using video cameras [18, 65, 66], leading to the development of methods for classifying sleep stages entirely from video input [6, 35, 60]. Camera-based methods have particular promise as part of a multipurpose sleep monitoring system. For example, within elderly care settings, in addition to classifying sleep stages, they could also be used to detect specific sleep movement disorders [18], which are linked with conditions such as Parkinson's, and which are typically distinguished via manual video review [51]. They could also be used to detect falls [12], one of the leading causes of injury and mortality amongst elderly individuals. + +State-of-the-art video-based sleep staging [6] has used the heart rate (HR) and breathing rate (BR) derived from video to perform sleep staging. However, these rates are typically measured as averages over time, losing higher frequency inter-pulse and inter-breath interval information. EEG- [40] and wearable-based [22] sleep staging methods have both been improved by using raw waveforms as inputs, rather than derived quantities e.g. signal features [37, 45]. + +![](images/a915d98d330c93a316f9c2cff731f64bbeacff570a48d1addd7b205a04b1e161.jpg) +Figure 1. Sleep staging from near-infrared video signals using SleepVST. (a) We first pre-train the model end-to-end on cardiac (heart) and respiratory (breathing) waveforms $x_{HW}(t)$ and $x_{BW}(t)$ derived from the electrocardiogram (ECG) and a thoracic respiratory belt (THX) respectively. (b) After pre-training, we use the model as a frozen feature extractor, applying it to cardio-respiratory waveforms derived from near-infrared (NIR) video to generate sequences of features. When transferring to video data, we additionally use a set of motion features $f_{i}(t)$ derived from an optical flow field $\pmb{u}(x,y,t)$ as inputs to the classifier C. This approach allows us to utilise much larger contact-sensor datasets to train SleepVST, whilst also enabling the incorporation of motion information when transferring to video data. (c) We evaluate SleepVST using overnight video polysomnography (vPSG) studies, comparing expert-labelled sleep stage sequences $\hat{y}_{1:T}$ against those generated entirely from near-infrared video $y_{1:T}$ using our method. + +In this paper, we introduce SleepVST1, a transformer model for sleep stage classification from cardio-respiratory waveforms. First, we pre-train the model on waveforms derived from contact sensors (Fig. 1a), after which it outperforms existing methods for cardio-respiratory sleep staging on two widely used public datasets, highlighting the effectiveness of our transformer architecture. + +After pre-training, we then show that SleepVST can be transferred to cardiac and respiratory waveforms measured entirely from near-infrared video, where our method additionally enables us to incorporate motion features derived from the camera data (Fig. 1b). Using this novel combination of inputs available from a near-infrared camera, we obtain state-of-the-art results for video-based sleep staging. Through a series of ablation studies we demonstrate the effectiveness of our approach. Our work further highlights the potential of camera-based sleep monitoring for important clinical applications. + +# 2. Background and Related Work + +Automated sleep staging. To reduce the manual effort of sleep assessment, deep learning methods have been developed for the automatic classification of sleep stages from polysomnography (PSG) recordings, achieving near-human accuracy [39]. Accuracy is commonly measured using Cohen's kappa statistic $(\kappa, [9])$ , which measures inter-rater agreement between two sources. Expert annotation of sleep stages is a subjective process, making $100\%$ agreement unachievable [59]. Prior work has shown that sleep scorers agree around $80\%$ of the time ( $\kappa = 0.71 - 0.81$ , [11, 24]) when classifying sleep stages according to American Academy of + +Sleep Medicine (AASM, [20]) scoring rules. These rules divide sleep into five classes: Wake, N1 (light), N2 (intermediate), N3 (deep) and rapid-eye-movement (REM) sleep. However, due to low inter-scorer agreement in identifying N1 [11], prior work has often treated sleep classification as a four-class problem, by merging N1 and N2 into a single class [22, 50]. + +Given the drawbacks of conventional polysomnography, prior work has investigated the potential of automatic sleep staging from a reduced number of contact sensors, including the electrocardiogram ([50], $\kappa = 0.66$ ), and the photoplethysmogram ([44], $\kappa = 0.65$ ). + +Camera-based physiological measurement. Respiratory measurement is possible from video by measuring the induced movements of the chest and abdomen [65]. This has been achieved through methods such as frame-differencing [2, 54], using dense optical flow magnitudes [32, 33] and by tracking the movement of sparse optical flow vectors [25] across video frames. + +Camera-based measurement of the pulse is possible using methods which measure specific cardiac phenomena [30]. These include remote photoplethysmography (rPPG, [42, 66]), which measures skin colour changes caused by variations in blood volume over the cardiac cycle, and ballistocardiography [4], which measures head movements induced as blood is pumped through the carotid arteries. More recently, deep learning methods have been developed which learn to extract a pulse signal directly from input video frames [7, 26, 70] rather than targeting specific underlying cardiac phenomena, including unsupervised approaches [49, 69]. + +Camera-based vital-sign measurement has been validated within clinical settings [55, 62], during overnight + +polysomnography recordings [58, 64], and has led to products with regulatory clearance for camera-based measurement of the heart and breathing rate, including the Gili Pro Biosensor [57] and Oxehealth Vital Signs software [56]. + +Camera-based sleep monitoring. Prior work has investigated the binary problem of distinguishing sleep from wake using video cameras, using measures of activity derived from the input frames [10, 33, 46]. Whilst activity information alone can be effective in distinguishing sleep from wake, cardio-respiratory signals carry important additional information for classifying specific sleep stages [19, 48]. The current state-of-the-art approach for video-based sleep staging ([6], $\kappa = 0.61$ ) combines multiple modalities derived from video: the heart rate, breathing rate and activity information, to perform sleep stage classification. This approach outperforms prior methods which had used only a single video-derived modality: motion [21, 35] or the cardiac pulse [60]. + +# 3. Datasets and Preprocessing + +# 3.1. Contact Sensor Data + +To train the SleepVST model, we used the SHHS [43, 71] and MESA [8] contact sensor datasets. These datasets contain PSG recordings and sleep stages annotated according to AASM scoring rules. The cardio-respiratory signals available in each recording include the electrocardiogram (ECG), which measures electrical activity of the heart, and respiratory signals, measured using respiratory bands placed around the abdomen and thorax. + +From the available signals, we seek to produce generic cardiac and respiratory waveforms to train the SleepVST model, preventing the model from learning to use information which may be present in signals derived from contact sensors but not present in those derived from video. This is to mitigate the problem of domain shift [28, 47], where the performance of a model is significantly reduced when applied to a different data distribution. + +To produce a cardiac (heart) waveform, we first applied the same filtering steps used by the Pan-Tompkins algorithm [36] to the ECG signal. We resampled the output from the moving-window integration step of the algorithm to $10\mathrm{Hz}$ , then applied a Butterworth bandpass filter with a passband of $0.66 - 2.8\mathrm{Hz}$ (40-168 beats per minute or BPM) and Gaussian smoothing. To produce a respiratory (breathing) waveform, we used the thoracic respiratory signal as the input source, as previously used in [3]. This was downsampled to $5\mathrm{Hz}$ and smoothed using a median filter with a kernel of length 5, to remove artefacts. + +# 3.2. Video Data + +We used the dataset previously introduced in [6] for our video-based sleep staging experiments. This dataset con + +sists of overnight vPSG recordings from 50 volunteers, with sleep stages annotated according to the AASM guidelines at 30-second intervals (epochs) by a sleep physiologist. Throughout this work, we refer to this as the Oxford Sleep Volunteers (OSV) dataset. Each recording includes $850\mathrm{nm}$ near-infrared (NIR) video data captured at 20 FPS. An example camera view can be seen in Figure 5. More information on the dataset is provided in $\S B$ . + +To derive cardiac and respiratory waveforms from the video data, we used components of FDA-cleared Oxhealth Vital Signs software [56]. For each overnight video recording, we used the software to obtain cardiac and respiratory waveforms sampled at $10\mathrm{Hz}$ and $5\mathrm{Hz}$ respectively. We then applied the same filtering steps applied to the contact sensor waveforms: Butterworth filtering and Gaussian smoothing to the heart waveforms, and median filtering to the breathing waveforms. + +Figure 2 shows examples of processed cardio-respiratory waveforms measured from contact sensors and video for a short section of a recording from the OSV dataset. These have additionally been normalised to zero mean and unit variance. + +![](images/13a2e06f84183abdde81392f75867ab555309f312b8bc93e075e587686f53b08.jpg) +Figure 2. Example (normalized) cardiac and respiratory waveforms, $x_{HW}(t)$ and $x_{BW}(t)$ , derived from contact sensors (blue) and video (orange) from the OSV dataset. + +# 4. Methods + +# 4.1. SleepVST Architecture + +Figure 3 illustrates the SleepVST model architecture, which turns cardiac and respiratory waveforms into sequences of feature vectors using a transformer encoder. These feature vectors are then used to classify each 30-second sleep epoch. During pre-training, we used a linear layer for classification. When transferring to video data, we trained a new classifier that additionally used motion features derived from video as inputs, discussed further in Section 4.2. + +Patchification. The continuous heart and breathing waveforms (HW and BW) are divided into N windows i.e. patches, drawing on ideas from PatchTST [34]. We used non-overlapping 30-second windows since this is the inter + +![](images/8d940f7e4bb5edf9b1e04902ea003124a51496629d71b795a736b64fcad25298.jpg) +Figure 3. SleepVST architecture. Each 30-second window of heart $(x_{HW})$ and breathing waveforms $(x_{BW})$ is passed to a patch encoder, which turns them into patch-level features. These features are concatenated and passed to a transformer encoder. During pre-training, a linear layer turns the output feature sequence $z_{o}$ of length N from SleepVST into sleep stage classifications. + +val at which sleep stages are classified according to AASM guidelines. We denote the patchified inputs as $\boldsymbol{x}_{HW} \in \mathbb{R}^{300 \times N}$ and $\boldsymbol{x}_{BW} \in \mathbb{R}^{150 \times N}$ respectively, where $N$ is the sequence length. + +Encodings. Each waveform patch is normalised to zero mean and unit variance before being passed to encoding layers $E_{HW}$ and $E_{BW}$ , which transform them into feature vectors $\pmb{z}_{HW} \in \mathbb{R}^{D_{HW} \times N}$ and $\pmb{z}_{BW} \in \mathbb{R}^{D_{BW} \times N}$ . The feature vectors from each patch are then concatenated, producing a sequence of feature vectors $\pmb{z}_i \in \mathbb{R}^{(D_{HW} + D_{BW}) \times N}$ which are passed to a transformer encoder. + +We used identical, shallow 1D ResNet-style [17] models for $E_{HW}$ and $E_{BW}$ , illustrated in Figure 4. This convolutional encoder design [1] compresses the waveform patches into lower-dimensional feature vectors. This allowed us to keep $D_{HW}$ and $D_{BW}$ small, reducing the computational and memory complexity of the downstream transformer and the likelihood of over-fitting, particularly when transferring the model to the smaller video dataset. + +This encoder design additionally provides a high degree of flexibility at the input level, such as the ability to use input signals of differing frequencies. In later ablation studies, this enables us to compare the performance of the SleepVST model with simpler variants which use 30-second patches of $1\mathrm{Hz}$ heart rate $\pmb{x}_{HR} \in \mathbb{R}^{30 \times N}$ and/or breathing rate $\pmb{x}_{BR} \in \mathbb{R}^{30 \times N}$ time-series as inputs, rather than underlying waveforms. When using either of these as inputs, we used linear projections [13, 34] for the input encoders, because of the lower dimensionality of $\pmb{x}_{HR}$ and $\pmb{x}_{BR}$ . Throughout our experiments we use the same feature dimension i.e. $D_{HW} = D_{BW} = D_{HR} = D_{BR} = 64$ . + +![](images/cd35a6b7e6e1cf582dc432a7f0dc609018fcec645f9926444d7cf0e0429209a1.jpg) +Figure 4. Waveform encoder design. Using a series of convolutional layers, the encoder turns sequences of signal patches into sequences of lower-dimensional feature vectors. + +Transformer layer. We used a Transformer encoder [61] with sinusoidal position encodings to turn the sequence of input features $\mathbf{z}_i$ into an output sequence of features $\mathbf{z}_o$ with an identical feature dimension. Normal sleep exhibits long-range structures, known as sleep cycles [38], which typically last around 90-120 minutes and typically occur four to five times per night. From its superior performance in many sequence modelling tasks, including EEG-based sleep staging [41], we expect that the inductive biases of the transformer model (i.e. the attention mechanism) can enable it to learn to use relative position within a sleep cycle to classify sleep stages. + +Classification. When training and testing on contact sensor data, we used a linear classification layer $\mathbf{W}_C \in \mathbb{R}^{C \times (D_{HW} + D_{BW})}$ to transform the sequence of output feature vectors $\mathbf{z}_o$ into a sequence of sleep stage probabilities $\mathbf{y} \in \mathbb{R}^{C \times N}$ , where the number of classes $C$ depends on the sleep classification strategy used. In line with prior work, we classified sleep stages into four classes during pretraining: Wake-N1/N2-N3-REM, i.e. $C = 4$ . + +Model training and inference. In all experiments, we used a sequence length of $N = 240$ windows i.e. two hours. During training, we created batches by sampling sequences of length $N$ with a step of 5 minutes from within each recording. The model was trained to minimise the multiclass cross-entropy loss using the AdamW [27] optimiser with a learning rate of 0.0005 and weight decay of 0.001. Other hyper-parameters, such as the no. encoder layers, are detailed in §D. To apply the model to arbitrary length input sequences during evaluation, we re-applied it at regular intervals. More details on this process are provided in §E. + +# 4.2. Motion features from Video + +Following prior work [35], we seek to produce motion features from video which encode for information which is pre + +dictive of the sleep stage, such as the rate and distribution of body movements [68]. To do so, we define a set of parameterised time-series features that collectively quantify motion over a range of spatial and temporal scales. + +Optical flow estimation. We first estimated a 2D optical flow field $\pmb{u}(t,x,y) \in \mathbb{R}^2$ over the bed region. Inspired by sleep pose detection work [31], we applied a homography transformation [16] to the individual video frames before estimating the flow field, to achieve camera viewpoint invariance (shown in Figure 5). + +![](images/d62adc07c5c05ad174f74c1c78cd6c8a43f496c74b8ac3715d3a5d004de9300b.jpg) +(a) + +![](images/5f43ab59badbe965d2739e0e1735184249d2d76bb15410cb12d4ee8051b20c94.jpg) +(b) + +![](images/35d473d2c27d3b528714d70946bd483a1a99b46eea75f67d05e4d225ce19bb26.jpg) +(c) +Figure 5. Example processing of an NIR video frame from the OSV dataset. (a) Real viewpoint. (b) Virtual viewpoint. (c) Head (H), body (B), and outer (O) bed regions. The distance from the camera to the head region is around $1.5\mathrm{m}$ . + +Feature definitions. Given an arbitrary region $\mathbf{R}$ of the optical flow field, we define the following parametrised signals: + +$$ +v (t; R) = \max _ {(x, y) \in R} | \boldsymbol {u} (t, x, y) | \tag {1} +$$ + +$$ +s (t; R) = \frac {1}{| R |} \sum_ {(x, y) \in R} | \boldsymbol {u} (t, x, y) | \tag {2} +$$ + +i.e. the maximum (1) and the average (2) of the optical flow magnitude in the region R at time $t$ . Using these signals, we define two time-series features that quantify motion within a time window $\Delta$ up to time $t$ : + +$$ +f _ {1} (t; R, \Delta) = \sum_ {t - \Delta} ^ {t} v (t; R) \tag {3} +$$ + +$$ +f _ {2} (t; R, \Delta) = \sum_ {t - \Delta} ^ {t} s (t; R) \tag {4} +$$ + +Plus two additional features which measure the time elapsed since either signal went above a threshold $\delta$ : + +$$ +f _ {3} (t; R, \delta) = \sum_ {t - \tau} ^ {t} 1 +$$ + +$$ +\text {w h e r e} \quad \tau = \max _ {t ^ {\prime} \leq t} t ^ {\prime} \tag {5} +$$ + +$$ +\begin{array}{c c} \text {s . t .} & v (t ^ {\prime}; R) > \delta \end{array} +$$ + +$$ +f _ {4} (t; R, \delta) = \sum_ {t - \tau} ^ {t} 1 +$$ + +$$ +\text {w h e r e} \quad \tau = \max _ {t ^ {\prime} \leq t} t ^ {\prime} \tag {6} +$$ + +$$ +\begin{array}{l l} \text {s . t .} & s (t ^ {\prime}; R) > \delta \end{array} +$$ + +We calculated each feature for all possible combinations of parameters $\Delta \in \{30,300\}$ seconds, $\delta \in \{0.01,0.1,1\}$ and $R \in \{H,B,O\}$ , where $\{H,B,O\}$ are approximate head, body and outer bed regions as illustrated in Figure 5. This resulted in $(2 + 2 + 3 + 3)$ time-series features per region, i.e. 30 unique time-series features. Additionally, we provided the video-based classifier (see Section 4.3) with time-shifted features $f_{i}(t + T)$ for $T \in \{-90,0, + 90\}$ seconds. This gave a total of 90 motion features for each sleep epoch. + +# 4.3. Transferring SleepVST to Video + +To transfer the SleepVST model to video using the OSV dataset, we removed the linear classification layer learnt during pre-training and used the model as a frozen feature extractor, applying it to sequences of cardiac and respiratory waveforms measured from video to produce sequences of output features $\mathbf{z}_o$ . We then trained a new classifier that used these features, plus the motion features derived in the previous section, to classify individual sleep epochs. This allowed us to leverage the representational capacity of the transformer model, by pre-training on the much larger contact-sensor datasets, whilst also enabling us to incorporate motion information from the video dataset. We followed the design choice of [6] and used a Random Forest [5] for our classifier due to its low computational cost and robustness to overfitting. + +# 5. Experimental Results + +Sleep staging performance has commonly been reported in terms of classification accuracy $(\mathrm{Acc}_T)$ and total Cohen's kappa $(\kappa_T)$ over all sleep epochs in the dataset, and/or using per-night mean statistics, $\mathrm{Acc}_{\mu}$ and $\kappa_{\mu}$ , i.e.: + +$$ +\kappa_ {T} = K \left(\sum_ {i = 1} ^ {N} c \left(\boldsymbol {y} _ {i}, \hat {\boldsymbol {y}} _ {i}\right)\right) \tag {7} +$$ + +$$ +\kappa_ {\mu} = \frac {1}{N} \sum_ {i = 1} ^ {N} K \left(c \left(\boldsymbol {y} _ {i}, \hat {\boldsymbol {y}} _ {i}\right)\right) \tag {8} +$$ + +where $\pmb{y}_i, \hat{\pmb{y}}_i \in \mathbb{R}^{C \times L_i}$ are the one-hot encoded model and scorer labels for recording $i$ of length $L_i$ and $c(\pmb{y}_i, \hat{\pmb{y}}_i) \in \mathbb{R}^{C \times C}$ is the confusion matrix between them. $K(\cdot)$ is the function which calculates the Cohen's kappa statistic [9] from a confusion matrix and is defined in §G. We report all of these statistics, to make fairer comparisons with prior work, and to aid future comparisons. + +![](images/77d3913bfd21dd10d4b8964c9f7266bc0c1d88709d6c0c2dfa38dfc559816add.jpg) +Figure 6. Example four-class sleep hypnograms from the OSV dataset. (top) Scored by an expert using signals from the vPSG recording. (bottom) Automatically generated from near-infrared video using SleepVST. (Cohen's $\kappa = 0.71$ between model and expert sleep stages.) + +# 5.1. Pre-training on Contact Sensors + +From the SHHS dataset, we randomly selected 500 participants who participated in both visits, to form a test set of 1000 nights of data. This ensured that no participant appeared in both the training/validation and test partitions. Our test set is listed in $\S H$ . From the MESA dataset, we used the same test set of 204 nights as Kotzen et al. [22]. The remaining nights from both the SHHS and MESA datasets were pooled together and randomly split into training and validation sets using a 75:25 split. + +Table 1 compares the overall performance of SleepVST with prior methods for cardio-respiratory sleep staging from contact sensors. These have used input modalities such as the photoplethysmogram (PPG) waveform and the ECG-derived heart rate. Using the ECG and thoracic respiratory signals, SleepVST outperforms prior methods for sleep staging from cardio-respiratory signals on both datasets. + +# 5.2. Video-based Sleep Staging + +In this section, we report the performance of the SleepVST model after transfer to the OSV video dataset. We follow the evaluation procedure of [6], training and evaluating on the 50 nights of data using 10-fold cross-validation with non-overlapping folds each containing five nights of data. + +In Table 2, we see that our method achieves significantly better performance in terms of both accuracy and Cohen's $\kappa$ statistic when compared with prior video-based methods evaluated on similar study populations. + +Figure 7 shows the total test confusion matrix between model and human expert (polysomnographer) classifications on the OSV dataset. We observe particularly high accuracy in distinguishing Wake and REM sleep stages. + +![](images/b9d84e28653074c427a8cd06f230fe54cc96db6dbead15932cabe78b06086e8f.jpg) +Figure 7. Confusion matrix between SleepVST and expert labels on four-class video based sleep staging. + +Population performance. Figure 8 shows the distribution of Cohen's $\kappa$ values across participants from the OSV dataset plotted against age, sex and Fitzpatrick skin type[2] [14]. We observe a decrease in performance with age, as observed in prior sleep staging work [6, 53]. In older adults, measures of cardio-respiratory activity (e.g. heart rate variability) are known to decline with age [72]. This likely reduces the distance between sleep stages in cardiorespiratory input space and thus makes them harder to distinguish. Additional video data from older subjects would likely help to improve this, which is discussed further in §C. + +Table 1. Comparison of four-class cardio-respiratory sleep staging methods using contact sensors. + +
DatasetNMethodModalitiesCohen's κAccuracy / %
κμκTAccμAccT
SHHS [43]296Bakker et al. [3]HR+Thor* +Air†-0.64-76.7
800Sridhar et al. [50]HR (ECG)0.650.6677.377.0
1000SleepVSTECG+Thor*0.730.7582.883.0
MESA [8]296Bakker et al. [3]HR+Thor* +Air†-0.68-79.8
194Sridhar et al. [50]HR (ECG)-0.69-80.0
204Kotzen et al. [22]PPG-0.73‡-82.6‡
204SleepVSTECG+Thor*0.760.7785.185.2
+ +*Thoracic respiratory effort. †Nasal airflow. ‡From reported confusion matrices. + +Table 2. Comparison of four-class sleep staging methods using video cameras. + +
DatasetNMethodModalities**Cohen's κAccuracy / %
κμκTAccμAccT
Healthy infants [21]8Kamon et al. [21]Motion0.26-48.0-
Healthy adults [35]6Nochino et al. [35]Motion0.19-40.5-
HealthBed Study [60]46van Meulen et al. [60]HW*0.490.49‡67.967.9‡
Oxford Sleep Volunteers50Carter et al. [6]HR+BR+Motion0.610.64‡73.474.3‡
50SleepVSTHW* + BW† + Motion0.680.7177.778.8
+ +**Camera-derived. *Cardiac pulse (heart) waveforms.† Respiration (breathing) waveforms.‡ From reported confusion matrices. + +![](images/2f4a18a5fbd1f584f58d720232e940c2dc5bdf04f9e91452845d9df4589524aa.jpg) +Figure 8. Scatter and box plot of Cohen's $\kappa$ against age, sex and Fitzpatrick skin type $^2$ for individuals from the OSV dataset. + +Example SleepVST output. Figure 6 shows an example hypnogram generated entirely from near-infrared video using our method. This example corresponds to the median Cohen's $\kappa$ between model and expert from the OSV dataset. The model correctly identifies all three REM cycles, periods of deep sleep, and brief awakenings throughout the night. Additional hypnogram examples can be found in §C. + +# 5.3. Ablation Studies + +Video transfer strategy. In Table 3, we compare the performance of three methods for transferring the SleepVST model to video data: + +1. Directly applying the pre-trained SleepVST model to cardiac and respiratory waveforms from video data. +2. Using SleepVST as a feature extractor and: + +(a) training a new classification layer without using motion features. +(b) training a new classification layer that also uses motion features i.e. our best-performing approach. + +Here we see that training a new classifier for video data is essential to the effectiveness of our approach. This step likely allows the classifier to ignore parts of the learnt feature space that do not transfer between contact-sensor and video-derived waveforms i.e. sensor-specific features. + +Table 3. SleepVST video-based sleep staging performance for different model transfer strategies. + +
κμκTAccμ / %AccT / %
Direct apply0.5100.53666.767.3
Transfer w/o motion0.6600.68976.477.4
Transfer w/ motion0.6770.70877.778.8
+ +![](images/c35a369e0fc4a80ccef27bd57c870129abc084cefc0256345880af757353eb8c.jpg) +Figure 9. Example four-class hypnograms during fragmented sleep from the OSV dataset. The use of motion features improves accuracy during short awakenings between 22.00 and 00.00. + +Motion features. Table 4 reports the effectiveness of the two main components of our motion feature set: the set of feature definitions $f_{i}$ and the set of regions used $R$ . Additional ablations can be found in §C. Here we define and report an additional metric $\kappa_{F}$ which is the Cohen's $\kappa$ statistic computed over epochs near transitions to/from Wake (W) i.e. where $\hat{y}_i^t\neq \hat{y}_i^{t + \Delta} = W$ or $\hat{y}_i^t = W\neq \hat{y}_i^{t + \Delta}$ for any $\Delta \in [-2,2]$ epochs. This metric better quantifies accuracy during fragmented sleep, such as in Figure 9. + +As measured by $\kappa_{F}$ , we see that motion features aid classification during fragmented sleep. The ability to accurately measure sleep fragmentation is particularly important since it is linked with daytime sleepiness and decreased cognitive performance [29]. + +Table 4. Ablation study of motion feature parameters. + +
AblationParameter Setnf*Cohen's κ
κTκF
Feature setfi ∈ {}00.6890.397
fi ∈ {f1, f2}360.6970.467
fi ∈ {f3, f4}540.7040.465
fi ∈ {f1, f2, f3, f4}900.7080.491
Region(s)R ∈ {H ∪ B}300.7020.461
R ∈ {H ∪ B, O}600.7040.468
R ∈ {H, B, O}900.7080.491
+ +\*No. features. + +Cardio-respiratory inputs. In Table 5, we compare the performance of the SleepVST model with variants which use heart rate $\boldsymbol{x}_{HR}$ or breathing rate $\boldsymbol{x}_{BR}$ as inputs instead of the underlying waveforms, as described in Section 4.1. We see that the use of waveforms, rather than derived heart and breathing rates, leads to improvements in sleep staging accuracy across all datasets. + +Table 5. Comparison of four-class sleep staging performance for different cardio-respiratory input combinations to SleepVST. + +
Input signalsCohen's κT
SHHSMESAOSV
xHR, xBR0.6830.7220.681
xHW, xBR0.7180.7360.684
xHR, xBW0.7230.7580.702
xHW, xBW0.7490.7650.708
+ +# 6. Conclusions + +We have introduced SleepVST, a transformer model for sleep stage classification from cardio-respiratory waveforms. Using a transfer learning approach, we have shown how SleepVST can be used to perform sleep stage classification entirely from near-infrared video. + +Our results advance the state-of-the-art in video-based sleep staging, narrowing the gap to expert-level performance and taking a significant step towards important clinical applications. More broadly, our work shows that a model pre-trained on a task using cardio-respiratory waveforms from large contact-sensor datasets can be effectively transferred to waveforms measured from video. This approach can benefit wider applications such as the detection of arrhythmias and sleep apnoes, from cardiac and respiratory waveforms respectively. + +Finally, we believe there is important future work in learning more expressive representations of motion during sleep. In addition to improving the performance of video-based sleep staging, this could also enable the detection of sleep movement disorders, such as REM behaviour disorder, removing the existing need for time-consuming, manual video review [51]. + +# Acknowledgements + +This work was supported by the EPSRC Centre for Doctoral Training in Autonomous Intelligent Machines and Systems [EP/S024050/1] and funded by Oxehealth Ltd. Figure 1 was created with BioRender.com. We kindly thank the National Sleep Research Resource for granting access to SHHS and MESA. + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 4 +[2] Ying-Wen Bai, Wen-Tai Li, and Cheng-Hsiang Yeh. Design and implementation of an embedded monitor system for body breath detection by using image processing methods. In 2010 Digest of Technical Papers International Conference on Consumer Electronics (ICCE), pages 193-194, 2010. 2 +[3] Jessie P. Bakker, Marco Ross, Ray Vasko, Andreas Cerny, Pedro Fonseca, Jeff Jasko, Edmund Shaw, David P. White, and Peter Anderer. Estimating sleep stages using cardiorespiratory signals: validation of a novel algorithm across a wide range of sleep-disordered breathing severity. Journal of Clinical Sleep Medicine, 17(7):1343-1354, 2021. 3, 7 +[4] Guha Balakrishnan, Fredo Durand, and John Guttag. Detecting Pulse from Head Motions in Video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, Portland, OR, USA, 2013. IEEE. 2 +[5] Leo Breiman. Random Forests. Machine Learning, 45(1): 5-32, 2001. 5 +[6] Jonathan Carter, João Jorge, Bindia Venugopal, Oliver Gibson, and Lionel Tarassenko. Deep Learning-Enabled Sleep Staging From Vital Signs and Activity Measured Using a Near-Infrared Video Camera. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 5940-5949, Vancouver, BC, Canada, 2023. IEEE. 1, 3, 5, 6, 7 +[7] Weixuan Chen and Daniel McDuff. Deepphys: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 349-365, 2018. 2 +[8] Xiaoli Chen, Rui Wang, Phyllis Zee, Pamela L. Lutsey, Sogol Javaheri, Carmela Alcantara, Chandra L. Jackson, Michelle A. Williams, and Susan Redline. Racial/Ethnic Differences in Sleep Disturbances: The Multi-Ethnic Study of Atherosclerosis (MESA). Sleep, 38(6):877-888, 2015. 3, 7 +[9] Jacob Cohen. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1): 37-46, 1960. 2, 5 +[10] Kris Cuppens, Lieven Lagae, Berten Ceulemans, Sabine Van Huffel, and Bart Vanrumste. Automatic video detection of body movement during sleep based on optical flow in pediatric patients with epilepsy. Medical & Biological Engineering & Computing, 48(9):923-931, 2010. 3 +[11] Heidi Danker-Hopfe, Peter Anderer, Josef Zeitlhofer, Marion Boeck, Hans Dorn, Georg Gruber, Esther Heller, Erna Loretz, Doris Moser, Silvia Parapatics, Bernd Saletu, Andrea Schmidt, and Georg Dorffner. _Interrater reliability for sleep scoring according to the Rechtschaffen & Kales and the new AASM standard_. Journal of Sleep Research, 18(1):74-84, 2009. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1365-2869.2008.00700.x. 2 + +[12] Koldo De Miguel, Alberto Brunete, Miguel Hernando, and Ernesto Gambao. Home Camera-Based Fall Detection System for the Elderly. Sensors, 17(12):2864, 2017. 1 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 4 +[14] T. B. Fitzpatrick. The validity and practicality of sun-reactive skin types I through VI. Archives of Dermatology, 124(6): 869-871, 1988. 6 +[15] Marco Hafner, Martin Stepanek, Jirka Taylor, Wendy M. Troxel, and Christian van Stolk. Why Sleep Matters—The Economic Costs of Insufficient Sleep. *Rand Health Quarterly*, 6(4):11, 2017. 1 +[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 5 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4 +[18] Adrienne Heinrich, Xavier Aubert, and Gerard de Haan. Body movement analysis during sleep based on video motion estimation. In 2013 IEEE 15th International Conference on e-Health Networking, Applications and Services (Healthcom 2013), pages 539–543, 2013. 1 +[19] D. W. Hudgel, R. J. Martin, B. Johnson, and P. Hill. Mechanics of the respiratory system and breathing pattern during sleep in normal humans. Journal of Applied Physiology: Respiratory, Environmental and Exercise Physiology, 56(1): 133-137, 1984. 1, 3 +[20] C. Iber. The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology, and Technical Specification. 2007. 1, 2 +[21] Masamitsu Kamon, Shima Okada, Masafumi Furuta, and Koki Yoshida. Development of a non-contact sleep monitoring system for children. Frontiers in Digital Health, 4, 2022. 3, 7 +[22] Kevin Kotzen, Peter H. Charlton, Sharon Salabi, Lea Amar, Amir Landesberg, and Joachim A. Behar. SleepPPG-Net: A Deep Learning Algorithm for Robust Sleep Staging From Continuous Photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 27(2):924–932, 2023. 1, 2, 6, 7 +[23] Yee Fun Lee, Dmitry Gerashchenko, Igor Timofeev, Brian J. Bacskai, and Ksenia V. Kastanenka. Slow Wave Sleep Is a Promising Intervention Target for Alzheimer's Disease. Frontiers in Neuroscience, 14, 2020. 1 +[24] Yun Ji Lee, Jae Yong Lee, Jae Hoon Cho, and Ji Ho Choi. Interrater reliability of sleep stage scoring: a meta-analysis. Journal of Clinical Sleep Medicine, 18(1):193-202, 2022. 2 +[25] Michael H. Li, Azadeh Yadollahi, and Babak Taati. A noncontact vision-based system for respiratory rate estimation. In 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pages 2119-2122, 2014. 2 + +[26] Xin Liu, Brian Hill, Ziheng Jiang, Shwetak Patel, and Daniel McDuff. EfficientPhys: Enabling Simple, Fast and Accurate Camera-Based Cardiac Measurement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5008-5017, 2023. 2 +[27] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, 2019. 4 +[28] Yawei Luo, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Taking a Closer Look at Domain Shift: Category-Level Adversaries for Semantics Consistent Domain Adaptation. pages 2507-2516, 2019. 3 +[29] S E Martin, H M Engleman, I J Deary, and N J Douglas. The effect of sleep fragmentation on daytime function. American Journal of Respiratory and Critical Care Medicine, 153(4): 1328-1332, 1996. Publisher: American Thoracic Society - AJRCCM. 8 +[30] Daniel McDuff. Camera Measurement of Physiological Vital Signs. ACM Computing Surveys, 55(9):1-40, 2023. 2 +[31] Sara Mahvash Mohammadi, Shirin Enshaeifar, Adrian Hilton, Derk-Jan Dijk, and Kevin Wells. Transfer Learning for Clinical Sleep Pose Detection Using a Single 2D IR Camera. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 29:290-299, 2021. 5 +[32] K. Nakajima, A. Osa, and H. Miike. A method for measuring respiration and physical activity in bed by optical flow analysis. In Proceedings of the 19th Annual International Conference of the IEEE Engineering in Medicine and Biology Society., pages 2054-2057 vol.5, 1997. 2 +[33] Kazuki Nakajima, Yoshiaki Matsumoto, and Toshiyo Tamura. Development of real-time image sequence analysis for evaluating posture change and respiratory rate of a subject in bed. Physiological Measurement, 22(3):N21-N28, 2001. 2, 3 +[34] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations, 2022. 3, 4 +[35] Teruaki Nochino, Yuko Ohno, Takafumi Kato, Masako Tani-ike, and Shima Okada. Sleep stage estimation method using a camera for home use. Biomedical Engineering Letters, 9 (2):257-265, 2019. 1, 3, 4, 7 +[36] Jiapu Pan and Willis J. Tompkins. A Real-Time QRS Detection Algorithm. IEEE Transactions on Biomedical Engineering, BME-32(3):230-236, 1985. 3 +[37] James Pardey, Stephen Roberts, Lionel Tarassenko, and John Stradling. A new approach to the analysis of the human sleep/wakefulness continuum. Journal of sleep research, 5 (4):201-210, 1996. 1 +[38] Aakash K. Patel, Vamsi Reddy, and John F. Araujo. Physiology, Sleep Stages. StatPearls Publishing, 2022. 4 +[39] Huy Phan and Kaare Mikkelsen. Automatic sleep staging of EEG signals: recent development, challenges, and future directions. Physiological Measurement, 43(4):04TR01, 2022. 2 +[40] Huy Phan, Oliver Y. Chén, Minh C. Tran, Philipp Koch, Alfred Mertins, and Maarten De Vos. XSleepNet: Multi-View + +Sequential Model for Automatic Sleep Staging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 5903-5915, 2022. 1 +[41] Huy Phan, Kaare Mikkelsen, Oliver Y. Chen, Philipp Koch, Alfred Mertins, and Maarten De Vos. SleepTransformer: Automatic Sleep Staging With Interpretability and Uncertainty Quantification. IEEE Transactions on Biomedical Engineering, 69(8):2456-2467, 2022. 4 +[42] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Advancements in Noncontact, Multiparameter Physiological Measurements Using a Webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. Conference Name: IEEE Transactions on Biomedical Engineering. 2 +[43] S. F. Quan, B. V. Howard, C. Iber, J. P. Kiley, F. J. Nieto, G. T. O'Connor, D. M. Rapoport, S. Redline, J. Robbins, J. M. Samet, and P. W. Wahl. The Sleep Heart Health Study: design, rationale, and methods. Sleep, 20(12):1077-1085, 1997. 3, 7 +[44] Mustafa Radha, Pedro Fonseca, Arnaud Moreau, Marco Ross, Andreas Cerny, Peter Anderer, Xi Long, and Ronald M. Aarts. A deep transfer learning approach for wearable sleep stage classification with photoplethysmography. npj Digital Medicine, 4(1):1-11, 2021. 2 +[45] S.J. Redmond and C. Heneghan. Cardiorespiratory-based sleep staging in subjects with obstructive sleep apnea. IEEE Transactions on Biomedical Engineering, 53(3):485-496, 2006. 1 +[46] A. J. Schwichtenberg, Jeehyun Choe, Ashleigh Kellerman, Emily A. Abel, and Edward J. Delp. Pediatric Videosomnography: Can Signal/Video Processing Distinguish Sleep and Wake States? Frontiers in Pediatrics, 6:158, 2018. 3 +[47] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90(2):227-244, 2000. 3 +[48] Z. Shinar, A. Baharav, Y. Dagan, and S. Akselrod. Automatic detection of slow-wave-sleep using heart rate variability. In Computers in Cardiology 2001. Vol.28, pages 593-596, 2001. 1, 3 +[49] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 2 +[50] Niranjan Sridhar, Ali Shoeb, Philip Stephens, Alaa Kharbouch, David Ben Shimol, Joshua Burkart, Atiyeh Ghoreyshi, and Lance Myers. Deep learning for automated sleep staging using instantaneous heart rate. npj Digital Medicine, 3(1):1-10, 2020. 2, 7 +[51] Ambra Stefani and Birgit Hogl. Sleep in Parkinson's disease. Neuropsychopharmacology, 45(1):121-128, 2020. 1, 8 +[52] Ambra Stefani, David Gabelaia, Thomas Mitterling, Werner Poewe, Birgit Hogl, and Birgit Frauscher. A Prospective Video-Polysomnographic Analysis of Movements during Physiological Sleep in 100 Healthy Sleepers. Sleep, 38 (9):1479-1487, 2015. 1 +[53] Haoqi Sun, Wolfgang Ganglberger, Ezhil Panneerselvam, Michael J Leone, Syed A Quadri, Balaji Goparaju, Ryan A + +Tesh, Oluwaseun Akeju, Robert J Thomas, and M Brandon Westover. Sleep staging from electrocardiography and respiration with deep learning. *Sleep*, 43(7):zsz306, 2020. 6 +[54] K. S. Tan, R. Saatchi, H. Elphick, and D. Burke. Real-time vision based respiration monitoring system. In 2010 7th International Symposium on Communication Systems, Networks & Digital Signal Processing (CSNDSP 2010), pages 770-774, 2010. 2 +[55] Alexander Trumpp, Johannes Lohr, Daniel Wedekind, Martin Schmidt, Matthias Burghardt, Axel R. Heller, Hagen Malberg, and Sebastian Zaunseder. Camera-based photoplethysmography in an intraoperative setting. *BioMedical Engineering OnLine*, 17(1):33, 2018. 2 +[56] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Oxhealth Vital Signs, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200019.pdf.3 +[57] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Gili Pro Biosensor, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200038.pdf.3 +[58] Mark van Gastel, Sander Stuijk, Sebastiaan Overeem, Johannes P. van Dijk, Merel M. van Gilst, and Gerard de Haan. Camera-Based Vital Signs Monitoring During Sleep - A Proof of Concept Study. IEEE Journal of Biomedical and Health Informatics, 25(5):1409-1418, 2021. 3 +[59] Hans van Gorp, Iris A M Huijben, Pedro Fonseca, Ruud J G van Sloun, Sebastiaan Overeem, and Merel M van Gilst. Certainty about uncertainty in sleep staging: a theoretical framework. Sleep, 45(8), 2022. 2 +[60] Fokke B. van Meulen, Angela Grassi, Leonie van den Heuvel, Sebastiaan Overeem, Merel M. van Gilst, Johannes P. van Dijk, Henning Maass, Mark J. H. van Gastel, and Pedro Fonseca. Contactless Camera-Based Sleep Staging: The HealthBed Study. Bioengineering, 10(1):109, 2023. 1, 3, 7 +[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. arXiv:1706.03762 [cs], 2017. arXiv: 1706.03762. 4 +[62] Mauricio Villarroel, Sitthichok Chaichulee, João Jorge, Sara Davis, Gabrielle Green, Carlos Arteta, Andrew Zisserman, Kenny McCormick, Peter Watkinson, and Lionel Tarassenko. Non-contact physiological monitoring of preterm infants in the Neonatal Intensive Care Unit. npj Digital Medicine, 2(1):1-18, 2019. 2 +[63] Olivia Walch, Yitong Huang, Daniel Forger, and Cathy Goldstein. Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device. *Sleep*, 42(12), 2019. 1 +[64] Qiongyan Wang, Hanrong Cheng, and Wenjin Wang. Feasibility of Exploiting Physiological and Motion Features for Camera-based Sleep Staging: A Clinical Study. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5, 2023. ISSN: 2694-0604. 3 + +[65] Wenjin Wang and Albertus C. den Brinker. Algorithmic insights of camera-based respiratory motion extraction. Physiological measurement, 43(7):075004, 2022. 1, 2 +[66] Wenjin Wang, Albertus C. den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic Principles of Remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 1, 2 +[67] William H. Ward, Fernando Lambreton, Neha Goel, Jian Q. Yu, and Jeffrey M. Farma. Clinical Presentation and Staging of Melanoma. In Cutaneous Melanoma: Etiology and Therapy. Codon Publications, Brisbane (AU), 2017. 6 +[68] Johanna Wilde-Frenz and Hartmut Schulz. Rate and Distribution of Body Movements during Sleep in Humans. Perceptual and Motor Skills, 56(1):275-283, 1983. 1, 5 +[69] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. SimPer: Simple Self-Supervised Learning of Periodic Targets, 2023. 2 +[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. PhysFormer: Facial Video-based Physiological Measurement with Temporal Difference Transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4186, New Orleans, LA, USA, 2022. IEEE. 2 +[71] Guo-Qiang Zhang, Licong Cui, Remo Mueller, Shiqiang Tao, Matthew Kim, Michael Rueschman, Sara Mariani, Daniel Mobley, and Susan Redline. The National Sleep Research Resource: towards a sleep data commons. Journal of the American Medical Informatics Association: JAMIA, 25 (10):1351-1358, 2018. 3 +[72] Dan Ziegler, G. Laux, K. Dannehl, M. Spüler, H. Mühlen, P. Mayer, and F.a. Gries. Assessment of Cardiovascular Autonomic Function: Age-related Normal Ranges and Reproducibility of Spectral Analysis, Vector Analysis, and Standard Tests of Heart Rate Variation and Blood Pressure Responses. Diabetic Medicine, 9(2):166-175, 1992. 6 \ No newline at end of file diff --git a/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/images.zip b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16f955bc4f8805152d7bbebd0694a57f87ac75fa --- /dev/null +++ b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a12df211851774301091db9df9a04d87c45cb2360597e46a800f1b4439e0e19e +size 506510 diff --git a/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/layout.json b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2071944a6631de140399fc8a14452ac025f4a2db --- /dev/null +++ b/2024/SleepVST_ Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers/layout.json @@ -0,0 +1,9741 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 102, + 479, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 102, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 479, + 138 + ], + "type": "text", + "content": "SleepVST: Sleep Staging from Near-Infrared Video Signals using Pre-Trained Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "spans": [ + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "text", + "content": "Jonathan F. Carter" + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "text", + "content": " João Jorge" + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "text", + "content": " Oliver Gibson" + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "text", + "content": " Lionel Tarassenko" + }, + { + "bbox": [ + 127, + 170, + 465, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 194, + 441, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 194, + 441, + 223 + ], + "spans": [ + { + "bbox": [ + 154, + 194, + 441, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 194, + 441, + 223 + ], + "type": "text", + "content": "Institute of Biomedical Engineering, University of Oxford " + }, + { + "bbox": [ + 154, + 194, + 441, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 194, + 441, + 223 + ], + "type": "text", + "content": "Oxehealth Ltd., Oxford" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 284, + 289, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 524 + ], + "type": "text", + "content": "Advances in camera-based physiological monitoring have enabled the robust, non-contact measurement of respiration and the cardiac pulse, which are known to be indicative of the sleep stage. This has led to research into camera-based sleep monitoring as a promising alternative to \"gold-standard\" polysomnography, which is cumbersome, expensive to administer, and hence unsuitable for longer-term clinical studies. In this paper, we introduce SleepVST, a transformer model which enables state-of-the-art performance in camera-based sleep stage classification (sleep staging). After pre-training on contact sensor data, SleepVST outperforms existing methods for cardiorespiratory sleep staging on the SHHS and MESA datasets, achieving total Cohen's kappa scores of 0.75 and 0.77 respectively. We then show that SleepVST can be successfully transferred to cardio-respiratory waveforms extracted from video, enabling fully contact-free sleep staging. Using a video dataset of 50 nights, we achieve a total accuracy of " + }, + { + "bbox": [ + 46, + 284, + 289, + 524 + ], + "type": "inline_equation", + "content": "78.8\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 524 + ], + "type": "text", + "content": " and a Cohen's " + }, + { + "bbox": [ + 46, + 284, + 289, + 524 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 284, + 289, + 524 + ], + "type": "text", + "content": " of 0.71 in four-class video-based sleep staging, setting a new state-of-the-art in the domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": "Accurate sleep monitoring is critical for the diagnosis of sleep disorders and the discovery of novel treatments and biomarkers. Poor sleep is broadly linked with a number of health conditions, including cardiovascular diseases such as diabetes and hypertension [15]. Additionally, there are links between specific sleep stages and neurodegenerative conditions, such as a decrease in non-rapid-eye-movement stage 3 (N3) sleep with Alzheimer's disease [23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 667, + 287, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 667, + 287, + 692 + ], + "spans": [ + { + "bbox": [ + 47, + 667, + 287, + 692 + ], + "type": "text", + "content": "Overnight video polysomnography (vPSG), the \"gold-standard\" [52] technique for sleep monitoring, requires the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 259, + 547, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 356 + ], + "type": "text", + "content": "use of a large number of contact sensors including electrodes placed on the scalp (EEG), near the eyes (EOG), and under the chin (EMG) of the patient. Human experts (polysomnographers) must then review the recorded data, classifying the patient's sleep into five stages at 30-second intervals (epochs) and annotating other important events such as leg movements and apnoeas. This manual process often takes multiple hours to complete." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 358, + 547, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 547, + 455 + ], + "type": "text", + "content": "Experts primarily rely on characteristic patterns in brain activity measured from the EEG to classify sleep stages [20]. However, sleep stage information is also encoded in measures of autonomic activity, including cardiac [48] and respiratory signals [19], and body movements [68]. This has led to the development of methods for automatically classifying sleep stages from sensors that measure these parameters, such as smartwatches [63]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 458, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 615 + ], + "type": "text", + "content": "Prior work has shown that these physiological parameters can also be measured using video cameras [18, 65, 66], leading to the development of methods for classifying sleep stages entirely from video input [6, 35, 60]. Camera-based methods have particular promise as part of a multipurpose sleep monitoring system. For example, within elderly care settings, in addition to classifying sleep stages, they could also be used to detect specific sleep movement disorders [18], which are linked with conditions such as Parkinson's, and which are typically distinguished via manual video review [51]. They could also be used to detect falls [12], one of the leading causes of injury and mortality amongst elderly individuals." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "State-of-the-art video-based sleep staging [6] has used the heart rate (HR) and breathing rate (BR) derived from video to perform sleep staging. However, these rates are typically measured as averages over time, losing higher frequency inter-pulse and inter-breath interval information. EEG- [40] and wearable-based [22] sleep staging methods have both been improved by using raw waveforms as inputs, rather than derived quantities e.g. signal features [37, 45]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 175, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 175, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 175, + 712 + ], + "type": "text", + "content": "*jcarter@robots.ox.ac.uk" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12479" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 534, + 194 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 534, + 194 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 534, + 194 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 534, + 194 + ], + "type": "image", + "image_path": "a915d98d330c93a316f9c2cff731f64bbeacff570a48d1addd7b205a04b1e161.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "text", + "content": "Figure 1. Sleep staging from near-infrared video signals using SleepVST. (a) We first pre-train the model end-to-end on cardiac (heart) and respiratory (breathing) waveforms " + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "inline_equation", + "content": "x_{HW}(t)" + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "inline_equation", + "content": "x_{BW}(t)" + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "text", + "content": " derived from the electrocardiogram (ECG) and a thoracic respiratory belt (THX) respectively. (b) After pre-training, we use the model as a frozen feature extractor, applying it to cardio-respiratory waveforms derived from near-infrared (NIR) video to generate sequences of features. When transferring to video data, we additionally use a set of motion features " + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "inline_equation", + "content": "f_{i}(t)" + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "text", + "content": " derived from an optical flow field " + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{u}(x,y,t)" + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "text", + "content": " as inputs to the classifier C. This approach allows us to utilise much larger contact-sensor datasets to train SleepVST, whilst also enabling the incorporation of motion information when transferring to video data. (c) We evaluate SleepVST using overnight video polysomnography (vPSG) studies, comparing expert-labelled sleep stage sequences " + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\hat{y}_{1:T}" + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "text", + "content": " against those generated entirely from near-infrared video " + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "inline_equation", + "content": "y_{1:T}" + }, + { + "bbox": [ + 46, + 205, + 545, + 293 + ], + "type": "text", + "content": " using our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "text", + "content": "In this paper, we introduce SleepVST1, a transformer model for sleep stage classification from cardio-respiratory waveforms. First, we pre-train the model on waveforms derived from contact sensors (Fig. 1a), after which it outperforms existing methods for cardio-respiratory sleep staging on two widely used public datasets, highlighting the effectiveness of our transformer architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 286, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 286, + 530 + ], + "type": "text", + "content": "After pre-training, we then show that SleepVST can be transferred to cardiac and respiratory waveforms measured entirely from near-infrared video, where our method additionally enables us to incorporate motion features derived from the camera data (Fig. 1b). Using this novel combination of inputs available from a near-infrared camera, we obtain state-of-the-art results for video-based sleep staging. Through a series of ablation studies we demonstrate the effectiveness of our approach. Our work further highlights the potential of camera-based sleep monitoring for important clinical applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 542, + 221, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 221, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 221, + 555 + ], + "type": "text", + "content": "2. Background and Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "text", + "content": "Automated sleep staging. To reduce the manual effort of sleep assessment, deep learning methods have been developed for the automatic classification of sleep stages from polysomnography (PSG) recordings, achieving near-human accuracy [39]. Accuracy is commonly measured using Cohen's kappa statistic " + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "inline_equation", + "content": "(\\kappa, [9])" + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "text", + "content": ", which measures inter-rater agreement between two sources. Expert annotation of sleep stages is a subjective process, making " + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "text", + "content": " agreement unachievable [59]. Prior work has shown that sleep scorers agree around " + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "text", + "content": " of the time (" + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "inline_equation", + "content": "\\kappa = 0.71 - 0.81" + }, + { + "bbox": [ + 46, + 562, + 286, + 693 + ], + "type": "text", + "content": ", [11, 24]) when classifying sleep stages according to American Academy of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 309, + 315, + 544, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 315, + 544, + 398 + ], + "spans": [ + { + "bbox": [ + 309, + 315, + 544, + 398 + ], + "type": "text", + "content": "Sleep Medicine (AASM, [20]) scoring rules. These rules divide sleep into five classes: Wake, N1 (light), N2 (intermediate), N3 (deep) and rapid-eye-movement (REM) sleep. However, due to low inter-scorer agreement in identifying N1 [11], prior work has often treated sleep classification as a four-class problem, by merging N1 and N2 into a single class [22, 50]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 399, + 544, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 544, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 544, + 459 + ], + "type": "text", + "content": "Given the drawbacks of conventional polysomnography, prior work has investigated the potential of automatic sleep staging from a reduced number of contact sensors, including the electrocardiogram ([50], " + }, + { + "bbox": [ + 304, + 399, + 544, + 459 + ], + "type": "inline_equation", + "content": "\\kappa = 0.66" + }, + { + "bbox": [ + 304, + 399, + 544, + 459 + ], + "type": "text", + "content": "), and the photoplethysmogram ([44], " + }, + { + "bbox": [ + 304, + 399, + 544, + 459 + ], + "type": "inline_equation", + "content": "\\kappa = 0.65" + }, + { + "bbox": [ + 304, + 399, + 544, + 459 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 460, + 544, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 544, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 544, + 544 + ], + "type": "text", + "content": "Camera-based physiological measurement. Respiratory measurement is possible from video by measuring the induced movements of the chest and abdomen [65]. This has been achieved through methods such as frame-differencing [2, 54], using dense optical flow magnitudes [32, 33] and by tracking the movement of sparse optical flow vectors [25] across video frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 544, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 544, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 544, + 688 + ], + "type": "text", + "content": "Camera-based measurement of the pulse is possible using methods which measure specific cardiac phenomena [30]. These include remote photoplethysmography (rPPG, [42, 66]), which measures skin colour changes caused by variations in blood volume over the cardiac cycle, and ballistocardiography [4], which measures head movements induced as blood is pumped through the carotid arteries. More recently, deep learning methods have been developed which learn to extract a pulse signal directly from input video frames [7, 26, 70] rather than targeting specific underlying cardiac phenomena, including unsupervised approaches [49, 69]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "content": "Camera-based vital-sign measurement has been validated within clinical settings [55, 62], during overnight" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 168, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 168, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 168, + 712 + ], + "type": "text", + "content": "Sleep Video-Signal Transformer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12480" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "polysomnography recordings [58, 64], and has led to products with regulatory clearance for camera-based measurement of the heart and breathing rate, including the Gili Pro Biosensor [57] and Oxehealth Vital Signs software [56]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 289 + ], + "type": "text", + "content": "Camera-based sleep monitoring. Prior work has investigated the binary problem of distinguishing sleep from wake using video cameras, using measures of activity derived from the input frames [10, 33, 46]. Whilst activity information alone can be effective in distinguishing sleep from wake, cardio-respiratory signals carry important additional information for classifying specific sleep stages [19, 48]. The current state-of-the-art approach for video-based sleep staging ([6], " + }, + { + "bbox": [ + 46, + 121, + 289, + 289 + ], + "type": "inline_equation", + "content": "\\kappa = 0.61" + }, + { + "bbox": [ + 46, + 121, + 289, + 289 + ], + "type": "text", + "content": ") combines multiple modalities derived from video: the heart rate, breathing rate and activity information, to perform sleep stage classification. This approach outperforms prior methods which had used only a single video-derived modality: motion [21, 35] or the cardiac pulse [60]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 300, + 203, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 203, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 203, + 314 + ], + "type": "text", + "content": "3. Datasets and Preprocessing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 319, + 167, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 167, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 167, + 331 + ], + "type": "text", + "content": "3.1. Contact Sensor Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 338, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 434 + ], + "type": "text", + "content": "To train the SleepVST model, we used the SHHS [43, 71] and MESA [8] contact sensor datasets. These datasets contain PSG recordings and sleep stages annotated according to AASM scoring rules. The cardio-respiratory signals available in each recording include the electrocardiogram (ECG), which measures electrical activity of the heart, and respiratory signals, measured using respiratory bands placed around the abdomen and thorax." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 529 + ], + "type": "text", + "content": "From the available signals, we seek to produce generic cardiac and respiratory waveforms to train the SleepVST model, preventing the model from learning to use information which may be present in signals derived from contact sensors but not present in those derived from video. This is to mitigate the problem of domain shift [28, 47], where the performance of a model is significantly reduced when applied to a different data distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "content": "To produce a cardiac (heart) waveform, we first applied the same filtering steps used by the Pan-Tompkins algorithm [36] to the ECG signal. We resampled the output from the moving-window integration step of the algorithm to " + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "content": ", then applied a Butterworth bandpass filter with a passband of " + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "inline_equation", + "content": "0.66 - 2.8\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "content": " (40-168 beats per minute or BPM) and Gaussian smoothing. To produce a respiratory (breathing) waveform, we used the thoracic respiratory signal as the input source, as previously used in [3]. This was downsampled to " + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "content": " and smoothed using a median filter with a kernel of length 5, to remove artefacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 122, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 122, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 122, + 682 + ], + "type": "text", + "content": "3.2. Video Data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "We used the dataset previously introduced in [6] for our video-based sleep staging experiments. This dataset con" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "sists of overnight vPSG recordings from 50 volunteers, with sleep stages annotated according to the AASM guidelines at 30-second intervals (epochs) by a sleep physiologist. Throughout this work, we refer to this as the Oxford Sleep Volunteers (OSV) dataset. Each recording includes " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "850\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " near-infrared (NIR) video data captured at 20 FPS. An example camera view can be seen in Figure 5. More information on the dataset is provided in " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\S B" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 168, + 546, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 275 + ], + "type": "text", + "content": "To derive cardiac and respiratory waveforms from the video data, we used components of FDA-cleared Oxhealth Vital Signs software [56]. For each overnight video recording, we used the software to obtain cardiac and respiratory waveforms sampled at " + }, + { + "bbox": [ + 304, + 168, + 546, + 275 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 168, + 546, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 168, + 546, + 275 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 168, + 546, + 275 + ], + "type": "text", + "content": " respectively. We then applied the same filtering steps applied to the contact sensor waveforms: Butterworth filtering and Gaussian smoothing to the heart waveforms, and median filtering to the breathing waveforms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 276, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 547, + 334 + ], + "type": "text", + "content": "Figure 2 shows examples of processed cardio-respiratory waveforms measured from contact sensors and video for a short section of a recording from the OSV dataset. These have additionally been normalised to zero mean and unit variance." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 347, + 541, + 460 + ], + "blocks": [ + { + "bbox": [ + 310, + 347, + 541, + 460 + ], + "lines": [ + { + "bbox": [ + 310, + 347, + 541, + 460 + ], + "spans": [ + { + "bbox": [ + 310, + 347, + 541, + 460 + ], + "type": "image", + "image_path": "13a2e06f84183abdde81392f75867ab555309f312b8bc93e075e587686f53b08.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "text", + "content": "Figure 2. Example (normalized) cardiac and respiratory waveforms, " + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "inline_equation", + "content": "x_{HW}(t)" + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "inline_equation", + "content": "x_{BW}(t)" + }, + { + "bbox": [ + 304, + 472, + 545, + 506 + ], + "type": "text", + "content": ", derived from contact sensors (blue) and video (orange) from the OSV dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 532, + 366, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 366, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 366, + 544 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 551, + 437, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 437, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 437, + 563 + ], + "type": "text", + "content": "4.1. SleepVST Architecture" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "type": "text", + "content": "Figure 3 illustrates the SleepVST model architecture, which turns cardiac and respiratory waveforms into sequences of feature vectors using a transformer encoder. These feature vectors are then used to classify each 30-second sleep epoch. During pre-training, we used a linear layer for classification. When transferring to video data, we trained a new classifier that additionally used motion features derived from video as inputs, discussed further in Section 4.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Patchification. The continuous heart and breathing waveforms (HW and BW) are divided into N windows i.e. patches, drawing on ideas from PatchTST [34]. We used non-overlapping 30-second windows since this is the inter" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12481" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 272, + 238 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 272, + 238 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 272, + 238 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 272, + 238 + ], + "type": "image", + "image_path": "8d940f7e4bb5edf9b1e04902ea003124a51496629d71b795a736b64fcad25298.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "type": "text", + "content": "Figure 3. SleepVST architecture. Each 30-second window of heart " + }, + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "type": "inline_equation", + "content": "(x_{HW})" + }, + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "type": "text", + "content": " and breathing waveforms " + }, + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "type": "inline_equation", + "content": "(x_{BW})" + }, + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "type": "text", + "content": " is passed to a patch encoder, which turns them into patch-level features. These features are concatenated and passed to a transformer encoder. During pre-training, a linear layer turns the output feature sequence " + }, + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "type": "inline_equation", + "content": "z_{o}" + }, + { + "bbox": [ + 46, + 249, + 288, + 316 + ], + "type": "text", + "content": " of length N from SleepVST into sleep stage classifications." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "text", + "content": "val at which sleep stages are classified according to AASM guidelines. We denote the patchified inputs as " + }, + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{HW} \\in \\mathbb{R}^{300 \\times N}" + }, + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{BW} \\in \\mathbb{R}^{150 \\times N}" + }, + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 338, + 287, + 387 + ], + "type": "text", + "content": " is the sequence length." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "text", + "content": "Encodings. Each waveform patch is normalised to zero mean and unit variance before being passed to encoding layers " + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "E_{HW}" + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "E_{BW}" + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "text", + "content": ", which transform them into feature vectors " + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{HW} \\in \\mathbb{R}^{D_{HW} \\times N}" + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{BW} \\in \\mathbb{R}^{D_{BW} \\times N}" + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "text", + "content": ". The feature vectors from each patch are then concatenated, producing a sequence of feature vectors " + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{z}_i \\in \\mathbb{R}^{(D_{HW} + D_{BW}) \\times N}" + }, + { + "bbox": [ + 47, + 388, + 287, + 471 + ], + "type": "text", + "content": " which are passed to a transformer encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "content": "We used identical, shallow 1D ResNet-style [17] models for " + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "inline_equation", + "content": "E_{HW}" + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "inline_equation", + "content": "E_{BW}" + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "content": ", illustrated in Figure 4. This convolutional encoder design [1] compresses the waveform patches into lower-dimensional feature vectors. This allowed us to keep " + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "inline_equation", + "content": "D_{HW}" + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "inline_equation", + "content": "D_{BW}" + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "content": " small, reducing the computational and memory complexity of the downstream transformer and the likelihood of over-fitting, particularly when transferring the model to the smaller video dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "This encoder design additionally provides a high degree of flexibility at the input level, such as the ability to use input signals of differing frequencies. In later ablation studies, this enables us to compare the performance of the SleepVST model with simpler variants which use 30-second patches of " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " heart rate " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{HR} \\in \\mathbb{R}^{30 \\times N}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " and/or breathing rate " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{BR} \\in \\mathbb{R}^{30 \\times N}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " time-series as inputs, rather than underlying waveforms. When using either of these as inputs, we used linear projections [13, 34] for the input encoders, because of the lower dimensionality of " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{HR}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{BR}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": ". Throughout our experiments we use the same feature dimension i.e. " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{HW} = D_{BW} = D_{HR} = D_{BR} = 64" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 77, + 535, + 220 + ], + "blocks": [ + { + "bbox": [ + 318, + 77, + 535, + 220 + ], + "lines": [ + { + "bbox": [ + 318, + 77, + 535, + 220 + ], + "spans": [ + { + "bbox": [ + 318, + 77, + 535, + 220 + ], + "type": "image", + "image_path": "cd35a6b7e6e1cf582dc432a7f0dc609018fcec645f9926444d7cf0e0429209a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 238, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 270 + ], + "type": "text", + "content": "Figure 4. Waveform encoder design. Using a series of convolutional layers, the encoder turns sequences of signal patches into sequences of lower-dimensional feature vectors." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 292, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 435 + ], + "type": "text", + "content": "Transformer layer. We used a Transformer encoder [61] with sinusoidal position encodings to turn the sequence of input features " + }, + { + "bbox": [ + 304, + 292, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 292, + 545, + 435 + ], + "type": "text", + "content": " into an output sequence of features " + }, + { + "bbox": [ + 304, + 292, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_o" + }, + { + "bbox": [ + 304, + 292, + 545, + 435 + ], + "type": "text", + "content": " with an identical feature dimension. Normal sleep exhibits long-range structures, known as sleep cycles [38], which typically last around 90-120 minutes and typically occur four to five times per night. From its superior performance in many sequence modelling tasks, including EEG-based sleep staging [41], we expect that the inductive biases of the transformer model (i.e. the attention mechanism) can enable it to learn to use relative position within a sleep cycle to classify sleep stages." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "text", + "content": "Classification. When training and testing on contact sensor data, we used a linear classification layer " + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_C \\in \\mathbb{R}^{C \\times (D_{HW} + D_{BW})}" + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "text", + "content": " to transform the sequence of output feature vectors " + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_o" + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "text", + "content": " into a sequence of sleep stage probabilities " + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{C \\times N}" + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "text", + "content": ", where the number of classes " + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "text", + "content": " depends on the sleep classification strategy used. In line with prior work, we classified sleep stages into four classes during pretraining: Wake-N1/N2-N3-REM, i.e. " + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "inline_equation", + "content": "C = 4" + }, + { + "bbox": [ + 304, + 436, + 545, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": "Model training and inference. In all experiments, we used a sequence length of " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "N = 240" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " windows i.e. two hours. During training, we created batches by sampling sequences of length " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " with a step of 5 minutes from within each recording. The model was trained to minimise the multiclass cross-entropy loss using the AdamW [27] optimiser with a learning rate of 0.0005 and weight decay of 0.001. Other hyper-parameters, such as the no. encoder layers, are detailed in §D. To apply the model to arbitrary length input sequences during evaluation, we re-applied it at regular intervals. More details on this process are provided in §E." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 458, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 458, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 458, + 682 + ], + "type": "text", + "content": "4.2. Motion features from Video" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Following prior work [35], we seek to produce motion features from video which encode for information which is pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12482" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "dictive of the sleep stage, such as the rate and distribution of body movements [68]. To do so, we define a set of parameterised time-series features that collectively quantify motion over a range of spatial and temporal scales." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 286, + 192 + ], + "type": "text", + "content": "Optical flow estimation. We first estimated a 2D optical flow field " + }, + { + "bbox": [ + 46, + 120, + 286, + 192 + ], + "type": "inline_equation", + "content": "\\pmb{u}(t,x,y) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 120, + 286, + 192 + ], + "type": "text", + "content": " over the bed region. Inspired by sleep pose detection work [31], we applied a homography transformation [16] to the individual video frames before estimating the flow field, to achieve camera viewpoint invariance (shown in Figure 5)." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 54, + 202, + 145, + 287 + ], + "blocks": [ + { + "bbox": [ + 54, + 202, + 145, + 287 + ], + "lines": [ + { + "bbox": [ + 54, + 202, + 145, + 287 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 145, + 287 + ], + "type": "image", + "image_path": "d62adc07c5c05ad174f74c1c78cd6c8a43f496c74b8ac3715d3a5d004de9300b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 292, + 106, + 301 + ], + "lines": [ + { + "bbox": [ + 94, + 292, + 106, + 301 + ], + "spans": [ + { + "bbox": [ + 94, + 292, + 106, + 301 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 202, + 213, + 288 + ], + "blocks": [ + { + "bbox": [ + 151, + 202, + 213, + 288 + ], + "lines": [ + { + "bbox": [ + 151, + 202, + 213, + 288 + ], + "spans": [ + { + "bbox": [ + 151, + 202, + 213, + 288 + ], + "type": "image", + "image_path": "5f43ab59badbe965d2739e0e1735184249d2d76bb15410cb12d4ee8051b20c94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 292, + 188, + 301 + ], + "lines": [ + { + "bbox": [ + 176, + 292, + 188, + 301 + ], + "spans": [ + { + "bbox": [ + 176, + 292, + 188, + 301 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 202, + 281, + 288 + ], + "blocks": [ + { + "bbox": [ + 219, + 202, + 281, + 288 + ], + "lines": [ + { + "bbox": [ + 219, + 202, + 281, + 288 + ], + "spans": [ + { + "bbox": [ + 219, + 202, + 281, + 288 + ], + "type": "image", + "image_path": "35d473d2c27d3b528714d70946bd483a1a99b46eea75f67d05e4d225ce19bb26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 292, + 256, + 301 + ], + "lines": [ + { + "bbox": [ + 244, + 292, + 256, + 301 + ], + "spans": [ + { + "bbox": [ + 244, + 292, + 256, + 301 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 314, + 287, + 358 + ], + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 358 + ], + "type": "text", + "content": "Figure 5. Example processing of an NIR video frame from the OSV dataset. (a) Real viewpoint. (b) Virtual viewpoint. (c) Head (H), body (B), and outer (O) bed regions. The distance from the camera to the head region is around " + }, + { + "bbox": [ + 46, + 314, + 287, + 358 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 314, + 287, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": "Feature definitions. Given an arbitrary region " + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": " of the optical flow field, we define the following parametrised signals:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 98, + 417, + 287, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 417, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 98, + 417, + 287, + 436 + ], + "type": "interline_equation", + "content": "v (t; R) = \\max _ {(x, y) \\in R} | \\boldsymbol {u} (t, x, y) | \\tag {1}", + "image_path": "1b4b948f90923f78b5ded1fd5601e0d8c23eb35ea81125fc41c09725a25da4be.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 438, + 287, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 438, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 99, + 438, + 287, + 468 + ], + "type": "interline_equation", + "content": "s (t; R) = \\frac {1}{| R |} \\sum_ {(x, y) \\in R} | \\boldsymbol {u} (t, x, y) | \\tag {2}", + "image_path": "99cde45b40e2623247a919b023a36ca081e6ef656462973186547578494506f0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "text", + "content": "i.e. the maximum (1) and the average (2) of the optical flow magnitude in the region R at time " + }, + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "text", + "content": ". Using these signals, we define two time-series features that quantify motion within a time window " + }, + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "text", + "content": " up to time " + }, + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 478, + 287, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 536, + 287, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 536, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 107, + 536, + 287, + 567 + ], + "type": "interline_equation", + "content": "f _ {1} (t; R, \\Delta) = \\sum_ {t - \\Delta} ^ {t} v (t; R) \\tag {3}", + "image_path": "c9e2f7dcdb7c229463f4ca0dd9929ad95ab8c9b502ef68c2eee37219934040e9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 108, + 571, + 287, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 571, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 108, + 571, + 287, + 603 + ], + "type": "interline_equation", + "content": "f _ {2} (t; R, \\Delta) = \\sum_ {t - \\Delta} ^ {t} s (t; R) \\tag {4}", + "image_path": "8d43c7a13aa093c0e904d5f7f078947a702957878ca5acbb242835d465989bf4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 613, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 636 + ], + "type": "text", + "content": "Plus two additional features which measure the time elapsed since either signal went above a threshold " + }, + { + "bbox": [ + 46, + 613, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 613, + 287, + 636 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 119, + 647, + 199, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 647, + 199, + 677 + ], + "spans": [ + { + "bbox": [ + 119, + 647, + 199, + 677 + ], + "type": "interline_equation", + "content": "f _ {3} (t; R, \\delta) = \\sum_ {t - \\tau} ^ {t} 1", + "image_path": "056d83a4d2b1a8156f5fa9623cfeb06f67db626c60b15512db0514774655a7ac.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 123, + 676, + 287, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 676, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 123, + 676, + 287, + 699 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad \\tau = \\max _ {t ^ {\\prime} \\leq t} t ^ {\\prime} \\tag {5}", + "image_path": "3a6512ce3f2bb58a7cab04acc299073c0db0fa2a24473f71ab2870017334a6c2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 702, + 215, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 702, + 215, + 715 + ], + "spans": [ + { + "bbox": [ + 140, + 702, + 215, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{c c} \\text {s . t .} & v (t ^ {\\prime}; R) > \\delta \\end{array}", + "image_path": "bc94057bc64ad191efe3de745bc56c7608ad9d049dd86bd2c2d1f59d46d116ba.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 379, + 82, + 458, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 82, + 458, + 113 + ], + "spans": [ + { + "bbox": [ + 379, + 82, + 458, + 113 + ], + "type": "interline_equation", + "content": "f _ {4} (t; R, \\delta) = \\sum_ {t - \\tau} ^ {t} 1", + "image_path": "2807b79f9b65024012bedbed1e57c79a06fbe92c4cbefa67235bcf8ae1f042a2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 382, + 111, + 545, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 111, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 382, + 111, + 545, + 135 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad \\tau = \\max _ {t ^ {\\prime} \\leq t} t ^ {\\prime} \\tag {6}", + "image_path": "1276a35f5553dd0dfcdbedea2a0729890118ddcde4df160eb24780f0b51d8f6c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 400, + 137, + 473, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 137, + 473, + 149 + ], + "spans": [ + { + "bbox": [ + 400, + 137, + 473, + 149 + ], + "type": "interline_equation", + "content": "\\begin{array}{l l} \\text {s . t .} & s (t ^ {\\prime}; R) > \\delta \\end{array}", + "image_path": "a4a7b2cda293ebedd16e9a473eda60639b7b78e6ad18e4fa67700616b48b2ebd.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": "We calculated each feature for all possible combinations of parameters " + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\{30,300\\}" + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": " seconds, " + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\delta \\in \\{0.01,0.1,1\\}" + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "inline_equation", + "content": "R \\in \\{H,B,O\\}" + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\{H,B,O\\}" + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": " are approximate head, body and outer bed regions as illustrated in Figure 5. This resulted in " + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "inline_equation", + "content": "(2 + 2 + 3 + 3)" + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": " time-series features per region, i.e. 30 unique time-series features. Additionally, we provided the video-based classifier (see Section 4.3) with time-shifted features " + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "inline_equation", + "content": "f_{i}(t + T)" + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "inline_equation", + "content": "T \\in \\{-90,0, + 90\\}" + }, + { + "bbox": [ + 304, + 154, + 545, + 262 + ], + "type": "text", + "content": " seconds. This gave a total of 90 motion features for each sleep epoch." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 269, + 479, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 479, + 281 + ], + "type": "text", + "content": "4.3. Transferring SleepVST to Video" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 287, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 467 + ], + "type": "text", + "content": "To transfer the SleepVST model to video using the OSV dataset, we removed the linear classification layer learnt during pre-training and used the model as a frozen feature extractor, applying it to sequences of cardiac and respiratory waveforms measured from video to produce sequences of output features " + }, + { + "bbox": [ + 304, + 287, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_o" + }, + { + "bbox": [ + 304, + 287, + 545, + 467 + ], + "type": "text", + "content": ". We then trained a new classifier that used these features, plus the motion features derived in the previous section, to classify individual sleep epochs. This allowed us to leverage the representational capacity of the transformer model, by pre-training on the much larger contact-sensor datasets, whilst also enabling us to incorporate motion information from the video dataset. We followed the design choice of [6] and used a Random Forest [5] for our classifier due to its low computational cost and robustness to overfitting." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 477, + 432, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 432, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 432, + 491 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "content": "Sleep staging performance has commonly been reported in terms of classification accuracy " + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "inline_equation", + "content": "(\\mathrm{Acc}_T)" + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "content": " and total Cohen's kappa " + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "inline_equation", + "content": "(\\kappa_T)" + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "content": " over all sleep epochs in the dataset, and/or using per-night mean statistics, " + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_{\\mu}" + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\kappa_{\\mu}" + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "content": ", i.e.:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 370, + 553, + 545, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 553, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 370, + 553, + 545, + 586 + ], + "type": "interline_equation", + "content": "\\kappa_ {T} = K \\left(\\sum_ {i = 1} ^ {N} c \\left(\\boldsymbol {y} _ {i}, \\hat {\\boldsymbol {y}} _ {i}\\right)\\right) \\tag {7}", + "image_path": "a09a3c8767e498c595d6cc13b71dc408c60297cd50dd82334e2ad5c4e6f97070.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 372, + 588, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 588, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 372, + 588, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\kappa_ {\\mu} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} K \\left(c \\left(\\boldsymbol {y} _ {i}, \\hat {\\boldsymbol {y}} _ {i}\\right)\\right) \\tag {8}", + "image_path": "ab5dc070784d8fd8b20dcc467f3f13201babf60ac6b1c7f3d4e1ad645f588638.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i, \\hat{\\pmb{y}}_i \\in \\mathbb{R}^{C \\times L_i}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " are the one-hot encoded model and scorer labels for recording " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_i" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "c(\\pmb{y}_i, \\hat{\\pmb{y}}_i) \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the confusion matrix between them. " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "K(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the function which calculates the Cohen's kappa statistic [9] from a confusion matrix and is defined in §G. We report all of these statistics, to make fairer comparisons with prior work, and to aid future comparisons." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12483" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 78, + 537, + 261 + ], + "blocks": [ + { + "bbox": [ + 59, + 78, + 537, + 261 + ], + "lines": [ + { + "bbox": [ + 59, + 78, + 537, + 261 + ], + "spans": [ + { + "bbox": [ + 59, + 78, + 537, + 261 + ], + "type": "image", + "image_path": "77d3913bfd21dd10d4b8964c9f7266bc0c1d88709d6c0c2dfa38dfc559816add.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 545, + 301 + ], + "type": "text", + "content": "Figure 6. Example four-class sleep hypnograms from the OSV dataset. (top) Scored by an expert using signals from the vPSG recording. (bottom) Automatically generated from near-infrared video using SleepVST. (Cohen's " + }, + { + "bbox": [ + 46, + 276, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\kappa = 0.71" + }, + { + "bbox": [ + 46, + 276, + 545, + 301 + ], + "type": "text", + "content": " between model and expert sleep stages.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 319, + 221, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 221, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 221, + 333 + ], + "type": "text", + "content": "5.1. Pre-training on Contact Sensors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "content": "From the SHHS dataset, we randomly selected 500 participants who participated in both visits, to form a test set of 1000 nights of data. This ensured that no participant appeared in both the training/validation and test partitions. Our test set is listed in " + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\S H" + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "content": ". From the MESA dataset, we used the same test set of 204 nights as Kotzen et al. [22]. The remaining nights from both the SHHS and MESA datasets were pooled together and randomly split into training and validation sets using a 75:25 split." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 446, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 531 + ], + "type": "text", + "content": "Table 1 compares the overall performance of SleepVST with prior methods for cardio-respiratory sleep staging from contact sensors. These have used input modalities such as the photoplethysmogram (PPG) waveform and the ECG-derived heart rate. Using the ECG and thoracic respiratory signals, SleepVST outperforms prior methods for sleep staging from cardio-respiratory signals on both datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 538, + 194, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 194, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 194, + 552 + ], + "type": "text", + "content": "5.2. Video-based Sleep Staging" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 617 + ], + "type": "text", + "content": "In this section, we report the performance of the SleepVST model after transfer to the OSV video dataset. We follow the evaluation procedure of [6], training and evaluating on the 50 nights of data using 10-fold cross-validation with non-overlapping folds each containing five nights of data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": "In Table 2, we see that our method achieves significantly better performance in terms of both accuracy and Cohen's " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " statistic when compared with prior video-based methods evaluated on similar study populations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Figure 7 shows the total test confusion matrix between model and human expert (polysomnographer) classifications on the OSV dataset. We observe particularly high accuracy in distinguishing Wake and REM sleep stages." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 323, + 539, + 492 + ], + "blocks": [ + { + "bbox": [ + 313, + 323, + 539, + 492 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 539, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 539, + 492 + ], + "type": "image", + "image_path": "b9d84e28653074c427a8cd06f230fe54cc96db6dbead15932cabe78b06086e8f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 504, + 545, + 528 + ], + "lines": [ + { + "bbox": [ + 305, + 504, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 545, + 528 + ], + "type": "text", + "content": "Figure 7. Confusion matrix between SleepVST and expert labels on four-class video based sleep staging." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 549, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 681 + ], + "type": "text", + "content": "Population performance. Figure 8 shows the distribution of Cohen's " + }, + { + "bbox": [ + 304, + 549, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 549, + 545, + 681 + ], + "type": "text", + "content": " values across participants from the OSV dataset plotted against age, sex and Fitzpatrick skin type[2] [14]. We observe a decrease in performance with age, as observed in prior sleep staging work [6, 53]. In older adults, measures of cardio-respiratory activity (e.g. heart rate variability) are known to decline with age [72]. This likely reduces the distance between sleep stages in cardiorespiratory input space and thus makes them harder to distinguish. Additional video data from older subjects would likely help to improve this, which is discussed further in §C." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 481, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 481, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 481, + 714 + ], + "type": "text", + "content": "2 Plotted using representative skin tones from [67]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12484" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 91, + 471, + 213 + ], + "blocks": [ + { + "bbox": [ + 119, + 70, + 474, + 83 + ], + "lines": [ + { + "bbox": [ + 119, + 70, + 474, + 83 + ], + "spans": [ + { + "bbox": [ + 119, + 70, + 474, + 83 + ], + "type": "text", + "content": "Table 1. Comparison of four-class cardio-respiratory sleep staging methods using contact sensors." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 123, + 91, + 471, + 213 + ], + "lines": [ + { + "bbox": [ + 123, + 91, + 471, + 213 + ], + "spans": [ + { + "bbox": [ + 123, + 91, + 471, + 213 + ], + "type": "table", + "html": "
DatasetNMethodModalitiesCohen's κAccuracy / %
κμκTAccμAccT
SHHS [43]296Bakker et al. [3]HR+Thor* +Air†-0.64-76.7
800Sridhar et al. [50]HR (ECG)0.650.6677.377.0
1000SleepVSTECG+Thor*0.730.7582.883.0
MESA [8]296Bakker et al. [3]HR+Thor* +Air†-0.68-79.8
194Sridhar et al. [50]HR (ECG)-0.69-80.0
204Kotzen et al. [22]PPG-0.73‡-82.6‡
204SleepVSTECG+Thor*0.760.7785.185.2
", + "image_path": "fa56b9fb2d63635d8828f99834ae8ca1250d4c34a24c702b992f8780027026ef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 129, + 213, + 361, + 223 + ], + "lines": [ + { + "bbox": [ + 129, + 213, + 361, + 223 + ], + "spans": [ + { + "bbox": [ + 129, + 213, + 361, + 223 + ], + "type": "text", + "content": "*Thoracic respiratory effort. †Nasal airflow. ‡From reported confusion matrices." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 89, + 265, + 504, + 361 + ], + "blocks": [ + { + "bbox": [ + 153, + 245, + 440, + 257 + ], + "lines": [ + { + "bbox": [ + 153, + 245, + 440, + 257 + ], + "spans": [ + { + "bbox": [ + 153, + 245, + 440, + 257 + ], + "type": "text", + "content": "Table 2. Comparison of four-class sleep staging methods using video cameras." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 265, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 89, + 265, + 504, + 361 + ], + "spans": [ + { + "bbox": [ + 89, + 265, + 504, + 361 + ], + "type": "table", + "html": "
DatasetNMethodModalities**Cohen's κAccuracy / %
κμκTAccμAccT
Healthy infants [21]8Kamon et al. [21]Motion0.26-48.0-
Healthy adults [35]6Nochino et al. [35]Motion0.19-40.5-
HealthBed Study [60]46van Meulen et al. [60]HW*0.490.49‡67.967.9‡
Oxford Sleep Volunteers50Carter et al. [6]HR+BR+Motion0.610.64‡73.474.3‡
50SleepVSTHW* + BW† + Motion0.680.7177.778.8
", + "image_path": "481ef70e474959d9f37e10f314b3ac5466a0e8ffa72418b317c30d4ffabcac6d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 94, + 361, + 459, + 370 + ], + "lines": [ + { + "bbox": [ + 94, + 361, + 459, + 370 + ], + "spans": [ + { + "bbox": [ + 94, + 361, + 459, + 370 + ], + "type": "text", + "content": "**Camera-derived. *Cardiac pulse (heart) waveforms.† Respiration (breathing) waveforms.‡ From reported confusion matrices." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 406, + 280, + 571 + ], + "blocks": [ + { + "bbox": [ + 55, + 406, + 280, + 571 + ], + "lines": [ + { + "bbox": [ + 55, + 406, + 280, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 280, + 571 + ], + "type": "image", + "image_path": "2f4a18a5fbd1f584f58d720232e940c2dc5bdf04f9e91452845d9df4589524aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 585, + 287, + 608 + ], + "lines": [ + { + "bbox": [ + 47, + 585, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 287, + 608 + ], + "type": "text", + "content": "Figure 8. Scatter and box plot of Cohen's " + }, + { + "bbox": [ + 47, + 585, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 47, + 585, + 287, + 608 + ], + "type": "text", + "content": " against age, sex and Fitzpatrick skin type" + }, + { + "bbox": [ + 47, + 585, + 287, + 608 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 47, + 585, + 287, + 608 + ], + "type": "text", + "content": " for individuals from the OSV dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "content": "Example SleepVST output. Figure 6 shows an example hypnogram generated entirely from near-infrared video using our method. This example corresponds to the median Cohen's " + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 47, + 629, + 288, + 715 + ], + "type": "text", + "content": " between model and expert from the OSV dataset. The model correctly identifies all three REM cycles, periods of deep sleep, and brief awakenings throughout the night. Additional hypnogram examples can be found in §C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 401, + 407, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 407, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 407, + 414 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 420, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 454 + ], + "type": "text", + "content": "Video transfer strategy. In Table 3, we compare the performance of three methods for transferring the SleepVST model to video data:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 456, + 545, + 491 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 456, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 545, + 479 + ], + "type": "text", + "content": "1. Directly applying the pre-trained SleepVST model to cardiac and respiratory waveforms from video data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 480, + 495, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 495, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 495, + 491 + ], + "type": "text", + "content": "2. Using SleepVST as a feature extractor and:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 492, + 545, + 539 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 321, + 492, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 492, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 321, + 492, + 545, + 514 + ], + "type": "text", + "content": "(a) training a new classification layer without using motion features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 516, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 516, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 321, + 516, + 545, + 539 + ], + "type": "text", + "content": "(b) training a new classification layer that also uses motion features i.e. our best-performing approach." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 545, + 599 + ], + "type": "text", + "content": "Here we see that training a new classifier for video data is essential to the effectiveness of our approach. This step likely allows the classifier to ignore parts of the learnt feature space that do not transfer between contact-sensor and video-derived waveforms i.e. sensor-specific features." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 308, + 641, + 542, + 700 + ], + "blocks": [ + { + "bbox": [ + 306, + 610, + 545, + 633 + ], + "lines": [ + { + "bbox": [ + 306, + 610, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 545, + 633 + ], + "type": "text", + "content": "Table 3. SleepVST video-based sleep staging performance for different model transfer strategies." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 641, + 542, + 700 + ], + "lines": [ + { + "bbox": [ + 308, + 641, + 542, + 700 + ], + "spans": [ + { + "bbox": [ + 308, + 641, + 542, + 700 + ], + "type": "table", + "html": "
κμκTAccμ / %AccT / %
Direct apply0.5100.53666.767.3
Transfer w/o motion0.6600.68976.477.4
Transfer w/ motion0.6770.70877.778.8
", + "image_path": "fc3c9db71f33499f3c73106eca00049cb66b6a354a60964391dd5214309dc893.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "12485" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 75, + 286, + 306 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 286, + 306 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 286, + 306 + ], + "type": "image", + "image_path": "c35a369e0fc4a80ccef27bd57c870129abc084cefc0256345880af757353eb8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 288, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 352 + ], + "type": "text", + "content": "Figure 9. Example four-class hypnograms during fragmented sleep from the OSV dataset. The use of motion features improves accuracy during short awakenings between 22.00 and 00.00." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": "Motion features. Table 4 reports the effectiveness of the two main components of our motion feature set: the set of feature definitions " + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": " and the set of regions used " + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": ". Additional ablations can be found in §C. Here we define and report an additional metric " + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\kappa_{F}" + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": " which is the Cohen's " + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": " statistic computed over epochs near transitions to/from Wake (W) i.e. where " + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^t\\neq \\hat{y}_i^{t + \\Delta} = W" + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^t = W\\neq \\hat{y}_i^{t + \\Delta}" + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\Delta \\in [-2,2]" + }, + { + "bbox": [ + 46, + 374, + 287, + 483 + ], + "type": "text", + "content": " epochs. This metric better quantifies accuracy during fragmented sleep, such as in Figure 9." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "type": "text", + "content": "As measured by " + }, + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\kappa_{F}" + }, + { + "bbox": [ + 46, + 484, + 287, + 544 + ], + "type": "text", + "content": ", we see that motion features aid classification during fragmented sleep. The ability to accurately measure sleep fragmentation is particularly important since it is linked with daytime sleepiness and decreased cognitive performance [29]." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 578, + 276, + 700 + ], + "blocks": [ + { + "bbox": [ + 69, + 559, + 265, + 571 + ], + "lines": [ + { + "bbox": [ + 69, + 559, + 265, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 559, + 265, + 571 + ], + "type": "text", + "content": "Table 4. Ablation study of motion feature parameters." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 578, + 276, + 700 + ], + "lines": [ + { + "bbox": [ + 57, + 578, + 276, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 276, + 700 + ], + "type": "table", + "html": "
AblationParameter Setnf*Cohen's κ
κTκF
Feature setfi ∈ {}00.6890.397
fi ∈ {f1, f2}360.6970.467
fi ∈ {f3, f4}540.7040.465
fi ∈ {f1, f2, f3, f4}900.7080.491
Region(s)R ∈ {H ∪ B}300.7020.461
R ∈ {H ∪ B, O}600.7040.468
R ∈ {H, B, O}900.7080.491
", + "image_path": "e3cf9649ad6d12d46f9ca2d4c69b5999fe1075eaacf3595a011c6ec1a0f2e4f8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 59, + 700, + 100, + 709 + ], + "lines": [ + { + "bbox": [ + 59, + 700, + 100, + 709 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 100, + 709 + ], + "type": "text", + "content": "\\*No. features." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "Cardio-respiratory inputs. In Table 5, we compare the performance of the SleepVST model with variants which use heart rate " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{HR}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " or breathing rate " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{BR}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " as inputs instead of the underlying waveforms, as described in Section 4.1. We see that the use of waveforms, rather than derived heart and breathing rates, leads to improvements in sleep staging accuracy across all datasets." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 347, + 203, + 504, + 289 + ], + "blocks": [ + { + "bbox": [ + 305, + 172, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 172, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 172, + 547, + 196 + ], + "type": "text", + "content": "Table 5. Comparison of four-class sleep staging performance for different cardio-respiratory input combinations to SleepVST." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 203, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 347, + 203, + 504, + 289 + ], + "spans": [ + { + "bbox": [ + 347, + 203, + 504, + 289 + ], + "type": "table", + "html": "
Input signalsCohen's κT
SHHSMESAOSV
xHR, xBR0.6830.7220.681
xHW, xBR0.7180.7360.684
xHR, xBW0.7230.7580.702
xHW, xBW0.7490.7650.708
", + "image_path": "cc4d5190bc6a3c064218d64c9ac02a83829fa96857a817a6a916fc3f62d75680.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 324, + 384, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 384, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 384, + 337 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 347, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 407 + ], + "type": "text", + "content": "We have introduced SleepVST, a transformer model for sleep stage classification from cardio-respiratory waveforms. Using a transfer learning approach, we have shown how SleepVST can be used to perform sleep stage classification entirely from near-infrared video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 410, + 547, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 547, + 530 + ], + "type": "text", + "content": "Our results advance the state-of-the-art in video-based sleep staging, narrowing the gap to expert-level performance and taking a significant step towards important clinical applications. More broadly, our work shows that a model pre-trained on a task using cardio-respiratory waveforms from large contact-sensor datasets can be effectively transferred to waveforms measured from video. This approach can benefit wider applications such as the detection of arrhythmias and sleep apnoes, from cardiac and respiratory waveforms respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 532, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 617 + ], + "type": "text", + "content": "Finally, we believe there is important future work in learning more expressive representations of motion during sleep. In addition to improving the performance of video-based sleep staging, this could also enable the detection of sleep movement disorders, such as REM behaviour disorder, removing the existing need for time-consuming, manual video review [51]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 635, + 409, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 409, + 649 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 409, + 649 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 658, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 547, + 715 + ], + "type": "text", + "content": "This work was supported by the EPSRC Centre for Doctoral Training in Autonomous Intelligent Machines and Systems [EP/S024050/1] and funded by Oxehealth Ltd. Figure 1 was created with BioRender.com. We kindly thank the National Sleep Research Resource for granting access to SHHS and MESA." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12486" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Ying-Wen Bai, Wen-Tai Li, and Cheng-Hsiang Yeh. Design and implementation of an embedded monitor system for body breath detection by using image processing methods. In 2010 Digest of Technical Papers International Conference on Consumer Electronics (ICCE), pages 193-194, 2010. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 270 + ], + "type": "text", + "content": "[3] Jessie P. Bakker, Marco Ross, Ray Vasko, Andreas Cerny, Pedro Fonseca, Jeff Jasko, Edmund Shaw, David P. White, and Peter Anderer. Estimating sleep stages using cardiorespiratory signals: validation of a novel algorithm across a wide range of sleep-disordered breathing severity. Journal of Clinical Sleep Medicine, 17(7):1343-1354, 2021. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "text", + "content": "[4] Guha Balakrishnan, Fredo Durand, and John Guttag. Detecting Pulse from Head Motions in Video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 3430-3437, Portland, OR, USA, 2013. IEEE. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 286, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 286, + 339 + ], + "type": "text", + "content": "[5] Leo Breiman. Random Forests. Machine Learning, 45(1): 5-32, 2001. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 418 + ], + "type": "text", + "content": "[6] Jonathan Carter, João Jorge, Bindia Venugopal, Oliver Gibson, and Lionel Tarassenko. Deep Learning-Enabled Sleep Staging From Vital Signs and Activity Measured Using a Near-Infrared Video Camera. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 5940-5949, Vancouver, BC, Canada, 2023. IEEE. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "text", + "content": "[7] Weixuan Chen and Daniel McDuff. Deepphys: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 349-365, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "text", + "content": "[8] Xiaoli Chen, Rui Wang, Phyllis Zee, Pamela L. Lutsey, Sogol Javaheri, Carmela Alcantara, Chandra L. Jackson, Michelle A. Williams, and Susan Redline. Racial/Ethnic Differences in Sleep Disturbances: The Multi-Ethnic Study of Atherosclerosis (MESA). Sleep, 38(6):877-888, 2015. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 287, + 555 + ], + "type": "text", + "content": "[9] Jacob Cohen. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1): 37-46, 1960. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[10] Kris Cuppens, Lieven Lagae, Berten Ceulemans, Sabine Van Huffel, and Bart Vanrumste. Automatic video detection of body movement during sleep based on optical flow in pediatric patients with epilepsy. Medical & Biological Engineering & Computing, 48(9):923-931, 2010. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 712 + ], + "type": "text", + "content": "[11] Heidi Danker-Hopfe, Peter Anderer, Josef Zeitlhofer, Marion Boeck, Hans Dorn, Georg Gruber, Esther Heller, Erna Loretz, Doris Moser, Silvia Parapatics, Bernd Saletu, Andrea Schmidt, and Georg Dorffner. _Interrater reliability for sleep scoring according to the Rechtschaffen & Kales and the new AASM standard_. Journal of Sleep Research, 18(1):74-84, 2009. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1365-2869.2008.00700.x. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[12] Koldo De Miguel, Alberto Brunete, Miguel Hernando, and Ernesto Gambao. Home Camera-Based Fall Detection System for the Elderly. Sensors, 17(12):2864, 2017. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[14] T. B. Fitzpatrick. The validity and practicality of sun-reactive skin types I through VI. Archives of Dermatology, 124(6): 869-871, 1988. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "text", + "content": "[15] Marco Hafner, Martin Stepanek, Jirka Taylor, Wendy M. Troxel, and Christian van Stolk. Why Sleep Matters—The Economic Costs of Insufficient Sleep. *Rand Health Quarterly*, 6(4):11, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 251, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 281 + ], + "type": "text", + "content": "[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "text", + "content": "[18] Adrienne Heinrich, Xavier Aubert, and Gerard de Haan. Body movement analysis during sleep based on video motion estimation. In 2013 IEEE 15th International Conference on e-Health Networking, Applications and Services (Healthcom 2013), pages 539–543, 2013. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 436 + ], + "type": "text", + "content": "[19] D. W. Hudgel, R. J. Martin, B. Johnson, and P. Hill. Mechanics of the respiratory system and breathing pattern during sleep in normal humans. Journal of Applied Physiology: Respiratory, Environmental and Exercise Physiology, 56(1): 133-137, 1984. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "text", + "content": "[20] C. Iber. The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology, and Technical Specification. 2007. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "text", + "content": "[21] Masamitsu Kamon, Shima Okada, Masafumi Furuta, and Koki Yoshida. Development of a non-contact sleep monitoring system for children. Frontiers in Digital Health, 4, 2022. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 515, + 547, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 547, + 580 + ], + "type": "text", + "content": "[22] Kevin Kotzen, Peter H. Charlton, Sharon Salabi, Lea Amar, Amir Landesberg, and Joachim A. Behar. SleepPPG-Net: A Deep Learning Algorithm for Robust Sleep Staging From Continuous Photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 27(2):924–932, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[23] Yee Fun Lee, Dmitry Gerashchenko, Igor Timofeev, Brian J. Bacskai, and Ksenia V. Kastanenka. Slow Wave Sleep Is a Promising Intervention Target for Alzheimer's Disease. Frontiers in Neuroscience, 14, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "text", + "content": "[24] Yun Ji Lee, Jae Yong Lee, Jae Hoon Cho, and Ji Ho Choi. Interrater reliability of sleep stage scoring: a meta-analysis. Journal of Clinical Sleep Medicine, 18(1):193-202, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Michael H. Li, Azadeh Yadollahi, and Babak Taati. A noncontact vision-based system for respiratory rate estimation. In 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pages 2119-2122, 2014. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12487" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Xin Liu, Brian Hill, Ziheng Jiang, Shwetak Patel, and Daniel McDuff. EfficientPhys: Enabling Simple, Fast and Accurate Camera-Based Cardiac Measurement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5008-5017, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[28] Yawei Luo, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Taking a Closer Look at Domain Shift: Category-Level Adversaries for Semantics Consistent Domain Adaptation. pages 2507-2516, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "text", + "content": "[29] S E Martin, H M Engleman, I J Deary, and N J Douglas. The effect of sleep fragmentation on daytime function. American Journal of Respiratory and Critical Care Medicine, 153(4): 1328-1332, 1996. Publisher: American Thoracic Society - AJRCCM. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 293 + ], + "type": "text", + "content": "[30] Daniel McDuff. Camera Measurement of Physiological Vital Signs. ACM Computing Surveys, 55(9):1-40, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "text", + "content": "[31] Sara Mahvash Mohammadi, Shirin Enshaeifar, Adrian Hilton, Derk-Jan Dijk, and Kevin Wells. Transfer Learning for Clinical Sleep Pose Detection Using a Single 2D IR Camera. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 29:290-299, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "text", + "content": "[32] K. Nakajima, A. Osa, and H. Miike. A method for measuring respiration and physical activity in bed by optical flow analysis. In Proceedings of the 19th Annual International Conference of the IEEE Engineering in Medicine and Biology Society., pages 2054-2057 vol.5, 1997. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "text", + "content": "[33] Kazuki Nakajima, Yoshiaki Matsumoto, and Toshiyo Tamura. Development of real-time image sequence analysis for evaluating posture change and respiratory rate of a subject in bed. Physiological Measurement, 22(3):N21-N28, 2001. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "text", + "content": "[34] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations, 2022. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "text", + "content": "[35] Teruaki Nochino, Yuko Ohno, Takafumi Kato, Masako Tani-ike, and Shima Okada. Sleep stage estimation method using a camera for home use. Biomedical Engineering Letters, 9 (2):257-265, 2019. 1, 3, 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[36] Jiapu Pan and Willis J. Tompkins. A Real-Time QRS Detection Algorithm. IEEE Transactions on Biomedical Engineering, BME-32(3):230-236, 1985. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "text", + "content": "[37] James Pardey, Stephen Roberts, Lionel Tarassenko, and John Stradling. A new approach to the analysis of the human sleep/wakefulness continuum. Journal of sleep research, 5 (4):201-210, 1996. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 647 + ], + "type": "text", + "content": "[38] Aakash K. Patel, Vamsi Reddy, and John F. Araujo. Physiology, Sleep Stages. StatPearls Publishing, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[39] Huy Phan and Kaare Mikkelsen. Automatic sleep staging of EEG signals: recent development, challenges, and future directions. Physiological Measurement, 43(4):04TR01, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[40] Huy Phan, Oliver Y. Chén, Minh C. Tran, Philipp Koch, Alfred Mertins, and Maarten De Vos. XSleepNet: Multi-View" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Sequential Model for Automatic Sleep Staging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 5903-5915, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "text", + "content": "[41] Huy Phan, Kaare Mikkelsen, Oliver Y. Chen, Philipp Koch, Alfred Mertins, and Maarten De Vos. SleepTransformer: Automatic Sleep Staging With Interpretability and Uncertainty Quantification. IEEE Transactions on Biomedical Engineering, 69(8):2456-2467, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 162, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 216 + ], + "type": "text", + "content": "[42] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Advancements in Noncontact, Multiparameter Physiological Measurements Using a Webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. Conference Name: IEEE Transactions on Biomedical Engineering. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "text", + "content": "[43] S. F. Quan, B. V. Howard, C. Iber, J. P. Kiley, F. J. Nieto, G. T. O'Connor, D. M. Rapoport, S. Redline, J. Robbins, J. M. Samet, and P. W. Wahl. The Sleep Heart Health Study: design, rationale, and methods. Sleep, 20(12):1077-1085, 1997. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 271, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 327 + ], + "type": "text", + "content": "[44] Mustafa Radha, Pedro Fonseca, Arnaud Moreau, Marco Ross, Andreas Cerny, Peter Anderer, Xi Long, and Ronald M. Aarts. A deep transfer learning approach for wearable sleep stage classification with photoplethysmography. npj Digital Medicine, 4(1):1-11, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "text", + "content": "[45] S.J. Redmond and C. Heneghan. Cardiorespiratory-based sleep staging in subjects with obstructive sleep apnea. IEEE Transactions on Biomedical Engineering, 53(3):485-496, 2006. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[46] A. J. Schwichtenberg, Jeehyun Choe, Ashleigh Kellerman, Emily A. Abel, and Edward J. Delp. Pediatric Videosomnography: Can Signal/Video Processing Distinguish Sleep and Wake States? Frontiers in Pediatrics, 6:158, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "text", + "content": "[47] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90(2):227-244, 2000. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "type": "text", + "content": "[48] Z. Shinar, A. Baharav, Y. Dagan, and S. Akselrod. Automatic detection of slow-wave-sleep using heart rate variability. In Computers in Cardiology 2001. Vol.28, pages 593-596, 2001. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "type": "text", + "content": "[49] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 558, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 613 + ], + "type": "text", + "content": "[50] Niranjan Sridhar, Ali Shoeb, Philip Stephens, Alaa Kharbouch, David Ben Shimol, Joshua Burkart, Atiyeh Ghoreyshi, and Lance Myers. Deep learning for automated sleep staging using instantaneous heart rate. npj Digital Medicine, 3(1):1-10, 2020. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 635 + ], + "type": "text", + "content": "[51] Ambra Stefani and Birgit Hogl. Sleep in Parkinson's disease. Neuropsychopharmacology, 45(1):121-128, 2020. 1, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 690 + ], + "type": "text", + "content": "[52] Ambra Stefani, David Gabelaia, Thomas Mitterling, Werner Poewe, Birgit Hogl, and Birgit Frauscher. A Prospective Video-Polysomnographic Analysis of Movements during Physiological Sleep in 100 Healthy Sleepers. Sleep, 38 (9):1479-1487, 2015. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[53] Haoqi Sun, Wolfgang Ganglberger, Ezhil Panneerselvam, Michael J Leone, Syed A Quadri, Balaji Goparaju, Ryan A" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12488" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Tesh, Oluwaseun Akeju, Robert J Thomas, and M Brandon Westover. Sleep staging from electrocardiography and respiration with deep learning. *Sleep*, 43(7):zsz306, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 163 + ], + "type": "text", + "content": "[54] K. S. Tan, R. Saatchi, H. Elphick, and D. Burke. Real-time vision based respiration monitoring system. In 2010 7th International Symposium on Communication Systems, Networks & Digital Signal Processing (CSNDSP 2010), pages 770-774, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 220 + ], + "type": "text", + "content": "[55] Alexander Trumpp, Johannes Lohr, Daniel Wedekind, Martin Schmidt, Matthias Burghardt, Axel R. Heller, Hagen Malberg, and Sebastian Zaunseder. Camera-based photoplethysmography in an intraoperative setting. *BioMedical Engineering OnLine*, 17(1):33, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "text", + "content": "[56] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Oxhealth Vital Signs, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200019.pdf.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 269, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 313 + ], + "type": "text", + "content": "[57] U.S. Food and Drug Administration (FDA). De Novo Classification Request for Gili Pro Biosensor, 2020, https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200038.pdf.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 370 + ], + "type": "text", + "content": "[58] Mark van Gastel, Sander Stuijk, Sebastiaan Overeem, Johannes P. van Dijk, Merel M. van Gilst, and Gerard de Haan. Camera-Based Vital Signs Monitoring During Sleep - A Proof of Concept Study. IEEE Journal of Biomedical and Health Informatics, 25(5):1409-1418, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 373, + 287, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 416 + ], + "type": "text", + "content": "[59] Hans van Gorp, Iris A M Huijben, Pedro Fonseca, Ruud J G van Sloun, Sebastiaan Overeem, and Merel M van Gilst. Certainty about uncertainty in sleep staging: a theoretical framework. Sleep, 45(8), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 418, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 483 + ], + "type": "text", + "content": "[60] Fokke B. van Meulen, Angela Grassi, Leonie van den Heuvel, Sebastiaan Overeem, Merel M. van Gilst, Johannes P. van Dijk, Henning Maass, Mark J. H. van Gastel, and Pedro Fonseca. Contactless Camera-Based Sleep Staging: The HealthBed Study. Bioengineering, 10(1):109, 2023. 1, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "text", + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. arXiv:1706.03762 [cs], 2017. arXiv: 1706.03762. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "text", + "content": "[62] Mauricio Villarroel, Sitthichok Chaichulee, João Jorge, Sara Davis, Gabrielle Green, Carlos Arteta, Andrew Zisserman, Kenny McCormick, Peter Watkinson, and Lionel Tarassenko. Non-contact physiological monitoring of preterm infants in the Neonatal Intensive Care Unit. npj Digital Medicine, 2(1):1-18, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[63] Olivia Walch, Yitong Huang, Daniel Forger, and Cathy Goldstein. Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device. *Sleep*, 42(12), 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[64] Qiongyan Wang, Hanrong Cheng, and Wenjin Wang. Feasibility of Exploiting Physiological and Motion Features for Camera-based Sleep Staging: A Clinical Study. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5, 2023. ISSN: 2694-0604. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 464 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[65] Wenjin Wang and Albertus C. den Brinker. Algorithmic insights of camera-based respiratory motion extraction. Physiological measurement, 43(7):075004, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[66] Wenjin Wang, Albertus C. den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic Principles of Remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[67] William H. Ward, Fernando Lambreton, Neha Goel, Jian Q. Yu, and Jeffrey M. Farma. Clinical Presentation and Staging of Melanoma. In Cutaneous Melanoma: Etiology and Therapy. Codon Publications, Brisbane (AU), 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 198, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 229 + ], + "type": "text", + "content": "[68] Johanna Wilde-Frenz and Hartmut Schulz. Rate and Distribution of Body Movements during Sleep in Humans. Perceptual and Motor Skills, 56(1):275-283, 1983. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "text", + "content": "[69] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. SimPer: Simple Self-Supervised Learning of Periodic Targets, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 265, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 330 + ], + "type": "text", + "content": "[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. PhysFormer: Facial Video-based Physiological Measurement with Temporal Difference Transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176-4186, New Orleans, LA, USA, 2022. IEEE. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 396 + ], + "type": "text", + "content": "[71] Guo-Qiang Zhang, Licong Cui, Remo Mueller, Shiqiang Tao, Matthew Kim, Michael Rueschman, Sara Mariani, Daniel Mobley, and Susan Redline. The National Sleep Research Resource: towards a sleep data commons. Journal of the American Medical Informatics Association: JAMIA, 25 (10):1351-1358, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 398, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 464 + ], + "type": "text", + "content": "[72] Dan Ziegler, G. Laux, K. Dannehl, M. Spüler, H. Mühlen, P. Mayer, and F.a. Gries. Assessment of Cardiovascular Autonomic Function: Age-related Normal Ranges and Reproducibility of Spectral Analysis, Vector Analysis, and Standard Tests of Heart Rate Variation and Blood Pressure Responses. Diabetic Medicine, 9(2):166-175, 1992. 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12489" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_content_list.json b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5dd20524d7c75c8814c7dd078617399b09bf30b --- /dev/null +++ b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_content_list.json @@ -0,0 +1,1450 @@ +[ + { + "type": "text", + "text": "Slice3D: Multi-Slice, Occlusion-Revealing, Single View 3D Reconstruction", + "text_level": 1, + "bbox": [ + 111, + 130, + 854, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yizhi Wang $^{1}$ , Wallace Lira $^{1}$ , Wenqi Wang $^{2}$ , Ali Mahdavi-Amiri $^{1}$ , Hao Zhang $^{1}$ $^{1}$ Simon Fraser University $^{2}$ Tsinghua University", + "bbox": [ + 176, + 179, + 789, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a9295f996fd930fc7e69d6c20c126c3bba43b6605a0962a1940c108fe147907.jpg", + "image_caption": [ + "Figure 1. Our single-view 3D reconstruction method, Slice3D, predicts multi-slice images to reveal occluded parts without changing the camera (in contrast to multi-view synthesis), and then lifts the slices into a 3D model. This leads to better recovery of occluded parts than both encoder-decoder based methods such as DISN [65] and 3D reconstructions from diffusion-synthesized multi-view images (via Zero-1-to-3 [29]) such as One-2-3-45 [28], where the key challenge of the latter is multi-view inconsistency, as shown on the right." + ], + "image_footnote": [], + "bbox": [ + 98, + 224, + 874, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 448, + 313, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce multi-slice reasoning, a new notion for single-view 3D reconstruction which challenges the current and prevailing belief that multi-view synthesis is the most natural conduit between single-view and 3D. Our key observation is that object slicing is a more direct, and hence more advantageous, means to reveal occluded structures than altering camera views. Specifically, slicing can peel through any occluder without obstruction, and in the limit (i.e., with infinitely many slices), it is guaranteed to unveil all hidden object parts. We realize our idea by developing Slice3D, a novel method for single-view 3D reconstruction which first predicts multi-slice images from a single RGB input image and then integrates the slices into a 3D model using a coordinate-based transformer network to product a signed distance function. The slice images can be regressed or generated, both through a U-Net based network. For the former, we inject a learnable slice indicator code to designate each decoded image into a spatial slice location, while the slice generator is a denoising diffusion model operating on the entirety of slice images stacked on the input channels. We conduct extensive evaluation against state-of-the-art alternatives to demonstrate superiority of our method, especially in recovering complex and severely occluded shape structures, amid ambiguities. All Slice3D results were produced by networks trained on a single Nvidia A40 GPU, with an inference time of less than 20 seconds.", + "bbox": [ + 75, + 489, + 473, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 448, + 630, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-view 3D reconstruction has been one of the extensively studied problems in computer vision. Despite the rapid advances on learning-based approaches to tackle this problem, including the latest attempts to leverage large foundational models and powerful diffusion-based generators, the most fundamental challenge still remains: how to faithfully reconstruct occluded parts from just one view?", + "bbox": [ + 496, + 476, + 890, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional encoder-decoder based methods, e.g., [9, 17, 25, 44, 53, 65], rely on direct 3D supervision to learn mappings from global or local image features to 3D, but such features from just a single view often do not have the capacity to handle severe occlusion. Also, recent evi", + "bbox": [ + 496, + 585, + 892, + 660 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdecd051ecd2b83eb9f88b108329a0f2d0d4a7fc73e11c1c9188906c70f6209d.jpg", + "image_caption": [ + "Figure 2. Multi-view vs. multi-slice for occlusion revelation, with orange edges showing revealed parts and black edges showing occluded parts. In the limit, multi-view (i.e., infinitely many views) may still leave undisclosed parts (a), e.g., the bulb, while infinitely many slices would guarantee complete structure revelation (b). With limited slices (c) and views (d), four each, the multi-slice approach tends to reveal more shape structures." + ], + "image_footnote": [], + "bbox": [ + 504, + 688, + 888, + 790 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9881", + "bbox": [ + 482, + 941, + 513, + 954 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41d5ff1055a14c2b84ff2ab10614200f1ef2d8c7149523c8261fdbfa7492bf64.jpg", + "image_caption": [ + "Figure 3. Slice3D for single-view 3D reconstruction from multi-slice images. The example shows 4 slice images per $X$ , $Y$ , $Z$ direction, which can be obtained by either a regressor or a generator. Regression is carried out by a U-Net encoder-decoder, where the decoder takes an extra slice indicator code that is appended to the encoder output, to obtain a slice image. There is one indicator code for each of the $3 \\times 4 = 12$ slices, that is optimized during training. Our generator is a denoising diffusion model which produces a concatenation of slice images following the distribution of the ground-truth (GT) images, when conditioned on the single-view input. Inference from random noises may output multiple plausible results. The 3D reconstruction stage inputs the predicted slice images and produces an implicit field. We first extract hierarchical CNN features from the input slices, then project a 3D query point onto the 2D planes of the feature maps, and finally gather all the query-related features and send them into a transformer [58] to predict the signed distance at the query point." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 478, + 357 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/828ffce824fa801b7df7fa2abd02ba932f556386e94ea75d0b90212101b549db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 88, + 880, + 357 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/843f584c7b0b5b95fa3bf7b21c3be1bdfceb46d7a1eca93635ef66b4b75fd95a.jpg", + "image_caption": [ + "Figure 4. Our slice image prediction networks are trained using occlusion-revealing multi-slice images produced from a GT 3D object mesh along a predetermined set of slicing directions (the $X$ , $Y$ , $Z$ axes of the object bounding box in the example above), while keeping the hollows as holes in the images, through which the back faces of the mesh can be visible. The 3D mesh may be textured, but the slice images are all rendered without texture." + ], + "image_footnote": [], + "bbox": [ + 112, + 496, + 434, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dence suggests that many of these methods are essentially performing image classification or retrieval [55]. As such, occlusion recovery is by and large a \"hit or miss\". Differentiable rendering enables 3D reasoning from images without 3D supervision [7, 38], but image-space rendering losses alone do not account for occluded structures.", + "bbox": [ + 75, + 809, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, more and more works model the reconstruction problem as a single-view conditioned generative task, aiming to learn a distribution of unseen structures. Typically, multi-view images are generated from single view via diffusion. This is followed by multi-view reconstruction, e.g., via neural radiance fields (NeRF) [6, 28, 42], to obtain 3D outputs. However, multi-view consistency is a reoccurring challenge (see Fig. 1), especially over disparate views.", + "bbox": [ + 496, + 500, + 890, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a novel approach to single-view 3D reconstruction which is occlusion-revealing (Fig. 2) by first predicting multi-slice images from a single-view input and then integrating the slices into a 3D model; see Fig. 3. Fig. 4 shows slice images provided during training along a set of slicing directions. Each slice represents the projected image, in the input camera view, of one slice of the 3D mesh model. For the TV example shown, a completely occluded leg, in the camera view, is revealed in the slice images.", + "bbox": [ + 496, + 622, + 890, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The motivations behind our multi-slice reconstruction are three-fold. First, slicing through a volume or 3D mesh is a natural way to reveal occluded parts. Second, with the slice images forming an intermediate representation, our approach breaks the difficult problem of mapping from single view to 3D into two comparatively simpler tasks: slice prediction followed by 3D reconstruction from slices.", + "bbox": [ + 496, + 762, + 890, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More importantly, when contrasting our multi-slice representation with the popular multi-view approach for occlu", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9882", + "bbox": [ + 482, + 941, + 514, + 953 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2dd1e33200e1c1aacc6cbde8ed9088be2e1712e62e14b04920e0b3203080b985.jpg", + "image_caption": [ + "Figure 5. Singe-view reconstruction of severely occluded parts in a lamp (e.g., the bulb) by Slice3D vs. One-2-3-45 [28], a state-of-the-art multi-view approach. The multi-view images in [28] were obtained by Zero-1-to-3 [29], which was trained on renderings of 800K 3D models [10]. It is conceivable that with a limited number of views (12, the same as our slice count), the bulbs were barely visible due to occlusion by the lamp shades and bases." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 449, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion revelation, we observe that slicing in the limit (i.e., with infinitely many slices) is guaranteed to unveil all hidden structures, while the same cannot be said about multi-view captures, as illustrated in Fig. 2 (a-b), for a 2D lamp example. In practice, when slice and view counts are both limited, the slicing approach is more occlusion-revealing in general since it can peel through any occluders without any obstruction, unlike the multi-view case; see (c-d). On the technical front, neither our slice prediction step nor the 3D reconstruction step is concerned with camera view changes or multi-view consistency, which facilitates the use of convolutional features. In Fig. 5, we show the above contrasting characteristics of the multi-slice vs. multi-view approaches manifested in a real 3D lamp reconstruction.", + "bbox": [ + 75, + 299, + 468, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As outlined in Fig. 3, our slice images can be produced by either a regressive network or a generative one, both employing a U-Net architecture, with network losses measured against ground-truth (GT) images obtained by slicing 3D models from the training set; see Fig. 4. In the regressive network, we inject a learnable slice indicator code into the encoder-decoder setup to designate each decoded image into the proper spatial slice location. The slice generator is a denoising diffusion model operating on separate channels for different slice images. In the second stage, 3D reconstruction from the obtained slice images, we design a coordinate-based transformer to obtain an implicit field.", + "bbox": [ + 75, + 510, + 468, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our network is coined Slice3D. We conduct extensive experiments on ShapeNet [5] and Objverse [10] for evaluation. Quantitative and qualitative comparisons are made to representative methods including encoder-decoders capturing local image features [65], autoregressive latent models [36], generative model inversion [42], and three recent methods with diffusion + NeRF [6, 28, 33], demonstrating superiority of our method in terms of reconstruction quality, generalizability, and handling of ambiguities.", + "bbox": [ + 75, + 691, + 468, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 844, + 215, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The earliest inspiration for our multi-slice 3D reconstruction idea was GANHopper by Lira et al. [26]. In the con", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text of unsupervised image-to-image translation, the key idea of their work is to gradually execute the translation between disparate domains by making small \"hops,\" rather than completing it in one go. We were then motivated to tackle the single-view 3D reconstruction problem as a domain translation task, hopping through the multi-slice images so as to bridge the gap between 2D and 3D.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural fields. Neural implicit functions [63] are learned by a neural network to predict such properties as color, occupancy, opacity, and surface normals. Earlier works such as IM-Net [8], occupancy networks [34], and DeepSDF [41] represent a 3D shape as a latent code and send it to a decoder formed by MLPs to predict occupancy or SDF values for any query point sampled in 3D. NeRF [35] and its variants [15] represent a 3D scene using a neural radiance field for novel view synthesis. Many works on neural 3D reconstruction [15, 63] take on single- or multi-view images as well as point clouds as input. In contrast, the reconstruction module in Slice3D is built on multi-slice images.", + "bbox": [ + 496, + 199, + 892, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single-view reconstruction from image features. An early work to exploit neural image features for 3D reconstruction was Pixel2Mesh [59], which uses a graph representation. Subsequently, neural implicit fields have been widely adopted for single-view reconstruction. IM-Net [8] first extracts a global latent code, from an input view, to learn a 3D implicit field. Auto-SDF [36] and ShapeFormer [66] employ vector quantization (VQ) [57] to improve the representation of latent features. More recent works [22, 37, 68] consider single-view reconstruction as a 3D generation task conditioned on the latent code of the input view. However, a global latent code is insufficient to capture finer details on a 3D object, often leading to overly smoothed or retrieved outcomes that do not faithfully represent the input. To alleviate this problem, both DISN [65] and $\\mathrm{D}^2\\mathrm{IM}$ -Net [25] opted to learn local images features, along with global ones, to improve the recovery of fine-grained topological (e.g., thin) features and surface details, respectively.", + "bbox": [ + 496, + 383, + 892, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D reconstruction via multi-view. Conventional 3D reconstruction approaches [14] involve multi-view reasoning. With recent advances in generative AI, a prevailing practice in single-view 3D reconstruction is to first employ generative models such as GANs [16, 23] and diffusion [11, 20, 45, 50] to generate several novel views from the input view, and then feed them to a multi-view 3D reconstruction module such as NeRF. NeRF-based surface reconstruction [39, 60, 67] can be further employed to output 3D meshes. EG3D [3] and NeRF-from-Image [42] build NeRFs or its variants upon the outputs of a style-GAN [23], while more recent works [4, 6, 18, 28-30, 33, 42, 43, 51, 62, 64] are favoring diffusion models such as Stable Diffusion (SD) [45]. To date, Zero-1-to-3 [29] represents the state of the art on single- to multi-view synthesis; it is built on SD and fine-tuned by 800K 3D models from Objaverse [10].", + "bbox": [ + 496, + 657, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9883", + "bbox": [ + 482, + 941, + 514, + 953 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the multi-view consistency problem, which is an reoccurring challenge for the image-space generative methods, recent methods [30, 31, 43, 52, 71] aim to enhance the consistency by performing spatial attention between different views, which dramatically increases the computational overhead. Nevertheless, the consistency of two disparate views is still far from satisfactory. A deeper reason is that standard convolutional neural networks (CNNs) are not inherently rotation-invariant, making it difficult to build correspondences between significantly different views.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Slice representation. There have been techniques to use object slices to reconstruct a 3D shape, e.g., as inspired by Magnetic Resonance Imaging (MRI). Early works in computer graphics employ either parallel cross-sectional curves, e.g., [24], or non-parallel curve networks, e.g., [27], as inputs. Recently, with the advent of machine learning, OREx [48] and Cut-and-Approximate [40] develop neural reconstruction models to obtain 3D meshes, also from a sparse collection of 2D planar cross-sections, while Neural-Slice [21] uses slicing to obtain 3D shapes from 4D tetrahedral meshes. However, none of these techniques, nor others that we are aware of, address the problem of single-view 3D reconstruction since their inputs all consist of 3D data such as curve slices or point clouds. Moreover, our method learns to predict much fewer slice images and use them as input to a transformer-based reconstruction module.", + "bbox": [ + 75, + 244, + 472, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dip. Dip transform [2] is an ingenious way to acquire occluded 3D structures by measuring fluid displacement when an object is repeatedly oriented and dipped into a fluid. Similar to slicing, Dip was also designed to \"reach\" occluded parts that optical captures via lines of sight cannot. However, both the occlusion revelation (liquid penetration vs. slicing) and target applications (acquisition vs. single-view reconstruction) are completely different.", + "bbox": [ + 75, + 489, + 470, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-plane image (MPI). Introduced by Zhou et al. [72], an MPI representation for a scene consists of a set of frontoparallel planes at fixed depths, with each image intended to capture scene appearance and visibility at a specific depth. Conceptually, both MPs and our slice images provide a layered scene representation for occlusion revelation. However, MPs [13, 19, 56, 69, 72] have only been employed for novel view synthesis, not 3D reconstruction, and the prediction networks were trained on multi-view images, e.g., from Youtube videos [72]. The generated MPs only reveal a small amount of occluded scenes to account for view shifts. In contrast, our method directly cuts through a 3D object to reveal all the occluded parts at several depth ranges.", + "bbox": [ + 75, + 614, + 472, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 828, + 169, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method is based on a slicing representation; see Section 3.1 and Fig. 4. Given a single-view RGB image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ of a 3D object, we first learn a slicing network", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(Section 3.2) which takes $I$ as input and outputs its corresponding multi-slice images. Then, an SDF prediction network is trained (Section 3.3) to leverage image features from the slices to produce an implicit field of the object.", + "bbox": [ + 496, + 90, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Slicing Representation", + "text_level": 1, + "bbox": [ + 498, + 160, + 709, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 4, given a 3D object mesh, we slice it along the $X$ , $Y$ , and $Z$ axes of its bounding box and render the sliced segments. The slice images can optionally include textures or have holes filled. By default, we remove the textures and keep the holes, which could reveal back faces of the mesh, and overall, it produces the best experimental results. The rendered slice images consist of $\\{S_i^j | 1 \\leq i \\leq N_s, j \\in \\{X,Y,Z\\}\\}$ , where $S_i^j \\in \\mathbb{R}^{H \\times W \\times 1}$ , $j$ denotes the axis and $N_s$ is the number of slices along each axis. In our current implementation, we produce $N_s = 4$ slices per direction. Compared to planar cross-sections or tomography scans which consist of densely stacked planar images, our slicing representation is more compact with each slice image revealing more object semantics.", + "bbox": [ + 496, + 183, + 893, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Slice Regression and Generation", + "text_level": 1, + "bbox": [ + 498, + 402, + 785, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an image $I$ , observing a 3D shape from a single view, we aim to train a slicing network $F_{s}$ that can produce the slice images from $I$ : $F_{s}(I) = \\{\\hat{S}_{i}^{j}\\}$ . We have two different designs for such a network: regressive and generative. The former aims to regress GT slices, assuming deterministic occluded parts based on the partial observation of the input view. The latter aims to generate slices that are consistent with the input view, considering various possible occluded parts that do not necessarily match the GT slices.", + "bbox": [ + 496, + 426, + 893, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Regression-based Slicing", + "text_level": 1, + "bbox": [ + 498, + 580, + 725, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a U-Net [46] encoder-decoder to regress the slices from $I$ . As shown in the top-left of Fig. 3, an indicator code $E_{i}^{j}\\in \\mathbb{R}^{N_{e}}$ is concatenated with the latent code of $I$ to indicate which slice to output: $\\hat{S}_i^j = F_d([F_e(I);E_i^j])$ , where $F_{e}$ and $F_{d}$ are the encoder and decoder of $F_{s}$ , respectively, and $[\\cdot ]$ denotes concatenation.", + "bbox": [ + 496, + 604, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The indicator codes $\\{E_i^j\\}$ are randomly initialized and jointly optimized with the whole network. The loss function $\\mathcal{L}_{reg}$ is the reconstruction error measured by the $\\mathcal{L}_1$ -Norm and perceptual loss [70] $(\\mathcal{L}_p)$ for all slices:", + "bbox": [ + 496, + 696, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\frac {\\sum_ {i , j} \\left| \\left| \\hat {S} _ {i} ^ {j} - S _ {i} ^ {j} \\right| \\right| _ {1} + \\mathcal {L} _ {p} \\left(\\hat {S} _ {i} ^ {j} , S _ {i} ^ {j}\\right)}{| S |}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 767, + 890, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $|S| = 3 \\cdot N_{s}$ is the total number of slices.", + "bbox": [ + 500, + 813, + 818, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Generation-based Slicing", + "text_level": 1, + "bbox": [ + 500, + 845, + 730, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single-view occlusion revelation is clearly ill-posed, with potential ambiguities leading to multiple plausible 3D re", + "bbox": [ + 496, + 869, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9884", + "bbox": [ + 482, + 941, + 514, + 953 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e97b9d5609c4e2ba6cd05f620fc452a8ffe9a50712b29d667405001a4dedbe2b.jpg", + "image_caption": [ + "Figure 6. The pipeline of generation-based slicing via DDPM. $\\bigoplus$ denotes the injection of noise to slice images." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 467, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "constructions. To this end, we consider slicing as an image-conditioned generation task and develop a generation-based slicing via Denoising Diffusion Probabilistic Models (DDPM) [20], whose detailed pipeline is shown in Fig. 6.", + "bbox": [ + 75, + 304, + 467, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key challenge of applying DDPM lies in how to generate consistent slices. Rather than diffusing each single slice, we propose to concatenate all slices into an aggregated structure $[S]$ , along either the color dimension $([S] \\in \\mathbb{R}^{H \\times W \\times |S|})$ or a spatial dimension such as height $([S] \\in \\mathbb{R}^{(|S| \\cdot H) \\times W \\times 1})$ . The concatenation of slices instructs DDPM to learn the correlation of slices, yielding consistent generation results. By default, we concatenate along the color dimension. The comparison of these two concatenation ways can be found in the supplementary.", + "bbox": [ + 75, + 366, + 467, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following DDPM, we first inject a noise $\\epsilon \\sim \\mathcal{N}(0,1)$ to $[S]$ to obtain $[\\bar{S}]$ . The noised slices $[\\bar{S}]$ are then sent to a U-Net based denoising network $F_{dn}$ that outputs the predicted noise conditioned on $I$ : $\\hat{\\epsilon} = F_{dn}([\\bar{S}], I)$ . In practice, the injection and prediction of noise $\\epsilon$ involve an iteration step $t$ , which we omit here for brevity. To better condition $I$ for denoising diffusion, we employ a pre-trained VGG to extract multi-layer features from $I$ and add them into the corresponding layers in the encoder of $F_{dn}$ (see the blue part of Fig. 6). The loss function is the denosing loss: $\\mathcal{L}_{gen} = ||\\hat{\\epsilon} - \\epsilon||_2^2$ . After training, we can sample random noises to generate a set of slice images $\\{\\hat{S}_i^j\\}$ .", + "bbox": [ + 75, + 517, + 467, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to regression-based slicing, generation-based slicing can produce multiple plausible outputs rather than a single deterministic result, but requires much longer inference time due to the nature of diffusion models.", + "bbox": [ + 75, + 700, + 467, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning Implicit Field from Slices", + "text_level": 1, + "bbox": [ + 76, + 770, + 382, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We learn an implicit field from slice images, by projecting a 3D query point onto each slice image, extracting query-specific features, and then aggregating them to predict a signed distance. For regression-based slicing, we train the SDF prediction network on the predicted slices from the previous section thus the whole pipeline can be trained end-to-end. For generation-based slicing, we train the SDF pre", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "diction network on GT slice images since the generated slices may not match the GT signed distances.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for a query point $q \\in \\mathbb{R}^3$ , we project it to a location $\\tilde{q} \\in \\mathbb{R}^2$ on the 2D image plane according to the camera parameters $\\Theta$ . For slice $\\hat{S}_i^j$ (regression) or $S_i^j$ (generation), we first extract its feature maps, then retrieve features from each feature map corresponding to location $\\tilde{q}$ , and finally concatenate them to obtain the image feature vector $V_i^j \\in \\mathbb{R}^{1 \\times 1 \\times N_c}$ . The retrieval is accomplished via bi-linear interpolation over the features grids that encompass the location of $\\tilde{q}$ . We design a transformer network [58] $F_t$ which learns to aggregate the feature vectors from all slices to predict the signed distance $\\hat{d}$ : $\\hat{d} = F_t([q', V_1^X, ..., V_{N_s}^Z])$ , where $q' \\in \\mathbb{R}^{N_c}$ is linearly projected from $q$ to be the same dimension with $V_i^j$ . The optimization objective is to minimize the $\\mathcal{L}_1$ distance between predicted and GT signed distance: $\\mathcal{L}_{imp} = ||\\hat{d} - d||_1$ . In regression-based slicing, $\\mathcal{L}_{imp}$ and $\\mathcal{L}_{reg}$ are jointly optimized.", + "bbox": [ + 496, + 119, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DISN [65] extracts from only the input view $I$ to obtain a feature vector $V$ . For query points located around occluded parts of the object (e.g., the query point in Fig. 3), $V$ is irrelevant to $q$ since it is extracted from the front parts that block $q$ , often leading to erroneous SDF prediction. In contrast, one or more of our slice images could reveal the occluded parts, yielding more accurate features fed to the transformer for SDF prediction. After learning the implicit field, we apply Marching Cubes [32] to obtain triangle meshes. Estimating the extrinsics (camera poses) of $\\Theta$ can output 3D shapes in canonical poses, which is optional and applied only on ShapeNet [5]; See more details in Supplemental.", + "bbox": [ + 496, + 366, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results and Evaluation", + "text_level": 1, + "bbox": [ + 498, + 558, + 718, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We utilize two well-known datasets, ShapeNet [5] and Objaverse [10], to train and test our network. For ShapeNet, our method is applied to Chairs and Airplanes separately, using the train/validation/test split as in SHREC16 [47]. Objaverse is a challenging dataset, where many objects may not belong to a clear category. In our experiment, we randomly selected $\\approx 40\\mathrm{K}$ shapes from Objaverse to keep resource and time costs manageable, while ensuring diversity of the collection. We follow a $95\\% -5\\%$ train-test split over the 40K selected shapes, ensuring that every training model we use is also from the training set of Zero-1-2-3 [29]. Note that Zero-1-2-3 used a train-test split of $99\\% -1\\%$ over the entire 800K models in Objaverse.", + "bbox": [ + 496, + 583, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering Slice Images. We employ the bisect function in Blender [1] to cut an object mesh into slices and then render them. By default, we keep internal cavities and holes produced by slicing. Since many 3D models in Objaverse are not in, or have no, canonical poses and feature random rotations, slicing them along the default $X$ , $Y$ , and $Z$ directions is not suitable. To this end, we opted for camera-aligned slicing instead. Specifically, we set the camera position, in", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9885", + "bbox": [ + 482, + 941, + 514, + 953 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9baad35e8564fd3567393bb8957ef0dac28fd2b4ed4bc73f6f76ad3db8203e8d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative results of single-view 3D reconstruction on ShapeNet Chairs and Airplanes. \\* denotes using ground-truth (GT) camera poses. Slice3D (ours) uses regression-based slicing." + ], + "table_body": "
Datasets MetricsShapeNet ChairsShapeNet Airplanes
CD↓F1↑HD↓CD↓F1↑HD↓
DISN [65]3.252.347.843.482.817.62
Auto-SDF [36]25.00.7715.8---
NeRF-Img [42]19.51.2713.416.01.4114.0
One-2-3-45 [28]12.51.0111.723.40.7512.6
Ours2.802.436.752.992.836.45
DISN* [65]3.022.427.863.192.967.56
SSDNeRF* [6]8.982.0610.910.42.0010.4
Ours*2.672.536.702.812.966.38
", + "bbox": [ + 81, + 88, + 460, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "troduce a random rotation to the object, position it in front of the camera, and utilize the camera's axes for slicing. Under this setting, estimating the rotation of objects is unnecessary; see more details in the supplementary material.", + "bbox": [ + 75, + 294, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. When preparing training data, we use the Vega FEM Library [49] for SDF computation, same as DISN [65] and AutoSDF [36]. For each object, the diagonal length of the object's 3D bounding box is normalized to 1. We set $H$ and $W$ to 128, $N_e$ to 128, $N_s$ to 4, and $N_p$ to 1,024. We train our model on only one Nvidia A40 GPU. Evaluation Metrics. We use Chamfer $\\mathcal{L}_2$ Distance (CD), F-score%1 (F1) [54], and Hausdorff Distance (HD) to evaluate reconstruction results. Besides these 3D metrics, we also evaluate using 2D metrics, PSNR, SSIM [61], and LPIPS [70], on 24 rendered views of reconstructed meshes.", + "bbox": [ + 75, + 354, + 468, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparisons to State Of The Art (SOTA)", + "text_level": 1, + "bbox": [ + 76, + 530, + 431, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with SOTA methods including DISN [65], AutoSDF [36], NeRF-From-Img [42], SSD-NeRF [6], RealFusion [33], and One-2-3-45 [28]. Note that when showing visual results, we always present two views to remove any view bias towards the input.", + "bbox": [ + 75, + 551, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ShapeNet. Figs. 7 and 8 demonstrate the effectiveness of our approach in reconstructing obscured sections of objects while precisely restoring visible ones. In contrast, DISN [65] can barely reconstruct the occluded parts. AutoSDF [36] and NeRF-From-Image [42] rely on latent features for reconstruction, which can yield rather clean results but they do not necessarily respect the input image, as the chair results from Fig. 7 evidently showed. SSDNeRF [6] tends to produce erroneous geometric details and the surfaces are not as smooth as from other methods, especially for the airplane examples in Fig. 8. Generally speaking, One-2-3-45 [28] yields rough and bulky geometries that clearly lack details. Additionally, our method outperforms the other methods in the reconstruction of unconventional inputs, as shown by the two airplanes in Fig. 8.", + "bbox": [ + 75, + 628, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 shows quantitative comparisons to the above methods on the two ShapeNet categories. When our method operates without using GT camera poses and relies on esti", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8a5ef2d94bde0cde0d504fd1906ba4fe2ebae112897b8af1daebabc37a53579a.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Quantitative results of single-view 3D reconstruction on the Objaverse dataset. 'R' and and 'G' denote regression-based and generation-based slicing in Slice3D, respectively." + ], + "table_body": "
MethodCD↓F1↑HD↓PSNR↑SSIM↑LPIPS↓
DISN [65]28.21.4723.021.98.7915.2
One-2-3-45 [28]35.40.8729.819.88.5519.8
Ours (G)25.01.5116.422.38.8413.4
Ours (R)16.61.5313.422.68.8812.8
", + "bbox": [ + 504, + 88, + 890, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mated poses, its performance sees only a marginal degradation. Notably, for Chair, our method with estimated poses even outperforms DISN and SSDNerf with GT poses. It is also notable that One-2-3-45 tends to output dilated shapes. Although it falls behind DISN in quantitative results, it usually generates more complete shapes and more accurate topologies than DISN; see Figs. 7 and 8.", + "bbox": [ + 496, + 219, + 890, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Objaverse. Since AutoSDF [36], NeRF-From-Image [42], and SSDNeRF [6] are all category-specific methods, they cannot be trained on Objaverse which consists of many 3D objects with no clear categories. Thus, we mainly compare our method with DISN [65], RealFusion [33], and One-2-3-45 [28]; see Fig. 9 for a qualitative comparison.", + "bbox": [ + 496, + 324, + 890, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that RealFusion [33] and One-2-3-45 [28] both leverage a significantly larger volume of training data compared to our method, as they both rely on a pre-trained Stable Diffusion [45] model. Despite this disparity in the scale of the training sets, our method still outperforms these two methods, as well as DISN; see Fig. 9. Quantitatively, Tab. 2 demonstrates our method's across-the-board advantage over DISN and One-2-3-45 (more so on Objaverse than on ShapeNet), in both 3D and 2D metrics, especially CD, HD, and LPIPS. These results also show that our method possesses the capacity to generalize across common shapes rather than being confined to specific categories.", + "bbox": [ + 496, + 416, + 890, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world Image Input. To further stress-test the methods, we perform Google image search on object keywords (e.g., \"chair\", \"airplane\", etc.) and take the top-1 image with occlusions as input. In addition, we use some examples from Google Scanned Objects (GSO) [12] dataset. Fig. 10 shows that our method performs the best in recovering occluded parts while preserving fine details, e.g., the turbines of the airplane, the heel, and the opening of the boot.", + "bbox": [ + 496, + 598, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-slice vs. Multi-view Reconstruction. Fig. 11 compares between multi-slice and multi-view reconstructions. We demonstrate two different sets of multi-slices and multi-views generated from our method and One-2-3-45 [28]. Our generated slice images depict two distinct configurations of chair legs, maintaining high consistency across slices. In contrast, the multiple views synthesized in One-2-3-45 do encounter inconsistency issues, compromising an accurate 3D reconstruction. We show three slices in each set and more can be found in the supplementary material.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Inference Costs. Unlike RealFusion [33] and One-2-3-45 [28], our method does not depend on a large", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9886", + "bbox": [ + 482, + 941, + 514, + 953 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee668f36d8e41fa6c5b91a23e97fd7c60e3563ef640863aec0b850e45831ac2f.jpg", + "image_caption": [ + "Figure 7. Visual comparison between single-view 3D reconstruction methods on two ShapeNet chairs. DISN and our method (based on regressive slicing) utilize the same estimated camera parameters. Two different views are displayed to remove view bias." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 885, + 228 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6c6bb659619caf5510ee89c559132ebff623f99fb567beade6977155083b4ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 270, + 883, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/425773c5acaf5dc177562335accb4b537a6e2384cce327a00aa5a6ba65223e67.jpg", + "image_caption": [ + "Figure 8. Visual comparison between single-view 3D reconstruction methods on two somewhat unconventional airplanes from ShapeNet." + ], + "image_footnote": [], + "bbox": [ + 86, + 405, + 883, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c25a9ef23af71e990c63a87a8bccf7cda25dc94b04bae448cc84138ab403f663.jpg", + "image_caption": [ + "Figure 9. Visual comparison between cross-category reconstruction methods on Objverse. Our method better recovers overall geometries, fine details, and is more faithful to the inputs. Except for the seaweed (squared), all the other models are in the training set of Zero-1-to-3.", + "Figure 10. On real-world image inputs. Top two from Google image search and the others from GSO [12]. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 84, + 625, + 464, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pre-trained model [45]. NeRF-based methods tend to have long inference times, as many of them require per-test optimization. Our method has much faster inference times: 10 or 20 seconds with slice regression or generation, com", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pared to 90 minutes for RealFusion and 45 seconds for One-2-3-45 and SSDNeRF, while delivering superior results. With similar inference times to DISN, Auto-SDF, and NeRF-From-Img, our method has better reconstruction quality. See supplementary material for more details.", + "bbox": [ + 496, + 628, + 893, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 720, + 653, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generative vs. Regressive Slicing. As shown in the first row of Fig. 11, for generation-based slicing, we can produce multiple plausible results for a given input view. Tab. 2 shows that regressive slicing outperforms generative slicing when compared to GT. However, it is notable that the 3D outputs may differ from GT when input views have ambiguous occluded parts. Regressive training struggles to handle such one-to-many mappings, leading to generation outperforming regression in reconstructing occluded parts.", + "bbox": [ + 496, + 747, + 893, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Slices. We test the performance of our method", + "bbox": [ + 500, + 885, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9887", + "bbox": [ + 482, + 941, + 514, + 954 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c32affa8f67f8e1d6326c19515f6cb14cac8392d6d72ea628c003e98c11de202.jpg", + "image_caption": [ + "Figure 11. Multi-slice vs. multi-view reconstructions amid ambiguities in the chair legs. Both One-2-3-45 (bottom) and Slice3D (top) can produce multiple results. Our results are both plausible from consistent slices, while One-2-3-45 suffers from multi-view inconsistencies." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 888, + 228 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/819a8da2a30ec7e1d68180c10824a9d910e689769f6814e97688107125be78bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingCD↓F1↑HD↓SettingCD↓F1↑HD↓
One-hot ind.2.682.536.82only X2.882.397.74
DISN+Depth2.992.437.81only Y2.752.456.84
Ns=22.902.467.31only Z3.082.387.60
Ns=42.672.536.70Filled2.732.507.37
Ns=82.562.566.93w/ textures2.822.507.70
", + "bbox": [ + 81, + 282, + 460, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on ShapeNet Chairs.", + "bbox": [ + 133, + 368, + 410, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with regression-based slicing by varying $N_{s}$ between 2, 4, and 8, on ShapeNet Chairs. From Tab. 3, it is evident that larger $N_{s}$ leads to better reconstruction quality. However, $N_{s} = 8$ will dramatically increase the computational cost of our method since its complexity is $O(N_{s}^{2})$ , as pair-wise attentions are performed on the slices in the transformer. As a trade-off between reconstruction quality and computational complexity, we set $N_{s} = 4$ by default.", + "bbox": [ + 75, + 393, + 468, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 also covers several other ablation studies.", + "bbox": [ + 96, + 516, + 413, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learned v.s. One-hot Slice Indicator. Learned indicator codes (the default setting) achieve slightly better performance than using one-hot codes to indicate slice locations. This can be seen by comparing numbers from the first half row of Tab. 3 to those in the last row of Tab. 1.", + "bbox": [ + 75, + 532, + 468, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Slices v.s. Depth Images. Since our slice images offer layered information, it is logical to perform a comparison against depth images as supplementary information. We compared the performance of DISN when concatenating estimated depth images with the input view. Results are shown in the \"DISN+Depth\" half row, which indicates that depth images can only provide marginal improvement.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Three Axes v.s. One Axis. We also investigated the option of slicing along only one axis. Not surprisingly, Tab. 3 shows that this underperforms against using three axes.", + "bbox": [ + 75, + 715, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Solid v.s Hollow Slices. Tab. 3 also demonstrates that not filling the holes improves the performance as more information about the inner/outer surfaces is revealed this way.", + "bbox": [ + 75, + 762, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using Textured Slices. We can optionally keep object textures in the slice images. Experimental results show that removing the textures in the slice images can achieve better reconstruction. The reasons could be: 1) learning to produce slice images without textures could help the model better distinguish geometries from textures; 2) for some 3D", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models the color of inner and outer surfaces could be different which may complicate the learning process.", + "bbox": [ + 498, + 285, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Future Work", + "text_level": 1, + "bbox": [ + 498, + 329, + 764, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our multi-slice idea represents a significant departure from conventional approaches to single-view 3D reconstruction. Through conceptual illustrations (Fig. 5) and extensive experiments, we demonstrate its compelling properties in recovering occluded object structures. Note again that all the Slice3D results on Objaverse [10] were obtained by networks trained on a very small subset (40K or $5\\%$ ) of the 800K 3D models therein, whereas One-2-3-45 [28] utilizes Zero-1-to-3 [29] for multi-view synthesis and the latter was fine-tuned on the entirety of the 800K. Yet our results (Table 2) show that Slice3D outperforms One-2-3-45 across the board, suggesting that our model generalizes better, likely owing to its superior occlusion-revealing abilities.", + "bbox": [ + 496, + 354, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our current implementation only uses four slices per direction, which is quite coarse. Since there is no camera rotation, parts of an object far away from the camera are in smaller scale, e.g., the rear leg of the chair in Fig. 10 and of the buck in Fig. 9. This can hinder our network's ability to extract detailed features. More refined and gradual slicing over those parts may be needed to improve matters.", + "bbox": [ + 496, + 551, + 890, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The foremost limitation of slicing however, is that realistically, it can only be executed at scale on digital 3D models, not physical ones. On the other hand, the multi-view approach can be applied to physical 3D objects by capturing photos only, without ever needing 3D supervision, as in RealFusion [33]. However, this is not without a high cost: 90 minutes by RealFusion vs. 45 seconds by One-2-3-45 (requiring 3D fine-tuning for multi-view synthesis) and less than 20 seconds by Slice3D. We envision synthesizing 3D models, not many but carefully crafted few, for the multislice approach as an intriguing direction to explore.", + "bbox": [ + 496, + 657, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank all the anonymous reviewers and area chairs for their valuable comments. We thank Daniel Cohen-Or and Xingguang Yan for helpful discussions. This work was supported in part by NSERC.", + "bbox": [ + 496, + 838, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9888", + "bbox": [ + 482, + 941, + 514, + 953 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bisect - blender manual. https://docs.blender.org/manual/en/2.80/modeling/meshes/editing/subdividing/bisect.html. Accessed: 2023-11-15.5", + "[2] Kfir Aberman, Oren Katzir, Qiang Zhou, Zegang Luo, Andrei Sharf, Chen Greif, Baoquan Chen, and Daniel Cohen-Or. Dip transform for 3D shape reconstruction. ACM Transactions on Graphics, 36(4):Article 79, 2017. 4", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3", + "[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 3", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5", + "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion NeRF: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714, 2023. 2, 3, 6", + "[7] Wenzheng Chen, Jun Gao, Huan Ling, Edward J. Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in Neural Information Processing Systems, 2019. 2", + "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3", + "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In CVPR, 2020. 1", + "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 3, 5, 8", + "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3", + "[12] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6, 7", + "[13] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 4", + "[14] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3", + "[15] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3", + "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3", + "[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 1", + "[18] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3", + "[19] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 4", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5", + "[21] Chenbo Jiang, Jie Yang, Shwai He, Yu-Kun Lai, and Lin Gao. Neuralslice: Neural 3d triangle mesh reconstruction via slicing 4d tetrahedral meshes. 2023. 4", + "[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 3", + "[24] Shay Kels and Nira Dyn. Reconstruction of 3d objects from 2d cross-sections with the 4-point subdivision scheme adapted to sets. Computers & Graphics, 35:741-746, 2011. 4", + "[25] Manyi Li and Hao Zhang. $\\mathrm{D}^2\\mathrm{IM}$ -Net: Learning detail disentangled implicit fields from single images. In CVPR, 2021. 1, 3", + "[26] Wallace Lira, Johannes Merz, Daniel Ritchies, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop gan for unsupervised image-to-image translation. In ECCV, 2020. 3", + "[27] Lu Liu, Chandrajit Bajaj, Joseph O. Deasy, Daniel A. Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. Computer Graphics Forum, 27:155-163, 2008. 4", + "[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 1, 2, 3, 6, 8" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9889", + "bbox": [ + 482, + 941, + 516, + 953 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 1, 3, 5, 8", + "[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image, 2023. 3, 4", + "[31] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023.4", + "[32] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 5", + "[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 3, 6, 8", + "[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 3", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3, 6", + "[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3", + "[38] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2", + "[39] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 3", + "[40] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 4", + "[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165–174, 2019. 3", + "[42] Dario Pavllo, David Joseph Tan, Marie-Julie Rakotosaona, and Federico Tombari. Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In CVPR, pages 4391-4401, 2023. 2, 3, 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 3, 4", + "[44] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3d geometry via nested shape layers. In CVPR, 2018. 1", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3, 6, 7", + "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4", + "[47] Manolis Savva, Fisher Yu, Hao Su, M Aono, B Chen, D Cohen-Or, W Deng, Hang Su, Song Bai, Xiang Bai, et al. Shrec16 track: largescale 3d shape retrieval from shapenet core55. In Proceedings of the eurographics workshop on 3D object retrieval, 2016. 5", + "[48] Haim Sawdayee, Amir Vaxman, and Amit H. Bermano. Orex: Object reconstruction from planar cross-sections using neural fields. In CVPR, pages 20854-20862, 2023. 4", + "[49] Fun Shing Sin, Daniel Schroeder, and Jernej Barbič. Vega: non-linear fem deformable object simulator. In Computer Graphics Forum, pages 36-48. Wiley Online Library, 2013. 6", + "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[51] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 3", + "[52] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 4", + "[53] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 1", + "[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, pages 3405-3414, 2019. 6", + "[55] Maxim Tatarchenko, Stephan R. Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2", + "[56] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9890", + "bbox": [ + 482, + 941, + 514, + 953 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 4", + "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3", + "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5", + "[59] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, pages 52-67, 2018. 3", + "[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3", + "[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[62] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3", + "[63] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Comput. Graph. Forum, 2022. 3", + "[64] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, pages 4479–4489, 2023. 3", + "[65] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 1, 3, 5, 6", + "[66] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In CVPR, pages 6239–6249, 2022. 3", + "[67] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3", + "[68] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3", + "[69] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 4", + "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "features as a perceptual metric. In CVPR, pages 586-595, 2018. 4, 6", + "[71] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023.4", + "[72] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 4" + ], + "bbox": [ + 501, + 92, + 890, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9891", + "bbox": [ + 482, + 941, + 513, + 953 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_model.json b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_model.json new file mode 100644 index 0000000000000000000000000000000000000000..96f92a965395ee3aee58fc6f660cb2718f08985e --- /dev/null +++ b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.131, + 0.856, + 0.152 + ], + "angle": 0, + "content": "Slice3D: Multi-Slice, Occlusion-Revealing, Single View 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.18, + 0.79, + 0.218 + ], + "angle": 0, + "content": "Yizhi Wang\\(^{1}\\), Wallace Lira\\(^{1}\\), Wenqi Wang\\(^{2}\\), Ali Mahdavi-Amiri\\(^{1}\\), Hao Zhang\\(^{1}\\) \n\\(^{1}\\)Simon Fraser University \\(^{2}\\)Tsinghua University" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.226, + 0.875, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.893, + 0.437 + ], + "angle": 0, + "content": "Figure 1. Our single-view 3D reconstruction method, Slice3D, predicts multi-slice images to reveal occluded parts without changing the camera (in contrast to multi-view synthesis), and then lifts the slices into a 3D model. This leads to better recovery of occluded parts than both encoder-decoder based methods such as DISN [65] and 3D reconstructions from diffusion-synthesized multi-view images (via Zero-1-to-3 [29]) such as One-2-3-45 [28], where the key challenge of the latter is multi-view inconsistency, as shown on the right." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.449, + 0.314, + 0.465 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.474, + 0.884 + ], + "angle": 0, + "content": "We introduce multi-slice reasoning, a new notion for single-view 3D reconstruction which challenges the current and prevailing belief that multi-view synthesis is the most natural conduit between single-view and 3D. Our key observation is that object slicing is a more direct, and hence more advantageous, means to reveal occluded structures than altering camera views. Specifically, slicing can peel through any occluder without obstruction, and in the limit (i.e., with infinitely many slices), it is guaranteed to unveil all hidden object parts. We realize our idea by developing Slice3D, a novel method for single-view 3D reconstruction which first predicts multi-slice images from a single RGB input image and then integrates the slices into a 3D model using a coordinate-based transformer network to product a signed distance function. The slice images can be regressed or generated, both through a U-Net based network. For the former, we inject a learnable slice indicator code to designate each decoded image into a spatial slice location, while the slice generator is a denoising diffusion model operating on the entirety of slice images stacked on the input channels. We conduct extensive evaluation against state-of-the-art alternatives to demonstrate superiority of our method, especially in recovering complex and severely occluded shape structures, amid ambiguities. All Slice3D results were produced by networks trained on a single Nvidia A40 GPU, with an inference time of less than 20 seconds." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.449, + 0.631, + 0.465 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Single-view 3D reconstruction has been one of the extensively studied problems in computer vision. Despite the rapid advances on learning-based approaches to tackle this problem, including the latest attempts to leverage large foundational models and powerful diffusion-based generators, the most fundamental challenge still remains: how to faithfully reconstruct occluded parts from just one view?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.893, + 0.661 + ], + "angle": 0, + "content": "Conventional encoder-decoder based methods, e.g., [9, 17, 25, 44, 53, 65], rely on direct 3D supervision to learn mappings from global or local image features to 3D, but such features from just a single view often do not have the capacity to handle severe occlusion. Also, recent evi" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.689, + 0.89, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.799, + 0.892, + 0.897 + ], + "angle": 0, + "content": "Figure 2. Multi-view vs. multi-slice for occlusion revelation, with orange edges showing revealed parts and black edges showing occluded parts. In the limit, multi-view (i.e., infinitely many views) may still leave undisclosed parts (a), e.g., the bulb, while infinitely many slices would guarantee complete structure revelation (b). With limited slices (c) and views (d), four each, the multi-slice approach tends to reveal more shape structures." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.514, + 0.955 + ], + "angle": 0, + "content": "9881" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.48, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.882, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.368, + 0.896, + 0.482 + ], + "angle": 0, + "content": "Figure 3. Slice3D for single-view 3D reconstruction from multi-slice images. The example shows 4 slice images per \\(X\\), \\(Y\\), \\(Z\\) direction, which can be obtained by either a regressor or a generator. Regression is carried out by a U-Net encoder-decoder, where the decoder takes an extra slice indicator code that is appended to the encoder output, to obtain a slice image. There is one indicator code for each of the \\(3 \\times 4 = 12\\) slices, that is optimized during training. Our generator is a denoising diffusion model which produces a concatenation of slice images following the distribution of the ground-truth (GT) images, when conditioned on the single-view input. Inference from random noises may output multiple plausible results. The 3D reconstruction stage inputs the predicted slice images and produces an implicit field. We first extract hierarchical CNN features from the input slices, then project a 3D query point onto the 2D planes of the feature maps, and finally gather all the query-related features and send them into a transformer [58] to predict the signed distance at the query point." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.497, + 0.436, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.695, + 0.471, + 0.792 + ], + "angle": 0, + "content": "Figure 4. Our slice image prediction networks are trained using occlusion-revealing multi-slice images produced from a GT 3D object mesh along a predetermined set of slicing directions (the \\(X\\), \\(Y\\), \\(Z\\) axes of the object bounding box in the example above), while keeping the hollows as holes in the images, through which the back faces of the mesh can be visible. The 3D mesh may be textured, but the slice images are all rendered without texture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.9 + ], + "angle": 0, + "content": "dence suggests that many of these methods are essentially performing image classification or retrieval [55]. As such, occlusion recovery is by and large a \"hit or miss\". Differentiable rendering enables 3D reasoning from images without 3D supervision [7, 38], but image-space rendering losses alone do not account for occluded structures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Recently, more and more works model the reconstruction problem as a single-view conditioned generative task, aiming to learn a distribution of unseen structures. Typically, multi-view images are generated from single view via diffusion. This is followed by multi-view reconstruction, e.g., via neural radiance fields (NeRF) [6, 28, 42], to obtain 3D outputs. However, multi-view consistency is a reoccurring challenge (see Fig. 1), especially over disparate views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.759 + ], + "angle": 0, + "content": "We introduce a novel approach to single-view 3D reconstruction which is occlusion-revealing (Fig. 2) by first predicting multi-slice images from a single-view input and then integrating the slices into a 3D model; see Fig. 3. Fig. 4 shows slice images provided during training along a set of slicing directions. Each slice represents the projected image, in the input camera view, of one slice of the 3D mesh model. For the TV example shown, a completely occluded leg, in the camera view, is revealed in the slice images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.868 + ], + "angle": 0, + "content": "The motivations behind our multi-slice reconstruction are three-fold. First, slicing through a volume or 3D mesh is a natural way to reveal occluded parts. Second, with the slice images forming an intermediate representation, our approach breaks the difficult problem of mapping from single view to 3D into two comparatively simpler tasks: slice prediction followed by 3D reconstruction from slices." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "More importantly, when contrasting our multi-slice representation with the popular multi-view approach for occlu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.516, + 0.954 + ], + "angle": 0, + "content": "9882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.45, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.192, + 0.473, + 0.29 + ], + "angle": 0, + "content": "Figure 5. Singe-view reconstruction of severely occluded parts in a lamp (e.g., the bulb) by Slice3D vs. One-2-3-45 [28], a state-of-the-art multi-view approach. The multi-view images in [28] were obtained by Zero-1-to-3 [29], which was trained on renderings of 800K 3D models [10]. It is conceivable that with a limited number of views (12, the same as our slice count), the bulbs were barely visible due to occlusion by the lamp shades and bases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.47, + 0.51 + ], + "angle": 0, + "content": "sion revelation, we observe that slicing in the limit (i.e., with infinitely many slices) is guaranteed to unveil all hidden structures, while the same cannot be said about multi-view captures, as illustrated in Fig. 2 (a-b), for a 2D lamp example. In practice, when slice and view counts are both limited, the slicing approach is more occlusion-revealing in general since it can peel through any occluders without any obstruction, unlike the multi-view case; see (c-d). On the technical front, neither our slice prediction step nor the 3D reconstruction step is concerned with camera view changes or multi-view consistency, which facilitates the use of convolutional features. In Fig. 5, we show the above contrasting characteristics of the multi-slice vs. multi-view approaches manifested in a real 3D lamp reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.692 + ], + "angle": 0, + "content": "As outlined in Fig. 3, our slice images can be produced by either a regressive network or a generative one, both employing a U-Net architecture, with network losses measured against ground-truth (GT) images obtained by slicing 3D models from the training set; see Fig. 4. In the regressive network, we inject a learnable slice indicator code into the encoder-decoder setup to designate each decoded image into the proper spatial slice location. The slice generator is a denoising diffusion model operating on separate channels for different slice images. In the second stage, 3D reconstruction from the obtained slice images, we design a coordinate-based transformer to obtain an implicit field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Our network is coined Slice3D. We conduct extensive experiments on ShapeNet [5] and Objverse [10] for evaluation. Quantitative and qualitative comparisons are made to representative methods including encoder-decoders capturing local image features [65], autoregressive latent models [36], generative model inversion [42], and three recent methods with diffusion + NeRF [6, 28, 33], demonstrating superiority of our method in terms of reconstruction quality, generalizability, and handling of ambiguities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.216, + 0.86 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The earliest inspiration for our multi-slice 3D reconstruction idea was GANHopper by Lira et al. [26]. In the con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "text of unsupervised image-to-image translation, the key idea of their work is to gradually execute the translation between disparate domains by making small \"hops,\" rather than completing it in one go. We were then motivated to tackle the single-view 3D reconstruction problem as a domain translation task, hopping through the multi-slice images so as to bridge the gap between 2D and 3D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.893, + 0.382 + ], + "angle": 0, + "content": "Neural fields. Neural implicit functions [63] are learned by a neural network to predict such properties as color, occupancy, opacity, and surface normals. Earlier works such as IM-Net [8], occupancy networks [34], and DeepSDF [41] represent a 3D shape as a latent code and send it to a decoder formed by MLPs to predict occupancy or SDF values for any query point sampled in 3D. NeRF [35] and its variants [15] represent a 3D scene using a neural radiance field for novel view synthesis. Many works on neural 3D reconstruction [15, 63] take on single- or multi-view images as well as point clouds as input. In contrast, the reconstruction module in Slice3D is built on multi-slice images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.656 + ], + "angle": 0, + "content": "Single-view reconstruction from image features. An early work to exploit neural image features for 3D reconstruction was Pixel2Mesh [59], which uses a graph representation. Subsequently, neural implicit fields have been widely adopted for single-view reconstruction. IM-Net [8] first extracts a global latent code, from an input view, to learn a 3D implicit field. Auto-SDF [36] and ShapeFormer [66] employ vector quantization (VQ) [57] to improve the representation of latent features. More recent works [22, 37, 68] consider single-view reconstruction as a 3D generation task conditioned on the latent code of the input view. However, a global latent code is insufficient to capture finer details on a 3D object, often leading to overly smoothed or retrieved outcomes that do not faithfully represent the input. To alleviate this problem, both DISN [65] and \\(\\mathrm{D}^2\\mathrm{IM}\\) -Net [25] opted to learn local images features, along with global ones, to improve the recovery of fine-grained topological (e.g., thin) features and surface details, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.901 + ], + "angle": 0, + "content": "3D reconstruction via multi-view. Conventional 3D reconstruction approaches [14] involve multi-view reasoning. With recent advances in generative AI, a prevailing practice in single-view 3D reconstruction is to first employ generative models such as GANs [16, 23] and diffusion [11, 20, 45, 50] to generate several novel views from the input view, and then feed them to a multi-view 3D reconstruction module such as NeRF. NeRF-based surface reconstruction [39, 60, 67] can be further employed to output 3D meshes. EG3D [3] and NeRF-from-Image [42] build NeRFs or its variants upon the outputs of a style-GAN [23], while more recent works [4, 6, 18, 28-30, 33, 42, 43, 51, 62, 64] are favoring diffusion models such as Stable Diffusion (SD) [45]. To date, Zero-1-to-3 [29] represents the state of the art on single- to multi-view synthesis; it is built on SD and fine-tuned by 800K 3D models from Objaverse [10]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.516, + 0.954 + ], + "angle": 0, + "content": "9883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "To alleviate the multi-view consistency problem, which is an reoccurring challenge for the image-space generative methods, recent methods [30, 31, 43, 52, 71] aim to enhance the consistency by performing spatial attention between different views, which dramatically increases the computational overhead. Nevertheless, the consistency of two disparate views is still far from satisfactory. A deeper reason is that standard convolutional neural networks (CNNs) are not inherently rotation-invariant, making it difficult to build correspondences between significantly different views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.488 + ], + "angle": 0, + "content": "Slice representation. There have been techniques to use object slices to reconstruct a 3D shape, e.g., as inspired by Magnetic Resonance Imaging (MRI). Early works in computer graphics employ either parallel cross-sectional curves, e.g., [24], or non-parallel curve networks, e.g., [27], as inputs. Recently, with the advent of machine learning, OREx [48] and Cut-and-Approximate [40] develop neural reconstruction models to obtain 3D meshes, also from a sparse collection of 2D planar cross-sections, while Neural-Slice [21] uses slicing to obtain 3D shapes from 4D tetrahedral meshes. However, none of these techniques, nor others that we are aware of, address the problem of single-view 3D reconstruction since their inputs all consist of 3D data such as curve slices or point clouds. Moreover, our method learns to predict much fewer slice images and use them as input to a transformer-based reconstruction module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.472, + 0.613 + ], + "angle": 0, + "content": "Dip. Dip transform [2] is an ingenious way to acquire occluded 3D structures by measuring fluid displacement when an object is repeatedly oriented and dipped into a fluid. Similar to slicing, Dip was also designed to \"reach\" occluded parts that optical captures via lines of sight cannot. However, both the occlusion revelation (liquid penetration vs. slicing) and target applications (acquisition vs. single-view reconstruction) are completely different." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.473, + 0.814 + ], + "angle": 0, + "content": "Multi-plane image (MPI). Introduced by Zhou et al. [72], an MPI representation for a scene consists of a set of frontoparallel planes at fixed depths, with each image intended to capture scene appearance and visibility at a specific depth. Conceptually, both MPs and our slice images provide a layered scene representation for occlusion revelation. However, MPs [13, 19, 56, 69, 72] have only been employed for novel view synthesis, not 3D reconstruction, and the prediction networks were trained on multi-view images, e.g., from Youtube videos [72]. The generated MPs only reveal a small amount of occluded scenes to account for view shifts. In contrast, our method directly cuts through a 3D object to reveal all the occluded parts at several depth ranges." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.829, + 0.17, + 0.845 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Our method is based on a slicing representation; see Section 3.1 and Fig. 4. Given a single-view RGB image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\) of a 3D object, we first learn a slicing network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "(Section 3.2) which takes \\( I \\) as input and outputs its corresponding multi-slice images. Then, an SDF prediction network is trained (Section 3.3) to leverage image features from the slices to produce an implicit field of the object." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.161, + 0.71, + 0.178 + ], + "angle": 0, + "content": "3.1. Slicing Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.895, + 0.397 + ], + "angle": 0, + "content": "As shown in Fig. 4, given a 3D object mesh, we slice it along the \\(X\\), \\(Y\\), and \\(Z\\) axes of its bounding box and render the sliced segments. The slice images can optionally include textures or have holes filled. By default, we remove the textures and keep the holes, which could reveal back faces of the mesh, and overall, it produces the best experimental results. The rendered slice images consist of \\(\\{S_i^j | 1 \\leq i \\leq N_s, j \\in \\{X,Y,Z\\}\\}\\), where \\(S_i^j \\in \\mathbb{R}^{H \\times W \\times 1}\\), \\(j\\) denotes the axis and \\(N_s\\) is the number of slices along each axis. In our current implementation, we produce \\(N_s = 4\\) slices per direction. Compared to planar cross-sections or tomography scans which consist of densely stacked planar images, our slicing representation is more compact with each slice image revealing more object semantics." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.404, + 0.787, + 0.421 + ], + "angle": 0, + "content": "3.2. Slice Regression and Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.895, + 0.565 + ], + "angle": 0, + "content": "Given an image \\(I\\), observing a 3D shape from a single view, we aim to train a slicing network \\(F_{s}\\) that can produce the slice images from \\(I\\): \\(F_{s}(I) = \\{\\hat{S}_{i}^{j}\\}\\). We have two different designs for such a network: regressive and generative. The former aims to regress GT slices, assuming deterministic occluded parts based on the partial observation of the input view. The latter aims to generate slices that are consistent with the input view, considering various possible occluded parts that do not necessarily match the GT slices." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.581, + 0.727, + 0.597 + ], + "angle": 0, + "content": "3.2.1 Regression-based Slicing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.698 + ], + "angle": 0, + "content": "We employ a U-Net [46] encoder-decoder to regress the slices from \\(I\\). As shown in the top-left of Fig. 3, an indicator code \\(E_{i}^{j}\\in \\mathbb{R}^{N_{e}}\\) is concatenated with the latent code of \\(I\\) to indicate which slice to output: \\(\\hat{S}_i^j = F_d([F_e(I);E_i^j])\\), where \\(F_{e}\\) and \\(F_{d}\\) are the encoder and decoder of \\(F_{s}\\), respectively, and \\([\\cdot ]\\) denotes concatenation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.759 + ], + "angle": 0, + "content": "The indicator codes \\(\\{E_i^j\\}\\) are randomly initialized and jointly optimized with the whole network. The loss function \\(\\mathcal{L}_{reg}\\) is the reconstruction error measured by the \\(\\mathcal{L}_1\\)-Norm and perceptual loss [70] \\((\\mathcal{L}_p)\\) for all slices:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.768, + 0.892, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\frac {\\sum_ {i , j} \\left| \\left| \\hat {S} _ {i} ^ {j} - S _ {i} ^ {j} \\right| \\right| _ {1} + \\mathcal {L} _ {p} \\left(\\hat {S} _ {i} ^ {j} , S _ {i} ^ {j}\\right)}{| S |}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.814, + 0.819, + 0.83 + ], + "angle": 0, + "content": "where \\( |S| = 3 \\cdot N_{s} \\) is the total number of slices." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.731, + 0.864 + ], + "angle": 0, + "content": "3.2.2 Generation-based Slicing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Single-view occlusion revelation is clearly ill-posed, with potential ambiguities leading to multiple plausible 3D re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.516, + 0.954 + ], + "angle": 0, + "content": "9884" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.468, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.47, + 0.297 + ], + "angle": 0, + "content": "Figure 6. The pipeline of generation-based slicing via DDPM. \\(\\bigoplus\\) denotes the injection of noise to slice images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.468, + 0.366 + ], + "angle": 0, + "content": "constructions. To this end, we consider slicing as an image-conditioned generation task and develop a generation-based slicing via Denoising Diffusion Probabilistic Models (DDPM) [20], whose detailed pipeline is shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.468, + 0.517 + ], + "angle": 0, + "content": "The key challenge of applying DDPM lies in how to generate consistent slices. Rather than diffusing each single slice, we propose to concatenate all slices into an aggregated structure \\([S]\\), along either the color dimension \\(([S] \\in \\mathbb{R}^{H \\times W \\times |S|})\\) or a spatial dimension such as height \\(([S] \\in \\mathbb{R}^{(|S| \\cdot H) \\times W \\times 1})\\). The concatenation of slices instructs DDPM to learn the correlation of slices, yielding consistent generation results. By default, we concatenate along the color dimension. The comparison of these two concatenation ways can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.468, + 0.7 + ], + "angle": 0, + "content": "Following DDPM, we first inject a noise \\(\\epsilon \\sim \\mathcal{N}(0,1)\\) to \\([S]\\) to obtain \\([\\bar{S}]\\). The noised slices \\([\\bar{S}]\\) are then sent to a U-Net based denoising network \\(F_{dn}\\) that outputs the predicted noise conditioned on \\(I\\): \\(\\hat{\\epsilon} = F_{dn}([\\bar{S}], I)\\). In practice, the injection and prediction of noise \\(\\epsilon\\) involve an iteration step \\(t\\), which we omit here for brevity. To better condition \\(I\\) for denoising diffusion, we employ a pre-trained VGG to extract multi-layer features from \\(I\\) and add them into the corresponding layers in the encoder of \\(F_{dn}\\) (see the blue part of Fig. 6). The loss function is the denosing loss: \\(\\mathcal{L}_{gen} = ||\\hat{\\epsilon} - \\epsilon||_2^2\\). After training, we can sample random noises to generate a set of slice images \\(\\{\\hat{S}_i^j\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.468, + 0.76 + ], + "angle": 0, + "content": "Compared to regression-based slicing, generation-based slicing can produce multiple plausible outputs rather than a single deterministic result, but requires much longer inference time due to the nature of diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.383, + 0.788 + ], + "angle": 0, + "content": "3.3. Learning Implicit Field from Slices" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We learn an implicit field from slice images, by projecting a 3D query point onto each slice image, extracting query-specific features, and then aggregating them to predict a signed distance. For regression-based slicing, we train the SDF prediction network on the predicted slices from the previous section thus the whole pipeline can be trained end-to-end. For generation-based slicing, we train the SDF pre" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "diction network on GT slice images since the generated slices may not match the GT signed distances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Specifically, for a query point \\(q \\in \\mathbb{R}^3\\), we project it to a location \\(\\tilde{q} \\in \\mathbb{R}^2\\) on the 2D image plane according to the camera parameters \\(\\Theta\\). For slice \\(\\hat{S}_i^j\\) (regression) or \\(S_i^j\\) (generation), we first extract its feature maps, then retrieve features from each feature map corresponding to location \\(\\tilde{q}\\), and finally concatenate them to obtain the image feature vector \\(V_i^j \\in \\mathbb{R}^{1 \\times 1 \\times N_c}\\). The retrieval is accomplished via bi-linear interpolation over the features grids that encompass the location of \\(\\tilde{q}\\). We design a transformer network [58] \\(F_t\\) which learns to aggregate the feature vectors from all slices to predict the signed distance \\(\\hat{d}\\): \\(\\hat{d} = F_t([q', V_1^X, ..., V_{N_s}^Z])\\), where \\(q' \\in \\mathbb{R}^{N_c}\\) is linearly projected from \\(q\\) to be the same dimension with \\(V_i^j\\). The optimization objective is to minimize the \\(\\mathcal{L}_1\\) distance between predicted and GT signed distance: \\(\\mathcal{L}_{imp} = ||\\hat{d} - d||_1\\). In regression-based slicing, \\(\\mathcal{L}_{imp}\\) and \\(\\mathcal{L}_{reg}\\) are jointly optimized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.548 + ], + "angle": 0, + "content": "DISN [65] extracts from only the input view \\(I\\) to obtain a feature vector \\(V\\). For query points located around occluded parts of the object (e.g., the query point in Fig. 3), \\(V\\) is irrelevant to \\(q\\) since it is extracted from the front parts that block \\(q\\), often leading to erroneous SDF prediction. In contrast, one or more of our slice images could reveal the occluded parts, yielding more accurate features fed to the transformer for SDF prediction. After learning the implicit field, we apply Marching Cubes [32] to obtain triangle meshes. Estimating the extrinsics (camera poses) of \\(\\Theta\\) can output 3D shapes in canonical poses, which is optional and applied only on ShapeNet [5]; See more details in Supplemental." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.559, + 0.719, + 0.574 + ], + "angle": 0, + "content": "4. Results and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Dataset. We utilize two well-known datasets, ShapeNet [5] and Objaverse [10], to train and test our network. For ShapeNet, our method is applied to Chairs and Airplanes separately, using the train/validation/test split as in SHREC16 [47]. Objaverse is a challenging dataset, where many objects may not belong to a clear category. In our experiment, we randomly selected \\(\\approx 40\\mathrm{K}\\) shapes from Objaverse to keep resource and time costs manageable, while ensuring diversity of the collection. We follow a \\(95\\% -5\\%\\) train-test split over the 40K selected shapes, ensuring that every training model we use is also from the training set of Zero-1-2-3 [29]. Note that Zero-1-2-3 used a train-test split of \\(99\\% -1\\%\\) over the entire 800K models in Objaverse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Rendering Slice Images. We employ the bisect function in Blender [1] to cut an object mesh into slices and then render them. By default, we keep internal cavities and holes produced by slicing. Since many 3D models in Objaverse are not in, or have no, canonical poses and feature random rotations, slicing them along the default \\(X\\), \\(Y\\), and \\(Z\\) directions is not suitable. To this end, we opted for camera-aligned slicing instead. Specifically, we set the camera position, in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.515, + 0.954 + ], + "angle": 0, + "content": "9885" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.462, + 0.242 + ], + "angle": 0, + "content": "
Datasets MetricsShapeNet ChairsShapeNet Airplanes
CD↓F1↑HD↓CD↓F1↑HD↓
DISN [65]3.252.347.843.482.817.62
Auto-SDF [36]25.00.7715.8---
NeRF-Img [42]19.51.2713.416.01.4114.0
One-2-3-45 [28]12.51.0111.723.40.7512.6
Ours2.802.436.752.992.836.45
DISN* [65]3.022.427.863.192.967.56
SSDNeRF* [6]8.982.0610.910.42.0010.4
Ours*2.672.536.702.812.966.38
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.245, + 0.47, + 0.288 + ], + "angle": 0, + "content": "Table 1. Quantitative results of single-view 3D reconstruction on ShapeNet Chairs and Airplanes. \\* denotes using ground-truth (GT) camera poses. Slice3D (ours) uses regression-based slicing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.469, + 0.356 + ], + "angle": 0, + "content": "troduce a random rotation to the object, position it in front of the camera, and utilize the camera's axes for slicing. Under this setting, estimating the rotation of objects is unnecessary; see more details in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Implementation Details. When preparing training data, we use the Vega FEM Library [49] for SDF computation, same as DISN [65] and AutoSDF [36]. For each object, the diagonal length of the object's 3D bounding box is normalized to 1. We set \\( H \\) and \\( W \\) to 128, \\( N_e \\) to 128, \\( N_s \\) to 4, and \\( N_p \\) to 1,024. We train our model on only one Nvidia A40 GPU. Evaluation Metrics. We use Chamfer \\( \\mathcal{L}_2 \\) Distance (CD), F-score%1 (F1) [54], and Hausdorff Distance (HD) to evaluate reconstruction results. Besides these 3D metrics, we also evaluate using 2D metrics, PSNR, SSIM [61], and LPIPS [70], on 24 rendered views of reconstructed meshes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.433, + 0.547 + ], + "angle": 0, + "content": "4.1. Comparisons to State Of The Art (SOTA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.629 + ], + "angle": 0, + "content": "We compare our method with SOTA methods including DISN [65], AutoSDF [36], NeRF-From-Img [42], SSD-NeRF [6], RealFusion [33], and One-2-3-45 [28]. Note that when showing visual results, we always present two views to remove any view bias towards the input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.856 + ], + "angle": 0, + "content": "ShapeNet. Figs. 7 and 8 demonstrate the effectiveness of our approach in reconstructing obscured sections of objects while precisely restoring visible ones. In contrast, DISN [65] can barely reconstruct the occluded parts. AutoSDF [36] and NeRF-From-Image [42] rely on latent features for reconstruction, which can yield rather clean results but they do not necessarily respect the input image, as the chair results from Fig. 7 evidently showed. SSDNeRF [6] tends to produce erroneous geometric details and the surfaces are not as smooth as from other methods, especially for the airplane examples in Fig. 8. Generally speaking, One-2-3-45 [28] yields rough and bulky geometries that clearly lack details. Additionally, our method outperforms the other methods in the reconstruction of unconventional inputs, as shown by the two airplanes in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 shows quantitative comparisons to the above methods on the two ShapeNet categories. When our method operates without using GT camera poses and relies on esti" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.164 + ], + "angle": 0, + "content": "
MethodCD↓F1↑HD↓PSNR↑SSIM↑LPIPS↓
DISN [65]28.21.4723.021.98.7915.2
One-2-3-45 [28]35.40.8729.819.88.5519.8
Ours (G)25.01.5116.422.38.8413.4
Ours (R)16.61.5313.422.68.8812.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.169, + 0.892, + 0.211 + ], + "angle": 0, + "content": "Table 2. Quantitative results of single-view 3D reconstruction on the Objaverse dataset. 'R' and and 'G' denote regression-based and generation-based slicing in Slice3D, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.325 + ], + "angle": 0, + "content": "mated poses, its performance sees only a marginal degradation. Notably, for Chair, our method with estimated poses even outperforms DISN and SSDNerf with GT poses. It is also notable that One-2-3-45 tends to output dilated shapes. Although it falls behind DISN in quantitative results, it usually generates more complete shapes and more accurate topologies than DISN; see Figs. 7 and 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Objaverse. Since AutoSDF [36], NeRF-From-Image [42], and SSDNeRF [6] are all category-specific methods, they cannot be trained on Objaverse which consists of many 3D objects with no clear categories. Thus, we mainly compare our method with DISN [65], RealFusion [33], and One-2-3-45 [28]; see Fig. 9 for a qualitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Note that RealFusion [33] and One-2-3-45 [28] both leverage a significantly larger volume of training data compared to our method, as they both rely on a pre-trained Stable Diffusion [45] model. Despite this disparity in the scale of the training sets, our method still outperforms these two methods, as well as DISN; see Fig. 9. Quantitatively, Tab. 2 demonstrates our method's across-the-board advantage over DISN and One-2-3-45 (more so on Objaverse than on ShapeNet), in both 3D and 2D metrics, especially CD, HD, and LPIPS. These results also show that our method possesses the capacity to generalize across common shapes rather than being confined to specific categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Real-world Image Input. To further stress-test the methods, we perform Google image search on object keywords (e.g., \"chair\", \"airplane\", etc.) and take the top-1 image with occlusions as input. In addition, we use some examples from Google Scanned Objects (GSO) [12] dataset. Fig. 10 shows that our method performs the best in recovering occluded parts while preserving fine details, e.g., the turbines of the airplane, the heel, and the opening of the boot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Multi-slice vs. Multi-view Reconstruction. Fig. 11 compares between multi-slice and multi-view reconstructions. We demonstrate two different sets of multi-slices and multi-views generated from our method and One-2-3-45 [28]. Our generated slice images depict two distinct configurations of chair legs, maintaining high consistency across slices. In contrast, the multiple views synthesized in One-2-3-45 do encounter inconsistency issues, compromising an accurate 3D reconstruction. We show three slices in each set and more can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training and Inference Costs. Unlike RealFusion [33] and One-2-3-45 [28], our method does not depend on a large" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.516, + 0.954 + ], + "angle": 0, + "content": "9886" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.887, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.895, + 0.266 + ], + "angle": 0, + "content": "Figure 7. Visual comparison between single-view 3D reconstruction methods on two ShapeNet chairs. DISN and our method (based on regressive slicing) utilize the same estimated camera parameters. Two different views are displayed to remove view bias." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.271, + 0.885, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.387, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Figure 8. Visual comparison between single-view 3D reconstruction methods on two somewhat unconventional airplanes from ShapeNet." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.406, + 0.885, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.893, + 0.611 + ], + "angle": 0, + "content": "Figure 9. Visual comparison between cross-category reconstruction methods on Objverse. Our method better recovers overall geometries, fine details, and is more faithful to the inputs. Except for the seaweed (squared), all the other models are in the training set of Zero-1-to-3." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.626, + 0.465, + 0.792 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.8, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Figure 10. On real-world image inputs. Top two from Google image search and the others from GSO [12]. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "pre-trained model [45]. NeRF-based methods tend to have long inference times, as many of them require per-test optimization. Our method has much faster inference times: 10 or 20 seconds with slice regression or generation, com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.894, + 0.707 + ], + "angle": 0, + "content": "pared to 90 minutes for RealFusion and 45 seconds for One-2-3-45 and SSDNeRF, while delivering superior results. With similar inference times to DISN, Auto-SDF, and NeRF-From-Img, our method has better reconstruction quality. See supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.722, + 0.655, + 0.739 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.894, + 0.884 + ], + "angle": 0, + "content": "Generative vs. Regressive Slicing. As shown in the first row of Fig. 11, for generation-based slicing, we can produce multiple plausible results for a given input view. Tab. 2 shows that regressive slicing outperforms generative slicing when compared to GT. However, it is notable that the 3D outputs may differ from GT when input views have ambiguous occluded parts. Regressive training struggles to handle such one-to-many mappings, leading to generation outperforming regression in reconstructing occluded parts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Number of Slices. We test the performance of our method" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.516, + 0.955 + ], + "angle": 0, + "content": "9887" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.889, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Figure 11. Multi-slice vs. multi-view reconstructions amid ambiguities in the chair legs. Both One-2-3-45 (bottom) and Slice3D (top) can produce multiple results. Our results are both plausible from consistent slices, while One-2-3-45 suffers from multi-view inconsistencies." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.284, + 0.462, + 0.366 + ], + "angle": 0, + "content": "
SettingCD↓F1↑HD↓SettingCD↓F1↑HD↓
One-hot ind.2.682.536.82only X2.882.397.74
DISN+Depth2.992.437.81only Y2.752.456.84
Ns=22.902.467.31only Z3.082.387.60
Ns=42.672.536.70Filled2.732.507.37
Ns=82.562.566.93w/ textures2.822.507.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.369, + 0.411, + 0.383 + ], + "angle": 0, + "content": "Table 3. Ablation studies on ShapeNet Chairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.516 + ], + "angle": 0, + "content": "with regression-based slicing by varying \\( N_{s} \\) between 2, 4, and 8, on ShapeNet Chairs. From Tab. 3, it is evident that larger \\( N_{s} \\) leads to better reconstruction quality. However, \\( N_{s} = 8 \\) will dramatically increase the computational cost of our method since its complexity is \\( O(N_{s}^{2}) \\), as pair-wise attentions are performed on the slices in the transformer. As a trade-off between reconstruction quality and computational complexity, we set \\( N_{s} = 4 \\) by default." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.517, + 0.415, + 0.531 + ], + "angle": 0, + "content": "Tab. 3 also covers several other ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.608 + ], + "angle": 0, + "content": "Learned v.s. One-hot Slice Indicator. Learned indicator codes (the default setting) achieve slightly better performance than using one-hot codes to indicate slice locations. This can be seen by comparing numbers from the first half row of Tab. 3 to those in the last row of Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.469, + 0.716 + ], + "angle": 0, + "content": "Slices v.s. Depth Images. Since our slice images offer layered information, it is logical to perform a comparison against depth images as supplementary information. We compared the performance of DISN when concatenating estimated depth images with the input view. Results are shown in the \"DISN+Depth\" half row, which indicates that depth images can only provide marginal improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Three Axes v.s. One Axis. We also investigated the option of slicing along only one axis. Not surprisingly, Tab. 3 shows that this underperforms against using three axes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Solid v.s Hollow Slices. Tab. 3 also demonstrates that not filling the holes improves the performance as more information about the inner/outer surfaces is revealed this way." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Using Textured Slices. We can optionally keep object textures in the slice images. Experimental results show that removing the textures in the slice images can achieve better reconstruction. The reasons could be: 1) learning to produce slice images without textures could help the model better distinguish geometries from textures; 2) for some 3D" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.286, + 0.892, + 0.318 + ], + "angle": 0, + "content": "models the color of inner and outer surfaces could be different which may complicate the learning process." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.33, + 0.766, + 0.347 + ], + "angle": 0, + "content": "5. Discussion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Our multi-slice idea represents a significant departure from conventional approaches to single-view 3D reconstruction. Through conceptual illustrations (Fig. 5) and extensive experiments, we demonstrate its compelling properties in recovering occluded object structures. Note again that all the Slice3D results on Objaverse [10] were obtained by networks trained on a very small subset (40K or \\(5\\%\\)) of the 800K 3D models therein, whereas One-2-3-45 [28] utilizes Zero-1-to-3 [29] for multi-view synthesis and the latter was fine-tuned on the entirety of the 800K. Yet our results (Table 2) show that Slice3D outperforms One-2-3-45 across the board, suggesting that our model generalizes better, likely owing to its superior occlusion-revealing abilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Our current implementation only uses four slices per direction, which is quite coarse. Since there is no camera rotation, parts of an object far away from the camera are in smaller scale, e.g., the rear leg of the chair in Fig. 10 and of the buck in Fig. 9. This can hinder our network's ability to extract detailed features. More refined and gradual slicing over those parts may be needed to improve matters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.825 + ], + "angle": 0, + "content": "The foremost limitation of slicing however, is that realistically, it can only be executed at scale on digital 3D models, not physical ones. On the other hand, the multi-view approach can be applied to physical 3D objects by capturing photos only, without ever needing 3D supervision, as in RealFusion [33]. However, this is not without a high cost: 90 minutes by RealFusion vs. 45 seconds by One-2-3-45 (requiring 3D fine-tuning for multi-view synthesis) and less than 20 seconds by Slice3D. We envision synthesizing 3D models, not many but carefully crafted few, for the multislice approach as an intriguing direction to explore." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. We thank all the anonymous reviewers and area chairs for their valuable comments. We thank Daniel Cohen-Or and Xingguang Yan for helpful discussions. This work was supported in part by NSERC." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.516, + 0.954 + ], + "angle": 0, + "content": "9888" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Bisect - blender manual. https://docs.blender.org/manual/en/2.80/modeling/meshes/editing/subdividing/bisect.html. Accessed: 2023-11-15.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Kfir Aberman, Oren Katzir, Qiang Zhou, Zegang Luo, Andrei Sharf, Chen Greif, Baoquan Chen, and Daniel Cohen-Or. Dip transform for 3D shape reconstruction. ACM Transactions on Graphics, 36(4):Article 79, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion NeRF: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[7] Wenzheng Chen, Jun Gao, Huan Ling, Edward J. Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[14] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.166, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[15] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.223, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.281, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.338, + 0.892, + 0.421 + ], + "angle": 0, + "content": "[18] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.424, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[19] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.481, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[21] Chenbo Jiang, Jie Yang, Shwai He, Yu-Kun Lai, and Lin Gao. Neuralslice: Neural 3d triangle mesh reconstruction via slicing 4d tetrahedral meshes. 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.569, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[24] Shay Kels and Nira Dyn. Reconstruction of 3d objects from 2d cross-sections with the 4-point subdivision scheme adapted to sets. Computers & Graphics, 35:741-746, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.713, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[25] Manyi Li and Hao Zhang. \\(\\mathrm{D}^2\\mathrm{IM}\\) -Net: Learning detail disentangled implicit fields from single images. In CVPR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.757, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[26] Wallace Lira, Johannes Merz, Daniel Ritchies, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop gan for unsupervised image-to-image translation. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Lu Liu, Chandrajit Bajaj, Joseph O. Deasy, Daniel A. Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. Computer Graphics Forum, 27:155-163, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.517, + 0.954 + ], + "angle": 0, + "content": "9889" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[31] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.469, + 0.335 + ], + "angle": 0, + "content": "[32] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.583, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[38] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[39] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[40] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165–174, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Dario Pavllo, David Joseph Tan, Marie-Julie Rakotosaona, and Federico Tombari. Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In CVPR, pages 4391-4401, 2023. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[43] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[44] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3d geometry via nested shape layers. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[47] Manolis Savva, Fisher Yu, Hao Su, M Aono, B Chen, D Cohen-Or, W Deng, Hang Su, Song Bai, Xiang Bai, et al. Shrec16 track: largescale 3d shape retrieval from shapenet core55. In Proceedings of the eurographics workshop on 3D object retrieval, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[48] Haim Sawdayee, Amir Vaxman, and Amit H. Bermano. Orex: Object reconstruction from planar cross-sections using neural fields. In CVPR, pages 20854-20862, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[49] Fun Shing Sin, Daniel Schroeder, and Jernej Barbič. Vega: non-linear fem deformable object simulator. In Computer Graphics Forum, pages 36-48. Wiley Online Library, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[51] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[52] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[53] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, pages 3405-3414, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[55] Maxim Tatarchenko, Stephan R. Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.516, + 0.954 + ], + "angle": 0, + "content": "9890" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[59] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, pages 52-67, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[62] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[63] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Comput. Graph. Forum, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[64] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, pages 4479–4489, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[65] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[66] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In CVPR, pages 6239–6249, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[67] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[68] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[69] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "features as a perceptual metric. In CVPR, pages 586-595, 2018. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[71] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[72] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.943, + 0.514, + 0.954 + ], + "angle": 0, + "content": "9891" + } + ] +] \ No newline at end of file diff --git a/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_origin.pdf b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b5f6962678ba1d597ba9052b0c6d16d11158ff75 --- /dev/null +++ b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/e6768b9a-cd0d-4c71-899b-0e380d81b135_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63ea6e64379cad4ade8935135d2a2ed9e1670a59baacf268b37a061b13f4c279 +size 6413836 diff --git a/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/full.md b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6f89d176bb957e2964a0eba793b697eb7984abd8 --- /dev/null +++ b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/full.md @@ -0,0 +1,293 @@ +# Slice3D: Multi-Slice, Occlusion-Revealing, Single View 3D Reconstruction + +Yizhi Wang $^{1}$ , Wallace Lira $^{1}$ , Wenqi Wang $^{2}$ , Ali Mahdavi-Amiri $^{1}$ , Hao Zhang $^{1}$ $^{1}$ Simon Fraser University $^{2}$ Tsinghua University + +![](images/2a9295f996fd930fc7e69d6c20c126c3bba43b6605a0962a1940c108fe147907.jpg) +Figure 1. Our single-view 3D reconstruction method, Slice3D, predicts multi-slice images to reveal occluded parts without changing the camera (in contrast to multi-view synthesis), and then lifts the slices into a 3D model. This leads to better recovery of occluded parts than both encoder-decoder based methods such as DISN [65] and 3D reconstructions from diffusion-synthesized multi-view images (via Zero-1-to-3 [29]) such as One-2-3-45 [28], where the key challenge of the latter is multi-view inconsistency, as shown on the right. + +# Abstract + +We introduce multi-slice reasoning, a new notion for single-view 3D reconstruction which challenges the current and prevailing belief that multi-view synthesis is the most natural conduit between single-view and 3D. Our key observation is that object slicing is a more direct, and hence more advantageous, means to reveal occluded structures than altering camera views. Specifically, slicing can peel through any occluder without obstruction, and in the limit (i.e., with infinitely many slices), it is guaranteed to unveil all hidden object parts. We realize our idea by developing Slice3D, a novel method for single-view 3D reconstruction which first predicts multi-slice images from a single RGB input image and then integrates the slices into a 3D model using a coordinate-based transformer network to product a signed distance function. The slice images can be regressed or generated, both through a U-Net based network. For the former, we inject a learnable slice indicator code to designate each decoded image into a spatial slice location, while the slice generator is a denoising diffusion model operating on the entirety of slice images stacked on the input channels. We conduct extensive evaluation against state-of-the-art alternatives to demonstrate superiority of our method, especially in recovering complex and severely occluded shape structures, amid ambiguities. All Slice3D results were produced by networks trained on a single Nvidia A40 GPU, with an inference time of less than 20 seconds. + +# 1. Introduction + +Single-view 3D reconstruction has been one of the extensively studied problems in computer vision. Despite the rapid advances on learning-based approaches to tackle this problem, including the latest attempts to leverage large foundational models and powerful diffusion-based generators, the most fundamental challenge still remains: how to faithfully reconstruct occluded parts from just one view? + +Conventional encoder-decoder based methods, e.g., [9, 17, 25, 44, 53, 65], rely on direct 3D supervision to learn mappings from global or local image features to 3D, but such features from just a single view often do not have the capacity to handle severe occlusion. Also, recent evi + +![](images/cdecd051ecd2b83eb9f88b108329a0f2d0d4a7fc73e11c1c9188906c70f6209d.jpg) +Figure 2. Multi-view vs. multi-slice for occlusion revelation, with orange edges showing revealed parts and black edges showing occluded parts. In the limit, multi-view (i.e., infinitely many views) may still leave undisclosed parts (a), e.g., the bulb, while infinitely many slices would guarantee complete structure revelation (b). With limited slices (c) and views (d), four each, the multi-slice approach tends to reveal more shape structures. + +![](images/41d5ff1055a14c2b84ff2ab10614200f1ef2d8c7149523c8261fdbfa7492bf64.jpg) +Figure 3. Slice3D for single-view 3D reconstruction from multi-slice images. The example shows 4 slice images per $X$ , $Y$ , $Z$ direction, which can be obtained by either a regressor or a generator. Regression is carried out by a U-Net encoder-decoder, where the decoder takes an extra slice indicator code that is appended to the encoder output, to obtain a slice image. There is one indicator code for each of the $3 \times 4 = 12$ slices, that is optimized during training. Our generator is a denoising diffusion model which produces a concatenation of slice images following the distribution of the ground-truth (GT) images, when conditioned on the single-view input. Inference from random noises may output multiple plausible results. The 3D reconstruction stage inputs the predicted slice images and produces an implicit field. We first extract hierarchical CNN features from the input slices, then project a 3D query point onto the 2D planes of the feature maps, and finally gather all the query-related features and send them into a transformer [58] to predict the signed distance at the query point. + +![](images/828ffce824fa801b7df7fa2abd02ba932f556386e94ea75d0b90212101b549db.jpg) + +![](images/843f584c7b0b5b95fa3bf7b21c3be1bdfceb46d7a1eca93635ef66b4b75fd95a.jpg) +Figure 4. Our slice image prediction networks are trained using occlusion-revealing multi-slice images produced from a GT 3D object mesh along a predetermined set of slicing directions (the $X$ , $Y$ , $Z$ axes of the object bounding box in the example above), while keeping the hollows as holes in the images, through which the back faces of the mesh can be visible. The 3D mesh may be textured, but the slice images are all rendered without texture. + +dence suggests that many of these methods are essentially performing image classification or retrieval [55]. As such, occlusion recovery is by and large a "hit or miss". Differentiable rendering enables 3D reasoning from images without 3D supervision [7, 38], but image-space rendering losses alone do not account for occluded structures. + +Recently, more and more works model the reconstruction problem as a single-view conditioned generative task, aiming to learn a distribution of unseen structures. Typically, multi-view images are generated from single view via diffusion. This is followed by multi-view reconstruction, e.g., via neural radiance fields (NeRF) [6, 28, 42], to obtain 3D outputs. However, multi-view consistency is a reoccurring challenge (see Fig. 1), especially over disparate views. + +We introduce a novel approach to single-view 3D reconstruction which is occlusion-revealing (Fig. 2) by first predicting multi-slice images from a single-view input and then integrating the slices into a 3D model; see Fig. 3. Fig. 4 shows slice images provided during training along a set of slicing directions. Each slice represents the projected image, in the input camera view, of one slice of the 3D mesh model. For the TV example shown, a completely occluded leg, in the camera view, is revealed in the slice images. + +The motivations behind our multi-slice reconstruction are three-fold. First, slicing through a volume or 3D mesh is a natural way to reveal occluded parts. Second, with the slice images forming an intermediate representation, our approach breaks the difficult problem of mapping from single view to 3D into two comparatively simpler tasks: slice prediction followed by 3D reconstruction from slices. + +More importantly, when contrasting our multi-slice representation with the popular multi-view approach for occlu + +![](images/2dd1e33200e1c1aacc6cbde8ed9088be2e1712e62e14b04920e0b3203080b985.jpg) +Figure 5. Singe-view reconstruction of severely occluded parts in a lamp (e.g., the bulb) by Slice3D vs. One-2-3-45 [28], a state-of-the-art multi-view approach. The multi-view images in [28] were obtained by Zero-1-to-3 [29], which was trained on renderings of 800K 3D models [10]. It is conceivable that with a limited number of views (12, the same as our slice count), the bulbs were barely visible due to occlusion by the lamp shades and bases. + +sion revelation, we observe that slicing in the limit (i.e., with infinitely many slices) is guaranteed to unveil all hidden structures, while the same cannot be said about multi-view captures, as illustrated in Fig. 2 (a-b), for a 2D lamp example. In practice, when slice and view counts are both limited, the slicing approach is more occlusion-revealing in general since it can peel through any occluders without any obstruction, unlike the multi-view case; see (c-d). On the technical front, neither our slice prediction step nor the 3D reconstruction step is concerned with camera view changes or multi-view consistency, which facilitates the use of convolutional features. In Fig. 5, we show the above contrasting characteristics of the multi-slice vs. multi-view approaches manifested in a real 3D lamp reconstruction. + +As outlined in Fig. 3, our slice images can be produced by either a regressive network or a generative one, both employing a U-Net architecture, with network losses measured against ground-truth (GT) images obtained by slicing 3D models from the training set; see Fig. 4. In the regressive network, we inject a learnable slice indicator code into the encoder-decoder setup to designate each decoded image into the proper spatial slice location. The slice generator is a denoising diffusion model operating on separate channels for different slice images. In the second stage, 3D reconstruction from the obtained slice images, we design a coordinate-based transformer to obtain an implicit field. + +Our network is coined Slice3D. We conduct extensive experiments on ShapeNet [5] and Objverse [10] for evaluation. Quantitative and qualitative comparisons are made to representative methods including encoder-decoders capturing local image features [65], autoregressive latent models [36], generative model inversion [42], and three recent methods with diffusion + NeRF [6, 28, 33], demonstrating superiority of our method in terms of reconstruction quality, generalizability, and handling of ambiguities. + +# 2. Related work + +The earliest inspiration for our multi-slice 3D reconstruction idea was GANHopper by Lira et al. [26]. In the con + +text of unsupervised image-to-image translation, the key idea of their work is to gradually execute the translation between disparate domains by making small "hops," rather than completing it in one go. We were then motivated to tackle the single-view 3D reconstruction problem as a domain translation task, hopping through the multi-slice images so as to bridge the gap between 2D and 3D. + +Neural fields. Neural implicit functions [63] are learned by a neural network to predict such properties as color, occupancy, opacity, and surface normals. Earlier works such as IM-Net [8], occupancy networks [34], and DeepSDF [41] represent a 3D shape as a latent code and send it to a decoder formed by MLPs to predict occupancy or SDF values for any query point sampled in 3D. NeRF [35] and its variants [15] represent a 3D scene using a neural radiance field for novel view synthesis. Many works on neural 3D reconstruction [15, 63] take on single- or multi-view images as well as point clouds as input. In contrast, the reconstruction module in Slice3D is built on multi-slice images. + +Single-view reconstruction from image features. An early work to exploit neural image features for 3D reconstruction was Pixel2Mesh [59], which uses a graph representation. Subsequently, neural implicit fields have been widely adopted for single-view reconstruction. IM-Net [8] first extracts a global latent code, from an input view, to learn a 3D implicit field. Auto-SDF [36] and ShapeFormer [66] employ vector quantization (VQ) [57] to improve the representation of latent features. More recent works [22, 37, 68] consider single-view reconstruction as a 3D generation task conditioned on the latent code of the input view. However, a global latent code is insufficient to capture finer details on a 3D object, often leading to overly smoothed or retrieved outcomes that do not faithfully represent the input. To alleviate this problem, both DISN [65] and $\mathrm{D}^2\mathrm{IM}$ -Net [25] opted to learn local images features, along with global ones, to improve the recovery of fine-grained topological (e.g., thin) features and surface details, respectively. + +3D reconstruction via multi-view. Conventional 3D reconstruction approaches [14] involve multi-view reasoning. With recent advances in generative AI, a prevailing practice in single-view 3D reconstruction is to first employ generative models such as GANs [16, 23] and diffusion [11, 20, 45, 50] to generate several novel views from the input view, and then feed them to a multi-view 3D reconstruction module such as NeRF. NeRF-based surface reconstruction [39, 60, 67] can be further employed to output 3D meshes. EG3D [3] and NeRF-from-Image [42] build NeRFs or its variants upon the outputs of a style-GAN [23], while more recent works [4, 6, 18, 28-30, 33, 42, 43, 51, 62, 64] are favoring diffusion models such as Stable Diffusion (SD) [45]. To date, Zero-1-to-3 [29] represents the state of the art on single- to multi-view synthesis; it is built on SD and fine-tuned by 800K 3D models from Objaverse [10]. + +To alleviate the multi-view consistency problem, which is an reoccurring challenge for the image-space generative methods, recent methods [30, 31, 43, 52, 71] aim to enhance the consistency by performing spatial attention between different views, which dramatically increases the computational overhead. Nevertheless, the consistency of two disparate views is still far from satisfactory. A deeper reason is that standard convolutional neural networks (CNNs) are not inherently rotation-invariant, making it difficult to build correspondences between significantly different views. + +Slice representation. There have been techniques to use object slices to reconstruct a 3D shape, e.g., as inspired by Magnetic Resonance Imaging (MRI). Early works in computer graphics employ either parallel cross-sectional curves, e.g., [24], or non-parallel curve networks, e.g., [27], as inputs. Recently, with the advent of machine learning, OREx [48] and Cut-and-Approximate [40] develop neural reconstruction models to obtain 3D meshes, also from a sparse collection of 2D planar cross-sections, while Neural-Slice [21] uses slicing to obtain 3D shapes from 4D tetrahedral meshes. However, none of these techniques, nor others that we are aware of, address the problem of single-view 3D reconstruction since their inputs all consist of 3D data such as curve slices or point clouds. Moreover, our method learns to predict much fewer slice images and use them as input to a transformer-based reconstruction module. + +Dip. Dip transform [2] is an ingenious way to acquire occluded 3D structures by measuring fluid displacement when an object is repeatedly oriented and dipped into a fluid. Similar to slicing, Dip was also designed to "reach" occluded parts that optical captures via lines of sight cannot. However, both the occlusion revelation (liquid penetration vs. slicing) and target applications (acquisition vs. single-view reconstruction) are completely different. + +Multi-plane image (MPI). Introduced by Zhou et al. [72], an MPI representation for a scene consists of a set of frontoparallel planes at fixed depths, with each image intended to capture scene appearance and visibility at a specific depth. Conceptually, both MPs and our slice images provide a layered scene representation for occlusion revelation. However, MPs [13, 19, 56, 69, 72] have only been employed for novel view synthesis, not 3D reconstruction, and the prediction networks were trained on multi-view images, e.g., from Youtube videos [72]. The generated MPs only reveal a small amount of occluded scenes to account for view shifts. In contrast, our method directly cuts through a 3D object to reveal all the occluded parts at several depth ranges. + +# 3. Method + +Our method is based on a slicing representation; see Section 3.1 and Fig. 4. Given a single-view RGB image $I \in \mathbb{R}^{H \times W \times 3}$ of a 3D object, we first learn a slicing network + +(Section 3.2) which takes $I$ as input and outputs its corresponding multi-slice images. Then, an SDF prediction network is trained (Section 3.3) to leverage image features from the slices to produce an implicit field of the object. + +# 3.1. Slicing Representation + +As shown in Fig. 4, given a 3D object mesh, we slice it along the $X$ , $Y$ , and $Z$ axes of its bounding box and render the sliced segments. The slice images can optionally include textures or have holes filled. By default, we remove the textures and keep the holes, which could reveal back faces of the mesh, and overall, it produces the best experimental results. The rendered slice images consist of $\{S_i^j | 1 \leq i \leq N_s, j \in \{X,Y,Z\}\}$ , where $S_i^j \in \mathbb{R}^{H \times W \times 1}$ , $j$ denotes the axis and $N_s$ is the number of slices along each axis. In our current implementation, we produce $N_s = 4$ slices per direction. Compared to planar cross-sections or tomography scans which consist of densely stacked planar images, our slicing representation is more compact with each slice image revealing more object semantics. + +# 3.2. Slice Regression and Generation + +Given an image $I$ , observing a 3D shape from a single view, we aim to train a slicing network $F_{s}$ that can produce the slice images from $I$ : $F_{s}(I) = \{\hat{S}_{i}^{j}\}$ . We have two different designs for such a network: regressive and generative. The former aims to regress GT slices, assuming deterministic occluded parts based on the partial observation of the input view. The latter aims to generate slices that are consistent with the input view, considering various possible occluded parts that do not necessarily match the GT slices. + +# 3.2.1 Regression-based Slicing + +We employ a U-Net [46] encoder-decoder to regress the slices from $I$ . As shown in the top-left of Fig. 3, an indicator code $E_{i}^{j}\in \mathbb{R}^{N_{e}}$ is concatenated with the latent code of $I$ to indicate which slice to output: $\hat{S}_i^j = F_d([F_e(I);E_i^j])$ , where $F_{e}$ and $F_{d}$ are the encoder and decoder of $F_{s}$ , respectively, and $[\cdot ]$ denotes concatenation. + +The indicator codes $\{E_i^j\}$ are randomly initialized and jointly optimized with the whole network. The loss function $\mathcal{L}_{reg}$ is the reconstruction error measured by the $\mathcal{L}_1$ -Norm and perceptual loss [70] $(\mathcal{L}_p)$ for all slices: + +$$ +\mathcal {L} _ {r e g} = \frac {\sum_ {i , j} \left| \left| \hat {S} _ {i} ^ {j} - S _ {i} ^ {j} \right| \right| _ {1} + \mathcal {L} _ {p} \left(\hat {S} _ {i} ^ {j} , S _ {i} ^ {j}\right)}{| S |}, \tag {1} +$$ + +where $|S| = 3 \cdot N_{s}$ is the total number of slices. + +# 3.2.2 Generation-based Slicing + +Single-view occlusion revelation is clearly ill-posed, with potential ambiguities leading to multiple plausible 3D re + +![](images/e97b9d5609c4e2ba6cd05f620fc452a8ffe9a50712b29d667405001a4dedbe2b.jpg) +Figure 6. The pipeline of generation-based slicing via DDPM. $\bigoplus$ denotes the injection of noise to slice images. + +constructions. To this end, we consider slicing as an image-conditioned generation task and develop a generation-based slicing via Denoising Diffusion Probabilistic Models (DDPM) [20], whose detailed pipeline is shown in Fig. 6. + +The key challenge of applying DDPM lies in how to generate consistent slices. Rather than diffusing each single slice, we propose to concatenate all slices into an aggregated structure $[S]$ , along either the color dimension $([S] \in \mathbb{R}^{H \times W \times |S|})$ or a spatial dimension such as height $([S] \in \mathbb{R}^{(|S| \cdot H) \times W \times 1})$ . The concatenation of slices instructs DDPM to learn the correlation of slices, yielding consistent generation results. By default, we concatenate along the color dimension. The comparison of these two concatenation ways can be found in the supplementary. + +Following DDPM, we first inject a noise $\epsilon \sim \mathcal{N}(0,1)$ to $[S]$ to obtain $[\bar{S}]$ . The noised slices $[\bar{S}]$ are then sent to a U-Net based denoising network $F_{dn}$ that outputs the predicted noise conditioned on $I$ : $\hat{\epsilon} = F_{dn}([\bar{S}], I)$ . In practice, the injection and prediction of noise $\epsilon$ involve an iteration step $t$ , which we omit here for brevity. To better condition $I$ for denoising diffusion, we employ a pre-trained VGG to extract multi-layer features from $I$ and add them into the corresponding layers in the encoder of $F_{dn}$ (see the blue part of Fig. 6). The loss function is the denosing loss: $\mathcal{L}_{gen} = ||\hat{\epsilon} - \epsilon||_2^2$ . After training, we can sample random noises to generate a set of slice images $\{\hat{S}_i^j\}$ . + +Compared to regression-based slicing, generation-based slicing can produce multiple plausible outputs rather than a single deterministic result, but requires much longer inference time due to the nature of diffusion models. + +# 3.3. Learning Implicit Field from Slices + +We learn an implicit field from slice images, by projecting a 3D query point onto each slice image, extracting query-specific features, and then aggregating them to predict a signed distance. For regression-based slicing, we train the SDF prediction network on the predicted slices from the previous section thus the whole pipeline can be trained end-to-end. For generation-based slicing, we train the SDF pre + +diction network on GT slice images since the generated slices may not match the GT signed distances. + +Specifically, for a query point $q \in \mathbb{R}^3$ , we project it to a location $\tilde{q} \in \mathbb{R}^2$ on the 2D image plane according to the camera parameters $\Theta$ . For slice $\hat{S}_i^j$ (regression) or $S_i^j$ (generation), we first extract its feature maps, then retrieve features from each feature map corresponding to location $\tilde{q}$ , and finally concatenate them to obtain the image feature vector $V_i^j \in \mathbb{R}^{1 \times 1 \times N_c}$ . The retrieval is accomplished via bi-linear interpolation over the features grids that encompass the location of $\tilde{q}$ . We design a transformer network [58] $F_t$ which learns to aggregate the feature vectors from all slices to predict the signed distance $\hat{d}$ : $\hat{d} = F_t([q', V_1^X, ..., V_{N_s}^Z])$ , where $q' \in \mathbb{R}^{N_c}$ is linearly projected from $q$ to be the same dimension with $V_i^j$ . The optimization objective is to minimize the $\mathcal{L}_1$ distance between predicted and GT signed distance: $\mathcal{L}_{imp} = ||\hat{d} - d||_1$ . In regression-based slicing, $\mathcal{L}_{imp}$ and $\mathcal{L}_{reg}$ are jointly optimized. + +DISN [65] extracts from only the input view $I$ to obtain a feature vector $V$ . For query points located around occluded parts of the object (e.g., the query point in Fig. 3), $V$ is irrelevant to $q$ since it is extracted from the front parts that block $q$ , often leading to erroneous SDF prediction. In contrast, one or more of our slice images could reveal the occluded parts, yielding more accurate features fed to the transformer for SDF prediction. After learning the implicit field, we apply Marching Cubes [32] to obtain triangle meshes. Estimating the extrinsics (camera poses) of $\Theta$ can output 3D shapes in canonical poses, which is optional and applied only on ShapeNet [5]; See more details in Supplemental. + +# 4. Results and Evaluation + +Dataset. We utilize two well-known datasets, ShapeNet [5] and Objaverse [10], to train and test our network. For ShapeNet, our method is applied to Chairs and Airplanes separately, using the train/validation/test split as in SHREC16 [47]. Objaverse is a challenging dataset, where many objects may not belong to a clear category. In our experiment, we randomly selected $\approx 40\mathrm{K}$ shapes from Objaverse to keep resource and time costs manageable, while ensuring diversity of the collection. We follow a $95\% -5\%$ train-test split over the 40K selected shapes, ensuring that every training model we use is also from the training set of Zero-1-2-3 [29]. Note that Zero-1-2-3 used a train-test split of $99\% -1\%$ over the entire 800K models in Objaverse. + +Rendering Slice Images. We employ the bisect function in Blender [1] to cut an object mesh into slices and then render them. By default, we keep internal cavities and holes produced by slicing. Since many 3D models in Objaverse are not in, or have no, canonical poses and feature random rotations, slicing them along the default $X$ , $Y$ , and $Z$ directions is not suitable. To this end, we opted for camera-aligned slicing instead. Specifically, we set the camera position, in + +
Datasets MetricsShapeNet ChairsShapeNet Airplanes
CD↓F1↑HD↓CD↓F1↑HD↓
DISN [65]3.252.347.843.482.817.62
Auto-SDF [36]25.00.7715.8---
NeRF-Img [42]19.51.2713.416.01.4114.0
One-2-3-45 [28]12.51.0111.723.40.7512.6
Ours2.802.436.752.992.836.45
DISN* [65]3.022.427.863.192.967.56
SSDNeRF* [6]8.982.0610.910.42.0010.4
Ours*2.672.536.702.812.966.38
+ +Table 1. Quantitative results of single-view 3D reconstruction on ShapeNet Chairs and Airplanes. \* denotes using ground-truth (GT) camera poses. Slice3D (ours) uses regression-based slicing. + +troduce a random rotation to the object, position it in front of the camera, and utilize the camera's axes for slicing. Under this setting, estimating the rotation of objects is unnecessary; see more details in the supplementary material. + +Implementation Details. When preparing training data, we use the Vega FEM Library [49] for SDF computation, same as DISN [65] and AutoSDF [36]. For each object, the diagonal length of the object's 3D bounding box is normalized to 1. We set $H$ and $W$ to 128, $N_e$ to 128, $N_s$ to 4, and $N_p$ to 1,024. We train our model on only one Nvidia A40 GPU. Evaluation Metrics. We use Chamfer $\mathcal{L}_2$ Distance (CD), F-score%1 (F1) [54], and Hausdorff Distance (HD) to evaluate reconstruction results. Besides these 3D metrics, we also evaluate using 2D metrics, PSNR, SSIM [61], and LPIPS [70], on 24 rendered views of reconstructed meshes. + +# 4.1. Comparisons to State Of The Art (SOTA) + +We compare our method with SOTA methods including DISN [65], AutoSDF [36], NeRF-From-Img [42], SSD-NeRF [6], RealFusion [33], and One-2-3-45 [28]. Note that when showing visual results, we always present two views to remove any view bias towards the input. + +ShapeNet. Figs. 7 and 8 demonstrate the effectiveness of our approach in reconstructing obscured sections of objects while precisely restoring visible ones. In contrast, DISN [65] can barely reconstruct the occluded parts. AutoSDF [36] and NeRF-From-Image [42] rely on latent features for reconstruction, which can yield rather clean results but they do not necessarily respect the input image, as the chair results from Fig. 7 evidently showed. SSDNeRF [6] tends to produce erroneous geometric details and the surfaces are not as smooth as from other methods, especially for the airplane examples in Fig. 8. Generally speaking, One-2-3-45 [28] yields rough and bulky geometries that clearly lack details. Additionally, our method outperforms the other methods in the reconstruction of unconventional inputs, as shown by the two airplanes in Fig. 8. + +Tab. 1 shows quantitative comparisons to the above methods on the two ShapeNet categories. When our method operates without using GT camera poses and relies on esti + +
MethodCD↓F1↑HD↓PSNR↑SSIM↑LPIPS↓
DISN [65]28.21.4723.021.98.7915.2
One-2-3-45 [28]35.40.8729.819.88.5519.8
Ours (G)25.01.5116.422.38.8413.4
Ours (R)16.61.5313.422.68.8812.8
+ +Table 2. Quantitative results of single-view 3D reconstruction on the Objaverse dataset. 'R' and and 'G' denote regression-based and generation-based slicing in Slice3D, respectively. + +mated poses, its performance sees only a marginal degradation. Notably, for Chair, our method with estimated poses even outperforms DISN and SSDNerf with GT poses. It is also notable that One-2-3-45 tends to output dilated shapes. Although it falls behind DISN in quantitative results, it usually generates more complete shapes and more accurate topologies than DISN; see Figs. 7 and 8. + +Objaverse. Since AutoSDF [36], NeRF-From-Image [42], and SSDNeRF [6] are all category-specific methods, they cannot be trained on Objaverse which consists of many 3D objects with no clear categories. Thus, we mainly compare our method with DISN [65], RealFusion [33], and One-2-3-45 [28]; see Fig. 9 for a qualitative comparison. + +Note that RealFusion [33] and One-2-3-45 [28] both leverage a significantly larger volume of training data compared to our method, as they both rely on a pre-trained Stable Diffusion [45] model. Despite this disparity in the scale of the training sets, our method still outperforms these two methods, as well as DISN; see Fig. 9. Quantitatively, Tab. 2 demonstrates our method's across-the-board advantage over DISN and One-2-3-45 (more so on Objaverse than on ShapeNet), in both 3D and 2D metrics, especially CD, HD, and LPIPS. These results also show that our method possesses the capacity to generalize across common shapes rather than being confined to specific categories. + +Real-world Image Input. To further stress-test the methods, we perform Google image search on object keywords (e.g., "chair", "airplane", etc.) and take the top-1 image with occlusions as input. In addition, we use some examples from Google Scanned Objects (GSO) [12] dataset. Fig. 10 shows that our method performs the best in recovering occluded parts while preserving fine details, e.g., the turbines of the airplane, the heel, and the opening of the boot. + +Multi-slice vs. Multi-view Reconstruction. Fig. 11 compares between multi-slice and multi-view reconstructions. We demonstrate two different sets of multi-slices and multi-views generated from our method and One-2-3-45 [28]. Our generated slice images depict two distinct configurations of chair legs, maintaining high consistency across slices. In contrast, the multiple views synthesized in One-2-3-45 do encounter inconsistency issues, compromising an accurate 3D reconstruction. We show three slices in each set and more can be found in the supplementary material. + +Training and Inference Costs. Unlike RealFusion [33] and One-2-3-45 [28], our method does not depend on a large + +![](images/ee668f36d8e41fa6c5b91a23e97fd7c60e3563ef640863aec0b850e45831ac2f.jpg) +Figure 7. Visual comparison between single-view 3D reconstruction methods on two ShapeNet chairs. DISN and our method (based on regressive slicing) utilize the same estimated camera parameters. Two different views are displayed to remove view bias. + +![](images/d6c6bb659619caf5510ee89c559132ebff623f99fb567beade6977155083b4ff.jpg) + +![](images/425773c5acaf5dc177562335accb4b537a6e2384cce327a00aa5a6ba65223e67.jpg) +Figure 8. Visual comparison between single-view 3D reconstruction methods on two somewhat unconventional airplanes from ShapeNet. + +![](images/c25a9ef23af71e990c63a87a8bccf7cda25dc94b04bae448cc84138ab403f663.jpg) +Figure 9. Visual comparison between cross-category reconstruction methods on Objverse. Our method better recovers overall geometries, fine details, and is more faithful to the inputs. Except for the seaweed (squared), all the other models are in the training set of Zero-1-to-3. +Figure 10. On real-world image inputs. Top two from Google image search and the others from GSO [12]. Zoom in for details. + +pre-trained model [45]. NeRF-based methods tend to have long inference times, as many of them require per-test optimization. Our method has much faster inference times: 10 or 20 seconds with slice regression or generation, com + +pared to 90 minutes for RealFusion and 45 seconds for One-2-3-45 and SSDNeRF, while delivering superior results. With similar inference times to DISN, Auto-SDF, and NeRF-From-Img, our method has better reconstruction quality. See supplementary material for more details. + +# 4.2. Ablation Study + +Generative vs. Regressive Slicing. As shown in the first row of Fig. 11, for generation-based slicing, we can produce multiple plausible results for a given input view. Tab. 2 shows that regressive slicing outperforms generative slicing when compared to GT. However, it is notable that the 3D outputs may differ from GT when input views have ambiguous occluded parts. Regressive training struggles to handle such one-to-many mappings, leading to generation outperforming regression in reconstructing occluded parts. + +Number of Slices. We test the performance of our method + +![](images/c32affa8f67f8e1d6326c19515f6cb14cac8392d6d72ea628c003e98c11de202.jpg) +Figure 11. Multi-slice vs. multi-view reconstructions amid ambiguities in the chair legs. Both One-2-3-45 (bottom) and Slice3D (top) can produce multiple results. Our results are both plausible from consistent slices, while One-2-3-45 suffers from multi-view inconsistencies. + +
SettingCD↓F1↑HD↓SettingCD↓F1↑HD↓
One-hot ind.2.682.536.82only X2.882.397.74
DISN+Depth2.992.437.81only Y2.752.456.84
Ns=22.902.467.31only Z3.082.387.60
Ns=42.672.536.70Filled2.732.507.37
Ns=82.562.566.93w/ textures2.822.507.70
+ +Table 3. Ablation studies on ShapeNet Chairs. + +with regression-based slicing by varying $N_{s}$ between 2, 4, and 8, on ShapeNet Chairs. From Tab. 3, it is evident that larger $N_{s}$ leads to better reconstruction quality. However, $N_{s} = 8$ will dramatically increase the computational cost of our method since its complexity is $O(N_{s}^{2})$ , as pair-wise attentions are performed on the slices in the transformer. As a trade-off between reconstruction quality and computational complexity, we set $N_{s} = 4$ by default. + +Tab. 3 also covers several other ablation studies. + +Learned v.s. One-hot Slice Indicator. Learned indicator codes (the default setting) achieve slightly better performance than using one-hot codes to indicate slice locations. This can be seen by comparing numbers from the first half row of Tab. 3 to those in the last row of Tab. 1. + +Slices v.s. Depth Images. Since our slice images offer layered information, it is logical to perform a comparison against depth images as supplementary information. We compared the performance of DISN when concatenating estimated depth images with the input view. Results are shown in the "DISN+Depth" half row, which indicates that depth images can only provide marginal improvement. + +Three Axes v.s. One Axis. We also investigated the option of slicing along only one axis. Not surprisingly, Tab. 3 shows that this underperforms against using three axes. + +Solid v.s Hollow Slices. Tab. 3 also demonstrates that not filling the holes improves the performance as more information about the inner/outer surfaces is revealed this way. + +Using Textured Slices. We can optionally keep object textures in the slice images. Experimental results show that removing the textures in the slice images can achieve better reconstruction. The reasons could be: 1) learning to produce slice images without textures could help the model better distinguish geometries from textures; 2) for some 3D + +models the color of inner and outer surfaces could be different which may complicate the learning process. + +# 5. Discussion and Future Work + +Our multi-slice idea represents a significant departure from conventional approaches to single-view 3D reconstruction. Through conceptual illustrations (Fig. 5) and extensive experiments, we demonstrate its compelling properties in recovering occluded object structures. Note again that all the Slice3D results on Objaverse [10] were obtained by networks trained on a very small subset (40K or $5\%$ ) of the 800K 3D models therein, whereas One-2-3-45 [28] utilizes Zero-1-to-3 [29] for multi-view synthesis and the latter was fine-tuned on the entirety of the 800K. Yet our results (Table 2) show that Slice3D outperforms One-2-3-45 across the board, suggesting that our model generalizes better, likely owing to its superior occlusion-revealing abilities. + +Our current implementation only uses four slices per direction, which is quite coarse. Since there is no camera rotation, parts of an object far away from the camera are in smaller scale, e.g., the rear leg of the chair in Fig. 10 and of the buck in Fig. 9. This can hinder our network's ability to extract detailed features. More refined and gradual slicing over those parts may be needed to improve matters. + +The foremost limitation of slicing however, is that realistically, it can only be executed at scale on digital 3D models, not physical ones. On the other hand, the multi-view approach can be applied to physical 3D objects by capturing photos only, without ever needing 3D supervision, as in RealFusion [33]. However, this is not without a high cost: 90 minutes by RealFusion vs. 45 seconds by One-2-3-45 (requiring 3D fine-tuning for multi-view synthesis) and less than 20 seconds by Slice3D. We envision synthesizing 3D models, not many but carefully crafted few, for the multislice approach as an intriguing direction to explore. + +Acknowledgements. We thank all the anonymous reviewers and area chairs for their valuable comments. We thank Daniel Cohen-Or and Xingguang Yan for helpful discussions. This work was supported in part by NSERC. + +# References + +[1] Bisect - blender manual. https://docs.blender.org/manual/en/2.80/modeling/meshes/editing/subdividing/bisect.html. Accessed: 2023-11-15.5 +[2] Kfir Aberman, Oren Katzir, Qiang Zhou, Zegang Luo, Andrei Sharf, Chen Greif, Baoquan Chen, and Daniel Cohen-Or. Dip transform for 3D shape reconstruction. ACM Transactions on Graphics, 36(4):Article 79, 2017. 4 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3 +[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 3 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5 +[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion NeRF: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714, 2023. 2, 3, 6 +[7] Wenzheng Chen, Jun Gao, Huan Ling, Edward J. Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in Neural Information Processing Systems, 2019. 2 +[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3 +[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In CVPR, 2020. 1 +[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 3, 5, 8 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3 +[12] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6, 7 +[13] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 4 +[14] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3 +[15] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3 +[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3 +[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 1 +[18] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3 +[19] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 4 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5 +[21] Chenbo Jiang, Jie Yang, Shwai He, Yu-Kun Lai, and Lin Gao. Neuralslice: Neural 3d triangle mesh reconstruction via slicing 4d tetrahedral meshes. 2023. 4 +[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 3 +[24] Shay Kels and Nira Dyn. Reconstruction of 3d objects from 2d cross-sections with the 4-point subdivision scheme adapted to sets. Computers & Graphics, 35:741-746, 2011. 4 +[25] Manyi Li and Hao Zhang. $\mathrm{D}^2\mathrm{IM}$ -Net: Learning detail disentangled implicit fields from single images. In CVPR, 2021. 1, 3 +[26] Wallace Lira, Johannes Merz, Daniel Ritchies, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop gan for unsupervised image-to-image translation. In ECCV, 2020. 3 +[27] Lu Liu, Chandrajit Bajaj, Joseph O. Deasy, Daniel A. Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. Computer Graphics Forum, 27:155-163, 2008. 4 +[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 1, 2, 3, 6, 8 + +[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 1, 3, 5, 8 +[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image, 2023. 3, 4 +[31] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023.4 +[32] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 5 +[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 3, 6, 8 +[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 3 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3, 6 +[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3 +[38] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2 +[39] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 3 +[40] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 4 +[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165–174, 2019. 3 +[42] Dario Pavllo, David Joseph Tan, Marie-Julie Rakotosaona, and Federico Tombari. Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In CVPR, pages 4391-4401, 2023. 2, 3, 6 + +[43] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 3, 4 +[44] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3d geometry via nested shape layers. In CVPR, 2018. 1 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3, 6, 7 +[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4 +[47] Manolis Savva, Fisher Yu, Hao Su, M Aono, B Chen, D Cohen-Or, W Deng, Hang Su, Song Bai, Xiang Bai, et al. Shrec16 track: largescale 3d shape retrieval from shapenet core55. In Proceedings of the eurographics workshop on 3D object retrieval, 2016. 5 +[48] Haim Sawdayee, Amir Vaxman, and Amit H. Bermano. Orex: Object reconstruction from planar cross-sections using neural fields. In CVPR, pages 20854-20862, 2023. 4 +[49] Fun Shing Sin, Daniel Schroeder, and Jernej Barbič. Vega: non-linear fem deformable object simulator. In Computer Graphics Forum, pages 36-48. Wiley Online Library, 2013. 6 +[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[51] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 3 +[52] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 4 +[53] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 1 +[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, pages 3405-3414, 2019. 6 +[55] Maxim Tatarchenko, Stephan R. Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2 +[56] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 4 +[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3 +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5 +[59] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, pages 52-67, 2018. 3 +[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3 +[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[62] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3 +[63] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Comput. Graph. Forum, 2022. 3 +[64] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, pages 4479–4489, 2023. 3 +[65] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 1, 3, 5, 6 +[66] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In CVPR, pages 6239–6249, 2022. 3 +[67] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3 +[68] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3 +[69] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 4 +[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep + +features as a perceptual metric. In CVPR, pages 586-595, 2018. 4, 6 +[71] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023.4 +[72] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 4 \ No newline at end of file diff --git a/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/images.zip b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6548a615cc1b881bb04db8c591dde92220679396 --- /dev/null +++ b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bed603f3507e23748381328ce52454cd31500d13b54428e331a0d27ee5e64684 +size 582440 diff --git a/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/layout.json b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..697be171ffcb7c40f0ef52413cf9c56299f732a1 --- /dev/null +++ b/2024/Slice3D_ Multi-Slice Occlusion-Revealing Single View 3D Reconstruction/layout.json @@ -0,0 +1,9012 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 103, + 523, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 103, + 523, + 120 + ], + "spans": [ + { + "bbox": [ + 68, + 103, + 523, + 120 + ], + "type": "text", + "content": "Slice3D: Multi-Slice, Occlusion-Revealing, Single View 3D Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "spans": [ + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "content": "Yizhi Wang" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "content": ", Wallace Lira" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "content": ", Wenqi Wang" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "content": ", Ali Mahdavi-Amiri" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "content": ", Hao Zhang" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "content": "Simon Fraser University " + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 483, + 172 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 178, + 535, + 293 + ], + "blocks": [ + { + "bbox": [ + 60, + 178, + 535, + 293 + ], + "lines": [ + { + "bbox": [ + 60, + 178, + 535, + 293 + ], + "spans": [ + { + "bbox": [ + 60, + 178, + 535, + 293 + ], + "type": "image", + "image_path": "2a9295f996fd930fc7e69d6c20c126c3bba43b6605a0962a1940c108fe147907.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 546, + 346 + ], + "type": "text", + "content": "Figure 1. Our single-view 3D reconstruction method, Slice3D, predicts multi-slice images to reveal occluded parts without changing the camera (in contrast to multi-view synthesis), and then lifts the slices into a 3D model. This leads to better recovery of occluded parts than both encoder-decoder based methods such as DISN [65] and 3D reconstructions from diffusion-synthesized multi-view images (via Zero-1-to-3 [29]) such as One-2-3-45 [28], where the key challenge of the latter is multi-view inconsistency, as shown on the right." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 355, + 192, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 355, + 192, + 368 + ], + "spans": [ + { + "bbox": [ + 143, + 355, + 192, + 368 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 388, + 290, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 290, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 290, + 700 + ], + "type": "text", + "content": "We introduce multi-slice reasoning, a new notion for single-view 3D reconstruction which challenges the current and prevailing belief that multi-view synthesis is the most natural conduit between single-view and 3D. Our key observation is that object slicing is a more direct, and hence more advantageous, means to reveal occluded structures than altering camera views. Specifically, slicing can peel through any occluder without obstruction, and in the limit (i.e., with infinitely many slices), it is guaranteed to unveil all hidden object parts. We realize our idea by developing Slice3D, a novel method for single-view 3D reconstruction which first predicts multi-slice images from a single RGB input image and then integrates the slices into a 3D model using a coordinate-based transformer network to product a signed distance function. The slice images can be regressed or generated, both through a U-Net based network. For the former, we inject a learnable slice indicator code to designate each decoded image into a spatial slice location, while the slice generator is a denoising diffusion model operating on the entirety of slice images stacked on the input channels. We conduct extensive evaluation against state-of-the-art alternatives to demonstrate superiority of our method, especially in recovering complex and severely occluded shape structures, amid ambiguities. All Slice3D results were produced by networks trained on a single Nvidia A40 GPU, with an inference time of less than 20 seconds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 355, + 386, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 386, + 368 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 386, + 368 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 377, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 461 + ], + "type": "text", + "content": "Single-view 3D reconstruction has been one of the extensively studied problems in computer vision. Despite the rapid advances on learning-based approaches to tackle this problem, including the latest attempts to leverage large foundational models and powerful diffusion-based generators, the most fundamental challenge still remains: how to faithfully reconstruct occluded parts from just one view?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 464, + 546, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 523 + ], + "type": "text", + "content": "Conventional encoder-decoder based methods, e.g., [9, 17, 25, 44, 53, 65], rely on direct 3D supervision to learn mappings from global or local image features to 3D, but such features from just a single view often do not have the capacity to handle severe occlusion. Also, recent evi" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 545, + 544, + 626 + ], + "blocks": [ + { + "bbox": [ + 309, + 545, + 544, + 626 + ], + "lines": [ + { + "bbox": [ + 309, + 545, + 544, + 626 + ], + "spans": [ + { + "bbox": [ + 309, + 545, + 544, + 626 + ], + "type": "image", + "image_path": "cdecd051ecd2b83eb9f88b108329a0f2d0d4a7fc73e11c1c9188906c70f6209d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 632, + 545, + 710 + ], + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 710 + ], + "type": "text", + "content": "Figure 2. Multi-view vs. multi-slice for occlusion revelation, with orange edges showing revealed parts and black edges showing occluded parts. In the limit, multi-view (i.e., infinitely many views) may still leave undisclosed parts (a), e.g., the bulb, while infinitely many slices would guarantee complete structure revelation (b). With limited slices (c) and views (d), four each, the multi-slice approach tends to reveal more shape structures." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 746, + 314, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 314, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 314, + 756 + ], + "type": "text", + "content": "9881" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 293, + 283 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 293, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 293, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 293, + 283 + ], + "type": "image", + "image_path": "41d5ff1055a14c2b84ff2ab10614200f1ef2d8c7149523c8261fdbfa7492bf64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "text", + "content": "Figure 3. Slice3D for single-view 3D reconstruction from multi-slice images. The example shows 4 slice images per " + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "text", + "content": " direction, which can be obtained by either a regressor or a generator. Regression is carried out by a U-Net encoder-decoder, where the decoder takes an extra slice indicator code that is appended to the encoder output, to obtain a slice image. There is one indicator code for each of the " + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "inline_equation", + "content": "3 \\times 4 = 12" + }, + { + "bbox": [ + 46, + 291, + 548, + 381 + ], + "type": "text", + "content": " slices, that is optimized during training. Our generator is a denoising diffusion model which produces a concatenation of slice images following the distribution of the ground-truth (GT) images, when conditioned on the single-view input. Inference from random noises may output multiple plausible results. The 3D reconstruction stage inputs the predicted slice images and produces an implicit field. We first extract hierarchical CNN features from the input slices, then project a 3D query point onto the 2D planes of the feature maps, and finally gather all the query-related features and send them into a transformer [58] to predict the signed distance at the query point." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 539, + 283 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 539, + 283 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 539, + 283 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 539, + 283 + ], + "type": "image", + "image_path": "828ffce824fa801b7df7fa2abd02ba932f556386e94ea75d0b90212101b549db.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 393, + 266, + 541 + ], + "blocks": [ + { + "bbox": [ + 69, + 393, + 266, + 541 + ], + "lines": [ + { + "bbox": [ + 69, + 393, + 266, + 541 + ], + "spans": [ + { + "bbox": [ + 69, + 393, + 266, + 541 + ], + "type": "image", + "image_path": "843f584c7b0b5b95fa3bf7b21c3be1bdfceb46d7a1eca93635ef66b4b75fd95a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "lines": [ + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "type": "text", + "content": "Figure 4. Our slice image prediction networks are trained using occlusion-revealing multi-slice images produced from a GT 3D object mesh along a predetermined set of slicing directions (the " + }, + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 550, + 288, + 627 + ], + "type": "text", + "content": " axes of the object bounding box in the example above), while keeping the hollows as holes in the images, through which the back faces of the mesh can be visible. The 3D mesh may be textured, but the slice images are all rendered without texture." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "content": "dence suggests that many of these methods are essentially performing image classification or retrieval [55]. As such, occlusion recovery is by and large a \"hit or miss\". Differentiable rendering enables 3D reasoning from images without 3D supervision [7, 38], but image-space rendering losses alone do not account for occluded structures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 396, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 491 + ], + "type": "text", + "content": "Recently, more and more works model the reconstruction problem as a single-view conditioned generative task, aiming to learn a distribution of unseen structures. Typically, multi-view images are generated from single view via diffusion. This is followed by multi-view reconstruction, e.g., via neural radiance fields (NeRF) [6, 28, 42], to obtain 3D outputs. However, multi-view consistency is a reoccurring challenge (see Fig. 1), especially over disparate views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 493, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 601 + ], + "type": "text", + "content": "We introduce a novel approach to single-view 3D reconstruction which is occlusion-revealing (Fig. 2) by first predicting multi-slice images from a single-view input and then integrating the slices into a 3D model; see Fig. 3. Fig. 4 shows slice images provided during training along a set of slicing directions. Each slice represents the projected image, in the input camera view, of one slice of the 3D mesh model. For the TV example shown, a completely occluded leg, in the camera view, is revealed in the slice images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "content": "The motivations behind our multi-slice reconstruction are three-fold. First, slicing through a volume or 3D mesh is a natural way to reveal occluded parts. Second, with the slice images forming an intermediate representation, our approach breaks the difficult problem of mapping from single view to 3D into two comparatively simpler tasks: slice prediction followed by 3D reconstruction from slices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "More importantly, when contrasting our multi-slice representation with the popular multi-view approach for occlu" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "text", + "content": "9882" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 275, + 144 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 275, + 144 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 275, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 275, + 144 + ], + "type": "image", + "image_path": "2dd1e33200e1c1aacc6cbde8ed9088be2e1712e62e14b04920e0b3203080b985.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 152, + 289, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 289, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 289, + 229 + ], + "type": "text", + "content": "Figure 5. Singe-view reconstruction of severely occluded parts in a lamp (e.g., the bulb) by Slice3D vs. One-2-3-45 [28], a state-of-the-art multi-view approach. The multi-view images in [28] were obtained by Zero-1-to-3 [29], which was trained on renderings of 800K 3D models [10]. It is conceivable that with a limited number of views (12, the same as our slice count), the bulbs were barely visible due to occlusion by the lamp shades and bases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 403 + ], + "type": "text", + "content": "sion revelation, we observe that slicing in the limit (i.e., with infinitely many slices) is guaranteed to unveil all hidden structures, while the same cannot be said about multi-view captures, as illustrated in Fig. 2 (a-b), for a 2D lamp example. In practice, when slice and view counts are both limited, the slicing approach is more occlusion-revealing in general since it can peel through any occluders without any obstruction, unlike the multi-view case; see (c-d). On the technical front, neither our slice prediction step nor the 3D reconstruction step is concerned with camera view changes or multi-view consistency, which facilitates the use of convolutional features. In Fig. 5, we show the above contrasting characteristics of the multi-slice vs. multi-view approaches manifested in a real 3D lamp reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "type": "text", + "content": "As outlined in Fig. 3, our slice images can be produced by either a regressive network or a generative one, both employing a U-Net architecture, with network losses measured against ground-truth (GT) images obtained by slicing 3D models from the training set; see Fig. 4. In the regressive network, we inject a learnable slice indicator code into the encoder-decoder setup to designate each decoded image into the proper spatial slice location. The slice generator is a denoising diffusion model operating on separate channels for different slice images. In the second stage, 3D reconstruction from the obtained slice images, we design a coordinate-based transformer to obtain an implicit field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "text", + "content": "Our network is coined Slice3D. We conduct extensive experiments on ShapeNet [5] and Objverse [10] for evaluation. Quantitative and qualitative comparisons are made to representative methods including encoder-decoders capturing local image features [65], autoregressive latent models [36], generative model inversion [42], and three recent methods with diffusion + NeRF [6, 28, 33], demonstrating superiority of our method in terms of reconstruction quality, generalizability, and handling of ambiguities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 132, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 132, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 132, + 681 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "The earliest inspiration for our multi-slice 3D reconstruction idea was GANHopper by Lira et al. [26]. In the con" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "text of unsupervised image-to-image translation, the key idea of their work is to gradually execute the translation between disparate domains by making small \"hops,\" rather than completing it in one go. We were then motivated to tackle the single-view 3D reconstruction problem as a domain translation task, hopping through the multi-slice images so as to bridge the gap between 2D and 3D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 158, + 546, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 546, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 546, + 302 + ], + "type": "text", + "content": "Neural fields. Neural implicit functions [63] are learned by a neural network to predict such properties as color, occupancy, opacity, and surface normals. Earlier works such as IM-Net [8], occupancy networks [34], and DeepSDF [41] represent a 3D shape as a latent code and send it to a decoder formed by MLPs to predict occupancy or SDF values for any query point sampled in 3D. NeRF [35] and its variants [15] represent a 3D scene using a neural radiance field for novel view synthesis. Many works on neural 3D reconstruction [15, 63] take on single- or multi-view images as well as point clouds as input. In contrast, the reconstruction module in Slice3D is built on multi-slice images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 519 + ], + "type": "text", + "content": "Single-view reconstruction from image features. An early work to exploit neural image features for 3D reconstruction was Pixel2Mesh [59], which uses a graph representation. Subsequently, neural implicit fields have been widely adopted for single-view reconstruction. IM-Net [8] first extracts a global latent code, from an input view, to learn a 3D implicit field. Auto-SDF [36] and ShapeFormer [66] employ vector quantization (VQ) [57] to improve the representation of latent features. More recent works [22, 37, 68] consider single-view reconstruction as a 3D generation task conditioned on the latent code of the input view. However, a global latent code is insufficient to capture finer details on a 3D object, often leading to overly smoothed or retrieved outcomes that do not faithfully represent the input. To alleviate this problem, both DISN [65] and " + }, + { + "bbox": [ + 304, + 304, + 546, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{IM}" + }, + { + "bbox": [ + 304, + 304, + 546, + 519 + ], + "type": "text", + "content": " -Net [25] opted to learn local images features, along with global ones, to improve the recovery of fine-grained topological (e.g., thin) features and surface details, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": "3D reconstruction via multi-view. Conventional 3D reconstruction approaches [14] involve multi-view reasoning. With recent advances in generative AI, a prevailing practice in single-view 3D reconstruction is to first employ generative models such as GANs [16, 23] and diffusion [11, 20, 45, 50] to generate several novel views from the input view, and then feed them to a multi-view 3D reconstruction module such as NeRF. NeRF-based surface reconstruction [39, 60, 67] can be further employed to output 3D meshes. EG3D [3] and NeRF-from-Image [42] build NeRFs or its variants upon the outputs of a style-GAN [23], while more recent works [4, 6, 18, 28-30, 33, 42, 43, 51, 62, 64] are favoring diffusion models such as Stable Diffusion (SD) [45]. To date, Zero-1-to-3 [29] represents the state of the art on single- to multi-view synthesis; it is built on SD and fine-tuned by 800K 3D models from Objaverse [10]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "text", + "content": "9883" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "To alleviate the multi-view consistency problem, which is an reoccurring challenge for the image-space generative methods, recent methods [30, 31, 43, 52, 71] aim to enhance the consistency by performing spatial attention between different views, which dramatically increases the computational overhead. Nevertheless, the consistency of two disparate views is still far from satisfactory. A deeper reason is that standard convolutional neural networks (CNNs) are not inherently rotation-invariant, making it difficult to build correspondences between significantly different views." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 386 + ], + "type": "text", + "content": "Slice representation. There have been techniques to use object slices to reconstruct a 3D shape, e.g., as inspired by Magnetic Resonance Imaging (MRI). Early works in computer graphics employ either parallel cross-sectional curves, e.g., [24], or non-parallel curve networks, e.g., [27], as inputs. Recently, with the advent of machine learning, OREx [48] and Cut-and-Approximate [40] develop neural reconstruction models to obtain 3D meshes, also from a sparse collection of 2D planar cross-sections, while Neural-Slice [21] uses slicing to obtain 3D shapes from 4D tetrahedral meshes. However, none of these techniques, nor others that we are aware of, address the problem of single-view 3D reconstruction since their inputs all consist of 3D data such as curve slices or point clouds. Moreover, our method learns to predict much fewer slice images and use them as input to a transformer-based reconstruction module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 388, + 288, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 485 + ], + "type": "text", + "content": "Dip. Dip transform [2] is an ingenious way to acquire occluded 3D structures by measuring fluid displacement when an object is repeatedly oriented and dipped into a fluid. Similar to slicing, Dip was also designed to \"reach\" occluded parts that optical captures via lines of sight cannot. However, both the occlusion revelation (liquid penetration vs. slicing) and target applications (acquisition vs. single-view reconstruction) are completely different." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 487, + 289, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 289, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 289, + 644 + ], + "type": "text", + "content": "Multi-plane image (MPI). Introduced by Zhou et al. [72], an MPI representation for a scene consists of a set of frontoparallel planes at fixed depths, with each image intended to capture scene appearance and visibility at a specific depth. Conceptually, both MPs and our slice images provide a layered scene representation for occlusion revelation. However, MPs [13, 19, 56, 69, 72] have only been employed for novel view synthesis, not 3D reconstruction, and the prediction networks were trained on multi-view images, e.g., from Youtube videos [72]. The generated MPs only reveal a small amount of occluded scenes to account for view shifts. In contrast, our method directly cuts through a 3D object to reveal all the occluded parts at several depth ranges." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 656, + 104, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 104, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 104, + 669 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "Our method is based on a slicing representation; see Section 3.1 and Fig. 4. Given a single-view RGB image " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " of a 3D object, we first learn a slicing network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": "(Section 3.2) which takes " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " as input and outputs its corresponding multi-slice images. Then, an SDF prediction network is trained (Section 3.3) to leverage image features from the slices to produce an implicit field of the object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 127, + 434, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 434, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 434, + 140 + ], + "type": "text", + "content": "3.1. Slicing Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": "As shown in Fig. 4, given a 3D object mesh, we slice it along the " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": " axes of its bounding box and render the sliced segments. The slice images can optionally include textures or have holes filled. By default, we remove the textures and keep the holes, which could reveal back faces of the mesh, and overall, it produces the best experimental results. The rendered slice images consist of " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "\\{S_i^j | 1 \\leq i \\leq N_s, j \\in \\{X,Y,Z\\}\\}" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "S_i^j \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": " denotes the axis and " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": " is the number of slices along each axis. In our current implementation, we produce " + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "inline_equation", + "content": "N_s = 4" + }, + { + "bbox": [ + 304, + 145, + 547, + 314 + ], + "type": "text", + "content": " slices per direction. Compared to planar cross-sections or tomography scans which consist of densely stacked planar images, our slicing representation is more compact with each slice image revealing more object semantics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 319, + 481, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 481, + 333 + ], + "type": "text", + "content": "3.2. Slice Regression and Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "text", + "content": ", observing a 3D shape from a single view, we aim to train a slicing network " + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "text", + "content": " that can produce the slice images from " + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "inline_equation", + "content": "F_{s}(I) = \\{\\hat{S}_{i}^{j}\\}" + }, + { + "bbox": [ + 304, + 338, + 547, + 447 + ], + "type": "text", + "content": ". We have two different designs for such a network: regressive and generative. The former aims to regress GT slices, assuming deterministic occluded parts based on the partial observation of the input view. The latter aims to generate slices that are consistent with the input view, considering various possible occluded parts that do not necessarily match the GT slices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 460, + 444, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 444, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 444, + 472 + ], + "type": "text", + "content": "3.2.1 Regression-based Slicing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": "We employ a U-Net [46] encoder-decoder to regress the slices from " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": ". As shown in the top-left of Fig. 3, an indicator code " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "E_{i}^{j}\\in \\mathbb{R}^{N_{e}}" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " is concatenated with the latent code of " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " to indicate which slice to output: " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\hat{S}_i^j = F_d([F_e(I);E_i^j])" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "F_{e}" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " are the encoder and decoder of " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " denotes concatenation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "text", + "content": "The indicator codes " + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\{E_i^j\\}" + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "text", + "content": " are randomly initialized and jointly optimized with the whole network. The loss function " + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "text", + "content": " is the reconstruction error measured by the " + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "text", + "content": "-Norm and perceptual loss [70] " + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_p)" + }, + { + "bbox": [ + 304, + 552, + 545, + 601 + ], + "type": "text", + "content": " for all slices:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 608, + 545, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 608, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 340, + 608, + 545, + 638 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\frac {\\sum_ {i , j} \\left| \\left| \\hat {S} _ {i} ^ {j} - S _ {i} ^ {j} \\right| \\right| _ {1} + \\mathcal {L} _ {p} \\left(\\hat {S} _ {i} ^ {j} , S _ {i} ^ {j}\\right)}{| S |}, \\tag {1}", + "image_path": "35fdb4224823e2c5c425e5cfb0f139f172b62b6baae07706d96edcc06e61e41e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 644, + 501, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 501, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 501, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 644, + 501, + 657 + ], + "type": "inline_equation", + "content": "|S| = 3 \\cdot N_{s}" + }, + { + "bbox": [ + 306, + 644, + 501, + 657 + ], + "type": "text", + "content": " is the total number of slices." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 670, + 447, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 447, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 447, + 684 + ], + "type": "text", + "content": "3.2.2 Generation-based Slicing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "Single-view occlusion revelation is clearly ill-posed, with potential ambiguities leading to multiple plausible 3D re" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "text", + "content": "9884" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 286, + 203 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 286, + 203 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 286, + 203 + ], + "type": "image", + "image_path": "e97b9d5609c4e2ba6cd05f620fc452a8ffe9a50712b29d667405001a4dedbe2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 287, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 235 + ], + "type": "text", + "content": "Figure 6. The pipeline of generation-based slicing via DDPM. " + }, + { + "bbox": [ + 46, + 212, + 287, + 235 + ], + "type": "inline_equation", + "content": "\\bigoplus" + }, + { + "bbox": [ + 46, + 212, + 287, + 235 + ], + "type": "text", + "content": " denotes the injection of noise to slice images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 286, + 289 + ], + "type": "text", + "content": "constructions. To this end, we consider slicing as an image-conditioned generation task and develop a generation-based slicing via Denoising Diffusion Probabilistic Models (DDPM) [20], whose detailed pipeline is shown in Fig. 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "text", + "content": "The key challenge of applying DDPM lies in how to generate consistent slices. Rather than diffusing each single slice, we propose to concatenate all slices into an aggregated structure " + }, + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "inline_equation", + "content": "[S]" + }, + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "text", + "content": ", along either the color dimension " + }, + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "inline_equation", + "content": "([S] \\in \\mathbb{R}^{H \\times W \\times |S|})" + }, + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "text", + "content": " or a spatial dimension such as height " + }, + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "inline_equation", + "content": "([S] \\in \\mathbb{R}^{(|S| \\cdot H) \\times W \\times 1})" + }, + { + "bbox": [ + 46, + 290, + 286, + 409 + ], + "type": "text", + "content": ". The concatenation of slices instructs DDPM to learn the correlation of slices, yielding consistent generation results. By default, we concatenate along the color dimension. The comparison of these two concatenation ways can be found in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": "Following DDPM, we first inject a noise " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "[S]" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "[\\bar{S}]" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": ". The noised slices " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "[\\bar{S}]" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " are then sent to a U-Net based denoising network " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "F_{dn}" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " that outputs the predicted noise conditioned on " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = F_{dn}([\\bar{S}], I)" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": ". In practice, the injection and prediction of noise " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " involve an iteration step " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": ", which we omit here for brevity. To better condition " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " for denoising diffusion, we employ a pre-trained VGG to extract multi-layer features from " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " and add them into the corresponding layers in the encoder of " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "F_{dn}" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": " (see the blue part of Fig. 6). The loss function is the denosing loss: " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gen} = ||\\hat{\\epsilon} - \\epsilon||_2^2" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": ". After training, we can sample random noises to generate a set of slice images " + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\{\\hat{S}_i^j\\}" + }, + { + "bbox": [ + 46, + 410, + 286, + 554 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 286, + 601 + ], + "type": "text", + "content": "Compared to regression-based slicing, generation-based slicing can produce multiple plausible outputs rather than a single deterministic result, but requires much longer inference time due to the nature of diffusion models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 610, + 234, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 234, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 234, + 624 + ], + "type": "text", + "content": "3.3. Learning Implicit Field from Slices" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "We learn an implicit field from slice images, by projecting a 3D query point onto each slice image, extracting query-specific features, and then aggregating them to predict a signed distance. For regression-based slicing, we train the SDF prediction network on the predicted slices from the previous section thus the whole pipeline can be trained end-to-end. For generation-based slicing, we train the SDF pre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "diction network on GT slice images since the generated slices may not match the GT signed distances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": "Specifically, for a query point " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ", we project it to a location " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\tilde{q} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " on the 2D image plane according to the camera parameters " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ". For slice " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\hat{S}_i^j" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " (regression) or " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "S_i^j" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " (generation), we first extract its feature maps, then retrieve features from each feature map corresponding to location " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\tilde{q}" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ", and finally concatenate them to obtain the image feature vector " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "V_i^j \\in \\mathbb{R}^{1 \\times 1 \\times N_c}" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ". The retrieval is accomplished via bi-linear interpolation over the features grids that encompass the location of " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\tilde{q}" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ". We design a transformer network [58] " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " which learns to aggregate the feature vectors from all slices to predict the signed distance " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\hat{d} = F_t([q', V_1^X, ..., V_{N_s}^Z])" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "q' \\in \\mathbb{R}^{N_c}" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " is linearly projected from " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " to be the same dimension with " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "V_i^j" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ". The optimization objective is to minimize the " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " distance between predicted and GT signed distance: " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{imp} = ||\\hat{d} - d||_1" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": ". In regression-based slicing, " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{imp}" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 95, + 545, + 289 + ], + "type": "text", + "content": " are jointly optimized." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": "DISN [65] extracts from only the input view " + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": " to obtain a feature vector " + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": ". For query points located around occluded parts of the object (e.g., the query point in Fig. 3), " + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": " is irrelevant to " + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": " since it is extracted from the front parts that block " + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": ", often leading to erroneous SDF prediction. In contrast, one or more of our slice images could reveal the occluded parts, yielding more accurate features fed to the transformer for SDF prediction. After learning the implicit field, we apply Marching Cubes [32] to obtain triangle meshes. Estimating the extrinsics (camera poses) of " + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": " can output 3D shapes in canonical poses, which is optional and applied only on ShapeNet [5]; See more details in Supplemental." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 442, + 440, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 440, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 440, + 454 + ], + "type": "text", + "content": "4. Results and Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": "Dataset. We utilize two well-known datasets, ShapeNet [5] and Objaverse [10], to train and test our network. For ShapeNet, our method is applied to Chairs and Airplanes separately, using the train/validation/test split as in SHREC16 [47]. Objaverse is a challenging dataset, where many objects may not belong to a clear category. In our experiment, we randomly selected " + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\approx 40\\mathrm{K}" + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": " shapes from Objaverse to keep resource and time costs manageable, while ensuring diversity of the collection. We follow a " + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "95\\% -5\\%" + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": " train-test split over the 40K selected shapes, ensuring that every training model we use is also from the training set of Zero-1-2-3 [29]. Note that Zero-1-2-3 used a train-test split of " + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "99\\% -1\\%" + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": " over the entire 800K models in Objaverse." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Rendering Slice Images. We employ the bisect function in Blender [1] to cut an object mesh into slices and then render them. By default, we keep internal cavities and holes produced by slicing. Since many 3D models in Objaverse are not in, or have no, canonical poses and feature random rotations, slicing them along the default " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " directions is not suitable. To this end, we opted for camera-aligned slicing instead. Specifically, we set the camera position, in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "text", + "content": "9885" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 282, + 191 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 282, + 191 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 282, + 191 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 282, + 191 + ], + "type": "table", + "html": "
Datasets MetricsShapeNet ChairsShapeNet Airplanes
CD↓F1↑HD↓CD↓F1↑HD↓
DISN [65]3.252.347.843.482.817.62
Auto-SDF [36]25.00.7715.8---
NeRF-Img [42]19.51.2713.416.01.4114.0
One-2-3-45 [28]12.51.0111.723.40.7512.6
Ours2.802.436.752.992.836.45
DISN* [65]3.022.427.863.192.967.56
SSDNeRF* [6]8.982.0610.910.42.0010.4
Ours*2.672.536.702.812.966.38
", + "image_path": "9baad35e8564fd3567393bb8957ef0dac28fd2b4ed4bc73f6f76ad3db8203e8d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 194, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 228 + ], + "type": "text", + "content": "Table 1. Quantitative results of single-view 3D reconstruction on ShapeNet Chairs and Airplanes. \\* denotes using ground-truth (GT) camera poses. Slice3D (ours) uses regression-based slicing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 233, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 281 + ], + "type": "text", + "content": "troduce a random rotation to the object, position it in front of the camera, and utilize the camera's axes for slicing. Under this setting, estimating the rotation of objects is unnecessary; see more details in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "content": "Implementation Details. When preparing training data, we use the Vega FEM Library [49] for SDF computation, same as DISN [65] and AutoSDF [36]. For each object, the diagonal length of the object's 3D bounding box is normalized to 1. We set " + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "content": " to 128, " + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "N_e" + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "content": " to 128, " + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "content": " to 4, and " + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "content": " to 1,024. We train our model on only one Nvidia A40 GPU. Evaluation Metrics. We use Chamfer " + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 281, + 287, + 413 + ], + "type": "text", + "content": " Distance (CD), F-score%1 (F1) [54], and Hausdorff Distance (HD) to evaluate reconstruction results. Besides these 3D metrics, we also evaluate using 2D metrics, PSNR, SSIM [61], and LPIPS [70], on 24 rendered views of reconstructed meshes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 420, + 264, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 264, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 264, + 433 + ], + "type": "text", + "content": "4.1. Comparisons to State Of The Art (SOTA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": "We compare our method with SOTA methods including DISN [65], AutoSDF [36], NeRF-From-Img [42], SSD-NeRF [6], RealFusion [33], and One-2-3-45 [28]. Note that when showing visual results, we always present two views to remove any view bias towards the input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "type": "text", + "content": "ShapeNet. Figs. 7 and 8 demonstrate the effectiveness of our approach in reconstructing obscured sections of objects while precisely restoring visible ones. In contrast, DISN [65] can barely reconstruct the occluded parts. AutoSDF [36] and NeRF-From-Image [42] rely on latent features for reconstruction, which can yield rather clean results but they do not necessarily respect the input image, as the chair results from Fig. 7 evidently showed. SSDNeRF [6] tends to produce erroneous geometric details and the surfaces are not as smooth as from other methods, especially for the airplane examples in Fig. 8. Generally speaking, One-2-3-45 [28] yields rough and bulky geometries that clearly lack details. Additionally, our method outperforms the other methods in the reconstruction of unconventional inputs, as shown by the two airplanes in Fig. 8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Tab. 1 shows quantitative comparisons to the above methods on the two ShapeNet categories. When our method operates without using GT camera poses and relies on esti" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 129 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 129 + ], + "type": "table", + "html": "
MethodCD↓F1↑HD↓PSNR↑SSIM↑LPIPS↓
DISN [65]28.21.4723.021.98.7915.2
One-2-3-45 [28]35.40.8729.819.88.5519.8
Ours (G)25.01.5116.422.38.8413.4
Ours (R)16.61.5313.422.68.8812.8
", + "image_path": "8a5ef2d94bde0cde0d504fd1906ba4fe2ebae112897b8af1daebabc37a53579a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 133, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 167 + ], + "type": "text", + "content": "Table 2. Quantitative results of single-view 3D reconstruction on the Objaverse dataset. 'R' and and 'G' denote regression-based and generation-based slicing in Slice3D, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 174, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 257 + ], + "type": "text", + "content": "mated poses, its performance sees only a marginal degradation. Notably, for Chair, our method with estimated poses even outperforms DISN and SSDNerf with GT poses. It is also notable that One-2-3-45 tends to output dilated shapes. Although it falls behind DISN in quantitative results, it usually generates more complete shapes and more accurate topologies than DISN; see Figs. 7 and 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "text", + "content": "Objaverse. Since AutoSDF [36], NeRF-From-Image [42], and SSDNeRF [6] are all category-specific methods, they cannot be trained on Objaverse which consists of many 3D objects with no clear categories. Thus, we mainly compare our method with DISN [65], RealFusion [33], and One-2-3-45 [28]; see Fig. 9 for a qualitative comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 330, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 474 + ], + "type": "text", + "content": "Note that RealFusion [33] and One-2-3-45 [28] both leverage a significantly larger volume of training data compared to our method, as they both rely on a pre-trained Stable Diffusion [45] model. Despite this disparity in the scale of the training sets, our method still outperforms these two methods, as well as DISN; see Fig. 9. Quantitatively, Tab. 2 demonstrates our method's across-the-board advantage over DISN and One-2-3-45 (more so on Objaverse than on ShapeNet), in both 3D and 2D metrics, especially CD, HD, and LPIPS. These results also show that our method possesses the capacity to generalize across common shapes rather than being confined to specific categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "text", + "content": "Real-world Image Input. To further stress-test the methods, we perform Google image search on object keywords (e.g., \"chair\", \"airplane\", etc.) and take the top-1 image with occlusions as input. In addition, we use some examples from Google Scanned Objects (GSO) [12] dataset. Fig. 10 shows that our method performs the best in recovering occluded parts while preserving fine details, e.g., the turbines of the airplane, the heel, and the opening of the boot." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "Multi-slice vs. Multi-view Reconstruction. Fig. 11 compares between multi-slice and multi-view reconstructions. We demonstrate two different sets of multi-slices and multi-views generated from our method and One-2-3-45 [28]. Our generated slice images depict two distinct configurations of chair legs, maintaining high consistency across slices. In contrast, the multiple views synthesized in One-2-3-45 do encounter inconsistency issues, compromising an accurate 3D reconstruction. We show three slices in each set and more can be found in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Training and Inference Costs. Unlike RealFusion [33] and One-2-3-45 [28], our method does not depend on a large" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "text", + "content": "9886" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 542, + 181 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 542, + 181 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 542, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 542, + 181 + ], + "type": "image", + "image_path": "ee668f36d8e41fa6c5b91a23e97fd7c60e3563ef640863aec0b850e45831ac2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 547, + 210 + ], + "type": "text", + "content": "Figure 7. Visual comparison between single-view 3D reconstruction methods on two ShapeNet chairs. DISN and our method (based on regressive slicing) utilize the same estimated camera parameters. Two different views are displayed to remove view bias." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 214, + 541, + 300 + ], + "blocks": [ + { + "bbox": [ + 53, + 214, + 541, + 300 + ], + "lines": [ + { + "bbox": [ + 53, + 214, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 541, + 300 + ], + "type": "image", + "image_path": "d6c6bb659619caf5510ee89c559132ebff623f99fb567beade6977155083b4ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 321, + 541, + 453 + ], + "blocks": [ + { + "bbox": [ + 47, + 306, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 306, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 545, + 319 + ], + "type": "text", + "content": "Figure 8. Visual comparison between single-view 3D reconstruction methods on two somewhat unconventional airplanes from ShapeNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 321, + 541, + 453 + ], + "lines": [ + { + "bbox": [ + 53, + 321, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 321, + 541, + 453 + ], + "type": "image", + "image_path": "425773c5acaf5dc177562335accb4b537a6e2384cce327a00aa5a6ba65223e67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 495, + 284, + 627 + ], + "blocks": [ + { + "bbox": [ + 46, + 460, + 546, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 546, + 483 + ], + "type": "text", + "content": "Figure 9. Visual comparison between cross-category reconstruction methods on Objverse. Our method better recovers overall geometries, fine details, and is more faithful to the inputs. Except for the seaweed (squared), all the other models are in the training set of Zero-1-to-3." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 495, + 284, + 627 + ], + "lines": [ + { + "bbox": [ + 52, + 495, + 284, + 627 + ], + "spans": [ + { + "bbox": [ + 52, + 495, + 284, + 627 + ], + "type": "image", + "image_path": "c25a9ef23af71e990c63a87a8bccf7cda25dc94b04bae448cc84138ab403f663.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 633, + 287, + 656 + ], + "lines": [ + { + "bbox": [ + 46, + 633, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 633, + 287, + 656 + ], + "type": "text", + "content": "Figure 10. On real-world image inputs. Top two from Google image search and the others from GSO [12]. Zoom in for details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "pre-trained model [45]. NeRF-based methods tend to have long inference times, as many of them require per-test optimization. Our method has much faster inference times: 10 or 20 seconds with slice regression or generation, com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "type": "text", + "content": "pared to 90 minutes for RealFusion and 45 seconds for One-2-3-45 and SSDNeRF, while delivering superior results. With similar inference times to DISN, Auto-SDF, and NeRF-From-Img, our method has better reconstruction quality. See supplementary material for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 571, + 400, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 400, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 400, + 585 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 592, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 547, + 700 + ], + "type": "text", + "content": "Generative vs. Regressive Slicing. As shown in the first row of Fig. 11, for generation-based slicing, we can produce multiple plausible results for a given input view. Tab. 2 shows that regressive slicing outperforms generative slicing when compared to GT. However, it is notable that the 3D outputs may differ from GT when input views have ambiguous occluded parts. Regressive training struggles to handle such one-to-many mappings, leading to generation outperforming regression in reconstructing occluded parts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 546, + 713 + ], + "type": "text", + "content": "Number of Slices. We test the performance of our method" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 756 + ], + "type": "text", + "content": "9887" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 544, + 181 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 544, + 181 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 544, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 544, + 181 + ], + "type": "image", + "image_path": "c32affa8f67f8e1d6326c19515f6cb14cac8392d6d72ea628c003e98c11de202.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "type": "text", + "content": "Figure 11. Multi-slice vs. multi-view reconstructions amid ambiguities in the chair legs. Both One-2-3-45 (bottom) and Slice3D (top) can produce multiple results. Our results are both plausible from consistent slices, while One-2-3-45 suffers from multi-view inconsistencies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 224, + 282, + 289 + ], + "blocks": [ + { + "bbox": [ + 50, + 224, + 282, + 289 + ], + "lines": [ + { + "bbox": [ + 50, + 224, + 282, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 224, + 282, + 289 + ], + "type": "table", + "html": "
SettingCD↓F1↑HD↓SettingCD↓F1↑HD↓
One-hot ind.2.682.536.82only X2.882.397.74
DISN+Depth2.992.437.81only Y2.752.456.84
Ns=22.902.467.31only Z3.082.387.60
Ns=42.672.536.70Filled2.732.507.37
Ns=82.562.566.93w/ textures2.822.507.70
", + "image_path": "819a8da2a30ec7e1d68180c10824a9d910e689769f6814e97688107125be78bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 292, + 251, + 303 + ], + "lines": [ + { + "bbox": [ + 82, + 292, + 251, + 303 + ], + "spans": [ + { + "bbox": [ + 82, + 292, + 251, + 303 + ], + "type": "text", + "content": "Table 3. Ablation studies on ShapeNet Chairs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "content": "with regression-based slicing by varying " + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "content": " between 2, 4, and 8, on ShapeNet Chairs. From Tab. 3, it is evident that larger " + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "content": " leads to better reconstruction quality. However, " + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "inline_equation", + "content": "N_{s} = 8" + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "content": " will dramatically increase the computational cost of our method since its complexity is " + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "inline_equation", + "content": "O(N_{s}^{2})" + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "content": ", as pair-wise attentions are performed on the slices in the transformer. As a trade-off between reconstruction quality and computational complexity, we set " + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "inline_equation", + "content": "N_{s} = 4" + }, + { + "bbox": [ + 46, + 312, + 287, + 408 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 409, + 253, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 409, + 253, + 420 + ], + "spans": [ + { + "bbox": [ + 59, + 409, + 253, + 420 + ], + "type": "text", + "content": "Tab. 3 also covers several other ablation studies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 422, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 481 + ], + "type": "text", + "content": "Learned v.s. One-hot Slice Indicator. Learned indicator codes (the default setting) achieve slightly better performance than using one-hot codes to indicate slice locations. This can be seen by comparing numbers from the first half row of Tab. 3 to those in the last row of Tab. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "Slices v.s. Depth Images. Since our slice images offer layered information, it is logical to perform a comparison against depth images as supplementary information. We compared the performance of DISN when concatenating estimated depth images with the input view. Results are shown in the \"DISN+Depth\" half row, which indicates that depth images can only provide marginal improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": "Three Axes v.s. One Axis. We also investigated the option of slicing along only one axis. Not surprisingly, Tab. 3 shows that this underperforms against using three axes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 640 + ], + "type": "text", + "content": "Solid v.s Hollow Slices. Tab. 3 also demonstrates that not filling the holes improves the performance as more information about the inner/outer surfaces is revealed this way." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Using Textured Slices. We can optionally keep object textures in the slice images. Experimental results show that removing the textures in the slice images can achieve better reconstruction. The reasons could be: 1) learning to produce slice images without textures could help the model better distinguish geometries from textures; 2) for some 3D" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 226, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 226, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 226, + 545, + 251 + ], + "type": "text", + "content": "models the color of inner and outer surfaces could be different which may complicate the learning process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 261, + 468, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 468, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 468, + 274 + ], + "type": "text", + "content": "5. Discussion and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": "Our multi-slice idea represents a significant departure from conventional approaches to single-view 3D reconstruction. Through conceptual illustrations (Fig. 5) and extensive experiments, we demonstrate its compelling properties in recovering occluded object structures. Note again that all the Slice3D results on Objaverse [10] were obtained by networks trained on a very small subset (40K or " + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": ") of the 800K 3D models therein, whereas One-2-3-45 [28] utilizes Zero-1-to-3 [29] for multi-view synthesis and the latter was fine-tuned on the entirety of the 800K. Yet our results (Table 2) show that Slice3D outperforms One-2-3-45 across the board, suggesting that our model generalizes better, likely owing to its superior occlusion-revealing abilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "text", + "content": "Our current implementation only uses four slices per direction, which is quite coarse. Since there is no camera rotation, parts of an object far away from the camera are in smaller scale, e.g., the rear leg of the chair in Fig. 10 and of the buck in Fig. 9. This can hinder our network's ability to extract detailed features. More refined and gradual slicing over those parts may be needed to improve matters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": "The foremost limitation of slicing however, is that realistically, it can only be executed at scale on digital 3D models, not physical ones. On the other hand, the multi-view approach can be applied to physical 3D objects by capturing photos only, without ever needing 3D supervision, as in RealFusion [33]. However, this is not without a high cost: 90 minutes by RealFusion vs. 45 seconds by One-2-3-45 (requiring 3D fine-tuning for multi-view synthesis) and less than 20 seconds by Slice3D. We envision synthesizing 3D models, not many but carefully crafted few, for the multislice approach as an intriguing direction to explore." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 664, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. We thank all the anonymous reviewers and area chairs for their valuable comments. We thank Daniel Cohen-Or and Xingguang Yan for helpful discussions. This work was supported in part by NSERC." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "text", + "content": "9888" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Bisect - blender manual. https://docs.blender.org/manual/en/2.80/modeling/meshes/editing/subdividing/bisect.html. Accessed: 2023-11-15.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "text", + "content": "[2] Kfir Aberman, Oren Katzir, Qiang Zhou, Zegang Luo, Andrei Sharf, Chen Greif, Baoquan Chen, and Daniel Cohen-Or. Dip transform for 3D shape reconstruction. ACM Transactions on Graphics, 36(4):Article 79, 2017. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[4] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 345 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "text", + "content": "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion NeRF: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714, 2023. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "type": "text", + "content": "[7] Wenzheng Chen, Jun Gao, Huan Ling, Edward J. Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "text", + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 512 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In CVPR, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "text", + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 3, 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "text", + "content": "[12] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[13] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[14] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 131, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 131, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 131, + 545, + 174 + ], + "type": "text", + "content": "[15] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 176, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 176, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 176, + 545, + 220 + ], + "type": "text", + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 222, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 222, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 308, + 222, + 545, + 265 + ], + "type": "text", + "content": "[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 267, + 545, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 267, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 308, + 267, + 545, + 333 + ], + "type": "text", + "content": "[18] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 335, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 335, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 545, + 379 + ], + "type": "text", + "content": "[19] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 380, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 380, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 545, + 414 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "text", + "content": "[21] Chenbo Jiang, Jie Yang, Shwai He, Yu-Kun Lai, and Lin Gao. Neuralslice: Neural 3d triangle mesh reconstruction via slicing 4d tetrahedral meshes. 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 450, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 450, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 450, + 545, + 483 + ], + "type": "text", + "content": "[22] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "type": "text", + "content": "[24] Shay Kels and Nira Dyn. Reconstruction of 3d objects from 2d cross-sections with the 4-point subdivision scheme adapted to sets. Computers & Graphics, 35:741-746, 2011. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "text", + "content": "[25] Manyi Li and Hao Zhang. " + }, + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{IM}" + }, + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "text", + "content": " -Net: Learning detail disentangled implicit fields from single images. In CVPR, 2021. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 599, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 599, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 599, + 545, + 633 + ], + "type": "text", + "content": "[26] Wallace Lira, Johannes Merz, Daniel Ritchies, Daniel Cohen-Or, and Hao Zhang. GANHopper: Multi-hop gan for unsupervised image-to-image translation. In ECCV, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "text", + "content": "[27] Lu Liu, Chandrajit Bajaj, Joseph O. Deasy, Daniel A. Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. Computer Graphics Forum, 27:155-163, 2008. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[28] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 316, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 316, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 316, + 755 + ], + "type": "text", + "content": "9889" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 1, 3, 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "text", + "content": "[30] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image, 2023. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 219 + ], + "type": "text", + "content": "[31] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 287, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 287, + 265 + ], + "type": "text", + "content": "[32] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353. 1998. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "type": "text", + "content": "[33] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In CVPR, pages 8446-8455, 2023. 3, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 313, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 287, + 357 + ], + "type": "text", + "content": "[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 359, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 287, + 413 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "type": "text", + "content": "[37] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 508, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 287, + 540 + ], + "type": "text", + "content": "[38] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[39] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, pages 5589-5599, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 588, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 621 + ], + "type": "text", + "content": "[40] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "text", + "content": "[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165–174, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] Dario Pavllo, David Joseph Tan, Marie-Julie Rakotosaona, and Federico Tombari. Shape, pose, and appearance from a single image via bootstrapped radiance field inversion. In CVPR, pages 4391-4401, 2023. 2, 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[43] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "text", + "content": "[44] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3d geometry via nested shape layers. In CVPR, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 285 + ], + "type": "text", + "content": "[46] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "text", + "content": "[47] Manolis Savva, Fisher Yu, Hao Su, M Aono, B Chen, D Cohen-Or, W Deng, Hang Su, Song Bai, Xiang Bai, et al. Shrec16 track: largescale 3d shape retrieval from shapenet core55. In Proceedings of the eurographics workshop on 3D object retrieval, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "text", + "content": "[48] Haim Sawdayee, Amir Vaxman, and Amit H. Bermano. Orex: Object reconstruction from planar cross-sections using neural fields. In CVPR, pages 20854-20862, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[49] Fun Shing Sin, Daniel Schroeder, and Jernej Barbič. Vega: non-linear fem deformable object simulator. In Computer Graphics Forum, pages 36-48. Wiley Online Library, 2013. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 474 + ], + "type": "text", + "content": "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "text", + "content": "[51] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[52] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv preprint arXiv:2307.01097, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "text", + "content": "[53] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In ICCV, 2017. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[54] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, pages 3405-3414, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[55] Maxim Tatarchenko, Stephan R. Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[56] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 315, + 755 + ], + "type": "text", + "content": "9890" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "text", + "content": "[59] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, pages 52-67, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "text", + "content": "[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "text", + "content": "[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "text", + "content": "[62] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 408 + ], + "type": "text", + "content": "[63] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Comput. Graph. Forum, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[64] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, pages 4479–4489, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[65] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 1, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[66] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In CVPR, pages 6239–6249, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 589 + ], + "type": "text", + "content": "[67] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[68] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. arXiv preprint arXiv:2301.11445, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[69] Mingfang Zhang, Jinglu Wang, Xiao Li, Yifei Huang, Yoichi Sato, and Yan Lu. Structural multiplane image: Bridging neural view synthesis and 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16707-16716, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 195 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "features as a perceptual metric. In CVPR, pages 586-595, 2018. 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "text", + "content": "[71] Minda Zhao, Chaoyi Zhao, Xinyue Liang, Lincheng Li, Zeng Zhao, Zhipeng Hu, Changjie Fan, and Xin Yu. Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion prior. arXiv preprint arXiv:2308.13223, 2023.4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[72] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 746, + 314, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 746, + 314, + 755 + ], + "spans": [ + { + "bbox": [ + 295, + 746, + 314, + 755 + ], + "type": "text", + "content": "9891" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_content_list.json b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18146152a015d965f7896235c96b60374d2c378a --- /dev/null +++ b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_content_list.json @@ -0,0 +1,1611 @@ +[ + { + "type": "text", + "text": "SlowFormer: Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers", + "text_level": 1, + "bbox": [ + 107, + 128, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "K L Navaneet *1", + "bbox": [ + 99, + 202, + 232, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Soroush Abbasi Koohpayegani *1", + "bbox": [ + 253, + 203, + 519, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of California, Davis", + "bbox": [ + 263, + 222, + 519, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Essam Sleiman *12†", + "bbox": [ + 537, + 203, + 694, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Harvard University", + "bbox": [ + 540, + 220, + 707, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hamed Piriaviash", + "bbox": [ + 717, + 203, + 867, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, there has been a lot of progress in reducing the computation of deep models at inference time. These methods can reduce both the computational needs and power usage of deep models. Some of these approaches adaptively scale the compute based on the input instance. We show that such models can be vulnerable to a universal adversarial patch attack, where the attacker optimizes for a patch that when pasted on any image, can increase the compute and power consumption of the model. We run experiments with three different efficient vision transformer methods showing that in some cases, the attacker can increase the computation to the maximum possible level by simply pasting a patch that occupies only $8\\%$ of the image area. We also show that a standard adversarial training defense method can reduce some of the attack's success. We believe adaptive efficient methods will be necessary in the future to lower the power usage of expensive deep models, so we hope our paper encourages the community to study the robustness of these methods and develop better defense methods for the proposed attack. Code is available at:", + "bbox": [ + 76, + 306, + 473, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/UCDvision/SlowFormer", + "bbox": [ + 76, + 608, + 465, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of deep learning has recently made significant progress in improving model efficiency for inference. Two broad categories of methods can be distinguished: 1) those that reduce computation regardless of input, and 2) those that reduce the computation depending on the input (adaptively). Most methods, such as weight pruning or model quantization, belong to the first category which reduces computation by a constant factor regardless of the input. However, in many applications, the complexity of the perception task may differ depending on the input. For example, when a self-driving car is driving between lanes in an empty street,", + "bbox": [ + 76, + 696, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the perception may be simpler and require less computation when compared to driving in a busy city street scene. Interestingly, in some applications, simple scenes such as highway driving may account for the majority of the time. Therefore, we believe that adaptive computation reduction will become an increasingly important research area in the future, especially when non-adaptive methods reach the lower bound of computation.", + "bbox": [ + 500, + 275, + 895, + 396 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We argue that reduction of compute usually reduces power usage, which is crucial, particularly in mobile devices that run on battery, e.g., AR/VR headsets, humanoid robots, and drones. For instance, increasing the size of the battery for a drone may lead to a drastic reduction in its range due to the increased battery weight. This is important since the improvement in battery technology is much slower than compute technology. For instance, the battery capacity from iPhone [1] (1st generation) in 2007 to iPhone 15 Pro Max in 2023 improved from 5.18 watt-hour to 17.32 watt-hour (less than 4 times) while the compute has increased by a much larger factor. As an example, a delivery robot like Starship uses a 1,200Wh battery and can run for 12 hours [2], so it uses almost 100 watts for compute and mobility. Hence, an adversary increasing the power consumption of the perception unit by 20 watts, will reduce the battery life by almost $20\\%$ , which can be significant. Note that 20 watts increase in power is realistic assuming that it uses two NVIDIA Jetson Xavier NX cards (almost 20 watts each).", + "bbox": [ + 496, + 397, + 895, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Key idea: Assuming that a perception method is reducing the computation adaptively with the input, an adversary can attack the model by modifying the input to increase the computation and power consumption. Our goal is to design a universal adversarial patch that when pasted on any input image, it will increase the computation of the model leading to increased power consumption. We believe this is an important vulnerability, particularly for safety-critical mobile systems that run on battery.", + "bbox": [ + 496, + 686, + 895, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Please note that in this paper, we do not experiment with real hardware to measure the power consumption. Instead, we report the change in FLOPs of the inference time assuming that the power consumption is positively correlated with the number of FLOPs, as studied in [73].", + "bbox": [ + 496, + 825, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 99, + 875, + 212, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done while he was an undergraduate student at UC Davis.", + "bbox": [ + 99, + 887, + 447, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24786", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08806d9dd9fae0c7d4e594a954d3ccfc9e86cf3bd11aa140d34f226b9a8cafc9.jpg", + "image_caption": [ + "Figure 1. Computation and Energy Attack on Vision Transformers: Given a pre-trained input-dependent computation efficient model, the adversary first attaches an adversarial patch to all images in a dataset and optimizes this patch with our method such that it maximizes the model's computation for each sample. During inference, the adversary modifies the input of the victim's model by applying the learnt patch to it. This results in an increase in compute in the victim's model. The attack will thus potentially slowdown and also lead to increased energy consumption and CPU/GPU usage on the victim's device." + ], + "image_footnote": [], + "bbox": [ + 76, + 92, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design our attack, SlowFormer, for three different methods (A-ViT [88], ATS [25], and Ada-VIT [63]) that reduce the computation of vision transformers. These methods generally identify the importance of each token for the final task and drop the insignificant ones to reduce the computation. We show that in all three cases, our attack can increase the computation by a large margin, returning it to the full-compute level (non-efficient baseline) for all images in some settings. Our threat model is agnostic to the accuracy of the model and attacks the computation and power consumption only. Figure 1 shows our attack.", + "bbox": [ + 75, + 393, + 472, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are some prior works that design a pixel-level perturbation attack to increase the compute of the model. We believe universal patch-based attacks that do not change with the input image (generalize from training data to test data) are much more practical in real applications. Note that to modify the pixel values on a real robot, the attacker needs to access and manipulate the image between the camera and compute modules, which is impossible in many applications.", + "bbox": [ + 75, + 560, + 472, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: We show that efficient vision transformer methods are vulnerable to a universal patch attack that can increase their compute and power usage. We demonstrate this through experiments on three different efficient transformer methods. We show that an adversarial training defense can reduce attack success to some extent.", + "bbox": [ + 75, + 680, + 470, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 784, + 218, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers: The popularity of transformers [79] in vision has grown rapidly since the introduction of the first vision transformer [20, 78]. Recent works demonstrate the strength of vision transformers on a variety of computer vision tasks [11, 18, 21, 56, 69, 77, 90, 93, 94, 96]. Moreover, transformers are the backbone of recent Self-Supervised", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning (SSL) models [12, 37], and vision-language models [67]. In our work, we design an attack to target the computation and energy efficiency of vision transformers.", + "bbox": [ + 496, + 393, + 893, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Vision Transformers: Due to the recent importance and popularity of vision transformers, many works have started to study the efficiency of vision transformers [7, 46, 89]. To accomplish this, some lines of work study token pruning with the goal of removing uninformative tokens in each layer [25, 62, 63, 68, 88]. ToMe [6] merges similar tokens in each layer to decrease the computation. Some works address quadratic computation of self-attention module by introducing linear attention [3, 45, 48, 58, 74]. Efficient architectures [40, 55] that limit the attention span of each token have been proposed to improve efficiency. In our paper, we attack token pruning based efficient transformers where the computation varies based on the input samples [25, 63, 88].", + "bbox": [ + 496, + 443, + 895, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Computation: There are different approaches to reducing the computation of vision models, including knowledge distillation to lighter network [39, 59], model quantization [54, 70] and model pruning [51]. In these methods, the computation is fixed during inference. In contrast to the above models, some works address efficiency by having variable computation based on the input. The intuition behind this direction is that not all samples require the same amount of computation. Several recent works have developed models that dynamically exit early or skip layers [5, 22, 28, 33, 34, 44, 76, 80, 82] and selectively activate neurons, channels or branches for dynamic width [4, 9, 13, 26, 31, 38, 43, 91] depending on the complexity of the input sample. Zhou et al. show that not all locations in an image contribute equally to the predictions of a CNN model [95], encouraging a new line of work to make CNNs more", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24787", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "efficient through spatially dynamic computation. Pixel-Wise dynamic architectures [10, 14, 23, 47, 71, 81, 85] learn to focus on the significant pixels for the required task while Region-Level dynamic architectures perform adaptive inference on the regions or patches of the input [29, 53]. Finally, lowering the resolution of inputs decreases computation, but at the cost of performance. Conventional CNNs process all regions of an image equally, however, this can be inefficient if some regions are \"easier\" to process than others [42]. Correspondingly, [86, 87] develop methods to adaptively scale the resolution of images.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformers have recently become extremely popular for vision tasks, resulting in the release of a few input-dynamic transformer architectures [24, 63, 88]. Fayyaz et al. [24] introduce a differentiable parameter-free Adaptive Token Sampler (ATS) module which scores and adaptively samples significant tokens. ATS can be plugged into any existing vision transformer architecture. A-ViT [88] reduces the number of tokens in vision transformers by discarding redundant spatial tokens. Meng et al. [63] propose AdaViT, which trains a decision network to dynamically choose which patch, head, and block to keep/activate throughout the backbone.", + "bbox": [ + 75, + 260, + 472, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Attack: Adversarial attacks are designed to fool models by applying a targeted perturbation or patch on an image sample during inference [32, 50, 75]. These methods can be incorporated into the training set and optimized to fool the model. Correspondingly, defenses have been proposed to mitigate the effects of these attacks [27, 52, 65, 84]. Patch-Fool [30] considers adversarial patch-based attacks on transformers. Some recent works [61, 83, 92] also study and design methods for the transferability of adversarial attacks on vision transformers. However, most prior adversarial attacks target model accuracy, ignoring model efficiency.", + "bbox": [ + 75, + 429, + 472, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Energy Attack: Very recently, there have been a few works on energy adversarial attacks on neural networks. In ILFO [35], Haque et al. attack two CNN-based input-dynamic methods: SkipNet [82] and SACT [28] using image specific perturbation. DeepSloth [41] attack focuses on slowing down early-exit methods, reducing their energy efficiency by $90 - 100\\%$ . GradAuto [64] successfully attacks methods that are both dynamic width and dynamic depth. NICGSlowDown and TransSlowDown [16, 17] attack neural image caption generation and neural machine translation methods, respectively. All these methods primarily employ image specific perturbation based adversarial attack. Sloth-Bomb injects efficiency backdoors to input-adaptive dynamic neural networks [15] and NodeAttack [36] attacks Neural Ordinary Differential Equation models, which use ordinary differential equation solving to dynamically predict the output of a neural network. Our work is closely related to ILFO [35], DeepSloth [41] and GradAuto [64] in that we attack the computational efficiency of networks. However, unlike these methods, we focus on designing an adversarial patch-based", + "bbox": [ + 75, + 598, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attack that is universal and on vision transformers. We additionally provide a potential defense for our attack. We use a patch that generalizes from train to test set and thus we do not optimize per sample during inference. Our patch-based attack is especially suited for transformer architectures [30].", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Computation and Energy Attack", + "text_level": 1, + "bbox": [ + 498, + 181, + 799, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Threat Model:", + "text_level": 1, + "bbox": [ + 500, + 207, + 648, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a scenario where the adversary has access to the victim's trained deep model and modifies its input such that the energy consumption and computational demand of the model is increased. The attack is agnostic to model accuracy. To make the setting more practical, instead of perturbing the entire image, we assume that the adversary can modify the input image by only pasting a patch [8, 72] on it and that the patch is universal, that is, image independent. During inference, a pretrained patch is pasted on the test image before propagating it through the network.", + "bbox": [ + 496, + 229, + 893, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we attack three state-of-the-art efficient transformers. Since the attacker manipulates only the input image and not the network parameters, the attacked model must have dynamic computation that depends on the input image. As stated earlier, several recent works have developed such adaptive efficient models and we believe that they will be more popular in the future due to the limits of non-adaptive efficiency improvement.", + "bbox": [ + 496, + 382, + 893, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Attack on Efficient Vision Transformers:", + "text_level": 1, + "bbox": [ + 498, + 513, + 848, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Universal Adversarial Patch: We use an adversarial patch to attack the computational efficiency of transforms. The learned patch is universal, that is, a single patch is trained and is used during inference on all test images. The patch generalizes across images but not across models. The patch optimization is performed only on the train set. The patch is pasted on an image by replacing the image pixels using the patch. We assume the patch location does not change from train to test. The patch pixels are initialized using i.i.d. samples from a uniform distribution over [0, 255]. During each training iteration, the patch is pasted on the mini-batch samples and is updated to increase the computation of the attacked network. The patch values are projected onto [0, 255] and quantized to 256 uniform levels after each iteration. Note that we use a pretrained network and do not update its parameters either in the training or in the evaluation of our attack. During inference, the trained patch is pasted on the test images and the computational efficiency of the network on the adversarial image is measured.", + "bbox": [ + 496, + 537, + 893, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we focus on three methods employing vision transformers for the task of image classification. All these methods modify the computational flow of the network based on the input image for faster inference. A pretrained model is used for the attack and is not modified during", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24788", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "our adversarial patch training. We first provide a brief background of each method before describing our attack.", + "bbox": [ + 75, + 90, + 472, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attacking A-ViT :", + "text_level": 1, + "bbox": [ + 76, + 136, + 207, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background: A-ViT [88] adaptively prunes image tokens to achieve speed-up in inference with minimal loss in accuracy. For a given image, a dropped token will not be used again in the succeeding layers of the network. Let $x$ be the input image and $\\{t^l\\}_{1:K}$ be the corresponding $K$ tokens at layer $l$ . An input-dependent halting score $h_k^l$ for a token $k$ at layer $l$ is calculated and the token is dropped at layer $N_k$ where its cumulative halting score exceeds a fixed threshold value $1 - \\epsilon$ for the first time. The token is propagated until the final layer if its score never exceeds the threshold. Instead of introducing a new parameter for $h_k^l$ , the first dimension of each token is used to predict the halting score for the corresponding token. The network is trained to maximize the cumulative halting score at each layer and thus drop the tokens earlier. The loss, termed ponder loss, is given by:", + "bbox": [ + 75, + 152, + 472, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p o n d e r}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(N _ {k} + r _ {k}\\right), \\quad r _ {k} = 1 - \\sum_ {l = 1} ^ {N _ {k - 1}} h _ {k} ^ {l} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 407, + 472, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, A-ViT enforces a Gaussian prior on the expected halting scores of all tokens via $KL$ -divergence based distribution loss, $\\mathcal{L}_{\\mathrm{distr}}$ . These loss terms are minimized along with the task-specific loss $\\mathcal{L}_{\\mathrm{task}}$ . Thus, the overall training objective is $\\mathcal{L} = \\mathcal{L}_{\\mathrm{task}} + \\alpha_d\\mathcal{L}_{\\mathrm{distr}} + \\alpha_p\\mathcal{L}_{\\mathrm{ponder}}$ where $\\alpha_{d}$ and $\\alpha_{p}$ are hyperparameters.", + "bbox": [ + 75, + 465, + 472, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attack: Here, we train the patch to increase the inference compute of a trained A-ViT model. Since we are interested in the compute and not task-specific performance, we simply use $-\\left(\\alpha_{d}\\mathcal{L}_{\\mathrm{distr.}} + \\alpha_{p}\\mathcal{L}_{\\mathrm{ponder}}\\right)$ as our loss. It is possible to preserve (or hurt) the task performance by additionally using $+\\mathcal{L}_{\\mathrm{task}}$ (or $-\\mathcal{L}_{\\mathrm{task}}$ ) in the loss formulation.", + "bbox": [ + 75, + 559, + 472, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attacking AdaViT:", + "text_level": 1, + "bbox": [ + 76, + 664, + 215, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Background:** To improve the inference efficiency of vision transformers, AdaViT [63] inserts and trains a decision network before each transformer block to dynamically decide which patches, self-attention heads, and transformer blocks to keep/activate throughout the backbone. The $l^{\\text{th}}$ block's decision network consists of three linear layers with parameters $W_{l} = W_{l}^{p}, W_{l}^{h}, W_{l}^{b}$ which are then multiplied by each block's input $Z_{l}$ to get $m$ .", + "bbox": [ + 75, + 680, + 470, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(m _ {l} ^ {p}, m _ {l} ^ {h}, m _ {l} ^ {b}\\right) = \\left(W _ {l} ^ {p}, W _ {l} ^ {h}, W _ {l} ^ {b}\\right) Z _ {l} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 811, + 470, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The value $m$ is then passed to sigmoid function to convert it to a probability value used to make the binary decision of keep/discard. Gumbel-Softmax trick [60] is used to make this decision differentiable during training. Let $M$ be the", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "keep/discard mask after applying Gumbel-Softmax on $m$ . The loss on computation is given by:", + "bbox": [ + 496, + 90, + 893, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u s a g e} = (\\frac {1}{D _ {p}} \\sum_ {d = 1} ^ {D _ {p}} M _ {d} ^ {p} - \\gamma_ {p}) ^ {2} + (\\frac {1}{D _ {h}} \\sum_ {d = 1} ^ {D _ {h}} M _ {d} ^ {h} - \\gamma_ {h}) ^ {2} \\\\ + \\left(\\frac {1}{D _ {b}} \\sum_ {d = 1} ^ {D _ {b}} M _ {d} ^ {b} - \\gamma_ {b}\\right) ^ {2} \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 128, + 893, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{p}, D_{h}, D_{b}$ represent the number of total patches, heads, and blocks of the entire transformer, respectively. $\\gamma_{p}, \\gamma_{h}, \\gamma_{b}$ denote the target computation budgets i.e. the percentage of patches/heads/blocks to keep. The total loss is a combination of task loss (cross-entropy) and computation loss: $\\mathcal{L} = \\mathcal{L}_{ce} + \\mathcal{L}_{usage}$ .", + "bbox": [ + 496, + 224, + 893, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attack: To attack this model, we train the patch to maximize the computation loss $\\mathcal{L}_{usage}$ . More specifically, we set the computation-target $\\gamma$ values to 0 and negate the $\\mathcal{L}_{usage}$ term in Eq. 3. As a result, the patch is optimized to maximize the probability of keeping the corresponding patch (p), attention head (h), and transformer block (b). We can also choose to attack the prediction performance by selectively including or excluding the $\\mathcal{L}_{ce}$ term.", + "bbox": [ + 496, + 316, + 893, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attacking ATS:", + "text_level": 1, + "bbox": [ + 498, + 450, + 612, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background: Given $N$ tokens with the first one as the classification token, the transformer attention matrix $\\mathcal{A}$ is calculated by the following dot product:", + "bbox": [ + 496, + 465, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{A} = \\mathrm{Softmax}\\left(\\mathcal{Q}\\mathcal{K}^T /\\sqrt{d}\\right)$ where $\\sqrt{d}$ is a scaling coefficient, $d$ is the dimension of tokens, $\\mathcal{Q},\\mathcal{K}$ and $\\nu$ are the query, key and value matrices, respectively. The value $\\mathcal{A}_{1,j}$ denotes the attention of the classification token to token $j$ . ATS [25] assigns importance score $S_{j}$ for each token $j$ by measuring how much the classification token attends to it:", + "bbox": [ + 496, + 512, + 893, + 608 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {j} = \\frac {\\mathcal {A} _ {1 , j} \\times \\left\\| \\mathcal {V} _ {j} \\right\\|}{\\sum_ {i = 2} \\mathcal {A} _ {1 , i} \\times \\left\\| \\mathcal {V} _ {i} \\right\\|} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 614, + 893, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The importance scores are converted to probabilities and are used to sample tokens, where tokens with a lower score have more of a chance of being dropped.", + "bbox": [ + 496, + 662, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attack: Since ATS uses inverse transform sampling, it results in fewer samples if the importance distribution is sharp. To maximize the computation in ATS, we aim to obtain a distribution of scores with high entropy to maximize the number of retained tokens. Therefore, we optimize the patch so that the attention of the classification token over other tokens is a uniform distribution using the following MSE loss:", + "bbox": [ + 496, + 708, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 2} ^ {N} \\left| \\left| \\mathcal {A} _ {1, i} - \\frac {1}{N} \\right| \\right| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 825, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that one can optimize $S$ to be uniform, but we found the above loss to be easier to optimize. For a multi-head", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24789", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "attention layer, we calculate the loss for each head and then sum the loss over all heads. Moreover, ATS can be applied to any layer of a vision transformer. For a given model, we apply our loss at all ATS layers and use a weighted summation for optimization.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Defense", + "text_level": 1, + "bbox": [ + 76, + 184, + 166, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An obvious defense, although weak, will be to use nondynamic efficient methods only, e.g., weight pruning, where the reduction in compute is deterministic and does not depend on the input. However, most such methods do not achieve high levels of computation efficiency since they do not take advantage of the simplicity of images.", + "bbox": [ + 75, + 210, + 470, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt standard adversarial training as a better defense method for our attack. In the standard way, at each iteration of training the model, one would load an image, attack it, and then use it with correct labels in training the model. We cannot adopt this out-of-the-box since our attack generalizes across images and is not dependent on a single image only. To do this, we maintain a set of adversarial patches, and at each iteration sample one of them randomly (uniformly), and use it at the input while optimizing the original loss of the efficient model to train a robust model. To adapt the set of adversarial patches to the model being trained, we interrupt the training at every $20\\%$ mark of each epoch and optimize for a new patch to be added to the set of patches. To limit the computational cost of training, we use only 500 iterations to optimize for a new patch, which results in an attack with reasonable accuracy compared to our main results.", + "bbox": [ + 75, + 301, + 473, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 560, + 209, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Attack on Efficient Vision Transformers", + "text_level": 1, + "bbox": [ + 76, + 585, + 419, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset: We evaluate the effectiveness of our attack on two datasets: ImageNet-1K [19] and CIFAR-10 [49]. ImageNet-1K contains 1.3M images in the train set and 50K images in the validation set with 1000 total categories. CIFAR-10 has 50K images for training and 10K images for validation with 10 total categories.", + "bbox": [ + 75, + 609, + 470, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: We report Top-1 accuracy and average computation in terms of GFLOPs for both attacked and unattacked models. Similar to Attack Success Rate in a standard adversarial attack, we introduce a metric: Attack Success to quantify the efficacy of the attack. We define Attack Success as the number of FLOPs increased by the attack divided by the number of FLOPs decreased by the efficient method. Attack Success = $\\frac{(\\mathrm{FLOPs}_{\\mathrm{attack}} - \\mathrm{FLOPs}_{\\mathrm{min}})}{(\\mathrm{FLOPs}_{\\mathrm{max}} - \\mathrm{FLOPs}_{\\mathrm{min}})}$ where $\\mathrm{FLOPs}_{\\mathrm{min}}$ is the compute of the efficient model and $\\mathrm{FLOPs}_{\\mathrm{max}}$ is that of the original inefficient model. Attack Success is thus capped at $100\\%$ while a negative value denotes a reduction in FLOPs. Note that our", + "bbox": [ + 75, + 717, + 472, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attack Success metric illustrates the effectiveness of an attack in reversing the FLOPs reduction of a particular method.", + "bbox": [ + 498, + 90, + 893, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines: We propose three alternative approaches to Slow-Former (ours) to generate the patch.", + "bbox": [ + 498, + 137, + 893, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Random Patch: A simple baseline is to generate a randomly initialized patch. We sample IID pixel values from a uniform distribution between 0 and 255 to create the patch.", + "bbox": [ + 498, + 169, + 890, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NTAP: We consider a standard adversarial patch that is trained to attack the model task performance instead of compute. We use a non-targeted universal adversarial patch (NTAP) to attack the model. We train the patch to fool the model by misclassifying the image it is pasted on. We use the negative of the cross-entropy loss with the predicted and ground-truth labels as the loss to optimize the patch.", + "bbox": [ + 498, + 215, + 893, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TAP: We train a universal targeted adversarial patch (TAP). The patch is optimized to classify all images in the train set to a single fixed category. Similar to NTAP, the adversarial attack here is on task performance and not computation. We experiment with ten randomly generated target category labels and report the averaged metrics.", + "bbox": [ + 498, + 321, + 893, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: We use PyTorch [66] for all experiments. Unless specified, we use a patch of size $64 \\times 64$ , train and test on $224 \\times 224$ images, and we paste the patch on the top-left corner. Note that our patch occupies just $8\\%$ of the total area of an input image. We use AdamW [57] optimizer to optimize the patches and use 4 NVIDIA RTX 3090 GPUs for each experiment. We use varying batch sizes and learning rates for each of the computation-efficient methods.", + "bbox": [ + 498, + 429, + 893, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ATS Details: As in ATS [25], we replace layers 3 through 9 of ViT networks with the ATS block and set the maximum limit for the number of tokens sampled to 197 for each layer. We train the patch for 2 epochs with a learning rate of 0.4 for ViT-Tiny and $lr = 0.2$ for ViT-Base and ViT-Small. We use a batch size of 1024 and different loss coefficients for each layer of ATS. For DeiT-Tiny we use [1.0, 0.2, 0.2, 0.2, 0.01, 0.01, 0.01], for DeiT-Small we use [1.0, 0.2, 0.05, 0.01, 0.005, 0.005, 0.005], and for DeiT-Base we use [2.0, 0.1, 0.02, 0.01, 0.005, 0.005, 0.005]. The weights are vastly different at initial and final layers to account for the difference in loss magnitudes across layers.", + "bbox": [ + 498, + 551, + 893, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A-ViT Details: When attacking A-VIT[88], the patches are optimized for one epoch with a learning rate of 0.2 and a batch size of 512 (128 × 4 GPUs) using AdamW [57] optimizer. We optimize the patches for 4 epochs for patch length 32 and below. For CIFAR-10 experiments, the images are resized from $32 \\times 32$ to $256 \\times 256$ and a $224 \\times 224$ crop is used as the input. For the training of adversarial defense, we generate 5 patches per epoch of adversarial training and limit the number of iterations for patch generation to 500. The learning rate for patch optimization is increased to 0.8 for faster convergence.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24790", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5dc90aac70e089a48fda267e378abc93bc5f77268d1bdd84b541dbd072177e79.jpg", + "table_caption": [ + "Table 1. Computation and Energy Attack on Efficient Vision Transformers: Comparison of the effect of our attack with baselines: No Attack, Random Patch, targeted (TAP), and non-targeted (NTAP) adversarial patches applied to three input-dynamic computation efficient pre-trained models of varying architectures. The maximum possible compute for a given architecture is provided in bold. On A-ViT, we completely undo the efficiency gains obtained by the efficient method through our attack, achieving Attack Success of $100\\%$ . We achieve high Attack Success on all approaches while the baselines expectedly do not contribute to increase in compute." + ], + "table_footnote": [], + "table_body": "
MethodAttackModel GFLOPsTop-1 AccAttack Success
A-ViTViT-Tiny1.3--
No attack0.8771.4%-
Random Patch0.8770.8%-1%
TAP0.850.1%-5%
NTAP0.830.1%-10%
SlowFormer (ours)1.34.7%100%
A-ViTViT-Small4.6--
No attack3.778.8%-
Random Patch3.778.4%-2%
TAP3.60.1%-12%
NTAP3.60.1%-7%
SlowFormer (ours)4.62.3%99%
ATSViT-Tiny1.3--
No attack0.8470.3%-
Random Patch0.8369.8%-2%
TAP0.760.1%-17%
NTAP0.610.1%-50%
SlowFormer (ours)1.01.2%35%
ATSViT-Small4.6--
No attack3.179.2%-
Random Patch3.178.6%-1%
TAP3.00.1%-7%
NTAP2.40.1%-47%
SlowFormer (ours)4.01.0%60%
ATSViT-Base17.6--
No attack12.681.3%-
Random Patch12.581.2%-2%
TAP12.00.1%-12%
NTAP11.00.1%-32%
SlowFormer (ours)15.40.2%52%
AdaViTViT-Small4.6--
No attack2.2577.3%-
Random Patch2.2076.9%-2%
TAP2.280.1%1%
NTAP2.150.1%-4%
SlowFormer (ours)3.20.4%40%
", + "bbox": [ + 99, + 224, + 506, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AdaViT Details: For AdaViT[63], we first freeze the weights and use a learning rate of 0.2 and a batch size of 128 with 4 GPUs for patch optimization. We use AdamW [57]", + "bbox": [ + 75, + 839, + 470, + 887 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0424ccce6be847edeaa91ff449a53701345e2f373aa6249232a05fc5e04a35c0.jpg", + "table_caption": [ + "Figure 2. Visualization of our Energy Attack on Vision Transformers: We visualize the A-ViT-Small with and without our attack. We use patch size of 32 for the attack (on the top-left corner). We show pruned tokens at layer 8 of A-ViT-Small. Our attack can recover most of the pruned tokens, resulting in increased computation and power consumption. Note that although the patch is reasonably small and is in the corner of the view, it can affect the whole computational flow of the network. This is probably due to the global attention mechanism in transformers." + ], + "table_footnote": [], + "table_body": "
AttackedEfficient ModelAttackedEfficient Model
", + "bbox": [ + 545, + 224, + 913, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimizer with no decay and train for 2 epochs with a patch size of $64 \\times 64$ . We train on the ImageNet-1k train dataset and evaluate it on the test set.", + "bbox": [ + 498, + 840, + 890, + 885 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6de61f8eae0584938f9a64c90199c3ffbd14dfc0f00b6d600c5bcee0e0cf0b98.jpg", + "table_caption": [ + "Table 2. Results on CIFAR10 dataset. We report results on CIFAR10 dataset to show that our attack is not specific to ImageNet alone. CIFAR-10 is a small dataset compared to ImageNet and thus results in an extremely efficient A-ViT model. Our attack increases the FLOPs from 0.11 to 0.58 which restores nearly $41\\%$ of the original reduction in the FLOPs." + ], + "table_footnote": [], + "table_body": "
MethodModel FLOPsTop-1 AccAttack Success
ViT-Tiny1.2695.9%-
A-ViT-Tiny0.1195.8%-
SlowFormer (ours)0.5860.2%41%
ATS-Tiny0.8594.7%-
SlowFormer (ours)0.9924.7%34.1%
", + "bbox": [ + 109, + 183, + 455, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84741892e84b8e36c7ca7cd0580b06f8d4308cb1212a4b01896dc10fe3f7e223.jpg", + "image_caption": [ + "Figure 3. Visualization of optimized patch: We show the learned universal patches for each of the three efficient methods." + ], + "image_footnote": [], + "bbox": [ + 80, + 325, + 171, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cff71724dc6dfc20de777051a6fffde590fd94fd51cf3f04e6e3dd981910b7f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 325, + 320, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e43cce4c5073a7ea8d1a9a2865440e96238cfed9feee7e47454ac716ec047a96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 325, + 467, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Effect of patch size: Analysis of the effect of adversarial patch size on the attack success rate on A-ViT. Our attach is reasonably successful even using a small patch size $(32 \\times 32)$ , which is only $2\\%$ of the image area. Interestingly, a small patch on the corner of the view affects the computational flow of the entire transformer model. This might be due to the global attention mechanism in transformers.", + "bbox": [ + 75, + 467, + 470, + 565 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a02167d8537d1dcadd52e190d6c719b3c4e6cbdd1323f64237e3536a305da35.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Patch Size (Area)Model GFLOPsTop-1 AccuracyAttack Success
ViT-Tiny1.26--
A-ViT-Tiny0.8771.4%-
64 (8%)1.264.7%100%
48 (5%)1.261.8%99%
32 (2%)1.2217.4%90%
16 (0.5%)0.9863.3%27%
ViT-Small4.6--
A-ViT-Small3.778.8%-
64 (8%)4.62.3%99%
48 (5%)4.65.1%98%
32 (2%)4.439.5%78%
16 (0.5%)3.878.2%16%
", + "bbox": [ + 78, + 575, + 416, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. The results of our attack, SlowFormer , on various methods on ImageNet dataset are shown in table 1. In A-ViT, we successfully recover $100\\%$ of the computation reduced", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/baedee61a8c1b91141487e2f457a4c368d873f1cbe5b46a6b64c01afda3f9bba.jpg", + "table_caption": [ + "Table 3. Accuracy controlled compute adversarial attack: We attack the efficiency of A-ViT while either maintaining or destroying its classification performance. We observe that our attack can achieve a huge variation in task performance without affecting the Attack Success. The ability to attack the computation without affecting the task performance might be crucial in some applications." + ], + "table_footnote": [], + "table_body": "
AttackModel GFLOPsAttack SuccessTop-1 Acc
ViT-Tiny1.26--
No attack0.87-71.4%
Acc agnostic1.26100%4.7%
Preserve acc1.2392%68.5%
Destroy acc1.26100%0.1%
", + "bbox": [ + 547, + 183, + 867, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The efficient methods are also susceptible universal perturbation based attacks. We use an $\\ell_{\\infty}$ bound on the perturbation.", + "bbox": [ + 498, + 340, + 890, + 369 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1290446887e70b76f6066558ada7a194eef6eebfaea57e87541a3f011a8e6156.jpg", + "table_caption": [ + "Table 5. Attack with adversarial perturbation on ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodEpsilon (/255.)AttackModel GFLOPsTop-1 AccuracyAttack Success
A-ViTViT-Tiny1.3--
-No attack0.8771.4%-
16SlowFormer1.156.1%73%
32SlowFormer1.250.5%98.4%
ATS-No attack0.8470.3%-
16SlowFormer0.9815.6%30.4%
32SlowFormer1.040.8%43.5%
A-ViTViT-Small4.6--
-No attack3.778.8%-
16SlowFormer4.4820%86.4%
32SlowFormer4.591%98.1%
ATS-No attack3.179.2%-
16SlowFormer3.631.0%33.3%
32SlowFormer3.83.6%46.7%
AdaVit-No attack2.2577.3%-
16SlowFormer3.026.1%31.9%
32SlowFormer3.22.8%40.4%
", + "bbox": [ + 503, + 378, + 908, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by A-ViT. Our attack has an Attack Success of $60\\%$ on ATS and $40\\%$ on AdaViT with ViT-Small. A random patch attack has little effect on both the accuracy and computation of the method. Both standard adversarial attack baselines, TAP and NTAP, reduce the accuracy to nearly $0\\%$ . Interestingly, these patches further decrease the computation of the efficient model being attacked. This might be because of the increased importance of adversarial patch tokens to the task and thus reduced importance of other tokens. Targeted patch (TAP) has a significant reduction in FLOPs on the ATS method. Since the token dropping in ATS relies on the distribution of attention values of classification tokens, a sharper distribution due to the increased importance of a token can result in a reduction in computation. The", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24792", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "computation increase with SlowFormer for AdaViT is comparatively low. To investigate, we ran a further experiment using a patch size of $224 \\times 224$ (entire image size) to find the maximum possible computation for an image. This resulted in 4.18 GFLOPs on the ImageNet-1K validation set, which is markedly lower than the limit of 4.6. Using this as an upper-bound of GFLOPs increase, SlowFormer achieves a $49\\%$ Attack Success.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report the results on CIFAR-10 dataset in Table 2. The efficient model (A-ViT) drastically reduces the computation from 1.26 GFLOPs to 0.11 GFLOPs. Most of the tokens are dropped as early as layer two in the efficient model. SlowFormer is able to effectively attack even in such extreme scenarios, achieving an Attack Success of $40\\%$ and increasing the mean depth of tokens from nearly one to five. SlowFormer is similarly effective on ATS with an Attack Success of $34\\%$ .", + "bbox": [ + 75, + 212, + 470, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We additionally visualize the effectiveness of our attack in Figure 2. The un-attacked efficient method retains only highly relevant tokens at the latter layers of the network. However, our attack results in nearly the entire image being passed through all layers of the model for all inputs. In Fig. 3, we visualize the optimized patches for each of the three efficient methods.", + "bbox": [ + 75, + 348, + 468, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablations:", + "text_level": 1, + "bbox": [ + 76, + 465, + 192, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform all ablations on the A-ViT approach using their pretrained ViT-Tiny architecture model.", + "bbox": [ + 75, + 489, + 468, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Accuracy controlled compute adversarial attack: As seen in Table 1, our attack can not only increase the computation, but also reduce the model accuracy. This can be desirable or hurtful based on the attacker's goals. A low-accuracy model might be an added benefit, similar to regular adversaries, but might also lead to the victim detecting the attack. We show that it is possible to attack the computation of the model while either preserving or destroying the task performance by additionally employing a task loss in the patch optimization. Table 3 indicates that the accuracy can be significantly modified while maintaining a high Attack Success.", + "bbox": [ + 75, + 536, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of patch size: We vary the patch size from $64 \\times 64$ to $16 \\times 16$ (just a single token) and report the results in Table 4. Interestingly, our attack with ViT-Small has a $73\\%$ Attack Success with a $32 \\times 32$ patch size, which occupies only $2\\%$ of the input image area.", + "bbox": [ + 75, + 702, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of patch location: We vary the location of the patch to study its effect on the Attack Success of the model. We randomly sample a location in the image for where we paste the patch on. We perform five such experiments and observe an Attack Success of $100\\%$ for all patch locations.", + "bbox": [ + 75, + 777, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perturbation attack: While we focus on patch based attacks in this paper, efficient transformers are also susceptible to perturbation based attacks (table 5). In perturbation attacks,", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a61897e858efd9e27d789e81775c6b4f4ed3c6b95222cd0ca748451e5000e5c9.jpg", + "table_caption": [ + "Table 6. Defense using adversarial training: We propose and show the impact of our defense for our adversarial attack on A-ViT. Our defense is simply maintaining a set of universal patches and training the model to be robust to a random sample of those at each iteration. The defense reduces the computation to some extent (1.26 to 1.01), but is still far from the unattacked model (0.87)." + ], + "table_footnote": [], + "table_body": "
MethodGFLOPsTop-1 Acc.Attack Success
No attack0.8771.4-
SlowFormer1.264.7%100%
Adv Defense + SlowFormer1.0165.8%34%
", + "bbox": [ + 532, + 183, + 859, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all pixels in the image can be modified, but with an upper bound on the $\\ell_{\\infty}$ norm of the perturbation.", + "bbox": [ + 496, + 324, + 890, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Adversarial training based defense", + "text_level": 1, + "bbox": [ + 498, + 375, + 803, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our simple defense that is adopted from standard adversarial training is explained in Section 4. The results for defending against attacking A-ViT are shown in Table 6. The original A-ViT reduces the GFLOPs from 1.26 to 0.87, our attack increases it back to 1.26 with $100\\%$ attack success. The proposed defense reduces the GFLOPs to 1.01 which is still higher than the original 0.87. We hope our paper encourages the community to develop better defense methods to reduce the vulnerability of efficient vision transformers.", + "bbox": [ + 496, + 402, + 890, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 565, + 619, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recently, we have seen efficient vision transformer models in which the computation is adaptively modified based on the input. We argue that this is an important research direction and that there will be more progress in this direction in the future. However, we show that the current methods are vulnerable to a universal adversarial patch that increases the computation and thus power consumption at inference time. Our experiments show promising results for three SOTA efficient transformer models, where a small patch that is optimized on the training data can increase the computation to the maximum possible level in the testing data in some settings. We also propose a defense that reduces the effectiveness of our attack. We hope that our paper will encourage the community to study such attacks and develop better defense methods on various machine learning methods, including generative models, that reduce the computation adaptively with the input.", + "bbox": [ + 496, + 594, + 893, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This work was partially supported by DARPA under Contract No. HR00112190135 and HR00112290115 and NSF grant 1845216.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24793", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] iphone battery capacity. https://bigthink.com/the-future/battery-technology-lags/.1", + "[2] Starship robot. https://www.wevolver.com/specs/starship-technologies-starship-robot.1", + "[3] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34, 2021. 2", + "[4] Babak Ehteshami Bejnordi, Tijmen Blankevoort, and Max Welling. Batch-shaping for learning conditional channel gated networks. arXiv preprint arXiv:1907.06627, 2019. 2", + "[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2", + "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2", + "[7] Jason Ross Brown, Yiren Zhao, Ilia Shumailov, and Robert D Mullins. Dartformer: Finding the best type of attention. arXiv preprint arXiv:2210.00641, 2022. 2", + "[8] Tom B Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017. 3", + "[9] Shaofeng Cai, Yao Shu, and Wei Wang. Dynamic routing networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3597, 2021. 2", + "[10] Shijie Cao, Lingxiao Ma, Wencong Xiao, Chen Zhang, Yunxin Liu, Lintao Zhang, Lanshun Nie, and Zhi Yang. Seernet: Predicting convolutional neural network feature-map sparsity through low-bit quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11216-11225, 2019. 3", + "[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2", + "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021. 2", + "[13] Guangyi Chen, Chunze Lin, Liangliang Ren, Jiwen Lu, and Jie Zhou. Self-critical attention learning for person re-identification. In ICCV, pages 9637–9646, 2019. 2", + "[14] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8064-8073, 2021. 3", + "[15] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. Slothbomb: Efficiency poisoning attack against dynamic neural networks. 3" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Simin Chen, Mirazul Haque, Zihe Song, Cong Liu, and Wei Yang. Transslowdown: Efficiency attacks on neural machine translation systems. 2021. 3", + "[17] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsglowsdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 3", + "[18] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2", + "[22] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv preprint arXiv:1910.10073, 2019. 2", + "[23] Quanfu Fan, Chun-Fu (Ricarhd) Chen, Hilde Kuehne, Marco Pistoia, and David Cox. More Is Less: Learning Efficient Video Representations by Temporal Aggregation Modules. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 3", + "[24] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers, 2021. 3", + "[25] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 396-414. Springer, 2022. 2, 4, 5", + "[26] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res, 23:1-40, 2021. 2", + "[27] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3", + "[28] Michael Figurnov, Maxwell D Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24794", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Spatially adaptive computation time for residual networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1039-1048, 2017. 2, 3", + "[29] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[30] Yonggan Fu, Shunyao Zhang, Shang Wu, Cheng Wan, and Yingyan Lin. Patch-fool: Are vision transformers always robust against adversarial perturbations? arXiv preprint arXiv:2203.08392, 2022. 3", + "[31] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. arXiv preprint arXiv:1810.05331, 2018. 2", + "[32] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In arXiv preprint arXiv:1412.6572, 2014. 3", + "[33] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2", + "[34] Jiaqi Guan, Yang Liu, Qiang Liu, and Jian Peng. Energy-efficient amortized inference with cascaded deep classifiers. arXiv preprint arXiv:1710.03368, 2017. 2", + "[35] Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14264-14273, 2020. 3", + "[36] Mirazul Haque, Simin Chen, Wasif Arman Haque, Cong Liu, and Wei Yang. Nodeattack: Adversarial attack on the energy consumption of neural odes. 2021. 3", + "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2", + "[38] Charles Herrmann, Richard Strong Bowen, and Ramin Zabih. Channel selection using gumbel softmax. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII, pages 241-257. Springer, 2020. 2", + "[39] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[40] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv preprint arXiv:1912.12180, 2019. 2", + "[41] Sanghyun Hong, Yigitcan Kaya, Ionut-Vlad Modoranu, and Tudor Dumitraş. A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432, 2020. 3", + "[42] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3", + "[43] Weizhe Hua, Yuan Zhou, Christopher M De Sa, Zhiru Zhang, and G Edward Suh. Channel gating neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2", + "[45] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 2", + "[46] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of self-attention. arXiv preprint arXiv:2209.04881, 2022. 2", + "[47] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 3", + "[48] Soroush Abbasi Koohpayegani and Hamed Piriavash. Sima: Simple softmax-free attention for vision transformers. arXiv preprint arXiv:2206.08898, 2022. 2", + "[49] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5", + "[50] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3", + "[51] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 2", + "[52] Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE international conference on computer vision, pages 5764-5772, 2017. 3", + "[53] Zhichao Li, Yi Yang, Xiao Liu, Feng Zhou, Shilei Wen, and Wei Xu. Dynamic computational time for visual attention. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1199-1209, 2017. 3", + "[54] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. arXiv preprint arXiv:2209.09004, 2022. 2", + "[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[56] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. 2019. 5, 6", + "[58] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[59] Wenhao Lu, Jian Jiao, and Ruofei Zhang. Twinbert: Distilling knowledge to twin-structured bert models for efficient retrieval. arXiv preprint arXiv:2002.06275, 2020. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24795", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 4", + "[61] Kaleel Mahmood, Rigel Mahmood, and Marten Van Dijk. On the robustness of vision transformers to adversarial examples. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7838-7847, 2021. 3", + "[62] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers. arXiv preprint arXiv:2110.03860, 2021. 2", + "[63] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022. 2, 3, 4, 6", + "[64] Jianhong Pan, Qichen Zheng, Zhipeng Fan, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Gradauto: Energy-oriented attack on dynamic neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 637-653. Springer, 2022. 3", + "[65] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 3", + "[66] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2", + "[68] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 2", + "[69] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[70] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks, 2016. 2", + "[71] Mengye Ren, Andrei Pokrovsky, Bin Yang, and Raquel Urtasun. Sbnet: Sparse blocks network for fast inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8711-8720, 2018. 3", + "[72] Aniruddha Saha, Akshayvarun Subramanya, Koninika B. Patil, and Hamed Pirsavash. Adversarial patches exploiting contextual reasoning in object detection. In ArXiv, 2019. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference, 2023. 1", + "[74] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3531-3539, 2021. 2", + "[75] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3", + "[76] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2", + "[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2", + "[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2", + "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[80] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2", + "[81] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the iee/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 3", + "[82] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2, 3", + "[83] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 3", + "[84] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3", + "[85] Zhenda Xie, Zheng Zhang, Xizhou Zhu, Gao Huang, and Stephen Lin. Spatially adaptive inference with stochastic feature sampling and interpolation. In Computer Vision-ECCV" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24796", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 531-548. Springer, 2020. 3", + "[86] Le Yang, Yizeng Han, Xi Chen, Shiji Song, Jifeng Dai, and Gao Huang. Resolution adaptive networks for efficient inference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2369-2378, 2020. 3", + "[87] Zerui Yang, Yuhui Xu, Wenrui Dai, and Hongkai Xiong. Dynamic-stride-net: Deep convolutional neural network with dynamic stride. In Optoelectronic Imaging and Multimedia Technology VI, pages 42-53. SPIE, 2019. 3", + "[88] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2, 3, 4, 5", + "[89] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819-10829, 2022. 2", + "[90] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[91] Zhihang Yuan, Bingzhe Wu, Guangyu Sun, Zheng Liang, Shiwan Zhao, and Weichen Bi. S2dnas: Transforming static cnn model for dynamic inference via neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 175-192. Springer, 2020. 2", + "[92] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 3", + "[93] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[94] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[95] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2", + "[96] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 866 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24797", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_model.json b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..73cb66ba30455d80cc81d7aafbd80e98274f9794 --- /dev/null +++ b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_model.json @@ -0,0 +1,2633 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.868, + 0.175 + ], + "angle": 0, + "content": "SlowFormer: Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.203, + 0.233, + 0.221 + ], + "angle": 0, + "content": "K L Navaneet *1" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.204, + 0.52, + 0.222 + ], + "angle": 0, + "content": "Soroush Abbasi Koohpayegani *1" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.223, + 0.521, + 0.24 + ], + "angle": 0, + "content": "1 University of California, Davis" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.204, + 0.696, + 0.221 + ], + "angle": 0, + "content": "Essam Sleiman *12†" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.222, + 0.708, + 0.24 + ], + "angle": 0, + "content": "2 Harvard University" + }, + { + "type": "text", + "bbox": [ + 0.718, + 0.204, + 0.868, + 0.221 + ], + "angle": 0, + "content": "Hamed Piriaviash" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.291 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.307, + 0.474, + 0.607 + ], + "angle": 0, + "content": "Recently, there has been a lot of progress in reducing the computation of deep models at inference time. These methods can reduce both the computational needs and power usage of deep models. Some of these approaches adaptively scale the compute based on the input instance. We show that such models can be vulnerable to a universal adversarial patch attack, where the attacker optimizes for a patch that when pasted on any image, can increase the compute and power consumption of the model. We run experiments with three different efficient vision transformer methods showing that in some cases, the attacker can increase the computation to the maximum possible level by simply pasting a patch that occupies only \\(8\\%\\) of the image area. We also show that a standard adversarial training defense method can reduce some of the attack's success. We believe adaptive efficient methods will be necessary in the future to lower the power usage of expensive deep models, so we hope our paper encourages the community to study the robustness of these methods and develop better defense methods for the proposed attack. Code is available at:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.609, + 0.466, + 0.624 + ], + "angle": 0, + "content": "https://github.com/UCDvision/SlowFormer" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.671 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.697, + 0.472, + 0.864 + ], + "angle": 0, + "content": "The field of deep learning has recently made significant progress in improving model efficiency for inference. Two broad categories of methods can be distinguished: 1) those that reduce computation regardless of input, and 2) those that reduce the computation depending on the input (adaptively). Most methods, such as weight pruning or model quantization, belong to the first category which reduces computation by a constant factor regardless of the input. However, in many applications, the complexity of the perception task may differ depending on the input. For example, when a self-driving car is driving between lanes in an empty street," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.276, + 0.896, + 0.397 + ], + "angle": 0, + "content": "the perception may be simpler and require less computation when compared to driving in a busy city street scene. Interestingly, in some applications, simple scenes such as highway driving may account for the majority of the time. Therefore, we believe that adaptive computation reduction will become an increasingly important research area in the future, especially when non-adaptive methods reach the lower bound of computation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.896, + 0.685 + ], + "angle": 0, + "content": "We argue that reduction of compute usually reduces power usage, which is crucial, particularly in mobile devices that run on battery, e.g., AR/VR headsets, humanoid robots, and drones. For instance, increasing the size of the battery for a drone may lead to a drastic reduction in its range due to the increased battery weight. This is important since the improvement in battery technology is much slower than compute technology. For instance, the battery capacity from iPhone [1] (1st generation) in 2007 to iPhone 15 Pro Max in 2023 improved from 5.18 watt-hour to 17.32 watt-hour (less than 4 times) while the compute has increased by a much larger factor. As an example, a delivery robot like Starship uses a 1,200Wh battery and can run for 12 hours [2], so it uses almost 100 watts for compute and mobility. Hence, an adversary increasing the power consumption of the perception unit by 20 watts, will reduce the battery life by almost \\(20\\%\\), which can be significant. Note that 20 watts increase in power is realistic assuming that it uses two NVIDIA Jetson Xavier NX cards (almost 20 watts each)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.897, + 0.824 + ], + "angle": 0, + "content": "Key idea: Assuming that a perception method is reducing the computation adaptively with the input, an adversary can attack the model by modifying the input to increase the computation and power consumption. Our goal is to design a universal adversarial patch that when pasted on any input image, it will increase the computation of the model leading to increased power consumption. We believe this is an important vulnerability, particularly for safety-critical mobile systems that run on battery." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Please note that in this paper, we do not experiment with real hardware to measure the power consumption. Instead, we report the change in FLOPs of the inference time assuming that the power consumption is positively correlated with the number of FLOPs, as studied in [73]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.876, + 0.213, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.448, + 0.9 + ], + "angle": 0, + "content": "† Work done while he was an undergraduate student at UC Davis." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.876, + 0.448, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24786" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.093, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.897, + 0.368 + ], + "angle": 0, + "content": "Figure 1. Computation and Energy Attack on Vision Transformers: Given a pre-trained input-dependent computation efficient model, the adversary first attaches an adversarial patch to all images in a dataset and optimizes this patch with our method such that it maximizes the model's computation for each sample. During inference, the adversary modifies the input of the victim's model by applying the learnt patch to it. This results in an increase in compute in the victim's model. The attack will thus potentially slowdown and also lead to increased energy consumption and CPU/GPU usage on the victim's device." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.56 + ], + "angle": 0, + "content": "We design our attack, SlowFormer, for three different methods (A-ViT [88], ATS [25], and Ada-VIT [63]) that reduce the computation of vision transformers. These methods generally identify the importance of each token for the final task and drop the insignificant ones to reduce the computation. We show that in all three cases, our attack can increase the computation by a large margin, returning it to the full-compute level (non-efficient baseline) for all images in some settings. Our threat model is agnostic to the accuracy of the model and attacks the computation and power consumption only. Figure 1 shows our attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.473, + 0.681 + ], + "angle": 0, + "content": "There are some prior works that design a pixel-level perturbation attack to increase the compute of the model. We believe universal patch-based attacks that do not change with the input image (generalize from training data to test data) are much more practical in real applications. Note that to modify the pixel values on a real robot, the attacker needs to access and manipulate the image between the camera and compute modules, which is impossible in many applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.472, + 0.772 + ], + "angle": 0, + "content": "Contributions: We show that efficient vision transformer methods are vulnerable to a universal patch attack that can increase their compute and power usage. We demonstrate this through experiments on three different efficient transformer methods. We show that an adversarial training defense can reduce attack success to some extent." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.22, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Vision Transformers: The popularity of transformers [79] in vision has grown rapidly since the introduction of the first vision transformer [20, 78]. Recent works demonstrate the strength of vision transformers on a variety of computer vision tasks [11, 18, 21, 56, 69, 77, 90, 93, 94, 96]. Moreover, transformers are the backbone of recent Self-Supervised" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Learning (SSL) models [12, 37], and vision-language models [67]. In our work, we design an attack to target the computation and energy efficiency of vision transformers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.897, + 0.655 + ], + "angle": 0, + "content": "Efficient Vision Transformers: Due to the recent importance and popularity of vision transformers, many works have started to study the efficiency of vision transformers [7, 46, 89]. To accomplish this, some lines of work study token pruning with the goal of removing uninformative tokens in each layer [25, 62, 63, 68, 88]. ToMe [6] merges similar tokens in each layer to decrease the computation. Some works address quadratic computation of self-attention module by introducing linear attention [3, 45, 48, 58, 74]. Efficient architectures [40, 55] that limit the attention span of each token have been proposed to improve efficiency. In our paper, we attack token pruning based efficient transformers where the computation varies based on the input samples [25, 63, 88]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Dynamic Computation: There are different approaches to reducing the computation of vision models, including knowledge distillation to lighter network [39, 59], model quantization [54, 70] and model pruning [51]. In these methods, the computation is fixed during inference. In contrast to the above models, some works address efficiency by having variable computation based on the input. The intuition behind this direction is that not all samples require the same amount of computation. Several recent works have developed models that dynamically exit early or skip layers [5, 22, 28, 33, 34, 44, 76, 80, 82] and selectively activate neurons, channels or branches for dynamic width [4, 9, 13, 26, 31, 38, 43, 91] depending on the complexity of the input sample. Zhou et al. show that not all locations in an image contribute equally to the predictions of a CNN model [95], encouraging a new line of work to make CNNs more" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24787" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "efficient through spatially dynamic computation. Pixel-Wise dynamic architectures [10, 14, 23, 47, 71, 81, 85] learn to focus on the significant pixels for the required task while Region-Level dynamic architectures perform adaptive inference on the regions or patches of the input [29, 53]. Finally, lowering the resolution of inputs decreases computation, but at the cost of performance. Conventional CNNs process all regions of an image equally, however, this can be inefficient if some regions are \"easier\" to process than others [42]. Correspondingly, [86, 87] develop methods to adaptively scale the resolution of images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.428 + ], + "angle": 0, + "content": "Transformers have recently become extremely popular for vision tasks, resulting in the release of a few input-dynamic transformer architectures [24, 63, 88]. Fayyaz et al. [24] introduce a differentiable parameter-free Adaptive Token Sampler (ATS) module which scores and adaptively samples significant tokens. ATS can be plugged into any existing vision transformer architecture. A-ViT [88] reduces the number of tokens in vision transformers by discarding redundant spatial tokens. Meng et al. [63] propose AdaViT, which trains a decision network to dynamically choose which patch, head, and block to keep/activate throughout the backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.597 + ], + "angle": 0, + "content": "Adversarial Attack: Adversarial attacks are designed to fool models by applying a targeted perturbation or patch on an image sample during inference [32, 50, 75]. These methods can be incorporated into the training set and optimized to fool the model. Correspondingly, defenses have been proposed to mitigate the effects of these attacks [27, 52, 65, 84]. Patch-Fool [30] considers adversarial patch-based attacks on transformers. Some recent works [61, 83, 92] also study and design methods for the transferability of adversarial attacks on vision transformers. However, most prior adversarial attacks target model accuracy, ignoring model efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Energy Attack: Very recently, there have been a few works on energy adversarial attacks on neural networks. In ILFO [35], Haque et al. attack two CNN-based input-dynamic methods: SkipNet [82] and SACT [28] using image specific perturbation. DeepSloth [41] attack focuses on slowing down early-exit methods, reducing their energy efficiency by \\(90 - 100\\%\\). GradAuto [64] successfully attacks methods that are both dynamic width and dynamic depth. NICGSlowDown and TransSlowDown [16, 17] attack neural image caption generation and neural machine translation methods, respectively. All these methods primarily employ image specific perturbation based adversarial attack. Sloth-Bomb injects efficiency backdoors to input-adaptive dynamic neural networks [15] and NodeAttack [36] attacks Neural Ordinary Differential Equation models, which use ordinary differential equation solving to dynamically predict the output of a neural network. Our work is closely related to ILFO [35], DeepSloth [41] and GradAuto [64] in that we attack the computational efficiency of networks. However, unlike these methods, we focus on designing an adversarial patch-based" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "attack that is universal and on vision transformers. We additionally provide a potential defense for our attack. We use a patch that generalizes from train to test set and thus we do not optimize per sample during inference. Our patch-based attack is especially suited for transformer architectures [30]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.182, + 0.8, + 0.2 + ], + "angle": 0, + "content": "3. Computation and Energy Attack" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.208, + 0.649, + 0.222 + ], + "angle": 0, + "content": "3.1. Threat Model:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.895, + 0.382 + ], + "angle": 0, + "content": "We consider a scenario where the adversary has access to the victim's trained deep model and modifies its input such that the energy consumption and computational demand of the model is increased. The attack is agnostic to model accuracy. To make the setting more practical, instead of perturbing the entire image, we assume that the adversary can modify the input image by only pasting a patch [8, 72] on it and that the patch is universal, that is, image independent. During inference, a pretrained patch is pasted on the test image before propagating it through the network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.895, + 0.504 + ], + "angle": 0, + "content": "In this paper, we attack three state-of-the-art efficient transformers. Since the attacker manipulates only the input image and not the network parameters, the attacked model must have dynamic computation that depends on the input image. As stated earlier, several recent works have developed such adaptive efficient models and we believe that they will be more popular in the future due to the limits of non-adaptive efficiency improvement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.849, + 0.529 + ], + "angle": 0, + "content": "3.2. Attack on Efficient Vision Transformers:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Universal Adversarial Patch: We use an adversarial patch to attack the computational efficiency of transforms. The learned patch is universal, that is, a single patch is trained and is used during inference on all test images. The patch generalizes across images but not across models. The patch optimization is performed only on the train set. The patch is pasted on an image by replacing the image pixels using the patch. We assume the patch location does not change from train to test. The patch pixels are initialized using i.i.d. samples from a uniform distribution over [0, 255]. During each training iteration, the patch is pasted on the mini-batch samples and is updated to increase the computation of the attacked network. The patch values are projected onto [0, 255] and quantized to 256 uniform levels after each iteration. Note that we use a pretrained network and do not update its parameters either in the training or in the evaluation of our attack. During inference, the trained patch is pasted on the test images and the computational efficiency of the network on the adversarial image is measured." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Here, we focus on three methods employing vision transformers for the task of image classification. All these methods modify the computational flow of the network based on the input image for faster inference. A pretrained model is used for the attack and is not modified during" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24788" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.123 + ], + "angle": 0, + "content": "our adversarial patch training. We first provide a brief background of each method before describing our attack." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.137, + 0.208, + 0.151 + ], + "angle": 0, + "content": "Attacking A-ViT :" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Background: A-ViT [88] adaptively prunes image tokens to achieve speed-up in inference with minimal loss in accuracy. For a given image, a dropped token will not be used again in the succeeding layers of the network. Let \\( x \\) be the input image and \\( \\{t^l\\}_{1:K} \\) be the corresponding \\( K \\) tokens at layer \\( l \\). An input-dependent halting score \\( h_k^l \\) for a token \\( k \\) at layer \\( l \\) is calculated and the token is dropped at layer \\( N_k \\) where its cumulative halting score exceeds a fixed threshold value \\( 1 - \\epsilon \\) for the first time. The token is propagated until the final layer if its score never exceeds the threshold. Instead of introducing a new parameter for \\( h_k^l \\), the first dimension of each token is used to predict the halting score for the corresponding token. The network is trained to maximize the cumulative halting score at each layer and thus drop the tokens earlier. The loss, termed ponder loss, is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.408, + 0.473, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p o n d e r}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(N _ {k} + r _ {k}\\right), \\quad r _ {k} = 1 - \\sum_ {l = 1} ^ {N _ {k - 1}} h _ {k} ^ {l} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.473, + 0.558 + ], + "angle": 0, + "content": "Additionally, A-ViT enforces a Gaussian prior on the expected halting scores of all tokens via \\(KL\\)-divergence based distribution loss, \\(\\mathcal{L}_{\\mathrm{distr}}\\). These loss terms are minimized along with the task-specific loss \\(\\mathcal{L}_{\\mathrm{task}}\\). Thus, the overall training objective is \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{task}} + \\alpha_d\\mathcal{L}_{\\mathrm{distr}} + \\alpha_p\\mathcal{L}_{\\mathrm{ponder}}\\) where \\(\\alpha_{d}\\) and \\(\\alpha_{p}\\) are hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.473, + 0.651 + ], + "angle": 0, + "content": "Attack: Here, we train the patch to increase the inference compute of a trained A-ViT model. Since we are interested in the compute and not task-specific performance, we simply use \\(-\\left(\\alpha_{d}\\mathcal{L}_{\\mathrm{distr.}} + \\alpha_{p}\\mathcal{L}_{\\mathrm{ponder}}\\right)\\) as our loss. It is possible to preserve (or hurt) the task performance by additionally using \\(+\\mathcal{L}_{\\mathrm{task}}\\) (or \\(-\\mathcal{L}_{\\mathrm{task}}\\)) in the loss formulation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.216, + 0.68 + ], + "angle": 0, + "content": "Attacking AdaViT:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.472, + 0.803 + ], + "angle": 0, + "content": "**Background:** To improve the inference efficiency of vision transformers, AdaViT [63] inserts and trains a decision network before each transformer block to dynamically decide which patches, self-attention heads, and transformer blocks to keep/activate throughout the backbone. The \\( l^{\\text{th}} \\) block's decision network consists of three linear layers with parameters \\( W_{l} = W_{l}^{p}, W_{l}^{h}, W_{l}^{b} \\) which are then multiplied by each block's input \\( Z_{l} \\) to get \\( m \\)." + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.812, + 0.472, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\left(m _ {l} ^ {p}, m _ {l} ^ {h}, m _ {l} ^ {b}\\right) = \\left(W _ {l} ^ {p}, W _ {l} ^ {h}, W _ {l} ^ {b}\\right) Z _ {l} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "The value \\( m \\) is then passed to sigmoid function to convert it to a probability value used to make the binary decision of keep/discard. Gumbel-Softmax trick [60] is used to make this decision differentiable during training. Let \\( M \\) be the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.123 + ], + "angle": 0, + "content": "keep/discard mask after applying Gumbel-Softmax on \\(m\\). The loss on computation is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.13, + 0.894, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {u s a g e} = (\\frac {1}{D _ {p}} \\sum_ {d = 1} ^ {D _ {p}} M _ {d} ^ {p} - \\gamma_ {p}) ^ {2} + (\\frac {1}{D _ {h}} \\sum_ {d = 1} ^ {D _ {h}} M _ {d} ^ {h} - \\gamma_ {h}) ^ {2} \\\\ + \\left(\\frac {1}{D _ {b}} \\sum_ {d = 1} ^ {D _ {b}} M _ {d} ^ {b} - \\gamma_ {b}\\right) ^ {2} \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.895, + 0.317 + ], + "angle": 0, + "content": "where \\(D_{p}, D_{h}, D_{b}\\) represent the number of total patches, heads, and blocks of the entire transformer, respectively. \\(\\gamma_{p}, \\gamma_{h}, \\gamma_{b}\\) denote the target computation budgets i.e. the percentage of patches/heads/blocks to keep. The total loss is a combination of task loss (cross-entropy) and computation loss: \\(\\mathcal{L} = \\mathcal{L}_{ce} + \\mathcal{L}_{usage}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.894, + 0.438 + ], + "angle": 0, + "content": "Attack: To attack this model, we train the patch to maximize the computation loss \\(\\mathcal{L}_{usage}\\). More specifically, we set the computation-target \\(\\gamma\\) values to 0 and negate the \\(\\mathcal{L}_{usage}\\) term in Eq. 3. As a result, the patch is optimized to maximize the probability of keeping the corresponding patch (p), attention head (h), and transformer block (b). We can also choose to attack the prediction performance by selectively including or excluding the \\(\\mathcal{L}_{ce}\\) term." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.452, + 0.613, + 0.467 + ], + "angle": 0, + "content": "Attacking ATS:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.512 + ], + "angle": 0, + "content": "Background: Given \\(N\\) tokens with the first one as the classification token, the transformer attention matrix \\(\\mathcal{A}\\) is calculated by the following dot product:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.894, + 0.609 + ], + "angle": 0, + "content": "\\(\\mathcal{A} = \\mathrm{Softmax}\\left(\\mathcal{Q}\\mathcal{K}^T /\\sqrt{d}\\right)\\) where \\(\\sqrt{d}\\) is a scaling coefficient, \\(d\\) is the dimension of tokens, \\(\\mathcal{Q},\\mathcal{K}\\) and \\(\\nu\\) are the query, key and value matrices, respectively. The value \\(\\mathcal{A}_{1,j}\\) denotes the attention of the classification token to token \\(j\\). ATS [25] assigns importance score \\(S_{j}\\) for each token \\(j\\) by measuring how much the classification token attends to it:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.615, + 0.894, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {j} = \\frac {\\mathcal {A} _ {1 , j} \\times \\left\\| \\mathcal {V} _ {j} \\right\\|}{\\sum_ {i = 2} \\mathcal {A} _ {1 , i} \\times \\left\\| \\mathcal {V} _ {i} \\right\\|} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.709 + ], + "angle": 0, + "content": "The importance scores are converted to probabilities and are used to sample tokens, where tokens with a lower score have more of a chance of being dropped." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Attack: Since ATS uses inverse transform sampling, it results in fewer samples if the importance distribution is sharp. To maximize the computation in ATS, we aim to obtain a distribution of scores with high entropy to maximize the number of retained tokens. Therefore, we optimize the patch so that the attention of the classification token over other tokens is a uniform distribution using the following MSE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.827, + 0.892, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 2} ^ {N} \\left| \\left| \\mathcal {A} _ {1, i} - \\frac {1}{N} \\right| \\right| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Note that one can optimize \\( S \\) to be uniform, but we found the above loss to be easier to optimize. For a multi-head" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24789" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "attention layer, we calculate the loss for each head and then sum the loss over all heads. Moreover, ATS can be applied to any layer of a vision transformer. For a given model, we apply our loss at all ATS layers and use a weighted summation for optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.185, + 0.168, + 0.2 + ], + "angle": 0, + "content": "4. Defense" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.302 + ], + "angle": 0, + "content": "An obvious defense, although weak, will be to use nondynamic efficient methods only, e.g., weight pruning, where the reduction in compute is deterministic and does not depend on the input. However, most such methods do not achieve high levels of computation efficiency since they do not take advantage of the simplicity of images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.474, + 0.545 + ], + "angle": 0, + "content": "We adopt standard adversarial training as a better defense method for our attack. In the standard way, at each iteration of training the model, one would load an image, attack it, and then use it with correct labels in training the model. We cannot adopt this out-of-the-box since our attack generalizes across images and is not dependent on a single image only. To do this, we maintain a set of adversarial patches, and at each iteration sample one of them randomly (uniformly), and use it at the input while optimizing the original loss of the efficient model to train a robust model. To adapt the set of adversarial patches to the model being trained, we interrupt the training at every \\(20\\%\\) mark of each epoch and optimize for a new patch to be added to the set of patches. To limit the computational cost of training, we use only 500 iterations to optimize for a new patch, which results in an attack with reasonable accuracy compared to our main results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.21, + 0.579 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.421, + 0.603 + ], + "angle": 0, + "content": "5.1. Attack on Efficient Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.703 + ], + "angle": 0, + "content": "Dataset: We evaluate the effectiveness of our attack on two datasets: ImageNet-1K [19] and CIFAR-10 [49]. ImageNet-1K contains 1.3M images in the train set and 50K images in the validation set with 1000 total categories. CIFAR-10 has 50K images for training and 10K images for validation with 10 total categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Metrics: We report Top-1 accuracy and average computation in terms of GFLOPs for both attacked and unattacked models. Similar to Attack Success Rate in a standard adversarial attack, we introduce a metric: Attack Success to quantify the efficacy of the attack. We define Attack Success as the number of FLOPs increased by the attack divided by the number of FLOPs decreased by the efficient method. Attack Success = \\(\\frac{(\\mathrm{FLOPs}_{\\mathrm{attack}} - \\mathrm{FLOPs}_{\\mathrm{min}})}{(\\mathrm{FLOPs}_{\\mathrm{max}} - \\mathrm{FLOPs}_{\\mathrm{min}})}\\) where \\(\\mathrm{FLOPs}_{\\mathrm{min}}\\) is the compute of the efficient model and \\(\\mathrm{FLOPs}_{\\mathrm{max}}\\) is that of the original inefficient model. Attack Success is thus capped at \\(100\\%\\) while a negative value denotes a reduction in FLOPs. Note that our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.122 + ], + "angle": 0, + "content": "Attack Success metric illustrates the effectiveness of an attack in reversing the FLOPs reduction of a particular method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.138, + 0.894, + 0.168 + ], + "angle": 0, + "content": "Baselines: We propose three alternative approaches to Slow-Former (ours) to generate the patch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.17, + 0.892, + 0.215 + ], + "angle": 0, + "content": "Random Patch: A simple baseline is to generate a randomly initialized patch. We sample IID pixel values from a uniform distribution between 0 and 255 to create the patch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.217, + 0.894, + 0.322 + ], + "angle": 0, + "content": "NTAP: We consider a standard adversarial patch that is trained to attack the model task performance instead of compute. We use a non-targeted universal adversarial patch (NTAP) to attack the model. We train the patch to fool the model by misclassifying the image it is pasted on. We use the negative of the cross-entropy loss with the predicted and ground-truth labels as the loss to optimize the patch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.323, + 0.894, + 0.414 + ], + "angle": 0, + "content": "TAP: We train a universal targeted adversarial patch (TAP). The patch is optimized to classify all images in the train set to a single fixed category. Similar to NTAP, the adversarial attack here is on task performance and not computation. We experiment with ten randomly generated target category labels and report the averaged metrics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.43, + 0.895, + 0.551 + ], + "angle": 0, + "content": "Implementation Details: We use PyTorch [66] for all experiments. Unless specified, we use a patch of size \\(64 \\times 64\\), train and test on \\(224 \\times 224\\) images, and we paste the patch on the top-left corner. Note that our patch occupies just \\(8\\%\\) of the total area of an input image. We use AdamW [57] optimizer to optimize the patches and use 4 NVIDIA RTX 3090 GPUs for each experiment. We use varying batch sizes and learning rates for each of the computation-efficient methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.895, + 0.734 + ], + "angle": 0, + "content": "ATS Details: As in ATS [25], we replace layers 3 through 9 of ViT networks with the ATS block and set the maximum limit for the number of tokens sampled to 197 for each layer. We train the patch for 2 epochs with a learning rate of 0.4 for ViT-Tiny and \\( lr = 0.2 \\) for ViT-Base and ViT-Small. We use a batch size of 1024 and different loss coefficients for each layer of ATS. For DeiT-Tiny we use [1.0, 0.2, 0.2, 0.2, 0.01, 0.01, 0.01], for DeiT-Small we use [1.0, 0.2, 0.05, 0.01, 0.005, 0.005, 0.005], and for DeiT-Base we use [2.0, 0.1, 0.02, 0.01, 0.005, 0.005, 0.005]. The weights are vastly different at initial and final layers to account for the difference in loss magnitudes across layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "A-ViT Details: When attacking A-VIT[88], the patches are optimized for one epoch with a learning rate of 0.2 and a batch size of 512 (128 × 4 GPUs) using AdamW [57] optimizer. We optimize the patches for 4 epochs for patch length 32 and below. For CIFAR-10 experiments, the images are resized from \\(32 \\times 32\\) to \\(256 \\times 256\\) and a \\(224 \\times 224\\) crop is used as the input. For the training of adversarial defense, we generate 5 patches per epoch of adversarial training and limit the number of iterations for patch generation to 500. The learning rate for patch optimization is increased to 0.8 for faster convergence." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24790" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.53, + 0.215 + ], + "angle": 0, + "content": "Table 1. Computation and Energy Attack on Efficient Vision Transformers: Comparison of the effect of our attack with baselines: No Attack, Random Patch, targeted (TAP), and non-targeted (NTAP) adversarial patches applied to three input-dynamic computation efficient pre-trained models of varying architectures. The maximum possible compute for a given architecture is provided in bold. On A-ViT, we completely undo the efficiency gains obtained by the efficient method through our attack, achieving Attack Success of \\(100\\%\\). We achieve high Attack Success on all approaches while the baselines expectedly do not contribute to increase in compute." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.226, + 0.508, + 0.821 + ], + "angle": 0, + "content": "
MethodAttackModel GFLOPsTop-1 AccAttack Success
A-ViTViT-Tiny1.3--
No attack0.8771.4%-
Random Patch0.8770.8%-1%
TAP0.850.1%-5%
NTAP0.830.1%-10%
SlowFormer (ours)1.34.7%100%
A-ViTViT-Small4.6--
No attack3.778.8%-
Random Patch3.778.4%-2%
TAP3.60.1%-12%
NTAP3.60.1%-7%
SlowFormer (ours)4.62.3%99%
ATSViT-Tiny1.3--
No attack0.8470.3%-
Random Patch0.8369.8%-2%
TAP0.760.1%-17%
NTAP0.610.1%-50%
SlowFormer (ours)1.01.2%35%
ATSViT-Small4.6--
No attack3.179.2%-
Random Patch3.178.6%-1%
TAP3.00.1%-7%
NTAP2.40.1%-47%
SlowFormer (ours)4.01.0%60%
ATSViT-Base17.6--
No attack12.681.3%-
Random Patch12.581.2%-2%
TAP12.00.1%-12%
NTAP11.00.1%-32%
SlowFormer (ours)15.40.2%52%
AdaViTViT-Small4.6--
No attack2.2577.3%-
Random Patch2.2076.9%-2%
TAP2.280.1%1%
NTAP2.150.1%-4%
SlowFormer (ours)3.20.4%40%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.888 + ], + "angle": 0, + "content": "AdaViT Details: For AdaViT[63], we first freeze the weights and use a learning rate of 0.2 and a batch size of 128 with 4 GPUs for patch optimization. We use AdamW [57]" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.089, + 0.923, + 0.216 + ], + "angle": 0, + "content": "Figure 2. Visualization of our Energy Attack on Vision Transformers: We visualize the A-ViT-Small with and without our attack. We use patch size of 32 for the attack (on the top-left corner). We show pruned tokens at layer 8 of A-ViT-Small. Our attack can recover most of the pruned tokens, resulting in increased computation and power consumption. Note that although the patch is reasonably small and is in the corner of the view, it can affect the whole computational flow of the network. This is probably due to the global attention mechanism in transformers." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.226, + 0.915, + 0.822 + ], + "angle": 0, + "content": "
AttackedEfficient ModelAttackedEfficient Model
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.886 + ], + "angle": 0, + "content": "optimizer with no decay and train for 2 epochs with a patch size of \\(64 \\times 64\\). We train on the ImageNet-1k train dataset and evaluate it on the test set." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24791" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.49, + 0.173 + ], + "angle": 0, + "content": "Table 2. Results on CIFAR10 dataset. We report results on CIFAR10 dataset to show that our attack is not specific to ImageNet alone. CIFAR-10 is a small dataset compared to ImageNet and thus results in an extremely efficient A-ViT model. Our attack increases the FLOPs from 0.11 to 0.58 which restores nearly \\(41\\%\\) of the original reduction in the FLOPs." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.184, + 0.457, + 0.314 + ], + "angle": 0, + "content": "
MethodModel FLOPsTop-1 AccAttack Success
ViT-Tiny1.2695.9%-
A-ViT-Tiny0.1195.8%-
SlowFormer (ours)0.5860.2%41%
ATS-Tiny0.8594.7%-
SlowFormer (ours)0.9924.7%34.1%
" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.326, + 0.172, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.327, + 0.321, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.327, + 0.468, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.426, + 0.471, + 0.455 + ], + "angle": 0, + "content": "Figure 3. Visualization of optimized patch: We show the learned universal patches for each of the three efficient methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.471, + 0.566 + ], + "angle": 0, + "content": "Table 4. Effect of patch size: Analysis of the effect of adversarial patch size on the attack success rate on A-ViT. Our attach is reasonably successful even using a small patch size \\((32 \\times 32)\\), which is only \\(2\\%\\) of the image area. Interestingly, a small patch on the corner of the view affects the computational flow of the entire transformer model. This might be due to the global attention mechanism in transformers." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.577, + 0.418, + 0.826 + ], + "angle": 0, + "content": "
Patch Size (Area)Model GFLOPsTop-1 AccuracyAttack Success
ViT-Tiny1.26--
A-ViT-Tiny0.8771.4%-
64 (8%)1.264.7%100%
48 (5%)1.261.8%99%
32 (2%)1.2217.4%90%
16 (0.5%)0.9863.3%27%
ViT-Small4.6--
A-ViT-Small3.778.8%-
64 (8%)4.62.3%99%
48 (5%)4.65.1%98%
32 (2%)4.439.5%78%
16 (0.5%)3.878.2%16%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Results. The results of our attack, SlowFormer , on various methods on ImageNet dataset are shown in table 1. In A-ViT, we successfully recover \\(100\\%\\) of the computation reduced" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.089, + 0.914, + 0.174 + ], + "angle": 0, + "content": "Table 3. Accuracy controlled compute adversarial attack: We attack the efficiency of A-ViT while either maintaining or destroying its classification performance. We observe that our attack can achieve a huge variation in task performance without affecting the Attack Success. The ability to attack the computation without affecting the task performance might be crucial in some applications." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.184, + 0.868, + 0.314 + ], + "angle": 0, + "content": "
AttackModel GFLOPsAttack SuccessTop-1 Acc
ViT-Tiny1.26--
No attack0.87-71.4%
Acc agnostic1.26100%4.7%
Preserve acc1.2392%68.5%
Destroy acc1.26100%0.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.341 + ], + "angle": 0, + "content": "Table 5. Attack with adversarial perturbation on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.37 + ], + "angle": 0, + "content": "The efficient methods are also susceptible universal perturbation based attacks. We use an \\(\\ell_{\\infty}\\) bound on the perturbation." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.38, + 0.91, + 0.661 + ], + "angle": 0, + "content": "
MethodEpsilon (/255.)AttackModel GFLOPsTop-1 AccuracyAttack Success
A-ViTViT-Tiny1.3--
-No attack0.8771.4%-
16SlowFormer1.156.1%73%
32SlowFormer1.250.5%98.4%
ATS-No attack0.8470.3%-
16SlowFormer0.9815.6%30.4%
32SlowFormer1.040.8%43.5%
A-ViTViT-Small4.6--
-No attack3.778.8%-
16SlowFormer4.4820%86.4%
32SlowFormer4.591%98.1%
ATS-No attack3.179.2%-
16SlowFormer3.631.0%33.3%
32SlowFormer3.83.6%46.7%
AdaVit-No attack2.2577.3%-
16SlowFormer3.026.1%31.9%
32SlowFormer3.22.8%40.4%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "by A-ViT. Our attack has an Attack Success of \\(60\\%\\) on ATS and \\(40\\%\\) on AdaViT with ViT-Small. A random patch attack has little effect on both the accuracy and computation of the method. Both standard adversarial attack baselines, TAP and NTAP, reduce the accuracy to nearly \\(0\\%\\). Interestingly, these patches further decrease the computation of the efficient model being attacked. This might be because of the increased importance of adversarial patch tokens to the task and thus reduced importance of other tokens. Targeted patch (TAP) has a significant reduction in FLOPs on the ATS method. Since the token dropping in ATS relies on the distribution of attention values of classification tokens, a sharper distribution due to the increased importance of a token can result in a reduction in computation. The" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24792" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.211 + ], + "angle": 0, + "content": "computation increase with SlowFormer for AdaViT is comparatively low. To investigate, we ran a further experiment using a patch size of \\(224 \\times 224\\) (entire image size) to find the maximum possible computation for an image. This resulted in 4.18 GFLOPs on the ImageNet-1K validation set, which is markedly lower than the limit of 4.6. Using this as an upper-bound of GFLOPs increase, SlowFormer achieves a \\(49\\%\\) Attack Success." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.348 + ], + "angle": 0, + "content": "We report the results on CIFAR-10 dataset in Table 2. The efficient model (A-ViT) drastically reduces the computation from 1.26 GFLOPs to 0.11 GFLOPs. Most of the tokens are dropped as early as layer two in the efficient model. SlowFormer is able to effectively attack even in such extreme scenarios, achieving an Attack Success of \\(40\\%\\) and increasing the mean depth of tokens from nearly one to five. SlowFormer is similarly effective on ATS with an Attack Success of \\(34\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.455 + ], + "angle": 0, + "content": "We additionally visualize the effectiveness of our attack in Figure 2. The un-attacked efficient method retains only highly relevant tokens at the latter layers of the network. However, our attack results in nearly the entire image being passed through all layers of the model for all inputs. In Fig. 3, we visualize the optimized patches for each of the three efficient methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.194, + 0.481 + ], + "angle": 0, + "content": "5.2. Ablations:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.522 + ], + "angle": 0, + "content": "We perform all ablations on the A-ViT approach using their pretrained ViT-Tiny architecture model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.702 + ], + "angle": 0, + "content": "Accuracy controlled compute adversarial attack: As seen in Table 1, our attack can not only increase the computation, but also reduce the model accuracy. This can be desirable or hurtful based on the attacker's goals. A low-accuracy model might be an added benefit, similar to regular adversaries, but might also lead to the victim detecting the attack. We show that it is possible to attack the computation of the model while either preserving or destroying the task performance by additionally employing a task loss in the patch optimization. Table 3 indicates that the accuracy can be significantly modified while maintaining a high Attack Success." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Effect of patch size: We vary the patch size from \\(64 \\times 64\\) to \\(16 \\times 16\\) (just a single token) and report the results in Table 4. Interestingly, our attack with ViT-Small has a \\(73\\%\\) Attack Success with a \\(32 \\times 32\\) patch size, which occupies only \\(2\\%\\) of the input image area." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Effect of patch location: We vary the location of the patch to study its effect on the Attack Success of the model. We randomly sample a location in the image for where we paste the patch on. We perform five such experiments and observe an Attack Success of \\(100\\%\\) for all patch locations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Perturbation attack: While we focus on patch based attacks in this paper, efficient transformers are also susceptible to perturbation based attacks (table 5). In perturbation attacks," + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.894, + 0.174 + ], + "angle": 0, + "content": "Table 6. Defense using adversarial training: We propose and show the impact of our defense for our adversarial attack on A-ViT. Our defense is simply maintaining a set of universal patches and training the model to be robust to a random sample of those at each iteration. The defense reduces the computation to some extent (1.26 to 1.01), but is still far from the unattacked model (0.87)." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.184, + 0.86, + 0.293 + ], + "angle": 0, + "content": "
MethodGFLOPsTop-1 Acc.Attack Success
No attack0.8771.4-
SlowFormer1.264.7%100%
Adv Defense + SlowFormer1.0165.8%34%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.356 + ], + "angle": 0, + "content": "all pixels in the image can be modified, but with an upper bound on the \\(\\ell_{\\infty}\\) norm of the perturbation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.376, + 0.805, + 0.393 + ], + "angle": 0, + "content": "5.3. Adversarial training based defense" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Our simple defense that is adopted from standard adversarial training is explained in Section 4. The results for defending against attacking A-ViT are shown in Table 6. The original A-ViT reduces the GFLOPs from 1.26 to 0.87, our attack increases it back to 1.26 with \\(100\\%\\) attack success. The proposed defense reduces the GFLOPs to 1.01 which is still higher than the original 0.87. We hope our paper encourages the community to develop better defense methods to reduce the vulnerability of efficient vision transformers." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.566, + 0.62, + 0.581 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.894, + 0.852 + ], + "angle": 0, + "content": "Recently, we have seen efficient vision transformer models in which the computation is adaptively modified based on the input. We argue that this is an important research direction and that there will be more progress in this direction in the future. However, we show that the current methods are vulnerable to a universal adversarial patch that increases the computation and thus power consumption at inference time. Our experiments show promising results for three SOTA efficient transformer models, where a small patch that is optimized on the training data can increase the computation to the maximum possible level in the testing data in some settings. We also propose a defense that reduces the effectiveness of our attack. We hope that our paper will encourage the community to study such attacks and develop better defense methods on various machine learning methods, including generative models, that reduce the computation adaptively with the input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment: This work was partially supported by DARPA under Contract No. HR00112190135 and HR00112290115 and NSF grant 1845216." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24793" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.143 + ], + "angle": 0, + "content": "[1] iphone battery capacity. https://bigthink.com/the-future/battery-technology-lags/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.473, + 0.185 + ], + "angle": 0, + "content": "[2] Starship robot. https://www.wevolver.com/specs/starship-technologies-starship-robot.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[3] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[4] Babak Ehteshami Bejnordi, Tijmen Blankevoort, and Max Welling. Batch-shaping for learning conditional channel gated networks. arXiv preprint arXiv:1907.06627, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[7] Jason Ross Brown, Yiren Zhao, Ilia Shumailov, and Robert D Mullins. Dartformer: Finding the best type of attention. arXiv preprint arXiv:2210.00641, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[8] Tom B Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[9] Shaofeng Cai, Yao Shu, and Wei Wang. Dynamic routing networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3597, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.472, + 0.629 + ], + "angle": 0, + "content": "[10] Shijie Cao, Lingxiao Ma, Wencong Xiao, Chen Zhang, Yunxin Liu, Lintao Zhang, Lanshun Nie, and Zhi Yang. Seernet: Predicting convolutional neural network feature-map sparsity through low-bit quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11216-11225, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.686 + ], + "angle": 0, + "content": "[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[13] Guangyi Chen, Chunze Lin, Liangliang Ren, Jiwen Lu, and Jie Zhou. Self-critical attention learning for person re-identification. In ICCV, pages 9637–9646, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[14] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8064-8073, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. Slothbomb: Efficiency poisoning attack against dynamic neural networks. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Simin Chen, Mirazul Haque, Zihe Song, Cong Liu, and Wei Yang. Transslowdown: Efficiency attacks on neural machine translation systems. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[17] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsglowsdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[18] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.894, + 0.501 + ], + "angle": 0, + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[22] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv preprint arXiv:1910.10073, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[23] Quanfu Fan, Chun-Fu (Ricarhd) Chen, Hilde Kuehne, Marco Pistoia, and David Cox. More Is Less: Learning Efficient Video Representations by Temporal Aggregation Modules. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[24] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.894, + 0.785 + ], + "angle": 0, + "content": "[25] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 396-414. Springer, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[26] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res, 23:1-40, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[27] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[28] Michael Figurnov, Maxwell D Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24794" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "Spatially adaptive computation time for residual networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1039-1048, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[29] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[30] Yonggan Fu, Shunyao Zhang, Shang Wu, Cheng Wan, and Yingyan Lin. Patch-fool: Are vision transformers always robust against adversarial perturbations? arXiv preprint arXiv:2203.08392, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[31] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. arXiv preprint arXiv:1810.05331, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[32] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In arXiv preprint arXiv:1412.6572, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.376 + ], + "angle": 0, + "content": "[33] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.471, + 0.418 + ], + "angle": 0, + "content": "[34] Jiaqi Guan, Yang Liu, Qiang Liu, and Jian Peng. Energy-efficient amortized inference with cascaded deep classifiers. arXiv preprint arXiv:1710.03368, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.476 + ], + "angle": 0, + "content": "[35] Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14264-14273, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.471, + 0.518 + ], + "angle": 0, + "content": "[36] Mirazul Haque, Simin Chen, Wasif Arman Haque, Cong Liu, and Wei Yang. Nodeattack: Adversarial attack on the energy consumption of neural odes. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.471, + 0.561 + ], + "angle": 0, + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.471, + 0.631 + ], + "angle": 0, + "content": "[38] Charles Herrmann, Richard Strong Bowen, and Ramin Zabih. Channel selection using gumbel softmax. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII, pages 241-257. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.471, + 0.674 + ], + "angle": 0, + "content": "[39] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[40] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv preprint arXiv:1912.12180, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[41] Sanghyun Hong, Yigitcan Kaya, Ionut-Vlad Modoranu, and Tudor Dumitraş. A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[42] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[43] Weizhe Hua, Yuan Zhou, Christopher M De Sa, Zhiru Zhang, and G Edward Suh. Channel gating neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[44] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.216 + ], + "angle": 0, + "content": "[45] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[46] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of self-attention. arXiv preprint arXiv:2209.04881, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[47] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[48] Soroush Abbasi Koohpayegani and Hamed Piriavash. Sima: Simple softmax-free attention for vision transformers. arXiv preprint arXiv:2206.08898, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[49] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.893, + 0.439 + ], + "angle": 0, + "content": "[50] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.893, + 0.482 + ], + "angle": 0, + "content": "[51] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.893, + 0.537 + ], + "angle": 0, + "content": "[52] Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE international conference on computer vision, pages 5764-5772, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[53] Zhichao Li, Yi Yang, Xiao Liu, Feng Zhou, Shilei Wen, and Wei Xu. Dynamic computational time for visual attention. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1199-1209, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.893, + 0.637 + ], + "angle": 0, + "content": "[54] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. arXiv preprint arXiv:2209.09004, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.893, + 0.705 + ], + "angle": 0, + "content": "[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.893, + 0.775 + ], + "angle": 0, + "content": "[56] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.893, + 0.859 + ], + "angle": 0, + "content": "[58] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[59] Wenhao Lu, Jian Jiao, and Ruofei Zhang. Twinbert: Distilling knowledge to twin-structured bert models for efficient retrieval. arXiv preprint arXiv:2002.06275, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24795" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[60] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[61] Kaleel Mahmood, Rigel Mahmood, and Marten Van Dijk. On the robustness of vision transformers to adversarial examples. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7838-7847, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[62] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers. arXiv preprint arXiv:2110.03860, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.472, + 0.329 + ], + "angle": 0, + "content": "[63] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.412 + ], + "angle": 0, + "content": "[64] Jianhong Pan, Qichen Zheng, Zhipeng Fan, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Gradauto: Energy-oriented attack on dynamic neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 637-653. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[65] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[66] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.472, + 0.636 + ], + "angle": 0, + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.472, + 0.692 + ], + "angle": 0, + "content": "[68] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.472, + 0.746 + ], + "angle": 0, + "content": "[69] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[70] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[71] Mengye Ren, Andrei Pokrovsky, Bin Yang, and Raquel Urtasun. Sbnet: Sparse blocks network for fast inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8711-8720, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[72] Aniruddha Saha, Akshayvarun Subramanya, Koninika B. Patil, and Hamed Pirsavash. Adversarial patches exploiting contextual reasoning in object detection. In ArXiv, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[73] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[74] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3531-3539, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[75] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.893, + 0.355 + ], + "angle": 0, + "content": "[76] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.893, + 0.426 + ], + "angle": 0, + "content": "[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.893, + 0.496 + ], + "angle": 0, + "content": "[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.893, + 0.552 + ], + "angle": 0, + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.893, + 0.607 + ], + "angle": 0, + "content": "[80] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.893, + 0.664 + ], + "angle": 0, + "content": "[81] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the iee/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.893, + 0.72 + ], + "angle": 0, + "content": "[82] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[83] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[84] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[85] Zhenda Xie, Zheng Zhang, Xizhou Zhu, Gao Huang, and Stephen Lin. Spatially adaptive inference with stochastic feature sampling and interpolation. In Computer Vision-ECCV" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24796" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 531-548. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[86] Le Yang, Yizeng Han, Xi Chen, Shiji Song, Jifeng Dai, and Gao Huang. Resolution adaptive networks for efficient inference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2369-2378, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[87] Zerui Yang, Yuhui Xu, Wenrui Dai, and Hongkai Xiong. Dynamic-stride-net: Deep convolutional neural network with dynamic stride. In Optoelectronic Imaging and Multimedia Technology VI, pages 42-53. SPIE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[88] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.472, + 0.402 + ], + "angle": 0, + "content": "[89] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819-10829, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[90] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[91] Zhihang Yuan, Bingzhe Wu, Guangyu Sun, Zheng Liang, Shiwan Zhao, and Weichen Bi. S2dnas: Transforming static cnn model for dynamic inference via neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 175-192. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[92] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.472, + 0.656 + ], + "angle": 0, + "content": "[93] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.472, + 0.74 + ], + "angle": 0, + "content": "[94] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.472, + 0.809 + ], + "angle": 0, + "content": "[95] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.812, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[96] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24797" + } + ] +] \ No newline at end of file diff --git a/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_origin.pdf b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c5ee0ac4cfb00338e19bde46c70dffd9ce06bf8 --- /dev/null +++ b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/93475bbf-f8a6-48bc-a54b-7b5297d3501e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3bce7c252167909fca5b274874854f50692cb0f9a6af28590f7c4e3455166eb +size 2297124 diff --git a/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/full.md b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a69caa8ff6b75d2f2cf2aa8f876b7051d58cf2f1 --- /dev/null +++ b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/full.md @@ -0,0 +1,342 @@ +# SlowFormer: Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers + +K L Navaneet *1 + +Soroush Abbasi Koohpayegani *1 + +1 University of California, Davis + +Essam Sleiman *12† + +2 Harvard University + +Hamed Piriaviash + +# Abstract + +Recently, there has been a lot of progress in reducing the computation of deep models at inference time. These methods can reduce both the computational needs and power usage of deep models. Some of these approaches adaptively scale the compute based on the input instance. We show that such models can be vulnerable to a universal adversarial patch attack, where the attacker optimizes for a patch that when pasted on any image, can increase the compute and power consumption of the model. We run experiments with three different efficient vision transformer methods showing that in some cases, the attacker can increase the computation to the maximum possible level by simply pasting a patch that occupies only $8\%$ of the image area. We also show that a standard adversarial training defense method can reduce some of the attack's success. We believe adaptive efficient methods will be necessary in the future to lower the power usage of expensive deep models, so we hope our paper encourages the community to study the robustness of these methods and develop better defense methods for the proposed attack. Code is available at: + +https://github.com/UCDvision/SlowFormer + +# 1. Introduction + +The field of deep learning has recently made significant progress in improving model efficiency for inference. Two broad categories of methods can be distinguished: 1) those that reduce computation regardless of input, and 2) those that reduce the computation depending on the input (adaptively). Most methods, such as weight pruning or model quantization, belong to the first category which reduces computation by a constant factor regardless of the input. However, in many applications, the complexity of the perception task may differ depending on the input. For example, when a self-driving car is driving between lanes in an empty street, + +the perception may be simpler and require less computation when compared to driving in a busy city street scene. Interestingly, in some applications, simple scenes such as highway driving may account for the majority of the time. Therefore, we believe that adaptive computation reduction will become an increasingly important research area in the future, especially when non-adaptive methods reach the lower bound of computation. + +We argue that reduction of compute usually reduces power usage, which is crucial, particularly in mobile devices that run on battery, e.g., AR/VR headsets, humanoid robots, and drones. For instance, increasing the size of the battery for a drone may lead to a drastic reduction in its range due to the increased battery weight. This is important since the improvement in battery technology is much slower than compute technology. For instance, the battery capacity from iPhone [1] (1st generation) in 2007 to iPhone 15 Pro Max in 2023 improved from 5.18 watt-hour to 17.32 watt-hour (less than 4 times) while the compute has increased by a much larger factor. As an example, a delivery robot like Starship uses a 1,200Wh battery and can run for 12 hours [2], so it uses almost 100 watts for compute and mobility. Hence, an adversary increasing the power consumption of the perception unit by 20 watts, will reduce the battery life by almost $20\%$ , which can be significant. Note that 20 watts increase in power is realistic assuming that it uses two NVIDIA Jetson Xavier NX cards (almost 20 watts each). + +Key idea: Assuming that a perception method is reducing the computation adaptively with the input, an adversary can attack the model by modifying the input to increase the computation and power consumption. Our goal is to design a universal adversarial patch that when pasted on any input image, it will increase the computation of the model leading to increased power consumption. We believe this is an important vulnerability, particularly for safety-critical mobile systems that run on battery. + +Please note that in this paper, we do not experiment with real hardware to measure the power consumption. Instead, we report the change in FLOPs of the inference time assuming that the power consumption is positively correlated with the number of FLOPs, as studied in [73]. + +![](images/08806d9dd9fae0c7d4e594a954d3ccfc9e86cf3bd11aa140d34f226b9a8cafc9.jpg) +Figure 1. Computation and Energy Attack on Vision Transformers: Given a pre-trained input-dependent computation efficient model, the adversary first attaches an adversarial patch to all images in a dataset and optimizes this patch with our method such that it maximizes the model's computation for each sample. During inference, the adversary modifies the input of the victim's model by applying the learnt patch to it. This results in an increase in compute in the victim's model. The attack will thus potentially slowdown and also lead to increased energy consumption and CPU/GPU usage on the victim's device. + +We design our attack, SlowFormer, for three different methods (A-ViT [88], ATS [25], and Ada-VIT [63]) that reduce the computation of vision transformers. These methods generally identify the importance of each token for the final task and drop the insignificant ones to reduce the computation. We show that in all three cases, our attack can increase the computation by a large margin, returning it to the full-compute level (non-efficient baseline) for all images in some settings. Our threat model is agnostic to the accuracy of the model and attacks the computation and power consumption only. Figure 1 shows our attack. + +There are some prior works that design a pixel-level perturbation attack to increase the compute of the model. We believe universal patch-based attacks that do not change with the input image (generalize from training data to test data) are much more practical in real applications. Note that to modify the pixel values on a real robot, the attacker needs to access and manipulate the image between the camera and compute modules, which is impossible in many applications. + +Contributions: We show that efficient vision transformer methods are vulnerable to a universal patch attack that can increase their compute and power usage. We demonstrate this through experiments on three different efficient transformer methods. We show that an adversarial training defense can reduce attack success to some extent. + +# 2. Related Work + +Vision Transformers: The popularity of transformers [79] in vision has grown rapidly since the introduction of the first vision transformer [20, 78]. Recent works demonstrate the strength of vision transformers on a variety of computer vision tasks [11, 18, 21, 56, 69, 77, 90, 93, 94, 96]. Moreover, transformers are the backbone of recent Self-Supervised + +Learning (SSL) models [12, 37], and vision-language models [67]. In our work, we design an attack to target the computation and energy efficiency of vision transformers. + +Efficient Vision Transformers: Due to the recent importance and popularity of vision transformers, many works have started to study the efficiency of vision transformers [7, 46, 89]. To accomplish this, some lines of work study token pruning with the goal of removing uninformative tokens in each layer [25, 62, 63, 68, 88]. ToMe [6] merges similar tokens in each layer to decrease the computation. Some works address quadratic computation of self-attention module by introducing linear attention [3, 45, 48, 58, 74]. Efficient architectures [40, 55] that limit the attention span of each token have been proposed to improve efficiency. In our paper, we attack token pruning based efficient transformers where the computation varies based on the input samples [25, 63, 88]. + +Dynamic Computation: There are different approaches to reducing the computation of vision models, including knowledge distillation to lighter network [39, 59], model quantization [54, 70] and model pruning [51]. In these methods, the computation is fixed during inference. In contrast to the above models, some works address efficiency by having variable computation based on the input. The intuition behind this direction is that not all samples require the same amount of computation. Several recent works have developed models that dynamically exit early or skip layers [5, 22, 28, 33, 34, 44, 76, 80, 82] and selectively activate neurons, channels or branches for dynamic width [4, 9, 13, 26, 31, 38, 43, 91] depending on the complexity of the input sample. Zhou et al. show that not all locations in an image contribute equally to the predictions of a CNN model [95], encouraging a new line of work to make CNNs more + +efficient through spatially dynamic computation. Pixel-Wise dynamic architectures [10, 14, 23, 47, 71, 81, 85] learn to focus on the significant pixels for the required task while Region-Level dynamic architectures perform adaptive inference on the regions or patches of the input [29, 53]. Finally, lowering the resolution of inputs decreases computation, but at the cost of performance. Conventional CNNs process all regions of an image equally, however, this can be inefficient if some regions are "easier" to process than others [42]. Correspondingly, [86, 87] develop methods to adaptively scale the resolution of images. + +Transformers have recently become extremely popular for vision tasks, resulting in the release of a few input-dynamic transformer architectures [24, 63, 88]. Fayyaz et al. [24] introduce a differentiable parameter-free Adaptive Token Sampler (ATS) module which scores and adaptively samples significant tokens. ATS can be plugged into any existing vision transformer architecture. A-ViT [88] reduces the number of tokens in vision transformers by discarding redundant spatial tokens. Meng et al. [63] propose AdaViT, which trains a decision network to dynamically choose which patch, head, and block to keep/activate throughout the backbone. + +Adversarial Attack: Adversarial attacks are designed to fool models by applying a targeted perturbation or patch on an image sample during inference [32, 50, 75]. These methods can be incorporated into the training set and optimized to fool the model. Correspondingly, defenses have been proposed to mitigate the effects of these attacks [27, 52, 65, 84]. Patch-Fool [30] considers adversarial patch-based attacks on transformers. Some recent works [61, 83, 92] also study and design methods for the transferability of adversarial attacks on vision transformers. However, most prior adversarial attacks target model accuracy, ignoring model efficiency. + +Energy Attack: Very recently, there have been a few works on energy adversarial attacks on neural networks. In ILFO [35], Haque et al. attack two CNN-based input-dynamic methods: SkipNet [82] and SACT [28] using image specific perturbation. DeepSloth [41] attack focuses on slowing down early-exit methods, reducing their energy efficiency by $90 - 100\%$ . GradAuto [64] successfully attacks methods that are both dynamic width and dynamic depth. NICGSlowDown and TransSlowDown [16, 17] attack neural image caption generation and neural machine translation methods, respectively. All these methods primarily employ image specific perturbation based adversarial attack. Sloth-Bomb injects efficiency backdoors to input-adaptive dynamic neural networks [15] and NodeAttack [36] attacks Neural Ordinary Differential Equation models, which use ordinary differential equation solving to dynamically predict the output of a neural network. Our work is closely related to ILFO [35], DeepSloth [41] and GradAuto [64] in that we attack the computational efficiency of networks. However, unlike these methods, we focus on designing an adversarial patch-based + +attack that is universal and on vision transformers. We additionally provide a potential defense for our attack. We use a patch that generalizes from train to test set and thus we do not optimize per sample during inference. Our patch-based attack is especially suited for transformer architectures [30]. + +# 3. Computation and Energy Attack + +# 3.1. Threat Model: + +We consider a scenario where the adversary has access to the victim's trained deep model and modifies its input such that the energy consumption and computational demand of the model is increased. The attack is agnostic to model accuracy. To make the setting more practical, instead of perturbing the entire image, we assume that the adversary can modify the input image by only pasting a patch [8, 72] on it and that the patch is universal, that is, image independent. During inference, a pretrained patch is pasted on the test image before propagating it through the network. + +In this paper, we attack three state-of-the-art efficient transformers. Since the attacker manipulates only the input image and not the network parameters, the attacked model must have dynamic computation that depends on the input image. As stated earlier, several recent works have developed such adaptive efficient models and we believe that they will be more popular in the future due to the limits of non-adaptive efficiency improvement. + +# 3.2. Attack on Efficient Vision Transformers: + +Universal Adversarial Patch: We use an adversarial patch to attack the computational efficiency of transforms. The learned patch is universal, that is, a single patch is trained and is used during inference on all test images. The patch generalizes across images but not across models. The patch optimization is performed only on the train set. The patch is pasted on an image by replacing the image pixels using the patch. We assume the patch location does not change from train to test. The patch pixels are initialized using i.i.d. samples from a uniform distribution over [0, 255]. During each training iteration, the patch is pasted on the mini-batch samples and is updated to increase the computation of the attacked network. The patch values are projected onto [0, 255] and quantized to 256 uniform levels after each iteration. Note that we use a pretrained network and do not update its parameters either in the training or in the evaluation of our attack. During inference, the trained patch is pasted on the test images and the computational efficiency of the network on the adversarial image is measured. + +Here, we focus on three methods employing vision transformers for the task of image classification. All these methods modify the computational flow of the network based on the input image for faster inference. A pretrained model is used for the attack and is not modified during + +our adversarial patch training. We first provide a brief background of each method before describing our attack. + +# Attacking A-ViT : + +Background: A-ViT [88] adaptively prunes image tokens to achieve speed-up in inference with minimal loss in accuracy. For a given image, a dropped token will not be used again in the succeeding layers of the network. Let $x$ be the input image and $\{t^l\}_{1:K}$ be the corresponding $K$ tokens at layer $l$ . An input-dependent halting score $h_k^l$ for a token $k$ at layer $l$ is calculated and the token is dropped at layer $N_k$ where its cumulative halting score exceeds a fixed threshold value $1 - \epsilon$ for the first time. The token is propagated until the final layer if its score never exceeds the threshold. Instead of introducing a new parameter for $h_k^l$ , the first dimension of each token is used to predict the halting score for the corresponding token. The network is trained to maximize the cumulative halting score at each layer and thus drop the tokens earlier. The loss, termed ponder loss, is given by: + +$$ +\mathcal {L} _ {\text {p o n d e r}} = \frac {1}{K} \sum_ {k = 1} ^ {K} \left(N _ {k} + r _ {k}\right), \quad r _ {k} = 1 - \sum_ {l = 1} ^ {N _ {k - 1}} h _ {k} ^ {l} \tag {1} +$$ + +Additionally, A-ViT enforces a Gaussian prior on the expected halting scores of all tokens via $KL$ -divergence based distribution loss, $\mathcal{L}_{\mathrm{distr}}$ . These loss terms are minimized along with the task-specific loss $\mathcal{L}_{\mathrm{task}}$ . Thus, the overall training objective is $\mathcal{L} = \mathcal{L}_{\mathrm{task}} + \alpha_d\mathcal{L}_{\mathrm{distr}} + \alpha_p\mathcal{L}_{\mathrm{ponder}}$ where $\alpha_{d}$ and $\alpha_{p}$ are hyperparameters. + +Attack: Here, we train the patch to increase the inference compute of a trained A-ViT model. Since we are interested in the compute and not task-specific performance, we simply use $-\left(\alpha_{d}\mathcal{L}_{\mathrm{distr.}} + \alpha_{p}\mathcal{L}_{\mathrm{ponder}}\right)$ as our loss. It is possible to preserve (or hurt) the task performance by additionally using $+\mathcal{L}_{\mathrm{task}}$ (or $-\mathcal{L}_{\mathrm{task}}$ ) in the loss formulation. + +# Attacking AdaViT: + +**Background:** To improve the inference efficiency of vision transformers, AdaViT [63] inserts and trains a decision network before each transformer block to dynamically decide which patches, self-attention heads, and transformer blocks to keep/activate throughout the backbone. The $l^{\text{th}}$ block's decision network consists of three linear layers with parameters $W_{l} = W_{l}^{p}, W_{l}^{h}, W_{l}^{b}$ which are then multiplied by each block's input $Z_{l}$ to get $m$ . + +$$ +\left(m _ {l} ^ {p}, m _ {l} ^ {h}, m _ {l} ^ {b}\right) = \left(W _ {l} ^ {p}, W _ {l} ^ {h}, W _ {l} ^ {b}\right) Z _ {l} \tag {2} +$$ + +The value $m$ is then passed to sigmoid function to convert it to a probability value used to make the binary decision of keep/discard. Gumbel-Softmax trick [60] is used to make this decision differentiable during training. Let $M$ be the + +keep/discard mask after applying Gumbel-Softmax on $m$ . The loss on computation is given by: + +$$ +\begin{array}{l} \mathcal {L} _ {u s a g e} = (\frac {1}{D _ {p}} \sum_ {d = 1} ^ {D _ {p}} M _ {d} ^ {p} - \gamma_ {p}) ^ {2} + (\frac {1}{D _ {h}} \sum_ {d = 1} ^ {D _ {h}} M _ {d} ^ {h} - \gamma_ {h}) ^ {2} \\ + \left(\frac {1}{D _ {b}} \sum_ {d = 1} ^ {D _ {b}} M _ {d} ^ {b} - \gamma_ {b}\right) ^ {2} \tag {3} \\ \end{array} +$$ + +where $D_{p}, D_{h}, D_{b}$ represent the number of total patches, heads, and blocks of the entire transformer, respectively. $\gamma_{p}, \gamma_{h}, \gamma_{b}$ denote the target computation budgets i.e. the percentage of patches/heads/blocks to keep. The total loss is a combination of task loss (cross-entropy) and computation loss: $\mathcal{L} = \mathcal{L}_{ce} + \mathcal{L}_{usage}$ . + +Attack: To attack this model, we train the patch to maximize the computation loss $\mathcal{L}_{usage}$ . More specifically, we set the computation-target $\gamma$ values to 0 and negate the $\mathcal{L}_{usage}$ term in Eq. 3. As a result, the patch is optimized to maximize the probability of keeping the corresponding patch (p), attention head (h), and transformer block (b). We can also choose to attack the prediction performance by selectively including or excluding the $\mathcal{L}_{ce}$ term. + +# Attacking ATS: + +Background: Given $N$ tokens with the first one as the classification token, the transformer attention matrix $\mathcal{A}$ is calculated by the following dot product: + +$\mathcal{A} = \mathrm{Softmax}\left(\mathcal{Q}\mathcal{K}^T /\sqrt{d}\right)$ where $\sqrt{d}$ is a scaling coefficient, $d$ is the dimension of tokens, $\mathcal{Q},\mathcal{K}$ and $\nu$ are the query, key and value matrices, respectively. The value $\mathcal{A}_{1,j}$ denotes the attention of the classification token to token $j$ . ATS [25] assigns importance score $S_{j}$ for each token $j$ by measuring how much the classification token attends to it: + +$$ +\mathcal {S} _ {j} = \frac {\mathcal {A} _ {1 , j} \times \left\| \mathcal {V} _ {j} \right\|}{\sum_ {i = 2} \mathcal {A} _ {1 , i} \times \left\| \mathcal {V} _ {i} \right\|} \tag {4} +$$ + +The importance scores are converted to probabilities and are used to sample tokens, where tokens with a lower score have more of a chance of being dropped. + +Attack: Since ATS uses inverse transform sampling, it results in fewer samples if the importance distribution is sharp. To maximize the computation in ATS, we aim to obtain a distribution of scores with high entropy to maximize the number of retained tokens. Therefore, we optimize the patch so that the attention of the classification token over other tokens is a uniform distribution using the following MSE loss: + +$$ +\mathcal {L} = \sum_ {i = 2} ^ {N} \left| \left| \mathcal {A} _ {1, i} - \frac {1}{N} \right| \right| _ {2} ^ {2} \tag {5} +$$ + +Note that one can optimize $S$ to be uniform, but we found the above loss to be easier to optimize. For a multi-head + +attention layer, we calculate the loss for each head and then sum the loss over all heads. Moreover, ATS can be applied to any layer of a vision transformer. For a given model, we apply our loss at all ATS layers and use a weighted summation for optimization. + +# 4. Defense + +An obvious defense, although weak, will be to use nondynamic efficient methods only, e.g., weight pruning, where the reduction in compute is deterministic and does not depend on the input. However, most such methods do not achieve high levels of computation efficiency since they do not take advantage of the simplicity of images. + +We adopt standard adversarial training as a better defense method for our attack. In the standard way, at each iteration of training the model, one would load an image, attack it, and then use it with correct labels in training the model. We cannot adopt this out-of-the-box since our attack generalizes across images and is not dependent on a single image only. To do this, we maintain a set of adversarial patches, and at each iteration sample one of them randomly (uniformly), and use it at the input while optimizing the original loss of the efficient model to train a robust model. To adapt the set of adversarial patches to the model being trained, we interrupt the training at every $20\%$ mark of each epoch and optimize for a new patch to be added to the set of patches. To limit the computational cost of training, we use only 500 iterations to optimize for a new patch, which results in an attack with reasonable accuracy compared to our main results. + +# 5. Experiments + +# 5.1. Attack on Efficient Vision Transformers + +Dataset: We evaluate the effectiveness of our attack on two datasets: ImageNet-1K [19] and CIFAR-10 [49]. ImageNet-1K contains 1.3M images in the train set and 50K images in the validation set with 1000 total categories. CIFAR-10 has 50K images for training and 10K images for validation with 10 total categories. + +Metrics: We report Top-1 accuracy and average computation in terms of GFLOPs for both attacked and unattacked models. Similar to Attack Success Rate in a standard adversarial attack, we introduce a metric: Attack Success to quantify the efficacy of the attack. We define Attack Success as the number of FLOPs increased by the attack divided by the number of FLOPs decreased by the efficient method. Attack Success = $\frac{(\mathrm{FLOPs}_{\mathrm{attack}} - \mathrm{FLOPs}_{\mathrm{min}})}{(\mathrm{FLOPs}_{\mathrm{max}} - \mathrm{FLOPs}_{\mathrm{min}})}$ where $\mathrm{FLOPs}_{\mathrm{min}}$ is the compute of the efficient model and $\mathrm{FLOPs}_{\mathrm{max}}$ is that of the original inefficient model. Attack Success is thus capped at $100\%$ while a negative value denotes a reduction in FLOPs. Note that our + +Attack Success metric illustrates the effectiveness of an attack in reversing the FLOPs reduction of a particular method. + +Baselines: We propose three alternative approaches to Slow-Former (ours) to generate the patch. + +Random Patch: A simple baseline is to generate a randomly initialized patch. We sample IID pixel values from a uniform distribution between 0 and 255 to create the patch. + +NTAP: We consider a standard adversarial patch that is trained to attack the model task performance instead of compute. We use a non-targeted universal adversarial patch (NTAP) to attack the model. We train the patch to fool the model by misclassifying the image it is pasted on. We use the negative of the cross-entropy loss with the predicted and ground-truth labels as the loss to optimize the patch. + +TAP: We train a universal targeted adversarial patch (TAP). The patch is optimized to classify all images in the train set to a single fixed category. Similar to NTAP, the adversarial attack here is on task performance and not computation. We experiment with ten randomly generated target category labels and report the averaged metrics. + +Implementation Details: We use PyTorch [66] for all experiments. Unless specified, we use a patch of size $64 \times 64$ , train and test on $224 \times 224$ images, and we paste the patch on the top-left corner. Note that our patch occupies just $8\%$ of the total area of an input image. We use AdamW [57] optimizer to optimize the patches and use 4 NVIDIA RTX 3090 GPUs for each experiment. We use varying batch sizes and learning rates for each of the computation-efficient methods. + +ATS Details: As in ATS [25], we replace layers 3 through 9 of ViT networks with the ATS block and set the maximum limit for the number of tokens sampled to 197 for each layer. We train the patch for 2 epochs with a learning rate of 0.4 for ViT-Tiny and $lr = 0.2$ for ViT-Base and ViT-Small. We use a batch size of 1024 and different loss coefficients for each layer of ATS. For DeiT-Tiny we use [1.0, 0.2, 0.2, 0.2, 0.01, 0.01, 0.01], for DeiT-Small we use [1.0, 0.2, 0.05, 0.01, 0.005, 0.005, 0.005], and for DeiT-Base we use [2.0, 0.1, 0.02, 0.01, 0.005, 0.005, 0.005]. The weights are vastly different at initial and final layers to account for the difference in loss magnitudes across layers. + +A-ViT Details: When attacking A-VIT[88], the patches are optimized for one epoch with a learning rate of 0.2 and a batch size of 512 (128 × 4 GPUs) using AdamW [57] optimizer. We optimize the patches for 4 epochs for patch length 32 and below. For CIFAR-10 experiments, the images are resized from $32 \times 32$ to $256 \times 256$ and a $224 \times 224$ crop is used as the input. For the training of adversarial defense, we generate 5 patches per epoch of adversarial training and limit the number of iterations for patch generation to 500. The learning rate for patch optimization is increased to 0.8 for faster convergence. + +Table 1. Computation and Energy Attack on Efficient Vision Transformers: Comparison of the effect of our attack with baselines: No Attack, Random Patch, targeted (TAP), and non-targeted (NTAP) adversarial patches applied to three input-dynamic computation efficient pre-trained models of varying architectures. The maximum possible compute for a given architecture is provided in bold. On A-ViT, we completely undo the efficiency gains obtained by the efficient method through our attack, achieving Attack Success of $100\%$ . We achieve high Attack Success on all approaches while the baselines expectedly do not contribute to increase in compute. + +
MethodAttackModel GFLOPsTop-1 AccAttack Success
A-ViTViT-Tiny1.3--
No attack0.8771.4%-
Random Patch0.8770.8%-1%
TAP0.850.1%-5%
NTAP0.830.1%-10%
SlowFormer (ours)1.34.7%100%
A-ViTViT-Small4.6--
No attack3.778.8%-
Random Patch3.778.4%-2%
TAP3.60.1%-12%
NTAP3.60.1%-7%
SlowFormer (ours)4.62.3%99%
ATSViT-Tiny1.3--
No attack0.8470.3%-
Random Patch0.8369.8%-2%
TAP0.760.1%-17%
NTAP0.610.1%-50%
SlowFormer (ours)1.01.2%35%
ATSViT-Small4.6--
No attack3.179.2%-
Random Patch3.178.6%-1%
TAP3.00.1%-7%
NTAP2.40.1%-47%
SlowFormer (ours)4.01.0%60%
ATSViT-Base17.6--
No attack12.681.3%-
Random Patch12.581.2%-2%
TAP12.00.1%-12%
NTAP11.00.1%-32%
SlowFormer (ours)15.40.2%52%
AdaViTViT-Small4.6--
No attack2.2577.3%-
Random Patch2.2076.9%-2%
TAP2.280.1%1%
NTAP2.150.1%-4%
SlowFormer (ours)3.20.4%40%
+ +AdaViT Details: For AdaViT[63], we first freeze the weights and use a learning rate of 0.2 and a batch size of 128 with 4 GPUs for patch optimization. We use AdamW [57] + +Figure 2. Visualization of our Energy Attack on Vision Transformers: We visualize the A-ViT-Small with and without our attack. We use patch size of 32 for the attack (on the top-left corner). We show pruned tokens at layer 8 of A-ViT-Small. Our attack can recover most of the pruned tokens, resulting in increased computation and power consumption. Note that although the patch is reasonably small and is in the corner of the view, it can affect the whole computational flow of the network. This is probably due to the global attention mechanism in transformers. + +
AttackedEfficient ModelAttackedEfficient Model
+ +optimizer with no decay and train for 2 epochs with a patch size of $64 \times 64$ . We train on the ImageNet-1k train dataset and evaluate it on the test set. + +Table 2. Results on CIFAR10 dataset. We report results on CIFAR10 dataset to show that our attack is not specific to ImageNet alone. CIFAR-10 is a small dataset compared to ImageNet and thus results in an extremely efficient A-ViT model. Our attack increases the FLOPs from 0.11 to 0.58 which restores nearly $41\%$ of the original reduction in the FLOPs. + +
MethodModel FLOPsTop-1 AccAttack Success
ViT-Tiny1.2695.9%-
A-ViT-Tiny0.1195.8%-
SlowFormer (ours)0.5860.2%41%
ATS-Tiny0.8594.7%-
SlowFormer (ours)0.9924.7%34.1%
+ +![](images/84741892e84b8e36c7ca7cd0580b06f8d4308cb1212a4b01896dc10fe3f7e223.jpg) +Figure 3. Visualization of optimized patch: We show the learned universal patches for each of the three efficient methods. + +![](images/cff71724dc6dfc20de777051a6fffde590fd94fd51cf3f04e6e3dd981910b7f1.jpg) + +![](images/e43cce4c5073a7ea8d1a9a2865440e96238cfed9feee7e47454ac716ec047a96.jpg) + +Table 4. Effect of patch size: Analysis of the effect of adversarial patch size on the attack success rate on A-ViT. Our attach is reasonably successful even using a small patch size $(32 \times 32)$ , which is only $2\%$ of the image area. Interestingly, a small patch on the corner of the view affects the computational flow of the entire transformer model. This might be due to the global attention mechanism in transformers. + +
Patch Size (Area)Model GFLOPsTop-1 AccuracyAttack Success
ViT-Tiny1.26--
A-ViT-Tiny0.8771.4%-
64 (8%)1.264.7%100%
48 (5%)1.261.8%99%
32 (2%)1.2217.4%90%
16 (0.5%)0.9863.3%27%
ViT-Small4.6--
A-ViT-Small3.778.8%-
64 (8%)4.62.3%99%
48 (5%)4.65.1%98%
32 (2%)4.439.5%78%
16 (0.5%)3.878.2%16%
+ +Results. The results of our attack, SlowFormer , on various methods on ImageNet dataset are shown in table 1. In A-ViT, we successfully recover $100\%$ of the computation reduced + +Table 3. Accuracy controlled compute adversarial attack: We attack the efficiency of A-ViT while either maintaining or destroying its classification performance. We observe that our attack can achieve a huge variation in task performance without affecting the Attack Success. The ability to attack the computation without affecting the task performance might be crucial in some applications. + +
AttackModel GFLOPsAttack SuccessTop-1 Acc
ViT-Tiny1.26--
No attack0.87-71.4%
Acc agnostic1.26100%4.7%
Preserve acc1.2392%68.5%
Destroy acc1.26100%0.1%
+ +The efficient methods are also susceptible universal perturbation based attacks. We use an $\ell_{\infty}$ bound on the perturbation. + +Table 5. Attack with adversarial perturbation on ImageNet. + +
MethodEpsilon (/255.)AttackModel GFLOPsTop-1 AccuracyAttack Success
A-ViTViT-Tiny1.3--
-No attack0.8771.4%-
16SlowFormer1.156.1%73%
32SlowFormer1.250.5%98.4%
ATS-No attack0.8470.3%-
16SlowFormer0.9815.6%30.4%
32SlowFormer1.040.8%43.5%
A-ViTViT-Small4.6--
-No attack3.778.8%-
16SlowFormer4.4820%86.4%
32SlowFormer4.591%98.1%
ATS-No attack3.179.2%-
16SlowFormer3.631.0%33.3%
32SlowFormer3.83.6%46.7%
AdaVit-No attack2.2577.3%-
16SlowFormer3.026.1%31.9%
32SlowFormer3.22.8%40.4%
+ +by A-ViT. Our attack has an Attack Success of $60\%$ on ATS and $40\%$ on AdaViT with ViT-Small. A random patch attack has little effect on both the accuracy and computation of the method. Both standard adversarial attack baselines, TAP and NTAP, reduce the accuracy to nearly $0\%$ . Interestingly, these patches further decrease the computation of the efficient model being attacked. This might be because of the increased importance of adversarial patch tokens to the task and thus reduced importance of other tokens. Targeted patch (TAP) has a significant reduction in FLOPs on the ATS method. Since the token dropping in ATS relies on the distribution of attention values of classification tokens, a sharper distribution due to the increased importance of a token can result in a reduction in computation. The + +computation increase with SlowFormer for AdaViT is comparatively low. To investigate, we ran a further experiment using a patch size of $224 \times 224$ (entire image size) to find the maximum possible computation for an image. This resulted in 4.18 GFLOPs on the ImageNet-1K validation set, which is markedly lower than the limit of 4.6. Using this as an upper-bound of GFLOPs increase, SlowFormer achieves a $49\%$ Attack Success. + +We report the results on CIFAR-10 dataset in Table 2. The efficient model (A-ViT) drastically reduces the computation from 1.26 GFLOPs to 0.11 GFLOPs. Most of the tokens are dropped as early as layer two in the efficient model. SlowFormer is able to effectively attack even in such extreme scenarios, achieving an Attack Success of $40\%$ and increasing the mean depth of tokens from nearly one to five. SlowFormer is similarly effective on ATS with an Attack Success of $34\%$ . + +We additionally visualize the effectiveness of our attack in Figure 2. The un-attacked efficient method retains only highly relevant tokens at the latter layers of the network. However, our attack results in nearly the entire image being passed through all layers of the model for all inputs. In Fig. 3, we visualize the optimized patches for each of the three efficient methods. + +# 5.2. Ablations: + +We perform all ablations on the A-ViT approach using their pretrained ViT-Tiny architecture model. + +Accuracy controlled compute adversarial attack: As seen in Table 1, our attack can not only increase the computation, but also reduce the model accuracy. This can be desirable or hurtful based on the attacker's goals. A low-accuracy model might be an added benefit, similar to regular adversaries, but might also lead to the victim detecting the attack. We show that it is possible to attack the computation of the model while either preserving or destroying the task performance by additionally employing a task loss in the patch optimization. Table 3 indicates that the accuracy can be significantly modified while maintaining a high Attack Success. + +Effect of patch size: We vary the patch size from $64 \times 64$ to $16 \times 16$ (just a single token) and report the results in Table 4. Interestingly, our attack with ViT-Small has a $73\%$ Attack Success with a $32 \times 32$ patch size, which occupies only $2\%$ of the input image area. + +Effect of patch location: We vary the location of the patch to study its effect on the Attack Success of the model. We randomly sample a location in the image for where we paste the patch on. We perform five such experiments and observe an Attack Success of $100\%$ for all patch locations. + +Perturbation attack: While we focus on patch based attacks in this paper, efficient transformers are also susceptible to perturbation based attacks (table 5). In perturbation attacks, + +Table 6. Defense using adversarial training: We propose and show the impact of our defense for our adversarial attack on A-ViT. Our defense is simply maintaining a set of universal patches and training the model to be robust to a random sample of those at each iteration. The defense reduces the computation to some extent (1.26 to 1.01), but is still far from the unattacked model (0.87). + +
MethodGFLOPsTop-1 Acc.Attack Success
No attack0.8771.4-
SlowFormer1.264.7%100%
Adv Defense + SlowFormer1.0165.8%34%
+ +all pixels in the image can be modified, but with an upper bound on the $\ell_{\infty}$ norm of the perturbation. + +# 5.3. Adversarial training based defense + +Our simple defense that is adopted from standard adversarial training is explained in Section 4. The results for defending against attacking A-ViT are shown in Table 6. The original A-ViT reduces the GFLOPs from 1.26 to 0.87, our attack increases it back to 1.26 with $100\%$ attack success. The proposed defense reduces the GFLOPs to 1.01 which is still higher than the original 0.87. We hope our paper encourages the community to develop better defense methods to reduce the vulnerability of efficient vision transformers. + +# 6. Conclusion + +Recently, we have seen efficient vision transformer models in which the computation is adaptively modified based on the input. We argue that this is an important research direction and that there will be more progress in this direction in the future. However, we show that the current methods are vulnerable to a universal adversarial patch that increases the computation and thus power consumption at inference time. Our experiments show promising results for three SOTA efficient transformer models, where a small patch that is optimized on the training data can increase the computation to the maximum possible level in the testing data in some settings. We also propose a defense that reduces the effectiveness of our attack. We hope that our paper will encourage the community to study such attacks and develop better defense methods on various machine learning methods, including generative models, that reduce the computation adaptively with the input. + +Acknowledgment: This work was partially supported by DARPA under Contract No. HR00112190135 and HR00112290115 and NSF grant 1845216. + +# References + +[1] iphone battery capacity. https://bigthink.com/the-future/battery-technology-lags/.1 +[2] Starship robot. https://www.wevolver.com/specs/starship-technologies-starship-robot.1 +[3] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34, 2021. 2 +[4] Babak Ehteshami Bejnordi, Tijmen Blankevoort, and Max Welling. Batch-shaping for learning conditional channel gated networks. arXiv preprint arXiv:1907.06627, 2019. 2 +[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2 +[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2 +[7] Jason Ross Brown, Yiren Zhao, Ilia Shumailov, and Robert D Mullins. Dartformer: Finding the best type of attention. arXiv preprint arXiv:2210.00641, 2022. 2 +[8] Tom B Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017. 3 +[9] Shaofeng Cai, Yao Shu, and Wei Wang. Dynamic routing networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3597, 2021. 2 +[10] Shijie Cao, Lingxiao Ma, Wencong Xiao, Chen Zhang, Yunxin Liu, Lintao Zhang, Lanshun Nie, and Zhi Yang. Seernet: Predicting convolutional neural network feature-map sparsity through low-bit quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11216-11225, 2019. 3 +[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2 +[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021. 2 +[13] Guangyi Chen, Chunze Lin, Liangliang Ren, Jiwen Lu, and Jie Zhou. Self-critical attention learning for person re-identification. In ICCV, pages 9637–9646, 2019. 2 +[14] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8064-8073, 2021. 3 +[15] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. Slothbomb: Efficiency poisoning attack against dynamic neural networks. 3 + +[16] Simin Chen, Mirazul Haque, Zihe Song, Cong Liu, and Wei Yang. Transslowdown: Efficiency attacks on neural machine translation systems. 2021. 3 +[17] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsglowsdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 3 +[18] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2 +[22] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv preprint arXiv:1910.10073, 2019. 2 +[23] Quanfu Fan, Chun-Fu (Ricarhd) Chen, Hilde Kuehne, Marco Pistoia, and David Cox. More Is Less: Learning Efficient Video Representations by Temporal Aggregation Modules. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 3 +[24] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers, 2021. 3 +[25] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 396-414. Springer, 2022. 2, 4, 5 +[26] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res, 23:1-40, 2021. 2 +[27] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3 +[28] Michael Figurnov, Maxwell D Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov. + +Spatially adaptive computation time for residual networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1039-1048, 2017. 2, 3 +[29] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[30] Yonggan Fu, Shunyao Zhang, Shang Wu, Cheng Wan, and Yingyan Lin. Patch-fool: Are vision transformers always robust against adversarial perturbations? arXiv preprint arXiv:2203.08392, 2022. 3 +[31] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. arXiv preprint arXiv:1810.05331, 2018. 2 +[32] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In arXiv preprint arXiv:1412.6572, 2014. 3 +[33] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2 +[34] Jiaqi Guan, Yang Liu, Qiang Liu, and Jian Peng. Energy-efficient amortized inference with cascaded deep classifiers. arXiv preprint arXiv:1710.03368, 2017. 2 +[35] Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14264-14273, 2020. 3 +[36] Mirazul Haque, Simin Chen, Wasif Arman Haque, Cong Liu, and Wei Yang. Nodeattack: Adversarial attack on the energy consumption of neural odes. 2021. 3 +[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2 +[38] Charles Herrmann, Richard Strong Bowen, and Ramin Zabih. Channel selection using gumbel softmax. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII, pages 241-257. Springer, 2020. 2 +[39] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[40] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv preprint arXiv:1912.12180, 2019. 2 +[41] Sanghyun Hong, Yigitcan Kaya, Ionut-Vlad Modoranu, and Tudor Dumitraş. A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432, 2020. 3 +[42] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3 +[43] Weizhe Hua, Yuan Zhou, Christopher M De Sa, Zhiru Zhang, and G Edward Suh. Channel gating neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2 + +[44] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2 +[45] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 2 +[46] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of self-attention. arXiv preprint arXiv:2209.04881, 2022. 2 +[47] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 3 +[48] Soroush Abbasi Koohpayegani and Hamed Piriavash. Sima: Simple softmax-free attention for vision transformers. arXiv preprint arXiv:2206.08898, 2022. 2 +[49] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5 +[50] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3 +[51] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 2 +[52] Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE international conference on computer vision, pages 5764-5772, 2017. 3 +[53] Zhichao Li, Yi Yang, Xiao Liu, Feng Zhou, Shilei Wen, and Wei Xu. Dynamic computational time for visual attention. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1199-1209, 2017. 3 +[54] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. arXiv preprint arXiv:2209.09004, 2022. 2 +[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[56] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. 2019. 5, 6 +[58] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34, 2021. 2 +[59] Wenhao Lu, Jian Jiao, and Ruofei Zhang. Twinbert: Distilling knowledge to twin-structured bert models for efficient retrieval. arXiv preprint arXiv:2002.06275, 2020. 2 + +[60] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 4 +[61] Kaleel Mahmood, Rigel Mahmood, and Marten Van Dijk. On the robustness of vision transformers to adversarial examples. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7838-7847, 2021. 3 +[62] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers. arXiv preprint arXiv:2110.03860, 2021. 2 +[63] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022. 2, 3, 4, 6 +[64] Jianhong Pan, Qichen Zheng, Zhipeng Fan, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Gradauto: Energy-oriented attack on dynamic neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 637-653. Springer, 2022. 3 +[65] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 3 +[66] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2 +[68] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 2 +[69] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[70] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks, 2016. 2 +[71] Mengye Ren, Andrei Pokrovsky, Bin Yang, and Raquel Urtasun. Sbnet: Sparse blocks network for fast inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8711-8720, 2018. 3 +[72] Aniruddha Saha, Akshayvarun Subramanya, Koninika B. Patil, and Hamed Pirsavash. Adversarial patches exploiting contextual reasoning in object detection. In ArXiv, 2019. 3 + +[73] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference, 2023. 1 +[74] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3531-3539, 2021. 2 +[75] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3 +[76] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2 +[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2 +[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2 +[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[80] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2 +[81] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the iee/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 3 +[82] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2, 3 +[83] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 3 +[84] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3 +[85] Zhenda Xie, Zheng Zhang, Xizhou Zhu, Gao Huang, and Stephen Lin. Spatially adaptive inference with stochastic feature sampling and interpolation. In Computer Vision-ECCV + +2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 531-548. Springer, 2020. 3 +[86] Le Yang, Yizeng Han, Xi Chen, Shiji Song, Jifeng Dai, and Gao Huang. Resolution adaptive networks for efficient inference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2369-2378, 2020. 3 +[87] Zerui Yang, Yuhui Xu, Wenrui Dai, and Hongkai Xiong. Dynamic-stride-net: Deep convolutional neural network with dynamic stride. In Optoelectronic Imaging and Multimedia Technology VI, pages 42-53. SPIE, 2019. 3 +[88] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2, 3, 4, 5 +[89] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819-10829, 2022. 2 +[90] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[91] Zhihang Yuan, Bingzhe Wu, Guangyu Sun, Zheng Liang, Shiwan Zhao, and Weichen Bi. S2dnas: Transforming static cnn model for dynamic inference via neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 175-192. Springer, 2020. 2 +[92] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 3 +[93] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[94] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[95] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2 +[96] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2 \ No newline at end of file diff --git a/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/images.zip b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3a0c54d8c7829cf5a83ebdf3aa9d8d7c35f418e0 --- /dev/null +++ b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0deff27756f1d4e8556852e75fb0cf396d853b295974da518c80a133ad22ffc +size 538639 diff --git a/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/layout.json b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec050799d949b68bf9775e4d1e66cc2e4edbae4a --- /dev/null +++ b/2024/SlowFormer_ Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers/layout.json @@ -0,0 +1,9567 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 531, + 138 + ], + "type": "text", + "content": "SlowFormer: Adversarial Attack on Compute and Energy Consumption of Efficient Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 142, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 142, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 142, + 175 + ], + "type": "text", + "content": "K L Navaneet *1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 161, + 318, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 161, + 318, + 175 + ], + "spans": [ + { + "bbox": [ + 155, + 161, + 318, + 175 + ], + "type": "text", + "content": "Soroush Abbasi Koohpayegani *1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 176, + 318, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 176, + 318, + 190 + ], + "spans": [ + { + "bbox": [ + 161, + 176, + 318, + 190 + ], + "type": "text", + "content": "1 University of California, Davis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 329, + 161, + 425, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 161, + 425, + 175 + ], + "spans": [ + { + "bbox": [ + 329, + 161, + 425, + 175 + ], + "type": "text", + "content": "Essam Sleiman *12†" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 331, + 175, + 433, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 175, + 433, + 190 + ], + "spans": [ + { + "bbox": [ + 331, + 175, + 433, + 190 + ], + "type": "text", + "content": "2 Harvard University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 439, + 161, + 531, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 439, + 161, + 531, + 175 + ], + "spans": [ + { + "bbox": [ + 439, + 161, + 531, + 175 + ], + "type": "text", + "content": "Hamed Piriaviash" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 217, + 192, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 230 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 230 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 243, + 290, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 290, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 290, + 480 + ], + "type": "text", + "content": "Recently, there has been a lot of progress in reducing the computation of deep models at inference time. These methods can reduce both the computational needs and power usage of deep models. Some of these approaches adaptively scale the compute based on the input instance. We show that such models can be vulnerable to a universal adversarial patch attack, where the attacker optimizes for a patch that when pasted on any image, can increase the compute and power consumption of the model. We run experiments with three different efficient vision transformer methods showing that in some cases, the attacker can increase the computation to the maximum possible level by simply pasting a patch that occupies only " + }, + { + "bbox": [ + 47, + 243, + 290, + 480 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 47, + 243, + 290, + 480 + ], + "type": "text", + "content": " of the image area. We also show that a standard adversarial training defense method can reduce some of the attack's success. We believe adaptive efficient methods will be necessary in the future to lower the power usage of expensive deep models, so we hope our paper encourages the community to study the robustness of these methods and develop better defense methods for the proposed attack. Code is available at:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 482, + 285, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 285, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 285, + 494 + ], + "type": "text", + "content": "https://github.com/UCDvision/SlowFormer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 518, + 128, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 531 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 552, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 288, + 684 + ], + "type": "text", + "content": "The field of deep learning has recently made significant progress in improving model efficiency for inference. Two broad categories of methods can be distinguished: 1) those that reduce computation regardless of input, and 2) those that reduce the computation depending on the input (adaptively). Most methods, such as weight pruning or model quantization, belong to the first category which reduces computation by a constant factor regardless of the input. However, in many applications, the complexity of the perception task may differ depending on the input. For example, when a self-driving car is driving between lanes in an empty street," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 218, + 548, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 218, + 548, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 218, + 548, + 314 + ], + "type": "text", + "content": "the perception may be simpler and require less computation when compared to driving in a busy city street scene. Interestingly, in some applications, simple scenes such as highway driving may account for the majority of the time. Therefore, we believe that adaptive computation reduction will become an increasingly important research area in the future, especially when non-adaptive methods reach the lower bound of computation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 315, + 548, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 548, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 548, + 542 + ], + "type": "text", + "content": "We argue that reduction of compute usually reduces power usage, which is crucial, particularly in mobile devices that run on battery, e.g., AR/VR headsets, humanoid robots, and drones. For instance, increasing the size of the battery for a drone may lead to a drastic reduction in its range due to the increased battery weight. This is important since the improvement in battery technology is much slower than compute technology. For instance, the battery capacity from iPhone [1] (1st generation) in 2007 to iPhone 15 Pro Max in 2023 improved from 5.18 watt-hour to 17.32 watt-hour (less than 4 times) while the compute has increased by a much larger factor. As an example, a delivery robot like Starship uses a 1,200Wh battery and can run for 12 hours [2], so it uses almost 100 watts for compute and mobility. Hence, an adversary increasing the power consumption of the perception unit by 20 watts, will reduce the battery life by almost " + }, + { + "bbox": [ + 304, + 315, + 548, + 542 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 315, + 548, + 542 + ], + "type": "text", + "content": ", which can be significant. Note that 20 watts increase in power is realistic assuming that it uses two NVIDIA Jetson Xavier NX cards (almost 20 watts each)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 544, + 548, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 548, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 548, + 652 + ], + "type": "text", + "content": "Key idea: Assuming that a perception method is reducing the computation adaptively with the input, an adversary can attack the model by modifying the input to increase the computation and power consumption. Our goal is to design a universal adversarial patch that when pasted on any input image, it will increase the computation of the model leading to increased power consumption. We believe this is an important vulnerability, particularly for safety-critical mobile systems that run on battery." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 713 + ], + "type": "text", + "content": "Please note that in this paper, we do not experiment with real hardware to measure the power consumption. Instead, we report the change in FLOPs of the inference time assuming that the power consumption is positively correlated with the number of FLOPs, as studied in [73]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 130, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 130, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 130, + 703 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 61, + 703, + 274, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 274, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 274, + 712 + ], + "type": "text", + "content": "† Work done while he was an undergraduate student at UC Davis." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24786" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 73, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 47, + 73, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 73, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 545, + 216 + ], + "type": "image", + "image_path": "08806d9dd9fae0c7d4e594a954d3ccfc9e86cf3bd11aa140d34f226b9a8cafc9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 235, + 548, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 548, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 548, + 291 + ], + "type": "text", + "content": "Figure 1. Computation and Energy Attack on Vision Transformers: Given a pre-trained input-dependent computation efficient model, the adversary first attaches an adversarial patch to all images in a dataset and optimizes this patch with our method such that it maximizes the model's computation for each sample. During inference, the adversary modifies the input of the victim's model by applying the learnt patch to it. This results in an increase in compute in the victim's model. The attack will thus potentially slowdown and also lead to increased energy consumption and CPU/GPU usage on the victim's device." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 443 + ], + "type": "text", + "content": "We design our attack, SlowFormer, for three different methods (A-ViT [88], ATS [25], and Ada-VIT [63]) that reduce the computation of vision transformers. These methods generally identify the importance of each token for the final task and drop the insignificant ones to reduce the computation. We show that in all three cases, our attack can increase the computation by a large margin, returning it to the full-compute level (non-efficient baseline) for all images in some settings. Our threat model is agnostic to the accuracy of the model and attacks the computation and power consumption only. Figure 1 shows our attack." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 444, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 289, + 539 + ], + "type": "text", + "content": "There are some prior works that design a pixel-level perturbation attack to increase the compute of the model. We believe universal patch-based attacks that do not change with the input image (generalize from training data to test data) are much more practical in real applications. Note that to modify the pixel values on a real robot, the attacker needs to access and manipulate the image between the camera and compute modules, which is impossible in many applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 539, + 288, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 611 + ], + "type": "text", + "content": "Contributions: We show that efficient vision transformer methods are vulnerable to a universal patch attack that can increase their compute and power usage. We demonstrate this through experiments on three different efficient transformer methods. We show that an adversarial training defense can reduce attack success to some extent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 621, + 134, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 134, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 134, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Vision Transformers: The popularity of transformers [79] in vision has grown rapidly since the introduction of the first vision transformer [20, 78]. Recent works demonstrate the strength of vision transformers on a variety of computer vision tasks [11, 18, 21, 56, 69, 77, 90, 93, 94, 96]. Moreover, transformers are the backbone of recent Self-Supervised" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 312, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 547, + 348 + ], + "type": "text", + "content": "Learning (SSL) models [12, 37], and vision-language models [67]. In our work, we design an attack to target the computation and energy efficiency of vision transformers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 351, + 548, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 548, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 548, + 518 + ], + "type": "text", + "content": "Efficient Vision Transformers: Due to the recent importance and popularity of vision transformers, many works have started to study the efficiency of vision transformers [7, 46, 89]. To accomplish this, some lines of work study token pruning with the goal of removing uninformative tokens in each layer [25, 62, 63, 68, 88]. ToMe [6] merges similar tokens in each layer to decrease the computation. Some works address quadratic computation of self-attention module by introducing linear attention [3, 45, 48, 58, 74]. Efficient architectures [40, 55] that limit the attention span of each token have been proposed to improve efficiency. In our paper, we attack token pruning based efficient transformers where the computation varies based on the input samples [25, 63, 88]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "Dynamic Computation: There are different approaches to reducing the computation of vision models, including knowledge distillation to lighter network [39, 59], model quantization [54, 70] and model pruning [51]. In these methods, the computation is fixed during inference. In contrast to the above models, some works address efficiency by having variable computation based on the input. The intuition behind this direction is that not all samples require the same amount of computation. Several recent works have developed models that dynamically exit early or skip layers [5, 22, 28, 33, 34, 44, 76, 80, 82] and selectively activate neurons, channels or branches for dynamic width [4, 9, 13, 26, 31, 38, 43, 91] depending on the complexity of the input sample. Zhou et al. show that not all locations in an image contribute equally to the predictions of a CNN model [95], encouraging a new line of work to make CNNs more" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24787" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "efficient through spatially dynamic computation. Pixel-Wise dynamic architectures [10, 14, 23, 47, 71, 81, 85] learn to focus on the significant pixels for the required task while Region-Level dynamic architectures perform adaptive inference on the regions or patches of the input [29, 53]. Finally, lowering the resolution of inputs decreases computation, but at the cost of performance. Conventional CNNs process all regions of an image equally, however, this can be inefficient if some regions are \"easier\" to process than others [42]. Correspondingly, [86, 87] develop methods to adaptively scale the resolution of images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 338 + ], + "type": "text", + "content": "Transformers have recently become extremely popular for vision tasks, resulting in the release of a few input-dynamic transformer architectures [24, 63, 88]. Fayyaz et al. [24] introduce a differentiable parameter-free Adaptive Token Sampler (ATS) module which scores and adaptively samples significant tokens. ATS can be plugged into any existing vision transformer architecture. A-ViT [88] reduces the number of tokens in vision transformers by discarding redundant spatial tokens. Meng et al. [63] propose AdaViT, which trains a decision network to dynamically choose which patch, head, and block to keep/activate throughout the backbone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 472 + ], + "type": "text", + "content": "Adversarial Attack: Adversarial attacks are designed to fool models by applying a targeted perturbation or patch on an image sample during inference [32, 50, 75]. These methods can be incorporated into the training set and optimized to fool the model. Correspondingly, defenses have been proposed to mitigate the effects of these attacks [27, 52, 65, 84]. Patch-Fool [30] considers adversarial patch-based attacks on transformers. Some recent works [61, 83, 92] also study and design methods for the transferability of adversarial attacks on vision transformers. However, most prior adversarial attacks target model accuracy, ignoring model efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": "Energy Attack: Very recently, there have been a few works on energy adversarial attacks on neural networks. In ILFO [35], Haque et al. attack two CNN-based input-dynamic methods: SkipNet [82] and SACT [28] using image specific perturbation. DeepSloth [41] attack focuses on slowing down early-exit methods, reducing their energy efficiency by " + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "inline_equation", + "content": "90 - 100\\%" + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": ". GradAuto [64] successfully attacks methods that are both dynamic width and dynamic depth. NICGSlowDown and TransSlowDown [16, 17] attack neural image caption generation and neural machine translation methods, respectively. All these methods primarily employ image specific perturbation based adversarial attack. Sloth-Bomb injects efficiency backdoors to input-adaptive dynamic neural networks [15] and NodeAttack [36] attacks Neural Ordinary Differential Equation models, which use ordinary differential equation solving to dynamically predict the output of a neural network. Our work is closely related to ILFO [35], DeepSloth [41] and GradAuto [64] in that we attack the computational efficiency of networks. However, unlike these methods, we focus on designing an adversarial patch-based" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "attack that is universal and on vision transformers. We additionally provide a potential defense for our attack. We use a patch that generalizes from train to test set and thus we do not optimize per sample during inference. Our patch-based attack is especially suited for transformer architectures [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 144, + 489, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 489, + 158 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 489, + 158 + ], + "type": "text", + "content": "3. Computation and Energy Attack" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 164, + 397, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 164, + 397, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 397, + 175 + ], + "type": "text", + "content": "3.1. Threat Model:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 182, + 547, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 302 + ], + "type": "text", + "content": "We consider a scenario where the adversary has access to the victim's trained deep model and modifies its input such that the energy consumption and computational demand of the model is increased. The attack is agnostic to model accuracy. To make the setting more practical, instead of perturbing the entire image, we assume that the adversary can modify the input image by only pasting a patch [8, 72] on it and that the patch is universal, that is, image independent. During inference, a pretrained patch is pasted on the test image before propagating it through the network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 303, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 547, + 399 + ], + "type": "text", + "content": "In this paper, we attack three state-of-the-art efficient transformers. Since the attacker manipulates only the input image and not the network parameters, the attacked model must have dynamic computation that depends on the input image. As stated earlier, several recent works have developed such adaptive efficient models and we believe that they will be more popular in the future due to the limits of non-adaptive efficiency improvement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 407, + 519, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 519, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 519, + 418 + ], + "type": "text", + "content": "3.2. Attack on Efficient Vision Transformers:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 653 + ], + "type": "text", + "content": "Universal Adversarial Patch: We use an adversarial patch to attack the computational efficiency of transforms. The learned patch is universal, that is, a single patch is trained and is used during inference on all test images. The patch generalizes across images but not across models. The patch optimization is performed only on the train set. The patch is pasted on an image by replacing the image pixels using the patch. We assume the patch location does not change from train to test. The patch pixels are initialized using i.i.d. samples from a uniform distribution over [0, 255]. During each training iteration, the patch is pasted on the mini-batch samples and is updated to increase the computation of the attacked network. The patch values are projected onto [0, 255] and quantized to 256 uniform levels after each iteration. Note that we use a pretrained network and do not update its parameters either in the training or in the evaluation of our attack. During inference, the trained patch is pasted on the test images and the computational efficiency of the network on the adversarial image is measured." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Here, we focus on three methods employing vision transformers for the task of image classification. All these methods modify the computational flow of the network based on the input image for faster inference. A pretrained model is used for the attack and is not modified during" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24788" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 97 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 97 + ], + "type": "text", + "content": "our adversarial patch training. We first provide a brief background of each method before describing our attack." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 127, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 127, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 127, + 119 + ], + "type": "text", + "content": "Attacking A-ViT :" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": "Background: A-ViT [88] adaptively prunes image tokens to achieve speed-up in inference with minimal loss in accuracy. For a given image, a dropped token will not be used again in the succeeding layers of the network. Let " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " be the input image and " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\{t^l\\}_{1:K}" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " be the corresponding " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " tokens at layer " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": ". An input-dependent halting score " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "h_k^l" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " for a token " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " is calculated and the token is dropped at layer " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "N_k" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " where its cumulative halting score exceeds a fixed threshold value " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": " for the first time. The token is propagated until the final layer if its score never exceeds the threshold. Instead of introducing a new parameter for " + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "inline_equation", + "content": "h_k^l" + }, + { + "bbox": [ + 46, + 121, + 289, + 312 + ], + "type": "text", + "content": ", the first dimension of each token is used to predict the halting score for the corresponding token. The network is trained to maximize the cumulative halting score at each layer and thus drop the tokens earlier. The loss, termed ponder loss, is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 323, + 289, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 289, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p o n d e r}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(N _ {k} + r _ {k}\\right), \\quad r _ {k} = 1 - \\sum_ {l = 1} ^ {N _ {k - 1}} h _ {k} ^ {l} \\tag {1}", + "image_path": "2d94ccfb00957d9cbaa9adae87ca2664ca4eb1e32a53e33fdeb0e6c9f6031fb9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "content": "Additionally, A-ViT enforces a Gaussian prior on the expected halting scores of all tokens via " + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "content": "-divergence based distribution loss, " + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{distr}}" + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "content": ". These loss terms are minimized along with the task-specific loss " + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "content": ". Thus, the overall training objective is " + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{task}} + \\alpha_d\\mathcal{L}_{\\mathrm{distr}} + \\alpha_p\\mathcal{L}_{\\mathrm{ponder}}" + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\alpha_{d}" + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\alpha_{p}" + }, + { + "bbox": [ + 46, + 369, + 289, + 441 + ], + "type": "text", + "content": " are hyperparameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "text", + "content": "Attack: Here, we train the patch to increase the inference compute of a trained A-ViT model. Since we are interested in the compute and not task-specific performance, we simply use " + }, + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "inline_equation", + "content": "-\\left(\\alpha_{d}\\mathcal{L}_{\\mathrm{distr.}} + \\alpha_{p}\\mathcal{L}_{\\mathrm{ponder}}\\right)" + }, + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "text", + "content": " as our loss. It is possible to preserve (or hurt) the task performance by additionally using " + }, + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "inline_equation", + "content": "+\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "inline_equation", + "content": "-\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 46, + 443, + 289, + 515 + ], + "type": "text", + "content": ") in the loss formulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 526, + 132, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 132, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 132, + 538 + ], + "type": "text", + "content": "Attacking AdaViT:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "content": "**Background:** To improve the inference efficiency of vision transformers, AdaViT [63] inserts and trains a decision network before each transformer block to dynamically decide which patches, self-attention heads, and transformer blocks to keep/activate throughout the backbone. The " + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "inline_equation", + "content": "l^{\\text{th}}" + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "content": " block's decision network consists of three linear layers with parameters " + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "inline_equation", + "content": "W_{l} = W_{l}^{p}, W_{l}^{h}, W_{l}^{b}" + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "content": " which are then multiplied by each block's input " + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "inline_equation", + "content": "Z_{l}" + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 643, + 288, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 643, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 94, + 643, + 288, + 658 + ], + "type": "interline_equation", + "content": "\\left(m _ {l} ^ {p}, m _ {l} ^ {h}, m _ {l} ^ {b}\\right) = \\left(W _ {l} ^ {p}, W _ {l} ^ {h}, W _ {l} ^ {b}\\right) Z _ {l} \\tag {2}", + "image_path": "eba2494250d2c7a78d1005a0d3ae8fa1d721974c6602544bb7d7c158b05c1b9a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "The value " + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": " is then passed to sigmoid function to convert it to a probability value used to make the binary decision of keep/discard. Gumbel-Softmax trick [60] is used to make this decision differentiable during training. Let " + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": " be the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 97 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 97 + ], + "type": "text", + "content": "keep/discard mask after applying Gumbel-Softmax on " + }, + { + "bbox": [ + 304, + 72, + 547, + 97 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 72, + 547, + 97 + ], + "type": "text", + "content": ". The loss on computation is given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 102, + 547, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 102, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 306, + 102, + 547, + 173 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {u s a g e} = (\\frac {1}{D _ {p}} \\sum_ {d = 1} ^ {D _ {p}} M _ {d} ^ {p} - \\gamma_ {p}) ^ {2} + (\\frac {1}{D _ {h}} \\sum_ {d = 1} ^ {D _ {h}} M _ {d} ^ {h} - \\gamma_ {h}) ^ {2} \\\\ + \\left(\\frac {1}{D _ {b}} \\sum_ {d = 1} ^ {D _ {b}} M _ {d} ^ {b} - \\gamma_ {b}\\right) ^ {2} \\tag {3} \\\\ \\end{array}", + "image_path": "2244df5b5fb7d21c3ff240b5e4a611ccd42b9810f9a3b20bfa40ef44ce878264.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "inline_equation", + "content": "D_{p}, D_{h}, D_{b}" + }, + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "text", + "content": " represent the number of total patches, heads, and blocks of the entire transformer, respectively. " + }, + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\gamma_{p}, \\gamma_{h}, \\gamma_{b}" + }, + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "text", + "content": " denote the target computation budgets i.e. the percentage of patches/heads/blocks to keep. The total loss is a combination of task loss (cross-entropy) and computation loss: " + }, + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{ce} + \\mathcal{L}_{usage}" + }, + { + "bbox": [ + 304, + 178, + 547, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "content": "Attack: To attack this model, we train the patch to maximize the computation loss " + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{usage}" + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "content": ". More specifically, we set the computation-target " + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "content": " values to 0 and negate the " + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{usage}" + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "content": " term in Eq. 3. As a result, the patch is optimized to maximize the probability of keeping the corresponding patch (p), attention head (h), and transformer block (b). We can also choose to attack the prediction performance by selectively including or excluding the " + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "content": " term." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 357, + 375, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 375, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 375, + 369 + ], + "type": "text", + "content": "Attacking ATS:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "text", + "content": "Background: Given " + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "text", + "content": " tokens with the first one as the classification token, the transformer attention matrix " + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 369, + 545, + 405 + ], + "type": "text", + "content": " is calculated by the following dot product:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\mathrm{Softmax}\\left(\\mathcal{Q}\\mathcal{K}^T /\\sqrt{d}\\right)" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\sqrt{d}" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " is a scaling coefficient, " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " is the dimension of tokens, " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{Q},\\mathcal{K}" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " are the query, key and value matrices, respectively. The value " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{1,j}" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " denotes the attention of the classification token to token " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": ". ATS [25] assigns importance score " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " for each token " + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 406, + 547, + 482 + ], + "type": "text", + "content": " by measuring how much the classification token attends to it:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 487, + 547, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 487, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 372, + 487, + 547, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {j} = \\frac {\\mathcal {A} _ {1 , j} \\times \\left\\| \\mathcal {V} _ {j} \\right\\|}{\\sum_ {i = 2} \\mathcal {A} _ {1 , i} \\times \\left\\| \\mathcal {V} _ {i} \\right\\|} \\tag {4}", + "image_path": "b1685543a17fbd4f370115969175dccd089d9e7237f11e61accef25ecc829767.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": "The importance scores are converted to probabilities and are used to sample tokens, where tokens with a lower score have more of a chance of being dropped." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 561, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 654 + ], + "type": "text", + "content": "Attack: Since ATS uses inverse transform sampling, it results in fewer samples if the importance distribution is sharp. To maximize the computation in ATS, we aim to obtain a distribution of scores with high entropy to maximize the number of retained tokens. Therefore, we optimize the patch so that the attention of the classification token over other tokens is a uniform distribution using the following MSE loss:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 377, + 654, + 545, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 654, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 377, + 654, + 545, + 686 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 2} ^ {N} \\left| \\left| \\mathcal {A} _ {1, i} - \\frac {1}{N} \\right| \\right| _ {2} ^ {2} \\tag {5}", + "image_path": "28ced09d577a77374f5049bce5ca5bf1bab49411dcd231089378e94ea4d70094.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Note that one can optimize " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " to be uniform, but we found the above loss to be easier to optimize. For a multi-head" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24789" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "attention layer, we calculate the loss for each head and then sum the loss over all heads. Moreover, ATS can be applied to any layer of a vision transformer. For a given model, we apply our loss at all ATS layers and use a weighted summation for optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 146, + 102, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 146, + 102, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 102, + 158 + ], + "type": "text", + "content": "4. Defense" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 167, + 288, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 239 + ], + "type": "text", + "content": "An obvious defense, although weak, will be to use nondynamic efficient methods only, e.g., weight pruning, where the reduction in compute is deterministic and does not depend on the input. However, most such methods do not achieve high levels of computation efficiency since they do not take advantage of the simplicity of images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "type": "text", + "content": "We adopt standard adversarial training as a better defense method for our attack. In the standard way, at each iteration of training the model, one would load an image, attack it, and then use it with correct labels in training the model. We cannot adopt this out-of-the-box since our attack generalizes across images and is not dependent on a single image only. To do this, we maintain a set of adversarial patches, and at each iteration sample one of them randomly (uniformly), and use it at the input while optimizing the original loss of the efficient model to train a robust model. To adapt the set of adversarial patches to the model being trained, we interrupt the training at every " + }, + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "type": "text", + "content": " mark of each epoch and optimize for a new patch to be added to the set of patches. To limit the computational cost of training, we use only 500 iterations to optimize for a new patch, which results in an attack with reasonable accuracy compared to our main results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 444, + 128, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 128, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 128, + 458 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 464, + 257, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 257, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 257, + 477 + ], + "type": "text", + "content": "5.1. Attack on Efficient Vision Transformers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 288, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 556 + ], + "type": "text", + "content": "Dataset: We evaluate the effectiveness of our attack on two datasets: ImageNet-1K [19] and CIFAR-10 [49]. ImageNet-1K contains 1.3M images in the train set and 50K images in the validation set with 1000 total categories. CIFAR-10 has 50K images for training and 10K images for validation with 10 total categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "text", + "content": "Metrics: We report Top-1 accuracy and average computation in terms of GFLOPs for both attacked and unattacked models. Similar to Attack Success Rate in a standard adversarial attack, we introduce a metric: Attack Success to quantify the efficacy of the attack. We define Attack Success as the number of FLOPs increased by the attack divided by the number of FLOPs decreased by the efficient method. Attack Success = " + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\frac{(\\mathrm{FLOPs}_{\\mathrm{attack}} - \\mathrm{FLOPs}_{\\mathrm{min}})}{(\\mathrm{FLOPs}_{\\mathrm{max}} - \\mathrm{FLOPs}_{\\mathrm{min}})}" + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FLOPs}_{\\mathrm{min}}" + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "text", + "content": " is the compute of the efficient model and " + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FLOPs}_{\\mathrm{max}}" + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "text", + "content": " is that of the original inefficient model. Attack Success is thus capped at " + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 568, + 289, + 713 + ], + "type": "text", + "content": " while a negative value denotes a reduction in FLOPs. Note that our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "content": "Attack Success metric illustrates the effectiveness of an attack in reversing the FLOPs reduction of a particular method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 109, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 547, + 133 + ], + "type": "text", + "content": "Baselines: We propose three alternative approaches to Slow-Former (ours) to generate the patch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 134, + 545, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 134, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 305, + 134, + 545, + 170 + ], + "type": "text", + "content": "Random Patch: A simple baseline is to generate a randomly initialized patch. We sample IID pixel values from a uniform distribution between 0 and 255 to create the patch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 171, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 547, + 255 + ], + "type": "text", + "content": "NTAP: We consider a standard adversarial patch that is trained to attack the model task performance instead of compute. We use a non-targeted universal adversarial patch (NTAP) to attack the model. We train the patch to fool the model by misclassifying the image it is pasted on. We use the negative of the cross-entropy loss with the predicted and ground-truth labels as the loss to optimize the patch." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 255, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 547, + 327 + ], + "type": "text", + "content": "TAP: We train a universal targeted adversarial patch (TAP). The patch is optimized to classify all images in the train set to a single fixed category. Similar to NTAP, the adversarial attack here is on task performance and not computation. We experiment with ten randomly generated target category labels and report the averaged metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "text", + "content": "Implementation Details: We use PyTorch [66] for all experiments. Unless specified, we use a patch of size " + }, + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "text", + "content": ", train and test on " + }, + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "text", + "content": " images, and we paste the patch on the top-left corner. Note that our patch occupies just " + }, + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 305, + 340, + 547, + 436 + ], + "type": "text", + "content": " of the total area of an input image. We use AdamW [57] optimizer to optimize the patches and use 4 NVIDIA RTX 3090 GPUs for each experiment. We use varying batch sizes and learning rates for each of the computation-efficient methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 437, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 547, + 581 + ], + "type": "text", + "content": "ATS Details: As in ATS [25], we replace layers 3 through 9 of ViT networks with the ATS block and set the maximum limit for the number of tokens sampled to 197 for each layer. We train the patch for 2 epochs with a learning rate of 0.4 for ViT-Tiny and " + }, + { + "bbox": [ + 305, + 437, + 547, + 581 + ], + "type": "inline_equation", + "content": "lr = 0.2" + }, + { + "bbox": [ + 305, + 437, + 547, + 581 + ], + "type": "text", + "content": " for ViT-Base and ViT-Small. We use a batch size of 1024 and different loss coefficients for each layer of ATS. For DeiT-Tiny we use [1.0, 0.2, 0.2, 0.2, 0.01, 0.01, 0.01], for DeiT-Small we use [1.0, 0.2, 0.05, 0.01, 0.005, 0.005, 0.005], and for DeiT-Base we use [2.0, 0.1, 0.02, 0.01, 0.005, 0.005, 0.005]. The weights are vastly different at initial and final layers to account for the difference in loss magnitudes across layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "A-ViT Details: When attacking A-VIT[88], the patches are optimized for one epoch with a learning rate of 0.2 and a batch size of 512 (128 × 4 GPUs) using AdamW [57] optimizer. We optimize the patches for 4 epochs for patch length 32 and below. For CIFAR-10 experiments, the images are resized from " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " crop is used as the input. For the training of adversarial defense, we generate 5 patches per epoch of adversarial training and limit the number of iterations for patch generation to 500. The learning rate for patch optimization is increased to 0.8 for faster convergence." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24790" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 178, + 310, + 650 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 324, + 170 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 324, + 170 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 324, + 170 + ], + "type": "text", + "content": "Table 1. Computation and Energy Attack on Efficient Vision Transformers: Comparison of the effect of our attack with baselines: No Attack, Random Patch, targeted (TAP), and non-targeted (NTAP) adversarial patches applied to three input-dynamic computation efficient pre-trained models of varying architectures. The maximum possible compute for a given architecture is provided in bold. On A-ViT, we completely undo the efficiency gains obtained by the efficient method through our attack, achieving Attack Success of " + }, + { + "bbox": [ + 46, + 70, + 324, + 170 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 70, + 324, + 170 + ], + "type": "text", + "content": ". We achieve high Attack Success on all approaches while the baselines expectedly do not contribute to increase in compute." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 178, + 310, + 650 + ], + "lines": [ + { + "bbox": [ + 61, + 178, + 310, + 650 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 310, + 650 + ], + "type": "table", + "html": "
MethodAttackModel GFLOPsTop-1 AccAttack Success
A-ViTViT-Tiny1.3--
No attack0.8771.4%-
Random Patch0.8770.8%-1%
TAP0.850.1%-5%
NTAP0.830.1%-10%
SlowFormer (ours)1.34.7%100%
A-ViTViT-Small4.6--
No attack3.778.8%-
Random Patch3.778.4%-2%
TAP3.60.1%-12%
NTAP3.60.1%-7%
SlowFormer (ours)4.62.3%99%
ATSViT-Tiny1.3--
No attack0.8470.3%-
Random Patch0.8369.8%-2%
TAP0.760.1%-17%
NTAP0.610.1%-50%
SlowFormer (ours)1.01.2%35%
ATSViT-Small4.6--
No attack3.179.2%-
Random Patch3.178.6%-1%
TAP3.00.1%-7%
NTAP2.40.1%-47%
SlowFormer (ours)4.01.0%60%
ATSViT-Base17.6--
No attack12.681.3%-
Random Patch12.581.2%-2%
TAP12.00.1%-12%
NTAP11.00.1%-32%
SlowFormer (ours)15.40.2%52%
AdaViTViT-Small4.6--
No attack2.2577.3%-
Random Patch2.2076.9%-2%
TAP2.280.1%1%
NTAP2.150.1%-4%
SlowFormer (ours)3.20.4%40%
", + "image_path": "5dc90aac70e089a48fda267e378abc93bc5f77268d1bdd84b541dbd072177e79.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 665, + 288, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 703 + ], + "type": "text", + "content": "AdaViT Details: For AdaViT[63], we first freeze the weights and use a learning rate of 0.2 and a batch size of 128 with 4 GPUs for patch optimization. We use AdamW [57]" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 334, + 178, + 559, + 651 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 564, + 171 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 564, + 171 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 564, + 171 + ], + "type": "text", + "content": "Figure 2. Visualization of our Energy Attack on Vision Transformers: We visualize the A-ViT-Small with and without our attack. We use patch size of 32 for the attack (on the top-left corner). We show pruned tokens at layer 8 of A-ViT-Small. Our attack can recover most of the pruned tokens, resulting in increased computation and power consumption. Note that although the patch is reasonably small and is in the corner of the view, it can affect the whole computational flow of the network. This is probably due to the global attention mechanism in transformers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 178, + 559, + 651 + ], + "lines": [ + { + "bbox": [ + 334, + 178, + 559, + 651 + ], + "spans": [ + { + "bbox": [ + 334, + 178, + 559, + 651 + ], + "type": "table", + "html": "
AttackedEfficient ModelAttackedEfficient Model
", + "image_path": "0424ccce6be847edeaa91ff449a53701345e2f373aa6249232a05fc5e04a35c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 666, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 701 + ], + "type": "text", + "content": "optimizer with no decay and train for 2 epochs with a patch size of " + }, + { + "bbox": [ + 305, + 666, + 545, + 701 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 666, + 545, + 701 + ], + "type": "text", + "content": ". We train on the ImageNet-1k train dataset and evaluate it on the test set." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24791" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 145, + 279, + 248 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 299, + 137 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 299, + 137 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 299, + 137 + ], + "type": "text", + "content": "Table 2. Results on CIFAR10 dataset. We report results on CIFAR10 dataset to show that our attack is not specific to ImageNet alone. CIFAR-10 is a small dataset compared to ImageNet and thus results in an extremely efficient A-ViT model. Our attack increases the FLOPs from 0.11 to 0.58 which restores nearly " + }, + { + "bbox": [ + 46, + 70, + 299, + 137 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 46, + 70, + 299, + 137 + ], + "type": "text", + "content": " of the original reduction in the FLOPs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 145, + 279, + 248 + ], + "lines": [ + { + "bbox": [ + 67, + 145, + 279, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 145, + 279, + 248 + ], + "type": "table", + "html": "
MethodModel FLOPsTop-1 AccAttack Success
ViT-Tiny1.2695.9%-
A-ViT-Tiny0.1195.8%-
SlowFormer (ours)0.5860.2%41%
ATS-Tiny0.8594.7%-
SlowFormer (ours)0.9924.7%34.1%
", + "image_path": "6de61f8eae0584938f9a64c90199c3ffbd14dfc0f00b6d600c5bcee0e0cf0b98.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 258, + 105, + 327 + ], + "blocks": [ + { + "bbox": [ + 49, + 258, + 105, + 327 + ], + "lines": [ + { + "bbox": [ + 49, + 258, + 105, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 258, + 105, + 327 + ], + "type": "image", + "image_path": "84741892e84b8e36c7ca7cd0580b06f8d4308cb1212a4b01896dc10fe3f7e223.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 337, + 288, + 360 + ], + "lines": [ + { + "bbox": [ + 47, + 337, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 288, + 360 + ], + "type": "text", + "content": "Figure 3. Visualization of optimized patch: We show the learned universal patches for each of the three efficient methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 258, + 196, + 327 + ], + "blocks": [ + { + "bbox": [ + 140, + 258, + 196, + 327 + ], + "lines": [ + { + "bbox": [ + 140, + 258, + 196, + 327 + ], + "spans": [ + { + "bbox": [ + 140, + 258, + 196, + 327 + ], + "type": "image", + "image_path": "cff71724dc6dfc20de777051a6fffde590fd94fd51cf3f04e6e3dd981910b7f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 230, + 258, + 286, + 327 + ], + "blocks": [ + { + "bbox": [ + 230, + 258, + 286, + 327 + ], + "lines": [ + { + "bbox": [ + 230, + 258, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 230, + 258, + 286, + 327 + ], + "type": "image", + "image_path": "e43cce4c5073a7ea8d1a9a2865440e96238cfed9feee7e47454ac716ec047a96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 370, + 288, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 288, + 448 + ], + "type": "text", + "content": "Table 4. Effect of patch size: Analysis of the effect of adversarial patch size on the attack success rate on A-ViT. Our attach is reasonably successful even using a small patch size " + }, + { + "bbox": [ + 46, + 370, + 288, + 448 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 46, + 370, + 288, + 448 + ], + "type": "text", + "content": ", which is only " + }, + { + "bbox": [ + 46, + 370, + 288, + 448 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 370, + 288, + 448 + ], + "type": "text", + "content": " of the image area. Interestingly, a small patch on the corner of the view affects the computational flow of the entire transformer model. This might be due to the global attention mechanism in transformers." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 456, + 255, + 654 + ], + "blocks": [ + { + "bbox": [ + 48, + 456, + 255, + 654 + ], + "lines": [ + { + "bbox": [ + 48, + 456, + 255, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 255, + 654 + ], + "type": "table", + "html": "
Patch Size (Area)Model GFLOPsTop-1 AccuracyAttack Success
ViT-Tiny1.26--
A-ViT-Tiny0.8771.4%-
64 (8%)1.264.7%100%
48 (5%)1.261.8%99%
32 (2%)1.2217.4%90%
16 (0.5%)0.9863.3%27%
ViT-Small4.6--
A-ViT-Small3.778.8%-
64 (8%)4.62.3%99%
48 (5%)4.65.1%98%
32 (2%)4.439.5%78%
16 (0.5%)3.878.2%16%
", + "image_path": "5a02167d8537d1dcadd52e190d6c719b3c4e6cbdd1323f64237e3536a305da35.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Results. The results of our attack, SlowFormer , on various methods on ImageNet dataset are shown in table 1. In A-ViT, we successfully recover " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " of the computation reduced" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 335, + 145, + 531, + 248 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 559, + 137 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 559, + 137 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 559, + 137 + ], + "type": "text", + "content": "Table 3. Accuracy controlled compute adversarial attack: We attack the efficiency of A-ViT while either maintaining or destroying its classification performance. We observe that our attack can achieve a huge variation in task performance without affecting the Attack Success. The ability to attack the computation without affecting the task performance might be crucial in some applications." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 145, + 531, + 248 + ], + "lines": [ + { + "bbox": [ + 335, + 145, + 531, + 248 + ], + "spans": [ + { + "bbox": [ + 335, + 145, + 531, + 248 + ], + "type": "table", + "html": "
AttackModel GFLOPsAttack SuccessTop-1 Acc
ViT-Tiny1.26--
No attack0.87-71.4%
Acc agnostic1.26100%4.7%
Preserve acc1.2392%68.5%
Destroy acc1.26100%0.1%
", + "image_path": "baedee61a8c1b91141487e2f457a4c368d873f1cbe5b46a6b64c01afda3f9bba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "type": "text", + "content": "The efficient methods are also susceptible universal perturbation based attacks. We use an " + }, + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "type": "text", + "content": " bound on the perturbation." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 300, + 556, + 523 + ], + "blocks": [ + { + "bbox": [ + 305, + 258, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 270 + ], + "type": "text", + "content": "Table 5. Attack with adversarial perturbation on ImageNet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 300, + 556, + 523 + ], + "lines": [ + { + "bbox": [ + 308, + 300, + 556, + 523 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 556, + 523 + ], + "type": "table", + "html": "
MethodEpsilon (/255.)AttackModel GFLOPsTop-1 AccuracyAttack Success
A-ViTViT-Tiny1.3--
-No attack0.8771.4%-
16SlowFormer1.156.1%73%
32SlowFormer1.250.5%98.4%
ATS-No attack0.8470.3%-
16SlowFormer0.9815.6%30.4%
32SlowFormer1.040.8%43.5%
A-ViTViT-Small4.6--
-No attack3.778.8%-
16SlowFormer4.4820%86.4%
32SlowFormer4.591%98.1%
ATS-No attack3.179.2%-
16SlowFormer3.631.0%33.3%
32SlowFormer3.83.6%46.7%
AdaVit-No attack2.2577.3%-
16SlowFormer3.026.1%31.9%
32SlowFormer3.22.8%40.4%
", + "image_path": "1290446887e70b76f6066558ada7a194eef6eebfaea57e87541a3f011a8e6156.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "by A-ViT. Our attack has an Attack Success of " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " on ATS and " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " on AdaViT with ViT-Small. A random patch attack has little effect on both the accuracy and computation of the method. Both standard adversarial attack baselines, TAP and NTAP, reduce the accuracy to nearly " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": ". Interestingly, these patches further decrease the computation of the efficient model being attacked. This might be because of the increased importance of adversarial patch tokens to the task and thus reduced importance of other tokens. Targeted patch (TAP) has a significant reduction in FLOPs on the ATS method. Since the token dropping in ATS relies on the distribution of attention values of classification tokens, a sharper distribution due to the increased importance of a token can result in a reduction in computation. The" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24792" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "computation increase with SlowFormer for AdaViT is comparatively low. To investigate, we ran a further experiment using a patch size of " + }, + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": " (entire image size) to find the maximum possible computation for an image. This resulted in 4.18 GFLOPs on the ImageNet-1K validation set, which is markedly lower than the limit of 4.6. Using this as an upper-bound of GFLOPs increase, SlowFormer achieves a " + }, + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": " Attack Success." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 275 + ], + "type": "text", + "content": "We report the results on CIFAR-10 dataset in Table 2. The efficient model (A-ViT) drastically reduces the computation from 1.26 GFLOPs to 0.11 GFLOPs. Most of the tokens are dropped as early as layer two in the efficient model. SlowFormer is able to effectively attack even in such extreme scenarios, achieving an Attack Success of " + }, + { + "bbox": [ + 46, + 168, + 288, + 275 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 168, + 288, + 275 + ], + "type": "text", + "content": " and increasing the mean depth of tokens from nearly one to five. SlowFormer is similarly effective on ATS with an Attack Success of " + }, + { + "bbox": [ + 46, + 168, + 288, + 275 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 46, + 168, + 288, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 276, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 360 + ], + "type": "text", + "content": "We additionally visualize the effectiveness of our attack in Figure 2. The un-attacked efficient method retains only highly relevant tokens at the latter layers of the network. However, our attack results in nearly the entire image being passed through all layers of the model for all inputs. In Fig. 3, we visualize the optimized patches for each of the three efficient methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 118, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 118, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 118, + 380 + ], + "type": "text", + "content": "5.2. Ablations:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 413 + ], + "type": "text", + "content": "We perform all ablations on the A-ViT approach using their pretrained ViT-Tiny architecture model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 555 + ], + "type": "text", + "content": "Accuracy controlled compute adversarial attack: As seen in Table 1, our attack can not only increase the computation, but also reduce the model accuracy. This can be desirable or hurtful based on the attacker's goals. A low-accuracy model might be an added benefit, similar to regular adversaries, but might also lead to the victim detecting the attack. We show that it is possible to attack the computation of the model while either preserving or destroying the task performance by additionally employing a task loss in the patch optimization. Table 3 indicates that the accuracy can be significantly modified while maintaining a high Attack Success." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": "Effect of patch size: We vary the patch size from " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " (just a single token) and report the results in Table 4. Interestingly, our attack with ViT-Small has a " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " Attack Success with a " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " patch size, which occupies only " + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 616 + ], + "type": "text", + "content": " of the input image area." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 616, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 676 + ], + "type": "text", + "content": "Effect of patch location: We vary the location of the patch to study its effect on the Attack Success of the model. We randomly sample a location in the image for where we paste the patch on. We perform five such experiments and observe an Attack Success of " + }, + { + "bbox": [ + 46, + 616, + 287, + 676 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 616, + 287, + 676 + ], + "type": "text", + "content": " for all patch locations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Perturbation attack: While we focus on patch based attacks in this paper, efficient transformers are also susceptible to perturbation based attacks (table 5). In perturbation attacks," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 145, + 526, + 232 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 137 + ], + "type": "text", + "content": "Table 6. Defense using adversarial training: We propose and show the impact of our defense for our adversarial attack on A-ViT. Our defense is simply maintaining a set of universal patches and training the model to be robust to a random sample of those at each iteration. The defense reduces the computation to some extent (1.26 to 1.01), but is still far from the unattacked model (0.87)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 145, + 526, + 232 + ], + "lines": [ + { + "bbox": [ + 326, + 145, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 326, + 145, + 526, + 232 + ], + "type": "table", + "html": "
MethodGFLOPsTop-1 Acc.Attack Success
No attack0.8771.4-
SlowFormer1.264.7%100%
Adv Defense + SlowFormer1.0165.8%34%
", + "image_path": "a61897e858efd9e27d789e81775c6b4f4ed3c6b95222cd0ca748451e5000e5c9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 281 + ], + "type": "text", + "content": "all pixels in the image can be modified, but with an upper bound on the " + }, + { + "bbox": [ + 304, + 257, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 304, + 257, + 545, + 281 + ], + "type": "text", + "content": " norm of the perturbation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 297, + 492, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 492, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 492, + 311 + ], + "type": "text", + "content": "5.3. Adversarial training based defense" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "type": "text", + "content": "Our simple defense that is adopted from standard adversarial training is explained in Section 4. The results for defending against attacking A-ViT are shown in Table 6. The original A-ViT reduces the GFLOPs from 1.26 to 0.87, our attack increases it back to 1.26 with " + }, + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "type": "text", + "content": " attack success. The proposed defense reduces the GFLOPs to 1.01 which is still higher than the original 0.87. We hope our paper encourages the community to develop better defense methods to reduce the vulnerability of efficient vision transformers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 448, + 379, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 379, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 379, + 460 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 471, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 674 + ], + "type": "text", + "content": "Recently, we have seen efficient vision transformer models in which the computation is adaptively modified based on the input. We argue that this is an important research direction and that there will be more progress in this direction in the future. However, we show that the current methods are vulnerable to a universal adversarial patch that increases the computation and thus power consumption at inference time. Our experiments show promising results for three SOTA efficient transformer models, where a small patch that is optimized on the training data can increase the computation to the maximum possible level in the testing data in some settings. We also propose a defense that reduces the effectiveness of our attack. We hope that our paper will encourage the community to study such attacks and develop better defense methods on various machine learning methods, including generative models, that reduce the computation adaptively with the input." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment: This work was partially supported by DARPA under Contract No. HR00112190135 and HR00112290115 and NSF grant 1845216." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24793" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "text", + "content": "[1] iphone battery capacity. https://bigthink.com/the-future/battery-technology-lags/.1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 289, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 289, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 289, + 146 + ], + "type": "text", + "content": "[2] Starship robot. https://www.wevolver.com/specs/starship-technologies-starship-robot.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[3] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "text", + "content": "[4] Babak Ehteshami Bejnordi, Tijmen Blankevoort, and Max Welling. Batch-shaping for learning conditional channel gated networks. arXiv preprint arXiv:1907.06627, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 317 + ], + "type": "text", + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 319, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 351 + ], + "type": "text", + "content": "[7] Jason Ross Brown, Yiren Zhao, Ilia Shumailov, and Robert D Mullins. Dartformer: Finding the best type of attention. arXiv preprint arXiv:2210.00641, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 353, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 288, + 385 + ], + "type": "text", + "content": "[8] Tom B Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 430 + ], + "type": "text", + "content": "[9] Shaofeng Cai, Yao Shu, and Wei Wang. Dynamic routing networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3597, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 432, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 288, + 498 + ], + "type": "text", + "content": "[10] Shijie Cao, Lingxiao Ma, Wencong Xiao, Chen Zhang, Yunxin Liu, Lintao Zhang, Lanshun Nie, and Zhi Yang. Seernet: Predicting convolutional neural network feature-map sparsity through low-bit quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11216-11225, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 543 + ], + "type": "text", + "content": "[11] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "type": "text", + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 634 + ], + "type": "text", + "content": "[13] Guangyi Chen, Chunze Lin, Liangliang Ren, Jiwen Lu, and Jie Zhou. Self-critical attention learning for person re-identification. In ICCV, pages 9637–9646, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[14] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8064-8073, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[15] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. Slothbomb: Efficiency poisoning attack against dynamic neural networks. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Simin Chen, Mirazul Haque, Zihe Song, Cong Liu, and Wei Yang. Transslowdown: Efficiency attacks on neural machine translation systems. 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "type": "text", + "content": "[17] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsglowsdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "text", + "content": "[18] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 547, + 252 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 254, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 319 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 320, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 547, + 396 + ], + "type": "text", + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "text", + "content": "[22] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv preprint arXiv:1910.10073, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "type": "text", + "content": "[23] Quanfu Fan, Chun-Fu (Ricarhd) Chen, Hilde Kuehne, Marco Pistoia, and David Cox. More Is Less: Learning Efficient Video Representations by Temporal Aggregation Modules. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "text", + "content": "[24] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 547, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 547, + 621 + ], + "type": "text", + "content": "[25] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 396-414. Springer, 2022. 2, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "text", + "content": "[26] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res, 23:1-40, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 689 + ], + "type": "text", + "content": "[27] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "type": "text", + "content": "[28] Michael Figurnov, Maxwell D Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24794" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Spatially adaptive computation time for residual networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1039-1048, 2017. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[29] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[30] Yonggan Fu, Shunyao Zhang, Shang Wu, Cheng Wan, and Yingyan Lin. Patch-fool: Are vision transformers always robust against adversarial perturbations? arXiv preprint arXiv:2203.08392, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "text", + "content": "[31] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. arXiv preprint arXiv:1810.05331, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 274 + ], + "type": "text", + "content": "[32] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In arXiv preprint arXiv:1412.6572, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 297 + ], + "type": "text", + "content": "[33] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 288, + 331 + ], + "type": "text", + "content": "[34] Jiaqi Guan, Yang Liu, Qiang Liu, and Jian Peng. Energy-efficient amortized inference with cascaded deep classifiers. arXiv preprint arXiv:1710.03368, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "text", + "content": "[35] Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14264-14273, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 377, + 288, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 410 + ], + "type": "text", + "content": "[36] Mirazul Haque, Simin Chen, Wasif Arman Haque, Cong Liu, and Wei Yang. Nodeattack: Adversarial attack on the energy consumption of neural odes. 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 288, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 444 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 444 + ], + "type": "text", + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 499 + ], + "type": "text", + "content": "[38] Charles Herrmann, Richard Strong Bowen, and Ramin Zabih. Channel selection using gumbel softmax. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII, pages 241-257. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 533 + ], + "type": "text", + "content": "[39] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 567 + ], + "type": "text", + "content": "[40] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv preprint arXiv:1912.12180, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[41] Sanghyun Hong, Yigitcan Kaya, Ionut-Vlad Modoranu, and Tudor Dumitraş. A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "text", + "content": "[42] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[43] Weizhe Hua, Yuan Zhou, Christopher M De Sa, Zhiru Zhang, and G Edward Suh. Channel gating neural networks. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[44] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[45] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 172, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 547, + 205 + ], + "type": "text", + "content": "[46] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of self-attention. arXiv preprint arXiv:2209.04881, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 205, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 547, + 248 + ], + "type": "text", + "content": "[47] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 250, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 546, + 282 + ], + "type": "text", + "content": "[48] Soroush Abbasi Koohpayegani and Hamed Piriavash. Sima: Simple softmax-free attention for vision transformers. arXiv preprint arXiv:2206.08898, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 283, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 546, + 304 + ], + "type": "text", + "content": "[49] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 305, + 546, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 546, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 546, + 347 + ], + "type": "text", + "content": "[50] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 349, + 546, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 546, + 381 + ], + "type": "text", + "content": "[51] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 383, + 546, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 546, + 425 + ], + "type": "text", + "content": "[52] Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE international conference on computer vision, pages 5764-5772, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "type": "text", + "content": "[53] Zhichao Li, Yi Yang, Xiao Liu, Feng Zhou, Shilei Wen, and Wei Xu. Dynamic computational time for visual attention. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1199-1209, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 471, + 546, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 546, + 504 + ], + "type": "text", + "content": "[54] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. arXiv preprint arXiv:2209.09004, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 504, + 546, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 546, + 558 + ], + "type": "text", + "content": "[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "text", + "content": "[56] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 614, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 635 + ], + "type": "text", + "content": "[57] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. 2019. 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 636, + 546, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 546, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 546, + 680 + ], + "type": "text", + "content": "[58] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "text", + "content": "[59] Wenhao Lu, Jian Jiao, and Ruofei Zhang. Twinbert: Distilling knowledge to twin-structured bert models for efficient retrieval. arXiv preprint arXiv:2002.06275, 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24795" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[60] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 288, + 150 + ], + "type": "text", + "content": "[61] Kaleel Mahmood, Rigel Mahmood, and Marten Van Dijk. On the robustness of vision transformers to adversarial examples. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7838-7847, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "text", + "content": "[62] Dmitrii Marin, Jen-Hao Rick Chang, Anurag Ranjan, Anish Prabhu, Mohammad Rastegari, and Oncel Tuzel. Token pooling in vision transformers. arXiv preprint arXiv:2110.03860, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 260 + ], + "type": "text", + "content": "[63] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "type": "text", + "content": "[64] Jianhong Pan, Qichen Zheng, Zhipeng Fan, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Gradauto: Energy-oriented attack on dynamic neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV, pages 637-653. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "text", + "content": "[65] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 446 + ], + "type": "text", + "content": "[66] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 448, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 288, + 503 + ], + "type": "text", + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "type": "text", + "content": "[68] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 590 + ], + "type": "text", + "content": "[69] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "text", + "content": "[70] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[71] Mengye Ren, Andrei Pokrovsky, Bin Yang, and Raquel Urtasun. Sbnet: Sparse blocks network for fast inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8711-8720, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[72] Aniruddha Saha, Akshayvarun Subramanya, Koninika B. Patil, and Hamed Pirsavash. Adversarial patches exploiting contextual reasoning in object detection. In ArXiv, 2019. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[73] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 128, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 547, + 183 + ], + "type": "text", + "content": "[74] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3531-3539, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 183, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 547, + 227 + ], + "type": "text", + "content": "[75] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "text", + "content": "[76] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 282, + 546, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 546, + 337 + ], + "type": "text", + "content": "[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 337, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 546, + 392 + ], + "type": "text", + "content": "[78] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 393, + 546, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 546, + 437 + ], + "type": "text", + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 437, + 546, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 546, + 480 + ], + "type": "text", + "content": "[80] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 481, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 546, + 525 + ], + "type": "text", + "content": "[81] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the iee/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 525, + 546, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 546, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 546, + 570 + ], + "type": "text", + "content": "[82] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "type": "text", + "content": "[83] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "type": "text", + "content": "[84] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "text", + "content": "[85] Zhenda Xie, Zheng Zhang, Xizhou Zhu, Gao Huang, and Stephen Lin. Spatially adaptive inference with stochastic feature sampling and interpolation. In Computer Vision-ECCV" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24796" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 686 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "text", + "content": "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 531-548. Springer, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 160 + ], + "type": "text", + "content": "[86] Le Yang, Yizeng Han, Xi Chen, Shiji Song, Jifeng Dai, and Gao Huang. Resolution adaptive networks for efficient inference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2369-2378, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "text", + "content": "[87] Zerui Yang, Yuhui Xu, Wenrui Dai, and Hongkai Xiong. Dynamic-stride-net: Deep convolutional neural network with dynamic stride. In Optoelectronic Imaging and Multimedia Technology VI, pages 42-53. SPIE, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "text", + "content": "[88] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 288, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 288, + 318 + ], + "type": "text", + "content": "[89] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819-10829, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "type": "text", + "content": "[90] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 430 + ], + "type": "text", + "content": "[91] Zhihang Yuan, Bingzhe Wu, Guangyu Sun, Zheng Liang, Shiwan Zhao, and Weichen Bi. S2dnas: Transforming static cnn model for dynamic inference via neural architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 175-192. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 288, + 485 + ], + "type": "text", + "content": "[92] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 288, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 519 + ], + "type": "text", + "content": "[93] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 521, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 288, + 586 + ], + "type": "text", + "content": "[94] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 587, + 288, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 288, + 640 + ], + "type": "text", + "content": "[95] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 643, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 288, + 686 + ], + "type": "text", + "content": "[96] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24797" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_content_list.json b/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8878615d176eac6246a6acd74be2912bd65dd719 --- /dev/null +++ b/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_content_list.json @@ -0,0 +1,1432 @@ +[ + { + "type": "text", + "text": "Small Scale Data-Free Knowledge Distillation", + "text_level": 1, + "bbox": [ + 253, + 130, + 715, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "He Liu $^{1,2*}$ , Yikai Wang $^{2*}$ , Huaping Liu $^{2}$ , Fuchun Sun $^{2}$ , Anbang Yao $^{3\\dagger}$", + "bbox": [ + 192, + 179, + 776, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1China Telecom Cloud Technology Co., Ltd.", + "bbox": [ + 305, + 204, + 661, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ BNRist Center, State Key Lab on Intelligent Technology and Systems,", + "bbox": [ + 202, + 223, + 766, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science and Technology, Tsinghua University $^{3}$ Intel Labs China", + "bbox": [ + 120, + 241, + 848, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liuhel7@mails.,yikaiw@,hpliu@,fuchuns@}tsinghua.edu.cn，anbang.yao@intel.com", + "bbox": [ + 142, + 261, + 823, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data, avoiding privacy, security, and proprietary risks in real applications. In this line of research, existing methods typically follow an inversion-and-distillation paradigm in which a generative adversarial network on-the-fly trained with the guidance of the pre-trained teacher network is used to synthesize a large-scale sample set for knowledge distillation. In this paper, we reexamine this common data-free knowledge distillation paradigm, showing that there is considerable room to improve the overall training efficiency through a lens of \"small-scale inverted data for knowledge distillation\". In light of three empirical observations indicating the importance of how to balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, we propose Small Scale Data-free Knowledge Distillation (SSD-KD). In formulation, SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples, facilitated by a dynamic replay buffer and a reinforcement learning strategy. As a result, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (e.g., $10 \\times$ less than the original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream methods while retaining superior or competitive model performance, as demonstrated on popular image classification and semantic segmentation benchmarks. The code is available at https://github.com/OSVAI/SSD-KD.", + "bbox": [ + 75, + 359, + 472, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 310, + 632, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For computer vision applications on resource-constrained devices, how to learn portable neural networks yet with satisfied prediction accuracy is the key problem. Knowledge distillation (KD) [2, 12, 17, 22, 25, 33], which leverages the information of a pre-trained large teacher network to promote the training of a smaller target student network on the same training data, has become a mainstream solution. Conventional KD methods assume that the original training data is always available. However, accessing the source dataset on which the teacher network was trained is usually not feasible in practice, due to its potential privacy or security or proprietary or huge-size concerns. To relax the constraint on training data, knowledge distillation under a data-free regime has recently attracted increasing attention.", + "bbox": [ + 496, + 338, + 893, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The basic idea of Data-free Knowledge Distillation (D-KD) [3, 18, 21] is to construct synthetic samples for knowledge distillation conditioned on the pre-trained teacher network, which would match the underlying distribution of the original training data. Existing top-performing D-KD methods [4, 7, 8, 10, 24, 30, 31] generally adopt an adversarial inversion-and-distillation paradigm. Under this paradigm, (1) during the inversion process, a generator is trained by taking the pre-trained teacher network as the discriminator; (2) during the subsequent knowledge distillation process, the on-the-fly learned generator will synthesize pseudo samples for training the student network. However, adversarial D-KD methods usually require generating a large number of synthetic samples (compared to the original training dataset size) in order to guarantee trustworthy knowledge distillation. This poses a heavy burden on training resource consumption, suppressing their use in real applications. In the recent work of [11], the authors present an effective meta-learning strategy that seeks common features and reuses them as initial priors to reduce the number of iteration steps required to reach the convergence of the generator. Although faster data synthesis can be attained, [11] still needs to generate a sufficiently large number of syn", + "bbox": [ + 496, + 551, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author. This work was done when He Liu was an intern at Intel Labs China, supervised by Anbang Yao who conceived the project.", + "bbox": [ + 75, + 862, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6008", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7f42998a74db30c7b8e815a1189712969ecef13a265d814d68391212031450a.jpg", + "image_caption": [ + "Figure 1. Comparison of knowledge distillation (KD) using original samples vs. synthetic samples, under the same training data scale: 5000 samples (10% of the CIFAR-10 training dataset size). In such small-scale KD regime, the student models (DeepInv [30] and SSD-KD) trained on synthetic samples always show much better accuracy than the counterparts (Vanilla KD [12]) trained on original samples." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 367, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9c240abb887b9a1ec2f64ca22ed33a7533dc2c8bd1ed8faa10f24cede29b72e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 89, + 598, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e18545f7d22b152170307a88c4da13443fd3a243b0468e8f156b885606daf5f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 90, + 826, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe139fdc34149eec03ccb445c88c9a599773d42a7ab2b60cd0f813f59566d066.jpg", + "image_caption": [ + "(a) Difficulty distribution." + ], + "image_footnote": [], + "bbox": [ + 81, + 265, + 277, + 381 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9b6198df7fadf4575acb67f2d97055cfc86844e36daf27473f20cf505ca95460.jpg", + "image_caption": [ + "Figure 2. Comparison of synthetic sample distributions collected from two top-performing adversarial D-KD methods (DeepInv [30] and Fast10 [11]) and our SSD-KD on CIFAR-100 dataset. Comparatively, Fig. 2a shows that our method can better balance the difficulty distribution of synthetic samples while encouraging the generator to invert more hard samples, and Fig. 2b further shows that our method can better balance the diversity distribution of synthetic samples across different categories." + ], + "image_footnote": [], + "bbox": [ + 282, + 266, + 480, + 381 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6607a92b3d6f91b26564e6bb5369f1fea0582bc266653b3d9ffdfc99ab03b42c.jpg", + "image_caption": [ + "(b) Diversity distribution." + ], + "image_footnote": [], + "bbox": [ + 488, + 263, + 687, + 381 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7448c10dad09d474c88480a5743946261a5faeceb2bd86f85d29f4ec14f9de9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 263, + 888, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thetic samples to ensure effective knowledge distillation, neglecting the efficiency of the following knowledge distillation process which will become the major bottleneck to the overall training efficiency. In a nutshell, there is no research effort made to improve the overall training efficiency of D-KD via jointly considering data inversion and knowledge distillation processes, to the best of our knowledge.", + "bbox": [ + 75, + 470, + 468, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To remedy this critical gap, this paper presents the first fully efficient D-KD approach termed Small Scale Data-free Knowledge Distillation (SSD-KD). Our SSD-KD improves the overall training efficiency of the adversarial inversion-and-distillation paradigm from a novel \"small data scale\" perspective. In this work, \"data scale\" refers to the total number of inverted samples used in knowledge distillation during a training epoch. The formulation of SSD-KD is inspired by three empirical observations. On different pairs of teacher-student networks, we first observe that the student networks trained on synthetic samples tend to show much better performance than their corresponding counterparts trained on original samples when significantly reducing data scales of synthetic samples and original samples to the same (e.g., $10\\%$ of the source training dataset size), as illustrated in Fig. 1. Note that synthetic samples are generated with the guidance of the teacher network pre-trained on the whole source dataset, which naturally reflect different views of the original data distribution. Under a small enough data scale, this makes synthetic samples have superior capability to original samples in fitting the underlying", + "bbox": [ + 75, + 583, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distribution of the whole source dataset. This inspiring observation indicates that if we can construct a small-scale set of high-quality synthetic samples, a promising way toward fully efficient D-KD would be created. In principle, we believe a high-quality small-scale synthetic dataset should have well-balanced class distributions in terms of both synthetic sample diversity and difficulty. However, our other two observations indicate that existing D-KD methods including both conventional and the most efficient designs do not have good capabilities to balance the aforementioned two class distributions of synthetic samples under a small data scale, as illustrated in Fig. 2. Note that there already exist a few D-KD methods to enhance the diversity of synthetic samples [7, 10, 15, 30], but the diversity of synthetic samples in terms of sample difficulty is not explored yet.", + "bbox": [ + 496, + 470, + 892, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Driven by the above observations and analysis, we come up with our SSD-KD which introduces two interdependent modules to significantly accelerate the overall training efficiency of the predominant adversarial inversion-and-distillation paradigm. The first module of SSD-KD relies on a novel modulating function that defines a diversity-aware term and a difficulty-aware term to jointly balance the class distributions of synthetic samples during both data synthesis and knowledge distillation processes in an explicit manner. The second module defines a novel priority sampling function facilitated by a reinforcement learning strategy that selects a small portion of proper synthetic samples from candidates stored in a dynamic replay buffer for knowledge", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6009", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distillation, further improving the end-to-end training efficiency. Benefiting from these two modules, SSD-KD has two appealing merits. On one side, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples ( $10 \\times$ less than original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream D-KD methods while retaining competitive model performance. On the other side, SSD-KD attains largely improved performance in student model accuracy and maintains overall training efficiency when relaxing the data scale of synthetic samples to a relatively large number (which is still smaller than those for existing D-KD methods). We validate these merits of our method by lots of experiments on popular image classification and semantic segmentation benchmarks.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 330, + 218, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-free knowledge distillation. D-KD is originally explored by [18] which assumes that layer-wise activation records of a well-trained teacher network pre-computed on training samples is available. The prevailing inversion-and-distillation paradigm for D-KD without reliance on the original training data or the recorded metadata is introduced in [21]. In the inversion process, synthetic training samples are generated by leveraging the information learned by the teacher network, whose distribution is expected to fit the underlying distribution of the original training dataset. In the distillation process, the student network is trained on synthetic samples by forcing it to match the predictions of the teacher network. Based on this paradigm, [4, 31] use generative adversarial networks for data inversion. Subsequent D-KD methods mostly follow this adversarial inversion-and-distillation paradigm. They try to improve the data inversion process from different aspects, such as enhancing synthetic sample discrimination with contrastive learning [10] or an ensemble of generators [19], combating distribution shift with momentum adversarial learning [8] or meta learning [24], and promoting adversarial learning with data augmentation [15, 32]. Our method intends to improve the overall training efficiency of the adversarial D-KD, and differs from these methods both in focus and formulation.", + "bbox": [ + 75, + 356, + 472, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient synthetic data sampling. How to select proper synthetic samples for knowledge distillation is essential in D-KD research. Existing methods [8, 10, 11, 24] commonly rely on a memory bank to store synthetic samples, and directly update synthetic samples without considering the efficiency of the following knowledge distillation process. In sharp contrast, our method introduces a reinforcement learning strategy that adaptively selects appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty, significantly improving knowledge dis", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tillation efficiency. As far as we know, our method made the first attempt to extend reinforcement learning methodology [1, 13, 26] to address data-free knowledge distillation.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 148, + 591, + 164 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: D-KD", + "text_level": 1, + "bbox": [ + 498, + 172, + 697, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $f_{t}(\\cdot ;\\theta_{t})$ be a teacher model pre-trained on the original task dataset that is no longer accessible, the goal of D-KD is to first construct a set of synthetic training samples $x$ via inverting the data distribution information learned by the teacher model, on which a target student model $f_{s}(\\cdot ;\\theta_{s})$ then can be trained by forcing it to mimic the teacher's function. Existing D-KD methods mostly use a generative adversarial network $g(\\cdot ;\\theta_{g})$ for producing synthetic training samples $x = g(z;\\theta_g)$ from the latent noise input $z$ , which is trained by taking the teacher model as the discriminator.", + "bbox": [ + 496, + 196, + 890, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The optimization of D-KD contains a common distillation regularization to minimize the teacher-student function discrepancy $\\mathcal{L}_{\\mathrm{KD}}(x) = D_{\\mathrm{KD}}(f_t(x;\\theta_t)\\| f_s(x;\\theta_s))^1$ based on the KL-divergence, and a task-oriented regularization $\\mathcal{L}_{\\mathrm{Task}}(x)$ , e.g., the cross-entropy loss using the teacher's predication as the ground truth. Besides, since D-KD is primarily based on the assumption that the teacher model has been optimized to be capable of capturing the source training data distribution after pre-training, recent D-KD methods [7, 29, 30] introduce an extra loss to regularize the statistics (Batch-Normalization (BN) parameters) of the training data distribution during the data inversion process,", + "bbox": [ + 496, + 348, + 892, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B N}} (x) = \\sum_ {l} \\left| \\left| \\mu_ {l} (x) - \\mathbb {E} (\\mu_ {l}) \\right| \\right| _ {2} + \\left| \\left| \\sigma_ {l} ^ {2} (x) - \\mathbb {E} (\\sigma_ {l} ^ {2}) \\right| \\right| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 537, + 890, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_l(\\cdot)$ and $\\sigma_l(\\cdot)$ denote the batch-wise mean and variance estimates of feature maps at the $l$ -th layer, respectively; $\\mathbb{E}(\\cdot)$ over the BN statistics can be approximately substituted by running mean or variance.", + "bbox": [ + 496, + 578, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The effectiveness of D-KD heavily depends on the quality of synthetic samples which are inverted from leveraging the knowledge of the pre-trained teacher model. The prevailing adversarial D-KD paradigm consists of two processes, namely data inversion and knowledge distillation. From both perspectives of efficiency and efficacy, on the one hand, the data inversion process largely affects the optimization performance of the student model; on the other hand, the training time cost of knowledge distillation emerges as a significant constraint to the overall training efficiency of D-KD.", + "bbox": [ + 496, + 638, + 890, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Our Design: SSD-KD", + "text_level": 1, + "bbox": [ + 500, + 813, + 702, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regarding our SSD-KD, we focuses on improving the adversarial D-KD paradigm through a lens of \"small-scale", + "bbox": [ + 496, + 835, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "${}^{1}$ Concretely, ${f}_{t}\\left( {x;{\\theta }_{t}}\\right)$ and ${f}_{s}\\left( {x;{\\theta }_{s}}\\right)$ are the outputted logits (before softmax) of the teacher model and the student model, respectively.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6010", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc45cced476971ff7311f2e7e0ee4076b0abfcf0eb660e02de925fa8f62e45cf.jpg", + "image_caption": [ + "Figure 3. Comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] (left) and more efficient family [10, 11] (middle), and our SSD-KD (right). Our SSD-KD formulates a reinforcement learning strategy that can flexibly seek appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty distributions. See the method section for notation definitions." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 289, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4dfee92936b674cac3945ad7aff790f96c23b78efcfe8e52c032a1b485a8ee8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 85, + 880, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inverted data for knowledge distillation\". SSD-KD lays emphasis on instructing the data inversion process with the feedback of both the pre-trained teacher and the knowledge distillation process, significantly accelerating the overall training efficiency. Following the notations in the prevision subsection, the optimization goal of our SSD-KD is defined as", + "bbox": [ + 75, + 296, + 472, + 401 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {f _ {s}} \\max _ {g} \\mathbb {E} _ {x = \\delta \\circ g (z)} \\left(\\mathcal {L} _ {\\mathrm {B N}} (x) + \\mathcal {L} _ {\\mathrm {K D}} (x) + \\phi (x) \\mathcal {L} _ {\\text {T a s k}} (x)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 414, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where,", + "text_level": 1, + "bbox": [ + 76, + 448, + 125, + 460 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We adopt a diversity-aware modulating function $\\phi(x)$ that allocates each synthetic sample with a different priority regarding its predicted category by the teacher model, as presented in Sec. 3.3.", + "- Under the constraint of BN estimates, with $\\phi(x)$ , we encourage the generator to explore as tough synthetic samples (w.r.t. the teacher model) as possible, as introduced in Sec. 3.3.", + "- Instead of applying a random sampling strategy to select samples for knowledge distillation, we adopt a reweighting strategy to control the sampling process. We abuse notation slightly with $\\circ$ to represent applying the strategy based on priority sampling function $\\delta$ , with more details in Sec. 3.4.", + "- Each synthetic sample is not only prioritized by its modulating function $\\phi(x)$ but also is reweighted at the sampling stage that reuses the same intermediate values as $\\phi(x)$ ." + ], + "bbox": [ + 76, + 462, + 468, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the D-KD pipeline allows training samples to be synthesized and served for training the student model on the same task. However, there is a large extent of data redundancy that hinders the training efficiency of D-KD methods. In the following sections, we detail our SSD-KD, a fully efficient D-KD method that is capable of using an extremely small scale of synthetic data yet achieving competitive performance compared to existing D-KD methods.", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pipeline of SSD-KD is summarized in Alg. 1 and the comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] and more efficient family [10, 11], and our", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SSD-KD is shown in Fig. 3.", + "bbox": [ + 500, + 296, + 689, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Inversion with Distribution Balancing", + "text_level": 1, + "bbox": [ + 500, + 319, + 867, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide Fig. 2 to demonstrate the data redundancy of D-KD that results from a large imbalance of the synthetic data. The left two sub-figures of Fig. 2 depict the distribution of categories predicted by the teacher model, indicating a significant imbalance in data categories. The right two sub-figures of Fig. 2 show sample accounts of different bars that correspond to different prediction difficulties (the difficulty is measured by the predicted probability by the teacher model). For D-KD, it indicates that generating samples with only the instruction of teacher-student discrepancy results in a sharp distribution over the sample difficulty and tends to obtain easily predicted data samples. We argue that for the D-KD task where data samples are all synthesized, the data-generating process needs to consider both the teacher-student discrepancy and the pre-trained knowledge of the teacher itself, by which our SSD-KD proposes diversity-aware and difficulty-aware data synthesis, as detailed below.", + "bbox": [ + 496, + 340, + 890, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diversity-aware balancing. We first propose to address the issue of imbalanced sample difficulty in the data inversion process. Specifically, we maintain a replay buffer $\\mathcal{B}$ that stores a constant amount (denoted as $|\\mathcal{B}|$ ) of synthetic data samples. For each data sample $x$ in $\\mathcal{B}$ , we penalize its total amount of samples that share the same predicted category (by the teacher model) with $x$ . To realize this, we adopt a diversity-aware balancing term that encourages the generator to synthesize samples with infrequent categories, which will be shown in Eq. (3).", + "bbox": [ + 496, + 598, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Difficulty-aware balancing. Drawing inspiration from the field of object detection that utilizes focal loss for largely imbalanced samples [16, 27], for each sample $x$ , we further introduce a difficulty-aware balancing term on the predicted probability $p_T(x)$ . Here, difficult synthetic samples are considered as those with low-confidence predictions by the teacher model, which are encouraged by the difficulty-aware balancing term as will be given in Eq. (3).", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, we introduce a modulating function $\\phi (x)$ to adjust the optimization of the generator based on the pre", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6011", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diction feedback from the pre-trained teacher model. $\\phi (x)$ is expected to balance the category distribution and dynamically distinguish between easy and hard synthetic samples, by which easy samples no longer overwhelm the distillation process. Formally, for a synthetic data sample $x\\in \\mathcal{B}$ , its modulating function $\\phi (x)$ is computed by", + "bbox": [ + 75, + 90, + 468, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (x) = \\underbrace {\\left(1 - \\frac {1}{| \\mathcal {B} |} \\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\mathbb {I} _ {c _ {T} \\left(x ^ {\\prime}\\right) = c _ {T} (x)}\\right)} _ {\\text {d i v e r s i t y - a w a r e b a l a n c i n g}} \\underbrace {\\left(1 - p _ {T} (x)\\right) ^ {\\gamma}} _ {\\text {d i f f i c u l t y - a w a r e b a l a n c i n g}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 189, + 468, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{T}(x)$ and $p_{T}(x)$ refer to the predicted category index and probability (or confidence) by the pre-trained teacher model, respectively; $\\mathbb{I}_{c_T(x') = c_T(x)}$ denotes an indicator function that equals to 1 if the predicted category of $x'$ is the same with that of $x$ , while otherwise 0; $\\gamma$ is a hyperparameter.", + "bbox": [ + 75, + 250, + 468, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We highlight two properties of the modulating function $\\phi(x)$ . Firstly, for the data sample $x$ with a high prediction certainty w.r.t. the teacher model (i.e., considered as an easy sample), $\\phi(x)$ approaches to a low value and thus results in low impact on the task-oriented loss $\\mathcal{L}_{\\mathrm{Task}}(x)$ in Eq. (2). Secondly, when the category distribution of synthetic data samples in $\\mathcal{B}$ is largely imbalanced as predicted by the teacher network, the sample $x$ of which the category shares with more samples in $\\mathcal{B}$ is penalized and thus the corresponding $\\mathcal{L}_{\\mathrm{Task}}(x)$ is weakened by $\\phi(x)$ .", + "bbox": [ + 75, + 340, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although Eq. (3) implies that the value of the modulating function $\\phi(x)$ is partially determined by the current replay buffer $\\mathcal{B}$ , note that $\\mathcal{B}$ changes dynamically and is also affected by $\\phi(x)$ . This is due to that the term $\\phi(x)\\mathcal{L}_{\\mathrm{Task}}(x)$ in Eq. (2) directly optimizes the generator that synthesizes data samples to compose $\\mathcal{B}$ . In this sense, the balancing in category diversity is maintained during training given the mutual effects of $\\mathcal{B}$ and $\\phi(x)$ . We find it particularly important in the data inversion process for balancing categories of data samples.", + "bbox": [ + 75, + 491, + 468, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With both balancing terms, as shown in Fig. 2, our method (SSD-KD) generates mild sample distributions in terms of both sample category and difficulty.", + "bbox": [ + 75, + 642, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Distillation with Priority Sampling", + "text_level": 1, + "bbox": [ + 76, + 695, + 379, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The original prioritized experience replay method [26] reuses important transitions more frequently and learns more efficiently. Differently, rather than obtaining the reward from the environment, our prioritized sampling method is designed to fit in data-free knowledge distillation and get feedback from the framework itself. In other words, the prioritized sampling method performs the opposite role of the previous data-free knowledge distillation methods: it focuses on training a sparse set of highly prioritized samples instead of uniform sampling to speed up training.", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By Eq. (3), we sample the synthetic data $x$ from the current replay buffer $\\mathcal{B}$ . Instead of uniformly sampling $x$ , we", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: SSD-KD" + ], + "code_body": "Input: Number of training epochs $E$ ; number of iterations for data inversion $T_{g}$ ; number of iterations for distillation $T_{kd}$ ; pre-trained teacher model $f_{t}(\\cdot;\\theta_{t})$ ; student network $f_{s}(\\cdot;\\theta_{s})$ \nOutput: The optimized student model $f_{s}(\\cdot;\\theta_{s})$ \n1 Initialize replay buffer $\\mathcal{B} \\gets \\emptyset$ and $f_{s}(\\cdot;\\theta_{s})$ \n2 for $e = 1$ to $E$ do \n3 (1) Data Inversion with Distribution Balancing: \n4 Initialize a generative network $g(\\cdot;\\theta_{g})$ \n5 $z \\gets \\mathcal{N}(0,1)$ \n6 for $i = 1$ to $T_{g}$ do \n7 $x \\gets g(z;\\theta_{g})$ \n8 Compute the overall loss by Eq. (2) \n9 Update $z$ and $\\theta_{g}$ \n10 end \n11 (2) Distillation with Priority Sampling: \n12 Compute the importance-sampling weight $w_{i-1}(x)$ \n13 Compute the sample priority $\\delta_{i}(x)$ \n14 $\\mathcal{B} \\gets \\mathcal{B} \\cup \\{x,\\delta_{i}(x)\\}$ \n15 Remove old samples to keep $|\\mathcal{B}|$ a constant \n16 for $j = 1$ to $T_{kd}$ do \n17 Sample a small-scale mini-batch $M$ from $\\mathcal{B}$ \n18 Compute the distillation loss $\\mathcal{L}_{\\mathrm{KD}}(x)$ in Eq. (2), where $x \\in M$ \n19 Compute the importance-sampling weight $w_{i}(x)$ \n20 Compute the sample priority $\\delta_{i+1}(x)$ \n21 Update old samples $\\{x,\\delta_{i}(x)\\}$ to $\\{x,\\delta_{i+1}(x)\\}$ in $\\mathcal{B}$ \n22 Update $\\theta_{s}$ \n23 end \n24 end \n25 Return $f_{s}(\\cdot;\\theta_{s})$", + "bbox": [ + 500, + 111, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "propose to modulate the sampling probability by a sampling strategy termed Priority Sampling (PS). The basic function of PS is to measure the importance of each sample $x$ in $\\mathcal{B}$ , by which we introduce a priority sampling function $\\delta_{i}(x)$ ,", + "bbox": [ + 496, + 678, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {i} (x) = w _ {i - 1} (x) K L \\left(f _ {t} (x; \\theta_ {t}) \\mid \\mid f _ {s} \\left(x; \\theta_ {s}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 751, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where as mentioned in Sec. 3.1, $KL$ denotes the KL-divergence between the softmax outputs of logits $f_{t}(x;\\theta_{t})$ and $f_{s}(x;\\theta_{s})$ ; $\\theta_t$ , $\\theta_s$ depend on the training step $i$ ; $w_{i}(x)$ is the calibration term [26] for normalizing samples in $\\mathcal{B}$ , as will be formalized in Eq. (5), especially, when $i = 0$ , $w_{-1}(x) = 1$ .", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training of knowledge distillation with random updates relies on those updates corresponding to the same dis", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6012", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c0696876f5e6796947af7c73ee2bfb4671840db42853c624c66a957cbf21d19c.jpg", + "table_caption": [ + "Table 1. Performance comparison of Fast2 [11] (the current most efficient D-KD method) and our SSD-KD, in terms of top-1 classification accuracy $(\\%)$ and overall training time cost (hours). Our SSD-KD performs with a very small training data scale: 5000 synthetic samples, i.e., $10\\%$ of the original training dataset size. All image classification results in this and the other tables are averaged over three independent runs, and the methods \"Teacher\" and \"Student\" are performed on the whole original training dataset." + ], + "table_footnote": [], + "table_body": "
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
Fast2 [11]92.62(3.92h)84.67(3.02h)88.36(1.46h)89.56(2.37h)89.68(1.70h)
SSD-KD92.92(1.11h)89.56(0.77h)88.07(0.77h)90.12(0.96h)91.84(0.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
Fast2 [11]69.76(4.02h)62.83(2.99h)41.77(1.44h)53.15(2.26h)57.08(1.72h)
SSD-KD73.05(1.12h)66.66(0.78h)51.96(0.77h)59.60(0.97h)61.44(0.78h)
", + "bbox": [ + 199, + 152, + 769, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "distribution as its expectation. Prioritized sampling data introduces bias since it might change the data distribution, and affect the solution that the estimates will converge to. Thus we correct the bias by introducing an importance-sampling (IS) weight $w_{i}(x)$ for the data sample $x$ :", + "bbox": [ + 75, + 304, + 470, + 380 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} (x) = \\left(N \\cdot P _ {i} (x)\\right) ^ {- \\beta}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 391, + 468, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyper-parameter; $P_{i}(x)$ is the probability of sampling transition defined by", + "bbox": [ + 75, + 420, + 470, + 452 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} (x) = \\frac {\\left(\\left| \\delta_ {i} (x) \\right| + \\epsilon\\right) ^ {\\alpha}}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\left(\\left| \\delta_ {i} \\left(x ^ {\\prime}\\right) \\right| + \\epsilon\\right) ^ {\\alpha}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 460, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a small positive constant which prevents the edge-case of transitions not being selected once their priority is zero.", + "bbox": [ + 75, + 506, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The priority sampling function $\\delta(x)$ has two noteworthy properties. Firstly, as the delta value increases, $\\delta(x)$ reflects a greater information discrepancy between the teacher and student models for the synthetic samples in the current $\\mathcal{B}$ . The student model should therefore be optimized from samples with greater information discrepancy, as this facilitates the faster acquisition of the teacher model. Secondly, $\\delta(x)$ dynamically changes with each update iteration of the student and generative models. Consequently, when the student model acquires the teacher model's capabilities on certain samples, it continues to learn from samples with larger differences relative to the teacher model based on the new sample distribution. This further enhances the performance of the student model.", + "bbox": [ + 75, + 551, + 470, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 776, + 202, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Details", + "text_level": 1, + "bbox": [ + 76, + 801, + 276, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments on image classification and semantic segmentation tasks to evaluate the effectiveness and study the design of our SSD-KD.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For image classification experiments, we use CIFAR-10 and CIFAR-100 [14], two most popular datasets", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for D-KD research. For semantic segmentation experiments, we use the NYUv2 dataset [28]. On each dataset, the baseline models are trained with its standard data split.", + "bbox": [ + 496, + 304, + 890, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Setups. We follow the settings of [11] for basic experiments and comparisons. Regarding experiments on the CIFAR-10 and CIFAR-100 datasets, we use 5 different teacher-student model pairs having either the same type or different type network architectures (see Table 1, 2). Regarding experiments on the NYUv2 dataset, we use two Deeplabv3 models [6] as a teacher-student model pair (see Table 3). Unless otherwise stated, we always adopt the same basic settings as in [11] for experiments, including the number of training epochs, the optimizer, the weight decay, etc. Detailed settings are put in the supplementary material.", + "bbox": [ + 496, + 349, + 892, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Besides comparing the student model accuracy (top-1 accuracy for image classification and mean Intersection over Union (IoU) for semantic segmentation), we also compare the overall training time cost of existing mainstream D-KD methods [4, 7, 10, 11, 20, 30] and our SSD-KD. For each teacher-student model pair, we meticulously record the total time cost (hours) for each run of the end-to-end training by all methods. In order to guarantee a fair comparison, all training speed assessments are performed on 1 NVIDIA V100 GPU using 12 cores of Intel Xeon Gold 6240R CPU. Our experiments are implemented with the PyTorch [23] library. For each teacher-student model pair, the experiment is conducted with three independent runs for all methods, and in comparison we report the averaged results, unless otherwise stated.", + "bbox": [ + 496, + 517, + 893, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 755, + 702, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on image classification task. As the main focus of our work is to improve the overall training efficiency of the adversarial data-free knowledge distillation paradigm, we first compare the proposed SSD-KD with the current most efficient D-KD method Fast2 [11] on the CIFAR-10 and CIFAR-100 datasets. From the results shown in Table 1, we can observe: (1) on the CIFAR-10 dataset, our SSD-KD (using a very small data scale: 5000 synthetic samples,", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6013", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/46674af9fe3ac0c73e7aa69ddad434e8ec6791fbb3a6f527594581d972325838.jpg", + "table_caption": [ + "Table 2. Performance comparison of existing top-performing D-KD methods and our SSD-KD, in terms of top-1 classification accuracy $(\\%)$ and overall training time cost (hours). For our SSD-KD, we relax the data scale of synthetic samples to be similar to that for Fast5." + ], + "table_footnote": [], + "table_body": "
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
DeepInv [30]93.26(25.62h)90.36(13.19h)83.04(10.62h)86.85(13.84h)89.72(11.67h)
CMI [10]94.84(15.49h)91.13(10.99h)90.01(9.87h)92.78(11.57h)92.52(10.58h)
DAFL [4]92.22(303.50h)81.10(209.78h)65.71(96.57h)81.33(157.44h)81.55(118.52h)
DFQ [7]94.61(152.16h)90.84(104.93h)86.14(48.40h)91.69(78.81h)92.01(59.40h)
ZSKT [20]93.32(304.11h)89.46(209.65h)83.74(96.75h)86.07(157.51h)89.66(118.82h)
Fast5 [11]93.63(4.34h)89.94(3.15h)88.90(1.61h)92.04(2.44h)91.96(1.89h)
Fast10 [11]94.05(4.72h)90.53(3.22h)89.29(1.72h)92.51(2.48h)92.45(2.05h)
SSD-KD94.26(4.20h)90.67(2.95h)89.96(1.48h)93.23(2.30h)93.11(1.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
DeepInv [30]61.32(25.88h)54.13(13.13h)53.77(10.62h)61.33(13.86h)61.34(11.68h)
CMI [10]77.04(15.42h)70.56(11.28h)57.91(10.01h)68.88(11.57h)68.75(10.53h)
DAFL [4]74.47(303.54h)54.16(209.82h)20.88(96.62h)42.83(157.49h)43.70(118.67h)
DFQ [7]77.01(152.18h)66.21(105.14h)51.27(48.47h)54.43(78.83h)64.79(59.48h)
ZSKT [20]67.74(304.13h)54.31(209.63h)36.66(96.87h)53.60(157.68h)54.59(118.84h)
Fast5 [11]72.82(4.35h)65.28(3.06h)52.90(1.56h)61.80(2.35h)63.83(1.86h)
Fast10 [11]74.34(4.50h)67.44(3.12h)54.02(1.65h)63.91(2.42h)65.12(1.95h)
SSD-KD75.16(4.22h)68.77(2.94h)55.61(1.52h)64.57(2.27h)65.28(1.78h)
", + "bbox": [ + 138, + 125, + 828, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "i.e., $10\\%$ of the original training dataset size) shows large improvements in training efficiency to Fast2 (at least $1.90\\times$ and at most $3.92\\times$ training speedup, on 5 teacher-student model pairs), and gets better student models on 4 out of 5 teacher-student model pairs; (2) on the more challenging CIFAR-100 dataset, our SSD-KD shows a very similar training speedup trend against Fast2 as on the CIFAR-10 dataset, but gets significantly better student models on all 5 teacher-student model pairs (at least $3.29\\%$ and at most $10.19\\%$ absolute top-1 accuracy gain to Fast2). The superior performance of SSD-KD against Fast2 validates the efficacy of our small-scale data inversion and sampling mechanism which can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes.", + "bbox": [ + 75, + 441, + 472, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we compare the proposed SSD-KD with existing top-performing D-KD methods including DeepInv [30], CMI [10], DAFL [4], DFQ [7], ZSKT [20], Fast5 and Fast10 [11]. In order to get improved student model performance, we relax the data scale of synthetic samples in our SSD-KD to be similar to that for Fast5. Experimental results are summarized in Table 2. Compared to these mainstream D-KD methods, our SSD-KD gets very competitive performance in student model accuracy, and significantly better performance in overall training efficiency, on all 5 teacher-student model pairs.", + "bbox": [ + 75, + 670, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In light of the results in Table 1, 2, it can be seen that our SSD-KD makes the overall training efficiency one or two orders of magnitude faster than most D-KD methods while retaining competitive student model performance.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/764582cddbb4a91d1e6abbc095f47ebc6138d04d894ceb4175f93c11df55f316.jpg", + "table_caption": [ + "Table 3. Performance comparison on the NYUv2 dataset. The teacher model is pre-trained on the ImageNet dataset and finetuned on the NYUv2 dataset, and the student model is trained from scratch. The results of reference methods are collected from [11]." + ], + "table_footnote": [], + "table_body": "
MethodTraining Data ScalemIoU (%)
Teacher: Deeplabv3-ResNet501,449 NYUv20.519
Student: Deeplabv3-Mobilenetv21,449 NYUv20.375
KD [12]1,449 NYUv20.380
DFND [5]14M (ImageNet)0.378
DFAD [9]960K (synthetic)0.364
DAFL [4]960K (synthetic)0.105
Fast10 [11]17K (synthetic)0.366
SSD-KD16K (synthetic)0.384
", + "bbox": [ + 508, + 502, + 879, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on semantic segmentation task. To further validate the generalization ability of our method, we compare the performance of SSD-KD with existing D-KD methods on the NYUv2 dataset. Table 3 summarizes the results. We can see that our SSD-KD not only achieves state-of-the-art performance in terms of model accuracy, but also is significantly more efficient than other D-KD methods in terms of the training data scale. The overall training time cost for our SSD-KD and Fast10 is 8.9 hours and 9.5 hours, respectively. Besides, the student model trained by SSD-KD even outperforms the baseline model trained on the original training data, as NYUv2 is a small-scale dataset.", + "bbox": [ + 496, + 626, + 890, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 816, + 663, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of the synthetic data scale. So far, we have already demonstrated that SSD-KD has the appealing capability to attain efficient and effective end-to-end distillation training using a small amount of synthetic samples. To bet-", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6014", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fe131909012347caf67092652fbfd4f1caf4b0101025930b6ce32e24131f577.jpg", + "image_caption": [ + "Figure 4. Performance comparison of SSD-KD under different synthetic data scales against the original training dataset size, in terms of top-1 classification accuracy $(\\%)$ and overall training time cost (hour)." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 277, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c9704e1d98650e5f15730a4013d278788f6eaa4fc5a2214531fa9a3dad6e6b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 87, + 480, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/898ebb0651ce575f43792e1a95ed211c7c757e0408b20edd57290557101d66a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 85, + 687, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf55f48331c8c1e62490291f08d16b98e3030536f036ffeac314115379f670b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 85, + 893, + 199 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/070bc5100ff68b804667b0def2d96b9eb50a38cc19f5f5f2887d2060748995a6.jpg", + "table_caption": [ + "Table 4. Effect of two core modules in SSD-KD. PS: the priority sampling function, Difficulty|Diversity: the difficulty|diversity-aware balancing term in the modulating function, see Eq. (3)." + ], + "table_footnote": [], + "table_body": "
DatasetBaselinePSDifficultyDiversityResNet-34 ResNet-18WRN40-2 WRN16-1
CIFAR-10✓87.35(1.19h)80.47(0.84h)
✓✓90.28(1.01h)87.19(0.69h)
✓✓88.57(1.20h)83.12(0.86h)
✓✓89.12(1.21h)84.31(0.87h)
✓✓✓89.95(1.21h)85.79(0.88h)
✓✓✓✓92.92(1.11h)88.07(0.77h)
CIFAR-100✓64.05(1.19h)41.67(0.84h)
✓✓71.70(1.01h)49.70(0.69h)
✓✓66.94(1.20h)43.96(0.86h)
✓✓68.52(1.21h)46.34(0.87h)
✓✓✓70.85(1.21h)47.82(0.88h)
✓✓✓✓73.05(1.12h)51.96(0.77h)
", + "bbox": [ + 259, + 276, + 707, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5d5d534d99b2af5096176bece819beed0552b8ab90d64915bec0d9892f33e6f.jpg", + "image_caption": [ + "Figure 5. Visualization examples of synthetic image samples generated by Fast10 and our SSD-KD for the NYUv2 dataset." + ], + "image_footnote": [], + "bbox": [ + 112, + 479, + 436, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ter explore the boundary of this capability, we conduct an ablation with three teacher-student model pairs, including ResNet34 $\\rightarrow$ ResNet18, VGG11 $\\rightarrow$ ResNet18, and WRN40-2 $\\rightarrow$ WRN16-1. In the experiments, we decrease the synthetic data scale used in SSD-KD [11] from 50,000(100% relative to the original training data size) to {10000(20%), 5000(10%), 2500(%5), 500(1%)} for the CIFAR-10 and CIFAR-100 datasets. As shown in Fig. 4, the accuracy of the student model trained by SSD-KD remains stable across a relatively large synthetic data scale range, for all teacher-student model pairs. When decreasing the synthetic data scale, the overall training time cost appears to decrease nearly linearly, while the student model accuracy drop is mild (less than 10% even for the synthetic data scale 500).", + "bbox": [ + 75, + 657, + 468, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the core modules. In Table 4, we provide an ablation to scrutinize our SSD-KD systematically. We", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "observe that: (1) the two modulating functions (consisting of a diversity-aware term and a difficulty-aware term, see Eq. (3)) and the priority sampling function, are both critical to our SSD-KD; (2) the combination of them strikes a good tradeoff between model accuracy and training efficiency.", + "bbox": [ + 496, + 479, + 890, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of data inversion. Fig. 5 shows examples of synthetic images inverted by Fast10 and our SSD-KD for the NYUv2 dataset. Compared to Fast10, our method can better invert texture information and has less noise.", + "bbox": [ + 496, + 556, + 890, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 628, + 619, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented SSD-KD, the first fully efficient method to advance adversarial data-free knowledge distillation research. Benefiting from a small-scale data inversion and sampling mechanism based on a modulating function and a priority sampling function, SSD-KD can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, attaining efficient and effective data-free knowledge distillation. Extensive experiments on image classification and semantic segmentation benchmarks validate the efficacy of SSD-KD. We hope our work can inspire future research on efficient D-KD designs.", + "bbox": [ + 496, + 648, + 890, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 844, + 679, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Fund for Key International Collaboration (62120106005).", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6015", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. NeurIPS, 2017. 3", + "[2] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 1", + "[3] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. arXiv preprint arXiv:1905.07072, 2019. 1", + "[4] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 1, 3, 4, 6, 7", + "[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In CVPR, 2021. 7", + "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6", + "[7] Yoojin Choi, Jihwan Choi, Mostafa El-Khamy, and Jungwon Lee. Data-free network quantization with adversarial knowledge distillation. In CVPR Workshops, 2020. 1, 2, 3, 4, 6, 7", + "[8] Kien Do, Thai Hung Le, Dung Nguyen, Dang Nguyen, Haripriya Harikumar, Truyen Tran, Santu Rana, and Svetha Venkatesh. Momentum adversarial distillation: Handling large distribution shifts in data-free knowledge distillation. NeurIPS, 2022. 1, 3", + "[9] Gongfan Fang, Jie Song, Chengchao Shen, Xinchao Wang, Da Chen, and Mingli Song. Data-free adversarial distillation. arXiv preprint arXiv:1912.11006, 2019. 7", + "[10] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. arXiv preprint arXiv:2105.08584, 2021. 1, 2, 3, 4, 6, 7", + "[11] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In AAAI, 2022. 1, 2, 3, 4, 6, 7, 8", + "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 7", + "[13] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 3", + "[14] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Tech Report, 2009. 6", + "[15] Yuhang Li, Feng Zhu, Ruihao Gong, Mingzhu Shen, Xin Dong, Fengwei Yu, Shaoqing Lu, and Shi Gu. Mixmix: All you need for data-free compression are feature and data mixing. In ICCV, 2021. 2, 3", + "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 4" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured knowledge distillation for semantic segmentation. In CVPR, 2019. 1", + "[18] Raphael Gontijo Lopes, Stefano Fenu, and Thad Starner. Data-free knowledge distillation for deep neural networks. arXiv preprint arXiv:1710.07535, 2017. 1, 3", + "[19] Liangchen Luo, Mark Sandler, Zi Lin, Andrey Zhmoginov, and Andrew Howard. Large-scale generative data-free distillation. arXiv preprint arXiv:2012.05578, 2020. 3", + "[20] Paul Micaelli and Amos J Storkey. Zero-shot knowledge transfer via adversarial belief matching. NeurIPS, 2019. 4, 6, 7", + "[21] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In ICML, 2019. 1, 3", + "[22] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 1", + "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6", + "[24] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In CVPR, 2023. 1, 3", + "[25] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1", + "[26] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. 3, 5", + "[27] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 4", + "[28] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6", + "[29] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 3", + "[30] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2, 3, 4, 6, 7", + "[31] Jaemin Yoo, Minyong Cho, Taebum Kim, and U Kang. Knowledge extraction with no observable data. NeurIPS, 2019. 1, 3", + "[32] Shikang Yu, Jiachen Chen, Hu Han, and Shuqiang Jiang. Data-free knowledge distillation via feature exchange and activation region constraint. In CVPR, 2023. 3", + "[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6016", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_model.json b/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f70146b14e590c1c12f3e35f4428aaa4fa6765c7 --- /dev/null +++ b/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_model.json @@ -0,0 +1,1857 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.254, + 0.131, + 0.717, + 0.154 + ], + "angle": 0, + "content": "Small Scale Data-Free Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.18, + 0.777, + 0.201 + ], + "angle": 0, + "content": "He Liu\\(^{1,2*}\\), Yikai Wang\\(^{2*}\\), Huaping Liu\\(^{2}\\), Fuchun Sun\\(^{2}\\), Anbang Yao\\(^{3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.205, + 0.662, + 0.223 + ], + "angle": 0, + "content": "1China Telecom Cloud Technology Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.224, + 0.767, + 0.242 + ], + "angle": 0, + "content": "\\(^{2}\\)BNRist Center, State Key Lab on Intelligent Technology and Systems," + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.242, + 0.849, + 0.26 + ], + "angle": 0, + "content": "Department of Computer Science and Technology, Tsinghua University \\(^{3}\\)Intel Labs China" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.262, + 0.825, + 0.277 + ], + "angle": 0, + "content": "{liuhel7@mails.,yikaiw@,hpliu@,fuchuns@}tsinghua.edu.cn，anbang.yao@intel.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.326, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.473, + 0.83 + ], + "angle": 0, + "content": "Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data, avoiding privacy, security, and proprietary risks in real applications. In this line of research, existing methods typically follow an inversion-and-distillation paradigm in which a generative adversarial network on-the-fly trained with the guidance of the pre-trained teacher network is used to synthesize a large-scale sample set for knowledge distillation. In this paper, we reexamine this common data-free knowledge distillation paradigm, showing that there is considerable room to improve the overall training efficiency through a lens of \"small-scale inverted data for knowledge distillation\". In light of three empirical observations indicating the importance of how to balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, we propose Small Scale Data-free Knowledge Distillation (SSD-KD). In formulation, SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples, facilitated by a dynamic replay buffer and a reinforcement learning strategy. As a result, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (e.g., \\(10 \\times\\) less than the original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream methods while retaining superior or competitive model performance, as demonstrated on popular image classification and semantic segmentation benchmarks. The code is available at https://github.com/OSVAI/SSD-KD." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.311, + 0.633, + 0.327 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.552 + ], + "angle": 0, + "content": "For computer vision applications on resource-constrained devices, how to learn portable neural networks yet with satisfied prediction accuracy is the key problem. Knowledge distillation (KD) [2, 12, 17, 22, 25, 33], which leverages the information of a pre-trained large teacher network to promote the training of a smaller target student network on the same training data, has become a mainstream solution. Conventional KD methods assume that the original training data is always available. However, accessing the source dataset on which the teacher network was trained is usually not feasible in practice, due to its potential privacy or security or proprietary or huge-size concerns. To relax the constraint on training data, knowledge distillation under a data-free regime has recently attracted increasing attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.896, + 0.903 + ], + "angle": 0, + "content": "The basic idea of Data-free Knowledge Distillation (D-KD) [3, 18, 21] is to construct synthetic samples for knowledge distillation conditioned on the pre-trained teacher network, which would match the underlying distribution of the original training data. Existing top-performing D-KD methods [4, 7, 8, 10, 24, 30, 31] generally adopt an adversarial inversion-and-distillation paradigm. Under this paradigm, (1) during the inversion process, a generator is trained by taking the pre-trained teacher network as the discriminator; (2) during the subsequent knowledge distillation process, the on-the-fly learned generator will synthesize pseudo samples for training the student network. However, adversarial D-KD methods usually require generating a large number of synthetic samples (compared to the original training dataset size) in order to guarantee trustworthy knowledge distillation. This poses a heavy burden on training resource consumption, suppressing their use in real applications. In the recent work of [11], the authors present an effective meta-learning strategy that seeks common features and reuses them as initial priors to reduce the number of iteration steps required to reach the convergence of the generator. Although faster data synthesis can be attained, [11] still needs to generate a sufficiently large number of syn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.903 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author. This work was done when He Liu was an intern at Intel Labs China, supervised by Anbang Yao who conceived the project." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6008" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.09, + 0.368, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.09, + 0.599, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.091, + 0.828, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.214, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Figure 1. Comparison of knowledge distillation (KD) using original samples vs. synthetic samples, under the same training data scale: 5000 samples (10% of the CIFAR-10 training dataset size). In such small-scale KD regime, the student models (DeepInv [30] and SSD-KD) trained on synthetic samples always show much better accuracy than the counterparts (Vanilla KD [12]) trained on original samples." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.266, + 0.278, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.267, + 0.482, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.385, + 0.349, + 0.398 + ], + "angle": 0, + "content": "(a) Difficulty distribution." + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.264, + 0.688, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.264, + 0.889, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.385, + 0.756, + 0.398 + ], + "angle": 0, + "content": "(b) Diversity distribution." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.401, + 0.893, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Comparison of synthetic sample distributions collected from two top-performing adversarial D-KD methods (DeepInv [30] and Fast10 [11]) and our SSD-KD on CIFAR-100 dataset. Comparatively, Fig. 2a shows that our method can better balance the difficulty distribution of synthetic samples while encouraging the generator to invert more hard samples, and Fig. 2b further shows that our method can better balance the diversity distribution of synthetic samples across different categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.47, + 0.577 + ], + "angle": 0, + "content": "thetic samples to ensure effective knowledge distillation, neglecting the efficiency of the following knowledge distillation process which will become the major bottleneck to the overall training efficiency. In a nutshell, there is no research effort made to improve the overall training efficiency of D-KD via jointly considering data inversion and knowledge distillation processes, to the best of our knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To remedy this critical gap, this paper presents the first fully efficient D-KD approach termed Small Scale Data-free Knowledge Distillation (SSD-KD). Our SSD-KD improves the overall training efficiency of the adversarial inversion-and-distillation paradigm from a novel \"small data scale\" perspective. In this work, \"data scale\" refers to the total number of inverted samples used in knowledge distillation during a training epoch. The formulation of SSD-KD is inspired by three empirical observations. On different pairs of teacher-student networks, we first observe that the student networks trained on synthetic samples tend to show much better performance than their corresponding counterparts trained on original samples when significantly reducing data scales of synthetic samples and original samples to the same (e.g., \\(10\\%\\) of the source training dataset size), as illustrated in Fig. 1. Note that synthetic samples are generated with the guidance of the teacher network pre-trained on the whole source dataset, which naturally reflect different views of the original data distribution. Under a small enough data scale, this makes synthetic samples have superior capability to original samples in fitting the underlying" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.699 + ], + "angle": 0, + "content": "distribution of the whole source dataset. This inspiring observation indicates that if we can construct a small-scale set of high-quality synthetic samples, a promising way toward fully efficient D-KD would be created. In principle, we believe a high-quality small-scale synthetic dataset should have well-balanced class distributions in terms of both synthetic sample diversity and difficulty. However, our other two observations indicate that existing D-KD methods including both conventional and the most efficient designs do not have good capabilities to balance the aforementioned two class distributions of synthetic samples under a small data scale, as illustrated in Fig. 2. Note that there already exist a few D-KD methods to enhance the diversity of synthetic samples [7, 10, 15, 30], but the diversity of synthetic samples in terms of sample difficulty is not explored yet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Driven by the above observations and analysis, we come up with our SSD-KD which introduces two interdependent modules to significantly accelerate the overall training efficiency of the predominant adversarial inversion-and-distillation paradigm. The first module of SSD-KD relies on a novel modulating function that defines a diversity-aware term and a difficulty-aware term to jointly balance the class distributions of synthetic samples during both data synthesis and knowledge distillation processes in an explicit manner. The second module defines a novel priority sampling function facilitated by a reinforcement learning strategy that selects a small portion of proper synthetic samples from candidates stored in a dynamic replay buffer for knowledge" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6009" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "distillation, further improving the end-to-end training efficiency. Benefiting from these two modules, SSD-KD has two appealing merits. On one side, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (\\(10 \\times\\) less than original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream D-KD methods while retaining competitive model performance. On the other side, SSD-KD attains largely improved performance in student model accuracy and maintains overall training efficiency when relaxing the data scale of synthetic samples to a relatively large number (which is still smaller than those for existing D-KD methods). We validate these merits of our method by lots of experiments on popular image classification and semantic segmentation benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.332, + 0.22, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.473, + 0.718 + ], + "angle": 0, + "content": "Data-free knowledge distillation. D-KD is originally explored by [18] which assumes that layer-wise activation records of a well-trained teacher network pre-computed on training samples is available. The prevailing inversion-and-distillation paradigm for D-KD without reliance on the original training data or the recorded metadata is introduced in [21]. In the inversion process, synthetic training samples are generated by leveraging the information learned by the teacher network, whose distribution is expected to fit the underlying distribution of the original training dataset. In the distillation process, the student network is trained on synthetic samples by forcing it to match the predictions of the teacher network. Based on this paradigm, [4, 31] use generative adversarial networks for data inversion. Subsequent D-KD methods mostly follow this adversarial inversion-and-distillation paradigm. They try to improve the data inversion process from different aspects, such as enhancing synthetic sample discrimination with contrastive learning [10] or an ensemble of generators [19], combating distribution shift with momentum adversarial learning [8] or meta learning [24], and promoting adversarial learning with data augmentation [15, 32]. Our method intends to improve the overall training efficiency of the adversarial D-KD, and differs from these methods both in focus and formulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Efficient synthetic data sampling. How to select proper synthetic samples for knowledge distillation is essential in D-KD research. Existing methods [8, 10, 11, 24] commonly rely on a memory bank to store synthetic samples, and directly update synthetic samples without considering the efficiency of the following knowledge distillation process. In sharp contrast, our method introduces a reinforcement learning strategy that adaptively selects appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty, significantly improving knowledge dis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "tillation efficiency. As far as we know, our method made the first attempt to extend reinforcement learning methodology [1, 13, 26] to address data-free knowledge distillation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.149, + 0.593, + 0.165 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.174, + 0.698, + 0.189 + ], + "angle": 0, + "content": "3.1. Preliminaries: D-KD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Let \\( f_{t}(\\cdot ;\\theta_{t}) \\) be a teacher model pre-trained on the original task dataset that is no longer accessible, the goal of D-KD is to first construct a set of synthetic training samples \\( x \\) via inverting the data distribution information learned by the teacher model, on which a target student model \\( f_{s}(\\cdot ;\\theta_{s}) \\) then can be trained by forcing it to mimic the teacher's function. Existing D-KD methods mostly use a generative adversarial network \\( g(\\cdot ;\\theta_{g}) \\) for producing synthetic training samples \\( x = g(z;\\theta_g) \\) from the latent noise input \\( z \\), which is trained by taking the teacher model as the discriminator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.893, + 0.53 + ], + "angle": 0, + "content": "The optimization of D-KD contains a common distillation regularization to minimize the teacher-student function discrepancy \\(\\mathcal{L}_{\\mathrm{KD}}(x) = D_{\\mathrm{KD}}(f_t(x;\\theta_t)\\| f_s(x;\\theta_s))^1\\) based on the KL-divergence, and a task-oriented regularization \\(\\mathcal{L}_{\\mathrm{Task}}(x)\\), e.g., the cross-entropy loss using the teacher's predication as the ground truth. Besides, since D-KD is primarily based on the assumption that the teacher model has been optimized to be capable of capturing the source training data distribution after pre-training, recent D-KD methods [7, 29, 30] introduce an extra loss to regularize the statistics (Batch-Normalization (BN) parameters) of the training data distribution during the data inversion process," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.538, + 0.892, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B N}} (x) = \\sum_ {l} \\left| \\left| \\mu_ {l} (x) - \\mathbb {E} (\\mu_ {l}) \\right| \\right| _ {2} + \\left| \\left| \\sigma_ {l} ^ {2} (x) - \\mathbb {E} (\\sigma_ {l} ^ {2}) \\right| \\right| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.639 + ], + "angle": 0, + "content": "where \\(\\mu_l(\\cdot)\\) and \\(\\sigma_l(\\cdot)\\) denote the batch-wise mean and variance estimates of feature maps at the \\(l\\)-th layer, respectively; \\(\\mathbb{E}(\\cdot)\\) over the BN statistics can be approximately substituted by running mean or variance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.805 + ], + "angle": 0, + "content": "The effectiveness of D-KD heavily depends on the quality of synthetic samples which are inverted from leveraging the knowledge of the pre-trained teacher model. The prevailing adversarial D-KD paradigm consists of two processes, namely data inversion and knowledge distillation. From both perspectives of efficiency and efficacy, on the one hand, the data inversion process largely affects the optimization performance of the student model; on the other hand, the training time cost of knowledge distillation emerges as a significant constraint to the overall training efficiency of D-KD." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.704, + 0.83 + ], + "angle": 0, + "content": "3.2. Our Design: SSD-KD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.837, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Regarding our SSD-KD, we focuses on improving the adversarial D-KD paradigm through a lens of \"small-scale" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{1} \\) Concretely, \\( {f}_{t}\\left( {x;{\\theta }_{t}}\\right) \\) and \\( {f}_{s}\\left( {x;{\\theta }_{s}}\\right) \\) are the outputted logits (before softmax) of the teacher model and the student model, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.087, + 0.29, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.087, + 0.882, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.894, + 0.285 + ], + "angle": 0, + "content": "Figure 3. Comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] (left) and more efficient family [10, 11] (middle), and our SSD-KD (right). Our SSD-KD formulates a reinforcement learning strategy that can flexibly seek appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty distributions. See the method section for notation definitions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.473, + 0.402 + ], + "angle": 0, + "content": "inverted data for knowledge distillation\". SSD-KD lays emphasis on instructing the data inversion process with the feedback of both the pre-trained teacher and the knowledge distillation process, significantly accelerating the overall training efficiency. Following the notations in the prevision subsection, the optimization goal of our SSD-KD is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.415, + 0.47, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\min _ {f _ {s}} \\max _ {g} \\mathbb {E} _ {x = \\delta \\circ g (z)} \\left(\\mathcal {L} _ {\\mathrm {B N}} (x) + \\mathcal {L} _ {\\mathrm {K D}} (x) + \\phi (x) \\mathcal {L} _ {\\text {T a s k}} (x)\\right), \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.449, + 0.127, + 0.461 + ], + "angle": 0, + "content": "where," + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.463, + 0.47, + 0.522 + ], + "angle": 0, + "content": "- We adopt a diversity-aware modulating function \\(\\phi(x)\\) that allocates each synthetic sample with a different priority regarding its predicted category by the teacher model, as presented in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.524, + 0.47, + 0.582 + ], + "angle": 0, + "content": "- Under the constraint of BN estimates, with \\(\\phi(x)\\), we encourage the generator to explore as tough synthetic samples (w.r.t. the teacher model) as possible, as introduced in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.584, + 0.47, + 0.673 + ], + "angle": 0, + "content": "- Instead of applying a random sampling strategy to select samples for knowledge distillation, we adopt a reweighting strategy to control the sampling process. We abuse notation slightly with \\(\\circ\\) to represent applying the strategy based on priority sampling function \\(\\delta\\), with more details in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.469, + 0.719 + ], + "angle": 0, + "content": "- Each synthetic sample is not only prioritized by its modulating function \\(\\phi(x)\\) but also is reweighted at the sampling stage that reuses the same intermediate values as \\(\\phi(x)\\)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.463, + 0.47, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Although the D-KD pipeline allows training samples to be synthesized and served for training the student model on the same task. However, there is a large extent of data redundancy that hinders the training efficiency of D-KD methods. In the following sections, we detail our SSD-KD, a fully efficient D-KD method that is capable of using an extremely small scale of synthetic data yet achieving competitive performance compared to existing D-KD methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The pipeline of SSD-KD is summarized in Alg. 1 and the comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] and more efficient family [10, 11], and our" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.297, + 0.691, + 0.312 + ], + "angle": 0, + "content": "SSD-KD is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.868, + 0.336 + ], + "angle": 0, + "content": "3.3. Data Inversion with Distribution Balancing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.599 + ], + "angle": 0, + "content": "We provide Fig. 2 to demonstrate the data redundancy of D-KD that results from a large imbalance of the synthetic data. The left two sub-figures of Fig. 2 depict the distribution of categories predicted by the teacher model, indicating a significant imbalance in data categories. The right two sub-figures of Fig. 2 show sample accounts of different bars that correspond to different prediction difficulties (the difficulty is measured by the predicted probability by the teacher model). For D-KD, it indicates that generating samples with only the instruction of teacher-student discrepancy results in a sharp distribution over the sample difficulty and tends to obtain easily predicted data samples. We argue that for the D-KD task where data samples are all synthesized, the data-generating process needs to consider both the teacher-student discrepancy and the pre-trained knowledge of the teacher itself, by which our SSD-KD proposes diversity-aware and difficulty-aware data synthesis, as detailed below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Diversity-aware balancing. We first propose to address the issue of imbalanced sample difficulty in the data inversion process. Specifically, we maintain a replay buffer \\(\\mathcal{B}\\) that stores a constant amount (denoted as \\(|\\mathcal{B}|\\)) of synthetic data samples. For each data sample \\(x\\) in \\(\\mathcal{B}\\), we penalize its total amount of samples that share the same predicted category (by the teacher model) with \\(x\\). To realize this, we adopt a diversity-aware balancing term that encourages the generator to synthesize samples with infrequent categories, which will be shown in Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Difficulty-aware balancing. Drawing inspiration from the field of object detection that utilizes focal loss for largely imbalanced samples [16, 27], for each sample \\( x \\), we further introduce a difficulty-aware balancing term on the predicted probability \\( p_T(x) \\). Here, difficult synthetic samples are considered as those with low-confidence predictions by the teacher model, which are encouraged by the difficulty-aware balancing term as will be given in Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In summary, we introduce a modulating function \\(\\phi (x)\\) to adjust the optimization of the generator based on the pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6011" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.184 + ], + "angle": 0, + "content": "diction feedback from the pre-trained teacher model. \\(\\phi (x)\\) is expected to balance the category distribution and dynamically distinguish between easy and hard synthetic samples, by which easy samples no longer overwhelm the distillation process. Formally, for a synthetic data sample \\(x\\in \\mathcal{B}\\), its modulating function \\(\\phi (x)\\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.19, + 0.47, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\phi (x) = \\underbrace {\\left(1 - \\frac {1}{| \\mathcal {B} |} \\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\mathbb {I} _ {c _ {T} \\left(x ^ {\\prime}\\right) = c _ {T} (x)}\\right)} _ {\\text {d i v e r s i t y - a w a r e b a l a n c i n g}} \\underbrace {\\left(1 - p _ {T} (x)\\right) ^ {\\gamma}} _ {\\text {d i f f i c u l t y - a w a r e b a l a n c i n g}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.34 + ], + "angle": 0, + "content": "where \\( c_{T}(x) \\) and \\( p_{T}(x) \\) refer to the predicted category index and probability (or confidence) by the pre-trained teacher model, respectively; \\( \\mathbb{I}_{c_T(x') = c_T(x)} \\) denotes an indicator function that equals to 1 if the predicted category of \\( x' \\) is the same with that of \\( x \\), while otherwise 0; \\( \\gamma \\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.492 + ], + "angle": 0, + "content": "We highlight two properties of the modulating function \\(\\phi(x)\\). Firstly, for the data sample \\(x\\) with a high prediction certainty w.r.t. the teacher model (i.e., considered as an easy sample), \\(\\phi(x)\\) approaches to a low value and thus results in low impact on the task-oriented loss \\(\\mathcal{L}_{\\mathrm{Task}}(x)\\) in Eq. (2). Secondly, when the category distribution of synthetic data samples in \\(\\mathcal{B}\\) is largely imbalanced as predicted by the teacher network, the sample \\(x\\) of which the category shares with more samples in \\(\\mathcal{B}\\) is penalized and thus the corresponding \\(\\mathcal{L}_{\\mathrm{Task}}(x)\\) is weakened by \\(\\phi(x)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Although Eq. (3) implies that the value of the modulating function \\(\\phi(x)\\) is partially determined by the current replay buffer \\(\\mathcal{B}\\), note that \\(\\mathcal{B}\\) changes dynamically and is also affected by \\(\\phi(x)\\). This is due to that the term \\(\\phi(x)\\mathcal{L}_{\\mathrm{Task}}(x)\\) in Eq. (2) directly optimizes the generator that synthesizes data samples to compose \\(\\mathcal{B}\\). In this sense, the balancing in category diversity is maintained during training given the mutual effects of \\(\\mathcal{B}\\) and \\(\\phi(x)\\). We find it particularly important in the data inversion process for balancing categories of data samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.689 + ], + "angle": 0, + "content": "With both balancing terms, as shown in Fig. 2, our method (SSD-KD) generates mild sample distributions in terms of both sample category and difficulty." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.38, + 0.714 + ], + "angle": 0, + "content": "3.4. Distillation with Priority Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.871 + ], + "angle": 0, + "content": "The original prioritized experience replay method [26] reuses important transitions more frequently and learns more efficiently. Differently, rather than obtaining the reward from the environment, our prioritized sampling method is designed to fit in data-free knowledge distillation and get feedback from the framework itself. In other words, the prioritized sampling method performs the opposite role of the previous data-free knowledge distillation methods: it focuses on training a sparse set of highly prioritized samples instead of uniform sampling to speed up training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "By Eq. (3), we sample the synthetic data \\( x \\) from the current replay buffer \\( \\mathcal{B} \\). Instead of uniformly sampling \\( x \\), we" + }, + { + "type": "code_caption", + "bbox": [ + 0.511, + 0.095, + 0.669, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: SSD-KD" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.112, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Input: Number of training epochs \\(E\\); number of iterations for data inversion \\(T_{g}\\); number of iterations for distillation \\(T_{kd}\\); pre-trained teacher model \\(f_{t}(\\cdot;\\theta_{t})\\); student network \\(f_{s}(\\cdot;\\theta_{s})\\) \nOutput: The optimized student model \\(f_{s}(\\cdot;\\theta_{s})\\) \n1 Initialize replay buffer \\(\\mathcal{B} \\gets \\emptyset\\) and \\(f_{s}(\\cdot;\\theta_{s})\\) \n2 for \\(e = 1\\) to \\(E\\) do \n3 (1) Data Inversion with Distribution Balancing: \n4 Initialize a generative network \\(g(\\cdot;\\theta_{g})\\) \n5 \\(z \\gets \\mathcal{N}(0,1)\\) \n6 for \\(i = 1\\) to \\(T_{g}\\) do \n7 \\(x \\gets g(z;\\theta_{g})\\) \n8 Compute the overall loss by Eq. (2) \n9 Update \\(z\\) and \\(\\theta_{g}\\) \n10 end \n11 (2) Distillation with Priority Sampling: \n12 Compute the importance-sampling weight \\(w_{i-1}(x)\\) \n13 Compute the sample priority \\(\\delta_{i}(x)\\) \n14 \\(\\mathcal{B} \\gets \\mathcal{B} \\cup \\{x,\\delta_{i}(x)\\}\\) \n15 Remove old samples to keep \\(|\\mathcal{B}|\\) a constant \n16 for \\(j = 1\\) to \\(T_{kd}\\) do \n17 Sample a small-scale mini-batch \\(M\\) from \\(\\mathcal{B}\\) \n18 Compute the distillation loss \\(\\mathcal{L}_{\\mathrm{KD}}(x)\\) in Eq. (2), where \\(x \\in M\\) \n19 Compute the importance-sampling weight \\(w_{i}(x)\\) \n20 Compute the sample priority \\(\\delta_{i+1}(x)\\) \n21 Update old samples \\(\\{x,\\delta_{i}(x)\\}\\) to \\(\\{x,\\delta_{i+1}(x)\\}\\) in \\(\\mathcal{B}\\) \n22 Update \\(\\theta_{s}\\) \n23 end \n24 end \n25 Return \\(f_{s}(\\cdot;\\theta_{s})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.741 + ], + "angle": 0, + "content": "propose to modulate the sampling probability by a sampling strategy termed Priority Sampling (PS). The basic function of PS is to measure the importance of each sample \\( x \\) in \\( \\mathcal{B} \\), by which we introduce a priority sampling function \\( \\delta_{i}(x) \\)," + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.752, + 0.892, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\delta_ {i} (x) = w _ {i - 1} (x) K L \\left(f _ {t} (x; \\theta_ {t}) \\mid \\mid f _ {s} \\left(x; \\theta_ {s}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where as mentioned in Sec. 3.1, \\(KL\\) denotes the KL-divergence between the softmax outputs of logits \\(f_{t}(x;\\theta_{t})\\) and \\(f_{s}(x;\\theta_{s})\\); \\(\\theta_t\\), \\(\\theta_s\\) depend on the training step \\(i\\); \\(w_{i}(x)\\) is the calibration term [26] for normalizing samples in \\(\\mathcal{B}\\), as will be formalized in Eq. (5), especially, when \\(i = 0\\), \\(w_{-1}(x) = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The training of knowledge distillation with random updates relies on those updates corresponding to the same dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6012" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.148 + ], + "angle": 0, + "content": "Table 1. Performance comparison of Fast2 [11] (the current most efficient D-KD method) and our SSD-KD, in terms of top-1 classification accuracy \\((\\%)\\) and overall training time cost (hours). Our SSD-KD performs with a very small training data scale: 5000 synthetic samples, i.e., \\(10\\%\\) of the original training dataset size. All image classification results in this and the other tables are averaged over three independent runs, and the methods \"Teacher\" and \"Student\" are performed on the whole original training dataset." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.153, + 0.77, + 0.284 + ], + "angle": 0, + "content": "
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
Fast2 [11]92.62(3.92h)84.67(3.02h)88.36(1.46h)89.56(2.37h)89.68(1.70h)
SSD-KD92.92(1.11h)89.56(0.77h)88.07(0.77h)90.12(0.96h)91.84(0.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
Fast2 [11]69.76(4.02h)62.83(2.99h)41.77(1.44h)53.15(2.26h)57.08(1.72h)
SSD-KD73.05(1.12h)66.66(0.78h)51.96(0.77h)59.60(0.97h)61.44(0.78h)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.471, + 0.381 + ], + "angle": 0, + "content": "distribution as its expectation. Prioritized sampling data introduces bias since it might change the data distribution, and affect the solution that the estimates will converge to. Thus we correct the bias by introducing an importance-sampling (IS) weight \\( w_{i}(x) \\) for the data sample \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.392, + 0.47, + 0.41 + ], + "angle": 0, + "content": "\\[\nw _ {i} (x) = \\left(N \\cdot P _ {i} (x)\\right) ^ {- \\beta}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.472, + 0.453 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyper-parameter; \\(P_{i}(x)\\) is the probability of sampling transition defined by" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.462, + 0.47, + 0.502 + ], + "angle": 0, + "content": "\\[\nP _ {i} (x) = \\frac {\\left(\\left| \\delta_ {i} (x) \\right| + \\epsilon\\right) ^ {\\alpha}}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\left(\\left| \\delta_ {i} \\left(x ^ {\\prime}\\right) \\right| + \\epsilon\\right) ^ {\\alpha}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.55 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a small positive constant which prevents the edge-case of transitions not being selected once their priority is zero." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.472, + 0.763 + ], + "angle": 0, + "content": "The priority sampling function \\(\\delta(x)\\) has two noteworthy properties. Firstly, as the delta value increases, \\(\\delta(x)\\) reflects a greater information discrepancy between the teacher and student models for the synthetic samples in the current \\(\\mathcal{B}\\). The student model should therefore be optimized from samples with greater information discrepancy, as this facilitates the faster acquisition of the teacher model. Secondly, \\(\\delta(x)\\) dynamically changes with each update iteration of the student and generative models. Consequently, when the student model acquires the teacher model's capabilities on certain samples, it continues to learn from samples with larger differences relative to the teacher model based on the new sample distribution. This further enhances the performance of the student model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.777, + 0.203, + 0.794 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.277, + 0.818 + ], + "angle": 0, + "content": "4.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.87 + ], + "angle": 0, + "content": "We conduct comprehensive experiments on image classification and semantic segmentation tasks to evaluate the effectiveness and study the design of our SSD-KD." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Datasets. For image classification experiments, we use CIFAR-10 and CIFAR-100 [14], two most popular datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.35 + ], + "angle": 0, + "content": "for D-KD research. For semantic segmentation experiments, we use the NYUv2 dataset [28]. On each dataset, the baseline models are trained with its standard data split." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.893, + 0.518 + ], + "angle": 0, + "content": "Training Setups. We follow the settings of [11] for basic experiments and comparisons. Regarding experiments on the CIFAR-10 and CIFAR-100 datasets, we use 5 different teacher-student model pairs having either the same type or different type network architectures (see Table 1, 2). Regarding experiments on the NYUv2 dataset, we use two Deeplabv3 models [6] as a teacher-student model pair (see Table 3). Unless otherwise stated, we always adopt the same basic settings as in [11] for experiments, including the number of training epochs, the optimizer, the weight decay, etc. Detailed settings are put in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.895, + 0.744 + ], + "angle": 0, + "content": "Evaluation Metrics. Besides comparing the student model accuracy (top-1 accuracy for image classification and mean Intersection over Union (IoU) for semantic segmentation), we also compare the overall training time cost of existing mainstream D-KD methods [4, 7, 10, 11, 20, 30] and our SSD-KD. For each teacher-student model pair, we meticulously record the total time cost (hours) for each run of the end-to-end training by all methods. In order to guarantee a fair comparison, all training speed assessments are performed on 1 NVIDIA V100 GPU using 12 cores of Intel Xeon Gold 6240R CPU. Our experiments are implemented with the PyTorch [23] library. For each teacher-student model pair, the experiment is conducted with three independent runs for all methods, and in comparison we report the averaged results, unless otherwise stated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.704, + 0.772 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Results on image classification task. As the main focus of our work is to improve the overall training efficiency of the adversarial data-free knowledge distillation paradigm, we first compare the proposed SSD-KD with the current most efficient D-KD method Fast2 [11] on the CIFAR-10 and CIFAR-100 datasets. From the results shown in Table 1, we can observe: (1) on the CIFAR-10 dataset, our SSD-KD (using a very small data scale: 5000 synthetic samples," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6013" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 2. Performance comparison of existing top-performing D-KD methods and our SSD-KD, in terms of top-1 classification accuracy \\((\\%)\\) and overall training time cost (hours). For our SSD-KD, we relax the data scale of synthetic samples to be similar to that for Fast5." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.125, + 0.83, + 0.421 + ], + "angle": 0, + "content": "
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
DeepInv [30]93.26(25.62h)90.36(13.19h)83.04(10.62h)86.85(13.84h)89.72(11.67h)
CMI [10]94.84(15.49h)91.13(10.99h)90.01(9.87h)92.78(11.57h)92.52(10.58h)
DAFL [4]92.22(303.50h)81.10(209.78h)65.71(96.57h)81.33(157.44h)81.55(118.52h)
DFQ [7]94.61(152.16h)90.84(104.93h)86.14(48.40h)91.69(78.81h)92.01(59.40h)
ZSKT [20]93.32(304.11h)89.46(209.65h)83.74(96.75h)86.07(157.51h)89.66(118.82h)
Fast5 [11]93.63(4.34h)89.94(3.15h)88.90(1.61h)92.04(2.44h)91.96(1.89h)
Fast10 [11]94.05(4.72h)90.53(3.22h)89.29(1.72h)92.51(2.48h)92.45(2.05h)
SSD-KD94.26(4.20h)90.67(2.95h)89.96(1.48h)93.23(2.30h)93.11(1.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
DeepInv [30]61.32(25.88h)54.13(13.13h)53.77(10.62h)61.33(13.86h)61.34(11.68h)
CMI [10]77.04(15.42h)70.56(11.28h)57.91(10.01h)68.88(11.57h)68.75(10.53h)
DAFL [4]74.47(303.54h)54.16(209.82h)20.88(96.62h)42.83(157.49h)43.70(118.67h)
DFQ [7]77.01(152.18h)66.21(105.14h)51.27(48.47h)54.43(78.83h)64.79(59.48h)
ZSKT [20]67.74(304.13h)54.31(209.63h)36.66(96.87h)53.60(157.68h)54.59(118.84h)
Fast5 [11]72.82(4.35h)65.28(3.06h)52.90(1.56h)61.80(2.35h)63.83(1.86h)
Fast10 [11]74.34(4.50h)67.44(3.12h)54.02(1.65h)63.91(2.42h)65.12(1.95h)
SSD-KD75.16(4.22h)68.77(2.94h)55.61(1.52h)64.57(2.27h)65.28(1.78h)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.67 + ], + "angle": 0, + "content": "i.e., \\(10\\%\\) of the original training dataset size) shows large improvements in training efficiency to Fast2 (at least \\(1.90\\times\\) and at most \\(3.92\\times\\) training speedup, on 5 teacher-student model pairs), and gets better student models on 4 out of 5 teacher-student model pairs; (2) on the more challenging CIFAR-100 dataset, our SSD-KD shows a very similar training speedup trend against Fast2 as on the CIFAR-10 dataset, but gets significantly better student models on all 5 teacher-student model pairs (at least \\(3.29\\%\\) and at most \\(10.19\\%\\) absolute top-1 accuracy gain to Fast2). The superior performance of SSD-KD against Fast2 validates the efficacy of our small-scale data inversion and sampling mechanism which can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Next, we compare the proposed SSD-KD with existing top-performing D-KD methods including DeepInv [30], CMI [10], DAFL [4], DFQ [7], ZSKT [20], Fast5 and Fast10 [11]. In order to get improved student model performance, we relax the data scale of synthetic samples in our SSD-KD to be similar to that for Fast5. Experimental results are summarized in Table 2. Compared to these mainstream D-KD methods, our SSD-KD gets very competitive performance in student model accuracy, and significantly better performance in overall training efficiency, on all 5 teacher-student model pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In light of the results in Table 1, 2, it can be seen that our SSD-KD makes the overall training efficiency one or two orders of magnitude faster than most D-KD methods while retaining competitive student model performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.441, + 0.892, + 0.497 + ], + "angle": 0, + "content": "Table 3. Performance comparison on the NYUv2 dataset. The teacher model is pre-trained on the ImageNet dataset and finetuned on the NYUv2 dataset, and the student model is trained from scratch. The results of reference methods are collected from [11]." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.503, + 0.88, + 0.617 + ], + "angle": 0, + "content": "
MethodTraining Data ScalemIoU (%)
Teacher: Deeplabv3-ResNet501,449 NYUv20.519
Student: Deeplabv3-Mobilenetv21,449 NYUv20.375
KD [12]1,449 NYUv20.380
DFND [5]14M (ImageNet)0.378
DFAD [9]960K (synthetic)0.364
DAFL [4]960K (synthetic)0.105
Fast10 [11]17K (synthetic)0.366
SSD-KD16K (synthetic)0.384
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Results on semantic segmentation task. To further validate the generalization ability of our method, we compare the performance of SSD-KD with existing D-KD methods on the NYUv2 dataset. Table 3 summarizes the results. We can see that our SSD-KD not only achieves state-of-the-art performance in terms of model accuracy, but also is significantly more efficient than other D-KD methods in terms of the training data scale. The overall training time cost for our SSD-KD and Fast10 is 8.9 hours and 9.5 hours, respectively. Besides, the student model trained by SSD-KD even outperforms the baseline model trained on the original training data, as NYUv2 is a small-scale dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.665, + 0.832 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Effect of the synthetic data scale. So far, we have already demonstrated that SSD-KD has the appealing capability to attain efficient and effective end-to-end distillation training using a small amount of synthetic samples. To bet-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6014" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.278, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.088, + 0.482, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.087, + 0.688, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.087, + 0.895, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.895, + 0.233 + ], + "angle": 0, + "content": "Figure 4. Performance comparison of SSD-KD under different synthetic data scales against the original training dataset size, in terms of top-1 classification accuracy \\((\\%)\\) and overall training time cost (hour)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.241, + 0.895, + 0.27 + ], + "angle": 0, + "content": "Table 4. Effect of two core modules in SSD-KD. PS: the priority sampling function, Difficulty|Diversity: the difficulty|diversity-aware balancing term in the modulating function, see Eq. (3)." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.277, + 0.709, + 0.457 + ], + "angle": 0, + "content": "
DatasetBaselinePSDifficultyDiversityResNet-34 ResNet-18WRN40-2 WRN16-1
CIFAR-10✓87.35(1.19h)80.47(0.84h)
✓✓90.28(1.01h)87.19(0.69h)
✓✓88.57(1.20h)83.12(0.86h)
✓✓89.12(1.21h)84.31(0.87h)
✓✓✓89.95(1.21h)85.79(0.88h)
✓✓✓✓92.92(1.11h)88.07(0.77h)
CIFAR-100✓64.05(1.19h)41.67(0.84h)
✓✓71.70(1.01h)49.70(0.69h)
✓✓66.94(1.20h)43.96(0.86h)
✓✓68.52(1.21h)46.34(0.87h)
✓✓✓70.85(1.21h)47.82(0.88h)
✓✓✓✓73.05(1.12h)51.96(0.77h)
" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.48, + 0.437, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.649 + ], + "angle": 0, + "content": "Figure 5. Visualization examples of synthetic image samples generated by Fast10 and our SSD-KD for the NYUv2 dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.869 + ], + "angle": 0, + "content": "ter explore the boundary of this capability, we conduct an ablation with three teacher-student model pairs, including ResNet34 \\(\\rightarrow\\) ResNet18, VGG11 \\(\\rightarrow\\) ResNet18, and WRN40-2 \\(\\rightarrow\\) WRN16-1. In the experiments, we decrease the synthetic data scale used in SSD-KD [11] from 50,000(100% relative to the original training data size) to {10000(20%), 5000(10%), 2500(%5), 500(1%)} for the CIFAR-10 and CIFAR-100 datasets. As shown in Fig. 4, the accuracy of the student model trained by SSD-KD remains stable across a relatively large synthetic data scale range, for all teacher-student model pairs. When decreasing the synthetic data scale, the overall training time cost appears to decrease nearly linearly, while the student model accuracy drop is mild (less than 10% even for the synthetic data scale 500)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Effect of the core modules. In Table 4, we provide an ablation to scrutinize our SSD-KD systematically. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.557 + ], + "angle": 0, + "content": "observe that: (1) the two modulating functions (consisting of a diversity-aware term and a difficulty-aware term, see Eq. (3)) and the priority sampling function, are both critical to our SSD-KD; (2) the combination of them strikes a good tradeoff between model accuracy and training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.616 + ], + "angle": 0, + "content": "Visualization of data inversion. Fig. 5 shows examples of synthetic images inverted by Fast10 and our SSD-KD for the NYUv2 dataset. Compared to Fast10, our method can better invert texture information and has less noise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.62, + 0.645 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.831 + ], + "angle": 0, + "content": "In this paper, we presented SSD-KD, the first fully efficient method to advance adversarial data-free knowledge distillation research. Benefiting from a small-scale data inversion and sampling mechanism based on a modulating function and a priority sampling function, SSD-KD can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, attaining efficient and effective data-free knowledge distillation. Extensive experiments on image classification and semantic segmentation benchmarks validate the efficacy of SSD-KD. We hope our work can inspire future research on efficient D-KD designs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.681, + 0.862 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.899 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Fund for Key International Collaboration (62120106005)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6015" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[2] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.468, + 0.242 + ], + "angle": 0, + "content": "[3] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. arXiv preprint arXiv:1905.07072, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.468, + 0.297 + ], + "angle": 0, + "content": "[4] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.339 + ], + "angle": 0, + "content": "[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.45 + ], + "angle": 0, + "content": "[7] Yoojin Choi, Jihwan Choi, Mostafa El-Khamy, and Jungwon Lee. Data-free network quantization with adversarial knowledge distillation. In CVPR Workshops, 2020. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.468, + 0.521 + ], + "angle": 0, + "content": "[8] Kien Do, Thai Hung Le, Dung Nguyen, Dang Nguyen, Haripriya Harikumar, Truyen Tran, Santu Rana, and Svetha Venkatesh. Momentum adversarial distillation: Handling large distribution shifts in data-free knowledge distillation. NeurIPS, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[9] Gongfan Fang, Jie Song, Chengchao Shen, Xinchao Wang, Da Chen, and Mingli Song. Data-free adversarial distillation. arXiv preprint arXiv:1912.11006, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[10] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. arXiv preprint arXiv:2105.08584, 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[11] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In AAAI, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.717 + ], + "angle": 0, + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[13] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[14] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Tech Report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[15] Yuhang Li, Feng Zhu, Ruihao Gong, Mingzhu Shen, Xin Dong, Fengwei Yu, Shaoqing Lu, and Shi Gu. Mixmix: All you need for data-free compression are feature and data mixing. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[17] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured knowledge distillation for semantic segmentation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[18] Raphael Gontijo Lopes, Stefano Fenu, and Thad Starner. Data-free knowledge distillation for deep neural networks. arXiv preprint arXiv:1710.07535, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[19] Liangchen Luo, Mark Sandler, Zi Lin, Andrey Zhmoginov, and Andrew Howard. Large-scale generative data-free distillation. arXiv preprint arXiv:2012.05578, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[20] Paul Micaelli and Amos J Storkey. Zero-shot knowledge transfer via adversarial belief matching. NeurIPS, 2019. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[21] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In ICML, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[22] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[24] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In CVPR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[25] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[26] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[27] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[28] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[29] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[30] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[31] Jaemin Yoo, Minyong Cho, Taebum Kim, and U Kang. Knowledge extraction with no observable data. NeurIPS, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[32] Shikang Yu, Jiachen Chen, Hu Han, and Shuqiang Jiang. Data-free knowledge distillation via feature exchange and activation region constraint. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6016" + } + ] +] \ No newline at end of file diff --git a/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_origin.pdf b/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6cdad7100a6cfd043f88f234e58a840147842c81 --- /dev/null +++ b/2024/Small Scale Data-Free Knowledge Distillation/9c1a5016-dfa0-4549-8999-c0090c1c7132_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c436caf1a825225ab01a9092287c59ba082cd9cd8337f46c16c087494552ae3f +size 895659 diff --git a/2024/Small Scale Data-Free Knowledge Distillation/full.md b/2024/Small Scale Data-Free Knowledge Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d01adb94a56e065318e5b98ba71163345c46b1ee --- /dev/null +++ b/2024/Small Scale Data-Free Knowledge Distillation/full.md @@ -0,0 +1,297 @@ +# Small Scale Data-Free Knowledge Distillation + +He Liu $^{1,2*}$ , Yikai Wang $^{2*}$ , Huaping Liu $^{2}$ , Fuchun Sun $^{2}$ , Anbang Yao $^{3\dagger}$ + +1China Telecom Cloud Technology Co., Ltd. + +$^{2}$ BNRist Center, State Key Lab on Intelligent Technology and Systems, + +Department of Computer Science and Technology, Tsinghua University $^{3}$ Intel Labs China + +{liuhel7@mails.,yikaiw@,hpliu@,fuchuns@}tsinghua.edu.cn，anbang.yao@intel.com + +# Abstract + +Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data, avoiding privacy, security, and proprietary risks in real applications. In this line of research, existing methods typically follow an inversion-and-distillation paradigm in which a generative adversarial network on-the-fly trained with the guidance of the pre-trained teacher network is used to synthesize a large-scale sample set for knowledge distillation. In this paper, we reexamine this common data-free knowledge distillation paradigm, showing that there is considerable room to improve the overall training efficiency through a lens of "small-scale inverted data for knowledge distillation". In light of three empirical observations indicating the importance of how to balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, we propose Small Scale Data-free Knowledge Distillation (SSD-KD). In formulation, SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples, facilitated by a dynamic replay buffer and a reinforcement learning strategy. As a result, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (e.g., $10 \times$ less than the original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream methods while retaining superior or competitive model performance, as demonstrated on popular image classification and semantic segmentation benchmarks. The code is available at https://github.com/OSVAI/SSD-KD. + +# 1. Introduction + +For computer vision applications on resource-constrained devices, how to learn portable neural networks yet with satisfied prediction accuracy is the key problem. Knowledge distillation (KD) [2, 12, 17, 22, 25, 33], which leverages the information of a pre-trained large teacher network to promote the training of a smaller target student network on the same training data, has become a mainstream solution. Conventional KD methods assume that the original training data is always available. However, accessing the source dataset on which the teacher network was trained is usually not feasible in practice, due to its potential privacy or security or proprietary or huge-size concerns. To relax the constraint on training data, knowledge distillation under a data-free regime has recently attracted increasing attention. + +The basic idea of Data-free Knowledge Distillation (D-KD) [3, 18, 21] is to construct synthetic samples for knowledge distillation conditioned on the pre-trained teacher network, which would match the underlying distribution of the original training data. Existing top-performing D-KD methods [4, 7, 8, 10, 24, 30, 31] generally adopt an adversarial inversion-and-distillation paradigm. Under this paradigm, (1) during the inversion process, a generator is trained by taking the pre-trained teacher network as the discriminator; (2) during the subsequent knowledge distillation process, the on-the-fly learned generator will synthesize pseudo samples for training the student network. However, adversarial D-KD methods usually require generating a large number of synthetic samples (compared to the original training dataset size) in order to guarantee trustworthy knowledge distillation. This poses a heavy burden on training resource consumption, suppressing their use in real applications. In the recent work of [11], the authors present an effective meta-learning strategy that seeks common features and reuses them as initial priors to reduce the number of iteration steps required to reach the convergence of the generator. Although faster data synthesis can be attained, [11] still needs to generate a sufficiently large number of syn + +![](images/b7f42998a74db30c7b8e815a1189712969ecef13a265d814d68391212031450a.jpg) +Figure 1. Comparison of knowledge distillation (KD) using original samples vs. synthetic samples, under the same training data scale: 5000 samples (10% of the CIFAR-10 training dataset size). In such small-scale KD regime, the student models (DeepInv [30] and SSD-KD) trained on synthetic samples always show much better accuracy than the counterparts (Vanilla KD [12]) trained on original samples. + +![](images/a9c240abb887b9a1ec2f64ca22ed33a7533dc2c8bd1ed8faa10f24cede29b72e.jpg) + +![](images/e18545f7d22b152170307a88c4da13443fd3a243b0468e8f156b885606daf5f5.jpg) + +![](images/fe139fdc34149eec03ccb445c88c9a599773d42a7ab2b60cd0f813f59566d066.jpg) +(a) Difficulty distribution. + +![](images/9b6198df7fadf4575acb67f2d97055cfc86844e36daf27473f20cf505ca95460.jpg) +Figure 2. Comparison of synthetic sample distributions collected from two top-performing adversarial D-KD methods (DeepInv [30] and Fast10 [11]) and our SSD-KD on CIFAR-100 dataset. Comparatively, Fig. 2a shows that our method can better balance the difficulty distribution of synthetic samples while encouraging the generator to invert more hard samples, and Fig. 2b further shows that our method can better balance the diversity distribution of synthetic samples across different categories. + +![](images/6607a92b3d6f91b26564e6bb5369f1fea0582bc266653b3d9ffdfc99ab03b42c.jpg) +(b) Diversity distribution. + +![](images/7448c10dad09d474c88480a5743946261a5faeceb2bd86f85d29f4ec14f9de9c.jpg) + +thetic samples to ensure effective knowledge distillation, neglecting the efficiency of the following knowledge distillation process which will become the major bottleneck to the overall training efficiency. In a nutshell, there is no research effort made to improve the overall training efficiency of D-KD via jointly considering data inversion and knowledge distillation processes, to the best of our knowledge. + +To remedy this critical gap, this paper presents the first fully efficient D-KD approach termed Small Scale Data-free Knowledge Distillation (SSD-KD). Our SSD-KD improves the overall training efficiency of the adversarial inversion-and-distillation paradigm from a novel "small data scale" perspective. In this work, "data scale" refers to the total number of inverted samples used in knowledge distillation during a training epoch. The formulation of SSD-KD is inspired by three empirical observations. On different pairs of teacher-student networks, we first observe that the student networks trained on synthetic samples tend to show much better performance than their corresponding counterparts trained on original samples when significantly reducing data scales of synthetic samples and original samples to the same (e.g., $10\%$ of the source training dataset size), as illustrated in Fig. 1. Note that synthetic samples are generated with the guidance of the teacher network pre-trained on the whole source dataset, which naturally reflect different views of the original data distribution. Under a small enough data scale, this makes synthetic samples have superior capability to original samples in fitting the underlying + +distribution of the whole source dataset. This inspiring observation indicates that if we can construct a small-scale set of high-quality synthetic samples, a promising way toward fully efficient D-KD would be created. In principle, we believe a high-quality small-scale synthetic dataset should have well-balanced class distributions in terms of both synthetic sample diversity and difficulty. However, our other two observations indicate that existing D-KD methods including both conventional and the most efficient designs do not have good capabilities to balance the aforementioned two class distributions of synthetic samples under a small data scale, as illustrated in Fig. 2. Note that there already exist a few D-KD methods to enhance the diversity of synthetic samples [7, 10, 15, 30], but the diversity of synthetic samples in terms of sample difficulty is not explored yet. + +Driven by the above observations and analysis, we come up with our SSD-KD which introduces two interdependent modules to significantly accelerate the overall training efficiency of the predominant adversarial inversion-and-distillation paradigm. The first module of SSD-KD relies on a novel modulating function that defines a diversity-aware term and a difficulty-aware term to jointly balance the class distributions of synthetic samples during both data synthesis and knowledge distillation processes in an explicit manner. The second module defines a novel priority sampling function facilitated by a reinforcement learning strategy that selects a small portion of proper synthetic samples from candidates stored in a dynamic replay buffer for knowledge + +distillation, further improving the end-to-end training efficiency. Benefiting from these two modules, SSD-KD has two appealing merits. On one side, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples ( $10 \times$ less than original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream D-KD methods while retaining competitive model performance. On the other side, SSD-KD attains largely improved performance in student model accuracy and maintains overall training efficiency when relaxing the data scale of synthetic samples to a relatively large number (which is still smaller than those for existing D-KD methods). We validate these merits of our method by lots of experiments on popular image classification and semantic segmentation benchmarks. + +# 2. Related Work + +Data-free knowledge distillation. D-KD is originally explored by [18] which assumes that layer-wise activation records of a well-trained teacher network pre-computed on training samples is available. The prevailing inversion-and-distillation paradigm for D-KD without reliance on the original training data or the recorded metadata is introduced in [21]. In the inversion process, synthetic training samples are generated by leveraging the information learned by the teacher network, whose distribution is expected to fit the underlying distribution of the original training dataset. In the distillation process, the student network is trained on synthetic samples by forcing it to match the predictions of the teacher network. Based on this paradigm, [4, 31] use generative adversarial networks for data inversion. Subsequent D-KD methods mostly follow this adversarial inversion-and-distillation paradigm. They try to improve the data inversion process from different aspects, such as enhancing synthetic sample discrimination with contrastive learning [10] or an ensemble of generators [19], combating distribution shift with momentum adversarial learning [8] or meta learning [24], and promoting adversarial learning with data augmentation [15, 32]. Our method intends to improve the overall training efficiency of the adversarial D-KD, and differs from these methods both in focus and formulation. + +Efficient synthetic data sampling. How to select proper synthetic samples for knowledge distillation is essential in D-KD research. Existing methods [8, 10, 11, 24] commonly rely on a memory bank to store synthetic samples, and directly update synthetic samples without considering the efficiency of the following knowledge distillation process. In sharp contrast, our method introduces a reinforcement learning strategy that adaptively selects appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty, significantly improving knowledge dis + +tillation efficiency. As far as we know, our method made the first attempt to extend reinforcement learning methodology [1, 13, 26] to address data-free knowledge distillation. + +# 3. Method + +# 3.1. Preliminaries: D-KD + +Let $f_{t}(\cdot ;\theta_{t})$ be a teacher model pre-trained on the original task dataset that is no longer accessible, the goal of D-KD is to first construct a set of synthetic training samples $x$ via inverting the data distribution information learned by the teacher model, on which a target student model $f_{s}(\cdot ;\theta_{s})$ then can be trained by forcing it to mimic the teacher's function. Existing D-KD methods mostly use a generative adversarial network $g(\cdot ;\theta_{g})$ for producing synthetic training samples $x = g(z;\theta_g)$ from the latent noise input $z$ , which is trained by taking the teacher model as the discriminator. + +The optimization of D-KD contains a common distillation regularization to minimize the teacher-student function discrepancy $\mathcal{L}_{\mathrm{KD}}(x) = D_{\mathrm{KD}}(f_t(x;\theta_t)\| f_s(x;\theta_s))^1$ based on the KL-divergence, and a task-oriented regularization $\mathcal{L}_{\mathrm{Task}}(x)$ , e.g., the cross-entropy loss using the teacher's predication as the ground truth. Besides, since D-KD is primarily based on the assumption that the teacher model has been optimized to be capable of capturing the source training data distribution after pre-training, recent D-KD methods [7, 29, 30] introduce an extra loss to regularize the statistics (Batch-Normalization (BN) parameters) of the training data distribution during the data inversion process, + +$$ +\mathcal {L} _ {\mathrm {B N}} (x) = \sum_ {l} \left| \left| \mu_ {l} (x) - \mathbb {E} (\mu_ {l}) \right| \right| _ {2} + \left| \left| \sigma_ {l} ^ {2} (x) - \mathbb {E} (\sigma_ {l} ^ {2}) \right| \right| _ {2}, \tag {1} +$$ + +where $\mu_l(\cdot)$ and $\sigma_l(\cdot)$ denote the batch-wise mean and variance estimates of feature maps at the $l$ -th layer, respectively; $\mathbb{E}(\cdot)$ over the BN statistics can be approximately substituted by running mean or variance. + +The effectiveness of D-KD heavily depends on the quality of synthetic samples which are inverted from leveraging the knowledge of the pre-trained teacher model. The prevailing adversarial D-KD paradigm consists of two processes, namely data inversion and knowledge distillation. From both perspectives of efficiency and efficacy, on the one hand, the data inversion process largely affects the optimization performance of the student model; on the other hand, the training time cost of knowledge distillation emerges as a significant constraint to the overall training efficiency of D-KD. + +# 3.2. Our Design: SSD-KD + +Regarding our SSD-KD, we focuses on improving the adversarial D-KD paradigm through a lens of "small-scale + +![](images/dc45cced476971ff7311f2e7e0ee4076b0abfcf0eb660e02de925fa8f62e45cf.jpg) +Figure 3. Comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] (left) and more efficient family [10, 11] (middle), and our SSD-KD (right). Our SSD-KD formulates a reinforcement learning strategy that can flexibly seek appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty distributions. See the method section for notation definitions. + +![](images/4dfee92936b674cac3945ad7aff790f96c23b78efcfe8e52c032a1b485a8ee8b.jpg) + +inverted data for knowledge distillation". SSD-KD lays emphasis on instructing the data inversion process with the feedback of both the pre-trained teacher and the knowledge distillation process, significantly accelerating the overall training efficiency. Following the notations in the prevision subsection, the optimization goal of our SSD-KD is defined as + +$$ +\min _ {f _ {s}} \max _ {g} \mathbb {E} _ {x = \delta \circ g (z)} \left(\mathcal {L} _ {\mathrm {B N}} (x) + \mathcal {L} _ {\mathrm {K D}} (x) + \phi (x) \mathcal {L} _ {\text {T a s k}} (x)\right), \tag {2} +$$ + +# where, + +- We adopt a diversity-aware modulating function $\phi(x)$ that allocates each synthetic sample with a different priority regarding its predicted category by the teacher model, as presented in Sec. 3.3. +- Under the constraint of BN estimates, with $\phi(x)$ , we encourage the generator to explore as tough synthetic samples (w.r.t. the teacher model) as possible, as introduced in Sec. 3.3. +- Instead of applying a random sampling strategy to select samples for knowledge distillation, we adopt a reweighting strategy to control the sampling process. We abuse notation slightly with $\circ$ to represent applying the strategy based on priority sampling function $\delta$ , with more details in Sec. 3.4. +- Each synthetic sample is not only prioritized by its modulating function $\phi(x)$ but also is reweighted at the sampling stage that reuses the same intermediate values as $\phi(x)$ . + +Although the D-KD pipeline allows training samples to be synthesized and served for training the student model on the same task. However, there is a large extent of data redundancy that hinders the training efficiency of D-KD methods. In the following sections, we detail our SSD-KD, a fully efficient D-KD method that is capable of using an extremely small scale of synthetic data yet achieving competitive performance compared to existing D-KD methods. + +The pipeline of SSD-KD is summarized in Alg. 1 and the comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] and more efficient family [10, 11], and our + +SSD-KD is shown in Fig. 3. + +# 3.3. Data Inversion with Distribution Balancing + +We provide Fig. 2 to demonstrate the data redundancy of D-KD that results from a large imbalance of the synthetic data. The left two sub-figures of Fig. 2 depict the distribution of categories predicted by the teacher model, indicating a significant imbalance in data categories. The right two sub-figures of Fig. 2 show sample accounts of different bars that correspond to different prediction difficulties (the difficulty is measured by the predicted probability by the teacher model). For D-KD, it indicates that generating samples with only the instruction of teacher-student discrepancy results in a sharp distribution over the sample difficulty and tends to obtain easily predicted data samples. We argue that for the D-KD task where data samples are all synthesized, the data-generating process needs to consider both the teacher-student discrepancy and the pre-trained knowledge of the teacher itself, by which our SSD-KD proposes diversity-aware and difficulty-aware data synthesis, as detailed below. + +Diversity-aware balancing. We first propose to address the issue of imbalanced sample difficulty in the data inversion process. Specifically, we maintain a replay buffer $\mathcal{B}$ that stores a constant amount (denoted as $|\mathcal{B}|$ ) of synthetic data samples. For each data sample $x$ in $\mathcal{B}$ , we penalize its total amount of samples that share the same predicted category (by the teacher model) with $x$ . To realize this, we adopt a diversity-aware balancing term that encourages the generator to synthesize samples with infrequent categories, which will be shown in Eq. (3). + +Difficulty-aware balancing. Drawing inspiration from the field of object detection that utilizes focal loss for largely imbalanced samples [16, 27], for each sample $x$ , we further introduce a difficulty-aware balancing term on the predicted probability $p_T(x)$ . Here, difficult synthetic samples are considered as those with low-confidence predictions by the teacher model, which are encouraged by the difficulty-aware balancing term as will be given in Eq. (3). + +In summary, we introduce a modulating function $\phi (x)$ to adjust the optimization of the generator based on the pre + +diction feedback from the pre-trained teacher model. $\phi (x)$ is expected to balance the category distribution and dynamically distinguish between easy and hard synthetic samples, by which easy samples no longer overwhelm the distillation process. Formally, for a synthetic data sample $x\in \mathcal{B}$ , its modulating function $\phi (x)$ is computed by + +$$ +\phi (x) = \underbrace {\left(1 - \frac {1}{| \mathcal {B} |} \sum_ {x ^ {\prime} \in \mathcal {B}} \mathbb {I} _ {c _ {T} \left(x ^ {\prime}\right) = c _ {T} (x)}\right)} _ {\text {d i v e r s i t y - a w a r e b a l a n c i n g}} \underbrace {\left(1 - p _ {T} (x)\right) ^ {\gamma}} _ {\text {d i f f i c u l t y - a w a r e b a l a n c i n g}}, \tag {3} +$$ + +where $c_{T}(x)$ and $p_{T}(x)$ refer to the predicted category index and probability (or confidence) by the pre-trained teacher model, respectively; $\mathbb{I}_{c_T(x') = c_T(x)}$ denotes an indicator function that equals to 1 if the predicted category of $x'$ is the same with that of $x$ , while otherwise 0; $\gamma$ is a hyperparameter. + +We highlight two properties of the modulating function $\phi(x)$ . Firstly, for the data sample $x$ with a high prediction certainty w.r.t. the teacher model (i.e., considered as an easy sample), $\phi(x)$ approaches to a low value and thus results in low impact on the task-oriented loss $\mathcal{L}_{\mathrm{Task}}(x)$ in Eq. (2). Secondly, when the category distribution of synthetic data samples in $\mathcal{B}$ is largely imbalanced as predicted by the teacher network, the sample $x$ of which the category shares with more samples in $\mathcal{B}$ is penalized and thus the corresponding $\mathcal{L}_{\mathrm{Task}}(x)$ is weakened by $\phi(x)$ . + +Although Eq. (3) implies that the value of the modulating function $\phi(x)$ is partially determined by the current replay buffer $\mathcal{B}$ , note that $\mathcal{B}$ changes dynamically and is also affected by $\phi(x)$ . This is due to that the term $\phi(x)\mathcal{L}_{\mathrm{Task}}(x)$ in Eq. (2) directly optimizes the generator that synthesizes data samples to compose $\mathcal{B}$ . In this sense, the balancing in category diversity is maintained during training given the mutual effects of $\mathcal{B}$ and $\phi(x)$ . We find it particularly important in the data inversion process for balancing categories of data samples. + +With both balancing terms, as shown in Fig. 2, our method (SSD-KD) generates mild sample distributions in terms of both sample category and difficulty. + +# 3.4. Distillation with Priority Sampling + +The original prioritized experience replay method [26] reuses important transitions more frequently and learns more efficiently. Differently, rather than obtaining the reward from the environment, our prioritized sampling method is designed to fit in data-free knowledge distillation and get feedback from the framework itself. In other words, the prioritized sampling method performs the opposite role of the previous data-free knowledge distillation methods: it focuses on training a sparse set of highly prioritized samples instead of uniform sampling to speed up training. + +By Eq. (3), we sample the synthetic data $x$ from the current replay buffer $\mathcal{B}$ . Instead of uniformly sampling $x$ , we + +Algorithm 1: SSD-KD +Input: Number of training epochs $E$ ; number of iterations for data inversion $T_{g}$ ; number of iterations for distillation $T_{kd}$ ; pre-trained teacher model $f_{t}(\cdot;\theta_{t})$ ; student network $f_{s}(\cdot;\theta_{s})$ +Output: The optimized student model $f_{s}(\cdot;\theta_{s})$ +1 Initialize replay buffer $\mathcal{B} \gets \emptyset$ and $f_{s}(\cdot;\theta_{s})$ +2 for $e = 1$ to $E$ do +3 (1) Data Inversion with Distribution Balancing: +4 Initialize a generative network $g(\cdot;\theta_{g})$ +5 $z \gets \mathcal{N}(0,1)$ +6 for $i = 1$ to $T_{g}$ do +7 $x \gets g(z;\theta_{g})$ +8 Compute the overall loss by Eq. (2) +9 Update $z$ and $\theta_{g}$ +10 end +11 (2) Distillation with Priority Sampling: +12 Compute the importance-sampling weight $w_{i-1}(x)$ +13 Compute the sample priority $\delta_{i}(x)$ +14 $\mathcal{B} \gets \mathcal{B} \cup \{x,\delta_{i}(x)\}$ +15 Remove old samples to keep $|\mathcal{B}|$ a constant +16 for $j = 1$ to $T_{kd}$ do +17 Sample a small-scale mini-batch $M$ from $\mathcal{B}$ +18 Compute the distillation loss $\mathcal{L}_{\mathrm{KD}}(x)$ in Eq. (2), where $x \in M$ +19 Compute the importance-sampling weight $w_{i}(x)$ +20 Compute the sample priority $\delta_{i+1}(x)$ +21 Update old samples $\{x,\delta_{i}(x)\}$ to $\{x,\delta_{i+1}(x)\}$ in $\mathcal{B}$ +22 Update $\theta_{s}$ +23 end +24 end +25 Return $f_{s}(\cdot;\theta_{s})$ + +propose to modulate the sampling probability by a sampling strategy termed Priority Sampling (PS). The basic function of PS is to measure the importance of each sample $x$ in $\mathcal{B}$ , by which we introduce a priority sampling function $\delta_{i}(x)$ , + +$$ +\delta_ {i} (x) = w _ {i - 1} (x) K L \left(f _ {t} (x; \theta_ {t}) \mid \mid f _ {s} \left(x; \theta_ {s}\right)\right), \tag {4} +$$ + +where as mentioned in Sec. 3.1, $KL$ denotes the KL-divergence between the softmax outputs of logits $f_{t}(x;\theta_{t})$ and $f_{s}(x;\theta_{s})$ ; $\theta_t$ , $\theta_s$ depend on the training step $i$ ; $w_{i}(x)$ is the calibration term [26] for normalizing samples in $\mathcal{B}$ , as will be formalized in Eq. (5), especially, when $i = 0$ , $w_{-1}(x) = 1$ . + +The training of knowledge distillation with random updates relies on those updates corresponding to the same dis + +Table 1. Performance comparison of Fast2 [11] (the current most efficient D-KD method) and our SSD-KD, in terms of top-1 classification accuracy $(\%)$ and overall training time cost (hours). Our SSD-KD performs with a very small training data scale: 5000 synthetic samples, i.e., $10\%$ of the original training dataset size. All image classification results in this and the other tables are averaged over three independent runs, and the methods "Teacher" and "Student" are performed on the whole original training dataset. + +
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
Fast2 [11]92.62(3.92h)84.67(3.02h)88.36(1.46h)89.56(2.37h)89.68(1.70h)
SSD-KD92.92(1.11h)89.56(0.77h)88.07(0.77h)90.12(0.96h)91.84(0.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
Fast2 [11]69.76(4.02h)62.83(2.99h)41.77(1.44h)53.15(2.26h)57.08(1.72h)
SSD-KD73.05(1.12h)66.66(0.78h)51.96(0.77h)59.60(0.97h)61.44(0.78h)
+ +distribution as its expectation. Prioritized sampling data introduces bias since it might change the data distribution, and affect the solution that the estimates will converge to. Thus we correct the bias by introducing an importance-sampling (IS) weight $w_{i}(x)$ for the data sample $x$ : + +$$ +w _ {i} (x) = \left(N \cdot P _ {i} (x)\right) ^ {- \beta}, \tag {5} +$$ + +where $\beta$ is a hyper-parameter; $P_{i}(x)$ is the probability of sampling transition defined by + +$$ +P _ {i} (x) = \frac {\left(\left| \delta_ {i} (x) \right| + \epsilon\right) ^ {\alpha}}{\sum_ {x ^ {\prime} \in \mathcal {B}} \left(\left| \delta_ {i} \left(x ^ {\prime}\right) \right| + \epsilon\right) ^ {\alpha}}, \tag {6} +$$ + +where $\epsilon$ is a small positive constant which prevents the edge-case of transitions not being selected once their priority is zero. + +The priority sampling function $\delta(x)$ has two noteworthy properties. Firstly, as the delta value increases, $\delta(x)$ reflects a greater information discrepancy between the teacher and student models for the synthetic samples in the current $\mathcal{B}$ . The student model should therefore be optimized from samples with greater information discrepancy, as this facilitates the faster acquisition of the teacher model. Secondly, $\delta(x)$ dynamically changes with each update iteration of the student and generative models. Consequently, when the student model acquires the teacher model's capabilities on certain samples, it continues to learn from samples with larger differences relative to the teacher model based on the new sample distribution. This further enhances the performance of the student model. + +# 4. Experiment + +# 4.1. Experimental Details + +We conduct comprehensive experiments on image classification and semantic segmentation tasks to evaluate the effectiveness and study the design of our SSD-KD. + +Datasets. For image classification experiments, we use CIFAR-10 and CIFAR-100 [14], two most popular datasets + +for D-KD research. For semantic segmentation experiments, we use the NYUv2 dataset [28]. On each dataset, the baseline models are trained with its standard data split. + +Training Setups. We follow the settings of [11] for basic experiments and comparisons. Regarding experiments on the CIFAR-10 and CIFAR-100 datasets, we use 5 different teacher-student model pairs having either the same type or different type network architectures (see Table 1, 2). Regarding experiments on the NYUv2 dataset, we use two Deeplabv3 models [6] as a teacher-student model pair (see Table 3). Unless otherwise stated, we always adopt the same basic settings as in [11] for experiments, including the number of training epochs, the optimizer, the weight decay, etc. Detailed settings are put in the supplementary material. + +Evaluation Metrics. Besides comparing the student model accuracy (top-1 accuracy for image classification and mean Intersection over Union (IoU) for semantic segmentation), we also compare the overall training time cost of existing mainstream D-KD methods [4, 7, 10, 11, 20, 30] and our SSD-KD. For each teacher-student model pair, we meticulously record the total time cost (hours) for each run of the end-to-end training by all methods. In order to guarantee a fair comparison, all training speed assessments are performed on 1 NVIDIA V100 GPU using 12 cores of Intel Xeon Gold 6240R CPU. Our experiments are implemented with the PyTorch [23] library. For each teacher-student model pair, the experiment is conducted with three independent runs for all methods, and in comparison we report the averaged results, unless otherwise stated. + +# 4.2. Experimental Results + +Results on image classification task. As the main focus of our work is to improve the overall training efficiency of the adversarial data-free knowledge distillation paradigm, we first compare the proposed SSD-KD with the current most efficient D-KD method Fast2 [11] on the CIFAR-10 and CIFAR-100 datasets. From the results shown in Table 1, we can observe: (1) on the CIFAR-10 dataset, our SSD-KD (using a very small data scale: 5000 synthetic samples, + +Table 2. Performance comparison of existing top-performing D-KD methods and our SSD-KD, in terms of top-1 classification accuracy $(\%)$ and overall training time cost (hours). For our SSD-KD, we relax the data scale of synthetic samples to be similar to that for Fast5. + +
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
DeepInv [30]93.26(25.62h)90.36(13.19h)83.04(10.62h)86.85(13.84h)89.72(11.67h)
CMI [10]94.84(15.49h)91.13(10.99h)90.01(9.87h)92.78(11.57h)92.52(10.58h)
DAFL [4]92.22(303.50h)81.10(209.78h)65.71(96.57h)81.33(157.44h)81.55(118.52h)
DFQ [7]94.61(152.16h)90.84(104.93h)86.14(48.40h)91.69(78.81h)92.01(59.40h)
ZSKT [20]93.32(304.11h)89.46(209.65h)83.74(96.75h)86.07(157.51h)89.66(118.82h)
Fast5 [11]93.63(4.34h)89.94(3.15h)88.90(1.61h)92.04(2.44h)91.96(1.89h)
Fast10 [11]94.05(4.72h)90.53(3.22h)89.29(1.72h)92.51(2.48h)92.45(2.05h)
SSD-KD94.26(4.20h)90.67(2.95h)89.96(1.48h)93.23(2.30h)93.11(1.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
DeepInv [30]61.32(25.88h)54.13(13.13h)53.77(10.62h)61.33(13.86h)61.34(11.68h)
CMI [10]77.04(15.42h)70.56(11.28h)57.91(10.01h)68.88(11.57h)68.75(10.53h)
DAFL [4]74.47(303.54h)54.16(209.82h)20.88(96.62h)42.83(157.49h)43.70(118.67h)
DFQ [7]77.01(152.18h)66.21(105.14h)51.27(48.47h)54.43(78.83h)64.79(59.48h)
ZSKT [20]67.74(304.13h)54.31(209.63h)36.66(96.87h)53.60(157.68h)54.59(118.84h)
Fast5 [11]72.82(4.35h)65.28(3.06h)52.90(1.56h)61.80(2.35h)63.83(1.86h)
Fast10 [11]74.34(4.50h)67.44(3.12h)54.02(1.65h)63.91(2.42h)65.12(1.95h)
SSD-KD75.16(4.22h)68.77(2.94h)55.61(1.52h)64.57(2.27h)65.28(1.78h)
+ +i.e., $10\%$ of the original training dataset size) shows large improvements in training efficiency to Fast2 (at least $1.90\times$ and at most $3.92\times$ training speedup, on 5 teacher-student model pairs), and gets better student models on 4 out of 5 teacher-student model pairs; (2) on the more challenging CIFAR-100 dataset, our SSD-KD shows a very similar training speedup trend against Fast2 as on the CIFAR-10 dataset, but gets significantly better student models on all 5 teacher-student model pairs (at least $3.29\%$ and at most $10.19\%$ absolute top-1 accuracy gain to Fast2). The superior performance of SSD-KD against Fast2 validates the efficacy of our small-scale data inversion and sampling mechanism which can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes. + +Next, we compare the proposed SSD-KD with existing top-performing D-KD methods including DeepInv [30], CMI [10], DAFL [4], DFQ [7], ZSKT [20], Fast5 and Fast10 [11]. In order to get improved student model performance, we relax the data scale of synthetic samples in our SSD-KD to be similar to that for Fast5. Experimental results are summarized in Table 2. Compared to these mainstream D-KD methods, our SSD-KD gets very competitive performance in student model accuracy, and significantly better performance in overall training efficiency, on all 5 teacher-student model pairs. + +In light of the results in Table 1, 2, it can be seen that our SSD-KD makes the overall training efficiency one or two orders of magnitude faster than most D-KD methods while retaining competitive student model performance. + +Table 3. Performance comparison on the NYUv2 dataset. The teacher model is pre-trained on the ImageNet dataset and finetuned on the NYUv2 dataset, and the student model is trained from scratch. The results of reference methods are collected from [11]. + +
MethodTraining Data ScalemIoU (%)
Teacher: Deeplabv3-ResNet501,449 NYUv20.519
Student: Deeplabv3-Mobilenetv21,449 NYUv20.375
KD [12]1,449 NYUv20.380
DFND [5]14M (ImageNet)0.378
DFAD [9]960K (synthetic)0.364
DAFL [4]960K (synthetic)0.105
Fast10 [11]17K (synthetic)0.366
SSD-KD16K (synthetic)0.384
+ +Results on semantic segmentation task. To further validate the generalization ability of our method, we compare the performance of SSD-KD with existing D-KD methods on the NYUv2 dataset. Table 3 summarizes the results. We can see that our SSD-KD not only achieves state-of-the-art performance in terms of model accuracy, but also is significantly more efficient than other D-KD methods in terms of the training data scale. The overall training time cost for our SSD-KD and Fast10 is 8.9 hours and 9.5 hours, respectively. Besides, the student model trained by SSD-KD even outperforms the baseline model trained on the original training data, as NYUv2 is a small-scale dataset. + +# 4.3. Ablation Studies + +Effect of the synthetic data scale. So far, we have already demonstrated that SSD-KD has the appealing capability to attain efficient and effective end-to-end distillation training using a small amount of synthetic samples. To bet- + +![](images/5fe131909012347caf67092652fbfd4f1caf4b0101025930b6ce32e24131f577.jpg) +Figure 4. Performance comparison of SSD-KD under different synthetic data scales against the original training dataset size, in terms of top-1 classification accuracy $(\%)$ and overall training time cost (hour). + +![](images/9c9704e1d98650e5f15730a4013d278788f6eaa4fc5a2214531fa9a3dad6e6b1.jpg) + +![](images/898ebb0651ce575f43792e1a95ed211c7c757e0408b20edd57290557101d66a2.jpg) + +![](images/cf55f48331c8c1e62490291f08d16b98e3030536f036ffeac314115379f670b9.jpg) + +Table 4. Effect of two core modules in SSD-KD. PS: the priority sampling function, Difficulty|Diversity: the difficulty|diversity-aware balancing term in the modulating function, see Eq. (3). + +
DatasetBaselinePSDifficultyDiversityResNet-34 ResNet-18WRN40-2 WRN16-1
CIFAR-10✓87.35(1.19h)80.47(0.84h)
✓✓90.28(1.01h)87.19(0.69h)
✓✓88.57(1.20h)83.12(0.86h)
✓✓89.12(1.21h)84.31(0.87h)
✓✓✓89.95(1.21h)85.79(0.88h)
✓✓✓✓92.92(1.11h)88.07(0.77h)
CIFAR-100✓64.05(1.19h)41.67(0.84h)
✓✓71.70(1.01h)49.70(0.69h)
✓✓66.94(1.20h)43.96(0.86h)
✓✓68.52(1.21h)46.34(0.87h)
✓✓✓70.85(1.21h)47.82(0.88h)
✓✓✓✓73.05(1.12h)51.96(0.77h)
+ +![](images/a5d5d534d99b2af5096176bece819beed0552b8ab90d64915bec0d9892f33e6f.jpg) +Figure 5. Visualization examples of synthetic image samples generated by Fast10 and our SSD-KD for the NYUv2 dataset. + +ter explore the boundary of this capability, we conduct an ablation with three teacher-student model pairs, including ResNet34 $\rightarrow$ ResNet18, VGG11 $\rightarrow$ ResNet18, and WRN40-2 $\rightarrow$ WRN16-1. In the experiments, we decrease the synthetic data scale used in SSD-KD [11] from 50,000(100% relative to the original training data size) to {10000(20%), 5000(10%), 2500(%5), 500(1%)} for the CIFAR-10 and CIFAR-100 datasets. As shown in Fig. 4, the accuracy of the student model trained by SSD-KD remains stable across a relatively large synthetic data scale range, for all teacher-student model pairs. When decreasing the synthetic data scale, the overall training time cost appears to decrease nearly linearly, while the student model accuracy drop is mild (less than 10% even for the synthetic data scale 500). + +Effect of the core modules. In Table 4, we provide an ablation to scrutinize our SSD-KD systematically. We + +observe that: (1) the two modulating functions (consisting of a diversity-aware term and a difficulty-aware term, see Eq. (3)) and the priority sampling function, are both critical to our SSD-KD; (2) the combination of them strikes a good tradeoff between model accuracy and training efficiency. + +Visualization of data inversion. Fig. 5 shows examples of synthetic images inverted by Fast10 and our SSD-KD for the NYUv2 dataset. Compared to Fast10, our method can better invert texture information and has less noise. + +# 5. Conclusion + +In this paper, we presented SSD-KD, the first fully efficient method to advance adversarial data-free knowledge distillation research. Benefiting from a small-scale data inversion and sampling mechanism based on a modulating function and a priority sampling function, SSD-KD can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, attaining efficient and effective data-free knowledge distillation. Extensive experiments on image classification and semantic segmentation benchmarks validate the efficacy of SSD-KD. We hope our work can inspire future research on efficient D-KD designs. + +# 6. Acknowledgement + +This work was supported by the National Natural Science Fund for Key International Collaboration (62120106005). + +# References + +[1] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. NeurIPS, 2017. 3 +[2] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 1 +[3] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. arXiv preprint arXiv:1905.07072, 2019. 1 +[4] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 1, 3, 4, 6, 7 +[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In CVPR, 2021. 7 +[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6 +[7] Yoojin Choi, Jihwan Choi, Mostafa El-Khamy, and Jungwon Lee. Data-free network quantization with adversarial knowledge distillation. In CVPR Workshops, 2020. 1, 2, 3, 4, 6, 7 +[8] Kien Do, Thai Hung Le, Dung Nguyen, Dang Nguyen, Haripriya Harikumar, Truyen Tran, Santu Rana, and Svetha Venkatesh. Momentum adversarial distillation: Handling large distribution shifts in data-free knowledge distillation. NeurIPS, 2022. 1, 3 +[9] Gongfan Fang, Jie Song, Chengchao Shen, Xinchao Wang, Da Chen, and Mingli Song. Data-free adversarial distillation. arXiv preprint arXiv:1912.11006, 2019. 7 +[10] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. arXiv preprint arXiv:2105.08584, 2021. 1, 2, 3, 4, 6, 7 +[11] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In AAAI, 2022. 1, 2, 3, 4, 6, 7, 8 +[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 7 +[13] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 3 +[14] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Tech Report, 2009. 6 +[15] Yuhang Li, Feng Zhu, Ruihao Gong, Mingzhu Shen, Xin Dong, Fengwei Yu, Shaoqing Lu, and Shi Gu. Mixmix: All you need for data-free compression are feature and data mixing. In ICCV, 2021. 2, 3 +[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 4 + +[17] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured knowledge distillation for semantic segmentation. In CVPR, 2019. 1 +[18] Raphael Gontijo Lopes, Stefano Fenu, and Thad Starner. Data-free knowledge distillation for deep neural networks. arXiv preprint arXiv:1710.07535, 2017. 1, 3 +[19] Liangchen Luo, Mark Sandler, Zi Lin, Andrey Zhmoginov, and Andrew Howard. Large-scale generative data-free distillation. arXiv preprint arXiv:2012.05578, 2020. 3 +[20] Paul Micaelli and Amos J Storkey. Zero-shot knowledge transfer via adversarial belief matching. NeurIPS, 2019. 4, 6, 7 +[21] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In ICML, 2019. 1, 3 +[22] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 1 +[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6 +[24] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In CVPR, 2023. 1, 3 +[25] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1 +[26] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. 3, 5 +[27] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 4 +[28] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6 +[29] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 3 +[30] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2, 3, 4, 6, 7 +[31] Jaemin Yoo, Minyong Cho, Taebum Kim, and U Kang. Knowledge extraction with no observable data. NeurIPS, 2019. 1, 3 +[32] Shikang Yu, Jiachen Chen, Hu Han, and Shuqiang Jiang. Data-free knowledge distillation via feature exchange and activation region constraint. In CVPR, 2023. 3 +[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 1 \ No newline at end of file diff --git a/2024/Small Scale Data-Free Knowledge Distillation/images.zip b/2024/Small Scale Data-Free Knowledge Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..340c5572a8a2ba408a39577360bd7b0b55b279a9 --- /dev/null +++ b/2024/Small Scale Data-Free Knowledge Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:640df7737f31e825ff33ba1b6b91a5d7383da504450e8dfbc177c24a5e3198b2 +size 597910 diff --git a/2024/Small Scale Data-Free Knowledge Distillation/layout.json b/2024/Small Scale Data-Free Knowledge Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..36ca89d1431c7f3637b8acb9d91ceab2e0f6d62a --- /dev/null +++ b/2024/Small Scale Data-Free Knowledge Distillation/layout.json @@ -0,0 +1,8478 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 155, + 103, + 438, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 103, + 438, + 121 + ], + "spans": [ + { + "bbox": [ + 155, + 103, + 438, + 121 + ], + "type": "text", + "content": "Small Scale Data-Free Knowledge Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "text", + "content": "He Liu" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "text", + "content": ", Yikai Wang" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "text", + "content": ", Huaping Liu" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "text", + "content": ", Fuchun Sun" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "text", + "content": ", Anbang Yao" + }, + { + "bbox": [ + 118, + 142, + 475, + 159 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 162, + 405, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 162, + 405, + 176 + ], + "spans": [ + { + "bbox": [ + 187, + 162, + 405, + 176 + ], + "type": "text", + "content": "1China Telecom Cloud Technology Co., Ltd." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 177, + 469, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 177, + 469, + 191 + ], + "spans": [ + { + "bbox": [ + 124, + 177, + 469, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 177, + 469, + 191 + ], + "type": "text", + "content": "BNRist Center, State Key Lab on Intelligent Technology and Systems," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 191, + 519, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 191, + 519, + 205 + ], + "spans": [ + { + "bbox": [ + 74, + 191, + 519, + 205 + ], + "type": "text", + "content": "Department of Computer Science and Technology, Tsinghua University " + }, + { + "bbox": [ + 74, + 191, + 519, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 74, + 191, + 519, + 205 + ], + "type": "text", + "content": "Intel Labs China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 207, + 504, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 207, + 504, + 219 + ], + "spans": [ + { + "bbox": [ + 87, + 207, + 504, + 219 + ], + "type": "text", + "content": "{liuhel7@mails.,yikaiw@,hpliu@,fuchuns@}tsinghua.edu.cn，anbang.yao@intel.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 285, + 289, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 657 + ], + "type": "text", + "content": "Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data, avoiding privacy, security, and proprietary risks in real applications. In this line of research, existing methods typically follow an inversion-and-distillation paradigm in which a generative adversarial network on-the-fly trained with the guidance of the pre-trained teacher network is used to synthesize a large-scale sample set for knowledge distillation. In this paper, we reexamine this common data-free knowledge distillation paradigm, showing that there is considerable room to improve the overall training efficiency through a lens of \"small-scale inverted data for knowledge distillation\". In light of three empirical observations indicating the importance of how to balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, we propose Small Scale Data-free Knowledge Distillation (SSD-KD). In formulation, SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples, facilitated by a dynamic replay buffer and a reinforcement learning strategy. As a result, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (e.g., " + }, + { + "bbox": [ + 46, + 285, + 289, + 657 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 285, + 289, + 657 + ], + "type": "text", + "content": " less than the original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream methods while retaining superior or competitive model performance, as demonstrated on popular image classification and semantic segmentation benchmarks. The code is available at https://github.com/OSVAI/SSD-KD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 246, + 387, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 387, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 387, + 258 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 268, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 437 + ], + "type": "text", + "content": "For computer vision applications on resource-constrained devices, how to learn portable neural networks yet with satisfied prediction accuracy is the key problem. Knowledge distillation (KD) [2, 12, 17, 22, 25, 33], which leverages the information of a pre-trained large teacher network to promote the training of a smaller target student network on the same training data, has become a mainstream solution. Conventional KD methods assume that the original training data is always available. However, accessing the source dataset on which the teacher network was trained is usually not feasible in practice, due to its potential privacy or security or proprietary or huge-size concerns. To relax the constraint on training data, knowledge distillation under a data-free regime has recently attracted increasing attention." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "text", + "content": "The basic idea of Data-free Knowledge Distillation (D-KD) [3, 18, 21] is to construct synthetic samples for knowledge distillation conditioned on the pre-trained teacher network, which would match the underlying distribution of the original training data. Existing top-performing D-KD methods [4, 7, 8, 10, 24, 30, 31] generally adopt an adversarial inversion-and-distillation paradigm. Under this paradigm, (1) during the inversion process, a generator is trained by taking the pre-trained teacher network as the discriminator; (2) during the subsequent knowledge distillation process, the on-the-fly learned generator will synthesize pseudo samples for training the student network. However, adversarial D-KD methods usually require generating a large number of synthetic samples (compared to the original training dataset size) in order to guarantee trustworthy knowledge distillation. This poses a heavy burden on training resource consumption, suppressing their use in real applications. In the recent work of [11], the authors present an effective meta-learning strategy that seeks common features and reuses them as initial priors to reduce the number of iteration steps required to reach the convergence of the generator. Although faster data synthesis can be attained, [11] still needs to generate a sufficiently large number of syn" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 715 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author. This work was done when He Liu was an intern at Intel Labs China, supervised by Anbang Yao who conceived the project." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6008" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 225, + 162 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 225, + 162 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 225, + 162 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 225, + 162 + ], + "type": "image", + "image_path": "b7f42998a74db30c7b8e815a1189712969ecef13a265d814d68391212031450a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 169, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 547, + 205 + ], + "type": "text", + "content": "Figure 1. Comparison of knowledge distillation (KD) using original samples vs. synthetic samples, under the same training data scale: 5000 samples (10% of the CIFAR-10 training dataset size). In such small-scale KD regime, the student models (DeepInv [30] and SSD-KD) trained on synthetic samples always show much better accuracy than the counterparts (Vanilla KD [12]) trained on original samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 71, + 366, + 162 + ], + "blocks": [ + { + "bbox": [ + 228, + 71, + 366, + 162 + ], + "lines": [ + { + "bbox": [ + 228, + 71, + 366, + 162 + ], + "spans": [ + { + "bbox": [ + 228, + 71, + 366, + 162 + ], + "type": "image", + "image_path": "a9c240abb887b9a1ec2f64ca22ed33a7533dc2c8bd1ed8faa10f24cede29b72e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 369, + 72, + 506, + 162 + ], + "blocks": [ + { + "bbox": [ + 369, + 72, + 506, + 162 + ], + "lines": [ + { + "bbox": [ + 369, + 72, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 369, + 72, + 506, + 162 + ], + "type": "image", + "image_path": "e18545f7d22b152170307a88c4da13443fd3a243b0468e8f156b885606daf5f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 210, + 170, + 302 + ], + "blocks": [ + { + "bbox": [ + 50, + 210, + 170, + 302 + ], + "lines": [ + { + "bbox": [ + 50, + 210, + 170, + 302 + ], + "spans": [ + { + "bbox": [ + 50, + 210, + 170, + 302 + ], + "type": "image", + "image_path": "fe139fdc34149eec03ccb445c88c9a599773d42a7ab2b60cd0f813f59566d066.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 304, + 213, + 315 + ], + "lines": [ + { + "bbox": [ + 129, + 304, + 213, + 315 + ], + "spans": [ + { + "bbox": [ + 129, + 304, + 213, + 315 + ], + "type": "text", + "content": "(a) Difficulty distribution." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 211, + 294, + 302 + ], + "blocks": [ + { + "bbox": [ + 173, + 211, + 294, + 302 + ], + "lines": [ + { + "bbox": [ + 173, + 211, + 294, + 302 + ], + "spans": [ + { + "bbox": [ + 173, + 211, + 294, + 302 + ], + "type": "image", + "image_path": "9b6198df7fadf4575acb67f2d97055cfc86844e36daf27473f20cf505ca95460.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 317, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 317, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 546, + 361 + ], + "type": "text", + "content": "Figure 2. Comparison of synthetic sample distributions collected from two top-performing adversarial D-KD methods (DeepInv [30] and Fast10 [11]) and our SSD-KD on CIFAR-100 dataset. Comparatively, Fig. 2a shows that our method can better balance the difficulty distribution of synthetic samples while encouraging the generator to invert more hard samples, and Fig. 2b further shows that our method can better balance the diversity distribution of synthetic samples across different categories." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 299, + 209, + 421, + 302 + ], + "blocks": [ + { + "bbox": [ + 299, + 209, + 421, + 302 + ], + "lines": [ + { + "bbox": [ + 299, + 209, + 421, + 302 + ], + "spans": [ + { + "bbox": [ + 299, + 209, + 421, + 302 + ], + "type": "image", + "image_path": "6607a92b3d6f91b26564e6bb5369f1fea0582bc266653b3d9ffdfc99ab03b42c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 304, + 462, + 315 + ], + "lines": [ + { + "bbox": [ + 380, + 304, + 462, + 315 + ], + "spans": [ + { + "bbox": [ + 380, + 304, + 462, + 315 + ], + "type": "text", + "content": "(b) Diversity distribution." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 424, + 209, + 544, + 302 + ], + "blocks": [ + { + "bbox": [ + 424, + 209, + 544, + 302 + ], + "lines": [ + { + "bbox": [ + 424, + 209, + 544, + 302 + ], + "spans": [ + { + "bbox": [ + 424, + 209, + 544, + 302 + ], + "type": "image", + "image_path": "7448c10dad09d474c88480a5743946261a5faeceb2bd86f85d29f4ec14f9de9c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 456 + ], + "type": "text", + "content": "thetic samples to ensure effective knowledge distillation, neglecting the efficiency of the following knowledge distillation process which will become the major bottleneck to the overall training efficiency. In a nutshell, there is no research effort made to improve the overall training efficiency of D-KD via jointly considering data inversion and knowledge distillation processes, to the best of our knowledge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": "To remedy this critical gap, this paper presents the first fully efficient D-KD approach termed Small Scale Data-free Knowledge Distillation (SSD-KD). Our SSD-KD improves the overall training efficiency of the adversarial inversion-and-distillation paradigm from a novel \"small data scale\" perspective. In this work, \"data scale\" refers to the total number of inverted samples used in knowledge distillation during a training epoch. The formulation of SSD-KD is inspired by three empirical observations. On different pairs of teacher-student networks, we first observe that the student networks trained on synthetic samples tend to show much better performance than their corresponding counterparts trained on original samples when significantly reducing data scales of synthetic samples and original samples to the same (e.g., " + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": " of the source training dataset size), as illustrated in Fig. 1. Note that synthetic samples are generated with the guidance of the teacher network pre-trained on the whole source dataset, which naturally reflect different views of the original data distribution. Under a small enough data scale, this makes synthetic samples have superior capability to original samples in fitting the underlying" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 373, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 553 + ], + "type": "text", + "content": "distribution of the whole source dataset. This inspiring observation indicates that if we can construct a small-scale set of high-quality synthetic samples, a promising way toward fully efficient D-KD would be created. In principle, we believe a high-quality small-scale synthetic dataset should have well-balanced class distributions in terms of both synthetic sample diversity and difficulty. However, our other two observations indicate that existing D-KD methods including both conventional and the most efficient designs do not have good capabilities to balance the aforementioned two class distributions of synthetic samples under a small data scale, as illustrated in Fig. 2. Note that there already exist a few D-KD methods to enhance the diversity of synthetic samples [7, 10, 15, 30], but the diversity of synthetic samples in terms of sample difficulty is not explored yet." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Driven by the above observations and analysis, we come up with our SSD-KD which introduces two interdependent modules to significantly accelerate the overall training efficiency of the predominant adversarial inversion-and-distillation paradigm. The first module of SSD-KD relies on a novel modulating function that defines a diversity-aware term and a difficulty-aware term to jointly balance the class distributions of synthetic samples during both data synthesis and knowledge distillation processes in an explicit manner. The second module defines a novel priority sampling function facilitated by a reinforcement learning strategy that selects a small portion of proper synthetic samples from candidates stored in a dynamic replay buffer for knowledge" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6009" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "distillation, further improving the end-to-end training efficiency. Benefiting from these two modules, SSD-KD has two appealing merits. On one side, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " less than original training data scale), making the overall training efficiency one or two orders of magnitude faster than many mainstream D-KD methods while retaining competitive model performance. On the other side, SSD-KD attains largely improved performance in student model accuracy and maintains overall training efficiency when relaxing the data scale of synthetic samples to a relatively large number (which is still smaller than those for existing D-KD methods). We validate these merits of our method by lots of experiments on popular image classification and semantic segmentation benchmarks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 262, + 134, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 134, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 134, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 282, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 289, + 568 + ], + "type": "text", + "content": "Data-free knowledge distillation. D-KD is originally explored by [18] which assumes that layer-wise activation records of a well-trained teacher network pre-computed on training samples is available. The prevailing inversion-and-distillation paradigm for D-KD without reliance on the original training data or the recorded metadata is introduced in [21]. In the inversion process, synthetic training samples are generated by leveraging the information learned by the teacher network, whose distribution is expected to fit the underlying distribution of the original training dataset. In the distillation process, the student network is trained on synthetic samples by forcing it to match the predictions of the teacher network. Based on this paradigm, [4, 31] use generative adversarial networks for data inversion. Subsequent D-KD methods mostly follow this adversarial inversion-and-distillation paradigm. They try to improve the data inversion process from different aspects, such as enhancing synthetic sample discrimination with contrastive learning [10] or an ensemble of generators [19], combating distribution shift with momentum adversarial learning [8] or meta learning [24], and promoting adversarial learning with data augmentation [15, 32]. Our method intends to improve the overall training efficiency of the adversarial D-KD, and differs from these methods both in focus and formulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Efficient synthetic data sampling. How to select proper synthetic samples for knowledge distillation is essential in D-KD research. Existing methods [8, 10, 11, 24] commonly rely on a memory bank to store synthetic samples, and directly update synthetic samples without considering the efficiency of the following knowledge distillation process. In sharp contrast, our method introduces a reinforcement learning strategy that adaptively selects appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty, significantly improving knowledge dis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "tillation efficiency. As far as we know, our method made the first attempt to extend reinforcement learning methodology [1, 13, 26] to address data-free knowledge distillation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 118, + 362, + 130 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 118, + 362, + 130 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 362, + 130 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 137, + 427, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 427, + 149 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 427, + 149 + ], + "type": "text", + "content": "3.1. Preliminaries: D-KD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "inline_equation", + "content": "f_{t}(\\cdot ;\\theta_{t})" + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "content": " be a teacher model pre-trained on the original task dataset that is no longer accessible, the goal of D-KD is to first construct a set of synthetic training samples " + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "content": " via inverting the data distribution information learned by the teacher model, on which a target student model " + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "inline_equation", + "content": "f_{s}(\\cdot ;\\theta_{s})" + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "content": " then can be trained by forcing it to mimic the teacher's function. Existing D-KD methods mostly use a generative adversarial network " + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "inline_equation", + "content": "g(\\cdot ;\\theta_{g})" + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "content": " for producing synthetic training samples " + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "inline_equation", + "content": "x = g(z;\\theta_g)" + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "content": " from the latent noise input " + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 156, + 545, + 275 + ], + "type": "text", + "content": ", which is trained by taking the teacher model as the discriminator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 276, + 546, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 546, + 419 + ], + "type": "text", + "content": "The optimization of D-KD contains a common distillation regularization to minimize the teacher-student function discrepancy " + }, + { + "bbox": [ + 304, + 276, + 546, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KD}}(x) = D_{\\mathrm{KD}}(f_t(x;\\theta_t)\\| f_s(x;\\theta_s))^1" + }, + { + "bbox": [ + 304, + 276, + 546, + 419 + ], + "type": "text", + "content": " based on the KL-divergence, and a task-oriented regularization " + }, + { + "bbox": [ + 304, + 276, + 546, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Task}}(x)" + }, + { + "bbox": [ + 304, + 276, + 546, + 419 + ], + "type": "text", + "content": ", e.g., the cross-entropy loss using the teacher's predication as the ground truth. Besides, since D-KD is primarily based on the assumption that the teacher model has been optimized to be capable of capturing the source training data distribution after pre-training, recent D-KD methods [7, 29, 30] introduce an extra loss to regularize the statistics (Batch-Normalization (BN) parameters) of the training data distribution during the data inversion process," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 426, + 545, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 426, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 310, + 426, + 545, + 450 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B N}} (x) = \\sum_ {l} \\left| \\left| \\mu_ {l} (x) - \\mathbb {E} (\\mu_ {l}) \\right| \\right| _ {2} + \\left| \\left| \\sigma_ {l} ^ {2} (x) - \\mathbb {E} (\\sigma_ {l} ^ {2}) \\right| \\right| _ {2}, \\tag {1}", + "image_path": "0aa5c8d7dff707e375ed776abf1667b2973f2292d7b2a8580127989526dc42aa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mu_l(\\cdot)" + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\sigma_l(\\cdot)" + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": " denote the batch-wise mean and variance estimates of feature maps at the " + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": "-th layer, respectively; " + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathbb{E}(\\cdot)" + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": " over the BN statistics can be approximately substituted by running mean or variance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "type": "text", + "content": "The effectiveness of D-KD heavily depends on the quality of synthetic samples which are inverted from leveraging the knowledge of the pre-trained teacher model. The prevailing adversarial D-KD paradigm consists of two processes, namely data inversion and knowledge distillation. From both perspectives of efficiency and efficacy, on the one hand, the data inversion process largely affects the optimization performance of the student model; on the other hand, the training time cost of knowledge distillation emerges as a significant constraint to the overall training efficiency of D-KD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 644, + 430, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 430, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 430, + 657 + ], + "type": "text", + "content": "3.2. Our Design: SSD-KD" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "text", + "content": "Regarding our SSD-KD, we focuses on improving the adversarial D-KD paradigm through a lens of \"small-scale" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " Concretely, " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "{f}_{t}\\left( {x;{\\theta }_{t}}\\right)" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "{f}_{s}\\left( {x;{\\theta }_{s}}\\right)" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " are the outputted logits (before softmax) of the teacher model and the student model, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6010" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 177, + 175 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 177, + 175 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 177, + 175 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 177, + 175 + ], + "type": "image", + "image_path": "dc45cced476971ff7311f2e7e0ee4076b0abfcf0eb660e02de925fa8f62e45cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 225 + ], + "type": "text", + "content": "Figure 3. Comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] (left) and more efficient family [10, 11] (middle), and our SSD-KD (right). Our SSD-KD formulates a reinforcement learning strategy that can flexibly seek appropriate synthetic samples to update a portion of existing samples in a dynamic replay buffer by explicitly measuring their priorities in terms of jointly balancing sample diversity and difficulty distributions. See the method section for notation definitions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 68, + 539, + 175 + ], + "blocks": [ + { + "bbox": [ + 187, + 68, + 539, + 175 + ], + "lines": [ + { + "bbox": [ + 187, + 68, + 539, + 175 + ], + "spans": [ + { + "bbox": [ + 187, + 68, + 539, + 175 + ], + "type": "image", + "image_path": "4dfee92936b674cac3945ad7aff790f96c23b78efcfe8e52c032a1b485a8ee8b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 235, + 289, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 289, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 289, + 318 + ], + "type": "text", + "content": "inverted data for knowledge distillation\". SSD-KD lays emphasis on instructing the data inversion process with the feedback of both the pre-trained teacher and the knowledge distillation process, significantly accelerating the overall training efficiency. Following the notations in the prevision subsection, the optimization goal of our SSD-KD is defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 328, + 287, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 287, + 346 + ], + "type": "interline_equation", + "content": "\\min _ {f _ {s}} \\max _ {g} \\mathbb {E} _ {x = \\delta \\circ g (z)} \\left(\\mathcal {L} _ {\\mathrm {B N}} (x) + \\mathcal {L} _ {\\mathrm {K D}} (x) + \\phi (x) \\mathcal {L} _ {\\text {T a s k}} (x)\\right), \\tag {2}", + "image_path": "0776ab5d315eb1ebe0234a3aa041b1ae7fd511838c062dd36ee37d32d208a8ea.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 355, + 77, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 77, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 77, + 365 + ], + "type": "text", + "content": "where," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 366, + 287, + 569 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 47, + 366, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 287, + 413 + ], + "type": "text", + "content": "- We adopt a diversity-aware modulating function " + }, + { + "bbox": [ + 47, + 366, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 47, + 366, + 287, + 413 + ], + "type": "text", + "content": " that allocates each synthetic sample with a different priority regarding its predicted category by the teacher model, as presented in Sec. 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 415, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 460 + ], + "type": "text", + "content": "- Under the constraint of BN estimates, with " + }, + { + "bbox": [ + 47, + 415, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 47, + 415, + 287, + 460 + ], + "type": "text", + "content": ", we encourage the generator to explore as tough synthetic samples (w.r.t. the teacher model) as possible, as introduced in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 462, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 287, + 533 + ], + "type": "text", + "content": "- Instead of applying a random sampling strategy to select samples for knowledge distillation, we adopt a reweighting strategy to control the sampling process. We abuse notation slightly with " + }, + { + "bbox": [ + 47, + 462, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 47, + 462, + 287, + 533 + ], + "type": "text", + "content": " to represent applying the strategy based on priority sampling function " + }, + { + "bbox": [ + 47, + 462, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 462, + 287, + 533 + ], + "type": "text", + "content": ", with more details in Sec. 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "text", + "content": "- Each synthetic sample is not only prioritized by its modulating function " + }, + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "text", + "content": " but also is reweighted at the sampling stage that reuses the same intermediate values as " + }, + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 47, + 534, + 287, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "Although the D-KD pipeline allows training samples to be synthesized and served for training the student model on the same task. However, there is a large extent of data redundancy that hinders the training efficiency of D-KD methods. In the following sections, we detail our SSD-KD, a fully efficient D-KD method that is capable of using an extremely small scale of synthetic data yet achieving competitive performance compared to existing D-KD methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The pipeline of SSD-KD is summarized in Alg. 1 and the comparison of optimization pipelines for existing adversarial D-KD methods including both conventional family [4, 7, 20, 30] and more efficient family [10, 11], and our" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 235, + 422, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 422, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 422, + 247 + ], + "type": "text", + "content": "SSD-KD is shown in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 253, + 531, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 531, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 531, + 266 + ], + "type": "text", + "content": "3.3. Data Inversion with Distribution Balancing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 270, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 474 + ], + "type": "text", + "content": "We provide Fig. 2 to demonstrate the data redundancy of D-KD that results from a large imbalance of the synthetic data. The left two sub-figures of Fig. 2 depict the distribution of categories predicted by the teacher model, indicating a significant imbalance in data categories. The right two sub-figures of Fig. 2 show sample accounts of different bars that correspond to different prediction difficulties (the difficulty is measured by the predicted probability by the teacher model). For D-KD, it indicates that generating samples with only the instruction of teacher-student discrepancy results in a sharp distribution over the sample difficulty and tends to obtain easily predicted data samples. We argue that for the D-KD task where data samples are all synthesized, the data-generating process needs to consider both the teacher-student discrepancy and the pre-trained knowledge of the teacher itself, by which our SSD-KD proposes diversity-aware and difficulty-aware data synthesis, as detailed below." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": "Diversity-aware balancing. We first propose to address the issue of imbalanced sample difficulty in the data inversion process. Specifically, we maintain a replay buffer " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " that stores a constant amount (denoted as " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": ") of synthetic data samples. For each data sample " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": ", we penalize its total amount of samples that share the same predicted category (by the teacher model) with " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": ". To realize this, we adopt a diversity-aware balancing term that encourages the generator to synthesize samples with infrequent categories, which will be shown in Eq. (3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "Difficulty-aware balancing. Drawing inspiration from the field of object detection that utilizes focal loss for largely imbalanced samples [16, 27], for each sample " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ", we further introduce a difficulty-aware balancing term on the predicted probability " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "p_T(x)" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ". Here, difficult synthetic samples are considered as those with low-confidence predictions by the teacher model, which are encouraged by the difficulty-aware balancing term as will be given in Eq. (3)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "In summary, we introduce a modulating function " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\phi (x)" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " to adjust the optimization of the generator based on the pre" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6011" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "text", + "content": "diction feedback from the pre-trained teacher model. " + }, + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "inline_equation", + "content": "\\phi (x)" + }, + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "text", + "content": " is expected to balance the category distribution and dynamically distinguish between easy and hard synthetic samples, by which easy samples no longer overwhelm the distillation process. Formally, for a synthetic data sample " + }, + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{B}" + }, + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "text", + "content": ", its modulating function " + }, + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "inline_equation", + "content": "\\phi (x)" + }, + { + "bbox": [ + 46, + 72, + 287, + 145 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 150, + 287, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 287, + 192 + ], + "type": "interline_equation", + "content": "\\phi (x) = \\underbrace {\\left(1 - \\frac {1}{| \\mathcal {B} |} \\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\mathbb {I} _ {c _ {T} \\left(x ^ {\\prime}\\right) = c _ {T} (x)}\\right)} _ {\\text {d i v e r s i t y - a w a r e b a l a n c i n g}} \\underbrace {\\left(1 - p _ {T} (x)\\right) ^ {\\gamma}} _ {\\text {d i f f i c u l t y - a w a r e b a l a n c i n g}}, \\tag {3}", + "image_path": "4a44923d86717231b59c37b18df1f1b9e1975fd027988394452c49b24b6d9f36.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "inline_equation", + "content": "c_{T}(x)" + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "inline_equation", + "content": "p_{T}(x)" + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": " refer to the predicted category index and probability (or confidence) by the pre-trained teacher model, respectively; " + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{c_T(x') = c_T(x)}" + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": " denotes an indicator function that equals to 1 if the predicted category of " + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": " is the same with that of " + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": ", while otherwise 0; " + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": "We highlight two properties of the modulating function " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": ". Firstly, for the data sample " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": " with a high prediction certainty w.r.t. the teacher model (i.e., considered as an easy sample), " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": " approaches to a low value and thus results in low impact on the task-oriented loss " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Task}}(x)" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": " in Eq. (2). Secondly, when the category distribution of synthetic data samples in " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": " is largely imbalanced as predicted by the teacher network, the sample " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": " of which the category shares with more samples in " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": " is penalized and thus the corresponding " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Task}}(x)" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": " is weakened by " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": "Although Eq. (3) implies that the value of the modulating function " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": " is partially determined by the current replay buffer " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": ", note that " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": " changes dynamically and is also affected by " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": ". This is due to that the term " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\phi(x)\\mathcal{L}_{\\mathrm{Task}}(x)" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": " in Eq. (2) directly optimizes the generator that synthesizes data samples to compose " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": ". In this sense, the balancing in category diversity is maintained during training given the mutual effects of " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": ". We find it particularly important in the data inversion process for balancing categories of data samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "text", + "content": "With both balancing terms, as shown in Fig. 2, our method (SSD-KD) generates mild sample distributions in terms of both sample category and difficulty." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 232, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 232, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 232, + 565 + ], + "type": "text", + "content": "3.4. Distillation with Priority Sampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "The original prioritized experience replay method [26] reuses important transitions more frequently and learns more efficiently. Differently, rather than obtaining the reward from the environment, our prioritized sampling method is designed to fit in data-free knowledge distillation and get feedback from the framework itself. In other words, the prioritized sampling method performs the opposite role of the previous data-free knowledge distillation methods: it focuses on training a sparse set of highly prioritized samples instead of uniform sampling to speed up training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "By Eq. (3), we sample the synthetic data " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " from the current replay buffer " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ". Instead of uniformly sampling " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", we" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 306, + 88, + 545, + 513 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 409, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 409, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 409, + 87 + ], + "type": "text", + "content": "Algorithm 1: SSD-KD" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "lines": [ + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": "Input: Number of training epochs " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": "; number of iterations for data inversion " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": "; number of iterations for distillation " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "T_{kd}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": "; pre-trained teacher model " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "f_{t}(\\cdot;\\theta_{t})" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": "; student network " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "f_{s}(\\cdot;\\theta_{s})" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \nOutput: The optimized student model " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "f_{s}(\\cdot;\\theta_{s})" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n1 Initialize replay buffer " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\emptyset" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "f_{s}(\\cdot;\\theta_{s})" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n2 for " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "e = 1" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " do \n3 (1) Data Inversion with Distribution Balancing: \n4 Initialize a generative network " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "g(\\cdot;\\theta_{g})" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n5 " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "z \\gets \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n6 for " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " do \n7 " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "x \\gets g(z;\\theta_{g})" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n8 Compute the overall loss by Eq. (2) \n9 Update " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n10 end \n11 (2) Distillation with Priority Sampling: \n12 Compute the importance-sampling weight " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "w_{i-1}(x)" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n13 Compute the sample priority " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\delta_{i}(x)" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n14 " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\mathcal{B} \\cup \\{x,\\delta_{i}(x)\\}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n15 Remove old samples to keep " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " a constant \n16 for " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "j = 1" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "T_{kd}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " do \n17 Sample a small-scale mini-batch " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n18 Compute the distillation loss " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KD}}(x)" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " in Eq. (2), where " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "x \\in M" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n19 Compute the importance-sampling weight " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "w_{i}(x)" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n20 Compute the sample priority " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\delta_{i+1}(x)" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n21 Update old samples " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\{x,\\delta_{i}(x)\\}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\{x,\\delta_{i+1}(x)\\}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n22 Update " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "text", + "content": " \n23 end \n24 end \n25 Return " + }, + { + "bbox": [ + 306, + 88, + 545, + 513 + ], + "type": "inline_equation", + "content": "f_{s}(\\cdot;\\theta_{s})" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "text", + "content": "propose to modulate the sampling probability by a sampling strategy termed Priority Sampling (PS). The basic function of PS is to measure the importance of each sample " + }, + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "text", + "content": ", by which we introduce a priority sampling function " + }, + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\delta_{i}(x)" + }, + { + "bbox": [ + 304, + 537, + 545, + 586 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 338, + 595, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 595, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 338, + 595, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\delta_ {i} (x) = w _ {i - 1} (x) K L \\left(f _ {t} (x; \\theta_ {t}) \\mid \\mid f _ {s} \\left(x; \\theta_ {s}\\right)\\right), \\tag {4}", + "image_path": "94aca3fce63f3e09d042bdd5b867c3ba2b08de4d63b8964fec5d589526909439.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "where as mentioned in Sec. 3.1, " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " denotes the KL-divergence between the softmax outputs of logits " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "f_{t}(x;\\theta_{t})" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "f_{s}(x;\\theta_{s})" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " depend on the training step " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "w_{i}(x)" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " is the calibration term [26] for normalizing samples in " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ", as will be formalized in Eq. (5), especially, when " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "i = 0" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "w_{-1}(x) = 1" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The training of knowledge distillation with random updates relies on those updates corresponding to the same dis" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6012" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 121, + 471, + 224 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "type": "text", + "content": "Table 1. Performance comparison of Fast2 [11] (the current most efficient D-KD method) and our SSD-KD, in terms of top-1 classification accuracy " + }, + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "type": "text", + "content": " and overall training time cost (hours). Our SSD-KD performs with a very small training data scale: 5000 synthetic samples, i.e., " + }, + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "type": "text", + "content": " of the original training dataset size. All image classification results in this and the other tables are averaged over three independent runs, and the methods \"Teacher\" and \"Student\" are performed on the whole original training dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 121, + 471, + 224 + ], + "lines": [ + { + "bbox": [ + 122, + 121, + 471, + 224 + ], + "spans": [ + { + "bbox": [ + 122, + 121, + 471, + 224 + ], + "type": "table", + "html": "
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
Fast2 [11]92.62(3.92h)84.67(3.02h)88.36(1.46h)89.56(2.37h)89.68(1.70h)
SSD-KD92.92(1.11h)89.56(0.77h)88.07(0.77h)90.12(0.96h)91.84(0.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
Fast2 [11]69.76(4.02h)62.83(2.99h)41.77(1.44h)53.15(2.26h)57.08(1.72h)
SSD-KD73.05(1.12h)66.66(0.78h)51.96(0.77h)59.60(0.97h)61.44(0.78h)
", + "image_path": "c0696876f5e6796947af7c73ee2bfb4671840db42853c624c66a957cbf21d19c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "content": "distribution as its expectation. Prioritized sampling data introduces bias since it might change the data distribution, and affect the solution that the estimates will converge to. Thus we correct the bias by introducing an importance-sampling (IS) weight " + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "inline_equation", + "content": "w_{i}(x)" + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "content": " for the data sample " + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 310, + 287, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 310, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 116, + 310, + 287, + 324 + ], + "type": "interline_equation", + "content": "w _ {i} (x) = \\left(N \\cdot P _ {i} (x)\\right) ^ {- \\beta}, \\tag {5}", + "image_path": "ede23f567dd3f263e445a04ed650ce989dba260f0a2a76670fec9cf2d1aab1ea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "text", + "content": " is a hyper-parameter; " + }, + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "inline_equation", + "content": "P_{i}(x)" + }, + { + "bbox": [ + 46, + 333, + 288, + 358 + ], + "type": "text", + "content": " is the probability of sampling transition defined by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 365, + 287, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 365, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 97, + 365, + 287, + 397 + ], + "type": "interline_equation", + "content": "P _ {i} (x) = \\frac {\\left(\\left| \\delta_ {i} (x) \\right| + \\epsilon\\right) ^ {\\alpha}}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\left(\\left| \\delta_ {i} \\left(x ^ {\\prime}\\right) \\right| + \\epsilon\\right) ^ {\\alpha}}, \\tag {6}", + "image_path": "2ab54ece646fd77c05fe9286b20ea206f8e732f5296a69e761f1bb5472ec138f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "type": "text", + "content": " is a small positive constant which prevents the edge-case of transitions not being selected once their priority is zero." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "text", + "content": "The priority sampling function " + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "inline_equation", + "content": "\\delta(x)" + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "text", + "content": " has two noteworthy properties. Firstly, as the delta value increases, " + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "inline_equation", + "content": "\\delta(x)" + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "text", + "content": " reflects a greater information discrepancy between the teacher and student models for the synthetic samples in the current " + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "text", + "content": ". The student model should therefore be optimized from samples with greater information discrepancy, as this facilitates the faster acquisition of the teacher model. Secondly, " + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "inline_equation", + "content": "\\delta(x)" + }, + { + "bbox": [ + 46, + 437, + 288, + 604 + ], + "type": "text", + "content": " dynamically changes with each update iteration of the student and generative models. Consequently, when the student model acquires the teacher model's capabilities on certain samples, it continues to learn from samples with larger differences relative to the teacher model based on the new sample distribution. This further enhances the performance of the student model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 615, + 124, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 124, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 124, + 628 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 635, + 169, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 169, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 169, + 647 + ], + "type": "text", + "content": "4.1. Experimental Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "We conduct comprehensive experiments on image classification and semantic segmentation tasks to evaluate the effectiveness and study the design of our SSD-KD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Datasets. For image classification experiments, we use CIFAR-10 and CIFAR-100 [14], two most popular datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 241, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 277 + ], + "type": "text", + "content": "for D-KD research. For semantic segmentation experiments, we use the NYUv2 dataset [28]. On each dataset, the baseline models are trained with its standard data split." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 277, + 546, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 546, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 546, + 410 + ], + "type": "text", + "content": "Training Setups. We follow the settings of [11] for basic experiments and comparisons. Regarding experiments on the CIFAR-10 and CIFAR-100 datasets, we use 5 different teacher-student model pairs having either the same type or different type network architectures (see Table 1, 2). Regarding experiments on the NYUv2 dataset, we use two Deeplabv3 models [6] as a teacher-student model pair (see Table 3). Unless otherwise stated, we always adopt the same basic settings as in [11] for experiments, including the number of training epochs, the optimizer, the weight decay, etc. Detailed settings are put in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 410, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 547, + 589 + ], + "type": "text", + "content": "Evaluation Metrics. Besides comparing the student model accuracy (top-1 accuracy for image classification and mean Intersection over Union (IoU) for semantic segmentation), we also compare the overall training time cost of existing mainstream D-KD methods [4, 7, 10, 11, 20, 30] and our SSD-KD. For each teacher-student model pair, we meticulously record the total time cost (hours) for each run of the end-to-end training by all methods. In order to guarantee a fair comparison, all training speed assessments are performed on 1 NVIDIA V100 GPU using 12 cores of Intel Xeon Gold 6240R CPU. Our experiments are implemented with the PyTorch [23] library. For each teacher-student model pair, the experiment is conducted with three independent runs for all methods, and in comparison we report the averaged results, unless otherwise stated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 598, + 430, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 430, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 430, + 611 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Results on image classification task. As the main focus of our work is to improve the overall training efficiency of the adversarial data-free knowledge distillation paradigm, we first compare the proposed SSD-KD with the current most efficient D-KD method Fast2 [11] on the CIFAR-10 and CIFAR-100 datasets. From the results shown in Table 1, we can observe: (1) on the CIFAR-10 dataset, our SSD-KD (using a very small data scale: 5000 synthetic samples," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6013" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 99, + 507, + 333 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Performance comparison of existing top-performing D-KD methods and our SSD-KD, in terms of top-1 classification accuracy " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": " and overall training time cost (hours). For our SSD-KD, we relax the data scale of synthetic samples to be similar to that for Fast5." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 99, + 507, + 333 + ], + "lines": [ + { + "bbox": [ + 85, + 99, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 85, + 99, + 507, + 333 + ], + "type": "table", + "html": "
DatasetMethodResNet-34 ResNet-18VGG-11 ResNet-18WRN40-2 WRN16-1WRN40-2 WRN40-1WRN40-2 WRN16-2
CIFAR-10Teacher95.7092.2594.8794.8794.87
Student95.2092.2091.1293.9493.95
DeepInv [30]93.26(25.62h)90.36(13.19h)83.04(10.62h)86.85(13.84h)89.72(11.67h)
CMI [10]94.84(15.49h)91.13(10.99h)90.01(9.87h)92.78(11.57h)92.52(10.58h)
DAFL [4]92.22(303.50h)81.10(209.78h)65.71(96.57h)81.33(157.44h)81.55(118.52h)
DFQ [7]94.61(152.16h)90.84(104.93h)86.14(48.40h)91.69(78.81h)92.01(59.40h)
ZSKT [20]93.32(304.11h)89.46(209.65h)83.74(96.75h)86.07(157.51h)89.66(118.82h)
Fast5 [11]93.63(4.34h)89.94(3.15h)88.90(1.61h)92.04(2.44h)91.96(1.89h)
Fast10 [11]94.05(4.72h)90.53(3.22h)89.29(1.72h)92.51(2.48h)92.45(2.05h)
SSD-KD94.26(4.20h)90.67(2.95h)89.96(1.48h)93.23(2.30h)93.11(1.78h)
CIFAR-100Teacher78.0571.3275.8375.8375.83
Student77.1077.1065.3172.1973.56
DeepInv [30]61.32(25.88h)54.13(13.13h)53.77(10.62h)61.33(13.86h)61.34(11.68h)
CMI [10]77.04(15.42h)70.56(11.28h)57.91(10.01h)68.88(11.57h)68.75(10.53h)
DAFL [4]74.47(303.54h)54.16(209.82h)20.88(96.62h)42.83(157.49h)43.70(118.67h)
DFQ [7]77.01(152.18h)66.21(105.14h)51.27(48.47h)54.43(78.83h)64.79(59.48h)
ZSKT [20]67.74(304.13h)54.31(209.63h)36.66(96.87h)53.60(157.68h)54.59(118.84h)
Fast5 [11]72.82(4.35h)65.28(3.06h)52.90(1.56h)61.80(2.35h)63.83(1.86h)
Fast10 [11]74.34(4.50h)67.44(3.12h)54.02(1.65h)63.91(2.42h)65.12(1.95h)
SSD-KD75.16(4.22h)68.77(2.94h)55.61(1.52h)64.57(2.27h)65.28(1.78h)
", + "image_path": "46674af9fe3ac0c73e7aa69ddad434e8ec6791fbb3a6f527594581d972325838.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "text", + "content": "i.e., " + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "text", + "content": " of the original training dataset size) shows large improvements in training efficiency to Fast2 (at least " + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "inline_equation", + "content": "1.90\\times" + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "text", + "content": " and at most " + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "inline_equation", + "content": "3.92\\times" + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "text", + "content": " training speedup, on 5 teacher-student model pairs), and gets better student models on 4 out of 5 teacher-student model pairs; (2) on the more challenging CIFAR-100 dataset, our SSD-KD shows a very similar training speedup trend against Fast2 as on the CIFAR-10 dataset, but gets significantly better student models on all 5 teacher-student model pairs (at least " + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "inline_equation", + "content": "3.29\\%" + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "text", + "content": " and at most " + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "inline_equation", + "content": "10.19\\%" + }, + { + "bbox": [ + 46, + 350, + 289, + 530 + ], + "type": "text", + "content": " absolute top-1 accuracy gain to Fast2). The superior performance of SSD-KD against Fast2 validates the efficacy of our small-scale data inversion and sampling mechanism which can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 531, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 664 + ], + "type": "text", + "content": "Next, we compare the proposed SSD-KD with existing top-performing D-KD methods including DeepInv [30], CMI [10], DAFL [4], DFQ [7], ZSKT [20], Fast5 and Fast10 [11]. In order to get improved student model performance, we relax the data scale of synthetic samples in our SSD-KD to be similar to that for Fast5. Experimental results are summarized in Table 2. Compared to these mainstream D-KD methods, our SSD-KD gets very competitive performance in student model accuracy, and significantly better performance in overall training efficiency, on all 5 teacher-student model pairs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "In light of the results in Table 1, 2, it can be seen that our SSD-KD makes the overall training efficiency one or two orders of magnitude faster than most D-KD methods while retaining competitive student model performance." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 311, + 398, + 538, + 488 + ], + "blocks": [ + { + "bbox": [ + 305, + 349, + 545, + 393 + ], + "lines": [ + { + "bbox": [ + 305, + 349, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 545, + 393 + ], + "type": "text", + "content": "Table 3. Performance comparison on the NYUv2 dataset. The teacher model is pre-trained on the ImageNet dataset and finetuned on the NYUv2 dataset, and the student model is trained from scratch. The results of reference methods are collected from [11]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 398, + 538, + 488 + ], + "lines": [ + { + "bbox": [ + 311, + 398, + 538, + 488 + ], + "spans": [ + { + "bbox": [ + 311, + 398, + 538, + 488 + ], + "type": "table", + "html": "
MethodTraining Data ScalemIoU (%)
Teacher: Deeplabv3-ResNet501,449 NYUv20.519
Student: Deeplabv3-Mobilenetv21,449 NYUv20.375
KD [12]1,449 NYUv20.380
DFND [5]14M (ImageNet)0.378
DFAD [9]960K (synthetic)0.364
DAFL [4]960K (synthetic)0.105
Fast10 [11]17K (synthetic)0.366
SSD-KD16K (synthetic)0.384
", + "image_path": "764582cddbb4a91d1e6abbc095f47ebc6138d04d894ceb4175f93c11df55f316.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 496, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 639 + ], + "type": "text", + "content": "Results on semantic segmentation task. To further validate the generalization ability of our method, we compare the performance of SSD-KD with existing D-KD methods on the NYUv2 dataset. Table 3 summarizes the results. We can see that our SSD-KD not only achieves state-of-the-art performance in terms of model accuracy, but also is significantly more efficient than other D-KD methods in terms of the training data scale. The overall training time cost for our SSD-KD and Fast10 is 8.9 hours and 9.5 hours, respectively. Besides, the student model trained by SSD-KD even outperforms the baseline model trained on the original training data, as NYUv2 is a small-scale dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 647, + 406, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 406, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 406, + 658 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Effect of the synthetic data scale. So far, we have already demonstrated that SSD-KD has the appealing capability to attain efficient and effective end-to-end distillation training using a small amount of synthetic samples. To bet-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6014" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 170, + 158 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 170, + 158 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 170, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 170, + 158 + ], + "type": "image", + "image_path": "5fe131909012347caf67092652fbfd4f1caf4b0101025930b6ce32e24131f577.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "type": "text", + "content": "Figure 4. Performance comparison of SSD-KD under different synthetic data scales against the original training dataset size, in terms of top-1 classification accuracy " + }, + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "type": "text", + "content": " and overall training time cost (hour)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 69, + 294, + 158 + ], + "blocks": [ + { + "bbox": [ + 181, + 69, + 294, + 158 + ], + "lines": [ + { + "bbox": [ + 181, + 69, + 294, + 158 + ], + "spans": [ + { + "bbox": [ + 181, + 69, + 294, + 158 + ], + "type": "image", + "image_path": "9c9704e1d98650e5f15730a4013d278788f6eaa4fc5a2214531fa9a3dad6e6b1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 421, + 158 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 421, + 158 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 421, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 421, + 158 + ], + "type": "image", + "image_path": "898ebb0651ce575f43792e1a95ed211c7c757e0408b20edd57290557101d66a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 432, + 68, + 547, + 158 + ], + "blocks": [ + { + "bbox": [ + 432, + 68, + 547, + 158 + ], + "lines": [ + { + "bbox": [ + 432, + 68, + 547, + 158 + ], + "spans": [ + { + "bbox": [ + 432, + 68, + 547, + 158 + ], + "type": "image", + "image_path": "cf55f48331c8c1e62490291f08d16b98e3030536f036ffeac314115379f670b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 219, + 433, + 361 + ], + "blocks": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "type": "text", + "content": "Table 4. Effect of two core modules in SSD-KD. PS: the priority sampling function, Difficulty|Diversity: the difficulty|diversity-aware balancing term in the modulating function, see Eq. (3)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 219, + 433, + 361 + ], + "lines": [ + { + "bbox": [ + 159, + 219, + 433, + 361 + ], + "spans": [ + { + "bbox": [ + 159, + 219, + 433, + 361 + ], + "type": "table", + "html": "
DatasetBaselinePSDifficultyDiversityResNet-34 ResNet-18WRN40-2 WRN16-1
CIFAR-10✓87.35(1.19h)80.47(0.84h)
✓✓90.28(1.01h)87.19(0.69h)
✓✓88.57(1.20h)83.12(0.86h)
✓✓89.12(1.21h)84.31(0.87h)
✓✓✓89.95(1.21h)85.79(0.88h)
✓✓✓✓92.92(1.11h)88.07(0.77h)
CIFAR-100✓64.05(1.19h)41.67(0.84h)
✓✓71.70(1.01h)49.70(0.69h)
✓✓66.94(1.20h)43.96(0.86h)
✓✓68.52(1.21h)46.34(0.87h)
✓✓✓70.85(1.21h)47.82(0.88h)
✓✓✓✓73.05(1.12h)51.96(0.77h)
", + "image_path": "070bc5100ff68b804667b0def2d96b9eb50a38cc19f5f5f2887d2060748995a6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 69, + 380, + 267, + 489 + ], + "blocks": [ + { + "bbox": [ + 69, + 380, + 267, + 489 + ], + "lines": [ + { + "bbox": [ + 69, + 380, + 267, + 489 + ], + "spans": [ + { + "bbox": [ + 69, + 380, + 267, + 489 + ], + "type": "image", + "image_path": "a5d5d534d99b2af5096176bece819beed0552b8ab90d64915bec0d9892f33e6f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "type": "text", + "content": "Figure 5. Visualization examples of synthetic image samples generated by Fast10 and our SSD-KD for the NYUv2 dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": "ter explore the boundary of this capability, we conduct an ablation with three teacher-student model pairs, including ResNet34 " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " ResNet18, VGG11 " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " ResNet18, and WRN40-2 " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " WRN16-1. In the experiments, we decrease the synthetic data scale used in SSD-KD [11] from 50,000(100% relative to the original training data size) to {10000(20%), 5000(10%), 2500(%5), 500(1%)} for the CIFAR-10 and CIFAR-100 datasets. As shown in Fig. 4, the accuracy of the student model trained by SSD-KD remains stable across a relatively large synthetic data scale range, for all teacher-student model pairs. When decreasing the synthetic data scale, the overall training time cost appears to decrease nearly linearly, while the student model accuracy drop is mild (less than 10% even for the synthetic data scale 500)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Effect of the core modules. In Table 4, we provide an ablation to scrutinize our SSD-KD systematically. We" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 441 + ], + "type": "text", + "content": "observe that: (1) the two modulating functions (consisting of a diversity-aware term and a difficulty-aware term, see Eq. (3)) and the priority sampling function, are both critical to our SSD-KD; (2) the combination of them strikes a good tradeoff between model accuracy and training efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 441, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 487 + ], + "type": "text", + "content": "Visualization of data inversion. Fig. 5 shows examples of synthetic images inverted by Fast10 and our SSD-KD for the NYUv2 dataset. Compared to Fast10, our method can better invert texture information and has less noise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 498, + 379, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 379, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 379, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": "In this paper, we presented SSD-KD, the first fully efficient method to advance adversarial data-free knowledge distillation research. Benefiting from a small-scale data inversion and sampling mechanism based on a modulating function and a priority sampling function, SSD-KD can flexibly balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes, attaining efficient and effective data-free knowledge distillation. Extensive experiments on image classification and semantic segmentation benchmarks validate the efficacy of SSD-KD. We hope our work can inspire future research on efficient D-KD designs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 669, + 416, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 416, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 416, + 682 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Fund for Key International Collaboration (62120106005)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6015" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI Pieter Abbeel, and Wojciech Zaremba. Hind-sight experience replay. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 157 + ], + "type": "text", + "content": "[2] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 286, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 191 + ], + "type": "text", + "content": "[3] Kartikeya Bhardwaj, Naveen Suda, and Radu Marculescu. Dream distillation: A data-independent model compression framework. arXiv preprint arXiv:1905.07072, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 286, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 235 + ], + "type": "text", + "content": "[4] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 268 + ], + "type": "text", + "content": "[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In CVPR, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 312 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 286, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 356 + ], + "type": "text", + "content": "[7] Yoojin Choi, Jihwan Choi, Mostafa El-Khamy, and Jungwon Lee. Data-free network quantization with adversarial knowledge distillation. In CVPR Workshops, 2020. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 286, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 286, + 412 + ], + "type": "text", + "content": "[8] Kien Do, Thai Hung Le, Dung Nguyen, Dang Nguyen, Haripriya Harikumar, Truyen Tran, Santu Rana, and Svetha Venkatesh. Momentum adversarial distillation: Handling large distribution shifts in data-free knowledge distillation. NeurIPS, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 286, + 446 + ], + "type": "text", + "content": "[9] Gongfan Fang, Jie Song, Chengchao Shen, Xinchao Wang, Da Chen, and Mingli Song. Data-free adversarial distillation. arXiv preprint arXiv:1912.11006, 2019. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 286, + 491 + ], + "type": "text", + "content": "[10] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. arXiv preprint arXiv:2105.08584, 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "text", + "content": "[11] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In AAAI, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "text", + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 286, + 612 + ], + "type": "text", + "content": "[13] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 635 + ], + "type": "text", + "content": "[14] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Tech Report, 2009. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 679 + ], + "type": "text", + "content": "[15] Yuhang Li, Feng Zhu, Ruihao Gong, Mingzhu Shen, Xin Dong, Fengwei Yu, Shaoqing Lu, and Shi Gu. Mixmix: All you need for data-free compression are feature and data mixing. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Yifan Liu, Ke Chen, Chris Liu, Zengchang Qin, Zhenbo Luo, and Jingdong Wang. Structured knowledge distillation for semantic segmentation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[18] Raphael Gontijo Lopes, Stefano Fenu, and Thad Starner. Data-free knowledge distillation for deep neural networks. arXiv preprint arXiv:1710.07535, 2017. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "text", + "content": "[19] Liangchen Luo, Mark Sandler, Zi Lin, Andrey Zhmoginov, and Andrew Howard. Large-scale generative data-free distillation. arXiv preprint arXiv:2012.05578, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[20] Paul Micaelli and Amos J Storkey. Zero-shot knowledge transfer via adversarial belief matching. NeurIPS, 2019. 4, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "text", + "content": "[21] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In ICML, 2019. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 250, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 272 + ], + "type": "text", + "content": "[22] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 272, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 358 + ], + "type": "text", + "content": "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "text", + "content": "[24] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In CVPR, 2023. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "text", + "content": "[25] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "text", + "content": "[26] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "text", + "content": "[27] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[28] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[29] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[30] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 625, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 656 + ], + "type": "text", + "content": "[31] Jaemin Yoo, Minyong Cho, Taebum Kim, and U Kang. Knowledge extraction with no observable data. NeurIPS, 2019. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[32] Shikang Yu, Jiachen Chen, Hu Han, and Shuqiang Jiang. Data-free knowledge distillation via feature exchange and activation region constraint. In CVPR, 2023. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6016" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file